diff --git a/.gitattributes b/.gitattributes index b9cdac5016ed25008fec12bbff17c640bd6b791f..b556a2be7c8a1f72f86c097ff99db6a57a55c051 100644 --- a/.gitattributes +++ b/.gitattributes @@ -796,3 +796,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text 2025/Embodied[[:space:]]Navigation[[:space:]]with[[:space:]]Auxiliary[[:space:]]Task[[:space:]]of[[:space:]]Action[[:space:]]Description[[:space:]]Prediction/a8259b3c-f4b9-4360-9aef-015762e621d2_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Embodied[[:space:]]Representation[[:space:]]Alignment[[:space:]]with[[:space:]]Mirror[[:space:]]Neurons/1a9ab08d-9a2c-43fd-94f5-5038fe26d6e1_origin.pdf filter=lfs diff=lfs merge=lfs -text 2025/Embodied[[:space:]]VideoAgent_[[:space:]]Persistent[[:space:]]Memory[[:space:]]from[[:space:]]Egocentric[[:space:]]Videos[[:space:]]and[[:space:]]Embodied[[:space:]]Sensors[[:space:]]Enables[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Understanding/a017345c-038c-4d4f-96f4-badf8dc4eef1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EmbodiedOcc_[[:space:]]Embodied[[:space:]]3D[[:space:]]Occupancy[[:space:]]Prediction[[:space:]]for[[:space:]]Vision-based[[:space:]]Online[[:space:]]Scene[[:space:]]Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EmbodiedSplat_[[:space:]]Personalized[[:space:]]Real-to-Sim-to-Real[[:space:]]Navigation[[:space:]]with[[:space:]]Gaussian[[:space:]]Splats[[:space:]]from[[:space:]]a[[:space:]]Mobile[[:space:]]Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EmotiCrafter_[[:space:]]Text-to-Emotional-Image[[:space:]]Generation[[:space:]]based[[:space:]]on[[:space:]]Valence-Arousal[[:space:]]Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Emulating[[:space:]]Self-attention[[:space:]]with[[:space:]]Convolution[[:space:]]for[[:space:]]Efficient[[:space:]]Image[[:space:]]Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/End-to-End[[:space:]]Driving[[:space:]]with[[:space:]]Online[[:space:]]Trajectory[[:space:]]Evaluation[[:space:]]via[[:space:]]BEV[[:space:]]World[[:space:]]Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/End-to-End[[:space:]]Entity-Predicate[[:space:]]Association[[:space:]]Reasoning[[:space:]]for[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Graph[[:space:]]Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/End-to-End[[:space:]]Multi-Modal[[:space:]]Diffusion[[:space:]]Mamba/229968bf-b390-4563-8550-16e090dc25cb_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Engage[[:space:]]for[[:space:]]All_[[:space:]]Making[[:space:]]Ordinary[[:space:]]Image[[:space:]]Descriptions[[:space:]]Appealing[[:space:]]Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhanced[[:space:]]Event-based[[:space:]]Dense[[:space:]]Stereo[[:space:]]via[[:space:]]Cross-Sensor[[:space:]]Knowledge[[:space:]]Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhanced[[:space:]]Pansharpening[[:space:]]via[[:space:]]Quaternion[[:space:]]Spatial-Spectral[[:space:]]Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Adversarial[[:space:]]Transferability[[:space:]]by[[:space:]]Balancing[[:space:]]Exploration[[:space:]]and[[:space:]]Exploitation[[:space:]]with[[:space:]]Gradient-Guided[[:space:]]Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Few-Shot[[:space:]]Vision-Language[[:space:]]Classification[[:space:]]with[[:space:]]Large[[:space:]]Multimodal[[:space:]]Model[[:space:]]Features/958199ee-2123-4605-9c2c-3f90d0a78ead_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Image[[:space:]]Restoration[[:space:]]Transformer[[:space:]]via[[:space:]]Adaptive[[:space:]]Translation[[:space:]]Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Mamba[[:space:]]Decoder[[:space:]]with[[:space:]]Bidirectional[[:space:]]Interaction[[:space:]]in[[:space:]]Multi-Task[[:space:]]Dense[[:space:]]Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Numerical[[:space:]]Prediction[[:space:]]of[[:space:]]MLLMs[[:space:]]with[[:space:]]Soft[[:space:]]Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Partially[[:space:]]Relevant[[:space:]]Video[[:space:]]Retrieval[[:space:]]with[[:space:]]Hyperbolic[[:space:]]Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Prompt[[:space:]]Generation[[:space:]]with[[:space:]]Adaptive[[:space:]]Refinement[[:space:]]for[[:space:]]Camouflaged[[:space:]]Object[[:space:]]Detection/31897efc-f494-43de-9990-b3b0b058743c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Reward[[:space:]]Models[[:space:]]for[[:space:]]High-quality[[:space:]]Image[[:space:]]Generation_[[:space:]]Beyond[[:space:]]Text-Image[[:space:]]Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Spatial[[:space:]]Reasoning[[:space:]]in[[:space:]]Multimodal[[:space:]]Large[[:space:]]Language[[:space:]]Models[[:space:]]through[[:space:]]Reasoning-based[[:space:]]Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Transferability[[:space:]]of[[:space:]]Targeted[[:space:]]Adversarial[[:space:]]Examples[[:space:]]via[[:space:]]Inverse[[:space:]]Target[[:space:]]Gradient[[:space:]]Competition[[:space:]]and[[:space:]]Spatial[[:space:]]Distance[[:space:]]Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Transformers[[:space:]]Through[[:space:]]Conditioned[[:space:]]Embedded[[:space:]]Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enhancing[[:space:]]Zero-shot[[:space:]]Object[[:space:]]Counting[[:space:]]via[[:space:]]Text-guided[[:space:]]Local[[:space:]]Ranking[[:space:]]and[[:space:]]Number-evoked[[:space:]]Global[[:space:]]Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enpowering[[:space:]]Your[[:space:]]Pansharpening[[:space:]]Models[[:space:]]with[[:space:]]Generalizability_[[:space:]]Unified[[:space:]]Distribution[[:space:]]is[[:space:]]All[[:space:]]You[[:space:]]Need/84a6c0dd-500e-40a7-b080-833571baa669_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Enrich[[:space:]]and[[:space:]]Detect_[[:space:]]Video[[:space:]]Temporal[[:space:]]Grounding[[:space:]]with[[:space:]]Multimodal[[:space:]]LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Ensemble[[:space:]]Foreground[[:space:]]Management[[:space:]]for[[:space:]]Unsupervised[[:space:]]Object[[:space:]]Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Entropy-Adaptive[[:space:]]Diffusion[[:space:]]Policy[[:space:]]Optimization[[:space:]]with[[:space:]]Dynamic[[:space:]]Step[[:space:]]Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Environment-Agnostic[[:space:]]Pose_[[:space:]]Generating[[:space:]]Environment-independent[[:space:]]Object[[:space:]]Representations[[:space:]]for[[:space:]]6D[[:space:]]Pose[[:space:]]Estimation/5521ae38-4b00-475c-a94f-31fec5839328_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Epipolar[[:space:]]Consistent[[:space:]]Attention[[:space:]]Aggregation[[:space:]]Network[[:space:]]for[[:space:]]Unsupervised[[:space:]]Light[[:space:]]Field[[:space:]]Disparity[[:space:]]Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Epona_[[:space:]]Autoregressive[[:space:]]Diffusion[[:space:]]World[[:space:]]Model[[:space:]]for[[:space:]]Autonomous[[:space:]]Driving/96492e1e-bfce-4329-bd89-949aa9fef588_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EquiCaps_[[:space:]]Predictor-Free[[:space:]]Pose-Aware[[:space:]]Pre-Trained[[:space:]]Capsule[[:space:]]Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Equipping[[:space:]]Vision[[:space:]]Foundation[[:space:]]Model[[:space:]]with[[:space:]]Mixture[[:space:]]of[[:space:]]Experts[[:space:]]for[[:space:]]Out-of-Distribution[[:space:]]Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Erasing[[:space:]]More[[:space:]]Than[[:space:]]Intended_[[:space:]]How[[:space:]]Concept[[:space:]]Erasure[[:space:]]Degrades[[:space:]]the[[:space:]]Generation[[:space:]]of[[:space:]]Non-Target[[:space:]]Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Error[[:space:]]Recognition[[:space:]]in[[:space:]]Procedural[[:space:]]Videos[[:space:]]using[[:space:]]Generalized[[:space:]]Task[[:space:]]Graph/5a14de37-2719-437f-aefb-79f29e2ef354_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Estimating[[:space:]]2D[[:space:]]Camera[[:space:]]Motion[[:space:]]with[[:space:]]Hybrid[[:space:]]Motion[[:space:]]Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EvRT-DETR_[[:space:]]Latent[[:space:]]Space[[:space:]]Adaptation[[:space:]]of[[:space:]]Image[[:space:]]Detectors[[:space:]]for[[:space:]]Event-based[[:space:]]Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EvaGaussians_[[:space:]]Event[[:space:]]Stream[[:space:]]Assisted[[:space:]]Gaussian[[:space:]]Splatting[[:space:]]from[[:space:]]Blurry[[:space:]]Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Evading[[:space:]]Data[[:space:]]Provenance[[:space:]]in[[:space:]]Deep[[:space:]]Neural[[:space:]]Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-Driven[[:space:]]Storytelling[[:space:]]with[[:space:]]Multiple[[:space:]]Lifelike[[:space:]]Humans[[:space:]]in[[:space:]]a[[:space:]]3D[[:space:]]Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-aided[[:space:]]Dense[[:space:]]and[[:space:]]Continuous[[:space:]]Point[[:space:]]Tracking_[[:space:]]Everywhere[[:space:]]and[[:space:]]Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-based[[:space:]]Tiny[[:space:]]Object[[:space:]]Detection_[[:space:]]A[[:space:]]Benchmark[[:space:]]Dataset[[:space:]]and[[:space:]]Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-based[[:space:]]Visual[[:space:]]Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-boosted[[:space:]]Deformable[[:space:]]3D[[:space:]]Gaussians[[:space:]]for[[:space:]]Dynamic[[:space:]]Scene[[:space:]]Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-guided[[:space:]]HDR[[:space:]]Reconstruction[[:space:]]with[[:space:]]Diffusion[[:space:]]Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Event-guided[[:space:]]Unified[[:space:]]Framework[[:space:]]for[[:space:]]Low-light[[:space:]]Video[[:space:]]Enhancement,[[:space:]]Frame[[:space:]]Interpolation,[[:space:]]and[[:space:]]Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EventUPS_[[:space:]]Uncalibrated[[:space:]]Photometric[[:space:]]Stereo[[:space:]]Using[[:space:]]an[[:space:]]Event[[:space:]]Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Everything[[:space:]]is[[:space:]]a[[:space:]]Video_[[:space:]]Unifying[[:space:]]Modalities[[:space:]]through[[:space:]]Next-Frame[[:space:]]Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Evidential[[:space:]]Knowledge[[:space:]]Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/EvolvingGrasp_[[:space:]]Evolutionary[[:space:]]Grasp[[:space:]]Generation[[:space:]]via[[:space:]]Efficient[[:space:]]Preference[[:space:]]Alignment/a5b9e976-b527-49c5-a219-2287871511f8_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ExCap3D_[[:space:]]Expressive[[:space:]]3D[[:space:]]Scene[[:space:]]Understanding[[:space:]]via[[:space:]]Object[[:space:]]Captioning[[:space:]]with[[:space:]]Varying[[:space:]]Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Explaining[[:space:]]Human[[:space:]]Preferences[[:space:]]via[[:space:]]Metrics[[:space:]]for[[:space:]]Structured[[:space:]]3D[[:space:]]Reconstruction/daa34d93-214e-4447-9b72-957738436c31_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Diffusion[[:space:]]Prior[[:space:]]for[[:space:]]Task-driven[[:space:]]Image[[:space:]]Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Domain[[:space:]]Properties[[:space:]]in[[:space:]]Language-Driven[[:space:]]Domain[[:space:]]Generalization[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Frequency[[:space:]]Dynamics[[:space:]]for[[:space:]]Enhanced[[:space:]]Multimodal[[:space:]]Event-based[[:space:]]Action[[:space:]]Recognition/ff885410-5312-499d-bf42-f2819adc9e95_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploiting[[:space:]]Vision[[:space:]]Language[[:space:]]Model[[:space:]]for[[:space:]]Training-Free[[:space:]]3D[[:space:]]Point[[:space:]]Cloud[[:space:]]OOD[[:space:]]Detection[[:space:]]via[[:space:]]Graph[[:space:]]Score[[:space:]]Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/ExploreGS_[[:space:]]Explorable[[:space:]]3D[[:space:]]Scene[[:space:]]Reconstruction[[:space:]]with[[:space:]]Virtual[[:space:]]Camera[[:space:]]Samplings[[:space:]]and[[:space:]]Diffusion[[:space:]]Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Multimodal[[:space:]]Diffusion[[:space:]]Transformers[[:space:]]for[[:space:]]Enhanced[[:space:]]Prompt-based[[:space:]]Image[[:space:]]Editing/db432361-cc2c-467e-9c54-baa03e561a43_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Probabilistic[[:space:]]Modeling[[:space:]]Beyond[[:space:]]Domain[[:space:]]Generalization[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]The[[:space:]]Visual[[:space:]]Feature[[:space:]]Space[[:space:]]for[[:space:]]Multimodal[[:space:]]Neural[[:space:]]Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]View[[:space:]]Consistency[[:space:]]for[[:space:]]Scene-Adaptive[[:space:]]Low-Light[[:space:]]Light[[:space:]]Field[[:space:]]Image[[:space:]]Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]Weather-aware[[:space:]]Aggregation[[:space:]]and[[:space:]]Adaptation[[:space:]]for[[:space:]]Semantic[[:space:]]Segmentation[[:space:]]under[[:space:]]Adverse[[:space:]]Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Exploring[[:space:]]the[[:space:]]Adversarial[[:space:]]Vulnerabilities[[:space:]]of[[:space:]]Vision-Language-Action[[:space:]]Models[[:space:]]in[[:space:]]Robotics/654da609-1a74-4dd3-8326-891848884c54_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Expressive[[:space:]]Talking[[:space:]]Human[[:space:]]from[[:space:]]Single-Image[[:space:]]with[[:space:]]Imperfect[[:space:]]Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/Extending[[:space:]]Foundational[[:space:]]Monocular[[:space:]]Depth[[:space:]]Estimators[[:space:]]to[[:space:]]Fisheye[[:space:]]Cameras[[:space:]]with[[:space:]]Calibration[[:space:]]Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_origin.pdf filter=lfs diff=lfs merge=lfs -text +2025/External[[:space:]]Knowledge[[:space:]]Injection[[:space:]]for[[:space:]]CLIP-Based[[:space:]]Class-Incremental[[:space:]]Learning/59db92d7-ff41-4079-a08f-c630a2445462_origin.pdf filter=lfs diff=lfs merge=lfs -text diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_content_list.json b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..397f2ae457324d9f2237b4628d8a2e53daf46a4e --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_content_list.json @@ -0,0 +1,1548 @@ +[ + { + "type": "text", + "text": "EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding", + "text_level": 1, + "bbox": [ + 235, + 87, + 763, + 133 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuqi Wu Wenzhao Zheng Sicheng Zuo Yuanhui Huang Jie Zhou Jiwen Lu Department of Automation, Tsinghua University, China", + "bbox": [ + 161, + 159, + 836, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wuyq24@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com", + "bbox": [ + 251, + 199, + 743, + 213 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4a6e8029c1021c3b3ea302c23baa602e5f0e9efa46d30e49178be860a011b80.jpg", + "image_caption": [ + "Figure 1. Given streaming monocular RGB inputs, our EmbodiedOcc conducts embodied occupancy prediction in an online manner for indoor scenes. Different from existing methods which focus on offline perception from monocular images, we focus on the scene-level occupancy prediction from embodied observations. We initialize the scene to be explored with uniform 3D semantic Gaussians and progressively update them based on new observations, similar to how humans explore unknown scenes." + ], + "image_footnote": [], + "bbox": [ + 99, + 227, + 903, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 593, + 328, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D", + "bbox": [ + 88, + 627, + 485, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOccScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc.", + "bbox": [ + 511, + 594, + 908, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 799, + 645, + 815 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With the rapid development of embodied intelligence and active agents [14, 17, 32], 3D scene perception [30, 34, 41, 42] has become a crucial task in computer vision. Intelligent agents first perceive their surrounding environments and then make decisions based on the perception results.", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding author.", + "bbox": [ + 112, + 886, + 250, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26360", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the low costs of camera sensors, vision-based 3D occupancy prediction is gaining increasing popularity and produces a comprehensive understanding of both semantics and structures of the scene [2, 11, 13, 46, 56].", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While vision-based 3D occupancy prediction has made significant progress in outdoor driving scenes [11, 13, 22, 40, 43, 45, 46, 52, 58, 59], the application to indoor scenarios is still challenging due to the diversity and complexity of indoor scenes. Most existing methods [2, 54, 56] still focus on local 3D occupancy prediction by integrating semantic and depth information extracted from the visual inputs. However, different from outdoor scenarios, it is important to obtain a global understanding of the room for indoor scenarios, as it usually requires multiple traversals for embodied agents. Also, it is more practical to progressively explore and update the global occupancy of the 3D scene in an online manner from embodied vision-based observations with different positions and perspectives.", + "bbox": [ + 88, + 152, + 482, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To bridge this gap, we formulate a new embodied 3D occupancy prediction task to evaluate the ability to progressively explore an unknown scene using only visual inputs. We propose an EmbodiedOcc framework based on Gaussian memories to accomplish this task, considering the explicit and structural nature of 3D Gaussians. We initialize the global scene with uniform 3D semantic Gaussians and progressively update the Gaussians within the field of view observed by the agent. Throughout the exploration process, we maintain an explicit global memory of 3D Gaussians as the global understanding and derive the global 3D occupancy with Gaussian-to-voxel splatting [13]. Specifically, we propose a structure-aware local refinement module to update the relevant Gaussians within the current frustum. We employs a simple yet effective depth-aware branch to introduce explicit structural information for each Gaussian, ensuring the update of these Gaussians to better align with the global representation. During the continuous exploration, we read out Gaussians within the current frustum from the memory as inputs to the local module for refinement. We assign high confidence values for updated Gaussians and use them to reweight information from the memory and the current input. This ensures the consistency of the 3D representation during the fusion and update process. We reorganize an EmbodiedOcc-ScanNet benchmark for the embodied 3D occupancy prediction task based on the locally annotated Occ-ScanNet dataset [3, 47, 56]. Experiments show that our EmbodiedOcc outperforms existing methods by a large margin and accomplishes embodied occupancy prediction with high accuracy and efficiency.", + "bbox": [ + 91, + 364, + 482, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 829, + 232, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Occupancy Prediction. Benefiting from its compactness and versatility, 3D occupancy prediction based on multi-view images or additional 3D information [1, 11-", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "13, 21, 37, 46] has gained great popularity over the last few years. MonoScene [2] was the first to derive 3D occupancy prediction from a single image, propelling the original 3D Semantic Scene Completion (SSC) [4, 8, 18, 19, 33, 37] into a more challenging stage with vision-only inputs and more universal scenarios (both indoor and outdoor scenes). Subsequent works [54, 56] further focused on addressing the depth ambiguity in this monocular setting. However, most of these efforts were confined to local and offline prediction. SCFusion [47] proposed an incremental framework based on RGB-D inputs. EmbodiedScan [28, 44] introduced an offline global prediction framework using multimodal sequential inputs. Differently, the proposed embodied 3D occupancy prediction aims at online prediction from RGB-only inputs, which is more challenging and practical.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Online 3D Scene Perception. Accurate comprehension of 3D scenes is an indispensable capability for embodied agents, such as 3D occupancy prediction [2, 56] and object detection [16, 31, 42]. Most existing works on indoor 3D scene perception [9, 30, 41, 55] take pre-acquired and reconstructed 3D data as inputs and perceive the scene in an offline manner. To achieve online perception, Online3D [49] introduced an adapter-based model that equips mainstream offline frameworks with the competence to perform online scene perception, enabling the process of real-time RGB-D sequences. However, this framework still requires depth information as inputs and mainly targets point segmentation and 3D detection. Differently, we target online vision-based 3D occupancy prediction which can provide a more comprehensive understanding of the scene.", + "bbox": [ + 511, + 319, + 903, + 545 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting. 3D Gaussian Splatting [15] uses 3D Gaussians to model a 3D scene and benefits from fast speed and high quality in the field of neural rendering. The physical characteristics of 3D Gaussians and the splat-based rasterization also motivated rapid advancements in research fields such as scene editing [10, 24, 29, 36], dynamic scenarios [7, 27, 38, 48, 53], and SLAM [5, 20, 50, 57]. GaussianFormer [13] pioneers the application of 3D Gaussians in outdoor 3D occupancy prediction and uses features from multi-view images to update 3D Gaussians, which can be converted into 3D occupancy prediction through a Gaussian-to-voxel splatting module. However, it is still unclear how to employ 3D Gaussians for online global indoor scene understanding from local observations. We achieve this by designing a Gaussian memory mechanism and progressively updating it with structure-aware interaction.", + "bbox": [ + 511, + 547, + 903, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Approach", + "text_level": 1, + "bbox": [ + 511, + 806, + 705, + 823 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Embodied 3D Occupancy Prediction", + "text_level": 1, + "bbox": [ + 511, + 832, + 828, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventional methods in indoor scenarios for occupancy prediction accepted RGB-D as inputs to predict the semantic occupancy of a 3D scene which requires depth sensors.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26361", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/83329bd7533bd4ca726b7428f30052ba3744d5342ae60be45c107bae2757ea56.jpg", + "image_caption": [ + "Figure 2. Framework of our EmbodiedOcc for embodied 3D occupancy prediction. We maintain an explicit global memory of 3D Gaussians during the exploration of the current scene. For each update, the Gaussians within the current frustum are taken from the memory and updated using semantic and structural features extracted from the monocular RGB input. Each Gaussian has a confidence value to integrate information from both the memory and the current input. Then we detach and put these updated Gaussians back into the memory. We can obtain the current 3D occupancy prediction using a Gaussian-to-voxel splatting module whenever we need." + ], + "image_footnote": [], + "bbox": [ + 93, + 99, + 877, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "However, we humans are capable of effortlessly processing the visual information from a single view to obtain 3D perception of our surroundings. Recent methods begin to consider endowing models with the same competence, which accept a monocular RGB image as input and derive a 3D occupancy prediction within the current frustum:", + "bbox": [ + 89, + 383, + 483, + 474 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {\\text {m o n o}} = \\mathcal {F} _ {\\text {m o n o}} \\left(I _ {\\text {m o n o}}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 479, + 482, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{mono}$ is the proposed monocular prediction model, $I_{mono} \\in \\mathbb{R}^{H \\times W \\times 3}$ and $\\mathbf{Y}_{mono} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}$ refer to the monocular RGB input and the obtained 3D occupancy prediction. $X, Y, Z$ represent the dimensions of the local 3D scene and $C$ represents the total number of semantics.", + "bbox": [ + 89, + 508, + 483, + 583 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This is only the initial step towards practical scenarios. The essence of human intelligence is the capacity to analyze and respond immediately based on real-time perception of the surroundings. Correspondingly, superior embodied agents are anticipated to process egocentrically gathered real-time visual input to update the 3D occupancy prediction of the current scene. This capability facilitates the execution of downstream tasks based on real-time perception.", + "bbox": [ + 89, + 584, + 482, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivated by this, we propose an embodied 3D occupancy prediction task in this paper. Let $\\mathcal{X}_t = \\{x_1, x_2, \\dots, x_t\\}$ be an RGB sequence and the corresponding extrinsics collected by the embodied agent up to the present, where $x_t = (I_t, M_t)$ , $I_t \\in \\mathbb{R}^{H \\times W \\times 3}$ , $M_t \\in \\mathbb{R}^{3 \\times 4}$ . It is worth noting that the variation in the subscripts merely represents the change in the position and perspective of the agent when exploring the current scene continuously. Different subscripts may correspond to similar positions and perspectives, indicating that the agent has returned to a previously explored location. In embodied occupancy prediction, re-exploration of the same area should maintain global consistency and even demonstrate improved performance, akin to we hu", + "bbox": [ + 89, + 705, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mans always possessing a more comprehensive understanding of sights that have been encountered repeatedly.", + "bbox": [ + 511, + 383, + 906, + 412 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We formulate the function of an embodied occupancy prediction model as follows:", + "bbox": [ + 511, + 414, + 906, + 444 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Y} _ {t} = \\mathcal {F} _ {\\text {e m b o d i e d}} \\left(\\mathbf {Y} _ {t - 1}, x _ {t}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 449, + 903, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\text{embodied}}$ is the embodied prediction model, $\\mathbf{Y}_t \\in \\mathbb{R}^{X_{\\text{room}} \\times Y_{\\text{room}} \\times Z_{\\text{room}} \\times C}$ refers to the current occupancy prediction of the whole scene ( $\\mathbf{Y}_0$ is the initialization). $X_{\\text{room}}, Y_{\\text{room}}, Z_{\\text{room}}$ denote the scene dimensions.", + "bbox": [ + 511, + 477, + 906, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Local Refinement Module", + "text_level": 1, + "bbox": [ + 511, + 546, + 750, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Different from conventional methods that conducted feature integration in a voxelized space, we use a set of 3D semantic Gaussians to represent an indoor scene [13]. In this subsection, we will first explain our local refinement module, which extracts semantic and structural features from the monocular input and integrates them to update the Gaussian-based representation of the current frustum.", + "bbox": [ + 511, + 568, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initialization. We first initialize a set of semantic Gaussians to represent the current frustum. Each semantic Gaussian $\\mathbf{G}$ is represented by a vector comprising mean $\\mathbf{m} \\in \\mathbb{R}^3$ , scale $\\mathbf{s} \\in \\mathbb{R}^3$ , rotation quaternion $\\mathbf{r} \\in \\mathbb{R}^4$ , opacity $\\mathbf{o} \\in \\mathbb{R}$ , and semantic logits $\\mathbf{c} \\in \\mathbb{R}^C$ ( $C$ denotes the total number of semantic categories). We use an embedding layer to lift each Gaussian vector $\\mathbf{G}$ to its corresponding high-dimensional feature vectors $\\mathbf{Q}$ , and derive $\\mathcal{Q} = \\{\\mathbf{Q}_i \\in \\mathbb{R}^m, i = 1, \\dots, N\\}$ , where $m$ is the dimension of $\\mathbf{Q}_i$ and $N$ is the total number of the Gaussians.", + "bbox": [ + 511, + 674, + 908, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Depth-Aware Branch. Due to the variable scales and tight arrangements of indoor objects, depth ambiguity has always been one of the core challenges limiting the performance of indoor occupancy prediction models in monocular settings. Previous work has consistently focused on how to", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26362", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a5f0e7fd302ed683f6f3396f6d471b6603310e1ecc57aa244de0512451669db9.jpg", + "image_caption": [ + "Figure 3. Motivation of the depth-aware branch. Along a specific ray, Gaussians distributed in front of the true depth point are likely to model the empty semantic (A). Gaussians distributed behind the true depth point closely are likely to model valid semantics (B). Gaussians that are distributed behind the true depth point but are too far away require more information to guide their updates (C). During the embodied exploration, the subsequent frames can make up for this lack of information in the current frame." + ], + "image_footnote": [], + "bbox": [ + 125, + 92, + 449, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "better extract and utilize depth information from the input image. We design a depth-aware branch to provide more accurate and effective guidance for the update of 3D semantic Gaussians in our local refinement module.", + "bbox": [ + 89, + 373, + 483, + 433 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first use a depth prediction network to obtain a relatively accurate depth map $D_{\\text{metric}}$ from input $I_{mono}$ . A naive approach can explicitly utilize this depth information when initializing the Gaussians, e.g., we can randomly sample some points from the pseudo point cloud recovered from the depth map and use these coordinates to initialize the means of some Gaussians. Although providing direct hints for the means of some Gaussians, this cannot exploit the potential of the depth information. We design a simple yet effective depth-aware layer to accomplish this. We still uniformly initialize a number of Gaussians within the current frustum. For each Gaussian, we project its mean $\\mathbf{m}$ into the pixel coordinate system through the intrinsics $K_{mono} \\in \\mathbb{R}^{3 \\times 3}$ and obtain the depth value $d$ . The sampled depth value $d$ , along with the z-component $z$ of the Gaussian mean in the camera coordinate system, are fed into the depth-aware layer, which is a multi-layer perceptron (MLP) that outputs the depth-aware feature $\\mathbf{Q}_{depth}$ for this Gaussian. Then we add the depth-aware feature to the original feature vector $\\mathbf{Q}$ , injecting additional information into the subsequent feature integration. In this way, depth information not only affects the means of the Gaussians but also promotes the update of other properties:", + "bbox": [ + 91, + 434, + 483, + 781 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {\\text {d e p t h}} = \\mathcal {M} _ {\\text {d e p t h a w a r e}} \\left(\\left(D _ {\\text {m e t r i c}} (u, v), z\\right), \\right.\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 782, + 434, + 799 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {Q}} = \\left\\{\\hat {\\mathbf {Q}} _ {i}, i = 1, \\dots , N \\mid \\hat {\\mathbf {Q}} _ {i} = \\mathbf {Q} _ {i} + \\mathbf {Q} _ {i} ^ {\\text {d e p t h}} \\right\\}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 799, + 480, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{depthaware}$ is the depth-aware layer, $(u,v)$ are pixel coordinates of each Gaussian. We illustrate our depth-aware branch in Figure 3.", + "bbox": [ + 89, + 825, + 482, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Feature Integration and Gaussian Refinement. Feature integration in our local refinement module includes the", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "interactions among Gaussians as well as the interactions between image features and Gaussians. We voxelize the Gaussian centers and conduct 3D sparse convolution on the generated grid to allow interactions among Gaussian vectors $\\hat{\\mathcal{Q}}$ . We project the Gaussian centers onto the image feature map and use the deformable attention function to integrate the queried features and the Gaussian vectors $\\hat{\\mathcal{Q}}$ . After the prior feature integration, these feature vectors with aggregated information will be used to obtain the update amounts $\\Delta \\mathbf{G} = (\\Delta \\mathbf{m}, \\Delta \\mathbf{s}, \\Delta \\mathbf{r}, \\Delta \\mathbf{o}, \\Delta \\mathbf{c})$ of each Gaussian. We use the update amounts $\\Delta \\mathbf{G}$ to refine the Gaussian properties:", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\text {n e w}} = (\\Delta \\mathbf {m} + \\mathbf {m}, \\Delta \\mathbf {s} + \\mathbf {s}, \\Delta \\mathbf {r} \\otimes \\mathbf {r}, \\Delta \\mathbf {o} + \\mathbf {o}, \\Delta \\mathbf {c} + \\mathbf {c}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 517, + 261, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ refers to the special composition of quaternions.", + "bbox": [ + 511, + 281, + 890, + 295 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We conduct the feature integration and the refinement of Gaussians multiple times. After the final refinement, we use a Gaussian-to-voxel splatting module [13] to obtain the final occupancy within the frustum.", + "bbox": [ + 511, + 296, + 906, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Gaussian Memory Updated Online", + "text_level": 1, + "bbox": [ + 511, + 364, + 818, + 381 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To explore unknown scenes, we humans continuously update the objects within the scene and their relationships to gradually construct a global scene memory. When revisiting this scene for further exploration, we use the visual information to refine this memory. Inspired by this, we design an online framework (shown in Figure 2) and maintain a Gaussian memory for global understanding.", + "bbox": [ + 511, + 387, + 905, + 492 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Initialization. Our local refinement module initializes and updates Gaussians in the camera coordinate system, as the extrinsics in indoor scenarios are constantly changing, which will pose additional difficulties for our local module. But in the final embodied framework, we initialize the entire scene with uniform Gaussians in the world coordinate system. For a novel scene to be explored, we have: $\\mathcal{G}_{\\text{room}} = \\{(G_i, \\gamma_i), i = 1, \\dots, N | G_i = (\\mathbf{m}_i, \\mathbf{s}_i, \\mathbf{r}_i, \\mathbf{o}_i, \\mathbf{c}_i), \\gamma_i = 0, 1\\}$ , where $N$ refers to the number of Gaussians to initialize this scene, $\\mathbf{m}_i$ and $\\mathbf{r}_i$ are the means and rotation quaternions of these Gaussians in the world coordinate system ( $\\mathbf{s}_i$ , $\\mathbf{o}_i$ and $\\mathbf{c}_i$ maintain consistency between the world and camera coordinate systems). We introduce an additional tag $\\gamma$ for all the Gaussians in the memory. When initializing a novel scene, tags of these Gaussians are set to 0. Every time we put some updated Gaussians back into the memory, their tags are set to 1.", + "bbox": [ + 511, + 493, + 905, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Memory Update. At the current step $t$ , our embodied occupancy prediction framework receives a posed visual input $x_{t} = (I_{t}, M_{t})$ to perform the update. During the current update, we use a mask from coordinate system transformation to get all Gaussians $\\mathcal{G}_{t}$ within the current frustum from the memory. These Gaussians will interact and be refined using a tailored confidence refinement module. Then we detach these Gaussians and put them back into the memory.", + "bbox": [ + 511, + 750, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Confidence Refinement. Apart from the initial update for each scene which is akin to the local refinement, sub", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26363", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/efc113c114e99c8a76bf3a224fc9b8a473dd484b7b70563fcff1d89c657f662f.jpg", + "image_caption": [ + "Figure 4. Illustration of our Gaussian memory. During each update, the Gaussians within the current frustum are taken from the memory. Confidence values of those well-updated Gaussians are used to integrate information from both the memory and the current input. Then we put these Gaussians back into the memory." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sequent exploration involves the update of Gaussians from the Gaussian memory, among which some have been well-updated by previous frames (if we can derive an acceptable local occupancy prediction from these Gaussians, we believe that they have been \"well-updated\") and some still remain random. It is unreasonable to update these Gaussians equally. For those Gaussians deemed well-updated, we only need to refine them slightly based on the semantic and structural features extracted from the current image, which is exactly the essence of maintaining the Gaussian memory. As for those random Gaussians that have never been updated, we can directly update them with a fresh perspective.", + "bbox": [ + 88, + 343, + 482, + 523 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To elaborate, we generate a confidence value $\\theta$ for each Gaussian taken from the memory. For those having been previously updated $(\\gamma = 1)$ , we set their confidence values to a certain value between 0 and 1, which means we will integrate information from both the memory and the newly observed image to update these Gaussians. For those that have never been updated, we set their confidence values to 0 and follow the same refinement module in Sec. 3.2:", + "bbox": [ + 89, + 525, + 483, + 647 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {G} _ {\\text {o n l i n e}} = (1 - \\theta) \\Delta \\mathbf {G}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 651, + 482, + 674 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} _ {\\text {a f t e r}} = \\Delta \\mathbf {G} _ {\\text {o n l i n e}} \\oplus \\mathbf {G} _ {\\text {b e f o r e}},\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 670, + 395, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we use $\\oplus$ to represent the composition of rotation quaternions and the add operation of other parts. Figure 4 illustrates how we maintain the Gaussian memory.", + "bbox": [ + 89, + 696, + 482, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Stopping Mechanism. We propose a simple stopping mechanism to consider a room as having been effectively explored. At the step $t$ , we first calculate a confidence ratio $\\alpha$ to measure the exploration of the current room:", + "bbox": [ + 89, + 743, + 482, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\alpha = \\sum_ {i = 0} ^ {N} \\mathbb {I} _ {\\gamma_ {i} = 1} / N, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 808, + 482, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{I}_{\\gamma_i = 1}$ takes the value of 1 if $\\gamma_{i} = 1$ . If $\\alpha$ exceeds a certain threshold we set before, the model can decide to enter an adjacent room to begin a new exploration or stay here to get a better perception of the current room.", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. EmbodiedOcc: An Embodied Framework", + "text_level": 1, + "bbox": [ + 511, + 90, + 872, + 106 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We present the training framework of our EmbodiedOcc model for indoor embodied occupancy prediction. During the whole prediction process, we use the current monocular input to update our Gaussian memory in real time, which can be easily converted into 3D occupancy prediction.", + "bbox": [ + 511, + 112, + 905, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first train our local refinement module using the focal loss $L_{focal}$ , the lovasz-softmax loss $L_{lov}$ , the scene-class affinity loss $L_{scal}^{geo}$ and $L_{scal}^{sem}$ following RetinaNet [23], TPV-Former [11] and MonoScene [2]. We use monocular occupancy within the frustum $\\mathbf{Y}_{mono}^{fov}$ and the corresponding ground truth $\\mathbf{Y}_{gt}^{fov}$ to compute the loss:", + "bbox": [ + 511, + 188, + 905, + 280 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) + \\mathcal {L} _ {\\text {l o v}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) \\tag {7} \\\\ + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}) + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 282, + 903, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{1}$ is a balance factor.", + "bbox": [ + 511, + 325, + 702, + 340 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We then use the trained local module to train our EmbodiedOcc. For efficient training, we initialize the Gaussian memory before the first update and compute the current loss following the equation 7 after each update. To ensure consistency, the local occupancy ground truth used here is obtained from the occupancy of the whole scene. After a certain number of updates, we re-initialize the memory and come to the next scene. Trained with such a pipeline, our EmbodiedOcc can effectively perform the embodied occupancy prediction task while ensuring consistency within the same scene. We expect that our EmbodiedOcc can have an improving prediction with continuous exploration rather than undermining previous predictions when encountering parts that have been explored before. Therefore, we conduct some tailored tests to validate the capability of our model.", + "bbox": [ + 511, + 340, + 906, + 568 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 580, + 645, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. EmbodiedOcc-ScanNet Benchmark", + "text_level": 1, + "bbox": [ + 511, + 606, + 823, + 621 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Task Descriptions. We conducted two tasks to evaluate our EmbodiedOcc framework: local occupancy prediction and embodied occupancy prediction. Local occupancy prediction shares the same setting with previous works, which accept monocular image as input and obtain the occupancy within the current frustum. Embodied occupancy prediction accepts real-time visual inputs continuously and updates the occupancy of the current scene online. The visual input at a certain step $t$ during embodied occupancy prediction is still monocular, which is a more challenging setting compared with multi-view input or input with 3D information.", + "bbox": [ + 511, + 628, + 905, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We trained and evaluated our local refinement module on the Occ-ScanNet dataset [56], which provides frames in $60 \\times 60 \\times 36$ voxel grids (a $4.8m \\times 4.8m \\times 2.88m$ box in front of the camera). These frames are labeled with 12 semantics, including 11 for valid semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects) and 1 for empty space.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26364", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c84ce06e560a9e88e6e2520fbdc3ac7031923561645a6aed81287d7503995779.jpg", + "table_caption": [ + "Table 1. Local Prediction Performance on the Occ-ScanNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodInputIoUceilingfloorwallwindowchairbedsofatablevtsfurnitureobjectsmIoU
TPVFormer [11]xrgb33.396.9632.9714.419.1024.0141.4945.4428.6110.6635.3725.3124.94
GaussianFormer [13]xrgb40.9120.7042.0023.4017.4027.044.3044.8032.7015.3036.7025.0029.93
MonoScene [2]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.8424.62
ISO [56]xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.6128.71
Surroundocc [46]xrgb42.5218.9049.3024.8018.0026.8042.0044.1032.9018.6036.8026.9030.83
Oursxrgb53.5539.6050.4041.4031.7040.9055.0061.4044.0036.1053.9042.2045.15
", + "bbox": [ + 107, + 103, + 890, + 261 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/405379bf8e3976726467a2cc92e895c36524ab0194e293c513d1daf1dd576885.jpg", + "table_caption": [ + "Table 2. Embodied Prediction Performance on the EmbodiedOcc-ScanNet dataset." + ], + "table_footnote": [], + "table_body": "
MethodDatasetIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjectsmIoU
TPVFormer [11]EmbodiedOcc35.881.6230.5412.0313.2235.4751.3949.7925.633.6043.1516.2325.70
SurroundOcc [46]EmbodiedOcc37.0412.7031.8022.5022.0029.9044.7036.5024.6011.5034.4018.2026.27
GaussianFormer [13]EmbodiedOcc38.0217.0033.6021.5021.7029.4047.8037.1024.3015.5036.2016.8027.36
SplicingOccEmbodiedOcc49.0131.6038.8035.5036.3047.1054.5057.2034.4032.5051.2029.1040.74
EmbodiedOccEmbodiedOcc51.5222.7044.6037.4038.0050.1056.7059.7035.4038.4052.0032.9042.53
", + "bbox": [ + 93, + 276, + 906, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Based on this dataset, we reorganized an EmbodiedOccScanNet dataset to train and evaluate our EmbodiedOcc framework [35, 56]. Our EmbodiedOcc-ScanNet comprises 537/137 scenes in the train/val splits. Each scene consists of 30 posed frames with their corresponding local occupancies. The resolutions of global occupancy of each scene are calculated by $l_{x} \\times l_{y} \\times l_{z} / (0.08m)^{3}$ , where $l_{x} \\times l_{y} \\times l_{z}$ is the range of this scene in the world coordinate system. In addition, we associate grid points that can be projected onto the camera plane for each frame as the global mask of this frame. By splicing the global mask of all processed frames, we can easily obtain the occupancy ground truth of the explored part in the current scene.", + "bbox": [ + 89, + 425, + 480, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Apart from Occ-ScanNet and EmbodiedOcc-ScanNet datasets in the original scale, we sampled a small set from the EmbodiedOcc-ScanNet dataset as the EmbodiedOcc-ScanNet-mini dataset which comprises 64/16 scenes in the train/val splits. We sampled from the Occ-ScanNet dataset accordingly and obtained an Occ-ScanNet-mini2 dataset, which comprises 5504/2376 frames in the train/val splits. We conducted the local occupancy prediction task on the Occ-ScanNet and Occ-ScanNet-mini2 datasets and conducted the embodied prediction task on the EmbodiedOcc-ScanNet and EmbodiedOcc-ScanNet-mini datasets.", + "bbox": [ + 89, + 625, + 482, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use mIoU and IoU as the evaluation metrics. For local occupancy prediction, we calculate the mIoU and IoU using the occupancy within the current frustum (same with the evaluation in ISO [56]). For embodied occupancy prediction, we calculate the mIoU and IoU using the global occupancy of the current scene. It is worth mentioning that the global occupancy used here is the union", + "bbox": [ + 89, + 795, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the frustums corresponding to 30 frames of each scene, which represents the region that has been explored.", + "bbox": [ + 511, + 425, + 903, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 468, + 730, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Local Refinement Module. Following existing works [13, 56], we use a pre-trained EfficientNet-B7 [39] to initialize the image encoder in our local module. The depth prediction network used in the depth-aware branch is a fine-tuned DepthAnything-V2 model [51] that remains frozen during the training, and the depth-aware layer is a 3-layer MLP. The resolutions of the monocular input are set to $480 \\times 640$ and the number of Gaussians used to conduct the local prediction is 16200. We utilize the AdamW [26] optimizer with a weight decay of 0.01. The learning rate warms up in the first 1000 iterations to a maximum value of 2e-4 and decreases according to a cosine schedule [25]. We train our local refinement module for 10 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the Occ-ScanNet dataset and 20 epochs on the Occ-ScanNet-mini2 dataset.", + "bbox": [ + 511, + 491, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EmbodiedOcc Framework. We initialize the Gaussians with a $0.16\\mathrm{m}$ interval to represent a novel scene. For each update, the confidence value $\\theta$ of well-updated Gaussians is set to 0 in the first two refinement layers (frozen) and 0.5 in the final refinement layer. We train our EmbodiedOcc for 5 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet dataset and 20 epochs using 4 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet-mini dataset. The maximum value of the learning rate is set to 2e-4 using 8 GPUs and 1e-4 using 4 GPUs. The other settings remain the same with the training of the local refinement module.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26365", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4fa24c05c76959a20c07f5c10386478c5b78ed836050924cb55cebe8e07335db.jpg", + "table_caption": [ + "Table 3. Look-Back Prediction vs First-Time Prediction. For $\\mathrm{K} = k$ , we simply select 0, 1, ..., k - 1th frames to evaluate our EmbodiedOcc framework and the occupancy ground truth used here is the union of the $k$ frustums. K was set to 3/5/8." + ], + "table_footnote": [], + "table_body": "
ModeKFrame ListIoUmIoU
First-Time3[0,1,2]49.3939.32
Look-Back3[0,1,2,1,0]49.5240.09
First-Time5[0,...,4]50.1340.03
Look-Back5[0,...,3,4,3,...,0]50.6440.98
First-Time8[0,...,7]50.9440.86
Look-Back8[0,...,6,7,6,...,0]51.1441.17
", + "bbox": [ + 91, + 146, + 480, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3c6074de3b7fe941098c5799c9e632025327f8bb247d5930e6b0525a35464491.jpg", + "image_caption": [ + "Figure 5. Performance with different stopping ratios." + ], + "image_footnote": [], + "bbox": [ + 91, + 258, + 281, + 369 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/439b848185db8c512927307636dc02c494ba2ea477da018d637dcf35c96f68d6.jpg", + "image_caption": [ + "Figure 6. Ablation study of the confidence refinement." + ], + "image_footnote": [], + "bbox": [ + 289, + 258, + 478, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 407, + 230, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Local Occupancy Prediction. We evaluated our local refinement module on the Occ-ScanNet dataset [56]. As shown in Table 1, the results indicate that our local refinement module outperforms ISO [56]. We also implemented several state-of-the-art driving scene methods [11, 13, 46] on this benchmark and our local refinement module outperforms them by a large margin. This is because they mainly focus on the coarse layout (e.g., positions of objects) while indoor scenes require modeling of the fine-grained structure (e.g., shapes of objects).", + "bbox": [ + 89, + 429, + 483, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Embodied Occupancy Prediction. We assessed the occupancy prediction for the entire scene after processing 30 frames, and the ground truth for calculating IoU and mIoU is the union of the frustums. We spliced the local occupancy obtained from our local module to serve as the main baseline (referred to as SplicingOcc), as our local module has achieved the best local performance to date. It can be observed in Table 2 that our EmbodiedOcc exhibits superior prediction of the scene, which is achieved through the integration of different views. We also compared our EmbodiedOcc with the driving scene methods mentioned before (we obtained their embodied results by voting from different local predictions). Their poor results are due to ignoring the continuity of the observations without a global memory.", + "bbox": [ + 89, + 580, + 483, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Experimental Analysis", + "text_level": 1, + "bbox": [ + 89, + 803, + 302, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Continuous Online Updating. We expect EmbodiedOcc to have better performance when encountering parts that have been explored before, and thus, we designed a Look-Back evaluation on the EmbodiedOccScanNet dataset. Specifically, after processing K frames,", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c4743897a0d11bff6abb3f6d42c4e7eb250f8eac67cf6dbcb0a9cd9e9347c45b.jpg", + "table_caption": [ + "Table 4. Analysis of the model design." + ], + "table_footnote": [], + "table_body": "
MethodGaussianStructureMemoryLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
EmbodiedOcc-Voxel×✓✓47.5038.1237.5326.99
EmbodiedOcc w/o memory✓✓×53.5545.1549.0140.74
EmbodiedOcc✓✓✓53.5545.1551.5242.53
", + "bbox": [ + 519, + 104, + 903, + 160 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/82893a84e7069fb57bf8228feed281baab5cd032e3abf782b255defbcb2d63a1.jpg", + "table_caption": [ + "Table 5. Analysis of the depth-aware branch." + ], + "table_footnote": [], + "table_body": "
Branch TypeDepth Estimation ModuleLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
Depth-aware branchDepthAnything-V253.9346.2050.7841.45
No-depth branch/48.1540.0737.5230.73
Naive-depth branchDepthAnything-V250.3242.73//
Depth-aware branchIndoorDepth51.2443.8746.4237.78
", + "bbox": [ + 519, + 176, + 903, + 250 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7ef74f83c4c624fe26a217c0af4e36091ee23924ea4c6c55ae32f7d626f8d77.jpg", + "table_caption": [ + "Table 6. Analysis of the Gaussian parameters." + ], + "table_footnote": [], + "table_body": "
Gaussian Number (In local box)Gaussian Scale Min(m)Max(m)Gaussian Interval(m) (In global scene)Local Prediction IoUmIoUEmbodied Prediction IoUmIoU
162000.010.08(0.16, 0.16, 0.16)53.9346.2050.7841.45
81000.010.08(0.20, 0.20, 0.20)50.4742.8246.2437.99
162000.010.20(0.16, 0.16, 0.16)51.5743.7448.0938.40
", + "bbox": [ + 519, + 266, + 903, + 323 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d0318b1588d0a95109726f6e192e01cd9c13020527a331fe537aaac7c71ab4a2.jpg", + "table_caption": [ + "Table 7. Runtime decomposition." + ], + "table_footnote": [], + "table_body": "
Scene level (ms)Scene init.6.626Occ head39.635
Frame level (ms)Load memory0.973Depth aware1.816
Img backbone61.478GS Encoder14.761
Depthanything34.687Update memory0.474
", + "bbox": [ + 514, + 339, + 903, + 411 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "we direct the model to reprocess the last $\\mathrm{K} - 1$ frames. By comparing this Look-Back result with the First-Time prediction, we verified that our EmbodiedOcc has met our expectations as shown in Table 3.", + "bbox": [ + 511, + 415, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of the Stopping Mechanism. We use Figure 5 to show the effectiveness of our stopping mechanism. The ground truth used here for calculating IoU and mIoU is the union occupancy of the 30 frustums in a global scene. We observed that using a larger threshold results in more observations and better performance.", + "bbox": [ + 511, + 476, + 905, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the Confidence Refinement. During each update, local Gaussians are refined through three refinement layers. For Gaussians that have been updated before, we froze the first two layers and updated them in the last refinement layer when training our EmbodiedOcc. Figure 6 on the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets shows the impact of different confidence values (determines the coefficient of each $\\Delta G$ ). We observe that moderate updates to those previously processed Gaussians yield the best embodied occupancy prediction.", + "bbox": [ + 511, + 566, + 905, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the Model Design. The essence of our EmbodiedOcc is the explicit Gaussian memory. We adopt object-centric Gaussians instead of grid-based voxels since Gaussians are more flexible for local-global interaction. We implemented a voxel version of our EmbodiedOcc and evaluated it on our benchmark. As shown in Table 4, the satisfactory local yet poor embodied performance of EmbodiedOcc in the voxel version verified our conclusion. Results in Table 4 were evaluated on the Occ-ScanNet and EmbodiedOcc-ScanNet datasets.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of the Depth-Aware Branch. We analyze the effect of our depth-aware branch in Table 5 using the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26366", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/115be923ff84fe4ad5d2263d128bc65fde7a8a543816755979996082009d3095.jpg", + "image_caption": [ + "Figure 7. Visualization of the embodied occupancy prediction. We visualize the update of Gaussian memory and corresponding global occupancy. As the Gaussians transition from random to ordered, the occupancy of the current scene becomes more accurate and complete." + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 903, + 329 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c70f89a4b4041829c48b0571c878debc79ee313bd4540506f3d2174e31b8167c.jpg", + "image_caption": [ + "Figure 8. Visualization of local occupancy prediction." + ], + "image_footnote": [], + "bbox": [ + 101, + 362, + 472, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets. We find that depth information will significantly benefit the local and embodied occupancy prediction. As shown in the second row, without the assistance of depth information, the performance of embodied occupancy prediction drops sharply. This indicates that the update of Gaussians within the current frustum may corrupt previous predictions without the guidance of depth information. The results in the third row suggest that the depth-aware branch we employ is more reasonable compared to the naive method of directly initializing a portion of Gaussians with the pseudo point cloud recovered from the predicted depth map, the latter also poses difficulties for the initialization of global Gaussians so we do not provide the embodied results. Besides, we replaced DepthAnything-V2 with IndoorDepth [6] in the last row to prove that our depth-aware branch does not rely on a specific depth prediction network.", + "bbox": [ + 88, + 598, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the Gaussian Parameters. We analyze the effect of different Gaussian parameters in Table 6 using the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini", + "bbox": [ + 89, + 854, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets. We see that decreasing the number or increasing the scale of the Gaussians can lead to a decrease in performance during local and embodied occupancy prediction. This is closely related to the physical properties of Gaussians. Gaussians initialized too sparse may lead to holes in occupancy prediction, while Gaussians with too large scale will overlap and influence each other which is also detrimental to the correct prediction of occupancy.", + "bbox": [ + 511, + 364, + 903, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Runtime Analysis. We present in Table 7 a runtime analysis on scene 0687-00 from the EmbodiedOcc-ScanNet dataset. The runtime decomposition details show that our method is efficient while the main bottleneck is the image and depth backbones, suggesting that the overall runtime of our EmbodiedOcc can be further reduced.", + "bbox": [ + 511, + 484, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations. Figure 7 and 8 visualize the global and local predictions, respectively. Our model demonstrates reasonable local perception ability and further achieves good online prediction with the Gaussian memory. Due to space limitations, we will use a more diverse set of samples to further show the visual effect of our EmbodiedOcc in the supplementary material.", + "bbox": [ + 511, + 575, + 905, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have presented an embodied 3D occupancy prediction task and proposed a Gaussian-based EmbodiedOcc framework accordingly. Our EmbodiedOcc maintains an explicit Gaussian memory of the current scene and updates this memory during the exploration of this scene. Both quantitative and visualization results have shown that our EmbodiedOcc outperforms existing methods in terms of local occupancy prediction and accomplishes the embodied occupancy prediction task with high accuracy and strong expandability. We believe that our EmbodiedOcc paves the way for enabling active agents to conduct accurate and flexible embodied occupancy prediction.", + "bbox": [ + 509, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26367", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Tianyu Hu for her valuable assistance with the experiments. This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, and Grant 62336004, and in part by the Beijing Natural Science Foundation under Grant No. L247009.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 217, + 187, + 233 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2", + "[2] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 2, 5, 6", + "[3] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2", + "[4] Angela Dai, Daniel Ritchie, Martin Bokeloh, Scott Reed, Jürgen Sturm, and Matthias Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In CVPR, pages 4578-4587, 2018. 2", + "[5] Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2", + "[6] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 8", + "[7] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2", + "[8] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPRW, pages 0–0, 2019. 2", + "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2", + "[10] Antoine Guédon and Vincent Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. arXiv preprint arXiv:2403.14554, 2024. 2", + "[11] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2, 5, 6, 7", + "[12] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In CVPR, pages 19946-19956, 2024.", + "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In ECCV, pages 376-393, 2025. 2, 3, 4, 6, 7" + ], + "bbox": [ + 93, + 243, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, pages 4065-4071. IEEE, 2022. 1", + "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2", + "[16] Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, and Anton Konushin. Unidet3d: Multi-dataset indoor 3d object detection. In AAAI, 2025. 2", + "[17] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In CVPR, pages 16329–16339, 2024. 1", + "[18] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2", + "[19] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. RAL, 5(1): 219-226, 2019. 2", + "[20] Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Semantic gaussian splatting for neural dense slam. In ECCV, pages 163-179, 2025. 2", + "[21] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, 2023. 2", + "[22] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 2", + "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5", + "[24] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024. 2", + "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[27] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, pages 8900-8910, 2024. 2", + "[28] Ruiyuan Lyu, Tai Wang, Jingli Lin, Shuai Yang, Xiaohan Mao, Yilun Chen, Runsen Xu, Haifeng Huang, Chenming Zhu, Dahua Lin, et al. Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. arXiv preprint arXiv:2406.09401, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26368", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2", + "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1, 2", + "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2", + "[32] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In WACV, pages 5763-5773, 2024. 1", + "[33] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2", + "[34] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493, 2022. 1", + "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgb images. In ECCV, pages 746-760, 2012. 6", + "[36] Myrna C Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering: Weakly supervised 3d scene segmentation. arXiv preprint arXiv:2404.12784, 2024. 2", + "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2", + "[38] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In CVPR, pages 20675-20685, 2024. 2", + "[39] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6", + "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 2", + "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2", + "[42] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 1, 2", + "[43] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2", + "[44] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 2", + "[45] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 2", + "[46] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 2, 6, 7", + "[47] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2", + "[48] Yuting Xiao, Xuan Wang, Jiafei Li, Hongrui Cai, Yanbo Fan, Nan Xue, Minghui Yang, Yujun Shen, and Shenghua Gao. Bridging 3d gaussian and mesh for freeview video rendering. arXiv preprint arXiv:2403.11453, 2024. 2", + "[49] Xiuwei Xu, Chong Xia, Ziwei Wang, Linqing Zhao, Yueqi Duan, Jie Zhou, and Jiwen Lu. Memory-based adapters for online 3d scene perception. arXiv preprint arXiv:2403.06974, 2024. 2", + "[50] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In CVPR, pages 19595-19604, 2024. 2", + "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 6", + "[52] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. arXiv preprint arXiv:2408.14197, 2024. 2", + "[53] Zeyu Yang, Zijie Pan, Xiatian Zhu, Li Zhang, Yu-Gang Jiang, and Philip HS Torr. 4d gaussian splatting: Modeling dynamic scenes with native 4d primitives. arXiv preprint arXiv:2412.20720, 2024. 2", + "[54] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9455-9465, 2023. 2", + "[55] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2", + "[56] Hongxiao Yu, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Monocular occupancy prediction for scalable indoor scenes. arXiv preprint arXiv:2407.11730, 2024. 2, 5, 6, 7", + "[57] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2", + "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26369", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "ref_text", + "text": "[59] Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023. 2", + "bbox": [ + 91, + 90, + 482, + 147 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "26370", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_model.json b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ca2bba43cf3aeefe9d90a091e4877dfb678f76fa --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.088, + 0.764, + 0.134 + ], + "angle": 0, + "content": "EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.163, + 0.16, + 0.837, + 0.198 + ], + "angle": 0, + "content": "Yuqi Wu Wenzhao Zheng Sicheng Zuo Yuanhui Huang Jie Zhou Jiwen Lu Department of Automation, Tsinghua University, China" + }, + { + "type": "text", + "bbox": [ + 0.253, + 0.2, + 0.745, + 0.214 + ], + "angle": 0, + "content": "wuyq24@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.228, + 0.905, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.525, + 0.908, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Given streaming monocular RGB inputs, our EmbodiedOcc conducts embodied occupancy prediction in an online manner for indoor scenes. Different from existing methods which focus on offline perception from monocular images, we focus on the scene-level occupancy prediction from embodied observations. We initialize the scene to be explored with uniform 3D semantic Gaussians and progressively update them based on new observations, similar to how humans explore unknown scenes." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.594, + 0.33, + 0.61 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.486, + 0.871 + ], + "angle": 0, + "content": "3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.794 + ], + "angle": 0, + "content": "occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOccScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.8, + 0.646, + 0.816 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "With the rapid development of embodied intelligence and active agents [14, 17, 32], 3D scene perception [30, 34, 41, 42] has become a crucial task in computer vision. Intelligent agents first perceive their surrounding environments and then make decisions based on the perception results." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.887, + 0.251, + 0.901 + ], + "angle": 0, + "content": "Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26360" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "Due to the low costs of camera sensors, vision-based 3D occupancy prediction is gaining increasing popularity and produces a comprehensive understanding of both semantics and structures of the scene [2, 11, 13, 46, 56]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.363 + ], + "angle": 0, + "content": "While vision-based 3D occupancy prediction has made significant progress in outdoor driving scenes [11, 13, 22, 40, 43, 45, 46, 52, 58, 59], the application to indoor scenarios is still challenging due to the diversity and complexity of indoor scenes. Most existing methods [2, 54, 56] still focus on local 3D occupancy prediction by integrating semantic and depth information extracted from the visual inputs. However, different from outdoor scenarios, it is important to obtain a global understanding of the room for indoor scenarios, as it usually requires multiple traversals for embodied agents. Also, it is more practical to progressively explore and update the global occupancy of the 3D scene in an online manner from embodied vision-based observations with different positions and perspectives." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.365, + 0.483, + 0.817 + ], + "angle": 0, + "content": "To bridge this gap, we formulate a new embodied 3D occupancy prediction task to evaluate the ability to progressively explore an unknown scene using only visual inputs. We propose an EmbodiedOcc framework based on Gaussian memories to accomplish this task, considering the explicit and structural nature of 3D Gaussians. We initialize the global scene with uniform 3D semantic Gaussians and progressively update the Gaussians within the field of view observed by the agent. Throughout the exploration process, we maintain an explicit global memory of 3D Gaussians as the global understanding and derive the global 3D occupancy with Gaussian-to-voxel splatting [13]. Specifically, we propose a structure-aware local refinement module to update the relevant Gaussians within the current frustum. We employs a simple yet effective depth-aware branch to introduce explicit structural information for each Gaussian, ensuring the update of these Gaussians to better align with the global representation. During the continuous exploration, we read out Gaussians within the current frustum from the memory as inputs to the local module for refinement. We assign high confidence values for updated Gaussians and use them to reweight information from the memory and the current input. This ensures the consistency of the 3D representation during the fusion and update process. We reorganize an EmbodiedOcc-ScanNet benchmark for the embodied 3D occupancy prediction task based on the locally annotated Occ-ScanNet dataset [3, 47, 56]. Experiments show that our EmbodiedOcc outperforms existing methods by a large margin and accomplishes embodied occupancy prediction with high accuracy and efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.233, + 0.847 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "3D Occupancy Prediction. Benefiting from its compactness and versatility, 3D occupancy prediction based on multi-view images or additional 3D information [1, 11-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "13, 21, 37, 46] has gained great popularity over the last few years. MonoScene [2] was the first to derive 3D occupancy prediction from a single image, propelling the original 3D Semantic Scene Completion (SSC) [4, 8, 18, 19, 33, 37] into a more challenging stage with vision-only inputs and more universal scenarios (both indoor and outdoor scenes). Subsequent works [54, 56] further focused on addressing the depth ambiguity in this monocular setting. However, most of these efforts were confined to local and offline prediction. SCFusion [47] proposed an incremental framework based on RGB-D inputs. EmbodiedScan [28, 44] introduced an offline global prediction framework using multimodal sequential inputs. Differently, the proposed embodied 3D occupancy prediction aims at online prediction from RGB-only inputs, which is more challenging and practical." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.905, + 0.546 + ], + "angle": 0, + "content": "Online 3D Scene Perception. Accurate comprehension of 3D scenes is an indispensable capability for embodied agents, such as 3D occupancy prediction [2, 56] and object detection [16, 31, 42]. Most existing works on indoor 3D scene perception [9, 30, 41, 55] take pre-acquired and reconstructed 3D data as inputs and perceive the scene in an offline manner. To achieve online perception, Online3D [49] introduced an adapter-based model that equips mainstream offline frameworks with the competence to perform online scene perception, enabling the process of real-time RGB-D sequences. However, this framework still requires depth information as inputs and mainly targets point segmentation and 3D detection. Differently, we target online vision-based 3D occupancy prediction which can provide a more comprehensive understanding of the scene." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.548, + 0.905, + 0.789 + ], + "angle": 0, + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [15] uses 3D Gaussians to model a 3D scene and benefits from fast speed and high quality in the field of neural rendering. The physical characteristics of 3D Gaussians and the splat-based rasterization also motivated rapid advancements in research fields such as scene editing [10, 24, 29, 36], dynamic scenarios [7, 27, 38, 48, 53], and SLAM [5, 20, 50, 57]. GaussianFormer [13] pioneers the application of 3D Gaussians in outdoor 3D occupancy prediction and uses features from multi-view images to update 3D Gaussians, which can be converted into 3D occupancy prediction through a Gaussian-to-voxel splatting module. However, it is still unclear how to employ 3D Gaussians for online global indoor scene understanding from local observations. We achieve this by designing a Gaussian memory mechanism and progressively updating it with structure-aware interaction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.707, + 0.824 + ], + "angle": 0, + "content": "3. Proposed Approach" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.83, + 0.849 + ], + "angle": 0, + "content": "3.1. Embodied 3D Occupancy Prediction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Conventional methods in indoor scenarios for occupancy prediction accepted RGB-D as inputs to predict the semantic occupancy of a 3D scene which requires depth sensors." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26361" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.101, + 0.878, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Figure 2. Framework of our EmbodiedOcc for embodied 3D occupancy prediction. We maintain an explicit global memory of 3D Gaussians during the exploration of the current scene. For each update, the Gaussians within the current frustum are taken from the memory and updated using semantic and structural features extracted from the monocular RGB input. Each Gaussian has a confidence value to integrate information from both the memory and the current input. Then we detach and put these updated Gaussians back into the memory. We can obtain the current 3D occupancy prediction using a Gaussian-to-voxel splatting module whenever we need." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.384, + 0.484, + 0.476 + ], + "angle": 0, + "content": "However, we humans are capable of effortlessly processing the visual information from a single view to obtain 3D perception of our surroundings. Recent methods begin to consider endowing models with the same competence, which accept a monocular RGB image as input and derive a 3D occupancy prediction within the current frustum:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.48, + 0.483, + 0.497 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {\\text {m o n o}} = \\mathcal {F} _ {\\text {m o n o}} \\left(I _ {\\text {m o n o}}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.584 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{mono}\\) is the proposed monocular prediction model, \\(I_{mono} \\in \\mathbb{R}^{H \\times W \\times 3}\\) and \\(\\mathbf{Y}_{mono} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}\\) refer to the monocular RGB input and the obtained 3D occupancy prediction. \\(X, Y, Z\\) represent the dimensions of the local 3D scene and \\(C\\) represents the total number of semantics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.705 + ], + "angle": 0, + "content": "This is only the initial step towards practical scenarios. The essence of human intelligence is the capacity to analyze and respond immediately based on real-time perception of the surroundings. Correspondingly, superior embodied agents are anticipated to process egocentrically gathered real-time visual input to update the 3D occupancy prediction of the current scene. This capability facilitates the execution of downstream tasks based on real-time perception." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.706, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Motivated by this, we propose an embodied 3D occupancy prediction task in this paper. Let \\(\\mathcal{X}_t = \\{x_1, x_2, \\dots, x_t\\}\\) be an RGB sequence and the corresponding extrinsics collected by the embodied agent up to the present, where \\(x_t = (I_t, M_t)\\), \\(I_t \\in \\mathbb{R}^{H \\times W \\times 3}\\), \\(M_t \\in \\mathbb{R}^{3 \\times 4}\\). It is worth noting that the variation in the subscripts merely represents the change in the position and perspective of the agent when exploring the current scene continuously. Different subscripts may correspond to similar positions and perspectives, indicating that the agent has returned to a previously explored location. In embodied occupancy prediction, re-exploration of the same area should maintain global consistency and even demonstrate improved performance, akin to we hu" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.907, + 0.414 + ], + "angle": 0, + "content": "mans always possessing a more comprehensive understanding of sights that have been encountered repeatedly." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.907, + 0.445 + ], + "angle": 0, + "content": "We formulate the function of an embodied occupancy prediction model as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.45, + 0.905, + 0.466 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Y} _ {t} = \\mathcal {F} _ {\\text {e m b o d i e d}} \\left(\\mathbf {Y} _ {t - 1}, x _ {t}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.478, + 0.907, + 0.539 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\text{embodied}}\\) is the embodied prediction model, \\(\\mathbf{Y}_t \\in \\mathbb{R}^{X_{\\text{room}} \\times Y_{\\text{room}} \\times Z_{\\text{room}} \\times C}\\) refers to the current occupancy prediction of the whole scene (\\(\\mathbf{Y}_0\\) is the initialization). \\(X_{\\text{room}}, Y_{\\text{room}}, Z_{\\text{room}}\\) denote the scene dimensions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.75, + 0.563 + ], + "angle": 0, + "content": "3.2. Local Refinement Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Different from conventional methods that conducted feature integration in a voxelized space, we use a set of 3D semantic Gaussians to represent an indoor scene [13]. In this subsection, we will first explain our local refinement module, which extracts semantic and structural features from the monocular input and integrates them to update the Gaussian-based representation of the current frustum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.825 + ], + "angle": 0, + "content": "Initialization. We first initialize a set of semantic Gaussians to represent the current frustum. Each semantic Gaussian \\(\\mathbf{G}\\) is represented by a vector comprising mean \\(\\mathbf{m} \\in \\mathbb{R}^3\\), scale \\(\\mathbf{s} \\in \\mathbb{R}^3\\), rotation quaternion \\(\\mathbf{r} \\in \\mathbb{R}^4\\), opacity \\(\\mathbf{o} \\in \\mathbb{R}\\), and semantic logits \\(\\mathbf{c} \\in \\mathbb{R}^C\\) (\\(C\\) denotes the total number of semantic categories). We use an embedding layer to lift each Gaussian vector \\(\\mathbf{G}\\) to its corresponding high-dimensional feature vectors \\(\\mathbf{Q}\\), and derive \\(\\mathcal{Q} = \\{\\mathbf{Q}_i \\in \\mathbb{R}^m, i = 1, \\dots, N\\}\\), where \\(m\\) is the dimension of \\(\\mathbf{Q}_i\\) and \\(N\\) is the total number of the Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Depth-Aware Branch. Due to the variable scales and tight arrangements of indoor objects, depth ambiguity has always been one of the core challenges limiting the performance of indoor occupancy prediction models in monocular settings. Previous work has consistently focused on how to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26362" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.093, + 0.45, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.257, + 0.483, + 0.368 + ], + "angle": 0, + "content": "Figure 3. Motivation of the depth-aware branch. Along a specific ray, Gaussians distributed in front of the true depth point are likely to model the empty semantic (A). Gaussians distributed behind the true depth point closely are likely to model valid semantics (B). Gaussians that are distributed behind the true depth point but are too far away require more information to guide their updates (C). During the embodied exploration, the subsequent frames can make up for this lack of information in the current frame." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.374, + 0.484, + 0.434 + ], + "angle": 0, + "content": "better extract and utilize depth information from the input image. We design a depth-aware branch to provide more accurate and effective guidance for the update of 3D semantic Gaussians in our local refinement module." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.782 + ], + "angle": 0, + "content": "We first use a depth prediction network to obtain a relatively accurate depth map \\( D_{\\text{metric}} \\) from input \\( I_{mono} \\). A naive approach can explicitly utilize this depth information when initializing the Gaussians, e.g., we can randomly sample some points from the pseudo point cloud recovered from the depth map and use these coordinates to initialize the means of some Gaussians. Although providing direct hints for the means of some Gaussians, this cannot exploit the potential of the depth information. We design a simple yet effective depth-aware layer to accomplish this. We still uniformly initialize a number of Gaussians within the current frustum. For each Gaussian, we project its mean \\( \\mathbf{m} \\) into the pixel coordinate system through the intrinsics \\( K_{mono} \\in \\mathbb{R}^{3 \\times 3} \\) and obtain the depth value \\( d \\). The sampled depth value \\( d \\), along with the z-component \\( z \\) of the Gaussian mean in the camera coordinate system, are fed into the depth-aware layer, which is a multi-layer perceptron (MLP) that outputs the depth-aware feature \\( \\mathbf{Q}_{depth} \\) for this Gaussian. Then we add the depth-aware feature to the original feature vector \\( \\mathbf{Q} \\), injecting additional information into the subsequent feature integration. In this way, depth information not only affects the means of the Gaussians but also promotes the update of other properties:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.784, + 0.435, + 0.8 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {\\text {d e p t h}} = \\mathcal {M} _ {\\text {d e p t h a w a r e}} \\left(\\left(D _ {\\text {m e t r i c}} (u, v), z\\right), \\right.\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.8, + 0.482, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {Q}} = \\left\\{\\hat {\\mathbf {Q}} _ {i}, i = 1, \\dots , N \\mid \\hat {\\mathbf {Q}} _ {i} = \\mathbf {Q} _ {i} + \\mathbf {Q} _ {i} ^ {\\text {d e p t h}} \\right\\}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{depthaware}\\) is the depth-aware layer, \\((u,v)\\) are pixel coordinates of each Gaussian. We illustrate our depth-aware branch in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Feature Integration and Gaussian Refinement. Feature integration in our local refinement module includes the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "interactions among Gaussians as well as the interactions between image features and Gaussians. We voxelize the Gaussian centers and conduct 3D sparse convolution on the generated grid to allow interactions among Gaussian vectors \\(\\hat{\\mathcal{Q}}\\). We project the Gaussian centers onto the image feature map and use the deformable attention function to integrate the queried features and the Gaussian vectors \\(\\hat{\\mathcal{Q}}\\). After the prior feature integration, these feature vectors with aggregated information will be used to obtain the update amounts \\(\\Delta \\mathbf{G} = (\\Delta \\mathbf{m}, \\Delta \\mathbf{s}, \\Delta \\mathbf{r}, \\Delta \\mathbf{o}, \\Delta \\mathbf{c})\\) of each Gaussian. We use the update amounts \\(\\Delta \\mathbf{G}\\) to refine the Gaussian properties:" + }, + { + "type": "equation", + "bbox": [ + 0.518, + 0.262, + 0.905, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\text {n e w}} = (\\Delta \\mathbf {m} + \\mathbf {m}, \\Delta \\mathbf {s} + \\mathbf {s}, \\Delta \\mathbf {r} \\otimes \\mathbf {r}, \\Delta \\mathbf {o} + \\mathbf {o}, \\Delta \\mathbf {c} + \\mathbf {c}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.282, + 0.892, + 0.296 + ], + "angle": 0, + "content": "where \\(\\otimes\\) refers to the special composition of quaternions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.907, + 0.357 + ], + "angle": 0, + "content": "We conduct the feature integration and the refinement of Gaussians multiple times. After the final refinement, we use a Gaussian-to-voxel splatting module [13] to obtain the final occupancy within the frustum." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.819, + 0.382 + ], + "angle": 0, + "content": "3.3. Gaussian Memory Updated Online" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.906, + 0.493 + ], + "angle": 0, + "content": "To explore unknown scenes, we humans continuously update the objects within the scene and their relationships to gradually construct a global scene memory. When revisiting this scene for further exploration, we use the visual information to refine this memory. Inspired by this, we design an online framework (shown in Figure 2) and maintain a Gaussian memory for global understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.906, + 0.75 + ], + "angle": 0, + "content": "Memory Initialization. Our local refinement module initializes and updates Gaussians in the camera coordinate system, as the extrinsics in indoor scenarios are constantly changing, which will pose additional difficulties for our local module. But in the final embodied framework, we initialize the entire scene with uniform Gaussians in the world coordinate system. For a novel scene to be explored, we have: \\(\\mathcal{G}_{\\text{room}} = \\{(G_i, \\gamma_i), i = 1, \\dots, N | G_i = (\\mathbf{m}_i, \\mathbf{s}_i, \\mathbf{r}_i, \\mathbf{o}_i, \\mathbf{c}_i), \\gamma_i = 0, 1\\}\\), where \\(N\\) refers to the number of Gaussians to initialize this scene, \\(\\mathbf{m}_i\\) and \\(\\mathbf{r}_i\\) are the means and rotation quaternions of these Gaussians in the world coordinate system (\\(\\mathbf{s}_i\\), \\(\\mathbf{o}_i\\) and \\(\\mathbf{c}_i\\) maintain consistency between the world and camera coordinate systems). We introduce an additional tag \\(\\gamma\\) for all the Gaussians in the memory. When initializing a novel scene, tags of these Gaussians are set to 0. Every time we put some updated Gaussians back into the memory, their tags are set to 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Memory Update. At the current step \\( t \\), our embodied occupancy prediction framework receives a posed visual input \\( x_{t} = (I_{t}, M_{t}) \\) to perform the update. During the current update, we use a mask from coordinate system transformation to get all Gaussians \\( \\mathcal{G}_{t} \\) within the current frustum from the memory. These Gaussians will interact and be refined using a tailored confidence refinement module. Then we detach these Gaussians and put them back into the memory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Confidence Refinement. Apart from the initial update for each scene which is akin to the local refinement, sub" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26363" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.482, + 0.26 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.263, + 0.483, + 0.334 + ], + "angle": 0, + "content": "Figure 4. Illustration of our Gaussian memory. During each update, the Gaussians within the current frustum are taken from the memory. Confidence values of those well-updated Gaussians are used to integrate information from both the memory and the current input. Then we put these Gaussians back into the memory." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.344, + 0.483, + 0.525 + ], + "angle": 0, + "content": "sequent exploration involves the update of Gaussians from the Gaussian memory, among which some have been well-updated by previous frames (if we can derive an acceptable local occupancy prediction from these Gaussians, we believe that they have been \"well-updated\") and some still remain random. It is unreasonable to update these Gaussians equally. For those Gaussians deemed well-updated, we only need to refine them slightly based on the semantic and structural features extracted from the current image, which is exactly the essence of maintaining the Gaussian memory. As for those random Gaussians that have never been updated, we can directly update them with a fresh perspective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.484, + 0.648 + ], + "angle": 0, + "content": "To elaborate, we generate a confidence value \\(\\theta\\) for each Gaussian taken from the memory. For those having been previously updated \\((\\gamma = 1)\\), we set their confidence values to a certain value between 0 and 1, which means we will integrate information from both the memory and the newly observed image to update these Gaussians. For those that have never been updated, we set their confidence values to 0 and follow the same refinement module in Sec. 3.2:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.652, + 0.483, + 0.675 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {G} _ {\\text {o n l i n e}} = (1 - \\theta) \\Delta \\mathbf {G}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.671, + 0.396, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} _ {\\text {a f t e r}} = \\Delta \\mathbf {G} _ {\\text {o n l i n e}} \\oplus \\mathbf {G} _ {\\text {b e f o r e}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.483, + 0.743 + ], + "angle": 0, + "content": "where we use \\(\\oplus\\) to represent the composition of rotation quaternions and the add operation of other parts. Figure 4 illustrates how we maintain the Gaussian memory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.483, + 0.805 + ], + "angle": 0, + "content": "Stopping Mechanism. We propose a simple stopping mechanism to consider a room as having been effectively explored. At the step \\( t \\), we first calculate a confidence ratio \\( \\alpha \\) to measure the exploration of the current room:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.809, + 0.483, + 0.83 + ], + "angle": 0, + "content": "\\[\n\\alpha = \\sum_ {i = 0} ^ {N} \\mathbb {I} _ {\\gamma_ {i} = 1} / N, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbb{I}_{\\gamma_i = 1}\\) takes the value of 1 if \\(\\gamma_{i} = 1\\). If \\(\\alpha\\) exceeds a certain threshold we set before, the model can decide to enter an adjacent room to begin a new exploration or stay here to get a better perception of the current room." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.874, + 0.107 + ], + "angle": 0, + "content": "3.4. EmbodiedOcc: An Embodied Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.906, + 0.188 + ], + "angle": 0, + "content": "We present the training framework of our EmbodiedOcc model for indoor embodied occupancy prediction. During the whole prediction process, we use the current monocular input to update our Gaussian memory in real time, which can be easily converted into 3D occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.906, + 0.281 + ], + "angle": 0, + "content": "We first train our local refinement module using the focal loss \\(L_{focal}\\), the lovasz-softmax loss \\(L_{lov}\\), the scene-class affinity loss \\(L_{scal}^{geo}\\) and \\(L_{scal}^{sem}\\) following RetinaNet [23], TPV-Former [11] and MonoScene [2]. We use monocular occupancy within the frustum \\(\\mathbf{Y}_{mono}^{fov}\\) and the corresponding ground truth \\(\\mathbf{Y}_{gt}^{fov}\\) to compute the loss:" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.283, + 0.905, + 0.325 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) + \\mathcal {L} _ {\\text {l o v}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) \\tag {7} \\\\ + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}) + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.327, + 0.704, + 0.341 + ], + "angle": 0, + "content": "where \\(\\lambda_{1}\\) is a balance factor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.569 + ], + "angle": 0, + "content": "We then use the trained local module to train our EmbodiedOcc. For efficient training, we initialize the Gaussian memory before the first update and compute the current loss following the equation 7 after each update. To ensure consistency, the local occupancy ground truth used here is obtained from the occupancy of the whole scene. After a certain number of updates, we re-initialize the memory and come to the next scene. Trained with such a pipeline, our EmbodiedOcc can effectively perform the embodied occupancy prediction task while ensuring consistency within the same scene. We expect that our EmbodiedOcc can have an improving prediction with continuous exploration rather than undermining previous predictions when encountering parts that have been explored before. Therefore, we conduct some tailored tests to validate the capability of our model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.582, + 0.646, + 0.599 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.824, + 0.622 + ], + "angle": 0, + "content": "4.1. EmbodiedOcc-ScanNet Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.795 + ], + "angle": 0, + "content": "Task Descriptions. We conducted two tasks to evaluate our EmbodiedOcc framework: local occupancy prediction and embodied occupancy prediction. Local occupancy prediction shares the same setting with previous works, which accept monocular image as input and obtain the occupancy within the current frustum. Embodied occupancy prediction accepts real-time visual inputs continuously and updates the occupancy of the current scene online. The visual input at a certain step \\( t \\) during embodied occupancy prediction is still monocular, which is a more challenging setting compared with multi-view input or input with 3D information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Datasets. We trained and evaluated our local refinement module on the Occ-ScanNet dataset [56], which provides frames in \\(60 \\times 60 \\times 36\\) voxel grids (a \\(4.8m \\times 4.8m \\times 2.88m\\) box in front of the camera). These frames are labeled with 12 semantics, including 11 for valid semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects) and 1 for empty space." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26364" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.288, + 0.09, + 0.71, + 0.102 + ], + "angle": 0, + "content": "Table 1. Local Prediction Performance on the Occ-ScanNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.104, + 0.892, + 0.262 + ], + "angle": 0, + "content": "
MethodInputIoUceilingfloorwallwindowchairbedsofatablevtsfurnitureobjectsmIoU
TPVFormer [11]xrgb33.396.9632.9714.419.1024.0141.4945.4428.6110.6635.3725.3124.94
GaussianFormer [13]xrgb40.9120.7042.0023.4017.4027.044.3044.8032.7015.3036.7025.0029.93
MonoScene [2]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.8424.62
ISO [56]xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.6128.71
Surroundocc [46]xrgb42.5218.9049.3024.8018.0026.8042.0044.1032.9018.6036.8026.9030.83
Oursxrgb53.5539.6050.4041.4031.7040.9055.0061.4044.0036.1053.9042.2045.15
" + }, + { + "type": "table_caption", + "bbox": [ + 0.241, + 0.264, + 0.757, + 0.275 + ], + "angle": 0, + "content": "Table 2. Embodied Prediction Performance on the EmbodiedOcc-ScanNet dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.277, + 0.907, + 0.42 + ], + "angle": 0, + "content": "
MethodDatasetIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjectsmIoU
TPVFormer [11]EmbodiedOcc35.881.6230.5412.0313.2235.4751.3949.7925.633.6043.1516.2325.70
SurroundOcc [46]EmbodiedOcc37.0412.7031.8022.5022.0029.9044.7036.5024.6011.5034.4018.2026.27
GaussianFormer [13]EmbodiedOcc38.0217.0033.6021.5021.7029.4047.8037.1024.3015.5036.2016.8027.36
SplicingOccEmbodiedOcc49.0131.6038.8035.5036.3047.1054.5057.2034.4032.5051.2029.1040.74
EmbodiedOccEmbodiedOcc51.5222.7044.6037.4038.0050.1056.7059.7035.4038.4052.0032.9042.53
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.482, + 0.621 + ], + "angle": 0, + "content": "Based on this dataset, we reorganized an EmbodiedOccScanNet dataset to train and evaluate our EmbodiedOcc framework [35, 56]. Our EmbodiedOcc-ScanNet comprises 537/137 scenes in the train/val splits. Each scene consists of 30 posed frames with their corresponding local occupancies. The resolutions of global occupancy of each scene are calculated by \\( l_{x} \\times l_{y} \\times l_{z} / (0.08m)^{3} \\), where \\( l_{x} \\times l_{y} \\times l_{z} \\) is the range of this scene in the world coordinate system. In addition, we associate grid points that can be projected onto the camera plane for each frame as the global mask of this frame. By splicing the global mask of all processed frames, we can easily obtain the occupancy ground truth of the explored part in the current scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.79 + ], + "angle": 0, + "content": "Apart from Occ-ScanNet and EmbodiedOcc-ScanNet datasets in the original scale, we sampled a small set from the EmbodiedOcc-ScanNet dataset as the EmbodiedOcc-ScanNet-mini dataset which comprises 64/16 scenes in the train/val splits. We sampled from the Occ-ScanNet dataset accordingly and obtained an Occ-ScanNet-mini2 dataset, which comprises 5504/2376 frames in the train/val splits. We conducted the local occupancy prediction task on the Occ-ScanNet and Occ-ScanNet-mini2 datasets and conducted the embodied prediction task on the EmbodiedOcc-ScanNet and EmbodiedOcc-ScanNet-mini datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics. We use mIoU and IoU as the evaluation metrics. For local occupancy prediction, we calculate the mIoU and IoU using the occupancy within the current frustum (same with the evaluation in ISO [56]). For embodied occupancy prediction, we calculate the mIoU and IoU using the global occupancy of the current scene. It is worth mentioning that the global occupancy used here is the union" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.426, + 0.905, + 0.456 + ], + "angle": 0, + "content": "of the frustums corresponding to 30 frames of each scene, which represents the region that has been explored." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.469, + 0.731, + 0.485 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Local Refinement Module. Following existing works [13, 56], we use a pre-trained EfficientNet-B7 [39] to initialize the image encoder in our local module. The depth prediction network used in the depth-aware branch is a fine-tuned DepthAnything-V2 model [51] that remains frozen during the training, and the depth-aware layer is a 3-layer MLP. The resolutions of the monocular input are set to \\(480 \\times 640\\) and the number of Gaussians used to conduct the local prediction is 16200. We utilize the AdamW [26] optimizer with a weight decay of 0.01. The learning rate warms up in the first 1000 iterations to a maximum value of 2e-4 and decreases according to a cosine schedule [25]. We train our local refinement module for 10 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the Occ-ScanNet dataset and 20 epochs on the Occ-ScanNet-mini2 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "EmbodiedOcc Framework. We initialize the Gaussians with a \\(0.16\\mathrm{m}\\) interval to represent a novel scene. For each update, the confidence value \\(\\theta\\) of well-updated Gaussians is set to 0 in the first two refinement layers (frozen) and 0.5 in the final refinement layer. We train our EmbodiedOcc for 5 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet dataset and 20 epochs using 4 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet-mini dataset. The maximum value of the learning rate is set to 2e-4 using 8 GPUs and 1e-4 using 4 GPUs. The other settings remain the same with the training of the local refinement module." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26365" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.145 + ], + "angle": 0, + "content": "Table 3. Look-Back Prediction vs First-Time Prediction. For \\( \\mathrm{K} = k \\), we simply select 0, 1, ..., k - 1th frames to evaluate our EmbodiedOcc framework and the occupancy ground truth used here is the union of the \\( k \\) frustums. K was set to 3/5/8." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.147, + 0.482, + 0.258 + ], + "angle": 0, + "content": "
ModeKFrame ListIoUmIoU
First-Time3[0,1,2]49.3939.32
Look-Back3[0,1,2,1,0]49.5240.09
First-Time5[0,...,4]50.1340.03
Look-Back5[0,...,3,4,3,...,0]50.6440.98
First-Time8[0,...,7]50.9440.86
Look-Back8[0,...,6,7,6,...,0]51.1441.17
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.26, + 0.282, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.374, + 0.282, + 0.403 + ], + "angle": 0, + "content": "Figure 5. Performance with different stopping ratios." + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.26, + 0.48, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.288, + 0.374, + 0.48, + 0.403 + ], + "angle": 0, + "content": "Figure 6. Ablation study of the confidence refinement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.408, + 0.232, + 0.423 + ], + "angle": 0, + "content": "4.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Local Occupancy Prediction. We evaluated our local refinement module on the Occ-ScanNet dataset [56]. As shown in Table 1, the results indicate that our local refinement module outperforms ISO [56]. We also implemented several state-of-the-art driving scene methods [11, 13, 46] on this benchmark and our local refinement module outperforms them by a large margin. This is because they mainly focus on the coarse layout (e.g., positions of objects) while indoor scenes require modeling of the fine-grained structure (e.g., shapes of objects)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.793 + ], + "angle": 0, + "content": "Embodied Occupancy Prediction. We assessed the occupancy prediction for the entire scene after processing 30 frames, and the ground truth for calculating IoU and mIoU is the union of the frustums. We spliced the local occupancy obtained from our local module to serve as the main baseline (referred to as SplicingOcc), as our local module has achieved the best local performance to date. It can be observed in Table 2 that our EmbodiedOcc exhibits superior prediction of the scene, which is achieved through the integration of different views. We also compared our EmbodiedOcc with the driving scene methods mentioned before (we obtained their embodied results by voting from different local predictions). Their poor results are due to ignoring the continuity of the observations without a global memory." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.303, + 0.82 + ], + "angle": 0, + "content": "4.4. Experimental Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Effect of Continuous Online Updating. We expect EmbodiedOcc to have better performance when encountering parts that have been explored before, and thus, we designed a Look-Back evaluation on the EmbodiedOccScanNet dataset. Specifically, after processing K frames," + }, + { + "type": "table_caption", + "bbox": [ + 0.591, + 0.09, + 0.829, + 0.104 + ], + "angle": 0, + "content": "Table 4. Analysis of the model design." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.105, + 0.905, + 0.161 + ], + "angle": 0, + "content": "
MethodGaussianStructureMemoryLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
EmbodiedOcc-Voxel×✓✓47.5038.1237.5326.99
EmbodiedOcc w/o memory✓✓×53.5545.1549.0140.74
EmbodiedOcc✓✓✓53.5545.1551.5242.53
" + }, + { + "type": "table_caption", + "bbox": [ + 0.569, + 0.163, + 0.851, + 0.175 + ], + "angle": 0, + "content": "Table 5. Analysis of the depth-aware branch." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.178, + 0.905, + 0.25 + ], + "angle": 0, + "content": "
Branch TypeDepth Estimation ModuleLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
Depth-aware branchDepthAnything-V253.9346.2050.7841.45
No-depth branch/48.1540.0737.5230.73
Naive-depth branchDepthAnything-V250.3242.73//
Depth-aware branchIndoorDepth51.2443.8746.4237.78
" + }, + { + "type": "table_caption", + "bbox": [ + 0.566, + 0.252, + 0.854, + 0.266 + ], + "angle": 0, + "content": "Table 6. Analysis of the Gaussian parameters." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.267, + 0.905, + 0.324 + ], + "angle": 0, + "content": "
Gaussian Number (In local box)Gaussian Scale Min(m)Max(m)Gaussian Interval(m) (In global scene)Local Prediction IoUmIoUEmbodied Prediction IoUmIoU
162000.010.08(0.16, 0.16, 0.16)53.9346.2050.7841.45
81000.010.08(0.20, 0.20, 0.20)50.4742.8246.2437.99
162000.010.20(0.16, 0.16, 0.16)51.5743.7448.0938.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.607, + 0.325, + 0.814, + 0.338 + ], + "angle": 0, + "content": "Table 7. Runtime decomposition." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.34, + 0.904, + 0.412 + ], + "angle": 0, + "content": "
Scene level (ms)Scene init.6.626Occ head39.635
Frame level (ms)Load memory0.973Depth aware1.816
Img backbone61.478GS Encoder14.761
Depthanything34.687Update memory0.474
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.476 + ], + "angle": 0, + "content": "we direct the model to reprocess the last \\( \\mathrm{K} - 1 \\) frames. By comparing this Look-Back result with the First-Time prediction, we verified that our EmbodiedOcc has met our expectations as shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.568 + ], + "angle": 0, + "content": "Effect of the Stopping Mechanism. We use Figure 5 to show the effectiveness of our stopping mechanism. The ground truth used here for calculating IoU and mIoU is the union occupancy of the 30 frustums in a global scene. We observed that using a larger threshold results in more observations and better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Analysis of the Confidence Refinement. During each update, local Gaussians are refined through three refinement layers. For Gaussians that have been updated before, we froze the first two layers and updated them in the last refinement layer when training our EmbodiedOcc. Figure 6 on the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets shows the impact of different confidence values (determines the coefficient of each \\(\\Delta G\\)). We observe that moderate updates to those previously processed Gaussians yield the best embodied occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Analysis of the Model Design. The essence of our EmbodiedOcc is the explicit Gaussian memory. We adopt object-centric Gaussians instead of grid-based voxels since Gaussians are more flexible for local-global interaction. We implemented a voxel version of our EmbodiedOcc and evaluated it on our benchmark. As shown in Table 4, the satisfactory local yet poor embodied performance of EmbodiedOcc in the voxel version verified our conclusion. Results in Table 4 were evaluated on the Occ-ScanNet and EmbodiedOcc-ScanNet datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Analysis of the Depth-Aware Branch. We analyze the effect of our depth-aware branch in Table 5 using the" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26366" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.09, + 0.905, + 0.33 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.332, + 0.907, + 0.361 + ], + "angle": 0, + "content": "Figure 7. Visualization of the embodied occupancy prediction. We visualize the update of Gaussian memory and corresponding global occupancy. As the Gaussians transition from random to ordered, the occupancy of the current scene becomes more accurate and complete." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.363, + 0.473, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.586, + 0.454, + 0.599 + ], + "angle": 0, + "content": "Figure 8. Visualization of local occupancy prediction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.856 + ], + "angle": 0, + "content": "Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets. We find that depth information will significantly benefit the local and embodied occupancy prediction. As shown in the second row, without the assistance of depth information, the performance of embodied occupancy prediction drops sharply. This indicates that the update of Gaussians within the current frustum may corrupt previous predictions without the guidance of depth information. The results in the third row suggest that the depth-aware branch we employ is more reasonable compared to the naive method of directly initializing a portion of Gaussians with the pseudo point cloud recovered from the predicted depth map, the latter also poses difficulties for the initialization of global Gaussians so we do not provide the embodied results. Besides, we replaced DepthAnything-V2 with IndoorDepth [6] in the last row to prove that our depth-aware branch does not rely on a specific depth prediction network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Analysis of the Gaussian Parameters. We analyze the effect of different Gaussian parameters in Table 6 using the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.905, + 0.486 + ], + "angle": 0, + "content": "datasets. We see that decreasing the number or increasing the scale of the Gaussians can lead to a decrease in performance during local and embodied occupancy prediction. This is closely related to the physical properties of Gaussians. Gaussians initialized too sparse may lead to holes in occupancy prediction, while Gaussians with too large scale will overlap and influence each other which is also detrimental to the correct prediction of occupancy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.576 + ], + "angle": 0, + "content": "Runtime Analysis. We present in Table 7 a runtime analysis on scene 0687-00 from the EmbodiedOcc-ScanNet dataset. The runtime decomposition details show that our method is efficient while the main bottleneck is the image and depth backbones, suggesting that the overall runtime of our EmbodiedOcc can be further reduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.906, + 0.682 + ], + "angle": 0, + "content": "Visualizations. Figure 7 and 8 visualize the global and local predictions, respectively. Our model demonstrates reasonable local perception ability and further achieves good online prediction with the Gaussian memory. Due to space limitations, we will use a more diverse set of samples to further show the visual effect of our EmbodiedOcc in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we have presented an embodied 3D occupancy prediction task and proposed a Gaussian-based EmbodiedOcc framework accordingly. Our EmbodiedOcc maintains an explicit Gaussian memory of the current scene and updates this memory during the exploration of this scene. Both quantitative and visualization results have shown that our EmbodiedOcc outperforms existing methods in terms of local occupancy prediction and accomplishes the embodied occupancy prediction task with high accuracy and strong expandability. We believe that our EmbodiedOcc paves the way for enabling active agents to conduct accurate and flexible embodied occupancy prediction." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26367" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.206 + ], + "angle": 0, + "content": "We would like to thank Tianyu Hu for her valuable assistance with the experiments. This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, and Grant 62336004, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.218, + 0.188, + 0.234 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.244, + 0.484, + 0.299 + ], + "angle": 0, + "content": "[1] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.3, + 0.484, + 0.34 + ], + "angle": 0, + "content": "[2] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.342, + 0.483, + 0.397 + ], + "angle": 0, + "content": "[3] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[4] Angela Dai, Daniel Ritchie, Martin Bokeloh, Scott Reed, Jürgen Sturm, and Matthias Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In CVPR, pages 4578-4587, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.454, + 0.483, + 0.509 + ], + "angle": 0, + "content": "[5] Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[6] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[7] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.483, + 0.649 + ], + "angle": 0, + "content": "[8] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPRW, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.65, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.483, + 0.747 + ], + "angle": 0, + "content": "[10] Antoine Guédon and Vincent Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. arXiv preprint arXiv:2403.14554, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[11] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In CVPR, pages 19946-19956, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In ECCV, pages 376-393, 2025. 2, 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.244, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[14] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, pages 4065-4071. IEEE, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.176, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[16] Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, and Anton Konushin. Unidet3d: Multi-dataset indoor 3d object detection. In AAAI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[17] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In CVPR, pages 16329–16339, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[18] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[19] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. RAL, 5(1): 219-226, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[20] Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Semantic gaussian splatting for neural dense slam. In ECCV, pages 163-179, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[21] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[22] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.705, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[27] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, pages 8900-8910, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Ruiyuan Lyu, Tai Wang, Jingli Lin, Shuai Yang, Xiaohan Mao, Yilun Chen, Runsen Xu, Haifeng Huang, Chenming Zhu, Dahua Lin, et al. Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. arXiv preprint arXiv:2406.09401, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26368" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[29] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[32] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In WACV, pages 5763-5773, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[33] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[34] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.433 + ], + "angle": 0, + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgb images. In ECCV, pages 746-760, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[36] Myrna C Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering: Weakly supervised 3d scene segmentation. arXiv preprint arXiv:2404.12784, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[38] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In CVPR, pages 20675-20685, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.605, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[39] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[42] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.869 + ], + "angle": 0, + "content": "[43] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[45] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[46] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[47] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[48] Yuting Xiao, Xuan Wang, Jiafei Li, Hongrui Cai, Yanbo Fan, Nan Xue, Minghui Yang, Yujun Shen, and Shenghua Gao. Bridging 3d gaussian and mesh for freeview video rendering. arXiv preprint arXiv:2403.11453, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.373, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[49] Xiuwei Xu, Chong Xia, Ziwei Wang, Linqing Zhao, Yueqi Duan, Jie Zhou, and Jiwen Lu. Memory-based adapters for online 3d scene perception. arXiv preprint arXiv:2403.06974, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[50] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In CVPR, pages 19595-19604, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.471, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[52] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. arXiv preprint arXiv:2408.14197, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[53] Zeyu Yang, Zijie Pan, Xiatian Zhu, Li Zhang, Yu-Gang Jiang, and Philip HS Torr. 4d gaussian splatting: Modeling dynamic scenes with native 4d primitives. arXiv preprint arXiv:2412.20720, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.703 + ], + "angle": 0, + "content": "[54] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9455-9465, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[55] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[56] Hongxiao Yu, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Monocular occupancy prediction for scalable indoor scenes. arXiv preprint arXiv:2407.11730, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[57] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "26369" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.148 + ], + "angle": 0, + "content": "[59] Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023. 2" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "26370" + } + ] +] \ No newline at end of file diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_origin.pdf b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..dbaedb533f0341a2c02b3a52e95388b1b7e6fbbc --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/ecffb22e-f04b-4d58-841a-1b052e2a208b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9fb9c6f49ce3fff69136317ff227b5488b3255f4928247e40cb405b11c18a962 +size 25270120 diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/full.md b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f27f24c62e3bc5677e03a2fa92a198e2b06abf51 --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/full.md @@ -0,0 +1,318 @@ +# EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding + +Yuqi Wu Wenzhao Zheng Sicheng Zuo Yuanhui Huang Jie Zhou Jiwen Lu Department of Automation, Tsinghua University, China + +wuyq24@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com + +![](images/b4a6e8029c1021c3b3ea302c23baa602e5f0e9efa46d30e49178be860a011b80.jpg) +Figure 1. Given streaming monocular RGB inputs, our EmbodiedOcc conducts embodied occupancy prediction in an online manner for indoor scenes. Different from existing methods which focus on offline perception from monocular images, we focus on the scene-level occupancy prediction from embodied observations. We initialize the scene to be explored with uniform 3D semantic Gaussians and progressively update them based on new observations, similar to how humans explore unknown scenes. + +# Abstract + +3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D + +occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOccScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc. + +# 1. Introduction + +With the rapid development of embodied intelligence and active agents [14, 17, 32], 3D scene perception [30, 34, 41, 42] has become a crucial task in computer vision. Intelligent agents first perceive their surrounding environments and then make decisions based on the perception results. + +Due to the low costs of camera sensors, vision-based 3D occupancy prediction is gaining increasing popularity and produces a comprehensive understanding of both semantics and structures of the scene [2, 11, 13, 46, 56]. + +While vision-based 3D occupancy prediction has made significant progress in outdoor driving scenes [11, 13, 22, 40, 43, 45, 46, 52, 58, 59], the application to indoor scenarios is still challenging due to the diversity and complexity of indoor scenes. Most existing methods [2, 54, 56] still focus on local 3D occupancy prediction by integrating semantic and depth information extracted from the visual inputs. However, different from outdoor scenarios, it is important to obtain a global understanding of the room for indoor scenarios, as it usually requires multiple traversals for embodied agents. Also, it is more practical to progressively explore and update the global occupancy of the 3D scene in an online manner from embodied vision-based observations with different positions and perspectives. + +To bridge this gap, we formulate a new embodied 3D occupancy prediction task to evaluate the ability to progressively explore an unknown scene using only visual inputs. We propose an EmbodiedOcc framework based on Gaussian memories to accomplish this task, considering the explicit and structural nature of 3D Gaussians. We initialize the global scene with uniform 3D semantic Gaussians and progressively update the Gaussians within the field of view observed by the agent. Throughout the exploration process, we maintain an explicit global memory of 3D Gaussians as the global understanding and derive the global 3D occupancy with Gaussian-to-voxel splatting [13]. Specifically, we propose a structure-aware local refinement module to update the relevant Gaussians within the current frustum. We employs a simple yet effective depth-aware branch to introduce explicit structural information for each Gaussian, ensuring the update of these Gaussians to better align with the global representation. During the continuous exploration, we read out Gaussians within the current frustum from the memory as inputs to the local module for refinement. We assign high confidence values for updated Gaussians and use them to reweight information from the memory and the current input. This ensures the consistency of the 3D representation during the fusion and update process. We reorganize an EmbodiedOcc-ScanNet benchmark for the embodied 3D occupancy prediction task based on the locally annotated Occ-ScanNet dataset [3, 47, 56]. Experiments show that our EmbodiedOcc outperforms existing methods by a large margin and accomplishes embodied occupancy prediction with high accuracy and efficiency. + +# 2. Related Work + +3D Occupancy Prediction. Benefiting from its compactness and versatility, 3D occupancy prediction based on multi-view images or additional 3D information [1, 11- + +13, 21, 37, 46] has gained great popularity over the last few years. MonoScene [2] was the first to derive 3D occupancy prediction from a single image, propelling the original 3D Semantic Scene Completion (SSC) [4, 8, 18, 19, 33, 37] into a more challenging stage with vision-only inputs and more universal scenarios (both indoor and outdoor scenes). Subsequent works [54, 56] further focused on addressing the depth ambiguity in this monocular setting. However, most of these efforts were confined to local and offline prediction. SCFusion [47] proposed an incremental framework based on RGB-D inputs. EmbodiedScan [28, 44] introduced an offline global prediction framework using multimodal sequential inputs. Differently, the proposed embodied 3D occupancy prediction aims at online prediction from RGB-only inputs, which is more challenging and practical. + +Online 3D Scene Perception. Accurate comprehension of 3D scenes is an indispensable capability for embodied agents, such as 3D occupancy prediction [2, 56] and object detection [16, 31, 42]. Most existing works on indoor 3D scene perception [9, 30, 41, 55] take pre-acquired and reconstructed 3D data as inputs and perceive the scene in an offline manner. To achieve online perception, Online3D [49] introduced an adapter-based model that equips mainstream offline frameworks with the competence to perform online scene perception, enabling the process of real-time RGB-D sequences. However, this framework still requires depth information as inputs and mainly targets point segmentation and 3D detection. Differently, we target online vision-based 3D occupancy prediction which can provide a more comprehensive understanding of the scene. + +3D Gaussian Splatting. 3D Gaussian Splatting [15] uses 3D Gaussians to model a 3D scene and benefits from fast speed and high quality in the field of neural rendering. The physical characteristics of 3D Gaussians and the splat-based rasterization also motivated rapid advancements in research fields such as scene editing [10, 24, 29, 36], dynamic scenarios [7, 27, 38, 48, 53], and SLAM [5, 20, 50, 57]. GaussianFormer [13] pioneers the application of 3D Gaussians in outdoor 3D occupancy prediction and uses features from multi-view images to update 3D Gaussians, which can be converted into 3D occupancy prediction through a Gaussian-to-voxel splatting module. However, it is still unclear how to employ 3D Gaussians for online global indoor scene understanding from local observations. We achieve this by designing a Gaussian memory mechanism and progressively updating it with structure-aware interaction. + +# 3. Proposed Approach + +# 3.1. Embodied 3D Occupancy Prediction + +Conventional methods in indoor scenarios for occupancy prediction accepted RGB-D as inputs to predict the semantic occupancy of a 3D scene which requires depth sensors. + +![](images/83329bd7533bd4ca726b7428f30052ba3744d5342ae60be45c107bae2757ea56.jpg) +Figure 2. Framework of our EmbodiedOcc for embodied 3D occupancy prediction. We maintain an explicit global memory of 3D Gaussians during the exploration of the current scene. For each update, the Gaussians within the current frustum are taken from the memory and updated using semantic and structural features extracted from the monocular RGB input. Each Gaussian has a confidence value to integrate information from both the memory and the current input. Then we detach and put these updated Gaussians back into the memory. We can obtain the current 3D occupancy prediction using a Gaussian-to-voxel splatting module whenever we need. + +However, we humans are capable of effortlessly processing the visual information from a single view to obtain 3D perception of our surroundings. Recent methods begin to consider endowing models with the same competence, which accept a monocular RGB image as input and derive a 3D occupancy prediction within the current frustum: + +$$ +\mathbf {Y} _ {\text {m o n o}} = \mathcal {F} _ {\text {m o n o}} \left(I _ {\text {m o n o}}\right), \tag {1} +$$ + +where $\mathcal{F}_{mono}$ is the proposed monocular prediction model, $I_{mono} \in \mathbb{R}^{H \times W \times 3}$ and $\mathbf{Y}_{mono} \in \mathbb{R}^{X \times Y \times Z \times C}$ refer to the monocular RGB input and the obtained 3D occupancy prediction. $X, Y, Z$ represent the dimensions of the local 3D scene and $C$ represents the total number of semantics. + +This is only the initial step towards practical scenarios. The essence of human intelligence is the capacity to analyze and respond immediately based on real-time perception of the surroundings. Correspondingly, superior embodied agents are anticipated to process egocentrically gathered real-time visual input to update the 3D occupancy prediction of the current scene. This capability facilitates the execution of downstream tasks based on real-time perception. + +Motivated by this, we propose an embodied 3D occupancy prediction task in this paper. Let $\mathcal{X}_t = \{x_1, x_2, \dots, x_t\}$ be an RGB sequence and the corresponding extrinsics collected by the embodied agent up to the present, where $x_t = (I_t, M_t)$ , $I_t \in \mathbb{R}^{H \times W \times 3}$ , $M_t \in \mathbb{R}^{3 \times 4}$ . It is worth noting that the variation in the subscripts merely represents the change in the position and perspective of the agent when exploring the current scene continuously. Different subscripts may correspond to similar positions and perspectives, indicating that the agent has returned to a previously explored location. In embodied occupancy prediction, re-exploration of the same area should maintain global consistency and even demonstrate improved performance, akin to we hu + +mans always possessing a more comprehensive understanding of sights that have been encountered repeatedly. + +We formulate the function of an embodied occupancy prediction model as follows: + +$$ +\mathbf {Y} _ {t} = \mathcal {F} _ {\text {e m b o d i e d}} \left(\mathbf {Y} _ {t - 1}, x _ {t}\right), \tag {2} +$$ + +where $\mathcal{F}_{\text{embodied}}$ is the embodied prediction model, $\mathbf{Y}_t \in \mathbb{R}^{X_{\text{room}} \times Y_{\text{room}} \times Z_{\text{room}} \times C}$ refers to the current occupancy prediction of the whole scene ( $\mathbf{Y}_0$ is the initialization). $X_{\text{room}}, Y_{\text{room}}, Z_{\text{room}}$ denote the scene dimensions. + +# 3.2. Local Refinement Module + +Different from conventional methods that conducted feature integration in a voxelized space, we use a set of 3D semantic Gaussians to represent an indoor scene [13]. In this subsection, we will first explain our local refinement module, which extracts semantic and structural features from the monocular input and integrates them to update the Gaussian-based representation of the current frustum. + +Initialization. We first initialize a set of semantic Gaussians to represent the current frustum. Each semantic Gaussian $\mathbf{G}$ is represented by a vector comprising mean $\mathbf{m} \in \mathbb{R}^3$ , scale $\mathbf{s} \in \mathbb{R}^3$ , rotation quaternion $\mathbf{r} \in \mathbb{R}^4$ , opacity $\mathbf{o} \in \mathbb{R}$ , and semantic logits $\mathbf{c} \in \mathbb{R}^C$ ( $C$ denotes the total number of semantic categories). We use an embedding layer to lift each Gaussian vector $\mathbf{G}$ to its corresponding high-dimensional feature vectors $\mathbf{Q}$ , and derive $\mathcal{Q} = \{\mathbf{Q}_i \in \mathbb{R}^m, i = 1, \dots, N\}$ , where $m$ is the dimension of $\mathbf{Q}_i$ and $N$ is the total number of the Gaussians. + +Depth-Aware Branch. Due to the variable scales and tight arrangements of indoor objects, depth ambiguity has always been one of the core challenges limiting the performance of indoor occupancy prediction models in monocular settings. Previous work has consistently focused on how to + +![](images/a5f0e7fd302ed683f6f3396f6d471b6603310e1ecc57aa244de0512451669db9.jpg) +Figure 3. Motivation of the depth-aware branch. Along a specific ray, Gaussians distributed in front of the true depth point are likely to model the empty semantic (A). Gaussians distributed behind the true depth point closely are likely to model valid semantics (B). Gaussians that are distributed behind the true depth point but are too far away require more information to guide their updates (C). During the embodied exploration, the subsequent frames can make up for this lack of information in the current frame. + +better extract and utilize depth information from the input image. We design a depth-aware branch to provide more accurate and effective guidance for the update of 3D semantic Gaussians in our local refinement module. + +We first use a depth prediction network to obtain a relatively accurate depth map $D_{\text{metric}}$ from input $I_{mono}$ . A naive approach can explicitly utilize this depth information when initializing the Gaussians, e.g., we can randomly sample some points from the pseudo point cloud recovered from the depth map and use these coordinates to initialize the means of some Gaussians. Although providing direct hints for the means of some Gaussians, this cannot exploit the potential of the depth information. We design a simple yet effective depth-aware layer to accomplish this. We still uniformly initialize a number of Gaussians within the current frustum. For each Gaussian, we project its mean $\mathbf{m}$ into the pixel coordinate system through the intrinsics $K_{mono} \in \mathbb{R}^{3 \times 3}$ and obtain the depth value $d$ . The sampled depth value $d$ , along with the z-component $z$ of the Gaussian mean in the camera coordinate system, are fed into the depth-aware layer, which is a multi-layer perceptron (MLP) that outputs the depth-aware feature $\mathbf{Q}_{depth}$ for this Gaussian. Then we add the depth-aware feature to the original feature vector $\mathbf{Q}$ , injecting additional information into the subsequent feature integration. In this way, depth information not only affects the means of the Gaussians but also promotes the update of other properties: + +$$ +\mathbf {Q} _ {\text {d e p t h}} = \mathcal {M} _ {\text {d e p t h a w a r e}} \left(\left(D _ {\text {m e t r i c}} (u, v), z\right), \right. +$$ + +$$ +\hat {\mathcal {Q}} = \left\{\hat {\mathbf {Q}} _ {i}, i = 1, \dots , N \mid \hat {\mathbf {Q}} _ {i} = \mathbf {Q} _ {i} + \mathbf {Q} _ {i} ^ {\text {d e p t h}} \right\}, \tag {3} +$$ + +where $\mathcal{M}_{depthaware}$ is the depth-aware layer, $(u,v)$ are pixel coordinates of each Gaussian. We illustrate our depth-aware branch in Figure 3. + +Feature Integration and Gaussian Refinement. Feature integration in our local refinement module includes the + +interactions among Gaussians as well as the interactions between image features and Gaussians. We voxelize the Gaussian centers and conduct 3D sparse convolution on the generated grid to allow interactions among Gaussian vectors $\hat{\mathcal{Q}}$ . We project the Gaussian centers onto the image feature map and use the deformable attention function to integrate the queried features and the Gaussian vectors $\hat{\mathcal{Q}}$ . After the prior feature integration, these feature vectors with aggregated information will be used to obtain the update amounts $\Delta \mathbf{G} = (\Delta \mathbf{m}, \Delta \mathbf{s}, \Delta \mathbf{r}, \Delta \mathbf{o}, \Delta \mathbf{c})$ of each Gaussian. We use the update amounts $\Delta \mathbf{G}$ to refine the Gaussian properties: + +$$ +\mathbf {G} _ {\text {n e w}} = (\Delta \mathbf {m} + \mathbf {m}, \Delta \mathbf {s} + \mathbf {s}, \Delta \mathbf {r} \otimes \mathbf {r}, \Delta \mathbf {o} + \mathbf {o}, \Delta \mathbf {c} + \mathbf {c}), \tag {4} +$$ + +where $\otimes$ refers to the special composition of quaternions. + +We conduct the feature integration and the refinement of Gaussians multiple times. After the final refinement, we use a Gaussian-to-voxel splatting module [13] to obtain the final occupancy within the frustum. + +# 3.3. Gaussian Memory Updated Online + +To explore unknown scenes, we humans continuously update the objects within the scene and their relationships to gradually construct a global scene memory. When revisiting this scene for further exploration, we use the visual information to refine this memory. Inspired by this, we design an online framework (shown in Figure 2) and maintain a Gaussian memory for global understanding. + +Memory Initialization. Our local refinement module initializes and updates Gaussians in the camera coordinate system, as the extrinsics in indoor scenarios are constantly changing, which will pose additional difficulties for our local module. But in the final embodied framework, we initialize the entire scene with uniform Gaussians in the world coordinate system. For a novel scene to be explored, we have: $\mathcal{G}_{\text{room}} = \{(G_i, \gamma_i), i = 1, \dots, N | G_i = (\mathbf{m}_i, \mathbf{s}_i, \mathbf{r}_i, \mathbf{o}_i, \mathbf{c}_i), \gamma_i = 0, 1\}$ , where $N$ refers to the number of Gaussians to initialize this scene, $\mathbf{m}_i$ and $\mathbf{r}_i$ are the means and rotation quaternions of these Gaussians in the world coordinate system ( $\mathbf{s}_i$ , $\mathbf{o}_i$ and $\mathbf{c}_i$ maintain consistency between the world and camera coordinate systems). We introduce an additional tag $\gamma$ for all the Gaussians in the memory. When initializing a novel scene, tags of these Gaussians are set to 0. Every time we put some updated Gaussians back into the memory, their tags are set to 1. + +Memory Update. At the current step $t$ , our embodied occupancy prediction framework receives a posed visual input $x_{t} = (I_{t}, M_{t})$ to perform the update. During the current update, we use a mask from coordinate system transformation to get all Gaussians $\mathcal{G}_{t}$ within the current frustum from the memory. These Gaussians will interact and be refined using a tailored confidence refinement module. Then we detach these Gaussians and put them back into the memory. + +Confidence Refinement. Apart from the initial update for each scene which is akin to the local refinement, sub + +![](images/efc113c114e99c8a76bf3a224fc9b8a473dd484b7b70563fcff1d89c657f662f.jpg) +Figure 4. Illustration of our Gaussian memory. During each update, the Gaussians within the current frustum are taken from the memory. Confidence values of those well-updated Gaussians are used to integrate information from both the memory and the current input. Then we put these Gaussians back into the memory. + +sequent exploration involves the update of Gaussians from the Gaussian memory, among which some have been well-updated by previous frames (if we can derive an acceptable local occupancy prediction from these Gaussians, we believe that they have been "well-updated") and some still remain random. It is unreasonable to update these Gaussians equally. For those Gaussians deemed well-updated, we only need to refine them slightly based on the semantic and structural features extracted from the current image, which is exactly the essence of maintaining the Gaussian memory. As for those random Gaussians that have never been updated, we can directly update them with a fresh perspective. + +To elaborate, we generate a confidence value $\theta$ for each Gaussian taken from the memory. For those having been previously updated $(\gamma = 1)$ , we set their confidence values to a certain value between 0 and 1, which means we will integrate information from both the memory and the newly observed image to update these Gaussians. For those that have never been updated, we set their confidence values to 0 and follow the same refinement module in Sec. 3.2: + +$$ +\Delta \mathbf {G} _ {\text {o n l i n e}} = (1 - \theta) \Delta \mathbf {G}, \tag {5} +$$ + +$$ +\mathbf {G} _ {\text {a f t e r}} = \Delta \mathbf {G} _ {\text {o n l i n e}} \oplus \mathbf {G} _ {\text {b e f o r e}}, +$$ + +where we use $\oplus$ to represent the composition of rotation quaternions and the add operation of other parts. Figure 4 illustrates how we maintain the Gaussian memory. + +Stopping Mechanism. We propose a simple stopping mechanism to consider a room as having been effectively explored. At the step $t$ , we first calculate a confidence ratio $\alpha$ to measure the exploration of the current room: + +$$ +\alpha = \sum_ {i = 0} ^ {N} \mathbb {I} _ {\gamma_ {i} = 1} / N, \tag {6} +$$ + +where $\mathbb{I}_{\gamma_i = 1}$ takes the value of 1 if $\gamma_{i} = 1$ . If $\alpha$ exceeds a certain threshold we set before, the model can decide to enter an adjacent room to begin a new exploration or stay here to get a better perception of the current room. + +# 3.4. EmbodiedOcc: An Embodied Framework + +We present the training framework of our EmbodiedOcc model for indoor embodied occupancy prediction. During the whole prediction process, we use the current monocular input to update our Gaussian memory in real time, which can be easily converted into 3D occupancy prediction. + +We first train our local refinement module using the focal loss $L_{focal}$ , the lovasz-softmax loss $L_{lov}$ , the scene-class affinity loss $L_{scal}^{geo}$ and $L_{scal}^{sem}$ following RetinaNet [23], TPV-Former [11] and MonoScene [2]. We use monocular occupancy within the frustum $\mathbf{Y}_{mono}^{fov}$ and the corresponding ground truth $\mathbf{Y}_{gt}^{fov}$ to compute the loss: + +$$ +\begin{array}{l} \mathcal {L} = \lambda_ {1} \mathcal {L} _ {\text {f o c a l}} \left(\mathbf {Y} _ {\text {m o n o}} ^ {\text {f o v}}, \mathbf {Y} _ {\text {g t}} ^ {\text {f o v}}\right) + \mathcal {L} _ {\text {l o v}} \left(\mathbf {Y} _ {\text {m o n o}} ^ {\text {f o v}}, \mathbf {Y} _ {\text {g t}} ^ {\text {f o v}}\right) \tag {7} \\ + \mathcal {L} _ {\text {s c a l}} ^ {\text {g e o}} (\mathbf {Y} _ {\text {m o n o}} ^ {f o v}, \mathbf {Y} _ {g t} ^ {f o v}) + \mathcal {L} _ {\text {s c a l}} ^ {\text {s e m}} (\mathbf {Y} _ {\text {m o n o}} ^ {f o v}, \mathbf {Y} _ {g t} ^ {f o v}), \\ \end{array} +$$ + +where $\lambda_{1}$ is a balance factor. + +We then use the trained local module to train our EmbodiedOcc. For efficient training, we initialize the Gaussian memory before the first update and compute the current loss following the equation 7 after each update. To ensure consistency, the local occupancy ground truth used here is obtained from the occupancy of the whole scene. After a certain number of updates, we re-initialize the memory and come to the next scene. Trained with such a pipeline, our EmbodiedOcc can effectively perform the embodied occupancy prediction task while ensuring consistency within the same scene. We expect that our EmbodiedOcc can have an improving prediction with continuous exploration rather than undermining previous predictions when encountering parts that have been explored before. Therefore, we conduct some tailored tests to validate the capability of our model. + +# 4. Experiments + +# 4.1. EmbodiedOcc-ScanNet Benchmark + +Task Descriptions. We conducted two tasks to evaluate our EmbodiedOcc framework: local occupancy prediction and embodied occupancy prediction. Local occupancy prediction shares the same setting with previous works, which accept monocular image as input and obtain the occupancy within the current frustum. Embodied occupancy prediction accepts real-time visual inputs continuously and updates the occupancy of the current scene online. The visual input at a certain step $t$ during embodied occupancy prediction is still monocular, which is a more challenging setting compared with multi-view input or input with 3D information. + +Datasets. We trained and evaluated our local refinement module on the Occ-ScanNet dataset [56], which provides frames in $60 \times 60 \times 36$ voxel grids (a $4.8m \times 4.8m \times 2.88m$ box in front of the camera). These frames are labeled with 12 semantics, including 11 for valid semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects) and 1 for empty space. + +Table 1. Local Prediction Performance on the Occ-ScanNet dataset. + +
MethodInputIoUceilingfloorwallwindowchairbedsofatablevtsfurnitureobjectsmIoU
TPVFormer [11]xrgb33.396.9632.9714.419.1024.0141.4945.4428.6110.6635.3725.3124.94
GaussianFormer [13]xrgb40.9120.7042.0023.4017.4027.044.3044.8032.7015.3036.7025.0029.93
MonoScene [2]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.8424.62
ISO [56]xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.6128.71
Surroundocc [46]xrgb42.5218.9049.3024.8018.0026.8042.0044.1032.9018.6036.8026.9030.83
Oursxrgb53.5539.6050.4041.4031.7040.9055.0061.4044.0036.1053.9042.2045.15
+ +Table 2. Embodied Prediction Performance on the EmbodiedOcc-ScanNet dataset. + +
MethodDatasetIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjectsmIoU
TPVFormer [11]EmbodiedOcc35.881.6230.5412.0313.2235.4751.3949.7925.633.6043.1516.2325.70
SurroundOcc [46]EmbodiedOcc37.0412.7031.8022.5022.0029.9044.7036.5024.6011.5034.4018.2026.27
GaussianFormer [13]EmbodiedOcc38.0217.0033.6021.5021.7029.4047.8037.1024.3015.5036.2016.8027.36
SplicingOccEmbodiedOcc49.0131.6038.8035.5036.3047.1054.5057.2034.4032.5051.2029.1040.74
EmbodiedOccEmbodiedOcc51.5222.7044.6037.4038.0050.1056.7059.7035.4038.4052.0032.9042.53
+ +Based on this dataset, we reorganized an EmbodiedOccScanNet dataset to train and evaluate our EmbodiedOcc framework [35, 56]. Our EmbodiedOcc-ScanNet comprises 537/137 scenes in the train/val splits. Each scene consists of 30 posed frames with their corresponding local occupancies. The resolutions of global occupancy of each scene are calculated by $l_{x} \times l_{y} \times l_{z} / (0.08m)^{3}$ , where $l_{x} \times l_{y} \times l_{z}$ is the range of this scene in the world coordinate system. In addition, we associate grid points that can be projected onto the camera plane for each frame as the global mask of this frame. By splicing the global mask of all processed frames, we can easily obtain the occupancy ground truth of the explored part in the current scene. + +Apart from Occ-ScanNet and EmbodiedOcc-ScanNet datasets in the original scale, we sampled a small set from the EmbodiedOcc-ScanNet dataset as the EmbodiedOcc-ScanNet-mini dataset which comprises 64/16 scenes in the train/val splits. We sampled from the Occ-ScanNet dataset accordingly and obtained an Occ-ScanNet-mini2 dataset, which comprises 5504/2376 frames in the train/val splits. We conducted the local occupancy prediction task on the Occ-ScanNet and Occ-ScanNet-mini2 datasets and conducted the embodied prediction task on the EmbodiedOcc-ScanNet and EmbodiedOcc-ScanNet-mini datasets. + +Evaluation Metrics. We use mIoU and IoU as the evaluation metrics. For local occupancy prediction, we calculate the mIoU and IoU using the occupancy within the current frustum (same with the evaluation in ISO [56]). For embodied occupancy prediction, we calculate the mIoU and IoU using the global occupancy of the current scene. It is worth mentioning that the global occupancy used here is the union + +of the frustums corresponding to 30 frames of each scene, which represents the region that has been explored. + +# 4.2. Implementation Details + +Local Refinement Module. Following existing works [13, 56], we use a pre-trained EfficientNet-B7 [39] to initialize the image encoder in our local module. The depth prediction network used in the depth-aware branch is a fine-tuned DepthAnything-V2 model [51] that remains frozen during the training, and the depth-aware layer is a 3-layer MLP. The resolutions of the monocular input are set to $480 \times 640$ and the number of Gaussians used to conduct the local prediction is 16200. We utilize the AdamW [26] optimizer with a weight decay of 0.01. The learning rate warms up in the first 1000 iterations to a maximum value of 2e-4 and decreases according to a cosine schedule [25]. We train our local refinement module for 10 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the Occ-ScanNet dataset and 20 epochs on the Occ-ScanNet-mini2 dataset. + +EmbodiedOcc Framework. We initialize the Gaussians with a $0.16\mathrm{m}$ interval to represent a novel scene. For each update, the confidence value $\theta$ of well-updated Gaussians is set to 0 in the first two refinement layers (frozen) and 0.5 in the final refinement layer. We train our EmbodiedOcc for 5 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet dataset and 20 epochs using 4 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet-mini dataset. The maximum value of the learning rate is set to 2e-4 using 8 GPUs and 1e-4 using 4 GPUs. The other settings remain the same with the training of the local refinement module. + +Table 3. Look-Back Prediction vs First-Time Prediction. For $\mathrm{K} = k$ , we simply select 0, 1, ..., k - 1th frames to evaluate our EmbodiedOcc framework and the occupancy ground truth used here is the union of the $k$ frustums. K was set to 3/5/8. + +
ModeKFrame ListIoUmIoU
First-Time3[0,1,2]49.3939.32
Look-Back3[0,1,2,1,0]49.5240.09
First-Time5[0,...,4]50.1340.03
Look-Back5[0,...,3,4,3,...,0]50.6440.98
First-Time8[0,...,7]50.9440.86
Look-Back8[0,...,6,7,6,...,0]51.1441.17
+ +![](images/3c6074de3b7fe941098c5799c9e632025327f8bb247d5930e6b0525a35464491.jpg) +Figure 5. Performance with different stopping ratios. + +![](images/439b848185db8c512927307636dc02c494ba2ea477da018d637dcf35c96f68d6.jpg) +Figure 6. Ablation study of the confidence refinement. + +# 4.3. Main Results + +Local Occupancy Prediction. We evaluated our local refinement module on the Occ-ScanNet dataset [56]. As shown in Table 1, the results indicate that our local refinement module outperforms ISO [56]. We also implemented several state-of-the-art driving scene methods [11, 13, 46] on this benchmark and our local refinement module outperforms them by a large margin. This is because they mainly focus on the coarse layout (e.g., positions of objects) while indoor scenes require modeling of the fine-grained structure (e.g., shapes of objects). + +Embodied Occupancy Prediction. We assessed the occupancy prediction for the entire scene after processing 30 frames, and the ground truth for calculating IoU and mIoU is the union of the frustums. We spliced the local occupancy obtained from our local module to serve as the main baseline (referred to as SplicingOcc), as our local module has achieved the best local performance to date. It can be observed in Table 2 that our EmbodiedOcc exhibits superior prediction of the scene, which is achieved through the integration of different views. We also compared our EmbodiedOcc with the driving scene methods mentioned before (we obtained their embodied results by voting from different local predictions). Their poor results are due to ignoring the continuity of the observations without a global memory. + +# 4.4. Experimental Analysis + +Effect of Continuous Online Updating. We expect EmbodiedOcc to have better performance when encountering parts that have been explored before, and thus, we designed a Look-Back evaluation on the EmbodiedOccScanNet dataset. Specifically, after processing K frames, + +Table 4. Analysis of the model design. + +
MethodGaussianStructureMemoryLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
EmbodiedOcc-Voxel×✓✓47.5038.1237.5326.99
EmbodiedOcc w/o memory✓✓×53.5545.1549.0140.74
EmbodiedOcc✓✓✓53.5545.1551.5242.53
+ +Table 5. Analysis of the depth-aware branch. + +
Branch TypeDepth Estimation ModuleLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
Depth-aware branchDepthAnything-V253.9346.2050.7841.45
No-depth branch/48.1540.0737.5230.73
Naive-depth branchDepthAnything-V250.3242.73//
Depth-aware branchIndoorDepth51.2443.8746.4237.78
+ +Table 6. Analysis of the Gaussian parameters. + +
Gaussian Number (In local box)Gaussian Scale Min(m)Max(m)Gaussian Interval(m) (In global scene)Local Prediction IoUmIoUEmbodied Prediction IoUmIoU
162000.010.08(0.16, 0.16, 0.16)53.9346.2050.7841.45
81000.010.08(0.20, 0.20, 0.20)50.4742.8246.2437.99
162000.010.20(0.16, 0.16, 0.16)51.5743.7448.0938.40
+ +Table 7. Runtime decomposition. + +
Scene level (ms)Scene init.6.626Occ head39.635
Frame level (ms)Load memory0.973Depth aware1.816
Img backbone61.478GS Encoder14.761
Depthanything34.687Update memory0.474
+ +we direct the model to reprocess the last $\mathrm{K} - 1$ frames. By comparing this Look-Back result with the First-Time prediction, we verified that our EmbodiedOcc has met our expectations as shown in Table 3. + +Effect of the Stopping Mechanism. We use Figure 5 to show the effectiveness of our stopping mechanism. The ground truth used here for calculating IoU and mIoU is the union occupancy of the 30 frustums in a global scene. We observed that using a larger threshold results in more observations and better performance. + +Analysis of the Confidence Refinement. During each update, local Gaussians are refined through three refinement layers. For Gaussians that have been updated before, we froze the first two layers and updated them in the last refinement layer when training our EmbodiedOcc. Figure 6 on the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets shows the impact of different confidence values (determines the coefficient of each $\Delta G$ ). We observe that moderate updates to those previously processed Gaussians yield the best embodied occupancy prediction. + +Analysis of the Model Design. The essence of our EmbodiedOcc is the explicit Gaussian memory. We adopt object-centric Gaussians instead of grid-based voxels since Gaussians are more flexible for local-global interaction. We implemented a voxel version of our EmbodiedOcc and evaluated it on our benchmark. As shown in Table 4, the satisfactory local yet poor embodied performance of EmbodiedOcc in the voxel version verified our conclusion. Results in Table 4 were evaluated on the Occ-ScanNet and EmbodiedOcc-ScanNet datasets. + +Analysis of the Depth-Aware Branch. We analyze the effect of our depth-aware branch in Table 5 using the + +![](images/115be923ff84fe4ad5d2263d128bc65fde7a8a543816755979996082009d3095.jpg) +Figure 7. Visualization of the embodied occupancy prediction. We visualize the update of Gaussian memory and corresponding global occupancy. As the Gaussians transition from random to ordered, the occupancy of the current scene becomes more accurate and complete. + +![](images/c70f89a4b4041829c48b0571c878debc79ee313bd4540506f3d2174e31b8167c.jpg) +Figure 8. Visualization of local occupancy prediction. + +Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets. We find that depth information will significantly benefit the local and embodied occupancy prediction. As shown in the second row, without the assistance of depth information, the performance of embodied occupancy prediction drops sharply. This indicates that the update of Gaussians within the current frustum may corrupt previous predictions without the guidance of depth information. The results in the third row suggest that the depth-aware branch we employ is more reasonable compared to the naive method of directly initializing a portion of Gaussians with the pseudo point cloud recovered from the predicted depth map, the latter also poses difficulties for the initialization of global Gaussians so we do not provide the embodied results. Besides, we replaced DepthAnything-V2 with IndoorDepth [6] in the last row to prove that our depth-aware branch does not rely on a specific depth prediction network. + +Analysis of the Gaussian Parameters. We analyze the effect of different Gaussian parameters in Table 6 using the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini + +datasets. We see that decreasing the number or increasing the scale of the Gaussians can lead to a decrease in performance during local and embodied occupancy prediction. This is closely related to the physical properties of Gaussians. Gaussians initialized too sparse may lead to holes in occupancy prediction, while Gaussians with too large scale will overlap and influence each other which is also detrimental to the correct prediction of occupancy. + +Runtime Analysis. We present in Table 7 a runtime analysis on scene 0687-00 from the EmbodiedOcc-ScanNet dataset. The runtime decomposition details show that our method is efficient while the main bottleneck is the image and depth backbones, suggesting that the overall runtime of our EmbodiedOcc can be further reduced. + +Visualizations. Figure 7 and 8 visualize the global and local predictions, respectively. Our model demonstrates reasonable local perception ability and further achieves good online prediction with the Gaussian memory. Due to space limitations, we will use a more diverse set of samples to further show the visual effect of our EmbodiedOcc in the supplementary material. + +# 5. Conclusion + +In this paper, we have presented an embodied 3D occupancy prediction task and proposed a Gaussian-based EmbodiedOcc framework accordingly. Our EmbodiedOcc maintains an explicit Gaussian memory of the current scene and updates this memory during the exploration of this scene. Both quantitative and visualization results have shown that our EmbodiedOcc outperforms existing methods in terms of local occupancy prediction and accomplishes the embodied occupancy prediction task with high accuracy and strong expandability. We believe that our EmbodiedOcc paves the way for enabling active agents to conduct accurate and flexible embodied occupancy prediction. + +# Acknowledgements + +We would like to thank Tianyu Hu for her valuable assistance with the experiments. This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, and Grant 62336004, and in part by the Beijing Natural Science Foundation under Grant No. L247009. + +# References + +[1] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2 +[2] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 2, 5, 6 +[3] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2 +[4] Angela Dai, Daniel Ritchie, Martin Bokeloh, Scott Reed, Jürgen Sturm, and Matthias Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In CVPR, pages 4578-4587, 2018. 2 +[5] Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2 +[6] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 8 +[7] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2 +[8] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPRW, pages 0–0, 2019. 2 +[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2 +[10] Antoine Guédon and Vincent Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. arXiv preprint arXiv:2403.14554, 2024. 2 +[11] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2, 5, 6, 7 +[12] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In CVPR, pages 19946-19956, 2024. +[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In ECCV, pages 376-393, 2025. 2, 3, 4, 6, 7 + +[14] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, pages 4065-4071. IEEE, 2022. 1 +[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2 +[16] Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, and Anton Konushin. Unidet3d: Multi-dataset indoor 3d object detection. In AAAI, 2025. 2 +[17] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In CVPR, pages 16329–16339, 2024. 1 +[18] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2 +[19] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. RAL, 5(1): 219-226, 2019. 2 +[20] Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Semantic gaussian splatting for neural dense slam. In ECCV, pages 163-179, 2025. 2 +[21] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, 2023. 2 +[22] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 2 +[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5 +[24] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024. 2 +[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[27] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, pages 8900-8910, 2024. 2 +[28] Ruiyuan Lyu, Tai Wang, Jingli Lin, Shuai Yang, Xiaohan Mao, Yilun Chen, Runsen Xu, Haifeng Huang, Chenming Zhu, Dahua Lin, et al. Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. arXiv preprint arXiv:2406.09401, 2024. 2 + +[29] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2 +[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1, 2 +[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2 +[32] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In WACV, pages 5763-5773, 2024. 1 +[33] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2 +[34] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493, 2022. 1 +[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgb images. In ECCV, pages 746-760, 2012. 6 +[36] Myrna C Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering: Weakly supervised 3d scene segmentation. arXiv preprint arXiv:2404.12784, 2024. 2 +[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2 +[38] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In CVPR, pages 20675-20685, 2024. 2 +[39] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6 +[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 2 +[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2 +[42] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 1, 2 +[43] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2 +[44] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai + +Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 2 +[45] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 2 +[46] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 2, 6, 7 +[47] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2 +[48] Yuting Xiao, Xuan Wang, Jiafei Li, Hongrui Cai, Yanbo Fan, Nan Xue, Minghui Yang, Yujun Shen, and Shenghua Gao. Bridging 3d gaussian and mesh for freeview video rendering. arXiv preprint arXiv:2403.11453, 2024. 2 +[49] Xiuwei Xu, Chong Xia, Ziwei Wang, Linqing Zhao, Yueqi Duan, Jie Zhou, and Jiwen Lu. Memory-based adapters for online 3d scene perception. arXiv preprint arXiv:2403.06974, 2024. 2 +[50] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In CVPR, pages 19595-19604, 2024. 2 +[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 6 +[52] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. arXiv preprint arXiv:2408.14197, 2024. 2 +[53] Zeyu Yang, Zijie Pan, Xiatian Zhu, Li Zhang, Yu-Gang Jiang, and Philip HS Torr. 4d gaussian splatting: Modeling dynamic scenes with native 4d primitives. arXiv preprint arXiv:2412.20720, 2024. 2 +[54] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9455-9465, 2023. 2 +[55] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2 +[56] Hongxiao Yu, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Monocular occupancy prediction for scalable indoor scenes. arXiv preprint arXiv:2407.11730, 2024. 2, 5, 6, 7 +[57] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2 +[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, 2024. 2 + +[59] Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023. 2 \ No newline at end of file diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/images.zip b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c08fe4220354dd275ba3058942f23ba9f7108427 --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a7e224f8ca080a68c251e7e0ac488e4fd50e71753f5ae63675f1ddc1ea5bd0af +size 826645 diff --git a/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/layout.json b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2a0a475618242f8519df9bb101c65188005a0f09 --- /dev/null +++ b/2025/EmbodiedOcc_ Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding/layout.json @@ -0,0 +1,8447 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 69, + 467, + 106 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 69, + 467, + 106 + ], + "spans": [ + { + "bbox": [ + 144, + 69, + 467, + 106 + ], + "type": "text", + "content": "EmbodiedOcc: Embodied 3D Occupancy Prediction for Vision-based Online Scene Understanding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 126, + 512, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 126, + 512, + 156 + ], + "spans": [ + { + "bbox": [ + 99, + 126, + 512, + 156 + ], + "type": "text", + "content": "Yuqi Wu Wenzhao Zheng Sicheng Zuo Yuanhui Huang Jie Zhou Jiwen Lu Department of Automation, Tsinghua University, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 154, + 158, + 455, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 154, + 158, + 455, + 169 + ], + "spans": [ + { + "bbox": [ + 154, + 158, + 455, + 169 + ], + "type": "text", + "content": "wuyq24@mails.tsinghua.edu.cn; wenzhao.zheng@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 61, + 180, + 553, + 413 + ], + "blocks": [ + { + "bbox": [ + 61, + 180, + 553, + 413 + ], + "lines": [ + { + "bbox": [ + 61, + 180, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 180, + 553, + 413 + ], + "type": "image", + "image_path": "b4a6e8029c1021c3b3ea302c23baa602e5f0e9efa46d30e49178be860a011b80.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Given streaming monocular RGB inputs, our EmbodiedOcc conducts embodied occupancy prediction in an online manner for indoor scenes. Different from existing methods which focus on offline perception from monocular images, we focus on the scene-level occupancy prediction from embodied observations. We initialize the scene to be explored with uniform 3D semantic Gaussians and progressively update them based on new observations, similar to how humans explore unknown scenes." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 470, + 201, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 470, + 201, + 483 + ], + "spans": [ + { + "bbox": [ + 151, + 470, + 201, + 483 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 497, + 297, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 297, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 297, + 689 + ], + "type": "text", + "content": "3D occupancy prediction provides a comprehensive description of the surrounding scenes and has become an essential task for 3D perception. Most existing methods focus on offline perception from one or a few views and cannot be applied to embodied agents that demand to gradually perceive the scene through progressive embodied exploration. In this paper, we formulate an embodied 3D occupancy prediction task to target this practical scenario and propose a Gaussian-based EmbodiedOcc framework to accomplish it. We initialize the global scene with uniform 3D semantic Gaussians and progressively update local regions observed by the embodied agent. For each update, we extract semantic and structural features from the observed image and efficiently incorporate them via deformable cross-attention to refine the regional Gaussians. Finally, we employ Gaussian-to-voxel splatting to obtain the global 3D" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 471, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 628 + ], + "type": "text", + "content": "occupancy from the updated 3D Gaussians. Our EmbodiedOcc assumes an unknown (i.e., uniformly distributed) environment and maintains an explicit global memory of it with 3D Gaussians. It gradually gains knowledge through the local refinement of regional Gaussians, which is consistent with how humans understand new scenes through embodied exploration. We reorganize an EmbodiedOccScanNet benchmark based on local annotations to facilitate the evaluation of the embodied 3D occupancy prediction task. Our EmbodiedOcc outperforms existing methods by a large margin and accomplishes the embodied occupancy prediction with high accuracy and efficiency. Code: https://github.com/YkiWu/EmbodiedOcc." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 633, + 395, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 395, + 646 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 395, + 646 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "With the rapid development of embodied intelligence and active agents [14, 17, 32], 3D scene perception [30, 34, 41, 42] has become a crucial task in computer vision. Intelligent agents first perceive their surrounding environments and then make decisions based on the perception results." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 702, + 153, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 702, + 153, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 702, + 153, + 713 + ], + "type": "text", + "content": "Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26360" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "Due to the low costs of camera sensors, vision-based 3D occupancy prediction is gaining increasing popularity and produces a comprehensive understanding of both semantics and structures of the scene [2, 11, 13, 46, 56]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 287 + ], + "type": "text", + "content": "While vision-based 3D occupancy prediction has made significant progress in outdoor driving scenes [11, 13, 22, 40, 43, 45, 46, 52, 58, 59], the application to indoor scenarios is still challenging due to the diversity and complexity of indoor scenes. Most existing methods [2, 54, 56] still focus on local 3D occupancy prediction by integrating semantic and depth information extracted from the visual inputs. However, different from outdoor scenarios, it is important to obtain a global understanding of the room for indoor scenarios, as it usually requires multiple traversals for embodied agents. Also, it is more practical to progressively explore and update the global occupancy of the 3D scene in an online manner from embodied vision-based observations with different positions and perspectives." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 289, + 295, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 289, + 295, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 289, + 295, + 647 + ], + "type": "text", + "content": "To bridge this gap, we formulate a new embodied 3D occupancy prediction task to evaluate the ability to progressively explore an unknown scene using only visual inputs. We propose an EmbodiedOcc framework based on Gaussian memories to accomplish this task, considering the explicit and structural nature of 3D Gaussians. We initialize the global scene with uniform 3D semantic Gaussians and progressively update the Gaussians within the field of view observed by the agent. Throughout the exploration process, we maintain an explicit global memory of 3D Gaussians as the global understanding and derive the global 3D occupancy with Gaussian-to-voxel splatting [13]. Specifically, we propose a structure-aware local refinement module to update the relevant Gaussians within the current frustum. We employs a simple yet effective depth-aware branch to introduce explicit structural information for each Gaussian, ensuring the update of these Gaussians to better align with the global representation. During the continuous exploration, we read out Gaussians within the current frustum from the memory as inputs to the local module for refinement. We assign high confidence values for updated Gaussians and use them to reweight information from the memory and the current input. This ensures the consistency of the 3D representation during the fusion and update process. We reorganize an EmbodiedOcc-ScanNet benchmark for the embodied 3D occupancy prediction task based on the locally annotated Occ-ScanNet dataset [3, 47, 56]. Experiments show that our EmbodiedOcc outperforms existing methods by a large margin and accomplishes embodied occupancy prediction with high accuracy and efficiency." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 142, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "3D Occupancy Prediction. Benefiting from its compactness and versatility, 3D occupancy prediction based on multi-view images or additional 3D information [1, 11-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "13, 21, 37, 46] has gained great popularity over the last few years. MonoScene [2] was the first to derive 3D occupancy prediction from a single image, propelling the original 3D Semantic Scene Completion (SSC) [4, 8, 18, 19, 33, 37] into a more challenging stage with vision-only inputs and more universal scenarios (both indoor and outdoor scenes). Subsequent works [54, 56] further focused on addressing the depth ambiguity in this monocular setting. However, most of these efforts were confined to local and offline prediction. SCFusion [47] proposed an incremental framework based on RGB-D inputs. EmbodiedScan [28, 44] introduced an offline global prediction framework using multimodal sequential inputs. Differently, the proposed embodied 3D occupancy prediction aims at online prediction from RGB-only inputs, which is more challenging and practical." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 432 + ], + "type": "text", + "content": "Online 3D Scene Perception. Accurate comprehension of 3D scenes is an indispensable capability for embodied agents, such as 3D occupancy prediction [2, 56] and object detection [16, 31, 42]. Most existing works on indoor 3D scene perception [9, 30, 41, 55] take pre-acquired and reconstructed 3D data as inputs and perceive the scene in an offline manner. To achieve online perception, Online3D [49] introduced an adapter-based model that equips mainstream offline frameworks with the competence to perform online scene perception, enabling the process of real-time RGB-D sequences. However, this framework still requires depth information as inputs and mainly targets point segmentation and 3D detection. Differently, we target online vision-based 3D occupancy prediction which can provide a more comprehensive understanding of the scene." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 434, + 553, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 553, + 624 + ], + "type": "text", + "content": "3D Gaussian Splatting. 3D Gaussian Splatting [15] uses 3D Gaussians to model a 3D scene and benefits from fast speed and high quality in the field of neural rendering. The physical characteristics of 3D Gaussians and the splat-based rasterization also motivated rapid advancements in research fields such as scene editing [10, 24, 29, 36], dynamic scenarios [7, 27, 38, 48, 53], and SLAM [5, 20, 50, 57]. GaussianFormer [13] pioneers the application of 3D Gaussians in outdoor 3D occupancy prediction and uses features from multi-view images to update 3D Gaussians, which can be converted into 3D occupancy prediction through a Gaussian-to-voxel splatting module. However, it is still unclear how to employ 3D Gaussians for online global indoor scene understanding from local observations. We achieve this by designing a Gaussian memory mechanism and progressively updating it with structure-aware interaction." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 639, + 432, + 652 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 432, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 432, + 652 + ], + "type": "text", + "content": "3. Proposed Approach" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 659, + 507, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 507, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 507, + 672 + ], + "type": "text", + "content": "3.1. Embodied 3D Occupancy Prediction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Conventional methods in indoor scenarios for occupancy prediction accepted RGB-D as inputs to predict the semantic occupancy of a 3D scene which requires depth sensors." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26361" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 79, + 537, + 240 + ], + "blocks": [ + { + "bbox": [ + 57, + 79, + 537, + 240 + ], + "lines": [ + { + "bbox": [ + 57, + 79, + 537, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 79, + 537, + 240 + ], + "type": "image", + "image_path": "83329bd7533bd4ca726b7428f30052ba3744d5342ae60be45c107bae2757ea56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 298 + ], + "type": "text", + "content": "Figure 2. Framework of our EmbodiedOcc for embodied 3D occupancy prediction. We maintain an explicit global memory of 3D Gaussians during the exploration of the current scene. For each update, the Gaussians within the current frustum are taken from the memory and updated using semantic and structural features extracted from the monocular RGB input. Each Gaussian has a confidence value to integrate information from both the memory and the current input. Then we detach and put these updated Gaussians back into the memory. We can obtain the current 3D occupancy prediction using a Gaussian-to-voxel splatting module whenever we need." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 304, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 304, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 304, + 296, + 376 + ], + "type": "text", + "content": "However, we humans are capable of effortlessly processing the visual information from a single view to obtain 3D perception of our surroundings. Recent methods begin to consider endowing models with the same competence, which accept a monocular RGB image as input and derive a 3D occupancy prediction within the current frustum:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 380, + 295, + 393 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 380, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 121, + 380, + 295, + 393 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {\\text {m o n o}} = \\mathcal {F} _ {\\text {m o n o}} \\left(I _ {\\text {m o n o}}\\right), \\tag {1}", + "image_path": "77b5f8e68f5fb7cd14f1cae66c27076ca65973b5b2015c083e201e87b04fac29.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{mono}" + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": " is the proposed monocular prediction model, " + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "inline_equation", + "content": "I_{mono} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{mono} \\in \\mathbb{R}^{X \\times Y \\times Z \\times C}" + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": " refer to the monocular RGB input and the obtained 3D occupancy prediction. " + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "inline_equation", + "content": "X, Y, Z" + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": " represent the dimensions of the local 3D scene and " + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 403, + 296, + 462 + ], + "type": "text", + "content": " represents the total number of semantics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 463, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 558 + ], + "type": "text", + "content": "This is only the initial step towards practical scenarios. The essence of human intelligence is the capacity to analyze and respond immediately based on real-time perception of the surroundings. Correspondingly, superior embodied agents are anticipated to process egocentrically gathered real-time visual input to update the 3D occupancy prediction of the current scene. This capability facilitates the execution of downstream tasks based on real-time perception." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": "Motivated by this, we propose an embodied 3D occupancy prediction task in this paper. Let " + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_t = \\{x_1, x_2, \\dots, x_t\\}" + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": " be an RGB sequence and the corresponding extrinsics collected by the embodied agent up to the present, where " + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_t = (I_t, M_t)" + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_t \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_t \\in \\mathbb{R}^{3 \\times 4}" + }, + { + "bbox": [ + 55, + 559, + 296, + 714 + ], + "type": "text", + "content": ". It is worth noting that the variation in the subscripts merely represents the change in the position and perspective of the agent when exploring the current scene continuously. Different subscripts may correspond to similar positions and perspectives, indicating that the agent has returned to a previously explored location. In embodied occupancy prediction, re-exploration of the same area should maintain global consistency and even demonstrate improved performance, akin to we hu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 327 + ], + "type": "text", + "content": "mans always possessing a more comprehensive understanding of sights that have been encountered repeatedly." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 328, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 352 + ], + "type": "text", + "content": "We formulate the function of an embodied occupancy prediction model as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 376, + 356, + 553, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 356, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 376, + 356, + 553, + 369 + ], + "type": "interline_equation", + "content": "\\mathbf {Y} _ {t} = \\mathcal {F} _ {\\text {e m b o d i e d}} \\left(\\mathbf {Y} _ {t - 1}, x _ {t}\\right), \\tag {2}", + "image_path": "be10b00dd392997167654e6e5cce96bf737df720f202f451dfe5b8f91cbbd158.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\text{embodied}}" + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "content": " is the embodied prediction model, " + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_t \\in \\mathbb{R}^{X_{\\text{room}} \\times Y_{\\text{room}} \\times Z_{\\text{room}} \\times C}" + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "content": " refers to the current occupancy prediction of the whole scene (" + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_0" + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "content": " is the initialization). " + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "inline_equation", + "content": "X_{\\text{room}}, Y_{\\text{room}}, Z_{\\text{room}}" + }, + { + "bbox": [ + 313, + 378, + 555, + 426 + ], + "type": "text", + "content": " denote the scene dimensions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 433, + 459, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 459, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 459, + 445 + ], + "type": "text", + "content": "3.2. Local Refinement Module" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 533 + ], + "type": "text", + "content": "Different from conventional methods that conducted feature integration in a voxelized space, we use a set of 3D semantic Gaussians to represent an indoor scene [13]. In this subsection, we will first explain our local refinement module, which extracts semantic and structural features from the monocular input and integrates them to update the Gaussian-based representation of the current frustum." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": "Initialization. We first initialize a set of semantic Gaussians to represent the current frustum. Each semantic Gaussian " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " is represented by a vector comprising mean " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{m} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", scale " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{s} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", rotation quaternion " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{r} \\in \\mathbb{R}^4" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", opacity " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{o} \\in \\mathbb{R}" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", and semantic logits " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{c} \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " denotes the total number of semantic categories). We use an embedding layer to lift each Gaussian vector " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " to its corresponding high-dimensional feature vectors " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", and derive " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{\\mathbf{Q}_i \\in \\mathbb{R}^m, i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " is the dimension of " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_i" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 534, + 556, + 653 + ], + "type": "text", + "content": " is the total number of the Gaussians." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Depth-Aware Branch. Due to the variable scales and tight arrangements of indoor objects, depth ambiguity has always been one of the core challenges limiting the performance of indoor occupancy prediction models in monocular settings. Previous work has consistently focused on how to" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26362" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 73, + 275, + 202 + ], + "blocks": [ + { + "bbox": [ + 77, + 73, + 275, + 202 + ], + "lines": [ + { + "bbox": [ + 77, + 73, + 275, + 202 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 275, + 202 + ], + "type": "image", + "image_path": "a5f0e7fd302ed683f6f3396f6d471b6603310e1ecc57aa244de0512451669db9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 203, + 295, + 291 + ], + "lines": [ + { + "bbox": [ + 55, + 203, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 295, + 291 + ], + "type": "text", + "content": "Figure 3. Motivation of the depth-aware branch. Along a specific ray, Gaussians distributed in front of the true depth point are likely to model the empty semantic (A). Gaussians distributed behind the true depth point closely are likely to model valid semantics (B). Gaussians that are distributed behind the true depth point but are too far away require more information to guide their updates (C). During the embodied exploration, the subsequent frames can make up for this lack of information in the current frame." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 296, + 343 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 296, + 343 + ], + "type": "text", + "content": "better extract and utilize depth information from the input image. We design a depth-aware branch to provide more accurate and effective guidance for the update of 3D semantic Gaussians in our local refinement module." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": "We first use a depth prediction network to obtain a relatively accurate depth map " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "D_{\\text{metric}}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": " from input " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "I_{mono}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": ". A naive approach can explicitly utilize this depth information when initializing the Gaussians, e.g., we can randomly sample some points from the pseudo point cloud recovered from the depth map and use these coordinates to initialize the means of some Gaussians. Although providing direct hints for the means of some Gaussians, this cannot exploit the potential of the depth information. We design a simple yet effective depth-aware layer to accomplish this. We still uniformly initialize a number of Gaussians within the current frustum. For each Gaussian, we project its mean " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{m}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": " into the pixel coordinate system through the intrinsics " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "K_{mono} \\in \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": " and obtain the depth value " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": ". The sampled depth value " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": ", along with the z-component " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": " of the Gaussian mean in the camera coordinate system, are fed into the depth-aware layer, which is a multi-layer perceptron (MLP) that outputs the depth-aware feature " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{depth}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": " for this Gaussian. Then we add the depth-aware feature to the original feature vector " + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}" + }, + { + "bbox": [ + 56, + 344, + 296, + 619 + ], + "type": "text", + "content": ", injecting additional information into the subsequent feature integration. In this way, depth information not only affects the means of the Gaussians but also promotes the update of other properties:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 620, + 266, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 620, + 266, + 633 + ], + "spans": [ + { + "bbox": [ + 89, + 620, + 266, + 633 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {\\text {d e p t h}} = \\mathcal {M} _ {\\text {d e p t h a w a r e}} \\left(\\left(D _ {\\text {m e t r i c}} (u, v), z\\right), \\right.", + "image_path": "509fdf2faab8fb5abdaf49064a8936ff204f799402429af66f492a42f677d721.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 633, + 294, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 633, + 294, + 651 + ], + "spans": [ + { + "bbox": [ + 86, + 633, + 294, + 651 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {Q}} = \\left\\{\\hat {\\mathbf {Q}} _ {i}, i = 1, \\dots , N \\mid \\hat {\\mathbf {Q}} _ {i} = \\mathbf {Q} _ {i} + \\mathbf {Q} _ {i} ^ {\\text {d e p t h}} \\right\\}, \\tag {3}", + "image_path": "c398bf64a429055af285fc7c2ed193438cd04c8361bbd0008493d20b8c60dae6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{depthaware}" + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": " is the depth-aware layer, " + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": " are pixel coordinates of each Gaussian. We illustrate our depth-aware branch in Figure 3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Feature Integration and Gaussian Refinement. Feature integration in our local refinement module includes the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "interactions among Gaussians as well as the interactions between image features and Gaussians. We voxelize the Gaussian centers and conduct 3D sparse convolution on the generated grid to allow interactions among Gaussian vectors " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ". We project the Gaussian centers onto the image feature map and use the deformable attention function to integrate the queried features and the Gaussian vectors " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Q}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ". After the prior feature integration, these feature vectors with aggregated information will be used to obtain the update amounts " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{G} = (\\Delta \\mathbf{m}, \\Delta \\mathbf{s}, \\Delta \\mathbf{r}, \\Delta \\mathbf{o}, \\Delta \\mathbf{c})" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " of each Gaussian. We use the update amounts " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " to refine the Gaussian properties:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 207, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 207, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 207, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\text {n e w}} = (\\Delta \\mathbf {m} + \\mathbf {m}, \\Delta \\mathbf {s} + \\mathbf {s}, \\Delta \\mathbf {r} \\otimes \\mathbf {r}, \\Delta \\mathbf {o} + \\mathbf {o}, \\Delta \\mathbf {c} + \\mathbf {c}), \\tag {4}", + "image_path": "9636b732a0d630ea0106a17b1aa2d6d50260fde5e19f0172bd0ec157f69f5058.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 223, + 545, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 545, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 545, + 234 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 223, + 545, + 234 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 223, + 545, + 234 + ], + "type": "text", + "content": " refers to the special composition of quaternions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 555, + 282 + ], + "type": "text", + "content": "We conduct the feature integration and the refinement of Gaussians multiple times. After the final refinement, we use a Gaussian-to-voxel splatting module [13] to obtain the final occupancy within the frustum." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 501, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 501, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 501, + 302 + ], + "type": "text", + "content": "3.3. Gaussian Memory Updated Online" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 307, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 554, + 390 + ], + "type": "text", + "content": "To explore unknown scenes, we humans continuously update the objects within the scene and their relationships to gradually construct a global scene memory. When revisiting this scene for further exploration, we use the visual information to refine this memory. Inspired by this, we design an online framework (shown in Figure 2) and maintain a Gaussian memory for global understanding." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": "Memory Initialization. Our local refinement module initializes and updates Gaussians in the camera coordinate system, as the extrinsics in indoor scenarios are constantly changing, which will pose additional difficulties for our local module. But in the final embodied framework, we initialize the entire scene with uniform Gaussians in the world coordinate system. For a novel scene to be explored, we have: " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{\\text{room}} = \\{(G_i, \\gamma_i), i = 1, \\dots, N | G_i = (\\mathbf{m}_i, \\mathbf{s}_i, \\mathbf{r}_i, \\mathbf{o}_i, \\mathbf{c}_i), \\gamma_i = 0, 1\\}" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " refers to the number of Gaussians to initialize this scene, " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{m}_i" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " are the means and rotation quaternions of these Gaussians in the world coordinate system (" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{o}_i" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\mathbf{c}_i" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " maintain consistency between the world and camera coordinate systems). We introduce an additional tag " + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 391, + 554, + 594 + ], + "type": "text", + "content": " for all the Gaussians in the memory. When initializing a novel scene, tags of these Gaussians are set to 0. Every time we put some updated Gaussians back into the memory, their tags are set to 1." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": "Memory Update. At the current step " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": ", our embodied occupancy prediction framework receives a posed visual input " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "x_{t} = (I_{t}, M_{t})" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": " to perform the update. During the current update, we use a mask from coordinate system transformation to get all Gaussians " + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{t}" + }, + { + "bbox": [ + 313, + 594, + 554, + 689 + ], + "type": "text", + "content": " within the current frustum from the memory. These Gaussians will interact and be refined using a tailored confidence refinement module. Then we detach these Gaussians and put them back into the memory." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Confidence Refinement. Apart from the initial update for each scene which is akin to the local refinement, sub" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26363" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 205 + ], + "type": "image", + "image_path": "efc113c114e99c8a76bf3a224fc9b8a473dd484b7b70563fcff1d89c657f662f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 208, + 295, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 208, + 295, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 295, + 264 + ], + "type": "text", + "content": "Figure 4. Illustration of our Gaussian memory. During each update, the Gaussians within the current frustum are taken from the memory. Confidence values of those well-updated Gaussians are used to integrate information from both the memory and the current input. Then we put these Gaussians back into the memory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 272, + 295, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 272, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 272, + 295, + 415 + ], + "type": "text", + "content": "sequent exploration involves the update of Gaussians from the Gaussian memory, among which some have been well-updated by previous frames (if we can derive an acceptable local occupancy prediction from these Gaussians, we believe that they have been \"well-updated\") and some still remain random. It is unreasonable to update these Gaussians equally. For those Gaussians deemed well-updated, we only need to refine them slightly based on the semantic and structural features extracted from the current image, which is exactly the essence of maintaining the Gaussian memory. As for those random Gaussians that have never been updated, we can directly update them with a fresh perspective." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": "To elaborate, we generate a confidence value " + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": " for each Gaussian taken from the memory. For those having been previously updated " + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "inline_equation", + "content": "(\\gamma = 1)" + }, + { + "bbox": [ + 55, + 416, + 296, + 513 + ], + "type": "text", + "content": ", we set their confidence values to a certain value between 0 and 1, which means we will integrate information from both the memory and the newly observed image to update these Gaussians. For those that have never been updated, we set their confidence values to 0 and follow the same refinement module in Sec. 3.2:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 137, + 516, + 295, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 516, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 137, + 516, + 295, + 534 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {G} _ {\\text {o n l i n e}} = (1 - \\theta) \\Delta \\mathbf {G}, \\tag {5}", + "image_path": "1a95773ee5ba3bf470e4f838e76be2c52ce177cca05f4d1ee459658defbd976d.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 109, + 531, + 242, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 531, + 242, + 544 + ], + "spans": [ + { + "bbox": [ + 109, + 531, + 242, + 544 + ], + "type": "interline_equation", + "content": "\\mathbf {G} _ {\\text {a f t e r}} = \\Delta \\mathbf {G} _ {\\text {o n l i n e}} \\oplus \\mathbf {G} _ {\\text {b e f o r e}},", + "image_path": "f4d18b51856cd4bae57b06ac896b16b64bcd8e6523a26946e14dd4f32505ec64.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "text", + "content": "where we use " + }, + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 55, + 552, + 295, + 588 + ], + "type": "text", + "content": " to represent the composition of rotation quaternions and the add operation of other parts. Figure 4 illustrates how we maintain the Gaussian memory." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "text", + "content": "Stopping Mechanism. We propose a simple stopping mechanism to consider a room as having been effectively explored. At the step " + }, + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "text", + "content": ", we first calculate a confidence ratio " + }, + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 589, + 295, + 637 + ], + "type": "text", + "content": " to measure the exploration of the current room:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 133, + 640, + 295, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 640, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 133, + 640, + 295, + 657 + ], + "type": "interline_equation", + "content": "\\alpha = \\sum_ {i = 0} ^ {N} \\mathbb {I} _ {\\gamma_ {i} = 1} / N, \\tag {6}", + "image_path": "e5b8c983e32383caa5e8c2e3dc71fd7486dd0ef8055e762ea330330ec66eb9c1.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{I}_{\\gamma_i = 1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " takes the value of 1 if " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\gamma_{i} = 1" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": ". If " + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": " exceeds a certain threshold we set before, the model can decide to enter an adjacent room to begin a new exploration or stay here to get a better perception of the current room." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 534, + 84 + ], + "type": "text", + "content": "3.4. EmbodiedOcc: An Embodied Framework" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 554, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 554, + 148 + ], + "type": "text", + "content": "We present the training framework of our EmbodiedOcc model for indoor embodied occupancy prediction. During the whole prediction process, we use the current monocular input to update our Gaussian memory in real time, which can be easily converted into 3D occupancy prediction." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": "We first train our local refinement module using the focal loss " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "L_{focal}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": ", the lovasz-softmax loss " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "L_{lov}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": ", the scene-class affinity loss " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "L_{scal}^{geo}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "L_{scal}^{sem}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": " following RetinaNet [23], TPV-Former [11] and MonoScene [2]. We use monocular occupancy within the frustum " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{mono}^{fov}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": " and the corresponding ground truth " + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_{gt}^{fov}" + }, + { + "bbox": [ + 313, + 149, + 554, + 222 + ], + "type": "text", + "content": " to compute the loss:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 320, + 224, + 553, + 257 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 224, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 320, + 224, + 553, + 257 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} = \\lambda_ {1} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) + \\mathcal {L} _ {\\text {l o v}} \\left(\\mathbf {Y} _ {\\text {m o n o}} ^ {\\text {f o v}}, \\mathbf {Y} _ {\\text {g t}} ^ {\\text {f o v}}\\right) \\tag {7} \\\\ + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {g e o}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}) + \\mathcal {L} _ {\\text {s c a l}} ^ {\\text {s e m}} (\\mathbf {Y} _ {\\text {m o n o}} ^ {f o v}, \\mathbf {Y} _ {g t} ^ {f o v}), \\\\ \\end{array}", + "image_path": "98c5f3a7269101275dec00f4574388b0eca9c7ebab7956e1ae37bc10a10ea4b7.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 258, + 430, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 430, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 430, + 270 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 258, + 430, + 270 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 313, + 258, + 430, + 270 + ], + "type": "text", + "content": " is a balance factor." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 450 + ], + "type": "text", + "content": "We then use the trained local module to train our EmbodiedOcc. For efficient training, we initialize the Gaussian memory before the first update and compute the current loss following the equation 7 after each update. To ensure consistency, the local occupancy ground truth used here is obtained from the occupancy of the whole scene. After a certain number of updates, we re-initialize the memory and come to the next scene. Trained with such a pipeline, our EmbodiedOcc can effectively perform the embodied occupancy prediction task while ensuring consistency within the same scene. We expect that our EmbodiedOcc can have an improving prediction with continuous exploration rather than undermining previous predictions when encountering parts that have been explored before. Therefore, we conduct some tailored tests to validate the capability of our model." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 395, + 474 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 504, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 504, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 504, + 492 + ], + "type": "text", + "content": "4.1. EmbodiedOcc-ScanNet Benchmark" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": "Task Descriptions. We conducted two tasks to evaluate our EmbodiedOcc framework: local occupancy prediction and embodied occupancy prediction. Local occupancy prediction shares the same setting with previous works, which accept monocular image as input and obtain the occupancy within the current frustum. Embodied occupancy prediction accepts real-time visual inputs continuously and updates the occupancy of the current scene online. The visual input at a certain step " + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 498, + 554, + 629 + ], + "type": "text", + "content": " during embodied occupancy prediction is still monocular, which is a more challenging setting compared with multi-view input or input with 3D information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Datasets. We trained and evaluated our local refinement module on the Occ-ScanNet dataset [56], which provides frames in " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "60 \\times 60 \\times 36" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " voxel grids (a " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "4.8m \\times 4.8m \\times 2.88m" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " box in front of the camera). These frames are labeled with 12 semantics, including 11 for valid semantics (ceiling, floor, wall, window, chair, bed, sofa, table, tvs, furniture, objects) and 1 for empty space." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26364" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 66, + 82, + 545, + 207 + ], + "blocks": [ + { + "bbox": [ + 176, + 71, + 434, + 80 + ], + "lines": [ + { + "bbox": [ + 176, + 71, + 434, + 80 + ], + "spans": [ + { + "bbox": [ + 176, + 71, + 434, + 80 + ], + "type": "text", + "content": "Table 1. Local Prediction Performance on the Occ-ScanNet dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 82, + 545, + 207 + ], + "lines": [ + { + "bbox": [ + 66, + 82, + 545, + 207 + ], + "spans": [ + { + "bbox": [ + 66, + 82, + 545, + 207 + ], + "type": "table", + "html": "
MethodInputIoUceilingfloorwallwindowchairbedsofatablevtsfurnitureobjectsmIoU
TPVFormer [11]xrgb33.396.9632.9714.419.1024.0141.4945.4428.6110.6635.3725.3124.94
GaussianFormer [13]xrgb40.9120.7042.0023.4017.4027.044.3044.8032.7015.3036.7025.0029.93
MonoScene [2]xrgb41.6015.1744.7122.4112.5526.1127.0335.9128.326.5732.1619.8424.62
ISO [56]xrgb42.1619.8841.8822.3716.9829.0942.4342.0029.6010.6236.3624.6128.71
Surroundocc [46]xrgb42.5218.9049.3024.8018.0026.8042.0044.1032.9018.6036.8026.9030.83
Oursxrgb53.5539.6050.4041.4031.7040.9055.0061.4044.0036.1053.9042.2045.15
", + "image_path": "c84ce06e560a9e88e6e2520fbdc3ac7031923561645a6aed81287d7503995779.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 57, + 219, + 555, + 332 + ], + "blocks": [ + { + "bbox": [ + 147, + 209, + 463, + 217 + ], + "lines": [ + { + "bbox": [ + 147, + 209, + 463, + 217 + ], + "spans": [ + { + "bbox": [ + 147, + 209, + 463, + 217 + ], + "type": "text", + "content": "Table 2. Embodied Prediction Performance on the EmbodiedOcc-ScanNet dataset." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 219, + 555, + 332 + ], + "lines": [ + { + "bbox": [ + 57, + 219, + 555, + 332 + ], + "spans": [ + { + "bbox": [ + 57, + 219, + 555, + 332 + ], + "type": "table", + "html": "
MethodDatasetIoUceilingfloorwallwindowchairbedsofatabletvsfurnitureobjectsmIoU
TPVFormer [11]EmbodiedOcc35.881.6230.5412.0313.2235.4751.3949.7925.633.6043.1516.2325.70
SurroundOcc [46]EmbodiedOcc37.0412.7031.8022.5022.0029.9044.7036.5024.6011.5034.4018.2026.27
GaussianFormer [13]EmbodiedOcc38.0217.0033.6021.5021.7029.4047.8037.1024.3015.5036.2016.8027.36
SplicingOccEmbodiedOcc49.0131.6038.8035.5036.3047.1054.5057.2034.4032.5051.2029.1040.74
EmbodiedOccEmbodiedOcc51.5222.7044.6037.4038.0050.1056.7059.7035.4038.4052.0032.9042.53
", + "image_path": "405379bf8e3976726467a2cc92e895c36524ab0194e293c513d1daf1dd576885.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "text", + "content": "Based on this dataset, we reorganized an EmbodiedOccScanNet dataset to train and evaluate our EmbodiedOcc framework [35, 56]. Our EmbodiedOcc-ScanNet comprises 537/137 scenes in the train/val splits. Each scene consists of 30 posed frames with their corresponding local occupancies. The resolutions of global occupancy of each scene are calculated by " + }, + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "inline_equation", + "content": "l_{x} \\times l_{y} \\times l_{z} / (0.08m)^{3}" + }, + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "inline_equation", + "content": "l_{x} \\times l_{y} \\times l_{z}" + }, + { + "bbox": [ + 55, + 337, + 294, + 491 + ], + "type": "text", + "content": " is the range of this scene in the world coordinate system. In addition, we associate grid points that can be projected onto the camera plane for each frame as the global mask of this frame. By splicing the global mask of all processed frames, we can easily obtain the occupancy ground truth of the explored part in the current scene." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 625 + ], + "type": "text", + "content": "Apart from Occ-ScanNet and EmbodiedOcc-ScanNet datasets in the original scale, we sampled a small set from the EmbodiedOcc-ScanNet dataset as the EmbodiedOcc-ScanNet-mini dataset which comprises 64/16 scenes in the train/val splits. We sampled from the Occ-ScanNet dataset accordingly and obtained an Occ-ScanNet-mini2 dataset, which comprises 5504/2376 frames in the train/val splits. We conducted the local occupancy prediction task on the Occ-ScanNet and Occ-ScanNet-mini2 datasets and conducted the embodied prediction task on the EmbodiedOcc-ScanNet and EmbodiedOcc-ScanNet-mini datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 713 + ], + "type": "text", + "content": "Evaluation Metrics. We use mIoU and IoU as the evaluation metrics. For local occupancy prediction, we calculate the mIoU and IoU using the occupancy within the current frustum (same with the evaluation in ISO [56]). For embodied occupancy prediction, we calculate the mIoU and IoU using the global occupancy of the current scene. It is worth mentioning that the global occupancy used here is the union" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 337, + 553, + 361 + ], + "type": "text", + "content": "of the frustums corresponding to 30 frames of each scene, which represents the region that has been explored." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 371, + 447, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 447, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 447, + 384 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 389, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 568 + ], + "type": "text", + "content": "Local Refinement Module. Following existing works [13, 56], we use a pre-trained EfficientNet-B7 [39] to initialize the image encoder in our local module. The depth prediction network used in the depth-aware branch is a fine-tuned DepthAnything-V2 model [51] that remains frozen during the training, and the depth-aware layer is a 3-layer MLP. The resolutions of the monocular input are set to " + }, + { + "bbox": [ + 313, + 389, + 555, + 568 + ], + "type": "inline_equation", + "content": "480 \\times 640" + }, + { + "bbox": [ + 313, + 389, + 555, + 568 + ], + "type": "text", + "content": " and the number of Gaussians used to conduct the local prediction is 16200. We utilize the AdamW [26] optimizer with a weight decay of 0.01. The learning rate warms up in the first 1000 iterations to a maximum value of 2e-4 and decreases according to a cosine schedule [25]. We train our local refinement module for 10 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the Occ-ScanNet dataset and 20 epochs on the Occ-ScanNet-mini2 dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "EmbodiedOcc Framework. We initialize the Gaussians with a " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "0.16\\mathrm{m}" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " interval to represent a novel scene. For each update, the confidence value " + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": " of well-updated Gaussians is set to 0 in the first two refinement layers (frozen) and 0.5 in the final refinement layer. We train our EmbodiedOcc for 5 epochs using 8 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet dataset and 20 epochs using 4 NVIDIA GeForce RTX 4090 GPUs on the EmbodiedOcc-ScanNet-mini dataset. The maximum value of the learning rate is set to 2e-4 using 8 GPUs and 1e-4 using 4 GPUs. The other settings remain the same with the training of the local refinement module." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26365" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 116, + 294, + 204 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": "Table 3. Look-Back Prediction vs First-Time Prediction. For " + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "inline_equation", + "content": "\\mathrm{K} = k" + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": ", we simply select 0, 1, ..., k - 1th frames to evaluate our EmbodiedOcc framework and the occupancy ground truth used here is the union of the " + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 70, + 297, + 114 + ], + "type": "text", + "content": " frustums. K was set to 3/5/8." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 116, + 294, + 204 + ], + "lines": [ + { + "bbox": [ + 56, + 116, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 294, + 204 + ], + "type": "table", + "html": "
ModeKFrame ListIoUmIoU
First-Time3[0,1,2]49.3939.32
Look-Back3[0,1,2,1,0]49.5240.09
First-Time5[0,...,4]50.1340.03
Look-Back5[0,...,3,4,3,...,0]50.6440.98
First-Time8[0,...,7]50.9440.86
Look-Back8[0,...,6,7,6,...,0]51.1441.17
", + "image_path": "4fa24c05c76959a20c07f5c10386478c5b78ed836050924cb55cebe8e07335db.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 205, + 172, + 293 + ], + "blocks": [ + { + "bbox": [ + 56, + 205, + 172, + 293 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 172, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 172, + 293 + ], + "type": "image", + "image_path": "3c6074de3b7fe941098c5799c9e632025327f8bb247d5930e6b0525a35464491.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 172, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 172, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 172, + 319 + ], + "type": "text", + "content": "Figure 5. Performance with different stopping ratios." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 177, + 205, + 293, + 293 + ], + "blocks": [ + { + "bbox": [ + 177, + 205, + 293, + 293 + ], + "lines": [ + { + "bbox": [ + 177, + 205, + 293, + 293 + ], + "spans": [ + { + "bbox": [ + 177, + 205, + 293, + 293 + ], + "type": "image", + "image_path": "439b848185db8c512927307636dc02c494ba2ea477da018d637dcf35c96f68d6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 296, + 293, + 319 + ], + "lines": [ + { + "bbox": [ + 176, + 296, + 293, + 319 + ], + "spans": [ + { + "bbox": [ + 176, + 296, + 293, + 319 + ], + "type": "text", + "content": "Figure 6. Ablation study of the confidence refinement." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 141, + 335 + ], + "type": "text", + "content": "4.3. Main Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 296, + 460 + ], + "type": "text", + "content": "Local Occupancy Prediction. We evaluated our local refinement module on the Occ-ScanNet dataset [56]. As shown in Table 1, the results indicate that our local refinement module outperforms ISO [56]. We also implemented several state-of-the-art driving scene methods [11, 13, 46] on this benchmark and our local refinement module outperforms them by a large margin. This is because they mainly focus on the coarse layout (e.g., positions of objects) while indoor scenes require modeling of the fine-grained structure (e.g., shapes of objects)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 628 + ], + "type": "text", + "content": "Embodied Occupancy Prediction. We assessed the occupancy prediction for the entire scene after processing 30 frames, and the ground truth for calculating IoU and mIoU is the union of the frustums. We spliced the local occupancy obtained from our local module to serve as the main baseline (referred to as SplicingOcc), as our local module has achieved the best local performance to date. It can be observed in Table 2 that our EmbodiedOcc exhibits superior prediction of the scene, which is achieved through the integration of different views. We also compared our EmbodiedOcc with the driving scene methods mentioned before (we obtained their embodied results by voting from different local predictions). Their poor results are due to ignoring the continuity of the observations without a global memory." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 185, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 185, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 185, + 649 + ], + "type": "text", + "content": "4.4. Experimental Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Effect of Continuous Online Updating. We expect EmbodiedOcc to have better performance when encountering parts that have been explored before, and thus, we designed a Look-Back evaluation on the EmbodiedOccScanNet dataset. Specifically, after processing K frames," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 318, + 83, + 553, + 127 + ], + "blocks": [ + { + "bbox": [ + 361, + 71, + 507, + 82 + ], + "lines": [ + { + "bbox": [ + 361, + 71, + 507, + 82 + ], + "spans": [ + { + "bbox": [ + 361, + 71, + 507, + 82 + ], + "type": "text", + "content": "Table 4. Analysis of the model design." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 83, + 553, + 127 + ], + "lines": [ + { + "bbox": [ + 318, + 83, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 318, + 83, + 553, + 127 + ], + "type": "table", + "html": "
MethodGaussianStructureMemoryLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
EmbodiedOcc-Voxel×✓✓47.5038.1237.5326.99
EmbodiedOcc w/o memory✓✓×53.5545.1549.0140.74
EmbodiedOcc✓✓✓53.5545.1551.5242.53
", + "image_path": "c4743897a0d11bff6abb3f6d42c4e7eb250f8eac67cf6dbcb0a9cd9e9347c45b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 318, + 140, + 553, + 198 + ], + "blocks": [ + { + "bbox": [ + 348, + 129, + 520, + 138 + ], + "lines": [ + { + "bbox": [ + 348, + 129, + 520, + 138 + ], + "spans": [ + { + "bbox": [ + 348, + 129, + 520, + 138 + ], + "type": "text", + "content": "Table 5. Analysis of the depth-aware branch." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 140, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 140, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 140, + 553, + 198 + ], + "type": "table", + "html": "
Branch TypeDepth Estimation ModuleLocal PredictionEmbodied Prediction
IoUmIoUIoUmIoU
Depth-aware branchDepthAnything-V253.9346.2050.7841.45
No-depth branch/48.1540.0737.5230.73
Naive-depth branchDepthAnything-V250.3242.73//
Depth-aware branchIndoorDepth51.2443.8746.4237.78
", + "image_path": "82893a84e7069fb57bf8228feed281baab5cd032e3abf782b255defbcb2d63a1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 318, + 211, + 553, + 256 + ], + "blocks": [ + { + "bbox": [ + 346, + 199, + 522, + 210 + ], + "lines": [ + { + "bbox": [ + 346, + 199, + 522, + 210 + ], + "spans": [ + { + "bbox": [ + 346, + 199, + 522, + 210 + ], + "type": "text", + "content": "Table 6. Analysis of the Gaussian parameters." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 211, + 553, + 256 + ], + "lines": [ + { + "bbox": [ + 318, + 211, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 318, + 211, + 553, + 256 + ], + "type": "table", + "html": "
Gaussian Number (In local box)Gaussian Scale Min(m)Max(m)Gaussian Interval(m) (In global scene)Local Prediction IoUmIoUEmbodied Prediction IoUmIoU
162000.010.08(0.16, 0.16, 0.16)53.9346.2050.7841.45
81000.010.08(0.20, 0.20, 0.20)50.4742.8246.2437.99
162000.010.20(0.16, 0.16, 0.16)51.5743.7448.0938.40
", + "image_path": "b7ef74f83c4c624fe26a217c0af4e36091ee23924ea4c6c55ae32f7d626f8d77.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "type": "table", + "bbox": [ + 315, + 269, + 553, + 326 + ], + "blocks": [ + { + "bbox": [ + 371, + 257, + 498, + 267 + ], + "lines": [ + { + "bbox": [ + 371, + 257, + 498, + 267 + ], + "spans": [ + { + "bbox": [ + 371, + 257, + 498, + 267 + ], + "type": "text", + "content": "Table 7. Runtime decomposition." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 269, + 553, + 326 + ], + "lines": [ + { + "bbox": [ + 315, + 269, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 315, + 269, + 553, + 326 + ], + "type": "table", + "html": "
Scene level (ms)Scene init.6.626Occ head39.635
Frame level (ms)Load memory0.973Depth aware1.816
Img backbone61.478GS Encoder14.761
Depthanything34.687Update memory0.474
", + "image_path": "d0318b1588d0a95109726f6e192e01cd9c13020527a331fe537aaac7c71ab4a2.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "content": "we direct the model to reprocess the last " + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "inline_equation", + "content": "\\mathrm{K} - 1" + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "content": " frames. By comparing this Look-Back result with the First-Time prediction, we verified that our EmbodiedOcc has met our expectations as shown in Table 3." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 377, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 449 + ], + "type": "text", + "content": "Effect of the Stopping Mechanism. We use Figure 5 to show the effectiveness of our stopping mechanism. The ground truth used here for calculating IoU and mIoU is the union occupancy of the 30 frustums in a global scene. We observed that using a larger threshold results in more observations and better performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 449, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 569 + ], + "type": "text", + "content": "Analysis of the Confidence Refinement. During each update, local Gaussians are refined through three refinement layers. For Gaussians that have been updated before, we froze the first two layers and updated them in the last refinement layer when training our EmbodiedOcc. Figure 6 on the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets shows the impact of different confidence values (determines the coefficient of each " + }, + { + "bbox": [ + 313, + 449, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\Delta G" + }, + { + "bbox": [ + 313, + 449, + 554, + 569 + ], + "type": "text", + "content": "). We observe that moderate updates to those previously processed Gaussians yield the best embodied occupancy prediction." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "Analysis of the Model Design. The essence of our EmbodiedOcc is the explicit Gaussian memory. We adopt object-centric Gaussians instead of grid-based voxels since Gaussians are more flexible for local-global interaction. We implemented a voxel version of our EmbodiedOcc and evaluated it on our benchmark. As shown in Table 4, the satisfactory local yet poor embodied performance of EmbodiedOcc in the voxel version verified our conclusion. Results in Table 4 were evaluated on the Occ-ScanNet and EmbodiedOcc-ScanNet datasets." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Analysis of the Depth-Aware Branch. We analyze the effect of our depth-aware branch in Table 5 using the" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26366" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 553, + 261 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 553, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 553, + 261 + ], + "type": "image", + "image_path": "115be923ff84fe4ad5d2263d128bc65fde7a8a543816755979996082009d3095.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 285 + ], + "type": "text", + "content": "Figure 7. Visualization of the embodied occupancy prediction. We visualize the update of Gaussian memory and corresponding global occupancy. As the Gaussians transition from random to ordered, the occupancy of the current scene becomes more accurate and complete." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 62, + 287, + 289, + 462 + ], + "blocks": [ + { + "bbox": [ + 62, + 287, + 289, + 462 + ], + "lines": [ + { + "bbox": [ + 62, + 287, + 289, + 462 + ], + "spans": [ + { + "bbox": [ + 62, + 287, + 289, + 462 + ], + "type": "image", + "image_path": "c70f89a4b4041829c48b0571c878debc79ee313bd4540506f3d2174e31b8167c.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 464, + 277, + 474 + ], + "lines": [ + { + "bbox": [ + 73, + 464, + 277, + 474 + ], + "spans": [ + { + "bbox": [ + 73, + 464, + 277, + 474 + ], + "type": "text", + "content": "Figure 8. Visualization of local occupancy prediction." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 474, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 677 + ], + "type": "text", + "content": "Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini datasets. We find that depth information will significantly benefit the local and embodied occupancy prediction. As shown in the second row, without the assistance of depth information, the performance of embodied occupancy prediction drops sharply. This indicates that the update of Gaussians within the current frustum may corrupt previous predictions without the guidance of depth information. The results in the third row suggest that the depth-aware branch we employ is more reasonable compared to the naive method of directly initializing a portion of Gaussians with the pseudo point cloud recovered from the predicted depth map, the latter also poses difficulties for the initialization of global Gaussians so we do not provide the embodied results. Besides, we replaced DepthAnything-V2 with IndoorDepth [6] in the last row to prove that our depth-aware branch does not rely on a specific depth prediction network." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 712 + ], + "type": "text", + "content": "Analysis of the Gaussian Parameters. We analyze the effect of different Gaussian parameters in Table 6 using the Occ-ScanNet-mini2 and the EmbodiedOcc-ScanNet-mini" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 289, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 384 + ], + "type": "text", + "content": "datasets. We see that decreasing the number or increasing the scale of the Gaussians can lead to a decrease in performance during local and embodied occupancy prediction. This is closely related to the physical properties of Gaussians. Gaussians initialized too sparse may lead to holes in occupancy prediction, while Gaussians with too large scale will overlap and influence each other which is also detrimental to the correct prediction of occupancy." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 456 + ], + "type": "text", + "content": "Runtime Analysis. We present in Table 7 a runtime analysis on scene 0687-00 from the EmbodiedOcc-ScanNet dataset. The runtime decomposition details show that our method is efficient while the main bottleneck is the image and depth backbones, suggesting that the overall runtime of our EmbodiedOcc can be further reduced." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 540 + ], + "type": "text", + "content": "Visualizations. Figure 7 and 8 visualize the global and local predictions, respectively. Our model demonstrates reasonable local perception ability and further achieves good online prediction with the Gaussian memory. Due to space limitations, we will use a more diverse set of samples to further show the visual effect of our EmbodiedOcc in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we have presented an embodied 3D occupancy prediction task and proposed a Gaussian-based EmbodiedOcc framework accordingly. Our EmbodiedOcc maintains an explicit Gaussian memory of the current scene and updates this memory during the exploration of this scene. Both quantitative and visualization results have shown that our EmbodiedOcc outperforms existing methods in terms of local occupancy prediction and accomplishes the embodied occupancy prediction task with high accuracy and strong expandability. We believe that our EmbodiedOcc paves the way for enabling active agents to conduct accurate and flexible embodied occupancy prediction." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26367" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We would like to thank Tianyu Hu for her valuable assistance with the experiments. This work was supported in part by the National Natural Science Foundation of China under Grant 62125603, Grant 62321005, and Grant 62336004, and in part by the Beijing Natural Science Foundation under Grant No. L247009." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 115, + 185 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 193, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 61, + 193, + 296, + 236 + ], + "type": "text", + "content": "[1] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In CVPR, pages 324-333, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 237, + 296, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 237, + 296, + 269 + ], + "spans": [ + { + "bbox": [ + 61, + 237, + 296, + 269 + ], + "type": "text", + "content": "[2] Anh-Quan Cao and Raoul de Charette. Monoscene: Monocular 3d semantic scene completion. In CVPR, pages 3991–4001, 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 295, + 314 + ], + "type": "text", + "content": "[3] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In CVPR, pages 5828-5839, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 295, + 358 + ], + "type": "text", + "content": "[4] Angela Dai, Daniel Ritchie, Martin Bokeloh, Scott Reed, Jürgen Sturm, and Matthias Nießner. Scancomplete: Large-scale scene completion and semantic segmentation for 3d scans. In CVPR, pages 4578-4587, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 359, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 359, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 359, + 295, + 403 + ], + "type": "text", + "content": "[5] Tianchen Deng, Yaohui Chen, Leyan Zhang, Jianfei Yang, Shenghai Yuan, Jiuming Liu, Danwei Wang, Hesheng Wang, and Weidong Chen. Compact 3d gaussian splatting for dense visual slam. arXiv preprint arXiv:2403.11247, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 295, + 435 + ], + "type": "text", + "content": "[6] Chao Fan, Zhenyu Yin, Yue Li, and Feiqing Zhang. Deeper into self-supervised monocular indoor depth estimation. arXiv preprint arXiv:2312.01283, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 480 + ], + "type": "text", + "content": "[7] Quankai Gao, Qiangeng Xu, Zhe Cao, Ben Mildenhall, Wenchao Ma, Le Chen, Danhang Tang, and Ulrich Neumann. Gaussianflow: Splitting gaussian dynamics for 4d content creation. arXiv preprint arXiv:2403.12365, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 481, + 295, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 295, + 514 + ], + "type": "text", + "content": "[8] Martin Garbade, Yueh-Tung Chen, Johann Sawatzky, and Juergen Gall. Two stream 3d semantic scene completion. In CVPRW, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 557 + ], + "type": "text", + "content": "[9] Benjamin Graham, Martin Engelcke, and Laurens Van Der Maaten. 3d semantic segmentation with submanifold sparse convolutional networks. In CVPR, pages 9224-9232, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 558, + 295, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 295, + 591 + ], + "type": "text", + "content": "[10] Antoine Guédon and Vincent Lepetit. Gaussian frosting: Editable complex radiance fields with real-time rendering. arXiv preprint arXiv:2403.14554, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 295, + 635 + ], + "type": "text", + "content": "[11] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Tri-perspective view for vision-based 3d semantic occupancy prediction. In CVPR, pages 9223-9232, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 669 + ], + "type": "text", + "content": "[12] Yuanhui Huang, Wenzhao Zheng, Borui Zhang, Jie Zhou, and Jiwen Lu. Selfocc: Self-supervised vision-based 3d occupancy prediction. In CVPR, pages 19946-19956, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Yuanhui Huang, Wenzhao Zheng, Yunpeng Zhang, Jie Zhou, and Jiwen Lu. Gaussianformer: Scene as gaussians for vision-based 3d semantic occupancy prediction. In ECCV, pages 376-393, 2025. 2, 3, 4, 6, 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[14] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In ICPR, pages 4065-4071. IEEE, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 139, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 139, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 139, + 553, + 172 + ], + "type": "text", + "content": "[15] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. TOG, 42(4), 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[16] Maksim Kolodiazhnyi, Anna Vorontsova, Matvey Skripkin, Danila Rukhovich, and Anton Konushin. Unidet3d: Multi-dataset indoor 3d object detection. In AAAI, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[17] Xiaohan Lei, Min Wang, Wengang Zhou, Li Li, and Houqiang Li. Instance-aware exploration-verification-exploitation for instance imagegoal navigation. In CVPR, pages 16329–16339, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 294 + ], + "type": "text", + "content": "[18] Jie Li, Yu Liu, Dong Gong, Qinfeng Shi, Xia Yuan, Chunxia Zhao, and Ian Reid. Rgbd based dimensional decomposition residual network for 3d semantic scene completion. In CVPR, pages 7693-7702, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 336 + ], + "type": "text", + "content": "[19] Jie Li, Yu Liu, Xia Yuan, Chunxia Zhao, Roland Siegwart, Ian Reid, and Cesar Cadena. Depth based semantic scene completion with position importance aware loss. RAL, 5(1): 219-226, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[20] Mingrui Li, Shuhong Liu, Heng Zhou, Guohao Zhu, Na Cheng, Tianchen Deng, and Hongyu Wang. Sgs-slam: Semantic gaussian splatting for neural dense slam. In ECCV, pages 163-179, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 426 + ], + "type": "text", + "content": "[21] Yiming Li, Zhiding Yu, Christopher Choy, Chaowei Xiao, Jose M Alvarez, Sanja Fidler, Chen Feng, and Anima Anandkumar. Voxformer: Sparse voxel transformer for camera-based 3d semantic scene completion. In CVPR, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 470 + ], + "type": "text", + "content": "[22] Zhiqi Li, Zhiding Yu, David Austin, Mingsheng Fang, Shiyi Lan, Jan Kautz, and Jose M Alvarez. Fb-occ: 3d occupancy prediction based on forward-backward view transformation. arXiv preprint arXiv:2307.01492, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 503 + ], + "type": "text", + "content": "[23] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dólar. Focal loss for dense object detection. In ICCV, pages 2980-2988, 2017. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[24] Zhiheng Liu, Hao Ouyang, Qiuyu Wang, Ka Leong Cheng, Jie Xiao, Kai Zhu, Nan Xue, Yu Liu, Yujun Shen, and Yang Cao. Infusion: Inpainting 3d gaussians via learning depth completion from diffusion prior. arXiv preprint arXiv:2404.11613, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 558, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 558, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 558, + 553, + 591 + ], + "type": "text", + "content": "[25] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. arXiv preprint arXiv:1608.03983, 2016. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 613 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[27] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, pages 8900-8910, 2024. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[28] Ruiyuan Lyu, Tai Wang, Jingli Lin, Shuai Yang, Xiaohan Mao, Yilun Chen, Runsen Xu, Haifeng Huang, Chenming Zhu, Dahua Lin, et al. Mmscan: A multi-modal 3d scene dataset with hierarchical grounded language annotations. arXiv preprint arXiv:2406.09401, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26368" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[29] Francesco Palandra, Andrea Sanchietti, Daniele Baieri, and Emanuele Rodola. Gsedit: Efficient text-guided editing of 3d objects via gaussian splatting. arXiv preprint arXiv:2403.05154, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[30] Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In CVPR, pages 652-660, 2017. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[31] Charles R Qi, Or Litany, Kaiming He, and Leonidas J Guibas. Deep hough voting for 3d object detection in point clouds. In ICCV, pages 9277-9286, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[32] Sonia Raychaudhuri, Tommaso Campari, Unnat Jain, Manolis Savva, and Angel X Chang. Mopa: Modular object navigation with pointgoal agents. In WACV, pages 5763-5773, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[33] Christoph B Rist, David Emmerichs, Markus Enzweiler, and Dariu M Gavrila. Semantic scene completion using local deep implicit functions on lidar data. TPAMI, 44(10):7205-7218, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[34] Danila Rukhovich, Anna Vorontsova, and Anton Konushin. Fcaf3d: Fully convolutional anchor-free 3d object detection. In ECCV, pages 477-493, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 342 + ], + "type": "text", + "content": "[35] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgb images. In ECCV, pages 746-760, 2012. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 385 + ], + "type": "text", + "content": "[36] Myrna C Silva, Mahtab Dahaghin, Matteo Toso, and Alessio Del Bue. Contrastive gaussian clustering: Weakly supervised 3d scene segmentation. arXiv preprint arXiv:2404.12784, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 430 + ], + "type": "text", + "content": "[37] Shuran Song, Fisher Yu, Andy Zeng, Angel X Chang, Manolis Savva, and Thomas Funkhouser. Semantic scene completion from a single depth image. In CVPR, pages 1746-1754, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 433, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 294, + 477 + ], + "type": "text", + "content": "[38] Jiakai Sun, Han Jiao, Guangyuan Li, Zhanjie Zhang, Lei Zhao, and Wei Xing. 3dgstream: On-the-fly training of 3d gaussians for efficient streaming of photo-realistic free-viewpoint videos. In CVPR, pages 20675-20685, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 479, + 294, + 510 + ], + "type": "text", + "content": "[39] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In ICML, pages 6105-6114, 2019. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 553 + ], + "type": "text", + "content": "[40] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In ICCV, pages 8406-8415, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 589 + ], + "type": "text", + "content": "[41] Thang Vu, Kookhoi Kim, Tung M Luu, Thanh Nguyen, and Chang D Yoo. Softgroup for 3d instance segmentation on point clouds. In CVPR, pages 2708-2717, 2022. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 634 + ], + "type": "text", + "content": "[42] Haiyang Wang, Lihe Ding, Shaocong Dong, Shaoshuai Shi, Aoxue Li, Jianan Li, Zhenguo Li, and Liwei Wang. Cagroup3d: Class-aware grouping for 3d object detection on point clouds. NeurIPS, 35:29975-29988, 2022. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 636, + 294, + 688 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 688 + ], + "type": "text", + "content": "[43] Lening Wang, Wenzhao Zheng, Yilong Ren, Han Jiang, Zhiyong Cui, Haiyang Yu, and Jiwen Lu. Occsora: 4d occupancy generation models as world simulators for autonomous driving. arXiv preprint arXiv:2405.20337, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[44] Tai Wang, Xiaohan Mao, Chenming Zhu, Runsen Xu, Ruiyuan Lyu, Peisen Li, Xiao Chen, Wenwei Zhang, Kai" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "Chen, Tianfan Xue, Xihui Liu, Cewu Lu, Dahua Lin, and Jiangmiao Pang. Embodiedscan: A holistic multi-modal 3d perception suite towards embodied ai. In CVPR, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 160 + ], + "type": "text", + "content": "[45] Xiaofeng Wang, Zheng Zhu, Wenbo Xu, Yunpeng Zhang, Yi Wei, Xu Chi, Yun Ye, Dalong Du, Jiwen Lu, and Xingang Wang. Openoccupancy: A large scale benchmark for surrounding semantic occupancy perception. arXiv preprint arXiv:2303.03991, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 204 + ], + "type": "text", + "content": "[46] Yi Wei, Linqing Zhao, Wenzhao Zheng, Zheng Zhu, Jie Zhou, and Jiwen Lu. Surroundocc: Multi-camera 3d occupancy prediction for autonomous driving. In ICCV, pages 21729-21740, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 248 + ], + "type": "text", + "content": "[47] Shun-Cheng Wu, Keisuke Tateno, Nassir Navab, and Federico Tombari. Scfusion: Real-time incremental scene reconstruction with semantic completion. In 3DV, pages 801-810. IEEE, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[48] Yuting Xiao, Xuan Wang, Jiafei Li, Hongrui Cai, Yanbo Fan, Nan Xue, Minghui Yang, Yujun Shen, and Shenghua Gao. Bridging 3d gaussian and mesh for freeview video rendering. arXiv preprint arXiv:2403.11453, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 295, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 295, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 295, + 553, + 336 + ], + "type": "text", + "content": "[49] Xiuwei Xu, Chong Xia, Ziwei Wang, Linqing Zhao, Yueqi Duan, Jie Zhou, and Jiwen Lu. Memory-based adapters for online 3d scene perception. arXiv preprint arXiv:2403.06974, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 371 + ], + "type": "text", + "content": "[50] Chi Yan, Delin Qu, Dan Xu, Bin Zhao, Zhigang Wang, Dong Wang, and Xuelong Li. Gs-slam: Dense visual slam with 3d gaussian splatting. In CVPR, pages 19595-19604, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 403 + ], + "type": "text", + "content": "[51] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2. arXiv preprint arXiv:2406.09414, 2024. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 459 + ], + "type": "text", + "content": "[52] Yu Yang, Jianbiao Mei, Yukai Ma, Siliang Du, Wenqing Chen, Yijie Qian, Yuxiang Feng, and Yong Liu. Driving in the occupancy world: Vision-centric 4d occupancy forecasting and planning via world models for autonomous driving. arXiv preprint arXiv:2408.14197, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 502 + ], + "type": "text", + "content": "[53] Zeyu Yang, Zijie Pan, Xiatian Zhu, Li Zhang, Yu-Gang Jiang, and Philip HS Torr. 4d gaussian splatting: Modeling dynamic scenes with native 4d primitives. arXiv preprint arXiv:2412.20720, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 504, + 553, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 556 + ], + "type": "text", + "content": "[54] Jiawei Yao, Chuming Li, Keqiang Sun, Yingjie Cai, Hao Li, Wanli Ouyang, and Hongsheng Li. Ndc-scene: Boost monocular 3d semantic scene completion in normalized device coordinates space. In ICCV, pages 9455-9465, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "text", + "content": "[55] Li Yi, Wang Zhao, He Wang, Minhyuk Sung, and Leonidas J Guibas. Gspn: Generative shape proposal network for 3d instance segmentation in point cloud. In CVPR, pages 3947-3956, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 635 + ], + "type": "text", + "content": "[56] Hongxiao Yu, Yuqi Wang, Yuntao Chen, and Zhaoxiang Zhang. Monocular occupancy prediction for scalable indoor scenes. arXiv preprint arXiv:2407.11730, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 669 + ], + "type": "text", + "content": "[57] Vladimir Yugay, Yue Li, Theo Gevers, and Martin R Oswald. Gaussian-slam: Photo-realistic dense slam with gaussian splatting. arXiv preprint arXiv:2312.10070, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[58] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In ECCV, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26369" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 295, + 117 + ], + "type": "text", + "content": "[59] Sicheng Zuo, Wenzhao Zheng, Yuanhui Huang, Jie Zhou, and Jiwen Lu. Pointocc: Cylindrical tri-perspective view for point-based 3d semantic occupancy prediction. arXiv preprint arXiv:2308.16896, 2023. 2" + } + ] + } + ], + "index": 0 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26370" + } + ] + } + ], + "index": 1 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_content_list.json b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cd7c41b931c7244b3caee69bef1ecaf8cd64dbc5 --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_content_list.json @@ -0,0 +1,1460 @@ +[ + { + "type": "text", + "text": "EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device", + "text_level": 1, + "bbox": [ + 112, + 128, + 883, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Gunjan Chhablani $^{1*}$ , Xiaomeng Ye $^{2}$ , Muhammad Zubair Irshad $^{3}$ , Zsolt Kira $^{4}$ , $^{1,2,4}$ Georgia Tech, $^{3}$ Toyota Research Institute", + "bbox": [ + 197, + 202, + 797, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chhablani.gunjan, muhammadzubairirshad}@gmail.com, {xye87, zkira}@gatech.edu", + "bbox": [ + 143, + 239, + 846, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of Embodied AI predominantly relies on simulation for training and evaluation, often using either fully synthetic environments that lack photorealism or high-fidelity real-world reconstructions captured with expensive hardware. As a result, sim-to-real transfer remains a major challenge. In this paper, we introduce EmbodiedSplat, a novel approach that personalizes policy training by efficiently capturing the deployment environment and fine-tuning policies within the reconstructed scenes. Our method leverages 3D Gaussian Splatting (GS) and the Habitat-Sim simulator to bridge the gap between realistic scene capture and effective training environments. Using iPhone-captured deployment scenes, we reconstruct meshes via GS, enabling training in settings that closely approximate real-world conditions. We conduct a comprehensive analysis of training strategies, pre-training datasets, and mesh reconstruction techniques, evaluating their impact on sim-to-real predictivity in real-world scenarios. Experimental results demonstrate that agents fine-tuned with EmbodiedSplat outperform both zero-shot baselines pre-trained on large-scale real-world datasets (HM3D) and synthetically generated datasets (HSSD), achieving absolute success rate improvements of $20\\%$ and $40\\%$ on real-world Image Navigation task. Moreover, our approach yields a high sim-vs-real correlation (0.87-0.97) for the reconstructed meshes, underscoring its effectiveness in adapting policies to diverse environments with minimal effort. Project page: https://gchhablani.github.io/embodied-splat.", + "bbox": [ + 88, + 321, + 483, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 775, + 220, + 791 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in Embodied AI have demonstrated impressive performance in simulated environments [17, 18, 32, 44, 46]. However, translating these capabilities to physical robots remains a significant challenge, primarily due to limitations in simulation fidelity and accessibility [21]. A", + "bbox": [ + 89, + 801, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7f16acf6b415ccc72ebda8492ade9cc1686285b2509ad2f81a0563857922492.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 290, + 903, + 356 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c7726c6f9d69e6904e81b6732783c27b6f3832e4f0bbee223deeb1d783a3554.jpg", + "image_caption": [ + "Figure 1. Overview of EmbodiedSplat: Mobile phone captures are used to generate reconstructed meshes via 3D Gaussian Splating (GS). Agents are trained within these reconstructed environments in simulation before being deployed in the real world, enabling effective sim-to-real transfer. Our analysis demonstrates a strong sim-to-real correlation across both types of generated meshes, highlighting their ability to bridge the gap between simulation and real-world performance." + ], + "image_footnote": [], + "bbox": [ + 517, + 358, + 901, + 539 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "key bottleneck is the sim-to-real gap, where handcrafted or synthetic simulation environments (e.g. HSSD [25]) struggle to capture the complexity and variability of real-world settings, necessitating the use of real-world reconstructions for effective policy training. On the other hand, real-world datasets such as Matterport3D [8] and HM3D [41] rely on expensive capture equipment and labor-intensive reconstruction pipelines, making large-scale scene collection and adaptation impractical for many applications. Furthermore, it is difficult to fully capture the variability of potential deployment environments with these datasets.", + "bbox": [ + 511, + 699, + 906, + 866 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Developments in 3D scene representations, particularly 3D Gaussian Splatting (GS) [24], have shown promise to", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done as a student at Georgia Tech, currently at Waymo.", + "bbox": [ + 107, + 886, + 439, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "25431", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/0f36f4c731378f24c97b882a5d20c1882f7e40d3038f47465d3ccda9c5b3db1f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
WorkGoal TypeE2ERealS2R
SplatNav [11]Point/Language×✓×
GaussNav [27]Image [26]×××
OursImage✓✓✓
", + "bbox": [ + 91, + 88, + 485, + 154 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Table 1. Comparison against recent works using GS for navigation. \"E2E\" indicates whether the policy is trained end-to-end, \"Real\" denotes evaluation on real-world robots, and \"S2R\" indicates demonstrated sim-to-real transfer.", + "bbox": [ + 89, + 164, + 482, + 220 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "wards reducing the effort needed to capture new scenes, enabling high-quality scene reconstruction from casual mobile phone captures. These approaches offer a strong ability to handle complex geometry, perform novel-view view synthesis, and provide high visual fidelity. Building upon this foundation, methods like DN-Splatter [49] have further enhanced mesh reconstruction quality through depth-and-normal regularization. However, their potential for training robot navigation policies and deploying them in the real-world has remained largely unexplored.", + "bbox": [ + 88, + 250, + 482, + 398 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we explore the following question: Can low-effort cellphone video captures be leveraged to generate meshes that facilitate the training of Embodied AI navigation policies for effective transfer to the target environment? In other words, we seek to personalize models by training them directly on the target distribution, as represented by 3D models built from low-cost and low-effort data collects of the deployment environment. Some recent works indeed explore GS for navigation - GaussNav [11, 27] explores Instance-Image Navigation [26] in simulation using GS; SplatNav [11] explores using GS-based collision meshes for collision avoidance and localization in a single real-world environment using a modular policy for a drone. In contrast, as shown in Tab. 1, we explore end-to-end policy learning for image-goal navigation and sim-to-real transfer in indoor environments. To the best of our knowledge, we are the first to explore an approach to solve the personalized real-to-sim-to-real problem using Gaussian Splats for indoor image-goal navigation.", + "bbox": [ + 91, + 400, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To achieve the above goal, we present a comprehensive framework for leveraging open-source 3D Gaussian Splating [24] (and compare with Polycam [38]). The central premise of our work is that it is possible to quickly capture the scenes in which a robot will be deployed, using readily available consumer-grade hardware, and seamlessly integrate them into simulation environments (i.e. Habitat-Sim [39]). Policies can then be trained in these simulated environments, leading to improved sim-to-real transfer. Our approach combines the accessibility of smartphone-based scene capture with recent advances in depth-aware 3D scene representation, enabling rapid training and deployment of navigation policies in new, realistic environments.", + "bbox": [ + 89, + 688, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, there are several challenges and open ques", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions, ranging from reconstruction quality to finetuning strategies, to enabling successful transfer. Building on the work of Silwal et al. [46] as a baseline, we test our framework across a range of scenes captured in a university environment, which is out-of-distribution for typical pre-trained policies. We analyze factors affecting transfer performance, including trade-offs between mesh generation pipelines, mesh quality for policy training, and training strategies (e.g., zero-shot vs. fine-tuning). Through rigorous evaluation and real-world robot experiments, we show that our approach yields significant gains in real-world image-goal navigation.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our key contributions can be summarized as follows:", + "bbox": [ + 511, + 272, + 864, + 287 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. An efficient and cost-effective pipeline for bridging the real-to-sim gap in navigation, enabling the creation of high-quality simulation scenes from low-cost, consumer-grade iPhone captures using depth-aware 3D Gaussian Splats (GS) and Polycam.", + "2. Comprehensive evaluations conducted in both simulation and real-world environments, assessing zero-shot and fine-tuned policies on our Captured scenes. Our results demonstrate substantial improvements in sim-to-real transfer, emphasizing the effectiveness of fine-tuning on high-fidelity reconstructions.", + "3. In-depth analysis addressing key research questions, exploring the relationship between reconstruction quality, pre-training scenes, downstream navigation performance, and training strategies. For example, a notable finding is that overfitting policies on a single high-fidelity scene reconstruction in simulation yields reasonable real-world performance. Our findings provide valuable insights into how reconstruction fidelity influences policy generalization and its applicability to real-world scenarios.", + "4. An open-source codebase and dataset to facilitate further research in this domain and reproducibility of results." + ], + "bbox": [ + 511, + 291, + 903, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through this work, we aim to make high-quality scene collection and agent training more accessible, as well as enable easy development of personalized agents.", + "bbox": [ + 511, + 641, + 903, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 702, + 653, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Scene Datasets", + "text_level": 1, + "bbox": [ + 511, + 727, + 663, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent years have seen rapid progress in the development of high-quality 3D scene datasets for embodied AI research. The Matterport3D dataset [8] provides 10,800 panoramic views from 194,400 RGB-D images across 90 buildings-scale scenes, complete with surface reconstructions and semantic annotations. Building upon this, HM3D [41] expanded the scale to 1,000 building-scale 3D reconstructions from diverse real-world locations, though some issues with mesh quality were noted. The Gibson dataset [53] introduced another collection of real-world scans, while", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "25432", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "synthetic datasets like HSSD [25] (with 211 realistic environments) demonstrated that smaller but higher-quality datasets can sometimes be better than larger ones for agent training. Another dataset that is commonly used is ReplicaCAD [47] which provides synthetic variations in layouts of a single scene. In this work, we use HM3D [41] for training our zero-shot baseline, as it is one of the largest photorealistic indoor-scene datasets captured using Matterport [1] cameras. Additionally, we use HSSD [25] as the synthetic counterpart for another zero-shot baseline. Note, however, that these scenes do have a bias towards apartments, and we therefore test our methods on scenes within a university environment, which is out-of-distribution for these datasets. The MuSHRoom dataset [42] provides multi-sensor captures of 10 real-world scenes, offering benchmarks for reconstruction and novel view synthesis methods. While this dataset has not been used for embodied tasks in prior works, our dataset collection method is inspired by the MuSHRoom dataset's iPhone captures. We evaluate our agents, as well as reconstruction strategies, on these scenes to benchmark our depth and normal encoders.", + "bbox": [ + 89, + 90, + 480, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Embodied Navigation in Indoor Environments", + "text_level": 1, + "bbox": [ + 89, + 421, + 480, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Embodied navigation encompasses various forms of goal-directed navigation tasks [2, 6, 9, 26, 31, 50, 55, 56]. This study primarily investigates ImageNav [2, 55, 56] in indoor environments. Recent advancements in image-goal navigation have explored a variety of approaches to enhance performance and generalization. Notably, pre-trained visual representations [32, 46, 55, 56] have shown substantial promise in improving performance on the ImageNav task. In particular, Silwal et al. [46] demonstrate remarkable zero-shot success rates (90%) on real-world ImageNav, leveraging VC-1 [32], fine-tuned end-to-end with data augmentation. Inspired by this, we use the setup in Silwal et al. [46] for our policy training and evaluation.", + "bbox": [ + 89, + 444, + 482, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sim-to-real transfer remains a core challenge in Embodied AI beyond leveraging pre-trained visual encoders. Kadian et al. [21] proposed the Sim-vs-Real Correlation Coefficient (SRCC) to quantify how well PointNav performance in simulation predicts real-world results. We build on this idea to evaluate the sim-to-real predictivity of our Captured meshes.", + "bbox": [ + 89, + 642, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A closely related work is Phone2Proc [13], which demonstrates enhanced sim-to-real ObjectNav performance by generating layouts from iPhone RoomPlan API captures. While similar to our approach in utilizing iPhone captures, our methodology differs in that we capture the entire room without focusing on layouts, nor do we perform any post-processing of scenes or generate multiple variations of the same scene for agent training. Instead, we generate meshes from our captures, fine-tune pre-trained policies for ImageNav task, and deploy them subsequently.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. 3D Scene and Mesh Reconstruction", + "text_level": 1, + "bbox": [ + 513, + 90, + 821, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we adopt DN-Splatter [49] for its simplicity and superior performance (and compare to meshes from Polycam [38], which is not open-source). DN-Splatter uses depth-and-normal regularization to improve mesh quality. While mesh reconstruction has also been explored with NeRFs [35], we choose GS for 3D reconstruction due to their fast training and rendering speeds, improving the overall efficiency of the pipeline. For an expanded list of related works on this topic, please see Appendix D.1.", + "bbox": [ + 511, + 113, + 906, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. 3D Representations for Embodied AI/Robotics", + "text_level": 1, + "bbox": [ + 511, + 261, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "There have been works which explore sim-to-real manipulation [7, 28, 52]. In this work, we attempt to solve the problem of sim-to-real navigation, laying the foundation for a diverse range of embodied tasks, such as object-goal navigation, rearrangement, and mobile manipulation, which require room-, apartment-, and building-level scene representations. Capturing high-quality 3D Gaussian Splats (GS) is relatively straightforward for tabletop manipulation tasks, whereas large-scale scene reconstruction poses greater challenges — a key aspect of this work. Recent works also explore the use of GS in solving the embodied navigation problem. GaussNav [27] presents a semantic Gaussian-based map reconstruction for the HM3D-Instance ImageNav [26] task. Splat-Nav [11] introduces a two-stage pipeline for planning and pose estimation through Gaussian Splat Maps. Tab. 1 highlights the differences between our approach and recent works utilizing GS for navigation. By capturing the entire scene using mobile phones, we enable personalization to specific scenes, as well as the potential to scale scene collection in the future. Some works [33] explore automated 3D scene capture using robot cameras in simulation, but these methods require complex, embodiment-dependent strategies that are time-consuming and difficult to scale. In contrast, our approach leverages human captures in the real-world, offering a fast and scalable solution without the need for intricate planning. For more related work on this topic, please see Appendix D.2.", + "bbox": [ + 511, + 285, + 906, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 513, + 708, + 648, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline for bridging and integrating a real-world scene with Habitat-Sim [39] is shown in Fig. 2. In the following subsections, we discuss each stage of the pipeline in detail. Sec. 3.1 discusses the datasets used and the details of scene captures. Sec. 3.2 discusses the second stage, where these captures are converted to meshes. Sec. 3.3 discusses how ImageNav episodes are created, which is a crucial part of integrating the scene with Habitat Simulator [39]. Appendix H discusses agent training and evaluation details. For details on real-world deployment setup, refer to Appendix A.", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "25433", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6002c9aa5b763a45cccefc9b1c847b2dec73bd4c8faa6fb56ed50c24a0f8cb27.jpg", + "image_caption": [ + "Figure 2. The EmbodiedSplat Pipeline: Pipeline for integration of real-world captures with Habitat-Sim [39] and subsequent deployment. The first stage (a.) involves capturing the scene using Polycam [38] and Nerfstudio [48] which produces RGB frames, associated iPhone GT depth maps, and poses. In the second stage (b.), we use DN-Splatter [49] to train GS using depth and normal regularization, with normals from Metric3D-V2 [16] monocular encoder. A mesh (.ply) is created using Poisson reconstruction from the GS. In the third stage (c.), the mesh is processed and loaded into Habitat-Sim [39] for training the agent in simulation. In the last stage (d.), the policy is deployed in the real-world in the same scene for image-goal navigation." + ], + "image_footnote": [], + "bbox": [ + 89, + 90, + 906, + 296 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Datasets and Scene Captures", + "text_level": 1, + "bbox": [ + 89, + 421, + 349, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Scene Datasets: We use two datasets for pre-training ImageNav policies, which serve as zero-shot baselines and pre-trained policies for our scenes, both in simulation and real-world evaluations. Specifically, we use the HM3D [41] dataset which consists of apartment-scale scenes split into 800 training and 100 validation scenes, and the HSSD [25] dataset split into 134 training and 33 validation scenes.", + "bbox": [ + 88, + 445, + 482, + 550 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Captured Scenes: To evaluate the feasibility of the pipeline and conduct real-world experiments, we capture scenes from a university environment (classroom, community lounges, conference rooms, etc). For custom data collection with an iPhone, we follow the procedure used for collecting the MuSHRoom dataset [42]. Specifically, we use an iPhone 13 Pro Max to record the iPhone RGB-D data using the Polycam application [38]. Polycam provides an assistive interface to ensure that all the details of the scene have been sufficiently covered during the capture. We use the default settings with Polycam and export the raw data exposed by the application. Subsequently, we use Nerf-studio [48] to process the RGB-D data and sample 1000 aligned RGB-depth frames with low blur scores and corresponding poses. Additionally, Polycam also provides a mesh with its exported data. We use this mesh for comparison purposes, referred to as POLYCAM meshes throughout the paper. Unlike Ren et al. [42], we use a manually-held iPhone for capture instead of a gimbal, enabling us to evaluate whether the process remains effective and easily replicable without relying on expensive capture equipment. Each capture requires 20-30 minutes of recording with Polycam. We repeat this process for different indoor scenes, which we", + "bbox": [ + 89, + 553, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "refer to as lounge, classroom, conf_a, and conf_b. The scale of these of these scenes is similar to those in the MuSHRoom dataset (1-3 rooms) [42].", + "bbox": [ + 511, + 421, + 906, + 467 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MuSHRoom Dataset: We use MuSHRoom [42] dataset for benchmarking. The dataset consists of long and short sequences of 10 indoor scenes, captured with iPhone, Kinect, and Faro Scanner. In our work, we only use the iPhone long-sequences for training and evaluation purposes. We benchmark different normal encoders on all 10 scenes (see Appendix G) for DN-Splatter [49]. For agent training and evaluation, we primarily use three scenes - honka, sauna, activity - which have varied scale and complexity.", + "bbox": [ + 511, + 468, + 908, + 604 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. 3D Scene and Mesh Reconstruction", + "text_level": 1, + "bbox": [ + 511, + 618, + 821, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For mesh reconstruction, we use DN-Splatter [49] as our method of choice for its superior performance on mesh reconstruction in comparison to others, and simplicity of its integration with Habitat-Sim [39]. DN-Splatter leverages depth-normal regularization, and smoothness losses to maintain geometric consistency during Gaussian Splat training. For mesh reconstruction, the rendered depth and normal maps are back-projected from training views to create a point-set for meshing.", + "bbox": [ + 509, + 642, + 906, + 777 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We use the default hyperparameters from DN-Splatter [49] and train the GS for 30,000 iterations. We use sensor depth $\\lambda_{d} = 0.2$ , and enable the depth smoothness and normal losses. Metric3D-V2 [16] is used as the normal encoder instead of Omnidata [14, 23], as empirical results demonstrate it yields higher-quality meshes. For further discussion on depth and normal encoder selection, see Appendix G.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "25434", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5740ae30a9454782880445ce517f7542d8303fbc3c6cb1be7b77fdf50b594d6c.jpg", + "image_caption": [ + "(a) conf.b" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 267, + 199 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/99562dac0ade027ceca080b7985028af6c15ec938a6c6ab5ce7f0b7746a1f523.jpg", + "image_caption": [ + "(b) conf_a" + ], + "image_footnote": [], + "bbox": [ + 308, + 103, + 475, + 199 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2273cddcecf21365cdc2119b8471557da8c581508fb16589a526f288bf99b219.jpg", + "image_caption": [ + "(c) classroom", + "Figure 3. Reconstructed meshes of the Captured scenes after DN-Splatter [49] training and Poisson reconstruction." + ], + "image_footnote": [], + "bbox": [ + 517, + 111, + 660, + 200 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ba08cb9ec672ff0d6dc02abb2166cbfb24d9c941ecc5d1ab7b4a8469c63f0ea.jpg", + "image_caption": [ + "(d) lounge" + ], + "image_footnote": [], + "bbox": [ + 730, + 90, + 895, + 200 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After training GS and converting into ply meshes, we convert the meshes to glb in Blender (see Appendix E) before loading them in Habitat-Sim [39]. The final meshes for our Captured scenes are shown in Fig. 3 and referred to as DN mesh hereafter to contrast against the POLYCAM mesh counterpart. It takes approximately 20-30 minutes per capture, and 1-2 hours of training with DN-Splatter [49] to generate these meshes, which is significantly lesser compared to the cost and several hours of capture and processing with Matterport [1] cameras.", + "bbox": [ + 89, + 268, + 483, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. ImageNav Episode Generation", + "text_level": 1, + "bbox": [ + 89, + 433, + 362, + 450 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Pre-training Scene Datasets: For the HM3D [41] and HSSD [25] datasets, we use the predefined train-validation split, generating 10,000 episodes for each training scene and 25 episodes for each validation scene (see Appendix B).", + "bbox": [ + 89, + 455, + 483, + 518 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Captured and MuSHRoom Scenes: For scenes with DN meshes and POLYCAM meshes, we select only the largest navmesh island within each scene. We ensure that the largest island covers most, if not all, of the navigable areas in the scene. This step is crucial due to the relatively smaller sizes of these scenes compared to HM3D scenes. Hence, we generate only 1000 training episodes and 100 evaluation episodes per scene. By selecting the largest navmesh island, we minimize the risk of incorrectly treating entities other than the floor as navigable.", + "bbox": [ + 89, + 518, + 483, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metric: We consider Success Rate (SR), the fraction of successful episodes out of all the episodes, as the primary metric for performance evaluation of our policies. An episode is considered successful if the agent stops within $1\\mathrm{m}$ of the goal location before the maximum number of steps (1000 for simulation, 100 for real) are over.", + "bbox": [ + 89, + 671, + 483, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 781, + 223, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section aims to answer the following research questions:", + "bbox": [ + 89, + 808, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. How does a pre-trained policy perform on Captured and MuSHRoom scenes? (Sec. 4.1)", + "2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance? (Sec. 4.2)." + ], + "bbox": [ + 91, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3. Does the performance in simulation transfer to the realworld? (Sec. 4.3).", + "bbox": [ + 511, + 268, + 903, + 297 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Further experiments and analyses are discussed in Sec. 5.", + "bbox": [ + 511, + 297, + 890, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. How does a pre-trained policy perform on Captured and MuSHRoom scenes?", + "text_level": 1, + "bbox": [ + 511, + 321, + 903, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, we evaluate the pre-trained policies that achieved the highest validation success rates on the HM3D [41] (83.08% val SR) and HSSD [25] (63.15% val SR), trained for 600M and 1200M steps, respectively. Note that the performance on HM3D is consistent with the simulation results reported in Silwal et al. [46] and represents the current state-of-the-art for this task. For agent embodiment, policy training and evaluation details, please refer to Appendix H. We assess the zero-shot generalization of these policies within simulation, evaluating their performance across different meshes for individual scenes.", + "bbox": [ + 511, + 359, + 906, + 525 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39d0b3ff99eade48b144ed040e17c4c787b65e637b9ec608ad8d786bd42002b5.jpg", + "image_caption": [ + "Figure 4. Zero-shot val. SR for HM3D pre-trained policy." + ], + "image_footnote": [], + "bbox": [ + 517, + 537, + 898, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 4 presents the results of the zero-shot evaluation across individual scenes for the HM3D pre-trained policy. Performance varies significantly based on scene size, complexity, and mesh type. For smaller scenes such as conf_a and conf_b, the policy demonstrates relatively high success rates in both DN mesh and POLYCAM settings, achieving $85\\%$ (DN) and $82\\%$ (POLYCAM) for conf_a and $88\\%$ (DN mesh) and $79\\%$ (POLYCAM) for conf_b.", + "bbox": [ + 511, + 719, + 905, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In contrast, performance declines in larger environments such as classroom and lounge. In classroom, success rates drop to $53\\%$ with DN mesh and $42\\%$ with POLYCAM, while in lounge, the policy achieves $50\\%$ (DN", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "25435", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cecb7120ff12bf0c11d540300add3e767fab168f86faedf352804101c4e67e1.jpg", + "image_caption": [ + "HSSD Zero-Shot Results on Captured and MuSHRoom Scenes" + ], + "image_footnote": [], + "bbox": [ + 94, + 103, + 468, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mesh) and $76\\%$ (POLYCAM). These results highlight the challenge of transferring the policy to more complex, large-scale environments that differ from the training data. Since HM3D consists primarily of apartment-style scenes, the policy has not been exposed to classrooms or community lounges during training. Both DN and POLYCAM meshes yield similar success rates on average $(\\sim 60\\%)$ , suggesting no clear advantage of one mesh type over the other in aligning with HM3D. Additionally, we evaluate the policies on three scenes from the MuSHRoom [42] dataset. Due to the unavailability of POLYCAM meshes for MuSHRoom, we conduct evaluations solely on DN meshes. The policy achieves a $95\\%$ success rate in the sauna and honka scenes, suggesting promising potential for performance improvements through additional scene refinements (such as using a gimbal during capture). However, in the larger activity scene, performance declines, indicating challenges in generalizing to larger environments.", + "bbox": [ + 88, + 280, + 482, + 551 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 5 reveals similar trends for the HSSD pre-trained policy, though with significantly lower success rates overall. This degradation in performance can be attributed to the synthetic nature of HSSD scenes, which lack the realism and scale necessary for effective generalization. The policy performs particularly poorly on classroom, achieving only $1\\%$ success on both DN mesh and POLYCAM, while performance in lounge is slightly better, reaching $14\\%$ success with POLYCAM. The policy exhibits slightly improved performance on other scenes, but it is hard to conclude which mesh type best aligns with HSSD.", + "bbox": [ + 89, + 551, + 482, + 719 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance?", + "text_level": 1, + "bbox": [ + 89, + 727, + 482, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we investigate whether fine-tuning the pre-trained policies for only 20M additional steps improves performance in simulation. The learning rate is set to $2.5e - 6$ for the LSTM policy and $6e - 7$ for the visual encoder, following a fine-tuning strategy similar to that of Deitke et al. [13]. We fine-tune the pre-trained policy on the training episodes for a single scene and evaluate on the corresponding validation episodes.", + "bbox": [ + 89, + 763, + 482, + 885 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results are presented in Fig. 6 and Fig. 7. We ob", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/601aeea520cda99e5bc897510d426123c502cb35e23164404e378baf52aa3931.jpg", + "image_caption": [ + "Figure 5. Zero-shot val. SR for HSSD pre-trained policy.", + "HM3D Fine-tuned Results on Captured and MuSHRoom Scenes", + "Figure 6. Fine-tuned validation SR from HM3D pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + ], + "image_footnote": [], + "bbox": [ + 516, + 103, + 888, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/219958e3f0fb2eab600e9c11125a03294961fcd86030cbd276ef00ddb761ca8c.jpg", + "image_caption": [ + "Figure 7. Fine-tuned validation SR from HSSD pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + ], + "image_footnote": [], + "bbox": [ + 516, + 284, + 898, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "serve that fine-tuning significantly improves performance across all tested scenes. For the pre-trained HM3D policy, fine-tuning for 20M steps results in success rates approaching $90\\%+$ across different meshes. Similarly, for the HSSD pre-trained policy, fine-tuning leads to substantial improvements, with most policies achieving success rates of $80\\%+$ in respective scenes. In particular, performance gains are particularly pronounced in larger, more complex environments such as classroom and lounge, which differ significantly from apartment-style scenes in HM3D and HSSD. These results indicate that our pipeline can effectively be used to collect diverse scene data at scale. Furthermore, they suggest that agents can be quickly finetuned (only 20M steps vs $100\\mathrm{M}+$ steps for training from scratch) on these generated meshes to improve performance on specific scene types, enhancing personalization beyond the original training distribution. We present results on additional scenes in Appendix J.", + "bbox": [ + 511, + 489, + 906, + 762 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Does the performance in simulation transfer to the real-world?", + "text_level": 1, + "bbox": [ + 511, + 772, + 903, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate both zero-shot and fine-tuned policies in the real-world lounge scene on a Stretch robot. During the evaluation, each episode is capped at 100 steps, with 10 distinct start-and-goal locations sampled within the scene (see Appendix A for deployment details). To assess performance, we record the number of steps taken and final", + "bbox": [ + 511, + 809, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "25436", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "distance to the goal at the end of each episode, determining whether the agent successfully completes the task.", + "bbox": [ + 89, + 90, + 485, + 122 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/20c70cec609ca492acba284f0e3cca7385aeda0d60c1bf9887857a2748d04df4.jpg", + "image_caption": [ + "Figure 8. Real world results of zero-shot and fine-tuned models on 1ounge scene. HM3D-ZS (real dataset) leads to a better success rate than HSSD-ZS (synthetic dataset). We observe increased real-world success with fine-tuning on DN and POLYCAM meshes for both HM3D and HSSD policies." + ], + "image_footnote": [], + "bbox": [ + 94, + 135, + 475, + 277 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 8 presents results across 10 evaluation episodes. The zero-shot HM3D policy achieves a $50\\%$ success rate, demonstrating our hypothesized lack of generalization. This is in contrast with the results reported in Silwal et al. [46] showing $90\\%$ zero-shot success rate in the real-world. We attribute this discrepancy to the structural and semantic differences between the lounge and the apartment-style scenes typically encountered in HM3D. Fine-tuning on the POLYCAM and DN mesh reconstructions of this scene improves performance, with success rates increasing to $70\\%$ . For HSSD, zero-shot performance is significantly lower at $10\\%$ , while fine-tuned policies improve success rates to $50\\%$ with POLYCAM and $40\\%$ with DN mesh.", + "bbox": [ + 89, + 377, + 482, + 571 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 1 illustrates the Sim-to-Real Correlation Coefficient (SRCC) [21] between simulation and real-world performance. The observation suggests that improvements in evaluation performance on DN and POLYCAM meshes in simulation translate to improved real-world performance. This demonstrates that our approach can efficiently adapt policies to novel real-world environments. We discuss real-world statistics (number of steps, distance) in Appendix I.", + "bbox": [ + 89, + 574, + 482, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Ablations and Analysis", + "text_level": 1, + "bbox": [ + 89, + 709, + 308, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we systematically investigate critical factors influencing policy performance. First, we assess whether pre-training on large-scale datasets such as HM3D or HSSD is necessary for effective real-world transfer (Sec. 5.1). Next, we analyze the impact of validation PSNR and scene geometry on the zero-shot performance (Sec. 5.2). Finally, we examine whether continuous training on large-scale datasets improves zero-shot performance on our Captured scenes, shedding light on the relationship between dataset characteristics and generalizability (Sec. 5.3).", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Is it necessary to pre-train over large-scale datasets?", + "text_level": 1, + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this experiment, we investigate the necessity of pretraining on large-scale datasets by \"overfitting\" policies directly on POLYCAM and DN meshes with a policy trained from scratch (not pre-trained on large-scale datasets such as HM3D or HSSD) for $\\sim 100\\mathrm{M}$ steps. Training and validation data are derived from ImageNav episode generation, as described in Sec. 3.3. To ensure diversity, start-goal locations for training and evaluation episodes are randomly sampled. The best validation success rates for all policies on their respective scenes are presented in Fig. 9. As expected, the overfitted policies achieve near-perfect performance in simulation, maintaining high success rates. For a detailed discussion on why overfitting results seem better than fine-tuning in this scenario, please see Appendix L.", + "bbox": [ + 511, + 127, + 906, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c837f32cfba2b5daf80ef7e04e8461909eff2a64da2a5af9ae3435a4b2ce8380.jpg", + "image_caption": [ + "Figure 9. Validation SR for overfitted policies on respective scenes: Polices are not pre-trained on HM3D/HSSD." + ], + "image_footnote": [], + "bbox": [ + 516, + 353, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We further evaluate these overfitted policies in the real-world lounge scene using POLYCAM and DN meshes. Surprisingly, the overfitted policy trained on the POLYCAM mesh achieves a $50\\%$ success rate in real-world evaluations. In contrast, the policy trained on the DN mesh achieves only $10\\%$ success. This result suggests that our mesh-based training approach can yield non-zero real-world performance, even without large-scale pre-training. We attribute the performance gap between POLYCAM and DN-trained policies to differences in visual fidelity—POLYCAM meshes preserve more visual detail by directly utilizing original images to reconstruct the scene, whereas DN meshes are based on GS which use the learned colors for the 3D Gaussians.", + "bbox": [ + 511, + 551, + 908, + 750 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. How is the scale and validation PSNR of the scene related to the final performance?", + "text_level": 1, + "bbox": [ + 511, + 757, + 906, + 789 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 10 illustrates the correlation between HM3D zero-shot validation success rates on DN meshes and two key factors: the Peak Signal-to-Noise Ratio (PSNR) of the corresponding 3D Gaussian Splats (GS) and the scale of the scene, measured as the average shortest distance between start-goal locations in validation episodes. We observe a negative correlation between success rate (SR) and average", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "25437", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01c3bf71b84b52d8d26c633cc7033e0ab77fa9b82b6d0dd85ce8527568927565.jpg", + "image_caption": [ + "Figure 10. HM3D zero-shot validation success rates on DN meshes vs. validation GS PSNRs vs. average shortest distances of validation episodes. The zero-shot SR is inversely correlated with the scale of the scene, while directly correlated with val. PSNR on the trained GS." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 480, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shortest distance—indicating that as the scale of the scene increases, the zero-shot success rate declines. Conversely, a positive correlation is seen between SR and PSNR, where higher validation PSNR values correspond to improved success rates. Notably, different trend lines emerge for MuSHRoom captures and our own Captured scenes. MuSHRoom captures generally exhibit higher validation PSNRs, likely due to the use of a stabilized gimbal during data collection. Future work will further investigate the impact of capture stability on policy performance.", + "bbox": [ + 88, + 348, + 482, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Does continuous training on large-scale datasets improve zero-shot performance?", + "text_level": 1, + "bbox": [ + 89, + 510, + 482, + 544 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e06aafad43397d2acb1bc4aee357c3c3b59607cdbbc99007fd5228834aaec64.jpg", + "image_caption": [ + "Figure 11. Average zero-shot success rates on respective validation sets over different stages of pre-training on HM3D." + ], + "image_footnote": [], + "bbox": [ + 94, + 565, + 478, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 11 presents average zero-shot success rates across HM3D validation set, POLYCAM meshes, and DN meshes at various stages of HM3D policy pre-training. This analysis aims to determine whether continuous performance improvements on HM3D validation scenes translate to improved zero-shot performance on our Captured scenes. We observe that while performance initially increases, it begins to deteriorate or plateau at approximately 400M steps, despite continued improvements on HM3D", + "bbox": [ + 88, + 763, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d2bad78c6e6af564e8a4d7a8097e0f59851c0beda5062ce327721f60df0c5b12.jpg", + "image_caption": [ + "Figure 12. Average zero-shot success rates on respective validation sets over different stages of pre-training on HSSD." + ], + "image_footnote": [], + "bbox": [ + 517, + 92, + 903, + 231 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "validation scenes.", + "bbox": [ + 511, + 300, + 633, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 12 shows a similar experiment conducted using the HSSD dataset. Up to 300M steps, improvements in HSSD validation performance correspond to slight improvements in zero-shot performance on our Captured scenes. However, performance plateaus for Captured, showing no further gains despite continued improvement on HSSD validation scenes. For HM3D pre-training, POLYCAM meshes outperform DN meshes in success rates, while for HSSD, the trend reverses slightly, highlighting the impact of pretraining dataset characteristics on zero-shot generalization.", + "bbox": [ + 511, + 315, + 906, + 467 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 481, + 633, + 496 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we introduced a comprehensive pipeline for bridging the gap between real-world and simulated environments in training embodied agents using 3D Gaussian Splats (GS) and Polycam. By leveraging the MuSHRoom dataset and custom iPhone-captured scenes, we demonstrated an efficient and scalable approach to policy personalization, leveraging 3D scene reconstruction from low-effort collects, enabling high-quality training for the ImageNav task. Our pipeline facilitates accessible and replicable scene collection without requiring specialized hardware or significant costs, making it a practical solution for large-scale embodied AI research. We evaluated overfitted, zero-shot, and fine-tuned policies, showing that fine-tuning pre-trained policies on real-world scene reconstructions improves sim-to-real transfer. We also analyzed the differences between GS-generated DN meshes and POLYCAM meshes, finding that POLYCAM more closely resemble real-world scenes.", + "bbox": [ + 511, + 507, + 906, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work lays the foundation for seamlessly integrating real-world scene captures into simulation, expanding the applicability of embodied AI systems. Another promising avenue is integrating GS directly into training, replacing visual observations and goal images to enhance learning efficiency. Furthermore, we aim to extend the use of Gaussian Splats to more complex embodied AI tasks, such as rearrangement and mobile manipulation, broadening its impact across diverse real-world applications.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "25438", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Capture, share, and collaborate the built world in immersive 3D — matterport.com. https://matterport.com/. [Accessed 08-03-2025]. 3, 5", + "[2] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation, 2022. 3", + "[3] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 4", + "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2", + "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 4", + "[6] Guillaume Bono, Leonid Antsfeld, Boris Chidlovskii, Philippe Weinzaepfel, and Christian Wolf. End-to-end (instance)-image goal navigation through correspondence as an emergent phenomenon, 2023. 3", + "[7] Arunkumar Byravan, Jan Humplik, Leonard Hasenclever, Arthur Brussee, Francesco Nori, Tuomas Haarnoja, Ben Moran, Steven Bohez, Fereshteh Sadeghi, Bojan Vujatovic, and Nicolas Heess. Nerf2real: Sim2real transfer of vision-guided bipedal motion skills using neural radiance fields, 2022. 3, 2", + "[8] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments, 2017. 1, 2", + "[9] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration, 2020. 3", + "[10] Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan, Muhammad Zubair Irshad, Richard Cheng, Kurt Keutzer, Masayoshi Tomizuka, Quan Vuong, and Ken Goldberg. Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning. In Conference on Robot Learning (CoRL), Munich, Germany, 2024. 2", + "[11] Timothy Chen, Ola Shorinwa, Joseph Bruno, Javier Yu, Wei-jia Zeng, Keiko Nagami, Philip Dames, and Mac Schwager. Splat-nav: Safe real-time robot navigation in gaussian splatting maps, 2024. 2, 3", + "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels, 2024. 2", + "[13] Matt Deitke, Rose Hendrix, Luca Weihs, Ali Farhadi, Kiana Ehsani, and Aniruddha Kembhavi. Phone2proc: Bringing robust robots into our chaotic world, 2022. 3, 6", + "[14] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4, 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering, 2023. 2", + "[16] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. 2024. 4, 3", + "[17] Muhammad Zubair Irshad, Chih-Yao Ma, and Zsolt Kira. Hierarchical cross-modal agent for robotics vision-and-language navigation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2021. 1", + "[18] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4065-4071, 2022. 1", + "[19] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. 2023. 2", + "[20] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv:2410.20220, 2024. 2", + "[21] Abhishek Kadian, Joanne Truong, Aaron Gokaslan, Alexander Clegg, Erik Wijmans, Stefan Lee, Manolis Savva, Sonia Chernova, and Dhruv Batra. Sim2real predictivity: Does evaluation in simulation predict real-world performance? IEEE Robotics and Automation Letters, 5(4):6670-6677, 2020. 1, 3, 7", + "[22] Ivan Kapelyukh, Yifei Ren, Ignacio Alzugaray, and Edward Johns. Dream2real: Zero-shot 3d object rearrangement with vision-language models, 2024. 2", + "[23] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4", + "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2", + "[25] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Schacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. arXiv preprint arXiv:2306.11290, 2023. 1, 3, 4, 5", + "[26] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 2, 3", + "[27] Xiaohan Lei, Min Wang, Wengang Zhou, and Houqiang Li. Gaussnav: Gaussian splatting for visual navigation, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "25439", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Yulong Li and Deepak Pathak. Object-aware gaussian splatt ing for robotic manipulation. In ICRA 2024 Workshop on 3D Visual Representations for Robot Manipulation, 2024. 3, 2", + "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 3", + "[30] JunEn Low, Maximilian Adang, Javier Yu, Keiko Nagami, and Mac Schwager. Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum, 2025. 2", + "[31] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings, 2023. 3", + "[32] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, Dhruv Batra, Yixin Lin, Oleksandr Maksymets, Aravind Rajeswaran, and Franziska Meier. Where are we in the search for an artificial visual cortex for embodied intelligence?, 2024. 1, 3", + "[33] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra Singh Chaplot. Autonerf: Training implicit scene representations with autonomous agents, 2023. 3, 2", + "[34] Wugang Meng, Tianfu Wu, Huan Yin, and Fumin Zhang. Beings: Bayesian embodied image-goal navigation with gaussian splatting, 2024. 2", + "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3, 2", + "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2", + "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4", + "[38] Polycam. Polycam. Accessed: 2024-11-10. 2, 3, 4", + "[39] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. In The Twelfth International Conference on Learning Representations, 2023. 2, 3, 4, 5, 1", + "[40] Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun, David Held, George Kantor, and Abhisesh Silwal. Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting, 2024. 2", + "[41] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai, 2021. 1, 2, 3, 4, 5" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Xuqian Ren, Wenjia Wang, Dingding Cai, Tuuli Tuominen, Juho Kannala, and Esa Rahtu. Mushroom: Multi-sensor hybrid room dataset for joint 3d reconstruction and novel view synthesis, 2024. 3, 4, 6, 2", + "[43] Meta AI Research. Home robot. https://github.com/facebookresearch/home-robot, 2023. GitHub repository. 1", + "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 1", + "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 3", + "[46] Sneha Silwal, Karmesh Yadav, Tingfan Wu, Jay Vakil, Arjun Majumdar, Sergio Arnaud, Claire Chen, Vincent-Pierre Berges, Dhruv Batra, Aravind Rajeswaran, Mrinal Kalakrishnan, Franziska Meier, and Oleksandr Maksymets. What do we learn from a large-scale study of pre-trained visual representations in sim and real environments?, 2024. 1, 2, 3, 5, 7, 4", + "[47] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3", + "[48] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 4", + "[49] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splattering and meshing, 2024. 2, 3, 4, 5", + "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In International Conference on Learning Representations, 2019. 3", + "[51] Yaniv Wolf, Amit Bracha, and Ron Kimmel. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views, 2024. 2", + "[52] Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, and Hesheng Wang. Rl-gsbridge: 3d gaussian splating based real2sim2real method for robotic manipulation learning, 2024. 3, 2", + "[53] Fei Xia, Amir Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents, 2018. 2", + "[54] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "25440", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "browser-compatible environment from a single video, 2024. 2", + "[55] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation, 2022. 3", + "[56] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav, 2023. 3, 4", + "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 4", + "[58] Justin Yu, Kush Hari, Kishore Srinivas, Karim El-Refai, Adam Rashid, Chung Min Kim, Justin Kerr, Richard Cheng, Muhammad Zubair Irshad, Ashwin Balakrishna, et al. Language-embedded gaussian splats (legs): incrementally building room-scale representations with a mobile robot. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13326–13332. IEEE, 2024. 2", + "[59] Justin Yu, Kush Hari, Karim El-Refai, Arnav Dalil, Justin Kerr, Chung-Min Kim, Richard Cheng, Muhammad Z. Irshad, and Ken Goldberg. Persistent object gaussian splat (pogs) for tracking human and robot manipulation of irregularly shaped objects. ICRA, 2025. 2", + "[60] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey, 2024. 2" + ], + "bbox": [ + 91, + 90, + 482, + 486 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "25441", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_model.json b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5555f6ab37764008736826a75812d03ca684f515 --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_model.json @@ -0,0 +1,2246 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.114, + 0.13, + 0.885, + 0.177 + ], + "angle": 0, + "content": "EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device" + }, + { + "type": "text", + "bbox": [ + 0.198, + 0.203, + 0.799, + 0.24 + ], + "angle": 0, + "content": "Gunjan Chhablani\\(^{1*}\\), Xiaomeng Ye\\(^{2}\\), Muhammad Zubair Irshad\\(^{3}\\), Zsolt Kira\\(^{4}\\), \\(^{1,2,4}\\)Georgia Tech, \\(^{3}\\)Toyota Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.145, + 0.241, + 0.848, + 0.258 + ], + "angle": 0, + "content": "{chhablani.gunjan, muhammadzubairirshad}@gmail.com, {xye87, zkira}@gatech.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.323, + 0.485, + 0.749 + ], + "angle": 0, + "content": "The field of Embodied AI predominantly relies on simulation for training and evaluation, often using either fully synthetic environments that lack photorealism or high-fidelity real-world reconstructions captured with expensive hardware. As a result, sim-to-real transfer remains a major challenge. In this paper, we introduce EmbodiedSplat, a novel approach that personalizes policy training by efficiently capturing the deployment environment and fine-tuning policies within the reconstructed scenes. Our method leverages 3D Gaussian Splatting (GS) and the Habitat-Sim simulator to bridge the gap between realistic scene capture and effective training environments. Using iPhone-captured deployment scenes, we reconstruct meshes via GS, enabling training in settings that closely approximate real-world conditions. We conduct a comprehensive analysis of training strategies, pre-training datasets, and mesh reconstruction techniques, evaluating their impact on sim-to-real predictivity in real-world scenarios. Experimental results demonstrate that agents fine-tuned with EmbodiedSplat outperform both zero-shot baselines pre-trained on large-scale real-world datasets (HM3D) and synthetically generated datasets (HSSD), achieving absolute success rate improvements of \\(20\\%\\) and \\(40\\%\\) on real-world Image Navigation task. Moreover, our approach yields a high sim-vs-real correlation (0.87-0.97) for the reconstructed meshes, underscoring its effectiveness in adapting policies to diverse environments with minimal effort. Project page: https://gchhablani.github.io/embodied-splat." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.222, + 0.792 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Recent advancements in Embodied AI have demonstrated impressive performance in simulated environments [17, 18, 32, 44, 46]. However, translating these capabilities to physical robots remains a significant challenge, primarily due to limitations in simulation fidelity and accessibility [21]. A" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.291, + 0.905, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.359, + 0.903, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.665 + ], + "angle": 0, + "content": "Figure 1. Overview of EmbodiedSplat: Mobile phone captures are used to generate reconstructed meshes via 3D Gaussian Splating (GS). Agents are trained within these reconstructed environments in simulation before being deployed in the real world, enabling effective sim-to-real transfer. Our analysis demonstrates a strong sim-to-real correlation across both types of generated meshes, highlighting their ability to bridge the gap between simulation and real-world performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.867 + ], + "angle": 0, + "content": "key bottleneck is the sim-to-real gap, where handcrafted or synthetic simulation environments (e.g. HSSD [25]) struggle to capture the complexity and variability of real-world settings, necessitating the use of real-world reconstructions for effective policy training. On the other hand, real-world datasets such as Matterport3D [8] and HM3D [41] rely on expensive capture equipment and labor-intensive reconstruction pipelines, making large-scale scene collection and adaptation impractical for many applications. Furthermore, it is difficult to fully capture the variability of potential deployment environments with these datasets." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Developments in 3D scene representations, particularly 3D Gaussian Splatting (GS) [24], have shown promise to" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.441, + 0.901 + ], + "angle": 0, + "content": "*Work done as a student at Georgia Tech, currently at Waymo." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25431" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.486, + 0.155 + ], + "angle": 0, + "content": "
WorkGoal TypeE2ERealS2R
SplatNav [11]Point/Language×✓×
GaussNav [27]Image [26]×××
OursImage✓✓✓
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.165, + 0.483, + 0.221 + ], + "angle": 0, + "content": "Table 1. Comparison against recent works using GS for navigation. \"E2E\" indicates whether the policy is trained end-to-end, \"Real\" denotes evaluation on real-world robots, and \"S2R\" indicates demonstrated sim-to-real transfer." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.25, + 0.483, + 0.4 + ], + "angle": 0, + "content": "wards reducing the effort needed to capture new scenes, enabling high-quality scene reconstruction from casual mobile phone captures. These approaches offer a strong ability to handle complex geometry, perform novel-view view synthesis, and provide high visual fidelity. Building upon this foundation, methods like DN-Splatter [49] have further enhanced mesh reconstruction quality through depth-and-normal regularization. However, their potential for training robot navigation policies and deploying them in the real-world has remained largely unexplored." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.401, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In this work, we explore the following question: Can low-effort cellphone video captures be leveraged to generate meshes that facilitate the training of Embodied AI navigation policies for effective transfer to the target environment? In other words, we seek to personalize models by training them directly on the target distribution, as represented by 3D models built from low-cost and low-effort data collects of the deployment environment. Some recent works indeed explore GS for navigation - GaussNav [11, 27] explores Instance-Image Navigation [26] in simulation using GS; SplatNav [11] explores using GS-based collision meshes for collision avoidance and localization in a single real-world environment using a modular policy for a drone. In contrast, as shown in Tab. 1, we explore end-to-end policy learning for image-goal navigation and sim-to-real transfer in indoor environments. To the best of our knowledge, we are the first to explore an approach to solve the personalized real-to-sim-to-real problem using Gaussian Splats for indoor image-goal navigation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.884 + ], + "angle": 0, + "content": "To achieve the above goal, we present a comprehensive framework for leveraging open-source 3D Gaussian Splating [24] (and compare with Polycam [38]). The central premise of our work is that it is possible to quickly capture the scenes in which a robot will be deployed, using readily available consumer-grade hardware, and seamlessly integrate them into simulation environments (i.e. Habitat-Sim [39]). Policies can then be trained in these simulated environments, leading to improved sim-to-real transfer. Our approach combines the accessibility of smartphone-based scene capture with recent advances in depth-aware 3D scene representation, enabling rapid training and deployment of navigation policies in new, realistic environments." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "However, there are several challenges and open ques" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "tions, ranging from reconstruction quality to finetuning strategies, to enabling successful transfer. Building on the work of Silwal et al. [46] as a baseline, we test our framework across a range of scenes captured in a university environment, which is out-of-distribution for typical pre-trained policies. We analyze factors affecting transfer performance, including trade-offs between mesh generation pipelines, mesh quality for policy training, and training strategies (e.g., zero-shot vs. fine-tuning). Through rigorous evaluation and real-world robot experiments, we show that our approach yields significant gains in real-world image-goal navigation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.866, + 0.288 + ], + "angle": 0, + "content": "Our key contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.292, + 0.905, + 0.367 + ], + "angle": 0, + "content": "1. An efficient and cost-effective pipeline for bridging the real-to-sim gap in navigation, enabling the creation of high-quality simulation scenes from low-cost, consumer-grade iPhone captures using depth-aware 3D Gaussian Splats (GS) and Polycam." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.457 + ], + "angle": 0, + "content": "2. Comprehensive evaluations conducted in both simulation and real-world environments, assessing zero-shot and fine-tuned policies on our Captured scenes. Our results demonstrate substantial improvements in sim-to-real transfer, emphasizing the effectiveness of fine-tuning on high-fidelity reconstructions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.458, + 0.905, + 0.607 + ], + "angle": 0, + "content": "3. In-depth analysis addressing key research questions, exploring the relationship between reconstruction quality, pre-training scenes, downstream navigation performance, and training strategies. For example, a notable finding is that overfitting policies on a single high-fidelity scene reconstruction in simulation yields reasonable real-world performance. Our findings provide valuable insights into how reconstruction fidelity influences policy generalization and its applicability to real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.904, + 0.639 + ], + "angle": 0, + "content": "4. An open-source codebase and dataset to facilitate further research in this domain and reproducibility of results." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.292, + 0.905, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.905, + 0.688 + ], + "angle": 0, + "content": "Through this work, we aim to make high-quality scene collection and agent training more accessible, as well as enable easy development of personalized agents." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.703, + 0.655, + 0.718 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.664, + 0.742 + ], + "angle": 0, + "content": "2.1. Scene Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Recent years have seen rapid progress in the development of high-quality 3D scene datasets for embodied AI research. The Matterport3D dataset [8] provides 10,800 panoramic views from 194,400 RGB-D images across 90 buildings-scale scenes, complete with surface reconstructions and semantic annotations. Building upon this, HM3D [41] expanded the scale to 1,000 building-scale 3D reconstructions from diverse real-world locations, though some issues with mesh quality were noted. The Gibson dataset [53] introduced another collection of real-world scans, while" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "25432" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.409 + ], + "angle": 0, + "content": "synthetic datasets like HSSD [25] (with 211 realistic environments) demonstrated that smaller but higher-quality datasets can sometimes be better than larger ones for agent training. Another dataset that is commonly used is ReplicaCAD [47] which provides synthetic variations in layouts of a single scene. In this work, we use HM3D [41] for training our zero-shot baseline, as it is one of the largest photorealistic indoor-scene datasets captured using Matterport [1] cameras. Additionally, we use HSSD [25] as the synthetic counterpart for another zero-shot baseline. Note, however, that these scenes do have a bias towards apartments, and we therefore test our methods on scenes within a university environment, which is out-of-distribution for these datasets. The MuSHRoom dataset [42] provides multi-sensor captures of 10 real-world scenes, offering benchmarks for reconstruction and novel view synthesis methods. While this dataset has not been used for embodied tasks in prior works, our dataset collection method is inspired by the MuSHRoom dataset's iPhone captures. We evaluate our agents, as well as reconstruction strategies, on these scenes to benchmark our depth and normal encoders." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.482, + 0.437 + ], + "angle": 0, + "content": "2.2. Embodied Navigation in Indoor Environments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Embodied navigation encompasses various forms of goal-directed navigation tasks [2, 6, 9, 26, 31, 50, 55, 56]. This study primarily investigates ImageNav [2, 55, 56] in indoor environments. Recent advancements in image-goal navigation have explored a variety of approaches to enhance performance and generalization. Notably, pre-trained visual representations [32, 46, 55, 56] have shown substantial promise in improving performance on the ImageNav task. In particular, Silwal et al. [46] demonstrate remarkable zero-shot success rates (90%) on real-world ImageNav, leveraging VC-1 [32], fine-tuned end-to-end with data augmentation. Inspired by this, we use the setup in Silwal et al. [46] for our policy training and evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Sim-to-real transfer remains a core challenge in Embodied AI beyond leveraging pre-trained visual encoders. Kadian et al. [21] proposed the Sim-vs-Real Correlation Coefficient (SRCC) to quantify how well PointNav performance in simulation predicts real-world results. We build on this idea to evaluate the sim-to-real predictivity of our Captured meshes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "A closely related work is Phone2Proc [13], which demonstrates enhanced sim-to-real ObjectNav performance by generating layouts from iPhone RoomPlan API captures. While similar to our approach in utilizing iPhone captures, our methodology differs in that we capture the entire room without focusing on layouts, nor do we perform any post-processing of scenes or generate multiple variations of the same scene for agent training. Instead, we generate meshes from our captures, fine-tune pre-trained policies for ImageNav task, and deploy them subsequently." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.822, + 0.106 + ], + "angle": 0, + "content": "2.3. 3D Scene and Mesh Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.114, + 0.907, + 0.25 + ], + "angle": 0, + "content": "In this work, we adopt DN-Splatter [49] for its simplicity and superior performance (and compare to meshes from Polycam [38], which is not open-source). DN-Splatter uses depth-and-normal regularization to improve mesh quality. While mesh reconstruction has also been explored with NeRFs [35], we choose GS for 3D reconstruction due to their fast training and rendering speeds, improving the overall efficiency of the pipeline. For an expanded list of related works on this topic, please see Appendix D.1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.262, + 0.905, + 0.278 + ], + "angle": 0, + "content": "2.4. 3D Representations for Embodied AI/Robotics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.286, + 0.908, + 0.693 + ], + "angle": 0, + "content": "There have been works which explore sim-to-real manipulation [7, 28, 52]. In this work, we attempt to solve the problem of sim-to-real navigation, laying the foundation for a diverse range of embodied tasks, such as object-goal navigation, rearrangement, and mobile manipulation, which require room-, apartment-, and building-level scene representations. Capturing high-quality 3D Gaussian Splats (GS) is relatively straightforward for tabletop manipulation tasks, whereas large-scale scene reconstruction poses greater challenges — a key aspect of this work. Recent works also explore the use of GS in solving the embodied navigation problem. GaussNav [27] presents a semantic Gaussian-based map reconstruction for the HM3D-Instance ImageNav [26] task. Splat-Nav [11] introduces a two-stage pipeline for planning and pose estimation through Gaussian Splat Maps. Tab. 1 highlights the differences between our approach and recent works utilizing GS for navigation. By capturing the entire scene using mobile phones, we enable personalization to specific scenes, as well as the potential to scale scene collection in the future. Some works [33] explore automated 3D scene capture using robot cameras in simulation, but these methods require complex, embodiment-dependent strategies that are time-consuming and difficult to scale. In contrast, our approach leverages human captures in the real-world, offering a fast and scalable solution without the need for intricate planning. For more related work on this topic, please see Appendix D.2." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.709, + 0.649, + 0.725 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The overall pipeline for bridging and integrating a real-world scene with Habitat-Sim [39] is shown in Fig. 2. In the following subsections, we discuss each stage of the pipeline in detail. Sec. 3.1 discusses the datasets used and the details of scene captures. Sec. 3.2 discusses the second stage, where these captures are converted to meshes. Sec. 3.3 discusses how ImageNav episodes are created, which is a crucial part of integrating the scene with Habitat Simulator [39]. Appendix H discusses agent training and evaluation details. For details on real-world deployment setup, refer to Appendix A." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "25433" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.092, + 0.907, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.312, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 2. The EmbodiedSplat Pipeline: Pipeline for integration of real-world captures with Habitat-Sim [39] and subsequent deployment. The first stage (a.) involves capturing the scene using Polycam [38] and Nerfstudio [48] which produces RGB frames, associated iPhone GT depth maps, and poses. In the second stage (b.), we use DN-Splatter [49] to train GS using depth and normal regularization, with normals from Metric3D-V2 [16] monocular encoder. A mesh (.ply) is created using Poisson reconstruction from the GS. In the third stage (c.), the mesh is processed and loaded into Habitat-Sim [39] for training the agent in simulation. In the last stage (d.), the policy is deployed in the real-world in the same scene for image-goal navigation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.35, + 0.437 + ], + "angle": 0, + "content": "3.1. Datasets and Scene Captures" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.446, + 0.483, + 0.551 + ], + "angle": 0, + "content": "Scene Datasets: We use two datasets for pre-training ImageNav policies, which serve as zero-shot baselines and pre-trained policies for our scenes, both in simulation and real-world evaluations. Specifically, we use the HM3D [41] dataset which consists of apartment-scale scenes split into 800 training and 100 validation scenes, and the HSSD [25] dataset split into 134 training and 33 validation scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Captured Scenes: To evaluate the feasibility of the pipeline and conduct real-world experiments, we capture scenes from a university environment (classroom, community lounges, conference rooms, etc). For custom data collection with an iPhone, we follow the procedure used for collecting the MuSHRoom dataset [42]. Specifically, we use an iPhone 13 Pro Max to record the iPhone RGB-D data using the Polycam application [38]. Polycam provides an assistive interface to ensure that all the details of the scene have been sufficiently covered during the capture. We use the default settings with Polycam and export the raw data exposed by the application. Subsequently, we use Nerf-studio [48] to process the RGB-D data and sample 1000 aligned RGB-depth frames with low blur scores and corresponding poses. Additionally, Polycam also provides a mesh with its exported data. We use this mesh for comparison purposes, referred to as POLYCAM meshes throughout the paper. Unlike Ren et al. [42], we use a manually-held iPhone for capture instead of a gimbal, enabling us to evaluate whether the process remains effective and easily replicable without relying on expensive capture equipment. Each capture requires 20-30 minutes of recording with Polycam. We repeat this process for different indoor scenes, which we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.468 + ], + "angle": 0, + "content": "refer to as lounge, classroom, conf_a, and conf_b. The scale of these of these scenes is similar to those in the MuSHRoom dataset (1-3 rooms) [42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.606 + ], + "angle": 0, + "content": "MuSHRoom Dataset: We use MuSHRoom [42] dataset for benchmarking. The dataset consists of long and short sequences of 10 indoor scenes, captured with iPhone, Kinect, and Faro Scanner. In our work, we only use the iPhone long-sequences for training and evaluation purposes. We benchmark different normal encoders on all 10 scenes (see Appendix G) for DN-Splatter [49]. For agent training and evaluation, we primarily use three scenes - honka, sauna, activity - which have varied scale and complexity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.822, + 0.634 + ], + "angle": 0, + "content": "3.2. 3D Scene and Mesh Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.779 + ], + "angle": 0, + "content": "For mesh reconstruction, we use DN-Splatter [49] as our method of choice for its superior performance on mesh reconstruction in comparison to others, and simplicity of its integration with Habitat-Sim [39]. DN-Splatter leverages depth-normal regularization, and smoothness losses to maintain geometric consistency during Gaussian Splat training. For mesh reconstruction, the rendered depth and normal maps are back-projected from training views to create a point-set for meshing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We use the default hyperparameters from DN-Splatter [49] and train the GS for 30,000 iterations. We use sensor depth \\(\\lambda_{d} = 0.2\\), and enable the depth smoothness and normal losses. Metric3D-V2 [16] is used as the normal encoder instead of Omnidata [14, 23], as empirical results demonstrate it yields higher-quality meshes. For further discussion on depth and normal encoder selection, see Appendix G." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25434" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.089, + 0.268, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.204, + 0.224, + 0.215 + ], + "angle": 0, + "content": "(a) conf.b" + }, + { + "type": "image", + "bbox": [ + 0.31, + 0.104, + 0.477, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.203, + 0.429, + 0.215 + ], + "angle": 0, + "content": "(b) conf_a" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.112, + 0.661, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.556, + 0.204, + 0.648, + 0.215 + ], + "angle": 0, + "content": "(c) classroom" + }, + { + "type": "image", + "bbox": [ + 0.731, + 0.092, + 0.897, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.204, + 0.84, + 0.215 + ], + "angle": 0, + "content": "(d) lounge" + }, + { + "type": "image_caption", + "bbox": [ + 0.15, + 0.227, + 0.846, + 0.243 + ], + "angle": 0, + "content": "Figure 3. Reconstructed meshes of the Captured scenes after DN-Splatter [49] training and Poisson reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.484, + 0.42 + ], + "angle": 0, + "content": "After training GS and converting into ply meshes, we convert the meshes to glb in Blender (see Appendix E) before loading them in Habitat-Sim [39]. The final meshes for our Captured scenes are shown in Fig. 3 and referred to as DN mesh hereafter to contrast against the POLYCAM mesh counterpart. It takes approximately 20-30 minutes per capture, and 1-2 hours of training with DN-Splatter [49] to generate these meshes, which is significantly lesser compared to the cost and several hours of capture and processing with Matterport [1] cameras." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.364, + 0.451 + ], + "angle": 0, + "content": "3.3. ImageNav Episode Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.519 + ], + "angle": 0, + "content": "Pre-training Scene Datasets: For the HM3D [41] and HSSD [25] datasets, we use the predefined train-validation split, generating 10,000 episodes for each training scene and 25 episodes for each validation scene (see Appendix B)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Captured and MuSHRoom Scenes: For scenes with DN meshes and POLYCAM meshes, we select only the largest navmesh island within each scene. We ensure that the largest island covers most, if not all, of the navigable areas in the scene. This step is crucial due to the relatively smaller sizes of these scenes compared to HM3D scenes. Hence, we generate only 1000 training episodes and 100 evaluation episodes per scene. By selecting the largest navmesh island, we minimize the risk of incorrectly treating entities other than the floor as navigable." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.764 + ], + "angle": 0, + "content": "Evaluation Metric: We consider Success Rate (SR), the fraction of successful episodes out of all the episodes, as the primary metric for performance evaluation of our policies. An episode is considered successful if the agent stops within \\(1\\mathrm{m}\\) of the goal location before the maximum number of steps (1000 for simulation, 100 for real) are over." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.782, + 0.224, + 0.799 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.483, + 0.838 + ], + "angle": 0, + "content": "This section aims to answer the following research questions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.84, + 0.483, + 0.871 + ], + "angle": 0, + "content": "1. How does a pre-trained policy perform on Captured and MuSHRoom scenes? (Sec. 4.1)" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance? (Sec. 4.2)." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.298 + ], + "angle": 0, + "content": "3. Does the performance in simulation transfer to the realworld? (Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.299, + 0.892, + 0.315 + ], + "angle": 0, + "content": "Further experiments and analyses are discussed in Sec. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.323, + 0.905, + 0.354 + ], + "angle": 0, + "content": "4.1. How does a pre-trained policy perform on Captured and MuSHRoom scenes?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.907, + 0.526 + ], + "angle": 0, + "content": "In this experiment, we evaluate the pre-trained policies that achieved the highest validation success rates on the HM3D [41] (83.08% val SR) and HSSD [25] (63.15% val SR), trained for 600M and 1200M steps, respectively. Note that the performance on HM3D is consistent with the simulation results reported in Silwal et al. [46] and represents the current state-of-the-art for this task. For agent embodiment, policy training and evaluation details, please refer to Appendix H. We assess the zero-shot generalization of these policies within simulation, evaluating their performance across different meshes for individual scenes." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.538, + 0.9, + 0.673 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.687, + 0.89, + 0.703 + ], + "angle": 0, + "content": "Figure 4. Zero-shot val. SR for HM3D pre-trained policy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Fig. 4 presents the results of the zero-shot evaluation across individual scenes for the HM3D pre-trained policy. Performance varies significantly based on scene size, complexity, and mesh type. For smaller scenes such as conf_a and conf_b, the policy demonstrates relatively high success rates in both DN mesh and POLYCAM settings, achieving \\(85\\%\\) (DN) and \\(82\\%\\) (POLYCAM) for conf_a and \\(88\\%\\) (DN mesh) and \\(79\\%\\) (POLYCAM) for conf_b." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In contrast, performance declines in larger environments such as classroom and lounge. In classroom, success rates drop to \\(53\\%\\) with DN mesh and \\(42\\%\\) with POLYCAM, while in lounge, the policy achieves \\(50\\%\\) (DN" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25435" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.091, + 0.478, + 0.103 + ], + "angle": 0, + "content": "HSSD Zero-Shot Results on Captured and MuSHRoom Scenes" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.104, + 0.47, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.241, + 0.465, + 0.256 + ], + "angle": 0, + "content": "Figure 5. Zero-shot val. SR for HSSD pre-trained policy." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.281, + 0.483, + 0.552 + ], + "angle": 0, + "content": "mesh) and \\(76\\%\\) (POLYCAM). These results highlight the challenge of transferring the policy to more complex, large-scale environments that differ from the training data. Since HM3D consists primarily of apartment-style scenes, the policy has not been exposed to classrooms or community lounges during training. Both DN and POLYCAM meshes yield similar success rates on average \\((\\sim 60\\%)\\), suggesting no clear advantage of one mesh type over the other in aligning with HM3D. Additionally, we evaluate the policies on three scenes from the MuSHRoom [42] dataset. Due to the unavailability of POLYCAM meshes for MuSHRoom, we conduct evaluations solely on DN meshes. The policy achieves a \\(95\\%\\) success rate in the sauna and honka scenes, suggesting promising potential for performance improvements through additional scene refinements (such as using a gimbal during capture). However, in the larger activity scene, performance declines, indicating challenges in generalizing to larger environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.72 + ], + "angle": 0, + "content": "Fig. 5 reveals similar trends for the HSSD pre-trained policy, though with significantly lower success rates overall. This degradation in performance can be attributed to the synthetic nature of HSSD scenes, which lack the realism and scale necessary for effective generalization. The policy performs particularly poorly on classroom, achieving only \\(1\\%\\) success on both DN mesh and POLYCAM, while performance in lounge is slightly better, reaching \\(14\\%\\) success with POLYCAM. The policy exhibits slightly improved performance on other scenes, but it is hard to conclude which mesh type best aligns with HSSD." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.759 + ], + "angle": 0, + "content": "4.2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.886 + ], + "angle": 0, + "content": "In this section, we investigate whether fine-tuning the pre-trained policies for only 20M additional steps improves performance in simulation. The learning rate is set to \\(2.5e - 6\\) for the LSTM policy and \\(6e - 7\\) for the visual encoder, following a fine-tuning strategy similar to that of Deitke et al. [13]. We fine-tune the pre-trained policy on the training episodes for a single scene and evaluate on the corresponding validation episodes." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The results are presented in Fig. 6 and Fig. 7. We ob" + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.091, + 0.901, + 0.103 + ], + "angle": 0, + "content": "HM3D Fine-tuned Results on Captured and MuSHRoom Scenes" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.104, + 0.89, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.24, + 0.905, + 0.269 + ], + "angle": 0, + "content": "Figure 6. Fine-tuned validation SR from HM3D pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.9, + 0.42 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.435, + 0.905, + 0.464 + ], + "angle": 0, + "content": "Figure 7. Fine-tuned validation SR from HSSD pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.763 + ], + "angle": 0, + "content": "serve that fine-tuning significantly improves performance across all tested scenes. For the pre-trained HM3D policy, fine-tuning for 20M steps results in success rates approaching \\(90\\%+\\) across different meshes. Similarly, for the HSSD pre-trained policy, fine-tuning leads to substantial improvements, with most policies achieving success rates of \\(80\\%+\\) in respective scenes. In particular, performance gains are particularly pronounced in larger, more complex environments such as classroom and lounge, which differ significantly from apartment-style scenes in HM3D and HSSD. These results indicate that our pipeline can effectively be used to collect diverse scene data at scale. Furthermore, they suggest that agents can be quickly finetuned (only 20M steps vs \\(100\\mathrm{M}+\\) steps for training from scratch) on these generated meshes to improve performance on specific scene types, enhancing personalization beyond the original training distribution. We present results on additional scenes in Appendix J." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.905, + 0.803 + ], + "angle": 0, + "content": "4.3. Does the performance in simulation transfer to the real-world?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We evaluate both zero-shot and fine-tuned policies in the real-world lounge scene on a Stretch robot. During the evaluation, each episode is capped at 100 steps, with 10 distinct start-and-goal locations sampled within the scene (see Appendix A for deployment details). To assess performance, we record the number of steps taken and final" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "25436" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.123 + ], + "angle": 0, + "content": "distance to the goal at the end of each episode, determining whether the agent successfully completes the task." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.136, + 0.476, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.291, + 0.485, + 0.362 + ], + "angle": 0, + "content": "Figure 8. Real world results of zero-shot and fine-tuned models on 1ounge scene. HM3D-ZS (real dataset) leads to a better success rate than HSSD-ZS (synthetic dataset). We observe increased real-world success with fine-tuning on DN and POLYCAM meshes for both HM3D and HSSD policies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.378, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Fig. 8 presents results across 10 evaluation episodes. The zero-shot HM3D policy achieves a \\(50\\%\\) success rate, demonstrating our hypothesized lack of generalization. This is in contrast with the results reported in Silwal et al. [46] showing \\(90\\%\\) zero-shot success rate in the real-world. We attribute this discrepancy to the structural and semantic differences between the lounge and the apartment-style scenes typically encountered in HM3D. Fine-tuning on the POLYCAM and DN mesh reconstructions of this scene improves performance, with success rates increasing to \\(70\\%\\). For HSSD, zero-shot performance is significantly lower at \\(10\\%\\), while fine-tuned policies improve success rates to \\(50\\%\\) with POLYCAM and \\(40\\%\\) with DN mesh." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.696 + ], + "angle": 0, + "content": "Fig. 1 illustrates the Sim-to-Real Correlation Coefficient (SRCC) [21] between simulation and real-world performance. The observation suggests that improvements in evaluation performance on DN and POLYCAM meshes in simulation translate to improved real-world performance. This demonstrates that our approach can efficiently adapt policies to novel real-world environments. We discuss real-world statistics (number of steps, distance) in Appendix I." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.31, + 0.727 + ], + "angle": 0, + "content": "5. Ablations and Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we systematically investigate critical factors influencing policy performance. First, we assess whether pre-training on large-scale datasets such as HM3D or HSSD is necessary for effective real-world transfer (Sec. 5.1). Next, we analyze the impact of validation PSNR and scene geometry on the zero-shot performance (Sec. 5.2). Finally, we examine whether continuous training on large-scale datasets improves zero-shot performance on our Captured scenes, shedding light on the relationship between dataset characteristics and generalizability (Sec. 5.3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "5.1. Is it necessary to pre-train over large-scale datasets?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.128, + 0.907, + 0.34 + ], + "angle": 0, + "content": "In this experiment, we investigate the necessity of pretraining on large-scale datasets by \"overfitting\" policies directly on POLYCAM and DN meshes with a policy trained from scratch (not pre-trained on large-scale datasets such as HM3D or HSSD) for \\(\\sim 100\\mathrm{M}\\) steps. Training and validation data are derived from ImageNav episode generation, as described in Sec. 3.3. To ensure diversity, start-goal locations for training and evaluation episodes are randomly sampled. The best validation success rates for all policies on their respective scenes are presented in Fig. 9. As expected, the overfitted policies achieve near-perfect performance in simulation, maintaining high success rates. For a detailed discussion on why overfitting results seem better than fine-tuning in this scenario, please see Appendix L." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.354, + 0.904, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.508, + 0.907, + 0.537 + ], + "angle": 0, + "content": "Figure 9. Validation SR for overfitted policies on respective scenes: Polices are not pre-trained on HM3D/HSSD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.909, + 0.75 + ], + "angle": 0, + "content": "We further evaluate these overfitted policies in the real-world lounge scene using POLYCAM and DN meshes. Surprisingly, the overfitted policy trained on the POLYCAM mesh achieves a \\(50\\%\\) success rate in real-world evaluations. In contrast, the policy trained on the DN mesh achieves only \\(10\\%\\) success. This result suggests that our mesh-based training approach can yield non-zero real-world performance, even without large-scale pre-training. We attribute the performance gap between POLYCAM and DN-trained policies to differences in visual fidelity—POLYCAM meshes preserve more visual detail by directly utilizing original images to reconstruct the scene, whereas DN meshes are based on GS which use the learned colors for the 3D Gaussians." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.758, + 0.907, + 0.79 + ], + "angle": 0, + "content": "5.2. How is the scale and validation PSNR of the scene related to the final performance?" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Fig. 10 illustrates the correlation between HM3D zero-shot validation success rates on DN meshes and two key factors: the Peak Signal-to-Noise Ratio (PSNR) of the corresponding 3D Gaussian Splats (GS) and the scale of the scene, measured as the average shortest distance between start-goal locations in validation episodes. We observe a negative correlation between success rate (SR) and average" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25437" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.093, + 0.482, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.321 + ], + "angle": 0, + "content": "Figure 10. HM3D zero-shot validation success rates on DN meshes vs. validation GS PSNRs vs. average shortest distances of validation episodes. The zero-shot SR is inversely correlated with the scale of the scene, while directly correlated with val. PSNR on the trained GS." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.349, + 0.483, + 0.501 + ], + "angle": 0, + "content": "shortest distance—indicating that as the scale of the scene increases, the zero-shot success rate declines. Conversely, a positive correlation is seen between SR and PSNR, where higher validation PSNR values correspond to improved success rates. Notably, different trend lines emerge for MuSHRoom captures and our own Captured scenes. MuSHRoom captures generally exhibit higher validation PSNRs, likely due to the use of a stabilized gimbal during data collection. Future work will further investigate the impact of capture stability on policy performance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.545 + ], + "angle": 0, + "content": "5.3. Does continuous training on large-scale datasets improve zero-shot performance?" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.566, + 0.48, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.748 + ], + "angle": 0, + "content": "Figure 11. Average zero-shot success rates on respective validation sets over different stages of pre-training on HM3D." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Fig. 11 presents average zero-shot success rates across HM3D validation set, POLYCAM meshes, and DN meshes at various stages of HM3D policy pre-training. This analysis aims to determine whether continuous performance improvements on HM3D validation scenes translate to improved zero-shot performance on our Captured scenes. We observe that while performance initially increases, it begins to deteriorate or plateau at approximately 400M steps, despite continued improvements on HM3D" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.232 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.245, + 0.907, + 0.274 + ], + "angle": 0, + "content": "Figure 12. Average zero-shot success rates on respective validation sets over different stages of pre-training on HSSD." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.301, + 0.634, + 0.314 + ], + "angle": 0, + "content": "validation scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.316, + 0.907, + 0.468 + ], + "angle": 0, + "content": "Fig. 12 shows a similar experiment conducted using the HSSD dataset. Up to 300M steps, improvements in HSSD validation performance correspond to slight improvements in zero-shot performance on our Captured scenes. However, performance plateaus for Captured, showing no further gains despite continued improvement on HSSD validation scenes. For HM3D pre-training, POLYCAM meshes outperform DN meshes in success rates, while for HSSD, the trend reverses slightly, highlighting the impact of pretraining dataset characteristics on zero-shot generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.634, + 0.497 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.765 + ], + "angle": 0, + "content": "In this work, we introduced a comprehensive pipeline for bridging the gap between real-world and simulated environments in training embodied agents using 3D Gaussian Splats (GS) and Polycam. By leveraging the MuSHRoom dataset and custom iPhone-captured scenes, we demonstrated an efficient and scalable approach to policy personalization, leveraging 3D scene reconstruction from low-effort collects, enabling high-quality training for the ImageNav task. Our pipeline facilitates accessible and replicable scene collection without requiring specialized hardware or significant costs, making it a practical solution for large-scale embodied AI research. We evaluated overfitted, zero-shot, and fine-tuned policies, showing that fine-tuning pre-trained policies on real-world scene reconstructions improves sim-to-real transfer. We also analyzed the differences between GS-generated DN meshes and POLYCAM meshes, finding that POLYCAM more closely resemble real-world scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work lays the foundation for seamlessly integrating real-world scene captures into simulation, expanding the applicability of embodied AI systems. Another promising avenue is integrating GS directly into training, replacing visual observations and goal images to enhance learning efficiency. Furthermore, we aim to extend the use of Gaussian Splats to more complex embodied AI tasks, such as rearrangement and mobile manipulation, broadening its impact across diverse real-world applications." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "25438" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.156 + ], + "angle": 0, + "content": "[1] Capture, share, and collaborate the built world in immersive 3D — matterport.com. https://matterport.com/. [Accessed 08-03-2025]. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[2] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[3] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.484, + 0.314 + ], + "angle": 0, + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.317, + 0.483, + 0.358 + ], + "angle": 0, + "content": "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[6] Guillaume Bono, Leonid Antsfeld, Boris Chidlovskii, Philippe Weinzaepfel, and Christian Wolf. End-to-end (instance)-image goal navigation through correspondence as an emergent phenomenon, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.484, + 0.499 + ], + "angle": 0, + "content": "[7] Arunkumar Byravan, Jan Humplik, Leonard Hasenclever, Arthur Brussee, Francesco Nori, Tuomas Haarnoja, Ben Moran, Steven Bohez, Fereshteh Sadeghi, Bojan Vujatovic, and Nicolas Heess. Nerf2real: Sim2real transfer of vision-guided bipedal motion skills using neural radiance fields, 2022. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.484, + 0.557 + ], + "angle": 0, + "content": "[8] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[9] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.603, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[10] Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan, Muhammad Zubair Irshad, Richard Cheng, Kurt Keutzer, Masayoshi Tomizuka, Quan Vuong, and Ken Goldberg. Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning. In Conference on Robot Learning (CoRL), Munich, Germany, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[11] Timothy Chen, Ola Shorinwa, Joseph Bruno, Javier Yu, Wei-jia Zeng, Keiko Nagami, Philip Dames, and Mac Schwager. Splat-nav: Safe real-time robot navigation in gaussian splatting maps, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[13] Matt Deitke, Rose Hendrix, Luca Weihs, Ali Farhadi, Kiana Ehsani, and Aniruddha Kembhavi. Phone2proc: Bringing robust robots into our chaotic world, 2022. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[16] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. 2024. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[17] Muhammad Zubair Irshad, Chih-Yao Ma, and Zsolt Kira. Hierarchical cross-modal agent for robotics vision-and-language navigation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[18] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4065-4071, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[19] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[20] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv:2410.20220, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[21] Abhishek Kadian, Joanne Truong, Aaron Gokaslan, Alexander Clegg, Erik Wijmans, Stefan Lee, Manolis Savva, Sonia Chernova, and Dhruv Batra. Sim2real predictivity: Does evaluation in simulation predict real-world performance? IEEE Robotics and Automation Letters, 5(4):6670-6677, 2020. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[22] Ivan Kapelyukh, Yifei Ren, Ignacio Alzugaray, and Edward Johns. Dream2real: Zero-shot 3d object rearrangement with vision-language models, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.659 + ], + "angle": 0, + "content": "[23] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.718, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[25] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Schacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. arXiv preprint arXiv:2306.11290, 2023. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[26] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[27] Xiaohan Lei, Min Wang, Wengang Zhou, and Houqiang Li. Gaussnav: Gaussian splatting for visual navigation, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25439" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[28] Yulong Li and Deepak Pathak. Object-aware gaussian splatt ing for robotic manipulation. In ICRA 2024 Workshop on 3D Visual Representations for Robot Manipulation, 2024. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.178 + ], + "angle": 0, + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.223 + ], + "angle": 0, + "content": "[30] JunEn Low, Maximilian Adang, Javier Yu, Keiko Nagami, and Mac Schwager. Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.225, + 0.482, + 0.267 + ], + "angle": 0, + "content": "[31] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.269, + 0.482, + 0.365 + ], + "angle": 0, + "content": "[32] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, Dhruv Batra, Yixin Lin, Oleksandr Maksymets, Aravind Rajeswaran, and Franziska Meier. Where are we in the search for an artificial visual cortex for embodied intelligence?, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.369, + 0.482, + 0.423 + ], + "angle": 0, + "content": "[33] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra Singh Chaplot. Autonerf: Training implicit scene representations with autonomous agents, 2023. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[34] Wugang Meng, Tianfu Wu, Huan Yin, and Fumin Zhang. Beings: Bayesian embodied image-goal navigation with gaussian splatting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.528, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.586, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.428, + 0.672 + ], + "angle": 0, + "content": "[38] Polycam. Polycam. Accessed: 2024-11-10. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.674, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[39] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. In The Twelfth International Conference on Learning Representations, 2023. 2, 3, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[40] Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun, David Held, George Kantor, and Abhisesh Silwal. Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[41] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai, 2021. 1, 2, 3, 4, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[42] Xuqian Ren, Wenjia Wang, Dingding Cai, Tuuli Tuominen, Juho Kannala, and Esa Rahtu. Mushroom: Multi-sensor hybrid room dataset for joint 3d reconstruction and novel view synthesis, 2024. 3, 4, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[43] Meta AI Research. Home robot. https://github.com/facebookresearch/home-robot, 2023. GitHub repository. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[46] Sneha Silwal, Karmesh Yadav, Tingfan Wu, Jay Vakil, Arjun Majumdar, Sergio Arnaud, Claire Chen, Vincent-Pierre Berges, Dhruv Batra, Aravind Rajeswaran, Mrinal Kalakrishnan, Franziska Meier, and Oleksandr Maksymets. What do we learn from a large-scale study of pre-trained visual representations in sim and real environments?, 2024. 1, 2, 3, 5, 7, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.522 + ], + "angle": 0, + "content": "[47] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.524, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[48] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.608, + 0.905, + 0.662 + ], + "angle": 0, + "content": "[49] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splattering and meshing, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[51] Yaniv Wolf, Amit Bracha, and Ron Kimmel. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[52] Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, and Hesheng Wang. Rl-gsbridge: 3d gaussian splating based real2sim2real method for robotic manipulation learning, 2024. 3, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[53] Fei Xia, Amir Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[54] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "25440" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "browser-compatible environment from a single video, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[55] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[56] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[58] Justin Yu, Kush Hari, Kishore Srinivas, Karim El-Refai, Adam Rashid, Chung Min Kim, Justin Kerr, Richard Cheng, Muhammad Zubair Irshad, Ashwin Balakrishna, et al. Language-embedded gaussian splats (legs): incrementally building room-scale representations with a mobile robot. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13326–13332. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[59] Justin Yu, Kush Hari, Karim El-Refai, Arnav Dalil, Justin Kerr, Chung-Min Kim, Richard Cheng, Muhammad Z. Irshad, and Ken Goldberg. Persistent object gaussian splat (pogs) for tracking human and robot manipulation of irregularly shaped objects. ICRA, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[60] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "25441" + } + ] +] \ No newline at end of file diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_origin.pdf b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5b1bcfa8b6ed70bb08e9006487062ee24379b976 --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/eb28cbcd-1806-4f8d-ad73-ef4f7f494673_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2d27dae3c25ee787f4b88a63dec56824b9ad0ee2938718d04619cdd3cf1b5f26 +size 7873921 diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/full.md b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9fa43f0934847d1e63d0f618f7ae920b5179b4e2 --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/full.md @@ -0,0 +1,283 @@ +# EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device + +Gunjan Chhablani $^{1*}$ , Xiaomeng Ye $^{2}$ , Muhammad Zubair Irshad $^{3}$ , Zsolt Kira $^{4}$ , $^{1,2,4}$ Georgia Tech, $^{3}$ Toyota Research Institute + +{chhablani.gunjan, muhammadzubairirshad}@gmail.com, {xye87, zkira}@gatech.edu + +# Abstract + +The field of Embodied AI predominantly relies on simulation for training and evaluation, often using either fully synthetic environments that lack photorealism or high-fidelity real-world reconstructions captured with expensive hardware. As a result, sim-to-real transfer remains a major challenge. In this paper, we introduce EmbodiedSplat, a novel approach that personalizes policy training by efficiently capturing the deployment environment and fine-tuning policies within the reconstructed scenes. Our method leverages 3D Gaussian Splatting (GS) and the Habitat-Sim simulator to bridge the gap between realistic scene capture and effective training environments. Using iPhone-captured deployment scenes, we reconstruct meshes via GS, enabling training in settings that closely approximate real-world conditions. We conduct a comprehensive analysis of training strategies, pre-training datasets, and mesh reconstruction techniques, evaluating their impact on sim-to-real predictivity in real-world scenarios. Experimental results demonstrate that agents fine-tuned with EmbodiedSplat outperform both zero-shot baselines pre-trained on large-scale real-world datasets (HM3D) and synthetically generated datasets (HSSD), achieving absolute success rate improvements of $20\%$ and $40\%$ on real-world Image Navigation task. Moreover, our approach yields a high sim-vs-real correlation (0.87-0.97) for the reconstructed meshes, underscoring its effectiveness in adapting policies to diverse environments with minimal effort. Project page: https://gchhablani.github.io/embodied-splat. + +# 1. Introduction + +Recent advancements in Embodied AI have demonstrated impressive performance in simulated environments [17, 18, 32, 44, 46]. However, translating these capabilities to physical robots remains a significant challenge, primarily due to limitations in simulation fidelity and accessibility [21]. A + +![](images/e7f16acf6b415ccc72ebda8492ade9cc1686285b2509ad2f81a0563857922492.jpg) + +![](images/8c7726c6f9d69e6904e81b6732783c27b6f3832e4f0bbee223deeb1d783a3554.jpg) +Figure 1. Overview of EmbodiedSplat: Mobile phone captures are used to generate reconstructed meshes via 3D Gaussian Splating (GS). Agents are trained within these reconstructed environments in simulation before being deployed in the real world, enabling effective sim-to-real transfer. Our analysis demonstrates a strong sim-to-real correlation across both types of generated meshes, highlighting their ability to bridge the gap between simulation and real-world performance. + +key bottleneck is the sim-to-real gap, where handcrafted or synthetic simulation environments (e.g. HSSD [25]) struggle to capture the complexity and variability of real-world settings, necessitating the use of real-world reconstructions for effective policy training. On the other hand, real-world datasets such as Matterport3D [8] and HM3D [41] rely on expensive capture equipment and labor-intensive reconstruction pipelines, making large-scale scene collection and adaptation impractical for many applications. Furthermore, it is difficult to fully capture the variability of potential deployment environments with these datasets. + +Developments in 3D scene representations, particularly 3D Gaussian Splatting (GS) [24], have shown promise to + +
WorkGoal TypeE2ERealS2R
SplatNav [11]Point/Language×✓×
GaussNav [27]Image [26]×××
OursImage✓✓✓
+ +Table 1. Comparison against recent works using GS for navigation. "E2E" indicates whether the policy is trained end-to-end, "Real" denotes evaluation on real-world robots, and "S2R" indicates demonstrated sim-to-real transfer. + +wards reducing the effort needed to capture new scenes, enabling high-quality scene reconstruction from casual mobile phone captures. These approaches offer a strong ability to handle complex geometry, perform novel-view view synthesis, and provide high visual fidelity. Building upon this foundation, methods like DN-Splatter [49] have further enhanced mesh reconstruction quality through depth-and-normal regularization. However, their potential for training robot navigation policies and deploying them in the real-world has remained largely unexplored. + +In this work, we explore the following question: Can low-effort cellphone video captures be leveraged to generate meshes that facilitate the training of Embodied AI navigation policies for effective transfer to the target environment? In other words, we seek to personalize models by training them directly on the target distribution, as represented by 3D models built from low-cost and low-effort data collects of the deployment environment. Some recent works indeed explore GS for navigation - GaussNav [11, 27] explores Instance-Image Navigation [26] in simulation using GS; SplatNav [11] explores using GS-based collision meshes for collision avoidance and localization in a single real-world environment using a modular policy for a drone. In contrast, as shown in Tab. 1, we explore end-to-end policy learning for image-goal navigation and sim-to-real transfer in indoor environments. To the best of our knowledge, we are the first to explore an approach to solve the personalized real-to-sim-to-real problem using Gaussian Splats for indoor image-goal navigation. + +To achieve the above goal, we present a comprehensive framework for leveraging open-source 3D Gaussian Splating [24] (and compare with Polycam [38]). The central premise of our work is that it is possible to quickly capture the scenes in which a robot will be deployed, using readily available consumer-grade hardware, and seamlessly integrate them into simulation environments (i.e. Habitat-Sim [39]). Policies can then be trained in these simulated environments, leading to improved sim-to-real transfer. Our approach combines the accessibility of smartphone-based scene capture with recent advances in depth-aware 3D scene representation, enabling rapid training and deployment of navigation policies in new, realistic environments. + +However, there are several challenges and open ques + +tions, ranging from reconstruction quality to finetuning strategies, to enabling successful transfer. Building on the work of Silwal et al. [46] as a baseline, we test our framework across a range of scenes captured in a university environment, which is out-of-distribution for typical pre-trained policies. We analyze factors affecting transfer performance, including trade-offs between mesh generation pipelines, mesh quality for policy training, and training strategies (e.g., zero-shot vs. fine-tuning). Through rigorous evaluation and real-world robot experiments, we show that our approach yields significant gains in real-world image-goal navigation. + +Our key contributions can be summarized as follows: + +1. An efficient and cost-effective pipeline for bridging the real-to-sim gap in navigation, enabling the creation of high-quality simulation scenes from low-cost, consumer-grade iPhone captures using depth-aware 3D Gaussian Splats (GS) and Polycam. +2. Comprehensive evaluations conducted in both simulation and real-world environments, assessing zero-shot and fine-tuned policies on our Captured scenes. Our results demonstrate substantial improvements in sim-to-real transfer, emphasizing the effectiveness of fine-tuning on high-fidelity reconstructions. +3. In-depth analysis addressing key research questions, exploring the relationship between reconstruction quality, pre-training scenes, downstream navigation performance, and training strategies. For example, a notable finding is that overfitting policies on a single high-fidelity scene reconstruction in simulation yields reasonable real-world performance. Our findings provide valuable insights into how reconstruction fidelity influences policy generalization and its applicability to real-world scenarios. +4. An open-source codebase and dataset to facilitate further research in this domain and reproducibility of results. + +Through this work, we aim to make high-quality scene collection and agent training more accessible, as well as enable easy development of personalized agents. + +# 2. Related Work + +# 2.1. Scene Datasets + +Recent years have seen rapid progress in the development of high-quality 3D scene datasets for embodied AI research. The Matterport3D dataset [8] provides 10,800 panoramic views from 194,400 RGB-D images across 90 buildings-scale scenes, complete with surface reconstructions and semantic annotations. Building upon this, HM3D [41] expanded the scale to 1,000 building-scale 3D reconstructions from diverse real-world locations, though some issues with mesh quality were noted. The Gibson dataset [53] introduced another collection of real-world scans, while + +synthetic datasets like HSSD [25] (with 211 realistic environments) demonstrated that smaller but higher-quality datasets can sometimes be better than larger ones for agent training. Another dataset that is commonly used is ReplicaCAD [47] which provides synthetic variations in layouts of a single scene. In this work, we use HM3D [41] for training our zero-shot baseline, as it is one of the largest photorealistic indoor-scene datasets captured using Matterport [1] cameras. Additionally, we use HSSD [25] as the synthetic counterpart for another zero-shot baseline. Note, however, that these scenes do have a bias towards apartments, and we therefore test our methods on scenes within a university environment, which is out-of-distribution for these datasets. The MuSHRoom dataset [42] provides multi-sensor captures of 10 real-world scenes, offering benchmarks for reconstruction and novel view synthesis methods. While this dataset has not been used for embodied tasks in prior works, our dataset collection method is inspired by the MuSHRoom dataset's iPhone captures. We evaluate our agents, as well as reconstruction strategies, on these scenes to benchmark our depth and normal encoders. + +# 2.2. Embodied Navigation in Indoor Environments + +Embodied navigation encompasses various forms of goal-directed navigation tasks [2, 6, 9, 26, 31, 50, 55, 56]. This study primarily investigates ImageNav [2, 55, 56] in indoor environments. Recent advancements in image-goal navigation have explored a variety of approaches to enhance performance and generalization. Notably, pre-trained visual representations [32, 46, 55, 56] have shown substantial promise in improving performance on the ImageNav task. In particular, Silwal et al. [46] demonstrate remarkable zero-shot success rates (90%) on real-world ImageNav, leveraging VC-1 [32], fine-tuned end-to-end with data augmentation. Inspired by this, we use the setup in Silwal et al. [46] for our policy training and evaluation. + +Sim-to-real transfer remains a core challenge in Embodied AI beyond leveraging pre-trained visual encoders. Kadian et al. [21] proposed the Sim-vs-Real Correlation Coefficient (SRCC) to quantify how well PointNav performance in simulation predicts real-world results. We build on this idea to evaluate the sim-to-real predictivity of our Captured meshes. + +A closely related work is Phone2Proc [13], which demonstrates enhanced sim-to-real ObjectNav performance by generating layouts from iPhone RoomPlan API captures. While similar to our approach in utilizing iPhone captures, our methodology differs in that we capture the entire room without focusing on layouts, nor do we perform any post-processing of scenes or generate multiple variations of the same scene for agent training. Instead, we generate meshes from our captures, fine-tune pre-trained policies for ImageNav task, and deploy them subsequently. + +# 2.3. 3D Scene and Mesh Reconstruction + +In this work, we adopt DN-Splatter [49] for its simplicity and superior performance (and compare to meshes from Polycam [38], which is not open-source). DN-Splatter uses depth-and-normal regularization to improve mesh quality. While mesh reconstruction has also been explored with NeRFs [35], we choose GS for 3D reconstruction due to their fast training and rendering speeds, improving the overall efficiency of the pipeline. For an expanded list of related works on this topic, please see Appendix D.1. + +# 2.4. 3D Representations for Embodied AI/Robotics + +There have been works which explore sim-to-real manipulation [7, 28, 52]. In this work, we attempt to solve the problem of sim-to-real navigation, laying the foundation for a diverse range of embodied tasks, such as object-goal navigation, rearrangement, and mobile manipulation, which require room-, apartment-, and building-level scene representations. Capturing high-quality 3D Gaussian Splats (GS) is relatively straightforward for tabletop manipulation tasks, whereas large-scale scene reconstruction poses greater challenges — a key aspect of this work. Recent works also explore the use of GS in solving the embodied navigation problem. GaussNav [27] presents a semantic Gaussian-based map reconstruction for the HM3D-Instance ImageNav [26] task. Splat-Nav [11] introduces a two-stage pipeline for planning and pose estimation through Gaussian Splat Maps. Tab. 1 highlights the differences between our approach and recent works utilizing GS for navigation. By capturing the entire scene using mobile phones, we enable personalization to specific scenes, as well as the potential to scale scene collection in the future. Some works [33] explore automated 3D scene capture using robot cameras in simulation, but these methods require complex, embodiment-dependent strategies that are time-consuming and difficult to scale. In contrast, our approach leverages human captures in the real-world, offering a fast and scalable solution without the need for intricate planning. For more related work on this topic, please see Appendix D.2. + +# 3. Methodology + +The overall pipeline for bridging and integrating a real-world scene with Habitat-Sim [39] is shown in Fig. 2. In the following subsections, we discuss each stage of the pipeline in detail. Sec. 3.1 discusses the datasets used and the details of scene captures. Sec. 3.2 discusses the second stage, where these captures are converted to meshes. Sec. 3.3 discusses how ImageNav episodes are created, which is a crucial part of integrating the scene with Habitat Simulator [39]. Appendix H discusses agent training and evaluation details. For details on real-world deployment setup, refer to Appendix A. + +![](images/6002c9aa5b763a45cccefc9b1c847b2dec73bd4c8faa6fb56ed50c24a0f8cb27.jpg) +Figure 2. The EmbodiedSplat Pipeline: Pipeline for integration of real-world captures with Habitat-Sim [39] and subsequent deployment. The first stage (a.) involves capturing the scene using Polycam [38] and Nerfstudio [48] which produces RGB frames, associated iPhone GT depth maps, and poses. In the second stage (b.), we use DN-Splatter [49] to train GS using depth and normal regularization, with normals from Metric3D-V2 [16] monocular encoder. A mesh (.ply) is created using Poisson reconstruction from the GS. In the third stage (c.), the mesh is processed and loaded into Habitat-Sim [39] for training the agent in simulation. In the last stage (d.), the policy is deployed in the real-world in the same scene for image-goal navigation. + +# 3.1. Datasets and Scene Captures + +Scene Datasets: We use two datasets for pre-training ImageNav policies, which serve as zero-shot baselines and pre-trained policies for our scenes, both in simulation and real-world evaluations. Specifically, we use the HM3D [41] dataset which consists of apartment-scale scenes split into 800 training and 100 validation scenes, and the HSSD [25] dataset split into 134 training and 33 validation scenes. + +Captured Scenes: To evaluate the feasibility of the pipeline and conduct real-world experiments, we capture scenes from a university environment (classroom, community lounges, conference rooms, etc). For custom data collection with an iPhone, we follow the procedure used for collecting the MuSHRoom dataset [42]. Specifically, we use an iPhone 13 Pro Max to record the iPhone RGB-D data using the Polycam application [38]. Polycam provides an assistive interface to ensure that all the details of the scene have been sufficiently covered during the capture. We use the default settings with Polycam and export the raw data exposed by the application. Subsequently, we use Nerf-studio [48] to process the RGB-D data and sample 1000 aligned RGB-depth frames with low blur scores and corresponding poses. Additionally, Polycam also provides a mesh with its exported data. We use this mesh for comparison purposes, referred to as POLYCAM meshes throughout the paper. Unlike Ren et al. [42], we use a manually-held iPhone for capture instead of a gimbal, enabling us to evaluate whether the process remains effective and easily replicable without relying on expensive capture equipment. Each capture requires 20-30 minutes of recording with Polycam. We repeat this process for different indoor scenes, which we + +refer to as lounge, classroom, conf_a, and conf_b. The scale of these of these scenes is similar to those in the MuSHRoom dataset (1-3 rooms) [42]. + +MuSHRoom Dataset: We use MuSHRoom [42] dataset for benchmarking. The dataset consists of long and short sequences of 10 indoor scenes, captured with iPhone, Kinect, and Faro Scanner. In our work, we only use the iPhone long-sequences for training and evaluation purposes. We benchmark different normal encoders on all 10 scenes (see Appendix G) for DN-Splatter [49]. For agent training and evaluation, we primarily use three scenes - honka, sauna, activity - which have varied scale and complexity. + +# 3.2. 3D Scene and Mesh Reconstruction + +For mesh reconstruction, we use DN-Splatter [49] as our method of choice for its superior performance on mesh reconstruction in comparison to others, and simplicity of its integration with Habitat-Sim [39]. DN-Splatter leverages depth-normal regularization, and smoothness losses to maintain geometric consistency during Gaussian Splat training. For mesh reconstruction, the rendered depth and normal maps are back-projected from training views to create a point-set for meshing. + +We use the default hyperparameters from DN-Splatter [49] and train the GS for 30,000 iterations. We use sensor depth $\lambda_{d} = 0.2$ , and enable the depth smoothness and normal losses. Metric3D-V2 [16] is used as the normal encoder instead of Omnidata [14, 23], as empirical results demonstrate it yields higher-quality meshes. For further discussion on depth and normal encoder selection, see Appendix G. + +![](images/5740ae30a9454782880445ce517f7542d8303fbc3c6cb1be7b77fdf50b594d6c.jpg) +(a) conf.b + +![](images/99562dac0ade027ceca080b7985028af6c15ec938a6c6ab5ce7f0b7746a1f523.jpg) +(b) conf_a + +![](images/2273cddcecf21365cdc2119b8471557da8c581508fb16589a526f288bf99b219.jpg) +(c) classroom +Figure 3. Reconstructed meshes of the Captured scenes after DN-Splatter [49] training and Poisson reconstruction. + +![](images/4ba08cb9ec672ff0d6dc02abb2166cbfb24d9c941ecc5d1ab7b4a8469c63f0ea.jpg) +(d) lounge + +After training GS and converting into ply meshes, we convert the meshes to glb in Blender (see Appendix E) before loading them in Habitat-Sim [39]. The final meshes for our Captured scenes are shown in Fig. 3 and referred to as DN mesh hereafter to contrast against the POLYCAM mesh counterpart. It takes approximately 20-30 minutes per capture, and 1-2 hours of training with DN-Splatter [49] to generate these meshes, which is significantly lesser compared to the cost and several hours of capture and processing with Matterport [1] cameras. + +# 3.3. ImageNav Episode Generation + +Pre-training Scene Datasets: For the HM3D [41] and HSSD [25] datasets, we use the predefined train-validation split, generating 10,000 episodes for each training scene and 25 episodes for each validation scene (see Appendix B). + +Captured and MuSHRoom Scenes: For scenes with DN meshes and POLYCAM meshes, we select only the largest navmesh island within each scene. We ensure that the largest island covers most, if not all, of the navigable areas in the scene. This step is crucial due to the relatively smaller sizes of these scenes compared to HM3D scenes. Hence, we generate only 1000 training episodes and 100 evaluation episodes per scene. By selecting the largest navmesh island, we minimize the risk of incorrectly treating entities other than the floor as navigable. + +Evaluation Metric: We consider Success Rate (SR), the fraction of successful episodes out of all the episodes, as the primary metric for performance evaluation of our policies. An episode is considered successful if the agent stops within $1\mathrm{m}$ of the goal location before the maximum number of steps (1000 for simulation, 100 for real) are over. + +# 4. Experiments + +This section aims to answer the following research questions: + +1. How does a pre-trained policy perform on Captured and MuSHRoom scenes? (Sec. 4.1) +2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance? (Sec. 4.2). + +3. Does the performance in simulation transfer to the realworld? (Sec. 4.3). + +Further experiments and analyses are discussed in Sec. 5. + +# 4.1. How does a pre-trained policy perform on Captured and MuSHRoom scenes? + +In this experiment, we evaluate the pre-trained policies that achieved the highest validation success rates on the HM3D [41] (83.08% val SR) and HSSD [25] (63.15% val SR), trained for 600M and 1200M steps, respectively. Note that the performance on HM3D is consistent with the simulation results reported in Silwal et al. [46] and represents the current state-of-the-art for this task. For agent embodiment, policy training and evaluation details, please refer to Appendix H. We assess the zero-shot generalization of these policies within simulation, evaluating their performance across different meshes for individual scenes. + +![](images/39d0b3ff99eade48b144ed040e17c4c787b65e637b9ec608ad8d786bd42002b5.jpg) +Figure 4. Zero-shot val. SR for HM3D pre-trained policy. + +Fig. 4 presents the results of the zero-shot evaluation across individual scenes for the HM3D pre-trained policy. Performance varies significantly based on scene size, complexity, and mesh type. For smaller scenes such as conf_a and conf_b, the policy demonstrates relatively high success rates in both DN mesh and POLYCAM settings, achieving $85\%$ (DN) and $82\%$ (POLYCAM) for conf_a and $88\%$ (DN mesh) and $79\%$ (POLYCAM) for conf_b. + +In contrast, performance declines in larger environments such as classroom and lounge. In classroom, success rates drop to $53\%$ with DN mesh and $42\%$ with POLYCAM, while in lounge, the policy achieves $50\%$ (DN + +![](images/0cecb7120ff12bf0c11d540300add3e767fab168f86faedf352804101c4e67e1.jpg) +HSSD Zero-Shot Results on Captured and MuSHRoom Scenes + +mesh) and $76\%$ (POLYCAM). These results highlight the challenge of transferring the policy to more complex, large-scale environments that differ from the training data. Since HM3D consists primarily of apartment-style scenes, the policy has not been exposed to classrooms or community lounges during training. Both DN and POLYCAM meshes yield similar success rates on average $(\sim 60\%)$ , suggesting no clear advantage of one mesh type over the other in aligning with HM3D. Additionally, we evaluate the policies on three scenes from the MuSHRoom [42] dataset. Due to the unavailability of POLYCAM meshes for MuSHRoom, we conduct evaluations solely on DN meshes. The policy achieves a $95\%$ success rate in the sauna and honka scenes, suggesting promising potential for performance improvements through additional scene refinements (such as using a gimbal during capture). However, in the larger activity scene, performance declines, indicating challenges in generalizing to larger environments. + +Fig. 5 reveals similar trends for the HSSD pre-trained policy, though with significantly lower success rates overall. This degradation in performance can be attributed to the synthetic nature of HSSD scenes, which lack the realism and scale necessary for effective generalization. The policy performs particularly poorly on classroom, achieving only $1\%$ success on both DN mesh and POLYCAM, while performance in lounge is slightly better, reaching $14\%$ success with POLYCAM. The policy exhibits slightly improved performance on other scenes, but it is hard to conclude which mesh type best aligns with HSSD. + +# 4.2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance? + +In this section, we investigate whether fine-tuning the pre-trained policies for only 20M additional steps improves performance in simulation. The learning rate is set to $2.5e - 6$ for the LSTM policy and $6e - 7$ for the visual encoder, following a fine-tuning strategy similar to that of Deitke et al. [13]. We fine-tune the pre-trained policy on the training episodes for a single scene and evaluate on the corresponding validation episodes. + +The results are presented in Fig. 6 and Fig. 7. We ob + +![](images/601aeea520cda99e5bc897510d426123c502cb35e23164404e378baf52aa3931.jpg) +Figure 5. Zero-shot val. SR for HSSD pre-trained policy. +HM3D Fine-tuned Results on Captured and MuSHRoom Scenes +Figure 6. Fine-tuned validation SR from HM3D pre-trained policy: The policy is fine-tuned and evaluated on the same mesh. + +![](images/219958e3f0fb2eab600e9c11125a03294961fcd86030cbd276ef00ddb761ca8c.jpg) +Figure 7. Fine-tuned validation SR from HSSD pre-trained policy: The policy is fine-tuned and evaluated on the same mesh. + +serve that fine-tuning significantly improves performance across all tested scenes. For the pre-trained HM3D policy, fine-tuning for 20M steps results in success rates approaching $90\%+$ across different meshes. Similarly, for the HSSD pre-trained policy, fine-tuning leads to substantial improvements, with most policies achieving success rates of $80\%+$ in respective scenes. In particular, performance gains are particularly pronounced in larger, more complex environments such as classroom and lounge, which differ significantly from apartment-style scenes in HM3D and HSSD. These results indicate that our pipeline can effectively be used to collect diverse scene data at scale. Furthermore, they suggest that agents can be quickly finetuned (only 20M steps vs $100\mathrm{M}+$ steps for training from scratch) on these generated meshes to improve performance on specific scene types, enhancing personalization beyond the original training distribution. We present results on additional scenes in Appendix J. + +# 4.3. Does the performance in simulation transfer to the real-world? + +We evaluate both zero-shot and fine-tuned policies in the real-world lounge scene on a Stretch robot. During the evaluation, each episode is capped at 100 steps, with 10 distinct start-and-goal locations sampled within the scene (see Appendix A for deployment details). To assess performance, we record the number of steps taken and final + +distance to the goal at the end of each episode, determining whether the agent successfully completes the task. + +![](images/20c70cec609ca492acba284f0e3cca7385aeda0d60c1bf9887857a2748d04df4.jpg) +Figure 8. Real world results of zero-shot and fine-tuned models on 1ounge scene. HM3D-ZS (real dataset) leads to a better success rate than HSSD-ZS (synthetic dataset). We observe increased real-world success with fine-tuning on DN and POLYCAM meshes for both HM3D and HSSD policies. + +Fig. 8 presents results across 10 evaluation episodes. The zero-shot HM3D policy achieves a $50\%$ success rate, demonstrating our hypothesized lack of generalization. This is in contrast with the results reported in Silwal et al. [46] showing $90\%$ zero-shot success rate in the real-world. We attribute this discrepancy to the structural and semantic differences between the lounge and the apartment-style scenes typically encountered in HM3D. Fine-tuning on the POLYCAM and DN mesh reconstructions of this scene improves performance, with success rates increasing to $70\%$ . For HSSD, zero-shot performance is significantly lower at $10\%$ , while fine-tuned policies improve success rates to $50\%$ with POLYCAM and $40\%$ with DN mesh. + +Fig. 1 illustrates the Sim-to-Real Correlation Coefficient (SRCC) [21] between simulation and real-world performance. The observation suggests that improvements in evaluation performance on DN and POLYCAM meshes in simulation translate to improved real-world performance. This demonstrates that our approach can efficiently adapt policies to novel real-world environments. We discuss real-world statistics (number of steps, distance) in Appendix I. + +# 5. Ablations and Analysis + +In this section, we systematically investigate critical factors influencing policy performance. First, we assess whether pre-training on large-scale datasets such as HM3D or HSSD is necessary for effective real-world transfer (Sec. 5.1). Next, we analyze the impact of validation PSNR and scene geometry on the zero-shot performance (Sec. 5.2). Finally, we examine whether continuous training on large-scale datasets improves zero-shot performance on our Captured scenes, shedding light on the relationship between dataset characteristics and generalizability (Sec. 5.3). + +# 5.1. Is it necessary to pre-train over large-scale datasets? + +In this experiment, we investigate the necessity of pretraining on large-scale datasets by "overfitting" policies directly on POLYCAM and DN meshes with a policy trained from scratch (not pre-trained on large-scale datasets such as HM3D or HSSD) for $\sim 100\mathrm{M}$ steps. Training and validation data are derived from ImageNav episode generation, as described in Sec. 3.3. To ensure diversity, start-goal locations for training and evaluation episodes are randomly sampled. The best validation success rates for all policies on their respective scenes are presented in Fig. 9. As expected, the overfitted policies achieve near-perfect performance in simulation, maintaining high success rates. For a detailed discussion on why overfitting results seem better than fine-tuning in this scenario, please see Appendix L. + +![](images/c837f32cfba2b5daf80ef7e04e8461909eff2a64da2a5af9ae3435a4b2ce8380.jpg) +Figure 9. Validation SR for overfitted policies on respective scenes: Polices are not pre-trained on HM3D/HSSD. + +We further evaluate these overfitted policies in the real-world lounge scene using POLYCAM and DN meshes. Surprisingly, the overfitted policy trained on the POLYCAM mesh achieves a $50\%$ success rate in real-world evaluations. In contrast, the policy trained on the DN mesh achieves only $10\%$ success. This result suggests that our mesh-based training approach can yield non-zero real-world performance, even without large-scale pre-training. We attribute the performance gap between POLYCAM and DN-trained policies to differences in visual fidelity—POLYCAM meshes preserve more visual detail by directly utilizing original images to reconstruct the scene, whereas DN meshes are based on GS which use the learned colors for the 3D Gaussians. + +# 5.2. How is the scale and validation PSNR of the scene related to the final performance? + +Fig. 10 illustrates the correlation between HM3D zero-shot validation success rates on DN meshes and two key factors: the Peak Signal-to-Noise Ratio (PSNR) of the corresponding 3D Gaussian Splats (GS) and the scale of the scene, measured as the average shortest distance between start-goal locations in validation episodes. We observe a negative correlation between success rate (SR) and average + +![](images/01c3bf71b84b52d8d26c633cc7033e0ab77fa9b82b6d0dd85ce8527568927565.jpg) +Figure 10. HM3D zero-shot validation success rates on DN meshes vs. validation GS PSNRs vs. average shortest distances of validation episodes. The zero-shot SR is inversely correlated with the scale of the scene, while directly correlated with val. PSNR on the trained GS. + +shortest distance—indicating that as the scale of the scene increases, the zero-shot success rate declines. Conversely, a positive correlation is seen between SR and PSNR, where higher validation PSNR values correspond to improved success rates. Notably, different trend lines emerge for MuSHRoom captures and our own Captured scenes. MuSHRoom captures generally exhibit higher validation PSNRs, likely due to the use of a stabilized gimbal during data collection. Future work will further investigate the impact of capture stability on policy performance. + +# 5.3. Does continuous training on large-scale datasets improve zero-shot performance? + +![](images/3e06aafad43397d2acb1bc4aee357c3c3b59607cdbbc99007fd5228834aaec64.jpg) +Figure 11. Average zero-shot success rates on respective validation sets over different stages of pre-training on HM3D. + +Fig. 11 presents average zero-shot success rates across HM3D validation set, POLYCAM meshes, and DN meshes at various stages of HM3D policy pre-training. This analysis aims to determine whether continuous performance improvements on HM3D validation scenes translate to improved zero-shot performance on our Captured scenes. We observe that while performance initially increases, it begins to deteriorate or plateau at approximately 400M steps, despite continued improvements on HM3D + +![](images/d2bad78c6e6af564e8a4d7a8097e0f59851c0beda5062ce327721f60df0c5b12.jpg) +Figure 12. Average zero-shot success rates on respective validation sets over different stages of pre-training on HSSD. + +validation scenes. + +Fig. 12 shows a similar experiment conducted using the HSSD dataset. Up to 300M steps, improvements in HSSD validation performance correspond to slight improvements in zero-shot performance on our Captured scenes. However, performance plateaus for Captured, showing no further gains despite continued improvement on HSSD validation scenes. For HM3D pre-training, POLYCAM meshes outperform DN meshes in success rates, while for HSSD, the trend reverses slightly, highlighting the impact of pretraining dataset characteristics on zero-shot generalization. + +# 6. Conclusion + +In this work, we introduced a comprehensive pipeline for bridging the gap between real-world and simulated environments in training embodied agents using 3D Gaussian Splats (GS) and Polycam. By leveraging the MuSHRoom dataset and custom iPhone-captured scenes, we demonstrated an efficient and scalable approach to policy personalization, leveraging 3D scene reconstruction from low-effort collects, enabling high-quality training for the ImageNav task. Our pipeline facilitates accessible and replicable scene collection without requiring specialized hardware or significant costs, making it a practical solution for large-scale embodied AI research. We evaluated overfitted, zero-shot, and fine-tuned policies, showing that fine-tuning pre-trained policies on real-world scene reconstructions improves sim-to-real transfer. We also analyzed the differences between GS-generated DN meshes and POLYCAM meshes, finding that POLYCAM more closely resemble real-world scenes. + +This work lays the foundation for seamlessly integrating real-world scene captures into simulation, expanding the applicability of embodied AI systems. Another promising avenue is integrating GS directly into training, replacing visual observations and goal images to enhance learning efficiency. Furthermore, we aim to extend the use of Gaussian Splats to more complex embodied AI tasks, such as rearrangement and mobile manipulation, broadening its impact across diverse real-world applications. + +# References + +[1] Capture, share, and collaborate the built world in immersive 3D — matterport.com. https://matterport.com/. [Accessed 08-03-2025]. 3, 5 +[2] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation, 2022. 3 +[3] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 4 +[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2 +[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 4 +[6] Guillaume Bono, Leonid Antsfeld, Boris Chidlovskii, Philippe Weinzaepfel, and Christian Wolf. End-to-end (instance)-image goal navigation through correspondence as an emergent phenomenon, 2023. 3 +[7] Arunkumar Byravan, Jan Humplik, Leonard Hasenclever, Arthur Brussee, Francesco Nori, Tuomas Haarnoja, Ben Moran, Steven Bohez, Fereshteh Sadeghi, Bojan Vujatovic, and Nicolas Heess. Nerf2real: Sim2real transfer of vision-guided bipedal motion skills using neural radiance fields, 2022. 3, 2 +[8] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments, 2017. 1, 2 +[9] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration, 2020. 3 +[10] Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan, Muhammad Zubair Irshad, Richard Cheng, Kurt Keutzer, Masayoshi Tomizuka, Quan Vuong, and Ken Goldberg. Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning. In Conference on Robot Learning (CoRL), Munich, Germany, 2024. 2 +[11] Timothy Chen, Ola Shorinwa, Joseph Bruno, Javier Yu, Wei-jia Zeng, Keiko Nagami, Philip Dames, and Mac Schwager. Splat-nav: Safe real-time robot navigation in gaussian splatting maps, 2024. 2, 3 +[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels, 2024. 2 +[13] Matt Deitke, Rose Hendrix, Luca Weihs, Ali Farhadi, Kiana Ehsani, and Aniruddha Kembhavi. Phone2proc: Bringing robust robots into our chaotic world, 2022. 3, 6 +[14] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4, 3 + +[15] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering, 2023. 2 +[16] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. 2024. 4, 3 +[17] Muhammad Zubair Irshad, Chih-Yao Ma, and Zsolt Kira. Hierarchical cross-modal agent for robotics vision-and-language navigation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2021. 1 +[18] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4065-4071, 2022. 1 +[19] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. 2023. 2 +[20] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv:2410.20220, 2024. 2 +[21] Abhishek Kadian, Joanne Truong, Aaron Gokaslan, Alexander Clegg, Erik Wijmans, Stefan Lee, Manolis Savva, Sonia Chernova, and Dhruv Batra. Sim2real predictivity: Does evaluation in simulation predict real-world performance? IEEE Robotics and Automation Letters, 5(4):6670-6677, 2020. 1, 3, 7 +[22] Ivan Kapelyukh, Yifei Ren, Ignacio Alzugaray, and Edward Johns. Dream2real: Zero-shot 3d object rearrangement with vision-language models, 2024. 2 +[23] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4 +[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2 +[25] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Schacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. arXiv preprint arXiv:2306.11290, 2023. 1, 3, 4, 5 +[26] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 2, 3 +[27] Xiaohan Lei, Min Wang, Wengang Zhou, and Houqiang Li. Gaussnav: Gaussian splatting for visual navigation, 2024. 2, 3 + +[28] Yulong Li and Deepak Pathak. Object-aware gaussian splatt ing for robotic manipulation. In ICRA 2024 Workshop on 3D Visual Representations for Robot Manipulation, 2024. 3, 2 +[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 3 +[30] JunEn Low, Maximilian Adang, Javier Yu, Keiko Nagami, and Mac Schwager. Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum, 2025. 2 +[31] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings, 2023. 3 +[32] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, Dhruv Batra, Yixin Lin, Oleksandr Maksymets, Aravind Rajeswaran, and Franziska Meier. Where are we in the search for an artificial visual cortex for embodied intelligence?, 2024. 1, 3 +[33] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra Singh Chaplot. Autonerf: Training implicit scene representations with autonomous agents, 2023. 3, 2 +[34] Wugang Meng, Tianfu Wu, Huan Yin, and Fumin Zhang. Beings: Bayesian embodied image-goal navigation with gaussian splatting, 2024. 2 +[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3, 2 +[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2 +[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4 +[38] Polycam. Polycam. Accessed: 2024-11-10. 2, 3, 4 +[39] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. In The Twelfth International Conference on Learning Representations, 2023. 2, 3, 4, 5, 1 +[40] Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun, David Held, George Kantor, and Abhisesh Silwal. Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting, 2024. 2 +[41] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai, 2021. 1, 2, 3, 4, 5 + +[42] Xuqian Ren, Wenjia Wang, Dingding Cai, Tuuli Tuominen, Juho Kannala, and Esa Rahtu. Mushroom: Multi-sensor hybrid room dataset for joint 3d reconstruction and novel view synthesis, 2024. 3, 4, 6, 2 +[43] Meta AI Research. Home robot. https://github.com/facebookresearch/home-robot, 2023. GitHub repository. 1 +[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 1 +[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 3 +[46] Sneha Silwal, Karmesh Yadav, Tingfan Wu, Jay Vakil, Arjun Majumdar, Sergio Arnaud, Claire Chen, Vincent-Pierre Berges, Dhruv Batra, Aravind Rajeswaran, Mrinal Kalakrishnan, Franziska Meier, and Oleksandr Maksymets. What do we learn from a large-scale study of pre-trained visual representations in sim and real environments?, 2024. 1, 2, 3, 5, 7, 4 +[47] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3 +[48] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 4 +[49] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splattering and meshing, 2024. 2, 3, 4, 5 +[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In International Conference on Learning Representations, 2019. 3 +[51] Yaniv Wolf, Amit Bracha, and Ron Kimmel. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views, 2024. 2 +[52] Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, and Hesheng Wang. Rl-gsbridge: 3d gaussian splating based real2sim2real method for robotic manipulation learning, 2024. 3, 2 +[53] Fei Xia, Amir Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents, 2018. 2 +[54] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and + +browser-compatible environment from a single video, 2024. 2 +[55] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation, 2022. 3 +[56] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav, 2023. 3, 4 +[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 4 +[58] Justin Yu, Kush Hari, Kishore Srinivas, Karim El-Refai, Adam Rashid, Chung Min Kim, Justin Kerr, Richard Cheng, Muhammad Zubair Irshad, Ashwin Balakrishna, et al. Language-embedded gaussian splats (legs): incrementally building room-scale representations with a mobile robot. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13326–13332. IEEE, 2024. 2 +[59] Justin Yu, Kush Hari, Karim El-Refai, Arnav Dalil, Justin Kerr, Chung-Min Kim, Richard Cheng, Muhammad Z. Irshad, and Ken Goldberg. Persistent object gaussian splat (pogs) for tracking human and robot manipulation of irregularly shaped objects. ICRA, 2025. 2 +[60] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey, 2024. 2 \ No newline at end of file diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/images.zip b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..acbfa4c20426e8d61c434f1c23e6de08ae48c965 --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82682d045314dfb7a51ffd00aca4e06034d4c317f071846b61c789b6d1e7caa4 +size 413981 diff --git a/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/layout.json b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c1dcca8f9c4ea157b47bb1f25cfda48682222d8f --- /dev/null +++ b/2025/EmbodiedSplat_ Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device/layout.json @@ -0,0 +1,7585 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 541, + 140 + ], + "type": "text", + "content": "EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "spans": [ + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": "Gunjan Chhablani" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Xiaomeng Ye" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Muhammad Zubair Irshad" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": ", Zsolt Kira" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": "Georgia Tech, " + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 121, + 160, + 488, + 190 + ], + "type": "text", + "content": "Toyota Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 88, + 190, + 518, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 190, + 518, + 204 + ], + "spans": [ + { + "bbox": [ + 88, + 190, + 518, + 204 + ], + "type": "text", + "content": "{chhablani.gunjan, muhammadzubairirshad}@gmail.com, {xye87, zkira}@gatech.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "text", + "content": "The field of Embodied AI predominantly relies on simulation for training and evaluation, often using either fully synthetic environments that lack photorealism or high-fidelity real-world reconstructions captured with expensive hardware. As a result, sim-to-real transfer remains a major challenge. In this paper, we introduce EmbodiedSplat, a novel approach that personalizes policy training by efficiently capturing the deployment environment and fine-tuning policies within the reconstructed scenes. Our method leverages 3D Gaussian Splatting (GS) and the Habitat-Sim simulator to bridge the gap between realistic scene capture and effective training environments. Using iPhone-captured deployment scenes, we reconstruct meshes via GS, enabling training in settings that closely approximate real-world conditions. We conduct a comprehensive analysis of training strategies, pre-training datasets, and mesh reconstruction techniques, evaluating their impact on sim-to-real predictivity in real-world scenarios. Experimental results demonstrate that agents fine-tuned with EmbodiedSplat outperform both zero-shot baselines pre-trained on large-scale real-world datasets (HM3D) and synthetically generated datasets (HSSD), achieving absolute success rate improvements of " + }, + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 54, + 255, + 296, + 593 + ], + "type": "text", + "content": " on real-world Image Navigation task. Moreover, our approach yields a high sim-vs-real correlation (0.87-0.97) for the reconstructed meshes, underscoring its effectiveness in adapting policies to diverse environments with minimal effort. Project page: https://gchhablani.github.io/embodied-splat." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 135, + 627 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 696 + ], + "type": "text", + "content": "Recent advancements in Embodied AI have demonstrated impressive performance in simulated environments [17, 18, 32, 44, 46]. However, translating these capabilities to physical robots remains a significant challenge, primarily due to limitations in simulation fidelity and accessibility [21]. A" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 230, + 553, + 282 + ], + "blocks": [ + { + "bbox": [ + 317, + 230, + 553, + 282 + ], + "lines": [ + { + "bbox": [ + 317, + 230, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 317, + 230, + 553, + 282 + ], + "type": "image", + "image_path": "e7f16acf6b415ccc72ebda8492ade9cc1686285b2509ad2f81a0563857922492.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 284, + 552, + 427 + ], + "blocks": [ + { + "bbox": [ + 317, + 284, + 552, + 427 + ], + "lines": [ + { + "bbox": [ + 317, + 284, + 552, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 552, + 427 + ], + "type": "image", + "image_path": "8c7726c6f9d69e6904e81b6732783c27b6f3832e4f0bbee223deeb1d783a3554.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 437, + 555, + 526 + ], + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 526 + ], + "type": "text", + "content": "Figure 1. Overview of EmbodiedSplat: Mobile phone captures are used to generate reconstructed meshes via 3D Gaussian Splating (GS). Agents are trained within these reconstructed environments in simulation before being deployed in the real world, enabling effective sim-to-real transfer. Our analysis demonstrates a strong sim-to-real correlation across both types of generated meshes, highlighting their ability to bridge the gap between simulation and real-world performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 554, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 686 + ], + "type": "text", + "content": "key bottleneck is the sim-to-real gap, where handcrafted or synthetic simulation environments (e.g. HSSD [25]) struggle to capture the complexity and variability of real-world settings, necessitating the use of real-world reconstructions for effective policy training. On the other hand, real-world datasets such as Matterport3D [8] and HM3D [41] rely on expensive capture equipment and labor-intensive reconstruction pipelines, making large-scale scene collection and adaptation impractical for many applications. Furthermore, it is difficult to fully capture the variability of potential deployment environments with these datasets." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Developments in 3D scene representations, particularly 3D Gaussian Splatting (GS) [24], have shown promise to" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 269, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 269, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 269, + 713 + ], + "type": "text", + "content": "*Work done as a student at Georgia Tech, currently at Waymo." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25431" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 297, + 122 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 297, + 122 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 297, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 297, + 122 + ], + "type": "table", + "html": "
WorkGoal TypeE2ERealS2R
SplatNav [11]Point/Language×✓×
GaussNav [27]Image [26]×××
OursImage✓✓✓
", + "image_path": "0f36f4c731378f24c97b882a5d20c1882f7e40d3038f47465d3ccda9c5b3db1f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 130, + 295, + 175 + ], + "type": "text", + "content": "Table 1. Comparison against recent works using GS for navigation. \"E2E\" indicates whether the policy is trained end-to-end, \"Real\" denotes evaluation on real-world robots, and \"S2R\" indicates demonstrated sim-to-real transfer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 198, + 295, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 295, + 316 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 295, + 316 + ], + "type": "text", + "content": "wards reducing the effort needed to capture new scenes, enabling high-quality scene reconstruction from casual mobile phone captures. These approaches offer a strong ability to handle complex geometry, perform novel-view view synthesis, and provide high visual fidelity. Building upon this foundation, methods like DN-Splatter [49] have further enhanced mesh reconstruction quality through depth-and-normal regularization. However, their potential for training robot navigation policies and deploying them in the real-world has remained largely unexplored." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 317, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 295, + 544 + ], + "type": "text", + "content": "In this work, we explore the following question: Can low-effort cellphone video captures be leveraged to generate meshes that facilitate the training of Embodied AI navigation policies for effective transfer to the target environment? In other words, we seek to personalize models by training them directly on the target distribution, as represented by 3D models built from low-cost and low-effort data collects of the deployment environment. Some recent works indeed explore GS for navigation - GaussNav [11, 27] explores Instance-Image Navigation [26] in simulation using GS; SplatNav [11] explores using GS-based collision meshes for collision avoidance and localization in a single real-world environment using a modular policy for a drone. In contrast, as shown in Tab. 1, we explore end-to-end policy learning for image-goal navigation and sim-to-real transfer in indoor environments. To the best of our knowledge, we are the first to explore an approach to solve the personalized real-to-sim-to-real problem using Gaussian Splats for indoor image-goal navigation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 700 + ], + "type": "text", + "content": "To achieve the above goal, we present a comprehensive framework for leveraging open-source 3D Gaussian Splating [24] (and compare with Polycam [38]). The central premise of our work is that it is possible to quickly capture the scenes in which a robot will be deployed, using readily available consumer-grade hardware, and seamlessly integrate them into simulation environments (i.e. Habitat-Sim [39]). Policies can then be trained in these simulated environments, leading to improved sim-to-real transfer. Our approach combines the accessibility of smartphone-based scene capture with recent advances in depth-aware 3D scene representation, enabling rapid training and deployment of navigation policies in new, realistic environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "However, there are several challenges and open ques" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "tions, ranging from reconstruction quality to finetuning strategies, to enabling successful transfer. Building on the work of Silwal et al. [46] as a baseline, we test our framework across a range of scenes captured in a university environment, which is out-of-distribution for typical pre-trained policies. We analyze factors affecting transfer performance, including trade-offs between mesh generation pipelines, mesh quality for policy training, and training strategies (e.g., zero-shot vs. fine-tuning). Through rigorous evaluation and real-world robot experiments, we show that our approach yields significant gains in real-world image-goal navigation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 216, + 529, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 529, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 529, + 228 + ], + "type": "text", + "content": "Our key contributions can be summarized as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 553, + 506 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 231, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 553, + 290 + ], + "type": "text", + "content": "1. An efficient and cost-effective pipeline for bridging the real-to-sim gap in navigation, enabling the creation of high-quality simulation scenes from low-cost, consumer-grade iPhone captures using depth-aware 3D Gaussian Splats (GS) and Polycam." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 361 + ], + "type": "text", + "content": "2. Comprehensive evaluations conducted in both simulation and real-world environments, assessing zero-shot and fine-tuned policies on our Captured scenes. Our results demonstrate substantial improvements in sim-to-real transfer, emphasizing the effectiveness of fine-tuning on high-fidelity reconstructions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 362, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 362, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 362, + 553, + 480 + ], + "type": "text", + "content": "3. In-depth analysis addressing key research questions, exploring the relationship between reconstruction quality, pre-training scenes, downstream navigation performance, and training strategies. For example, a notable finding is that overfitting policies on a single high-fidelity scene reconstruction in simulation yields reasonable real-world performance. Our findings provide valuable insights into how reconstruction fidelity influences policy generalization and its applicability to real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 482, + 553, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 553, + 506 + ], + "type": "text", + "content": "4. An open-source codebase and dataset to facilitate further research in this domain and reproducibility of results." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 508, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 553, + 544 + ], + "type": "text", + "content": "Through this work, we aim to make high-quality scene collection and agent training more accessible, as well as enable easy development of personalized agents." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 556, + 400, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 400, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 400, + 568 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 576, + 406, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 406, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 406, + 587 + ], + "type": "text", + "content": "2.1. Scene Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "Recent years have seen rapid progress in the development of high-quality 3D scene datasets for embodied AI research. The Matterport3D dataset [8] provides 10,800 panoramic views from 194,400 RGB-D images across 90 buildings-scale scenes, complete with surface reconstructions and semantic annotations. Building upon this, HM3D [41] expanded the scale to 1,000 building-scale 3D reconstructions from diverse real-world locations, though some issues with mesh quality were noted. The Gibson dataset [53] introduced another collection of real-world scans, while" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "25432" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 323 + ], + "type": "text", + "content": "synthetic datasets like HSSD [25] (with 211 realistic environments) demonstrated that smaller but higher-quality datasets can sometimes be better than larger ones for agent training. Another dataset that is commonly used is ReplicaCAD [47] which provides synthetic variations in layouts of a single scene. In this work, we use HM3D [41] for training our zero-shot baseline, as it is one of the largest photorealistic indoor-scene datasets captured using Matterport [1] cameras. Additionally, we use HSSD [25] as the synthetic counterpart for another zero-shot baseline. Note, however, that these scenes do have a bias towards apartments, and we therefore test our methods on scenes within a university environment, which is out-of-distribution for these datasets. The MuSHRoom dataset [42] provides multi-sensor captures of 10 real-world scenes, offering benchmarks for reconstruction and novel view synthesis methods. While this dataset has not been used for embodied tasks in prior works, our dataset collection method is inspired by the MuSHRoom dataset's iPhone captures. We evaluate our agents, as well as reconstruction strategies, on these scenes to benchmark our depth and normal encoders." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 294, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 294, + 346 + ], + "type": "text", + "content": "2.2. Embodied Navigation in Indoor Environments" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 508 + ], + "type": "text", + "content": "Embodied navigation encompasses various forms of goal-directed navigation tasks [2, 6, 9, 26, 31, 50, 55, 56]. This study primarily investigates ImageNav [2, 55, 56] in indoor environments. Recent advancements in image-goal navigation have explored a variety of approaches to enhance performance and generalization. Notably, pre-trained visual representations [32, 46, 55, 56] have shown substantial promise in improving performance on the ImageNav task. In particular, Silwal et al. [46] demonstrate remarkable zero-shot success rates (90%) on real-world ImageNav, leveraging VC-1 [32], fine-tuned end-to-end with data augmentation. Inspired by this, we use the setup in Silwal et al. [46] for our policy training and evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 592 + ], + "type": "text", + "content": "Sim-to-real transfer remains a core challenge in Embodied AI beyond leveraging pre-trained visual encoders. Kadian et al. [21] proposed the Sim-vs-Real Correlation Coefficient (SRCC) to quantify how well PointNav performance in simulation predicts real-world results. We build on this idea to evaluate the sim-to-real predictivity of our Captured meshes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "A closely related work is Phone2Proc [13], which demonstrates enhanced sim-to-real ObjectNav performance by generating layouts from iPhone RoomPlan API captures. While similar to our approach in utilizing iPhone captures, our methodology differs in that we capture the entire room without focusing on layouts, nor do we perform any post-processing of scenes or generate multiple variations of the same scene for agent training. Instead, we generate meshes from our captures, fine-tune pre-trained policies for ImageNav task, and deploy them subsequently." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 72, + 503, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 503, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 503, + 83 + ], + "type": "text", + "content": "2.3. 3D Scene and Mesh Reconstruction" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 90, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 90, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 90, + 555, + 198 + ], + "type": "text", + "content": "In this work, we adopt DN-Splatter [49] for its simplicity and superior performance (and compare to meshes from Polycam [38], which is not open-source). DN-Splatter uses depth-and-normal regularization to improve mesh quality. While mesh reconstruction has also been explored with NeRFs [35], we choose GS for 3D reconstruction due to their fast training and rendering speeds, improving the overall efficiency of the pipeline. For an expanded list of related works on this topic, please see Appendix D.1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 220 + ], + "type": "text", + "content": "2.4. 3D Representations for Embodied AI/Robotics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 226, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 226, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 226, + 555, + 548 + ], + "type": "text", + "content": "There have been works which explore sim-to-real manipulation [7, 28, 52]. In this work, we attempt to solve the problem of sim-to-real navigation, laying the foundation for a diverse range of embodied tasks, such as object-goal navigation, rearrangement, and mobile manipulation, which require room-, apartment-, and building-level scene representations. Capturing high-quality 3D Gaussian Splats (GS) is relatively straightforward for tabletop manipulation tasks, whereas large-scale scene reconstruction poses greater challenges — a key aspect of this work. Recent works also explore the use of GS in solving the embodied navigation problem. GaussNav [27] presents a semantic Gaussian-based map reconstruction for the HM3D-Instance ImageNav [26] task. Splat-Nav [11] introduces a two-stage pipeline for planning and pose estimation through Gaussian Splat Maps. Tab. 1 highlights the differences between our approach and recent works utilizing GS for navigation. By capturing the entire scene using mobile phones, we enable personalization to specific scenes, as well as the potential to scale scene collection in the future. Some works [33] explore automated 3D scene capture using robot cameras in simulation, but these methods require complex, embodiment-dependent strategies that are time-consuming and difficult to scale. In contrast, our approach leverages human captures in the real-world, offering a fast and scalable solution without the need for intricate planning. For more related work on this topic, please see Appendix D.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 561, + 397, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 561, + 397, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 561, + 397, + 574 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "The overall pipeline for bridging and integrating a real-world scene with Habitat-Sim [39] is shown in Fig. 2. In the following subsections, we discuss each stage of the pipeline in detail. Sec. 3.1 discusses the datasets used and the details of scene captures. Sec. 3.2 discusses the second stage, where these captures are converted to meshes. Sec. 3.3 discusses how ImageNav episodes are created, which is a crucial part of integrating the scene with Habitat Simulator [39]. Appendix H discusses agent training and evaluation details. For details on real-world deployment setup, refer to Appendix A." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25433" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 72, + 555, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 72, + 555, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 72, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 555, + 235 + ], + "type": "image", + "image_path": "6002c9aa5b763a45cccefc9b1c847b2dec73bd4c8faa6fb56ed50c24a0f8cb27.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 555, + 314 + ], + "type": "text", + "content": "Figure 2. The EmbodiedSplat Pipeline: Pipeline for integration of real-world captures with Habitat-Sim [39] and subsequent deployment. The first stage (a.) involves capturing the scene using Polycam [38] and Nerfstudio [48] which produces RGB frames, associated iPhone GT depth maps, and poses. In the second stage (b.), we use DN-Splatter [49] to train GS using depth and normal regularization, with normals from Metric3D-V2 [16] monocular encoder. A mesh (.ply) is created using Poisson reconstruction from the GS. In the third stage (c.), the mesh is processed and loaded into Habitat-Sim [39] for training the agent in simulation. In the last stage (d.), the policy is deployed in the real-world in the same scene for image-goal navigation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 214, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 214, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 214, + 346 + ], + "type": "text", + "content": "3.1. Datasets and Scene Captures" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 353, + 295, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 353, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 54, + 353, + 295, + 436 + ], + "type": "text", + "content": "Scene Datasets: We use two datasets for pre-training ImageNav policies, which serve as zero-shot baselines and pre-trained policies for our scenes, both in simulation and real-world evaluations. Specifically, we use the HM3D [41] dataset which consists of apartment-scale scenes split into 800 training and 100 validation scenes, and the HSSD [25] dataset split into 134 training and 33 validation scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 714 + ], + "type": "text", + "content": "Captured Scenes: To evaluate the feasibility of the pipeline and conduct real-world experiments, we capture scenes from a university environment (classroom, community lounges, conference rooms, etc). For custom data collection with an iPhone, we follow the procedure used for collecting the MuSHRoom dataset [42]. Specifically, we use an iPhone 13 Pro Max to record the iPhone RGB-D data using the Polycam application [38]. Polycam provides an assistive interface to ensure that all the details of the scene have been sufficiently covered during the capture. We use the default settings with Polycam and export the raw data exposed by the application. Subsequently, we use Nerf-studio [48] to process the RGB-D data and sample 1000 aligned RGB-depth frames with low blur scores and corresponding poses. Additionally, Polycam also provides a mesh with its exported data. We use this mesh for comparison purposes, referred to as POLYCAM meshes throughout the paper. Unlike Ren et al. [42], we use a manually-held iPhone for capture instead of a gimbal, enabling us to evaluate whether the process remains effective and easily replicable without relying on expensive capture equipment. Each capture requires 20-30 minutes of recording with Polycam. We repeat this process for different indoor scenes, which we" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 334, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 370 + ], + "type": "text", + "content": "refer to as lounge, classroom, conf_a, and conf_b. The scale of these of these scenes is similar to those in the MuSHRoom dataset (1-3 rooms) [42]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 371, + 556, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 479 + ], + "type": "text", + "content": "MuSHRoom Dataset: We use MuSHRoom [42] dataset for benchmarking. The dataset consists of long and short sequences of 10 indoor scenes, captured with iPhone, Kinect, and Faro Scanner. In our work, we only use the iPhone long-sequences for training and evaluation purposes. We benchmark different normal encoders on all 10 scenes (see Appendix G) for DN-Splatter [49]. For agent training and evaluation, we primarily use three scenes - honka, sauna, activity - which have varied scale and complexity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 490, + 503, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 503, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 503, + 502 + ], + "type": "text", + "content": "3.2. 3D Scene and Mesh Reconstruction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 509, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 616 + ], + "type": "text", + "content": "For mesh reconstruction, we use DN-Splatter [49] as our method of choice for its superior performance on mesh reconstruction in comparison to others, and simplicity of its integration with Habitat-Sim [39]. DN-Splatter leverages depth-normal regularization, and smoothness losses to maintain geometric consistency during Gaussian Splat training. For mesh reconstruction, the rendered depth and normal maps are back-projected from training views to create a point-set for meshing." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "We use the default hyperparameters from DN-Splatter [49] and train the GS for 30,000 iterations. We use sensor depth " + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{d} = 0.2" + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": ", and enable the depth smoothness and normal losses. Metric3D-V2 [16] is used as the normal encoder instead of Omnidata [14, 23], as empirical results demonstrate it yields higher-quality meshes. For further discussion on depth and normal encoder selection, see Appendix G." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25434" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 164, + 158 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 164, + 158 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 164, + 158 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 164, + 158 + ], + "type": "image", + "image_path": "5740ae30a9454782880445ce517f7542d8303fbc3c6cb1be7b77fdf50b594d6c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 161, + 137, + 170 + ], + "lines": [ + { + "bbox": [ + 97, + 161, + 137, + 170 + ], + "spans": [ + { + "bbox": [ + 97, + 161, + 137, + 170 + ], + "type": "text", + "content": "(a) conf.b" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 189, + 82, + 291, + 158 + ], + "blocks": [ + { + "bbox": [ + 189, + 82, + 291, + 158 + ], + "lines": [ + { + "bbox": [ + 189, + 82, + 291, + 158 + ], + "spans": [ + { + "bbox": [ + 189, + 82, + 291, + 158 + ], + "type": "image", + "image_path": "99562dac0ade027ceca080b7985028af6c15ec938a6c6ab5ce7f0b7746a1f523.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 160, + 262, + 170 + ], + "lines": [ + { + "bbox": [ + 223, + 160, + 262, + 170 + ], + "spans": [ + { + "bbox": [ + 223, + 160, + 262, + 170 + ], + "type": "text", + "content": "(b) conf_a" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 317, + 88, + 404, + 159 + ], + "blocks": [ + { + "bbox": [ + 317, + 88, + 404, + 159 + ], + "lines": [ + { + "bbox": [ + 317, + 88, + 404, + 159 + ], + "spans": [ + { + "bbox": [ + 317, + 88, + 404, + 159 + ], + "type": "image", + "image_path": "2273cddcecf21365cdc2119b8471557da8c581508fb16589a526f288bf99b219.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 161, + 396, + 170 + ], + "lines": [ + { + "bbox": [ + 340, + 161, + 396, + 170 + ], + "spans": [ + { + "bbox": [ + 340, + 161, + 396, + 170 + ], + "type": "text", + "content": "(c) classroom" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 91, + 179, + 517, + 192 + ], + "lines": [ + { + "bbox": [ + 91, + 179, + 517, + 192 + ], + "spans": [ + { + "bbox": [ + 91, + 179, + 517, + 192 + ], + "type": "text", + "content": "Figure 3. Reconstructed meshes of the Captured scenes after DN-Splatter [49] training and Poisson reconstruction." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 447, + 72, + 548, + 159 + ], + "blocks": [ + { + "bbox": [ + 447, + 72, + 548, + 159 + ], + "lines": [ + { + "bbox": [ + 447, + 72, + 548, + 159 + ], + "spans": [ + { + "bbox": [ + 447, + 72, + 548, + 159 + ], + "type": "image", + "image_path": "4ba08cb9ec672ff0d6dc02abb2166cbfb24d9c941ecc5d1ab7b4a8469c63f0ea.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 161, + 514, + 170 + ], + "lines": [ + { + "bbox": [ + 473, + 161, + 514, + 170 + ], + "spans": [ + { + "bbox": [ + 473, + 161, + 514, + 170 + ], + "type": "text", + "content": "(d) lounge" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 213, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 332 + ], + "type": "text", + "content": "After training GS and converting into ply meshes, we convert the meshes to glb in Blender (see Appendix E) before loading them in Habitat-Sim [39]. The final meshes for our Captured scenes are shown in Fig. 3 and referred to as DN mesh hereafter to contrast against the POLYCAM mesh counterpart. It takes approximately 20-30 minutes per capture, and 1-2 hours of training with DN-Splatter [49] to generate these meshes, which is significantly lesser compared to the cost and several hours of capture and processing with Matterport [1] cameras." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 343, + 222, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 222, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 222, + 357 + ], + "type": "text", + "content": "3.3. ImageNav Episode Generation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 361, + 296, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 411 + ], + "type": "text", + "content": "Pre-training Scene Datasets: For the HM3D [41] and HSSD [25] datasets, we use the predefined train-validation split, generating 10,000 episodes for each training scene and 25 episodes for each validation scene (see Appendix B)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": "Captured and MuSHRoom Scenes: For scenes with DN meshes and POLYCAM meshes, we select only the largest navmesh island within each scene. We ensure that the largest island covers most, if not all, of the navigable areas in the scene. This step is crucial due to the relatively smaller sizes of these scenes compared to HM3D scenes. Hence, we generate only 1000 training episodes and 100 evaluation episodes per scene. By selecting the largest navmesh island, we minimize the risk of incorrectly treating entities other than the floor as navigable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": "Evaluation Metric: We consider Success Rate (SR), the fraction of successful episodes out of all the episodes, as the primary metric for performance evaluation of our policies. An episode is considered successful if the agent stops within " + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "inline_equation", + "content": "1\\mathrm{m}" + }, + { + "bbox": [ + 55, + 532, + 296, + 605 + ], + "type": "text", + "content": " of the goal location before the maximum number of steps (1000 for simulation, 100 for real) are over." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 619, + 137, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 137, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 137, + 632 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 640, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 295, + 663 + ], + "type": "text", + "content": "This section aims to answer the following research questions:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 665, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 665, + 295, + 689 + ], + "type": "text", + "content": "1. How does a pre-trained policy perform on Captured and MuSHRoom scenes? (Sec. 4.1)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 689, + 296, + 714 + ], + "type": "text", + "content": "2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance? (Sec. 4.2)." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 236 + ], + "type": "text", + "content": "3. Does the performance in simulation transfer to the realworld? (Sec. 4.3)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 236, + 545, + 249 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 545, + 249 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 545, + 249 + ], + "type": "text", + "content": "Further experiments and analyses are discussed in Sec. 5." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 255, + 553, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 280 + ], + "type": "text", + "content": "4.1. How does a pre-trained policy perform on Captured and MuSHRoom scenes?" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 285, + 555, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 416 + ], + "type": "text", + "content": "In this experiment, we evaluate the pre-trained policies that achieved the highest validation success rates on the HM3D [41] (83.08% val SR) and HSSD [25] (63.15% val SR), trained for 600M and 1200M steps, respectively. Note that the performance on HM3D is consistent with the simulation results reported in Silwal et al. [46] and represents the current state-of-the-art for this task. For agent embodiment, policy training and evaluation details, please refer to Appendix H. We assess the zero-shot generalization of these policies within simulation, evaluating their performance across different meshes for individual scenes." + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 426, + 550, + 533 + ], + "blocks": [ + { + "bbox": [ + 317, + 426, + 550, + 533 + ], + "lines": [ + { + "bbox": [ + 317, + 426, + 550, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 550, + 533 + ], + "type": "image", + "image_path": "39d0b3ff99eade48b144ed040e17c4c787b65e637b9ec608ad8d786bd42002b5.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 544, + 544, + 556 + ], + "lines": [ + { + "bbox": [ + 323, + 544, + 544, + 556 + ], + "spans": [ + { + "bbox": [ + 323, + 544, + 544, + 556 + ], + "type": "text", + "content": "Figure 4. Zero-shot val. SR for HM3D pre-trained policy." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": "Fig. 4 presents the results of the zero-shot evaluation across individual scenes for the HM3D pre-trained policy. Performance varies significantly based on scene size, complexity, and mesh type. For smaller scenes such as conf_a and conf_b, the policy demonstrates relatively high success rates in both DN mesh and POLYCAM settings, achieving " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " (DN) and " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "82\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " (POLYCAM) for conf_a and " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "88\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " (DN mesh) and " + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "inline_equation", + "content": "79\\%" + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": " (POLYCAM) for conf_b." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "In contrast, performance declines in larger environments such as classroom and lounge. In classroom, success rates drop to " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "53\\%" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " with DN mesh and " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "42\\%" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " with POLYCAM, while in lounge, the policy achieves " + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": " (DN" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25435" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 82, + 287, + 178 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 292, + 81 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 292, + 81 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 292, + 81 + ], + "type": "text", + "content": "HSSD Zero-Shot Results on Captured and MuSHRoom Scenes" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 82, + 287, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 82, + 287, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 82, + 287, + 178 + ], + "type": "image", + "image_path": "0cecb7120ff12bf0c11d540300add3e767fab168f86faedf352804101c4e67e1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "text", + "content": "mesh) and " + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "text", + "content": " (POLYCAM). These results highlight the challenge of transferring the policy to more complex, large-scale environments that differ from the training data. Since HM3D consists primarily of apartment-style scenes, the policy has not been exposed to classrooms or community lounges during training. Both DN and POLYCAM meshes yield similar success rates on average " + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "inline_equation", + "content": "(\\sim 60\\%)" + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "text", + "content": ", suggesting no clear advantage of one mesh type over the other in aligning with HM3D. Additionally, we evaluate the policies on three scenes from the MuSHRoom [42] dataset. Due to the unavailability of POLYCAM meshes for MuSHRoom, we conduct evaluations solely on DN meshes. The policy achieves a " + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 54, + 222, + 295, + 437 + ], + "type": "text", + "content": " success rate in the sauna and honka scenes, suggesting promising potential for performance improvements through additional scene refinements (such as using a gimbal during capture). However, in the larger activity scene, performance declines, indicating challenges in generalizing to larger environments." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "content": "Fig. 5 reveals similar trends for the HSSD pre-trained policy, though with significantly lower success rates overall. This degradation in performance can be attributed to the synthetic nature of HSSD scenes, which lack the realism and scale necessary for effective generalization. The policy performs particularly poorly on classroom, achieving only " + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "content": " success on both DN mesh and POLYCAM, while performance in lounge is slightly better, reaching " + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 55, + 437, + 295, + 570 + ], + "type": "text", + "content": " success with POLYCAM. The policy exhibits slightly improved performance on other scenes, but it is hard to conclude which mesh type best aligns with HSSD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 601 + ], + "type": "text", + "content": "4.2. Does fine-tuning on the Captured and MuSHRoom scenes improve performance?" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "content": "In this section, we investigate whether fine-tuning the pre-trained policies for only 20M additional steps improves performance in simulation. The learning rate is set to " + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "inline_equation", + "content": "2.5e - 6" + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "content": " for the LSTM policy and " + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "inline_equation", + "content": "6e - 7" + }, + { + "bbox": [ + 55, + 605, + 295, + 701 + ], + "type": "text", + "content": " for the visual encoder, following a fine-tuning strategy similar to that of Deitke et al. [13]. We fine-tune the pre-trained policy on the training episodes for a single scene and evaluate on the corresponding validation episodes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "The results are presented in Fig. 6 and Fig. 7. We ob" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 316, + 82, + 544, + 178 + ], + "blocks": [ + { + "bbox": [ + 66, + 190, + 284, + 202 + ], + "lines": [ + { + "bbox": [ + 66, + 190, + 284, + 202 + ], + "spans": [ + { + "bbox": [ + 66, + 190, + 284, + 202 + ], + "type": "text", + "content": "Figure 5. Zero-shot val. SR for HSSD pre-trained policy." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 72, + 551, + 81 + ], + "lines": [ + { + "bbox": [ + 334, + 72, + 551, + 81 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 551, + 81 + ], + "type": "text", + "content": "HM3D Fine-tuned Results on Captured and MuSHRoom Scenes" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 82, + 544, + 178 + ], + "lines": [ + { + "bbox": [ + 316, + 82, + 544, + 178 + ], + "spans": [ + { + "bbox": [ + 316, + 82, + 544, + 178 + ], + "type": "image", + "image_path": "601aeea520cda99e5bc897510d426123c502cb35e23164404e378baf52aa3931.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 213 + ], + "type": "text", + "content": "Figure 6. Fine-tuned validation SR from HM3D pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 550, + 332 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 550, + 332 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 550, + 332 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 550, + 332 + ], + "type": "image", + "image_path": "219958e3f0fb2eab600e9c11125a03294961fcd86030cbd276ef00ddb761ca8c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 344, + 553, + 367 + ], + "lines": [ + { + "bbox": [ + 313, + 344, + 553, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 553, + 367 + ], + "type": "text", + "content": "Figure 7. Fine-tuned validation SR from HSSD pre-trained policy: The policy is fine-tuned and evaluated on the same mesh." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "content": "serve that fine-tuning significantly improves performance across all tested scenes. For the pre-trained HM3D policy, fine-tuning for 20M steps results in success rates approaching " + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "inline_equation", + "content": "90\\%+" + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "content": " across different meshes. Similarly, for the HSSD pre-trained policy, fine-tuning leads to substantial improvements, with most policies achieving success rates of " + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "inline_equation", + "content": "80\\%+" + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "content": " in respective scenes. In particular, performance gains are particularly pronounced in larger, more complex environments such as classroom and lounge, which differ significantly from apartment-style scenes in HM3D and HSSD. These results indicate that our pipeline can effectively be used to collect diverse scene data at scale. Furthermore, they suggest that agents can be quickly finetuned (only 20M steps vs " + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "inline_equation", + "content": "100\\mathrm{M}+" + }, + { + "bbox": [ + 313, + 388, + 555, + 604 + ], + "type": "text", + "content": " steps for training from scratch) on these generated meshes to improve performance on specific scene types, enhancing personalization beyond the original training distribution. We present results on additional scenes in Appendix J." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 553, + 635 + ], + "type": "text", + "content": "4.3. Does the performance in simulation transfer to the real-world?" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 713 + ], + "type": "text", + "content": "We evaluate both zero-shot and fine-tuned policies in the real-world lounge scene on a Stretch robot. During the evaluation, each episode is capped at 100 steps, with 10 distinct start-and-goal locations sampled within the scene (see Appendix A for deployment details). To assess performance, we record the number of steps taken and final" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25436" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 97 + ], + "type": "text", + "content": "distance to the goal at the end of each episode, determining whether the agent successfully completes the task." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 107, + 291, + 220 + ], + "blocks": [ + { + "bbox": [ + 58, + 107, + 291, + 220 + ], + "lines": [ + { + "bbox": [ + 58, + 107, + 291, + 220 + ], + "spans": [ + { + "bbox": [ + 58, + 107, + 291, + 220 + ], + "type": "image", + "image_path": "20c70cec609ca492acba284f0e3cca7385aeda0d60c1bf9887857a2748d04df4.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 230, + 296, + 286 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 296, + 286 + ], + "type": "text", + "content": "Figure 8. Real world results of zero-shot and fine-tuned models on 1ounge scene. HM3D-ZS (real dataset) leads to a better success rate than HSSD-ZS (synthetic dataset). We observe increased real-world success with fine-tuning on DN and POLYCAM meshes for both HM3D and HSSD policies." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": "Fig. 8 presents results across 10 evaluation episodes. The zero-shot HM3D policy achieves a " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": " success rate, demonstrating our hypothesized lack of generalization. This is in contrast with the results reported in Silwal et al. [46] showing " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": " zero-shot success rate in the real-world. We attribute this discrepancy to the structural and semantic differences between the lounge and the apartment-style scenes typically encountered in HM3D. Fine-tuning on the POLYCAM and DN mesh reconstructions of this scene improves performance, with success rates increasing to " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "70\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": ". For HSSD, zero-shot performance is significantly lower at " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": ", while fine-tuned policies improve success rates to " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": " with POLYCAM and " + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 55, + 299, + 295, + 453 + ], + "type": "text", + "content": " with DN mesh." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 295, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 551 + ], + "type": "text", + "content": "Fig. 1 illustrates the Sim-to-Real Correlation Coefficient (SRCC) [21] between simulation and real-world performance. The observation suggests that improvements in evaluation performance on DN and POLYCAM meshes in simulation translate to improved real-world performance. This demonstrates that our approach can efficiently adapt policies to novel real-world environments. We discuss real-world statistics (number of steps, distance) in Appendix I." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 562, + 189, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 189, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 189, + 575 + ], + "type": "text", + "content": "5. Ablations and Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "In this section, we systematically investigate critical factors influencing policy performance. First, we assess whether pre-training on large-scale datasets such as HM3D or HSSD is necessary for effective real-world transfer (Sec. 5.1). Next, we analyze the impact of validation PSNR and scene geometry on the zero-shot performance (Sec. 5.2). Finally, we examine whether continuous training on large-scale datasets improves zero-shot performance on our Captured scenes, shedding light on the relationship between dataset characteristics and generalizability (Sec. 5.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "5.1. Is it necessary to pre-train over large-scale datasets?" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 101, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 101, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 101, + 555, + 269 + ], + "type": "text", + "content": "In this experiment, we investigate the necessity of pretraining on large-scale datasets by \"overfitting\" policies directly on POLYCAM and DN meshes with a policy trained from scratch (not pre-trained on large-scale datasets such as HM3D or HSSD) for " + }, + { + "bbox": [ + 313, + 101, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\sim 100\\mathrm{M}" + }, + { + "bbox": [ + 313, + 101, + 555, + 269 + ], + "type": "text", + "content": " steps. Training and validation data are derived from ImageNav episode generation, as described in Sec. 3.3. To ensure diversity, start-goal locations for training and evaluation episodes are randomly sampled. The best validation success rates for all policies on their respective scenes are presented in Fig. 9. As expected, the overfitted policies achieve near-perfect performance in simulation, maintaining high success rates. For a detailed discussion on why overfitting results seem better than fine-tuning in this scenario, please see Appendix L." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 280, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 316, + 280, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 392 + ], + "type": "image", + "image_path": "c837f32cfba2b5daf80ef7e04e8461909eff2a64da2a5af9ae3435a4b2ce8380.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 425 + ], + "type": "text", + "content": "Figure 9. Validation SR for overfitted policies on respective scenes: Polices are not pre-trained on HM3D/HSSD." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "content": "We further evaluate these overfitted policies in the real-world lounge scene using POLYCAM and DN meshes. Surprisingly, the overfitted policy trained on the POLYCAM mesh achieves a " + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "content": " success rate in real-world evaluations. In contrast, the policy trained on the DN mesh achieves only " + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 437, + 556, + 594 + ], + "type": "text", + "content": " success. This result suggests that our mesh-based training approach can yield non-zero real-world performance, even without large-scale pre-training. We attribute the performance gap between POLYCAM and DN-trained policies to differences in visual fidelity—POLYCAM meshes preserve more visual detail by directly utilizing original images to reconstruct the scene, whereas DN meshes are based on GS which use the learned colors for the 3D Gaussians." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 600, + 555, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 600, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 600, + 555, + 625 + ], + "type": "text", + "content": "5.2. How is the scale and validation PSNR of the scene related to the final performance?" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Fig. 10 illustrates the correlation between HM3D zero-shot validation success rates on DN meshes and two key factors: the Peak Signal-to-Noise Ratio (PSNR) of the corresponding 3D Gaussian Splats (GS) and the scale of the scene, measured as the average shortest distance between start-goal locations in validation episodes. We observe a negative correlation between success rate (SR) and average" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25437" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 294, + 185 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 185 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 185 + ], + "type": "image", + "image_path": "01c3bf71b84b52d8d26c633cc7033e0ab77fa9b82b6d0dd85ce8527568927565.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 254 + ], + "type": "text", + "content": "Figure 10. HM3D zero-shot validation success rates on DN meshes vs. validation GS PSNRs vs. average shortest distances of validation episodes. The zero-shot SR is inversely correlated with the scale of the scene, while directly correlated with val. PSNR on the trained GS." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 276, + 295, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 276, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 54, + 276, + 295, + 396 + ], + "type": "text", + "content": "shortest distance—indicating that as the scale of the scene increases, the zero-shot success rate declines. Conversely, a positive correlation is seen between SR and PSNR, where higher validation PSNR values correspond to improved success rates. Notably, different trend lines emerge for MuSHRoom captures and our own Captured scenes. MuSHRoom captures generally exhibit higher validation PSNRs, likely due to the use of a stabilized gimbal during data collection. Future work will further investigate the impact of capture stability on policy performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 404, + 295, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 431 + ], + "type": "text", + "content": "5.3. Does continuous training on large-scale datasets improve zero-shot performance?" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 448, + 293, + 559 + ], + "blocks": [ + { + "bbox": [ + 58, + 448, + 293, + 559 + ], + "lines": [ + { + "bbox": [ + 58, + 448, + 293, + 559 + ], + "spans": [ + { + "bbox": [ + 58, + 448, + 293, + 559 + ], + "type": "image", + "image_path": "3e06aafad43397d2acb1bc4aee357c3c3b59607cdbbc99007fd5228834aaec64.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 592 + ], + "type": "text", + "content": "Figure 11. Average zero-shot success rates on respective validation sets over different stages of pre-training on HM3D." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 715 + ], + "type": "text", + "content": "Fig. 11 presents average zero-shot success rates across HM3D validation set, POLYCAM meshes, and DN meshes at various stages of HM3D policy pre-training. This analysis aims to determine whether continuous performance improvements on HM3D validation scenes translate to improved zero-shot performance on our Captured scenes. We observe that while performance initially increases, it begins to deteriorate or plateau at approximately 400M steps, despite continued improvements on HM3D" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 73, + 553, + 183 + ], + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 183 + ], + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 183 + ], + "type": "image", + "image_path": "d2bad78c6e6af564e8a4d7a8097e0f59851c0beda5062ce327721f60df0c5b12.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 194, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 217 + ], + "type": "text", + "content": "Figure 12. Average zero-shot success rates on respective validation sets over different stages of pre-training on HSSD." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 238, + 388, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 388, + 248 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 388, + 248 + ], + "type": "text", + "content": "validation scenes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 250, + 555, + 370 + ], + "type": "text", + "content": "Fig. 12 shows a similar experiment conducted using the HSSD dataset. Up to 300M steps, improvements in HSSD validation performance correspond to slight improvements in zero-shot performance on our Captured scenes. However, performance plateaus for Captured, showing no further gains despite continued improvement on HSSD validation scenes. For HM3D pre-training, POLYCAM meshes outperform DN meshes in success rates, while for HSSD, the trend reverses slightly, highlighting the impact of pretraining dataset characteristics on zero-shot generalization." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 381, + 388, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 388, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 388, + 393 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 402, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 605 + ], + "type": "text", + "content": "In this work, we introduced a comprehensive pipeline for bridging the gap between real-world and simulated environments in training embodied agents using 3D Gaussian Splats (GS) and Polycam. By leveraging the MuSHRoom dataset and custom iPhone-captured scenes, we demonstrated an efficient and scalable approach to policy personalization, leveraging 3D scene reconstruction from low-effort collects, enabling high-quality training for the ImageNav task. Our pipeline facilitates accessible and replicable scene collection without requiring specialized hardware or significant costs, making it a practical solution for large-scale embodied AI research. We evaluated overfitted, zero-shot, and fine-tuned policies, showing that fine-tuning pre-trained policies on real-world scene reconstructions improves sim-to-real transfer. We also analyzed the differences between GS-generated DN meshes and POLYCAM meshes, finding that POLYCAM more closely resemble real-world scenes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "This work lays the foundation for seamlessly integrating real-world scene captures into simulation, expanding the applicability of embodied AI systems. Another promising avenue is integrating GS directly into training, replacing visual observations and goal images to enhance learning efficiency. Furthermore, we aim to extend the use of Gaussian Splats to more complex embodied AI tasks, such as rearrangement and mobile manipulation, broadening its impact across diverse real-world applications." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "25438" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 123 + ], + "type": "text", + "content": "[1] Capture, share, and collaborate the built world in immersive 3D — matterport.com. https://matterport.com/. [Accessed 08-03-2025]. 3, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 158 + ], + "type": "text", + "content": "[2] Ziad Al-Halah, Santhosh K. Ramakrishnan, and Kristen Grauman. Zero experience required: Plug & play modular transfer learning for semantic visual navigation, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 296, + 203 + ], + "type": "text", + "content": "[3] Gwangbin Bae and Andrew J. Davison. Rethinking inductive biases for surface normal estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 296, + 248 + ], + "type": "text", + "content": "[4] Jonathan T. Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P. Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 295, + 283 + ], + "type": "text", + "content": "[5] Shariq Farooq Bhat, Reiner Birkl, Diana Wofk, Peter Wonka, and Matthias Müller. Zoedepth: Zero-shot transfer by combining relative and metric depth, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "text", + "content": "[6] Guillaume Bono, Leonid Antsfeld, Boris Chidlovskii, Philippe Weinzaepfel, and Christian Wolf. End-to-end (instance)-image goal navigation through correspondence as an emergent phenomenon, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 296, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 296, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 296, + 395 + ], + "type": "text", + "content": "[7] Arunkumar Byravan, Jan Humplik, Leonard Hasenclever, Arthur Brussee, Francesco Nori, Tuomas Haarnoja, Ben Moran, Steven Bohez, Fereshteh Sadeghi, Bojan Vujatovic, and Nicolas Heess. Nerf2real: Sim2real transfer of vision-guided bipedal motion skills using neural radiance fields, 2022. 3, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 296, + 441 + ], + "type": "text", + "content": "[8] Angel Chang, Angela Dai, Thomas Funkhouser, Maciej Halber, Matthias Nießner, Manolis Savva, Shuran Song, Andy Zeng, and Yinda Zhang. Matterport3d: Learning from rgb-d data in indoor environments, 2017. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 295, + 475 + ], + "type": "text", + "content": "[9] Devendra Singh Chaplot, Dhiraj Gandhi, Abhinav Gupta, and Ruslan Salakhutdinov. Object goal navigation using goal-oriented semantic exploration, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 477, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 477, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 477, + 296, + 542 + ], + "type": "text", + "content": "[10] Lawrence Yunliang Chen, Chenfeng Xu, Karthik Dharmarajan, Muhammad Zubair Irshad, Richard Cheng, Kurt Keutzer, Masayoshi Tomizuka, Quan Vuong, and Ken Goldberg. Rovi-aug: Robot and viewpoint augmentation for cross-embodiment robot learning. In Conference on Robot Learning (CoRL), Munich, Germany, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 296, + 588 + ], + "type": "text", + "content": "[11] Timothy Chen, Ola Shorinwa, Joseph Bruno, Javier Yu, Wei-jia Zeng, Keiko Nagami, Philip Dames, and Mac Schwager. Splat-nav: Safe real-time robot navigation in gaussian splatting maps, 2024. 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 622 + ], + "type": "text", + "content": "[12] Pinxuan Dai, Jiamin Xu, Wenxiang Xie, Xinguo Liu, Huamin Wang, and Weiwei Xu. High-quality surface reconstruction using gaussian surfels, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 656 + ], + "type": "text", + "content": "[13] Matt Deitke, Rose Hendrix, Luca Weihs, Ali Farhadi, Kiana Ehsani, and Aniruddha Kembhavi. Phone2proc: Bringing robust robots into our chaotic world, 2022. 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 295, + 713 + ], + "type": "text", + "content": "[14] Ainaz Eftekhar, Alexander Sax, Jitendra Malik, and Amir Zamir. Omnidata: A scalable pipeline for making multitask mid-level vision datasets from 3d scans. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10786-10796, 2021. 4, 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Antoine Guédon and Vincent Lepetit. Sugar: Surface-aligned gaussian splatting for efficient 3d mesh reconstruction and high-quality mesh rendering, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 553, + 161 + ], + "type": "text", + "content": "[16] Mu Hu, Wei Yin, Chi Zhang, Zhipeng Cai, Xiaoxiao Long, Hao Chen, Kaixuan Wang, Gang Yu, Chunhua Shen, and Shaojie Shen. Metric3d v2: A versatile monocular geometric foundation model for zero-shot metric depth and surface normal estimation. 2024. 4, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 217 + ], + "type": "text", + "content": "[17] Muhammad Zubair Irshad, Chih-Yao Ma, and Zsolt Kira. Hierarchical cross-modal agent for robotics vision-and-language navigation. In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 285 + ], + "type": "text", + "content": "[18] Muhammad Zubair Irshad, Niluthpol Chowdhury Mithun, Zachary Seymour, Han-Pang Chiu, Supun Samarasekera, and Rakesh Kumar. Semantically-aware spatio-temporal reasoning agent for vision-and-language navigation in continuous environments. In 2022 26th International Conference on Pattern Recognition (ICPR), pages 4065-4071, 2022. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[19] Muhammad Zubair Irshad, Sergey Zakharov, Katherine Liu, Vitor Guizilini, Thomas Kollar, Adrien Gaidon, Zsolt Kira, and Rares Ambrus. Neo 360: Neural fields for sparse view synthesis of outdoor scenes. 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 374 + ], + "type": "text", + "content": "[20] Muhammad Zubair Irshad, Mauro Comi, Yen-Chen Lin, Nick Heppert, Abhinav Valada, Rares Ambrus, Zsolt Kira, and Jonathan Tremblay. Neural fields in robotics: A survey. arXiv preprint arXiv:2410.20220, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 441 + ], + "type": "text", + "content": "[21] Abhishek Kadian, Joanne Truong, Aaron Gokaslan, Alexander Clegg, Erik Wijmans, Stefan Lee, Manolis Savva, Sonia Chernova, and Dhruv Batra. Sim2real predictivity: Does evaluation in simulation predict real-world performance? IEEE Robotics and Automation Letters, 5(4):6670-6677, 2020. 1, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 476 + ], + "type": "text", + "content": "[22] Ivan Kapelyukh, Yifei Ren, Ignacio Alzugaray, and Edward Johns. Dream2real: Zero-shot 3d object rearrangement with vision-language models, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[23] Oğuzhan Fatih Kar, Teresa Yeo, Andrei Atanov, and Amir Zamir. 3d common corruptions and data augmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18963-18974, 2022. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 523, + 553, + 567 + ], + "type": "text", + "content": "[24] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4), 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 568, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 568, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 568, + 553, + 634 + ], + "type": "text", + "content": "[25] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Schacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. arXiv preprint arXiv:2306.11290, 2023. 1, 3, 4, 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 678 + ], + "type": "text", + "content": "[26] Jacob Krantz, Stefan Lee, Jitendra Malik, Dhruv Batra, and Devendra Singh Chaplot. Instance-specific image goal navigation: Training embodied agents to find object instances, 2022. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[27] Xiaohan Lei, Min Wang, Wengang Zhou, and Houqiang Li. Gaussnav: Gaussian splatting for visual navigation, 2024. 2, 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25439" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[28] Yulong Li and Deepak Pathak. Object-aware gaussian splatt ing for robotic manipulation. In ICRA 2024 Workshop on 3D Visual Representations for Robot Manipulation, 2024. 3, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 119, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 140 + ], + "type": "text", + "content": "[29] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "text", + "content": "[30] JunEn Low, Maximilian Adang, Javier Yu, Keiko Nagami, and Mac Schwager. Sous vide: Cooking visual drone navigation policies in a gaussian splatting vacuum, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 178, + 294, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 294, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 294, + 211 + ], + "type": "text", + "content": "[31] Arjun Majumdar, Gunjan Aggarwal, Bhavika Devnani, Judy Hoffman, and Dhruv Batra. Zson: Zero-shot object-goal navigation using multimodal goal embeddings, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 213, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 294, + 289 + ], + "type": "text", + "content": "[32] Arjun Majumdar, Karmesh Yadav, Sergio Arnaud, Yecheng Jason Ma, Claire Chen, Sneha Silwal, Aryan Jain, Vincent-Pierre Berges, Pieter Abbeel, Jitendra Malik, Dhruv Batra, Yixin Lin, Oleksandr Maksymets, Aravind Rajeswaran, and Franziska Meier. Where are we in the search for an artificial visual cortex for embodied intelligence?, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 294, + 335 + ], + "type": "text", + "content": "[33] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra Singh Chaplot. Autonerf: Training implicit scene representations with autonomous agents, 2023. 3, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 370 + ], + "type": "text", + "content": "[34] Wugang Meng, Tianfu Wu, Huan Yin, and Fumin Zhang. Beings: Bayesian embodied image-goal navigation with gaussian splatting, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 415 + ], + "type": "text", + "content": "[35] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis, 2020. 3, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 418, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 294, + 460 + ], + "type": "text", + "content": "[36] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph., 41(4):102:1-102:15, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "text", + "content": "[37] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. UniDepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 520, + 261, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 261, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 261, + 532 + ], + "type": "text", + "content": "[38] Polycam. Polycam. Accessed: 2024-11-10. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 533, + 294, + 599 + ], + "type": "text", + "content": "[39] Xavier Puig, Eric Undersander, Andrew Szot, Mikael Dallaire Cote, Tsung-Yen Yang, Ruslan Partsey, Ruta Desai, Alexander Clegg, Michal Hlavac, So Yeon Min, et al. Habitat 3.0: A co-habitat for humans, avatars, and robots. In The Twelfth International Conference on Learning Representations, 2023. 2, 3, 4, 5, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[40] Mohammad Nomaan Qureshi, Sparsh Garg, Francisco Yandun, David Held, George Kantor, and Abhisesh Silwal. Splatsim: Zero-shot sim2real transfer of rgb manipulation policies using gaussian splatting, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[41] Santhosh K. Ramakrishnan, Aaron Gokaslan, Erik Wijmans, Oleksandr Maksymets, Alex Clegg, John Turner, Eric Undersander, Wojciech Galuba, Andrew Westbury, Angel X. Chang, Manolis Savva, Yili Zhao, and Dhruv Batra. Habitatmatterport 3d dataset (hm3d): 1000 large-scale 3d environments for embodied ai, 2021. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[42] Xuqian Ren, Wenjia Wang, Dingding Cai, Tuuli Tuominen, Juho Kannala, and Esa Rahtu. Mushroom: Multi-sensor hybrid room dataset for joint 3d reconstruction and novel view synthesis, 2024. 3, 4, 6, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[43] Meta AI Research. Home robot. https://github.com/facebookresearch/home-robot, 2023. GitHub repository. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 204 + ], + "type": "text", + "content": "[44] Manolis Savva, Abhishek Kadian, Oleksandr Maksymets, Yili Zhao, Erik Wijmans, Bhavana Jain, Julian Straub, Jia Liu, Vladlen Koltun, Jitendra Malik, Devi Parikh, and Dhruv Batra. Habitat: A platform for embodied ai research, 2019. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 237 + ], + "type": "text", + "content": "[45] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms, 2017. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "text", + "content": "[46] Sneha Silwal, Karmesh Yadav, Tingfan Wu, Jay Vakil, Arjun Majumdar, Sergio Arnaud, Claire Chen, Vincent-Pierre Berges, Dhruv Batra, Aravind Rajeswaran, Mrinal Kalakrishnan, Franziska Meier, and Oleksandr Maksymets. What do we learn from a large-scale study of pre-trained visual representations in sim and real environments?, 2024. 1, 2, 3, 5, 7, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 413 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 413 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 413 + ], + "type": "text", + "content": "[47] Andrew Szot, Alex Clegg, Eric Undersander, Erik Wijmans, Yili Zhao, John Turner, Noah Maestre, Mustafa Mukadam, Devendra Chaplot, Oleksandr Maksymets, Aaron Gokaslan, Vladimir Vondrus, Sameer Dharur, Franziska Meier, Wojciech Galuba, Angel Chang, Zsolt Kira, Vladlen Koltun, Jitendra Malik, Manolis Savva, and Dhruv Batra. Habitat 2.0: Training home assistants to rearrange their habitat. In Advances in Neural Information Processing Systems (NeurIPS), 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 480 + ], + "type": "text", + "content": "[48] Matthew Tancik, Ethan Weber, Evonne Ng, Ruilong Li, Brent Yi, Justin Kerr, Terrance Wang, Alexander Kristoffersen, Jake Austin, Kamyar Salahi, Abhik Ahuja, David McAllister, and Angjoo Kanazawa. Nerfstudio: A modular framework for neural radiance field development. In ACM SIGGRAPH 2023 Conference Proceedings, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 481, + 553, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 481, + 553, + 524 + ], + "spans": [ + { + "bbox": [ + 316, + 481, + 553, + 524 + ], + "type": "text", + "content": "[49] Matias Turkulainen, Xuqian Ren, Iaroslav Melekhov, Otto Seiskari, Esa Rahtu, and Juho Kannala. Dn-splatter: Depth and normal priors for gaussian splattering and meshing, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 553, + 579 + ], + "type": "text", + "content": "[50] Erik Wijmans, Abhishek Kadian, Ari Morcos, Stefan Lee, Irfan Essa, Devi Parikh, Manolis Savva, and Dhruv Batra. Dppo: Learning near-perfect pointgoal navigators from 2.5 billion frames. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 612 + ], + "type": "text", + "content": "[51] Yaniv Wolf, Amit Bracha, and Ron Kimmel. Gs2mesh: Surface reconstruction from gaussian splatting via novel stereo views, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "text", + "content": "[52] Yuxuan Wu, Lei Pan, Wenhua Wu, Guangming Wang, Yanzi Miao, and Hesheng Wang. Rl-gsbridge: 3d gaussian splating based real2sim2real method for robotic manipulation learning, 2024. 3, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[53] Fei Xia, Amir Zamir, Zhi-Yang He, Alexander Sax, Jitendra Malik, and Silvio Savarese. Gibson env: Real-world perception for embodied agents, 2018. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[54] Hongchi Xia, Zhi-Hao Lin, Wei-Chiu Ma, and Shenlong Wang. Video2game: Real-time, interactive, realistic and" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "25440" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 385 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "browser-compatible environment from a single video, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[55] Karmesh Yadav, Ram Ramrakhya, Arjun Majumdar, Vincent-Pierre Berges, Sachit Kuhar, Dhruv Batra, Alexei Baevski, and Oleksandr Maksymets. Offline visual representation learning for embodied navigation, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 184 + ], + "type": "text", + "content": "[56] Karmesh Yadav, Arjun Majumdar, Ram Ramrakhya, Naoki Yokoyama, Alexei Baevski, Zsolt Kira, Oleksandr Maksymets, and Dhruv Batra. Ovrl-v2: A simple state-of-art baseline forImagenav and objectnav, 2023. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 218 + ], + "type": "text", + "content": "[57] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 306 + ], + "type": "text", + "content": "[58] Justin Yu, Kush Hari, Kishore Srinivas, Karim El-Refai, Adam Rashid, Chung Min Kim, Justin Kerr, Richard Cheng, Muhammad Zubair Irshad, Ashwin Balakrishna, et al. Language-embedded gaussian splats (legs): incrementally building room-scale representations with a mobile robot. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13326–13332. IEEE, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 363 + ], + "type": "text", + "content": "[59] Justin Yu, Kush Hari, Karim El-Refai, Arnav Dalil, Justin Kerr, Chung-Min Kim, Richard Cheng, Muhammad Z. Irshad, and Ken Goldberg. Persistent object gaussian splat (pogs) for tracking human and robot manipulation of irregularly shaped objects. ICRA, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 385 + ], + "type": "text", + "content": "[60] Siting Zhu, Guangming Wang, Dezhi Kong, and Hesheng Wang. 3d gaussian splatting in robotics: A survey, 2024. 2" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "25441" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_content_list.json b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d614321a0d7e05e6c311992539926edcf36b5662 --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_content_list.json @@ -0,0 +1,1872 @@ +[ + { + "type": "text", + "text": "EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model", + "text_level": 1, + "bbox": [ + 187, + 130, + 810, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shengqi Dang $^{1,2\\dagger}$ , Yi He $^{1\\dagger}$ , Long Ling $^{1}$ , Ziqing Qian $^{1}$ , Nanxuan Zhao $^{3}$ , Nan Cao $^{1,2*}$", + "bbox": [ + 156, + 202, + 836, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Tongji University $^{2}$ Shanghai Innovation Institute $^{3}$ Adobe Research", + "bbox": [ + 212, + 237, + 779, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9dc9533db4779021a908e9427fceb336619f5bdf630b50450e8af0d377fc860.jpg", + "image_caption": [ + "Figure 1. Generated images of the EmotiCrafter. Given a prompt and Valence-Arousal (V-A) values, our method can generate an emotional image that reflects the input prompt and aligns with the specified emotion values. By adjusting the V-A values, our method can also generate images that evoke discrete emotions, such as sadness, relaxation, anger, and amusement." + ], + "image_footnote": [], + "bbox": [ + 93, + 292, + 517, + 636 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55dce0c3547afc774b1fb55f839f4a82b439cc06a2e13168250b292622a033ff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 315, + 705, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/797cad69fe04c21a3acb13042d7e645b30333b9fab05414714b18187e0604097.jpg", + "image_caption": [ + "A man is running fast", + "A cartoon shows children playing football" + ], + "image_footnote": [], + "bbox": [ + 517, + 460, + 704, + 623 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6207cb4e237e7a95d06c221e56f1eded15b1e3f069220485469848612e8967ac.jpg", + "image_caption": [ + "A group of people are reading books" + ], + "image_footnote": [], + "bbox": [ + 709, + 315, + 900, + 460 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/133bea0b51c8c34e94453b5de36915ec32431c45a14430b101e0d7aeccff4e3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 477, + 901, + 623 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 700, + 326, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent research shows that emotions can enhance users' cognition and influence information communication. While", + "bbox": [ + 89, + 733, + 485, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "research on visual emotion analysis is extensive, limited work has been done on helping users generate emotionally rich image content. Existing work on emotional image generation relies on discrete emotion categories, making it challenging to capture complex and subtle emotional nuances accurately. Additionally, these methods struggle to control the specific content of generated images based on text prompts. In this paper, we introduce the task of continuous emotional image content generation (C-EICG) and present EmotiCrafter, a general emotional image generation model that generates images based on free text prompts and Valence-Arousal (V-A) values. It leverages a novel emotion-embedding mapping network to fuse V-A values", + "bbox": [ + 511, + 702, + 908, + 898 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Shengqi Dang and Yi He contributed equally to this work.", + "bbox": [ + 114, + 779, + 426, + 792 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\* Nan Cao is the corresponding author.", + "bbox": [ + 116, + 792, + 318, + 804 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Shengqi Dang, Yi He, Long Ling, Ziqing Qian, and Nan Cao are with the Intelligent Big Data Visualization Lab, Tongji University. Shengqi Dang and Nan Cao are also with the Shanghai Innovation Institute. Email: {dangsq123, heyi_11}@tongji.edu.cn, lucyling0224@gmail.com, 2411920@tongji.edu.cn, nan.cao@gmail.com.", + "bbox": [ + 91, + 804, + 482, + 863 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Nanxuan Zhao is with Adobe Research. Email: nanxuanzhao@g-mail.com.", + "bbox": [ + 91, + 864, + 480, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Code: https://github.com/idvxlab/EmotiCrafter", + "bbox": [ + 114, + 887, + 465, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "15218", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "into textual features, enabling the capture of emotions in alignment with intended input prompts. A novel loss function is also proposed to enhance emotion expression. The experimental results show that our method effectively generates images representing specific emotions with the desired content and outperforms existing techniques.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 210, + 222, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Emotions are fundamental to human experiences and play a critical role in shaping how people perceive and interact with the world. Research has shown that emotions affect memory [15, 19, 25, 45] and comprehension [16, 38, 40], which are crucial for effective communication. As a result, content creators increasingly recognize the importance of incorporating emotions to enhance audience engagement.", + "bbox": [ + 89, + 236, + 483, + 340 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While research on visual emotion analysis is extensive [13, 49, 58], there is limited work on generating emotionally rich image content. Some early studies explored emotional content generation techniques within specific domains such as facial expressions [2, 46] or landscapes [24]. EmoGen [50] generates images based on a given emotion tag (e.g., happy or sad), which pioneered the domain-free emotional image content generation (EICG) task. However, it has two critical limitations: (1) the image is generated from an emotion tag instead of a text prompt, making the generated content difficult to control; (2) while the discrete emotion tags used in EmoGen are easy to understand, psychologists have not achieved a consensus on the emotion categories [60]. The limited scope of discrete emotion tags falls short of capturing nuanced emotions.", + "bbox": [ + 89, + 342, + 483, + 566 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issues, we propose the continuous emotional image content generation (C-EICG) task and present EmotiCrafter, the first C-EIGC model that generates emotional images using free-text prompts and continuous Valence-Arousal (V-A) values defined in the V-A model (a well-known psychological continuous emotion model) [33]. V-A model represents emotions in a two-dimensional Cartesian space (Figure 2), where Valence quantifies pleasantness (negative to positive) and Arousal measures intensity (calm to excited). We utilize a [-3, 3] range for Valence and Arousal [10, 21]. The continuous V-A space enables smooth transitions and nuanced emotional shifts beyond the capability of discrete labels (e.g., shifts from \"bored\" to \"tired\" of a character's state in a video). Specific emotional values in this space have been investigated in prior work [34], and such fine-grained modeling is particularly beneficial for human-computer interaction [2]. Leveraging this model, EmotiCrafter captures subtle affective variations via precise (V, A) positioning. The contributions of this paper are as follows:", + "bbox": [ + 89, + 568, + 483, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a novel task, continuous emotional image content generation (C-EICG), and develop the first ded", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/9624c89c88a2a32c6ec937c2b1a82f28ed6403ab72b50c5a17fd755a11f5f52c.jpg", + "image_caption": [ + "Figure 2. Valence-Arousal model." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 268 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "icated model for this purpose. Our model introduces an emotion-embedding network that integrates continuous Valence-Arousal (V-A) values with text prompts. These fused features are then injected into Stable Diffusion XL [28] using cross-attention mechanisms, enabling precise control over both content and emotional expressions in the generated images.", + "bbox": [ + 524, + 316, + 903, + 422 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel loss function that enhances the emotional resonance of generated images. By amplifying the difference between neutral and emotional text features, our approach enables the model to capture more distinct emotional variations. Additionally, the loss function incorporates the V-A distribution to address data imbalance, further refining the model's ability to generate images with rich, accurate emotional expressions.", + "- We constructed an emotional prompts dataset to train the emotion-embedding network, where each sample consists of a neutral prompt and an emotional prompt that share the same core meaning but express a specific emotion corresponding to a given pair of V-A values." + ], + "bbox": [ + 511, + 424, + 903, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 633, + 651, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we present a review of the related work, specifically focusing on visual emotion analysis, image emotion transfer, and conditional image generation.", + "bbox": [ + 511, + 657, + 903, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Visual Emotion Analysis", + "text_level": 1, + "bbox": [ + 511, + 712, + 736, + 728 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Visual emotion analysis refers to the computational recognition and interpretation of human emotion [57] in visual media, such as images or videos. It has been a prominent research area, with most efforts focusing on the classification of discrete emotions [4, 42, 47, 48, 52]. However, discrete emotion categories limit the ability to capture the nuanced emotions [7], leading to an increased focus on continuous emotion analysis in images [59].", + "bbox": [ + 511, + 734, + 905, + 854 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Much of the continuous emotion analysis remains centered on facial expression analysis [9, 11, 22, 37, 54]. While effective, facial analysis can overlook crucial con", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "15219", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "textual information that influences emotional interpretation. This drives studies to focus on emotions in objects or individuals within their environment, rather than just faces[12, 13, 20, 21]. For example, Kosti et al. [12] combined person-specific characteristics with the context of the scene to predict continuous emotional dimensions such as arousal, valence, and dominance. Kragel et al. [13] proposed EmoNet to extract visual features such as facial expressions, body posture, and scene elements from images to predict valence and arousal values. Recently, Mertens et al.[21] tested multiple backbones (such as ResNet, CLIP, DINO, etc.) for prediction of valence and arousal, demonstrating strong performance across various architectures.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous studies have achieved high accuracy in predicting continuous emotions in images, demonstrating a correlation between continuous emotions and visual elements. Building on this insight, our work moves beyond prediction to explore how continuous emotions can be actively embedded within generated content.", + "bbox": [ + 89, + 287, + 480, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Image Emotion Transfer", + "text_level": 1, + "bbox": [ + 89, + 390, + 316, + 405 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image Emotion Transfer (IET) focuses on editing the content of the images to evoke different emotions [17, 27, 43, 61]. For instance, Peng et al. [27] achieved emotion transfer by adjusting color tones and texture-related features. Zhu et al. [61] introduced a method that separates high-level emotion-relevant features (e.g., object shapes and scene layout) from low-level emotion-relevant features (e.g., brightness). By applying GANs, they transferred emotions between images while preserving their original structure. Building on these methods, Weng et al. [43] proposed the Affective Image Filter, which uses a multi-modal transformer to process both text and image inputs. With the emergence of text-to-image models, IET has expanded into new applications based on instructive commands. For example, EmoEdit [51] used GPT-4V to build emotion factor trees that map abstract emotions to specific visual elements and employed the InstructPix2Pix model to apply emotion-driven content and color adjustments to images.", + "bbox": [ + 89, + 412, + 482, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current methods primarily extract specific features, focusing on certain visual elements or emotional cues, which can limit the depth of emotional expression. In contrast, our method broadly learns a variety of emotion-influencing features, and it could accept natural language prompts as input.", + "bbox": [ + 89, + 685, + 482, + 761 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Conditional Image Generation", + "text_level": 1, + "bbox": [ + 89, + 771, + 362, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditional image generation aims to create images that align with specific input conditions, such as text [3, 28, 30, 31], reference images [53], subjects [6, 32], and depth maps [23, 55]. To enhance quality, researchers have developed specialized approaches, such as Diffusion Transformers (DiT)[26], which use transformer-based diffusion for denoising, and Visual Autoregressive Modeling (VAR)[36],", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which encodes images into discrete tokens for autoregressive prediction across scales. However, these state-of-the-art methods rely on discrete labels, limiting flexibility and control. In contrast, our approach offers greater generalization by leveraging Stable Diffusion XL [28] to generate emotional images from free-text prompts.", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite these progresses, incorporating emotion as a condition for image generation remains underexplored. EmoGen [50] pioneered the generation of emotional image content (EIGC) by mapping emotional features to semantic features to generate emotional images. However, EmoGen struggles to understand natural language, limiting its capacity to effectively control specific content. Its reliance on discrete emotions also limits its practical applicability.", + "bbox": [ + 511, + 181, + 903, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method bridges this gap by embedding continuous emotion into textual features, enabling image generation models to use continuous emotion for emotion control. Unlike label-based methods, our method supports free-text prompts for flexible content control.", + "bbox": [ + 511, + 303, + 903, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 392, + 604, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce the technique details of the proposed EmotiCrafter.", + "bbox": [ + 511, + 417, + 903, + 448 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 458, + 624, + 472 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our method generates emotional images $I_{emo}$ from two inputs (Figure 3): a free-text prompt describing the desired content, and a pair of V-A values $(v, a)$ specifying the emotion. First the prompt encoder $\\mathcal{E}$ converts the text prompt into feature $f_{n}$ , which is then processed by an emotion-embedding network $\\mathcal{M}$ to produce emotional prompt feature $\\hat{f}_{e}$ that integrate the V-A values:", + "bbox": [ + 511, + 479, + 905, + 585 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {e} = \\mathcal {M} \\left(f _ {n} | (v, a)\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 595, + 903, + 616 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Next, this feature is injected into Stable Diffusion XL $\\mathcal{G}$ via its cross-attention mechanism to generate the emotional images: $I_{emo} = \\mathcal{G}(\\hat{f}_e)$ . To enhance emotional expressiveness, we introduce a loss function (Fig.3(b.2)) that leverages the V-A distribution and emphasizes the differences between neutral and emotional prompt features. This ensures that the generated images accurately convey both the intended emotions and content. Additionally, we construct a dataset (Fig.3(a)) that pairs neutral and emotional prompts with the corresponding V-A values.", + "bbox": [ + 511, + 626, + 905, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Emotion-Embedding Network", + "text_level": 1, + "bbox": [ + 511, + 787, + 782, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The emotion-embedding network $\\mathcal{M}$ generates the emotional prompt feature by integrating a pair of V-A values with a neutral prompt feature (Fig.3(b.1)). First, a V-A encoder converts the V-A values into feature vectors. Then, an emotion injection transformer—modified from GPT-2[29]—fuses these vectors with the neutral prompt", + "bbox": [ + 511, + 810, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "15220", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b6efb014286d027a882699f6303429d453f12f20c4e9240f0d8501500385d5dd.jpg", + "image_caption": [ + "Figure 3. Overview of our method. Specifically, we take the following steps. (a) We collect an image dataset annotated with V-A values, neutral prompts, and emotional prompts. These prompts are then encoded into features by prompt encoder $\\mathcal{E}$ . (b) Next, we design (b.1) an emotion-embedding network $\\mathcal{M}$ to embed V/A values into textual features based on the transformer architecture, and (b.2) a specialized loss function to enhance the emotional resonance of generated images. The output of the mapping network serves as the condition for the image generation model $\\mathcal{G}$ to generate emotional images." + ], + "image_footnote": [], + "bbox": [ + 96, + 89, + 903, + 277 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86bc912c030cf236269f88e37e87c6e9258c5e203b081b03668ec9abea346b32.jpg", + "image_caption": [ + "Figure 4. Structure of Emotion Injection Block. It accepts hidden state $h_{i-1}$ as input and produces $h_i$ as output. The V-feature $e_v$ and A-feature $e_a$ represent the emotion features, which are injected through the cross-attention module." + ], + "image_footnote": [], + "bbox": [ + 96, + 364, + 478, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "feature, preserving the original textual context while infusing emotional content.", + "bbox": [ + 89, + 529, + 482, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "V-A Encoder. The V-A Encoder converts a pair of V-A values into feature vectors using two separate multilayer perceptrons (MLPs). One MLP processes the Valence value to produce V-feature $e_v$ , and the other processes the Arousal value to generate A-feature $e_a$ . These features are then fed into the emotion injection transformer network for further emotion infusion.", + "bbox": [ + 89, + 560, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Emotion Injection Transformer (EIT). The Emotion Injection Transformer (EIT) leverages a modified GPT-2 architecture to seamlessly integrate V/A-features into textual features. Its process consists of three stages: input projection, emotion injection, and output projection.", + "bbox": [ + 89, + 665, + 482, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "First, we project the input neutral prompt feature $f_{n}$ into the transformer's feature space:", + "bbox": [ + 89, + 739, + 482, + 770 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {0} = \\mathrm {P} _ {\\text {i n}} \\left(f _ {n}\\right) + \\mathrm {P E}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 777, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_0$ represents the initial hidden state; $\\mathrm{P_{in}(\\cdot)}$ is a linear projection layer, and PE denotes positional embedding [39].", + "bbox": [ + 89, + 801, + 482, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the next, we inject emotion into $f_{n}$ via 12 sequential Emotion Injection Blocks (EIBs) corresponding to 12 transformer blocks. Each block outputs a hidden state:", + "bbox": [ + 89, + 832, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\operatorname {E I B} \\left(h _ {i - 1}, e _ {v}, e _ {a}\\right), \\quad i \\in \\{1, \\dots , 1 2 \\} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 885, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h_i$ is the output of the $i$ -th EIB( $\\cdot$ ). As shown in Figure 4, each EIB enhances the transformer block through a cross-attention mechanism:", + "bbox": [ + 511, + 366, + 906, + 411 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {\\prime} = \\text {s e l f - a t t n} (\\mathrm {L N} (h _ {i - 1})) + h _ {i - 1} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 420, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {(v)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {\\prime}\\right), e _ {v}\\right) + h _ {i} ^ {\\prime} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 440, + 903, + 460 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} ^ {(v, a)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v)}\\right), e _ {a}\\right) + h _ {i} ^ {(v)} \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 462, + 903, + 483 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nh _ {i} = \\operatorname {f n n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v, a)}\\right)\\right) + h _ {i} ^ {(v, a)} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 484, + 903, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $h^{\\prime}, h^{(v)}, h^{(v,a)}$ are the intermediate hidden variables; $\\mathrm{LN}(\\cdot)$ is the LayerNorm; self-attn( $\\cdot$ ) denotes self-attention, employed to capture context dependencies; cross-attn( $\\cdot$ ) is cross-attention for injecting $e_v$ and $e_a$ ; $\\mathrm{fnn}(\\cdot)$ is a feedforward network that adapts the complexity of the emotional embedding process. We also remove the causal mask typically used for next-token prediction from the original transformer model to fit our task.", + "bbox": [ + 511, + 513, + 905, + 635 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, the output of the last (i.e., the 12th) injection block $h_{12}$ is projected back to SDXL's prompt feature space via $\\mathbf{P}_{\\mathrm{out}}$ (a linear layer and a LayerNorm layer) to predict the residual between emotional and neutral prompt features, which represents a semantic shift between emotional and neutral prompts:", + "bbox": [ + 511, + 636, + 905, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {r} = \\mathrm {P} _ {\\text {o u t}} \\left(\\mathrm {L N} \\left(h _ {1 2}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 734, + 903, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final emotional prompt feature is obtained by adding this residual to the original neutral prompt feature:", + "bbox": [ + 511, + 760, + 906, + 791 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {e} = \\hat {f} _ {r} + f _ {n} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 797, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The above emotion embedding network is trained by minimizing the averaged expectation of the difference between the predicted emotional prompt feature $\\hat{f}_e = \\mathcal{M}(f_n|(v,a))$ and the scaled target emotional prompt feature $f_e^t$ , using the loss function described in Equation 10.", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "15221", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\frac {1}{n} \\mathbb {E} \\left(\\frac {1}{d (v , a)} \\| \\hat {f} _ {e} - f _ {e} ^ {t} \\| ^ {2}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 102, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $n$ is the number of feature elements; $\\mathbb{E}(\\cdot)$ denotes the expectation; $d(v, a)$ is a density function that describes the distribution of V-A values in the training sample.", + "bbox": [ + 89, + 140, + 482, + 185 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively embed emotions and address the challenges posed by the uneven distribution of V-A values in the dataset, this loss function incorporates two key strategies to improve the model's performance:", + "bbox": [ + 89, + 185, + 482, + 246 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Scaled Residual Learning. To better capture pronounced emotional changes in generated images, we enlarge the target residuals:", + "bbox": [ + 89, + 246, + 482, + 291 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {e} ^ {t} = f _ {n} + \\alpha \\underbrace {\\left(f _ {e} - f _ {n}\\right)} _ {\\text {r e s i d u a l f e a t u r e}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 301, + 482, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{e}$ is the emotional prompt feature, $f_{n}$ is the neutral prompt feature, and $\\alpha$ is a scale factor, we set its value to 1.5 based on the ablation study.", + "bbox": [ + 89, + 345, + 482, + 391 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "V-A Density Weighting. To mitigate the effects of the imbalanced distribution of the training samples, we weigh the loss inversely proportional to the density of training samples in the V-A space. The density is estimated using Kernel Density Estimation (KDE) [5] with a Gaussian kernel, denoted as $d(v,a)$ :", + "bbox": [ + 89, + 391, + 483, + 482 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nd (v, a) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} K _ {H} ((v, a) - (v _ {i}, a _ {i})), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 494, + 482, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $K_{H}$ is a 2D Gaussian kernel with bandwidth $H$ ; $n$ is the number of training samples; $(v_{i},a_{i})$ are the V-A values of the $i$ -th training sample. The bandwidth $H$ is selected using Silverman's rule of thumb to provide optimal smoothing of the density estimation.", + "bbox": [ + 89, + 537, + 483, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Dataset and Training", + "text_level": 1, + "bbox": [ + 89, + 622, + 290, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To train the emotion-embedding network, we constructed a dataset of paired neutral and emotional prompts with corresponding Valence-Arousal (V-A) values (Figure 3(a)). These pairs were automatically generated using GPT-4 based on 39,843 images with human-annotated V-A values from publicly available datasets, including OASIS [14], EMOTIC [12], and FindingEmo [21]. Specifically, GPT-4 was used to generate neutral prompts with objective image descriptions and emotional prompts that emphasize affective attributes such as color, lighting, and texture, which influence emotional perception. To ensure data reliability, all LLM-generated prompts underwent crowd-sourced verification, with disagreements resolved through a voting mechanism among annotators.", + "bbox": [ + 89, + 643, + 482, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The proposed emotion embedding network is trained on two NVIDIA A800 GPUs using the aforementioned dataset. We employ the AdamW [18] optimizer with a weight decay", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of 1e-5 and a learning rate of 1e-3. The training process spans 200 epochs with a batch size of 768, completing in approximately 7 to 8 hours.", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 511, + 152, + 630, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Generation Results", + "text_level": 1, + "bbox": [ + 511, + 176, + 697, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 1 shows the proposed technique's ability to achieve continuous and effective control over both emotion and content during image generation. Meanwhile, Figure 5 highlights four key capabilities: (a) emotion-content decoupling, where V-A values override emotional cues in the prompt, allowing typically positive concepts to be rendered with negative emotions; (b) compatibility with discrete emotion categories; (c) content-independent generation, where images generated from empty prompts and specified V-A values maintain emotional consistency without semantic constraints; and (d) fine-grained emotional control, demonstrated through V-A increments of 0.2, showcasing the model's sensitivity to subtle emotional variations.", + "bbox": [ + 511, + 200, + 906, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 407, + 651, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To estimate the effectiveness of the proposed method, we built four baselines based on existing techniques.", + "bbox": [ + 511, + 430, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We established baselines for comparison using two strategies: (1) directly injecting emotion features into the image generation modules of SDXL, such as UNet, and (2) modifying the input text prompt using emotion features to influence the generated image's content, similar to the proposed method.", + "bbox": [ + 511, + 460, + 905, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As a result, four different baselines were built: (1) Cross Attention: inject the emotion features $(e_v, e_a)$ into the UNet in SDXL via its cross-attention mechanism based on IP-Adapter [53]; (2) Time Embedding: directly add emotion features $(e_v, e_a)$ to the time embedding of the UNet in SDXL. (3) Textual Inversion: use the text inversion technique [6] to embed emotion features $(e_v, e_a)$ into prompt templates with predefined emotion placeholders. (4) GPT-4+SDXL (GPT-SD): use GPT-4 [1] to rewrite the input text according to (v,a) values to generate an emotional SDXL prompt for image generation.", + "bbox": [ + 511, + 551, + 905, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Comparison. We evaluate the generated images based on three criteria: (1) the effectiveness of emotion embedding, (2) image-prompt similarity, and (3) the continuity of emotional variations as V-A values change. Figure 6 demonstrates that the baseline methods—Cross Attention, Time Embedding, and Textual Inversion—tend to produce nearly identical outputs regardless of emotional variation. This is because the loss terms of these baselines primarily align low-level image features (e.g., SDXL's latent space), which are highly correlated with prompt content but struggle to capture subtle emotional cues when both prompts and V-A values are provided. Notably, Tex-", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "15222", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bba03a1446d97cabe0e071e1851f7434ab9605a8260b4e84c5e83090ecf6fd34.jpg", + "image_caption": [ + "(a)", + "(b)" + ], + "image_footnote": [], + "bbox": [ + 125, + 88, + 210, + 155 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8c02341b7eb8fe02f47bdb2914dcda120616909acc139cea56b3820fc8bced53.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 241, + 88, + 475, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c3ea3c2ca75d16339a200ea498be0e290ed9f8f7898b6df493694aeed647d402.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 488, + 88, + 571, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7246a3cd59b6b304cac1ac54cf48949b2ece02cb95fb761f3d677a85d52e2f62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 573, + 89, + 736, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/04b476acc97c862b300ac66bd4d4db2f86f5a04a77cef2867e7a552d59f2c69e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 89, + 903, + 152 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4a7e7a5c16f462a4c1676792605a3a2c0e990deedbc06e8892190c2ffaecfe3f.jpg", + "image_caption": [ + "1)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 125, + 157, + 459, + 223 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/741da1258ba440b8884678442a21ec33513b45280af59afc71bb6eecba5e314b.jpg", + "image_caption": [ + "$\\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : \\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = x$" + ], + "image_footnote": [], + "bbox": [ + 509, + 157, + 653, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/26f2a0cb69f8bb32498994df30781659074ec362a6b4b19631d06fa9a3e3a91f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 653, + 157, + 736, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75770b5fdbd7dade7d4f1cbd2317355a6922ed0bb85ff199c0feb74a9546c741.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 736, + 157, + 821, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/853b3c65492d661868b046a816bdb95e1f56d2909eb2e5e6ac93f8ac86b60afe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 821, + 157, + 903, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f17bff65dc980cc74376cc6e18e6e732917c86e995f3106e793171699f33ebc.jpg", + "image_caption": [ + "Figure 5. Results under multiple inputs. (a) Overriding semantic content ('a child in the amusement park') with sad V-A (-2,-2); (b) Discrete emotion mapping in V-A space as emotion input; (c) Empty-prompt generation with pure emotion condition; (d) Fine-grained control of V-A variations with a granularity of 0.2.", + "Figure 6. Qualitative comparisons with baselines. These images are generated at varying V-A values, specifically -1.5, 0, and 1.5. Only our approach and the GPT-4+SDXL successfully generate images that clearly reflect emotional variations. Notably, our results show enhanced continuity, indicating superior controllability over continuous V-A values compared to the GPT-4+SDXL." + ], + "image_footnote": [], + "bbox": [ + 117, + 277, + 879, + 554 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "tual Inversion often produces images with a persistent purple tint. Furthermore, all methods faithfully generate content aligned with the given prompts. However, for image continuity, we focus on comparing our method with GPT-4+SDXL, as they are the only two capable of generating distinct emotional variations. Figure 7 provides a more detailed comparison, showing that our method maintains smooth emotional transitions even under extreme V-A conditions, whereas GPT-4+SDXL introduces noticeable discontinuities (e.g., in the $\\mathrm{V} = 3$ column).", + "bbox": [ + 89, + 619, + 482, + 770 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Comparison. We compare our method against several baselines using the following metrics: (1) $\\underline{\\mathrm{V} / \\mathrm{A} - \\mathrm{Error}}$ evaluates the absolute error between the predicted V/A of the generated images and the input V/A. (2) CLIPScore [8] assesses the similarity between the input text and the generated images. (3) CLIP-IQA [41] leverages a pre-trained CLIP model to evaluate image quality without requiring reference images. (4) LPIPS-Continuous utilizes", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b2ae7573125fb0afe061cdf77f2d4a0072ccbe10815c6d272f441480cf24d30e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
A-Error ↓V-Error ↓CLIPScore ↑CLIP-IQA ↑
Cross Attention1.923±1.1532.080±1.43826.266±2.3810.949±0.046
Time Embedding1.941±1.1682.031±1.34826.566±2.1250.786±0.164
Textual Inversion1.958±1.1881.923±1.17022.346±3.5940.370±0.111
GPT-4+SDXL1.860±1.0901.517±1.06025.907±1.9490.906±0.066
Ours1.828±1.0851.510±1.07423.067±2.6550.881±0.099
", + "bbox": [ + 514, + 617, + 903, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparison on emotion accuracy, prompt fidelity, and image quality across different baselines, evaluated on 3,300 images per method (132 prompts $\\times$ 5 V values $\\times$ 5 A values). Our method achieves the highest performance in emotion accuracy while maintaining comparable results in prompt fidelity and image quality. The slight decrease in prompt fidelity is expected, as modifying emotional content affects semantic alignment.", + "bbox": [ + 511, + 685, + 906, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the Learned Perceptual Image Patch Similarity [56] to measure the continuity of the change in image as V/A changes.", + "bbox": [ + 511, + 806, + 903, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, our method achieves the lowest (best) $\\underline{V / A - Error}$ on average. While Cross Attention and Time Embedding achieve the highest $\\underline{CLIP - IQA}$ and $\\underline{CLIPScore}$ , respectively, these methods fail to generate", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "15223", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/16012fcf819496df2d55baca64f95cd2d704ca4f067f84b38ee5f5ed4b266430.jpg", + "image_caption": [ + "Figure 7. A more comprehensive comparison with the GPT-4+SDXL for the prompt \"A group of people is skiing in a snow hill.\" Our approach maintains continuity even under extreme V/A conditions. Conversely, GPT-4+SDXL displays noticeable discontinuities (e.g., in the $\\mathrm{V} = 3$ column)." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 480, + 670 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "emotionally expressive images (Figure 6).", + "bbox": [ + 89, + 776, + 372, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Our method exhibits a slight decrease in CLIPScore compared to the baselines, which we attribute to the inherent trade-off between emotional modulation and strict semantic alignment. However, the high CLIP-IQA score indicates that our method produces high-quality images. Additionally, as shown in Table 2, our approach demonstrates superior continuity compared to GPT-4 + SDXL.", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/be69039cc467c64114134922759098bbc41242868f6712ebd8c09286548a03d4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OursGPT-4+SDXL
LPIPS-Continuous↓0.220±0.0640.361±0.059
", + "bbox": [ + 535, + 87, + 880, + 119 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45564c4c38295c6a329a9f7fb1eca106ca0c88452c74fb41afc51ae21b2be65d.jpg", + "table_caption": [ + "Table 2. Continuity comparison between our method and baseline." + ], + "table_footnote": [], + "table_body": "
OursGPT-4+SDXL
Study IA-Ranking Consistency ↑0.759±0.2730.165±0.379
V-Ranking Consistency ↑0.887±0.2450.584±0.259
A-Error ↓1.327±1.1202.029±1.446
V-Error ↓0.692±0.6821.229±1.026
Study IIEmotion Consistency ↑4.215±0.7153.525±1.065
Emotion Smoothness ↑4.240±0.8283.195±1.163
", + "bbox": [ + 514, + 141, + 919, + 247 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Our method outperformed the baseline.", + "bbox": [ + 563, + 253, + 854, + 266 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/652a060b2d3ed81b304f1994db701d6714a9b96b18dbc10ef2e6c96670ef485e.jpg", + "image_caption": [ + "Figure 8. User Study Results ( $^{*}\\mathsf{p} < 0.05$ $^{**}\\mathsf{p} < 0.01$ $^{**}p < 0.001$" + ], + "image_footnote": [], + "bbox": [ + 514, + 276, + 903, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. User Study", + "text_level": 1, + "bbox": [ + 513, + 431, + 635, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conducted two user studies with 20 college students to evaluate the effectiveness of our method by comparing it to $GPT-4 + SDXL$ (the baseline).", + "bbox": [ + 511, + 455, + 906, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Study I. In the first experiment, we evaluated whether the generated images' emotions align with human perception. Two image collections were prepared—one for Arousal (A) and one for Valence (V)—each comprising 20 sets (10 from our method and 10 from the baseline) of 5 randomized images with varying A or V values. Participants reordered the images by perceived intensity and estimated each image's V and A values. We then computed Kendall's $\\tau_{b}$ (with $\\tau_{b} = 1$ indicating perfect alignment) to assess ordering accuracy and calculated the absolute error between the estimated and ground truth values.", + "bbox": [ + 511, + 503, + 908, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Study II. In the second experiment, we assessed whether the generated images could effectively reflect continuous emotional changes (i.e., V-A values). We generated 20 image sets (10 per method), each containing 25 images with V-A values varying gradually from -3 to +3 (Figure 7). Participants rated each set on a 5-point Likert scale regarding (1) the alignment between V-A changes and image content, and (2) the smoothness of the content transition.", + "bbox": [ + 511, + 672, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis & Results. We conducted a Shapiro-Wilk test [35] to assess normality and applied the Wilcoxon Signed Rank test [44] to evaluate statistical significance. Our results indicate that our method outperformed the baseline across all metrics (Table 3), with statistically significant improvements in V/A Ranking, V/A Error, Consistency, and Smoothness (Figure 8).", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "15224", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/49d7a374a28ece4a1b71699e296223fc2af1c98585540d696136733c83d8f2ba.jpg", + "image_caption": [ + "Figure 9. Abalation Study. Images are generated from the prompt \"An oil painting shows an astronaut.\" As $\\alpha$ increases, image-prompt similarity decreases, while emotional variations increase. The usage of $d(a,v)$ enhances the accuracy of emotional changes." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 906, + 354 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3b75a7c34fb19af0a35185edcbb56db21e04b2aff75fff5dc97d181fe1b6946.jpg", + "image_caption": [ + "Figure 10. The effectiveness of the scaling factor $\\alpha$ . The method with $\\alpha = 1.5$ ( $\\star$ ) surpasses the performance indicated by the regression line $(- - )$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 412, + 475, + 551 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c17b1b21d0dd0ff49275e51e53ac12390fed95c1319811997436134c9b7ff770.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
A-Error ↓V-Error ↓CLIPScore ↑
Ours1.828±1.0851.510±1.07423.067±2.655
w/o d(v, a)1.829±1.0831.546±1.08221.977±0.066
", + "bbox": [ + 93, + 608, + 480, + 655 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The effectiveness of $d(v,a)$", + "bbox": [ + 174, + 657, + 397, + 670 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 691, + 243, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We performed ablation experiments to assess the contribution and effectiveness of the proposed loss function.", + "bbox": [ + 89, + 715, + 482, + 744 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Scaling Factor $\\alpha$ . We evaluated how varying $\\alpha$ affects CLIPScore and V/A-Error (Figure 10). As $\\alpha$ increases, CLIPScore decreases (indicating reduced semantic alignment), while V/A-Error also decreases (indicating improved emotional accuracy). This trend is further illustrated by the examples in Figure 9. Based on these findings, we set $\\alpha = 1.5$ as an optimal tradeoff, though users can adjust $\\alpha$ to suit their specific needs.", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of the Density Weighting $d(v,a)$ . We compare our full method with a variant that omits d(v,a)", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from the loss function. As shown in Table 4, including $\\mathrm{d}(\\mathrm{v},\\mathrm{a})$ leads to improvements in both CLIPScore and V/A-Error. This positive effect is further illustrated in Figure 9.", + "bbox": [ + 511, + 410, + 906, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 467, + 769, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce continuous emotional image content generation (C-EICG) and present EmotiCrafter, a novel method that generates emotionally expressive images using continuous Valence-Arousal (V-A) values. Our emotion-embedding network integrates V-A values into textual features, and extensive experiments show that our approach reliably aligns images with both user prompts and specified emotions. We believe this work will advance affective computing and image generation, and we will release our code and data to foster further research.", + "bbox": [ + 511, + 493, + 906, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our method achieved promising results, it still has some limitations need to be addressed in the future. First, controlling image generation based on arousal remains more challenging than controlling based on valence. This is consistent with prior research in visual emotion analysis, which has found that arousal is harder to predict due to lower inter-annotator agreement [21]. Second, our approach frequently generates images featuring human activities even when such activities are not mentioned in the prompts. This likely stems from the limited representation of non-human scenes in our training data and could be mitigated by incorporating a more diverse range of non-human scenarios. Third, our method occasionally modifies users' input prompts to better align with the specified emotional prompts, resulting in a slight semantic shift that affects the generated images. We believe this issue could be addressed by adding a semantic preservation term to the loss function.", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "15225", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "Nan Cao is the corresponding author. This work was supported by the National Key Research and Development Program of China (2023YFB3107100).", + "bbox": [ + 89, + 116, + 483, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 180, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5", + "[2] Bita Azari and Angelica Lim. Emostyle: One-shot facial expression editing using continuous emotion parameters. In Proceedings of WACV, pages 6385-6394, 2024. 2", + "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3", + "[4] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of ACM MM, pages 223-232, 2013. 2", + "[5] Yen-Chi Chen. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187, 2017. 5", + "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, et al. An image is worth one word: Personalizing text-to-image generation using textual inversion. In Proceedings of ICLR, 2023. 3, 5", + "[7] Hatrice Gunes, Björn Schuller, Maja Pantic, and Roddy Cowie. Emotion representation, analysis and synthesis in continuous space: A survey. In Proceedings of IEEE international conference on automatic face & gesture recognition, pages 827-834, 2011. 2", + "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6", + "[9] Stephen Khor Wen Hwoo, Alice Othmani, and Aznul Qalid Md. Sabri. Deep learning-based approach for continuous affect prediction from facial expression images in valence-arousal space. IEEE Access, 10:96053-96065, 2022. 2", + "[10] Ahmed Khota, Eric Cooper, Yu Yan, and Mate Kovacs. Modelling emotional valence and arousal of non-linguistic utterances for sound design support. In Proceedings of the International Conference on Kansei Engineering and Emotion Research, pages 507–516, 2022. 2", + "[11] Dimitrios Kollias. Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges. In Proceedings of CVPR, pages 2328-2336, 2022. 2", + "[12] Ronak Kosti, Jose M. Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of CVPR, pages 1667-1675, 2017. 3, 5" + ], + "bbox": [ + 93, + 205, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Philip A Kragel, Marianne C Reddan, Kevin S LaBar, and Tor D Wager. Emotion schemas are embedded in the human visual system. Science advances, 5(7):eaaw4358, 2019. 2, 3", + "[14] Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. Introducing the open affective standardized image set (oasis). Behavior research methods, 49:457-470, 2017. 5", + "[15] Angela Y Lee and Brian Sternthal. The effects of positive mood on memory. Journal of consumer research, 26(2):115-127, 1999. 2", + "[16] Jiansheng Li, Chuanlan Luo, Qi Zhang, and Rustam Shadiev. Can emotional design really evoke emotion in multimedia learning? International Journal of Educational Technology in Higher Education, 17:1-18, 2020. 2", + "[17] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3", + "[18] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[19] Olga Megalakaki, Ugo Ballenghein, and Thierry Baccino. Effects of valence and emotional intensity on the comprehension and memorization of texts. Frontiers in Psychology, 10:179, 2019. 2", + "[20] Liyu Meng, Yuchen Liu, Xiaolong Liu, Zhaopei Huang, Yuan Cheng, Meng Wang, Chuanhe Liu, and Qin Jin. Multimodal emotion estimation for in-the-wild videos. arXiv preprint arXiv:2203.13032, 2022. 3", + "[21] Laurent Mertens, Elahe Yargholi, Hans Op de Beeck, Jan Van den Stock, and Joost Vennekens. Findingemo: An image dataset for emotion recognition in the wild. Advances in Neural Information Processing Systems, 37:4956-4996, 2024. 2, 3, 5, 8", + "[22] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 2", + "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3", + "[24] Chanjong Park and In-Kwon Lee. Emotional landscape image generation using generative adversarial networks. In Proceedings of ACCV, pages 573-590, 2020. 2", + "[25] Monika Pawlowska and Ewa Magier-Lakomy. The influence of emotional and non-emotional concepts activation on information processing and unintentional memorizing. Polish Psychological Bulletin, pages 150-159, 2011. 2", + "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of ICCV, pages 4195-4205, 2023. 3", + "[27] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of CVPR, pages 860-868, 2015. 3", + "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, et al. SDXL: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of ICLR, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "15226", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3", + "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3", + "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of CVPR, pages 10684-10695, 2022. 3", + "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of CVPR, pages 22500-22510, 2023. 3", + "[33] James A Russell. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161, 1980. 2", + "[34] James A Russell and Albert Mehrabian. Evidence for a three-factor theory of emotions. Journal of research in Personality, 11(3):273-294, 1977. 2", + "[35] S Shaphiro and MBJB Wilk. An analysis of variance test for normality. Biometrika, 52(3):591-611, 1965. 7", + "[36] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839-84865, 2025. 3", + "[37] Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos, and Maja Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions. Nature Machine Intelligence, 3(1):42-50, 2021. 2", + "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 8:235933, 2017. 2", + "[39] Ashish Vaswani. Attention is all you need. In Proceedings of NeurIPS, page 6000-6010, 2017. 4", + "[40] Manuelde Vega. The representation of changing emotions in reading comprehension. Cognition & emotion, 10(3):303-322, 1996. 2", + "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of AAAI, pages 2555-2563, 2023. 6", + "[42] Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li, et al. A systematic review on affective computing: Emotion models, databases, and recent advances. Information Fusion, 83:19-52, 2022. 2", + "[43] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of ICCV, pages 10810-10819, 2023. 3", + "[44] Frank Wilcoxon. Individual comparisons by ranking methods. In Breakthroughs in statistics: Methodology and distribution, pages 196-202. 1992. 7", + "[45] Weizhen Xie and Weiwei Zhang. Negative emotion enhances mnemonic precision and subjective feelings of remembering in visual long-term memory. Cognition, 166:73-83, 2017. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Chao Xu, Junwei Zhu, Jiangning Zhang, Yue Han, Wenqing Chu, Ying Tai, et al. High-fidelity generalized emotional talking face generation with multi-modal emotion space learning. In Proceedings of CVPR, pages 6609-6619, 2023. 2", + "[47] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of CVPR, pages 9479-9488, 2022. 2", + "[48] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, et al. Emotion recognition for multiple context awareness. In Proceedings of ECCV, pages 144–162, 2022. 2", + "[49] Jufeng Yang, Dongyu She, and Ming Sun. Joint image emotion classification and distribution learning via deep convolutional neural network. In Proceedings of IJCAI, pages 3266-3272, 2017. 2", + "[50] Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emotional image content generation with text-to-image diffusion models. In Proceedings of CVPR, pages 6358-6368, 2024. 2, 3", + "[51] Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. arXiv preprint arXiv:2405.12661, 2024. 3", + "[52] Xiaocui Yang, Shi Feng, Daling Wang, and Yifei Zhang. Image-text multimodal emotion classification via multi-view attentional network. IEEE Transactions on Multimedia, 23: 4014-4026, 2021. 2", + "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 5", + "[54] Ligang Zhang, Dian Tjondronegoro, and Vinod Chandran. Representation of facial expression categories in continuous arousal-valence space: feature and correlation. Image and Vision Computing, 32(12):1067-1079, 2014. 2", + "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of ICCV, pages 3836-3847, 2023. 3", + "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of CVPR, pages 586-595, 2018. 6", + "[57] Zixing Zhang, Liyizhe Peng, Tao Pang, Jing Han, Huan Zhao, and Bjorn W Schuller. Refashioning emotion recognition modelling: The advent of generalised large models. IEEE Transactions on Computational Social Systems, pages 6690-6704, 2024. 2", + "[58] Sicheng Zhao, Yue Gao, Xiaolei Jiang, Hongxun Yao, TatSeng Chua, and Xiaoshuai Sun. Exploring principles-of-art features for image emotion recognition. In Proceedings of ACM MM, pages 47-56, 2014. 2", + "[59] Sicheng Zhao, Hongxun Yao, and Xiaolei Jiang. Predicting continuous probability distribution of image emotions in valence-arousal space. In Proceedings of ACM MM, page 879-882, 2015. 2" + ], + "bbox": [ + 514, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "15227", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Sicheng Zhao, Xingxu Yao, Jufeng Yang, et al. Affective image content analysis: Two decades review and new perspectives. IEEE TPAMI, 44(10):6729-6751, 2022. 2", + "[61] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + ], + "bbox": [ + 91, + 90, + 480, + 189 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "15228", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_model.json b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..f16a7d4707ee4329edd1dd6882024e2925671aef --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_model.json @@ -0,0 +1,2609 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.189, + 0.131, + 0.812, + 0.175 + ], + "angle": 0, + "content": "EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model" + }, + { + "type": "text", + "bbox": [ + 0.158, + 0.203, + 0.837, + 0.224 + ], + "angle": 0, + "content": "Shengqi Dang \\(^{1,2\\dagger}\\), Yi He \\(^{1\\dagger}\\), Long Ling \\(^{1}\\), Ziqing Qian \\(^{1}\\), Nanxuan Zhao \\(^{3}\\), Nan Cao \\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.213, + 0.238, + 0.78, + 0.258 + ], + "angle": 0, + "content": "\\(^{1}\\)Tongji University \\(^{2}\\)Shanghai Innovation Institute \\(^{3}\\)Adobe Research" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.293, + 0.518, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.316, + 0.706, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.462, + 0.705, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.625, + 0.666, + 0.636 + ], + "angle": 0, + "content": "A man is running fast" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.316, + 0.901, + 0.461 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.714, + 0.462, + 0.898, + 0.474 + ], + "angle": 0, + "content": "A group of people are reading books" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.478, + 0.902, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.699, + 0.625, + 0.905, + 0.637 + ], + "angle": 0, + "content": "A cartoon shows children playing football" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.646, + 0.908, + 0.687 + ], + "angle": 0, + "content": "Figure 1. Generated images of the EmotiCrafter. Given a prompt and Valence-Arousal (V-A) values, our method can generate an emotional image that reflects the input prompt and aligns with the specified emotion values. By adjusting the V-A values, our method can also generate images that evoke discrete emotions, such as sadness, relaxation, anger, and amusement." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.701, + 0.327, + 0.716 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.734, + 0.486, + 0.764 + ], + "angle": 0, + "content": "Recent research shows that emotions can enhance users' cognition and influence information communication. While" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.909, + 0.899 + ], + "angle": 0, + "content": "research on visual emotion analysis is extensive, limited work has been done on helping users generate emotionally rich image content. Existing work on emotional image generation relies on discrete emotion categories, making it challenging to capture complex and subtle emotional nuances accurately. Additionally, these methods struggle to control the specific content of generated images based on text prompts. In this paper, we introduce the task of continuous emotional image content generation (C-EICG) and present EmotiCrafter, a general emotional image generation model that generates images based on free text prompts and Valence-Arousal (V-A) values. It leverages a novel emotion-embedding mapping network to fuse V-A values" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.78, + 0.428, + 0.793 + ], + "angle": 0, + "content": "† Shengqi Dang and Yi He contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.793, + 0.32, + 0.805 + ], + "angle": 0, + "content": "\\* Nan Cao is the corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.805, + 0.483, + 0.864 + ], + "angle": 0, + "content": "*Shengqi Dang, Yi He, Long Ling, Ziqing Qian, and Nan Cao are with the Intelligent Big Data Visualization Lab, Tongji University. Shengqi Dang and Nan Cao are also with the Shanghai Innovation Institute. Email: {dangsq123, heyi_11}@tongji.edu.cn, lucyling0224@gmail.com, 2411920@tongji.edu.cn, nan.cao@gmail.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.865, + 0.482, + 0.888 + ], + "angle": 0, + "content": "*Nanxuan Zhao is with Adobe Research. Email: nanxuanzhao@g-mail.com." + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.888, + 0.466, + 0.9 + ], + "angle": 0, + "content": "\\*Code: https://github.com/idvxlab/EmotiCrafter" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.78, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15218" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "into textual features, enabling the capture of emotions in alignment with intended input prompts. A novel loss function is also proposed to enhance emotion expression. The experimental results show that our method effectively generates images representing specific emotions with the desired content and outperforms existing techniques." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.223, + 0.227 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.237, + 0.485, + 0.342 + ], + "angle": 0, + "content": "Emotions are fundamental to human experiences and play a critical role in shaping how people perceive and interact with the world. Research has shown that emotions affect memory [15, 19, 25, 45] and comprehension [16, 38, 40], which are crucial for effective communication. As a result, content creators increasingly recognize the importance of incorporating emotions to enhance audience engagement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.568 + ], + "angle": 0, + "content": "While research on visual emotion analysis is extensive [13, 49, 58], there is limited work on generating emotionally rich image content. Some early studies explored emotional content generation techniques within specific domains such as facial expressions [2, 46] or landscapes [24]. EmoGen [50] generates images based on a given emotion tag (e.g., happy or sad), which pioneered the domain-free emotional image content generation (EICG) task. However, it has two critical limitations: (1) the image is generated from an emotion tag instead of a text prompt, making the generated content difficult to control; (2) while the discrete emotion tags used in EmoGen are easy to understand, psychologists have not achieved a consensus on the emotion categories [60]. The limited scope of discrete emotion tags falls short of capturing nuanced emotions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.87 + ], + "angle": 0, + "content": "To address the above issues, we propose the continuous emotional image content generation (C-EICG) task and present EmotiCrafter, the first C-EIGC model that generates emotional images using free-text prompts and continuous Valence-Arousal (V-A) values defined in the V-A model (a well-known psychological continuous emotion model) [33]. V-A model represents emotions in a two-dimensional Cartesian space (Figure 2), where Valence quantifies pleasantness (negative to positive) and Arousal measures intensity (calm to excited). We utilize a [-3, 3] range for Valence and Arousal [10, 21]. The continuous V-A space enables smooth transitions and nuanced emotional shifts beyond the capability of discrete labels (e.g., shifts from \"bored\" to \"tired\" of a character's state in a video). Specific emotional values in this space have been investigated in prior work [34], and such fine-grained modeling is particularly beneficial for human-computer interaction [2]. Leveraging this model, EmotiCrafter captures subtle affective variations via precise (V, A) positioning. The contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a novel task, continuous emotional image content generation (C-EICG), and develop the first ded" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.279, + 0.811, + 0.293 + ], + "angle": 0, + "content": "Figure 2. Valence-Arousal model." + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.318, + 0.905, + 0.424 + ], + "angle": 0, + "content": "icated model for this purpose. Our model introduces an emotion-embedding network that integrates continuous Valence-Arousal (V-A) values with text prompts. These fused features are then injected into Stable Diffusion XL [28] using cross-attention mechanisms, enabling precise control over both content and emotional expressions in the generated images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.425, + 0.905, + 0.544 + ], + "angle": 0, + "content": "- We propose a novel loss function that enhances the emotional resonance of generated images. By amplifying the difference between neutral and emotional text features, our approach enables the model to capture more distinct emotional variations. Additionally, the loss function incorporates the V-A distribution to address data imbalance, further refining the model's ability to generate images with rich, accurate emotional expressions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.545, + 0.905, + 0.62 + ], + "angle": 0, + "content": "- We constructed an emotional prompts dataset to train the emotion-embedding network, where each sample consists of a neutral prompt and an emotional prompt that share the same core meaning but express a specific emotion corresponding to a given pair of V-A values." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.425, + 0.905, + 0.62 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.634, + 0.652, + 0.649 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.905, + 0.705 + ], + "angle": 0, + "content": "In this section, we present a review of the related work, specifically focusing on visual emotion analysis, image emotion transfer, and conditional image generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.738, + 0.729 + ], + "angle": 0, + "content": "2.1. Visual Emotion Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.856 + ], + "angle": 0, + "content": "Visual emotion analysis refers to the computational recognition and interpretation of human emotion [57] in visual media, such as images or videos. It has been a prominent research area, with most efforts focusing on the classification of discrete emotions [4, 42, 47, 48, 52]. However, discrete emotion categories limit the ability to capture the nuanced emotions [7], leading to an increased focus on continuous emotion analysis in images [59]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Much of the continuous emotion analysis remains centered on facial expression analysis [9, 11, 22, 37, 54]. While effective, facial analysis can overlook crucial con" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15219" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "textual information that influences emotional interpretation. This drives studies to focus on emotions in objects or individuals within their environment, rather than just faces[12, 13, 20, 21]. For example, Kosti et al. [12] combined person-specific characteristics with the context of the scene to predict continuous emotional dimensions such as arousal, valence, and dominance. Kragel et al. [13] proposed EmoNet to extract visual features such as facial expressions, body posture, and scene elements from images to predict valence and arousal values. Recently, Mertens et al.[21] tested multiple backbones (such as ResNet, CLIP, DINO, etc.) for prediction of valence and arousal, demonstrating strong performance across various architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.482, + 0.38 + ], + "angle": 0, + "content": "Previous studies have achieved high accuracy in predicting continuous emotions in images, demonstrating a correlation between continuous emotions and visual elements. Building on this insight, our work moves beyond prediction to explore how continuous emotions can be actively embedded within generated content." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.317, + 0.406 + ], + "angle": 0, + "content": "2.2. Image Emotion Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Image Emotion Transfer (IET) focuses on editing the content of the images to evoke different emotions [17, 27, 43, 61]. For instance, Peng et al. [27] achieved emotion transfer by adjusting color tones and texture-related features. Zhu et al. [61] introduced a method that separates high-level emotion-relevant features (e.g., object shapes and scene layout) from low-level emotion-relevant features (e.g., brightness). By applying GANs, they transferred emotions between images while preserving their original structure. Building on these methods, Weng et al. [43] proposed the Affective Image Filter, which uses a multi-modal transformer to process both text and image inputs. With the emergence of text-to-image models, IET has expanded into new applications based on instructive commands. For example, EmoEdit [51] used GPT-4V to build emotion factor trees that map abstract emotions to specific visual elements and employed the InstructPix2Pix model to apply emotion-driven content and color adjustments to images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.762 + ], + "angle": 0, + "content": "Current methods primarily extract specific features, focusing on certain visual elements or emotional cues, which can limit the depth of emotional expression. In contrast, our method broadly learns a variety of emotion-influencing features, and it could accept natural language prompts as input." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.363, + 0.789 + ], + "angle": 0, + "content": "2.3. Conditional Image Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Conditional image generation aims to create images that align with specific input conditions, such as text [3, 28, 30, 31], reference images [53], subjects [6, 32], and depth maps [23, 55]. To enhance quality, researchers have developed specialized approaches, such as Diffusion Transformers (DiT)[26], which use transformer-based diffusion for denoising, and Visual Autoregressive Modeling (VAR)[36]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "which encodes images into discrete tokens for autoregressive prediction across scales. However, these state-of-the-art methods rely on discrete labels, limiting flexibility and control. In contrast, our approach offers greater generalization by leveraging Stable Diffusion XL [28] to generate emotional images from free-text prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.303 + ], + "angle": 0, + "content": "Despite these progresses, incorporating emotion as a condition for image generation remains underexplored. EmoGen [50] pioneered the generation of emotional image content (EIGC) by mapping emotional features to semantic features to generate emotional images. However, EmoGen struggles to understand natural language, limiting its capacity to effectively control specific content. Its reliance on discrete emotions also limits its practical applicability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Our method bridges this gap by embedding continuous emotion into textual features, enabling image generation models to use continuous emotion for emotion control. Unlike label-based methods, our method supports free-text prompts for flexible content control." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.393, + 0.605, + 0.408 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.905, + 0.449 + ], + "angle": 0, + "content": "In this section, we introduce the technique details of the proposed EmotiCrafter." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.625, + 0.473 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Our method generates emotional images \\( I_{emo} \\) from two inputs (Figure 3): a free-text prompt describing the desired content, and a pair of V-A values \\( (v, a) \\) specifying the emotion. First the prompt encoder \\( \\mathcal{E} \\) converts the text prompt into feature \\( f_{n} \\), which is then processed by an emotion-embedding network \\( \\mathcal{M} \\) to produce emotional prompt feature \\( \\hat{f}_{e} \\) that integrate the V-A values:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.597, + 0.905, + 0.617 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {e} = \\mathcal {M} \\left(f _ {n} | (v, a)\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Next, this feature is injected into Stable Diffusion XL \\(\\mathcal{G}\\) via its cross-attention mechanism to generate the emotional images: \\(I_{emo} = \\mathcal{G}(\\hat{f}_e)\\). To enhance emotional expressiveness, we introduce a loss function (Fig.3(b.2)) that leverages the V-A distribution and emphasizes the differences between neutral and emotional prompt features. This ensures that the generated images accurately convey both the intended emotions and content. Additionally, we construct a dataset (Fig.3(a)) that pairs neutral and emotional prompts with the corresponding V-A values." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.783, + 0.804 + ], + "angle": 0, + "content": "3.2. Emotion-Embedding Network" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The emotion-embedding network \\(\\mathcal{M}\\) generates the emotional prompt feature by integrating a pair of V-A values with a neutral prompt feature (Fig.3(b.1)). First, a V-A encoder converts the V-A values into feature vectors. Then, an emotion injection transformer—modified from GPT-2[29]—fuses these vectors with the neutral prompt" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15220" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.09, + 0.904, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.288, + 0.908, + 0.359 + ], + "angle": 0, + "content": "Figure 3. Overview of our method. Specifically, we take the following steps. (a) We collect an image dataset annotated with V-A values, neutral prompts, and emotional prompts. These prompts are then encoded into features by prompt encoder \\(\\mathcal{E}\\). (b) Next, we design (b.1) an emotion-embedding network \\(\\mathcal{M}\\) to embed V/A values into textual features based on the transformer architecture, and (b.2) a specialized loss function to enhance the emotional resonance of generated images. The output of the mapping network serves as the condition for the image generation model \\(\\mathcal{G}\\) to generate emotional images." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.366, + 0.48, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.469, + 0.483, + 0.525 + ], + "angle": 0, + "content": "Figure 4. Structure of Emotion Injection Block. It accepts hidden state \\( h_{i-1} \\) as input and produces \\( h_i \\) as output. The V-feature \\( e_v \\) and A-feature \\( e_a \\) represent the emotion features, which are injected through the cross-attention module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.559 + ], + "angle": 0, + "content": "feature, preserving the original textual context while infusing emotional content." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.664 + ], + "angle": 0, + "content": "V-A Encoder. The V-A Encoder converts a pair of V-A values into feature vectors using two separate multilayer perceptrons (MLPs). One MLP processes the Valence value to produce V-feature \\( e_v \\), and the other processes the Arousal value to generate A-feature \\( e_a \\). These features are then fed into the emotion injection transformer network for further emotion infusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.741 + ], + "angle": 0, + "content": "Emotion Injection Transformer (EIT). The Emotion Injection Transformer (EIT) leverages a modified GPT-2 architecture to seamlessly integrate V/A-features into textual features. Its process consists of three stages: input projection, emotion injection, and output projection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.483, + 0.771 + ], + "angle": 0, + "content": "First, we project the input neutral prompt feature \\( f_{n} \\) into the transformer's feature space:" + }, + { + "type": "equation", + "bbox": [ + 0.219, + 0.779, + 0.483, + 0.795 + ], + "angle": 0, + "content": "\\[\nh _ {0} = \\mathrm {P} _ {\\text {i n}} \\left(f _ {n}\\right) + \\mathrm {P E}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.483, + 0.832 + ], + "angle": 0, + "content": "where \\( h_0 \\) represents the initial hidden state; \\( \\mathrm{P_{in}(\\cdot)} \\) is a linear projection layer, and PE denotes positional embedding [39]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In the next, we inject emotion into \\( f_{n} \\) via 12 sequential Emotion Injection Blocks (EIBs) corresponding to 12 transformer blocks. Each block outputs a hidden state:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\operatorname {E I B} \\left(h _ {i - 1}, e _ {v}, e _ {a}\\right), \\quad i \\in \\{1, \\dots , 1 2 \\} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.907, + 0.412 + ], + "angle": 0, + "content": "where \\( h_i \\) is the output of the \\( i \\)-th EIB(\\(\\cdot\\)). As shown in Figure 4, each EIB enhances the transformer block through a cross-attention mechanism:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.421, + 0.905, + 0.437 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {\\prime} = \\text {s e l f - a t t n} (\\mathrm {L N} (h _ {i - 1})) + h _ {i - 1} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.441, + 0.905, + 0.461 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {(v)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {\\prime}\\right), e _ {v}\\right) + h _ {i} ^ {\\prime} \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.463, + 0.905, + 0.484 + ], + "angle": 0, + "content": "\\[\nh _ {i} ^ {(v, a)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v)}\\right), e _ {a}\\right) + h _ {i} ^ {(v)} \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.486, + 0.905, + 0.507 + ], + "angle": 0, + "content": "\\[\nh _ {i} = \\operatorname {f n n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v, a)}\\right)\\right) + h _ {i} ^ {(v, a)} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.636 + ], + "angle": 0, + "content": "where \\( h^{\\prime}, h^{(v)}, h^{(v,a)} \\) are the intermediate hidden variables; \\( \\mathrm{LN}(\\cdot) \\) is the LayerNorm; self-attn(\\(\\cdot\\)) denotes self-attention, employed to capture context dependencies; cross-attn(\\(\\cdot\\)) is cross-attention for injecting \\( e_v \\) and \\( e_a \\); \\( \\mathrm{fnn}(\\cdot) \\) is a feedforward network that adapts the complexity of the emotional embedding process. We also remove the causal mask typically used for next-token prediction from the original transformer model to fit our task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.906, + 0.728 + ], + "angle": 0, + "content": "Finally, the output of the last (i.e., the 12th) injection block \\( h_{12} \\) is projected back to SDXL's prompt feature space via \\( \\mathbf{P}_{\\mathrm{out}} \\) (a linear layer and a LayerNorm layer) to predict the residual between emotional and neutral prompt features, which represents a semantic shift between emotional and neutral prompts:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.735, + 0.905, + 0.754 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {r} = \\mathrm {P} _ {\\text {o u t}} \\left(\\mathrm {L N} \\left(h _ {1 2}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.761, + 0.907, + 0.792 + ], + "angle": 0, + "content": "The final emotional prompt feature is obtained by adding this residual to the original neutral prompt feature:" + }, + { + "type": "equation", + "bbox": [ + 0.663, + 0.799, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {e} = \\hat {f} _ {r} + f _ {n} \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The above emotion embedding network is trained by minimizing the averaged expectation of the difference between the predicted emotional prompt feature \\(\\hat{f}_e = \\mathcal{M}(f_n|(v,a))\\) and the scaled target emotional prompt feature \\(f_e^t\\), using the loss function described in Equation 10." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15221" + } + ], + [ + { + "type": "equation", + "bbox": [ + 0.18, + 0.103, + 0.483, + 0.137 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\frac {1}{n} \\mathbb {E} \\left(\\frac {1}{d (v , a)} \\| \\hat {f} _ {e} - f _ {e} ^ {t} \\| ^ {2}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.141, + 0.483, + 0.186 + ], + "angle": 0, + "content": "where \\( n \\) is the number of feature elements; \\( \\mathbb{E}(\\cdot) \\) denotes the expectation; \\( d(v, a) \\) is a density function that describes the distribution of V-A values in the training sample." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.247 + ], + "angle": 0, + "content": "To effectively embed emotions and address the challenges posed by the uneven distribution of V-A values in the dataset, this loss function incorporates two key strategies to improve the model's performance:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Scaled Residual Learning. To better capture pronounced emotional changes in generated images, we enlarge the target residuals:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.302, + 0.483, + 0.337 + ], + "angle": 0, + "content": "\\[\nf _ {e} ^ {t} = f _ {n} + \\alpha \\underbrace {\\left(f _ {e} - f _ {n}\\right)} _ {\\text {r e s i d u a l f e a t u r e}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.483, + 0.392 + ], + "angle": 0, + "content": "where \\( f_{e} \\) is the emotional prompt feature, \\( f_{n} \\) is the neutral prompt feature, and \\( \\alpha \\) is a scale factor, we set its value to 1.5 based on the ablation study." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.483 + ], + "angle": 0, + "content": "V-A Density Weighting. To mitigate the effects of the imbalanced distribution of the training samples, we weigh the loss inversely proportional to the density of training samples in the V-A space. The density is estimated using Kernel Density Estimation (KDE) [5] with a Gaussian kernel, denoted as \\( d(v,a) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.495, + 0.483, + 0.533 + ], + "angle": 0, + "content": "\\[\nd (v, a) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} K _ {H} ((v, a) - (v _ {i}, a _ {i})), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.538, + 0.484, + 0.614 + ], + "angle": 0, + "content": "where \\( K_{H} \\) is a 2D Gaussian kernel with bandwidth \\( H \\); \\( n \\) is the number of training samples; \\( (v_{i},a_{i}) \\) are the V-A values of the \\( i \\)-th training sample. The bandwidth \\( H \\) is selected using Silverman's rule of thumb to provide optimal smoothing of the density estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.623, + 0.292, + 0.639 + ], + "angle": 0, + "content": "3.3. Dataset and Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.855 + ], + "angle": 0, + "content": "To train the emotion-embedding network, we constructed a dataset of paired neutral and emotional prompts with corresponding Valence-Arousal (V-A) values (Figure 3(a)). These pairs were automatically generated using GPT-4 based on 39,843 images with human-annotated V-A values from publicly available datasets, including OASIS [14], EMOTIC [12], and FindingEmo [21]. Specifically, GPT-4 was used to generate neutral prompts with objective image descriptions and emotional prompts that emphasize affective attributes such as color, lighting, and texture, which influence emotional perception. To ensure data reliability, all LLM-generated prompts underwent crowd-sourced verification, with disagreements resolved through a voting mechanism among annotators." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The proposed emotion embedding network is trained on two NVIDIA A800 GPUs using the aforementioned dataset. We employ the AdamW [18] optimizer with a weight decay" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "of 1e-5 and a learning rate of 1e-3. The training process spans 200 epochs with a batch size of 768, completing in approximately 7 to 8 hours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.153, + 0.631, + 0.168 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.698, + 0.193 + ], + "angle": 0, + "content": "4.1. Generation Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.201, + 0.907, + 0.397 + ], + "angle": 0, + "content": "Figure 1 shows the proposed technique's ability to achieve continuous and effective control over both emotion and content during image generation. Meanwhile, Figure 5 highlights four key capabilities: (a) emotion-content decoupling, where V-A values override emotional cues in the prompt, allowing typically positive concepts to be rendered with negative emotions; (b) compatibility with discrete emotion categories; (c) content-independent generation, where images generated from empty prompts and specified V-A values maintain emotional consistency without semantic constraints; and (d) fine-grained emotional control, demonstrated through V-A increments of 0.2, showcasing the model's sensitivity to subtle emotional variations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.409, + 0.653, + 0.424 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.461 + ], + "angle": 0, + "content": "To estimate the effectiveness of the proposed method, we built four baselines based on existing techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Baselines. We established baselines for comparison using two strategies: (1) directly injecting emotion features into the image generation modules of SDXL, such as UNet, and (2) modifying the input text prompt using emotion features to influence the generated image's content, similar to the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.719 + ], + "angle": 0, + "content": "As a result, four different baselines were built: (1) Cross Attention: inject the emotion features \\((e_v, e_a)\\) into the UNet in SDXL via its cross-attention mechanism based on IP-Adapter [53]; (2) Time Embedding: directly add emotion features \\((e_v, e_a)\\) to the time embedding of the UNet in SDXL. (3) Textual Inversion: use the text inversion technique [6] to embed emotion features \\((e_v, e_a)\\) into prompt templates with predefined emotion placeholders. (4) GPT-4+SDXL (GPT-SD): use GPT-4 [1] to rewrite the input text according to (v,a) values to generate an emotional SDXL prompt for image generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Qualitative Comparison. We evaluate the generated images based on three criteria: (1) the effectiveness of emotion embedding, (2) image-prompt similarity, and (3) the continuity of emotional variations as V-A values change. Figure 6 demonstrates that the baseline methods—Cross Attention, Time Embedding, and Textual Inversion—tend to produce nearly identical outputs regardless of emotional variation. This is because the loss terms of these baselines primarily align low-level image features (e.g., SDXL's latent space), which are highly correlated with prompt content but struggle to capture subtle emotional cues when both prompts and V-A values are provided. Notably, Tex-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15222" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.117, + 0.12, + 0.13 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.089, + 0.212, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.215, + 0.117, + 0.237, + 0.13 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.242, + 0.089, + 0.477, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.478, + 0.185, + 0.487, + 0.2 + ], + "angle": 0, + "content": "1)" + }, + { + "type": "image", + "bbox": [ + 0.489, + 0.089, + 0.573, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.574, + 0.09, + 0.738, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.09, + 0.905, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.185, + 0.12, + 0.2 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.158, + 0.46, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.158, + 0.509, + 0.171 + ], + "angle": 0, + "content": "\\( \\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : \\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = x \\)" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.158, + 0.655, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.655, + 0.158, + 0.738, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.738, + 0.159, + 0.822, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.822, + 0.159, + 0.905, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.232, + 0.907, + 0.274 + ], + "angle": 0, + "content": "Figure 5. Results under multiple inputs. (a) Overriding semantic content ('a child in the amusement park') with sad V-A (-2,-2); (b) Discrete emotion mapping in V-A space as emotion input; (c) Empty-prompt generation with pure emotion condition; (d) Fine-grained control of V-A variations with a granularity of 0.2." + }, + { + "type": "image", + "bbox": [ + 0.119, + 0.278, + 0.88, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.908, + 0.608 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparisons with baselines. These images are generated at varying V-A values, specifically -1.5, 0, and 1.5. Only our approach and the GPT-4+SDXL successfully generate images that clearly reflect emotional variations. Notably, our results show enhanced continuity, indicating superior controllability over continuous V-A values compared to the GPT-4+SDXL." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.771 + ], + "angle": 0, + "content": "tual Inversion often produces images with a persistent purple tint. Furthermore, all methods faithfully generate content aligned with the given prompts. However, for image continuity, we focus on comparing our method with GPT-4+SDXL, as they are the only two capable of generating distinct emotional variations. Figure 7 provides a more detailed comparison, showing that our method maintains smooth emotional transitions even under extreme V-A conditions, whereas GPT-4+SDXL introduces noticeable discontinuities (e.g., in the \\( \\mathrm{V} = 3 \\) column)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Quantitative Comparison. We compare our method against several baselines using the following metrics: (1) \\( \\underline{\\mathrm{V} / \\mathrm{A} - \\mathrm{Error}} \\) evaluates the absolute error between the predicted V/A of the generated images and the input V/A. (2) CLIPScore [8] assesses the similarity between the input text and the generated images. (3) CLIP-IQA [41] leverages a pre-trained CLIP model to evaluate image quality without requiring reference images. (4) LPIPS-Continuous utilizes" + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.618, + 0.905, + 0.685 + ], + "angle": 0, + "content": "
A-Error ↓V-Error ↓CLIPScore ↑CLIP-IQA ↑
Cross Attention1.923±1.1532.080±1.43826.266±2.3810.949±0.046
Time Embedding1.941±1.1682.031±1.34826.566±2.1250.786±0.164
Textual Inversion1.958±1.1881.923±1.17022.346±3.5940.370±0.111
GPT-4+SDXL1.860±1.0901.517±1.06025.907±1.9490.906±0.066
Ours1.828±1.0851.510±1.07423.067±2.6550.881±0.099
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.686, + 0.907, + 0.784 + ], + "angle": 0, + "content": "Table 1. Comparison on emotion accuracy, prompt fidelity, and image quality across different baselines, evaluated on 3,300 images per method (132 prompts \\(\\times\\) 5 V values \\(\\times\\) 5 A values). Our method achieves the highest performance in emotion accuracy while maintaining comparable results in prompt fidelity and image quality. The slight decrease in prompt fidelity is expected, as modifying emotional content affects semantic alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.905, + 0.838 + ], + "angle": 0, + "content": "the Learned Perceptual Image Patch Similarity [56] to measure the continuity of the change in image as V/A changes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As shown in Table 1, our method achieves the lowest (best) \\( \\underline{V / A - Error} \\) on average. While Cross Attention and Time Embedding achieve the highest \\( \\underline{CLIP - IQA} \\) and \\( \\underline{CLIPScore} \\), respectively, these methods fail to generate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15223" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.482, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.754 + ], + "angle": 0, + "content": "Figure 7. A more comprehensive comparison with the GPT-4+SDXL for the prompt \"A group of people is skiing in a snow hill.\" Our approach maintains continuity even under extreme V/A conditions. Conversely, GPT-4+SDXL displays noticeable discontinuities (e.g., in the \\( \\mathrm{V} = 3 \\) column)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.777, + 0.373, + 0.793 + ], + "angle": 0, + "content": "emotionally expressive images (Figure 6)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Our method exhibits a slight decrease in CLIPScore compared to the baselines, which we attribute to the inherent trade-off between emotional modulation and strict semantic alignment. However, the high CLIP-IQA score indicates that our method produces high-quality images. Additionally, as shown in Table 2, our approach demonstrates superior continuity compared to GPT-4 + SDXL." + }, + { + "type": "table", + "bbox": [ + 0.537, + 0.088, + 0.882, + 0.121 + ], + "angle": 0, + "content": "
OursGPT-4+SDXL
LPIPS-Continuous↓0.220±0.0640.361±0.059
" + }, + { + "type": "table_caption", + "bbox": [ + 0.515, + 0.124, + 0.907, + 0.138 + ], + "angle": 0, + "content": "Table 2. Continuity comparison between our method and baseline." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.142, + 0.92, + 0.248 + ], + "angle": 0, + "content": "
OursGPT-4+SDXL
Study IA-Ranking Consistency ↑0.759±0.2730.165±0.379
V-Ranking Consistency ↑0.887±0.2450.584±0.259
A-Error ↓1.327±1.1202.029±1.446
V-Error ↓0.692±0.6821.229±1.026
Study IIEmotion Consistency ↑4.215±0.7153.525±1.065
Emotion Smoothness ↑4.240±0.8283.195±1.163
" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.254, + 0.855, + 0.267 + ], + "angle": 0, + "content": "Table 3. Our method outperformed the baseline." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.277, + 0.905, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.4, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Figure 8. User Study Results (\\(^{*}\\mathsf{p} < 0.05\\) \\(^{**}\\mathsf{p} < 0.01\\) \\(^{**}p < 0.001\\)" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.433, + 0.637, + 0.45 + ], + "angle": 0, + "content": "4.3. User Study" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.907, + 0.503 + ], + "angle": 0, + "content": "We conducted two user studies with 20 college students to evaluate the effectiveness of our method by comparing it to \\( GPT-4 + SDXL \\) (the baseline)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.505, + 0.909, + 0.672 + ], + "angle": 0, + "content": "Study I. In the first experiment, we evaluated whether the generated images' emotions align with human perception. Two image collections were prepared—one for Arousal (A) and one for Valence (V)—each comprising 20 sets (10 from our method and 10 from the baseline) of 5 randomized images with varying A or V values. Participants reordered the images by perceived intensity and estimated each image's V and A values. We then computed Kendall's \\(\\tau_{b}\\) (with \\(\\tau_{b} = 1\\) indicating perfect alignment) to assess ordering accuracy and calculated the absolute error between the estimated and ground truth values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Study II. In the second experiment, we assessed whether the generated images could effectively reflect continuous emotional changes (i.e., V-A values). We generated 20 image sets (10 per method), each containing 25 images with V-A values varying gradually from -3 to +3 (Figure 7). Participants rated each set on a 5-point Likert scale regarding (1) the alignment between V-A changes and image content, and (2) the smoothness of the content transition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis & Results. We conducted a Shapiro-Wilk test [35] to assess normality and applied the Wilcoxon Signed Rank test [44] to evaluate statistical significance. Our results indicate that our method outperformed the baseline across all metrics (Table 3), with statistically significant improvements in V/A Ranking, V/A Error, Consistency, and Smoothness (Figure 8)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "15224" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.907, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.907, + 0.404 + ], + "angle": 0, + "content": "Figure 9. Abalation Study. Images are generated from the prompt \"An oil painting shows an astronaut.\" As \\(\\alpha\\) increases, image-prompt similarity decreases, while emotional variations increase. The usage of \\(d(a,v)\\) enhances the accuracy of emotional changes." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.414, + 0.476, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.558, + 0.483, + 0.601 + ], + "angle": 0, + "content": "Figure 10. The effectiveness of the scaling factor \\(\\alpha\\). The method with \\(\\alpha = 1.5\\) (\\(\\star\\)) surpasses the performance indicated by the regression line \\((- - )\\)." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.609, + 0.481, + 0.656 + ], + "angle": 0, + "content": "
A-Error ↓V-Error ↓CLIPScore ↑
Ours1.828±1.0851.510±1.07423.067±2.655
w/o d(v, a)1.829±1.0831.546±1.08221.977±0.066
" + }, + { + "type": "table_caption", + "bbox": [ + 0.175, + 0.658, + 0.398, + 0.671 + ], + "angle": 0, + "content": "Table 4. The effectiveness of \\(d(v,a)\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.693, + 0.245, + 0.71 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.746 + ], + "angle": 0, + "content": "We performed ablation experiments to assess the contribution and effectiveness of the proposed loss function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Effectiveness of the Scaling Factor \\(\\alpha\\). We evaluated how varying \\(\\alpha\\) affects CLIPScore and V/A-Error (Figure 10). As \\(\\alpha\\) increases, CLIPScore decreases (indicating reduced semantic alignment), while V/A-Error also decreases (indicating improved emotional accuracy). This trend is further illustrated by the examples in Figure 9. Based on these findings, we set \\(\\alpha = 1.5\\) as an optimal tradeoff, though users can adjust \\(\\alpha\\) to suit their specific needs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of the Density Weighting \\(d(v,a)\\). We compare our full method with a variant that omits d(v,a)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.907, + 0.457 + ], + "angle": 0, + "content": "from the loss function. As shown in Table 4, including \\( \\mathrm{d}(\\mathrm{v},\\mathrm{a}) \\) leads to improvements in both CLIPScore and V/A-Error. This positive effect is further illustrated in Figure 9." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.771, + 0.484 + ], + "angle": 0, + "content": "5. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.644 + ], + "angle": 0, + "content": "In this paper, we introduce continuous emotional image content generation (C-EICG) and present EmotiCrafter, a novel method that generates emotionally expressive images using continuous Valence-Arousal (V-A) values. Our emotion-embedding network integrates V-A values into textual features, and extensive experiments show that our approach reliably aligns images with both user prompts and specified emotions. We believe this work will advance affective computing and image generation, and we will release our code and data to foster further research." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Although our method achieved promising results, it still has some limitations need to be addressed in the future. First, controlling image generation based on arousal remains more challenging than controlling based on valence. This is consistent with prior research in visual emotion analysis, which has found that arousal is harder to predict due to lower inter-annotator agreement [21]. Second, our approach frequently generates images featuring human activities even when such activities are not mentioned in the prompts. This likely stems from the limited representation of non-human scenes in our training data and could be mitigated by incorporating a more diverse range of non-human scenarios. Third, our method occasionally modifies users' input prompts to better align with the specified emotional prompts, resulting in a slight semantic shift that affects the generated images. We believe this issue could be addressed by adding a semantic preservation term to the loss function." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "15225" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.117, + 0.484, + 0.162 + ], + "angle": 0, + "content": "Nan Cao is the corresponding author. This work was supported by the National Key Research and Development Program of China (2023YFB3107100)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.181, + 0.188, + 0.197 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.206, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.484, + 0.292 + ], + "angle": 0, + "content": "[2] Bita Azari and Angelica Lim. Emostyle: One-shot facial expression editing using continuous emotion parameters. In Proceedings of WACV, pages 6385-6394, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.294, + 0.482, + 0.35 + ], + "angle": 0, + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.353, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[4] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of ACM MM, pages 223-232, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.451 + ], + "angle": 0, + "content": "[5] Yen-Chi Chen. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.456, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, et al. An image is worth one word: Personalizing text-to-image generation using textual inversion. In Proceedings of ICLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.483, + 0.582 + ], + "angle": 0, + "content": "[7] Hatrice Gunes, Björn Schuller, Maja Pantic, and Roddy Cowie. Emotion representation, analysis and synthesis in continuous space: A survey. In Proceedings of IEEE international conference on automatic face & gesture recognition, pages 827-834, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.483, + 0.655 + ], + "angle": 0, + "content": "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.483, + 0.725 + ], + "angle": 0, + "content": "[9] Stephen Khor Wen Hwoo, Alice Othmani, and Aznul Qalid Md. Sabri. Deep learning-based approach for continuous affect prediction from facial expression images in valence-arousal space. IEEE Access, 10:96053-96065, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.729, + 0.483, + 0.799 + ], + "angle": 0, + "content": "[10] Ahmed Khota, Eric Cooper, Yu Yan, and Mate Kovacs. Modelling emotional valence and arousal of non-linguistic utterances for sound design support. In Proceedings of the International Conference on Kansei Engineering and Emotion Research, pages 507–516, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.856 + ], + "angle": 0, + "content": "[11] Dimitrios Kollias. Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges. In Proceedings of CVPR, pages 2328-2336, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Ronak Kosti, Jose M. Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of CVPR, pages 1667-1675, 2017. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.206, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[13] Philip A Kragel, Marianne C Reddan, Kevin S LaBar, and Tor D Wager. Emotion schemas are embedded in the human visual system. Science advances, 5(7):eaaw4358, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[14] Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. Introducing the open affective standardized image set (oasis). Behavior research methods, 49:457-470, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[15] Angela Y Lee and Brian Sternthal. The effects of positive mood on memory. Journal of consumer research, 26(2):115-127, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[16] Jiansheng Li, Chuanlan Luo, Qi Zhang, and Rustam Shadiev. Can emotional design really evoke emotion in multimedia learning? International Journal of Educational Technology in Higher Education, 17:1-18, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[17] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.348 + ], + "angle": 0, + "content": "[18] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.906, + 0.404 + ], + "angle": 0, + "content": "[19] Olga Megalakaki, Ugo Ballenghein, and Thierry Baccino. Effects of valence and emotional intensity on the comprehension and memorization of texts. Frontiers in Psychology, 10:179, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.906, + 0.462 + ], + "angle": 0, + "content": "[20] Liyu Meng, Yuchen Liu, Xiaolong Liu, Zhaopei Huang, Yuan Cheng, Meng Wang, Chuanhe Liu, and Qin Jin. Multimodal emotion estimation for in-the-wild videos. arXiv preprint arXiv:2203.13032, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.464, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[21] Laurent Mertens, Elahe Yargholi, Hans Op de Beeck, Jan Van den Stock, and Joost Vennekens. Findingemo: An image dataset for emotion recognition in the wild. Advances in Neural Information Processing Systems, 37:4956-4996, 2024. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.906, + 0.588 + ], + "angle": 0, + "content": "[22] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.906, + 0.646 + ], + "angle": 0, + "content": "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.906, + 0.688 + ], + "angle": 0, + "content": "[24] Chanjong Park and In-Kwon Lee. Emotional landscape image generation using generative adversarial networks. In Proceedings of ACCV, pages 573-590, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.906, + 0.746 + ], + "angle": 0, + "content": "[25] Monika Pawlowska and Ewa Magier-Lakomy. The influence of emotional and non-emotional concepts activation on information processing and unintentional memorizing. Polish Psychological Bulletin, pages 150-159, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of ICCV, pages 4195-4205, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[27] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of CVPR, pages 860-868, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, et al. SDXL: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of ICLR, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "15226" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of CVPR, pages 10684-10695, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of CVPR, pages 22500-22510, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.351 + ], + "angle": 0, + "content": "[33] James A Russell. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161, 1980. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.352, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[34] James A Russell and Albert Mehrabian. Evidence for a three-factor theory of emotions. Journal of research in Personality, 11(3):273-294, 1977. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.395, + 0.482, + 0.424 + ], + "angle": 0, + "content": "[35] S Shaphiro and MBJB Wilk. An analysis of variance test for normality. Biometrika, 52(3):591-611, 1965. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.425, + 0.482, + 0.481 + ], + "angle": 0, + "content": "[36] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839-84865, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[37] Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos, and Maja Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions. Nature Machine Intelligence, 3(1):42-50, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.54, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 8:235933, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.584, + 0.482, + 0.612 + ], + "angle": 0, + "content": "[39] Ashish Vaswani. Attention is all you need. In Proceedings of NeurIPS, page 6000-6010, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.654 + ], + "angle": 0, + "content": "[40] Manuelde Vega. The representation of changing emotions in reading comprehension. Cognition & emotion, 10(3):303-322, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of AAAI, pages 2555-2563, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.482, + 0.756 + ], + "angle": 0, + "content": "[42] Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li, et al. A systematic review on affective computing: Emotion models, databases, and recent advances. Information Fusion, 83:19-52, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.758, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[43] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of ICCV, pages 10810-10819, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[44] Frank Wilcoxon. Individual comparisons by ranking methods. In Breakthroughs in statistics: Methodology and distribution, pages 196-202. 1992. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[45] Weizhen Xie and Weiwei Zhang. Negative emotion enhances mnemonic precision and subjective feelings of remembering in visual long-term memory. Cognition, 166:73-83, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[46] Chao Xu, Junwei Zhu, Jiangning Zhang, Yue Han, Wenqing Chu, Ying Tai, et al. High-fidelity generalized emotional talking face generation with multi-modal emotion space learning. In Proceedings of CVPR, pages 6609-6619, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.164, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[47] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of CVPR, pages 9479-9488, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.209, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[48] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, et al. Emotion recognition for multiple context awareness. In Proceedings of ECCV, pages 144–162, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.265, + 0.905, + 0.321 + ], + "angle": 0, + "content": "[49] Jufeng Yang, Dongyu She, and Ming Sun. Joint image emotion classification and distribution learning via deep convolutional neural network. In Proceedings of IJCAI, pages 3266-3272, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.324, + 0.905, + 0.379 + ], + "angle": 0, + "content": "[50] Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emotional image content generation with text-to-image diffusion models. In Proceedings of CVPR, pages 6358-6368, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.382, + 0.905, + 0.437 + ], + "angle": 0, + "content": "[51] Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. arXiv preprint arXiv:2405.12661, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.44, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[52] Xiaocui Yang, Shi Feng, Daling Wang, and Yifei Zhang. Image-text multimodal emotion classification via multi-view attentional network. IEEE Transactions on Multimedia, 23: 4014-4026, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.556, + 0.905, + 0.612 + ], + "angle": 0, + "content": "[54] Ligang Zhang, Dian Tjondronegoro, and Vinod Chandran. Representation of facial expression categories in continuous arousal-valence space: feature and correlation. Image and Vision Computing, 32(12):1067-1079, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.614, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of ICCV, pages 3836-3847, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.657, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of CVPR, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.716, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[57] Zixing Zhang, Liyizhe Peng, Tao Pang, Jing Han, Huan Zhao, and Bjorn W Schuller. Refashioning emotion recognition modelling: The advent of generalised large models. IEEE Transactions on Computational Social Systems, pages 6690-6704, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.787, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[58] Sicheng Zhao, Yue Gao, Xiaolei Jiang, Hongxun Yao, TatSeng Chua, and Xiaoshuai Sun. Exploring principles-of-art features for image emotion recognition. In Proceedings of ACM MM, pages 47-56, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[59] Sicheng Zhao, Hongxun Yao, and Xiaolei Jiang. Predicting continuous probability distribution of image emotions in valence-arousal space. In Proceedings of ACM MM, page 879-882, 2015. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "15227" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[60] Sicheng Zhao, Xingxu Yao, Jufeng Yang, et al. Affective image content analysis: Two decades review and new perspectives. IEEE TPAMI, 44(10):6729-6751, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[61] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.19 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "15228" + } + ] +] \ No newline at end of file diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_origin.pdf b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d2d334d155057b39614e20fdda9e6cb9ff00c887 --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/64d56063-afb5-47eb-8d97-ccd27f1db0d7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9efb65929a850d584c48f517b6522651e42057e986f7b65dbad9dc6262b88435 +size 3641451 diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/full.md b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ffac558f59128add4a3b796af359f58762c9a1a9 --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/full.md @@ -0,0 +1,368 @@ +# EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model + +Shengqi Dang $^{1,2\dagger}$ , Yi He $^{1\dagger}$ , Long Ling $^{1}$ , Ziqing Qian $^{1}$ , Nanxuan Zhao $^{3}$ , Nan Cao $^{1,2*}$ + +$^{1}$ Tongji University $^{2}$ Shanghai Innovation Institute $^{3}$ Adobe Research + +![](images/f9dc9533db4779021a908e9427fceb336619f5bdf630b50450e8af0d377fc860.jpg) +Figure 1. Generated images of the EmotiCrafter. Given a prompt and Valence-Arousal (V-A) values, our method can generate an emotional image that reflects the input prompt and aligns with the specified emotion values. By adjusting the V-A values, our method can also generate images that evoke discrete emotions, such as sadness, relaxation, anger, and amusement. + +![](images/55dce0c3547afc774b1fb55f839f4a82b439cc06a2e13168250b292622a033ff.jpg) + +![](images/797cad69fe04c21a3acb13042d7e645b30333b9fab05414714b18187e0604097.jpg) +A man is running fast +A cartoon shows children playing football + +![](images/6207cb4e237e7a95d06c221e56f1eded15b1e3f069220485469848612e8967ac.jpg) +A group of people are reading books + +![](images/133bea0b51c8c34e94453b5de36915ec32431c45a14430b101e0d7aeccff4e3a.jpg) + +# Abstract + +Recent research shows that emotions can enhance users' cognition and influence information communication. While + +research on visual emotion analysis is extensive, limited work has been done on helping users generate emotionally rich image content. Existing work on emotional image generation relies on discrete emotion categories, making it challenging to capture complex and subtle emotional nuances accurately. Additionally, these methods struggle to control the specific content of generated images based on text prompts. In this paper, we introduce the task of continuous emotional image content generation (C-EICG) and present EmotiCrafter, a general emotional image generation model that generates images based on free text prompts and Valence-Arousal (V-A) values. It leverages a novel emotion-embedding mapping network to fuse V-A values + +into textual features, enabling the capture of emotions in alignment with intended input prompts. A novel loss function is also proposed to enhance emotion expression. The experimental results show that our method effectively generates images representing specific emotions with the desired content and outperforms existing techniques. + +# 1. Introduction + +Emotions are fundamental to human experiences and play a critical role in shaping how people perceive and interact with the world. Research has shown that emotions affect memory [15, 19, 25, 45] and comprehension [16, 38, 40], which are crucial for effective communication. As a result, content creators increasingly recognize the importance of incorporating emotions to enhance audience engagement. + +While research on visual emotion analysis is extensive [13, 49, 58], there is limited work on generating emotionally rich image content. Some early studies explored emotional content generation techniques within specific domains such as facial expressions [2, 46] or landscapes [24]. EmoGen [50] generates images based on a given emotion tag (e.g., happy or sad), which pioneered the domain-free emotional image content generation (EICG) task. However, it has two critical limitations: (1) the image is generated from an emotion tag instead of a text prompt, making the generated content difficult to control; (2) while the discrete emotion tags used in EmoGen are easy to understand, psychologists have not achieved a consensus on the emotion categories [60]. The limited scope of discrete emotion tags falls short of capturing nuanced emotions. + +To address the above issues, we propose the continuous emotional image content generation (C-EICG) task and present EmotiCrafter, the first C-EIGC model that generates emotional images using free-text prompts and continuous Valence-Arousal (V-A) values defined in the V-A model (a well-known psychological continuous emotion model) [33]. V-A model represents emotions in a two-dimensional Cartesian space (Figure 2), where Valence quantifies pleasantness (negative to positive) and Arousal measures intensity (calm to excited). We utilize a [-3, 3] range for Valence and Arousal [10, 21]. The continuous V-A space enables smooth transitions and nuanced emotional shifts beyond the capability of discrete labels (e.g., shifts from "bored" to "tired" of a character's state in a video). Specific emotional values in this space have been investigated in prior work [34], and such fine-grained modeling is particularly beneficial for human-computer interaction [2]. Leveraging this model, EmotiCrafter captures subtle affective variations via precise (V, A) positioning. The contributions of this paper are as follows: + +- We propose a novel task, continuous emotional image content generation (C-EICG), and develop the first ded + +![](images/9624c89c88a2a32c6ec937c2b1a82f28ed6403ab72b50c5a17fd755a11f5f52c.jpg) +Figure 2. Valence-Arousal model. + +icated model for this purpose. Our model introduces an emotion-embedding network that integrates continuous Valence-Arousal (V-A) values with text prompts. These fused features are then injected into Stable Diffusion XL [28] using cross-attention mechanisms, enabling precise control over both content and emotional expressions in the generated images. + +- We propose a novel loss function that enhances the emotional resonance of generated images. By amplifying the difference between neutral and emotional text features, our approach enables the model to capture more distinct emotional variations. Additionally, the loss function incorporates the V-A distribution to address data imbalance, further refining the model's ability to generate images with rich, accurate emotional expressions. +- We constructed an emotional prompts dataset to train the emotion-embedding network, where each sample consists of a neutral prompt and an emotional prompt that share the same core meaning but express a specific emotion corresponding to a given pair of V-A values. + +# 2. Related work + +In this section, we present a review of the related work, specifically focusing on visual emotion analysis, image emotion transfer, and conditional image generation. + +# 2.1. Visual Emotion Analysis + +Visual emotion analysis refers to the computational recognition and interpretation of human emotion [57] in visual media, such as images or videos. It has been a prominent research area, with most efforts focusing on the classification of discrete emotions [4, 42, 47, 48, 52]. However, discrete emotion categories limit the ability to capture the nuanced emotions [7], leading to an increased focus on continuous emotion analysis in images [59]. + +Much of the continuous emotion analysis remains centered on facial expression analysis [9, 11, 22, 37, 54]. While effective, facial analysis can overlook crucial con + +textual information that influences emotional interpretation. This drives studies to focus on emotions in objects or individuals within their environment, rather than just faces[12, 13, 20, 21]. For example, Kosti et al. [12] combined person-specific characteristics with the context of the scene to predict continuous emotional dimensions such as arousal, valence, and dominance. Kragel et al. [13] proposed EmoNet to extract visual features such as facial expressions, body posture, and scene elements from images to predict valence and arousal values. Recently, Mertens et al.[21] tested multiple backbones (such as ResNet, CLIP, DINO, etc.) for prediction of valence and arousal, demonstrating strong performance across various architectures. + +Previous studies have achieved high accuracy in predicting continuous emotions in images, demonstrating a correlation between continuous emotions and visual elements. Building on this insight, our work moves beyond prediction to explore how continuous emotions can be actively embedded within generated content. + +# 2.2. Image Emotion Transfer + +Image Emotion Transfer (IET) focuses on editing the content of the images to evoke different emotions [17, 27, 43, 61]. For instance, Peng et al. [27] achieved emotion transfer by adjusting color tones and texture-related features. Zhu et al. [61] introduced a method that separates high-level emotion-relevant features (e.g., object shapes and scene layout) from low-level emotion-relevant features (e.g., brightness). By applying GANs, they transferred emotions between images while preserving their original structure. Building on these methods, Weng et al. [43] proposed the Affective Image Filter, which uses a multi-modal transformer to process both text and image inputs. With the emergence of text-to-image models, IET has expanded into new applications based on instructive commands. For example, EmoEdit [51] used GPT-4V to build emotion factor trees that map abstract emotions to specific visual elements and employed the InstructPix2Pix model to apply emotion-driven content and color adjustments to images. + +Current methods primarily extract specific features, focusing on certain visual elements or emotional cues, which can limit the depth of emotional expression. In contrast, our method broadly learns a variety of emotion-influencing features, and it could accept natural language prompts as input. + +# 2.3. Conditional Image Generation + +Conditional image generation aims to create images that align with specific input conditions, such as text [3, 28, 30, 31], reference images [53], subjects [6, 32], and depth maps [23, 55]. To enhance quality, researchers have developed specialized approaches, such as Diffusion Transformers (DiT)[26], which use transformer-based diffusion for denoising, and Visual Autoregressive Modeling (VAR)[36], + +which encodes images into discrete tokens for autoregressive prediction across scales. However, these state-of-the-art methods rely on discrete labels, limiting flexibility and control. In contrast, our approach offers greater generalization by leveraging Stable Diffusion XL [28] to generate emotional images from free-text prompts. + +Despite these progresses, incorporating emotion as a condition for image generation remains underexplored. EmoGen [50] pioneered the generation of emotional image content (EIGC) by mapping emotional features to semantic features to generate emotional images. However, EmoGen struggles to understand natural language, limiting its capacity to effectively control specific content. Its reliance on discrete emotions also limits its practical applicability. + +Our method bridges this gap by embedding continuous emotion into textual features, enabling image generation models to use continuous emotion for emotion control. Unlike label-based methods, our method supports free-text prompts for flexible content control. + +# 3. Method + +In this section, we introduce the technique details of the proposed EmotiCrafter. + +# 3.1. Overview + +Our method generates emotional images $I_{emo}$ from two inputs (Figure 3): a free-text prompt describing the desired content, and a pair of V-A values $(v, a)$ specifying the emotion. First the prompt encoder $\mathcal{E}$ converts the text prompt into feature $f_{n}$ , which is then processed by an emotion-embedding network $\mathcal{M}$ to produce emotional prompt feature $\hat{f}_{e}$ that integrate the V-A values: + +$$ +\hat {f} _ {e} = \mathcal {M} \left(f _ {n} | (v, a)\right) \tag {1} +$$ + +Next, this feature is injected into Stable Diffusion XL $\mathcal{G}$ via its cross-attention mechanism to generate the emotional images: $I_{emo} = \mathcal{G}(\hat{f}_e)$ . To enhance emotional expressiveness, we introduce a loss function (Fig.3(b.2)) that leverages the V-A distribution and emphasizes the differences between neutral and emotional prompt features. This ensures that the generated images accurately convey both the intended emotions and content. Additionally, we construct a dataset (Fig.3(a)) that pairs neutral and emotional prompts with the corresponding V-A values. + +# 3.2. Emotion-Embedding Network + +The emotion-embedding network $\mathcal{M}$ generates the emotional prompt feature by integrating a pair of V-A values with a neutral prompt feature (Fig.3(b.1)). First, a V-A encoder converts the V-A values into feature vectors. Then, an emotion injection transformer—modified from GPT-2[29]—fuses these vectors with the neutral prompt + +![](images/b6efb014286d027a882699f6303429d453f12f20c4e9240f0d8501500385d5dd.jpg) +Figure 3. Overview of our method. Specifically, we take the following steps. (a) We collect an image dataset annotated with V-A values, neutral prompts, and emotional prompts. These prompts are then encoded into features by prompt encoder $\mathcal{E}$ . (b) Next, we design (b.1) an emotion-embedding network $\mathcal{M}$ to embed V/A values into textual features based on the transformer architecture, and (b.2) a specialized loss function to enhance the emotional resonance of generated images. The output of the mapping network serves as the condition for the image generation model $\mathcal{G}$ to generate emotional images. + +![](images/86bc912c030cf236269f88e37e87c6e9258c5e203b081b03668ec9abea346b32.jpg) +Figure 4. Structure of Emotion Injection Block. It accepts hidden state $h_{i-1}$ as input and produces $h_i$ as output. The V-feature $e_v$ and A-feature $e_a$ represent the emotion features, which are injected through the cross-attention module. + +feature, preserving the original textual context while infusing emotional content. + +V-A Encoder. The V-A Encoder converts a pair of V-A values into feature vectors using two separate multilayer perceptrons (MLPs). One MLP processes the Valence value to produce V-feature $e_v$ , and the other processes the Arousal value to generate A-feature $e_a$ . These features are then fed into the emotion injection transformer network for further emotion infusion. + +Emotion Injection Transformer (EIT). The Emotion Injection Transformer (EIT) leverages a modified GPT-2 architecture to seamlessly integrate V/A-features into textual features. Its process consists of three stages: input projection, emotion injection, and output projection. + +First, we project the input neutral prompt feature $f_{n}$ into the transformer's feature space: + +$$ +h _ {0} = \mathrm {P} _ {\text {i n}} \left(f _ {n}\right) + \mathrm {P E}, \tag {2} +$$ + +where $h_0$ represents the initial hidden state; $\mathrm{P_{in}(\cdot)}$ is a linear projection layer, and PE denotes positional embedding [39]. + +In the next, we inject emotion into $f_{n}$ via 12 sequential Emotion Injection Blocks (EIBs) corresponding to 12 transformer blocks. Each block outputs a hidden state: + +$$ +h _ {i} = \operatorname {E I B} \left(h _ {i - 1}, e _ {v}, e _ {a}\right), \quad i \in \{1, \dots , 1 2 \} \tag {3} +$$ + +where $h_i$ is the output of the $i$ -th EIB( $\cdot$ ). As shown in Figure 4, each EIB enhances the transformer block through a cross-attention mechanism: + +$$ +h _ {i} ^ {\prime} = \text {s e l f - a t t n} (\mathrm {L N} (h _ {i - 1})) + h _ {i - 1} \tag {4} +$$ + +$$ +h _ {i} ^ {(v)} = \operatorname {c r o s s - a t t n} \left(\mathrm {L N} \left(h _ {i} ^ {\prime}\right), e _ {v}\right) + h _ {i} ^ {\prime} \tag {5} +$$ + +$$ +h _ {i} ^ {(v, a)} = \operatorname {c r o s s - a t t n} \left(\mathrm {L N} \left(h _ {i} ^ {(v)}\right), e _ {a}\right) + h _ {i} ^ {(v)} \tag {6} +$$ + +$$ +h _ {i} = \operatorname {f n n} \left(\mathrm {L N} \left(h _ {i} ^ {(v, a)}\right)\right) + h _ {i} ^ {(v, a)} \tag {7} +$$ + +where $h^{\prime}, h^{(v)}, h^{(v,a)}$ are the intermediate hidden variables; $\mathrm{LN}(\cdot)$ is the LayerNorm; self-attn( $\cdot$ ) denotes self-attention, employed to capture context dependencies; cross-attn( $\cdot$ ) is cross-attention for injecting $e_v$ and $e_a$ ; $\mathrm{fnn}(\cdot)$ is a feedforward network that adapts the complexity of the emotional embedding process. We also remove the causal mask typically used for next-token prediction from the original transformer model to fit our task. + +Finally, the output of the last (i.e., the 12th) injection block $h_{12}$ is projected back to SDXL's prompt feature space via $\mathbf{P}_{\mathrm{out}}$ (a linear layer and a LayerNorm layer) to predict the residual between emotional and neutral prompt features, which represents a semantic shift between emotional and neutral prompts: + +$$ +\hat {f} _ {r} = \mathrm {P} _ {\text {o u t}} \left(\mathrm {L N} \left(h _ {1 2}\right)\right) \tag {8} +$$ + +The final emotional prompt feature is obtained by adding this residual to the original neutral prompt feature: + +$$ +\hat {f} _ {e} = \hat {f} _ {r} + f _ {n} \tag {9} +$$ + +The above emotion embedding network is trained by minimizing the averaged expectation of the difference between the predicted emotional prompt feature $\hat{f}_e = \mathcal{M}(f_n|(v,a))$ and the scaled target emotional prompt feature $f_e^t$ , using the loss function described in Equation 10. + +$$ +\mathcal {L} = \frac {1}{n} \mathbb {E} \left(\frac {1}{d (v , a)} \| \hat {f} _ {e} - f _ {e} ^ {t} \| ^ {2}\right) \tag {10} +$$ + +where $n$ is the number of feature elements; $\mathbb{E}(\cdot)$ denotes the expectation; $d(v, a)$ is a density function that describes the distribution of V-A values in the training sample. + +To effectively embed emotions and address the challenges posed by the uneven distribution of V-A values in the dataset, this loss function incorporates two key strategies to improve the model's performance: + +Scaled Residual Learning. To better capture pronounced emotional changes in generated images, we enlarge the target residuals: + +$$ +f _ {e} ^ {t} = f _ {n} + \alpha \underbrace {\left(f _ {e} - f _ {n}\right)} _ {\text {r e s i d u a l f e a t u r e}}, \tag {11} +$$ + +where $f_{e}$ is the emotional prompt feature, $f_{n}$ is the neutral prompt feature, and $\alpha$ is a scale factor, we set its value to 1.5 based on the ablation study. + +V-A Density Weighting. To mitigate the effects of the imbalanced distribution of the training samples, we weigh the loss inversely proportional to the density of training samples in the V-A space. The density is estimated using Kernel Density Estimation (KDE) [5] with a Gaussian kernel, denoted as $d(v,a)$ : + +$$ +d (v, a) = \frac {1}{n} \sum_ {i = 1} ^ {n} K _ {H} ((v, a) - (v _ {i}, a _ {i})), \tag {12} +$$ + +where $K_{H}$ is a 2D Gaussian kernel with bandwidth $H$ ; $n$ is the number of training samples; $(v_{i},a_{i})$ are the V-A values of the $i$ -th training sample. The bandwidth $H$ is selected using Silverman's rule of thumb to provide optimal smoothing of the density estimation. + +# 3.3. Dataset and Training + +To train the emotion-embedding network, we constructed a dataset of paired neutral and emotional prompts with corresponding Valence-Arousal (V-A) values (Figure 3(a)). These pairs were automatically generated using GPT-4 based on 39,843 images with human-annotated V-A values from publicly available datasets, including OASIS [14], EMOTIC [12], and FindingEmo [21]. Specifically, GPT-4 was used to generate neutral prompts with objective image descriptions and emotional prompts that emphasize affective attributes such as color, lighting, and texture, which influence emotional perception. To ensure data reliability, all LLM-generated prompts underwent crowd-sourced verification, with disagreements resolved through a voting mechanism among annotators. + +The proposed emotion embedding network is trained on two NVIDIA A800 GPUs using the aforementioned dataset. We employ the AdamW [18] optimizer with a weight decay + +of 1e-5 and a learning rate of 1e-3. The training process spans 200 epochs with a batch size of 768, completing in approximately 7 to 8 hours. + +# 4. Evaluation + +# 4.1. Generation Results + +Figure 1 shows the proposed technique's ability to achieve continuous and effective control over both emotion and content during image generation. Meanwhile, Figure 5 highlights four key capabilities: (a) emotion-content decoupling, where V-A values override emotional cues in the prompt, allowing typically positive concepts to be rendered with negative emotions; (b) compatibility with discrete emotion categories; (c) content-independent generation, where images generated from empty prompts and specified V-A values maintain emotional consistency without semantic constraints; and (d) fine-grained emotional control, demonstrated through V-A increments of 0.2, showcasing the model's sensitivity to subtle emotional variations. + +# 4.2. Comparisons + +To estimate the effectiveness of the proposed method, we built four baselines based on existing techniques. + +Baselines. We established baselines for comparison using two strategies: (1) directly injecting emotion features into the image generation modules of SDXL, such as UNet, and (2) modifying the input text prompt using emotion features to influence the generated image's content, similar to the proposed method. + +As a result, four different baselines were built: (1) Cross Attention: inject the emotion features $(e_v, e_a)$ into the UNet in SDXL via its cross-attention mechanism based on IP-Adapter [53]; (2) Time Embedding: directly add emotion features $(e_v, e_a)$ to the time embedding of the UNet in SDXL. (3) Textual Inversion: use the text inversion technique [6] to embed emotion features $(e_v, e_a)$ into prompt templates with predefined emotion placeholders. (4) GPT-4+SDXL (GPT-SD): use GPT-4 [1] to rewrite the input text according to (v,a) values to generate an emotional SDXL prompt for image generation. + +Qualitative Comparison. We evaluate the generated images based on three criteria: (1) the effectiveness of emotion embedding, (2) image-prompt similarity, and (3) the continuity of emotional variations as V-A values change. Figure 6 demonstrates that the baseline methods—Cross Attention, Time Embedding, and Textual Inversion—tend to produce nearly identical outputs regardless of emotional variation. This is because the loss terms of these baselines primarily align low-level image features (e.g., SDXL's latent space), which are highly correlated with prompt content but struggle to capture subtle emotional cues when both prompts and V-A values are provided. Notably, Tex- + +![](images/bba03a1446d97cabe0e071e1851f7434ab9605a8260b4e84c5e83090ecf6fd34.jpg) +(a) +(b) + +![](images/8c02341b7eb8fe02f47bdb2914dcda120616909acc139cea56b3820fc8bced53.jpg) + +![](images/c3ea3c2ca75d16339a200ea498be0e290ed9f8f7898b6df493694aeed647d402.jpg) + +![](images/7246a3cd59b6b304cac1ac54cf48949b2ece02cb95fb761f3d677a85d52e2f62.jpg) + +![](images/04b476acc97c862b300ac66bd4d4db2f86f5a04a77cef2867e7a552d59f2c69e.jpg) + +![](images/4a7e7a5c16f462a4c1676792605a3a2c0e990deedbc06e8892190c2ffaecfe3f.jpg) +1) +(c) + +![](images/741da1258ba440b8884678442a21ec33513b45280af59afc71bb6eecba5e314b.jpg) +$\left\{ {2 : {x}^{\prime }\left( x\right) = {y}^{\prime }\left( x\right) }\right\} : \left\{ {2 : {x}^{\prime }\left( x\right) = {y}^{\prime }\left( x\right) }\right\} : y = x$ + +![](images/26f2a0cb69f8bb32498994df30781659074ec362a6b4b19631d06fa9a3e3a91f.jpg) + +![](images/75770b5fdbd7dade7d4f1cbd2317355a6922ed0bb85ff199c0feb74a9546c741.jpg) + +![](images/853b3c65492d661868b046a816bdb95e1f56d2909eb2e5e6ac93f8ac86b60afe.jpg) + +![](images/9f17bff65dc980cc74376cc6e18e6e732917c86e995f3106e793171699f33ebc.jpg) +Figure 5. Results under multiple inputs. (a) Overriding semantic content ('a child in the amusement park') with sad V-A (-2,-2); (b) Discrete emotion mapping in V-A space as emotion input; (c) Empty-prompt generation with pure emotion condition; (d) Fine-grained control of V-A variations with a granularity of 0.2. +Figure 6. Qualitative comparisons with baselines. These images are generated at varying V-A values, specifically -1.5, 0, and 1.5. Only our approach and the GPT-4+SDXL successfully generate images that clearly reflect emotional variations. Notably, our results show enhanced continuity, indicating superior controllability over continuous V-A values compared to the GPT-4+SDXL. + +tual Inversion often produces images with a persistent purple tint. Furthermore, all methods faithfully generate content aligned with the given prompts. However, for image continuity, we focus on comparing our method with GPT-4+SDXL, as they are the only two capable of generating distinct emotional variations. Figure 7 provides a more detailed comparison, showing that our method maintains smooth emotional transitions even under extreme V-A conditions, whereas GPT-4+SDXL introduces noticeable discontinuities (e.g., in the $\mathrm{V} = 3$ column). + +Quantitative Comparison. We compare our method against several baselines using the following metrics: (1) $\underline{\mathrm{V} / \mathrm{A} - \mathrm{Error}}$ evaluates the absolute error between the predicted V/A of the generated images and the input V/A. (2) CLIPScore [8] assesses the similarity between the input text and the generated images. (3) CLIP-IQA [41] leverages a pre-trained CLIP model to evaluate image quality without requiring reference images. (4) LPIPS-Continuous utilizes + +
A-Error ↓V-Error ↓CLIPScore ↑CLIP-IQA ↑
Cross Attention1.923±1.1532.080±1.43826.266±2.3810.949±0.046
Time Embedding1.941±1.1682.031±1.34826.566±2.1250.786±0.164
Textual Inversion1.958±1.1881.923±1.17022.346±3.5940.370±0.111
GPT-4+SDXL1.860±1.0901.517±1.06025.907±1.9490.906±0.066
Ours1.828±1.0851.510±1.07423.067±2.6550.881±0.099
+ +Table 1. Comparison on emotion accuracy, prompt fidelity, and image quality across different baselines, evaluated on 3,300 images per method (132 prompts $\times$ 5 V values $\times$ 5 A values). Our method achieves the highest performance in emotion accuracy while maintaining comparable results in prompt fidelity and image quality. The slight decrease in prompt fidelity is expected, as modifying emotional content affects semantic alignment. + +the Learned Perceptual Image Patch Similarity [56] to measure the continuity of the change in image as V/A changes. + +As shown in Table 1, our method achieves the lowest (best) $\underline{V / A - Error}$ on average. While Cross Attention and Time Embedding achieve the highest $\underline{CLIP - IQA}$ and $\underline{CLIPScore}$ , respectively, these methods fail to generate + +![](images/16012fcf819496df2d55baca64f95cd2d704ca4f067f84b38ee5f5ed4b266430.jpg) +Figure 7. A more comprehensive comparison with the GPT-4+SDXL for the prompt "A group of people is skiing in a snow hill." Our approach maintains continuity even under extreme V/A conditions. Conversely, GPT-4+SDXL displays noticeable discontinuities (e.g., in the $\mathrm{V} = 3$ column). + +emotionally expressive images (Figure 6). + +Our method exhibits a slight decrease in CLIPScore compared to the baselines, which we attribute to the inherent trade-off between emotional modulation and strict semantic alignment. However, the high CLIP-IQA score indicates that our method produces high-quality images. Additionally, as shown in Table 2, our approach demonstrates superior continuity compared to GPT-4 + SDXL. + +
OursGPT-4+SDXL
LPIPS-Continuous↓0.220±0.0640.361±0.059
+ +Table 2. Continuity comparison between our method and baseline. + +
OursGPT-4+SDXL
Study IA-Ranking Consistency ↑0.759±0.2730.165±0.379
V-Ranking Consistency ↑0.887±0.2450.584±0.259
A-Error ↓1.327±1.1202.029±1.446
V-Error ↓0.692±0.6821.229±1.026
Study IIEmotion Consistency ↑4.215±0.7153.525±1.065
Emotion Smoothness ↑4.240±0.8283.195±1.163
+ +Table 3. Our method outperformed the baseline. + +![](images/652a060b2d3ed81b304f1994db701d6714a9b96b18dbc10ef2e6c96670ef485e.jpg) +Figure 8. User Study Results ( $^{*}\mathsf{p} < 0.05$ $^{**}\mathsf{p} < 0.01$ $^{**}p < 0.001$ + +# 4.3. User Study + +We conducted two user studies with 20 college students to evaluate the effectiveness of our method by comparing it to $GPT-4 + SDXL$ (the baseline). + +Study I. In the first experiment, we evaluated whether the generated images' emotions align with human perception. Two image collections were prepared—one for Arousal (A) and one for Valence (V)—each comprising 20 sets (10 from our method and 10 from the baseline) of 5 randomized images with varying A or V values. Participants reordered the images by perceived intensity and estimated each image's V and A values. We then computed Kendall's $\tau_{b}$ (with $\tau_{b} = 1$ indicating perfect alignment) to assess ordering accuracy and calculated the absolute error between the estimated and ground truth values. + +Study II. In the second experiment, we assessed whether the generated images could effectively reflect continuous emotional changes (i.e., V-A values). We generated 20 image sets (10 per method), each containing 25 images with V-A values varying gradually from -3 to +3 (Figure 7). Participants rated each set on a 5-point Likert scale regarding (1) the alignment between V-A changes and image content, and (2) the smoothness of the content transition. + +Analysis & Results. We conducted a Shapiro-Wilk test [35] to assess normality and applied the Wilcoxon Signed Rank test [44] to evaluate statistical significance. Our results indicate that our method outperformed the baseline across all metrics (Table 3), with statistically significant improvements in V/A Ranking, V/A Error, Consistency, and Smoothness (Figure 8). + +![](images/49d7a374a28ece4a1b71699e296223fc2af1c98585540d696136733c83d8f2ba.jpg) +Figure 9. Abalation Study. Images are generated from the prompt "An oil painting shows an astronaut." As $\alpha$ increases, image-prompt similarity decreases, while emotional variations increase. The usage of $d(a,v)$ enhances the accuracy of emotional changes. + +![](images/d3b75a7c34fb19af0a35185edcbb56db21e04b2aff75fff5dc97d181fe1b6946.jpg) +Figure 10. The effectiveness of the scaling factor $\alpha$ . The method with $\alpha = 1.5$ ( $\star$ ) surpasses the performance indicated by the regression line $(- - )$ . + +
A-Error ↓V-Error ↓CLIPScore ↑
Ours1.828±1.0851.510±1.07423.067±2.655
w/o d(v, a)1.829±1.0831.546±1.08221.977±0.066
+ +Table 4. The effectiveness of $d(v,a)$ + +# 4.4. Ablation Study + +We performed ablation experiments to assess the contribution and effectiveness of the proposed loss function. + +Effectiveness of the Scaling Factor $\alpha$ . We evaluated how varying $\alpha$ affects CLIPScore and V/A-Error (Figure 10). As $\alpha$ increases, CLIPScore decreases (indicating reduced semantic alignment), while V/A-Error also decreases (indicating improved emotional accuracy). This trend is further illustrated by the examples in Figure 9. Based on these findings, we set $\alpha = 1.5$ as an optimal tradeoff, though users can adjust $\alpha$ to suit their specific needs. + +Effectiveness of the Density Weighting $d(v,a)$ . We compare our full method with a variant that omits d(v,a) + +from the loss function. As shown in Table 4, including $\mathrm{d}(\mathrm{v},\mathrm{a})$ leads to improvements in both CLIPScore and V/A-Error. This positive effect is further illustrated in Figure 9. + +# 5. Conclusion and Limitations + +In this paper, we introduce continuous emotional image content generation (C-EICG) and present EmotiCrafter, a novel method that generates emotionally expressive images using continuous Valence-Arousal (V-A) values. Our emotion-embedding network integrates V-A values into textual features, and extensive experiments show that our approach reliably aligns images with both user prompts and specified emotions. We believe this work will advance affective computing and image generation, and we will release our code and data to foster further research. + +Although our method achieved promising results, it still has some limitations need to be addressed in the future. First, controlling image generation based on arousal remains more challenging than controlling based on valence. This is consistent with prior research in visual emotion analysis, which has found that arousal is harder to predict due to lower inter-annotator agreement [21]. Second, our approach frequently generates images featuring human activities even when such activities are not mentioned in the prompts. This likely stems from the limited representation of non-human scenes in our training data and could be mitigated by incorporating a more diverse range of non-human scenarios. Third, our method occasionally modifies users' input prompts to better align with the specified emotional prompts, resulting in a slight semantic shift that affects the generated images. We believe this issue could be addressed by adding a semantic preservation term to the loss function. + +# 6. Acknowledgments + +Nan Cao is the corresponding author. This work was supported by the National Key Research and Development Program of China (2023YFB3107100). + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5 +[2] Bita Azari and Angelica Lim. Emostyle: One-shot facial expression editing using continuous emotion parameters. In Proceedings of WACV, pages 6385-6394, 2024. 2 +[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3 +[4] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of ACM MM, pages 223-232, 2013. 2 +[5] Yen-Chi Chen. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187, 2017. 5 +[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, et al. An image is worth one word: Personalizing text-to-image generation using textual inversion. In Proceedings of ICLR, 2023. 3, 5 +[7] Hatrice Gunes, Björn Schuller, Maja Pantic, and Roddy Cowie. Emotion representation, analysis and synthesis in continuous space: A survey. In Proceedings of IEEE international conference on automatic face & gesture recognition, pages 827-834, 2011. 2 +[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6 +[9] Stephen Khor Wen Hwoo, Alice Othmani, and Aznul Qalid Md. Sabri. Deep learning-based approach for continuous affect prediction from facial expression images in valence-arousal space. IEEE Access, 10:96053-96065, 2022. 2 +[10] Ahmed Khota, Eric Cooper, Yu Yan, and Mate Kovacs. Modelling emotional valence and arousal of non-linguistic utterances for sound design support. In Proceedings of the International Conference on Kansei Engineering and Emotion Research, pages 507–516, 2022. 2 +[11] Dimitrios Kollias. Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges. In Proceedings of CVPR, pages 2328-2336, 2022. 2 +[12] Ronak Kosti, Jose M. Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of CVPR, pages 1667-1675, 2017. 3, 5 + +[13] Philip A Kragel, Marianne C Reddan, Kevin S LaBar, and Tor D Wager. Emotion schemas are embedded in the human visual system. Science advances, 5(7):eaaw4358, 2019. 2, 3 +[14] Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. Introducing the open affective standardized image set (oasis). Behavior research methods, 49:457-470, 2017. 5 +[15] Angela Y Lee and Brian Sternthal. The effects of positive mood on memory. Journal of consumer research, 26(2):115-127, 1999. 2 +[16] Jiansheng Li, Chuanlan Luo, Qi Zhang, and Rustam Shadiev. Can emotional design really evoke emotion in multimedia learning? International Journal of Educational Technology in Higher Education, 17:1-18, 2020. 2 +[17] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3 +[18] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[19] Olga Megalakaki, Ugo Ballenghein, and Thierry Baccino. Effects of valence and emotional intensity on the comprehension and memorization of texts. Frontiers in Psychology, 10:179, 2019. 2 +[20] Liyu Meng, Yuchen Liu, Xiaolong Liu, Zhaopei Huang, Yuan Cheng, Meng Wang, Chuanhe Liu, and Qin Jin. Multimodal emotion estimation for in-the-wild videos. arXiv preprint arXiv:2203.13032, 2022. 3 +[21] Laurent Mertens, Elahe Yargholi, Hans Op de Beeck, Jan Van den Stock, and Joost Vennekens. Findingemo: An image dataset for emotion recognition in the wild. Advances in Neural Information Processing Systems, 37:4956-4996, 2024. 2, 3, 5, 8 +[22] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 2 +[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3 +[24] Chanjong Park and In-Kwon Lee. Emotional landscape image generation using generative adversarial networks. In Proceedings of ACCV, pages 573-590, 2020. 2 +[25] Monika Pawlowska and Ewa Magier-Lakomy. The influence of emotional and non-emotional concepts activation on information processing and unintentional memorizing. Polish Psychological Bulletin, pages 150-159, 2011. 2 +[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of ICCV, pages 4195-4205, 2023. 3 +[27] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of CVPR, pages 860-868, 2015. 3 +[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, et al. SDXL: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of ICLR, 2024. 2, 3 + +[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3 +[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3 +[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of CVPR, pages 10684-10695, 2022. 3 +[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of CVPR, pages 22500-22510, 2023. 3 +[33] James A Russell. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161, 1980. 2 +[34] James A Russell and Albert Mehrabian. Evidence for a three-factor theory of emotions. Journal of research in Personality, 11(3):273-294, 1977. 2 +[35] S Shaphiro and MBJB Wilk. An analysis of variance test for normality. Biometrika, 52(3):591-611, 1965. 7 +[36] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839-84865, 2025. 3 +[37] Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos, and Maja Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions. Nature Machine Intelligence, 3(1):42-50, 2021. 2 +[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 8:235933, 2017. 2 +[39] Ashish Vaswani. Attention is all you need. In Proceedings of NeurIPS, page 6000-6010, 2017. 4 +[40] Manuelde Vega. The representation of changing emotions in reading comprehension. Cognition & emotion, 10(3):303-322, 1996. 2 +[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of AAAI, pages 2555-2563, 2023. 6 +[42] Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li, et al. A systematic review on affective computing: Emotion models, databases, and recent advances. Information Fusion, 83:19-52, 2022. 2 +[43] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of ICCV, pages 10810-10819, 2023. 3 +[44] Frank Wilcoxon. Individual comparisons by ranking methods. In Breakthroughs in statistics: Methodology and distribution, pages 196-202. 1992. 7 +[45] Weizhen Xie and Weiwei Zhang. Negative emotion enhances mnemonic precision and subjective feelings of remembering in visual long-term memory. Cognition, 166:73-83, 2017. 2 + +[46] Chao Xu, Junwei Zhu, Jiangning Zhang, Yue Han, Wenqing Chu, Ying Tai, et al. High-fidelity generalized emotional talking face generation with multi-modal emotion space learning. In Proceedings of CVPR, pages 6609-6619, 2023. 2 +[47] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of CVPR, pages 9479-9488, 2022. 2 +[48] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, et al. Emotion recognition for multiple context awareness. In Proceedings of ECCV, pages 144–162, 2022. 2 +[49] Jufeng Yang, Dongyu She, and Ming Sun. Joint image emotion classification and distribution learning via deep convolutional neural network. In Proceedings of IJCAI, pages 3266-3272, 2017. 2 +[50] Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emotional image content generation with text-to-image diffusion models. In Proceedings of CVPR, pages 6358-6368, 2024. 2, 3 +[51] Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. arXiv preprint arXiv:2405.12661, 2024. 3 +[52] Xiaocui Yang, Shi Feng, Daling Wang, and Yifei Zhang. Image-text multimodal emotion classification via multi-view attentional network. IEEE Transactions on Multimedia, 23: 4014-4026, 2021. 2 +[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 5 +[54] Ligang Zhang, Dian Tjondronegoro, and Vinod Chandran. Representation of facial expression categories in continuous arousal-valence space: feature and correlation. Image and Vision Computing, 32(12):1067-1079, 2014. 2 +[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of ICCV, pages 3836-3847, 2023. 3 +[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of CVPR, pages 586-595, 2018. 6 +[57] Zixing Zhang, Liyizhe Peng, Tao Pang, Jing Han, Huan Zhao, and Bjorn W Schuller. Refashioning emotion recognition modelling: The advent of generalised large models. IEEE Transactions on Computational Social Systems, pages 6690-6704, 2024. 2 +[58] Sicheng Zhao, Yue Gao, Xiaolei Jiang, Hongxun Yao, TatSeng Chua, and Xiaoshuai Sun. Exploring principles-of-art features for image emotion recognition. In Proceedings of ACM MM, pages 47-56, 2014. 2 +[59] Sicheng Zhao, Hongxun Yao, and Xiaolei Jiang. Predicting continuous probability distribution of image emotions in valence-arousal space. In Proceedings of ACM MM, page 879-882, 2015. 2 + +[60] Sicheng Zhao, Xingxu Yao, Jufeng Yang, et al. Affective image content analysis: Two decades review and new perspectives. IEEE TPAMI, 44(10):6729-6751, 2022. 2 +[61] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3 \ No newline at end of file diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/images.zip b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4174e559bd1a4c6ed32178a2cb2b1565f541f9aa --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2501cdba8000660edba070978fccd17b65b58c613c97c92d4083048ec9e16abb +size 1056091 diff --git a/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/layout.json b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68e1bceacc4826f1674a975ded23d10f4e6525c5 --- /dev/null +++ b/2025/EmotiCrafter_ Text-to-Emotional-Image Generation based on Valence-Arousal Model/layout.json @@ -0,0 +1,9795 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "spans": [ + { + "bbox": [ + 115, + 103, + 496, + 138 + ], + "type": "text", + "content": "EmotiCrafter: Text-to-Emotional-Image Generation based on Valence-Arousal Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "spans": [ + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": "Shengqi Dang " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": ", Yi He " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": ", Long Ling " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": ", Ziqing Qian " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": ", Nanxuan Zhao " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "text", + "content": ", Nan Cao " + }, + { + "bbox": [ + 96, + 160, + 512, + 177 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "spans": [ + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "text", + "content": "Tongji University " + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "text", + "content": "Shanghai Innovation Institute " + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 130, + 188, + 477, + 204 + ], + "type": "text", + "content": "Adobe Research" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 57, + 232, + 317, + 504 + ], + "blocks": [ + { + "bbox": [ + 57, + 232, + 317, + 504 + ], + "lines": [ + { + "bbox": [ + 57, + 232, + 317, + 504 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 317, + 504 + ], + "type": "image", + "image_path": "f9dc9533db4779021a908e9427fceb336619f5bdf630b50450e8af0d377fc860.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 511, + 555, + 544 + ], + "lines": [ + { + "bbox": [ + 55, + 511, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 555, + 544 + ], + "type": "text", + "content": "Figure 1. Generated images of the EmotiCrafter. Given a prompt and Valence-Arousal (V-A) values, our method can generate an emotional image that reflects the input prompt and aligns with the specified emotion values. By adjusting the V-A values, our method can also generate images that evoke discrete emotions, such as sadness, relaxation, anger, and amusement." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 250, + 432, + 365 + ], + "blocks": [ + { + "bbox": [ + 317, + 250, + 432, + 365 + ], + "lines": [ + { + "bbox": [ + 317, + 250, + 432, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 250, + 432, + 365 + ], + "type": "image", + "image_path": "55dce0c3547afc774b1fb55f839f4a82b439cc06a2e13168250b292622a033ff.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 365, + 431, + 494 + ], + "blocks": [ + { + "bbox": [ + 317, + 365, + 431, + 494 + ], + "lines": [ + { + "bbox": [ + 317, + 365, + 431, + 494 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 431, + 494 + ], + "type": "image", + "image_path": "797cad69fe04c21a3acb13042d7e645b30333b9fab05414714b18187e0604097.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 495, + 407, + 503 + ], + "lines": [ + { + "bbox": [ + 341, + 495, + 407, + 503 + ], + "spans": [ + { + "bbox": [ + 341, + 495, + 407, + 503 + ], + "type": "text", + "content": "A man is running fast" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 495, + 553, + 504 + ], + "lines": [ + { + "bbox": [ + 427, + 495, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 427, + 495, + 553, + 504 + ], + "type": "text", + "content": "A cartoon shows children playing football" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 434, + 250, + 551, + 365 + ], + "blocks": [ + { + "bbox": [ + 434, + 250, + 551, + 365 + ], + "lines": [ + { + "bbox": [ + 434, + 250, + 551, + 365 + ], + "spans": [ + { + "bbox": [ + 434, + 250, + 551, + 365 + ], + "type": "image", + "image_path": "6207cb4e237e7a95d06c221e56f1eded15b1e3f069220485469848612e8967ac.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 365, + 549, + 375 + ], + "lines": [ + { + "bbox": [ + 436, + 365, + 549, + 375 + ], + "spans": [ + { + "bbox": [ + 436, + 365, + 549, + 375 + ], + "type": "text", + "content": "A group of people are reading books" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 378, + 552, + 494 + ], + "blocks": [ + { + "bbox": [ + 434, + 378, + 552, + 494 + ], + "lines": [ + { + "bbox": [ + 434, + 378, + 552, + 494 + ], + "spans": [ + { + "bbox": [ + 434, + 378, + 552, + 494 + ], + "type": "image", + "image_path": "133bea0b51c8c34e94453b5de36915ec32431c45a14430b101e0d7aeccff4e3a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 555, + 200, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 555, + 200, + 567 + ], + "spans": [ + { + "bbox": [ + 152, + 555, + 200, + 567 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 297, + 605 + ], + "type": "text", + "content": "Recent research shows that emotions can enhance users' cognition and influence information communication. While" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 556, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 556, + 712 + ], + "type": "text", + "content": "research on visual emotion analysis is extensive, limited work has been done on helping users generate emotionally rich image content. Existing work on emotional image generation relies on discrete emotion categories, making it challenging to capture complex and subtle emotional nuances accurately. Additionally, these methods struggle to control the specific content of generated images based on text prompts. In this paper, we introduce the task of continuous emotional image content generation (C-EICG) and present EmotiCrafter, a general emotional image generation model that generates images based on free text prompts and Valence-Arousal (V-A) values. It leverages a novel emotion-embedding mapping network to fuse V-A values" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 617, + 261, + 628 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 617, + 261, + 628 + ], + "spans": [ + { + "bbox": [ + 70, + 617, + 261, + 628 + ], + "type": "text", + "content": "† Shengqi Dang and Yi He contributed equally to this work." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 71, + 628, + 195, + 637 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 628, + 195, + 637 + ], + "spans": [ + { + "bbox": [ + 71, + 628, + 195, + 637 + ], + "type": "text", + "content": "\\* Nan Cao is the corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 56, + 637, + 295, + 684 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 295, + 684 + ], + "type": "text", + "content": "*Shengqi Dang, Yi He, Long Ling, Ziqing Qian, and Nan Cao are with the Intelligent Big Data Visualization Lab, Tongji University. Shengqi Dang and Nan Cao are also with the Shanghai Innovation Institute. Email: {dangsq123, heyi_11}@tongji.edu.cn, lucyling0224@gmail.com, 2411920@tongji.edu.cn, nan.cao@gmail.com." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 56, + 685, + 294, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 685, + 294, + 703 + ], + "spans": [ + { + "bbox": [ + 56, + 685, + 294, + 703 + ], + "type": "text", + "content": "*Nanxuan Zhao is with Adobe Research. Email: nanxuanzhao@g-mail.com." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 70, + 703, + 285, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 703, + 285, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 703, + 285, + 712 + ], + "type": "text", + "content": "\\*Code: https://github.com/idvxlab/EmotiCrafter" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15218" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "into textual features, enabling the capture of emotions in alignment with intended input prompts. A novel loss function is also proposed to enhance emotion expression. The experimental results show that our method effectively generates images representing specific emotions with the desired content and outperforms existing techniques." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 136, + 179 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 187, + 296, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 270 + ], + "type": "text", + "content": "Emotions are fundamental to human experiences and play a critical role in shaping how people perceive and interact with the world. Research has shown that emotions affect memory [15, 19, 25, 45] and comprehension [16, 38, 40], which are crucial for effective communication. As a result, content creators increasingly recognize the importance of incorporating emotions to enhance audience engagement." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 271, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 449 + ], + "type": "text", + "content": "While research on visual emotion analysis is extensive [13, 49, 58], there is limited work on generating emotionally rich image content. Some early studies explored emotional content generation techniques within specific domains such as facial expressions [2, 46] or landscapes [24]. EmoGen [50] generates images based on a given emotion tag (e.g., happy or sad), which pioneered the domain-free emotional image content generation (EICG) task. However, it has two critical limitations: (1) the image is generated from an emotion tag instead of a text prompt, making the generated content difficult to control; (2) while the discrete emotion tags used in EmoGen are easy to understand, psychologists have not achieved a consensus on the emotion categories [60]. The limited scope of discrete emotion tags falls short of capturing nuanced emotions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 450, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 689 + ], + "type": "text", + "content": "To address the above issues, we propose the continuous emotional image content generation (C-EICG) task and present EmotiCrafter, the first C-EIGC model that generates emotional images using free-text prompts and continuous Valence-Arousal (V-A) values defined in the V-A model (a well-known psychological continuous emotion model) [33]. V-A model represents emotions in a two-dimensional Cartesian space (Figure 2), where Valence quantifies pleasantness (negative to positive) and Arousal measures intensity (calm to excited). We utilize a [-3, 3] range for Valence and Arousal [10, 21]. The continuous V-A space enables smooth transitions and nuanced emotional shifts beyond the capability of discrete labels (e.g., shifts from \"bored\" to \"tired\" of a character's state in a video). Specific emotional values in this space have been investigated in prior work [34], and such fine-grained modeling is particularly beneficial for human-computer interaction [2]. Leveraging this model, EmotiCrafter captures subtle affective variations via precise (V, A) positioning. The contributions of this paper are as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose a novel task, continuous emotional image content generation (C-EICG), and develop the first ded" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 213 + ], + "type": "image", + "image_path": "9624c89c88a2a32c6ec937c2b1a82f28ed6403ab72b50c5a17fd755a11f5f52c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 220, + 496, + 232 + ], + "lines": [ + { + "bbox": [ + 371, + 220, + 496, + 232 + ], + "spans": [ + { + "bbox": [ + 371, + 220, + 496, + 232 + ], + "type": "text", + "content": "Figure 2. Valence-Arousal model." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 321, + 251, + 553, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 251, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 321, + 251, + 553, + 335 + ], + "type": "text", + "content": "icated model for this purpose. Our model introduces an emotion-embedding network that integrates continuous Valence-Arousal (V-A) values with text prompts. These fused features are then injected into Stable Diffusion XL [28] using cross-attention mechanisms, enabling precise control over both content and emotional expressions in the generated images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 336, + 553, + 491 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 336, + 553, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 336, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 336, + 553, + 430 + ], + "type": "text", + "content": "- We propose a novel loss function that enhances the emotional resonance of generated images. By amplifying the difference between neutral and emotional text features, our approach enables the model to capture more distinct emotional variations. Additionally, the loss function incorporates the V-A distribution to address data imbalance, further refining the model's ability to generate images with rich, accurate emotional expressions." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 553, + 491 + ], + "type": "text", + "content": "- We constructed an emotional prompts dataset to train the emotion-embedding network, where each sample consists of a neutral prompt and an emotional prompt that share the same core meaning but express a specific emotion corresponding to a given pair of V-A values." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "spans": [ + { + "bbox": [ + 314, + 502, + 399, + 514 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 553, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 558 + ], + "type": "text", + "content": "In this section, we present a review of the related work, specifically focusing on visual emotion analysis, image emotion transfer, and conditional image generation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 564, + 451, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 451, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 451, + 577 + ], + "type": "text", + "content": "2.1. Visual Emotion Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 677 + ], + "type": "text", + "content": "Visual emotion analysis refers to the computational recognition and interpretation of human emotion [57] in visual media, such as images or videos. It has been a prominent research area, with most efforts focusing on the classification of discrete emotions [4, 42, 47, 48, 52]. However, discrete emotion categories limit the ability to capture the nuanced emotions [7], leading to an increased focus on continuous emotion analysis in images [59]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Much of the continuous emotion analysis remains centered on facial expression analysis [9, 11, 22, 37, 54]. While effective, facial analysis can overlook crucial con" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15219" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "textual information that influences emotional interpretation. This drives studies to focus on emotions in objects or individuals within their environment, rather than just faces[12, 13, 20, 21]. For example, Kosti et al. [12] combined person-specific characteristics with the context of the scene to predict continuous emotional dimensions such as arousal, valence, and dominance. Kragel et al. [13] proposed EmoNet to extract visual features such as facial expressions, body posture, and scene elements from images to predict valence and arousal values. Recently, Mertens et al.[21] tested multiple backbones (such as ResNet, CLIP, DINO, etc.) for prediction of valence and arousal, demonstrating strong performance across various architectures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 294, + 300 + ], + "type": "text", + "content": "Previous studies have achieved high accuracy in predicting continuous emotions in images, demonstrating a correlation between continuous emotions and visual elements. Building on this insight, our work moves beyond prediction to explore how continuous emotions can be actively embedded within generated content." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 309, + 194, + 321 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 194, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 194, + 321 + ], + "type": "text", + "content": "2.2. Image Emotion Transfer" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 295, + 542 + ], + "type": "text", + "content": "Image Emotion Transfer (IET) focuses on editing the content of the images to evoke different emotions [17, 27, 43, 61]. For instance, Peng et al. [27] achieved emotion transfer by adjusting color tones and texture-related features. Zhu et al. [61] introduced a method that separates high-level emotion-relevant features (e.g., object shapes and scene layout) from low-level emotion-relevant features (e.g., brightness). By applying GANs, they transferred emotions between images while preserving their original structure. Building on these methods, Weng et al. [43] proposed the Affective Image Filter, which uses a multi-modal transformer to process both text and image inputs. With the emergence of text-to-image models, IET has expanded into new applications based on instructive commands. For example, EmoEdit [51] used GPT-4V to build emotion factor trees that map abstract emotions to specific visual elements and employed the InstructPix2Pix model to apply emotion-driven content and color adjustments to images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 543, + 295, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 603 + ], + "type": "text", + "content": "Current methods primarily extract specific features, focusing on certain visual elements or emotional cues, which can limit the depth of emotional expression. In contrast, our method broadly learns a variety of emotion-influencing features, and it could accept natural language prompts as input." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 611, + 222, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 222, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 222, + 624 + ], + "type": "text", + "content": "2.3. Conditional Image Generation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Conditional image generation aims to create images that align with specific input conditions, such as text [3, 28, 30, 31], reference images [53], subjects [6, 32], and depth maps [23, 55]. To enhance quality, researchers have developed specialized approaches, such as Diffusion Transformers (DiT)[26], which use transformer-based diffusion for denoising, and Visual Autoregressive Modeling (VAR)[36]," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "which encodes images into discrete tokens for autoregressive prediction across scales. However, these state-of-the-art methods rely on discrete labels, limiting flexibility and control. In contrast, our approach offers greater generalization by leveraging Stable Diffusion XL [28] to generate emotional images from free-text prompts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 239 + ], + "type": "text", + "content": "Despite these progresses, incorporating emotion as a condition for image generation remains underexplored. EmoGen [50] pioneered the generation of emotional image content (EIGC) by mapping emotional features to semantic features to generate emotional images. However, EmoGen struggles to understand natural language, limiting its capacity to effectively control specific content. Its reliance on discrete emotions also limits its practical applicability." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 240, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 553, + 300 + ], + "type": "text", + "content": "Our method bridges this gap by embedding continuous emotion into textual features, enabling image generation models to use continuous emotion for emotion control. Unlike label-based methods, our method supports free-text prompts for flexible content control." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 311, + 370, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 311, + 370, + 323 + ], + "spans": [ + { + "bbox": [ + 314, + 311, + 370, + 323 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "content": "In this section, we introduce the technique details of the proposed EmotiCrafter." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 363, + 382, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 382, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 382, + 374 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": "Our method generates emotional images " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "I_{emo}" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": " from two inputs (Figure 3): a free-text prompt describing the desired content, and a pair of V-A values " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "(v, a)" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": " specifying the emotion. First the prompt encoder " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": " converts the text prompt into feature " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": ", which is then processed by an emotion-embedding network " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": " to produce emotional prompt feature " + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\hat{f}_{e}" + }, + { + "bbox": [ + 313, + 380, + 554, + 464 + ], + "type": "text", + "content": " that integrate the V-A values:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 472, + 553, + 488 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 472, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 388, + 472, + 553, + 488 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {e} = \\mathcal {M} \\left(f _ {n} | (v, a)\\right) \\tag {1}", + "image_path": "5ad16f62a236661bd3d96aa6c28a96fbe3333fc8267b2fbc6a901da06e705610.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "text", + "content": "Next, this feature is injected into Stable Diffusion XL " + }, + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "text", + "content": " via its cross-attention mechanism to generate the emotional images: " + }, + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "inline_equation", + "content": "I_{emo} = \\mathcal{G}(\\hat{f}_e)" + }, + { + "bbox": [ + 313, + 496, + 554, + 616 + ], + "type": "text", + "content": ". To enhance emotional expressiveness, we introduce a loss function (Fig.3(b.2)) that leverages the V-A distribution and emphasizes the differences between neutral and emotional prompt features. This ensures that the generated images accurately convey both the intended emotions and content. Additionally, we construct a dataset (Fig.3(a)) that pairs neutral and emotional prompts with the corresponding V-A values." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 479, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 479, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 479, + 636 + ], + "type": "text", + "content": "3.2. Emotion-Embedding Network" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": "The emotion-embedding network " + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 642, + 554, + 714 + ], + "type": "text", + "content": " generates the emotional prompt feature by integrating a pair of V-A values with a neutral prompt feature (Fig.3(b.1)). First, a V-A encoder converts the V-A values into feature vectors. Then, an emotion injection transformer—modified from GPT-2[29]—fuses these vectors with the neutral prompt" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15220" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 71, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 71, + 553, + 220 + ], + "type": "image", + "image_path": "b6efb014286d027a882699f6303429d453f12f20c4e9240f0d8501500385d5dd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "lines": [ + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "spans": [ + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "text", + "content": "Figure 3. Overview of our method. Specifically, we take the following steps. (a) We collect an image dataset annotated with V-A values, neutral prompts, and emotional prompts. These prompts are then encoded into features by prompt encoder " + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "text", + "content": ". (b) Next, we design (b.1) an emotion-embedding network " + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "text", + "content": " to embed V/A values into textual features based on the transformer architecture, and (b.2) a specialized loss function to enhance the emotional resonance of generated images. The output of the mapping network serves as the condition for the image generation model " + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 54, + 228, + 555, + 284 + ], + "type": "text", + "content": " to generate emotional images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 289, + 293, + 365 + ], + "blocks": [ + { + "bbox": [ + 59, + 289, + 293, + 365 + ], + "lines": [ + { + "bbox": [ + 59, + 289, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 289, + 293, + 365 + ], + "type": "image", + "image_path": "86bc912c030cf236269f88e37e87c6e9258c5e203b081b03668ec9abea346b32.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "lines": [ + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "text", + "content": "Figure 4. Structure of Emotion Injection Block. It accepts hidden state " + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "inline_equation", + "content": "h_{i-1}" + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "text", + "content": " as input and produces " + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "text", + "content": " as output. The V-feature " + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "text", + "content": " and A-feature " + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "inline_equation", + "content": "e_a" + }, + { + "bbox": [ + 55, + 371, + 295, + 415 + ], + "type": "text", + "content": " represent the emotion features, which are injected through the cross-attention module." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 419, + 295, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 442 + ], + "type": "text", + "content": "feature, preserving the original textual context while infusing emotional content." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "text", + "content": "V-A Encoder. The V-A Encoder converts a pair of V-A values into feature vectors using two separate multilayer perceptrons (MLPs). One MLP processes the Valence value to produce V-feature " + }, + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "text", + "content": ", and the other processes the Arousal value to generate A-feature " + }, + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "inline_equation", + "content": "e_a" + }, + { + "bbox": [ + 55, + 444, + 295, + 525 + ], + "type": "text", + "content": ". These features are then fed into the emotion injection transformer network for further emotion infusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 527, + 295, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 586 + ], + "type": "text", + "content": "Emotion Injection Transformer (EIT). The Emotion Injection Transformer (EIT) leverages a modified GPT-2 architecture to seamlessly integrate V/A-features into textual features. Its process consists of three stages: input projection, emotion injection, and output projection." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "content": "First, we project the input neutral prompt feature " + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 55, + 586, + 295, + 610 + ], + "type": "text", + "content": " into the transformer's feature space:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 134, + 616, + 295, + 629 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 616, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 134, + 616, + 295, + 629 + ], + "type": "interline_equation", + "content": "h _ {0} = \\mathrm {P} _ {\\text {i n}} \\left(f _ {n}\\right) + \\mathrm {P E}, \\tag {2}", + "image_path": "a4781f23bb6ff5e4036d347e2aacdac7bec73194b8a8063ebb21cd754741ea32.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "inline_equation", + "content": "h_0" + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "content": " represents the initial hidden state; " + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "inline_equation", + "content": "\\mathrm{P_{in}(\\cdot)}" + }, + { + "bbox": [ + 55, + 635, + 295, + 658 + ], + "type": "text", + "content": " is a linear projection layer, and PE denotes positional embedding [39]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": "In the next, we inject emotion into " + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 55, + 659, + 295, + 696 + ], + "type": "text", + "content": " via 12 sequential Emotion Injection Blocks (EIBs) corresponding to 12 transformer blocks. Each block outputs a hidden state:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 88, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 88, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "h _ {i} = \\operatorname {E I B} \\left(h _ {i - 1}, e _ {v}, e _ {a}\\right), \\quad i \\in \\{1, \\dots , 1 2 \\} \\tag {3}", + "image_path": "04457892e9c392a57c25a2010e2d981418d95c7b321e896cdf42182742776e14.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "inline_equation", + "content": "h_i" + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "content": " is the output of the " + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "content": "-th EIB(" + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 290, + 555, + 326 + ], + "type": "text", + "content": "). As shown in Figure 4, each EIB enhances the transformer block through a cross-attention mechanism:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 333, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 333, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 350, + 333, + 553, + 346 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {\\prime} = \\text {s e l f - a t t n} (\\mathrm {L N} (h _ {i - 1})) + h _ {i - 1} \\tag {4}", + "image_path": "3b8c9efb8bd1ac3e144ee4eb003b32147140d52ac11a87e7cc57b255a8d08454.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 350, + 349, + 553, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 349, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 350, + 349, + 553, + 365 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {(v)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {\\prime}\\right), e _ {v}\\right) + h _ {i} ^ {\\prime} \\tag {5}", + "image_path": "ddd21babf2d0a177d11a587a533db48696610b74b87661a7406f82b50829f586.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 350, + 366, + 553, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 366, + 553, + 383 + ], + "spans": [ + { + "bbox": [ + 350, + 366, + 553, + 383 + ], + "type": "interline_equation", + "content": "h _ {i} ^ {(v, a)} = \\operatorname {c r o s s - a t t n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v)}\\right), e _ {a}\\right) + h _ {i} ^ {(v)} \\tag {6}", + "image_path": "573d693ac431aa44dd5f1cf2e75f48207ee9bf2bed130143cdb8270067f2df0d.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 351, + 384, + 553, + 401 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 384, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 351, + 384, + 553, + 401 + ], + "type": "interline_equation", + "content": "h _ {i} = \\operatorname {f n n} \\left(\\mathrm {L N} \\left(h _ {i} ^ {(v, a)}\\right)\\right) + h _ {i} ^ {(v, a)} \\tag {7}", + "image_path": "2f614d552511fd3aa797d1915165ff4cc3f9fbc7fb3070c67f0c3475feaa9d08.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "h^{\\prime}, h^{(v)}, h^{(v,a)}" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": " are the intermediate hidden variables; " + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{LN}(\\cdot)" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": " is the LayerNorm; self-attn(" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": ") denotes self-attention, employed to capture context dependencies; cross-attn(" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": ") is cross-attention for injecting " + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "e_v" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "e_a" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{fnn}(\\cdot)" + }, + { + "bbox": [ + 313, + 407, + 554, + 503 + ], + "type": "text", + "content": " is a feedforward network that adapts the complexity of the emotional embedding process. We also remove the causal mask typically used for next-token prediction from the original transformer model to fit our task." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "content": "Finally, the output of the last (i.e., the 12th) injection block " + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "inline_equation", + "content": "h_{12}" + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "content": " is projected back to SDXL's prompt feature space via " + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{\\mathrm{out}}" + }, + { + "bbox": [ + 313, + 504, + 554, + 576 + ], + "type": "text", + "content": " (a linear layer and a LayerNorm layer) to predict the residual between emotional and neutral prompt features, which represents a semantic shift between emotional and neutral prompts:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 392, + 582, + 553, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 582, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 392, + 582, + 553, + 597 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {r} = \\mathrm {P} _ {\\text {o u t}} \\left(\\mathrm {L N} \\left(h _ {1 2}\\right)\\right) \\tag {8}", + "image_path": "b536b669ad9390d3417b2bdaaa80007208d9442fb50f08642dfad96dd5ad502d.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 602, + 555, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 627 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 627 + ], + "type": "text", + "content": "The final emotional prompt feature is obtained by adding this residual to the original neutral prompt feature:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 405, + 632, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 632, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 405, + 632, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {e} = \\hat {f} _ {r} + f _ {n} \\tag {9}", + "image_path": "54dd846434791e737ee8c34def4a4fb243cece0585979216f3af820a3a6dbb10.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "The above emotion embedding network is trained by minimizing the averaged expectation of the difference between the predicted emotional prompt feature " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_e = \\mathcal{M}(f_n|(v,a))" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " and the scaled target emotional prompt feature " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "f_e^t" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ", using the loss function described in Equation 10." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15221" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 110, + 81, + 295, + 108 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 81, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 110, + 81, + 295, + 108 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\frac {1}{n} \\mathbb {E} \\left(\\frac {1}{d (v , a)} \\| \\hat {f} _ {e} - f _ {e} ^ {t} \\| ^ {2}\\right) \\tag {10}", + "image_path": "9e24504d9a15067a6bb45d32c4c562a2af68b0e153ecdc239a6edbb03c01c029.jpg" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "text", + "content": " is the number of feature elements; " + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "inline_equation", + "content": "\\mathbb{E}(\\cdot)" + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "text", + "content": " denotes the expectation; " + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "inline_equation", + "content": "d(v, a)" + }, + { + "bbox": [ + 55, + 111, + 295, + 147 + ], + "type": "text", + "content": " is a density function that describes the distribution of V-A values in the training sample." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 147, + 295, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 195 + ], + "type": "text", + "content": "To effectively embed emotions and address the challenges posed by the uneven distribution of V-A values in the dataset, this loss function incorporates two key strategies to improve the model's performance:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 295, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 231 + ], + "type": "text", + "content": "Scaled Residual Learning. To better capture pronounced emotional changes in generated images, we enlarge the target residuals:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 239, + 295, + 266 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 239, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 123, + 239, + 295, + 266 + ], + "type": "interline_equation", + "content": "f _ {e} ^ {t} = f _ {n} + \\alpha \\underbrace {\\left(f _ {e} - f _ {n}\\right)} _ {\\text {r e s i d u a l f e a t u r e}}, \\tag {11}", + "image_path": "17569d6bad649495a5c0aab1bdb59aa256324a0af50d32d0f414df6e1d28e470.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "inline_equation", + "content": "f_{e}" + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "content": " is the emotional prompt feature, " + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "inline_equation", + "content": "f_{n}" + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "content": " is the neutral prompt feature, and " + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "content": " is a scale factor, we set its value to 1.5 based on the ablation study." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 382 + ], + "type": "text", + "content": "V-A Density Weighting. To mitigate the effects of the imbalanced distribution of the training samples, we weigh the loss inversely proportional to the density of training samples in the V-A space. The density is estimated using Kernel Density Estimation (KDE) [5] with a Gaussian kernel, denoted as " + }, + { + "bbox": [ + 55, + 310, + 296, + 382 + ], + "type": "inline_equation", + "content": "d(v,a)" + }, + { + "bbox": [ + 55, + 310, + 296, + 382 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 94, + 392, + 295, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 392, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 94, + 392, + 295, + 422 + ], + "type": "interline_equation", + "content": "d (v, a) = \\frac {1}{n} \\sum_ {i = 1} ^ {n} K _ {H} ((v, a) - (v _ {i}, a _ {i})), \\tag {12}", + "image_path": "86413b3ce70c698ede806773d6f1bf1feca6500a39ce801195790c59ab4eb6a2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "K_{H}" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " is a 2D Gaussian kernel with bandwidth " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " is the number of training samples; " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "(v_{i},a_{i})" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " are the V-A values of the " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": "-th training sample. The bandwidth " + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 426, + 296, + 486 + ], + "type": "text", + "content": " is selected using Silverman's rule of thumb to provide optimal smoothing of the density estimation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 493, + 178, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 178, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 178, + 506 + ], + "type": "text", + "content": "3.3. Dataset and Training" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 677 + ], + "type": "text", + "content": "To train the emotion-embedding network, we constructed a dataset of paired neutral and emotional prompts with corresponding Valence-Arousal (V-A) values (Figure 3(a)). These pairs were automatically generated using GPT-4 based on 39,843 images with human-annotated V-A values from publicly available datasets, including OASIS [14], EMOTIC [12], and FindingEmo [21]. Specifically, GPT-4 was used to generate neutral prompts with objective image descriptions and emotional prompts that emphasize affective attributes such as color, lighting, and texture, which influence emotional perception. To ensure data reliability, all LLM-generated prompts underwent crowd-sourced verification, with disagreements resolved through a voting mechanism among annotators." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The proposed emotion embedding network is trained on two NVIDIA A800 GPUs using the aforementioned dataset. We employ the AdamW [18] optimizer with a weight decay" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "of 1e-5 and a learning rate of 1e-3. The training process spans 200 epochs with a batch size of 768, completing in approximately 7 to 8 hours." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 386, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 386, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 386, + 133 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 140, + 427, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 427, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 427, + 152 + ], + "type": "text", + "content": "4.1. Generation Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 159, + 555, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 555, + 314 + ], + "type": "text", + "content": "Figure 1 shows the proposed technique's ability to achieve continuous and effective control over both emotion and content during image generation. Meanwhile, Figure 5 highlights four key capabilities: (a) emotion-content decoupling, where V-A values override emotional cues in the prompt, allowing typically positive concepts to be rendered with negative emotions; (b) compatibility with discrete emotion categories; (c) content-independent generation, where images generated from empty prompts and specified V-A values maintain emotional consistency without semantic constraints; and (d) fine-grained emotional control, demonstrated through V-A increments of 0.2, showcasing the model's sensitivity to subtle emotional variations." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 399, + 335 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 341, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 365 + ], + "type": "text", + "content": "To estimate the effectiveness of the proposed method, we built four baselines based on existing techniques." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 437 + ], + "type": "text", + "content": "Baselines. We established baselines for comparison using two strategies: (1) directly injecting emotion features into the image generation modules of SDXL, such as UNet, and (2) modifying the input text prompt using emotion features to influence the generated image's content, similar to the proposed method." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": "As a result, four different baselines were built: (1) Cross Attention: inject the emotion features " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "(e_v, e_a)" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " into the UNet in SDXL via its cross-attention mechanism based on IP-Adapter [53]; (2) Time Embedding: directly add emotion features " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "(e_v, e_a)" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " to the time embedding of the UNet in SDXL. (3) Textual Inversion: use the text inversion technique [6] to embed emotion features " + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "inline_equation", + "content": "(e_v, e_a)" + }, + { + "bbox": [ + 313, + 437, + 554, + 569 + ], + "type": "text", + "content": " into prompt templates with predefined emotion placeholders. (4) GPT-4+SDXL (GPT-SD): use GPT-4 [1] to rewrite the input text according to (v,a) values to generate an emotional SDXL prompt for image generation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Qualitative Comparison. We evaluate the generated images based on three criteria: (1) the effectiveness of emotion embedding, (2) image-prompt similarity, and (3) the continuity of emotional variations as V-A values change. Figure 6 demonstrates that the baseline methods—Cross Attention, Time Embedding, and Textual Inversion—tend to produce nearly identical outputs regardless of emotional variation. This is because the loss terms of these baselines primarily align low-level image features (e.g., SDXL's latent space), which are highly correlated with prompt content but struggle to capture subtle emotional cues when both prompts and V-A values are provided. Notably, Tex-" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15222" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 70, + 129, + 123 + ], + "blocks": [ + { + "bbox": [ + 59, + 92, + 73, + 102 + ], + "lines": [ + { + "bbox": [ + 59, + 92, + 73, + 102 + ], + "spans": [ + { + "bbox": [ + 59, + 92, + 73, + 102 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 70, + 129, + 123 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 129, + 123 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 129, + 123 + ], + "type": "image", + "image_path": "bba03a1446d97cabe0e071e1851f7434ab9605a8260b4e84c5e83090ecf6fd34.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 131, + 92, + 145, + 102 + ], + "lines": [ + { + "bbox": [ + 131, + 92, + 145, + 102 + ], + "spans": [ + { + "bbox": [ + 131, + 92, + 145, + 102 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 148, + 70, + 291, + 121 + ], + "blocks": [ + { + "bbox": [ + 148, + 70, + 291, + 121 + ], + "lines": [ + { + "bbox": [ + 148, + 70, + 291, + 121 + ], + "spans": [ + { + "bbox": [ + 148, + 70, + 291, + 121 + ], + "type": "image", + "image_path": "8c02341b7eb8fe02f47bdb2914dcda120616909acc139cea56b3820fc8bced53.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 350, + 121 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 350, + 121 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 350, + 121 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 350, + 121 + ], + "type": "image", + "image_path": "c3ea3c2ca75d16339a200ea498be0e290ed9f8f7898b6df493694aeed647d402.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 351, + 71, + 451, + 121 + ], + "blocks": [ + { + "bbox": [ + 351, + 71, + 451, + 121 + ], + "lines": [ + { + "bbox": [ + 351, + 71, + 451, + 121 + ], + "spans": [ + { + "bbox": [ + 351, + 71, + 451, + 121 + ], + "type": "image", + "image_path": "7246a3cd59b6b304cac1ac54cf48949b2ece02cb95fb761f3d677a85d52e2f62.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 451, + 71, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 451, + 71, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 451, + 71, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 451, + 71, + 553, + 121 + ], + "type": "image", + "image_path": "04b476acc97c862b300ac66bd4d4db2f86f5a04a77cef2867e7a552d59f2c69e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 77, + 125, + 281, + 177 + ], + "blocks": [ + { + "bbox": [ + 292, + 146, + 298, + 158 + ], + "lines": [ + { + "bbox": [ + 292, + 146, + 298, + 158 + ], + "spans": [ + { + "bbox": [ + 292, + 146, + 298, + 158 + ], + "type": "text", + "content": "1)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 146, + 73, + 158 + ], + "lines": [ + { + "bbox": [ + 59, + 146, + 73, + 158 + ], + "spans": [ + { + "bbox": [ + 59, + 146, + 73, + 158 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 125, + 281, + 177 + ], + "lines": [ + { + "bbox": [ + 77, + 125, + 281, + 177 + ], + "spans": [ + { + "bbox": [ + 77, + 125, + 281, + 177 + ], + "type": "image", + "image_path": "4a7e7a5c16f462a4c1676792605a3a2c0e990deedbc06e8892190c2ffaecfe3f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 312, + 125, + 400, + 176 + ], + "blocks": [ + { + "bbox": [ + 299, + 125, + 311, + 135 + ], + "lines": [ + { + "bbox": [ + 299, + 125, + 311, + 135 + ], + "spans": [ + { + "bbox": [ + 299, + 125, + 311, + 135 + ], + "type": "inline_equation", + "content": "\\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : \\left\\{ {2 : {x}^{\\prime }\\left( x\\right) = {y}^{\\prime }\\left( x\\right) }\\right\\} : y = x" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 312, + 125, + 400, + 176 + ], + "lines": [ + { + "bbox": [ + 312, + 125, + 400, + 176 + ], + "spans": [ + { + "bbox": [ + 312, + 125, + 400, + 176 + ], + "type": "image", + "image_path": "741da1258ba440b8884678442a21ec33513b45280af59afc71bb6eecba5e314b.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 400, + 125, + 451, + 175 + ], + "blocks": [ + { + "bbox": [ + 400, + 125, + 451, + 175 + ], + "lines": [ + { + "bbox": [ + 400, + 125, + 451, + 175 + ], + "spans": [ + { + "bbox": [ + 400, + 125, + 451, + 175 + ], + "type": "image", + "image_path": "26f2a0cb69f8bb32498994df30781659074ec362a6b4b19631d06fa9a3e3a91f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 451, + 125, + 503, + 175 + ], + "blocks": [ + { + "bbox": [ + 451, + 125, + 503, + 175 + ], + "lines": [ + { + "bbox": [ + 451, + 125, + 503, + 175 + ], + "spans": [ + { + "bbox": [ + 451, + 125, + 503, + 175 + ], + "type": "image", + "image_path": "75770b5fdbd7dade7d4f1cbd2317355a6922ed0bb85ff199c0feb74a9546c741.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 503, + 125, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 503, + 125, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 503, + 125, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 503, + 125, + 553, + 175 + ], + "type": "image", + "image_path": "853b3c65492d661868b046a816bdb95e1f56d2909eb2e5e6ac93f8ac86b60afe.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 72, + 220, + 538, + 439 + ], + "blocks": [ + { + "bbox": [ + 55, + 183, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 183, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 183, + 555, + 217 + ], + "type": "text", + "content": "Figure 5. Results under multiple inputs. (a) Overriding semantic content ('a child in the amusement park') with sad V-A (-2,-2); (b) Discrete emotion mapping in V-A space as emotion input; (c) Empty-prompt generation with pure emotion condition; (d) Fine-grained control of V-A variations with a granularity of 0.2." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 72, + 220, + 538, + 439 + ], + "lines": [ + { + "bbox": [ + 72, + 220, + 538, + 439 + ], + "spans": [ + { + "bbox": [ + 72, + 220, + 538, + 439 + ], + "type": "image", + "image_path": "9f17bff65dc980cc74376cc6e18e6e732917c86e995f3106e793171699f33ebc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 555, + 481 + ], + "type": "text", + "content": "Figure 6. Qualitative comparisons with baselines. These images are generated at varying V-A values, specifically -1.5, 0, and 1.5. Only our approach and the GPT-4+SDXL successfully generate images that clearly reflect emotional variations. Notably, our results show enhanced continuity, indicating superior controllability over continuous V-A values compared to the GPT-4+SDXL." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 491, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 610 + ], + "type": "text", + "content": "tual Inversion often produces images with a persistent purple tint. Furthermore, all methods faithfully generate content aligned with the given prompts. However, for image continuity, we focus on comparing our method with GPT-4+SDXL, as they are the only two capable of generating distinct emotional variations. Figure 7 provides a more detailed comparison, showing that our method maintains smooth emotional transitions even under extreme V-A conditions, whereas GPT-4+SDXL introduces noticeable discontinuities (e.g., in the " + }, + { + "bbox": [ + 55, + 491, + 295, + 610 + ], + "type": "inline_equation", + "content": "\\mathrm{V} = 3" + }, + { + "bbox": [ + 55, + 491, + 295, + 610 + ], + "type": "text", + "content": " column)." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Quantitative Comparison. We compare our method against several baselines using the following metrics: (1) " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\underline{\\mathrm{V} / \\mathrm{A} - \\mathrm{Error}}" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " evaluates the absolute error between the predicted V/A of the generated images and the input V/A. (2) CLIPScore [8] assesses the similarity between the input text and the generated images. (3) CLIP-IQA [41] leverages a pre-trained CLIP model to evaluate image quality without requiring reference images. (4) LPIPS-Continuous utilizes" + } + ] + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 315, + 489, + 553, + 542 + ], + "blocks": [ + { + "bbox": [ + 315, + 489, + 553, + 542 + ], + "lines": [ + { + "bbox": [ + 315, + 489, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 489, + 553, + 542 + ], + "type": "table", + "html": "
A-Error ↓V-Error ↓CLIPScore ↑CLIP-IQA ↑
Cross Attention1.923±1.1532.080±1.43826.266±2.3810.949±0.046
Time Embedding1.941±1.1682.031±1.34826.566±2.1250.786±0.164
Textual Inversion1.958±1.1881.923±1.17022.346±3.5940.370±0.111
GPT-4+SDXL1.860±1.0901.517±1.06025.907±1.9490.906±0.066
Ours1.828±1.0851.510±1.07423.067±2.6550.881±0.099
", + "image_path": "b2ae7573125fb0afe061cdf77f2d4a0072ccbe10815c6d272f441480cf24d30e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "text", + "content": "Table 1. Comparison on emotion accuracy, prompt fidelity, and image quality across different baselines, evaluated on 3,300 images per method (132 prompts " + }, + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "text", + "content": " 5 V values " + }, + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 543, + 555, + 620 + ], + "type": "text", + "content": " 5 A values). Our method achieves the highest performance in emotion accuracy while maintaining comparable results in prompt fidelity and image quality. The slight decrease in prompt fidelity is expected, as modifying emotional content affects semantic alignment." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 639, + 553, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 553, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 553, + 663 + ], + "type": "text", + "content": "the Learned Perceptual Image Patch Similarity [56] to measure the continuity of the change in image as V/A changes." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "As shown in Table 1, our method achieves the lowest (best) " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\underline{V / A - Error}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " on average. While Cross Attention and Time Embedding achieve the highest " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\underline{CLIP - IQA}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\underline{CLIPScore}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": ", respectively, these methods fail to generate" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15223" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 294, + 531 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 294, + 531 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 294, + 531 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 294, + 531 + ], + "type": "image", + "image_path": "16012fcf819496df2d55baca64f95cd2d704ca4f067f84b38ee5f5ed4b266430.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 540, + 295, + 597 + ], + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 597 + ], + "type": "text", + "content": "Figure 7. A more comprehensive comparison with the GPT-4+SDXL for the prompt \"A group of people is skiing in a snow hill.\" Our approach maintains continuity even under extreme V/A conditions. Conversely, GPT-4+SDXL displays noticeable discontinuities (e.g., in the " + }, + { + "bbox": [ + 55, + 540, + 295, + 597 + ], + "type": "inline_equation", + "content": "\\mathrm{V} = 3" + }, + { + "bbox": [ + 55, + 540, + 295, + 597 + ], + "type": "text", + "content": " column)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 615, + 228, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 228, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 228, + 628 + ], + "type": "text", + "content": "emotionally expressive images (Figure 6)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Our method exhibits a slight decrease in CLIPScore compared to the baselines, which we attribute to the inherent trade-off between emotional modulation and strict semantic alignment. However, the high CLIP-IQA score indicates that our method produces high-quality images. Additionally, as shown in Table 2, our approach demonstrates superior continuity compared to GPT-4 + SDXL." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 328, + 69, + 539, + 95 + ], + "blocks": [ + { + "bbox": [ + 328, + 69, + 539, + 95 + ], + "lines": [ + { + "bbox": [ + 328, + 69, + 539, + 95 + ], + "spans": [ + { + "bbox": [ + 328, + 69, + 539, + 95 + ], + "type": "table", + "html": "
OursGPT-4+SDXL
LPIPS-Continuous↓0.220±0.0640.361±0.059
", + "image_path": "be69039cc467c64114134922759098bbc41242868f6712ebd8c09286548a03d4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 112, + 563, + 196 + ], + "blocks": [ + { + "bbox": [ + 315, + 98, + 555, + 109 + ], + "lines": [ + { + "bbox": [ + 315, + 98, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 315, + 98, + 555, + 109 + ], + "type": "text", + "content": "Table 2. Continuity comparison between our method and baseline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 112, + 563, + 196 + ], + "lines": [ + { + "bbox": [ + 315, + 112, + 563, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 112, + 563, + 196 + ], + "type": "table", + "html": "
OursGPT-4+SDXL
Study IA-Ranking Consistency ↑0.759±0.2730.165±0.379
V-Ranking Consistency ↑0.887±0.2450.584±0.259
A-Error ↓1.327±1.1202.029±1.446
V-Error ↓0.692±0.6821.229±1.026
Study IIEmotion Consistency ↑4.215±0.7153.525±1.065
Emotion Smoothness ↑4.240±0.8283.195±1.163
", + "image_path": "45564c4c38295c6a329a9f7fb1eca106ca0c88452c74fb41afc51ae21b2be65d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 345, + 201, + 523, + 211 + ], + "lines": [ + { + "bbox": [ + 345, + 201, + 523, + 211 + ], + "spans": [ + { + "bbox": [ + 345, + 201, + 523, + 211 + ], + "type": "text", + "content": "Table 3. Our method outperformed the baseline." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 315, + 219, + 553, + 314 + ], + "blocks": [ + { + "bbox": [ + 315, + 219, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 315, + 219, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 315, + 219, + 553, + 314 + ], + "type": "image", + "image_path": "652a060b2d3ed81b304f1994db701d6714a9b96b18dbc10ef2e6c96670ef485e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "type": "text", + "content": "Figure 8. User Study Results (" + }, + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "type": "inline_equation", + "content": "^{*}\\mathsf{p} < 0.05" + }, + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "type": "inline_equation", + "content": "^{**}\\mathsf{p} < 0.01" + }, + { + "bbox": [ + 314, + 316, + 555, + 328 + ], + "type": "inline_equation", + "content": "^{**}p < 0.001" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 342, + 389, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 389, + 356 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 389, + 356 + ], + "type": "text", + "content": "4.3. User Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "content": "We conducted two user studies with 20 college students to evaluate the effectiveness of our method by comparing it to " + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "inline_equation", + "content": "GPT-4 + SDXL" + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "content": " (the baseline)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "content": "Study I. In the first experiment, we evaluated whether the generated images' emotions align with human perception. Two image collections were prepared—one for Arousal (A) and one for Valence (V)—each comprising 20 sets (10 from our method and 10 from the baseline) of 5 randomized images with varying A or V values. Participants reordered the images by perceived intensity and estimated each image's V and A values. We then computed Kendall's " + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "inline_equation", + "content": "\\tau_{b}" + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "inline_equation", + "content": "\\tau_{b} = 1" + }, + { + "bbox": [ + 313, + 399, + 556, + 532 + ], + "type": "text", + "content": " indicating perfect alignment) to assess ordering accuracy and calculated the absolute error between the estimated and ground truth values." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 628 + ], + "type": "text", + "content": "Study II. In the second experiment, we assessed whether the generated images could effectively reflect continuous emotional changes (i.e., V-A values). We generated 20 image sets (10 per method), each containing 25 images with V-A values varying gradually from -3 to +3 (Figure 7). Participants rated each set on a 5-point Likert scale regarding (1) the alignment between V-A changes and image content, and (2) the smoothness of the content transition." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Analysis & Results. We conducted a Shapiro-Wilk test [35] to assess normality and applied the Wilcoxon Signed Rank test [44] to evaluate statistical significance. Our results indicate that our method outperformed the baseline across all metrics (Table 3), with statistically significant improvements in V/A Ranking, V/A Error, Consistency, and Smoothness (Figure 8)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "15224" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 555, + 281 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 555, + 281 + ], + "type": "image", + "image_path": "49d7a374a28ece4a1b71699e296223fc2af1c98585540d696136733c83d8f2ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "text", + "content": "Figure 9. Abalation Study. Images are generated from the prompt \"An oil painting shows an astronaut.\" As " + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "text", + "content": " increases, image-prompt similarity decreases, while emotional variations increase. The usage of " + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "inline_equation", + "content": "d(a,v)" + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "text", + "content": " enhances the accuracy of emotional changes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 327, + 291, + 437 + ], + "blocks": [ + { + "bbox": [ + 58, + 327, + 291, + 437 + ], + "lines": [ + { + "bbox": [ + 58, + 327, + 291, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 291, + 437 + ], + "type": "image", + "image_path": "d3b75a7c34fb19af0a35185edcbb56db21e04b2aff75fff5dc97d181fe1b6946.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": "Figure 10. The effectiveness of the scaling factor " + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": ". The method with " + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\alpha = 1.5" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": ") surpasses the performance indicated by the regression line " + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "inline_equation", + "content": "(- - )" + }, + { + "bbox": [ + 55, + 441, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 482, + 294, + 519 + ], + "blocks": [ + { + "bbox": [ + 57, + 482, + 294, + 519 + ], + "lines": [ + { + "bbox": [ + 57, + 482, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 57, + 482, + 294, + 519 + ], + "type": "table", + "html": "
A-Error ↓V-Error ↓CLIPScore ↑
Ours1.828±1.0851.510±1.07423.067±2.655
w/o d(v, a)1.829±1.0831.546±1.08221.977±0.066
", + "image_path": "c17b1b21d0dd0ff49275e51e53ac12390fed95c1319811997436134c9b7ff770.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 521, + 243, + 531 + ], + "lines": [ + { + "bbox": [ + 107, + 521, + 243, + 531 + ], + "spans": [ + { + "bbox": [ + 107, + 521, + 243, + 531 + ], + "type": "text", + "content": "Table 4. The effectiveness of " + }, + { + "bbox": [ + 107, + 521, + 243, + 531 + ], + "type": "inline_equation", + "content": "d(v,a)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 548, + 149, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 548, + 149, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 548, + 149, + 562 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 567, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 590 + ], + "type": "text", + "content": "We performed ablation experiments to assess the contribution and effectiveness of the proposed loss function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "Effectiveness of the Scaling Factor " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": ". We evaluated how varying " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " affects CLIPScore and V/A-Error (Figure 10). As " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " increases, CLIPScore decreases (indicating reduced semantic alignment), while V/A-Error also decreases (indicating improved emotional accuracy). This trend is further illustrated by the examples in Figure 9. Based on these findings, we set " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha = 1.5" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " as an optimal tradeoff, though users can adjust " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " to suit their specific needs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of the Density Weighting " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "d(v,a)" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": ". We compare our full method with a variant that omits d(v,a)" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 325, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 555, + 361 + ], + "type": "text", + "content": "from the loss function. As shown in Table 4, including " + }, + { + "bbox": [ + 313, + 325, + 555, + 361 + ], + "type": "inline_equation", + "content": "\\mathrm{d}(\\mathrm{v},\\mathrm{a})" + }, + { + "bbox": [ + 313, + 325, + 555, + 361 + ], + "type": "text", + "content": " leads to improvements in both CLIPScore and V/A-Error. This positive effect is further illustrated in Figure 9." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 370, + 471, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 471, + 383 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 471, + 383 + ], + "type": "text", + "content": "5. Conclusion and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 555, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 510 + ], + "type": "text", + "content": "In this paper, we introduce continuous emotional image content generation (C-EICG) and present EmotiCrafter, a novel method that generates emotionally expressive images using continuous Valence-Arousal (V-A) values. Our emotion-embedding network integrates V-A values into textual features, and extensive experiments show that our approach reliably aligns images with both user prompts and specified emotions. We believe this work will advance affective computing and image generation, and we will release our code and data to foster further research." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "Although our method achieved promising results, it still has some limitations need to be addressed in the future. First, controlling image generation based on arousal remains more challenging than controlling based on valence. This is consistent with prior research in visual emotion analysis, which has found that arousal is harder to predict due to lower inter-annotator agreement [21]. Second, our approach frequently generates images featuring human activities even when such activities are not mentioned in the prompts. This likely stems from the limited representation of non-human scenes in our training data and could be mitigated by incorporating a more diverse range of non-human scenarios. Third, our method occasionally modifies users' input prompts to better align with the specified emotional prompts, resulting in a slight semantic shift that affects the generated images. We believe this issue could be addressed by adding a semantic preservation term to the loss function." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15225" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 128 + ], + "type": "text", + "content": "Nan Cao is the corresponding author. This work was supported by the National Key Research and Development Program of China (2023YFB3107100)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 163, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 195 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 198, + 296, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 296, + 231 + ], + "type": "text", + "content": "[2] Bita Azari and Angelica Lim. Emostyle: One-shot facial expression editing using continuous emotion parameters. In Proceedings of WACV, pages 6385-6394, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 232, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 232, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 62, + 232, + 294, + 277 + ], + "type": "text", + "content": "[3] James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, et al. Improving image generation with better captions. Computer Science. https://cdn.openaia.com/papers/dall-e-3.pdf, 2(3):8, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 279, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 279, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 279, + 295, + 322 + ], + "type": "text", + "content": "[4] Damian Borth, Rongrong Ji, Tao Chen, Thomas Breuel, and Shih-Fu Chang. Large-scale visual sentiment ontology and detectors using adjective noun pairs. In Proceedings of ACM MM, pages 223-232, 2013. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 357 + ], + "type": "text", + "content": "[5] Yen-Chi Chen. A tutorial on kernel density estimation and recent advances. Biostatistics & Epidemiology, 1(1):161-187, 2017. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 361, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 295, + 403 + ], + "type": "text", + "content": "[6] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, et al. An image is worth one word: Personalizing text-to-image generation using textual inversion. In Proceedings of ICLR, 2023. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 406, + 295, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 295, + 460 + ], + "type": "text", + "content": "[7] Hatrice Gunes, Björn Schuller, Maja Pantic, and Roddy Cowie. Emotion representation, analysis and synthesis in continuous space: A survey. In Proceedings of IEEE international conference on automatic face & gesture recognition, pages 827-834, 2011. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 295, + 518 + ], + "type": "text", + "content": "[8] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In Proceedings of the Conference on Empirical Methods in Natural Language Processing, pages 7514-7528, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 295, + 574 + ], + "type": "text", + "content": "[9] Stephen Khor Wen Hwoo, Alice Othmani, and Aznul Qalid Md. Sabri. Deep learning-based approach for continuous affect prediction from facial expression images in valence-arousal space. IEEE Access, 10:96053-96065, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 577, + 295, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 577, + 295, + 632 + ], + "spans": [ + { + "bbox": [ + 57, + 577, + 295, + 632 + ], + "type": "text", + "content": "[10] Ahmed Khota, Eric Cooper, Yu Yan, and Mate Kovacs. Modelling emotional valence and arousal of non-linguistic utterances for sound design support. In Proceedings of the International Conference on Kansei Engineering and Emotion Research, pages 507–516, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 677 + ], + "type": "text", + "content": "[11] Dimitrios Kollias. Abaw: Valence-arousal estimation, expression recognition, action unit detection & multi-task learning challenges. In Proceedings of CVPR, pages 2328-2336, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[12] Ronak Kosti, Jose M. Alvarez, Adria Recasens, and Agata Lapedriza. Emotion recognition in context. In Proceedings of CVPR, pages 1667-1675, 2017. 3, 5" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[13] Philip A Kragel, Marianne C Reddan, Kevin S LaBar, and Tor D Wager. Emotion schemas are embedded in the human visual system. Science advances, 5(7):eaaw4358, 2019. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 554, + 139 + ], + "type": "text", + "content": "[14] Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. Introducing the open affective standardized image set (oasis). Behavior research methods, 49:457-470, 2017. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 554, + 173 + ], + "type": "text", + "content": "[15] Angela Y Lee and Brian Sternthal. The effects of positive mood on memory. Journal of consumer research, 26(2):115-127, 1999. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 218 + ], + "type": "text", + "content": "[16] Jiansheng Li, Chuanlan Luo, Qi Zhang, and Rustam Shadiev. Can emotional design really evoke emotion in multimedia learning? International Journal of Educational Technology in Higher Education, 17:1-18, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 252 + ], + "type": "text", + "content": "[17] Da Liu, Yaxi Jiang, Min Pei, and Shiguang Liu. Emotional image color transfer via deep learning. Pattern Recognition Letters, 110:16-22, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 254, + 554, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 275 + ], + "type": "text", + "content": "[18] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 277, + 554, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 554, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 554, + 319 + ], + "type": "text", + "content": "[19] Olga Megalakaki, Ugo Ballenghein, and Thierry Baccino. Effects of valence and emotional intensity on the comprehension and memorization of texts. Frontiers in Psychology, 10:179, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 322, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 554, + 365 + ], + "type": "text", + "content": "[20] Liyu Meng, Yuchen Liu, Xiaolong Liu, Zhaopei Huang, Yuan Cheng, Meng Wang, Chuanhe Liu, and Qin Jin. Multimodal emotion estimation for in-the-wild videos. arXiv preprint arXiv:2203.13032, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 367, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 367, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 367, + 554, + 421 + ], + "type": "text", + "content": "[21] Laurent Mertens, Elahe Yargholi, Hans Op de Beeck, Jan Van den Stock, and Joost Vennekens. Findingemo: An image dataset for emotion recognition in the wild. Advances in Neural Information Processing Systems, 37:4956-4996, 2024. 2, 3, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 422, + 554, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 554, + 465 + ], + "type": "text", + "content": "[22] Ali Mollahosseini, Behzad Hasani, and Mohammad H. Mahoor. Affectnet: A database for facial expression, valence, and arousal computing in the wild. IEEE Transactions on Affective Computing, 10(1):18-31, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 468, + 554, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 554, + 511 + ], + "type": "text", + "content": "[23] Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, et al. T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models. arXiv preprint arXiv:2302.08453, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 513, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 554, + 544 + ], + "type": "text", + "content": "[24] Chanjong Park and In-Kwon Lee. Emotional landscape image generation using generative adversarial networks. In Proceedings of ACCV, pages 573-590, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 546, + 554, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 554, + 590 + ], + "type": "text", + "content": "[25] Monika Pawlowska and Ewa Magier-Lakomy. The influence of emotional and non-emotional concepts activation on information processing and unintentional memorizing. Polish Psychological Bulletin, pages 150-159, 2011. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 591, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 554, + 623 + ], + "type": "text", + "content": "[26] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of ICCV, pages 4195-4205, 2023. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 554, + 668 + ], + "type": "text", + "content": "[27] Kuan-Chuan Peng, Tsuhan Chen, Amir Sadovnik, and Andrew C Gallagher. A mixed bag of emotions: Model, predict, and transfer emotion distributions. In Proceedings of CVPR, pages 860-868, 2015. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[28] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, et al. SDXL: Improving latent diffusion models for high-resolution image synthesis. In Proceedings of ICLR, 2024. 2, 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "15226" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 152 + ], + "type": "text", + "content": "[30] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 294, + 196 + ], + "type": "text", + "content": "[31] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of CVPR, pages 10684-10695, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[32] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of CVPR, pages 22500-22510, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 295, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 277 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 277 + ], + "type": "text", + "content": "[33] James A Russell. A circumplex model of affect. Journal of personality and social psychology, 39(6):1161, 1980. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 294, + 311 + ], + "type": "text", + "content": "[34] James A Russell and Albert Mehrabian. Evidence for a three-factor theory of emotions. Journal of research in Personality, 11(3):273-294, 1977. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 312, + 294, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 294, + 335 + ], + "type": "text", + "content": "[35] S Shaphiro and MBJB Wilk. An analysis of variance test for normality. Biometrika, 52(3):591-611, 1965. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 56, + 336, + 294, + 380 + ], + "type": "text", + "content": "[36] Keyu Tian, Yi Jiang, Zehuan Yuan, Bingyue Peng, and Liwei Wang. Visual autoregressive modeling: Scalable image generation via next-scale prediction. Advances in neural information processing systems, 37:84839-84865, 2025. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "text", + "content": "[37] Antoine Toisoul, Jean Kossaifi, Adrian Bulat, Georgios Tzimiropoulos, and Maja Pantic. Estimation of continuous valence and arousal levels from faces in naturalistic conditions. Nature Machine Intelligence, 3(1):42-50, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 427, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 427, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 56, + 427, + 294, + 460 + ], + "type": "text", + "content": "[38] Chai M Tyng, Hafeez U Amin, Mohammad NM Saad, and Aamir S Malik. The influences of emotion on learning and memory. Frontiers in psychology, 8:235933, 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 462, + 294, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 462, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 56, + 462, + 294, + 484 + ], + "type": "text", + "content": "[39] Ashish Vaswani. Attention is all you need. In Proceedings of NeurIPS, page 6000-6010, 2017. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 486, + 294, + 517 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 517 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 517 + ], + "type": "text", + "content": "[40] Manuelde Vega. The representation of changing emotions in reading comprehension. Cognition & emotion, 10(3):303-322, 1996. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 553 + ], + "type": "text", + "content": "[41] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In Proceedings of AAAI, pages 2555-2563, 2023. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 598 + ], + "type": "text", + "content": "[42] Yan Wang, Wei Song, Wei Tao, Antonio Liotta, Dawei Yang, Xinlei Li, et al. A systematic review on affective computing: Emotion models, databases, and recent advances. Information Fusion, 83:19-52, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 600, + 294, + 644 + ], + "type": "text", + "content": "[43] Shuchen Weng, Peixuan Zhang, Zheng Chang, Xinlong Wang, Si Li, and Boxin Shi. Affective image filter: Reflecting emotions from text to images. In Proceedings of ICCV, pages 10810-10819, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 678 + ], + "type": "text", + "content": "[44] Frank Wilcoxon. Individual comparisons by ranking methods. In Breakthroughs in statistics: Methodology and distribution, pages 196-202. 1992. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[45] Weizhen Xie and Weiwei Zhang. Negative emotion enhances mnemonic precision and subjective feelings of remembering in visual long-term memory. Cognition, 166:73-83, 2017. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 127 + ], + "type": "text", + "content": "[46] Chao Xu, Junwei Zhu, Jiangning Zhang, Yue Han, Wenqing Chu, Ying Tai, et al. High-fidelity generalized emotional talking face generation with multi-modal emotion space learning. In Proceedings of CVPR, pages 6609-6619, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 129, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 163 + ], + "type": "text", + "content": "[47] Liwen Xu, Zhengtao Wang, Bin Wu, and Simon Lui. Mdan: Multi-level dependent attention network for visual emotion analysis. In Proceedings of CVPR, pages 9479-9488, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 165, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 165, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 315, + 165, + 553, + 207 + ], + "type": "text", + "content": "[48] Dingkang Yang, Shuai Huang, Shunli Wang, Yang Liu, Peng Zhai, Liuzhen Su, et al. Emotion recognition for multiple context awareness. In Proceedings of ECCV, pages 144–162, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 209, + 553, + 254 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 209, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 315, + 209, + 553, + 254 + ], + "type": "text", + "content": "[49] Jufeng Yang, Dongyu She, and Ming Sun. Joint image emotion classification and distribution learning via deep convolutional neural network. In Proceedings of IJCAI, pages 3266-3272, 2017. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 256, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 256, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 315, + 256, + 553, + 300 + ], + "type": "text", + "content": "[50] Jingyuan Yang, Jiawei Feng, and Hui Huang. Emogen: Emotional image content generation with text-to-image diffusion models. In Proceedings of CVPR, pages 6358-6368, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 302, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 302, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 315, + 302, + 553, + 346 + ], + "type": "text", + "content": "[51] Jingyuan Yang, Jiawei Feng, Weibin Luo, Dani Lischinski, Daniel Cohen-Or, and Hui Huang. Emoedit: Evoking emotions through image manipulation. arXiv preprint arXiv:2405.12661, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 348, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 348, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 348, + 553, + 392 + ], + "type": "text", + "content": "[52] Xiaocui Yang, Shi Feng, Daling Wang, and Yifei Zhang. Image-text multimodal emotion classification via multi-view attentional network. IEEE Transactions on Multimedia, 23: 4014-4026, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 315, + 394, + 553, + 437 + ], + "type": "text", + "content": "[53] Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. Ip-adapter: Text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721, 2023. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 440, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 440, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 315, + 440, + 553, + 484 + ], + "type": "text", + "content": "[54] Ligang Zhang, Dian Tjondronegoro, and Vinod Chandran. Representation of facial expression categories in continuous arousal-valence space: feature and correlation. Image and Vision Computing, 32(12):1067-1079, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 486, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 486, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 315, + 486, + 553, + 518 + ], + "type": "text", + "content": "[55] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In Proceedings of ICCV, pages 3836-3847, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 315, + 520, + 553, + 564 + ], + "type": "text", + "content": "[56] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of CVPR, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 567, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 567, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 315, + 567, + 553, + 620 + ], + "type": "text", + "content": "[57] Zixing Zhang, Liyizhe Peng, Tao Pang, Jing Han, Huan Zhao, and Bjorn W Schuller. Refashioning emotion recognition modelling: The advent of generalised large models. IEEE Transactions on Computational Social Systems, pages 6690-6704, 2024. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 623, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 623, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 315, + 623, + 553, + 667 + ], + "type": "text", + "content": "[58] Sicheng Zhao, Yue Gao, Xiaolei Jiang, Hongxun Yao, TatSeng Chua, and Xiaoshuai Sun. Exploring principles-of-art features for image emotion recognition. In Proceedings of ACM MM, pages 47-56, 2014. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 670, + 553, + 712 + ], + "type": "text", + "content": "[59] Sicheng Zhao, Hongxun Yao, and Xiaolei Jiang. Predicting continuous probability distribution of image emotions in valence-arousal space. In Proceedings of ACM MM, page 879-882, 2015. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "15227" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 150 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[60] Sicheng Zhao, Xingxu Yao, Jufeng Yang, et al. Affective image content analysis: Two decades review and new perspectives. IEEE TPAMI, 44(10):6729-6751, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[61] Siqi Zhu, Chunmei Qing, Canqiang Chen, and Xiangmin Xu. Emotional generative adversarial network for image emotion transfer. Expert Systems with Applications, 216:119485, 2023. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "15228" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_content_list.json b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a98b6ab058e56726a8920e1bd823caf5be6b2732 --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_content_list.json @@ -0,0 +1,1989 @@ +[ + { + "type": "text", + "text": "Emulating Self-attention with Convolution for Efficient Image Super-Resolution", + "text_level": 1, + "bbox": [ + 96, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dongheon Lee, Seokju Yun, Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea", + "bbox": [ + 256, + 180, + 738, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{dslisleedh, wsz871, youngmin.ro}@uos.ac.kr", + "bbox": [ + 307, + 218, + 687, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Code: https://github.com/dslisleedh/ESC", + "bbox": [ + 374, + 236, + 617, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution (SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention (ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to $32 \\times 32$ with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100×2 while reducing latency and memory usage by $16 \\times$ and $12.2 \\times$ . Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution (ESC), notably improves PSNR by 0.27 dB on Urban100×4 compared to HiT-SRF, reducing the latency and memory usage by $3.7 \\times$ and $6.2 \\times$ , respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module.", + "bbox": [ + 88, + 316, + 485, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 734, + 220, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image Super-Resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs and remains an active area of research in computer vision. Recently, the demand for multimedia content and generative models has increased significantly, making SR particularly noteworthy as it enables users to enjoy high-quality content under resource-constrained conditions. Consequently, practical deployment has emerged as a critical consideration in", + "bbox": [ + 89, + 755, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/57e4709eed99ce2895f22728acd5884b2b7ad30facbd5a3f946ef13f48d8a708.jpg", + "image_caption": [ + "Figure 1. Comparison of performance, latency, and memory usage. Our methods are evaluated against representative SR models, including CNNs, Transformers, and SSMs." + ], + "image_footnote": [], + "bbox": [ + 516, + 286, + 903, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SR tasks, motivating numerous SR studies to enhance performance while reducing computational complexity (FLOATing point OPerations; FLOPs) and parameter size.", + "bbox": [ + 511, + 537, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Therefore, Transformers have gained significant attention in the SR task since they have achieved superior performance over Convolutional Neural Networks (CNNs) while requiring lower FLOPs and fewer parameters. By capturing long-range dependencies and performing input-dependent weighting through self-attention, Transformers exhibit high representational capacity with enhanced performance, especially as the training data volume increases [5, 15]. Accordingly, numerous transformer-based methods have been proposed [47, 56, 75] to capitalize on the advantages of self-attention while reducing FLOPs and parameter counts.", + "bbox": [ + 509, + 583, + 908, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, many of these works overlook the excessive memory access caused by self-attention, which arises from materializing the score matrix $(S = QK^T)$ [10] and leveraging memory-bound operations such as tensor reshaping and window masking [39, 68]. The memory access issue is further worsened in the SR architecture, which deals with large feature maps without patchify stem or downsampling stage. For example, SwinIR-light [36] exhibits $4.7\\times$ slower latency and $2\\times$ higher memory usage than CNNs [77] reconstructing an HD image at the $\\times 2$ scale, even though it", + "bbox": [ + 509, + 750, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24467", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "requires $14.5 \\times$ and $17 \\times$ fewer FLOPs and parameter sizes, respectively. Consequently, despite their promising performance, Transformers are challenging to deploy on resource-constrained devices, such as consumer-level GPUs.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we aim to design an efficiency-driven Transformer tailored for lightweight SR tasks, achieving both improved performance and memory overhead. We start by conducting a preliminary analysis of self-attention, observing that the similarity modeling performed by self-attention and resulting extracted features remain highly consistent across several layers, as shown in Figure 2. This finding exhibits that self-attention may extract overlapping features, suggesting the possibility of reducing computational overhead without compromising representational capability by utilizing efficient alternatives. Building on this finding, we establish a design strategy that retains self-attention only in the first layer of each block while replacing the remaining layers with our proposed efficient alternative, the Convolutional Attention (ConvAttn) module. To effectively emulate self-attention's long-range modeling and instance-dependent weighting, the ConvAttn module operates with a twofold mechanism. First, it simplifies self-attention's long-range interaction by applying convolution with a shared $13 \\times 13$ large kernel throughout the entire network, targeting only a subset of channels. Second, dynamic kernels are generated to capture input-dependent weighting, mimicking the adaptive nature of self-attention. By combining these components, the ConvAttn module significantly reduces the reliance on memory-intensive self-attention while maintaining the representational capability of Transformers.", + "bbox": [ + 91, + 154, + 483, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With most self-attention layers replaced by ConvAttn, we leverage this efficiency to further enhance the remaining self-attention layers. Specifically, we enlarge the window size of self-attention, significantly improving performance with only a slight increase in FLOPs. However, increasing the window size leads to an enlarged score matrix, substantially raising peak memory usage. To address this, we introduce Flash Attention [10, 14] into lightweight SR tasks to avoid materializing the score matrix. Our optimized implementation allows us to scale up the window size to $32 \\times 32$ while reducing latency and memory usage by $16 \\times$ and $12.2 \\times$ , respectively, as illustrated in Table 1.", + "bbox": [ + 89, + 550, + 482, + 731 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on these approaches, we introduce the lightweight SR network called Emulating Self-attention with Convolution (ESC). The proposed ESC outperforms ATD-light [73] on Urban $100 \\times 4$ with 0.1dB improvements on PSNR while being $8.9 \\times$ faster, as shown in Figure 1. Additionally, ESC-light surpasses ELAN-light [75] by a large margin of 0.29dB in PSNR on Urban $100 \\times 2$ while also reducing latency by $22\\%$ . We further validate our ESC in scenarios where reducing FLOPs and parameter size is essential by introducing ESC-FP, which outperforms MambaIRV2-light [21] on Manga $109 \\times 4$ with reductions of FLOPs and", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "parameter size by $20\\%$ and $32\\%$ , respectively. Through our extensive experiments, we demonstrate that our ESC fully leverages the advantages of Transformers – including their large receptive fields, representational capacity, and scalability with respect to data volume – even though most self-attention are replaced by the ConvAttn module. We support these results with in-depth experiments, as illustrated in Figure 6, suggesting that the proposed ConvAttn module extracts similar features with self-attention.", + "bbox": [ + 511, + 90, + 903, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 531, + 227, + 834, + 241 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We demonstrate that well-designed convolution partially replaces self-attention, significantly improving efficiency without sacrificing the advantages of Transformers.", + "- We mitigate the memory overhead of self-attention by integrating Flash Attention, thereby enlarging the window size to $32 \\times 32$ without excessive memory usage. To the best of our knowledge, this is the first successful application of Flash Attention in lightweight SR tasks.", + "- We fully exploit the Transformer's benefits in lightweight SR tasks by making them more efficient and simpler." + ], + "bbox": [ + 511, + 243, + 903, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 404, + 655, + 420 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional CNNs Early deep learning-based SR research [13, 29, 37, 77, 78] primarily employed convolution that extracts local features from images. These approaches typically relied on stacked convolutions, which often resulted in over-parameterization. To address this issue, numerous studies [28, 35, 46, 53] have shared their weights across multiple layers to reduce parameter size. However, many of them suffer from performance degradation since the input-independent weights are shared across multiple layers. In this paper, we share a large kernel through the entire network while leveraging input-dependently generated dynamic Depth-Wise Convolution (DWC) [22] by each layer in parallel. This approach efficiently mitigates the parameter growth of the large kernel while maintaining each layer's representational power.", + "bbox": [ + 511, + 431, + 906, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformers Pioneering studies that significantly reduce computational costs by computing self-attention within local windows [36] or channel-wise [69] have sparked considerable interest in applying Transformers [15, 55] to the SR task. Subsequent studies have proposed methods that enhance self-attention to strengthen its long-range modeling capabilities while reducing both FLOPs and parameters [7, 47, 56, 73, 75, 76, 82]. Nonetheless, Transformers remain challenging to deploy due to the excessive memory access caused by self-attention and involved memory-bound operations. In this paper, we address this limitation by replacing self-attention with convolutions that emulate its advantages, based on the observations of self-attention's inter-layer repetition. Our approach preserves the advantages of Transformers while reducing reliance on self-attention and associated memory-bound operations.", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24468", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a270a1cbcef1320720f80ffc80268667ef698c571b0fa59d655c777c60961f22.jpg", + "image_caption": [ + "(a) CKA Similarity of $F$", + "(b) Cosine Similarity of $A^{avg}$" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 477, + 213 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Large Kernel CNNs and State-Space Models Several recent studies have highlighted the computational burden of Transformers and investigated alternative core operators to replicate their advantages. Specifically, CNN-based models [52, 62, 63, 81] attempt to mimic the advantages of Transformers by leveraging large kernels and pixel attention [79], while State-Space Models (SSM) [20, 21] adapt their verified long-range modeling abilities from the 1D sequence domain [17, 18] to the 2D image domain. However, these approaches often overlook either the exceptional representational capability brought by self-attention's sophisticated similarity modeling or the additional memory access caused by supplementary mechanisms to adapting 2D domains like multi-directional or attentive scanning. In this paper, we fully leverage the advantages of both self-attention and convolution while utilizing Flash Attention [10, 14] to address the memory access of self-attention without relying on complex additional mechanisms.", + "bbox": [ + 91, + 325, + 483, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Methods", + "text_level": 1, + "bbox": [ + 89, + 607, + 272, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we start by representing our preliminary analysis and then describe our proposed network.", + "bbox": [ + 89, + 628, + 482, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary analysis", + "text_level": 1, + "bbox": [ + 89, + 670, + 285, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transformers often achieve improved performance by utilizing self-attention's long-range modeling ability and instance-dependent weighting. However, each self-attention incurs excessive memory overhead, resulting in significant latency and memory usage. Interestingly, our preliminary analysis demonstrates that the sophisticated similarity modeling of self-attention $(A^{avg})$ and the resulting extracted features $(F)$ exhibit a high degree of inter-layer similarities, with averages of $89\\%$ and $87\\%$ . This observation not only aligns with recent research [80] but also explains the success of studies that reduced computational cost by sharing the attention map across layers [47, 75]. Inspired by this observation, we investigate improving efficiency without compromising representational power.", + "bbox": [ + 89, + 688, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6c89fe3b2a2a5d7d51bd5c02340abdca3996a0849c2c522af739c538816d8674.jpg", + "image_caption": [ + "Figure 2. Visualized inter-layer similarities of the feature extracted by self-attention $(F)$ before being added to the skip path and corresponding attention map $(A^{avg})$ from SwinIR-light [36]. Both CKA [31] and cosine similarities are measured on the Urban $100\\times 4$ dataset. The window and head dimensions of the attention map are averaged for visualization.", + "Figure 3. Visualization of the overall architecture." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 901, + 212 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overall Structure", + "text_level": 1, + "bbox": [ + 511, + 236, + 687, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our network is composed of four main components, as shown in Figure 3. First, the LR input image $I_{LR} \\in \\mathbb{R}^{H \\times W \\times 3}$ is transformed into a shallow feature $F_0 \\in \\mathbb{R}^{H \\times W \\times C}$ through a $3 \\times 3$ convolution $(\\mathrm{Conv}_{3 \\times 3})$ . Here, $H$ and $W$ denote height and width of $I_{LR}$ , and $C$ denotes feature dimension. Next, $F_0$ and a large kernel $LK \\in \\mathbb{R}^{13 \\times 13 \\times 16 \\times 16}$ that is learned end-to-end manner and shared across the entire model are fed into the deep feature extractor $H$ , and the resulting deep feature is added back to $F_0$ to produce the final feature $F \\in \\mathbb{R}^{H \\times W \\times C}$ . In parallel, the image-wise skip module $S$ takes $I_{LR}$ as input to generate a skipped feature $F_s$ , and the upsampler $U$ utilize both $F$ and $F_s$ as inputs to return the upsampled image $I_{SR} \\in \\mathbb{R}^{rH \\times rW \\times 3}$ where $r$ denotes upscaling factor. The entire process is formulated as follows:", + "bbox": [ + 511, + 257, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {0} = \\operatorname {C o n v} _ {3 \\times 3} \\left(I _ {L R}\\right), \\\\ F = H \\left(F _ {0}, L K\\right) + F _ {0}, \\tag {1} \\\\ F _ {s} = S \\left(I _ {L R}\\right), \\\\ I _ {S R} = U (F, F _ {s}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 493, + 903, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The $H$ is composed of the $N$ ESCBlocks where $N$ denotes the number of ESCBlocks and a $3 \\times 3$ convolution, and is formulated as follows:", + "bbox": [ + 511, + 571, + 905, + 616 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\begin{array}{l} F _ {i} = \\operatorname {E S C B l o c k} _ {i} \\left(F _ {i - 1}, L K\\right), \\quad i = 1, \\dots , N, \\\\ F _ {i} = G _ {i - 1} \\left(F _ {i}\\right) \\end{array} \\tag {2} \\\\ F _ {d} = \\operatorname {C o n v} _ {3 \\times 3} \\left(F _ {N}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 622, + 903, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{i-1}$ indicates the feature input to the $i$ th ESCBlock together with $LK$ , and $F_i$ represents the extracted feature. The $U$ and $S$ will be described in the implementation detail section because they will be changed depending on the task.", + "bbox": [ + 511, + 664, + 905, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. ESC Block", + "text_level": 1, + "bbox": [ + 511, + 732, + 633, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall process of the ESCBlock is illustrated in Figure 4 and is formulated as follows:", + "bbox": [ + 511, + 753, + 905, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i} ^ {i n} = \\operatorname {C o n v F F N} \\big (\\mathrm {L N} (F _ {i - 1}) \\big),\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 789, + 730, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i, 0} = F _ {i} ^ {i n} + \\operatorname {S e l f A t t n} \\left(\\mathrm {L N} \\left(F _ {i} ^ {i n}\\right)\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 810, + 759, + 828 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {i, j} = F _ {i, j - 1} + \\operatorname {C o n v A t t n} _ {j} \\left(\\operatorname {C o n v F F N} _ {j} \\left(F _ {i, j - 1}\\right), L K\\right), \\\\ j = 1, \\dots , M, \\\\ F _ {i} = F _ {i - 1} + \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {L N} \\left(F _ {i, M}\\right)\\right). \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 830, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24469", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/65d86a19ecc2491781bd546dee653a97e6cc843f672b761b396f27107055e837.jpg", + "image_caption": [ + "Figure 4. Visualization of the proposed ESC block." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 478, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, the $F_{i-1}$ is converted to the $F_{i}^{in}$ by layer normalization (LN) and convolutional FFN (ConvFFN) [82] with the kernel size of three. Then, LN and naive window self-attention [36] with the window size of $32 \\times 32$ are employed in a residual connection to extract $F_{i,0}$ . Despite its large window size, we effectively reduce memory overhead by utilizing Flash Attention. Furthermore, since ConvFFN is placed before, the self-attention extracts features considering local information without the complicated QKV projection module. After establishing $F_{i,0}$ , the block iterates residual paths composed of ConvFFN and proposed ConvAttn module (ConvAttn) $M$ times. Although our methods do not employ shifted-window self-attention, these convolution-based modules after the self-attention will extract inter-window features. Finally, the result of the last iteration $F_{i,M}$ go through LN and $\\mathrm{Conv}_{3 \\times 3}$ , then is added to $F_{i-1}$ , resulting in $F_{i}$ .", + "bbox": [ + 89, + 310, + 483, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Convolutional Attention Module The ConvAttn is designed to emulate self-attention's two key advantages: capturing long-range dependencies and instance-dependent weighting. For this purpose, ConvAttn leverages two types of convolution formulated as:", + "bbox": [ + 89, + 566, + 483, + 642 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} F _ {i, j} ^ {a t t}, F _ {i, j} ^ {i d t} = \\mathrm {S p l i t} _ {1 6: C - 1 6} (F _ {i, j} ^ {\\mathrm {C F}}), \\\\ D K _ {i, j} = \\operatorname {C o n v} _ {1 \\times 1} ^ {u p} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} ^ {\\text {d o w n}} \\left(\\operatorname {G A P} \\left(F _ {i, j} ^ {\\text {a t t}}\\right)\\right)\\right)\\right), \\\\ F _ {i, j} ^ {r e s} = (F _ {i, j} ^ {a t t} \\circledast D K _ {i, j}) + (F _ {i, j} ^ {a t t} \\circledast L K), \\\\ F _ {i, j} ^ {f u s e} = \\operatorname {C o n v} _ {1 \\times 1} ^ {f u s e} \\left(\\operatorname {C o n c a t} \\left(F _ {i, j} ^ {r e s}, F _ {i, j} ^ {i d t}\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 648, + 480, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F_{i,j}^{\\mathrm{CF}}$ denotes the features extracted by $\\mathrm{ConvFFN}_{i,j}$ before. We first split the $F_{i,j}^{\\mathrm{CF}}$ channel-wise, resulting in $F_{i,j}^{att} \\in \\mathbb{R}^{H \\times W \\times 16}$ , which consists of the 16 channels in front, and $F_{i,j}^{idt} \\in \\mathbb{R}^{H \\times W \\times C - 16}$ , which contains the remaining part. Two types of convolutions are only applied within $F_{i,j}^{att}$ to reduce memory access cost, unlike the methods that leverage large kernels considering only computational complexity [52, 62, 63]. Next, dynamic depth-wise kernel $DK_{i,j} \\in \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16}$ [22] is created by two layers of $\\mathrm{Conv}_{1 \\times 1}$ with global average pooling (GAP) and gaussian", + "bbox": [ + 89, + 743, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "error linear unit activation $(\\phi)$ . Since enlarging $DK$ 's kernel size increases the parameter size multiplied by the hidden dimension $(h)$ of the kernel estimator, we fix its kernel size to $3\\times 3$ . Then both $DK_{i,j}$ and $LK$ are convolved on $F_{i,j}^{att}$ and sum up those, resulting convolved feature $F_{i,j}^{res}$ . Based on our preliminary analysis, we hypothesize that long-range interactions do not significantly vary across layers. Therefore, we share $LK$ across the entire ESCBlocks and ConvAttns, mitigating parameter growth and optimization difficulty [6, 38]. Finally, $F_{i,j}^{res}$ and $F_{i,j}^{idt}$ are concatenated channel-wise, followed by the last convolution $\\mathrm{Conv}_{1\\times 1}^{fuse}$ to fuse them, resulting in fused feature $F_{i,j}^{use}$ . In this way, both the local features extracted by ConvFFN and the long-range features extracted by ConvAttn can be leveraged to capture multi-scale information.", + "bbox": [ + 511, + 90, + 903, + 321 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While $DK_{i,j}$ can be merged into $LK$ by padding both spatial and channel dimensions to reduce FLOPs, we do not because it slows down latency. On the other hand, in situations where reducing FLOPs and parameter size is crucial rather than reducing latency, $LK$ can be decomposed depthwise separable manner [25] into point-wise large kernel filter $LK^{\\mathrm{c}}\\in \\mathbb{R}^{1\\times 1\\times 16\\times 16}$ and depth-wise large kernel filter $LK^{\\mathrm{s}}\\in \\mathbb{R}^{13\\times 13\\times 1\\times 16}$ . This decomposition redefines line 3 in Equation 4 as follows:", + "bbox": [ + 511, + 323, + 906, + 459 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {i, j} ^ {r e s} = \\left(F _ {i, j} ^ {a t t} * L K ^ {c}\\right) * \\left(Z P \\left(D K _ {i, j}\\right) + L K ^ {s}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 469, + 906, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $ZP(\\cdot): \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16} \\mapsto \\mathbb{R}^{13 \\times 13 \\times 1 \\times 16}$ denotes spatial zero padding.", + "bbox": [ + 511, + 496, + 903, + 527 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, ConvAttn efficiently emulates self-attention's two advantages by leveraging a shared $LK$ and layer-specific $DKs$ . In contrast to Transformers that share attention maps across several layers [47, 75, 80], which still require a full materialized attention map for every layer, our module only leverages a single shared $LK$ , significantly reducing memory footprints. Moreover, unlike recent CNNs [52, 62, 63, 81] that use large kernel convolutions and optionally employ pixel-wise attention to inject input dependency, we employ layer-specific $DKs$ , emulating self-attention's input-dependent weight more exactly in a memory-efficient manner.", + "bbox": [ + 511, + 527, + 905, + 708 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e3539a100e0566def5b8b0eb4860f342bbce17457c261128675a2b5d6e35169a.jpg", + "table_caption": [ + "Table 1. Comparison of latency and memory usage between naive self-attention and our Flash Attention (FA) [10, 14]. WS and #Pixels denote the window size of self-attention and the corresponding number of pixels. $X$ refers to the input tensor." + ], + "table_footnote": [], + "table_body": "
WS(#Pixels)TypeX ∈ R64×64×64X ∈ R288×288×64
Latency(ms)Memory Usage(mb)Latency(ms)Memory Usage(mb)
16×16(256)w/o FA2.0464.9758
w/ FA0.5 (4×)11 (4.2×)1.3 (3.8×)223 (3.4×)
24×24(576)w/o FA7.611114.51575
w/ FA0.5 (15.2×)14 (7.9×)2.0 (7.3×)223 (7.1×)
32×32(1024)w/o FA27.315744.72717
w/ FA0.5 (54.6×)11 (14.3×)2.8 (16.0×)223 (12.2×)
", + "bbox": [ + 514, + 781, + 906, + 898 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24470", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/cc215dab4958cfd0343f535d715af742396274e9278927044953af084d3af5ba.jpg", + "table_caption": [ + "Table 2. Comparisons of classic SR methods trained on the DIV2K dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. † denotes that we re-calculate the statistics of the indicated methods. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + ], + "table_footnote": [], + "table_body": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)Set5Set14PSNR / SSIM B100Urban100Manga109
SwinIR-lt [36]1409.81287244.291038.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-lt [75]94.5887203.162138.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
OmniSR [56]120.31031194.577238.22/0.961333.98/0.921032.36/0.902033.05/0.936339.28/0.9784
SRFormer-lt [82]1456.31184236.385338.23/0.961333.94/0.920932.36/0.901932.91/0.935339.28/0.9785
ATD-lt [73]733.52839380.075338.29/0.961634.10/0.921732.39/0.902333.27/0.937539.52/0.9789
HiT-SRF [76]268.11804226.584738.26/0.961534.01/0.921432.37/0.902333.13/0.937239.47/0.9787
ASID-D8 [47] ×2131.2999190.5†73238.32/0.961834.24/0.923232.40/0.902833.35/0.9387-/-
MambaIR-lt [20]277.11695334.290538.13/0.961033.95/0.920832.31/0.901332.85/0.934939.20/0.9782
MambaIRV2-lt [21]580.42824286.377438.26/0.961534.09/0.922132.36/0.901933.26/0.937839.35/0.9785
RDN [78]279.320585096.22212338.24/0.961434.01/0.921232.34/0.901732.89/0.935339.18/0.9780
RCAN [77]299.36263529.71544538.27/0.961434.12/0.921632.41/0.902733.34/0.938439.44/0.9786
ESC-FP (Ours)94.3627239.852438.27/0.961734.09/0.921932.37/0.902233.22/0.937539.40/0.9784
ESC-lt (Ours)73.2830359.460338.24/0.961533.98/0.921132.35/0.902033.05/0.936339.33/0.9786
ESC (Ours)120.9831592.094738.35/0.961934.11/0.922332.41/0.902733.46/0.939539.54/0.9790
SwinIR-lt [36] ×3331.7596110.891834.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-lt [75]32.539990.162934.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
OmniSR [56]41.247688.478034.70/0.929430.57/0.846929.28/0.809428.84/0.865634.22/0.9487
SRFormer-lt [82]530.5537105.486134.67/0.929630.57/0.846929.26/0.809928.81/0.865534.19/0.9489
ATD-lt [73]274.41258168.076034.74/0.930030.68/0.848529.32/0.810929.17/0.870934.60/0.9506
HiT-SRF [76]124.91464101.685534.75/0.930030.61/0.847529.29/0.810628.99/0.868734.53/0.9502
ASID-D8 [47] ×361.946086.4†73934.84/0.930730.66/0.849129.32/0.811929.08/0.8706-/-
MambaIR-lt [20]109.3760148.591334.63/0.928830.54/0.845929.23/0.808428.70/0.863134.12/0.9479
MambaIRV2-lt [21]259.01250126.778134.71/0.929830.68/0.848329.26/0.809829.01/0.868934.41/0.9497
RDN [78]146.59852281.22230834.71/0.929630.57/0.846829.26/0.809328.80/0.865334.13/0.9484
RCAN [77]85.15601586.11562934.74/0.929930.65/0.848229.32/0.811129.09/0.870234.44/0.9499
ESC-FP (Ours)34.8291110.053034.72/0.930030.67/0.848329.30/0.810729.12/0.870634.56/0.8706
ESC-lt (Ours)25.2384162.861234.61/0.929530.52/0.847529.26/0.810228.93/0.867934.33/0.9495
ESC (Ours)41.4385267.695534.84/0.930830.74/0.849329.34/0.811829.28/0.873934.66/0.9512
SwinIR-lt [36] ×4222.935163.693032.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-lt [75]18.024154.164032.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
OmniSR [56]22.527350.979232.49/0.898828.78/0.785927.71/0.741526.64/0.801831.02/0.9151
SRFormer-lt [82]287.232962.887332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
ATD-lt [73]189.7753100.176932.63/0.899828.89/0.788627.79/0.744026.97/0.810731.48/0.9198
HiT-SRF [76]82.1133158.086632.55/0.899928.87/0.788027.75/0.743226.80/0.806931.26/0.9171
ASID-D8 [47] ×461.826549.6†74832.57/0.899028.89/0.789827.78/0.744926.89/0.8096-/-
MambaIR-lt [20]55.843884.692432.42/0.897728.74/0.784727.68/0.740026.52/0.798330.94/0.9135
MambaIRV2-lt [21]153.474875.679032.51/0.899228.84/0.787827.75/0.742626.82/0.807931.24/0.9182
RDN [78]66.07911309.22227132.47/0.899028.81/0.787127.72/0.741926.61/0.802831.00/0.9151
RCAN [78]52.2540917.61559232.63/0.900228.87/0.788927.77/0.743626.82/0.808731.22/0.9173
ESC-FP (Ours)21.715860.853932.56/0.889728.87/0.788927.75/0.743526.90/0.809831.40/0.9192
ESC-lt (Ours)13.321391.062432.52/0.899528.87/0.787827.72/0.742326.76/0.805831.26/0.9173
ESC (Ours)21.9215149.296832.68/0.901128.93/0.790227.80/0.744727.07/0.814431.54/0.9207
", + "bbox": [ + 89, + 132, + 908, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 681, + 222, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This section demonstrates the effectiveness of our flash attention and our extensive experiments on various tasks. For implementation details for proposed networks, please refer to Section 7 in the supplementary.", + "bbox": [ + 89, + 708, + 482, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Effectiveness of Flash Attention", + "text_level": 1, + "bbox": [ + 89, + 785, + 369, + 800 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We address the excessive memory overhead of self-attention by leveraging Flash Attention [10, 14]. Implementation details for Flash Attention are described in Section 6 in the supplementary. As shown in Table 1, our self-attention implementation accelerated by Flash Attention significantly reduces the latency and memory usage of", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-attention at the window size of 32 by $16 \\times$ and $12.2 \\times$ , respectively. We do not consider window sizes above 32, as it may cause training instability due to the necessity of padding, since the training patch size is $64 \\times 64$ .", + "bbox": [ + 511, + 681, + 906, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Classic SR Results", + "text_level": 1, + "bbox": [ + 511, + 756, + 692, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To evaluate our proposed networks on the classic SR task, we leverage five commonly used datasets (Set5 [3], Set14 [70], B100 [42], Urban100 [26], Manga109 [43]). We measure peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) on the Y channel after cropping the boundary corresponding to the upscaling factor and converting them into YCbCr space. All statistics are measured by reconstructing an HD im", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24471", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad3af948326f3342cb5f8005975d510b3e9a14d0eac45efea5a7fb215ff5b0a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 233, + 176 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c153873211bd2a23aaebb7757db2b6027b83da5d9c15fb68504e9f830e6ae64e.jpg", + "image_caption": [ + "SRFormer-lt" + ], + "image_footnote": [], + "bbox": [ + 233, + 88, + 316, + 121 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/65391aea2f9fc40c1adbc21c5730896859bba76ac65aa04776a69d740d5ea47b.jpg", + "image_caption": [ + "24.62", + "ATD-lt" + ], + "image_footnote": [], + "bbox": [ + 233, + 138, + 316, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/839e2db0b400b76afa72833c903b2c38de879c159c427015545f94a521294572.jpg", + "image_caption": [ + "HiT-SRF" + ], + "image_footnote": [], + "bbox": [ + 316, + 88, + 398, + 121 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/849b55c8467b6952edcedf3182ff0c9c2b813ccb9e505b2feabfae42733e2c29.jpg", + "image_caption": [ + "25.21", + "MambaIRV2-lt" + ], + "image_footnote": [], + "bbox": [ + 318, + 138, + 398, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/142d00e0098cd39b24e6eb5eae1735479f93132f6ba4f23b25f7b27becbbacb1.jpg", + "image_caption": [ + "HR Crop" + ], + "image_footnote": [], + "bbox": [ + 400, + 88, + 480, + 121 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e3c115592736133a2bbb758c1c4ed9e3c1d88e18cf2733f141e5e5adb67b4c17.jpg", + "image_caption": [ + "PSNR", + "ESC (Ours)" + ], + "image_footnote": [], + "bbox": [ + 400, + 138, + 480, + 170 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/591253c6e9a584a32ecb6167fd1829964a297e1b1cbb504dd9de242fff00d60e.jpg", + "image_caption": [ + "Urban100 - img017", + "Urban100 - img059", + "Figure 5. Visual comparisons for classic SR $(\\times 2)$ . The best result on PSNR is bolded." + ], + "image_footnote": [], + "bbox": [ + 93, + 188, + 233, + 275 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ec536c400451b99d049a1c245819215e8478d8031988c4aede1121b48a55c11e.jpg", + "image_caption": [ + "24.45", + "SRFormer-lt" + ], + "image_footnote": [], + "bbox": [ + 233, + 186, + 316, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/33e8d3ff8330afda30bf591b69bc1126290ab1bf4a30e53381895c8dbaa2a0a1.jpg", + "image_caption": [ + "26.56", + "ATD-1t", + "28.90" + ], + "image_footnote": [], + "bbox": [ + 233, + 236, + 316, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8acf3361c1d5ea1ca3991aecd08d03e6d04f799e4471dd2342c2595f99b0c454.jpg", + "image_caption": [ + "25.23", + "HiT-SRI" + ], + "image_footnote": [], + "bbox": [ + 318, + 186, + 398, + 219 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a9a38dd95e565080fad042693ed43cbeaa87f51163b4c9e3adfaacfebc7cf722.jpg", + "image_caption": [ + "2671", + "MambaIRV2-lt", + "26.41" + ], + "image_footnote": [], + "bbox": [ + 318, + 236, + 398, + 268 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c584f3c1cd8088185615ed38bc49c594861712f5ee195443eca248de5ceea6ca.jpg", + "image_caption": [ + "26.50", + "HR Crops" + ], + "image_footnote": [], + "bbox": [ + 400, + 186, + 480, + 218 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/881497e03039a1651738526853488b53a3e959cd0973d27d2511327e57ac4951.jpg", + "image_caption": [ + "PSNR", + "ESC (Ours)", + "37.75" + ], + "image_footnote": [], + "bbox": [ + 400, + 237, + 480, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "age (1280×720) image using an RTX4090 GPU at FP32 precision 100 times. Memory usage is calculated by torch.cuda.max_memory_allocated function and FLOPs are measured by fvcore library. For the classic SR task, we present three variants: ESC, ESC-light, and ESC-FP. ESC and ESC-light prioritize reducing latency, while ESC-FP focuses on reducing FLOPs and parameter size.", + "bbox": [ + 89, + 323, + 483, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Results on the DIV2K dataset", + "text_level": 1, + "bbox": [ + 89, + 436, + 434, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our methods with various representative lightweight SR methods, including CNN-based models [77, 78], Transformers [36, 47, 56, 73, 75, 76, 82], and SSMs [20, 21] trained on the DIV2K dataset [1]. Notably, as shown in Table 2, our ESC achieves a notable 0.1dB PSNR improvement on the Urban $100 \\times 4$ dataset, while being $8.7 \\times$ faster than ATD-light. Moreover, ESC-light is $1.4 \\times$ faster than ELAN-light, exhibiting an impressive performance improvement of 0.34dB PSNR in the Manga109 $\\times 4$ dataset. These results exhibit that state-of-the-art performance can be achieved even when most self-attention are replaced with convolutions, suggesting that replacing self-attention with convolutions does not compromise representational power. Furthermore, ESC-FP outperforms MambaIRV2-light over 0.18dB on the Urban $100 \\times 4$ dataset while reducing FLOPs and parameters by $20\\%$ and $31\\%$ , respectively. This result underscores that our design is effective even when reducing FLOPs and parameter size.", + "bbox": [ + 89, + 455, + 483, + 728 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.2. Visual Results on the DIV2K dataset", + "text_level": 1, + "bbox": [ + 89, + 734, + 390, + 748 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We also visually compare our ESC with representative lightweight SR methods, including SRFormer-light [82], HiT-SRF [76], ATD-light [73], and MambaIRV2-light [21]. As shown in Figure 5, our ESC effectively restores details and achieves a high PSNR, demonstrating that our network also produces visually pleasant results.", + "bbox": [ + 89, + 753, + 482, + 844 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 851, + 241, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct an ablation study to justify our network's composition, as shown in Table 3. First, we introduce two vari-", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8cd7aa61a6c2c2d442316bed03a12112c3b1c9709e49de097d34a31ee6d65923.jpg", + "table_caption": [ + "Table 3. Ablation studies on our ESC network. All comparisons are designed to have similar latencies to ESC. $LK$ , $DK$ , and WS denote large kernel, dynamic kernel, and window size, respectively. All statistics and metrics are measured as detailed in Section 4.2." + ], + "table_footnote": [], + "table_body": "
CasesLatency (ms)Memory (mb)#FLOPs (G)#Params (K)Set5×2Urban100×2
Only Self-attention128.2830456.673738.27/0.961733.23/0.9378
Only ConvAttns126.3514962.4121038.18/0.961332.91/0.9353
9×9 LK117.1831462.292438.33/0.961833.42/0.9392
17×17 LK126.6832768.997738.32/0.961833.40/0.9390
w/o LK Share and DK119.7835591.1195038.32/0.961833.37/0.9389
w/o LK Share120.9835592.0198538.32/0.961933.43/0.9392
w/o DK119.7835591.191138.31/0.961833.36/0.9388
WS16 and more layers118.0807603.6105438.24/0.961433.05/0.9360
WS24 and more layers116.5801583.599538.33/0.961833.28/0.9383
ESC (Ours)120.9831592.094738.35/0.961933.46/0.9395
", + "bbox": [ + 514, + 162, + 906, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ants, one that uses only the self-attention module and another that uses only the ConvAttn. The number of blocks and layers is adjusted to match the latency on par with the ESC. Both variants exhibit a performance drop compared to ESC, suggesting that leveraging both self-attention and ConvAttn offers the best performance-efficiency trade-off. Next, we vary the $LK$ 's kernel size to $9 \\times 9$ and $17 \\times 17$ to determine the optimal size. In both cases, we observe a drop in performance, confirming that a kernel size of $13 \\times 13$ is optimal. Then, we evaluate performance changes by altering the convolution type used in the ConvAttn. First, we observe performance drops when neither $LK$ sharing nor $DK$ is used. Then, we add $DK$ without sharing $LK$ and observe some performance improvement, though it still falls short of ESC. This result suggests that addressing the optimization difficulty that comes from the massive parameter size is more critical than the layer-specific long-range modeling. Finally, we leverage only $LK$ sharing without $DK$ and observe significant performance drops, affirming that harmonizing both $LK$ and $DK$ is crucial. For the next step, we investigate the effect of the self-attention window size by setting it to 16 and 24 and adding extra layers. In every case, we observe a performance drop, which confirms that enlarging the window size to enhance self-attention is effective even when only a single self-attention is leveraged in each block.", + "bbox": [ + 511, + 304, + 906, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.4. Analysis on ConvAttn Module", + "text_level": 1, + "bbox": [ + 511, + 707, + 769, + 722 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section verifies that the proposed ConvAttn operates similarly to self-attention. Self-attention is known to extract shape-biased structural features [12, 48], unlike $3 \\times 3$ convolution, which mainly extracts local texture and edges. Indeed, as shown in Figure 6 (a), $F^{idt}$ focuses on local textures and edges while $F^{att}$ concentrates on structural components. Next, to verify each convolution's receptive field, we visualize positive gradients $\\bar{G}^{+} = \\frac{1}{NMC} \\sum_{p=1}^{NMC} \\max \\left(0, \\frac{\\partial I_{i,j}^{SR}}{\\partial F_p}\\right)$ propagated from the output image's center pixel $(i,j)$ to the inputs feature $(F)$ of both $DK$ and $LK$ across $N$ blocks, $M$ modules, and $C$", + "bbox": [ + 511, + 727, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24472", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09cbfbc1d699ba553d3c35a6fc58f4946202d716ef6cc0dd048381099c86487f.jpg", + "table_caption": [ + "Table 4. Comparisons of classic SR methods trained on the large-scale DFLIP dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. $\\S$ denotes that we train the indicated methods following their descriptions. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + ], + "table_footnote": [], + "table_body": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)PSNR / SSIM
Set5Set14B100Urban100Manga109
SRFormer-lt$ [82]×21838.11184236.385338.24/0.961534.13/0.921832.42/0.902633.37/0.938639.36/0.9787
ATD-lt$ [73]733.52839380.075338.29/0.961634.30/0.923032.43/0.902733.62/0.940139.60/0.9791
HiT-SRF$ [76]268.11804226.584738.31/0.961634.31/0.923032.45/0.903133.58/0.940439.69/0.9793
ESC (Ours)120.9831592.094738.34/0.961834.42/0.923532.50/0.903633.86/0.942439.73/0.9795
SRFormer-lt$ [82]×3668.3537105.486134.67/0.929730.75/0.848429.30/0.810829.10/0.870134.26/0.9498
ATD-lt$ [73]274.41258168.076034.71/0.930030.77/0.849329.33/0.811629.42/0.874334.61/0.9509
HiT-SRF$ [76]124.91464101.685534.69/0.929830.81/0.849329.32/0.811529.28/0.872934.72/0.9511
ESC (Ours)41.4385267.695534.85/0.931230.97/0.851129.41/0.813529.70/0.879934.94/0.9525
SRFormer-lt$ [82]×4327.832962.887332.49/0.899328.89/0.788727.76/0.742926.90/0.808631.25/0.9189
ATD-lt$ [73]189.7753100.176932.52/0.899528.93/0.789627.79/0.744327.18/0.815031.47/0.9208
HiT-SRF$ [76]82.1133158.086632.55/0.899728.96/0.789727.77/0.744327.07/0.813031.59/0.9208
ESC (Ours)21.9215149.296832.79/0.902529.06/0.792727.85/0.746627.45/0.822931.87/0.9239
", + "bbox": [ + 91, + 132, + 906, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3d9a970861a0d3d661c8844eeebcb52b27d54d432db42f6c4f04d6f355e91048.jpg", + "image_caption": [ + "$F^{idt}$" + ], + "image_footnote": [], + "bbox": [ + 91, + 343, + 199, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/35affe7f1e3902deff6d3ea76d8b7e4369be2f9d84d91f5a23356019d3c62836.jpg", + "image_caption": [ + "$\\bar{G}_{DK}^{+}$ (DI: 48.77)" + ], + "image_footnote": [], + "bbox": [ + 223, + 343, + 321, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a0a20bee06f34aed40681fb8d02503e30ec673ab4bd2c9479bae7f053a2caa77.jpg", + "image_caption": [ + "$S_{DK}$" + ], + "image_footnote": [], + "bbox": [ + 352, + 343, + 477, + 422 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ef548769d0b4d3f8ee5f1b118810a9e2b7f74654dd6cbc0a74e55214a0879ada.jpg", + "image_caption": [ + "(a) Feature Map Visualization", + "Figure 6. Visualized feature map, gradients, and input-wise similarity. $F$ is averaged across entire ConvAttn while processing the Lenna image in Set14. $\\tilde{G}^{+}$ and $S$ are calculated on the Urban100, and 10 samples in front of the Urban100, respectively." + ], + "image_footnote": [], + "bbox": [ + 93, + 431, + 199, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bfef0cf15bb9711874b8c09eae430559285602e1f9b8a67ecc615433292d6520.jpg", + "image_caption": [ + "$G_{LK}^{+}$ (DI: 62.23)", + "(b) Gradient Visualization" + ], + "image_footnote": [], + "bbox": [ + 223, + 433, + 348, + 513 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4cd9179fe74725d498817dbaf4140e31bc437341ecdf6431922a4dd3488366b9.jpg", + "image_caption": [ + "(c) Similarity Visualization" + ], + "image_footnote": [], + "bbox": [ + 354, + 433, + 477, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "channels, and corresponding diffusion index (DI). Figure 6 (b) clearly demonstrates that $LK$ extracts more long-range features than $DK$ . Lastly, we visualize input-wise similarity $(S)$ between average-pooled features extracted by $LK$ and $DK$ . As demonstrated in Figure 6 (c), $S_{DK}$ exhibits higher input-dependent diversity than $S_{LK}$ , confirming that $DK$ extracts more input-dependent features than $LK$ .", + "bbox": [ + 89, + 601, + 483, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.5. Local Attribution Map", + "text_level": 1, + "bbox": [ + 89, + 715, + 295, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use the Local Attribution Map (LAM) [19] to verify that our networks can extract long-range dependencies despite using less self-attention. We compare our ESC with studies of various self-attention implementations that aim to reduce FLOPs while expanding the receptive field, such as global-level attention [56], permuted self-attention [82], and spatial-channel correlation [76]. Notably, as described in Figure 7, our ESC references the widest range of pixels to restore the red bounding box along with the highest diffusion index. This result exhibits that our network effectively captures long-range dependencies even though most self", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28a54a66a3da98e00f804e36dd4a842bdff47d01e20c549a2bb0ab6a91691fb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 342, + 642, + 439 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8e9f9d3d98d47e5514293184ba7feecfbfd459092ac30ad7f0b42be2f224e20.jpg", + "image_caption": [ + "Urban100 - img063x4 Diffusion Index", + "Figure 7. Comparisons of Local Attribution Map (LAM) [19] and the corresponding area of contribution and diffusion index. The best result on the diffusion index is bolded." + ], + "image_footnote": [], + "bbox": [ + 516, + 439, + 642, + 551 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a7133d47f6035acf0854cc473d0099708571c6bb8846fe78e982d669b40d647b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 642, + 342, + 746, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d65bda4de02635cd7d89c52dfb7ea503d9b222953dbd368470102f5454cef3e.jpg", + "image_caption": [ + "OmniSR" + ], + "image_footnote": [], + "bbox": [ + 650, + 391, + 714, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c56951aa9bc9019df0c7e4b8842da8444b4e3deee09e1e81cd5ef9bc40c8501d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 720, + 342, + 776, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e985bc46b05a8fc2ae735c7fe1d73f6f84a7cd0407a095ac1925d7160a7fad8.jpg", + "image_caption": [ + "SRFormer-It" + ], + "image_footnote": [], + "bbox": [ + 718, + 391, + 776, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f69aa6221f006f8faf78404a3d331261e07ff16362d3433e7ddfff1d175f304f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 342, + 839, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/cf1db73828c316408c14858af8efd92a98eb797abe052ad76d913da64ea9dbd9.jpg", + "image_caption": [ + "iT-SRF" + ], + "image_footnote": [], + "bbox": [ + 779, + 391, + 841, + 438 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8dff5e1b36b6cc9d7cecc4c919c7bbbead87b13a5ba5dcc8369ccc162d1723d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 841, + 342, + 903, + 391 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/68a3b37e09115a0c7939e7431866efa9835c738ff170463f2c558338dd6797e8.jpg", + "image_caption": [ + "ESC (Ours)" + ], + "image_footnote": [], + "bbox": [ + 841, + 391, + 903, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": " attentions are replaced by the ConvAttn module, suggesting that our ConvAttn module contributes to retaining the large receptive field of Transformers.", + "bbox": [ + 511, + 632, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.6. Quantitative Results on the DFLIP dataset", + "text_level": 1, + "bbox": [ + 511, + 696, + 856, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Transformers exhibit significant performance improvements when trained on large-scale datasets [5, 15]. To verify our model also retains this benefit, we train and evaluate various representative Transformers – including SRFormerlight [82], ATD-light [73], and HiT-SRF [76] – alongside our ESC on a large dataset. For this purpose, we leverage DIV2K+Flickr2K+LSDIR+DiverSeg-IP (DFLIP) [1, 34, 45, 54] dataset. As illustrated in Table 4, our approach outperforms the other methods by a substantial margin across all scales and evaluation datasets. These results indicate that replacing self-attention with ConvAttn does not sacrifice the data scaling capability of Transformers.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24473", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Arbitrary-Scale SR", + "text_level": 1, + "bbox": [ + 89, + 90, + 276, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To evaluate our proposed networks on arbitrary-scale SR tasks, we leverage the DIV2K-validation [1] dataset. Following previous research [8, 32], we assess PSNR in the RGB channel after cropping the boundaries according to the upscaling factor plus six. For arbitrary-scale SR tasks, we change $U$ of our network and representative lightweight SR transforems [73, 76] into LTE [32].", + "bbox": [ + 89, + 112, + 483, + 219 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.1. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 227, + 279, + 242 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the outstanding representational capability of the proposed network, we train and evaluate the network on an arbitrary-scale SR task, which requires processing various upscaling factors with a single network. For quantitative comparison, we leverage commonly used RDN [78] and additional representative Transformers [73, 76] that we train. As described in Tables 5, our network continually achieves high performance at the scale both seen and unseen during training, even achieving a $0.09\\mathrm{dB}$ improvement over $\\mathrm{RDN + LTE}$ at $\\times 2$ . This result underscores our ESC's exceptional representational capability. For quantitative results measured on other datasets, including Set5, Set14, B100, and Urban100, refer to Section 8 in the supplementary.", + "bbox": [ + 89, + 246, + 483, + 443 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/de11fa0f491338a31aee418424ede1b031fb0879e648dba3ac1aba0c351f383d.jpg", + "table_caption": [ + "Table 5. Quantitative comparison on arbitrary-scale SR task employing LTE [32] as upsampler. The best result on PSNR is bolded." + ], + "table_footnote": [], + "table_body": "
MethodsSeenUnseen
×2×3×4×6×12×18×24×30
RDN+LTE [78]35.0431.3229.3327.0423.9522.4021.3620.64
ATD-lt+LTE§ [73]34.9831.2529.2727.0123.9422.3821.3720.64
HiT-SRF+LTE§ [76]35.1031.3529.3527.0523.9422.3921.3620.64
ESC+LTE (Ours)35.1331.3829.3727.0823.9722.4021.3820.64
", + "bbox": [ + 91, + 486, + 483, + 563 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3.2. Visual Results", + "text_level": 1, + "bbox": [ + 89, + 575, + 235, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visually compare our network with RDN, HiT-SRF, and ATD-light. As shown in Figure 8, Our network effectively restores fine details that other methods fail on both $\\times 4$ and $\\times 12$ scales. This result demonstrates that our network also excels at restoring visually present results on both in-distribution and out-of-distribution scales.", + "bbox": [ + 89, + 595, + 483, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Real-World Super-Resolution", + "text_level": 1, + "bbox": [ + 89, + 696, + 354, + 713 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further test our network on real-world SR tasks that aim to restore images with unknown degradations. To this end, we introduce a variant called ESC-Real and compare them to representative real-world SR methods, including RealESRGAN+ [59], SwinIR-Real [36], and DASR [61]. For our evaluation, we use the commonly employed RealSRSet dataset [72]. As shown in Table 9, our ESC-Real recovers more fine-grained details than the other methods while significantly reducing unpleasant artifacts in both graphic and natural images. This result demonstrates that our proposed methods are effective even when dealing with complex unknown degradations, underscoring our methods'", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e48203c369aeef8b5aceb8276a16d186b4609294a80ad356b7dc6ebdb383a22f.jpg", + "image_caption": [ + "Urban100 - img011" + ], + "image_footnote": [], + "bbox": [ + 527, + 90, + 705, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/69599c31241637d91e9c41a8477276e7ff0f8f475fc33b2a309fc997a4a63956.jpg", + "image_caption": [ + "RDN/17.71 HiT-SRF/17.56" + ], + "image_footnote": [], + "bbox": [ + 527, + 199, + 704, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5261f96ba48392c1cc34fe5cbbeab63e52dc9b89901d2e3dd11516fa527d6e8b.jpg", + "image_caption": [ + "ATD-lt / 17.91 ESC (Ours) / 20.70", + "(a) Scale $\\times 4$" + ], + "image_footnote": [], + "bbox": [ + 527, + 244, + 704, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/004309304b2629b400b6dd09514120fb4ff0293752614acdf976ae2006651900.jpg", + "image_caption": [ + "DIV2K-valid - 0845" + ], + "image_footnote": [], + "bbox": [ + 712, + 90, + 890, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5aae8ff9766fd6649fa991747e5884d70dc73248b0214f6d1217292993674d8c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 200, + 803, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9ca1c858354b8bb8f7d7f76d144d632ff1ad8a21a4aa2771a8445659bc2f233d.jpg", + "image_caption": [ + "RDN/16.25 HiT-SRF/17.28", + "ATD-lt / 16.77", + "(b) Scale $\\times 12$" + ], + "image_footnote": [], + "bbox": [ + 712, + 244, + 802, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4454287fc45e67358b390d1ed12973c9fddb5d8d08b416b79b0c5a7fa82588d1.jpg", + "image_caption": [ + "ESC (Ours) / 17.82" + ], + "image_footnote": [], + "bbox": [ + 805, + 244, + 890, + 279 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ce69f6c552a458a04e971142c286f89dcb172cf8a519dc1efa8c5cdbe994c86f.jpg", + "image_caption": [ + "Figure 8. Visual comparisons on arbitrary-scale SR for both indistribution $(\\times 4)$ and out-of-distribution $(\\times 12)$ scales. The best result on PSNR is bolded.", + "(a) RealSRSet - 0030" + ], + "image_footnote": [], + "bbox": [ + 517, + 357, + 901, + 479 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/13859fafde256993ad2a62390e20e941d31a546487edda44d7296f90d2e04ca8.jpg", + "image_caption": [ + "(b) RealSRSet - dog", + "Figure 9. Visual comparisons on real-world SR $(\\times 4)$ tasks." + ], + "image_footnote": [], + "bbox": [ + 517, + 494, + 901, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "outstanding representational capability. Quantitative results on real-world SR tasks are presented in Section 9 in the supplementary.", + "bbox": [ + 511, + 670, + 905, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 724, + 633, + 739 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose ESC, a novel lightweight SR network designed to mitigate the excessive memory overhead of Transformers. Our extensive experiments demonstrate that most self-attention layers can be effectively replaced with carefully designed convolutions to improve efficiency. Moreover, we introduce Flash Attention into lightweight SR tasks, significantly enhancing performance without complicated modifications to self-attention. As a result, we fully exploit the advantages of Transformers by prioritizing improvements in their efficiency.", + "bbox": [ + 511, + 750, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24474", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400].", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 148, + 187, + 164 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6, 7, 8, 1", + "[2] Yang Ai, Xiaoqiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. NeurIPS, 37:55443-55469, 2024. 2", + "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 2", + "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, pages 3086-3095, 2019. 2", + "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 1, 7", + "[6] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In CVPR, pages 5557-5567, 2024. 4", + "[7] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In CVPR, pages 22367-22377, 2023. 2", + "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, pages 8628-8638, 2021. 8", + "[9] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NeurIPS, 2022. 3", + "[10] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 1, 2, 3, 4, 5", + "[11] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 44(5):2567-2581, 2020. 2", + "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, pages 11963-11975, 2022. 6", + "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE TPAMI, 38(2):295-307, 2015. 2", + "[14] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A programming model for generating optimized attention kernels. arXiv preprint arXiv:2412.05496, 2024. 2, 3, 4, 5, 1" + ], + "bbox": [ + 93, + 172, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 7", + "[16] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In CVPRW, pages 2494-2502, 2021. 1", + "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint, 2023. 3", + "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. *ICLR*, 2022. 3", + "[19] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9199-9208, 2021. 7", + "[20] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 3, 5, 6", + "[21] Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. Mambairv2: Attentive state space restoration. CVPR, 2025. 2, 3, 5, 6", + "[22] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between local attention and dynamic depth-wise convolution. *ICLR*, 2022, 2, 4", + "[23] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In ECCV, pages 289-305. Springer, 2024. 1", + "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 2", + "[25] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 4", + "[26] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 5, 2", + "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021. 2", + "[28] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, pages 1637-1645, 2016. 2", + "[29] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 2", + "[30] Diederik P Kingma and Jimmy Ba. Adam: a method for stochastic optimization. In ICLR, 2014. 1", + "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representa" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24475", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tions revisited. In ICML, pages 3519-3529. PMLR, 2019. 3", + "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In CVPR, pages 1929-1938, 2022. 8, 1, 2", + "[33] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In IJCAI, pages 1089-1097, 2023. 3", + "[34] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPRW, pages 1775-1787, 2023. 7, 1", + "[35] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image superresolution. In CVPR, pages 3867-3876, 2019. 2", + "[36] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 8", + "[37] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 136-144, 2017. 2", + "[38] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. ICLR, 2023. 4", + "[39] Yong Liu, Hang Dong, Boyang Liang, Songwei Liu, Qingji Dong, Kai Chen, Fangmin Chen, Lean Fu, and Fei Wang. Unfolding once is enough: A deployment-friendly transformer unit for super-resolution. In ACMMM, pages 7952-7960, 2023. 1", + "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1", + "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 1", + "[42] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423. IEEE, 2001. 5, 2", + "[43] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 5", + "[44] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 20(3): 209-212, 2012. 2", + "[45] Go Ohtani, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka, and Yoshimitsu Aoki. Rethinking image super-resolution from training data perspectives. In ECCV, pages 19-36. Springer, 2024. 7, 1", + "[46] Karam Park and Nam Ik Cho. Partial filter-sharing: Improved parameter-sharing method for single image super-resolution networks. In WACV, pages 2653-2663. IEEE, 2025. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Karam Park, Jae Woong Soh, and Nam Ik Cho. Efficient attention-sharing information distillation transformer for lightweight single image super-resolution. In AAAI, 2025. 1, 2, 3, 4, 5, 6", + "[48] Namuk Park and Songkuk Kim. How do vision transformers work? In ICLR, 2022. 6", + "[49] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1", + "[50] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 1", + "[51] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1", + "[52] Long Sun, Jinshan Pan, and Jinhui Tang. Shufflemixer: An efficient convnet for image super-resolution. NeurIPS, 35: 17314-17326, 2022. 3, 4", + "[53] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, pages 3147-3155, 2017. 2", + "[54] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 114-125, 2017. 7, 1", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2", + "[56] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Liu jinfan. Omni aggregation networks for lightweight image super-resolution. In CVPR, 2023. 1, 2, 5, 6, 7", + "[57] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, pages 2555-2563, 2023. 2", + "[58] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, pages 606-615, 2018. 2", + "[59] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, pages 1905-1914, 2021. 8, 2", + "[60] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, pages 101-117. Springer, 2020. 2", + "[61] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In CVPR, pages 13385-13394, 2021. 8", + "[62] Gang Wu, Junjun Jiang, Junpeng Jiang, and Xianming Liu. Transforming image super-resolution: a convformer-based efficient approach. IEEE Transactions on Image Processing, 2024. 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24476", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[63] Chengxing Xie, Xiaoming Zhang, Linze Li, Haiteng Meng, Tianlin Zhang, Tianrui Li, and Xiaole Zhao. Large kernel distillation network for efficient single image superresolution. In CVPRW, pages 1283-1292, 2023. 3, 4", + "[64] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, pages 1191-1200, 2022. 2", + "[65] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1", + "[66] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, pages 15703-15712, 2025. 1", + "[67] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In WACV, pages 4956-4965, 2023. 3", + "[68] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5756-5767, 2024. 1", + "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: efficient attention transformer for high-resolution image restoration. In CVPR, pages 5728-5739, 2022. 2", + "[70] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 2", + "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 3", + "[72] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, pages 4791-4800, 2021. 8, 2", + "[73] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In CVPR, pages 2856-2865, 2024. 2, 5, 6, 7, 8, 1, 3", + "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 2", + "[75] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In ECCV, pages 649-667. Springer, 2022. 1, 2, 3, 4, 5, 6", + "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient attention image super-resolution. In ECCV, 2024. 2, 5, 6, 7, 8, 1" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[77] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 1, 2, 5, 6", + "[78] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2, 5, 6, 8", + "[79] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCVW, pages 56-72. Springer, 2020. 3", + "[80] Kun Zhou, Xinyu Lin, LIU Zhonghang, Xiaoguang Han, and Jiangbo Lu. Ups: Unified projection sharing for lightweight single-image super-resolution and beyond. In NeurIPS, 2024. 3, 4", + "[81] Lin Zhou, Haoming Cai, Jinjin Gu, Zheyuan Li, Yingqi Liu, Xiangyu Chen, Yu Qiao, and Chao Dong. Efficient image super-resolution using vast-receptive-field attention. In EC-CVW, pages 256-272. Springer, 2022. 3, 4", + "[82] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In ICCV, pages 12780–12791, 2023. 2, 4, 5, 6, 7, 1, 3" + ], + "bbox": [ + 514, + 92, + 903, + 402 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24477", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_model.json b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7f772b1d7c3b972f60506e6a3dbf1a29c20ac0ab --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_model.json @@ -0,0 +1,3335 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.098, + 0.131, + 0.9, + 0.154 + ], + "angle": 0, + "content": "Emulating Self-attention with Convolution for Efficient Image Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.181, + 0.74, + 0.217 + ], + "angle": 0, + "content": "Dongheon Lee, Seokju Yun, Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea" + }, + { + "type": "text", + "bbox": [ + 0.308, + 0.219, + 0.688, + 0.234 + ], + "angle": 0, + "content": "{dslisleedh, wsz871, youngmin.ro}@uos.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.375, + 0.237, + 0.619, + 0.251 + ], + "angle": 0, + "content": "Code: https://github.com/dslisleedh/ESC" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.726 + ], + "angle": 0, + "content": "In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution (SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention (ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to \\(32 \\times 32\\) with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100×2 while reducing latency and memory usage by \\(16 \\times\\) and \\(12.2 \\times\\). Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution (ESC), notably improves PSNR by 0.27 dB on Urban100×4 compared to HiT-SRF, reducing the latency and memory usage by \\(3.7 \\times\\) and \\(6.2 \\times\\), respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.222, + 0.75 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Image Super-Resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs and remains an active area of research in computer vision. Recently, the demand for multimedia content and generative models has increased significantly, making SR particularly noteworthy as it enables users to enjoy high-quality content under resource-constrained conditions. Consequently, practical deployment has emerged as a critical consideration in" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 1. Comparison of performance, latency, and memory usage. Our methods are evaluated against representative SR models, including CNNs, Transformers, and SSMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.583 + ], + "angle": 0, + "content": "SR tasks, motivating numerous SR studies to enhance performance while reducing computational complexity (FLOATing point OPerations; FLOPs) and parameter size." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.909, + 0.75 + ], + "angle": 0, + "content": "Therefore, Transformers have gained significant attention in the SR task since they have achieved superior performance over Convolutional Neural Networks (CNNs) while requiring lower FLOPs and fewer parameters. By capturing long-range dependencies and performing input-dependent weighting through self-attention, Transformers exhibit high representational capacity with enhanced performance, especially as the training data volume increases [5, 15]. Accordingly, numerous transformer-based methods have been proposed [47, 56, 75] to capitalize on the advantages of self-attention while reducing FLOPs and parameter counts." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.911, + 0.903 + ], + "angle": 0, + "content": "However, many of these works overlook the excessive memory access caused by self-attention, which arises from materializing the score matrix \\((S = QK^T)\\) [10] and leveraging memory-bound operations such as tensor reshaping and window masking [39, 68]. The memory access issue is further worsened in the SR architecture, which deals with large feature maps without patchify stem or downsampling stage. For example, SwinIR-light [36] exhibits \\(4.7\\times\\) slower latency and \\(2\\times\\) higher memory usage than CNNs [77] reconstructing an HD image at the \\(\\times 2\\) scale, even though it" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24467" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "requires \\(14.5 \\times\\) and \\(17 \\times\\) fewer FLOPs and parameter sizes, respectively. Consequently, despite their promising performance, Transformers are challenging to deploy on resource-constrained devices, such as consumer-level GPUs." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.485, + 0.548 + ], + "angle": 0, + "content": "In this paper, we aim to design an efficiency-driven Transformer tailored for lightweight SR tasks, achieving both improved performance and memory overhead. We start by conducting a preliminary analysis of self-attention, observing that the similarity modeling performed by self-attention and resulting extracted features remain highly consistent across several layers, as shown in Figure 2. This finding exhibits that self-attention may extract overlapping features, suggesting the possibility of reducing computational overhead without compromising representational capability by utilizing efficient alternatives. Building on this finding, we establish a design strategy that retains self-attention only in the first layer of each block while replacing the remaining layers with our proposed efficient alternative, the Convolutional Attention (ConvAttn) module. To effectively emulate self-attention's long-range modeling and instance-dependent weighting, the ConvAttn module operates with a twofold mechanism. First, it simplifies self-attention's long-range interaction by applying convolution with a shared \\(13 \\times 13\\) large kernel throughout the entire network, targeting only a subset of channels. Second, dynamic kernels are generated to capture input-dependent weighting, mimicking the adaptive nature of self-attention. By combining these components, the ConvAttn module significantly reduces the reliance on memory-intensive self-attention while maintaining the representational capability of Transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.551, + 0.483, + 0.732 + ], + "angle": 0, + "content": "With most self-attention layers replaced by ConvAttn, we leverage this efficiency to further enhance the remaining self-attention layers. Specifically, we enlarge the window size of self-attention, significantly improving performance with only a slight increase in FLOPs. However, increasing the window size leads to an enlarged score matrix, substantially raising peak memory usage. To address this, we introduce Flash Attention [10, 14] into lightweight SR tasks to avoid materializing the score matrix. Our optimized implementation allows us to scale up the window size to \\(32 \\times 32\\) while reducing latency and memory usage by \\(16 \\times\\) and \\(12.2 \\times\\), respectively, as illustrated in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Based on these approaches, we introduce the lightweight SR network called Emulating Self-attention with Convolution (ESC). The proposed ESC outperforms ATD-light [73] on Urban \\(100 \\times 4\\) with 0.1dB improvements on PSNR while being \\(8.9 \\times\\) faster, as shown in Figure 1. Additionally, ESC-light surpasses ELAN-light [75] by a large margin of 0.29dB in PSNR on Urban \\(100 \\times 2\\) while also reducing latency by \\(22\\%\\). We further validate our ESC in scenarios where reducing FLOPs and parameter size is essential by introducing ESC-FP, which outperforms MambaIRV2-light [21] on Manga \\(109 \\times 4\\) with reductions of FLOPs and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.226 + ], + "angle": 0, + "content": "parameter size by \\(20\\%\\) and \\(32\\%\\), respectively. Through our extensive experiments, we demonstrate that our ESC fully leverages the advantages of Transformers – including their large receptive fields, representational capacity, and scalability with respect to data volume – even though most self-attention are replaced by the ConvAttn module. We support these results with in-depth experiments, as illustrated in Figure 6, suggesting that the proposed ConvAttn module extracts similar features with self-attention." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.228, + 0.836, + 0.242 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.244, + 0.905, + 0.289 + ], + "angle": 0, + "content": "- We demonstrate that well-designed convolution partially replaces self-attention, significantly improving efficiency without sacrificing the advantages of Transformers." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.29, + 0.905, + 0.365 + ], + "angle": 0, + "content": "- We mitigate the memory overhead of self-attention by integrating Flash Attention, thereby enlarging the window size to \\(32 \\times 32\\) without excessive memory usage. To the best of our knowledge, this is the first successful application of Flash Attention in lightweight SR tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.905, + 0.395 + ], + "angle": 0, + "content": "- We fully exploit the Transformer's benefits in lightweight SR tasks by making them more efficient and simpler." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.405, + 0.656, + 0.421 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Traditional CNNs Early deep learning-based SR research [13, 29, 37, 77, 78] primarily employed convolution that extracts local features from images. These approaches typically relied on stacked convolutions, which often resulted in over-parameterization. To address this issue, numerous studies [28, 35, 46, 53] have shared their weights across multiple layers to reduce parameter size. However, many of them suffer from performance degradation since the input-independent weights are shared across multiple layers. In this paper, we share a large kernel through the entire network while leveraging input-dependently generated dynamic Depth-Wise Convolution (DWC) [22] by each layer in parallel. This approach efficiently mitigates the parameter growth of the large kernel while maintaining each layer's representational power." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Transformers Pioneering studies that significantly reduce computational costs by computing self-attention within local windows [36] or channel-wise [69] have sparked considerable interest in applying Transformers [15, 55] to the SR task. Subsequent studies have proposed methods that enhance self-attention to strengthen its long-range modeling capabilities while reducing both FLOPs and parameters [7, 47, 56, 73, 75, 76, 82]. Nonetheless, Transformers remain challenging to deploy due to the excessive memory access caused by self-attention and involved memory-bound operations. In this paper, we address this limitation by replacing self-attention with convolutions that emulate its advantages, based on the observations of self-attention's inter-layer repetition. Our approach preserves the advantages of Transformers while reducing reliance on self-attention and associated memory-bound operations." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24468" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.478, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.125, + 0.215, + 0.243, + 0.227 + ], + "angle": 0, + "content": "(a) CKA Similarity of \\(F\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.215, + 0.445, + 0.227 + ], + "angle": 0, + "content": "(b) Cosine Similarity of \\(A^{avg}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Figure 2. Visualized inter-layer similarities of the feature extracted by self-attention \\((F)\\) before being added to the skip path and corresponding attention map \\((A^{avg})\\) from SwinIR-light [36]. Both CKA [31] and cosine similarities are measured on the Urban \\(100\\times 4\\) dataset. The window and head dimensions of the attention map are averaged for visualization." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.326, + 0.485, + 0.599 + ], + "angle": 0, + "content": "Large Kernel CNNs and State-Space Models Several recent studies have highlighted the computational burden of Transformers and investigated alternative core operators to replicate their advantages. Specifically, CNN-based models [52, 62, 63, 81] attempt to mimic the advantages of Transformers by leveraging large kernels and pixel attention [79], while State-Space Models (SSM) [20, 21] adapt their verified long-range modeling abilities from the 1D sequence domain [17, 18] to the 2D image domain. However, these approaches often overlook either the exceptional representational capability brought by self-attention's sophisticated similarity modeling or the additional memory access caused by supplementary mechanisms to adapting 2D domains like multi-directional or attentive scanning. In this paper, we fully leverage the advantages of both self-attention and convolution while utilizing Flash Attention [10, 14] to address the memory access of self-attention without relying on complex additional mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.608, + 0.273, + 0.625 + ], + "angle": 0, + "content": "3. Proposed Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.661 + ], + "angle": 0, + "content": "In this section, we start by representing our preliminary analysis and then describe our proposed network." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.671, + 0.287, + 0.687 + ], + "angle": 0, + "content": "3.1. Preliminary analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Transformers often achieve improved performance by utilizing self-attention's long-range modeling ability and instance-dependent weighting. However, each self-attention incurs excessive memory overhead, resulting in significant latency and memory usage. Interestingly, our preliminary analysis demonstrates that the sophisticated similarity modeling of self-attention \\((A^{avg})\\) and the resulting extracted features \\((F)\\) exhibit a high degree of inter-layer similarities, with averages of \\(89\\%\\) and \\(87\\%\\). This observation not only aligns with recent research [80] but also explains the success of studies that reduced computational cost by sharing the attention map across layers [47, 75]. Inspired by this observation, we investigate improving efficiency without compromising representational power." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.903, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.561, + 0.217, + 0.858, + 0.231 + ], + "angle": 0, + "content": "Figure 3. Visualization of the overall architecture." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.237, + 0.688, + 0.251 + ], + "angle": 0, + "content": "3.2. Overall Structure" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.258, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Our network is composed of four main components, as shown in Figure 3. First, the LR input image \\( I_{LR} \\in \\mathbb{R}^{H \\times W \\times 3} \\) is transformed into a shallow feature \\( F_0 \\in \\mathbb{R}^{H \\times W \\times C} \\) through a \\( 3 \\times 3 \\) convolution \\( (\\mathrm{Conv}_{3 \\times 3}) \\). Here, \\( H \\) and \\( W \\) denote height and width of \\( I_{LR} \\), and \\( C \\) denotes feature dimension. Next, \\( F_0 \\) and a large kernel \\( LK \\in \\mathbb{R}^{13 \\times 13 \\times 16 \\times 16} \\) that is learned end-to-end manner and shared across the entire model are fed into the deep feature extractor \\( H \\), and the resulting deep feature is added back to \\( F_0 \\) to produce the final feature \\( F \\in \\mathbb{R}^{H \\times W \\times C} \\). In parallel, the image-wise skip module \\( S \\) takes \\( I_{LR} \\) as input to generate a skipped feature \\( F_s \\), and the upsampler \\( U \\) utilize both \\( F \\) and \\( F_s \\) as inputs to return the upsampled image \\( I_{SR} \\in \\mathbb{R}^{rH \\times rW \\times 3} \\) where \\( r \\) denotes upscaling factor. The entire process is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.494, + 0.904, + 0.565 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {0} = \\operatorname {C o n v} _ {3 \\times 3} \\left(I _ {L R}\\right), \\\\ F = H \\left(F _ {0}, L K\\right) + F _ {0}, \\tag {1} \\\\ F _ {s} = S \\left(I _ {L R}\\right), \\\\ I _ {S R} = U (F, F _ {s}). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.906, + 0.617 + ], + "angle": 0, + "content": "The \\(H\\) is composed of the \\(N\\) ESCBlocks where \\(N\\) denotes the number of ESCBlocks and a \\(3 \\times 3\\) convolution, and is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.623, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\begin{array}{l} F _ {i} = \\operatorname {E S C B l o c k} _ {i} \\left(F _ {i - 1}, L K\\right), \\quad i = 1, \\dots , N, \\\\ F _ {i} = G _ {i - 1} \\left(F _ {i}\\right) \\end{array} \\tag {2} \\\\ F _ {d} = \\operatorname {C o n v} _ {3 \\times 3} \\left(F _ {N}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.727 + ], + "angle": 0, + "content": "where \\( F_{i-1} \\) indicates the feature input to the \\( i \\)th ESCBlock together with \\( LK \\), and \\( F_i \\) represents the extracted feature. The \\( U \\) and \\( S \\) will be described in the implementation detail section because they will be changed depending on the task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.733, + 0.634, + 0.748 + ], + "angle": 0, + "content": "3.3. ESC Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.755, + 0.906, + 0.785 + ], + "angle": 0, + "content": "The overall process of the ESCBlock is illustrated in Figure 4 and is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.79, + 0.732, + 0.809 + ], + "angle": 0, + "content": "\\[\nF _ {i} ^ {i n} = \\operatorname {C o n v F F N} \\big (\\mathrm {L N} (F _ {i - 1}) \\big),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.811, + 0.76, + 0.829 + ], + "angle": 0, + "content": "\\[\nF _ {i, 0} = F _ {i} ^ {i n} + \\operatorname {S e l f A t t n} \\left(\\mathrm {L N} \\left(F _ {i} ^ {i n}\\right)\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {i, j} = F _ {i, j - 1} + \\operatorname {C o n v A t t n} _ {j} \\left(\\operatorname {C o n v F F N} _ {j} \\left(F _ {i, j - 1}\\right), L K\\right), \\\\ j = 1, \\dots , M, \\\\ F _ {i} = F _ {i - 1} + \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {L N} \\left(F _ {i, M}\\right)\\right). \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24469" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.48, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.295, + 0.438, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Visualization of the proposed ESC block." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.311, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Firstly, the \\( F_{i-1} \\) is converted to the \\( F_{i}^{in} \\) by layer normalization (LN) and convolutional FFN (ConvFFN) [82] with the kernel size of three. Then, LN and naive window self-attention [36] with the window size of \\( 32 \\times 32 \\) are employed in a residual connection to extract \\( F_{i,0} \\). Despite its large window size, we effectively reduce memory overhead by utilizing Flash Attention. Furthermore, since ConvFFN is placed before, the self-attention extracts features considering local information without the complicated QKV projection module. After establishing \\( F_{i,0} \\), the block iterates residual paths composed of ConvFFN and proposed ConvAttn module (ConvAttn) \\( M \\) times. Although our methods do not employ shifted-window self-attention, these convolution-based modules after the self-attention will extract inter-window features. Finally, the result of the last iteration \\( F_{i,M} \\) go through LN and \\( \\mathrm{Conv}_{3 \\times 3} \\), then is added to \\( F_{i-1} \\), resulting in \\( F_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Convolutional Attention Module The ConvAttn is designed to emulate self-attention's two key advantages: capturing long-range dependencies and instance-dependent weighting. For this purpose, ConvAttn leverages two types of convolution formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.649, + 0.482, + 0.743 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} F _ {i, j} ^ {a t t}, F _ {i, j} ^ {i d t} = \\mathrm {S p l i t} _ {1 6: C - 1 6} (F _ {i, j} ^ {\\mathrm {C F}}), \\\\ D K _ {i, j} = \\operatorname {C o n v} _ {1 \\times 1} ^ {u p} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} ^ {\\text {d o w n}} \\left(\\operatorname {G A P} \\left(F _ {i, j} ^ {\\text {a t t}}\\right)\\right)\\right)\\right), \\\\ F _ {i, j} ^ {r e s} = (F _ {i, j} ^ {a t t} \\circledast D K _ {i, j}) + (F _ {i, j} ^ {a t t} \\circledast L K), \\\\ F _ {i, j} ^ {f u s e} = \\operatorname {C o n v} _ {1 \\times 1} ^ {f u s e} \\left(\\operatorname {C o n c a t} \\left(F _ {i, j} ^ {r e s}, F _ {i, j} ^ {i d t}\\right)\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(F_{i,j}^{\\mathrm{CF}}\\) denotes the features extracted by \\(\\mathrm{ConvFFN}_{i,j}\\) before. We first split the \\(F_{i,j}^{\\mathrm{CF}}\\) channel-wise, resulting in \\(F_{i,j}^{att} \\in \\mathbb{R}^{H \\times W \\times 16}\\), which consists of the 16 channels in front, and \\(F_{i,j}^{idt} \\in \\mathbb{R}^{H \\times W \\times C - 16}\\), which contains the remaining part. Two types of convolutions are only applied within \\(F_{i,j}^{att}\\) to reduce memory access cost, unlike the methods that leverage large kernels considering only computational complexity [52, 62, 63]. Next, dynamic depth-wise kernel \\(DK_{i,j} \\in \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16}\\) [22] is created by two layers of \\(\\mathrm{Conv}_{1 \\times 1}\\) with global average pooling (GAP) and gaussian" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.323 + ], + "angle": 0, + "content": "error linear unit activation \\((\\phi)\\). Since enlarging \\(DK\\)'s kernel size increases the parameter size multiplied by the hidden dimension \\((h)\\) of the kernel estimator, we fix its kernel size to \\(3\\times 3\\). Then both \\(DK_{i,j}\\) and \\(LK\\) are convolved on \\(F_{i,j}^{att}\\) and sum up those, resulting convolved feature \\(F_{i,j}^{res}\\). Based on our preliminary analysis, we hypothesize that long-range interactions do not significantly vary across layers. Therefore, we share \\(LK\\) across the entire ESCBlocks and ConvAttns, mitigating parameter growth and optimization difficulty [6, 38]. Finally, \\(F_{i,j}^{res}\\) and \\(F_{i,j}^{idt}\\) are concatenated channel-wise, followed by the last convolution \\(\\mathrm{Conv}_{1\\times 1}^{fuse}\\) to fuse them, resulting in fused feature \\(F_{i,j}^{use}\\). In this way, both the local features extracted by ConvFFN and the long-range features extracted by ConvAttn can be leveraged to capture multi-scale information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.324, + 0.907, + 0.46 + ], + "angle": 0, + "content": "While \\(DK_{i,j}\\) can be merged into \\(LK\\) by padding both spatial and channel dimensions to reduce FLOPs, we do not because it slows down latency. On the other hand, in situations where reducing FLOPs and parameter size is crucial rather than reducing latency, \\(LK\\) can be decomposed depthwise separable manner [25] into point-wise large kernel filter \\(LK^{\\mathrm{c}}\\in \\mathbb{R}^{1\\times 1\\times 16\\times 16}\\) and depth-wise large kernel filter \\(LK^{\\mathrm{s}}\\in \\mathbb{R}^{13\\times 13\\times 1\\times 16}\\). This decomposition redefines line 3 in Equation 4 as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.47, + 0.907, + 0.49 + ], + "angle": 0, + "content": "\\[\nF _ {i, j} ^ {r e s} = \\left(F _ {i, j} ^ {a t t} * L K ^ {c}\\right) * \\left(Z P \\left(D K _ {i, j}\\right) + L K ^ {s}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.904, + 0.528 + ], + "angle": 0, + "content": "where \\( ZP(\\cdot): \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16} \\mapsto \\mathbb{R}^{13 \\times 13 \\times 1 \\times 16} \\) denotes spatial zero padding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.906, + 0.709 + ], + "angle": 0, + "content": "In summary, ConvAttn efficiently emulates self-attention's two advantages by leveraging a shared \\(LK\\) and layer-specific \\(DKs\\). In contrast to Transformers that share attention maps across several layers [47, 75, 80], which still require a full materialized attention map for every layer, our module only leverages a single shared \\(LK\\), significantly reducing memory footprints. Moreover, unlike recent CNNs [52, 62, 63, 81] that use large kernel convolutions and optionally employ pixel-wise attention to inject input dependency, we employ layer-specific \\(DKs\\), emulating self-attention's input-dependent weight more exactly in a memory-efficient manner." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.722, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Table 1. Comparison of latency and memory usage between naive self-attention and our Flash Attention (FA) [10, 14]. WS and #Pixels denote the window size of self-attention and the corresponding number of pixels. \\(X\\) refers to the input tensor." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.782, + 0.907, + 0.899 + ], + "angle": 0, + "content": "
WS(#Pixels)TypeX ∈ R64×64×64X ∈ R288×288×64
Latency(ms)Memory Usage(mb)Latency(ms)Memory Usage(mb)
16×16(256)w/o FA2.0464.9758
w/ FA0.5 (4×)11 (4.2×)1.3 (3.8×)223 (3.4×)
24×24(576)w/o FA7.611114.51575
w/ FA0.5 (15.2×)14 (7.9×)2.0 (7.3×)223 (7.1×)
32×32(1024)w/o FA27.315744.72717
w/ FA0.5 (54.6×)11 (14.3×)2.8 (16.0×)223 (12.2×)
" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24470" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparisons of classic SR methods trained on the DIV2K dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. † denotes that we re-calculate the statistics of the indicated methods. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + }, + { + "type": "table", + "bbox": [ + 0.09, + 0.133, + 0.909, + 0.675 + ], + "angle": 0, + "content": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)Set5Set14PSNR / SSIM B100Urban100Manga109
SwinIR-lt [36]1409.81287244.291038.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-lt [75]94.5887203.162138.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
OmniSR [56]120.31031194.577238.22/0.961333.98/0.921032.36/0.902033.05/0.936339.28/0.9784
SRFormer-lt [82]1456.31184236.385338.23/0.961333.94/0.920932.36/0.901932.91/0.935339.28/0.9785
ATD-lt [73]733.52839380.075338.29/0.961634.10/0.921732.39/0.902333.27/0.937539.52/0.9789
HiT-SRF [76]268.11804226.584738.26/0.961534.01/0.921432.37/0.902333.13/0.937239.47/0.9787
ASID-D8 [47] ×2131.2999190.5†73238.32/0.961834.24/0.923232.40/0.902833.35/0.9387-/-
MambaIR-lt [20]277.11695334.290538.13/0.961033.95/0.920832.31/0.901332.85/0.934939.20/0.9782
MambaIRV2-lt [21]580.42824286.377438.26/0.961534.09/0.922132.36/0.901933.26/0.937839.35/0.9785
RDN [78]279.320585096.22212338.24/0.961434.01/0.921232.34/0.901732.89/0.935339.18/0.9780
RCAN [77]299.36263529.71544538.27/0.961434.12/0.921632.41/0.902733.34/0.938439.44/0.9786
ESC-FP (Ours)94.3627239.852438.27/0.961734.09/0.921932.37/0.902233.22/0.937539.40/0.9784
ESC-lt (Ours)73.2830359.460338.24/0.961533.98/0.921132.35/0.902033.05/0.936339.33/0.9786
ESC (Ours)120.9831592.094738.35/0.961934.11/0.922332.41/0.902733.46/0.939539.54/0.9790
SwinIR-lt [36] ×3331.7596110.891834.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-lt [75]32.539990.162934.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
OmniSR [56]41.247688.478034.70/0.929430.57/0.846929.28/0.809428.84/0.865634.22/0.9487
SRFormer-lt [82]530.5537105.486134.67/0.929630.57/0.846929.26/0.809928.81/0.865534.19/0.9489
ATD-lt [73]274.41258168.076034.74/0.930030.68/0.848529.32/0.810929.17/0.870934.60/0.9506
HiT-SRF [76]124.91464101.685534.75/0.930030.61/0.847529.29/0.810628.99/0.868734.53/0.9502
ASID-D8 [47] ×361.946086.4†73934.84/0.930730.66/0.849129.32/0.811929.08/0.8706-/-
MambaIR-lt [20]109.3760148.591334.63/0.928830.54/0.845929.23/0.808428.70/0.863134.12/0.9479
MambaIRV2-lt [21]259.01250126.778134.71/0.929830.68/0.848329.26/0.809829.01/0.868934.41/0.9497
RDN [78]146.59852281.22230834.71/0.929630.57/0.846829.26/0.809328.80/0.865334.13/0.9484
RCAN [77]85.15601586.11562934.74/0.929930.65/0.848229.32/0.811129.09/0.870234.44/0.9499
ESC-FP (Ours)34.8291110.053034.72/0.930030.67/0.848329.30/0.810729.12/0.870634.56/0.8706
ESC-lt (Ours)25.2384162.861234.61/0.929530.52/0.847529.26/0.810228.93/0.867934.33/0.9495
ESC (Ours)41.4385267.695534.84/0.930830.74/0.849329.34/0.811829.28/0.873934.66/0.9512
SwinIR-lt [36] ×4222.935163.693032.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-lt [75]18.024154.164032.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
OmniSR [56]22.527350.979232.49/0.898828.78/0.785927.71/0.741526.64/0.801831.02/0.9151
SRFormer-lt [82]287.232962.887332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
ATD-lt [73]189.7753100.176932.63/0.899828.89/0.788627.79/0.744026.97/0.810731.48/0.9198
HiT-SRF [76]82.1133158.086632.55/0.899928.87/0.788027.75/0.743226.80/0.806931.26/0.9171
ASID-D8 [47] ×461.826549.6†74832.57/0.899028.89/0.789827.78/0.744926.89/0.8096-/-
MambaIR-lt [20]55.843884.692432.42/0.897728.74/0.784727.68/0.740026.52/0.798330.94/0.9135
MambaIRV2-lt [21]153.474875.679032.51/0.899228.84/0.787827.75/0.742626.82/0.807931.24/0.9182
RDN [78]66.07911309.22227132.47/0.899028.81/0.787127.72/0.741926.61/0.802831.00/0.9151
RCAN [78]52.2540917.61559232.63/0.900228.87/0.788927.77/0.743626.82/0.808731.22/0.9173
ESC-FP (Ours)21.715860.853932.56/0.889728.87/0.788927.75/0.743526.90/0.809831.40/0.9192
ESC-lt (Ours)13.321391.062432.52/0.899528.87/0.787827.72/0.742326.76/0.805831.26/0.9173
ESC (Ours)21.9215149.296832.68/0.901128.93/0.790227.80/0.744727.07/0.814431.54/0.9207
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.223, + 0.699 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.483, + 0.771 + ], + "angle": 0, + "content": "This section demonstrates the effectiveness of our flash attention and our extensive experiments on various tasks. For implementation details for proposed networks, please refer to Section 7 in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.786, + 0.37, + 0.801 + ], + "angle": 0, + "content": "4.1. Effectiveness of Flash Attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We address the excessive memory overhead of self-attention by leveraging Flash Attention [10, 14]. Implementation details for Flash Attention are described in Section 6 in the supplementary. As shown in Table 1, our self-attention implementation accelerated by Flash Attention significantly reduces the latency and memory usage of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.745 + ], + "angle": 0, + "content": "self-attention at the window size of 32 by \\(16 \\times\\) and \\(12.2 \\times\\), respectively. We do not consider window sizes above 32, as it may cause training instability due to the necessity of padding, since the training patch size is \\(64 \\times 64\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.757, + 0.693, + 0.771 + ], + "angle": 0, + "content": "4.2. Classic SR Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "To evaluate our proposed networks on the classic SR task, we leverage five commonly used datasets (Set5 [3], Set14 [70], B100 [42], Urban100 [26], Manga109 [43]). We measure peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) on the Y channel after cropping the boundary corresponding to the upscaling factor and converting them into YCbCr space. All statistics are measured by reconstructing an HD im" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24471" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.235, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.179, + 0.21, + 0.189 + ], + "angle": 0, + "content": "Urban100 - img017" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.089, + 0.318, + 0.122 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.123, + 0.306, + 0.13 + ], + "angle": 0, + "content": "SRFormer-lt" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.131, + 0.289, + 0.138 + ], + "angle": 0, + "content": "24.62" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.139, + 0.318, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.171, + 0.293, + 0.179 + ], + "angle": 0, + "content": "ATD-lt" + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.18, + 0.289, + 0.187 + ], + "angle": 0, + "content": "24.45" + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.089, + 0.4, + 0.122 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.339, + 0.123, + 0.38, + 0.129 + ], + "angle": 0, + "content": "HiT-SRF" + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.13, + 0.371, + 0.138 + ], + "angle": 0, + "content": "25.21" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.139, + 0.4, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.171, + 0.393, + 0.179 + ], + "angle": 0, + "content": "MambaIRV2-lt" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.18, + 0.371, + 0.187 + ], + "angle": 0, + "content": "25.23" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.089, + 0.482, + 0.122 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.123, + 0.462, + 0.129 + ], + "angle": 0, + "content": "HR Crop" + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.13, + 0.455, + 0.138 + ], + "angle": 0, + "content": "PSNR" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.139, + 0.482, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.417, + 0.171, + 0.468, + 0.179 + ], + "angle": 0, + "content": "ESC (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.18, + 0.454, + 0.187 + ], + "angle": 0, + "content": "26.50" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.189, + 0.235, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.277, + 0.209, + 0.286 + ], + "angle": 0, + "content": "Urban100 - img059" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.188, + 0.318, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.221, + 0.305, + 0.228 + ], + "angle": 0, + "content": "SRFormer-lt" + }, + { + "type": "image_caption", + "bbox": [ + 0.248, + 0.229, + 0.29, + 0.236 + ], + "angle": 0, + "content": "26.56" + }, + { + "type": "image", + "bbox": [ + 0.235, + 0.237, + 0.318, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.27, + 0.291, + 0.277 + ], + "angle": 0, + "content": "ATD-1t" + }, + { + "type": "image_caption", + "bbox": [ + 0.261, + 0.278, + 0.29, + 0.286 + ], + "angle": 0, + "content": "28.90" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.188, + 0.4, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.221, + 0.376, + 0.228 + ], + "angle": 0, + "content": "HiT-SRI" + }, + { + "type": "image_caption", + "bbox": [ + 0.345, + 0.229, + 0.371, + 0.236 + ], + "angle": 0, + "content": "2671" + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.237, + 0.4, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.27, + 0.393, + 0.277 + ], + "angle": 0, + "content": "MambaIRV2-lt" + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.278, + 0.37, + 0.286 + ], + "angle": 0, + "content": "26.41" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.188, + 0.482, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.221, + 0.462, + 0.229 + ], + "angle": 0, + "content": "HR Crops" + }, + { + "type": "image_caption", + "bbox": [ + 0.425, + 0.23, + 0.455, + 0.237 + ], + "angle": 0, + "content": "PSNR" + }, + { + "type": "image", + "bbox": [ + 0.401, + 0.238, + 0.482, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.27, + 0.468, + 0.277 + ], + "angle": 0, + "content": "ESC (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.278, + 0.454, + 0.286 + ], + "angle": 0, + "content": "37.75" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.291, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 5. Visual comparisons for classic SR \\((\\times 2)\\). The best result on PSNR is bolded." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.484, + 0.43 + ], + "angle": 0, + "content": "age (1280×720) image using an RTX4090 GPU at FP32 precision 100 times. Memory usage is calculated by torch.cuda.max_memory_allocated function and FLOPs are measured by fvcore library. For the classic SR task, we present three variants: ESC, ESC-light, and ESC-FP. ESC and ESC-light prioritize reducing latency, while ESC-FP focuses on reducing FLOPs and parameter size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.437, + 0.435, + 0.451 + ], + "angle": 0, + "content": "4.2.1. Quantitative Results on the DIV2K dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.456, + 0.484, + 0.729 + ], + "angle": 0, + "content": "We compare our methods with various representative lightweight SR methods, including CNN-based models [77, 78], Transformers [36, 47, 56, 73, 75, 76, 82], and SSMs [20, 21] trained on the DIV2K dataset [1]. Notably, as shown in Table 2, our ESC achieves a notable 0.1dB PSNR improvement on the Urban \\(100 \\times 4\\) dataset, while being \\(8.7 \\times\\) faster than ATD-light. Moreover, ESC-light is \\(1.4 \\times\\) faster than ELAN-light, exhibiting an impressive performance improvement of 0.34dB PSNR in the Manga109 \\(\\times 4\\) dataset. These results exhibit that state-of-the-art performance can be achieved even when most self-attention are replaced with convolutions, suggesting that replacing self-attention with convolutions does not compromise representational power. Furthermore, ESC-FP outperforms MambaIRV2-light over 0.18dB on the Urban \\(100 \\times 4\\) dataset while reducing FLOPs and parameters by \\(20\\%\\) and \\(31\\%\\), respectively. This result underscores that our design is effective even when reducing FLOPs and parameter size." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.392, + 0.749 + ], + "angle": 0, + "content": "4.2.2. Visual Results on the DIV2K dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.754, + 0.483, + 0.845 + ], + "angle": 0, + "content": "We also visually compare our ESC with representative lightweight SR methods, including SRFormer-light [82], HiT-SRF [76], ATD-light [73], and MambaIRV2-light [21]. As shown in Figure 5, our ESC effectively restores details and achieves a high PSNR, demonstrating that our network also produces visually pleasant results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.243, + 0.867 + ], + "angle": 0, + "content": "4.2.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conduct an ablation study to justify our network's composition, as shown in Table 3. First, we introduce two vari-" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.906, + 0.158 + ], + "angle": 0, + "content": "Table 3. Ablation studies on our ESC network. All comparisons are designed to have similar latencies to ESC. \\(LK\\), \\(DK\\), and WS denote large kernel, dynamic kernel, and window size, respectively. All statistics and metrics are measured as detailed in Section 4.2." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.164, + 0.907, + 0.294 + ], + "angle": 0, + "content": "
CasesLatency (ms)Memory (mb)#FLOPs (G)#Params (K)Set5×2Urban100×2
Only Self-attention128.2830456.673738.27/0.961733.23/0.9378
Only ConvAttns126.3514962.4121038.18/0.961332.91/0.9353
9×9 LK117.1831462.292438.33/0.961833.42/0.9392
17×17 LK126.6832768.997738.32/0.961833.40/0.9390
w/o LK Share and DK119.7835591.1195038.32/0.961833.37/0.9389
w/o LK Share120.9835592.0198538.32/0.961933.43/0.9392
w/o DK119.7835591.191138.31/0.961833.36/0.9388
WS16 and more layers118.0807603.6105438.24/0.961433.05/0.9360
WS24 and more layers116.5801583.599538.33/0.961833.28/0.9383
ESC (Ours)120.9831592.094738.35/0.961933.46/0.9395
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.697 + ], + "angle": 0, + "content": "ants, one that uses only the self-attention module and another that uses only the ConvAttn. The number of blocks and layers is adjusted to match the latency on par with the ESC. Both variants exhibit a performance drop compared to ESC, suggesting that leveraging both self-attention and ConvAttn offers the best performance-efficiency trade-off. Next, we vary the \\(LK\\)'s kernel size to \\(9 \\times 9\\) and \\(17 \\times 17\\) to determine the optimal size. In both cases, we observe a drop in performance, confirming that a kernel size of \\(13 \\times 13\\) is optimal. Then, we evaluate performance changes by altering the convolution type used in the ConvAttn. First, we observe performance drops when neither \\(LK\\) sharing nor \\(DK\\) is used. Then, we add \\(DK\\) without sharing \\(LK\\) and observe some performance improvement, though it still falls short of ESC. This result suggests that addressing the optimization difficulty that comes from the massive parameter size is more critical than the layer-specific long-range modeling. Finally, we leverage only \\(LK\\) sharing without \\(DK\\) and observe significant performance drops, affirming that harmonizing both \\(LK\\) and \\(DK\\) is crucial. For the next step, we investigate the effect of the self-attention window size by setting it to 16 and 24 and adding extra layers. In every case, we observe a performance drop, which confirms that enlarging the window size to enhance self-attention is effective even when only a single self-attention is leveraged in each block." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.708, + 0.771, + 0.723 + ], + "angle": 0, + "content": "4.2.4. Analysis on ConvAttn Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This section verifies that the proposed ConvAttn operates similarly to self-attention. Self-attention is known to extract shape-biased structural features [12, 48], unlike \\(3 \\times 3\\) convolution, which mainly extracts local texture and edges. Indeed, as shown in Figure 6 (a), \\(F^{idt}\\) focuses on local textures and edges while \\(F^{att}\\) concentrates on structural components. Next, to verify each convolution's receptive field, we visualize positive gradients \\(\\bar{G}^{+} = \\frac{1}{NMC} \\sum_{p=1}^{NMC} \\max \\left(0, \\frac{\\partial I_{i,j}^{SR}}{\\partial F_p}\\right)\\) propagated from the output image's center pixel \\((i,j)\\) to the inputs feature \\((F)\\) of both \\(DK\\) and \\(LK\\) across \\(N\\) blocks, \\(M\\) modules, and \\(C\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24472" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.907, + 0.131 + ], + "angle": 0, + "content": "Table 4. Comparisons of classic SR methods trained on the large-scale DFLIP dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. \\(\\S\\) denotes that we train the indicated methods following their descriptions. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.133, + 0.907, + 0.332 + ], + "angle": 0, + "content": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)PSNR / SSIM
Set5Set14B100Urban100Manga109
SRFormer-lt$ [82]×21838.11184236.385338.24/0.961534.13/0.921832.42/0.902633.37/0.938639.36/0.9787
ATD-lt$ [73]733.52839380.075338.29/0.961634.30/0.923032.43/0.902733.62/0.940139.60/0.9791
HiT-SRF$ [76]268.11804226.584738.31/0.961634.31/0.923032.45/0.903133.58/0.940439.69/0.9793
ESC (Ours)120.9831592.094738.34/0.961834.42/0.923532.50/0.903633.86/0.942439.73/0.9795
SRFormer-lt$ [82]×3668.3537105.486134.67/0.929730.75/0.848429.30/0.810829.10/0.870134.26/0.9498
ATD-lt$ [73]274.41258168.076034.71/0.930030.77/0.849329.33/0.811629.42/0.874334.61/0.9509
HiT-SRF$ [76]124.91464101.685534.69/0.929830.81/0.849329.32/0.811529.28/0.872934.72/0.9511
ESC (Ours)41.4385267.695534.85/0.931230.97/0.851129.41/0.813529.70/0.879934.94/0.9525
SRFormer-lt$ [82]×4327.832962.887332.49/0.899328.89/0.788727.76/0.742926.90/0.808631.25/0.9189
ATD-lt$ [73]189.7753100.176932.52/0.899528.93/0.789627.79/0.744327.18/0.815031.47/0.9208
HiT-SRF$ [76]82.1133158.086632.55/0.899728.96/0.789727.77/0.744327.07/0.813031.59/0.9208
ESC (Ours)21.9215149.296832.79/0.902529.06/0.792727.85/0.746627.45/0.822931.87/0.9239
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.344, + 0.2, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.424, + 0.151, + 0.432 + ], + "angle": 0, + "content": "\\(F^{idt}\\)" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.344, + 0.322, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.247, + 0.424, + 0.3, + 0.432 + ], + "angle": 0, + "content": "\\(\\bar{G}_{DK}^{+}\\) (DI: 48.77)" + }, + { + "type": "image", + "bbox": [ + 0.354, + 0.344, + 0.478, + 0.424 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.395, + 0.424, + 0.411, + 0.434 + ], + "angle": 0, + "content": "\\(S_{DK}\\)" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.433, + 0.2, + 0.514 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.523, + 0.212, + 0.53 + ], + "angle": 0, + "content": "(a) Feature Map Visualization" + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.434, + 0.349, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.515, + 0.3, + 0.522 + ], + "angle": 0, + "content": "\\(G_{LK}^{+}\\) (DI: 62.23)" + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.523, + 0.333, + 0.53 + ], + "angle": 0, + "content": "(b) Gradient Visualization" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.434, + 0.478, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.36, + 0.523, + 0.461, + 0.53 + ], + "angle": 0, + "content": "(c) Similarity Visualization" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.533, + 0.483, + 0.588 + ], + "angle": 0, + "content": "Figure 6. Visualized feature map, gradients, and input-wise similarity. \\(F\\) is averaged across entire ConvAttn while processing the Lenna image in Set14. \\(\\tilde{G}^{+}\\) and \\(S\\) are calculated on the Urban100, and 10 samples in front of the Urban100, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.484, + 0.707 + ], + "angle": 0, + "content": "channels, and corresponding diffusion index (DI). Figure 6 (b) clearly demonstrates that \\(LK\\) extracts more long-range features than \\(DK\\). Lastly, we visualize input-wise similarity \\((S)\\) between average-pooled features extracted by \\(LK\\) and \\(DK\\). As demonstrated in Figure 6 (c), \\(S_{DK}\\) exhibits higher input-dependent diversity than \\(S_{LK}\\), confirming that \\(DK\\) extracts more input-dependent features than \\(LK\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.716, + 0.296, + 0.731 + ], + "angle": 0, + "content": "4.2.5. Local Attribution Map" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We use the Local Attribution Map (LAM) [19] to verify that our networks can extract long-range dependencies despite using less self-attention. We compare our ESC with studies of various self-attention implementations that aim to reduce FLOPs while expanding the receptive field, such as global-level attention [56], permuted self-attention [82], and spatial-channel correlation [76]. Notably, as described in Figure 7, our ESC references the widest range of pixels to restore the red bounding box along with the highest diffusion index. This result exhibits that our network effectively captures long-range dependencies even though most self" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.343, + 0.643, + 0.44 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.44, + 0.643, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.553, + 0.62, + 0.567 + ], + "angle": 0, + "content": "Urban100 - img063x4 Diffusion Index" + }, + { + "type": "image", + "bbox": [ + 0.643, + 0.343, + 0.748, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.392, + 0.715, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.44, + 0.696, + 0.447 + ], + "angle": 0, + "content": "OmniSR" + }, + { + "type": "image", + "bbox": [ + 0.721, + 0.343, + 0.777, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.392, + 0.777, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.44, + 0.769, + 0.447 + ], + "angle": 0, + "content": "SRFormer-It" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.343, + 0.841, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.78, + 0.392, + 0.842, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.799, + 0.44, + 0.824, + 0.447 + ], + "angle": 0, + "content": "iT-SRF" + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.343, + 0.905, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.842, + 0.392, + 0.905, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.852, + 0.44, + 0.895, + 0.447 + ], + "angle": 0, + "content": "ESC (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.57, + 0.907, + 0.611 + ], + "angle": 0, + "content": "Figure 7. Comparisons of Local Attribution Map (LAM) [19] and the corresponding area of contribution and diffusion index. The best result on the diffusion index is bolded." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.633, + 0.907, + 0.678 + ], + "angle": 0, + "content": " attentions are replaced by the ConvAttn module, suggesting that our ConvAttn module contributes to retaining the large receptive field of Transformers." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.857, + 0.712 + ], + "angle": 0, + "content": "4.2.6. Quantitative Results on the DFLIP dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Transformers exhibit significant performance improvements when trained on large-scale datasets [5, 15]. To verify our model also retains this benefit, we train and evaluate various representative Transformers – including SRFormerlight [82], ATD-light [73], and HiT-SRF [76] – alongside our ESC on a large dataset. For this purpose, we leverage DIV2K+Flickr2K+LSDIR+DiverSeg-IP (DFLIP) [1, 34, 45, 54] dataset. As illustrated in Table 4, our approach outperforms the other methods by a substantial margin across all scales and evaluation datasets. These results indicate that replacing self-attention with ConvAttn does not sacrifice the data scaling capability of Transformers." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24473" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.277, + 0.108 + ], + "angle": 0, + "content": "4.3. Arbitrary-Scale SR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.22 + ], + "angle": 0, + "content": "To evaluate our proposed networks on arbitrary-scale SR tasks, we leverage the DIV2K-validation [1] dataset. Following previous research [8, 32], we assess PSNR in the RGB channel after cropping the boundaries according to the upscaling factor plus six. For arbitrary-scale SR tasks, we change \\(U\\) of our network and representative lightweight SR transforems [73, 76] into LTE [32]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.228, + 0.281, + 0.243 + ], + "angle": 0, + "content": "4.3.1. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.484, + 0.444 + ], + "angle": 0, + "content": "To demonstrate the outstanding representational capability of the proposed network, we train and evaluate the network on an arbitrary-scale SR task, which requires processing various upscaling factors with a single network. For quantitative comparison, we leverage commonly used RDN [78] and additional representative Transformers [73, 76] that we train. As described in Tables 5, our network continually achieves high performance at the scale both seen and unseen during training, even achieving a \\(0.09\\mathrm{dB}\\) improvement over \\(\\mathrm{RDN + LTE}\\) at \\(\\times 2\\). This result underscores our ESC's exceptional representational capability. For quantitative results measured on other datasets, including Set5, Set14, B100, and Urban100, refer to Section 8 in the supplementary." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.458, + 0.483, + 0.485 + ], + "angle": 0, + "content": "Table 5. Quantitative comparison on arbitrary-scale SR task employing LTE [32] as upsampler. The best result on PSNR is bolded." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.487, + 0.485, + 0.564 + ], + "angle": 0, + "content": "
MethodsSeenUnseen
×2×3×4×6×12×18×24×30
RDN+LTE [78]35.0431.3229.3327.0423.9522.4021.3620.64
ATD-lt+LTE§ [73]34.9831.2529.2727.0123.9422.3821.3720.64
HiT-SRF+LTE§ [76]35.1031.3529.3527.0523.9422.3921.3620.64
ESC+LTE (Ours)35.1331.3829.3727.0823.9722.4021.3820.64
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.236, + 0.59 + ], + "angle": 0, + "content": "4.3.2. Visual Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.687 + ], + "angle": 0, + "content": "We also visually compare our network with RDN, HiT-SRF, and ATD-light. As shown in Figure 8, Our network effectively restores fine details that other methods fail on both \\(\\times 4\\) and \\(\\times 12\\) scales. This result demonstrates that our network also excels at restoring visually present results on both in-distribution and out-of-distribution scales." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.355, + 0.714 + ], + "angle": 0, + "content": "4.4. Real-World Super-Resolution" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We further test our network on real-world SR tasks that aim to restore images with unknown degradations. To this end, we introduce a variant called ESC-Real and compare them to representative real-world SR methods, including RealESRGAN+ [59], SwinIR-Real [36], and DASR [61]. For our evaluation, we use the commonly employed RealSRSet dataset [72]. As shown in Table 9, our ESC-Real recovers more fine-grained details than the other methods while significantly reducing unpleasant artifacts in both graphic and natural images. This result demonstrates that our proposed methods are effective even when dealing with complex unknown degradations, underscoring our methods'" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.091, + 0.707, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.571, + 0.191, + 0.664, + 0.2 + ], + "angle": 0, + "content": "Urban100 - img011" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.2, + 0.705, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.546, + 0.236, + 0.696, + 0.244 + ], + "angle": 0, + "content": "RDN/17.71 HiT-SRF/17.56" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.245, + 0.705, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.542, + 0.281, + 0.704, + 0.29 + ], + "angle": 0, + "content": "ATD-lt / 17.91 ESC (Ours) / 20.70" + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.293, + 0.653, + 0.305 + ], + "angle": 0, + "content": "(a) Scale \\(\\times 4\\)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.091, + 0.891, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.755, + 0.191, + 0.854, + 0.2 + ], + "angle": 0, + "content": "DIV2K-valid - 0845" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.201, + 0.805, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.731, + 0.237, + 0.884, + 0.245 + ], + "angle": 0, + "content": "RDN/16.25 HiT-SRF/17.28" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.245, + 0.803, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.726, + 0.281, + 0.79, + 0.29 + ], + "angle": 0, + "content": "ATD-lt / 16.77" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.245, + 0.892, + 0.28 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.281, + 0.888, + 0.29 + ], + "angle": 0, + "content": "ESC (Ours) / 17.82" + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.293, + 0.842, + 0.305 + ], + "angle": 0, + "content": "(b) Scale \\(\\times 12\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.31, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 8. Visual comparisons on arbitrary-scale SR for both indistribution \\((\\times 4)\\) and out-of-distribution \\((\\times 12)\\) scales. The best result on PSNR is bolded." + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.358, + 0.903, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.481, + 0.769, + 0.491 + ], + "angle": 0, + "content": "(a) RealSRSet - 0030" + }, + { + "type": "image", + "bbox": [ + 0.519, + 0.496, + 0.903, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.619, + 0.767, + 0.631 + ], + "angle": 0, + "content": "(b) RealSRSet - dog" + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.648, + 0.884, + 0.662 + ], + "angle": 0, + "content": "Figure 9. Visual comparisons on real-world SR \\((\\times 4)\\) tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.717 + ], + "angle": 0, + "content": "outstanding representational capability. Quantitative results on real-world SR tasks are presented in Section 9 in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.634, + 0.74 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we propose ESC, a novel lightweight SR network designed to mitigate the excessive memory overhead of Transformers. Our extensive experiments demonstrate that most self-attention layers can be effectively replaced with carefully designed convolutions to improve efficiency. Moreover, we introduce Flash Attention into lightweight SR tasks, significantly enhancing performance without complicated modifications to self-attention. As a result, we fully exploit the advantages of Transformers by prioritizing improvements in their efficiency." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24474" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400]." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.149, + 0.188, + 0.165 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.174, + 0.482, + 0.215 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.482, + 0.284 + ], + "angle": 0, + "content": "[2] Yang Ai, Xiaoqiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. NeurIPS, 37:55443-55469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.286, + 0.482, + 0.34 + ], + "angle": 0, + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, pages 3086-3095, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.398, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[6] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In CVPR, pages 5557-5567, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.51, + 0.482, + 0.563 + ], + "angle": 0, + "content": "[7] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In CVPR, pages 22367-22377, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.566, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, pages 8628-8638, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.608, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[9] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NeurIPS, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.65, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[10] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 1, 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.706, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[11] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 44(5):2567-2581, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, pages 11963-11975, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE TPAMI, 38(2):295-307, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A programming model for generating optimized attention kernels. arXiv preprint arXiv:2412.05496, 2024. 2, 3, 4, 5, 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.174, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[16] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In CVPRW, pages 2494-2502, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. *ICLR*, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[19] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9199-9208, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.402 + ], + "angle": 0, + "content": "[20] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.446 + ], + "angle": 0, + "content": "[21] Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. Mambairv2: Attentive state space restoration. CVPR, 2025. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[22] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between local attention and dynamic depth-wise convolution. *ICLR*, 2022, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[23] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In ECCV, pages 289-305. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[25] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.674, + 0.905, + 0.715 + ], + "angle": 0, + "content": "[26] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[28] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, pages 1637-1645, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[29] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: a method for stochastic optimization. In ICLR, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representa" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24475" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "tions revisited. In ICML, pages 3519-3529. PMLR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In CVPR, pages 1929-1938, 2022. 8, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[33] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In IJCAI, pages 1089-1097, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[34] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPRW, pages 1775-1787, 2023. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[35] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image superresolution. In CVPR, pages 3867-3876, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[36] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.359, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[37] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 136-144, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.401, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[38] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. ICLR, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[39] Yong Liu, Hang Dong, Boyang Liang, Songwei Liu, Qingji Dong, Kai Chen, Fangmin Chen, Lean Fu, and Fei Wang. Unfolding once is enough: A deployment-friendly transformer unit for super-resolution. In ACMMM, pages 7952-7960, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[42] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423. IEEE, 2001. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[43] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.749, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[44] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 20(3): 209-212, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.859 + ], + "angle": 0, + "content": "[45] Go Ohtani, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka, and Yoshimitsu Aoki. Rethinking image super-resolution from training data perspectives. In ECCV, pages 19-36. Springer, 2024. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[46] Karam Park and Nam Ik Cho. Partial filter-sharing: Improved parameter-sharing method for single image super-resolution networks. In WACV, pages 2653-2663. IEEE, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[47] Karam Park, Jae Woong Soh, and Nam Ik Cho. Efficient attention-sharing information distillation transformer for lightweight single image super-resolution. In AAAI, 2025. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[48] Namuk Park and Songkuk Kim. How do vision transformers work? In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[49] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[50] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.37 + ], + "angle": 0, + "content": "[51] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[52] Long Sun, Jinshan Pan, and Jinhui Tang. Shufflemixer: An efficient convnet for image super-resolution. NeurIPS, 35: 17314-17326, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[53] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, pages 3147-3155, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[54] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 114-125, 2017. 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[56] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Liu jinfan. Omni aggregation networks for lightweight image super-resolution. In CVPR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[57] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, pages 2555-2563, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[58] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, pages 606-615, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[59] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, pages 1905-1914, 2021. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[60] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, pages 101-117. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[61] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In CVPR, pages 13385-13394, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[62] Gang Wu, Junjun Jiang, Junpeng Jiang, and Xianming Liu. Transforming image super-resolution: a convformer-based efficient approach. IEEE Transactions on Image Processing, 2024. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24476" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[63] Chengxing Xie, Xiaoming Zhang, Linze Li, Haiteng Meng, Tianlin Zhang, Tianrui Li, and Xiaole Zhao. Large kernel distillation network for efficient single image superresolution. In CVPRW, pages 1283-1292, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[64] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, pages 1191-1200, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[65] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.333 + ], + "angle": 0, + "content": "[66] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, pages 15703-15712, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[67] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In WACV, pages 4956-4965, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[68] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5756-5767, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: efficient attention transformer for high-resolution image restoration. In CVPR, pages 5728-5739, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[70] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[72] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, pages 4791-4800, 2021. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[73] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In CVPR, pages 2856-2865, 2024. 2, 5, 6, 7, 8, 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[75] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In ECCV, pages 649-667. Springer, 2022. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient attention image super-resolution. In ECCV, 2024. 2, 5, 6, 7, 8, 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[77] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[78] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.192, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[79] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCVW, pages 56-72. Springer, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.235, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[80] Kun Zhou, Xinyu Lin, LIU Zhonghang, Xiaoguang Han, and Jiangbo Lu. Ups: Unified projection sharing for lightweight single-image super-resolution and beyond. In NeurIPS, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[81] Lin Zhou, Haoming Cai, Jinjin Gu, Zheyuan Li, Yingqi Liu, Xiangyu Chen, Yu Qiao, and Chao Dong. Efficient image super-resolution using vast-receptive-field attention. In EC-CVW, pages 256-272. Springer, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[82] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In ICCV, pages 12780–12791, 2023. 2, 4, 5, 6, 7, 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24477" + } + ] +] \ No newline at end of file diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_origin.pdf b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b4980bd0f0343422a51510ea0dd0982e91bd4f9c --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/28032f25-1d68-496c-851a-6c9d0a64d207_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3a5234a5a97d27f07d243bba12cdd83933b5d9f2033f5941c432b9d3272b5e4b +size 8227120 diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/full.md b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9a6e26e545543d5bf22ae7264ba9180b2daa2da --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/full.md @@ -0,0 +1,434 @@ +# Emulating Self-attention with Convolution for Efficient Image Super-Resolution + +Dongheon Lee, Seokju Yun, Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea + +{dslisleedh, wsz871, youngmin.ro}@uos.ac.kr + +Code: https://github.com/dslisleedh/ESC + +# Abstract + +In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution (SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention (ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to $32 \times 32$ with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100×2 while reducing latency and memory usage by $16 \times$ and $12.2 \times$ . Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution (ESC), notably improves PSNR by 0.27 dB on Urban100×4 compared to HiT-SRF, reducing the latency and memory usage by $3.7 \times$ and $6.2 \times$ , respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module. + +# 1. Introduction + +Image Super-Resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs and remains an active area of research in computer vision. Recently, the demand for multimedia content and generative models has increased significantly, making SR particularly noteworthy as it enables users to enjoy high-quality content under resource-constrained conditions. Consequently, practical deployment has emerged as a critical consideration in + +![](images/57e4709eed99ce2895f22728acd5884b2b7ad30facbd5a3f946ef13f48d8a708.jpg) +Figure 1. Comparison of performance, latency, and memory usage. Our methods are evaluated against representative SR models, including CNNs, Transformers, and SSMs. + +SR tasks, motivating numerous SR studies to enhance performance while reducing computational complexity (FLOATing point OPerations; FLOPs) and parameter size. + +Therefore, Transformers have gained significant attention in the SR task since they have achieved superior performance over Convolutional Neural Networks (CNNs) while requiring lower FLOPs and fewer parameters. By capturing long-range dependencies and performing input-dependent weighting through self-attention, Transformers exhibit high representational capacity with enhanced performance, especially as the training data volume increases [5, 15]. Accordingly, numerous transformer-based methods have been proposed [47, 56, 75] to capitalize on the advantages of self-attention while reducing FLOPs and parameter counts. + +However, many of these works overlook the excessive memory access caused by self-attention, which arises from materializing the score matrix $(S = QK^T)$ [10] and leveraging memory-bound operations such as tensor reshaping and window masking [39, 68]. The memory access issue is further worsened in the SR architecture, which deals with large feature maps without patchify stem or downsampling stage. For example, SwinIR-light [36] exhibits $4.7\times$ slower latency and $2\times$ higher memory usage than CNNs [77] reconstructing an HD image at the $\times 2$ scale, even though it + +requires $14.5 \times$ and $17 \times$ fewer FLOPs and parameter sizes, respectively. Consequently, despite their promising performance, Transformers are challenging to deploy on resource-constrained devices, such as consumer-level GPUs. + +In this paper, we aim to design an efficiency-driven Transformer tailored for lightweight SR tasks, achieving both improved performance and memory overhead. We start by conducting a preliminary analysis of self-attention, observing that the similarity modeling performed by self-attention and resulting extracted features remain highly consistent across several layers, as shown in Figure 2. This finding exhibits that self-attention may extract overlapping features, suggesting the possibility of reducing computational overhead without compromising representational capability by utilizing efficient alternatives. Building on this finding, we establish a design strategy that retains self-attention only in the first layer of each block while replacing the remaining layers with our proposed efficient alternative, the Convolutional Attention (ConvAttn) module. To effectively emulate self-attention's long-range modeling and instance-dependent weighting, the ConvAttn module operates with a twofold mechanism. First, it simplifies self-attention's long-range interaction by applying convolution with a shared $13 \times 13$ large kernel throughout the entire network, targeting only a subset of channels. Second, dynamic kernels are generated to capture input-dependent weighting, mimicking the adaptive nature of self-attention. By combining these components, the ConvAttn module significantly reduces the reliance on memory-intensive self-attention while maintaining the representational capability of Transformers. + +With most self-attention layers replaced by ConvAttn, we leverage this efficiency to further enhance the remaining self-attention layers. Specifically, we enlarge the window size of self-attention, significantly improving performance with only a slight increase in FLOPs. However, increasing the window size leads to an enlarged score matrix, substantially raising peak memory usage. To address this, we introduce Flash Attention [10, 14] into lightweight SR tasks to avoid materializing the score matrix. Our optimized implementation allows us to scale up the window size to $32 \times 32$ while reducing latency and memory usage by $16 \times$ and $12.2 \times$ , respectively, as illustrated in Table 1. + +Based on these approaches, we introduce the lightweight SR network called Emulating Self-attention with Convolution (ESC). The proposed ESC outperforms ATD-light [73] on Urban $100 \times 4$ with 0.1dB improvements on PSNR while being $8.9 \times$ faster, as shown in Figure 1. Additionally, ESC-light surpasses ELAN-light [75] by a large margin of 0.29dB in PSNR on Urban $100 \times 2$ while also reducing latency by $22\%$ . We further validate our ESC in scenarios where reducing FLOPs and parameter size is essential by introducing ESC-FP, which outperforms MambaIRV2-light [21] on Manga $109 \times 4$ with reductions of FLOPs and + +parameter size by $20\%$ and $32\%$ , respectively. Through our extensive experiments, we demonstrate that our ESC fully leverages the advantages of Transformers – including their large receptive fields, representational capacity, and scalability with respect to data volume – even though most self-attention are replaced by the ConvAttn module. We support these results with in-depth experiments, as illustrated in Figure 6, suggesting that the proposed ConvAttn module extracts similar features with self-attention. + +Our contributions are summarized as follows: + +- We demonstrate that well-designed convolution partially replaces self-attention, significantly improving efficiency without sacrificing the advantages of Transformers. +- We mitigate the memory overhead of self-attention by integrating Flash Attention, thereby enlarging the window size to $32 \times 32$ without excessive memory usage. To the best of our knowledge, this is the first successful application of Flash Attention in lightweight SR tasks. +- We fully exploit the Transformer's benefits in lightweight SR tasks by making them more efficient and simpler. + +# 2. Related Work + +Traditional CNNs Early deep learning-based SR research [13, 29, 37, 77, 78] primarily employed convolution that extracts local features from images. These approaches typically relied on stacked convolutions, which often resulted in over-parameterization. To address this issue, numerous studies [28, 35, 46, 53] have shared their weights across multiple layers to reduce parameter size. However, many of them suffer from performance degradation since the input-independent weights are shared across multiple layers. In this paper, we share a large kernel through the entire network while leveraging input-dependently generated dynamic Depth-Wise Convolution (DWC) [22] by each layer in parallel. This approach efficiently mitigates the parameter growth of the large kernel while maintaining each layer's representational power. + +Transformers Pioneering studies that significantly reduce computational costs by computing self-attention within local windows [36] or channel-wise [69] have sparked considerable interest in applying Transformers [15, 55] to the SR task. Subsequent studies have proposed methods that enhance self-attention to strengthen its long-range modeling capabilities while reducing both FLOPs and parameters [7, 47, 56, 73, 75, 76, 82]. Nonetheless, Transformers remain challenging to deploy due to the excessive memory access caused by self-attention and involved memory-bound operations. In this paper, we address this limitation by replacing self-attention with convolutions that emulate its advantages, based on the observations of self-attention's inter-layer repetition. Our approach preserves the advantages of Transformers while reducing reliance on self-attention and associated memory-bound operations. + +![](images/a270a1cbcef1320720f80ffc80268667ef698c571b0fa59d655c777c60961f22.jpg) +(a) CKA Similarity of $F$ +(b) Cosine Similarity of $A^{avg}$ + +Large Kernel CNNs and State-Space Models Several recent studies have highlighted the computational burden of Transformers and investigated alternative core operators to replicate their advantages. Specifically, CNN-based models [52, 62, 63, 81] attempt to mimic the advantages of Transformers by leveraging large kernels and pixel attention [79], while State-Space Models (SSM) [20, 21] adapt their verified long-range modeling abilities from the 1D sequence domain [17, 18] to the 2D image domain. However, these approaches often overlook either the exceptional representational capability brought by self-attention's sophisticated similarity modeling or the additional memory access caused by supplementary mechanisms to adapting 2D domains like multi-directional or attentive scanning. In this paper, we fully leverage the advantages of both self-attention and convolution while utilizing Flash Attention [10, 14] to address the memory access of self-attention without relying on complex additional mechanisms. + +# 3. Proposed Methods + +In this section, we start by representing our preliminary analysis and then describe our proposed network. + +# 3.1. Preliminary analysis + +Transformers often achieve improved performance by utilizing self-attention's long-range modeling ability and instance-dependent weighting. However, each self-attention incurs excessive memory overhead, resulting in significant latency and memory usage. Interestingly, our preliminary analysis demonstrates that the sophisticated similarity modeling of self-attention $(A^{avg})$ and the resulting extracted features $(F)$ exhibit a high degree of inter-layer similarities, with averages of $89\%$ and $87\%$ . This observation not only aligns with recent research [80] but also explains the success of studies that reduced computational cost by sharing the attention map across layers [47, 75]. Inspired by this observation, we investigate improving efficiency without compromising representational power. + +![](images/6c89fe3b2a2a5d7d51bd5c02340abdca3996a0849c2c522af739c538816d8674.jpg) +Figure 2. Visualized inter-layer similarities of the feature extracted by self-attention $(F)$ before being added to the skip path and corresponding attention map $(A^{avg})$ from SwinIR-light [36]. Both CKA [31] and cosine similarities are measured on the Urban $100\times 4$ dataset. The window and head dimensions of the attention map are averaged for visualization. +Figure 3. Visualization of the overall architecture. + +# 3.2. Overall Structure + +Our network is composed of four main components, as shown in Figure 3. First, the LR input image $I_{LR} \in \mathbb{R}^{H \times W \times 3}$ is transformed into a shallow feature $F_0 \in \mathbb{R}^{H \times W \times C}$ through a $3 \times 3$ convolution $(\mathrm{Conv}_{3 \times 3})$ . Here, $H$ and $W$ denote height and width of $I_{LR}$ , and $C$ denotes feature dimension. Next, $F_0$ and a large kernel $LK \in \mathbb{R}^{13 \times 13 \times 16 \times 16}$ that is learned end-to-end manner and shared across the entire model are fed into the deep feature extractor $H$ , and the resulting deep feature is added back to $F_0$ to produce the final feature $F \in \mathbb{R}^{H \times W \times C}$ . In parallel, the image-wise skip module $S$ takes $I_{LR}$ as input to generate a skipped feature $F_s$ , and the upsampler $U$ utilize both $F$ and $F_s$ as inputs to return the upsampled image $I_{SR} \in \mathbb{R}^{rH \times rW \times 3}$ where $r$ denotes upscaling factor. The entire process is formulated as follows: + +$$ +\begin{array}{l} F _ {0} = \operatorname {C o n v} _ {3 \times 3} \left(I _ {L R}\right), \\ F = H \left(F _ {0}, L K\right) + F _ {0}, \tag {1} \\ F _ {s} = S \left(I _ {L R}\right), \\ I _ {S R} = U (F, F _ {s}). \\ \end{array} +$$ + +The $H$ is composed of the $N$ ESCBlocks where $N$ denotes the number of ESCBlocks and a $3 \times 3$ convolution, and is formulated as follows: + +$$ +\begin{array}{l} \begin{array}{l} F _ {i} = \operatorname {E S C B l o c k} _ {i} \left(F _ {i - 1}, L K\right), \quad i = 1, \dots , N, \\ F _ {i} = G _ {i - 1} \left(F _ {i}\right) \end{array} \tag {2} \\ F _ {d} = \operatorname {C o n v} _ {3 \times 3} \left(F _ {N}\right), \\ \end{array} +$$ + +where $F_{i-1}$ indicates the feature input to the $i$ th ESCBlock together with $LK$ , and $F_i$ represents the extracted feature. The $U$ and $S$ will be described in the implementation detail section because they will be changed depending on the task. + +# 3.3. ESC Block + +The overall process of the ESCBlock is illustrated in Figure 4 and is formulated as follows: + +$$ +F _ {i} ^ {i n} = \operatorname {C o n v F F N} \big (\mathrm {L N} (F _ {i - 1}) \big), +$$ + +$$ +F _ {i, 0} = F _ {i} ^ {i n} + \operatorname {S e l f A t t n} \left(\mathrm {L N} \left(F _ {i} ^ {i n}\right)\right), +$$ + +$$ +\begin{array}{l} F _ {i, j} = F _ {i, j - 1} + \operatorname {C o n v A t t n} _ {j} \left(\operatorname {C o n v F F N} _ {j} \left(F _ {i, j - 1}\right), L K\right), \\ j = 1, \dots , M, \\ F _ {i} = F _ {i - 1} + \operatorname {C o n v} _ {3 \times 3} \left(\mathrm {L N} \left(F _ {i, M}\right)\right). \tag {3} \\ \end{array} +$$ + +![](images/65d86a19ecc2491781bd546dee653a97e6cc843f672b761b396f27107055e837.jpg) +Figure 4. Visualization of the proposed ESC block. + +Firstly, the $F_{i-1}$ is converted to the $F_{i}^{in}$ by layer normalization (LN) and convolutional FFN (ConvFFN) [82] with the kernel size of three. Then, LN and naive window self-attention [36] with the window size of $32 \times 32$ are employed in a residual connection to extract $F_{i,0}$ . Despite its large window size, we effectively reduce memory overhead by utilizing Flash Attention. Furthermore, since ConvFFN is placed before, the self-attention extracts features considering local information without the complicated QKV projection module. After establishing $F_{i,0}$ , the block iterates residual paths composed of ConvFFN and proposed ConvAttn module (ConvAttn) $M$ times. Although our methods do not employ shifted-window self-attention, these convolution-based modules after the self-attention will extract inter-window features. Finally, the result of the last iteration $F_{i,M}$ go through LN and $\mathrm{Conv}_{3 \times 3}$ , then is added to $F_{i-1}$ , resulting in $F_{i}$ . + +Convolutional Attention Module The ConvAttn is designed to emulate self-attention's two key advantages: capturing long-range dependencies and instance-dependent weighting. For this purpose, ConvAttn leverages two types of convolution formulated as: + +$$ +\begin{array}{l} F _ {i, j} ^ {a t t}, F _ {i, j} ^ {i d t} = \mathrm {S p l i t} _ {1 6: C - 1 6} (F _ {i, j} ^ {\mathrm {C F}}), \\ D K _ {i, j} = \operatorname {C o n v} _ {1 \times 1} ^ {u p} \left(\phi \left(\operatorname {C o n v} _ {1 \times 1} ^ {\text {d o w n}} \left(\operatorname {G A P} \left(F _ {i, j} ^ {\text {a t t}}\right)\right)\right)\right), \\ F _ {i, j} ^ {r e s} = (F _ {i, j} ^ {a t t} \circledast D K _ {i, j}) + (F _ {i, j} ^ {a t t} \circledast L K), \\ F _ {i, j} ^ {f u s e} = \operatorname {C o n v} _ {1 \times 1} ^ {f u s e} \left(\operatorname {C o n c a t} \left(F _ {i, j} ^ {r e s}, F _ {i, j} ^ {i d t}\right)\right), \tag {4} \\ \end{array} +$$ + +where $F_{i,j}^{\mathrm{CF}}$ denotes the features extracted by $\mathrm{ConvFFN}_{i,j}$ before. We first split the $F_{i,j}^{\mathrm{CF}}$ channel-wise, resulting in $F_{i,j}^{att} \in \mathbb{R}^{H \times W \times 16}$ , which consists of the 16 channels in front, and $F_{i,j}^{idt} \in \mathbb{R}^{H \times W \times C - 16}$ , which contains the remaining part. Two types of convolutions are only applied within $F_{i,j}^{att}$ to reduce memory access cost, unlike the methods that leverage large kernels considering only computational complexity [52, 62, 63]. Next, dynamic depth-wise kernel $DK_{i,j} \in \mathbb{R}^{3 \times 3 \times 1 \times 16}$ [22] is created by two layers of $\mathrm{Conv}_{1 \times 1}$ with global average pooling (GAP) and gaussian + +error linear unit activation $(\phi)$ . Since enlarging $DK$ 's kernel size increases the parameter size multiplied by the hidden dimension $(h)$ of the kernel estimator, we fix its kernel size to $3\times 3$ . Then both $DK_{i,j}$ and $LK$ are convolved on $F_{i,j}^{att}$ and sum up those, resulting convolved feature $F_{i,j}^{res}$ . Based on our preliminary analysis, we hypothesize that long-range interactions do not significantly vary across layers. Therefore, we share $LK$ across the entire ESCBlocks and ConvAttns, mitigating parameter growth and optimization difficulty [6, 38]. Finally, $F_{i,j}^{res}$ and $F_{i,j}^{idt}$ are concatenated channel-wise, followed by the last convolution $\mathrm{Conv}_{1\times 1}^{fuse}$ to fuse them, resulting in fused feature $F_{i,j}^{use}$ . In this way, both the local features extracted by ConvFFN and the long-range features extracted by ConvAttn can be leveraged to capture multi-scale information. + +While $DK_{i,j}$ can be merged into $LK$ by padding both spatial and channel dimensions to reduce FLOPs, we do not because it slows down latency. On the other hand, in situations where reducing FLOPs and parameter size is crucial rather than reducing latency, $LK$ can be decomposed depthwise separable manner [25] into point-wise large kernel filter $LK^{\mathrm{c}}\in \mathbb{R}^{1\times 1\times 16\times 16}$ and depth-wise large kernel filter $LK^{\mathrm{s}}\in \mathbb{R}^{13\times 13\times 1\times 16}$ . This decomposition redefines line 3 in Equation 4 as follows: + +$$ +F _ {i, j} ^ {r e s} = \left(F _ {i, j} ^ {a t t} * L K ^ {c}\right) * \left(Z P \left(D K _ {i, j}\right) + L K ^ {s}\right), \tag {5} +$$ + +where $ZP(\cdot): \mathbb{R}^{3 \times 3 \times 1 \times 16} \mapsto \mathbb{R}^{13 \times 13 \times 1 \times 16}$ denotes spatial zero padding. + +In summary, ConvAttn efficiently emulates self-attention's two advantages by leveraging a shared $LK$ and layer-specific $DKs$ . In contrast to Transformers that share attention maps across several layers [47, 75, 80], which still require a full materialized attention map for every layer, our module only leverages a single shared $LK$ , significantly reducing memory footprints. Moreover, unlike recent CNNs [52, 62, 63, 81] that use large kernel convolutions and optionally employ pixel-wise attention to inject input dependency, we employ layer-specific $DKs$ , emulating self-attention's input-dependent weight more exactly in a memory-efficient manner. + +Table 1. Comparison of latency and memory usage between naive self-attention and our Flash Attention (FA) [10, 14]. WS and #Pixels denote the window size of self-attention and the corresponding number of pixels. $X$ refers to the input tensor. + +
WS(#Pixels)TypeX ∈ R64×64×64X ∈ R288×288×64
Latency(ms)Memory Usage(mb)Latency(ms)Memory Usage(mb)
16×16(256)w/o FA2.0464.9758
w/ FA0.5 (4×)11 (4.2×)1.3 (3.8×)223 (3.4×)
24×24(576)w/o FA7.611114.51575
w/ FA0.5 (15.2×)14 (7.9×)2.0 (7.3×)223 (7.1×)
32×32(1024)w/o FA27.315744.72717
w/ FA0.5 (54.6×)11 (14.3×)2.8 (16.0×)223 (12.2×)
+ +Table 2. Comparisons of classic SR methods trained on the DIV2K dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. † denotes that we re-calculate the statistics of the indicated methods. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded. + +
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)Set5Set14PSNR / SSIM B100Urban100Manga109
SwinIR-lt [36]1409.81287244.291038.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-lt [75]94.5887203.162138.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
OmniSR [56]120.31031194.577238.22/0.961333.98/0.921032.36/0.902033.05/0.936339.28/0.9784
SRFormer-lt [82]1456.31184236.385338.23/0.961333.94/0.920932.36/0.901932.91/0.935339.28/0.9785
ATD-lt [73]733.52839380.075338.29/0.961634.10/0.921732.39/0.902333.27/0.937539.52/0.9789
HiT-SRF [76]268.11804226.584738.26/0.961534.01/0.921432.37/0.902333.13/0.937239.47/0.9787
ASID-D8 [47] ×2131.2999190.5†73238.32/0.961834.24/0.923232.40/0.902833.35/0.9387-/-
MambaIR-lt [20]277.11695334.290538.13/0.961033.95/0.920832.31/0.901332.85/0.934939.20/0.9782
MambaIRV2-lt [21]580.42824286.377438.26/0.961534.09/0.922132.36/0.901933.26/0.937839.35/0.9785
RDN [78]279.320585096.22212338.24/0.961434.01/0.921232.34/0.901732.89/0.935339.18/0.9780
RCAN [77]299.36263529.71544538.27/0.961434.12/0.921632.41/0.902733.34/0.938439.44/0.9786
ESC-FP (Ours)94.3627239.852438.27/0.961734.09/0.921932.37/0.902233.22/0.937539.40/0.9784
ESC-lt (Ours)73.2830359.460338.24/0.961533.98/0.921132.35/0.902033.05/0.936339.33/0.9786
ESC (Ours)120.9831592.094738.35/0.961934.11/0.922332.41/0.902733.46/0.939539.54/0.9790
SwinIR-lt [36] ×3331.7596110.891834.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-lt [75]32.539990.162934.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
OmniSR [56]41.247688.478034.70/0.929430.57/0.846929.28/0.809428.84/0.865634.22/0.9487
SRFormer-lt [82]530.5537105.486134.67/0.929630.57/0.846929.26/0.809928.81/0.865534.19/0.9489
ATD-lt [73]274.41258168.076034.74/0.930030.68/0.848529.32/0.810929.17/0.870934.60/0.9506
HiT-SRF [76]124.91464101.685534.75/0.930030.61/0.847529.29/0.810628.99/0.868734.53/0.9502
ASID-D8 [47] ×361.946086.4†73934.84/0.930730.66/0.849129.32/0.811929.08/0.8706-/-
MambaIR-lt [20]109.3760148.591334.63/0.928830.54/0.845929.23/0.808428.70/0.863134.12/0.9479
MambaIRV2-lt [21]259.01250126.778134.71/0.929830.68/0.848329.26/0.809829.01/0.868934.41/0.9497
RDN [78]146.59852281.22230834.71/0.929630.57/0.846829.26/0.809328.80/0.865334.13/0.9484
RCAN [77]85.15601586.11562934.74/0.929930.65/0.848229.32/0.811129.09/0.870234.44/0.9499
ESC-FP (Ours)34.8291110.053034.72/0.930030.67/0.848329.30/0.810729.12/0.870634.56/0.8706
ESC-lt (Ours)25.2384162.861234.61/0.929530.52/0.847529.26/0.810228.93/0.867934.33/0.9495
ESC (Ours)41.4385267.695534.84/0.930830.74/0.849329.34/0.811829.28/0.873934.66/0.9512
SwinIR-lt [36] ×4222.935163.693032.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-lt [75]18.024154.164032.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
OmniSR [56]22.527350.979232.49/0.898828.78/0.785927.71/0.741526.64/0.801831.02/0.9151
SRFormer-lt [82]287.232962.887332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
ATD-lt [73]189.7753100.176932.63/0.899828.89/0.788627.79/0.744026.97/0.810731.48/0.9198
HiT-SRF [76]82.1133158.086632.55/0.899928.87/0.788027.75/0.743226.80/0.806931.26/0.9171
ASID-D8 [47] ×461.826549.6†74832.57/0.899028.89/0.789827.78/0.744926.89/0.8096-/-
MambaIR-lt [20]55.843884.692432.42/0.897728.74/0.784727.68/0.740026.52/0.798330.94/0.9135
MambaIRV2-lt [21]153.474875.679032.51/0.899228.84/0.787827.75/0.742626.82/0.807931.24/0.9182
RDN [78]66.07911309.22227132.47/0.899028.81/0.787127.72/0.741926.61/0.802831.00/0.9151
RCAN [78]52.2540917.61559232.63/0.900228.87/0.788927.77/0.743626.82/0.808731.22/0.9173
ESC-FP (Ours)21.715860.853932.56/0.889728.87/0.788927.75/0.743526.90/0.809831.40/0.9192
ESC-lt (Ours)13.321391.062432.52/0.899528.87/0.787827.72/0.742326.76/0.805831.26/0.9173
ESC (Ours)21.9215149.296832.68/0.901128.93/0.790227.80/0.744727.07/0.814431.54/0.9207
+ +# 4. Experiments + +This section demonstrates the effectiveness of our flash attention and our extensive experiments on various tasks. For implementation details for proposed networks, please refer to Section 7 in the supplementary. + +# 4.1. Effectiveness of Flash Attention + +We address the excessive memory overhead of self-attention by leveraging Flash Attention [10, 14]. Implementation details for Flash Attention are described in Section 6 in the supplementary. As shown in Table 1, our self-attention implementation accelerated by Flash Attention significantly reduces the latency and memory usage of + +self-attention at the window size of 32 by $16 \times$ and $12.2 \times$ , respectively. We do not consider window sizes above 32, as it may cause training instability due to the necessity of padding, since the training patch size is $64 \times 64$ . + +# 4.2. Classic SR Results + +To evaluate our proposed networks on the classic SR task, we leverage five commonly used datasets (Set5 [3], Set14 [70], B100 [42], Urban100 [26], Manga109 [43]). We measure peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) on the Y channel after cropping the boundary corresponding to the upscaling factor and converting them into YCbCr space. All statistics are measured by reconstructing an HD im + +![](images/ad3af948326f3342cb5f8005975d510b3e9a14d0eac45efea5a7fb215ff5b0a6.jpg) + +![](images/c153873211bd2a23aaebb7757db2b6027b83da5d9c15fb68504e9f830e6ae64e.jpg) +SRFormer-lt + +![](images/65391aea2f9fc40c1adbc21c5730896859bba76ac65aa04776a69d740d5ea47b.jpg) +24.62 +ATD-lt + +![](images/839e2db0b400b76afa72833c903b2c38de879c159c427015545f94a521294572.jpg) +HiT-SRF + +![](images/849b55c8467b6952edcedf3182ff0c9c2b813ccb9e505b2feabfae42733e2c29.jpg) +25.21 +MambaIRV2-lt + +![](images/142d00e0098cd39b24e6eb5eae1735479f93132f6ba4f23b25f7b27becbbacb1.jpg) +HR Crop + +![](images/e3c115592736133a2bbb758c1c4ed9e3c1d88e18cf2733f141e5e5adb67b4c17.jpg) +PSNR +ESC (Ours) + +![](images/591253c6e9a584a32ecb6167fd1829964a297e1b1cbb504dd9de242fff00d60e.jpg) +Urban100 - img017 +Urban100 - img059 +Figure 5. Visual comparisons for classic SR $(\times 2)$ . The best result on PSNR is bolded. + +![](images/ec536c400451b99d049a1c245819215e8478d8031988c4aede1121b48a55c11e.jpg) +24.45 +SRFormer-lt + +![](images/33e8d3ff8330afda30bf591b69bc1126290ab1bf4a30e53381895c8dbaa2a0a1.jpg) +26.56 +ATD-1t +28.90 + +![](images/8acf3361c1d5ea1ca3991aecd08d03e6d04f799e4471dd2342c2595f99b0c454.jpg) +25.23 +HiT-SRI + +![](images/a9a38dd95e565080fad042693ed43cbeaa87f51163b4c9e3adfaacfebc7cf722.jpg) +2671 +MambaIRV2-lt +26.41 + +![](images/c584f3c1cd8088185615ed38bc49c594861712f5ee195443eca248de5ceea6ca.jpg) +26.50 +HR Crops + +![](images/881497e03039a1651738526853488b53a3e959cd0973d27d2511327e57ac4951.jpg) +PSNR +ESC (Ours) +37.75 + +age (1280×720) image using an RTX4090 GPU at FP32 precision 100 times. Memory usage is calculated by torch.cuda.max_memory_allocated function and FLOPs are measured by fvcore library. For the classic SR task, we present three variants: ESC, ESC-light, and ESC-FP. ESC and ESC-light prioritize reducing latency, while ESC-FP focuses on reducing FLOPs and parameter size. + +# 4.2.1. Quantitative Results on the DIV2K dataset + +We compare our methods with various representative lightweight SR methods, including CNN-based models [77, 78], Transformers [36, 47, 56, 73, 75, 76, 82], and SSMs [20, 21] trained on the DIV2K dataset [1]. Notably, as shown in Table 2, our ESC achieves a notable 0.1dB PSNR improvement on the Urban $100 \times 4$ dataset, while being $8.7 \times$ faster than ATD-light. Moreover, ESC-light is $1.4 \times$ faster than ELAN-light, exhibiting an impressive performance improvement of 0.34dB PSNR in the Manga109 $\times 4$ dataset. These results exhibit that state-of-the-art performance can be achieved even when most self-attention are replaced with convolutions, suggesting that replacing self-attention with convolutions does not compromise representational power. Furthermore, ESC-FP outperforms MambaIRV2-light over 0.18dB on the Urban $100 \times 4$ dataset while reducing FLOPs and parameters by $20\%$ and $31\%$ , respectively. This result underscores that our design is effective even when reducing FLOPs and parameter size. + +# 4.2.2. Visual Results on the DIV2K dataset + +We also visually compare our ESC with representative lightweight SR methods, including SRFormer-light [82], HiT-SRF [76], ATD-light [73], and MambaIRV2-light [21]. As shown in Figure 5, our ESC effectively restores details and achieves a high PSNR, demonstrating that our network also produces visually pleasant results. + +# 4.2.3. Ablation Study + +We conduct an ablation study to justify our network's composition, as shown in Table 3. First, we introduce two vari- + +Table 3. Ablation studies on our ESC network. All comparisons are designed to have similar latencies to ESC. $LK$ , $DK$ , and WS denote large kernel, dynamic kernel, and window size, respectively. All statistics and metrics are measured as detailed in Section 4.2. + +
CasesLatency (ms)Memory (mb)#FLOPs (G)#Params (K)Set5×2Urban100×2
Only Self-attention128.2830456.673738.27/0.961733.23/0.9378
Only ConvAttns126.3514962.4121038.18/0.961332.91/0.9353
9×9 LK117.1831462.292438.33/0.961833.42/0.9392
17×17 LK126.6832768.997738.32/0.961833.40/0.9390
w/o LK Share and DK119.7835591.1195038.32/0.961833.37/0.9389
w/o LK Share120.9835592.0198538.32/0.961933.43/0.9392
w/o DK119.7835591.191138.31/0.961833.36/0.9388
WS16 and more layers118.0807603.6105438.24/0.961433.05/0.9360
WS24 and more layers116.5801583.599538.33/0.961833.28/0.9383
ESC (Ours)120.9831592.094738.35/0.961933.46/0.9395
+ +ants, one that uses only the self-attention module and another that uses only the ConvAttn. The number of blocks and layers is adjusted to match the latency on par with the ESC. Both variants exhibit a performance drop compared to ESC, suggesting that leveraging both self-attention and ConvAttn offers the best performance-efficiency trade-off. Next, we vary the $LK$ 's kernel size to $9 \times 9$ and $17 \times 17$ to determine the optimal size. In both cases, we observe a drop in performance, confirming that a kernel size of $13 \times 13$ is optimal. Then, we evaluate performance changes by altering the convolution type used in the ConvAttn. First, we observe performance drops when neither $LK$ sharing nor $DK$ is used. Then, we add $DK$ without sharing $LK$ and observe some performance improvement, though it still falls short of ESC. This result suggests that addressing the optimization difficulty that comes from the massive parameter size is more critical than the layer-specific long-range modeling. Finally, we leverage only $LK$ sharing without $DK$ and observe significant performance drops, affirming that harmonizing both $LK$ and $DK$ is crucial. For the next step, we investigate the effect of the self-attention window size by setting it to 16 and 24 and adding extra layers. In every case, we observe a performance drop, which confirms that enlarging the window size to enhance self-attention is effective even when only a single self-attention is leveraged in each block. + +# 4.2.4. Analysis on ConvAttn Module + +This section verifies that the proposed ConvAttn operates similarly to self-attention. Self-attention is known to extract shape-biased structural features [12, 48], unlike $3 \times 3$ convolution, which mainly extracts local texture and edges. Indeed, as shown in Figure 6 (a), $F^{idt}$ focuses on local textures and edges while $F^{att}$ concentrates on structural components. Next, to verify each convolution's receptive field, we visualize positive gradients $\bar{G}^{+} = \frac{1}{NMC} \sum_{p=1}^{NMC} \max \left(0, \frac{\partial I_{i,j}^{SR}}{\partial F_p}\right)$ propagated from the output image's center pixel $(i,j)$ to the inputs feature $(F)$ of both $DK$ and $LK$ across $N$ blocks, $M$ modules, and $C$ + +Table 4. Comparisons of classic SR methods trained on the large-scale DFLIP dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. $\S$ denotes that we train the indicated methods following their descriptions. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded. + +
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)PSNR / SSIM
Set5Set14B100Urban100Manga109
SRFormer-lt$ [82]×21838.11184236.385338.24/0.961534.13/0.921832.42/0.902633.37/0.938639.36/0.9787
ATD-lt$ [73]733.52839380.075338.29/0.961634.30/0.923032.43/0.902733.62/0.940139.60/0.9791
HiT-SRF$ [76]268.11804226.584738.31/0.961634.31/0.923032.45/0.903133.58/0.940439.69/0.9793
ESC (Ours)120.9831592.094738.34/0.961834.42/0.923532.50/0.903633.86/0.942439.73/0.9795
SRFormer-lt$ [82]×3668.3537105.486134.67/0.929730.75/0.848429.30/0.810829.10/0.870134.26/0.9498
ATD-lt$ [73]274.41258168.076034.71/0.930030.77/0.849329.33/0.811629.42/0.874334.61/0.9509
HiT-SRF$ [76]124.91464101.685534.69/0.929830.81/0.849329.32/0.811529.28/0.872934.72/0.9511
ESC (Ours)41.4385267.695534.85/0.931230.97/0.851129.41/0.813529.70/0.879934.94/0.9525
SRFormer-lt$ [82]×4327.832962.887332.49/0.899328.89/0.788727.76/0.742926.90/0.808631.25/0.9189
ATD-lt$ [73]189.7753100.176932.52/0.899528.93/0.789627.79/0.744327.18/0.815031.47/0.9208
HiT-SRF$ [76]82.1133158.086632.55/0.899728.96/0.789727.77/0.744327.07/0.813031.59/0.9208
ESC (Ours)21.9215149.296832.79/0.902529.06/0.792727.85/0.746627.45/0.822931.87/0.9239
+ +![](images/3d9a970861a0d3d661c8844eeebcb52b27d54d432db42f6c4f04d6f355e91048.jpg) +$F^{idt}$ + +![](images/35affe7f1e3902deff6d3ea76d8b7e4369be2f9d84d91f5a23356019d3c62836.jpg) +$\bar{G}_{DK}^{+}$ (DI: 48.77) + +![](images/a0a20bee06f34aed40681fb8d02503e30ec673ab4bd2c9479bae7f053a2caa77.jpg) +$S_{DK}$ + +![](images/ef548769d0b4d3f8ee5f1b118810a9e2b7f74654dd6cbc0a74e55214a0879ada.jpg) +(a) Feature Map Visualization +Figure 6. Visualized feature map, gradients, and input-wise similarity. $F$ is averaged across entire ConvAttn while processing the Lenna image in Set14. $\tilde{G}^{+}$ and $S$ are calculated on the Urban100, and 10 samples in front of the Urban100, respectively. + +![](images/bfef0cf15bb9711874b8c09eae430559285602e1f9b8a67ecc615433292d6520.jpg) +$G_{LK}^{+}$ (DI: 62.23) +(b) Gradient Visualization + +![](images/4cd9179fe74725d498817dbaf4140e31bc437341ecdf6431922a4dd3488366b9.jpg) +(c) Similarity Visualization + +channels, and corresponding diffusion index (DI). Figure 6 (b) clearly demonstrates that $LK$ extracts more long-range features than $DK$ . Lastly, we visualize input-wise similarity $(S)$ between average-pooled features extracted by $LK$ and $DK$ . As demonstrated in Figure 6 (c), $S_{DK}$ exhibits higher input-dependent diversity than $S_{LK}$ , confirming that $DK$ extracts more input-dependent features than $LK$ . + +# 4.2.5. Local Attribution Map + +We use the Local Attribution Map (LAM) [19] to verify that our networks can extract long-range dependencies despite using less self-attention. We compare our ESC with studies of various self-attention implementations that aim to reduce FLOPs while expanding the receptive field, such as global-level attention [56], permuted self-attention [82], and spatial-channel correlation [76]. Notably, as described in Figure 7, our ESC references the widest range of pixels to restore the red bounding box along with the highest diffusion index. This result exhibits that our network effectively captures long-range dependencies even though most self + +![](images/28a54a66a3da98e00f804e36dd4a842bdff47d01e20c549a2bb0ab6a91691fb1.jpg) + +![](images/c8e9f9d3d98d47e5514293184ba7feecfbfd459092ac30ad7f0b42be2f224e20.jpg) +Urban100 - img063x4 Diffusion Index +Figure 7. Comparisons of Local Attribution Map (LAM) [19] and the corresponding area of contribution and diffusion index. The best result on the diffusion index is bolded. + +![](images/a7133d47f6035acf0854cc473d0099708571c6bb8846fe78e982d669b40d647b.jpg) + +![](images/5d65bda4de02635cd7d89c52dfb7ea503d9b222953dbd368470102f5454cef3e.jpg) +OmniSR + +![](images/c56951aa9bc9019df0c7e4b8842da8444b4e3deee09e1e81cd5ef9bc40c8501d.jpg) + +![](images/6e985bc46b05a8fc2ae735c7fe1d73f6f84a7cd0407a095ac1925d7160a7fad8.jpg) +SRFormer-It + +![](images/f69aa6221f006f8faf78404a3d331261e07ff16362d3433e7ddfff1d175f304f.jpg) + +![](images/cf1db73828c316408c14858af8efd92a98eb797abe052ad76d913da64ea9dbd9.jpg) +iT-SRF + +![](images/8dff5e1b36b6cc9d7cecc4c919c7bbbead87b13a5ba5dcc8369ccc162d1723d9.jpg) + +![](images/68a3b37e09115a0c7939e7431866efa9835c738ff170463f2c558338dd6797e8.jpg) +ESC (Ours) + +attentions are replaced by the ConvAttn module, suggesting that our ConvAttn module contributes to retaining the large receptive field of Transformers. + +# 4.2.6. Quantitative Results on the DFLIP dataset + +Transformers exhibit significant performance improvements when trained on large-scale datasets [5, 15]. To verify our model also retains this benefit, we train and evaluate various representative Transformers – including SRFormerlight [82], ATD-light [73], and HiT-SRF [76] – alongside our ESC on a large dataset. For this purpose, we leverage DIV2K+Flickr2K+LSDIR+DiverSeg-IP (DFLIP) [1, 34, 45, 54] dataset. As illustrated in Table 4, our approach outperforms the other methods by a substantial margin across all scales and evaluation datasets. These results indicate that replacing self-attention with ConvAttn does not sacrifice the data scaling capability of Transformers. + +# 4.3. Arbitrary-Scale SR + +To evaluate our proposed networks on arbitrary-scale SR tasks, we leverage the DIV2K-validation [1] dataset. Following previous research [8, 32], we assess PSNR in the RGB channel after cropping the boundaries according to the upscaling factor plus six. For arbitrary-scale SR tasks, we change $U$ of our network and representative lightweight SR transforems [73, 76] into LTE [32]. + +# 4.3.1. Quantitative Results + +To demonstrate the outstanding representational capability of the proposed network, we train and evaluate the network on an arbitrary-scale SR task, which requires processing various upscaling factors with a single network. For quantitative comparison, we leverage commonly used RDN [78] and additional representative Transformers [73, 76] that we train. As described in Tables 5, our network continually achieves high performance at the scale both seen and unseen during training, even achieving a $0.09\mathrm{dB}$ improvement over $\mathrm{RDN + LTE}$ at $\times 2$ . This result underscores our ESC's exceptional representational capability. For quantitative results measured on other datasets, including Set5, Set14, B100, and Urban100, refer to Section 8 in the supplementary. + +Table 5. Quantitative comparison on arbitrary-scale SR task employing LTE [32] as upsampler. The best result on PSNR is bolded. + +
MethodsSeenUnseen
×2×3×4×6×12×18×24×30
RDN+LTE [78]35.0431.3229.3327.0423.9522.4021.3620.64
ATD-lt+LTE§ [73]34.9831.2529.2727.0123.9422.3821.3720.64
HiT-SRF+LTE§ [76]35.1031.3529.3527.0523.9422.3921.3620.64
ESC+LTE (Ours)35.1331.3829.3727.0823.9722.4021.3820.64
+ +# 4.3.2. Visual Results + +We also visually compare our network with RDN, HiT-SRF, and ATD-light. As shown in Figure 8, Our network effectively restores fine details that other methods fail on both $\times 4$ and $\times 12$ scales. This result demonstrates that our network also excels at restoring visually present results on both in-distribution and out-of-distribution scales. + +# 4.4. Real-World Super-Resolution + +We further test our network on real-world SR tasks that aim to restore images with unknown degradations. To this end, we introduce a variant called ESC-Real and compare them to representative real-world SR methods, including RealESRGAN+ [59], SwinIR-Real [36], and DASR [61]. For our evaluation, we use the commonly employed RealSRSet dataset [72]. As shown in Table 9, our ESC-Real recovers more fine-grained details than the other methods while significantly reducing unpleasant artifacts in both graphic and natural images. This result demonstrates that our proposed methods are effective even when dealing with complex unknown degradations, underscoring our methods' + +![](images/e48203c369aeef8b5aceb8276a16d186b4609294a80ad356b7dc6ebdb383a22f.jpg) +Urban100 - img011 + +![](images/69599c31241637d91e9c41a8477276e7ff0f8f475fc33b2a309fc997a4a63956.jpg) +RDN/17.71 HiT-SRF/17.56 + +![](images/5261f96ba48392c1cc34fe5cbbeab63e52dc9b89901d2e3dd11516fa527d6e8b.jpg) +ATD-lt / 17.91 ESC (Ours) / 20.70 +(a) Scale $\times 4$ + +![](images/004309304b2629b400b6dd09514120fb4ff0293752614acdf976ae2006651900.jpg) +DIV2K-valid - 0845 + +![](images/5aae8ff9766fd6649fa991747e5884d70dc73248b0214f6d1217292993674d8c.jpg) + +![](images/9ca1c858354b8bb8f7d7f76d144d632ff1ad8a21a4aa2771a8445659bc2f233d.jpg) +RDN/16.25 HiT-SRF/17.28 +ATD-lt / 16.77 +(b) Scale $\times 12$ + +![](images/4454287fc45e67358b390d1ed12973c9fddb5d8d08b416b79b0c5a7fa82588d1.jpg) +ESC (Ours) / 17.82 + +![](images/ce69f6c552a458a04e971142c286f89dcb172cf8a519dc1efa8c5cdbe994c86f.jpg) +Figure 8. Visual comparisons on arbitrary-scale SR for both indistribution $(\times 4)$ and out-of-distribution $(\times 12)$ scales. The best result on PSNR is bolded. +(a) RealSRSet - 0030 + +![](images/13859fafde256993ad2a62390e20e941d31a546487edda44d7296f90d2e04ca8.jpg) +(b) RealSRSet - dog +Figure 9. Visual comparisons on real-world SR $(\times 4)$ tasks. + +outstanding representational capability. Quantitative results on real-world SR tasks are presented in Section 9 in the supplementary. + +# 5. Conclusion + +In this paper, we propose ESC, a novel lightweight SR network designed to mitigate the excessive memory overhead of Transformers. Our extensive experiments demonstrate that most self-attention layers can be effectively replaced with carefully designed convolutions to improve efficiency. Moreover, we introduce Flash Attention into lightweight SR tasks, significantly enhancing performance without complicated modifications to self-attention. As a result, we fully exploit the advantages of Transformers by prioritizing improvements in their efficiency. + +Acknowledgments This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400]. + +# References + +[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6, 7, 8, 1 +[2] Yang Ai, Xiaoqiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. NeurIPS, 37:55443-55469, 2024. 2 +[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 2 +[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, pages 3086-3095, 2019. 2 +[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 1, 7 +[6] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In CVPR, pages 5557-5567, 2024. 4 +[7] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In CVPR, pages 22367-22377, 2023. 2 +[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, pages 8628-8638, 2021. 8 +[9] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NeurIPS, 2022. 3 +[10] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 1, 2, 3, 4, 5 +[11] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 44(5):2567-2581, 2020. 2 +[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, pages 11963-11975, 2022. 6 +[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE TPAMI, 38(2):295-307, 2015. 2 +[14] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A programming model for generating optimized attention kernels. arXiv preprint arXiv:2412.05496, 2024. 2, 3, 4, 5, 1 + +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 7 +[16] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In CVPRW, pages 2494-2502, 2021. 1 +[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint, 2023. 3 +[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. *ICLR*, 2022. 3 +[19] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9199-9208, 2021. 7 +[20] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 3, 5, 6 +[21] Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. Mambairv2: Attentive state space restoration. CVPR, 2025. 2, 3, 5, 6 +[22] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between local attention and dynamic depth-wise convolution. *ICLR*, 2022, 2, 4 +[23] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In ECCV, pages 289-305. Springer, 2024. 1 +[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 2 +[25] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 4 +[26] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 5, 2 +[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021. 2 +[28] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, pages 1637-1645, 2016. 2 +[29] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 2 +[30] Diederik P Kingma and Jimmy Ba. Adam: a method for stochastic optimization. In ICLR, 2014. 1 +[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representa + +tions revisited. In ICML, pages 3519-3529. PMLR, 2019. 3 +[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In CVPR, pages 1929-1938, 2022. 8, 1, 2 +[33] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In IJCAI, pages 1089-1097, 2023. 3 +[34] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPRW, pages 1775-1787, 2023. 7, 1 +[35] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image superresolution. In CVPR, pages 3867-3876, 2019. 2 +[36] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 8 +[37] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 136-144, 2017. 2 +[38] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. ICLR, 2023. 4 +[39] Yong Liu, Hang Dong, Boyang Liang, Songwei Liu, Qingji Dong, Kai Chen, Fangmin Chen, Lean Fu, and Fei Wang. Unfolding once is enough: A deployment-friendly transformer unit for super-resolution. In ACMMM, pages 7952-7960, 2023. 1 +[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1 +[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 1 +[42] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423. IEEE, 2001. 5, 2 +[43] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 5 +[44] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. SPL, 20(3): 209-212, 2012. 2 +[45] Go Ohtani, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka, and Yoshimitsu Aoki. Rethinking image super-resolution from training data perspectives. In ECCV, pages 19-36. Springer, 2024. 7, 1 +[46] Karam Park and Nam Ik Cho. Partial filter-sharing: Improved parameter-sharing method for single image super-resolution networks. In WACV, pages 2653-2663. IEEE, 2025. 2 + +[47] Karam Park, Jae Woong Soh, and Nam Ik Cho. Efficient attention-sharing information distillation transformer for lightweight single image super-resolution. In AAAI, 2025. 1, 2, 3, 4, 5, 6 +[48] Namuk Park and Songkuk Kim. How do vision transformers work? In ICLR, 2022. 6 +[49] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1 +[50] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 1 +[51] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1 +[52] Long Sun, Jinshan Pan, and Jinhui Tang. Shufflemixer: An efficient convnet for image super-resolution. NeurIPS, 35: 17314-17326, 2022. 3, 4 +[53] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, pages 3147-3155, 2017. 2 +[54] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 114-125, 2017. 7, 1 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2 +[56] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Liu jinfan. Omni aggregation networks for lightweight image super-resolution. In CVPR, 2023. 1, 2, 5, 6, 7 +[57] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, pages 2555-2563, 2023. 2 +[58] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, pages 606-615, 2018. 2 +[59] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, pages 1905-1914, 2021. 8, 2 +[60] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, pages 101-117. Springer, 2020. 2 +[61] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In CVPR, pages 13385-13394, 2021. 8 +[62] Gang Wu, Junjun Jiang, Junpeng Jiang, and Xianming Liu. Transforming image super-resolution: a convformer-based efficient approach. IEEE Transactions on Image Processing, 2024. 3, 4 + +[63] Chengxing Xie, Xiaoming Zhang, Linze Li, Haiteng Meng, Tianlin Zhang, Tianrui Li, and Xiaole Zhao. Large kernel distillation network for efficient single image superresolution. In CVPRW, pages 1283-1292, 2023. 3, 4 +[64] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, pages 1191-1200, 2022. 2 +[65] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1 +[66] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, pages 15703-15712, 2025. 1 +[67] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In WACV, pages 4956-4965, 2023. 3 +[68] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5756-5767, 2024. 1 +[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: efficient attention transformer for high-resolution image restoration. In CVPR, pages 5728-5739, 2022. 2 +[70] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 2 +[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 3 +[72] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, pages 4791-4800, 2021. 8, 2 +[73] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In CVPR, pages 2856-2865, 2024. 2, 5, 6, 7, 8, 1, 3 +[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 2 +[75] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In ECCV, pages 649-667. Springer, 2022. 1, 2, 3, 4, 5, 6 +[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient attention image super-resolution. In ECCV, 2024. 2, 5, 6, 7, 8, 1 + +[77] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 1, 2, 5, 6 +[78] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2, 5, 6, 8 +[79] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCVW, pages 56-72. Springer, 2020. 3 +[80] Kun Zhou, Xinyu Lin, LIU Zhonghang, Xiaoguang Han, and Jiangbo Lu. Ups: Unified projection sharing for lightweight single-image super-resolution and beyond. In NeurIPS, 2024. 3, 4 +[81] Lin Zhou, Haoming Cai, Jinjin Gu, Zheyuan Li, Yingqi Liu, Xiangyu Chen, Yu Qiao, and Chao Dong. Efficient image super-resolution using vast-receptive-field attention. In EC-CVW, pages 256-272. Springer, 2022. 3, 4 +[82] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In ICCV, pages 12780–12791, 2023. 2, 4, 5, 6, 7, 1, 3 \ No newline at end of file diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/images.zip b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..595d4554921397408e0d904998a41d3bc39189e8 --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:25b9466523f73e0333009fd8a9d72efee40c3fefbc09fefd8472a0a4cd73e71b +size 1119384 diff --git a/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/layout.json b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..49723ca432ad8f97064d9f3b9602577261050803 --- /dev/null +++ b/2025/Emulating Self-attention with Convolution for Efficient Image Super-Resolution/layout.json @@ -0,0 +1,13785 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 59, + 103, + 550, + 121 + ], + "type": "text", + "content": "Emulating Self-attention with Convolution for Efficient Image Super-Resolution" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 143, + 452, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 143, + 452, + 171 + ], + "spans": [ + { + "bbox": [ + 157, + 143, + 452, + 171 + ], + "type": "text", + "content": "Dongheon Lee, Seokju Yun, Youngmin Ro* Machine Intelligence Laboratory, University of Seoul, Korea" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 188, + 173, + 421, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 188, + 173, + 421, + 185 + ], + "spans": [ + { + "bbox": [ + 188, + 173, + 421, + 185 + ], + "type": "text", + "content": "{dslisleedh, wsz871, youngmin.ro}@uos.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 229, + 187, + 378, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 229, + 187, + 378, + 198 + ], + "spans": [ + { + "bbox": [ + 229, + 187, + 378, + 198 + ], + "type": "text", + "content": "Code: https://github.com/dslisleedh/ESC" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": "In this paper, we tackle the high computational overhead of Transformers for efficient image super-resolution (SR). Motivated by the observations of self-attention's inter-layer repetition, we introduce a convolutionized self-attention module named Convolutional Attention (ConvAttn) that emulates self-attention's long-range modeling capability and instance-dependent weighting with a single shared large kernel and dynamic kernels. By utilizing the ConvAttn module, we significantly reduce the reliance on self-attention and its involved memory-bound operations while maintaining the representational capability of Transformers. Furthermore, we overcome the challenge of integrating flash attention into the lightweight SR regime, effectively mitigating self-attention's inherent memory bottleneck. We scale up the window size to " + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": " with flash attention rather than proposing an intricate self-attention module, significantly improving PSNR by 0.31dB on Urban100×2 while reducing latency and memory usage by " + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "inline_equation", + "content": "12.2 \\times" + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": ". Building on these approaches, our proposed network, termed Emulating Self-attention with Convolution (ESC), notably improves PSNR by 0.27 dB on Urban100×4 compared to HiT-SRF, reducing the latency and memory usage by " + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "inline_equation", + "content": "3.7 \\times" + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "inline_equation", + "content": "6.2 \\times" + }, + { + "bbox": [ + 54, + 251, + 297, + 574 + ], + "type": "text", + "content": ", respectively. Extensive experiments demonstrate that our ESC maintains the ability for long-range modeling, data scalability, and the representational power of Transformers despite most self-attention being replaced by the ConvAttn module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 696 + ], + "type": "text", + "content": "Image Super-Resolution (SR) aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs and remains an active area of research in computer vision. Recently, the demand for multimedia content and generative models has increased significantly, making SR particularly noteworthy as it enables users to enjoy high-quality content under resource-constrained conditions. Consequently, practical deployment has emerged as a critical consideration in" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 227, + 553, + 384 + ], + "blocks": [ + { + "bbox": [ + 316, + 227, + 553, + 384 + ], + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 384 + ], + "type": "image", + "image_path": "57e4709eed99ce2895f22728acd5884b2b7ad30facbd5a3f946ef13f48d8a708.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 388, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 422 + ], + "type": "text", + "content": "Figure 1. Comparison of performance, latency, and memory usage. Our methods are evaluated against representative SR models, including CNNs, Transformers, and SSMs." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 426, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 461 + ], + "type": "text", + "content": "SR tasks, motivating numerous SR studies to enhance performance while reducing computational complexity (FLOATing point OPerations; FLOPs) and parameter size." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 462, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 556, + 594 + ], + "type": "text", + "content": "Therefore, Transformers have gained significant attention in the SR task since they have achieved superior performance over Convolutional Neural Networks (CNNs) while requiring lower FLOPs and fewer parameters. By capturing long-range dependencies and performing input-dependent weighting through self-attention, Transformers exhibit high representational capacity with enhanced performance, especially as the training data volume increases [5, 15]. Accordingly, numerous transformer-based methods have been proposed [47, 56, 75] to capitalize on the advantages of self-attention while reducing FLOPs and parameter counts." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": "However, many of these works overlook the excessive memory access caused by self-attention, which arises from materializing the score matrix " + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "(S = QK^T)" + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": " [10] and leveraging memory-bound operations such as tensor reshaping and window masking [39, 68]. The memory access issue is further worsened in the SR architecture, which deals with large feature maps without patchify stem or downsampling stage. For example, SwinIR-light [36] exhibits " + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "4.7\\times" + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": " slower latency and " + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": " higher memory usage than CNNs [77] reconstructing an HD image at the " + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 312, + 594, + 557, + 715 + ], + "type": "text", + "content": " scale, even though it" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24467" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "requires " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "14.5 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "inline_equation", + "content": "17 \\times" + }, + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": " fewer FLOPs and parameter sizes, respectively. Consequently, despite their promising performance, Transformers are challenging to deploy on resource-constrained devices, such as consumer-level GPUs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "content": "In this paper, we aim to design an efficiency-driven Transformer tailored for lightweight SR tasks, achieving both improved performance and memory overhead. We start by conducting a preliminary analysis of self-attention, observing that the similarity modeling performed by self-attention and resulting extracted features remain highly consistent across several layers, as shown in Figure 2. This finding exhibits that self-attention may extract overlapping features, suggesting the possibility of reducing computational overhead without compromising representational capability by utilizing efficient alternatives. Building on this finding, we establish a design strategy that retains self-attention only in the first layer of each block while replacing the remaining layers with our proposed efficient alternative, the Convolutional Attention (ConvAttn) module. To effectively emulate self-attention's long-range modeling and instance-dependent weighting, the ConvAttn module operates with a twofold mechanism. First, it simplifies self-attention's long-range interaction by applying convolution with a shared " + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "inline_equation", + "content": "13 \\times 13" + }, + { + "bbox": [ + 56, + 122, + 296, + 434 + ], + "type": "text", + "content": " large kernel throughout the entire network, targeting only a subset of channels. Second, dynamic kernels are generated to capture input-dependent weighting, mimicking the adaptive nature of self-attention. By combining these components, the ConvAttn module significantly reduces the reliance on memory-intensive self-attention while maintaining the representational capability of Transformers." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": "With most self-attention layers replaced by ConvAttn, we leverage this efficiency to further enhance the remaining self-attention layers. Specifically, we enlarge the window size of self-attention, significantly improving performance with only a slight increase in FLOPs. However, increasing the window size leads to an enlarged score matrix, substantially raising peak memory usage. To address this, we introduce Flash Attention [10, 14] into lightweight SR tasks to avoid materializing the score matrix. Our optimized implementation allows us to scale up the window size to " + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": " while reducing latency and memory usage by " + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "inline_equation", + "content": "12.2 \\times" + }, + { + "bbox": [ + 55, + 436, + 295, + 579 + ], + "type": "text", + "content": ", respectively, as illustrated in Table 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Based on these approaches, we introduce the lightweight SR network called Emulating Self-attention with Convolution (ESC). The proposed ESC outperforms ATD-light [73] on Urban " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "100 \\times 4" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " with 0.1dB improvements on PSNR while being " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "8.9 \\times" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " faster, as shown in Figure 1. Additionally, ESC-light surpasses ELAN-light [75] by a large margin of 0.29dB in PSNR on Urban " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "100 \\times 2" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " while also reducing latency by " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "22\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": ". We further validate our ESC in scenarios where reducing FLOPs and parameter size is essential by introducing ESC-FP, which outperforms MambaIRV2-light [21] on Manga " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "109 \\times 4" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " with reductions of FLOPs and" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "text", + "content": "parameter size by " + }, + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "inline_equation", + "content": "32\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 178 + ], + "type": "text", + "content": ", respectively. Through our extensive experiments, we demonstrate that our ESC fully leverages the advantages of Transformers – including their large receptive fields, representational capacity, and scalability with respect to data volume – even though most self-attention are replaced by the ConvAttn module. We support these results with in-depth experiments, as illustrated in Figure 6, suggesting that the proposed ConvAttn module extracts similar features with self-attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 325, + 180, + 511, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 180, + 511, + 191 + ], + "spans": [ + { + "bbox": [ + 325, + 180, + 511, + 191 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 553, + 312 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 193, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 193, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 193, + 553, + 228 + ], + "type": "text", + "content": "- We demonstrate that well-designed convolution partially replaces self-attention, significantly improving efficiency without sacrificing the advantages of Transformers." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 229, + 553, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 553, + 289 + ], + "type": "text", + "content": "- We mitigate the memory overhead of self-attention by integrating Flash Attention, thereby enlarging the window size to " + }, + { + "bbox": [ + 313, + 229, + 553, + 289 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 229, + 553, + 289 + ], + "type": "text", + "content": " without excessive memory usage. To the best of our knowledge, this is the first successful application of Flash Attention in lightweight SR tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 553, + 312 + ], + "type": "text", + "content": "- We fully exploit the Transformer's benefits in lightweight SR tasks by making them more efficient and simpler." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 320, + 401, + 333 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 320, + 401, + 333 + ], + "spans": [ + { + "bbox": [ + 314, + 320, + 401, + 333 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 521 + ], + "type": "text", + "content": "Traditional CNNs Early deep learning-based SR research [13, 29, 37, 77, 78] primarily employed convolution that extracts local features from images. These approaches typically relied on stacked convolutions, which often resulted in over-parameterization. To address this issue, numerous studies [28, 35, 46, 53] have shared their weights across multiple layers to reduce parameter size. However, many of them suffer from performance degradation since the input-independent weights are shared across multiple layers. In this paper, we share a large kernel through the entire network while leveraging input-dependently generated dynamic Depth-Wise Convolution (DWC) [22] by each layer in parallel. This approach efficiently mitigates the parameter growth of the large kernel while maintaining each layer's representational power." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Transformers Pioneering studies that significantly reduce computational costs by computing self-attention within local windows [36] or channel-wise [69] have sparked considerable interest in applying Transformers [15, 55] to the SR task. Subsequent studies have proposed methods that enhance self-attention to strengthen its long-range modeling capabilities while reducing both FLOPs and parameters [7, 47, 56, 73, 75, 76, 82]. Nonetheless, Transformers remain challenging to deploy due to the excessive memory access caused by self-attention and involved memory-bound operations. In this paper, we address this limitation by replacing self-attention with convolutions that emulate its advantages, based on the observations of self-attention's inter-layer repetition. Our approach preserves the advantages of Transformers while reducing reliance on self-attention and associated memory-bound operations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24468" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 292, + 169 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 292, + 169 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 292, + 169 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 292, + 169 + ], + "type": "image", + "image_path": "a270a1cbcef1320720f80ffc80268667ef698c571b0fa59d655c777c60961f22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 76, + 170, + 148, + 179 + ], + "lines": [ + { + "bbox": [ + 76, + 170, + 148, + 179 + ], + "spans": [ + { + "bbox": [ + 76, + 170, + 148, + 179 + ], + "type": "text", + "content": "(a) CKA Similarity of " + }, + { + "bbox": [ + 76, + 170, + 148, + 179 + ], + "type": "inline_equation", + "content": "F" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 184, + 170, + 272, + 179 + ], + "lines": [ + { + "bbox": [ + 184, + 170, + 272, + 179 + ], + "spans": [ + { + "bbox": [ + 184, + 170, + 272, + 179 + ], + "type": "text", + "content": "(b) Cosine Similarity of " + }, + { + "bbox": [ + 184, + 170, + 272, + 179 + ], + "type": "inline_equation", + "content": "A^{avg}" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 258, + 296, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 296, + 474 + ], + "type": "text", + "content": "Large Kernel CNNs and State-Space Models Several recent studies have highlighted the computational burden of Transformers and investigated alternative core operators to replicate their advantages. Specifically, CNN-based models [52, 62, 63, 81] attempt to mimic the advantages of Transformers by leveraging large kernels and pixel attention [79], while State-Space Models (SSM) [20, 21] adapt their verified long-range modeling abilities from the 1D sequence domain [17, 18] to the 2D image domain. However, these approaches often overlook either the exceptional representational capability brought by self-attention's sophisticated similarity modeling or the additional memory access caused by supplementary mechanisms to adapting 2D domains like multi-directional or attentive scanning. In this paper, we fully leverage the advantages of both self-attention and convolution while utilizing Flash Attention [10, 14] to address the memory access of self-attention without relying on complex additional mechanisms." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 481, + 167, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 167, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 167, + 495 + ], + "type": "text", + "content": "3. Proposed Methods" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 523 + ], + "type": "text", + "content": "In this section, we start by representing our preliminary analysis and then describe our proposed network." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 531, + 175, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 531, + 175, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 531, + 175, + 544 + ], + "type": "text", + "content": "3.1. Preliminary analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": "Transformers often achieve improved performance by utilizing self-attention's long-range modeling ability and instance-dependent weighting. However, each self-attention incurs excessive memory overhead, resulting in significant latency and memory usage. Interestingly, our preliminary analysis demonstrates that the sophisticated similarity modeling of self-attention " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "(A^{avg})" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " and the resulting extracted features " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " exhibit a high degree of inter-layer similarities, with averages of " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "89\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "87\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": ". This observation not only aligns with recent research [80] but also explains the success of studies that reduced computational cost by sharing the attention map across layers [47, 75]. Inspired by this observation, we investigate improving efficiency without compromising representational power." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 552, + 168 + ], + "blocks": [ + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "text", + "content": "Figure 2. Visualized inter-layer similarities of the feature extracted by self-attention " + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "text", + "content": " before being added to the skip path and corresponding attention map " + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "inline_equation", + "content": "(A^{avg})" + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "text", + "content": " from SwinIR-light [36]. Both CKA [31] and cosine similarities are measured on the Urban " + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "inline_equation", + "content": "100\\times 4" + }, + { + "bbox": [ + 55, + 187, + 295, + 253 + ], + "type": "text", + "content": " dataset. The window and head dimensions of the attention map are averaged for visualization." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 70, + 552, + 168 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 552, + 168 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 552, + 168 + ], + "type": "image", + "image_path": "6c89fe3b2a2a5d7d51bd5c02340abdca3996a0849c2c522af739c538816d8674.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 171, + 525, + 182 + ], + "lines": [ + { + "bbox": [ + 343, + 171, + 525, + 182 + ], + "spans": [ + { + "bbox": [ + 343, + 171, + 525, + 182 + ], + "type": "text", + "content": "Figure 3. Visualization of the overall architecture." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 187, + 421, + 198 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 187, + 421, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 187, + 421, + 198 + ], + "type": "text", + "content": "3.2. Overall Structure" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": "Our network is composed of four main components, as shown in Figure 3. First, the LR input image " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "I_{LR} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " is transformed into a shallow feature " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F_0 \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " through a " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " convolution " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "(\\mathrm{Conv}_{3 \\times 3})" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " denote height and width of " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " denotes feature dimension. Next, " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " and a large kernel " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "LK \\in \\mathbb{R}^{13 \\times 13 \\times 16 \\times 16}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " that is learned end-to-end manner and shared across the entire model are fed into the deep feature extractor " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": ", and the resulting deep feature is added back to " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F_0" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " to produce the final feature " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": ". In parallel, the image-wise skip module " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " takes " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "I_{LR}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " as input to generate a skipped feature " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": ", and the upsampler " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " utilize both " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "F_s" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " as inputs to return the upsampled image " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "I_{SR} \\in \\mathbb{R}^{rH \\times rW \\times 3}" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 204, + 555, + 384 + ], + "type": "text", + "content": " denotes upscaling factor. The entire process is formulated as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 391, + 553, + 447 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 391, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 383, + 391, + 553, + 447 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {0} = \\operatorname {C o n v} _ {3 \\times 3} \\left(I _ {L R}\\right), \\\\ F = H \\left(F _ {0}, L K\\right) + F _ {0}, \\tag {1} \\\\ F _ {s} = S \\left(I _ {L R}\\right), \\\\ I _ {S R} = U (F, F _ {s}). \\\\ \\end{array}", + "image_path": "23c0ecf3617aa823f9ce511d3866230a941cf60876bdc9c2ba65011b262a96e8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "content": " is composed of the " + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "content": " ESCBlocks where " + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "content": " denotes the number of ESCBlocks and a " + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 453, + 554, + 488 + ], + "type": "text", + "content": " convolution, and is formulated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 339, + 493, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 493, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 339, + 493, + 553, + 521 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\begin{array}{l} F _ {i} = \\operatorname {E S C B l o c k} _ {i} \\left(F _ {i - 1}, L K\\right), \\quad i = 1, \\dots , N, \\\\ F _ {i} = G _ {i - 1} \\left(F _ {i}\\right) \\end{array} \\tag {2} \\\\ F _ {d} = \\operatorname {C o n v} _ {3 \\times 3} \\left(F _ {N}\\right), \\\\ \\end{array}", + "image_path": "b943dfda699bbe69e5ccf23a6c9635db72528b087127b6c5dbc309ff56559836.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "F_{i-1}" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": " indicates the feature input to the " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": "th ESCBlock together with " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "F_i" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": " represents the extracted feature. The " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 526, + 554, + 575 + ], + "type": "text", + "content": " will be described in the implementation detail section because they will be changed depending on the task." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 580, + 388, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 388, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 388, + 592 + ], + "type": "text", + "content": "3.3. ESC Block" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 554, + 621 + ], + "type": "text", + "content": "The overall process of the ESCBlock is illustrated in Figure 4 and is formulated as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 318, + 625, + 447, + 640 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 625, + 447, + 640 + ], + "spans": [ + { + "bbox": [ + 318, + 625, + 447, + 640 + ], + "type": "interline_equation", + "content": "F _ {i} ^ {i n} = \\operatorname {C o n v F F N} \\big (\\mathrm {L N} (F _ {i - 1}) \\big),", + "image_path": "31b478ccf8955b208c8cd131b4ef0e687c2976b082a8062bf6c70b1848febd3a.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 642, + 465, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 642, + 465, + 656 + ], + "spans": [ + { + "bbox": [ + 320, + 642, + 465, + 656 + ], + "type": "interline_equation", + "content": "F _ {i, 0} = F _ {i} ^ {i n} + \\operatorname {S e l f A t t n} \\left(\\mathrm {L N} \\left(F _ {i} ^ {i n}\\right)\\right),", + "image_path": "82c9ead44134084a1c73f5df06b720c9ef4d1b5aacf1029c4c8cf9968e9f97b0.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {i, j} = F _ {i, j - 1} + \\operatorname {C o n v A t t n} _ {j} \\left(\\operatorname {C o n v F F N} _ {j} \\left(F _ {i, j - 1}\\right), L K\\right), \\\\ j = 1, \\dots , M, \\\\ F _ {i} = F _ {i - 1} + \\operatorname {C o n v} _ {3 \\times 3} \\left(\\mathrm {L N} \\left(F _ {i, M}\\right)\\right). \\tag {3} \\\\ \\end{array}", + "image_path": "eb14290e5bbaf448464d123b1f776962c0850ba180f85d9830feb7936e115eb6.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24469" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 293, + 228 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 293, + 228 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 293, + 228 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 293, + 228 + ], + "type": "image", + "image_path": "65d86a19ecc2491781bd546dee653a97e6cc843f672b761b396f27107055e837.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 233, + 268, + 243 + ], + "lines": [ + { + "bbox": [ + 82, + 233, + 268, + 243 + ], + "spans": [ + { + "bbox": [ + 82, + 233, + 268, + 243 + ], + "type": "text", + "content": "Figure 4. Visualization of the proposed ESC block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": "Firstly, the " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i-1}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": " is converted to the " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i}^{in}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": " by layer normalization (LN) and convolutional FFN (ConvFFN) [82] with the kernel size of three. Then, LN and naive window self-attention [36] with the window size of " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": " are employed in a residual connection to extract " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i,0}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": ". Despite its large window size, we effectively reduce memory overhead by utilizing Flash Attention. Furthermore, since ConvFFN is placed before, the self-attention extracts features considering local information without the complicated QKV projection module. After establishing " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i,0}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": ", the block iterates residual paths composed of ConvFFN and proposed ConvAttn module (ConvAttn) " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": " times. Although our methods do not employ shifted-window self-attention, these convolution-based modules after the self-attention will extract inter-window features. Finally, the result of the last iteration " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i,M}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": " go through LN and " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{3 \\times 3}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": ", then is added to " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i-1}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": ", resulting in " + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "inline_equation", + "content": "F_{i}" + }, + { + "bbox": [ + 55, + 246, + 296, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 509 + ], + "type": "text", + "content": "Convolutional Attention Module The ConvAttn is designed to emulate self-attention's two key advantages: capturing long-range dependencies and instance-dependent weighting. For this purpose, ConvAttn leverages two types of convolution formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 514, + 294, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 514, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 67, + 514, + 294, + 588 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} F _ {i, j} ^ {a t t}, F _ {i, j} ^ {i d t} = \\mathrm {S p l i t} _ {1 6: C - 1 6} (F _ {i, j} ^ {\\mathrm {C F}}), \\\\ D K _ {i, j} = \\operatorname {C o n v} _ {1 \\times 1} ^ {u p} \\left(\\phi \\left(\\operatorname {C o n v} _ {1 \\times 1} ^ {\\text {d o w n}} \\left(\\operatorname {G A P} \\left(F _ {i, j} ^ {\\text {a t t}}\\right)\\right)\\right)\\right), \\\\ F _ {i, j} ^ {r e s} = (F _ {i, j} ^ {a t t} \\circledast D K _ {i, j}) + (F _ {i, j} ^ {a t t} \\circledast L K), \\\\ F _ {i, j} ^ {f u s e} = \\operatorname {C o n v} _ {1 \\times 1} ^ {f u s e} \\left(\\operatorname {C o n c a t} \\left(F _ {i, j} ^ {r e s}, F _ {i, j} ^ {i d t}\\right)\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "5b2c75bb7530ae4e3f44583f17f00bd0d6bf2971f8f3c97f21d647b31941f4c0.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i,j}^{\\mathrm{CF}}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " denotes the features extracted by " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{ConvFFN}_{i,j}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " before. We first split the " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i,j}^{\\mathrm{CF}}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " channel-wise, resulting in " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i,j}^{att} \\in \\mathbb{R}^{H \\times W \\times 16}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": ", which consists of the 16 channels in front, and " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i,j}^{idt} \\in \\mathbb{R}^{H \\times W \\times C - 16}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": ", which contains the remaining part. Two types of convolutions are only applied within " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{i,j}^{att}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " to reduce memory access cost, unlike the methods that leverage large kernels considering only computational complexity [52, 62, 63]. Next, dynamic depth-wise kernel " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "DK_{i,j} \\in \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " [22] is created by two layers of " + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1 \\times 1}" + }, + { + "bbox": [ + 55, + 589, + 296, + 714 + ], + "type": "text", + "content": " with global average pooling (GAP) and gaussian" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": "error linear unit activation " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "(\\phi)" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": ". Since enlarging " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": "'s kernel size increases the parameter size multiplied by the hidden dimension " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "(h)" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " of the kernel estimator, we fix its kernel size to " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": ". Then both " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "DK_{i,j}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " are convolved on " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "F_{i,j}^{att}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " and sum up those, resulting convolved feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "F_{i,j}^{res}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": ". Based on our preliminary analysis, we hypothesize that long-range interactions do not significantly vary across layers. Therefore, we share " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " across the entire ESCBlocks and ConvAttns, mitigating parameter growth and optimization difficulty [6, 38]. Finally, " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "F_{i,j}^{res}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "F_{i,j}^{idt}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " are concatenated channel-wise, followed by the last convolution " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_{1\\times 1}^{fuse}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": " to fuse them, resulting in fused feature " + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "inline_equation", + "content": "F_{i,j}^{use}" + }, + { + "bbox": [ + 313, + 72, + 553, + 255 + ], + "type": "text", + "content": ". In this way, both the local features extracted by ConvFFN and the long-range features extracted by ConvAttn can be leveraged to capture multi-scale information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": "While " + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "inline_equation", + "content": "DK_{i,j}" + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": " can be merged into " + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": " by padding both spatial and channel dimensions to reduce FLOPs, we do not because it slows down latency. On the other hand, in situations where reducing FLOPs and parameter size is crucial rather than reducing latency, " + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": " can be decomposed depthwise separable manner [25] into point-wise large kernel filter " + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "inline_equation", + "content": "LK^{\\mathrm{c}}\\in \\mathbb{R}^{1\\times 1\\times 16\\times 16}" + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": " and depth-wise large kernel filter " + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "inline_equation", + "content": "LK^{\\mathrm{s}}\\in \\mathbb{R}^{13\\times 13\\times 1\\times 16}" + }, + { + "bbox": [ + 313, + 256, + 555, + 364 + ], + "type": "text", + "content": ". This decomposition redefines line 3 in Equation 4 as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 330, + 372, + 555, + 388 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 372, + 555, + 388 + ], + "spans": [ + { + "bbox": [ + 330, + 372, + 555, + 388 + ], + "type": "interline_equation", + "content": "F _ {i, j} ^ {r e s} = \\left(F _ {i, j} ^ {a t t} * L K ^ {c}\\right) * \\left(Z P \\left(D K _ {i, j}\\right) + L K ^ {s}\\right), \\tag {5}", + "image_path": "78a4c0b6a36d3a49e8c28452bea6ec98356929580eae7724539230e6a1ca2b76.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 393, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 553, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 393, + 553, + 418 + ], + "type": "inline_equation", + "content": "ZP(\\cdot): \\mathbb{R}^{3 \\times 3 \\times 1 \\times 16} \\mapsto \\mathbb{R}^{13 \\times 13 \\times 1 \\times 16}" + }, + { + "bbox": [ + 313, + 393, + 553, + 418 + ], + "type": "text", + "content": " denotes spatial zero padding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "content": "In summary, ConvAttn efficiently emulates self-attention's two advantages by leveraging a shared " + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "content": " and layer-specific " + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "inline_equation", + "content": "DKs" + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "content": ". In contrast to Transformers that share attention maps across several layers [47, 75, 80], which still require a full materialized attention map for every layer, our module only leverages a single shared " + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "content": ", significantly reducing memory footprints. Moreover, unlike recent CNNs [52, 62, 63, 81] that use large kernel convolutions and optionally employ pixel-wise attention to inject input dependency, we employ layer-specific " + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "inline_equation", + "content": "DKs" + }, + { + "bbox": [ + 313, + 418, + 554, + 561 + ], + "type": "text", + "content": ", emulating self-attention's input-dependent weight more exactly in a memory-efficient manner." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 619, + 555, + 712 + ], + "blocks": [ + { + "bbox": [ + 313, + 571, + 555, + 616 + ], + "lines": [ + { + "bbox": [ + 313, + 571, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 555, + 616 + ], + "type": "text", + "content": "Table 1. Comparison of latency and memory usage between naive self-attention and our Flash Attention (FA) [10, 14]. WS and #Pixels denote the window size of self-attention and the corresponding number of pixels. " + }, + { + "bbox": [ + 313, + 571, + 555, + 616 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 571, + 555, + 616 + ], + "type": "text", + "content": " refers to the input tensor." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 619, + 555, + 712 + ], + "lines": [ + { + "bbox": [ + 315, + 619, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 315, + 619, + 555, + 712 + ], + "type": "table", + "html": "
WS(#Pixels)TypeX ∈ R64×64×64X ∈ R288×288×64
Latency(ms)Memory Usage(mb)Latency(ms)Memory Usage(mb)
16×16(256)w/o FA2.0464.9758
w/ FA0.5 (4×)11 (4.2×)1.3 (3.8×)223 (3.4×)
24×24(576)w/o FA7.611114.51575
w/ FA0.5 (15.2×)14 (7.9×)2.0 (7.3×)223 (7.1×)
32×32(1024)w/o FA27.315744.72717
w/ FA0.5 (54.6×)11 (14.3×)2.8 (16.0×)223 (12.2×)
", + "image_path": "e3539a100e0566def5b8b0eb4860f342bbce17457c261128675a2b5d6e35169a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24470" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 105, + 556, + 534 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 2. Comparisons of classic SR methods trained on the DIV2K dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. † denotes that we re-calculate the statistics of the indicated methods. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 105, + 556, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 105, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 105, + 556, + 534 + ], + "type": "table", + "html": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)Set5Set14PSNR / SSIM B100Urban100Manga109
SwinIR-lt [36]1409.81287244.291038.14/0.961133.86/0.920632.31/0.901232.76/0.934039.12/0.9783
ELAN-lt [75]94.5887203.162138.17/0.961133.94/0.920732.30/0.901232.76/0.934039.11/0.9782
OmniSR [56]120.31031194.577238.22/0.961333.98/0.921032.36/0.902033.05/0.936339.28/0.9784
SRFormer-lt [82]1456.31184236.385338.23/0.961333.94/0.920932.36/0.901932.91/0.935339.28/0.9785
ATD-lt [73]733.52839380.075338.29/0.961634.10/0.921732.39/0.902333.27/0.937539.52/0.9789
HiT-SRF [76]268.11804226.584738.26/0.961534.01/0.921432.37/0.902333.13/0.937239.47/0.9787
ASID-D8 [47] ×2131.2999190.5†73238.32/0.961834.24/0.923232.40/0.902833.35/0.9387-/-
MambaIR-lt [20]277.11695334.290538.13/0.961033.95/0.920832.31/0.901332.85/0.934939.20/0.9782
MambaIRV2-lt [21]580.42824286.377438.26/0.961534.09/0.922132.36/0.901933.26/0.937839.35/0.9785
RDN [78]279.320585096.22212338.24/0.961434.01/0.921232.34/0.901732.89/0.935339.18/0.9780
RCAN [77]299.36263529.71544538.27/0.961434.12/0.921632.41/0.902733.34/0.938439.44/0.9786
ESC-FP (Ours)94.3627239.852438.27/0.961734.09/0.921932.37/0.902233.22/0.937539.40/0.9784
ESC-lt (Ours)73.2830359.460338.24/0.961533.98/0.921132.35/0.902033.05/0.936339.33/0.9786
ESC (Ours)120.9831592.094738.35/0.961934.11/0.922332.41/0.902733.46/0.939539.54/0.9790
SwinIR-lt [36] ×3331.7596110.891834.62/0.928930.54/0.846329.20/0.808228.66/0.862433.98/0.9478
ELAN-lt [75]32.539990.162934.61/0.928830.55/0.846329.21/0.808128.69/0.862434.00/0.9478
OmniSR [56]41.247688.478034.70/0.929430.57/0.846929.28/0.809428.84/0.865634.22/0.9487
SRFormer-lt [82]530.5537105.486134.67/0.929630.57/0.846929.26/0.809928.81/0.865534.19/0.9489
ATD-lt [73]274.41258168.076034.74/0.930030.68/0.848529.32/0.810929.17/0.870934.60/0.9506
HiT-SRF [76]124.91464101.685534.75/0.930030.61/0.847529.29/0.810628.99/0.868734.53/0.9502
ASID-D8 [47] ×361.946086.4†73934.84/0.930730.66/0.849129.32/0.811929.08/0.8706-/-
MambaIR-lt [20]109.3760148.591334.63/0.928830.54/0.845929.23/0.808428.70/0.863134.12/0.9479
MambaIRV2-lt [21]259.01250126.778134.71/0.929830.68/0.848329.26/0.809829.01/0.868934.41/0.9497
RDN [78]146.59852281.22230834.71/0.929630.57/0.846829.26/0.809328.80/0.865334.13/0.9484
RCAN [77]85.15601586.11562934.74/0.929930.65/0.848229.32/0.811129.09/0.870234.44/0.9499
ESC-FP (Ours)34.8291110.053034.72/0.930030.67/0.848329.30/0.810729.12/0.870634.56/0.8706
ESC-lt (Ours)25.2384162.861234.61/0.929530.52/0.847529.26/0.810228.93/0.867934.33/0.9495
ESC (Ours)41.4385267.695534.84/0.930830.74/0.849329.34/0.811829.28/0.873934.66/0.9512
SwinIR-lt [36] ×4222.935163.693032.44/0.897628.77/0.785827.69/0.740626.47/0.798030.92/0.9151
ELAN-lt [75]18.024154.164032.43/0.897528.78/0.785827.69/0.740626.54/0.798230.92/0.9150
OmniSR [56]22.527350.979232.49/0.898828.78/0.785927.71/0.741526.64/0.801831.02/0.9151
SRFormer-lt [82]287.232962.887332.51/0.898828.82/0.787227.73/0.742226.67/0.803231.17/0.9165
ATD-lt [73]189.7753100.176932.63/0.899828.89/0.788627.79/0.744026.97/0.810731.48/0.9198
HiT-SRF [76]82.1133158.086632.55/0.899928.87/0.788027.75/0.743226.80/0.806931.26/0.9171
ASID-D8 [47] ×461.826549.6†74832.57/0.899028.89/0.789827.78/0.744926.89/0.8096-/-
MambaIR-lt [20]55.843884.692432.42/0.897728.74/0.784727.68/0.740026.52/0.798330.94/0.9135
MambaIRV2-lt [21]153.474875.679032.51/0.899228.84/0.787827.75/0.742626.82/0.807931.24/0.9182
RDN [78]66.07911309.22227132.47/0.899028.81/0.787127.72/0.741926.61/0.802831.00/0.9151
RCAN [78]52.2540917.61559232.63/0.900228.87/0.788927.77/0.743626.82/0.808731.22/0.9173
ESC-FP (Ours)21.715860.853932.56/0.889728.87/0.788927.75/0.743526.90/0.809831.40/0.9192
ESC-lt (Ours)13.321391.062432.52/0.899528.87/0.787827.72/0.742326.76/0.805831.26/0.9173
ESC (Ours)21.9215149.296832.68/0.901128.93/0.790227.80/0.744727.07/0.814431.54/0.9207
", + "image_path": "cc215dab4958cfd0343f535d715af742396274e9278927044953af084d3af5ba.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 540, + 136, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 136, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 136, + 553 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 561, + 295, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 295, + 610 + ], + "type": "text", + "content": "This section demonstrates the effectiveness of our flash attention and our extensive experiments on various tasks. For implementation details for proposed networks, please refer to Section 7 in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 622, + 226, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 622, + 226, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 622, + 226, + 634 + ], + "type": "text", + "content": "4.1. Effectiveness of Flash Attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We address the excessive memory overhead of self-attention by leveraging Flash Attention [10, 14]. Implementation details for Flash Attention are described in Section 6 in the supplementary. As shown in Table 1, our self-attention implementation accelerated by Flash Attention significantly reduces the latency and memory usage of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "content": "self-attention at the window size of 32 by " + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "inline_equation", + "content": "12.2 \\times" + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "content": ", respectively. We do not consider window sizes above 32, as it may cause training instability due to the necessity of padding, since the training patch size is " + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 540, + 555, + 590 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 599, + 424, + 610 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 599, + 424, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 599, + 424, + 610 + ], + "type": "text", + "content": "4.2. Classic SR Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "To evaluate our proposed networks on the classic SR task, we leverage five commonly used datasets (Set5 [3], Set14 [70], B100 [42], Urban100 [26], Manga109 [43]). We measure peak signal-to-noise ratio (PSNR) and structural similarity index measure (SSIM) on the Y channel after cropping the boundary corresponding to the upscaling factor and converting them into YCbCr space. All statistics are measured by reconstructing an HD im" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24471" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 143, + 140 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 143, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 143, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 143, + 140 + ], + "type": "image", + "image_path": "ad3af948326f3342cb5f8005975d510b3e9a14d0eac45efea5a7fb215ff5b0a6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 143, + 70, + 194, + 96 + ], + "blocks": [ + { + "bbox": [ + 143, + 70, + 194, + 96 + ], + "lines": [ + { + "bbox": [ + 143, + 70, + 194, + 96 + ], + "spans": [ + { + "bbox": [ + 143, + 70, + 194, + 96 + ], + "type": "image", + "image_path": "c153873211bd2a23aaebb7757db2b6027b83da5d9c15fb68504e9f830e6ae64e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 97, + 187, + 102 + ], + "lines": [ + { + "bbox": [ + 151, + 97, + 187, + 102 + ], + "spans": [ + { + "bbox": [ + 151, + 97, + 187, + 102 + ], + "type": "text", + "content": "SRFormer-lt" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 143, + 110, + 194, + 135 + ], + "blocks": [ + { + "bbox": [ + 151, + 103, + 176, + 109 + ], + "lines": [ + { + "bbox": [ + 151, + 103, + 176, + 109 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 176, + 109 + ], + "type": "text", + "content": "24.62" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 110, + 194, + 135 + ], + "lines": [ + { + "bbox": [ + 143, + 110, + 194, + 135 + ], + "spans": [ + { + "bbox": [ + 143, + 110, + 194, + 135 + ], + "type": "image", + "image_path": "65391aea2f9fc40c1adbc21c5730896859bba76ac65aa04776a69d740d5ea47b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 135, + 179, + 141 + ], + "lines": [ + { + "bbox": [ + 159, + 135, + 179, + 141 + ], + "spans": [ + { + "bbox": [ + 159, + 135, + 179, + 141 + ], + "type": "text", + "content": "ATD-lt" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 194, + 70, + 244, + 96 + ], + "blocks": [ + { + "bbox": [ + 194, + 70, + 244, + 96 + ], + "lines": [ + { + "bbox": [ + 194, + 70, + 244, + 96 + ], + "spans": [ + { + "bbox": [ + 194, + 70, + 244, + 96 + ], + "type": "image", + "image_path": "839e2db0b400b76afa72833c903b2c38de879c159c427015545f94a521294572.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 207, + 97, + 232, + 102 + ], + "lines": [ + { + "bbox": [ + 207, + 97, + 232, + 102 + ], + "spans": [ + { + "bbox": [ + 207, + 97, + 232, + 102 + ], + "type": "text", + "content": "HiT-SRF" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 195, + 110, + 244, + 135 + ], + "blocks": [ + { + "bbox": [ + 211, + 102, + 227, + 109 + ], + "lines": [ + { + "bbox": [ + 211, + 102, + 227, + 109 + ], + "spans": [ + { + "bbox": [ + 211, + 102, + 227, + 109 + ], + "type": "text", + "content": "25.21" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 110, + 244, + 135 + ], + "lines": [ + { + "bbox": [ + 195, + 110, + 244, + 135 + ], + "spans": [ + { + "bbox": [ + 195, + 110, + 244, + 135 + ], + "type": "image", + "image_path": "849b55c8467b6952edcedf3182ff0c9c2b813ccb9e505b2feabfae42733e2c29.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 135, + 240, + 141 + ], + "lines": [ + { + "bbox": [ + 198, + 135, + 240, + 141 + ], + "spans": [ + { + "bbox": [ + 198, + 135, + 240, + 141 + ], + "type": "text", + "content": "MambaIRV2-lt" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 245, + 70, + 294, + 96 + ], + "blocks": [ + { + "bbox": [ + 245, + 70, + 294, + 96 + ], + "lines": [ + { + "bbox": [ + 245, + 70, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 245, + 70, + 294, + 96 + ], + "type": "image", + "image_path": "142d00e0098cd39b24e6eb5eae1735479f93132f6ba4f23b25f7b27becbbacb1.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 97, + 282, + 102 + ], + "lines": [ + { + "bbox": [ + 256, + 97, + 282, + 102 + ], + "spans": [ + { + "bbox": [ + 256, + 97, + 282, + 102 + ], + "type": "text", + "content": "HR Crop" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 245, + 110, + 294, + 135 + ], + "blocks": [ + { + "bbox": [ + 260, + 102, + 278, + 109 + ], + "lines": [ + { + "bbox": [ + 260, + 102, + 278, + 109 + ], + "spans": [ + { + "bbox": [ + 260, + 102, + 278, + 109 + ], + "type": "text", + "content": "PSNR" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 110, + 294, + 135 + ], + "lines": [ + { + "bbox": [ + 245, + 110, + 294, + 135 + ], + "spans": [ + { + "bbox": [ + 245, + 110, + 294, + 135 + ], + "type": "image", + "image_path": "e3c115592736133a2bbb758c1c4ed9e3c1d88e18cf2733f141e5e5adb67b4c17.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 135, + 286, + 141 + ], + "lines": [ + { + "bbox": [ + 255, + 135, + 286, + 141 + ], + "spans": [ + { + "bbox": [ + 255, + 135, + 286, + 141 + ], + "type": "text", + "content": "ESC (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 57, + 149, + 143, + 218 + ], + "blocks": [ + { + "bbox": [ + 72, + 141, + 128, + 149 + ], + "lines": [ + { + "bbox": [ + 72, + 141, + 128, + 149 + ], + "spans": [ + { + "bbox": [ + 72, + 141, + 128, + 149 + ], + "type": "text", + "content": "Urban100 - img017" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 57, + 149, + 143, + 218 + ], + "lines": [ + { + "bbox": [ + 57, + 149, + 143, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 149, + 143, + 218 + ], + "type": "image", + "image_path": "591253c6e9a584a32ecb6167fd1829964a297e1b1cbb504dd9de242fff00d60e.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 219, + 127, + 226 + ], + "lines": [ + { + "bbox": [ + 72, + 219, + 127, + 226 + ], + "spans": [ + { + "bbox": [ + 72, + 219, + 127, + 226 + ], + "type": "text", + "content": "Urban100 - img059" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 230, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 251 + ], + "type": "text", + "content": "Figure 5. Visual comparisons for classic SR " + }, + { + "bbox": [ + 55, + 230, + 295, + 251 + ], + "type": "inline_equation", + "content": "(\\times 2)" + }, + { + "bbox": [ + 55, + 230, + 295, + 251 + ], + "type": "text", + "content": ". The best result on PSNR is bolded." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 143, + 148, + 194, + 174 + ], + "blocks": [ + { + "bbox": [ + 159, + 142, + 176, + 148 + ], + "lines": [ + { + "bbox": [ + 159, + 142, + 176, + 148 + ], + "spans": [ + { + "bbox": [ + 159, + 142, + 176, + 148 + ], + "type": "text", + "content": "24.45" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 148, + 194, + 174 + ], + "lines": [ + { + "bbox": [ + 143, + 148, + 194, + 174 + ], + "spans": [ + { + "bbox": [ + 143, + 148, + 194, + 174 + ], + "type": "image", + "image_path": "ec536c400451b99d049a1c245819215e8478d8031988c4aede1121b48a55c11e.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 175, + 186, + 180 + ], + "lines": [ + { + "bbox": [ + 151, + 175, + 186, + 180 + ], + "spans": [ + { + "bbox": [ + 151, + 175, + 186, + 180 + ], + "type": "text", + "content": "SRFormer-lt" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 143, + 187, + 194, + 213 + ], + "blocks": [ + { + "bbox": [ + 151, + 181, + 177, + 186 + ], + "lines": [ + { + "bbox": [ + 151, + 181, + 177, + 186 + ], + "spans": [ + { + "bbox": [ + 151, + 181, + 177, + 186 + ], + "type": "text", + "content": "26.56" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 143, + 187, + 194, + 213 + ], + "lines": [ + { + "bbox": [ + 143, + 187, + 194, + 213 + ], + "spans": [ + { + "bbox": [ + 143, + 187, + 194, + 213 + ], + "type": "image", + "image_path": "33e8d3ff8330afda30bf591b69bc1126290ab1bf4a30e53381895c8dbaa2a0a1.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 159, + 213, + 178, + 219 + ], + "lines": [ + { + "bbox": [ + 159, + 213, + 178, + 219 + ], + "spans": [ + { + "bbox": [ + 159, + 213, + 178, + 219 + ], + "type": "text", + "content": "ATD-1t" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 159, + 220, + 177, + 226 + ], + "lines": [ + { + "bbox": [ + 159, + 220, + 177, + 226 + ], + "spans": [ + { + "bbox": [ + 159, + 220, + 177, + 226 + ], + "type": "text", + "content": "28.90" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 195, + 148, + 244, + 174 + ], + "blocks": [ + { + "bbox": [ + 211, + 142, + 227, + 148 + ], + "lines": [ + { + "bbox": [ + 211, + 142, + 227, + 148 + ], + "spans": [ + { + "bbox": [ + 211, + 142, + 227, + 148 + ], + "type": "text", + "content": "25.23" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 148, + 244, + 174 + ], + "lines": [ + { + "bbox": [ + 195, + 148, + 244, + 174 + ], + "spans": [ + { + "bbox": [ + 195, + 148, + 244, + 174 + ], + "type": "image", + "image_path": "8acf3361c1d5ea1ca3991aecd08d03e6d04f799e4471dd2342c2595f99b0c454.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 175, + 230, + 180 + ], + "lines": [ + { + "bbox": [ + 206, + 175, + 230, + 180 + ], + "spans": [ + { + "bbox": [ + 206, + 175, + 230, + 180 + ], + "type": "text", + "content": "HiT-SRI" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 195, + 187, + 244, + 213 + ], + "blocks": [ + { + "bbox": [ + 211, + 181, + 227, + 186 + ], + "lines": [ + { + "bbox": [ + 211, + 181, + 227, + 186 + ], + "spans": [ + { + "bbox": [ + 211, + 181, + 227, + 186 + ], + "type": "text", + "content": "2671" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 195, + 187, + 244, + 213 + ], + "lines": [ + { + "bbox": [ + 195, + 187, + 244, + 213 + ], + "spans": [ + { + "bbox": [ + 195, + 187, + 244, + 213 + ], + "type": "image", + "image_path": "a9a38dd95e565080fad042693ed43cbeaa87f51163b4c9e3adfaacfebc7cf722.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 213, + 240, + 219 + ], + "lines": [ + { + "bbox": [ + 198, + 213, + 240, + 219 + ], + "spans": [ + { + "bbox": [ + 198, + 213, + 240, + 219 + ], + "type": "text", + "content": "MambaIRV2-lt" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 198, + 220, + 226, + 226 + ], + "lines": [ + { + "bbox": [ + 198, + 220, + 226, + 226 + ], + "spans": [ + { + "bbox": [ + 198, + 220, + 226, + 226 + ], + "type": "text", + "content": "26.41" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 245, + 148, + 294, + 173 + ], + "blocks": [ + { + "bbox": [ + 260, + 142, + 277, + 148 + ], + "lines": [ + { + "bbox": [ + 260, + 142, + 277, + 148 + ], + "spans": [ + { + "bbox": [ + 260, + 142, + 277, + 148 + ], + "type": "text", + "content": "26.50" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 148, + 294, + 173 + ], + "lines": [ + { + "bbox": [ + 245, + 148, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 245, + 148, + 294, + 173 + ], + "type": "image", + "image_path": "c584f3c1cd8088185615ed38bc49c594861712f5ee195443eca248de5ceea6ca.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 260, + 175, + 282, + 181 + ], + "lines": [ + { + "bbox": [ + 260, + 175, + 282, + 181 + ], + "spans": [ + { + "bbox": [ + 260, + 175, + 282, + 181 + ], + "type": "text", + "content": "HR Crops" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 245, + 188, + 294, + 213 + ], + "blocks": [ + { + "bbox": [ + 260, + 182, + 278, + 187 + ], + "lines": [ + { + "bbox": [ + 260, + 182, + 278, + 187 + ], + "spans": [ + { + "bbox": [ + 260, + 182, + 278, + 187 + ], + "type": "text", + "content": "PSNR" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 245, + 188, + 294, + 213 + ], + "lines": [ + { + "bbox": [ + 245, + 188, + 294, + 213 + ], + "spans": [ + { + "bbox": [ + 245, + 188, + 294, + 213 + ], + "type": "image", + "image_path": "881497e03039a1651738526853488b53a3e959cd0973d27d2511327e57ac4951.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 213, + 286, + 219 + ], + "lines": [ + { + "bbox": [ + 253, + 213, + 286, + 219 + ], + "spans": [ + { + "bbox": [ + 253, + 213, + 286, + 219 + ], + "type": "text", + "content": "ESC (Ours)" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 253, + 220, + 277, + 226 + ], + "lines": [ + { + "bbox": [ + 253, + 220, + 277, + 226 + ], + "spans": [ + { + "bbox": [ + 253, + 220, + 277, + 226 + ], + "type": "text", + "content": "37.75" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 296, + 340 + ], + "type": "text", + "content": "age (1280×720) image using an RTX4090 GPU at FP32 precision 100 times. Memory usage is calculated by torch.cuda.max_memory_allocated function and FLOPs are measured by fvcore library. For the classic SR task, we present three variants: ESC, ESC-light, and ESC-FP. ESC and ESC-light prioritize reducing latency, while ESC-FP focuses on reducing FLOPs and parameter size." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 55, + 346, + 266, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 266, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 266, + 357 + ], + "type": "text", + "content": "4.2.1. Quantitative Results on the DIV2K dataset" + } + ] + } + ], + "index": 42 + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": "We compare our methods with various representative lightweight SR methods, including CNN-based models [77, 78], Transformers [36, 47, 56, 73, 75, 76, 82], and SSMs [20, 21] trained on the DIV2K dataset [1]. Notably, as shown in Table 2, our ESC achieves a notable 0.1dB PSNR improvement on the Urban " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "100 \\times 4" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " dataset, while being " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "8.7 \\times" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " faster than ATD-light. Moreover, ESC-light is " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "1.4 \\times" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " faster than ELAN-light, exhibiting an impressive performance improvement of 0.34dB PSNR in the Manga109 " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " dataset. These results exhibit that state-of-the-art performance can be achieved even when most self-attention are replaced with convolutions, suggesting that replacing self-attention with convolutions does not compromise representational power. Furthermore, ESC-FP outperforms MambaIRV2-light over 0.18dB on the Urban " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "100 \\times 4" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " dataset while reducing FLOPs and parameters by " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "inline_equation", + "content": "31\\%" + }, + { + "bbox": [ + 55, + 361, + 296, + 577 + ], + "type": "text", + "content": ", respectively. This result underscores that our design is effective even when reducing FLOPs and parameter size." + } + ] + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 582, + 239, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 239, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 239, + 593 + ], + "type": "text", + "content": "4.2.2. Visual Results on the DIV2K dataset" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 295, + 669 + ], + "type": "text", + "content": "We also visually compare our ESC with representative lightweight SR methods, including SRFormer-light [82], HiT-SRF [76], ATD-light [73], and MambaIRV2-light [21]. As shown in Figure 5, our ESC effectively restores details and achieves a high PSNR, demonstrating that our network also produces visually pleasant results." + } + ] + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 674, + 148, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 148, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 148, + 686 + ], + "type": "text", + "content": "4.2.3. Ablation Study" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We conduct an ablation study to justify our network's composition, as shown in Table 3. First, we introduce two vari-" + } + ] + } + ], + "index": 47 + }, + { + "type": "table", + "bbox": [ + 315, + 129, + 555, + 232 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "type": "text", + "content": "Table 3. Ablation studies on our ESC network. All comparisons are designed to have similar latencies to ESC. " + }, + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 70, + 554, + 125 + ], + "type": "text", + "content": ", and WS denote large kernel, dynamic kernel, and window size, respectively. All statistics and metrics are measured as detailed in Section 4.2." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 129, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 315, + 129, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 555, + 232 + ], + "type": "table", + "html": "
CasesLatency (ms)Memory (mb)#FLOPs (G)#Params (K)Set5×2Urban100×2
Only Self-attention128.2830456.673738.27/0.961733.23/0.9378
Only ConvAttns126.3514962.4121038.18/0.961332.91/0.9353
9×9 LK117.1831462.292438.33/0.961833.42/0.9392
17×17 LK126.6832768.997738.32/0.961833.40/0.9390
w/o LK Share and DK119.7835591.1195038.32/0.961833.37/0.9389
w/o LK Share120.9835592.0198538.32/0.961933.43/0.9392
w/o DK119.7835591.191138.31/0.961833.36/0.9388
WS16 and more layers118.0807603.6105438.24/0.961433.05/0.9360
WS24 and more layers116.5801583.599538.33/0.961833.28/0.9383
ESC (Ours)120.9831592.094738.35/0.961933.46/0.9395
", + "image_path": "8cd7aa61a6c2c2d442316bed03a12112c3b1c9709e49de097d34a31ee6d65923.jpg" + } + ] + } + ], + "index": 49, + "angle": 0, + "type": "table_body" + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": "ants, one that uses only the self-attention module and another that uses only the ConvAttn. The number of blocks and layers is adjusted to match the latency on par with the ESC. Both variants exhibit a performance drop compared to ESC, suggesting that leveraging both self-attention and ConvAttn offers the best performance-efficiency trade-off. Next, we vary the " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": "'s kernel size to " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "9 \\times 9" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "17 \\times 17" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " to determine the optimal size. In both cases, we observe a drop in performance, confirming that a kernel size of " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "13 \\times 13" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " is optimal. Then, we evaluate performance changes by altering the convolution type used in the ConvAttn. First, we observe performance drops when neither " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " sharing nor " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " is used. Then, we add " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " without sharing " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " and observe some performance improvement, though it still falls short of ESC. This result suggests that addressing the optimization difficulty that comes from the massive parameter size is more critical than the layer-specific long-range modeling. Finally, we leverage only " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " sharing without " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " and observe significant performance drops, affirming that harmonizing both " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 241, + 555, + 552 + ], + "type": "text", + "content": " is crucial. For the next step, we investigate the effect of the self-attention window size by setting it to 16 and 24 and adding extra layers. In every case, we observe a performance drop, which confirms that enlarging the window size to enhance self-attention is effective even when only a single self-attention is leveraged in each block." + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 560, + 471, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 471, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 471, + 572 + ], + "type": "text", + "content": "4.2.4. Analysis on ConvAttn Module" + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": "This section verifies that the proposed ConvAttn operates similarly to self-attention. Self-attention is known to extract shape-biased structural features [12, 48], unlike " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " convolution, which mainly extracts local texture and edges. Indeed, as shown in Figure 6 (a), " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "F^{idt}" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " focuses on local textures and edges while " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "F^{att}" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " concentrates on structural components. Next, to verify each convolution's receptive field, we visualize positive gradients " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{G}^{+} = \\frac{1}{NMC} \\sum_{p=1}^{NMC} \\max \\left(0, \\frac{\\partial I_{i,j}^{SR}}{\\partial F_p}\\right)" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " propagated from the output image's center pixel " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "(i,j)" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " to the inputs feature " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "(F)" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " of both " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " across " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " blocks, " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "text", + "content": " modules, and " + }, + { + "bbox": [ + 313, + 576, + 555, + 713 + ], + "type": "inline_equation", + "content": "C" + } + ] + } + ], + "index": 52 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24472" + } + ] + } + ], + "index": 53 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 105, + 555, + 262 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "text", + "content": "Table 4. Comparisons of classic SR methods trained on the large-scale DFLIP dataset. Mem, #FLOPs, and #Params denote memory usage, the number of FLOPs, and parameter size, respectively. " + }, + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "inline_equation", + "content": "\\S" + }, + { + "bbox": [ + 55, + 71, + 555, + 103 + ], + "type": "text", + "content": " denotes that we train the indicated methods following their descriptions. All statistics and metrics are measured as detailed in Section 4.2. The best result on each PSNR and SSIM is bolded." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 105, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 56, + 105, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 105, + 555, + 262 + ], + "type": "table", + "html": "
MethodScaleLatency (ms)Mem (mb)#FLOPs (G)#Params (K)PSNR / SSIM
Set5Set14B100Urban100Manga109
SRFormer-lt$ [82]×21838.11184236.385338.24/0.961534.13/0.921832.42/0.902633.37/0.938639.36/0.9787
ATD-lt$ [73]733.52839380.075338.29/0.961634.30/0.923032.43/0.902733.62/0.940139.60/0.9791
HiT-SRF$ [76]268.11804226.584738.31/0.961634.31/0.923032.45/0.903133.58/0.940439.69/0.9793
ESC (Ours)120.9831592.094738.34/0.961834.42/0.923532.50/0.903633.86/0.942439.73/0.9795
SRFormer-lt$ [82]×3668.3537105.486134.67/0.929730.75/0.848429.30/0.810829.10/0.870134.26/0.9498
ATD-lt$ [73]274.41258168.076034.71/0.930030.77/0.849329.33/0.811629.42/0.874334.61/0.9509
HiT-SRF$ [76]124.91464101.685534.69/0.929830.81/0.849329.32/0.811529.28/0.872934.72/0.9511
ESC (Ours)41.4385267.695534.85/0.931230.97/0.851129.41/0.813529.70/0.879934.94/0.9525
SRFormer-lt$ [82]×4327.832962.887332.49/0.899328.89/0.788727.76/0.742926.90/0.808631.25/0.9189
ATD-lt$ [73]189.7753100.176932.52/0.899528.93/0.789627.79/0.744327.18/0.815031.47/0.9208
HiT-SRF$ [76]82.1133158.086632.55/0.899728.96/0.789727.77/0.744327.07/0.813031.59/0.9208
ESC (Ours)21.9215149.296832.79/0.902529.06/0.792727.85/0.746627.45/0.822931.87/0.9239
", + "image_path": "09cbfbc1d699ba553d3c35a6fc58f4946202d716ef6cc0dd048381099c86487f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 272, + 122, + 335 + ], + "blocks": [ + { + "bbox": [ + 56, + 272, + 122, + 335 + ], + "lines": [ + { + "bbox": [ + 56, + 272, + 122, + 335 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 122, + 335 + ], + "type": "image", + "image_path": "3d9a970861a0d3d661c8844eeebcb52b27d54d432db42f6c4f04d6f355e91048.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 335, + 92, + 342 + ], + "lines": [ + { + "bbox": [ + 83, + 335, + 92, + 342 + ], + "spans": [ + { + "bbox": [ + 83, + 335, + 92, + 342 + ], + "type": "inline_equation", + "content": "F^{idt}" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 137, + 272, + 197, + 335 + ], + "blocks": [ + { + "bbox": [ + 137, + 272, + 197, + 335 + ], + "lines": [ + { + "bbox": [ + 137, + 272, + 197, + 335 + ], + "spans": [ + { + "bbox": [ + 137, + 272, + 197, + 335 + ], + "type": "image", + "image_path": "35affe7f1e3902deff6d3ea76d8b7e4369be2f9d84d91f5a23356019d3c62836.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 151, + 335, + 183, + 342 + ], + "lines": [ + { + "bbox": [ + 151, + 335, + 183, + 342 + ], + "spans": [ + { + "bbox": [ + 151, + 335, + 183, + 342 + ], + "type": "inline_equation", + "content": "\\bar{G}_{DK}^{+}" + }, + { + "bbox": [ + 151, + 335, + 183, + 342 + ], + "type": "text", + "content": " (DI: 48.77)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 216, + 272, + 292, + 335 + ], + "blocks": [ + { + "bbox": [ + 216, + 272, + 292, + 335 + ], + "lines": [ + { + "bbox": [ + 216, + 272, + 292, + 335 + ], + "spans": [ + { + "bbox": [ + 216, + 272, + 292, + 335 + ], + "type": "image", + "image_path": "a0a20bee06f34aed40681fb8d02503e30ec673ab4bd2c9479bae7f053a2caa77.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 335, + 251, + 343 + ], + "lines": [ + { + "bbox": [ + 241, + 335, + 251, + 343 + ], + "spans": [ + { + "bbox": [ + 241, + 335, + 251, + 343 + ], + "type": "inline_equation", + "content": "S_{DK}" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 57, + 342, + 122, + 407 + ], + "blocks": [ + { + "bbox": [ + 57, + 342, + 122, + 407 + ], + "lines": [ + { + "bbox": [ + 57, + 342, + 122, + 407 + ], + "spans": [ + { + "bbox": [ + 57, + 342, + 122, + 407 + ], + "type": "image", + "image_path": "ef548769d0b4d3f8ee5f1b118810a9e2b7f74654dd6cbc0a74e55214a0879ada.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 414, + 129, + 419 + ], + "lines": [ + { + "bbox": [ + 62, + 414, + 129, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 129, + 419 + ], + "type": "text", + "content": "(a) Feature Map Visualization" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "lines": [ + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "text", + "content": "Figure 6. Visualized feature map, gradients, and input-wise similarity. " + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "text", + "content": " is averaged across entire ConvAttn while processing the Lenna image in Set14. " + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "inline_equation", + "content": "\\tilde{G}^{+}" + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 422, + 295, + 465 + ], + "type": "text", + "content": " are calculated on the Urban100, and 10 samples in front of the Urban100, respectively." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 137, + 343, + 213, + 407 + ], + "blocks": [ + { + "bbox": [ + 137, + 343, + 213, + 407 + ], + "lines": [ + { + "bbox": [ + 137, + 343, + 213, + 407 + ], + "spans": [ + { + "bbox": [ + 137, + 343, + 213, + 407 + ], + "type": "image", + "image_path": "bfef0cf15bb9711874b8c09eae430559285602e1f9b8a67ecc615433292d6520.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 407, + 183, + 413 + ], + "lines": [ + { + "bbox": [ + 150, + 407, + 183, + 413 + ], + "spans": [ + { + "bbox": [ + 150, + 407, + 183, + 413 + ], + "type": "inline_equation", + "content": "G_{LK}^{+}" + }, + { + "bbox": [ + 150, + 407, + 183, + 413 + ], + "type": "text", + "content": " (DI: 62.23)" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 145, + 414, + 203, + 419 + ], + "lines": [ + { + "bbox": [ + 145, + 414, + 203, + 419 + ], + "spans": [ + { + "bbox": [ + 145, + 414, + 203, + 419 + ], + "type": "text", + "content": "(b) Gradient Visualization" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 217, + 343, + 292, + 407 + ], + "blocks": [ + { + "bbox": [ + 217, + 343, + 292, + 407 + ], + "lines": [ + { + "bbox": [ + 217, + 343, + 292, + 407 + ], + "spans": [ + { + "bbox": [ + 217, + 343, + 292, + 407 + ], + "type": "image", + "image_path": "4cd9179fe74725d498817dbaf4140e31bc437341ecdf6431922a4dd3488366b9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 220, + 414, + 282, + 419 + ], + "lines": [ + { + "bbox": [ + 220, + 414, + 282, + 419 + ], + "spans": [ + { + "bbox": [ + 220, + 414, + 282, + 419 + ], + "type": "text", + "content": "(c) Similarity Visualization" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": "channels, and corresponding diffusion index (DI). Figure 6 (b) clearly demonstrates that " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " extracts more long-range features than " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": ". Lastly, we visualize input-wise similarity " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "(S)" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " between average-pooled features extracted by " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": ". As demonstrated in Figure 6 (c), " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "S_{DK}" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " exhibits higher input-dependent diversity than " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "S_{LK}" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": ", confirming that " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "DK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": " extracts more input-dependent features than " + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "inline_equation", + "content": "LK" + }, + { + "bbox": [ + 55, + 476, + 296, + 559 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 567, + 181, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 181, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 181, + 578 + ], + "type": "text", + "content": "4.2.5. Local Attribution Map" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "We use the Local Attribution Map (LAM) [19] to verify that our networks can extract long-range dependencies despite using less self-attention. We compare our ESC with studies of various self-attention implementations that aim to reduce FLOPs while expanding the receptive field, such as global-level attention [56], permuted self-attention [82], and spatial-channel correlation [76]. Notably, as described in Figure 7, our ESC references the widest range of pixels to restore the red bounding box along with the highest diffusion index. This result exhibits that our network effectively captures long-range dependencies even though most self" + } + ] + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 316, + 271, + 393, + 348 + ], + "blocks": [ + { + "bbox": [ + 316, + 271, + 393, + 348 + ], + "lines": [ + { + "bbox": [ + 316, + 271, + 393, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 393, + 348 + ], + "type": "image", + "image_path": "28a54a66a3da98e00f804e36dd4a842bdff47d01e20c549a2bb0ab6a91691fb1.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 316, + 348, + 393, + 437 + ], + "blocks": [ + { + "bbox": [ + 316, + 348, + 393, + 437 + ], + "lines": [ + { + "bbox": [ + 316, + 348, + 393, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 348, + 393, + 437 + ], + "type": "image", + "image_path": "c8e9f9d3d98d47e5514293184ba7feecfbfd459092ac30ad7f0b42be2f224e20.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 329, + 437, + 379, + 449 + ], + "lines": [ + { + "bbox": [ + 329, + 437, + 379, + 449 + ], + "spans": [ + { + "bbox": [ + 329, + 437, + 379, + 449 + ], + "type": "text", + "content": "Urban100 - img063x4 Diffusion Index" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 314, + 451, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 314, + 451, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 314, + 451, + 555, + 483 + ], + "type": "text", + "content": "Figure 7. Comparisons of Local Attribution Map (LAM) [19] and the corresponding area of contribution and diffusion index. The best result on the diffusion index is bolded." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 393, + 271, + 457, + 310 + ], + "blocks": [ + { + "bbox": [ + 393, + 271, + 457, + 310 + ], + "lines": [ + { + "bbox": [ + 393, + 271, + 457, + 310 + ], + "spans": [ + { + "bbox": [ + 393, + 271, + 457, + 310 + ], + "type": "image", + "image_path": "a7133d47f6035acf0854cc473d0099708571c6bb8846fe78e982d669b40d647b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 398, + 310, + 437, + 347 + ], + "blocks": [ + { + "bbox": [ + 398, + 310, + 437, + 347 + ], + "lines": [ + { + "bbox": [ + 398, + 310, + 437, + 347 + ], + "spans": [ + { + "bbox": [ + 398, + 310, + 437, + 347 + ], + "type": "image", + "image_path": "5d65bda4de02635cd7d89c52dfb7ea503d9b222953dbd368470102f5454cef3e.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 348, + 425, + 354 + ], + "lines": [ + { + "bbox": [ + 407, + 348, + 425, + 354 + ], + "spans": [ + { + "bbox": [ + 407, + 348, + 425, + 354 + ], + "type": "text", + "content": "OmniSR" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 441, + 271, + 475, + 310 + ], + "blocks": [ + { + "bbox": [ + 441, + 271, + 475, + 310 + ], + "lines": [ + { + "bbox": [ + 441, + 271, + 475, + 310 + ], + "spans": [ + { + "bbox": [ + 441, + 271, + 475, + 310 + ], + "type": "image", + "image_path": "c56951aa9bc9019df0c7e4b8842da8444b4e3deee09e1e81cd5ef9bc40c8501d.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 440, + 310, + 475, + 347 + ], + "blocks": [ + { + "bbox": [ + 440, + 310, + 475, + 347 + ], + "lines": [ + { + "bbox": [ + 440, + 310, + 475, + 347 + ], + "spans": [ + { + "bbox": [ + 440, + 310, + 475, + 347 + ], + "type": "image", + "image_path": "6e985bc46b05a8fc2ae735c7fe1d73f6f84a7cd0407a095ac1925d7160a7fad8.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 348, + 470, + 354 + ], + "lines": [ + { + "bbox": [ + 443, + 348, + 470, + 354 + ], + "spans": [ + { + "bbox": [ + 443, + 348, + 470, + 354 + ], + "type": "text", + "content": "SRFormer-It" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 480, + 271, + 514, + 310 + ], + "blocks": [ + { + "bbox": [ + 480, + 271, + 514, + 310 + ], + "lines": [ + { + "bbox": [ + 480, + 271, + 514, + 310 + ], + "spans": [ + { + "bbox": [ + 480, + 271, + 514, + 310 + ], + "type": "image", + "image_path": "f69aa6221f006f8faf78404a3d331261e07ff16362d3433e7ddfff1d175f304f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 477, + 310, + 515, + 347 + ], + "blocks": [ + { + "bbox": [ + 477, + 310, + 515, + 347 + ], + "lines": [ + { + "bbox": [ + 477, + 310, + 515, + 347 + ], + "spans": [ + { + "bbox": [ + 477, + 310, + 515, + 347 + ], + "type": "image", + "image_path": "cf1db73828c316408c14858af8efd92a98eb797abe052ad76d913da64ea9dbd9.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 348, + 504, + 354 + ], + "lines": [ + { + "bbox": [ + 488, + 348, + 504, + 354 + ], + "spans": [ + { + "bbox": [ + 488, + 348, + 504, + 354 + ], + "type": "text", + "content": "iT-SRF" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 515, + 271, + 553, + 310 + ], + "blocks": [ + { + "bbox": [ + 515, + 271, + 553, + 310 + ], + "lines": [ + { + "bbox": [ + 515, + 271, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 515, + 271, + 553, + 310 + ], + "type": "image", + "image_path": "8dff5e1b36b6cc9d7cecc4c919c7bbbead87b13a5ba5dcc8369ccc162d1723d9.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 515, + 310, + 553, + 347 + ], + "blocks": [ + { + "bbox": [ + 515, + 310, + 553, + 347 + ], + "lines": [ + { + "bbox": [ + 515, + 310, + 553, + 347 + ], + "spans": [ + { + "bbox": [ + 515, + 310, + 553, + 347 + ], + "type": "image", + "image_path": "68a3b37e09115a0c7939e7431866efa9835c738ff170463f2c558338dd6797e8.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 521, + 348, + 547, + 354 + ], + "lines": [ + { + "bbox": [ + 521, + 348, + 547, + 354 + ], + "spans": [ + { + "bbox": [ + 521, + 348, + 547, + 354 + ], + "type": "text", + "content": "ESC (Ours)" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 501, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 536 + ], + "type": "text", + "content": " attentions are replaced by the ConvAttn module, suggesting that our ConvAttn module contributes to retaining the large receptive field of Transformers." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 552, + 524, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 524, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 524, + 563 + ], + "type": "text", + "content": "4.2.6. Quantitative Results on the DFLIP dataset" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Transformers exhibit significant performance improvements when trained on large-scale datasets [5, 15]. To verify our model also retains this benefit, we train and evaluate various representative Transformers – including SRFormerlight [82], ATD-light [73], and HiT-SRF [76] – alongside our ESC on a large dataset. For this purpose, we leverage DIV2K+Flickr2K+LSDIR+DiverSeg-IP (DFLIP) [1, 34, 45, 54] dataset. As illustrated in Table 4, our approach outperforms the other methods by a substantial margin across all scales and evaluation datasets. These results indicate that replacing self-attention with ConvAttn does not sacrifice the data scaling capability of Transformers." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24473" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 169, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 169, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 169, + 85 + ], + "type": "text", + "content": "4.3. Arbitrary-Scale SR" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 174 + ], + "type": "text", + "content": "To evaluate our proposed networks on arbitrary-scale SR tasks, we leverage the DIV2K-validation [1] dataset. Following previous research [8, 32], we assess PSNR in the RGB channel after cropping the boundaries according to the upscaling factor plus six. For arbitrary-scale SR tasks, we change " + }, + { + "bbox": [ + 55, + 89, + 296, + 174 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 89, + 296, + 174 + ], + "type": "text", + "content": " of our network and representative lightweight SR transforems [73, 76] into LTE [32]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 180, + 171, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 171, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 171, + 192 + ], + "type": "text", + "content": "4.3.1. Quantitative Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "text", + "content": "To demonstrate the outstanding representational capability of the proposed network, we train and evaluate the network on an arbitrary-scale SR task, which requires processing various upscaling factors with a single network. For quantitative comparison, we leverage commonly used RDN [78] and additional representative Transformers [73, 76] that we train. As described in Tables 5, our network continually achieves high performance at the scale both seen and unseen during training, even achieving a " + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "inline_equation", + "content": "0.09\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "text", + "content": " improvement over " + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\mathrm{RDN + LTE}" + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "text", + "content": " at " + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "inline_equation", + "content": "\\times 2" + }, + { + "bbox": [ + 55, + 195, + 296, + 351 + ], + "type": "text", + "content": ". This result underscores our ESC's exceptional representational capability. For quantitative results measured on other datasets, including Set5, Set14, B100, and Urban100, refer to Section 8 in the supplementary." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 385, + 296, + 446 + ], + "blocks": [ + { + "bbox": [ + 55, + 362, + 295, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 362, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 295, + 384 + ], + "type": "text", + "content": "Table 5. Quantitative comparison on arbitrary-scale SR task employing LTE [32] as upsampler. The best result on PSNR is bolded." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 385, + 296, + 446 + ], + "lines": [ + { + "bbox": [ + 56, + 385, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 385, + 296, + 446 + ], + "type": "table", + "html": "
MethodsSeenUnseen
×2×3×4×6×12×18×24×30
RDN+LTE [78]35.0431.3229.3327.0423.9522.4021.3620.64
ATD-lt+LTE§ [73]34.9831.2529.2727.0123.9422.3821.3720.64
HiT-SRF+LTE§ [76]35.1031.3529.3527.0523.9422.3921.3620.64
ESC+LTE (Ours)35.1331.3829.3727.0823.9722.4021.3820.64
", + "image_path": "de11fa0f491338a31aee418424ede1b031fb0879e648dba3ac1aba0c351f383d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 144, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 144, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 144, + 467 + ], + "type": "text", + "content": "4.3.2. Visual Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": "We also visually compare our network with RDN, HiT-SRF, and ATD-light. As shown in Figure 8, Our network effectively restores fine details that other methods fail on both " + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\times 12" + }, + { + "bbox": [ + 55, + 472, + 296, + 544 + ], + "type": "text", + "content": " scales. This result demonstrates that our network also excels at restoring visually present results on both in-distribution and out-of-distribution scales." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 217, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 217, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 217, + 565 + ], + "type": "text", + "content": "4.4. Real-World Super-Resolution" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "We further test our network on real-world SR tasks that aim to restore images with unknown degradations. To this end, we introduce a variant called ESC-Real and compare them to representative real-world SR methods, including RealESRGAN+ [59], SwinIR-Real [36], and DASR [61]. For our evaluation, we use the commonly employed RealSRSet dataset [72]. As shown in Table 9, our ESC-Real recovers more fine-grained details than the other methods while significantly reducing unpleasant artifacts in both graphic and natural images. This result demonstrates that our proposed methods are effective even when dealing with complex unknown degradations, underscoring our methods'" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 72, + 432, + 151 + ], + "blocks": [ + { + "bbox": [ + 323, + 72, + 432, + 151 + ], + "lines": [ + { + "bbox": [ + 323, + 72, + 432, + 151 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 432, + 151 + ], + "type": "image", + "image_path": "e48203c369aeef8b5aceb8276a16d186b4609294a80ad356b7dc6ebdb383a22f.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 349, + 151, + 406, + 158 + ], + "lines": [ + { + "bbox": [ + 349, + 151, + 406, + 158 + ], + "spans": [ + { + "bbox": [ + 349, + 151, + 406, + 158 + ], + "type": "text", + "content": "Urban100 - img011" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 158, + 431, + 186 + ], + "blocks": [ + { + "bbox": [ + 323, + 158, + 431, + 186 + ], + "lines": [ + { + "bbox": [ + 323, + 158, + 431, + 186 + ], + "spans": [ + { + "bbox": [ + 323, + 158, + 431, + 186 + ], + "type": "image", + "image_path": "69599c31241637d91e9c41a8477276e7ff0f8f475fc33b2a309fc997a4a63956.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 186, + 425, + 193 + ], + "lines": [ + { + "bbox": [ + 334, + 186, + 425, + 193 + ], + "spans": [ + { + "bbox": [ + 334, + 186, + 425, + 193 + ], + "type": "text", + "content": "RDN/17.71 HiT-SRF/17.56" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 323, + 194, + 431, + 221 + ], + "blocks": [ + { + "bbox": [ + 323, + 194, + 431, + 221 + ], + "lines": [ + { + "bbox": [ + 323, + 194, + 431, + 221 + ], + "spans": [ + { + "bbox": [ + 323, + 194, + 431, + 221 + ], + "type": "image", + "image_path": "5261f96ba48392c1cc34fe5cbbeab63e52dc9b89901d2e3dd11516fa527d6e8b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 222, + 430, + 229 + ], + "lines": [ + { + "bbox": [ + 331, + 222, + 430, + 229 + ], + "spans": [ + { + "bbox": [ + 331, + 222, + 430, + 229 + ], + "type": "text", + "content": "ATD-lt / 17.91 ESC (Ours) / 20.70" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 356, + 232, + 399, + 241 + ], + "lines": [ + { + "bbox": [ + 356, + 232, + 399, + 241 + ], + "spans": [ + { + "bbox": [ + 356, + 232, + 399, + 241 + ], + "type": "text", + "content": "(a) Scale " + }, + { + "bbox": [ + 356, + 232, + 399, + 241 + ], + "type": "inline_equation", + "content": "\\times 4" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 545, + 151 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 545, + 151 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 545, + 151 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 545, + 151 + ], + "type": "image", + "image_path": "004309304b2629b400b6dd09514120fb4ff0293752614acdf976ae2006651900.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 462, + 151, + 522, + 158 + ], + "lines": [ + { + "bbox": [ + 462, + 151, + 522, + 158 + ], + "spans": [ + { + "bbox": [ + 462, + 151, + 522, + 158 + ], + "type": "text", + "content": "DIV2K-valid - 0845" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 437, + 159, + 492, + 186 + ], + "blocks": [ + { + "bbox": [ + 437, + 159, + 492, + 186 + ], + "lines": [ + { + "bbox": [ + 437, + 159, + 492, + 186 + ], + "spans": [ + { + "bbox": [ + 437, + 159, + 492, + 186 + ], + "type": "image", + "image_path": "5aae8ff9766fd6649fa991747e5884d70dc73248b0214f6d1217292993674d8c.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 436, + 194, + 491, + 221 + ], + "blocks": [ + { + "bbox": [ + 447, + 187, + 541, + 194 + ], + "lines": [ + { + "bbox": [ + 447, + 187, + 541, + 194 + ], + "spans": [ + { + "bbox": [ + 447, + 187, + 541, + 194 + ], + "type": "text", + "content": "RDN/16.25 HiT-SRF/17.28" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 436, + 194, + 491, + 221 + ], + "lines": [ + { + "bbox": [ + 436, + 194, + 491, + 221 + ], + "spans": [ + { + "bbox": [ + 436, + 194, + 491, + 221 + ], + "type": "image", + "image_path": "9ca1c858354b8bb8f7d7f76d144d632ff1ad8a21a4aa2771a8445659bc2f233d.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 444, + 222, + 483, + 229 + ], + "lines": [ + { + "bbox": [ + 444, + 222, + 483, + 229 + ], + "spans": [ + { + "bbox": [ + 444, + 222, + 483, + 229 + ], + "type": "text", + "content": "ATD-lt / 16.77" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 466, + 232, + 515, + 241 + ], + "lines": [ + { + "bbox": [ + 466, + 232, + 515, + 241 + ], + "spans": [ + { + "bbox": [ + 466, + 232, + 515, + 241 + ], + "type": "text", + "content": "(b) Scale " + }, + { + "bbox": [ + 466, + 232, + 515, + 241 + ], + "type": "inline_equation", + "content": "\\times 12" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 493, + 194, + 545, + 221 + ], + "blocks": [ + { + "bbox": [ + 493, + 194, + 545, + 221 + ], + "lines": [ + { + "bbox": [ + 493, + 194, + 545, + 221 + ], + "spans": [ + { + "bbox": [ + 493, + 194, + 545, + 221 + ], + "type": "image", + "image_path": "4454287fc45e67358b390d1ed12973c9fddb5d8d08b416b79b0c5a7fa82588d1.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 493, + 222, + 543, + 229 + ], + "lines": [ + { + "bbox": [ + 493, + 222, + 543, + 229 + ], + "spans": [ + { + "bbox": [ + 493, + 222, + 543, + 229 + ], + "type": "text", + "content": "ESC (Ours) / 17.82" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 317, + 283, + 552, + 380 + ], + "blocks": [ + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "type": "text", + "content": "Figure 8. Visual comparisons on arbitrary-scale SR for both indistribution " + }, + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "type": "text", + "content": " and out-of-distribution " + }, + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "type": "inline_equation", + "content": "(\\times 12)" + }, + { + "bbox": [ + 313, + 245, + 555, + 277 + ], + "type": "text", + "content": " scales. The best result on PSNR is bolded." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 283, + 552, + 380 + ], + "lines": [ + { + "bbox": [ + 317, + 283, + 552, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 283, + 552, + 380 + ], + "type": "image", + "image_path": "ce69f6c552a458a04e971142c286f89dcb172cf8a519dc1efa8c5cdbe994c86f.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 407, + 380, + 470, + 388 + ], + "lines": [ + { + "bbox": [ + 407, + 380, + 470, + 388 + ], + "spans": [ + { + "bbox": [ + 407, + 380, + 470, + 388 + ], + "type": "text", + "content": "(a) RealSRSet - 0030" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 317, + 392, + 552, + 488 + ], + "blocks": [ + { + "bbox": [ + 317, + 392, + 552, + 488 + ], + "lines": [ + { + "bbox": [ + 317, + 392, + 552, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 392, + 552, + 488 + ], + "type": "image", + "image_path": "13859fafde256993ad2a62390e20e941d31a546487edda44d7296f90d2e04ca8.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 408, + 490, + 469, + 499 + ], + "lines": [ + { + "bbox": [ + 408, + 490, + 469, + 499 + ], + "spans": [ + { + "bbox": [ + 408, + 490, + 469, + 499 + ], + "type": "text", + "content": "(b) RealSRSet - dog" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 326, + 513, + 541, + 524 + ], + "lines": [ + { + "bbox": [ + 326, + 513, + 541, + 524 + ], + "spans": [ + { + "bbox": [ + 326, + 513, + 541, + 524 + ], + "type": "text", + "content": "Figure 9. Visual comparisons on real-world SR " + }, + { + "bbox": [ + 326, + 513, + 541, + 524 + ], + "type": "inline_equation", + "content": "(\\times 4)" + }, + { + "bbox": [ + 326, + 513, + 541, + 524 + ], + "type": "text", + "content": " tasks." + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 567 + ], + "type": "text", + "content": "outstanding representational capability. Quantitative results on real-world SR tasks are presented in Section 9 in the supplementary." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 388, + 586 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we propose ESC, a novel lightweight SR network designed to mitigate the excessive memory overhead of Transformers. Our extensive experiments demonstrate that most self-attention layers can be effectively replaced with carefully designed convolutions to improve efficiency. Moreover, we introduce Flash Attention into lightweight SR tasks, significantly enhancing performance without complicated modifications to self-attention. As a result, we fully exploit the advantages of Transformers by prioritizing improvements in their efficiency." + } + ] + } + ], + "index": 34 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24474" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the National Research Foundation (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 115, + 130 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 115, + 130 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 115, + 130 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 137, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 137, + 294, + 170 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 294, + 170 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 294, + 170 + ], + "type": "text", + "content": "[1] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In CVPRW, 2017. 6, 7, 8, 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 171, + 294, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 294, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 294, + 224 + ], + "type": "text", + "content": "[2] Yang Ai, Xiaoqiang Zhou, Huaibo Huang, Xiaotian Han, Zhengyu Chen, Quanzeng You, and Hongxia Yang. Dreamclear: High-capacity real-world image restoration with privacy-safe dataset curation. NeurIPS, 37:55443-55469, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 226, + 294, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 226, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 226, + 294, + 269 + ], + "type": "text", + "content": "[3] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. 2012. 5, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 295, + 313 + ], + "type": "text", + "content": "[4] Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang. Toward real-world single image super-resolution: A new benchmark and a new model. In ICCV, pages 3086-3095, 2019. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 315, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 315, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 62, + 315, + 294, + 358 + ], + "type": "text", + "content": "[5] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, pages 12299-12310, 2021. 1, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "type": "text", + "content": "[6] Honghao Chen, Xiangxiang Chu, Yongjian Ren, Xin Zhao, and Kaiqi Huang. Pelk: Parameter-efficient large kernel convnets with peripheral convolution. In CVPR, pages 5557-5567, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 294, + 445 + ], + "type": "text", + "content": "[7] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In CVPR, pages 22367-22377, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 448, + 294, + 480 + ], + "type": "text", + "content": "[8] Yinbo Chen, Sifei Liu, and Xiaolong Wang. Learning continuous image representation with local implicit image function. In CVPR, pages 8628-8638, 2021. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 481, + 294, + 514 + ], + "type": "text", + "content": "[9] Zheng Chen, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Cross aggregation transformer for image restoration. In NeurIPS, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 557 + ], + "type": "text", + "content": "[10] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. NeurIPS, 35:16344-16359, 2022. 1, 2, 3, 4, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 57, + 559, + 294, + 591 + ], + "type": "text", + "content": "[11] Keyan Ding, Kede Ma, Shiqi Wang, and Eero P Simoncelli. Image quality assessment: Unifying structure and texture similarity. TPAMI, 44(5):2567-2581, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 635 + ], + "type": "text", + "content": "[12] Xiaohan Ding, Xiangyu Zhang, Jungong Han, and Guiguang Ding. Scaling up your kernels to 31x31: Revisiting large kernel design in cnns. In CVPR, pages 11963-11975, 2022. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "text", + "content": "[13] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE TPAMI, 38(2):295-307, 2015. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[14] Juechu Dong, Boyuan Feng, Driss Guessous, Yanbo Liang, and Horace He. Flex attention: A programming model for generating optimized attention kernels. arXiv preprint arXiv:2412.05496, 2024. 2, 3, 4, 5, 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 137 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2020. 1, 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[16] Zongcai Du, Jie Liu, Jie Tang, and Gangshan Wu. Anchors-based plain net for mobile image super-resolution. In CVPRW, pages 2494-2502, 2021. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 196 + ], + "type": "text", + "content": "[17] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 228 + ], + "type": "text", + "content": "[18] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. *ICLR*, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 555, + 274 + ], + "type": "text", + "content": "[19] Jinjin Gu and Chao Dong. Interpreting super-resolution networks with local attribution maps. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9199-9208, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 318 + ], + "type": "text", + "content": "[20] Hang Guo, Jinmin Li, Tao Dai, Zhihao Ouyang, Xudong Ren, and Shu-Tao Xia. Mambair: A simple baseline for image restoration with state-space model. In ECCV, 2024. 3, 5, 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 353 + ], + "type": "text", + "content": "[21] Hang Guo, Yong Guo, Yaohua Zha, Yulun Zhang, Wenbo Li, Tao Dai, Shu-Tao Xia, and Yawei Li. Mambairv2: Attentive state space restoration. CVPR, 2025. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[22] Qi Han, Zejia Fan, Qi Dai, Lei Sun, Ming-Ming Cheng, Jiaying Liu, and Jingdong Wang. On the connection between local attention and dynamic depth-wise convolution. *ICLR*, 2022, 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 432 + ], + "type": "text", + "content": "[23] Byeongho Heo, Song Park, Dongyoon Han, and Sangwoo Yun. Rotary position embedding for vision transformer. In ECCV, pages 289-305. Springer, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 475 + ], + "type": "text", + "content": "[24] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 477, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 553, + 532 + ], + "type": "text", + "content": "[25] Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andretto, and Hartwig Adam. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 566 + ], + "type": "text", + "content": "[26] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In CVPR, pages 5197-5206, 2015. 5, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 600 + ], + "type": "text", + "content": "[27] Junjie Ke, Qifei Wang, Yilin Wang, Peyman Milanfar, and Feng Yang. Musiq: Multi-scale image quality transformer. In ICCV, pages 5148-5157, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 634 + ], + "type": "text", + "content": "[28] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Deeplyrecursive convolutional network for image super-resolution. In CVPR, pages 1637-1645, 2016. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 668 + ], + "type": "text", + "content": "[29] Jiwon Kim, Jung Kwon Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, pages 1646-1654, 2016. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 690 + ], + "type": "text", + "content": "[30] Diederik P Kingma and Jimmy Ba. Adam: a method for stochastic optimization. In ICLR, 2014. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[31] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representa" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24475" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "tions revisited. In ICML, pages 3519-3529. PMLR, 2019. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 128 + ], + "type": "text", + "content": "[32] Jaewon Lee and Kyong Hwan Jin. Local texture estimator for implicit representation function. In CVPR, pages 1929-1938, 2022. 8, 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 161 + ], + "type": "text", + "content": "[33] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In IJCAI, pages 1089-1097, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 205 + ], + "type": "text", + "content": "[34] Yawei Li, Kai Zhang, Jingyun Liang, Jiezhang Cao, Ce Liu, Rui Gong, Yulun Zhang, Hao Tang, Yun Liu, Denis Demandolx, et al. Lsdir: A large scale dataset for image restoration. In CVPRW, pages 1775-1787, 2023. 7, 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 239 + ], + "type": "text", + "content": "[35] Zhen Li, Jinglei Yang, Zheng Liu, Xiaomin Yang, Gwang-gil Jeon, and Wei Wu. Feedback network for image superresolution. In CVPR, pages 3867-3876, 2019. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[36] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In ICCVW, pages 1833-1844, 2021. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 284, + 294, + 316 + ], + "type": "text", + "content": "[37] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPRW, pages 136-144, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 317, + 294, + 370 + ], + "type": "text", + "content": "[38] Shiwei Liu, Tianlong Chen, Xiaohan Chen, Xuxi Chen, Qiao Xiao, Boqian Wu, Tommi Kärkkäinen, Mykola Pechenizkiy, Decebal Mocanu, and Zhangyang Wang. More convnets in the 2020s: Scaling up kernels beyond 51x51 using sparsity. ICLR, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 294, + 426 + ], + "type": "text", + "content": "[39] Yong Liu, Hang Dong, Boyang Liang, Songwei Liu, Qingji Dong, Kai Chen, Fangmin Chen, Lean Fu, and Fei Wang. Unfolding once is enough: A deployment-friendly transformer unit for super-resolution. In ACMMM, pages 7952-7960, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 492 + ], + "type": "text", + "content": "[41] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 547 + ], + "type": "text", + "content": "[42] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, pages 416-423. IEEE, 2001. 5, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 592 + ], + "type": "text", + "content": "[43] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia tools and applications, 76:21811-21838, 2017. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 593, + 294, + 624 + ], + "type": "text", + "content": "[44] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 20(3): 209-212, 2012. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 680 + ], + "type": "text", + "content": "[45] Go Ohtani, Ryu Tadokoro, Ryosuke Yamada, Yuki M Asano, Iro Laina, Christian Rupprecht, Nakamasa Inoue, Rio Yokota, Hirokatsu Kataoka, and Yoshimitsu Aoki. Rethinking image super-resolution from training data perspectives. In ECCV, pages 19-36. Springer, 2024. 7, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[46] Karam Park and Nam Ik Cho. Partial filter-sharing: Improved parameter-sharing method for single image super-resolution networks. In WACV, pages 2653-2663. IEEE, 2025. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[47] Karam Park, Jae Woong Soh, and Nam Ik Cho. Efficient attention-sharing information distillation transformer for lightweight single image super-resolution. In AAAI, 2025. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 138 + ], + "type": "text", + "content": "[48] Namuk Park and Songkuk Kim. How do vision transformers work? In ICLR, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 194 + ], + "type": "text", + "content": "[49] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 249 + ], + "type": "text", + "content": "[50] Wenzhe Shi, Jose Caballero, Ferenc Huszár, Johannes Totz, Andrew P Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In CVPR, pages 1874-1883, 2016. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[51] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 326 + ], + "type": "text", + "content": "[52] Long Sun, Jinshan Pan, and Jinhui Tang. Shufflemixer: An efficient convnet for image super-resolution. NeurIPS, 35: 17314-17326, 2022. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 360 + ], + "type": "text", + "content": "[53] Ying Tai, Jian Yang, and Xiaoming Liu. Image superresolution via deep recursive residual network. In CVPR, pages 3147-3155, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[54] Radu Timofte, Eirikur Agustsson, Luc Van Gool, Ming-Hsuan Yang, and Lei Zhang. Ntire 2017 challenge on single image super-resolution: Methods and results. In CVPRW, pages 114-125, 2017. 7, 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. NeurIPS, 30, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 470 + ], + "type": "text", + "content": "[56] Hang Wang, Xuanhong Chen, Bingbing Ni, Yutian Liu, and Liu jinfan. Omni aggregation networks for lightweight image super-resolution. In CVPR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "text", + "content": "[57] Jianyi Wang, Kelvin CK Chan, and Chen Change Loy. Exploring clip for assessing the look and feel of images. In AAAI, pages 2555-2563, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 553, + 547 + ], + "type": "text", + "content": "[58] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, pages 606-615, 2018. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 581 + ], + "type": "text", + "content": "[59] Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. Real-esrgan: Training real-world blind super-resolution with pure synthetic data. In ICCV, pages 1905-1914, 2021. 8, 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 624 + ], + "type": "text", + "content": "[60] Pengxu Wei, Ziwei Xie, Hannan Lu, Zongyuan Zhan, Qixiang Ye, Wangmeng Zuo, and Liang Lin. Component divide-and-conquer for real-world image super-resolution. In ECCV, pages 101-117. Springer, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[61] Yunxuan Wei, Shuhang Gu, Yawei Li, Radu Timofte, Longcun Jin, and Hengjie Song. Unsupervised real-world image super resolution via domain-distance aware training. In CVPR, pages 13385-13394, 2021. 8" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[62] Gang Wu, Junjun Jiang, Junpeng Jiang, and Xianming Liu. Transforming image super-resolution: a convformer-based efficient approach. IEEE Transactions on Image Processing, 2024. 3, 4" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24476" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[63] Chengxing Xie, Xiaoming Zhang, Linze Li, Haiteng Meng, Tianlin Zhang, Tianrui Li, and Xiaole Zhao. Large kernel distillation network for efficient single image superresolution. In CVPRW, pages 1283-1292, 2023. 3, 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[64] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Maniaq: Multi-dimension attention network for no-reference image quality assessment. In CVPR, pages 1191-1200, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[65] Zhuoyi Yang, Jiayan Teng, Wendi Zheng, Ming Ding, Shiyu Huang, Jiazheng Xu, Yuanming Yang, Wenyi Hong, Xiaohan Zhang, Guanyu Feng, et al. Cogvideox: Text-to-video diffusion models with an expert transformer. arXiv preprint arXiv:2408.06072, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 263 + ], + "type": "text", + "content": "[66] Jingfeng Yao, Bin Yang, and Xinggang Wang. Reconstruction vs. generation: Taming optimization dilemma in latent diffusion models. In CVPR, pages 15703-15712, 2025. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 308 + ], + "type": "text", + "content": "[67] Jinsu Yoo, Taehoon Kim, Sihaeng Lee, Seung Hwan Kim, Honglak Lee, and Tae Hyun Kim. Enriched cnn-transformer feature aggregation networks for super-resolution. In WACV, pages 4956-4965, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[68] Seokju Yun and Youngmin Ro. Shvit: Single-head vision transformer with memory efficient macro design. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5756-5767, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 399 + ], + "type": "text", + "content": "[69] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: efficient attention transformer for high-resolution image restoration. In CVPR, pages 5728-5739, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[70] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In Curves and Surfaces: 7th International Conference, Avignon, France, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[71] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[72] Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. Designing a practical degradation model for deep blind image super-resolution. In CVPR, pages 4791-4800, 2021. 8, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 588 + ], + "type": "text", + "content": "[73] Leheng Zhang, Yawei Li, Xingyu Zhou, Xiaorui Zhao, and Shuhang Gu. Transcending the limit of local window: Advanced super-resolution transformer with adaptive token dictionary. In CVPR, pages 2856-2865, 2024. 2, 5, 6, 7, 8, 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 633 + ], + "type": "text", + "content": "[74] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, pages 586-595, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 678 + ], + "type": "text", + "content": "[75] Xindong Zhang, Hui Zeng, Shi Guo, and Lei Zhang. Efficient long-range attention network for image superresolution. In ECCV, pages 649-667. Springer, 2022. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[76] Xiang Zhang, Yulun Zhang, and Fisher Yu. Hit-sr: Hierarchical transformer for efficient attention image super-resolution. In ECCV, 2024. 2, 5, 6, 7, 8, 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 319 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[77] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, pages 286-301, 2018. 1, 2, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 150 + ], + "type": "text", + "content": "[78] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, pages 2472-2481, 2018. 2, 5, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 152, + 553, + 185 + ], + "type": "text", + "content": "[79] Hengyuan Zhao, Xiangtao Kong, Jingwen He, Yu Qiao, and Chao Dong. Efficient image super-resolution using pixel attention. In ECCVW, pages 56-72. Springer, 2020. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 186, + 553, + 228 + ], + "type": "text", + "content": "[80] Kun Zhou, Xinyu Lin, LIU Zhonghang, Xiaoguang Han, and Jiangbo Lu. Ups: Unified projection sharing for lightweight single-image super-resolution and beyond. In NeurIPS, 2024. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "type": "text", + "content": "[81] Lin Zhou, Haoming Cai, Jinjin Gu, Zheyuan Li, Yingqi Liu, Xiangyu Chen, Yu Qiao, and Chao Dong. Efficient image super-resolution using vast-receptive-field attention. In EC-CVW, pages 256-272. Springer, 2022. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 553, + 319 + ], + "type": "text", + "content": "[82] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In ICCV, pages 12780–12791, 2023. 2, 4, 5, 6, 7, 1, 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24477" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_content_list.json b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c97c688bc611ff4c4260c0d6a87d73f7df710ba --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_content_list.json @@ -0,0 +1,1576 @@ +[ + { + "type": "text", + "text": "End-to-End Driving with Online Trajectory Evaluation via BEV World Model", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yingyan Li* Yuqi Wang* Yang Liu Jiawei He Lue Fan* Zhaoxiang Zhang* NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA) State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 151, + 180, + 844, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving has achieved remarkable progress by integrating perception, prediction, and planning into a fully differentiable framework. Yet, to fully realize its potential, an effective online trajectory evaluation is in dispensable to ensure safety. By forecasting the future outcomes of a given trajectory, trajectory evaluation becomes much more effective. This goal can be achieved by employing a world model to capture environmental dynamic and predict future states. Therefore, we propose an end-to-end driving framework WoTE, which leverages a BE World model to predict future BEV states for Trajectory Evaluation. The proposed BEV world model is latency efficient compared to image-level world models and can be seamlessly supervised using off-the-shelf BEV-space traffic simulators. We validate our framework on both the NAVSIM benchmark and the closed-loop Bench2Drive benchmark based on the CARLA simulator, achieving state-of-the-art performance. Code is released at https://github.com/liyingyanUCAS/WoTE.", + "bbox": [ + 89, + 316, + 483, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 636, + 220, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "End-to-end autonomous driving [6, 30, 35] has achieved remarkable success in recent years and garnered widespread attention. While most approaches [21, 23, 25] primarily focus on predicting a high-quality trajectory, recent works [7, 27] suggest that predicting multiple trajectories simultaneously better reflects the multi-modal nature of driving and leads to improved performance. Given these multiple possible trajectories, it becomes crucial to effectively evaluate them in order to ensure the safety and reliability of end-to-end autonomous driving systems.", + "bbox": [ + 89, + 662, + 482, + 813 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional trajectory evaluation methods are rule-based [10, 14, 20, 37] and rely on perception results, such as bounding boxes and maps, to assess trajectories. These methods are sensitive to perception inaccuracies and may", + "bbox": [ + 89, + 814, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e3c1fc45288ff454935d5c641bb01389b75cbc08501912a7f1adb1ddf7dd0df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 287, + 901, + 352 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d73d50b4b000a502be8c230e7e71a5b7e1b34d3f498c0b249cdf8fbc45eb9fbb.jpg", + "image_caption": [ + "Figure 1. The illustration of end-to-end trajectory evaluation with BEV world model. (a) Previous end-to-end driving methods primarily focus on learning a high-quality trajectory. (b) Our WoTE consists of end-to-end trajectory prediction and evaluation. First, we predict multi-modal trajectories. Next, given these trajectories, we leverage the BEV world model to predict the corresponding future BEV states. The reward model evaluates these trajectories based on the predicted states and assigns rewards, selecting the highest-reward trajectory." + ], + "image_footnote": [], + "bbox": [ + 519, + 353, + 901, + 460 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "not be directly optimized in an end-to-end manner. Recently, some end-to-end driving approaches [7, 27] have incorporated trajectory evaluation. Their evaluations rely on the current state. However, evaluating a trajectory candidate is particularly challenging without knowledge of the future led by this trajectory. Similar to human drivers, who anticipate futures before making decisions, a trajectory evaluation network would benefit from predicting the corresponding future states of trajectory candidates.", + "bbox": [ + 511, + 628, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively predict future states, it is intuitive to adopt methodologies from reinforcement learning, where world models have successfully facilitated future state predictions. In reinforcement learning, model-based methods [5, 18, 32] have consistently shown superior performance compared to model-free approaches [17, 29], largely due to their ability of modeling the environmental dynamics and reasoning about future scenarios through world models. Inspired by this success, we propose adopting a world model [15, 19, 38-40]", + "bbox": [ + 511, + 763, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}\\ast$ : Co-first author, $^{2}$ : Co-advisor. Email: liyingyan2021@ia.ac.cn", + "bbox": [ + 107, + 886, + 464, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27137", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "within end-to-end autonomous driving frameworks to facilitate accurate prediction of future states, thus significantly improving trajectory evaluation.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, leveraging the world model for trajectory evaluation presents some technical issues. First, a good representation of future scenarios is needed. Many existing world models for driving [19, 38, 40] predict the images of future scenarios with diffusion models [33], which can be time-consuming and not suitable for real-time driving. Second, we lack supervision for future states. Supervising the predicted future states of the world model is difficult because world models need to imagine multiple future states based on the multiple trajectory candidates, while only one future state is available in the real-world datasets.", + "bbox": [ + 89, + 141, + 485, + 308 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these technical issues, we propose WoTE as shown in Fig.1. We propose to predict future states in the BEV space. The information in BEV space is much more compact than in raw images, which enables a more efficient single-step feed-forward future prediction than the time-consuming multi-step denoising in previous driving world models [19, 38, 40, 45]. This characteristic of BEV space makes WoTE highly suitable for real-time driving applications. Additionally, the predicted future BEV states can be easily supervised using off-the-shelf BEV-space traffic simulators such as nuPlan [3, 16]. These simulators model the scenario in BEV space by representing the map and objects using BEV semantic maps, creating a dynamic simulation of the driving environment. Consequently, we can use the BEV semantic maps from these simulated scenarios as supervision for the prediction of the BEV world model. Moreover, these simulators are equipped with well-established evaluation rules to assess the simulated future BEV scenarios, which provide target rewards.", + "bbox": [ + 91, + 313, + 485, + 599 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our framework on both NAVSIM [11] benchmark and closed-loop Bench2Drive [24] benchmark based on the CARLA [13] simulator and achieve state-of-the-art performance. Our experiments show that incorporating future states into trajectory evaluation is essential for improved performance. Additionally, the trajectory evaluation module, built on the world model shows great generalization ability across diverse trajectories. Our contributions are as follows:", + "bbox": [ + 89, + 604, + 485, + 727 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We highlight the importance of utilizing imagined future states for trajectory evaluation in end-to-end autonomous driving.", + "- We introduce WoTE, which consists of an end-to-end trajectory evaluation module based on a BEV world model. The world model in BEV space enables efficient real-time future prediction and overcomes the scarcity of multi-future supervisions.", + "- WoTE is validated on both the NAVSIM [11] benchmark and the closed-loop CARLA-based Bench2Drive [24] benchmark, achieving state-of-the-art performance." + ], + "bbox": [ + 91, + 731, + 482, + 896 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. End-to-end Autonomous Driving", + "text_level": 1, + "bbox": [ + 513, + 114, + 802, + 132 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "End-to-end autonomous driving [7, 22, 25, 43] is to train a model that directly maps sensor inputs to trajectories or control signals instead of decomposing the driving task into traditional sub-tasks. These works can be divided into two categories according to the training paradigm: imitation-learning-based and reinforcement-learning-based. Imitation learning is the most common practice in end-to-end autonomous driving. Their predicted trajectories are supervised by the expert trajectories. For instance, P3 [34], takes a differentiable semantic occupancy as the intermediate representation and learns actions from human trajectories. Following this, ST-P3 [21] learns spatial-temporal features for perception, prediction and planning tasks simultaneously. UniAD [22] proposed goal-oriented planning, combining and leveraging advantages of perception and motion prediction modules. VAD [25] introduces vectorized representations for end-to-end planning. There are also some reinforcement-learning-based methods. For example, MaRLn [36] utilizes implicit affordances to develop a reinforcement learning algorithm. LBC [4] trained a vision-based sensorimotor agent from the privileged agent. Combining imitation learning and reinforcement learning, Hydra-MDP [27] utilizes knowledge distillation to obtain supervisions from both rule-based planners and human drivers. It is a promising research direction as more supervisions are utilized.", + "bbox": [ + 511, + 138, + 906, + 517 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Trajectory Evaluation", + "text_level": 1, + "bbox": [ + 513, + 529, + 720, + 546 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Trajectory evaluation is essential for ensuring the reliability and safety of autonomous driving systems. We divide the main trajectory evaluation methods into two categories: model-free and model-based. Model-free methods directly assess the learned policy using observed data without explicitly modeling environmental dynamics. For instance, Monte Carlo policy evaluation [1] uses the empirical mean return as the metric. The value function approximates the expectation based on sampled returns. Temporal-difference learning [9] uses bootstrapping, where the current estimate of the value function is used to generate target values for updating the value function. Subsequently, model-based trajectory evaluation methods [12] have gained increasing popularity in recent years. By leveraging learned world models that capture environmental dynamics, these methods enable more accurate and predictive policy evaluation. For example, PPUU [18] measures policy performance by quantifying uncertainty in predicted future states. Rails [5] evaluates trajectories through a tabular dynamic programming mechanism, and MARL-CCE [32] employs learned environmental dynamics to assess multi-agent policies in driving scenarios. However, these methods often rely on explicit trajectory representations and non-differentiable metrics, which limits", + "bbox": [ + 511, + 553, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27138", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "their end-to-end optimization capabilities. In contrast, our approach evaluates trajectories using BEV features and encoded trajectory embeddings, enabling fully differentiable and thus end-to-end trajectory evaluation.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 165, + 181, + 181 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce four key components of WoTE: Trajectory Predictor (Sec. 3.1), BEV World Model (Sec. 3.2), Reward Model (Sec. 3.3) and BEV Space Supervision (Sec. 3.4). As illustrated in Figure 2, our method integrates future state prediction via the BEV World Model and reward prediction via the Reward Model within a unified framework. Specifically, the Trajectory Predictor encodes the multi-modal observations into BEV state, represented as a BEV feature map, and provides multiple trajectory proposals. The BEV World Model predicts corresponding future BEV states based on these trajectory proposals. With the help of future states, the Reward Model predicts the reward of each trajectory. Finally, we choose the trajectory with the highest reward as the final trajectory.", + "bbox": [ + 89, + 191, + 483, + 402 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Trajectory Predictor", + "text_level": 1, + "bbox": [ + 89, + 412, + 285, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2 (a), the Trajectory Predictor aims to predict multiple trajectory candidates. First, the BEV encoder encodes the multi-modal sensor inputs into a unified BEV feature map. Then, given trajectory anchors, the trajectory refinement module refines trajectories based on the BEV feature map. We will now describe each part in detail.", + "bbox": [ + 89, + 434, + 485, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal BEV Encoder. Following TransFuser [31], our method leverages multi-modal sensor inputs, including LiDAR and multi-view RGB images. We use a BEV encoder [31] to transform the current multi-modal inputs into a unified BEV feature map $\\mathbf{B} \\in \\mathbb{R}^{h \\times w \\times c}$ . For convenience in the following text, we refer to the BEV feature map as the BEV state. $h$ and $w$ represent the height and width of the BEV state, while $c$ denotes the number of channels.", + "bbox": [ + 89, + 544, + 483, + 664 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory Anchors. Following methods in [7, 27], we generate trajectory anchors $\\tau$ by K-Means clustering. This clustering is performed on all the expert trajectories collected from the training dataset. The number of trajectory anchors is denoted as $N$ .", + "bbox": [ + 89, + 684, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory Refinement. As illustrated in Figure 2 (a), the trajectory refinement module takes the trajectory anchors and the current BEV states as inputs to predict the refined trajectories. Specifically, we first use a trajectory encoder TE, which is implemented by an MLP, to encode the trajectory anchors $\\tau$ into feature vectors. Then, these feature vectors are used as queries to perform cross attention with the current BEV state B, which serves as both the key and", + "bbox": [ + 89, + 779, + 485, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "value. The output of the cross-attention is then fed into an MLP head to predict the offsets. The offset is added to the trajectory anchors to produce the refined trajectories. The entire process can be expressed as:", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\tau} = \\tau + \\operatorname {M L P} (\\text {C r o s s A t t e n t i o n} (\\mathbf {T E} (\\tau), \\mathbf {B}, \\mathbf {B})), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 162, + 906, + 178 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{\\tau}$ is the refined trajectories.", + "bbox": [ + 511, + 186, + 740, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. BEV World Model", + "text_level": 1, + "bbox": [ + 511, + 210, + 692, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we use the BEV World Model to predict the future BEV states, as illustrated in Figure 2 (b). The process is divided into two steps. First, the current BEV state and refined trajectories are combined to form the inputs for the world model. Next, the world model recurrently predicts future states over multiple time steps.", + "bbox": [ + 511, + 233, + 905, + 324 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input of World Model. Given $N$ refined trajectories $\\hat{\\tau}$ and the current BEV state $\\mathbf{B}_t$ at time $t$ , we construct $N$ state-action pairs as the inputs of world model. We use the trajectory encoder TE in Eq. 1, with shared parameters, to encode $\\hat{\\tau}$ into action embeddings $\\mathbf{A}_t$ . Given $\\mathbf{A}_t = \\{\\mathbf{a}_t^1, \\mathbf{a}_t^2, \\dots, \\mathbf{a}_t^N\\}$ , each state-action pair is represented as $(\\mathbf{B}_t, \\mathbf{a}_t^i)$ , where $i \\in \\{1, \\dots, N\\}$ .", + "bbox": [ + 511, + 340, + 906, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recurrent Future Prediction. Given the state-action pairs, we use a world model to predict their corresponding future states. In contrast to prior work [40], which predict the future in image space, we build a world model in BEV space as it is much more efficient. In detail, given a state-action pair $(\\mathbf{B}_t,\\mathbf{a}_t^i)$ , we flatten $\\mathbf{B}_t$ into $hw$ feature vectors. The dimension of each vector is $c$ . Then we concatenate these vectors with the action embedding $\\mathbf{a}_t^i$ , resulting in a total of $hw + 1$ feature vectors, namely $\\mathbf{F}_i\\in \\mathbb{R}^{(hw + 1)\\times c}$ . $\\mathbf{F}_i$ are then fed into the world model, whose network architecture is a transformer encoder, as $\\mathbf{F}_{i + 1} = \\mathrm{TransformerEncoder}(\\mathbf{F}_i)$ . The output of the world model consists of $h\\times w + 1$ feature vectors, corresponding to the predicted future states $\\mathbf{B}_{t + 1}^i$ and future actions $\\mathbf{a}_{t + 1}^i$ . In summary, we can formulate this process in the following equation:", + "bbox": [ + 511, + 465, + 908, + 693 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {B} _ {t + 1} ^ {i}, \\mathbf {a} _ {t + 1} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t}, \\mathbf {a} _ {t} ^ {i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 700, + 906, + 720 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Furthermore, this world model is able to predict the states $\\mathbf{B}_{t + K}^{i}$ of the future $K$ steps in a recurrent manner, as shown by the following equation:", + "bbox": [ + 511, + 728, + 905, + 773 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathbf {B} _ {t + K} ^ {i}, \\mathbf {a} _ {t + K} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t + K - 1} ^ {i}, \\mathbf {a} _ {t + K - 1} ^ {i}\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 782, + 906, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The process above is conducted for every state-action pair in a parallel manner. It is important to note that the size of BEV states is relatively small (e.g., $h = w = 8$ ), so the world model does not introduce significant computational overhead. Next, we introduce the Reward Model, which predicts rewards for a trajectory based on its future states.", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27139", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b7d10dd316364701d4253cc9d914004d67f33f6eb35bbeaa4c4ce313239d36eb.jpg", + "image_caption": [ + "Figure 2. Overall structure of WoTE. Our model predicts and evaluates trajectory in an end-to-end manner. It is mainly divided into two parts. (a) Trajectory Prediction: the BEV encoder encodes the multi-modal observations into a BEV feature map (state) and proposes multiple trajectory candidates. (b) Trajectory Evaluation: Given the current BEV state and trajectory candidates, we leverage the BEV World Model to predict the corresponding future BEV states of these trajectories. The Reward Model then predicts rewards with the help of future BEV states. The trajectory with the highest reward is selected as the final trajectory." + ], + "image_footnote": [], + "bbox": [ + 96, + 93, + 898, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Reward Model", + "text_level": 1, + "bbox": [ + 89, + 521, + 243, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first present the reward types and then describe how they are learned.", + "bbox": [ + 89, + 542, + 482, + 574 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reward Types. Following [27], we have two types of rewards: the imitation reward and the simulation reward. The imitation reward $r_{\\mathrm{im}}$ measures how well the predicted trajectory mimics an expert trajectory. The detailed implementation will be introduced in Sec. 3.4. The simulation rewards $r_{\\mathrm{sim}}$ measures trajectory quality based on simulator-defined criteria. Following [11], we assess trajectories using the following five criteria: no collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf), and ego progress (EP). As a result, we have $r_{\\mathrm{sim}} = \\{r_{\\mathrm{sim}}^{\\mathrm{NC}}, r_{\\mathrm{sim}}^{\\mathrm{DAC}}, r_{\\mathrm{sim}}^{\\mathrm{TTC}}, r_{\\mathrm{sim}}^{\\mathrm{Comf}}, r_{\\mathrm{sim}}^{\\mathrm{EP}}\\}$ . Given $r_{\\mathrm{im}}$ and $r_{\\mathrm{sim}}$ , the final reward $r_{\\mathrm{final}}$ is defined as", + "bbox": [ + 89, + 590, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} r _ {\\text {f i n a l}} = - \\left(w _ {1} \\log r _ {\\mathrm {i m}} + w _ {2} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {N C}} + w _ {3} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {D A C}} \\right. \\tag {4} \\\\ \\left. + w _ {4} \\log \\left(5 r _ {\\mathrm {s i m}} ^ {\\mathrm {T T C}} + 2 r _ {\\mathrm {s i m}} ^ {\\mathrm {C o m f}} + 5 r _ {\\mathrm {s i m}} ^ {\\mathrm {E P}}\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 777, + 482, + 832 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $w_{i}$ are hyper-parameter.", + "bbox": [ + 89, + 838, + 294, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Reward Prediction. Our method needs these rewards for trajectory evaluation in the inference time, so we propose", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reward prediction. Take the $i$ -th refined trajectory $\\hat{\\tau}_i$ for illustration, the previous approaches [27] typically predict a reward based solely on the current state. With the help of the world model, our reward model predicts the reward $\\mathbf{r}_i$ of $i$ -th trajectory based on both the current BEV state and future $K$ steps BEV states. This process is illustrated in the following equation:", + "bbox": [ + 511, + 522, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {r} _ {i} = \\operatorname {R e w a r d M o d e l} \\left(\\mathbf {B} _ {\\text {a l l}} ^ {i}, \\mathbf {a} _ {\\text {a l l}} ^ {i}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 641, + 906, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{B}_{\\mathrm{all}}^i = \\mathrm{cat}([\\mathbf{B}_t,\\mathbf{B}_{t + 1}^i,\\dots,\\mathbf{B}_{t + K}^i])$ and $\\mathbf{a}_{\\mathrm{all}}^i = \\mathrm{MLP}(\\mathrm{cat}([\\mathbf{a}_t^i,\\mathbf{a}_{t + 1}^i,\\dots,\\mathbf{a}_{t + K}^i]))$ . $\\mathrm{cat}([\\cdot])$ represents the concatenate operation along the channel dimension. In detail, Reward Model consists of three parts. It first applies 2D convolutions to $\\mathbf{B}_{\\mathrm{all}}^i$ to aggregate information from different regions and time steps of the BEV states. Next, a global average pooling is applied to form $\\mathbf{B}_{\\mathrm{pool}}\\in \\mathbb{R}^{c}$ , $c$ is the channel dimension. We then fed $\\mathbf{S}^i = \\mathrm{cat}([\\mathbf{B}_{\\mathrm{pool}},\\mathbf{a}_{\\mathrm{all}}^i])$ , where $\\mathbf{S}^i\\in \\mathbb{R}^{2c}$ , into an MLP head to predict the reward $\\mathbf{r}_i\\in \\mathbb{R}^m$ , where $m$ denotes the number of reward. Next, we introduction the definition of these rewards and select a final trajectory based on these rewards.", + "bbox": [ + 511, + 671, + 908, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The refined trajectory with the highest final reward is selected as the final trajectory. Next, we introduce the supervision of these predicted rewards and BEV states.", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27140", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Supervision in BEV Space", + "text_level": 1, + "bbox": [ + 89, + 90, + 328, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The world model predicts multiple future feature states corresponding to different trajectories. However, supervising these predictions in the image space is challenging, as driving logs provide only a single future trajectory. To address this, we propose supervising in the BEV space. Supervising in BEV space alleviates the difficulty of accurately modeling future states in image space while also improving computational efficiency. Many real-world traffic simulators support traffic simulations in the BEV space. They generate realistic and reliable future BEV scenarios along with rewards. We adopt nuPlan [3] as our traffic simulator for its broad adoption and realistic scenario simulations.", + "bbox": [ + 89, + 112, + 485, + 294 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervision of BEV States. For BEV state supervision, we decode a semantic BEV map based on the BEV state to enable explicit supervision as Figure 2 (b) shows. We utilize upsampling and transposed convolution layers to decode a BEV semantic map from the BEV state. The dimension of the BEV semantic map is $H \\times W \\times L$ , where $L$ denotes the number of semantic classes. This BEV semantic map provides a comprehensive representation of a BEV scenario, covering classes including background, road, walkways, centerline, static objects, vehicles and pedestrians. Each element is represented as a one-hot encoded channel in $L$ . For each time step $t + k$ , We use Focal Loss [28] to supervise the predicted BEV semantic map $\\mathbb{B}_{t + k}$ as", + "bbox": [ + 89, + 311, + 483, + 508 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {B E V}} = \\operatorname {F o c a l L o s s} \\left(\\mathbb {B} _ {t + k}, \\mathbb {B} _ {t + k} ^ {*}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 521, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbb{B}_{t + k}^{*}$ is the ground truth simulated BEV semantic map at time $t + k$ provided by the simulator.", + "bbox": [ + 89, + 549, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervision of Simulation Rewards. Regarding the simulation rewards, we leverage the simulator to produce five rewards for evaluating a trajectory. Specifically, given the $i$ -th trajectory $\\hat{\\tau}^i$ , the simulator simulates the future location of the other agents and ego vehicle and produces a simulated future BEV semantic map with agents. Next, it uses a rule-based evaluator to evaluate this simulated future BEV scenarios for producing the corresponding target simulation rewards of $\\hat{\\tau}^i$ . Given the rewards provided by the simulator, we use the Binary Cross Entropy (BCE) loss to supervise the predicted simulation reward $r_{\\mathrm{sim}}$ . As a result, the loss is formulated as", + "bbox": [ + 89, + 597, + 485, + 777 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} = \\mathrm {B C E} \\left(r _ {\\text {s i m}}, r _ {\\text {s i m}} ^ {*}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 199, + 790, + 483, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervision of Imitation Reward. In addition to the supervision from the simulator, our framework is also guided by the expert driver.", + "bbox": [ + 89, + 825, + 483, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We compute the target of imitation reward by first calculating the L2 distances $d_{i}$ between the $i$ -th trajectory anchor", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and an expert trajectory. Then, we apply the negative of these distances from $N$ trajectory anchors to the softmax function, yielding: $r_{\\mathrm{im},i}^{*} = \\frac{\\exp(-d_{i})}{\\sum_{j=1}^{N}\\exp(-d_{j})}$ . Intuitively, a trajectory anchor closer to the expert trajectory than others will receive a higher reward. As a result, the imitation reward $r_{\\mathrm{im}}^{*} \\in \\mathbb{R}^{N}$ quantifies the similarity of each trajectory anchor to the human trajectory. The predicted imitation reward $r_{\\mathrm{im}}$ , which is produced by Eq. 5, are supervised using the Cross Entropy loss as", + "bbox": [ + 511, + 90, + 908, + 233 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\mathrm {i m}} = \\operatorname {C r o s s E n t r o p y} \\left(r _ {\\mathrm {i m}}, r _ {\\mathrm {i m}} ^ {*}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 243, + 906, + 262 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Supervision of Trajectory. For supervising the refined trajectories, we adopt the commonly used winner-take-all strategy [46]. In this strategy, we select the trajectory anchor that is closest to the expert trajectory. Only the corresponding refined trajectory of this anchor will be supervised. Let this trajectory anchor be denoted as $\\tau_{i}$ . We use the L1 loss to measure the difference between the expert trajectory $\\tau^{*}$ and the refined trajectory $\\hat{\\tau}_{i}$ as", + "bbox": [ + 511, + 277, + 908, + 400 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t r a j}} = \\left| \\hat {\\tau} _ {i} - \\tau^ {*} \\right|. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 411, + 906, + 426 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Instead of supervising all refined trajectories, we focus solely on the most optimal one. This strategy helps the model capture a diverse range of trajectory patterns, with each anchor specializing in a particular trajectory modality.", + "bbox": [ + 511, + 438, + 906, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall training loss $\\mathcal{L}_{\\mathrm{total}}$ is as follow:", + "bbox": [ + 532, + 500, + 815, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {B E V}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {i m}} + \\mathcal {L} _ {\\text {t r a j}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 525, + 906, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 566, + 643, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Benchmark", + "text_level": 1, + "bbox": [ + 511, + 590, + 640, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NAVSIM Dataset The NAVSIM dataset [11] is constructed based on nuPlan [3]. Specifically, OpenScene [8] first downsampled the nuPlan data from $10\\mathrm{Hz}$ to $2\\mathrm{Hz}$ to condense it into 120 hours of driving logs. NAVSIM resampled the data from OpenScene to emphasize challenging scenarios, reducing simple situations like straight-line driving. The dataset is divided into two parts: Navtrain and Navtest, comprising 1192 scenarios for training and validation and 136 scenarios for testing. Unlike nuScenes [2], which collects data in simpler, slower driving scenarios, NAVSIM ensures that challenging scenarios are prioritized, making simply fitting ego status insufficient for planning.", + "bbox": [ + 511, + 613, + 908, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "NAVSIM Metrics The original end-to-end driving metrics were primarily designed to evaluate the deviation between the predicted trajectories and those driven by human experts. However, this evaluation approach has proven inadequate, resulting in poor performance in real-world driving contexts. NAVSIM introduces more practical and reliable metrics, focusing on aligning open-loop and closed-loop metrics.", + "bbox": [ + 511, + 795, + 910, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27141", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/a0bf8737b09b4588402d615262edc5bc6a9945e761d56b9d781b99c028f325b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodInput#Traj.Traj. Eval.NC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
Human---10010087.510099.994.8
Constant Velocity-1×69.958.849.349.3100.021.6
Ego Status MLP-1×93.077.362.883.6100.065.6
VADv2 [41]C8192Rule-based97.991.777.692.9100.083.0
UniAD [22]C1Rule-based97.891.978.892.9100.083.4
LTF [31]C1×97.492.879.092.4100.083.8
PARA-Drive [41]C1Rule-based97.992.479.393.099.884.0
TransFuser [31]C & L1×97.792.879.292.8100.084.0
LAW [26]C1×96.495.481.788.799.984.6
DRAMA [43]C & L1×98.093.180.194.8100.085.5
Hydra-MDP [27]C & L8192Model-free98.396.078.794.6100.086.5
WoTEC & L256Model-based98.596.881.994.999.988.3
", + "bbox": [ + 114, + 88, + 883, + 316 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e7455fc63dbfcff680297a389d3c8eaf49119b9721173f37838a247f62fbf342.jpg", + "table_caption": [ + "Table 1. Comparison with the SOTA approaches on NAVSIM test set. Rule-based: UniAD and PARA-Drive use occupancy maps, while VADv2 relies on vectorized maps for trajectory evaluation following specific rules. Model-free: Hydra-MDP evaluates trajectory without world modeling. Model-based: our WoTE evaluates trajectory with the assistance of the BEV world model. NC: no at-fault collision. DAC: drivable area compliance. EP: ego progress. TTC: time-to-collision. Conf.: comfort. PDMS: the predictive driver model score. Traj.: Trajectory. Eval.: Evaluation. C: Camera. L: LiDAR." + ], + "table_footnote": [], + "table_body": "
MethodTraj. Eval.Success Rate↑Driving Score↑
AD-MLP [44]-0.0018.05
UniAD [22]Rule-based16.3645.81
VAD [25]Rule-based15.0042.35
TCP [42]-30.0059.90
WoTEModel-based31.3661.71
", + "bbox": [ + 91, + 431, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with SOTA approaches on closed-loop Bench2Drive [24] benchmark on CARLA simulator. Traj.: Trajectory. Eval.: Evaluation.", + "bbox": [ + 89, + 535, + 483, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, NAVSIM evaluates model performance using the Predictive Driver Model Score (PDMS), which is calculated based on five factors: No At-Fault Collision (NC), Drivable Area Compliance (DAC), Time-to-Collision (TTC), Comfort (Comf.), and Ego Progress (EP). The PDMS is calculated as $\\mathrm{PDMS} = \\mathrm{NC} \\times \\mathrm{DAC} \\times \\frac{5 \\times \\mathrm{EP} + 5 \\times \\mathrm{TTC} + 2 \\times \\mathrm{C}}{12}$ .", + "bbox": [ + 89, + 595, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bench2Drive Dataset and Metrics To further evaluate our framework, we conduct closed-loop experiments in the CARLA simulator [13]. Specifically, we adopt the recently proposed Bench2Drive [24] benchmark as our closed-loop evaluation benchmark. Bench2Drive provides a standardized training dataset on CARLA, ensuring fair comparisons across different methods. The benchmark consists of 220 short evaluation routes (around $150\\mathrm{m}$ each) distributed across all CARLA towns, with each route containing one safety-critical scenario. This design effectively reduces the evaluation variance. For metric, it utilizes the standard CARLA metric, Driving Score (DS), as the primary metric. Additionally, Bench2Drive reports the success rate, which represents the proportion of successfully completed routes.", + "bbox": [ + 89, + 688, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "A route is considered successful if and only if DS reaches $100\\%$ on this route.", + "bbox": [ + 511, + 434, + 906, + 464 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 481, + 730, + 496 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "NAVSIM For input data, it is aligned with TransFuser [31]. We concatenate the front-view image with center-cropped front-left and front-right images, resulting in a combined resolution of $256 \\times 1024$ pixels. For LiDAR, the $64\\mathrm{m} \\times 64\\mathrm{m}$ point cloud surrounding the ego vehicles is used. For network architecture, we employ ResNet34 as the backbone for BEV feature extraction following TransFuser [31]. The world model consists of two transformer decoder layers. We use $N = 256$ trajectory anchors. The predicted future states are of time $t + 2s, t + 4s$ when the current is of time $t$ . The reward weights are set as follows: $w_{1} = 0.1, w_{2} = w_{3} = 0.5, w_{4} = 1.0$ . The training is conducted on the Navtrain split using 8 NVIDIA L20 GPUs with a total batch size of 256, distributed across 50 epochs. In the ablation study section, we train our model for 20 epochs instead of 50 to accelerate the training process. For fast training, we pre-compute the simulation results (BEV semantic maps and scores) based on the trajectory anchors. We then feed trajectory anchors into the trajectory evaluation module during training. During testing, the trajectory evaluation module evaluates the refined trajectory instead of the anchor trajectories. We utilize the Adam optimizer with a learning rate of 1e-4.", + "bbox": [ + 511, + 505, + 908, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Bench2Drive For Bench2Drive, we adopt TCP [42] as our baseline and integrate the trajectory evaluation module. We choose TCP as it is the best open-source framework officially", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27142", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/80d729add5851c43b89f1dc69c81ba530ce2e696512e1e4dad2f187f63ba9b0a.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Traj. Eval.Future StatesNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
××96.491.576.290.399.981.0
✓×97.192.978.291.9100.083.2
✓✓98.094.779.993.4100.085.6
", + "bbox": [ + 94, + 89, + 480, + 141 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "provided by Bench2Drive. Similar to NAVSIM, we utilize imitation rewards along with simulation-based rewards, including No Collision (NC) and Drivable Area Compliance (DAC). The number of trajectory anchors is set to 256. Notably, rather than employing the winner-take-all strategy (Eq. 9) to supervise only the best trajectory, we find that supervising all trajectories leads to improved performance. The model is trained for 27 epochs with a batch size of 300. We use the Adam optimizer with a learning rate of 1e-4.", + "bbox": [ + 88, + 218, + 486, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA", + "text_level": 1, + "bbox": [ + 89, + 363, + 312, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NASIM Table 1 compares our method with state-of-the-art approaches on the NAVSIM test set. By leveraging the world model to simulate future scenarios and guide trajectory evaluation, our method achieves a PDMS of 87.1. WoTE outperforms the previous model-free approach, Hydra-MDP, highlighting the advantages of model-based trajectory evaluation.", + "bbox": [ + 89, + 385, + 483, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bench2Drive To further validate our approach, we conduct a closed-loop evaluation on the Bench2Drive benchmark within the CARLA simulator. As shown in Table 2, our method improves the Driving Score by 1.81 points, demonstrating its effectiveness in various evaluation settings.", + "bbox": [ + 89, + 491, + 483, + 568 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 575, + 256, + 590 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trajectory evaluation and future state prediction matter. In this experiment, we investigate the importance of the trajectory evaluation and future state prediction as Table 3 shows. The first row in Table 3 represents our baseline TransFuser [31], which performs trajectory prediction only. In the second row, we change to our framework, which consists of both trajectory prediction and trajectory evaluation. However, we do not used the predicted future states as the input of Reward Model. That is, $\\mathbf{B}_{\\mathrm{all}} = \\mathbf{B}_t$ and $\\mathbf{a}_{\\mathrm{all}} = \\mathbf{a}_t$ in Eq. 5. In the third row, we integrate the BEV world model to predict future states, leading to notable performance improvements across multiple metrics.", + "bbox": [ + 89, + 598, + 485, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Imitation and simulation rewards are complementary. To investigate how each type of reward affects driving performance, we conduct an ablation study, as shown in Table 4. During inference, we adjust the reward weights in Eq. 4 to isolate their effects: setting $w_{1} = 0$ enables evaluation with only simulation rewards, while setting $w_{2} = w_{3} = w_{4} = 0$ tests performance with only an imitation reward. The results indicate that imitation and simulation rewards emphasize", + "bbox": [ + 89, + 780, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8e4344f6de648e46009081368beee9ba37677b235e5ce049b53109b9e1c000c2.jpg", + "table_caption": [ + "Table 3. Ablation study on trajectory evaluation and future state prediction. Traj: Trajectory. Eval.: Evaluation." + ], + "table_footnote": [], + "table_body": "
Imi. RewardSim. RewardsNC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
✓×97.992.577.793.499.983.5
×✓96.594.779.290.199.483.6
✓✓98.094.779.993.4100.085.6
", + "bbox": [ + 517, + 90, + 903, + 138 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9b2cae285c4bbdadaef0e9f89e4456b5e7f3e5c98f34af501717baefa47340b3.jpg", + "table_caption": [ + "Table 4. Ablation study on the imitation and simulation rewards. The experiment shows that imitation and simulation rewards are complementary. Imi: Imitation. Sim.: Simulation." + ], + "table_footnote": [], + "table_body": "
Future Prediction StepsNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
0s → 4s97.493.679.191.9100.084.0
0s → 2s → 4s98.094.779.993.4100.085.6
", + "bbox": [ + 516, + 220, + 903, + 263 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/297e3ee86002da8512d05626266b644a4a587e8342a4cdb4d66b664414ecbcf5.jpg", + "table_caption": [ + "Table 5. Ablation study on the different number of time steps in the future prediction. Our BEV world model predicts future states in a recurrent manner. $\\mathrm{A} \\rightarrow \\mathrm{B}$ : the predicted future state at time B is based on the state at time A." + ], + "table_footnote": [], + "table_body": "
#Trajs.NC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑Latency ↓
6497.494.179.691.9100.084.517.2 ms
12897.994.680.093.0100.085.417.9 ms
25698.094.779.993.4100.085.618.7 ms
", + "bbox": [ + 517, + 359, + 903, + 415 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Ablation study on the different number of Trajectories. The number of trajectories is the same during training and testing. The latency is evaluated on an NVIDIA L20 GPU. Trajs.: Trajectories.", + "bbox": [ + 511, + 424, + 908, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "different aspects of planning. Trajectory evaluation using imitation rewards performs better in NC and TTC, while using simulation rewards excels in DAC and EP. By integrating both, the model leverages their respective strengths, achieving improved overall performance across all key metrics.", + "bbox": [ + 511, + 521, + 908, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Recurrent future state prediction helps. Our framework supports predicting future states recurrently, as detailed in Eq. 3. In Table 5, we present the results obtained by varying the number of BEV states predicted by the world model. Our findings indicate that using a finer time step for future state prediction significantly improves performance. This is because a finer time step provides richer temporal information, which aids trajectory evaluation.", + "bbox": [ + 511, + 612, + 908, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More trajectories help. This ablation study explores the effect of varying the number of trajectories on overall performance. Specifically, the model in each row is trained and tested under the same number of trajectories, and the world model is fed with refined trajectories. As shown in Table 6, using a small number of trajectories, such as 64, significantly degrades performance. As the number of trajectories increases, performance improves notably, with a marked improvement observed when increasing from 64 to 128 trajectories. However, as the number of trajectories increases from 128 to 256, the performance gains diminish,", + "bbox": [ + 511, + 734, + 910, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27143", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6f338334e2475f36eeba07ee9ed11cdeeb840316694a4f123dd8776c1ceb5cb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Train Traj.Test Traj.Traj. Refine†NC ↑DAC ↑EP ↑
256256×96.694.679.0
2561024×97.295.679.3
256256✓98.094.779.9
", + "bbox": [ + 93, + 88, + 357, + 132 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Generalization of the trajectory across different trajectory configurations.", + "bbox": [ + 89, + 143, + 357, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the input trajectories of BEV world model a trajectory.", + "bbox": [ + 89, + 171, + 357, + 199 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "indicating that the performance of the most optimal level. Based on these results, we default number of trajectories for our ex- Latency analysis. A common concern is of our framework. To address this, we run on an NVIDIA L20 GPU, as shown in GPU processes state-action pairs in para remains efficient even as the number of t The total latency is only $18.7\\mathrm{ms}$ well $\\mathsf{V}$ requirements for end-to-end autonomous", + "bbox": [ + 89, + 239, + 357, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization ability across different", + "text_level": 1, + "bbox": [ + 89, + 402, + 357, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "shown in Table 7, we investigate the generalization ability of our trajectory evaluation module. To be specific, we sample $N = 1024$ instead of $N = 256$ trajectory anchors from the dataset. Since our model is trained under the setting of $N = 256$ trajectory anchors, the model has never seen these $N = 1024$ trajectories. Surprisingly, we find a 1.3 PDMS increase when using these unseen trajectory anchors. Besides, the model performs even better when using refined trajectories as inputs. These refined trajectories continuously change based on varying sensor inputs. This shows that the proposed world model has great generalization ability when dealing with unseen trajectories.", + "bbox": [ + 88, + 419, + 483, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visual Analysis", + "text_level": 1, + "bbox": [ + 89, + 612, + 245, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "End-to-end planning trajectories. To highlight the improvements in end-to-end driving achieved through trajectory evaluation, we present visualizations of end-to-end planning trajectories in Figure 3.", + "bbox": [ + 89, + 635, + 483, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Trajectory rewards. Figure 4 illustrates the predicted rewards for different trajectories.", + "bbox": [ + 89, + 705, + 483, + 736 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 753, + 209, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel framework that leverages a BEV world model for end-to-end trajectory evaluation in autonomous driving. By integrating a BEV world model, we enable a more informed evaluation process that considers the dynamic evolution of driving scenarios, leading to more effective trajectory evaluation. Our approach benefits from dense supervision provided by BEV-space traffic simulators, which supply both semantic future states and rule-based", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a81390ca3df63e0c7551fcba665b66c1848c10cd21449e8f9f189b80987ad348.jpg", + "image_caption": [ + "Figure 3. Visualization of end-to-end planning trajectories. We compare our method with TransFuser [31], which serves as our trajectory prediction baseline without a trajectory evaluation module. Our trajectory evaluation module effectively filters out low-quality trajectories and helps avoid collisions." + ], + "image_footnote": [], + "bbox": [ + 524, + 90, + 898, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3edd753908390f9104df1a78bed1563b0620bf91fbf95d995bd72770970f2b46.jpg", + "image_caption": [ + "(a) Multi-modality trajectory preservation" + ], + "image_footnote": [], + "bbox": [ + 519, + 380, + 893, + 500 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e9903960bdf2540bf5e55a7e97e3140fc1b6eb009ec09ab904f0ec0d00d5937d.jpg", + "image_caption": [ + "(b) Low-quality trajectory suppression", + "Figure 4. Visualization of rewards of all trajectories. Brighter colors indicate higher trajectory rewards, while gray trajectories correspond to those with very low rewards. (a): WoTE effectively retains diverse multi-modal trajectories, as long as they comply with traffic rules and do not result in unsafe outcomes. (b): it demonstrates that trajectories considered unreasonable receive significantly lower rewards, highlighting the effectiveness of our trajectory evaluation module." + ], + "image_footnote": [], + "bbox": [ + 519, + 513, + 893, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "reward targets. Our method achieves state-of-the-art performance on the NAVSIM and Bench2Drive benchmarks while maintaining real-time efficiency. Overall, our work establishes trajectory evaluation as an important research direction for end-to-end autonomous driving. We hope our framework serves as a strong baseline and inspires further research in end-to-end online trajectory evaluation.", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27144", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Science and Technology Major Project (No. 2022ZD0116500), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010).", + "bbox": [ + 89, + 114, + 485, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 191, + 187, + 207 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sébastien MR Arnold, Pierre L'Ecuyer, Liyu Chen, Yi-fan Chen, and Fei Sha. Policy learning and evaluation with randomized quasi-monte carlo. arXiv preprint arXiv:2202.07808, 2022. 2", + "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. CVPR, 2020. 5", + "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 5", + "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 2", + "[5] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In CVPR, pages 15590-15599, 2021. 1, 2", + "[6] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1", + "[7] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3", + "[8] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving, 2023. 5", + "[9] Christoph Dann, Gerhard Neumann, and Jan Peters. Policy evaluation with temporal differences: A survey and comparison. The Journal of Machine Learning Research, 15(1): 809-883, 2014. 2", + "[10] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1", + "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. arXiv preprint arXiv:2406.15349, 2024. 2, 4, 5", + "[12] Christopher Diehl, Timo Sievernich, Martin Krüger, Frank Hoffmann, and Torsten Bertram. Umbrella: Uncertainty-aware model-based offline reinforcement learning leveraging planning. arXiv preprint arXiv:2111.11097, 2021. 2" + ], + "bbox": [ + 93, + 215, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 2, 6", + "[14] Haoyang Fan, Fan Zhu, Changchun Liu, Liangliang Zhang, Li Zhuang, Dong Li, Weicheng Zhu, Jiangtao Hu, Hongye Li, and Qi Kong. Baidu Apollo em motion planner. arXiv preprint arXiv:1807.08048, 2018. 1", + "[15] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 1", + "[16] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[17] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018. 1", + "[18] Mikael Henaff, Alfredo Canziani, and Yann LeCun. Model-predictive policy learning with uncertainty regularization for driving in dense traffic. In International Conference on Learning Representations, 2019. 1, 2", + "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 1, 2", + "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1", + "[21] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 1, 2", + "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 2, 6", + "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 1", + "[24] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 2, 6", + "[25] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6", + "[26] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27145", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 6", + "[27] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 3, 4, 6", + "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dolkar. Focal loss for dense object detection. In ICCV, 2017. 5", + "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 1", + "[30] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, pages 7077-7087, 2021. 1", + "[31] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 3, 6, 7, 8", + "[32] Guanren Qiao, Guorui Quan, Rongxiao Qu, and Guiliang Liu. Modelling competitive behaviors in autonomous driving under generative world model. In European Conference on Computer Vision, pages 19-36. Springer, 2024. 1, 2", + "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[34] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In ECCV, 2020. 2", + "[35] Ardi Tampuu, Tambet Matiisen, Maksym Semikin, Dmytro Fishman, and Naveed Muhammad. A survey of end-to-end driving: Architectures and training methods. IEEE Transactions on Neural Networks and Learning Systems, 33(4): 1364-1384, 2020. 1", + "[36] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2", + "[37] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1", + "[38] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2", + "[39] Yuqi Wang, Ke Cheng, Jiawei He, Qitai Wang, Hengchen Dai, Yuntao Chen, Fei Xia, and Zhaoxiang Zhang. Drivingndojo dataset: Advancing interactive and knowledge-enriched driving world model. arXiv preprint arXiv:2410.10738, 2024.", + "[40] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Con-" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ference on Computer Vision and Pattern Recognition, pages 14749-14759, 2024. 1, 2, 3", + "[41] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 6", + "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: a simple yet strong baseline. NeurIPS, 2022. 6", + "[43] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 2, 6", + "[44] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 6", + "[45] Yumeng Zhang, Shi Gong, Kaixin Xiong, Xiaoqing Ye, Xiao Tan, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Bevworld: A multimodal world model for autonomous driving via unified bev latent space. arXiv preprint arXiv:2407.05679, 2024. 2", + "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 5" + ], + "bbox": [ + 516, + 92, + 906, + 515 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27146", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_model.json b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c3c36a17b852d4f6cc9ac03b4ae02d79fbdde8f0 --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.152 + ], + "angle": 0, + "content": "End-to-End Driving with Online Trajectory Evaluation via BEV World Model" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.181, + 0.846, + 0.253 + ], + "angle": 0, + "content": "Yingyan Li* Yuqi Wang* Yang Liu Jiawei He Lue Fan* Zhaoxiang Zhang* NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA) State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.485, + 0.606 + ], + "angle": 0, + "content": "End-to-end autonomous driving has achieved remarkable progress by integrating perception, prediction, and planning into a fully differentiable framework. Yet, to fully realize its potential, an effective online trajectory evaluation is in dispensable to ensure safety. By forecasting the future outcomes of a given trajectory, trajectory evaluation becomes much more effective. This goal can be achieved by employing a world model to capture environmental dynamic and predict future states. Therefore, we propose an end-to-end driving framework WoTE, which leverages a BE World model to predict future BEV states for Trajectory Evaluation. The proposed BEV world model is latency efficient compared to image-level world models and can be seamlessly supervised using off-the-shelf BEV-space traffic simulators. We validate our framework on both the NAVSIM benchmark and the closed-loop Bench2Drive benchmark based on the CARLA simulator, achieving state-of-the-art performance. Code is released at https://github.com/liyingyanUCAS/WoTE." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.637, + 0.222, + 0.652 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.483, + 0.814 + ], + "angle": 0, + "content": "End-to-end autonomous driving [6, 30, 35] has achieved remarkable success in recent years and garnered widespread attention. While most approaches [21, 23, 25] primarily focus on predicting a high-quality trajectory, recent works [7, 27] suggest that predicting multiple trajectories simultaneously better reflects the multi-modal nature of driving and leads to improved performance. Given these multiple possible trajectories, it becomes crucial to effectively evaluate them in order to ensure the safety and reliability of end-to-end autonomous driving systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Traditional trajectory evaluation methods are rule-based [10, 14, 20, 37] and rely on perception results, such as bounding boxes and maps, to assess trajectories. These methods are sensitive to perception inaccuracies and may" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.288, + 0.902, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.354, + 0.902, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.6 + ], + "angle": 0, + "content": "Figure 1. The illustration of end-to-end trajectory evaluation with BEV world model. (a) Previous end-to-end driving methods primarily focus on learning a high-quality trajectory. (b) Our WoTE consists of end-to-end trajectory prediction and evaluation. First, we predict multi-modal trajectories. Next, given these trajectories, we leverage the BEV world model to predict the corresponding future BEV states. The reward model evaluates these trajectories based on the predicted states and assigns rewards, selecting the highest-reward trajectory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.764 + ], + "angle": 0, + "content": "not be directly optimized in an end-to-end manner. Recently, some end-to-end driving approaches [7, 27] have incorporated trajectory evaluation. Their evaluations rely on the current state. However, evaluating a trajectory candidate is particularly challenging without knowledge of the future led by this trajectory. Similar to human drivers, who anticipate futures before making decisions, a trajectory evaluation network would benefit from predicting the corresponding future states of trajectory candidates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To effectively predict future states, it is intuitive to adopt methodologies from reinforcement learning, where world models have successfully facilitated future state predictions. In reinforcement learning, model-based methods [5, 18, 32] have consistently shown superior performance compared to model-free approaches [17, 29], largely due to their ability of modeling the environmental dynamics and reasoning about future scenarios through world models. Inspired by this success, we propose adopting a world model [15, 19, 38-40]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.465, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\ast\\): Co-first author, \\(^{2}\\): Co-advisor. Email: liyingyan2021@ia.ac.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27137" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "within end-to-end autonomous driving frameworks to facilitate accurate prediction of future states, thus significantly improving trajectory evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.142, + 0.486, + 0.309 + ], + "angle": 0, + "content": "However, leveraging the world model for trajectory evaluation presents some technical issues. First, a good representation of future scenarios is needed. Many existing world models for driving [19, 38, 40] predict the images of future scenarios with diffusion models [33], which can be time-consuming and not suitable for real-time driving. Second, we lack supervision for future states. Supervising the predicted future states of the world model is difficult because world models need to imagine multiple future states based on the multiple trajectory candidates, while only one future state is available in the real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.314, + 0.486, + 0.601 + ], + "angle": 0, + "content": "To address these technical issues, we propose WoTE as shown in Fig.1. We propose to predict future states in the BEV space. The information in BEV space is much more compact than in raw images, which enables a more efficient single-step feed-forward future prediction than the time-consuming multi-step denoising in previous driving world models [19, 38, 40, 45]. This characteristic of BEV space makes WoTE highly suitable for real-time driving applications. Additionally, the predicted future BEV states can be easily supervised using off-the-shelf BEV-space traffic simulators such as nuPlan [3, 16]. These simulators model the scenario in BEV space by representing the map and objects using BEV semantic maps, creating a dynamic simulation of the driving environment. Consequently, we can use the BEV semantic maps from these simulated scenarios as supervision for the prediction of the BEV world model. Moreover, these simulators are equipped with well-established evaluation rules to assess the simulated future BEV scenarios, which provide target rewards." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.606, + 0.486, + 0.728 + ], + "angle": 0, + "content": "We validate our framework on both NAVSIM [11] benchmark and closed-loop Bench2Drive [24] benchmark based on the CARLA [13] simulator and achieve state-of-the-art performance. Our experiments show that incorporating future states into trajectory evaluation is essential for improved performance. Additionally, the trajectory evaluation module, built on the world model shows great generalization ability across diverse trajectories. Our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.732, + 0.483, + 0.775 + ], + "angle": 0, + "content": "- We highlight the importance of utilizing imagined future states for trajectory evaluation in end-to-end autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.778, + 0.483, + 0.851 + ], + "angle": 0, + "content": "- We introduce WoTE, which consists of an end-to-end trajectory evaluation module based on a BEV world model. The world model in BEV space enables efficient real-time future prediction and overcomes the scarcity of multi-future supervisions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.853, + 0.482, + 0.897 + ], + "angle": 0, + "content": "- WoTE is validated on both the NAVSIM [11] benchmark and the closed-loop CARLA-based Bench2Drive [24] benchmark, achieving state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.732, + 0.483, + 0.897 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.665, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.116, + 0.803, + 0.133 + ], + "angle": 0, + "content": "2.1. End-to-end Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.139, + 0.908, + 0.518 + ], + "angle": 0, + "content": "End-to-end autonomous driving [7, 22, 25, 43] is to train a model that directly maps sensor inputs to trajectories or control signals instead of decomposing the driving task into traditional sub-tasks. These works can be divided into two categories according to the training paradigm: imitation-learning-based and reinforcement-learning-based. Imitation learning is the most common practice in end-to-end autonomous driving. Their predicted trajectories are supervised by the expert trajectories. For instance, P3 [34], takes a differentiable semantic occupancy as the intermediate representation and learns actions from human trajectories. Following this, ST-P3 [21] learns spatial-temporal features for perception, prediction and planning tasks simultaneously. UniAD [22] proposed goal-oriented planning, combining and leveraging advantages of perception and motion prediction modules. VAD [25] introduces vectorized representations for end-to-end planning. There are also some reinforcement-learning-based methods. For example, MaRLn [36] utilizes implicit affordances to develop a reinforcement learning algorithm. LBC [4] trained a vision-based sensorimotor agent from the privileged agent. Combining imitation learning and reinforcement learning, Hydra-MDP [27] utilizes knowledge distillation to obtain supervisions from both rule-based planners and human drivers. It is a promising research direction as more supervisions are utilized." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.53, + 0.722, + 0.547 + ], + "angle": 0, + "content": "2.2. Trajectory Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.554, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Trajectory evaluation is essential for ensuring the reliability and safety of autonomous driving systems. We divide the main trajectory evaluation methods into two categories: model-free and model-based. Model-free methods directly assess the learned policy using observed data without explicitly modeling environmental dynamics. For instance, Monte Carlo policy evaluation [1] uses the empirical mean return as the metric. The value function approximates the expectation based on sampled returns. Temporal-difference learning [9] uses bootstrapping, where the current estimate of the value function is used to generate target values for updating the value function. Subsequently, model-based trajectory evaluation methods [12] have gained increasing popularity in recent years. By leveraging learned world models that capture environmental dynamics, these methods enable more accurate and predictive policy evaluation. For example, PPUU [18] measures policy performance by quantifying uncertainty in predicted future states. Rails [5] evaluates trajectories through a tabular dynamic programming mechanism, and MARL-CCE [32] employs learned environmental dynamics to assess multi-agent policies in driving scenarios. However, these methods often rely on explicit trajectory representations and non-differentiable metrics, which limits" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27138" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "their end-to-end optimization capabilities. In contrast, our approach evaluates trajectories using BEV features and encoded trajectory embeddings, enabling fully differentiable and thus end-to-end trajectory evaluation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.182, + 0.182 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.485, + 0.403 + ], + "angle": 0, + "content": "In this section, we introduce four key components of WoTE: Trajectory Predictor (Sec. 3.1), BEV World Model (Sec. 3.2), Reward Model (Sec. 3.3) and BEV Space Supervision (Sec. 3.4). As illustrated in Figure 2, our method integrates future state prediction via the BEV World Model and reward prediction via the Reward Model within a unified framework. Specifically, the Trajectory Predictor encodes the multi-modal observations into BEV state, represented as a BEV feature map, and provides multiple trajectory proposals. The BEV World Model predicts corresponding future BEV states based on these trajectory proposals. With the help of future states, the Reward Model predicts the reward of each trajectory. Finally, we choose the trajectory with the highest reward as the final trajectory." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.413, + 0.287, + 0.429 + ], + "angle": 0, + "content": "3.1. Trajectory Predictor" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.486, + 0.527 + ], + "angle": 0, + "content": "As shown in Figure 2 (a), the Trajectory Predictor aims to predict multiple trajectory candidates. First, the BEV encoder encodes the multi-modal sensor inputs into a unified BEV feature map. Then, given trajectory anchors, the trajectory refinement module refines trajectories based on the BEV feature map. We will now describe each part in detail." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.545, + 0.485, + 0.665 + ], + "angle": 0, + "content": "Multi-modal BEV Encoder. Following TransFuser [31], our method leverages multi-modal sensor inputs, including LiDAR and multi-view RGB images. We use a BEV encoder [31] to transform the current multi-modal inputs into a unified BEV feature map \\(\\mathbf{B} \\in \\mathbb{R}^{h \\times w \\times c}\\). For convenience in the following text, we refer to the BEV feature map as the BEV state. \\(h\\) and \\(w\\) represent the height and width of the BEV state, while \\(c\\) denotes the number of channels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.485, + 0.76 + ], + "angle": 0, + "content": "Trajectory Anchors. Following methods in [7, 27], we generate trajectory anchors \\(\\tau\\) by K-Means clustering. This clustering is performed on all the expert trajectories collected from the training dataset. The number of trajectory anchors is denoted as \\(N\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Trajectory Refinement. As illustrated in Figure 2 (a), the trajectory refinement module takes the trajectory anchors and the current BEV states as inputs to predict the refined trajectories. Specifically, we first use a trajectory encoder TE, which is implemented by an MLP, to encode the trajectory anchors \\(\\tau\\) into feature vectors. Then, these feature vectors are used as queries to perform cross attention with the current BEV state B, which serves as both the key and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "value. The output of the cross-attention is then fed into an MLP head to predict the offsets. The offset is added to the trajectory anchors to produce the refined trajectories. The entire process can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.163, + 0.907, + 0.179 + ], + "angle": 0, + "content": "\\[\n\\hat {\\tau} = \\tau + \\operatorname {M L P} (\\text {C r o s s A t t e n t i o n} (\\mathbf {T E} (\\tau), \\mathbf {B}, \\mathbf {B})), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.188, + 0.741, + 0.204 + ], + "angle": 0, + "content": "where \\(\\hat{\\tau}\\) is the refined trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.212, + 0.694, + 0.227 + ], + "angle": 0, + "content": "3.2. BEV World Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.234, + 0.906, + 0.325 + ], + "angle": 0, + "content": "In this section, we use the BEV World Model to predict the future BEV states, as illustrated in Figure 2 (b). The process is divided into two steps. First, the current BEV state and refined trajectories are combined to form the inputs for the world model. Next, the world model recurrently predicts future states over multiple time steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.908, + 0.45 + ], + "angle": 0, + "content": "Input of World Model. Given \\(N\\) refined trajectories \\(\\hat{\\tau}\\) and the current BEV state \\(\\mathbf{B}_t\\) at time \\(t\\), we construct \\(N\\) state-action pairs as the inputs of world model. We use the trajectory encoder TE in Eq. 1, with shared parameters, to encode \\(\\hat{\\tau}\\) into action embeddings \\(\\mathbf{A}_t\\). Given \\(\\mathbf{A}_t = \\{\\mathbf{a}_t^1, \\mathbf{a}_t^2, \\dots, \\mathbf{a}_t^N\\}\\), each state-action pair is represented as \\((\\mathbf{B}_t, \\mathbf{a}_t^i)\\), where \\(i \\in \\{1, \\dots, N\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.466, + 0.909, + 0.694 + ], + "angle": 0, + "content": "Recurrent Future Prediction. Given the state-action pairs, we use a world model to predict their corresponding future states. In contrast to prior work [40], which predict the future in image space, we build a world model in BEV space as it is much more efficient. In detail, given a state-action pair \\((\\mathbf{B}_t,\\mathbf{a}_t^i)\\), we flatten \\(\\mathbf{B}_t\\) into \\(hw\\) feature vectors. The dimension of each vector is \\(c\\). Then we concatenate these vectors with the action embedding \\(\\mathbf{a}_t^i\\), resulting in a total of \\(hw + 1\\) feature vectors, namely \\(\\mathbf{F}_i\\in \\mathbb{R}^{(hw + 1)\\times c}\\). \\(\\mathbf{F}_i\\) are then fed into the world model, whose network architecture is a transformer encoder, as \\(\\mathbf{F}_{i + 1} = \\mathrm{TransformerEncoder}(\\mathbf{F}_i)\\). The output of the world model consists of \\(h\\times w + 1\\) feature vectors, corresponding to the predicted future states \\(\\mathbf{B}_{t + 1}^i\\) and future actions \\(\\mathbf{a}_{t + 1}^i\\). In summary, we can formulate this process in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.583, + 0.702, + 0.907, + 0.721 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {B} _ {t + 1} ^ {i}, \\mathbf {a} _ {t + 1} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t}, \\mathbf {a} _ {t} ^ {i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.729, + 0.906, + 0.774 + ], + "angle": 0, + "content": "Furthermore, this world model is able to predict the states \\(\\mathbf{B}_{t + K}^{i}\\) of the future \\(K\\) steps in a recurrent manner, as shown by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.784, + 0.907, + 0.802 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathbf {B} _ {t + K} ^ {i}, \\mathbf {a} _ {t + K} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t + K - 1} ^ {i}, \\mathbf {a} _ {t + K - 1} ^ {i}\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "The process above is conducted for every state-action pair in a parallel manner. It is important to note that the size of BEV states is relatively small (e.g., \\( h = w = 8 \\)), so the world model does not introduce significant computational overhead. Next, we introduce the Reward Model, which predicts rewards for a trajectory based on its future states." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27139" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.094, + 0.9, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.426, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Figure 2. Overall structure of WoTE. Our model predicts and evaluates trajectory in an end-to-end manner. It is mainly divided into two parts. (a) Trajectory Prediction: the BEV encoder encodes the multi-modal observations into a BEV feature map (state) and proposes multiple trajectory candidates. (b) Trajectory Evaluation: Given the current BEV state and trajectory candidates, we leverage the BEV World Model to predict the corresponding future BEV states of these trajectories. The Reward Model then predicts rewards with the help of future BEV states. The trajectory with the highest reward is selected as the final trajectory." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.522, + 0.245, + 0.537 + ], + "angle": 0, + "content": "3.3. Reward Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.575 + ], + "angle": 0, + "content": "In this section, we first present the reward types and then describe how they are learned." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.775 + ], + "angle": 0, + "content": "Reward Types. Following [27], we have two types of rewards: the imitation reward and the simulation reward. The imitation reward \\( r_{\\mathrm{im}} \\) measures how well the predicted trajectory mimics an expert trajectory. The detailed implementation will be introduced in Sec. 3.4. The simulation rewards \\( r_{\\mathrm{sim}} \\) measures trajectory quality based on simulator-defined criteria. Following [11], we assess trajectories using the following five criteria: no collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf), and ego progress (EP). As a result, we have \\( r_{\\mathrm{sim}} = \\{r_{\\mathrm{sim}}^{\\mathrm{NC}}, r_{\\mathrm{sim}}^{\\mathrm{DAC}}, r_{\\mathrm{sim}}^{\\mathrm{TTC}}, r_{\\mathrm{sim}}^{\\mathrm{Comf}}, r_{\\mathrm{sim}}^{\\mathrm{EP}}\\} \\). Given \\( r_{\\mathrm{im}} \\) and \\( r_{\\mathrm{sim}} \\), the final reward \\( r_{\\mathrm{final}} \\) is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.779, + 0.483, + 0.833 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} r _ {\\text {f i n a l}} = - \\left(w _ {1} \\log r _ {\\mathrm {i m}} + w _ {2} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {N C}} + w _ {3} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {D A C}} \\right. \\tag {4} \\\\ \\left. + w _ {4} \\log \\left(5 r _ {\\mathrm {s i m}} ^ {\\mathrm {T T C}} + 2 r _ {\\mathrm {s i m}} ^ {\\mathrm {C o m f}} + 5 r _ {\\mathrm {s i m}} ^ {\\mathrm {E P}}\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.839, + 0.295, + 0.854 + ], + "angle": 0, + "content": "where \\(w_{i}\\) are hyper-parameter." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Reward Prediction. Our method needs these rewards for trajectory evaluation in the inference time, so we propose" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.907, + 0.629 + ], + "angle": 0, + "content": "reward prediction. Take the \\(i\\)-th refined trajectory \\(\\hat{\\tau}_i\\) for illustration, the previous approaches [27] typically predict a reward based solely on the current state. With the help of the world model, our reward model predicts the reward \\(\\mathbf{r}_i\\) of \\(i\\)-th trajectory based on both the current BEV state and future \\(K\\) steps BEV states. This process is illustrated in the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.642, + 0.907, + 0.661 + ], + "angle": 0, + "content": "\\[\n\\mathbf {r} _ {i} = \\operatorname {R e w a r d M o d e l} \\left(\\mathbf {B} _ {\\text {a l l}} ^ {i}, \\mathbf {a} _ {\\text {a l l}} ^ {i}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.909, + 0.854 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}_{\\mathrm{all}}^i = \\mathrm{cat}([\\mathbf{B}_t,\\mathbf{B}_{t + 1}^i,\\dots,\\mathbf{B}_{t + K}^i])\\) and \\(\\mathbf{a}_{\\mathrm{all}}^i = \\mathrm{MLP}(\\mathrm{cat}([\\mathbf{a}_t^i,\\mathbf{a}_{t + 1}^i,\\dots,\\mathbf{a}_{t + K}^i]))\\). \\(\\mathrm{cat}([\\cdot])\\) represents the concatenate operation along the channel dimension. In detail, Reward Model consists of three parts. It first applies 2D convolutions to \\(\\mathbf{B}_{\\mathrm{all}}^i\\) to aggregate information from different regions and time steps of the BEV states. Next, a global average pooling is applied to form \\(\\mathbf{B}_{\\mathrm{pool}}\\in \\mathbb{R}^{c}\\), \\(c\\) is the channel dimension. We then fed \\(\\mathbf{S}^i = \\mathrm{cat}([\\mathbf{B}_{\\mathrm{pool}},\\mathbf{a}_{\\mathrm{all}}^i])\\), where \\(\\mathbf{S}^i\\in \\mathbb{R}^{2c}\\), into an MLP head to predict the reward \\(\\mathbf{r}_i\\in \\mathbb{R}^m\\), where \\(m\\) denotes the number of reward. Next, we introduction the definition of these rewards and select a final trajectory based on these rewards." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "The refined trajectory with the highest final reward is selected as the final trajectory. Next, we introduce the supervision of these predicted rewards and BEV states." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27140" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.33, + 0.108 + ], + "angle": 0, + "content": "3.4. Supervision in BEV Space" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.295 + ], + "angle": 0, + "content": "The world model predicts multiple future feature states corresponding to different trajectories. However, supervising these predictions in the image space is challenging, as driving logs provide only a single future trajectory. To address this, we propose supervising in the BEV space. Supervising in BEV space alleviates the difficulty of accurately modeling future states in image space while also improving computational efficiency. Many real-world traffic simulators support traffic simulations in the BEV space. They generate realistic and reliable future BEV scenarios along with rewards. We adopt nuPlan [3] as our traffic simulator for its broad adoption and realistic scenario simulations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.313, + 0.485, + 0.51 + ], + "angle": 0, + "content": "Supervision of BEV States. For BEV state supervision, we decode a semantic BEV map based on the BEV state to enable explicit supervision as Figure 2 (b) shows. We utilize upsampling and transposed convolution layers to decode a BEV semantic map from the BEV state. The dimension of the BEV semantic map is \\( H \\times W \\times L \\), where \\( L \\) denotes the number of semantic classes. This BEV semantic map provides a comprehensive representation of a BEV scenario, covering classes including background, road, walkways, centerline, static objects, vehicles and pedestrians. Each element is represented as a one-hot encoded channel in \\( L \\). For each time step \\( t + k \\), We use Focal Loss [28] to supervise the predicted BEV semantic map \\( \\mathbb{B}_{t + k} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.522, + 0.484, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {B E V}} = \\operatorname {F o c a l L o s s} \\left(\\mathbb {B} _ {t + k}, \\mathbb {B} _ {t + k} ^ {*}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.581 + ], + "angle": 0, + "content": "where \\(\\mathbb{B}_{t + k}^{*}\\) is the ground truth simulated BEV semantic map at time \\(t + k\\) provided by the simulator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.486, + 0.779 + ], + "angle": 0, + "content": "Supervision of Simulation Rewards. Regarding the simulation rewards, we leverage the simulator to produce five rewards for evaluating a trajectory. Specifically, given the \\(i\\)-th trajectory \\(\\hat{\\tau}^i\\), the simulator simulates the future location of the other agents and ego vehicle and produces a simulated future BEV semantic map with agents. Next, it uses a rule-based evaluator to evaluate this simulated future BEV scenarios for producing the corresponding target simulation rewards of \\(\\hat{\\tau}^i\\). Given the rewards provided by the simulator, we use the Binary Cross Entropy (BCE) loss to supervise the predicted simulation reward \\(r_{\\mathrm{sim}}\\). As a result, the loss is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.2, + 0.791, + 0.484, + 0.81 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} = \\mathrm {B C E} \\left(r _ {\\text {s i m}}, r _ {\\text {s i m}} ^ {*}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.485, + 0.87 + ], + "angle": 0, + "content": "Supervision of Imitation Reward. In addition to the supervision from the simulator, our framework is also guided by the expert driver." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "We compute the target of imitation reward by first calculating the L2 distances \\( d_{i} \\) between the \\( i \\)-th trajectory anchor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.234 + ], + "angle": 0, + "content": "and an expert trajectory. Then, we apply the negative of these distances from \\(N\\) trajectory anchors to the softmax function, yielding: \\(r_{\\mathrm{im},i}^{*} = \\frac{\\exp(-d_{i})}{\\sum_{j=1}^{N}\\exp(-d_{j})}\\). Intuitively, a trajectory anchor closer to the expert trajectory than others will receive a higher reward. As a result, the imitation reward \\(r_{\\mathrm{im}}^{*} \\in \\mathbb{R}^{N}\\) quantifies the similarity of each trajectory anchor to the human trajectory. The predicted imitation reward \\(r_{\\mathrm{im}}\\), which is produced by Eq. 5, are supervised using the Cross Entropy loss as" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.244, + 0.907, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\mathrm {i m}} = \\operatorname {C r o s s E n t r o p y} \\left(r _ {\\mathrm {i m}}, r _ {\\mathrm {i m}} ^ {*}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.909, + 0.401 + ], + "angle": 0, + "content": "Supervision of Trajectory. For supervising the refined trajectories, we adopt the commonly used winner-take-all strategy [46]. In this strategy, we select the trajectory anchor that is closest to the expert trajectory. Only the corresponding refined trajectory of this anchor will be supervised. Let this trajectory anchor be denoted as \\(\\tau_{i}\\). We use the L1 loss to measure the difference between the expert trajectory \\(\\tau^{*}\\) and the refined trajectory \\(\\hat{\\tau}_{i}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.412, + 0.907, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t r a j}} = \\left| \\hat {\\tau} _ {i} - \\tau^ {*} \\right|. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.5 + ], + "angle": 0, + "content": "Instead of supervising all refined trajectories, we focus solely on the most optimal one. This strategy helps the model capture a diverse range of trajectory patterns, with each anchor specializing in a particular trajectory modality." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.5, + 0.816, + 0.515 + ], + "angle": 0, + "content": "The overall training loss \\(\\mathcal{L}_{\\mathrm{total}}\\) is as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.526, + 0.907, + 0.543 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {B E V}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {i m}} + \\mathcal {L} _ {\\text {t r a j}}. \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.568, + 0.645, + 0.585 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.641, + 0.607 + ], + "angle": 0, + "content": "4.1. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.795 + ], + "angle": 0, + "content": "NAVSIM Dataset The NAVSIM dataset [11] is constructed based on nuPlan [3]. Specifically, OpenScene [8] first downsampled the nuPlan data from \\(10\\mathrm{Hz}\\) to \\(2\\mathrm{Hz}\\) to condense it into 120 hours of driving logs. NAVSIM resampled the data from OpenScene to emphasize challenging scenarios, reducing simple situations like straight-line driving. The dataset is divided into two parts: Navtrain and Navtest, comprising 1192 scenarios for training and validation and 136 scenarios for testing. Unlike nuScenes [2], which collects data in simpler, slower driving scenarios, NAVSIM ensures that challenging scenarios are prioritized, making simply fitting ego status insufficient for planning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.911, + 0.902 + ], + "angle": 0, + "content": "NAVSIM Metrics The original end-to-end driving metrics were primarily designed to evaluate the deviation between the predicted trajectories and those driven by human experts. However, this evaluation approach has proven inadequate, resulting in poor performance in real-world driving contexts. NAVSIM introduces more practical and reliable metrics, focusing on aligning open-loop and closed-loop metrics." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27141" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.885, + 0.317 + ], + "angle": 0, + "content": "
MethodInput#Traj.Traj. Eval.NC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
Human---10010087.510099.994.8
Constant Velocity-1×69.958.849.349.3100.021.6
Ego Status MLP-1×93.077.362.883.6100.065.6
VADv2 [41]C8192Rule-based97.991.777.692.9100.083.0
UniAD [22]C1Rule-based97.891.978.892.9100.083.4
LTF [31]C1×97.492.879.092.4100.083.8
PARA-Drive [41]C1Rule-based97.992.479.393.099.884.0
TransFuser [31]C & L1×97.792.879.292.8100.084.0
LAW [26]C1×96.495.481.788.799.984.6
DRAMA [43]C & L1×98.093.180.194.8100.085.5
Hydra-MDP [27]C & L8192Model-free98.396.078.794.6100.086.5
WoTEC & L256Model-based98.596.881.994.999.988.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.327, + 0.908, + 0.397 + ], + "angle": 0, + "content": "Table 1. Comparison with the SOTA approaches on NAVSIM test set. Rule-based: UniAD and PARA-Drive use occupancy maps, while VADv2 relies on vectorized maps for trajectory evaluation following specific rules. Model-free: Hydra-MDP evaluates trajectory without world modeling. Model-based: our WoTE evaluates trajectory with the assistance of the BEV world model. NC: no at-fault collision. DAC: drivable area compliance. EP: ego progress. TTC: time-to-collision. Conf.: comfort. PDMS: the predictive driver model score. Traj.: Trajectory. Eval.: Evaluation. C: Camera. L: LiDAR." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.433, + 0.485, + 0.525 + ], + "angle": 0, + "content": "
MethodTraj. Eval.Success Rate↑Driving Score↑
AD-MLP [44]-0.0018.05
UniAD [22]Rule-based16.3645.81
VAD [25]Rule-based15.0042.35
TCP [42]-30.0059.90
WoTEModel-based31.3661.71
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.536, + 0.484, + 0.579 + ], + "angle": 0, + "content": "Table 2. Comparison with SOTA approaches on closed-loop Bench2Drive [24] benchmark on CARLA simulator. Traj.: Trajectory. Eval.: Evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Specifically, NAVSIM evaluates model performance using the Predictive Driver Model Score (PDMS), which is calculated based on five factors: No At-Fault Collision (NC), Drivable Area Compliance (DAC), Time-to-Collision (TTC), Comfort (Comf.), and Ego Progress (EP). The PDMS is calculated as \\(\\mathrm{PDMS} = \\mathrm{NC} \\times \\mathrm{DAC} \\times \\frac{5 \\times \\mathrm{EP} + 5 \\times \\mathrm{TTC} + 2 \\times \\mathrm{C}}{12}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Bench2Drive Dataset and Metrics To further evaluate our framework, we conduct closed-loop experiments in the CARLA simulator [13]. Specifically, we adopt the recently proposed Bench2Drive [24] benchmark as our closed-loop evaluation benchmark. Bench2Drive provides a standardized training dataset on CARLA, ensuring fair comparisons across different methods. The benchmark consists of 220 short evaluation routes (around \\(150\\mathrm{m}\\) each) distributed across all CARLA towns, with each route containing one safety-critical scenario. This design effectively reduces the evaluation variance. For metric, it utilizes the standard CARLA metric, Driving Score (DS), as the primary metric. Additionally, Bench2Drive reports the success rate, which represents the proportion of successfully completed routes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.435, + 0.907, + 0.465 + ], + "angle": 0, + "content": "A route is considered successful if and only if DS reaches \\(100\\%\\) on this route." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.731, + 0.497 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.506, + 0.909, + 0.852 + ], + "angle": 0, + "content": "NAVSIM For input data, it is aligned with TransFuser [31]. We concatenate the front-view image with center-cropped front-left and front-right images, resulting in a combined resolution of \\(256 \\times 1024\\) pixels. For LiDAR, the \\(64\\mathrm{m} \\times 64\\mathrm{m}\\) point cloud surrounding the ego vehicles is used. For network architecture, we employ ResNet34 as the backbone for BEV feature extraction following TransFuser [31]. The world model consists of two transformer decoder layers. We use \\(N = 256\\) trajectory anchors. The predicted future states are of time \\(t + 2s, t + 4s\\) when the current is of time \\(t\\). The reward weights are set as follows: \\(w_{1} = 0.1, w_{2} = w_{3} = 0.5, w_{4} = 1.0\\). The training is conducted on the Navtrain split using 8 NVIDIA L20 GPUs with a total batch size of 256, distributed across 50 epochs. In the ablation study section, we train our model for 20 epochs instead of 50 to accelerate the training process. For fast training, we pre-compute the simulation results (BEV semantic maps and scores) based on the trajectory anchors. We then feed trajectory anchors into the trajectory evaluation module during training. During testing, the trajectory evaluation module evaluates the refined trajectory instead of the anchor trajectories. We utilize the Adam optimizer with a learning rate of 1e-4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Bench2Drive For Bench2Drive, we adopt TCP [42] as our baseline and integrate the trajectory evaluation module. We choose TCP as it is the best open-source framework officially" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27142" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.142 + ], + "angle": 0, + "content": "
Traj. Eval.Future StatesNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
××96.491.576.290.399.981.0
✓×97.192.978.291.9100.083.2
✓✓98.094.779.993.4100.085.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.152, + 0.483, + 0.181 + ], + "angle": 0, + "content": "Table 3. Ablation study on trajectory evaluation and future state prediction. Traj: Trajectory. Eval.: Evaluation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.219, + 0.487, + 0.356 + ], + "angle": 0, + "content": "provided by Bench2Drive. Similar to NAVSIM, we utilize imitation rewards along with simulation-based rewards, including No Collision (NC) and Drivable Area Compliance (DAC). The number of trajectory anchors is set to 256. Notably, rather than employing the winner-take-all strategy (Eq. 9) to supervise only the best trajectory, we find that supervising all trajectories leads to improved performance. The model is trained for 27 epochs with a batch size of 300. We use the Adam optimizer with a learning rate of 1e-4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.364, + 0.313, + 0.381 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.485, + 0.491 + ], + "angle": 0, + "content": "NASIM Table 1 compares our method with state-of-the-art approaches on the NAVSIM test set. By leveraging the world model to simulate future scenarios and guide trajectory evaluation, our method achieves a PDMS of 87.1. WoTE outperforms the previous model-free approach, Hydra-MDP, highlighting the advantages of model-based trajectory evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.485, + 0.569 + ], + "angle": 0, + "content": "Bench2Drive To further validate our approach, we conduct a closed-loop evaluation on the Bench2Drive benchmark within the CARLA simulator. As shown in Table 2, our method improves the Driving Score by 1.81 points, demonstrating its effectiveness in various evaluation settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.577, + 0.258, + 0.592 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.486, + 0.78 + ], + "angle": 0, + "content": "Trajectory evaluation and future state prediction matter. In this experiment, we investigate the importance of the trajectory evaluation and future state prediction as Table 3 shows. The first row in Table 3 represents our baseline TransFuser [31], which performs trajectory prediction only. In the second row, we change to our framework, which consists of both trajectory prediction and trajectory evaluation. However, we do not used the predicted future states as the input of Reward Model. That is, \\(\\mathbf{B}_{\\mathrm{all}} = \\mathbf{B}_t\\) and \\(\\mathbf{a}_{\\mathrm{all}} = \\mathbf{a}_t\\) in Eq. 5. In the third row, we integrate the BEV world model to predict future states, leading to notable performance improvements across multiple metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Imitation and simulation rewards are complementary. To investigate how each type of reward affects driving performance, we conduct an ablation study, as shown in Table 4. During inference, we adjust the reward weights in Eq. 4 to isolate their effects: setting \\( w_{1} = 0 \\) enables evaluation with only simulation rewards, while setting \\( w_{2} = w_{3} = w_{4} = 0 \\) tests performance with only an imitation reward. The results indicate that imitation and simulation rewards emphasize" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.905, + 0.14 + ], + "angle": 0, + "content": "
Imi. RewardSim. RewardsNC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
✓×97.992.577.793.499.983.5
×✓96.594.779.290.199.483.6
✓✓98.094.779.993.4100.085.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.15, + 0.908, + 0.193 + ], + "angle": 0, + "content": "Table 4. Ablation study on the imitation and simulation rewards. The experiment shows that imitation and simulation rewards are complementary. Imi: Imitation. Sim.: Simulation." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.265 + ], + "angle": 0, + "content": "
Future Prediction StepsNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
0s → 4s97.493.679.191.9100.084.0
0s → 2s → 4s98.094.779.993.4100.085.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.275, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Table 5. Ablation study on the different number of time steps in the future prediction. Our BEV world model predicts future states in a recurrent manner. \\( \\mathrm{A} \\rightarrow \\mathrm{B} \\): the predicted future state at time B is based on the state at time A." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.36, + 0.905, + 0.416 + ], + "angle": 0, + "content": "
#Trajs.NC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑Latency ↓
6497.494.179.691.9100.084.517.2 ms
12897.994.680.093.0100.085.417.9 ms
25698.094.779.993.4100.085.618.7 ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.425, + 0.91, + 0.482 + ], + "angle": 0, + "content": "Table 6. Ablation study on the different number of Trajectories. The number of trajectories is the same during training and testing. The latency is evaluated on an NVIDIA L20 GPU. Trajs.: Trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.6 + ], + "angle": 0, + "content": "different aspects of planning. Trajectory evaluation using imitation rewards performs better in NC and TTC, while using simulation rewards excels in DAC and EP. By integrating both, the model leverages their respective strengths, achieving improved overall performance across all key metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Recurrent future state prediction helps. Our framework supports predicting future states recurrently, as detailed in Eq. 3. In Table 5, we present the results obtained by varying the number of BEV states predicted by the world model. Our findings indicate that using a finer time step for future state prediction significantly improves performance. This is because a finer time step provides richer temporal information, which aids trajectory evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.911, + 0.903 + ], + "angle": 0, + "content": "More trajectories help. This ablation study explores the effect of varying the number of trajectories on overall performance. Specifically, the model in each row is trained and tested under the same number of trajectories, and the world model is fed with refined trajectories. As shown in Table 6, using a small number of trajectories, such as 64, significantly degrades performance. As the number of trajectories increases, performance improves notably, with a marked improvement observed when increasing from 64 to 128 trajectories. However, as the number of trajectories increases from 128 to 256, the performance gains diminish," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27143" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.358, + 0.133 + ], + "angle": 0, + "content": "
Train Traj.Test Traj.Traj. Refine†NC ↑DAC ↑EP ↑
256256×96.694.679.0
2561024×97.295.679.3
256256✓98.094.779.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.144, + 0.358, + 0.171 + ], + "angle": 0, + "content": "Table 7. Generalization of the trajectory across different trajectory configurations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.358, + 0.2 + ], + "angle": 0, + "content": "the input trajectories of BEV world model a trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.24, + 0.359, + 0.392 + ], + "angle": 0, + "content": "indicating that the performance of the most optimal level. Based on these results, we default number of trajectories for our ex- Latency analysis. A common concern is of our framework. To address this, we run on an NVIDIA L20 GPU, as shown in GPU processes state-action pairs in para remains efficient even as the number of t The total latency is only \\(18.7\\mathrm{ms}\\) well \\(\\mathsf{V}\\) requirements for end-to-end autonomous" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.404, + 0.358, + 0.419 + ], + "angle": 0, + "content": "Generalization ability across different" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.42, + 0.484, + 0.601 + ], + "angle": 0, + "content": "shown in Table 7, we investigate the generalization ability of our trajectory evaluation module. To be specific, we sample \\( N = 1024 \\) instead of \\( N = 256 \\) trajectory anchors from the dataset. Since our model is trained under the setting of \\( N = 256 \\) trajectory anchors, the model has never seen these \\( N = 1024 \\) trajectories. Surprisingly, we find a 1.3 PDMS increase when using these unseen trajectory anchors. Besides, the model performs even better when using refined trajectories as inputs. These refined trajectories continuously change based on varying sensor inputs. This shows that the proposed world model has great generalization ability when dealing with unseen trajectories." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.613, + 0.246, + 0.629 + ], + "angle": 0, + "content": "4.5. Visual Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.484, + 0.696 + ], + "angle": 0, + "content": "End-to-end planning trajectories. To highlight the improvements in end-to-end driving achieved through trajectory evaluation, we present visualizations of end-to-end planning trajectories in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.707, + 0.485, + 0.737 + ], + "angle": 0, + "content": "Trajectory rewards. Figure 4 illustrates the predicted rewards for different trajectories." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.754, + 0.21, + 0.769 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this paper, we introduced a novel framework that leverages a BEV world model for end-to-end trajectory evaluation in autonomous driving. By integrating a BEV world model, we enable a more informed evaluation process that considers the dynamic evolution of driving scenarios, leading to more effective trajectory evaluation. Our approach benefits from dense supervision provided by BEV-space traffic simulators, which supply both semantic future states and rule-based" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.091, + 0.9, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.295, + 0.908, + 0.365 + ], + "angle": 0, + "content": "Figure 3. Visualization of end-to-end planning trajectories. We compare our method with TransFuser [31], which serves as our trajectory prediction baseline without a trajectory evaluation module. Our trajectory evaluation module effectively filters out low-quality trajectories and helps avoid collisions." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.381, + 0.895, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.589, + 0.501, + 0.792, + 0.513 + ], + "angle": 0, + "content": "(a) Multi-modality trajectory preservation" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.514, + 0.895, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.595, + 0.633, + 0.783, + 0.645 + ], + "angle": 0, + "content": "(b) Low-quality trajectory suppression" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.662, + 0.908, + 0.773 + ], + "angle": 0, + "content": "Figure 4. Visualization of rewards of all trajectories. Brighter colors indicate higher trajectory rewards, while gray trajectories correspond to those with very low rewards. (a): WoTE effectively retains diverse multi-modal trajectories, as long as they comply with traffic rules and do not result in unsafe outcomes. (b): it demonstrates that trajectories considered unreasonable receive significantly lower rewards, highlighting the effectiveness of our trajectory evaluation module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "reward targets. Our method achieves state-of-the-art performance on the NAVSIM and Bench2Drive benchmarks while maintaining real-time efficiency. Overall, our work establishes trajectory evaluation as an important research direction for end-to-end autonomous driving. We hope our framework serves as a strong baseline and inspires further research in end-to-end online trajectory evaluation." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27144" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.176 + ], + "angle": 0, + "content": "This work was supported in part by the National Science and Technology Major Project (No. 2022ZD0116500), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.192, + 0.188, + 0.208 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.217, + 0.486, + 0.271 + ], + "angle": 0, + "content": "[1] Sébastien MR Arnold, Pierre L'Ecuyer, Liyu Chen, Yi-fan Chen, and Fei Sha. Policy learning and evaluation with randomized quasi-monte carlo. arXiv preprint arXiv:2202.07808, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.274, + 0.485, + 0.33 + ], + "angle": 0, + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. CVPR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.332, + 0.484, + 0.4 + ], + "angle": 0, + "content": "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.485, + 0.444 + ], + "angle": 0, + "content": "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.485, + 0.487 + ], + "angle": 0, + "content": "[5] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In CVPR, pages 15590-15599, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.484, + 0.544 + ], + "angle": 0, + "content": "[6] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.484, + 0.614 + ], + "angle": 0, + "content": "[7] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.715 + ], + "angle": 0, + "content": "[9] Christoph Dann, Gerhard Neumann, and Jan Peters. Policy evaluation with temporal differences: A survey and comparison. The Journal of Machine Learning Research, 15(1): 809-883, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[10] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. arXiv preprint arXiv:2406.15349, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Christopher Diehl, Timo Sievernich, Martin Krüger, Frank Hoffmann, and Torsten Bertram. Umbrella: Uncertainty-aware model-based offline reinforcement learning leveraging planning. arXiv preprint arXiv:2111.11097, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.217, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.137, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[14] Haoyang Fan, Fan Zhu, Changchun Liu, Liangliang Zhang, Li Zhuang, Dong Li, Weicheng Zhu, Jiangtao Hu, Hongye Li, and Qi Kong. Baidu Apollo em motion planner. arXiv preprint arXiv:1807.08048, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[15] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.907, + 0.332 + ], + "angle": 0, + "content": "[16] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.907, + 0.389 + ], + "angle": 0, + "content": "[17] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[18] Mikael Henaff, Alfredo Canziani, and Yann LeCun. Model-predictive policy learning with uncertainty regularization for driving in dense traffic. In International Conference on Learning Representations, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.504 + ], + "angle": 0, + "content": "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.907, + 0.574 + ], + "angle": 0, + "content": "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[21] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.907, + 0.757 + ], + "angle": 0, + "content": "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[24] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.872 + ], + "angle": 0, + "content": "[25] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27145" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[27] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.233 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dolkar. Focal loss for dense object detection. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.303 + ], + "angle": 0, + "content": "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.305, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[30] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, pages 7077-7087, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.389 + ], + "angle": 0, + "content": "[31] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.484, + 0.447 + ], + "angle": 0, + "content": "[32] Guanren Qiao, Guorui Quan, Rongxiao Qu, and Guiliang Liu. Modelling competitive behaviors in autonomous driving under generative world model. In European Conference on Computer Vision, pages 19-36. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.484, + 0.517 + ], + "angle": 0, + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[34] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[35] Ardi Tampuu, Tambet Matiisen, Maksym Semikin, Dmytro Fishman, and Naveed Muhammad. A survey of end-to-end driving: Architectures and training methods. IEEE Transactions on Neural Networks and Learning Systems, 33(4): 1364-1384, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[36] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[37] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.786 + ], + "angle": 0, + "content": "[38] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[39] Yuqi Wang, Ke Cheng, Jiawei He, Qitai Wang, Hengchen Dai, Yuntao Chen, Fei Xia, and Zhaoxiang Zhang. Drivingndojo dataset: Advancing interactive and knowledge-enriched driving world model. arXiv preprint arXiv:2410.10738, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Con-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ference on Computer Vision and Pattern Recognition, pages 14749-14759, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[41] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.246 + ], + "angle": 0, + "content": "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: a simple yet strong baseline. NeurIPS, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.318 + ], + "angle": 0, + "content": "[43] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[44] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.907, + 0.457 + ], + "angle": 0, + "content": "[45] Yumeng Zhang, Shi Gong, Kaixin Xiong, Xiaoqing Ye, Xiao Tan, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Bevworld: A multimodal world model for autonomous driving via unified bev latent space. arXiv preprint arXiv:2407.05679, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.46, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27146" + } + ] +] \ No newline at end of file diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_origin.pdf b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0a267a0d7b7ef6090cb15711f5531ec15067da1d --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/24c54d04-9c44-4257-9691-2d6e9dd0c2c8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1de1791bc8141749dac37ae43d8c7ff121d02d2b489695f9aff773e8873b73cf +size 1007625 diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/full.md b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1c8e76cfeda0a8d33ee362a6abd1d9ed93ee4855 --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/full.md @@ -0,0 +1,313 @@ +# End-to-End Driving with Online Trajectory Evaluation via BEV World Model + +Yingyan Li* Yuqi Wang* Yang Liu Jiawei He Lue Fan* Zhaoxiang Zhang* NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA) State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) University of Chinese Academy of Sciences (UCAS) + +# Abstract + +End-to-end autonomous driving has achieved remarkable progress by integrating perception, prediction, and planning into a fully differentiable framework. Yet, to fully realize its potential, an effective online trajectory evaluation is in dispensable to ensure safety. By forecasting the future outcomes of a given trajectory, trajectory evaluation becomes much more effective. This goal can be achieved by employing a world model to capture environmental dynamic and predict future states. Therefore, we propose an end-to-end driving framework WoTE, which leverages a BE World model to predict future BEV states for Trajectory Evaluation. The proposed BEV world model is latency efficient compared to image-level world models and can be seamlessly supervised using off-the-shelf BEV-space traffic simulators. We validate our framework on both the NAVSIM benchmark and the closed-loop Bench2Drive benchmark based on the CARLA simulator, achieving state-of-the-art performance. Code is released at https://github.com/liyingyanUCAS/WoTE. + +# 1. Introduction + +End-to-end autonomous driving [6, 30, 35] has achieved remarkable success in recent years and garnered widespread attention. While most approaches [21, 23, 25] primarily focus on predicting a high-quality trajectory, recent works [7, 27] suggest that predicting multiple trajectories simultaneously better reflects the multi-modal nature of driving and leads to improved performance. Given these multiple possible trajectories, it becomes crucial to effectively evaluate them in order to ensure the safety and reliability of end-to-end autonomous driving systems. + +Traditional trajectory evaluation methods are rule-based [10, 14, 20, 37] and rely on perception results, such as bounding boxes and maps, to assess trajectories. These methods are sensitive to perception inaccuracies and may + +![](images/0e3c1fc45288ff454935d5c641bb01389b75cbc08501912a7f1adb1ddf7dd0df.jpg) + +![](images/d73d50b4b000a502be8c230e7e71a5b7e1b34d3f498c0b249cdf8fbc45eb9fbb.jpg) +Figure 1. The illustration of end-to-end trajectory evaluation with BEV world model. (a) Previous end-to-end driving methods primarily focus on learning a high-quality trajectory. (b) Our WoTE consists of end-to-end trajectory prediction and evaluation. First, we predict multi-modal trajectories. Next, given these trajectories, we leverage the BEV world model to predict the corresponding future BEV states. The reward model evaluates these trajectories based on the predicted states and assigns rewards, selecting the highest-reward trajectory. + +not be directly optimized in an end-to-end manner. Recently, some end-to-end driving approaches [7, 27] have incorporated trajectory evaluation. Their evaluations rely on the current state. However, evaluating a trajectory candidate is particularly challenging without knowledge of the future led by this trajectory. Similar to human drivers, who anticipate futures before making decisions, a trajectory evaluation network would benefit from predicting the corresponding future states of trajectory candidates. + +To effectively predict future states, it is intuitive to adopt methodologies from reinforcement learning, where world models have successfully facilitated future state predictions. In reinforcement learning, model-based methods [5, 18, 32] have consistently shown superior performance compared to model-free approaches [17, 29], largely due to their ability of modeling the environmental dynamics and reasoning about future scenarios through world models. Inspired by this success, we propose adopting a world model [15, 19, 38-40] + +within end-to-end autonomous driving frameworks to facilitate accurate prediction of future states, thus significantly improving trajectory evaluation. + +However, leveraging the world model for trajectory evaluation presents some technical issues. First, a good representation of future scenarios is needed. Many existing world models for driving [19, 38, 40] predict the images of future scenarios with diffusion models [33], which can be time-consuming and not suitable for real-time driving. Second, we lack supervision for future states. Supervising the predicted future states of the world model is difficult because world models need to imagine multiple future states based on the multiple trajectory candidates, while only one future state is available in the real-world datasets. + +To address these technical issues, we propose WoTE as shown in Fig.1. We propose to predict future states in the BEV space. The information in BEV space is much more compact than in raw images, which enables a more efficient single-step feed-forward future prediction than the time-consuming multi-step denoising in previous driving world models [19, 38, 40, 45]. This characteristic of BEV space makes WoTE highly suitable for real-time driving applications. Additionally, the predicted future BEV states can be easily supervised using off-the-shelf BEV-space traffic simulators such as nuPlan [3, 16]. These simulators model the scenario in BEV space by representing the map and objects using BEV semantic maps, creating a dynamic simulation of the driving environment. Consequently, we can use the BEV semantic maps from these simulated scenarios as supervision for the prediction of the BEV world model. Moreover, these simulators are equipped with well-established evaluation rules to assess the simulated future BEV scenarios, which provide target rewards. + +We validate our framework on both NAVSIM [11] benchmark and closed-loop Bench2Drive [24] benchmark based on the CARLA [13] simulator and achieve state-of-the-art performance. Our experiments show that incorporating future states into trajectory evaluation is essential for improved performance. Additionally, the trajectory evaluation module, built on the world model shows great generalization ability across diverse trajectories. Our contributions are as follows: + +- We highlight the importance of utilizing imagined future states for trajectory evaluation in end-to-end autonomous driving. +- We introduce WoTE, which consists of an end-to-end trajectory evaluation module based on a BEV world model. The world model in BEV space enables efficient real-time future prediction and overcomes the scarcity of multi-future supervisions. +- WoTE is validated on both the NAVSIM [11] benchmark and the closed-loop CARLA-based Bench2Drive [24] benchmark, achieving state-of-the-art performance. + +# 2. Related Works + +# 2.1. End-to-end Autonomous Driving + +End-to-end autonomous driving [7, 22, 25, 43] is to train a model that directly maps sensor inputs to trajectories or control signals instead of decomposing the driving task into traditional sub-tasks. These works can be divided into two categories according to the training paradigm: imitation-learning-based and reinforcement-learning-based. Imitation learning is the most common practice in end-to-end autonomous driving. Their predicted trajectories are supervised by the expert trajectories. For instance, P3 [34], takes a differentiable semantic occupancy as the intermediate representation and learns actions from human trajectories. Following this, ST-P3 [21] learns spatial-temporal features for perception, prediction and planning tasks simultaneously. UniAD [22] proposed goal-oriented planning, combining and leveraging advantages of perception and motion prediction modules. VAD [25] introduces vectorized representations for end-to-end planning. There are also some reinforcement-learning-based methods. For example, MaRLn [36] utilizes implicit affordances to develop a reinforcement learning algorithm. LBC [4] trained a vision-based sensorimotor agent from the privileged agent. Combining imitation learning and reinforcement learning, Hydra-MDP [27] utilizes knowledge distillation to obtain supervisions from both rule-based planners and human drivers. It is a promising research direction as more supervisions are utilized. + +# 2.2. Trajectory Evaluation + +Trajectory evaluation is essential for ensuring the reliability and safety of autonomous driving systems. We divide the main trajectory evaluation methods into two categories: model-free and model-based. Model-free methods directly assess the learned policy using observed data without explicitly modeling environmental dynamics. For instance, Monte Carlo policy evaluation [1] uses the empirical mean return as the metric. The value function approximates the expectation based on sampled returns. Temporal-difference learning [9] uses bootstrapping, where the current estimate of the value function is used to generate target values for updating the value function. Subsequently, model-based trajectory evaluation methods [12] have gained increasing popularity in recent years. By leveraging learned world models that capture environmental dynamics, these methods enable more accurate and predictive policy evaluation. For example, PPUU [18] measures policy performance by quantifying uncertainty in predicted future states. Rails [5] evaluates trajectories through a tabular dynamic programming mechanism, and MARL-CCE [32] employs learned environmental dynamics to assess multi-agent policies in driving scenarios. However, these methods often rely on explicit trajectory representations and non-differentiable metrics, which limits + +their end-to-end optimization capabilities. In contrast, our approach evaluates trajectories using BEV features and encoded trajectory embeddings, enabling fully differentiable and thus end-to-end trajectory evaluation. + +# 3. Method + +In this section, we introduce four key components of WoTE: Trajectory Predictor (Sec. 3.1), BEV World Model (Sec. 3.2), Reward Model (Sec. 3.3) and BEV Space Supervision (Sec. 3.4). As illustrated in Figure 2, our method integrates future state prediction via the BEV World Model and reward prediction via the Reward Model within a unified framework. Specifically, the Trajectory Predictor encodes the multi-modal observations into BEV state, represented as a BEV feature map, and provides multiple trajectory proposals. The BEV World Model predicts corresponding future BEV states based on these trajectory proposals. With the help of future states, the Reward Model predicts the reward of each trajectory. Finally, we choose the trajectory with the highest reward as the final trajectory. + +# 3.1. Trajectory Predictor + +As shown in Figure 2 (a), the Trajectory Predictor aims to predict multiple trajectory candidates. First, the BEV encoder encodes the multi-modal sensor inputs into a unified BEV feature map. Then, given trajectory anchors, the trajectory refinement module refines trajectories based on the BEV feature map. We will now describe each part in detail. + +Multi-modal BEV Encoder. Following TransFuser [31], our method leverages multi-modal sensor inputs, including LiDAR and multi-view RGB images. We use a BEV encoder [31] to transform the current multi-modal inputs into a unified BEV feature map $\mathbf{B} \in \mathbb{R}^{h \times w \times c}$ . For convenience in the following text, we refer to the BEV feature map as the BEV state. $h$ and $w$ represent the height and width of the BEV state, while $c$ denotes the number of channels. + +Trajectory Anchors. Following methods in [7, 27], we generate trajectory anchors $\tau$ by K-Means clustering. This clustering is performed on all the expert trajectories collected from the training dataset. The number of trajectory anchors is denoted as $N$ . + +Trajectory Refinement. As illustrated in Figure 2 (a), the trajectory refinement module takes the trajectory anchors and the current BEV states as inputs to predict the refined trajectories. Specifically, we first use a trajectory encoder TE, which is implemented by an MLP, to encode the trajectory anchors $\tau$ into feature vectors. Then, these feature vectors are used as queries to perform cross attention with the current BEV state B, which serves as both the key and + +value. The output of the cross-attention is then fed into an MLP head to predict the offsets. The offset is added to the trajectory anchors to produce the refined trajectories. The entire process can be expressed as: + +$$ +\hat {\tau} = \tau + \operatorname {M L P} (\text {C r o s s A t t e n t i o n} (\mathbf {T E} (\tau), \mathbf {B}, \mathbf {B})), \tag {1} +$$ + +where $\hat{\tau}$ is the refined trajectories. + +# 3.2. BEV World Model + +In this section, we use the BEV World Model to predict the future BEV states, as illustrated in Figure 2 (b). The process is divided into two steps. First, the current BEV state and refined trajectories are combined to form the inputs for the world model. Next, the world model recurrently predicts future states over multiple time steps. + +Input of World Model. Given $N$ refined trajectories $\hat{\tau}$ and the current BEV state $\mathbf{B}_t$ at time $t$ , we construct $N$ state-action pairs as the inputs of world model. We use the trajectory encoder TE in Eq. 1, with shared parameters, to encode $\hat{\tau}$ into action embeddings $\mathbf{A}_t$ . Given $\mathbf{A}_t = \{\mathbf{a}_t^1, \mathbf{a}_t^2, \dots, \mathbf{a}_t^N\}$ , each state-action pair is represented as $(\mathbf{B}_t, \mathbf{a}_t^i)$ , where $i \in \{1, \dots, N\}$ . + +Recurrent Future Prediction. Given the state-action pairs, we use a world model to predict their corresponding future states. In contrast to prior work [40], which predict the future in image space, we build a world model in BEV space as it is much more efficient. In detail, given a state-action pair $(\mathbf{B}_t,\mathbf{a}_t^i)$ , we flatten $\mathbf{B}_t$ into $hw$ feature vectors. The dimension of each vector is $c$ . Then we concatenate these vectors with the action embedding $\mathbf{a}_t^i$ , resulting in a total of $hw + 1$ feature vectors, namely $\mathbf{F}_i\in \mathbb{R}^{(hw + 1)\times c}$ . $\mathbf{F}_i$ are then fed into the world model, whose network architecture is a transformer encoder, as $\mathbf{F}_{i + 1} = \mathrm{TransformerEncoder}(\mathbf{F}_i)$ . The output of the world model consists of $h\times w + 1$ feature vectors, corresponding to the predicted future states $\mathbf{B}_{t + 1}^i$ and future actions $\mathbf{a}_{t + 1}^i$ . In summary, we can formulate this process in the following equation: + +$$ +\left(\mathbf {B} _ {t + 1} ^ {i}, \mathbf {a} _ {t + 1} ^ {i}\right) = \operatorname {W o r l d M o d e l} \left(\mathbf {B} _ {t}, \mathbf {a} _ {t} ^ {i}\right), \tag {2} +$$ + +Furthermore, this world model is able to predict the states $\mathbf{B}_{t + K}^{i}$ of the future $K$ steps in a recurrent manner, as shown by the following equation: + +$$ +\left(\mathbf {B} _ {t + K} ^ {i}, \mathbf {a} _ {t + K} ^ {i}\right) = \operatorname {W o r l d M o d e l} \left(\mathbf {B} _ {t + K - 1} ^ {i}, \mathbf {a} _ {t + K - 1} ^ {i}\right). \tag {3} +$$ + +The process above is conducted for every state-action pair in a parallel manner. It is important to note that the size of BEV states is relatively small (e.g., $h = w = 8$ ), so the world model does not introduce significant computational overhead. Next, we introduce the Reward Model, which predicts rewards for a trajectory based on its future states. + +![](images/b7d10dd316364701d4253cc9d914004d67f33f6eb35bbeaa4c4ce313239d36eb.jpg) +Figure 2. Overall structure of WoTE. Our model predicts and evaluates trajectory in an end-to-end manner. It is mainly divided into two parts. (a) Trajectory Prediction: the BEV encoder encodes the multi-modal observations into a BEV feature map (state) and proposes multiple trajectory candidates. (b) Trajectory Evaluation: Given the current BEV state and trajectory candidates, we leverage the BEV World Model to predict the corresponding future BEV states of these trajectories. The Reward Model then predicts rewards with the help of future BEV states. The trajectory with the highest reward is selected as the final trajectory. + +# 3.3. Reward Model + +In this section, we first present the reward types and then describe how they are learned. + +Reward Types. Following [27], we have two types of rewards: the imitation reward and the simulation reward. The imitation reward $r_{\mathrm{im}}$ measures how well the predicted trajectory mimics an expert trajectory. The detailed implementation will be introduced in Sec. 3.4. The simulation rewards $r_{\mathrm{sim}}$ measures trajectory quality based on simulator-defined criteria. Following [11], we assess trajectories using the following five criteria: no collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf), and ego progress (EP). As a result, we have $r_{\mathrm{sim}} = \{r_{\mathrm{sim}}^{\mathrm{NC}}, r_{\mathrm{sim}}^{\mathrm{DAC}}, r_{\mathrm{sim}}^{\mathrm{TTC}}, r_{\mathrm{sim}}^{\mathrm{Comf}}, r_{\mathrm{sim}}^{\mathrm{EP}}\}$ . Given $r_{\mathrm{im}}$ and $r_{\mathrm{sim}}$ , the final reward $r_{\mathrm{final}}$ is defined as + +$$ +\begin{array}{l} r _ {\text {f i n a l}} = - \left(w _ {1} \log r _ {\mathrm {i m}} + w _ {2} \log r _ {\mathrm {s i m}} ^ {\mathrm {N C}} + w _ {3} \log r _ {\mathrm {s i m}} ^ {\mathrm {D A C}} \right. \tag {4} \\ \left. + w _ {4} \log \left(5 r _ {\mathrm {s i m}} ^ {\mathrm {T T C}} + 2 r _ {\mathrm {s i m}} ^ {\mathrm {C o m f}} + 5 r _ {\mathrm {s i m}} ^ {\mathrm {E P}}\right)\right), \\ \end{array} +$$ + +where $w_{i}$ are hyper-parameter. + +Reward Prediction. Our method needs these rewards for trajectory evaluation in the inference time, so we propose + +reward prediction. Take the $i$ -th refined trajectory $\hat{\tau}_i$ for illustration, the previous approaches [27] typically predict a reward based solely on the current state. With the help of the world model, our reward model predicts the reward $\mathbf{r}_i$ of $i$ -th trajectory based on both the current BEV state and future $K$ steps BEV states. This process is illustrated in the following equation: + +$$ +\mathbf {r} _ {i} = \operatorname {R e w a r d M o d e l} \left(\mathbf {B} _ {\text {a l l}} ^ {i}, \mathbf {a} _ {\text {a l l}} ^ {i}\right), \tag {5} +$$ + +where $\mathbf{B}_{\mathrm{all}}^i = \mathrm{cat}([\mathbf{B}_t,\mathbf{B}_{t + 1}^i,\dots,\mathbf{B}_{t + K}^i])$ and $\mathbf{a}_{\mathrm{all}}^i = \mathrm{MLP}(\mathrm{cat}([\mathbf{a}_t^i,\mathbf{a}_{t + 1}^i,\dots,\mathbf{a}_{t + K}^i]))$ . $\mathrm{cat}([\cdot])$ represents the concatenate operation along the channel dimension. In detail, Reward Model consists of three parts. It first applies 2D convolutions to $\mathbf{B}_{\mathrm{all}}^i$ to aggregate information from different regions and time steps of the BEV states. Next, a global average pooling is applied to form $\mathbf{B}_{\mathrm{pool}}\in \mathbb{R}^{c}$ , $c$ is the channel dimension. We then fed $\mathbf{S}^i = \mathrm{cat}([\mathbf{B}_{\mathrm{pool}},\mathbf{a}_{\mathrm{all}}^i])$ , where $\mathbf{S}^i\in \mathbb{R}^{2c}$ , into an MLP head to predict the reward $\mathbf{r}_i\in \mathbb{R}^m$ , where $m$ denotes the number of reward. Next, we introduction the definition of these rewards and select a final trajectory based on these rewards. + +The refined trajectory with the highest final reward is selected as the final trajectory. Next, we introduce the supervision of these predicted rewards and BEV states. + +# 3.4. Supervision in BEV Space + +The world model predicts multiple future feature states corresponding to different trajectories. However, supervising these predictions in the image space is challenging, as driving logs provide only a single future trajectory. To address this, we propose supervising in the BEV space. Supervising in BEV space alleviates the difficulty of accurately modeling future states in image space while also improving computational efficiency. Many real-world traffic simulators support traffic simulations in the BEV space. They generate realistic and reliable future BEV scenarios along with rewards. We adopt nuPlan [3] as our traffic simulator for its broad adoption and realistic scenario simulations. + +Supervision of BEV States. For BEV state supervision, we decode a semantic BEV map based on the BEV state to enable explicit supervision as Figure 2 (b) shows. We utilize upsampling and transposed convolution layers to decode a BEV semantic map from the BEV state. The dimension of the BEV semantic map is $H \times W \times L$ , where $L$ denotes the number of semantic classes. This BEV semantic map provides a comprehensive representation of a BEV scenario, covering classes including background, road, walkways, centerline, static objects, vehicles and pedestrians. Each element is represented as a one-hot encoded channel in $L$ . For each time step $t + k$ , We use Focal Loss [28] to supervise the predicted BEV semantic map $\mathbb{B}_{t + k}$ as + +$$ +\mathcal {L} _ {\mathrm {B E V}} = \operatorname {F o c a l L o s s} \left(\mathbb {B} _ {t + k}, \mathbb {B} _ {t + k} ^ {*}\right), \tag {6} +$$ + +where $\mathbb{B}_{t + k}^{*}$ is the ground truth simulated BEV semantic map at time $t + k$ provided by the simulator. + +Supervision of Simulation Rewards. Regarding the simulation rewards, we leverage the simulator to produce five rewards for evaluating a trajectory. Specifically, given the $i$ -th trajectory $\hat{\tau}^i$ , the simulator simulates the future location of the other agents and ego vehicle and produces a simulated future BEV semantic map with agents. Next, it uses a rule-based evaluator to evaluate this simulated future BEV scenarios for producing the corresponding target simulation rewards of $\hat{\tau}^i$ . Given the rewards provided by the simulator, we use the Binary Cross Entropy (BCE) loss to supervise the predicted simulation reward $r_{\mathrm{sim}}$ . As a result, the loss is formulated as + +$$ +\mathcal {L} _ {\text {r e w a r d}} ^ {\text {s i m}} = \mathrm {B C E} \left(r _ {\text {s i m}}, r _ {\text {s i m}} ^ {*}\right). \tag {7} +$$ + +Supervision of Imitation Reward. In addition to the supervision from the simulator, our framework is also guided by the expert driver. + +We compute the target of imitation reward by first calculating the L2 distances $d_{i}$ between the $i$ -th trajectory anchor + +and an expert trajectory. Then, we apply the negative of these distances from $N$ trajectory anchors to the softmax function, yielding: $r_{\mathrm{im},i}^{*} = \frac{\exp(-d_{i})}{\sum_{j=1}^{N}\exp(-d_{j})}$ . Intuitively, a trajectory anchor closer to the expert trajectory than others will receive a higher reward. As a result, the imitation reward $r_{\mathrm{im}}^{*} \in \mathbb{R}^{N}$ quantifies the similarity of each trajectory anchor to the human trajectory. The predicted imitation reward $r_{\mathrm{im}}$ , which is produced by Eq. 5, are supervised using the Cross Entropy loss as + +$$ +\mathcal {L} _ {\text {r e w a r d}} ^ {\mathrm {i m}} = \operatorname {C r o s s E n t r o p y} \left(r _ {\mathrm {i m}}, r _ {\mathrm {i m}} ^ {*}\right). \tag {8} +$$ + +Supervision of Trajectory. For supervising the refined trajectories, we adopt the commonly used winner-take-all strategy [46]. In this strategy, we select the trajectory anchor that is closest to the expert trajectory. Only the corresponding refined trajectory of this anchor will be supervised. Let this trajectory anchor be denoted as $\tau_{i}$ . We use the L1 loss to measure the difference between the expert trajectory $\tau^{*}$ and the refined trajectory $\hat{\tau}_{i}$ as + +$$ +\mathcal {L} _ {\text {t r a j}} = \left| \hat {\tau} _ {i} - \tau^ {*} \right|. \tag {9} +$$ + +Instead of supervising all refined trajectories, we focus solely on the most optimal one. This strategy helps the model capture a diverse range of trajectory patterns, with each anchor specializing in a particular trajectory modality. + +The overall training loss $\mathcal{L}_{\mathrm{total}}$ is as follow: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {B E V}} + \mathcal {L} _ {\text {r e w a r d}} ^ {\text {s i m}} + \mathcal {L} _ {\text {r e w a r d}} ^ {\text {i m}} + \mathcal {L} _ {\text {t r a j}}. \tag {10} +$$ + +# 4. Experiments + +# 4.1. Benchmark + +NAVSIM Dataset The NAVSIM dataset [11] is constructed based on nuPlan [3]. Specifically, OpenScene [8] first downsampled the nuPlan data from $10\mathrm{Hz}$ to $2\mathrm{Hz}$ to condense it into 120 hours of driving logs. NAVSIM resampled the data from OpenScene to emphasize challenging scenarios, reducing simple situations like straight-line driving. The dataset is divided into two parts: Navtrain and Navtest, comprising 1192 scenarios for training and validation and 136 scenarios for testing. Unlike nuScenes [2], which collects data in simpler, slower driving scenarios, NAVSIM ensures that challenging scenarios are prioritized, making simply fitting ego status insufficient for planning. + +NAVSIM Metrics The original end-to-end driving metrics were primarily designed to evaluate the deviation between the predicted trajectories and those driven by human experts. However, this evaluation approach has proven inadequate, resulting in poor performance in real-world driving contexts. NAVSIM introduces more practical and reliable metrics, focusing on aligning open-loop and closed-loop metrics. + +
MethodInput#Traj.Traj. Eval.NC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
Human---10010087.510099.994.8
Constant Velocity-1×69.958.849.349.3100.021.6
Ego Status MLP-1×93.077.362.883.6100.065.6
VADv2 [41]C8192Rule-based97.991.777.692.9100.083.0
UniAD [22]C1Rule-based97.891.978.892.9100.083.4
LTF [31]C1×97.492.879.092.4100.083.8
PARA-Drive [41]C1Rule-based97.992.479.393.099.884.0
TransFuser [31]C & L1×97.792.879.292.8100.084.0
LAW [26]C1×96.495.481.788.799.984.6
DRAMA [43]C & L1×98.093.180.194.8100.085.5
Hydra-MDP [27]C & L8192Model-free98.396.078.794.6100.086.5
WoTEC & L256Model-based98.596.881.994.999.988.3
+ +Table 1. Comparison with the SOTA approaches on NAVSIM test set. Rule-based: UniAD and PARA-Drive use occupancy maps, while VADv2 relies on vectorized maps for trajectory evaluation following specific rules. Model-free: Hydra-MDP evaluates trajectory without world modeling. Model-based: our WoTE evaluates trajectory with the assistance of the BEV world model. NC: no at-fault collision. DAC: drivable area compliance. EP: ego progress. TTC: time-to-collision. Conf.: comfort. PDMS: the predictive driver model score. Traj.: Trajectory. Eval.: Evaluation. C: Camera. L: LiDAR. + +
MethodTraj. Eval.Success Rate↑Driving Score↑
AD-MLP [44]-0.0018.05
UniAD [22]Rule-based16.3645.81
VAD [25]Rule-based15.0042.35
TCP [42]-30.0059.90
WoTEModel-based31.3661.71
+ +Table 2. Comparison with SOTA approaches on closed-loop Bench2Drive [24] benchmark on CARLA simulator. Traj.: Trajectory. Eval.: Evaluation. + +Specifically, NAVSIM evaluates model performance using the Predictive Driver Model Score (PDMS), which is calculated based on five factors: No At-Fault Collision (NC), Drivable Area Compliance (DAC), Time-to-Collision (TTC), Comfort (Comf.), and Ego Progress (EP). The PDMS is calculated as $\mathrm{PDMS} = \mathrm{NC} \times \mathrm{DAC} \times \frac{5 \times \mathrm{EP} + 5 \times \mathrm{TTC} + 2 \times \mathrm{C}}{12}$ . + +Bench2Drive Dataset and Metrics To further evaluate our framework, we conduct closed-loop experiments in the CARLA simulator [13]. Specifically, we adopt the recently proposed Bench2Drive [24] benchmark as our closed-loop evaluation benchmark. Bench2Drive provides a standardized training dataset on CARLA, ensuring fair comparisons across different methods. The benchmark consists of 220 short evaluation routes (around $150\mathrm{m}$ each) distributed across all CARLA towns, with each route containing one safety-critical scenario. This design effectively reduces the evaluation variance. For metric, it utilizes the standard CARLA metric, Driving Score (DS), as the primary metric. Additionally, Bench2Drive reports the success rate, which represents the proportion of successfully completed routes. + +A route is considered successful if and only if DS reaches $100\%$ on this route. + +# 4.2. Implementation Details + +NAVSIM For input data, it is aligned with TransFuser [31]. We concatenate the front-view image with center-cropped front-left and front-right images, resulting in a combined resolution of $256 \times 1024$ pixels. For LiDAR, the $64\mathrm{m} \times 64\mathrm{m}$ point cloud surrounding the ego vehicles is used. For network architecture, we employ ResNet34 as the backbone for BEV feature extraction following TransFuser [31]. The world model consists of two transformer decoder layers. We use $N = 256$ trajectory anchors. The predicted future states are of time $t + 2s, t + 4s$ when the current is of time $t$ . The reward weights are set as follows: $w_{1} = 0.1, w_{2} = w_{3} = 0.5, w_{4} = 1.0$ . The training is conducted on the Navtrain split using 8 NVIDIA L20 GPUs with a total batch size of 256, distributed across 50 epochs. In the ablation study section, we train our model for 20 epochs instead of 50 to accelerate the training process. For fast training, we pre-compute the simulation results (BEV semantic maps and scores) based on the trajectory anchors. We then feed trajectory anchors into the trajectory evaluation module during training. During testing, the trajectory evaluation module evaluates the refined trajectory instead of the anchor trajectories. We utilize the Adam optimizer with a learning rate of 1e-4. + +Bench2Drive For Bench2Drive, we adopt TCP [42] as our baseline and integrate the trajectory evaluation module. We choose TCP as it is the best open-source framework officially + +
Traj. Eval.Future StatesNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
××96.491.576.290.399.981.0
✓×97.192.978.291.9100.083.2
✓✓98.094.779.993.4100.085.6
+ +provided by Bench2Drive. Similar to NAVSIM, we utilize imitation rewards along with simulation-based rewards, including No Collision (NC) and Drivable Area Compliance (DAC). The number of trajectory anchors is set to 256. Notably, rather than employing the winner-take-all strategy (Eq. 9) to supervise only the best trajectory, we find that supervising all trajectories leads to improved performance. The model is trained for 27 epochs with a batch size of 300. We use the Adam optimizer with a learning rate of 1e-4. + +# 4.3. Comparison with SOTA + +NASIM Table 1 compares our method with state-of-the-art approaches on the NAVSIM test set. By leveraging the world model to simulate future scenarios and guide trajectory evaluation, our method achieves a PDMS of 87.1. WoTE outperforms the previous model-free approach, Hydra-MDP, highlighting the advantages of model-based trajectory evaluation. + +Bench2Drive To further validate our approach, we conduct a closed-loop evaluation on the Bench2Drive benchmark within the CARLA simulator. As shown in Table 2, our method improves the Driving Score by 1.81 points, demonstrating its effectiveness in various evaluation settings. + +# 4.4. Ablation Studies + +Trajectory evaluation and future state prediction matter. In this experiment, we investigate the importance of the trajectory evaluation and future state prediction as Table 3 shows. The first row in Table 3 represents our baseline TransFuser [31], which performs trajectory prediction only. In the second row, we change to our framework, which consists of both trajectory prediction and trajectory evaluation. However, we do not used the predicted future states as the input of Reward Model. That is, $\mathbf{B}_{\mathrm{all}} = \mathbf{B}_t$ and $\mathbf{a}_{\mathrm{all}} = \mathbf{a}_t$ in Eq. 5. In the third row, we integrate the BEV world model to predict future states, leading to notable performance improvements across multiple metrics. + +Imitation and simulation rewards are complementary. To investigate how each type of reward affects driving performance, we conduct an ablation study, as shown in Table 4. During inference, we adjust the reward weights in Eq. 4 to isolate their effects: setting $w_{1} = 0$ enables evaluation with only simulation rewards, while setting $w_{2} = w_{3} = w_{4} = 0$ tests performance with only an imitation reward. The results indicate that imitation and simulation rewards emphasize + +Table 3. Ablation study on trajectory evaluation and future state prediction. Traj: Trajectory. Eval.: Evaluation. + +
Imi. RewardSim. RewardsNC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
✓×97.992.577.793.499.983.5
×✓96.594.779.290.199.483.6
✓✓98.094.779.993.4100.085.6
+ +Table 4. Ablation study on the imitation and simulation rewards. The experiment shows that imitation and simulation rewards are complementary. Imi: Imitation. Sim.: Simulation. + +
Future Prediction StepsNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
0s → 4s97.493.679.191.9100.084.0
0s → 2s → 4s98.094.779.993.4100.085.6
+ +Table 5. Ablation study on the different number of time steps in the future prediction. Our BEV world model predicts future states in a recurrent manner. $\mathrm{A} \rightarrow \mathrm{B}$ : the predicted future state at time B is based on the state at time A. + +
#Trajs.NC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑Latency ↓
6497.494.179.691.9100.084.517.2 ms
12897.994.680.093.0100.085.417.9 ms
25698.094.779.993.4100.085.618.7 ms
+ +Table 6. Ablation study on the different number of Trajectories. The number of trajectories is the same during training and testing. The latency is evaluated on an NVIDIA L20 GPU. Trajs.: Trajectories. + +different aspects of planning. Trajectory evaluation using imitation rewards performs better in NC and TTC, while using simulation rewards excels in DAC and EP. By integrating both, the model leverages their respective strengths, achieving improved overall performance across all key metrics. + +Recurrent future state prediction helps. Our framework supports predicting future states recurrently, as detailed in Eq. 3. In Table 5, we present the results obtained by varying the number of BEV states predicted by the world model. Our findings indicate that using a finer time step for future state prediction significantly improves performance. This is because a finer time step provides richer temporal information, which aids trajectory evaluation. + +More trajectories help. This ablation study explores the effect of varying the number of trajectories on overall performance. Specifically, the model in each row is trained and tested under the same number of trajectories, and the world model is fed with refined trajectories. As shown in Table 6, using a small number of trajectories, such as 64, significantly degrades performance. As the number of trajectories increases, performance improves notably, with a marked improvement observed when increasing from 64 to 128 trajectories. However, as the number of trajectories increases from 128 to 256, the performance gains diminish, + +
Train Traj.Test Traj.Traj. Refine†NC ↑DAC ↑EP ↑
256256×96.694.679.0
2561024×97.295.679.3
256256✓98.094.779.9
+ +Table 7. Generalization of the trajectory across different trajectory configurations. + +the input trajectories of BEV world model a trajectory. + +indicating that the performance of the most optimal level. Based on these results, we default number of trajectories for our ex- Latency analysis. A common concern is of our framework. To address this, we run on an NVIDIA L20 GPU, as shown in GPU processes state-action pairs in para remains efficient even as the number of t The total latency is only $18.7\mathrm{ms}$ well $\mathsf{V}$ requirements for end-to-end autonomous + +# Generalization ability across different + +shown in Table 7, we investigate the generalization ability of our trajectory evaluation module. To be specific, we sample $N = 1024$ instead of $N = 256$ trajectory anchors from the dataset. Since our model is trained under the setting of $N = 256$ trajectory anchors, the model has never seen these $N = 1024$ trajectories. Surprisingly, we find a 1.3 PDMS increase when using these unseen trajectory anchors. Besides, the model performs even better when using refined trajectories as inputs. These refined trajectories continuously change based on varying sensor inputs. This shows that the proposed world model has great generalization ability when dealing with unseen trajectories. + +# 4.5. Visual Analysis + +End-to-end planning trajectories. To highlight the improvements in end-to-end driving achieved through trajectory evaluation, we present visualizations of end-to-end planning trajectories in Figure 3. + +Trajectory rewards. Figure 4 illustrates the predicted rewards for different trajectories. + +# 5. Conclusion + +In this paper, we introduced a novel framework that leverages a BEV world model for end-to-end trajectory evaluation in autonomous driving. By integrating a BEV world model, we enable a more informed evaluation process that considers the dynamic evolution of driving scenarios, leading to more effective trajectory evaluation. Our approach benefits from dense supervision provided by BEV-space traffic simulators, which supply both semantic future states and rule-based + +![](images/a81390ca3df63e0c7551fcba665b66c1848c10cd21449e8f9f189b80987ad348.jpg) +Figure 3. Visualization of end-to-end planning trajectories. We compare our method with TransFuser [31], which serves as our trajectory prediction baseline without a trajectory evaluation module. Our trajectory evaluation module effectively filters out low-quality trajectories and helps avoid collisions. + +![](images/3edd753908390f9104df1a78bed1563b0620bf91fbf95d995bd72770970f2b46.jpg) +(a) Multi-modality trajectory preservation + +![](images/e9903960bdf2540bf5e55a7e97e3140fc1b6eb009ec09ab904f0ec0d00d5937d.jpg) +(b) Low-quality trajectory suppression +Figure 4. Visualization of rewards of all trajectories. Brighter colors indicate higher trajectory rewards, while gray trajectories correspond to those with very low rewards. (a): WoTE effectively retains diverse multi-modal trajectories, as long as they comply with traffic rules and do not result in unsafe outcomes. (b): it demonstrates that trajectories considered unreasonable receive significantly lower rewards, highlighting the effectiveness of our trajectory evaluation module. + +reward targets. Our method achieves state-of-the-art performance on the NAVSIM and Bench2Drive benchmarks while maintaining real-time efficiency. Overall, our work establishes trajectory evaluation as an important research direction for end-to-end autonomous driving. We hope our framework serves as a strong baseline and inspires further research in end-to-end online trajectory evaluation. + +# 6. Acknowledgements + +This work was supported in part by the National Science and Technology Major Project (No. 2022ZD0116500), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010). + +# References + +[1] Sébastien MR Arnold, Pierre L'Ecuyer, Liyu Chen, Yi-fan Chen, and Fei Sha. Policy learning and evaluation with randomized quasi-monte carlo. arXiv preprint arXiv:2202.07808, 2022. 2 +[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. CVPR, 2020. 5 +[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 5 +[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 2 +[5] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In CVPR, pages 15590-15599, 2021. 1, 2 +[6] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1 +[7] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3 +[8] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving, 2023. 5 +[9] Christoph Dann, Gerhard Neumann, and Jan Peters. Policy evaluation with temporal differences: A survey and comparison. The Journal of Machine Learning Research, 15(1): 809-883, 2014. 2 +[10] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1 +[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. arXiv preprint arXiv:2406.15349, 2024. 2, 4, 5 +[12] Christopher Diehl, Timo Sievernich, Martin Krüger, Frank Hoffmann, and Torsten Bertram. Umbrella: Uncertainty-aware model-based offline reinforcement learning leveraging planning. arXiv preprint arXiv:2111.11097, 2021. 2 + +[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 2, 6 +[14] Haoyang Fan, Fan Zhu, Changchun Liu, Liangliang Zhang, Li Zhuang, Dong Li, Weicheng Zhu, Jiangtao Hu, Hongye Li, and Qi Kong. Baidu Apollo em motion planner. arXiv preprint arXiv:1807.08048, 2018. 1 +[15] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 1 +[16] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems, 36, 2024. 2 +[17] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018. 1 +[18] Mikael Henaff, Alfredo Canziani, and Yann LeCun. Model-predictive policy learning with uncertainty regularization for driving in dense traffic. In International Conference on Learning Representations, 2019. 1, 2 +[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 1, 2 +[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1 +[21] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 1, 2 +[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 2, 6 +[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 1 +[24] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 2, 6 +[25] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6 +[26] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end + +autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 6 +[27] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 3, 4, 6 +[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dolkar. Focal loss for dense object detection. In ICCV, 2017. 5 +[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 1 +[30] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, pages 7077-7087, 2021. 1 +[31] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 3, 6, 7, 8 +[32] Guanren Qiao, Guorui Quan, Rongxiao Qu, and Guiliang Liu. Modelling competitive behaviors in autonomous driving under generative world model. In European Conference on Computer Vision, pages 19-36. Springer, 2024. 1, 2 +[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[34] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In ECCV, 2020. 2 +[35] Ardi Tampuu, Tambet Matiisen, Maksym Semikin, Dmytro Fishman, and Naveed Muhammad. A survey of end-to-end driving: Architectures and training methods. IEEE Transactions on Neural Networks and Learning Systems, 33(4): 1364-1384, 2020. 1 +[36] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2 +[37] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1 +[38] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2 +[39] Yuqi Wang, Ke Cheng, Jiawei He, Qitai Wang, Hengchen Dai, Yuntao Chen, Fei Xia, and Zhaoxiang Zhang. Drivingndojo dataset: Advancing interactive and knowledge-enriched driving world model. arXiv preprint arXiv:2410.10738, 2024. +[40] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Con- + +ference on Computer Vision and Pattern Recognition, pages 14749-14759, 2024. 1, 2, 3 +[41] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 6 +[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: a simple yet strong baseline. NeurIPS, 2022. 6 +[43] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 2, 6 +[44] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 6 +[45] Yumeng Zhang, Shi Gong, Kaixin Xiong, Xiaoqing Ye, Xiao Tan, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Bevworld: A multimodal world model for autonomous driving via unified bev latent space. arXiv preprint arXiv:2407.05679, 2024. 2 +[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 5 \ No newline at end of file diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/images.zip b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7a1129d3557262a384bebf8e72cb78819b8cc73a --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8f4b74834b68edb28793f249f849146172457c3a4b215f4b463180d902e785e6 +size 435210 diff --git a/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/layout.json b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0ea0f54b49cabd6f1dcab511eb7e6874a4178bf2 --- /dev/null +++ b/2025/End-to-End Driving with Online Trajectory Evaluation via BEV World Model/layout.json @@ -0,0 +1,8487 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 120 + ], + "type": "text", + "content": "End-to-End Driving with Online Trajectory Evaluation via BEV World Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 143, + 517, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 143, + 517, + 200 + ], + "spans": [ + { + "bbox": [ + 93, + 143, + 517, + 200 + ], + "type": "text", + "content": "Yingyan Li* Yuqi Wang* Yang Liu Jiawei He Lue Fan* Zhaoxiang Zhang* NLPR, Institute of Automation, Chinese Academy of Sciences (CASIA) State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS) University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 251, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 479 + ], + "type": "text", + "content": "End-to-end autonomous driving has achieved remarkable progress by integrating perception, prediction, and planning into a fully differentiable framework. Yet, to fully realize its potential, an effective online trajectory evaluation is in dispensable to ensure safety. By forecasting the future outcomes of a given trajectory, trajectory evaluation becomes much more effective. This goal can be achieved by employing a world model to capture environmental dynamic and predict future states. Therefore, we propose an end-to-end driving framework WoTE, which leverages a BE World model to predict future BEV states for Trajectory Evaluation. The proposed BEV world model is latency efficient compared to image-level world models and can be seamlessly supervised using off-the-shelf BEV-space traffic simulators. We validate our framework on both the NAVSIM benchmark and the closed-loop Bench2Drive benchmark based on the CARLA simulator, achieving state-of-the-art performance. Code is released at https://github.com/liyingyanUCAS/WoTE." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 295, + 644 + ], + "type": "text", + "content": "End-to-end autonomous driving [6, 30, 35] has achieved remarkable success in recent years and garnered widespread attention. While most approaches [21, 23, 25] primarily focus on predicting a high-quality trajectory, recent works [7, 27] suggest that predicting multiple trajectories simultaneously better reflects the multi-modal nature of driving and leads to improved performance. Given these multiple possible trajectories, it becomes crucial to effectively evaluate them in order to ensure the safety and reliability of end-to-end autonomous driving systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 296, + 693 + ], + "type": "text", + "content": "Traditional trajectory evaluation methods are rule-based [10, 14, 20, 37] and rely on perception results, such as bounding boxes and maps, to assess trajectories. These methods are sensitive to perception inaccuracies and may" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 228, + 552, + 279 + ], + "blocks": [ + { + "bbox": [ + 318, + 228, + 552, + 279 + ], + "lines": [ + { + "bbox": [ + 318, + 228, + 552, + 279 + ], + "spans": [ + { + "bbox": [ + 318, + 228, + 552, + 279 + ], + "type": "image", + "image_path": "0e3c1fc45288ff454935d5c641bb01389b75cbc08501912a7f1adb1ddf7dd0df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 280, + 552, + 365 + ], + "blocks": [ + { + "bbox": [ + 318, + 280, + 552, + 365 + ], + "lines": [ + { + "bbox": [ + 318, + 280, + 552, + 365 + ], + "spans": [ + { + "bbox": [ + 318, + 280, + 552, + 365 + ], + "type": "image", + "image_path": "d73d50b4b000a502be8c230e7e71a5b7e1b34d3f498c0b249cdf8fbc45eb9fbb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 376, + 555, + 475 + ], + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 475 + ], + "type": "text", + "content": "Figure 1. The illustration of end-to-end trajectory evaluation with BEV world model. (a) Previous end-to-end driving methods primarily focus on learning a high-quality trajectory. (b) Our WoTE consists of end-to-end trajectory prediction and evaluation. First, we predict multi-modal trajectories. Next, given these trajectories, we leverage the BEV world model to predict the corresponding future BEV states. The reward model evaluates these trajectories based on the predicted states and assigns rewards, selecting the highest-reward trajectory." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 605 + ], + "type": "text", + "content": "not be directly optimized in an end-to-end manner. Recently, some end-to-end driving approaches [7, 27] have incorporated trajectory evaluation. Their evaluations rely on the current state. However, evaluating a trajectory candidate is particularly challenging without knowledge of the future led by this trajectory. Similar to human drivers, who anticipate futures before making decisions, a trajectory evaluation network would benefit from predicting the corresponding future states of trajectory candidates." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 714 + ], + "type": "text", + "content": "To effectively predict future states, it is intuitive to adopt methodologies from reinforcement learning, where world models have successfully facilitated future state predictions. In reinforcement learning, model-based methods [5, 18, 32] have consistently shown superior performance compared to model-free approaches [17, 29], largely due to their ability of modeling the environmental dynamics and reasoning about future scenarios through world models. Inspired by this success, we propose adopting a world model [15, 19, 38-40]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "inline_equation", + "content": "^{1}\\ast" + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "text", + "content": ": Co-first author, " + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 66, + 702, + 284, + 713 + ], + "type": "text", + "content": ": Co-advisor. Email: liyingyan2021@ia.ac.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27137" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "within end-to-end autonomous driving frameworks to facilitate accurate prediction of future states, thus significantly improving trajectory evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 112, + 297, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 112, + 297, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 112, + 297, + 244 + ], + "type": "text", + "content": "However, leveraging the world model for trajectory evaluation presents some technical issues. First, a good representation of future scenarios is needed. Many existing world models for driving [19, 38, 40] predict the images of future scenarios with diffusion models [33], which can be time-consuming and not suitable for real-time driving. Second, we lack supervision for future states. Supervising the predicted future states of the world model is difficult because world models need to imagine multiple future states based on the multiple trajectory candidates, while only one future state is available in the real-world datasets." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 248, + 297, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 248, + 297, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 297, + 475 + ], + "type": "text", + "content": "To address these technical issues, we propose WoTE as shown in Fig.1. We propose to predict future states in the BEV space. The information in BEV space is much more compact than in raw images, which enables a more efficient single-step feed-forward future prediction than the time-consuming multi-step denoising in previous driving world models [19, 38, 40, 45]. This characteristic of BEV space makes WoTE highly suitable for real-time driving applications. Additionally, the predicted future BEV states can be easily supervised using off-the-shelf BEV-space traffic simulators such as nuPlan [3, 16]. These simulators model the scenario in BEV space by representing the map and objects using BEV semantic maps, creating a dynamic simulation of the driving environment. Consequently, we can use the BEV semantic maps from these simulated scenarios as supervision for the prediction of the BEV world model. Moreover, these simulators are equipped with well-established evaluation rules to assess the simulated future BEV scenarios, which provide target rewards." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 297, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 297, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 297, + 576 + ], + "type": "text", + "content": "We validate our framework on both NAVSIM [11] benchmark and closed-loop Bench2Drive [24] benchmark based on the CARLA [13] simulator and achieve state-of-the-art performance. Our experiments show that incorporating future states into trajectory evaluation is essential for improved performance. Additionally, the trajectory evaluation module, built on the world model shows great generalization ability across diverse trajectories. Our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 579, + 295, + 710 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 579, + 295, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 613 + ], + "type": "text", + "content": "- We highlight the importance of utilizing imagined future states for trajectory evaluation in end-to-end autonomous driving." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 616, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 616, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 56, + 616, + 295, + 673 + ], + "type": "text", + "content": "- We introduce WoTE, which consists of an end-to-end trajectory evaluation module based on a BEV world model. The world model in BEV space enables efficient real-time future prediction and overcomes the scarcity of multi-future supervisions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 675, + 294, + 710 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 675, + 294, + 710 + ], + "spans": [ + { + "bbox": [ + 56, + 675, + 294, + 710 + ], + "type": "text", + "content": "- WoTE is validated on both the NAVSIM [11] benchmark and the closed-loop CARLA-based Bench2Drive [24] benchmark, achieving state-of-the-art performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 91, + 491, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 491, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 491, + 105 + ], + "type": "text", + "content": "2.1. End-to-end Autonomous Driving" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 110, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 410 + ], + "type": "text", + "content": "End-to-end autonomous driving [7, 22, 25, 43] is to train a model that directly maps sensor inputs to trajectories or control signals instead of decomposing the driving task into traditional sub-tasks. These works can be divided into two categories according to the training paradigm: imitation-learning-based and reinforcement-learning-based. Imitation learning is the most common practice in end-to-end autonomous driving. Their predicted trajectories are supervised by the expert trajectories. For instance, P3 [34], takes a differentiable semantic occupancy as the intermediate representation and learns actions from human trajectories. Following this, ST-P3 [21] learns spatial-temporal features for perception, prediction and planning tasks simultaneously. UniAD [22] proposed goal-oriented planning, combining and leveraging advantages of perception and motion prediction modules. VAD [25] introduces vectorized representations for end-to-end planning. There are also some reinforcement-learning-based methods. For example, MaRLn [36] utilizes implicit affordances to develop a reinforcement learning algorithm. LBC [4] trained a vision-based sensorimotor agent from the privileged agent. Combining imitation learning and reinforcement learning, Hydra-MDP [27] utilizes knowledge distillation to obtain supervisions from both rule-based planners and human drivers. It is a promising research direction as more supervisions are utilized." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 419, + 441, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 419, + 441, + 433 + ], + "spans": [ + { + "bbox": [ + 314, + 419, + 441, + 433 + ], + "type": "text", + "content": "2.2. Trajectory Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 556, + 715 + ], + "type": "text", + "content": "Trajectory evaluation is essential for ensuring the reliability and safety of autonomous driving systems. We divide the main trajectory evaluation methods into two categories: model-free and model-based. Model-free methods directly assess the learned policy using observed data without explicitly modeling environmental dynamics. For instance, Monte Carlo policy evaluation [1] uses the empirical mean return as the metric. The value function approximates the expectation based on sampled returns. Temporal-difference learning [9] uses bootstrapping, where the current estimate of the value function is used to generate target values for updating the value function. Subsequently, model-based trajectory evaluation methods [12] have gained increasing popularity in recent years. By leveraging learned world models that capture environmental dynamics, these methods enable more accurate and predictive policy evaluation. For example, PPUU [18] measures policy performance by quantifying uncertainty in predicted future states. Rails [5] evaluates trajectories through a tabular dynamic programming mechanism, and MARL-CCE [32] employs learned environmental dynamics to assess multi-agent policies in driving scenarios. However, these methods often rely on explicit trajectory representations and non-differentiable metrics, which limits" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27138" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "their end-to-end optimization capabilities. In contrast, our approach evaluates trajectories using BEV features and encoded trajectory embeddings, enabling fully differentiable and thus end-to-end trajectory evaluation." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 111, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 111, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 111, + 144 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 319 + ], + "type": "text", + "content": "In this section, we introduce four key components of WoTE: Trajectory Predictor (Sec. 3.1), BEV World Model (Sec. 3.2), Reward Model (Sec. 3.3) and BEV Space Supervision (Sec. 3.4). As illustrated in Figure 2, our method integrates future state prediction via the BEV World Model and reward prediction via the Reward Model within a unified framework. Specifically, the Trajectory Predictor encodes the multi-modal observations into BEV state, represented as a BEV feature map, and provides multiple trajectory proposals. The BEV World Model predicts corresponding future BEV states based on these trajectory proposals. With the help of future states, the Reward Model predicts the reward of each trajectory. Finally, we choose the trajectory with the highest reward as the final trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 175, + 339 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 175, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 175, + 339 + ], + "type": "text", + "content": "3.1. Trajectory Predictor" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 344, + 297, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 297, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 297, + 417 + ], + "type": "text", + "content": "As shown in Figure 2 (a), the Trajectory Predictor aims to predict multiple trajectory candidates. First, the BEV encoder encodes the multi-modal sensor inputs into a unified BEV feature map. Then, given trajectory anchors, the trajectory refinement module refines trajectories based on the BEV feature map. We will now describe each part in detail." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "content": "Multi-modal BEV Encoder. Following TransFuser [31], our method leverages multi-modal sensor inputs, including LiDAR and multi-view RGB images. We use a BEV encoder [31] to transform the current multi-modal inputs into a unified BEV feature map " + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{B} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "content": ". For convenience in the following text, we refer to the BEV feature map as the BEV state. " + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "content": " represent the height and width of the BEV state, while " + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 431, + 296, + 526 + ], + "type": "text", + "content": " denotes the number of channels." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": "Trajectory Anchors. Following methods in [7, 27], we generate trajectory anchors " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": " by K-Means clustering. This clustering is performed on all the expert trajectories collected from the training dataset. The number of trajectory anchors is denoted as " + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 542, + 296, + 601 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": "Trajectory Refinement. As illustrated in Figure 2 (a), the trajectory refinement module takes the trajectory anchors and the current BEV states as inputs to predict the refined trajectories. Specifically, we first use a trajectory encoder TE, which is implemented by an MLP, to encode the trajectory anchors " + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 617, + 297, + 713 + ], + "type": "text", + "content": " into feature vectors. Then, these feature vectors are used as queries to perform cross attention with the current BEV state B, which serves as both the key and" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "value. The output of the cross-attention is then fed into an MLP head to predict the offsets. The offset is added to the trajectory anchors to produce the refined trajectories. The entire process can be expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 337, + 129, + 555, + 141 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 129, + 555, + 141 + ], + "spans": [ + { + "bbox": [ + 337, + 129, + 555, + 141 + ], + "type": "interline_equation", + "content": "\\hat {\\tau} = \\tau + \\operatorname {M L P} (\\text {C r o s s A t t e n t i o n} (\\mathbf {T E} (\\tau), \\mathbf {B}, \\mathbf {B})), \\tag {1}", + "image_path": "b4cbd52387f5a5589b1eefaea8c5e2170d3fe132d234760840290c7b6ce5128a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 148, + 453, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 453, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 453, + 161 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 148, + 453, + 161 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 313, + 148, + 453, + 161 + ], + "type": "text", + "content": " is the refined trajectories." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 167, + 424, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 424, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 424, + 179 + ], + "type": "text", + "content": "3.2. BEV World Model" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 185, + 554, + 257 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 185, + 554, + 257 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 554, + 257 + ], + "type": "text", + "content": "In this section, we use the BEV World Model to predict the future BEV states, as illustrated in Figure 2 (b). The process is divided into two steps. First, the current BEV state and refined trajectories are combined to form the inputs for the world model. Next, the world model recurrently predicts future states over multiple time steps." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": "Input of World Model. Given " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": " refined trajectories " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": " and the current BEV state " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": ", we construct " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": " state-action pairs as the inputs of world model. We use the trajectory encoder TE in Eq. 1, with shared parameters, to encode " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": " into action embeddings " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_t = \\{\\mathbf{a}_t^1, \\mathbf{a}_t^2, \\dots, \\mathbf{a}_t^N\\}" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": ", each state-action pair is represented as " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_t, \\mathbf{a}_t^i)" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "inline_equation", + "content": "i \\in \\{1, \\dots, N\\}" + }, + { + "bbox": [ + 313, + 270, + 555, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": "Recurrent Future Prediction. Given the state-action pairs, we use a world model to predict their corresponding future states. In contrast to prior work [40], which predict the future in image space, we build a world model in BEV space as it is much more efficient. In detail, given a state-action pair " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "(\\mathbf{B}_t,\\mathbf{a}_t^i)" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ", we flatten " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_t" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "hw" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " feature vectors. The dimension of each vector is " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ". Then we concatenate these vectors with the action embedding " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_t^i" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ", resulting in a total of " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "hw + 1" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " feature vectors, namely " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i\\in \\mathbb{R}^{(hw + 1)\\times c}" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_i" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " are then fed into the world model, whose network architecture is a transformer encoder, as " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i + 1} = \\mathrm{TransformerEncoder}(\\mathbf{F}_i)" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ". The output of the world model consists of " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "h\\times w + 1" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " feature vectors, corresponding to the predicted future states " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t + 1}^i" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": " and future actions " + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t + 1}^i" + }, + { + "bbox": [ + 313, + 369, + 556, + 549 + ], + "type": "text", + "content": ". In summary, we can formulate this process in the following equation:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 356, + 555, + 555, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 555, + 555, + 571 + ], + "spans": [ + { + "bbox": [ + 356, + 555, + 555, + 571 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {B} _ {t + 1} ^ {i}, \\mathbf {a} _ {t + 1} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t}, \\mathbf {a} _ {t} ^ {i}\\right), \\tag {2}", + "image_path": "19b0af36a82ffbaa44b5049d282f581e112d2103472f10e4edb79a89f1662f0a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "text", + "content": "Furthermore, this world model is able to predict the states " + }, + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{t + K}^{i}" + }, + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "text", + "content": " of the future " + }, + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 577, + 554, + 613 + ], + "type": "text", + "content": " steps in a recurrent manner, as shown by the following equation:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 620, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 620, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 324, + 620, + 555, + 635 + ], + "type": "interline_equation", + "content": "\\left(\\mathbf {B} _ {t + K} ^ {i}, \\mathbf {a} _ {t + K} ^ {i}\\right) = \\operatorname {W o r l d M o d e l} \\left(\\mathbf {B} _ {t + K - 1} ^ {i}, \\mathbf {a} _ {t + K - 1} ^ {i}\\right). \\tag {3}", + "image_path": "da9ca674eaa5ed68d8913eba27faa1cd4ce72042e837b379f6acf1da511b1a9d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "The process above is conducted for every state-action pair in a parallel manner. It is important to note that the size of BEV states is relatively small (e.g., " + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "inline_equation", + "content": "h = w = 8" + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "), so the world model does not introduce significant computational overhead. Next, we introduce the Reward Model, which predicts rewards for a trajectory based on its future states." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27139" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 74, + 550, + 324 + ], + "blocks": [ + { + "bbox": [ + 59, + 74, + 550, + 324 + ], + "lines": [ + { + "bbox": [ + 59, + 74, + 550, + 324 + ], + "spans": [ + { + "bbox": [ + 59, + 74, + 550, + 324 + ], + "type": "image", + "image_path": "b7d10dd316364701d4253cc9d914004d67f33f6eb35bbeaa4c4ce313239d36eb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 337, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 337, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. Overall structure of WoTE. Our model predicts and evaluates trajectory in an end-to-end manner. It is mainly divided into two parts. (a) Trajectory Prediction: the BEV encoder encodes the multi-modal observations into a BEV feature map (state) and proposes multiple trajectory candidates. (b) Trajectory Evaluation: Given the current BEV state and trajectory candidates, we leverage the BEV World Model to predict the corresponding future BEV states of these trajectories. The Reward Model then predicts rewards with the help of future BEV states. The trajectory with the highest reward is selected as the final trajectory." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 413, + 149, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 149, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 149, + 425 + ], + "type": "text", + "content": "3.3. Reward Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": "In this section, we first present the reward types and then describe how they are learned." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": "Reward Types. Following [27], we have two types of rewards: the imitation reward and the simulation reward. The imitation reward " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{im}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": " measures how well the predicted trajectory mimics an expert trajectory. The detailed implementation will be introduced in Sec. 3.4. The simulation rewards " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{sim}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": " measures trajectory quality based on simulator-defined criteria. Following [11], we assess trajectories using the following five criteria: no collisions (NC), drivable area compliance (DAC), time-to-collision (TTC), comfort (Comf), and ego progress (EP). As a result, we have " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{sim}} = \\{r_{\\mathrm{sim}}^{\\mathrm{NC}}, r_{\\mathrm{sim}}^{\\mathrm{DAC}}, r_{\\mathrm{sim}}^{\\mathrm{TTC}}, r_{\\mathrm{sim}}^{\\mathrm{Comf}}, r_{\\mathrm{sim}}^{\\mathrm{EP}}\\}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": ". Given " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{im}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{sim}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": ", the final reward " + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{final}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 613 + ], + "type": "text", + "content": " is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 616, + 295, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 616, + 295, + 659 + ], + "spans": [ + { + "bbox": [ + 67, + 616, + 295, + 659 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} r _ {\\text {f i n a l}} = - \\left(w _ {1} \\log r _ {\\mathrm {i m}} + w _ {2} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {N C}} + w _ {3} \\log r _ {\\mathrm {s i m}} ^ {\\mathrm {D A C}} \\right. \\tag {4} \\\\ \\left. + w _ {4} \\log \\left(5 r _ {\\mathrm {s i m}} ^ {\\mathrm {T T C}} + 2 r _ {\\mathrm {s i m}} ^ {\\mathrm {C o m f}} + 5 r _ {\\mathrm {s i m}} ^ {\\mathrm {E P}}\\right)\\right), \\\\ \\end{array}", + "image_path": "805d4bc2519824310d8a7b7ff96c36de3c2d9d3c71b82b8be5ab1de5cab752ca.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 664, + 180, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 664, + 180, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 664, + 180, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 664, + 180, + 676 + ], + "type": "inline_equation", + "content": "w_{i}" + }, + { + "bbox": [ + 55, + 664, + 180, + 676 + ], + "type": "text", + "content": " are hyper-parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Reward Prediction. Our method needs these rewards for trajectory evaluation in the inference time, so we propose" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": "reward prediction. Take the " + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": "-th refined trajectory " + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_i" + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": " for illustration, the previous approaches [27] typically predict a reward based solely on the current state. With the help of the world model, our reward model predicts the reward " + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i" + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": "-th trajectory based on both the current BEV state and future " + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 414, + 555, + 498 + ], + "type": "text", + "content": " steps BEV states. This process is illustrated in the following equation:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 371, + 508, + 555, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 508, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 371, + 508, + 555, + 523 + ], + "type": "interline_equation", + "content": "\\mathbf {r} _ {i} = \\operatorname {R e w a r d M o d e l} \\left(\\mathbf {B} _ {\\text {a l l}} ^ {i}, \\mathbf {a} _ {\\text {a l l}} ^ {i}\\right), \\tag {5}", + "image_path": "d091e953f9e713df309cec8307bb3659d9891fc599d8ff00109b56e1788960d4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{all}}^i = \\mathrm{cat}([\\mathbf{B}_t,\\mathbf{B}_{t + 1}^i,\\dots,\\mathbf{B}_{t + K}^i])" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathrm{all}}^i = \\mathrm{MLP}(\\mathrm{cat}([\\mathbf{a}_t^i,\\mathbf{a}_{t + 1}^i,\\dots,\\mathbf{a}_{t + K}^i]))" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{cat}([\\cdot])" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": " represents the concatenate operation along the channel dimension. In detail, Reward Model consists of three parts. It first applies 2D convolutions to " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{all}}^i" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": " to aggregate information from different regions and time steps of the BEV states. Next, a global average pooling is applied to form " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{pool}}\\in \\mathbb{R}^{c}" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": " is the channel dimension. We then fed " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^i = \\mathrm{cat}([\\mathbf{B}_{\\mathrm{pool}},\\mathbf{a}_{\\mathrm{all}}^i])" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{S}^i\\in \\mathbb{R}^{2c}" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": ", into an MLP head to predict the reward " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_i\\in \\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 532, + 556, + 676 + ], + "type": "text", + "content": " denotes the number of reward. Next, we introduction the definition of these rewards and select a final trajectory based on these rewards." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "The refined trajectory with the highest final reward is selected as the final trajectory. Next, we introduce the supervision of these predicted rewards and BEV states." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27140" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 201, + 85 + ], + "type": "text", + "content": "3.4. Supervision in BEV Space" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 233 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 233 + ], + "type": "text", + "content": "The world model predicts multiple future feature states corresponding to different trajectories. However, supervising these predictions in the image space is challenging, as driving logs provide only a single future trajectory. To address this, we propose supervising in the BEV space. Supervising in BEV space alleviates the difficulty of accurately modeling future states in image space while also improving computational efficiency. Many real-world traffic simulators support traffic simulations in the BEV space. They generate realistic and reliable future BEV scenarios along with rewards. We adopt nuPlan [3] as our traffic simulator for its broad adoption and realistic scenario simulations." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": "Supervision of BEV States. For BEV state supervision, we decode a semantic BEV map based on the BEV state to enable explicit supervision as Figure 2 (b) shows. We utilize upsampling and transposed convolution layers to decode a BEV semantic map from the BEV state. The dimension of the BEV semantic map is " + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "H \\times W \\times L" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": " denotes the number of semantic classes. This BEV semantic map provides a comprehensive representation of a BEV scenario, covering classes including background, road, walkways, centerline, static objects, vehicles and pedestrians. Each element is represented as a one-hot encoded channel in " + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": ". For each time step " + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "t + k" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": ", We use Focal Loss [28] to supervise the predicted BEV semantic map " + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{t + k}" + }, + { + "bbox": [ + 55, + 247, + 296, + 403 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 108, + 413, + 296, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 413, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 108, + 413, + 296, + 426 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {B E V}} = \\operatorname {F o c a l L o s s} \\left(\\mathbb {B} _ {t + k}, \\mathbb {B} _ {t + k} ^ {*}\\right), \\tag {6}", + "image_path": "d59b2d651a9228701aba34ce062b7cfdebb5d1d13164c5dff3a1cc0dbc180144.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "inline_equation", + "content": "\\mathbb{B}_{t + k}^{*}" + }, + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "text", + "content": " is the ground truth simulated BEV semantic map at time " + }, + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "inline_equation", + "content": "t + k" + }, + { + "bbox": [ + 55, + 435, + 295, + 460 + ], + "type": "text", + "content": " provided by the simulator." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "content": "Supervision of Simulation Rewards. Regarding the simulation rewards, we leverage the simulator to produce five rewards for evaluating a trajectory. Specifically, given the " + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "content": "-th trajectory " + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}^i" + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "content": ", the simulator simulates the future location of the other agents and ego vehicle and produces a simulated future BEV semantic map with agents. Next, it uses a rule-based evaluator to evaluate this simulated future BEV scenarios for producing the corresponding target simulation rewards of " + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}^i" + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "content": ". Given the rewards provided by the simulator, we use the Binary Cross Entropy (BCE) loss to supervise the predicted simulation reward " + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{sim}}" + }, + { + "bbox": [ + 55, + 473, + 297, + 616 + ], + "type": "text", + "content": ". As a result, the loss is formulated as" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 122, + 626, + 296, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 122, + 626, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 122, + 626, + 296, + 641 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} = \\mathrm {B C E} \\left(r _ {\\text {s i m}}, r _ {\\text {s i m}} ^ {*}\\right). \\tag {7}", + "image_path": "fdf02714e2cfd7977975887b18822de0b71af8e1c2fc8687c899d07da517b7cd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 689 + ], + "type": "text", + "content": "Supervision of Imitation Reward. In addition to the supervision from the simulator, our framework is also guided by the expert driver." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "We compute the target of imitation reward by first calculating the L2 distances " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " between the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "-th trajectory anchor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "content": "and an expert trajectory. Then, we apply the negative of these distances from " + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "content": " trajectory anchors to the softmax function, yielding: " + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{im},i}^{*} = \\frac{\\exp(-d_{i})}{\\sum_{j=1}^{N}\\exp(-d_{j})}" + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "content": ". Intuitively, a trajectory anchor closer to the expert trajectory than others will receive a higher reward. As a result, the imitation reward " + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{im}}^{*} \\in \\mathbb{R}^{N}" + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "content": " quantifies the similarity of each trajectory anchor to the human trajectory. The predicted imitation reward " + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{im}}" + }, + { + "bbox": [ + 313, + 72, + 556, + 185 + ], + "type": "text", + "content": ", which is produced by Eq. 5, are supervised using the Cross Entropy loss as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 365, + 193, + 555, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 193, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 365, + 193, + 555, + 208 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e w a r d}} ^ {\\mathrm {i m}} = \\operatorname {C r o s s E n t r o p y} \\left(r _ {\\mathrm {i m}}, r _ {\\mathrm {i m}} ^ {*}\\right). \\tag {8}", + "image_path": "28d0fba002d68606420c204b4e8366323fa29e5d008e79a1f0e8ff683ca0ab4a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "text", + "content": "Supervision of Trajectory. For supervising the refined trajectories, we adopt the commonly used winner-take-all strategy [46]. In this strategy, we select the trajectory anchor that is closest to the expert trajectory. Only the corresponding refined trajectory of this anchor will be supervised. Let this trajectory anchor be denoted as " + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "inline_equation", + "content": "\\tau_{i}" + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "text", + "content": ". We use the L1 loss to measure the difference between the expert trajectory " + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "inline_equation", + "content": "\\tau^{*}" + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "text", + "content": " and the refined trajectory " + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "inline_equation", + "content": "\\hat{\\tau}_{i}" + }, + { + "bbox": [ + 313, + 220, + 556, + 317 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 326, + 555, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 326, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 399, + 326, + 555, + 338 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t r a j}} = \\left| \\hat {\\tau} _ {i} - \\tau^ {*} \\right|. \\tag {9}", + "image_path": "521d7407a39611854b54249156c683f23169c8dfaa3b24c7caf265d57e21245e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 396 + ], + "type": "text", + "content": "Instead of supervising all refined trajectories, we focus solely on the most optimal one. This strategy helps the model capture a diverse range of trajectory patterns, with each anchor specializing in a particular trajectory modality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 396, + 499, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 396, + 499, + 407 + ], + "spans": [ + { + "bbox": [ + 326, + 396, + 499, + 407 + ], + "type": "text", + "content": "The overall training loss " + }, + { + "bbox": [ + 326, + 396, + 499, + 407 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{total}}" + }, + { + "bbox": [ + 326, + 396, + 499, + 407 + ], + "type": "text", + "content": " is as follow:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 351, + 416, + 555, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 416, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 351, + 416, + 555, + 430 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {B E V}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {s i m}} + \\mathcal {L} _ {\\text {r e w a r d}} ^ {\\text {i m}} + \\mathcal {L} _ {\\text {t r a j}}. \\tag {10}", + "image_path": "a6b96f239a4f7c5f424d14ab2f1d67c6dc72b5ba29c699436ab8773677860535.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 449, + 394, + 463 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 394, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 394, + 463 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 468, + 392, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 392, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 392, + 480 + ], + "type": "text", + "content": "4.1. Benchmark" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "content": "NAVSIM Dataset The NAVSIM dataset [11] is constructed based on nuPlan [3]. Specifically, OpenScene [8] first downsampled the nuPlan data from " + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "inline_equation", + "content": "10\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "inline_equation", + "content": "2\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "content": " to condense it into 120 hours of driving logs. NAVSIM resampled the data from OpenScene to emphasize challenging scenarios, reducing simple situations like straight-line driving. The dataset is divided into two parts: Navtrain and Navtest, comprising 1192 scenarios for training and validation and 136 scenarios for testing. Unlike nuScenes [2], which collects data in simpler, slower driving scenarios, NAVSIM ensures that challenging scenarios are prioritized, making simply fitting ego status insufficient for planning." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 557, + 714 + ], + "type": "text", + "content": "NAVSIM Metrics The original end-to-end driving metrics were primarily designed to evaluate the deviation between the predicted trajectories and those driven by human experts. However, this evaluation approach has proven inadequate, resulting in poor performance in real-world driving contexts. NAVSIM introduces more practical and reliable metrics, focusing on aligning open-loop and closed-loop metrics." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27141" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 541, + 251 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 541, + 251 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 541, + 251 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 541, + 251 + ], + "type": "table", + "html": "
MethodInput#Traj.Traj. Eval.NC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
Human---10010087.510099.994.8
Constant Velocity-1×69.958.849.349.3100.021.6
Ego Status MLP-1×93.077.362.883.6100.065.6
VADv2 [41]C8192Rule-based97.991.777.692.9100.083.0
UniAD [22]C1Rule-based97.891.978.892.9100.083.4
LTF [31]C1×97.492.879.092.4100.083.8
PARA-Drive [41]C1Rule-based97.992.479.393.099.884.0
TransFuser [31]C & L1×97.792.879.292.8100.084.0
LAW [26]C1×96.495.481.788.799.984.6
DRAMA [43]C & L1×98.093.180.194.8100.085.5
Hydra-MDP [27]C & L8192Model-free98.396.078.794.6100.086.5
WoTEC & L256Model-based98.596.881.994.999.988.3
", + "image_path": "a0bf8737b09b4588402d615262edc5bc6a9945e761d56b9d781b99c028f325b3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 342, + 296, + 415 + ], + "blocks": [ + { + "bbox": [ + 55, + 258, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 55, + 258, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 555, + 314 + ], + "type": "text", + "content": "Table 1. Comparison with the SOTA approaches on NAVSIM test set. Rule-based: UniAD and PARA-Drive use occupancy maps, while VADv2 relies on vectorized maps for trajectory evaluation following specific rules. Model-free: Hydra-MDP evaluates trajectory without world modeling. Model-based: our WoTE evaluates trajectory with the assistance of the BEV world model. NC: no at-fault collision. DAC: drivable area compliance. EP: ego progress. TTC: time-to-collision. Conf.: comfort. PDMS: the predictive driver model score. Traj.: Trajectory. Eval.: Evaluation. C: Camera. L: LiDAR." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 342, + 296, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 342, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 296, + 415 + ], + "type": "table", + "html": "
MethodTraj. Eval.Success Rate↑Driving Score↑
AD-MLP [44]-0.0018.05
UniAD [22]Rule-based16.3645.81
VAD [25]Rule-based15.0042.35
TCP [42]-30.0059.90
WoTEModel-based31.3661.71
", + "image_path": "e7455fc63dbfcff680297a389d3c8eaf49119b9721173f37838a247f62fbf342.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 424, + 296, + 458 + ], + "lines": [ + { + "bbox": [ + 55, + 424, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 424, + 296, + 458 + ], + "type": "text", + "content": "Table 2. Comparison with SOTA approaches on closed-loop Bench2Drive [24] benchmark on CARLA simulator. Traj.: Trajectory. Eval.: Evaluation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 545 + ], + "type": "text", + "content": "Specifically, NAVSIM evaluates model performance using the Predictive Driver Model Score (PDMS), which is calculated based on five factors: No At-Fault Collision (NC), Drivable Area Compliance (DAC), Time-to-Collision (TTC), Comfort (Comf.), and Ego Progress (EP). The PDMS is calculated as " + }, + { + "bbox": [ + 55, + 472, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathrm{PDMS} = \\mathrm{NC} \\times \\mathrm{DAC} \\times \\frac{5 \\times \\mathrm{EP} + 5 \\times \\mathrm{TTC} + 2 \\times \\mathrm{C}}{12}" + }, + { + "bbox": [ + 55, + 472, + 295, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": "Bench2Drive Dataset and Metrics To further evaluate our framework, we conduct closed-loop experiments in the CARLA simulator [13]. Specifically, we adopt the recently proposed Bench2Drive [24] benchmark as our closed-loop evaluation benchmark. Bench2Drive provides a standardized training dataset on CARLA, ensuring fair comparisons across different methods. The benchmark consists of 220 short evaluation routes (around " + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "inline_equation", + "content": "150\\mathrm{m}" + }, + { + "bbox": [ + 55, + 545, + 297, + 713 + ], + "type": "text", + "content": " each) distributed across all CARLA towns, with each route containing one safety-critical scenario. This design effectively reduces the evaluation variance. For metric, it utilizes the standard CARLA metric, Driving Score (DS), as the primary metric. Additionally, Bench2Drive reports the success rate, which represents the proportion of successfully completed routes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 344, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 555, + 368 + ], + "type": "text", + "content": "A route is considered successful if and only if DS reaches " + }, + { + "bbox": [ + 313, + 344, + 555, + 368 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 344, + 555, + 368 + ], + "type": "text", + "content": " on this route." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 381, + 447, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 447, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 447, + 393 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": "NAVSIM For input data, it is aligned with TransFuser [31]. We concatenate the front-view image with center-cropped front-left and front-right images, resulting in a combined resolution of " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "256 \\times 1024" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": " pixels. For LiDAR, the " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "64\\mathrm{m} \\times 64\\mathrm{m}" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": " point cloud surrounding the ego vehicles is used. For network architecture, we employ ResNet34 as the backbone for BEV feature extraction following TransFuser [31]. The world model consists of two transformer decoder layers. We use " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": " trajectory anchors. The predicted future states are of time " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "t + 2s, t + 4s" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": " when the current is of time " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": ". The reward weights are set as follows: " + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "inline_equation", + "content": "w_{1} = 0.1, w_{2} = w_{3} = 0.5, w_{4} = 1.0" + }, + { + "bbox": [ + 313, + 400, + 556, + 674 + ], + "type": "text", + "content": ". The training is conducted on the Navtrain split using 8 NVIDIA L20 GPUs with a total batch size of 256, distributed across 50 epochs. In the ablation study section, we train our model for 20 epochs instead of 50 to accelerate the training process. For fast training, we pre-compute the simulation results (BEV semantic maps and scores) based on the trajectory anchors. We then feed trajectory anchors into the trajectory evaluation module during training. During testing, the trajectory evaluation module evaluates the refined trajectory instead of the anchor trajectories. We utilize the Adam optimizer with a learning rate of 1e-4." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Bench2Drive For Bench2Drive, we adopt TCP [42] as our baseline and integrate the trajectory evaluation module. We choose TCP as it is the best open-source framework officially" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27142" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 294, + 112 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 112 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 112 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 112 + ], + "type": "table", + "html": "
Traj. Eval.Future StatesNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
××96.491.576.290.399.981.0
✓×97.192.978.291.9100.083.2
✓✓98.094.779.993.4100.085.6
", + "image_path": "80d729add5851c43b89f1dc69c81ba530ce2e696512e1e4dad2f187f63ba9b0a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 173, + 298, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 173, + 298, + 281 + ], + "spans": [ + { + "bbox": [ + 54, + 173, + 298, + 281 + ], + "type": "text", + "content": "provided by Bench2Drive. Similar to NAVSIM, we utilize imitation rewards along with simulation-based rewards, including No Collision (NC) and Drivable Area Compliance (DAC). The number of trajectory anchors is set to 256. Notably, rather than employing the winner-take-all strategy (Eq. 9) to supervise only the best trajectory, we find that supervising all trajectories leads to improved performance. The model is trained for 27 epochs with a batch size of 300. We use the Adam optimizer with a learning rate of 1e-4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 288, + 191, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 288, + 191, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 191, + 301 + ], + "type": "text", + "content": "4.3. Comparison with SOTA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 305, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 388 + ], + "type": "text", + "content": "NASIM Table 1 compares our method with state-of-the-art approaches on the NAVSIM test set. By leveraging the world model to simulate future scenarios and guide trajectory evaluation, our method achieves a PDMS of 87.1. WoTE outperforms the previous model-free approach, Hydra-MDP, highlighting the advantages of model-based trajectory evaluation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 450 + ], + "type": "text", + "content": "Bench2Drive To further validate our approach, we conduct a closed-loop evaluation on the Bench2Drive benchmark within the CARLA simulator. As shown in Table 2, our method improves the Driving Score by 1.81 points, demonstrating its effectiveness in various evaluation settings." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 456, + 157, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 157, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 157, + 468 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "text", + "content": "Trajectory evaluation and future state prediction matter. In this experiment, we investigate the importance of the trajectory evaluation and future state prediction as Table 3 shows. The first row in Table 3 represents our baseline TransFuser [31], which performs trajectory prediction only. In the second row, we change to our framework, which consists of both trajectory prediction and trajectory evaluation. However, we do not used the predicted future states as the input of Reward Model. That is, " + }, + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{B}_{\\mathrm{all}} = \\mathbf{B}_t" + }, + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{\\mathrm{all}} = \\mathbf{a}_t" + }, + { + "bbox": [ + 55, + 474, + 297, + 617 + ], + "type": "text", + "content": " in Eq. 5. In the third row, we integrate the BEV world model to predict future states, leading to notable performance improvements across multiple metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": "Imitation and simulation rewards are complementary. To investigate how each type of reward affects driving performance, we conduct an ablation study, as shown in Table 4. During inference, we adjust the reward weights in Eq. 4 to isolate their effects: setting " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "w_{1} = 0" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " enables evaluation with only simulation rewards, while setting " + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "inline_equation", + "content": "w_{2} = w_{3} = w_{4} = 0" + }, + { + "bbox": [ + 55, + 618, + 297, + 714 + ], + "type": "text", + "content": " tests performance with only an imitation reward. The results indicate that imitation and simulation rewards emphasize" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 553, + 110 + ], + "blocks": [ + { + "bbox": [ + 55, + 120, + 295, + 143 + ], + "lines": [ + { + "bbox": [ + 55, + 120, + 295, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 120, + 295, + 143 + ], + "type": "text", + "content": "Table 3. Ablation study on trajectory evaluation and future state prediction. Traj: Trajectory. Eval.: Evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 72, + 553, + 110 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 110 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 110 + ], + "type": "table", + "html": "
Imi. RewardSim. RewardsNC↑DAC↑EP↑TTC↑Comf.↑PDMS↑
✓×97.992.577.793.499.983.5
×✓96.594.779.290.199.483.6
✓✓98.094.779.993.4100.085.6
", + "image_path": "8e4344f6de648e46009081368beee9ba37677b235e5ce049b53109b9e1c000c2.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 316, + 175, + 553, + 209 + ], + "blocks": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 152 + ], + "type": "text", + "content": "Table 4. Ablation study on the imitation and simulation rewards. The experiment shows that imitation and simulation rewards are complementary. Imi: Imitation. Sim.: Simulation." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 175, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 209 + ], + "type": "table", + "html": "
Future Prediction StepsNC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑
0s → 4s97.493.679.191.9100.084.0
0s → 2s → 4s98.094.779.993.4100.085.6
", + "image_path": "9b2cae285c4bbdadaef0e9f89e4456b5e7f3e5c98f34af501717baefa47340b3.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 285, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 313, + 217, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 262 + ], + "type": "text", + "content": "Table 5. Ablation study on the different number of time steps in the future prediction. Our BEV world model predicts future states in a recurrent manner. " + }, + { + "bbox": [ + 313, + 217, + 555, + 262 + ], + "type": "inline_equation", + "content": "\\mathrm{A} \\rightarrow \\mathrm{B}" + }, + { + "bbox": [ + 313, + 217, + 555, + 262 + ], + "type": "text", + "content": ": the predicted future state at time B is based on the state at time A." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 285, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 317, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 285, + 553, + 329 + ], + "type": "table", + "html": "
#Trajs.NC ↑DAC ↑EP ↑TTC ↑Comf. ↑PDMS ↑Latency ↓
6497.494.179.691.9100.084.517.2 ms
12897.994.680.093.0100.085.417.9 ms
25698.094.779.993.4100.085.618.7 ms
", + "image_path": "297e3ee86002da8512d05626266b644a4a587e8342a4cdb4d66b664414ecbcf5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 556, + 381 + ], + "lines": [ + { + "bbox": [ + 313, + 336, + 556, + 381 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 556, + 381 + ], + "type": "text", + "content": "Table 6. Ablation study on the different number of Trajectories. The number of trajectories is the same during training and testing. The latency is evaluated on an NVIDIA L20 GPU. Trajs.: Trajectories." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 413, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 475 + ], + "type": "text", + "content": "different aspects of planning. Trajectory evaluation using imitation rewards performs better in NC and TTC, while using simulation rewards excels in DAC and EP. By integrating both, the model leverages their respective strengths, achieving improved overall performance across all key metrics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 485, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 556, + 582 + ], + "type": "text", + "content": "Recurrent future state prediction helps. Our framework supports predicting future states recurrently, as detailed in Eq. 3. In Table 5, we present the results obtained by varying the number of BEV states predicted by the world model. Our findings indicate that using a finer time step for future state prediction significantly improves performance. This is because a finer time step provides richer temporal information, which aids trajectory evaluation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 557, + 715 + ], + "type": "text", + "content": "More trajectories help. This ablation study explores the effect of varying the number of trajectories on overall performance. Specifically, the model in each row is trained and tested under the same number of trajectories, and the world model is fed with refined trajectories. As shown in Table 6, using a small number of trajectories, such as 64, significantly degrades performance. As the number of trajectories increases, performance improves notably, with a marked improvement observed when increasing from 64 to 128 trajectories. However, as the number of trajectories increases from 128 to 256, the performance gains diminish," + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27143" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 219, + 105 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 219, + 105 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 219, + 105 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 219, + 105 + ], + "type": "table", + "html": "
Train Traj.Test Traj.Traj. Refine†NC ↑DAC ↑EP ↑
256256×96.694.679.0
2561024×97.295.679.3
256256✓98.094.779.9
", + "image_path": "c6f338334e2475f36eeba07ee9ed11cdeeb840316694a4f123dd8776c1ceb5cb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 219, + 135 + ], + "lines": [ + { + "bbox": [ + 55, + 114, + 219, + 135 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 219, + 135 + ], + "type": "text", + "content": "Table 7. Generalization of the trajectory across different trajectory configurations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 136, + 219, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 219, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 219, + 158 + ], + "type": "text", + "content": "the input trajectories of BEV world model a trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "text", + "content": "indicating that the performance of the most optimal level. Based on these results, we default number of trajectories for our ex- Latency analysis. A common concern is of our framework. To address this, we run on an NVIDIA L20 GPU, as shown in GPU processes state-action pairs in para remains efficient even as the number of t The total latency is only " + }, + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "inline_equation", + "content": "18.7\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "text", + "content": " well " + }, + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "inline_equation", + "content": "\\mathsf{V}" + }, + { + "bbox": [ + 55, + 190, + 219, + 310 + ], + "type": "text", + "content": " requirements for end-to-end autonomous" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 319, + 219, + 331 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 219, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 219, + 331 + ], + "type": "text", + "content": "Generalization ability across different" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "content": "shown in Table 7, we investigate the generalization ability of our trajectory evaluation module. To be specific, we sample " + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "inline_equation", + "content": "N = 1024" + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "content": " instead of " + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "content": " trajectory anchors from the dataset. Since our model is trained under the setting of " + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "inline_equation", + "content": "N = 256" + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "content": " trajectory anchors, the model has never seen these " + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "inline_equation", + "content": "N = 1024" + }, + { + "bbox": [ + 54, + 332, + 296, + 475 + ], + "type": "text", + "content": " trajectories. Surprisingly, we find a 1.3 PDMS increase when using these unseen trajectory anchors. Besides, the model performs even better when using refined trajectories as inputs. These refined trajectories continuously change based on varying sensor inputs. This shows that the proposed world model has great generalization ability when dealing with unseen trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 485, + 150, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 150, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 150, + 498 + ], + "type": "text", + "content": "4.5. Visual Analysis" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 503, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 296, + 551 + ], + "type": "text", + "content": "End-to-end planning trajectories. To highlight the improvements in end-to-end driving achieved through trajectory evaluation, we present visualizations of end-to-end planning trajectories in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 296, + 583 + ], + "type": "text", + "content": "Trajectory rewards. Figure 4 illustrates the predicted rewards for different trajectories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 128, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "In this paper, we introduced a novel framework that leverages a BEV world model for end-to-end trajectory evaluation in autonomous driving. By integrating a BEV world model, we enable a more informed evaluation process that considers the dynamic evolution of driving scenarios, leading to more effective trajectory evaluation. Our approach benefits from dense supervision provided by BEV-space traffic simulators, which supply both semantic future states and rule-based" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 72, + 550, + 223 + ], + "blocks": [ + { + "bbox": [ + 321, + 72, + 550, + 223 + ], + "lines": [ + { + "bbox": [ + 321, + 72, + 550, + 223 + ], + "spans": [ + { + "bbox": [ + 321, + 72, + 550, + 223 + ], + "type": "image", + "image_path": "a81390ca3df63e0c7551fcba665b66c1848c10cd21449e8f9f189b80987ad348.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 233, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 313, + 233, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 555, + 289 + ], + "type": "text", + "content": "Figure 3. Visualization of end-to-end planning trajectories. We compare our method with TransFuser [31], which serves as our trajectory prediction baseline without a trajectory evaluation module. Our trajectory evaluation module effectively filters out low-quality trajectories and helps avoid collisions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 318, + 301, + 547, + 396 + ], + "blocks": [ + { + "bbox": [ + 318, + 301, + 547, + 396 + ], + "lines": [ + { + "bbox": [ + 318, + 301, + 547, + 396 + ], + "spans": [ + { + "bbox": [ + 318, + 301, + 547, + 396 + ], + "type": "image", + "image_path": "3edd753908390f9104df1a78bed1563b0620bf91fbf95d995bd72770970f2b46.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 360, + 396, + 484, + 406 + ], + "lines": [ + { + "bbox": [ + 360, + 396, + 484, + 406 + ], + "spans": [ + { + "bbox": [ + 360, + 396, + 484, + 406 + ], + "type": "text", + "content": "(a) Multi-modality trajectory preservation" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 318, + 407, + 547, + 501 + ], + "blocks": [ + { + "bbox": [ + 318, + 407, + 547, + 501 + ], + "lines": [ + { + "bbox": [ + 318, + 407, + 547, + 501 + ], + "spans": [ + { + "bbox": [ + 318, + 407, + 547, + 501 + ], + "type": "image", + "image_path": "e9903960bdf2540bf5e55a7e97e3140fc1b6eb009ec09ab904f0ec0d00d5937d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 364, + 501, + 479, + 510 + ], + "lines": [ + { + "bbox": [ + 364, + 501, + 479, + 510 + ], + "spans": [ + { + "bbox": [ + 364, + 501, + 479, + 510 + ], + "type": "text", + "content": "(b) Low-quality trajectory suppression" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 524, + 555, + 612 + ], + "lines": [ + { + "bbox": [ + 313, + 524, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 555, + 612 + ], + "type": "text", + "content": "Figure 4. Visualization of rewards of all trajectories. Brighter colors indicate higher trajectory rewards, while gray trajectories correspond to those with very low rewards. (a): WoTE effectively retains diverse multi-modal trajectories, as long as they comply with traffic rules and do not result in unsafe outcomes. (b): it demonstrates that trajectories considered unreasonable receive significantly lower rewards, highlighting the effectiveness of our trajectory evaluation module." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "reward targets. Our method achieves state-of-the-art performance on the NAVSIM and Bench2Drive benchmarks while maintaining real-time efficiency. Overall, our work establishes trajectory evaluation as an important research direction for end-to-end autonomous driving. We hope our framework serves as a strong baseline and inspires further research in end-to-end online trajectory evaluation." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27144" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "6. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 139 + ], + "type": "text", + "content": "This work was supported in part by the National Science and Technology Major Project (No. 2022ZD0116500), the National Natural Science Foundation of China (No. U21B2042, No. 62320106010)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 115, + 164 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 171, + 297, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 297, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 297, + 214 + ], + "type": "text", + "content": "[1] Sébastien MR Arnold, Pierre L'Ecuyer, Liyu Chen, Yi-fan Chen, and Fei Sha. Policy learning and evaluation with randomized quasi-monte carlo. arXiv preprint arXiv:2202.07808, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 296, + 261 + ], + "type": "text", + "content": "[2] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. CVPR, 2020. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 262, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 262, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 262, + 296, + 316 + ], + "type": "text", + "content": "[3] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 318, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 296, + 351 + ], + "type": "text", + "content": "[4] Dian Chen, Brady Zhou, Vladlen Koltun, and Philipp Krahenbuhl. Learning by cheating. In Conference on Robot Learning, pages 66-75. PMLR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 353, + 296, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 296, + 385 + ], + "type": "text", + "content": "[5] Dian Chen, Vladlen Koltun, and Philipp Krahenbuhl. Learning to drive from a world on rails. In CVPR, pages 15590-15599, 2021. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 296, + 430 + ], + "type": "text", + "content": "[6] Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 296, + 486 + ], + "type": "text", + "content": "[7] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024. 1, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 520 + ], + "type": "text", + "content": "[8] OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction benchmark in autonomous driving, 2023. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 566 + ], + "type": "text", + "content": "[9] Christoph Dann, Gerhard Neumann, and Jan Peters. Policy evaluation with temporal differences: A survey and comparison. The Journal of Machine Learning Research, 15(1): 809-883, 2014. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 296, + 612 + ], + "type": "text", + "content": "[10] Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning-based vehicle motion planning. In Conference on Robot Learning, pages 1268-1281. PMLR, 2023. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "type": "text", + "content": "[11] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. arXiv preprint arXiv:2406.15349, 2024. 2, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "text", + "content": "[12] Christopher Diehl, Timo Sievernich, Martin Krüger, Frank Hoffmann, and Torsten Bertram. Umbrella: Uncertainty-aware model-based offline reinforcement learning leveraging planning. arXiv preprint arXiv:2111.11097, 2021. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[13] Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In CoRL, 2017. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 108, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 108, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 108, + 554, + 152 + ], + "type": "text", + "content": "[14] Haoyang Fan, Fan Zhu, Changchun Liu, Liangliang Zhang, Li Zhuang, Dong Li, Weicheng Zhu, Jiangtao Hu, Hongye Li, and Qi Kong. Baidu Apollo em motion planner. arXiv preprint arXiv:1807.08048, 2018. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 153, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 206 + ], + "type": "text", + "content": "[15] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 555, + 262 + ], + "type": "text", + "content": "[16] Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, et al. Waymax: An accelerated, data-driven simulator for large-scale autonomous driving research. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 555, + 308 + ], + "type": "text", + "content": "[17] Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 555, + 353 + ], + "type": "text", + "content": "[18] Mikael Henaff, Alfredo Canziani, and Yann LeCun. Model-predictive policy learning with uncertainty regularization for driving in dense traffic. In International Conference on Learning Representations, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 399 + ], + "type": "text", + "content": "[19] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 399, + 555, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 555, + 454 + ], + "type": "text", + "content": "[20] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 456, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 554, + 498 + ], + "type": "text", + "content": "[21] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In ECCV, 2022. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 554, + 544 + ], + "type": "text", + "content": "[22] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Goal-oriented autonomous driving. arXiv preprint arXiv:2212.10156, 2022. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 555, + 599 + ], + "type": "text", + "content": "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, Lewei Lu, Xiaosong Jia, Qiang Liu, Jifeng Dai, Yu Qiao, and Hongyang Li. Planning-oriented autonomous driving. In CVPR, 2023. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 555, + 645 + ], + "type": "text", + "content": "[24] Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. arXiv preprint arXiv:2406.03877, 2024. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 690 + ], + "type": "text", + "content": "[25] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In ICCV, 2023. 1, 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 554, + 713 + ], + "type": "text", + "content": "[26] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27145" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "autonomous driving with latent world model. arXiv preprint arXiv:2406.08481, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 150 + ], + "type": "text", + "content": "[27] Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978, 2024. 1, 2, 3, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 184 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dolkar. Focal loss for dense object detection. In ICCV, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 239 + ], + "type": "text", + "content": "[29] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learning. nature, 518(7540):529-533, 2015. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 274 + ], + "type": "text", + "content": "[30] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, pages 7077-7087, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 308 + ], + "type": "text", + "content": "[31] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 296, + 354 + ], + "type": "text", + "content": "[32] Guanren Qiao, Guorui Quan, Rongxiao Qu, and Guiliang Liu. Modelling competitive behaviors in autonomous driving under generative world model. In European Conference on Computer Vision, pages 19-36. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 296, + 409 + ], + "type": "text", + "content": "[33] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 411, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 296, + 453 + ], + "type": "text", + "content": "[34] Abbas Sadat, Sergio Casas, Mengye Ren, Xinyu Wu, Pranaab Dhawan, and Raquel Urtasun. Perceive, predict, and plan: Safe motion planning through interpretable semantic representations. In ECCV, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 510 + ], + "type": "text", + "content": "[35] Ardi Tampuu, Tambet Matiisen, Maksym Semikin, Dmytro Fishman, and Naveed Muhammad. A survey of end-to-end driving: Architectures and training methods. IEEE Transactions on Neural Networks and Learning Systems, 33(4): 1364-1384, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 296, + 544 + ], + "type": "text", + "content": "[36] Marin Toromanoff, Emilie Wirbel, and Fabien Moutarde. End-to-end model-free reinforcement learning for urban driving using implicit affordances. In CVPR, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 296, + 578 + ], + "type": "text", + "content": "[37] Martin Treiber, Ansgar Hennecke, and Dirk Helbing. Congested traffic states in empirical observations and microscopic simulations. Physical review E, 62(2):1805, 2000. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 622 + ], + "type": "text", + "content": "[38] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "text", + "content": "[39] Yuqi Wang, Ke Cheng, Jiawei He, Qitai Wang, Hengchen Dai, Yuntao Chen, Fei Xia, and Zhaoxiang Zhang. Drivingndojo dataset: Advancing interactive and knowledge-enriched driving world model. arXiv preprint arXiv:2410.10738, 2024." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[40] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Con-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 408 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ference on Computer Vision and Pattern Recognition, pages 14749-14759, 2024. 1, 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 150 + ], + "type": "text", + "content": "[41] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15449-15458, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 194 + ], + "type": "text", + "content": "[42] Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided control prediction for end-to-end autonomous driving: a simple yet strong baseline. NeurIPS, 2022. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 251 + ], + "type": "text", + "content": "[43] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, et al. Drama: An efficient end-to-end motion planner for autonomous driving with mamba. arXiv preprint arXiv:2408.03601, 2024. 2, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "type": "text", + "content": "[44] Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 361 + ], + "type": "text", + "content": "[45] Yumeng Zhang, Shi Gong, Kaixin Xiong, Xiaoqing Ye, Xiao Tan, Fan Wang, Jizhou Huang, Hua Wu, and Haifeng Wang. Bevworld: A multimodal world model for autonomous driving via unified bev latent space. arXiv preprint arXiv:2407.05679, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 364, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 364, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 364, + 554, + 408 + ], + "type": "text", + "content": "[46] Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863-17873, 2023. 5" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27146" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_content_list.json b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f8ca7ce999d82e2e2a1deb72ae2cbc604a485bb1 --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_content_list.json @@ -0,0 +1,1587 @@ +[ + { + "type": "text", + "text": "End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation", + "text_level": 1, + "bbox": [ + 114, + 138, + 883, + 181 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liwei Wang $^{1}$ , Yanduo Zhang $^{1,2\\text{串}}$ , Tao Lu $^{1}$ , Fang Liu $^{1}$ , Huiqin Zhang $^{1}$ , Jiayi Ma $^{3}$ , Huabing Zhou $^{1\\text{串}}$ $^{1}$ Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology \n $^{2}$ Hubei University of Arts and Science $^{3}$ Wuhan University \nwlw951226@163.com, {ydzhang,lut}@wit.edu.cn, {fangliu8822,666zhanghuiqin,jyma2010,Zhouhuabing}@gmail.com", + "bbox": [ + 207, + 210, + 789, + 313 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 254, + 348, + 330, + 363 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets $<$ subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in https://github.com/wlw951226/ARN.", + "bbox": [ + 99, + 380, + 483, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 102, + 763, + 230, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Scene graph generation is a promising approach to scene understanding and has emerged as a frontier topic. Dynamic Scene Graph Generation (DSGG) advances this by capturing dynamic visual relationships in videos, contributing to a comprehensive understanding of video content. DSGG has", + "bbox": [ + 99, + 789, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53fe100a31e68d35bffe889123a41815bebc3c60c5f2ac1d9c69979803d6e177.jpg", + "image_caption": [ + "Figure 1. (a) The association frequency matrix between entities and predicates in the dataset [13]; (b) the association frequency matrix between predicate subclasses, where the vertical axis represents the pairwise combinations of two predicate subclasses; (c) an example and the pipeline of our associative reasoning network." + ], + "image_footnote": [], + "bbox": [ + 517, + 348, + 890, + 579 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "also opened up new possibilities in areas such as video analysis [9, 18, 42], video captioning [12, 38], and video question answering [5, 43], enabling computers to better simulate human perception and cognitive processes in real-world scenarios.", + "bbox": [ + 509, + 683, + 893, + 755 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic scene graphs encode implicitly co-occurrences and affordance relations among entities and predicates. We analyze the frequency co-occurrence of predicates and entities in the Action Genome [13] dataset. As shown in Fig. 1(a), there is a strong association between certain predicates and entities; for instance, the predicate \"writing on\" often implies that the object is a notebook. Similarly, when the object is a notebook, it is likely in front of a person who is looking at it. Fig. 1(b) highlights the association be", + "bbox": [ + 509, + 757, + 895, + 890 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 116, + 16, + 189, + 55 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 245, + 13, + 799, + 56 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Corresponding authors.", + "bbox": [ + 124, + 876, + 264, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "17729", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tween different predicates, and Fig. 1(c) provides an example demonstrating that holding an object strongly suggests its position in front of the person. However, existing methods overlook the dependencies between entities and predicates. Modeling these visual dependencies enables a multiperspective understanding of scenes, resulting in more accurate scene graphs.", + "bbox": [ + 99, + 102, + 480, + 205 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In addition, the currently prevalent DSGG paradigm [4, 6, 15, 29, 31, 36] are built upon a two-stage pipeline, focusing on capturing the temporal dependencies of videos. This two-stage paradigm independently optimizes the subtasks of entity localization [33] and predicate prediction, hindering end-to-end learning. Additionally, these methods generate redundant candidate boxes and predict relationships for all detected entity pairs, increasing computational cost.", + "bbox": [ + 99, + 205, + 480, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, the one-stage detector DETR [2] has transformed the pattern of traditional detection frameworks by adopting a set-based prediction approach for one-to-one matching, eliminating the need for computationally expensive hand-designed components such as non-maximum suppression. Meanwhile, DETR-based methods [3, 25, 30, 35, 41] have demonstrated remarkable performance in image understanding.", + "bbox": [ + 99, + 340, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we extend DETR to DSGG and propose an end-to-end Association Reasoning Network (ARN). ARN explores visual relation dependencies within scene graphs and performs association reasoning with fine-grained entity context. ARN leverages Contrastive Language-Image Pretraining (CLIP) [32] to bridge vision and language, modeling triplet cues through textual prompts, and extending predicate prediction to triplet prediction. This approach enriches the semantic representation of predicate prediction while ensuring semantic consistency between paired entities and predicates in DSGG.", + "bbox": [ + 99, + 459, + 480, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We further design a Predicate Association Parsing (PAP) module to link fine-grained representations from paired entities and different predicate subclasses. It employs a conditional weight mapping mechanism to dynamically organize predicate representations across different branches. Additionally, we develop a Hierarchical Attention (HA) mechanism that integrates spatio-temporal context with complementary predicate and entity cues in a structured hierarchy. This method achieves association reasoning between entities and predicates while maintaining stable performance during end-to-end training. Our experiments on the widely used Action Genome dataset demonstrate that ARN achieves state-of-the-art performance in DSGG.", + "bbox": [ + 99, + 621, + 480, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper can be summarized as follows:", + "bbox": [ + 101, + 814, + 480, + 842 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose an end-to-end ARN model for DSGG, which explores visual relation dependencies between entities and predicates and performs association reasoning to gen", + "bbox": [ + 101, + 845, + 480, + 888 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erate accurate scene graphs.", + "bbox": [ + 526, + 103, + 707, + 116 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a Predicate Association Parsing (PAP) module to associate and organize fine-grained representations of different predicate subclasses and paired entities.", + "- We designed a Hierarchical Attention (HA) module to hierarchically perform visual relation association reasoning between paired entity representations and predicate relational context.", + "- We conducted a comprehensive evaluation of the proposed method, demonstrating its superior performance." + ], + "bbox": [ + 511, + 117, + 893, + 248 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 262, + 648, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image scene graph generation. Image scene graph generation aims to identify entities and predict the predicates of image in form of triplets, which initially proposed by [14]. Most early works [8, 19, 23, 27, 44, 45], focused on refining entity and visual relationship representations in images using various network architectures. Subsequently, several works have attempted to design end-to-end image SGG models. Cong et al. [7] treated SGG as a set prediction problem and proposed an end-to-end image SGG model that directly predicts sparse scene graphs based solely on visual appearance. Furthermore, Li et al. [20] designed a graph assembling module that infers the connectivity of a bipartite graph based on entities features, generating scene graphs through graph matching. Meanwhile, some researchers [28] have focused on addressing the long-tailed distribution of predicates in benchmark datasets [16] to generate unbiased scene graphs.", + "bbox": [ + 509, + 287, + 893, + 535 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dynamic scene graph generation. The dynamic visual relationships in videos are more complex and have temporal dependencies, making DSGG more challenging compared to image scene graph generation. Previous works have focused on capturing temporal dependencies in videos and addressing data imbalance caused by long-tail distributions. Some works focus on designing more efficient temporal dependencies reasoning networks. Cong et al. [6] employed an encoder-decoder architecture to extract spatiotemporal context from video frames and proposed a two-stage baseline. Building on this, Pu et al. [31] injected prior spatio-temporal knowledge into DSGG via statistical co-occurrence, proposing a spatio-temporal Knowledge-Embedded Transformer. However, this method heavily relies on the inherent distribution of the data, such as the long-tail phenomenon. Differently, ARN does not rely on statistical priors.", + "bbox": [ + 509, + 537, + 893, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of the data's inherent distribution, Nag et al. [29] proposed a memory-guided training method to alleviate the long-tailed distribution. Lin et al. [24] introduced an asymmetric re-weighting loss that relieve the issue of label bias by adjusting the weights of tail predicates. Despite some progress, these methods sacrifice the performance of head predicates to focus on tail", + "bbox": [ + 509, + 786, + 893, + 888 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "17730", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "predicates. Additionally, they overlook the visual relational dependencies within the scene graph, which encompass a broader understanding of interaction context.", + "bbox": [ + 99, + 102, + 483, + 145 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most relevant work to ours is OED [37], which introduces a progressive refined module to aggregate temporal context. But this module relies on independently optimized spatial context, resulting in degraded performance during end-to-end training. In contrast to OED, we focus on exploring visual dependencies between entities to enhance holistic scene understanding and enabling fully end-to-end implementation.", + "bbox": [ + 99, + 146, + 483, + 263 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 101, + 276, + 191, + 291 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we extend the end-to-end detector DETR[2] to DSGG and propose an Association Reasoning Network (ARN), as illustrated in Fig. 2. ARN sequentially extracts visual features and feeds them into two cascaded decoder modules, the Entity Decoder and the Predicate Decoder, to perform two subtasks: entity detection and triplet prediction (Sec 3.1). Subsequently, we introduce triplet cues modeling using CLIP's semantic knowledge (Sec 3.2). Moreover, the Predicate Association Parsing (PAP) Module in predicate decoder organizes predicate cues from different branches and fine-grained paired entity features (Sec 3.3). The Hierarchical Attention (HA) mechanism then aggregates these with spatio-temporal features to perform association reasoning, enhancing broader contextual understanding of video scenes (Sec 3.4). Finally, we introduce the training loss and inference process of ARN (Sec 3.5).", + "bbox": [ + 99, + 300, + 483, + 536 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 101, + 544, + 294, + 559 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of our ARN is illustrated in Fig. 2. The ARN is composed of a Backbone, an Encoder, an Entity decoder and a Predicate Decoder.", + "bbox": [ + 99, + 565, + 483, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Backbone.** CNN-based backbone maps an input frame $I_{i}$ to a spatial feature map $\\mathbf{f}_i\\in \\mathbb{R}^{H\\times W\\times C}$ .", + "bbox": [ + 101, + 609, + 482, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Encoder. The transformer-based encoder [2, 3, 30] refines image features and captures spatial contextual information. First, a $1 \\times 1$ convolution and flatten operation are applied to reduce the dimension and transform it into a serialized feature representation $\\mathbf{f}_i \\in \\mathbb{R}^{HW \\times d}$ . Next, we feed the $\\mathbf{f}_i$ and fixed positional encodings $\\mathbf{p}_e$ [1] to the encoder and output the enhanced spatial feature $\\mathbf{F}_{vi} \\in \\mathbb{R}^{HW \\times d}$ .", + "bbox": [ + 99, + 640, + 483, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Entity Decoder. Entity decoder consists of multiple transformer decoder layers and several FFN heads and introduces an entity query set $\\mathbf{Q}^E\\in \\mathbb{R}^{N_q\\times d}$ , where $N_{q}$ is the number of queries. Unlike previous methods, our query set has two subsets $\\mathbf{Q}^E = [\\mathbf{Q}^s,\\mathbf{Q}^o ]\\in \\mathbb{R}^{N_q\\times d}$ that aim to separate the subject and the object.", + "bbox": [ + 99, + 743, + 483, + 830 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Moreover, a pair of learnable entity position embeddings $\\{\\mathbf{e}^s,\\mathbf{e}^o\\} \\in \\mathbb{R}^{N_q\\times d}$ is designed to assign the subject and object at the same position to form pairwise entities. Furthermore, $\\mathbf{e}^o$ is employed to align the query sets for the two", + "bbox": [ + 101, + 830, + 483, + 890 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "decoders and update them during training. Overall, the implementation of the entity decoder layer is as follows:", + "bbox": [ + 511, + 102, + 893, + 132 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {E} = \\operatorname {C o n c a t} \\left[ \\mathbf {Q} ^ {s} + \\mathbf {e} ^ {s}, \\mathbf {Q} ^ {o} + \\mathbf {e} ^ {o} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 143, + 893, + 162 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {Q}} _ {(l)} ^ {E} = \\operatorname {D e c} _ {(l)} ^ {E} \\left(\\mathbf {Q} ^ {E}, \\mathbf {F} _ {v i}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 179, + 893, + 199 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Specifically, at each layer $\\mathrm{Dec}_{(l)}^{E}$ , the queries first undergoes self-attention to enable interactions among queries. Subsequently, cross-attention aggregates visual features $\\mathbf{F}_{vi}$ into the queries to obtain updated queries $\\tilde{\\mathbf{Q}}_{(l)}^{E}$ . The output of the entity decoder, denoted as $\\mathbf{F}^{E}$ , is derived from the final updated queries $\\tilde{\\mathbf{Q}}_{(l)}^{E}$ after $l$ layers.", + "bbox": [ + 511, + 205, + 893, + 300 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, the entity representation $\\mathbf{F}^E$ passes through the FFN head and object classifier at the top of the entity decoder to obtain the bounding box coordinates $\\mathbf{B}^{E} = \\{\\mathbf{b}^{s},\\mathbf{b}^{o}\\} \\in \\mathbb{R}^{N_{q}\\times 4}$ and object categories $\\mathbf{C}^o\\in \\mathbb{R}^{N_q}$ for the paired entities.", + "bbox": [ + 511, + 301, + 893, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Predicate Decoder The predicate decoder is tailored to capture temporal dependencies and facilitate associative reasoning between entities and multiple predicate subclasses. In detail, the structure of the predicate decoder is a multi-branch decoder architecture, each branch corresponding to the prediction of a predicate subclass.", + "bbox": [ + 511, + 375, + 893, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each branch consists of multiple decoder layers and initializes predicate queries $\\mathbf{Q}^{predi} \\in \\mathbb{R}^{N_q \\times d}$ in $i$ -th branch. Unlike the entity decoder, we capture temporal dependencies in visual features $\\mathbf{F}_{vi}$ . Specifically, we initialized a learnable frame encoding $\\mathbf{p}_t \\in \\mathbb{R}^{N_q \\times t \\times d}$ to identify different frames. Subsequently, we transpose the $\\mathbf{F}_{vi}$ in the temporal dimension and update it by self-attention (SA) operation to obtain the spatio-temporal context feature $\\mathbf{F}_{vit}$ .", + "bbox": [ + 511, + 462, + 893, + 580 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {v i t} = S A (\\text {t r a n s p o s e} (\\mathbf {F} _ {v i})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 590, + 893, + 607 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ultimately, $\\mathbf{Q}^{predi}$ aggregates the $\\mathbf{F}_{vit}$ through crossattention to generate fine-grained triplet representations $\\mathbf{F}^{predi}$ . Furthermore, to encourage associative reasoning across predicates, we design a predicate association embedding $\\mathbf{e}^{predi} \\in \\mathbb{R}^{N_q \\times d}$ , which is dynamically updated along with the $\\mathbf{Q}^{predi}$ in the predicate decoder. The decoding process of the predicate representation in the $l$ -th layer of predicate decoder $Dec_l^{predi}$ is as follows:", + "bbox": [ + 511, + 616, + 893, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {e}} ^ {p r e d i} = \\mathbf {e} ^ {o} + \\mathbf {e} ^ {p r e d i}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 746, + 893, + 763 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {(l)} ^ {p r e d i} = \\operatorname {D e c} _ {(l)} ^ {p r e d i} \\left(\\mathbf {Q} ^ {p r e d i}, \\tilde {\\mathbf {e}} ^ {p r e d i}, \\mathbf {F} _ {v i t}, \\mathbf {p} _ {t}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 794, + 893, + 815 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Semantic Triplet Cue Modeling", + "text_level": 1, + "bbox": [ + 511, + 823, + 782, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We use CLIP [32] to extending traditional predicate prediction to triplet prediction and inject domain-agnostic prior knowledge into our ARN.", + "bbox": [ + 511, + 844, + 893, + 888 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "17731", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b3eea5138604925be438f43914f39ef99f113dd5ccd301f055c5792835e38afd.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed Association Reasoning Network (ARN), which first extracts visual features from the video, and the entity decoder is used to locate entities. Subsequently, the multi-branch predicate decoder captures the temporal dependencies in visual features and infers representations for each predicate subclasses. Finally, the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) associates different predicate subclasses cues and aggregates fine-grained entity features to refine the prediction results." + ], + "image_footnote": [], + "bbox": [ + 143, + 101, + 854, + 404 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we follow the ActionGenome [13] setting and divide predicates into three subclasses: Attention, Spatial, and Contact. Then we design distinct textual prompts for each predicate subclass to build a bridge for interaction between CLIP and visual triplets :(1) 'The person is [attention predicate] at the [object]'; (2) 'The [object] is [spatial predicate] the person'; (3) 'a photo of a person maybe [contacting predicate] the [object]', an example is illustrated on the right side of Fig. 2. Next, we use the CLIP text encoder to extract text embeddings $\\mathbf{E}^{predi} \\in \\mathbb{R}^{N_{predi} \\times d}$ for each predicate subclass triplets, where $\\mathbf{E}^{predi}$ serves as semantic cues for the visual triplets, $N_{predi}$ denotes the number of triplet categories in $i$ -th predicate decoder branch.", + "bbox": [ + 99, + 510, + 483, + 731 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on this, we use the $\\mathbf{E}^{predi}$ to initialize the classifier weights $w^{predi}$ . For the $n$ -th query in $i$ -th branch, it computes the similarity score $s_n^i$ between the query representation and the triplet embeddings to align the visual and semantic information, then apply softmax activation to compute the focal loss [22], yielding the predicate prediction loss $\\mathcal{L}_{\\mathrm{pred}}$ .", + "bbox": [ + 99, + 736, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Moreover, for the entity decoder, we design textual prompts for object categories `a photo of person and a [object]''. We extract text embeddings $E^o \\in$", + "bbox": [ + 99, + 845, + 483, + 890 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{N_{obj}\\times d}$ in the same manner and use them to initialize the classifier weights $w^o$ of the entity decoder. The object classifier is optimized using cross-entropy loss.", + "bbox": [ + 511, + 508, + 895, + 555 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {n} ^ {i} = \\left[ \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {1} ^ {p r e d i}\\right), \\dots , \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {N _ {p r e d i}} ^ {p r e d i}\\right) \\right] \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 580, + 893, + 613 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r e d}} = \\frac {1}{N} \\sum_ {n \\in N _ {q}} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\sigma \\left(s _ {n} ^ {i}\\right), y _ {r} ^ {\\text {p r e d i}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 627, + 893, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Predicate Association Parsing Module", + "text_level": 1, + "bbox": [ + 511, + 675, + 831, + 691 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As mentioned previously, the pairwise entities can be associated with multiple predicates, and these entity-predicate associative cues are crucial for DSGG. Therefore, we propose a Predicate Association Parsing (PAP) Module that first leverages a conditional weight mapping mechanism to differentiate and organize multiple predicate subclass cues, then performs associative reasoning by aggregating global spatio-temporal context and fine-grained entity representations.", + "bbox": [ + 509, + 696, + 895, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the outputs of different branches in the predicate decoder as $\\mathbf{F}^{predA}$ , $\\mathbf{F}^{predS}$ , and $\\mathbf{F}^{predC}$ , using the initials of predicate subclasses for simplicity. For $\\mathbf{F}^{predA}$ , we treat $\\mathbf{F}^{predS}$ and $\\mathbf{F}^{predC}$ as complementary predicate", + "bbox": [ + 511, + 830, + 895, + 890 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "17732", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/484f43c2ac52c30acb5c9c7b3c75eb7e17cd7b7ff62c521e5162c07aa12417a9.jpg", + "image_caption": [ + "Predicate Association Parsing Module", + "Figure 3. The architecture of the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) Mechanism. The PAP organizes complementary predicate cues and embeds entity context into the current predicate representation. Building on this, the HA aggregates spatio-temporal context, explores visual relation dependencies, and progressively refines the predicate representation." + ], + "image_footnote": [], + "bbox": [ + 125, + 119, + 478, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "representations and use a linear layer to reduce the dimensionality of them. Then we concatenate them as complementary predicate cues for the current branch, denoted as $\\mathbf{F}^{compA} = \\mathrm{Concat}[\\mathbf{F}^{\\mathrm{predS}},\\mathbf{F}^{\\mathrm{predC}}]$ . During the dot product operation, the cross-attention weights consist of two components, balancing the contributions of different predicate subclasses and preventing the model from being biased toward any single predicate representation.", + "bbox": [ + 99, + 487, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} = \\left(\\mathbf {Q} _ {S}, \\mathbf {Q} _ {C}\\right) = \\left(\\mathbf {F} ^ {p r e d S}, \\mathbf {F} ^ {p r e d C}\\right), \\mathbf {K} = \\mathbf {F} _ {v i t} + \\mathbf {P} _ {t}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 612, + 482, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} ^ {T} \\mathbf {K} = \\mathbf {Q} _ {S} ^ {T} \\mathbf {K} + \\mathbf {Q} _ {C} ^ {T} \\mathbf {K}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 652, + 482, + 670 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we integrate the fine-grained paired entity features $\\mathbf{F}^E$ into $\\mathbf{F}^{predA}$ .", + "bbox": [ + 101, + 675, + 482, + 704 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {f u s i o n} ^ {p r e d A} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left[ \\mathbf {F} ^ {p r e d A} + \\mathbf {e} ^ {p r e d A}, \\mathbf {F} ^ {E} \\right]\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 726, + 482, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{e}^{predA}$ denotes the predicate association embedding for attention predicates. Finally, we perform associative reasoning between the entity representation and various predicate representations through the proposed hierarchical attention module.", + "bbox": [ + 99, + 755, + 483, + 829 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Hierarchical Attention", + "text_level": 1, + "bbox": [ + 101, + 838, + 305, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Visual relational dependencies are crucial for understanding scene context, yet traditional cross attention mechanisms", + "bbox": [ + 101, + 859, + 483, + 888 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "struggle to associate multiple semantic representations. To this end, we designed a Hierarchical Attention (HA) Mechanism, building upon the predicate association parsing module. The HA mechanism enables fine-grained integration of entity features with diverse predicate representations, facilitating associative reasoning and improving the prediction of tail predicates. The concept is illustrated in Fig. 3.", + "bbox": [ + 511, + 102, + 893, + 204 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, HA is a quadruplet attention paradigm $(\\mathbf{Q}_{cur},\\mathbf{Q}_{comp},\\mathbf{K},\\mathbf{V})$ , where $\\mathbf{Q}_{cur}$ is the predicate subclass representation of the current branch, and $\\mathbf{Q}_{comp}$ represents a complementary predicate cues, which is the cues representation of the other predicate subclasses.", + "bbox": [ + 511, + 205, + 893, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Taking the branch of the attention predicate subclass as an example, we perform the first-level cross-attention operation to associate complementary predicate cues $\\mathbf{F}^{compA}$ , where $\\mathbf{F}^{compA}$ serves as the query $\\mathbf{Q}_{comp}$ , and $\\mathbf{F}_{vit}$ is used as the key and value, $\\mathbf{e}^o$ as the query pos. This process highlights regions of interest related to other predicate subclasses and updates $\\mathbf{Q}_{comp}$ to obtain $\\mathbf{F}_{comp}$ .", + "bbox": [ + 511, + 277, + 893, + 382 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {c o m p}} = \\operatorname {A t t} _ {1} ^ {H A} \\left(\\mathbf {Q} _ {\\text {c o m p}}, \\mathbf {K} = \\mathbf {F} _ {\\text {v i t}} + \\mathbf {p} _ {t}, \\mathbf {V} = \\mathbf {F} _ {\\text {v i t}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 404, + 893, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, we perform the second-level cross-attention operation, where $\\mathbf{F}^{predA}$ serves as the query $\\mathbf{Q}_{cur}$ , $\\mathbf{Q}_{comp}$ as the key and $\\mathbf{F}_{comp}$ as the value. This operation first establishes the association between the current predicate subclass and complementary predicate cues, then propagates the spatio-temporal visual features to the query of the current subclass, highlighting the most relevant categories and updating the predicate representation $\\mathbf{F}^{predA}$ . Finally, the $\\mathbf{F}^{predA}$ passes through the linear mapping layer and is added to the original input $\\mathbf{Q}_{cur}$ , yielding the final output $\\mathbf{F}_{assoc}^{predA}$ .", + "bbox": [ + 511, + 431, + 893, + 593 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} ^ {p r e d A} = \\operatorname {A t t} _ {2} ^ {H A} \\left(\\mathbf {Q} _ {c u r} = \\mathbf {F} ^ {p r e d A}, \\mathbf {Q} _ {c o m p}, \\mathbf {F} _ {s u p}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 614, + 893, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {\\text {a s s o c}} ^ {\\text {p r e d} A} = \\operatorname {L i n e a r} \\left(\\mathbf {F} ^ {\\text {p r e d} A}\\right) + \\mathbf {F} ^ {\\text {p r e d} A}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 655, + 892, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 511, + 681, + 720, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training. We compute the matching cost using the Hungarian matching algorithm [17], thereby obtaining the query indices that best match the ground truth within the query set. Then we compute the loss between the corresponding queries and the ground truth across three aspects: object category loss $\\mathcal{L}_{obj}$ , paired entity bounding box loss $\\mathcal{L}_{box}$ and predicate category loss $\\mathcal{L}_{pred}$ . The first two aspects follow the DETR [2] framework. Besides, we follow [11] to apply an $\\mathcal{L}_1$ loss between the predicate and image embeddings to align the visual and textual features. The overall loss function is formulated as:", + "bbox": [ + 511, + 702, + 893, + 862 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {o b j} \\mathcal {L} _ {o b j} + \\lambda_ {b o x} \\mathcal {L} _ {b o x} + \\lambda_ {p r e d} \\mathcal {L} _ {p r e d} + \\lambda_ {l _ {1}} \\mathcal {L} _ {1}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 875, + 893, + 891 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "17733", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{obj},\\lambda_{box},\\lambda_{pred}$ and $\\lambda_{l_1}$ are hyper-parameters.", + "bbox": [ + 119, + 103, + 467, + 117 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During inference, the score of object category $\\mathbf{s}^o$ is simply the top-1 score from the softmax distribution over objects, and the triplet score $(\\mathbf{s}^{triA},\\mathbf{s}^{triS},\\mathbf{s}^{triC})$ of each predicate subclass is obtained using sigmoid function. Subsequently, we calculate the predicate score. For the $i$ -th query and the $j$ -th predicate class, we take the maximum score among the triplet scores belonging to the predicate subclasses associated with that predicate class to obtain the score for the $j$ -th predicate class.", + "bbox": [ + 99, + 118, + 483, + 250 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {s} _ {i j} ^ {p r e d A} = \\operatorname {M a x} \\left\\{\\mathbf {s} _ {i 1} ^ {t r i A}, \\mathbf {s} _ {i 2} ^ {t r i A}, \\dots , \\mathbf {s} _ {i k} ^ {t r i A} \\right\\}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 258, + 483, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $1,2,\\ldots,k$ are the score indices in the triplet scores of the $j$ -th query. Finally, we compute $\\mathbf{s}_i^o*\\mathbf{s}_{ij}^{predA}$ as the final score for the $j$ -th predicate of the $i$ -th query to generate the scene graph.", + "bbox": [ + 99, + 287, + 483, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 101, + 359, + 230, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 101, + 383, + 295, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluated our method on the Action Genome dataset [13], which is a large-scale video scene graph dataset built on the Charades dataset [34]. This dataset covers 234,253 frames from 9,848 videos, with a total of 476,229 object bounding boxes and 1,715,568 relationship entities. There are 36 entity categories and 26 predicate categories in these annotations. Predicates are further divided into three categories: attention, spatial, and contacting.", + "bbox": [ + 99, + 404, + 483, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluated the Scene Graph Detection (SGDet) and Predicate Classification (PredCls) task on the AG dataset. SGDet aims to detect entities and predict predicates without labels. PredCls is a relatively simple task, focused on predicting predicates based on given oracle detection results. We primarily evaluate the performance of SGDet task. Following previous works, we adopt Recall@k and mean Recall@K as evaluation metric to measure the ratio of correct entities among the top K predicted entities with the highest confidence $(\\mathrm{K} = 10,20,50)$ . To make a sufficient comparison with the baseline, the experiment was conducted under two settings: With Constraint and No Constraints. With Constraint only selects one predicate for each entity pair. No Constraints allows multiple predictions of the relationship between each entity pair.", + "bbox": [ + 99, + 522, + 483, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Details. We employ ResNet-50 as the CNN backbone and The number of layers for the Transformer encoder and decoder are set to 6 and 3, respectively. For training our model, we initialize the network with the parameters of DETR trained with the COCO dataset. ARN train for 20 epochs using the AdamW optimizer [26] with the batch size 8, initial learning rate of $1e^{-4}$ of the backbone. For each decoder, we set the number of queries $N_{q} = 100$ and use fine-tune the CLIP text embeddings initialized classifier with a small learning rate of $1e^{-5}$ . Moreover, the last layer", + "bbox": [ + 99, + 742, + 483, + 888 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fa4b47066d415b7b387a77857f5cb6054d444c2b98ab80b49c817f3b766d760b.jpg", + "image_caption": [ + "Figure 4. Comparative per predicate class performance for SGDET task. Results are in terms of mR@10 under \"No constraint\" setup. Predicates are arranged from left to right in order from head to tail." + ], + "image_footnote": [], + "bbox": [ + 517, + 101, + 893, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of the predicate decoder is our proposed predicate association parsing module. The hyper-parameters for the Hungarian costs $\\lambda_{obj}$ , $\\lambda_{pred}$ , $\\lambda_{box}$ , $\\lambda_{l_1}$ , and the loss weights of them are set to 1, 2, 2.5, 20, respectively. We conduct our experiments on one NVIDIA 4090 GPU.", + "bbox": [ + 511, + 396, + 893, + 469 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison to state-of-the-art", + "text_level": 1, + "bbox": [ + 511, + 482, + 777, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To evaluate the performance of our model, we compare it with several state-of-the-art DSGG approaches [6, 10, 21, 23, 27, 29, 36, 37, 39, 40]. Tab. 1 shows the comparative results for SGDET task in terms of R@K under both No constraints and With Constraint settings.", + "bbox": [ + 511, + 505, + 893, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In SGDet task, our method outperforms all existing two stage and one stage methods, with ARN improves the performance of the second-best method by $2.3\\%$ , $2.4\\%$ , and $2.3\\%$ under R@K metric of No constraints setting, respectively. Under the With Constraint setting, ARN improves performance by $1.6\\%$ on R@10 and $0.9\\%$ on R@20 compared to the current best method, however, it is slightly lower than the current best method on R@50. We conjecture the reason is as follows: In triplet prediction, ARN may generate more semantically similar results, and under the With Constraints setting, redundant results may occupy the top-ranking positions, thereby affecting the results in this setting. However, the results become reliable with fewer predictions when $K = [10,20]$ .", + "bbox": [ + 509, + 580, + 893, + 785 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluated the mR@K metric for each predicate category under No constraints setting, selecting several representative methods for comparison, as shown in Fig. 4. Our method achieves the best performance across most predicate categories, with varying shades of green indicate instances where other methods slightly outperform ours in a few categories. This demonstrates that our proposed ARN", + "bbox": [ + 511, + 786, + 893, + 888 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "17734", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/61b2bf1f5e181df3bf5df37a79f9e17ccf8c7dfb5a18a3ca5b4a0e3fbf6a8601.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodWith ConstraintNo Constraint
R@10R@20R@50mR@10mR@20mR@50R@10R@20R@50mR@10mR@20mR@50
Two Stage Method
VRD[27]19.224.526.0---19.128.840.5---
GPS-Net[23]24.733.135.1---24.435.747.3---
TRACE[36]13.914.514.58.28.28.226.535.645.322.831.341.8
STTran[6]25.234.137.016.620.822.224.636.248.820.929.739.2
APT[21]26.336.138.3---25.737.950.1---
STTran-TPI[40]26.234.637.415.620.221.8------
TR^2[39]26.835.538.3---27.839.250.0---
TEMPURA[29]28.133.434.918.522.623.729.838.146.424.733.943.7
TD^2[24]28.7-37.120.4-26.130.5-49.327.9-46.3
DSG-DETR[10]30.334.836.116.419.320.132.140.948.321.328.1-
OED[37]33.540.948.921.527.433.435.344.451.827.240.550.4
One Stage Method
RelTR[7]19.723.425.9---20.924.628.2---
OED*[37]31.537.743.720.626.130.233.441.349.026.138.547.7
Ours35.141.847.223.229.033.637.646.854.129.741.951.9
", + "bbox": [ + 106, + 99, + 893, + 330 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee48c0e481e2e5d3cd57f055b430c43ee5c77c5a5d36539072a9efb4e1ff9ea6.jpg", + "image_caption": [ + "Figure 5. Ablation results for per class performance in SGDET task. Results are in terms of mR@50 under \"No constraint\"." + ], + "image_footnote": [], + "bbox": [ + 102, + 393, + 480, + 539 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/618420f2be9cd9ebf25e96d458de146469e5d9705655166d0c85c818fe7f4e13.jpg", + "table_caption": [ + "Table 1. Comparative results for SGDet task, on Action Genome dataset [13], in terms of Recall@K and mean Recall@K metric, best and second best results under each setting are respectively marked in bold and underline. where * indicates end-to-end training results." + ], + "table_footnote": [], + "table_body": "
MethodWith ConstraintNo ConstraintFPS#Params(M)
R@10R@20R@50R@10R@20R@50
TRACE[36]27.527.527.572.691.696.4--
STTran[6]68.671.871.877.994.299.1-151.03
TEMPURA[29]68.871.571.580.494.299.4-248.8
TD2'2[24]70.1-73.181.7-98.8--
OED[37]73.076.176.183.395.399.212154.54
Ours74.674.674.683.796.799.916254.06
", + "bbox": [ + 104, + 590, + 482, + 665 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison results for PredCls task, on Action Genome dataset [13], in terms of Recall@K metric. Best and second best results under each setting are respectively marked in bold and underline.", + "bbox": [ + 99, + 675, + 482, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "framework effectively mitigates the long-tail problem, improving accuracy on tail predicates while maintaining performance on head predicates.", + "bbox": [ + 99, + 756, + 482, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In PredCls task, as shown in Tab. 2, ARN imporves the performance of the second-best methods by $0.4\\%$ , $1.4\\%$ , $0.5\\%$ on the R@K metric under the No Constraints setting. It also achieves comparable performance under the With Constraints setting. Although ARN is designed for SGDet task, its performance on the PredCls task also demonstrates", + "bbox": [ + 99, + 801, + 482, + 888 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "itsversatilityacrossmultipletasks.", + "bbox": [ + 511, + 393, + 740, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, as shown in Tab. 2, our model exhibits a significant advantage in parameter efficiency compared to two-stage methods [6, 29]. In terms of inference speed, it surpasses the end-to-end algorithm OED. Notably, OED relies on adjacent frames during inference, requiring the processing of more images. Our experiments show that our method achieves an inference time of 8 minutes per round, whereas OED requires 22.7 minutes per round.", + "bbox": [ + 511, + 410, + 893, + 527 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablative Study", + "text_level": 1, + "bbox": [ + 511, + 540, + 660, + 556 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We proposed two modules in our ARN, Predicate Association Parsing Module (PAP) and Hierarchical Attention (HA), and performed ablation studies on different components to clarify their contributions to performance.", + "bbox": [ + 511, + 564, + 893, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Association Parsing Network. We evaluate the effectiveness of several components in the proposed ARN framework. We extend DETR for DSGG and establish it as our baseline. We choose focal loss as the loss function for predicate prediction and compare it with cross-entropy loss for fairness, as shown in the first and second rows of Tab. 3. The results indicate that focal loss is better suited for DSGG. We leverage CLIP to extend predicate prediction to triplet prediction, using CLIP's semantic knowledge to guide the model in learning more robust triplet semantic context. As demonstrated in the third row of Tab. 3, this is an effective approach for enhancing scene understanding.", + "bbox": [ + 511, + 623, + 895, + 799 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Predicate Association Parsing Module. We examine the impact of the proposed Predicate Association Parsing Module on model performance, as shown in the fourth row of Tab. 3. Without incorporating HA, associating fine-grained paired entity features enhances the model's performance, demonstrating the positive role of visual relation", + "bbox": [ + 511, + 801, + 895, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "17735", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0bde642e36a35b094ba0adf480ca0cc0d3f996d3508d568445d3a001ee897cac.jpg", + "image_caption": [ + "Figure 6. Comparative qualitative results. From left to right: input video frames, dynamic scene graphs generated by STTran [6], TEMPURA [29], OED [37] and ours. Incorrect and correct predicate predictions are shown in green and red, respectively." + ], + "image_footnote": [], + "bbox": [ + 112, + 99, + 890, + 344 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/fd0dc09adc97c03b627f4d9508284d96be253db8a8a53f4cbb496543e4e10d51.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Baseline\\( \\mathcal{L}_{focal} \\)CLIPPAPHANo Constraint
R@10R@50mR@10mR@50
✓28.444.319.639.3
✓✓31.745.425.644.8
✓✓✓34.451.727.148.5
✓✓✓✓35.352.227.949.8
✓✓✓✓✓37.654.129.751.9
", + "bbox": [ + 106, + 405, + 483, + 481 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/56b8f7a9e06d64a36cfd2ee45a1211197fcc40bbd02b0e7fec575a29a80c4848.jpg", + "table_caption": [ + "Table 3. Ablative studies on our framework for SGDet task." + ], + "table_footnote": [], + "table_body": "
MethodNo Constraint
R@10R@50mR@10mR@50
Model I35.051.527.348.9
Model II36.953.029.351.4
Ours37.654.129.751.9
", + "bbox": [ + 107, + 517, + 483, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablative studies on Conditional weight mapping.", + "bbox": [ + 124, + 613, + 459, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dependencies between entities and predicates in predicate prediction. With Hierarchical Attention included, the results further improve (fifth row of Tab. 3), indicating that visual relation dependencies between predicate subclasses also facilitate a broader understanding of interactive context, validating our motivation.", + "bbox": [ + 99, + 654, + 482, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conditional weight mapping. We evaluate the effectiveness of the proposed conditional weight mapping mechanism for association reasoning. Following the cross-attention structure, we directly add paired entity representations to the predicate query, denoted as Model I. Additionally, we experiment with handling features from different sources through nonlinear mapping, denoted as Model II. As observed in the first and second rows of Tab. 4, introducing entity information without proper organization negatively impacts the original performance of ARN. Nonlinear", + "bbox": [ + 99, + 742, + 483, + 888 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mapping yields some improvement, but conditional weight mapping provides a more elegant approach to achieving effective association reasoning.", + "bbox": [ + 511, + 409, + 893, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More analyze. To further analyze the impact of the proposed modules, we present the average recall for each predicate class in the SGDet task, as shown in Fig. 5. It can be observed that after removing the PAP and HA modules, the model's performance drops in $92\\%$ of predicate categories, with only slight improvements in the \"standing on\" and \"wearing\" categories $(0.48\\%)$ and $0.17\\%$ , respectively).", + "bbox": [ + 511, + 453, + 893, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 565, + 684, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Fig. 6 presents the qualitative results in SGDet task. The results show that existing two-stage methods (i.e., STTran, TEMPURA) are clearly at a disadvantage in entity detection (erroneous detections result in incorrect triplets). In contrast, one-stage methods excel in entity detection. Additionally, ARN accurately detects more tail predicates, demonstrating a broader capability for scene understanding.", + "bbox": [ + 511, + 585, + 893, + 689 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 703, + 627, + 717 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN is the first to model reasoning over visual relation dependencies. We extend predicate prediction to triplet prediction and introduce a Predicate Association Parsing Module (PAP) to organize and interpret multiple predicate subclasses and fine-grained entity representations. Additionally, we design a Hierarchical Attention (HA) mechanism to effectively aggregate multi-source information for association reasoning. Extensive experiments on the Action Genome dataset demonstrate the effectiveness of our method.", + "bbox": [ + 509, + 728, + 893, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "17736", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China (62171328,62171327,62072350); in part by the Key Program of Hubei Provincial Natural Science Foundation (Xiangyang Innovation and Development Joint Fund) under Grand (2025AFD050); by the Central Government's Special Project for Guiding Local Development in Hubei Province (2022BGE242); by the Hubei Science and technology innovation team (T2023009); by the National Natural Science Foundation of Hubei (2023AFB158).", + "bbox": [ + 99, + 102, + 485, + 250 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 102, + 262, + 194, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3286-3295, 2019. 3", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229, 2020. 2, 3, 5", + "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9004-9013, 2021. 2, 3", + "[4] Siqi Chen, Jun Xiao, and Long Chen. Video scene graph generation from single-frame weak supervision. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2", + "[5] Anoop Cherian, Chiori Hori, Tim K Marks, and Jonathan Le Roux. $(2.5 + 1)$ d spatio-temporal scene graphs for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 444-453, 2022. 1", + "[6] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16372-16382, 2021. 2, 6, 7, 8", + "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 7, 1", + "[8] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2150-2159, 2021. 2", + "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjm Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 88-98, 2023. 1", + "[10] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term de" + ], + "bbox": [ + 102, + 286, + 483, + 888 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "pendencies for generating dynamic scene graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5130-5139, 2023. 6, 7", + "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Eleventh International Conference on Learning Representations (ICLR), 2021. 5", + "[12] Xia Hua, Xinqing Wang, Ting Rui, Faming Shao, and Dong Wang. Adversarial reinforcement learning with object-scene relational graph for video captioning. IEEE Transactions on Image Processing (TIP), 31:2004-2016, 2022. 1", + "[13] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10236–10247, 2020. 1, 4, 6, 7", + "[14] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668-3678, 2015. 2", + "[15] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debiasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2516-2526, 2024. 2", + "[16] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision (IJCV), 123:32-73, 2017. 2", + "[17] Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics quarterly (NRL), 2(1-2): 83-97, 1955. 5", + "[18] Stan Weixian Lei, Difei Gao, Jay Zhangjie Wu, Yuxuan Wang, Wei Liu, Mengmi Zhang, and Mike Zheng Shou. Symbolic replay: Scene graph as prompt for continual learning on vqa task. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1250-1259, 2023. 1", + "[19] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28369-28379, 2024. 2", + "[20] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19486-19496, 2022. 2", + "[21] Yiming Li, Xiaoshan Yang, and Changsheng Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13874-13883, 2022. 6, 7", + "[22] T Lin. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2998, 2017. 4" + ], + "bbox": [ + 514, + 103, + 893, + 888 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "17737", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3746-3753, 2020. 2, 6, 7", + "[24] Xin Lin, Chong Shi, Yibing Zhan, Zuopeng Yang, Yaqi Wu, and Dacheng Tao. $\\mathrm{Td}^2$ -net: Toward denoising and debiasing for video scene graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3495-3503, 2024. 2, 7, 1", + "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2", + "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Eleventh International Conference on Learning Representations (ICLR), 2017. 6", + "[27] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In European Conference on Computer Vision (ECCV), pages 852–869, 2016. 2, 6, 7", + "[28] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2", + "[29] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22803-22813, 2023. 2, 6, 7, 8, 1", + "[30] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2, 3", + "[31] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for video scene graph generation. IEEE Transactions on Image Processing (TIP), 33:556-568, 2023. 2, 1", + "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 3", + "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 2", + "[34] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, 2016. 6", + "[35] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detec" + ], + "bbox": [ + 102, + 103, + 483, + 888 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10410-10419, 2021. 2", + "[36] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13688-13697, 2021. 2, 6, 7", + "[37] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27938-27947, 2024. 3, 6, 7, 8, 1", + "[38] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cross-modal graph with meta concepts for video captioning. IEEE Transactions on Image Processing (TIP), 31: 5150-5162, 2022. 1", + "[39] Jingyi Wang, Jina Huang, Can Zhang, and Zhidong Deng. Cross-modality time-variant relation learning for generating dynamic scene graphs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8231–8238. IEEE, 2023. 6, 7", + "[40] Shuang Wang, Lianli Gao, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, and Jingkuan Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 5793-5801, 2022. 6, 7", + "[41] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2", + "[42] Xu Yang, Chongyang Gao, Hanwang Zhang, and Jianfei Cai. Hierarchical scene graph encoder-decoder for image paragraph captioning. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 4181-4189, 2020. 1", + "[43] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1556-1565, 2020. 1", + "[44] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5831-5840, 2018. 2", + "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2" + ], + "bbox": [ + 514, + 103, + 893, + 813 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "17738", + "bbox": [ + 480, + 926, + 517, + 938 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_model.json b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_model.json new file mode 100644 index 0000000000000000000000000000000000000000..35601e47bbaa826a67b2b62238b6e64ecd9f9847 --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_model.json @@ -0,0 +1,2123 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.117, + 0.017, + 0.191, + 0.056 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.246, + 0.015, + 0.8, + 0.058 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.14, + 0.885, + 0.183 + ], + "angle": 0, + "content": "End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.211, + 0.79, + 0.314 + ], + "angle": 0, + "content": "Liwei Wang\\(^{1}\\), Yanduo Zhang\\(^{1,2\\text{串}}\\), Tao Lu\\(^{1}\\), Fang Liu\\(^{1}\\), Huiqin Zhang\\(^{1}\\), Jiayi Ma\\(^{3}\\), Huabing Zhou\\(^{1\\text{串}}\\) \n\\(^{1}\\)Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology \n\\(^{2}\\)Hubei University of Arts and Science \\(^{3}\\)Wuhan University \nwlw951226@163.com, {ydzhang,lut}@wit.edu.cn, {fangliu8822,666zhanghuiqin,jyma2010,Zhouhuabing}@gmail.com" + }, + { + "type": "title", + "bbox": [ + 0.256, + 0.349, + 0.331, + 0.364 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.381, + 0.484, + 0.733 + ], + "angle": 0, + "content": "Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets \\(<\\) subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in https://github.com/wlw951226/ARN." + }, + { + "type": "title", + "bbox": [ + 0.103, + 0.764, + 0.231, + 0.78 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.79, + 0.484, + 0.864 + ], + "angle": 0, + "content": "Scene graph generation is a promising approach to scene understanding and has emerged as a frontier topic. Dynamic Scene Graph Generation (DSGG) advances this by capturing dynamic visual relationships in videos, contributing to a comprehensive understanding of video content. DSGG has" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.349, + 0.891, + 0.58 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.59, + 0.896, + 0.658 + ], + "angle": 0, + "content": "Figure 1. (a) The association frequency matrix between entities and predicates in the dataset [13]; (b) the association frequency matrix between predicate subclasses, where the vertical axis represents the pairwise combinations of two predicate subclasses; (c) an example and the pipeline of our associative reasoning network." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.684, + 0.895, + 0.756 + ], + "angle": 0, + "content": "also opened up new possibilities in areas such as video analysis [9, 18, 42], video captioning [12, 38], and video question answering [5, 43], enabling computers to better simulate human perception and cognitive processes in real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.758, + 0.896, + 0.891 + ], + "angle": 0, + "content": "Dynamic scene graphs encode implicitly co-occurrences and affordance relations among entities and predicates. We analyze the frequency co-occurrence of predicates and entities in the Action Genome [13] dataset. As shown in Fig. 1(a), there is a strong association between certain predicates and entities; for instance, the predicate \"writing on\" often implies that the object is a notebook. Similarly, when the object is a notebook, it is likely in front of a person who is looking at it. Fig. 1(b) highlights the association be" + }, + { + "type": "page_footnote", + "bbox": [ + 0.125, + 0.877, + 0.265, + 0.89 + ], + "angle": 0, + "content": "Corresponding authors." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "17729" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.482, + 0.206 + ], + "angle": 0, + "content": "tween different predicates, and Fig. 1(c) provides an example demonstrating that holding an object strongly suggests its position in front of the person. However, existing methods overlook the dependencies between entities and predicates. Modeling these visual dependencies enables a multiperspective understanding of scenes, resulting in more accurate scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.207, + 0.482, + 0.337 + ], + "angle": 0, + "content": "In addition, the currently prevalent DSGG paradigm [4, 6, 15, 29, 31, 36] are built upon a two-stage pipeline, focusing on capturing the temporal dependencies of videos. This two-stage paradigm independently optimizes the subtasks of entity localization [33] and predicate prediction, hindering end-to-end learning. Additionally, these methods generate redundant candidate boxes and predict relationships for all detected entity pairs, increasing computational cost." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.341, + 0.482, + 0.457 + ], + "angle": 0, + "content": "Recently, the one-stage detector DETR [2] has transformed the pattern of traditional detection frameworks by adopting a set-based prediction approach for one-to-one matching, eliminating the need for computationally expensive hand-designed components such as non-maximum suppression. Meanwhile, DETR-based methods [3, 25, 30, 35, 41] have demonstrated remarkable performance in image understanding." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.46, + 0.482, + 0.62 + ], + "angle": 0, + "content": "In this paper, we extend DETR to DSGG and propose an end-to-end Association Reasoning Network (ARN). ARN explores visual relation dependencies within scene graphs and performs association reasoning with fine-grained entity context. ARN leverages Contrastive Language-Image Pretraining (CLIP) [32] to bridge vision and language, modeling triplet cues through textual prompts, and extending predicate prediction to triplet prediction. This approach enriches the semantic representation of predicate prediction while ensuring semantic consistency between paired entities and predicates in DSGG." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.622, + 0.482, + 0.812 + ], + "angle": 0, + "content": "We further design a Predicate Association Parsing (PAP) module to link fine-grained representations from paired entities and different predicate subclasses. It employs a conditional weight mapping mechanism to dynamically organize predicate representations across different branches. Additionally, we develop a Hierarchical Attention (HA) mechanism that integrates spatio-temporal context with complementary predicate and entity cues in a structured hierarchy. This method achieves association reasoning between entities and predicates while maintaining stable performance during end-to-end training. Our experiments on the widely used Action Genome dataset demonstrate that ARN achieves state-of-the-art performance in DSGG." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.815, + 0.482, + 0.843 + ], + "angle": 0, + "content": "The contributions of this paper can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.846, + 0.482, + 0.89 + ], + "angle": 0, + "content": "- We propose an end-to-end ARN model for DSGG, which explores visual relation dependencies between entities and predicates and performs association reasoning to gen" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.104, + 0.708, + 0.117 + ], + "angle": 0, + "content": "erate accurate scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.118, + 0.892, + 0.16 + ], + "angle": 0, + "content": "- We introduce a Predicate Association Parsing (PAP) module to associate and organize fine-grained representations of different predicate subclasses and paired entities." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.162, + 0.894, + 0.219 + ], + "angle": 0, + "content": "- We designed a Hierarchical Attention (HA) module to hierarchically perform visual relation association reasoning between paired entity representations and predicate relational context." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.221, + 0.892, + 0.249 + ], + "angle": 0, + "content": "- We conducted a comprehensive evaluation of the proposed method, demonstrating its superior performance." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.118, + 0.894, + 0.249 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.263, + 0.65, + 0.278 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.288, + 0.895, + 0.536 + ], + "angle": 0, + "content": "Image scene graph generation. Image scene graph generation aims to identify entities and predict the predicates of image in form of triplets, which initially proposed by [14]. Most early works [8, 19, 23, 27, 44, 45], focused on refining entity and visual relationship representations in images using various network architectures. Subsequently, several works have attempted to design end-to-end image SGG models. Cong et al. [7] treated SGG as a set prediction problem and proposed an end-to-end image SGG model that directly predicts sparse scene graphs based solely on visual appearance. Furthermore, Li et al. [20] designed a graph assembling module that infers the connectivity of a bipartite graph based on entities features, generating scene graphs through graph matching. Meanwhile, some researchers [28] have focused on addressing the long-tailed distribution of predicates in benchmark datasets [16] to generate unbiased scene graphs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.894, + 0.786 + ], + "angle": 0, + "content": "Dynamic scene graph generation. The dynamic visual relationships in videos are more complex and have temporal dependencies, making DSGG more challenging compared to image scene graph generation. Previous works have focused on capturing temporal dependencies in videos and addressing data imbalance caused by long-tail distributions. Some works focus on designing more efficient temporal dependencies reasoning networks. Cong et al. [6] employed an encoder-decoder architecture to extract spatiotemporal context from video frames and proposed a two-stage baseline. Building on this, Pu et al. [31] injected prior spatio-temporal knowledge into DSGG via statistical co-occurrence, proposing a spatio-temporal Knowledge-Embedded Transformer. However, this method heavily relies on the inherent distribution of the data, such as the long-tail phenomenon. Differently, ARN does not rely on statistical priors." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.787, + 0.894, + 0.889 + ], + "angle": 0, + "content": "To overcome the limitations of the data's inherent distribution, Nag et al. [29] proposed a memory-guided training method to alleviate the long-tailed distribution. Lin et al. [24] introduced an asymmetric re-weighting loss that relieve the issue of label bias by adjusting the weights of tail predicates. Despite some progress, these methods sacrifice the performance of head predicates to focus on tail" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17730" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.484, + 0.146 + ], + "angle": 0, + "content": "predicates. Additionally, they overlook the visual relational dependencies within the scene graph, which encompass a broader understanding of interaction context." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.147, + 0.484, + 0.265 + ], + "angle": 0, + "content": "The most relevant work to ours is OED [37], which introduces a progressive refined module to aggregate temporal context. But this module relies on independently optimized spatial context, resulting in degraded performance during end-to-end training. In contrast to OED, we focus on exploring visual dependencies between entities to enhance holistic scene understanding and enabling fully end-to-end implementation." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.277, + 0.192, + 0.292 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.301, + 0.485, + 0.537 + ], + "angle": 0, + "content": "In this work, we extend the end-to-end detector DETR[2] to DSGG and propose an Association Reasoning Network (ARN), as illustrated in Fig. 2. ARN sequentially extracts visual features and feeds them into two cascaded decoder modules, the Entity Decoder and the Predicate Decoder, to perform two subtasks: entity detection and triplet prediction (Sec 3.1). Subsequently, we introduce triplet cues modeling using CLIP's semantic knowledge (Sec 3.2). Moreover, the Predicate Association Parsing (PAP) Module in predicate decoder organizes predicate cues from different branches and fine-grained paired entity features (Sec 3.3). The Hierarchical Attention (HA) mechanism then aggregates these with spatio-temporal features to perform association reasoning, enhancing broader contextual understanding of video scenes (Sec 3.4). Finally, we introduce the training loss and inference process of ARN (Sec 3.5)." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.545, + 0.295, + 0.56 + ], + "angle": 0, + "content": "3.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.566, + 0.484, + 0.61 + ], + "angle": 0, + "content": "The overall architecture of our ARN is illustrated in Fig. 2. The ARN is composed of a Backbone, an Encoder, an Entity decoder and a Predicate Decoder." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.611, + 0.483, + 0.64 + ], + "angle": 0, + "content": "**Backbone.** CNN-based backbone maps an input frame \\(I_{i}\\) to a spatial feature map \\(\\mathbf{f}_i\\in \\mathbb{R}^{H\\times W\\times C}\\)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.641, + 0.484, + 0.744 + ], + "angle": 0, + "content": "Encoder. The transformer-based encoder [2, 3, 30] refines image features and captures spatial contextual information. First, a \\(1 \\times 1\\) convolution and flatten operation are applied to reduce the dimension and transform it into a serialized feature representation \\(\\mathbf{f}_i \\in \\mathbb{R}^{HW \\times d}\\). Next, we feed the \\(\\mathbf{f}_i\\) and fixed positional encodings \\(\\mathbf{p}_e\\) [1] to the encoder and output the enhanced spatial feature \\(\\mathbf{F}_{vi} \\in \\mathbb{R}^{HW \\times d}\\)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.744, + 0.484, + 0.831 + ], + "angle": 0, + "content": "Entity Decoder. Entity decoder consists of multiple transformer decoder layers and several FFN heads and introduces an entity query set \\(\\mathbf{Q}^E\\in \\mathbb{R}^{N_q\\times d}\\), where \\(N_{q}\\) is the number of queries. Unlike previous methods, our query set has two subsets \\(\\mathbf{Q}^E = [\\mathbf{Q}^s,\\mathbf{Q}^o ]\\in \\mathbb{R}^{N_q\\times d}\\) that aim to separate the subject and the object." + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.832, + 0.484, + 0.891 + ], + "angle": 0, + "content": "Moreover, a pair of learnable entity position embeddings \\(\\{\\mathbf{e}^s,\\mathbf{e}^o\\} \\in \\mathbb{R}^{N_q\\times d}\\) is designed to assign the subject and object at the same position to form pairwise entities. Furthermore, \\(\\mathbf{e}^o\\) is employed to align the query sets for the two" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.894, + 0.133 + ], + "angle": 0, + "content": "decoders and update them during training. Overall, the implementation of the entity decoder layer is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.145, + 0.895, + 0.163 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {E} = \\operatorname {C o n c a t} \\left[ \\mathbf {Q} ^ {s} + \\mathbf {e} ^ {s}, \\mathbf {Q} ^ {o} + \\mathbf {e} ^ {o} \\right], \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.18, + 0.895, + 0.2 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {Q}} _ {(l)} ^ {E} = \\operatorname {D e c} _ {(l)} ^ {E} \\left(\\mathbf {Q} ^ {E}, \\mathbf {F} _ {v i}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.207, + 0.895, + 0.301 + ], + "angle": 0, + "content": "Specifically, at each layer \\(\\mathrm{Dec}_{(l)}^{E}\\), the queries first undergoes self-attention to enable interactions among queries. Subsequently, cross-attention aggregates visual features \\(\\mathbf{F}_{vi}\\) into the queries to obtain updated queries \\(\\tilde{\\mathbf{Q}}_{(l)}^{E}\\). The output of the entity decoder, denoted as \\(\\mathbf{F}^{E}\\), is derived from the final updated queries \\(\\tilde{\\mathbf{Q}}_{(l)}^{E}\\) after \\(l\\) layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.302, + 0.894, + 0.375 + ], + "angle": 0, + "content": "Finally, the entity representation \\(\\mathbf{F}^E\\) passes through the FFN head and object classifier at the top of the entity decoder to obtain the bounding box coordinates \\(\\mathbf{B}^{E} = \\{\\mathbf{b}^{s},\\mathbf{b}^{o}\\} \\in \\mathbb{R}^{N_{q}\\times 4}\\) and object categories \\(\\mathbf{C}^o\\in \\mathbb{R}^{N_q}\\) for the paired entities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.894, + 0.463 + ], + "angle": 0, + "content": "Predicate Decoder The predicate decoder is tailored to capture temporal dependencies and facilitate associative reasoning between entities and multiple predicate subclasses. In detail, the structure of the predicate decoder is a multi-branch decoder architecture, each branch corresponding to the prediction of a predicate subclass." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.894, + 0.581 + ], + "angle": 0, + "content": "Each branch consists of multiple decoder layers and initializes predicate queries \\(\\mathbf{Q}^{predi} \\in \\mathbb{R}^{N_q \\times d}\\) in \\(i\\)-th branch. Unlike the entity decoder, we capture temporal dependencies in visual features \\(\\mathbf{F}_{vi}\\). Specifically, we initialized a learnable frame encoding \\(\\mathbf{p}_t \\in \\mathbb{R}^{N_q \\times t \\times d}\\) to identify different frames. Subsequently, we transpose the \\(\\mathbf{F}_{vi}\\) in the temporal dimension and update it by self-attention (SA) operation to obtain the spatio-temporal context feature \\(\\mathbf{F}_{vit}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.591, + 0.894, + 0.608 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {v i t} = S A (\\text {t r a n s p o s e} (\\mathbf {F} _ {v i})), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.617, + 0.894, + 0.736 + ], + "angle": 0, + "content": "Ultimately, \\(\\mathbf{Q}^{predi}\\) aggregates the \\(\\mathbf{F}_{vit}\\) through crossattention to generate fine-grained triplet representations \\(\\mathbf{F}^{predi}\\). Furthermore, to encourage associative reasoning across predicates, we design a predicate association embedding \\(\\mathbf{e}^{predi} \\in \\mathbb{R}^{N_q \\times d}\\), which is dynamically updated along with the \\(\\mathbf{Q}^{predi}\\) in the predicate decoder. The decoding process of the predicate representation in the \\(l\\)-th layer of predicate decoder \\(Dec_l^{predi}\\) is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.632, + 0.747, + 0.895, + 0.765 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {e}} ^ {p r e d i} = \\mathbf {e} ^ {o} + \\mathbf {e} ^ {p r e d i}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.795, + 0.895, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {(l)} ^ {p r e d i} = \\operatorname {D e c} _ {(l)} ^ {p r e d i} \\left(\\mathbf {Q} ^ {p r e d i}, \\tilde {\\mathbf {e}} ^ {p r e d i}, \\mathbf {F} _ {v i t}, \\mathbf {p} _ {t}\\right), \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.783, + 0.84 + ], + "angle": 0, + "content": "3.2. Semantic Triplet Cue Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.895, + 0.89 + ], + "angle": 0, + "content": "We use CLIP [32] to extending traditional predicate prediction to triplet prediction and inject domain-agnostic prior knowledge into our ARN." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17731" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.102, + 0.856, + 0.405 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.1, + 0.417, + 0.897, + 0.485 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed Association Reasoning Network (ARN), which first extracts visual features from the video, and the entity decoder is used to locate entities. Subsequently, the multi-branch predicate decoder captures the temporal dependencies in visual features and infers representations for each predicate subclasses. Finally, the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) associates different predicate subclasses cues and aggregates fine-grained entity features to refine the prediction results." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.511, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Specifically, we follow the ActionGenome [13] setting and divide predicates into three subclasses: Attention, Spatial, and Contact. Then we design distinct textual prompts for each predicate subclass to build a bridge for interaction between CLIP and visual triplets :(1) 'The person is [attention predicate] at the [object]'; (2) 'The [object] is [spatial predicate] the person'; (3) 'a photo of a person maybe [contacting predicate] the [object]', an example is illustrated on the right side of Fig. 2. Next, we use the CLIP text encoder to extract text embeddings \\(\\mathbf{E}^{predi} \\in \\mathbb{R}^{N_{predi} \\times d}\\) for each predicate subclass triplets, where \\(\\mathbf{E}^{predi}\\) serves as semantic cues for the visual triplets, \\(N_{predi}\\) denotes the number of triplet categories in \\(i\\)-th predicate decoder branch." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.737, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Based on this, we use the \\(\\mathbf{E}^{predi}\\) to initialize the classifier weights \\(w^{predi}\\). For the \\(n\\)-th query in \\(i\\)-th branch, it computes the similarity score \\(s_n^i\\) between the query representation and the triplet embeddings to align the visual and semantic information, then apply softmax activation to compute the focal loss [22], yielding the predicate prediction loss \\(\\mathcal{L}_{\\mathrm{pred}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.846, + 0.484, + 0.891 + ], + "angle": 0, + "content": "Moreover, for the entity decoder, we design textual prompts for object categories `a photo of person and a [object]''. We extract text embeddings \\( E^o \\in \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.896, + 0.556 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{N_{obj}\\times d}\\) in the same manner and use them to initialize the classifier weights \\(w^o\\) of the entity decoder. The object classifier is optimized using cross-entropy loss." + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.581, + 0.894, + 0.614 + ], + "angle": 0, + "content": "\\[\ns _ {n} ^ {i} = \\left[ \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {1} ^ {p r e d i}\\right), \\dots , \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {N _ {p r e d i}} ^ {p r e d i}\\right) \\right] \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.628, + 0.895, + 0.665 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r e d}} = \\frac {1}{N} \\sum_ {n \\in N _ {q}} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\sigma \\left(s _ {n} ^ {i}\\right), y _ {r} ^ {\\text {p r e d i}}\\right), \\tag {7}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.676, + 0.833, + 0.692 + ], + "angle": 0, + "content": "3.3. Predicate Association Parsing Module" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.698, + 0.896, + 0.829 + ], + "angle": 0, + "content": "As mentioned previously, the pairwise entities can be associated with multiple predicates, and these entity-predicate associative cues are crucial for DSGG. Therefore, we propose a Predicate Association Parsing (PAP) Module that first leverages a conditional weight mapping mechanism to differentiate and organize multiple predicate subclass cues, then performs associative reasoning by aggregating global spatio-temporal context and fine-grained entity representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.831, + 0.897, + 0.891 + ], + "angle": 0, + "content": "We denote the outputs of different branches in the predicate decoder as \\(\\mathbf{F}^{predA}\\), \\(\\mathbf{F}^{predS}\\), and \\(\\mathbf{F}^{predC}\\), using the initials of predicate subclasses for simplicity. For \\(\\mathbf{F}^{predA}\\), we treat \\(\\mathbf{F}^{predS}\\) and \\(\\mathbf{F}^{predC}\\) as complementary predicate" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "17732" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.106, + 0.328, + 0.12 + ], + "angle": 0, + "content": "Predicate Association Parsing Module" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.121, + 0.48, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.368, + 0.484, + 0.463 + ], + "angle": 0, + "content": "Figure 3. The architecture of the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) Mechanism. The PAP organizes complementary predicate cues and embeds entity context into the current predicate representation. Building on this, the HA aggregates spatio-temporal context, explores visual relation dependencies, and progressively refines the predicate representation." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.488, + 0.484, + 0.605 + ], + "angle": 0, + "content": "representations and use a linear layer to reduce the dimensionality of them. Then we concatenate them as complementary predicate cues for the current branch, denoted as \\(\\mathbf{F}^{compA} = \\mathrm{Concat}[\\mathbf{F}^{\\mathrm{predS}},\\mathbf{F}^{\\mathrm{predC}}]\\). During the dot product operation, the cross-attention weights consist of two components, balancing the contributions of different predicate subclasses and preventing the model from being biased toward any single predicate representation." + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.613, + 0.483, + 0.632 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} = \\left(\\mathbf {Q} _ {S}, \\mathbf {Q} _ {C}\\right) = \\left(\\mathbf {F} ^ {p r e d S}, \\mathbf {F} ^ {p r e d C}\\right), \\mathbf {K} = \\mathbf {F} _ {v i t} + \\mathbf {P} _ {t}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.653, + 0.483, + 0.671 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} ^ {T} \\mathbf {K} = \\mathbf {Q} _ {S} ^ {T} \\mathbf {K} + \\mathbf {Q} _ {C} ^ {T} \\mathbf {K}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.705 + ], + "angle": 0, + "content": "Next, we integrate the fine-grained paired entity features \\(\\mathbf{F}^E\\) into \\(\\mathbf{F}^{predA}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.727, + 0.483, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {f u s i o n} ^ {p r e d A} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left[ \\mathbf {F} ^ {p r e d A} + \\mathbf {e} ^ {p r e d A}, \\mathbf {F} ^ {E} \\right]\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.756, + 0.484, + 0.83 + ], + "angle": 0, + "content": "where \\(\\mathbf{e}^{predA}\\) denotes the predicate association embedding for attention predicates. Finally, we perform associative reasoning between the entity representation and various predicate representations through the proposed hierarchical attention module." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.839, + 0.307, + 0.854 + ], + "angle": 0, + "content": "3.4. Hierarchical Attention" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.86, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Visual relational dependencies are crucial for understanding scene context, yet traditional cross attention mechanisms" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.103, + 0.895, + 0.205 + ], + "angle": 0, + "content": "struggle to associate multiple semantic representations. To this end, we designed a Hierarchical Attention (HA) Mechanism, building upon the predicate association parsing module. The HA mechanism enables fine-grained integration of entity features with diverse predicate representations, facilitating associative reasoning and improving the prediction of tail predicates. The concept is illustrated in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.206, + 0.895, + 0.278 + ], + "angle": 0, + "content": "Specifically, HA is a quadruplet attention paradigm \\((\\mathbf{Q}_{cur},\\mathbf{Q}_{comp},\\mathbf{K},\\mathbf{V})\\), where \\(\\mathbf{Q}_{cur}\\) is the predicate subclass representation of the current branch, and \\(\\mathbf{Q}_{comp}\\) represents a complementary predicate cues, which is the cues representation of the other predicate subclasses." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.279, + 0.895, + 0.383 + ], + "angle": 0, + "content": "Taking the branch of the attention predicate subclass as an example, we perform the first-level cross-attention operation to associate complementary predicate cues \\(\\mathbf{F}^{compA}\\), where \\(\\mathbf{F}^{compA}\\) serves as the query \\(\\mathbf{Q}_{comp}\\), and \\(\\mathbf{F}_{vit}\\) is used as the key and value, \\(\\mathbf{e}^o\\) as the query pos. This process highlights regions of interest related to other predicate subclasses and updates \\(\\mathbf{Q}_{comp}\\) to obtain \\(\\mathbf{F}_{comp}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.405, + 0.894, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {c o m p}} = \\operatorname {A t t} _ {1} ^ {H A} \\left(\\mathbf {Q} _ {\\text {c o m p}}, \\mathbf {K} = \\mathbf {F} _ {\\text {v i t}} + \\mathbf {p} _ {t}, \\mathbf {V} = \\mathbf {F} _ {\\text {v i t}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.895, + 0.594 + ], + "angle": 0, + "content": "Second, we perform the second-level cross-attention operation, where \\(\\mathbf{F}^{predA}\\) serves as the query \\(\\mathbf{Q}_{cur}\\), \\(\\mathbf{Q}_{comp}\\) as the key and \\(\\mathbf{F}_{comp}\\) as the value. This operation first establishes the association between the current predicate subclass and complementary predicate cues, then propagates the spatio-temporal visual features to the query of the current subclass, highlighting the most relevant categories and updating the predicate representation \\(\\mathbf{F}^{predA}\\). Finally, the \\(\\mathbf{F}^{predA}\\) passes through the linear mapping layer and is added to the original input \\(\\mathbf{Q}_{cur}\\), yielding the final output \\(\\mathbf{F}_{assoc}^{predA}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.616, + 0.894, + 0.636 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} ^ {p r e d A} = \\operatorname {A t t} _ {2} ^ {H A} \\left(\\mathbf {Q} _ {c u r} = \\mathbf {F} ^ {p r e d A}, \\mathbf {Q} _ {c o m p}, \\mathbf {F} _ {s u p}\\right), \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.656, + 0.893, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {\\text {a s s o c}} ^ {\\text {p r e d} A} = \\operatorname {L i n e a r} \\left(\\mathbf {F} ^ {\\text {p r e d} A}\\right) + \\mathbf {F} ^ {\\text {p r e d} A}. \\tag {13}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.722, + 0.698 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.895, + 0.863 + ], + "angle": 0, + "content": "Training. We compute the matching cost using the Hungarian matching algorithm [17], thereby obtaining the query indices that best match the ground truth within the query set. Then we compute the loss between the corresponding queries and the ground truth across three aspects: object category loss \\(\\mathcal{L}_{obj}\\), paired entity bounding box loss \\(\\mathcal{L}_{box}\\) and predicate category loss \\(\\mathcal{L}_{pred}\\). The first two aspects follow the DETR [2] framework. Besides, we follow [11] to apply an \\(\\mathcal{L}_1\\) loss between the predicate and image embeddings to align the visual and textual features. The overall loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.875, + 0.894, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {o b j} \\mathcal {L} _ {o b j} + \\lambda_ {b o x} \\mathcal {L} _ {b o x} + \\lambda_ {p r e d} \\mathcal {L} _ {p r e d} + \\lambda_ {l _ {1}} \\mathcal {L} _ {1}, \\tag {14}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17733" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.12, + 0.104, + 0.468, + 0.118 + ], + "angle": 0, + "content": "where \\(\\lambda_{obj},\\lambda_{box},\\lambda_{pred}\\) and \\(\\lambda_{l_1}\\) are hyper-parameters." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.119, + 0.484, + 0.25 + ], + "angle": 0, + "content": "Inference. During inference, the score of object category \\(\\mathbf{s}^o\\) is simply the top-1 score from the softmax distribution over objects, and the triplet score \\((\\mathbf{s}^{triA},\\mathbf{s}^{triS},\\mathbf{s}^{triC})\\) of each predicate subclass is obtained using sigmoid function. Subsequently, we calculate the predicate score. For the \\(i\\)-th query and the \\(j\\)-th predicate class, we take the maximum score among the triplet scores belonging to the predicate subclasses associated with that predicate class to obtain the score for the \\(j\\)-th predicate class." + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.259, + 0.484, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\mathbf {s} _ {i j} ^ {p r e d A} = \\operatorname {M a x} \\left\\{\\mathbf {s} _ {i 1} ^ {t r i A}, \\mathbf {s} _ {i 2} ^ {t r i A}, \\dots , \\mathbf {s} _ {i k} ^ {t r i A} \\right\\}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.288, + 0.484, + 0.348 + ], + "angle": 0, + "content": "where \\(1,2,\\ldots,k\\) are the score indices in the triplet scores of the \\(j\\)-th query. Finally, we compute \\(\\mathbf{s}_i^o*\\mathbf{s}_{ij}^{predA}\\) as the final score for the \\(j\\)-th predicate of the \\(i\\)-th query to generate the scene graph." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.36, + 0.232, + 0.377 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.384, + 0.297, + 0.4 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.405, + 0.484, + 0.522 + ], + "angle": 0, + "content": "Dataset. We evaluated our method on the Action Genome dataset [13], which is a large-scale video scene graph dataset built on the Charades dataset [34]. This dataset covers 234,253 frames from 9,848 videos, with a total of 476,229 object bounding boxes and 1,715,568 relationship entities. There are 36 entity categories and 26 predicate categories in these annotations. Predicates are further divided into three categories: attention, spatial, and contacting." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.523, + 0.484, + 0.743 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluated the Scene Graph Detection (SGDet) and Predicate Classification (PredCls) task on the AG dataset. SGDet aims to detect entities and predict predicates without labels. PredCls is a relatively simple task, focused on predicting predicates based on given oracle detection results. We primarily evaluate the performance of SGDet task. Following previous works, we adopt Recall@k and mean Recall@K as evaluation metric to measure the ratio of correct entities among the top K predicted entities with the highest confidence \\((\\mathrm{K} = 10,20,50)\\). To make a sufficient comparison with the baseline, the experiment was conducted under two settings: With Constraint and No Constraints. With Constraint only selects one predicate for each entity pair. No Constraints allows multiple predictions of the relationship between each entity pair." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Training Details. We employ ResNet-50 as the CNN backbone and The number of layers for the Transformer encoder and decoder are set to 6 and 3, respectively. For training our model, we initialize the network with the parameters of DETR trained with the COCO dataset. ARN train for 20 epochs using the AdamW optimizer [26] with the batch size 8, initial learning rate of \\(1e^{-4}\\) of the backbone. For each decoder, we set the number of queries \\(N_{q} = 100\\) and use fine-tune the CLIP text embeddings initialized classifier with a small learning rate of \\(1e^{-5}\\). Moreover, the last layer" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.102, + 0.895, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.313, + 0.896, + 0.368 + ], + "angle": 0, + "content": "Figure 4. Comparative per predicate class performance for SGDET task. Results are in terms of mR@10 under \"No constraint\" setup. Predicates are arranged from left to right in order from head to tail." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.895, + 0.47 + ], + "angle": 0, + "content": "of the predicate decoder is our proposed predicate association parsing module. The hyper-parameters for the Hungarian costs \\(\\lambda_{obj}\\), \\(\\lambda_{pred}\\), \\(\\lambda_{box}\\), \\(\\lambda_{l_1}\\), and the loss weights of them are set to 1, 2, 2.5, 20, respectively. We conduct our experiments on one NVIDIA 4090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.483, + 0.779, + 0.498 + ], + "angle": 0, + "content": "4.2. Comparison to state-of-the-art" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.895, + 0.579 + ], + "angle": 0, + "content": "To evaluate the performance of our model, we compare it with several state-of-the-art DSGG approaches [6, 10, 21, 23, 27, 29, 36, 37, 39, 40]. Tab. 1 shows the comparative results for SGDET task in terms of R@K under both No constraints and With Constraint settings." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.581, + 0.895, + 0.786 + ], + "angle": 0, + "content": "In SGDet task, our method outperforms all existing two stage and one stage methods, with ARN improves the performance of the second-best method by \\(2.3\\%\\), \\(2.4\\%\\), and \\(2.3\\%\\) under R@K metric of No constraints setting, respectively. Under the With Constraint setting, ARN improves performance by \\(1.6\\%\\) on R@10 and \\(0.9\\%\\) on R@20 compared to the current best method, however, it is slightly lower than the current best method on R@50. We conjecture the reason is as follows: In triplet prediction, ARN may generate more semantically similar results, and under the With Constraints setting, redundant results may occupy the top-ranking positions, thereby affecting the results in this setting. However, the results become reliable with fewer predictions when \\(K = [10,20]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.787, + 0.895, + 0.89 + ], + "angle": 0, + "content": "We evaluated the mR@K metric for each predicate category under No constraints setting, selecting several representative methods for comparison, as shown in Fig. 4. Our method achieves the best performance across most predicate categories, with varying shades of green indicate instances where other methods slightly outperform ours in a few categories. This demonstrates that our proposed ARN" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "17734" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.1, + 0.894, + 0.332 + ], + "angle": 0, + "content": "
MethodWith ConstraintNo Constraint
R@10R@20R@50mR@10mR@20mR@50R@10R@20R@50mR@10mR@20mR@50
Two Stage Method
VRD[27]19.224.526.0---19.128.840.5---
GPS-Net[23]24.733.135.1---24.435.747.3---
TRACE[36]13.914.514.58.28.28.226.535.645.322.831.341.8
STTran[6]25.234.137.016.620.822.224.636.248.820.929.739.2
APT[21]26.336.138.3---25.737.950.1---
STTran-TPI[40]26.234.637.415.620.221.8------
TR^2[39]26.835.538.3---27.839.250.0---
TEMPURA[29]28.133.434.918.522.623.729.838.146.424.733.943.7
TD^2[24]28.7-37.120.4-26.130.5-49.327.9-46.3
DSG-DETR[10]30.334.836.116.419.320.132.140.948.321.328.1-
OED[37]33.540.948.921.527.433.435.344.451.827.240.550.4
One Stage Method
RelTR[7]19.723.425.9---20.924.628.2---
OED*[37]31.537.743.720.626.130.233.441.349.026.138.547.7
Ours35.141.847.223.229.033.637.646.854.129.741.951.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.341, + 0.896, + 0.37 + ], + "angle": 0, + "content": "Table 1. Comparative results for SGDet task, on Action Genome dataset [13], in terms of Recall@K and mean Recall@K metric, best and second best results under each setting are respectively marked in bold and underline. where * indicates end-to-end training results." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.394, + 0.481, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.102, + 0.55, + 0.483, + 0.577 + ], + "angle": 0, + "content": "Figure 5. Ablation results for per class performance in SGDET task. Results are in terms of mR@50 under \"No constraint\"." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.591, + 0.483, + 0.666 + ], + "angle": 0, + "content": "
MethodWith ConstraintNo ConstraintFPS#Params(M)
R@10R@20R@50R@10R@20R@50
TRACE[36]27.527.527.572.691.696.4--
STTran[6]68.671.871.877.994.299.1-151.03
TEMPURA[29]68.871.571.580.494.299.4-248.8
TD2'2[24]70.1-73.181.7-98.8--
OED[37]73.076.176.183.395.399.212154.54
Ours74.674.674.683.796.799.916254.06
" + }, + { + "type": "table_caption", + "bbox": [ + 0.101, + 0.676, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Table 2. Comparison results for PredCls task, on Action Genome dataset [13], in terms of Recall@K metric. Best and second best results under each setting are respectively marked in bold and underline." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.757, + 0.483, + 0.801 + ], + "angle": 0, + "content": "framework effectively mitigates the long-tail problem, improving accuracy on tail predicates while maintaining performance on head predicates." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.802, + 0.483, + 0.89 + ], + "angle": 0, + "content": "In PredCls task, as shown in Tab. 2, ARN imporves the performance of the second-best methods by \\(0.4\\%\\), \\(1.4\\%\\), \\(0.5\\%\\) on the R@K metric under the No Constraints setting. It also achieves comparable performance under the With Constraints setting. Although ARN is designed for SGDet task, its performance on the PredCls task also demonstrates" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.395, + 0.741, + 0.41 + ], + "angle": 0, + "content": "itsversatilityacrossmultipletasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.895, + 0.529 + ], + "angle": 0, + "content": "Additionally, as shown in Tab. 2, our model exhibits a significant advantage in parameter efficiency compared to two-stage methods [6, 29]. In terms of inference speed, it surpasses the end-to-end algorithm OED. Notably, OED relies on adjacent frames during inference, requiring the processing of more images. Our experiments show that our method achieves an inference time of 8 minutes per round, whereas OED requires 22.7 minutes per round." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.541, + 0.661, + 0.558 + ], + "angle": 0, + "content": "4.3. Ablative Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.895, + 0.623 + ], + "angle": 0, + "content": "We proposed two modules in our ARN, Predicate Association Parsing Module (PAP) and Hierarchical Attention (HA), and performed ablation studies on different components to clarify their contributions to performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.624, + 0.896, + 0.8 + ], + "angle": 0, + "content": "Association Parsing Network. We evaluate the effectiveness of several components in the proposed ARN framework. We extend DETR for DSGG and establish it as our baseline. We choose focal loss as the loss function for predicate prediction and compare it with cross-entropy loss for fairness, as shown in the first and second rows of Tab. 3. The results indicate that focal loss is better suited for DSGG. We leverage CLIP to extend predicate prediction to triplet prediction, using CLIP's semantic knowledge to guide the model in learning more robust triplet semantic context. As demonstrated in the third row of Tab. 3, this is an effective approach for enhancing scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.802, + 0.897, + 0.89 + ], + "angle": 0, + "content": "Predicate Association Parsing Module. We examine the impact of the proposed Predicate Association Parsing Module on model performance, as shown in the fourth row of Tab. 3. Without incorporating HA, associating fine-grained paired entity features enhances the model's performance, demonstrating the positive role of visual relation" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17735" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.101, + 0.891, + 0.345 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.101, + 0.356, + 0.895, + 0.385 + ], + "angle": 0, + "content": "Figure 6. Comparative qualitative results. From left to right: input video frames, dynamic scene graphs generated by STTran [6], TEMPURA [29], OED [37] and ours. Incorrect and correct predicate predictions are shown in green and red, respectively." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.406, + 0.485, + 0.482 + ], + "angle": 0, + "content": "
Baseline\\( \\mathcal{L}_{focal} \\)CLIPPAPHANo Constraint
R@10R@50mR@10mR@50
✓28.444.319.639.3
✓✓31.745.425.644.8
✓✓✓34.451.727.148.5
✓✓✓✓35.352.227.949.8
✓✓✓✓✓37.654.129.751.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.492, + 0.465, + 0.504 + ], + "angle": 0, + "content": "Table 3. Ablative studies on our framework for SGDet task." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.518, + 0.484, + 0.604 + ], + "angle": 0, + "content": "
MethodNo Constraint
R@10R@50mR@10mR@50
Model I35.051.527.348.9
Model II36.953.029.351.4
Ours37.654.129.751.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.614, + 0.46, + 0.628 + ], + "angle": 0, + "content": "Table 4. Ablative studies on Conditional weight mapping." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.655, + 0.483, + 0.742 + ], + "angle": 0, + "content": "dependencies between entities and predicates in predicate prediction. With Hierarchical Attention included, the results further improve (fifth row of Tab. 3), indicating that visual relation dependencies between predicate subclasses also facilitate a broader understanding of interactive context, validating our motivation." + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.743, + 0.484, + 0.89 + ], + "angle": 0, + "content": "Conditional weight mapping. We evaluate the effectiveness of the proposed conditional weight mapping mechanism for association reasoning. Following the cross-attention structure, we directly add paired entity representations to the predicate query, denoted as Model I. Additionally, we experiment with handling features from different sources through nonlinear mapping, denoted as Model II. As observed in the first and second rows of Tab. 4, introducing entity information without proper organization negatively impacts the original performance of ARN. Nonlinear" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.895, + 0.453 + ], + "angle": 0, + "content": "mapping yields some improvement, but conditional weight mapping provides a more elegant approach to achieving effective association reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.454, + 0.895, + 0.557 + ], + "angle": 0, + "content": "More analyze. To further analyze the impact of the proposed modules, we present the average recall for each predicate class in the SGDet task, as shown in Fig. 5. It can be observed that after removing the PAP and HA modules, the model's performance drops in \\(92\\%\\) of predicate categories, with only slight improvements in the \"standing on\" and \"wearing\" categories \\((0.48\\%)\\) and \\(0.17\\%\\), respectively)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.686, + 0.581 + ], + "angle": 0, + "content": "4.4. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.587, + 0.895, + 0.69 + ], + "angle": 0, + "content": "Fig. 6 presents the qualitative results in SGDet task. The results show that existing two-stage methods (i.e., STTran, TEMPURA) are clearly at a disadvantage in entity detection (erroneous detections result in incorrect triplets). In contrast, one-stage methods excel in entity detection. Additionally, ARN accurately detects more tail predicates, demonstrating a broader capability for scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.704, + 0.629, + 0.718 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.729, + 0.895, + 0.889 + ], + "angle": 0, + "content": "In this paper, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN is the first to model reasoning over visual relation dependencies. We extend predicate prediction to triplet prediction and introduce a Predicate Association Parsing Module (PAP) to organize and interpret multiple predicate subclasses and fine-grained entity representations. Additionally, we design a Hierarchical Attention (HA) mechanism to effectively aggregate multi-source information for association reasoning. Extensive experiments on the Action Genome dataset demonstrate the effectiveness of our method." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17736" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.101, + 0.103, + 0.486, + 0.25 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China (62171328,62171327,62072350); in part by the Key Program of Hubei Provincial Natural Science Foundation (Xiangyang Innovation and Development Joint Fund) under Grand (2025AFD050); by the Central Government's Special Project for Guiding Local Development in Hubei Province (2022BGE242); by the Hubei Science and technology innovation team (T2023009); by the National Natural Science Foundation of Hubei (2023AFB158)." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.263, + 0.196, + 0.278 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.287, + 0.484, + 0.354 + ], + "angle": 0, + "content": "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3286-3295, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.357, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229, 2020. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.424, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9004-9013, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.493, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[4] Siqi Chen, Jun Xiao, and Long Chen. Video scene graph generation from single-frame weak supervision. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.548, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[5] Anoop Cherian, Chiori Hori, Tim K Marks, and Jonathan Le Roux. \\((2.5 + 1)\\) d spatio-temporal scene graphs for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 444-453, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.603, + 0.484, + 0.669 + ], + "angle": 0, + "content": "[6] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16372-16382, 2021. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.672, + 0.484, + 0.724 + ], + "angle": 0, + "content": "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.726, + 0.484, + 0.792 + ], + "angle": 0, + "content": "[8] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2150-2159, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.111, + 0.794, + 0.484, + 0.861 + ], + "angle": 0, + "content": "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjm Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 88-98, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.863, + 0.484, + 0.89 + ], + "angle": 0, + "content": "[10] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term de" + }, + { + "type": "list", + "bbox": [ + 0.104, + 0.287, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.895, + 0.157 + ], + "angle": 0, + "content": "pendencies for generating dynamic scene graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5130-5139, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.159, + 0.895, + 0.212 + ], + "angle": 0, + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Eleventh International Conference on Learning Representations (ICLR), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.213, + 0.895, + 0.267 + ], + "angle": 0, + "content": "[12] Xia Hua, Xinqing Wang, Ting Rui, Faming Shao, and Dong Wang. Adversarial reinforcement learning with object-scene relational graph for video captioning. IEEE Transactions on Image Processing (TIP), 31:2004-2016, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.268, + 0.895, + 0.333 + ], + "angle": 0, + "content": "[13] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10236–10247, 2020. 1, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.335, + 0.895, + 0.402 + ], + "angle": 0, + "content": "[14] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668-3678, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.403, + 0.895, + 0.468 + ], + "angle": 0, + "content": "[15] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debiasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2516-2526, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.47, + 0.895, + 0.55 + ], + "angle": 0, + "content": "[16] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision (IJCV), 123:32-73, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.551, + 0.895, + 0.59 + ], + "angle": 0, + "content": "[17] Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics quarterly (NRL), 2(1-2): 83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.592, + 0.895, + 0.659 + ], + "angle": 0, + "content": "[18] Stan Weixian Lei, Difei Gao, Jay Zhangjie Wu, Yuxuan Wang, Wei Liu, Mengmi Zhang, and Mike Zheng Shou. Symbolic replay: Scene graph as prompt for continual learning on vqa task. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1250-1259, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.66, + 0.895, + 0.727 + ], + "angle": 0, + "content": "[19] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28369-28379, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.728, + 0.895, + 0.781 + ], + "angle": 0, + "content": "[20] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19486-19496, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.782, + 0.895, + 0.848 + ], + "angle": 0, + "content": "[21] Yiming Li, Xiaoshan Yang, and Changsheng Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13874-13883, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.849, + 0.895, + 0.89 + ], + "angle": 0, + "content": "[22] T Lin. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2998, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.518, + 0.939 + ], + "angle": 0, + "content": "17737" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.104, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[23] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3746-3753, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.172, + 0.484, + 0.238 + ], + "angle": 0, + "content": "[24] Xin Lin, Chong Shi, Yibing Zhan, Zuopeng Yang, Yaqi Wu, and Dacheng Tao. \\(\\mathrm{Td}^2\\)-net: Toward denoising and debiasing for video scene graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3495-3503, 2024. 2, 7, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.24, + 0.483, + 0.306 + ], + "angle": 0, + "content": "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.308, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Eleventh International Conference on Learning Representations (ICLR), 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.349, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[27] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In European Conference on Computer Vision (ECCV), pages 852–869, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.403, + 0.483, + 0.468 + ], + "angle": 0, + "content": "[28] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.47, + 0.483, + 0.536 + ], + "angle": 0, + "content": "[29] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22803-22813, 2023. 2, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.538, + 0.483, + 0.605 + ], + "angle": 0, + "content": "[30] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.607, + 0.483, + 0.659 + ], + "angle": 0, + "content": "[31] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for video scene graph generation. IEEE Transactions on Image Processing (TIP), 33:556-568, 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.66, + 0.483, + 0.74 + ], + "angle": 0, + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.742, + 0.483, + 0.794 + ], + "angle": 0, + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.796, + 0.483, + 0.862 + ], + "angle": 0, + "content": "[34] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.104, + 0.864, + 0.483, + 0.89 + ], + "angle": 0, + "content": "[35] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detec" + }, + { + "type": "list", + "bbox": [ + 0.103, + 0.104, + 0.484, + 0.89 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.104, + 0.895, + 0.145 + ], + "angle": 0, + "content": "tion with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10410-10419, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.147, + 0.895, + 0.212 + ], + "angle": 0, + "content": "[36] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13688-13697, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.215, + 0.895, + 0.28 + ], + "angle": 0, + "content": "[37] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27938-27947, 2024. 3, 6, 7, 8, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.283, + 0.894, + 0.335 + ], + "angle": 0, + "content": "[38] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cross-modal graph with meta concepts for video captioning. IEEE Transactions on Image Processing (TIP), 31: 5150-5162, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.338, + 0.894, + 0.404 + ], + "angle": 0, + "content": "[39] Jingyi Wang, Jina Huang, Can Zhang, and Zhidong Deng. Cross-modality time-variant relation learning for generating dynamic scene graphs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8231–8238. IEEE, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.407, + 0.894, + 0.473 + ], + "angle": 0, + "content": "[40] Shuang Wang, Lianli Gao, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, and Jingkuan Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 5793-5801, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.476, + 0.895, + 0.542 + ], + "angle": 0, + "content": "[41] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.544, + 0.894, + 0.609 + ], + "angle": 0, + "content": "[42] Xu Yang, Chongyang Gao, Hanwang Zhang, and Jianfei Cai. Hierarchical scene graph encoder-decoder for image paragraph captioning. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 4181-4189, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.612, + 0.895, + 0.679 + ], + "angle": 0, + "content": "[43] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1556-1565, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.681, + 0.895, + 0.746 + ], + "angle": 0, + "content": "[44] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5831-5840, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.749, + 0.895, + 0.814 + ], + "angle": 0, + "content": "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.104, + 0.895, + 0.814 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.928, + 0.519, + 0.939 + ], + "angle": 0, + "content": "17738" + } + ] +] \ No newline at end of file diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_origin.pdf b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..95ac8206b2ffedd62e42d315dfbd1e7aafe6511f --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/331916de-5a77-43d5-b8b1-867a812cb4ed_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc47f0992ff5cd76203d895372fd6de03f1b7092fbad362b7451539bad84c107 +size 1243929 diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/full.md b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c3a1a281161fd1d5f3163357f8a1d5b996a444f5 --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/full.md @@ -0,0 +1,325 @@ +# End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation + +Liwei Wang $^{1}$ , Yanduo Zhang $^{1,2\text{串}}$ , Tao Lu $^{1}$ , Fang Liu $^{1}$ , Huiqin Zhang $^{1}$ , Jiayi Ma $^{3}$ , Huabing Zhou $^{1\text{串}}$ $^{1}$ Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology + $^{2}$ Hubei University of Arts and Science $^{3}$ Wuhan University +wlw951226@163.com, {ydzhang,lut}@wit.edu.cn, {fangliu8822,666zhanghuiqin,jyma2010,Zhouhuabing}@gmail.com + +# Abstract + +Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets $<$ subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in https://github.com/wlw951226/ARN. + +# 1. Introduction + +Scene graph generation is a promising approach to scene understanding and has emerged as a frontier topic. Dynamic Scene Graph Generation (DSGG) advances this by capturing dynamic visual relationships in videos, contributing to a comprehensive understanding of video content. DSGG has + +![](images/53fe100a31e68d35bffe889123a41815bebc3c60c5f2ac1d9c69979803d6e177.jpg) +Figure 1. (a) The association frequency matrix between entities and predicates in the dataset [13]; (b) the association frequency matrix between predicate subclasses, where the vertical axis represents the pairwise combinations of two predicate subclasses; (c) an example and the pipeline of our associative reasoning network. + +also opened up new possibilities in areas such as video analysis [9, 18, 42], video captioning [12, 38], and video question answering [5, 43], enabling computers to better simulate human perception and cognitive processes in real-world scenarios. + +Dynamic scene graphs encode implicitly co-occurrences and affordance relations among entities and predicates. We analyze the frequency co-occurrence of predicates and entities in the Action Genome [13] dataset. As shown in Fig. 1(a), there is a strong association between certain predicates and entities; for instance, the predicate "writing on" often implies that the object is a notebook. Similarly, when the object is a notebook, it is likely in front of a person who is looking at it. Fig. 1(b) highlights the association be + +tween different predicates, and Fig. 1(c) provides an example demonstrating that holding an object strongly suggests its position in front of the person. However, existing methods overlook the dependencies between entities and predicates. Modeling these visual dependencies enables a multiperspective understanding of scenes, resulting in more accurate scene graphs. + +In addition, the currently prevalent DSGG paradigm [4, 6, 15, 29, 31, 36] are built upon a two-stage pipeline, focusing on capturing the temporal dependencies of videos. This two-stage paradigm independently optimizes the subtasks of entity localization [33] and predicate prediction, hindering end-to-end learning. Additionally, these methods generate redundant candidate boxes and predict relationships for all detected entity pairs, increasing computational cost. + +Recently, the one-stage detector DETR [2] has transformed the pattern of traditional detection frameworks by adopting a set-based prediction approach for one-to-one matching, eliminating the need for computationally expensive hand-designed components such as non-maximum suppression. Meanwhile, DETR-based methods [3, 25, 30, 35, 41] have demonstrated remarkable performance in image understanding. + +In this paper, we extend DETR to DSGG and propose an end-to-end Association Reasoning Network (ARN). ARN explores visual relation dependencies within scene graphs and performs association reasoning with fine-grained entity context. ARN leverages Contrastive Language-Image Pretraining (CLIP) [32] to bridge vision and language, modeling triplet cues through textual prompts, and extending predicate prediction to triplet prediction. This approach enriches the semantic representation of predicate prediction while ensuring semantic consistency between paired entities and predicates in DSGG. + +We further design a Predicate Association Parsing (PAP) module to link fine-grained representations from paired entities and different predicate subclasses. It employs a conditional weight mapping mechanism to dynamically organize predicate representations across different branches. Additionally, we develop a Hierarchical Attention (HA) mechanism that integrates spatio-temporal context with complementary predicate and entity cues in a structured hierarchy. This method achieves association reasoning between entities and predicates while maintaining stable performance during end-to-end training. Our experiments on the widely used Action Genome dataset demonstrate that ARN achieves state-of-the-art performance in DSGG. + +The contributions of this paper can be summarized as follows: + +- We propose an end-to-end ARN model for DSGG, which explores visual relation dependencies between entities and predicates and performs association reasoning to gen + +erate accurate scene graphs. + +- We introduce a Predicate Association Parsing (PAP) module to associate and organize fine-grained representations of different predicate subclasses and paired entities. +- We designed a Hierarchical Attention (HA) module to hierarchically perform visual relation association reasoning between paired entity representations and predicate relational context. +- We conducted a comprehensive evaluation of the proposed method, demonstrating its superior performance. + +# 2. Related Work + +Image scene graph generation. Image scene graph generation aims to identify entities and predict the predicates of image in form of triplets, which initially proposed by [14]. Most early works [8, 19, 23, 27, 44, 45], focused on refining entity and visual relationship representations in images using various network architectures. Subsequently, several works have attempted to design end-to-end image SGG models. Cong et al. [7] treated SGG as a set prediction problem and proposed an end-to-end image SGG model that directly predicts sparse scene graphs based solely on visual appearance. Furthermore, Li et al. [20] designed a graph assembling module that infers the connectivity of a bipartite graph based on entities features, generating scene graphs through graph matching. Meanwhile, some researchers [28] have focused on addressing the long-tailed distribution of predicates in benchmark datasets [16] to generate unbiased scene graphs. + +Dynamic scene graph generation. The dynamic visual relationships in videos are more complex and have temporal dependencies, making DSGG more challenging compared to image scene graph generation. Previous works have focused on capturing temporal dependencies in videos and addressing data imbalance caused by long-tail distributions. Some works focus on designing more efficient temporal dependencies reasoning networks. Cong et al. [6] employed an encoder-decoder architecture to extract spatiotemporal context from video frames and proposed a two-stage baseline. Building on this, Pu et al. [31] injected prior spatio-temporal knowledge into DSGG via statistical co-occurrence, proposing a spatio-temporal Knowledge-Embedded Transformer. However, this method heavily relies on the inherent distribution of the data, such as the long-tail phenomenon. Differently, ARN does not rely on statistical priors. + +To overcome the limitations of the data's inherent distribution, Nag et al. [29] proposed a memory-guided training method to alleviate the long-tailed distribution. Lin et al. [24] introduced an asymmetric re-weighting loss that relieve the issue of label bias by adjusting the weights of tail predicates. Despite some progress, these methods sacrifice the performance of head predicates to focus on tail + +predicates. Additionally, they overlook the visual relational dependencies within the scene graph, which encompass a broader understanding of interaction context. + +The most relevant work to ours is OED [37], which introduces a progressive refined module to aggregate temporal context. But this module relies on independently optimized spatial context, resulting in degraded performance during end-to-end training. In contrast to OED, we focus on exploring visual dependencies between entities to enhance holistic scene understanding and enabling fully end-to-end implementation. + +# 3. Method + +In this work, we extend the end-to-end detector DETR[2] to DSGG and propose an Association Reasoning Network (ARN), as illustrated in Fig. 2. ARN sequentially extracts visual features and feeds them into two cascaded decoder modules, the Entity Decoder and the Predicate Decoder, to perform two subtasks: entity detection and triplet prediction (Sec 3.1). Subsequently, we introduce triplet cues modeling using CLIP's semantic knowledge (Sec 3.2). Moreover, the Predicate Association Parsing (PAP) Module in predicate decoder organizes predicate cues from different branches and fine-grained paired entity features (Sec 3.3). The Hierarchical Attention (HA) mechanism then aggregates these with spatio-temporal features to perform association reasoning, enhancing broader contextual understanding of video scenes (Sec 3.4). Finally, we introduce the training loss and inference process of ARN (Sec 3.5). + +# 3.1. Overall Architecture + +The overall architecture of our ARN is illustrated in Fig. 2. The ARN is composed of a Backbone, an Encoder, an Entity decoder and a Predicate Decoder. + +**Backbone.** CNN-based backbone maps an input frame $I_{i}$ to a spatial feature map $\mathbf{f}_i\in \mathbb{R}^{H\times W\times C}$ . + +Encoder. The transformer-based encoder [2, 3, 30] refines image features and captures spatial contextual information. First, a $1 \times 1$ convolution and flatten operation are applied to reduce the dimension and transform it into a serialized feature representation $\mathbf{f}_i \in \mathbb{R}^{HW \times d}$ . Next, we feed the $\mathbf{f}_i$ and fixed positional encodings $\mathbf{p}_e$ [1] to the encoder and output the enhanced spatial feature $\mathbf{F}_{vi} \in \mathbb{R}^{HW \times d}$ . + +Entity Decoder. Entity decoder consists of multiple transformer decoder layers and several FFN heads and introduces an entity query set $\mathbf{Q}^E\in \mathbb{R}^{N_q\times d}$ , where $N_{q}$ is the number of queries. Unlike previous methods, our query set has two subsets $\mathbf{Q}^E = [\mathbf{Q}^s,\mathbf{Q}^o ]\in \mathbb{R}^{N_q\times d}$ that aim to separate the subject and the object. + +Moreover, a pair of learnable entity position embeddings $\{\mathbf{e}^s,\mathbf{e}^o\} \in \mathbb{R}^{N_q\times d}$ is designed to assign the subject and object at the same position to form pairwise entities. Furthermore, $\mathbf{e}^o$ is employed to align the query sets for the two + +decoders and update them during training. Overall, the implementation of the entity decoder layer is as follows: + +$$ +\mathbf {Q} ^ {E} = \operatorname {C o n c a t} \left[ \mathbf {Q} ^ {s} + \mathbf {e} ^ {s}, \mathbf {Q} ^ {o} + \mathbf {e} ^ {o} \right], \tag {1} +$$ + +$$ +\tilde {\mathbf {Q}} _ {(l)} ^ {E} = \operatorname {D e c} _ {(l)} ^ {E} \left(\mathbf {Q} ^ {E}, \mathbf {F} _ {v i}\right), \tag {2} +$$ + +Specifically, at each layer $\mathrm{Dec}_{(l)}^{E}$ , the queries first undergoes self-attention to enable interactions among queries. Subsequently, cross-attention aggregates visual features $\mathbf{F}_{vi}$ into the queries to obtain updated queries $\tilde{\mathbf{Q}}_{(l)}^{E}$ . The output of the entity decoder, denoted as $\mathbf{F}^{E}$ , is derived from the final updated queries $\tilde{\mathbf{Q}}_{(l)}^{E}$ after $l$ layers. + +Finally, the entity representation $\mathbf{F}^E$ passes through the FFN head and object classifier at the top of the entity decoder to obtain the bounding box coordinates $\mathbf{B}^{E} = \{\mathbf{b}^{s},\mathbf{b}^{o}\} \in \mathbb{R}^{N_{q}\times 4}$ and object categories $\mathbf{C}^o\in \mathbb{R}^{N_q}$ for the paired entities. + +Predicate Decoder The predicate decoder is tailored to capture temporal dependencies and facilitate associative reasoning between entities and multiple predicate subclasses. In detail, the structure of the predicate decoder is a multi-branch decoder architecture, each branch corresponding to the prediction of a predicate subclass. + +Each branch consists of multiple decoder layers and initializes predicate queries $\mathbf{Q}^{predi} \in \mathbb{R}^{N_q \times d}$ in $i$ -th branch. Unlike the entity decoder, we capture temporal dependencies in visual features $\mathbf{F}_{vi}$ . Specifically, we initialized a learnable frame encoding $\mathbf{p}_t \in \mathbb{R}^{N_q \times t \times d}$ to identify different frames. Subsequently, we transpose the $\mathbf{F}_{vi}$ in the temporal dimension and update it by self-attention (SA) operation to obtain the spatio-temporal context feature $\mathbf{F}_{vit}$ . + +$$ +\mathbf {F} _ {v i t} = S A (\text {t r a n s p o s e} (\mathbf {F} _ {v i})), \tag {3} +$$ + +Ultimately, $\mathbf{Q}^{predi}$ aggregates the $\mathbf{F}_{vit}$ through crossattention to generate fine-grained triplet representations $\mathbf{F}^{predi}$ . Furthermore, to encourage associative reasoning across predicates, we design a predicate association embedding $\mathbf{e}^{predi} \in \mathbb{R}^{N_q \times d}$ , which is dynamically updated along with the $\mathbf{Q}^{predi}$ in the predicate decoder. The decoding process of the predicate representation in the $l$ -th layer of predicate decoder $Dec_l^{predi}$ is as follows: + +$$ +\tilde {\mathbf {e}} ^ {p r e d i} = \mathbf {e} ^ {o} + \mathbf {e} ^ {p r e d i}, \tag {4} +$$ + +$$ +\mathbf {F} _ {(l)} ^ {p r e d i} = \operatorname {D e c} _ {(l)} ^ {p r e d i} \left(\mathbf {Q} ^ {p r e d i}, \tilde {\mathbf {e}} ^ {p r e d i}, \mathbf {F} _ {v i t}, \mathbf {p} _ {t}\right), \tag {5} +$$ + +# 3.2. Semantic Triplet Cue Modeling + +We use CLIP [32] to extending traditional predicate prediction to triplet prediction and inject domain-agnostic prior knowledge into our ARN. + +![](images/b3eea5138604925be438f43914f39ef99f113dd5ccd301f055c5792835e38afd.jpg) +Figure 2. Illustration of the proposed Association Reasoning Network (ARN), which first extracts visual features from the video, and the entity decoder is used to locate entities. Subsequently, the multi-branch predicate decoder captures the temporal dependencies in visual features and infers representations for each predicate subclasses. Finally, the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) associates different predicate subclasses cues and aggregates fine-grained entity features to refine the prediction results. + +Specifically, we follow the ActionGenome [13] setting and divide predicates into three subclasses: Attention, Spatial, and Contact. Then we design distinct textual prompts for each predicate subclass to build a bridge for interaction between CLIP and visual triplets :(1) 'The person is [attention predicate] at the [object]'; (2) 'The [object] is [spatial predicate] the person'; (3) 'a photo of a person maybe [contacting predicate] the [object]', an example is illustrated on the right side of Fig. 2. Next, we use the CLIP text encoder to extract text embeddings $\mathbf{E}^{predi} \in \mathbb{R}^{N_{predi} \times d}$ for each predicate subclass triplets, where $\mathbf{E}^{predi}$ serves as semantic cues for the visual triplets, $N_{predi}$ denotes the number of triplet categories in $i$ -th predicate decoder branch. + +Based on this, we use the $\mathbf{E}^{predi}$ to initialize the classifier weights $w^{predi}$ . For the $n$ -th query in $i$ -th branch, it computes the similarity score $s_n^i$ between the query representation and the triplet embeddings to align the visual and semantic information, then apply softmax activation to compute the focal loss [22], yielding the predicate prediction loss $\mathcal{L}_{\mathrm{pred}}$ . + +Moreover, for the entity decoder, we design textual prompts for object categories `a photo of person and a [object]''. We extract text embeddings $E^o \in$ + +$\mathbb{R}^{N_{obj}\times d}$ in the same manner and use them to initialize the classifier weights $w^o$ of the entity decoder. The object classifier is optimized using cross-entropy loss. + +$$ +s _ {n} ^ {i} = \left[ \operatorname {s i m} \left(\mathbf {F} _ {n} ^ {p r e d i}, \mathbf {E} _ {1} ^ {p r e d i}\right), \dots , \operatorname {s i m} \left(\mathbf {F} _ {n} ^ {p r e d i}, \mathbf {E} _ {N _ {p r e d i}} ^ {p r e d i}\right) \right] \tag {6} +$$ + +$$ +\mathcal {L} _ {\text {p r e d}} = \frac {1}{N} \sum_ {n \in N _ {q}} \mathcal {L} _ {\text {f o c a l}} \left(\sigma \left(s _ {n} ^ {i}\right), y _ {r} ^ {\text {p r e d i}}\right), \tag {7} +$$ + +# 3.3. Predicate Association Parsing Module + +As mentioned previously, the pairwise entities can be associated with multiple predicates, and these entity-predicate associative cues are crucial for DSGG. Therefore, we propose a Predicate Association Parsing (PAP) Module that first leverages a conditional weight mapping mechanism to differentiate and organize multiple predicate subclass cues, then performs associative reasoning by aggregating global spatio-temporal context and fine-grained entity representations. + +We denote the outputs of different branches in the predicate decoder as $\mathbf{F}^{predA}$ , $\mathbf{F}^{predS}$ , and $\mathbf{F}^{predC}$ , using the initials of predicate subclasses for simplicity. For $\mathbf{F}^{predA}$ , we treat $\mathbf{F}^{predS}$ and $\mathbf{F}^{predC}$ as complementary predicate + +![](images/484f43c2ac52c30acb5c9c7b3c75eb7e17cd7b7ff62c521e5162c07aa12417a9.jpg) +Predicate Association Parsing Module +Figure 3. The architecture of the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) Mechanism. The PAP organizes complementary predicate cues and embeds entity context into the current predicate representation. Building on this, the HA aggregates spatio-temporal context, explores visual relation dependencies, and progressively refines the predicate representation. + +representations and use a linear layer to reduce the dimensionality of them. Then we concatenate them as complementary predicate cues for the current branch, denoted as $\mathbf{F}^{compA} = \mathrm{Concat}[\mathbf{F}^{\mathrm{predS}},\mathbf{F}^{\mathrm{predC}}]$ . During the dot product operation, the cross-attention weights consist of two components, balancing the contributions of different predicate subclasses and preventing the model from being biased toward any single predicate representation. + +$$ +\mathbf {Q} = \left(\mathbf {Q} _ {S}, \mathbf {Q} _ {C}\right) = \left(\mathbf {F} ^ {p r e d S}, \mathbf {F} ^ {p r e d C}\right), \mathbf {K} = \mathbf {F} _ {v i t} + \mathbf {P} _ {t}, \tag {8} +$$ + +$$ +\mathbf {Q} ^ {T} \mathbf {K} = \mathbf {Q} _ {S} ^ {T} \mathbf {K} + \mathbf {Q} _ {C} ^ {T} \mathbf {K}, \tag {9} +$$ + +Next, we integrate the fine-grained paired entity features $\mathbf{F}^E$ into $\mathbf{F}^{predA}$ . + +$$ +\mathbf {F} _ {f u s i o n} ^ {p r e d A} = \operatorname {M L P} \left(\operatorname {c o n c a t} \left[ \mathbf {F} ^ {p r e d A} + \mathbf {e} ^ {p r e d A}, \mathbf {F} ^ {E} \right]\right), \tag {10} +$$ + +where $\mathbf{e}^{predA}$ denotes the predicate association embedding for attention predicates. Finally, we perform associative reasoning between the entity representation and various predicate representations through the proposed hierarchical attention module. + +# 3.4. Hierarchical Attention + +Visual relational dependencies are crucial for understanding scene context, yet traditional cross attention mechanisms + +struggle to associate multiple semantic representations. To this end, we designed a Hierarchical Attention (HA) Mechanism, building upon the predicate association parsing module. The HA mechanism enables fine-grained integration of entity features with diverse predicate representations, facilitating associative reasoning and improving the prediction of tail predicates. The concept is illustrated in Fig. 3. + +Specifically, HA is a quadruplet attention paradigm $(\mathbf{Q}_{cur},\mathbf{Q}_{comp},\mathbf{K},\mathbf{V})$ , where $\mathbf{Q}_{cur}$ is the predicate subclass representation of the current branch, and $\mathbf{Q}_{comp}$ represents a complementary predicate cues, which is the cues representation of the other predicate subclasses. + +Taking the branch of the attention predicate subclass as an example, we perform the first-level cross-attention operation to associate complementary predicate cues $\mathbf{F}^{compA}$ , where $\mathbf{F}^{compA}$ serves as the query $\mathbf{Q}_{comp}$ , and $\mathbf{F}_{vit}$ is used as the key and value, $\mathbf{e}^o$ as the query pos. This process highlights regions of interest related to other predicate subclasses and updates $\mathbf{Q}_{comp}$ to obtain $\mathbf{F}_{comp}$ . + +$$ +\mathbf {F} _ {\text {c o m p}} = \operatorname {A t t} _ {1} ^ {H A} \left(\mathbf {Q} _ {\text {c o m p}}, \mathbf {K} = \mathbf {F} _ {\text {v i t}} + \mathbf {p} _ {t}, \mathbf {V} = \mathbf {F} _ {\text {v i t}}\right), \tag {11} +$$ + +Second, we perform the second-level cross-attention operation, where $\mathbf{F}^{predA}$ serves as the query $\mathbf{Q}_{cur}$ , $\mathbf{Q}_{comp}$ as the key and $\mathbf{F}_{comp}$ as the value. This operation first establishes the association between the current predicate subclass and complementary predicate cues, then propagates the spatio-temporal visual features to the query of the current subclass, highlighting the most relevant categories and updating the predicate representation $\mathbf{F}^{predA}$ . Finally, the $\mathbf{F}^{predA}$ passes through the linear mapping layer and is added to the original input $\mathbf{Q}_{cur}$ , yielding the final output $\mathbf{F}_{assoc}^{predA}$ . + +$$ +\mathbf {F} ^ {p r e d A} = \operatorname {A t t} _ {2} ^ {H A} \left(\mathbf {Q} _ {c u r} = \mathbf {F} ^ {p r e d A}, \mathbf {Q} _ {c o m p}, \mathbf {F} _ {s u p}\right), \tag {12} +$$ + +$$ +\mathbf {F} _ {\text {a s s o c}} ^ {\text {p r e d} A} = \operatorname {L i n e a r} \left(\mathbf {F} ^ {\text {p r e d} A}\right) + \mathbf {F} ^ {\text {p r e d} A}. \tag {13} +$$ + +# 3.5. Training and Inference + +Training. We compute the matching cost using the Hungarian matching algorithm [17], thereby obtaining the query indices that best match the ground truth within the query set. Then we compute the loss between the corresponding queries and the ground truth across three aspects: object category loss $\mathcal{L}_{obj}$ , paired entity bounding box loss $\mathcal{L}_{box}$ and predicate category loss $\mathcal{L}_{pred}$ . The first two aspects follow the DETR [2] framework. Besides, we follow [11] to apply an $\mathcal{L}_1$ loss between the predicate and image embeddings to align the visual and textual features. The overall loss function is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {o b j} \mathcal {L} _ {o b j} + \lambda_ {b o x} \mathcal {L} _ {b o x} + \lambda_ {p r e d} \mathcal {L} _ {p r e d} + \lambda_ {l _ {1}} \mathcal {L} _ {1}, \tag {14} +$$ + +where $\lambda_{obj},\lambda_{box},\lambda_{pred}$ and $\lambda_{l_1}$ are hyper-parameters. + +Inference. During inference, the score of object category $\mathbf{s}^o$ is simply the top-1 score from the softmax distribution over objects, and the triplet score $(\mathbf{s}^{triA},\mathbf{s}^{triS},\mathbf{s}^{triC})$ of each predicate subclass is obtained using sigmoid function. Subsequently, we calculate the predicate score. For the $i$ -th query and the $j$ -th predicate class, we take the maximum score among the triplet scores belonging to the predicate subclasses associated with that predicate class to obtain the score for the $j$ -th predicate class. + +$$ +\mathbf {s} _ {i j} ^ {p r e d A} = \operatorname {M a x} \left\{\mathbf {s} _ {i 1} ^ {t r i A}, \mathbf {s} _ {i 2} ^ {t r i A}, \dots , \mathbf {s} _ {i k} ^ {t r i A} \right\}, \tag {15} +$$ + +where $1,2,\ldots,k$ are the score indices in the triplet scores of the $j$ -th query. Finally, we compute $\mathbf{s}_i^o*\mathbf{s}_{ij}^{predA}$ as the final score for the $j$ -th predicate of the $i$ -th query to generate the scene graph. + +# 4. Experiments + +# 4.1. Experimental Setting + +Dataset. We evaluated our method on the Action Genome dataset [13], which is a large-scale video scene graph dataset built on the Charades dataset [34]. This dataset covers 234,253 frames from 9,848 videos, with a total of 476,229 object bounding boxes and 1,715,568 relationship entities. There are 36 entity categories and 26 predicate categories in these annotations. Predicates are further divided into three categories: attention, spatial, and contacting. + +Evaluation Metrics. We evaluated the Scene Graph Detection (SGDet) and Predicate Classification (PredCls) task on the AG dataset. SGDet aims to detect entities and predict predicates without labels. PredCls is a relatively simple task, focused on predicting predicates based on given oracle detection results. We primarily evaluate the performance of SGDet task. Following previous works, we adopt Recall@k and mean Recall@K as evaluation metric to measure the ratio of correct entities among the top K predicted entities with the highest confidence $(\mathrm{K} = 10,20,50)$ . To make a sufficient comparison with the baseline, the experiment was conducted under two settings: With Constraint and No Constraints. With Constraint only selects one predicate for each entity pair. No Constraints allows multiple predictions of the relationship between each entity pair. + +Training Details. We employ ResNet-50 as the CNN backbone and The number of layers for the Transformer encoder and decoder are set to 6 and 3, respectively. For training our model, we initialize the network with the parameters of DETR trained with the COCO dataset. ARN train for 20 epochs using the AdamW optimizer [26] with the batch size 8, initial learning rate of $1e^{-4}$ of the backbone. For each decoder, we set the number of queries $N_{q} = 100$ and use fine-tune the CLIP text embeddings initialized classifier with a small learning rate of $1e^{-5}$ . Moreover, the last layer + +![](images/fa4b47066d415b7b387a77857f5cb6054d444c2b98ab80b49c817f3b766d760b.jpg) +Figure 4. Comparative per predicate class performance for SGDET task. Results are in terms of mR@10 under "No constraint" setup. Predicates are arranged from left to right in order from head to tail. + +of the predicate decoder is our proposed predicate association parsing module. The hyper-parameters for the Hungarian costs $\lambda_{obj}$ , $\lambda_{pred}$ , $\lambda_{box}$ , $\lambda_{l_1}$ , and the loss weights of them are set to 1, 2, 2.5, 20, respectively. We conduct our experiments on one NVIDIA 4090 GPU. + +# 4.2. Comparison to state-of-the-art + +To evaluate the performance of our model, we compare it with several state-of-the-art DSGG approaches [6, 10, 21, 23, 27, 29, 36, 37, 39, 40]. Tab. 1 shows the comparative results for SGDET task in terms of R@K under both No constraints and With Constraint settings. + +In SGDet task, our method outperforms all existing two stage and one stage methods, with ARN improves the performance of the second-best method by $2.3\%$ , $2.4\%$ , and $2.3\%$ under R@K metric of No constraints setting, respectively. Under the With Constraint setting, ARN improves performance by $1.6\%$ on R@10 and $0.9\%$ on R@20 compared to the current best method, however, it is slightly lower than the current best method on R@50. We conjecture the reason is as follows: In triplet prediction, ARN may generate more semantically similar results, and under the With Constraints setting, redundant results may occupy the top-ranking positions, thereby affecting the results in this setting. However, the results become reliable with fewer predictions when $K = [10,20]$ . + +We evaluated the mR@K metric for each predicate category under No constraints setting, selecting several representative methods for comparison, as shown in Fig. 4. Our method achieves the best performance across most predicate categories, with varying shades of green indicate instances where other methods slightly outperform ours in a few categories. This demonstrates that our proposed ARN + +
MethodWith ConstraintNo Constraint
R@10R@20R@50mR@10mR@20mR@50R@10R@20R@50mR@10mR@20mR@50
Two Stage Method
VRD[27]19.224.526.0---19.128.840.5---
GPS-Net[23]24.733.135.1---24.435.747.3---
TRACE[36]13.914.514.58.28.28.226.535.645.322.831.341.8
STTran[6]25.234.137.016.620.822.224.636.248.820.929.739.2
APT[21]26.336.138.3---25.737.950.1---
STTran-TPI[40]26.234.637.415.620.221.8------
TR^2[39]26.835.538.3---27.839.250.0---
TEMPURA[29]28.133.434.918.522.623.729.838.146.424.733.943.7
TD^2[24]28.7-37.120.4-26.130.5-49.327.9-46.3
DSG-DETR[10]30.334.836.116.419.320.132.140.948.321.328.1-
OED[37]33.540.948.921.527.433.435.344.451.827.240.550.4
One Stage Method
RelTR[7]19.723.425.9---20.924.628.2---
OED*[37]31.537.743.720.626.130.233.441.349.026.138.547.7
Ours35.141.847.223.229.033.637.646.854.129.741.951.9
+ +![](images/ee48c0e481e2e5d3cd57f055b430c43ee5c77c5a5d36539072a9efb4e1ff9ea6.jpg) +Figure 5. Ablation results for per class performance in SGDET task. Results are in terms of mR@50 under "No constraint". + +Table 1. Comparative results for SGDet task, on Action Genome dataset [13], in terms of Recall@K and mean Recall@K metric, best and second best results under each setting are respectively marked in bold and underline. where * indicates end-to-end training results. + +
MethodWith ConstraintNo ConstraintFPS#Params(M)
R@10R@20R@50R@10R@20R@50
TRACE[36]27.527.527.572.691.696.4--
STTran[6]68.671.871.877.994.299.1-151.03
TEMPURA[29]68.871.571.580.494.299.4-248.8
TD2'2[24]70.1-73.181.7-98.8--
OED[37]73.076.176.183.395.399.212154.54
Ours74.674.674.683.796.799.916254.06
+ +Table 2. Comparison results for PredCls task, on Action Genome dataset [13], in terms of Recall@K metric. Best and second best results under each setting are respectively marked in bold and underline. + +framework effectively mitigates the long-tail problem, improving accuracy on tail predicates while maintaining performance on head predicates. + +In PredCls task, as shown in Tab. 2, ARN imporves the performance of the second-best methods by $0.4\%$ , $1.4\%$ , $0.5\%$ on the R@K metric under the No Constraints setting. It also achieves comparable performance under the With Constraints setting. Although ARN is designed for SGDet task, its performance on the PredCls task also demonstrates + +itsversatilityacrossmultipletasks. + +Additionally, as shown in Tab. 2, our model exhibits a significant advantage in parameter efficiency compared to two-stage methods [6, 29]. In terms of inference speed, it surpasses the end-to-end algorithm OED. Notably, OED relies on adjacent frames during inference, requiring the processing of more images. Our experiments show that our method achieves an inference time of 8 minutes per round, whereas OED requires 22.7 minutes per round. + +# 4.3. Ablative Study + +We proposed two modules in our ARN, Predicate Association Parsing Module (PAP) and Hierarchical Attention (HA), and performed ablation studies on different components to clarify their contributions to performance. + +Association Parsing Network. We evaluate the effectiveness of several components in the proposed ARN framework. We extend DETR for DSGG and establish it as our baseline. We choose focal loss as the loss function for predicate prediction and compare it with cross-entropy loss for fairness, as shown in the first and second rows of Tab. 3. The results indicate that focal loss is better suited for DSGG. We leverage CLIP to extend predicate prediction to triplet prediction, using CLIP's semantic knowledge to guide the model in learning more robust triplet semantic context. As demonstrated in the third row of Tab. 3, this is an effective approach for enhancing scene understanding. + +Predicate Association Parsing Module. We examine the impact of the proposed Predicate Association Parsing Module on model performance, as shown in the fourth row of Tab. 3. Without incorporating HA, associating fine-grained paired entity features enhances the model's performance, demonstrating the positive role of visual relation + +![](images/0bde642e36a35b094ba0adf480ca0cc0d3f996d3508d568445d3a001ee897cac.jpg) +Figure 6. Comparative qualitative results. From left to right: input video frames, dynamic scene graphs generated by STTran [6], TEMPURA [29], OED [37] and ours. Incorrect and correct predicate predictions are shown in green and red, respectively. + +
Baseline\( \mathcal{L}_{focal} \)CLIPPAPHANo Constraint
R@10R@50mR@10mR@50
✓28.444.319.639.3
✓✓31.745.425.644.8
✓✓✓34.451.727.148.5
✓✓✓✓35.352.227.949.8
✓✓✓✓✓37.654.129.751.9
+ +Table 3. Ablative studies on our framework for SGDet task. + +
MethodNo Constraint
R@10R@50mR@10mR@50
Model I35.051.527.348.9
Model II36.953.029.351.4
Ours37.654.129.751.9
+ +Table 4. Ablative studies on Conditional weight mapping. + +dependencies between entities and predicates in predicate prediction. With Hierarchical Attention included, the results further improve (fifth row of Tab. 3), indicating that visual relation dependencies between predicate subclasses also facilitate a broader understanding of interactive context, validating our motivation. + +Conditional weight mapping. We evaluate the effectiveness of the proposed conditional weight mapping mechanism for association reasoning. Following the cross-attention structure, we directly add paired entity representations to the predicate query, denoted as Model I. Additionally, we experiment with handling features from different sources through nonlinear mapping, denoted as Model II. As observed in the first and second rows of Tab. 4, introducing entity information without proper organization negatively impacts the original performance of ARN. Nonlinear + +mapping yields some improvement, but conditional weight mapping provides a more elegant approach to achieving effective association reasoning. + +More analyze. To further analyze the impact of the proposed modules, we present the average recall for each predicate class in the SGDet task, as shown in Fig. 5. It can be observed that after removing the PAP and HA modules, the model's performance drops in $92\%$ of predicate categories, with only slight improvements in the "standing on" and "wearing" categories $(0.48\%)$ and $0.17\%$ , respectively). + +# 4.4. Qualitative Results + +Fig. 6 presents the qualitative results in SGDet task. The results show that existing two-stage methods (i.e., STTran, TEMPURA) are clearly at a disadvantage in entity detection (erroneous detections result in incorrect triplets). In contrast, one-stage methods excel in entity detection. Additionally, ARN accurately detects more tail predicates, demonstrating a broader capability for scene understanding. + +# 5. Conclusion + +In this paper, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN is the first to model reasoning over visual relation dependencies. We extend predicate prediction to triplet prediction and introduce a Predicate Association Parsing Module (PAP) to organize and interpret multiple predicate subclasses and fine-grained entity representations. Additionally, we design a Hierarchical Attention (HA) mechanism to effectively aggregate multi-source information for association reasoning. Extensive experiments on the Action Genome dataset demonstrate the effectiveness of our method. + +Acknowledgments. This work was supported in part by the National Natural Science Foundation of China (62171328,62171327,62072350); in part by the Key Program of Hubei Provincial Natural Science Foundation (Xiangyang Innovation and Development Joint Fund) under Grand (2025AFD050); by the Central Government's Special Project for Guiding Local Development in Hubei Province (2022BGE242); by the Hubei Science and technology innovation team (T2023009); by the National Natural Science Foundation of Hubei (2023AFB158). + +# References + +[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3286-3295, 2019. 3 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229, 2020. 2, 3, 5 +[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9004-9013, 2021. 2, 3 +[4] Siqi Chen, Jun Xiao, and Long Chen. Video scene graph generation from single-frame weak supervision. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2 +[5] Anoop Cherian, Chiori Hori, Tim K Marks, and Jonathan Le Roux. $(2.5 + 1)$ d spatio-temporal scene graphs for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 444-453, 2022. 1 +[6] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16372-16382, 2021. 2, 6, 7, 8 +[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 7, 1 +[8] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2150-2159, 2021. 2 +[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjm Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 88-98, 2023. 1 +[10] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term de + +pendencies for generating dynamic scene graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5130-5139, 2023. 6, 7 +[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Eleventh International Conference on Learning Representations (ICLR), 2021. 5 +[12] Xia Hua, Xinqing Wang, Ting Rui, Faming Shao, and Dong Wang. Adversarial reinforcement learning with object-scene relational graph for video captioning. IEEE Transactions on Image Processing (TIP), 31:2004-2016, 2022. 1 +[13] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10236–10247, 2020. 1, 4, 6, 7 +[14] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668-3678, 2015. 2 +[15] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debiasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2516-2526, 2024. 2 +[16] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision (IJCV), 123:32-73, 2017. 2 +[17] Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics quarterly (NRL), 2(1-2): 83-97, 1955. 5 +[18] Stan Weixian Lei, Difei Gao, Jay Zhangjie Wu, Yuxuan Wang, Wei Liu, Mengmi Zhang, and Mike Zheng Shou. Symbolic replay: Scene graph as prompt for continual learning on vqa task. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1250-1259, 2023. 1 +[19] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28369-28379, 2024. 2 +[20] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19486-19496, 2022. 2 +[21] Yiming Li, Xiaoshan Yang, and Changsheng Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13874-13883, 2022. 6, 7 +[22] T Lin. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2998, 2017. 4 + +[23] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3746-3753, 2020. 2, 6, 7 +[24] Xin Lin, Chong Shi, Yibing Zhan, Zuopeng Yang, Yaqi Wu, and Dacheng Tao. $\mathrm{Td}^2$ -net: Toward denoising and debiasing for video scene graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3495-3503, 2024. 2, 7, 1 +[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2 +[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Eleventh International Conference on Learning Representations (ICLR), 2017. 6 +[27] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In European Conference on Computer Vision (ECCV), pages 852–869, 2016. 2, 6, 7 +[28] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2 +[29] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22803-22813, 2023. 2, 6, 7, 8, 1 +[30] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2, 3 +[31] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for video scene graph generation. IEEE Transactions on Image Processing (TIP), 33:556-568, 2023. 2, 1 +[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 3 +[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 2 +[34] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, 2016. 6 +[35] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detec + +tion with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10410-10419, 2021. 2 +[36] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13688-13697, 2021. 2, 6, 7 +[37] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27938-27947, 2024. 3, 6, 7, 8, 1 +[38] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cross-modal graph with meta concepts for video captioning. IEEE Transactions on Image Processing (TIP), 31: 5150-5162, 2022. 1 +[39] Jingyi Wang, Jina Huang, Can Zhang, and Zhidong Deng. Cross-modality time-variant relation learning for generating dynamic scene graphs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8231–8238. IEEE, 2023. 6, 7 +[40] Shuang Wang, Lianli Gao, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, and Jingkuan Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 5793-5801, 2022. 6, 7 +[41] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2 +[42] Xu Yang, Chongyang Gao, Hanwang Zhang, and Jianfei Cai. Hierarchical scene graph encoder-decoder for image paragraph captioning. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 4181-4189, 2020. 1 +[43] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1556-1565, 2020. 1 +[44] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5831-5840, 2018. 2 +[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2 \ No newline at end of file diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/images.zip b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ea82c9cae85079e003e2d690afce21aa284788a7 --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4eeaf71c6e6c6c0c34bb36809790b87b5e256cab19b2d0b3fb48f2233040a8c4 +size 578342 diff --git a/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/layout.json b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6da90fb5342e5d374e8a034f3b07037ed61d8de8 --- /dev/null +++ b/2025/End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation/layout.json @@ -0,0 +1,8913 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 110, + 541, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 110, + 541, + 144 + ], + "spans": [ + { + "bbox": [ + 70, + 110, + 541, + 144 + ], + "type": "text", + "content": "End-to-End Entity-Predicate Association Reasoning for Dynamic Scene Graph Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "spans": [ + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": "Liwei Wang" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Yanduo Zhang" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1,2\\text{串}}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Tao Lu" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Fang Liu" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Huiqin Zhang" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Jiayi Ma" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": ", Huabing Zhou" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1\\text{串}}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": "Hubei Key Laboratory of Intelligent Robot, Wuhan Institute of Technology \n" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": "Hubei University of Arts and Science " + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 167, + 483, + 248 + ], + "type": "text", + "content": "Wuhan University \nwlw951226@163.com, {ydzhang,lut}@wit.edu.cn, {fangliu8822,666zhanghuiqin,jyma2010,Zhouhuabing}@gmail.com" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 156, + 276, + 202, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 156, + 276, + 202, + 288 + ], + "spans": [ + { + "bbox": [ + 156, + 276, + 202, + 288 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 301, + 296, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 301, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 61, + 301, + 296, + 580 + ], + "type": "text", + "content": "Dynamic Scene Graph Generation (DSGG) aims to comprehensively understand videos by abstracting them into visual triplets " + }, + { + "bbox": [ + 61, + 301, + 296, + 580 + ], + "type": "inline_equation", + "content": "<" + }, + { + "bbox": [ + 61, + 301, + 296, + 580 + ], + "type": "text", + "content": " subject, predicate, object>. Most existing methods focus on capturing temporal dependencies, but overlook crucial visual relationship dependencies between entities and predicates, as well as among predicate subclasses. These dependencies are essential for a deeper contextual understanding of scenarios. Additionally, current approaches do not support end-to-end training and instead rely on a two-stage pipeline, which incurs higher computational costs. To address these issues, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN leverages CLIP's semantic priors to model fine-grained triplet cues to generate scene graph. In addition, we design a Predicate Association Parsing (PAP) module that employs a conditional weight mapping mechanism to structure entity and predicate representations. We further introduce a Hierarchical Attention (HA) mechanism to integrate spatio-temporal context with entity and predicate representations, enabling effective associative reasoning. Extensive experiments on the Action Genome dataset demonstrate significant performance improvements over existing methods. The source code is available in https://github.com/wlw951226/ARN." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 605, + 141, + 617 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 605, + 141, + 617 + ], + "spans": [ + { + "bbox": [ + 63, + 605, + 141, + 617 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 625, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 625, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 61, + 625, + 296, + 684 + ], + "type": "text", + "content": "Scene graph generation is a promising approach to scene understanding and has emerged as a frontier topic. Dynamic Scene Graph Generation (DSGG) advances this by capturing dynamic visual relationships in videos, contributing to a comprehensive understanding of video content. DSGG has" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 276, + 545, + 459 + ], + "blocks": [ + { + "bbox": [ + 317, + 276, + 545, + 459 + ], + "lines": [ + { + "bbox": [ + 317, + 276, + 545, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 545, + 459 + ], + "type": "image", + "image_path": "53fe100a31e68d35bffe889123a41815bebc3c60c5f2ac1d9c69979803d6e177.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 467, + 548, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 467, + 548, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 548, + 521 + ], + "type": "text", + "content": "Figure 1. (a) The association frequency matrix between entities and predicates in the dataset [13]; (b) the association frequency matrix between predicate subclasses, where the vertical axis represents the pairwise combinations of two predicate subclasses; (c) an example and the pipeline of our associative reasoning network." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 541, + 547, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 541, + 547, + 598 + ], + "spans": [ + { + "bbox": [ + 312, + 541, + 547, + 598 + ], + "type": "text", + "content": "also opened up new possibilities in areas such as video analysis [9, 18, 42], video captioning [12, 38], and video question answering [5, 43], enabling computers to better simulate human perception and cognitive processes in real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 312, + 600, + 548, + 705 + ], + "type": "text", + "content": "Dynamic scene graphs encode implicitly co-occurrences and affordance relations among entities and predicates. We analyze the frequency co-occurrence of predicates and entities in the Action Genome [13] dataset. As shown in Fig. 1(a), there is a strong association between certain predicates and entities; for instance, the predicate \"writing on\" often implies that the object is a notebook. Similarly, when the object is a notebook, it is likely in front of a person who is looking at it. Fig. 1(b) highlights the association be" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "spans": [ + { + "bbox": [ + 71, + 13, + 116, + 44 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "spans": [ + { + "bbox": [ + 150, + 11, + 489, + 45 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 76, + 694, + 162, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 694, + 162, + 704 + ], + "spans": [ + { + "bbox": [ + 76, + 694, + 162, + 704 + ], + "type": "text", + "content": "Corresponding authors." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17729" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 294, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 294, + 163 + ], + "type": "text", + "content": "tween different predicates, and Fig. 1(c) provides an example demonstrating that holding an object strongly suggests its position in front of the person. However, existing methods overlook the dependencies between entities and predicates. Modeling these visual dependencies enables a multiperspective understanding of scenes, resulting in more accurate scene graphs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 163, + 294, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 294, + 266 + ], + "type": "text", + "content": "In addition, the currently prevalent DSGG paradigm [4, 6, 15, 29, 31, 36] are built upon a two-stage pipeline, focusing on capturing the temporal dependencies of videos. This two-stage paradigm independently optimizes the subtasks of entity localization [33] and predicate prediction, hindering end-to-end learning. Additionally, these methods generate redundant candidate boxes and predict relationships for all detected entity pairs, increasing computational cost." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 270, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 270, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 294, + 361 + ], + "type": "text", + "content": "Recently, the one-stage detector DETR [2] has transformed the pattern of traditional detection frameworks by adopting a set-based prediction approach for one-to-one matching, eliminating the need for computationally expensive hand-designed components such as non-maximum suppression. Meanwhile, DETR-based methods [3, 25, 30, 35, 41] have demonstrated remarkable performance in image understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 364, + 294, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 364, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 61, + 364, + 294, + 491 + ], + "type": "text", + "content": "In this paper, we extend DETR to DSGG and propose an end-to-end Association Reasoning Network (ARN). ARN explores visual relation dependencies within scene graphs and performs association reasoning with fine-grained entity context. ARN leverages Contrastive Language-Image Pretraining (CLIP) [32] to bridge vision and language, modeling triplet cues through textual prompts, and extending predicate prediction to triplet prediction. This approach enriches the semantic representation of predicate prediction while ensuring semantic consistency between paired entities and predicates in DSGG." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 492, + 294, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 492, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 61, + 492, + 294, + 643 + ], + "type": "text", + "content": "We further design a Predicate Association Parsing (PAP) module to link fine-grained representations from paired entities and different predicate subclasses. It employs a conditional weight mapping mechanism to dynamically organize predicate representations across different branches. Additionally, we develop a Hierarchical Attention (HA) mechanism that integrates spatio-temporal context with complementary predicate and entity cues in a structured hierarchy. This method achieves association reasoning between entities and predicates while maintaining stable performance during end-to-end training. Our experiments on the widely used Action Genome dataset demonstrate that ARN achieves state-of-the-art performance in DSGG." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 645, + 294, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 645, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 645, + 294, + 667 + ], + "type": "text", + "content": "The contributions of this paper can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 670, + 294, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 294, + 704 + ], + "type": "text", + "content": "- We propose an end-to-end ARN model for DSGG, which explores visual relation dependencies between entities and predicates and performs association reasoning to gen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 322, + 82, + 433, + 92 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 82, + 433, + 92 + ], + "spans": [ + { + "bbox": [ + 322, + 82, + 433, + 92 + ], + "type": "text", + "content": "erate accurate scene graphs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 93, + 547, + 197 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 93, + 545, + 126 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 93, + 545, + 126 + ], + "spans": [ + { + "bbox": [ + 314, + 93, + 545, + 126 + ], + "type": "text", + "content": "- We introduce a Predicate Association Parsing (PAP) module to associate and organize fine-grained representations of different predicate subclasses and paired entities." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 128, + 547, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 128, + 547, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 128, + 547, + 173 + ], + "type": "text", + "content": "- We designed a Hierarchical Attention (HA) module to hierarchically perform visual relation association reasoning between paired entity representations and predicate relational context." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 545, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 545, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 545, + 197 + ], + "type": "text", + "content": "- We conducted a comprehensive evaluation of the proposed method, demonstrating its superior performance." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 208, + 397, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 397, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 397, + 220 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 228, + 547, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 228, + 547, + 424 + ], + "spans": [ + { + "bbox": [ + 312, + 228, + 547, + 424 + ], + "type": "text", + "content": "Image scene graph generation. Image scene graph generation aims to identify entities and predict the predicates of image in form of triplets, which initially proposed by [14]. Most early works [8, 19, 23, 27, 44, 45], focused on refining entity and visual relationship representations in images using various network architectures. Subsequently, several works have attempted to design end-to-end image SGG models. Cong et al. [7] treated SGG as a set prediction problem and proposed an end-to-end image SGG model that directly predicts sparse scene graphs based solely on visual appearance. Furthermore, Li et al. [20] designed a graph assembling module that infers the connectivity of a bipartite graph based on entities features, generating scene graphs through graph matching. Meanwhile, some researchers [28] have focused on addressing the long-tailed distribution of predicates in benchmark datasets [16] to generate unbiased scene graphs." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 426, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 547, + 622 + ], + "type": "text", + "content": "Dynamic scene graph generation. The dynamic visual relationships in videos are more complex and have temporal dependencies, making DSGG more challenging compared to image scene graph generation. Previous works have focused on capturing temporal dependencies in videos and addressing data imbalance caused by long-tail distributions. Some works focus on designing more efficient temporal dependencies reasoning networks. Cong et al. [6] employed an encoder-decoder architecture to extract spatiotemporal context from video frames and proposed a two-stage baseline. Building on this, Pu et al. [31] injected prior spatio-temporal knowledge into DSGG via statistical co-occurrence, proposing a spatio-temporal Knowledge-Embedded Transformer. However, this method heavily relies on the inherent distribution of the data, such as the long-tail phenomenon. Differently, ARN does not rely on statistical priors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 312, + 623, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 623, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 623, + 547, + 704 + ], + "type": "text", + "content": "To overcome the limitations of the data's inherent distribution, Nag et al. [29] proposed a memory-guided training method to alleviate the long-tailed distribution. Lin et al. [24] introduced an asymmetric re-weighting loss that relieve the issue of label bias by adjusting the weights of tail predicates. Despite some progress, these methods sacrifice the performance of head predicates to focus on tail" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17730" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 296, + 115 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 296, + 115 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 296, + 115 + ], + "type": "text", + "content": "predicates. Additionally, they overlook the visual relational dependencies within the scene graph, which encompass a broader understanding of interaction context." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 116, + 296, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 116, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 116, + 296, + 209 + ], + "type": "text", + "content": "The most relevant work to ours is OED [37], which introduces a progressive refined module to aggregate temporal context. But this module relies on independently optimized spatial context, resulting in degraded performance during end-to-end training. In contrast to OED, we focus on exploring visual dependencies between entities to enhance holistic scene understanding and enabling fully end-to-end implementation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 219, + 117, + 231 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 219, + 117, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 219, + 117, + 231 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 238, + 296, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 296, + 425 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 296, + 425 + ], + "type": "text", + "content": "In this work, we extend the end-to-end detector DETR[2] to DSGG and propose an Association Reasoning Network (ARN), as illustrated in Fig. 2. ARN sequentially extracts visual features and feeds them into two cascaded decoder modules, the Entity Decoder and the Predicate Decoder, to perform two subtasks: entity detection and triplet prediction (Sec 3.1). Subsequently, we introduce triplet cues modeling using CLIP's semantic knowledge (Sec 3.2). Moreover, the Predicate Association Parsing (PAP) Module in predicate decoder organizes predicate cues from different branches and fine-grained paired entity features (Sec 3.3). The Hierarchical Attention (HA) mechanism then aggregates these with spatio-temporal features to perform association reasoning, enhancing broader contextual understanding of video scenes (Sec 3.4). Finally, we introduce the training loss and inference process of ARN (Sec 3.5)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 431, + 180, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 180, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 180, + 443 + ], + "type": "text", + "content": "3.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 448, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 448, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 61, + 448, + 296, + 483 + ], + "type": "text", + "content": "The overall architecture of our ARN is illustrated in Fig. 2. The ARN is composed of a Backbone, an Encoder, an Entity decoder and a Predicate Decoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "text", + "content": "**Backbone.** CNN-based backbone maps an input frame " + }, + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "text", + "content": " to a spatial feature map " + }, + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i\\in \\mathbb{R}^{H\\times W\\times C}" + }, + { + "bbox": [ + 62, + 483, + 295, + 506 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": "Encoder. The transformer-based encoder [2, 3, 30] refines image features and captures spatial contextual information. First, a " + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": " convolution and flatten operation are applied to reduce the dimension and transform it into a serialized feature representation " + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i \\in \\mathbb{R}^{HW \\times d}" + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": ". Next, we feed the " + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i" + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": " and fixed positional encodings " + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_e" + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": " [1] to the encoder and output the enhanced spatial feature " + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vi} \\in \\mathbb{R}^{HW \\times d}" + }, + { + "bbox": [ + 61, + 507, + 296, + 589 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "text", + "content": "Entity Decoder. Entity decoder consists of multiple transformer decoder layers and several FFN heads and introduces an entity query set " + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^E\\in \\mathbb{R}^{N_q\\times d}" + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "text", + "content": " is the number of queries. Unlike previous methods, our query set has two subsets " + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^E = [\\mathbf{Q}^s,\\mathbf{Q}^o ]\\in \\mathbb{R}^{N_q\\times d}" + }, + { + "bbox": [ + 61, + 589, + 296, + 658 + ], + "type": "text", + "content": " that aim to separate the subject and the object." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "text", + "content": "Moreover, a pair of learnable entity position embeddings " + }, + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{e}^s,\\mathbf{e}^o\\} \\in \\mathbb{R}^{N_q\\times d}" + }, + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "text", + "content": " is designed to assign the subject and object at the same position to form pairwise entities. Furthermore, " + }, + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^o" + }, + { + "bbox": [ + 62, + 658, + 296, + 705 + ], + "type": "text", + "content": " is employed to align the query sets for the two" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 81, + 547, + 105 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 105 + ], + "type": "text", + "content": "decoders and update them during training. Overall, the implementation of the entity decoder layer is as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 363, + 114, + 547, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 114, + 547, + 129 + ], + "spans": [ + { + "bbox": [ + 363, + 114, + 547, + 129 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {E} = \\operatorname {C o n c a t} \\left[ \\mathbf {Q} ^ {s} + \\mathbf {e} ^ {s}, \\mathbf {Q} ^ {o} + \\mathbf {e} ^ {o} \\right], \\tag {1}", + "image_path": "46d6fa2e0b163fae19e94b384fbc46d07d674c298d22f42f1a49fc30ee50dfaa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 379, + 142, + 547, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 142, + 547, + 158 + ], + "spans": [ + { + "bbox": [ + 379, + 142, + 547, + 158 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {Q}} _ {(l)} ^ {E} = \\operatorname {D e c} _ {(l)} ^ {E} \\left(\\mathbf {Q} ^ {E}, \\mathbf {F} _ {v i}\\right), \\tag {2}", + "image_path": "a684205558c1d60c3f7fb995ae5ba04b7779af4fd37e6b8c3f4d6f898058ee75.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": "Specifically, at each layer " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "\\mathrm{Dec}_{(l)}^{E}" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": ", the queries first undergoes self-attention to enable interactions among queries. Subsequently, cross-attention aggregates visual features " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vi}" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": " into the queries to obtain updated queries " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{(l)}^{E}" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": ". The output of the entity decoder, denoted as " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{E}" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": ", is derived from the final updated queries " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{Q}}_{(l)}^{E}" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 163, + 547, + 238 + ], + "type": "text", + "content": " layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "text", + "content": "Finally, the entity representation " + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^E" + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "text", + "content": " passes through the FFN head and object classifier at the top of the entity decoder to obtain the bounding box coordinates " + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^{E} = \\{\\mathbf{b}^{s},\\mathbf{b}^{o}\\} \\in \\mathbb{R}^{N_{q}\\times 4}" + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "text", + "content": " and object categories " + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "inline_equation", + "content": "\\mathbf{C}^o\\in \\mathbb{R}^{N_q}" + }, + { + "bbox": [ + 313, + 239, + 547, + 297 + ], + "type": "text", + "content": " for the paired entities." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 297, + 547, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 547, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 547, + 366 + ], + "type": "text", + "content": "Predicate Decoder The predicate decoder is tailored to capture temporal dependencies and facilitate associative reasoning between entities and multiple predicate subclasses. In detail, the structure of the predicate decoder is a multi-branch decoder architecture, each branch corresponding to the prediction of a predicate subclass." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": "Each branch consists of multiple decoder layers and initializes predicate queries " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{predi} \\in \\mathbb{R}^{N_q \\times d}" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": "-th branch. Unlike the entity decoder, we capture temporal dependencies in visual features " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vi}" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": ". Specifically, we initialized a learnable frame encoding " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{p}_t \\in \\mathbb{R}^{N_q \\times t \\times d}" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": " to identify different frames. Subsequently, we transpose the " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vi}" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": " in the temporal dimension and update it by self-attention (SA) operation to obtain the spatio-temporal context feature " + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vit}" + }, + { + "bbox": [ + 313, + 366, + 547, + 460 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 369, + 468, + 547, + 481 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 468, + 547, + 481 + ], + "spans": [ + { + "bbox": [ + 369, + 468, + 547, + 481 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {v i t} = S A (\\text {t r a n s p o s e} (\\mathbf {F} _ {v i})), \\tag {3}", + "image_path": "b50e83de4405650f71d97383d82af7c097901acdd519e51d7e0d8837233c4c07.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": "Ultimately, " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{predi}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": " aggregates the " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vit}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": " through crossattention to generate fine-grained triplet representations " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predi}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": ". Furthermore, to encourage associative reasoning across predicates, we design a predicate association embedding " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{predi} \\in \\mathbb{R}^{N_q \\times d}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": ", which is dynamically updated along with the " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{predi}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": " in the predicate decoder. The decoding process of the predicate representation in the " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": "-th layer of predicate decoder " + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "inline_equation", + "content": "Dec_l^{predi}" + }, + { + "bbox": [ + 313, + 488, + 547, + 582 + ], + "type": "text", + "content": " is as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 386, + 591, + 547, + 605 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 591, + 547, + 605 + ], + "spans": [ + { + "bbox": [ + 386, + 591, + 547, + 605 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {e}} ^ {p r e d i} = \\mathbf {e} ^ {o} + \\mathbf {e} ^ {p r e d i}, \\tag {4}", + "image_path": "a71a982fa24b9ad004f940f8cc074c618487df1df375e42c02207c6f0ddaab27.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 341, + 629, + 547, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 629, + 547, + 646 + ], + "spans": [ + { + "bbox": [ + 341, + 629, + 547, + 646 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {(l)} ^ {p r e d i} = \\operatorname {D e c} _ {(l)} ^ {p r e d i} \\left(\\mathbf {Q} ^ {p r e d i}, \\tilde {\\mathbf {e}} ^ {p r e d i}, \\mathbf {F} _ {v i t}, \\mathbf {p} _ {t}\\right), \\tag {5}", + "image_path": "326b4f421c7f95bf7701e7385f7996edbe19d4c8b965e43e4514782539ac6267.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 652, + 479, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 479, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 479, + 665 + ], + "type": "text", + "content": "3.2. Semantic Triplet Cue Modeling" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 547, + 704 + ], + "type": "text", + "content": "We use CLIP [32] to extending traditional predicate prediction to triplet prediction and inject domain-agnostic prior knowledge into our ARN." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17731" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 80, + 523, + 320 + ], + "blocks": [ + { + "bbox": [ + 88, + 80, + 523, + 320 + ], + "lines": [ + { + "bbox": [ + 88, + 80, + 523, + 320 + ], + "spans": [ + { + "bbox": [ + 88, + 80, + 523, + 320 + ], + "type": "image", + "image_path": "b3eea5138604925be438f43914f39ef99f113dd5ccd301f055c5792835e38afd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 330, + 548, + 384 + ], + "lines": [ + { + "bbox": [ + 61, + 330, + 548, + 384 + ], + "spans": [ + { + "bbox": [ + 61, + 330, + 548, + 384 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed Association Reasoning Network (ARN), which first extracts visual features from the video, and the entity decoder is used to locate entities. Subsequently, the multi-branch predicate decoder captures the temporal dependencies in visual features and infers representations for each predicate subclasses. Finally, the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) associates different predicate subclasses cues and aggregates fine-grained entity features to refine the prediction results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "content": "Specifically, we follow the ActionGenome [13] setting and divide predicates into three subclasses: Attention, Spatial, and Contact. Then we design distinct textual prompts for each predicate subclass to build a bridge for interaction between CLIP and visual triplets :(1) 'The person is [attention predicate] at the [object]'; (2) 'The [object] is [spatial predicate] the person'; (3) 'a photo of a person maybe [contacting predicate] the [object]', an example is illustrated on the right side of Fig. 2. Next, we use the CLIP text encoder to extract text embeddings " + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{predi} \\in \\mathbb{R}^{N_{predi} \\times d}" + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "content": " for each predicate subclass triplets, where " + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{predi}" + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "content": " serves as semantic cues for the visual triplets, " + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "inline_equation", + "content": "N_{predi}" + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "content": " denotes the number of triplet categories in " + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 61, + 404, + 296, + 579 + ], + "type": "text", + "content": "-th predicate decoder branch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": "Based on this, we use the " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathbf{E}^{predi}" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": " to initialize the classifier weights " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "w^{predi}" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": ". For the " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": "-th query in " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": "-th branch, it computes the similarity score " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "s_n^i" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": " between the query representation and the triplet embeddings to align the visual and semantic information, then apply softmax activation to compute the focal loss [22], yielding the predicate prediction loss " + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{pred}}" + }, + { + "bbox": [ + 61, + 583, + 296, + 666 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 670, + 296, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 670, + 296, + 705 + ], + "spans": [ + { + "bbox": [ + 61, + 670, + 296, + 705 + ], + "type": "text", + "content": "Moreover, for the entity decoder, we design textual prompts for object categories `a photo of person and a [object]''. We extract text embeddings " + }, + { + "bbox": [ + 61, + 670, + 296, + 705 + ], + "type": "inline_equation", + "content": "E^o \\in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{N_{obj}\\times d}" + }, + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "type": "text", + "content": " in the same manner and use them to initialize the classifier weights " + }, + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "type": "inline_equation", + "content": "w^o" + }, + { + "bbox": [ + 313, + 403, + 548, + 440 + ], + "type": "text", + "content": " of the entity decoder. The object classifier is optimized using cross-entropy loss." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 460, + 547, + 486 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 460, + 547, + 486 + ], + "spans": [ + { + "bbox": [ + 320, + 460, + 547, + 486 + ], + "type": "interline_equation", + "content": "s _ {n} ^ {i} = \\left[ \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {1} ^ {p r e d i}\\right), \\dots , \\operatorname {s i m} \\left(\\mathbf {F} _ {n} ^ {p r e d i}, \\mathbf {E} _ {N _ {p r e d i}} ^ {p r e d i}\\right) \\right] \\tag {6}", + "image_path": "4579bd04e86196bb5eac929291c5346afdfd80c3250d79ea666606ecc4344e87.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 354, + 497, + 547, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 497, + 547, + 526 + ], + "spans": [ + { + "bbox": [ + 354, + 497, + 547, + 526 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r e d}} = \\frac {1}{N} \\sum_ {n \\in N _ {q}} \\mathcal {L} _ {\\text {f o c a l}} \\left(\\sigma \\left(s _ {n} ^ {i}\\right), y _ {r} ^ {\\text {p r e d i}}\\right), \\tag {7}", + "image_path": "0db13da41b0e42e1215cd873f7b9a23977f940b08070b3aafcf1b141c959d687.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 535, + 509, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 535, + 509, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 535, + 509, + 548 + ], + "type": "text", + "content": "3.3. Predicate Association Parsing Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 552, + 548, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 552, + 548, + 656 + ], + "spans": [ + { + "bbox": [ + 312, + 552, + 548, + 656 + ], + "type": "text", + "content": "As mentioned previously, the pairwise entities can be associated with multiple predicates, and these entity-predicate associative cues are crucial for DSGG. Therefore, we propose a Predicate Association Parsing (PAP) Module that first leverages a conditional weight mapping mechanism to differentiate and organize multiple predicate subclass cues, then performs associative reasoning by aggregating global spatio-temporal context and fine-grained entity representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": "We denote the outputs of different branches in the predicate decoder as " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predS}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predC}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": ", using the initials of predicate subclasses for simplicity. For " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": ", we treat " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predS}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predC}" + }, + { + "bbox": [ + 313, + 658, + 548, + 705 + ], + "type": "text", + "content": " as complementary predicate" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17732" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 95, + 293, + 283 + ], + "blocks": [ + { + "bbox": [ + 72, + 83, + 200, + 95 + ], + "lines": [ + { + "bbox": [ + 72, + 83, + 200, + 95 + ], + "spans": [ + { + "bbox": [ + 72, + 83, + 200, + 95 + ], + "type": "text", + "content": "Predicate Association Parsing Module" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 95, + 293, + 283 + ], + "lines": [ + { + "bbox": [ + 77, + 95, + 293, + 283 + ], + "spans": [ + { + "bbox": [ + 77, + 95, + 293, + 283 + ], + "type": "image", + "image_path": "484f43c2ac52c30acb5c9c7b3c75eb7e17cd7b7ff62c521e5162c07aa12417a9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 291, + 296, + 366 + ], + "lines": [ + { + "bbox": [ + 61, + 291, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 61, + 291, + 296, + 366 + ], + "type": "text", + "content": "Figure 3. The architecture of the proposed Predicate Association Parsing (PAP) Module and Hierarchical Attention (HA) Mechanism. The PAP organizes complementary predicate cues and embeds entity context into the current predicate representation. Building on this, the HA aggregates spatio-temporal context, explores visual relation dependencies, and progressively refines the predicate representation." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 386, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 386, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 61, + 386, + 296, + 479 + ], + "type": "text", + "content": "representations and use a linear layer to reduce the dimensionality of them. Then we concatenate them as complementary predicate cues for the current branch, denoted as " + }, + { + "bbox": [ + 61, + 386, + 296, + 479 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{compA} = \\mathrm{Concat}[\\mathbf{F}^{\\mathrm{predS}},\\mathbf{F}^{\\mathrm{predC}}]" + }, + { + "bbox": [ + 61, + 386, + 296, + 479 + ], + "type": "text", + "content": ". During the dot product operation, the cross-attention weights consist of two components, balancing the contributions of different predicate subclasses and preventing the model from being biased toward any single predicate representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 485, + 295, + 500 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 485, + 295, + 500 + ], + "spans": [ + { + "bbox": [ + 67, + 485, + 295, + 500 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} = \\left(\\mathbf {Q} _ {S}, \\mathbf {Q} _ {C}\\right) = \\left(\\mathbf {F} ^ {p r e d S}, \\mathbf {F} ^ {p r e d C}\\right), \\mathbf {K} = \\mathbf {F} _ {v i t} + \\mathbf {P} _ {t}, \\tag {8}", + "image_path": "07e435b97e9877e9f3a3b5e8e93854d3602aa472875b0bc659d1acd99bae5860.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 517, + 295, + 531 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 517, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 129, + 517, + 295, + 531 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} ^ {T} \\mathbf {K} = \\mathbf {Q} _ {S} ^ {T} \\mathbf {K} + \\mathbf {Q} _ {C} ^ {T} \\mathbf {K}, \\tag {9}", + "image_path": "09972c769b8c80cb33de3cd556756e795a661c2344e9df310ca2cbabbd8923da.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "text", + "content": "Next, we integrate the fine-grained paired entity features " + }, + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^E" + }, + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 62, + 535, + 295, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 73, + 575, + 295, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 575, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 73, + 575, + 295, + 592 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {f u s i o n} ^ {p r e d A} = \\operatorname {M L P} \\left(\\operatorname {c o n c a t} \\left[ \\mathbf {F} ^ {p r e d A} + \\mathbf {e} ^ {p r e d A}, \\mathbf {F} ^ {E} \\right]\\right), \\tag {10}", + "image_path": "11f0d19d376915a32aaba21e9409c1372a95b69d79fb48552fba60a1a67e4398.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 598, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 598, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 61, + 598, + 296, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 598, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^{predA}" + }, + { + "bbox": [ + 61, + 598, + 296, + 657 + ], + "type": "text", + "content": " denotes the predicate association embedding for attention predicates. Finally, we perform associative reasoning between the entity representation and various predicate representations through the proposed hierarchical attention module." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 664, + 187, + 676 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 664, + 187, + 676 + ], + "spans": [ + { + "bbox": [ + 62, + 664, + 187, + 676 + ], + "type": "text", + "content": "3.4. Hierarchical Attention" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 296, + 704 + ], + "type": "text", + "content": "Visual relational dependencies are crucial for understanding scene context, yet traditional cross attention mechanisms" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 81, + 547, + 162 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 81, + 547, + 162 + ], + "spans": [ + { + "bbox": [ + 313, + 81, + 547, + 162 + ], + "type": "text", + "content": "struggle to associate multiple semantic representations. To this end, we designed a Hierarchical Attention (HA) Mechanism, building upon the predicate association parsing module. The HA mechanism enables fine-grained integration of entity features with diverse predicate representations, facilitating associative reasoning and improving the prediction of tail predicates. The concept is illustrated in Fig. 3." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "text", + "content": "Specifically, HA is a quadruplet attention paradigm " + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "inline_equation", + "content": "(\\mathbf{Q}_{cur},\\mathbf{Q}_{comp},\\mathbf{K},\\mathbf{V})" + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cur}" + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "text", + "content": " is the predicate subclass representation of the current branch, and " + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{comp}" + }, + { + "bbox": [ + 313, + 163, + 547, + 220 + ], + "type": "text", + "content": " represents a complementary predicate cues, which is the cues representation of the other predicate subclasses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": "Taking the branch of the attention predicate subclass as an example, we perform the first-level cross-attention operation to associate complementary predicate cues " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{compA}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{compA}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": " serves as the query " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{comp}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{vit}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": " is used as the key and value, " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{e}^o" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": " as the query pos. This process highlights regions of interest related to other predicate subclasses and updates " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{comp}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{comp}" + }, + { + "bbox": [ + 313, + 220, + 547, + 303 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 320, + 320, + 547, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 320, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 320, + 320, + 547, + 335 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {c o m p}} = \\operatorname {A t t} _ {1} ^ {H A} \\left(\\mathbf {Q} _ {\\text {c o m p}}, \\mathbf {K} = \\mathbf {F} _ {\\text {v i t}} + \\mathbf {p} _ {t}, \\mathbf {V} = \\mathbf {F} _ {\\text {v i t}}\\right), \\tag {11}", + "image_path": "6a8adbf4dadbd977427d686ec51bf84bad9a54ee3291310f551e485dd57e3d14.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": "Second, we perform the second-level cross-attention operation, where " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": " serves as the query " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cur}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{comp}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": " as the key and " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{comp}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": " as the value. This operation first establishes the association between the current predicate subclass and complementary predicate cues, then propagates the spatio-temporal visual features to the query of the current subclass, highlighting the most relevant categories and updating the predicate representation " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": ". Finally, the " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^{predA}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": " passes through the linear mapping layer and is added to the original input " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}_{cur}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": ", yielding the final output " + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{assoc}^{predA}" + }, + { + "bbox": [ + 313, + 342, + 547, + 470 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 487, + 547, + 503 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 487, + 547, + 503 + ], + "spans": [ + { + "bbox": [ + 322, + 487, + 547, + 503 + ], + "type": "interline_equation", + "content": "\\mathbf {F} ^ {p r e d A} = \\operatorname {A t t} _ {2} ^ {H A} \\left(\\mathbf {Q} _ {c u r} = \\mathbf {F} ^ {p r e d A}, \\mathbf {Q} _ {c o m p}, \\mathbf {F} _ {s u p}\\right), \\tag {12}", + "image_path": "5854762fc1bd156bf875001e164ed637dd136675726ea1ef83b8af976b5cac78.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 356, + 519, + 546, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 519, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 356, + 519, + 546, + 533 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {\\text {a s s o c}} ^ {\\text {p r e d} A} = \\operatorname {L i n e a r} \\left(\\mathbf {F} ^ {\\text {p r e d} A}\\right) + \\mathbf {F} ^ {\\text {p r e d} A}. \\tag {13}", + "image_path": "749d2608a1e2bcc40f3a36bd486d68d1ec2edc8fa34d2a0eda9c8cd77f41c47b.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 540, + 441, + 552 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 441, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 441, + 552 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "content": "Training. We compute the matching cost using the Hungarian matching algorithm [17], thereby obtaining the query indices that best match the ground truth within the query set. Then we compute the loss between the corresponding queries and the ground truth across three aspects: object category loss " + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{obj}" + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "content": ", paired entity bounding box loss " + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{box}" + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "content": " and predicate category loss " + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{pred}" + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "content": ". The first two aspects follow the DETR [2] framework. Besides, we follow [11] to apply an " + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_1" + }, + { + "bbox": [ + 313, + 556, + 547, + 683 + ], + "type": "text", + "content": " loss between the predicate and image embeddings to align the visual and textual features. The overall loss function is formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 693, + 547, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 693, + 547, + 706 + ], + "spans": [ + { + "bbox": [ + 318, + 693, + 547, + 706 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {o b j} \\mathcal {L} _ {o b j} + \\lambda_ {b o x} \\mathcal {L} _ {b o x} + \\lambda_ {p r e d} \\mathcal {L} _ {p r e d} + \\lambda_ {l _ {1}} \\mathcal {L} _ {1}, \\tag {14}", + "image_path": "6a8904593fd751ecddc2a634f398742692b231e60d7044e56b2e79641108eb4d.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17733" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "spans": [ + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "inline_equation", + "content": "\\lambda_{obj},\\lambda_{box},\\lambda_{pred}" + }, + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "inline_equation", + "content": "\\lambda_{l_1}" + }, + { + "bbox": [ + 73, + 82, + 286, + 93 + ], + "type": "text", + "content": " are hyper-parameters." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": "Inference. During inference, the score of object category " + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "inline_equation", + "content": "\\mathbf{s}^o" + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": " is simply the top-1 score from the softmax distribution over objects, and the triplet score " + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "inline_equation", + "content": "(\\mathbf{s}^{triA},\\mathbf{s}^{triS},\\mathbf{s}^{triC})" + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": " of each predicate subclass is obtained using sigmoid function. Subsequently, we calculate the predicate score. For the " + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": "-th query and the " + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": "-th predicate class, we take the maximum score among the triplet scores belonging to the predicate subclasses associated with that predicate class to obtain the score for the " + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 94, + 296, + 198 + ], + "type": "text", + "content": "-th predicate class." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 97, + 205, + 296, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 205, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 97, + 205, + 296, + 222 + ], + "type": "interline_equation", + "content": "\\mathbf {s} _ {i j} ^ {p r e d A} = \\operatorname {M a x} \\left\\{\\mathbf {s} _ {i 1} ^ {t r i A}, \\mathbf {s} _ {i 2} ^ {t r i A}, \\dots , \\mathbf {s} _ {i k} ^ {t r i A} \\right\\}, \\tag {15}", + "image_path": "c1dcda2cb31c0d6072837910760cad1cc5b1f024621cfd1f17212b55c918dec0.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "1,2,\\ldots,k" + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": " are the score indices in the triplet scores of the " + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": "-th query. Finally, we compute " + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_i^o*\\mathbf{s}_{ij}^{predA}" + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": " as the final score for the " + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": "-th predicate of the " + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 61, + 228, + 296, + 275 + ], + "type": "text", + "content": "-th query to generate the scene graph." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 285, + 141, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 141, + 298 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 141, + 298 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 304, + 181, + 316 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 304, + 181, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 304, + 181, + 316 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 320, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 320, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 61, + 320, + 296, + 413 + ], + "type": "text", + "content": "Dataset. We evaluated our method on the Action Genome dataset [13], which is a large-scale video scene graph dataset built on the Charades dataset [34]. This dataset covers 234,253 frames from 9,848 videos, with a total of 476,229 object bounding boxes and 1,715,568 relationship entities. There are 36 entity categories and 26 predicate categories in these annotations. Predicates are further divided into three categories: attention, spatial, and contacting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 414, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 414, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 61, + 414, + 296, + 588 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluated the Scene Graph Detection (SGDet) and Predicate Classification (PredCls) task on the AG dataset. SGDet aims to detect entities and predict predicates without labels. PredCls is a relatively simple task, focused on predicting predicates based on given oracle detection results. We primarily evaluate the performance of SGDet task. Following previous works, we adopt Recall@k and mean Recall@K as evaluation metric to measure the ratio of correct entities among the top K predicted entities with the highest confidence " + }, + { + "bbox": [ + 61, + 414, + 296, + 588 + ], + "type": "inline_equation", + "content": "(\\mathrm{K} = 10,20,50)" + }, + { + "bbox": [ + 61, + 414, + 296, + 588 + ], + "type": "text", + "content": ". To make a sufficient comparison with the baseline, the experiment was conducted under two settings: With Constraint and No Constraints. With Constraint only selects one predicate for each entity pair. No Constraints allows multiple predictions of the relationship between each entity pair." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": "Training Details. We employ ResNet-50 as the CNN backbone and The number of layers for the Transformer encoder and decoder are set to 6 and 3, respectively. For training our model, we initialize the network with the parameters of DETR trained with the COCO dataset. ARN train for 20 epochs using the AdamW optimizer [26] with the batch size 8, initial learning rate of " + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": " of the backbone. For each decoder, we set the number of queries " + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "inline_equation", + "content": "N_{q} = 100" + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": " and use fine-tune the CLIP text embeddings initialized classifier with a small learning rate of " + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "inline_equation", + "content": "1e^{-5}" + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": ". Moreover, the last layer" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 80, + 547, + 241 + ], + "blocks": [ + { + "bbox": [ + 317, + 80, + 547, + 241 + ], + "lines": [ + { + "bbox": [ + 317, + 80, + 547, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 80, + 547, + 241 + ], + "type": "image", + "image_path": "fa4b47066d415b7b387a77857f5cb6054d444c2b98ab80b49c817f3b766d760b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 548, + 291 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 548, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 548, + 291 + ], + "type": "text", + "content": "Figure 4. Comparative per predicate class performance for SGDET task. Results are in terms of mR@10 under \"No constraint\" setup. Predicates are arranged from left to right in order from head to tail." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "content": "of the predicate decoder is our proposed predicate association parsing module. The hyper-parameters for the Hungarian costs " + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{obj}" + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{pred}" + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{box}" + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "inline_equation", + "content": "\\lambda_{l_1}" + }, + { + "bbox": [ + 313, + 314, + 547, + 372 + ], + "type": "text", + "content": ", and the loss weights of them are set to 1, 2, 2.5, 20, respectively. We conduct our experiments on one NVIDIA 4090 GPU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 382, + 476, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 476, + 394 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 476, + 394 + ], + "type": "text", + "content": "4.2. Comparison to state-of-the-art" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 547, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 547, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 547, + 458 + ], + "type": "text", + "content": "To evaluate the performance of our model, we compare it with several state-of-the-art DSGG approaches [6, 10, 21, 23, 27, 29, 36, 37, 39, 40]. Tab. 1 shows the comparative results for SGDET task in terms of R@K under both No constraints and With Constraint settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": "In SGDet task, our method outperforms all existing two stage and one stage methods, with ARN improves the performance of the second-best method by " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "2.3\\%" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": " under R@K metric of No constraints setting, respectively. Under the With Constraint setting, ARN improves performance by " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": " on R@10 and " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "0.9\\%" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": " on R@20 compared to the current best method, however, it is slightly lower than the current best method on R@50. We conjecture the reason is as follows: In triplet prediction, ARN may generate more semantically similar results, and under the With Constraints setting, redundant results may occupy the top-ranking positions, thereby affecting the results in this setting. However, the results become reliable with fewer predictions when " + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "inline_equation", + "content": "K = [10,20]" + }, + { + "bbox": [ + 312, + 460, + 547, + 622 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 623, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 547, + 704 + ], + "type": "text", + "content": "We evaluated the mR@K metric for each predicate category under No constraints setting, selecting several representative methods for comparison, as shown in Fig. 4. Our method achieves the best performance across most predicate categories, with varying shades of green indicate instances where other methods slightly outperform ours in a few categories. This demonstrates that our proposed ARN" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17734" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 79, + 547, + 262 + ], + "blocks": [ + { + "bbox": [ + 65, + 79, + 547, + 262 + ], + "lines": [ + { + "bbox": [ + 65, + 79, + 547, + 262 + ], + "spans": [ + { + "bbox": [ + 65, + 79, + 547, + 262 + ], + "type": "table", + "html": "
MethodWith ConstraintNo Constraint
R@10R@20R@50mR@10mR@20mR@50R@10R@20R@50mR@10mR@20mR@50
Two Stage Method
VRD[27]19.224.526.0---19.128.840.5---
GPS-Net[23]24.733.135.1---24.435.747.3---
TRACE[36]13.914.514.58.28.28.226.535.645.322.831.341.8
STTran[6]25.234.137.016.620.822.224.636.248.820.929.739.2
APT[21]26.336.138.3---25.737.950.1---
STTran-TPI[40]26.234.637.415.620.221.8------
TR^2[39]26.835.538.3---27.839.250.0---
TEMPURA[29]28.133.434.918.522.623.729.838.146.424.733.943.7
TD^2[24]28.7-37.120.4-26.130.5-49.327.9-46.3
DSG-DETR[10]30.334.836.116.419.320.132.140.948.321.328.1-
OED[37]33.540.948.921.527.433.435.344.451.827.240.550.4
One Stage Method
RelTR[7]19.723.425.9---20.924.628.2---
OED*[37]31.537.743.720.626.130.233.441.349.026.138.547.7
Ours35.141.847.223.229.033.637.646.854.129.741.951.9
", + "image_path": "61b2bf1f5e181df3bf5df37a79f9e17ccf8c7dfb5a18a3ca5b4a0e3fbf6a8601.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 312, + 294, + 427 + ], + "blocks": [ + { + "bbox": [ + 63, + 312, + 294, + 427 + ], + "lines": [ + { + "bbox": [ + 63, + 312, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 63, + 312, + 294, + 427 + ], + "type": "image", + "image_path": "ee48c0e481e2e5d3cd57f055b430c43ee5c77c5a5d36539072a9efb4e1ff9ea6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 62, + 435, + 295, + 456 + ], + "lines": [ + { + "bbox": [ + 62, + 435, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 295, + 456 + ], + "type": "text", + "content": "Figure 5. Ablation results for per class performance in SGDET task. Results are in terms of mR@50 under \"No constraint\"." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 64, + 468, + 295, + 527 + ], + "blocks": [ + { + "bbox": [ + 61, + 270, + 548, + 293 + ], + "lines": [ + { + "bbox": [ + 61, + 270, + 548, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 270, + 548, + 293 + ], + "type": "text", + "content": "Table 1. Comparative results for SGDet task, on Action Genome dataset [13], in terms of Recall@K and mean Recall@K metric, best and second best results under each setting are respectively marked in bold and underline. where * indicates end-to-end training results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 64, + 468, + 295, + 527 + ], + "lines": [ + { + "bbox": [ + 64, + 468, + 295, + 527 + ], + "spans": [ + { + "bbox": [ + 64, + 468, + 295, + 527 + ], + "type": "table", + "html": "
MethodWith ConstraintNo ConstraintFPS#Params(M)
R@10R@20R@50R@10R@20R@50
TRACE[36]27.527.527.572.691.696.4--
STTran[6]68.671.871.877.994.299.1-151.03
TEMPURA[29]68.871.571.580.494.299.4-248.8
TD2'2[24]70.1-73.181.7-98.8--
OED[37]73.076.176.183.395.399.212154.54
Ours74.674.674.683.796.799.916254.06
", + "image_path": "618420f2be9cd9ebf25e96d458de146469e5d9705655166d0c85c818fe7f4e13.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 535, + 295, + 578 + ], + "lines": [ + { + "bbox": [ + 61, + 535, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 61, + 535, + 295, + 578 + ], + "type": "text", + "content": "Table 2. Comparison results for PredCls task, on Action Genome dataset [13], in terms of Recall@K metric. Best and second best results under each setting are respectively marked in bold and underline." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 599, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 599, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 61, + 599, + 295, + 634 + ], + "type": "text", + "content": "framework effectively mitigates the long-tail problem, improving accuracy on tail predicates while maintaining performance on head predicates." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "text", + "content": "In PredCls task, as shown in Tab. 2, ARN imporves the performance of the second-best methods by " + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "inline_equation", + "content": "0.4\\%" + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 61, + 635, + 295, + 704 + ], + "type": "text", + "content": " on the R@K metric under the No Constraints setting. It also achieves comparable performance under the With Constraints setting. Although ARN is designed for SGDet task, its performance on the PredCls task also demonstrates" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 312, + 453, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 453, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 453, + 324 + ], + "type": "text", + "content": "itsversatilityacrossmultipletasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 325, + 547, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 547, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 547, + 418 + ], + "type": "text", + "content": "Additionally, as shown in Tab. 2, our model exhibits a significant advantage in parameter efficiency compared to two-stage methods [6, 29]. In terms of inference speed, it surpasses the end-to-end algorithm OED. Notably, OED relies on adjacent frames during inference, requiring the processing of more images. Our experiments show that our method achieves an inference time of 8 minutes per round, whereas OED requires 22.7 minutes per round." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 428, + 404, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 404, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 404, + 441 + ], + "type": "text", + "content": "4.3. Ablative Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 447, + 547, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 547, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 547, + 493 + ], + "type": "text", + "content": "We proposed two modules in our ARN, Predicate Association Parsing Module (PAP) and Hierarchical Attention (HA), and performed ablation studies on different components to clarify their contributions to performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 494, + 548, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 494, + 548, + 633 + ], + "spans": [ + { + "bbox": [ + 313, + 494, + 548, + 633 + ], + "type": "text", + "content": "Association Parsing Network. We evaluate the effectiveness of several components in the proposed ARN framework. We extend DETR for DSGG and establish it as our baseline. We choose focal loss as the loss function for predicate prediction and compare it with cross-entropy loss for fairness, as shown in the first and second rows of Tab. 3. The results indicate that focal loss is better suited for DSGG. We leverage CLIP to extend predicate prediction to triplet prediction, using CLIP's semantic knowledge to guide the model in learning more robust triplet semantic context. As demonstrated in the third row of Tab. 3, this is an effective approach for enhancing scene understanding." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 635, + 548, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 548, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 548, + 704 + ], + "type": "text", + "content": "Predicate Association Parsing Module. We examine the impact of the proposed Predicate Association Parsing Module on model performance, as shown in the fourth row of Tab. 3. Without incorporating HA, associating fine-grained paired entity features enhances the model's performance, demonstrating the positive role of visual relation" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17735" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 79, + 545, + 273 + ], + "blocks": [ + { + "bbox": [ + 69, + 79, + 545, + 273 + ], + "lines": [ + { + "bbox": [ + 69, + 79, + 545, + 273 + ], + "spans": [ + { + "bbox": [ + 69, + 79, + 545, + 273 + ], + "type": "image", + "image_path": "0bde642e36a35b094ba0adf480ca0cc0d3f996d3508d568445d3a001ee897cac.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 61, + 281, + 547, + 304 + ], + "lines": [ + { + "bbox": [ + 61, + 281, + 547, + 304 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 547, + 304 + ], + "type": "text", + "content": "Figure 6. Comparative qualitative results. From left to right: input video frames, dynamic scene graphs generated by STTran [6], TEMPURA [29], OED [37] and ours. Incorrect and correct predicate predictions are shown in green and red, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 321, + 296, + 381 + ], + "blocks": [ + { + "bbox": [ + 65, + 321, + 296, + 381 + ], + "lines": [ + { + "bbox": [ + 65, + 321, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 65, + 321, + 296, + 381 + ], + "type": "table", + "html": "
Baseline\\( \\mathcal{L}_{focal} \\)CLIPPAPHANo Constraint
R@10R@50mR@10mR@50
✓28.444.319.639.3
✓✓31.745.425.644.8
✓✓✓34.451.727.148.5
✓✓✓✓35.352.227.949.8
✓✓✓✓✓37.654.129.751.9
", + "image_path": "fd0dc09adc97c03b627f4d9508284d96be253db8a8a53f4cbb496543e4e10d51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 66, + 410, + 296, + 478 + ], + "blocks": [ + { + "bbox": [ + 73, + 389, + 284, + 399 + ], + "lines": [ + { + "bbox": [ + 73, + 389, + 284, + 399 + ], + "spans": [ + { + "bbox": [ + 73, + 389, + 284, + 399 + ], + "type": "text", + "content": "Table 3. Ablative studies on our framework for SGDet task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 410, + 296, + 478 + ], + "lines": [ + { + "bbox": [ + 66, + 410, + 296, + 478 + ], + "spans": [ + { + "bbox": [ + 66, + 410, + 296, + 478 + ], + "type": "table", + "html": "
MethodNo Constraint
R@10R@50mR@10mR@50
Model I35.051.527.348.9
Model II36.953.029.351.4
Ours37.654.129.751.9
", + "image_path": "56b8f7a9e06d64a36cfd2ee45a1211197fcc40bbd02b0e7fec575a29a80c4848.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 76, + 486, + 281, + 497 + ], + "lines": [ + { + "bbox": [ + 76, + 486, + 281, + 497 + ], + "spans": [ + { + "bbox": [ + 76, + 486, + 281, + 497 + ], + "type": "text", + "content": "Table 4. Ablative studies on Conditional weight mapping." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 61, + 518, + 295, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 518, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 61, + 518, + 295, + 587 + ], + "type": "text", + "content": "dependencies between entities and predicates in predicate prediction. With Hierarchical Attention included, the results further improve (fifth row of Tab. 3), indicating that visual relation dependencies between predicate subclasses also facilitate a broader understanding of interactive context, validating our motivation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 61, + 588, + 296, + 704 + ], + "type": "text", + "content": "Conditional weight mapping. We evaluate the effectiveness of the proposed conditional weight mapping mechanism for association reasoning. Following the cross-attention structure, we directly add paired entity representations to the predicate query, denoted as Model I. Additionally, we experiment with handling features from different sources through nonlinear mapping, denoted as Model II. As observed in the first and second rows of Tab. 4, introducing entity information without proper organization negatively impacts the original performance of ARN. Nonlinear" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 324, + 547, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 547, + 358 + ], + "type": "text", + "content": "mapping yields some improvement, but conditional weight mapping provides a more elegant approach to achieving effective association reasoning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "text", + "content": "More analyze. To further analyze the impact of the proposed modules, we present the average recall for each predicate class in the SGDet task, as shown in Fig. 5. It can be observed that after removing the PAP and HA modules, the model's performance drops in " + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "inline_equation", + "content": "92\\%" + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "text", + "content": " of predicate categories, with only slight improvements in the \"standing on\" and \"wearing\" categories " + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "inline_equation", + "content": "(0.48\\%)" + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "inline_equation", + "content": "0.17\\%" + }, + { + "bbox": [ + 313, + 359, + 547, + 441 + ], + "type": "text", + "content": ", respectively)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 448, + 419, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 419, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 419, + 460 + ], + "type": "text", + "content": "4.4. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 464, + 547, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 547, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 547, + 546 + ], + "type": "text", + "content": "Fig. 6 presents the qualitative results in SGDet task. The results show that existing two-stage methods (i.e., STTran, TEMPURA) are clearly at a disadvantage in entity detection (erroneous detections result in incorrect triplets). In contrast, one-stage methods excel in entity detection. Additionally, ARN accurately detects more tail predicates, demonstrating a broader capability for scene understanding." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 557, + 384, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 384, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 384, + 568 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 577, + 547, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 577, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 312, + 577, + 547, + 704 + ], + "type": "text", + "content": "In this paper, we propose an end-to-end Association Reasoning Network (ARN) for DSGG. ARN is the first to model reasoning over visual relation dependencies. We extend predicate prediction to triplet prediction and introduce a Predicate Association Parsing Module (PAP) to organize and interpret multiple predicate subclasses and fine-grained entity representations. Additionally, we design a Hierarchical Attention (HA) mechanism to effectively aggregate multi-source information for association reasoning. Extensive experiments on the Action Genome dataset demonstrate the effectiveness of our method." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17736" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 81, + 297, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 81, + 297, + 198 + ], + "spans": [ + { + "bbox": [ + 61, + 81, + 297, + 198 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the National Natural Science Foundation of China (62171328,62171327,62072350); in part by the Key Program of Hubei Provincial Natural Science Foundation (Xiangyang Innovation and Development Joint Fund) under Grand (2025AFD050); by the Central Government's Special Project for Guiding Local Development in Hubei Province (2022BGE242); by the Hubei Science and technology innovation team (T2023009); by the National Natural Science Foundation of Hubei (2023AFB158)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 208, + 119, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 208, + 119, + 220 + ], + "spans": [ + { + "bbox": [ + 63, + 208, + 119, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 227, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 67, + 227, + 296, + 280 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 227, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 67, + 227, + 296, + 280 + ], + "type": "text", + "content": "[1] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 3286-3295, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 282, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 282, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 67, + 282, + 296, + 334 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European Conference on Computer Vision (ECCV), pages 213-229, 2020. 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 335, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 335, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 67, + 335, + 296, + 388 + ], + "type": "text", + "content": "[3] Mingfei Chen, Yue Liao, Si Liu, Zhiyuan Chen, Fei Wang, and Chen Qian. Reformulating hoi detection as adaptive set prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9004-9013, 2021. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 390, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 390, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 67, + 390, + 296, + 431 + ], + "type": "text", + "content": "[4] Siqi Chen, Jun Xiao, and Long Chen. Video scene graph generation from single-frame weak supervision. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 434, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 434, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 67, + 434, + 296, + 475 + ], + "type": "text", + "content": "[5] Anoop Cherian, Chiori Hori, Tim K Marks, and Jonathan Le Roux. " + }, + { + "bbox": [ + 67, + 434, + 296, + 475 + ], + "type": "inline_equation", + "content": "(2.5 + 1)" + }, + { + "bbox": [ + 67, + 434, + 296, + 475 + ], + "type": "text", + "content": " d spatio-temporal scene graphs for video question answering. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 444-453, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 477, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 477, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 67, + 477, + 296, + 529 + ], + "type": "text", + "content": "[6] Yuren Cong, Wentong Liao, Hanno Ackermann, Bodo Rosenhahn, and Michael Ying Yang. Spatial-temporal transformer for dynamic scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16372-16382, 2021. 2, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 532, + 296, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 532, + 296, + 573 + ], + "spans": [ + { + "bbox": [ + 67, + 532, + 296, + 573 + ], + "type": "text", + "content": "[7] Yuren Cong, Michael Ying Yang, and Bodo Rosenhahn. Reltr: Relation transformer for scene graph generation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2023. 2, 7, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 67, + 574, + 296, + 627 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 574, + 296, + 627 + ], + "spans": [ + { + "bbox": [ + 67, + 574, + 296, + 627 + ], + "type": "text", + "content": "[8] Naina Dhingra, Florian Ritter, and Andreas Kunz. Bgt-net: Bidirectional gru transformer network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2150-2159, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 67, + 628, + 296, + 681 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 628, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 67, + 628, + 296, + 681 + ], + "type": "text", + "content": "[9] Azade Farshad, Yousef Yeganeh, Yu Chi, Chengzhi Shen, Böjm Ommer, and Nassir Navab. Scenegenie: Scene graph guided diffusion models for image synthesis. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 88-98, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 683, + 296, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 683, + 296, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 683, + 296, + 704 + ], + "type": "text", + "content": "[10] Shengyu Feng, Hesham Mostafa, Marcel Nassar, Somdeb Majumdar, and Subarna Tripathi. Exploiting long-term de" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 704 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 124 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 124 + ], + "type": "text", + "content": "pendencies for generating dynamic scene graphs. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 5130-5139, 2023. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 125, + 547, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 125, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 125, + 547, + 167 + ], + "type": "text", + "content": "[11] Xiuye Gu, Tsung-Yi Lin, Weicheng Kuo, and Yin Cui. Open-vocabulary object detection via vision and language knowledge distillation. In Eleventh International Conference on Learning Representations (ICLR), 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 168, + 547, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 168, + 547, + 211 + ], + "spans": [ + { + "bbox": [ + 315, + 168, + 547, + 211 + ], + "type": "text", + "content": "[12] Xia Hua, Xinqing Wang, Ting Rui, Faming Shao, and Dong Wang. Adversarial reinforcement learning with object-scene relational graph for video captioning. IEEE Transactions on Image Processing (TIP), 31:2004-2016, 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 212, + 547, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 212, + 547, + 263 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 547, + 263 + ], + "type": "text", + "content": "[13] Jingwei Ji, Ranjay Krishna, Li Fei-Fei, and Juan Carlos Niebles. Action genome: Actions as compositions of spatiotemporal scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10236–10247, 2020. 1, 4, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 265, + 547, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 265, + 547, + 318 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 547, + 318 + ], + "type": "text", + "content": "[14] Justin Johnson, Ranjay Krishna, Michael Stark, Li-Jia Li, David Shamma, Michael Bernstein, and Li Fei-Fei. Image retrieval using scene graphs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3668-3678, 2015. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 319, + 547, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 319, + 547, + 370 + ], + "spans": [ + { + "bbox": [ + 315, + 319, + 547, + 370 + ], + "type": "text", + "content": "[15] Anant Khandelwal. Flocode: Unbiased dynamic scene graph generation with temporal consistency and correlation debiasing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2516-2526, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 372, + 547, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 372, + 547, + 435 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 547, + 435 + ], + "type": "text", + "content": "[16] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision (IJCV), 123:32-73, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 436, + 547, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 436, + 547, + 467 + ], + "spans": [ + { + "bbox": [ + 315, + 436, + 547, + 467 + ], + "type": "text", + "content": "[17] Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics quarterly (NRL), 2(1-2): 83-97, 1955. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 468, + 547, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 468, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 315, + 468, + 547, + 521 + ], + "type": "text", + "content": "[18] Stan Weixian Lei, Difei Gao, Jay Zhangjie Wu, Yuxuan Wang, Wei Liu, Mengmi Zhang, and Mike Zheng Shou. Symbolic replay: Scene graph as prompt for continual learning on vqa task. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 1250-1259, 2023. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 522, + 547, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 522, + 547, + 575 + ], + "spans": [ + { + "bbox": [ + 315, + 522, + 547, + 575 + ], + "type": "text", + "content": "[19] Jiankai Li, Yunhong Wang, Xiefan Guo, Ruijie Yang, and Weixin Li. Leveraging predicate and triplet learning for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 28369-28379, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 576, + 547, + 618 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 576, + 547, + 618 + ], + "spans": [ + { + "bbox": [ + 315, + 576, + 547, + 618 + ], + "type": "text", + "content": "[20] Rongjie Li, Songyang Zhang, and Xuming He. Sgtr: End-to-end scene graph generation with transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19486-19496, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 619, + 547, + 671 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 619, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 315, + 619, + 547, + 671 + ], + "type": "text", + "content": "[21] Yiming Li, Xiaoshan Yang, and Changsheng Xu. Dynamic scene graph generation via anticipatory pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13874-13883, 2022. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "spans": [ + { + "bbox": [ + 315, + 672, + 547, + 704 + ], + "type": "text", + "content": "[22] T Lin. Focal loss for dense object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 2980-2998, 2017. 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17737" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 704 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 63, + 82, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 82, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 63, + 82, + 296, + 134 + ], + "type": "text", + "content": "[23] Xin Lin, Changxing Ding, Jinquan Zeng, and Dacheng Tao. Gps-net: Graph property sensing network for scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3746-3753, 2020. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 136, + 296, + 188 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 136, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 63, + 136, + 296, + 188 + ], + "type": "text", + "content": "[24] Xin Lin, Chong Shi, Yibing Zhan, Zuopeng Yang, Yaqi Wu, and Dacheng Tao. " + }, + { + "bbox": [ + 63, + 136, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\mathrm{Td}^2" + }, + { + "bbox": [ + 63, + 136, + 296, + 188 + ], + "type": "text", + "content": "-net: Toward denoising and debiasing for video scene graph generation. In Proceedings of the AAAI Conference on Artificial Intelligence (AAAI), pages 3495-3503, 2024. 2, 7, 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 190, + 295, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 190, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 63, + 190, + 295, + 242 + ], + "type": "text", + "content": "[25] Shilong Liu, Feng Li, Hao Zhang, Xiao Yang, Xianbiao Qi, Hang Su, Jun Zhu, and Lei Zhang. Dab-detr: Dynamic anchor boxes are better queries for detr. In Eleventh International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 243, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 243, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 63, + 243, + 295, + 275 + ], + "type": "text", + "content": "[26] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Eleventh International Conference on Learning Representations (ICLR), 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 276, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 276, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 63, + 276, + 295, + 317 + ], + "type": "text", + "content": "[27] Cewu Lu, Ranjay Krishna, Michael Bernstein, and Li Fei-Fei. Visual relationship detection with language priors. In European Conference on Computer Vision (ECCV), pages 852–869, 2016. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 319, + 295, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 319, + 295, + 370 + ], + "spans": [ + { + "bbox": [ + 63, + 319, + 295, + 370 + ], + "type": "text", + "content": "[28] Yukuan Min, Among Wu, and Cheng Deng. Environment-invariant curriculum relation learning for fine-grained scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13296-13307, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 372, + 295, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 372, + 295, + 424 + ], + "spans": [ + { + "bbox": [ + 63, + 372, + 295, + 424 + ], + "type": "text", + "content": "[29] Sayak Nag, Kyle Min, Subarna Tripathi, and Amit K Roy-Chowdhury. Unbiased scene graph generation in videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22803-22813, 2023. 2, 6, 7, 8, 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 426, + 295, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 426, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 63, + 426, + 295, + 479 + ], + "type": "text", + "content": "[30] Shan Ning, Longtian Qiu, Yongfei Liu, and Xuming He. Hoiclip: Efficient knowledge transfer for hoi detection with vision-language models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23507-23517, 2023. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 480, + 295, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 480, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 63, + 480, + 295, + 521 + ], + "type": "text", + "content": "[31] Tao Pu, Tianshui Chen, Hefeng Wu, Yongyi Lu, and Liang Lin. Spatial-temporal knowledge-embedded transformer for video scene graph generation. IEEE Transactions on Image Processing (TIP), 33:556-568, 2023. 2, 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 522, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 522, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 63, + 522, + 295, + 586 + ], + "type": "text", + "content": "[32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 587, + 295, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 587, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 63, + 587, + 295, + 628 + ], + "type": "text", + "content": "[33] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems (NeurIPS), 2015. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 630, + 295, + 682 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 630, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 63, + 630, + 295, + 682 + ], + "type": "text", + "content": "[34] Gunnar A Sigurdsson, Gül Varol, Xiaolong Wang, Ali Farhadi, Ivan Laptev, and Abhinav Gupta. Hollywood in homes: Crowdsourcing data collection for activity understanding. In European Conference on Computer Vision (ECCV), pages 510-526, 2016. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "spans": [ + { + "bbox": [ + 63, + 684, + 295, + 704 + ], + "type": "text", + "content": "[35] Masato Tamura, Hiroki Ohashi, and Tomoaki Yoshinaga. Qpic: Query-based pairwise human-object interaction detec" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 82, + 547, + 644 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 333, + 82, + 547, + 114 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 82, + 547, + 114 + ], + "spans": [ + { + "bbox": [ + 333, + 82, + 547, + 114 + ], + "type": "text", + "content": "tion with image-wide contextual information. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10410-10419, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 116, + 547, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 116, + 547, + 167 + ], + "spans": [ + { + "bbox": [ + 315, + 116, + 547, + 167 + ], + "type": "text", + "content": "[36] Yao Teng, Limin Wang, Zhifeng Li, and Gangshan Wu. Target adaptive context aggregation for video scene graph generation. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 13688-13697, 2021. 2, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 170, + 547, + 221 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 170, + 547, + 221 + ], + "spans": [ + { + "bbox": [ + 315, + 170, + 547, + 221 + ], + "type": "text", + "content": "[37] Guan Wang, Zhimin Li, Qingchao Chen, and Yang Liu. Oed: Towards one-stage end-to-end dynamic scene graph generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27938-27947, 2024. 3, 6, 7, 8, 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 224, + 547, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 224, + 547, + 265 + ], + "spans": [ + { + "bbox": [ + 315, + 224, + 547, + 265 + ], + "type": "text", + "content": "[38] Hao Wang, Guosheng Lin, Steven CH Hoi, and Chunyan Miao. Cross-modal graph with meta concepts for video captioning. IEEE Transactions on Image Processing (TIP), 31: 5150-5162, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 267, + 547, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 267, + 547, + 319 + ], + "spans": [ + { + "bbox": [ + 315, + 267, + 547, + 319 + ], + "type": "text", + "content": "[39] Jingyi Wang, Jina Huang, Can Zhang, and Zhidong Deng. Cross-modality time-variant relation learning for generating dynamic scene graphs. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8231–8238. IEEE, 2023. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 322, + 547, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 322, + 547, + 374 + ], + "spans": [ + { + "bbox": [ + 315, + 322, + 547, + 374 + ], + "type": "text", + "content": "[40] Shuang Wang, Lianli Gao, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, and Jingkuan Song. Dynamic scene graph generation via temporal prior inference. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 5793-5801, 2022. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 376, + 547, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 376, + 547, + 429 + ], + "spans": [ + { + "bbox": [ + 315, + 376, + 547, + 429 + ], + "type": "text", + "content": "[41] Xiaoshi Wu, Feng Zhu, Rui Zhao, and Hongsheng Li. Cora: Adapting clip for open-vocabulary detection with region prompting and anchor pre-matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7031-7040, 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 430, + 547, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 430, + 547, + 482 + ], + "spans": [ + { + "bbox": [ + 315, + 430, + 547, + 482 + ], + "type": "text", + "content": "[42] Xu Yang, Chongyang Gao, Hanwang Zhang, and Jianfei Cai. Hierarchical scene graph encoder-decoder for image paragraph captioning. In Proceedings of the 30th ACM International Conference on Multimedia (ACMMM), pages 4181-4189, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 484, + 547, + 537 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 484, + 547, + 537 + ], + "spans": [ + { + "bbox": [ + 315, + 484, + 547, + 537 + ], + "type": "text", + "content": "[43] Zekun Yang, Noa Garcia, Chenhui Chu, Mayu Otani, Yuta Nakashima, and Haruo Takemura. Bert representations for video question answering. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1556-1565, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 539, + 547, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 539, + 547, + 590 + ], + "spans": [ + { + "bbox": [ + 315, + 539, + 547, + 590 + ], + "type": "text", + "content": "[44] Rowan Zellers, Mark Yatskar, Sam Thomson, and Yejin Choi. Neural motifs: Scene graph parsing with global context. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5831-5840, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 593, + 547, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 593, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 315, + 593, + 547, + 644 + ], + "type": "text", + "content": "[45] Chaofan Zheng, Xinyu Lyu, Yuyu Guo, Pengpeng Zeng, Jingkuan Song, and Lianli Gao. Learning to generate scene graph from head to tail. In 2022 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6, 2022. 2" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "spans": [ + { + "bbox": [ + 294, + 734, + 317, + 743 + ], + "type": "text", + "content": "17738" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_content_list.json b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..42723673e6cd6d0356a00efe24019551eeeea82b --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_content_list.json @@ -0,0 +1,1912 @@ +[ + { + "type": "text", + "text": "End-to-End Multi-Modal Diffusion Mamba", + "text_level": 1, + "bbox": [ + 277, + 130, + 718, + 150 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chunhao Lu1 Qiang Lu1⊗ Meichen Dong1,2 Jake Luo3", + "bbox": [ + 269, + 179, + 743, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1China University of Petroleum-Beijing, 2Leyard Optoelectronic, 3University of Wisconsin-Milwaukee", + "bbox": [ + 102, + 202, + 913, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{luchunhao,meichen.dong}@student.cup.edu.cn luqiang@cup.edu.cn jakeluo@uwm.edu", + "bbox": [ + 150, + 227, + 862, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 277, + 326, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current end-to-end multi-modal models utilize different encoders and decoders to process input and output information. This separation hinders the joint representation learning of various modalities. To unify multi-modal processing, we propose a novel architecture called MDM (Multi-modal Diffusion Mamba). MDM utilizes a Mamba-based multistep selection diffusion model to progressively generate and refine modality-specific information through a unified variational autoencoder for both encoding and decoding. This innovative approach allows MDM to achieve superior performance when processing high-dimensional data, particularly in generating high-resolution images and extended text sequences simultaneously. Our evaluations in areas such as image generation, image captioning, visual question answering, text comprehension, and reasoning tasks demonstrate that MDM significantly outperforms existing end-to-end models (MonoFormer, LlamaGen, and Chameleon etc.) and competes effectively with SOTA models like GPT-4V, Gemini Pro, and Mistral. Our results validate MDM's effectiveness in unifying multi-modal processes while maintaining computational efficiency, establishing a new direction for end-to-end multi-modal architectures.", + "bbox": [ + 88, + 311, + 485, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 652, + 220, + 667 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional large-scale multi-modal models [2, 4, 17, 21, 41, 43, 46, 48, 58, 62, 64, 86] typically use multiple encoders and decoders to process multi-modal data. This approach makes learning a unified joint representation of the multi-modal data difficult and can significantly slow inference time (as shown in Fig. 1A). To alleviate these problems, end-to-end models without modal-fusion en(de)coder architecture have been proposed (as shown in Fig. 1B). This approach offers a streamlined, unified processing framework that enhances efficiency and consistency in multimodal representation learning. Existing end-to-end models follow three primary strategies: (1) Autoregressive models [5, 31, 69, 71] leverage a single Transformer for both text and image generation, but struggle with the inherent sequential dependency of autoregressive decoding. (2) Hy", + "bbox": [ + 89, + 672, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/69bf2ddb13c63877bd884b7241cb6f3ce9497cdd50221666359c3fee6333d1ed.jpg", + "image_caption": [ + "(A) Traditional Multi-Modal Model" + ], + "image_footnote": [], + "bbox": [ + 517, + 276, + 645, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/93064fba34e137e7c3670503d22b47d150f89cf7ea5f93a5cdb1197d8ccd1e28.jpg", + "image_caption": [ + "(B) End-To-End Multi-Modal Model" + ], + "image_footnote": [], + "bbox": [ + 645, + 276, + 776, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/68c2974f2548c1b20ccbd9c5d1a25a22dfe803286cf6563818771aed8a644328.jpg", + "image_caption": [ + "(C) Ours", + "Figure 1. Comparison of three types of models." + ], + "image_footnote": [], + "bbox": [ + 777, + 276, + 903, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "brid image generation models [23, 78] integrate an additional image synthesis module, improving image quality but introducing extra complexity. (3) Mixed autoregressive-diffusion models [89, 90] employ diffusion-based image generation while maintaining an autoregressive framework for text, yet still struggles with unifying multi-modal.", + "bbox": [ + 511, + 455, + 906, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite recent advancements, Transformer-based end-to-end models face several critical challenges: (1) their quadratic computational complexity makes them inefficient for generating high-resolution image and long-sequence text. Although various studies have attempted to optimize this computational complexity [1, 3, 12, 27, 29, 56, 59, 66, 74, 75], the challenge remains substantial. (2) their reliance on multi-objective learning introduces conflicting optimization goals, impeding convergence and hindering effective joint representation learning. In contrast, statespace models like Mamba [26, 61] offer a compelling alternative due to their ability to scale linearly with sequence length while effectively capturing long-range dependencies. However, the current multi-modal implementations of Mamba [18, 22, 30, 37, 49, 60, 73, 76, 80, 82] still adopt a multi-objective approach, limiting their capacity for end-to-end joint representation learning.", + "bbox": [ + 511, + 550, + 908, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To effectively process multi-modal data, we propose an end-to-end model called the Multi-Modal Diffusion Mamba (MDM) (as shown in Fig. 1c). MDM first employs patchify [19] and embedding to pre-process multi-modal data. Then, it uses a variational autoencoder (VAE) [42] as a multi-modal encoder, which uniformly maps the multi", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20529", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/15d2a8e5c773c4f13c81be3b9ad8a35e1b558ae70f915c32dc1c8b0a3a6b323d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 90, + 903, + 188 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/52cba87c0cc5afb41e3e3c8706d008e2f63434b17c4ee4ed22fe46a8d9c49508.jpg", + "image_caption": [ + "Figure 2. Framework of Multi-Modal Diffusion Mamba. MDM first encodes inputs (caption, VQVAE-processed image, question) using VAE (a), while performing padding (class, diffusion timestep, token completion) and flatten operations (d, e). Next, data reconstruction is progressively completed via diffusion mamba operations (b), modeling images and text temporally through scanning processes (f, g) for efficient information selection (red boxes indicate selection). Selected data undergoes computation (i) guided by (h) within the Mamba-2 framework to update model parameters. Finally, the MDM output passes through the VAE decoder (c) to reconstruct real data." + ], + "image_footnote": [], + "bbox": [ + 96, + 190, + 903, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "modal data to a noisy latent space (as illustrated in Fig. 2a). MDM constructs a multi-step selection diffusion model based on the Mamba architecture as a uniform decoder for the rapid generation of multi-modal information.", + "bbox": [ + 89, + 440, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This decoder generates the target text or image step-by-step based on the diffusion process through the multi-step selection diffusion model (as shown in Fig. 2b). To enhance decoding speed, the decoder employs the Score Entropy Loss [50] as the objective function instead of Markov chain-based [35] methods for updating the network to handle multi-modal data throughout the diffusion process. The decoder comprises two components: an image and text scan switch, and a Mamba-2 block [26]. The text scan switch has two modes for sequence modeling (as shown in Fig. 2f), while the image scan switch has four, based on the settings of DiM [73] (as shown in Fig. 2e). The scan switches enable the model to capture sequential relationships across various temporal directions in the data. The selection state-space structure in Mamba then analyzes these sequential relationships within the current denoising step. This analysis guides the selection of relevant information to focus on and irrelevant information to ignore, effectively directing the model's denoising process at each step.", + "bbox": [ + 89, + 503, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since MDM unifies the modality encoder and decoder, the model is capable of generating an image and text simultaneously. For example, as shown in Fig. 2h, when generating an image of a dog alongside its description, the scan switch in the decoder first assesses whether the description contains conditions that necessitate image generation. If such conditions exist, the image scan switch is", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "activated. Consequently, the model directs its selection to the image patches corresponding to the dog during each denoising step. This targeted focus guides the model to effectively denoise relevant pixels while disregarding other areas of the image. A similar selection process is employed for text data. Ultimately, the data, once denoised via the t-step diffusion process, is reconstructed into authentic text (or an image) through the VAE decoder simultaneously. The main contributions of this paper are as follows.", + "bbox": [ + 511, + 440, + 906, + 575 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We introduce the Multi-Modal Diffusion Mamba (MDM), an end-to-end model that achieves a computational complexity of $\\mathcal{O}(MLN^2)$ , outperforming previous end-to-end models like MonoFormer [89], which operate at $\\mathcal{O}(ML^2 N / G)$ . This advancement enables the efficient generation of long-sequence text and high-resolution images.", + "2) We propose a novel multi-step selection diffusion model that combines autoregressive and diffusion-based generative paradigms into a unified learning objective. This method effectively integrates both paradigms within a diffusion process, generating multi-modal data simultaneously.", + "3) Our experimental results demonstrate MDM's superior performance in image generation on the ImageNet [13] and COCO datasets [40]. Additionally, it excels in various tasks, including image captioning on Flickr30K [84] and COCO [40], VQA on VQAv2 [25], VizWiz [28], and OKVQA [53], as well as text comprehension and reasoning on seven datasets [7, 9, 10, 54, 65, 87]. Furthermore, MDM shows strong results in math-related world knowledge tasks on GSM8k [11], MATH [33], and MMLU [32]." + ], + "bbox": [ + 511, + 584, + 908, + 906 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20530", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 89, + 240, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Traditional large multi-modal model", + "text_level": 1, + "bbox": [ + 89, + 108, + 406, + 123 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most existing LMMs are built by integrating architectures from multiple modalities [2, 4, 41, 46, 62, 86]. SOTA image and video generation models employ pre-trained text encoders to represent input prompts in latent space, which then condition a diffusion model for generating videos and images [64]. Many researchers have adopted this approach, fusing feature representations from multiple pretrained encoders to enhance model performance across different modalities [21, 58]. This pattern is also prevalent in visual language models, where pre-trained language models are typically augmented with linear projection layers from other pre-trained en/decoders for training in the text space. Examples include Flamingo [2] and LLaVA [48] for visual understanding, GILL [43] for visual generation, and DreamLLM [17] for both understanding and generation.", + "bbox": [ + 89, + 125, + 483, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. End-to-End multi-modal model", + "text_level": 1, + "bbox": [ + 89, + 352, + 369, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "End-to-end models have emerged recently to facilitate joint representation learning while improving training and inference efficiency. It can be categorized into three main types: 1) The autoregressive model [5, 31, 69, 71] utilizes one Transformer with an autoregressive approach to generate images and text. For instance, the Fuyu model [5] processes image patches directly as input to achieve visual comprehension. Models like Chameleon [71], Mars [31], and LlamaGen [69] convert images into discrete sequence tokens, then concatenate them with text.", + "bbox": [ + 89, + 367, + 482, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2) The hybrid image generation model [23, 78] addresses the limitations of autoregressive approaches in image generation. While maintaining an autoregressive structure for text generation, the models enhance image quality by incorporating an image-generation network. For example, Seedx model [23] focuses on enhancing specific aspects of image generation, while Next-GPT [78] aims to expand multimodal capabilities within an end-to-end framework.", + "bbox": [ + 89, + 518, + 483, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3) The mixed autoregressive-diffusion model [89, 90] combines the strengths of previous approaches. It performs text autoregressive generation and image diffusion restoration simultaneously. Models like MonoFormer [89] and Transfusion [90] achieve this by incorporating causal self-attention [81] for text tokens and bidirectional self-attention [14] for image patches, enabling high-quality multi-modal understanding and generation.", + "bbox": [ + 89, + 640, + 483, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Mamba in multi-modal model", + "text_level": 1, + "bbox": [ + 89, + 763, + 357, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mamba has emerged as a powerful alternative to Transformer for multi-modal data alignment [18, 49, 76, 77, 82]. Recent works showcase Mamba's capabilities across different multi-modal applications. VL-Mamba [60] combines a pre-trained Mamba model for language understanding with a connector module to align visual patches and language tokens. However, these models lack end-to-end training capabilities and struggle to learn unified joint representation", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tions. MDM provides a truly end-to-end architecture, enabling rapid generation of high-quality, long sequences.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Multi-step Selection Diffusion Model", + "text_level": 1, + "bbox": [ + 511, + 125, + 846, + 142 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The multi-step selection diffusion model enables rapid generation of multi-modal information through two key processes: diffusion & denoising and selection. During the diffusion & denoising, the model employs a unified Score Entropy Loss [50](SE) to gradually reconstruct target data from noise through a series of denoising steps (as illustrated in Fig. 2b). The selection process enables the model to capture sequential relationships across different temporal dimensions in the latent space, determining which information should be focused on or ignored during each diffusion denoising step (as shown in Fig. 2h).", + "bbox": [ + 511, + 143, + 906, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion & Denoising", + "text_level": 1, + "bbox": [ + 511, + 311, + 720, + 327 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The diffusion & denoising process comprises two main components: diffusion and denoising. The diffusion component can be expressed by the following equation:", + "bbox": [ + 511, + 328, + 905, + 375 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n, t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t} ^ {g}} \\epsilon_ {n, t} ^ {g}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 380, + 906, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $g$ denotes either image patch or text embedding, and $z_{n,0}^{g}$ represents the latent space vector of the $n$ -th image patch or text embedding, obtained through VAE sampling [42]. $z_{n,t}^{g}$ is derived from $z_{n,0}^{g}$ after $t$ steps of noise addition; $\\epsilon_{n,t}^{g} \\sim \\mathcal{N}(0, I)$ represents the added noise; $\\bar{\\alpha}_{t}^{g} = \\prod_{k=1}^{t} \\alpha_{k}^{g}$ , $\\alpha_{k}^{g} = 1 - \\beta_{k}^{g}$ , and $\\{\\beta_{k}^{g} \\in (0,1)\\}_{k=1}^{T}$ are Gaussian distribution hyperparameters controlling the forward diffusion noise. Following the diffusion Markov principle [35], $t$ -step forward diffusion process can be characterized by conditional probabilities as follows:", + "bbox": [ + 511, + 412, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np (z _ {n, t} ^ {g} | z _ {n, 0} ^ {g}) = \\mathcal {N} (z _ {n, t} ^ {g}; \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g}, (1 - \\bar {\\alpha} _ {t} ^ {g}) I), \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 574, + 906, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which means that given $z_{n,0}^{g}, z_{n,t}^{g}$ follows a Gaussian distribution with $\\sqrt{\\bar{\\alpha}_t^g} z_{n,0}^g$ as mean and $(1 - \\bar{\\alpha}_t^g)I$ as variance.", + "bbox": [ + 511, + 606, + 903, + 640 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the classic diffusion denoising component [35], the model needs to learn the posterior $p(z_{n,t-1}^g | z_{n,t}^g)$ to gradually reconstruct the data. Since $p(z_{n,t}^g | z_{n,0}^g)$ follows a Gaussian distribution, we can assume that the approximate distribution of the denoising process is:", + "bbox": [ + 511, + 641, + 905, + 715 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} (z _ {n, t - 1} ^ {g} | z _ {n, t} ^ {g}) = \\mathcal {N} (z _ {n, t - 1} ^ {g}; \\mu_ {\\theta} (z _ {n, t} ^ {g}), (\\sigma_ {\\theta , n} ^ {g}) ^ {2}). \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 722, + 906, + 742 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(z_{n,t}^{g})$ and $\\sigma_{\\theta ,n}^{g}$ represent the model predicted noise mean and variance at the $t$ -th denoising step.", + "bbox": [ + 511, + 750, + 905, + 779 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This method achieves the gradual recovery of data by optimizing the conditional probability of each time step by maximum likelihood. However, Markov chain-based [35] methods limit computational efficiency in high-dimensional spaces and are difficult to extend to discrete data.", + "bbox": [ + 511, + 780, + 905, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To further optimize the denoising process, this paper uses SE [50] as the optimization target. It is a generalized score matching objective that aims to directly learn the", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20531", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "probability density ratio between discrete states. The SE can not only stabilize the diffusion denoising process but also improve the sampling quality through the global information of data distribution. In general form, for any state pair $(z_{n,t}^{g}, z_{n,0}^{g})$ , define the model's score ratio $s_{\\theta}(z_{n,t}^{g})$ which represents the relative probability of transferring from $z_{n,t}^{g}$ to $z_{n,0}^{g}$ . SE is defined as:", + "bbox": [ + 89, + 90, + 483, + 198 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} s e = \\sum_ {y \\in z _ {n, 0: t - 1} ^ {g}} \\omega_ {z _ {n, t} ^ {g}} ^ {g} \\left(s _ {\\theta} (z _ {n, t} ^ {g}) - \\frac {p _ {d a t a} (y)}{p _ {d a t a} (z _ {n , t} ^ {g})} \\log s _ {\\theta} (z _ {n, t} ^ {g}) \\right. \\\\ + K \\left(\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)}\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 204, + 480, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega_{z_{n,t}}^{g}$ is the weight of the loss term, which is used to balance the loss of different states. $K(a) = a(\\log a - 1)$ is a normalization term that ensures the loss is non-negative. $\\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}$ represents the actual score ratio. $p_{data}(y)$ and $p_{data}(z_{n,t}^{g})$ are the actual data distributions of the former noisy state and the current noisy state. The actual score ratio calculation relationship is shown in Theorem 1.", + "bbox": [ + 89, + 305, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 1. According to Bayes' theorem and the Gaussian distribution density formula, the following calculation relationship of $\\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}$ is obtained:", + "bbox": [ + 89, + 428, + 483, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)} = \\exp \\left(\\frac {\\| z _ {n , t} ^ {g} \\| ^ {2}}{2} - \\frac {\\| z _ {n , t} ^ {g} - \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n , 0} ^ {g} \\| ^ {2}}{2 \\left(1 - \\bar {\\alpha} _ {t} ^ {g}\\right)}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 484, + 480, + 539 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is provided in Appendix A.", + "bbox": [ + 109, + 545, + 361, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the SE [50], the model predicted score ratio indicates how the model adjusts the probability of the current state to tend to the original data distribution during the denoising process. The definition is as follows:", + "bbox": [ + 89, + 560, + 483, + 619 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {p _ {\\theta} \\left(z _ {n , 0} ^ {g}\\right)}{p _ {\\theta} \\left(z _ {n , t} ^ {g}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 214, + 626, + 482, + 662 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the denominator represents the probability of the current noise state and the numerator represents the original state probability estimated by the model. According to Theorem 2, the model uses softmax for normalization ensuring numerical stability and enabling gradient optimization when predicting the score ratio.", + "bbox": [ + 89, + 667, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 2. Given the denoising process modelled by a score-based probability ratio function $s_{\\theta}(z_{n,t}^{g})$ , defined as Eq. (6), this paper defines a learnable approximation using a parameterized score function $f_{\\theta}$ , such that the probability ratio can be estimated as:", + "bbox": [ + 89, + 763, + 483, + 839 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {\\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , z _ {n , 0} ^ {g}\\right)\\right)}{\\sum_ {y \\in z _ {n , 0 : t - 1} ^ {g}} \\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , y\\right)\\right)}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 844, + 482, + 883 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is provided in Appendix A.", + "bbox": [ + 109, + 888, + 362, + 905 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Selection", + "text_level": 1, + "bbox": [ + 513, + 90, + 619, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection process comprises two key steps: scan switch and selection. The scan switch mechanism captures temporal relationships between adjacent image patches (or text embeddings) by generating latent space representations with $k$ different sequential relationships, such as four image patch sequences and two text embedding sequences illustrated in Fig. 2fg. The mechanism creates $k$ temporal sequences $S = \\{\\langle z_{1,t}^{g},z_{2,t}^{g},\\dots ,z_{i,t}^{g}\\rangle \\}_{k}$ .", + "bbox": [ + 511, + 106, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The selection step then analyzes these different sequential relationships at the current denoising step $t$ to determine which information should be focused on or ignored, thereby guiding the model's denoising direction in each diffusion step. The selection step chooses $j$ items $z_{n,t}^{g}$ from each sequence in $S$ according to the following Theorem 3. So, the selection step obtain $k$ selection sequences with different lengths, i.e., $S' = \\{\\langle z_{j_1,t}^g, z_{j_2,t}^g, \\ldots, z_{j,t}^g \\rangle\\}_k$ and $S' \\in S$ .", + "bbox": [ + 511, + 227, + 905, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3. To achieve the optimal score entropy [50] which is demonstrated on Eq. (4), the selection step choose $j$ items where each $z_{n,t}^{g}$ satisfies $se = 0$ , i.e.,", + "bbox": [ + 511, + 354, + 906, + 401 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) \\approx \\frac {p _ {d a t a} (y)}{p _ {d a t a} \\left(z _ {n , t} ^ {g}\\right)} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 410, + 903, + 444 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof is provided in Appendix A.", + "bbox": [ + 532, + 450, + 784, + 465 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Architecture", + "text_level": 1, + "bbox": [ + 511, + 470, + 643, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The neural network architecture consists of two primary components: a VAE noisy latent encoder [42] and a multistep selection diffusion decoder, as illustrated in Fig. 2ab. The encoder first processes image data $X_{img}$ through patchify [19] operations and processes text data $X_{txt}$ through tokenization based on SentencePiece with Unigram BPE [45] and embedding operations, then uniformly maps them to the latent space before applying forward noise.", + "bbox": [ + 511, + 489, + 905, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The decoder, based on the multi-step selection diffusion model, leverages Mamba to achieve unified learning objectives while enhancing computational efficiency for processing long sequence data. It employs the SE [50] as the unified objective for both image and text modalities during the diffusion process. During selection, the model captures sequential relationships across different temporal dimensions using various scan switches. These relationships are then efficiently processed through the selection state-space structure in the Mamba Block determining which information to focus on or ignore according to Eq. (8), thereby guiding subsequent diffusion denoising steps (as shown in Fig. 2h). Finally, the reconstructed image patches and text embeddings are transformed back into their original data formats through a VAE noisy latent decoder [42].", + "bbox": [ + 511, + 611, + 905, + 837 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. The noisy latent encoder", + "text_level": 1, + "bbox": [ + 511, + 839, + 738, + 854 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The noisy latent encoder first processes input image $X_{img}$ through patchify and processes text $X_{txt}$ through tokenization and embedding operations to obtain the patch sequence", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20532", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$G(X_{img} / X_{txt}) = \\langle g_1, g_2, \\dots, g_i \\rangle$ , where $g_n$ represents the $n$ -th image patch or text embedding, respectively. The encoder VAE [42] generates Gaussian distribution parameters (mean $\\mu$ and variance $\\sigma$ ) for these patches, with a similar process applied to text embeddings, i.e., $VAE(G) = (\\mu, \\sigma)$ . For each image patch or text embedding $g_n$ , its noise $z_n$ is a sample $s_n$ from the distribution $\\mathcal{N}(\\mu, \\sigma)$ with the addition noise $\\epsilon_n \\sim \\mathcal{N}(0, 1)$ , i.e., $z_n = s_n + \\epsilon_n$ . Finally, the image $X_{img}$ and text $X_{txt}$ are transformed into the noise sequence $\\langle z_1, \\dots, z_i \\rangle$ through the above process.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Moreover, three types of learnable padding tokens, time, category, and pad, are inserted into these noise sequences, as illustrated in Fig. 2de. The time token encodes the current diffusion step, the class token is used to learn the data category, and the pad token represents the start or end position for splitting these noise sequences.", + "bbox": [ + 89, + 243, + 483, + 335 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. The multi-step selection diffusion decoder", + "text_level": 1, + "bbox": [ + 89, + 340, + 447, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The decoder aims at progressively recovering the image $X_{img}$ or text $X_{txt}$ from noise sequences through two main modules: 1) the multi-step selection diffusion Mamba and 2) the VAE noisy latent decoder. 1) The Mamba is used to recover the patch sequence $\\langle g_1, \\dots, g_i \\rangle$ from the noise sequence $\\langle z_1, \\dots, z_i \\rangle$ . 2) The VAE noisy latent decoder assembles patches and generates the image $\\hat{X}_{img}$ or text $\\hat{X}_{txt}$ .", + "bbox": [ + 89, + 358, + 483, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.1. Multi-step selection diffusion Mamba", + "text_level": 1, + "bbox": [ + 89, + 469, + 397, + 483 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The module leverages two components, image/text scan switch and Mamba Block, to implement each denoising step in the multi-step selection diffusion model (Sec. 3).", + "bbox": [ + 89, + 483, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The image/text scan switch component establishes sequences with different directions to capture different temporal relationships between patches. Following Dim [73], we implement four distinct scan switches for images (as shown in Fig. 2f) and two for text (as shown in Fig. 2g).", + "bbox": [ + 89, + 530, + 483, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The Mamba block is used to select patches from these different scan switch sequences and denoise the input noise $z_{n,t}^{g}$ . The block adopts the state space architecture from Mamba-2 [26]. According to Sec. 3.2, it is $s_{\\theta}$ , where $\\theta = \\{H_{n,t}^{g}, A, B, C, D, \\Delta\\}$ represent the state space in the block. The block comprises six key components: 1) linear input and output projection layers, 2) convolution kernel layer, 3) nonlinear activation layer, 4) state space model (SSM), 5) skip connection layer, and 6) normalization layer.", + "bbox": [ + 89, + 606, + 483, + 742 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "1) The linear input projection layer reduces the dimensionality of the latent space noise vector while simultaneously applying initial state matrices $A$ , $B$ , $C$ to the linear projection of input data $z_{n,t}^{g}$ . Additionally, the linear output projection layer represents the denoising step, which transforms the selection noise $z_{n,t}^{g}$ into $z_{n,t - \\Delta t}^{g}$ and outputs it to the next Mamba block according to the following equation.", + "bbox": [ + 89, + 743, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n, t - \\Delta t} ^ {g} = z _ {n, t} ^ {g} - \\frac {\\Delta t}{2} \\left[ f _ {\\theta} \\left(z _ {n, t} ^ {g}, t\\right) + f _ {\\theta} \\left(z _ {n, t - \\Delta t} ^ {g}, t - \\Delta t\\right) \\right] (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 875, + 483, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the equation adopts the second-order numerical method of DPM-Solver [51] to improve sampling accuracy. Details are provided in Appendix B.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "2) The convolution kernel layer implements parallel scan switches, routing the initial linear projection of the input and the state matrix's linear projection through the SSM, as shown in Fig. 2i. The sweep down and sweep up [26] enable parallel computation between Eqs. (10) to (13).", + "3) The nonlinear layer enhances model generalization.", + "4) The SSM lets the Mamba block $s_{\\theta}$ approximate the actual score ratio based on Theorem 3. To implement the target, SSM updates the state space $\\theta$ by the following equations (based on Theorem 3 and details in Appendix A)." + ], + "bbox": [ + 511, + 136, + 905, + 287 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH _ {n, t} ^ {g} = \\bar {A} H _ {n, t - 1} ^ {g} + \\bar {B} z _ {n, t} ^ {g} \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 301, + 903, + 320 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n - 1, t} ^ {g} = C H _ {n, t} ^ {g} + D z _ {n, t} ^ {g} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 339, + 903, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {A} = \\exp (\\Delta A) \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 656, + 378, + 903, + 395 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {B} = (\\Delta A) ^ {- 1} \\cdot \\left(\\exp (\\Delta A) - I\\right) \\cdot \\Delta B \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 416, + 903, + 433 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $H_{n,t}^{g}$ represents the hidden state representation, $A$ and $B$ control the evolution of hidden states and latent space noise vector inputs, respectively, $C$ governs the hidden state representation of the target output and $D$ manages the nonlinear skip connection for latent space noise vector inputs. $\\Delta$ denotes the learnable time parameter.", + "bbox": [ + 511, + 439, + 905, + 530 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "5) The skip connection layer facilitates input feature reuse and mitigates model degradation.", + "6) The Normalization layer ensures training stability.", + "According to Eq. (8) in Theorem 3 and Eq. (4), the goal of training the Mamba block is:" + ], + "bbox": [ + 511, + 531, + 903, + 606 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s e} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim p _ {0}, z _ {n} ^ {g} \\sim p (\\cdot | z _ {n, 0} ^ {g})} s e = 0 \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 619, + 903, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2.2. The noisy latent decoder", + "text_level": 1, + "bbox": [ + 511, + 645, + 730, + 660 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After applying the diffusion-based denoising process, the recovered latent variable $z_{n,0}^{g}$ is passed to the VAE decoder [42] as illustrated in Fig. 2c. For image reconstruction, the decoder applies an $\\ell_2$ loss:", + "bbox": [ + 511, + 665, + 905, + 726 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c} ^ {i m g} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\| X _ {i m g} - \\hat {X} _ {i m g} \\| ^ {2}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 736, + 903, + 756 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $q_{\\phi}(z|X)$ represents the posterior distribution of the VAE encoder.", + "bbox": [ + 511, + 768, + 903, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For text, the decoder minimizes the cross-entropy loss:", + "bbox": [ + 532, + 797, + 893, + 813 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {r e c} ^ {t x t} = - \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\sum_ {t} p \\left(X _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right) \\log p _ {\\psi} \\left(\\hat {X} _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right). \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 824, + 908, + 867 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $p(X_{txt}^{(t)}|z_{n,0}^{g})$ represents the probability distribution of real text data under the condition of latent variable $z_{n,0}^{g}$ .", + "bbox": [ + 511, + 868, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20533", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/953d44ded220231e3b4ec9deadbc4fb7cc804a29d77460b4fb29182ebc1593d1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 480, + 238 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d18d89bf12c3d61711c87fa4e2fa0800653c74b98ff3378c3001a1654b9e784a.jpg", + "image_caption": [ + "Figure 3. VQA, QA and Multi-Modal generation test from MDM. The results of VQA are part of VQAv2 [25]. The QA results are part of PIQA [7] and MMLU [32]. The Multi-Modal generation results are tested with ground-truth data.", + "Figure 4. Comparison between each model on generating captioning and image results on COCO dataset. Unlike other models, MDM generates both image and caption data simultaneously." + ], + "image_footnote": [], + "bbox": [ + 94, + 309, + 480, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "And $p_{\\psi}(\\hat{X}_{txt}^{(t)}|z_{n,0}^g)$ represents the probability distribution of the text token generated by the VAE decoder under the condition of the latent variable $z_{n,0}^g$", + "bbox": [ + 89, + 594, + 482, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, a KL divergence regularizes the latent space:", + "bbox": [ + 109, + 643, + 465, + 657 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {K L} = D _ {K L} \\left(q _ {\\phi} (z | X) \\| p (z)\\right). \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 670, + 482, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $p(z)$ represents the prior distribution of the latent variable by VAE, which is assumed to be a standard Gaussian distribution $\\mathcal{N}(0,I)$ to regularize the latent variable space and enable it to have smooth generation capabilities.", + "bbox": [ + 89, + 699, + 482, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final optimization objective integrates VAE reconstruction, KL divergence and SE:", + "bbox": [ + 89, + 760, + 482, + 789 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {r e c} ^ {i m g} + L _ {r e c} ^ {t x t} + \\beta L _ {K L} + \\lambda L _ {s e}. \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 801, + 482, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 829, + 223, + 845 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 849, + 279, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model configuration. Our model applies a VAE [42] as the noisy latent encoder and decoder. Moreover, it integrates", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the DiM selection state space [73] in each Mamba block as the diffusion decoder. The resulting model contains 7 billion parameters, with 49 Mamba blocks in the multi-step selection diffusion decoder, each having a dimension of 2048 (Details of parameter settings listed in Appendix C).", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Before the training MDM process, we trained a tokenization model based on SentencePiece (Unigram BPE) [45]. The tokenization model can help the model construct a stable text latent variable representation, thereby optimizing the forward diffusion and reverse denoising process. See Appendix D for detailed experimental settings.", + "bbox": [ + 511, + 167, + 903, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In the training process, we import the DDPM scheduler [35] and DPM-Solver [51] to improve the sampling efficiency in the diffusion model. We then use the AdamW optimizer without weight decay, maintaining a constant learning rate of 0.0001. Meanwhile, we keep an EMA of the model weights with a coefficient of 0.9999.", + "bbox": [ + 511, + 260, + 905, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baseline and dataset. Our evaluation encompasses four tasks: image generation with classifier-free guidance [34] (CFG), text-to-image, image-to-text, and text-to-text generation. For the baseline model training, we train MDM on ImageNet [13], JourneyDB [68] and UltraChat [16].", + "bbox": [ + 511, + 352, + 905, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the image generation and the text-to-image task at $256 \\times 256$ resolution, we compare the MDM baseline model against established baselines across three categories: diffusion models (Imagen [64], ADM [15], CDM [36], LDM [63], DiT-XL/2 [57], SDXL [58], and SD-3 [21]), autoregressive models (VQGAN [20] and ViT-VQGAN [85]), and end-to-end multi-modal models (NExT-GPT [78], Chameleon [71], LlamaGen [69], Transfusion [90], MonoFormer [89], Dual-DiT [47], JanusFlow [52] and ShowO [79]). For the image generation task, we evaluate performance on ImageNet [13] using four metrics: Frechet Inception Distance (FID), Inception Score (IS), and Precision/Recall. For the text-to-image task, we evaluate performance on COCO [40] using FID and Gen Eval [24].", + "bbox": [ + 511, + 430, + 905, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the image-to-text task (image captioning and vision question answering, VQA) and text-to-text task, we employ MDM baseline model and MDM instruction model by visual instruction tuning [48] on multiple datasets: COCO [40], GQA [38], OCR-VQA [55], TextVQA [67], and VisualGenome [44]. We evaluate the model against two groups of baselines: traditional models and end-to-end multi-modal models. Performance evaluation of image captioning is conducted on Flickr 30K [84] and COCO [40] datasets using the Consensus-based Image Description Evaluation (CIDEr) metric. And performance evaluation of VQA is conducted on VQAv2 [25], VizWiz [28], and OKVQA [53] using answer accuracy rate as the evaluation metric.", + "bbox": [ + 511, + 642, + 905, + 852 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the text-to-text task, we evaluate the model on text comprehension and reasoning tasks using HellaSwag [87], OpenBookQA [54], Wino-Grande [65], ARCEasy, ARC", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20534", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b14bce0ea13faa1f24f79fe8c13237667ea93db9215c6b87490ce7345a5cf576.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelArcParamsImage Generation with CFGText-to-Image Generation
FID ↓IS ↑Pre ↑Re ↑FID ↓Gen Eval ↑
Imagen [64]Diff7.3B----7.27-
ADM [15]Diff554M10.94101.00.690.63--
CDM [36]Diff-4.88158.7----
LDM [63]Diff400M3.60147.60.870.68-0.43
DiT-XL/2 [57]Diff675M2.27278.20.830.57--
SDXL [58]Diff3.4B----4.400.55
SD-3 [21]Diff12.7B-----0.68
VQGAN [20]AR227M18.6580.40.780.26--
ViT-VQGAN [85]AR1.7B4.17175.1----
NExT-GPT [78]AR7B----10.07-
Chameleon [71]AR7B----26.740.39
LlamaGen [69]AR3.1B2.81311.50.840.544.19-
Transfusion [90]AR+Diff7.3B----6.780.63
MonoFormer [89]AR+Diff1.1B2.57272.60.840.56--
Dual-DiT [47]Diff2B----9.400.65
JanusFlow [52]AR+Diff1.3B-----0.70
Show-O [79]AR+Diff1.3B----9.240.68
MDMDiff7B2.49281.40.860.595.910.68
", + "bbox": [ + 91, + 88, + 906, + 378 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1edc2bbedcc758710c393608f62476f248d17b216bfbcc796f3980b9fd8c42d1.jpg", + "table_caption": [ + "Table 1. Performance on ImageNet and COCO ${256} \\times {256}$ . FID,IS,Pre,and Re stands for Frechet Inception Distance,Inception Score, Precision, and Recall, respectively." + ], + "table_footnote": [], + "table_body": "
ModelICVQAText Comprehension and ReasoningMath and World
FlickrCOCOVQAv2VizWizOKHSOBQAWGARCEARCCBoolQPIQAGSM8kMATHMMLU
Llama-2 [74] (7B)-----77.258.678.575.245.977.478.814.62.545.3
Mistral [39] (7B)-----81.3-75.380.055.584.783.052.113.160.1
Flamingo [2] (80B)75.1113.867.6------------
Gemini Pro [72]82.299.871.2--84.7------86.532.671.8
GPT4V [8]55.378.577.2--95.3------92.052.986.4
InstructBLIP [48] (7B)82.4102.2-33.433.9----------
mPLUG-Owl [83] (7B)80.3119.3-39.0-----------
TinyLlama [88] (1.1B)-----59.236.059.155.330.157.873.3---
Pythia [6] (12B)-----52.033.257.454.028.563.370.9---
DREAMLLM [17](7B)-115.456.645.844.3----------
Emu [70](7B)-117.740.035.434.7----------
Chameleon [71](34B)74.7120.266.0--74.251.070.476.146.581.479.641.611.552.1
NExT-GPT [78](7B)84.5124.966.748.452.1----------
Transfusion [90](7B)-33.7-------------
MonoFormer [89](1.1B)-----50.637.256.948.231.562.371.2---
Dual-DiT [47](2B)-56.260.129.925.3----------
JanusFlow [52](1.3B)--79.8------------
Show-O [79](1.3B)67.6-74.7------------
MDM (7B)62.4109.660.339.847.170.641.568.855.146.265.779.940.512.154.4
InstructMDM (7B)75.2122.166.746.351.674.848.374.965.447.171.583.746.013.159.2
", + "bbox": [ + 91, + 422, + 908, + 732 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Performance on image-to-text and text-to-text tasks. The evaluation of image captioning (IC) and VQA is CIDEr and answer accuracy % (Flickr is evaluated on 30K and OK represents OKVQA).", + "bbox": [ + 89, + 743, + 906, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Challenge [10], BoolQ [9], and PIQA [7]. We also evaluate the model on math and world knowledge tasks using GSM8K [11], MATH [33], and MMLU [32]. The evaluation metrics for all the tasks are accuracy rates.", + "bbox": [ + 89, + 781, + 480, + 842 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 848, + 292, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Image Generation. In the image generation task on ImageNet, MDM achieves top-three rankings across all eval", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "uation metrics: second in FID, IS, and Precision, and third in Recall when compared against one-modal diffusion models and end-to-end multi-modal models (see Tab. 1). MDM demonstrates superior overall performance, notably surpassing other end-to-end multi-modal models in three of the four metrics. In the text-to-image task, we tested the model on the COCO dataset to generate both image and", + "bbox": [ + 511, + 781, + 906, + 888 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20535", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de077a502ef4c2b76e1729731b18c36fbbfd4c0b9fe749887897a94cf1ad89c3.jpg", + "image_caption": [ + "Figure 5. Comparison between Mamba Baseline, MonoFormer, and ours MDM on inference speed test. The left shows the inference speed of the model FPS at different resolutions. The right shows the inference speed of the model latency at different token lengths." + ], + "image_footnote": [], + "bbox": [ + 174, + 88, + 496, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a8790f8903f155935ff66a3aae2ea185867caf22ff191eac4dba5c23dbcb7b2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 498, + 88, + 821, + 222 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d03eeac77510c16f6fe0d7736c7a175436383b3257c0048d9640cf7d9628ed57.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelImage/Text \nScan SwitchFPS w log \nscale↑FID↓
Model w Mamba①②③④/①②1.3572.49
Model w Mamba①②/①1.4053.96
Model w Transformer-1.9146.72
", + "bbox": [ + 91, + 273, + 483, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation on ImageNet ${256} \\times {256}$ image generation.", + "bbox": [ + 107, + 357, + 464, + 372 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "caption data. For the image generation results, we evaluated the FID and Gen Eval performance indicators of the model-generated images. MDM still achieved the top three performance levels and achieved SOTA on Gen Eval.", + "bbox": [ + 89, + 390, + 482, + 450 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Text Generation. In the image-to-text task image captioning, according to the settings on image generation on the COCO dataset, we tested the caption data of the model based on the model outputting both text and image data using the CIDEr indicator. The results showed that MDM ranked second among all models, as shown in Tab. 2. While in task VQA, MDM achieves competitive performance, surpassing several traditional models including InstructBLIP, mPLUG-Owl, DREAMLLM, and Emu, although it still trails behind top-performing models in the field as shown in Tab. 2. In the text-to-text generation task, as shown in Tab. 2, MDM and the other end-to-end multi-modal models perform worse than well-known traditional models. This discrepancy may be attributed to the fact that these end-to-end models have some deviations in multimodal fusion and learning because they abandon multiple language encoders, visual encoders, and multimodal fusion encoders. However, when compared with the other two end-to-end models, MDM excels, outperforming MonoFormer and surpassing Chameleon on seven out of ten datasets.", + "bbox": [ + 89, + 450, + 482, + 752 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Discussion", + "text_level": 1, + "bbox": [ + 89, + 757, + 209, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.1. Performance Analysis", + "text_level": 1, + "bbox": [ + 89, + 776, + 290, + 791 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As demonstrated in Fig. 3, MDM shows the ability to generate image and text simultaneously in multiple rounds of dialogue and perform well in QA&VQA. Some results even exceed those of GPT-4V, particularly evident in the second and third rows of Fig. 4 which is a hybrid output process for the MDM model. Due to this, we set the model to generate corresponding images for the description text while simul", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "taneously generating image captioning.", + "bbox": [ + 511, + 277, + 772, + 291 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This enhanced performance stems from MDM's multi-step selection diffusion decoder, which leverages Mamba's integrated selection and denoising capabilities to maintain focused attention on both textual and visual details. Validating our complexity analysis in Appendix E, MDM demonstrates superior efficiency compared to end-to-end Transformer models when processing long sequences, as shown in Fig. 5, particularly outperforming other end-to-end multimodal models for sequences exceeding 1280 tokens.", + "bbox": [ + 511, + 292, + 906, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.2. Ablations", + "text_level": 1, + "bbox": [ + 511, + 431, + 625, + 443 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation studies examine the impact of both the selection process and Mamba block components. Reducing the number of image/text scan switch sequences from 6 ('①②③④/①②') to 3 ('①②/①'), as shown in Tab. 3, improves inference speed but degrades image quality, as fewer scan switch sequences limit the model's ability to capture accurate information in complex sequences. Additionally, replacing the Mamba block with the Transformer further deteriorates output image quality, suggesting Mamba's temporal network architecture is better suited for representing diffusion relationships during the denoising process.", + "bbox": [ + 511, + 446, + 908, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 619, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces MDM (Multi-Modal Diffusion Mamba), a novel end-to-end architecture that significantly enhances multi-modal processing through two key innovations: a unified diffusion objective and an efficient selection mechanism leveraging Mamba's state-space structure. By integrating variational autoencoder with multi-step selection diffusion, MDM achieves SOTA overall performance in image generation and demonstrates remarkable versatility across various tasks, including image-to-text, text-to-text and text-image-to-text-image. Our comprehensive experiments illustrate that MDM consistently surpasses traditional end-to-end multi-modal models, particularly in processing high-resolution images and long-sequence text, while maintaining computational efficiency. The model's ability to unify different modalities under a single objective, coupled with its superior management of temporal relationships in the diffusion process, establishes a promising direction for future multi-modal architecture.", + "bbox": [ + 511, + 636, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20536", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 1", + "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1, 3, 7", + "[3] Silas Alberti, Niclas Dern, Laura Thesing, and Gitta Kutyniok. Sumformer: Universal approximation for efficient transformers. In Topological, Algebraic and Geometric Learning Workshops 2023, pages 72-86. PMLR, 2023. 1", + "[4] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. Advances in Neural Information Processing Systems, 35:32897-32912, 2022. 1, 3", + "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sağnak Taşırlar. Introducing our multimodal models, 2023. 1, 3", + "[6] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 7", + "[7] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, pages 7432-7439, 2020. 2, 6, 7", + "[8] GPTV System Card. Openai, 2023. 7", + "[9] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 2, 7", + "[10] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 2, 7", + "[11] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 7", + "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 6", + "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3", + "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 7", + "[16] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233, 2023. 6", + "[17] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3, 7", + "[18] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-mamba for cross-modality object detection. arXiv preprint arXiv:2404.09146, 2024. 1, 3", + "[19] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 4", + "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6, 7", + "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 7", + "[22] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Youqiang Zhang, and Junshi Huang. Dimba: Transformermamba diffusion models. arXiv preprint arXiv:2406.01159, 2024. 1", + "[23] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 3", + "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 6", + "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 6", + "[26] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 5" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20537", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Ahan Gupta, Yueming Yuan, Yanqi Zhou, and Charith Mendis. Flurka: Fast fused low-rank & kernel attention. arXiv preprint arXiv:2306.15799, 2023. 1", + "[28] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 2, 6", + "[29] Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David P Woodruff, and Amir Zandieh. Hyperattention: Long-context attention in near-linear time. arXiv preprint arXiv:2310.05869, 2023. 1", + "[30] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 1", + "[31] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 3", + "[32] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 2, 6, 7", + "[33] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021. 2, 7", + "[34] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6", + "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 6", + "[36] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 6, 7", + "[37] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 1", + "[38] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6", + "[39] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7", + "[40] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 2, 6" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1, 3", + "[42] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 3, 4, 5, 6", + "[43] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3", + "[44] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6", + "[45] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4, 6", + "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1, 3", + "[47] Zijie Li, Henry Li, Yichun Shi, Amir Barati Farimani, Yuval Kluger, Linjie Yang, and Peng Wang. Dual diffusion for unified image generation and understanding. arXiv preprint arXiv:2501.00289, 2024. 6, 7", + "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7", + "[49] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 1, 3", + "[50] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4", + "[51] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 5, 6", + "[52] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv preprint arXiv:2411.07975, 2024.6.7", + "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 2, 6", + "[54] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new" + ], + "bbox": [ + 514, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20538", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 2, 6", + "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6", + "[56] Matteo Pagliardini, Daniele Paliotta, Martin Jaggi, and François Fleuret. Fast attention over long sequences with dynamic sparse flash attention. Advances in Neural Information Processing Systems, 36:59808-59831, 2023. 1", + "[57] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 7", + "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 6, 7", + "[59] Ramya Prabhu, Ajay Nayak, Jayashree Mohan, Ramachandran Ramjee, and Ashish Panwar. vattention: Dynamic memory management for serving llms without pagedattention. arXiv preprint arXiv:2405.04437, 2024. 1", + "[60] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 1, 3", + "[61] Haohao Qu, Liangbo Ning, Rui An, Wenqi Fan, Tyler Derr, Xin Xu, and Qing Li. A survey of mamba. arXiv preprint arXiv:2408.01129, 2024. 1", + "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3", + "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7", + "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3, 6, 7", + "[65] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 2, 6", + "[66] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 1", + "[67] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[68] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. Advances in Neural Information Processing Systems, 36, 2024. 6", + "[69] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 3, 6, 7", + "[70] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 7", + "[71] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 3, 6, 7", + "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7", + "[73] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 1, 2, 5, 6", + "[74] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 7", + "[75] Zhongwei Wan, Xin Wang, Che Liu, Samiul Alam, Yu Zheng, Jiachen Liu, Zhongnan Qu, Shen Yan, Yi Zhu, Quanlu Zhang, et al. Efficient large language models: A survey. arXiv preprint arXiv:2312.03863, 2023. 1", + "[76] Zifu Wan, Pingping Zhang, Yuhao Wang, Silong Yong, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Sigma: Siamese mamba network for multi-modal semantic segmentation. arXiv preprint arXiv:2404.04256, 2024. 1, 3", + "[77] Xinghan Wang, Zixi Kang, and Yadong Mu. Text-controlled motion mamba: Text-instructed temporal grounding of human motion. arXiv preprint arXiv:2404.11375, 2024. 3", + "[78] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 1, 3, 6, 7", + "[79] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6, 7", + "[80] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8239-8249, 2024. 1", + "[81] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20539", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[82] Zhe Yang, Wenrui Li, and Guanghui Cheng. Shmamba: Structured hyperbolic state space model for audio-visual question answering. arXiv preprint arXiv:2406.09833, 2024. 1, 3", + "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7", + "[84] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 2, 6", + "[85] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 6, 7", + "[86] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 3", + "[87] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 2, 6", + "[88] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 7", + "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 1, 2, 3, 6, 7", + "[90] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 642 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20540", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_model.json b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b0ecbcb5470f3600f0b9a4c1dd2cb37a1a8f7bca --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_model.json @@ -0,0 +1,2941 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.278, + 0.131, + 0.72, + 0.151 + ], + "angle": 0, + "content": "End-to-End Multi-Modal Diffusion Mamba" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.18, + 0.744, + 0.2 + ], + "angle": 0, + "content": "Chunhao Lu1 Qiang Lu1⊗ Meichen Dong1,2 Jake Luo3" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.203, + 0.914, + 0.222 + ], + "angle": 0, + "content": "1China University of Petroleum-Beijing, 2Leyard Optoelectronic, 3University of Wisconsin-Milwaukee" + }, + { + "type": "text", + "bbox": [ + 0.151, + 0.228, + 0.864, + 0.243 + ], + "angle": 0, + "content": "{luchunhao,meichen.dong}@student.cup.edu.cn luqiang@cup.edu.cn jakeluo@uwm.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.278, + 0.327, + 0.294 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.312, + 0.486, + 0.645 + ], + "angle": 0, + "content": "Current end-to-end multi-modal models utilize different encoders and decoders to process input and output information. This separation hinders the joint representation learning of various modalities. To unify multi-modal processing, we propose a novel architecture called MDM (Multi-modal Diffusion Mamba). MDM utilizes a Mamba-based multistep selection diffusion model to progressively generate and refine modality-specific information through a unified variational autoencoder for both encoding and decoding. This innovative approach allows MDM to achieve superior performance when processing high-dimensional data, particularly in generating high-resolution images and extended text sequences simultaneously. Our evaluations in areas such as image generation, image captioning, visual question answering, text comprehension, and reasoning tasks demonstrate that MDM significantly outperforms existing end-to-end models (MonoFormer, LlamaGen, and Chameleon etc.) and competes effectively with SOTA models like GPT-4V, Gemini Pro, and Mistral. Our results validate MDM's effectiveness in unifying multi-modal processes while maintaining computational efficiency, establishing a new direction for end-to-end multi-modal architectures." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.654, + 0.222, + 0.669 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Traditional large-scale multi-modal models [2, 4, 17, 21, 41, 43, 46, 48, 58, 62, 64, 86] typically use multiple encoders and decoders to process multi-modal data. This approach makes learning a unified joint representation of the multi-modal data difficult and can significantly slow inference time (as shown in Fig. 1A). To alleviate these problems, end-to-end models without modal-fusion en(de)coder architecture have been proposed (as shown in Fig. 1B). This approach offers a streamlined, unified processing framework that enhances efficiency and consistency in multimodal representation learning. Existing end-to-end models follow three primary strategies: (1) Autoregressive models [5, 31, 69, 71] leverage a single Transformer for both text and image generation, but struggle with the inherent sequential dependency of autoregressive decoding. (2) Hy" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.277, + 0.646, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.426, + 0.635, + 0.434 + ], + "angle": 0, + "content": "(A) Traditional Multi-Modal Model" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.277, + 0.777, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.657, + 0.426, + 0.765, + 0.434 + ], + "angle": 0, + "content": "(B) End-To-End Multi-Modal Model" + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.277, + 0.905, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.427, + 0.855, + 0.434 + ], + "angle": 0, + "content": "(C) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.436, + 0.851, + 0.45 + ], + "angle": 0, + "content": "Figure 1. Comparison of three types of models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.548 + ], + "angle": 0, + "content": "brid image generation models [23, 78] integrate an additional image synthesis module, improving image quality but introducing extra complexity. (3) Mixed autoregressive-diffusion models [89, 90] employ diffusion-based image generation while maintaining an autoregressive framework for text, yet still struggles with unifying multi-modal." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Despite recent advancements, Transformer-based end-to-end models face several critical challenges: (1) their quadratic computational complexity makes them inefficient for generating high-resolution image and long-sequence text. Although various studies have attempted to optimize this computational complexity [1, 3, 12, 27, 29, 56, 59, 66, 74, 75], the challenge remains substantial. (2) their reliance on multi-objective learning introduces conflicting optimization goals, impeding convergence and hindering effective joint representation learning. In contrast, statespace models like Mamba [26, 61] offer a compelling alternative due to their ability to scale linearly with sequence length while effectively capturing long-range dependencies. However, the current multi-modal implementations of Mamba [18, 22, 30, 37, 49, 60, 73, 76, 80, 82] still adopt a multi-objective approach, limiting their capacity for end-to-end joint representation learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To effectively process multi-modal data, we propose an end-to-end model called the Multi-Modal Diffusion Mamba (MDM) (as shown in Fig. 1c). MDM first employs patchify [19] and embedding to pre-process multi-modal data. Then, it uses a variational autoencoder (VAE) [42] as a multi-modal encoder, which uniformly maps the multi" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20529" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.091, + 0.904, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.191, + 0.904, + 0.36 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.363, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Framework of Multi-Modal Diffusion Mamba. MDM first encodes inputs (caption, VQVAE-processed image, question) using VAE (a), while performing padding (class, diffusion timestep, token completion) and flatten operations (d, e). Next, data reconstruction is progressively completed via diffusion mamba operations (b), modeling images and text temporally through scanning processes (f, g) for efficient information selection (red boxes indicate selection). Selected data undergoes computation (i) guided by (h) within the Mamba-2 framework to update model parameters. Finally, the MDM output passes through the VAE decoder (c) to reconstruct real data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.483, + 0.502 + ], + "angle": 0, + "content": "modal data to a noisy latent space (as illustrated in Fig. 2a). MDM constructs a multi-step selection diffusion model based on the Mamba architecture as a uniform decoder for the rapid generation of multi-modal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.483, + 0.793 + ], + "angle": 0, + "content": "This decoder generates the target text or image step-by-step based on the diffusion process through the multi-step selection diffusion model (as shown in Fig. 2b). To enhance decoding speed, the decoder employs the Score Entropy Loss [50] as the objective function instead of Markov chain-based [35] methods for updating the network to handle multi-modal data throughout the diffusion process. The decoder comprises two components: an image and text scan switch, and a Mamba-2 block [26]. The text scan switch has two modes for sequence modeling (as shown in Fig. 2f), while the image scan switch has four, based on the settings of DiM [73] (as shown in Fig. 2e). The scan switches enable the model to capture sequential relationships across various temporal directions in the data. The selection state-space structure in Mamba then analyzes these sequential relationships within the current denoising step. This analysis guides the selection of relevant information to focus on and irrelevant information to ignore, effectively directing the model's denoising process at each step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Since MDM unifies the modality encoder and decoder, the model is capable of generating an image and text simultaneously. For example, as shown in Fig. 2h, when generating an image of a dog alongside its description, the scan switch in the decoder first assesses whether the description contains conditions that necessitate image generation. If such conditions exist, the image scan switch is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.907, + 0.577 + ], + "angle": 0, + "content": "activated. Consequently, the model directs its selection to the image patches corresponding to the dog during each denoising step. This targeted focus guides the model to effectively denoise relevant pixels while disregarding other areas of the image. A similar selection process is employed for text data. Ultimately, the data, once denoised via the t-step diffusion process, is reconstructed into authentic text (or an image) through the VAE decoder simultaneously. The main contributions of this paper are as follows." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.585, + 0.907, + 0.678 + ], + "angle": 0, + "content": "1) We introduce the Multi-Modal Diffusion Mamba (MDM), an end-to-end model that achieves a computational complexity of \\(\\mathcal{O}(MLN^2)\\), outperforming previous end-to-end models like MonoFormer [89], which operate at \\(\\mathcal{O}(ML^2 N / G)\\). This advancement enables the efficient generation of long-sequence text and high-resolution images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.909, + 0.761 + ], + "angle": 0, + "content": "2) We propose a novel multi-step selection diffusion model that combines autoregressive and diffusion-based generative paradigms into a unified learning objective. This method effectively integrates both paradigms within a diffusion process, generating multi-modal data simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.909, + 0.907 + ], + "angle": 0, + "content": "3) Our experimental results demonstrate MDM's superior performance in image generation on the ImageNet [13] and COCO datasets [40]. Additionally, it excels in various tasks, including image captioning on Flickr30K [84] and COCO [40], VQA on VQAv2 [25], VizWiz [28], and OKVQA [53], as well as text comprehension and reasoning on seven datasets [7, 9, 10, 54, 65, 87]. Furthermore, MDM shows strong results in math-related world knowledge tasks on GSM8k [11], MATH [33], and MMLU [32]." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.585, + 0.909, + 0.907 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20530" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.241, + 0.105 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.109, + 0.408, + 0.124 + ], + "angle": 0, + "content": "2.1. Traditional large multi-modal model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.125, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Most existing LMMs are built by integrating architectures from multiple modalities [2, 4, 41, 46, 62, 86]. SOTA image and video generation models employ pre-trained text encoders to represent input prompts in latent space, which then condition a diffusion model for generating videos and images [64]. Many researchers have adopted this approach, fusing feature representations from multiple pretrained encoders to enhance model performance across different modalities [21, 58]. This pattern is also prevalent in visual language models, where pre-trained language models are typically augmented with linear projection layers from other pre-trained en/decoders for training in the text space. Examples include Flamingo [2] and LLaVA [48] for visual understanding, GILL [43] for visual generation, and DreamLLM [17] for both understanding and generation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.353, + 0.37, + 0.367 + ], + "angle": 0, + "content": "2.2. End-to-End multi-modal model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.518 + ], + "angle": 0, + "content": "End-to-end models have emerged recently to facilitate joint representation learning while improving training and inference efficiency. It can be categorized into three main types: 1) The autoregressive model [5, 31, 69, 71] utilizes one Transformer with an autoregressive approach to generate images and text. For instance, the Fuyu model [5] processes image patches directly as input to achieve visual comprehension. Models like Chameleon [71], Mars [31], and LlamaGen [69] convert images into discrete sequence tokens, then concatenate them with text." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.64 + ], + "angle": 0, + "content": "2) The hybrid image generation model [23, 78] addresses the limitations of autoregressive approaches in image generation. While maintaining an autoregressive structure for text generation, the models enhance image quality by incorporating an image-generation network. For example, Seedx model [23] focuses on enhancing specific aspects of image generation, while Next-GPT [78] aims to expand multimodal capabilities within an end-to-end framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.484, + 0.761 + ], + "angle": 0, + "content": "3) The mixed autoregressive-diffusion model [89, 90] combines the strengths of previous approaches. It performs text autoregressive generation and image diffusion restoration simultaneously. Models like MonoFormer [89] and Transfusion [90] achieve this by incorporating causal self-attention [81] for text tokens and bidirectional self-attention [14] for image patches, enabling high-quality multi-modal understanding and generation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.765, + 0.358, + 0.779 + ], + "angle": 0, + "content": "2.3. Mamba in multi-modal model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Mamba has emerged as a powerful alternative to Transformer for multi-modal data alignment [18, 49, 76, 77, 82]. Recent works showcase Mamba's capabilities across different multi-modal applications. VL-Mamba [60] combines a pre-trained Mamba model for language understanding with a connector module to align visual patches and language tokens. However, these models lack end-to-end training capabilities and struggle to learn unified joint representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "tions. MDM provides a truly end-to-end architecture, enabling rapid generation of high-quality, long sequences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.126, + 0.847, + 0.143 + ], + "angle": 0, + "content": "3. Multi-step Selection Diffusion Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.145, + 0.907, + 0.312 + ], + "angle": 0, + "content": "The multi-step selection diffusion model enables rapid generation of multi-modal information through two key processes: diffusion & denoising and selection. During the diffusion & denoising, the model employs a unified Score Entropy Loss [50](SE) to gradually reconstruct target data from noise through a series of denoising steps (as illustrated in Fig. 2b). The selection process enables the model to capture sequential relationships across different temporal dimensions in the latent space, determining which information should be focused on or ignored during each diffusion denoising step (as shown in Fig. 2h)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.313, + 0.722, + 0.328 + ], + "angle": 0, + "content": "3.1. Diffusion & Denoising" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.906, + 0.375 + ], + "angle": 0, + "content": "The diffusion & denoising process comprises two main components: diffusion and denoising. The diffusion component can be expressed by the following equation:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.381, + 0.907, + 0.408 + ], + "angle": 0, + "content": "\\[\nz _ {n, t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t} ^ {g}} \\epsilon_ {n, t} ^ {g}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.413, + 0.907, + 0.567 + ], + "angle": 0, + "content": "where \\(g\\) denotes either image patch or text embedding, and \\(z_{n,0}^{g}\\) represents the latent space vector of the \\(n\\)-th image patch or text embedding, obtained through VAE sampling [42]. \\(z_{n,t}^{g}\\) is derived from \\(z_{n,0}^{g}\\) after \\(t\\) steps of noise addition; \\(\\epsilon_{n,t}^{g} \\sim \\mathcal{N}(0, I)\\) represents the added noise; \\(\\bar{\\alpha}_{t}^{g} = \\prod_{k=1}^{t} \\alpha_{k}^{g}\\), \\(\\alpha_{k}^{g} = 1 - \\beta_{k}^{g}\\), and \\(\\{\\beta_{k}^{g} \\in (0,1)\\}_{k=1}^{T}\\) are Gaussian distribution hyperparameters controlling the forward diffusion noise. Following the diffusion Markov principle [35], \\(t\\)-step forward diffusion process can be characterized by conditional probabilities as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.575, + 0.907, + 0.601 + ], + "angle": 0, + "content": "\\[\np (z _ {n, t} ^ {g} | z _ {n, 0} ^ {g}) = \\mathcal {N} (z _ {n, t} ^ {g}; \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g}, (1 - \\bar {\\alpha} _ {t} ^ {g}) I), \\quad (2)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.904, + 0.641 + ], + "angle": 0, + "content": "which means that given \\(z_{n,0}^{g}, z_{n,t}^{g}\\) follows a Gaussian distribution with \\(\\sqrt{\\bar{\\alpha}_t^g} z_{n,0}^g\\) as mean and \\((1 - \\bar{\\alpha}_t^g)I\\) as variance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.906, + 0.716 + ], + "angle": 0, + "content": "In the classic diffusion denoising component [35], the model needs to learn the posterior \\( p(z_{n,t-1}^g | z_{n,t}^g) \\) to gradually reconstruct the data. Since \\( p(z_{n,t}^g | z_{n,0}^g) \\) follows a Gaussian distribution, we can assume that the approximate distribution of the denoising process is:" + }, + { + "type": "equation", + "bbox": [ + 0.538, + 0.723, + 0.907, + 0.743 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} (z _ {n, t - 1} ^ {g} | z _ {n, t} ^ {g}) = \\mathcal {N} (z _ {n, t - 1} ^ {g}; \\mu_ {\\theta} (z _ {n, t} ^ {g}), (\\sigma_ {\\theta , n} ^ {g}) ^ {2}). \\qquad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.78 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(z_{n,t}^{g})\\) and \\(\\sigma_{\\theta ,n}^{g}\\) represent the model predicted noise mean and variance at the \\(t\\) -th denoising step." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.855 + ], + "angle": 0, + "content": "This method achieves the gradual recovery of data by optimizing the conditional probability of each time step by maximum likelihood. However, Markov chain-based [35] methods limit computational efficiency in high-dimensional spaces and are difficult to extend to discrete data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To further optimize the denoising process, this paper uses SE [50] as the optimization target. It is a generalized score matching objective that aims to directly learn the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20531" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.199 + ], + "angle": 0, + "content": "probability density ratio between discrete states. The SE can not only stabilize the diffusion denoising process but also improve the sampling quality through the global information of data distribution. In general form, for any state pair \\((z_{n,t}^{g}, z_{n,0}^{g})\\), define the model's score ratio \\(s_{\\theta}(z_{n,t}^{g})\\) which represents the relative probability of transferring from \\(z_{n,t}^{g}\\) to \\(z_{n,0}^{g}\\). SE is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.205, + 0.482, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} s e = \\sum_ {y \\in z _ {n, 0: t - 1} ^ {g}} \\omega_ {z _ {n, t} ^ {g}} ^ {g} \\left(s _ {\\theta} (z _ {n, t} ^ {g}) - \\frac {p _ {d a t a} (y)}{p _ {d a t a} (z _ {n , t} ^ {g})} \\log s _ {\\theta} (z _ {n, t} ^ {g}) \\right. \\\\ + K \\left(\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)}\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.424 + ], + "angle": 0, + "content": "where \\(\\omega_{z_{n,t}}^{g}\\) is the weight of the loss term, which is used to balance the loss of different states. \\(K(a) = a(\\log a - 1)\\) is a normalization term that ensures the loss is non-negative. \\(\\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}\\) represents the actual score ratio. \\(p_{data}(y)\\) and \\(p_{data}(z_{n,t}^{g})\\) are the actual data distributions of the former noisy state and the current noisy state. The actual score ratio calculation relationship is shown in Theorem 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.484, + 0.481 + ], + "angle": 0, + "content": "Theorem 1. According to Bayes' theorem and the Gaussian distribution density formula, the following calculation relationship of \\( \\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})} \\) is obtained:" + }, + { + "type": "equation", + "bbox": [ + 0.103, + 0.486, + 0.482, + 0.54 + ], + "angle": 0, + "content": "\\[\n\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)} = \\exp \\left(\\frac {\\| z _ {n , t} ^ {g} \\| ^ {2}}{2} - \\frac {\\| z _ {n , t} ^ {g} - \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n , 0} ^ {g} \\| ^ {2}}{2 \\left(1 - \\bar {\\alpha} _ {t} ^ {g}\\right)}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.546, + 0.362, + 0.561 + ], + "angle": 0, + "content": "The proof is provided in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.621 + ], + "angle": 0, + "content": "Based on the SE [50], the model predicted score ratio indicates how the model adjusts the probability of the current state to tend to the original data distribution during the denoising process. The definition is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.215, + 0.627, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {p _ {\\theta} \\left(z _ {n , 0} ^ {g}\\right)}{p _ {\\theta} \\left(z _ {n , t} ^ {g}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.668, + 0.484, + 0.759 + ], + "angle": 0, + "content": "where the denominator represents the probability of the current noise state and the numerator represents the original state probability estimated by the model. According to Theorem 2, the model uses softmax for normalization ensuring numerical stability and enabling gradient optimization when predicting the score ratio." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.84 + ], + "angle": 0, + "content": "Theorem 2. Given the denoising process modelled by a score-based probability ratio function \\( s_{\\theta}(z_{n,t}^{g}) \\), defined as Eq. (6), this paper defines a learnable approximation using a parameterized score function \\( f_{\\theta} \\), such that the probability ratio can be estimated as:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.845, + 0.483, + 0.884 + ], + "angle": 0, + "content": "\\[\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {\\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , z _ {n , 0} ^ {g}\\right)\\right)}{\\sum_ {y \\in z _ {n , 0 : t - 1} ^ {g}} \\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , y\\right)\\right)}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.89, + 0.363, + 0.906 + ], + "angle": 0, + "content": "The proof is provided in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.62, + 0.105 + ], + "angle": 0, + "content": "3.2. Selection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.107, + 0.905, + 0.228 + ], + "angle": 0, + "content": "The selection process comprises two key steps: scan switch and selection. The scan switch mechanism captures temporal relationships between adjacent image patches (or text embeddings) by generating latent space representations with \\( k \\) different sequential relationships, such as four image patch sequences and two text embedding sequences illustrated in Fig. 2fg. The mechanism creates \\( k \\) temporal sequences \\( S = \\{\\langle z_{1,t}^{g},z_{2,t}^{g},\\dots ,z_{i,t}^{g}\\rangle \\}_{k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.906, + 0.351 + ], + "angle": 0, + "content": "The selection step then analyzes these different sequential relationships at the current denoising step \\( t \\) to determine which information should be focused on or ignored, thereby guiding the model's denoising direction in each diffusion step. The selection step chooses \\( j \\) items \\( z_{n,t}^{g} \\) from each sequence in \\( S \\) according to the following Theorem 3. So, the selection step obtain \\( k \\) selection sequences with different lengths, i.e., \\( S' = \\{\\langle z_{j_1,t}^g, z_{j_2,t}^g, \\ldots, z_{j,t}^g \\rangle\\}_k \\) and \\( S' \\in S \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.356, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Theorem 3. To achieve the optimal score entropy [50] which is demonstrated on Eq. (4), the selection step choose \\(j\\) items where each \\(z_{n,t}^{g}\\) satisfies \\(se = 0\\), i.e.," + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.411, + 0.905, + 0.445 + ], + "angle": 0, + "content": "\\[\ns _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) \\approx \\frac {p _ {d a t a} (y)}{p _ {d a t a} \\left(z _ {n , t} ^ {g}\\right)} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.451, + 0.785, + 0.466 + ], + "angle": 0, + "content": "The proof is provided in Appendix A." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.472, + 0.645, + 0.487 + ], + "angle": 0, + "content": "4. Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.612 + ], + "angle": 0, + "content": "The neural network architecture consists of two primary components: a VAE noisy latent encoder [42] and a multistep selection diffusion decoder, as illustrated in Fig. 2ab. The encoder first processes image data \\( X_{img} \\) through patchify [19] operations and processes text data \\( X_{txt} \\) through tokenization based on SentencePiece with Unigram BPE [45] and embedding operations, then uniformly maps them to the latent space before applying forward noise." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.838 + ], + "angle": 0, + "content": "The decoder, based on the multi-step selection diffusion model, leverages Mamba to achieve unified learning objectives while enhancing computational efficiency for processing long sequence data. It employs the SE [50] as the unified objective for both image and text modalities during the diffusion process. During selection, the model captures sequential relationships across different temporal dimensions using various scan switches. These relationships are then efficiently processed through the selection state-space structure in the Mamba Block determining which information to focus on or ignore according to Eq. (8), thereby guiding subsequent diffusion denoising steps (as shown in Fig. 2h). Finally, the reconstructed image patches and text embeddings are transformed back into their original data formats through a VAE noisy latent decoder [42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.84, + 0.74, + 0.855 + ], + "angle": 0, + "content": "4.1. The noisy latent encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The noisy latent encoder first processes input image \\( X_{img} \\) through patchify and processes text \\( X_{txt} \\) through tokenization and embedding operations to obtain the patch sequence" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "20532" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.484, + 0.243 + ], + "angle": 0, + "content": "\\(G(X_{img} / X_{txt}) = \\langle g_1, g_2, \\dots, g_i \\rangle\\), where \\(g_n\\) represents the \\(n\\)-th image patch or text embedding, respectively. The encoder VAE [42] generates Gaussian distribution parameters (mean \\(\\mu\\) and variance \\(\\sigma\\)) for these patches, with a similar process applied to text embeddings, i.e., \\(VAE(G) = (\\mu, \\sigma)\\). For each image patch or text embedding \\(g_n\\), its noise \\(z_n\\) is a sample \\(s_n\\) from the distribution \\(\\mathcal{N}(\\mu, \\sigma)\\) with the addition noise \\(\\epsilon_n \\sim \\mathcal{N}(0, 1)\\), i.e., \\(z_n = s_n + \\epsilon_n\\). Finally, the image \\(X_{img}\\) and text \\(X_{txt}\\) are transformed into the noise sequence \\(\\langle z_1, \\dots, z_i \\rangle\\) through the above process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.336 + ], + "angle": 0, + "content": "Moreover, three types of learnable padding tokens, time, category, and pad, are inserted into these noise sequences, as illustrated in Fig. 2de. The time token encodes the current diffusion step, the class token is used to learn the data category, and the pad token represents the start or end position for splitting these noise sequences." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.342, + 0.449, + 0.357 + ], + "angle": 0, + "content": "4.2. The multi-step selection diffusion decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.484, + 0.466 + ], + "angle": 0, + "content": "The decoder aims at progressively recovering the image \\( X_{img} \\) or text \\( X_{txt} \\) from noise sequences through two main modules: 1) the multi-step selection diffusion Mamba and 2) the VAE noisy latent decoder. 1) The Mamba is used to recover the patch sequence \\( \\langle g_1, \\dots, g_i \\rangle \\) from the noise sequence \\( \\langle z_1, \\dots, z_i \\rangle \\). 2) The VAE noisy latent decoder assembles patches and generates the image \\( \\hat{X}_{img} \\) or text \\( \\hat{X}_{txt} \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.47, + 0.398, + 0.484 + ], + "angle": 0, + "content": "4.2.1. Multi-step selection diffusion Mamba" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.529 + ], + "angle": 0, + "content": "The module leverages two components, image/text scan switch and Mamba Block, to implement each denoising step in the multi-step selection diffusion model (Sec. 3)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.606 + ], + "angle": 0, + "content": "The image/text scan switch component establishes sequences with different directions to capture different temporal relationships between patches. Following Dim [73], we implement four distinct scan switches for images (as shown in Fig. 2f) and two for text (as shown in Fig. 2g)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.743 + ], + "angle": 0, + "content": "The Mamba block is used to select patches from these different scan switch sequences and denoise the input noise \\( z_{n,t}^{g} \\). The block adopts the state space architecture from Mamba-2 [26]. According to Sec. 3.2, it is \\( s_{\\theta} \\), where \\( \\theta = \\{H_{n,t}^{g}, A, B, C, D, \\Delta\\} \\) represent the state space in the block. The block comprises six key components: 1) linear input and output projection layers, 2) convolution kernel layer, 3) nonlinear activation layer, 4) state space model (SSM), 5) skip connection layer, and 6) normalization layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.851 + ], + "angle": 0, + "content": "1) The linear input projection layer reduces the dimensionality of the latent space noise vector while simultaneously applying initial state matrices \\( A \\), \\( B \\), \\( C \\) to the linear projection of input data \\( z_{n,t}^{g} \\). Additionally, the linear output projection layer represents the denoising step, which transforms the selection noise \\( z_{n,t}^{g} \\) into \\( z_{n,t - \\Delta t}^{g} \\) and outputs it to the next Mamba block according to the following equation." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.875, + 0.484, + 0.906 + ], + "angle": 0, + "content": "\\[\nz _ {n, t - \\Delta t} ^ {g} = z _ {n, t} ^ {g} - \\frac {\\Delta t}{2} \\left[ f _ {\\theta} \\left(z _ {n, t} ^ {g}, t\\right) + f _ {\\theta} \\left(z _ {n, t - \\Delta t} ^ {g}, t - \\Delta t\\right) \\right] (9)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "where the equation adopts the second-order numerical method of DPM-Solver [51] to improve sampling accuracy. Details are provided in Appendix B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.213 + ], + "angle": 0, + "content": "2) The convolution kernel layer implements parallel scan switches, routing the initial linear projection of the input and the state matrix's linear projection through the SSM, as shown in Fig. 2i. The sweep down and sweep up [26] enable parallel computation between Eqs. (10) to (13)." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.213, + 0.892, + 0.228 + ], + "angle": 0, + "content": "3) The nonlinear layer enhances model generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.289 + ], + "angle": 0, + "content": "4) The SSM lets the Mamba block \\( s_{\\theta} \\) approximate the actual score ratio based on Theorem 3. To implement the target, SSM updates the state space \\( \\theta \\) by the following equations (based on Theorem 3 and details in Appendix A)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.302, + 0.905, + 0.321 + ], + "angle": 0, + "content": "\\[\nH _ {n, t} ^ {g} = \\bar {A} H _ {n, t - 1} ^ {g} + \\bar {B} z _ {n, t} ^ {g} \\tag {10}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.34, + 0.905, + 0.36 + ], + "angle": 0, + "content": "\\[\nz _ {n - 1, t} ^ {g} = C H _ {n, t} ^ {g} + D z _ {n, t} ^ {g} \\tag {11}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.658, + 0.379, + 0.905, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\bar {A} = \\exp (\\Delta A) \\tag {12}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.417, + 0.905, + 0.434 + ], + "angle": 0, + "content": "\\[\n\\bar {B} = (\\Delta A) ^ {- 1} \\cdot \\left(\\exp (\\Delta A) - I\\right) \\cdot \\Delta B \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.44, + 0.906, + 0.531 + ], + "angle": 0, + "content": "where \\( H_{n,t}^{g} \\) represents the hidden state representation, \\( A \\) and \\( B \\) control the evolution of hidden states and latent space noise vector inputs, respectively, \\( C \\) governs the hidden state representation of the target output and \\( D \\) manages the nonlinear skip connection for latent space noise vector inputs. \\( \\Delta \\) denotes the learnable time parameter." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.532, + 0.905, + 0.562 + ], + "angle": 0, + "content": "5) The skip connection layer facilitates input feature reuse and mitigates model degradation." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.562, + 0.884, + 0.577 + ], + "angle": 0, + "content": "6) The Normalization layer ensures training stability." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.577, + 0.905, + 0.607 + ], + "angle": 0, + "content": "According to Eq. (8) in Theorem 3 and Eq. (4), the goal of training the Mamba block is:" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.532, + 0.905, + 0.607 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.62, + 0.905, + 0.639 + ], + "angle": 0, + "content": "\\[\nL _ {s e} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim p _ {0}, z _ {n} ^ {g} \\sim p (\\cdot | z _ {n, 0} ^ {g})} s e = 0 \\tag {14}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.731, + 0.661 + ], + "angle": 0, + "content": "4.2.2. The noisy latent decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.906, + 0.727 + ], + "angle": 0, + "content": "After applying the diffusion-based denoising process, the recovered latent variable \\( z_{n,0}^{g} \\) is passed to the VAE decoder [42] as illustrated in Fig. 2c. For image reconstruction, the decoder applies an \\( \\ell_2 \\) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.737, + 0.905, + 0.757 + ], + "angle": 0, + "content": "\\[\nL _ {r e c} ^ {i m g} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\| X _ {i m g} - \\hat {X} _ {i m g} \\| ^ {2}. \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.769, + 0.905, + 0.797 + ], + "angle": 0, + "content": "where \\(q_{\\phi}(z|X)\\) represents the posterior distribution of the VAE encoder." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.799, + 0.895, + 0.814 + ], + "angle": 0, + "content": "For text, the decoder minimizes the cross-entropy loss:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.825, + 0.91, + 0.868 + ], + "angle": 0, + "content": "\\[\nL _ {r e c} ^ {t x t} = - \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\sum_ {t} p \\left(X _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right) \\log p _ {\\psi} \\left(\\hat {X} _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right). \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.869, + 0.905, + 0.903 + ], + "angle": 0, + "content": "where \\(p(X_{txt}^{(t)}|z_{n,0}^{g})\\) represents the probability distribution of real text data under the condition of latent variable \\(z_{n,0}^{g}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20533" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.482, + 0.239 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.306 + ], + "angle": 0, + "content": "Figure 3. VQA, QA and Multi-Modal generation test from MDM. The results of VQA are part of VQAv2 [25]. The QA results are part of PIQA [7] and MMLU [32]. The Multi-Modal generation results are tested with ground-truth data." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.31, + 0.482, + 0.537 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.547, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Figure 4. Comparison between each model on generating captioning and image results on COCO dataset. Unlike other models, MDM generates both image and caption data simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.643 + ], + "angle": 0, + "content": "And \\(p_{\\psi}(\\hat{X}_{txt}^{(t)}|z_{n,0}^g)\\) represents the probability distribution of the text token generated by the VAE decoder under the condition of the latent variable \\(z_{n,0}^g\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.644, + 0.466, + 0.658 + ], + "angle": 0, + "content": "Besides, a KL divergence regularizes the latent space:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.671, + 0.483, + 0.687 + ], + "angle": 0, + "content": "\\[\nL _ {K L} = D _ {K L} \\left(q _ {\\phi} (z | X) \\| p (z)\\right). \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.76 + ], + "angle": 0, + "content": "where \\(p(z)\\) represents the prior distribution of the latent variable by VAE, which is assumed to be a standard Gaussian distribution \\(\\mathcal{N}(0,I)\\) to regularize the latent variable space and enable it to have smooth generation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.761, + 0.483, + 0.79 + ], + "angle": 0, + "content": "The final optimization objective integrates VAE reconstruction, KL divergence and SE:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.802, + 0.483, + 0.82 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {r e c} ^ {i m g} + L _ {r e c} ^ {t x t} + \\beta L _ {K L} + \\lambda L _ {s e}. \\tag {18}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.224, + 0.847 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.281, + 0.868 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Model configuration. Our model applies a VAE [42] as the noisy latent encoder and decoder. Moreover, it integrates" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "the DiM selection state space [73] in each Mamba block as the diffusion decoder. The resulting model contains 7 billion parameters, with 49 Mamba blocks in the multi-step selection diffusion decoder, each having a dimension of 2048 (Details of parameter settings listed in Appendix C)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Before the training MDM process, we trained a tokenization model based on SentencePiece (Unigram BPE) [45]. The tokenization model can help the model construct a stable text latent variable representation, thereby optimizing the forward diffusion and reverse denoising process. See Appendix D for detailed experimental settings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.261, + 0.906, + 0.351 + ], + "angle": 0, + "content": "In the training process, we import the DDPM scheduler [35] and DPM-Solver [51] to improve the sampling efficiency in the diffusion model. We then use the AdamW optimizer without weight decay, maintaining a constant learning rate of 0.0001. Meanwhile, we keep an EMA of the model weights with a coefficient of 0.9999." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.906, + 0.429 + ], + "angle": 0, + "content": "Baseline and dataset. Our evaluation encompasses four tasks: image generation with classifier-free guidance [34] (CFG), text-to-image, image-to-text, and text-to-text generation. For the baseline model training, we train MDM on ImageNet [13], JourneyDB [68] and UltraChat [16]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.642 + ], + "angle": 0, + "content": "For the image generation and the text-to-image task at \\(256 \\times 256\\) resolution, we compare the MDM baseline model against established baselines across three categories: diffusion models (Imagen [64], ADM [15], CDM [36], LDM [63], DiT-XL/2 [57], SDXL [58], and SD-3 [21]), autoregressive models (VQGAN [20] and ViT-VQGAN [85]), and end-to-end multi-modal models (NExT-GPT [78], Chameleon [71], LlamaGen [69], Transfusion [90], MonoFormer [89], Dual-DiT [47], JanusFlow [52] and ShowO [79]). For the image generation task, we evaluate performance on ImageNet [13] using four metrics: Frechet Inception Distance (FID), Inception Score (IS), and Precision/Recall. For the text-to-image task, we evaluate performance on COCO [40] using FID and Gen Eval [24]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.906, + 0.853 + ], + "angle": 0, + "content": "For the image-to-text task (image captioning and vision question answering, VQA) and text-to-text task, we employ MDM baseline model and MDM instruction model by visual instruction tuning [48] on multiple datasets: COCO [40], GQA [38], OCR-VQA [55], TextVQA [67], and VisualGenome [44]. We evaluate the model against two groups of baselines: traditional models and end-to-end multi-modal models. Performance evaluation of image captioning is conducted on Flickr 30K [84] and COCO [40] datasets using the Consensus-based Image Description Evaluation (CIDEr) metric. And performance evaluation of VQA is conducted on VQAv2 [25], VizWiz [28], and OKVQA [53] using answer accuracy rate as the evaluation metric." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "For the text-to-text task, we evaluate the model on text comprehension and reasoning tasks using HellaSwag [87], OpenBookQA [54], Wino-Grande [65], ARCEasy, ARC" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20534" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.907, + 0.38 + ], + "angle": 0, + "content": "
ModelArcParamsImage Generation with CFGText-to-Image Generation
FID ↓IS ↑Pre ↑Re ↑FID ↓Gen Eval ↑
Imagen [64]Diff7.3B----7.27-
ADM [15]Diff554M10.94101.00.690.63--
CDM [36]Diff-4.88158.7----
LDM [63]Diff400M3.60147.60.870.68-0.43
DiT-XL/2 [57]Diff675M2.27278.20.830.57--
SDXL [58]Diff3.4B----4.400.55
SD-3 [21]Diff12.7B-----0.68
VQGAN [20]AR227M18.6580.40.780.26--
ViT-VQGAN [85]AR1.7B4.17175.1----
NExT-GPT [78]AR7B----10.07-
Chameleon [71]AR7B----26.740.39
LlamaGen [69]AR3.1B2.81311.50.840.544.19-
Transfusion [90]AR+Diff7.3B----6.780.63
MonoFormer [89]AR+Diff1.1B2.57272.60.840.56--
Dual-DiT [47]Diff2B----9.400.65
JanusFlow [52]AR+Diff1.3B-----0.70
Show-O [79]AR+Diff1.3B----9.240.68
MDMDiff7B2.49281.40.860.595.910.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.39, + 0.907, + 0.418 + ], + "angle": 0, + "content": "Table 1. Performance on ImageNet and COCO \\( {256} \\times {256} \\) . FID,IS,Pre,and Re stands for Frechet Inception Distance,Inception Score, Precision, and Recall, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.424, + 0.909, + 0.733 + ], + "angle": 0, + "content": "
ModelICVQAText Comprehension and ReasoningMath and World
FlickrCOCOVQAv2VizWizOKHSOBQAWGARCEARCCBoolQPIQAGSM8kMATHMMLU
Llama-2 [74] (7B)-----77.258.678.575.245.977.478.814.62.545.3
Mistral [39] (7B)-----81.3-75.380.055.584.783.052.113.160.1
Flamingo [2] (80B)75.1113.867.6------------
Gemini Pro [72]82.299.871.2--84.7------86.532.671.8
GPT4V [8]55.378.577.2--95.3------92.052.986.4
InstructBLIP [48] (7B)82.4102.2-33.433.9----------
mPLUG-Owl [83] (7B)80.3119.3-39.0-----------
TinyLlama [88] (1.1B)-----59.236.059.155.330.157.873.3---
Pythia [6] (12B)-----52.033.257.454.028.563.370.9---
DREAMLLM [17](7B)-115.456.645.844.3----------
Emu [70](7B)-117.740.035.434.7----------
Chameleon [71](34B)74.7120.266.0--74.251.070.476.146.581.479.641.611.552.1
NExT-GPT [78](7B)84.5124.966.748.452.1----------
Transfusion [90](7B)-33.7-------------
MonoFormer [89](1.1B)-----50.637.256.948.231.562.371.2---
Dual-DiT [47](2B)-56.260.129.925.3----------
JanusFlow [52](1.3B)--79.8------------
Show-O [79](1.3B)67.6-74.7------------
MDM (7B)62.4109.660.339.847.170.641.568.855.146.265.779.940.512.154.4
InstructMDM (7B)75.2122.166.746.351.674.848.374.965.447.171.583.746.013.159.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.744, + 0.908, + 0.772 + ], + "angle": 0, + "content": "Table 2. Performance on image-to-text and text-to-text tasks. The evaluation of image captioning (IC) and VQA is CIDEr and answer accuracy % (Flickr is evaluated on 30K and OK represents OKVQA)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.782, + 0.482, + 0.843 + ], + "angle": 0, + "content": "Challenge [10], BoolQ [9], and PIQA [7]. We also evaluate the model on math and world knowledge tasks using GSM8K [11], MATH [33], and MMLU [32]. The evaluation metrics for all the tasks are accuracy rates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.293, + 0.866 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Image Generation. In the image generation task on ImageNet, MDM achieves top-three rankings across all eval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.782, + 0.907, + 0.889 + ], + "angle": 0, + "content": "uation metrics: second in FID, IS, and Precision, and third in Recall when compared against one-modal diffusion models and end-to-end multi-modal models (see Tab. 1). MDM demonstrates superior overall performance, notably surpassing other end-to-end multi-modal models in three of the four metrics. In the text-to-image task, we tested the model on the COCO dataset to generate both image and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20535" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.176, + 0.089, + 0.498, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.499, + 0.089, + 0.822, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.907, + 0.264 + ], + "angle": 0, + "content": "Figure 5. Comparison between Mamba Baseline, MonoFormer, and ours MDM on inference speed test. The left shows the inference speed of the model FPS at different resolutions. The right shows the inference speed of the model latency at different token lengths." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.274, + 0.485, + 0.348 + ], + "angle": 0, + "content": "
ModelImage/Text \nScan SwitchFPS w log \nscale↑FID↓
Model w Mamba①②③④/①②1.3572.49
Model w Mamba①②/①1.4053.96
Model w Transformer-1.9146.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.109, + 0.358, + 0.465, + 0.373 + ], + "angle": 0, + "content": "Table 3. Ablation on ImageNet \\( {256} \\times {256} \\) image generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.483, + 0.451 + ], + "angle": 0, + "content": "caption data. For the image generation results, we evaluated the FID and Gen Eval performance indicators of the model-generated images. MDM still achieved the top three performance levels and achieved SOTA on Gen Eval." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.452, + 0.483, + 0.753 + ], + "angle": 0, + "content": "Text Generation. In the image-to-text task image captioning, according to the settings on image generation on the COCO dataset, we tested the caption data of the model based on the model outputting both text and image data using the CIDEr indicator. The results showed that MDM ranked second among all models, as shown in Tab. 2. While in task VQA, MDM achieves competitive performance, surpassing several traditional models including InstructBLIP, mPLUG-Owl, DREAMLLM, and Emu, although it still trails behind top-performing models in the field as shown in Tab. 2. In the text-to-text generation task, as shown in Tab. 2, MDM and the other end-to-end multi-modal models perform worse than well-known traditional models. This discrepancy may be attributed to the fact that these end-to-end models have some deviations in multimodal fusion and learning because they abandon multiple language encoders, visual encoders, and multimodal fusion encoders. However, when compared with the other two end-to-end models, MDM excels, outperforming MonoFormer and surpassing Chameleon on seven out of ten datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.21, + 0.772 + ], + "angle": 0, + "content": "5.3. Discussion" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.777, + 0.291, + 0.792 + ], + "angle": 0, + "content": "5.3.1. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As demonstrated in Fig. 3, MDM shows the ability to generate image and text simultaneously in multiple rounds of dialogue and perform well in QA&VQA. Some results even exceed those of GPT-4V, particularly evident in the second and third rows of Fig. 4 which is a hybrid output process for the MDM model. Due to this, we set the model to generate corresponding images for the description text while simul" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.774, + 0.292 + ], + "angle": 0, + "content": "taneously generating image captioning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.429 + ], + "angle": 0, + "content": "This enhanced performance stems from MDM's multi-step selection diffusion decoder, which leverages Mamba's integrated selection and denoising capabilities to maintain focused attention on both textual and visual details. Validating our complexity analysis in Appendix E, MDM demonstrates superior efficiency compared to end-to-end Transformer models when processing long sequences, as shown in Fig. 5, particularly outperforming other end-to-end multimodal models for sequences exceeding 1280 tokens." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.432, + 0.627, + 0.444 + ], + "angle": 0, + "content": "5.3.2. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.909, + 0.614 + ], + "angle": 0, + "content": "Our ablation studies examine the impact of both the selection process and Mamba block components. Reducing the number of image/text scan switch sequences from 6 ('①②③④/①②') to 3 ('①②/①'), as shown in Tab. 3, improves inference speed but degrades image quality, as fewer scan switch sequences limit the model's ability to capture accurate information in complex sequences. Additionally, replacing the Mamba block with the Transformer further deteriorates output image quality, suggesting Mamba's temporal network architecture is better suited for representing diffusion relationships during the denoising process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.634, + 0.634 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.889 + ], + "angle": 0, + "content": "This paper introduces MDM (Multi-Modal Diffusion Mamba), a novel end-to-end architecture that significantly enhances multi-modal processing through two key innovations: a unified diffusion objective and an efficient selection mechanism leveraging Mamba's state-space structure. By integrating variational autoencoder with multi-step selection diffusion, MDM achieves SOTA overall performance in image generation and demonstrates remarkable versatility across various tasks, including image-to-text, text-to-text and text-image-to-text-image. Our comprehensive experiments illustrate that MDM consistently surpasses traditional end-to-end multi-modal models, particularly in processing high-resolution images and long-sequence text, while maintaining computational efficiency. The model's ability to unify different modalities under a single objective, coupled with its superior management of temporal relationships in the diffusion process, establishes a promising direction for future multi-modal architecture." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20536" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.268 + ], + "angle": 0, + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.269, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Silas Alberti, Niclas Dern, Laura Thesing, and Gitta Kutyniok. Sumformer: Universal approximation for efficient transformers. In Topological, Algebraic and Geometric Learning Workshops 2023, pages 72-86. PMLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[4] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. Advances in Neural Information Processing Systems, 35:32897-32912, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.409, + 0.482, + 0.45 + ], + "angle": 0, + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sağnak Taşırlar. Introducing our multimodal models, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.452, + 0.483, + 0.548 + ], + "angle": 0, + "content": "[6] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.55, + 0.482, + 0.605 + ], + "angle": 0, + "content": "[7] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, pages 7432-7439, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.35, + 0.619 + ], + "angle": 0, + "content": "[8] GPTV System Card. Openai, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[9] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[10] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[11] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.906, + 0.164 + ], + "angle": 0, + "content": "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.209 + ], + "angle": 0, + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.211, + 0.906, + 0.279 + ], + "angle": 0, + "content": "[16] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.905, + 0.352 + ], + "angle": 0, + "content": "[17] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.41 + ], + "angle": 0, + "content": "[18] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-mamba for cross-modality object detection. arXiv preprint arXiv:2404.09146, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.512 + ], + "angle": 0, + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.514, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[22] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Youqiang Zhang, and Junshi Huang. Dimba: Transformermamba diffusion models. arXiv preprint arXiv:2406.01159, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[23] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20537" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Ahan Gupta, Yueming Yuan, Yanqi Zhou, and Charith Mendis. Flurka: Fast fused low-rank & kernel attention. arXiv preprint arXiv:2306.15799, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[28] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[29] Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David P Woodruff, and Amir Zandieh. Hyperattention: Long-context attention in near-linear time. arXiv preprint arXiv:2310.05869, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[30] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[31] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[32] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[33] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[34] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.646 + ], + "angle": 0, + "content": "[36] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.648, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[37] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[38] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[41] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.149, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[42] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[43] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[44] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.356 + ], + "angle": 0, + "content": "[45] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[47] Zijie Li, Henry Li, Yichun Shi, Amir Barati Farimani, Yuval Kluger, Linjie Yang, and Peng Wang. Dual diffusion for unified image generation and understanding. arXiv preprint arXiv:2501.00289, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[49] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[50] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[51] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[52] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv preprint arXiv:2411.07975, 2024.6.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20538" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.483, + 0.12 + ], + "angle": 0, + "content": "dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.164 + ], + "angle": 0, + "content": "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.165, + 0.483, + 0.22 + ], + "angle": 0, + "content": "[56] Matteo Pagliardini, Daniele Paliotta, Martin Jaggi, and François Fleuret. Fast attention over long sequences with dynamic sparse flash attention. Advances in Neural Information Processing Systems, 36:59808-59831, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.222, + 0.483, + 0.276 + ], + "angle": 0, + "content": "[57] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.483, + 0.348 + ], + "angle": 0, + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.483, + 0.404 + ], + "angle": 0, + "content": "[59] Ramya Prabhu, Ajay Nayak, Jayashree Mohan, Ramachandran Ramjee, and Ashish Panwar. vattention: Dynamic memory management for serving llms without pagedattention. arXiv preprint arXiv:2405.04437, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[60] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.463, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[61] Haohao Qu, Liangbo Ning, Rui An, Wenqi Fan, Tyler Derr, Xin Xu, and Qing Li. A survey of mamba. arXiv preprint arXiv:2408.01129, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.483, + 0.588 + ], + "angle": 0, + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[65] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.83 + ], + "angle": 0, + "content": "[66] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[67] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[68] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. Advances in Neural Information Processing Systems, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[69] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.905, + 0.274 + ], + "angle": 0, + "content": "[70] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[71] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[73] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[74] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[75] Zhongwei Wan, Xin Wang, Che Liu, Samiul Alam, Yu Zheng, Jiachen Liu, Zhongnan Qu, Shen Yan, Yi Zhu, Quanlu Zhang, et al. Efficient large language models: A survey. arXiv preprint arXiv:2312.03863, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[76] Zifu Wan, Pingping Zhang, Yuhao Wang, Silong Yong, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Sigma: Siamese mamba network for multi-modal semantic segmentation. arXiv preprint arXiv:2404.04256, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[77] Xinghan Wang, Zixi Kang, and Yadong Mu. Text-controlled motion mamba: Text-instructed temporal grounding of human motion. arXiv preprint arXiv:2404.11375, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[78] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[79] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[80] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8239-8249, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[81] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20539" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[82] Zhe Yang, Wenrui Li, and Guanghui Cheng. Shmamba: Structured hyperbolic state space model for audio-visual question answering. arXiv preprint arXiv:2406.09833, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[84] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[85] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.361, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[86] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.418, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[87] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[88] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[90] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.643 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20540" + } + ] +] \ No newline at end of file diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_origin.pdf b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..040e375a97cc832e67433623440326ad0470d251 --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/229968bf-b390-4563-8550-16e090dc25cb_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e66238e0a46d437438b875baa17e1328bab16adafed99364050ff789f590f632 +size 1799426 diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/full.md b/2025/End-to-End Multi-Modal Diffusion Mamba/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a0e9d811cedcfe2d09fe1f0d10440558e6c2056c --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/full.md @@ -0,0 +1,420 @@ +# End-to-End Multi-Modal Diffusion Mamba + +Chunhao Lu1 Qiang Lu1⊗ Meichen Dong1,2 Jake Luo3 + +1China University of Petroleum-Beijing, 2Leyard Optoelectronic, 3University of Wisconsin-Milwaukee + +{luchunhao,meichen.dong}@student.cup.edu.cn luqiang@cup.edu.cn jakeluo@uwm.edu + +# Abstract + +Current end-to-end multi-modal models utilize different encoders and decoders to process input and output information. This separation hinders the joint representation learning of various modalities. To unify multi-modal processing, we propose a novel architecture called MDM (Multi-modal Diffusion Mamba). MDM utilizes a Mamba-based multistep selection diffusion model to progressively generate and refine modality-specific information through a unified variational autoencoder for both encoding and decoding. This innovative approach allows MDM to achieve superior performance when processing high-dimensional data, particularly in generating high-resolution images and extended text sequences simultaneously. Our evaluations in areas such as image generation, image captioning, visual question answering, text comprehension, and reasoning tasks demonstrate that MDM significantly outperforms existing end-to-end models (MonoFormer, LlamaGen, and Chameleon etc.) and competes effectively with SOTA models like GPT-4V, Gemini Pro, and Mistral. Our results validate MDM's effectiveness in unifying multi-modal processes while maintaining computational efficiency, establishing a new direction for end-to-end multi-modal architectures. + +# 1. Introduction + +Traditional large-scale multi-modal models [2, 4, 17, 21, 41, 43, 46, 48, 58, 62, 64, 86] typically use multiple encoders and decoders to process multi-modal data. This approach makes learning a unified joint representation of the multi-modal data difficult and can significantly slow inference time (as shown in Fig. 1A). To alleviate these problems, end-to-end models without modal-fusion en(de)coder architecture have been proposed (as shown in Fig. 1B). This approach offers a streamlined, unified processing framework that enhances efficiency and consistency in multimodal representation learning. Existing end-to-end models follow three primary strategies: (1) Autoregressive models [5, 31, 69, 71] leverage a single Transformer for both text and image generation, but struggle with the inherent sequential dependency of autoregressive decoding. (2) Hy + +![](images/69bf2ddb13c63877bd884b7241cb6f3ce9497cdd50221666359c3fee6333d1ed.jpg) +(A) Traditional Multi-Modal Model + +![](images/93064fba34e137e7c3670503d22b47d150f89cf7ea5f93a5cdb1197d8ccd1e28.jpg) +(B) End-To-End Multi-Modal Model + +![](images/68c2974f2548c1b20ccbd9c5d1a25a22dfe803286cf6563818771aed8a644328.jpg) +(C) Ours +Figure 1. Comparison of three types of models. + +brid image generation models [23, 78] integrate an additional image synthesis module, improving image quality but introducing extra complexity. (3) Mixed autoregressive-diffusion models [89, 90] employ diffusion-based image generation while maintaining an autoregressive framework for text, yet still struggles with unifying multi-modal. + +Despite recent advancements, Transformer-based end-to-end models face several critical challenges: (1) their quadratic computational complexity makes them inefficient for generating high-resolution image and long-sequence text. Although various studies have attempted to optimize this computational complexity [1, 3, 12, 27, 29, 56, 59, 66, 74, 75], the challenge remains substantial. (2) their reliance on multi-objective learning introduces conflicting optimization goals, impeding convergence and hindering effective joint representation learning. In contrast, statespace models like Mamba [26, 61] offer a compelling alternative due to their ability to scale linearly with sequence length while effectively capturing long-range dependencies. However, the current multi-modal implementations of Mamba [18, 22, 30, 37, 49, 60, 73, 76, 80, 82] still adopt a multi-objective approach, limiting their capacity for end-to-end joint representation learning. + +To effectively process multi-modal data, we propose an end-to-end model called the Multi-Modal Diffusion Mamba (MDM) (as shown in Fig. 1c). MDM first employs patchify [19] and embedding to pre-process multi-modal data. Then, it uses a variational autoencoder (VAE) [42] as a multi-modal encoder, which uniformly maps the multi + +![](images/15d2a8e5c773c4f13c81be3b9ad8a35e1b558ae70f915c32dc1c8b0a3a6b323d.jpg) + +![](images/52cba87c0cc5afb41e3e3c8706d008e2f63434b17c4ee4ed22fe46a8d9c49508.jpg) +Figure 2. Framework of Multi-Modal Diffusion Mamba. MDM first encodes inputs (caption, VQVAE-processed image, question) using VAE (a), while performing padding (class, diffusion timestep, token completion) and flatten operations (d, e). Next, data reconstruction is progressively completed via diffusion mamba operations (b), modeling images and text temporally through scanning processes (f, g) for efficient information selection (red boxes indicate selection). Selected data undergoes computation (i) guided by (h) within the Mamba-2 framework to update model parameters. Finally, the MDM output passes through the VAE decoder (c) to reconstruct real data. + +modal data to a noisy latent space (as illustrated in Fig. 2a). MDM constructs a multi-step selection diffusion model based on the Mamba architecture as a uniform decoder for the rapid generation of multi-modal information. + +This decoder generates the target text or image step-by-step based on the diffusion process through the multi-step selection diffusion model (as shown in Fig. 2b). To enhance decoding speed, the decoder employs the Score Entropy Loss [50] as the objective function instead of Markov chain-based [35] methods for updating the network to handle multi-modal data throughout the diffusion process. The decoder comprises two components: an image and text scan switch, and a Mamba-2 block [26]. The text scan switch has two modes for sequence modeling (as shown in Fig. 2f), while the image scan switch has four, based on the settings of DiM [73] (as shown in Fig. 2e). The scan switches enable the model to capture sequential relationships across various temporal directions in the data. The selection state-space structure in Mamba then analyzes these sequential relationships within the current denoising step. This analysis guides the selection of relevant information to focus on and irrelevant information to ignore, effectively directing the model's denoising process at each step. + +Since MDM unifies the modality encoder and decoder, the model is capable of generating an image and text simultaneously. For example, as shown in Fig. 2h, when generating an image of a dog alongside its description, the scan switch in the decoder first assesses whether the description contains conditions that necessitate image generation. If such conditions exist, the image scan switch is + +activated. Consequently, the model directs its selection to the image patches corresponding to the dog during each denoising step. This targeted focus guides the model to effectively denoise relevant pixels while disregarding other areas of the image. A similar selection process is employed for text data. Ultimately, the data, once denoised via the t-step diffusion process, is reconstructed into authentic text (or an image) through the VAE decoder simultaneously. The main contributions of this paper are as follows. + +1) We introduce the Multi-Modal Diffusion Mamba (MDM), an end-to-end model that achieves a computational complexity of $\mathcal{O}(MLN^2)$ , outperforming previous end-to-end models like MonoFormer [89], which operate at $\mathcal{O}(ML^2 N / G)$ . This advancement enables the efficient generation of long-sequence text and high-resolution images. +2) We propose a novel multi-step selection diffusion model that combines autoregressive and diffusion-based generative paradigms into a unified learning objective. This method effectively integrates both paradigms within a diffusion process, generating multi-modal data simultaneously. +3) Our experimental results demonstrate MDM's superior performance in image generation on the ImageNet [13] and COCO datasets [40]. Additionally, it excels in various tasks, including image captioning on Flickr30K [84] and COCO [40], VQA on VQAv2 [25], VizWiz [28], and OKVQA [53], as well as text comprehension and reasoning on seven datasets [7, 9, 10, 54, 65, 87]. Furthermore, MDM shows strong results in math-related world knowledge tasks on GSM8k [11], MATH [33], and MMLU [32]. + +# 2. Related Works + +# 2.1. Traditional large multi-modal model + +Most existing LMMs are built by integrating architectures from multiple modalities [2, 4, 41, 46, 62, 86]. SOTA image and video generation models employ pre-trained text encoders to represent input prompts in latent space, which then condition a diffusion model for generating videos and images [64]. Many researchers have adopted this approach, fusing feature representations from multiple pretrained encoders to enhance model performance across different modalities [21, 58]. This pattern is also prevalent in visual language models, where pre-trained language models are typically augmented with linear projection layers from other pre-trained en/decoders for training in the text space. Examples include Flamingo [2] and LLaVA [48] for visual understanding, GILL [43] for visual generation, and DreamLLM [17] for both understanding and generation. + +# 2.2. End-to-End multi-modal model + +End-to-end models have emerged recently to facilitate joint representation learning while improving training and inference efficiency. It can be categorized into three main types: 1) The autoregressive model [5, 31, 69, 71] utilizes one Transformer with an autoregressive approach to generate images and text. For instance, the Fuyu model [5] processes image patches directly as input to achieve visual comprehension. Models like Chameleon [71], Mars [31], and LlamaGen [69] convert images into discrete sequence tokens, then concatenate them with text. + +2) The hybrid image generation model [23, 78] addresses the limitations of autoregressive approaches in image generation. While maintaining an autoregressive structure for text generation, the models enhance image quality by incorporating an image-generation network. For example, Seedx model [23] focuses on enhancing specific aspects of image generation, while Next-GPT [78] aims to expand multimodal capabilities within an end-to-end framework. + +3) The mixed autoregressive-diffusion model [89, 90] combines the strengths of previous approaches. It performs text autoregressive generation and image diffusion restoration simultaneously. Models like MonoFormer [89] and Transfusion [90] achieve this by incorporating causal self-attention [81] for text tokens and bidirectional self-attention [14] for image patches, enabling high-quality multi-modal understanding and generation. + +# 2.3. Mamba in multi-modal model + +Mamba has emerged as a powerful alternative to Transformer for multi-modal data alignment [18, 49, 76, 77, 82]. Recent works showcase Mamba's capabilities across different multi-modal applications. VL-Mamba [60] combines a pre-trained Mamba model for language understanding with a connector module to align visual patches and language tokens. However, these models lack end-to-end training capabilities and struggle to learn unified joint representation + +tions. MDM provides a truly end-to-end architecture, enabling rapid generation of high-quality, long sequences. + +# 3. Multi-step Selection Diffusion Model + +The multi-step selection diffusion model enables rapid generation of multi-modal information through two key processes: diffusion & denoising and selection. During the diffusion & denoising, the model employs a unified Score Entropy Loss [50](SE) to gradually reconstruct target data from noise through a series of denoising steps (as illustrated in Fig. 2b). The selection process enables the model to capture sequential relationships across different temporal dimensions in the latent space, determining which information should be focused on or ignored during each diffusion denoising step (as shown in Fig. 2h). + +# 3.1. Diffusion & Denoising + +The diffusion & denoising process comprises two main components: diffusion and denoising. The diffusion component can be expressed by the following equation: + +$$ +z _ {n, t} ^ {g} = \sqrt {\bar {\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g} + \sqrt {1 - \bar {\alpha} _ {t} ^ {g}} \epsilon_ {n, t} ^ {g}, \tag {1} +$$ + +where $g$ denotes either image patch or text embedding, and $z_{n,0}^{g}$ represents the latent space vector of the $n$ -th image patch or text embedding, obtained through VAE sampling [42]. $z_{n,t}^{g}$ is derived from $z_{n,0}^{g}$ after $t$ steps of noise addition; $\epsilon_{n,t}^{g} \sim \mathcal{N}(0, I)$ represents the added noise; $\bar{\alpha}_{t}^{g} = \prod_{k=1}^{t} \alpha_{k}^{g}$ , $\alpha_{k}^{g} = 1 - \beta_{k}^{g}$ , and $\{\beta_{k}^{g} \in (0,1)\}_{k=1}^{T}$ are Gaussian distribution hyperparameters controlling the forward diffusion noise. Following the diffusion Markov principle [35], $t$ -step forward diffusion process can be characterized by conditional probabilities as follows: + +$$ +p (z _ {n, t} ^ {g} | z _ {n, 0} ^ {g}) = \mathcal {N} (z _ {n, t} ^ {g}; \sqrt {\bar {\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g}, (1 - \bar {\alpha} _ {t} ^ {g}) I), \quad (2) +$$ + +which means that given $z_{n,0}^{g}, z_{n,t}^{g}$ follows a Gaussian distribution with $\sqrt{\bar{\alpha}_t^g} z_{n,0}^g$ as mean and $(1 - \bar{\alpha}_t^g)I$ as variance. + +In the classic diffusion denoising component [35], the model needs to learn the posterior $p(z_{n,t-1}^g | z_{n,t}^g)$ to gradually reconstruct the data. Since $p(z_{n,t}^g | z_{n,0}^g)$ follows a Gaussian distribution, we can assume that the approximate distribution of the denoising process is: + +$$ +p _ {\theta} (z _ {n, t - 1} ^ {g} | z _ {n, t} ^ {g}) = \mathcal {N} (z _ {n, t - 1} ^ {g}; \mu_ {\theta} (z _ {n, t} ^ {g}), (\sigma_ {\theta , n} ^ {g}) ^ {2}). \qquad (3) +$$ + +where $\mu_{\theta}(z_{n,t}^{g})$ and $\sigma_{\theta ,n}^{g}$ represent the model predicted noise mean and variance at the $t$ -th denoising step. + +This method achieves the gradual recovery of data by optimizing the conditional probability of each time step by maximum likelihood. However, Markov chain-based [35] methods limit computational efficiency in high-dimensional spaces and are difficult to extend to discrete data. + +To further optimize the denoising process, this paper uses SE [50] as the optimization target. It is a generalized score matching objective that aims to directly learn the + +probability density ratio between discrete states. The SE can not only stabilize the diffusion denoising process but also improve the sampling quality through the global information of data distribution. In general form, for any state pair $(z_{n,t}^{g}, z_{n,0}^{g})$ , define the model's score ratio $s_{\theta}(z_{n,t}^{g})$ which represents the relative probability of transferring from $z_{n,t}^{g}$ to $z_{n,0}^{g}$ . SE is defined as: + +$$ +\begin{array}{l} s e = \sum_ {y \in z _ {n, 0: t - 1} ^ {g}} \omega_ {z _ {n, t} ^ {g}} ^ {g} \left(s _ {\theta} (z _ {n, t} ^ {g}) - \frac {p _ {d a t a} (y)}{p _ {d a t a} (z _ {n , t} ^ {g})} \log s _ {\theta} (z _ {n, t} ^ {g}) \right. \\ + K \left(\frac {p _ {\text {d a t a}} (y)}{p _ {\text {d a t a}} \left(z _ {n , t} ^ {g}\right)}\right), \tag {4} \\ \end{array} +$$ + +where $\omega_{z_{n,t}}^{g}$ is the weight of the loss term, which is used to balance the loss of different states. $K(a) = a(\log a - 1)$ is a normalization term that ensures the loss is non-negative. $\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}$ represents the actual score ratio. $p_{data}(y)$ and $p_{data}(z_{n,t}^{g})$ are the actual data distributions of the former noisy state and the current noisy state. The actual score ratio calculation relationship is shown in Theorem 1. + +Theorem 1. According to Bayes' theorem and the Gaussian distribution density formula, the following calculation relationship of $\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}$ is obtained: + +$$ +\frac {p _ {\text {d a t a}} (y)}{p _ {\text {d a t a}} \left(z _ {n , t} ^ {g}\right)} = \exp \left(\frac {\| z _ {n , t} ^ {g} \| ^ {2}}{2} - \frac {\| z _ {n , t} ^ {g} - \sqrt {\bar {\alpha} _ {t} ^ {g}} z _ {n , 0} ^ {g} \| ^ {2}}{2 \left(1 - \bar {\alpha} _ {t} ^ {g}\right)}\right). \tag {5} +$$ + +The proof is provided in Appendix A. + +Based on the SE [50], the model predicted score ratio indicates how the model adjusts the probability of the current state to tend to the original data distribution during the denoising process. The definition is as follows: + +$$ +s _ {\theta} \left(z _ {n, t} ^ {g}\right) = \frac {p _ {\theta} \left(z _ {n , 0} ^ {g}\right)}{p _ {\theta} \left(z _ {n , t} ^ {g}\right)}, \tag {6} +$$ + +where the denominator represents the probability of the current noise state and the numerator represents the original state probability estimated by the model. According to Theorem 2, the model uses softmax for normalization ensuring numerical stability and enabling gradient optimization when predicting the score ratio. + +Theorem 2. Given the denoising process modelled by a score-based probability ratio function $s_{\theta}(z_{n,t}^{g})$ , defined as Eq. (6), this paper defines a learnable approximation using a parameterized score function $f_{\theta}$ , such that the probability ratio can be estimated as: + +$$ +s _ {\theta} \left(z _ {n, t} ^ {g}\right) = \frac {\exp \left(f _ {\theta} \left(z _ {n , t} ^ {g} , z _ {n , 0} ^ {g}\right)\right)}{\sum_ {y \in z _ {n , 0 : t - 1} ^ {g}} \exp \left(f _ {\theta} \left(z _ {n , t} ^ {g} , y\right)\right)}, \tag {7} +$$ + +The proof is provided in Appendix A. + +# 3.2. Selection + +The selection process comprises two key steps: scan switch and selection. The scan switch mechanism captures temporal relationships between adjacent image patches (or text embeddings) by generating latent space representations with $k$ different sequential relationships, such as four image patch sequences and two text embedding sequences illustrated in Fig. 2fg. The mechanism creates $k$ temporal sequences $S = \{\langle z_{1,t}^{g},z_{2,t}^{g},\dots ,z_{i,t}^{g}\rangle \}_{k}$ . + +The selection step then analyzes these different sequential relationships at the current denoising step $t$ to determine which information should be focused on or ignored, thereby guiding the model's denoising direction in each diffusion step. The selection step chooses $j$ items $z_{n,t}^{g}$ from each sequence in $S$ according to the following Theorem 3. So, the selection step obtain $k$ selection sequences with different lengths, i.e., $S' = \{\langle z_{j_1,t}^g, z_{j_2,t}^g, \ldots, z_{j,t}^g \rangle\}_k$ and $S' \in S$ . + +Theorem 3. To achieve the optimal score entropy [50] which is demonstrated on Eq. (4), the selection step choose $j$ items where each $z_{n,t}^{g}$ satisfies $se = 0$ , i.e., + +$$ +s _ {\theta} \left(z _ {n, t} ^ {g}\right) \approx \frac {p _ {d a t a} (y)}{p _ {d a t a} \left(z _ {n , t} ^ {g}\right)} \tag {8} +$$ + +The proof is provided in Appendix A. + +# 4. Architecture + +The neural network architecture consists of two primary components: a VAE noisy latent encoder [42] and a multistep selection diffusion decoder, as illustrated in Fig. 2ab. The encoder first processes image data $X_{img}$ through patchify [19] operations and processes text data $X_{txt}$ through tokenization based on SentencePiece with Unigram BPE [45] and embedding operations, then uniformly maps them to the latent space before applying forward noise. + +The decoder, based on the multi-step selection diffusion model, leverages Mamba to achieve unified learning objectives while enhancing computational efficiency for processing long sequence data. It employs the SE [50] as the unified objective for both image and text modalities during the diffusion process. During selection, the model captures sequential relationships across different temporal dimensions using various scan switches. These relationships are then efficiently processed through the selection state-space structure in the Mamba Block determining which information to focus on or ignore according to Eq. (8), thereby guiding subsequent diffusion denoising steps (as shown in Fig. 2h). Finally, the reconstructed image patches and text embeddings are transformed back into their original data formats through a VAE noisy latent decoder [42]. + +# 4.1. The noisy latent encoder + +The noisy latent encoder first processes input image $X_{img}$ through patchify and processes text $X_{txt}$ through tokenization and embedding operations to obtain the patch sequence + +$G(X_{img} / X_{txt}) = \langle g_1, g_2, \dots, g_i \rangle$ , where $g_n$ represents the $n$ -th image patch or text embedding, respectively. The encoder VAE [42] generates Gaussian distribution parameters (mean $\mu$ and variance $\sigma$ ) for these patches, with a similar process applied to text embeddings, i.e., $VAE(G) = (\mu, \sigma)$ . For each image patch or text embedding $g_n$ , its noise $z_n$ is a sample $s_n$ from the distribution $\mathcal{N}(\mu, \sigma)$ with the addition noise $\epsilon_n \sim \mathcal{N}(0, 1)$ , i.e., $z_n = s_n + \epsilon_n$ . Finally, the image $X_{img}$ and text $X_{txt}$ are transformed into the noise sequence $\langle z_1, \dots, z_i \rangle$ through the above process. + +Moreover, three types of learnable padding tokens, time, category, and pad, are inserted into these noise sequences, as illustrated in Fig. 2de. The time token encodes the current diffusion step, the class token is used to learn the data category, and the pad token represents the start or end position for splitting these noise sequences. + +# 4.2. The multi-step selection diffusion decoder + +The decoder aims at progressively recovering the image $X_{img}$ or text $X_{txt}$ from noise sequences through two main modules: 1) the multi-step selection diffusion Mamba and 2) the VAE noisy latent decoder. 1) The Mamba is used to recover the patch sequence $\langle g_1, \dots, g_i \rangle$ from the noise sequence $\langle z_1, \dots, z_i \rangle$ . 2) The VAE noisy latent decoder assembles patches and generates the image $\hat{X}_{img}$ or text $\hat{X}_{txt}$ . + +# 4.2.1. Multi-step selection diffusion Mamba + +The module leverages two components, image/text scan switch and Mamba Block, to implement each denoising step in the multi-step selection diffusion model (Sec. 3). + +The image/text scan switch component establishes sequences with different directions to capture different temporal relationships between patches. Following Dim [73], we implement four distinct scan switches for images (as shown in Fig. 2f) and two for text (as shown in Fig. 2g). + +The Mamba block is used to select patches from these different scan switch sequences and denoise the input noise $z_{n,t}^{g}$ . The block adopts the state space architecture from Mamba-2 [26]. According to Sec. 3.2, it is $s_{\theta}$ , where $\theta = \{H_{n,t}^{g}, A, B, C, D, \Delta\}$ represent the state space in the block. The block comprises six key components: 1) linear input and output projection layers, 2) convolution kernel layer, 3) nonlinear activation layer, 4) state space model (SSM), 5) skip connection layer, and 6) normalization layer. + +1) The linear input projection layer reduces the dimensionality of the latent space noise vector while simultaneously applying initial state matrices $A$ , $B$ , $C$ to the linear projection of input data $z_{n,t}^{g}$ . Additionally, the linear output projection layer represents the denoising step, which transforms the selection noise $z_{n,t}^{g}$ into $z_{n,t - \Delta t}^{g}$ and outputs it to the next Mamba block according to the following equation. + +$$ +z _ {n, t - \Delta t} ^ {g} = z _ {n, t} ^ {g} - \frac {\Delta t}{2} \left[ f _ {\theta} \left(z _ {n, t} ^ {g}, t\right) + f _ {\theta} \left(z _ {n, t - \Delta t} ^ {g}, t - \Delta t\right) \right] (9) +$$ + +where the equation adopts the second-order numerical method of DPM-Solver [51] to improve sampling accuracy. Details are provided in Appendix B. + +2) The convolution kernel layer implements parallel scan switches, routing the initial linear projection of the input and the state matrix's linear projection through the SSM, as shown in Fig. 2i. The sweep down and sweep up [26] enable parallel computation between Eqs. (10) to (13). +3) The nonlinear layer enhances model generalization. +4) The SSM lets the Mamba block $s_{\theta}$ approximate the actual score ratio based on Theorem 3. To implement the target, SSM updates the state space $\theta$ by the following equations (based on Theorem 3 and details in Appendix A). + +$$ +H _ {n, t} ^ {g} = \bar {A} H _ {n, t - 1} ^ {g} + \bar {B} z _ {n, t} ^ {g} \tag {10} +$$ + +$$ +z _ {n - 1, t} ^ {g} = C H _ {n, t} ^ {g} + D z _ {n, t} ^ {g} \tag {11} +$$ + +$$ +\bar {A} = \exp (\Delta A) \tag {12} +$$ + +$$ +\bar {B} = (\Delta A) ^ {- 1} \cdot \left(\exp (\Delta A) - I\right) \cdot \Delta B \tag {13} +$$ + +where $H_{n,t}^{g}$ represents the hidden state representation, $A$ and $B$ control the evolution of hidden states and latent space noise vector inputs, respectively, $C$ governs the hidden state representation of the target output and $D$ manages the nonlinear skip connection for latent space noise vector inputs. $\Delta$ denotes the learnable time parameter. + +5) The skip connection layer facilitates input feature reuse and mitigates model degradation. +6) The Normalization layer ensures training stability. +According to Eq. (8) in Theorem 3 and Eq. (4), the goal of training the Mamba block is: + +$$ +L _ {s e} = \mathbb {E} _ {z _ {n, 0} ^ {g} \sim p _ {0}, z _ {n} ^ {g} \sim p (\cdot | z _ {n, 0} ^ {g})} s e = 0 \tag {14} +$$ + +# 4.2.2. The noisy latent decoder + +After applying the diffusion-based denoising process, the recovered latent variable $z_{n,0}^{g}$ is passed to the VAE decoder [42] as illustrated in Fig. 2c. For image reconstruction, the decoder applies an $\ell_2$ loss: + +$$ +L _ {r e c} ^ {i m g} = \mathbb {E} _ {z _ {n, 0} ^ {g} \sim q _ {\phi} (z | X)} \| X _ {i m g} - \hat {X} _ {i m g} \| ^ {2}. \tag {15} +$$ + +where $q_{\phi}(z|X)$ represents the posterior distribution of the VAE encoder. + +For text, the decoder minimizes the cross-entropy loss: + +$$ +L _ {r e c} ^ {t x t} = - \mathbb {E} _ {z _ {n, 0} ^ {g} \sim q _ {\phi} (z | X)} \sum_ {t} p \left(X _ {t x t} ^ {(t)} \mid z _ {n, 0} ^ {g}\right) \log p _ {\psi} \left(\hat {X} _ {t x t} ^ {(t)} \mid z _ {n, 0} ^ {g}\right). \tag {16} +$$ + +where $p(X_{txt}^{(t)}|z_{n,0}^{g})$ represents the probability distribution of real text data under the condition of latent variable $z_{n,0}^{g}$ . + +![](images/953d44ded220231e3b4ec9deadbc4fb7cc804a29d77460b4fb29182ebc1593d1.jpg) + +![](images/d18d89bf12c3d61711c87fa4e2fa0800653c74b98ff3378c3001a1654b9e784a.jpg) +Figure 3. VQA, QA and Multi-Modal generation test from MDM. The results of VQA are part of VQAv2 [25]. The QA results are part of PIQA [7] and MMLU [32]. The Multi-Modal generation results are tested with ground-truth data. +Figure 4. Comparison between each model on generating captioning and image results on COCO dataset. Unlike other models, MDM generates both image and caption data simultaneously. + +And $p_{\psi}(\hat{X}_{txt}^{(t)}|z_{n,0}^g)$ represents the probability distribution of the text token generated by the VAE decoder under the condition of the latent variable $z_{n,0}^g$ + +Besides, a KL divergence regularizes the latent space: + +$$ +L _ {K L} = D _ {K L} \left(q _ {\phi} (z | X) \| p (z)\right). \tag {17} +$$ + +where $p(z)$ represents the prior distribution of the latent variable by VAE, which is assumed to be a standard Gaussian distribution $\mathcal{N}(0,I)$ to regularize the latent variable space and enable it to have smooth generation capabilities. + +The final optimization objective integrates VAE reconstruction, KL divergence and SE: + +$$ +L _ {t o t a l} = L _ {r e c} ^ {i m g} + L _ {r e c} ^ {t x t} + \beta L _ {K L} + \lambda L _ {s e}. \tag {18} +$$ + +# 5. Experiments + +# 5.1. Experimental Setup + +Model configuration. Our model applies a VAE [42] as the noisy latent encoder and decoder. Moreover, it integrates + +the DiM selection state space [73] in each Mamba block as the diffusion decoder. The resulting model contains 7 billion parameters, with 49 Mamba blocks in the multi-step selection diffusion decoder, each having a dimension of 2048 (Details of parameter settings listed in Appendix C). + +Before the training MDM process, we trained a tokenization model based on SentencePiece (Unigram BPE) [45]. The tokenization model can help the model construct a stable text latent variable representation, thereby optimizing the forward diffusion and reverse denoising process. See Appendix D for detailed experimental settings. + +In the training process, we import the DDPM scheduler [35] and DPM-Solver [51] to improve the sampling efficiency in the diffusion model. We then use the AdamW optimizer without weight decay, maintaining a constant learning rate of 0.0001. Meanwhile, we keep an EMA of the model weights with a coefficient of 0.9999. + +Baseline and dataset. Our evaluation encompasses four tasks: image generation with classifier-free guidance [34] (CFG), text-to-image, image-to-text, and text-to-text generation. For the baseline model training, we train MDM on ImageNet [13], JourneyDB [68] and UltraChat [16]. + +For the image generation and the text-to-image task at $256 \times 256$ resolution, we compare the MDM baseline model against established baselines across three categories: diffusion models (Imagen [64], ADM [15], CDM [36], LDM [63], DiT-XL/2 [57], SDXL [58], and SD-3 [21]), autoregressive models (VQGAN [20] and ViT-VQGAN [85]), and end-to-end multi-modal models (NExT-GPT [78], Chameleon [71], LlamaGen [69], Transfusion [90], MonoFormer [89], Dual-DiT [47], JanusFlow [52] and ShowO [79]). For the image generation task, we evaluate performance on ImageNet [13] using four metrics: Frechet Inception Distance (FID), Inception Score (IS), and Precision/Recall. For the text-to-image task, we evaluate performance on COCO [40] using FID and Gen Eval [24]. + +For the image-to-text task (image captioning and vision question answering, VQA) and text-to-text task, we employ MDM baseline model and MDM instruction model by visual instruction tuning [48] on multiple datasets: COCO [40], GQA [38], OCR-VQA [55], TextVQA [67], and VisualGenome [44]. We evaluate the model against two groups of baselines: traditional models and end-to-end multi-modal models. Performance evaluation of image captioning is conducted on Flickr 30K [84] and COCO [40] datasets using the Consensus-based Image Description Evaluation (CIDEr) metric. And performance evaluation of VQA is conducted on VQAv2 [25], VizWiz [28], and OKVQA [53] using answer accuracy rate as the evaluation metric. + +For the text-to-text task, we evaluate the model on text comprehension and reasoning tasks using HellaSwag [87], OpenBookQA [54], Wino-Grande [65], ARCEasy, ARC + +
ModelArcParamsImage Generation with CFGText-to-Image Generation
FID ↓IS ↑Pre ↑Re ↑FID ↓Gen Eval ↑
Imagen [64]Diff7.3B----7.27-
ADM [15]Diff554M10.94101.00.690.63--
CDM [36]Diff-4.88158.7----
LDM [63]Diff400M3.60147.60.870.68-0.43
DiT-XL/2 [57]Diff675M2.27278.20.830.57--
SDXL [58]Diff3.4B----4.400.55
SD-3 [21]Diff12.7B-----0.68
VQGAN [20]AR227M18.6580.40.780.26--
ViT-VQGAN [85]AR1.7B4.17175.1----
NExT-GPT [78]AR7B----10.07-
Chameleon [71]AR7B----26.740.39
LlamaGen [69]AR3.1B2.81311.50.840.544.19-
Transfusion [90]AR+Diff7.3B----6.780.63
MonoFormer [89]AR+Diff1.1B2.57272.60.840.56--
Dual-DiT [47]Diff2B----9.400.65
JanusFlow [52]AR+Diff1.3B-----0.70
Show-O [79]AR+Diff1.3B----9.240.68
MDMDiff7B2.49281.40.860.595.910.68
+ +Table 1. Performance on ImageNet and COCO ${256} \times {256}$ . FID,IS,Pre,and Re stands for Frechet Inception Distance,Inception Score, Precision, and Recall, respectively. + +
ModelICVQAText Comprehension and ReasoningMath and World
FlickrCOCOVQAv2VizWizOKHSOBQAWGARCEARCCBoolQPIQAGSM8kMATHMMLU
Llama-2 [74] (7B)-----77.258.678.575.245.977.478.814.62.545.3
Mistral [39] (7B)-----81.3-75.380.055.584.783.052.113.160.1
Flamingo [2] (80B)75.1113.867.6------------
Gemini Pro [72]82.299.871.2--84.7------86.532.671.8
GPT4V [8]55.378.577.2--95.3------92.052.986.4
InstructBLIP [48] (7B)82.4102.2-33.433.9----------
mPLUG-Owl [83] (7B)80.3119.3-39.0-----------
TinyLlama [88] (1.1B)-----59.236.059.155.330.157.873.3---
Pythia [6] (12B)-----52.033.257.454.028.563.370.9---
DREAMLLM [17](7B)-115.456.645.844.3----------
Emu [70](7B)-117.740.035.434.7----------
Chameleon [71](34B)74.7120.266.0--74.251.070.476.146.581.479.641.611.552.1
NExT-GPT [78](7B)84.5124.966.748.452.1----------
Transfusion [90](7B)-33.7-------------
MonoFormer [89](1.1B)-----50.637.256.948.231.562.371.2---
Dual-DiT [47](2B)-56.260.129.925.3----------
JanusFlow [52](1.3B)--79.8------------
Show-O [79](1.3B)67.6-74.7------------
MDM (7B)62.4109.660.339.847.170.641.568.855.146.265.779.940.512.154.4
InstructMDM (7B)75.2122.166.746.351.674.848.374.965.447.171.583.746.013.159.2
+ +Table 2. Performance on image-to-text and text-to-text tasks. The evaluation of image captioning (IC) and VQA is CIDEr and answer accuracy % (Flickr is evaluated on 30K and OK represents OKVQA). + +Challenge [10], BoolQ [9], and PIQA [7]. We also evaluate the model on math and world knowledge tasks using GSM8K [11], MATH [33], and MMLU [32]. The evaluation metrics for all the tasks are accuracy rates. + +# 5.2. Experimental Results + +Image Generation. In the image generation task on ImageNet, MDM achieves top-three rankings across all eval + +uation metrics: second in FID, IS, and Precision, and third in Recall when compared against one-modal diffusion models and end-to-end multi-modal models (see Tab. 1). MDM demonstrates superior overall performance, notably surpassing other end-to-end multi-modal models in three of the four metrics. In the text-to-image task, we tested the model on the COCO dataset to generate both image and + +![](images/de077a502ef4c2b76e1729731b18c36fbbfd4c0b9fe749887897a94cf1ad89c3.jpg) +Figure 5. Comparison between Mamba Baseline, MonoFormer, and ours MDM on inference speed test. The left shows the inference speed of the model FPS at different resolutions. The right shows the inference speed of the model latency at different token lengths. + +![](images/2a8790f8903f155935ff66a3aae2ea185867caf22ff191eac4dba5c23dbcb7b2.jpg) + +
ModelImage/Text +Scan SwitchFPS w log +scale↑FID↓
Model w Mamba①②③④/①②1.3572.49
Model w Mamba①②/①1.4053.96
Model w Transformer-1.9146.72
+ +Table 3. Ablation on ImageNet ${256} \times {256}$ image generation. + +caption data. For the image generation results, we evaluated the FID and Gen Eval performance indicators of the model-generated images. MDM still achieved the top three performance levels and achieved SOTA on Gen Eval. + +Text Generation. In the image-to-text task image captioning, according to the settings on image generation on the COCO dataset, we tested the caption data of the model based on the model outputting both text and image data using the CIDEr indicator. The results showed that MDM ranked second among all models, as shown in Tab. 2. While in task VQA, MDM achieves competitive performance, surpassing several traditional models including InstructBLIP, mPLUG-Owl, DREAMLLM, and Emu, although it still trails behind top-performing models in the field as shown in Tab. 2. In the text-to-text generation task, as shown in Tab. 2, MDM and the other end-to-end multi-modal models perform worse than well-known traditional models. This discrepancy may be attributed to the fact that these end-to-end models have some deviations in multimodal fusion and learning because they abandon multiple language encoders, visual encoders, and multimodal fusion encoders. However, when compared with the other two end-to-end models, MDM excels, outperforming MonoFormer and surpassing Chameleon on seven out of ten datasets. + +# 5.3. Discussion + +# 5.3.1. Performance Analysis + +As demonstrated in Fig. 3, MDM shows the ability to generate image and text simultaneously in multiple rounds of dialogue and perform well in QA&VQA. Some results even exceed those of GPT-4V, particularly evident in the second and third rows of Fig. 4 which is a hybrid output process for the MDM model. Due to this, we set the model to generate corresponding images for the description text while simul + +taneously generating image captioning. + +This enhanced performance stems from MDM's multi-step selection diffusion decoder, which leverages Mamba's integrated selection and denoising capabilities to maintain focused attention on both textual and visual details. Validating our complexity analysis in Appendix E, MDM demonstrates superior efficiency compared to end-to-end Transformer models when processing long sequences, as shown in Fig. 5, particularly outperforming other end-to-end multimodal models for sequences exceeding 1280 tokens. + +# 5.3.2. Ablations + +Our ablation studies examine the impact of both the selection process and Mamba block components. Reducing the number of image/text scan switch sequences from 6 ('①②③④/①②') to 3 ('①②/①'), as shown in Tab. 3, improves inference speed but degrades image quality, as fewer scan switch sequences limit the model's ability to capture accurate information in complex sequences. Additionally, replacing the Mamba block with the Transformer further deteriorates output image quality, suggesting Mamba's temporal network architecture is better suited for representing diffusion relationships during the denoising process. + +# 6. Conclusion + +This paper introduces MDM (Multi-Modal Diffusion Mamba), a novel end-to-end architecture that significantly enhances multi-modal processing through two key innovations: a unified diffusion objective and an efficient selection mechanism leveraging Mamba's state-space structure. By integrating variational autoencoder with multi-step selection diffusion, MDM achieves SOTA overall performance in image generation and demonstrates remarkable versatility across various tasks, including image-to-text, text-to-text and text-image-to-text-image. Our comprehensive experiments illustrate that MDM consistently surpasses traditional end-to-end multi-modal models, particularly in processing high-resolution images and long-sequence text, while maintaining computational efficiency. The model's ability to unify different modalities under a single objective, coupled with its superior management of temporal relationships in the diffusion process, establishes a promising direction for future multi-modal architecture. + +# References + +[1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 1 +[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1, 3, 7 +[3] Silas Alberti, Niclas Dern, Laura Thesing, and Gitta Kutyniok. Sumformer: Universal approximation for efficient transformers. In Topological, Algebraic and Geometric Learning Workshops 2023, pages 72-86. PMLR, 2023. 1 +[4] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. Advances in Neural Information Processing Systems, 35:32897-32912, 2022. 1, 3 +[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sağnak Taşırlar. Introducing our multimodal models, 2023. 1, 3 +[6] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 7 +[7] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, pages 7432-7439, 2020. 2, 6, 7 +[8] GPTV System Card. Openai, 2023. 7 +[9] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 2, 7 +[10] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 2, 7 +[11] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 7 +[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image + +database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 6 +[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3 +[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 7 +[16] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233, 2023. 6 +[17] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3, 7 +[18] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-mamba for cross-modality object detection. arXiv preprint arXiv:2404.09146, 2024. 1, 3 +[19] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 4 +[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6, 7 +[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 7 +[22] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Youqiang Zhang, and Junshi Huang. Dimba: Transformermamba diffusion models. arXiv preprint arXiv:2406.01159, 2024. 1 +[23] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 3 +[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 6 +[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 6 +[26] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 5 + +[27] Ahan Gupta, Yueming Yuan, Yanqi Zhou, and Charith Mendis. Flurka: Fast fused low-rank & kernel attention. arXiv preprint arXiv:2306.15799, 2023. 1 +[28] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 2, 6 +[29] Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David P Woodruff, and Amir Zandieh. Hyperattention: Long-context attention in near-linear time. arXiv preprint arXiv:2310.05869, 2023. 1 +[30] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 1 +[31] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 3 +[32] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 2, 6, 7 +[33] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021. 2, 7 +[34] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6 +[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 6 +[36] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 6, 7 +[37] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 1 +[38] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6 +[39] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7 +[40] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 2, 6 + +[41] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1, 3 +[42] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 3, 4, 5, 6 +[43] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3 +[44] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6 +[45] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4, 6 +[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1, 3 +[47] Zijie Li, Henry Li, Yichun Shi, Amir Barati Farimani, Yuval Kluger, Linjie Yang, and Peng Wang. Dual diffusion for unified image generation and understanding. arXiv preprint arXiv:2501.00289, 2024. 6, 7 +[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7 +[49] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 1, 3 +[50] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4 +[51] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 5, 6 +[52] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv preprint arXiv:2411.07975, 2024.6.7 +[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 2, 6 +[54] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new + +dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 2, 6 +[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6 +[56] Matteo Pagliardini, Daniele Paliotta, Martin Jaggi, and François Fleuret. Fast attention over long sequences with dynamic sparse flash attention. Advances in Neural Information Processing Systems, 36:59808-59831, 2023. 1 +[57] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 7 +[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 6, 7 +[59] Ramya Prabhu, Ajay Nayak, Jayashree Mohan, Ramachandran Ramjee, and Ashish Panwar. vattention: Dynamic memory management for serving llms without pagedattention. arXiv preprint arXiv:2405.04437, 2024. 1 +[60] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 1, 3 +[61] Haohao Qu, Liangbo Ning, Rui An, Wenqi Fan, Tyler Derr, Xin Xu, and Qing Li. A survey of mamba. arXiv preprint arXiv:2408.01129, 2024. 1 +[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3 +[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7 +[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3, 6, 7 +[65] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 2, 6 +[66] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 1 +[67] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6 + +[68] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. Advances in Neural Information Processing Systems, 36, 2024. 6 +[69] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 3, 6, 7 +[70] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 7 +[71] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 3, 6, 7 +[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7 +[73] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 1, 2, 5, 6 +[74] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 7 +[75] Zhongwei Wan, Xin Wang, Che Liu, Samiul Alam, Yu Zheng, Jiachen Liu, Zhongnan Qu, Shen Yan, Yi Zhu, Quanlu Zhang, et al. Efficient large language models: A survey. arXiv preprint arXiv:2312.03863, 2023. 1 +[76] Zifu Wan, Pingping Zhang, Yuhao Wang, Silong Yong, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Sigma: Siamese mamba network for multi-modal semantic segmentation. arXiv preprint arXiv:2404.04256, 2024. 1, 3 +[77] Xinghan Wang, Zixi Kang, and Yadong Mu. Text-controlled motion mamba: Text-instructed temporal grounding of human motion. arXiv preprint arXiv:2404.11375, 2024. 3 +[78] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 1, 3, 6, 7 +[79] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6, 7 +[80] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8239-8249, 2024. 1 +[81] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 3 + +[82] Zhe Yang, Wenrui Li, and Guanghui Cheng. Shmamba: Structured hyperbolic state space model for audio-visual question answering. arXiv preprint arXiv:2406.09833, 2024. 1, 3 +[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7 +[84] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 2, 6 +[85] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 6, 7 +[86] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 3 +[87] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 2, 6 +[88] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 7 +[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 1, 2, 3, 6, 7 +[90] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 3, 6, 7 \ No newline at end of file diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/images.zip b/2025/End-to-End Multi-Modal Diffusion Mamba/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1e10d80611b52ac1dd7fe107bb60d02f9def3890 --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:145dc136697a0430398ce8c652d9068f4bfc36009bb22a9a4a0109e7ab63661f +size 627022 diff --git a/2025/End-to-End Multi-Modal Diffusion Mamba/layout.json b/2025/End-to-End Multi-Modal Diffusion Mamba/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e6d74003253d5f6ef4ece13b12ab2634800cc87a --- /dev/null +++ b/2025/End-to-End Multi-Modal Diffusion Mamba/layout.json @@ -0,0 +1,10838 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 170, + 103, + 440, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 103, + 440, + 119 + ], + "spans": [ + { + "bbox": [ + 170, + 103, + 440, + 119 + ], + "type": "text", + "content": "End-to-End Multi-Modal Diffusion Mamba" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 165, + 142, + 455, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 142, + 455, + 158 + ], + "spans": [ + { + "bbox": [ + 165, + 142, + 455, + 158 + ], + "type": "text", + "content": "Chunhao Lu1 Qiang Lu1⊗ Meichen Dong1,2 Jake Luo3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 160, + 559, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 160, + 559, + 175 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 559, + 175 + ], + "type": "text", + "content": "1China University of Petroleum-Beijing, 2Leyard Optoelectronic, 3University of Wisconsin-Milwaukee" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 92, + 180, + 528, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 180, + 528, + 192 + ], + "spans": [ + { + "bbox": [ + 92, + 180, + 528, + 192 + ], + "type": "text", + "content": "{luchunhao,meichen.dong}@student.cup.edu.cn luqiang@cup.edu.cn jakeluo@uwm.edu" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 220, + 200, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 220, + 200, + 232 + ], + "spans": [ + { + "bbox": [ + 151, + 220, + 200, + 232 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 247, + 297, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 510 + ], + "type": "text", + "content": "Current end-to-end multi-modal models utilize different encoders and decoders to process input and output information. This separation hinders the joint representation learning of various modalities. To unify multi-modal processing, we propose a novel architecture called MDM (Multi-modal Diffusion Mamba). MDM utilizes a Mamba-based multistep selection diffusion model to progressively generate and refine modality-specific information through a unified variational autoencoder for both encoding and decoding. This innovative approach allows MDM to achieve superior performance when processing high-dimensional data, particularly in generating high-resolution images and extended text sequences simultaneously. Our evaluations in areas such as image generation, image captioning, visual question answering, text comprehension, and reasoning tasks demonstrate that MDM significantly outperforms existing end-to-end models (MonoFormer, LlamaGen, and Chameleon etc.) and competes effectively with SOTA models like GPT-4V, Gemini Pro, and Mistral. Our results validate MDM's effectiveness in unifying multi-modal processes while maintaining computational efficiency, establishing a new direction for end-to-end multi-modal architectures." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 135, + 529 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 296, + 714 + ], + "type": "text", + "content": "Traditional large-scale multi-modal models [2, 4, 17, 21, 41, 43, 46, 48, 58, 62, 64, 86] typically use multiple encoders and decoders to process multi-modal data. This approach makes learning a unified joint representation of the multi-modal data difficult and can significantly slow inference time (as shown in Fig. 1A). To alleviate these problems, end-to-end models without modal-fusion en(de)coder architecture have been proposed (as shown in Fig. 1B). This approach offers a streamlined, unified processing framework that enhances efficiency and consistency in multimodal representation learning. Existing end-to-end models follow three primary strategies: (1) Autoregressive models [5, 31, 69, 71] leverage a single Transformer for both text and image generation, but struggle with the inherent sequential dependency of autoregressive decoding. (2) Hy" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 219, + 395, + 336 + ], + "blocks": [ + { + "bbox": [ + 317, + 219, + 395, + 336 + ], + "lines": [ + { + "bbox": [ + 317, + 219, + 395, + 336 + ], + "spans": [ + { + "bbox": [ + 317, + 219, + 395, + 336 + ], + "type": "image", + "image_path": "69bf2ddb13c63877bd884b7241cb6f3ce9497cdd50221666359c3fee6333d1ed.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 337, + 388, + 343 + ], + "lines": [ + { + "bbox": [ + 323, + 337, + 388, + 343 + ], + "spans": [ + { + "bbox": [ + 323, + 337, + 388, + 343 + ], + "type": "text", + "content": "(A) Traditional Multi-Modal Model" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 395, + 219, + 475, + 336 + ], + "blocks": [ + { + "bbox": [ + 395, + 219, + 475, + 336 + ], + "lines": [ + { + "bbox": [ + 395, + 219, + 475, + 336 + ], + "spans": [ + { + "bbox": [ + 395, + 219, + 475, + 336 + ], + "type": "image", + "image_path": "93064fba34e137e7c3670503d22b47d150f89cf7ea5f93a5cdb1197d8ccd1e28.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 337, + 468, + 343 + ], + "lines": [ + { + "bbox": [ + 402, + 337, + 468, + 343 + ], + "spans": [ + { + "bbox": [ + 402, + 337, + 468, + 343 + ], + "type": "text", + "content": "(B) End-To-End Multi-Modal Model" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 476, + 219, + 553, + 336 + ], + "blocks": [ + { + "bbox": [ + 476, + 219, + 553, + 336 + ], + "lines": [ + { + "bbox": [ + 476, + 219, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 476, + 219, + 553, + 336 + ], + "type": "image", + "image_path": "68c2974f2548c1b20ccbd9c5d1a25a22dfe803286cf6563818771aed8a644328.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 338, + 523, + 343 + ], + "lines": [ + { + "bbox": [ + 505, + 338, + 523, + 343 + ], + "spans": [ + { + "bbox": [ + 505, + 338, + 523, + 343 + ], + "type": "text", + "content": "(C) Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 347, + 345, + 520, + 356 + ], + "lines": [ + { + "bbox": [ + 347, + 345, + 520, + 356 + ], + "spans": [ + { + "bbox": [ + 347, + 345, + 520, + 356 + ], + "type": "text", + "content": "Figure 1. Comparison of three types of models." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 555, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 434 + ], + "type": "text", + "content": "brid image generation models [23, 78] integrate an additional image synthesis module, improving image quality but introducing extra complexity. (3) Mixed autoregressive-diffusion models [89, 90] employ diffusion-based image generation while maintaining an autoregressive framework for text, yet still struggles with unifying multi-modal." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 556, + 640 + ], + "type": "text", + "content": "Despite recent advancements, Transformer-based end-to-end models face several critical challenges: (1) their quadratic computational complexity makes them inefficient for generating high-resolution image and long-sequence text. Although various studies have attempted to optimize this computational complexity [1, 3, 12, 27, 29, 56, 59, 66, 74, 75], the challenge remains substantial. (2) their reliance on multi-objective learning introduces conflicting optimization goals, impeding convergence and hindering effective joint representation learning. In contrast, statespace models like Mamba [26, 61] offer a compelling alternative due to their ability to scale linearly with sequence length while effectively capturing long-range dependencies. However, the current multi-modal implementations of Mamba [18, 22, 30, 37, 49, 60, 73, 76, 80, 82] still adopt a multi-objective approach, limiting their capacity for end-to-end joint representation learning." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "To effectively process multi-modal data, we propose an end-to-end model called the Multi-Modal Diffusion Mamba (MDM) (as shown in Fig. 1c). MDM first employs patchify [19] and embedding to pre-process multi-modal data. Then, it uses a variational autoencoder (VAE) [42] as a multi-modal encoder, which uniformly maps the multi" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20529" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 553, + 149 + ], + "type": "image", + "image_path": "15d2a8e5c773c4f13c81be3b9ad8a35e1b558ae70f915c32dc1c8b0a3a6b323d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 151, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 59, + 151, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 59, + 151, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 59, + 151, + 553, + 285 + ], + "type": "image", + "image_path": "52cba87c0cc5afb41e3e3c8706d008e2f63434b17c4ee4ed22fe46a8d9c49508.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 287, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 55, + 287, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Framework of Multi-Modal Diffusion Mamba. MDM first encodes inputs (caption, VQVAE-processed image, question) using VAE (a), while performing padding (class, diffusion timestep, token completion) and flatten operations (d, e). Next, data reconstruction is progressively completed via diffusion mamba operations (b), modeling images and text temporally through scanning processes (f, g) for efficient information selection (red boxes indicate selection). Selected data undergoes computation (i) guided by (h) within the Mamba-2 framework to update model parameters. Finally, the MDM output passes through the VAE decoder (c) to reconstruct real data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 295, + 397 + ], + "type": "text", + "content": "modal data to a noisy latent space (as illustrated in Fig. 2a). MDM constructs a multi-step selection diffusion model based on the Mamba architecture as a uniform decoder for the rapid generation of multi-modal information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 399, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 628 + ], + "type": "text", + "content": "This decoder generates the target text or image step-by-step based on the diffusion process through the multi-step selection diffusion model (as shown in Fig. 2b). To enhance decoding speed, the decoder employs the Score Entropy Loss [50] as the objective function instead of Markov chain-based [35] methods for updating the network to handle multi-modal data throughout the diffusion process. The decoder comprises two components: an image and text scan switch, and a Mamba-2 block [26]. The text scan switch has two modes for sequence modeling (as shown in Fig. 2f), while the image scan switch has four, based on the settings of DiM [73] (as shown in Fig. 2e). The scan switches enable the model to capture sequential relationships across various temporal directions in the data. The selection state-space structure in Mamba then analyzes these sequential relationships within the current denoising step. This analysis guides the selection of relevant information to focus on and irrelevant information to ignore, effectively directing the model's denoising process at each step." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Since MDM unifies the modality encoder and decoder, the model is capable of generating an image and text simultaneously. For example, as shown in Fig. 2h, when generating an image of a dog alongside its description, the scan switch in the decoder first assesses whether the description contains conditions that necessitate image generation. If such conditions exist, the image scan switch is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 555, + 456 + ], + "type": "text", + "content": "activated. Consequently, the model directs its selection to the image patches corresponding to the dog during each denoising step. This targeted focus guides the model to effectively denoise relevant pixels while disregarding other areas of the image. A similar selection process is employed for text data. Ultimately, the data, once denoised via the t-step diffusion process, is reconstructed into authentic text (or an image) through the VAE decoder simultaneously. The main contributions of this paper are as follows." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 463, + 556, + 718 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "content": "1) We introduce the Multi-Modal Diffusion Mamba (MDM), an end-to-end model that achieves a computational complexity of " + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(MLN^2)" + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "content": ", outperforming previous end-to-end models like MonoFormer [89], which operate at " + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(ML^2 N / G)" + }, + { + "bbox": [ + 313, + 463, + 555, + 536 + ], + "type": "text", + "content": ". This advancement enables the efficient generation of long-sequence text and high-resolution images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 542, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 556, + 602 + ], + "type": "text", + "content": "2) We propose a novel multi-step selection diffusion model that combines autoregressive and diffusion-based generative paradigms into a unified learning objective. This method effectively integrates both paradigms within a diffusion process, generating multi-modal data simultaneously." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 609, + 556, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 556, + 718 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 556, + 718 + ], + "type": "text", + "content": "3) Our experimental results demonstrate MDM's superior performance in image generation on the ImageNet [13] and COCO datasets [40]. Additionally, it excels in various tasks, including image captioning on Flickr30K [84] and COCO [40], VQA on VQAv2 [25], VizWiz [28], and OKVQA [53], as well as text comprehension and reasoning on seven datasets [7, 9, 10, 54, 65, 87]. Furthermore, MDM shows strong results in math-related world knowledge tasks on GSM8k [11], MATH [33], and MMLU [32]." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20530" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 147, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 249, + 98 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 249, + 98 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 249, + 98 + ], + "type": "text", + "content": "2.1. Traditional large multi-modal model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 99, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 99, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 99, + 296, + 277 + ], + "type": "text", + "content": "Most existing LMMs are built by integrating architectures from multiple modalities [2, 4, 41, 46, 62, 86]. SOTA image and video generation models employ pre-trained text encoders to represent input prompts in latent space, which then condition a diffusion model for generating videos and images [64]. Many researchers have adopted this approach, fusing feature representations from multiple pretrained encoders to enhance model performance across different modalities [21, 58]. This pattern is also prevalent in visual language models, where pre-trained language models are typically augmented with linear projection layers from other pre-trained en/decoders for training in the text space. Examples include Flamingo [2] and LLaVA [48] for visual understanding, GILL [43] for visual generation, and DreamLLM [17] for both understanding and generation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 279, + 226, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 226, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 226, + 290 + ], + "type": "text", + "content": "2.2. End-to-End multi-modal model" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 291, + 295, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 410 + ], + "type": "text", + "content": "End-to-end models have emerged recently to facilitate joint representation learning while improving training and inference efficiency. It can be categorized into three main types: 1) The autoregressive model [5, 31, 69, 71] utilizes one Transformer with an autoregressive approach to generate images and text. For instance, the Fuyu model [5] processes image patches directly as input to achieve visual comprehension. Models like Chameleon [71], Mars [31], and LlamaGen [69] convert images into discrete sequence tokens, then concatenate them with text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 411, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 506 + ], + "type": "text", + "content": "2) The hybrid image generation model [23, 78] addresses the limitations of autoregressive approaches in image generation. While maintaining an autoregressive structure for text generation, the models enhance image quality by incorporating an image-generation network. For example, Seedx model [23] focuses on enhancing specific aspects of image generation, while Next-GPT [78] aims to expand multimodal capabilities within an end-to-end framework." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 507, + 296, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 296, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 296, + 602 + ], + "type": "text", + "content": "3) The mixed autoregressive-diffusion model [89, 90] combines the strengths of previous approaches. It performs text autoregressive generation and image diffusion restoration simultaneously. Models like MonoFormer [89] and Transfusion [90] achieve this by incorporating causal self-attention [81] for text tokens and bidirectional self-attention [14] for image patches, enabling high-quality multi-modal understanding and generation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 219, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 219, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 219, + 616 + ], + "type": "text", + "content": "2.3. Mamba in multi-modal model" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Mamba has emerged as a powerful alternative to Transformer for multi-modal data alignment [18, 49, 76, 77, 82]. Recent works showcase Mamba's capabilities across different multi-modal applications. VL-Mamba [60] combines a pre-trained Mamba model for language understanding with a connector module to align visual patches and language tokens. However, these models lack end-to-end training capabilities and struggle to learn unified joint representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "tions. MDM provides a truly end-to-end architecture, enabling rapid generation of high-quality, long sequences." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 99, + 518, + 113 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 518, + 113 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 518, + 113 + ], + "type": "text", + "content": "3. Multi-step Selection Diffusion Model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 114, + 555, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 555, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 555, + 247 + ], + "type": "text", + "content": "The multi-step selection diffusion model enables rapid generation of multi-modal information through two key processes: diffusion & denoising and selection. During the diffusion & denoising, the model employs a unified Score Entropy Loss [50](SE) to gradually reconstruct target data from noise through a series of denoising steps (as illustrated in Fig. 2b). The selection process enables the model to capture sequential relationships across different temporal dimensions in the latent space, determining which information should be focused on or ignored during each diffusion denoising step (as shown in Fig. 2h)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 247, + 441, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 441, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 441, + 259 + ], + "type": "text", + "content": "3.1. Diffusion & Denoising" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 260, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 297 + ], + "type": "text", + "content": "The diffusion & denoising process comprises two main components: diffusion and denoising. The diffusion component can be expressed by the following equation:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 365, + 301, + 555, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 301, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 365, + 301, + 555, + 323 + ], + "type": "interline_equation", + "content": "z _ {n, t} ^ {g} = \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g} + \\sqrt {1 - \\bar {\\alpha} _ {t} ^ {g}} \\epsilon_ {n, t} ^ {g}, \\tag {1}", + "image_path": "4c230f24a35a071e3e546928c93c525d3495a0586f7ea57ae7a5efbcf699cd59.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "g" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " denotes either image patch or text embedding, and " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " represents the latent space vector of the " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": "-th image patch or text embedding, obtained through VAE sampling [42]. " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " after " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " steps of noise addition; " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\epsilon_{n,t}^{g} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " represents the added noise; " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t}^{g} = \\prod_{k=1}^{t} \\alpha_{k}^{g}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\alpha_{k}^{g} = 1 - \\beta_{k}^{g}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\{\\beta_{k}^{g} \\in (0,1)\\}_{k=1}^{T}" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": " are Gaussian distribution hyperparameters controlling the forward diffusion noise. Following the diffusion Markov principle [35], " + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 327, + 555, + 449 + ], + "type": "text", + "content": "-step forward diffusion process can be characterized by conditional probabilities as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 340, + 455, + 555, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 455, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 340, + 455, + 555, + 475 + ], + "type": "interline_equation", + "content": "p (z _ {n, t} ^ {g} | z _ {n, 0} ^ {g}) = \\mathcal {N} (z _ {n, t} ^ {g}; \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n, 0} ^ {g}, (1 - \\bar {\\alpha} _ {t} ^ {g}) I), \\quad (2)", + "image_path": "6f710daabb627fa3283ce5238c7a3a4a83f4d258b0f3587d1f5c325ade036c01.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "text", + "content": "which means that given " + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}, z_{n,t}^{g}" + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "text", + "content": " follows a Gaussian distribution with " + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "inline_equation", + "content": "\\sqrt{\\bar{\\alpha}_t^g} z_{n,0}^g" + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "text", + "content": " as mean and " + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "inline_equation", + "content": "(1 - \\bar{\\alpha}_t^g)I" + }, + { + "bbox": [ + 313, + 480, + 553, + 507 + ], + "type": "text", + "content": " as variance." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "text", + "content": "In the classic diffusion denoising component [35], the model needs to learn the posterior " + }, + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "inline_equation", + "content": "p(z_{n,t-1}^g | z_{n,t}^g)" + }, + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "text", + "content": " to gradually reconstruct the data. Since " + }, + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "inline_equation", + "content": "p(z_{n,t}^g | z_{n,0}^g)" + }, + { + "bbox": [ + 313, + 508, + 554, + 567 + ], + "type": "text", + "content": " follows a Gaussian distribution, we can assume that the approximate distribution of the denoising process is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 572, + 555, + 588 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 572, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 329, + 572, + 555, + 588 + ], + "type": "interline_equation", + "content": "p _ {\\theta} (z _ {n, t - 1} ^ {g} | z _ {n, t} ^ {g}) = \\mathcal {N} (z _ {n, t - 1} ^ {g}; \\mu_ {\\theta} (z _ {n, t} ^ {g}), (\\sigma_ {\\theta , n} ^ {g}) ^ {2}). \\qquad (3)", + "image_path": "937d384df1bd14303807362842301674fcf4c0a6fbcd1a849e6eba90d14b74dd.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(z_{n,t}^{g})" + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "inline_equation", + "content": "\\sigma_{\\theta ,n}^{g}" + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "text", + "content": " represent the model predicted noise mean and variance at the " + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 594, + 554, + 617 + ], + "type": "text", + "content": " -th denoising step." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 677 + ], + "type": "text", + "content": "This method achieves the gradual recovery of data by optimizing the conditional probability of each time step by maximum likelihood. However, Markov chain-based [35] methods limit computational efficiency in high-dimensional spaces and are difficult to extend to discrete data." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To further optimize the denoising process, this paper uses SE [50] as the optimization target. It is a generalized score matching objective that aims to directly learn the" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20531" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": "probability density ratio between discrete states. The SE can not only stabilize the diffusion denoising process but also improve the sampling quality through the global information of data distribution. In general form, for any state pair " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "(z_{n,t}^{g}, z_{n,0}^{g})" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": ", define the model's score ratio " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_{n,t}^{g})" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " which represents the relative probability of transferring from " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}" + }, + { + "bbox": [ + 55, + 72, + 296, + 157 + ], + "type": "text", + "content": ". SE is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 162, + 294, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 162, + 294, + 236 + ], + "spans": [ + { + "bbox": [ + 58, + 162, + 294, + 236 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} s e = \\sum_ {y \\in z _ {n, 0: t - 1} ^ {g}} \\omega_ {z _ {n, t} ^ {g}} ^ {g} \\left(s _ {\\theta} (z _ {n, t} ^ {g}) - \\frac {p _ {d a t a} (y)}{p _ {d a t a} (z _ {n , t} ^ {g})} \\log s _ {\\theta} (z _ {n, t} ^ {g}) \\right. \\\\ + K \\left(\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)}\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "aca92db9cb24941331738673bfa85f6f2cc10495deda475e20393d4f2efe0f14.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\omega_{z_{n,t}}^{g}" + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": " is the weight of the loss term, which is used to balance the loss of different states. " + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "inline_equation", + "content": "K(a) = a(\\log a - 1)" + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": " is a normalization term that ensures the loss is non-negative. " + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}" + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": " represents the actual score ratio. " + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "inline_equation", + "content": "p_{data}(y)" + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "inline_equation", + "content": "p_{data}(z_{n,t}^{g})" + }, + { + "bbox": [ + 55, + 242, + 296, + 335 + ], + "type": "text", + "content": " are the actual data distributions of the former noisy state and the current noisy state. The actual score ratio calculation relationship is shown in Theorem 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 339, + 296, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 380 + ], + "type": "text", + "content": "Theorem 1. According to Bayes' theorem and the Gaussian distribution density formula, the following calculation relationship of " + }, + { + "bbox": [ + 55, + 339, + 296, + 380 + ], + "type": "inline_equation", + "content": "\\frac{p_{data}(y)}{p_{data}(z_{n,t}^{g})}" + }, + { + "bbox": [ + 55, + 339, + 296, + 380 + ], + "type": "text", + "content": " is obtained:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 384, + 294, + 427 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 384, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 63, + 384, + 294, + 427 + ], + "type": "interline_equation", + "content": "\\frac {p _ {\\text {d a t a}} (y)}{p _ {\\text {d a t a}} \\left(z _ {n , t} ^ {g}\\right)} = \\exp \\left(\\frac {\\| z _ {n , t} ^ {g} \\| ^ {2}}{2} - \\frac {\\| z _ {n , t} ^ {g} - \\sqrt {\\bar {\\alpha} _ {t} ^ {g}} z _ {n , 0} ^ {g} \\| ^ {2}}{2 \\left(1 - \\bar {\\alpha} _ {t} ^ {g}\\right)}\\right). \\tag {5}", + "image_path": "34d8dbb2fb02eb452a37b09724f3af8aa25ab482a6e2b561cedfc9d3b42e6c8b.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 432, + 221, + 444 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 432, + 221, + 444 + ], + "spans": [ + { + "bbox": [ + 67, + 432, + 221, + 444 + ], + "type": "text", + "content": "The proof is provided in Appendix A." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 491 + ], + "type": "text", + "content": "Based on the SE [50], the model predicted score ratio indicates how the model adjusts the probability of the current state to tend to the original data distribution during the denoising process. The definition is as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 131, + 496, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 496, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 131, + 496, + 295, + 525 + ], + "type": "interline_equation", + "content": "s _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {p _ {\\theta} \\left(z _ {n , 0} ^ {g}\\right)}{p _ {\\theta} \\left(z _ {n , t} ^ {g}\\right)}, \\tag {6}", + "image_path": "5f71e9a91a9de63343801d925b8b26f1f3714d4dd4bb564a688c304bf8067c00.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 296, + 601 + ], + "type": "text", + "content": "where the denominator represents the probability of the current noise state and the numerator represents the original state probability estimated by the model. According to Theorem 2, the model uses softmax for normalization ensuring numerical stability and enabling gradient optimization when predicting the score ratio." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": "Theorem 2. Given the denoising process modelled by a score-based probability ratio function " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "s_{\\theta}(z_{n,t}^{g})" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": ", defined as Eq. (6), this paper defines a learnable approximation using a parameterized score function " + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 605, + 296, + 665 + ], + "type": "text", + "content": ", such that the probability ratio can be estimated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 89, + 669, + 295, + 700 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 669, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 89, + 669, + 295, + 700 + ], + "type": "interline_equation", + "content": "s _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) = \\frac {\\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , z _ {n , 0} ^ {g}\\right)\\right)}{\\sum_ {y \\in z _ {n , 0 : t - 1} ^ {g}} \\exp \\left(f _ {\\theta} \\left(z _ {n , t} ^ {g} , y\\right)\\right)}, \\tag {7}", + "image_path": "56e9047e2b4954ad069d2ba8acac3b101dbda53b57b9b133e2ea8e5b9c047fcf.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 67, + 704, + 222, + 717 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 222, + 717 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 222, + 717 + ], + "type": "text", + "content": "The proof is provided in Appendix A." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 379, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 379, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 379, + 83 + ], + "type": "text", + "content": "3.2. Selection" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "text", + "content": "The selection process comprises two key steps: scan switch and selection. The scan switch mechanism captures temporal relationships between adjacent image patches (or text embeddings) by generating latent space representations with " + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "text", + "content": " different sequential relationships, such as four image patch sequences and two text embedding sequences illustrated in Fig. 2fg. The mechanism creates " + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "text", + "content": " temporal sequences " + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "inline_equation", + "content": "S = \\{\\langle z_{1,t}^{g},z_{2,t}^{g},\\dots ,z_{i,t}^{g}\\rangle \\}_{k}" + }, + { + "bbox": [ + 313, + 84, + 553, + 180 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": "The selection step then analyzes these different sequential relationships at the current denoising step " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " to determine which information should be focused on or ignored, thereby guiding the model's denoising direction in each diffusion step. The selection step chooses " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " items " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " from each sequence in " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " according to the following Theorem 3. So, the selection step obtain " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " selection sequences with different lengths, i.e., " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "S' = \\{\\langle z_{j_1,t}^g, z_{j_2,t}^g, \\ldots, z_{j,t}^g \\rangle\\}_k" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "inline_equation", + "content": "S' \\in S" + }, + { + "bbox": [ + 313, + 180, + 554, + 277 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "text", + "content": "Theorem 3. To achieve the optimal score entropy [50] which is demonstrated on Eq. (4), the selection step choose " + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "text", + "content": " items where each " + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "text", + "content": " satisfies " + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "inline_equation", + "content": "se = 0" + }, + { + "bbox": [ + 313, + 281, + 555, + 318 + ], + "type": "text", + "content": ", i.e.," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 386, + 325, + 553, + 352 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 325, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 386, + 325, + 553, + 352 + ], + "type": "interline_equation", + "content": "s _ {\\theta} \\left(z _ {n, t} ^ {g}\\right) \\approx \\frac {p _ {d a t a} (y)}{p _ {d a t a} \\left(z _ {n , t} ^ {g}\\right)} \\tag {8}", + "image_path": "5e17e8e223818561223f19e57a68a7bbe9986620090f058c3ccd15443b057968.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 326, + 357, + 480, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 357, + 480, + 369 + ], + "spans": [ + { + "bbox": [ + 326, + 357, + 480, + 369 + ], + "type": "text", + "content": "The proof is provided in Appendix A." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 373, + 394, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 394, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 394, + 385 + ], + "type": "text", + "content": "4. Architecture" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "text", + "content": "The neural network architecture consists of two primary components: a VAE noisy latent encoder [42] and a multistep selection diffusion decoder, as illustrated in Fig. 2ab. The encoder first processes image data " + }, + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "inline_equation", + "content": "X_{img}" + }, + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "text", + "content": " through patchify [19] operations and processes text data " + }, + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "inline_equation", + "content": "X_{txt}" + }, + { + "bbox": [ + 313, + 388, + 554, + 484 + ], + "type": "text", + "content": " through tokenization based on SentencePiece with Unigram BPE [45] and embedding operations, then uniformly maps them to the latent space before applying forward noise." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 663 + ], + "type": "text", + "content": "The decoder, based on the multi-step selection diffusion model, leverages Mamba to achieve unified learning objectives while enhancing computational efficiency for processing long sequence data. It employs the SE [50] as the unified objective for both image and text modalities during the diffusion process. During selection, the model captures sequential relationships across different temporal dimensions using various scan switches. These relationships are then efficiently processed through the selection state-space structure in the Mamba Block determining which information to focus on or ignore according to Eq. (8), thereby guiding subsequent diffusion denoising steps (as shown in Fig. 2h). Finally, the reconstructed image patches and text embeddings are transformed back into their original data formats through a VAE noisy latent decoder [42]." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 452, + 677 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 452, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 452, + 677 + ], + "type": "text", + "content": "4.1. The noisy latent encoder" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "The noisy latent encoder first processes input image " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "X_{img}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " through patchify and processes text " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "X_{txt}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " through tokenization and embedding operations to obtain the patch sequence" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20532" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "G(X_{img} / X_{txt}) = \\langle g_1, g_2, \\dots, g_i \\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "g_n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "-th image patch or text embedding, respectively. The encoder VAE [42] generates Gaussian distribution parameters (mean " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ") for these patches, with a similar process applied to text embeddings, i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "VAE(G) = (\\mu, \\sigma)" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ". For each image patch or text embedding " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "g_n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ", its noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "z_n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " is a sample " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "s_n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " from the distribution " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mu, \\sigma)" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " with the addition noise " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\epsilon_n \\sim \\mathcal{N}(0, 1)" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "z_n = s_n + \\epsilon_n" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": ". Finally, the image " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "X_{img}" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "X_{txt}" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " are transformed into the noise sequence " + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "inline_equation", + "content": "\\langle z_1, \\dots, z_i \\rangle" + }, + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": " through the above process." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 266 + ], + "type": "text", + "content": "Moreover, three types of learnable padding tokens, time, category, and pad, are inserted into these noise sequences, as illustrated in Fig. 2de. The time token encodes the current diffusion step, the class token is used to learn the data category, and the pad token represents the start or end position for splitting these noise sequences." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 270, + 274, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 274, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 274, + 282 + ], + "type": "text", + "content": "4.2. The multi-step selection diffusion decoder" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": "The decoder aims at progressively recovering the image " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "X_{img}" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": " or text " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "X_{txt}" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": " from noise sequences through two main modules: 1) the multi-step selection diffusion Mamba and 2) the VAE noisy latent decoder. 1) The Mamba is used to recover the patch sequence " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\langle g_1, \\dots, g_i \\rangle" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": " from the noise sequence " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\langle z_1, \\dots, z_i \\rangle" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": ". 2) The VAE noisy latent decoder assembles patches and generates the image " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\hat{X}_{img}" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": " or text " + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "inline_equation", + "content": "\\hat{X}_{txt}" + }, + { + "bbox": [ + 55, + 284, + 296, + 369 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 372, + 243, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 243, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 243, + 383 + ], + "type": "text", + "content": "4.2.1. Multi-step selection diffusion Mamba" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 418 + ], + "type": "text", + "content": "The module leverages two components, image/text scan switch and Mamba Block, to implement each denoising step in the multi-step selection diffusion model (Sec. 3)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 479 + ], + "type": "text", + "content": "The image/text scan switch component establishes sequences with different directions to capture different temporal relationships between patches. Following Dim [73], we implement four distinct scan switches for images (as shown in Fig. 2f) and two for text (as shown in Fig. 2g)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "content": "The Mamba block is used to select patches from these different scan switch sequences and denoise the input noise " + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "content": ". The block adopts the state space architecture from Mamba-2 [26]. According to Sec. 3.2, it is " + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "inline_equation", + "content": "s_{\\theta}" + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "inline_equation", + "content": "\\theta = \\{H_{n,t}^{g}, A, B, C, D, \\Delta\\}" + }, + { + "bbox": [ + 55, + 480, + 296, + 588 + ], + "type": "text", + "content": " represent the state space in the block. The block comprises six key components: 1) linear input and output projection layers, 2) convolution kernel layer, 3) nonlinear activation layer, 4) state space model (SSM), 5) skip connection layer, and 6) normalization layer." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": "1) The linear input projection layer reduces the dimensionality of the latent space noise vector while simultaneously applying initial state matrices " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": " to the linear projection of input data " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": ". Additionally, the linear output projection layer represents the denoising step, which transforms the selection noise " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "z_{n,t}^{g}" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "inline_equation", + "content": "z_{n,t - \\Delta t}^{g}" + }, + { + "bbox": [ + 55, + 589, + 296, + 673 + ], + "type": "text", + "content": " and outputs it to the next Mamba block according to the following equation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 693, + 296, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 693, + 296, + 717 + ], + "spans": [ + { + "bbox": [ + 61, + 693, + 296, + 717 + ], + "type": "interline_equation", + "content": "z _ {n, t - \\Delta t} ^ {g} = z _ {n, t} ^ {g} - \\frac {\\Delta t}{2} \\left[ f _ {\\theta} \\left(z _ {n, t} ^ {g}, t\\right) + f _ {\\theta} \\left(z _ {n, t - \\Delta t} ^ {g}, t - \\Delta t\\right) \\right] (9)", + "image_path": "d2ca3085efc2e7d8827a7b02b75f55a8ecfe06538e2f1c4b215a5c159ee464cf.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "where the equation adopts the second-order numerical method of DPM-Solver [51] to improve sampling accuracy. Details are provided in Appendix B." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 108, + 554, + 228 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 168 + ], + "type": "text", + "content": "2) The convolution kernel layer implements parallel scan switches, routing the initial linear projection of the input and the state matrix's linear projection through the SSM, as shown in Fig. 2i. The sweep down and sweep up [26] enable parallel computation between Eqs. (10) to (13)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 168, + 545, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 168, + 545, + 180 + ], + "spans": [ + { + "bbox": [ + 326, + 168, + 545, + 180 + ], + "type": "text", + "content": "3) The nonlinear layer enhances model generalization." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "content": "4) The SSM lets the Mamba block " + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "inline_equation", + "content": "s_{\\theta}" + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "content": " approximate the actual score ratio based on Theorem 3. To implement the target, SSM updates the state space " + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 180, + 553, + 228 + ], + "type": "text", + "content": " by the following equations (based on Theorem 3 and details in Appendix A)." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 380, + 239, + 553, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 239, + 553, + 254 + ], + "spans": [ + { + "bbox": [ + 380, + 239, + 553, + 254 + ], + "type": "interline_equation", + "content": "H _ {n, t} ^ {g} = \\bar {A} H _ {n, t - 1} ^ {g} + \\bar {B} z _ {n, t} ^ {g} \\tag {10}", + "image_path": "5195509fa2c85e7604a82d4acbe399374eca88fbb283aba32bdae942030d47b0.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 382, + 269, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 269, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 382, + 269, + 553, + 285 + ], + "type": "interline_equation", + "content": "z _ {n - 1, t} ^ {g} = C H _ {n, t} ^ {g} + D z _ {n, t} ^ {g} \\tag {11}", + "image_path": "ca322df9e6c415af46706e7501876b4dd2059a15843191c85c162ff3183a6fb7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 402, + 300, + 553, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 402, + 300, + 553, + 313 + ], + "spans": [ + { + "bbox": [ + 402, + 300, + 553, + 313 + ], + "type": "interline_equation", + "content": "\\bar {A} = \\exp (\\Delta A) \\tag {12}", + "image_path": "758a1c01f00351e5f01f0c8bc7ed0858f2c03994de4639ee36e5148a1c50c43d.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 357, + 330, + 553, + 343 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 330, + 553, + 343 + ], + "spans": [ + { + "bbox": [ + 357, + 330, + 553, + 343 + ], + "type": "interline_equation", + "content": "\\bar {B} = (\\Delta A) ^ {- 1} \\cdot \\left(\\exp (\\Delta A) - I\\right) \\cdot \\Delta B \\tag {13}", + "image_path": "ba4b60f9276e3e11e6256e95f4f0face0e7b1b75ed9fcb535bdb6ca7e5ca9093.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "H_{n,t}^{g}" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " represents the hidden state representation, " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " control the evolution of hidden states and latent space noise vector inputs, respectively, " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " governs the hidden state representation of the target output and " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " manages the nonlinear skip connection for latent space noise vector inputs. " + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 348, + 554, + 420 + ], + "type": "text", + "content": " denotes the learnable time parameter." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 421, + 553, + 480 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 313, + 421, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 445 + ], + "type": "text", + "content": "5) The skip connection layer facilitates input feature reuse and mitigates model degradation." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 326, + 445, + 541, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 445, + 541, + 456 + ], + "spans": [ + { + "bbox": [ + 326, + 445, + 541, + 456 + ], + "type": "text", + "content": "6) The Normalization layer ensures training stability." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 456, + 553, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 553, + 480 + ], + "type": "text", + "content": "According to Eq. (8) in Theorem 3 and Eq. (4), the goal of training the Mamba block is:" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 364, + 491, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 491, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 364, + 491, + 553, + 506 + ], + "type": "interline_equation", + "content": "L _ {s e} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim p _ {0}, z _ {n} ^ {g} \\sim p (\\cdot | z _ {n, 0} ^ {g})} s e = 0 \\tag {14}", + "image_path": "cda40c2ac5759be1ad1096eff187338698491482c5b3970a4bba6cc8fd4ebffc.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 511, + 447, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 447, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 447, + 523 + ], + "type": "text", + "content": "4.2.2. The noisy latent decoder" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "text", + "content": "After applying the diffusion-based denoising process, the recovered latent variable " + }, + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}" + }, + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "text", + "content": " is passed to the VAE decoder [42] as illustrated in Fig. 2c. For image reconstruction, the decoder applies an " + }, + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\ell_2" + }, + { + "bbox": [ + 313, + 527, + 554, + 575 + ], + "type": "text", + "content": " loss:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 350, + 583, + 553, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 583, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 350, + 583, + 553, + 599 + ], + "type": "interline_equation", + "content": "L _ {r e c} ^ {i m g} = \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\| X _ {i m g} - \\hat {X} _ {i m g} \\| ^ {2}. \\tag {15}", + "image_path": "3175c52c16f295a5e18d28f5efc9bb49fb09844b05845a894a6f4c6eaeead6a7.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 609, + 553, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 553, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 609, + 553, + 631 + ], + "type": "inline_equation", + "content": "q_{\\phi}(z|X)" + }, + { + "bbox": [ + 313, + 609, + 553, + 631 + ], + "type": "text", + "content": " represents the posterior distribution of the VAE encoder." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 326, + 632, + 547, + 644 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 632, + 547, + 644 + ], + "spans": [ + { + "bbox": [ + 326, + 632, + 547, + 644 + ], + "type": "text", + "content": "For text, the decoder minimizes the cross-entropy loss:" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 314, + 653, + 556, + 687 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 653, + 556, + 687 + ], + "spans": [ + { + "bbox": [ + 314, + 653, + 556, + 687 + ], + "type": "interline_equation", + "content": "L _ {r e c} ^ {t x t} = - \\mathbb {E} _ {z _ {n, 0} ^ {g} \\sim q _ {\\phi} (z | X)} \\sum_ {t} p \\left(X _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right) \\log p _ {\\psi} \\left(\\hat {X} _ {t x t} ^ {(t)} \\mid z _ {n, 0} ^ {g}\\right). \\tag {16}", + "image_path": "4cc2404310a600e4757c2d073f26a70d61a1ca46bf04f0dbb8fb8c21420ca11f.jpg" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "p(X_{txt}^{(t)}|z_{n,0}^{g})" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": " represents the probability distribution of real text data under the condition of latent variable " + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "inline_equation", + "content": "z_{n,0}^{g}" + }, + { + "bbox": [ + 313, + 688, + 553, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20533" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 294, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 294, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 294, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 294, + 189 + ], + "type": "image", + "image_path": "953d44ded220231e3b4ec9deadbc4fb7cc804a29d77460b4fb29182ebc1593d1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 245, + 294, + 425 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 242 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 242 + ], + "type": "text", + "content": "Figure 3. VQA, QA and Multi-Modal generation test from MDM. The results of VQA are part of VQAv2 [25]. The QA results are part of PIQA [7] and MMLU [32]. The Multi-Modal generation results are tested with ground-truth data." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 245, + 294, + 425 + ], + "lines": [ + { + "bbox": [ + 58, + 245, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 58, + 245, + 294, + 425 + ], + "type": "image", + "image_path": "d18d89bf12c3d61711c87fa4e2fa0800653c74b98ff3378c3001a1654b9e784a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 433, + 295, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 433, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 295, + 466 + ], + "type": "text", + "content": "Figure 4. Comparison between each model on generating captioning and image results on COCO dataset. Unlike other models, MDM generates both image and caption data simultaneously." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "type": "text", + "content": "And " + }, + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "type": "inline_equation", + "content": "p_{\\psi}(\\hat{X}_{txt}^{(t)}|z_{n,0}^g)" + }, + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "type": "text", + "content": " represents the probability distribution of the text token generated by the VAE decoder under the condition of the latent variable " + }, + { + "bbox": [ + 55, + 471, + 295, + 509 + ], + "type": "inline_equation", + "content": "z_{n,0}^g" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 510, + 285, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 510, + 285, + 521 + ], + "spans": [ + { + "bbox": [ + 67, + 510, + 285, + 521 + ], + "type": "text", + "content": "Besides, a KL divergence regularizes the latent space:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 111, + 531, + 295, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 531, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 111, + 531, + 295, + 544 + ], + "type": "interline_equation", + "content": "L _ {K L} = D _ {K L} \\left(q _ {\\phi} (z | X) \\| p (z)\\right). \\tag {17}", + "image_path": "d720e67dc094d798cf3ed24b5a02ab6a01b6d3a5a95e512a1916d4f11713903a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "inline_equation", + "content": "p(z)" + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "content": " represents the prior distribution of the latent variable by VAE, which is assumed to be a standard Gaussian distribution " + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 554, + 295, + 601 + ], + "type": "text", + "content": " to regularize the latent variable space and enable it to have smooth generation capabilities." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 602, + 295, + 625 + ], + "type": "text", + "content": "The final optimization objective integrates VAE reconstruction, KL divergence and SE:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 92, + 635, + 295, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 635, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 92, + 635, + 295, + 649 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {r e c} ^ {i m g} + L _ {r e c} ^ {t x t} + \\beta L _ {K L} + \\lambda L _ {s e}. \\tag {18}", + "image_path": "0f22c7328e25772b9af5491c33734b2e2abaa7c44da4affaf60499eb498fbaff.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 137, + 670 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 673, + 171, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 171, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 171, + 687 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Model configuration. Our model applies a VAE [42] as the noisy latent encoder and decoder. Moreover, it integrates" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "the DiM selection state space [73] in each Mamba block as the diffusion decoder. The resulting model contains 7 billion parameters, with 49 Mamba blocks in the multi-step selection diffusion decoder, each having a dimension of 2048 (Details of parameter settings listed in Appendix C)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 133, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 205 + ], + "type": "text", + "content": "Before the training MDM process, we trained a tokenization model based on SentencePiece (Unigram BPE) [45]. The tokenization model can help the model construct a stable text latent variable representation, thereby optimizing the forward diffusion and reverse denoising process. See Appendix D for detailed experimental settings." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 206, + 554, + 277 + ], + "type": "text", + "content": "In the training process, we import the DDPM scheduler [35] and DPM-Solver [51] to improve the sampling efficiency in the diffusion model. We then use the AdamW optimizer without weight decay, maintaining a constant learning rate of 0.0001. Meanwhile, we keep an EMA of the model weights with a coefficient of 0.9999." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 279, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 554, + 339 + ], + "type": "text", + "content": "Baseline and dataset. Our evaluation encompasses four tasks: image generation with classifier-free guidance [34] (CFG), text-to-image, image-to-text, and text-to-text generation. For the baseline model training, we train MDM on ImageNet [13], JourneyDB [68] and UltraChat [16]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "content": "For the image generation and the text-to-image task at " + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 313, + 341, + 554, + 508 + ], + "type": "text", + "content": " resolution, we compare the MDM baseline model against established baselines across three categories: diffusion models (Imagen [64], ADM [15], CDM [36], LDM [63], DiT-XL/2 [57], SDXL [58], and SD-3 [21]), autoregressive models (VQGAN [20] and ViT-VQGAN [85]), and end-to-end multi-modal models (NExT-GPT [78], Chameleon [71], LlamaGen [69], Transfusion [90], MonoFormer [89], Dual-DiT [47], JanusFlow [52] and ShowO [79]). For the image generation task, we evaluate performance on ImageNet [13] using four metrics: Frechet Inception Distance (FID), Inception Score (IS), and Precision/Recall. For the text-to-image task, we evaluate performance on COCO [40] using FID and Gen Eval [24]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 554, + 675 + ], + "type": "text", + "content": "For the image-to-text task (image captioning and vision question answering, VQA) and text-to-text task, we employ MDM baseline model and MDM instruction model by visual instruction tuning [48] on multiple datasets: COCO [40], GQA [38], OCR-VQA [55], TextVQA [67], and VisualGenome [44]. We evaluate the model against two groups of baselines: traditional models and end-to-end multi-modal models. Performance evaluation of image captioning is conducted on Flickr 30K [84] and COCO [40] datasets using the Consensus-based Image Description Evaluation (CIDEr) metric. And performance evaluation of VQA is conducted on VQAv2 [25], VizWiz [28], and OKVQA [53] using answer accuracy rate as the evaluation metric." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "For the text-to-text task, we evaluate the model on text comprehension and reasoning tasks using HellaSwag [87], OpenBookQA [54], Wino-Grande [65], ARCEasy, ARC" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20534" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 300 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 300 + ], + "type": "table", + "html": "
ModelArcParamsImage Generation with CFGText-to-Image Generation
FID ↓IS ↑Pre ↑Re ↑FID ↓Gen Eval ↑
Imagen [64]Diff7.3B----7.27-
ADM [15]Diff554M10.94101.00.690.63--
CDM [36]Diff-4.88158.7----
LDM [63]Diff400M3.60147.60.870.68-0.43
DiT-XL/2 [57]Diff675M2.27278.20.830.57--
SDXL [58]Diff3.4B----4.400.55
SD-3 [21]Diff12.7B-----0.68
VQGAN [20]AR227M18.6580.40.780.26--
ViT-VQGAN [85]AR1.7B4.17175.1----
NExT-GPT [78]AR7B----10.07-
Chameleon [71]AR7B----26.740.39
LlamaGen [69]AR3.1B2.81311.50.840.544.19-
Transfusion [90]AR+Diff7.3B----6.780.63
MonoFormer [89]AR+Diff1.1B2.57272.60.840.56--
Dual-DiT [47]Diff2B----9.400.65
JanusFlow [52]AR+Diff1.3B-----0.70
Show-O [79]AR+Diff1.3B----9.240.68
MDMDiff7B2.49281.40.860.595.910.68
", + "image_path": "b14bce0ea13faa1f24f79fe8c13237667ea93db9215c6b87490ce7345a5cf576.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 335, + 556, + 580 + ], + "blocks": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": "Table 1. Performance on ImageNet and COCO " + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 55, + 308, + 555, + 331 + ], + "type": "text", + "content": " . FID,IS,Pre,and Re stands for Frechet Inception Distance,Inception Score, Precision, and Recall, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 335, + 556, + 580 + ], + "lines": [ + { + "bbox": [ + 56, + 335, + 556, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 335, + 556, + 580 + ], + "type": "table", + "html": "
ModelICVQAText Comprehension and ReasoningMath and World
FlickrCOCOVQAv2VizWizOKHSOBQAWGARCEARCCBoolQPIQAGSM8kMATHMMLU
Llama-2 [74] (7B)-----77.258.678.575.245.977.478.814.62.545.3
Mistral [39] (7B)-----81.3-75.380.055.584.783.052.113.160.1
Flamingo [2] (80B)75.1113.867.6------------
Gemini Pro [72]82.299.871.2--84.7------86.532.671.8
GPT4V [8]55.378.577.2--95.3------92.052.986.4
InstructBLIP [48] (7B)82.4102.2-33.433.9----------
mPLUG-Owl [83] (7B)80.3119.3-39.0-----------
TinyLlama [88] (1.1B)-----59.236.059.155.330.157.873.3---
Pythia [6] (12B)-----52.033.257.454.028.563.370.9---
DREAMLLM [17](7B)-115.456.645.844.3----------
Emu [70](7B)-117.740.035.434.7----------
Chameleon [71](34B)74.7120.266.0--74.251.070.476.146.581.479.641.611.552.1
NExT-GPT [78](7B)84.5124.966.748.452.1----------
Transfusion [90](7B)-33.7-------------
MonoFormer [89](1.1B)-----50.637.256.948.231.562.371.2---
Dual-DiT [47](2B)-56.260.129.925.3----------
JanusFlow [52](1.3B)--79.8------------
Show-O [79](1.3B)67.6-74.7------------
MDM (7B)62.4109.660.339.847.170.641.568.855.146.265.779.940.512.154.4
InstructMDM (7B)75.2122.166.746.351.674.848.374.965.447.171.583.746.013.159.2
", + "image_path": "1edc2bbedcc758710c393608f62476f248d17b216bfbcc796f3980b9fd8c42d1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 589, + 555, + 611 + ], + "lines": [ + { + "bbox": [ + 55, + 589, + 555, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 555, + 611 + ], + "type": "text", + "content": "Table 2. Performance on image-to-text and text-to-text tasks. The evaluation of image captioning (IC) and VQA is CIDEr and answer accuracy % (Flickr is evaluated on 30K and OK represents OKVQA)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 619, + 294, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 619, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 619, + 294, + 667 + ], + "type": "text", + "content": "Challenge [10], BoolQ [9], and PIQA [7]. We also evaluate the model on math and world knowledge tasks using GSM8K [11], MATH [33], and MMLU [32]. The evaluation metrics for all the tasks are accuracy rates." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 672, + 179, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 179, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 179, + 685 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Image Generation. In the image generation task on ImageNet, MDM achieves top-three rankings across all eval" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 619, + 555, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 704 + ], + "type": "text", + "content": "uation metrics: second in FID, IS, and Precision, and third in Recall when compared against one-modal diffusion models and end-to-end multi-modal models (see Tab. 1). MDM demonstrates superior overall performance, notably surpassing other end-to-end multi-modal models in three of the four metrics. In the text-to-image task, we tested the model on the COCO dataset to generate both image and" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20535" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 70, + 304, + 176 + ], + "blocks": [ + { + "bbox": [ + 107, + 70, + 304, + 176 + ], + "lines": [ + { + "bbox": [ + 107, + 70, + 304, + 176 + ], + "spans": [ + { + "bbox": [ + 107, + 70, + 304, + 176 + ], + "type": "image", + "image_path": "de077a502ef4c2b76e1729731b18c36fbbfd4c0b9fe749887897a94cf1ad89c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 209 + ], + "type": "text", + "content": "Figure 5. Comparison between Mamba Baseline, MonoFormer, and ours MDM on inference speed test. The left shows the inference speed of the model FPS at different resolutions. The right shows the inference speed of the model latency at different token lengths." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 305, + 70, + 503, + 176 + ], + "blocks": [ + { + "bbox": [ + 305, + 70, + 503, + 176 + ], + "lines": [ + { + "bbox": [ + 305, + 70, + 503, + 176 + ], + "spans": [ + { + "bbox": [ + 305, + 70, + 503, + 176 + ], + "type": "image", + "image_path": "2a8790f8903f155935ff66a3aae2ea185867caf22ff191eac4dba5c23dbcb7b2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 217, + 296, + 275 + ], + "blocks": [ + { + "bbox": [ + 56, + 217, + 296, + 275 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 275 + ], + "type": "table", + "html": "
ModelImage/Text \nScan SwitchFPS w log \nscale↑FID↓
Model w Mamba①②③④/①②1.3572.49
Model w Mamba①②/①1.4053.96
Model w Transformer-1.9146.72
", + "image_path": "d03eeac77510c16f6fe0d7736c7a175436383b3257c0048d9640cf7d9628ed57.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 283, + 284, + 295 + ], + "lines": [ + { + "bbox": [ + 66, + 283, + 284, + 295 + ], + "spans": [ + { + "bbox": [ + 66, + 283, + 284, + 295 + ], + "type": "text", + "content": "Table 3. Ablation on ImageNet " + }, + { + "bbox": [ + 66, + 283, + 284, + 295 + ], + "type": "inline_equation", + "content": "{256} \\times {256}" + }, + { + "bbox": [ + 66, + 283, + 284, + 295 + ], + "type": "text", + "content": " image generation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 295, + 357 + ], + "type": "text", + "content": "caption data. For the image generation results, we evaluated the FID and Gen Eval performance indicators of the model-generated images. MDM still achieved the top three performance levels and achieved SOTA on Gen Eval." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 357, + 295, + 596 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 295, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 295, + 596 + ], + "type": "text", + "content": "Text Generation. In the image-to-text task image captioning, according to the settings on image generation on the COCO dataset, we tested the caption data of the model based on the model outputting both text and image data using the CIDEr indicator. The results showed that MDM ranked second among all models, as shown in Tab. 2. While in task VQA, MDM achieves competitive performance, surpassing several traditional models including InstructBLIP, mPLUG-Owl, DREAMLLM, and Emu, although it still trails behind top-performing models in the field as shown in Tab. 2. In the text-to-text generation task, as shown in Tab. 2, MDM and the other end-to-end multi-modal models perform worse than well-known traditional models. This discrepancy may be attributed to the fact that these end-to-end models have some deviations in multimodal fusion and learning because they abandon multiple language encoders, visual encoders, and multimodal fusion encoders. However, when compared with the other two end-to-end models, MDM excels, outperforming MonoFormer and surpassing Chameleon on seven out of ten datasets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 600, + 128, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 128, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 128, + 611 + ], + "type": "text", + "content": "5.3. Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 615, + 178, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 178, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 178, + 627 + ], + "type": "text", + "content": "5.3.1. Performance Analysis" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "As demonstrated in Fig. 3, MDM shows the ability to generate image and text simultaneously in multiple rounds of dialogue and perform well in QA&VQA. Some results even exceed those of GPT-4V, particularly evident in the second and third rows of Fig. 4 which is a hybrid output process for the MDM model. Due to this, we set the model to generate corresponding images for the description text while simul" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 220, + 473, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 473, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 473, + 231 + ], + "type": "text", + "content": "taneously generating image captioning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 339 + ], + "type": "text", + "content": "This enhanced performance stems from MDM's multi-step selection diffusion decoder, which leverages Mamba's integrated selection and denoising capabilities to maintain focused attention on both textual and visual details. Validating our complexity analysis in Appendix E, MDM demonstrates superior efficiency compared to end-to-end Transformer models when processing long sequences, as shown in Fig. 5, particularly outperforming other end-to-end multimodal models for sequences exceeding 1280 tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 342, + 383, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 383, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 383, + 351 + ], + "type": "text", + "content": "5.3.2. Ablations" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 354, + 556, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 486 + ], + "type": "text", + "content": "Our ablation studies examine the impact of both the selection process and Mamba block components. Reducing the number of image/text scan switch sequences from 6 ('①②③④/①②') to 3 ('①②/①'), as shown in Tab. 3, improves inference speed but degrades image quality, as fewer scan switch sequences limit the model's ability to capture accurate information in complex sequences. Additionally, replacing the Mamba block with the Transformer further deteriorates output image quality, suggesting Mamba's temporal network architecture is better suited for representing diffusion relationships during the denoising process." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 504, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 704 + ], + "type": "text", + "content": "This paper introduces MDM (Multi-Modal Diffusion Mamba), a novel end-to-end architecture that significantly enhances multi-modal processing through two key innovations: a unified diffusion objective and an efficient selection mechanism leveraging Mamba's state-space structure. By integrating variational autoencoder with multi-step selection diffusion, MDM achieves SOTA overall performance in image generation and demonstrates remarkable versatility across various tasks, including image-to-text, text-to-text and text-image-to-text-image. Our comprehensive experiments illustrate that MDM consistently surpasses traditional end-to-end multi-modal models, particularly in processing high-resolution images and long-sequence text, while maintaining computational efficiency. The model's ability to unify different modalities under a single objective, coupled with its superior management of temporal relationships in the diffusion process, establishes a promising direction for future multi-modal architecture." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "20536" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Joshua Ainslie, James Lee-Thorp, Michiel de Jong, Yury Zemlyanskiy, Federico Lebrón, and Sumit Sanghai. Gqa: Training generalized multi-query transformer models from multi-head checkpoints. arXiv preprint arXiv:2305.13245, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 212 + ], + "type": "text", + "content": "[2] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 213, + 295, + 257 + ], + "type": "text", + "content": "[3] Silas Alberti, Niclas Dern, Laura Thesing, and Gitta Kutyniok. Sumformer: Universal approximation for efficient transformers. In Topological, Algebraic and Geometric Learning Workshops 2023, pages 72-86. PMLR, 2023. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 322 + ], + "type": "text", + "content": "[4] Hangbo Bao, Wenhui Wang, Li Dong, Qiang Liu, Owais Khan Mohammed, Kriti Aggarwal, Subhojit Som, Songhao Piao, and Furu Wei. Vlmo: Unified vision-language pre-training with mixture-of-modality-experts. Advances in Neural Information Processing Systems, 35:32897-32912, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 323, + 294, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 323, + 294, + 356 + ], + "spans": [ + { + "bbox": [ + 62, + 323, + 294, + 356 + ], + "type": "text", + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Sağnak Taşırlar. Introducing our multimodal models, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 295, + 434 + ], + "type": "text", + "content": "[6] Stella Biderman, Hailey Schoelkopf, Quentin Gregory Anthony, Herbie Bradley, Kyle O'Brien, Eric Hallahan, Mohammad Aflah Khan, Shivanshu Purohit, USVSN Sai Prashanth, Edward Raff, et al. Pythia: A suite for analyzing large language models across training and scaling. In International Conference on Machine Learning, pages 2397-2430. PMLR, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 435, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 435, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 435, + 294, + 479 + ], + "type": "text", + "content": "[7] Yonatan Bisk, Rowan Zellers, Jianfeng Gao, Yejin Choi, et al. Piqa: Reasoning about physical commonsense in natural language. In Proceedings of the AAAI conference on artificial intelligence, pages 7432-7439, 2020. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 479, + 214, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 214, + 490 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 214, + 490 + ], + "type": "text", + "content": "[8] GPTV System Card. Openai, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 294, + 535 + ], + "type": "text", + "content": "[9] Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. arXiv preprint arXiv:1905.10044, 2019. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 294, + 590 + ], + "type": "text", + "content": "[10] Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457, 2018. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 645 + ], + "type": "text", + "content": "[11] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168, 2021. 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[12] Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in Neural Information Processing Systems, 35:16344-16359, 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 554, + 95 + ], + "type": "text", + "content": "database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 98, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 554, + 129 + ], + "type": "text", + "content": "[14] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805, 2018. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 165 + ], + "type": "text", + "content": "[15] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 167, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 167, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 167, + 554, + 220 + ], + "type": "text", + "content": "[16] Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing chat language models by scaling high-quality instructional conversations. arXiv preprint arXiv:2305.14233, 2023. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 224, + 553, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 278 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 278 + ], + "type": "text", + "content": "[17] Runpei Dong, Chunrui Han, Yuang Peng, Zekun Qi, Zheng Ge, Jinrong Yang, Liang Zhao, Jianjian Sun, Hongyu Zhou, Haoran Wei, et al. Dreamllm: Synergistic multimodal comprehension and creation. arXiv preprint arXiv:2309.11499, 2023. 1, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 280, + 553, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 324 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 324 + ], + "type": "text", + "content": "[18] Wenhao Dong, Haodong Zhu, Shaohui Lin, Xiaoyan Luo, Yunhang Shen, Xuhui Liu, Juan Zhang, Guodong Guo, and Baochang Zhang. Fusion-mamba for cross-modality object detection. arXiv preprint arXiv:2404.09146, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 1, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 405 + ], + "type": "text", + "content": "[20] Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12873-12883, 2021. 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 407, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 553, + 473 + ], + "type": "text", + "content": "[21] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In *Forty-first International Conference on Machine Learning*, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "text", + "content": "[22] Zhengcong Fei, Mingyuan Fan, Changqian Yu, Debang Li, Youqiang Zhang, and Junshi Huang. Dimba: Transformermamba diffusion models. arXiv preprint arXiv:2406.01159, 2024. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 574 + ], + "type": "text", + "content": "[23] Yuying Ge, Sijie Zhao, Jinguo Zhu, Yixiao Ge, Kun Yi, Lin Song, Chen Li, Xiaohan Ding, and Ying Shan. Seed-x: Multimodal models with unified multi-granularity comprehension and generation. arXiv preprint arXiv:2404.14396, 2024. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 621 + ], + "type": "text", + "content": "[24] Dhruba Ghosh, Hannaneh Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. Advances in Neural Information Processing Systems, 36:52132-52152, 2023. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 623, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 678 + ], + "type": "text", + "content": "[25] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6904-6913, 2017. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[26] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 1, 2, 5" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20537" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[27] Ahan Gupta, Yueming Yuan, Yanqi Zhou, and Charith Mendis. Flurka: Fast fused low-rank & kernel attention. arXiv preprint arXiv:2306.15799, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 172 + ], + "type": "text", + "content": "[28] Danna Gurari, Qing Li, Abigale J Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P Bigham. Vizwiz grand challenge: Answering visual questions from blind people. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3608-3617, 2018. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[29] Insu Han, Rajesh Jayaram, Amin Karbasi, Vahab Mirrokni, David P Woodruff, and Amir Zandieh. Hyperattention: Long-context attention in near-linear time. arXiv preprint arXiv:2310.05869, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[30] Ali Hatamizadeh and Jan Kautz. Mambavision: A hybrid mamba-transformer vision backbone. arXiv preprint arXiv:2407.08083, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[31] Wanggui He, Siming Fu, Mushui Liu, Xierui Wang, Wenyi Xiao, Fangxun Shu, Yi Wang, Lei Zhang, Zhelun Yu, Haoyuan Li, et al. Mars: Mixture of auto-regressive models for fine-grained text-to-image synthesis. arXiv preprint arXiv:2407.07614, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[32] Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding. arXiv preprint arXiv:2009.03300, 2020. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[33] Dan Hendrycks, Collin Burns, Saurav Kadavath, Akul Arora, Steven Basart, Eric Tang, Dawn Song, and Jacob Steinhardt. Measuring mathematical problem solving with the math dataset. arXiv preprint arXiv:2103.03874, 2021. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 432 + ], + "type": "text", + "content": "[34] Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 466 + ], + "type": "text", + "content": "[35] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 511 + ], + "type": "text", + "content": "[36] Jonathan Ho, Chitwan Sahara, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans. Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research, 23(47):1-33, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 513, + 294, + 555 + ], + "type": "text", + "content": "[37] Vincent Tao Hu, Stefan Andreas Baumann, Ming Gui, Olga Grebenkova, Pingchuan Ma, Johannes S Fischer, and Björn Ommer. *Zigma: A dit-style zigzag mamba diffusion model.* arXiv preprint arXiv:2403.13802, 2024. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 558, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 611 + ], + "type": "text", + "content": "[38] Drew A Hudson and Christopher D Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 6700-6709, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[39] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[40] Andrej Karpathy and Li Fei-Fei. Deep visual-semantic alignments for generating image descriptions. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3128-3137, 2015. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 117 + ], + "type": "text", + "content": "[41] Wonjae Kim, Bokyung Son, and Ildoo Kim. Vilt: Vision-and-language transformer without convolution or region supervision. In International conference on machine learning, pages 5583-5594. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 139 + ], + "type": "text", + "content": "[42] Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 1, 3, 4, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 172 + ], + "type": "text", + "content": "[43] Jing Yu Koh, Daniel Fried, and Russ R Salakhutdinov. Generating images with multimodal language models. Advances in Neural Information Processing Systems, 36, 2024. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 174, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 237 + ], + "type": "text", + "content": "[44] Ranjay Krishna, Yuke Zhu, Oliver Groth, Justin Johnson, Kenji Hata, Joshua Kravitz, Stephanie Chen, Yannis Kalantidis, Li-Jia Li, David A Shamma, et al. Visual genome: Connecting language and vision using crowdsourced dense image annotations. International journal of computer vision, 123:32-73, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[45] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and tokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 337 + ], + "type": "text", + "content": "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[47] Zijie Li, Henry Li, Yichun Shi, Amir Barati Farimani, Yuval Kluger, Linjie Yang, and Peng Wang. Dual diffusion for unified image generation and understanding. arXiv preprint arXiv:2501.00289, 2024. 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 415 + ], + "type": "text", + "content": "[48] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. Advances in neural information processing systems, 36, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[49] Jiaming Liu, Mengzhen Liu, Zhenyu Wang, Lily Lee, Kaichen Zhou, Pengju An, Senqiao Yang, Renrui Zhang, Yandong Guo, and Shanghang Zhang. Robomamba: Multimodal state space model for efficient robot reasoning and manipulation. arXiv preprint arXiv:2406.04339, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[50] Aaron Lou, Chenlin Meng, and Stefano Ermon. Discrete diffusion modeling by estimating the ratios of the data distribution. In *Forty-first International Conference on Machine Learning*, 2024. 2, 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 553, + 568 + ], + "type": "text", + "content": "[51] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 571, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 634 + ], + "type": "text", + "content": "[52] Yiyang Ma, Xingchao Liu, Xiaokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Liang Zhao, et al. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. arXiv preprint arXiv:2411.07975, 2024.6.7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 690 + ], + "type": "text", + "content": "[53] Kenneth Marino, Mohammad Rastegari, Ali Farhadi, and Roozbeh Mottaghi. Ok-vqa: A visual question answering benchmark requiring external knowledge. In Proceedings of the IEEE/cvf conference on computer vision and pattern recognition, pages 3195-3204, 2019. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[54] Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20538" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 295, + 95 + ], + "type": "text", + "content": "dataset for open book question answering. arXiv preprint arXiv:1809.02789, 2018. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 129 + ], + "type": "text", + "content": "[55] Anand Mishra, Shashank Shekhar, Ajeet Kumar Singh, and Anirban Chakraborty. Ocr-vqa: Visual question answering by reading text in images. In ICDAR, 2019. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 130, + 295, + 174 + ], + "type": "text", + "content": "[56] Matteo Pagliardini, Daniele Paliotta, Martin Jaggi, and François Fleuret. Fast attention over long sequences with dynamic sparse flash attention. Advances in Neural Information Processing Systems, 36:59808-59831, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 175, + 295, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 218 + ], + "type": "text", + "content": "[57] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4195-4205, 2023. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 295, + 275 + ], + "type": "text", + "content": "[58] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 277, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 295, + 319 + ], + "type": "text", + "content": "[59] Ramya Prabhu, Ajay Nayak, Jayashree Mohan, Ramachandran Ramjee, and Ashish Panwar. vattention: Dynamic memory management for serving llms without pagedattention. arXiv preprint arXiv:2405.04437, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 295, + 365 + ], + "type": "text", + "content": "[60] Yanyuan Qiao, Zheng Yu, Longteng Guo, Sihan Chen, Zijia Zhao, Mingzhen Sun, Qi Wu, and Jing Liu. Vl-mamba: Exploring state space models for multimodal learning. arXiv preprint arXiv:2403.13600, 2024. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 366, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 366, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 366, + 295, + 399 + ], + "type": "text", + "content": "[61] Haohao Qu, Liangbo Ning, Rui An, Wenqi Fan, Tyler Derr, Xin Xu, and Qing Li. A survey of mamba. arXiv preprint arXiv:2408.01129, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 295, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 295, + 465 + ], + "type": "text", + "content": "[62] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 295, + 522 + ], + "type": "text", + "content": "[63] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 295, + 589 + ], + "type": "text", + "content": "[64] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "text", + "content": "[65] Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. Winogrande: An adversarial winograd schema challenge at scale. Communications of the ACM, 64 (9):99-106, 2021. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 657 + ], + "type": "text", + "content": "[66] Noam Shazeer. Fast transformer decoding: One write-head is all you need. arXiv preprint arXiv:1911.02150, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[67] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8317-8326, 2019. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[68] Keqiang Sun, Junting Pan, Yuying Ge, Hao Li, Haodong Duan, Xiaoshi Wu, Renrui Zhang, Aojun Zhou, Zipeng Qin, Yi Wang, et al. Journeydb: A benchmark for generative image understanding. Advances in Neural Information Processing Systems, 36, 2024. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[69] Peize Sun, Yi Jiang, Shoufa Chen, Shilong Zhang, Bingyue Peng, Ping Luo, and Zehuan Yuan. Autoregressive model beats diffusion: Llama for scalable image generation. arXiv preprint arXiv:2406.06525, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 553, + 217 + ], + "type": "text", + "content": "[70] Quan Sun, Qiying Yu, Yufeng Cui, Fan Zhang, Xiaosong Zhang, Yueze Wang, Hongcheng Gao, Jingjing Liu, Tiejun Huang, and Xinlong Wang. Generative pretraining in multimodality. arXiv preprint arXiv:2307.05222, 2023. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[71] Chameleon Team. Chameleon: Mixed-modal early-fusion foundation models. arXiv preprint arXiv:2405.09818, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 304 + ], + "type": "text", + "content": "[72] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 348 + ], + "type": "text", + "content": "[73] Yao Teng, Yue Wu, Han Shi, Xuefei Ning, Guohao Dai, Yu Wang, Zhenguo Li, and Xihui Liu. Dim: Diffusion mamba for efficient high-resolution image synthesis. arXiv preprint arXiv:2405.14224, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 403 + ], + "type": "text", + "content": "[74] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023. 1, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 448 + ], + "type": "text", + "content": "[75] Zhongwei Wan, Xin Wang, Che Liu, Samiul Alam, Yu Zheng, Jiachen Liu, Zhongnan Qu, Shen Yan, Yi Zhu, Quanlu Zhang, et al. Efficient large language models: A survey. arXiv preprint arXiv:2312.03863, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[76] Zifu Wan, Pingping Zhang, Yuhao Wang, Silong Yong, Simon Stepputtis, Katia Sycara, and Yaqi Xie. Sigma: Siamese mamba network for multi-modal semantic segmentation. arXiv preprint arXiv:2404.04256, 2024. 1, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 525 + ], + "type": "text", + "content": "[77] Xinghan Wang, Zixi Kang, and Yadong Mu. Text-controlled motion mamba: Text-instructed temporal grounding of human motion. arXiv preprint arXiv:2404.11375, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 557 + ], + "type": "text", + "content": "[78] Shengqiong Wu, Hao Fei, Leigang Qu, Wei Ji, and Tat-Seng Chua. Next-gpt: Any-to-any multimodal llm. arXiv preprint arXiv:2309.05519, 2023. 1, 3, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 623 + ], + "type": "text", + "content": "[79] Jinheng Xie, Weijia Mao, Zechen Bai, David Junhao Zhang, Weihao Wang, Kevin Qinghong Lin, Yuchao Gu, Zhijie Chen, Zhenheng Yang, and Mike Zheng Shou. Show-o: One single transformer to unify multimodal understanding and generation. arXiv preprint arXiv:2408.12528, 2024. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[80] Jing Nathan Yan, Jiatao Gu, and Alexander M Rush. Diffusion models without attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8239-8249, 2024. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[81] Xu Yang, Hanwang Zhang, Guojun Qi, and Jianfei Cai. Causal attention for vision-language tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9847-9857, 2021. 3" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20539" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 509 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[82] Zhe Yang, Wenrui Li, and Guanghui Cheng. Shmamba: Structured hyperbolic state space model for audio-visual question answering. arXiv preprint arXiv:2406.09833, 2024. 1, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[83] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[84] Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 282 + ], + "type": "text", + "content": "[85] Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. arXiv preprint arXiv:2110.04627, 2021. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 328 + ], + "type": "text", + "content": "[86] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. arXiv preprint arXiv:2205.01917, 2022. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 363 + ], + "type": "text", + "content": "[87] Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. Hellaswag: Can a machine really finish your sentence? arXiv preprint arXiv:1905.07830, 2019. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 397 + ], + "type": "text", + "content": "[88] Peiyuan Zhang, Guangtao Zeng, Tianduo Wang, and Wei Lu. Tinyllama: An open-source small language model. arXiv preprint arXiv:2401.02385, 2024. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[89] Chuyang Zhao, Yuxing Song, Wenhao Wang, Haocheng Feng, Errui Ding, Yifan Sun, Xinyan Xiao, and Jingdong Wang. Monoformer: One transformer for both diffusion and autoregression. arXiv preprint arXiv:2409.16280, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 509 + ], + "type": "text", + "content": "[90] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. arXiv preprint arXiv:2408.11039, 2024. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20540" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_content_list.json b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..35259b191944f78d35f6e53982008aff476bdb3b --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_content_list.json @@ -0,0 +1,1753 @@ +[ + { + "type": "text", + "text": "Engage for All: Making Ordinary Image Descriptions Appealing Again!", + "text_level": 1, + "bbox": [ + 135, + 130, + 859, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuyan Chen $^{1*}$ , Yifan Jiang $^{2}$ , Li Zhou $^{3}$ , Jinghan Cao $^{4}$ , Yu Guan $^{5}$ , Ming Yang $^{6}$ , Qingpei Guo $^{5,6\\dagger}$", + "bbox": [ + 109, + 178, + 890, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Cornell University, 2University of Southern California,", + "bbox": [ + 279, + 198, + 718, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Morgan Stanley Service Canada, $^{4}$ San Francisco State University,", + "bbox": [ + 235, + 215, + 761, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ University of Warwick, $^{6}$ Ant Group", + "bbox": [ + 352, + 233, + 645, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yolandachen0313@gmail.com, yifjia@isi.edu, lzhou1068@gmail.com,", + "bbox": [ + 217, + 252, + 777, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jcao3@alumni.sfsu.edu, yu.guan@warwick.ac.uk, {m.yang, qingpei.gqp} @antgroup.com", + "bbox": [ + 143, + 268, + 851, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/725b0b6e4590d8a73dc00f0fd2a727e430d513cb42c0405c4fb6736498134f62.jpg", + "image_caption": [ + "(b)", + "(c)" + ], + "image_footnote": [], + "bbox": [ + 145, + 299, + 542, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fc41a6ce9281e528e215569e232081d81be0e3fbe15bfeb569b1cc94b0868f2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 299, + 661, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f00537d9fe38944fe3e2eb8972a9c5e38d0821876e9060cac6b6d46d4cdacb7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 299, + 779, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0718fc342ef018db28f0919097d7427eb6a17d1dded5cc7099a182117b5aaa3e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 299, + 897, + 435 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/392d130013441ceb1476320e40bd8c1ec5857014b0a16d8ff273c24422312720.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 96, + 446, + 542, + 588 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0fe1aa93994b15778cba576facf1e8c0f528d4635359c7393420d20e32766c4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 439, + 661, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/84b566dd215d0710233ae04071977e5769bf23b7993cb1d00a185b86fe4f8da2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 439, + 758, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d09bbe516a3b9f4eb14f84e13ad41361adf84137f5bd15e0cc17c422c4a82226.jpg", + "image_caption": [ + "(e)" + ], + "image_footnote": [], + "bbox": [ + 761, + 439, + 897, + 530 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f9eddc224e6aabed8efcb3ebd824c499df1c9340e8b8301b5a87f356b9356fb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 545, + 542, + 563, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CharmNet:", + "bbox": [ + 566, + 545, + 604, + 551 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The tram is leading a parade, bringing life and color to the city as it glides by with effortless charm!", + "bbox": [ + 545, + 553, + 903, + 561 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1d641931a72eb9503e920177c0ea7b90cf536e115e4be74c049311bce9ffae59.jpg", + "image_caption": [ + "Figure 1. (a) to (c) present Conventional descriptions (highlighted in red) and appealing descriptions (highlighted in blue) for three ordinary images. (d) presents a sample from AppealBench with complete object descriptions. (e) presents four tasks applied to a sample from AppealBench across single-object, multi-object, and global perspectives." + ], + "image_footnote": [], + "bbox": [ + 545, + 565, + 563, + 575 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ground Truth:", + "bbox": [ + 566, + 566, + 616, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This tram must be on a seafood diet because it's definitely a shell of its former self!", + "bbox": [ + 566, + 575, + 849, + 583 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 688, + 326, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, multi-modal large language models (MLLMs) have been successfully adopted to generate humorous and engaging descriptions for internet memes. While, it is challenging for the same approaches to apply to ordinary images which lack of inherent funny or exaggerated contents. Thus, crafting appealing descriptions for ordinary image demands imaginative efforts to discover or create intriguing connections between words to image contents. To address this gap, we introduce Ap", + "bbox": [ + 88, + 723, + 485, + 859 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pealImage, a large-scale dataset consisting of ordinary images paired with appealing descriptions. AppealImage allows us to define four distinct tasks with quantitative metrics to enable objective evaluation. Subsequently, we propose CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet combines instruction tuning with heuristic active learning, guided by a referee model. Experimental results demonstrate that CharmNet outperforms the state-of-the-art method by $11.4\\%$ in generating appealing descriptions. Furthermore, CharmNet delivers impressive performance across various creative applications, including visual storytelling and situational dialogue generation. These results", + "bbox": [ + 511, + 690, + 906, + 886 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation.", + "bbox": [ + 238, + 0, + 807, + 18 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "Except for this watermark, it is identical to the accepted version;", + "bbox": [ + 325, + 16, + 722, + 30 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 292, + 31, + 754, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Work done during an internship at Ant Group.", + "bbox": [ + 107, + 875, + 372, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Qingpei Guo is the corresponding author.", + "bbox": [ + 109, + 887, + 348, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19342", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "highlight CharmNet's potential to enhance social media engagement and to empower strong brand presence in competitive markets.", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advancements in multi-modal large language models (MLLMs) have greatly expanded their capability in generating multi-modal humor. However, much of the existing research has concentrated on crafting exaggerated, humorous descriptions for internet memes, without attending to the issue of creating appealing descriptions for ordinary images [6, 44]. Unlike memes, which are inherently humorous, ordinary images typically lack superficial engaging features. Producing appealing descriptions for such images requires a sense of creativity and imaginative effort. The goal is to transform ordinary images into captivating content with compelling descriptions, an essential factor for boosting engagement on social media and enabling brand-oriented images to stand out in competitive markets.", + "bbox": [ + 89, + 191, + 482, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide several illustrative examples to demonstrate the differences between appealing descriptions and their conventional counterparts. For instance, in Fig.1 (a), phrases like \"trunk-call\" and \"splashy surprise\" employ creative wordplay to impress viewers. In Fig.1 (b), terms such as \"man-wich\" and \"hero of my lunch\" use phonetic similarity to capture attention. In Fig. 1 (c), expressions like \"sharp roof\" and \"slice through a sermon\" bring in irony through reversal to make the description appealing. In contrast, the corresponding conventional descriptions are plain and factual, simply listing the specific objects or actions in the images without any of these appealing elements.", + "bbox": [ + 89, + 402, + 482, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous research has primarily focused on generating accurate descriptions for ordinary images using fine-tuning or in-context learning [30, 32, 33, 41]. Other studies have emphasized creating humorous captions for images with inherent funny elements by leveraging fine-grained contrasts within the images alongside commonsense, bridging the gap between funny images and conventional descriptions [6, 34, 38, 44]. While funny images readily lend themselves to generating engaging and humorous descriptions, creating appealing descriptions for ordinary images remains a challenge [44]. We further identify two key challenges in identifying the appealing perspectives of objects/contents in ordinary images: (i) The current level of integration between visual and language clues is insufficient, limiting MLLMs' ability to capture the latent interest in ordinary images; (ii) MLLMs lack of supervised training or tuning dedicated to distinguish supervised training to distinguish between appealing and non-appealing descriptions for these types of images.", + "bbox": [ + 89, + 584, + 482, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For ordinary images, existing methods often struggle to generate sufficiently appealing descriptions. Chan-", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "drasekaran et al. [4] focus solely on puns, which offer only a limited form of wordplay in descriptions. Gao et al. [12] primarily produce promotional descriptions that highlight the connection between an object and human, rather than revealing the engaging features of objects. Shuster et al. [35] generate captions from diverse human perspectives, making the content engaging through shared viewpoints rather than by enhancing the intrinsic appeal of the object itself.", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this work, we first construct a large-scale dataset named AppealImage1, comprising ordinary images with appealing descriptions crafted using three strategies: wordplay, phonetic similarity, and reversal, drawing inspiration from meme theory. To thoroughly evaluate and enhance MLLMs' ability to produce appealing descriptions, we define four tasks, including discrimination, selection, ranking, and generation, targeting single-object, multi-object, and global perspectives. Each task is paired with a quantitative metric for objective assessment. After that, we introduce a simple-yet-effective framework named CharmNet, designed to generate appealing descriptions for ordinary images through instruction tuning and heuristic active learning, supervised by a referee model. We conduct a comprehensive evaluation of CharmNet on our AppealImage, where the experiments show that CharmNet consistently surpasses other competing methods, achieving an $11.4\\%$ improvement in generating appealing descriptions. Additionally, CharmNet demonstrates robust performance and generalization on various creative tasks, such as visual storytelling and situational dialogue generation.", + "bbox": [ + 511, + 212, + 903, + 529 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 546, + 653, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image Caption Generation. Rotstein et al. [33] fine-tuned a large language model (LLM) to produce natural and coherent image captions. Huang et al. [18] combined vision embeddings with language models to improve caption generation quality. Ma et al. [29] employed a multi-context data generation pipeline using condensed captions and diffusion models to advance image captioning. Dwibedi et al. [11] introduced FlexCap for generating localized captions targeting specific image regions. Hu et al. [15] developed Instruct-Imagen, which integrates style images with captions using multi-modal instructions. Al-Buraihy and Dan [1] enhanced cross-lingual image descriptions through a multi-modal approach. Li et al. [24] introduced EVCap, which employs retrieval-augmented captioning with an external visual-name memory. Chandrasekaran et al. [4] generated witty descriptions for given images by leveraging linguistic puns. Gao et al. [12] proposed the GEM framework for generating promotional multi-modal content. Shuster et al. [35] focused on generating personalized image captions. In contrast, our work focuses on crafting appealing", + "bbox": [ + 511, + 571, + 903, + 875 + ], + "page_idx": 1 + }, + { + "type": "page_footnote", + "text": "1https://github.com/Yukyin/AppealImage.", + "bbox": [ + 529, + 887, + 834, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19343", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/d924b733c3c171f36b0427f7da4dfa7765b99da92e4a7b5e0fc8665fe1d9a3bf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
SingleLDes\nMultipleGlobalNObjNRObjCategoryNIma
WPPSRe
14.816.019.43.51.033.0%3.4%63.6%74,942
", + "bbox": [ + 94, + 89, + 478, + 131 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Table 1. AppealImage Statistics. $\\mathbf{L_{Des}}$ : Average description length in AppealImage, $\\mathbf{N_{Obj}}$ and $\\mathbf{N_{RObj}}$ : Number of objects and related objects per image, $\\mathbf{N_{Ima}}$ : Total number of images in AppealImage. WP: Wordplay, PS: Phonetic Similarity, Re:Reversal.", + "bbox": [ + 89, + 143, + 482, + 202 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "descriptions for images themselves, rather than discussing the relationship between image contents and human.", + "bbox": [ + 89, + 227, + 482, + 258 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Humor Generation. In addition to research on textual humor understanding [5] and generation [7], there has been growing interest in multi-modal humor generation. Wang and Lee [38] developed MemeCraft, a meme generator that leverages LLMs and VLMs to create humorous, socially impactful memes. Zhong et al. [44] introduced the Leap-of-Thought paradigm to enhance LLMs' ability to generate humorous captions. Ruan et al. [34] focused on generating cross-modal sarcasm in multimodal dialogue systems. Chen et al. [6] proposed XMeCap, a novel approach to generating humorous captions across both single-image and multi-image memes. In contrast, our focus is on generating appealing descriptions for ordinary images rather than improving the humorous captions for inherently funny images.", + "bbox": [ + 91, + 258, + 483, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. AppealImage", + "text_level": 1, + "bbox": [ + 89, + 500, + 228, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we introduce a novel dataset named AppealImage and design four tasks to evaluate on this dataset.", + "bbox": [ + 89, + 525, + 482, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Data Construction", + "text_level": 1, + "bbox": [ + 89, + 566, + 269, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first collect 70K images from GQA [19], COCO [27], and ImageNet [9], representing objects and scenes grounded in the physical world. Inspired by the rise of internet memes like \"cash me outside\" $^{2}$ , we define three types of appealing descriptions, including wordplay, phonetic similarity, and reversal. Wordplay creatively leverages multiple meanings to produce a clever effect, as seen in terms like \"pedestrian\", which can mean both a person walking and something ordinary. Phonetic similarity plays on homophones or near-homophones, such as \"leafing\" versus \"leaving\". Reversal subverts the expected meaning for humor or irony, as in \"tired of the street life\", which personifies a car as if it were human. To generate detailed appealing descriptions, we define three perspectives, including single-object, multi-object, and global. The single-object description focuses on a specific object, while the multi-object description emphasizes the relationships among multiple objects within the image. The global description provides a comprehensive depiction of the entire image, addressing the", + "bbox": [ + 89, + 588, + 482, + 876 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "main scene and overall atmosphere. Next, we employ GPT-4o $^3$ to generate appealing descriptions for the collected images based on the three definitions across these perspectives with a sample shown in Fig. 1 (d).", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After that, GPT-4o is used to assess the quality of descriptions, first filtering out toxic contents related to politics, race, etc. For the remaining descriptions, each is scored on a 1-5 scale across four dimensions, including informativeness, factual accuracy, appeal, and fluency, retaining only those that score at least 4 in each dimension. Descriptions receiving scores below 4 are collected for training the referee, which will be introduced in the next section, and are then regenerated. To guarantee that the output is a scalar value, we employ GLM4 [13] to transform the text, ensuring the final score is appropriately scaled. Then, $10\\%$ of these descriptions are randomly selected for validation by three volunteer graduate students. They are first trained by scoring multiple 50-sample batches, with feedback after each round, until their scores reaching a Pearson correlation above 0.9 with us. Next, they are requested to score descriptions from 1 to 5 across the same dimensions. If more than $1\\%$ of the reviewed descriptions score below 4 in any dimension, the entire set is re-evaluated by GPT-4o. After re-scoring, a new random $10\\%$ sample is selected and subjected to the same manual validation process until fewer than $1\\%$ of descriptions score below 4. Ultimately, the GPT-4o and human scores show a correlation of 0.96, validating the final AppealImage dataset. Table 1 summarizes its statistics.", + "bbox": [ + 511, + 152, + 906, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Task Setup", + "text_level": 1, + "bbox": [ + 513, + 537, + 635, + 555 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing inspiration from Zhong et al. [44], in the discrimination task, an MLLM is presented with an image and two description options and is required to identify which description is more appealing. For the selection task, the MLLM receives an image and three description options, choosing the most appealing one. The ranking task involves the MLLM ranking four given descriptions of an image according to their level of appeal. In the generation task, the MLLM is tasked with generating an appealing description for an ordinary image. Moreover, for the first three tasks, we introduce three levels of difficulty for the description options, distributed equally. Difficulty levels, evaluated by GPT-4o, are determined by the variable $S$ , which quantifies the differentiation between options. These levels are categorized as follows: $[S_{min}, S_{min} + \\frac{1}{3}(S_{max} - S_{min})]$ , $[S_{min} + \\frac{1}{3}(S_{max} - S_{min})$ , $S_{min} + \\frac{2}{3}(S_{max} - S_{min})$ ), and $[S_{min} + \\frac{2}{3}(S_{max} - S_{min})$ , $S_{max}]$ , representing low, medium, and high difficulty, respectively, where $S_{min}$ and $S_{max}$ are the minimum and maximum scores. A sample covering all four tasks from single-object, multi-object, and global perspective is shown in Fig. 1(e).", + "bbox": [ + 511, + 560, + 906, + 878 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2https://www.dictionary.com/e/memes/cash-me-outside/", + "bbox": [ + 107, + 886, + 408, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "3https://chatgpt.com", + "bbox": [ + 531, + 886, + 643, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19344", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Methods", + "text_level": 1, + "bbox": [ + 89, + 89, + 189, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After constructing AppealImage, we introduce a robust method, CharmNet, for generating appealing descriptions, as illustrated in Fig. 2.", + "bbox": [ + 89, + 114, + 483, + 161 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Instruction Tuning", + "text_level": 1, + "bbox": [ + 89, + 172, + 272, + 188 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We first employ a two-stage instruction tuning strategy inspired by Li et al. [25]. In the first stage, we use a pretrained visual encoder, Qwen-VL [2], to extract image features, keeping the visual encoder initially frozen to maintain feature extraction stability. We then train a QFormer model to translates visual features into textual descriptions by minimizing the cross-entropy loss between predicted and ground-truth appealing descriptions. In the second stage, we refine the image-text alignment. We unfreeze the visual encoder and conduct supervised fine-tuning with LoRA [14] to further optimize QFormer.", + "bbox": [ + 89, + 195, + 483, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "After that, we use multi-task learning, enabling the QFormer to leverage diverse representations from various tasks. We normalize the weights of each task to ensure consistency in scale, as shown in the following formula:", + "bbox": [ + 89, + 362, + 483, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T}} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\mathrm {g}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 436, + 482, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{d}}, \\lambda_{\\mathrm{s}}, \\lambda_{\\mathrm{r}}$ , and $\\lambda_{\\mathrm{g}}$ represent the normalized weights for the discrimination, selection, ranking, and generation tasks, respectively. These parameters are initially set to be proportional and will be adjusted during training. $\\mathcal{L}$ denotes each task's specific cross-entropy loss function. The trained QFormer serves as the precursor to CharmNet and is referred to as pre-CharmNet.", + "bbox": [ + 89, + 465, + 483, + 571 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Referee Construction", + "text_level": 1, + "bbox": [ + 89, + 583, + 290, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Second, to effectively assess and supervise the appeal level of generated descriptions, we construct a referee model. We first create a warm-up dataset, $D_0$ , containing 1,000 ordinary images and corresponding candidate descriptions, each paired with appeal scores as outlined in the previous section. These scores undergo further validation through manual annotation. Using $D_0$ , we then use LLaVA-1.6 [28] to train a referee model. This referee takes an image and its descriptions as input and outputs scalar scores representing each description's appeal level.", + "bbox": [ + 89, + 606, + 483, + 757 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We employ two types of loss functions in this process. The first is a pairwise ranking loss, which guides the referee to assign higher scores to more appealing descriptions while assigning lower scores to less appealing ones. The loss function is defined as:", + "bbox": [ + 89, + 758, + 483, + 833 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {R}} = - \\frac {1}{\\binom {k} {2}} E _ {(I _ {i}, d _ {i, j}, d _ {i, j ^ {\\prime}}) \\sim D _ {0}} [ \\log (\\sigma (r _ {\\theta} (I _ {i}, d _ {i, j}) - \\\\ \\left. r _ {\\theta} \\left(I _ {i}, d _ {i, j ^ {\\prime}}\\right)\\right) \\bigg ], \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 843, + 482, + 897 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{i,j}$ and $d_{i,j'}$ are descriptions of image $I_i$ with higher and lower appeal scores, respectively. $r_\\theta$ is the scalar score assigned by the referee. $k$ denotes the number of descriptions per image, and $\\sigma$ is the sigmoid function, converting score differences into probability values. The second is an alignment loss function, which ensures the referee's scores align closely with human-annotated scores. To create these annotations, three volunteers provide scores from 1 to 5 for each description in $D_0$ , and inter-rater agreement, measured by Krippendorff's Alpha (IRA), is calculated to ensure the reliability of these scores. Descriptions with low agreement (IRA < 0.7) are discarded as Taylor and Watkinson [37] states that an IRA exceeding 0.7 is an excellent agreement. The alignment loss is defined as:", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {A}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) - r _ {\\mathrm {h}} \\left(I _ {i}, d _ {i, j}\\right)\\right) ^ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 315, + 906, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N$ is the total number of descriptions for each image, and $r_{\\mathrm{h}}(I_i,d_{i,j})$ is the human-annotated score for description $d_{i,j}$ of image $I_{i}$ . Finally, we combine the two loss functions in a weighted sum to create a referee that is both accurate and aligned with human preferences. The combined loss function is defined as:", + "bbox": [ + 511, + 369, + 906, + 459 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\alpha \\mathcal {L} _ {\\mathrm {R}} + \\beta \\mathcal {L} _ {\\mathrm {A}} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 650, + 474, + 906, + 491 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are trainable coefficients that adjust the contributions of the two loss components.", + "bbox": [ + 511, + 503, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Heuristic Active Learning", + "text_level": 1, + "bbox": [ + 511, + 547, + 750, + 564 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we adopt a heuristic active learning strategy ${}^{4}$ to further enhance pre-CharmNet's ability to generate appealing descriptions for ordinary images. This strategy is to identify the more challenging samples (i.e. those with low scores evaluated by the referee). The chosen challenging data points are re-incorporated into the training process, enabling pre-CharmNet to learn which descriptions are perceived as non-appealing. We also validate this observation with comprehensive experiments in Table 8.", + "bbox": [ + 511, + 570, + 905, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, we first create a validation dataset, denoted as $D_V$ , by selecting $10\\%$ of the data entries from AppealImage. The purpose of $D_V$ is to provide a benchmark for evaluating pre-CharmNet's performance during training iterations. In addition, we construct a hard dataset, $D_1$ , consisting of $X$ images whose descriptions received scores below a predefined threshold $\\alpha$ evaluated by the referee. The motivation of $D_1$ is to train pre-CharmNet on samples where the model currently underperforms, enabling it to learn from its weaknesses and refine its ability to generate more appealing descriptions.", + "bbox": [ + 511, + 707, + 906, + 873 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "4https://en.wikipedia.org/wiki/Active_learning_(machine_learning)", + "bbox": [ + 529, + 886, + 883, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19345", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/de7ed2fb34524a9abc9d0610ff38257c5b4292f70c559c10253d3bba76250c6b.jpg", + "image_caption": [ + "Figure 2. The framework of CharmNet, including Instruction Tuning which extract image features and refine image-text alignment with supervised fine-tuning, Referee Construction which assess and modify the appeal level of generated descriptions, and Heuristic Active Learning which further enhance CharmNet's ability to generate appealing descriptions for ordinary images." + ], + "image_footnote": [], + "bbox": [ + 135, + 90, + 859, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After each iteration of training, pre-CharmNet is re-evaluated on $D_V$ to assess whether its performance has improved measured by the referee. If no improvement is observed, the current set of $X$ images in $D_1$ is replaced with a new set of $X$ images that also fall below the threshold $\\mu$ . This replacement process ensures that pre-CharmNet is consistently exposed to new challenging examples, avoiding overfitting to specific samples in $D_1$ . The iterative training continues until pre-CharmNet demonstrates consistent improvement on $D_V$ over five consecutive evaluations, indicating that pre-CharmNet has effectively internalized insights from the difficult examples. The update process for constructing $D_1$ is defined as follows:", + "bbox": [ + 88, + 316, + 485, + 512 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nD _ {1} = \\left\\{\\left(I _ {i}, d _ {i, j}\\right) \\mid r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) < \\mu \\right\\}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 525, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu$ is the predefined performance threshold, $I_{i}$ represents an image, $d_{i,j}$ denotes its corresponding description, and $r_{\\theta}$ is the score assigned by the referee. Through this optimization process, pre-CharmNet is refined into the final CharmNet, equipped with an enhanced ability to generate more appealing descriptions.", + "bbox": [ + 89, + 554, + 483, + 645 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 660, + 225, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to evaluate the effectiveness of our proposed CharmNet.", + "bbox": [ + 89, + 686, + 482, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 727, + 282, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our experiments are conducted using PyTorch on four Nvidia A100 GPUs, each with 80GB of memory. We allocate $60\\%$ , $30\\%$ , and $10\\%$ of AppealImage for training, validation, and testing, respectively. The datasets $D_0$ and $D_1$ are derived from the validation set and do not overlap, while $D_V$ denotes the test set. The maximum sequence lengths for input and output sequences are set to 2048 and 512 tokens, respectively. We use the AdamW optimizer with $\\beta_{1}$ set to 0.9, $\\beta_{2}$ to 0.95, a learning rate of 1e-5, and a weight decay of 1e-1. The batch size is set to 8 during training. LoRA's", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "rank is set to 32, with a normalization parameter of 16 and a dropout rate of 0.05. Training is conducted for 10 epochs. Experiments validate Qwen-VL for instruction tuning and LLaVA-1.6 for referee construction as optimal choices. In referee construction, each image is paired with $k = 5$ descriptions, and the predefined evaluation threshold $\\mu = 0.2$ , both yielding optimal performance.", + "bbox": [ + 511, + 316, + 906, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Datasets, Baselines, and Metrics", + "text_level": 1, + "bbox": [ + 511, + 436, + 795, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize AppealImage and other creative tasks, including visual story creation and situational dialogue generation, to assess CharmNet's performance. For the visual story creation task, we employ the VIST dataset [17], where CharmNet generates appealing narratives based on a sequence of images. CharmNet first creates an appealing description for each individual image, which are then integrated into a cohesive story with the addition of transitional words like \"first,\" \"next,\" and \"then\", etc. In the situational dialogue generation task, we use the DialogCC dataset [21] to evaluate CharmNet's ability to produce appealing responses within dialogue histories that include images.", + "bbox": [ + 511, + 459, + 906, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The comparison baselines for generating appealing descriptions are shown in Table 2. For the VIST and DialogCC datasets, CharmNet's performance is evaluated using metrics that measure the distance between human and model outputs in terms of coherence (denoted as $d_{HM}^{C}$ ), visual grounding (denoted as $d_{HM}^{G}$ ), and repetition (denoted as $d_{HM}^{R}$ ), as well as R@1, R@5, R@10, and MRR, respectively, consistent with the corresponding research [21, 36].", + "bbox": [ + 511, + 642, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For evaluation criteria and metrics, the discrimination task is assessed based on whether the MLLM's response matches the ground truth, with top-1 accuracy as the metric. The selection task is evaluated similarly, using top-1 accuracy to determine the MLLM's selected response. For the ranking task, evaluation is based on how closely the MLLM's rankings align with the relative positions of the ground truth, with higher scores awarded for closer matches. The generation task is assessed with automated", + "bbox": [ + 511, + 765, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19346", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/47a4775277e2f4e0c421fcf9f507caad6f39d2f3465f8b7d16a63d2a4b703b98.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(Total)DSRG
ACCACCPCBROCMReH
CogVLM-7B [39]57.445.274.436.541.561.437.30.243.1
GPT-4o376.457.783.7------
MiniGPT-4-7B [45]41.128.471.131.136.456.132.90.171.8
LLaVA-1.6-7B [28]44.956.68045.351.470.746.50.364.3
InstructBLIP-7B [8]33.624.166.627.932.651.229.90.111.2
mPLUG-Owl-I [42]50.533.467.633.438.358.634.20.192.4
Otter-I [22]36.324.969.629.534.453.931.20.151.3
Qwen-VL-7B+CLoT [44]60.850.376.940.244.865.442.30.303.6
Qwen-VL-Chat-7B [2]54.643.57338.843.363.440.10.283.4
VisualGLM-6B [10]47.738.374.934.640.36035.90.222.9
XMeCap [6]67.542.77543.549.568.545.50.333.7
CharmNet77.559.583.947.253.373.148.00.384.4
↑1.11.80.21.91.92.41.50.020.1
↑(%)1.43.10.24.23.73.43.25.62.3
", + "bbox": [ + 94, + 89, + 478, + 256 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/dfe0695d6d3961fc0b73e59d2536b3025c09470b56c28979db127ae562b6bd2e.jpg", + "table_caption": [ + "Table 2. Comparison of MLLMs and CharmNet across various tasks based on AppealImage. D: Discrimination task, S: Selection task, R: Ranking task, G: Generation task. ACC: Accuracy. PC: Position Consistency. B: BLEU, RO:ROUGE, C: CIDEr, M: METEOR. Re: Referee score. H: Human score." + ], + "table_footnote": [], + "table_body": "
DSRDSRG
SiMuGlSiMuGl
CogVLM-7B [39]65.057.454.748.845.744.3
GPT-4o377.470.069.7---
MiniGPT-4-7B [45]53.047.739.842.239.936.5
LLaVA-1.6-7B [28]65.558.657.453.350.549.5
InstructBLIP-7B [8]47.342.035.037.935.732.4
mPLUG-Owl-I [42]57.050.843.844.341.838.7
Otter-I [22]49.543.637.739.737.334.5
Qwen-VL-7B+CLoT [44]68.461.558.052.049.347.5
Qwen-VL-Chat-7B [2]63.455.452.249.245.944.2
VisualGLM-6B [10]59.552.948.646.043.341.3
XMeCap [6]66.259.859.253.050.549.8
CharmNet79.770.171.159.455.755.5
↑2.30.11.46.15.25.7
↑(%)3.00.12.011.410.311.4
", + "bbox": [ + 142, + 349, + 433, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metrics, including BLEU [31], ROUGE [26], CIDEr, and METEOR [3], alongside appealing scores from 1 to 5 in both referee and human evaluations. For human evaluations, three volunteers are recruited to rate 1,000 randomly selected images and their generated descriptions. To ensure the reliability of human ratings, we calculate inter-rater agreement using Krippendorff's Alpha (IRA). Descriptions with ratings that fall below an IRA threshold of 0.7 are considered controversial, and these are discarded and replaced with descriptions from other images.", + "bbox": [ + 89, + 626, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Main Results", + "text_level": 1, + "bbox": [ + 89, + 787, + 228, + 801 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We begin by comparing MLLMs and CharmNet across various tasks on AppealImage, as shown in Table 2. Results indicate that the selected baseline MLLMs perform poorly in generating appealing descriptions for ordinary images, while CharmNet consistently outperforms these MLLMs across all evaluated tasks. Given that our dataset was gen", + "bbox": [ + 89, + 809, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1bad674f0c7ab8963ce4b3aad73bb752a0c0fc051cb34acf3b2b3ed1d4e1a164.jpg", + "table_caption": [ + "Table 3. Comparison of MLLMs and CharmNet across three perspectives in AppealImage. DSR: Discrimination, Selection, and Ranking tasks. DSRG: Discrimination, Selection, Ranking, and Generation tasks. Si: Single-object perspective, Mu: Multi-object perspective, Gl: Global perspective." + ], + "table_footnote": [], + "table_body": "
(Global)MaReHTime
Human-written48.60.314.330.5
CharmNet52.70.354.53.2
↑4.10.040.227.3
↑(%)8.412.94.7853.1
", + "bbox": [ + 614, + 89, + 805, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Comparison of CharmNet-generated and human-written appealing descriptions. Ma: Average of BLEU, ROUGE, CIDEr, and METEOR. Re: Referee score. H: Human score. Time: The average time to generate one description.", + "bbox": [ + 511, + 162, + 903, + 219 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "erated by GPT-4o, we exclude GPT-4o from the generation task to avoid inflated scores due to stylistic similarity with the original dataset. Notably, CharmNet achieves an accuracy of $77.5\\%$ in the discrimination task, surpassing GPT-4o's accuracy of $76.4\\%$ and showing an approximately $130.7\\%$ improvement over the lowest-performing MLLM, InstructBLIP-7B, which scored $33.6\\%$ . Similarly, CharmNet outperforms GPT-4o and demonstrates a marked improvement over InstructBLIP-7B in the selection task. In the ranking task, both CharmNet and GPT-4o exhibit top performance. In the generation task, CharmNet achieves a human-rated appeal score of 4.4, ranking the highest.", + "bbox": [ + 511, + 247, + 906, + 428 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we conduct a detailed comparison of MLLMs and CharmNet across different task perspectives, as shown in Table 3, once again excluding GPT-4o's performance in the generation task. We calculate the average performance across all MLLMs for the discrimination, selection, and ranking tasks, as well as for all MLLMs excluding GPT-4o. Results show that CharmNet consistently outperforms other MLLMs, with particularly strong results in single-object perspective generation.", + "bbox": [ + 511, + 429, + 908, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we invite three literary experts to write appealing descriptions for images, aiming to compare human performance with machine-generated descriptions in the task of in generating appealing descriptions. Due to the relatively long time required for human writing (approximately 30 seconds per description), we randomly select 1,000 images from the test set of AppealImage for evaluation. The results are shown in Table 4, demonstrating that CharmNet has surpassed human experts in both effectiveness and efficiency.", + "bbox": [ + 511, + 566, + 908, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Additionally, we evaluate CharmNet on two creative tasks, visual story creation and situational dialogue generation, with results shown in Table 5 and Table 6, respectively. Baseline results, excluding Qwen-VL-7B+CLoT, are drawn from previous studies [21, 36]. The findings indicate that CharmNet significantly outperforms the baselines, with only a slight improvement over Qwen-VL-7B+CLoT, validating CharmNet's generalizability across a variety of creative tasks.", + "bbox": [ + 511, + 718, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Last but not least, while maintaining creativity, we further evaluate CharmNet on non-creative tasks to investigate whether training on the AppealImage dataset leads to the", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19347", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1983ff5918e063b090c6420a41217e5bfca7f7e1883a1d69f9f75f413fb402e8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
(DialogCC)R@1R@5R@10MRR
Image Only8.2222.633.5216.94
Dialogue Only34.4165.3677.3748.67
Dialogue + Image40.6471.4681.9954.61
Owen-VL-7B+CLoT [44]52.6778.6788.2462.46
CharmNet56.3482.2191.4668.11
↑3.673.543.225.65
↑(%)7.04.53.69.0
", + "bbox": [ + 155, + 89, + 419, + 181 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e97b5e78fbc2b11dcbd71c51e00fcf44794ffbc6c7d5989f94c152dbe58c95ad.jpg", + "table_caption": [ + "Table 5. Performance of CharmNet on the contextual dialogue generation task." + ], + "table_footnote": [], + "table_body": "
(VIST)dCM(-)dGM(-)dRHM(-)dHM(-)
AREL [40]0.1100.0580.0400.208
GLAC Net [20]0.0700.1070.0070.184
TAPM [43]0.0950.0150.0400.150
BLIP-2 [23]0.1900.0500.0940.334
LLaVA [28]0.0480.0040.0470.099
Qwen-VL-7B+CLoT [44]0.0350.0030.0310.069
CharmNet0.0190.0020.0240.045
↓0.0160.0010.0070.024
↓(%)84.250.029.253.3
", + "bbox": [ + 138, + 234, + 434, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Performance of CharmNet on the visual storytelling task.", + "bbox": [ + 89, + 358, + 482, + 373 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "generation of irrelevant information. We employ an image captioning task using 500 randomly selected samples of the MSCOCO dataset, adjusting instructions to guide CharmNet in generating captions that are devoid of creative elements. We design three criteria based on key factors in image captioning as proposed in Huang et al. [16], and conduct manual evaluation with the results shown in Table 7. The manual evaluation show that $89.6\\%$ of the generated results correctly described the main objects and their relationships in the image, indicating that hallucination issues are relatively infrequent when CharmNet is used for non-creative tasks.", + "bbox": [ + 88, + 398, + 482, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The influence of different dataset size in heuristic active learning. We further examine the influence of varying the size of dataset $D_{1}$ in heuristic active learning, as illustrated in Fig. 3(a). The results show a steady increase in performance as the size of $D_{1}$ grows, peaking at $30\\%$ . Beyond this threshold, performance begins to decrease, indicating that a $D_{1}$ size of $30\\%$ achieves an optimal balance between data scale and quality.", + "bbox": [ + 89, + 580, + 482, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 712, + 243, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The impact of each module in CharmNet. We first make analysis of the impact of each module in CharmNet as shown in Table 8. All results are averaged for three times to avoid randomness. When the first stage of instruction tuning is removed (denoted as \"w/o IT1\"), we observe a substantial decline of $9.2\\%$ across all metrics. The possible reason may be that directly unfreezing both image and text together for learning mixed their features, preventing the model from learning useful information. Similarly, the removal of heuristic active learning (denoted as \"w/o HAL\") has a pronounced effect, indicating that HAL fa", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dad9c75132d4a4663037012fbdd55ac6c13975085e537f40b2bff4d0735837e8.jpg", + "image_caption": [ + "Figure 3. (a) Effect of the proportion of difficult data in heuristic active learning. GenerationM, GenerationR, and GenerationH represent machine scores (i.e., BLEU, ROUGE,CIDEr, METEOR), referee scores, and human scores, respectively. (b) Performance across tasks at varying difficulty levels." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 707, + 196 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9cc84b20e7e424f18259c29022017a6b4b4b0be5794b3527c4efb8b0dbcae0e6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 90, + 900, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cilitates targeted training and improvements through challenging samples. Furthermore, replacing HAL with reinforcement learning (denoted as \"HAL to RL\", we use direct preference optimization here) does not yield significant performance improvements but has substantially higher training costs. Moreover, substituting multi-task learning with single-task learning (denoted as \"MTL to STL\"), where only the generation task is used, does not enhance performance in the generation task. This suggests that complementary tasks contribute to CharmNet's understanding of what makes descriptions appealing, thereby improving the quality of generated outputs.", + "bbox": [ + 511, + 308, + 905, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The function of AppealImage to CharmNet. Next, we use Qwen-VL-7B+CLoT for instruction tuning on AppealImage. While it underperforms on AppealImage, it excels in the Oogiri game, which requires creativity and associative thinking for generating humorous responses. We compare its performance with CharmNet in global perspective generation task, as shown in Table 9, where CharmNet consistently outperforms the AppealImage-tuned Qwen-VL-7B+CLoT. This highlights that the effectiveness of CharmNet is not solely attributable to training on AppealImage but also to its combined strategies like multi-task learning and heuristic active learning.", + "bbox": [ + 511, + 489, + 905, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The effect of different difficulty levels in each tasks. After that, we investigate the effect of different difficulty levels on CharmNet's performance in the discrimination, selection, and ranking tasks, as shown in Fig. 3(b). Results reveal a near-linear performance decline across all tasks as difficulty increases from easy to hard. Data difficulty has the greatest impact on the selection task, followed by the discrimination task, likely due to the fewer options in discrimination tasks, which simplifies the identification of correct answers. The ranking task is not directly comparable due to its unique evaluation metrics.", + "bbox": [ + 511, + 672, + 905, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The correlation between referee and human scores for CharmNet. Since ground-truth labels for these tasks are provided by the referee, we employ the Pearson correlation coefficient to ensure alignment between referee and", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19348", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9c1268476380869415f87742eced307c1b67044964d2a0290baea98e9f3da5d8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CriteriaPortion (%)
Descriptions of the main objects in the image are incorrect or the relationships between objects in the image are wrong.52.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, but the descriptions are straightforward and unappealing.26.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, and the descriptions are creative and appealing.42.0
", + "bbox": [ + 91, + 89, + 903, + 140 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f68f91f6e04be7f89042ad83b55efc2bc87a0b69218f32905fa444897efe038f.jpg", + "table_caption": [ + "Table 7. Three criteria and the corresponding portion based on manual evaluation on 500 samples from the MSCOCO dataset." + ], + "table_footnote": [], + "table_body": "
(Total)D (↓%)S (↓%)R (↓%)GBRCM (↓%)GRe (↓%)GH (↓%)A (↓%)
w/o IT171.8 (7.9)54.0 (10.2)78.0 (7.6)39.0 (13.8)0.29 (31.0)4.0 (10.0)41.2 (9.2)
MTL to STL---43.1 (3.0)0.34 (11.8)4.2 (4.8)-
HAL to RL75.8 (2.2)56.9 (4.6)81.2 (3.3)44.1 (0.7)0.37 (2.7)4.2 (4.8)43.8 (2.7)
w/o HAL66.0 (17.4)48.1 (23.7)73.8 (13.7)33.7 (31.8)0.28 (35.7)3.5 (25.7)37.6 (19.7)
CharmNet77.559.583.944.40.384.445.0
", + "bbox": [ + 264, + 178, + 733, + 250 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/858db653aab95c0af75fbaf453d71a6b48b1678ed83f37f3d8bb6ba8bb58d000.jpg", + "image_caption": [ + "Figure 4. A running case of CharmNet in generating appealing descriptions." + ], + "image_footnote": [], + "bbox": [ + 200, + 304, + 797, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/307faffe276bea8c98615444fe2712f69e91b950f560b11bac732b26b67c802b.jpg", + "table_caption": [ + "Table 8. Impact of each module in CharmNet. ${\\mathrm{G}}_{BRCM}$ , ${\\mathrm{G}}_{Re}$ ,and ${\\mathrm{G}}_{H}$ denote machine,referee,and human scores,respectively. A: Average results." + ], + "table_footnote": [], + "table_body": "
(Total)BROCMReH
Qwen-VL-7B+CLoT45.550.271.647.10.344.1
CharmNet47.253.373.148.00.384.4
↑1.73.11.50.90.040.3
↑(%)3.76.22.11.911.87.3
", + "bbox": [ + 155, + 464, + 419, + 525 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/87c24bf9a1bbf9bf7a8bb62ca2d430c4a2c5daed61191dfdf8fa3111f6ffb7ba.jpg", + "table_caption": [ + "Table 9. Comparison of AppealImage-tuned Qwen-VL-7B+CLoT and CharmNet in the generation task." + ], + "table_footnote": [], + "table_body": "
DSRG
Single-object0.820.740.710.89
Multiple-object0.800.750.700.87
Global0.850.780.730.86
Total0.820.760.710.87
", + "bbox": [ + 194, + 578, + 379, + 638 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 10. Pearson correlation coefficients between referee and human scores for CharmNet.", + "bbox": [ + 89, + 648, + 482, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "human scores for CharmNet. The results are shown in Table 10, indicating a high level of consistency with all values exceeding 0.8 (p<0.05), especially in the generation task. We also test with a larger dataset, but the correlation coefficient does not show significant improvement. This high correlation underscores the reliability of the referee, demonstrating its potential to reduce the need for extensive human involvement while enhancing CharmNet's ability to generate appealing descriptions.", + "bbox": [ + 89, + 703, + 482, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Case Study", + "text_level": 1, + "bbox": [ + 89, + 848, + 212, + 864 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a case study of the appealing descriptions generated by CharmNet, illustrated in Fig. 4. For instance, for the", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "single-object description, CharmNet uses terms like “zips” and “exciting trail” to describe the tram, conveying a sense of speed. In the multi-object description, phrases such as “races by”, “vibrant carnival”, “swaying”, and “flashing” depict a static scene as a lively city celebration. Finally, in the global description, words like “leading a parade”, “bringing life and color”, and “effortless charm” enhance the image’s atmosphere, making it appear more dynamic and engaging. We also observe some less appealing descriptions, such as “a boy walking with a man in a crowd …”, which lacks vivid detail and emotional depth.", + "bbox": [ + 511, + 465, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 513, + 648, + 640, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "How to craft appealing descriptions by MLLMs for ordinary images beyond those internet memes is quite underinvestigated before. To bridge this gap, we present AppealImage, a large-scale dataset containing ordinary images paired with appealing descriptions. We define four tasks, each paired with quantitative metrics, to enable thorough evaluation. Upon AppealImage, we introduce CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet leverages instruction tuning and heuristic active learning, guided by a referee model. Our experiments show CharmNet outperforms the state-of-the-art method by $11.4\\%$ in generating appealing descriptions and has impressive results across related creative tasks. Future work will aim to broaden CharmNet's applicability and customize it for personalized contents.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19349", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by Ant Group Research Intern Program.", + "bbox": [ + 89, + 114, + 482, + 147 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 161, + 187, + 176 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Emran Al-Buraihy and W Dan. Enhancing crosslingual image description: A multimodal approach for semantic relevance and stylistic alignment. Comput. Mater. Contin, 2, 2024. 2", + "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 4, 6", + "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6", + "[4] Arjun Chandrasekaran, Devi Parikh, and Mohit Bansal. Funny captions: Witty wordplay in image descriptions. arXiv preprint arXiv:1704.08224, 2017. 2", + "[5] Yuyan Chen, Zhixu Li, Jiaqing Liang, Yanghua Xiao, Bang Liu, and Yunwen Chen. Can pre-trained language models understand chinese humor? 2023. 3", + "[6] Yuyan Chen, Songzhou Yan, Zhihong Zhu, Zhixu Li, and Yanghua Xiao. Xmecap: Meme caption generation with sub-image adaptability. arXiv preprint arXiv:2407.17152, 2024. 2, 3, 6", + "[7] Yuyan Chen, Yichen Yuan, Panjun Liu, Dayiheng Liu, Qinghao Guan, Mengfei Guo, Haiming Peng, Bang Liu, Zhixu Li, and Yanghua Xiao. Talk funny! a largescale humor response dataset with chain-of-humor interpretation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17826-17834, 2024. 3", + "[8] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3", + "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 6" + ], + "bbox": [ + 93, + 186, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. arXiv preprint arXiv:2403.12026, 2024. 2", + "[12] Chongyang Gao, Yiren Jian, Natalia Denisenko, Soroush Vosoughi, and VS Subrahmanian. Gem: generating engaging multimodal content. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 7654–7662, 2024. 2", + "[13] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Jingyu Sun, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 3", + "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 4", + "[15] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4754-4763, 2024. 2", + "[16] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4634-4643, 2019. 7", + "[17] Ting-Hao Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, et al. Visual storytelling. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: Human language technologies, pages 1233-1239, 2016. 5", + "[18] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings" + ], + "bbox": [ + 516, + 90, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19350", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 2", + "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3", + "[20] Taehyeong Kim, Min-Oh Heo, Seonil Son, Kyoung-Wha Park, and Byoung-Tak Zhang. Glac net: Glocal attention cascading networks for multi-image cued story generation. arXiv preprint arXiv:1805.10973, 2018. 7", + "[21] Young-Jun Lee, Byungsoo Ko, Han-Gyu Kim, and Ho-Jin Choi. Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119, 2022. 5, 6", + "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multimodal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 6", + "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models, 2023. 7", + "[24] Jiaxuan Li, Duc Minh Vo, Akihiro Sugimoto, and Hideki Nakayama. Evcap: Retrieval-augmented image captioning with external visual-name memory for open-world comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2024. 2", + "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 4", + "[26] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, 2004. Association for Computational Linguistics. 6", + "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 3", + "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 4, 6, 7", + "[29] Feipeng Ma, Yizhou Zhou, Fengyun Rao, Yueyi Zhang, and Xiaoyan Sun. Image captioning with multi-context synthetic data. In Proceedings of the" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "AAAI Conference on Artificial Intelligence, pages 4089-4097, 2024. 2", + "[30] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6", + "[32] Suzanne Petryk, Spencer Whitehead, Joseph E Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus Rohrbach. Simple token-level confidence improves caption correctness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5742-5752, 2024. 2", + "[33] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 2", + "[34] Jie Ruan, Yue Wu, Xiaojun Wan, and Yuesheng Zhu. Describe images in a boring way: Towards cross-modal sarcasm generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5701-5710, 2024. 2, 3", + "[35] Kurt Shuster, Samuel Humeau, Hexiang Hu, Antoine Bordes, and Jason Weston. Engaging image captioning via personality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12516-12526, 2019. 2", + "[36] Aditya K Surikuchi, Raquel Fernandez, and Sandro Pezzelle. Not (yet) the whole story: Evaluating visual storytelling requires more than measuring coherence, grounding, and repetition. arXiv preprint arXiv:2407.04559, 2024. 5, 6", + "[37] Joel Taylor and David Watkinson. Indexing reliability for condition survey data. The Conservator, 30(1):49-62, 2007. 4", + "[38] Han Wang and Roy Ka-Wei Lee. Memecraft: Contextual and stance-driven multimodal meme generation. In Proceedings of the ACM on Web Conference 2024, pages 4642–4652, 2024. 2, 3", + "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19351", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Xin Wang, Wenhu Chen, Yuan-Fang Wang, and William Yang Wang. No metrics are perfect: Adversarial reward learning for visual storytelling. arXiv preprint arXiv:1804.09160, 2018.7", + "[41] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[42] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 6", + "[43] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jongseok Kim, and Gunhee Kim. Transitional adaptation of pretrained models for visual storytelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12658-12668, 2021. 7", + "[44] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 2, 3, 6, 7", + "[45] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 6" + ], + "bbox": [ + 89, + 90, + 482, + 534 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19352", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_model.json b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..420094f468153f807e2723bbe7d090b2029cf843 --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_model.json @@ -0,0 +1,2290 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.02 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + }, + { + "type": "header", + "bbox": [ + 0.326, + 0.017, + 0.723, + 0.031 + ], + "angle": 0, + "content": "Except for this watermark, it is identical to the accepted version;" + }, + { + "type": "header", + "bbox": [ + 0.294, + 0.032, + 0.755, + 0.047 + ], + "angle": 0, + "content": "the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.861, + 0.154 + ], + "angle": 0, + "content": "Engage for All: Making Ordinary Image Descriptions Appealing Again!" + }, + { + "type": "text", + "bbox": [ + 0.111, + 0.179, + 0.892, + 0.199 + ], + "angle": 0, + "content": "Yuyan Chen\\(^{1*}\\), Yifan Jiang\\(^{2}\\), Li Zhou\\(^{3}\\), Jinghan Cao\\(^{4}\\), Yu Guan\\(^{5}\\), Ming Yang\\(^{6}\\), Qingpei Guo\\(^{5,6\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.28, + 0.199, + 0.719, + 0.217 + ], + "angle": 0, + "content": "1Cornell University, 2University of Southern California," + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.217, + 0.762, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)Morgan Stanley Service Canada, \\(^{4}\\)San Francisco State University," + }, + { + "type": "text", + "bbox": [ + 0.353, + 0.234, + 0.646, + 0.253 + ], + "angle": 0, + "content": "\\(^{5}\\)University of Warwick, \\(^{6}\\)Ant Group" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.253, + 0.778, + 0.27 + ], + "angle": 0, + "content": "yolandachen0313@gmail.com, yifjia@isi.edu, lzhou1068@gmail.com," + }, + { + "type": "text", + "bbox": [ + 0.144, + 0.27, + 0.852, + 0.288 + ], + "angle": 0, + "content": "jcao3@alumni.sfsu.edu, yu.guan@warwick.ac.uk, {m.yang, qingpei.gqp} @antgroup.com" + }, + { + "type": "image", + "bbox": [ + 0.147, + 0.3, + 0.543, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.437, + 0.331, + 0.446 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.3, + 0.662, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.459, + 0.437, + 0.472, + 0.445 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.3, + 0.78, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.3, + 0.898, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.448, + 0.543, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.59, + 0.337, + 0.599 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.44, + 0.662, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.667, + 0.44, + 0.759, + 0.491 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.762, + 0.44, + 0.898, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.533, + 0.746, + 0.54 + ], + "angle": 0, + "content": "(e)" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.543, + 0.565, + 0.553 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.546, + 0.606, + 0.553 + ], + "angle": 0, + "content": "CharmNet:" + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.554, + 0.904, + 0.563 + ], + "angle": 0, + "content": "The tram is leading a parade, bringing life and color to the city as it glides by with effortless charm!" + }, + { + "type": "image", + "bbox": [ + 0.547, + 0.566, + 0.565, + 0.576 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.567, + 0.617, + 0.575 + ], + "angle": 0, + "content": "Ground Truth:" + }, + { + "type": "text", + "bbox": [ + 0.567, + 0.576, + 0.851, + 0.584 + ], + "angle": 0, + "content": "This tram must be on a seafood diet because it's definitely a shell of its former self!" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.613, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 1. (a) to (c) present Conventional descriptions (highlighted in red) and appealing descriptions (highlighted in blue) for three ordinary images. (d) presents a sample from AppealBench with complete object descriptions. (e) presents four tasks applied to a sample from AppealBench across single-object, multi-object, and global perspectives." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.689, + 0.327, + 0.705 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.724, + 0.486, + 0.86 + ], + "angle": 0, + "content": "In recent years, multi-modal large language models (MLLMs) have been successfully adopted to generate humorous and engaging descriptions for internet memes. While, it is challenging for the same approaches to apply to ordinary images which lack of inherent funny or exaggerated contents. Thus, crafting appealing descriptions for ordinary image demands imaginative efforts to discover or create intriguing connections between words to image contents. To address this gap, we introduce Ap" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.691, + 0.907, + 0.887 + ], + "angle": 0, + "content": "pealImage, a large-scale dataset consisting of ordinary images paired with appealing descriptions. AppealImage allows us to define four distinct tasks with quantitative metrics to enable objective evaluation. Subsequently, we propose CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet combines instruction tuning with heuristic active learning, guided by a referee model. Experimental results demonstrate that CharmNet outperforms the state-of-the-art method by \\(11.4\\%\\) in generating appealing descriptions. Furthermore, CharmNet delivers impressive performance across various creative applications, including visual storytelling and situational dialogue generation. These results" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.373, + 0.888 + ], + "angle": 0, + "content": "* Work done during an internship at Ant Group." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.349, + 0.9 + ], + "angle": 0, + "content": "† Qingpei Guo is the corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.373, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19342" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "highlight CharmNet's potential to enhance social media engagement and to empower strong brand presence in competitive markets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.192, + 0.483, + 0.403 + ], + "angle": 0, + "content": "Recent advancements in multi-modal large language models (MLLMs) have greatly expanded their capability in generating multi-modal humor. However, much of the existing research has concentrated on crafting exaggerated, humorous descriptions for internet memes, without attending to the issue of creating appealing descriptions for ordinary images [6, 44]. Unlike memes, which are inherently humorous, ordinary images typically lack superficial engaging features. Producing appealing descriptions for such images requires a sense of creativity and imaginative effort. The goal is to transform ordinary images into captivating content with compelling descriptions, an essential factor for boosting engagement on social media and enabling brand-oriented images to stand out in competitive markets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.483, + 0.584 + ], + "angle": 0, + "content": "We provide several illustrative examples to demonstrate the differences between appealing descriptions and their conventional counterparts. For instance, in Fig.1 (a), phrases like \"trunk-call\" and \"splashy surprise\" employ creative wordplay to impress viewers. In Fig.1 (b), terms such as \"man-wich\" and \"hero of my lunch\" use phonetic similarity to capture attention. In Fig. 1 (c), expressions like \"sharp roof\" and \"slice through a sermon\" bring in irony through reversal to make the description appealing. In contrast, the corresponding conventional descriptions are plain and factual, simply listing the specific objects or actions in the images without any of these appealing elements." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Previous research has primarily focused on generating accurate descriptions for ordinary images using fine-tuning or in-context learning [30, 32, 33, 41]. Other studies have emphasized creating humorous captions for images with inherent funny elements by leveraging fine-grained contrasts within the images alongside commonsense, bridging the gap between funny images and conventional descriptions [6, 34, 38, 44]. While funny images readily lend themselves to generating engaging and humorous descriptions, creating appealing descriptions for ordinary images remains a challenge [44]. We further identify two key challenges in identifying the appealing perspectives of objects/contents in ordinary images: (i) The current level of integration between visual and language clues is insufficient, limiting MLLMs' ability to capture the latent interest in ordinary images; (ii) MLLMs lack of supervised training or tuning dedicated to distinguish supervised training to distinguish between appealing and non-appealing descriptions for these types of images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "For ordinary images, existing methods often struggle to generate sufficiently appealing descriptions. Chan-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "drasekaran et al. [4] focus solely on puns, which offer only a limited form of wordplay in descriptions. Gao et al. [12] primarily produce promotional descriptions that highlight the connection between an object and human, rather than revealing the engaging features of objects. Shuster et al. [35] generate captions from diverse human perspectives, making the content engaging through shared viewpoints rather than by enhancing the intrinsic appeal of the object itself." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.905, + 0.53 + ], + "angle": 0, + "content": "Therefore, in this work, we first construct a large-scale dataset named AppealImage1, comprising ordinary images with appealing descriptions crafted using three strategies: wordplay, phonetic similarity, and reversal, drawing inspiration from meme theory. To thoroughly evaluate and enhance MLLMs' ability to produce appealing descriptions, we define four tasks, including discrimination, selection, ranking, and generation, targeting single-object, multi-object, and global perspectives. Each task is paired with a quantitative metric for objective assessment. After that, we introduce a simple-yet-effective framework named CharmNet, designed to generate appealing descriptions for ordinary images through instruction tuning and heuristic active learning, supervised by a referee model. We conduct a comprehensive evaluation of CharmNet on our AppealImage, where the experiments show that CharmNet consistently surpasses other competing methods, achieving an \\(11.4\\%\\) improvement in generating appealing descriptions. Additionally, CharmNet demonstrates robust performance and generalization on various creative tasks, such as visual storytelling and situational dialogue generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.655, + 0.563 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.905, + 0.875 + ], + "angle": 0, + "content": "Image Caption Generation. Rotstein et al. [33] fine-tuned a large language model (LLM) to produce natural and coherent image captions. Huang et al. [18] combined vision embeddings with language models to improve caption generation quality. Ma et al. [29] employed a multi-context data generation pipeline using condensed captions and diffusion models to advance image captioning. Dwibedi et al. [11] introduced FlexCap for generating localized captions targeting specific image regions. Hu et al. [15] developed Instruct-Imagen, which integrates style images with captions using multi-modal instructions. Al-Buraihy and Dan [1] enhanced cross-lingual image descriptions through a multi-modal approach. Li et al. [24] introduced EVCap, which employs retrieval-augmented captioning with an external visual-name memory. Chandrasekaran et al. [4] generated witty descriptions for given images by leveraging linguistic puns. Gao et al. [12] proposed the GEM framework for generating promotional multi-modal content. Shuster et al. [35] focused on generating personalized image captions. In contrast, our work focuses on crafting appealing" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.888, + 0.835, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/Yukyin/AppealImage." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19343" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.48, + 0.132 + ], + "angle": 0, + "content": "
SingleLDes\nMultipleGlobalNObjNRObjCategoryNIma
WPPSRe
14.816.019.43.51.033.0%3.4%63.6%74,942
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.144, + 0.483, + 0.203 + ], + "angle": 0, + "content": "Table 1. AppealImage Statistics. \\( \\mathbf{L_{Des}} \\): Average description length in AppealImage, \\( \\mathbf{N_{Obj}} \\) and \\( \\mathbf{N_{RObj}} \\): Number of objects and related objects per image, \\( \\mathbf{N_{Ima}} \\): Total number of images in AppealImage. WP: Wordplay, PS: Phonetic Similarity, Re:Reversal." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.228, + 0.483, + 0.259 + ], + "angle": 0, + "content": "descriptions for images themselves, rather than discussing the relationship between image contents and human." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.26, + 0.485, + 0.486 + ], + "angle": 0, + "content": "Multi-modal Humor Generation. In addition to research on textual humor understanding [5] and generation [7], there has been growing interest in multi-modal humor generation. Wang and Lee [38] developed MemeCraft, a meme generator that leverages LLMs and VLMs to create humorous, socially impactful memes. Zhong et al. [44] introduced the Leap-of-Thought paradigm to enhance LLMs' ability to generate humorous captions. Ruan et al. [34] focused on generating cross-modal sarcasm in multimodal dialogue systems. Chen et al. [6] proposed XMeCap, a novel approach to generating humorous captions across both single-image and multi-image memes. In contrast, our focus is on generating appealing descriptions for ordinary images rather than improving the humorous captions for inherently funny images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.501, + 0.229, + 0.518 + ], + "angle": 0, + "content": "3. AppealImage" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.526, + 0.483, + 0.557 + ], + "angle": 0, + "content": "In this section, we introduce a novel dataset named AppealImage and design four tasks to evaluate on this dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.567, + 0.271, + 0.582 + ], + "angle": 0, + "content": "3.1. Data Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.589, + 0.483, + 0.877 + ], + "angle": 0, + "content": "We first collect 70K images from GQA [19], COCO [27], and ImageNet [9], representing objects and scenes grounded in the physical world. Inspired by the rise of internet memes like \"cash me outside\" \\(^{2}\\), we define three types of appealing descriptions, including wordplay, phonetic similarity, and reversal. Wordplay creatively leverages multiple meanings to produce a clever effect, as seen in terms like \"pedestrian\", which can mean both a person walking and something ordinary. Phonetic similarity plays on homophones or near-homophones, such as \"leafing\" versus \"leaving\". Reversal subverts the expected meaning for humor or irony, as in \"tired of the street life\", which personifies a car as if it were human. To generate detailed appealing descriptions, we define three perspectives, including single-object, multi-object, and global. The single-object description focuses on a specific object, while the multi-object description emphasizes the relationships among multiple objects within the image. The global description provides a comprehensive depiction of the entire image, addressing the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "main scene and overall atmosphere. Next, we employ GPT-4o \\(^3\\) to generate appealing descriptions for the collected images based on the three definitions across these perspectives with a sample shown in Fig. 1 (d)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.529 + ], + "angle": 0, + "content": "After that, GPT-4o is used to assess the quality of descriptions, first filtering out toxic contents related to politics, race, etc. For the remaining descriptions, each is scored on a 1-5 scale across four dimensions, including informativeness, factual accuracy, appeal, and fluency, retaining only those that score at least 4 in each dimension. Descriptions receiving scores below 4 are collected for training the referee, which will be introduced in the next section, and are then regenerated. To guarantee that the output is a scalar value, we employ GLM4 [13] to transform the text, ensuring the final score is appropriately scaled. Then, \\(10\\%\\) of these descriptions are randomly selected for validation by three volunteer graduate students. They are first trained by scoring multiple 50-sample batches, with feedback after each round, until their scores reaching a Pearson correlation above 0.9 with us. Next, they are requested to score descriptions from 1 to 5 across the same dimensions. If more than \\(1\\%\\) of the reviewed descriptions score below 4 in any dimension, the entire set is re-evaluated by GPT-4o. After re-scoring, a new random \\(10\\%\\) sample is selected and subjected to the same manual validation process until fewer than \\(1\\%\\) of descriptions score below 4. Ultimately, the GPT-4o and human scores show a correlation of 0.96, validating the final AppealImage dataset. Table 1 summarizes its statistics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.539, + 0.637, + 0.556 + ], + "angle": 0, + "content": "3.2. Task Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Drawing inspiration from Zhong et al. [44], in the discrimination task, an MLLM is presented with an image and two description options and is required to identify which description is more appealing. For the selection task, the MLLM receives an image and three description options, choosing the most appealing one. The ranking task involves the MLLM ranking four given descriptions of an image according to their level of appeal. In the generation task, the MLLM is tasked with generating an appealing description for an ordinary image. Moreover, for the first three tasks, we introduce three levels of difficulty for the description options, distributed equally. Difficulty levels, evaluated by GPT-4o, are determined by the variable \\(S\\), which quantifies the differentiation between options. These levels are categorized as follows: \\([S_{min}, S_{min} + \\frac{1}{3}(S_{max} - S_{min})]\\), \\([S_{min} + \\frac{1}{3}(S_{max} - S_{min})\\), \\(S_{min} + \\frac{2}{3}(S_{max} - S_{min})\\)), and \\([S_{min} + \\frac{2}{3}(S_{max} - S_{min})\\), \\(S_{max}]\\), representing low, medium, and high difficulty, respectively, where \\(S_{min}\\) and \\(S_{max}\\) are the minimum and maximum scores. A sample covering all four tasks from single-object, multi-object, and global perspective is shown in Fig. 1(e)." + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.409, + 0.901 + ], + "angle": 0, + "content": "2https://www.dictionary.com/e/memes/cash-me-outside/" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.644, + 0.901 + ], + "angle": 0, + "content": "3https://chatgpt.com" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19344" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.191, + 0.106 + ], + "angle": 0, + "content": "4. Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.485, + 0.162 + ], + "angle": 0, + "content": "After constructing AppealImage, we introduce a robust method, CharmNet, for generating appealing descriptions, as illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.173, + 0.274, + 0.189 + ], + "angle": 0, + "content": "4.1. Instruction Tuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.196, + 0.484, + 0.362 + ], + "angle": 0, + "content": "We first employ a two-stage instruction tuning strategy inspired by Li et al. [25]. In the first stage, we use a pretrained visual encoder, Qwen-VL [2], to extract image features, keeping the visual encoder initially frozen to maintain feature extraction stability. We then train a QFormer model to translates visual features into textual descriptions by minimizing the cross-entropy loss between predicted and ground-truth appealing descriptions. In the second stage, we refine the image-text alignment. We unfreeze the visual encoder and conduct supervised fine-tuning with LoRA [14] to further optimize QFormer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.363, + 0.484, + 0.424 + ], + "angle": 0, + "content": "After that, we use multi-task learning, enabling the QFormer to leverage diverse representations from various tasks. We normalize the weights of each task to ensure consistency in scale, as shown in the following formula:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.437, + 0.483, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T}} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\mathrm {g}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.467, + 0.484, + 0.573 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{d}}, \\lambda_{\\mathrm{s}}, \\lambda_{\\mathrm{r}}\\), and \\(\\lambda_{\\mathrm{g}}\\) represent the normalized weights for the discrimination, selection, ranking, and generation tasks, respectively. These parameters are initially set to be proportional and will be adjusted during training. \\(\\mathcal{L}\\) denotes each task's specific cross-entropy loss function. The trained QFormer serves as the precursor to CharmNet and is referred to as pre-CharmNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.584, + 0.292, + 0.599 + ], + "angle": 0, + "content": "4.2. Referee Construction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.758 + ], + "angle": 0, + "content": "Second, to effectively assess and supervise the appeal level of generated descriptions, we construct a referee model. We first create a warm-up dataset, \\(D_0\\), containing 1,000 ordinary images and corresponding candidate descriptions, each paired with appeal scores as outlined in the previous section. These scores undergo further validation through manual annotation. Using \\(D_0\\), we then use LLaVA-1.6 [28] to train a referee model. This referee takes an image and its descriptions as input and outputs scalar scores representing each description's appeal level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.484, + 0.834 + ], + "angle": 0, + "content": "We employ two types of loss functions in this process. The first is a pairwise ranking loss, which guides the referee to assign higher scores to more appealing descriptions while assigning lower scores to less appealing ones. The loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.119, + 0.844, + 0.483, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {R}} = - \\frac {1}{\\binom {k} {2}} E _ {(I _ {i}, d _ {i, j}, d _ {i, j ^ {\\prime}}) \\sim D _ {0}} [ \\log (\\sigma (r _ {\\theta} (I _ {i}, d _ {i, j}) - \\\\ \\left. r _ {\\theta} \\left(I _ {i}, d _ {i, j ^ {\\prime}}\\right)\\right) \\bigg ], \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "where \\( d_{i,j} \\) and \\( d_{i,j'} \\) are descriptions of image \\( I_i \\) with higher and lower appeal scores, respectively. \\( r_\\theta \\) is the scalar score assigned by the referee. \\( k \\) denotes the number of descriptions per image, and \\( \\sigma \\) is the sigmoid function, converting score differences into probability values. The second is an alignment loss function, which ensures the referee's scores align closely with human-annotated scores. To create these annotations, three volunteers provide scores from 1 to 5 for each description in \\( D_0 \\), and inter-rater agreement, measured by Krippendorff's Alpha (IRA), is calculated to ensure the reliability of these scores. Descriptions with low agreement (IRA < 0.7) are discarded as Taylor and Watkinson [37] states that an IRA exceeding 0.7 is an excellent agreement. The alignment loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.316, + 0.907, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {A}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) - r _ {\\mathrm {h}} \\left(I _ {i}, d _ {i, j}\\right)\\right) ^ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.907, + 0.46 + ], + "angle": 0, + "content": "where \\(N\\) is the total number of descriptions for each image, and \\(r_{\\mathrm{h}}(I_i,d_{i,j})\\) is the human-annotated score for description \\(d_{i,j}\\) of image \\(I_{i}\\). Finally, we combine the two loss functions in a weighted sum to create a referee that is both accurate and aligned with human preferences. The combined loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.651, + 0.475, + 0.907, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\alpha \\mathcal {L} _ {\\mathrm {R}} + \\beta \\mathcal {L} _ {\\mathrm {A}} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.505, + 0.906, + 0.536 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are trainable coefficients that adjust the contributions of the two loss components." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.548, + 0.75, + 0.565 + ], + "angle": 0, + "content": "4.3. Heuristic Active Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.906, + 0.707 + ], + "angle": 0, + "content": "In this section, we adopt a heuristic active learning strategy \\( {}^{4} \\) to further enhance pre-CharmNet's ability to generate appealing descriptions for ordinary images. This strategy is to identify the more challenging samples (i.e. those with low scores evaluated by the referee). The chosen challenging data points are re-incorporated into the training process, enabling pre-CharmNet to learn which descriptions are perceived as non-appealing. We also validate this observation with comprehensive experiments in Table 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.708, + 0.907, + 0.874 + ], + "angle": 0, + "content": "Specifically, we first create a validation dataset, denoted as \\( D_V \\), by selecting \\( 10\\% \\) of the data entries from AppealImage. The purpose of \\( D_V \\) is to provide a benchmark for evaluating pre-CharmNet's performance during training iterations. In addition, we construct a hard dataset, \\( D_1 \\), consisting of \\( X \\) images whose descriptions received scores below a predefined threshold \\( \\alpha \\) evaluated by the referee. The motivation of \\( D_1 \\) is to train pre-CharmNet on samples where the model currently underperforms, enabling it to learn from its weaknesses and refine its ability to generate more appealing descriptions." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.884, + 0.902 + ], + "angle": 0, + "content": "4https://en.wikipedia.org/wiki/Active_learning_(machine_learning)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.137, + 0.092, + 0.861, + 0.235 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.248, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 2. The framework of CharmNet, including Instruction Tuning which extract image features and refine image-text alignment with supervised fine-tuning, Referee Construction which assess and modify the appeal level of generated descriptions, and Heuristic Active Learning which further enhance CharmNet's ability to generate appealing descriptions for ordinary images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.317, + 0.486, + 0.513 + ], + "angle": 0, + "content": "After each iteration of training, pre-CharmNet is re-evaluated on \\( D_V \\) to assess whether its performance has improved measured by the referee. If no improvement is observed, the current set of \\( X \\) images in \\( D_1 \\) is replaced with a new set of \\( X \\) images that also fall below the threshold \\( \\mu \\). This replacement process ensures that pre-CharmNet is consistently exposed to new challenging examples, avoiding overfitting to specific samples in \\( D_1 \\). The iterative training continues until pre-CharmNet demonstrates consistent improvement on \\( D_V \\) over five consecutive evaluations, indicating that pre-CharmNet has effectively internalized insights from the difficult examples. The update process for constructing \\( D_1 \\) is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.526, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\nD _ {1} = \\left\\{\\left(I _ {i}, d _ {i, j}\\right) \\mid r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) < \\mu \\right\\}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.646 + ], + "angle": 0, + "content": "where \\(\\mu\\) is the predefined performance threshold, \\(I_{i}\\) represents an image, \\(d_{i,j}\\) denotes its corresponding description, and \\(r_{\\theta}\\) is the score assigned by the referee. Through this optimization process, pre-CharmNet is refined into the final CharmNet, equipped with an enhanced ability to generate more appealing descriptions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.661, + 0.226, + 0.679 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.718 + ], + "angle": 0, + "content": "In this section, we conduct experiments to evaluate the effectiveness of our proposed CharmNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.283, + 0.745 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Our experiments are conducted using PyTorch on four Nvidia A100 GPUs, each with 80GB of memory. We allocate \\(60\\%\\), \\(30\\%\\), and \\(10\\%\\) of AppealImage for training, validation, and testing, respectively. The datasets \\(D_0\\) and \\(D_1\\) are derived from the validation set and do not overlap, while \\(D_V\\) denotes the test set. The maximum sequence lengths for input and output sequences are set to 2048 and 512 tokens, respectively. We use the AdamW optimizer with \\(\\beta_{1}\\) set to 0.9, \\(\\beta_{2}\\) to 0.95, a learning rate of 1e-5, and a weight decay of 1e-1. The batch size is set to 8 during training. LoRA's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.317, + 0.907, + 0.424 + ], + "angle": 0, + "content": "rank is set to 32, with a normalization parameter of 16 and a dropout rate of 0.05. Training is conducted for 10 epochs. Experiments validate Qwen-VL for instruction tuning and LLaVA-1.6 for referee construction as optimal choices. In referee construction, each image is paired with \\( k = 5 \\) descriptions, and the predefined evaluation threshold \\( \\mu = 0.2 \\), both yielding optimal performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.437, + 0.796, + 0.453 + ], + "angle": 0, + "content": "5.2. Datasets, Baselines, and Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.642 + ], + "angle": 0, + "content": "We utilize AppealImage and other creative tasks, including visual story creation and situational dialogue generation, to assess CharmNet's performance. For the visual story creation task, we employ the VIST dataset [17], where CharmNet generates appealing narratives based on a sequence of images. CharmNet first creates an appealing description for each individual image, which are then integrated into a cohesive story with the addition of transitional words like \"first,\" \"next,\" and \"then\", etc. In the situational dialogue generation task, we use the DialogCC dataset [21] to evaluate CharmNet's ability to produce appealing responses within dialogue histories that include images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.765 + ], + "angle": 0, + "content": "The comparison baselines for generating appealing descriptions are shown in Table 2. For the VIST and DialogCC datasets, CharmNet's performance is evaluated using metrics that measure the distance between human and model outputs in terms of coherence (denoted as \\( d_{HM}^{C} \\)), visual grounding (denoted as \\( d_{HM}^{G} \\)), and repetition (denoted as \\( d_{HM}^{R} \\)), as well as R@1, R@5, R@10, and MRR, respectively, consistent with the corresponding research [21, 36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.902 + ], + "angle": 0, + "content": "For evaluation criteria and metrics, the discrimination task is assessed based on whether the MLLM's response matches the ground truth, with top-1 accuracy as the metric. The selection task is evaluated similarly, using top-1 accuracy to determine the MLLM's selected response. For the ranking task, evaluation is based on how closely the MLLM's rankings align with the relative positions of the ground truth, with higher scores awarded for closer matches. The generation task is assessed with automated" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19346" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.48, + 0.257 + ], + "angle": 0, + "content": "
(Total)DSRG
ACCACCPCBROCMReH
CogVLM-7B [39]57.445.274.436.541.561.437.30.243.1
GPT-4o376.457.783.7------
MiniGPT-4-7B [45]41.128.471.131.136.456.132.90.171.8
LLaVA-1.6-7B [28]44.956.68045.351.470.746.50.364.3
InstructBLIP-7B [8]33.624.166.627.932.651.229.90.111.2
mPLUG-Owl-I [42]50.533.467.633.438.358.634.20.192.4
Otter-I [22]36.324.969.629.534.453.931.20.151.3
Qwen-VL-7B+CLoT [44]60.850.376.940.244.865.442.30.303.6
Qwen-VL-Chat-7B [2]54.643.57338.843.363.440.10.283.4
VisualGLM-6B [10]47.738.374.934.640.36035.90.222.9
XMeCap [6]67.542.77543.549.568.545.50.333.7
CharmNet77.559.583.947.253.373.148.00.384.4
↑1.11.80.21.91.92.41.50.020.1
↑(%)1.43.10.24.23.73.43.25.62.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.337 + ], + "angle": 0, + "content": "Table 2. Comparison of MLLMs and CharmNet across various tasks based on AppealImage. D: Discrimination task, S: Selection task, R: Ranking task, G: Generation task. ACC: Accuracy. PC: Position Consistency. B: BLEU, RO:ROUGE, C: CIDEr, M: METEOR. Re: Referee score. H: Human score." + }, + { + "type": "table", + "bbox": [ + 0.143, + 0.351, + 0.434, + 0.519 + ], + "angle": 0, + "content": "
DSRDSRG
SiMuGlSiMuGl
CogVLM-7B [39]65.057.454.748.845.744.3
GPT-4o377.470.069.7---
MiniGPT-4-7B [45]53.047.739.842.239.936.5
LLaVA-1.6-7B [28]65.558.657.453.350.549.5
InstructBLIP-7B [8]47.342.035.037.935.732.4
mPLUG-Owl-I [42]57.050.843.844.341.838.7
Otter-I [22]49.543.637.739.737.334.5
Qwen-VL-7B+CLoT [44]68.461.558.052.049.347.5
Qwen-VL-Chat-7B [2]63.455.452.249.245.944.2
VisualGLM-6B [10]59.552.948.646.043.341.3
XMeCap [6]66.259.859.253.050.549.8
CharmNet79.770.171.159.455.755.5
↑2.30.11.46.15.25.7
↑(%)3.00.12.011.410.311.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.603 + ], + "angle": 0, + "content": "Table 3. Comparison of MLLMs and CharmNet across three perspectives in AppealImage. DSR: Discrimination, Selection, and Ranking tasks. DSRG: Discrimination, Selection, Ranking, and Generation tasks. Si: Single-object perspective, Mu: Multi-object perspective, Gl: Global perspective." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.627, + 0.483, + 0.779 + ], + "angle": 0, + "content": "metrics, including BLEU [31], ROUGE [26], CIDEr, and METEOR [3], alongside appealing scores from 1 to 5 in both referee and human evaluations. For human evaluations, three volunteers are recruited to rate 1,000 randomly selected images and their generated descriptions. To ensure the reliability of human ratings, we calculate inter-rater agreement using Krippendorff's Alpha (IRA). Descriptions with ratings that fall below an IRA threshold of 0.7 are considered controversial, and these are discarded and replaced with descriptions from other images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.788, + 0.229, + 0.803 + ], + "angle": 0, + "content": "5.3. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We begin by comparing MLLMs and CharmNet across various tasks on AppealImage, as shown in Table 2. Results indicate that the selected baseline MLLMs perform poorly in generating appealing descriptions for ordinary images, while CharmNet consistently outperforms these MLLMs across all evaluated tasks. Given that our dataset was gen" + }, + { + "type": "table", + "bbox": [ + 0.615, + 0.09, + 0.807, + 0.152 + ], + "angle": 0, + "content": "
(Global)MaReHTime
Human-written48.60.314.330.5
CharmNet52.70.354.53.2
↑4.10.040.227.3
↑(%)8.412.94.7853.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.163, + 0.905, + 0.22 + ], + "angle": 0, + "content": "Table 4. Comparison of CharmNet-generated and human-written appealing descriptions. Ma: Average of BLEU, ROUGE, CIDEr, and METEOR. Re: Referee score. H: Human score. Time: The average time to generate one description." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.429 + ], + "angle": 0, + "content": "erated by GPT-4o, we exclude GPT-4o from the generation task to avoid inflated scores due to stylistic similarity with the original dataset. Notably, CharmNet achieves an accuracy of \\(77.5\\%\\) in the discrimination task, surpassing GPT-4o's accuracy of \\(76.4\\%\\) and showing an approximately \\(130.7\\%\\) improvement over the lowest-performing MLLM, InstructBLIP-7B, which scored \\(33.6\\%\\). Similarly, CharmNet outperforms GPT-4o and demonstrates a marked improvement over InstructBLIP-7B in the selection task. In the ranking task, both CharmNet and GPT-4o exhibit top performance. In the generation task, CharmNet achieves a human-rated appeal score of 4.4, ranking the highest." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.43, + 0.909, + 0.567 + ], + "angle": 0, + "content": "Next, we conduct a detailed comparison of MLLMs and CharmNet across different task perspectives, as shown in Table 3, once again excluding GPT-4o's performance in the generation task. We calculate the average performance across all MLLMs for the discrimination, selection, and ranking tasks, as well as for all MLLMs excluding GPT-4o. Results show that CharmNet consistently outperforms other MLLMs, with particularly strong results in single-object perspective generation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.718 + ], + "angle": 0, + "content": "Furthermore, we invite three literary experts to write appealing descriptions for images, aiming to compare human performance with machine-generated descriptions in the task of in generating appealing descriptions. Due to the relatively long time required for human writing (approximately 30 seconds per description), we randomly select 1,000 images from the test set of AppealImage for evaluation. The results are shown in Table 4, demonstrating that CharmNet has surpassed human experts in both effectiveness and efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Additionally, we evaluate CharmNet on two creative tasks, visual story creation and situational dialogue generation, with results shown in Table 5 and Table 6, respectively. Baseline results, excluding Qwen-VL-7B+CLoT, are drawn from previous studies [21, 36]. The findings indicate that CharmNet significantly outperforms the baselines, with only a slight improvement over Qwen-VL-7B+CLoT, validating CharmNet's generalizability across a variety of creative tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Last but not least, while maintaining creativity, we further evaluate CharmNet on non-creative tasks to investigate whether training on the AppealImage dataset leads to the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19347" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.09, + 0.42, + 0.182 + ], + "angle": 0, + "content": "
(DialogCC)R@1R@5R@10MRR
Image Only8.2222.633.5216.94
Dialogue Only34.4165.3677.3748.67
Dialogue + Image40.6471.4681.9954.61
Owen-VL-7B+CLoT [44]52.6778.6788.2462.46
CharmNet56.3482.2191.4668.11
↑3.673.543.225.65
↑(%)7.04.53.69.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.194, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Table 5. Performance of CharmNet on the contextual dialogue generation task." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.236, + 0.436, + 0.349 + ], + "angle": 0, + "content": "
(VIST)dCM(-)dGM(-)dRHM(-)dHM(-)
AREL [40]0.1100.0580.0400.208
GLAC Net [20]0.0700.1070.0070.184
TAPM [43]0.0950.0150.0400.150
BLIP-2 [23]0.1900.0500.0940.334
LLaVA [28]0.0480.0040.0470.099
Qwen-VL-7B+CLoT [44]0.0350.0030.0310.069
CharmNet0.0190.0020.0240.045
↓0.0160.0010.0070.024
↓(%)84.250.029.253.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.359, + 0.483, + 0.374 + ], + "angle": 0, + "content": "Table 6. Performance of CharmNet on the visual storytelling task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.4, + 0.483, + 0.581 + ], + "angle": 0, + "content": "generation of irrelevant information. We employ an image captioning task using 500 randomly selected samples of the MSCOCO dataset, adjusting instructions to guide CharmNet in generating captions that are devoid of creative elements. We design three criteria based on key factors in image captioning as proposed in Huang et al. [16], and conduct manual evaluation with the results shown in Table 7. The manual evaluation show that \\(89.6\\%\\) of the generated results correctly described the main objects and their relationships in the image, indicating that hallucination issues are relatively infrequent when CharmNet is used for non-creative tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.705 + ], + "angle": 0, + "content": "The influence of different dataset size in heuristic active learning. We further examine the influence of varying the size of dataset \\( D_{1} \\) in heuristic active learning, as illustrated in Fig. 3(a). The results show a steady increase in performance as the size of \\( D_{1} \\) grows, peaking at \\( 30\\% \\). Beyond this threshold, performance begins to decrease, indicating that a \\( D_{1} \\) size of \\( 30\\% \\) achieves an optimal balance between data scale and quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.245, + 0.729 + ], + "angle": 0, + "content": "5.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The impact of each module in CharmNet. We first make analysis of the impact of each module in CharmNet as shown in Table 8. All results are averaged for three times to avoid randomness. When the first stage of instruction tuning is removed (denoted as \"w/o IT1\"), we observe a substantial decline of \\(9.2\\%\\) across all metrics. The possible reason may be that directly unfreezing both image and text together for learning mixed their features, preventing the model from learning useful information. Similarly, the removal of heuristic active learning (denoted as \"w/o HAL\") has a pronounced effect, indicating that HAL fa" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.709, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.091, + 0.901, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.211, + 0.906, + 0.281 + ], + "angle": 0, + "content": "Figure 3. (a) Effect of the proportion of difficult data in heuristic active learning. GenerationM, GenerationR, and GenerationH represent machine scores (i.e., BLEU, ROUGE,CIDEr, METEOR), referee scores, and human scores, respectively. (b) Performance across tasks at varying difficulty levels." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.309, + 0.906, + 0.49 + ], + "angle": 0, + "content": "cilitates targeted training and improvements through challenging samples. Furthermore, replacing HAL with reinforcement learning (denoted as \"HAL to RL\", we use direct preference optimization here) does not yield significant performance improvements but has substantially higher training costs. Moreover, substituting multi-task learning with single-task learning (denoted as \"MTL to STL\"), where only the generation task is used, does not enhance performance in the generation task. This suggests that complementary tasks contribute to CharmNet's understanding of what makes descriptions appealing, thereby improving the quality of generated outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.672 + ], + "angle": 0, + "content": "The function of AppealImage to CharmNet. Next, we use Qwen-VL-7B+CLoT for instruction tuning on AppealImage. While it underperforms on AppealImage, it excels in the Oogiri game, which requires creativity and associative thinking for generating humorous responses. We compare its performance with CharmNet in global perspective generation task, as shown in Table 9, where CharmNet consistently outperforms the AppealImage-tuned Qwen-VL-7B+CLoT. This highlights that the effectiveness of CharmNet is not solely attributable to training on AppealImage but also to its combined strategies like multi-task learning and heuristic active learning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.839 + ], + "angle": 0, + "content": "The effect of different difficulty levels in each tasks. After that, we investigate the effect of different difficulty levels on CharmNet's performance in the discrimination, selection, and ranking tasks, as shown in Fig. 3(b). Results reveal a near-linear performance decline across all tasks as difficulty increases from easy to hard. Data difficulty has the greatest impact on the selection task, followed by the discrimination task, likely due to the fewer options in discrimination tasks, which simplifies the identification of correct answers. The ranking task is not directly comparable due to its unique evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The correlation between referee and human scores for CharmNet. Since ground-truth labels for these tasks are provided by the referee, we employ the Pearson correlation coefficient to ensure alignment between referee and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19348" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.09, + 0.905, + 0.141 + ], + "angle": 0, + "content": "
CriteriaPortion (%)
Descriptions of the main objects in the image are incorrect or the relationships between objects in the image are wrong.52.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, but the descriptions are straightforward and unappealing.26.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, and the descriptions are creative and appealing.42.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.125, + 0.15, + 0.87, + 0.165 + ], + "angle": 0, + "content": "Table 7. Three criteria and the corresponding portion based on manual evaluation on 500 samples from the MSCOCO dataset." + }, + { + "type": "table", + "bbox": [ + 0.266, + 0.179, + 0.734, + 0.251 + ], + "angle": 0, + "content": "
(Total)D (↓%)S (↓%)R (↓%)GBRCM (↓%)GRe (↓%)GH (↓%)A (↓%)
w/o IT171.8 (7.9)54.0 (10.2)78.0 (7.6)39.0 (13.8)0.29 (31.0)4.0 (10.0)41.2 (9.2)
MTL to STL---43.1 (3.0)0.34 (11.8)4.2 (4.8)-
HAL to RL75.8 (2.2)56.9 (4.6)81.2 (3.3)44.1 (0.7)0.37 (2.7)4.2 (4.8)43.8 (2.7)
w/o HAL66.0 (17.4)48.1 (23.7)73.8 (13.7)33.7 (31.8)0.28 (35.7)3.5 (25.7)37.6 (19.7)
CharmNet77.559.583.944.40.384.445.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.261, + 0.905, + 0.29 + ], + "angle": 0, + "content": "Table 8. Impact of each module in CharmNet. \\( {\\mathrm{G}}_{BRCM} \\) , \\( {\\mathrm{G}}_{Re} \\) ,and \\( {\\mathrm{G}}_{H} \\) denote machine,referee,and human scores,respectively. A: Average results." + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.305, + 0.799, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.272, + 0.427, + 0.724, + 0.442 + ], + "angle": 0, + "content": "Figure 4. A running case of CharmNet in generating appealing descriptions." + }, + { + "type": "table", + "bbox": [ + 0.156, + 0.465, + 0.421, + 0.526 + ], + "angle": 0, + "content": "
(Total)BROCMReH
Qwen-VL-7B+CLoT45.550.271.647.10.344.1
CharmNet47.253.373.148.00.384.4
↑1.73.11.50.90.040.3
↑(%)3.76.22.11.911.87.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.566 + ], + "angle": 0, + "content": "Table 9. Comparison of AppealImage-tuned Qwen-VL-7B+CLoT and CharmNet in the generation task." + }, + { + "type": "table", + "bbox": [ + 0.196, + 0.579, + 0.38, + 0.64 + ], + "angle": 0, + "content": "
DSRG
Single-object0.820.740.710.89
Multiple-object0.800.750.700.87
Global0.850.780.730.86
Total0.820.760.710.87
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.65, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 10. Pearson correlation coefficients between referee and human scores for CharmNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.84 + ], + "angle": 0, + "content": "human scores for CharmNet. The results are shown in Table 10, indicating a high level of consistency with all values exceeding 0.8 (p<0.05), especially in the generation task. We also test with a larger dataset, but the correlation coefficient does not show significant improvement. This high correlation underscores the reliability of the referee, demonstrating its potential to reduce the need for extensive human involvement while enhancing CharmNet's ability to generate appealing descriptions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.214, + 0.866 + ], + "angle": 0, + "content": "5.5. Case Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We present a case study of the appealing descriptions generated by CharmNet, illustrated in Fig. 4. For instance, for the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.634 + ], + "angle": 0, + "content": "single-object description, CharmNet uses terms like “zips” and “exciting trail” to describe the tram, conveying a sense of speed. In the multi-object description, phrases such as “races by”, “vibrant carnival”, “swaying”, and “flashing” depict a static scene as a lively city celebration. Finally, in the global description, words like “leading a parade”, “bringing life and color”, and “effortless charm” enhance the image’s atmosphere, making it appear more dynamic and engaging. We also observe some less appealing descriptions, such as “a boy walking with a man in a crowd …”, which lacks vivid detail and emotional depth." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.641, + 0.665 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "How to craft appealing descriptions by MLLMs for ordinary images beyond those internet memes is quite underinvestigated before. To bridge this gap, we present AppealImage, a large-scale dataset containing ordinary images paired with appealing descriptions. We define four tasks, each paired with quantitative metrics, to enable thorough evaluation. Upon AppealImage, we introduce CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet leverages instruction tuning and heuristic active learning, guided by a referee model. Our experiments show CharmNet outperforms the state-of-the-art method by \\(11.4\\%\\) in generating appealing descriptions and has impressive results across related creative tasks. Future work will aim to broaden CharmNet's applicability and customize it for personalized contents." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19349" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.148 + ], + "angle": 0, + "content": "This work is supported by Ant Group Research Intern Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.162, + 0.188, + 0.178 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.188, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[1] Emran Al-Buraihy and W Dan. Enhancing crosslingual image description: A multimodal approach for semantic relevance and stylistic alignment. Comput. Mater. Contin, 2, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.251, + 0.484, + 0.326 + ], + "angle": 0, + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.419 + ], + "angle": 0, + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.421, + 0.483, + 0.467 + ], + "angle": 0, + "content": "[4] Arjun Chandrasekaran, Devi Parikh, and Mohit Bansal. Funny captions: Witty wordplay in image descriptions. arXiv preprint arXiv:1704.08224, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.469, + 0.484, + 0.514 + ], + "angle": 0, + "content": "[5] Yuyan Chen, Zhixu Li, Jiaqing Liang, Yanghua Xiao, Bang Liu, and Yunwen Chen. Can pre-trained language models understand chinese humor? 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.516, + 0.484, + 0.576 + ], + "angle": 0, + "content": "[6] Yuyan Chen, Songzhou Yan, Zhihong Zhu, Zhixu Li, and Yanghua Xiao. Xmecap: Meme caption generation with sub-image adaptability. arXiv preprint arXiv:2407.17152, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.682 + ], + "angle": 0, + "content": "[7] Yuyan Chen, Yichen Yuan, Panjun Liu, Dayiheng Liu, Qinghao Guan, Mengfei Guo, Haiming Peng, Bang Liu, Zhixu Li, and Yanghua Xiao. Talk funny! a largescale humor response dataset with chain-of-humor interpretation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17826-17834, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[8] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.822 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.826, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.188, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "[11] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. arXiv preprint arXiv:2403.12026, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.156, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[12] Chongyang Gao, Yiren Jian, Natalia Denisenko, Soroush Vosoughi, and VS Subrahmanian. Gem: generating engaging multimodal content. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 7654–7662, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[13] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Jingyu Sun, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.495, + 0.905, + 0.556 + ], + "angle": 0, + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.559, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[15] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4754-4763, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.668, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[16] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4634-4643, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.905, + 0.853 + ], + "angle": 0, + "content": "[17] Ting-Hao Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, et al. Visual storytelling. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: Human language technologies, pages 1233-1239, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.856, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[18] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19350" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.128, + 0.092, + 0.484, + 0.122 + ], + "angle": 0, + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.124, + 0.484, + 0.2 + ], + "angle": 0, + "content": "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.202, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[20] Taehyeong Kim, Min-Oh Heo, Seonil Son, Kyoung-Wha Park, and Byoung-Tak Zhang. Glac net: Glocal attention cascading networks for multi-image cued story generation. arXiv preprint arXiv:1805.10973, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.34 + ], + "angle": 0, + "content": "[21] Young-Jun Lee, Byungsoo Ko, Han-Gyu Kim, and Ho-Jin Choi. Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.342, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multimodal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.482, + 0.465 + ], + "angle": 0, + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.467, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[24] Jiaxuan Li, Duc Minh Vo, Akihiro Sugimoto, and Hideki Nakayama. Evcap: Retrieval-augmented image captioning with external visual-name memory for open-world comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.651 + ], + "angle": 0, + "content": "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.653, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[26] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, 2004. Association for Computational Linguistics. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.806 + ], + "angle": 0, + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.808, + 0.482, + 0.853 + ], + "angle": 0, + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[29] Feipeng Ma, Yizhou Zhou, Fengyun Rao, Yueyi Zhang, and Xiaoyan Sun. Image captioning with multi-context synthetic data. In Proceedings of the" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.55, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "AAAI Conference on Artificial Intelligence, pages 4089-4097, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.184 + ], + "angle": 0, + "content": "[30] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.187, + 0.905, + 0.293 + ], + "angle": 0, + "content": "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.386 + ], + "angle": 0, + "content": "[32] Suzanne Petryk, Spencer Whitehead, Joseph E Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus Rohrbach. Simple token-level confidence improves caption correctness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5742-5752, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.388, + 0.905, + 0.477 + ], + "angle": 0, + "content": "[33] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.481, + 0.908, + 0.558 + ], + "angle": 0, + "content": "[34] Jie Ruan, Yue Wu, Xiaojun Wan, and Yuesheng Zhu. Describe images in a boring way: Towards cross-modal sarcasm generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5701-5710, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[35] Kurt Shuster, Samuel Humeau, Hexiang Hu, Antoine Bordes, and Jason Weston. Engaging image captioning via personality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12516-12526, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[36] Aditya K Surikuchi, Raquel Fernandez, and Sandro Pezzelle. Not (yet) the whole story: Evaluating visual storytelling requires more than measuring coherence, grounding, and repetition. arXiv preprint arXiv:2407.04559, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.715, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[37] Joel Taylor and David Watkinson. Indexing reliability for condition survey data. The Conservator, 30(1):49-62, 2007. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.823 + ], + "angle": 0, + "content": "[38] Han Wang and Roy Ka-Wei Lee. Memecraft: Contextual and stance-driven multimodal meme generation. In Proceedings of the ACM on Web Conference 2024, pages 4642–4652, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19351" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.091, + 0.482, + 0.152 + ], + "angle": 0, + "content": "[40] Xin Wang, Wenhu Chen, Yuan-Fang Wang, and William Yang Wang. No metrics are perfect: Adversarial reward learning for visual storytelling. arXiv preprint arXiv:1804.09160, 2018.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.153, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[41] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.215, + 0.483, + 0.291 + ], + "angle": 0, + "content": "[42] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.293, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[43] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jongseok Kim, and Gunhee Kim. Transitional adaptation of pretrained models for visual storytelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12658-12668, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.369, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[44] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[45] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "19352" + } + ] +] \ No newline at end of file diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_origin.pdf b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..57143b0b04a2b730926dcf7fda98f2ce70158292 --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/ad64f5b0-b3c7-476c-8285-9a78b446995d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f91d4349dca6b9dcbc7c7f1c6ca8c6de2479e7bde0803e4b7c37f91f8ded89d8 +size 2445634 diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/full.md b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/full.md new file mode 100644 index 0000000000000000000000000000000000000000..39fc6fa34a8a407badeea8739bf0b71400da0d2b --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/full.md @@ -0,0 +1,318 @@ +# Engage for All: Making Ordinary Image Descriptions Appealing Again! + +Yuyan Chen $^{1*}$ , Yifan Jiang $^{2}$ , Li Zhou $^{3}$ , Jinghan Cao $^{4}$ , Yu Guan $^{5}$ , Ming Yang $^{6}$ , Qingpei Guo $^{5,6\dagger}$ + +1Cornell University, 2University of Southern California, + +$^{3}$ Morgan Stanley Service Canada, $^{4}$ San Francisco State University, + +$^{5}$ University of Warwick, $^{6}$ Ant Group + +yolandachen0313@gmail.com, yifjia@isi.edu, lzhou1068@gmail.com, + +jcao3@alumni.sfsu.edu, yu.guan@warwick.ac.uk, {m.yang, qingpei.gqp} @antgroup.com + +![](images/725b0b6e4590d8a73dc00f0fd2a727e430d513cb42c0405c4fb6736498134f62.jpg) +(b) +(c) + +![](images/fc41a6ce9281e528e215569e232081d81be0e3fbe15bfeb569b1cc94b0868f2d.jpg) + +![](images/f00537d9fe38944fe3e2eb8972a9c5e38d0821876e9060cac6b6d46d4cdacb7e.jpg) + +![](images/0718fc342ef018db28f0919097d7427eb6a17d1dded5cc7099a182117b5aaa3e.jpg) + +![](images/392d130013441ceb1476320e40bd8c1ec5857014b0a16d8ff273c24422312720.jpg) +(d) + +![](images/0fe1aa93994b15778cba576facf1e8c0f528d4635359c7393420d20e32766c4b.jpg) + +![](images/84b566dd215d0710233ae04071977e5769bf23b7993cb1d00a185b86fe4f8da2.jpg) + +![](images/d09bbe516a3b9f4eb14f84e13ad41361adf84137f5bd15e0cc17c422c4a82226.jpg) +(e) + +![](images/f9eddc224e6aabed8efcb3ebd824c499df1c9340e8b8301b5a87f356b9356fb0.jpg) + +CharmNet: + +The tram is leading a parade, bringing life and color to the city as it glides by with effortless charm! + +![](images/1d641931a72eb9503e920177c0ea7b90cf536e115e4be74c049311bce9ffae59.jpg) +Figure 1. (a) to (c) present Conventional descriptions (highlighted in red) and appealing descriptions (highlighted in blue) for three ordinary images. (d) presents a sample from AppealBench with complete object descriptions. (e) presents four tasks applied to a sample from AppealBench across single-object, multi-object, and global perspectives. + +Ground Truth: + +This tram must be on a seafood diet because it's definitely a shell of its former self! + +# Abstract + +In recent years, multi-modal large language models (MLLMs) have been successfully adopted to generate humorous and engaging descriptions for internet memes. While, it is challenging for the same approaches to apply to ordinary images which lack of inherent funny or exaggerated contents. Thus, crafting appealing descriptions for ordinary image demands imaginative efforts to discover or create intriguing connections between words to image contents. To address this gap, we introduce Ap + +pealImage, a large-scale dataset consisting of ordinary images paired with appealing descriptions. AppealImage allows us to define four distinct tasks with quantitative metrics to enable objective evaluation. Subsequently, we propose CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet combines instruction tuning with heuristic active learning, guided by a referee model. Experimental results demonstrate that CharmNet outperforms the state-of-the-art method by $11.4\%$ in generating appealing descriptions. Furthermore, CharmNet delivers impressive performance across various creative applications, including visual storytelling and situational dialogue generation. These results + +highlight CharmNet's potential to enhance social media engagement and to empower strong brand presence in competitive markets. + +# 1. Introduction + +Recent advancements in multi-modal large language models (MLLMs) have greatly expanded their capability in generating multi-modal humor. However, much of the existing research has concentrated on crafting exaggerated, humorous descriptions for internet memes, without attending to the issue of creating appealing descriptions for ordinary images [6, 44]. Unlike memes, which are inherently humorous, ordinary images typically lack superficial engaging features. Producing appealing descriptions for such images requires a sense of creativity and imaginative effort. The goal is to transform ordinary images into captivating content with compelling descriptions, an essential factor for boosting engagement on social media and enabling brand-oriented images to stand out in competitive markets. + +We provide several illustrative examples to demonstrate the differences between appealing descriptions and their conventional counterparts. For instance, in Fig.1 (a), phrases like "trunk-call" and "splashy surprise" employ creative wordplay to impress viewers. In Fig.1 (b), terms such as "man-wich" and "hero of my lunch" use phonetic similarity to capture attention. In Fig. 1 (c), expressions like "sharp roof" and "slice through a sermon" bring in irony through reversal to make the description appealing. In contrast, the corresponding conventional descriptions are plain and factual, simply listing the specific objects or actions in the images without any of these appealing elements. + +Previous research has primarily focused on generating accurate descriptions for ordinary images using fine-tuning or in-context learning [30, 32, 33, 41]. Other studies have emphasized creating humorous captions for images with inherent funny elements by leveraging fine-grained contrasts within the images alongside commonsense, bridging the gap between funny images and conventional descriptions [6, 34, 38, 44]. While funny images readily lend themselves to generating engaging and humorous descriptions, creating appealing descriptions for ordinary images remains a challenge [44]. We further identify two key challenges in identifying the appealing perspectives of objects/contents in ordinary images: (i) The current level of integration between visual and language clues is insufficient, limiting MLLMs' ability to capture the latent interest in ordinary images; (ii) MLLMs lack of supervised training or tuning dedicated to distinguish supervised training to distinguish between appealing and non-appealing descriptions for these types of images. + +For ordinary images, existing methods often struggle to generate sufficiently appealing descriptions. Chan- + +drasekaran et al. [4] focus solely on puns, which offer only a limited form of wordplay in descriptions. Gao et al. [12] primarily produce promotional descriptions that highlight the connection between an object and human, rather than revealing the engaging features of objects. Shuster et al. [35] generate captions from diverse human perspectives, making the content engaging through shared viewpoints rather than by enhancing the intrinsic appeal of the object itself. + +Therefore, in this work, we first construct a large-scale dataset named AppealImage1, comprising ordinary images with appealing descriptions crafted using three strategies: wordplay, phonetic similarity, and reversal, drawing inspiration from meme theory. To thoroughly evaluate and enhance MLLMs' ability to produce appealing descriptions, we define four tasks, including discrimination, selection, ranking, and generation, targeting single-object, multi-object, and global perspectives. Each task is paired with a quantitative metric for objective assessment. After that, we introduce a simple-yet-effective framework named CharmNet, designed to generate appealing descriptions for ordinary images through instruction tuning and heuristic active learning, supervised by a referee model. We conduct a comprehensive evaluation of CharmNet on our AppealImage, where the experiments show that CharmNet consistently surpasses other competing methods, achieving an $11.4\%$ improvement in generating appealing descriptions. Additionally, CharmNet demonstrates robust performance and generalization on various creative tasks, such as visual storytelling and situational dialogue generation. + +# 2. Related Work + +Image Caption Generation. Rotstein et al. [33] fine-tuned a large language model (LLM) to produce natural and coherent image captions. Huang et al. [18] combined vision embeddings with language models to improve caption generation quality. Ma et al. [29] employed a multi-context data generation pipeline using condensed captions and diffusion models to advance image captioning. Dwibedi et al. [11] introduced FlexCap for generating localized captions targeting specific image regions. Hu et al. [15] developed Instruct-Imagen, which integrates style images with captions using multi-modal instructions. Al-Buraihy and Dan [1] enhanced cross-lingual image descriptions through a multi-modal approach. Li et al. [24] introduced EVCap, which employs retrieval-augmented captioning with an external visual-name memory. Chandrasekaran et al. [4] generated witty descriptions for given images by leveraging linguistic puns. Gao et al. [12] proposed the GEM framework for generating promotional multi-modal content. Shuster et al. [35] focused on generating personalized image captions. In contrast, our work focuses on crafting appealing + +
SingleLDes +MultipleGlobalNObjNRObjCategoryNIma
WPPSRe
14.816.019.43.51.033.0%3.4%63.6%74,942
+ +Table 1. AppealImage Statistics. $\mathbf{L_{Des}}$ : Average description length in AppealImage, $\mathbf{N_{Obj}}$ and $\mathbf{N_{RObj}}$ : Number of objects and related objects per image, $\mathbf{N_{Ima}}$ : Total number of images in AppealImage. WP: Wordplay, PS: Phonetic Similarity, Re:Reversal. + +descriptions for images themselves, rather than discussing the relationship between image contents and human. + +Multi-modal Humor Generation. In addition to research on textual humor understanding [5] and generation [7], there has been growing interest in multi-modal humor generation. Wang and Lee [38] developed MemeCraft, a meme generator that leverages LLMs and VLMs to create humorous, socially impactful memes. Zhong et al. [44] introduced the Leap-of-Thought paradigm to enhance LLMs' ability to generate humorous captions. Ruan et al. [34] focused on generating cross-modal sarcasm in multimodal dialogue systems. Chen et al. [6] proposed XMeCap, a novel approach to generating humorous captions across both single-image and multi-image memes. In contrast, our focus is on generating appealing descriptions for ordinary images rather than improving the humorous captions for inherently funny images. + +# 3. AppealImage + +In this section, we introduce a novel dataset named AppealImage and design four tasks to evaluate on this dataset. + +# 3.1. Data Construction + +We first collect 70K images from GQA [19], COCO [27], and ImageNet [9], representing objects and scenes grounded in the physical world. Inspired by the rise of internet memes like "cash me outside" $^{2}$ , we define three types of appealing descriptions, including wordplay, phonetic similarity, and reversal. Wordplay creatively leverages multiple meanings to produce a clever effect, as seen in terms like "pedestrian", which can mean both a person walking and something ordinary. Phonetic similarity plays on homophones or near-homophones, such as "leafing" versus "leaving". Reversal subverts the expected meaning for humor or irony, as in "tired of the street life", which personifies a car as if it were human. To generate detailed appealing descriptions, we define three perspectives, including single-object, multi-object, and global. The single-object description focuses on a specific object, while the multi-object description emphasizes the relationships among multiple objects within the image. The global description provides a comprehensive depiction of the entire image, addressing the + +main scene and overall atmosphere. Next, we employ GPT-4o $^3$ to generate appealing descriptions for the collected images based on the three definitions across these perspectives with a sample shown in Fig. 1 (d). + +After that, GPT-4o is used to assess the quality of descriptions, first filtering out toxic contents related to politics, race, etc. For the remaining descriptions, each is scored on a 1-5 scale across four dimensions, including informativeness, factual accuracy, appeal, and fluency, retaining only those that score at least 4 in each dimension. Descriptions receiving scores below 4 are collected for training the referee, which will be introduced in the next section, and are then regenerated. To guarantee that the output is a scalar value, we employ GLM4 [13] to transform the text, ensuring the final score is appropriately scaled. Then, $10\%$ of these descriptions are randomly selected for validation by three volunteer graduate students. They are first trained by scoring multiple 50-sample batches, with feedback after each round, until their scores reaching a Pearson correlation above 0.9 with us. Next, they are requested to score descriptions from 1 to 5 across the same dimensions. If more than $1\%$ of the reviewed descriptions score below 4 in any dimension, the entire set is re-evaluated by GPT-4o. After re-scoring, a new random $10\%$ sample is selected and subjected to the same manual validation process until fewer than $1\%$ of descriptions score below 4. Ultimately, the GPT-4o and human scores show a correlation of 0.96, validating the final AppealImage dataset. Table 1 summarizes its statistics. + +# 3.2. Task Setup + +Drawing inspiration from Zhong et al. [44], in the discrimination task, an MLLM is presented with an image and two description options and is required to identify which description is more appealing. For the selection task, the MLLM receives an image and three description options, choosing the most appealing one. The ranking task involves the MLLM ranking four given descriptions of an image according to their level of appeal. In the generation task, the MLLM is tasked with generating an appealing description for an ordinary image. Moreover, for the first three tasks, we introduce three levels of difficulty for the description options, distributed equally. Difficulty levels, evaluated by GPT-4o, are determined by the variable $S$ , which quantifies the differentiation between options. These levels are categorized as follows: $[S_{min}, S_{min} + \frac{1}{3}(S_{max} - S_{min})]$ , $[S_{min} + \frac{1}{3}(S_{max} - S_{min})$ , $S_{min} + \frac{2}{3}(S_{max} - S_{min})$ ), and $[S_{min} + \frac{2}{3}(S_{max} - S_{min})$ , $S_{max}]$ , representing low, medium, and high difficulty, respectively, where $S_{min}$ and $S_{max}$ are the minimum and maximum scores. A sample covering all four tasks from single-object, multi-object, and global perspective is shown in Fig. 1(e). + +# 4. Methods + +After constructing AppealImage, we introduce a robust method, CharmNet, for generating appealing descriptions, as illustrated in Fig. 2. + +# 4.1. Instruction Tuning + +We first employ a two-stage instruction tuning strategy inspired by Li et al. [25]. In the first stage, we use a pretrained visual encoder, Qwen-VL [2], to extract image features, keeping the visual encoder initially frozen to maintain feature extraction stability. We then train a QFormer model to translates visual features into textual descriptions by minimizing the cross-entropy loss between predicted and ground-truth appealing descriptions. In the second stage, we refine the image-text alignment. We unfreeze the visual encoder and conduct supervised fine-tuning with LoRA [14] to further optimize QFormer. + +After that, we use multi-task learning, enabling the QFormer to leverage diverse representations from various tasks. We normalize the weights of each task to ensure consistency in scale, as shown in the following formula: + +$$ +\mathcal {L} _ {\mathrm {T}} = \lambda_ {\mathrm {d}} \mathcal {L} _ {\mathrm {d}} + \lambda_ {\mathrm {s}} \mathcal {L} _ {\mathrm {s}} + \lambda_ {\mathrm {r}} \mathcal {L} _ {\mathrm {r}} + \lambda_ {\mathrm {g}} \mathcal {L} _ {\mathrm {g}}, \tag {1} +$$ + +where $\lambda_{\mathrm{d}}, \lambda_{\mathrm{s}}, \lambda_{\mathrm{r}}$ , and $\lambda_{\mathrm{g}}$ represent the normalized weights for the discrimination, selection, ranking, and generation tasks, respectively. These parameters are initially set to be proportional and will be adjusted during training. $\mathcal{L}$ denotes each task's specific cross-entropy loss function. The trained QFormer serves as the precursor to CharmNet and is referred to as pre-CharmNet. + +# 4.2. Referee Construction + +Second, to effectively assess and supervise the appeal level of generated descriptions, we construct a referee model. We first create a warm-up dataset, $D_0$ , containing 1,000 ordinary images and corresponding candidate descriptions, each paired with appeal scores as outlined in the previous section. These scores undergo further validation through manual annotation. Using $D_0$ , we then use LLaVA-1.6 [28] to train a referee model. This referee takes an image and its descriptions as input and outputs scalar scores representing each description's appeal level. + +We employ two types of loss functions in this process. The first is a pairwise ranking loss, which guides the referee to assign higher scores to more appealing descriptions while assigning lower scores to less appealing ones. The loss function is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {R}} = - \frac {1}{\binom {k} {2}} E _ {(I _ {i}, d _ {i, j}, d _ {i, j ^ {\prime}}) \sim D _ {0}} [ \log (\sigma (r _ {\theta} (I _ {i}, d _ {i, j}) - \\ \left. r _ {\theta} \left(I _ {i}, d _ {i, j ^ {\prime}}\right)\right) \bigg ], \tag {2} \\ \end{array} +$$ + +where $d_{i,j}$ and $d_{i,j'}$ are descriptions of image $I_i$ with higher and lower appeal scores, respectively. $r_\theta$ is the scalar score assigned by the referee. $k$ denotes the number of descriptions per image, and $\sigma$ is the sigmoid function, converting score differences into probability values. The second is an alignment loss function, which ensures the referee's scores align closely with human-annotated scores. To create these annotations, three volunteers provide scores from 1 to 5 for each description in $D_0$ , and inter-rater agreement, measured by Krippendorff's Alpha (IRA), is calculated to ensure the reliability of these scores. Descriptions with low agreement (IRA < 0.7) are discarded as Taylor and Watkinson [37] states that an IRA exceeding 0.7 is an excellent agreement. The alignment loss is defined as: + +$$ +\mathcal {L} _ {\mathrm {A}} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left(r _ {\theta} \left(I _ {i}, d _ {i, j}\right) - r _ {\mathrm {h}} \left(I _ {i}, d _ {i, j}\right)\right) ^ {2}, \tag {3} +$$ + +where $N$ is the total number of descriptions for each image, and $r_{\mathrm{h}}(I_i,d_{i,j})$ is the human-annotated score for description $d_{i,j}$ of image $I_{i}$ . Finally, we combine the two loss functions in a weighted sum to create a referee that is both accurate and aligned with human preferences. The combined loss function is defined as: + +$$ +\mathcal {L} = \alpha \mathcal {L} _ {\mathrm {R}} + \beta \mathcal {L} _ {\mathrm {A}} \tag {4} +$$ + +where $\alpha$ and $\beta$ are trainable coefficients that adjust the contributions of the two loss components. + +# 4.3. Heuristic Active Learning + +In this section, we adopt a heuristic active learning strategy ${}^{4}$ to further enhance pre-CharmNet's ability to generate appealing descriptions for ordinary images. This strategy is to identify the more challenging samples (i.e. those with low scores evaluated by the referee). The chosen challenging data points are re-incorporated into the training process, enabling pre-CharmNet to learn which descriptions are perceived as non-appealing. We also validate this observation with comprehensive experiments in Table 8. + +Specifically, we first create a validation dataset, denoted as $D_V$ , by selecting $10\%$ of the data entries from AppealImage. The purpose of $D_V$ is to provide a benchmark for evaluating pre-CharmNet's performance during training iterations. In addition, we construct a hard dataset, $D_1$ , consisting of $X$ images whose descriptions received scores below a predefined threshold $\alpha$ evaluated by the referee. The motivation of $D_1$ is to train pre-CharmNet on samples where the model currently underperforms, enabling it to learn from its weaknesses and refine its ability to generate more appealing descriptions. + +![](images/de7ed2fb34524a9abc9d0610ff38257c5b4292f70c559c10253d3bba76250c6b.jpg) +Figure 2. The framework of CharmNet, including Instruction Tuning which extract image features and refine image-text alignment with supervised fine-tuning, Referee Construction which assess and modify the appeal level of generated descriptions, and Heuristic Active Learning which further enhance CharmNet's ability to generate appealing descriptions for ordinary images. + +After each iteration of training, pre-CharmNet is re-evaluated on $D_V$ to assess whether its performance has improved measured by the referee. If no improvement is observed, the current set of $X$ images in $D_1$ is replaced with a new set of $X$ images that also fall below the threshold $\mu$ . This replacement process ensures that pre-CharmNet is consistently exposed to new challenging examples, avoiding overfitting to specific samples in $D_1$ . The iterative training continues until pre-CharmNet demonstrates consistent improvement on $D_V$ over five consecutive evaluations, indicating that pre-CharmNet has effectively internalized insights from the difficult examples. The update process for constructing $D_1$ is defined as follows: + +$$ +D _ {1} = \left\{\left(I _ {i}, d _ {i, j}\right) \mid r _ {\theta} \left(I _ {i}, d _ {i, j}\right) < \mu \right\}, \tag {5} +$$ + +where $\mu$ is the predefined performance threshold, $I_{i}$ represents an image, $d_{i,j}$ denotes its corresponding description, and $r_{\theta}$ is the score assigned by the referee. Through this optimization process, pre-CharmNet is refined into the final CharmNet, equipped with an enhanced ability to generate more appealing descriptions. + +# 5. Experiments + +In this section, we conduct experiments to evaluate the effectiveness of our proposed CharmNet. + +# 5.1. Experimental Setup + +Our experiments are conducted using PyTorch on four Nvidia A100 GPUs, each with 80GB of memory. We allocate $60\%$ , $30\%$ , and $10\%$ of AppealImage for training, validation, and testing, respectively. The datasets $D_0$ and $D_1$ are derived from the validation set and do not overlap, while $D_V$ denotes the test set. The maximum sequence lengths for input and output sequences are set to 2048 and 512 tokens, respectively. We use the AdamW optimizer with $\beta_{1}$ set to 0.9, $\beta_{2}$ to 0.95, a learning rate of 1e-5, and a weight decay of 1e-1. The batch size is set to 8 during training. LoRA's + +rank is set to 32, with a normalization parameter of 16 and a dropout rate of 0.05. Training is conducted for 10 epochs. Experiments validate Qwen-VL for instruction tuning and LLaVA-1.6 for referee construction as optimal choices. In referee construction, each image is paired with $k = 5$ descriptions, and the predefined evaluation threshold $\mu = 0.2$ , both yielding optimal performance. + +# 5.2. Datasets, Baselines, and Metrics + +We utilize AppealImage and other creative tasks, including visual story creation and situational dialogue generation, to assess CharmNet's performance. For the visual story creation task, we employ the VIST dataset [17], where CharmNet generates appealing narratives based on a sequence of images. CharmNet first creates an appealing description for each individual image, which are then integrated into a cohesive story with the addition of transitional words like "first," "next," and "then", etc. In the situational dialogue generation task, we use the DialogCC dataset [21] to evaluate CharmNet's ability to produce appealing responses within dialogue histories that include images. + +The comparison baselines for generating appealing descriptions are shown in Table 2. For the VIST and DialogCC datasets, CharmNet's performance is evaluated using metrics that measure the distance between human and model outputs in terms of coherence (denoted as $d_{HM}^{C}$ ), visual grounding (denoted as $d_{HM}^{G}$ ), and repetition (denoted as $d_{HM}^{R}$ ), as well as R@1, R@5, R@10, and MRR, respectively, consistent with the corresponding research [21, 36]. + +For evaluation criteria and metrics, the discrimination task is assessed based on whether the MLLM's response matches the ground truth, with top-1 accuracy as the metric. The selection task is evaluated similarly, using top-1 accuracy to determine the MLLM's selected response. For the ranking task, evaluation is based on how closely the MLLM's rankings align with the relative positions of the ground truth, with higher scores awarded for closer matches. The generation task is assessed with automated + +
(Total)DSRG
ACCACCPCBROCMReH
CogVLM-7B [39]57.445.274.436.541.561.437.30.243.1
GPT-4o376.457.783.7------
MiniGPT-4-7B [45]41.128.471.131.136.456.132.90.171.8
LLaVA-1.6-7B [28]44.956.68045.351.470.746.50.364.3
InstructBLIP-7B [8]33.624.166.627.932.651.229.90.111.2
mPLUG-Owl-I [42]50.533.467.633.438.358.634.20.192.4
Otter-I [22]36.324.969.629.534.453.931.20.151.3
Qwen-VL-7B+CLoT [44]60.850.376.940.244.865.442.30.303.6
Qwen-VL-Chat-7B [2]54.643.57338.843.363.440.10.283.4
VisualGLM-6B [10]47.738.374.934.640.36035.90.222.9
XMeCap [6]67.542.77543.549.568.545.50.333.7
CharmNet77.559.583.947.253.373.148.00.384.4
↑1.11.80.21.91.92.41.50.020.1
↑(%)1.43.10.24.23.73.43.25.62.3
+ +Table 2. Comparison of MLLMs and CharmNet across various tasks based on AppealImage. D: Discrimination task, S: Selection task, R: Ranking task, G: Generation task. ACC: Accuracy. PC: Position Consistency. B: BLEU, RO:ROUGE, C: CIDEr, M: METEOR. Re: Referee score. H: Human score. + +
DSRDSRG
SiMuGlSiMuGl
CogVLM-7B [39]65.057.454.748.845.744.3
GPT-4o377.470.069.7---
MiniGPT-4-7B [45]53.047.739.842.239.936.5
LLaVA-1.6-7B [28]65.558.657.453.350.549.5
InstructBLIP-7B [8]47.342.035.037.935.732.4
mPLUG-Owl-I [42]57.050.843.844.341.838.7
Otter-I [22]49.543.637.739.737.334.5
Qwen-VL-7B+CLoT [44]68.461.558.052.049.347.5
Qwen-VL-Chat-7B [2]63.455.452.249.245.944.2
VisualGLM-6B [10]59.552.948.646.043.341.3
XMeCap [6]66.259.859.253.050.549.8
CharmNet79.770.171.159.455.755.5
↑2.30.11.46.15.25.7
↑(%)3.00.12.011.410.311.4
+ +metrics, including BLEU [31], ROUGE [26], CIDEr, and METEOR [3], alongside appealing scores from 1 to 5 in both referee and human evaluations. For human evaluations, three volunteers are recruited to rate 1,000 randomly selected images and their generated descriptions. To ensure the reliability of human ratings, we calculate inter-rater agreement using Krippendorff's Alpha (IRA). Descriptions with ratings that fall below an IRA threshold of 0.7 are considered controversial, and these are discarded and replaced with descriptions from other images. + +# 5.3. Main Results + +We begin by comparing MLLMs and CharmNet across various tasks on AppealImage, as shown in Table 2. Results indicate that the selected baseline MLLMs perform poorly in generating appealing descriptions for ordinary images, while CharmNet consistently outperforms these MLLMs across all evaluated tasks. Given that our dataset was gen + +Table 3. Comparison of MLLMs and CharmNet across three perspectives in AppealImage. DSR: Discrimination, Selection, and Ranking tasks. DSRG: Discrimination, Selection, Ranking, and Generation tasks. Si: Single-object perspective, Mu: Multi-object perspective, Gl: Global perspective. + +
(Global)MaReHTime
Human-written48.60.314.330.5
CharmNet52.70.354.53.2
↑4.10.040.227.3
↑(%)8.412.94.7853.1
+ +Table 4. Comparison of CharmNet-generated and human-written appealing descriptions. Ma: Average of BLEU, ROUGE, CIDEr, and METEOR. Re: Referee score. H: Human score. Time: The average time to generate one description. + +erated by GPT-4o, we exclude GPT-4o from the generation task to avoid inflated scores due to stylistic similarity with the original dataset. Notably, CharmNet achieves an accuracy of $77.5\%$ in the discrimination task, surpassing GPT-4o's accuracy of $76.4\%$ and showing an approximately $130.7\%$ improvement over the lowest-performing MLLM, InstructBLIP-7B, which scored $33.6\%$ . Similarly, CharmNet outperforms GPT-4o and demonstrates a marked improvement over InstructBLIP-7B in the selection task. In the ranking task, both CharmNet and GPT-4o exhibit top performance. In the generation task, CharmNet achieves a human-rated appeal score of 4.4, ranking the highest. + +Next, we conduct a detailed comparison of MLLMs and CharmNet across different task perspectives, as shown in Table 3, once again excluding GPT-4o's performance in the generation task. We calculate the average performance across all MLLMs for the discrimination, selection, and ranking tasks, as well as for all MLLMs excluding GPT-4o. Results show that CharmNet consistently outperforms other MLLMs, with particularly strong results in single-object perspective generation. + +Furthermore, we invite three literary experts to write appealing descriptions for images, aiming to compare human performance with machine-generated descriptions in the task of in generating appealing descriptions. Due to the relatively long time required for human writing (approximately 30 seconds per description), we randomly select 1,000 images from the test set of AppealImage for evaluation. The results are shown in Table 4, demonstrating that CharmNet has surpassed human experts in both effectiveness and efficiency. + +Additionally, we evaluate CharmNet on two creative tasks, visual story creation and situational dialogue generation, with results shown in Table 5 and Table 6, respectively. Baseline results, excluding Qwen-VL-7B+CLoT, are drawn from previous studies [21, 36]. The findings indicate that CharmNet significantly outperforms the baselines, with only a slight improvement over Qwen-VL-7B+CLoT, validating CharmNet's generalizability across a variety of creative tasks. + +Last but not least, while maintaining creativity, we further evaluate CharmNet on non-creative tasks to investigate whether training on the AppealImage dataset leads to the + +
(DialogCC)R@1R@5R@10MRR
Image Only8.2222.633.5216.94
Dialogue Only34.4165.3677.3748.67
Dialogue + Image40.6471.4681.9954.61
Owen-VL-7B+CLoT [44]52.6778.6788.2462.46
CharmNet56.3482.2191.4668.11
↑3.673.543.225.65
↑(%)7.04.53.69.0
+ +Table 5. Performance of CharmNet on the contextual dialogue generation task. + +
(VIST)dCM(-)dGM(-)dRHM(-)dHM(-)
AREL [40]0.1100.0580.0400.208
GLAC Net [20]0.0700.1070.0070.184
TAPM [43]0.0950.0150.0400.150
BLIP-2 [23]0.1900.0500.0940.334
LLaVA [28]0.0480.0040.0470.099
Qwen-VL-7B+CLoT [44]0.0350.0030.0310.069
CharmNet0.0190.0020.0240.045
↓0.0160.0010.0070.024
↓(%)84.250.029.253.3
+ +Table 6. Performance of CharmNet on the visual storytelling task. + +generation of irrelevant information. We employ an image captioning task using 500 randomly selected samples of the MSCOCO dataset, adjusting instructions to guide CharmNet in generating captions that are devoid of creative elements. We design three criteria based on key factors in image captioning as proposed in Huang et al. [16], and conduct manual evaluation with the results shown in Table 7. The manual evaluation show that $89.6\%$ of the generated results correctly described the main objects and their relationships in the image, indicating that hallucination issues are relatively infrequent when CharmNet is used for non-creative tasks. + +The influence of different dataset size in heuristic active learning. We further examine the influence of varying the size of dataset $D_{1}$ in heuristic active learning, as illustrated in Fig. 3(a). The results show a steady increase in performance as the size of $D_{1}$ grows, peaking at $30\%$ . Beyond this threshold, performance begins to decrease, indicating that a $D_{1}$ size of $30\%$ achieves an optimal balance between data scale and quality. + +# 5.4. Ablation Study + +The impact of each module in CharmNet. We first make analysis of the impact of each module in CharmNet as shown in Table 8. All results are averaged for three times to avoid randomness. When the first stage of instruction tuning is removed (denoted as "w/o IT1"), we observe a substantial decline of $9.2\%$ across all metrics. The possible reason may be that directly unfreezing both image and text together for learning mixed their features, preventing the model from learning useful information. Similarly, the removal of heuristic active learning (denoted as "w/o HAL") has a pronounced effect, indicating that HAL fa + +![](images/dad9c75132d4a4663037012fbdd55ac6c13975085e537f40b2bff4d0735837e8.jpg) +Figure 3. (a) Effect of the proportion of difficult data in heuristic active learning. GenerationM, GenerationR, and GenerationH represent machine scores (i.e., BLEU, ROUGE,CIDEr, METEOR), referee scores, and human scores, respectively. (b) Performance across tasks at varying difficulty levels. + +![](images/9cc84b20e7e424f18259c29022017a6b4b4b0be5794b3527c4efb8b0dbcae0e6.jpg) + +cilitates targeted training and improvements through challenging samples. Furthermore, replacing HAL with reinforcement learning (denoted as "HAL to RL", we use direct preference optimization here) does not yield significant performance improvements but has substantially higher training costs. Moreover, substituting multi-task learning with single-task learning (denoted as "MTL to STL"), where only the generation task is used, does not enhance performance in the generation task. This suggests that complementary tasks contribute to CharmNet's understanding of what makes descriptions appealing, thereby improving the quality of generated outputs. + +The function of AppealImage to CharmNet. Next, we use Qwen-VL-7B+CLoT for instruction tuning on AppealImage. While it underperforms on AppealImage, it excels in the Oogiri game, which requires creativity and associative thinking for generating humorous responses. We compare its performance with CharmNet in global perspective generation task, as shown in Table 9, where CharmNet consistently outperforms the AppealImage-tuned Qwen-VL-7B+CLoT. This highlights that the effectiveness of CharmNet is not solely attributable to training on AppealImage but also to its combined strategies like multi-task learning and heuristic active learning. + +The effect of different difficulty levels in each tasks. After that, we investigate the effect of different difficulty levels on CharmNet's performance in the discrimination, selection, and ranking tasks, as shown in Fig. 3(b). Results reveal a near-linear performance decline across all tasks as difficulty increases from easy to hard. Data difficulty has the greatest impact on the selection task, followed by the discrimination task, likely due to the fewer options in discrimination tasks, which simplifies the identification of correct answers. The ranking task is not directly comparable due to its unique evaluation metrics. + +The correlation between referee and human scores for CharmNet. Since ground-truth labels for these tasks are provided by the referee, we employ the Pearson correlation coefficient to ensure alignment between referee and + +
CriteriaPortion (%)
Descriptions of the main objects in the image are incorrect or the relationships between objects in the image are wrong.52.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, but the descriptions are straightforward and unappealing.26.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, and the descriptions are creative and appealing.42.0
+ +Table 7. Three criteria and the corresponding portion based on manual evaluation on 500 samples from the MSCOCO dataset. + +
(Total)D (↓%)S (↓%)R (↓%)GBRCM (↓%)GRe (↓%)GH (↓%)A (↓%)
w/o IT171.8 (7.9)54.0 (10.2)78.0 (7.6)39.0 (13.8)0.29 (31.0)4.0 (10.0)41.2 (9.2)
MTL to STL---43.1 (3.0)0.34 (11.8)4.2 (4.8)-
HAL to RL75.8 (2.2)56.9 (4.6)81.2 (3.3)44.1 (0.7)0.37 (2.7)4.2 (4.8)43.8 (2.7)
w/o HAL66.0 (17.4)48.1 (23.7)73.8 (13.7)33.7 (31.8)0.28 (35.7)3.5 (25.7)37.6 (19.7)
CharmNet77.559.583.944.40.384.445.0
+ +![](images/858db653aab95c0af75fbaf453d71a6b48b1678ed83f37f3d8bb6ba8bb58d000.jpg) +Figure 4. A running case of CharmNet in generating appealing descriptions. + +Table 8. Impact of each module in CharmNet. ${\mathrm{G}}_{BRCM}$ , ${\mathrm{G}}_{Re}$ ,and ${\mathrm{G}}_{H}$ denote machine,referee,and human scores,respectively. A: Average results. + +
(Total)BROCMReH
Qwen-VL-7B+CLoT45.550.271.647.10.344.1
CharmNet47.253.373.148.00.384.4
↑1.73.11.50.90.040.3
↑(%)3.76.22.11.911.87.3
+ +Table 9. Comparison of AppealImage-tuned Qwen-VL-7B+CLoT and CharmNet in the generation task. + +
DSRG
Single-object0.820.740.710.89
Multiple-object0.800.750.700.87
Global0.850.780.730.86
Total0.820.760.710.87
+ +Table 10. Pearson correlation coefficients between referee and human scores for CharmNet. + +human scores for CharmNet. The results are shown in Table 10, indicating a high level of consistency with all values exceeding 0.8 (p<0.05), especially in the generation task. We also test with a larger dataset, but the correlation coefficient does not show significant improvement. This high correlation underscores the reliability of the referee, demonstrating its potential to reduce the need for extensive human involvement while enhancing CharmNet's ability to generate appealing descriptions. + +# 5.5. Case Study + +We present a case study of the appealing descriptions generated by CharmNet, illustrated in Fig. 4. For instance, for the + +single-object description, CharmNet uses terms like “zips” and “exciting trail” to describe the tram, conveying a sense of speed. In the multi-object description, phrases such as “races by”, “vibrant carnival”, “swaying”, and “flashing” depict a static scene as a lively city celebration. Finally, in the global description, words like “leading a parade”, “bringing life and color”, and “effortless charm” enhance the image’s atmosphere, making it appear more dynamic and engaging. We also observe some less appealing descriptions, such as “a boy walking with a man in a crowd …”, which lacks vivid detail and emotional depth. + +# 6. Conclusions + +How to craft appealing descriptions by MLLMs for ordinary images beyond those internet memes is quite underinvestigated before. To bridge this gap, we present AppealImage, a large-scale dataset containing ordinary images paired with appealing descriptions. We define four tasks, each paired with quantitative metrics, to enable thorough evaluation. Upon AppealImage, we introduce CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet leverages instruction tuning and heuristic active learning, guided by a referee model. Our experiments show CharmNet outperforms the state-of-the-art method by $11.4\%$ in generating appealing descriptions and has impressive results across related creative tasks. Future work will aim to broaden CharmNet's applicability and customize it for personalized contents. + +# Acknowledgements + +This work is supported by Ant Group Research Intern Program. + +# References + +[1] Emran Al-Buraihy and W Dan. Enhancing crosslingual image description: A multimodal approach for semantic relevance and stylistic alignment. Comput. Mater. Contin, 2, 2024. 2 +[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 4, 6 +[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6 +[4] Arjun Chandrasekaran, Devi Parikh, and Mohit Bansal. Funny captions: Witty wordplay in image descriptions. arXiv preprint arXiv:1704.08224, 2017. 2 +[5] Yuyan Chen, Zhixu Li, Jiaqing Liang, Yanghua Xiao, Bang Liu, and Yunwen Chen. Can pre-trained language models understand chinese humor? 2023. 3 +[6] Yuyan Chen, Songzhou Yan, Zhihong Zhu, Zhixu Li, and Yanghua Xiao. Xmecap: Meme caption generation with sub-image adaptability. arXiv preprint arXiv:2407.17152, 2024. 2, 3, 6 +[7] Yuyan Chen, Yichen Yuan, Panjun Liu, Dayiheng Liu, Qinghao Guan, Mengfei Guo, Haiming Peng, Bang Liu, Zhixu Li, and Yanghua Xiao. Talk funny! a largescale humor response dataset with chain-of-humor interpretation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17826-17834, 2024. 3 +[8] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3 +[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 6 + +[11] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. arXiv preprint arXiv:2403.12026, 2024. 2 +[12] Chongyang Gao, Yiren Jian, Natalia Denisenko, Soroush Vosoughi, and VS Subrahmanian. Gem: generating engaging multimodal content. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 7654–7662, 2024. 2 +[13] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Jingyu Sun, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 3 +[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 4 +[15] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4754-4763, 2024. 2 +[16] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4634-4643, 2019. 7 +[17] Ting-Hao Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, et al. Visual storytelling. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: Human language technologies, pages 1233-1239, 2016. 5 +[18] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings + +of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 2 +[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3 +[20] Taehyeong Kim, Min-Oh Heo, Seonil Son, Kyoung-Wha Park, and Byoung-Tak Zhang. Glac net: Glocal attention cascading networks for multi-image cued story generation. arXiv preprint arXiv:1805.10973, 2018. 7 +[21] Young-Jun Lee, Byungsoo Ko, Han-Gyu Kim, and Ho-Jin Choi. Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119, 2022. 5, 6 +[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multimodal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 6 +[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models, 2023. 7 +[24] Jiaxuan Li, Duc Minh Vo, Akihiro Sugimoto, and Hideki Nakayama. Evcap: Retrieval-augmented image captioning with external visual-name memory for open-world comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2024. 2 +[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 4 +[26] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, 2004. Association for Computational Linguistics. 6 +[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 3 +[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 4, 6, 7 +[29] Feipeng Ma, Yizhou Zhou, Fengyun Rao, Yueyi Zhang, and Xiaoyan Sun. Image captioning with multi-context synthetic data. In Proceedings of the + +AAAI Conference on Artificial Intelligence, pages 4089-4097, 2024. 2 +[30] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2 +[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6 +[32] Suzanne Petryk, Spencer Whitehead, Joseph E Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus Rohrbach. Simple token-level confidence improves caption correctness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5742-5752, 2024. 2 +[33] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 2 +[34] Jie Ruan, Yue Wu, Xiaojun Wan, and Yuesheng Zhu. Describe images in a boring way: Towards cross-modal sarcasm generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5701-5710, 2024. 2, 3 +[35] Kurt Shuster, Samuel Humeau, Hexiang Hu, Antoine Bordes, and Jason Weston. Engaging image captioning via personality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12516-12526, 2019. 2 +[36] Aditya K Surikuchi, Raquel Fernandez, and Sandro Pezzelle. Not (yet) the whole story: Evaluating visual storytelling requires more than measuring coherence, grounding, and repetition. arXiv preprint arXiv:2407.04559, 2024. 5, 6 +[37] Joel Taylor and David Watkinson. Indexing reliability for condition survey data. The Conservator, 30(1):49-62, 2007. 4 +[38] Han Wang and Roy Ka-Wei Lee. Memecraft: Contextual and stance-driven multimodal meme generation. In Proceedings of the ACM on Web Conference 2024, pages 4642–4652, 2024. 2, 3 +[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6 + +[40] Xin Wang, Wenhu Chen, Yuan-Fang Wang, and William Yang Wang. No metrics are perfect: Adversarial reward learning for visual storytelling. arXiv preprint arXiv:1804.09160, 2018.7 +[41] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2 +[42] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 6 +[43] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jongseok Kim, and Gunhee Kim. Transitional adaptation of pretrained models for visual storytelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12658-12668, 2021. 7 +[44] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 2, 3, 6, 7 +[45] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 6 \ No newline at end of file diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/images.zip b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c527d7725441a570b5fde5a72b043047f5e4807f --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cd7f5a516c98661c8a4c2a9fc06910125e974c5720ccc82a7e70c82e03f461b1 +size 541046 diff --git a/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/layout.json b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2ea1a4975346b710eeb8eac7343886b56b7299ab --- /dev/null +++ b/2025/Engage for All_ Making Ordinary Image Descriptions Appealing Again!/layout.json @@ -0,0 +1,9010 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 526, + 121 + ], + "type": "text", + "content": "Engage for All: Making Ordinary Image Descriptions Appealing Again!" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "spans": [ + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": "Yuyan Chen" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Yifan Jiang" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Li Zhou" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Jinghan Cao" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Yu Guan" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Ming Yang" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "text", + "content": ", Qingpei Guo" + }, + { + "bbox": [ + 67, + 141, + 545, + 157 + ], + "type": "inline_equation", + "content": "^{5,6\\dagger}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 171, + 157, + 440, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 171, + 157, + 440, + 171 + ], + "spans": [ + { + "bbox": [ + 171, + 157, + 440, + 171 + ], + "type": "text", + "content": "1Cornell University, 2University of Southern California," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "spans": [ + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "type": "text", + "content": "Morgan Stanley Service Canada, " + }, + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 144, + 171, + 466, + 186 + ], + "type": "text", + "content": "San Francisco State University," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "spans": [ + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "type": "text", + "content": "University of Warwick, " + }, + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 216, + 185, + 395, + 200 + ], + "type": "text", + "content": "Ant Group" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 133, + 200, + 476, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 200, + 476, + 213 + ], + "spans": [ + { + "bbox": [ + 133, + 200, + 476, + 213 + ], + "type": "text", + "content": "yolandachen0313@gmail.com, yifjia@isi.edu, lzhou1068@gmail.com," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 88, + 213, + 521, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 213, + 521, + 228 + ], + "spans": [ + { + "bbox": [ + 88, + 213, + 521, + 228 + ], + "type": "text", + "content": "jcao3@alumni.sfsu.edu, yu.guan@warwick.ac.uk, {m.yang, qingpei.gqp} @antgroup.com" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 89, + 237, + 332, + 345 + ], + "blocks": [ + { + "bbox": [ + 89, + 237, + 332, + 345 + ], + "lines": [ + { + "bbox": [ + 89, + 237, + 332, + 345 + ], + "spans": [ + { + "bbox": [ + 89, + 237, + 332, + 345 + ], + "type": "image", + "image_path": "725b0b6e4590d8a73dc00f0fd2a727e430d513cb42c0405c4fb6736498134f62.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 346, + 202, + 353 + ], + "lines": [ + { + "bbox": [ + 194, + 346, + 202, + 353 + ], + "spans": [ + { + "bbox": [ + 194, + 346, + 202, + 353 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 280, + 346, + 288, + 352 + ], + "lines": [ + { + "bbox": [ + 280, + 346, + 288, + 352 + ], + "spans": [ + { + "bbox": [ + 280, + 346, + 288, + 352 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 237, + 405, + 345 + ], + "blocks": [ + { + "bbox": [ + 334, + 237, + 405, + 345 + ], + "lines": [ + { + "bbox": [ + 334, + 237, + 405, + 345 + ], + "spans": [ + { + "bbox": [ + 334, + 237, + 405, + 345 + ], + "type": "image", + "image_path": "fc41a6ce9281e528e215569e232081d81be0e3fbe15bfeb569b1cc94b0868f2d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 408, + 237, + 477, + 345 + ], + "blocks": [ + { + "bbox": [ + 408, + 237, + 477, + 345 + ], + "lines": [ + { + "bbox": [ + 408, + 237, + 477, + 345 + ], + "spans": [ + { + "bbox": [ + 408, + 237, + 477, + 345 + ], + "type": "image", + "image_path": "f00537d9fe38944fe3e2eb8972a9c5e38d0821876e9060cac6b6d46d4cdacb7e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 478, + 237, + 549, + 345 + ], + "blocks": [ + { + "bbox": [ + 478, + 237, + 549, + 345 + ], + "lines": [ + { + "bbox": [ + 478, + 237, + 549, + 345 + ], + "spans": [ + { + "bbox": [ + 478, + 237, + 549, + 345 + ], + "type": "image", + "image_path": "0718fc342ef018db28f0919097d7427eb6a17d1dded5cc7099a182117b5aaa3e.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 59, + 354, + 332, + 466 + ], + "blocks": [ + { + "bbox": [ + 59, + 354, + 332, + 466 + ], + "lines": [ + { + "bbox": [ + 59, + 354, + 332, + 466 + ], + "spans": [ + { + "bbox": [ + 59, + 354, + 332, + 466 + ], + "type": "image", + "image_path": "392d130013441ceb1476320e40bd8c1ec5857014b0a16d8ff273c24422312720.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 467, + 206, + 474 + ], + "lines": [ + { + "bbox": [ + 198, + 467, + 206, + 474 + ], + "spans": [ + { + "bbox": [ + 198, + 467, + 206, + 474 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 334, + 348, + 405, + 420 + ], + "blocks": [ + { + "bbox": [ + 334, + 348, + 405, + 420 + ], + "lines": [ + { + "bbox": [ + 334, + 348, + 405, + 420 + ], + "spans": [ + { + "bbox": [ + 334, + 348, + 405, + 420 + ], + "type": "image", + "image_path": "0fe1aa93994b15778cba576facf1e8c0f528d4635359c7393420d20e32766c4b.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 408, + 348, + 464, + 388 + ], + "blocks": [ + { + "bbox": [ + 408, + 348, + 464, + 388 + ], + "lines": [ + { + "bbox": [ + 408, + 348, + 464, + 388 + ], + "spans": [ + { + "bbox": [ + 408, + 348, + 464, + 388 + ], + "type": "image", + "image_path": "84b566dd215d0710233ae04071977e5769bf23b7993cb1d00a185b86fe4f8da2.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 466, + 348, + 549, + 420 + ], + "blocks": [ + { + "bbox": [ + 466, + 348, + 549, + 420 + ], + "lines": [ + { + "bbox": [ + 466, + 348, + 549, + 420 + ], + "spans": [ + { + "bbox": [ + 466, + 348, + 549, + 420 + ], + "type": "image", + "image_path": "d09bbe516a3b9f4eb14f84e13ad41361adf84137f5bd15e0cc17c422c4a82226.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 422, + 456, + 427 + ], + "lines": [ + { + "bbox": [ + 449, + 422, + 456, + 427 + ], + "spans": [ + { + "bbox": [ + 449, + 422, + 456, + 427 + ], + "type": "text", + "content": "(e)" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 334, + 430, + 345, + 437 + ], + "blocks": [ + { + "bbox": [ + 334, + 430, + 345, + 437 + ], + "lines": [ + { + "bbox": [ + 334, + 430, + 345, + 437 + ], + "spans": [ + { + "bbox": [ + 334, + 430, + 345, + 437 + ], + "type": "image", + "image_path": "f9eddc224e6aabed8efcb3ebd824c499df1c9340e8b8301b5a87f356b9356fb0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 432, + 370, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 432, + 370, + 437 + ], + "spans": [ + { + "bbox": [ + 347, + 432, + 370, + 437 + ], + "type": "text", + "content": "CharmNet:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 334, + 438, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 438, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 334, + 438, + 553, + 445 + ], + "type": "text", + "content": "The tram is leading a parade, bringing life and color to the city as it glides by with effortless charm!" + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 334, + 448, + 345, + 456 + ], + "blocks": [ + { + "bbox": [ + 334, + 448, + 345, + 456 + ], + "lines": [ + { + "bbox": [ + 334, + 448, + 345, + 456 + ], + "spans": [ + { + "bbox": [ + 334, + 448, + 345, + 456 + ], + "type": "image", + "image_path": "1d641931a72eb9503e920177c0ea7b90cf536e115e4be74c049311bce9ffae59.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 485, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 54, + 485, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. (a) to (c) present Conventional descriptions (highlighted in red) and appealing descriptions (highlighted in blue) for three ordinary images. (d) presents a sample from AppealBench with complete object descriptions. (e) presents four tasks applied to a sample from AppealBench across single-object, multi-object, and global perspectives." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 347, + 449, + 377, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 449, + 377, + 455 + ], + "spans": [ + { + "bbox": [ + 347, + 449, + 377, + 455 + ], + "type": "text", + "content": "Ground Truth:" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 347, + 456, + 520, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 456, + 520, + 462 + ], + "spans": [ + { + "bbox": [ + 347, + 456, + 520, + 462 + ], + "type": "text", + "content": "This tram must be on a seafood diet because it's definitely a shell of its former self!" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "spans": [ + { + "bbox": [ + 152, + 545, + 200, + 558 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 54, + 573, + 297, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 573, + 297, + 681 + ], + "spans": [ + { + "bbox": [ + 54, + 573, + 297, + 681 + ], + "type": "text", + "content": "In recent years, multi-modal large language models (MLLMs) have been successfully adopted to generate humorous and engaging descriptions for internet memes. While, it is challenging for the same approaches to apply to ordinary images which lack of inherent funny or exaggerated contents. Thus, crafting appealing descriptions for ordinary image demands imaginative efforts to discover or create intriguing connections between words to image contents. To address this gap, we introduce Ap" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 547, + 555, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 702 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 702 + ], + "type": "text", + "content": "pealImage, a large-scale dataset consisting of ordinary images paired with appealing descriptions. AppealImage allows us to define four distinct tasks with quantitative metrics to enable objective evaluation. Subsequently, we propose CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet combines instruction tuning with heuristic active learning, guided by a referee model. Experimental results demonstrate that CharmNet outperforms the state-of-the-art method by " + }, + { + "bbox": [ + 313, + 547, + 555, + 702 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 313, + 547, + 555, + 702 + ], + "type": "text", + "content": " in generating appealing descriptions. Furthermore, CharmNet delivers impressive performance across various creative applications, including visual storytelling and situational dialogue generation. These results" + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 15 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "spans": [ + { + "bbox": [ + 199, + 13, + 442, + 24 + ], + "type": "text", + "content": "Except for this watermark, it is identical to the accepted version;" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "spans": [ + { + "bbox": [ + 179, + 25, + 462, + 37 + ], + "type": "text", + "content": "the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 693, + 228, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 228, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 228, + 703 + ], + "type": "text", + "content": "* Work done during an internship at Ant Group." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 67, + 703, + 213, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 213, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 213, + 712 + ], + "type": "text", + "content": "† Qingpei Guo is the corresponding author." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19342" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "highlight CharmNet's potential to enhance social media engagement and to empower strong brand presence in competitive markets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 319 + ], + "type": "text", + "content": "Recent advancements in multi-modal large language models (MLLMs) have greatly expanded their capability in generating multi-modal humor. However, much of the existing research has concentrated on crafting exaggerated, humorous descriptions for internet memes, without attending to the issue of creating appealing descriptions for ordinary images [6, 44]. Unlike memes, which are inherently humorous, ordinary images typically lack superficial engaging features. Producing appealing descriptions for such images requires a sense of creativity and imaginative effort. The goal is to transform ordinary images into captivating content with compelling descriptions, an essential factor for boosting engagement on social media and enabling brand-oriented images to stand out in competitive markets." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 462 + ], + "type": "text", + "content": "We provide several illustrative examples to demonstrate the differences between appealing descriptions and their conventional counterparts. For instance, in Fig.1 (a), phrases like \"trunk-call\" and \"splashy surprise\" employ creative wordplay to impress viewers. In Fig.1 (b), terms such as \"man-wich\" and \"hero of my lunch\" use phonetic similarity to capture attention. In Fig. 1 (c), expressions like \"sharp roof\" and \"slice through a sermon\" bring in irony through reversal to make the description appealing. In contrast, the corresponding conventional descriptions are plain and factual, simply listing the specific objects or actions in the images without any of these appealing elements." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 689 + ], + "type": "text", + "content": "Previous research has primarily focused on generating accurate descriptions for ordinary images using fine-tuning or in-context learning [30, 32, 33, 41]. Other studies have emphasized creating humorous captions for images with inherent funny elements by leveraging fine-grained contrasts within the images alongside commonsense, bridging the gap between funny images and conventional descriptions [6, 34, 38, 44]. While funny images readily lend themselves to generating engaging and humorous descriptions, creating appealing descriptions for ordinary images remains a challenge [44]. We further identify two key challenges in identifying the appealing perspectives of objects/contents in ordinary images: (i) The current level of integration between visual and language clues is insufficient, limiting MLLMs' ability to capture the latent interest in ordinary images; (ii) MLLMs lack of supervised training or tuning dedicated to distinguish supervised training to distinguish between appealing and non-appealing descriptions for these types of images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "For ordinary images, existing methods often struggle to generate sufficiently appealing descriptions. Chan-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "drasekaran et al. [4] focus solely on puns, which offer only a limited form of wordplay in descriptions. Gao et al. [12] primarily produce promotional descriptions that highlight the connection between an object and human, rather than revealing the engaging features of objects. Shuster et al. [35] generate captions from diverse human perspectives, making the content engaging through shared viewpoints rather than by enhancing the intrinsic appeal of the object itself." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 168, + 553, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 553, + 419 + ], + "type": "text", + "content": "Therefore, in this work, we first construct a large-scale dataset named AppealImage1, comprising ordinary images with appealing descriptions crafted using three strategies: wordplay, phonetic similarity, and reversal, drawing inspiration from meme theory. To thoroughly evaluate and enhance MLLMs' ability to produce appealing descriptions, we define four tasks, including discrimination, selection, ranking, and generation, targeting single-object, multi-object, and global perspectives. Each task is paired with a quantitative metric for objective assessment. After that, we introduce a simple-yet-effective framework named CharmNet, designed to generate appealing descriptions for ordinary images through instruction tuning and heuristic active learning, supervised by a referee model. We conduct a comprehensive evaluation of CharmNet on our AppealImage, where the experiments show that CharmNet consistently surpasses other competing methods, achieving an " + }, + { + "bbox": [ + 313, + 168, + 553, + 419 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 313, + 168, + 553, + 419 + ], + "type": "text", + "content": " improvement in generating appealing descriptions. Additionally, CharmNet demonstrates robust performance and generalization on various creative tasks, such as visual storytelling and situational dialogue generation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 400, + 445 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 553, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 553, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 553, + 693 + ], + "type": "text", + "content": "Image Caption Generation. Rotstein et al. [33] fine-tuned a large language model (LLM) to produce natural and coherent image captions. Huang et al. [18] combined vision embeddings with language models to improve caption generation quality. Ma et al. [29] employed a multi-context data generation pipeline using condensed captions and diffusion models to advance image captioning. Dwibedi et al. [11] introduced FlexCap for generating localized captions targeting specific image regions. Hu et al. [15] developed Instruct-Imagen, which integrates style images with captions using multi-modal instructions. Al-Buraihy and Dan [1] enhanced cross-lingual image descriptions through a multi-modal approach. Li et al. [24] introduced EVCap, which employs retrieval-augmented captioning with an external visual-name memory. Chandrasekaran et al. [4] generated witty descriptions for given images by leveraging linguistic puns. Gao et al. [12] proposed the GEM framework for generating promotional multi-modal content. Shuster et al. [35] focused on generating personalized image captions. In contrast, our work focuses on crafting appealing" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 703, + 511, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 703, + 511, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 703, + 511, + 713 + ], + "type": "text", + "content": "1https://github.com/Yukyin/AppealImage." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19343" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 293, + 104 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 104 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 104 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 104 + ], + "type": "table", + "html": "
SingleLDes\nMultipleGlobalNObjNRObjCategoryNIma
WPPSRe
14.816.019.43.51.033.0%3.4%63.6%74,942
", + "image_path": "d924b733c3c171f36b0427f7da4dfa7765b99da92e4a7b5e0fc8665fe1d9a3bf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "text", + "content": "Table 1. AppealImage Statistics. " + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{L_{Des}}" + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "text", + "content": ": Average description length in AppealImage, " + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{N_{Obj}}" + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{N_{RObj}}" + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "text", + "content": ": Number of objects and related objects per image, " + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "inline_equation", + "content": "\\mathbf{N_{Ima}}" + }, + { + "bbox": [ + 55, + 114, + 295, + 160 + ], + "type": "text", + "content": ": Total number of images in AppealImage. WP: Wordplay, PS: Phonetic Similarity, Re:Reversal." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 180, + 295, + 205 + ], + "type": "text", + "content": "descriptions for images themselves, rather than discussing the relationship between image contents and human." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 384 + ], + "type": "text", + "content": "Multi-modal Humor Generation. In addition to research on textual humor understanding [5] and generation [7], there has been growing interest in multi-modal humor generation. Wang and Lee [38] developed MemeCraft, a meme generator that leverages LLMs and VLMs to create humorous, socially impactful memes. Zhong et al. [44] introduced the Leap-of-Thought paradigm to enhance LLMs' ability to generate humorous captions. Ruan et al. [34] focused on generating cross-modal sarcasm in multimodal dialogue systems. Chen et al. [6] proposed XMeCap, a novel approach to generating humorous captions across both single-image and multi-image memes. In contrast, our focus is on generating appealing descriptions for ordinary images rather than improving the humorous captions for inherently funny images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 396, + 140, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 140, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 140, + 410 + ], + "type": "text", + "content": "3. AppealImage" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 416, + 295, + 441 + ], + "type": "text", + "content": "In this section, we introduce a novel dataset named AppealImage and design four tasks to evaluate on this dataset." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 165, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 165, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 165, + 460 + ], + "type": "text", + "content": "3.1. Data Construction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 466, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 466, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 466, + 295, + 694 + ], + "type": "text", + "content": "We first collect 70K images from GQA [19], COCO [27], and ImageNet [9], representing objects and scenes grounded in the physical world. Inspired by the rise of internet memes like \"cash me outside\" " + }, + { + "bbox": [ + 55, + 466, + 295, + 694 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 466, + 295, + 694 + ], + "type": "text", + "content": ", we define three types of appealing descriptions, including wordplay, phonetic similarity, and reversal. Wordplay creatively leverages multiple meanings to produce a clever effect, as seen in terms like \"pedestrian\", which can mean both a person walking and something ordinary. Phonetic similarity plays on homophones or near-homophones, such as \"leafing\" versus \"leaving\". Reversal subverts the expected meaning for humor or irony, as in \"tired of the street life\", which personifies a car as if it were human. To generate detailed appealing descriptions, we define three perspectives, including single-object, multi-object, and global. The single-object description focuses on a specific object, while the multi-object description emphasizes the relationships among multiple objects within the image. The global description provides a comprehensive depiction of the entire image, addressing the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "main scene and overall atmosphere. Next, we employ GPT-4o " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " to generate appealing descriptions for the collected images based on the three definitions across these perspectives with a sample shown in Fig. 1 (d)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "content": "After that, GPT-4o is used to assess the quality of descriptions, first filtering out toxic contents related to politics, race, etc. For the remaining descriptions, each is scored on a 1-5 scale across four dimensions, including informativeness, factual accuracy, appeal, and fluency, retaining only those that score at least 4 in each dimension. Descriptions receiving scores below 4 are collected for training the referee, which will be introduced in the next section, and are then regenerated. To guarantee that the output is a scalar value, we employ GLM4 [13] to transform the text, ensuring the final score is appropriately scaled. Then, " + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "content": " of these descriptions are randomly selected for validation by three volunteer graduate students. They are first trained by scoring multiple 50-sample batches, with feedback after each round, until their scores reaching a Pearson correlation above 0.9 with us. Next, they are requested to score descriptions from 1 to 5 across the same dimensions. If more than " + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "content": " of the reviewed descriptions score below 4 in any dimension, the entire set is re-evaluated by GPT-4o. After re-scoring, a new random " + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "content": " sample is selected and subjected to the same manual validation process until fewer than " + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 121, + 555, + 418 + ], + "type": "text", + "content": " of descriptions score below 4. Ultimately, the GPT-4o and human scores show a correlation of 0.96, validating the final AppealImage dataset. Table 1 summarizes its statistics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 426, + 389, + 440 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 426, + 389, + 440 + ], + "spans": [ + { + "bbox": [ + 314, + 426, + 389, + 440 + ], + "type": "text", + "content": "3.2. Task Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": "Drawing inspiration from Zhong et al. [44], in the discrimination task, an MLLM is presented with an image and two description options and is required to identify which description is more appealing. For the selection task, the MLLM receives an image and three description options, choosing the most appealing one. The ranking task involves the MLLM ranking four given descriptions of an image according to their level of appeal. In the generation task, the MLLM is tasked with generating an appealing description for an ordinary image. Moreover, for the first three tasks, we introduce three levels of difficulty for the description options, distributed equally. Difficulty levels, evaluated by GPT-4o, are determined by the variable " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": ", which quantifies the differentiation between options. These levels are categorized as follows: " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "[S_{min}, S_{min} + \\frac{1}{3}(S_{max} - S_{min})]" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "[S_{min} + \\frac{1}{3}(S_{max} - S_{min})" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{min} + \\frac{2}{3}(S_{max} - S_{min})" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "[S_{min} + \\frac{2}{3}(S_{max} - S_{min})" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{max}]" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": ", representing low, medium, and high difficulty, respectively, where " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{min}" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "inline_equation", + "content": "S_{max}" + }, + { + "bbox": [ + 313, + 444, + 555, + 696 + ], + "type": "text", + "content": " are the minimum and maximum scores. A sample covering all four tasks from single-object, multi-object, and global perspective is shown in Fig. 1(e)." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 250, + 713 + ], + "type": "text", + "content": "2https://www.dictionary.com/e/memes/cash-me-outside/" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 394, + 713 + ], + "type": "text", + "content": "3https://chatgpt.com" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19344" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 116, + 83 + ], + "type": "text", + "content": "4. Methods" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 128 + ], + "type": "text", + "content": "After constructing AppealImage, we introduce a robust method, CharmNet, for generating appealing descriptions, as illustrated in Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 137, + 167, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 137, + 167, + 149 + ], + "spans": [ + { + "bbox": [ + 55, + 137, + 167, + 149 + ], + "type": "text", + "content": "4.1. Instruction Tuning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 155, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 155, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 296, + 286 + ], + "type": "text", + "content": "We first employ a two-stage instruction tuning strategy inspired by Li et al. [25]. In the first stage, we use a pretrained visual encoder, Qwen-VL [2], to extract image features, keeping the visual encoder initially frozen to maintain feature extraction stability. We then train a QFormer model to translates visual features into textual descriptions by minimizing the cross-entropy loss between predicted and ground-truth appealing descriptions. In the second stage, we refine the image-text alignment. We unfreeze the visual encoder and conduct supervised fine-tuning with LoRA [14] to further optimize QFormer." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 287, + 296, + 335 + ], + "type": "text", + "content": "After that, we use multi-task learning, enabling the QFormer to leverage diverse representations from various tasks. We normalize the weights of each task to ensure consistency in scale, as shown in the following formula:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 346, + 295, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 346, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 102, + 346, + 295, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T}} = \\lambda_ {\\mathrm {d}} \\mathcal {L} _ {\\mathrm {d}} + \\lambda_ {\\mathrm {s}} \\mathcal {L} _ {\\mathrm {s}} + \\lambda_ {\\mathrm {r}} \\mathcal {L} _ {\\mathrm {r}} + \\lambda_ {\\mathrm {g}} \\mathcal {L} _ {\\mathrm {g}}, \\tag {1}", + "image_path": "668cf485feb18cc61f4f79ab650855daba75a2a6a19d803690c28256f33a153e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{d}}, \\lambda_{\\mathrm{s}}, \\lambda_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{g}}" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " represent the normalized weights for the discrimination, selection, ranking, and generation tasks, respectively. These parameters are initially set to be proportional and will be adjusted during training. " + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 369, + 296, + 453 + ], + "type": "text", + "content": " denotes each task's specific cross-entropy loss function. The trained QFormer serves as the precursor to CharmNet and is referred to as pre-CharmNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 178, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 178, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 178, + 474 + ], + "type": "text", + "content": "4.2. Referee Construction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "text", + "content": "Second, to effectively assess and supervise the appeal level of generated descriptions, we construct a referee model. We first create a warm-up dataset, " + }, + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "text", + "content": ", containing 1,000 ordinary images and corresponding candidate descriptions, each paired with appeal scores as outlined in the previous section. These scores undergo further validation through manual annotation. Using " + }, + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 55, + 480, + 296, + 600 + ], + "type": "text", + "content": ", we then use LLaVA-1.6 [28] to train a referee model. This referee takes an image and its descriptions as input and outputs scalar scores representing each description's appeal level." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 601, + 296, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 660 + ], + "type": "text", + "content": "We employ two types of loss functions in this process. The first is a pairwise ranking loss, which guides the referee to assign higher scores to more appealing descriptions while assigning lower scores to less appealing ones. The loss function is defined as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 72, + 668, + 295, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 668, + 295, + 711 + ], + "spans": [ + { + "bbox": [ + 72, + 668, + 295, + 711 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {R}} = - \\frac {1}{\\binom {k} {2}} E _ {(I _ {i}, d _ {i, j}, d _ {i, j ^ {\\prime}}) \\sim D _ {0}} [ \\log (\\sigma (r _ {\\theta} (I _ {i}, d _ {i, j}) - \\\\ \\left. r _ {\\theta} \\left(I _ {i}, d _ {i, j ^ {\\prime}}\\right)\\right) \\bigg ], \\tag {2} \\\\ \\end{array}", + "image_path": "68ecefd045ad98fb59f9fa744eabd4c18fc481cb6e4307cf2f0d7406a7d4ae8d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "d_{i,j'}" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " are descriptions of image " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " with higher and lower appeal scores, respectively. " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "r_\\theta" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is the scalar score assigned by the referee. " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " denotes the number of descriptions per image, and " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": " is the sigmoid function, converting score differences into probability values. The second is an alignment loss function, which ensures the referee's scores align closely with human-annotated scores. To create these annotations, three volunteers provide scores from 1 to 5 for each description in " + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": ", and inter-rater agreement, measured by Krippendorff's Alpha (IRA), is calculated to ensure the reliability of these scores. Descriptions with low agreement (IRA < 0.7) are discarded as Taylor and Watkinson [37] states that an IRA exceeding 0.7 is an excellent agreement. The alignment loss is defined as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 250, + 555, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 250, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 349, + 250, + 555, + 283 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {A}} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left(r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) - r _ {\\mathrm {h}} \\left(I _ {i}, d _ {i, j}\\right)\\right) ^ {2}, \\tag {3}", + "image_path": "5e12dc1251c0f7f757ea0d97f0b5b9b87bf2d0541e4cbc99f76011ff3421be9e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "content": " is the total number of descriptions for each image, and " + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "inline_equation", + "content": "r_{\\mathrm{h}}(I_i,d_{i,j})" + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "content": " is the human-annotated score for description " + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "content": " of image " + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 293, + 555, + 364 + ], + "type": "text", + "content": ". Finally, we combine the two loss functions in a weighted sum to create a referee that is both accurate and aligned with human preferences. The combined loss function is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 398, + 376, + 555, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 376, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 398, + 376, + 555, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\alpha \\mathcal {L} _ {\\mathrm {R}} + \\beta \\mathcal {L} _ {\\mathrm {A}} \\tag {4}", + "image_path": "7405f4b885eb83efe92aa1134b70c2aa6e6f3121dbeb3a9857ca509ed5b530a5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 399, + 554, + 424 + ], + "type": "text", + "content": " are trainable coefficients that adjust the contributions of the two loss components." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 434, + 459, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 459, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 459, + 447 + ], + "type": "text", + "content": "4.3. Heuristic Active Learning" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 452, + 554, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 554, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 554, + 559 + ], + "type": "text", + "content": "In this section, we adopt a heuristic active learning strategy " + }, + { + "bbox": [ + 313, + 452, + 554, + 559 + ], + "type": "inline_equation", + "content": "{}^{4}" + }, + { + "bbox": [ + 313, + 452, + 554, + 559 + ], + "type": "text", + "content": " to further enhance pre-CharmNet's ability to generate appealing descriptions for ordinary images. This strategy is to identify the more challenging samples (i.e. those with low scores evaluated by the referee). The chosen challenging data points are re-incorporated into the training process, enabling pre-CharmNet to learn which descriptions are perceived as non-appealing. We also validate this observation with comprehensive experiments in Table 8." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "spans": [ + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": "Specifically, we first create a validation dataset, denoted as " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "D_V" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": ", by selecting " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": " of the data entries from AppealImage. The purpose of " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "D_V" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": " is to provide a benchmark for evaluating pre-CharmNet's performance during training iterations. In addition, we construct a hard dataset, " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": ", consisting of " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": " images whose descriptions received scores below a predefined threshold " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": " evaluated by the referee. The motivation of " + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 313, + 560, + 555, + 692 + ], + "type": "text", + "content": " is to train pre-CharmNet on samples where the model currently underperforms, enabling it to learn from its weaknesses and refine its ability to generate more appealing descriptions." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 541, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 541, + 714 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 541, + 714 + ], + "type": "text", + "content": "4https://en.wikipedia.org/wiki/Active_learning_(machine_learning)" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19345" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 72, + 526, + 186 + ], + "blocks": [ + { + "bbox": [ + 83, + 72, + 526, + 186 + ], + "lines": [ + { + "bbox": [ + 83, + 72, + 526, + 186 + ], + "spans": [ + { + "bbox": [ + 83, + 72, + 526, + 186 + ], + "type": "image", + "image_path": "de7ed2fb34524a9abc9d0610ff38257c5b4292f70c559c10253d3bba76250c6b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 555, + 231 + ], + "type": "text", + "content": "Figure 2. The framework of CharmNet, including Instruction Tuning which extract image features and refine image-text alignment with supervised fine-tuning, Referee Construction which assess and modify the appeal level of generated descriptions, and Heuristic Active Learning which further enhance CharmNet's ability to generate appealing descriptions for ordinary images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": "After each iteration of training, pre-CharmNet is re-evaluated on " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "D_V" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " to assess whether its performance has improved measured by the referee. If no improvement is observed, the current set of " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " images in " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " is replaced with a new set of " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " images that also fall below the threshold " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": ". This replacement process ensures that pre-CharmNet is consistently exposed to new challenging examples, avoiding overfitting to specific samples in " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": ". The iterative training continues until pre-CharmNet demonstrates consistent improvement on " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "D_V" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " over five consecutive evaluations, indicating that pre-CharmNet has effectively internalized insights from the difficult examples. The update process for constructing " + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 54, + 251, + 297, + 406 + ], + "type": "text", + "content": " is defined as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 416, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 416, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 104, + 416, + 295, + 430 + ], + "type": "interline_equation", + "content": "D _ {1} = \\left\\{\\left(I _ {i}, d _ {i, j}\\right) \\mid r _ {\\theta} \\left(I _ {i}, d _ {i, j}\\right) < \\mu \\right\\}, \\tag {5}", + "image_path": "4e94e0960a270093ec435524b124b4e17ad49dc3b2f081ae69ae269fc6df86c3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " is the predefined performance threshold, " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " represents an image, " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "d_{i,j}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " denotes its corresponding description, and " + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "inline_equation", + "content": "r_{\\theta}" + }, + { + "bbox": [ + 55, + 439, + 296, + 511 + ], + "type": "text", + "content": " is the score assigned by the referee. Through this optimization process, pre-CharmNet is refined into the final CharmNet, equipped with an enhanced ability to generate more appealing descriptions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 523, + 138, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 138, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 138, + 537 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 568 + ], + "type": "text", + "content": "In this section, we conduct experiments to evaluate the effectiveness of our proposed CharmNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 173, + 590 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Our experiments are conducted using PyTorch on four Nvidia A100 GPUs, each with 80GB of memory. We allocate " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " of AppealImage for training, validation, and testing, respectively. The datasets " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_0" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_1" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " are derived from the validation set and do not overlap, while " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "D_V" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " denotes the test set. The maximum sequence lengths for input and output sequences are set to 2048 and 512 tokens, respectively. We use the AdamW optimizer with " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{1}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " set to 0.9, " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta_{2}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " to 0.95, a learning rate of 1e-5, and a weight decay of 1e-1. The batch size is set to 8 during training. LoRA's" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "text", + "content": "rank is set to 32, with a normalization parameter of 16 and a dropout rate of 0.05. Training is conducted for 10 epochs. Experiments validate Qwen-VL for instruction tuning and LLaVA-1.6 for referee construction as optimal choices. In referee construction, each image is paired with " + }, + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "text", + "content": " descriptions, and the predefined evaluation threshold " + }, + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "inline_equation", + "content": "\\mu = 0.2" + }, + { + "bbox": [ + 313, + 251, + 555, + 335 + ], + "type": "text", + "content": ", both yielding optimal performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 346, + 487, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 487, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 487, + 358 + ], + "type": "text", + "content": "5.2. Datasets, Baselines, and Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 364, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 508 + ], + "type": "text", + "content": "We utilize AppealImage and other creative tasks, including visual story creation and situational dialogue generation, to assess CharmNet's performance. For the visual story creation task, we employ the VIST dataset [17], where CharmNet generates appealing narratives based on a sequence of images. CharmNet first creates an appealing description for each individual image, which are then integrated into a cohesive story with the addition of transitional words like \"first,\" \"next,\" and \"then\", etc. In the situational dialogue generation task, we use the DialogCC dataset [21] to evaluate CharmNet's ability to produce appealing responses within dialogue histories that include images." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "The comparison baselines for generating appealing descriptions are shown in Table 2. For the VIST and DialogCC datasets, CharmNet's performance is evaluated using metrics that measure the distance between human and model outputs in terms of coherence (denoted as " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "d_{HM}^{C}" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "), visual grounding (denoted as " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "d_{HM}^{G}" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "), and repetition (denoted as " + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "inline_equation", + "content": "d_{HM}^{R}" + }, + { + "bbox": [ + 313, + 509, + 555, + 605 + ], + "type": "text", + "content": "), as well as R@1, R@5, R@10, and MRR, respectively, consistent with the corresponding research [21, 36]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 714 + ], + "type": "text", + "content": "For evaluation criteria and metrics, the discrimination task is assessed based on whether the MLLM's response matches the ground truth, with top-1 accuracy as the metric. The selection task is evaluated similarly, using top-1 accuracy to determine the MLLM's selected response. For the ranking task, evaluation is based on how closely the MLLM's rankings align with the relative positions of the ground truth, with higher scores awarded for closer matches. The generation task is assessed with automated" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19346" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 293, + 203 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 203 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 203 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 203 + ], + "type": "table", + "html": "
(Total)DSRG
ACCACCPCBROCMReH
CogVLM-7B [39]57.445.274.436.541.561.437.30.243.1
GPT-4o376.457.783.7------
MiniGPT-4-7B [45]41.128.471.131.136.456.132.90.171.8
LLaVA-1.6-7B [28]44.956.68045.351.470.746.50.364.3
InstructBLIP-7B [8]33.624.166.627.932.651.229.90.111.2
mPLUG-Owl-I [42]50.533.467.633.438.358.634.20.192.4
Otter-I [22]36.324.969.629.534.453.931.20.151.3
Qwen-VL-7B+CLoT [44]60.850.376.940.244.865.442.30.303.6
Qwen-VL-Chat-7B [2]54.643.57338.843.363.440.10.283.4
VisualGLM-6B [10]47.738.374.934.640.36035.90.222.9
XMeCap [6]67.542.77543.549.568.545.50.333.7
CharmNet77.559.583.947.253.373.148.00.384.4
↑1.11.80.21.91.92.41.50.020.1
↑(%)1.43.10.24.23.73.43.25.62.3
", + "image_path": "47a4775277e2f4e0c421fcf9f507caad6f39d2f3465f8b7d16a63d2a4b703b98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 87, + 277, + 265, + 411 + ], + "blocks": [ + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 266 + ], + "type": "text", + "content": "Table 2. Comparison of MLLMs and CharmNet across various tasks based on AppealImage. D: Discrimination task, S: Selection task, R: Ranking task, G: Generation task. ACC: Accuracy. PC: Position Consistency. B: BLEU, RO:ROUGE, C: CIDEr, M: METEOR. Re: Referee score. H: Human score." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 87, + 277, + 265, + 411 + ], + "lines": [ + { + "bbox": [ + 87, + 277, + 265, + 411 + ], + "spans": [ + { + "bbox": [ + 87, + 277, + 265, + 411 + ], + "type": "table", + "html": "
DSRDSRG
SiMuGlSiMuGl
CogVLM-7B [39]65.057.454.748.845.744.3
GPT-4o377.470.069.7---
MiniGPT-4-7B [45]53.047.739.842.239.936.5
LLaVA-1.6-7B [28]65.558.657.453.350.549.5
InstructBLIP-7B [8]47.342.035.037.935.732.4
mPLUG-Owl-I [42]57.050.843.844.341.838.7
Otter-I [22]49.543.637.739.737.334.5
Qwen-VL-7B+CLoT [44]68.461.558.052.049.347.5
Qwen-VL-Chat-7B [2]63.455.452.249.245.944.2
VisualGLM-6B [10]59.552.948.646.043.341.3
XMeCap [6]66.259.859.253.050.549.8
CharmNet79.770.171.159.455.755.5
↑2.30.11.46.15.25.7
↑(%)3.00.12.011.410.311.4
", + "image_path": "dfe0695d6d3961fc0b73e59d2536b3025c09470b56c28979db127ae562b6bd2e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 496, + 295, + 616 + ], + "type": "text", + "content": "metrics, including BLEU [31], ROUGE [26], CIDEr, and METEOR [3], alongside appealing scores from 1 to 5 in both referee and human evaluations. For human evaluations, three volunteers are recruited to rate 1,000 randomly selected images and their generated descriptions. To ensure the reliability of human ratings, we calculate inter-rater agreement using Krippendorff's Alpha (IRA). Descriptions with ratings that fall below an IRA threshold of 0.7 are considered controversial, and these are discarded and replaced with descriptions from other images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 140, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 140, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 140, + 635 + ], + "type": "text", + "content": "5.3. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 714 + ], + "type": "text", + "content": "We begin by comparing MLLMs and CharmNet across various tasks on AppealImage, as shown in Table 2. Results indicate that the selected baseline MLLMs perform poorly in generating appealing descriptions for ordinary images, while CharmNet consistently outperforms these MLLMs across all evaluated tasks. Given that our dataset was gen" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 376, + 71, + 493, + 120 + ], + "blocks": [ + { + "bbox": [ + 55, + 419, + 295, + 477 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 477 + ], + "type": "text", + "content": "Table 3. Comparison of MLLMs and CharmNet across three perspectives in AppealImage. DSR: Discrimination, Selection, and Ranking tasks. DSRG: Discrimination, Selection, Ranking, and Generation tasks. Si: Single-object perspective, Mu: Multi-object perspective, Gl: Global perspective." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 376, + 71, + 493, + 120 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 493, + 120 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 493, + 120 + ], + "type": "table", + "html": "
(Global)MaReHTime
Human-written48.60.314.330.5
CharmNet52.70.354.53.2
↑4.10.040.227.3
↑(%)8.412.94.7853.1
", + "image_path": "1bad674f0c7ab8963ce4b3aad73bb752a0c0fc051cb34acf3b2b3ed1d4e1a164.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 129, + 553, + 174 + ], + "lines": [ + { + "bbox": [ + 313, + 129, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 129, + 553, + 174 + ], + "type": "text", + "content": "Table 4. Comparison of CharmNet-generated and human-written appealing descriptions. Ma: Average of BLEU, ROUGE, CIDEr, and METEOR. Re: Referee score. H: Human score. Time: The average time to generate one description." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "content": "erated by GPT-4o, we exclude GPT-4o from the generation task to avoid inflated scores due to stylistic similarity with the original dataset. Notably, CharmNet achieves an accuracy of " + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "inline_equation", + "content": "77.5\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "content": " in the discrimination task, surpassing GPT-4o's accuracy of " + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "inline_equation", + "content": "76.4\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "content": " and showing an approximately " + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "inline_equation", + "content": "130.7\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "content": " improvement over the lowest-performing MLLM, InstructBLIP-7B, which scored " + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "inline_equation", + "content": "33.6\\%" + }, + { + "bbox": [ + 313, + 196, + 555, + 339 + ], + "type": "text", + "content": ". Similarly, CharmNet outperforms GPT-4o and demonstrates a marked improvement over InstructBLIP-7B in the selection task. In the ranking task, both CharmNet and GPT-4o exhibit top performance. In the generation task, CharmNet achieves a human-rated appeal score of 4.4, ranking the highest." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 556, + 449 + ], + "type": "text", + "content": "Next, we conduct a detailed comparison of MLLMs and CharmNet across different task perspectives, as shown in Table 3, once again excluding GPT-4o's performance in the generation task. We calculate the average performance across all MLLMs for the discrimination, selection, and ranking tasks, as well as for all MLLMs excluding GPT-4o. Results show that CharmNet consistently outperforms other MLLMs, with particularly strong results in single-object perspective generation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 568 + ], + "type": "text", + "content": "Furthermore, we invite three literary experts to write appealing descriptions for images, aiming to compare human performance with machine-generated descriptions in the task of in generating appealing descriptions. Due to the relatively long time required for human writing (approximately 30 seconds per description), we randomly select 1,000 images from the test set of AppealImage for evaluation. The results are shown in Table 4, demonstrating that CharmNet has surpassed human experts in both effectiveness and efficiency." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 556, + 676 + ], + "type": "text", + "content": "Additionally, we evaluate CharmNet on two creative tasks, visual story creation and situational dialogue generation, with results shown in Table 5 and Table 6, respectively. Baseline results, excluding Qwen-VL-7B+CLoT, are drawn from previous studies [21, 36]. The findings indicate that CharmNet significantly outperforms the baselines, with only a slight improvement over Qwen-VL-7B+CLoT, validating CharmNet's generalizability across a variety of creative tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Last but not least, while maintaining creativity, we further evaluate CharmNet on non-creative tasks to investigate whether training on the AppealImage dataset leads to the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19347" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 71, + 257, + 144 + ], + "blocks": [ + { + "bbox": [ + 95, + 71, + 257, + 144 + ], + "lines": [ + { + "bbox": [ + 95, + 71, + 257, + 144 + ], + "spans": [ + { + "bbox": [ + 95, + 71, + 257, + 144 + ], + "type": "table", + "html": "
(DialogCC)R@1R@5R@10MRR
Image Only8.2222.633.5216.94
Dialogue Only34.4165.3677.3748.67
Dialogue + Image40.6471.4681.9954.61
Owen-VL-7B+CLoT [44]52.6778.6788.2462.46
CharmNet56.3482.2191.4668.11
↑3.673.543.225.65
↑(%)7.04.53.69.0
", + "image_path": "1983ff5918e063b090c6420a41217e5bfca7f7e1883a1d69f9f75f413fb402e8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 85, + 186, + 266, + 276 + ], + "blocks": [ + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 153, + 295, + 175 + ], + "type": "text", + "content": "Table 5. Performance of CharmNet on the contextual dialogue generation task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 186, + 266, + 276 + ], + "lines": [ + { + "bbox": [ + 85, + 186, + 266, + 276 + ], + "spans": [ + { + "bbox": [ + 85, + 186, + 266, + 276 + ], + "type": "table", + "html": "
(VIST)dCM(-)dGM(-)dRHM(-)dHM(-)
AREL [40]0.1100.0580.0400.208
GLAC Net [20]0.0700.1070.0070.184
TAPM [43]0.0950.0150.0400.150
BLIP-2 [23]0.1900.0500.0940.334
LLaVA [28]0.0480.0040.0470.099
Qwen-VL-7B+CLoT [44]0.0350.0030.0310.069
CharmNet0.0190.0020.0240.045
↓0.0160.0010.0070.024
↓(%)84.250.029.253.3
", + "image_path": "e97b5e78fbc2b11dcbd71c51e00fcf44794ffbc6c7d5989f94c152dbe58c95ad.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 295, + 296 + ], + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 296 + ], + "type": "text", + "content": "Table 6. Performance of CharmNet on the visual storytelling task." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": "generation of irrelevant information. We employ an image captioning task using 500 randomly selected samples of the MSCOCO dataset, adjusting instructions to guide CharmNet in generating captions that are devoid of creative elements. We design three criteria based on key factors in image captioning as proposed in Huang et al. [16], and conduct manual evaluation with the results shown in Table 7. The manual evaluation show that " + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "inline_equation", + "content": "89.6\\%" + }, + { + "bbox": [ + 54, + 316, + 295, + 460 + ], + "type": "text", + "content": " of the generated results correctly described the main objects and their relationships in the image, indicating that hallucination issues are relatively infrequent when CharmNet is used for non-creative tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": "The influence of different dataset size in heuristic active learning. We further examine the influence of varying the size of dataset " + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": " in heuristic active learning, as illustrated in Fig. 3(a). The results show a steady increase in performance as the size of " + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": " grows, peaking at " + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": ". Beyond this threshold, performance begins to decrease, indicating that a " + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "inline_equation", + "content": "D_{1}" + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": " size of " + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 55, + 460, + 295, + 558 + ], + "type": "text", + "content": " achieves an optimal balance between data scale and quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 149, + 577 + ], + "type": "text", + "content": "5.4. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "The impact of each module in CharmNet. We first make analysis of the impact of each module in CharmNet as shown in Table 8. All results are averaged for three times to avoid randomness. When the first stage of instruction tuning is removed (denoted as \"w/o IT1\"), we observe a substantial decline of " + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "inline_equation", + "content": "9.2\\%" + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": " across all metrics. The possible reason may be that directly unfreezing both image and text together for learning mixed their features, preventing the model from learning useful information. Similarly, the removal of heuristic active learning (denoted as \"w/o HAL\") has a pronounced effect, indicating that HAL fa" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 433, + 156 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 433, + 156 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 433, + 156 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 433, + 156 + ], + "type": "image", + "image_path": "dad9c75132d4a4663037012fbdd55ac6c13975085e537f40b2bff4d0735837e8.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 167, + 554, + 222 + ], + "lines": [ + { + "bbox": [ + 313, + 167, + 554, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 554, + 222 + ], + "type": "text", + "content": "Figure 3. (a) Effect of the proportion of difficult data in heuristic active learning. GenerationM, GenerationR, and GenerationH represent machine scores (i.e., BLEU, ROUGE,CIDEr, METEOR), referee scores, and human scores, respectively. (b) Performance across tasks at varying difficulty levels." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 436, + 72, + 551, + 156 + ], + "blocks": [ + { + "bbox": [ + 436, + 72, + 551, + 156 + ], + "lines": [ + { + "bbox": [ + 436, + 72, + 551, + 156 + ], + "spans": [ + { + "bbox": [ + 436, + 72, + 551, + 156 + ], + "type": "image", + "image_path": "9cc84b20e7e424f18259c29022017a6b4b4b0be5794b3527c4efb8b0dbcae0e6.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 244, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 554, + 388 + ], + "type": "text", + "content": "cilitates targeted training and improvements through challenging samples. Furthermore, replacing HAL with reinforcement learning (denoted as \"HAL to RL\", we use direct preference optimization here) does not yield significant performance improvements but has substantially higher training costs. Moreover, substituting multi-task learning with single-task learning (denoted as \"MTL to STL\"), where only the generation task is used, does not enhance performance in the generation task. This suggests that complementary tasks contribute to CharmNet's understanding of what makes descriptions appealing, thereby improving the quality of generated outputs." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 388, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 532 + ], + "type": "text", + "content": "The function of AppealImage to CharmNet. Next, we use Qwen-VL-7B+CLoT for instruction tuning on AppealImage. While it underperforms on AppealImage, it excels in the Oogiri game, which requires creativity and associative thinking for generating humorous responses. We compare its performance with CharmNet in global perspective generation task, as shown in Table 9, where CharmNet consistently outperforms the AppealImage-tuned Qwen-VL-7B+CLoT. This highlights that the effectiveness of CharmNet is not solely attributable to training on AppealImage but also to its combined strategies like multi-task learning and heuristic active learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 664 + ], + "type": "text", + "content": "The effect of different difficulty levels in each tasks. After that, we investigate the effect of different difficulty levels on CharmNet's performance in the discrimination, selection, and ranking tasks, as shown in Fig. 3(b). Results reveal a near-linear performance decline across all tasks as difficulty increases from easy to hard. Data difficulty has the greatest impact on the selection task, followed by the discrimination task, likely due to the fewer options in discrimination tasks, which simplifies the identification of correct answers. The ranking task is not directly comparable due to its unique evaluation metrics." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "The correlation between referee and human scores for CharmNet. Since ground-truth labels for these tasks are provided by the referee, we employ the Pearson correlation coefficient to ensure alignment between referee and" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19348" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 71, + 553, + 111 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 553, + 111 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 553, + 111 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 553, + 111 + ], + "type": "table", + "html": "
CriteriaPortion (%)
Descriptions of the main objects in the image are incorrect or the relationships between objects in the image are wrong.52.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, but the descriptions are straightforward and unappealing.26.0
Descriptions of the main objects in the image are correct, the relationships between objects are accurately described, and the descriptions are creative and appealing.42.0
", + "image_path": "9c1268476380869415f87742eced307c1b67044964d2a0290baea98e9f3da5d8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 162, + 141, + 449, + 198 + ], + "blocks": [ + { + "bbox": [ + 76, + 118, + 532, + 130 + ], + "lines": [ + { + "bbox": [ + 76, + 118, + 532, + 130 + ], + "spans": [ + { + "bbox": [ + 76, + 118, + 532, + 130 + ], + "type": "text", + "content": "Table 7. Three criteria and the corresponding portion based on manual evaluation on 500 samples from the MSCOCO dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 162, + 141, + 449, + 198 + ], + "lines": [ + { + "bbox": [ + 162, + 141, + 449, + 198 + ], + "spans": [ + { + "bbox": [ + 162, + 141, + 449, + 198 + ], + "type": "table", + "html": "
(Total)D (↓%)S (↓%)R (↓%)GBRCM (↓%)GRe (↓%)GH (↓%)A (↓%)
w/o IT171.8 (7.9)54.0 (10.2)78.0 (7.6)39.0 (13.8)0.29 (31.0)4.0 (10.0)41.2 (9.2)
MTL to STL---43.1 (3.0)0.34 (11.8)4.2 (4.8)-
HAL to RL75.8 (2.2)56.9 (4.6)81.2 (3.3)44.1 (0.7)0.37 (2.7)4.2 (4.8)43.8 (2.7)
w/o HAL66.0 (17.4)48.1 (23.7)73.8 (13.7)33.7 (31.8)0.28 (35.7)3.5 (25.7)37.6 (19.7)
CharmNet77.559.583.944.40.384.445.0
", + "image_path": "f68f91f6e04be7f89042ad83b55efc2bc87a0b69218f32905fa444897efe038f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 123, + 241, + 488, + 327 + ], + "blocks": [ + { + "bbox": [ + 123, + 241, + 488, + 327 + ], + "lines": [ + { + "bbox": [ + 123, + 241, + 488, + 327 + ], + "spans": [ + { + "bbox": [ + 123, + 241, + 488, + 327 + ], + "type": "image", + "image_path": "858db653aab95c0af75fbaf453d71a6b48b1678ed83f37f3d8bb6ba8bb58d000.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 166, + 338, + 443, + 350 + ], + "lines": [ + { + "bbox": [ + 166, + 338, + 443, + 350 + ], + "spans": [ + { + "bbox": [ + 166, + 338, + 443, + 350 + ], + "type": "text", + "content": "Figure 4. A running case of CharmNet in generating appealing descriptions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 95, + 368, + 257, + 416 + ], + "blocks": [ + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "text", + "content": "Table 8. Impact of each module in CharmNet. " + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "inline_equation", + "content": "{\\mathrm{G}}_{BRCM}" + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "text", + "content": " , " + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "inline_equation", + "content": "{\\mathrm{G}}_{Re}" + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "text", + "content": " ,and " + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "inline_equation", + "content": "{\\mathrm{G}}_{H}" + }, + { + "bbox": [ + 55, + 206, + 553, + 229 + ], + "type": "text", + "content": " denote machine,referee,and human scores,respectively. A: Average results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 95, + 368, + 257, + 416 + ], + "lines": [ + { + "bbox": [ + 95, + 368, + 257, + 416 + ], + "spans": [ + { + "bbox": [ + 95, + 368, + 257, + 416 + ], + "type": "table", + "html": "
(Total)BROCMReH
Qwen-VL-7B+CLoT45.550.271.647.10.344.1
CharmNet47.253.373.148.00.384.4
↑1.73.11.50.90.040.3
↑(%)3.76.22.11.911.87.3
", + "image_path": "307faffe276bea8c98615444fe2712f69e91b950f560b11bac732b26b67c802b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 119, + 458, + 232, + 506 + ], + "blocks": [ + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 448 + ], + "type": "text", + "content": "Table 9. Comparison of AppealImage-tuned Qwen-VL-7B+CLoT and CharmNet in the generation task." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 119, + 458, + 232, + 506 + ], + "lines": [ + { + "bbox": [ + 119, + 458, + 232, + 506 + ], + "spans": [ + { + "bbox": [ + 119, + 458, + 232, + 506 + ], + "type": "table", + "html": "
DSRG
Single-object0.820.740.710.89
Multiple-object0.800.750.700.87
Global0.850.780.730.86
Total0.820.760.710.87
", + "image_path": "87c24bf9a1bbf9bf7a8bb62ca2d430c4a2c5daed61191dfdf8fa3111f6ffb7ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 514, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 536 + ], + "type": "text", + "content": "Table 10. Pearson correlation coefficients between referee and human scores for CharmNet." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 665 + ], + "type": "text", + "content": "human scores for CharmNet. The results are shown in Table 10, indicating a high level of consistency with all values exceeding 0.8 (p<0.05), especially in the generation task. We also test with a larger dataset, but the correlation coefficient does not show significant improvement. This high correlation underscores the reliability of the referee, demonstrating its potential to reduce the need for extensive human involvement while enhancing CharmNet's ability to generate appealing descriptions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 130, + 685 + ], + "type": "text", + "content": "5.5. Case Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We present a case study of the appealing descriptions generated by CharmNet, illustrated in Fig. 4. For instance, for the" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 502 + ], + "type": "text", + "content": "single-object description, CharmNet uses terms like “zips” and “exciting trail” to describe the tram, conveying a sense of speed. In the multi-object description, phrases such as “races by”, “vibrant carnival”, “swaying”, and “flashing” depict a static scene as a lively city celebration. Finally, in the global description, words like “leading a parade”, “bringing life and color”, and “effortless charm” enhance the image’s atmosphere, making it appear more dynamic and engaging. We also observe some less appealing descriptions, such as “a boy walking with a man in a crowd …”, which lacks vivid detail and emotional depth." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 392, + 526 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "How to craft appealing descriptions by MLLMs for ordinary images beyond those internet memes is quite underinvestigated before. To bridge this gap, we present AppealImage, a large-scale dataset containing ordinary images paired with appealing descriptions. We define four tasks, each paired with quantitative metrics, to enable thorough evaluation. Upon AppealImage, we introduce CharmNet, an innovative framework designed to generate appealing descriptions for ordinary images. CharmNet leverages instruction tuning and heuristic active learning, guided by a referee model. Our experiments show CharmNet outperforms the state-of-the-art method by " + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "inline_equation", + "content": "11.4\\%" + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": " in generating appealing descriptions and has impressive results across related creative tasks. Future work will aim to broaden CharmNet's applicability and customize it for personalized contents." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19349" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 117 + ], + "type": "text", + "content": "This work is supported by Ant Group Research Intern Program." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 115, + 140 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 148, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 62, + 148, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 148, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 148, + 296, + 196 + ], + "type": "text", + "content": "[1] Emran Al-Buraihy and W Dan. Enhancing crosslingual image description: A multimodal approach for semantic relevance and stylistic alignment. Comput. Mater. Contin, 2, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 198, + 296, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 296, + 258 + ], + "type": "text", + "content": "[2] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond, 2023. 4, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 260, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 331 + ], + "type": "text", + "content": "[3] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In Proceedings of the acl workshop on intrinsic and extrinsic evaluation measures for machine translation and/or summarization, pages 65-72, 2005. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 333, + 295, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 333, + 295, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 333, + 295, + 369 + ], + "type": "text", + "content": "[4] Arjun Chandrasekaran, Devi Parikh, and Mohit Bansal. Funny captions: Witty wordplay in image descriptions. arXiv preprint arXiv:1704.08224, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 371, + 296, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 371, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 371, + 296, + 407 + ], + "type": "text", + "content": "[5] Yuyan Chen, Zhixu Li, Jiaqing Liang, Yanghua Xiao, Bang Liu, and Yunwen Chen. Can pre-trained language models understand chinese humor? 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 408, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 408, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 408, + 296, + 456 + ], + "type": "text", + "content": "[6] Yuyan Chen, Songzhou Yan, Zhihong Zhu, Zhixu Li, and Yanghua Xiao. Xmecap: Meme caption generation with sub-image adaptability. arXiv preprint arXiv:2407.17152, 2024. 2, 3, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 457, + 296, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 540 + ], + "type": "text", + "content": "[7] Yuyan Chen, Yichen Yuan, Panjun Liu, Dayiheng Liu, Qinghao Guan, Mengfei Guo, Haiming Peng, Bang Liu, Zhixu Li, and Yanghua Xiao. Talk funny! a largescale humor response dataset with chain-of-humor interpretation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 17826-17834, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 543, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 296, + 591 + ], + "type": "text", + "content": "[8] W Dai, J Li, D Li, AMH Tiong, J Zhao, W Wang, B Li, P Fung, and S Hoi. Instructclip: Towards general-purpose vision-language models with instruction tuning. arXiv preprint arXiv:2305.06500, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 593, + 296, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 651 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE Conference on Computer Vision and Pattern Recognition, pages 248–255, 2009. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 654, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 654, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 654, + 296, + 712 + ], + "type": "text", + "content": "[10] Zhengxiao Du, Yujie Qian, Xiao Liu, Ming Ding, Jiezhong Qiu, Zhilin Yang, and Jie Tang. Glm: General language model pretraining with autoregressive blank infilling. arXiv preprint arXiv:2103.10360, 2021. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 120 + ], + "type": "text", + "content": "[11] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. arXiv preprint arXiv:2403.12026, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 123, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 123, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 553, + 183 + ], + "type": "text", + "content": "[12] Chongyang Gao, Yiren Jian, Natalia Denisenko, Soroush Vosoughi, and VS Subrahmanian. Gem: generating engaging multimodal content. In Proceedings of the Thirty-Third International Joint Conference on Artificial Intelligence, pages 7654–7662, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 186, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 388 + ], + "type": "text", + "content": "[13] Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Dan Zhang, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Jingyu Sun, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. Chatglm: A family of large language models from glm-130b to glm-4 all tools, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 392, + 553, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 392, + 553, + 440 + ], + "spans": [ + { + "bbox": [ + 316, + 392, + 553, + 440 + ], + "type": "text", + "content": "[14] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 442, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 442, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 442, + 553, + 525 + ], + "type": "text", + "content": "[15] Hexiang Hu, Kelvin CK Chan, Yu-Chuan Su, Wenhu Chen, Yandong Li, Kihyuk Sohn, Yang Zhao, Xue Ben, Boqing Gong, William Cohen, et al. Instruct-imagen: Image generation with multi-modal instruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4754-4763, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 529, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 529, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 317, + 529, + 553, + 577 + ], + "type": "text", + "content": "[16] Lun Huang, Wenmin Wang, Jie Chen, and Xiao-Yong Wei. Attention on attention for image captioning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4634-4643, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 579, + 553, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 553, + 675 + ], + "type": "text", + "content": "[17] Ting-Hao Huang, Francis Ferraro, Nasrin Mostafazadeh, Ishan Misra, Aishwarya Agrawal, Jacob Devlin, Ross Girshick, Xiaodong He, Pushmeet Kohli, Dhruv Batra, et al. Visual storytelling. In Proceedings of the 2016 conference of the North American chapter of the association for computational linguistics: Human language technologies, pages 1233-1239, 2016. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 677, + 553, + 714 + ], + "type": "text", + "content": "[18] Xiaoke Huang, Jianfeng Wang, Yansong Tang, Zheng Zhang, Han Hu, Jiwen Lu, Lijuan Wang, and Zicheng Liu. Segment and caption anything. In Proceedings" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19350" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 296, + 96 + ], + "type": "text", + "content": "of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13405-13417, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 98, + 296, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 98, + 296, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 98, + 296, + 158 + ], + "type": "text", + "content": "[19] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 159, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 294, + 218 + ], + "type": "text", + "content": "[20] Taehyeong Kim, Min-Oh Heo, Seonil Son, Kyoung-Wha Park, and Byoung-Tak Zhang. Glac net: Glocal attention cascading networks for multi-image cued story generation. arXiv preprint arXiv:1805.10973, 2018. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 269 + ], + "type": "text", + "content": "[21] Young-Jun Lee, Byungsoo Ko, Han-Gyu Kim, and Ho-Jin Choi. Dialogcc: Large-scale multi-modal dialogue dataset. arXiv preprint arXiv:2212.04119, 2022. 5, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 270, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 270, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 270, + 294, + 318 + ], + "type": "text", + "content": "[22] Bo Li, Yuanhan Zhang, Liangyu Chen, Jinghao Wang, Jingkang Yang, and Ziwei Liu. Otter: A multimodal model with in-context instruction tuning. corr abs/2305.03726 (2023), 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 368 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 368 + ], + "type": "text", + "content": "[23] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pretraining with frozen image encoders and large language models, 2023. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 369, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 369, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 294, + 441 + ], + "type": "text", + "content": "[24] Jiaxuan Li, Duc Minh Vo, Akihiro Sugimoto, and Hideki Nakayama. Evcap: Retrieval-augmented image captioning with external visual-name memory for open-world comprehension. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13733-13742, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 443, + 294, + 515 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 515 + ], + "type": "text", + "content": "[25] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22195-22206, 2024. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 517, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 294, + 564 + ], + "type": "text", + "content": "[26] Chin-Yew Lin. ROUGE: A package for automatic evaluation of summaries. In Text Summarization Branches Out, pages 74-81, Barcelona, Spain, 2004. Association for Computational Linguistics. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 566, + 294, + 638 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 638 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 638 + ], + "type": "text", + "content": "[27] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision - ECCV 2014, pages 740-755, Cham, 2014. Springer International Publishing. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 639, + 294, + 675 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 639, + 294, + 675 + ], + "spans": [ + { + "bbox": [ + 56, + 639, + 294, + 675 + ], + "type": "text", + "content": "[28] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 4, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 677, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 294, + 713 + ], + "type": "text", + "content": "[29] Feipeng Ma, Yizhou Zhou, Fengyun Rao, Yueyi Zhang, and Xiaoyan Sun. Image captioning with multi-context synthetic data. In Proceedings of the" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 336, + 72, + 553, + 95 + ], + "type": "text", + "content": "AAAI Conference on Artificial Intelligence, pages 4089-4097, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 98, + 553, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 145 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 145 + ], + "type": "text", + "content": "[30] Thao Nguyen, Samir Yitzhak Gadre, Gabriel Ilharco, Sewoong Oh, and Ludwig Schmidt. Improving multimodal datasets with image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 148, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 148, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 148, + 553, + 232 + ], + "type": "text", + "content": "[31] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics, pages 311-318, Philadelphia, Pennsylvania, USA, 2002. Association for Computational Linguistics. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 234, + 553, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 305 + ], + "type": "text", + "content": "[32] Suzanne Petryk, Spencer Whitehead, Joseph E Gonzalez, Trevor Darrell, Anna Rohrbach, and Marcus Rohrbach. Simple token-level confidence improves caption correctness. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5742-5752, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 307, + 553, + 377 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 307, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 316, + 307, + 553, + 377 + ], + "type": "text", + "content": "[33] Noam Rotstein, David Bensaïd, Shaked Brody, Roy Ganz, and Ron Kimmel. Fusecap: Leveraging large language models for enriched fused image captions. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5689-5700, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 380, + 555, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 380, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 380, + 555, + 441 + ], + "type": "text", + "content": "[34] Jie Ruan, Yue Wu, Xiaojun Wan, and Yuesheng Zhu. Describe images in a boring way: Towards cross-modal sarcasm generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 5701-5710, 2024. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 442, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 502 + ], + "type": "text", + "content": "[35] Kurt Shuster, Samuel Humeau, Hexiang Hu, Antoine Bordes, and Jason Weston. Engaging image captioning via personality. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12516-12526, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 504, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 564 + ], + "type": "text", + "content": "[36] Aditya K Surikuchi, Raquel Fernandez, and Sandro Pezzelle. Not (yet) the whole story: Evaluating visual storytelling requires more than measuring coherence, grounding, and repetition. arXiv preprint arXiv:2407.04559, 2024. 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 566, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 566, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 566, + 553, + 601 + ], + "type": "text", + "content": "[37] Joel Taylor and David Watkinson. Indexing reliability for condition survey data. The Conservator, 30(1):49-62, 2007. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 603, + 553, + 651 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 651 + ], + "type": "text", + "content": "[38] Han Wang and Roy Ka-Wei Lee. Memecraft: Contextual and stance-driven multimodal meme generation. In Proceedings of the ACM on Web Conference 2024, pages 4642–4652, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 654, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 654, + 553, + 713 + ], + "type": "text", + "content": "[39] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, et al. Cogvlm: Visual expert for pretrained language models. arXiv preprint arXiv:2311.03079, 2023. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19351" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 423 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 120 + ], + "type": "text", + "content": "[40] Xin Wang, Wenhu Chen, Yuan-Fang Wang, and William Yang Wang. No metrics are perfect: Adversarial reward learning for visual storytelling. arXiv preprint arXiv:1804.09160, 2018.7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 121, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 121, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 121, + 295, + 168 + ], + "type": "text", + "content": "[41] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 170, + 295, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 170, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 295, + 230 + ], + "type": "text", + "content": "[42] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. mplug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 232, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 232, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 57, + 232, + 295, + 290 + ], + "type": "text", + "content": "[43] Youngjae Yu, Jiwan Chung, Heeseung Yun, Jongseok Kim, and Gunhee Kim. Transitional adaptation of pretrained models for visual storytelling. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12658-12668, 2021. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 292, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 292, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 57, + 292, + 294, + 374 + ], + "type": "text", + "content": "[44] Shanshan Zhong, Zhongzhan Huang, Shanghua Gao, Wushao Wen, Liang Lin, Marinka Zitnik, and Pan Zhou. Let's think outside the box: Exploring leap-of-thought in large language models with creative humor generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13246-13257, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 376, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 294, + 423 + ], + "type": "text", + "content": "[45] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models, 2023. 6" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19352" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_content_list.json b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ea9541517798fc2bb172207698eff2ce2469ddee --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_content_list.json @@ -0,0 +1,2050 @@ +[ + { + "type": "text", + "text": "Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation", + "text_level": 1, + "bbox": [ + 109, + 130, + 887, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haihao Zhang $^{1,2}$ Yunjian Zhang $^{3\\dagger}$ Jianing Li $^{3\\dagger}$ Lin Zhu $^{2}$ Meng Lv $^{2}$ Yao Zhu $^{3}$ Yanwei Liu $^{1}$ Xiangyang Ji $^{3}$", + "bbox": [ + 187, + 178, + 803, + 219 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences $^{2}$ Beijing Institute of Technology $^{3}$ Tsinghua University", + "bbox": [ + 233, + 215, + 763, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "haihaozhang@bit.edu.cn sdtczyj@gmail.com lijianing@pku.edu.cn", + "bbox": [ + 214, + 253, + 777, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 319, + 326, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Accurate stereo matching under fast motion and extreme lighting conditions is a challenge for many vision applications. Event cameras have the advantages of low latency and high dynamic range, thus providing a reliable solution to this challenge. However, since events are sparse, this makes it an ill-posed problem to obtain dense disparity using only events. In this work, we propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. Specifically, a multi-level intensity-to-event distillation strategy is designed to maximize the potential of long-range information, local texture details, and task-related knowledge of the intensity images. Simultaneously, to enforce the cross-view consistency, an intensity-event joint left-right consistency module is proposed. With our framework, extensive dense and structural information contained in intensity images is distilled to the event branch. Therefore, retaining only the events can predict dense disparities during inference, preserving the low latency characteristics of the events. Adequate experiments conducted on the MVSEC and DSEC datasets demonstrate that our method exhibits superior stereo matching performance than baselines, both quantitatively and qualitatively.", + "bbox": [ + 88, + 351, + 483, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 713, + 220, + 728 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Depth estimation based on binocular stereo vision is a classic problem in computer vision [15, 16, 20, 24, 29, 36, 44, 46, 59], which aims to determine the correspondence between points in a rectified image pair. Deep learning-based stereo matching methods perform well on various public benchmarks [2, 6, 9, 23, 24, 31-33, 52, 54-57]. However, due to the shortcomings of frame-based sensors in terms of frame rate and dynamic range, it is very difficult to apply traditional cameras in more challenging scenarios such as", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b89b1ea31765c0322264d0b680275c6d4170439a658e2bba8cc46e1dd580d9b8.jpg", + "image_caption": [ + "(a) Event-based Method" + ], + "image_footnote": [], + "bbox": [ + 527, + 324, + 895, + 372 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e7334d2c544679a3eac1aa96ae1da71845bbf97f4ab5d71784764ed71c9187f0.jpg", + "image_caption": [ + "(b) Event-Intensity Fusion Method" + ], + "image_footnote": [], + "bbox": [ + 522, + 388, + 895, + 497 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da46527f76d8da6623574e2a3331938f3948b3844f218d84fdd922ae91755e64.jpg", + "image_caption": [ + "(c) Our Proposed Method", + "Figure 1. Comparisons on stereo matching frameworks (event-based, event-intensity fusion methods) and our proposed IED $^2$ S that presents an effective intensity-to-event distillation strategy. Texture and dense knowledge is transferred to the event branch, leading to dense stereo matching in the inference stage." + ], + "image_footnote": [], + "bbox": [ + 521, + 521, + 898, + 652 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high speed and extreme lighting conditions.", + "bbox": [ + 511, + 750, + 803, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event cameras [3, 7, 30, 35, 40, 47, 64] are bio-inspired sensors and report intensity changes asynchronously in the form of an event stream, with each containing spatial location, polarity, and timestamps. They have the advantages of low latency and high dynamic range, making them immune to high-speed motion blur and can operate in extreme lighting scenarios. Therefore, they offer a satisfying solution to overcome the limitations of frame-based sensors.", + "bbox": [ + 509, + 763, + 905, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent event-based stereo matching studies [1, 10, 11,", + "bbox": [ + 531, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding authors: Jianing Li, Yunjian Zhang", + "bbox": [ + 114, + 886, + 392, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5437", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "13, 18, 37, 38, 49] embed asynchronous event data into image-like representations, then fine-grained features are extracted and subsequent disparity regression is performed. While disparity can be effectively estimated, since event cameras can only perceive changes in intensity, when the scene is stationary or there is little relative motion between the camera and the scene, the event stream will be very sparse. This makes it an ill-posed problem to obtain dense disparity maps using only sparse asynchronous event data.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Considering the dense features provided by intensity images, there raises a trend to combine events and intensity images for dense stereo matching [10, 11, 37]. They design algorithms that simultaneously input intensity images and events, using their complementary features to calculate disparities, and the results are outstanding advances for event-based stereo matching. Nevertheless, these approaches require the events and intensity images simultaneously during the inference stage, suffering from two problems: (1) The low latency characteristics of events is sacrificed unintentionally; (2) Processing both modalities concurrently consumes more computing resources, which is a heavy burden for real-time applications.", + "bbox": [ + 89, + 229, + 480, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above problems, we propose a novel framework named Intensity-to-Event Distillation for Dense Stereo matching method (IED $^2$ S), which explores the potential of an efficient event-based stereo model for dense disparity prediction. As shown in Fig. 1(c), we implement cross-sensor distillations from the intensity branch to the event branch. The dense and structural information of the intensity image provides useful regularization for asynchronous events, especially for ill-posed regions. In our framework, we design a novel feature alignment strategy, achieving binocular-to-binocular distillation for the first time. Specifically, a comprehensive multi-level intensity-to-event distillation strategy is designed, fully leveraging the modality-specific characteristics of shallow features and the task-related contextual understanding involved in deep abstract representations. Moreover, a warping-based intensity-event joint left-right consistency module is designed to ensure cross-view consistency. The intensity images are only used in the training stage. For inference, only event data is needed for obtaining dense disparity. Compared with fusion-based methods shown in Fig. 1(b), it has greater flexibility, preserving the low latency characteristics of event data and greatly reducing the inference burden. In summary, the main contributions of this work are:", + "bbox": [ + 89, + 428, + 482, + 789 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1) We propose a dense event-based stereo matching method, $\\mathbf{IED}^2\\mathbf{S}$ , which is the first trial to implement binocular intensity-to-event distillation. The proposed $\\mathbf{IED}^2\\mathbf{S}$ enables the event branch to learn accurate dense information, greatly improving its disparity prediction capability.", + "2) We design a multi-level intensity-to-event distillation, fully distilling long-range information, local texture details," + ], + "bbox": [ + 89, + 792, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and task-related knowledge into the event branch while avoiding the contamination of the original event modality.", + "bbox": [ + 511, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3) Extensive experiments demonstrate that the proposed $\\mathbf{IED}^2\\mathbf{S}$ achieves significantly better performance than the prior works for event-based stereo matching.", + "bbox": [ + 511, + 122, + 903, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 183, + 653, + 198 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Event-Based Stereo", + "text_level": 1, + "bbox": [ + 513, + 209, + 699, + 223 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early attempts used traditional hand-crafted methods [4, 12, 42-45, 68, 71] to determine the correspondence between two event streams. However, there are no fixed patterns in the event data, which makes it troublesome to use traditional methods to obtain dense disparity.", + "bbox": [ + 511, + 231, + 903, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent learning-based approaches [1, 10, 11, 13, 37, 38, 49, 51] are able to estimate disparities using sparse events. The network usually includes feature extraction, cost aggregation, disparity regression, and disparity refinement modules. Tulyakov et al. [49] proposed a 4D queue embedding method that includes the spatiotemporal information of event data. Nam et al. [38] proposed an attention-based event concentration network to display scene details by omitting fewer details without covering events. DTC [62] used discrete time convolution modules to aggregate sequential event information at the feature level, accelerating the stereo matching process at a lower computational cost. Cho et al. [13] proposed the concept of stereo flow to aggregate past features and cost volumes. Accurate stereo matching is achieved by training the stereo matching and stereo flow simultaneously. However, it is an ill-posed problem to obtain dense disparity using only sparse asynchronous events. We therefore explore cross-sensor distillation for event-based dense stereo matching.", + "bbox": [ + 511, + 308, + 903, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Cross-Sensor Distillation", + "text_level": 1, + "bbox": [ + 513, + 606, + 740, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge distillation is an effective method for compressing models while maintaining accuracy [21, 41, 50, 63]. In particular, cross-sensor knowledge distillation has received increasing attention and applied in many tasks, such as 3D object detection [8, 14, 22, 53, 65, 66], multispectral detection [34, 61], 3D hand pose estimation [60], and LiDAR semantic segmentation [25, 26, 48, 58]. Typically, a well-trained teacher model guides the student model's feature extraction or result prediction that receives input from other modalities. Specifically, Hong et al. [22] first proposed cross-sensor knowledge distillation based on BEV, mimicking LiDAR BEV's representation to achieve improved performance. Chong et al. [14] achieved the transfer of structural cues by projecting point clouds onto the plane of the image and maintaining similar local region affinities. Liu et al. [34] adopted different losses at multiple task levels to transfer coarse-grained and fine-grained features covering feature, detection, and segmentation. Tang et al. [48] estab", + "bbox": [ + 511, + 628, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5438", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5cb4d299c5a05b21db20b69663dfa9d4d855dbd173833363253a0d9de51376c3.jpg", + "image_caption": [ + "Figure 2. The pipeline of the proposed Intensity-to-Event Distillation for Dense Stereo matching method (IED $^2$ S). The proposed approach enables the event branch to learn a massive amount of texture knowledge and dense structure information under the guidance of the intensity branch. After training, the intensity branch is removed without introducing additional computational burden during inference." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 893, + 422 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "lished a prototype-like library from strictly aligned fused features, and encouraged all point cloud features to learn from the prototypes.", + "bbox": [ + 88, + 486, + 482, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inspired by this, we propose an enhanced event-based dense stereo method via cross-sensor knowledge distillation. It obtains a massive amount of texture knowledge from the intensity images in the training stage, while in the inference stage, only events are needed to obtain dense and accurate disparity, greatly reducing the computational burden.", + "bbox": [ + 88, + 532, + 482, + 625 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 638, + 225, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Definition", + "text_level": 1, + "bbox": [ + 89, + 662, + 272, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The events of the left and right event cameras between $t - 1$ and $t$ are denoted as $E_{t}^{L}$ and $E_{t}^{R}$ , respectively, and the corresponding intensity images are indicated as $I_{t}^{L}$ and $I_{t}^{R}$ . The event data is a sequence of four-dimensional vectors $(x,y,p,\\tau)$ , where $(x,y)$ means pixel coordinates, $p\\in \\{-1,1\\}$ indicates polarity, and $\\tau \\in (t - 1,t]$ represents continuous timestamps. We represent event data $E_{t}^{L,R}\\in \\{(x,y,p,\\tau)|t - 1 < \\tau \\leq t\\}$ in voxel grid format following [70]. Specifically, we first scale the timestamps to bin indexes $[0:B - 1]$ , after which the events $E_{t}^{L,R}$ are stacked to generate an voxel grid $V_{t}^{L,R}(b,x,y)\\in R^{B\\times H\\times W}$ with discretized time dimension B, where H and W represent the height and the width, respectively.", + "bbox": [ + 89, + 685, + 482, + 883 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given continuous stereo intensity image pair and stereo", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "event pair, with the same ground truth disparity, the training goal is to input four data in the training stage, which refer to $V_{t}^{L}, V_{t}^{R}, I_{t}^{L}$ , and $I_{t}^{R}$ respectively. By making full use of the information of the two modalities, in the inference phase, only events are needed to predict the dense disparity.", + "bbox": [ + 511, + 487, + 906, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Framework Overview", + "text_level": 1, + "bbox": [ + 511, + 574, + 718, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fig. 2 shows a pipeline of our proposed framework. The framework consists of two identical branches, named the event branch and the intensity branch. The pre-trained intensity branch is used to distill knowledge to the event branch so that it can learn a massive amount of texture knowledge and dense structure information of intensity images. The distillation process includes three levels: shallow distillation, deep distillation, and logit-level distillation.", + "bbox": [ + 511, + 597, + 906, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Shallow features contain much modality-specific information [53, 67], and distilling them into the event branch allows it to obtain complementary cross-sensor information. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polarity and timestamps of events. The simple alignment would contaminate the original modality-specific information and introduce noise. To avoid this problem, we adopt a comprehensive distillation strategy from both long-range and focal perspective, thus the event branch can fully capture both long-range information and local texture details. For the deep features, we adopt a direct alignment strategy", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5439", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "because they mainly contain task-related contextual information. Furthermore, the logit-level distillation makes the disparity prediction of the event branch closer to the inspection prediction of the intensity branch. Besides the distillation strategy, to ensure cross-view consistency, we design an intensity-event joint left-right consistency module. This module attempts to enforce pixel-level consistency of the left intensity image with the warped one, which implicitly indicates the quality of the estimated disparity. In the inference stage, only keeping the event branch enables efficient and dense disparity prediction, and most importantly, preserving the low latency characteristics of events.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Multi-Level Intensity-to-Event Distillation", + "text_level": 1, + "bbox": [ + 89, + 282, + 452, + 297 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed intensity-to-event distillation scheme enables the event branch to learn semantic cues from intensity images. It consists of three levels, including shallow distillation, deep distillation, and logit-level distillation. Shallow and deep distillation are both performed at the feature level, allowing the event branch to fully obtain long-range information, local texture details, and task-related knowledge. The last level drives the event branch to be close to the distribution of dense disparity of the intensity branch.", + "bbox": [ + 89, + 304, + 483, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shallow Distillation. In event-based stereo matching, sparse events lack detailed texture minutiae and rich semantic information, whether from the overall or local area details. As shown in Fig. 3, we extract long-range and focal knowledge from the event and intensity branch respectively, and then align them so that the event branch can obtain both long-range information and local texture details.", + "bbox": [ + 89, + 440, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0de1b51fa0ef79cd57c42585cf0ee3cd99d630a72e9701e2c7754b6271290b2f.jpg", + "image_caption": [ + "Figure 3. Illustration of shallow distillation. Features (cost volumes) of the two modalities are fed to generate relationship maps with long-range guidance. Simultaneous focal distillation is performed via channel attention and spatial attention. Shallow distillation is achieved through these two losses." + ], + "image_footnote": [], + "bbox": [ + 96, + 560, + 478, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long-range information distillation can transfer dark knowledge from the intensity branch to the event branch. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polar", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ity and timestamps of events. To avoid the noise introduced by simple alignment, we use gcblock [5] to extract long-range relations ensuring that two branches have the same long-range relationships. The feature mapping with a long-range relation generated by gcblock is shown below:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM _ {l o n g} (x) = x \\oplus \\left(F \\left(\\ln \\left(F \\left(\\sum_ {i = 1} ^ {N} \\frac {e ^ {F \\left(x _ {i}\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {F \\left(x _ {j}\\right)} x _ {i}}\\right)\\right)\\right)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 176, + 906, + 219 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $F$ denotes the $1 \\times 1$ convolution layer, $N$ means the number of pixels of the feature $x$ , $ln$ indicates the layer normalization, and $\\oplus$ denotes the broadcast element-wise addition. The long-range extraction loss can be obtained by computing the $L_{2}$ norm between feature maps with long-range relationships as follows:", + "bbox": [ + 511, + 227, + 905, + 318 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l o n g} = \\left\\| M _ {l o n g} \\left(f _ {E}\\right) - M _ {l o n g} \\left(f _ {I}\\right) \\right\\| _ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 329, + 906, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{E}$ and $f_{I}$ represent the binocular features of the event branch and the intensity branch, respectively.", + "bbox": [ + 511, + 354, + 905, + 386 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Long-range information distillation achieves the transfer of long-range information but still lacks local texture details. For stereo matching, focusing only on long-range relations is not sufficient to obtain dense disparity. We introduce spatial attention and channel attention to achieve intensity-to-event focal alignment. First, we compute the spatial and channel attention maps as follows:", + "bbox": [ + 511, + 386, + 905, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {s} (x) = H \\cdot W \\cdot s o f t m a x (\\frac {1}{C} \\sum_ {c = 1} ^ {C} | f _ {c} |), \\qquad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 501, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA _ {c} (x) = C \\cdot \\operatorname {s o f t m a x} \\left(\\frac {1}{H \\cdot W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} | f _ {i, j} |\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 553, + 905, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $A_{s}$ and $A_{c}$ represent the spatial attention and channel attention maps derived from the features $f$ . In order to enable the event branch to learn spatial and channel attention from the intensity branch comprehensively, we design the focal loss as follows:", + "bbox": [ + 511, + 602, + 905, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {f o c a l} = \\beta_ {1} L 1 \\left(A _ {s} ^ {E}, A _ {s} ^ {I}\\right) + \\beta_ {2} L 1 \\left(A _ {c} ^ {E}, A _ {c} ^ {I}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 686, + 906, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\beta_{1},\\beta_{2}$ are hyper-parameters to balance the loss, and $L1$ denotes the $L_{1}$ norm.", + "bbox": [ + 511, + 715, + 905, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Deep Distillation. As known from previous studies [48, 67], deep features are high-dimensional abstract representations with fewer modality variances. They retain rich task-related global contextual embeddings. In order to enable the event branch to effectively approximate the modality-independent intensity branch, we employ mean square error loss to facilitate deep distillation as follows:", + "bbox": [ + 511, + 744, + 906, + 851 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d e e p}} = \\sum_ {l} ^ {L} \\left| \\left| F _ {E} ^ {l} - F _ {I} ^ {l} \\right| \\right| _ {2}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 862, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5440", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $l$ is the index where the distillation is performed.", + "bbox": [ + 89, + 90, + 455, + 104 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Logit-Level Distillation. The prediction of the disparity regression layer represents the probability distribution of each modality branch. We propose the logit-level distillation, and then the disparity prediction of the event branch is learned from the soft labels generated by the intensity branch. We use Kullback-Leibler Divergence to measure the similarity in the probability distribution, making the disparity prediction of the event branch closer to the inspection prediction of the intensity branch, formulated as:", + "bbox": [ + 89, + 107, + 483, + 243 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {l o g i t}} = K L \\left(P _ {E}, P _ {I}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 256, + 482, + 273 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In summary, the overall distillation loss $L_{distill}$ is the sum of the long-range loss $L_{long}$ (2) and focal loss $L_{focal}$ (5) in the shallow distillation, the deep distillation loss $L_{deep}$ (6), and the logit-level loss $L_{logit}$ (7), as shown below:", + "bbox": [ + 89, + 286, + 483, + 348 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {d i s t i l l}} = L _ {\\text {l o n g}} + L _ {\\text {f o c a l}} + L _ {\\text {d e e p}} + L _ {\\text {l o g i t}}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 361, + 482, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Intensity-Event Joint Left-Right Consistency", + "text_level": 1, + "bbox": [ + 89, + 390, + 470, + 407 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In order to implicitly indicate the quality of the estimated disparity and bring more explicit prompts, we introduce the Intensity-Event joint Left-Right Consistency module (IE-LRC). For a stereo image pair, right image pixels can be shifted using disparity along each epipolar line to construct a warped left intensity image. Minimizing the difference between the warped left image and the real image reduces the ambiguity of the disparity [19]. The proposed IE-LRC module takes the disparity $D^{L} \\in R^{1 \\times H \\times W}$ predicted by the event branch, the left intensity image $I^{L} \\in R^{1 \\times H \\times W}$ , and the right intensity image $I^{R} \\in R^{1 \\times H \\times W}$ as input, as shown in Fig. 4. For stereo intensity image pairs, the right intensity image pixels are shifted using disparity predicted by the event branch along each epipolar to construct the warped left intensity image $\\hat{I}^{L} \\in R^{1 \\times H \\times W}$ . The local appearance matching loss is constructed to minimize the reconstruction loss between the warped and real intensity image. We utilize both the structural similarity index loss and edge-aware smoothness loss to promote local smoothness of disparity, which can be expressed as follows:", + "bbox": [ + 89, + 412, + 483, + 717 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum_ {i, j} \\left| \\partial_ {x} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {x} I _ {i j} ^ {l} \\right\\|} + \\left| \\partial_ {y} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {y} I _ {i j} ^ {l} \\right\\|}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 739, + 482, + 790 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {l r c} = L _ {s m o o t h} + \\alpha \\frac {1}{N} \\sum_ {i, j} \\frac {1 - S S I M \\left(I _ {i j} ^ {l} , \\hat {I} _ {i j} ^ {l}\\right)}{2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 816, + 482, + 858 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We use a simplified SSIM with a $3 \\times 3$ block filter, and set $\\alpha = 0.85$ , following [19].", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bb36dd1cb9826f473882601499fabab372f89adc6c6cf8a9ec7ccfba24fbf4ae.jpg", + "image_caption": [ + "Figure 4. Illustration of intensity-event joint left-right consistency module. The input disparity comes from the event branch, and the input intensity images come from the intensity branch. The cross-view consistency across modalities is enforced through edge-aware smoothness loss and SSIM loss." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 897, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training and Inference", + "text_level": 1, + "bbox": [ + 513, + 292, + 728, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall Loss for Training. The total training loss of our proposed framework consists of three components, namely the stereo matching loss $L_{disp}$ , the intensity-to-event distillation loss $L_{distill}$ (8), and the left-right consistency loss $L_{lrc}$ (10), as shown below:", + "bbox": [ + 511, + 314, + 905, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {t o t a l} = L _ {d i s p} + L _ {d i s t i l l} + L _ {l r c}. \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 402, + 903, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The stereo matching loss $L_{disp}$ can be defined as:", + "bbox": [ + 513, + 431, + 841, + 448 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {d i s p} (D, D _ {E} ^ {*}) = \\frac {1}{V} \\sum_ {i = 0} ^ {V} S m o o t h _ {L _ {1}} \\left(d _ {i} - d _ {i} ^ {*}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 458, + 903, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D$ is the ground truth disparity, $D_{E}^{*}$ is the disparity predicted by the event branch, and $V$ is the number of valid pixels with ground truths, $d_{i}$ for $i\\in \\{1,2,3,\\dots ,V\\}$ is the pixels in the ground truth disparity $D$ that can be used for training, and $d_{i}^{*}$ is the pixel at the corresponding position in the predicted disparity $D_{E}^{*}$ . The smooth $L_{1}$ loss function can be expressed as follows:", + "bbox": [ + 511, + 511, + 905, + 617 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S m o o t h} _ {L _ {1}} (x) = \\left\\{ \\begin{array}{l l} 0. 5 x ^ {2}, & \\text {i f} | x | < 1, \\\\ | x | - 0. 5, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 628, + 903, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inference. The event branch is fully optimized during training, capturing a massive amount of texture knowledge and dense structure information from intensity images. During inference, retaining only the event branch can achieve efficient and dense disparity prediction without increasing the computational cost, and most importantly, preserving the low latency characteristics of events.", + "bbox": [ + 511, + 672, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 792, + 638, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 816, + 714, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To evaluate the proposed method's effectiveness, the experiments are performed on two publicly available datasets, MVSEC [69] and DSEC [17], both from the real world. The MVSEC dataset has two DAVIS cameras", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5441", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/fb3af3f91612f635ef99ca882a88a4a037a2f06d1973900dc3ac545e2c112d40.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodMean disp. error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓Median depth error [cm] ↓
Split1Spilt3Split1Spilt3Split1Spilt3Split1Spilt3
EIS-E [37]--80.668.313.325.7--
DDES [49]0.590.9489.874.816.727.86.814.7
DTC-PDS [62]0.560.6591.588.715.318.66.48.7
CTC-PDS [62]0.530.7391.688.214.920.66.410.6
EITNet [1]0.550.7592.189.614.219.45.910.4
DTC-SPADE [62]0.460.6093.089.713.517.15.27.9
TES [13]0.460.4992.992.613.015.05.05.8
Ours0.320.3794.893.411.213.24.24.7
", + "bbox": [ + 101, + 88, + 893, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "with stereo setups that provide intensity images and pairing events. We use indoor flying from the MVSEC dataset, which was captured from a drone flying in a room with various objects. Following [10, 13, 49, 62], we split the MVSEC dataset indoor flying into three and use two of them, split 1 and split 3. For a fair comparison, we use mean disparity error, one-pixel accuracy, mean depth error, and median depth error as metrics. The DSEC dataset is a stereo event camera dataset for outdoor driving scenes. The intensity images and events are captured by different devices with different resolutions. All cameras are at the same height, so we approximately map the intensity images to the corresponding positions of the events with the provided camera matrix. For evaluation, we use the mean absolute error (MAE), one-pixel error (1PE), two-pixel error (2PE), and root mean square error (RMSE) as the metrics. See the supplementary material for more details on the dataset split.", + "bbox": [ + 91, + 304, + 483, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. The proposed method is implemented using PyTorch and trained from scratch. We adopt the Adam [27] ( $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.99$ ) optimizer to train the model with a batch size of 8 using $256 \\times 256$ random crops. We set the maximum disparity to 192. The initial learning rate is $1 \\times 10^{-3}$ , and the weight decay is $1 \\times 10^{-4}$ . All experiments are conducted on an NVIDIA RTX A6000 GPU. For more details about the implementation, please refer to the supplementary material.", + "bbox": [ + 89, + 561, + 483, + 699 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 710, + 284, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the results of the proposed $\\mathbf{IED}^2\\mathbf{S}$ with state-of-the-art methods. The comparison results are shown in Table 1. EIS-E [37], DDES [49], DTC-PDS [62], EITNet [1], and TES[13] only use asynchronous and sparse events, which limits the effectiveness of stereo matching due to the lack of sufficient texture knowledge and dense structure information. The proposed $\\mathbf{IED}^2\\mathbf{S}$ fully extracts extensive dense and structural information contained in intensity images and distills it into the event branch during the training, making its performance significantly better than existing event-based stereo matching methods. Specifically,", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9cc6f1e05a6610678a70a28db4a6109babdcc5bfc67feb7132486c3b2a5b129a.jpg", + "table_caption": [ + "Table 1. Disparity estimation results on the MVSEC [69] dataset. - indicates that results are not provided in the original paper. The best is in bold and the second best is in underlined." + ], + "table_footnote": [], + "table_body": "
MethodsMAE ↓1PE ↓2PE ↓RMSE ↓
DDES [49]0.57610.9152.9051.381
DTC-PDS [62]0.5269.5342.3531.263
Se-CFF [38]0.5219.5862.6231.235
Ours0.4938.8232.1341.069
", + "bbox": [ + 514, + 301, + 905, + 398 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Disparity estimation results on the DSEC [17] dataset.", + "bbox": [ + 519, + 405, + 895, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "compared with the best-performing event-based method, TES[13], the proposed method outperforms it by $13.85\\%$ and $12.00\\%$ in terms of mean depth error for split1 and split3, respectively. We believe that this improvement comes from fully considering the modality gap and the novel design of the distillation strategy. Meanwhile, it is worth noting that the combined intensity images and events methods require both events and intensity images to be input at the inference stage, which limits the application of some scenarios, such as high-speed motion scenes or when intensity images are unavailable. However, ours only requires events as input during the inference phase, which fully utilizes the low-latency characteristics of events in special scenarios (high speed and extreme lighting).", + "bbox": [ + 511, + 431, + 906, + 642 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Furthermore, we evaluate the proposed model on the DSEC dataset, and the results are shown in Table 2. Similarly, compared with the state-of-the-art baselines, ours achieves the best performance on all metrics and outperforms existing methods by a large margin, thanks to the designed cross-modal distillation in Sec. 3. Specifically, it outperforms Se-CFF [38] by $5.37\\%$ and $7.46\\%$ in terms of mean absolute error and one-pixel error, respectively, which verifies the effectiveness of the proposed method.", + "bbox": [ + 511, + 643, + 908, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 787, + 694, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we present a qualitative comparison of our method with the baseline in various scenarios on the MVSEC dataset. DDES [49] and TES [13] struggle to output dense disparity due to the lack of detailed information caused by the inherent sparsity of events. They are not ideal in some locations, such as the stripe lines dis", + "bbox": [ + 511, + 810, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5442", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/58cb30f1dfd709951ac92ff87597200830eb71f08dd763da73ab73c153a332d6.jpg", + "image_caption": [ + "(a) Intensity" + ], + "image_footnote": [], + "bbox": [ + 143, + 104, + 259, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/37ed519bf0c3aab5a98a0c72294d7aae677da59230bc058f4fb24273c7888eb3.jpg", + "image_caption": [ + "(b) Events" + ], + "image_footnote": [], + "bbox": [ + 261, + 104, + 377, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5d4d4fa603bdc1d2aa5c16a7a4032593d67c7ebaefeea9b6d5e2df7b683c9309.jpg", + "image_caption": [ + "(c) DDES" + ], + "image_footnote": [], + "bbox": [ + 380, + 104, + 495, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6892d8279cb9b270df1f96c35857e8fe448420c5d2de80668283cbcbb1430858.jpg", + "image_caption": [ + "(d) TES" + ], + "image_footnote": [], + "bbox": [ + 496, + 104, + 612, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/43560092a0f315aac53e7f503813a769e8a90e55b729eddbde288cc0f1b34f9c.jpg", + "image_caption": [ + "(e) Ours" + ], + "image_footnote": [], + "bbox": [ + 614, + 104, + 730, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/39a8daf66e7229aa4a6be50e46228e7c7a104517e1d5e11fd4cfa3833fcdaba7.jpg", + "image_caption": [ + "(f) Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 733, + 104, + 849, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42929fc331a6eafdc51542d0926137ab52682cb5a4eb1d0e94282dc9cf4b3101.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 181, + 258, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e6b5b457e4762ee87d6443594f6fdb049ec003eb0d8ba48e898a69d754c350b5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 181, + 377, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5876d02dc277ea3d3e54e61c7497cf0ae89b4db37e8cc014ba891ce2fa2ab95f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 181, + 495, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1fb4cdc86ef1da8c60fd57272bb0fc518162a135d48e5a356c734ccebdb7947b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 181, + 612, + 256 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a84e706cd9a7e9bfc47cde48ab88d2596e8fbe7e5e236b96fa2dfdf81a8be524.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 181, + 730, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3c51fe3b9432df8432c28c62156e9af10a44248377824b125417c6ba9cfd264.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 181, + 849, + 257 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b74272c89e4772ee380c51224711f604ce2aa6f77146bd4cdfc5617bd81a724.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 143, + 258, + 258, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e718529e166f432e7ad36b7c1eadf5325c587a5f2c64c325a02d1475748c3ce9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 258, + 377, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/551349293e567a1a6737d62d9aa13fea7ea4db1807bbf199e4df9fb52f6e6ac3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 258, + 495, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/014947b9143a97eb457e2a0509b99b9d7bf82c896e3a68c119e6d83c4394ce30.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 258, + 612, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d58eddf5ab90983210cf7f52e303049e393067d62f2af813cf8ac00b488e8ea7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 258, + 730, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da0c009a26bddc00f9b337cb826dac440dc05ae5893c323cf0095c32d902276f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 258, + 849, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8092e78bbe8e8877750e5d3da848cbf2bda2e1de7c6322229c69c7e58df471a.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison of dense disparity estimation for indoor flying scenes in the MVSEC dataset. The first two columns are the intensity images and the corresponding events for visual reference, where the red and blue discrete points represent positive and negative events, respectively. (c), (d), and (e) are the results of DDES [49], TES [13], and ours, respectively. Our proposed method fully uses the intensity image's detail and texture information during the training stage, better displaying fine details in the region highlighted by the red box. Detailed disparity information is framed by the red box for comparison." + ], + "image_footnote": [], + "bbox": [ + 143, + 335, + 258, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d0bc4d6091119dfb1833c5b4ce7d4fd585919feb9a6c14912974c1fb52c49783.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 261, + 335, + 377, + 410 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/067a424c94053d08475bf63ee39263493a094936bc1f71dd9cf1052d21a50093.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 335, + 495, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0db7a3c8dc200af2701023449c63174e342bc7c9844a2dec3b02290f87f0dbf4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 496, + 335, + 612, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45ebc35ede989c22063ab41d59476f6d8e239ff4ceb2dabc15a57d3f1fca5efe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 335, + 730, + 409 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/943260f303efdc1ed17023ce53b9594c1bd4d62cdf4e5510fafa9bc2f942dd1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 733, + 335, + 849, + 409 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "appearing where the disparity should change dramatically. For example, blur appears at the junction of objects such as barrels and chairs with the background. In contrast, our method addresses these challenges by enabling the model to capture more intricate features, thereby offering more reliable information for accurate disparity estimation in ill-posed regions. For example, for a chair placed far away, the proposed method achieves fine-grained disparity prediction. Consequently, even in areas where event data is sparse and closely resembles the surrounding environment, our approach produces disparities that are not only clearer and denser but also exhibit more precise boundaries, such as the area outlined by the red box in Fig. 5.", + "bbox": [ + 89, + 493, + 483, + 690 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e13e1e32e265aeba5dd28b46dad203975f0b4cf1ea5aee7ece49da17b56936be.jpg", + "image_caption": [ + "(a) Intensity" + ], + "image_footnote": [], + "bbox": [ + 112, + 718, + 197, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bb06f6a66af0ab5a77c4e1fcdcf3bcfd64c642e6fd3548cee46cd8e6ec68efda.jpg", + "image_caption": [ + "(b) Sc-CFF" + ], + "image_footnote": [], + "bbox": [ + 200, + 719, + 284, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3b4b9ecc6f8dccce080218c9ea65fe23a1b82dd51f832d2e3d83efa7de2d801.jpg", + "image_caption": [ + "(c) TES" + ], + "image_footnote": [], + "bbox": [ + 285, + 719, + 370, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dc6816f2934fa445ba481159e62fdcc8166dd8005f43e8993582e02c5f3f5403.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 370, + 719, + 455, + 765 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/536721dbf39190a30b3f10f50c64b24e61043025db68a1a8ee3a40520210dca0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 767, + 197, + 814 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dcee0aece8a6b08245878424c5fce1a5bb1037ae4b921ee0a765012a06a04674.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 767, + 284, + 814 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/63e5e939c434f807f199d17dbd2536ba6de73a9015e9dc06bb7c26d273955f95.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 767, + 370, + 814 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2044afabb9c685fe9f62eeb33436b1d8c3cbe090190568668c32d6c5779cd330.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 767, + 455, + 814 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/da3466bc082759a2277aa57255176c8ae03d453448ff1c5eaa992e68cb6903e9.jpg", + "image_caption": [ + "Figure 6. Qualitative results on the DSEC dataset. Our method can construct more dense disparities than others." + ], + "image_footnote": [], + "bbox": [ + 114, + 816, + 197, + 863 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4f4af12a9a16ea7eba34307b40017842a0967681425f458d3ca86543ae2305ac.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 200, + 816, + 284, + 863 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/62a34741376023e3666e55e748a83db532bb751505286888d00a4916178e186a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 285, + 816, + 370, + 863 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0ca9ba52097ce76f5357d9e62279a3796ea7290c93b335cbd684b8f91866cfa1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 816, + 455, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additionally, we present results on the DSEC datasets in Fig. 6. We highlight regions with complex textures, where our method outperforms the others in predicting disparity with higher accuracy. For example, for objects such as trees and guardrails on the roadside, we obtain much sharper, more detailed, and artifact-free disparities.", + "bbox": [ + 511, + 493, + 906, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 590, + 678, + 606 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contribution of each component. We perform extensive experiments on the MVSEC dataset to fully validate the effectiveness of each proposed component and related design choices in Sec. 3. We first present the results of the ablation study in Table 3. It can be observed that all the proposed components have contributed to the superior performance. This indicates that the information derived from intensity images is effective for dense event-based disparity estimation without significantly altering the network structure. When only one component is added, we see the most improvement comes from the feature distillation, which indicates that modality-specific and task-related features of the intensity branch are needed. When it is applied solely, the one-pixel accuracy is increased by $3.74\\%$ and $2.91\\%$ for split1 and split3, respectively. It is worth noting that intensity-event joint left-right consistency improves the one-pixel accuracy of split1 and split3 by $0.22\\%$ and $1.01\\%$ , respectively, which verifies the feasibility of improving the matching effect through ensuring cross-view consistency in", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5443", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bf4f33ec5a2a408f05ef9f0bf63ba922efa65d5da877a5037d4e6ea030e4ee2d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
F-distill.L-distill.LRCMean disparity error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓
Split 1Split 3Split 1Split 3Split1Split3
XXX0.540.6791.089.415.018.5
✓XX0.380.4694.492.012.113.9
X✓X0.490.5391.991.613.817.5
XX✓0.500.6491.290.314.618.0
X✓✓0.490.5292.191.613.717.4
✓X✓0.370.4194.592.711.613.7
✓✓X0.340.3994.893.211.313.4
✓✓✓0.320.3794.893.411.213.2
", + "bbox": [ + 93, + 88, + 897, + 258 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "stereo matching. When all components are combined, the performance is further improved without any conflicts, resulting in the best performance with a significant improvement of the one-pixel accuracy by $4.18\\%$ and $4.47\\%$ for split1 and split3, respectively.", + "bbox": [ + 89, + 306, + 482, + 383 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e80ec51c3c0e1de92b6979de03a102ca2e68bbe09780dfc9102e9cc1a3857966.jpg", + "table_caption": [ + "Table 3. The result of our intensity-to-event distillation scheme with different settings on the MVSEC dataset. F-distill. denotes the feature distillation including shallow and deep distillation, L-distill. represents logit-level distillation, and LRC means left-right consistency." + ], + "table_footnote": [], + "table_body": "
MethodsMean disparity error [pix] ↓One-pixel accuracy [%] ↑
Split1Split3Split1Split3
MSE0.450.5092.291.8
Si2eD0.390.4693.092.5
MSE+Di2eD0.370.4194.093.0
Si2eD+Di2eD0.320.3794.893.4
", + "bbox": [ + 93, + 397, + 478, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature alignment. Although feature distillation has been a widely used loss term in many knowledge distillation studies [28, 39], we are the first to implement that for event-based stereo. For further exploration and discussion, we present our experiments to validate the design choice of the feature distillation module, as it brings the most improvement among all the proposed components. Specifically, to solely examine the effectiveness of our proposed shallow distillation, we compare it to the simple alignment feature-level distillation methods. MSE is a baseline method [22] that simply minimizes the element-wise $L_{2}$ distance between the features of the teacher and student. As shown in Table 4, directly aligning the features of the two modalities contaminates the original modality-specific information and introduces noise to the event branch. Specifically, the one-pixel accuracy is reduced by $2.74\\%$ and $1.71\\%$ for split1 and split3, respectively. A possible reason for this performance gain comes from the large modality difference between events and intensity images, and the novel design fully exploits these modality-specific features while effectively avoiding cross-sensor contamination.", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c22bc89ff60edd7e2ed75dcce41621c6a320481757041aec80c6e6baf5a75c87.jpg", + "table_caption": [ + "Table 4. The effectiveness of our designed feature distillation on the MVSEC dataset. Si2eD represents shallow distillation, and Di2eD denotes deep distillation." + ], + "table_footnote": [], + "table_body": "
SettingMVSEC -> DSECDSEC -> MVSEC
MetricsMAE ↓1PE ↓Mean disp. error [pix] ↓One-pixel acc. [%] ↑
w/o distill.0.78212.4130.7479.8
Full model0.5639.9730.4691.4
", + "bbox": [ + 516, + 303, + 903, + 383 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Cross-validation experimental results between two real-world datasets, verifying good generalization.", + "bbox": [ + 511, + 388, + 903, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Cross-Domain Generalization Performance", + "text_level": 1, + "bbox": [ + 511, + 426, + 880, + 441 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The modality-specific knowledge obtained by cross-sensor distillation is independent of the dataset and has good generalization under different domain distributions. In this section, we discuss the generalization performance of the proposed method on unseen domains. Specifically, we perform cross-validation on two real-world datasets, i.e., the model is trained on MVSEC (DSEC) and evaluated on DSEC (MVSEC). As shown in Table 5, the designed module is less affected by the domain gap. This is because the designed cross-modal distillation enables the event branch to acquire general knowledge for dense stereo matching.", + "bbox": [ + 511, + 450, + 906, + 617 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 633, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. We design a multi-level intensity-to-event distillation strategy, achieving binocular-to-binocular distillation for the first time. A massive amount of texture knowledge and dense structure information contained in the intensity image is distilled to the event branch through the designed strategy. At the same time, an intensity-event joint left-right consistency module is proposed to enforce cross-view consistency. Extensive experiments on the MVSEC and DSEC datasets demonstrate the effectiveness of the proposed framework.", + "bbox": [ + 511, + 660, + 906, + 828 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 844, + 671, + 859 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under grant No.62371450.", + "bbox": [ + 511, + 869, + 903, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5444", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Soikat Hasan Ahmed, Hae Woong Jang, SM Nadim Uddin, and Yong Ju Jung. Deep event stereo leveraged by event-to-image translation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 882-890, 2021. 1, 2, 6", + "[2] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In British Machine Vision Conference, pages 1-11, 2011. 1", + "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \\times 180$ 130 db $3\\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1", + "[4] Luis A Camunas-Mesa, Teresa Serrano-Gotarredona, Sio H Ieng, Ryad B Benosman, and Bernabe Linares-Barranco. On the use of orientation filters for 3d reconstruction in eventdriven stereo vision. Frontiers in neuroscience, 8:48, 2014. 2", + "[5] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4", + "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1", + "[7] Shoushun Chen and Menghan Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1682-1683. IEEE, 2019. 1", + "[8] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. 2", + "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 1", + "[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 1, 2, 6", + "[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In Proceedings of the European Conference on Computer Vision, pages 470-486. Springer, 2022. 1, 2", + "[12] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2", + "[13] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In Proceedings of the European Conference on Computer Vision, pages 294–314. Springer, 2025. 2, 6, 7", + "[14] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learn" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022. 2", + "[15] Weiqin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, David Suter, and Alireza Bab-Hadiashar. Semantic guided long range stereo depth estimation for safer autonomous vehicle applications. IEEE Transactions on Intelligent Transportation Systems, 23(10):18916-18926, 2022. 1", + "[16] Boyu Gao, Haoxiang Lang, and Jing Ren. Stereo visual slam for autonomous vehicles: A review. In IEEE International Conference on Systems, Man, and Cybernetics, pages 1316-1322. IEEE, 2020. 1", + "[17] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 5, 6", + "[18] Suman Ghosh and Guillermo Gallego. Event-based stereo depth estimation: A survey. arXiv preprint arXiv:2409.17680, 2024. 2", + "[19] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 5", + "[20] Rostam Affendi Hamzah, A Fauzan Kadmin, M Saad Hamid, S Fakhar A Ghani, and Haidi Ibrahim. Improvement of stereo matching algorithm for 3d surface reconstruction. Signal Processing: Image Communication, 65:165-172, 2018. 1", + "[21] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[22] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 87-104. Springer, 2022. 2, 8", + "[23] Asmaa Hosni, Michael Bleyer, Margrit Gelautz, and Christoph Rhemann. Local stereo matching using geodesic support weights. In 16th IEEE International Conference on Image Processing, pages 2093-2096. IEEE, 2009. 1", + "[24] Asmaa Hosni, Christoph Rhemann, Michael Bleyer, Carsten Rother, and Margrit Gelautz. Fast cost-volume filtering for visual correspondence and beyond. IEEE transactions on pattern analysis and machine intelligence, 35(2):504-511, 2012. 1", + "[25] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12605-12614, 2020. 2", + "[26] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Émilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2022. 2", + "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5445", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 8", + "[29] Hamid Laga, Laurent Valentin Jospin, Farid Boussaid, and Mohammed Bennamoun. A survey on deep learning techniques for stereo-based depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(4):1738-1764, 2020. 1", + "[30] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 1", + "[31] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1", + "[32] Xineng Li, Chen Zhang, Wanjuan Su, and Wenbing Tao. Inet: Implicit intra-inter information fusion for real-time stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3225-3233, 2024.", + "[33] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI conference on artificial intelligence, pages 1647-1655, 2022. 1", + "[34] Tianshan Liu, Kin-Man Lam, Rui Zhao, and Guoping Qiu. Deep cross-modal representation learning and distillation for illumination-invariant pedestrian detection. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 315-329, 2021. 2", + "[35] Xu Liu, Jianing Li, Jinqiao Shi, Xiaopeng Fan, Yonghong Tian, and Debin Zhao. Event-based monocular depth estimation with recurrent transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1", + "[36] Qian Long, Qiwei Xie, Seiichi Mita, Kazuhisa Ishimaru, and Noriaki Shirai. A real-time dense stereo matching method for critical environment sensing in autonomous driving. In 17th International IEEE Conference on Intelligent Transportation Systems, pages 853-860. IEEE, 2014. 1", + "[37] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021. 2, 6", + "[38] Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concentrate and focus on the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6114-6123, 2022. 2, 6", + "[39] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5007-5016, 2019. 8", + "[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen," + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1", + "[41] Mary Phuong and Christoph H Lampert. Distillation-based training for multi-exit architectures. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1355-1364, 2019. 2", + "[42] Ewa Piatkowska, Ahmed Belbachir, and Margrit Gelautz. Asynchronous stereo vision for event-driven dynamic stereo sensor using an adaptive cooperative approach. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 45-50, 2013. 2", + "[43] Ewa Piatkowska, Jurgen Kogler, Nabil Belbachir, and Margrit Gelautz. Improved cooperative stereo matching for dynamic vision sensors with ground truth evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 53-60, 2017.", + "[44] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision, 126(12):1394-1414, 2018. 1", + "[45] Paul Rogister, Ryad Benosman, Sio-Hoi Ieng, Patrick Lichtsteiner, and Tobi Delbruck. Asynchronous event-based binocular stereo matching. IEEE Transactions on Neural Networks and Learning Systems, 23(2):347-353, 2011. 2", + "[46] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47:7-42, 2002. 1", + "[47] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 2024. 1", + "[48] Pin Tang, Hai-Ming Xu, and Chao Ma. Prototransfer: Cross-modal prototype transfer for point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3337-3347, 2023. 2, 4", + "[49] Stepan Tulyakov, Francois Fleuret, Martin Kiefel, Peter Gehler, and Michael Hirsch. Learning an event sequence embedding for dense event-based deep stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1527-1537, 2019. 2, 6, 7", + "[50] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1365-1374, 2019. 2", + "[51] SM Nadim Uddin, Soikat Hasan Ahmed, and Yong Ju Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 2", + "[52] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5446", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8637-8646, 2023. 2, 3", + "[54] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 1", + "[55] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12981-12990, 2022.", + "[56] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023.", + "[57] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1", + "[58] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European conference on computer vision, pages 677-695. Springer, 2022. 2", + "[59] Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2019. 1", + "[60] Shanxin Yuan, Bjorn Stenger, and Tae-Kyun Kim. 3d hand pose estimation from rgb using privileged learning with depth data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2", + "[61] Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Low-cost multispectral scene analysis with modality distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 803-812, 2022. 2", + "[62] Kaixuan Zhang, Kaiwei Che, Jianguo Zhang, Jie Cheng, Ziyang Zhang, Qinghai Guo, and Luziwei Leng. Discrete time convolution for fast event-based stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8676-8686, 2022. 2, 6", + "[63] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3713-3722, 2019. 2", + "[64] Bo Zhao, Ruoxi Ding, Shoushun Chen, Bernabe Linares-Barranco, and Huajin Tang. Feedforward categorization on" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "aer motion events using cortex-like features in a spiking neural network. IEEE transactions on neural networks and learning systems, 26(9):1963-1978, 2014. 1", + "[65] Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao. Simdistill: Simulated multimodal distillation for bev 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 7460-7468, 2024. 2", + "[66] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 2", + "[67] Runkai Zhao, Heng Wang, and Weidong Cai. Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer. In ACM Multimedia, 2024. 3, 4", + "[68] Alex Zihao Zhu, Yibo Chen, and Kostas Daniilidis. Realtime time synchronized event-based stereo. In Proceedings of the European Conference on Computer Vision, pages 433-447, 2018. 2", + "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 5, 6", + "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3", + "[71] Dongqing Zou, Feng Shi, Wei-Heng Liu, Jia Li, Qiang Wang, Paul KJ Park, and Hyunsurk Ryu. Robust dense depth maps generations from sparse dvs stereos. In British Machine Vision Conference, 2017. 2" + ], + "bbox": [ + 516, + 92, + 906, + 571 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5447", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_model.json b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_model.json new file mode 100644 index 0000000000000000000000000000000000000000..aa0c24ee4003e387f0e0a0323550f1014bc788c9 --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_model.json @@ -0,0 +1,2884 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.131, + 0.888, + 0.154 + ], + "angle": 0, + "content": "Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.179, + 0.804, + 0.22 + ], + "angle": 0, + "content": "Haihao Zhang\\(^{1,2}\\) Yunjian Zhang\\(^{3\\dagger}\\) Jianing Li\\(^{3\\dagger}\\) Lin Zhu\\(^{2}\\) Meng Lv\\(^{2}\\) Yao Zhu\\(^{3}\\) Yanwei Liu\\(^{1}\\) Xiangyang Ji\\(^{3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.217, + 0.764, + 0.253 + ], + "angle": 0, + "content": "\\(^{1}\\)Institute of Information Engineering, Chinese Academy of Sciences \\(^{2}\\)Beijing Institute of Technology \\(^{3}\\)Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.215, + 0.255, + 0.778, + 0.27 + ], + "angle": 0, + "content": "haihaozhang@bit.edu.cn sdtczyj@gmail.com lijianing@pku.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.32, + 0.327, + 0.335 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.352, + 0.485, + 0.685 + ], + "angle": 0, + "content": "Accurate stereo matching under fast motion and extreme lighting conditions is a challenge for many vision applications. Event cameras have the advantages of low latency and high dynamic range, thus providing a reliable solution to this challenge. However, since events are sparse, this makes it an ill-posed problem to obtain dense disparity using only events. In this work, we propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. Specifically, a multi-level intensity-to-event distillation strategy is designed to maximize the potential of long-range information, local texture details, and task-related knowledge of the intensity images. Simultaneously, to enforce the cross-view consistency, an intensity-event joint left-right consistency module is proposed. With our framework, extensive dense and structural information contained in intensity images is distilled to the event branch. Therefore, retaining only the events can predict dense disparities during inference, preserving the low latency characteristics of the events. Adequate experiments conducted on the MVSEC and DSEC datasets demonstrate that our method exhibits superior stereo matching performance than baselines, both quantitatively and qualitatively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.714, + 0.222, + 0.729 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Depth estimation based on binocular stereo vision is a classic problem in computer vision [15, 16, 20, 24, 29, 36, 44, 46, 59], which aims to determine the correspondence between points in a rectified image pair. Deep learning-based stereo matching methods perform well on various public benchmarks [2, 6, 9, 23, 24, 31-33, 52, 54-57]. However, due to the shortcomings of frame-based sensors in terms of frame rate and dynamic range, it is very difficult to apply traditional cameras in more challenging scenarios such as" + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.309, + 0.665, + 0.321 + ], + "angle": 0, + "content": "(a) Event-based Method" + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.325, + 0.897, + 0.373 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.375, + 0.722, + 0.387 + ], + "angle": 0, + "content": "(b) Event-Intensity Fusion Method" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.39, + 0.897, + 0.498 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.503, + 0.669, + 0.515 + ], + "angle": 0, + "content": "(c) Our Proposed Method" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.522, + 0.899, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.665, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Figure 1. Comparisons on stereo matching frameworks (event-based, event-intensity fusion methods) and our proposed IED\\(^2\\)S that presents an effective intensity-to-event distillation strategy. Texture and dense knowledge is transferred to the event branch, leading to dense stereo matching in the inference stage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.804, + 0.765 + ], + "angle": 0, + "content": "high speed and extreme lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.906, + 0.884 + ], + "angle": 0, + "content": "Event cameras [3, 7, 30, 35, 40, 47, 64] are bio-inspired sensors and report intensity changes asynchronously in the form of an event stream, with each containing spatial location, polarity, and timestamps. They have the advantages of low latency and high dynamic range, making them immune to high-speed motion blur and can operate in extreme lighting scenarios. Therefore, they offer a satisfying solution to overcome the limitations of frame-based sensors." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Recent event-based stereo matching studies [1, 10, 11," + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.393, + 0.901 + ], + "angle": 0, + "content": "† Corresponding authors: Jianing Li, Yunjian Zhang" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5437" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "13, 18, 37, 38, 49] embed asynchronous event data into image-like representations, then fine-grained features are extracted and subsequent disparity regression is performed. While disparity can be effectively estimated, since event cameras can only perceive changes in intensity, when the scene is stationary or there is little relative motion between the camera and the scene, the event stream will be very sparse. This makes it an ill-posed problem to obtain dense disparity maps using only sparse asynchronous event data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.482, + 0.426 + ], + "angle": 0, + "content": "Considering the dense features provided by intensity images, there raises a trend to combine events and intensity images for dense stereo matching [10, 11, 37]. They design algorithms that simultaneously input intensity images and events, using their complementary features to calculate disparities, and the results are outstanding advances for event-based stereo matching. Nevertheless, these approaches require the events and intensity images simultaneously during the inference stage, suffering from two problems: (1) The low latency characteristics of events is sacrificed unintentionally; (2) Processing both modalities concurrently consumes more computing resources, which is a heavy burden for real-time applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.429, + 0.483, + 0.79 + ], + "angle": 0, + "content": "To address the above problems, we propose a novel framework named Intensity-to-Event Distillation for Dense Stereo matching method (IED\\(^2\\)S), which explores the potential of an efficient event-based stereo model for dense disparity prediction. As shown in Fig. 1(c), we implement cross-sensor distillations from the intensity branch to the event branch. The dense and structural information of the intensity image provides useful regularization for asynchronous events, especially for ill-posed regions. In our framework, we design a novel feature alignment strategy, achieving binocular-to-binocular distillation for the first time. Specifically, a comprehensive multi-level intensity-to-event distillation strategy is designed, fully leveraging the modality-specific characteristics of shallow features and the task-related contextual understanding involved in deep abstract representations. Moreover, a warping-based intensity-event joint left-right consistency module is designed to ensure cross-view consistency. The intensity images are only used in the training stage. For inference, only event data is needed for obtaining dense disparity. Compared with fusion-based methods shown in Fig. 1(b), it has greater flexibility, preserving the low latency characteristics of event data and greatly reducing the inference burden. In summary, the main contributions of this work are:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.482, + 0.868 + ], + "angle": 0, + "content": "1) We propose a dense event-based stereo matching method, \\(\\mathbf{IED}^2\\mathbf{S}\\), which is the first trial to implement binocular intensity-to-event distillation. The proposed \\(\\mathbf{IED}^2\\mathbf{S}\\) enables the event branch to learn accurate dense information, greatly improving its disparity prediction capability." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "2) We design a multi-level intensity-to-event distillation, fully distilling long-range information, local texture details," + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.794, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "and task-related knowledge into the event branch while avoiding the contamination of the original event modality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.123, + 0.905, + 0.168 + ], + "angle": 0, + "content": "3) Extensive experiments demonstrate that the proposed \\(\\mathbf{IED}^2\\mathbf{S}\\) achieves significantly better performance than the prior works for event-based stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.184, + 0.655, + 0.199 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.21, + 0.7, + 0.224 + ], + "angle": 0, + "content": "2.1. Event-Based Stereo" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.232, + 0.905, + 0.307 + ], + "angle": 0, + "content": "Early attempts used traditional hand-crafted methods [4, 12, 42-45, 68, 71] to determine the correspondence between two event streams. However, there are no fixed patterns in the event data, which makes it troublesome to use traditional methods to obtain dense disparity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.595 + ], + "angle": 0, + "content": "Recent learning-based approaches [1, 10, 11, 13, 37, 38, 49, 51] are able to estimate disparities using sparse events. The network usually includes feature extraction, cost aggregation, disparity regression, and disparity refinement modules. Tulyakov et al. [49] proposed a 4D queue embedding method that includes the spatiotemporal information of event data. Nam et al. [38] proposed an attention-based event concentration network to display scene details by omitting fewer details without covering events. DTC [62] used discrete time convolution modules to aggregate sequential event information at the feature level, accelerating the stereo matching process at a lower computational cost. Cho et al. [13] proposed the concept of stereo flow to aggregate past features and cost volumes. Accurate stereo matching is achieved by training the stereo matching and stereo flow simultaneously. However, it is an ill-posed problem to obtain dense disparity using only sparse asynchronous events. We therefore explore cross-sensor distillation for event-based dense stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.607, + 0.741, + 0.621 + ], + "angle": 0, + "content": "2.2. Cross-Sensor Distillation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Knowledge distillation is an effective method for compressing models while maintaining accuracy [21, 41, 50, 63]. In particular, cross-sensor knowledge distillation has received increasing attention and applied in many tasks, such as 3D object detection [8, 14, 22, 53, 65, 66], multispectral detection [34, 61], 3D hand pose estimation [60], and LiDAR semantic segmentation [25, 26, 48, 58]. Typically, a well-trained teacher model guides the student model's feature extraction or result prediction that receives input from other modalities. Specifically, Hong et al. [22] first proposed cross-sensor knowledge distillation based on BEV, mimicking LiDAR BEV's representation to achieve improved performance. Chong et al. [14] achieved the transfer of structural cues by projecting point clouds onto the plane of the image and maintaining similar local region affinities. Liu et al. [34] adopted different losses at multiple task levels to transfer coarse-grained and fine-grained features covering feature, detection, and segmentation. Tang et al. [48] estab" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5438" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.895, + 0.423 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.43, + 0.908, + 0.475 + ], + "angle": 0, + "content": "Figure 2. The pipeline of the proposed Intensity-to-Event Distillation for Dense Stereo matching method (IED\\(^2\\)S). The proposed approach enables the event branch to learn a massive amount of texture knowledge and dense structure information under the guidance of the intensity branch. After training, the intensity branch is removed without introducing additional computational burden during inference." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.487, + 0.483, + 0.533 + ], + "angle": 0, + "content": "lished a prototype-like library from strictly aligned fused features, and encouraged all point cloud features to learn from the prototypes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.533, + 0.483, + 0.625 + ], + "angle": 0, + "content": "Inspired by this, we propose an enhanced event-based dense stereo method via cross-sensor knowledge distillation. It obtains a massive amount of texture knowledge from the intensity images in the training stage, while in the inference stage, only events are needed to obtain dense and accurate disparity, greatly reducing the computational burden." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.639, + 0.227, + 0.656 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.664, + 0.274, + 0.679 + ], + "angle": 0, + "content": "3.1. Problem Definition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.686, + 0.483, + 0.885 + ], + "angle": 0, + "content": "The events of the left and right event cameras between \\( t - 1 \\) and \\( t \\) are denoted as \\( E_{t}^{L} \\) and \\( E_{t}^{R} \\), respectively, and the corresponding intensity images are indicated as \\( I_{t}^{L} \\) and \\( I_{t}^{R} \\). The event data is a sequence of four-dimensional vectors \\( (x,y,p,\\tau) \\), where \\( (x,y) \\) means pixel coordinates, \\( p\\in \\{-1,1\\} \\) indicates polarity, and \\( \\tau \\in (t - 1,t] \\) represents continuous timestamps. We represent event data \\( E_{t}^{L,R}\\in \\{(x,y,p,\\tau)|t - 1 < \\tau \\leq t\\} \\) in voxel grid format following [70]. Specifically, we first scale the timestamps to bin indexes \\( [0:B - 1] \\), after which the events \\( E_{t}^{L,R} \\) are stacked to generate an voxel grid \\( V_{t}^{L,R}(b,x,y)\\in R^{B\\times H\\times W} \\) with discretized time dimension B, where H and W represent the height and the width, respectively." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given continuous stereo intensity image pair and stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.565 + ], + "angle": 0, + "content": "event pair, with the same ground truth disparity, the training goal is to input four data in the training stage, which refer to \\( V_{t}^{L}, V_{t}^{R}, I_{t}^{L} \\), and \\( I_{t}^{R} \\) respectively. By making full use of the information of the two modalities, in the inference phase, only events are needed to predict the dense disparity." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.575, + 0.719, + 0.59 + ], + "angle": 0, + "content": "3.2. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Fig. 2 shows a pipeline of our proposed framework. The framework consists of two identical branches, named the event branch and the intensity branch. The pre-trained intensity branch is used to distill knowledge to the event branch so that it can learn a massive amount of texture knowledge and dense structure information of intensity images. The distillation process includes three levels: shallow distillation, deep distillation, and logit-level distillation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Shallow features contain much modality-specific information [53, 67], and distilling them into the event branch allows it to obtain complementary cross-sensor information. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polarity and timestamps of events. The simple alignment would contaminate the original modality-specific information and introduce noise. To avoid this problem, we adopt a comprehensive distillation strategy from both long-range and focal perspective, thus the event branch can fully capture both long-range information and local texture details. For the deep features, we adopt a direct alignment strategy" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5439" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "because they mainly contain task-related contextual information. Furthermore, the logit-level distillation makes the disparity prediction of the event branch closer to the inspection prediction of the intensity branch. Besides the distillation strategy, to ensure cross-view consistency, we design an intensity-event joint left-right consistency module. This module attempts to enforce pixel-level consistency of the left intensity image with the warped one, which implicitly indicates the quality of the estimated disparity. In the inference stage, only keeping the event branch enables efficient and dense disparity prediction, and most importantly, preserving the low latency characteristics of events." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.283, + 0.453, + 0.299 + ], + "angle": 0, + "content": "3.3. Multi-Level Intensity-to-Event Distillation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.305, + 0.484, + 0.44 + ], + "angle": 0, + "content": "The proposed intensity-to-event distillation scheme enables the event branch to learn semantic cues from intensity images. It consists of three levels, including shallow distillation, deep distillation, and logit-level distillation. Shallow and deep distillation are both performed at the feature level, allowing the event branch to fully obtain long-range information, local texture details, and task-related knowledge. The last level drives the event branch to be close to the distribution of dense disparity of the intensity branch." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.441, + 0.484, + 0.547 + ], + "angle": 0, + "content": "Shallow Distillation. In event-based stereo matching, sparse events lack detailed texture minutiae and rich semantic information, whether from the overall or local area details. As shown in Fig. 3, we extract long-range and focal knowledge from the event and intensity branch respectively, and then align them so that the event branch can obtain both long-range information and local texture details." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.561, + 0.48, + 0.749 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.76, + 0.483, + 0.83 + ], + "angle": 0, + "content": "Figure 3. Illustration of shallow distillation. Features (cost volumes) of the two modalities are fed to generate relationship maps with long-range guidance. Simultaneous focal distillation is performed via channel attention and spatial attention. Shallow distillation is achieved through these two losses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Long-range information distillation can transfer dark knowledge from the intensity branch to the event branch. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "ity and timestamps of events. To avoid the noise introduced by simple alignment, we use gcblock [5] to extract long-range relations ensuring that two branches have the same long-range relationships. The feature mapping with a long-range relation generated by gcblock is shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.177, + 0.907, + 0.22 + ], + "angle": 0, + "content": "\\[\nM _ {l o n g} (x) = x \\oplus \\left(F \\left(\\ln \\left(F \\left(\\sum_ {i = 1} ^ {N} \\frac {e ^ {F \\left(x _ {i}\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {F \\left(x _ {j}\\right)} x _ {i}}\\right)\\right)\\right)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.906, + 0.319 + ], + "angle": 0, + "content": "where \\(F\\) denotes the \\(1 \\times 1\\) convolution layer, \\(N\\) means the number of pixels of the feature \\(x\\), \\(ln\\) indicates the layer normalization, and \\(\\oplus\\) denotes the broadcast element-wise addition. The long-range extraction loss can be obtained by computing the \\(L_{2}\\) norm between feature maps with long-range relationships as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.33, + 0.907, + 0.348 + ], + "angle": 0, + "content": "\\[\nL _ {l o n g} = \\left\\| M _ {l o n g} \\left(f _ {E}\\right) - M _ {l o n g} \\left(f _ {I}\\right) \\right\\| _ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.356, + 0.906, + 0.387 + ], + "angle": 0, + "content": "where \\( f_{E} \\) and \\( f_{I} \\) represent the binocular features of the event branch and the intensity branch, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Long-range information distillation achieves the transfer of long-range information but still lacks local texture details. For stereo matching, focusing only on long-range relations is not sufficient to obtain dense disparity. We introduce spatial attention and channel attention to achieve intensity-to-event focal alignment. First, we compute the spatial and channel attention maps as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.502, + 0.907, + 0.545 + ], + "angle": 0, + "content": "\\[\nA _ {s} (x) = H \\cdot W \\cdot s o f t m a x (\\frac {1}{C} \\sum_ {c = 1} ^ {C} | f _ {c} |), \\qquad (3)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.556, + 0.554, + 0.906, + 0.597 + ], + "angle": 0, + "content": "\\[\nA _ {c} (x) = C \\cdot \\operatorname {s o f t m a x} \\left(\\frac {1}{H \\cdot W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} | f _ {i, j} |\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.603, + 0.906, + 0.678 + ], + "angle": 0, + "content": "where \\( A_{s} \\) and \\( A_{c} \\) represent the spatial attention and channel attention maps derived from the features \\( f \\). In order to enable the event branch to learn spatial and channel attention from the intensity branch comprehensively, we design the focal loss as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.687, + 0.907, + 0.707 + ], + "angle": 0, + "content": "\\[\nL _ {f o c a l} = \\beta_ {1} L 1 \\left(A _ {s} ^ {E}, A _ {s} ^ {I}\\right) + \\beta_ {2} L 1 \\left(A _ {c} ^ {E}, A _ {c} ^ {I}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.716, + 0.906, + 0.746 + ], + "angle": 0, + "content": "where \\(\\beta_{1},\\beta_{2}\\) are hyper-parameters to balance the loss, and \\(L1\\) denotes the \\(L_{1}\\) norm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.746, + 0.907, + 0.852 + ], + "angle": 0, + "content": "Deep Distillation. As known from previous studies [48, 67], deep features are high-dimensional abstract representations with fewer modality variances. They retain rich task-related global contextual embeddings. In order to enable the event branch to effectively approximate the modality-independent intensity branch, we employ mean square error loss to facilitate deep distillation as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.863, + 0.907, + 0.904 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d e e p}} = \\sum_ {l} ^ {L} \\left| \\left| F _ {E} ^ {l} - F _ {I} ^ {l} \\right| \\right| _ {2}, \\tag {6}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5440" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.457, + 0.106 + ], + "angle": 0, + "content": "where \\(l\\) is the index where the distillation is performed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.108, + 0.484, + 0.244 + ], + "angle": 0, + "content": "Logit-Level Distillation. The prediction of the disparity regression layer represents the probability distribution of each modality branch. We propose the logit-level distillation, and then the disparity prediction of the event branch is learned from the soft labels generated by the intensity branch. We use Kullback-Leibler Divergence to measure the similarity in the probability distribution, making the disparity prediction of the event branch closer to the inspection prediction of the intensity branch, formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.257, + 0.483, + 0.274 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {l o g i t}} = K L \\left(P _ {E}, P _ {I}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.349 + ], + "angle": 0, + "content": "In summary, the overall distillation loss \\( L_{distill} \\) is the sum of the long-range loss \\( L_{long} \\) (2) and focal loss \\( L_{focal} \\) (5) in the shallow distillation, the deep distillation loss \\( L_{deep} \\) (6), and the logit-level loss \\( L_{logit} \\) (7), as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.362, + 0.483, + 0.379 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {d i s t i l l}} = L _ {\\text {l o n g}} + L _ {\\text {f o c a l}} + L _ {\\text {d e e p}} + L _ {\\text {l o g i t}}. \\tag {8}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.471, + 0.408 + ], + "angle": 0, + "content": "3.4. Intensity-Event Joint Left-Right Consistency" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.414, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In order to implicitly indicate the quality of the estimated disparity and bring more explicit prompts, we introduce the Intensity-Event joint Left-Right Consistency module (IE-LRC). For a stereo image pair, right image pixels can be shifted using disparity along each epipolar line to construct a warped left intensity image. Minimizing the difference between the warped left image and the real image reduces the ambiguity of the disparity [19]. The proposed IE-LRC module takes the disparity \\( D^{L} \\in R^{1 \\times H \\times W} \\) predicted by the event branch, the left intensity image \\( I^{L} \\in R^{1 \\times H \\times W} \\), and the right intensity image \\( I^{R} \\in R^{1 \\times H \\times W} \\) as input, as shown in Fig. 4. For stereo intensity image pairs, the right intensity image pixels are shifted using disparity predicted by the event branch along each epipolar to construct the warped left intensity image \\( \\hat{I}^{L} \\in R^{1 \\times H \\times W} \\). The local appearance matching loss is constructed to minimize the reconstruction loss between the warped and real intensity image. We utilize both the structural similarity index loss and edge-aware smoothness loss to promote local smoothness of disparity, which can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.741, + 0.483, + 0.791 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum_ {i, j} \\left| \\partial_ {x} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {x} I _ {i j} ^ {l} \\right\\|} + \\left| \\partial_ {y} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {y} I _ {i j} ^ {l} \\right\\|}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.817, + 0.483, + 0.859 + ], + "angle": 0, + "content": "\\[\nL _ {l r c} = L _ {s m o o t h} + \\alpha \\frac {1}{N} \\sum_ {i, j} \\frac {1 - S S I M \\left(I _ {i j} ^ {l} , \\hat {I} _ {i j} ^ {l}\\right)}{2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We use a simplified SSIM with a \\(3 \\times 3\\) block filter, and set \\(\\alpha = 0.85\\), following [19]." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.898, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.209, + 0.906, + 0.279 + ], + "angle": 0, + "content": "Figure 4. Illustration of intensity-event joint left-right consistency module. The input disparity comes from the event branch, and the input intensity images come from the intensity branch. The cross-view consistency across modalities is enforced through edge-aware smoothness loss and SSIM loss." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.293, + 0.729, + 0.309 + ], + "angle": 0, + "content": "3.5. Training and Inference" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.315, + 0.906, + 0.39 + ], + "angle": 0, + "content": "Overall Loss for Training. The total training loss of our proposed framework consists of three components, namely the stereo matching loss \\( L_{disp} \\), the intensity-to-event distillation loss \\( L_{distill} \\) (8), and the left-right consistency loss \\( L_{lrc} \\) (10), as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.404, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\nL _ {t o t a l} = L _ {d i s p} + L _ {d i s t i l l} + L _ {l r c}. \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.432, + 0.843, + 0.449 + ], + "angle": 0, + "content": "The stereo matching loss \\( L_{disp} \\) can be defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.459, + 0.905, + 0.502 + ], + "angle": 0, + "content": "\\[\nL _ {d i s p} (D, D _ {E} ^ {*}) = \\frac {1}{V} \\sum_ {i = 0} ^ {V} S m o o t h _ {L _ {1}} \\left(d _ {i} - d _ {i} ^ {*}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.906, + 0.618 + ], + "angle": 0, + "content": "where \\(D\\) is the ground truth disparity, \\(D_{E}^{*}\\) is the disparity predicted by the event branch, and \\(V\\) is the number of valid pixels with ground truths, \\(d_{i}\\) for \\(i\\in \\{1,2,3,\\dots ,V\\}\\) is the pixels in the ground truth disparity \\(D\\) that can be used for training, and \\(d_{i}^{*}\\) is the pixel at the corresponding position in the predicted disparity \\(D_{E}^{*}\\). The smooth \\(L_{1}\\) loss function can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.542, + 0.629, + 0.905, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S m o o t h} _ {L _ {1}} (x) = \\left\\{ \\begin{array}{l l} 0. 5 x ^ {2}, & \\text {i f} | x | < 1, \\\\ | x | - 0. 5, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.779 + ], + "angle": 0, + "content": "Inference. The event branch is fully optimized during training, capturing a massive amount of texture knowledge and dense structure information from intensity images. During inference, retaining only the event branch can achieve efficient and dense disparity prediction without increasing the computational cost, and most importantly, preserving the low latency characteristics of events." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.794, + 0.64, + 0.81 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.715, + 0.835 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. To evaluate the proposed method's effectiveness, the experiments are performed on two publicly available datasets, MVSEC [69] and DSEC [17], both from the real world. The MVSEC dataset has two DAVIS cameras" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5441" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.102, + 0.089, + 0.895, + 0.26 + ], + "angle": 0, + "content": "
MethodMean disp. error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓Median depth error [cm] ↓
Split1Spilt3Split1Spilt3Split1Spilt3Split1Spilt3
EIS-E [37]--80.668.313.325.7--
DDES [49]0.590.9489.874.816.727.86.814.7
DTC-PDS [62]0.560.6591.588.715.318.66.48.7
CTC-PDS [62]0.530.7391.688.214.920.66.410.6
EITNet [1]0.550.7592.189.614.219.45.910.4
DTC-SPADE [62]0.460.6093.089.713.517.15.27.9
TES [13]0.460.4992.992.613.015.05.05.8
Ours0.320.3794.893.411.213.24.24.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Table 1. Disparity estimation results on the MVSEC [69] dataset. - indicates that results are not provided in the original paper. The best is in bold and the second best is in underlined." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.305, + 0.485, + 0.562 + ], + "angle": 0, + "content": "with stereo setups that provide intensity images and pairing events. We use indoor flying from the MVSEC dataset, which was captured from a drone flying in a room with various objects. Following [10, 13, 49, 62], we split the MVSEC dataset indoor flying into three and use two of them, split 1 and split 3. For a fair comparison, we use mean disparity error, one-pixel accuracy, mean depth error, and median depth error as metrics. The DSEC dataset is a stereo event camera dataset for outdoor driving scenes. The intensity images and events are captured by different devices with different resolutions. All cameras are at the same height, so we approximately map the intensity images to the corresponding positions of the events with the provided camera matrix. For evaluation, we use the mean absolute error (MAE), one-pixel error (1PE), two-pixel error (2PE), and root mean square error (RMSE) as the metrics. See the supplementary material for more details on the dataset split." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.7 + ], + "angle": 0, + "content": "Implementation Details. The proposed method is implemented using PyTorch and trained from scratch. We adopt the Adam [27] (\\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.99\\)) optimizer to train the model with a batch size of 8 using \\(256 \\times 256\\) random crops. We set the maximum disparity to 192. The initial learning rate is \\(1 \\times 10^{-3}\\), and the weight decay is \\(1 \\times 10^{-4}\\). All experiments are conducted on an NVIDIA RTX A6000 GPU. For more details about the implementation, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.285, + 0.727 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare the results of the proposed \\(\\mathbf{IED}^2\\mathbf{S}\\) with state-of-the-art methods. The comparison results are shown in Table 1. EIS-E [37], DDES [49], DTC-PDS [62], EITNet [1], and TES[13] only use asynchronous and sparse events, which limits the effectiveness of stereo matching due to the lack of sufficient texture knowledge and dense structure information. The proposed \\(\\mathbf{IED}^2\\mathbf{S}\\) fully extracts extensive dense and structural information contained in intensity images and distills it into the event branch during the training, making its performance significantly better than existing event-based stereo matching methods. Specifically," + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.302, + 0.906, + 0.4 + ], + "angle": 0, + "content": "
MethodsMAE ↓1PE ↓2PE ↓RMSE ↓
DDES [49]0.57610.9152.9051.381
DTC-PDS [62]0.5269.5342.3531.263
Se-CFF [38]0.5219.5862.6231.235
Ours0.4938.8232.1341.069
" + }, + { + "type": "table_caption", + "bbox": [ + 0.521, + 0.406, + 0.897, + 0.421 + ], + "angle": 0, + "content": "Table 2. Disparity estimation results on the DSEC [17] dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.907, + 0.643 + ], + "angle": 0, + "content": "compared with the best-performing event-based method, TES[13], the proposed method outperforms it by \\(13.85\\%\\) and \\(12.00\\%\\) in terms of mean depth error for split1 and split3, respectively. We believe that this improvement comes from fully considering the modality gap and the novel design of the distillation strategy. Meanwhile, it is worth noting that the combined intensity images and events methods require both events and intensity images to be input at the inference stage, which limits the application of some scenarios, such as high-speed motion scenes or when intensity images are unavailable. However, ours only requires events as input during the inference phase, which fully utilizes the low-latency characteristics of events in special scenarios (high speed and extreme lighting)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Furthermore, we evaluate the proposed model on the DSEC dataset, and the results are shown in Table 2. Similarly, compared with the state-of-the-art baselines, ours achieves the best performance on all metrics and outperforms existing methods by a large margin, thanks to the designed cross-modal distillation in Sec. 3. Specifically, it outperforms Se-CFF [38] by \\(5.37\\%\\) and \\(7.46\\%\\) in terms of mean absolute error and one-pixel error, respectively, which verifies the effectiveness of the proposed method." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.789, + 0.696, + 0.804 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 5, we present a qualitative comparison of our method with the baseline in various scenarios on the MVSEC dataset. DDES [49] and TES [13] struggle to output dense disparity due to the lack of detailed information caused by the inherent sparsity of events. They are not ideal in some locations, such as the stripe lines dis" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5442" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.175, + 0.095, + 0.235, + 0.104 + ], + "angle": 0, + "content": "(a) Intensity" + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.105, + 0.26, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.296, + 0.095, + 0.345, + 0.104 + ], + "angle": 0, + "content": "(b) Events" + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.105, + 0.379, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.095, + 0.461, + 0.104 + ], + "angle": 0, + "content": "(c) DDES" + }, + { + "type": "image", + "bbox": [ + 0.381, + 0.105, + 0.496, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.537, + 0.095, + 0.576, + 0.104 + ], + "angle": 0, + "content": "(d) TES" + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.105, + 0.614, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.654, + 0.095, + 0.694, + 0.104 + ], + "angle": 0, + "content": "(e) Ours" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.105, + 0.732, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.752, + 0.095, + 0.832, + 0.104 + ], + "angle": 0, + "content": "(f) Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.105, + 0.85, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.182, + 0.259, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.182, + 0.378, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.182, + 0.496, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.182, + 0.614, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.182, + 0.732, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.182, + 0.85, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.26, + 0.259, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.259, + 0.378, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.259, + 0.496, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.259, + 0.614, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.259, + 0.732, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.259, + 0.85, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.144, + 0.337, + 0.259, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.262, + 0.336, + 0.378, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.336, + 0.496, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.498, + 0.336, + 0.614, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.336, + 0.732, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.735, + 0.337, + 0.85, + 0.41 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.415, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison of dense disparity estimation for indoor flying scenes in the MVSEC dataset. The first two columns are the intensity images and the corresponding events for visual reference, where the red and blue discrete points represent positive and negative events, respectively. (c), (d), and (e) are the results of DDES [49], TES [13], and ours, respectively. Our proposed method fully uses the intensity image's detail and texture information during the training stage, better displaying fine details in the region highlighted by the red box. Detailed disparity information is framed by the red box for comparison." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.691 + ], + "angle": 0, + "content": "appearing where the disparity should change dramatically. For example, blur appears at the junction of objects such as barrels and chairs with the background. In contrast, our method addresses these challenges by enabling the model to capture more intricate features, thereby offering more reliable information for accurate disparity estimation in ill-posed regions. For example, for a chair placed far away, the proposed method achieves fine-grained disparity prediction. Consequently, even in areas where event data is sparse and closely resembles the surrounding environment, our approach produces disparities that are not only clearer and denser but also exhibit more precise boundaries, such as the area outlined by the red box in Fig. 5." + }, + { + "type": "image_caption", + "bbox": [ + 0.133, + 0.711, + 0.182, + 0.719 + ], + "angle": 0, + "content": "(a) Intensity" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.719, + 0.199, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.222, + 0.712, + 0.267, + 0.719 + ], + "angle": 0, + "content": "(b) Sc-CFF" + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.72, + 0.285, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.311, + 0.712, + 0.341, + 0.719 + ], + "angle": 0, + "content": "(c) TES" + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.72, + 0.371, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.397, + 0.712, + 0.432, + 0.719 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.72, + 0.456, + 0.766 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.768, + 0.199, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.202, + 0.768, + 0.285, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.768, + 0.371, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.768, + 0.456, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.115, + 0.817, + 0.199, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.201, + 0.817, + 0.285, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.287, + 0.817, + 0.371, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.817, + 0.456, + 0.864 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.869, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on the DSEC dataset. Our method can construct more dense disparities than others." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.585 + ], + "angle": 0, + "content": "Additionally, we present results on the DSEC datasets in Fig. 6. We highlight regions with complex textures, where our method outperforms the others in predicting disparity with higher accuracy. For example, for objects such as trees and guardrails on the roadside, we obtain much sharper, more detailed, and artifact-free disparities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.679, + 0.607 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Contribution of each component. We perform extensive experiments on the MVSEC dataset to fully validate the effectiveness of each proposed component and related design choices in Sec. 3. We first present the results of the ablation study in Table 3. It can be observed that all the proposed components have contributed to the superior performance. This indicates that the information derived from intensity images is effective for dense event-based disparity estimation without significantly altering the network structure. When only one component is added, we see the most improvement comes from the feature distillation, which indicates that modality-specific and task-related features of the intensity branch are needed. When it is applied solely, the one-pixel accuracy is increased by \\(3.74\\%\\) and \\(2.91\\%\\) for split1 and split3, respectively. It is worth noting that intensity-event joint left-right consistency improves the one-pixel accuracy of split1 and split3 by \\(0.22\\%\\) and \\(1.01\\%\\), respectively, which verifies the feasibility of improving the matching effect through ensuring cross-view consistency in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5443" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.898, + 0.26 + ], + "angle": 0, + "content": "
F-distill.L-distill.LRCMean disparity error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓
Split 1Split 3Split 1Split 3Split1Split3
XXX0.540.6791.089.415.018.5
✓XX0.380.4694.492.012.113.9
X✓X0.490.5391.991.613.817.5
XX✓0.500.6491.290.314.618.0
X✓✓0.490.5292.191.613.717.4
✓X✓0.370.4194.592.711.613.7
✓✓X0.340.3994.893.211.313.4
✓✓✓0.320.3794.893.411.213.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.264, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Table 3. The result of our intensity-to-event distillation scheme with different settings on the MVSEC dataset. F-distill. denotes the feature distillation including shallow and deep distillation, L-distill. represents logit-level distillation, and LRC means left-right consistency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.483, + 0.385 + ], + "angle": 0, + "content": "stereo matching. When all components are combined, the performance is further improved without any conflicts, resulting in the best performance with a significant improvement of the one-pixel accuracy by \\(4.18\\%\\) and \\(4.47\\%\\) for split1 and split3, respectively." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.398, + 0.479, + 0.525 + ], + "angle": 0, + "content": "
MethodsMean disparity error [pix] ↓One-pixel accuracy [%] ↑
Split1Split3Split1Split3
MSE0.450.5092.291.8
Si2eD0.390.4693.092.5
MSE+Di2eD0.370.4194.093.0
Si2eD+Di2eD0.320.3794.893.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.53, + 0.483, + 0.573 + ], + "angle": 0, + "content": "Table 4. The effectiveness of our designed feature distillation on the MVSEC dataset. Si2eD represents shallow distillation, and Di2eD denotes deep distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Feature alignment. Although feature distillation has been a widely used loss term in many knowledge distillation studies [28, 39], we are the first to implement that for event-based stereo. For further exploration and discussion, we present our experiments to validate the design choice of the feature distillation module, as it brings the most improvement among all the proposed components. Specifically, to solely examine the effectiveness of our proposed shallow distillation, we compare it to the simple alignment feature-level distillation methods. MSE is a baseline method [22] that simply minimizes the element-wise \\( L_{2} \\) distance between the features of the teacher and student. As shown in Table 4, directly aligning the features of the two modalities contaminates the original modality-specific information and introduces noise to the event branch. Specifically, the one-pixel accuracy is reduced by \\( 2.74\\% \\) and \\( 1.71\\% \\) for split1 and split3, respectively. A possible reason for this performance gain comes from the large modality difference between events and intensity images, and the novel design fully exploits these modality-specific features while effectively avoiding cross-sensor contamination." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.384 + ], + "angle": 0, + "content": "
SettingMVSEC -> DSECDSEC -> MVSEC
MetricsMAE ↓1PE ↓Mean disp. error [pix] ↓One-pixel acc. [%] ↑
w/o distill.0.78212.4130.7479.8
Full model0.5639.9730.4691.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.389, + 0.905, + 0.419 + ], + "angle": 0, + "content": "Table 5. Cross-validation experimental results between two real-world datasets, verifying good generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.428, + 0.882, + 0.443 + ], + "angle": 0, + "content": "4.5. Cross-Domain Generalization Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.618 + ], + "angle": 0, + "content": "The modality-specific knowledge obtained by cross-sensor distillation is independent of the dataset and has good generalization under different domain distributions. In this section, we discuss the generalization performance of the proposed method on unseen domains. Specifically, we perform cross-validation on two real-world datasets, i.e., the model is trained on MVSEC (DSEC) and evaluated on DSEC (MVSEC). As shown in Table 5, the designed module is less affected by the domain gap. This is because the designed cross-modal distillation enables the event branch to acquire general knowledge for dense stereo matching." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.634, + 0.634, + 0.65 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.829 + ], + "angle": 0, + "content": "We propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. We design a multi-level intensity-to-event distillation strategy, achieving binocular-to-binocular distillation for the first time. A massive amount of texture knowledge and dense structure information contained in the intensity image is distilled to the event branch through the designed strategy. At the same time, an intensity-event joint left-right consistency module is proposed to enforce cross-view consistency. Extensive experiments on the MVSEC and DSEC datasets demonstrate the effectiveness of the proposed framework." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.845, + 0.673, + 0.861 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.903 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under grant No.62371450." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5444" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Soikat Hasan Ahmed, Hae Woong Jang, SM Nadim Uddin, and Yong Ju Jung. Deep event stereo leveraged by event-to-image translation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 882-890, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.225 + ], + "angle": 0, + "content": "[2] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In British Machine Vision Conference, pages 1-11, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A \\(240 \\times 180\\) 130 db \\(3\\mu s\\) latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.352 + ], + "angle": 0, + "content": "[4] Luis A Camunas-Mesa, Teresa Serrano-Gotarredona, Sio H Ieng, Ryad B Benosman, and Bernabe Linares-Barranco. On the use of orientation filters for 3d reconstruction in eventdriven stereo vision. Frontiers in neuroscience, 8:48, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[5] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[7] Shoushun Chen and Menghan Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1682-1683. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.59 + ], + "angle": 0, + "content": "[8] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.661 + ], + "angle": 0, + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.719, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In Proceedings of the European Conference on Computer Vision, pages 470-486. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[12] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In Proceedings of the European Conference on Computer Vision, pages 294–314. Springer, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ing spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.124, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[15] Weiqin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, David Suter, and Alireza Bab-Hadiashar. Semantic guided long range stereo depth estimation for safer autonomous vehicle applications. IEEE Transactions on Intelligent Transportation Systems, 23(10):18916-18926, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.195, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[16] Boyu Gao, Haoxiang Lang, and Jing Ren. Stereo visual slam for autonomous vehicles: A review. In IEEE International Conference on Systems, Man, and Cybernetics, pages 1316-1322. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.253, + 0.904, + 0.307 + ], + "angle": 0, + "content": "[17] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.351 + ], + "angle": 0, + "content": "[18] Suman Ghosh and Guillermo Gallego. Event-based stereo depth estimation: A survey. arXiv preprint arXiv:2409.17680, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.905, + 0.422 + ], + "angle": 0, + "content": "[19] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[20] Rostam Affendi Hamzah, A Fauzan Kadmin, M Saad Hamid, S Fakhar A Ghani, and Haidi Ibrahim. Improvement of stereo matching algorithm for 3d surface reconstruction. Signal Processing: Image Communication, 65:165-172, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[21] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[22] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 87-104. Springer, 2022. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[23] Asmaa Hosni, Michael Bleyer, Margrit Gelautz, and Christoph Rhemann. Local stereo matching using geodesic support weights. In 16th IEEE International Conference on Image Processing, pages 2093-2096. IEEE, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.905, + 0.726 + ], + "angle": 0, + "content": "[24] Asmaa Hosni, Christoph Rhemann, Michael Bleyer, Carsten Rother, and Margrit Gelautz. Fast cost-volume filtering for visual correspondence and beyond. IEEE transactions on pattern analysis and machine intelligence, 35(2):504-511, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[25] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12605-12614, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[26] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Émilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5445" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[28] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[29] Hamid Laga, Laurent Valentin Jospin, Farid Boussaid, and Mohammed Bennamoun. A survey on deep learning techniques for stereo-based depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(4):1738-1764, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[30] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] Xineng Li, Chen Zhang, Wanjuan Su, and Wenbing Tao. Inet: Implicit intra-inter information fusion for real-time stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3225-3233, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[33] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI conference on artificial intelligence, pages 1647-1655, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[34] Tianshan Liu, Kin-Man Lam, Rui Zhao, and Guoping Qiu. Deep cross-modal representation learning and distillation for illumination-invariant pedestrian detection. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 315-329, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[35] Xu Liu, Jianing Li, Jinqiao Shi, Xiaopeng Fan, Yonghong Tian, and Debin Zhao. Event-based monocular depth estimation with recurrent transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[36] Qian Long, Qiwei Xie, Seiichi Mita, Kazuhisa Ishimaru, and Noriaki Shirai. A real-time dense stereo matching method for critical environment sensing in autonomous driving. In 17th International IEEE Conference on Intelligent Transportation Systems, pages 853-860. IEEE, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[37] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[38] Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concentrate and focus on the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6114-6123, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[39] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5007-5016, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen," + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.218 + ], + "angle": 0, + "content": "[41] Mary Phuong and Christoph H Lampert. Distillation-based training for multi-exit architectures. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1355-1364, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[42] Ewa Piatkowska, Ahmed Belbachir, and Margrit Gelautz. Asynchronous stereo vision for event-driven dynamic stereo sensor using an adaptive cooperative approach. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 45-50, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[43] Ewa Piatkowska, Jurgen Kogler, Nabil Belbachir, and Margrit Gelautz. Improved cooperative stereo matching for dynamic vision sensors with ground truth evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 53-60, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[44] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision, 126(12):1394-1414, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[45] Paul Rogister, Ryad Benosman, Sio-Hoi Ieng, Patrick Lichtsteiner, and Tobi Delbruck. Asynchronous event-based binocular stereo matching. IEEE Transactions on Neural Networks and Learning Systems, 23(2):347-353, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[46] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47:7-42, 2002. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[47] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.905, + 0.65 + ], + "angle": 0, + "content": "[48] Pin Tang, Hai-Ming Xu, and Chao Ma. Prototransfer: Cross-modal prototype transfer for point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3337-3347, 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.652, + 0.905, + 0.719 + ], + "angle": 0, + "content": "[49] Stepan Tulyakov, Francois Fleuret, Martin Kiefel, Peter Gehler, and Michael Hirsch. Learning an event sequence embedding for dense event-based deep stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1527-1537, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.721, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[50] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1365-1374, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[51] SM Nadim Uddin, Soikat Hasan Ahmed, and Yong Ju Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.833, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "5446" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[53] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8637-8646, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[54] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[55] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12981-12990, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[56] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[57] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[58] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European conference on computer vision, pages 677-695. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.519, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[59] Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[60] Shanxin Yuan, Bjorn Stenger, and Tae-Kyun Kim. 3d hand pose estimation from rgb using privileged learning with depth data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[61] Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Low-cost multispectral scene analysis with modality distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 803-812, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[62] Kaixuan Zhang, Kaiwei Che, Jianguo Zhang, Jie Cheng, Ziyang Zhang, Qinghai Guo, and Luziwei Leng. Discrete time convolution for fast event-based stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8676-8686, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[63] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3713-3722, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[64] Bo Zhao, Ruoxi Ding, Shoushun Chen, Bernabe Linares-Barranco, and Huajin Tang. Feedforward categorization on" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "aer motion events using cortex-like features in a spiking neural network. IEEE transactions on neural networks and learning systems, 26(9):1963-1978, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[65] Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao. Simdistill: Simulated multimodal distillation for bev 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 7460-7468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[66] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[67] Runkai Zhao, Heng Wang, and Weidong Cai. Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer. In ACM Multimedia, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[68] Alex Zihao Zhu, Yibo Chen, and Kostas Daniilidis. Realtime time synchronized event-based stereo. In Proceedings of the European Conference on Computer Vision, pages 433-447, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.907, + 0.515 + ], + "angle": 0, + "content": "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.517, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[71] Dongqing Zou, Feng Shi, Wei-Heng Liu, Jia Li, Qiang Wang, Paul KJ Park, and Hyunsurk Ryu. Robust dense depth maps generations from sparse dvs stereos. In British Machine Vision Conference, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.572 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5447" + } + ] +] \ No newline at end of file diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_origin.pdf b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..084f58ccba2faa6e14c1998b65ff9ed1d91fb301 --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/f6b9298e-6b4d-4df1-b0c3-014f084ff091_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:aeddb7482ab1011e59ba54ff4e10e90364e9dce3b7e2963943ce4c5ae8f2915b +size 1730026 diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/full.md b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..354c8ebccf3a63e624790112d6794c81230478c1 --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/full.md @@ -0,0 +1,415 @@ +# Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation + +Haihao Zhang $^{1,2}$ Yunjian Zhang $^{3\dagger}$ Jianing Li $^{3\dagger}$ Lin Zhu $^{2}$ Meng Lv $^{2}$ Yao Zhu $^{3}$ Yanwei Liu $^{1}$ Xiangyang Ji $^{3}$ + +$^{1}$ Institute of Information Engineering, Chinese Academy of Sciences $^{2}$ Beijing Institute of Technology $^{3}$ Tsinghua University + +haihaozhang@bit.edu.cn sdtczyj@gmail.com lijianing@pku.edu.cn + +# Abstract + +Accurate stereo matching under fast motion and extreme lighting conditions is a challenge for many vision applications. Event cameras have the advantages of low latency and high dynamic range, thus providing a reliable solution to this challenge. However, since events are sparse, this makes it an ill-posed problem to obtain dense disparity using only events. In this work, we propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. Specifically, a multi-level intensity-to-event distillation strategy is designed to maximize the potential of long-range information, local texture details, and task-related knowledge of the intensity images. Simultaneously, to enforce the cross-view consistency, an intensity-event joint left-right consistency module is proposed. With our framework, extensive dense and structural information contained in intensity images is distilled to the event branch. Therefore, retaining only the events can predict dense disparities during inference, preserving the low latency characteristics of the events. Adequate experiments conducted on the MVSEC and DSEC datasets demonstrate that our method exhibits superior stereo matching performance than baselines, both quantitatively and qualitatively. + +# 1. Introduction + +Depth estimation based on binocular stereo vision is a classic problem in computer vision [15, 16, 20, 24, 29, 36, 44, 46, 59], which aims to determine the correspondence between points in a rectified image pair. Deep learning-based stereo matching methods perform well on various public benchmarks [2, 6, 9, 23, 24, 31-33, 52, 54-57]. However, due to the shortcomings of frame-based sensors in terms of frame rate and dynamic range, it is very difficult to apply traditional cameras in more challenging scenarios such as + +![](images/b89b1ea31765c0322264d0b680275c6d4170439a658e2bba8cc46e1dd580d9b8.jpg) +(a) Event-based Method + +![](images/e7334d2c544679a3eac1aa96ae1da71845bbf97f4ab5d71784764ed71c9187f0.jpg) +(b) Event-Intensity Fusion Method + +![](images/da46527f76d8da6623574e2a3331938f3948b3844f218d84fdd922ae91755e64.jpg) +(c) Our Proposed Method +Figure 1. Comparisons on stereo matching frameworks (event-based, event-intensity fusion methods) and our proposed IED $^2$ S that presents an effective intensity-to-event distillation strategy. Texture and dense knowledge is transferred to the event branch, leading to dense stereo matching in the inference stage. + +high speed and extreme lighting conditions. + +Event cameras [3, 7, 30, 35, 40, 47, 64] are bio-inspired sensors and report intensity changes asynchronously in the form of an event stream, with each containing spatial location, polarity, and timestamps. They have the advantages of low latency and high dynamic range, making them immune to high-speed motion blur and can operate in extreme lighting scenarios. Therefore, they offer a satisfying solution to overcome the limitations of frame-based sensors. + +Recent event-based stereo matching studies [1, 10, 11, + +13, 18, 37, 38, 49] embed asynchronous event data into image-like representations, then fine-grained features are extracted and subsequent disparity regression is performed. While disparity can be effectively estimated, since event cameras can only perceive changes in intensity, when the scene is stationary or there is little relative motion between the camera and the scene, the event stream will be very sparse. This makes it an ill-posed problem to obtain dense disparity maps using only sparse asynchronous event data. + +Considering the dense features provided by intensity images, there raises a trend to combine events and intensity images for dense stereo matching [10, 11, 37]. They design algorithms that simultaneously input intensity images and events, using their complementary features to calculate disparities, and the results are outstanding advances for event-based stereo matching. Nevertheless, these approaches require the events and intensity images simultaneously during the inference stage, suffering from two problems: (1) The low latency characteristics of events is sacrificed unintentionally; (2) Processing both modalities concurrently consumes more computing resources, which is a heavy burden for real-time applications. + +To address the above problems, we propose a novel framework named Intensity-to-Event Distillation for Dense Stereo matching method (IED $^2$ S), which explores the potential of an efficient event-based stereo model for dense disparity prediction. As shown in Fig. 1(c), we implement cross-sensor distillations from the intensity branch to the event branch. The dense and structural information of the intensity image provides useful regularization for asynchronous events, especially for ill-posed regions. In our framework, we design a novel feature alignment strategy, achieving binocular-to-binocular distillation for the first time. Specifically, a comprehensive multi-level intensity-to-event distillation strategy is designed, fully leveraging the modality-specific characteristics of shallow features and the task-related contextual understanding involved in deep abstract representations. Moreover, a warping-based intensity-event joint left-right consistency module is designed to ensure cross-view consistency. The intensity images are only used in the training stage. For inference, only event data is needed for obtaining dense disparity. Compared with fusion-based methods shown in Fig. 1(b), it has greater flexibility, preserving the low latency characteristics of event data and greatly reducing the inference burden. In summary, the main contributions of this work are: + +1) We propose a dense event-based stereo matching method, $\mathbf{IED}^2\mathbf{S}$ , which is the first trial to implement binocular intensity-to-event distillation. The proposed $\mathbf{IED}^2\mathbf{S}$ enables the event branch to learn accurate dense information, greatly improving its disparity prediction capability. +2) We design a multi-level intensity-to-event distillation, fully distilling long-range information, local texture details, + +and task-related knowledge into the event branch while avoiding the contamination of the original event modality. + +3) Extensive experiments demonstrate that the proposed $\mathbf{IED}^2\mathbf{S}$ achieves significantly better performance than the prior works for event-based stereo matching. + +# 2. Related Work + +# 2.1. Event-Based Stereo + +Early attempts used traditional hand-crafted methods [4, 12, 42-45, 68, 71] to determine the correspondence between two event streams. However, there are no fixed patterns in the event data, which makes it troublesome to use traditional methods to obtain dense disparity. + +Recent learning-based approaches [1, 10, 11, 13, 37, 38, 49, 51] are able to estimate disparities using sparse events. The network usually includes feature extraction, cost aggregation, disparity regression, and disparity refinement modules. Tulyakov et al. [49] proposed a 4D queue embedding method that includes the spatiotemporal information of event data. Nam et al. [38] proposed an attention-based event concentration network to display scene details by omitting fewer details without covering events. DTC [62] used discrete time convolution modules to aggregate sequential event information at the feature level, accelerating the stereo matching process at a lower computational cost. Cho et al. [13] proposed the concept of stereo flow to aggregate past features and cost volumes. Accurate stereo matching is achieved by training the stereo matching and stereo flow simultaneously. However, it is an ill-posed problem to obtain dense disparity using only sparse asynchronous events. We therefore explore cross-sensor distillation for event-based dense stereo matching. + +# 2.2. Cross-Sensor Distillation + +Knowledge distillation is an effective method for compressing models while maintaining accuracy [21, 41, 50, 63]. In particular, cross-sensor knowledge distillation has received increasing attention and applied in many tasks, such as 3D object detection [8, 14, 22, 53, 65, 66], multispectral detection [34, 61], 3D hand pose estimation [60], and LiDAR semantic segmentation [25, 26, 48, 58]. Typically, a well-trained teacher model guides the student model's feature extraction or result prediction that receives input from other modalities. Specifically, Hong et al. [22] first proposed cross-sensor knowledge distillation based on BEV, mimicking LiDAR BEV's representation to achieve improved performance. Chong et al. [14] achieved the transfer of structural cues by projecting point clouds onto the plane of the image and maintaining similar local region affinities. Liu et al. [34] adopted different losses at multiple task levels to transfer coarse-grained and fine-grained features covering feature, detection, and segmentation. Tang et al. [48] estab + +![](images/5cb4d299c5a05b21db20b69663dfa9d4d855dbd173833363253a0d9de51376c3.jpg) +Figure 2. The pipeline of the proposed Intensity-to-Event Distillation for Dense Stereo matching method (IED $^2$ S). The proposed approach enables the event branch to learn a massive amount of texture knowledge and dense structure information under the guidance of the intensity branch. After training, the intensity branch is removed without introducing additional computational burden during inference. + +lished a prototype-like library from strictly aligned fused features, and encouraged all point cloud features to learn from the prototypes. + +Inspired by this, we propose an enhanced event-based dense stereo method via cross-sensor knowledge distillation. It obtains a massive amount of texture knowledge from the intensity images in the training stage, while in the inference stage, only events are needed to obtain dense and accurate disparity, greatly reducing the computational burden. + +# 3. Methodology + +# 3.1. Problem Definition + +The events of the left and right event cameras between $t - 1$ and $t$ are denoted as $E_{t}^{L}$ and $E_{t}^{R}$ , respectively, and the corresponding intensity images are indicated as $I_{t}^{L}$ and $I_{t}^{R}$ . The event data is a sequence of four-dimensional vectors $(x,y,p,\tau)$ , where $(x,y)$ means pixel coordinates, $p\in \{-1,1\}$ indicates polarity, and $\tau \in (t - 1,t]$ represents continuous timestamps. We represent event data $E_{t}^{L,R}\in \{(x,y,p,\tau)|t - 1 < \tau \leq t\}$ in voxel grid format following [70]. Specifically, we first scale the timestamps to bin indexes $[0:B - 1]$ , after which the events $E_{t}^{L,R}$ are stacked to generate an voxel grid $V_{t}^{L,R}(b,x,y)\in R^{B\times H\times W}$ with discretized time dimension B, where H and W represent the height and the width, respectively. + +Given continuous stereo intensity image pair and stereo + +event pair, with the same ground truth disparity, the training goal is to input four data in the training stage, which refer to $V_{t}^{L}, V_{t}^{R}, I_{t}^{L}$ , and $I_{t}^{R}$ respectively. By making full use of the information of the two modalities, in the inference phase, only events are needed to predict the dense disparity. + +# 3.2. Framework Overview + +Fig. 2 shows a pipeline of our proposed framework. The framework consists of two identical branches, named the event branch and the intensity branch. The pre-trained intensity branch is used to distill knowledge to the event branch so that it can learn a massive amount of texture knowledge and dense structure information of intensity images. The distillation process includes three levels: shallow distillation, deep distillation, and logit-level distillation. + +Shallow features contain much modality-specific information [53, 67], and distilling them into the event branch allows it to obtain complementary cross-sensor information. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polarity and timestamps of events. The simple alignment would contaminate the original modality-specific information and introduce noise. To avoid this problem, we adopt a comprehensive distillation strategy from both long-range and focal perspective, thus the event branch can fully capture both long-range information and local texture details. For the deep features, we adopt a direct alignment strategy + +because they mainly contain task-related contextual information. Furthermore, the logit-level distillation makes the disparity prediction of the event branch closer to the inspection prediction of the intensity branch. Besides the distillation strategy, to ensure cross-view consistency, we design an intensity-event joint left-right consistency module. This module attempts to enforce pixel-level consistency of the left intensity image with the warped one, which implicitly indicates the quality of the estimated disparity. In the inference stage, only keeping the event branch enables efficient and dense disparity prediction, and most importantly, preserving the low latency characteristics of events. + +# 3.3. Multi-Level Intensity-to-Event Distillation + +The proposed intensity-to-event distillation scheme enables the event branch to learn semantic cues from intensity images. It consists of three levels, including shallow distillation, deep distillation, and logit-level distillation. Shallow and deep distillation are both performed at the feature level, allowing the event branch to fully obtain long-range information, local texture details, and task-related knowledge. The last level drives the event branch to be close to the distribution of dense disparity of the intensity branch. + +Shallow Distillation. In event-based stereo matching, sparse events lack detailed texture minutiae and rich semantic information, whether from the overall or local area details. As shown in Fig. 3, we extract long-range and focal knowledge from the event and intensity branch respectively, and then align them so that the event branch can obtain both long-range information and local texture details. + +![](images/0de1b51fa0ef79cd57c42585cf0ee3cd99d630a72e9701e2c7754b6271290b2f.jpg) +Figure 3. Illustration of shallow distillation. Features (cost volumes) of the two modalities are fed to generate relationship maps with long-range guidance. Simultaneous focal distillation is performed via channel attention and spatial attention. Shallow distillation is achieved through these two losses. + +Long-range information distillation can transfer dark knowledge from the intensity branch to the event branch. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polar + +ity and timestamps of events. To avoid the noise introduced by simple alignment, we use gcblock [5] to extract long-range relations ensuring that two branches have the same long-range relationships. The feature mapping with a long-range relation generated by gcblock is shown below: + +$$ +M _ {l o n g} (x) = x \oplus \left(F \left(\ln \left(F \left(\sum_ {i = 1} ^ {N} \frac {e ^ {F \left(x _ {i}\right)}}{\sum_ {j = 1} ^ {N} e ^ {F \left(x _ {j}\right)} x _ {i}}\right)\right)\right)\right), \tag {1} +$$ + +where $F$ denotes the $1 \times 1$ convolution layer, $N$ means the number of pixels of the feature $x$ , $ln$ indicates the layer normalization, and $\oplus$ denotes the broadcast element-wise addition. The long-range extraction loss can be obtained by computing the $L_{2}$ norm between feature maps with long-range relationships as follows: + +$$ +L _ {l o n g} = \left\| M _ {l o n g} \left(f _ {E}\right) - M _ {l o n g} \left(f _ {I}\right) \right\| _ {2}, \tag {2} +$$ + +where $f_{E}$ and $f_{I}$ represent the binocular features of the event branch and the intensity branch, respectively. + +Long-range information distillation achieves the transfer of long-range information but still lacks local texture details. For stereo matching, focusing only on long-range relations is not sufficient to obtain dense disparity. We introduce spatial attention and channel attention to achieve intensity-to-event focal alignment. First, we compute the spatial and channel attention maps as follows: + +$$ +A _ {s} (x) = H \cdot W \cdot s o f t m a x (\frac {1}{C} \sum_ {c = 1} ^ {C} | f _ {c} |), \qquad (3) +$$ + +$$ +A _ {c} (x) = C \cdot \operatorname {s o f t m a x} \left(\frac {1}{H \cdot W} \sum_ {i = 1} ^ {H} \sum_ {j = 1} ^ {W} | f _ {i, j} |\right), \tag {4} +$$ + +where $A_{s}$ and $A_{c}$ represent the spatial attention and channel attention maps derived from the features $f$ . In order to enable the event branch to learn spatial and channel attention from the intensity branch comprehensively, we design the focal loss as follows: + +$$ +L _ {f o c a l} = \beta_ {1} L 1 \left(A _ {s} ^ {E}, A _ {s} ^ {I}\right) + \beta_ {2} L 1 \left(A _ {c} ^ {E}, A _ {c} ^ {I}\right), \tag {5} +$$ + +where $\beta_{1},\beta_{2}$ are hyper-parameters to balance the loss, and $L1$ denotes the $L_{1}$ norm. + +Deep Distillation. As known from previous studies [48, 67], deep features are high-dimensional abstract representations with fewer modality variances. They retain rich task-related global contextual embeddings. In order to enable the event branch to effectively approximate the modality-independent intensity branch, we employ mean square error loss to facilitate deep distillation as follows: + +$$ +L _ {\text {d e e p}} = \sum_ {l} ^ {L} \left| \left| F _ {E} ^ {l} - F _ {I} ^ {l} \right| \right| _ {2}, \tag {6} +$$ + +where $l$ is the index where the distillation is performed. + +Logit-Level Distillation. The prediction of the disparity regression layer represents the probability distribution of each modality branch. We propose the logit-level distillation, and then the disparity prediction of the event branch is learned from the soft labels generated by the intensity branch. We use Kullback-Leibler Divergence to measure the similarity in the probability distribution, making the disparity prediction of the event branch closer to the inspection prediction of the intensity branch, formulated as: + +$$ +L _ {\text {l o g i t}} = K L \left(P _ {E}, P _ {I}\right). \tag {7} +$$ + +In summary, the overall distillation loss $L_{distill}$ is the sum of the long-range loss $L_{long}$ (2) and focal loss $L_{focal}$ (5) in the shallow distillation, the deep distillation loss $L_{deep}$ (6), and the logit-level loss $L_{logit}$ (7), as shown below: + +$$ +L _ {\text {d i s t i l l}} = L _ {\text {l o n g}} + L _ {\text {f o c a l}} + L _ {\text {d e e p}} + L _ {\text {l o g i t}}. \tag {8} +$$ + +# 3.4. Intensity-Event Joint Left-Right Consistency + +In order to implicitly indicate the quality of the estimated disparity and bring more explicit prompts, we introduce the Intensity-Event joint Left-Right Consistency module (IE-LRC). For a stereo image pair, right image pixels can be shifted using disparity along each epipolar line to construct a warped left intensity image. Minimizing the difference between the warped left image and the real image reduces the ambiguity of the disparity [19]. The proposed IE-LRC module takes the disparity $D^{L} \in R^{1 \times H \times W}$ predicted by the event branch, the left intensity image $I^{L} \in R^{1 \times H \times W}$ , and the right intensity image $I^{R} \in R^{1 \times H \times W}$ as input, as shown in Fig. 4. For stereo intensity image pairs, the right intensity image pixels are shifted using disparity predicted by the event branch along each epipolar to construct the warped left intensity image $\hat{I}^{L} \in R^{1 \times H \times W}$ . The local appearance matching loss is constructed to minimize the reconstruction loss between the warped and real intensity image. We utilize both the structural similarity index loss and edge-aware smoothness loss to promote local smoothness of disparity, which can be expressed as follows: + +$$ +L _ {\text {s m o o t h}} = \frac {1}{N} \sum_ {i, j} \left| \partial_ {x} D _ {i j} ^ {l} \right| e ^ {- \left\| \partial_ {x} I _ {i j} ^ {l} \right\|} + \left| \partial_ {y} D _ {i j} ^ {l} \right| e ^ {- \left\| \partial_ {y} I _ {i j} ^ {l} \right\|}, \tag {9} +$$ + +$$ +L _ {l r c} = L _ {s m o o t h} + \alpha \frac {1}{N} \sum_ {i, j} \frac {1 - S S I M \left(I _ {i j} ^ {l} , \hat {I} _ {i j} ^ {l}\right)}{2}. \tag {10} +$$ + +We use a simplified SSIM with a $3 \times 3$ block filter, and set $\alpha = 0.85$ , following [19]. + +![](images/bb36dd1cb9826f473882601499fabab372f89adc6c6cf8a9ec7ccfba24fbf4ae.jpg) +Figure 4. Illustration of intensity-event joint left-right consistency module. The input disparity comes from the event branch, and the input intensity images come from the intensity branch. The cross-view consistency across modalities is enforced through edge-aware smoothness loss and SSIM loss. + +# 3.5. Training and Inference + +Overall Loss for Training. The total training loss of our proposed framework consists of three components, namely the stereo matching loss $L_{disp}$ , the intensity-to-event distillation loss $L_{distill}$ (8), and the left-right consistency loss $L_{lrc}$ (10), as shown below: + +$$ +L _ {t o t a l} = L _ {d i s p} + L _ {d i s t i l l} + L _ {l r c}. \tag {11} +$$ + +The stereo matching loss $L_{disp}$ can be defined as: + +$$ +L _ {d i s p} (D, D _ {E} ^ {*}) = \frac {1}{V} \sum_ {i = 0} ^ {V} S m o o t h _ {L _ {1}} \left(d _ {i} - d _ {i} ^ {*}\right), \tag {12} +$$ + +where $D$ is the ground truth disparity, $D_{E}^{*}$ is the disparity predicted by the event branch, and $V$ is the number of valid pixels with ground truths, $d_{i}$ for $i\in \{1,2,3,\dots ,V\}$ is the pixels in the ground truth disparity $D$ that can be used for training, and $d_{i}^{*}$ is the pixel at the corresponding position in the predicted disparity $D_{E}^{*}$ . The smooth $L_{1}$ loss function can be expressed as follows: + +$$ +\operatorname {S m o o t h} _ {L _ {1}} (x) = \left\{ \begin{array}{l l} 0. 5 x ^ {2}, & \text {i f} | x | < 1, \\ | x | - 0. 5, & \text {o t h e r w i s e .} \end{array} \right. \tag {13} +$$ + +Inference. The event branch is fully optimized during training, capturing a massive amount of texture knowledge and dense structure information from intensity images. During inference, retaining only the event branch can achieve efficient and dense disparity prediction without increasing the computational cost, and most importantly, preserving the low latency characteristics of events. + +# 4. Experiment + +# 4.1. Experimental Setting + +Datasets. To evaluate the proposed method's effectiveness, the experiments are performed on two publicly available datasets, MVSEC [69] and DSEC [17], both from the real world. The MVSEC dataset has two DAVIS cameras + +
MethodMean disp. error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓Median depth error [cm] ↓
Split1Spilt3Split1Spilt3Split1Spilt3Split1Spilt3
EIS-E [37]--80.668.313.325.7--
DDES [49]0.590.9489.874.816.727.86.814.7
DTC-PDS [62]0.560.6591.588.715.318.66.48.7
CTC-PDS [62]0.530.7391.688.214.920.66.410.6
EITNet [1]0.550.7592.189.614.219.45.910.4
DTC-SPADE [62]0.460.6093.089.713.517.15.27.9
TES [13]0.460.4992.992.613.015.05.05.8
Ours0.320.3794.893.411.213.24.24.7
+ +with stereo setups that provide intensity images and pairing events. We use indoor flying from the MVSEC dataset, which was captured from a drone flying in a room with various objects. Following [10, 13, 49, 62], we split the MVSEC dataset indoor flying into three and use two of them, split 1 and split 3. For a fair comparison, we use mean disparity error, one-pixel accuracy, mean depth error, and median depth error as metrics. The DSEC dataset is a stereo event camera dataset for outdoor driving scenes. The intensity images and events are captured by different devices with different resolutions. All cameras are at the same height, so we approximately map the intensity images to the corresponding positions of the events with the provided camera matrix. For evaluation, we use the mean absolute error (MAE), one-pixel error (1PE), two-pixel error (2PE), and root mean square error (RMSE) as the metrics. See the supplementary material for more details on the dataset split. + +Implementation Details. The proposed method is implemented using PyTorch and trained from scratch. We adopt the Adam [27] ( $\beta_{1} = 0.9$ , $\beta_{2} = 0.99$ ) optimizer to train the model with a batch size of 8 using $256 \times 256$ random crops. We set the maximum disparity to 192. The initial learning rate is $1 \times 10^{-3}$ , and the weight decay is $1 \times 10^{-4}$ . All experiments are conducted on an NVIDIA RTX A6000 GPU. For more details about the implementation, please refer to the supplementary material. + +# 4.2. Quantitative Results + +We compare the results of the proposed $\mathbf{IED}^2\mathbf{S}$ with state-of-the-art methods. The comparison results are shown in Table 1. EIS-E [37], DDES [49], DTC-PDS [62], EITNet [1], and TES[13] only use asynchronous and sparse events, which limits the effectiveness of stereo matching due to the lack of sufficient texture knowledge and dense structure information. The proposed $\mathbf{IED}^2\mathbf{S}$ fully extracts extensive dense and structural information contained in intensity images and distills it into the event branch during the training, making its performance significantly better than existing event-based stereo matching methods. Specifically, + +Table 1. Disparity estimation results on the MVSEC [69] dataset. - indicates that results are not provided in the original paper. The best is in bold and the second best is in underlined. + +
MethodsMAE ↓1PE ↓2PE ↓RMSE ↓
DDES [49]0.57610.9152.9051.381
DTC-PDS [62]0.5269.5342.3531.263
Se-CFF [38]0.5219.5862.6231.235
Ours0.4938.8232.1341.069
+ +Table 2. Disparity estimation results on the DSEC [17] dataset. + +compared with the best-performing event-based method, TES[13], the proposed method outperforms it by $13.85\%$ and $12.00\%$ in terms of mean depth error for split1 and split3, respectively. We believe that this improvement comes from fully considering the modality gap and the novel design of the distillation strategy. Meanwhile, it is worth noting that the combined intensity images and events methods require both events and intensity images to be input at the inference stage, which limits the application of some scenarios, such as high-speed motion scenes or when intensity images are unavailable. However, ours only requires events as input during the inference phase, which fully utilizes the low-latency characteristics of events in special scenarios (high speed and extreme lighting). + +Furthermore, we evaluate the proposed model on the DSEC dataset, and the results are shown in Table 2. Similarly, compared with the state-of-the-art baselines, ours achieves the best performance on all metrics and outperforms existing methods by a large margin, thanks to the designed cross-modal distillation in Sec. 3. Specifically, it outperforms Se-CFF [38] by $5.37\%$ and $7.46\%$ in terms of mean absolute error and one-pixel error, respectively, which verifies the effectiveness of the proposed method. + +# 4.3. Qualitative Results + +As shown in Fig. 5, we present a qualitative comparison of our method with the baseline in various scenarios on the MVSEC dataset. DDES [49] and TES [13] struggle to output dense disparity due to the lack of detailed information caused by the inherent sparsity of events. They are not ideal in some locations, such as the stripe lines dis + +![](images/58cb30f1dfd709951ac92ff87597200830eb71f08dd763da73ab73c153a332d6.jpg) +(a) Intensity + +![](images/37ed519bf0c3aab5a98a0c72294d7aae677da59230bc058f4fb24273c7888eb3.jpg) +(b) Events + +![](images/5d4d4fa603bdc1d2aa5c16a7a4032593d67c7ebaefeea9b6d5e2df7b683c9309.jpg) +(c) DDES + +![](images/6892d8279cb9b270df1f96c35857e8fe448420c5d2de80668283cbcbb1430858.jpg) +(d) TES + +![](images/43560092a0f315aac53e7f503813a769e8a90e55b729eddbde288cc0f1b34f9c.jpg) +(e) Ours + +![](images/39a8daf66e7229aa4a6be50e46228e7c7a104517e1d5e11fd4cfa3833fcdaba7.jpg) +(f) Ground Truth + +![](images/42929fc331a6eafdc51542d0926137ab52682cb5a4eb1d0e94282dc9cf4b3101.jpg) + +![](images/e6b5b457e4762ee87d6443594f6fdb049ec003eb0d8ba48e898a69d754c350b5.jpg) + +![](images/5876d02dc277ea3d3e54e61c7497cf0ae89b4db37e8cc014ba891ce2fa2ab95f.jpg) + +![](images/1fb4cdc86ef1da8c60fd57272bb0fc518162a135d48e5a356c734ccebdb7947b.jpg) + +![](images/a84e706cd9a7e9bfc47cde48ab88d2596e8fbe7e5e236b96fa2dfdf81a8be524.jpg) + +![](images/c3c51fe3b9432df8432c28c62156e9af10a44248377824b125417c6ba9cfd264.jpg) + +![](images/5b74272c89e4772ee380c51224711f604ce2aa6f77146bd4cdfc5617bd81a724.jpg) + +![](images/e718529e166f432e7ad36b7c1eadf5325c587a5f2c64c325a02d1475748c3ce9.jpg) + +![](images/551349293e567a1a6737d62d9aa13fea7ea4db1807bbf199e4df9fb52f6e6ac3.jpg) + +![](images/014947b9143a97eb457e2a0509b99b9d7bf82c896e3a68c119e6d83c4394ce30.jpg) + +![](images/d58eddf5ab90983210cf7f52e303049e393067d62f2af813cf8ac00b488e8ea7.jpg) + +![](images/da0c009a26bddc00f9b337cb826dac440dc05ae5893c323cf0095c32d902276f.jpg) + +![](images/b8092e78bbe8e8877750e5d3da848cbf2bda2e1de7c6322229c69c7e58df471a.jpg) +Figure 5. Qualitative comparison of dense disparity estimation for indoor flying scenes in the MVSEC dataset. The first two columns are the intensity images and the corresponding events for visual reference, where the red and blue discrete points represent positive and negative events, respectively. (c), (d), and (e) are the results of DDES [49], TES [13], and ours, respectively. Our proposed method fully uses the intensity image's detail and texture information during the training stage, better displaying fine details in the region highlighted by the red box. Detailed disparity information is framed by the red box for comparison. + +![](images/d0bc4d6091119dfb1833c5b4ce7d4fd585919feb9a6c14912974c1fb52c49783.jpg) + +![](images/067a424c94053d08475bf63ee39263493a094936bc1f71dd9cf1052d21a50093.jpg) + +![](images/0db7a3c8dc200af2701023449c63174e342bc7c9844a2dec3b02290f87f0dbf4.jpg) + +![](images/45ebc35ede989c22063ab41d59476f6d8e239ff4ceb2dabc15a57d3f1fca5efe.jpg) + +![](images/943260f303efdc1ed17023ce53b9594c1bd4d62cdf4e5510fafa9bc2f942dd1e.jpg) + +appearing where the disparity should change dramatically. For example, blur appears at the junction of objects such as barrels and chairs with the background. In contrast, our method addresses these challenges by enabling the model to capture more intricate features, thereby offering more reliable information for accurate disparity estimation in ill-posed regions. For example, for a chair placed far away, the proposed method achieves fine-grained disparity prediction. Consequently, even in areas where event data is sparse and closely resembles the surrounding environment, our approach produces disparities that are not only clearer and denser but also exhibit more precise boundaries, such as the area outlined by the red box in Fig. 5. + +![](images/e13e1e32e265aeba5dd28b46dad203975f0b4cf1ea5aee7ece49da17b56936be.jpg) +(a) Intensity + +![](images/bb06f6a66af0ab5a77c4e1fcdcf3bcfd64c642e6fd3548cee46cd8e6ec68efda.jpg) +(b) Sc-CFF + +![](images/c3b4b9ecc6f8dccce080218c9ea65fe23a1b82dd51f832d2e3d83efa7de2d801.jpg) +(c) TES + +![](images/dc6816f2934fa445ba481159e62fdcc8166dd8005f43e8993582e02c5f3f5403.jpg) +(d) Ours + +![](images/536721dbf39190a30b3f10f50c64b24e61043025db68a1a8ee3a40520210dca0.jpg) + +![](images/dcee0aece8a6b08245878424c5fce1a5bb1037ae4b921ee0a765012a06a04674.jpg) + +![](images/63e5e939c434f807f199d17dbd2536ba6de73a9015e9dc06bb7c26d273955f95.jpg) + +![](images/2044afabb9c685fe9f62eeb33436b1d8c3cbe090190568668c32d6c5779cd330.jpg) + +![](images/da3466bc082759a2277aa57255176c8ae03d453448ff1c5eaa992e68cb6903e9.jpg) +Figure 6. Qualitative results on the DSEC dataset. Our method can construct more dense disparities than others. + +![](images/4f4af12a9a16ea7eba34307b40017842a0967681425f458d3ca86543ae2305ac.jpg) + +![](images/62a34741376023e3666e55e748a83db532bb751505286888d00a4916178e186a.jpg) + +![](images/0ca9ba52097ce76f5357d9e62279a3796ea7290c93b335cbd684b8f91866cfa1.jpg) + +Additionally, we present results on the DSEC datasets in Fig. 6. We highlight regions with complex textures, where our method outperforms the others in predicting disparity with higher accuracy. For example, for objects such as trees and guardrails on the roadside, we obtain much sharper, more detailed, and artifact-free disparities. + +# 4.4. Ablation Studies + +Contribution of each component. We perform extensive experiments on the MVSEC dataset to fully validate the effectiveness of each proposed component and related design choices in Sec. 3. We first present the results of the ablation study in Table 3. It can be observed that all the proposed components have contributed to the superior performance. This indicates that the information derived from intensity images is effective for dense event-based disparity estimation without significantly altering the network structure. When only one component is added, we see the most improvement comes from the feature distillation, which indicates that modality-specific and task-related features of the intensity branch are needed. When it is applied solely, the one-pixel accuracy is increased by $3.74\%$ and $2.91\%$ for split1 and split3, respectively. It is worth noting that intensity-event joint left-right consistency improves the one-pixel accuracy of split1 and split3 by $0.22\%$ and $1.01\%$ , respectively, which verifies the feasibility of improving the matching effect through ensuring cross-view consistency in + +
F-distill.L-distill.LRCMean disparity error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓
Split 1Split 3Split 1Split 3Split1Split3
XXX0.540.6791.089.415.018.5
✓XX0.380.4694.492.012.113.9
X✓X0.490.5391.991.613.817.5
XX✓0.500.6491.290.314.618.0
X✓✓0.490.5292.191.613.717.4
✓X✓0.370.4194.592.711.613.7
✓✓X0.340.3994.893.211.313.4
✓✓✓0.320.3794.893.411.213.2
+ +stereo matching. When all components are combined, the performance is further improved without any conflicts, resulting in the best performance with a significant improvement of the one-pixel accuracy by $4.18\%$ and $4.47\%$ for split1 and split3, respectively. + +Table 3. The result of our intensity-to-event distillation scheme with different settings on the MVSEC dataset. F-distill. denotes the feature distillation including shallow and deep distillation, L-distill. represents logit-level distillation, and LRC means left-right consistency. + +
MethodsMean disparity error [pix] ↓One-pixel accuracy [%] ↑
Split1Split3Split1Split3
MSE0.450.5092.291.8
Si2eD0.390.4693.092.5
MSE+Di2eD0.370.4194.093.0
Si2eD+Di2eD0.320.3794.893.4
+ +Feature alignment. Although feature distillation has been a widely used loss term in many knowledge distillation studies [28, 39], we are the first to implement that for event-based stereo. For further exploration and discussion, we present our experiments to validate the design choice of the feature distillation module, as it brings the most improvement among all the proposed components. Specifically, to solely examine the effectiveness of our proposed shallow distillation, we compare it to the simple alignment feature-level distillation methods. MSE is a baseline method [22] that simply minimizes the element-wise $L_{2}$ distance between the features of the teacher and student. As shown in Table 4, directly aligning the features of the two modalities contaminates the original modality-specific information and introduces noise to the event branch. Specifically, the one-pixel accuracy is reduced by $2.74\%$ and $1.71\%$ for split1 and split3, respectively. A possible reason for this performance gain comes from the large modality difference between events and intensity images, and the novel design fully exploits these modality-specific features while effectively avoiding cross-sensor contamination. + +Table 4. The effectiveness of our designed feature distillation on the MVSEC dataset. Si2eD represents shallow distillation, and Di2eD denotes deep distillation. + +
SettingMVSEC -> DSECDSEC -> MVSEC
MetricsMAE ↓1PE ↓Mean disp. error [pix] ↓One-pixel acc. [%] ↑
w/o distill.0.78212.4130.7479.8
Full model0.5639.9730.4691.4
+ +Table 5. Cross-validation experimental results between two real-world datasets, verifying good generalization. + +# 4.5. Cross-Domain Generalization Performance + +The modality-specific knowledge obtained by cross-sensor distillation is independent of the dataset and has good generalization under different domain distributions. In this section, we discuss the generalization performance of the proposed method on unseen domains. Specifically, we perform cross-validation on two real-world datasets, i.e., the model is trained on MVSEC (DSEC) and evaluated on DSEC (MVSEC). As shown in Table 5, the designed module is less affected by the domain gap. This is because the designed cross-modal distillation enables the event branch to acquire general knowledge for dense stereo matching. + +# 5. Conclusion + +We propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. We design a multi-level intensity-to-event distillation strategy, achieving binocular-to-binocular distillation for the first time. A massive amount of texture knowledge and dense structure information contained in the intensity image is distilled to the event branch through the designed strategy. At the same time, an intensity-event joint left-right consistency module is proposed to enforce cross-view consistency. Extensive experiments on the MVSEC and DSEC datasets demonstrate the effectiveness of the proposed framework. + +# Acknowledgments + +This work was supported in part by National Natural Science Foundation of China under grant No.62371450. + +# References + +[1] Soikat Hasan Ahmed, Hae Woong Jang, SM Nadim Uddin, and Yong Ju Jung. Deep event stereo leveraged by event-to-image translation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 882-890, 2021. 1, 2, 6 +[2] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In British Machine Vision Conference, pages 1-11, 2011. 1 +[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A $240 \times 180$ 130 db $3\mu s$ latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1 +[4] Luis A Camunas-Mesa, Teresa Serrano-Gotarredona, Sio H Ieng, Ryad B Benosman, and Bernabe Linares-Barranco. On the use of orientation filters for 3d reconstruction in eventdriven stereo vision. Frontiers in neuroscience, 8:48, 2014. 2 +[5] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4 +[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1 +[7] Shoushun Chen and Menghan Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1682-1683. IEEE, 2019. 1 +[8] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. 2 +[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 1 +[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 1, 2, 6 +[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In Proceedings of the European Conference on Computer Vision, pages 470-486. Springer, 2022. 1, 2 +[12] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2 +[13] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In Proceedings of the European Conference on Computer Vision, pages 294–314. Springer, 2025. 2, 6, 7 +[14] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learn + +ing spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022. 2 +[15] Weiqin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, David Suter, and Alireza Bab-Hadiashar. Semantic guided long range stereo depth estimation for safer autonomous vehicle applications. IEEE Transactions on Intelligent Transportation Systems, 23(10):18916-18926, 2022. 1 +[16] Boyu Gao, Haoxiang Lang, and Jing Ren. Stereo visual slam for autonomous vehicles: A review. In IEEE International Conference on Systems, Man, and Cybernetics, pages 1316-1322. IEEE, 2020. 1 +[17] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 5, 6 +[18] Suman Ghosh and Guillermo Gallego. Event-based stereo depth estimation: A survey. arXiv preprint arXiv:2409.17680, 2024. 2 +[19] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 5 +[20] Rostam Affendi Hamzah, A Fauzan Kadmin, M Saad Hamid, S Fakhar A Ghani, and Haidi Ibrahim. Improvement of stereo matching algorithm for 3d surface reconstruction. Signal Processing: Image Communication, 65:165-172, 2018. 1 +[21] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[22] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 87-104. Springer, 2022. 2, 8 +[23] Asmaa Hosni, Michael Bleyer, Margrit Gelautz, and Christoph Rhemann. Local stereo matching using geodesic support weights. In 16th IEEE International Conference on Image Processing, pages 2093-2096. IEEE, 2009. 1 +[24] Asmaa Hosni, Christoph Rhemann, Michael Bleyer, Carsten Rother, and Margrit Gelautz. Fast cost-volume filtering for visual correspondence and beyond. IEEE transactions on pattern analysis and machine intelligence, 35(2):504-511, 2012. 1 +[25] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12605-12614, 2020. 2 +[26] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Émilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2022. 2 +[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 + +[28] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 8 +[29] Hamid Laga, Laurent Valentin Jospin, Farid Boussaid, and Mohammed Bennamoun. A survey on deep learning techniques for stereo-based depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(4):1738-1764, 2020. 1 +[30] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 1 +[31] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1 +[32] Xineng Li, Chen Zhang, Wanjuan Su, and Wenbing Tao. Inet: Implicit intra-inter information fusion for real-time stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3225-3233, 2024. +[33] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI conference on artificial intelligence, pages 1647-1655, 2022. 1 +[34] Tianshan Liu, Kin-Man Lam, Rui Zhao, and Guoping Qiu. Deep cross-modal representation learning and distillation for illumination-invariant pedestrian detection. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 315-329, 2021. 2 +[35] Xu Liu, Jianing Li, Jinqiao Shi, Xiaopeng Fan, Yonghong Tian, and Debin Zhao. Event-based monocular depth estimation with recurrent transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[36] Qian Long, Qiwei Xie, Seiichi Mita, Kazuhisa Ishimaru, and Noriaki Shirai. A real-time dense stereo matching method for critical environment sensing in autonomous driving. In 17th International IEEE Conference on Intelligent Transportation Systems, pages 853-860. IEEE, 2014. 1 +[37] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021. 2, 6 +[38] Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concentrate and focus on the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6114-6123, 2022. 2, 6 +[39] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5007-5016, 2019. 8 +[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen, + +and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1 +[41] Mary Phuong and Christoph H Lampert. Distillation-based training for multi-exit architectures. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1355-1364, 2019. 2 +[42] Ewa Piatkowska, Ahmed Belbachir, and Margrit Gelautz. Asynchronous stereo vision for event-driven dynamic stereo sensor using an adaptive cooperative approach. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 45-50, 2013. 2 +[43] Ewa Piatkowska, Jurgen Kogler, Nabil Belbachir, and Margrit Gelautz. Improved cooperative stereo matching for dynamic vision sensors with ground truth evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 53-60, 2017. +[44] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision, 126(12):1394-1414, 2018. 1 +[45] Paul Rogister, Ryad Benosman, Sio-Hoi Ieng, Patrick Lichtsteiner, and Tobi Delbruck. Asynchronous event-based binocular stereo matching. IEEE Transactions on Neural Networks and Learning Systems, 23(2):347-353, 2011. 2 +[46] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47:7-42, 2002. 1 +[47] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 2024. 1 +[48] Pin Tang, Hai-Ming Xu, and Chao Ma. Prototransfer: Cross-modal prototype transfer for point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3337-3347, 2023. 2, 4 +[49] Stepan Tulyakov, Francois Fleuret, Martin Kiefel, Peter Gehler, and Michael Hirsch. Learning an event sequence embedding for dense event-based deep stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1527-1537, 2019. 2, 6, 7 +[50] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1365-1374, 2019. 2 +[51] SM Nadim Uddin, Soikat Hasan Ahmed, and Yong Ju Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 2 +[52] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1 + +[53] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8637-8646, 2023. 2, 3 +[54] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 1 +[55] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12981-12990, 2022. +[56] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023. +[57] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1 +[58] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European conference on computer vision, pages 677-695. Springer, 2022. 2 +[59] Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2019. 1 +[60] Shanxin Yuan, Bjorn Stenger, and Tae-Kyun Kim. 3d hand pose estimation from rgb using privileged learning with depth data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2 +[61] Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Low-cost multispectral scene analysis with modality distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 803-812, 2022. 2 +[62] Kaixuan Zhang, Kaiwei Che, Jianguo Zhang, Jie Cheng, Ziyang Zhang, Qinghai Guo, and Luziwei Leng. Discrete time convolution for fast event-based stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8676-8686, 2022. 2, 6 +[63] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3713-3722, 2019. 2 +[64] Bo Zhao, Ruoxi Ding, Shoushun Chen, Bernabe Linares-Barranco, and Huajin Tang. Feedforward categorization on + +aer motion events using cortex-like features in a spiking neural network. IEEE transactions on neural networks and learning systems, 26(9):1963-1978, 2014. 1 +[65] Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao. Simdistill: Simulated multimodal distillation for bev 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 7460-7468, 2024. 2 +[66] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 2 +[67] Runkai Zhao, Heng Wang, and Weidong Cai. Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer. In ACM Multimedia, 2024. 3, 4 +[68] Alex Zihao Zhu, Yibo Chen, and Kostas Daniilidis. Realtime time synchronized event-based stereo. In Proceedings of the European Conference on Computer Vision, pages 433-447, 2018. 2 +[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 5, 6 +[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3 +[71] Dongqing Zou, Feng Shi, Wei-Heng Liu, Jia Li, Qiang Wang, Paul KJ Park, and Hyunsurk Ryu. Robust dense depth maps generations from sparse dvs stereos. In British Machine Vision Conference, 2017. 2 \ No newline at end of file diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/images.zip b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..010acd373289615aa1b08842ee467bb18677d3e7 --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1bb1198732227fa2a28b8c794467f9eb54fb8e23056efcc8ddd8aacab91e8a51 +size 681134 diff --git a/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/layout.json b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c98327a6552779df06a1909570cefd10ef1464bd --- /dev/null +++ b/2025/Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation/layout.json @@ -0,0 +1,11036 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "spans": [ + { + "bbox": [ + 67, + 103, + 543, + 121 + ], + "type": "text", + "content": "Enhanced Event-based Dense Stereo via Cross-Sensor Knowledge Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": "Haihao Zhang" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Yunjian Zhang" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Jianing Li" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{3\\dagger}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Lin Zhu" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Meng Lv" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Yao Zhu" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Yanwei Liu" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "text", + "content": " Xiangyang Ji" + }, + { + "bbox": [ + 115, + 141, + 492, + 174 + ], + "type": "inline_equation", + "content": "^{3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "spans": [ + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "text", + "content": "Institute of Information Engineering, Chinese Academy of Sciences " + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "text", + "content": "Beijing Institute of Technology " + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 171, + 467, + 200 + ], + "type": "text", + "content": "Tsinghua University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 131, + 201, + 476, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 131, + 201, + 476, + 213 + ], + "spans": [ + { + "bbox": [ + 131, + 201, + 476, + 213 + ], + "type": "text", + "content": "haihaozhang@bit.edu.cn sdtczyj@gmail.com lijianing@pku.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 253, + 200, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 253, + 200, + 265 + ], + "spans": [ + { + "bbox": [ + 152, + 253, + 200, + 265 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 278, + 296, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 278, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 296, + 542 + ], + "type": "text", + "content": "Accurate stereo matching under fast motion and extreme lighting conditions is a challenge for many vision applications. Event cameras have the advantages of low latency and high dynamic range, thus providing a reliable solution to this challenge. However, since events are sparse, this makes it an ill-posed problem to obtain dense disparity using only events. In this work, we propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. Specifically, a multi-level intensity-to-event distillation strategy is designed to maximize the potential of long-range information, local texture details, and task-related knowledge of the intensity images. Simultaneously, to enforce the cross-view consistency, an intensity-event joint left-right consistency module is proposed. With our framework, extensive dense and structural information contained in intensity images is distilled to the event branch. Therefore, retaining only the events can predict dense disparities during inference, preserving the low latency characteristics of the events. Adequate experiments conducted on the MVSEC and DSEC datasets demonstrate that our method exhibits superior stereo matching performance than baselines, both quantitatively and qualitatively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 135, + 577 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Depth estimation based on binocular stereo vision is a classic problem in computer vision [15, 16, 20, 24, 29, 36, 44, 46, 59], which aims to determine the correspondence between points in a rectified image pair. Deep learning-based stereo matching methods perform well on various public benchmarks [2, 6, 9, 23, 24, 31-33, 52, 54-57]. However, due to the shortcomings of frame-based sensors in terms of frame rate and dynamic range, it is very difficult to apply traditional cameras in more challenging scenarios such as" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 323, + 257, + 548, + 295 + ], + "blocks": [ + { + "bbox": [ + 325, + 244, + 406, + 254 + ], + "lines": [ + { + "bbox": [ + 325, + 244, + 406, + 254 + ], + "spans": [ + { + "bbox": [ + 325, + 244, + 406, + 254 + ], + "type": "text", + "content": "(a) Event-based Method" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 323, + 257, + 548, + 295 + ], + "lines": [ + { + "bbox": [ + 323, + 257, + 548, + 295 + ], + "spans": [ + { + "bbox": [ + 323, + 257, + 548, + 295 + ], + "type": "image", + "image_path": "b89b1ea31765c0322264d0b680275c6d4170439a658e2bba8cc46e1dd580d9b8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 320, + 308, + 548, + 394 + ], + "blocks": [ + { + "bbox": [ + 326, + 297, + 441, + 306 + ], + "lines": [ + { + "bbox": [ + 326, + 297, + 441, + 306 + ], + "spans": [ + { + "bbox": [ + 326, + 297, + 441, + 306 + ], + "type": "text", + "content": "(b) Event-Intensity Fusion Method" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 320, + 308, + 548, + 394 + ], + "lines": [ + { + "bbox": [ + 320, + 308, + 548, + 394 + ], + "spans": [ + { + "bbox": [ + 320, + 308, + 548, + 394 + ], + "type": "image", + "image_path": "e7334d2c544679a3eac1aa96ae1da71845bbf97f4ab5d71784764ed71c9187f0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 319, + 413, + 550, + 517 + ], + "blocks": [ + { + "bbox": [ + 323, + 398, + 409, + 407 + ], + "lines": [ + { + "bbox": [ + 323, + 398, + 409, + 407 + ], + "spans": [ + { + "bbox": [ + 323, + 398, + 409, + 407 + ], + "type": "text", + "content": "(c) Our Proposed Method" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 413, + 550, + 517 + ], + "lines": [ + { + "bbox": [ + 319, + 413, + 550, + 517 + ], + "spans": [ + { + "bbox": [ + 319, + 413, + 550, + 517 + ], + "type": "image", + "image_path": "da46527f76d8da6623574e2a3331938f3948b3844f218d84fdd922ae91755e64.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 526, + 554, + 583 + ], + "lines": [ + { + "bbox": [ + 313, + 526, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 554, + 583 + ], + "type": "text", + "content": "Figure 1. Comparisons on stereo matching frameworks (event-based, event-intensity fusion methods) and our proposed IED" + }, + { + "bbox": [ + 313, + 526, + 554, + 583 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 526, + 554, + 583 + ], + "type": "text", + "content": "S that presents an effective intensity-to-event distillation strategy. Texture and dense knowledge is transferred to the event branch, leading to dense stereo matching in the inference stage." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 492, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 492, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 492, + 605 + ], + "type": "text", + "content": "high speed and extreme lighting conditions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 605, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 554, + 700 + ], + "type": "text", + "content": "Event cameras [3, 7, 30, 35, 40, 47, 64] are bio-inspired sensors and report intensity changes asynchronously in the form of an event stream, with each containing spatial location, polarity, and timestamps. They have the advantages of low latency and high dynamic range, making them immune to high-speed motion blur and can operate in extreme lighting scenarios. Therefore, they offer a satisfying solution to overcome the limitations of frame-based sensors." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 714 + ], + "type": "text", + "content": "Recent event-based stereo matching studies [1, 10, 11," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 240, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 240, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 240, + 713 + ], + "type": "text", + "content": "† Corresponding authors: Jianing Li, Yunjian Zhang" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5437" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "13, 18, 37, 38, 49] embed asynchronous event data into image-like representations, then fine-grained features are extracted and subsequent disparity regression is performed. While disparity can be effectively estimated, since event cameras can only perceive changes in intensity, when the scene is stationary or there is little relative motion between the camera and the scene, the event stream will be very sparse. This makes it an ill-posed problem to obtain dense disparity maps using only sparse asynchronous event data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 182, + 294, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 294, + 337 + ], + "type": "text", + "content": "Considering the dense features provided by intensity images, there raises a trend to combine events and intensity images for dense stereo matching [10, 11, 37]. They design algorithms that simultaneously input intensity images and events, using their complementary features to calculate disparities, and the results are outstanding advances for event-based stereo matching. Nevertheless, these approaches require the events and intensity images simultaneously during the inference stage, suffering from two problems: (1) The low latency characteristics of events is sacrificed unintentionally; (2) Processing both modalities concurrently consumes more computing resources, which is a heavy burden for real-time applications." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 339, + 295, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 295, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 295, + 625 + ], + "type": "text", + "content": "To address the above problems, we propose a novel framework named Intensity-to-Event Distillation for Dense Stereo matching method (IED" + }, + { + "bbox": [ + 55, + 339, + 295, + 625 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 339, + 295, + 625 + ], + "type": "text", + "content": "S), which explores the potential of an efficient event-based stereo model for dense disparity prediction. As shown in Fig. 1(c), we implement cross-sensor distillations from the intensity branch to the event branch. The dense and structural information of the intensity image provides useful regularization for asynchronous events, especially for ill-posed regions. In our framework, we design a novel feature alignment strategy, achieving binocular-to-binocular distillation for the first time. Specifically, a comprehensive multi-level intensity-to-event distillation strategy is designed, fully leveraging the modality-specific characteristics of shallow features and the task-related contextual understanding involved in deep abstract representations. Moreover, a warping-based intensity-event joint left-right consistency module is designed to ensure cross-view consistency. The intensity images are only used in the training stage. For inference, only event data is needed for obtaining dense disparity. Compared with fusion-based methods shown in Fig. 1(b), it has greater flexibility, preserving the low latency characteristics of event data and greatly reducing the inference burden. In summary, the main contributions of this work are:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 628, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "text", + "content": "1) We propose a dense event-based stereo matching method, " + }, + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{IED}^2\\mathbf{S}" + }, + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "text", + "content": ", which is the first trial to implement binocular intensity-to-event distillation. The proposed " + }, + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{IED}^2\\mathbf{S}" + }, + { + "bbox": [ + 55, + 628, + 294, + 687 + ], + "type": "text", + "content": " enables the event branch to learn accurate dense information, greatly improving its disparity prediction capability." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "2) We design a multi-level intensity-to-event distillation, fully distilling long-range information, local texture details," + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 96 + ], + "type": "text", + "content": "and task-related knowledge into the event branch while avoiding the contamination of the original event modality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": "3) Extensive experiments demonstrate that the proposed " + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{IED}^2\\mathbf{S}" + }, + { + "bbox": [ + 313, + 97, + 553, + 133 + ], + "type": "text", + "content": " achieves significantly better performance than the prior works for event-based stereo matching." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "spans": [ + { + "bbox": [ + 314, + 145, + 400, + 157 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 166, + 428, + 177 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 166, + 428, + 177 + ], + "spans": [ + { + "bbox": [ + 314, + 166, + 428, + 177 + ], + "type": "text", + "content": "2.1. Event-Based Stereo" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 183, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 183, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 183, + 553, + 243 + ], + "type": "text", + "content": "Early attempts used traditional hand-crafted methods [4, 12, 42-45, 68, 71] to determine the correspondence between two event streams. However, there are no fixed patterns in the event data, which makes it troublesome to use traditional methods to obtain dense disparity." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 244, + 553, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 471 + ], + "type": "text", + "content": "Recent learning-based approaches [1, 10, 11, 13, 37, 38, 49, 51] are able to estimate disparities using sparse events. The network usually includes feature extraction, cost aggregation, disparity regression, and disparity refinement modules. Tulyakov et al. [49] proposed a 4D queue embedding method that includes the spatiotemporal information of event data. Nam et al. [38] proposed an attention-based event concentration network to display scene details by omitting fewer details without covering events. DTC [62] used discrete time convolution modules to aggregate sequential event information at the feature level, accelerating the stereo matching process at a lower computational cost. Cho et al. [13] proposed the concept of stereo flow to aggregate past features and cost volumes. Accurate stereo matching is achieved by training the stereo matching and stereo flow simultaneously. However, it is an ill-posed problem to obtain dense disparity using only sparse asynchronous events. We therefore explore cross-sensor distillation for event-based dense stereo matching." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 480, + 453, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 480, + 453, + 491 + ], + "spans": [ + { + "bbox": [ + 314, + 480, + 453, + 491 + ], + "type": "text", + "content": "2.2. Cross-Sensor Distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 713 + ], + "type": "text", + "content": "Knowledge distillation is an effective method for compressing models while maintaining accuracy [21, 41, 50, 63]. In particular, cross-sensor knowledge distillation has received increasing attention and applied in many tasks, such as 3D object detection [8, 14, 22, 53, 65, 66], multispectral detection [34, 61], 3D hand pose estimation [60], and LiDAR semantic segmentation [25, 26, 48, 58]. Typically, a well-trained teacher model guides the student model's feature extraction or result prediction that receives input from other modalities. Specifically, Hong et al. [22] first proposed cross-sensor knowledge distillation based on BEV, mimicking LiDAR BEV's representation to achieve improved performance. Chong et al. [14] achieved the transfer of structural cues by projecting point clouds onto the plane of the image and maintaining similar local region affinities. Liu et al. [34] adopted different losses at multiple task levels to transfer coarse-grained and fine-grained features covering feature, detection, and segmentation. Tang et al. [48] estab" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5438" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 547, + 335 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 547, + 335 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 547, + 335 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 547, + 335 + ], + "type": "image", + "image_path": "5cb4d299c5a05b21db20b69663dfa9d4d855dbd173833363253a0d9de51376c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 340, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 54, + 340, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. The pipeline of the proposed Intensity-to-Event Distillation for Dense Stereo matching method (IED" + }, + { + "bbox": [ + 54, + 340, + 555, + 376 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 54, + 340, + 555, + 376 + ], + "type": "text", + "content": "S). The proposed approach enables the event branch to learn a massive amount of texture knowledge and dense structure information under the guidance of the intensity branch. After training, the intensity branch is removed without introducing additional computational burden during inference." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 385, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 385, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 54, + 385, + 295, + 422 + ], + "type": "text", + "content": "lished a prototype-like library from strictly aligned fused features, and encouraged all point cloud features to learn from the prototypes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 422, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 422, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 422, + 295, + 495 + ], + "type": "text", + "content": "Inspired by this, we propose an enhanced event-based dense stereo method via cross-sensor knowledge distillation. It obtains a massive amount of texture knowledge from the intensity images in the training stage, while in the inference stage, only events are needed to obtain dense and accurate disparity, greatly reducing the computational burden." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 506, + 138, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 138, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 138, + 519 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 525, + 167, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 167, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 167, + 537 + ], + "type": "text", + "content": "3.1. Problem Definition" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": "The events of the left and right event cameras between " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " are denoted as " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "E_{t}^{L}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "E_{t}^{R}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": ", respectively, and the corresponding intensity images are indicated as " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "I_{t}^{L}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "I_{t}^{R}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": ". The event data is a sequence of four-dimensional vectors " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "(x,y,p,\\tau)" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " means pixel coordinates, " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "p\\in \\{-1,1\\}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " indicates polarity, and " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\tau \\in (t - 1,t]" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " represents continuous timestamps. We represent event data " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "E_{t}^{L,R}\\in \\{(x,y,p,\\tau)|t - 1 < \\tau \\leq t\\}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " in voxel grid format following [70]. Specifically, we first scale the timestamps to bin indexes " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "[0:B - 1]" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": ", after which the events " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "E_{t}^{L,R}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " are stacked to generate an voxel grid " + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "inline_equation", + "content": "V_{t}^{L,R}(b,x,y)\\in R^{B\\times H\\times W}" + }, + { + "bbox": [ + 55, + 543, + 295, + 700 + ], + "type": "text", + "content": " with discretized time dimension B, where H and W represent the height and the width, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "Given continuous stereo intensity image pair and stereo" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": "event pair, with the same ground truth disparity, the training goal is to input four data in the training stage, which refer to " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "V_{t}^{L}, V_{t}^{R}, I_{t}^{L}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "inline_equation", + "content": "I_{t}^{R}" + }, + { + "bbox": [ + 313, + 386, + 555, + 447 + ], + "type": "text", + "content": " respectively. By making full use of the information of the two modalities, in the inference phase, only events are needed to predict the dense disparity." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 455, + 440, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 455, + 440, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 455, + 440, + 467 + ], + "type": "text", + "content": "3.2. Framework Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 568 + ], + "type": "text", + "content": "Fig. 2 shows a pipeline of our proposed framework. The framework consists of two identical branches, named the event branch and the intensity branch. The pre-trained intensity branch is used to distill knowledge to the event branch so that it can learn a massive amount of texture knowledge and dense structure information of intensity images. The distillation process includes three levels: shallow distillation, deep distillation, and logit-level distillation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "Shallow features contain much modality-specific information [53, 67], and distilling them into the event branch allows it to obtain complementary cross-sensor information. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polarity and timestamps of events. The simple alignment would contaminate the original modality-specific information and introduce noise. To avoid this problem, we adopt a comprehensive distillation strategy from both long-range and focal perspective, thus the event branch can fully capture both long-range information and local texture details. For the deep features, we adopt a direct alignment strategy" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5439" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "because they mainly contain task-related contextual information. Furthermore, the logit-level distillation makes the disparity prediction of the event branch closer to the inspection prediction of the intensity branch. Besides the distillation strategy, to ensure cross-view consistency, we design an intensity-event joint left-right consistency module. This module attempts to enforce pixel-level consistency of the left intensity image with the warped one, which implicitly indicates the quality of the estimated disparity. In the inference stage, only keeping the event branch enables efficient and dense disparity prediction, and most importantly, preserving the low latency characteristics of events." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 224, + 277, + 236 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 277, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 277, + 236 + ], + "type": "text", + "content": "3.3. Multi-Level Intensity-to-Event Distillation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 296, + 348 + ], + "type": "text", + "content": "The proposed intensity-to-event distillation scheme enables the event branch to learn semantic cues from intensity images. It consists of three levels, including shallow distillation, deep distillation, and logit-level distillation. Shallow and deep distillation are both performed at the feature level, allowing the event branch to fully obtain long-range information, local texture details, and task-related knowledge. The last level drives the event branch to be close to the distribution of dense disparity of the intensity branch." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 296, + 433 + ], + "type": "text", + "content": "Shallow Distillation. In event-based stereo matching, sparse events lack detailed texture minutiae and rich semantic information, whether from the overall or local area details. As shown in Fig. 3, we extract long-range and focal knowledge from the event and intensity branch respectively, and then align them so that the event branch can obtain both long-range information and local texture details." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 444, + 293, + 593 + ], + "blocks": [ + { + "bbox": [ + 59, + 444, + 293, + 593 + ], + "lines": [ + { + "bbox": [ + 59, + 444, + 293, + 593 + ], + "spans": [ + { + "bbox": [ + 59, + 444, + 293, + 593 + ], + "type": "image", + "image_path": "0de1b51fa0ef79cd57c42585cf0ee3cd99d630a72e9701e2c7754b6271290b2f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 601, + 295, + 657 + ], + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 657 + ], + "type": "text", + "content": "Figure 3. Illustration of shallow distillation. Features (cost volumes) of the two modalities are fed to generate relationship maps with long-range guidance. Simultaneous focal distillation is performed via channel attention and spatial attention. Shallow distillation is achieved through these two losses." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Long-range information distillation can transfer dark knowledge from the intensity branch to the event branch. However, there are significant modality gaps between shallow features, such as textures of intensity images, and polar" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "ity and timestamps of events. To avoid the noise introduced by simple alignment, we use gcblock [5] to extract long-range relations ensuring that two branches have the same long-range relationships. The feature mapping with a long-range relation generated by gcblock is shown below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 322, + 140, + 555, + 174 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 140, + 555, + 174 + ], + "spans": [ + { + "bbox": [ + 322, + 140, + 555, + 174 + ], + "type": "interline_equation", + "content": "M _ {l o n g} (x) = x \\oplus \\left(F \\left(\\ln \\left(F \\left(\\sum_ {i = 1} ^ {N} \\frac {e ^ {F \\left(x _ {i}\\right)}}{\\sum_ {j = 1} ^ {N} e ^ {F \\left(x _ {j}\\right)} x _ {i}}\\right)\\right)\\right)\\right), \\tag {1}", + "image_path": "7780831b4a39849a296d54ba8c643e7f2d62d98a933f70700d3aaed5267933cd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " convolution layer, " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " means the number of pixels of the feature " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "ln" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " indicates the layer normalization, and " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " denotes the broadcast element-wise addition. The long-range extraction loss can be obtained by computing the " + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 180, + 554, + 252 + ], + "type": "text", + "content": " norm between feature maps with long-range relationships as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 261, + 555, + 275 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 261, + 555, + 275 + ], + "spans": [ + { + "bbox": [ + 354, + 261, + 555, + 275 + ], + "type": "interline_equation", + "content": "L _ {l o n g} = \\left\\| M _ {l o n g} \\left(f _ {E}\\right) - M _ {l o n g} \\left(f _ {I}\\right) \\right\\| _ {2}, \\tag {2}", + "image_path": "8963e3588aee0fe631fbfd7fa9d21f4e424c5f3a6caf3966cbfce7288c9726ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "inline_equation", + "content": "f_{E}" + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "inline_equation", + "content": "f_{I}" + }, + { + "bbox": [ + 313, + 281, + 554, + 306 + ], + "type": "text", + "content": " represent the binocular features of the event branch and the intensity branch, respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 389 + ], + "type": "text", + "content": "Long-range information distillation achieves the transfer of long-range information but still lacks local texture details. For stereo matching, focusing only on long-range relations is not sufficient to obtain dense disparity. We introduce spatial attention and channel attention to achieve intensity-to-event focal alignment. First, we compute the spatial and channel attention maps as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 397, + 555, + 431 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 397, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 350, + 397, + 555, + 431 + ], + "type": "interline_equation", + "content": "A _ {s} (x) = H \\cdot W \\cdot s o f t m a x (\\frac {1}{C} \\sum_ {c = 1} ^ {C} | f _ {c} |), \\qquad (3)", + "image_path": "6c7d1cc0c1793a4cb745ed6fe9abf381c11aa5e84273c0619f68627e7179d4a2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 340, + 438, + 554, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 438, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 340, + 438, + 554, + 472 + ], + "type": "interline_equation", + "content": "A _ {c} (x) = C \\cdot \\operatorname {s o f t m a x} \\left(\\frac {1}{H \\cdot W} \\sum_ {i = 1} ^ {H} \\sum_ {j = 1} ^ {W} | f _ {i, j} |\\right), \\tag {4}", + "image_path": "cabe381182a8a55f23f9241f50e49eef44bbe948b1d665a76c450823d7948006.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "inline_equation", + "content": "A_{s}" + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "inline_equation", + "content": "A_{c}" + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "text", + "content": " represent the spatial attention and channel attention maps derived from the features " + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 477, + 554, + 536 + ], + "type": "text", + "content": ". In order to enable the event branch to learn spatial and channel attention from the intensity branch comprehensively, we design the focal loss as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 345, + 544, + 555, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 544, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 345, + 544, + 555, + 559 + ], + "type": "interline_equation", + "content": "L _ {f o c a l} = \\beta_ {1} L 1 \\left(A _ {s} ^ {E}, A _ {s} ^ {I}\\right) + \\beta_ {2} L 1 \\left(A _ {c} ^ {E}, A _ {c} ^ {I}\\right), \\tag {5}", + "image_path": "4e0a7835d91eac9dfa3e823ae32df11df4194d9942c3dd46a72390aa644a8e49.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "inline_equation", + "content": "\\beta_{1},\\beta_{2}" + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "text", + "content": " are hyper-parameters to balance the loss, and " + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 567, + 554, + 590 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 555, + 674 + ], + "type": "text", + "content": "Deep Distillation. As known from previous studies [48, 67], deep features are high-dimensional abstract representations with fewer modality variances. They retain rich task-related global contextual embeddings. In order to enable the event branch to effectively approximate the modality-independent intensity branch, we employ mean square error loss to facilitate deep distillation as follows:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 376, + 683, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 683, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 376, + 683, + 555, + 715 + ], + "type": "interline_equation", + "content": "L _ {\\text {d e e p}} = \\sum_ {l} ^ {L} \\left| \\left| F _ {E} ^ {l} - F _ {I} ^ {l} \\right| \\right| _ {2}, \\tag {6}", + "image_path": "8d6389ccaf2db88bef2bc168deaeff9fe9229eb4d6ec435850382d23342eaeed.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5440" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 279, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 279, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 279, + 83 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 279, + 83 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 72, + 279, + 83 + ], + "type": "text", + "content": " is the index where the distillation is performed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 85, + 296, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 85, + 296, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 85, + 296, + 193 + ], + "type": "text", + "content": "Logit-Level Distillation. The prediction of the disparity regression layer represents the probability distribution of each modality branch. We propose the logit-level distillation, and then the disparity prediction of the event branch is learned from the soft labels generated by the intensity branch. We use Kullback-Leibler Divergence to measure the similarity in the probability distribution, making the disparity prediction of the event branch closer to the inspection prediction of the intensity branch, formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 203, + 295, + 217 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 203, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 127, + 203, + 295, + 217 + ], + "type": "interline_equation", + "content": "L _ {\\text {l o g i t}} = K L \\left(P _ {E}, P _ {I}\\right). \\tag {7}", + "image_path": "96caa729ae4819e3a2a414f1a3de6c6c522d597b884d0dcfee53ccadd1c661d3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": "In summary, the overall distillation loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "inline_equation", + "content": "L_{distill}" + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": " is the sum of the long-range loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "inline_equation", + "content": "L_{long}" + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": " (2) and focal loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "inline_equation", + "content": "L_{focal}" + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": " (5) in the shallow distillation, the deep distillation loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "inline_equation", + "content": "L_{deep}" + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": " (6), and the logit-level loss " + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "inline_equation", + "content": "L_{logit}" + }, + { + "bbox": [ + 55, + 227, + 296, + 276 + ], + "type": "text", + "content": " (7), as shown below:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 286, + 295, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 286, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 85, + 286, + 295, + 300 + ], + "type": "interline_equation", + "content": "L _ {\\text {d i s t i l l}} = L _ {\\text {l o n g}} + L _ {\\text {f o c a l}} + L _ {\\text {d e e p}} + L _ {\\text {l o g i t}}. \\tag {8}", + "image_path": "e3c72465da997d325a975652d580ab8a31cfffeb9a64db2a56296ad5c3bd8982.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 309, + 288, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 288, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 288, + 323 + ], + "type": "text", + "content": "3.4. Intensity-Event Joint Left-Right Consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "content": "In order to implicitly indicate the quality of the estimated disparity and bring more explicit prompts, we introduce the Intensity-Event joint Left-Right Consistency module (IE-LRC). For a stereo image pair, right image pixels can be shifted using disparity along each epipolar line to construct a warped left intensity image. Minimizing the difference between the warped left image and the real image reduces the ambiguity of the disparity [19]. The proposed IE-LRC module takes the disparity " + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "D^{L} \\in R^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "content": " predicted by the event branch, the left intensity image " + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "I^{L} \\in R^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "content": ", and the right intensity image " + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "I^{R} \\in R^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "content": " as input, as shown in Fig. 4. For stereo intensity image pairs, the right intensity image pixels are shifted using disparity predicted by the event branch along each epipolar to construct the warped left intensity image " + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "inline_equation", + "content": "\\hat{I}^{L} \\in R^{1 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 327, + 296, + 568 + ], + "type": "text", + "content": ". The local appearance matching loss is constructed to minimize the reconstruction loss between the warped and real intensity image. We utilize both the structural similarity index loss and edge-aware smoothness loss to promote local smoothness of disparity, which can be expressed as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 586, + 295, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 295, + 626 + ], + "type": "interline_equation", + "content": "L _ {\\text {s m o o t h}} = \\frac {1}{N} \\sum_ {i, j} \\left| \\partial_ {x} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {x} I _ {i j} ^ {l} \\right\\|} + \\left| \\partial_ {y} D _ {i j} ^ {l} \\right| e ^ {- \\left\\| \\partial_ {y} I _ {i j} ^ {l} \\right\\|}, \\tag {9}", + "image_path": "54a8c9320c641d430ec3a78a3187f990134b41546d06548601ec3fdd10a7978f.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 647, + 295, + 680 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 647, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 66, + 647, + 295, + 680 + ], + "type": "interline_equation", + "content": "L _ {l r c} = L _ {s m o o t h} + \\alpha \\frac {1}{N} \\sum_ {i, j} \\frac {1 - S S I M \\left(I _ {i j} ^ {l} , \\hat {I} _ {i j} ^ {l}\\right)}{2}. \\tag {10}", + "image_path": "446b07efbfbfdbc48ec8daf729da96104ad5af22ecb85a7a7740b66b02a9b12f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "We use a simplified SSIM with a " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " block filter, and set " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\alpha = 0.85" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": ", following [19]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 549, + 160 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 549, + 160 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 549, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 549, + 160 + ], + "type": "image", + "image_path": "bb36dd1cb9826f473882601499fabab372f89adc6c6cf8a9ec7ccfba24fbf4ae.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 165, + 554, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 165, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 554, + 220 + ], + "type": "text", + "content": "Figure 4. Illustration of intensity-event joint left-right consistency module. The input disparity comes from the event branch, and the input intensity images come from the intensity branch. The cross-view consistency across modalities is enforced through edge-aware smoothness loss and SSIM loss." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 232, + 446, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 232, + 446, + 244 + ], + "spans": [ + { + "bbox": [ + 314, + 232, + 446, + 244 + ], + "type": "text", + "content": "3.5. Training and Inference" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "text", + "content": "Overall Loss for Training. The total training loss of our proposed framework consists of three components, namely the stereo matching loss " + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "inline_equation", + "content": "L_{disp}" + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "text", + "content": ", the intensity-to-event distillation loss " + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "inline_equation", + "content": "L_{distill}" + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "text", + "content": " (8), and the left-right consistency loss " + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "inline_equation", + "content": "L_{lrc}" + }, + { + "bbox": [ + 313, + 249, + 554, + 308 + ], + "type": "text", + "content": " (10), as shown below:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 366, + 319, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 319, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 366, + 319, + 553, + 333 + ], + "type": "interline_equation", + "content": "L _ {t o t a l} = L _ {d i s p} + L _ {d i s t i l l} + L _ {l r c}. \\tag {11}", + "image_path": "a0580d3fa1803c01770362aa8379cdc8e7efeba4ea67285c10c5e01ad6c720fd.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 342, + 515, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 515, + 355 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 515, + 355 + ], + "type": "text", + "content": "The stereo matching loss " + }, + { + "bbox": [ + 314, + 342, + 515, + 355 + ], + "type": "inline_equation", + "content": "L_{disp}" + }, + { + "bbox": [ + 314, + 342, + 515, + 355 + ], + "type": "text", + "content": " can be defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 331, + 363, + 553, + 397 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 363, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 331, + 363, + 553, + 397 + ], + "type": "interline_equation", + "content": "L _ {d i s p} (D, D _ {E} ^ {*}) = \\frac {1}{V} \\sum_ {i = 0} ^ {V} S m o o t h _ {L _ {1}} \\left(d _ {i} - d _ {i} ^ {*}\\right), \\tag {12}", + "image_path": "e5bcd6714929a79e1d9010694a150ef395030451e05e1cd8eceebab5c7c34089.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " is the ground truth disparity, " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "D_{E}^{*}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " is the disparity predicted by the event branch, and " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " is the number of valid pixels with ground truths, " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "i\\in \\{1,2,3,\\dots ,V\\}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " is the pixels in the ground truth disparity " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " that can be used for training, and " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "d_{i}^{*}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " is the pixel at the corresponding position in the predicted disparity " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "D_{E}^{*}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": ". The smooth " + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 405, + 554, + 489 + ], + "type": "text", + "content": " loss function can be expressed as follows:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 331, + 498, + 553, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 498, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 331, + 498, + 553, + 525 + ], + "type": "interline_equation", + "content": "\\operatorname {S m o o t h} _ {L _ {1}} (x) = \\left\\{ \\begin{array}{l l} 0. 5 x ^ {2}, & \\text {i f} | x | < 1, \\\\ | x | - 0. 5, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {13}", + "image_path": "431cbe3116c0e2d12a9f5b9cd8eb1f1284fec0bf994f51cf3427f9e3ac5c61f9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 616 + ], + "type": "text", + "content": "Inference. The event branch is fully optimized during training, capturing a massive amount of texture knowledge and dense structure information from intensity images. During inference, retaining only the event branch can achieve efficient and dense disparity prediction without increasing the computational cost, and most importantly, preserving the low latency characteristics of events." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 628, + 391, + 641 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 628, + 391, + 641 + ], + "spans": [ + { + "bbox": [ + 314, + 628, + 391, + 641 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 647, + 437, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 437, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 437, + 661 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Datasets. To evaluate the proposed method's effectiveness, the experiments are performed on two publicly available datasets, MVSEC [69] and DSEC [17], both from the real world. The MVSEC dataset has two DAVIS cameras" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5441" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 62, + 70, + 547, + 205 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 547, + 205 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 547, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 547, + 205 + ], + "type": "table", + "html": "
MethodMean disp. error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓Median depth error [cm] ↓
Split1Spilt3Split1Spilt3Split1Spilt3Split1Spilt3
EIS-E [37]--80.668.313.325.7--
DDES [49]0.590.9489.874.816.727.86.814.7
DTC-PDS [62]0.560.6591.588.715.318.66.48.7
CTC-PDS [62]0.530.7391.688.214.920.66.410.6
EITNet [1]0.550.7592.189.614.219.45.910.4
DTC-SPADE [62]0.460.6093.089.713.517.15.27.9
TES [13]0.460.4992.992.613.015.05.05.8
Ours0.320.3794.893.411.213.24.24.7
", + "image_path": "fb3af3f91612f635ef99ca882a88a4a037a2f06d1973900dc3ac545e2c112d40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 241, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 241, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 296, + 445 + ], + "type": "text", + "content": "with stereo setups that provide intensity images and pairing events. We use indoor flying from the MVSEC dataset, which was captured from a drone flying in a room with various objects. Following [10, 13, 49, 62], we split the MVSEC dataset indoor flying into three and use two of them, split 1 and split 3. For a fair comparison, we use mean disparity error, one-pixel accuracy, mean depth error, and median depth error as metrics. The DSEC dataset is a stereo event camera dataset for outdoor driving scenes. The intensity images and events are captured by different devices with different resolutions. All cameras are at the same height, so we approximately map the intensity images to the corresponding positions of the events with the provided camera matrix. For evaluation, we use the mean absolute error (MAE), one-pixel error (1PE), two-pixel error (2PE), and root mean square error (RMSE) as the metrics. See the supplementary material for more details on the dataset split." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": "Implementation Details. The proposed method is implemented using PyTorch and trained from scratch. We adopt the Adam [27] (" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.99" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": ") optimizer to train the model with a batch size of 8 using " + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "inline_equation", + "content": "256 \\times 256" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": " random crops. We set the maximum disparity to 192. The initial learning rate is " + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-3}" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": ", and the weight decay is " + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 445, + 296, + 554 + ], + "type": "text", + "content": ". All experiments are conducted on an NVIDIA RTX A6000 GPU. For more details about the implementation, please refer to the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 174, + 575 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "We compare the results of the proposed " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{IED}^2\\mathbf{S}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " with state-of-the-art methods. The comparison results are shown in Table 1. EIS-E [37], DDES [49], DTC-PDS [62], EITNet [1], and TES[13] only use asynchronous and sparse events, which limits the effectiveness of stereo matching due to the lack of sufficient texture knowledge and dense structure information. The proposed " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{IED}^2\\mathbf{S}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " fully extracts extensive dense and structural information contained in intensity images and distills it into the event branch during the training, making its performance significantly better than existing event-based stereo matching methods. Specifically," + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 315, + 239, + 554, + 316 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "type": "text", + "content": "Table 1. Disparity estimation results on the MVSEC [69] dataset. - indicates that results are not provided in the original paper. The best is in bold and the second best is in underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 239, + 554, + 316 + ], + "lines": [ + { + "bbox": [ + 315, + 239, + 554, + 316 + ], + "spans": [ + { + "bbox": [ + 315, + 239, + 554, + 316 + ], + "type": "table", + "html": "
MethodsMAE ↓1PE ↓2PE ↓RMSE ↓
DDES [49]0.57610.9152.9051.381
DTC-PDS [62]0.5269.5342.3531.263
Se-CFF [38]0.5219.5862.6231.235
Ours0.4938.8232.1341.069
", + "image_path": "9cc6f1e05a6610678a70a28db4a6109babdcc5bfc67feb7132486c3b2a5b129a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 321, + 548, + 333 + ], + "lines": [ + { + "bbox": [ + 318, + 321, + 548, + 333 + ], + "spans": [ + { + "bbox": [ + 318, + 321, + 548, + 333 + ], + "type": "text", + "content": "Table 2. Disparity estimation results on the DSEC [17] dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "text", + "content": "compared with the best-performing event-based method, TES[13], the proposed method outperforms it by " + }, + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "inline_equation", + "content": "13.85\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "inline_equation", + "content": "12.00\\%" + }, + { + "bbox": [ + 313, + 342, + 555, + 509 + ], + "type": "text", + "content": " in terms of mean depth error for split1 and split3, respectively. We believe that this improvement comes from fully considering the modality gap and the novel design of the distillation strategy. Meanwhile, it is worth noting that the combined intensity images and events methods require both events and intensity images to be input at the inference stage, which limits the application of some scenarios, such as high-speed motion scenes or when intensity images are unavailable. However, ours only requires events as input during the inference phase, which fully utilizes the low-latency characteristics of events in special scenarios (high speed and extreme lighting)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "text", + "content": "Furthermore, we evaluate the proposed model on the DSEC dataset, and the results are shown in Table 2. Similarly, compared with the state-of-the-art baselines, ours achieves the best performance on all metrics and outperforms existing methods by a large margin, thanks to the designed cross-modal distillation in Sec. 3. Specifically, it outperforms Se-CFF [38] by " + }, + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "inline_equation", + "content": "5.37\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "inline_equation", + "content": "7.46\\%" + }, + { + "bbox": [ + 313, + 510, + 556, + 617 + ], + "type": "text", + "content": " in terms of mean absolute error and one-pixel error, respectively, which verifies the effectiveness of the proposed method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 425, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 425, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 425, + 636 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 713 + ], + "type": "text", + "content": "As shown in Fig. 5, we present a qualitative comparison of our method with the baseline in various scenarios on the MVSEC dataset. DDES [49] and TES [13] struggle to output dense disparity due to the lack of detailed information caused by the inherent sparsity of events. They are not ideal in some locations, such as the stripe lines dis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5442" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 83, + 159, + 142 + ], + "blocks": [ + { + "bbox": [ + 107, + 75, + 143, + 82 + ], + "lines": [ + { + "bbox": [ + 107, + 75, + 143, + 82 + ], + "spans": [ + { + "bbox": [ + 107, + 75, + 143, + 82 + ], + "type": "text", + "content": "(a) Intensity" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 88, + 83, + 159, + 142 + ], + "lines": [ + { + "bbox": [ + 88, + 83, + 159, + 142 + ], + "spans": [ + { + "bbox": [ + 88, + 83, + 159, + 142 + ], + "type": "image", + "image_path": "58cb30f1dfd709951ac92ff87597200830eb71f08dd763da73ab73c153a332d6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 160, + 83, + 231, + 141 + ], + "blocks": [ + { + "bbox": [ + 181, + 75, + 211, + 82 + ], + "lines": [ + { + "bbox": [ + 181, + 75, + 211, + 82 + ], + "spans": [ + { + "bbox": [ + 181, + 75, + 211, + 82 + ], + "type": "text", + "content": "(b) Events" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 160, + 83, + 231, + 141 + ], + "lines": [ + { + "bbox": [ + 160, + 83, + 231, + 141 + ], + "spans": [ + { + "bbox": [ + 160, + 83, + 231, + 141 + ], + "type": "image", + "image_path": "37ed519bf0c3aab5a98a0c72294d7aae677da59230bc058f4fb24273c7888eb3.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 233, + 83, + 303, + 142 + ], + "blocks": [ + { + "bbox": [ + 253, + 75, + 282, + 82 + ], + "lines": [ + { + "bbox": [ + 253, + 75, + 282, + 82 + ], + "spans": [ + { + "bbox": [ + 253, + 75, + 282, + 82 + ], + "type": "text", + "content": "(c) DDES" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 83, + 303, + 142 + ], + "lines": [ + { + "bbox": [ + 233, + 83, + 303, + 142 + ], + "spans": [ + { + "bbox": [ + 233, + 83, + 303, + 142 + ], + "type": "image", + "image_path": "5d4d4fa603bdc1d2aa5c16a7a4032593d67c7ebaefeea9b6d5e2df7b683c9309.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 304, + 83, + 375, + 142 + ], + "blocks": [ + { + "bbox": [ + 328, + 75, + 352, + 82 + ], + "lines": [ + { + "bbox": [ + 328, + 75, + 352, + 82 + ], + "spans": [ + { + "bbox": [ + 328, + 75, + 352, + 82 + ], + "type": "text", + "content": "(d) TES" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 83, + 375, + 142 + ], + "lines": [ + { + "bbox": [ + 304, + 83, + 375, + 142 + ], + "spans": [ + { + "bbox": [ + 304, + 83, + 375, + 142 + ], + "type": "image", + "image_path": "6892d8279cb9b270df1f96c35857e8fe448420c5d2de80668283cbcbb1430858.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 376, + 83, + 447, + 142 + ], + "blocks": [ + { + "bbox": [ + 400, + 75, + 424, + 82 + ], + "lines": [ + { + "bbox": [ + 400, + 75, + 424, + 82 + ], + "spans": [ + { + "bbox": [ + 400, + 75, + 424, + 82 + ], + "type": "text", + "content": "(e) Ours" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 376, + 83, + 447, + 142 + ], + "lines": [ + { + "bbox": [ + 376, + 83, + 447, + 142 + ], + "spans": [ + { + "bbox": [ + 376, + 83, + 447, + 142 + ], + "type": "image", + "image_path": "43560092a0f315aac53e7f503813a769e8a90e55b729eddbde288cc0f1b34f9c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 449, + 83, + 520, + 142 + ], + "blocks": [ + { + "bbox": [ + 460, + 75, + 509, + 82 + ], + "lines": [ + { + "bbox": [ + 460, + 75, + 509, + 82 + ], + "spans": [ + { + "bbox": [ + 460, + 75, + 509, + 82 + ], + "type": "text", + "content": "(f) Ground Truth" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 449, + 83, + 520, + 142 + ], + "lines": [ + { + "bbox": [ + 449, + 83, + 520, + 142 + ], + "spans": [ + { + "bbox": [ + 449, + 83, + 520, + 142 + ], + "type": "image", + "image_path": "39a8daf66e7229aa4a6be50e46228e7c7a104517e1d5e11fd4cfa3833fcdaba7.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 88, + 144, + 158, + 203 + ], + "blocks": [ + { + "bbox": [ + 88, + 144, + 158, + 203 + ], + "lines": [ + { + "bbox": [ + 88, + 144, + 158, + 203 + ], + "spans": [ + { + "bbox": [ + 88, + 144, + 158, + 203 + ], + "type": "image", + "image_path": "42929fc331a6eafdc51542d0926137ab52682cb5a4eb1d0e94282dc9cf4b3101.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 160, + 144, + 231, + 204 + ], + "blocks": [ + { + "bbox": [ + 160, + 144, + 231, + 204 + ], + "lines": [ + { + "bbox": [ + 160, + 144, + 231, + 204 + ], + "spans": [ + { + "bbox": [ + 160, + 144, + 231, + 204 + ], + "type": "image", + "image_path": "e6b5b457e4762ee87d6443594f6fdb049ec003eb0d8ba48e898a69d754c350b5.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 232, + 144, + 303, + 203 + ], + "blocks": [ + { + "bbox": [ + 232, + 144, + 303, + 203 + ], + "lines": [ + { + "bbox": [ + 232, + 144, + 303, + 203 + ], + "spans": [ + { + "bbox": [ + 232, + 144, + 303, + 203 + ], + "type": "image", + "image_path": "5876d02dc277ea3d3e54e61c7497cf0ae89b4db37e8cc014ba891ce2fa2ab95f.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 304, + 144, + 375, + 203 + ], + "blocks": [ + { + "bbox": [ + 304, + 144, + 375, + 203 + ], + "lines": [ + { + "bbox": [ + 304, + 144, + 375, + 203 + ], + "spans": [ + { + "bbox": [ + 304, + 144, + 375, + 203 + ], + "type": "image", + "image_path": "1fb4cdc86ef1da8c60fd57272bb0fc518162a135d48e5a356c734ccebdb7947b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 376, + 144, + 447, + 204 + ], + "blocks": [ + { + "bbox": [ + 376, + 144, + 447, + 204 + ], + "lines": [ + { + "bbox": [ + 376, + 144, + 447, + 204 + ], + "spans": [ + { + "bbox": [ + 376, + 144, + 447, + 204 + ], + "type": "image", + "image_path": "a84e706cd9a7e9bfc47cde48ab88d2596e8fbe7e5e236b96fa2dfdf81a8be524.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 449, + 144, + 520, + 204 + ], + "blocks": [ + { + "bbox": [ + 449, + 144, + 520, + 204 + ], + "lines": [ + { + "bbox": [ + 449, + 144, + 520, + 204 + ], + "spans": [ + { + "bbox": [ + 449, + 144, + 520, + 204 + ], + "type": "image", + "image_path": "c3c51fe3b9432df8432c28c62156e9af10a44248377824b125417c6ba9cfd264.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 88, + 205, + 158, + 264 + ], + "blocks": [ + { + "bbox": [ + 88, + 205, + 158, + 264 + ], + "lines": [ + { + "bbox": [ + 88, + 205, + 158, + 264 + ], + "spans": [ + { + "bbox": [ + 88, + 205, + 158, + 264 + ], + "type": "image", + "image_path": "5b74272c89e4772ee380c51224711f604ce2aa6f77146bd4cdfc5617bd81a724.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 160, + 205, + 231, + 264 + ], + "blocks": [ + { + "bbox": [ + 160, + 205, + 231, + 264 + ], + "lines": [ + { + "bbox": [ + 160, + 205, + 231, + 264 + ], + "spans": [ + { + "bbox": [ + 160, + 205, + 231, + 264 + ], + "type": "image", + "image_path": "e718529e166f432e7ad36b7c1eadf5325c587a5f2c64c325a02d1475748c3ce9.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 232, + 205, + 303, + 264 + ], + "blocks": [ + { + "bbox": [ + 232, + 205, + 303, + 264 + ], + "lines": [ + { + "bbox": [ + 232, + 205, + 303, + 264 + ], + "spans": [ + { + "bbox": [ + 232, + 205, + 303, + 264 + ], + "type": "image", + "image_path": "551349293e567a1a6737d62d9aa13fea7ea4db1807bbf199e4df9fb52f6e6ac3.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 304, + 205, + 375, + 264 + ], + "blocks": [ + { + "bbox": [ + 304, + 205, + 375, + 264 + ], + "lines": [ + { + "bbox": [ + 304, + 205, + 375, + 264 + ], + "spans": [ + { + "bbox": [ + 304, + 205, + 375, + 264 + ], + "type": "image", + "image_path": "014947b9143a97eb457e2a0509b99b9d7bf82c896e3a68c119e6d83c4394ce30.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 376, + 205, + 447, + 264 + ], + "blocks": [ + { + "bbox": [ + 376, + 205, + 447, + 264 + ], + "lines": [ + { + "bbox": [ + 376, + 205, + 447, + 264 + ], + "spans": [ + { + "bbox": [ + 376, + 205, + 447, + 264 + ], + "type": "image", + "image_path": "d58eddf5ab90983210cf7f52e303049e393067d62f2af813cf8ac00b488e8ea7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 449, + 205, + 520, + 264 + ], + "blocks": [ + { + "bbox": [ + 449, + 205, + 520, + 264 + ], + "lines": [ + { + "bbox": [ + 449, + 205, + 520, + 264 + ], + "spans": [ + { + "bbox": [ + 449, + 205, + 520, + 264 + ], + "type": "image", + "image_path": "da0c009a26bddc00f9b337cb826dac440dc05ae5893c323cf0095c32d902276f.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 88, + 266, + 158, + 325 + ], + "blocks": [ + { + "bbox": [ + 88, + 266, + 158, + 325 + ], + "lines": [ + { + "bbox": [ + 88, + 266, + 158, + 325 + ], + "spans": [ + { + "bbox": [ + 88, + 266, + 158, + 325 + ], + "type": "image", + "image_path": "b8092e78bbe8e8877750e5d3da848cbf2bda2e1de7c6322229c69c7e58df471a.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 328, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 55, + 328, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 555, + 384 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison of dense disparity estimation for indoor flying scenes in the MVSEC dataset. The first two columns are the intensity images and the corresponding events for visual reference, where the red and blue discrete points represent positive and negative events, respectively. (c), (d), and (e) are the results of DDES [49], TES [13], and ours, respectively. Our proposed method fully uses the intensity image's detail and texture information during the training stage, better displaying fine details in the region highlighted by the red box. Detailed disparity information is framed by the red box for comparison." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 160, + 266, + 231, + 325 + ], + "blocks": [ + { + "bbox": [ + 160, + 266, + 231, + 325 + ], + "lines": [ + { + "bbox": [ + 160, + 266, + 231, + 325 + ], + "spans": [ + { + "bbox": [ + 160, + 266, + 231, + 325 + ], + "type": "image", + "image_path": "d0bc4d6091119dfb1833c5b4ce7d4fd585919feb9a6c14912974c1fb52c49783.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 232, + 266, + 303, + 324 + ], + "blocks": [ + { + "bbox": [ + 232, + 266, + 303, + 324 + ], + "lines": [ + { + "bbox": [ + 232, + 266, + 303, + 324 + ], + "spans": [ + { + "bbox": [ + 232, + 266, + 303, + 324 + ], + "type": "image", + "image_path": "067a424c94053d08475bf63ee39263493a094936bc1f71dd9cf1052d21a50093.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 304, + 266, + 375, + 324 + ], + "blocks": [ + { + "bbox": [ + 304, + 266, + 375, + 324 + ], + "lines": [ + { + "bbox": [ + 304, + 266, + 375, + 324 + ], + "spans": [ + { + "bbox": [ + 304, + 266, + 375, + 324 + ], + "type": "image", + "image_path": "0db7a3c8dc200af2701023449c63174e342bc7c9844a2dec3b02290f87f0dbf4.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 376, + 266, + 447, + 324 + ], + "blocks": [ + { + "bbox": [ + 376, + 266, + 447, + 324 + ], + "lines": [ + { + "bbox": [ + 376, + 266, + 447, + 324 + ], + "spans": [ + { + "bbox": [ + 376, + 266, + 447, + 324 + ], + "type": "image", + "image_path": "45ebc35ede989c22063ab41d59476f6d8e239ff4ceb2dabc15a57d3f1fca5efe.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 449, + 266, + 520, + 324 + ], + "blocks": [ + { + "bbox": [ + 449, + 266, + 520, + 324 + ], + "lines": [ + { + "bbox": [ + 449, + 266, + 520, + 324 + ], + "spans": [ + { + "bbox": [ + 449, + 266, + 520, + 324 + ], + "type": "image", + "image_path": "943260f303efdc1ed17023ce53b9594c1bd4d62cdf4e5510fafa9bc2f942dd1e.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "bbox": [ + 55, + 391, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 547 + ], + "type": "text", + "content": "appearing where the disparity should change dramatically. For example, blur appears at the junction of objects such as barrels and chairs with the background. In contrast, our method addresses these challenges by enabling the model to capture more intricate features, thereby offering more reliable information for accurate disparity estimation in ill-posed regions. For example, for a chair placed far away, the proposed method achieves fine-grained disparity prediction. Consequently, even in areas where event data is sparse and closely resembles the surrounding environment, our approach produces disparities that are not only clearer and denser but also exhibit more precise boundaries, such as the area outlined by the red box in Fig. 5." + } + ] + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 69, + 569, + 121, + 606 + ], + "blocks": [ + { + "bbox": [ + 81, + 563, + 111, + 569 + ], + "lines": [ + { + "bbox": [ + 81, + 563, + 111, + 569 + ], + "spans": [ + { + "bbox": [ + 81, + 563, + 111, + 569 + ], + "type": "text", + "content": "(a) Intensity" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 69, + 569, + 121, + 606 + ], + "lines": [ + { + "bbox": [ + 69, + 569, + 121, + 606 + ], + "spans": [ + { + "bbox": [ + 69, + 569, + 121, + 606 + ], + "type": "image", + "image_path": "e13e1e32e265aeba5dd28b46dad203975f0b4cf1ea5aee7ece49da17b56936be.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 123, + 570, + 174, + 606 + ], + "blocks": [ + { + "bbox": [ + 135, + 563, + 163, + 569 + ], + "lines": [ + { + "bbox": [ + 135, + 563, + 163, + 569 + ], + "spans": [ + { + "bbox": [ + 135, + 563, + 163, + 569 + ], + "type": "text", + "content": "(b) Sc-CFF" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 123, + 570, + 174, + 606 + ], + "lines": [ + { + "bbox": [ + 123, + 570, + 174, + 606 + ], + "spans": [ + { + "bbox": [ + 123, + 570, + 174, + 606 + ], + "type": "image", + "image_path": "bb06f6a66af0ab5a77c4e1fcdcf3bcfd64c642e6fd3548cee46cd8e6ec68efda.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 175, + 570, + 227, + 606 + ], + "blocks": [ + { + "bbox": [ + 190, + 563, + 208, + 569 + ], + "lines": [ + { + "bbox": [ + 190, + 563, + 208, + 569 + ], + "spans": [ + { + "bbox": [ + 190, + 563, + 208, + 569 + ], + "type": "text", + "content": "(c) TES" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 175, + 570, + 227, + 606 + ], + "lines": [ + { + "bbox": [ + 175, + 570, + 227, + 606 + ], + "spans": [ + { + "bbox": [ + 175, + 570, + 227, + 606 + ], + "type": "image", + "image_path": "c3b4b9ecc6f8dccce080218c9ea65fe23a1b82dd51f832d2e3d83efa7de2d801.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 227, + 570, + 279, + 606 + ], + "blocks": [ + { + "bbox": [ + 242, + 563, + 264, + 569 + ], + "lines": [ + { + "bbox": [ + 242, + 563, + 264, + 569 + ], + "spans": [ + { + "bbox": [ + 242, + 563, + 264, + 569 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 227, + 570, + 279, + 606 + ], + "lines": [ + { + "bbox": [ + 227, + 570, + 279, + 606 + ], + "spans": [ + { + "bbox": [ + 227, + 570, + 279, + 606 + ], + "type": "image", + "image_path": "dc6816f2934fa445ba481159e62fdcc8166dd8005f43e8993582e02c5f3f5403.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 70, + 608, + 121, + 645 + ], + "blocks": [ + { + "bbox": [ + 70, + 608, + 121, + 645 + ], + "lines": [ + { + "bbox": [ + 70, + 608, + 121, + 645 + ], + "spans": [ + { + "bbox": [ + 70, + 608, + 121, + 645 + ], + "type": "image", + "image_path": "536721dbf39190a30b3f10f50c64b24e61043025db68a1a8ee3a40520210dca0.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 123, + 608, + 174, + 645 + ], + "blocks": [ + { + "bbox": [ + 123, + 608, + 174, + 645 + ], + "lines": [ + { + "bbox": [ + 123, + 608, + 174, + 645 + ], + "spans": [ + { + "bbox": [ + 123, + 608, + 174, + 645 + ], + "type": "image", + "image_path": "dcee0aece8a6b08245878424c5fce1a5bb1037ae4b921ee0a765012a06a04674.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 175, + 608, + 227, + 645 + ], + "blocks": [ + { + "bbox": [ + 175, + 608, + 227, + 645 + ], + "lines": [ + { + "bbox": [ + 175, + 608, + 227, + 645 + ], + "spans": [ + { + "bbox": [ + 175, + 608, + 227, + 645 + ], + "type": "image", + "image_path": "63e5e939c434f807f199d17dbd2536ba6de73a9015e9dc06bb7c26d273955f95.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "image", + "bbox": [ + 227, + 608, + 279, + 645 + ], + "blocks": [ + { + "bbox": [ + 227, + 608, + 279, + 645 + ], + "lines": [ + { + "bbox": [ + 227, + 608, + 279, + 645 + ], + "spans": [ + { + "bbox": [ + 227, + 608, + 279, + 645 + ], + "type": "image", + "image_path": "2044afabb9c685fe9f62eeb33436b1d8c3cbe090190568668c32d6c5779cd330.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + } + ], + "index": 43 + }, + { + "type": "image", + "bbox": [ + 70, + 647, + 121, + 684 + ], + "blocks": [ + { + "bbox": [ + 70, + 647, + 121, + 684 + ], + "lines": [ + { + "bbox": [ + 70, + 647, + 121, + 684 + ], + "spans": [ + { + "bbox": [ + 70, + 647, + 121, + 684 + ], + "type": "image", + "image_path": "da3466bc082759a2277aa57255176c8ae03d453448ff1c5eaa992e68cb6903e9.jpg" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 688, + 295, + 710 + ], + "type": "text", + "content": "Figure 6. Qualitative results on the DSEC dataset. Our method can construct more dense disparities than others." + } + ] + } + ], + "index": 48, + "angle": 0, + "type": "image_caption" + } + ], + "index": 44 + }, + { + "type": "image", + "bbox": [ + 123, + 647, + 174, + 684 + ], + "blocks": [ + { + "bbox": [ + 123, + 647, + 174, + 684 + ], + "lines": [ + { + "bbox": [ + 123, + 647, + 174, + 684 + ], + "spans": [ + { + "bbox": [ + 123, + 647, + 174, + 684 + ], + "type": "image", + "image_path": "4f4af12a9a16ea7eba34307b40017842a0967681425f458d3ca86543ae2305ac.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_body" + } + ], + "index": 45 + }, + { + "type": "image", + "bbox": [ + 175, + 647, + 227, + 684 + ], + "blocks": [ + { + "bbox": [ + 175, + 647, + 227, + 684 + ], + "lines": [ + { + "bbox": [ + 175, + 647, + 227, + 684 + ], + "spans": [ + { + "bbox": [ + 175, + 647, + 227, + 684 + ], + "type": "image", + "image_path": "62a34741376023e3666e55e748a83db532bb751505286888d00a4916178e186a.jpg" + } + ] + } + ], + "index": 46, + "angle": 0, + "type": "image_body" + } + ], + "index": 46 + }, + { + "type": "image", + "bbox": [ + 227, + 647, + 279, + 684 + ], + "blocks": [ + { + "bbox": [ + 227, + 647, + 279, + 684 + ], + "lines": [ + { + "bbox": [ + 227, + 647, + 279, + 684 + ], + "spans": [ + { + "bbox": [ + 227, + 647, + 279, + 684 + ], + "type": "image", + "image_path": "0ca9ba52097ce76f5357d9e62279a3796ea7290c93b335cbd684b8f91866cfa1.jpg" + } + ] + } + ], + "index": 47, + "angle": 0, + "type": "image_body" + } + ], + "index": 47 + }, + { + "bbox": [ + 313, + 391, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 463 + ], + "type": "text", + "content": "Additionally, we present results on the DSEC datasets in Fig. 6. We highlight regions with complex textures, where our method outperforms the others in predicting disparity with higher accuracy. For example, for objects such as trees and guardrails on the roadside, we obtain much sharper, more detailed, and artifact-free disparities." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 468, + 415, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 415, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 415, + 480 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Contribution of each component. We perform extensive experiments on the MVSEC dataset to fully validate the effectiveness of each proposed component and related design choices in Sec. 3. We first present the results of the ablation study in Table 3. It can be observed that all the proposed components have contributed to the superior performance. This indicates that the information derived from intensity images is effective for dense event-based disparity estimation without significantly altering the network structure. When only one component is added, we see the most improvement comes from the feature distillation, which indicates that modality-specific and task-related features of the intensity branch are needed. When it is applied solely, the one-pixel accuracy is increased by " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "3.74\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "2.91\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " for split1 and split3, respectively. It is worth noting that intensity-event joint left-right consistency improves the one-pixel accuracy of split1 and split3 by " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "0.22\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.01\\%" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": ", respectively, which verifies the feasibility of improving the matching effect through ensuring cross-view consistency in" + } + ] + } + ], + "index": 51 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5443" + } + ] + } + ], + "index": 52 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 549, + 205 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 549, + 205 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 549, + 205 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 549, + 205 + ], + "type": "table", + "html": "
F-distill.L-distill.LRCMean disparity error [pix] ↓One-pixel accuracy [%] ↑Mean depth error [cm] ↓
Split 1Split 3Split 1Split 3Split1Split3
XXX0.540.6791.089.415.018.5
✓XX0.380.4694.492.012.113.9
X✓X0.490.5391.991.613.817.5
XX✓0.500.6491.290.314.618.0
X✓✓0.490.5292.191.613.717.4
✓X✓0.370.4194.592.711.613.7
✓✓X0.340.3994.893.211.313.4
✓✓✓0.320.3794.893.411.213.2
", + "image_path": "bf4f33ec5a2a408f05ef9f0bf63ba922efa65d5da877a5037d4e6ea030e4ee2d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "text", + "content": "stereo matching. When all components are combined, the performance is further improved without any conflicts, resulting in the best performance with a significant improvement of the one-pixel accuracy by " + }, + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "inline_equation", + "content": "4.18\\%" + }, + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "inline_equation", + "content": "4.47\\%" + }, + { + "bbox": [ + 55, + 243, + 295, + 304 + ], + "type": "text", + "content": " for split1 and split3, respectively." + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 315, + 293, + 415 + ], + "blocks": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 553, + 232 + ], + "type": "text", + "content": "Table 3. The result of our intensity-to-event distillation scheme with different settings on the MVSEC dataset. F-distill. denotes the feature distillation including shallow and deep distillation, L-distill. represents logit-level distillation, and LRC means left-right consistency." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 315, + 293, + 415 + ], + "lines": [ + { + "bbox": [ + 57, + 315, + 293, + 415 + ], + "spans": [ + { + "bbox": [ + 57, + 315, + 293, + 415 + ], + "type": "table", + "html": "
MethodsMean disparity error [pix] ↓One-pixel accuracy [%] ↑
Split1Split3Split1Split3
MSE0.450.5092.291.8
Si2eD0.390.4693.092.5
MSE+Di2eD0.370.4194.093.0
Si2eD+Di2eD0.320.3794.893.4
", + "image_path": "e80ec51c3c0e1de92b6979de03a102ca2e68bbe09780dfc9102e9cc1a3857966.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Feature alignment. Although feature distillation has been a widely used loss term in many knowledge distillation studies [28, 39], we are the first to implement that for event-based stereo. For further exploration and discussion, we present our experiments to validate the design choice of the feature distillation module, as it brings the most improvement among all the proposed components. Specifically, to solely examine the effectiveness of our proposed shallow distillation, we compare it to the simple alignment feature-level distillation methods. MSE is a baseline method [22] that simply minimizes the element-wise " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " distance between the features of the teacher and student. As shown in Table 4, directly aligning the features of the two modalities contaminates the original modality-specific information and introduces noise to the event branch. Specifically, the one-pixel accuracy is reduced by " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.74\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.71\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": " for split1 and split3, respectively. A possible reason for this performance gain comes from the large modality difference between events and intensity images, and the novel design fully exploits these modality-specific features while effectively avoiding cross-sensor contamination." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 316, + 240, + 553, + 304 + ], + "blocks": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "lines": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 295, + 453 + ], + "type": "text", + "content": "Table 4. The effectiveness of our designed feature distillation on the MVSEC dataset. Si2eD represents shallow distillation, and Di2eD denotes deep distillation." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 240, + 553, + 304 + ], + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 304 + ], + "type": "table", + "html": "
SettingMVSEC -> DSECDSEC -> MVSEC
MetricsMAE ↓1PE ↓Mean disp. error [pix] ↓One-pixel acc. [%] ↑
w/o distill.0.78212.4130.7479.8
Full model0.5639.9730.4691.4
", + "image_path": "c22bc89ff60edd7e2ed75dcce41621c6a320481757041aec80c6e6baf5a75c87.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 331 + ], + "type": "text", + "content": "Table 5. Cross-validation experimental results between two real-world datasets, verifying good generalization." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 338, + 539, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 539, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 539, + 350 + ], + "type": "text", + "content": "4.5. Cross-Domain Generalization Performance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 489 + ], + "type": "text", + "content": "The modality-specific knowledge obtained by cross-sensor distillation is independent of the dataset and has good generalization under different domain distributions. In this section, we discuss the generalization performance of the proposed method on unseen domains. Specifically, we perform cross-validation on two real-world datasets, i.e., the model is trained on MVSEC (DSEC) and evaluated on DSEC (MVSEC). As shown in Table 5, the designed module is less affected by the domain gap. This is because the designed cross-modal distillation enables the event branch to acquire general knowledge for dense stereo matching." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 523, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 656 + ], + "type": "text", + "content": "We propose a novel framework for event-based dense stereo via cross-sensor knowledge distillation. We design a multi-level intensity-to-event distillation strategy, achieving binocular-to-binocular distillation for the first time. A massive amount of texture knowledge and dense structure information contained in the intensity image is distilled to the event branch through the designed strategy. At the same time, an intensity-event joint left-right consistency module is proposed to enforce cross-view consistency. Extensive experiments on the MVSEC and DSEC datasets demonstrate the effectiveness of the proposed framework." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 669, + 411, + 681 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 411, + 681 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 411, + 681 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 715 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under grant No.62371450." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5444" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Soikat Hasan Ahmed, Hae Woong Jang, SM Nadim Uddin, and Yong Ju Jung. Deep event stereo leveraged by event-to-image translation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 882-890, 2021. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Michael Bleyer, Christoph Rhemann, and Carsten Rother. Patchmatch stereo-stereo matching with slanted support windows. In British Machine Vision Conference, pages 1-11, 2011. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "text", + "content": "[3] Christian Brandli, Raphael Berner, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. A " + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "text", + "content": " 130 db " + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "inline_equation", + "content": "3\\mu s" + }, + { + "bbox": [ + 62, + 180, + 296, + 223 + ], + "type": "text", + "content": " latency global shutter spatiotemporal vision sensor. IEEE Journal of Solid-State Circuits, 49(10):2333-2341, 2014. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 278 + ], + "type": "text", + "content": "[4] Luis A Camunas-Mesa, Teresa Serrano-Gotarredona, Sio H Ieng, Ryad B Benosman, and Bernabe Linares-Barranco. On the use of orientation filters for 3d reconstruction in eventdriven stereo vision. Frontiers in neuroscience, 8:48, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 294, + 334 + ], + "type": "text", + "content": "[5] Yue Cao, Jiarui Xu, Stephen Lin, Fangyun Wei, and Han Hu. Gcnet: Non-local networks meet squeeze-excitation networks and beyond. In Proceedings of the IEEE/CVF international conference on computer vision workshops, pages 0–0, 2019. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "text", + "content": "[6] Jia-Ren Chang and Yong-Sheng Chen. Pyramid stereo matching network. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5410-5418, 2018. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 423 + ], + "type": "text", + "content": "[7] Shoushun Chen and Menghan Guo. Live demonstration: Celex-v: A 1m pixel multi-mode event-based sensor. In IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 1682-1683. IEEE, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 425, + 294, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 467 + ], + "type": "text", + "content": "[8] Zehui Chen, Zhenyu Li, Shiquan Zhang, Liangji Fang, Qin-hong Jiang, and Feng Zhao. Bevdistill: Cross-modal bev distillation for multi-view 3d object detection. arXiv preprint arXiv:2211.09386, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 523 + ], + "type": "text", + "content": "[9] Xuelian Cheng, Yiran Zhong, Mehrtash Harandi, Yuchao Dai, Xiaojun Chang, Hongdong Li, Tom Drummond, and Zongyuan Ge. Hierarchical neural architecture search for deep stereo matching. Advances in neural information processing systems, 33:22158-22169, 2020. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 296, + 567 + ], + "type": "text", + "content": "[10] Hoonhee Cho and Kuk-Jin Yoon. Event-image fusion stereo using cross-modality feature propagation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 454-462, 2022. 1, 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 569, + 294, + 612 + ], + "type": "text", + "content": "[11] Hoonhee Cho and Kuk-Jin Yoon. Selection and cross similarity for event-image deep stereo. In Proceedings of the European Conference on Computer Vision, pages 470-486. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[12] Hoonhee Cho, Jaeseok Jeong, and Kuk-Jin Yoon. Eomvs: Event-based omnidirectional multi-view stereo. IEEE Robotics and Automation Letters, 6(4):6709-6716, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[13] Hoonhee Cho, Jae-Young Kang, and Kuk-Jin Yoon. Temporal event stereo via joint learning with stereoscopic flow. In Proceedings of the European Conference on Computer Vision, pages 294–314. Springer, 2025. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[14] Zhiyu Chong, Xinzhu Ma, Hong Zhang, Yuxin Yue, Haojie Li, Zhihui Wang, and Wanli Ouyang. Monodistill: Learn" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "ing spatial features for monocular 3d object detection. arXiv preprint arXiv:2201.10830, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 98, + 553, + 152 + ], + "type": "text", + "content": "[15] Weiqin Chuah, Ruwan Tennakoon, Reza Hoseinnezhad, David Suter, and Alireza Bab-Hadiashar. Semantic guided long range stereo depth estimation for safer autonomous vehicle applications. IEEE Transactions on Intelligent Transportation Systems, 23(10):18916-18926, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 154, + 553, + 196 + ], + "type": "text", + "content": "[16] Boyu Gao, Haoxiang Lang, and Jing Ren. Stereo visual slam for autonomous vehicles: A review. In IEEE International Conference on Systems, Man, and Cybernetics, pages 1316-1322. IEEE, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 200, + 553, + 243 + ], + "type": "text", + "content": "[17] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 277 + ], + "type": "text", + "content": "[18] Suman Ghosh and Guillermo Gallego. Event-based stereo depth estimation: A survey. arXiv preprint arXiv:2409.17680, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 553, + 334 + ], + "type": "text", + "content": "[19] Clément Godard, Oisin Mac Aodha, and Gabriel J Brostow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 392 + ], + "type": "text", + "content": "[20] Rostam Affendi Hamzah, A Fauzan Kadmin, M Saad Hamid, S Fakhar A Ghani, and Haidi Ibrahim. Improvement of stereo matching algorithm for 3d surface reconstruction. Signal Processing: Image Communication, 65:165-172, 2018. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "text", + "content": "[21] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 473 + ], + "type": "text", + "content": "[22] Yu Hong, Hang Dai, and Yong Ding. Cross-modality knowledge distillation network for monocular 3d object detection. In Proceedings of the European Conference on Computer Vision, pages 87-104. Springer, 2022. 2, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "text", + "content": "[23] Asmaa Hosni, Michael Bleyer, Margrit Gelautz, and Christoph Rhemann. Local stereo matching using geodesic support weights. In 16th IEEE International Conference on Image Processing, pages 2093-2096. IEEE, 2009. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 521, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 574 + ], + "type": "text", + "content": "[24] Asmaa Hosni, Christoph Rhemann, Michael Bleyer, Carsten Rother, and Margrit Gelautz. Fast cost-volume filtering for visual correspondence and beyond. IEEE transactions on pattern analysis and machine intelligence, 35(2):504-511, 2012. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 632 + ], + "type": "text", + "content": "[25] Maximilian Jaritz, Tuan-Hung Vu, Raoul de Charette, Emilie Wirbel, and Patrick Pérez. xmuda: Cross-modal unsupervised domain adaptation for 3d semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12605-12614, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[26] Maximilian Jaritz, Tuan-Hung Vu, Raoul De Charette, Émilie Wirbel, and Patrick Pérez. Cross-modal learning for domain adaptation in 3d semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1533-1544, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5445" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[28] Nikos Komodakis and Sergey Zagoruyko. Paying more attention to attention: improving the performance of convolutional neural networks via attention transfer. In International Conference on Learning Representations, 2017. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[29] Hamid Laga, Laurent Valentin Jospin, Farid Boussaid, and Mohammed Bennamoun. A survey on deep learning techniques for stereo-based depth estimation. IEEE transactions on pattern analysis and machine intelligence, 44(4):1738-1764, 2020. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[30] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Transactions on Image Processing, 31:2975-2987, 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[31] Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Zwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent network with adaptive correlation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16263-16272, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[32] Xineng Li, Chen Zhang, Wanjuan Su, and Wenbing Tao. Inet: Implicit intra-inter information fusion for real-time stereo matching. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 3225-3233, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 376 + ], + "type": "text", + "content": "[33] Biyang Liu, Huimin Yu, and Yangqi Long. Local similarity pattern and cost self-reassembling for deep stereo matching networks. In Proceedings of the AAAI conference on artificial intelligence, pages 1647-1655, 2022. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 294, + 431 + ], + "type": "text", + "content": "[34] Tianshan Liu, Kin-Man Lam, Rui Zhao, and Guoping Qiu. Deep cross-modal representation learning and distillation for illumination-invariant pedestrian detection. IEEE Transactions on Circuits and Systems for Video Technology, 32(1): 315-329, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 476 + ], + "type": "text", + "content": "[35] Xu Liu, Jianing Li, Jinqiao Shi, Xiaopeng Fan, Yonghong Tian, and Debin Zhao. Event-based monocular depth estimation with recurrent transformers. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 533 + ], + "type": "text", + "content": "[36] Qian Long, Qiwei Xie, Seiichi Mita, Kazuhisa Ishimaru, and Noriaki Shirai. A real-time dense stereo matching method for critical environment sensing in autonomous driving. In 17th International IEEE Conference on Intelligent Transportation Systems, pages 853-860. IEEE, 2014. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[37] Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Event-intensity stereo: Estimating depth by the best of both worlds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4258-4267, 2021. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 634 + ], + "type": "text", + "content": "[38] Yeongwoo Nam, Mohammad Mostafavi, Kuk-Jin Yoon, and Jonghyun Choi. Stereo depth from events cameras: Concentrate and focus on the future. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6114-6123, 2022. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[39] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5007-5016, 2019. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[40] José Antonio Pérez-Carrasco, Bo Zhao, Carmen Serrano, Begona Acha, Teresa Serrano-Gotarredona, Shouchun Chen," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 128 + ], + "type": "text", + "content": "and Bernabé Linares-Barranco. Mapping from frame-driven to frame-free event-driven vision systems by low-rate rate coding and coincidence processing—application to feedforward convnets. IEEE transactions on pattern analysis and machine intelligence, 35(11):2706-2719, 2013. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 172 + ], + "type": "text", + "content": "[41] Mary Phuong and Christoph H Lampert. Distillation-based training for multi-exit architectures. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1355-1364, 2019. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[42] Ewa Piatkowska, Ahmed Belbachir, and Margrit Gelautz. Asynchronous stereo vision for event-driven dynamic stereo sensor using an adaptive cooperative approach. In Proceedings of the IEEE International Conference on Computer Vision Workshops, pages 45-50, 2013. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 282 + ], + "type": "text", + "content": "[43] Ewa Piatkowska, Jurgen Kogler, Nabil Belbachir, and Margrit Gelautz. Improved cooperative stereo matching for dynamic vision sensors with ground truth evaluation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, pages 53-60, 2017." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 336 + ], + "type": "text", + "content": "[44] Henri Rebecq, Guillermo Gallego, Elias Mueggler, and Davide Scaramuzza. Emvs: Event-based multi-view stereo—3d reconstruction with an event camera in real-time. International Journal of Computer Vision, 126(12):1394-1414, 2018. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 381 + ], + "type": "text", + "content": "[45] Paul Rogister, Ryad Benosman, Sio-Hoi Ieng, Patrick Lichtsteiner, and Tobi Delbruck. Asynchronous event-based binocular stereo matching. IEEE Transactions on Neural Networks and Learning Systems, 23(2):347-353, 2011. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[46] Daniel Scharstein and Richard Szeliski. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47:7-42, 2002. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 469 + ], + "type": "text", + "content": "[47] Waseem Shariff, Mehdi Sefidgar Dilmaghani, Paul Kielty, Mohamed Moustafa, Joe Lemley, and Peter Corcoran. Event cameras in automotive sensing: A review. IEEE Access, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 553, + 514 + ], + "type": "text", + "content": "[48] Pin Tang, Hai-Ming Xu, and Chao Ma. Prototransfer: Cross-modal prototype transfer for point cloud segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3337-3347, 2023. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 516, + 553, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 516, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 316, + 516, + 553, + 569 + ], + "type": "text", + "content": "[49] Stepan Tulyakov, Francois Fleuret, Martin Kiefel, Peter Gehler, and Michael Hirsch. Learning an event sequence embedding for dense event-based deep stereo. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1527-1537, 2019. 2, 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 571, + 553, + 613 + ], + "type": "text", + "content": "[50] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1365-1374, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 658 + ], + "type": "text", + "content": "[51] SM Nadim Uddin, Soikat Hasan Ahmed, and Yong Ju Jung. Unsupervised deep event stereo for depth estimation. IEEE Transactions on Circuits and Systems for Video Technology, 32(11):7489-7504, 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 659, + 553, + 712 + ], + "type": "text", + "content": "[52] Xianqi Wang, Gangwei Xu, Hao Jia, and Xin Yang. Selective-stereo: Adaptive frequency information selection for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19701-19710, 2024. 1" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5446" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[53] Zeyu Wang, Dingwen Li, Chenxu Luo, Cihang Xie, and Xiaodong Yang. Distillbev: Boosting multi-camera 3d object detection with cross-modal knowledge distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8637-8646, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 205 + ], + "type": "text", + "content": "[54] Philippe Weinzaepfel, Thomas Lucas, Vincent Leroy, Yohann Cabon, Vaibhav Arora, Romain Brégier, Gabriela Csurka, Leonid Antsfeld, Boris Chidlovskii, and Jérôme Revaud. Croco v2: Improved cross-view completion pretraining for stereo matching and optical flow. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17969-17980, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 261 + ], + "type": "text", + "content": "[55] Gangwei Xu, Junda Cheng, Peng Guo, and Xin Yang. Attention concatenation volume for accurate and efficient stereo matching. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12981-12990, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "type": "text", + "content": "[56] Gangwei Xu, Xianqi Wang, Xiaohuan Ding, and Xin Yang. Iterative geometry encoding volume for stereo matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21919-21928, 2023." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[57] Gangwei Xu, Yun Wang, Junda Cheng, Jinhui Tang, and Xin Yang. Accurate and efficient stereo matching via attention concatenation volume. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[58] Xu Yan, Jiantao Gao, Chaoda Zheng, Chao Zheng, Ruimao Zhang, Shuguang Cui, and Zhen Li. 2dpass: 2d priors assisted semantic segmentation on lidar point clouds. In European conference on computer vision, pages 677-695. Springer, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[59] Guorun Yang, Xiao Song, Chaoqin Huang, Zhidong Deng, Jianping Shi, and Bolei Zhou. Drivingstereo: A large-scale dataset for stereo matching in autonomous driving scenarios. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 899-908, 2019. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "text", + "content": "[60] Shanxin Yuan, Bjorn Stenger, and Tae-Kyun Kim. 3d hand pose estimation from rgb using privileged learning with depth data. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 576 + ], + "type": "text", + "content": "[61] Heng Zhang, Elisa Fromont, Sébastien Lefèvre, and Bruno Avignon. Low-cost multispectral scene analysis with modality distillation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 803-812, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[62] Kaixuan Zhang, Kaiwei Che, Jianguo Zhang, Jie Cheng, Ziyang Zhang, Qinghai Guo, and Luziwei Leng. Discrete time convolution for fast event-based stereo. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8676-8686, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[63] Linfeng Zhang, Jiebo Song, Anni Gao, Jingwei Chen, Chenglong Bao, and Kaisheng Ma. Be your own teacher: Improve the performance of convolutional neural networks via self distillation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3713-3722, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[64] Bo Zhao, Ruoxi Ding, Shoushun Chen, Bernabe Linares-Barranco, and Huajin Tang. Feedforward categorization on" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 453 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "aer motion events using cortex-like features in a spiking neural network. IEEE transactions on neural networks and learning systems, 26(9):1963-1978, 2014. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[65] Haimei Zhao, Qiming Zhang, Shanshan Zhao, Zhe Chen, Jing Zhang, and Dacheng Tao. Simdistill: Simulated multimodal distillation for bev 3d object detection. In Proceedings of the AAAI conference on artificial intelligence, pages 7460-7468, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 217 + ], + "type": "text", + "content": "[66] Lingjun Zhao, Jingyu Song, and Katherine A Skinner. Crkd: Enhanced camera-radar object detection with cross-modality knowledge distillation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15470-15480, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 251 + ], + "type": "text", + "content": "[67] Runkai Zhao, Heng Wang, and Weidong Cai. Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer. In ACM Multimedia, 2024. 3, 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 295 + ], + "type": "text", + "content": "[68] Alex Zihao Zhu, Yibo Chen, and Kostas Daniilidis. Realtime time synchronized event-based stereo. In Proceedings of the European Conference on Computer Vision, pages 433-447, 2018. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 351 + ], + "type": "text", + "content": "[69] Alex Zihao Zhu, Dinesh Thakur, Tolga Özaslan, Bernd Pfrommer, Vijay Kumar, and Kostas Daniilidis. The multi-vehicle stereo event camera dataset: An event camera dataset for 3d perception. IEEE Robotics and Automation Letters, 3 (3):2032–2039, 2018. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 354, + 555, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 407 + ], + "type": "text", + "content": "[70] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 409, + 553, + 453 + ], + "type": "text", + "content": "[71] Dongqing Zou, Feng Shi, Wei-Heng Liu, Jia Li, Qiang Wang, Paul KJ Park, and Hyunsurk Ryu. Robust dense depth maps generations from sparse dvs stereos. In British Machine Vision Conference, 2017. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5447" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_content_list.json b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..f59b64b6dd50e2d74cac418b5619b957dfeca266 --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_content_list.json @@ -0,0 +1,1797 @@ +[ + { + "type": "text", + "text": "Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions", + "text_level": 1, + "bbox": [ + 142, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dong Li, Chunhui Luo, Yuanfei Bao, Gang Yang, Jie Xiao, Xueyang Fu, Zheng-Jun Zha* \nUniversity of Science and Technology of China, Hefei, China", + "bbox": [ + 148, + 180, + 852, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dongli6@mail.ustc.edu.cn，zhazj@ustc.edu.cn", + "bbox": [ + 303, + 219, + 687, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pansharpening aims to generate high-resolution multispectral (MS) images by fusing panchromatic (PAN) images with corresponding low-resolution MS images. However, many existing methods struggle to fully capture spatial and spectral interactions, limiting their effectiveness. To address this, we propose a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging the compact representation capabilities of quaternions for high-dimensional data. Our method consists of three key components: the quaternion global spectral interaction branch, the quaternion local spatial structure awareness branch, and the quaternion spatial-spectral interaction branch. The first applies the quaternion Fourier transform to convert multi-channel features into the frequency domain as a whole, enabling global information interaction while preserving inter-channel dependencies, which aids spectral fidelity. The second uses a customized spatial quaternion representation, combined with a window-shifting strategy, to maintain local spatial dependencies while promoting spatial interactions, which helps inject spatial details. The last integrates the two pathways within the quaternion framework to enrich spatial-spectral interactions for richer representations. By utilizing quaternion's multi-dimensional representation and parameter-sharing properties, our method achieves a more compact and efficient cross-resolution, multi-band information integration, significantly improving the quality of the fused image. Extensive experiments validate the proposed method's effectiveness and its superior performance over current SOTA techniques. The code is available at https://github.com/dongli8/QuatPanNet.", + "bbox": [ + 89, + 301, + 485, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In remote sensing, satellite sensors face physical limitations that prevent the direct acquisition of high-resolution multispectral (MS) images [45]. Pansharpening addresses", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/022bd077a824460e477d738dedbcbbe71250b2914c9c9ad1f8d7f2d6b16ad737.jpg", + "image_caption": [ + "Figure 1. Comparison between standard real-valued layer and quaternion-valued layer. The real-valued operation typically stacks the input multi-channel features, while the quaternion-valued layer, due to quaternion weight sharing, has the ability to learn the latent relationships within the input features." + ], + "image_footnote": [], + "bbox": [ + 522, + 270, + 901, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "this by merging low-resolution multispectral (LRMS) images with panchromatic (PAN) images to produce high-resolution multispectral (HRMS) images. As a crucial preprocessing step, pansharpening is widely applied across fields [38, 43]. Conceptually, it is a fusion super-resolution task, focused on incorporating fine spatial details while preserving spectral fidelity, requiring strong inter-channel dependency maintenance and the preservation of subtle local spatial structures.", + "bbox": [ + 511, + 577, + 906, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional methods often rely on mathematical models that integrate spatial and spectral data, typically assuming that the pansharpened image is a linear combination of various spectral bands from a high-resolution multispectral image [40]. However, these approaches are limited by their dependence on prior knowledge, which restricts flexibility and adaptability. With the rise of deep learning, convolutional neural networks (CNNs) have been introduced into pansharpening, leading to substantial breakthroughs in performance with increased model complexity [1, 5, 9, 23].", + "bbox": [ + 511, + 715, + 908, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite notable successes, existing methods still struggle to capture the intricate spatial-spectral interactions essential", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10908", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "for high-quality pansharpening. These limitations hinder simultaneous feature interaction and inter-feature dependency preservation, leading to compromises in both spectral fidelity and spatial detail. Unlike conventional real-valued networks, quaternion networks treat multidimensional features as unified entities, allowing them to effectively encode internal dependencies [26, 27]. Additionally, quaternion networks demonstrate strong modeling capabilities for both inter- and intra-correlations within the quaternion latent space [3]. Thus, we aim to harness quaternion operations to further elevate pansharpening performance.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our Motivation. PAN images provide high-resolution spatial details, while MS images offer rich multispectral information [11, 24]. A successful pansharpening model must preserve both spatial details and spectral fidelity during fusion. This requires capturing the unique features of PAN and MS images while leveraging their complementary information to produce accurate and comprehensive fused outputs. Traditional approaches often handle multi-channel inputs by stacking and mapping them to a single output channel, which causes information loss. Consequently, they struggle to capture inter-channel correlations and interactions between spatial and spectral data effectively. To compensate, some methods increase the number of output channels to improve representational capacity [39], but this also increases model complexity. Furthermore, these methods typically extract features in real-number space, limiting their ability to capture directional information such as edges and structures. In contrast, quaternions represent multiple spectral channels as distinct components of a compact quaternion representation, as shown in Figure 1. This approach integrates multi-band information efficiently while reducing information loss through parameter-sharing quaternion operations. These features make quaternions particularly well-suited for pansharpening tasks that involve multi-source data inputs.", + "bbox": [ + 91, + 261, + 483, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Based on the analysis above, we propose a novel quaternion-based pansharpening framework called QuatPanNet. QuatPanNet is designed to balance spectral fidelity and spatial detail richness by leveraging quaternion-based representations and operations to enhance interactions among multispectral, spatial structure, and spatial-spectral features. The framework consists of three core components, i.e., Quaternion Global Spectral Interaction Branch,Quaternion Local Spatial Structure Awareness Branch, andQuaternion Spatial-Spectral Interaction Branch. The first one uses the Quaternion Fourier Transform (QFT) to process multi-channel features holistically in the frequency domain. By enabling global interactions while preserving inter-channel dependencies, it effectively extracts global spectral representations, which are critical for maintaining spectral fidelity. The second one employs a customized spatial quaternion transformation method and a window", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "shifting strategy, this branch focuses on preserving local structural dependencies and enhancing spatial interactions. It ensures that spatial details are well-integrated while maintaining the structural integrity of the input data. The last one, which built on the outputs of the first two branches, integrates dual-path features into a unified quaternion representation. This integration enhances the complementarity and coherence of information during the spatial-spectral fusion process. By utilizing the compact and efficient representation of PAN and MS image data in quaternion space, QuatPanNet facilitates effective interactions among spatial and spectral features. This approach reduces parameter complexity while achieving high spectral fidelity and rich spatial resolution, leading to superior fusion quality. In summary, the contributions of this work are as follows:", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel quaternion-based pansharpening framework, QuatPanNet, designed to comprehensively harness quaternions for this task.", + "- We design customized quaternion-based spatial and spectral interaction modules for pansharpening. The former facilitates global information interaction while preserving inter-spectral dependencies, and the latter promotes spatial interaction while maintaining local structural dependencies.", + "- We introduce a quaternion-based spatial-spectral interaction branch, leveraging the high-dimensional representational power of quaternions to improve the fusion of spectral and spatial features." + ], + "bbox": [ + 513, + 321, + 903, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on multiple satellite datasets demonstrate that our method achieves superior qualitative and quantitative results with fewer parameters and effectively generalizes to full-resolution scenes.", + "bbox": [ + 511, + 520, + 903, + 580 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 595, + 661, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pansharpening", + "text_level": 1, + "bbox": [ + 511, + 621, + 665, + 638 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing pansharpening methods are divided into traditional ones and deep - learning - based ones. Traditional ones mainly consist of Component Substitution (CS) [31, 34], Multi - Resolution Analysis (MRA) [25, 29, 35], and Variational Optimization (VO) [8, 15, 21]. However, their heavy dependence on manually designed features and prior knowledge restricts their applicability [37]. PNN [23] is the first to introduce CNN into the field of pansharpening, achieving better results than traditional methods. Since then, deep learning-based approaches have dominated this field. MSDCNN [42] combines multi-scale feature extraction and residual learning, enhancing the quality of the fused images. INNformer [47], based on the Transformer architecture, integrates an invertible neural network module for lossless information to model long-range dependencies and efficient feature fusion, successfully introducing the Transformer architecture into this domain. The fusion of frequency do", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10909", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "main features in images is also a key focus for researchers [36, 48]. For instance, FAME-Net [16] proposes a method that combines MOE with frequency domain information, enabling the network to dynamically learn high-frequency information in remote sensing images. HFIN [33] explores the relationship between PAN and LRMS images using local Fourier information integration. In addition to designing pansharpening models, Zhu et al. developed the first upsampling module specifically for pansharpening methods, called PGCU [52].", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2.Quaternion", + "text_level": 1, + "bbox": [ + 89, + 251, + 217, + 266 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quaternions are the natural extension of complex numbers to four dimensions, proposed by Hamilton in 1843 [13]. A quaternion consists of four components, including one real part and three imaginary parts. Quaternions, with their four components, are particularly well-suited for three-dimensional and four-dimensional feature vectors, as seen in applications such as image processing and robotic kinematics [2, 30]. In 2001, Pei et al. [28] introduced the quaternion model for color images, marking the beginning of quaternion applications in the field of image processing. The advantages of quaternion wavelet transforms [18], quaternion principal component analysis [44], and other quaternion-based color image processing techniques [41] have been demonstrated to extract more representative features for color images. These techniques have achieved promising results in advanced visual tasks such as color image classification. With the wave of deep learning, many researchers have begun to explore the application of quaternions in the field of deep learning [10, 26]. Xu et al. [51] proposes QCNN, which combines quaternions with CNNs, achieving faster convergence of the loss function while preserving color information more effectively than real-valued CNNs. Kusamichi et al. [19] studies color night vision and finds that using quaternion networks yields better results in extracting color information from dim images.", + "bbox": [ + 89, + 273, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 664, + 181, + 679 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce the relevant properties of quaternions and their operators, and then present the proposed quaternion pan-sharpening network, as shown in Figures 2 and 3. We then provide a detailed explanation of the fundamental components of our method, including three quaternion-based branches: a) the global interaction branch utilizes the Quaternion Fourier Transform and quaternion convolution to maintain channel interdependencies while representing global information; b) the local structure-aware branch models local spatial dependencies through a specially designed spatial quaternion transformation method combined with a window-shifting strategy; and c) the dual-domain interaction branch leverages the orthogonal relationships inherent in quaternions to model", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "dual-domain correlations and promote complementary feature learning. Finally, we present the optimization strategy employed in our approach.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.Quaternion Operators", + "text_level": 1, + "bbox": [ + 511, + 143, + 722, + 161 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quaternion, proposed by Hamilton in 1843, extends complex numbers to a four-dimensional space. A quaternion $Q \\in \\mathbb{H}$ can be expressed as follows [32]:", + "bbox": [ + 511, + 166, + 905, + 212 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nQ = w 1 + x i + y j + z k, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 222, + 906, + 238 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w, x, y, z$ are all real numbers, and $1, i, j, k$ are the quaternion unit basis. $w1$ refers to the real component of the number, and $xi + yj + zk$ to the imaginary component. In the definition of quaternions, they must satisfy the following identities, known as the Hamilton rules:", + "bbox": [ + 511, + 247, + 905, + 321 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ni ^ {2} = j ^ {2} = k ^ {2} = i j k = - 1. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 330, + 906, + 348 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose the use of quaternions and their associated operators to uncover both intra- and inter-relations between the features of high spatial resolution PAN and MS images. This approach is inspired by the unique feature-handling capabilities inherent in quaternion representation [3]. Consequently, we introduce two quaternion operators: quaternion convolution and the quaternion Fourier transform.", + "bbox": [ + 511, + 357, + 905, + 462 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quaternion Convolution", + "text_level": 1, + "bbox": [ + 532, + 464, + 709, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the real-valued space, convolution involves convolving a filter matrix with a vector. Extending this operation to quaternions, quaternion convolution is achieved by convolving a quaternion filter matrix with a quaternion vector. To maintain consistency with the defined quaternions, we represent the computation of quaternion convolution using real-valued matrices. This computation is performed by convolving the quaternion filter matrix $W = A + Bi + Cj + Dk$ with the quaternion vector $Q = w + xi + yj + zk$ , where $A, B, C$ , and $D$ are real-valued matrices. The convolution operation between $W$ and $Q$ is denoted as:", + "bbox": [ + 511, + 478, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} W * Q = (A * w - B * x - C * y - D * z) + \\\\ i (A * x + B * w + C * z - D * y) + \\\\ j (A * y - B * z + C * w + D * x) + \\tag {3} \\\\ K (A * z + B * y - C * x + D * w). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 652, + 903, + 726 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It can be seen that the result of quaternion convolution produces a unique linear combination of each coordinate axis for every single-axis result. This stems from the structure of quaternion multiplication and enforces interactions between each axis of the kernel and each axis of the image. This contrasts with real-valued convolution, which simply multiplies each channel of the kernel with the corresponding channel of the image. The advantage of this approach is that quaternion convolution promotes cross-axis interactions and information extraction through the reuse of filters across each axis and combination [27].", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10910", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4360aa7e7cba208c962cf624fc35b89f8447ca4e5995689294eef91955104070.jpg", + "image_caption": [ + "Figure 2. The framework of our proposed pan-sharpening network QuatPanNet. It consists of several Quaternion Spatial-Spectral Interaction Blocks (QSSIB). The Quaternion Spatial-Spectral Interaction Block is shown in Figure 3." + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 906, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Quaternion Fourier Transform", + "text_level": 1, + "bbox": [ + 109, + 401, + 331, + 415 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Fourier Transform plays a crucial role in various image processing tasks, including pansharpening [50]. After decomposing the input image into multiple channels, the Discrete Fourier Transform (DFT) can be applied to each channel individually. However, this approach neglects the mutual information among different channels, which can hinder the performance of pansharpening. The Discrete Quaternion Fourier Transform (DQFT) offers the possibility of simultaneously transforming multi-channel images, as it can transform quaternions holistically [4].", + "bbox": [ + 88, + 416, + 482, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Due to the non-commutative nature of quaternion multiplication, there are various types of Quaternion Fourier Transforms (QFT), such as left-side, right-side and two-side QFT [6]. In this paper, we utilize a typical QFT, specifically the left-sided QFT. For an image $\\dot{f}$ of size $H\\times W$ , the Discrete Quaternion Fourier Transform is defined as follows.", + "bbox": [ + 89, + 566, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} (\\dot {f}) (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} e ^ {- \\dot {\\mu} 2 \\pi \\left(\\frac {x}{H} u + \\frac {y}{H} v\\right)} \\dot {f} (x, y), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 678, + 480, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\dot{f} (x,y)$ is pixel of $\\dot{f}$ and $\\dot{\\mu}$ is unit pure quaternion. $(u,v)$ is the coordinate in the quaternion frequency domain. $\\mathcal{F}^{-1}(\\dot{f})$ defines the inverse quaternion Fourier transform accordingly. Inspired by the traditional Fourier transform, the modulus of $\\mathcal{F}(\\dot{f})(u,v)$ is taken as the amplitude component, and the angles with the three imaginary axes are taken as the phase components.", + "bbox": [ + 89, + 733, + 482, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is worth noting that the QFT is not a simple substitution of the imaginary unit in the two-dimensional Fourier Transform with quaternions; rather, its basis functions possess orthogonal imaginary components, which preserve the", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "information and mutual relationships among different components in the frequency domain. Consequently, some studies leverage QFT to treat multiple features as a unified whole to enhance performance [4, 46]. Furthermore, the efficiency of QFT is also noteworthy. For example, when comparing the use of QFT with three separate conventional Fourier Transforms for processing color images, the number of real multiplications and additions required by QFT is approximately $75\\%$ of that needed for the three separate transforms [30]. For pansharpening, we propose utilizing the QFT to treat multispectral data as a unified whole, facilitating global information interactions in the frequency domain while preserving inter-channel dependencies.", + "bbox": [ + 511, + 401, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Framework", + "text_level": 1, + "bbox": [ + 513, + 617, + 640, + 633 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the above analysis, we propose a novel pansharpening approach based on quaternion information interaction, as illustrated in Figure 2. Given a panchromatic (PAN) image $P \\in \\mathbb{R}^{H \\times W \\times 1}$ and a low-resolution multispectral (LRMS) image $L \\in \\mathbb{R}^{H / r \\times W / r \\times C}$ , the network first uses convolution to project the bicubic upsampled r-times $L$ into a shallow feature representation. Simultaneously, $P$ is processed through multiple cascaded convolution layers to extract effective information for high-resolution multispectral (HRMS) reconstruction. Then, the features of the PAN and LRMS images are jointly processed through the core building module, the quaternion spatial-spectral interaction block (QSSIB), for continuous global information interaction and local structure-aware integration. Next, a convolution layer transforms all features gathered by the N QSSIBs back into the image space, and finally, these features are combined with the residual to obtain the final output.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10911", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5cbe67bf3e36207152cfbadca3c2bf3991aa580420568ebb7b7d872a29ba923a.jpg", + "image_caption": [ + "Figure 3. The detailed flowchart of the proposed core building module quaternion spatial-spectral interaction block (QSSIB), consisting of three components: quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch (QLSA) and quaternion spatial-spectral interaction branch (QSSI)." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 908, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Key Building Components", + "text_level": 1, + "bbox": [ + 89, + 421, + 331, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 3, the fundamental building block of our method comprises three key elements: the quaternion global spectral interaction branch (QGSI), the quaternion local spatial structure awareness branch (QLSA), and the quaternion spatial-spectral interaction branch (QSSI). To convert images into quaternion representations, we design customized methods, including channel quaternion representation (CQR), spatial quaternion representation (SQR), and shifted window quaternion representation (SWQR). When transforming quaternion representations back into image features, we denote the corresponding inverse operations as ICQR, ISQR, and ISWQR, respectively. (For more details on quaternion representation, refer to the supplementary material.) These three key components, in conjunction with the quaternion representations of the images, jointly facilitate channel interactions during multispectral fusion and local interactions for spatial resolution enhancement.", + "bbox": [ + 88, + 444, + 485, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Quaternion Global Spectral Interaction", + "text_level": 1, + "bbox": [ + 109, + 704, + 392, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The global interaction branch utilizes the Quaternion Fourier Transform (QFT) to transform multi-channel features as a whole into the frequency domain, facilitating global information interactions. Since the input to QFT must be in quaternion form and we aim to maintain interchannel interdependencies, we design a channel quaternion representation (CQR). This approach divides the image features equally along the channel dimension into four parts, which are then used as the four components of a quaternion. Let the features of the MS and PAN images be denoted as $F_{ms}$ and $F_{p}$ , respectively; their channel-dimensional quaternion representations are denoted as $Q_{ms}^{c}$ and $Q_{p}^{c}$ . We", + "bbox": [ + 89, + 719, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "then use QFT to transform the MS and PAN features into the frequency domain. Some studies on image restoration indicate that decomposing amplitude and phase components facilitates the decoupling of degradation [17, 50]. Therefore, we decompose the frequency domain features into amplitude and phase components, with the corresponding process represented as", + "bbox": [ + 511, + 421, + 906, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {p} ^ {c}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 532, + 906, + 551 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {m s} ^ {c}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 554, + 903, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{A}(\\cdot)$ represents the amplitude, and $\\mathcal{P}_1(\\cdot),\\mathcal{P}_2(\\cdot)$ and $\\mathcal{P}_3(\\cdot)$ represent the phases corresponding to the three imaginary axes. We then concatenate the amplitude and phase spectra of the PAN and MS images accordingly to form new quaternions. Next, we process the newly formed quaternions using quaternion convolution, thereby enhancing global representations in the frequency domain while promoting interactions between channels.", + "bbox": [ + 511, + 575, + 906, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q _ {f} ^ {c} = \\mathcal {O Q} (\\mathcal {C} [ \\mathcal {A} (Q _ {p} ^ {c}), \\mathcal {A} (Q _ {m s} ^ {c}) ], \\mathcal {C} [ \\mathcal {P} _ {1} (Q _ {p} ^ {c}), \\mathcal {P} _ {1} (Q _ {m s} ^ {c}) ] \\\\ , \\mathcal {C} \\left[ \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right) \\right], \\mathcal {C} \\left[ \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) \\right]) \\tag {7} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 700, + 903, + 755 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{C}$ denotes concatenation along the channel dimension, $\\mathcal{OQ}$ is quaternion convolution. The fused amplitude and phase components are the transformed back to the spatial domain using the inverse discrete quaternion Fourier transform.", + "bbox": [ + 511, + 760, + 906, + 834 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {f r e} ^ {c} = \\mathcal {F} ^ {- 1} \\left(Q _ {f} ^ {c}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 642, + 835, + 906, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By representing the image in a channel-dimensional quaternion form, the quaternion operator treats the multichannel data as a unified whole for information interaction.", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10912", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, according to spectral convolution theorem [7], processing information in the Fourier space enables capturing global frequency representations. In summary, this branch generates the global information representation with strengthened channel interactions $Q_{fre}^{c}$ .", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quaternion Local Spatial Structure Awareness", + "text_level": 1, + "bbox": [ + 109, + 167, + 441, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Pansharpening is essentially a fusion-based superresolution task, making the injection of spatial resolution critically important. Enhancing spatial resolution requires not only global information from the frequency domain but also local information interactions in the spatial domain. For pansharpening, preserving local structures during information interactions is particularly significant, as it enhances the network's ability to recover fine details.", + "bbox": [ + 89, + 184, + 482, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We aim to leverage the balance of enhanced information interaction and specific information preservation offered by quaternions to improve the network's spatial superresolution capabilities. To this end, we design the spatial quaternion representation (SQR). It divides the image into several $2 \\times 2$ windows in the spatial dimension, with each local window's four pixels serving as the four components of a quaternion, as illustrated in Figure 3. This approach explicitly ensures dependencies among local pixels, preserving the local structure. Based on this representation, we use quaternion convolutions (QCNN) to process the features.", + "bbox": [ + 89, + 305, + 482, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, this representation method lacks crosswindow connections, which limits spatial interactions. Inspired by [22], we design the shifted window quaternion representation (SWQR) based on SQR. This allows the network to alternately employ two partitioning configurations, thereby enhancing the effectiveness of spatial quaternion representation. Through the combined effect of this method and quaternion convolution, this branch preserves local spatial dependencies while promoting spatial interactions, thereby obtaining the local representation $Q_{spa}^{s}$ .", + "bbox": [ + 89, + 472, + 482, + 625 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The information representations of the two branches are highly complementary, with one branch responsible for global, frequency-domain, and channel interactions, and the other focused on local, spatial-domain, and spatial interactions. Therefore, further interaction and integration between them facilitate mutual compensation.", + "bbox": [ + 89, + 625, + 482, + 714 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quaternion Spatial-Spectral Interaction", + "text_level": 1, + "bbox": [ + 109, + 715, + 393, + 729 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The two branches described above respectively obtain quaternion representations in the channel dimension, $Q_{fre}^{c} \\in \\mathbb{H}^{H \\times W \\times C / 4}$ , and in the spatial dimension, $Q_{spa}^{s} \\in \\mathbb{H}^{H / 2 \\times W / 2 \\times C}$ . The quaternion dual-domain information interaction branch mainly consists of two parts: one is a quaternion-based global fusion, and the other is the channel-aware quaternion spatial feature injection.", + "bbox": [ + 89, + 731, + 482, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quaternion-based global fusion merges the features obtained from the two branches into a single quaternion for unified processing, thereby enhancing the correlation between the two through the structure of the quaternion. First,", + "bbox": [ + 89, + 840, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "both $Q_{fre}^{c}$ and $Q_{spa}^{s}$ are converted back into image feature representations. $Q_{fre}^{c}$ is combined along the channel dimension to obtain $F_{fre} \\in \\mathbb{R}^{H \\times W \\times C}$ , and $Q_{spa}^{s}$ is rearranged spatially to obtain $F_{spa} \\in \\mathbb{R}^{H \\times W \\times C}$ . To achieve a richer feature representation, we apply convolutions to $F_{fre}$ and $F_{spa}$ to obtain $F_{fre}'$ and $F_{spa}'$ , respectively. We then combine these four features into a quaternion $Q_{f} = F_{fre} + F_{spa}i + F_{fre}^{\\prime}j + F_{spa}^{\\prime}k$ . We use quaternion convolution to process $Q_{f} \\in \\mathbb{H}^{H \\times W \\times C}$ as a whole, producing $Q_{a} \\in \\mathbb{H}^{H \\times W \\times C/4}$ .", + "bbox": [ + 511, + 90, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Channel-aware quaternion spatial feature injection involves injecting spatial features $F_{spa}$ from the local structure-aware branch into the four components of the quaternion representation $Q_{fre}^{c}$ in the channel dimension, thereby facilitating information fusion in both spatial and channel domains. Specifically, we replicate the image feature $F_{spa}$ four times and concatenate each copy with one of the four components of the quaternion feature $Q_{fre}^{c}$ to obtain $Q_{i} \\in \\mathbb{H}^{H \\times W \\times 5C / 4}$ :", + "bbox": [ + 511, + 248, + 903, + 386 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {i} = \\mathcal {C} \\left[ Q _ {f r e} ^ {c}, \\left(F _ {s p a} + F _ {s p a} i + F _ {s p a} j + F _ {s p a} k\\right) \\right]. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 392, + 903, + 411 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then feed $Q_{i}$ into quaternion convolution to obtain $Q_{b} \\in \\mathbb{H}^{H \\times W \\times C / 4}$ . Afterward, we concatenate $Q_{a}$ and $Q_{b}$ , and apply quaternion convolution for further fusion to obtain $Q_{out} \\in \\mathbb{H}^{H \\times W \\times C / 4}$ .", + "bbox": [ + 511, + 415, + 903, + 476 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {o u t} = \\operatorname {Q C o n v} \\left(\\mathcal {C} \\left[ Q _ {a}, Q _ {b} \\right]\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 482, + 903, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where QConv denotes quaternion convolution. We then convert $Q_{out}$ into the image feature representation $F_{out} \\in \\mathbb{R}^{H \\times W \\times C}$ . Finally, a residual learning mechanism is applied, adding the input feature $F_{ms}$ to the fused feature. The proposed network achieves effective fusion through the seamless integration of the two quaternion representations.", + "bbox": [ + 511, + 506, + 903, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 511, + 604, + 653, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Let $Y$ and $GT$ denote the network's output and the corresponding ground truth, respectively. We use a joint spatial-frequency domain loss to supervise the network training. In the frequency domain, we first apply the quaternion Fourier transform to convert $Y$ and $GT$ into the quaternion Fourier space. Then, we compute the L1 norm of the real and imaginary differences between $Y$ and $GT$ , and sum them to obtain the quaternion Fourier domain loss.", + "bbox": [ + 511, + 626, + 903, + 746 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {f r e} = \\| R (Y) - R (G T) \\| _ {1} + \\| I (Y) - I (G T) \\| _ {1} \\tag {11} \\\\ + \\| J (Y) - J (G T) \\| _ {1} + \\| K (Y) - K (G T) \\| _ {1}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 752, + 903, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $R(\\cdot), I(\\cdot), J(\\cdot)$ , and $K(\\cdot)$ represent the real part and three imaginary parts of the quaternion respectively. In the spatial domain, we employ the L1 loss. Finally, the overall loss function is as follows", + "bbox": [ + 511, + 795, + 903, + 854 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\| Y - G T \\| _ {1} + \\lambda \\mathcal {L} _ {f r e}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 862, + 903, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda$ is the weighting factor and set to 0.1 empirically.", + "bbox": [ + 511, + 885, + 893, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10913", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0d58f5574228d307d2321d3d5d8c388c6620057a5ecbbc892c7ebc15abcf844c.jpg", + "image_caption": [ + "LRMS" + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 212, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/38072c22dc00240d02d7e6ed65f037633e9d00ce7651e752f71d89ab1acad415.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 230, + 88, + 352, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ffef7e3696608666d2933be36000362a1f6e0f6d9de01a923f225dc9f11d2cbb.jpg", + "image_caption": [ + "GS" + ], + "image_footnote": [], + "bbox": [ + 367, + 88, + 490, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87ab71925aaaba337c145cbcb0cd9044d4ff2687e234c8d0a4d11107d43b92fc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 627, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5065df451e37e817546e64992c3cca13443bedf43a23c3d79998c594042cfa73.jpg", + "image_caption": [ + "MSDCNN" + ], + "image_footnote": [], + "bbox": [ + 645, + 88, + 767, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99bc74a635a76c945eea1533db2d5ee4d6b23d610e6c2253dab76896a0d81de4.jpg", + "image_caption": [ + "GPPNN" + ], + "image_footnote": [], + "bbox": [ + 784, + 88, + 905, + 181 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f4cce0d1da5c8e8b4cd038352abc5bbb4dd90c14de8800013673f3a5d31a8000.jpg", + "image_caption": [ + "MutNet" + ], + "image_footnote": [], + "bbox": [ + 91, + 196, + 212, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a9d0c937fd2566ead0393ffd7b16b92077b50abf607409f9c1681efc351c03d5.jpg", + "image_caption": [ + "PAN", + "INNformer" + ], + "image_footnote": [], + "bbox": [ + 230, + 196, + 352, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1919873bdbc6fd9d15f08e2e747dbc178aadc3b91e380fe5b7c656d0e4d293f7.jpg", + "image_caption": [ + "HFIN" + ], + "image_footnote": [], + "bbox": [ + 367, + 196, + 490, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c8130270ab584a8b7065bfdc9df831381e9876bad7a81e34f5c9bba3822a9c0b.jpg", + "image_caption": [ + "PNN", + "FAME-Net" + ], + "image_footnote": [], + "bbox": [ + 506, + 196, + 627, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f8b8c85f3af244ebc58911bfa5dc2e7da3b3b25155bae08f62b295c0d8814d5a.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 645, + 196, + 767, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b9059ccf5e428aa1f113a6c11bf7e7bd144a5f8fa9c3a879f0f14acb7af998e6.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 784, + 196, + 905, + 291 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ca471fb721fb0893d11f23de893307781a0fbb459e18c1b16884346cc7ad1384.jpg", + "image_caption": [ + "GS", + "Figure 4. The visual comparisons between other pan-sharpening methods and our method on WorldView-II satellite" + ], + "image_footnote": [], + "bbox": [ + 91, + 309, + 176, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ad31d5f2834d3ce6e8ac79aaaae08a8e04c991bde3c783f9875b8ec8b543bcd0.jpg", + "image_caption": [ + "PNN" + ], + "image_footnote": [], + "bbox": [ + 181, + 310, + 264, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/13399ce9c97b7e0b3fdc734b48adfd4f3f58d92e79e0226e605edd676d45a5d6.jpg", + "image_caption": [ + "MSDCNN" + ], + "image_footnote": [], + "bbox": [ + 269, + 310, + 352, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9f34b258e8df8c4e0ecb0cbc917a61f783996d8baac216bc9908eb89bbafb680.jpg", + "image_caption": [ + "GPPNN" + ], + "image_footnote": [], + "bbox": [ + 357, + 310, + 439, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/463b8bb99ce9834c6f14903c20eacbabbc5ab291dda3178919fbecd78d0eaa50.jpg", + "image_caption": [ + "MutNet" + ], + "image_footnote": [], + "bbox": [ + 444, + 310, + 527, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c09ebd78fae13c0505879aee7ed21b324bd273d2efbb21a84c992d4a2537b410.jpg", + "image_caption": [ + "INNformer" + ], + "image_footnote": [], + "bbox": [ + 532, + 310, + 616, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e79a805be7f28a2d1cce5c85af39f15ab3f14ea6854a027cc99286d8283968af.jpg", + "image_caption": [ + "HFIN" + ], + "image_footnote": [], + "bbox": [ + 620, + 310, + 704, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9310c22fa58de36945cb88cde5aa81f9c634aeb7803e8cd8ea8841e6303ad23b.jpg", + "image_caption": [ + "FAME-Net" + ], + "image_footnote": [], + "bbox": [ + 710, + 310, + 792, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0610e756f4fc676d5c1e1c351a1e8801b131cf8cd72021c3f28973e6435a2457.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 799, + 310, + 880, + 367 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5526a513eba49cfd46396bac07528bedb800c2e7411d2d5547b33a5ad4d030f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 882, + 305, + 901, + 385 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ed6bc0c71bc31dee74ec77590c850d24b53eb30a250961e460bc24fcba9d40e9.jpg", + "table_caption": [ + "Table 1. Experimental results of all the competing methods on the three benchmark datasets. The best and the second best values are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
MethodsParams (M)WorldView IIWorldView IIIGaoFen2
PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓
Brovey-35.86460.92160.04031.823822.50600.54660.11598.233137.79740.90260.02181.3720
IHS-35.29620.90270.04612.027822.55790.53540.12668.361638.17540.91000.02431.5336
GS-35.63760.91760.04231.877422.56080.54700.12178.243337.22600.90340.03091.6736
PNN0.068940.75500.96240.02591.064629.94180.91210.08243.320643.12080.97040.01720.8528
MSDCNN0.239041.33550.96640.02420.99430.30380.91840.07823.188445.68740.98270.01350.6389
GPPNN0.119841.16220.96840.02441.031530.17850.91750.07763.259344.21450.98150.01370.7361
MutNet0.071441.67730.97050.02240.951930.49070.92230.07493.112547.30420.98920.01020.5481
INNformer0.070641.69030.97040.02270.951430.53650.92250.07473.099747.35280.98930.01020.5479
HFIN0.077242.23190.97140.02150.880730.61470.92030.07423.078648.87830.98980.00930.4591
FAME-Net0.140842.02620.97230.02150.917230.99030.92870.06972.953147.67210.98980.00980.5242
Ours0.038742.48460.97380.02080.855431.17660.93320.05212.576149.14230.99010.00900.4474
", + "bbox": [ + 91, + 465, + 906, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 672, + 223, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.1. Dataset and benchmarks", + "text_level": 1, + "bbox": [ + 89, + 696, + 316, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct our experiments on three classic datasets. WorldView-II (WV2), Gaofen2 (GF2) and WorldView-III (WV3). The WorldView-II dataset consists of 760 image pairs for training and 80 image pairs for testing. The WorldView-III dataset includes 2150 image pairs for training and 200 image pairs for testing. The GaoFen-2 dataset comprises 2712 image pairs for training and 200 image pairs for testing. To validate the effectiveness of our approach, we compare it with several state-of-the-art pansharpening methods, including PNN [23], MSDCNN [42], GPPNN [39], MutNet [49], INNformer [47], HFIN [33] and FAME-Net [16], as well as traditional pansharpening meth", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods such as GS [20], Brovey [12], and IHS [14].", + "bbox": [ + 513, + 672, + 833, + 688 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our experiments, all deep learning models are implemented using PyTorch and trained on an NVIDIA GeForce GTX 3090 GPU. For each set, the multispectral (MS) images are cropped into patches with the size of $32 \\times 32$ , and the corresponding panchromatic (PAN) images are of size $128 \\times 128$ . We employ common evaluation metrics, including PSNR, SSIM, SAM, and ERGAS. Additionally, we utilize three widely-used no-reference IQA metrics for real-world full-resolution scenes: $D_{\\lambda}, D_{S}$ and QNR. More descriptions of the implementation details and dataset are provided in the supplementary material.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10914", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ccf3eeed65c27bc98634f20f00894c79462a82ff6ad8bf60c81437e988b106fa.jpg", + "table_caption": [ + "Table 2. Non-reference metrics on full-resolution dataset." + ], + "table_footnote": [], + "table_body": "
PNNMSDCNNGPPNNINNformerMutNetHFINFAME-NetOurs
Dλ↓0.07460.07340.07820.06970.06940.07100.06740.0566
Ds↓0.11640.11510.12530.11280.11180.10980.11210.1055
QNRT0.81910.82150.80730.82530.82470.82610.82910.8447
", + "bbox": [ + 91, + 114, + 480, + 154 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2. Comparison with state-of-the-art methods", + "text_level": 1, + "bbox": [ + 89, + 179, + 450, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on reduced-resolution scene. To evaluate the effectiveness of our method, we conducted experiments on three datasets and compared it with ten other state-of-the-art pansharpening methods. Table 1 summarizes the evaluation metrics for the three datasets, where the best result for each metric is highlighted in red and the second-best result in blue. The results demonstrate that our method significantly outperforms the others across all evaluation metrics, achieving the best performance on all three datasets. Compared to the second-best method in terms of PSNR for each dataset, our method shows an improvement of $0.25\\mathrm{dB}$ on the WV2 dataset, $0.18\\mathrm{dB}$ on the GF2 dataset, and $0.27\\mathrm{dB}$ on the WV3 dataset, indicating the consistency of our method with the true high-resolution images. We also conducted qualitative experiments, Figures 4 present samples from the WV2 datasets. The top two rows show the comparison of experimental results, while the last row displays the mean squared error maps of each network's results compared to the ground truth. It is evident that the results obtained by our network are closest to the ground truth, with the smallest discrepancies.", + "bbox": [ + 91, + 200, + 483, + 517 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Evaluation on full-resolution scene. We also conduct full-resolution analysis in real-world scenarios to validate the generalization capability of our method. We perform experiments on an additional 200 sets of the GF2 dataset. Due to the absence of high-resolution multispectral images for actual real-world scenarios, we employ three commonly used no-reference metrics, $D_{\\lambda}$ , $D_{S}$ , and QNR for evaluation. The experimental results, as shown in Table 2, demonstrate that our method achieves the best performance across all three metrics.", + "bbox": [ + 89, + 518, + 483, + 670 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/9f235abea61f798aa9dac7479be34653e46a3344647f40c4c7e0acbc7adb6cd5.jpg", + "table_caption": [ + "Table 3. The results of different configurations on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
ConfigurationQGSIQLSAQSSIPSNR ↑SSIM ↑SAM ↓ERGAS ↓
IX✓✓41.73560.96730.02470.9337
II✓X✓42.02170.97060.02190.9134
III✓✓X42.17340.97130.02210.9012
Ours✓✓✓42.48460.97380.02080.8554
", + "bbox": [ + 102, + 729, + 470, + 787 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Experiments", + "text_level": 1, + "bbox": [ + 89, + 816, + 297, + 834 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We perform an ablation study using the WorldView-II dataset to further demonstrate the effectiveness of our method. The quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c8151d29cd07d8f843456ce1fd88d494c780da07f63c9d5c3c5708523fb9462.jpg", + "table_caption": [ + "Table 4. PSNR values of Our model with different number of blocks on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
Stages (K)PSNR ↑SSIM ↑SAM ↓ERGAS ↓
141.83780.96030.02490.9879
241.90020.96470.02380.9725
342.09130.96980.02340.9578
442.23690.97240.02150.9007
542.48460.97380.02080.8554
", + "bbox": [ + 516, + 142, + 903, + 247 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(QLSA), and quaternion spatial-spectral interaction branch (QSSI) are the core components of the proposed method, and we conduct separate ablations for each. Additionally, we also explore the impact of the number of basic building blocks used in the network through ablation studies.", + "bbox": [ + 511, + 272, + 903, + 347 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of the three key quaternion components, QGSI, QLSA, and QSSI, we replace their quaternion operations with real-valued operations, as shown in Table 3. The results demonstrate that removing each quaternion component leads to a decline in performance, thus confirming the effectiveness of the quaternion's multidimensional representation and parameter-sharing mechanism for pansharpening.", + "bbox": [ + 511, + 347, + 905, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct experiments on networks with varying numbers of core building blocks to evaluate the impact of the number of core modules, as shown in Table 4. As observed, the model performance improves with the increase in the number of Quaternion Spatial-Spectral Interaction Blocks (QSSIB), though this comes at the cost of increased resource consumption. We balance performance and resource consumption by using the default setting of 5. More ablation experiments are shown in the supplementary materials.", + "bbox": [ + 511, + 469, + 905, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 618, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose QuatPanNet, a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging quaternions' compact representation for high-dimensional data. Our work demonstrates the advantages of a unified quaternion framework for effectively addressing the pansharpening problem. We custom-design quaternion-based spectral and spatial interaction modules: the former promotes global information interaction while preserving spectral dependencies, and the latter fosters spatial interaction while maintaining local structural dependencies. Building on these, we integrate the dual-path features into a unified quaternion representation. By utilizing the compact and efficient representation of PAN and MS image data in the quaternion space, QuatPanNet facilitates effective interactions between spatial and spectral features. Our method achieves exceptional fusion quality with a reduced number of parameters.", + "bbox": [ + 511, + 643, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10915", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243.", + "bbox": [ + 89, + 114, + 485, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Paolo Addesso, Gemine Vivone, Rocco Restaino, and Jocelyn Chanussot. A data-driven model-based regression applied to panchromatic sharpening. IEEE Transactions on Image Processing, 29:7779-7794, 2020. 1", + "[2] Nicholas A Asparagus and John K Dimitros. A comparative study of three methods for robot kinematics. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(2):135-145, 1998. 3", + "[3] Qinglong Cao, Zhengqin Xu, Yuntian Chen, Chao Ma, and Xiaokang Yang. Domain prompt learning with quaternion networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26637-26646, 2024. 2, 3", + "[4] Ping Cui, Cui-Ming Zou, Bang-Ju Wang, and Feng Shi. Color image fusion in quaternion spectral domain. In 2023 International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR), pages 56-61. IEEE, 2023. 4", + "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1", + "[6] Todd A Ell, Nicolas Le Bihan, and Stephen J Sangwine. *Quaternion Fourier transforms for signal and image processing*. John Wiley & Sons, 2014. 4", + "[7] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), pages 1381-1384. IEEE, 1998. 6", + "[8] Xueyang Fu, Zihuang Lin, Yue Huang, and Xinghao Ding. A variational pan-sharpening with local gradient constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10265-10274, 2019. 2", + "[9] Xueyang Fu, Wu Wang, Yue Huang, Xinghao Ding, and John Paisley. Deep multiscale detail networks for multiband spectral image sharpening. IEEE Transactions on Neural Networks and Learning Systems, 32(5):2090-2104, 2020. 1", + "[10] Chase J Gaudet and Anthony S Maida. Deep quaternion networks. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2018. 3", + "[11] Pedram Ghamisi, Behnood Rasti, Naoto Yokoya, Qunming Wang, Bernhard Hofle, Lorenzo Bruzzone, Francesca Bovolo, Mingmin Chi, Katharina Anders, Richard Gloaguen, et al. Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine, 7(1):6-39, 2019. 2", + "[12] Alan R Gillespie, Anne B Kahle, and Richard E Walker. Color enhancement of highly correlated images. ii. channel" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ratio and \"chromaticity\" transformation techniques. Remote Sensing of Environment, 22(3):343-365, 1987. 7", + "[13] William Rowan Hamilton. Lectures on Quaternions: Containing a Systematic Statement of a New Mathematical Method; of which the Principles Were Communicated in 1843 to the Royal Irish Academy; and which Has Since Formed the Subject of Successive Courses of Lectures, Delivered in 1848 and Subsequent Years, in the Halls of Trinity College, Dublin: with Numerous Illustrative Diagrams, and with Some Geometrical and Physical Applications. Hodges and Smith, 1853. 3", + "[14] R Haydn. Application of the ihs color transform to the processing of multisensor data and image enhancement. In Proc. of the International Symposium on Remote Sensing of Arid and Semi-Arid Lands, Cairo, Egypt, 1982, 1982. 7", + "[15] Xiyan He, Laurent Condat, José M Bioucas-Dias, Jocelyn Chanussot, and Junshi Xia. A new pansharpening method based on spatial and spectral sparsity priors. IEEE Transactions on Image Processing, 23(9):4160-4174, 2014. 2", + "[16] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2121-2129, 2024. 3, 7", + "[17] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 5", + "[18] Creed F Jones III. Color face recognition using quaternionic Gabor filters. PhD thesis, Virginia Polytechnic Institute and State University, 2005. 3", + "[19] Hiromi Kusamichi, Teijiro Isokawa, Nobuyuki Matsui, Yuzo Ogawa, and Kazuaki Maeda. A new scheme for color night vision by quaternion neural network. In Proceedings of the 2nd international conference on autonomous robots and agents. CiteSeer, 2004. 3", + "[20] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening, 2000. US Patent 6,011,875. 7", + "[21] Pengfei Liu, Liang Xiao, and Tao Li. A variational pan-sharpening method based on spatial fractional-order geometry and spectral-spatial low-rank priors. IEEE Transactions on Geoscience and Remote Sensing, 56(3):1788-1802, 2017. 2", + "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6", + "[23] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. Remote Sensing, 8(7):594, 2016. 1, 2, 7", + "[24] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE Geoscience and Remote Sensing Magazine, 9(1):18-52, 2020. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10916", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jorge Nunez, Xavier Otazu, Octavi Fors, Albert Prades, Vicenc Pala, and Roman Arbiol. Multiresolution-based image fusion with additive wavelet decomposition. IEEE Transactions on Geoscience and Remote sensing, 37(3):1204-1211, 1999. 2", + "[26] Titouan Parcollet, Mirco Ravanelli, Mohamed Morchid, Georges Linares, Chiheb Trabelsi, Renato De Mori, and Yoshua Bengio. Quaternion recurrent neural networks. arXiv preprint arXiv:1806.04418, 2018. 2, 3", + "[27] Titouan Parcollet, Ying Zhang, Mohamed Morchid, Chiheb Trabelsi, Georges Linares, Renato De Mori, and Yoshua Bengio. *Quaternion convolutional neural networks for end-to-end automatic speech recognition*. arXiv preprint arXiv:1806.07789, 2018. 2, 3", + "[28] Soo-Chang Pei, Jian-Jiun Ding, and Ja-Han Chang. Efficient implementation of quaternion fourier transform, convolution, and correlation by 2-d complex FFT. IEEE transactions on signal processing, 49(11):2783-2797, 2001. 3", + "[29] Thierry Ranchin and Lucien Wald. Fusion of high spatial and spectral resolution images: The arsis concept and its implementation. Photogrammetric engineering and remote sensing, 66(1):49-61, 2000. 2", + "[30] Stephen John Sangwine. Fourier transforms of colour images using quaternion or hypercomplex, numbers. *Electronics letters*, 32(21):1979–1980, 1996. 3, 4", + "[31] Vijay P Shah, Nicolas H Younan, and Roger L King. An efficient pan-sharpening method via a combined adaptive pca approach and contourlets. IEEE transactions on geoscience and remote sensing, 46(5):1323-1335, 2008. 2", + "[32] Lilong Shi and Brian Funt. Quaternion color texture segmentation. Computer Vision and image understanding, 107(1-2): 88-96, 2007. 3", + "[33] Jiangtong Tan, Jie Huang, Naishan Zheng, Man Zhou, Keyu Yan, Danfeng Hong, and Feng Zhao. Revisiting spatial-frequency information integration from a hierarchical perspective for panchromatic and multi-spectral image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25922-25931, 2024. 3, 7", + "[34] Te-Ming Tu, Shun-Chi Su, Hsuen-Chyun Shyu, and Ping S Huang. A new look at ihs-like image fusion methods. Information fusion, 2(3):177-186, 2001. 2", + "[35] Gemine Vivone, Rocco Restaino, Mauro Dalla Mura, Giorgio Licciardi, and Jocelyn Chanussot. Contrast and error-based fusion schemes for multispectral image pansharpening. IEEE Geoscience and Remote Sensing Letters, 11(5): 930-934, 2013. 2", + "[36] Dong Wang, Yunpeng Bai, and Ying Li. Multispectral pan-sharpening via dual-channel convolutional network with convolutional LSTM based hierarchical spatial-spectral feature fusion, 2020. 3", + "[37] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2", + "[38] Chen Wu, Bo Du, Xiaohui Cui, and Liangpei Zhang. A post-classification change detection method based on itera" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive slow feature analysis and bayesian soft fusion. Remote Sensing of Environment, 199:241-255, 2017. 1", + "[39] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 2, 7", + "[40] Gang Yang, Xiangyong Cao, Wenzhe Xiao, Man Zhou, Aiping Liu, Xun Chen, and Deyu Meng. Panflownet: A flow-based deep network for pan-sharpening. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16857-16867, 2023. 1", + "[41] Licheng Yu, Yi Xu, Hongteng Xu, and Hao Zhang. Quaternion-based sparse representation of color image. In 2013 IEEE International Conference on Multimedia and Expo (ICME), pages 1-7. IEEE, 2013. 3", + "[42] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 2, 7", + "[43] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1", + "[44] Rui Zeng, Jiasong Wu, Zhuhong Shao, Yang Chen, Beijing Chen, Lotfi Senhadji, and Huazhong Shu. Color image classification via quaternion principal component analysis network. Neurocomputing, 216:416-428, 2016. 3", + "[45] Hao Zhang and Jiayi Ma. Gtp-pnet: A residual learning network based on gradient transformation prior for pansharpening. ISPRS Journal of Photogrammetry and Remote Sensing, 172:223-239, 2021. 1", + "[46] Yunzuo Zhang, Jiayu Zhang, Ruixue Liu, Pengfei Zhu, and Yameng Liu. Key frame extraction based on quaternion fourier transform with multiple features fusion. Expert Systems with Applications, 216:119467, 2023. 4", + "[47] Man Zhou, Jie Huang, Yanchi Fang, Xueyang Fu, and Aiping Liu. Pan-sharpening with customized transformer and invertible neural network. In Proceedings of the AAAI conference on artificial intelligence, pages 3553-3561, 2022. 2, 7", + "[48] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In European conference on computer vision, pages 274-291. Springer, 2022. 3", + "[49] Man Zhou, Keyu Yan, Jie Huang, Zihe Yang, Xueyang Fu, and Feng Zhao. Mutual information-driven pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1798-1808, 2022. 7", + "[50] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. FOurmer: An efficient global modeling paradigm for image restoration. In International conference on machine learning, pages 42589-42601. PMLR, 2023. 4, 5", + "[51] Xuanyu Zhu, Yi Xu, Hongteng Xu, and Changjian Chen. Quaternion convolutional neural networks. In Proceedings of" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10917", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "the European conference on computer vision (ECCV), pages 631-647, 2018. 3", + "[52] Zeyu Zhu, Xiangyong Cao, Man Zhou, Junhao Huang, and Deyu Meng. Probability-based global cross-modal upsampling for pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14039-14048, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 191 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10918", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_model.json b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_model.json new file mode 100644 index 0000000000000000000000000000000000000000..6e0f6d5e745bda954d525f6af78cd93dbbaf39d6 --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_model.json @@ -0,0 +1,2554 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.143, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions" + }, + { + "type": "text", + "bbox": [ + 0.15, + 0.181, + 0.854, + 0.217 + ], + "angle": 0, + "content": "Dong Li, Chunhui Luo, Yuanfei Bao, Gang Yang, Jie Xiao, Xueyang Fu, Zheng-Jun Zha* \nUniversity of Science and Technology of China, Hefei, China" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.22, + 0.688, + 0.234 + ], + "angle": 0, + "content": "dongli6@mail.ustc.edu.cn，zhazj@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.771 + ], + "angle": 0, + "content": "Pansharpening aims to generate high-resolution multispectral (MS) images by fusing panchromatic (PAN) images with corresponding low-resolution MS images. However, many existing methods struggle to fully capture spatial and spectral interactions, limiting their effectiveness. To address this, we propose a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging the compact representation capabilities of quaternions for high-dimensional data. Our method consists of three key components: the quaternion global spectral interaction branch, the quaternion local spatial structure awareness branch, and the quaternion spatial-spectral interaction branch. The first applies the quaternion Fourier transform to convert multi-channel features into the frequency domain as a whole, enabling global information interaction while preserving inter-channel dependencies, which aids spectral fidelity. The second uses a customized spatial quaternion representation, combined with a window-shifting strategy, to maintain local spatial dependencies while promoting spatial interactions, which helps inject spatial details. The last integrates the two pathways within the quaternion framework to enrich spatial-spectral interactions for richer representations. By utilizing quaternion's multi-dimensional representation and parameter-sharing properties, our method achieves a more compact and efficient cross-resolution, multi-band information integration, significantly improving the quality of the fused image. Extensive experiments validate the proposed method's effectiveness and its superior performance over current SOTA techniques. The code is available at https://github.com/dongli8/QuatPanNet." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.485, + 0.875 + ], + "angle": 0, + "content": "In remote sensing, satellite sensors face physical limitations that prevent the direct acquisition of high-resolution multispectral (MS) images [45]. Pansharpening addresses" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.271, + 0.903, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.483, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 1. Comparison between standard real-valued layer and quaternion-valued layer. The real-valued operation typically stacks the input multi-channel features, while the quaternion-valued layer, due to quaternion weight sharing, has the ability to learn the latent relationships within the input features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.713 + ], + "angle": 0, + "content": "this by merging low-resolution multispectral (LRMS) images with panchromatic (PAN) images to produce high-resolution multispectral (HRMS) images. As a crucial preprocessing step, pansharpening is widely applied across fields [38, 43]. Conceptually, it is a fusion super-resolution task, focused on incorporating fine spatial details while preserving spectral fidelity, requiring strong inter-channel dependency maintenance and the preservation of subtle local spatial structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.909, + 0.869 + ], + "angle": 0, + "content": "Traditional methods often rely on mathematical models that integrate spatial and spectral data, typically assuming that the pansharpened image is a linear combination of various spectral bands from a high-resolution multispectral image [40]. However, these approaches are limited by their dependence on prior knowledge, which restricts flexibility and adaptability. With the rise of deep learning, convolutional neural networks (CNNs) have been introduced into pansharpening, leading to substantial breakthroughs in performance with increased model complexity [1, 5, 9, 23]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Despite notable successes, existing methods still struggle to capture the intricate spatial-spectral interactions essential" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10908" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "for high-quality pansharpening. These limitations hinder simultaneous feature interaction and inter-feature dependency preservation, leading to compromises in both spectral fidelity and spatial detail. Unlike conventional real-valued networks, quaternion networks treat multidimensional features as unified entities, allowing them to effectively encode internal dependencies [26, 27]. Additionally, quaternion networks demonstrate strong modeling capabilities for both inter- and intra-correlations within the quaternion latent space [3]. Thus, we aim to harness quaternion operations to further elevate pansharpening performance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.262, + 0.485, + 0.64 + ], + "angle": 0, + "content": "Our Motivation. PAN images provide high-resolution spatial details, while MS images offer rich multispectral information [11, 24]. A successful pansharpening model must preserve both spatial details and spectral fidelity during fusion. This requires capturing the unique features of PAN and MS images while leveraging their complementary information to produce accurate and comprehensive fused outputs. Traditional approaches often handle multi-channel inputs by stacking and mapping them to a single output channel, which causes information loss. Consequently, they struggle to capture inter-channel correlations and interactions between spatial and spectral data effectively. To compensate, some methods increase the number of output channels to improve representational capacity [39], but this also increases model complexity. Furthermore, these methods typically extract features in real-number space, limiting their ability to capture directional information such as edges and structures. In contrast, quaternions represent multiple spectral channels as distinct components of a compact quaternion representation, as shown in Figure 1. This approach integrates multi-band information efficiently while reducing information loss through parameter-sharing quaternion operations. These features make quaternions particularly well-suited for pansharpening tasks that involve multi-source data inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Based on the analysis above, we propose a novel quaternion-based pansharpening framework called QuatPanNet. QuatPanNet is designed to balance spectral fidelity and spatial detail richness by leveraging quaternion-based representations and operations to enhance interactions among multispectral, spatial structure, and spatial-spectral features. The framework consists of three core components, i.e., Quaternion Global Spectral Interaction Branch,Quaternion Local Spatial Structure Awareness Branch, andQuaternion Spatial-Spectral Interaction Branch. The first one uses the Quaternion Fourier Transform (QFT) to process multi-channel features holistically in the frequency domain. By enabling global interactions while preserving inter-channel dependencies, it effectively extracts global spectral representations, which are critical for maintaining spectral fidelity. The second one employs a customized spatial quaternion transformation method and a window" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "shifting strategy, this branch focuses on preserving local structural dependencies and enhancing spatial interactions. It ensures that spatial details are well-integrated while maintaining the structural integrity of the input data. The last one, which built on the outputs of the first two branches, integrates dual-path features into a unified quaternion representation. This integration enhances the complementarity and coherence of information during the spatial-spectral fusion process. By utilizing the compact and efficient representation of PAN and MS image data in quaternion space, QuatPanNet facilitates effective interactions among spatial and spectral features. This approach reduces parameter complexity while achieving high spectral fidelity and rich spatial resolution, leading to superior fusion quality. In summary, the contributions of this work are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.322, + 0.905, + 0.367 + ], + "angle": 0, + "content": "- We propose a novel quaternion-based pansharpening framework, QuatPanNet, designed to comprehensively harness quaternions for this task." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.367, + 0.905, + 0.457 + ], + "angle": 0, + "content": "- We design customized quaternion-based spatial and spectral interaction modules for pansharpening. The former facilitates global information interaction while preserving inter-spectral dependencies, and the latter promotes spatial interaction while maintaining local structural dependencies." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.457, + 0.905, + 0.517 + ], + "angle": 0, + "content": "- We introduce a quaternion-based spatial-spectral interaction branch, leveraging the high-dimensional representational power of quaternions to improve the fusion of spectral and spatial features." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.322, + 0.905, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.905, + 0.581 + ], + "angle": 0, + "content": "Extensive experiments on multiple satellite datasets demonstrate that our method achieves superior qualitative and quantitative results with fewer parameters and effectively generalizes to full-resolution scenes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.597, + 0.663, + 0.612 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.622, + 0.666, + 0.639 + ], + "angle": 0, + "content": "2.1. Pansharpening" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Existing pansharpening methods are divided into traditional ones and deep - learning - based ones. Traditional ones mainly consist of Component Substitution (CS) [31, 34], Multi - Resolution Analysis (MRA) [25, 29, 35], and Variational Optimization (VO) [8, 15, 21]. However, their heavy dependence on manually designed features and prior knowledge restricts their applicability [37]. PNN [23] is the first to introduce CNN into the field of pansharpening, achieving better results than traditional methods. Since then, deep learning-based approaches have dominated this field. MSDCNN [42] combines multi-scale feature extraction and residual learning, enhancing the quality of the fused images. INNformer [47], based on the Transformer architecture, integrates an invertible neural network module for lossless information to model long-range dependencies and efficient feature fusion, successfully introducing the Transformer architecture into this domain. The fusion of frequency do" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10909" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "main features in images is also a key focus for researchers [36, 48]. For instance, FAME-Net [16] proposes a method that combines MOE with frequency domain information, enabling the network to dynamically learn high-frequency information in remote sensing images. HFIN [33] explores the relationship between PAN and LRMS images using local Fourier information integration. In addition to designing pansharpening models, Zhu et al. developed the first upsampling module specifically for pansharpening methods, called PGCU [52]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.252, + 0.218, + 0.267 + ], + "angle": 0, + "content": "2.2.Quaternion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.485, + 0.651 + ], + "angle": 0, + "content": "Quaternions are the natural extension of complex numbers to four dimensions, proposed by Hamilton in 1843 [13]. A quaternion consists of four components, including one real part and three imaginary parts. Quaternions, with their four components, are particularly well-suited for three-dimensional and four-dimensional feature vectors, as seen in applications such as image processing and robotic kinematics [2, 30]. In 2001, Pei et al. [28] introduced the quaternion model for color images, marking the beginning of quaternion applications in the field of image processing. The advantages of quaternion wavelet transforms [18], quaternion principal component analysis [44], and other quaternion-based color image processing techniques [41] have been demonstrated to extract more representative features for color images. These techniques have achieved promising results in advanced visual tasks such as color image classification. With the wave of deep learning, many researchers have begun to explore the application of quaternions in the field of deep learning [10, 26]. Xu et al. [51] proposes QCNN, which combines quaternions with CNNs, achieving faster convergence of the loss function while preserving color information more effectively than real-valued CNNs. Kusamichi et al. [19] studies color night vision and finds that using quaternion networks yields better results in extracting color information from dim images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.665, + 0.182, + 0.68 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this section, we first introduce the relevant properties of quaternions and their operators, and then present the proposed quaternion pan-sharpening network, as shown in Figures 2 and 3. We then provide a detailed explanation of the fundamental components of our method, including three quaternion-based branches: a) the global interaction branch utilizes the Quaternion Fourier Transform and quaternion convolution to maintain channel interdependencies while representing global information; b) the local structure-aware branch models local spatial dependencies through a specially designed spatial quaternion transformation method combined with a window-shifting strategy; and c) the dual-domain interaction branch leverages the orthogonal relationships inherent in quaternions to model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "dual-domain correlations and promote complementary feature learning. Finally, we present the optimization strategy employed in our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.145, + 0.723, + 0.162 + ], + "angle": 0, + "content": "3.1.Quaternion Operators" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.167, + 0.906, + 0.213 + ], + "angle": 0, + "content": "Quaternion, proposed by Hamilton in 1843, extends complex numbers to a four-dimensional space. A quaternion \\(Q \\in \\mathbb{H}\\) can be expressed as follows [32]:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.223, + 0.907, + 0.239 + ], + "angle": 0, + "content": "\\[\nQ = w 1 + x i + y j + z k, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.906, + 0.322 + ], + "angle": 0, + "content": "where \\( w, x, y, z \\) are all real numbers, and \\( 1, i, j, k \\) are the quaternion unit basis. \\( w1 \\) refers to the real component of the number, and \\( xi + yj + zk \\) to the imaginary component. In the definition of quaternions, they must satisfy the following identities, known as the Hamilton rules:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.331, + 0.907, + 0.349 + ], + "angle": 0, + "content": "\\[\ni ^ {2} = j ^ {2} = k ^ {2} = i j k = - 1. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.463 + ], + "angle": 0, + "content": "We propose the use of quaternions and their associated operators to uncover both intra- and inter-relations between the features of high spatial resolution PAN and MS images. This approach is inspired by the unique feature-handling capabilities inherent in quaternion representation [3]. Consequently, we introduce two quaternion operators: quaternion convolution and the quaternion Fourier transform." + }, + { + "type": "title", + "bbox": [ + 0.534, + 0.465, + 0.71, + 0.478 + ], + "angle": 0, + "content": "Quaternion Convolution" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.645 + ], + "angle": 0, + "content": "In the real-valued space, convolution involves convolving a filter matrix with a vector. Extending this operation to quaternions, quaternion convolution is achieved by convolving a quaternion filter matrix with a quaternion vector. To maintain consistency with the defined quaternions, we represent the computation of quaternion convolution using real-valued matrices. This computation is performed by convolving the quaternion filter matrix \\( W = A + Bi + Cj + Dk \\) with the quaternion vector \\( Q = w + xi + yj + zk \\), where \\( A, B, C \\), and \\( D \\) are real-valued matrices. The convolution operation between \\( W \\) and \\( Q \\) is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.654, + 0.905, + 0.727 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} W * Q = (A * w - B * x - C * y - D * z) + \\\\ i (A * x + B * w + C * z - D * y) + \\\\ j (A * y - B * z + C * w + D * x) + \\tag {3} \\\\ K (A * z + B * y - C * x + D * w). \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "It can be seen that the result of quaternion convolution produces a unique linear combination of each coordinate axis for every single-axis result. This stems from the structure of quaternion multiplication and enforces interactions between each axis of the kernel and each axis of the image. This contrasts with real-valued convolution, which simply multiplies each channel of the kernel with the corresponding channel of the image. The advantage of this approach is that quaternion convolution promotes cross-axis interactions and information extraction through the reuse of filters across each axis and combination [27]." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10910" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.088, + 0.907, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.353, + 0.907, + 0.381 + ], + "angle": 0, + "content": "Figure 2. The framework of our proposed pan-sharpening network QuatPanNet. It consists of several Quaternion Spatial-Spectral Interaction Blocks (QSSIB). The Quaternion Spatial-Spectral Interaction Block is shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.402, + 0.333, + 0.416 + ], + "angle": 0, + "content": "Quaternion Fourier Transform" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.417, + 0.483, + 0.568 + ], + "angle": 0, + "content": "The Fourier Transform plays a crucial role in various image processing tasks, including pansharpening [50]. After decomposing the input image into multiple channels, the Discrete Fourier Transform (DFT) can be applied to each channel individually. However, this approach neglects the mutual information among different channels, which can hinder the performance of pansharpening. The Discrete Quaternion Fourier Transform (DQFT) offers the possibility of simultaneously transforming multi-channel images, as it can transform quaternions holistically [4]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.659 + ], + "angle": 0, + "content": "Due to the non-commutative nature of quaternion multiplication, there are various types of Quaternion Fourier Transforms (QFT), such as left-side, right-side and two-side QFT [6]. In this paper, we utilize a typical QFT, specifically the left-sided QFT. For an image \\(\\dot{f}\\) of size \\(H\\times W\\), the Discrete Quaternion Fourier Transform is defined as follows." + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.679, + 0.482, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} (\\dot {f}) (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} e ^ {- \\dot {\\mu} 2 \\pi \\left(\\frac {x}{H} u + \\frac {y}{H} v\\right)} \\dot {f} (x, y), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.84 + ], + "angle": 0, + "content": "where \\(\\dot{f} (x,y)\\) is pixel of \\(\\dot{f}\\) and \\(\\dot{\\mu}\\) is unit pure quaternion. \\((u,v)\\) is the coordinate in the quaternion frequency domain. \\(\\mathcal{F}^{-1}(\\dot{f})\\) defines the inverse quaternion Fourier transform accordingly. Inspired by the traditional Fourier transform, the modulus of \\(\\mathcal{F}(\\dot{f})(u,v)\\) is taken as the amplitude component, and the angles with the three imaginary axes are taken as the phase components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "It is worth noting that the QFT is not a simple substitution of the imaginary unit in the two-dimensional Fourier Transform with quaternions; rather, its basis functions possess orthogonal imaginary components, which preserve the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.599 + ], + "angle": 0, + "content": "information and mutual relationships among different components in the frequency domain. Consequently, some studies leverage QFT to treat multiple features as a unified whole to enhance performance [4, 46]. Furthermore, the efficiency of QFT is also noteworthy. For example, when comparing the use of QFT with three separate conventional Fourier Transforms for processing color images, the number of real multiplications and additions required by QFT is approximately \\(75\\%\\) of that needed for the three separate transforms [30]. For pansharpening, we propose utilizing the QFT to treat multispectral data as a unified whole, facilitating global information interactions in the frequency domain while preserving inter-channel dependencies." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.618, + 0.642, + 0.634 + ], + "angle": 0, + "content": "3.2. Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Based on the above analysis, we propose a novel pansharpening approach based on quaternion information interaction, as illustrated in Figure 2. Given a panchromatic (PAN) image \\( P \\in \\mathbb{R}^{H \\times W \\times 1} \\) and a low-resolution multispectral (LRMS) image \\( L \\in \\mathbb{R}^{H / r \\times W / r \\times C} \\), the network first uses convolution to project the bicubic upsampled r-times \\( L \\) into a shallow feature representation. Simultaneously, \\( P \\) is processed through multiple cascaded convolution layers to extract effective information for high-resolution multispectral (HRMS) reconstruction. Then, the features of the PAN and LRMS images are jointly processed through the core building module, the quaternion spatial-spectral interaction block (QSSIB), for continuous global information interaction and local structure-aware integration. Next, a convolution layer transforms all features gathered by the N QSSIBs back into the image space, and finally, these features are combined with the residual to obtain the final output." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10911" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.91, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.416 + ], + "angle": 0, + "content": "Figure 3. The detailed flowchart of the proposed core building module quaternion spatial-spectral interaction block (QSSIB), consisting of three components: quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch (QLSA) and quaternion spatial-spectral interaction branch (QSSI)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.422, + 0.332, + 0.439 + ], + "angle": 0, + "content": "3.3. Key Building Components" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.445, + 0.486, + 0.703 + ], + "angle": 0, + "content": "As shown in Figure 3, the fundamental building block of our method comprises three key elements: the quaternion global spectral interaction branch (QGSI), the quaternion local spatial structure awareness branch (QLSA), and the quaternion spatial-spectral interaction branch (QSSI). To convert images into quaternion representations, we design customized methods, including channel quaternion representation (CQR), spatial quaternion representation (SQR), and shifted window quaternion representation (SWQR). When transforming quaternion representations back into image features, we denote the corresponding inverse operations as ICQR, ISQR, and ISWQR, respectively. (For more details on quaternion representation, refer to the supplementary material.) These three key components, in conjunction with the quaternion representations of the images, jointly facilitate channel interactions during multispectral fusion and local interactions for spatial resolution enhancement." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.705, + 0.393, + 0.719 + ], + "angle": 0, + "content": "Quaternion Global Spectral Interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.485, + 0.904 + ], + "angle": 0, + "content": "The global interaction branch utilizes the Quaternion Fourier Transform (QFT) to transform multi-channel features as a whole into the frequency domain, facilitating global information interactions. Since the input to QFT must be in quaternion form and we aim to maintain interchannel interdependencies, we design a channel quaternion representation (CQR). This approach divides the image features equally along the channel dimension into four parts, which are then used as the four components of a quaternion. Let the features of the MS and PAN images be denoted as \\( F_{ms} \\) and \\( F_{p} \\), respectively; their channel-dimensional quaternion representations are denoted as \\( Q_{ms}^{c} \\) and \\( Q_{p}^{c} \\). We" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.529 + ], + "angle": 0, + "content": "then use QFT to transform the MS and PAN features into the frequency domain. Some studies on image restoration indicate that decomposing amplitude and phase components facilitates the decoupling of degradation [17, 50]. Therefore, we decompose the frequency domain features into amplitude and phase components, with the corresponding process represented as" + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.534, + 0.907, + 0.553 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {p} ^ {c}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.555, + 0.905, + 0.572 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {m s} ^ {c}\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.698 + ], + "angle": 0, + "content": "where \\(\\mathcal{A}(\\cdot)\\) represents the amplitude, and \\(\\mathcal{P}_1(\\cdot),\\mathcal{P}_2(\\cdot)\\) and \\(\\mathcal{P}_3(\\cdot)\\) represent the phases corresponding to the three imaginary axes. We then concatenate the amplitude and phase spectra of the PAN and MS images accordingly to form new quaternions. Next, we process the newly formed quaternions using quaternion convolution, thereby enhancing global representations in the frequency domain while promoting interactions between channels." + }, + { + "type": "equation", + "bbox": [ + 0.521, + 0.702, + 0.905, + 0.756 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q _ {f} ^ {c} = \\mathcal {O Q} (\\mathcal {C} [ \\mathcal {A} (Q _ {p} ^ {c}), \\mathcal {A} (Q _ {m s} ^ {c}) ], \\mathcal {C} [ \\mathcal {P} _ {1} (Q _ {p} ^ {c}), \\mathcal {P} _ {1} (Q _ {m s} ^ {c}) ] \\\\ , \\mathcal {C} \\left[ \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right) \\right], \\mathcal {C} \\left[ \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) \\right]) \\tag {7} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.907, + 0.835 + ], + "angle": 0, + "content": "where \\(\\mathcal{C}\\) denotes concatenation along the channel dimension, \\(\\mathcal{OQ}\\) is quaternion convolution. The fused amplitude and phase components are the transformed back to the spatial domain using the inverse discrete quaternion Fourier transform." + }, + { + "type": "equation", + "bbox": [ + 0.643, + 0.836, + 0.907, + 0.855 + ], + "angle": 0, + "content": "\\[\nQ _ {f r e} ^ {c} = \\mathcal {F} ^ {- 1} \\left(Q _ {f} ^ {c}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "By representing the image in a channel-dimensional quaternion form, the quaternion operator treats the multichannel data as a unified whole for information interaction." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10912" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.169 + ], + "angle": 0, + "content": "Additionally, according to spectral convolution theorem [7], processing information in the Fourier space enables capturing global frequency representations. In summary, this branch generates the global information representation with strengthened channel interactions \\( Q_{fre}^{c} \\)." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.169, + 0.442, + 0.183 + ], + "angle": 0, + "content": "Quaternion Local Spatial Structure Awareness" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Pansharpening is essentially a fusion-based superresolution task, making the injection of spatial resolution critically important. Enhancing spatial resolution requires not only global information from the frequency domain but also local information interactions in the spatial domain. For pansharpening, preserving local structures during information interactions is particularly significant, as it enhances the network's ability to recover fine details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.483, + 0.472 + ], + "angle": 0, + "content": "We aim to leverage the balance of enhanced information interaction and specific information preservation offered by quaternions to improve the network's spatial superresolution capabilities. To this end, we design the spatial quaternion representation (SQR). It divides the image into several \\(2 \\times 2\\) windows in the spatial dimension, with each local window's four pixels serving as the four components of a quaternion, as illustrated in Figure 3. This approach explicitly ensures dependencies among local pixels, preserving the local structure. Based on this representation, we use quaternion convolutions (QCNN) to process the features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.473, + 0.483, + 0.625 + ], + "angle": 0, + "content": "However, this representation method lacks crosswindow connections, which limits spatial interactions. Inspired by [22], we design the shifted window quaternion representation (SWQR) based on SQR. This allows the network to alternately employ two partitioning configurations, thereby enhancing the effectiveness of spatial quaternion representation. Through the combined effect of this method and quaternion convolution, this branch preserves local spatial dependencies while promoting spatial interactions, thereby obtaining the local representation \\( Q_{spa}^{s} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.625, + 0.483, + 0.715 + ], + "angle": 0, + "content": "The information representations of the two branches are highly complementary, with one branch responsible for global, frequency-domain, and channel interactions, and the other focused on local, spatial-domain, and spatial interactions. Therefore, further interaction and integration between them facilitate mutual compensation." + }, + { + "type": "title", + "bbox": [ + 0.11, + 0.716, + 0.395, + 0.73 + ], + "angle": 0, + "content": "Quaternion Spatial-Spectral Interaction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.483, + 0.84 + ], + "angle": 0, + "content": "The two branches described above respectively obtain quaternion representations in the channel dimension, \\( Q_{fre}^{c} \\in \\mathbb{H}^{H \\times W \\times C / 4} \\), and in the spatial dimension, \\( Q_{spa}^{s} \\in \\mathbb{H}^{H / 2 \\times W / 2 \\times C} \\). The quaternion dual-domain information interaction branch mainly consists of two parts: one is a quaternion-based global fusion, and the other is the channel-aware quaternion spatial feature injection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Quaternion-based global fusion merges the features obtained from the two branches into a single quaternion for unified processing, thereby enhancing the correlation between the two through the structure of the quaternion. First," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.248 + ], + "angle": 0, + "content": "both \\(Q_{fre}^{c}\\) and \\(Q_{spa}^{s}\\) are converted back into image feature representations. \\(Q_{fre}^{c}\\) is combined along the channel dimension to obtain \\(F_{fre} \\in \\mathbb{R}^{H \\times W \\times C}\\), and \\(Q_{spa}^{s}\\) is rearranged spatially to obtain \\(F_{spa} \\in \\mathbb{R}^{H \\times W \\times C}\\). To achieve a richer feature representation, we apply convolutions to \\(F_{fre}\\) and \\(F_{spa}\\) to obtain \\(F_{fre}'\\) and \\(F_{spa}'\\), respectively. We then combine these four features into a quaternion \\(Q_{f} = F_{fre} + F_{spa}i + F_{fre}^{\\prime}j + F_{spa}^{\\prime}k\\). We use quaternion convolution to process \\(Q_{f} \\in \\mathbb{H}^{H \\times W \\times C}\\) as a whole, producing \\(Q_{a} \\in \\mathbb{H}^{H \\times W \\times C/4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.249, + 0.905, + 0.387 + ], + "angle": 0, + "content": "Channel-aware quaternion spatial feature injection involves injecting spatial features \\(F_{spa}\\) from the local structure-aware branch into the four components of the quaternion representation \\(Q_{fre}^{c}\\) in the channel dimension, thereby facilitating information fusion in both spatial and channel domains. Specifically, we replicate the image feature \\(F_{spa}\\) four times and concatenate each copy with one of the four components of the quaternion feature \\(Q_{fre}^{c}\\) to obtain \\(Q_{i} \\in \\mathbb{H}^{H \\times W \\times 5C / 4}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.393, + 0.905, + 0.412 + ], + "angle": 0, + "content": "\\[\nQ _ {i} = \\mathcal {C} \\left[ Q _ {f r e} ^ {c}, \\left(F _ {s p a} + F _ {s p a} i + F _ {s p a} j + F _ {s p a} k\\right) \\right]. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.416, + 0.905, + 0.477 + ], + "angle": 0, + "content": "We then feed \\( Q_{i} \\) into quaternion convolution to obtain \\( Q_{b} \\in \\mathbb{H}^{H \\times W \\times C / 4} \\). Afterward, we concatenate \\( Q_{a} \\) and \\( Q_{b} \\), and apply quaternion convolution for further fusion to obtain \\( Q_{out} \\in \\mathbb{H}^{H \\times W \\times C / 4} \\)." + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.483, + 0.905, + 0.501 + ], + "angle": 0, + "content": "\\[\nQ _ {o u t} = \\operatorname {Q C o n v} \\left(\\mathcal {C} \\left[ Q _ {a}, Q _ {b} \\right]\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.905, + 0.598 + ], + "angle": 0, + "content": "where QConv denotes quaternion convolution. We then convert \\(Q_{out}\\) into the image feature representation \\(F_{out} \\in \\mathbb{R}^{H \\times W \\times C}\\). Finally, a residual learning mechanism is applied, adding the input feature \\(F_{ms}\\) to the fused feature. The proposed network achieves effective fusion through the seamless integration of the two quaternion representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.605, + 0.654, + 0.621 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.627, + 0.905, + 0.747 + ], + "angle": 0, + "content": "Let \\( Y \\) and \\( GT \\) denote the network's output and the corresponding ground truth, respectively. We use a joint spatial-frequency domain loss to supervise the network training. In the frequency domain, we first apply the quaternion Fourier transform to convert \\( Y \\) and \\( GT \\) into the quaternion Fourier space. Then, we compute the L1 norm of the real and imaginary differences between \\( Y \\) and \\( GT \\), and sum them to obtain the quaternion Fourier domain loss." + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.753, + 0.905, + 0.79 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {f r e} = \\| R (Y) - R (G T) \\| _ {1} + \\| I (Y) - I (G T) \\| _ {1} \\tag {11} \\\\ + \\| J (Y) - J (G T) \\| _ {1} + \\| K (Y) - K (G T) \\| _ {1}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.855 + ], + "angle": 0, + "content": "where \\( R(\\cdot), I(\\cdot), J(\\cdot) \\), and \\( K(\\cdot) \\) represent the real part and three imaginary parts of the quaternion respectively. In the spatial domain, we employ the L1 loss. Finally, the overall loss function is as follows" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.863, + 0.905, + 0.881 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\| Y - G T \\| _ {1} + \\lambda \\mathcal {L} _ {f r e}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.895, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is the weighting factor and set to 0.1 empirically." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10913" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.214, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.184, + 0.179, + 0.196 + ], + "angle": 0, + "content": "LRMS" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.089, + 0.353, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.276, + 0.184, + 0.31, + 0.197 + ], + "angle": 0, + "content": "PAN" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.089, + 0.491, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.419, + 0.184, + 0.442, + 0.196 + ], + "angle": 0, + "content": "GS" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.629, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.184, + 0.586, + 0.197 + ], + "angle": 0, + "content": "PNN" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.089, + 0.768, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.184, + 0.743, + 0.196 + ], + "angle": 0, + "content": "MSDCNN" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.089, + 0.906, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.184, + 0.873, + 0.196 + ], + "angle": 0, + "content": "GPPNN" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.198, + 0.214, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.294, + 0.182, + 0.307 + ], + "angle": 0, + "content": "MutNet" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.198, + 0.353, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.294, + 0.331, + 0.307 + ], + "angle": 0, + "content": "INNformer" + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.198, + 0.491, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.411, + 0.294, + 0.451, + 0.306 + ], + "angle": 0, + "content": "HFIN" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.198, + 0.629, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.531, + 0.294, + 0.608, + 0.306 + ], + "angle": 0, + "content": "FAME-Net" + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.198, + 0.768, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.69, + 0.294, + 0.724, + 0.306 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.198, + 0.906, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.833, + 0.294, + 0.856, + 0.306 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.31, + 0.177, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.372, + 0.146, + 0.384 + ], + "angle": 0, + "content": "GS" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.311, + 0.265, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.208, + 0.372, + 0.24, + 0.384 + ], + "angle": 0, + "content": "PNN" + }, + { + "type": "image", + "bbox": [ + 0.27, + 0.311, + 0.353, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.28, + 0.372, + 0.343, + 0.384 + ], + "angle": 0, + "content": "MSDCNN" + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.311, + 0.441, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.372, + 0.425, + 0.384 + ], + "angle": 0, + "content": "GPPNN" + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.311, + 0.528, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.465, + 0.372, + 0.512, + 0.384 + ], + "angle": 0, + "content": "MutNet" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.311, + 0.617, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.543, + 0.372, + 0.611, + 0.384 + ], + "angle": 0, + "content": "INNformer" + }, + { + "type": "image", + "bbox": [ + 0.622, + 0.311, + 0.705, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.647, + 0.372, + 0.682, + 0.384 + ], + "angle": 0, + "content": "HFIN" + }, + { + "type": "image", + "bbox": [ + 0.711, + 0.311, + 0.793, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.72, + 0.372, + 0.786, + 0.384 + ], + "angle": 0, + "content": "FAME-Net" + }, + { + "type": "image", + "bbox": [ + 0.8, + 0.311, + 0.882, + 0.368 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.827, + 0.372, + 0.856, + 0.384 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.883, + 0.306, + 0.902, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.155, + 0.398, + 0.842, + 0.413 + ], + "angle": 0, + "content": "Figure 4. The visual comparisons between other pan-sharpening methods and our method on WorldView-II satellite" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.428, + 0.907, + 0.456 + ], + "angle": 0, + "content": "Table 1. Experimental results of all the competing methods on the three benchmark datasets. The best and the second best values are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.466, + 0.907, + 0.649 + ], + "angle": 0, + "content": "
MethodsParams (M)WorldView IIWorldView IIIGaoFen2
PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓
Brovey-35.86460.92160.04031.823822.50600.54660.11598.233137.79740.90260.02181.3720
IHS-35.29620.90270.04612.027822.55790.53540.12668.361638.17540.91000.02431.5336
GS-35.63760.91760.04231.877422.56080.54700.12178.243337.22600.90340.03091.6736
PNN0.068940.75500.96240.02591.064629.94180.91210.08243.320643.12080.97040.01720.8528
MSDCNN0.239041.33550.96640.02420.99430.30380.91840.07823.188445.68740.98270.01350.6389
GPPNN0.119841.16220.96840.02441.031530.17850.91750.07763.259344.21450.98150.01370.7361
MutNet0.071441.67730.97050.02240.951930.49070.92230.07493.112547.30420.98920.01020.5481
INNformer0.070641.69030.97040.02270.951430.53650.92250.07473.099747.35280.98930.01020.5479
HFIN0.077242.23190.97140.02150.880730.61470.92030.07423.078648.87830.98980.00930.4591
FAME-Net0.140842.02620.97230.02150.917230.99030.92870.06972.953147.67210.98980.00980.5242
Ours0.038742.48460.97380.02080.855431.17660.93320.05212.576149.14230.99010.00900.4474
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.673, + 0.224, + 0.689 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.698, + 0.318, + 0.712 + ], + "angle": 0, + "content": "4.1. Dataset and benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We conduct our experiments on three classic datasets. WorldView-II (WV2), Gaofen2 (GF2) and WorldView-III (WV3). The WorldView-II dataset consists of 760 image pairs for training and 80 image pairs for testing. The WorldView-III dataset includes 2150 image pairs for training and 200 image pairs for testing. The GaoFen-2 dataset comprises 2712 image pairs for training and 200 image pairs for testing. To validate the effectiveness of our approach, we compare it with several state-of-the-art pansharpening methods, including PNN [23], MSDCNN [42], GPPNN [39], MutNet [49], INNformer [47], HFIN [33] and FAME-Net [16], as well as traditional pansharpening meth" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.674, + 0.834, + 0.689 + ], + "angle": 0, + "content": "ods such as GS [20], Brovey [12], and IHS [14]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In our experiments, all deep learning models are implemented using PyTorch and trained on an NVIDIA GeForce GTX 3090 GPU. For each set, the multispectral (MS) images are cropped into patches with the size of \\(32 \\times 32\\), and the corresponding panchromatic (PAN) images are of size \\(128 \\times 128\\). We employ common evaluation metrics, including PSNR, SSIM, SAM, and ERGAS. Additionally, we utilize three widely-used no-reference IQA metrics for real-world full-resolution scenes: \\(D_{\\lambda}, D_{S}\\) and QNR. More descriptions of the implementation details and dataset are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10914" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.116, + 0.09, + 0.458, + 0.103 + ], + "angle": 0, + "content": "Table 2. Non-reference metrics on full-resolution dataset." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.115, + 0.482, + 0.155 + ], + "angle": 0, + "content": "
PNNMSDCNNGPPNNINNformerMutNetHFINFAME-NetOurs
Dλ↓0.07460.07340.07820.06970.06940.07100.06740.0566
Ds↓0.11640.11510.12530.11280.11180.10980.11210.1055
QNRT0.81910.82150.80730.82530.82470.82610.82910.8447
" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.18, + 0.451, + 0.196 + ], + "angle": 0, + "content": "4.2. Comparison with state-of-the-art methods" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.485, + 0.518 + ], + "angle": 0, + "content": "Evaluation on reduced-resolution scene. To evaluate the effectiveness of our method, we conducted experiments on three datasets and compared it with ten other state-of-the-art pansharpening methods. Table 1 summarizes the evaluation metrics for the three datasets, where the best result for each metric is highlighted in red and the second-best result in blue. The results demonstrate that our method significantly outperforms the others across all evaluation metrics, achieving the best performance on all three datasets. Compared to the second-best method in terms of PSNR for each dataset, our method shows an improvement of \\(0.25\\mathrm{dB}\\) on the WV2 dataset, \\(0.18\\mathrm{dB}\\) on the GF2 dataset, and \\(0.27\\mathrm{dB}\\) on the WV3 dataset, indicating the consistency of our method with the true high-resolution images. We also conducted qualitative experiments, Figures 4 present samples from the WV2 datasets. The top two rows show the comparison of experimental results, while the last row displays the mean squared error maps of each network's results compared to the ground truth. It is evident that the results obtained by our network are closest to the ground truth, with the smallest discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.671 + ], + "angle": 0, + "content": "Evaluation on full-resolution scene. We also conduct full-resolution analysis in real-world scenarios to validate the generalization capability of our method. We perform experiments on an additional 200 sets of the GF2 dataset. Due to the absence of high-resolution multispectral images for actual real-world scenarios, we employ three commonly used no-reference metrics, \\( D_{\\lambda} \\), \\( D_{S} \\), and QNR for evaluation. The experimental results, as shown in Table 2, demonstrate that our method achieves the best performance across all three metrics." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.683, + 0.483, + 0.725 + ], + "angle": 0, + "content": "Table 3. The results of different configurations on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.73, + 0.472, + 0.789 + ], + "angle": 0, + "content": "
ConfigurationQGSIQLSAQSSIPSNR ↑SSIM ↑SAM ↓ERGAS ↓
IX✓✓41.73560.96730.02470.9337
II✓X✓42.02170.97060.02190.9134
III✓✓X42.17340.97130.02210.9012
Ours✓✓✓42.48460.97380.02080.8554
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.298, + 0.835 + ], + "angle": 0, + "content": "4.3. Ablation Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We perform an ablation study using the WorldView-II dataset to further demonstrate the effectiveness of our method. The quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 4. PSNR values of Our model with different number of blocks on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.143, + 0.905, + 0.248 + ], + "angle": 0, + "content": "
Stages (K)PSNR ↑SSIM ↑SAM ↓ERGAS ↓
141.83780.96030.02490.9879
241.90020.96470.02380.9725
342.09130.96980.02340.9578
442.23690.97240.02150.9007
542.48460.97380.02080.8554
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.348 + ], + "angle": 0, + "content": "(QLSA), and quaternion spatial-spectral interaction branch (QSSI) are the core components of the proposed method, and we conduct separate ablations for each. Additionally, we also explore the impact of the number of basic building blocks used in the network through ablation studies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.906, + 0.469 + ], + "angle": 0, + "content": "To validate the effectiveness of the three key quaternion components, QGSI, QLSA, and QSSI, we replace their quaternion operations with real-valued operations, as shown in Table 3. The results demonstrate that removing each quaternion component leads to a decline in performance, thus confirming the effectiveness of the quaternion's multidimensional representation and parameter-sharing mechanism for pansharpening." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.47, + 0.906, + 0.606 + ], + "angle": 0, + "content": "We conduct experiments on networks with varying numbers of core building blocks to evaluate the impact of the number of core modules, as shown in Table 4. As observed, the model performance improves with the increase in the number of Quaternion Spatial-Spectral Interaction Blocks (QSSIB), though this comes at the cost of increased resource consumption. We balance performance and resource consumption by using the default setting of 5. More ablation experiments are shown in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose QuatPanNet, a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging quaternions' compact representation for high-dimensional data. Our work demonstrates the advantages of a unified quaternion framework for effectively addressing the pansharpening problem. We custom-design quaternion-based spectral and spatial interaction modules: the former promotes global information interaction while preserving spectral dependencies, and the latter fosters spatial interaction while maintaining local structural dependencies. Building on these, we integrate the dual-path features into a unified quaternion representation. By utilizing the compact and efficient representation of PAN and MS image data in the quaternion space, QuatPanNet facilitates effective interactions between spatial and spectral features. Our method achieves exceptional fusion quality with a reduced number of parameters." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10915" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.486, + 0.161 + ], + "angle": 0, + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.254 + ], + "angle": 0, + "content": "[1] Paolo Addesso, Gemine Vivone, Rocco Restaino, and Jocelyn Chanussot. A data-driven model-based regression applied to panchromatic sharpening. IEEE Transactions on Image Processing, 29:7779-7794, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Nicholas A Asparagus and John K Dimitros. A comparative study of three methods for robot kinematics. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(2):135-145, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.484, + 0.38 + ], + "angle": 0, + "content": "[3] Qinglong Cao, Zhengqin Xu, Yuntian Chen, Chao Ma, and Xiaokang Yang. Domain prompt learning with quaternion networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26637-26646, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.382, + 0.484, + 0.436 + ], + "angle": 0, + "content": "[4] Ping Cui, Cui-Ming Zou, Bang-Ju Wang, and Feng Shi. Color image fusion in quaternion spectral domain. In 2023 International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR), pages 56-61. IEEE, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.493 + ], + "angle": 0, + "content": "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.495, + 0.484, + 0.536 + ], + "angle": 0, + "content": "[6] Todd A Ell, Nicolas Le Bihan, and Stephen J Sangwine. *Quaternion Fourier transforms for signal and image processing*. John Wiley & Sons, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.484, + 0.605 + ], + "angle": 0, + "content": "[7] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), pages 1381-1384. IEEE, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[8] Xueyang Fu, Zihuang Lin, Yue Huang, and Xinghao Ding. A variational pan-sharpening with local gradient constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10265-10274, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[9] Xueyang Fu, Wu Wang, Yue Huang, Xinghao Ding, and John Paisley. Deep multiscale detail networks for multiband spectral image sharpening. IEEE Transactions on Neural Networks and Learning Systems, 32(5):2090-2104, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.484, + 0.774 + ], + "angle": 0, + "content": "[10] Chase J Gaudet and Anthony S Maida. Deep quaternion networks. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[11] Pedram Ghamisi, Behnood Rasti, Naoto Yokoya, Qunming Wang, Bernhard Hofle, Lorenzo Bruzzone, Francesca Bovolo, Mingmin Chi, Katharina Anders, Richard Gloaguen, et al. Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine, 7(1):6-39, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[12] Alan R Gillespie, Anne B Kahle, and Richard E Walker. Color enhancement of highly correlated images. ii. channel" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "ratio and \"chromaticity\" transformation techniques. Remote Sensing of Environment, 22(3):343-365, 1987. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.244 + ], + "angle": 0, + "content": "[13] William Rowan Hamilton. Lectures on Quaternions: Containing a Systematic Statement of a New Mathematical Method; of which the Principles Were Communicated in 1843 to the Royal Irish Academy; and which Has Since Formed the Subject of Successive Courses of Lectures, Delivered in 1848 and Subsequent Years, in the Halls of Trinity College, Dublin: with Numerous Illustrative Diagrams, and with Some Geometrical and Physical Applications. Hodges and Smith, 1853. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.906, + 0.301 + ], + "angle": 0, + "content": "[14] R Haydn. Application of the ihs color transform to the processing of multisensor data and image enhancement. In Proc. of the International Symposium on Remote Sensing of Arid and Semi-Arid Lands, Cairo, Egypt, 1982, 1982. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[15] Xiyan He, Laurent Condat, José M Bioucas-Dias, Jocelyn Chanussot, and Junshi Xia. A new pansharpening method based on spatial and spectral sparsity priors. IEEE Transactions on Image Processing, 23(9):4160-4174, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[16] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2121-2129, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.483 + ], + "angle": 0, + "content": "[17] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.525 + ], + "angle": 0, + "content": "[18] Creed F Jones III. Color face recognition using quaternionic Gabor filters. PhD thesis, Virginia Polytechnic Institute and State University, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.907, + 0.593 + ], + "angle": 0, + "content": "[19] Hiromi Kusamichi, Teijiro Isokawa, Nobuyuki Matsui, Yuzo Ogawa, and Kazuaki Maeda. A new scheme for color night vision by quaternion neural network. In Proceedings of the 2nd international conference on autonomous robots and agents. CiteSeer, 2004. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.596, + 0.906, + 0.637 + ], + "angle": 0, + "content": "[20] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening, 2000. US Patent 6,011,875. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.638, + 0.906, + 0.704 + ], + "angle": 0, + "content": "[21] Pengfei Liu, Liang Xiao, and Tao Li. A variational pan-sharpening method based on spatial fractional-order geometry and spectral-spatial low-rank priors. IEEE Transactions on Geoscience and Remote Sensing, 56(3):1788-1802, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.906, + 0.775 + ], + "angle": 0, + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.817 + ], + "angle": 0, + "content": "[23] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. Remote Sensing, 8(7):594, 2016. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[24] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE Geoscience and Remote Sensing Magazine, 9(1):18-52, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10916" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[25] Jorge Nunez, Xavier Otazu, Octavi Fors, Albert Prades, Vicenc Pala, and Roman Arbiol. Multiresolution-based image fusion with additive wavelet decomposition. IEEE Transactions on Geoscience and Remote sensing, 37(3):1204-1211, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[26] Titouan Parcollet, Mirco Ravanelli, Mohamed Morchid, Georges Linares, Chiheb Trabelsi, Renato De Mori, and Yoshua Bengio. Quaternion recurrent neural networks. arXiv preprint arXiv:1806.04418, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[27] Titouan Parcollet, Ying Zhang, Mohamed Morchid, Chiheb Trabelsi, Georges Linares, Renato De Mori, and Yoshua Bengio. *Quaternion convolutional neural networks for end-to-end automatic speech recognition*. arXiv preprint arXiv:1806.07789, 2018. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[28] Soo-Chang Pei, Jian-Jiun Ding, and Ja-Han Chang. Efficient implementation of quaternion fourier transform, convolution, and correlation by 2-d complex FFT. IEEE transactions on signal processing, 49(11):2783-2797, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[29] Thierry Ranchin and Lucien Wald. Fusion of high spatial and spectral resolution images: The arsis concept and its implementation. Photogrammetric engineering and remote sensing, 66(1):49-61, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Stephen John Sangwine. Fourier transforms of colour images using quaternion or hypercomplex, numbers. *Electronics letters*, 32(21):1979–1980, 1996. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[31] Vijay P Shah, Nicolas H Younan, and Roger L King. An efficient pan-sharpening method via a combined adaptive pca approach and contourlets. IEEE transactions on geoscience and remote sensing, 46(5):1323-1335, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[32] Lilong Shi and Brian Funt. Quaternion color texture segmentation. Computer Vision and image understanding, 107(1-2): 88-96, 2007. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[33] Jiangtong Tan, Jie Huang, Naishan Zheng, Man Zhou, Keyu Yan, Danfeng Hong, and Feng Zhao. Revisiting spatial-frequency information integration from a hierarchical perspective for panchromatic and multi-spectral image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25922-25931, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[34] Te-Ming Tu, Shun-Chi Su, Hsuen-Chyun Shyu, and Ping S Huang. A new look at ihs-like image fusion methods. Information fusion, 2(3):177-186, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[35] Gemine Vivone, Rocco Restaino, Mauro Dalla Mura, Giorgio Licciardi, and Jocelyn Chanussot. Contrast and error-based fusion schemes for multispectral image pansharpening. IEEE Geoscience and Remote Sensing Letters, 11(5): 930-934, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[36] Dong Wang, Yunpeng Bai, and Ying Li. Multispectral pan-sharpening via dual-channel convolutional network with convolutional LSTM based hierarchical spatial-spectral feature fusion, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[37] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Chen Wu, Bo Du, Xiaohui Cui, and Liangpei Zhang. A post-classification change detection method based on itera" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tive slow feature analysis and bayesian soft fusion. Remote Sensing of Environment, 199:241-255, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.188 + ], + "angle": 0, + "content": "[39] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[40] Gang Yang, Xiangyong Cao, Wenzhe Xiao, Man Zhou, Aiping Liu, Xun Chen, and Deyu Meng. Panflownet: A flow-based deep network for pan-sharpening. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16857-16867, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[41] Licheng Yu, Yi Xu, Hongteng Xu, and Hao Zhang. Quaternion-based sparse representation of color image. In 2013 IEEE International Conference on Multimedia and Expo (ICME), pages 1-7. IEEE, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[42] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[43] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[44] Rui Zeng, Jiasong Wu, Zhuhong Shao, Yang Chen, Beijing Chen, Lotfi Senhadji, and Huazhong Shu. Color image classification via quaternion principal component analysis network. Neurocomputing, 216:416-428, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[45] Hao Zhang and Jiayi Ma. Gtp-pnet: A residual learning network based on gradient transformation prior for pansharpening. ISPRS Journal of Photogrammetry and Remote Sensing, 172:223-239, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[46] Yunzuo Zhang, Jiayu Zhang, Ruixue Liu, Pengfei Zhu, and Yameng Liu. Key frame extraction based on quaternion fourier transform with multiple features fusion. Expert Systems with Applications, 216:119467, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.69 + ], + "angle": 0, + "content": "[47] Man Zhou, Jie Huang, Yanchi Fang, Xueyang Fu, and Aiping Liu. Pan-sharpening with customized transformer and invertible neural network. In Proceedings of the AAAI conference on artificial intelligence, pages 3553-3561, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[48] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In European conference on computer vision, pages 274-291. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[49] Man Zhou, Keyu Yan, Jie Huang, Zihe Yang, Xueyang Fu, and Feng Zhao. Mutual information-driven pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1798-1808, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[50] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. FOurmer: An efficient global modeling paradigm for image restoration. In International conference on machine learning, pages 42589-42601. PMLR, 2023. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[51] Xuanyu Zhu, Yi Xu, Hongteng Xu, and Changjian Chen. Quaternion convolutional neural networks. In Proceedings of" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10917" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "the European conference on computer vision (ECCV), pages 631-647, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[52] Zeyu Zhu, Xiangyong Cao, Man Zhou, Junhao Huang, and Deyu Meng. Probability-based global cross-modal upsampling for pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14039-14048, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10918" + } + ] +] \ No newline at end of file diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_origin.pdf b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4db7e05bb7300a4dfcd56457ee8419bc172fa6bf --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/ead32439-3d1c-41ba-b332-4dc5bcc158be_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f7db9723ef6afc53373681808e66c647a159b62a7afc12166d212adca1d858b0 +size 1356065 diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/full.md b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1ee0193ebf2dfa401151f927891db669f860b170 --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/full.md @@ -0,0 +1,367 @@ +# Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions + +Dong Li, Chunhui Luo, Yuanfei Bao, Gang Yang, Jie Xiao, Xueyang Fu, Zheng-Jun Zha* +University of Science and Technology of China, Hefei, China + +dongli6@mail.ustc.edu.cn，zhazj@ustc.edu.cn + +# Abstract + +Pansharpening aims to generate high-resolution multispectral (MS) images by fusing panchromatic (PAN) images with corresponding low-resolution MS images. However, many existing methods struggle to fully capture spatial and spectral interactions, limiting their effectiveness. To address this, we propose a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging the compact representation capabilities of quaternions for high-dimensional data. Our method consists of three key components: the quaternion global spectral interaction branch, the quaternion local spatial structure awareness branch, and the quaternion spatial-spectral interaction branch. The first applies the quaternion Fourier transform to convert multi-channel features into the frequency domain as a whole, enabling global information interaction while preserving inter-channel dependencies, which aids spectral fidelity. The second uses a customized spatial quaternion representation, combined with a window-shifting strategy, to maintain local spatial dependencies while promoting spatial interactions, which helps inject spatial details. The last integrates the two pathways within the quaternion framework to enrich spatial-spectral interactions for richer representations. By utilizing quaternion's multi-dimensional representation and parameter-sharing properties, our method achieves a more compact and efficient cross-resolution, multi-band information integration, significantly improving the quality of the fused image. Extensive experiments validate the proposed method's effectiveness and its superior performance over current SOTA techniques. The code is available at https://github.com/dongli8/QuatPanNet. + +# 1. Introduction + +In remote sensing, satellite sensors face physical limitations that prevent the direct acquisition of high-resolution multispectral (MS) images [45]. Pansharpening addresses + +![](images/022bd077a824460e477d738dedbcbbe71250b2914c9c9ad1f8d7f2d6b16ad737.jpg) +Figure 1. Comparison between standard real-valued layer and quaternion-valued layer. The real-valued operation typically stacks the input multi-channel features, while the quaternion-valued layer, due to quaternion weight sharing, has the ability to learn the latent relationships within the input features. + +this by merging low-resolution multispectral (LRMS) images with panchromatic (PAN) images to produce high-resolution multispectral (HRMS) images. As a crucial preprocessing step, pansharpening is widely applied across fields [38, 43]. Conceptually, it is a fusion super-resolution task, focused on incorporating fine spatial details while preserving spectral fidelity, requiring strong inter-channel dependency maintenance and the preservation of subtle local spatial structures. + +Traditional methods often rely on mathematical models that integrate spatial and spectral data, typically assuming that the pansharpened image is a linear combination of various spectral bands from a high-resolution multispectral image [40]. However, these approaches are limited by their dependence on prior knowledge, which restricts flexibility and adaptability. With the rise of deep learning, convolutional neural networks (CNNs) have been introduced into pansharpening, leading to substantial breakthroughs in performance with increased model complexity [1, 5, 9, 23]. + +Despite notable successes, existing methods still struggle to capture the intricate spatial-spectral interactions essential + +for high-quality pansharpening. These limitations hinder simultaneous feature interaction and inter-feature dependency preservation, leading to compromises in both spectral fidelity and spatial detail. Unlike conventional real-valued networks, quaternion networks treat multidimensional features as unified entities, allowing them to effectively encode internal dependencies [26, 27]. Additionally, quaternion networks demonstrate strong modeling capabilities for both inter- and intra-correlations within the quaternion latent space [3]. Thus, we aim to harness quaternion operations to further elevate pansharpening performance. + +Our Motivation. PAN images provide high-resolution spatial details, while MS images offer rich multispectral information [11, 24]. A successful pansharpening model must preserve both spatial details and spectral fidelity during fusion. This requires capturing the unique features of PAN and MS images while leveraging their complementary information to produce accurate and comprehensive fused outputs. Traditional approaches often handle multi-channel inputs by stacking and mapping them to a single output channel, which causes information loss. Consequently, they struggle to capture inter-channel correlations and interactions between spatial and spectral data effectively. To compensate, some methods increase the number of output channels to improve representational capacity [39], but this also increases model complexity. Furthermore, these methods typically extract features in real-number space, limiting their ability to capture directional information such as edges and structures. In contrast, quaternions represent multiple spectral channels as distinct components of a compact quaternion representation, as shown in Figure 1. This approach integrates multi-band information efficiently while reducing information loss through parameter-sharing quaternion operations. These features make quaternions particularly well-suited for pansharpening tasks that involve multi-source data inputs. + +Based on the analysis above, we propose a novel quaternion-based pansharpening framework called QuatPanNet. QuatPanNet is designed to balance spectral fidelity and spatial detail richness by leveraging quaternion-based representations and operations to enhance interactions among multispectral, spatial structure, and spatial-spectral features. The framework consists of three core components, i.e., Quaternion Global Spectral Interaction Branch,Quaternion Local Spatial Structure Awareness Branch, andQuaternion Spatial-Spectral Interaction Branch. The first one uses the Quaternion Fourier Transform (QFT) to process multi-channel features holistically in the frequency domain. By enabling global interactions while preserving inter-channel dependencies, it effectively extracts global spectral representations, which are critical for maintaining spectral fidelity. The second one employs a customized spatial quaternion transformation method and a window + +shifting strategy, this branch focuses on preserving local structural dependencies and enhancing spatial interactions. It ensures that spatial details are well-integrated while maintaining the structural integrity of the input data. The last one, which built on the outputs of the first two branches, integrates dual-path features into a unified quaternion representation. This integration enhances the complementarity and coherence of information during the spatial-spectral fusion process. By utilizing the compact and efficient representation of PAN and MS image data in quaternion space, QuatPanNet facilitates effective interactions among spatial and spectral features. This approach reduces parameter complexity while achieving high spectral fidelity and rich spatial resolution, leading to superior fusion quality. In summary, the contributions of this work are as follows: + +- We propose a novel quaternion-based pansharpening framework, QuatPanNet, designed to comprehensively harness quaternions for this task. +- We design customized quaternion-based spatial and spectral interaction modules for pansharpening. The former facilitates global information interaction while preserving inter-spectral dependencies, and the latter promotes spatial interaction while maintaining local structural dependencies. +- We introduce a quaternion-based spatial-spectral interaction branch, leveraging the high-dimensional representational power of quaternions to improve the fusion of spectral and spatial features. + +Extensive experiments on multiple satellite datasets demonstrate that our method achieves superior qualitative and quantitative results with fewer parameters and effectively generalizes to full-resolution scenes. + +# 2. Related Works + +# 2.1. Pansharpening + +Existing pansharpening methods are divided into traditional ones and deep - learning - based ones. Traditional ones mainly consist of Component Substitution (CS) [31, 34], Multi - Resolution Analysis (MRA) [25, 29, 35], and Variational Optimization (VO) [8, 15, 21]. However, their heavy dependence on manually designed features and prior knowledge restricts their applicability [37]. PNN [23] is the first to introduce CNN into the field of pansharpening, achieving better results than traditional methods. Since then, deep learning-based approaches have dominated this field. MSDCNN [42] combines multi-scale feature extraction and residual learning, enhancing the quality of the fused images. INNformer [47], based on the Transformer architecture, integrates an invertible neural network module for lossless information to model long-range dependencies and efficient feature fusion, successfully introducing the Transformer architecture into this domain. The fusion of frequency do + +main features in images is also a key focus for researchers [36, 48]. For instance, FAME-Net [16] proposes a method that combines MOE with frequency domain information, enabling the network to dynamically learn high-frequency information in remote sensing images. HFIN [33] explores the relationship between PAN and LRMS images using local Fourier information integration. In addition to designing pansharpening models, Zhu et al. developed the first upsampling module specifically for pansharpening methods, called PGCU [52]. + +# 2.2.Quaternion + +Quaternions are the natural extension of complex numbers to four dimensions, proposed by Hamilton in 1843 [13]. A quaternion consists of four components, including one real part and three imaginary parts. Quaternions, with their four components, are particularly well-suited for three-dimensional and four-dimensional feature vectors, as seen in applications such as image processing and robotic kinematics [2, 30]. In 2001, Pei et al. [28] introduced the quaternion model for color images, marking the beginning of quaternion applications in the field of image processing. The advantages of quaternion wavelet transforms [18], quaternion principal component analysis [44], and other quaternion-based color image processing techniques [41] have been demonstrated to extract more representative features for color images. These techniques have achieved promising results in advanced visual tasks such as color image classification. With the wave of deep learning, many researchers have begun to explore the application of quaternions in the field of deep learning [10, 26]. Xu et al. [51] proposes QCNN, which combines quaternions with CNNs, achieving faster convergence of the loss function while preserving color information more effectively than real-valued CNNs. Kusamichi et al. [19] studies color night vision and finds that using quaternion networks yields better results in extracting color information from dim images. + +# 3. Method + +In this section, we first introduce the relevant properties of quaternions and their operators, and then present the proposed quaternion pan-sharpening network, as shown in Figures 2 and 3. We then provide a detailed explanation of the fundamental components of our method, including three quaternion-based branches: a) the global interaction branch utilizes the Quaternion Fourier Transform and quaternion convolution to maintain channel interdependencies while representing global information; b) the local structure-aware branch models local spatial dependencies through a specially designed spatial quaternion transformation method combined with a window-shifting strategy; and c) the dual-domain interaction branch leverages the orthogonal relationships inherent in quaternions to model + +dual-domain correlations and promote complementary feature learning. Finally, we present the optimization strategy employed in our approach. + +# 3.1.Quaternion Operators + +Quaternion, proposed by Hamilton in 1843, extends complex numbers to a four-dimensional space. A quaternion $Q \in \mathbb{H}$ can be expressed as follows [32]: + +$$ +Q = w 1 + x i + y j + z k, \tag {1} +$$ + +where $w, x, y, z$ are all real numbers, and $1, i, j, k$ are the quaternion unit basis. $w1$ refers to the real component of the number, and $xi + yj + zk$ to the imaginary component. In the definition of quaternions, they must satisfy the following identities, known as the Hamilton rules: + +$$ +i ^ {2} = j ^ {2} = k ^ {2} = i j k = - 1. \tag {2} +$$ + +We propose the use of quaternions and their associated operators to uncover both intra- and inter-relations between the features of high spatial resolution PAN and MS images. This approach is inspired by the unique feature-handling capabilities inherent in quaternion representation [3]. Consequently, we introduce two quaternion operators: quaternion convolution and the quaternion Fourier transform. + +# Quaternion Convolution + +In the real-valued space, convolution involves convolving a filter matrix with a vector. Extending this operation to quaternions, quaternion convolution is achieved by convolving a quaternion filter matrix with a quaternion vector. To maintain consistency with the defined quaternions, we represent the computation of quaternion convolution using real-valued matrices. This computation is performed by convolving the quaternion filter matrix $W = A + Bi + Cj + Dk$ with the quaternion vector $Q = w + xi + yj + zk$ , where $A, B, C$ , and $D$ are real-valued matrices. The convolution operation between $W$ and $Q$ is denoted as: + +$$ +\begin{array}{l} W * Q = (A * w - B * x - C * y - D * z) + \\ i (A * x + B * w + C * z - D * y) + \\ j (A * y - B * z + C * w + D * x) + \tag {3} \\ K (A * z + B * y - C * x + D * w). \\ \end{array} +$$ + +It can be seen that the result of quaternion convolution produces a unique linear combination of each coordinate axis for every single-axis result. This stems from the structure of quaternion multiplication and enforces interactions between each axis of the kernel and each axis of the image. This contrasts with real-valued convolution, which simply multiplies each channel of the kernel with the corresponding channel of the image. The advantage of this approach is that quaternion convolution promotes cross-axis interactions and information extraction through the reuse of filters across each axis and combination [27]. + +![](images/4360aa7e7cba208c962cf624fc35b89f8447ca4e5995689294eef91955104070.jpg) +Figure 2. The framework of our proposed pan-sharpening network QuatPanNet. It consists of several Quaternion Spatial-Spectral Interaction Blocks (QSSIB). The Quaternion Spatial-Spectral Interaction Block is shown in Figure 3. + +# Quaternion Fourier Transform + +The Fourier Transform plays a crucial role in various image processing tasks, including pansharpening [50]. After decomposing the input image into multiple channels, the Discrete Fourier Transform (DFT) can be applied to each channel individually. However, this approach neglects the mutual information among different channels, which can hinder the performance of pansharpening. The Discrete Quaternion Fourier Transform (DQFT) offers the possibility of simultaneously transforming multi-channel images, as it can transform quaternions holistically [4]. + +Due to the non-commutative nature of quaternion multiplication, there are various types of Quaternion Fourier Transforms (QFT), such as left-side, right-side and two-side QFT [6]. In this paper, we utilize a typical QFT, specifically the left-sided QFT. For an image $\dot{f}$ of size $H\times W$ , the Discrete Quaternion Fourier Transform is defined as follows. + +$$ +\mathcal {F} (\dot {f}) (u, v) = \frac {1}{\sqrt {H W}} \sum_ {x = 0} ^ {H - 1} \sum_ {y = 0} ^ {W - 1} e ^ {- \dot {\mu} 2 \pi \left(\frac {x}{H} u + \frac {y}{H} v\right)} \dot {f} (x, y), \tag {4} +$$ + +where $\dot{f} (x,y)$ is pixel of $\dot{f}$ and $\dot{\mu}$ is unit pure quaternion. $(u,v)$ is the coordinate in the quaternion frequency domain. $\mathcal{F}^{-1}(\dot{f})$ defines the inverse quaternion Fourier transform accordingly. Inspired by the traditional Fourier transform, the modulus of $\mathcal{F}(\dot{f})(u,v)$ is taken as the amplitude component, and the angles with the three imaginary axes are taken as the phase components. + +It is worth noting that the QFT is not a simple substitution of the imaginary unit in the two-dimensional Fourier Transform with quaternions; rather, its basis functions possess orthogonal imaginary components, which preserve the + +information and mutual relationships among different components in the frequency domain. Consequently, some studies leverage QFT to treat multiple features as a unified whole to enhance performance [4, 46]. Furthermore, the efficiency of QFT is also noteworthy. For example, when comparing the use of QFT with three separate conventional Fourier Transforms for processing color images, the number of real multiplications and additions required by QFT is approximately $75\%$ of that needed for the three separate transforms [30]. For pansharpening, we propose utilizing the QFT to treat multispectral data as a unified whole, facilitating global information interactions in the frequency domain while preserving inter-channel dependencies. + +# 3.2. Framework + +Based on the above analysis, we propose a novel pansharpening approach based on quaternion information interaction, as illustrated in Figure 2. Given a panchromatic (PAN) image $P \in \mathbb{R}^{H \times W \times 1}$ and a low-resolution multispectral (LRMS) image $L \in \mathbb{R}^{H / r \times W / r \times C}$ , the network first uses convolution to project the bicubic upsampled r-times $L$ into a shallow feature representation. Simultaneously, $P$ is processed through multiple cascaded convolution layers to extract effective information for high-resolution multispectral (HRMS) reconstruction. Then, the features of the PAN and LRMS images are jointly processed through the core building module, the quaternion spatial-spectral interaction block (QSSIB), for continuous global information interaction and local structure-aware integration. Next, a convolution layer transforms all features gathered by the N QSSIBs back into the image space, and finally, these features are combined with the residual to obtain the final output. + +![](images/5cbe67bf3e36207152cfbadca3c2bf3991aa580420568ebb7b7d872a29ba923a.jpg) +Figure 3. The detailed flowchart of the proposed core building module quaternion spatial-spectral interaction block (QSSIB), consisting of three components: quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch (QLSA) and quaternion spatial-spectral interaction branch (QSSI). + +# 3.3. Key Building Components + +As shown in Figure 3, the fundamental building block of our method comprises three key elements: the quaternion global spectral interaction branch (QGSI), the quaternion local spatial structure awareness branch (QLSA), and the quaternion spatial-spectral interaction branch (QSSI). To convert images into quaternion representations, we design customized methods, including channel quaternion representation (CQR), spatial quaternion representation (SQR), and shifted window quaternion representation (SWQR). When transforming quaternion representations back into image features, we denote the corresponding inverse operations as ICQR, ISQR, and ISWQR, respectively. (For more details on quaternion representation, refer to the supplementary material.) These three key components, in conjunction with the quaternion representations of the images, jointly facilitate channel interactions during multispectral fusion and local interactions for spatial resolution enhancement. + +# Quaternion Global Spectral Interaction + +The global interaction branch utilizes the Quaternion Fourier Transform (QFT) to transform multi-channel features as a whole into the frequency domain, facilitating global information interactions. Since the input to QFT must be in quaternion form and we aim to maintain interchannel interdependencies, we design a channel quaternion representation (CQR). This approach divides the image features equally along the channel dimension into four parts, which are then used as the four components of a quaternion. Let the features of the MS and PAN images be denoted as $F_{ms}$ and $F_{p}$ , respectively; their channel-dimensional quaternion representations are denoted as $Q_{ms}^{c}$ and $Q_{p}^{c}$ . We + +then use QFT to transform the MS and PAN features into the frequency domain. Some studies on image restoration indicate that decomposing amplitude and phase components facilitates the decoupling of degradation [17, 50]. Therefore, we decompose the frequency domain features into amplitude and phase components, with the corresponding process represented as + +$$ +\mathcal {A} \left(Q _ {p} ^ {c}\right), \mathcal {P} _ {1} \left(Q _ {p} ^ {c}\right), \mathcal {P} _ {2} \left(Q _ {p} ^ {c}\right), \mathcal {P} _ {3} \left(Q _ {p} ^ {c}\right) = \mathcal {F} \left(Q _ {p} ^ {c}\right), \tag {5} +$$ + +$$ +\mathcal {A} \left(Q _ {m s} ^ {c}\right), \mathcal {P} _ {1} \left(Q _ {m s} ^ {c}\right), \mathcal {P} _ {2} \left(Q _ {m s} ^ {c}\right), \mathcal {P} _ {3} \left(Q _ {m s} ^ {c}\right) = \mathcal {F} \left(Q _ {m s} ^ {c}\right), \tag {6} +$$ + +where $\mathcal{A}(\cdot)$ represents the amplitude, and $\mathcal{P}_1(\cdot),\mathcal{P}_2(\cdot)$ and $\mathcal{P}_3(\cdot)$ represent the phases corresponding to the three imaginary axes. We then concatenate the amplitude and phase spectra of the PAN and MS images accordingly to form new quaternions. Next, we process the newly formed quaternions using quaternion convolution, thereby enhancing global representations in the frequency domain while promoting interactions between channels. + +$$ +\begin{array}{l} Q _ {f} ^ {c} = \mathcal {O Q} (\mathcal {C} [ \mathcal {A} (Q _ {p} ^ {c}), \mathcal {A} (Q _ {m s} ^ {c}) ], \mathcal {C} [ \mathcal {P} _ {1} (Q _ {p} ^ {c}), \mathcal {P} _ {1} (Q _ {m s} ^ {c}) ] \\ , \mathcal {C} \left[ \mathcal {P} _ {2} \left(Q _ {p} ^ {c}\right), \mathcal {P} _ {2} \left(Q _ {m s} ^ {c}\right) \right], \mathcal {C} \left[ \mathcal {P} _ {3} \left(Q _ {p} ^ {c}\right), \mathcal {P} _ {3} \left(Q _ {m s} ^ {c}\right) \right]) \tag {7} \\ \end{array} +$$ + +where $\mathcal{C}$ denotes concatenation along the channel dimension, $\mathcal{OQ}$ is quaternion convolution. The fused amplitude and phase components are the transformed back to the spatial domain using the inverse discrete quaternion Fourier transform. + +$$ +Q _ {f r e} ^ {c} = \mathcal {F} ^ {- 1} \left(Q _ {f} ^ {c}\right). \tag {8} +$$ + +By representing the image in a channel-dimensional quaternion form, the quaternion operator treats the multichannel data as a unified whole for information interaction. + +Additionally, according to spectral convolution theorem [7], processing information in the Fourier space enables capturing global frequency representations. In summary, this branch generates the global information representation with strengthened channel interactions $Q_{fre}^{c}$ . + +# Quaternion Local Spatial Structure Awareness + +Pansharpening is essentially a fusion-based superresolution task, making the injection of spatial resolution critically important. Enhancing spatial resolution requires not only global information from the frequency domain but also local information interactions in the spatial domain. For pansharpening, preserving local structures during information interactions is particularly significant, as it enhances the network's ability to recover fine details. + +We aim to leverage the balance of enhanced information interaction and specific information preservation offered by quaternions to improve the network's spatial superresolution capabilities. To this end, we design the spatial quaternion representation (SQR). It divides the image into several $2 \times 2$ windows in the spatial dimension, with each local window's four pixels serving as the four components of a quaternion, as illustrated in Figure 3. This approach explicitly ensures dependencies among local pixels, preserving the local structure. Based on this representation, we use quaternion convolutions (QCNN) to process the features. + +However, this representation method lacks crosswindow connections, which limits spatial interactions. Inspired by [22], we design the shifted window quaternion representation (SWQR) based on SQR. This allows the network to alternately employ two partitioning configurations, thereby enhancing the effectiveness of spatial quaternion representation. Through the combined effect of this method and quaternion convolution, this branch preserves local spatial dependencies while promoting spatial interactions, thereby obtaining the local representation $Q_{spa}^{s}$ . + +The information representations of the two branches are highly complementary, with one branch responsible for global, frequency-domain, and channel interactions, and the other focused on local, spatial-domain, and spatial interactions. Therefore, further interaction and integration between them facilitate mutual compensation. + +# Quaternion Spatial-Spectral Interaction + +The two branches described above respectively obtain quaternion representations in the channel dimension, $Q_{fre}^{c} \in \mathbb{H}^{H \times W \times C / 4}$ , and in the spatial dimension, $Q_{spa}^{s} \in \mathbb{H}^{H / 2 \times W / 2 \times C}$ . The quaternion dual-domain information interaction branch mainly consists of two parts: one is a quaternion-based global fusion, and the other is the channel-aware quaternion spatial feature injection. + +Quaternion-based global fusion merges the features obtained from the two branches into a single quaternion for unified processing, thereby enhancing the correlation between the two through the structure of the quaternion. First, + +both $Q_{fre}^{c}$ and $Q_{spa}^{s}$ are converted back into image feature representations. $Q_{fre}^{c}$ is combined along the channel dimension to obtain $F_{fre} \in \mathbb{R}^{H \times W \times C}$ , and $Q_{spa}^{s}$ is rearranged spatially to obtain $F_{spa} \in \mathbb{R}^{H \times W \times C}$ . To achieve a richer feature representation, we apply convolutions to $F_{fre}$ and $F_{spa}$ to obtain $F_{fre}'$ and $F_{spa}'$ , respectively. We then combine these four features into a quaternion $Q_{f} = F_{fre} + F_{spa}i + F_{fre}^{\prime}j + F_{spa}^{\prime}k$ . We use quaternion convolution to process $Q_{f} \in \mathbb{H}^{H \times W \times C}$ as a whole, producing $Q_{a} \in \mathbb{H}^{H \times W \times C/4}$ . + +Channel-aware quaternion spatial feature injection involves injecting spatial features $F_{spa}$ from the local structure-aware branch into the four components of the quaternion representation $Q_{fre}^{c}$ in the channel dimension, thereby facilitating information fusion in both spatial and channel domains. Specifically, we replicate the image feature $F_{spa}$ four times and concatenate each copy with one of the four components of the quaternion feature $Q_{fre}^{c}$ to obtain $Q_{i} \in \mathbb{H}^{H \times W \times 5C / 4}$ : + +$$ +Q _ {i} = \mathcal {C} \left[ Q _ {f r e} ^ {c}, \left(F _ {s p a} + F _ {s p a} i + F _ {s p a} j + F _ {s p a} k\right) \right]. \tag {9} +$$ + +We then feed $Q_{i}$ into quaternion convolution to obtain $Q_{b} \in \mathbb{H}^{H \times W \times C / 4}$ . Afterward, we concatenate $Q_{a}$ and $Q_{b}$ , and apply quaternion convolution for further fusion to obtain $Q_{out} \in \mathbb{H}^{H \times W \times C / 4}$ . + +$$ +Q _ {o u t} = \operatorname {Q C o n v} \left(\mathcal {C} \left[ Q _ {a}, Q _ {b} \right]\right), \tag {10} +$$ + +where QConv denotes quaternion convolution. We then convert $Q_{out}$ into the image feature representation $F_{out} \in \mathbb{R}^{H \times W \times C}$ . Finally, a residual learning mechanism is applied, adding the input feature $F_{ms}$ to the fused feature. The proposed network achieves effective fusion through the seamless integration of the two quaternion representations. + +# 3.4. Optimization + +Let $Y$ and $GT$ denote the network's output and the corresponding ground truth, respectively. We use a joint spatial-frequency domain loss to supervise the network training. In the frequency domain, we first apply the quaternion Fourier transform to convert $Y$ and $GT$ into the quaternion Fourier space. Then, we compute the L1 norm of the real and imaginary differences between $Y$ and $GT$ , and sum them to obtain the quaternion Fourier domain loss. + +$$ +\begin{array}{l} \mathcal {L} _ {f r e} = \| R (Y) - R (G T) \| _ {1} + \| I (Y) - I (G T) \| _ {1} \tag {11} \\ + \| J (Y) - J (G T) \| _ {1} + \| K (Y) - K (G T) \| _ {1}, \\ \end{array} +$$ + +where $R(\cdot), I(\cdot), J(\cdot)$ , and $K(\cdot)$ represent the real part and three imaginary parts of the quaternion respectively. In the spatial domain, we employ the L1 loss. Finally, the overall loss function is as follows + +$$ +\mathcal {L} = \| Y - G T \| _ {1} + \lambda \mathcal {L} _ {f r e}, \tag {12} +$$ + +where $\lambda$ is the weighting factor and set to 0.1 empirically. + +![](images/0d58f5574228d307d2321d3d5d8c388c6620057a5ecbbc892c7ebc15abcf844c.jpg) +LRMS + +![](images/38072c22dc00240d02d7e6ed65f037633e9d00ce7651e752f71d89ab1acad415.jpg) + +![](images/ffef7e3696608666d2933be36000362a1f6e0f6d9de01a923f225dc9f11d2cbb.jpg) +GS + +![](images/87ab71925aaaba337c145cbcb0cd9044d4ff2687e234c8d0a4d11107d43b92fc.jpg) + +![](images/5065df451e37e817546e64992c3cca13443bedf43a23c3d79998c594042cfa73.jpg) +MSDCNN + +![](images/99bc74a635a76c945eea1533db2d5ee4d6b23d610e6c2253dab76896a0d81de4.jpg) +GPPNN + +![](images/f4cce0d1da5c8e8b4cd038352abc5bbb4dd90c14de8800013673f3a5d31a8000.jpg) +MutNet + +![](images/a9d0c937fd2566ead0393ffd7b16b92077b50abf607409f9c1681efc351c03d5.jpg) +PAN +INNformer + +![](images/1919873bdbc6fd9d15f08e2e747dbc178aadc3b91e380fe5b7c656d0e4d293f7.jpg) +HFIN + +![](images/c8130270ab584a8b7065bfdc9df831381e9876bad7a81e34f5c9bba3822a9c0b.jpg) +PNN +FAME-Net + +![](images/f8b8c85f3af244ebc58911bfa5dc2e7da3b3b25155bae08f62b295c0d8814d5a.jpg) +Ours + +![](images/b9059ccf5e428aa1f113a6c11bf7e7bd144a5f8fa9c3a879f0f14acb7af998e6.jpg) +GT + +![](images/ca471fb721fb0893d11f23de893307781a0fbb459e18c1b16884346cc7ad1384.jpg) +GS +Figure 4. The visual comparisons between other pan-sharpening methods and our method on WorldView-II satellite + +![](images/ad31d5f2834d3ce6e8ac79aaaae08a8e04c991bde3c783f9875b8ec8b543bcd0.jpg) +PNN + +![](images/13399ce9c97b7e0b3fdc734b48adfd4f3f58d92e79e0226e605edd676d45a5d6.jpg) +MSDCNN + +![](images/9f34b258e8df8c4e0ecb0cbc917a61f783996d8baac216bc9908eb89bbafb680.jpg) +GPPNN + +![](images/463b8bb99ce9834c6f14903c20eacbabbc5ab291dda3178919fbecd78d0eaa50.jpg) +MutNet + +![](images/c09ebd78fae13c0505879aee7ed21b324bd273d2efbb21a84c992d4a2537b410.jpg) +INNformer + +![](images/e79a805be7f28a2d1cce5c85af39f15ab3f14ea6854a027cc99286d8283968af.jpg) +HFIN + +![](images/9310c22fa58de36945cb88cde5aa81f9c634aeb7803e8cd8ea8841e6303ad23b.jpg) +FAME-Net + +![](images/0610e756f4fc676d5c1e1c351a1e8801b131cf8cd72021c3f28973e6435a2457.jpg) +Ours + +![](images/5526a513eba49cfd46396bac07528bedb800c2e7411d2d5547b33a5ad4d030f8.jpg) + +Table 1. Experimental results of all the competing methods on the three benchmark datasets. The best and the second best values are highlighted in bold and underline, respectively. + +
MethodsParams (M)WorldView IIWorldView IIIGaoFen2
PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓
Brovey-35.86460.92160.04031.823822.50600.54660.11598.233137.79740.90260.02181.3720
IHS-35.29620.90270.04612.027822.55790.53540.12668.361638.17540.91000.02431.5336
GS-35.63760.91760.04231.877422.56080.54700.12178.243337.22600.90340.03091.6736
PNN0.068940.75500.96240.02591.064629.94180.91210.08243.320643.12080.97040.01720.8528
MSDCNN0.239041.33550.96640.02420.99430.30380.91840.07823.188445.68740.98270.01350.6389
GPPNN0.119841.16220.96840.02441.031530.17850.91750.07763.259344.21450.98150.01370.7361
MutNet0.071441.67730.97050.02240.951930.49070.92230.07493.112547.30420.98920.01020.5481
INNformer0.070641.69030.97040.02270.951430.53650.92250.07473.099747.35280.98930.01020.5479
HFIN0.077242.23190.97140.02150.880730.61470.92030.07423.078648.87830.98980.00930.4591
FAME-Net0.140842.02620.97230.02150.917230.99030.92870.06972.953147.67210.98980.00980.5242
Ours0.038742.48460.97380.02080.855431.17660.93320.05212.576149.14230.99010.00900.4474
+ +# 4. Experiments + +# 4.1. Dataset and benchmarks + +We conduct our experiments on three classic datasets. WorldView-II (WV2), Gaofen2 (GF2) and WorldView-III (WV3). The WorldView-II dataset consists of 760 image pairs for training and 80 image pairs for testing. The WorldView-III dataset includes 2150 image pairs for training and 200 image pairs for testing. The GaoFen-2 dataset comprises 2712 image pairs for training and 200 image pairs for testing. To validate the effectiveness of our approach, we compare it with several state-of-the-art pansharpening methods, including PNN [23], MSDCNN [42], GPPNN [39], MutNet [49], INNformer [47], HFIN [33] and FAME-Net [16], as well as traditional pansharpening meth + +ods such as GS [20], Brovey [12], and IHS [14]. + +In our experiments, all deep learning models are implemented using PyTorch and trained on an NVIDIA GeForce GTX 3090 GPU. For each set, the multispectral (MS) images are cropped into patches with the size of $32 \times 32$ , and the corresponding panchromatic (PAN) images are of size $128 \times 128$ . We employ common evaluation metrics, including PSNR, SSIM, SAM, and ERGAS. Additionally, we utilize three widely-used no-reference IQA metrics for real-world full-resolution scenes: $D_{\lambda}, D_{S}$ and QNR. More descriptions of the implementation details and dataset are provided in the supplementary material. + +Table 2. Non-reference metrics on full-resolution dataset. + +
PNNMSDCNNGPPNNINNformerMutNetHFINFAME-NetOurs
Dλ↓0.07460.07340.07820.06970.06940.07100.06740.0566
Ds↓0.11640.11510.12530.11280.11180.10980.11210.1055
QNRT0.81910.82150.80730.82530.82470.82610.82910.8447
+ +# 4.2. Comparison with state-of-the-art methods + +Evaluation on reduced-resolution scene. To evaluate the effectiveness of our method, we conducted experiments on three datasets and compared it with ten other state-of-the-art pansharpening methods. Table 1 summarizes the evaluation metrics for the three datasets, where the best result for each metric is highlighted in red and the second-best result in blue. The results demonstrate that our method significantly outperforms the others across all evaluation metrics, achieving the best performance on all three datasets. Compared to the second-best method in terms of PSNR for each dataset, our method shows an improvement of $0.25\mathrm{dB}$ on the WV2 dataset, $0.18\mathrm{dB}$ on the GF2 dataset, and $0.27\mathrm{dB}$ on the WV3 dataset, indicating the consistency of our method with the true high-resolution images. We also conducted qualitative experiments, Figures 4 present samples from the WV2 datasets. The top two rows show the comparison of experimental results, while the last row displays the mean squared error maps of each network's results compared to the ground truth. It is evident that the results obtained by our network are closest to the ground truth, with the smallest discrepancies. + +Evaluation on full-resolution scene. We also conduct full-resolution analysis in real-world scenarios to validate the generalization capability of our method. We perform experiments on an additional 200 sets of the GF2 dataset. Due to the absence of high-resolution multispectral images for actual real-world scenarios, we employ three commonly used no-reference metrics, $D_{\lambda}$ , $D_{S}$ , and QNR for evaluation. The experimental results, as shown in Table 2, demonstrate that our method achieves the best performance across all three metrics. + +Table 3. The results of different configurations on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively. + +
ConfigurationQGSIQLSAQSSIPSNR ↑SSIM ↑SAM ↓ERGAS ↓
IX✓✓41.73560.96730.02470.9337
II✓X✓42.02170.97060.02190.9134
III✓✓X42.17340.97130.02210.9012
Ours✓✓✓42.48460.97380.02080.8554
+ +# 4.3. Ablation Experiments + +We perform an ablation study using the WorldView-II dataset to further demonstrate the effectiveness of our method. The quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch + +Table 4. PSNR values of Our model with different number of blocks on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively. + +
Stages (K)PSNR ↑SSIM ↑SAM ↓ERGAS ↓
141.83780.96030.02490.9879
241.90020.96470.02380.9725
342.09130.96980.02340.9578
442.23690.97240.02150.9007
542.48460.97380.02080.8554
+ +(QLSA), and quaternion spatial-spectral interaction branch (QSSI) are the core components of the proposed method, and we conduct separate ablations for each. Additionally, we also explore the impact of the number of basic building blocks used in the network through ablation studies. + +To validate the effectiveness of the three key quaternion components, QGSI, QLSA, and QSSI, we replace their quaternion operations with real-valued operations, as shown in Table 3. The results demonstrate that removing each quaternion component leads to a decline in performance, thus confirming the effectiveness of the quaternion's multidimensional representation and parameter-sharing mechanism for pansharpening. + +We conduct experiments on networks with varying numbers of core building blocks to evaluate the impact of the number of core modules, as shown in Table 4. As observed, the model performance improves with the increase in the number of Quaternion Spatial-Spectral Interaction Blocks (QSSIB), though this comes at the cost of increased resource consumption. We balance performance and resource consumption by using the default setting of 5. More ablation experiments are shown in the supplementary materials. + +# 5. Conclusion + +In this paper, we propose QuatPanNet, a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging quaternions' compact representation for high-dimensional data. Our work demonstrates the advantages of a unified quaternion framework for effectively addressing the pansharpening problem. We custom-design quaternion-based spectral and spatial interaction modules: the former promotes global information interaction while preserving spectral dependencies, and the latter fosters spatial interaction while maintaining local structural dependencies. Building on these, we integrate the dual-path features into a unified quaternion representation. By utilizing the compact and efficient representation of PAN and MS image data in the quaternion space, QuatPanNet facilitates effective interactions between spatial and spectral features. Our method achieves exceptional fusion quality with a reduced number of parameters. + +# Acknowledgments + +This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243. + +# References + +[1] Paolo Addesso, Gemine Vivone, Rocco Restaino, and Jocelyn Chanussot. A data-driven model-based regression applied to panchromatic sharpening. IEEE Transactions on Image Processing, 29:7779-7794, 2020. 1 +[2] Nicholas A Asparagus and John K Dimitros. A comparative study of three methods for robot kinematics. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(2):135-145, 1998. 3 +[3] Qinglong Cao, Zhengqin Xu, Yuntian Chen, Chao Ma, and Xiaokang Yang. Domain prompt learning with quaternion networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26637-26646, 2024. 2, 3 +[4] Ping Cui, Cui-Ming Zou, Bang-Ju Wang, and Feng Shi. Color image fusion in quaternion spectral domain. In 2023 International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR), pages 56-61. IEEE, 2023. 4 +[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1 +[6] Todd A Ell, Nicolas Le Bihan, and Stephen J Sangwine. *Quaternion Fourier transforms for signal and image processing*. John Wiley & Sons, 2014. 4 +[7] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), pages 1381-1384. IEEE, 1998. 6 +[8] Xueyang Fu, Zihuang Lin, Yue Huang, and Xinghao Ding. A variational pan-sharpening with local gradient constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10265-10274, 2019. 2 +[9] Xueyang Fu, Wu Wang, Yue Huang, Xinghao Ding, and John Paisley. Deep multiscale detail networks for multiband spectral image sharpening. IEEE Transactions on Neural Networks and Learning Systems, 32(5):2090-2104, 2020. 1 +[10] Chase J Gaudet and Anthony S Maida. Deep quaternion networks. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2018. 3 +[11] Pedram Ghamisi, Behnood Rasti, Naoto Yokoya, Qunming Wang, Bernhard Hofle, Lorenzo Bruzzone, Francesca Bovolo, Mingmin Chi, Katharina Anders, Richard Gloaguen, et al. Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine, 7(1):6-39, 2019. 2 +[12] Alan R Gillespie, Anne B Kahle, and Richard E Walker. Color enhancement of highly correlated images. ii. channel + +ratio and "chromaticity" transformation techniques. Remote Sensing of Environment, 22(3):343-365, 1987. 7 +[13] William Rowan Hamilton. Lectures on Quaternions: Containing a Systematic Statement of a New Mathematical Method; of which the Principles Were Communicated in 1843 to the Royal Irish Academy; and which Has Since Formed the Subject of Successive Courses of Lectures, Delivered in 1848 and Subsequent Years, in the Halls of Trinity College, Dublin: with Numerous Illustrative Diagrams, and with Some Geometrical and Physical Applications. Hodges and Smith, 1853. 3 +[14] R Haydn. Application of the ihs color transform to the processing of multisensor data and image enhancement. In Proc. of the International Symposium on Remote Sensing of Arid and Semi-Arid Lands, Cairo, Egypt, 1982, 1982. 7 +[15] Xiyan He, Laurent Condat, José M Bioucas-Dias, Jocelyn Chanussot, and Junshi Xia. A new pansharpening method based on spatial and spectral sparsity priors. IEEE Transactions on Image Processing, 23(9):4160-4174, 2014. 2 +[16] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2121-2129, 2024. 3, 7 +[17] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 5 +[18] Creed F Jones III. Color face recognition using quaternionic Gabor filters. PhD thesis, Virginia Polytechnic Institute and State University, 2005. 3 +[19] Hiromi Kusamichi, Teijiro Isokawa, Nobuyuki Matsui, Yuzo Ogawa, and Kazuaki Maeda. A new scheme for color night vision by quaternion neural network. In Proceedings of the 2nd international conference on autonomous robots and agents. CiteSeer, 2004. 3 +[20] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening, 2000. US Patent 6,011,875. 7 +[21] Pengfei Liu, Liang Xiao, and Tao Li. A variational pan-sharpening method based on spatial fractional-order geometry and spectral-spatial low-rank priors. IEEE Transactions on Geoscience and Remote Sensing, 56(3):1788-1802, 2017. 2 +[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6 +[23] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. Remote Sensing, 8(7):594, 2016. 1, 2, 7 +[24] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE Geoscience and Remote Sensing Magazine, 9(1):18-52, 2020. 2 + +[25] Jorge Nunez, Xavier Otazu, Octavi Fors, Albert Prades, Vicenc Pala, and Roman Arbiol. Multiresolution-based image fusion with additive wavelet decomposition. IEEE Transactions on Geoscience and Remote sensing, 37(3):1204-1211, 1999. 2 +[26] Titouan Parcollet, Mirco Ravanelli, Mohamed Morchid, Georges Linares, Chiheb Trabelsi, Renato De Mori, and Yoshua Bengio. Quaternion recurrent neural networks. arXiv preprint arXiv:1806.04418, 2018. 2, 3 +[27] Titouan Parcollet, Ying Zhang, Mohamed Morchid, Chiheb Trabelsi, Georges Linares, Renato De Mori, and Yoshua Bengio. *Quaternion convolutional neural networks for end-to-end automatic speech recognition*. arXiv preprint arXiv:1806.07789, 2018. 2, 3 +[28] Soo-Chang Pei, Jian-Jiun Ding, and Ja-Han Chang. Efficient implementation of quaternion fourier transform, convolution, and correlation by 2-d complex FFT. IEEE transactions on signal processing, 49(11):2783-2797, 2001. 3 +[29] Thierry Ranchin and Lucien Wald. Fusion of high spatial and spectral resolution images: The arsis concept and its implementation. Photogrammetric engineering and remote sensing, 66(1):49-61, 2000. 2 +[30] Stephen John Sangwine. Fourier transforms of colour images using quaternion or hypercomplex, numbers. *Electronics letters*, 32(21):1979–1980, 1996. 3, 4 +[31] Vijay P Shah, Nicolas H Younan, and Roger L King. An efficient pan-sharpening method via a combined adaptive pca approach and contourlets. IEEE transactions on geoscience and remote sensing, 46(5):1323-1335, 2008. 2 +[32] Lilong Shi and Brian Funt. Quaternion color texture segmentation. Computer Vision and image understanding, 107(1-2): 88-96, 2007. 3 +[33] Jiangtong Tan, Jie Huang, Naishan Zheng, Man Zhou, Keyu Yan, Danfeng Hong, and Feng Zhao. Revisiting spatial-frequency information integration from a hierarchical perspective for panchromatic and multi-spectral image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25922-25931, 2024. 3, 7 +[34] Te-Ming Tu, Shun-Chi Su, Hsuen-Chyun Shyu, and Ping S Huang. A new look at ihs-like image fusion methods. Information fusion, 2(3):177-186, 2001. 2 +[35] Gemine Vivone, Rocco Restaino, Mauro Dalla Mura, Giorgio Licciardi, and Jocelyn Chanussot. Contrast and error-based fusion schemes for multispectral image pansharpening. IEEE Geoscience and Remote Sensing Letters, 11(5): 930-934, 2013. 2 +[36] Dong Wang, Yunpeng Bai, and Ying Li. Multispectral pan-sharpening via dual-channel convolutional network with convolutional LSTM based hierarchical spatial-spectral feature fusion, 2020. 3 +[37] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2 +[38] Chen Wu, Bo Du, Xiaohui Cui, and Liangpei Zhang. A post-classification change detection method based on itera + +tive slow feature analysis and bayesian soft fusion. Remote Sensing of Environment, 199:241-255, 2017. 1 +[39] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 2, 7 +[40] Gang Yang, Xiangyong Cao, Wenzhe Xiao, Man Zhou, Aiping Liu, Xun Chen, and Deyu Meng. Panflownet: A flow-based deep network for pan-sharpening. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16857-16867, 2023. 1 +[41] Licheng Yu, Yi Xu, Hongteng Xu, and Hao Zhang. Quaternion-based sparse representation of color image. In 2013 IEEE International Conference on Multimedia and Expo (ICME), pages 1-7. IEEE, 2013. 3 +[42] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 2, 7 +[43] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1 +[44] Rui Zeng, Jiasong Wu, Zhuhong Shao, Yang Chen, Beijing Chen, Lotfi Senhadji, and Huazhong Shu. Color image classification via quaternion principal component analysis network. Neurocomputing, 216:416-428, 2016. 3 +[45] Hao Zhang and Jiayi Ma. Gtp-pnet: A residual learning network based on gradient transformation prior for pansharpening. ISPRS Journal of Photogrammetry and Remote Sensing, 172:223-239, 2021. 1 +[46] Yunzuo Zhang, Jiayu Zhang, Ruixue Liu, Pengfei Zhu, and Yameng Liu. Key frame extraction based on quaternion fourier transform with multiple features fusion. Expert Systems with Applications, 216:119467, 2023. 4 +[47] Man Zhou, Jie Huang, Yanchi Fang, Xueyang Fu, and Aiping Liu. Pan-sharpening with customized transformer and invertible neural network. In Proceedings of the AAAI conference on artificial intelligence, pages 3553-3561, 2022. 2, 7 +[48] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In European conference on computer vision, pages 274-291. Springer, 2022. 3 +[49] Man Zhou, Keyu Yan, Jie Huang, Zihe Yang, Xueyang Fu, and Feng Zhao. Mutual information-driven pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1798-1808, 2022. 7 +[50] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. FOurmer: An efficient global modeling paradigm for image restoration. In International conference on machine learning, pages 42589-42601. PMLR, 2023. 4, 5 +[51] Xuanyu Zhu, Yi Xu, Hongteng Xu, and Changjian Chen. Quaternion convolutional neural networks. In Proceedings of + +the European conference on computer vision (ECCV), pages 631-647, 2018. 3 +[52] Zeyu Zhu, Xiangyong Cao, Man Zhou, Junhao Huang, and Deyu Meng. Probability-based global cross-modal upsampling for pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14039-14048, 2023. 3 \ No newline at end of file diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/images.zip b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..9b1963bbaae9a1a9c0bb323b8c1e9abd41a078a3 --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ec6b45d57e0fbab9245750e61c1da10dfc2a86bc8c1167b56dbd801503d84458 +size 615261 diff --git a/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/layout.json b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..46335617a192c63f679689c3b717f2f059fddec8 --- /dev/null +++ b/2025/Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions/layout.json @@ -0,0 +1,9587 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 87, + 103, + 523, + 121 + ], + "type": "text", + "content": "Enhanced Pansharpening via Quaternion Spatial-Spectral Interactions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 143, + 522, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 143, + 522, + 171 + ], + "spans": [ + { + "bbox": [ + 91, + 143, + 522, + 171 + ], + "type": "text", + "content": "Dong Li, Chunhui Luo, Yuanfei Bao, Gang Yang, Jie Xiao, Xueyang Fu, Zheng-Jun Zha* \nUniversity of Science and Technology of China, Hefei, China" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 186, + 174, + 421, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 174, + 421, + 185 + ], + "spans": [ + { + "bbox": [ + 186, + 174, + 421, + 185 + ], + "type": "text", + "content": "dongli6@mail.ustc.edu.cn，zhazj@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 610 + ], + "type": "text", + "content": "Pansharpening aims to generate high-resolution multispectral (MS) images by fusing panchromatic (PAN) images with corresponding low-resolution MS images. However, many existing methods struggle to fully capture spatial and spectral interactions, limiting their effectiveness. To address this, we propose a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging the compact representation capabilities of quaternions for high-dimensional data. Our method consists of three key components: the quaternion global spectral interaction branch, the quaternion local spatial structure awareness branch, and the quaternion spatial-spectral interaction branch. The first applies the quaternion Fourier transform to convert multi-channel features into the frequency domain as a whole, enabling global information interaction while preserving inter-channel dependencies, which aids spectral fidelity. The second uses a customized spatial quaternion representation, combined with a window-shifting strategy, to maintain local spatial dependencies while promoting spatial interactions, which helps inject spatial details. The last integrates the two pathways within the quaternion framework to enrich spatial-spectral interactions for richer representations. By utilizing quaternion's multi-dimensional representation and parameter-sharing properties, our method achieves a more compact and efficient cross-resolution, multi-band information integration, significantly improving the quality of the fused image. Extensive experiments validate the proposed method's effectiveness and its superior performance over current SOTA techniques. The code is available at https://github.com/dongli8/QuatPanNet." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "In remote sensing, satellite sensors face physical limitations that prevent the direct acquisition of high-resolution multispectral (MS) images [45]. Pansharpening addresses" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 214, + 552, + 371 + ], + "blocks": [ + { + "bbox": [ + 320, + 214, + 552, + 371 + ], + "lines": [ + { + "bbox": [ + 320, + 214, + 552, + 371 + ], + "spans": [ + { + "bbox": [ + 320, + 214, + 552, + 371 + ], + "type": "image", + "image_path": "022bd077a824460e477d738dedbcbbe71250b2914c9c9ad1f8d7f2d6b16ad737.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 382, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 382, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. Comparison between standard real-valued layer and quaternion-valued layer. The real-valued operation typically stacks the input multi-channel features, while the quaternion-valued layer, due to quaternion weight sharing, has the ability to learn the latent relationships within the input features." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 457, + 555, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 564 + ], + "type": "text", + "content": "this by merging low-resolution multispectral (LRMS) images with panchromatic (PAN) images to produce high-resolution multispectral (HRMS) images. As a crucial preprocessing step, pansharpening is widely applied across fields [38, 43]. Conceptually, it is a fusion super-resolution task, focused on incorporating fine spatial details while preserving spectral fidelity, requiring strong inter-channel dependency maintenance and the preservation of subtle local spatial structures." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 556, + 688 + ], + "type": "text", + "content": "Traditional methods often rely on mathematical models that integrate spatial and spectral data, typically assuming that the pansharpened image is a linear combination of various spectral bands from a high-resolution multispectral image [40]. However, these approaches are limited by their dependence on prior knowledge, which restricts flexibility and adaptability. With the rise of deep learning, convolutional neural networks (CNNs) have been introduced into pansharpening, leading to substantial breakthroughs in performance with increased model complexity [1, 5, 9, 23]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Despite notable successes, existing methods still struggle to capture the intricate spatial-spectral interactions essential" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10908" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "for high-quality pansharpening. These limitations hinder simultaneous feature interaction and inter-feature dependency preservation, leading to compromises in both spectral fidelity and spatial detail. Unlike conventional real-valued networks, quaternion networks treat multidimensional features as unified entities, allowing them to effectively encode internal dependencies [26, 27]. Additionally, quaternion networks demonstrate strong modeling capabilities for both inter- and intra-correlations within the quaternion latent space [3]. Thus, we aim to harness quaternion operations to further elevate pansharpening performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 207, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 296, + 506 + ], + "type": "text", + "content": "Our Motivation. PAN images provide high-resolution spatial details, while MS images offer rich multispectral information [11, 24]. A successful pansharpening model must preserve both spatial details and spectral fidelity during fusion. This requires capturing the unique features of PAN and MS images while leveraging their complementary information to produce accurate and comprehensive fused outputs. Traditional approaches often handle multi-channel inputs by stacking and mapping them to a single output channel, which causes information loss. Consequently, they struggle to capture inter-channel correlations and interactions between spatial and spectral data effectively. To compensate, some methods increase the number of output channels to improve representational capacity [39], but this also increases model complexity. Furthermore, these methods typically extract features in real-number space, limiting their ability to capture directional information such as edges and structures. In contrast, quaternions represent multiple spectral channels as distinct components of a compact quaternion representation, as shown in Figure 1. This approach integrates multi-band information efficiently while reducing information loss through parameter-sharing quaternion operations. These features make quaternions particularly well-suited for pansharpening tasks that involve multi-source data inputs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "Based on the analysis above, we propose a novel quaternion-based pansharpening framework called QuatPanNet. QuatPanNet is designed to balance spectral fidelity and spatial detail richness by leveraging quaternion-based representations and operations to enhance interactions among multispectral, spatial structure, and spatial-spectral features. The framework consists of three core components, i.e., Quaternion Global Spectral Interaction Branch,Quaternion Local Spatial Structure Awareness Branch, andQuaternion Spatial-Spectral Interaction Branch. The first one uses the Quaternion Fourier Transform (QFT) to process multi-channel features holistically in the frequency domain. By enabling global interactions while preserving inter-channel dependencies, it effectively extracts global spectral representations, which are critical for maintaining spectral fidelity. The second one employs a customized spatial quaternion transformation method and a window" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "shifting strategy, this branch focuses on preserving local structural dependencies and enhancing spatial interactions. It ensures that spatial details are well-integrated while maintaining the structural integrity of the input data. The last one, which built on the outputs of the first two branches, integrates dual-path features into a unified quaternion representation. This integration enhances the complementarity and coherence of information during the spatial-spectral fusion process. By utilizing the compact and efficient representation of PAN and MS image data in quaternion space, QuatPanNet facilitates effective interactions among spatial and spectral features. This approach reduces parameter complexity while achieving high spectral fidelity and rich spatial resolution, leading to superior fusion quality. In summary, the contributions of this work are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 255, + 553, + 409 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 255, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 255, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 314, + 255, + 553, + 290 + ], + "type": "text", + "content": "- We propose a novel quaternion-based pansharpening framework, QuatPanNet, designed to comprehensively harness quaternions for this task." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 290, + 553, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 290, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 290, + 553, + 361 + ], + "type": "text", + "content": "- We design customized quaternion-based spatial and spectral interaction modules for pansharpening. The former facilitates global information interaction while preserving inter-spectral dependencies, and the latter promotes spatial interaction while maintaining local structural dependencies." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 361, + 553, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 553, + 409 + ], + "type": "text", + "content": "- We introduce a quaternion-based spatial-spectral interaction branch, leveraging the high-dimensional representational power of quaternions to improve the fusion of spectral and spatial features." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 412, + 553, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 553, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 553, + 460 + ], + "type": "text", + "content": "Extensive experiments on multiple satellite datasets demonstrate that our method achieves superior qualitative and quantitative results with fewer parameters and effectively generalizes to full-resolution scenes." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 405, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 405, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 405, + 484 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 492, + 407, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 407, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 407, + 506 + ], + "type": "text", + "content": "2.1. Pansharpening" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "Existing pansharpening methods are divided into traditional ones and deep - learning - based ones. Traditional ones mainly consist of Component Substitution (CS) [31, 34], Multi - Resolution Analysis (MRA) [25, 29, 35], and Variational Optimization (VO) [8, 15, 21]. However, their heavy dependence on manually designed features and prior knowledge restricts their applicability [37]. PNN [23] is the first to introduce CNN into the field of pansharpening, achieving better results than traditional methods. Since then, deep learning-based approaches have dominated this field. MSDCNN [42] combines multi-scale feature extraction and residual learning, enhancing the quality of the fused images. INNformer [47], based on the Transformer architecture, integrates an invertible neural network module for lossless information to model long-range dependencies and efficient feature fusion, successfully introducing the Transformer architecture into this domain. The fusion of frequency do" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10909" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "main features in images is also a key focus for researchers [36, 48]. For instance, FAME-Net [16] proposes a method that combines MOE with frequency domain information, enabling the network to dynamically learn high-frequency information in remote sensing images. HFIN [33] explores the relationship between PAN and LRMS images using local Fourier information integration. In addition to designing pansharpening models, Zhu et al. developed the first upsampling module specifically for pansharpening methods, called PGCU [52]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 199, + 133, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 199, + 133, + 211 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 133, + 211 + ], + "type": "text", + "content": "2.2.Quaternion" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 515 + ], + "type": "text", + "content": "Quaternions are the natural extension of complex numbers to four dimensions, proposed by Hamilton in 1843 [13]. A quaternion consists of four components, including one real part and three imaginary parts. Quaternions, with their four components, are particularly well-suited for three-dimensional and four-dimensional feature vectors, as seen in applications such as image processing and robotic kinematics [2, 30]. In 2001, Pei et al. [28] introduced the quaternion model for color images, marking the beginning of quaternion applications in the field of image processing. The advantages of quaternion wavelet transforms [18], quaternion principal component analysis [44], and other quaternion-based color image processing techniques [41] have been demonstrated to extract more representative features for color images. These techniques have achieved promising results in advanced visual tasks such as color image classification. With the wave of deep learning, many researchers have begun to explore the application of quaternions in the field of deep learning [10, 26]. Xu et al. [51] proposes QCNN, which combines quaternions with CNNs, achieving faster convergence of the loss function while preserving color information more effectively than real-valued CNNs. Kusamichi et al. [19] studies color night vision and finds that using quaternion networks yields better results in extracting color information from dim images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 111, + 538 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "In this section, we first introduce the relevant properties of quaternions and their operators, and then present the proposed quaternion pan-sharpening network, as shown in Figures 2 and 3. We then provide a detailed explanation of the fundamental components of our method, including three quaternion-based branches: a) the global interaction branch utilizes the Quaternion Fourier Transform and quaternion convolution to maintain channel interdependencies while representing global information; b) the local structure-aware branch models local spatial dependencies through a specially designed spatial quaternion transformation method combined with a window-shifting strategy; and c) the dual-domain interaction branch leverages the orthogonal relationships inherent in quaternions to model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "dual-domain correlations and promote complementary feature learning. Finally, we present the optimization strategy employed in our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 114, + 442, + 128 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 442, + 128 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 442, + 128 + ], + "type": "text", + "content": "3.1.Quaternion Operators" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 132, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 132, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 132, + 554, + 168 + ], + "type": "text", + "content": "Quaternion, proposed by Hamilton in 1843, extends complex numbers to a four-dimensional space. A quaternion " + }, + { + "bbox": [ + 313, + 132, + 554, + 168 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{H}" + }, + { + "bbox": [ + 313, + 132, + 554, + 168 + ], + "type": "text", + "content": " can be expressed as follows [32]:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 381, + 176, + 555, + 189 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 176, + 555, + 189 + ], + "spans": [ + { + "bbox": [ + 381, + 176, + 555, + 189 + ], + "type": "interline_equation", + "content": "Q = w 1 + x i + y j + z k, \\tag {1}", + "image_path": "c350246cb60ab04027baa0ef7e71e9d867d189bcc07f472a6daeb872c3065759.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "inline_equation", + "content": "w, x, y, z" + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "content": " are all real numbers, and " + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "inline_equation", + "content": "1, i, j, k" + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "content": " are the quaternion unit basis. " + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "inline_equation", + "content": "w1" + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "content": " refers to the real component of the number, and " + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "inline_equation", + "content": "xi + yj + zk" + }, + { + "bbox": [ + 313, + 196, + 554, + 255 + ], + "type": "text", + "content": " to the imaginary component. In the definition of quaternions, they must satisfy the following identities, known as the Hamilton rules:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 377, + 262, + 555, + 276 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 262, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 377, + 262, + 555, + 276 + ], + "type": "interline_equation", + "content": "i ^ {2} = j ^ {2} = k ^ {2} = i j k = - 1. \\tag {2}", + "image_path": "c1a369cf87d4dd0dfc401f1cf16ae78768a52264bbed7d284cf0926478b49674.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 283, + 554, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 366 + ], + "type": "text", + "content": "We propose the use of quaternions and their associated operators to uncover both intra- and inter-relations between the features of high spatial resolution PAN and MS images. This approach is inspired by the unique feature-handling capabilities inherent in quaternion representation [3]. Consequently, we introduce two quaternion operators: quaternion convolution and the quaternion Fourier transform." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 368, + 434, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 368, + 434, + 378 + ], + "spans": [ + { + "bbox": [ + 326, + 368, + 434, + 378 + ], + "type": "text", + "content": "Quaternion Convolution" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": "In the real-valued space, convolution involves convolving a filter matrix with a vector. Extending this operation to quaternions, quaternion convolution is achieved by convolving a quaternion filter matrix with a quaternion vector. To maintain consistency with the defined quaternions, we represent the computation of quaternion convolution using real-valued matrices. This computation is performed by convolving the quaternion filter matrix " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "W = A + Bi + Cj + Dk" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": " with the quaternion vector " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "Q = w + xi + yj + zk" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "A, B, C" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": " are real-valued matrices. The convolution operation between " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 379, + 554, + 510 + ], + "type": "text", + "content": " is denoted as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 338, + 517, + 553, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 517, + 553, + 575 + ], + "spans": [ + { + "bbox": [ + 338, + 517, + 553, + 575 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} W * Q = (A * w - B * x - C * y - D * z) + \\\\ i (A * x + B * w + C * z - D * y) + \\\\ j (A * y - B * z + C * w + D * x) + \\tag {3} \\\\ K (A * z + B * y - C * x + D * w). \\\\ \\end{array}", + "image_path": "da8646b1e0ebe8b6774e8e4b3c4de8f5542b2c89a84797ba6765f9b4623c5b0f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "It can be seen that the result of quaternion convolution produces a unique linear combination of each coordinate axis for every single-axis result. This stems from the structure of quaternion multiplication and enforces interactions between each axis of the kernel and each axis of the image. This contrasts with real-valued convolution, which simply multiplies each channel of the kernel with the corresponding channel of the image. The advantage of this approach is that quaternion convolution promotes cross-axis interactions and information extraction through the reuse of filters across each axis and combination [27]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10910" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 555, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 555, + 270 + ], + "type": "image", + "image_path": "4360aa7e7cba208c962cf624fc35b89f8447ca4e5995689294eef91955104070.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 555, + 301 + ], + "type": "text", + "content": "Figure 2. The framework of our proposed pan-sharpening network QuatPanNet. It consists of several Quaternion Spatial-Spectral Interaction Blocks (QSSIB). The Quaternion Spatial-Spectral Interaction Block is shown in Figure 3." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 318, + 203, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 318, + 203, + 329 + ], + "spans": [ + { + "bbox": [ + 67, + 318, + 203, + 329 + ], + "type": "text", + "content": "Quaternion Fourier Transform" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 330, + 295, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 330, + 295, + 449 + ], + "spans": [ + { + "bbox": [ + 54, + 330, + 295, + 449 + ], + "type": "text", + "content": "The Fourier Transform plays a crucial role in various image processing tasks, including pansharpening [50]. After decomposing the input image into multiple channels, the Discrete Fourier Transform (DFT) can be applied to each channel individually. However, this approach neglects the mutual information among different channels, which can hinder the performance of pansharpening. The Discrete Quaternion Fourier Transform (DQFT) offers the possibility of simultaneously transforming multi-channel images, as it can transform quaternions holistically [4]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "text", + "content": "Due to the non-commutative nature of quaternion multiplication, there are various types of Quaternion Fourier Transforms (QFT), such as left-side, right-side and two-side QFT [6]. In this paper, we utilize a typical QFT, specifically the left-sided QFT. For an image " + }, + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "inline_equation", + "content": "\\dot{f}" + }, + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "text", + "content": " of size " + }, + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "inline_equation", + "content": "H\\times W" + }, + { + "bbox": [ + 55, + 449, + 295, + 521 + ], + "type": "text", + "content": ", the Discrete Quaternion Fourier Transform is defined as follows." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 537, + 294, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 537, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 61, + 537, + 294, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {F} (\\dot {f}) (u, v) = \\frac {1}{\\sqrt {H W}} \\sum_ {x = 0} ^ {H - 1} \\sum_ {y = 0} ^ {W - 1} e ^ {- \\dot {\\mu} 2 \\pi \\left(\\frac {x}{H} u + \\frac {y}{H} v\\right)} \\dot {f} (x, y), \\tag {4}", + "image_path": "93999489df412e2ae176df486632f4dc1352b3bb8cea6c66aff95d96d08866a7.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\dot{f} (x,y)" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " is pixel of " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\dot{f}" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\dot{\\mu}" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " is unit pure quaternion. " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " is the coordinate in the quaternion frequency domain. " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}(\\dot{f})" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " defines the inverse quaternion Fourier transform accordingly. Inspired by the traditional Fourier transform, the modulus of " + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\dot{f})(u,v)" + }, + { + "bbox": [ + 55, + 581, + 295, + 665 + ], + "type": "text", + "content": " is taken as the amplitude component, and the angles with the three imaginary axes are taken as the phase components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "It is worth noting that the QFT is not a simple substitution of the imaginary unit in the two-dimensional Fourier Transform with quaternions; rather, its basis functions possess orthogonal imaginary components, which preserve the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 318, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 474 + ], + "type": "text", + "content": "information and mutual relationships among different components in the frequency domain. Consequently, some studies leverage QFT to treat multiple features as a unified whole to enhance performance [4, 46]. Furthermore, the efficiency of QFT is also noteworthy. For example, when comparing the use of QFT with three separate conventional Fourier Transforms for processing color images, the number of real multiplications and additions required by QFT is approximately " + }, + { + "bbox": [ + 313, + 318, + 555, + 474 + ], + "type": "inline_equation", + "content": "75\\%" + }, + { + "bbox": [ + 313, + 318, + 555, + 474 + ], + "type": "text", + "content": " of that needed for the three separate transforms [30]. For pansharpening, we propose utilizing the QFT to treat multispectral data as a unified whole, facilitating global information interactions in the frequency domain while preserving inter-channel dependencies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 489, + 392, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 489, + 392, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 489, + 392, + 502 + ], + "type": "text", + "content": "3.2. Framework" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Based on the above analysis, we propose a novel pansharpening approach based on quaternion information interaction, as illustrated in Figure 2. Given a panchromatic (PAN) image " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " and a low-resolution multispectral (LRMS) image " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "L \\in \\mathbb{R}^{H / r \\times W / r \\times C}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": ", the network first uses convolution to project the bicubic upsampled r-times " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " into a shallow feature representation. Simultaneously, " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " is processed through multiple cascaded convolution layers to extract effective information for high-resolution multispectral (HRMS) reconstruction. Then, the features of the PAN and LRMS images are jointly processed through the core building module, the quaternion spatial-spectral interaction block (QSSIB), for continuous global information interaction and local structure-aware integration. Next, a convolution layer transforms all features gathered by the N QSSIBs back into the image space, and finally, these features are combined with the residual to obtain the final output." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10911" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 556, + 289 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 289 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 289 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 289 + ], + "type": "image", + "image_path": "5cbe67bf3e36207152cfbadca3c2bf3991aa580420568ebb7b7d872a29ba923a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 295, + 555, + 329 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 329 + ], + "type": "text", + "content": "Figure 3. The detailed flowchart of the proposed core building module quaternion spatial-spectral interaction block (QSSIB), consisting of three components: quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch (QLSA) and quaternion spatial-spectral interaction branch (QSSI)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 334, + 203, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 203, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 203, + 347 + ], + "type": "text", + "content": "3.3. Key Building Components" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 352, + 297, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 352, + 297, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 352, + 297, + 556 + ], + "type": "text", + "content": "As shown in Figure 3, the fundamental building block of our method comprises three key elements: the quaternion global spectral interaction branch (QGSI), the quaternion local spatial structure awareness branch (QLSA), and the quaternion spatial-spectral interaction branch (QSSI). To convert images into quaternion representations, we design customized methods, including channel quaternion representation (CQR), spatial quaternion representation (SQR), and shifted window quaternion representation (SWQR). When transforming quaternion representations back into image features, we denote the corresponding inverse operations as ICQR, ISQR, and ISWQR, respectively. (For more details on quaternion representation, refer to the supplementary material.) These three key components, in conjunction with the quaternion representations of the images, jointly facilitate channel interactions during multispectral fusion and local interactions for spatial resolution enhancement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 558, + 240, + 569 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 558, + 240, + 569 + ], + "spans": [ + { + "bbox": [ + 67, + 558, + 240, + 569 + ], + "type": "text", + "content": "Quaternion Global Spectral Interaction" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": "The global interaction branch utilizes the Quaternion Fourier Transform (QFT) to transform multi-channel features as a whole into the frequency domain, facilitating global information interactions. Since the input to QFT must be in quaternion form and we aim to maintain interchannel interdependencies, we design a channel quaternion representation (CQR). This approach divides the image features equally along the channel dimension into four parts, which are then used as the four components of a quaternion. Let the features of the MS and PAN images be denoted as " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{ms}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{p}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ", respectively; their channel-dimensional quaternion representations are denoted as " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "Q_{ms}^{c}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "inline_equation", + "content": "Q_{p}^{c}" + }, + { + "bbox": [ + 55, + 570, + 296, + 715 + ], + "type": "text", + "content": ". We" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 334, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 418 + ], + "type": "text", + "content": "then use QFT to transform the MS and PAN features into the frequency domain. Some studies on image restoration indicate that decomposing amplitude and phase components facilitates the decoupling of degradation [17, 50]. Therefore, we decompose the frequency domain features into amplitude and phase components, with the corresponding process represented as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 352, + 422, + 555, + 437 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 422, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 352, + 422, + 555, + 437 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {p} ^ {c}\\right), \\tag {5}", + "image_path": "44a7895943e1675dba1f847a884f0042d67633e8b18afdeae485097aa7b1d636.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 319, + 439, + 553, + 453 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 439, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 319, + 439, + 553, + 453 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {1} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) = \\mathcal {F} \\left(Q _ {m s} ^ {c}\\right), \\tag {6}", + "image_path": "57c832eee0216b47fc0141f249f65489e393722f15d4d48d198bc46d558bb41b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\cdot)" + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": " represents the amplitude, and " + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_1(\\cdot),\\mathcal{P}_2(\\cdot)" + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_3(\\cdot)" + }, + { + "bbox": [ + 313, + 456, + 555, + 552 + ], + "type": "text", + "content": " represent the phases corresponding to the three imaginary axes. We then concatenate the amplitude and phase spectra of the PAN and MS images accordingly to form new quaternions. Next, we process the newly formed quaternions using quaternion convolution, thereby enhancing global representations in the frequency domain while promoting interactions between channels." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 318, + 555, + 553, + 598 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 555, + 553, + 598 + ], + "spans": [ + { + "bbox": [ + 318, + 555, + 553, + 598 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q _ {f} ^ {c} = \\mathcal {O Q} (\\mathcal {C} [ \\mathcal {A} (Q _ {p} ^ {c}), \\mathcal {A} (Q _ {m s} ^ {c}) ], \\mathcal {C} [ \\mathcal {P} _ {1} (Q _ {p} ^ {c}), \\mathcal {P} _ {1} (Q _ {m s} ^ {c}) ] \\\\ , \\mathcal {C} \\left[ \\mathcal {P} _ {2} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {2} \\left(Q _ {m s} ^ {c}\\right) \\right], \\mathcal {C} \\left[ \\mathcal {P} _ {3} \\left(Q _ {p} ^ {c}\\right), \\mathcal {P} _ {3} \\left(Q _ {m s} ^ {c}\\right) \\right]) \\tag {7} \\\\ \\end{array}", + "image_path": "285c841eaa170be7ce1289d2412976e409ee04b9c38466e8c5277792abec4571.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "text", + "content": " denotes concatenation along the channel dimension, " + }, + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "inline_equation", + "content": "\\mathcal{OQ}" + }, + { + "bbox": [ + 313, + 602, + 555, + 661 + ], + "type": "text", + "content": " is quaternion convolution. The fused amplitude and phase components are the transformed back to the spatial domain using the inverse discrete quaternion Fourier transform." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 393, + 662, + 555, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 393, + 662, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 393, + 662, + 555, + 677 + ], + "type": "interline_equation", + "content": "Q _ {f r e} ^ {c} = \\mathcal {F} ^ {- 1} \\left(Q _ {f} ^ {c}\\right). \\tag {8}", + "image_path": "a8512900476b8e16e05a75469ea59013ff360f7de0d3103cea8a516f78badc0a.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "By representing the image in a channel-dimensional quaternion form, the quaternion operator treats the multichannel data as a unified whole for information interaction." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "10912" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "Additionally, according to spectral convolution theorem [7], processing information in the Fourier space enables capturing global frequency representations. In summary, this branch generates the global information representation with strengthened channel interactions " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c}" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 133, + 270, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 133, + 270, + 144 + ], + "spans": [ + { + "bbox": [ + 67, + 133, + 270, + 144 + ], + "type": "text", + "content": "Quaternion Local Spatial Structure Awareness" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 146, + 295, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 240 + ], + "type": "text", + "content": "Pansharpening is essentially a fusion-based superresolution task, making the injection of spatial resolution critically important. Enhancing spatial resolution requires not only global information from the frequency domain but also local information interactions in the spatial domain. For pansharpening, preserving local structures during information interactions is particularly significant, as it enhances the network's ability to recover fine details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "text", + "content": "We aim to leverage the balance of enhanced information interaction and specific information preservation offered by quaternions to improve the network's spatial superresolution capabilities. To this end, we design the spatial quaternion representation (SQR). It divides the image into several " + }, + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 55, + 242, + 295, + 373 + ], + "type": "text", + "content": " windows in the spatial dimension, with each local window's four pixels serving as the four components of a quaternion, as illustrated in Figure 3. This approach explicitly ensures dependencies among local pixels, preserving the local structure. Based on this representation, we use quaternion convolutions (QCNN) to process the features." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 374, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 295, + 495 + ], + "type": "text", + "content": "However, this representation method lacks crosswindow connections, which limits spatial interactions. Inspired by [22], we design the shifted window quaternion representation (SWQR) based on SQR. This allows the network to alternately employ two partitioning configurations, thereby enhancing the effectiveness of spatial quaternion representation. Through the combined effect of this method and quaternion convolution, this branch preserves local spatial dependencies while promoting spatial interactions, thereby obtaining the local representation " + }, + { + "bbox": [ + 55, + 374, + 295, + 495 + ], + "type": "inline_equation", + "content": "Q_{spa}^{s}" + }, + { + "bbox": [ + 55, + 374, + 295, + 495 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 295, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 566 + ], + "type": "text", + "content": "The information representations of the two branches are highly complementary, with one branch responsible for global, frequency-domain, and channel interactions, and the other focused on local, spatial-domain, and spatial interactions. Therefore, further interaction and integration between them facilitate mutual compensation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 567, + 241, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 567, + 241, + 578 + ], + "spans": [ + { + "bbox": [ + 67, + 567, + 241, + 578 + ], + "type": "text", + "content": "Quaternion Spatial-Spectral Interaction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "text", + "content": "The two branches described above respectively obtain quaternion representations in the channel dimension, " + }, + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c} \\in \\mathbb{H}^{H \\times W \\times C / 4}" + }, + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "text", + "content": ", and in the spatial dimension, " + }, + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "inline_equation", + "content": "Q_{spa}^{s} \\in \\mathbb{H}^{H / 2 \\times W / 2 \\times C}" + }, + { + "bbox": [ + 55, + 579, + 295, + 665 + ], + "type": "text", + "content": ". The quaternion dual-domain information interaction branch mainly consists of two parts: one is a quaternion-based global fusion, and the other is the channel-aware quaternion spatial feature injection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 713 + ], + "type": "text", + "content": "Quaternion-based global fusion merges the features obtained from the two branches into a single quaternion for unified processing, thereby enhancing the correlation between the two through the structure of the quaternion. First," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": "both " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{spa}^{s}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " are converted back into image feature representations. " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " is combined along the channel dimension to obtain " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{fre} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{spa}^{s}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " is rearranged spatially to obtain " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{spa} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": ". To achieve a richer feature representation, we apply convolutions to " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{fre}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{fre}'" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "F_{spa}'" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": ", respectively. We then combine these four features into a quaternion " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{f} = F_{fre} + F_{spa}i + F_{fre}^{\\prime}j + F_{spa}^{\\prime}k" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": ". We use quaternion convolution to process " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{f} \\in \\mathbb{H}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": " as a whole, producing " + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "inline_equation", + "content": "Q_{a} \\in \\mathbb{H}^{H \\times W \\times C/4}" + }, + { + "bbox": [ + 313, + 72, + 553, + 196 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": "Channel-aware quaternion spatial feature injection involves injecting spatial features " + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": " from the local structure-aware branch into the four components of the quaternion representation " + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c}" + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": " in the channel dimension, thereby facilitating information fusion in both spatial and channel domains. Specifically, we replicate the image feature " + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "inline_equation", + "content": "F_{spa}" + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": " four times and concatenate each copy with one of the four components of the quaternion feature " + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "inline_equation", + "content": "Q_{fre}^{c}" + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": " to obtain " + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "inline_equation", + "content": "Q_{i} \\in \\mathbb{H}^{H \\times W \\times 5C / 4}" + }, + { + "bbox": [ + 313, + 197, + 553, + 306 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 324, + 311, + 553, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 311, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 324, + 311, + 553, + 326 + ], + "type": "interline_equation", + "content": "Q _ {i} = \\mathcal {C} \\left[ Q _ {f r e} ^ {c}, \\left(F _ {s p a} + F _ {s p a} i + F _ {s p a} j + F _ {s p a} k\\right) \\right]. \\tag {9}", + "image_path": "41957efab28c911dceb9ff6b2a436f8f529a8aac142334366aa3a7d350c17377.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": "We then feed " + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": " into quaternion convolution to obtain " + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "inline_equation", + "content": "Q_{b} \\in \\mathbb{H}^{H \\times W \\times C / 4}" + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": ". Afterward, we concatenate " + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "inline_equation", + "content": "Q_{a}" + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "inline_equation", + "content": "Q_{b}" + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": ", and apply quaternion convolution for further fusion to obtain " + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "inline_equation", + "content": "Q_{out} \\in \\mathbb{H}^{H \\times W \\times C / 4}" + }, + { + "bbox": [ + 313, + 329, + 553, + 377 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 374, + 382, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 382, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 374, + 382, + 553, + 396 + ], + "type": "interline_equation", + "content": "Q _ {o u t} = \\operatorname {Q C o n v} \\left(\\mathcal {C} \\left[ Q _ {a}, Q _ {b} \\right]\\right), \\tag {10}", + "image_path": "cf1244ff0e2a57f1587902619392483ea03320fb10972cfc1e937466baac0e80.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": "where QConv denotes quaternion convolution. We then convert " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "Q_{out}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " into the image feature representation " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "F_{out} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": ". Finally, a residual learning mechanism is applied, adding the input feature " + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "inline_equation", + "content": "F_{ms}" + }, + { + "bbox": [ + 313, + 401, + 553, + 473 + ], + "type": "text", + "content": " to the fused feature. The proposed network achieves effective fusion through the seamless integration of the two quaternion representations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 400, + 491 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": " denote the network's output and the corresponding ground truth, respectively. We use a joint spatial-frequency domain loss to supervise the network training. In the frequency domain, we first apply the quaternion Fourier transform to convert " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": " into the quaternion Fourier space. Then, we compute the L1 norm of the real and imaginary differences between " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "inline_equation", + "content": "GT" + }, + { + "bbox": [ + 313, + 496, + 553, + 591 + ], + "type": "text", + "content": ", and sum them to obtain the quaternion Fourier domain loss." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 323, + 596, + 553, + 625 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 596, + 553, + 625 + ], + "spans": [ + { + "bbox": [ + 323, + 596, + 553, + 625 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {f r e} = \\| R (Y) - R (G T) \\| _ {1} + \\| I (Y) - I (G T) \\| _ {1} \\tag {11} \\\\ + \\| J (Y) - J (G T) \\| _ {1} + \\| K (Y) - K (G T) \\| _ {1}, \\\\ \\end{array}", + "image_path": "3731e140a6ae67ff6d6e0ba2ea3c2a3231b22196050115dca6ffab2ebffdfed7.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "inline_equation", + "content": "R(\\cdot), I(\\cdot), J(\\cdot)" + }, + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "inline_equation", + "content": "K(\\cdot)" + }, + { + "bbox": [ + 313, + 630, + 553, + 677 + ], + "type": "text", + "content": " represent the real part and three imaginary parts of the quaternion respectively. In the spatial domain, we employ the L1 loss. Finally, the overall loss function is as follows" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 377, + 683, + 553, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 683, + 553, + 697 + ], + "spans": [ + { + "bbox": [ + 377, + 683, + 553, + 697 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\| Y - G T \\| _ {1} + \\lambda \\mathcal {L} _ {f r e}, \\tag {12}", + "image_path": "335fb4be8506c44f27aff934382ff304f228e780daf5a5bef37cd61cafde28c2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 701, + 547, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 547, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 547, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 701, + 547, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 701, + 547, + 714 + ], + "type": "text", + "content": " is the weighting factor and set to 0.1 empirically." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10913" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 130, + 144 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 130, + 144 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 130, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 130, + 144 + ], + "type": "image", + "image_path": "0d58f5574228d307d2321d3d5d8c388c6620057a5ecbbc892c7ebc15abcf844c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 145, + 109, + 155 + ], + "lines": [ + { + "bbox": [ + 80, + 145, + 109, + 155 + ], + "spans": [ + { + "bbox": [ + 80, + 145, + 109, + 155 + ], + "type": "text", + "content": "LRMS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 141, + 70, + 216, + 144 + ], + "blocks": [ + { + "bbox": [ + 141, + 70, + 216, + 144 + ], + "lines": [ + { + "bbox": [ + 141, + 70, + 216, + 144 + ], + "spans": [ + { + "bbox": [ + 141, + 70, + 216, + 144 + ], + "type": "image", + "image_path": "38072c22dc00240d02d7e6ed65f037633e9d00ce7651e752f71d89ab1acad415.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 225, + 70, + 300, + 144 + ], + "blocks": [ + { + "bbox": [ + 225, + 70, + 300, + 144 + ], + "lines": [ + { + "bbox": [ + 225, + 70, + 300, + 144 + ], + "spans": [ + { + "bbox": [ + 225, + 70, + 300, + 144 + ], + "type": "image", + "image_path": "ffef7e3696608666d2933be36000362a1f6e0f6d9de01a923f225dc9f11d2cbb.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 256, + 145, + 270, + 155 + ], + "lines": [ + { + "bbox": [ + 256, + 145, + 270, + 155 + ], + "spans": [ + { + "bbox": [ + 256, + 145, + 270, + 155 + ], + "type": "text", + "content": "GS" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 384, + 144 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 384, + 144 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 384, + 144 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 384, + 144 + ], + "type": "image", + "image_path": "87ab71925aaaba337c145cbcb0cd9044d4ff2687e234c8d0a4d11107d43b92fc.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 395, + 70, + 470, + 144 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 470, + 144 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 470, + 144 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 470, + 144 + ], + "type": "image", + "image_path": "5065df451e37e817546e64992c3cca13443bedf43a23c3d79998c594042cfa73.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 145, + 454, + 155 + ], + "lines": [ + { + "bbox": [ + 410, + 145, + 454, + 155 + ], + "spans": [ + { + "bbox": [ + 410, + 145, + 454, + 155 + ], + "type": "text", + "content": "MSDCNN" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 480, + 70, + 554, + 144 + ], + "blocks": [ + { + "bbox": [ + 480, + 70, + 554, + 144 + ], + "lines": [ + { + "bbox": [ + 480, + 70, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 480, + 70, + 554, + 144 + ], + "type": "image", + "image_path": "99bc74a635a76c945eea1533db2d5ee4d6b23d610e6c2253dab76896a0d81de4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 145, + 534, + 155 + ], + "lines": [ + { + "bbox": [ + 500, + 145, + 534, + 155 + ], + "spans": [ + { + "bbox": [ + 500, + 145, + 534, + 155 + ], + "type": "text", + "content": "GPPNN" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 56, + 156, + 130, + 231 + ], + "blocks": [ + { + "bbox": [ + 56, + 156, + 130, + 231 + ], + "lines": [ + { + "bbox": [ + 56, + 156, + 130, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 130, + 231 + ], + "type": "image", + "image_path": "f4cce0d1da5c8e8b4cd038352abc5bbb4dd90c14de8800013673f3a5d31a8000.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 232, + 111, + 243 + ], + "lines": [ + { + "bbox": [ + 78, + 232, + 111, + 243 + ], + "spans": [ + { + "bbox": [ + 78, + 232, + 111, + 243 + ], + "type": "text", + "content": "MutNet" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 141, + 156, + 216, + 231 + ], + "blocks": [ + { + "bbox": [ + 168, + 145, + 189, + 156 + ], + "lines": [ + { + "bbox": [ + 168, + 145, + 189, + 156 + ], + "spans": [ + { + "bbox": [ + 168, + 145, + 189, + 156 + ], + "type": "text", + "content": "PAN" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 141, + 156, + 216, + 231 + ], + "lines": [ + { + "bbox": [ + 141, + 156, + 216, + 231 + ], + "spans": [ + { + "bbox": [ + 141, + 156, + 216, + 231 + ], + "type": "image", + "image_path": "a9d0c937fd2566ead0393ffd7b16b92077b50abf607409f9c1681efc351c03d5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 156, + 232, + 202, + 243 + ], + "lines": [ + { + "bbox": [ + 156, + 232, + 202, + 243 + ], + "spans": [ + { + "bbox": [ + 156, + 232, + 202, + 243 + ], + "type": "text", + "content": "INNformer" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 225, + 156, + 300, + 231 + ], + "blocks": [ + { + "bbox": [ + 225, + 156, + 300, + 231 + ], + "lines": [ + { + "bbox": [ + 225, + 156, + 300, + 231 + ], + "spans": [ + { + "bbox": [ + 225, + 156, + 300, + 231 + ], + "type": "image", + "image_path": "1919873bdbc6fd9d15f08e2e747dbc178aadc3b91e380fe5b7c656d0e4d293f7.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 251, + 232, + 276, + 242 + ], + "lines": [ + { + "bbox": [ + 251, + 232, + 276, + 242 + ], + "spans": [ + { + "bbox": [ + 251, + 232, + 276, + 242 + ], + "type": "text", + "content": "HFIN" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 156, + 384, + 231 + ], + "blocks": [ + { + "bbox": [ + 337, + 145, + 358, + 156 + ], + "lines": [ + { + "bbox": [ + 337, + 145, + 358, + 156 + ], + "spans": [ + { + "bbox": [ + 337, + 145, + 358, + 156 + ], + "type": "text", + "content": "PNN" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 156, + 384, + 231 + ], + "lines": [ + { + "bbox": [ + 310, + 156, + 384, + 231 + ], + "spans": [ + { + "bbox": [ + 310, + 156, + 384, + 231 + ], + "type": "image", + "image_path": "c8130270ab584a8b7065bfdc9df831381e9876bad7a81e34f5c9bba3822a9c0b.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 324, + 232, + 372, + 242 + ], + "lines": [ + { + "bbox": [ + 324, + 232, + 372, + 242 + ], + "spans": [ + { + "bbox": [ + 324, + 232, + 372, + 242 + ], + "type": "text", + "content": "FAME-Net" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 395, + 156, + 470, + 231 + ], + "blocks": [ + { + "bbox": [ + 395, + 156, + 470, + 231 + ], + "lines": [ + { + "bbox": [ + 395, + 156, + 470, + 231 + ], + "spans": [ + { + "bbox": [ + 395, + 156, + 470, + 231 + ], + "type": "image", + "image_path": "f8b8c85f3af244ebc58911bfa5dc2e7da3b3b25155bae08f62b295c0d8814d5a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 232, + 443, + 242 + ], + "lines": [ + { + "bbox": [ + 422, + 232, + 443, + 242 + ], + "spans": [ + { + "bbox": [ + 422, + 232, + 443, + 242 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 480, + 156, + 554, + 231 + ], + "blocks": [ + { + "bbox": [ + 480, + 156, + 554, + 231 + ], + "lines": [ + { + "bbox": [ + 480, + 156, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 480, + 156, + 554, + 231 + ], + "type": "image", + "image_path": "b9059ccf5e428aa1f113a6c11bf7e7bd144a5f8fa9c3a879f0f14acb7af998e6.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 509, + 232, + 523, + 242 + ], + "lines": [ + { + "bbox": [ + 509, + 232, + 523, + 242 + ], + "spans": [ + { + "bbox": [ + 509, + 232, + 523, + 242 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 56, + 245, + 108, + 291 + ], + "blocks": [ + { + "bbox": [ + 56, + 245, + 108, + 291 + ], + "lines": [ + { + "bbox": [ + 56, + 245, + 108, + 291 + ], + "spans": [ + { + "bbox": [ + 56, + 245, + 108, + 291 + ], + "type": "image", + "image_path": "ca471fb721fb0893d11f23de893307781a0fbb459e18c1b16884346cc7ad1384.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 294, + 89, + 304 + ], + "lines": [ + { + "bbox": [ + 77, + 294, + 89, + 304 + ], + "spans": [ + { + "bbox": [ + 77, + 294, + 89, + 304 + ], + "type": "text", + "content": "GS" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 94, + 315, + 515, + 327 + ], + "lines": [ + { + "bbox": [ + 94, + 315, + 515, + 327 + ], + "spans": [ + { + "bbox": [ + 94, + 315, + 515, + 327 + ], + "type": "text", + "content": "Figure 4. The visual comparisons between other pan-sharpening methods and our method on WorldView-II satellite" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 111, + 246, + 162, + 291 + ], + "blocks": [ + { + "bbox": [ + 111, + 246, + 162, + 291 + ], + "lines": [ + { + "bbox": [ + 111, + 246, + 162, + 291 + ], + "spans": [ + { + "bbox": [ + 111, + 246, + 162, + 291 + ], + "type": "image", + "image_path": "ad31d5f2834d3ce6e8ac79aaaae08a8e04c991bde3c783f9875b8ec8b543bcd0.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 127, + 294, + 146, + 304 + ], + "lines": [ + { + "bbox": [ + 127, + 294, + 146, + 304 + ], + "spans": [ + { + "bbox": [ + 127, + 294, + 146, + 304 + ], + "type": "text", + "content": "PNN" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 165, + 246, + 216, + 291 + ], + "blocks": [ + { + "bbox": [ + 165, + 246, + 216, + 291 + ], + "lines": [ + { + "bbox": [ + 165, + 246, + 216, + 291 + ], + "spans": [ + { + "bbox": [ + 165, + 246, + 216, + 291 + ], + "type": "image", + "image_path": "13399ce9c97b7e0b3fdc734b48adfd4f3f58d92e79e0226e605edd676d45a5d6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 171, + 294, + 209, + 304 + ], + "lines": [ + { + "bbox": [ + 171, + 294, + 209, + 304 + ], + "spans": [ + { + "bbox": [ + 171, + 294, + 209, + 304 + ], + "type": "text", + "content": "MSDCNN" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 219, + 246, + 269, + 291 + ], + "blocks": [ + { + "bbox": [ + 219, + 246, + 269, + 291 + ], + "lines": [ + { + "bbox": [ + 219, + 246, + 269, + 291 + ], + "spans": [ + { + "bbox": [ + 219, + 246, + 269, + 291 + ], + "type": "image", + "image_path": "9f34b258e8df8c4e0ecb0cbc917a61f783996d8baac216bc9908eb89bbafb680.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 294, + 260, + 304 + ], + "lines": [ + { + "bbox": [ + 229, + 294, + 260, + 304 + ], + "spans": [ + { + "bbox": [ + 229, + 294, + 260, + 304 + ], + "type": "text", + "content": "GPPNN" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 272, + 246, + 323, + 291 + ], + "blocks": [ + { + "bbox": [ + 272, + 246, + 323, + 291 + ], + "lines": [ + { + "bbox": [ + 272, + 246, + 323, + 291 + ], + "spans": [ + { + "bbox": [ + 272, + 246, + 323, + 291 + ], + "type": "image", + "image_path": "463b8bb99ce9834c6f14903c20eacbabbc5ab291dda3178919fbecd78d0eaa50.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 284, + 294, + 313, + 304 + ], + "lines": [ + { + "bbox": [ + 284, + 294, + 313, + 304 + ], + "spans": [ + { + "bbox": [ + 284, + 294, + 313, + 304 + ], + "type": "text", + "content": "MutNet" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 326, + 246, + 377, + 291 + ], + "blocks": [ + { + "bbox": [ + 326, + 246, + 377, + 291 + ], + "lines": [ + { + "bbox": [ + 326, + 246, + 377, + 291 + ], + "spans": [ + { + "bbox": [ + 326, + 246, + 377, + 291 + ], + "type": "image", + "image_path": "c09ebd78fae13c0505879aee7ed21b324bd273d2efbb21a84c992d4a2537b410.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 332, + 294, + 373, + 304 + ], + "lines": [ + { + "bbox": [ + 332, + 294, + 373, + 304 + ], + "spans": [ + { + "bbox": [ + 332, + 294, + 373, + 304 + ], + "type": "text", + "content": "INNformer" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 380, + 246, + 431, + 291 + ], + "blocks": [ + { + "bbox": [ + 380, + 246, + 431, + 291 + ], + "lines": [ + { + "bbox": [ + 380, + 246, + 431, + 291 + ], + "spans": [ + { + "bbox": [ + 380, + 246, + 431, + 291 + ], + "type": "image", + "image_path": "e79a805be7f28a2d1cce5c85af39f15ab3f14ea6854a027cc99286d8283968af.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 294, + 417, + 304 + ], + "lines": [ + { + "bbox": [ + 395, + 294, + 417, + 304 + ], + "spans": [ + { + "bbox": [ + 395, + 294, + 417, + 304 + ], + "type": "text", + "content": "HFIN" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 435, + 246, + 485, + 291 + ], + "blocks": [ + { + "bbox": [ + 435, + 246, + 485, + 291 + ], + "lines": [ + { + "bbox": [ + 435, + 246, + 485, + 291 + ], + "spans": [ + { + "bbox": [ + 435, + 246, + 485, + 291 + ], + "type": "image", + "image_path": "9310c22fa58de36945cb88cde5aa81f9c634aeb7803e8cd8ea8841e6303ad23b.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 440, + 294, + 481, + 304 + ], + "lines": [ + { + "bbox": [ + 440, + 294, + 481, + 304 + ], + "spans": [ + { + "bbox": [ + 440, + 294, + 481, + 304 + ], + "type": "text", + "content": "FAME-Net" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 489, + 246, + 539, + 291 + ], + "blocks": [ + { + "bbox": [ + 489, + 246, + 539, + 291 + ], + "lines": [ + { + "bbox": [ + 489, + 246, + 539, + 291 + ], + "spans": [ + { + "bbox": [ + 489, + 246, + 539, + 291 + ], + "type": "image", + "image_path": "0610e756f4fc676d5c1e1c351a1e8801b131cf8cd72021c3f28973e6435a2457.jpg" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 506, + 294, + 523, + 304 + ], + "lines": [ + { + "bbox": [ + 506, + 294, + 523, + 304 + ], + "spans": [ + { + "bbox": [ + 506, + 294, + 523, + 304 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_caption" + } + ], + "index": 40 + }, + { + "type": "image", + "bbox": [ + 540, + 242, + 552, + 305 + ], + "blocks": [ + { + "bbox": [ + 540, + 242, + 552, + 305 + ], + "lines": [ + { + "bbox": [ + 540, + 242, + 552, + 305 + ], + "spans": [ + { + "bbox": [ + 540, + 242, + 552, + 305 + ], + "type": "image", + "image_path": "5526a513eba49cfd46396bac07528bedb800c2e7411d2d5547b33a5ad4d030f8.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_body" + } + ], + "index": 42 + }, + { + "type": "table", + "bbox": [ + 56, + 369, + 555, + 514 + ], + "blocks": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 555, + 361 + ], + "type": "text", + "content": "Table 1. Experimental results of all the competing methods on the three benchmark datasets. The best and the second best values are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 369, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 56, + 369, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 369, + 555, + 514 + ], + "type": "table", + "html": "
MethodsParams (M)WorldView IIWorldView IIIGaoFen2
PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓PSNR ↑SSIM ↑SAM ↓ERGAS ↓
Brovey-35.86460.92160.04031.823822.50600.54660.11598.233137.79740.90260.02181.3720
IHS-35.29620.90270.04612.027822.55790.53540.12668.361638.17540.91000.02431.5336
GS-35.63760.91760.04231.877422.56080.54700.12178.243337.22600.90340.03091.6736
PNN0.068940.75500.96240.02591.064629.94180.91210.08243.320643.12080.97040.01720.8528
MSDCNN0.239041.33550.96640.02420.99430.30380.91840.07823.188445.68740.98270.01350.6389
GPPNN0.119841.16220.96840.02441.031530.17850.91750.07763.259344.21450.98150.01370.7361
MutNet0.071441.67730.97050.02240.951930.49070.92230.07493.112547.30420.98920.01020.5481
INNformer0.070641.69030.97040.02270.951430.53650.92250.07473.099747.35280.98930.01020.5479
HFIN0.077242.23190.97140.02150.880730.61470.92030.07423.078648.87830.98980.00930.4591
FAME-Net0.140842.02620.97230.02150.917230.99030.92870.06972.953147.67210.98980.00980.5242
Ours0.038742.48460.97380.02080.855431.17660.93320.05212.576149.14230.99010.00900.4474
", + "image_path": "ed6bc0c71bc31dee74ec77590c850d24b53eb30a250961e460bc24fcba9d40e9.jpg" + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "table_body" + } + ], + "index": 45 + }, + { + "bbox": [ + 55, + 533, + 137, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 137, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 137, + 545 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 194, + 563 + ], + "type": "text", + "content": "4.1. Dataset and benchmarks" + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "We conduct our experiments on three classic datasets. WorldView-II (WV2), Gaofen2 (GF2) and WorldView-III (WV3). The WorldView-II dataset consists of 760 image pairs for training and 80 image pairs for testing. The WorldView-III dataset includes 2150 image pairs for training and 200 image pairs for testing. The GaoFen-2 dataset comprises 2712 image pairs for training and 200 image pairs for testing. To validate the effectiveness of our approach, we compare it with several state-of-the-art pansharpening methods, including PNN [23], MSDCNN [42], GPPNN [39], MutNet [49], INNformer [47], HFIN [33] and FAME-Net [16], as well as traditional pansharpening meth" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 314, + 533, + 510, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 533, + 510, + 545 + ], + "spans": [ + { + "bbox": [ + 314, + 533, + 510, + 545 + ], + "type": "text", + "content": "ods such as GS [20], Brovey [12], and IHS [14]." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "In our experiments, all deep learning models are implemented using PyTorch and trained on an NVIDIA GeForce GTX 3090 GPU. For each set, the multispectral (MS) images are cropped into patches with the size of " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ", and the corresponding panchromatic (PAN) images are of size " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": ". We employ common evaluation metrics, including PSNR, SSIM, SAM, and ERGAS. Additionally, we utilize three widely-used no-reference IQA metrics for real-world full-resolution scenes: " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "D_{\\lambda}, D_{S}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and QNR. More descriptions of the implementation details and dataset are provided in the supplementary material." + } + ] + } + ], + "index": 50 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10914" + } + ] + } + ], + "index": 51 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 91, + 294, + 122 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 280, + 81 + ], + "type": "text", + "content": "Table 2. Non-reference metrics on full-resolution dataset." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 91, + 294, + 122 + ], + "lines": [ + { + "bbox": [ + 56, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 294, + 122 + ], + "type": "table", + "html": "
PNNMSDCNNGPPNNINNformerMutNetHFINFAME-NetOurs
Dλ↓0.07460.07340.07820.06970.06940.07100.06740.0566
Ds↓0.11640.11510.12530.11280.11180.10980.11210.1055
QNRT0.81910.82150.80730.82530.82470.82610.82910.8447
", + "image_path": "ccf3eeed65c27bc98634f20f00894c79462a82ff6ad8bf60c81437e988b106fa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 276, + 155 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 276, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 276, + 155 + ], + "type": "text", + "content": "4.2. Comparison with state-of-the-art methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "text", + "content": "Evaluation on reduced-resolution scene. To evaluate the effectiveness of our method, we conducted experiments on three datasets and compared it with ten other state-of-the-art pansharpening methods. Table 1 summarizes the evaluation metrics for the three datasets, where the best result for each metric is highlighted in red and the second-best result in blue. The results demonstrate that our method significantly outperforms the others across all evaluation metrics, achieving the best performance on all three datasets. Compared to the second-best method in terms of PSNR for each dataset, our method shows an improvement of " + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "inline_equation", + "content": "0.25\\mathrm{dB}" + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "text", + "content": " on the WV2 dataset, " + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "inline_equation", + "content": "0.18\\mathrm{dB}" + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "text", + "content": " on the GF2 dataset, and " + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "inline_equation", + "content": "0.27\\mathrm{dB}" + }, + { + "bbox": [ + 56, + 159, + 296, + 410 + ], + "type": "text", + "content": " on the WV3 dataset, indicating the consistency of our method with the true high-resolution images. We also conducted qualitative experiments, Figures 4 present samples from the WV2 datasets. The top two rows show the comparison of experimental results, while the last row displays the mean squared error maps of each network's results compared to the ground truth. It is evident that the results obtained by our network are closest to the ground truth, with the smallest discrepancies." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": "Evaluation on full-resolution scene. We also conduct full-resolution analysis in real-world scenarios to validate the generalization capability of our method. We perform experiments on an additional 200 sets of the GF2 dataset. Due to the absence of high-resolution multispectral images for actual real-world scenarios, we employ three commonly used no-reference metrics, " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "D_{\\lambda}" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 55, + 411, + 296, + 531 + ], + "type": "text", + "content": ", and QNR for evaluation. The experimental results, as shown in Table 2, demonstrate that our method achieves the best performance across all three metrics." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 63, + 578, + 288, + 624 + ], + "blocks": [ + { + "bbox": [ + 55, + 540, + 295, + 574 + ], + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 574 + ], + "type": "text", + "content": "Table 3. The results of different configurations on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 578, + 288, + 624 + ], + "lines": [ + { + "bbox": [ + 63, + 578, + 288, + 624 + ], + "spans": [ + { + "bbox": [ + 63, + 578, + 288, + 624 + ], + "type": "table", + "html": "
ConfigurationQGSIQLSAQSSIPSNR ↑SSIM ↑SAM ↓ERGAS ↓
IX✓✓41.73560.96730.02470.9337
II✓X✓42.02170.97060.02190.9134
III✓✓X42.17340.97130.02210.9012
Ours✓✓✓42.48460.97380.02080.8554
", + "image_path": "9f235abea61f798aa9dac7479be34653e46a3344647f40c4c7e0acbc7adb6cd5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 647, + 182, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 182, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 182, + 661 + ], + "type": "text", + "content": "4.3. Ablation Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "We perform an ablation study using the WorldView-II dataset to further demonstrate the effectiveness of our method. The quaternion global spectral interaction branch (QGSI), quaternion local spatial structure awareness branch" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 113, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 555, + 105 + ], + "type": "text", + "content": "Table 4. PSNR values of Our model with different number of blocks on WorldViewII. The best and the second best values are highlighted in bold and underline, respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 113, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 316, + 113, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 113, + 553, + 196 + ], + "type": "table", + "html": "
Stages (K)PSNR ↑SSIM ↑SAM ↓ERGAS ↓
141.83780.96030.02490.9879
241.90020.96470.02380.9725
342.09130.96980.02340.9578
442.23690.97240.02150.9007
542.48460.97380.02080.8554
", + "image_path": "0c8151d29cd07d8f843456ce1fd88d494c780da07f63c9d5c3c5708523fb9462.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 275 + ], + "type": "text", + "content": "(QLSA), and quaternion spatial-spectral interaction branch (QSSI) are the core components of the proposed method, and we conduct separate ablations for each. Additionally, we also explore the impact of the number of basic building blocks used in the network through ablation studies." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 275, + 554, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 554, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 554, + 371 + ], + "type": "text", + "content": "To validate the effectiveness of the three key quaternion components, QGSI, QLSA, and QSSI, we replace their quaternion operations with real-valued operations, as shown in Table 3. The results demonstrate that removing each quaternion component leads to a decline in performance, thus confirming the effectiveness of the quaternion's multidimensional representation and parameter-sharing mechanism for pansharpening." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 554, + 479 + ], + "type": "text", + "content": "We conduct experiments on networks with varying numbers of core building blocks to evaluate the impact of the number of core modules, as shown in Table 4. As observed, the model performance improves with the increase in the number of Quaternion Spatial-Spectral Interaction Blocks (QSSIB), though this comes at the cost of increased resource consumption. We balance performance and resource consumption by using the default setting of 5. More ablation experiments are shown in the supplementary materials." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we propose QuatPanNet, a novel quaternion-based spatial-spectral interaction network that enhances pansharpening by leveraging quaternions' compact representation for high-dimensional data. Our work demonstrates the advantages of a unified quaternion framework for effectively addressing the pansharpening problem. We custom-design quaternion-based spectral and spatial interaction modules: the former promotes global information interaction while preserving spectral dependencies, and the latter fosters spatial interaction while maintaining local structural dependencies. Building on these, we integrate the dual-path features into a unified quaternion representation. By utilizing the compact and efficient representation of PAN and MS image data in the quaternion space, QuatPanNet facilitates effective interactions between spatial and spectral features. Our method achieves exceptional fusion quality with a reduced number of parameters." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10915" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 127 + ], + "type": "text", + "content": "This work is supported by the National Natural Science Foundation of China (NSFC) under Grants 62225207, 62436008, 62422609 and 62276243." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 201 + ], + "type": "text", + "content": "[1] Paolo Addesso, Gemine Vivone, Rocco Restaino, and Jocelyn Chanussot. A data-driven model-based regression applied to panchromatic sharpening. IEEE Transactions on Image Processing, 29:7779-7794, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 296, + 245 + ], + "type": "text", + "content": "[2] Nicholas A Asparagus and John K Dimitros. A comparative study of three methods for robot kinematics. IEEE Transactions on Systems, Man, and Cybernetics, Part B (Cybernetics), 28(2):135-145, 1998. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 296, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 296, + 300 + ], + "type": "text", + "content": "[3] Qinglong Cao, Zhengqin Xu, Yuntian Chen, Chao Ma, and Xiaokang Yang. Domain prompt learning with quaternion networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26637-26646, 2024. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 302, + 296, + 345 + ], + "type": "text", + "content": "[4] Ping Cui, Cui-Ming Zou, Bang-Ju Wang, and Feng Shi. Color image fusion in quaternion spectral domain. In 2023 International Conference on Wavelet Analysis and Pattern Recognition (ICWAPR), pages 56-61. IEEE, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 390 + ], + "type": "text", + "content": "[5] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. IEEE transactions on pattern analysis and machine intelligence, 38(2):295-307, 2015. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 296, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 296, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 296, + 424 + ], + "type": "text", + "content": "[6] Todd A Ell, Nicolas Le Bihan, and Stephen J Sangwine. *Quaternion Fourier transforms for signal and image processing*. John Wiley & Sons, 2014. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 296, + 479 + ], + "type": "text", + "content": "[7] Matteo Frigo and Steven G Johnson. Fftw: An adaptive software architecture for the fft. In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), pages 1381-1384. IEEE, 1998. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 480, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 533 + ], + "type": "text", + "content": "[8] Xueyang Fu, Zihuang Lin, Yue Huang, and Xinghao Ding. A variational pan-sharpening with local gradient constraints. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10265-10274, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 579 + ], + "type": "text", + "content": "[9] Xueyang Fu, Wu Wang, Yue Huang, Xinghao Ding, and John Paisley. Deep multiscale detail networks for multiband spectral image sharpening. IEEE Transactions on Neural Networks and Learning Systems, 32(5):2090-2104, 2020. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 296, + 613 + ], + "type": "text", + "content": "[10] Chase J Gaudet and Anthony S Maida. Deep quaternion networks. In 2018 International Joint Conference on Neural Networks (IJCNN), pages 1-8. IEEE, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 614, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 296, + 689 + ], + "type": "text", + "content": "[11] Pedram Ghamisi, Behnood Rasti, Naoto Yokoya, Qunming Wang, Bernhard Hofle, Lorenzo Bruzzone, Francesca Bovolo, Mingmin Chi, Katharina Anders, Richard Gloaguen, et al. Multisource and multitemporal data fusion in remote sensing: A comprehensive review of the state of the art. IEEE Geoscience and Remote Sensing Magazine, 7(1):6-39, 2019. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[12] Alan R Gillespie, Anne B Kahle, and Richard E Walker. Color enhancement of highly correlated images. ii. channel" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "ratio and \"chromaticity\" transformation techniques. Remote Sensing of Environment, 22(3):343-365, 1987. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 555, + 193 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 193 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 193 + ], + "type": "text", + "content": "[13] William Rowan Hamilton. Lectures on Quaternions: Containing a Systematic Statement of a New Mathematical Method; of which the Principles Were Communicated in 1843 to the Royal Irish Academy; and which Has Since Formed the Subject of Successive Courses of Lectures, Delivered in 1848 and Subsequent Years, in the Halls of Trinity College, Dublin: with Numerous Illustrative Diagrams, and with Some Geometrical and Physical Applications. Hodges and Smith, 1853. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 554, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 554, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 554, + 238 + ], + "type": "text", + "content": "[14] R Haydn. Application of the ihs color transform to the processing of multisensor data and image enhancement. In Proc. of the International Symposium on Remote Sensing of Arid and Semi-Arid Lands, Cairo, Egypt, 1982, 1982. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 554, + 282 + ], + "type": "text", + "content": "[15] Xiyan He, Laurent Condat, José M Bioucas-Dias, Jocelyn Chanussot, and Junshi Xia. A new pansharpening method based on spatial and spectral sparsity priors. IEEE Transactions on Image Processing, 23(9):4160-4174, 2014. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 327 + ], + "type": "text", + "content": "[16] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2121-2129, 2024. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 382 + ], + "type": "text", + "content": "[17] Jie Huang, Yajing Liu, Feng Zhao, Keyu Yan, Jinghao Zhang, Yukun Huang, Man Zhou, and Zhiwei Xiong. Deep fourier-based exposure correction network with spatial-frequency interaction. In European Conference on Computer Vision, pages 163-180. Springer, 2022. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 415 + ], + "type": "text", + "content": "[18] Creed F Jones III. Color face recognition using quaternionic Gabor filters. PhD thesis, Virginia Polytechnic Institute and State University, 2005. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 555, + 469 + ], + "type": "text", + "content": "[19] Hiromi Kusamichi, Teijiro Isokawa, Nobuyuki Matsui, Yuzo Ogawa, and Kazuaki Maeda. A new scheme for color night vision by quaternion neural network. In Proceedings of the 2nd international conference on autonomous robots and agents. CiteSeer, 2004. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 554, + 504 + ], + "type": "text", + "content": "[20] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening, 2000. US Patent 6,011,875. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 505, + 554, + 557 + ], + "type": "text", + "content": "[21] Pengfei Liu, Liang Xiao, and Tao Li. A variational pan-sharpening method based on spatial fractional-order geometry and spectral-spatial low-rank priors. IEEE Transactions on Geoscience and Remote Sensing, 56(3):1788-1802, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 613 + ], + "type": "text", + "content": "[22] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 647 + ], + "type": "text", + "content": "[23] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. Remote Sensing, 8(7):594, 2016. 1, 2, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 712 + ], + "type": "text", + "content": "[24] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE Geoscience and Remote Sensing Magazine, 9(1):18-52, 2020. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10916" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[25] Jorge Nunez, Xavier Otazu, Octavi Fors, Albert Prades, Vicenc Pala, and Roman Arbiol. Multiresolution-based image fusion with additive wavelet decomposition. IEEE Transactions on Geoscience and Remote sensing, 37(3):1204-1211, 1999. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[26] Titouan Parcollet, Mirco Ravanelli, Mohamed Morchid, Georges Linares, Chiheb Trabelsi, Renato De Mori, and Yoshua Bengio. Quaternion recurrent neural networks. arXiv preprint arXiv:1806.04418, 2018. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[27] Titouan Parcollet, Ying Zhang, Mohamed Morchid, Chiheb Trabelsi, Georges Linares, Renato De Mori, and Yoshua Bengio. *Quaternion convolutional neural networks for end-to-end automatic speech recognition*. arXiv preprint arXiv:1806.07789, 2018. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[28] Soo-Chang Pei, Jian-Jiun Ding, and Ja-Han Chang. Efficient implementation of quaternion fourier transform, convolution, and correlation by 2-d complex FFT. IEEE transactions on signal processing, 49(11):2783-2797, 2001. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[29] Thierry Ranchin and Lucien Wald. Fusion of high spatial and spectral resolution images: The arsis concept and its implementation. Photogrammetric engineering and remote sensing, 66(1):49-61, 2000. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 353 + ], + "type": "text", + "content": "[30] Stephen John Sangwine. Fourier transforms of colour images using quaternion or hypercomplex, numbers. *Electronics letters*, 32(21):1979–1980, 1996. 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 398 + ], + "type": "text", + "content": "[31] Vijay P Shah, Nicolas H Younan, and Roger L King. An efficient pan-sharpening method via a combined adaptive pca approach and contourlets. IEEE transactions on geoscience and remote sensing, 46(5):1323-1335, 2008. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 431 + ], + "type": "text", + "content": "[32] Lilong Shi and Brian Funt. Quaternion color texture segmentation. Computer Vision and image understanding, 107(1-2): 88-96, 2007. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 509 + ], + "type": "text", + "content": "[33] Jiangtong Tan, Jie Huang, Naishan Zheng, Man Zhou, Keyu Yan, Danfeng Hong, and Feng Zhao. Revisiting spatial-frequency information integration from a hierarchical perspective for panchromatic and multi-spectral image fusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25922-25931, 2024. 3, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 544 + ], + "type": "text", + "content": "[34] Te-Ming Tu, Shun-Chi Su, Hsuen-Chyun Shyu, and Ping S Huang. A new look at ihs-like image fusion methods. Information fusion, 2(3):177-186, 2001. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 599 + ], + "type": "text", + "content": "[35] Gemine Vivone, Rocco Restaino, Mauro Dalla Mura, Giorgio Licciardi, and Jocelyn Chanussot. Contrast and error-based fusion schemes for multispectral image pansharpening. IEEE Geoscience and Remote Sensing Letters, 11(5): 930-934, 2013. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[36] Dong Wang, Yunpeng Bai, and Ying Li. Multispectral pan-sharpening via dual-channel convolutional network with convolutional LSTM based hierarchical spatial-spectral feature fusion, 2020. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[37] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE Transactions on Geoscience and Remote Sensing, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[38] Chen Wu, Bo Du, Xiaohui Cui, and Liangpei Zhang. A post-classification change detection method based on itera" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tive slow feature analysis and bayesian soft fusion. Remote Sensing of Environment, 199:241-255, 2017. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 148 + ], + "type": "text", + "content": "[39] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1366-1375, 2021. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 205 + ], + "type": "text", + "content": "[40] Gang Yang, Xiangyong Cao, Wenzhe Xiao, Man Zhou, Aiping Liu, Xun Chen, and Deyu Meng. Panflownet: A flow-based deep network for pan-sharpening. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16857-16867, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 249 + ], + "type": "text", + "content": "[41] Licheng Yu, Yi Xu, Hongteng Xu, and Hao Zhang. Quaternion-based sparse representation of color image. In 2013 IEEE International Conference on Multimedia and Expo (ICME), pages 1-7. IEEE, 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 315 + ], + "type": "text", + "content": "[42] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 11(3):978-989, 2018. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[43] Xiaohui Yuan, Jianfang Shi, and Lichuan Gu. A review of deep learning methods for semantic segmentation of remote sensing imagery. Expert Systems with Applications, 169: 114417, 2021. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 403 + ], + "type": "text", + "content": "[44] Rui Zeng, Jiasong Wu, Zhuhong Shao, Yang Chen, Beijing Chen, Lotfi Senhadji, and Huazhong Shu. Color image classification via quaternion principal component analysis network. Neurocomputing, 216:416-428, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 447 + ], + "type": "text", + "content": "[45] Hao Zhang and Jiayi Ma. Gtp-pnet: A residual learning network based on gradient transformation prior for pansharpening. ISPRS Journal of Photogrammetry and Remote Sensing, 172:223-239, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[46] Yunzuo Zhang, Jiayu Zhang, Ruixue Liu, Pengfei Zhu, and Yameng Liu. Key frame extraction based on quaternion fourier transform with multiple features fusion. Expert Systems with Applications, 216:119467, 2023. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 546 + ], + "type": "text", + "content": "[47] Man Zhou, Jie Huang, Yanchi Fang, Xueyang Fu, and Aiping Liu. Pan-sharpening with customized transformer and invertible neural network. In Proceedings of the AAAI conference on artificial intelligence, pages 3553-3561, 2022. 2, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 601 + ], + "type": "text", + "content": "[48] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In European conference on computer vision, pages 274-291. Springer, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 646 + ], + "type": "text", + "content": "[49] Man Zhou, Keyu Yan, Jie Huang, Zihe Yang, Xueyang Fu, and Feng Zhao. Mutual information-driven pan-sharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1798-1808, 2022. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[50] Man Zhou, Jie Huang, Chun-Le Guo, and Chongyi Li. FOurmer: An efficient global modeling paradigm for image restoration. In International conference on machine learning, pages 42589-42601. PMLR, 2023. 4, 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[51] Xuanyu Zhu, Yi Xu, Hongteng Xu, and Changjian Chen. Quaternion convolutional neural networks. In Proceedings of" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10917" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 152 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 296, + 94 + ], + "type": "text", + "content": "the European conference on computer vision (ECCV), pages 631-647, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 152 + ], + "type": "text", + "content": "[52] Zeyu Zhu, Xiangyong Cao, Man Zhou, Junhao Huang, and Deyu Meng. Probability-based global cross-modal upsampling for pansharpening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14039-14048, 2023. 3" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10918" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_content_list.json b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24ace71ad9f9d08e6decec5948c78b05ad951393 --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_content_list.json @@ -0,0 +1,1583 @@ +[ + { + "type": "text", + "text": "Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling", + "text_level": 1, + "bbox": [ + 148, + 130, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zenghao Niu $^{1}$ Weicheng Xie $^{1,2,3}$ * Siyang Song $^{4}$ Zitong Yu $^{5}$ Feng Liu $^{1}$ Linlin Shen $^{6}$", + "bbox": [ + 138, + 202, + 893, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Computer Science & Software Engineering, Shenzhen University, China", + "bbox": [ + 191, + 220, + 851, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China", + "bbox": [ + 148, + 238, + 890, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Guangdong Provincial Key Laboratory of Intelligent Information Processing, Shenzhen University, China", + "bbox": [ + 102, + 256, + 939, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{4}$ School of Computer Science, University of Exeter, U.K.", + "bbox": [ + 295, + 273, + 745, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{5}$ Department of Computing and Information Technology, Great Bay University, China", + "bbox": [ + 179, + 291, + 857, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6Computer Vision Institute, School of Artificial Intelligence, Shenzhen University, China", + "bbox": [ + 171, + 309, + 870, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{2300271072@mail., wcxie@, feng.liu@, llshen@}szu.edu.cn", + "bbox": [ + 269, + 329, + 763, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "s song@exeter.ac.uk, zitong.yu@ieee.org", + "bbox": [ + 341, + 348, + 686, + 361 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 397, + 326, + 412 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attacks present a critical challenge to deep neural networks' robustness, particularly in transfer scenarios across different model architectures. However, the transferability of adversarial attacks faces a fundamental dilemma between Exploitation (maximizing attack potency) and Exploration (enhancing cross-model generalization). Traditional momentum-based methods over-prioritize Exploitation, i.e., higher loss maxima for attack potency but weakened generalization (narrow loss surface). Conversely, recent methods with inner-iteration sampling over-prioritize Exploration, i.e., flatter loss surfaces for cross-model generalization but weakened attack potency (suboptimal local maxima). To resolve this dilemma, we propose a simple yet effective Gradient-Guided Sampling (GGS), which harmonizes both objectives through guiding sampling along the gradient ascent direction to improve both sampling efficiency and stability. Specifically, based on MI-FGSM, GGS introduces inner-iteration random sampling and guides the sampling direction using the gradient from the previous inner-iteration (the sampling's magnitude is determined by a random distribution). This mechanism encourages adversarial examples to reside in balanced regions with both flatness for cross-model generalization and higher local maxima for strong attack potency. Comprehensive experiments across multiple DNN architectures and multimodal large language models (MLLMs) demonstrate the superiority of our method over state-of-the-art transfer attacks. Code is made available at https://github.com/anuin-cat/GGS.", + "bbox": [ + 89, + 420, + 483, + 859 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5ce4302b0c4f1e496a52dafe22ed071a9152d827637a64291246dde7d2d47fb4.jpg", + "image_caption": [ + "Figure 1. The loss surfaces of (a) MI-FGSM [1] (Momentum iterative fast gradient sign method), (b) RS (Base inner-iteration Random Sampling defined in section 3.2) for enhancing exploration within the neighborhood, (c) PGN [7] (Penalizing Gradient Norm) with Random Sampling (RS) for stable gradient estimation to enhance exploration, and (d) Our GGS (Gradient-Guided Sampling) for efficient sampling to generate gradients that stably towards flat regions with higher local maxima. It shows that, building upon RS, our approach not only maintains a flat loss surface, but also improves the local maximum loss value, compared to PGN, achieving a balance between exploration and exploitation." + ], + "image_footnote": [], + "bbox": [ + 516, + 395, + 910, + 648 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 840, + 645, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Researches have indicated that adding imperceptible perturbations to input data can easily affect deep neural networks", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3885", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(DNNs), leading to erroneous predictions in safety-critical domains such as autonomous driving [13, 35], and cybersecurity [15, 28]. To facilitate the development of relevant defense mechanisms, in-depth research on adversarial attack methods is urgently needed.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While adversarial attacks are categorized into white-box and black-box settings, the black-box scenario is more realistic in practical applications, where the attacker lacks knowledge of the target model and relies solely on a surrogate model, aiming to generate adversarial examples that are effective across multiple target models. To enhance the transferability of adversarial examples in this setting, researchers have explored gradient-based methods [1, 7, 32], input transformation methods [2, 33, 37], and generative methods [16, 20, 36] to mitigate overfitting of adversarial examples on surrogate models. Despite significant progress, the effectiveness of attacks on target models remains substantially lower than that on surrogate models, driving the need for more effective attack strategies.", + "bbox": [ + 89, + 169, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In black-box adversarial attacks, RAP [24] adopted Sharpness-Aware Minimization (SAM [6]) to improve transferability by flattening the loss surface. Building on this, PGN [7] further enhanced the transferability of adversarial examples by penalizing gradient norm and flattening the loss surface. The efficacy of flat loss surface in improving adversarial attack transferability is increasingly supported [4, 7, 38]. To enhance exploration through local flatness, recent approaches, building on MI-FGSM's [1] momentum-based outer-iterations, often incorporates inner-iterations before each example update [7, 40], i.e., performing multi-step neighborhood sampling for stabilization.", + "bbox": [ + 89, + 382, + 482, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these methods enhance Exploration (cross-model generalization) at the cost of compromising Exploitation (attack potency). This is because, although inner-iterations can significantly enhance exploration capability through neighborhood sampling, independent random sampling struggles to obtain stable and effective gradient directions, while potential noise may cause the averaged gradient to deviate from the optimal ascent toward target regions, which exhibit flatter loss surface and higher local maxima. To optimize sampling efficiency, a guided sampling constraint is essential, which should satisfy two criteria: (1) ensuring stable gradient ascent direction and (2) directing gradients toward flatter regions, thereby enabling samples to reach flatter local maxima regions with enhanced transferability, balancing exploitation and exploration.", + "bbox": [ + 89, + 580, + 482, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For (1) stable gradient direction, a sufficiently small neighborhood sampling followed by gradient calculation generally yields correct ascent directions. Yet, how can we further stabilize this direction? The Nesterov Accelerated Gradient (NAG) [23] enhances stability by firstly updating samples along the current momentum (lookahead)", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "before acquiring gradients, providing more stable directions. Therefore, momentum-based direction constraints are promising for sampling, as the additional random magnitude can be used to preserve the sampling randomness. (2) However, directly applying NAG's momentum mechanism during inner-iteration sampling imposes strong constraints on subsequent directions, restricts the exploration capability for flat regions. To address this, we replace momentum with the gradient from the previous inner-iteration, named Gradient-Guided Sampling (GGS). This adjustment ensures that sampling to align with gradient ascent directions while reducing the impact from early-stage instability. Our GGS can achieve stable sampling directions after brief oscillations, consistently targeting flat loss regions with larger local maxima regardless of the initial sampling quality.", + "bbox": [ + 511, + 90, + 903, + 318 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As shown in Fig. 1(d), our GGS generates adversarial examples with flatter loss surface and higher local maxima compared to RS, i.e., it does not sacrifice local maxima in the pursuit of a flatter surface compared to PGN, thereby achieving a balance between exploration and exploitation to facilitate attack transferability. Additionally, since our approach represents a complementary improvement over Random Sampling (RS), it is compatible with RS-based methods, such as inner-iteration random sampling and gradient averaging, including PGN [7] and GRA [40]. In summary, this study makes the following contributions:", + "bbox": [ + 511, + 319, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose Gradient-Guided Sampling (GGS), a novel inner-iteration sampling strategy that effectively balances exploration (cross-model generalization) and exploitation (attack potency). By leveraging the gradient from the previous inner-iteration to guide sampling directions, GGS achieves stable gradient ascent toward flat loss regions with higher local maxima.", + "- The GGS framework is compatible with existing inner-iteration random sampling-based methods, enhancing their sampling efficiency and further improving the transferability of the adversarial examples they generate.", + "- Extensive experimental results as well as comprehensive visualizations demonstrate the efficacy of our GGS, covering targeted and non-targeted attacks in cross-architecture black-box scenarios, non-targeted attacks on multimodal large language models (MLLMs) and commercial cloud functions." + ], + "bbox": [ + 513, + 491, + 903, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 767, + 653, + 782 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transferability is a critical property in adversarial attacks, allowing adversarial examples generated from a surrogate model to deceive unseen target models in black-box settings. This characteristic makes transfer-based attacks highly practical in real-world applications where target models are inaccessible, motivating our focus on transferable adversarial attacks.", + "bbox": [ + 511, + 794, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3886", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Transferable Attack", + "text_level": 1, + "bbox": [ + 89, + 90, + 282, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gradient-based Attack. Gradient-based attacks aim to generate adversarial examples by exploiting the gradients of a surrogate model to maximize the loss. Classic methods such as Fast Gradient Sign Method (FGSM) [9] and Projected Gradient Descent (PGD) [22] optimize adversarial perturbations to fool the surrogate model. However, these attacks often overfit the surrogate model, resulting in limited transferability to target models. To address this, advanced techniques such as momentum-based (MI-FGSM [1]), Nesterov Iterative-based (NI-FGSM [19]), variance-tuned (VMI-FGSM [32]), gradient relevance-based (GRA [40]), momentum initialization (GI-FGSM [30]) and distribution-based methods (ANDA [5]) have been introduced to smooth out the optimization process and improve transferability across models.", + "bbox": [ + 89, + 112, + 483, + 338 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Input Transformation. To mitigate the overfitting issue inherent in gradient-based methods, input transformation techniques have been proposed to enhance the diversity of adversarial examples. Techniques like translation, resizing, padding, image mixup (e.g., DIM [37], SIM [19], TIM [2], Admix [33]) and other advanced image transformation methods (e.g., SSM [21], SIA [34], STM [8], L2T [41], BSR [31]) introduce randomness to the input data before generating perturbations. By applying such transformations during the attack process, more effective transferable adversarial examples can be generated.", + "bbox": [ + 89, + 339, + 483, + 505 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Flat Maxima", + "text_level": 1, + "bbox": [ + 89, + 515, + 225, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The generalization ability of models has been suggested to possibly have a certain association with flat minima [12]. More in-depth research and exploration have gradually confirmed this perspective [6, 17, 18, 39]. Moreover, flat maxima also have been validated to be effective for enhancing the generalization and transferability of adversarial examples [24]. Flat maxima refers to regions in the loss landscape, where small changes in the model parameters result in minimal changes in the loss, making adversarial examples generated in these regions less sensitive to the specific decision boundaries of models. This property contrasts with \"sharp maxima\", which are highly sensitive to small perturbations and can lead to overfitting to the surrogate model, reducing transferability capacity.", + "bbox": [ + 89, + 537, + 482, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the context of transfer-based attacks, works like Reverse Adversarial Perturbation (RAP) [24] leverage flat maxima to generate adversarial examples, PGN [7] adopts a first-order procedure to approximate the Hessian/vector product, largely improving computational efficiency. They are not only effective on the surrogate model but also more resilient when attacking unseen target models. These methods mitigate overfitting, via making adversarial examples lie in regions where the loss function remains stable across different models (e.g., flat maxima regions), improving at", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tack success rates in black-box settings. However, achieving flat maxima requires not only attention to \"flatness\" but also preservation of the \"maxima\" to ensure sufficient attack strength during transfer attack, which poses a significant challenge.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Inner-Iteration Sampling", + "text_level": 1, + "bbox": [ + 511, + 176, + 743, + 193 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To enhance the generalization capability of adversarial examples, recent methods have introduced inner-iteration sampling [7, 24, 32]. Specifically, an inner iterative process is inserted before updating the examples. This process involves neighborhood sampling to acquire domain information during inner-iterations which are used to adjust gradients [32] or neighborhood search is incorporated to locate local minima and enhance their loss values [24]. Furthermore, works like [7, 40] enhance the flatness of the loss landscape by averaging gradients from inner-iterations to update the examples.", + "bbox": [ + 511, + 199, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, neighborhood sampling can facilitate examples to escape from sharp local maxima regions, thereby avoiding overfitting to the surrogate model. However, directly averaging the gradients from inner-iterations will reduce the maximum value of the loss surface, since random sampling produces unstable gradient directions that fail to consistently direct to flat regions with higher local maxima. Thus, enhancing inner-iteration sampling efficiency to stabilize the final gradient has become a critical challenge.", + "bbox": [ + 511, + 366, + 905, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 516, + 648, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 541, + 653, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Transferable Adversarial Attacks. For a given example $(x,y)$ , a surrogate model $f_{\\theta}$ and multiple target models $f_{\\varphi_k}$ for $k \\in \\{1,2,\\dots,K\\}$ , the attacker's goal is to find $x^{adv} = x + \\delta$ using $f_{\\theta}$ to make $f_{\\varphi_k}(x^{adv}) \\neq y$ , where $\\| \\delta \\|_p < \\epsilon$ , $\\epsilon$ denotes the maximum magnitude of the specified perturbation and $\\| \\cdot \\|_p$ denotes the $\\ell_p$ -norm. For targeted attacks, we simply adapt the symbols and labels in the target function: $f_{\\varphi_k}(x^{adv}) = y_t$ , where $y_t$ denotes the target label. Meanwhile, to facilitate a clearer representation of the norm constraints against $x^{adv}$ as below, we define $\\mathcal{B}_{\\epsilon}(x) = \\{x': \\| x' - x \\|_p \\leq \\epsilon\\}$ to denote the $\\epsilon$ -ball of an input image $x$ . For the purpose of the above attack, we generally need to maximize the following objective function to generate the transferable adversarial examples:", + "bbox": [ + 511, + 564, + 906, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\max _ {x ^ {a d v} \\in \\mathcal {B} _ {\\epsilon} (x)} \\mathcal {L} \\left(x ^ {a d v}, y\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 629, + 786, + 906, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{L}(\\cdot)$ represents the loss function. In the above flow, we assume that $f_{\\theta}$ is known as a white-box model and $f_{\\varphi_k}$ is unknown as a black-box model. Similar to the process of training a neural network, the white-box model can be seen as the training data, with the adversarial examples serving", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3887", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as the optimization parameters, while the black-box model acts as the test data.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Nesterov Accelerated Gradient (NAG). Due to the similarity between model training task and adversarial example generation, improving performance in black-box settings can be viewed as enhancing model generalization in model training tasks. Consequently, some commonly used optimization methods, such as Momentum [25] and Nesterov Accelerated Gradient (NAG) [23], can be introduced to improve the generalization of adversarial examples across black-box models through a more stable optimization process.", + "bbox": [ + 89, + 121, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "NI-FGSM [19] was the first to incorporate the principles of NAG into adversarial attack tasks. This enables the algorithm to acquire directional information in advance, reducing oscillations and achieving stable convergence. Specifically, momentum is computed at each step as follows:", + "bbox": [ + 89, + 273, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} = \\gamma \\cdot v _ {t - 1} + \\frac {\\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y)}{\\left\\| \\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y) \\right\\| _ {1}}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 359, + 482, + 404 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} _ {t} = \\underbrace {x _ {t - 1} ^ {a d v} + \\alpha \\cdot \\gamma \\cdot v _ {t - 1}} _ {\\text {l o o k a h e a d p o i n t}},\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 398, + 344, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $v_{t}$ represents the momentum term, $x_0^{adv} = x$ , $\\alpha$ stands for the step size, $\\gamma$ denotes the momentum decay factor, and $\\| \\cdot \\|_1$ denotes the $\\ell_1$ -norm of the given variable. Unlike traditional momentum-based methods, NI-FGSM utilizes the estimated position of examples rather than their current actual position, during gradient computation, i.e., introducing a lookahead mechanism into the algorithm.", + "bbox": [ + 89, + 450, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Motivation", + "text_level": 1, + "bbox": [ + 89, + 566, + 210, + 580 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recent transfer-based adversarial attack methods have introduced inner-iteration random sampling, to enhance algorithm stability [7, 40] or acquire neighborhood information [24, 32]. It can be noted that these methods have increased the flatness of the loss surface, as illustrated in Fig.1(c), with additional loss surface plots for other methods provided in the supplementary materials. If we isolate the inner-iteration random sampling (RS) component, it can be expressed in the following form:", + "bbox": [ + 89, + 590, + 483, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nv _ {t} = \\gamma v _ {t - 1} + \\frac {\\sum_ {i = 1} ^ {N} \\tilde {g} _ {i}}{\\left\\| \\sum_ {i = 1} ^ {N} \\tilde {g} _ {i} \\right\\| _ {1}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 738, + 482, + 787 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {g} _ {i} = \\nabla_ {x} \\mathcal {L} \\big (\\tilde {x} _ {i}, y \\big), \\text {w i t h} \\tilde {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + \\tilde {p}} _ {\\text {s a m p l i n g p o i n t}},\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 786, + 424, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\sum_{i=1}^{N}$ denotes the output of the inner-iteration process; $N$ denotes the number of inner-iterations; $\\gamma$ denotes the momentum decay factor; $\\tilde{g}_0, \\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)$ , are uniform random noises with the same size as $x$ .", + "bbox": [ + 89, + 837, + 483, + 898 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/69edcd1e097eb143568264f1d7471dc58e31a8c36fa14f47250086412fa7dfe5.jpg", + "image_caption": [ + "(a) Optimization Trajectories of Various Methods" + ], + "image_footnote": [], + "bbox": [ + 527, + 89, + 893, + 234 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9db4dea360f9239cc1a12af075c01f27d6c466212e25fa60152bf5ff3f53e839.jpg", + "image_caption": [ + "(b) 1st step process of RS", + "Figure 2. (a) Outer-Iteration processes of MI-FGSM (green), RS (purple) and our GGS (red). (i) MI-FGSM prioritizes rapid ascent directions, thereby enhancing exploitation and facilitating access to sharp local maxima regions. (ii) RS introduces inner-iterations and enhances exploration capability through neighborhood search, thereby enabling access to flat local maxima regions. (iii) GGS incorporates gradient constraints into RS, simultaneously enhancing both exploration and exploitation capabilities, thereby enabling the faster convergence to the centers of flat local maxima regions. (b) RS is performed within a neighborhood of the current example, and uses the average gradient of all sampled points as the final gradient. (c) Building upon RS, GGS enables examples to have stable gradient directions (② ③ ④) toward the centers of flat local maxima regions, following an initial period of brief oscillation (①). It uses the gradient direction from the previous inner-iteration as the guidance, while the randomness is maintained by setting the gradient magnitude with a random distribution." + ], + "image_footnote": [], + "bbox": [ + 526, + 252, + 702, + 385 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/73bc083d01e3f442ec4f2e455d7d7f8559ec04893f7b9c360b51a491edd85f45.jpg", + "image_caption": [ + "(c) 1st step process of GGS" + ], + "image_footnote": [], + "bbox": [ + 714, + 252, + 890, + 385 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 1, the inner-iteration RS possesses a strong capability to enhance the flatness of the loss surface compared to MI-FGSM. Compared to RS, current methods such as PGN [7] can further improve the flatness of the loss surface, however, they unwillingly reduce the local maxima of the loss surface, limiting its attack potential. This reduction in local maxima can be attributed to the unstable gradients generated by the completely random sampling strategy, as illustrated in Fig. 2(b). Specifically, the averaged gradient roughly points towards flat regions, while it struggles to consistently align with the stable direction to the center of flat maxima regions, thereby reducing updating efficiency.", + "bbox": [ + 511, + 684, + 906, + 866 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To acquire stable gradient ascent direction and improve the efficiency of inner-iteration sampling, we define two es", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3888", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3ec578967648e41ed05fb22a323a14cc8103e97a0e4306a015df76a59cb6252d.jpg", + "image_caption": [ + "(a) Random Sampling" + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 483, + 172 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/07765863c5e00293a3429595b4e2a249f1b2e931802bf9b57d29c27f29d916ca.jpg", + "image_caption": [ + "(b) Momentum-Guided Sampling" + ], + "image_footnote": [], + "bbox": [ + 94, + 189, + 483, + 266 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/edaa2140ae66b9d6d89b7006827312421d9d38d09d4611bc1f4c6f6722165bc2.jpg", + "image_caption": [ + "(c) Gradient-Guided Sampling", + "Figure 3. The differences among three inner-iteration sampling guidance methods: (a) Random sampling in Eq. 3, with each inner-iteration sampling being independent; (b) Momentum-guided sampling in Eq. 4, where sampling direction depends on the cumulative average of all previous gradients, creating long-chain dependencies, while using random sampling for maintaining the randomness; (c) Gradient-guided sampling in Eq. 5, where sampling direction relies solely on the gradient direction of the previous iteration, establishing single-step dependencies." + ], + "image_footnote": [], + "bbox": [ + 94, + 284, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sential characteristics that a well-balanced gradient direction should possess: (1) alignment with the stable gradient ascent direction for improving Exploitation capacity, and (2) pointing to loss regions with flatter surface for improving Exploration capacity.", + "bbox": [ + 89, + 521, + 483, + 597 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Gradient-Guided Sampling", + "text_level": 1, + "bbox": [ + 89, + 613, + 338, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(1) Stable gradient ascent direction: A sampled gradient in a sufficiently small neighborhood of the current example is probably a gradient ascent direction, while how can we ensure that this gradient ascent direction is stable? The Nesterov Accelerated Gradient (NAG) technique can precisely address this issue. It first allows the example to \"lookahead\" along the current momentum direction, and then acquires the gradient based on this new position, thereby identifying a more stable ascent direction [19, 23]. To leverage the superior stability of NAG based on random sampling, we perform the magnitude sampling along the momentum direction with an uniform distribution, and named this approach as Momentum-Guided Sampling (MGS). Compared to RS in Eq. 3, we only need to replace $\\tilde{x}_i$ with $\\bar{x}_i$ as follows:", + "bbox": [ + 89, + 638, + 483, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(m _ {i - 1}\\right)} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 864, + 482, + 891 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "lookahead sampling point", + "bbox": [ + 243, + 892, + 364, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $|\\tilde{p}|$ denotes the element-wise absolute value of $\\tilde{p}$ , $\\mathrm{sign}(\\cdot)$ denotes the sign function, which outputs either 1 or -1, and momentum decay is set to 1 for simplicity, so that $m_{i} = \\sum_{k=1}^{i} \\tilde{g}_{k}$ .", + "bbox": [ + 511, + 90, + 905, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(2) Pointing to regions with better flatness: As shown in Fig. 3(b), if we use momentum for 'lookahead', the early unstable sampling in the inner-iterations will impose excessive constraints on subsequent sampling directions, leading to long-chain dependencies that severely impair RS's ability in exploring flat regions. Thus, for our GGS, we substitute the momentum term in Eq. 4 with the gradient obtained from the previous iteration, where the lookahead mechanism is maintained with the gradient ascent direction. Specifically, we only need to replace $\\tilde{x}_i$ in Eq. 3 with $\\hat{x}_i$ as follows:", + "bbox": [ + 511, + 154, + 906, + 319 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i - 1}\\right)} _ {\\text {l o o k a h e a d s a m p l i n g p o i n t}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 327, + 905, + 364 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Fig. 3 further sheds light on the differences among three inner-iteration sampling guidance methods. As shown in Fig. 3(c), GGS largely alleviates long-chain dependencies compared with Momentum-Guided Sampling in Fig. 3(b), while ensuring the sampling direction to be align with the gradient ascent direction, compared with random sampling in Fig. 3(a). In addition, as shown in Fig. 2(c), GGS will achieve stable sampling directions (② ③ ④) after a brief period of oscillation (①), when the initial sampling points are unstable.", + "bbox": [ + 511, + 382, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For clarity, our attack process is shown in Algorithm 1.", + "bbox": [ + 531, + 536, + 895, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1 The attacking process of our GGS.", + "text_level": 1, + "bbox": [ + 514, + 571, + 833, + 587 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input: A clean image $x^{C\\times H\\times W}$ with ground-truth label $y$ , and the loss function $\\mathcal{L}(\\cdot)$ ; the magnitude of perturbation $\\epsilon$ ; the sampling radius $\\zeta$ ; the number of outer iterations, $T$ ; the decay factor $\\gamma$ ; the number of inner iterations, $N$ ; $\\tilde{g}_0,\\tilde{p}\\in \\mathbb{R}^{C\\times H\\times W}$", + "bbox": [ + 514, + 592, + 905, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Output: An adversarial example $x^{adv}$", + "bbox": [ + 514, + 669, + 777, + 684 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: Randomly sample $\\tilde{g}_0\\sim \\mathrm{Uniform}(-\\zeta ,\\zeta)$", + "2: $v_{0} = 0$ , $x_{0}^{adv} = x$ , $\\alpha = \\epsilon / T$ ;", + "3: while $t \\gets 1$ to $T$ (outer-iteration) do" + ], + "bbox": [ + 524, + 685, + 813, + 729 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "4: while $i\\gets 1$ to $N$ (inner-iteration) do", + "5: Randomly sample noises $\\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)$ ;", + "6: Get the lookahead example $\\hat{x}_i$ by Eq. (5);", + "7: Get the gradient $\\tilde{g}_i = \\nabla_x\\mathcal{L}(\\hat{x}_i,y)$", + "8: end while", + "9: Update $v_{t} = \\gamma \\cdot v_{t - 1} + \\frac{\\sum_{i = 1}^{N}\\tilde{g}_{i}}{\\left\\|\\sum_{i = 1}^{N}\\tilde{g}_{i}\\right\\|_{1}}$", + "0: Update $x_{t}^{adv} = \\Pi_{\\mathcal{B}_{\\epsilon}(x)}\\left[x_{t - 1}^{adv} + \\alpha \\cdot \\mathrm{sign}(v_{t})\\right]$ ;", + "1: end while", + "12: $x^{adv} = x_T^{adv}$" + ], + "bbox": [ + 519, + 685, + 900, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3889", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1d551a3653a2aeb83378c449f7c51a96492a870e8f73be17d2b511c090532a0d.jpg", + "table_caption": [ + "Table 1. The average untargeted and targeted ASR (%) on the nine models (Res50 [10], Dense121 [14], Inc-v3 [26], IncRes-v2 [27], ViT-B [3], PiT-B [11], and Inc-v3ens3, Inc-v3ens4, IncRes-v2ens [29]), with adversarial examples generated on a single model (Res50, Inc-v3 or ViT-B). Each data pair $(u/w)$ corresponds to the performances under (untargeted/targeted) attacks. The best and second best results are labeled in bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
AttackRes50Dense121Inc-v3IncRes-v2ViT-BPiT-BInc-v3ens3Inc-v3ens4IncRes-v2ensAvg.
Res50MI CVPR'1899.8 / 98.154.9 / 0.244.2 / 0.028.0 / 0.011.8 / 0.022.4 / 0.022.8 / 0.025.3 / 0.019.6 / 0.036.53 / 10.92
N ICLR'20100.0 / 95.163.1 / 0.247.2 / 0.032.3 / 0.012.9 / 0.023.7 / 0.023.0 / 0.027.5 / 0.019.0 / 0.038.74 / 10.59
VMI CVPR'2199.8 / 96.671.2 / 2.858.1 / 0.549.4 / 0.231.3 / 0.147.7 / 0.543.9 / 0.045.4 / 0.137.7 / 0.153.83 / 11.21
RAP NeurIPS'2299.6 / 26.486.8 / 0.568.4 / 0.052.6 / 0.022.2 / 0.040.5 / 0.136.2 / 0.036.7 / 0.025.5 / 0.052.06 / 3.00
GRA ICCV'2396.9 / 67.788.6 / 3.981.8 / 0.975.8 / 1.745.3 / 0.262.6 / 0.671.5 / 0.770.9 / 1.167.3 / 0.773.41 / 8.61
PGN NeurIPS'2398.6 / 49.591.3 / 4.785.0 / 1.478.5 / 1.749.7 / 0.467.8 / 1.074.9 / 0.672.9 / 1.170.1 / 1.276.53 / 6.84
ANDA CVPR'2499.9 / 97.687.5 / 5.473.6 / 0.466.3 / 0.842.3 / 0.161.8 / 1.151.0 / 0.152.9 / 0.146.7 / 0.164.67 / 11.74
GI ESWA'24100.0 / 98.568.7 / 0.455.6 / 0.137.6 / 0.115.3 / 0.029.0 / 0.028.8 / 0.030.4 / 0.122.4 / 0.043.09 / 11.02
GGS99.3 / 90.795.9 / 28.689.6 / 7.286.4 / 8.660.2 / 3.481.0 / 8.577.5 / 3.777.4 / 4.371.4 / 4.082.08 / 17.67
Inc-v3MI32.1 / 0.049.6 / 0.1100.0 / 95.356.1 / 0.113.0 / 0.017.3 / 0.034.3 / 0.038.0 / 0.026.3 / 0.040.74 / 10.61
NI41.4 / 0.060.5 / 0.1100.0 / 75.966.9 / 0.113.4 / 0.019.2 / 0.036.8 / 0.040.6 / 0.129.5 / 0.045.37 / 8.47
VMI44.9 / 0.165.6 / 0.199.9 / 76.672.7 / 0.418.3 / 0.025.3 / 0.048.9 / 0.152.5 / 0.040.9 / 0.052.11 / 8.59
RAP58.0 / 0.078.4 / 0.2100.0 / 16.482.1 / 0.115.4 / 0.023.8 / 0.044.1 / 0.147.0 / 0.033.9 / 0.053.63 / 1.87
GRA53.3 / 0.175.4 / 0.499.8 / 67.585.4 / 1.821.4 / 0.030.5 / 0.164.9 / 0.165.9 / 0.356.1 / 0.161.41 / 7.82
PGN56.0 / 0.079.4 / 0.2100.0 / 50.387.9 / 0.923.2 / 0.033.4 / 0.065.9 / 0.067.0 / 0.155.6 / 0.063.16 / 5.72
ANDA51.3 / 0.376.3 / 0.4100.0 / 87.482.5 / 0.620.9 / 0.031.8 / 0.158.2 / 0.261.2 / 0.249.3 / 0.059.06 / 9.91
GI39.9 / 0.059.3 / 0.3100.0 / 97.162.6 / 0.514.6 / 0.019.9 / 0.039.2 / 0.040.9 / 0.031.9 / 0.045.37 / 10.88
GGS69.1 / 1.186.9 / 1.9100.0 / 76.095.8 / 6.527.9 / 0.039.7 / 0.170.0 / 0.773.9 / 1.460.1 / 0.569.27 / 9.80
ViT-BMI41.7 / 0.055.6 / 0.058.1 / 0.039.4 / 0.0100.0 / 99.950.1 / 0.034.2 / 0.036.3 / 0.029.9 / 0.049.48 / 11.10
NI42.4 / 0.058.3 / 0.059.5 / 0.044.7 / 0.1100.0 / 99.050.0 / 0.236.2 / 0.037.6 / 0.031.0 / 0.051.08 / 11.03
VMI54.8 / 0.665.9 / 0.463.1 / 0.150.9 / 0.3100.0 / 99.067.6 / 1.545.3 / 0.144.7 / 0.038.7 / 0.259.00 / 11.36
RAP62.1 / 0.179.6 / 0.273.7 / 0.062.3 / 0.099.9 / 57.066.2 / 0.148.5 / 0.051.2 / 0.042.6 / 0.065.12 / 6.38
GRA64.9 / 1.077.6 / 1.174.7 / 0.868.8 / 0.899.3 / 91.780.9 / 4.761.5 / 0.264.0 / 0.157.7 / 0.272.16 / 11.18
PGN69.3 / 0.481.1 / 0.778.6 / 0.470.2 / 0.599.1 / 82.284.8 / 4.465.6 / 0.167.5 / 0.361.4 / 0.275.29 / 9.91
ANDA66.2 / 0.379.0 / 0.575.5 / 0.165.7 / 0.3100.0 / 99.979.0 / 1.455.3 / 0.156.7 / 0.149.9 / 0.269.70 / 11.44
GI55.8 / 0.169.6 / 0.267.1 / 0.052.5 / 0.1100.0 / 100.061.0 / 0.345.4 / 0.046.3 / 0.038.0 / 0.059.52 / 11.19
GGS80.8 / 6.089.9 / 6.587.4 / 4.682.3 / 4.6100.0 / 99.292.7 / 27.374.0 / 2.174.5 / 2.368.4 / 2.483.33 / 17.22
", + "bbox": [ + 91, + 157, + 910, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 549, + 223, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 574, + 297, + 590 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: Our experiments are conducted on an ImageNet-compatible dataset that is widely used in adversarial attack research. The dataset includes 1,000 images, each with a size of $299 \\times 299 \\times 3$ .", + "bbox": [ + 89, + 597, + 482, + 656 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Models: We select six normally pre-trained models, i.e., ResNet50 (Res50) [10], DenseNet121 (Dense121) [14], Inception-v3 (Inc-v3) [26], InceptionResNet-v2 (IncRes-v2) [27], VisionTransformer-Base (ViT-B) [3], and Pooling-based VisionTransformer-Base (PiT-B) [11], along with three models pre-trained on adversarial examples Inception-v3_ens3 (Inc-v3_ens3), Inception-v3_ens4 (Inc-v3_ens4) and InceptionResNet-v2_ens (IncRes-v2_ens) [29], to evaluate the performance of different methods.", + "bbox": [ + 89, + 657, + 482, + 794 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines: To provide a comprehensive evaluation of our method, we choose eight popular gradient-based adversarial attack methods as baselines: MI-FGSM [1], NIFGSM [19], VMI-FGSM [32], RAP [24], GRA [40], PGN [7], ANDA [5] and GI [30]. Additionally, we integrate our method with five input transformation-based methods, i.e., DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "to verify the generalizability of our approach.", + "bbox": [ + 511, + 550, + 813, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hyperparameters: For our GGS, we set the maximum perturbation $\\epsilon = 16 / 255$ , the number of outer-iterations $T = 10$ , the step size $\\alpha = \\epsilon / T$ , the number of inner-iterations $N = 20$ , and the upper bound factor of sample range $\\zeta = 2.0 \\times \\epsilon$ . More hyperparameter explanation for other methods is provided in the Supplementary Material.", + "bbox": [ + 511, + 566, + 905, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Attack on Different Models", + "text_level": 1, + "bbox": [ + 511, + 674, + 759, + 689 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Classification model", + "text_level": 1, + "bbox": [ + 511, + 698, + 700, + 712 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Table 1, our method consistently outperforms state-of-the-art (SOTA) approaches. Compared to the best-performing competitor, it improves the average attack success rate (ASR) by over $5\\%$ in both untargeted and targeted attacks across most architectures, except for Inception-v3. Furthermore, our GGS significantly enhances attack performance on the surrogate model, achieving $4\\% -25\\%$ higher ASR in targeted attacks than other sampling methods like PGN [7] and GRA [40]. This demonstrates GGS's ability to boost attack potency while maintaining strong cross-model generalization. Additional details on ensemble setting can be found in the supplementary materials.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3890", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/be989cca6b14e70b2dce74843ac61ef3c16bb5de095c8a60ed785a800b77d7a2.jpg", + "image_caption": [ + "Figure 4. The loss surfaces of adversarial examples generated by different methods (MI-FGSM [1], GRA [40], PGN [7], ANDA [5]), based on three different model architectures (ResNet50 [10], Inception-v3 [26], and ViT-B [3]) along random directions with varying strengths like [7]. In the images marked with a star in the upper left corner, adversarial examples are generated and tested on the same model, indicating white-box testing. In contrast, unmarked images represent adversarial examples generated on one model and tested on another, indicating black-box testing. We have also highlighted the regions covered by our loss surface using a red background for easier visualization." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 911, + 349 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.2. Multimodal Large Language Models", + "text_level": 1, + "bbox": [ + 89, + 467, + 393, + 483 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The widespread adoption of Multimodal Large Language Models (MLLMs) has raised significant security concerns. We evaluated our method on five prominent MLLMs using adversarial examples generated under an ensemble setting, as shown in Table 2. Our GGS method achieved superior attack performance, reducing the average CSR by over $9\\%$ compared to the strongest baseline, which demonstrates its excellent cross-model applicability.", + "bbox": [ + 89, + 487, + 483, + 608 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6150c92b71d4ee276a0d3f7a351efd891f22fb6acdf02e0688b6cb87f66b6823.jpg", + "table_caption": [ + "Table 2. Classification success rate (CSR, %) of adversarial examples generated by different methods in Multimodal Large Language Models (MLLMs) from three major providers (OpenAI, Google, Anthropic) under ensemble settings (Res50 [10], Inc-v3 [26], ViT-B [3]). Row \"clean\" indicates the CSR on the clean samples. Avg. represents the average CSR across the five MLLMs." + ], + "table_footnote": [], + "table_body": "
MethodGPTGeminiClaude
4ominiproflashsonnetAvg. ↓
Clean77.184.085.680.968.379.18
MI_CVPR'18 [1]77.077.381.876.561.074.72
NI_ICLR'20 [19]75.073.983.176.862.474.24
VMI_CVPR'21 [32]68.866.678.473.263.170.02
RAP_NeurIPS'22 [24]64.362.075.166.160.265.54
GRA_ICCV'23 [40]57.152.270.965.646.758.50
PGN_NeurIPS'23 [7]56.249.069.664.145.156.80
ANDA_CVPR'24 [5]59.956.874.969.350.562.28
GI_ESWA'24 [30]71.572.879.473.263.872.14
GGS43.137.861.155.740.047.54
", + "bbox": [ + 94, + 709, + 480, + 897 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.3. Loss Surface Visualization against Noise Levels", + "text_level": 1, + "bbox": [ + 511, + 467, + 888, + 482 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To shed light on the capacity of our GGS in locating flat maxima regions for adversarial examples, we compared the loss surfaces of adversarial examples generated by different attack methods in Fig. 4. Each curve represents the average loss of 32 randomly selected adversarial examples, while the center of each image (noise level $= 0$ ) indicates the average loss of clean adversarial examples. In white-box settings (marked with a star), GGS consistently finds flatter maxima compared to baselines. Consequently, in black-box testing, GGS achieves the highest loss values (images without star), indicating its superior transfer attack capability. Additionally, the GGS loss surface (highlighted in red) almost entirely encompasses those of other methods, confirming its effectiveness in balancing strong attack potency with cross-model generalization.", + "bbox": [ + 511, + 487, + 906, + 714 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Compatibility with Other Methods", + "text_level": 1, + "bbox": [ + 511, + 724, + 816, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attack with Inner-Iteration Sampling Methods: Our Gradient-Guided Sampling (GGS) is compatible with existing inner-iteration RS-based approaches. When integrated with SOTA approaches like GRA [40] and PGN [7], GGS largely improves attack success rates (ASR). As shown in Table 3, untargeted ASR improved by $5.2\\%$ (GRA) and $6.7\\%$ (PGN), while targeted ASR increased by $5\\%$ (GRA) and $6\\%$ (PGN), demonstrating GGS's compatibility.", + "bbox": [ + 511, + 748, + 906, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Attack with Input Transformations Methods: To study the compatibility of our GGS with existing input", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3891", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cb8d7a553ac8ca343bd28868d690ad6b0905d3d7be442b34ceaaa8dbc101b08.jpg", + "image_caption": [ + "(a) Random (RS)" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 294, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb30b3b6476906dc0618bc9aeb22f4f3ba923781b64172d9eec5cfcd879634dc.jpg", + "image_caption": [ + "(b) Momentum-Guided (MGS)", + "Figure 5. Loss surfaces of adversarial examples generated by different sampling strategies, Random Sampling (RS), MGS and our GGS, with increasing inner-iteration on Resnet50. The red line on the left side of part (a)~(c) represents the maximum value of the loss surface in each iteration, and the blue dashed line highlights these values in early-stage inner iteration. For (d), it represents the cosine similarity between the gradient $\\tilde{g}_i$ generated in each inner-iteration and the average of gradients $\\sum_{i=1}^{N} \\tilde{g}_i / N$ ." + ], + "image_footnote": [], + "bbox": [ + 295, + 88, + 488, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9232b6d667d31188f1c864ee7a648e4c116d986e9440d9b7af55415a25b0f791.jpg", + "image_caption": [ + "(c) Gradient-Guided (GGS)" + ], + "image_footnote": [], + "bbox": [ + 488, + 88, + 684, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4788acac99032229f83e43fa9ddedeb10c116d6066817e4027a343a1098a4465.jpg", + "image_caption": [ + "(d) The cosine similarity between the gradient and the gradients averaging in inner-iterations" + ], + "image_footnote": [], + "bbox": [ + 686, + 88, + 903, + 186 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d324da4f8d411323608fcac32ff8430a80c9ce9b6c6e67897a1a1180a45157c7.jpg", + "table_caption": [ + "Table 3. The average untargeted and targeted ASR $(\\%)$ on all nine models with the adversarial examples generated on ResNet50, when combined our GGS with input transformation-based methods (DIM [37], TIM [2], SIM [19], Admix [33], SSM [21]) and inner-iteration RS-based methods (GRA [40], PGN [7])." + ], + "table_footnote": [], + "table_body": "
ClassMethodUntargeted Avg.Targeted Avg.
Random Sampling (RS)GRA / +ours73.41 / 78.698.61 / 13.64
PGN / +ours76.53 / 83.236.84 / 12.79
Input TransformationDIM / +ours51.30 / 90.138.63 / 19.37
TIM / +ours40.82 / 84.9411.03 / 18.48
SIM / +ours46.80 / 90.1211.14 / 28.44
Admix / +ours54.78 / 85.5011.28 / 30.50
SSM / +ours74.22 / 80.748.40 / 12.54
", + "bbox": [ + 91, + 382, + 486, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "transformation techniques, we integrated it with five methods of this type: DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]. Adversarial examples generated on ResNet50 were tested on nine diverse models. Table 3 shows GGS significantly improved adversarial transferability, increasing untargeted ASR by $6\\% -43\\%$ and targeted ASR by $4\\% -19\\%$ . For instance, SIM's untargeted ASR improved by $43\\%$ when combined with GGS. More details are in the supplementary material.", + "bbox": [ + 89, + 539, + 483, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Study of Proposed Components", + "text_level": 1, + "bbox": [ + 89, + 683, + 439, + 700 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e6d30d42f25c2b0268fe6df2fe6c8509ed4690a9c700ba45ea85c092c0b5d1e5.jpg", + "table_caption": [ + "Table 4. The average untargeted ASR on ResNet50 and eight target models for Random Sampling (RS), Momentum-Guided Sampling (MGS), and our Gradient-Guided Sampling (GGS)." + ], + "table_footnote": [], + "table_body": "
Sampling Guided byRes50Others
randmomentumgradient
(a) RS✓--97.363.74
(b) MGS-✓-97.558.79
(c) GGS--✓99.379.93
", + "bbox": [ + 99, + 770, + 473, + 857 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study compares RS, MGS, and GGS via loss surface analysis (Fig. 5) and attack success rates (Table 4).", + "bbox": [ + 89, + 869, + 480, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Fig. 5(a), Random Sampling (RS) yields a relatively flat loss surface due to its exploration capability but it suffers from significant oscillation in loss values. Introducing Momentum-Guided Sampling (MGS) (Fig. 5(b)) stabilizes the loss surface over iterations, but its long-chain dependency causes high similarity between later gradients and the final average gradient (Fig. 5(d)), severely limiting exploration. Our GGS resolves MGS's long-chain dependency and mitigates early-stage noisy gradients (Fig. 5(b)(c), blue dashed lines). The low gradient similarity in Fig. 5(d) reflects the enhanced exploration, while the flat yet higher-loss surface in Fig. 5(c) demonstrates the exceptional sampling efficiency of our GGS.", + "bbox": [ + 511, + 303, + 906, + 500 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Quantitatively, as shown in Table 4, MGS marginally improves ASR of the surrogate model by $0.2\\%$ , but reducing transfer attack capability by $5\\%$ . In contrast, GGS increases both white-box and transfer ASR by $2\\%$ and $16\\%$ , respectively, validating its superior sampling efficiency.", + "bbox": [ + 511, + 500, + 905, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Perspectives", + "text_level": 1, + "bbox": [ + 511, + 588, + 774, + 606 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Gradient-Guided Sampling (GGS), an easy-to-implement inner-iteration sampling strategy, for transferable adversarial attacks that effectively resolves the dilemma between Exploitation (attack potency) and Exploration (cross-model generalization). By guiding the sampling direction with the gradient from the previous inner iteration, GGS enables stable ascent toward flat loss regions with higher local maxima. Our GGS has the merit of enhancing exploitation capability without compromising exploration capability, effectively balancing these two aspects and addressing the inherent limitations of random sampling (RS) methods, as verified by extensive experiments and visualizations as well as the comparison with related state-of-the-art methods. GGS is also revealed to be highly compatible with existing methods.", + "bbox": [ + 511, + 613, + 906, + 839 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our GGS is highly compatible with gradient-averaging methods (e.g., GRA [40], PGN [7]), we aim to refine it to also support non-gradient-averaging techniques (e.g., VMI-FGSM [32], RAP [24]).", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3892", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported by the National Natural Science Foundation of China under grants no. 62276170, 82261138629, 62306061, the Science and Technology Project of Guangdong Province under grants no. 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grant no. JCYJ20220531101412030, Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) under Grant No. GML-KF-24-11, and Guangdong Provincial Key Laboratory under grant no. 2023B1212060076.", + "bbox": [ + 89, + 114, + 485, + 282 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 295, + 187, + 311 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 1, 2, 3, 6, 7", + "[2] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4312-4321, 2019. 2, 3, 6, 8", + "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6, 7", + "[4] Yan Fang, Zhongyuan Wang, Jikang Cheng, Ruoxi Wang, and Chao Liang. Promoting adversarial transferability with enhanced loss flatness. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1217-1222. IEEE, 2023. 2", + "[5] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 3, 6, 7", + "[6] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, ICLR, 2021. 2, 3", + "[7] Zhijin Ge, Hongying Liu, Wang Xiaosen, Fanhua Shang, and Yuanyuan Liu. Boosting adversarial transferability by achieving flat local maxima. Advances in Neural Information Processing Systems, 36:70141-70161, 2023. 1, 2, 3, 4, 6, 7, 8", + "[8] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 3", + "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy." + ], + "bbox": [ + 99, + 320, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Explaining and harnessing adversarial examples. In International Conference on Learning Representations, ICLR, 2015. 3", + "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7", + "[11] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11936-11945, 2021. 6", + "[12] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997. 3", + "[13] SM Mostaq Hossain, Shampa Banik, Trapa Banik, and Ashfak Md Shibli. Survey on security attacks in connected and autonomous vehicular systems. In 2023 IEEE International Conference on Computing (ICOCO), pages 295-300. IEEE, 2023. 2", + "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6", + "[15] Martin Husák, Jana Komárková, Elias Bou-Harb, and Pavel Čeleda. Survey of attack projection, prediction, and forecasting in cyber security. IEEE Communications Surveys & Tutorials, 21(1):640-660, 2018. 2", + "[16] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. In Advances in Neural Information Processing Systems, pages 73919-73942. Curran Associates, Inc., 2023. 2", + "[17] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. ICLR, 2017. 3", + "[18] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 3", + "[19] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, ICLR, 2020. 3, 4, 5, 6, 7, 8", + "[20] Decheng Liu, Xijun Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Adv-diffusion: Imperceptible adversarial face identity attack via latent diffusion model. Proceedings of the AAAI Conference on Artificial Intelligence, 38(4):3585-3593, 2024. 2", + "[21] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European conference on computer vision, pages 549-566. Springer, 2022. 3, 6, 8", + "[22] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, ICLR, 2018. 3" + ], + "bbox": [ + 516, + 92, + 908, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3893", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Dokl. Akad. Nauk. SSSR, page 543, 1983. 2, 4, 5", + "[24] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in neural information processing systems, 35:29845-29858, 2022. 2, 3, 4, 6, 7, 8", + "[25] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. 4", + "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6, 7", + "[27] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-reset and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6", + "[28] Chee-Wooi Ten, Govindarasu Manimaran, and Chen-Ching Liu. Cybersecurity for critical infrastructures: Attack and defense modeling. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 40(4):853–865, 2010. 2", + "[29] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. *ICLR*, 2018. 6", + "[30] Jiafeng Wang, Zhaoyu Chen, Kaixun Jiang, Dingkang Yang, Lingyi Hong, Pinxue Guo, Haijing Guo, and Wenqiang Zhang. Boosting the transferability of adversarial attacks with global momentum initialization. Expert Systems with Applications, 255:124757, 2024. 3, 6, 7", + "[31] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3", + "[32] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1924-1933, 2021. 2, 3, 4, 6, 7, 8", + "[33] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 2, 3, 6, 8", + "[34] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 3", + "[35] Han Wu, Syed Yunas, Sareh Rowlands, Wenjie Ruan, and Johan Wahlström. Adversarial driving: Attacking end-to-end autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7. IEEE, 2023. 2", + "[36] Chaowei Xiao, Bo Li, Jun yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 3905-3911, 2018. 2", + "[37] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 2, 3, 6, 8", + "[38] Yinhu Xu, Qi Chu, Haojie Yuan, Zixiang Luo, Bin Liu, and Nenghai Yu. Enhancing adversarial transferability from the perspective of input loss landscape. In International Conference on Image and Graphics, pages 254-266. Springer, 2023. 2", + "[39] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 3", + "[40] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4741-4750, 2023. 2, 3, 4, 6, 7, 8", + "[41] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 3" + ], + "bbox": [ + 516, + 92, + 903, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3894", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_model.json b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a202e753f599ca4afc73d0e1fa49bc2dff9ef74c --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_model.json @@ -0,0 +1,2277 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.85, + 0.177 + ], + "angle": 0, + "content": "Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.203, + 0.895, + 0.222 + ], + "angle": 0, + "content": "Zenghao Niu\\(^{1}\\) Weicheng Xie\\(^{1,2,3}\\)* Siyang Song\\(^{4}\\) Zitong Yu\\(^{5}\\) Feng Liu\\(^{1}\\) Linlin Shen\\(^{6}\\)" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.222, + 0.852, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Computer Science & Software Engineering, Shenzhen University, China" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.239, + 0.892, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.103, + 0.257, + 0.94, + 0.275 + ], + "angle": 0, + "content": "3Guangdong Provincial Key Laboratory of Intelligent Information Processing, Shenzhen University, China" + }, + { + "type": "text", + "bbox": [ + 0.297, + 0.275, + 0.746, + 0.292 + ], + "angle": 0, + "content": "\\(^{4}\\)School of Computer Science, University of Exeter, U.K." + }, + { + "type": "text", + "bbox": [ + 0.18, + 0.292, + 0.859, + 0.31 + ], + "angle": 0, + "content": "\\(^{5}\\)Department of Computing and Information Technology, Great Bay University, China" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.31, + 0.871, + 0.327 + ], + "angle": 0, + "content": "6Computer Vision Institute, School of Artificial Intelligence, Shenzhen University, China" + }, + { + "type": "text", + "bbox": [ + 0.27, + 0.33, + 0.764, + 0.345 + ], + "angle": 0, + "content": "{2300271072@mail., wcxie@, feng.liu@, llshen@}szu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.349, + 0.687, + 0.362 + ], + "angle": 0, + "content": "s song@exeter.ac.uk, zitong.yu@ieee.org" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.398, + 0.327, + 0.413 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.421, + 0.485, + 0.86 + ], + "angle": 0, + "content": "Adversarial attacks present a critical challenge to deep neural networks' robustness, particularly in transfer scenarios across different model architectures. However, the transferability of adversarial attacks faces a fundamental dilemma between Exploitation (maximizing attack potency) and Exploration (enhancing cross-model generalization). Traditional momentum-based methods over-prioritize Exploitation, i.e., higher loss maxima for attack potency but weakened generalization (narrow loss surface). Conversely, recent methods with inner-iteration sampling over-prioritize Exploration, i.e., flatter loss surfaces for cross-model generalization but weakened attack potency (suboptimal local maxima). To resolve this dilemma, we propose a simple yet effective Gradient-Guided Sampling (GGS), which harmonizes both objectives through guiding sampling along the gradient ascent direction to improve both sampling efficiency and stability. Specifically, based on MI-FGSM, GGS introduces inner-iteration random sampling and guides the sampling direction using the gradient from the previous inner-iteration (the sampling's magnitude is determined by a random distribution). This mechanism encourages adversarial examples to reside in balanced regions with both flatness for cross-model generalization and higher local maxima for strong attack potency. Comprehensive experiments across multiple DNN architectures and multimodal large language models (MLLMs) demonstrate the superiority of our method over state-of-the-art transfer attacks. Code is made available at https://github.com/anuin-cat/GGS." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.396, + 0.911, + 0.649 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.657, + 0.908, + 0.812 + ], + "angle": 0, + "content": "Figure 1. The loss surfaces of (a) MI-FGSM [1] (Momentum iterative fast gradient sign method), (b) RS (Base inner-iteration Random Sampling defined in section 3.2) for enhancing exploration within the neighborhood, (c) PGN [7] (Penalizing Gradient Norm) with Random Sampling (RS) for stable gradient estimation to enhance exploration, and (d) Our GGS (Gradient-Guided Sampling) for efficient sampling to generate gradients that stably towards flat regions with higher local maxima. It shows that, building upon RS, our approach not only maintains a flat loss surface, but also improves the local maximum loss value, compared to PGN, achieving a balance between exploration and exploitation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.842, + 0.646, + 0.858 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Researches have indicated that adding imperceptible perturbations to input data can easily affect deep neural networks" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.237, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3885" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "(DNNs), leading to erroneous predictions in safety-critical domains such as autonomous driving [13, 35], and cybersecurity [15, 28]. To facilitate the development of relevant defense mechanisms, in-depth research on adversarial attack methods is urgently needed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.381 + ], + "angle": 0, + "content": "While adversarial attacks are categorized into white-box and black-box settings, the black-box scenario is more realistic in practical applications, where the attacker lacks knowledge of the target model and relies solely on a surrogate model, aiming to generate adversarial examples that are effective across multiple target models. To enhance the transferability of adversarial examples in this setting, researchers have explored gradient-based methods [1, 7, 32], input transformation methods [2, 33, 37], and generative methods [16, 20, 36] to mitigate overfitting of adversarial examples on surrogate models. Despite significant progress, the effectiveness of attacks on target models remains substantially lower than that on surrogate models, driving the need for more effective attack strategies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.579 + ], + "angle": 0, + "content": "In black-box adversarial attacks, RAP [24] adopted Sharpness-Aware Minimization (SAM [6]) to improve transferability by flattening the loss surface. Building on this, PGN [7] further enhanced the transferability of adversarial examples by penalizing gradient norm and flattening the loss surface. The efficacy of flat loss surface in improving adversarial attack transferability is increasingly supported [4, 7, 38]. To enhance exploration through local flatness, recent approaches, building on MI-FGSM's [1] momentum-based outer-iterations, often incorporates inner-iterations before each example update [7, 40], i.e., performing multi-step neighborhood sampling for stabilization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.809 + ], + "angle": 0, + "content": "However, these methods enhance Exploration (cross-model generalization) at the cost of compromising Exploitation (attack potency). This is because, although inner-iterations can significantly enhance exploration capability through neighborhood sampling, independent random sampling struggles to obtain stable and effective gradient directions, while potential noise may cause the averaged gradient to deviate from the optimal ascent toward target regions, which exhibit flatter loss surface and higher local maxima. To optimize sampling efficiency, a guided sampling constraint is essential, which should satisfy two criteria: (1) ensuring stable gradient ascent direction and (2) directing gradients toward flatter regions, thereby enabling samples to reach flatter local maxima regions with enhanced transferability, balancing exploitation and exploration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "For (1) stable gradient direction, a sufficiently small neighborhood sampling followed by gradient calculation generally yields correct ascent directions. Yet, how can we further stabilize this direction? The Nesterov Accelerated Gradient (NAG) [23] enhances stability by firstly updating samples along the current momentum (lookahead)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.319 + ], + "angle": 0, + "content": "before acquiring gradients, providing more stable directions. Therefore, momentum-based direction constraints are promising for sampling, as the additional random magnitude can be used to preserve the sampling randomness. (2) However, directly applying NAG's momentum mechanism during inner-iteration sampling imposes strong constraints on subsequent directions, restricts the exploration capability for flat regions. To address this, we replace momentum with the gradient from the previous inner-iteration, named Gradient-Guided Sampling (GGS). This adjustment ensures that sampling to align with gradient ascent directions while reducing the impact from early-stage instability. Our GGS can achieve stable sampling directions after brief oscillations, consistently targeting flat loss regions with larger local maxima regardless of the initial sampling quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.32, + 0.905, + 0.486 + ], + "angle": 0, + "content": "As shown in Fig. 1(d), our GGS generates adversarial examples with flatter loss surface and higher local maxima compared to RS, i.e., it does not sacrifice local maxima in the pursuit of a flatter surface compared to PGN, thereby achieving a balance between exploration and exploitation to facilitate attack transferability. Additionally, since our approach represents a complementary improvement over Random Sampling (RS), it is compatible with RS-based methods, such as inner-iteration random sampling and gradient averaging, including PGN [7] and GRA [40]. In summary, this study makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.492, + 0.904, + 0.597 + ], + "angle": 0, + "content": "- We propose Gradient-Guided Sampling (GGS), a novel inner-iteration sampling strategy that effectively balances exploration (cross-model generalization) and exploitation (attack potency). By leveraging the gradient from the previous inner-iteration to guide sampling directions, GGS achieves stable gradient ascent toward flat loss regions with higher local maxima." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.598, + 0.905, + 0.658 + ], + "angle": 0, + "content": "- The GGS framework is compatible with existing inner-iteration random sampling-based methods, enhancing their sampling efficiency and further improving the transferability of the adversarial examples they generate." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.659, + 0.905, + 0.747 + ], + "angle": 0, + "content": "- Extensive experimental results as well as comprehensive visualizations demonstrate the efficacy of our GGS, covering targeted and non-targeted attacks in cross-architecture black-box scenarios, non-targeted attacks on multimodal large language models (MLLMs) and commercial cloud functions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.492, + 0.905, + 0.747 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.768, + 0.655, + 0.784 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Transferability is a critical property in adversarial attacks, allowing adversarial examples generated from a surrogate model to deceive unseen target models in black-box settings. This characteristic makes transfer-based attacks highly practical in real-world applications where target models are inaccessible, motivating our focus on transferable adversarial attacks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3886" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.283, + 0.105 + ], + "angle": 0, + "content": "2.1. Transferable Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.339 + ], + "angle": 0, + "content": "Gradient-based Attack. Gradient-based attacks aim to generate adversarial examples by exploiting the gradients of a surrogate model to maximize the loss. Classic methods such as Fast Gradient Sign Method (FGSM) [9] and Projected Gradient Descent (PGD) [22] optimize adversarial perturbations to fool the surrogate model. However, these attacks often overfit the surrogate model, resulting in limited transferability to target models. To address this, advanced techniques such as momentum-based (MI-FGSM [1]), Nesterov Iterative-based (NI-FGSM [19]), variance-tuned (VMI-FGSM [32]), gradient relevance-based (GRA [40]), momentum initialization (GI-FGSM [30]) and distribution-based methods (ANDA [5]) have been introduced to smooth out the optimization process and improve transferability across models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.484, + 0.506 + ], + "angle": 0, + "content": "Input Transformation. To mitigate the overfitting issue inherent in gradient-based methods, input transformation techniques have been proposed to enhance the diversity of adversarial examples. Techniques like translation, resizing, padding, image mixup (e.g., DIM [37], SIM [19], TIM [2], Admix [33]) and other advanced image transformation methods (e.g., SSM [21], SIA [34], STM [8], L2T [41], BSR [31]) introduce randomness to the input data before generating perturbations. By applying such transformations during the attack process, more effective transferable adversarial examples can be generated." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.516, + 0.227, + 0.531 + ], + "angle": 0, + "content": "2.2. Flat Maxima" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.75 + ], + "angle": 0, + "content": "The generalization ability of models has been suggested to possibly have a certain association with flat minima [12]. More in-depth research and exploration have gradually confirmed this perspective [6, 17, 18, 39]. Moreover, flat maxima also have been validated to be effective for enhancing the generalization and transferability of adversarial examples [24]. Flat maxima refers to regions in the loss landscape, where small changes in the model parameters result in minimal changes in the loss, making adversarial examples generated in these regions less sensitive to the specific decision boundaries of models. This property contrasts with \"sharp maxima\", which are highly sensitive to small perturbations and can lead to overfitting to the surrogate model, reducing transferability capacity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In the context of transfer-based attacks, works like Reverse Adversarial Perturbation (RAP) [24] leverage flat maxima to generate adversarial examples, PGN [7] adopts a first-order procedure to approximate the Hessian/vector product, largely improving computational efficiency. They are not only effective on the surrogate model but also more resilient when attacking unseen target models. These methods mitigate overfitting, via making adversarial examples lie in regions where the loss function remains stable across different models (e.g., flat maxima regions), improving at" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "tack success rates in black-box settings. However, achieving flat maxima requires not only attention to \"flatness\" but also preservation of the \"maxima\" to ensure sufficient attack strength during transfer attack, which poses a significant challenge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.178, + 0.745, + 0.194 + ], + "angle": 0, + "content": "2.3. Inner-Iteration Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.2, + 0.907, + 0.366 + ], + "angle": 0, + "content": "To enhance the generalization capability of adversarial examples, recent methods have introduced inner-iteration sampling [7, 24, 32]. Specifically, an inner iterative process is inserted before updating the examples. This process involves neighborhood sampling to acquire domain information during inner-iterations which are used to adjust gradients [32] or neighborhood search is incorporated to locate local minima and enhance their loss values [24]. Furthermore, works like [7, 40] enhance the flatness of the loss landscape by averaging gradients from inner-iterations to update the examples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.503 + ], + "angle": 0, + "content": "Additionally, neighborhood sampling can facilitate examples to escape from sharp local maxima regions, thereby avoiding overfitting to the surrogate model. However, directly averaging the gradients from inner-iterations will reduce the maximum value of the loss surface, since random sampling produces unstable gradient directions that fail to consistently direct to flat regions with higher local maxima. Thus, enhancing inner-iteration sampling efficiency to stabilize the final gradient has become a critical challenge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.649, + 0.534 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.655, + 0.557 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.565, + 0.907, + 0.777 + ], + "angle": 0, + "content": "Transferable Adversarial Attacks. For a given example \\((x,y)\\), a surrogate model \\(f_{\\theta}\\) and multiple target models \\(f_{\\varphi_k}\\) for \\(k \\in \\{1,2,\\dots,K\\}\\), the attacker's goal is to find \\(x^{adv} = x + \\delta\\) using \\(f_{\\theta}\\) to make \\(f_{\\varphi_k}(x^{adv}) \\neq y\\), where \\(\\| \\delta \\|_p < \\epsilon\\), \\(\\epsilon\\) denotes the maximum magnitude of the specified perturbation and \\(\\| \\cdot \\|_p\\) denotes the \\(\\ell_p\\)-norm. For targeted attacks, we simply adapt the symbols and labels in the target function: \\(f_{\\varphi_k}(x^{adv}) = y_t\\), where \\(y_t\\) denotes the target label. Meanwhile, to facilitate a clearer representation of the norm constraints against \\(x^{adv}\\) as below, we define \\(\\mathcal{B}_{\\epsilon}(x) = \\{x': \\| x' - x \\|_p \\leq \\epsilon\\}\\) to denote the \\(\\epsilon\\)-ball of an input image \\(x\\). For the purpose of the above attack, we generally need to maximize the following objective function to generate the transferable adversarial examples:" + }, + { + "type": "equation", + "bbox": [ + 0.63, + 0.787, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\max _ {x ^ {a d v} \\in \\mathcal {B} _ {\\epsilon} (x)} \\mathcal {L} \\left(x ^ {a d v}, y\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}(\\cdot)\\) represents the loss function. In the above flow, we assume that \\(f_{\\theta}\\) is known as a white-box model and \\(f_{\\varphi_k}\\) is unknown as a black-box model. Similar to the process of training a neural network, the white-box model can be seen as the training data, with the adversarial examples serving" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3887" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.12 + ], + "angle": 0, + "content": "as the optimization parameters, while the black-box model acts as the test data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.273 + ], + "angle": 0, + "content": "Nesterov Accelerated Gradient (NAG). Due to the similarity between model training task and adversarial example generation, improving performance in black-box settings can be viewed as enhancing model generalization in model training tasks. Consequently, some commonly used optimization methods, such as Momentum [25] and Nesterov Accelerated Gradient (NAG) [23], can be introduced to improve the generalization of adversarial examples across black-box models through a more stable optimization process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.484, + 0.351 + ], + "angle": 0, + "content": "NI-FGSM [19] was the first to incorporate the principles of NAG into adversarial attack tasks. This enables the algorithm to acquire directional information in advance, reducing oscillations and achieving stable convergence. Specifically, momentum is computed at each step as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.361, + 0.483, + 0.405 + ], + "angle": 0, + "content": "\\[\nv _ {t} = \\gamma \\cdot v _ {t - 1} + \\frac {\\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y)}{\\left\\| \\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y) \\right\\| _ {1}}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.4, + 0.346, + 0.437 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} _ {t} = \\underbrace {x _ {t - 1} ^ {a d v} + \\alpha \\cdot \\gamma \\cdot v _ {t - 1}} _ {\\text {l o o k a h e a d p o i n t}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.557 + ], + "angle": 0, + "content": "where \\( v_{t} \\) represents the momentum term, \\( x_0^{adv} = x \\), \\( \\alpha \\) stands for the step size, \\( \\gamma \\) denotes the momentum decay factor, and \\( \\| \\cdot \\|_1 \\) denotes the \\( \\ell_1 \\)-norm of the given variable. Unlike traditional momentum-based methods, NI-FGSM utilizes the estimated position of examples rather than their current actual position, during gradient computation, i.e., introducing a lookahead mechanism into the algorithm." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.212, + 0.582 + ], + "angle": 0, + "content": "3.2. Motivation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.727 + ], + "angle": 0, + "content": "Recent transfer-based adversarial attack methods have introduced inner-iteration random sampling, to enhance algorithm stability [7, 40] or acquire neighborhood information [24, 32]. It can be noted that these methods have increased the flatness of the loss surface, as illustrated in Fig.1(c), with additional loss surface plots for other methods provided in the supplementary materials. If we isolate the inner-iteration random sampling (RS) component, it can be expressed in the following form:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.739, + 0.483, + 0.789 + ], + "angle": 0, + "content": "\\[\nv _ {t} = \\gamma v _ {t - 1} + \\frac {\\sum_ {i = 1} ^ {N} \\tilde {g} _ {i}}{\\left\\| \\sum_ {i = 1} ^ {N} \\tilde {g} _ {i} \\right\\| _ {1}}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.787, + 0.426, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\tilde {g} _ {i} = \\nabla_ {x} \\mathcal {L} \\big (\\tilde {x} _ {i}, y \\big), \\text {w i t h} \\tilde {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + \\tilde {p}} _ {\\text {s a m p l i n g p o i n t}},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.838, + 0.484, + 0.9 + ], + "angle": 0, + "content": "where \\(\\sum_{i=1}^{N}\\) denotes the output of the inner-iteration process; \\(N\\) denotes the number of inner-iterations; \\(\\gamma\\) denotes the momentum decay factor; \\(\\tilde{g}_0, \\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)\\), are uniform random noises with the same size as \\(x\\)." + }, + { + "type": "image", + "bbox": [ + 0.529, + 0.09, + 0.895, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.237, + 0.858, + 0.251 + ], + "angle": 0, + "content": "(a) Optimization Trajectories of Various Methods" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.253, + 0.703, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.541, + 0.388, + 0.688, + 0.402 + ], + "angle": 0, + "content": "(b) 1st step process of RS" + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.253, + 0.891, + 0.386 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.724, + 0.389, + 0.88, + 0.402 + ], + "angle": 0, + "content": "(c) 1st step process of GGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.649 + ], + "angle": 0, + "content": "Figure 2. (a) Outer-Iteration processes of MI-FGSM (green), RS (purple) and our GGS (red). (i) MI-FGSM prioritizes rapid ascent directions, thereby enhancing exploitation and facilitating access to sharp local maxima regions. (ii) RS introduces inner-iterations and enhances exploration capability through neighborhood search, thereby enabling access to flat local maxima regions. (iii) GGS incorporates gradient constraints into RS, simultaneously enhancing both exploration and exploitation capabilities, thereby enabling the faster convergence to the centers of flat local maxima regions. (b) RS is performed within a neighborhood of the current example, and uses the average gradient of all sampled points as the final gradient. (c) Building upon RS, GGS enables examples to have stable gradient directions (② ③ ④) toward the centers of flat local maxima regions, following an initial period of brief oscillation (①). It uses the gradient direction from the previous inner-iteration as the guidance, while the randomness is maintained by setting the gradient magnitude with a random distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.907, + 0.867 + ], + "angle": 0, + "content": "As shown in Fig. 1, the inner-iteration RS possesses a strong capability to enhance the flatness of the loss surface compared to MI-FGSM. Compared to RS, current methods such as PGN [7] can further improve the flatness of the loss surface, however, they unwillingly reduce the local maxima of the loss surface, limiting its attack potential. This reduction in local maxima can be attributed to the unstable gradients generated by the completely random sampling strategy, as illustrated in Fig. 2(b). Specifically, the averaged gradient roughly points towards flat regions, while it struggles to consistently align with the stable direction to the center of flat maxima regions, thereby reducing updating efficiency." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "To acquire stable gradient ascent direction and improve the efficiency of inner-iteration sampling, we define two es" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3888" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.485, + 0.173 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.175, + 0.355, + 0.188 + ], + "angle": 0, + "content": "(a) Random Sampling" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.19, + 0.484, + 0.267 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.198, + 0.269, + 0.392, + 0.282 + ], + "angle": 0, + "content": "(b) Momentum-Guided Sampling" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.285, + 0.484, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.205, + 0.341, + 0.384, + 0.355 + ], + "angle": 0, + "content": "(c) Gradient-Guided Sampling" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.491 + ], + "angle": 0, + "content": "Figure 3. The differences among three inner-iteration sampling guidance methods: (a) Random sampling in Eq. 3, with each inner-iteration sampling being independent; (b) Momentum-guided sampling in Eq. 4, where sampling direction depends on the cumulative average of all previous gradients, creating long-chain dependencies, while using random sampling for maintaining the randomness; (c) Gradient-guided sampling in Eq. 5, where sampling direction relies solely on the gradient direction of the previous iteration, establishing single-step dependencies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.598 + ], + "angle": 0, + "content": "sential characteristics that a well-balanced gradient direction should possess: (1) alignment with the stable gradient ascent direction for improving Exploitation capacity, and (2) pointing to loss regions with flatter surface for improving Exploration capacity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.614, + 0.339, + 0.631 + ], + "angle": 0, + "content": "3.3. Gradient-Guided Sampling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.484, + 0.851 + ], + "angle": 0, + "content": "(1) Stable gradient ascent direction: A sampled gradient in a sufficiently small neighborhood of the current example is probably a gradient ascent direction, while how can we ensure that this gradient ascent direction is stable? The Nesterov Accelerated Gradient (NAG) technique can precisely address this issue. It first allows the example to \"lookahead\" along the current momentum direction, and then acquires the gradient based on this new position, thereby identifying a more stable ascent direction [19, 23]. To leverage the superior stability of NAG based on random sampling, we perform the magnitude sampling along the momentum direction with an uniform distribution, and named this approach as Momentum-Guided Sampling (MGS). Compared to RS in Eq. 3, we only need to replace \\(\\tilde{x}_i\\) with \\(\\bar{x}_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.866, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\bar {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(m _ {i - 1}\\right)} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.244, + 0.893, + 0.365, + 0.904 + ], + "angle": 0, + "content": "lookahead sampling point" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.154 + ], + "angle": 0, + "content": "where \\( |\\tilde{p}| \\) denotes the element-wise absolute value of \\( \\tilde{p} \\), \\( \\mathrm{sign}(\\cdot) \\) denotes the sign function, which outputs either 1 or -1, and momentum decay is set to 1 for simplicity, so that \\( m_{i} = \\sum_{k=1}^{i} \\tilde{g}_{k} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.155, + 0.907, + 0.32 + ], + "angle": 0, + "content": "(2) Pointing to regions with better flatness: As shown in Fig. 3(b), if we use momentum for 'lookahead', the early unstable sampling in the inner-iterations will impose excessive constraints on subsequent sampling directions, leading to long-chain dependencies that severely impair RS's ability in exploring flat regions. Thus, for our GGS, we substitute the momentum term in Eq. 4 with the gradient obtained from the previous iteration, where the lookahead mechanism is maintained with the gradient ascent direction. Specifically, we only need to replace \\(\\tilde{x}_i\\) in Eq. 3 with \\(\\hat{x}_i\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.611, + 0.328, + 0.906, + 0.366 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i - 1}\\right)} _ {\\text {l o o k a h e a d s a m p l i n g p o i n t}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.534 + ], + "angle": 0, + "content": "Fig. 3 further sheds light on the differences among three inner-iteration sampling guidance methods. As shown in Fig. 3(c), GGS largely alleviates long-chain dependencies compared with Momentum-Guided Sampling in Fig. 3(b), while ensuring the sampling direction to be align with the gradient ascent direction, compared with random sampling in Fig. 3(a). In addition, as shown in Fig. 2(c), GGS will achieve stable sampling directions (② ③ ④) after a brief period of oscillation (①), when the initial sampling points are unstable." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.537, + 0.897, + 0.553 + ], + "angle": 0, + "content": "For clarity, our attack process is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.516, + 0.573, + 0.834, + 0.588 + ], + "angle": 0, + "content": "Algorithm 1 The attacking process of our GGS." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.593, + 0.906, + 0.67 + ], + "angle": 0, + "content": "Input: A clean image \\(x^{C\\times H\\times W}\\) with ground-truth label \\(y\\), and the loss function \\(\\mathcal{L}(\\cdot)\\); the magnitude of perturbation \\(\\epsilon\\); the sampling radius \\(\\zeta\\); the number of outer iterations, \\(T\\); the decay factor \\(\\gamma\\); the number of inner iterations, \\(N\\); \\(\\tilde{g}_0,\\tilde{p}\\in \\mathbb{R}^{C\\times H\\times W}\\)" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.67, + 0.778, + 0.685 + ], + "angle": 0, + "content": "Output: An adversarial example \\( x^{adv} \\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.686, + 0.815, + 0.7 + ], + "angle": 0, + "content": "1: Randomly sample \\(\\tilde{g}_0\\sim \\mathrm{Uniform}(-\\zeta ,\\zeta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.701, + 0.733, + 0.716 + ], + "angle": 0, + "content": "2: \\( v_{0} = 0 \\), \\( x_{0}^{adv} = x \\), \\( \\alpha = \\epsilon / T \\);" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.717, + 0.789, + 0.73 + ], + "angle": 0, + "content": "3: while \\( t \\gets 1 \\) to \\( T \\) (outer-iteration) do" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.686, + 0.815, + 0.73 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.732, + 0.816, + 0.745 + ], + "angle": 0, + "content": "4: while \\(i\\gets 1\\) to \\(N\\) (inner-iteration) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.746, + 0.901, + 0.761 + ], + "angle": 0, + "content": "5: Randomly sample noises \\(\\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)\\);" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.762, + 0.867, + 0.776 + ], + "angle": 0, + "content": "6: Get the lookahead example \\(\\hat{x}_i\\) by Eq. (5);" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.777, + 0.816, + 0.792 + ], + "angle": 0, + "content": "7: Get the gradient \\(\\tilde{g}_i = \\nabla_x\\mathcal{L}(\\hat{x}_i,y)\\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.793, + 0.638, + 0.804 + ], + "angle": 0, + "content": "8: end while" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.804, + 0.8, + 0.829 + ], + "angle": 0, + "content": "9: Update \\( v_{t} = \\gamma \\cdot v_{t - 1} + \\frac{\\sum_{i = 1}^{N}\\tilde{g}_{i}}{\\left\\|\\sum_{i = 1}^{N}\\tilde{g}_{i}\\right\\|_{1}} \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.829, + 0.871, + 0.846 + ], + "angle": 0, + "content": "0: Update \\( x_{t}^{adv} = \\Pi_{\\mathcal{B}_{\\epsilon}(x)}\\left[x_{t - 1}^{adv} + \\alpha \\cdot \\mathrm{sign}(v_{t})\\right] \\);" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.846, + 0.613, + 0.858 + ], + "angle": 0, + "content": "1: end while" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.859, + 0.629, + 0.875 + ], + "angle": 0, + "content": "12: \\(x^{adv} = x_T^{adv}\\)" + }, + { + "type": "list", + "bbox": [ + 0.521, + 0.686, + 0.901, + 0.875 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3889" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Table 1. The average untargeted and targeted ASR (%) on the nine models (Res50 [10], Dense121 [14], Inc-v3 [26], IncRes-v2 [27], ViT-B [3], PiT-B [11], and Inc-v3ens3, Inc-v3ens4, IncRes-v2ens [29]), with adversarial examples generated on a single model (Res50, Inc-v3 or ViT-B). Each data pair \\((u/w)\\) corresponds to the performances under (untargeted/targeted) attacks. The best and second best results are labeled in bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.158, + 0.911, + 0.525 + ], + "angle": 0, + "content": "
AttackRes50Dense121Inc-v3IncRes-v2ViT-BPiT-BInc-v3ens3Inc-v3ens4IncRes-v2ensAvg.
Res50MI CVPR'1899.8 / 98.154.9 / 0.244.2 / 0.028.0 / 0.011.8 / 0.022.4 / 0.022.8 / 0.025.3 / 0.019.6 / 0.036.53 / 10.92
N ICLR'20100.0 / 95.163.1 / 0.247.2 / 0.032.3 / 0.012.9 / 0.023.7 / 0.023.0 / 0.027.5 / 0.019.0 / 0.038.74 / 10.59
VMI CVPR'2199.8 / 96.671.2 / 2.858.1 / 0.549.4 / 0.231.3 / 0.147.7 / 0.543.9 / 0.045.4 / 0.137.7 / 0.153.83 / 11.21
RAP NeurIPS'2299.6 / 26.486.8 / 0.568.4 / 0.052.6 / 0.022.2 / 0.040.5 / 0.136.2 / 0.036.7 / 0.025.5 / 0.052.06 / 3.00
GRA ICCV'2396.9 / 67.788.6 / 3.981.8 / 0.975.8 / 1.745.3 / 0.262.6 / 0.671.5 / 0.770.9 / 1.167.3 / 0.773.41 / 8.61
PGN NeurIPS'2398.6 / 49.591.3 / 4.785.0 / 1.478.5 / 1.749.7 / 0.467.8 / 1.074.9 / 0.672.9 / 1.170.1 / 1.276.53 / 6.84
ANDA CVPR'2499.9 / 97.687.5 / 5.473.6 / 0.466.3 / 0.842.3 / 0.161.8 / 1.151.0 / 0.152.9 / 0.146.7 / 0.164.67 / 11.74
GI ESWA'24100.0 / 98.568.7 / 0.455.6 / 0.137.6 / 0.115.3 / 0.029.0 / 0.028.8 / 0.030.4 / 0.122.4 / 0.043.09 / 11.02
GGS99.3 / 90.795.9 / 28.689.6 / 7.286.4 / 8.660.2 / 3.481.0 / 8.577.5 / 3.777.4 / 4.371.4 / 4.082.08 / 17.67
Inc-v3MI32.1 / 0.049.6 / 0.1100.0 / 95.356.1 / 0.113.0 / 0.017.3 / 0.034.3 / 0.038.0 / 0.026.3 / 0.040.74 / 10.61
NI41.4 / 0.060.5 / 0.1100.0 / 75.966.9 / 0.113.4 / 0.019.2 / 0.036.8 / 0.040.6 / 0.129.5 / 0.045.37 / 8.47
VMI44.9 / 0.165.6 / 0.199.9 / 76.672.7 / 0.418.3 / 0.025.3 / 0.048.9 / 0.152.5 / 0.040.9 / 0.052.11 / 8.59
RAP58.0 / 0.078.4 / 0.2100.0 / 16.482.1 / 0.115.4 / 0.023.8 / 0.044.1 / 0.147.0 / 0.033.9 / 0.053.63 / 1.87
GRA53.3 / 0.175.4 / 0.499.8 / 67.585.4 / 1.821.4 / 0.030.5 / 0.164.9 / 0.165.9 / 0.356.1 / 0.161.41 / 7.82
PGN56.0 / 0.079.4 / 0.2100.0 / 50.387.9 / 0.923.2 / 0.033.4 / 0.065.9 / 0.067.0 / 0.155.6 / 0.063.16 / 5.72
ANDA51.3 / 0.376.3 / 0.4100.0 / 87.482.5 / 0.620.9 / 0.031.8 / 0.158.2 / 0.261.2 / 0.249.3 / 0.059.06 / 9.91
GI39.9 / 0.059.3 / 0.3100.0 / 97.162.6 / 0.514.6 / 0.019.9 / 0.039.2 / 0.040.9 / 0.031.9 / 0.045.37 / 10.88
GGS69.1 / 1.186.9 / 1.9100.0 / 76.095.8 / 6.527.9 / 0.039.7 / 0.170.0 / 0.773.9 / 1.460.1 / 0.569.27 / 9.80
ViT-BMI41.7 / 0.055.6 / 0.058.1 / 0.039.4 / 0.0100.0 / 99.950.1 / 0.034.2 / 0.036.3 / 0.029.9 / 0.049.48 / 11.10
NI42.4 / 0.058.3 / 0.059.5 / 0.044.7 / 0.1100.0 / 99.050.0 / 0.236.2 / 0.037.6 / 0.031.0 / 0.051.08 / 11.03
VMI54.8 / 0.665.9 / 0.463.1 / 0.150.9 / 0.3100.0 / 99.067.6 / 1.545.3 / 0.144.7 / 0.038.7 / 0.259.00 / 11.36
RAP62.1 / 0.179.6 / 0.273.7 / 0.062.3 / 0.099.9 / 57.066.2 / 0.148.5 / 0.051.2 / 0.042.6 / 0.065.12 / 6.38
GRA64.9 / 1.077.6 / 1.174.7 / 0.868.8 / 0.899.3 / 91.780.9 / 4.761.5 / 0.264.0 / 0.157.7 / 0.272.16 / 11.18
PGN69.3 / 0.481.1 / 0.778.6 / 0.470.2 / 0.599.1 / 82.284.8 / 4.465.6 / 0.167.5 / 0.361.4 / 0.275.29 / 9.91
ANDA66.2 / 0.379.0 / 0.575.5 / 0.165.7 / 0.3100.0 / 99.979.0 / 1.455.3 / 0.156.7 / 0.149.9 / 0.269.70 / 11.44
GI55.8 / 0.169.6 / 0.267.1 / 0.052.5 / 0.1100.0 / 100.061.0 / 0.345.4 / 0.046.3 / 0.038.0 / 0.059.52 / 11.19
GGS80.8 / 6.089.9 / 6.587.4 / 4.682.3 / 4.6100.0 / 99.292.7 / 27.374.0 / 2.174.5 / 2.368.4 / 2.483.33 / 17.22
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.55, + 0.224, + 0.567 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.299, + 0.592 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.657 + ], + "angle": 0, + "content": "Dataset: Our experiments are conducted on an ImageNet-compatible dataset that is widely used in adversarial attack research. The dataset includes 1,000 images, each with a size of \\(299 \\times 299 \\times 3\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.795 + ], + "angle": 0, + "content": "Models: We select six normally pre-trained models, i.e., ResNet50 (Res50) [10], DenseNet121 (Dense121) [14], Inception-v3 (Inc-v3) [26], InceptionResNet-v2 (IncRes-v2) [27], VisionTransformer-Base (ViT-B) [3], and Pooling-based VisionTransformer-Base (PiT-B) [11], along with three models pre-trained on adversarial examples Inception-v3_ens3 (Inc-v3_ens3), Inception-v3_ens4 (Inc-v3_ens4) and InceptionResNet-v2_ens (IncRes-v2_ens) [29], to evaluate the performance of different methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Baselines: To provide a comprehensive evaluation of our method, we choose eight popular gradient-based adversarial attack methods as baselines: MI-FGSM [1], NIFGSM [19], VMI-FGSM [32], RAP [24], GRA [40], PGN [7], ANDA [5] and GI [30]. Additionally, we integrate our method with five input transformation-based methods, i.e., DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.551, + 0.814, + 0.567 + ], + "angle": 0, + "content": "to verify the generalizability of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Hyperparameters: For our GGS, we set the maximum perturbation \\(\\epsilon = 16 / 255\\), the number of outer-iterations \\(T = 10\\), the step size \\(\\alpha = \\epsilon / T\\), the number of inner-iterations \\(N = 20\\), and the upper bound factor of sample range \\(\\zeta = 2.0 \\times \\epsilon\\). More hyperparameter explanation for other methods is provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.675, + 0.761, + 0.69 + ], + "angle": 0, + "content": "4.2. Attack on Different Models" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.699, + 0.702, + 0.713 + ], + "angle": 0, + "content": "4.2.1. Classification model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "As shown in Table 1, our method consistently outperforms state-of-the-art (SOTA) approaches. Compared to the best-performing competitor, it improves the average attack success rate (ASR) by over \\(5\\%\\) in both untargeted and targeted attacks across most architectures, except for Inception-v3. Furthermore, our GGS significantly enhances attack performance on the surrogate model, achieving \\(4\\% -25\\%\\) higher ASR in targeted attacks than other sampling methods like PGN [7] and GRA [40]. This demonstrates GGS's ability to boost attack potency while maintaining strong cross-model generalization. Additional details on ensemble setting can be found in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3890" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.09, + 0.086, + 0.912, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.358, + 0.908, + 0.442 + ], + "angle": 0, + "content": "Figure 4. The loss surfaces of adversarial examples generated by different methods (MI-FGSM [1], GRA [40], PGN [7], ANDA [5]), based on three different model architectures (ResNet50 [10], Inception-v3 [26], and ViT-B [3]) along random directions with varying strengths like [7]. In the images marked with a star in the upper left corner, adversarial examples are generated and tested on the same model, indicating white-box testing. In contrast, unmarked images represent adversarial examples generated on one model and tested on another, indicating black-box testing. We have also highlighted the regions covered by our loss surface using a red background for easier visualization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.395, + 0.484 + ], + "angle": 0, + "content": "4.2.2. Multimodal Large Language Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.488, + 0.484, + 0.609 + ], + "angle": 0, + "content": "The widespread adoption of Multimodal Large Language Models (MLLMs) has raised significant security concerns. We evaluated our method on five prominent MLLMs using adversarial examples generated under an ensemble setting, as shown in Table 2. Our GGS method achieved superior attack performance, reducing the average CSR by over \\(9\\%\\) compared to the strongest baseline, which demonstrates its excellent cross-model applicability." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.623, + 0.484, + 0.707 + ], + "angle": 0, + "content": "Table 2. Classification success rate (CSR, %) of adversarial examples generated by different methods in Multimodal Large Language Models (MLLMs) from three major providers (OpenAI, Google, Anthropic) under ensemble settings (Res50 [10], Inc-v3 [26], ViT-B [3]). Row \"clean\" indicates the CSR on the clean samples. Avg. represents the average CSR across the five MLLMs." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.71, + 0.481, + 0.898 + ], + "angle": 0, + "content": "
MethodGPTGeminiClaude
4ominiproflashsonnetAvg. ↓
Clean77.184.085.680.968.379.18
MI_CVPR'18 [1]77.077.381.876.561.074.72
NI_ICLR'20 [19]75.073.983.176.862.474.24
VMI_CVPR'21 [32]68.866.678.473.263.170.02
RAP_NeurIPS'22 [24]64.362.075.166.160.265.54
GRA_ICCV'23 [40]57.152.270.965.646.758.50
PGN_NeurIPS'23 [7]56.249.069.664.145.156.80
ANDA_CVPR'24 [5]59.956.874.969.350.562.28
GI_ESWA'24 [30]71.572.879.473.263.872.14
GGS43.137.861.155.740.047.54
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.468, + 0.889, + 0.483 + ], + "angle": 0, + "content": "4.2.3. Loss Surface Visualization against Noise Levels" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.907, + 0.715 + ], + "angle": 0, + "content": "To shed light on the capacity of our GGS in locating flat maxima regions for adversarial examples, we compared the loss surfaces of adversarial examples generated by different attack methods in Fig. 4. Each curve represents the average loss of 32 randomly selected adversarial examples, while the center of each image (noise level \\(= 0\\)) indicates the average loss of clean adversarial examples. In white-box settings (marked with a star), GGS consistently finds flatter maxima compared to baselines. Consequently, in black-box testing, GGS achieves the highest loss values (images without star), indicating its superior transfer attack capability. Additionally, the GGS loss surface (highlighted in red) almost entirely encompasses those of other methods, confirming its effectiveness in balancing strong attack potency with cross-model generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.726, + 0.818, + 0.743 + ], + "angle": 0, + "content": "4.3. Compatibility with Other Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Attack with Inner-Iteration Sampling Methods: Our Gradient-Guided Sampling (GGS) is compatible with existing inner-iteration RS-based approaches. When integrated with SOTA approaches like GRA [40] and PGN [7], GGS largely improves attack success rates (ASR). As shown in Table 3, untargeted ASR improved by \\(5.2\\%\\) (GRA) and \\(6.7\\%\\) (PGN), while targeted ASR increased by \\(5\\%\\) (GRA) and \\(6\\%\\) (PGN), demonstrating GGS's compatibility." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Attack with Input Transformations Methods: To study the compatibility of our GGS with existing input" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3891" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.087, + 0.295, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.198, + 0.243, + 0.21 + ], + "angle": 0, + "content": "(a) Random (RS)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.089, + 0.49, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.318, + 0.198, + 0.476, + 0.21 + ], + "angle": 0, + "content": "(b) Momentum-Guided (MGS)" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.089, + 0.685, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.532, + 0.197, + 0.672, + 0.209 + ], + "angle": 0, + "content": "(c) Gradient-Guided (GGS)" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.089, + 0.905, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.189, + 0.902, + 0.209 + ], + "angle": 0, + "content": "(d) The cosine similarity between the gradient and the gradients averaging in inner-iterations" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.905, + 0.28 + ], + "angle": 0, + "content": "Figure 5. Loss surfaces of adversarial examples generated by different sampling strategies, Random Sampling (RS), MGS and our GGS, with increasing inner-iteration on Resnet50. The red line on the left side of part (a)~(c) represents the maximum value of the loss surface in each iteration, and the blue dashed line highlights these values in early-stage inner iteration. For (d), it represents the cosine similarity between the gradient \\(\\tilde{g}_i\\) generated in each inner-iteration and the average of gradients \\(\\sum_{i=1}^{N} \\tilde{g}_i / N\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.302, + 0.486, + 0.372 + ], + "angle": 0, + "content": "Table 3. The average untargeted and targeted ASR \\((\\%)\\) on all nine models with the adversarial examples generated on ResNet50, when combined our GGS with input transformation-based methods (DIM [37], TIM [2], SIM [19], Admix [33], SSM [21]) and inner-iteration RS-based methods (GRA [40], PGN [7])." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.383, + 0.487, + 0.517 + ], + "angle": 0, + "content": "
ClassMethodUntargeted Avg.Targeted Avg.
Random Sampling (RS)GRA / +ours73.41 / 78.698.61 / 13.64
PGN / +ours76.53 / 83.236.84 / 12.79
Input TransformationDIM / +ours51.30 / 90.138.63 / 19.37
TIM / +ours40.82 / 84.9411.03 / 18.48
SIM / +ours46.80 / 90.1211.14 / 28.44
Admix / +ours54.78 / 85.5011.28 / 30.50
SSM / +ours74.22 / 80.748.40 / 12.54
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.676 + ], + "angle": 0, + "content": "transformation techniques, we integrated it with five methods of this type: DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]. Adversarial examples generated on ResNet50 were tested on nine diverse models. Table 3 shows GGS significantly improved adversarial transferability, increasing untargeted ASR by \\(6\\% -43\\%\\) and targeted ASR by \\(4\\% -19\\%\\). For instance, SIM's untargeted ASR improved by \\(43\\%\\) when combined with GGS. More details are in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.684, + 0.441, + 0.701 + ], + "angle": 0, + "content": "4.4. Ablation Study of Proposed Components" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.718, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Table 4. The average untargeted ASR on ResNet50 and eight target models for Random Sampling (RS), Momentum-Guided Sampling (MGS), and our Gradient-Guided Sampling (GGS)." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.771, + 0.475, + 0.858 + ], + "angle": 0, + "content": "
Sampling Guided byRes50Others
randmomentumgradient
(a) RS✓--97.363.74
(b) MGS-✓-97.558.79
(c) GGS--✓99.379.93
" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Ablation study compares RS, MGS, and GGS via loss surface analysis (Fig. 5) and attack success rates (Table 4)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.907, + 0.501 + ], + "angle": 0, + "content": "As shown in Fig. 5(a), Random Sampling (RS) yields a relatively flat loss surface due to its exploration capability but it suffers from significant oscillation in loss values. Introducing Momentum-Guided Sampling (MGS) (Fig. 5(b)) stabilizes the loss surface over iterations, but its long-chain dependency causes high similarity between later gradients and the final average gradient (Fig. 5(d)), severely limiting exploration. Our GGS resolves MGS's long-chain dependency and mitigates early-stage noisy gradients (Fig. 5(b)(c), blue dashed lines). The low gradient similarity in Fig. 5(d) reflects the enhanced exploration, while the flat yet higher-loss surface in Fig. 5(c) demonstrates the exceptional sampling efficiency of our GGS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.501, + 0.906, + 0.577 + ], + "angle": 0, + "content": "Quantitatively, as shown in Table 4, MGS marginally improves ASR of the surrogate model by \\(0.2\\%\\), but reducing transfer attack capability by \\(5\\%\\). In contrast, GGS increases both white-box and transfer ASR by \\(2\\%\\) and \\(16\\%\\), respectively, validating its superior sampling efficiency." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.776, + 0.607 + ], + "angle": 0, + "content": "5. Conclusion and Perspectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.84 + ], + "angle": 0, + "content": "This paper introduces Gradient-Guided Sampling (GGS), an easy-to-implement inner-iteration sampling strategy, for transferable adversarial attacks that effectively resolves the dilemma between Exploitation (attack potency) and Exploration (cross-model generalization). By guiding the sampling direction with the gradient from the previous inner iteration, GGS enables stable ascent toward flat loss regions with higher local maxima. Our GGS has the merit of enhancing exploitation capability without compromising exploration capability, effectively balancing these two aspects and addressing the inherent limitations of random sampling (RS) methods, as verified by extensive experiments and visualizations as well as the comparison with related state-of-the-art methods. GGS is also revealed to be highly compatible with existing methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "While our GGS is highly compatible with gradient-averaging methods (e.g., GRA [40], PGN [7]), we aim to refine it to also support non-gradient-averaging techniques (e.g., VMI-FGSM [32], RAP [24])." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3892" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.283 + ], + "angle": 0, + "content": "The work was supported by the National Natural Science Foundation of China under grants no. 62276170, 82261138629, 62306061, the Science and Technology Project of Guangdong Province under grants no. 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grant no. JCYJ20220531101412030, Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) under Grant No. GML-KF-24-11, and Guangdong Provincial Key Laboratory under grant no. 2023B1212060076." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.296, + 0.188, + 0.312 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.321, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[1] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.392, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[2] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4312-4321, 2019. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.463, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[4] Yan Fang, Zhongyuan Wang, Jikang Cheng, Ruoxi Wang, and Chao Liang. Promoting adversarial transferability with enhanced loss flatness. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1217-1222. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.485, + 0.687 + ], + "angle": 0, + "content": "[5] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[6] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[7] Zhijin Ge, Hongying Liu, Wang Xiaosen, Fanhua Shang, and Yuanyuan Liu. Boosting adversarial transferability by achieving flat local maxima. Advances in Neural Information Processing Systems, 36:70141-70161, 2023. 1, 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.484, + 0.885 + ], + "angle": 0, + "content": "[8] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.888, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy." + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.321, + 0.485, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Explaining and harnessing adversarial examples. In International Conference on Learning Representations, ICLR, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.191 + ], + "angle": 0, + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.909, + 0.262 + ], + "angle": 0, + "content": "[11] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11936-11945, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.29 + ], + "angle": 0, + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.907, + 0.36 + ], + "angle": 0, + "content": "[13] SM Mostaq Hossain, Shampa Banik, Trapa Banik, and Ashfak Md Shibli. Survey on security attacks in connected and autonomous vehicular systems. In 2023 IEEE International Conference on Computing (ICOCO), pages 295-300. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.906, + 0.419 + ], + "angle": 0, + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.421, + 0.906, + 0.475 + ], + "angle": 0, + "content": "[15] Martin Husák, Jana Komárková, Elias Bou-Harb, and Pavel Čeleda. Survey of attack projection, prediction, and forecasting in cyber security. IEEE Communications Surveys & Tutorials, 21(1):640-660, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.906, + 0.532 + ], + "angle": 0, + "content": "[16] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. In Advances in Neural Information Processing Systems, pages 73919-73942. Curran Associates, Inc., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.534, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[17] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. ICLR, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[18] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.906, + 0.703 + ], + "angle": 0, + "content": "[19] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, ICLR, 2020. 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.906, + 0.773 + ], + "angle": 0, + "content": "[20] Decheng Liu, Xijun Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Adv-diffusion: Imperceptible adversarial face identity attack via latent diffusion model. Proceedings of the AAAI Conference on Artificial Intelligence, 38(4):3585-3593, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[21] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European conference on computer vision, pages 549-566. Springer, 2022. 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[22] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, ICLR, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3893" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[23] Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Dokl. Akad. Nauk. SSSR, page 543, 1983. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[24] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in neural information processing systems, 35:29845-29858, 2022. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[25] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.318 + ], + "angle": 0, + "content": "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[27] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-reset and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[28] Chee-Wooi Ten, Govindarasu Manimaran, and Chen-Ching Liu. Cybersecurity for critical infrastructures: Attack and defense modeling. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 40(4):853–865, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.483, + 0.489 + ], + "angle": 0, + "content": "[29] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. *ICLR*, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[30] Jiafeng Wang, Zhaoyu Chen, Kaixun Jiang, Dingkang Yang, Lingyi Hong, Pinxue Guo, Haijing Guo, and Wenqiang Zhang. Boosting the transferability of adversarial attacks with global momentum initialization. Expert Systems with Applications, 255:124757, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[31] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[32] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1924-1933, 2021. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[33] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[34] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[35] Han Wu, Syed Yunas, Sareh Rowlands, Wenjie Ruan, and Johan Wahlström. Adversarial driving: Attacking end-to-end autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[36] Chaowei Xiao, Bo Li, Jun yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 3905-3911, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[37] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[38] Yinhu Xu, Qi Chu, Haojie Yuan, Zixiang Luo, Bin Liu, and Nenghai Yu. Enhancing adversarial transferability from the perspective of input loss landscape. In International Conference on Image and Graphics, pages 254-266. Springer, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[39] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[40] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4741-4750, 2023. 2, 3, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[41] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3894" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_origin.pdf b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9f83202054e5aa322ac6f9c8a2132731a7deb2f3 --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/6da66727-c69d-4b80-9f15-0e506723c7cf_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6e977711927f20b74429c6ec5cbc6248c137c00ed483c5fb2bb38d37e4cad271 +size 6375099 diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/full.md b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1f05e643cded032e1dac80288ce66fade0b854cf --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/full.md @@ -0,0 +1,317 @@ +# Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling + +Zenghao Niu $^{1}$ Weicheng Xie $^{1,2,3}$ * Siyang Song $^{4}$ Zitong Yu $^{5}$ Feng Liu $^{1}$ Linlin Shen $^{6}$ + +$^{1}$ School of Computer Science & Software Engineering, Shenzhen University, China + +$^{2}$ Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China + +3Guangdong Provincial Key Laboratory of Intelligent Information Processing, Shenzhen University, China + +$^{4}$ School of Computer Science, University of Exeter, U.K. + +$^{5}$ Department of Computing and Information Technology, Great Bay University, China + +6Computer Vision Institute, School of Artificial Intelligence, Shenzhen University, China + +{2300271072@mail., wcxie@, feng.liu@, llshen@}szu.edu.cn + +s song@exeter.ac.uk, zitong.yu@ieee.org + +# Abstract + +Adversarial attacks present a critical challenge to deep neural networks' robustness, particularly in transfer scenarios across different model architectures. However, the transferability of adversarial attacks faces a fundamental dilemma between Exploitation (maximizing attack potency) and Exploration (enhancing cross-model generalization). Traditional momentum-based methods over-prioritize Exploitation, i.e., higher loss maxima for attack potency but weakened generalization (narrow loss surface). Conversely, recent methods with inner-iteration sampling over-prioritize Exploration, i.e., flatter loss surfaces for cross-model generalization but weakened attack potency (suboptimal local maxima). To resolve this dilemma, we propose a simple yet effective Gradient-Guided Sampling (GGS), which harmonizes both objectives through guiding sampling along the gradient ascent direction to improve both sampling efficiency and stability. Specifically, based on MI-FGSM, GGS introduces inner-iteration random sampling and guides the sampling direction using the gradient from the previous inner-iteration (the sampling's magnitude is determined by a random distribution). This mechanism encourages adversarial examples to reside in balanced regions with both flatness for cross-model generalization and higher local maxima for strong attack potency. Comprehensive experiments across multiple DNN architectures and multimodal large language models (MLLMs) demonstrate the superiority of our method over state-of-the-art transfer attacks. Code is made available at https://github.com/anuin-cat/GGS. + +![](images/5ce4302b0c4f1e496a52dafe22ed071a9152d827637a64291246dde7d2d47fb4.jpg) +Figure 1. The loss surfaces of (a) MI-FGSM [1] (Momentum iterative fast gradient sign method), (b) RS (Base inner-iteration Random Sampling defined in section 3.2) for enhancing exploration within the neighborhood, (c) PGN [7] (Penalizing Gradient Norm) with Random Sampling (RS) for stable gradient estimation to enhance exploration, and (d) Our GGS (Gradient-Guided Sampling) for efficient sampling to generate gradients that stably towards flat regions with higher local maxima. It shows that, building upon RS, our approach not only maintains a flat loss surface, but also improves the local maximum loss value, compared to PGN, achieving a balance between exploration and exploitation. + +# 1. Introduction + +Researches have indicated that adding imperceptible perturbations to input data can easily affect deep neural networks + +(DNNs), leading to erroneous predictions in safety-critical domains such as autonomous driving [13, 35], and cybersecurity [15, 28]. To facilitate the development of relevant defense mechanisms, in-depth research on adversarial attack methods is urgently needed. + +While adversarial attacks are categorized into white-box and black-box settings, the black-box scenario is more realistic in practical applications, where the attacker lacks knowledge of the target model and relies solely on a surrogate model, aiming to generate adversarial examples that are effective across multiple target models. To enhance the transferability of adversarial examples in this setting, researchers have explored gradient-based methods [1, 7, 32], input transformation methods [2, 33, 37], and generative methods [16, 20, 36] to mitigate overfitting of adversarial examples on surrogate models. Despite significant progress, the effectiveness of attacks on target models remains substantially lower than that on surrogate models, driving the need for more effective attack strategies. + +In black-box adversarial attacks, RAP [24] adopted Sharpness-Aware Minimization (SAM [6]) to improve transferability by flattening the loss surface. Building on this, PGN [7] further enhanced the transferability of adversarial examples by penalizing gradient norm and flattening the loss surface. The efficacy of flat loss surface in improving adversarial attack transferability is increasingly supported [4, 7, 38]. To enhance exploration through local flatness, recent approaches, building on MI-FGSM's [1] momentum-based outer-iterations, often incorporates inner-iterations before each example update [7, 40], i.e., performing multi-step neighborhood sampling for stabilization. + +However, these methods enhance Exploration (cross-model generalization) at the cost of compromising Exploitation (attack potency). This is because, although inner-iterations can significantly enhance exploration capability through neighborhood sampling, independent random sampling struggles to obtain stable and effective gradient directions, while potential noise may cause the averaged gradient to deviate from the optimal ascent toward target regions, which exhibit flatter loss surface and higher local maxima. To optimize sampling efficiency, a guided sampling constraint is essential, which should satisfy two criteria: (1) ensuring stable gradient ascent direction and (2) directing gradients toward flatter regions, thereby enabling samples to reach flatter local maxima regions with enhanced transferability, balancing exploitation and exploration. + +For (1) stable gradient direction, a sufficiently small neighborhood sampling followed by gradient calculation generally yields correct ascent directions. Yet, how can we further stabilize this direction? The Nesterov Accelerated Gradient (NAG) [23] enhances stability by firstly updating samples along the current momentum (lookahead) + +before acquiring gradients, providing more stable directions. Therefore, momentum-based direction constraints are promising for sampling, as the additional random magnitude can be used to preserve the sampling randomness. (2) However, directly applying NAG's momentum mechanism during inner-iteration sampling imposes strong constraints on subsequent directions, restricts the exploration capability for flat regions. To address this, we replace momentum with the gradient from the previous inner-iteration, named Gradient-Guided Sampling (GGS). This adjustment ensures that sampling to align with gradient ascent directions while reducing the impact from early-stage instability. Our GGS can achieve stable sampling directions after brief oscillations, consistently targeting flat loss regions with larger local maxima regardless of the initial sampling quality. + +As shown in Fig. 1(d), our GGS generates adversarial examples with flatter loss surface and higher local maxima compared to RS, i.e., it does not sacrifice local maxima in the pursuit of a flatter surface compared to PGN, thereby achieving a balance between exploration and exploitation to facilitate attack transferability. Additionally, since our approach represents a complementary improvement over Random Sampling (RS), it is compatible with RS-based methods, such as inner-iteration random sampling and gradient averaging, including PGN [7] and GRA [40]. In summary, this study makes the following contributions: + +- We propose Gradient-Guided Sampling (GGS), a novel inner-iteration sampling strategy that effectively balances exploration (cross-model generalization) and exploitation (attack potency). By leveraging the gradient from the previous inner-iteration to guide sampling directions, GGS achieves stable gradient ascent toward flat loss regions with higher local maxima. +- The GGS framework is compatible with existing inner-iteration random sampling-based methods, enhancing their sampling efficiency and further improving the transferability of the adversarial examples they generate. +- Extensive experimental results as well as comprehensive visualizations demonstrate the efficacy of our GGS, covering targeted and non-targeted attacks in cross-architecture black-box scenarios, non-targeted attacks on multimodal large language models (MLLMs) and commercial cloud functions. + +# 2. Related Work + +Transferability is a critical property in adversarial attacks, allowing adversarial examples generated from a surrogate model to deceive unseen target models in black-box settings. This characteristic makes transfer-based attacks highly practical in real-world applications where target models are inaccessible, motivating our focus on transferable adversarial attacks. + +# 2.1. Transferable Attack + +Gradient-based Attack. Gradient-based attacks aim to generate adversarial examples by exploiting the gradients of a surrogate model to maximize the loss. Classic methods such as Fast Gradient Sign Method (FGSM) [9] and Projected Gradient Descent (PGD) [22] optimize adversarial perturbations to fool the surrogate model. However, these attacks often overfit the surrogate model, resulting in limited transferability to target models. To address this, advanced techniques such as momentum-based (MI-FGSM [1]), Nesterov Iterative-based (NI-FGSM [19]), variance-tuned (VMI-FGSM [32]), gradient relevance-based (GRA [40]), momentum initialization (GI-FGSM [30]) and distribution-based methods (ANDA [5]) have been introduced to smooth out the optimization process and improve transferability across models. + +Input Transformation. To mitigate the overfitting issue inherent in gradient-based methods, input transformation techniques have been proposed to enhance the diversity of adversarial examples. Techniques like translation, resizing, padding, image mixup (e.g., DIM [37], SIM [19], TIM [2], Admix [33]) and other advanced image transformation methods (e.g., SSM [21], SIA [34], STM [8], L2T [41], BSR [31]) introduce randomness to the input data before generating perturbations. By applying such transformations during the attack process, more effective transferable adversarial examples can be generated. + +# 2.2. Flat Maxima + +The generalization ability of models has been suggested to possibly have a certain association with flat minima [12]. More in-depth research and exploration have gradually confirmed this perspective [6, 17, 18, 39]. Moreover, flat maxima also have been validated to be effective for enhancing the generalization and transferability of adversarial examples [24]. Flat maxima refers to regions in the loss landscape, where small changes in the model parameters result in minimal changes in the loss, making adversarial examples generated in these regions less sensitive to the specific decision boundaries of models. This property contrasts with "sharp maxima", which are highly sensitive to small perturbations and can lead to overfitting to the surrogate model, reducing transferability capacity. + +In the context of transfer-based attacks, works like Reverse Adversarial Perturbation (RAP) [24] leverage flat maxima to generate adversarial examples, PGN [7] adopts a first-order procedure to approximate the Hessian/vector product, largely improving computational efficiency. They are not only effective on the surrogate model but also more resilient when attacking unseen target models. These methods mitigate overfitting, via making adversarial examples lie in regions where the loss function remains stable across different models (e.g., flat maxima regions), improving at + +tack success rates in black-box settings. However, achieving flat maxima requires not only attention to "flatness" but also preservation of the "maxima" to ensure sufficient attack strength during transfer attack, which poses a significant challenge. + +# 2.3. Inner-Iteration Sampling + +To enhance the generalization capability of adversarial examples, recent methods have introduced inner-iteration sampling [7, 24, 32]. Specifically, an inner iterative process is inserted before updating the examples. This process involves neighborhood sampling to acquire domain information during inner-iterations which are used to adjust gradients [32] or neighborhood search is incorporated to locate local minima and enhance their loss values [24]. Furthermore, works like [7, 40] enhance the flatness of the loss landscape by averaging gradients from inner-iterations to update the examples. + +Additionally, neighborhood sampling can facilitate examples to escape from sharp local maxima regions, thereby avoiding overfitting to the surrogate model. However, directly averaging the gradients from inner-iterations will reduce the maximum value of the loss surface, since random sampling produces unstable gradient directions that fail to consistently direct to flat regions with higher local maxima. Thus, enhancing inner-iteration sampling efficiency to stabilize the final gradient has become a critical challenge. + +# 3. Methodology + +# 3.1. Preliminaries + +Transferable Adversarial Attacks. For a given example $(x,y)$ , a surrogate model $f_{\theta}$ and multiple target models $f_{\varphi_k}$ for $k \in \{1,2,\dots,K\}$ , the attacker's goal is to find $x^{adv} = x + \delta$ using $f_{\theta}$ to make $f_{\varphi_k}(x^{adv}) \neq y$ , where $\| \delta \|_p < \epsilon$ , $\epsilon$ denotes the maximum magnitude of the specified perturbation and $\| \cdot \|_p$ denotes the $\ell_p$ -norm. For targeted attacks, we simply adapt the symbols and labels in the target function: $f_{\varphi_k}(x^{adv}) = y_t$ , where $y_t$ denotes the target label. Meanwhile, to facilitate a clearer representation of the norm constraints against $x^{adv}$ as below, we define $\mathcal{B}_{\epsilon}(x) = \{x': \| x' - x \|_p \leq \epsilon\}$ to denote the $\epsilon$ -ball of an input image $x$ . For the purpose of the above attack, we generally need to maximize the following objective function to generate the transferable adversarial examples: + +$$ +\max _ {x ^ {a d v} \in \mathcal {B} _ {\epsilon} (x)} \mathcal {L} \left(x ^ {a d v}, y\right), \tag {1} +$$ + +where $\mathcal{L}(\cdot)$ represents the loss function. In the above flow, we assume that $f_{\theta}$ is known as a white-box model and $f_{\varphi_k}$ is unknown as a black-box model. Similar to the process of training a neural network, the white-box model can be seen as the training data, with the adversarial examples serving + +as the optimization parameters, while the black-box model acts as the test data. + +Nesterov Accelerated Gradient (NAG). Due to the similarity between model training task and adversarial example generation, improving performance in black-box settings can be viewed as enhancing model generalization in model training tasks. Consequently, some commonly used optimization methods, such as Momentum [25] and Nesterov Accelerated Gradient (NAG) [23], can be introduced to improve the generalization of adversarial examples across black-box models through a more stable optimization process. + +NI-FGSM [19] was the first to incorporate the principles of NAG into adversarial attack tasks. This enables the algorithm to acquire directional information in advance, reducing oscillations and achieving stable convergence. Specifically, momentum is computed at each step as follows: + +$$ +v _ {t} = \gamma \cdot v _ {t - 1} + \frac {\nabla_ {x} \mathcal {L} (\tilde {x} _ {t} , y)}{\left\| \nabla_ {x} \mathcal {L} (\tilde {x} _ {t} , y) \right\| _ {1}}, \tag {2} +$$ + +$$ +\tilde {x} _ {t} = \underbrace {x _ {t - 1} ^ {a d v} + \alpha \cdot \gamma \cdot v _ {t - 1}} _ {\text {l o o k a h e a d p o i n t}}, +$$ + +where $v_{t}$ represents the momentum term, $x_0^{adv} = x$ , $\alpha$ stands for the step size, $\gamma$ denotes the momentum decay factor, and $\| \cdot \|_1$ denotes the $\ell_1$ -norm of the given variable. Unlike traditional momentum-based methods, NI-FGSM utilizes the estimated position of examples rather than their current actual position, during gradient computation, i.e., introducing a lookahead mechanism into the algorithm. + +# 3.2. Motivation + +Recent transfer-based adversarial attack methods have introduced inner-iteration random sampling, to enhance algorithm stability [7, 40] or acquire neighborhood information [24, 32]. It can be noted that these methods have increased the flatness of the loss surface, as illustrated in Fig.1(c), with additional loss surface plots for other methods provided in the supplementary materials. If we isolate the inner-iteration random sampling (RS) component, it can be expressed in the following form: + +$$ +v _ {t} = \gamma v _ {t - 1} + \frac {\sum_ {i = 1} ^ {N} \tilde {g} _ {i}}{\left\| \sum_ {i = 1} ^ {N} \tilde {g} _ {i} \right\| _ {1}}, \tag {3} +$$ + +$$ +\tilde {g} _ {i} = \nabla_ {x} \mathcal {L} \big (\tilde {x} _ {i}, y \big), \text {w i t h} \tilde {x} _ {i} = \underbrace {x _ {t - 1} ^ {a d v} + \tilde {p}} _ {\text {s a m p l i n g p o i n t}}, +$$ + +where $\sum_{i=1}^{N}$ denotes the output of the inner-iteration process; $N$ denotes the number of inner-iterations; $\gamma$ denotes the momentum decay factor; $\tilde{g}_0, \tilde{p} \sim \mathrm{Uniform}(-\zeta, \zeta)$ , are uniform random noises with the same size as $x$ . + +![](images/69edcd1e097eb143568264f1d7471dc58e31a8c36fa14f47250086412fa7dfe5.jpg) +(a) Optimization Trajectories of Various Methods + +![](images/9db4dea360f9239cc1a12af075c01f27d6c466212e25fa60152bf5ff3f53e839.jpg) +(b) 1st step process of RS +Figure 2. (a) Outer-Iteration processes of MI-FGSM (green), RS (purple) and our GGS (red). (i) MI-FGSM prioritizes rapid ascent directions, thereby enhancing exploitation and facilitating access to sharp local maxima regions. (ii) RS introduces inner-iterations and enhances exploration capability through neighborhood search, thereby enabling access to flat local maxima regions. (iii) GGS incorporates gradient constraints into RS, simultaneously enhancing both exploration and exploitation capabilities, thereby enabling the faster convergence to the centers of flat local maxima regions. (b) RS is performed within a neighborhood of the current example, and uses the average gradient of all sampled points as the final gradient. (c) Building upon RS, GGS enables examples to have stable gradient directions (② ③ ④) toward the centers of flat local maxima regions, following an initial period of brief oscillation (①). It uses the gradient direction from the previous inner-iteration as the guidance, while the randomness is maintained by setting the gradient magnitude with a random distribution. + +![](images/73bc083d01e3f442ec4f2e455d7d7f8559ec04893f7b9c360b51a491edd85f45.jpg) +(c) 1st step process of GGS + +As shown in Fig. 1, the inner-iteration RS possesses a strong capability to enhance the flatness of the loss surface compared to MI-FGSM. Compared to RS, current methods such as PGN [7] can further improve the flatness of the loss surface, however, they unwillingly reduce the local maxima of the loss surface, limiting its attack potential. This reduction in local maxima can be attributed to the unstable gradients generated by the completely random sampling strategy, as illustrated in Fig. 2(b). Specifically, the averaged gradient roughly points towards flat regions, while it struggles to consistently align with the stable direction to the center of flat maxima regions, thereby reducing updating efficiency. + +To acquire stable gradient ascent direction and improve the efficiency of inner-iteration sampling, we define two es + +![](images/3ec578967648e41ed05fb22a323a14cc8103e97a0e4306a015df76a59cb6252d.jpg) +(a) Random Sampling + +![](images/07765863c5e00293a3429595b4e2a249f1b2e931802bf9b57d29c27f29d916ca.jpg) +(b) Momentum-Guided Sampling + +![](images/edaa2140ae66b9d6d89b7006827312421d9d38d09d4611bc1f4c6f6722165bc2.jpg) +(c) Gradient-Guided Sampling +Figure 3. The differences among three inner-iteration sampling guidance methods: (a) Random sampling in Eq. 3, with each inner-iteration sampling being independent; (b) Momentum-guided sampling in Eq. 4, where sampling direction depends on the cumulative average of all previous gradients, creating long-chain dependencies, while using random sampling for maintaining the randomness; (c) Gradient-guided sampling in Eq. 5, where sampling direction relies solely on the gradient direction of the previous iteration, establishing single-step dependencies. + +sential characteristics that a well-balanced gradient direction should possess: (1) alignment with the stable gradient ascent direction for improving Exploitation capacity, and (2) pointing to loss regions with flatter surface for improving Exploration capacity. + +# 3.3. Gradient-Guided Sampling + +(1) Stable gradient ascent direction: A sampled gradient in a sufficiently small neighborhood of the current example is probably a gradient ascent direction, while how can we ensure that this gradient ascent direction is stable? The Nesterov Accelerated Gradient (NAG) technique can precisely address this issue. It first allows the example to "lookahead" along the current momentum direction, and then acquires the gradient based on this new position, thereby identifying a more stable ascent direction [19, 23]. To leverage the superior stability of NAG based on random sampling, we perform the magnitude sampling along the momentum direction with an uniform distribution, and named this approach as Momentum-Guided Sampling (MGS). Compared to RS in Eq. 3, we only need to replace $\tilde{x}_i$ with $\bar{x}_i$ as follows: + +$$ +\bar {x} _ {i} = \underbrace {x _ {t - 1} ^ {a d v} + | \tilde {p} | \cdot \operatorname {s i g n} \left(m _ {i - 1}\right)} \tag {4} +$$ + +lookahead sampling point + +where $|\tilde{p}|$ denotes the element-wise absolute value of $\tilde{p}$ , $\mathrm{sign}(\cdot)$ denotes the sign function, which outputs either 1 or -1, and momentum decay is set to 1 for simplicity, so that $m_{i} = \sum_{k=1}^{i} \tilde{g}_{k}$ . + +(2) Pointing to regions with better flatness: As shown in Fig. 3(b), if we use momentum for 'lookahead', the early unstable sampling in the inner-iterations will impose excessive constraints on subsequent sampling directions, leading to long-chain dependencies that severely impair RS's ability in exploring flat regions. Thus, for our GGS, we substitute the momentum term in Eq. 4 with the gradient obtained from the previous iteration, where the lookahead mechanism is maintained with the gradient ascent direction. Specifically, we only need to replace $\tilde{x}_i$ in Eq. 3 with $\hat{x}_i$ as follows: + +$$ +\hat {x} _ {i} = \underbrace {x _ {t - 1} ^ {a d v} + | \tilde {p} | \cdot \operatorname {s i g n} \left(\tilde {g} _ {i - 1}\right)} _ {\text {l o o k a h e a d s a m p l i n g p o i n t}}, \tag {5} +$$ + +Fig. 3 further sheds light on the differences among three inner-iteration sampling guidance methods. As shown in Fig. 3(c), GGS largely alleviates long-chain dependencies compared with Momentum-Guided Sampling in Fig. 3(b), while ensuring the sampling direction to be align with the gradient ascent direction, compared with random sampling in Fig. 3(a). In addition, as shown in Fig. 2(c), GGS will achieve stable sampling directions (② ③ ④) after a brief period of oscillation (①), when the initial sampling points are unstable. + +For clarity, our attack process is shown in Algorithm 1. + +# Algorithm 1 The attacking process of our GGS. + +Input: A clean image $x^{C\times H\times W}$ with ground-truth label $y$ , and the loss function $\mathcal{L}(\cdot)$ ; the magnitude of perturbation $\epsilon$ ; the sampling radius $\zeta$ ; the number of outer iterations, $T$ ; the decay factor $\gamma$ ; the number of inner iterations, $N$ ; $\tilde{g}_0,\tilde{p}\in \mathbb{R}^{C\times H\times W}$ + +Output: An adversarial example $x^{adv}$ + +1: Randomly sample $\tilde{g}_0\sim \mathrm{Uniform}(-\zeta ,\zeta)$ +2: $v_{0} = 0$ , $x_{0}^{adv} = x$ , $\alpha = \epsilon / T$ ; +3: while $t \gets 1$ to $T$ (outer-iteration) do + +4: while $i\gets 1$ to $N$ (inner-iteration) do +5: Randomly sample noises $\tilde{p} \sim \mathrm{Uniform}(-\zeta, \zeta)$ ; +6: Get the lookahead example $\hat{x}_i$ by Eq. (5); +7: Get the gradient $\tilde{g}_i = \nabla_x\mathcal{L}(\hat{x}_i,y)$ +8: end while +9: Update $v_{t} = \gamma \cdot v_{t - 1} + \frac{\sum_{i = 1}^{N}\tilde{g}_{i}}{\left\|\sum_{i = 1}^{N}\tilde{g}_{i}\right\|_{1}}$ +0: Update $x_{t}^{adv} = \Pi_{\mathcal{B}_{\epsilon}(x)}\left[x_{t - 1}^{adv} + \alpha \cdot \mathrm{sign}(v_{t})\right]$ ; +1: end while +12: $x^{adv} = x_T^{adv}$ + +Table 1. The average untargeted and targeted ASR (%) on the nine models (Res50 [10], Dense121 [14], Inc-v3 [26], IncRes-v2 [27], ViT-B [3], PiT-B [11], and Inc-v3ens3, Inc-v3ens4, IncRes-v2ens [29]), with adversarial examples generated on a single model (Res50, Inc-v3 or ViT-B). Each data pair $(u/w)$ corresponds to the performances under (untargeted/targeted) attacks. The best and second best results are labeled in bold and underline, respectively. + +
AttackRes50Dense121Inc-v3IncRes-v2ViT-BPiT-BInc-v3ens3Inc-v3ens4IncRes-v2ensAvg.
Res50MI CVPR'1899.8 / 98.154.9 / 0.244.2 / 0.028.0 / 0.011.8 / 0.022.4 / 0.022.8 / 0.025.3 / 0.019.6 / 0.036.53 / 10.92
N ICLR'20100.0 / 95.163.1 / 0.247.2 / 0.032.3 / 0.012.9 / 0.023.7 / 0.023.0 / 0.027.5 / 0.019.0 / 0.038.74 / 10.59
VMI CVPR'2199.8 / 96.671.2 / 2.858.1 / 0.549.4 / 0.231.3 / 0.147.7 / 0.543.9 / 0.045.4 / 0.137.7 / 0.153.83 / 11.21
RAP NeurIPS'2299.6 / 26.486.8 / 0.568.4 / 0.052.6 / 0.022.2 / 0.040.5 / 0.136.2 / 0.036.7 / 0.025.5 / 0.052.06 / 3.00
GRA ICCV'2396.9 / 67.788.6 / 3.981.8 / 0.975.8 / 1.745.3 / 0.262.6 / 0.671.5 / 0.770.9 / 1.167.3 / 0.773.41 / 8.61
PGN NeurIPS'2398.6 / 49.591.3 / 4.785.0 / 1.478.5 / 1.749.7 / 0.467.8 / 1.074.9 / 0.672.9 / 1.170.1 / 1.276.53 / 6.84
ANDA CVPR'2499.9 / 97.687.5 / 5.473.6 / 0.466.3 / 0.842.3 / 0.161.8 / 1.151.0 / 0.152.9 / 0.146.7 / 0.164.67 / 11.74
GI ESWA'24100.0 / 98.568.7 / 0.455.6 / 0.137.6 / 0.115.3 / 0.029.0 / 0.028.8 / 0.030.4 / 0.122.4 / 0.043.09 / 11.02
GGS99.3 / 90.795.9 / 28.689.6 / 7.286.4 / 8.660.2 / 3.481.0 / 8.577.5 / 3.777.4 / 4.371.4 / 4.082.08 / 17.67
Inc-v3MI32.1 / 0.049.6 / 0.1100.0 / 95.356.1 / 0.113.0 / 0.017.3 / 0.034.3 / 0.038.0 / 0.026.3 / 0.040.74 / 10.61
NI41.4 / 0.060.5 / 0.1100.0 / 75.966.9 / 0.113.4 / 0.019.2 / 0.036.8 / 0.040.6 / 0.129.5 / 0.045.37 / 8.47
VMI44.9 / 0.165.6 / 0.199.9 / 76.672.7 / 0.418.3 / 0.025.3 / 0.048.9 / 0.152.5 / 0.040.9 / 0.052.11 / 8.59
RAP58.0 / 0.078.4 / 0.2100.0 / 16.482.1 / 0.115.4 / 0.023.8 / 0.044.1 / 0.147.0 / 0.033.9 / 0.053.63 / 1.87
GRA53.3 / 0.175.4 / 0.499.8 / 67.585.4 / 1.821.4 / 0.030.5 / 0.164.9 / 0.165.9 / 0.356.1 / 0.161.41 / 7.82
PGN56.0 / 0.079.4 / 0.2100.0 / 50.387.9 / 0.923.2 / 0.033.4 / 0.065.9 / 0.067.0 / 0.155.6 / 0.063.16 / 5.72
ANDA51.3 / 0.376.3 / 0.4100.0 / 87.482.5 / 0.620.9 / 0.031.8 / 0.158.2 / 0.261.2 / 0.249.3 / 0.059.06 / 9.91
GI39.9 / 0.059.3 / 0.3100.0 / 97.162.6 / 0.514.6 / 0.019.9 / 0.039.2 / 0.040.9 / 0.031.9 / 0.045.37 / 10.88
GGS69.1 / 1.186.9 / 1.9100.0 / 76.095.8 / 6.527.9 / 0.039.7 / 0.170.0 / 0.773.9 / 1.460.1 / 0.569.27 / 9.80
ViT-BMI41.7 / 0.055.6 / 0.058.1 / 0.039.4 / 0.0100.0 / 99.950.1 / 0.034.2 / 0.036.3 / 0.029.9 / 0.049.48 / 11.10
NI42.4 / 0.058.3 / 0.059.5 / 0.044.7 / 0.1100.0 / 99.050.0 / 0.236.2 / 0.037.6 / 0.031.0 / 0.051.08 / 11.03
VMI54.8 / 0.665.9 / 0.463.1 / 0.150.9 / 0.3100.0 / 99.067.6 / 1.545.3 / 0.144.7 / 0.038.7 / 0.259.00 / 11.36
RAP62.1 / 0.179.6 / 0.273.7 / 0.062.3 / 0.099.9 / 57.066.2 / 0.148.5 / 0.051.2 / 0.042.6 / 0.065.12 / 6.38
GRA64.9 / 1.077.6 / 1.174.7 / 0.868.8 / 0.899.3 / 91.780.9 / 4.761.5 / 0.264.0 / 0.157.7 / 0.272.16 / 11.18
PGN69.3 / 0.481.1 / 0.778.6 / 0.470.2 / 0.599.1 / 82.284.8 / 4.465.6 / 0.167.5 / 0.361.4 / 0.275.29 / 9.91
ANDA66.2 / 0.379.0 / 0.575.5 / 0.165.7 / 0.3100.0 / 99.979.0 / 1.455.3 / 0.156.7 / 0.149.9 / 0.269.70 / 11.44
GI55.8 / 0.169.6 / 0.267.1 / 0.052.5 / 0.1100.0 / 100.061.0 / 0.345.4 / 0.046.3 / 0.038.0 / 0.059.52 / 11.19
GGS80.8 / 6.089.9 / 6.587.4 / 4.682.3 / 4.6100.0 / 99.292.7 / 27.374.0 / 2.174.5 / 2.368.4 / 2.483.33 / 17.22
+ +# 4. Experiments + +# 4.1. Experimental Settings + +Dataset: Our experiments are conducted on an ImageNet-compatible dataset that is widely used in adversarial attack research. The dataset includes 1,000 images, each with a size of $299 \times 299 \times 3$ . + +Models: We select six normally pre-trained models, i.e., ResNet50 (Res50) [10], DenseNet121 (Dense121) [14], Inception-v3 (Inc-v3) [26], InceptionResNet-v2 (IncRes-v2) [27], VisionTransformer-Base (ViT-B) [3], and Pooling-based VisionTransformer-Base (PiT-B) [11], along with three models pre-trained on adversarial examples Inception-v3_ens3 (Inc-v3_ens3), Inception-v3_ens4 (Inc-v3_ens4) and InceptionResNet-v2_ens (IncRes-v2_ens) [29], to evaluate the performance of different methods. + +Baselines: To provide a comprehensive evaluation of our method, we choose eight popular gradient-based adversarial attack methods as baselines: MI-FGSM [1], NIFGSM [19], VMI-FGSM [32], RAP [24], GRA [40], PGN [7], ANDA [5] and GI [30]. Additionally, we integrate our method with five input transformation-based methods, i.e., DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21] + +to verify the generalizability of our approach. + +Hyperparameters: For our GGS, we set the maximum perturbation $\epsilon = 16 / 255$ , the number of outer-iterations $T = 10$ , the step size $\alpha = \epsilon / T$ , the number of inner-iterations $N = 20$ , and the upper bound factor of sample range $\zeta = 2.0 \times \epsilon$ . More hyperparameter explanation for other methods is provided in the Supplementary Material. + +# 4.2. Attack on Different Models + +# 4.2.1. Classification model + +As shown in Table 1, our method consistently outperforms state-of-the-art (SOTA) approaches. Compared to the best-performing competitor, it improves the average attack success rate (ASR) by over $5\%$ in both untargeted and targeted attacks across most architectures, except for Inception-v3. Furthermore, our GGS significantly enhances attack performance on the surrogate model, achieving $4\% -25\%$ higher ASR in targeted attacks than other sampling methods like PGN [7] and GRA [40]. This demonstrates GGS's ability to boost attack potency while maintaining strong cross-model generalization. Additional details on ensemble setting can be found in the supplementary materials. + +![](images/be989cca6b14e70b2dce74843ac61ef3c16bb5de095c8a60ed785a800b77d7a2.jpg) +Figure 4. The loss surfaces of adversarial examples generated by different methods (MI-FGSM [1], GRA [40], PGN [7], ANDA [5]), based on three different model architectures (ResNet50 [10], Inception-v3 [26], and ViT-B [3]) along random directions with varying strengths like [7]. In the images marked with a star in the upper left corner, adversarial examples are generated and tested on the same model, indicating white-box testing. In contrast, unmarked images represent adversarial examples generated on one model and tested on another, indicating black-box testing. We have also highlighted the regions covered by our loss surface using a red background for easier visualization. + +# 4.2.2. Multimodal Large Language Models + +The widespread adoption of Multimodal Large Language Models (MLLMs) has raised significant security concerns. We evaluated our method on five prominent MLLMs using adversarial examples generated under an ensemble setting, as shown in Table 2. Our GGS method achieved superior attack performance, reducing the average CSR by over $9\%$ compared to the strongest baseline, which demonstrates its excellent cross-model applicability. + +Table 2. Classification success rate (CSR, %) of adversarial examples generated by different methods in Multimodal Large Language Models (MLLMs) from three major providers (OpenAI, Google, Anthropic) under ensemble settings (Res50 [10], Inc-v3 [26], ViT-B [3]). Row "clean" indicates the CSR on the clean samples. Avg. represents the average CSR across the five MLLMs. + +
MethodGPTGeminiClaude
4ominiproflashsonnetAvg. ↓
Clean77.184.085.680.968.379.18
MI_CVPR'18 [1]77.077.381.876.561.074.72
NI_ICLR'20 [19]75.073.983.176.862.474.24
VMI_CVPR'21 [32]68.866.678.473.263.170.02
RAP_NeurIPS'22 [24]64.362.075.166.160.265.54
GRA_ICCV'23 [40]57.152.270.965.646.758.50
PGN_NeurIPS'23 [7]56.249.069.664.145.156.80
ANDA_CVPR'24 [5]59.956.874.969.350.562.28
GI_ESWA'24 [30]71.572.879.473.263.872.14
GGS43.137.861.155.740.047.54
+ +# 4.2.3. Loss Surface Visualization against Noise Levels + +To shed light on the capacity of our GGS in locating flat maxima regions for adversarial examples, we compared the loss surfaces of adversarial examples generated by different attack methods in Fig. 4. Each curve represents the average loss of 32 randomly selected adversarial examples, while the center of each image (noise level $= 0$ ) indicates the average loss of clean adversarial examples. In white-box settings (marked with a star), GGS consistently finds flatter maxima compared to baselines. Consequently, in black-box testing, GGS achieves the highest loss values (images without star), indicating its superior transfer attack capability. Additionally, the GGS loss surface (highlighted in red) almost entirely encompasses those of other methods, confirming its effectiveness in balancing strong attack potency with cross-model generalization. + +# 4.3. Compatibility with Other Methods + +Attack with Inner-Iteration Sampling Methods: Our Gradient-Guided Sampling (GGS) is compatible with existing inner-iteration RS-based approaches. When integrated with SOTA approaches like GRA [40] and PGN [7], GGS largely improves attack success rates (ASR). As shown in Table 3, untargeted ASR improved by $5.2\%$ (GRA) and $6.7\%$ (PGN), while targeted ASR increased by $5\%$ (GRA) and $6\%$ (PGN), demonstrating GGS's compatibility. + +Attack with Input Transformations Methods: To study the compatibility of our GGS with existing input + +![](images/8cb8d7a553ac8ca343bd28868d690ad6b0905d3d7be442b34ceaaa8dbc101b08.jpg) +(a) Random (RS) + +![](images/eb30b3b6476906dc0618bc9aeb22f4f3ba923781b64172d9eec5cfcd879634dc.jpg) +(b) Momentum-Guided (MGS) +Figure 5. Loss surfaces of adversarial examples generated by different sampling strategies, Random Sampling (RS), MGS and our GGS, with increasing inner-iteration on Resnet50. The red line on the left side of part (a)~(c) represents the maximum value of the loss surface in each iteration, and the blue dashed line highlights these values in early-stage inner iteration. For (d), it represents the cosine similarity between the gradient $\tilde{g}_i$ generated in each inner-iteration and the average of gradients $\sum_{i=1}^{N} \tilde{g}_i / N$ . + +![](images/9232b6d667d31188f1c864ee7a648e4c116d986e9440d9b7af55415a25b0f791.jpg) +(c) Gradient-Guided (GGS) + +![](images/4788acac99032229f83e43fa9ddedeb10c116d6066817e4027a343a1098a4465.jpg) +(d) The cosine similarity between the gradient and the gradients averaging in inner-iterations + +Table 3. The average untargeted and targeted ASR $(\%)$ on all nine models with the adversarial examples generated on ResNet50, when combined our GGS with input transformation-based methods (DIM [37], TIM [2], SIM [19], Admix [33], SSM [21]) and inner-iteration RS-based methods (GRA [40], PGN [7]). + +
ClassMethodUntargeted Avg.Targeted Avg.
Random Sampling (RS)GRA / +ours73.41 / 78.698.61 / 13.64
PGN / +ours76.53 / 83.236.84 / 12.79
Input TransformationDIM / +ours51.30 / 90.138.63 / 19.37
TIM / +ours40.82 / 84.9411.03 / 18.48
SIM / +ours46.80 / 90.1211.14 / 28.44
Admix / +ours54.78 / 85.5011.28 / 30.50
SSM / +ours74.22 / 80.748.40 / 12.54
+ +transformation techniques, we integrated it with five methods of this type: DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]. Adversarial examples generated on ResNet50 were tested on nine diverse models. Table 3 shows GGS significantly improved adversarial transferability, increasing untargeted ASR by $6\% -43\%$ and targeted ASR by $4\% -19\%$ . For instance, SIM's untargeted ASR improved by $43\%$ when combined with GGS. More details are in the supplementary material. + +# 4.4. Ablation Study of Proposed Components + +Table 4. The average untargeted ASR on ResNet50 and eight target models for Random Sampling (RS), Momentum-Guided Sampling (MGS), and our Gradient-Guided Sampling (GGS). + +
Sampling Guided byRes50Others
randmomentumgradient
(a) RS✓--97.363.74
(b) MGS-✓-97.558.79
(c) GGS--✓99.379.93
+ +Ablation study compares RS, MGS, and GGS via loss surface analysis (Fig. 5) and attack success rates (Table 4). + +As shown in Fig. 5(a), Random Sampling (RS) yields a relatively flat loss surface due to its exploration capability but it suffers from significant oscillation in loss values. Introducing Momentum-Guided Sampling (MGS) (Fig. 5(b)) stabilizes the loss surface over iterations, but its long-chain dependency causes high similarity between later gradients and the final average gradient (Fig. 5(d)), severely limiting exploration. Our GGS resolves MGS's long-chain dependency and mitigates early-stage noisy gradients (Fig. 5(b)(c), blue dashed lines). The low gradient similarity in Fig. 5(d) reflects the enhanced exploration, while the flat yet higher-loss surface in Fig. 5(c) demonstrates the exceptional sampling efficiency of our GGS. + +Quantitatively, as shown in Table 4, MGS marginally improves ASR of the surrogate model by $0.2\%$ , but reducing transfer attack capability by $5\%$ . In contrast, GGS increases both white-box and transfer ASR by $2\%$ and $16\%$ , respectively, validating its superior sampling efficiency. + +# 5. Conclusion and Perspectives + +This paper introduces Gradient-Guided Sampling (GGS), an easy-to-implement inner-iteration sampling strategy, for transferable adversarial attacks that effectively resolves the dilemma between Exploitation (attack potency) and Exploration (cross-model generalization). By guiding the sampling direction with the gradient from the previous inner iteration, GGS enables stable ascent toward flat loss regions with higher local maxima. Our GGS has the merit of enhancing exploitation capability without compromising exploration capability, effectively balancing these two aspects and addressing the inherent limitations of random sampling (RS) methods, as verified by extensive experiments and visualizations as well as the comparison with related state-of-the-art methods. GGS is also revealed to be highly compatible with existing methods. + +While our GGS is highly compatible with gradient-averaging methods (e.g., GRA [40], PGN [7]), we aim to refine it to also support non-gradient-averaging techniques (e.g., VMI-FGSM [32], RAP [24]). + +# Acknowledgements + +The work was supported by the National Natural Science Foundation of China under grants no. 62276170, 82261138629, 62306061, the Science and Technology Project of Guangdong Province under grants no. 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grant no. JCYJ20220531101412030, Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) under Grant No. GML-KF-24-11, and Guangdong Provincial Key Laboratory under grant no. 2023B1212060076. + +# References + +[1] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 1, 2, 3, 6, 7 +[2] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4312-4321, 2019. 2, 3, 6, 8 +[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6, 7 +[4] Yan Fang, Zhongyuan Wang, Jikang Cheng, Ruoxi Wang, and Chao Liang. Promoting adversarial transferability with enhanced loss flatness. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1217-1222. IEEE, 2023. 2 +[5] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 3, 6, 7 +[6] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, ICLR, 2021. 2, 3 +[7] Zhijin Ge, Hongying Liu, Wang Xiaosen, Fanhua Shang, and Yuanyuan Liu. Boosting adversarial transferability by achieving flat local maxima. Advances in Neural Information Processing Systems, 36:70141-70161, 2023. 1, 2, 3, 4, 6, 7, 8 +[8] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 3 +[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. + +Explaining and harnessing adversarial examples. In International Conference on Learning Representations, ICLR, 2015. 3 +[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7 +[11] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11936-11945, 2021. 6 +[12] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997. 3 +[13] SM Mostaq Hossain, Shampa Banik, Trapa Banik, and Ashfak Md Shibli. Survey on security attacks in connected and autonomous vehicular systems. In 2023 IEEE International Conference on Computing (ICOCO), pages 295-300. IEEE, 2023. 2 +[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6 +[15] Martin Husák, Jana Komárková, Elias Bou-Harb, and Pavel Čeleda. Survey of attack projection, prediction, and forecasting in cyber security. IEEE Communications Surveys & Tutorials, 21(1):640-660, 2018. 2 +[16] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. In Advances in Neural Information Processing Systems, pages 73919-73942. Curran Associates, Inc., 2023. 2 +[17] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. ICLR, 2017. 3 +[18] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 3 +[19] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, ICLR, 2020. 3, 4, 5, 6, 7, 8 +[20] Decheng Liu, Xijun Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Adv-diffusion: Imperceptible adversarial face identity attack via latent diffusion model. Proceedings of the AAAI Conference on Artificial Intelligence, 38(4):3585-3593, 2024. 2 +[21] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European conference on computer vision, pages 549-566. Springer, 2022. 3, 6, 8 +[22] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, ICLR, 2018. 3 + +[23] Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Dokl. Akad. Nauk. SSSR, page 543, 1983. 2, 4, 5 +[24] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in neural information processing systems, 35:29845-29858, 2022. 2, 3, 4, 6, 7, 8 +[25] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. 4 +[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6, 7 +[27] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-reset and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6 +[28] Chee-Wooi Ten, Govindarasu Manimaran, and Chen-Ching Liu. Cybersecurity for critical infrastructures: Attack and defense modeling. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 40(4):853–865, 2010. 2 +[29] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. *ICLR*, 2018. 6 +[30] Jiafeng Wang, Zhaoyu Chen, Kaixun Jiang, Dingkang Yang, Lingyi Hong, Pinxue Guo, Haijing Guo, and Wenqiang Zhang. Boosting the transferability of adversarial attacks with global momentum initialization. Expert Systems with Applications, 255:124757, 2024. 3, 6, 7 +[31] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3 +[32] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1924-1933, 2021. 2, 3, 4, 6, 7, 8 +[33] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 2, 3, 6, 8 +[34] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 3 +[35] Han Wu, Syed Yunas, Sareh Rowlands, Wenjie Ruan, and Johan Wahlström. Adversarial driving: Attacking end-to-end autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7. IEEE, 2023. 2 +[36] Chaowei Xiao, Bo Li, Jun yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty- + +Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 3905-3911, 2018. 2 +[37] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 2, 3, 6, 8 +[38] Yinhu Xu, Qi Chu, Haojie Yuan, Zixiang Luo, Bin Liu, and Nenghai Yu. Enhancing adversarial transferability from the perspective of input loss landscape. In International Conference on Image and Graphics, pages 254-266. Springer, 2023. 2 +[39] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 3 +[40] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4741-4750, 2023. 2, 3, 4, 6, 7, 8 +[41] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 3 \ No newline at end of file diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/images.zip b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..743afbdffd7f2a8ba1e2a45f304523616a4fd48d --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5bbe42774480e4cfe07316419cd45c2512bb397bab98313e7f2212f62444708d +size 760018 diff --git a/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/layout.json b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5770bfc7bed1609e00f06ec33743f8cb04e07697 --- /dev/null +++ b/2025/Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling/layout.json @@ -0,0 +1,8801 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 520, + 140 + ], + "type": "text", + "content": "Enhancing Adversarial Transferability by Balancing Exploration and Exploitation with Gradient-Guided Sampling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "spans": [ + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": "Zenghao Niu" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": " Weicheng Xie" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": "* Siyang Song" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": " Zitong Yu" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": " Feng Liu" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "text", + "content": " Linlin Shen" + }, + { + "bbox": [ + 85, + 160, + 547, + 175 + ], + "type": "inline_equation", + "content": "^{6}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 117, + 175, + 521, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 175, + 521, + 189 + ], + "spans": [ + { + "bbox": [ + 117, + 175, + 521, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 117, + 175, + 521, + 189 + ], + "type": "text", + "content": "School of Computer Science & Software Engineering, Shenzhen University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 91, + 189, + 545, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 189, + 545, + 203 + ], + "spans": [ + { + "bbox": [ + 91, + 189, + 545, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 91, + 189, + 545, + 203 + ], + "type": "text", + "content": "Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ), Shenzhen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 203, + 575, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 203, + 575, + 217 + ], + "spans": [ + { + "bbox": [ + 63, + 203, + 575, + 217 + ], + "type": "text", + "content": "3Guangdong Provincial Key Laboratory of Intelligent Information Processing, Shenzhen University, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 181, + 217, + 456, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 217, + 456, + 231 + ], + "spans": [ + { + "bbox": [ + 181, + 217, + 456, + 231 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 181, + 217, + 456, + 231 + ], + "type": "text", + "content": "School of Computer Science, University of Exeter, U.K." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 110, + 231, + 525, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 231, + 525, + 245 + ], + "spans": [ + { + "bbox": [ + 110, + 231, + 525, + 245 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 110, + 231, + 525, + 245 + ], + "type": "text", + "content": "Department of Computing and Information Technology, Great Bay University, China" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 105, + 245, + 533, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 245, + 533, + 258 + ], + "spans": [ + { + "bbox": [ + 105, + 245, + 533, + 258 + ], + "type": "text", + "content": "6Computer Vision Institute, School of Artificial Intelligence, Shenzhen University, China" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 165, + 261, + 467, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 261, + 467, + 273 + ], + "spans": [ + { + "bbox": [ + 165, + 261, + 467, + 273 + ], + "type": "text", + "content": "{2300271072@mail., wcxie@, feng.liu@, llshen@}szu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 209, + 276, + 420, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 276, + 420, + 286 + ], + "spans": [ + { + "bbox": [ + 209, + 276, + 420, + 286 + ], + "type": "text", + "content": "s song@exeter.ac.uk, zitong.yu@ieee.org" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "spans": [ + { + "bbox": [ + 151, + 315, + 200, + 327 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 333, + 296, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 333, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 296, + 681 + ], + "type": "text", + "content": "Adversarial attacks present a critical challenge to deep neural networks' robustness, particularly in transfer scenarios across different model architectures. However, the transferability of adversarial attacks faces a fundamental dilemma between Exploitation (maximizing attack potency) and Exploration (enhancing cross-model generalization). Traditional momentum-based methods over-prioritize Exploitation, i.e., higher loss maxima for attack potency but weakened generalization (narrow loss surface). Conversely, recent methods with inner-iteration sampling over-prioritize Exploration, i.e., flatter loss surfaces for cross-model generalization but weakened attack potency (suboptimal local maxima). To resolve this dilemma, we propose a simple yet effective Gradient-Guided Sampling (GGS), which harmonizes both objectives through guiding sampling along the gradient ascent direction to improve both sampling efficiency and stability. Specifically, based on MI-FGSM, GGS introduces inner-iteration random sampling and guides the sampling direction using the gradient from the previous inner-iteration (the sampling's magnitude is determined by a random distribution). This mechanism encourages adversarial examples to reside in balanced regions with both flatness for cross-model generalization and higher local maxima for strong attack potency. Comprehensive experiments across multiple DNN architectures and multimodal large language models (MLLMs) demonstrate the superiority of our method over state-of-the-art transfer attacks. Code is made available at https://github.com/anuin-cat/GGS." + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 313, + 557, + 514 + ], + "blocks": [ + { + "bbox": [ + 316, + 313, + 557, + 514 + ], + "lines": [ + { + "bbox": [ + 316, + 313, + 557, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 557, + 514 + ], + "type": "image", + "image_path": "5ce4302b0c4f1e496a52dafe22ed071a9152d827637a64291246dde7d2d47fb4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 520, + 555, + 643 + ], + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 643 + ], + "type": "text", + "content": "Figure 1. The loss surfaces of (a) MI-FGSM [1] (Momentum iterative fast gradient sign method), (b) RS (Base inner-iteration Random Sampling defined in section 3.2) for enhancing exploration within the neighborhood, (c) PGN [7] (Penalizing Gradient Norm) with Random Sampling (RS) for stable gradient estimation to enhance exploration, and (d) Our GGS (Gradient-Guided Sampling) for efficient sampling to generate gradients that stably towards flat regions with higher local maxima. It shows that, building upon RS, our approach not only maintains a flat loss surface, but also improves the local maximum loss value, compared to PGN, achieving a balance between exploration and exploitation." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Researches have indicated that adding imperceptible perturbations to input data can easily affect deep neural networks" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3885" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "(DNNs), leading to erroneous predictions in safety-critical domains such as autonomous driving [13, 35], and cybersecurity [15, 28]. To facilitate the development of relevant defense mechanisms, in-depth research on adversarial attack methods is urgently needed." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 301 + ], + "type": "text", + "content": "While adversarial attacks are categorized into white-box and black-box settings, the black-box scenario is more realistic in practical applications, where the attacker lacks knowledge of the target model and relies solely on a surrogate model, aiming to generate adversarial examples that are effective across multiple target models. To enhance the transferability of adversarial examples in this setting, researchers have explored gradient-based methods [1, 7, 32], input transformation methods [2, 33, 37], and generative methods [16, 20, 36] to mitigate overfitting of adversarial examples on surrogate models. Despite significant progress, the effectiveness of attacks on target models remains substantially lower than that on surrogate models, driving the need for more effective attack strategies." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 303, + 295, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 458 + ], + "type": "text", + "content": "In black-box adversarial attacks, RAP [24] adopted Sharpness-Aware Minimization (SAM [6]) to improve transferability by flattening the loss surface. Building on this, PGN [7] further enhanced the transferability of adversarial examples by penalizing gradient norm and flattening the loss surface. The efficacy of flat loss surface in improving adversarial attack transferability is increasingly supported [4, 7, 38]. To enhance exploration through local flatness, recent approaches, building on MI-FGSM's [1] momentum-based outer-iterations, often incorporates inner-iterations before each example update [7, 40], i.e., performing multi-step neighborhood sampling for stabilization." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 640 + ], + "type": "text", + "content": "However, these methods enhance Exploration (cross-model generalization) at the cost of compromising Exploitation (attack potency). This is because, although inner-iterations can significantly enhance exploration capability through neighborhood sampling, independent random sampling struggles to obtain stable and effective gradient directions, while potential noise may cause the averaged gradient to deviate from the optimal ascent toward target regions, which exhibit flatter loss surface and higher local maxima. To optimize sampling efficiency, a guided sampling constraint is essential, which should satisfy two criteria: (1) ensuring stable gradient ascent direction and (2) directing gradients toward flatter regions, thereby enabling samples to reach flatter local maxima regions with enhanced transferability, balancing exploitation and exploration." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "For (1) stable gradient direction, a sufficiently small neighborhood sampling followed by gradient calculation generally yields correct ascent directions. Yet, how can we further stabilize this direction? The Nesterov Accelerated Gradient (NAG) [23] enhances stability by firstly updating samples along the current momentum (lookahead)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 252 + ], + "type": "text", + "content": "before acquiring gradients, providing more stable directions. Therefore, momentum-based direction constraints are promising for sampling, as the additional random magnitude can be used to preserve the sampling randomness. (2) However, directly applying NAG's momentum mechanism during inner-iteration sampling imposes strong constraints on subsequent directions, restricts the exploration capability for flat regions. To address this, we replace momentum with the gradient from the previous inner-iteration, named Gradient-Guided Sampling (GGS). This adjustment ensures that sampling to align with gradient ascent directions while reducing the impact from early-stage instability. Our GGS can achieve stable sampling directions after brief oscillations, consistently targeting flat loss regions with larger local maxima regardless of the initial sampling quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 253, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 553, + 384 + ], + "type": "text", + "content": "As shown in Fig. 1(d), our GGS generates adversarial examples with flatter loss surface and higher local maxima compared to RS, i.e., it does not sacrifice local maxima in the pursuit of a flatter surface compared to PGN, thereby achieving a balance between exploration and exploitation to facilitate attack transferability. Additionally, since our approach represents a complementary improvement over Random Sampling (RS), it is compatible with RS-based methods, such as inner-iteration random sampling and gradient averaging, including PGN [7] and GRA [40]. In summary, this study makes the following contributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 389, + 553, + 591 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 389, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 389, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 553, + 472 + ], + "type": "text", + "content": "- We propose Gradient-Guided Sampling (GGS), a novel inner-iteration sampling strategy that effectively balances exploration (cross-model generalization) and exploitation (attack potency). By leveraging the gradient from the previous inner-iteration to guide sampling directions, GGS achieves stable gradient ascent toward flat loss regions with higher local maxima." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 473, + 553, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 473, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 473, + 553, + 521 + ], + "type": "text", + "content": "- The GGS framework is compatible with existing inner-iteration random sampling-based methods, enhancing their sampling efficiency and further improving the transferability of the adversarial examples they generate." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 521, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 521, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 314, + 521, + 553, + 591 + ], + "type": "text", + "content": "- Extensive experimental results as well as comprehensive visualizations demonstrate the efficacy of our GGS, covering targeted and non-targeted attacks in cross-architecture black-box scenarios, non-targeted attacks on multimodal large language models (MLLMs) and commercial cloud functions." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 608, + 400, + 620 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 608, + 400, + 620 + ], + "spans": [ + { + "bbox": [ + 314, + 608, + 400, + 620 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 553, + 712 + ], + "type": "text", + "content": "Transferability is a critical property in adversarial attacks, allowing adversarial examples generated from a surrogate model to deceive unseen target models in black-box settings. This characteristic makes transfer-based attacks highly practical in real-world applications where target models are inaccessible, motivating our focus on transferable adversarial attacks." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3886" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 173, + 83 + ], + "type": "text", + "content": "2.1. Transferable Attack" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 268 + ], + "type": "text", + "content": "Gradient-based Attack. Gradient-based attacks aim to generate adversarial examples by exploiting the gradients of a surrogate model to maximize the loss. Classic methods such as Fast Gradient Sign Method (FGSM) [9] and Projected Gradient Descent (PGD) [22] optimize adversarial perturbations to fool the surrogate model. However, these attacks often overfit the surrogate model, resulting in limited transferability to target models. To address this, advanced techniques such as momentum-based (MI-FGSM [1]), Nesterov Iterative-based (NI-FGSM [19]), variance-tuned (VMI-FGSM [32]), gradient relevance-based (GRA [40]), momentum initialization (GI-FGSM [30]) and distribution-based methods (ANDA [5]) have been introduced to smooth out the optimization process and improve transferability across models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 269, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 296, + 400 + ], + "type": "text", + "content": "Input Transformation. To mitigate the overfitting issue inherent in gradient-based methods, input transformation techniques have been proposed to enhance the diversity of adversarial examples. Techniques like translation, resizing, padding, image mixup (e.g., DIM [37], SIM [19], TIM [2], Admix [33]) and other advanced image transformation methods (e.g., SSM [21], SIA [34], STM [8], L2T [41], BSR [31]) introduce randomness to the input data before generating perturbations. By applying such transformations during the attack process, more effective transferable adversarial examples can be generated." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 408, + 138, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 138, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 138, + 420 + ], + "type": "text", + "content": "2.2. Flat Maxima" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 594 + ], + "type": "text", + "content": "The generalization ability of models has been suggested to possibly have a certain association with flat minima [12]. More in-depth research and exploration have gradually confirmed this perspective [6, 17, 18, 39]. Moreover, flat maxima also have been validated to be effective for enhancing the generalization and transferability of adversarial examples [24]. Flat maxima refers to regions in the loss landscape, where small changes in the model parameters result in minimal changes in the loss, making adversarial examples generated in these regions less sensitive to the specific decision boundaries of models. This property contrasts with \"sharp maxima\", which are highly sensitive to small perturbations and can lead to overfitting to the surrogate model, reducing transferability capacity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "In the context of transfer-based attacks, works like Reverse Adversarial Perturbation (RAP) [24] leverage flat maxima to generate adversarial examples, PGN [7] adopts a first-order procedure to approximate the Hessian/vector product, largely improving computational efficiency. They are not only effective on the surrogate model but also more resilient when attacking unseen target models. These methods mitigate overfitting, via making adversarial examples lie in regions where the loss function remains stable across different models (e.g., flat maxima regions), improving at" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "tack success rates in black-box settings. However, achieving flat maxima requires not only attention to \"flatness\" but also preservation of the \"maxima\" to ensure sufficient attack strength during transfer attack, which poses a significant challenge." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 140, + 455, + 153 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 140, + 455, + 153 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 455, + 153 + ], + "type": "text", + "content": "2.3. Inner-Iteration Sampling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 158, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 158, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 158, + 555, + 289 + ], + "type": "text", + "content": "To enhance the generalization capability of adversarial examples, recent methods have introduced inner-iteration sampling [7, 24, 32]. Specifically, an inner iterative process is inserted before updating the examples. This process involves neighborhood sampling to acquire domain information during inner-iterations which are used to adjust gradients [32] or neighborhood search is incorporated to locate local minima and enhance their loss values [24]. Furthermore, works like [7, 40] enhance the flatness of the loss landscape by averaging gradients from inner-iterations to update the examples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 398 + ], + "type": "text", + "content": "Additionally, neighborhood sampling can facilitate examples to escape from sharp local maxima regions, thereby avoiding overfitting to the surrogate model. However, directly averaging the gradients from inner-iterations will reduce the maximum value of the loss surface, since random sampling produces unstable gradient directions that fail to consistently direct to flat regions with higher local maxima. Thus, enhancing inner-iteration sampling efficiency to stabilize the final gradient has become a critical challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 409, + 397, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 397, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 397, + 422 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 400, + 441 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": "Transferable Adversarial Attacks. For a given example " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ", a surrogate model " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " and multiple target models " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "f_{\\varphi_k}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "k \\in \\{1,2,\\dots,K\\}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ", the attacker's goal is to find " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "x^{adv} = x + \\delta" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " to make " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "f_{\\varphi_k}(x^{adv}) \\neq y" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\| \\delta \\|_p < \\epsilon" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " denotes the maximum magnitude of the specified perturbation and " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_p" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": "-norm. For targeted attacks, we simply adapt the symbols and labels in the target function: " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "f_{\\varphi_k}(x^{adv}) = y_t" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "y_t" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " denotes the target label. Meanwhile, to facilitate a clearer representation of the norm constraints against " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "x^{adv}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " as below, we define " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_{\\epsilon}(x) = \\{x': \\| x' - x \\|_p \\leq \\epsilon\\}" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": " to denote the " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": "-ball of an input image " + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 447, + 555, + 615 + ], + "type": "text", + "content": ". For the purpose of the above attack, we generally need to maximize the following objective function to generate the transferable adversarial examples:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 385, + 623, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 385, + 623, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 385, + 623, + 555, + 644 + ], + "type": "interline_equation", + "content": "\\max _ {x ^ {a d v} \\in \\mathcal {B} _ {\\epsilon} (x)} \\mathcal {L} \\left(x ^ {a d v}, y\\right), \\tag {1}", + "image_path": "535e8e640e00bf418be84510277dc7ca8decb3f85a9a26764feb02ac43209bc8.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " represents the loss function. In the above flow, we assume that " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " is known as a white-box model and " + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "inline_equation", + "content": "f_{\\varphi_k}" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": " is unknown as a black-box model. Similar to the process of training a neural network, the white-box model can be seen as the training data, with the adversarial examples serving" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3887" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "as the optimization parameters, while the black-box model acts as the test data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 216 + ], + "type": "text", + "content": "Nesterov Accelerated Gradient (NAG). Due to the similarity between model training task and adversarial example generation, improving performance in black-box settings can be viewed as enhancing model generalization in model training tasks. Consequently, some commonly used optimization methods, such as Momentum [25] and Nesterov Accelerated Gradient (NAG) [23], can be introduced to improve the generalization of adversarial examples across black-box models through a more stable optimization process." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "type": "text", + "content": "NI-FGSM [19] was the first to incorporate the principles of NAG into adversarial attack tasks. This enables the algorithm to acquire directional information in advance, reducing oscillations and achieving stable convergence. Specifically, momentum is computed at each step as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 106, + 285, + 295, + 320 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 285, + 295, + 320 + ], + "spans": [ + { + "bbox": [ + 106, + 285, + 295, + 320 + ], + "type": "interline_equation", + "content": "v _ {t} = \\gamma \\cdot v _ {t - 1} + \\frac {\\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y)}{\\left\\| \\nabla_ {x} \\mathcal {L} (\\tilde {x} _ {t} , y) \\right\\| _ {1}}, \\tag {2}", + "image_path": "bdbf84f061eaf9bde98e1cb3cd3fda0e704d38f183d0da05ca07d2fa3ca52448.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 106, + 316, + 211, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 316, + 211, + 346 + ], + "spans": [ + { + "bbox": [ + 106, + 316, + 211, + 346 + ], + "type": "interline_equation", + "content": "\\tilde {x} _ {t} = \\underbrace {x _ {t - 1} ^ {a d v} + \\alpha \\cdot \\gamma \\cdot v _ {t - 1}} _ {\\text {l o o k a h e a d p o i n t}},", + "image_path": "b12656c649703ba4ae495c0f4b1591157a368afc900b5551eab47b0bc237e768.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "v_{t}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " represents the momentum term, " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "x_0^{adv} = x" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " stands for the step size, " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " denotes the momentum decay factor, and " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_1" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\ell_1" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": "-norm of the given variable. Unlike traditional momentum-based methods, NI-FGSM utilizes the estimated position of examples rather than their current actual position, during gradient computation, i.e., introducing a lookahead mechanism into the algorithm." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 129, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 129, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 129, + 460 + ], + "type": "text", + "content": "3.2. Motivation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 296, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 575 + ], + "type": "text", + "content": "Recent transfer-based adversarial attack methods have introduced inner-iteration random sampling, to enhance algorithm stability [7, 40] or acquire neighborhood information [24, 32]. It can be noted that these methods have increased the flatness of the loss surface, as illustrated in Fig.1(c), with additional loss surface plots for other methods provided in the supplementary materials. If we isolate the inner-iteration random sampling (RS) component, it can be expressed in the following form:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 90, + 585, + 295, + 624 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 585, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 90, + 585, + 295, + 624 + ], + "type": "interline_equation", + "content": "v _ {t} = \\gamma v _ {t - 1} + \\frac {\\sum_ {i = 1} ^ {N} \\tilde {g} _ {i}}{\\left\\| \\sum_ {i = 1} ^ {N} \\tilde {g} _ {i} \\right\\| _ {1}}, \\tag {3}", + "image_path": "e6fce36b25f586de364752170164dd669fceeb784f7aeaddd1eb97d344f3de64.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 91, + 623, + 260, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 623, + 260, + 653 + ], + "spans": [ + { + "bbox": [ + 91, + 623, + 260, + 653 + ], + "type": "interline_equation", + "content": "\\tilde {g} _ {i} = \\nabla_ {x} \\mathcal {L} \\big (\\tilde {x} _ {i}, y \\big), \\text {w i t h} \\tilde {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + \\tilde {p}} _ {\\text {s a m p l i n g p o i n t}},", + "image_path": "df168d365b4a578582c8907efa7b4fed1ac23be40778771cc7baffeddb7b2426.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N}" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": " denotes the output of the inner-iteration process; " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": " denotes the number of inner-iterations; " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": " denotes the momentum decay factor; " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "\\tilde{g}_0, \\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": ", are uniform random noises with the same size as " + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 663, + 296, + 712 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 323, + 71, + 547, + 186 + ], + "blocks": [ + { + "bbox": [ + 323, + 71, + 547, + 186 + ], + "lines": [ + { + "bbox": [ + 323, + 71, + 547, + 186 + ], + "spans": [ + { + "bbox": [ + 323, + 71, + 547, + 186 + ], + "type": "image", + "image_path": "69edcd1e097eb143568264f1d7471dc58e31a8c36fa14f47250086412fa7dfe5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 187, + 525, + 198 + ], + "lines": [ + { + "bbox": [ + 343, + 187, + 525, + 198 + ], + "spans": [ + { + "bbox": [ + 343, + 187, + 525, + 198 + ], + "type": "text", + "content": "(a) Optimization Trajectories of Various Methods" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 322, + 200, + 430, + 305 + ], + "blocks": [ + { + "bbox": [ + 322, + 200, + 430, + 305 + ], + "lines": [ + { + "bbox": [ + 322, + 200, + 430, + 305 + ], + "spans": [ + { + "bbox": [ + 322, + 200, + 430, + 305 + ], + "type": "image", + "image_path": "9db4dea360f9239cc1a12af075c01f27d6c466212e25fa60152bf5ff3f53e839.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 307, + 421, + 318 + ], + "lines": [ + { + "bbox": [ + 331, + 307, + 421, + 318 + ], + "spans": [ + { + "bbox": [ + 331, + 307, + 421, + 318 + ], + "type": "text", + "content": "(b) 1st step process of RS" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 327, + 555, + 514 + ], + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 514 + ], + "type": "text", + "content": "Figure 2. (a) Outer-Iteration processes of MI-FGSM (green), RS (purple) and our GGS (red). (i) MI-FGSM prioritizes rapid ascent directions, thereby enhancing exploitation and facilitating access to sharp local maxima regions. (ii) RS introduces inner-iterations and enhances exploration capability through neighborhood search, thereby enabling access to flat local maxima regions. (iii) GGS incorporates gradient constraints into RS, simultaneously enhancing both exploration and exploitation capabilities, thereby enabling the faster convergence to the centers of flat local maxima regions. (b) RS is performed within a neighborhood of the current example, and uses the average gradient of all sampled points as the final gradient. (c) Building upon RS, GGS enables examples to have stable gradient directions (② ③ ④) toward the centers of flat local maxima regions, following an initial period of brief oscillation (①). It uses the gradient direction from the previous inner-iteration as the guidance, while the randomness is maintained by setting the gradient magnitude with a random distribution." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 437, + 200, + 545, + 305 + ], + "blocks": [ + { + "bbox": [ + 437, + 200, + 545, + 305 + ], + "lines": [ + { + "bbox": [ + 437, + 200, + 545, + 305 + ], + "spans": [ + { + "bbox": [ + 437, + 200, + 545, + 305 + ], + "type": "image", + "image_path": "73bc083d01e3f442ec4f2e455d7d7f8559ec04893f7b9c360b51a491edd85f45.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 443, + 308, + 538, + 318 + ], + "lines": [ + { + "bbox": [ + 443, + 308, + 538, + 318 + ], + "spans": [ + { + "bbox": [ + 443, + 308, + 538, + 318 + ], + "type": "text", + "content": "(c) 1st step process of GGS" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 542, + 555, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 555, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 555, + 686 + ], + "type": "text", + "content": "As shown in Fig. 1, the inner-iteration RS possesses a strong capability to enhance the flatness of the loss surface compared to MI-FGSM. Compared to RS, current methods such as PGN [7] can further improve the flatness of the loss surface, however, they unwillingly reduce the local maxima of the loss surface, limiting its attack potential. This reduction in local maxima can be attributed to the unstable gradients generated by the completely random sampling strategy, as illustrated in Fig. 2(b). Specifically, the averaged gradient roughly points towards flat regions, while it struggles to consistently align with the stable direction to the center of flat maxima regions, thereby reducing updating efficiency." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "To acquire stable gradient ascent direction and improve the efficiency of inner-iteration sampling, we define two es" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3888" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 296, + 137 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 296, + 137 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 296, + 137 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 296, + 137 + ], + "type": "image", + "image_path": "3ec578967648e41ed05fb22a323a14cc8103e97a0e4306a015df76a59cb6252d.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 138, + 217, + 148 + ], + "lines": [ + { + "bbox": [ + 138, + 138, + 217, + 148 + ], + "spans": [ + { + "bbox": [ + 138, + 138, + 217, + 148 + ], + "type": "text", + "content": "(a) Random Sampling" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 150, + 296, + 211 + ], + "blocks": [ + { + "bbox": [ + 58, + 150, + 296, + 211 + ], + "lines": [ + { + "bbox": [ + 58, + 150, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 58, + 150, + 296, + 211 + ], + "type": "image", + "image_path": "07765863c5e00293a3429595b4e2a249f1b2e931802bf9b57d29c27f29d916ca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 121, + 213, + 239, + 223 + ], + "lines": [ + { + "bbox": [ + 121, + 213, + 239, + 223 + ], + "spans": [ + { + "bbox": [ + 121, + 213, + 239, + 223 + ], + "type": "text", + "content": "(b) Momentum-Guided Sampling" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 225, + 296, + 268 + ], + "blocks": [ + { + "bbox": [ + 58, + 225, + 296, + 268 + ], + "lines": [ + { + "bbox": [ + 58, + 225, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 58, + 225, + 296, + 268 + ], + "type": "image", + "image_path": "edaa2140ae66b9d6d89b7006827312421d9d38d09d4611bc1f4c6f6722165bc2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 125, + 270, + 235, + 281 + ], + "lines": [ + { + "bbox": [ + 125, + 270, + 235, + 281 + ], + "spans": [ + { + "bbox": [ + 125, + 270, + 235, + 281 + ], + "type": "text", + "content": "(c) Gradient-Guided Sampling" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 289, + 296, + 388 + ], + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 388 + ], + "type": "text", + "content": "Figure 3. The differences among three inner-iteration sampling guidance methods: (a) Random sampling in Eq. 3, with each inner-iteration sampling being independent; (b) Momentum-guided sampling in Eq. 4, where sampling direction depends on the cumulative average of all previous gradients, creating long-chain dependencies, while using random sampling for maintaining the randomness; (c) Gradient-guided sampling in Eq. 5, where sampling direction relies solely on the gradient direction of the previous iteration, establishing single-step dependencies." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 296, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 473 + ], + "type": "text", + "content": "sential characteristics that a well-balanced gradient direction should possess: (1) alignment with the stable gradient ascent direction for improving Exploitation capacity, and (2) pointing to loss regions with flatter surface for improving Exploration capacity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 486, + 207, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 207, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 207, + 499 + ], + "type": "text", + "content": "3.3. Gradient-Guided Sampling" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "text", + "content": "(1) Stable gradient ascent direction: A sampled gradient in a sufficiently small neighborhood of the current example is probably a gradient ascent direction, while how can we ensure that this gradient ascent direction is stable? The Nesterov Accelerated Gradient (NAG) technique can precisely address this issue. It first allows the example to \"lookahead\" along the current momentum direction, and then acquires the gradient based on this new position, thereby identifying a more stable ascent direction [19, 23]. To leverage the superior stability of NAG based on random sampling, we perform the magnitude sampling along the momentum direction with an uniform distribution, and named this approach as Momentum-Guided Sampling (MGS). Compared to RS in Eq. 3, we only need to replace " + }, + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "inline_equation", + "content": "\\bar{x}_i" + }, + { + "bbox": [ + 55, + 506, + 296, + 673 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 685, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 685, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 113, + 685, + 295, + 706 + ], + "type": "interline_equation", + "content": "\\bar {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(m _ {i - 1}\\right)} \\tag {4}", + "image_path": "cfced295a3b4c42ec5879bd03a289bb1761d959f8a05b11ecdb1a2f375a721be.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 149, + 707, + 223, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 707, + 223, + 715 + ], + "spans": [ + { + "bbox": [ + 149, + 707, + 223, + 715 + ], + "type": "text", + "content": "lookahead sampling point" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "|\\tilde{p}|" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " denotes the element-wise absolute value of " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\tilde{p}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "\\mathrm{sign}(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": " denotes the sign function, which outputs either 1 or -1, and momentum decay is set to 1 for simplicity, so that " + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "inline_equation", + "content": "m_{i} = \\sum_{k=1}^{i} \\tilde{g}_{k}" + }, + { + "bbox": [ + 313, + 72, + 554, + 121 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "text", + "content": "(2) Pointing to regions with better flatness: As shown in Fig. 3(b), if we use momentum for 'lookahead', the early unstable sampling in the inner-iterations will impose excessive constraints on subsequent sampling directions, leading to long-chain dependencies that severely impair RS's ability in exploring flat regions. Thus, for our GGS, we substitute the momentum term in Eq. 4 with the gradient obtained from the previous iteration, where the lookahead mechanism is maintained with the gradient ascent direction. Specifically, we only need to replace " + }, + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "inline_equation", + "content": "\\tilde{x}_i" + }, + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "text", + "content": " in Eq. 3 with " + }, + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 313, + 122, + 555, + 253 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 373, + 259, + 554, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 259, + 554, + 289 + ], + "spans": [ + { + "bbox": [ + 373, + 259, + 554, + 289 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {i} = \\underbrace {x _ {t - 1} ^ {a d v} + | \\tilde {p} | \\cdot \\operatorname {s i g n} \\left(\\tilde {g} _ {i - 1}\\right)} _ {\\text {l o o k a h e a d s a m p l i n g p o i n t}}, \\tag {5}", + "image_path": "93e17a787e7c46c374a015ad369f494f1b96114339959da291ae757be7f2652f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 422 + ], + "type": "text", + "content": "Fig. 3 further sheds light on the differences among three inner-iteration sampling guidance methods. As shown in Fig. 3(c), GGS largely alleviates long-chain dependencies compared with Momentum-Guided Sampling in Fig. 3(b), while ensuring the sampling direction to be align with the gradient ascent direction, compared with random sampling in Fig. 3(a). In addition, as shown in Fig. 2(c), GGS will achieve stable sampling directions (② ③ ④) after a brief period of oscillation (①), when the initial sampling points are unstable." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 425, + 548, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 425, + 548, + 437 + ], + "spans": [ + { + "bbox": [ + 325, + 425, + 548, + 437 + ], + "type": "text", + "content": "For clarity, our attack process is shown in Algorithm 1." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 453, + 510, + 465 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 453, + 510, + 465 + ], + "spans": [ + { + "bbox": [ + 315, + 453, + 510, + 465 + ], + "type": "text", + "content": "Algorithm 1 The attacking process of our GGS." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "spans": [ + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "Input: A clean image " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "x^{C\\times H\\times W}" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": " with ground-truth label " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": ", and the loss function " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(\\cdot)" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; the magnitude of perturbation " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; the sampling radius " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\zeta" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; the number of outer iterations, " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; the decay factor " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; the number of inner iterations, " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 315, + 469, + 554, + 530 + ], + "type": "inline_equation", + "content": "\\tilde{g}_0,\\tilde{p}\\in \\mathbb{R}^{C\\times H\\times W}" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 530, + 476, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 530, + 476, + 542 + ], + "spans": [ + { + "bbox": [ + 315, + 530, + 476, + 542 + ], + "type": "text", + "content": "Output: An adversarial example " + }, + { + "bbox": [ + 315, + 530, + 476, + 542 + ], + "type": "inline_equation", + "content": "x^{adv}" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 543, + 498, + 578 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 321, + 543, + 498, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 543, + 498, + 554 + ], + "spans": [ + { + "bbox": [ + 321, + 543, + 498, + 554 + ], + "type": "text", + "content": "1: Randomly sample " + }, + { + "bbox": [ + 321, + 543, + 498, + 554 + ], + "type": "inline_equation", + "content": "\\tilde{g}_0\\sim \\mathrm{Uniform}(-\\zeta ,\\zeta)" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "text", + "content": "2: " + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "inline_equation", + "content": "v_{0} = 0" + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "inline_equation", + "content": "x_{0}^{adv} = x" + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon / T" + }, + { + "bbox": [ + 321, + 555, + 448, + 567 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "spans": [ + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "text", + "content": "3: while " + }, + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "inline_equation", + "content": "t \\gets 1" + }, + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 321, + 567, + 482, + 578 + ], + "type": "text", + "content": " (outer-iteration) do" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 318, + 543, + 551, + 693 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "spans": [ + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "text", + "content": "4: while " + }, + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "inline_equation", + "content": "i\\gets 1" + }, + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 321, + 579, + 499, + 590 + ], + "type": "text", + "content": " (inner-iteration) do" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 590, + 551, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 590, + 551, + 602 + ], + "spans": [ + { + "bbox": [ + 321, + 590, + 551, + 602 + ], + "type": "text", + "content": "5: Randomly sample noises " + }, + { + "bbox": [ + 321, + 590, + 551, + 602 + ], + "type": "inline_equation", + "content": "\\tilde{p} \\sim \\mathrm{Uniform}(-\\zeta, \\zeta)" + }, + { + "bbox": [ + 321, + 590, + 551, + 602 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 603, + 530, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 603, + 530, + 614 + ], + "spans": [ + { + "bbox": [ + 321, + 603, + 530, + 614 + ], + "type": "text", + "content": "6: Get the lookahead example " + }, + { + "bbox": [ + 321, + 603, + 530, + 614 + ], + "type": "inline_equation", + "content": "\\hat{x}_i" + }, + { + "bbox": [ + 321, + 603, + 530, + 614 + ], + "type": "text", + "content": " by Eq. (5);" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 615, + 499, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 615, + 499, + 627 + ], + "spans": [ + { + "bbox": [ + 321, + 615, + 499, + 627 + ], + "type": "text", + "content": "7: Get the gradient " + }, + { + "bbox": [ + 321, + 615, + 499, + 627 + ], + "type": "inline_equation", + "content": "\\tilde{g}_i = \\nabla_x\\mathcal{L}(\\hat{x}_i,y)" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 628, + 390, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 628, + 390, + 636 + ], + "spans": [ + { + "bbox": [ + 321, + 628, + 390, + 636 + ], + "type": "text", + "content": "8: end while" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 636, + 489, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 636, + 489, + 656 + ], + "spans": [ + { + "bbox": [ + 321, + 636, + 489, + 656 + ], + "type": "text", + "content": "9: Update " + }, + { + "bbox": [ + 321, + 636, + 489, + 656 + ], + "type": "inline_equation", + "content": "v_{t} = \\gamma \\cdot v_{t - 1} + \\frac{\\sum_{i = 1}^{N}\\tilde{g}_{i}}{\\left\\|\\sum_{i = 1}^{N}\\tilde{g}_{i}\\right\\|_{1}}" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 656, + 533, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 656, + 533, + 670 + ], + "spans": [ + { + "bbox": [ + 321, + 656, + 533, + 670 + ], + "type": "text", + "content": "0: Update " + }, + { + "bbox": [ + 321, + 656, + 533, + 670 + ], + "type": "inline_equation", + "content": "x_{t}^{adv} = \\Pi_{\\mathcal{B}_{\\epsilon}(x)}\\left[x_{t - 1}^{adv} + \\alpha \\cdot \\mathrm{sign}(v_{t})\\right]" + }, + { + "bbox": [ + 321, + 656, + 533, + 670 + ], + "type": "text", + "content": ";" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 670, + 375, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 670, + 375, + 679 + ], + "spans": [ + { + "bbox": [ + 321, + 670, + 375, + 679 + ], + "type": "text", + "content": "1: end while" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 318, + 680, + 384, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 680, + 384, + 693 + ], + "spans": [ + { + "bbox": [ + 318, + 680, + 384, + 693 + ], + "type": "text", + "content": "12: " + }, + { + "bbox": [ + 318, + 680, + 384, + 693 + ], + "type": "inline_equation", + "content": "x^{adv} = x_T^{adv}" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3889" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 125, + 557, + 415 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "text", + "content": "Table 1. The average untargeted and targeted ASR (%) on the nine models (Res50 [10], Dense121 [14], Inc-v3 [26], IncRes-v2 [27], ViT-B [3], PiT-B [11], and Inc-v3ens3, Inc-v3ens4, IncRes-v2ens [29]), with adversarial examples generated on a single model (Res50, Inc-v3 or ViT-B). Each data pair " + }, + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "inline_equation", + "content": "(u/w)" + }, + { + "bbox": [ + 55, + 70, + 555, + 116 + ], + "type": "text", + "content": " corresponds to the performances under (untargeted/targeted) attacks. The best and second best results are labeled in bold and underline, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 125, + 557, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 125, + 557, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 125, + 557, + 415 + ], + "type": "table", + "html": "
AttackRes50Dense121Inc-v3IncRes-v2ViT-BPiT-BInc-v3ens3Inc-v3ens4IncRes-v2ensAvg.
Res50MI CVPR'1899.8 / 98.154.9 / 0.244.2 / 0.028.0 / 0.011.8 / 0.022.4 / 0.022.8 / 0.025.3 / 0.019.6 / 0.036.53 / 10.92
N ICLR'20100.0 / 95.163.1 / 0.247.2 / 0.032.3 / 0.012.9 / 0.023.7 / 0.023.0 / 0.027.5 / 0.019.0 / 0.038.74 / 10.59
VMI CVPR'2199.8 / 96.671.2 / 2.858.1 / 0.549.4 / 0.231.3 / 0.147.7 / 0.543.9 / 0.045.4 / 0.137.7 / 0.153.83 / 11.21
RAP NeurIPS'2299.6 / 26.486.8 / 0.568.4 / 0.052.6 / 0.022.2 / 0.040.5 / 0.136.2 / 0.036.7 / 0.025.5 / 0.052.06 / 3.00
GRA ICCV'2396.9 / 67.788.6 / 3.981.8 / 0.975.8 / 1.745.3 / 0.262.6 / 0.671.5 / 0.770.9 / 1.167.3 / 0.773.41 / 8.61
PGN NeurIPS'2398.6 / 49.591.3 / 4.785.0 / 1.478.5 / 1.749.7 / 0.467.8 / 1.074.9 / 0.672.9 / 1.170.1 / 1.276.53 / 6.84
ANDA CVPR'2499.9 / 97.687.5 / 5.473.6 / 0.466.3 / 0.842.3 / 0.161.8 / 1.151.0 / 0.152.9 / 0.146.7 / 0.164.67 / 11.74
GI ESWA'24100.0 / 98.568.7 / 0.455.6 / 0.137.6 / 0.115.3 / 0.029.0 / 0.028.8 / 0.030.4 / 0.122.4 / 0.043.09 / 11.02
GGS99.3 / 90.795.9 / 28.689.6 / 7.286.4 / 8.660.2 / 3.481.0 / 8.577.5 / 3.777.4 / 4.371.4 / 4.082.08 / 17.67
Inc-v3MI32.1 / 0.049.6 / 0.1100.0 / 95.356.1 / 0.113.0 / 0.017.3 / 0.034.3 / 0.038.0 / 0.026.3 / 0.040.74 / 10.61
NI41.4 / 0.060.5 / 0.1100.0 / 75.966.9 / 0.113.4 / 0.019.2 / 0.036.8 / 0.040.6 / 0.129.5 / 0.045.37 / 8.47
VMI44.9 / 0.165.6 / 0.199.9 / 76.672.7 / 0.418.3 / 0.025.3 / 0.048.9 / 0.152.5 / 0.040.9 / 0.052.11 / 8.59
RAP58.0 / 0.078.4 / 0.2100.0 / 16.482.1 / 0.115.4 / 0.023.8 / 0.044.1 / 0.147.0 / 0.033.9 / 0.053.63 / 1.87
GRA53.3 / 0.175.4 / 0.499.8 / 67.585.4 / 1.821.4 / 0.030.5 / 0.164.9 / 0.165.9 / 0.356.1 / 0.161.41 / 7.82
PGN56.0 / 0.079.4 / 0.2100.0 / 50.387.9 / 0.923.2 / 0.033.4 / 0.065.9 / 0.067.0 / 0.155.6 / 0.063.16 / 5.72
ANDA51.3 / 0.376.3 / 0.4100.0 / 87.482.5 / 0.620.9 / 0.031.8 / 0.158.2 / 0.261.2 / 0.249.3 / 0.059.06 / 9.91
GI39.9 / 0.059.3 / 0.3100.0 / 97.162.6 / 0.514.6 / 0.019.9 / 0.039.2 / 0.040.9 / 0.031.9 / 0.045.37 / 10.88
GGS69.1 / 1.186.9 / 1.9100.0 / 76.095.8 / 6.527.9 / 0.039.7 / 0.170.0 / 0.773.9 / 1.460.1 / 0.569.27 / 9.80
ViT-BMI41.7 / 0.055.6 / 0.058.1 / 0.039.4 / 0.0100.0 / 99.950.1 / 0.034.2 / 0.036.3 / 0.029.9 / 0.049.48 / 11.10
NI42.4 / 0.058.3 / 0.059.5 / 0.044.7 / 0.1100.0 / 99.050.0 / 0.236.2 / 0.037.6 / 0.031.0 / 0.051.08 / 11.03
VMI54.8 / 0.665.9 / 0.463.1 / 0.150.9 / 0.3100.0 / 99.067.6 / 1.545.3 / 0.144.7 / 0.038.7 / 0.259.00 / 11.36
RAP62.1 / 0.179.6 / 0.273.7 / 0.062.3 / 0.099.9 / 57.066.2 / 0.148.5 / 0.051.2 / 0.042.6 / 0.065.12 / 6.38
GRA64.9 / 1.077.6 / 1.174.7 / 0.868.8 / 0.899.3 / 91.780.9 / 4.761.5 / 0.264.0 / 0.157.7 / 0.272.16 / 11.18
PGN69.3 / 0.481.1 / 0.778.6 / 0.470.2 / 0.599.1 / 82.284.8 / 4.465.6 / 0.167.5 / 0.361.4 / 0.275.29 / 9.91
ANDA66.2 / 0.379.0 / 0.575.5 / 0.165.7 / 0.3100.0 / 99.979.0 / 1.455.3 / 0.156.7 / 0.149.9 / 0.269.70 / 11.44
GI55.8 / 0.169.6 / 0.267.1 / 0.052.5 / 0.1100.0 / 100.061.0 / 0.345.4 / 0.046.3 / 0.038.0 / 0.059.52 / 11.19
GGS80.8 / 6.089.9 / 6.587.4 / 4.682.3 / 4.6100.0 / 99.292.7 / 27.374.0 / 2.174.5 / 2.368.4 / 2.483.33 / 17.22
", + "image_path": "1d551a3653a2aeb83378c449f7c51a96492a870e8f73be17d2b511c090532a0d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 137, + 449 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 182, + 468 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 295, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 520 + ], + "type": "text", + "content": "Dataset: Our experiments are conducted on an ImageNet-compatible dataset that is widely used in adversarial attack research. The dataset includes 1,000 images, each with a size of " + }, + { + "bbox": [ + 55, + 473, + 295, + 520 + ], + "type": "inline_equation", + "content": "299 \\times 299 \\times 3" + }, + { + "bbox": [ + 55, + 473, + 295, + 520 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 629 + ], + "type": "text", + "content": "Models: We select six normally pre-trained models, i.e., ResNet50 (Res50) [10], DenseNet121 (Dense121) [14], Inception-v3 (Inc-v3) [26], InceptionResNet-v2 (IncRes-v2) [27], VisionTransformer-Base (ViT-B) [3], and Pooling-based VisionTransformer-Base (PiT-B) [11], along with three models pre-trained on adversarial examples Inception-v3_ens3 (Inc-v3_ens3), Inception-v3_ens4 (Inc-v3_ens4) and InceptionResNet-v2_ens (IncRes-v2_ens) [29], to evaluate the performance of different methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Baselines: To provide a comprehensive evaluation of our method, we choose eight popular gradient-based adversarial attack methods as baselines: MI-FGSM [1], NIFGSM [19], VMI-FGSM [32], RAP [24], GRA [40], PGN [7], ANDA [5] and GI [30]. Additionally, we integrate our method with five input transformation-based methods, i.e., DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 436, + 498, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 498, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 498, + 449 + ], + "type": "text", + "content": "to verify the generalizability of our approach." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": "Hyperparameters: For our GGS, we set the maximum perturbation " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\epsilon = 16 / 255" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": ", the number of outer-iterations " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "T = 10" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": ", the step size " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\alpha = \\epsilon / T" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": ", the number of inner-iterations " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "N = 20" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": ", and the upper bound factor of sample range " + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\zeta = 2.0 \\times \\epsilon" + }, + { + "bbox": [ + 313, + 449, + 554, + 521 + ], + "type": "text", + "content": ". More hyperparameter explanation for other methods is provided in the Supplementary Material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 534, + 465, + 546 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 465, + 546 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 465, + 546 + ], + "type": "text", + "content": "4.2. Attack on Different Models" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 553, + 429, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 429, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 429, + 564 + ], + "type": "text", + "content": "4.2.1. Classification model" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "As shown in Table 1, our method consistently outperforms state-of-the-art (SOTA) approaches. Compared to the best-performing competitor, it improves the average attack success rate (ASR) by over " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " in both untargeted and targeted attacks across most architectures, except for Inception-v3. Furthermore, our GGS significantly enhances attack performance on the surrogate model, achieving " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "4\\% -25\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " higher ASR in targeted attacks than other sampling methods like PGN [7] and GRA [40]. This demonstrates GGS's ability to boost attack potency while maintaining strong cross-model generalization. Additional details on ensemble setting can be found in the supplementary materials." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3890" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 558, + 277 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 558, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 558, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 558, + 277 + ], + "type": "image", + "image_path": "be989cca6b14e70b2dce74843ac61ef3c16bb5de095c8a60ed785a800b77d7a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "lines": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "spans": [ + { + "bbox": [ + 54, + 283, + 555, + 350 + ], + "type": "text", + "content": "Figure 4. The loss surfaces of adversarial examples generated by different methods (MI-FGSM [1], GRA [40], PGN [7], ANDA [5]), based on three different model architectures (ResNet50 [10], Inception-v3 [26], and ViT-B [3]) along random directions with varying strengths like [7]. In the images marked with a star in the upper left corner, adversarial examples are generated and tested on the same model, indicating white-box testing. In contrast, unmarked images represent adversarial examples generated on one model and tested on another, indicating black-box testing. We have also highlighted the regions covered by our loss surface using a red background for easier visualization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 370, + 241, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 241, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 241, + 383 + ], + "type": "text", + "content": "4.2.2. Multimodal Large Language Models" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 386, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 386, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 386, + 296, + 482 + ], + "type": "text", + "content": "The widespread adoption of Multimodal Large Language Models (MLLMs) has raised significant security concerns. We evaluated our method on five prominent MLLMs using adversarial examples generated under an ensemble setting, as shown in Table 2. Our GGS method achieved superior attack performance, reducing the average CSR by over " + }, + { + "bbox": [ + 55, + 386, + 296, + 482 + ], + "type": "inline_equation", + "content": "9\\%" + }, + { + "bbox": [ + 55, + 386, + 296, + 482 + ], + "type": "text", + "content": " compared to the strongest baseline, which demonstrates its excellent cross-model applicability." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 58, + 562, + 294, + 711 + ], + "blocks": [ + { + "bbox": [ + 55, + 493, + 296, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 559 + ], + "type": "text", + "content": "Table 2. Classification success rate (CSR, %) of adversarial examples generated by different methods in Multimodal Large Language Models (MLLMs) from three major providers (OpenAI, Google, Anthropic) under ensemble settings (Res50 [10], Inc-v3 [26], ViT-B [3]). Row \"clean\" indicates the CSR on the clean samples. Avg. represents the average CSR across the five MLLMs." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 562, + 294, + 711 + ], + "lines": [ + { + "bbox": [ + 58, + 562, + 294, + 711 + ], + "spans": [ + { + "bbox": [ + 58, + 562, + 294, + 711 + ], + "type": "table", + "html": "
MethodGPTGeminiClaude
4ominiproflashsonnetAvg. ↓
Clean77.184.085.680.968.379.18
MI_CVPR'18 [1]77.077.381.876.561.074.72
NI_ICLR'20 [19]75.073.983.176.862.474.24
VMI_CVPR'21 [32]68.866.678.473.263.170.02
RAP_NeurIPS'22 [24]64.362.075.166.160.265.54
GRA_ICCV'23 [40]57.152.270.965.646.758.50
PGN_NeurIPS'23 [7]56.249.069.664.145.156.80
ANDA_CVPR'24 [5]59.956.874.969.350.562.28
GI_ESWA'24 [30]71.572.879.473.263.872.14
GGS43.137.861.155.740.047.54
", + "image_path": "6150c92b71d4ee276a0d3f7a351efd891f22fb6acdf02e0688b6cb87f66b6823.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 370, + 544, + 382 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 544, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 544, + 382 + ], + "type": "text", + "content": "4.2.3. Loss Surface Visualization against Noise Levels" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 386, + 555, + 566 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 566 + ], + "type": "text", + "content": "To shed light on the capacity of our GGS in locating flat maxima regions for adversarial examples, we compared the loss surfaces of adversarial examples generated by different attack methods in Fig. 4. Each curve represents the average loss of 32 randomly selected adversarial examples, while the center of each image (noise level " + }, + { + "bbox": [ + 313, + 386, + 555, + 566 + ], + "type": "inline_equation", + "content": "= 0" + }, + { + "bbox": [ + 313, + 386, + 555, + 566 + ], + "type": "text", + "content": ") indicates the average loss of clean adversarial examples. In white-box settings (marked with a star), GGS consistently finds flatter maxima compared to baselines. Consequently, in black-box testing, GGS achieves the highest loss values (images without star), indicating its superior transfer attack capability. Additionally, the GGS loss surface (highlighted in red) almost entirely encompasses those of other methods, confirming its effectiveness in balancing strong attack potency with cross-model generalization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 574, + 500, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 500, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 500, + 588 + ], + "type": "text", + "content": "4.3. Compatibility with Other Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "content": "Attack with Inner-Iteration Sampling Methods: Our Gradient-Guided Sampling (GGS) is compatible with existing inner-iteration RS-based approaches. When integrated with SOTA approaches like GRA [40] and PGN [7], GGS largely improves attack success rates (ASR). As shown in Table 3, untargeted ASR improved by " + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "inline_equation", + "content": "5.2\\%" + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "content": " (GRA) and " + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "inline_equation", + "content": "6.7\\%" + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "content": " (PGN), while targeted ASR increased by " + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "content": " (GRA) and " + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "inline_equation", + "content": "6\\%" + }, + { + "bbox": [ + 313, + 593, + 555, + 689 + ], + "type": "text", + "content": " (PGN), demonstrating GGS's compatibility." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Attack with Input Transformations Methods: To study the compatibility of our GGS with existing input" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3891" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 180, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 180, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 180, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 180, + 155 + ], + "type": "image", + "image_path": "8cb8d7a553ac8ca343bd28868d690ad6b0905d3d7be442b34ceaaa8dbc101b08.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 156, + 148, + 166 + ], + "lines": [ + { + "bbox": [ + 94, + 156, + 148, + 166 + ], + "spans": [ + { + "bbox": [ + 94, + 156, + 148, + 166 + ], + "type": "text", + "content": "(a) Random (RS)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 181, + 70, + 299, + 155 + ], + "blocks": [ + { + "bbox": [ + 181, + 70, + 299, + 155 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 299, + 155 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 299, + 155 + ], + "type": "image", + "image_path": "eb30b3b6476906dc0618bc9aeb22f4f3ba923781b64172d9eec5cfcd879634dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 194, + 156, + 291, + 166 + ], + "lines": [ + { + "bbox": [ + 194, + 156, + 291, + 166 + ], + "spans": [ + { + "bbox": [ + 194, + 156, + 291, + 166 + ], + "type": "text", + "content": "(b) Momentum-Guided (MGS)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "type": "text", + "content": "Figure 5. Loss surfaces of adversarial examples generated by different sampling strategies, Random Sampling (RS), MGS and our GGS, with increasing inner-iteration on Resnet50. The red line on the left side of part (a)~(c) represents the maximum value of the loss surface in each iteration, and the blue dashed line highlights these values in early-stage inner iteration. For (d), it represents the cosine similarity between the gradient " + }, + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "type": "inline_equation", + "content": "\\tilde{g}_i" + }, + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "type": "text", + "content": " generated in each inner-iteration and the average of gradients " + }, + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{N} \\tilde{g}_i / N" + }, + { + "bbox": [ + 55, + 175, + 553, + 221 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 299, + 70, + 419, + 155 + ], + "blocks": [ + { + "bbox": [ + 299, + 70, + 419, + 155 + ], + "lines": [ + { + "bbox": [ + 299, + 70, + 419, + 155 + ], + "spans": [ + { + "bbox": [ + 299, + 70, + 419, + 155 + ], + "type": "image", + "image_path": "9232b6d667d31188f1c864ee7a648e4c116d986e9440d9b7af55415a25b0f791.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 325, + 156, + 411, + 165 + ], + "lines": [ + { + "bbox": [ + 325, + 156, + 411, + 165 + ], + "spans": [ + { + "bbox": [ + 325, + 156, + 411, + 165 + ], + "type": "text", + "content": "(c) Gradient-Guided (GGS)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 420, + 70, + 553, + 148 + ], + "blocks": [ + { + "bbox": [ + 420, + 70, + 553, + 148 + ], + "lines": [ + { + "bbox": [ + 420, + 70, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 420, + 70, + 553, + 148 + ], + "type": "image", + "image_path": "4788acac99032229f83e43fa9ddedeb10c116d6066817e4027a343a1098a4465.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 149, + 552, + 165 + ], + "lines": [ + { + "bbox": [ + 429, + 149, + 552, + 165 + ], + "spans": [ + { + "bbox": [ + 429, + 149, + 552, + 165 + ], + "type": "text", + "content": "(d) The cosine similarity between the gradient and the gradients averaging in inner-iterations" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 303, + 298, + 409 + ], + "blocks": [ + { + "bbox": [ + 55, + 239, + 297, + 294 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 297, + 294 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 297, + 294 + ], + "type": "text", + "content": "Table 3. The average untargeted and targeted ASR " + }, + { + "bbox": [ + 55, + 239, + 297, + 294 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 239, + 297, + 294 + ], + "type": "text", + "content": " on all nine models with the adversarial examples generated on ResNet50, when combined our GGS with input transformation-based methods (DIM [37], TIM [2], SIM [19], Admix [33], SSM [21]) and inner-iteration RS-based methods (GRA [40], PGN [7])." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 303, + 298, + 409 + ], + "lines": [ + { + "bbox": [ + 56, + 303, + 298, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 303, + 298, + 409 + ], + "type": "table", + "html": "
ClassMethodUntargeted Avg.Targeted Avg.
Random Sampling (RS)GRA / +ours73.41 / 78.698.61 / 13.64
PGN / +ours76.53 / 83.236.84 / 12.79
Input TransformationDIM / +ours51.30 / 90.138.63 / 19.37
TIM / +ours40.82 / 84.9411.03 / 18.48
SIM / +ours46.80 / 90.1211.14 / 28.44
Admix / +ours54.78 / 85.5011.28 / 30.50
SSM / +ours74.22 / 80.748.40 / 12.54
", + "image_path": "d324da4f8d411323608fcac32ff8430a80c9ce9b6c6e67897a1a1180a45157c7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": "transformation techniques, we integrated it with five methods of this type: DIM [37], TIM [2], SIM [19], Admix [33], and SSM [21]. Adversarial examples generated on ResNet50 were tested on nine diverse models. Table 3 shows GGS significantly improved adversarial transferability, increasing untargeted ASR by " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "6\\% -43\\%" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " and targeted ASR by " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "4\\% -19\\%" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": ". For instance, SIM's untargeted ASR improved by " + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 55, + 427, + 296, + 535 + ], + "type": "text", + "content": " when combined with GGS. More details are in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 541, + 269, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 541, + 269, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 541, + 269, + 555 + ], + "type": "text", + "content": "4.4. Ablation Study of Proposed Components" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 61, + 610, + 290, + 679 + ], + "blocks": [ + { + "bbox": [ + 55, + 568, + 295, + 601 + ], + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 601 + ], + "type": "text", + "content": "Table 4. The average untargeted ASR on ResNet50 and eight target models for Random Sampling (RS), Momentum-Guided Sampling (MGS), and our Gradient-Guided Sampling (GGS)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 610, + 290, + 679 + ], + "lines": [ + { + "bbox": [ + 61, + 610, + 290, + 679 + ], + "spans": [ + { + "bbox": [ + 61, + 610, + 290, + 679 + ], + "type": "table", + "html": "
Sampling Guided byRes50Others
randmomentumgradient
(a) RS✓--97.363.74
(b) MGS-✓-97.558.79
(c) GGS--✓99.379.93
", + "image_path": "e6d30d42f25c2b0268fe6df2fe6c8509ed4690a9c700ba45ea85c092c0b5d1e5.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 714 + ], + "type": "text", + "content": "Ablation study compares RS, MGS, and GGS via loss surface analysis (Fig. 5) and attack success rates (Table 4)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 396 + ], + "type": "text", + "content": "As shown in Fig. 5(a), Random Sampling (RS) yields a relatively flat loss surface due to its exploration capability but it suffers from significant oscillation in loss values. Introducing Momentum-Guided Sampling (MGS) (Fig. 5(b)) stabilizes the loss surface over iterations, but its long-chain dependency causes high similarity between later gradients and the final average gradient (Fig. 5(d)), severely limiting exploration. Our GGS resolves MGS's long-chain dependency and mitigates early-stage noisy gradients (Fig. 5(b)(c), blue dashed lines). The low gradient similarity in Fig. 5(d) reflects the enhanced exploration, while the flat yet higher-loss surface in Fig. 5(c) demonstrates the exceptional sampling efficiency of our GGS." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": "Quantitatively, as shown in Table 4, MGS marginally improves ASR of the surrogate model by " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "0.2\\%" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": ", but reducing transfer attack capability by " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": ". In contrast, GGS increases both white-box and transfer ASR by " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "inline_equation", + "content": "16\\%" + }, + { + "bbox": [ + 313, + 396, + 554, + 456 + ], + "type": "text", + "content": ", respectively, validating its superior sampling efficiency." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 466, + 474, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 474, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 474, + 480 + ], + "type": "text", + "content": "5. Conclusion and Perspectives" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 665 + ], + "type": "text", + "content": "This paper introduces Gradient-Guided Sampling (GGS), an easy-to-implement inner-iteration sampling strategy, for transferable adversarial attacks that effectively resolves the dilemma between Exploitation (attack potency) and Exploration (cross-model generalization). By guiding the sampling direction with the gradient from the previous inner iteration, GGS enables stable ascent toward flat loss regions with higher local maxima. Our GGS has the merit of enhancing exploitation capability without compromising exploration capability, effectively balancing these two aspects and addressing the inherent limitations of random sampling (RS) methods, as verified by extensive experiments and visualizations as well as the comparison with related state-of-the-art methods. GGS is also revealed to be highly compatible with existing methods." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "While our GGS is highly compatible with gradient-averaging methods (e.g., GRA [40], PGN [7]), we aim to refine it to also support non-gradient-averaging techniques (e.g., VMI-FGSM [32], RAP [24])." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3892" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 224 + ], + "type": "text", + "content": "The work was supported by the National Natural Science Foundation of China under grants no. 62276170, 82261138629, 62306061, the Science and Technology Project of Guangdong Province under grants no. 2023A1515010688, the Science and Technology Innovation Commission of Shenzhen under grant no. JCYJ20220531101412030, Open Research Fund from Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) under Grant No. GML-KF-24-11, and Guangdong Provincial Key Laboratory under grant no. 2023B1212060076." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "spans": [ + { + "bbox": [ + 56, + 234, + 115, + 247 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 254, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 61, + 254, + 294, + 308 + ], + "type": "text", + "content": "[1] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 310, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 310, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 61, + 310, + 296, + 365 + ], + "type": "text", + "content": "[2] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4312-4321, 2019. 2, 3, 6, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 366, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 366, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 61, + 366, + 296, + 431 + ], + "type": "text", + "content": "[3] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 487 + ], + "type": "text", + "content": "[4] Yan Fang, Zhongyuan Wang, Jikang Cheng, Ruoxi Wang, and Chao Liang. Promoting adversarial transferability with enhanced loss flatness. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1217-1222. IEEE, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 544 + ], + "type": "text", + "content": "[5] Zhengwei Fang, Rui Wang, Tao Huang, and Liping Jing. Strong transferable adversarial attacks via ensembled asymptotically normal distribution learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24841-24850, 2024. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 589 + ], + "type": "text", + "content": "[6] Pierre Foret, Ariel Kleiner, Hossein Mobahi, and Behnam Neyshabur. Sharpness-aware minimization for efficiently improving generalization. In International Conference on Learning Representations, ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 644 + ], + "type": "text", + "content": "[7] Zhijin Ge, Hongying Liu, Wang Xiaosen, Fanhua Shang, and Yuanyuan Liu. Boosting adversarial transferability by achieving flat local maxima. Advances in Neural Information Processing Systems, 36:70141-70161, 2023. 1, 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 296, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 296, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 296, + 700 + ], + "type": "text", + "content": "[8] Zhijin Ge, Fanhua Shang, Hongying Liu, Yuanyuan Liu, Liang Wan, Wei Feng, and Xiaosen Wang. Improving the transferability of adversarial examples with arbitrary style transfer. In Proceedings of the 31st ACM International Conference on Multimedia, pages 4440-4449, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 703, + 296, + 714 + ], + "type": "text", + "content": "[9] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 105 + ], + "type": "text", + "content": "Explaining and harnessing adversarial examples. In International Conference on Learning Representations, ICLR, 2015. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 151 + ], + "type": "text", + "content": "[10] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 556, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 556, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 556, + 207 + ], + "type": "text", + "content": "[11] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 11936-11945, 2021. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 209, + 554, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 229 + ], + "type": "text", + "content": "[12] Sepp Hochreiter and Jürgen Schmidhuber. Flat minima. Neural computation, 9(1):1-42, 1997. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 555, + 285 + ], + "type": "text", + "content": "[13] SM Mostaq Hossain, Shampa Banik, Trapa Banik, and Ashfak Md Shibli. Survey on security attacks in connected and autonomous vehicular systems. In 2023 IEEE International Conference on Computing (ICOCO), pages 295-300. IEEE, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 554, + 331 + ], + "type": "text", + "content": "[14] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 333, + 554, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 333, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 333, + 554, + 376 + ], + "type": "text", + "content": "[15] Martin Husák, Jana Komárková, Elias Bou-Harb, and Pavel Čeleda. Survey of attack projection, prediction, and forecasting in cyber security. IEEE Communications Surveys & Tutorials, 21(1):640-660, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 554, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 554, + 421 + ], + "type": "text", + "content": "[16] Mintong Kang, Dawn Song, and Bo Li. Diffattack: Evasion attacks against diffusion-based adversarial purification. In Advances in Neural Information Processing Systems, pages 73919-73942. Curran Associates, Inc., 2023. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 554, + 466 + ], + "type": "text", + "content": "[17] Nitish Shirish Keskar, Dheevatsa Mudigere, Jorge Nocedal, Mikhail Smelyanskiy, and Ping Tak Peter Tang. On large-batch training for deep learning: Generalization gap and sharp minima. ICLR, 2017. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 468, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 554, + 510 + ], + "type": "text", + "content": "[18] Hao Li, Zheng Xu, Gavin Taylor, Christoph Studer, and Tom Goldstein. Visualizing the loss landscape of neural nets. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 513, + 554, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 554, + 556 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 554, + 556 + ], + "type": "text", + "content": "[19] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, ICLR, 2020. 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 558, + 554, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 554, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 554, + 612 + ], + "type": "text", + "content": "[20] Decheng Liu, Xijun Wang, Chunlei Peng, Nannan Wang, Ruimin Hu, and Xinbo Gao. Adv-diffusion: Imperceptible adversarial face identity attack via latent diffusion model. Proceedings of the AAAI Conference on Artificial Intelligence, 38(4):3585-3593, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 554, + 668 + ], + "type": "text", + "content": "[21] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European conference on computer vision, pages 549-566. Springer, 2022. 3, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[22] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In International Conference on Learning Representations, ICLR, 2018. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3893" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[23] Yurii Nesterov. A method for unconstrained convex minimization problem with the rate of convergence o (1/k2). In Dokl. Akad. Nauk. SSSR, page 543, 1983. 2, 4, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[24] Zeyu Qin, Yanbo Fan, Yi Liu, Li Shen, Yong Zhang, Jue Wang, and Baoyuan Wu. Boosting the transferability of adversarial attacks with reverse adversarial perturbation. Advances in neural information processing systems, 35:29845-29858, 2022. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 196 + ], + "type": "text", + "content": "[25] David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. Learning representations by back-propagating errors. nature, 323(6088):533-536, 1986. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 251 + ], + "type": "text", + "content": "[26] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 295, + 297 + ], + "type": "text", + "content": "[27] Christian Szegedy, Sergey Ioffe, Vincent Vanhoucke, and Alexander Alemi. Inception-v4, inception-reset and the impact of residual connections on learning. In Proceedings of the AAAI conference on artificial intelligence, 2017. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 352 + ], + "type": "text", + "content": "[28] Chee-Wooi Ten, Govindarasu Manimaran, and Chen-Ching Liu. Cybersecurity for critical infrastructures: Attack and defense modeling. IEEE Transactions on Systems, Man, and Cybernetics-Part A: Systems and Humans, 40(4):853–865, 2010. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 295, + 387 + ], + "type": "text", + "content": "[29] Florian Tramér, Alexey Kurakin, Nicolas Papernot, Ian Goodfellow, Dan Boneh, and Patrick McDaniel. Ensemble adversarial training: Attacks and defenses. *ICLR*, 2018. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 443 + ], + "type": "text", + "content": "[30] Jiafeng Wang, Zhaoyu Chen, Kaixun Jiang, Dingkang Yang, Lingyi Hong, Pinxue Guo, Haijing Guo, and Wenqiang Zhang. Boosting the transferability of adversarial attacks with global momentum initialization. Expert Systems with Applications, 255:124757, 2024. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "text", + "content": "[31] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[32] Xiaosen Wang and Kun He. Enhancing the transferability of adversarial attacks through variance tuning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1924-1933, 2021. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 589 + ], + "type": "text", + "content": "[33] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 635 + ], + "type": "text", + "content": "[34] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 679 + ], + "type": "text", + "content": "[35] Han Wu, Syed Yunas, Sareh Rowlands, Wenjie Ruan, and Johan Wahlström. Adversarial driving: Attacking end-to-end autonomous driving. In 2023 IEEE Intelligent Vehicles Symposium (IV), pages 1-7. IEEE, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[36] Chaowei Xiao, Bo Li, Jun yan Zhu, Warren He, Mingyan Liu, and Dawn Song. Generating adversarial examples with adversarial networks. In Proceedings of the Twenty-" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 363 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Seventh International Joint Conference on Artificial Intelligence, IJCAI-18, pages 3905-3911, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 150 + ], + "type": "text", + "content": "[37] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 2, 3, 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[38] Yinhu Xu, Qi Chu, Haojie Yuan, Zixiang Luo, Bin Liu, and Nenghai Yu. Enhancing adversarial transferability from the perspective of input loss landscape. In International Conference on Image and Graphics, pages 254-266. Springer, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[39] Yang Zhao, Hao Zhang, and Xiuyuan Hu. Penalizing gradient norm for efficiently improving generalization in deep learning. In International Conference on Machine Learning, pages 26982-26992. PMLR, 2022. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[40] Hegui Zhu, Yuchen Ren, Xiaoyan Sui, Lianping Yang, and Wuming Jiang. Boosting adversarial transferability via gradient relevance attack. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4741-4750, 2023. 2, 3, 4, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[41] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 3" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3894" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_content_list.json b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..56bc5ad7de157a9ebed08e38f3a427fdb3640052 --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_content_list.json @@ -0,0 +1,1726 @@ +[ + { + "type": "text", + "text": "Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features", + "text_level": 1, + "bbox": [ + 114, + 128, + 883, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chancharik Mitra $^{1*}$ Brandon Huang $^{2*}$ Tianning Chai $^{2}$ Zhiqiu Lin $^{1}$ Assaf Arbelle $^{3}$ Rogerio Feris $^{4}$ Leonid Karlinsky $^{4}$ Trevor Darrell $^{2}$ Deva Ramanan $^{1}$ Roei Herzig $^{2,4}$", + "bbox": [ + 135, + 202, + 859, + 241 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Carnegie Mellon University $^{2}$ University of California, Berkeley $^{3}$ IBM Research $^{4}$ MIT-IBM Watson AI Lab", + "bbox": [ + 233, + 255, + 764, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs)—a finetuning-free method that leverages sparse attention head activations (fewer than $5\\%$ of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations.", + "bbox": [ + 88, + 358, + 485, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 708, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generative Large Multimodal Models (LMMs) such as GPT-4V [73], LLaVA [59, 60], and QwenVL [3] demonstrate state-of-the-art performance on open-ended vision-language (VL) tasks like image captioning [54, 107], visual question answering [2, 32, 43], and language grounding [38, 65]. However, despite their remarkable performance on generative tasks, these models struggle on vision-language classification tasks, where responses are a discrete set of labels [8, 111]. Indeed, LMMs with billions of parameters and trained on trillions more tokens of data underperform smaller encoder VLMs like CLIP and SigLIP [8, 111]", + "bbox": [ + 89, + 733, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/666b706b0f887295d28038f80fd97824ec1f34ab370bb4485266bf6b575b1e24.jpg", + "image_caption": [ + "Figure 1. Sparse Attention Vectors (SAVs) Overview. We develop a method for extracting features from a generative LMM without finetuning. We first extract a sparse set of attention vectors for each task given a set of few-shot examples, and then, we utilize these attention vectors directly as features for downstream vision-language classification tasks." + ], + "image_footnote": [], + "bbox": [ + 516, + 325, + 705, + 588 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/86cf6fc0bf9307c7557fbc168f79bbd5df228af602254180e66b6440ec931564.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 325, + 901, + 602 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and even classical machine learning methods [6] on image classification tasks. One reason that CLIP-like models excel at classification is the ease of extracting visual or textual features from modality-specific encoders. However, such models cannot process joint vision-language inputs, as LMMs can. Our goal is to extract compact features from generative LMMs, enabling downstream discrete labeling tasks such as classification and multiple-choice VQA.", + "bbox": [ + 511, + 729, + 906, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Feature extraction is a well-explored field in both vision-only [15, 83, 94] and language-only encoder models [13, 81], but such is not the case with generative models. Most", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2760", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "current methods for extracting features from generative models require carefully constructed prompts [41, 67, 104], specialized architectures [51], few-shot prompting [7, 11, 112], and finetuning [8, 64, 111]. Recent work [111] shows that prompting approaches fail to close the gap with encoder VLMs, while finetuning requires training-scale data for every unseen task, which is inefficient. However, as we observe in our work, generative models still offer the promise of more flexible, truly multimodal features as compared to modality-specific features from CLIP-like models.", + "bbox": [ + 89, + 90, + 480, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A source of inspiration for our method is long-standing work in neuroscience that suggests certain areas of the brain are reserved for specific tasks [17, 37] (i.e. functional specificity). Motivated by this idea, we refer to recent interpretability research that has focused on identifying specific heads in transformer-based models that correspond to particular tasks [72]. The most prominent of these methods is a line of work that looks to enhance vision-language capabilities using task vectors [24, 27, 30, 93], which are compact implicit representations of tasks encoded in the activations of a transformer model. While promising, these methods ultimately use these representations to augment a model's generative capabilities. On the other hand, we seek to use feature representations directly as classifiers. Nevertheless, this intuition from interpretability informs our work on Sparse Attention Vectors (SAVs), which are sparse features in an LMMs activation space that can be directly exploited for few-shot VL classification.", + "bbox": [ + 91, + 250, + 482, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our method has three steps: First, we extract features (i.e., attention vectors) from the output of each head of the LMM for some few-shot labeled examples ( $\\approx 20$ per label). Second, we average these attention vectors over the examples in each class and evaluate their accuracy as centroids in a class centroid classifier. We then select the top 20 heads by classification accuracy as our SAVs. In this way, we identify a very sparse set of attention vectors (less than $5\\%$ of the total number of heads) that can be used for discriminative tasks. Finally, we perform inference on the given task by doing a majority vote across this sparse set of attention vectors for each new query. This approach requires only few-shot examples at test-time to extract effective multimodal embeddings. An overview is shown in Figure 1.", + "bbox": [ + 89, + 530, + 482, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We summarize our main contributions as follows: (i) We introduce a novel method that yields a sparse set of attention vectors to serve as highly effective features for individual VL classification tasks; (ii) Our method surpasses zero-shot, few-shot, and LoRA fine-tuned baselines across multiple tasks ( $+7\\%$ improvement on average over LoRA on challenging benchmarks like BLINK [18], VLGuard [114], and NaturalBench [44]); (iii) We establish several advantageous properties of our approach, including strong generalization capabilities and favorable scaling characteristics.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Controllable Generation for Classification. Controllable text generation in LMMs guides model outputs to meet specific constraints. For classification tasks with generative LMMs, several approaches exist: test-time hard prompting [7, 104] uses prompt engineering or few-shot examples to elicit class label outputs [11, 88, 102, 105, 111]; direct probability analysis of generated class labels [55, 58] for image-text retrieval; and soft prompting methods that finetune learnable tokens [42, 49]. Other techniques include instruction finetuning [103] on labeled data [6, 111] and preference modeling like DPO and RLHF [16, 74, 79]. Our method, however, is finetuning-free and directly selects class labels without preference data. Most related work shows that internal transformer representations called task vectors [23, 27, 30, 92] can encapsulate ICL example tasks. Beyond previous approaches, we use a sparse set of attention vectors directly as VL classification task task features.", + "bbox": [ + 511, + 114, + 903, + 372 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision-Language Features. Feature extraction seeks useful representations for diverse downstream tasks. Early embedding techniques including autoencoders [4, 39, 62, 63, 82], Word2Vec [66] and GloVe [77] transformed inputs into vector representations, followed by advances in NLP [13, 19, 68, 81] and vision [15, 83, 94]. Recent methods like CLIP and SigCLIP [46, 47, 50, 78, 108, 109] explore multimodal correlations through contrastive learning or sigmoid loss. These representations offer flexibility across tasks [12, 25, 33, 40, 80, 95] and domains [53, 101, 110].", + "bbox": [ + 511, + 373, + 903, + 523 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Interestingly, extracting features from generative models poses unique challenges in identifying optimal extraction points. Some approaches finetune encoder VLMs on LLM-generated data [48, 97] or finetune LMMs directly on specific tasks [31, 111]. More efficient methods finetune encoder-decoder representations for better alignment [36, 64, 69, 70]. Finetuning-free approaches use distillation prompts to extract representations from model weights or activations [34, 35, 41, 61], while other methods employ mixture-of-experts [52], expert models [99], or embedding reranking [21].", + "bbox": [ + 511, + 525, + 903, + 691 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thus, current SOTA faces the following challenges: (1) modality-specific rather than multimodal features, (2) requiring finetuning, (3) limited flexibility due to specialized prompts, and (4) dependence on multiple models. Our approach provides effective multimodal embeddings without any gradient-based finetuning and flexibly apply to various VL classification tasks without additional models.", + "bbox": [ + 511, + 691, + 903, + 797 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 814, + 611, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we outline our approach for using sparse attention vectors from the activation space of a transformer-based large multimodal model (LMM) as features for any VL classification task. The method consists of three main", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2761", + "bbox": [ + 482, + 944, + 511, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2cd40e211ec3da687d4de37aa8df1d752ce1602a2582550ee7c4548b4ee2391b.jpg", + "image_caption": [ + "1. Extract Attention Vectors", + "2. Find Sparse Attention Vectors (SAVs):", + "Select Sparse Attention Vectors" + ], + "image_footnote": [], + "bbox": [ + 106, + 102, + 903, + 333 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4902bb64a69ee271e40c1203d2849b6ad4c020c4b2c3f22df58fca0e8bbbcaa5.jpg", + "image_caption": [ + "3. SAVs for Discriminative Tasks:", + "Figure 2. Sparse Attention Vectors (SAVs) Detailed View. Our method is broken into the following three parts: (1) Given a set of few shot examples to be classified by a frozen LMM, we extract attention vectors across all heads and layers. (2) These attention vectors are averaged across the set of examples for each class. For each head, we use these mean attention vectors as centroid classifiers, which are then used to select a sparse set of $k$ heads with the highest classification accuracy. (3) Finally, we use these sparse attention vectors to directly classify new inputs via majority vote (denoted by the \"Classifier\" block)." + ], + "image_footnote": [], + "bbox": [ + 116, + 351, + 861, + 414 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "steps: (i) extracting the attention vectors from all attention heads in the model, (ii) identifying a sparse set of vectors based on their ability to consistently return the correct label for some support set of examples, and (iii) using these sparse features to classify new queries. We begin with a formal description of the transformer decoder LLM and its attention mechanism, followed by the detailed methodology for sparse attention vector selection and classification. A detailed view of our method is shown in Figure 2.", + "bbox": [ + 88, + 523, + 480, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 686, + 230, + 700 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A transformer-based large language model (LLM) with $L$ layers and $H$ attention heads per layer processes input sequences through multi-head self-attention mechanisms. Each layer combines multiple attention heads to capture different aspects of the input sequence, followed by feedforward networks for further processing.", + "bbox": [ + 89, + 713, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-Head Attention. Let $x = \\{x_{1},x_{2},\\ldots ,x_{T}\\}$ represent a sequence of input tokens, where $x_{i}$ is the $i^{\\text{th}}$ token. For each layer $l\\in \\{1,\\dots ,L\\}$ , the input sequence is projected into queries, keys, and values for each attention head $m\\in \\{1,\\dots ,H\\}$ . Each head performs the following scaled dot-product attention:", + "bbox": [ + 88, + 809, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "", + "bbox": [ + 586, + 130, + 903, + 332 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {h} _ {l} ^ {m} \\left(x _ {i}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {m}}}\\right) V\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 539, + 816, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $Q$ , $K$ , and $V$ are the query, key, and value matrices respectively, and the dimensionality of each head $d_{m}$ which is given by $\\frac{d}{H}$ (the embedding dimension divided by the number of heads). We denote $\\mathbf{h}_l^m(x_i)$ as an attention vector for head $m$ in layer $l$ .", + "bbox": [ + 511, + 573, + 906, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The outputs of all heads are concatenated and projected to form the layer output:", + "bbox": [ + 511, + 648, + 905, + 679 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M u l t i H e a d} \\left(x _ {i}\\right) = \\operatorname {C o n c a t} \\left(\\mathbf {h} _ {l} ^ {1} \\left(x _ {i}\\right), \\dots , \\mathbf {h} _ {l} ^ {H} \\left(x _ {i}\\right)\\right) W ^ {O}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 688, + 877, + 708 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W^{O}$ is the output projection matrix.", + "bbox": [ + 511, + 717, + 795, + 732 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our work, we look to leverage attention vectors for the purpose of vision-language classification tasks. Specifically, the attention vectors are used as latent representations of the inputs to both find attention heads in an LMM suited for a classification task and then perform downstream inference using those selected attention heads. We describe our method in detail in the sections that follow.", + "bbox": [ + 511, + 734, + 905, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Sparse Attention Vectors", + "text_level": 1, + "bbox": [ + 511, + 848, + 741, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our key insight is that within the many attention heads and transformer layers of an LMM, there exists a sparse subset", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2762", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "that can serve as effective features for vision-language classification tasks. We present a three-step method to identify and utilize these features to build lightweight classifiers.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "Step 1: Extracting Attention Vectors. Given a frozen LMM and few-shot examples of sequence-label pairs $\\{(x_1, y_1), (x_2, y_2), \\ldots, (x_N, y_N)\\}$ we first extract the attention vectors for each sequence $x_i$ . Specifically, we compute the attention vector $\\mathbf{h}_l^m(x_i^T)$ for head $m$ from layer $l$ for the final token $x_i^T$ . This yields a set of attention vectors $\\{\\mathbf{h}_l^m(x_i^T) \\mid i = 1, \\dots, N\\}$ for each head $m$ and layer $l$ .", + "Step 2: Identifying Relevant Vectors. The central question is how to identify which attention vectors are naturally suited for the classification task at hand. We evaluate each vector's ability as a local classifier by computing its performance under a nearest class centroid classifier." + ], + "bbox": [ + 89, + 136, + 482, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, for each class $c \\in \\mathcal{C}$ , compute its centroid (or mean) attention vector across the few shot examples:", + "bbox": [ + 89, + 316, + 482, + 348 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu_ {c} ^ {l, m} = \\frac {1}{| N _ {c} |} \\sum_ {j: y _ {j} = c} \\mathbf {h} _ {l} ^ {m} (x _ {j} ^ {T})\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 356, + 380, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{c} = \\{j:y_{j} = c\\}$ is the set of indices of examples with label $c$ . For each input $x_{i}$ , we compute its cosine similarity to each class centroid head:", + "bbox": [ + 89, + 402, + 482, + 449 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {l, m} \\left(x _ {i}, c\\right) = \\frac {\\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\cdot \\mu_ {c} ^ {l , m}}{\\left\\| \\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\right\\| \\left\\| \\mu_ {c} ^ {l , m} \\right\\|}, \\quad \\forall c \\in \\mathcal {C}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 457, + 431, + 494 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we measure the classification accuracy of each head (i.e. local classifier):", + "bbox": [ + 89, + 503, + 482, + 532 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s c o r e} (l, m) = \\sum_ {i = 1} ^ {N} \\mathbf {1} [ \\hat {y} = y _ {i} ]\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 541, + 380, + 582 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where the nearest class centroid label is given as $\\hat{y} = \\operatorname*{argmax}_{c\\in \\mathcal{C}}s_{l,m}(x_i,c)$ , and $\\mathbf{1}[\\cdot ]$ is the indicator function that", + "bbox": [ + 89, + 590, + 482, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "evaluates to 1 when the condition is true (and 0 otherwise).", + "bbox": [ + 89, + 630, + 480, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We denote the set of $k$ top-scoring heads as $\\mathcal{H}_{\\mathrm{SAV}}$ :", + "bbox": [ + 89, + 643, + 423, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\mathcal{H}_{\\mathrm{SAV}} = \\{(l,m)\\mid \\mathrm{score}(l,m)$ is among $k$ highest scores}", + "bbox": [ + 96, + 667, + 473, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 3: Classification with Sparse Attention Vectors.", + "text_level": 1, + "bbox": [ + 89, + 695, + 457, + 709 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given a query sequence $Q$ to classify, we leverage our sparse set of heads $\\mathcal{H}_{\\mathrm{SAV}}$ for prediction. For each head $(l,m)\\in \\mathcal{H}_{\\mathrm{SAV}}$ , we compute the class centroid $\\mu_c^{l,m}$ closest to the query as follows:", + "bbox": [ + 89, + 710, + 482, + 771 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat{y}_{l,m} = \\operatorname *{arg max}_{c\\in \\mathcal{C}}s_{l,m}(Q^{T},c)\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 779, + 383, + 805 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $s_{l,m}(\\cdot, \\cdot)$ is defined as in Step 2. Our final class prediction is that of a global classifier that counts the majority vote across all local classifiers (heads) in $\\mathcal{H}_{\\mathrm{SAV}}$ :", + "bbox": [ + 89, + 814, + 482, + 859 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\underset {y \\in \\mathcal {C}} {\\arg \\max } \\sum_ {(l, m) \\in \\mathcal {H} _ {\\mathrm {S A V}}} \\mathbf {1} [ \\hat {y} _ {l, m} = y ]\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 869, + 392, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This approach reveals a surprising capability of LMMs: with just a few carefully selected attention heads ( $|\\mathcal{H}_{\\mathrm{SAV}}| \\ll LH$ ), we can transform a generative language model into a lightweight vision-language classifier. This finding suggests that classification-relevant features naturally emerge within specific attention heads during model pretraining. In theory, one could replace our local and global classifiers with more complex variants (e.g. linear models, MLPs), something we explore in 3a.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Evaluation", + "text_level": 1, + "bbox": [ + 513, + 239, + 630, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We apply SAVs to two state-of-the-art LMMs—LLaVA-OneVision [45] and Qwen2-VL [98]. We also do a rigorous comparison of our method to strong few-shot and finetuning baselines on a variety of different image-text and image-only vision-language classification tasks covering safety, VQA, image-text retrieval, and simple classification.", + "bbox": [ + 511, + 265, + 903, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 364, + 730, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We implemented our approach in PyTorch [76]. We use the official implementations of each model, and all of our experiments can be run on a single NVIDIA A100 GPU. We use the default hyperparameters provided by the original models' (LLaVA-OV and Qwen2-VL) codebases. These hyperparameters are detailed in Table 8. More details of the implementation is included in the supplementary in Section 8.", + "bbox": [ + 511, + 386, + 905, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Benchmarks", + "text_level": 1, + "bbox": [ + 513, + 516, + 645, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Safety. (1) LMM-Halucination [9] is a dataset which evaluates the hallucinations of the models when answering multimodal tasks. We report the raw classification accuracy of our method. Thus, the set of class labels for this task is given by $\\mathcal{C} = \\{\\text{\"hallucinating\",\"not hallucinating\"}\\}$ . (2) VLGuard [114] is a dataset focusing on vision-language safety which identifies 4 main categories of harmful content: Privacy, Risky Behavior, Deception and Hateful Speech. VLGuard proposes Attack Success Rate (ASR) for evaluating unsafe inputs and Helpfulness for evaluating safe inputs. We simply reformat it as a classification task, where the set of class labels is given by $\\mathcal{C} = \\{\\text{\"safe\",\"unsafe\"}\\}$ .", + "bbox": [ + 511, + 537, + 905, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "VQA Datasets. In our work, we evaluate on VQA benchmarks, many of which can be formulated as a classification task. (1) BLINK [18] contains many tasks that are intuitive for humans but complicated for multimodal models such as multi-view reasoning, and visual similarity comparison. Since potential answers in BLINK are multiple choice, the class labels would be given as $\\mathcal{C} = \\{\\text{\"A\", \"B\", \"C\", \"D\"}\\}$ (note: the number of labels depends on the possible number of options allowed for a task). (2) NaturalBench-VQA [44] is a compositional dataset collected from natural image-text corpora but validated with human filtering. Each sample of the dataset contains two questions on compositional", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2763", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8cdfe9004740d5cd677fd2358c3698958316fd4f461d85805df5c7df17f8b785.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelImage-Text TasksImage-Only Tasks
SafetyVQAI-T RetrievalClassification
MHaluVLGuardBLINKVizWizNB (T)NB (D)NB (G)NB-r (T)NB-r (D)NB-r (G)SCEuroSATPetsImagenet-IkFlowerCUB
CLIP-------41.845.023.235.364.088.196.192.897.8
SigLip-------54.554.931.242.763.998.397.695.898.4
GPT-4o†39.445.759.058.164.666.439.665.067.040.548.370.598.398.399.099.3
LLaVA-1.544.331.038.050.037.743.812.736.742.712.224.926.343.020.651.849.0
Instruct-BLIP44.018.538.734.520.224.24.019.521.31.14.720.456.011.018.233.8
LLaVA-OV-7B34.731.445.060.452.053.327.056.258.032.115.366.588.192.583.285.3
+4-shot-ICL25.035.038.947.847.650.422.248.249.631.516.147.163.949.063.860.6
+MTV37.332.944.561.156.258.030.758.159.533.428.765.588.595.983.285.6
+LoRA78.390.047.063.158.660.932.459.460.335.430.485.096.893.291.291.8
+SAVs80.894.351.866.160.362.335.172.773.053.137.686.797.097.599.697.5
+46.1+62.9+6.8+5.7+8.3+9.0+8.1+16.5+15.0+20.5+22.3+20.2+8.9+5.0+16.4+12.2
Qwen2-VL-7B24.026.943.368.353.856.628.560.261.935.624.954.792.684.493.793.2
+4-shot-ICL40.452.937.667.138.241.315.242.445.622.725.229.443.543.859.448.4
+MTV32.321.941.968.554.857.329.763.564.037.040.752.391.791.294.394.1
+LoRA84.887.746.370.855.357.428.865.266.140.438.472.998.496.497.195.0
+SAVs85.196.047.268.357.660.932.370.071.042.547.579.998.197.699.898.7
+61.1+69.1+3.9+0+3.8+4.3+3.8+9.8+9.1+13.8+22.6+25.2+5.5+13.2+6.1+5.5
", + "bbox": [ + 91, + 88, + 903, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Results evaluation on Safety, Visual Question Answering (VQA), Image-Text Retrieval (I-T Retrieval), and Image Classification benchmarks. The best result for each generative model is shown in bold and the second best in underline. We gray out additional baselines. We note that CLIP and SigLIP cannot be evaluating directly on tasks with interleaved image-text queries. † GPT-4o is a close-source model and as such, is shown just as an upperbound (since SAVs are not directly applicable). Key: NB - NaturalBench, SC - SUGARCREPE.", + "bbox": [ + 89, + 425, + 906, + 481 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "differences between two similar images, making Natural-Bench especially challenging for any existing VL models. The class labels are $\\mathcal{C} = \\{\\text{\"A\", \"B\"}\\}$ . As suggested in the paper, we evaluate \"question accuracy\" which awards one point if a model correctly answers a question for both images, \"image accuracy\" which awards a point when a model answers both questions for an image, and finally \"group accuracy\" awards one point when a model correctly answers all four pairs. (3) Vizwiz [20] is VQA dataset designed to progress research in vision systems to assist blind and vision-impaired individuals. The dataset was collected by asking blind people to take pictures and record questions about the image. Although typically a generative task, we reformulate Vizwiz as first a classification task distinguishing answerable and unanswerable questions followed by standard response generation. The class labels are $\\mathcal{C} = \\{\\text{\"answerable\", \"unanswerable\"}\\}$ .", + "bbox": [ + 88, + 507, + 482, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image-Text Retrieval. NaturalBench-Retrieval [44] and SUGARCREPE [28] both measure fine-grained semantic understanding in image-text pairs, with class labels $\\mathcal{C} = \\{\\text{\"Yes\",\"No\"}\\}$ indicating whether an image-text pair is correctly matched. While SUGARCREPE presents one image with two captions (original and altered), NaturalBench-Retrieval adds complexity by using two similar images with two corresponding captions, effectively eliminating lan", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "guage bias and requiring models to capture more nuanced visual-semantic relationships.", + "bbox": [ + 511, + 508, + 903, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image Classification. Image classification tasks evaluate a model's ability to categorize images into predefined classes. We evaluate on several standard classification benchmarks: EuroSAT [22] (satellite imagery for land use classification), Oxford-IIIT-Pets [75] (pet breed identification with visually similar categories), Flowers [71] (flower species recognition), Caltech Birds (CUB) [96] (fine-grained bird species classification), and ImageNet-1k [10] (general object recognition). For each dataset except ImageNet, we formulate the task as a 4-way multiple choice question with labels $\\mathcal{C} = \\{\\text{\"Class 1\", \"Class 2\", \"Class 3\", \"Class 4\"}\\}$ . Due to model oversaturation in simpler formulations, ImageNet-1k is formulated as a 16-way classification problem.", + "bbox": [ + 511, + 539, + 906, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Baselines", + "text_level": 1, + "bbox": [ + 511, + 742, + 622, + 757 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For our results, we utilized SAVs with 20 examples per label. We compared our method with multiple SOTA baselines, including GPT-4o [73]. As a closed-source model, GPT-4o is presented simply as a strong closed-source baseline. Furthermore, we present the results of classic LMMs, LLaVA-1.5 and InstructBLIP, which have been previously applied to image classification [111]. Zero-shot (ZS) baselines are implemented by querying the model directly and generating a response. For image-text retrieval, ZS uses the", + "bbox": [ + 511, + 763, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2764", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/05605116cb7f6c8d344467a0426d347ec8e9163f545fc879c8a595883b12cdb0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 87, + 354, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b1c9470e44c6ce7d0af80d2f8e8caffe3444b5bfc10907139ff66cb2f83db399.jpg", + "image_caption": [ + "Hallucination Top Head Att. Vec.", + "Figure 3. Head Visualization. On the top row, we show the top-20 attention head locations for a given task. All the heads are indexed by their (head, layer) position, and the selected heads are highlighted. On the bottom row, we visualize the attention vector of few-shot examples for the top head of the given class with t-SNE clustering [26]. Each point represents the embedding of an input sample." + ], + "image_footnote": [], + "bbox": [ + 125, + 329, + 352, + 518 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4712e0fc0c0a1557f5d1850a96de892b23ef1e18a91749bd57466faabd0d933b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 380, + 88, + 630, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dd94dbdbfae19ace29169120ccef96730682d57d58d77d2113727445158b6e17.jpg", + "image_caption": [ + "Relative Depth Top Head Att. Vec." + ], + "image_footnote": [], + "bbox": [ + 398, + 329, + 627, + 518 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/885ef9438ea4cda3ac22178c71e7606c33e55761b9be9b3d81f963995b6a0575.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 88, + 903, + 314 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d007f3842e214eaaf127dc025ec5b879ba4526d7ef4ca0bf10d8ed961a8477ab.jpg", + "image_caption": [ + "Image Cls. Top Head Att. Vec." + ], + "image_footnote": [], + "bbox": [ + 674, + 329, + 903, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SOTA generative scoring method VQAScore [57]. In addition to ZS, we also compare to several test-time and finetuning few-shot methods (all with the exact same sample complexity as SAVs). For instance, we compare to the current SOTA multimodal few-shot method, MTV [30] as well as doing 4-shot ICL and LoRA [29] finetuning for each task.", + "bbox": [ + 88, + 599, + 482, + 691 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Results", + "text_level": 1, + "bbox": [ + 89, + 709, + 184, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results are shown in Table 1. An advantage of our method is its adaptability to any VL classification task that has image, text, or interleaved image-text inputs. Thus, we apply SAVs to a wide range of tasks in safety, VQA, imagetext retrieval and image classification. Furthermore, our approach even significantly improves over SOTA ZS, few-shot, and finetuning methods. One interesting observation is that few-shot ICL consistently deteriorates performance. We hypothesize that instructioning tuning of LMMs disrupts the few-shot prompting capabilities gained during pretraining in favor of structured instructions as shown", + "bbox": [ + 88, + 734, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in [14, 86]. On the other hand, SAVs overcome this by directly operating on the internal activations. Finally, not only is our method successful across a wide-range of tasks, but it also improves on challenging visual perception and compositional reasoning tasks (e.g., BLINK and NaturalBench) that all VL models struggle with. Please refer to our supplementary section in Section 7 for mor results.", + "bbox": [ + 511, + 599, + 906, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Ablations", + "text_level": 1, + "bbox": [ + 511, + 722, + 624, + 736 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We perform a comprehensive ablation study of our method on MHaluBench, NaturalBench, and EuroSAT (see Table 2). For more ablations, please refer to Section 7.1 in the Supp. For all ablations, we use LLaVA-OneVision-7B.", + "bbox": [ + 511, + 747, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Varying number of examples. In Figure 4 (left), we examine the impact of varying the number of few-shot examples used in our method. Our primary results in Table 1 indicate that just 20 examples per label are necessary to yield state-of-the-art performance on a variety of VL classification tasks. This ablation shows that accuracy on these tasks", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2765", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4db5e75369ac156a903e916af66682b50f468d69d42a5bb067797b93b7fa1f98.jpg", + "image_caption": [ + "Figure 4. Scaling Property of SAVs. Performance of LLaVA-OneVision-7B + SAVs on varying number of few-shot examples per label (left). Performance of of LLaVA-OneVision-7B + SAVs on varying numbers of attention vectors used (right)." + ], + "image_footnote": [], + "bbox": [ + 106, + 97, + 491, + 276 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5033e9e9949bf23a4d849a305d38d697674b415c51e7364dbd5cbd3df9cb5afa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 98, + 890, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "can scale with increasing numbers of examples per label.", + "bbox": [ + 91, + 354, + 467, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Varying number of attention vectors. In order to extract attention vectors, we select a very sparse set of heads from hundreds. We vary the number of attention vectors selected in Step 2 of our method and demonstrate that just 20 vectors are enough to realize nearly all of the classification accuracy of our method. Results are shown in Figure 4 (right).", + "bbox": [ + 91, + 369, + 482, + 460 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "SAVs are flexible to different evaluation strategies. In our method, we leverage class centroid classification as the evaluation method for selecting sparse features. We view this flexibility of the sparsification method to be a key feature of our work. As such, we compare our class centroid classification approach to k-nearest neighbors (KNN) and linear probing. For linear probing, we train a lightweight MLP module for 20 epochs using the top heads' features. All methods make use of the same 20 examples per label for consistency. Our results in Table 3a show that class centroid classification outperforms both KNN classification and is comparable with linear probing.", + "bbox": [ + 91, + 460, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing heads vs. layers. Based on prior work and transformer-architecture intuition, we treat the attention vectors outputted by the heads as a viable set of features for classification tasks. We verify this intuition by comparing the performance of selecting 2 sparse layers to selecting sparse heads as feature maps for our tasks. As shown in Table 3b, head-based attention vectors outperform concatenated layer features on all three benchmarks.", + "bbox": [ + 89, + 642, + 482, + 762 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Additional Experiments", + "text_level": 1, + "bbox": [ + 89, + 772, + 310, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we present experiments that demonstrate additional properties and capabilities of SAVs, beyond its use as features for VL classification tasks. Additional experiments can be found in Section 7.2 of the Supplementary. For all experiments, we use LLaVA-OneVision-7B.", + "bbox": [ + 89, + 794, + 482, + 869 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualizing SAVs. SAVs are both an efficient and interpretable method for leveraging generative LMMs for VL", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "classification tasks. To emphasize this point, we show the selected heads for hallucination detection, relative depth, and image classification in the first row of Figure 3. The visualizations demonstrate both the sparsity and specificity of the SAVs that are used for each task. Unlike prompting and finetuning methods, our approach clearly outlines exactly where in the model's activation space informative attention vector features are extracted from. We furthermore show that the features extracted from these heads are useful for the given task in the second row of the figure, where we visualize the features outputted by the top selected head for each few-shot sample via t-SNE [26]. The clear clustering of examples of the same label indicates that even with a single head, high-quality features are being selected as SAVs.", + "bbox": [ + 511, + 354, + 903, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluating SAVs generalizability. Here, we ask whether SAVs extracted from one task, can generalize to another similar task. We utilize SAV heads from VLGuard to evaluate on MhaluBench and vice versa. We do a similar approach with LoRA, by applying LoRA finetuned on VLGuard to MHaluBench. Interestingly, our results in Table 2a show that SAVs generalize between tasks, while LoRA weights, as expected, overfit to the finetuned task.", + "bbox": [ + 511, + 566, + 903, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparing SAVs to CLIP/SigLIP on interleaved imagetext tasks. As discussed in Section 1, SAVs are fully multimodal features able to represent inputs that are image-only, text-only, and even interleaved image-text. This is something that is not possible to directly replicate with CLIP and SigLIP models which have separate image and text encoders. Nevertheless, we compare our method to both CLIP and SigLIP on tasks that require interleaved imagetext inputs. While SAVs can do this natively, we enable this comparison by concatenating the separate image and text features of CLIP and SigLIP in order to evaluate on MHaluBench and NaturalBench. We find that our method vastly outperforms concatenated CLIP and SigLIP features on both benchmarks as shown in Figure 2b. This result", + "bbox": [ + 511, + 688, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2766", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e5fc9b2cae56711f480565a887a0833d39d3b34481188fa311460bb29776dd61.jpg", + "table_caption": [ + "(a) Generalization" + ], + "table_footnote": [], + "table_body": "
MHBVLG
Zero-Shot34.731.4
VLG LoRA34.190.0
VLG SAV55.394.3
", + "bbox": [ + 96, + 99, + 336, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/303d1f8de34fb67ccc42d68ed35051b4db279a26dbc9f5eacf0b16a74d4aeebd.jpg", + "table_caption": [ + "(b) Interleaved Tasks" + ], + "table_footnote": [], + "table_body": "
MHBNB
CLIP51.91.2
SigLIP48.61.2
SAVs80.835.1
", + "bbox": [ + 370, + 99, + 575, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c6aa54d27ee23fa49161c55874ffaa1453f85f1e37cb8626cc3d4cb1191c0de0.jpg", + "table_caption": [ + "(c) SAVs vs Few-Shot SigLIP" + ], + "table_footnote": [], + "table_body": "
NBES
SigLIP Few-shot Mean1.288.2
SigLIP Few-shot Cross-Modal6.886.4
SAVs35.186.7
", + "bbox": [ + 609, + 99, + 905, + 183 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ab6d622f49bdd0496b14f6f4f56a01132e5f41d868d757a7890ae7a0b04266da.jpg", + "table_caption": [ + "Table 2. Additional SAVs Experiments. We (a) demonstrate the generalization of SAV heads to similar tasks, (b) show the effectiveness of SAV for tasks with interleaved image-text inputs, and (c) compare SAVs with few-shot formulations of CLIP and SigLIP.", + "(a) Classification Methods" + ], + "table_footnote": [], + "table_body": "
MHBNBES
Class Centroid80.835.186.7
KNN53.011.078.1
Linear Probe82.532.983.1
", + "bbox": [ + 96, + 251, + 455, + 325 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/08629bc2dd45de0c1d409823c6a452668f868b703631a723f94d2d2872350fab.jpg", + "table_caption": [ + "(b) Sparse Configurations" + ], + "table_footnote": [], + "table_body": "
MHBNBES
Sparse Heads80.835.186.7
Sparse Layers79.028.481.8
", + "bbox": [ + 514, + 252, + 870, + 324 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. SAVs Ablations. We perform several ablations to identify the important aspects of our method that contribute to its effectiveness. In particular, we compare the effectiveness of (a) different classification methods and (b) head feature sparsification versus layer feature sparsification. Note: MHB represents MHaluBench, NB represents NaturalBench Group Score, and ES represents EuroSAT. For more ablations, please refer to Section 7.1 in the Supplementary.", + "bbox": [ + 89, + 334, + 906, + 391 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "demonstrates the adaptability of our method to any VL classification regardless of the input's modality.", + "bbox": [ + 88, + 417, + 480, + 446 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparing SAVs to few-shot SigCLIP. As SAVs are extracted with few-shot examples, we compare our method to an analogous version of few-shot SigLIP. However, because SigLIP cannot be directly made few-shot, we adapt SigLIP as a few-shot class centroid classifier. One method used is the current SOTA few-shot classification method for encoder models Cross-Modal Adaptation [56]. We apply a second method where we aggregate CLIP/SigLIP embeddings into a mean embedding for each label. Then, just as in SAVs, we perform class centroid classification for each query using our set of mean SigLIP embeddings. We note that for image classification like EuroSAT, only image embeddings are needed, but for MHaluBench, multimodal embeddings are necessary. SAVs are inherently multimodal and so can be flexibly applied to both, but CLIP/SigLIP only have image-only or text-only embeddings. To overcome this, we use SigLIP image features for EuroSAT and concatenate image and text features for MHaluBench. Interestingly, while SiGLIP is comparable to SAVs on EuroSAT, our method vastly outperforms SigLIP in the few-shot setting for MHaluBench, suggesting generalizability of our method for a variety of multimodal tasks that CLIP-like struggle with. Our results are shown in Table 2c.", + "bbox": [ + 91, + 449, + 483, + 796 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 89, + 813, + 209, + 829 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our research demonstrates the effectiveness of extracting Sparse Attention Vectors (SAVs) from the heads of an LMM and utilizing them directly for vision-language classification. Our method stands out by using only few-shot exam", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "plies per label and only less than $1\\%$ of the heads to outperform zero-shot, few-shot, and fine-tuned baselines on a variety of image-text and image-only tasks. In addition, SAVs allows generative LMMs to close the gap with closed-source GPT-4o while also being an interpretable method that can generalize to similar tasks. Our ablations reveal the flexibility of using any classification method as a sparsification method for attention vectors and also shows that features are found as outputs of heads rather than layers. Overall, these results show that SAVs is a lightweight, performant, and generalizable method for extending generative LMMs' multimodal classification abilities. We are encouraged by the outcomes, and anticipate many directions for future work. In addition to methodological improvements, we look forward to the application of SAVs as features for multimodal retrieval, data compression, or more generally as a distilled representation for downstream models.", + "bbox": [ + 511, + 417, + 906, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 511, + 691, + 635, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Sparse Attention Vectors are a significant step in generalizing the capabilities of generative LMMs to classification tasks. Nevertheless, it is valuable to consider certain limitations of our approach. SAVs are a method that requires access to the model's internal architecture and so may not be directly applicable to closed-source models like GPT-4 [73] and Gemini [89, 90]. Additionally, some tasks like image-text retrieval [28, 91] can benefit from more fine-grained confidence metrics attached to each label than proportion of voting heads per label. These challenges prompt future work in these directions as well as exciting questions about how to use SAVs as feature embeddings for other tasks.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2767", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements.", + "text_level": 1, + "bbox": [ + 91, + 90, + 235, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We would like to thank Abrar Anwar and Tyler Bonnen for helpful feedback and discussions. This project was supported in part by DoD, including PTG and/or LwLL programs, as well as BAIR's industrial alliance programs.", + "bbox": [ + 89, + 109, + 485, + 172 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 184, + 187, + 199 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] S. R. Bowman A. Williams, N. Nangia. A broad-coverage challenge corpus for sentence understanding through inference. ArXiv, 2017. 2", + "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1", + "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv, abs/2308.12966, 2023. 1", + "[4] Pierre Baldi. Autoencoders, unsupervised learning, and deep architectures. In ICML 2011 Unsupervised and Transfer Learning Workshop, 2011. 2", + "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3", + "[6] Matyas Bohacek and Michal Bravansky. When XGBoost outperforms GPT-4 on text classification: A case study. In Trustworthy Natural Language Processing (TrustNLP) 4th Workshop, pages 51–60, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 2", + "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. *ArXiv*, abs/2005.14165, 2020. 2", + "[8] Martin Juan José E Bucher and Marco Martini. Finetuned 'small' llms (still) significantly outperform zero-shot generative ai models in text classification. ArXiv, abs/2406.08660, 2024. 1, 2", + "[9] Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Jinjie Gu, and Huajun Chen. Unified hallucination detection for multimodal large language models. arXiv preprint arXiv:2402.03190, 2024. 4, 3", + "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5", + "[11] Xiang Deng, Vasilisa Bashlovkina, Feng Han, Simon Baumgartner, and Michael Bendersky. Llms to the moon?" + ], + "bbox": [ + 101, + 209, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reddit market sentiment analysis with large language models. Companion Proceedings of the ACM Web Conference 2023, 2023. 2", + "[12] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11162-11173, 2021. 2", + "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805, 2019. 1, 2", + "[14] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. ArXiv, abs/2403.12736, 2024. 6", + "[15] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Herv'e J'egou. Training vision transformers for image retrieval. ArXiv, abs/2102.05644, 2021. 1, 2", + "[16] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2", + "[17] Evelina Fedorenko, Michael K. Behr, and Nancy Kanwisher. Functional specificity for high-level linguistic processing in the human brain. Proceedings of the National Academy of Sciences, 108(39):16428-16433, 2011. 2", + "[18] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 2, 4, 3", + "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2", + "[20] Danna Gurari, Qing Li, Abigale Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 5", + "[21] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2", + "[22] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5", + "[23] Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors. ArXiv, abs/2310.15916, 2023. 2", + "[24] Roee Hendel, Mor Geva, and Amir Globerson. Incontext learning creates task vectors. arXiv preprint arXiv:2310.15916, 2023. 2", + "[25] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson." + ], + "bbox": [ + 524, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2768", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Incorporating structured representations into pretrained vision $\\&$ language models using scene graphs. In 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2", + "[26] L. Hinton G, van der Maaten. Visualizing data using t-sne. Journal of machine learning research, 2008. 6, 7", + "[27] Alberto Hojel, Yutong Bai, Trevor Darrell, Amir Globerson, and Amir Bar. Finding visual task vectors. In European Conference on Computer Vision (ECCV), pages 257-273. Springer, 2025. 2", + "[28] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. ArXiv, abs/2306.14610, 2023. 5, 8", + "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6", + "[30] Brandon Huang, Chancharik Mitra, Assaf Arbelle, Leonid Karlinsky, Trevor Darrell, and Roei Herzig. Multimodal task vectors enable many-shot multimodal in-context learning. arXiv preprint arXiv:2406.15334, 2024. 2, 6, 3", + "[31] Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2", + "[32] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6693-6702, 2019. 1", + "[33] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 2", + "[34] Ting Jiang, Shaohan Huang, Zi qiang Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Liangjie Zhang, and Qi Zhang. Promptbert: Improving bert sentence embeddings with prompts. In Conference on Empirical Methods in Natural Language Processing, 2022. 2", + "[35] Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. Scaling sentence embeddings with large language models. ArXiv, abs/2307.16645, 2023. 2", + "[36] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. ArXiv, abs/2407.12580, 2024. 2", + "[37] Nancy Kanwisher. Domain specificity in face perception. Nature neuroscience, 3(8):759-763, 2000. 2", + "[38] Sahar Kazemzadeh, Vicente Ordonez, Marc andre Matten, and Tamara L. Berg. Referitgame: Referring to objects in photographs of natural scenes. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 1", + "[39] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 2" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2", + "[41] Yibin Lei, Di Wu, Tianyi Zhou, Tao Shen, Yu Cao, Chongyang Tao, and Andrew Yates. Meta-task prompting elicits embeddings from large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 2", + "[42] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2", + "[43] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv, abs/2307.16125, 2023. 1", + "[44] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Natural-bench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 2, 4, 5", + "[45] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, abs/2408.03326, 2024. 4", + "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2", + "[47] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2", + "[48] Shiyu Li, Yang Tang, Shizhe Chen, and Xi Chen. Conan-embedding: General text embedding with more and better negative samples. ArXiv, abs/2408.15710, 2024. 2", + "[49] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2", + "[50] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23390-23400, 2022. 2", + "[51] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2", + "[52] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2", + "[53] Jiacheng Lin, Kun Qian, Haoyu Han, Nurendra Choudhary, Tianxin Wei, Zhongruo Wang, Sahika Genc, Edward W" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2769", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Huang, Sheng Wang, Karthik Subbian, et al. Unleashing the power of lms as multi-modal encoders for text and graph-structured data. arXiv preprint arXiv:2410.11235, 2024. 2", + "[54] Tsung-Yi Lin, M. Maire, Serge J. Belongie, James Hays, P. Perona, D. Ramanan, Piotr Dollar, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), 2014. 1, 3", + "[55] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 2", + "[56] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 8", + "[57] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2024. 6", + "[58] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2025. 2", + "[59] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1", + "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3", + "[61] Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera, Luca Zancato, and Stefan O Soatto. Meaning representations from trajectories in autoregressive models. ArXiv, abs/2310.18348, 2023. 2", + "[62] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 2020. 2", + "[63] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 33:5081–5092, 2020. 2", + "[64] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. ArXiv, abs/2405.20797, 2024. 2", + "[65] Junhua Mao, Jonathan Huang, Alexander Toshev, OanaMaria Camburu, Alan Loddon Yuille, and Kevin P. Murphy. Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2015. 1", + "[66] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 2", + "[67] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain of thought prompting" + ], + "bbox": [ + 99, + 90, + 482, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for large multimodal models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[68] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. In Conference of the European Chapter of the Association for Computational Linguistics, 2022. 2", + "[69] Jianmo Ni, Gustavo Hernández Abrego, Noah Constant, Ji Ma, Keith B. Hall, Daniel Matthew Cer, and Yinfei Yang. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. ArXiv, abs/2108.08877, 2021. 2", + "[70] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith B. Hall, Ming-Wei Chang, and Yinfei Yang. Large dual encoders are generalizable retrievers. ArXiv, abs/2112.07899, 2021. 2", + "[71] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5", + "[72] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 2", + "[73] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 1, 5, 8, 3", + "[74] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 2", + "[75] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5, 6", + "[76] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 3", + "[77] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2", + "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2", + "[79] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a" + ], + "bbox": [ + 522, + 92, + 903, + 901 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2770", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reward model. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2", + "[81] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 1, 2", + "[82] David E Rumelhart, James L McClelland, PDP Research Group, et al. Parallel distributed processing, volume 1: Explorations in the microstructure of cognition: Foundations. The MIT press, 1986. 2", + "[83] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815-823, 2015. 1, 2", + "[84] Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014. 2", + "[85] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8309-8318, 2019. 3", + "[86] Aaditya K. Singh, Stephanie C.Y. Chan, Ted Moskovitz, Erin Grant, Andrew M. Saxe, and Felix Hill. The transient nature of emergent in-context learning in transformers. In Advances in neural information processing systems (NeurIPS), 2023. 6", + "[87] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2013. 2", + "[88] Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. Text classification via large language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2", + "[89] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. ArXiv, abs/2403.05530, 2024. 8", + "[90] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 8", + "[91] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 8", + "[92] Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau. Function vectors in large language models. ArXiv, abs/2310.15213, 2023. 2" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[93] Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. Function vectors in large language models. arXiv preprint arXiv:2310.15213, 2023. 2", + "[94] Matthew A. Turk and Alex Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3:71-86, 1991. 1, 2", + "[95] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. ArXiv, abs/1711.00937, 2017. 2", + "[96] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5", + "[97] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. ArXiv, abs/2401.00368, 2023. 2", + "[98] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Ke-Yang Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. ArXiv, abs/2409.12191, 2024. 4", + "[99] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. arXiv preprint arXiv:2401.11206, 2024. 2", + "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3", + "[101] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. Internvideo2: Scaling video foundation models for multimodal video understanding. ArXiv, abs/2403.15377, 2024. 2", + "[102] Zhiqiang Wang, Yiran Pang, and Yanbin Lin. Large language models are zero-shot text classifiers. ArXiv, abs/2312.01044, 2023. 2", + "[103] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. ArXiv, abs/2109.01652, 2021. 2", + "[104] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 2", + "[105] Dean Wyatte, Fatemeh Tahmasbi, Ming Li, and Thomas Markovich. Scaling laws for discriminative classification in large language models. ArXiv, abs/2405.15765, 2024. 2", + "[106] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "2771", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 3", + "[107] Peter Young, Alice Lai, Micah Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1", + "[108] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 2", + "[109] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2", + "[110] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and C. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine Learning in Health Care, 2020. 2", + "[111] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 1, 2, 5", + "[112] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. ArXiv, abs/2309.07915, 2023. 2", + "[113] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3", + "[114] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 2, 4" + ], + "bbox": [ + 91, + 90, + 482, + 614 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "2772", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_model.json b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5144df2f5d255e9a85da0d872dfedf6422215452 --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_model.json @@ -0,0 +1,3042 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.115, + 0.13, + 0.885, + 0.175 + ], + "angle": 0, + "content": "Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.203, + 0.861, + 0.242 + ], + "angle": 0, + "content": "Chancharik Mitra\\(^{1*}\\) Brandon Huang\\(^{2*}\\) Tianning Chai\\(^{2}\\) Zhiqiu Lin\\(^{1}\\) Assaf Arbelle\\(^{3}\\) Rogerio Feris\\(^{4}\\) Leonid Karlinsky\\(^{4}\\) Trevor Darrell\\(^{2}\\) Deva Ramanan\\(^{1}\\) Roei Herzig\\(^{2,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.256, + 0.765, + 0.293 + ], + "angle": 0, + "content": "\\(^{1}\\)Carnegie Mellon University \\(^{2}\\)University of California, Berkeley \\(^{3}\\)IBM Research \\(^{4}\\)MIT-IBM Watson AI Lab" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.678 + ], + "angle": 0, + "content": "Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs)—a finetuning-free method that leverages sparse attention head activations (fewer than \\(5\\%\\) of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.709, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Generative Large Multimodal Models (LMMs) such as GPT-4V [73], LLaVA [59, 60], and QwenVL [3] demonstrate state-of-the-art performance on open-ended vision-language (VL) tasks like image captioning [54, 107], visual question answering [2, 32, 43], and language grounding [38, 65]. However, despite their remarkable performance on generative tasks, these models struggle on vision-language classification tasks, where responses are a discrete set of labels [8, 111]. Indeed, LMMs with billions of parameters and trained on trillions more tokens of data underperform smaller encoder VLMs like CLIP and SigLIP [8, 111]" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.326, + 0.707, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.326, + 0.902, + 0.603 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.615, + 0.907, + 0.7 + ], + "angle": 0, + "content": "Figure 1. Sparse Attention Vectors (SAVs) Overview. We develop a method for extracting features from a generative LMM without finetuning. We first extract a sparse set of attention vectors for each task given a set of few-shot examples, and then, we utilize these attention vectors directly as features for downstream vision-language classification tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.731, + 0.907, + 0.853 + ], + "angle": 0, + "content": "and even classical machine learning methods [6] on image classification tasks. One reason that CLIP-like models excel at classification is the ease of extracting visual or textual features from modality-specific encoders. However, such models cannot process joint vision-language inputs, as LMMs can. Our goal is to extract compact features from generative LMMs, enabling downstream discrete labeling tasks such as classification and multiple-choice VQA." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Feature extraction is a well-explored field in both vision-only [15, 83, 94] and language-only encoder models [13, 81], but such is not the case with generative models. Most" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2760" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.243 + ], + "angle": 0, + "content": "current methods for extracting features from generative models require carefully constructed prompts [41, 67, 104], specialized architectures [51], few-shot prompting [7, 11, 112], and finetuning [8, 64, 111]. Recent work [111] shows that prompting approaches fail to close the gap with encoder VLMs, while finetuning requires training-scale data for every unseen task, which is inefficient. However, as we observe in our work, generative models still offer the promise of more flexible, truly multimodal features as compared to modality-specific features from CLIP-like models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.251, + 0.483, + 0.522 + ], + "angle": 0, + "content": "A source of inspiration for our method is long-standing work in neuroscience that suggests certain areas of the brain are reserved for specific tasks [17, 37] (i.e. functional specificity). Motivated by this idea, we refer to recent interpretability research that has focused on identifying specific heads in transformer-based models that correspond to particular tasks [72]. The most prominent of these methods is a line of work that looks to enhance vision-language capabilities using task vectors [24, 27, 30, 93], which are compact implicit representations of tasks encoded in the activations of a transformer model. While promising, these methods ultimately use these representations to augment a model's generative capabilities. On the other hand, we seek to use feature representations directly as classifiers. Nevertheless, this intuition from interpretability informs our work on Sparse Attention Vectors (SAVs), which are sparse features in an LMMs activation space that can be directly exploited for few-shot VL classification." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.483, + 0.743 + ], + "angle": 0, + "content": "Our method has three steps: First, we extract features (i.e., attention vectors) from the output of each head of the LMM for some few-shot labeled examples (\\(\\approx 20\\) per label). Second, we average these attention vectors over the examples in each class and evaluate their accuracy as centroids in a class centroid classifier. We then select the top 20 heads by classification accuracy as our SAVs. In this way, we identify a very sparse set of attention vectors (less than \\(5\\%\\) of the total number of heads) that can be used for discriminative tasks. Finally, we perform inference on the given task by doing a majority vote across this sparse set of attention vectors for each new query. This approach requires only few-shot examples at test-time to extract effective multimodal embeddings. An overview is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We summarize our main contributions as follows: (i) We introduce a novel method that yields a sparse set of attention vectors to serve as highly effective features for individual VL classification tasks; (ii) Our method surpasses zero-shot, few-shot, and LoRA fine-tuned baselines across multiple tasks (\\(+7\\%\\) improvement on average over LoRA on challenging benchmarks like BLINK [18], VLGuard [114], and NaturalBench [44]); (iii) We establish several advantageous properties of our approach, including strong generalization capabilities and favorable scaling characteristics." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.664, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.116, + 0.905, + 0.373 + ], + "angle": 0, + "content": "Controllable Generation for Classification. Controllable text generation in LMMs guides model outputs to meet specific constraints. For classification tasks with generative LMMs, several approaches exist: test-time hard prompting [7, 104] uses prompt engineering or few-shot examples to elicit class label outputs [11, 88, 102, 105, 111]; direct probability analysis of generated class labels [55, 58] for image-text retrieval; and soft prompting methods that finetune learnable tokens [42, 49]. Other techniques include instruction finetuning [103] on labeled data [6, 111] and preference modeling like DPO and RLHF [16, 74, 79]. Our method, however, is finetuning-free and directly selects class labels without preference data. Most related work shows that internal transformer representations called task vectors [23, 27, 30, 92] can encapsulate ICL example tasks. Beyond previous approaches, we use a sparse set of attention vectors directly as VL classification task task features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.374, + 0.905, + 0.525 + ], + "angle": 0, + "content": "Vision-Language Features. Feature extraction seeks useful representations for diverse downstream tasks. Early embedding techniques including autoencoders [4, 39, 62, 63, 82], Word2Vec [66] and GloVe [77] transformed inputs into vector representations, followed by advances in NLP [13, 19, 68, 81] and vision [15, 83, 94]. Recent methods like CLIP and SigCLIP [46, 47, 50, 78, 108, 109] explore multimodal correlations through contrastive learning or sigmoid loss. These representations offer flexibility across tasks [12, 25, 33, 40, 80, 95] and domains [53, 101, 110]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.905, + 0.692 + ], + "angle": 0, + "content": "Interestingly, extracting features from generative models poses unique challenges in identifying optimal extraction points. Some approaches finetune encoder VLMs on LLM-generated data [48, 97] or finetune LMMs directly on specific tasks [31, 111]. More efficient methods finetune encoder-decoder representations for better alignment [36, 64, 69, 70]. Finetuning-free approaches use distillation prompts to extract representations from model weights or activations [34, 35, 41, 61], while other methods employ mixture-of-experts [52], expert models [99], or embedding reranking [21]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.693, + 0.905, + 0.798 + ], + "angle": 0, + "content": "Thus, current SOTA faces the following challenges: (1) modality-specific rather than multimodal features, (2) requiring finetuning, (3) limited flexibility due to specialized prompts, and (4) dependence on multiple models. Our approach provides effective multimodal embeddings without any gradient-based finetuning and flexibly apply to various VL classification tasks without additional models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.815, + 0.612, + 0.83 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we outline our approach for using sparse attention vectors from the activation space of a transformer-based large multimodal model (LMM) as features for any VL classification task. The method consists of three main" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.513, + 0.957 + ], + "angle": 0, + "content": "2761" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.089, + 0.297, + 0.101 + ], + "angle": 0, + "content": "1. Extract Attention Vectors" + }, + { + "type": "image", + "bbox": [ + 0.107, + 0.103, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.387, + 0.337, + 0.6, + 0.351 + ], + "angle": 0, + "content": "3. SAVs for Discriminative Tasks:" + }, + { + "type": "image", + "bbox": [ + 0.117, + 0.352, + 0.862, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.429, + 0.908, + 0.499 + ], + "angle": 0, + "content": "Figure 2. Sparse Attention Vectors (SAVs) Detailed View. Our method is broken into the following three parts: (1) Given a set of few shot examples to be classified by a frozen LMM, we extract attention vectors across all heads and layers. (2) These attention vectors are averaged across the set of examples for each class. For each head, we use these mean attention vectors as centroid classifiers, which are then used to select a sparse set of \\( k \\) heads with the highest classification accuracy. (3) Finally, we use these sparse attention vectors to directly classify new inputs via majority vote (denoted by the \"Classifier\" block)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.525, + 0.482, + 0.661 + ], + "angle": 0, + "content": "steps: (i) extracting the attention vectors from all attention heads in the model, (ii) identifying a sparse set of vectors based on their ability to consistently return the correct label for some support set of examples, and (iii) using these sparse features to classify new queries. We begin with a formal description of the transformer decoder LLM and its attention mechanism, followed by the detailed methodology for sparse attention vector selection and classification. A detailed view of our method is shown in Figure 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.232, + 0.701 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.483, + 0.806 + ], + "angle": 0, + "content": "A transformer-based large language model (LLM) with \\( L \\) layers and \\( H \\) attention heads per layer processes input sequences through multi-head self-attention mechanisms. Each layer combines multiple attention heads to capture different aspects of the input sequence, followed by feedforward networks for further processing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Multi-Head Attention. Let \\( x = \\{x_{1},x_{2},\\ldots ,x_{T}\\} \\) represent a sequence of input tokens, where \\( x_{i} \\) is the \\( i^{\\text{th}} \\) token. For each layer \\( l\\in \\{1,\\dots ,L\\} \\), the input sequence is projected into queries, keys, and values for each attention head \\( m\\in \\{1,\\dots ,H\\} \\). Each head performs the following scaled dot-product attention:" + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.089, + 0.757, + 0.102 + ], + "angle": 0, + "content": "2. Find Sparse Attention Vectors (SAVs):" + }, + { + "type": "image_caption", + "bbox": [ + 0.497, + 0.113, + 0.845, + 0.126 + ], + "angle": 0, + "content": "Select Sparse Attention Vectors" + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.131, + 0.905, + 0.333 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.54, + 0.818, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathbf {h} _ {l} ^ {m} \\left(x _ {i}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {m}}}\\right) V\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.649 + ], + "angle": 0, + "content": "where \\( Q \\), \\( K \\), and \\( V \\) are the query, key, and value matrices respectively, and the dimensionality of each head \\( d_{m} \\) which is given by \\( \\frac{d}{H} \\) (the embedding dimension divided by the number of heads). We denote \\( \\mathbf{h}_l^m(x_i) \\) as an attention vector for head \\( m \\) in layer \\( l \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.906, + 0.68 + ], + "angle": 0, + "content": "The outputs of all heads are concatenated and projected to form the layer output:" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.689, + 0.878, + 0.709 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M u l t i H e a d} \\left(x _ {i}\\right) = \\operatorname {C o n c a t} \\left(\\mathbf {h} _ {l} ^ {1} \\left(x _ {i}\\right), \\dots , \\mathbf {h} _ {l} ^ {H} \\left(x _ {i}\\right)\\right) W ^ {O}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.796, + 0.733 + ], + "angle": 0, + "content": "where \\( W^{O} \\) is the output projection matrix." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.839 + ], + "angle": 0, + "content": "In our work, we look to leverage attention vectors for the purpose of vision-language classification tasks. Specifically, the attention vectors are used as latent representations of the inputs to both find attention heads in an LMM suited for a classification task and then perform downstream inference using those selected attention heads. We describe our method in detail in the sections that follow." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.742, + 0.865 + ], + "angle": 0, + "content": "3.2. Sparse Attention Vectors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Our key insight is that within the many attention heads and transformer layers of an LMM, there exists a sparse subset" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2762" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "that can serve as effective features for vision-language classification tasks. We present a three-step method to identify and utilize these features to build lightweight classifiers." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.137, + 0.483, + 0.243 + ], + "angle": 0, + "content": "Step 1: Extracting Attention Vectors. Given a frozen LMM and few-shot examples of sequence-label pairs \\(\\{(x_1, y_1), (x_2, y_2), \\ldots, (x_N, y_N)\\}\\) we first extract the attention vectors for each sequence \\(x_i\\). Specifically, we compute the attention vector \\(\\mathbf{h}_l^m(x_i^T)\\) for head \\(m\\) from layer \\(l\\) for the final token \\(x_i^T\\). This yields a set of attention vectors \\(\\{\\mathbf{h}_l^m(x_i^T) \\mid i = 1, \\dots, N\\}\\) for each head \\(m\\) and layer \\(l\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.317 + ], + "angle": 0, + "content": "Step 2: Identifying Relevant Vectors. The central question is how to identify which attention vectors are naturally suited for the classification task at hand. We evaluate each vector's ability as a local classifier by computing its performance under a nearest class centroid classifier." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.349 + ], + "angle": 0, + "content": "Specifically, for each class \\( c \\in \\mathcal{C} \\), compute its centroid (or mean) attention vector across the few shot examples:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.357, + 0.382, + 0.395 + ], + "angle": 0, + "content": "\\[\n\\mu_ {c} ^ {l, m} = \\frac {1}{| N _ {c} |} \\sum_ {j: y _ {j} = c} \\mathbf {h} _ {l} ^ {m} (x _ {j} ^ {T})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.45 + ], + "angle": 0, + "content": "where \\( N_{c} = \\{j:y_{j} = c\\} \\) is the set of indices of examples with label \\( c \\). For each input \\( x_{i} \\), we compute its cosine similarity to each class centroid head:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.458, + 0.432, + 0.496 + ], + "angle": 0, + "content": "\\[\ns _ {l, m} \\left(x _ {i}, c\\right) = \\frac {\\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\cdot \\mu_ {c} ^ {l , m}}{\\left\\| \\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\right\\| \\left\\| \\mu_ {c} ^ {l , m} \\right\\|}, \\quad \\forall c \\in \\mathcal {C}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.504, + 0.483, + 0.533 + ], + "angle": 0, + "content": "Next, we measure the classification accuracy of each head (i.e. local classifier):" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.542, + 0.382, + 0.583 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s c o r e} (l, m) = \\sum_ {i = 1} ^ {N} \\mathbf {1} [ \\hat {y} = y _ {i} ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.483, + 0.631 + ], + "angle": 0, + "content": "where the nearest class centroid label is given as \\(\\hat{y} = \\operatorname*{argmax}_{c\\in \\mathcal{C}}s_{l,m}(x_i,c)\\), and \\(\\mathbf{1}[\\cdot ]\\) is the indicator function that" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.631, + 0.482, + 0.644 + ], + "angle": 0, + "content": "evaluates to 1 when the condition is true (and 0 otherwise)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.424, + 0.66 + ], + "angle": 0, + "content": "We denote the set of \\(k\\) top-scoring heads as \\(\\mathcal{H}_{\\mathrm{SAV}}\\):" + }, + { + "type": "text", + "bbox": [ + 0.098, + 0.669, + 0.475, + 0.687 + ], + "angle": 0, + "content": "\\(\\mathcal{H}_{\\mathrm{SAV}} = \\{(l,m)\\mid \\mathrm{score}(l,m)\\) is among \\(k\\) highest scores}" + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.696, + 0.458, + 0.71 + ], + "angle": 0, + "content": "Step 3: Classification with Sparse Attention Vectors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.772 + ], + "angle": 0, + "content": "Given a query sequence \\(Q\\) to classify, we leverage our sparse set of heads \\(\\mathcal{H}_{\\mathrm{SAV}}\\) for prediction. For each head \\((l,m)\\in \\mathcal{H}_{\\mathrm{SAV}}\\), we compute the class centroid \\(\\mu_c^{l,m}\\) closest to the query as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.78, + 0.384, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\hat{y}_{l,m} = \\operatorname *{arg max}_{c\\in \\mathcal{C}}s_{l,m}(Q^{T},c)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.815, + 0.483, + 0.861 + ], + "angle": 0, + "content": "where \\( s_{l,m}(\\cdot, \\cdot) \\) is defined as in Step 2. Our final class prediction is that of a global classifier that counts the majority vote across all local classifiers (heads) in \\( \\mathcal{H}_{\\mathrm{SAV}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.87, + 0.393, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\underset {y \\in \\mathcal {C}} {\\arg \\max } \\sum_ {(l, m) \\in \\mathcal {H} _ {\\mathrm {S A V}}} \\mathbf {1} [ \\hat {y} _ {l, m} = y ]\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "This approach reveals a surprising capability of LMMs: with just a few carefully selected attention heads (\\(|\\mathcal{H}_{\\mathrm{SAV}}| \\ll LH\\)), we can transform a generative language model into a lightweight vision-language classifier. This finding suggests that classification-relevant features naturally emerge within specific attention heads during model pretraining. In theory, one could replace our local and global classifiers with more complex variants (e.g. linear models, MLPs), something we explore in 3a." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.24, + 0.631, + 0.256 + ], + "angle": 0, + "content": "4. Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.266, + 0.905, + 0.357 + ], + "angle": 0, + "content": "We apply SAVs to two state-of-the-art LMMs—LLaVA-OneVision [45] and Qwen2-VL [98]. We also do a rigorous comparison of our method to strong few-shot and finetuning baselines on a variety of different image-text and image-only vision-language classification tasks covering safety, VQA, image-text retrieval, and simple classification." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.366, + 0.731, + 0.381 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.906, + 0.507 + ], + "angle": 0, + "content": "We implemented our approach in PyTorch [76]. We use the official implementations of each model, and all of our experiments can be run on a single NVIDIA A100 GPU. We use the default hyperparameters provided by the original models' (LLaVA-OV and Qwen2-VL) codebases. These hyperparameters are detailed in Table 8. More details of the implementation is included in the supplementary in Section 8." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.517, + 0.647, + 0.531 + ], + "angle": 0, + "content": "4.2. Benchmarks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Safety. (1) LMM-Halucination [9] is a dataset which evaluates the hallucinations of the models when answering multimodal tasks. We report the raw classification accuracy of our method. Thus, the set of class labels for this task is given by \\(\\mathcal{C} = \\{\\text{\"hallucinating\",\"not hallucinating\"}\\}\\). (2) VLGuard [114] is a dataset focusing on vision-language safety which identifies 4 main categories of harmful content: Privacy, Risky Behavior, Deception and Hateful Speech. VLGuard proposes Attack Success Rate (ASR) for evaluating unsafe inputs and Helpfulness for evaluating safe inputs. We simply reformat it as a classification task, where the set of class labels is given by \\(\\mathcal{C} = \\{\\text{\"safe\",\"unsafe\"}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "VQA Datasets. In our work, we evaluate on VQA benchmarks, many of which can be formulated as a classification task. (1) BLINK [18] contains many tasks that are intuitive for humans but complicated for multimodal models such as multi-view reasoning, and visual similarity comparison. Since potential answers in BLINK are multiple choice, the class labels would be given as \\(\\mathcal{C} = \\{\\text{\"A\", \"B\", \"C\", \"D\"}\\}\\) (note: the number of labels depends on the possible number of options allowed for a task). (2) NaturalBench-VQA [44] is a compositional dataset collected from natural image-text corpora but validated with human filtering. Each sample of the dataset contains two questions on compositional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2763" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.415 + ], + "angle": 0, + "content": "
ModelImage-Text TasksImage-Only Tasks
SafetyVQAI-T RetrievalClassification
MHaluVLGuardBLINKVizWizNB (T)NB (D)NB (G)NB-r (T)NB-r (D)NB-r (G)SCEuroSATPetsImagenet-IkFlowerCUB
CLIP-------41.845.023.235.364.088.196.192.897.8
SigLip-------54.554.931.242.763.998.397.695.898.4
GPT-4o†39.445.759.058.164.666.439.665.067.040.548.370.598.398.399.099.3
LLaVA-1.544.331.038.050.037.743.812.736.742.712.224.926.343.020.651.849.0
Instruct-BLIP44.018.538.734.520.224.24.019.521.31.14.720.456.011.018.233.8
LLaVA-OV-7B34.731.445.060.452.053.327.056.258.032.115.366.588.192.583.285.3
+4-shot-ICL25.035.038.947.847.650.422.248.249.631.516.147.163.949.063.860.6
+MTV37.332.944.561.156.258.030.758.159.533.428.765.588.595.983.285.6
+LoRA78.390.047.063.158.660.932.459.460.335.430.485.096.893.291.291.8
+SAVs80.894.351.866.160.362.335.172.773.053.137.686.797.097.599.697.5
+46.1+62.9+6.8+5.7+8.3+9.0+8.1+16.5+15.0+20.5+22.3+20.2+8.9+5.0+16.4+12.2
Qwen2-VL-7B24.026.943.368.353.856.628.560.261.935.624.954.792.684.493.793.2
+4-shot-ICL40.452.937.667.138.241.315.242.445.622.725.229.443.543.859.448.4
+MTV32.321.941.968.554.857.329.763.564.037.040.752.391.791.294.394.1
+LoRA84.887.746.370.855.357.428.865.266.140.438.472.998.496.497.195.0
+SAVs85.196.047.268.357.660.932.370.071.042.547.579.998.197.699.898.7
+61.1+69.1+3.9+0+3.8+4.3+3.8+9.8+9.1+13.8+22.6+25.2+5.5+13.2+6.1+5.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.426, + 0.907, + 0.482 + ], + "angle": 0, + "content": "Table 1. Results evaluation on Safety, Visual Question Answering (VQA), Image-Text Retrieval (I-T Retrieval), and Image Classification benchmarks. The best result for each generative model is shown in bold and the second best in underline. We gray out additional baselines. We note that CLIP and SigLIP cannot be evaluating directly on tasks with interleaved image-text queries. † GPT-4o is a close-source model and as such, is shown just as an upperbound (since SAVs are not directly applicable). Key: NB - NaturalBench, SC - SUGARCREPE." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.766 + ], + "angle": 0, + "content": "differences between two similar images, making Natural-Bench especially challenging for any existing VL models. The class labels are \\(\\mathcal{C} = \\{\\text{\"A\", \"B\"}\\}\\). As suggested in the paper, we evaluate \"question accuracy\" which awards one point if a model correctly answers a question for both images, \"image accuracy\" which awards a point when a model answers both questions for an image, and finally \"group accuracy\" awards one point when a model correctly answers all four pairs. (3) Vizwiz [20] is VQA dataset designed to progress research in vision systems to assist blind and vision-impaired individuals. The dataset was collected by asking blind people to take pictures and record questions about the image. Although typically a generative task, we reformulate Vizwiz as first a classification task distinguishing answerable and unanswerable questions followed by standard response generation. The class labels are \\(\\mathcal{C} = \\{\\text{\"answerable\", \"unanswerable\"}\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Image-Text Retrieval. NaturalBench-Retrieval [44] and SUGARCREPE [28] both measure fine-grained semantic understanding in image-text pairs, with class labels \\(\\mathcal{C} = \\{\\text{\"Yes\",\"No\"}\\}\\) indicating whether an image-text pair is correctly matched. While SUGARCREPE presents one image with two captions (original and altered), NaturalBench-Retrieval adds complexity by using two similar images with two corresponding captions, effectively eliminating lan" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.539 + ], + "angle": 0, + "content": "guage bias and requiring models to capture more nuanced visual-semantic relationships." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.735 + ], + "angle": 0, + "content": "Image Classification. Image classification tasks evaluate a model's ability to categorize images into predefined classes. We evaluate on several standard classification benchmarks: EuroSAT [22] (satellite imagery for land use classification), Oxford-IIIT-Pets [75] (pet breed identification with visually similar categories), Flowers [71] (flower species recognition), Caltech Birds (CUB) [96] (fine-grained bird species classification), and ImageNet-1k [10] (general object recognition). For each dataset except ImageNet, we formulate the task as a 4-way multiple choice question with labels \\(\\mathcal{C} = \\{\\text{\"Class 1\", \"Class 2\", \"Class 3\", \"Class 4\"}\\}\\). Due to model oversaturation in simpler formulations, ImageNet-1k is formulated as a 16-way classification problem." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.623, + 0.758 + ], + "angle": 0, + "content": "4.3. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.903 + ], + "angle": 0, + "content": "For our results, we utilized SAVs with 20 examples per label. We compared our method with multiple SOTA baselines, including GPT-4o [73]. As a closed-source model, GPT-4o is presented simply as a strong closed-source baseline. Furthermore, we present the results of classic LMMs, LLaVA-1.5 and InstructBLIP, which have been previously applied to image classification [111]. Zero-shot (ZS) baselines are implemented by querying the model directly and generating a response. For image-text retrieval, ZS uses the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2764" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.355, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.316, + 0.347, + 0.33 + ], + "angle": 0, + "content": "Hallucination Top Head Att. Vec." + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.33, + 0.353, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.089, + 0.631, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.401, + 0.316, + 0.627, + 0.33 + ], + "angle": 0, + "content": "Relative Depth Top Head Att. Vec." + }, + { + "type": "image", + "bbox": [ + 0.399, + 0.33, + 0.629, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.089, + 0.905, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.316, + 0.888, + 0.33 + ], + "angle": 0, + "content": "Image Cls. Top Head Att. Vec." + }, + { + "type": "image", + "bbox": [ + 0.676, + 0.33, + 0.904, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.532, + 0.908, + 0.577 + ], + "angle": 0, + "content": "Figure 3. Head Visualization. On the top row, we show the top-20 attention head locations for a given task. All the heads are indexed by their (head, layer) position, and the selected heads are highlighted. On the bottom row, we visualize the attention vector of few-shot examples for the top head of the given class with t-SNE clustering [26]. Each point represents the embedding of an input sample." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.601, + 0.483, + 0.692 + ], + "angle": 0, + "content": "SOTA generative scoring method VQAScore [57]. In addition to ZS, we also compare to several test-time and finetuning few-shot methods (all with the exact same sample complexity as SAVs). For instance, we compare to the current SOTA multimodal few-shot method, MTV [30] as well as doing 4-shot ICL and LoRA [29] finetuning for each task." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.185, + 0.725 + ], + "angle": 0, + "content": "4.4. Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Results are shown in Table 1. An advantage of our method is its adaptability to any VL classification task that has image, text, or interleaved image-text inputs. Thus, we apply SAVs to a wide range of tasks in safety, VQA, imagetext retrieval and image classification. Furthermore, our approach even significantly improves over SOTA ZS, few-shot, and finetuning methods. One interesting observation is that few-shot ICL consistently deteriorates performance. We hypothesize that instructioning tuning of LMMs disrupts the few-shot prompting capabilities gained during pretraining in favor of structured instructions as shown" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.707 + ], + "angle": 0, + "content": "in [14, 86]. On the other hand, SAVs overcome this by directly operating on the internal activations. Finally, not only is our method successful across a wide-range of tasks, but it also improves on challenging visual perception and compositional reasoning tasks (e.g., BLINK and NaturalBench) that all VL models struggle with. Please refer to our supplementary section in Section 7 for mor results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.723, + 0.625, + 0.737 + ], + "angle": 0, + "content": "4.5. Ablations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.907, + 0.809 + ], + "angle": 0, + "content": "We perform a comprehensive ablation study of our method on MHaluBench, NaturalBench, and EuroSAT (see Table 2). For more ablations, please refer to Section 7.1 in the Supp. For all ablations, we use LLaVA-OneVision-7B." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Varying number of examples. In Figure 4 (left), we examine the impact of varying the number of few-shot examples used in our method. Our primary results in Table 1 indicate that just 20 examples per label are necessary to yield state-of-the-art performance on a variety of VL classification tasks. This ablation shows that accuracy on these tasks" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2765" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.098, + 0.493, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.099, + 0.891, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.3, + 0.908, + 0.331 + ], + "angle": 0, + "content": "Figure 4. Scaling Property of SAVs. Performance of LLaVA-OneVision-7B + SAVs on varying number of few-shot examples per label (left). Performance of of LLaVA-OneVision-7B + SAVs on varying numbers of attention vectors used (right)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.356, + 0.468, + 0.371 + ], + "angle": 0, + "content": "can scale with increasing numbers of examples per label." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.371, + 0.483, + 0.461 + ], + "angle": 0, + "content": "Varying number of attention vectors. In order to extract attention vectors, we select a very sparse set of heads from hundreds. We vary the number of attention vectors selected in Step 2 of our method and demonstrate that just 20 vectors are enough to realize nearly all of the classification accuracy of our method. Results are shown in Figure 4 (right)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.462, + 0.483, + 0.642 + ], + "angle": 0, + "content": "SAVs are flexible to different evaluation strategies. In our method, we leverage class centroid classification as the evaluation method for selecting sparse features. We view this flexibility of the sparsification method to be a key feature of our work. As such, we compare our class centroid classification approach to k-nearest neighbors (KNN) and linear probing. For linear probing, we train a lightweight MLP module for 20 epochs using the top heads' features. All methods make use of the same 20 examples per label for consistency. Our results in Table 3a show that class centroid classification outperforms both KNN classification and is comparable with linear probing." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.763 + ], + "angle": 0, + "content": "Comparing heads vs. layers. Based on prior work and transformer-architecture intuition, we treat the attention vectors outputted by the heads as a viable set of features for classification tasks. We verify this intuition by comparing the performance of selecting 2 sparse layers to selecting sparse heads as feature maps for our tasks. As shown in Table 3b, head-based attention vectors outperform concatenated layer features on all three benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.312, + 0.789 + ], + "angle": 0, + "content": "4.6. Additional Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In this subsection, we present experiments that demonstrate additional properties and capabilities of SAVs, beyond its use as features for VL classification tasks. Additional experiments can be found in Section 7.2 of the Supplementary. For all experiments, we use LLaVA-OneVision-7B." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visualizing SAVs. SAVs are both an efficient and interpretable method for leveraging generative LMMs for VL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.355, + 0.905, + 0.567 + ], + "angle": 0, + "content": "classification tasks. To emphasize this point, we show the selected heads for hallucination detection, relative depth, and image classification in the first row of Figure 3. The visualizations demonstrate both the sparsity and specificity of the SAVs that are used for each task. Unlike prompting and finetuning methods, our approach clearly outlines exactly where in the model's activation space informative attention vector features are extracted from. We furthermore show that the features extracted from these heads are useful for the given task in the second row of the figure, where we visualize the features outputted by the top selected head for each few-shot sample via t-SNE [26]. The clear clustering of examples of the same label indicates that even with a single head, high-quality features are being selected as SAVs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.905, + 0.688 + ], + "angle": 0, + "content": "Evaluating SAVs generalizability. Here, we ask whether SAVs extracted from one task, can generalize to another similar task. We utilize SAV heads from VLGuard to evaluate on MhaluBench and vice versa. We do a similar approach with LoRA, by applying LoRA finetuned on VLGuard to MHaluBench. Interestingly, our results in Table 2a show that SAVs generalize between tasks, while LoRA weights, as expected, overfit to the finetuned task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Comparing SAVs to CLIP/SigLIP on interleaved imagetext tasks. As discussed in Section 1, SAVs are fully multimodal features able to represent inputs that are image-only, text-only, and even interleaved image-text. This is something that is not possible to directly replicate with CLIP and SigLIP models which have separate image and text encoders. Nevertheless, we compare our method to both CLIP and SigLIP on tasks that require interleaved imagetext inputs. While SAVs can do this natively, we enable this comparison by concatenating the separate image and text features of CLIP and SigLIP in order to evaluate on MHaluBench and NaturalBench. We find that our method vastly outperforms concatenated CLIP and SigLIP features on both benchmarks as shown in Figure 2b. This result" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2766" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.167, + 0.088, + 0.265, + 0.099 + ], + "angle": 0, + "content": "(a) Generalization" + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.1, + 0.337, + 0.184 + ], + "angle": 0, + "content": "
MHBVLG
Zero-Shot34.731.4
VLG LoRA34.190.0
VLG SAV55.394.3
" + }, + { + "type": "table_caption", + "bbox": [ + 0.417, + 0.088, + 0.529, + 0.099 + ], + "angle": 0, + "content": "(b) Interleaved Tasks" + }, + { + "type": "table", + "bbox": [ + 0.371, + 0.1, + 0.576, + 0.184 + ], + "angle": 0, + "content": "
MHBNB
CLIP51.91.2
SigLIP48.61.2
SAVs80.835.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.678, + 0.088, + 0.838, + 0.099 + ], + "angle": 0, + "content": "(c) SAVs vs Few-Shot SigLIP" + }, + { + "type": "table", + "bbox": [ + 0.61, + 0.1, + 0.906, + 0.184 + ], + "angle": 0, + "content": "
NBES
SigLIP Few-shot Mean1.288.2
SigLIP Few-shot Cross-Modal6.886.4
SAVs35.186.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.193, + 0.907, + 0.222 + ], + "angle": 0, + "content": "Table 2. Additional SAVs Experiments. We (a) demonstrate the generalization of SAV heads to similar tasks, (b) show the effectiveness of SAV for tasks with interleaved image-text inputs, and (c) compare SAVs with few-shot formulations of CLIP and SigLIP." + }, + { + "type": "table_caption", + "bbox": [ + 0.206, + 0.24, + 0.347, + 0.251 + ], + "angle": 0, + "content": "(a) Classification Methods" + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.252, + 0.457, + 0.326 + ], + "angle": 0, + "content": "
MHBNBES
Class Centroid80.835.186.7
KNN53.011.078.1
Linear Probe82.532.983.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.624, + 0.24, + 0.762, + 0.252 + ], + "angle": 0, + "content": "(b) Sparse Configurations" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.253, + 0.871, + 0.325 + ], + "angle": 0, + "content": "
MHBNBES
Sparse Heads80.835.186.7
Sparse Layers79.028.481.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.335, + 0.907, + 0.392 + ], + "angle": 0, + "content": "Table 3. SAVs Ablations. We perform several ablations to identify the important aspects of our method that contribute to its effectiveness. In particular, we compare the effectiveness of (a) different classification methods and (b) head feature sparsification versus layer feature sparsification. Note: MHB represents MHaluBench, NB represents NaturalBench Group Score, and ES represents EuroSAT. For more ablations, please refer to Section 7.1 in the Supplementary." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.482, + 0.448 + ], + "angle": 0, + "content": "demonstrates the adaptability of our method to any VL classification regardless of the input's modality." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.45, + 0.485, + 0.797 + ], + "angle": 0, + "content": "Comparing SAVs to few-shot SigCLIP. As SAVs are extracted with few-shot examples, we compare our method to an analogous version of few-shot SigLIP. However, because SigLIP cannot be directly made few-shot, we adapt SigLIP as a few-shot class centroid classifier. One method used is the current SOTA few-shot classification method for encoder models Cross-Modal Adaptation [56]. We apply a second method where we aggregate CLIP/SigLIP embeddings into a mean embedding for each label. Then, just as in SAVs, we perform class centroid classification for each query using our set of mean SigLIP embeddings. We note that for image classification like EuroSAT, only image embeddings are needed, but for MHaluBench, multimodal embeddings are necessary. SAVs are inherently multimodal and so can be flexibly applied to both, but CLIP/SigLIP only have image-only or text-only embeddings. To overcome this, we use SigLIP image features for EuroSAT and concatenate image and text features for MHaluBench. Interestingly, while SiGLIP is comparable to SAVs on EuroSAT, our method vastly outperforms SigLIP in the few-shot setting for MHaluBench, suggesting generalizability of our method for a variety of multimodal tasks that CLIP-like struggle with. Our results are shown in Table 2c." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.814, + 0.21, + 0.83 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Our research demonstrates the effectiveness of extracting Sparse Attention Vectors (SAVs) from the heads of an LMM and utilizing them directly for vision-language classification. Our method stands out by using only few-shot exam" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.907, + 0.675 + ], + "angle": 0, + "content": "plies per label and only less than \\(1\\%\\) of the heads to outperform zero-shot, few-shot, and fine-tuned baselines on a variety of image-text and image-only tasks. In addition, SAVs allows generative LMMs to close the gap with closed-source GPT-4o while also being an interpretable method that can generalize to similar tasks. Our ablations reveal the flexibility of using any classification method as a sparsification method for attention vectors and also shows that features are found as outputs of heads rather than layers. Overall, these results show that SAVs is a lightweight, performant, and generalizable method for extending generative LMMs' multimodal classification abilities. We are encouraged by the outcomes, and anticipate many directions for future work. In addition to methodological improvements, we look forward to the application of SAVs as features for multimodal retrieval, data compression, or more generally as a distilled representation for downstream models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.637, + 0.708 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Sparse Attention Vectors are a significant step in generalizing the capabilities of generative LMMs to classification tasks. Nevertheless, it is valuable to consider certain limitations of our approach. SAVs are a method that requires access to the model's internal architecture and so may not be directly applicable to closed-source models like GPT-4 [73] and Gemini [89, 90]. Additionally, some tasks like image-text retrieval [28, 91] can benefit from more fine-grained confidence metrics attached to each label than proportion of voting heads per label. These challenges prompt future work in these directions as well as exciting questions about how to use SAVs as feature embeddings for other tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2767" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.092, + 0.236, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.11, + 0.486, + 0.173 + ], + "angle": 0, + "content": "We would like to thank Abrar Anwar and Tyler Bonnen for helpful feedback and discussions. This project was supported in part by DoD, including PTG and/or LwLL programs, as well as BAIR's industrial alliance programs." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.185, + 0.188, + 0.2 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.21, + 0.482, + 0.25 + ], + "angle": 0, + "content": "[1] S. R. Bowman A. Williams, N. Nangia. A broad-coverage challenge corpus for sentence understanding through inference. ArXiv, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.253, + 0.483, + 0.32 + ], + "angle": 0, + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.323, + 0.482, + 0.377 + ], + "angle": 0, + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv, abs/2308.12966, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.38, + 0.482, + 0.421 + ], + "angle": 0, + "content": "[4] Pierre Baldi. Autoencoders, unsupervised learning, and deep architectures. In ICML 2011 Unsupervised and Transfer Learning Workshop, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.423, + 0.482, + 0.463 + ], + "angle": 0, + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.466, + 0.482, + 0.534 + ], + "angle": 0, + "content": "[6] Matyas Bohacek and Michal Bravansky. When XGBoost outperforms GPT-4 on text classification: A case study. In Trustworthy Natural Language Processing (TrustNLP) 4th Workshop, pages 51–60, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.536, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. *ArXiv*, abs/2005.14165, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.69, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[8] Martin Juan José E Bucher and Marco Martini. Finetuned 'small' llms (still) significantly outperform zero-shot generative ai models in text classification. ArXiv, abs/2406.08660, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.747, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[9] Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Jinjie Gu, and Huajun Chen. Unified hallucination detection for multimodal large language models. arXiv preprint arXiv:2402.03190, 2024. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[11] Xiang Deng, Vasilisa Bashlovkina, Feng Han, Simon Baumgartner, and Michael Bendersky. Llms to the moon?" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.21, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "reddit market sentiment analysis with large language models. Companion Proceedings of the ACM Web Conference 2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[12] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11162-11173, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.193, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.25, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[14] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. ArXiv, abs/2403.12736, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.307, + 0.904, + 0.346 + ], + "angle": 0, + "content": "[15] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Herv'e J'egou. Training vision transformers for image retrieval. ArXiv, abs/2102.05644, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[16] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[17] Evelina Fedorenko, Michael K. Behr, and Nancy Kanwisher. Functional specificity for high-level linguistic processing in the human brain. Proceedings of the National Academy of Sciences, 108(39):16428-16433, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.463, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[18] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 2, 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.904, + 0.573 + ], + "angle": 0, + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.576, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Danna Gurari, Qing Li, Abigale Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[21] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.717, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[22] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.788, + 0.904, + 0.828 + ], + "angle": 0, + "content": "[23] Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors. ArXiv, abs/2310.15916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.831, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Roee Hendel, Mor Geva, and Amir Globerson. Incontext learning creates task vectors. arXiv preprint arXiv:2310.15916, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.874, + 0.904, + 0.901 + ], + "angle": 0, + "content": "[25] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson." + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2768" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.131, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "Incorporating structured representations into pretrained vision \\(\\&\\) language models using scene graphs. In 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[26] L. Hinton G, van der Maaten. Visualizing data using t-sne. Journal of machine learning research, 2008. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[27] Alberto Hojel, Yutong Bai, Trevor Darrell, Amir Globerson, and Amir Bar. Finding visual task vectors. In European Conference on Computer Vision (ECCV), pages 257-273. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.287 + ], + "angle": 0, + "content": "[28] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. ArXiv, abs/2306.14610, 2023. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[30] Brandon Huang, Chancharik Mitra, Assaf Arbelle, Leonid Karlinsky, Trevor Darrell, and Roei Herzig. Multimodal task vectors enable many-shot multimodal in-context learning. arXiv preprint arXiv:2406.15334, 2024. 2, 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[31] Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.47, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[32] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6693-6702, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.539, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[33] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[34] Ting Jiang, Shaohan Huang, Zi qiang Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Liangjie Zhang, and Qi Zhang. Promptbert: Improving bert sentence embeddings with prompts. In Conference on Empirical Methods in Natural Language Processing, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[35] Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. Scaling sentence embeddings with large language models. ArXiv, abs/2307.16645, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. ArXiv, abs/2407.12580, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[37] Nancy Kanwisher. Domain specificity in face perception. Nature neuroscience, 3(8):759-763, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[38] Sahar Kazemzadeh, Vicente Ordonez, Marc andre Matten, and Tamara L. Berg. Referitgame: Referring to objects in photographs of natural scenes. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[39] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[40] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.163, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[41] Yibin Lei, Di Wu, Tianyi Zhou, Tao Shen, Yu Cao, Chongyang Tao, and Andrew Yates. Meta-task prompting elicits embeddings from large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.232, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[42] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[43] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv, abs/2307.16125, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.344, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[44] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Natural-bench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.427, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[45] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, abs/2408.03326, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.484, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.539, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[47] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.595, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[48] Shiyu Li, Yang Tang, Shizhe Chen, and Xi Chen. Conan-embedding: General text embedding with more and better negative samples. ArXiv, abs/2408.15710, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.637, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[49] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.72, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[50] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23390-23400, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[51] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.832, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[52] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[53] Jiacheng Lin, Kun Qian, Haoyu Han, Nurendra Choudhary, Tianxin Wei, Zhongruo Wang, Sahika Genc, Edward W" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2769" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Huang, Sheng Wang, Karthik Subbian, et al. Unleashing the power of lms as multi-modal encoders for text and graph-structured data. arXiv preprint arXiv:2410.11235, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.135, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[54] Tsung-Yi Lin, M. Maire, Serge J. Belongie, James Hays, P. Perona, D. Ramanan, Piotr Dollar, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), 2014. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.19, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[55] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.246, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[56] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[57] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[58] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[59] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.525 + ], + "angle": 0, + "content": "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[61] Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera, Luca Zancato, and Stefan O Soatto. Meaning representations from trajectories in autoregressive models. ArXiv, abs/2310.18348, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.58, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[62] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.636, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[63] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 33:5081–5092, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[64] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. ArXiv, abs/2405.20797, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[65] Junhua Mao, Jonathan Huang, Alexander Toshev, OanaMaria Camburu, Alan Loddon Yuille, and Kevin P. Murphy. Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[66] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[67] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain of thought prompting" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "for large multimodal models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[68] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. In Conference of the European Chapter of the Association for Computational Linguistics, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.179, + 0.905, + 0.234 + ], + "angle": 0, + "content": "[69] Jianmo Ni, Gustavo Hernández Abrego, Noah Constant, Ji Ma, Keith B. Hall, Daniel Matthew Cer, and Yinfei Yang. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. ArXiv, abs/2108.08877, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.236, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[70] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith B. Hall, Ming-Wei Chang, and Yinfei Yang. Large dual encoders are generalizable retrievers. ArXiv, abs/2112.07899, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.307, + 0.905, + 0.362 + ], + "angle": 0, + "content": "[71] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.363, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[72] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.434, + 0.904, + 0.462 + ], + "angle": 0, + "content": "[73] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 1, 5, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.463, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[74] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.576, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[75] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.632, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[76] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.718, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[77] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.774, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.859, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[79] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2770" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "reward model. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.121, + 0.483, + 0.175 + ], + "angle": 0, + "content": "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[81] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[82] David E Rumelhart, James L McClelland, PDP Research Group, et al. Parallel distributed processing, volume 1: Explorations in the microstructure of cognition: Foundations. The MIT press, 1986. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.289, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[83] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815-823, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[84] Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.386, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[85] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8309-8318, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[86] Aaditya K. Singh, Stephanie C.Y. Chan, Ted Moskovitz, Erin Grant, Andrew M. Saxe, and Felix Hill. The transient nature of emergent in-context learning in transformers. In Advances in neural information processing systems (NeurIPS), 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.525, + 0.483, + 0.608 + ], + "angle": 0, + "content": "[87] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.483, + 0.664 + ], + "angle": 0, + "content": "[88] Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. Text classification via large language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.483, + 0.705 + ], + "angle": 0, + "content": "[89] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. ArXiv, abs/2403.05530, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.483, + 0.775 + ], + "angle": 0, + "content": "[90] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[91] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[92] Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau. Function vectors in large language models. ArXiv, abs/2310.15213, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[93] Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. Function vectors in large language models. arXiv preprint arXiv:2310.15213, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.149, + 0.906, + 0.19 + ], + "angle": 0, + "content": "[94] Matthew A. Turk and Alex Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3:71-86, 1991. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.192, + 0.906, + 0.234 + ], + "angle": 0, + "content": "[95] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. ArXiv, abs/1711.00937, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.235, + 0.906, + 0.276 + ], + "angle": 0, + "content": "[96] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.278, + 0.906, + 0.332 + ], + "angle": 0, + "content": "[97] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. ArXiv, abs/2401.00368, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.334, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[98] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Ke-Yang Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. ArXiv, abs/2409.12191, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.433, + 0.906, + 0.502 + ], + "angle": 0, + "content": "[99] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. arXiv preprint arXiv:2401.11206, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.56 + ], + "angle": 0, + "content": "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.906, + 0.657 + ], + "angle": 0, + "content": "[101] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. Internvideo2: Scaling video foundation models for multimodal video understanding. ArXiv, abs/2403.15377, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.906, + 0.7 + ], + "angle": 0, + "content": "[102] Zhiqiang Wang, Yiran Pang, and Yanbin Lin. Large language models are zero-shot text classifiers. ArXiv, abs/2312.01044, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.757 + ], + "angle": 0, + "content": "[103] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. ArXiv, abs/2109.01652, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[104] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.858 + ], + "angle": 0, + "content": "[105] Dean Wyatte, Fatemeh Tahmasbi, Ming Li, and Thomas Markovich. Scaling laws for discriminative classification in large language models. ArXiv, abs/2405.15765, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[106] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "2771" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[107] Peter Young, Alice Lai, Micah Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[108] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[109] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[110] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and C. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine Learning in Health Care, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[111] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[112] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. ArXiv, abs/2309.07915, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[113] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[114] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2772" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_origin.pdf b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c37d7c5af6a5ddd407456504c8e9c6036f78f9c7 --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/958199ee-2123-4605-9c2c-3f90d0a78ead_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e3084c931725b4279b90f5dc70bf21926bc2901fddee1ce7135f4685ea246df7 +size 1593010 diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/full.md b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/full.md new file mode 100644 index 0000000000000000000000000000000000000000..439caa06d0aef68372e7bc0d9d3f824c1bd8ed87 --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/full.md @@ -0,0 +1,385 @@ +# Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features + +Chancharik Mitra $^{1*}$ Brandon Huang $^{2*}$ Tianning Chai $^{2}$ Zhiqiu Lin $^{1}$ Assaf Arbelle $^{3}$ Rogerio Feris $^{4}$ Leonid Karlinsky $^{4}$ Trevor Darrell $^{2}$ Deva Ramanan $^{1}$ Roei Herzig $^{2,4}$ + +$^{1}$ Carnegie Mellon University $^{2}$ University of California, Berkeley $^{3}$ IBM Research $^{4}$ MIT-IBM Watson AI Lab + +# Abstract + +Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs)—a finetuning-free method that leverages sparse attention head activations (fewer than $5\%$ of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations. + +# 1. Introduction + +Generative Large Multimodal Models (LMMs) such as GPT-4V [73], LLaVA [59, 60], and QwenVL [3] demonstrate state-of-the-art performance on open-ended vision-language (VL) tasks like image captioning [54, 107], visual question answering [2, 32, 43], and language grounding [38, 65]. However, despite their remarkable performance on generative tasks, these models struggle on vision-language classification tasks, where responses are a discrete set of labels [8, 111]. Indeed, LMMs with billions of parameters and trained on trillions more tokens of data underperform smaller encoder VLMs like CLIP and SigLIP [8, 111] + +![](images/666b706b0f887295d28038f80fd97824ec1f34ab370bb4485266bf6b575b1e24.jpg) +Figure 1. Sparse Attention Vectors (SAVs) Overview. We develop a method for extracting features from a generative LMM without finetuning. We first extract a sparse set of attention vectors for each task given a set of few-shot examples, and then, we utilize these attention vectors directly as features for downstream vision-language classification tasks. + +![](images/86cf6fc0bf9307c7557fbc168f79bbd5df228af602254180e66b6440ec931564.jpg) + +and even classical machine learning methods [6] on image classification tasks. One reason that CLIP-like models excel at classification is the ease of extracting visual or textual features from modality-specific encoders. However, such models cannot process joint vision-language inputs, as LMMs can. Our goal is to extract compact features from generative LMMs, enabling downstream discrete labeling tasks such as classification and multiple-choice VQA. + +Feature extraction is a well-explored field in both vision-only [15, 83, 94] and language-only encoder models [13, 81], but such is not the case with generative models. Most + +current methods for extracting features from generative models require carefully constructed prompts [41, 67, 104], specialized architectures [51], few-shot prompting [7, 11, 112], and finetuning [8, 64, 111]. Recent work [111] shows that prompting approaches fail to close the gap with encoder VLMs, while finetuning requires training-scale data for every unseen task, which is inefficient. However, as we observe in our work, generative models still offer the promise of more flexible, truly multimodal features as compared to modality-specific features from CLIP-like models. + +A source of inspiration for our method is long-standing work in neuroscience that suggests certain areas of the brain are reserved for specific tasks [17, 37] (i.e. functional specificity). Motivated by this idea, we refer to recent interpretability research that has focused on identifying specific heads in transformer-based models that correspond to particular tasks [72]. The most prominent of these methods is a line of work that looks to enhance vision-language capabilities using task vectors [24, 27, 30, 93], which are compact implicit representations of tasks encoded in the activations of a transformer model. While promising, these methods ultimately use these representations to augment a model's generative capabilities. On the other hand, we seek to use feature representations directly as classifiers. Nevertheless, this intuition from interpretability informs our work on Sparse Attention Vectors (SAVs), which are sparse features in an LMMs activation space that can be directly exploited for few-shot VL classification. + +Our method has three steps: First, we extract features (i.e., attention vectors) from the output of each head of the LMM for some few-shot labeled examples ( $\approx 20$ per label). Second, we average these attention vectors over the examples in each class and evaluate their accuracy as centroids in a class centroid classifier. We then select the top 20 heads by classification accuracy as our SAVs. In this way, we identify a very sparse set of attention vectors (less than $5\%$ of the total number of heads) that can be used for discriminative tasks. Finally, we perform inference on the given task by doing a majority vote across this sparse set of attention vectors for each new query. This approach requires only few-shot examples at test-time to extract effective multimodal embeddings. An overview is shown in Figure 1. + +We summarize our main contributions as follows: (i) We introduce a novel method that yields a sparse set of attention vectors to serve as highly effective features for individual VL classification tasks; (ii) Our method surpasses zero-shot, few-shot, and LoRA fine-tuned baselines across multiple tasks ( $+7\%$ improvement on average over LoRA on challenging benchmarks like BLINK [18], VLGuard [114], and NaturalBench [44]); (iii) We establish several advantageous properties of our approach, including strong generalization capabilities and favorable scaling characteristics. + +# 2. Related Works + +Controllable Generation for Classification. Controllable text generation in LMMs guides model outputs to meet specific constraints. For classification tasks with generative LMMs, several approaches exist: test-time hard prompting [7, 104] uses prompt engineering or few-shot examples to elicit class label outputs [11, 88, 102, 105, 111]; direct probability analysis of generated class labels [55, 58] for image-text retrieval; and soft prompting methods that finetune learnable tokens [42, 49]. Other techniques include instruction finetuning [103] on labeled data [6, 111] and preference modeling like DPO and RLHF [16, 74, 79]. Our method, however, is finetuning-free and directly selects class labels without preference data. Most related work shows that internal transformer representations called task vectors [23, 27, 30, 92] can encapsulate ICL example tasks. Beyond previous approaches, we use a sparse set of attention vectors directly as VL classification task task features. + +Vision-Language Features. Feature extraction seeks useful representations for diverse downstream tasks. Early embedding techniques including autoencoders [4, 39, 62, 63, 82], Word2Vec [66] and GloVe [77] transformed inputs into vector representations, followed by advances in NLP [13, 19, 68, 81] and vision [15, 83, 94]. Recent methods like CLIP and SigCLIP [46, 47, 50, 78, 108, 109] explore multimodal correlations through contrastive learning or sigmoid loss. These representations offer flexibility across tasks [12, 25, 33, 40, 80, 95] and domains [53, 101, 110]. + +Interestingly, extracting features from generative models poses unique challenges in identifying optimal extraction points. Some approaches finetune encoder VLMs on LLM-generated data [48, 97] or finetune LMMs directly on specific tasks [31, 111]. More efficient methods finetune encoder-decoder representations for better alignment [36, 64, 69, 70]. Finetuning-free approaches use distillation prompts to extract representations from model weights or activations [34, 35, 41, 61], while other methods employ mixture-of-experts [52], expert models [99], or embedding reranking [21]. + +Thus, current SOTA faces the following challenges: (1) modality-specific rather than multimodal features, (2) requiring finetuning, (3) limited flexibility due to specialized prompts, and (4) dependence on multiple models. Our approach provides effective multimodal embeddings without any gradient-based finetuning and flexibly apply to various VL classification tasks without additional models. + +# 3. Methods + +In this section, we outline our approach for using sparse attention vectors from the activation space of a transformer-based large multimodal model (LMM) as features for any VL classification task. The method consists of three main + +![](images/2cd40e211ec3da687d4de37aa8df1d752ce1602a2582550ee7c4548b4ee2391b.jpg) +1. Extract Attention Vectors +2. Find Sparse Attention Vectors (SAVs): +Select Sparse Attention Vectors + +![](images/4902bb64a69ee271e40c1203d2849b6ad4c020c4b2c3f22df58fca0e8bbbcaa5.jpg) +3. SAVs for Discriminative Tasks: +Figure 2. Sparse Attention Vectors (SAVs) Detailed View. Our method is broken into the following three parts: (1) Given a set of few shot examples to be classified by a frozen LMM, we extract attention vectors across all heads and layers. (2) These attention vectors are averaged across the set of examples for each class. For each head, we use these mean attention vectors as centroid classifiers, which are then used to select a sparse set of $k$ heads with the highest classification accuracy. (3) Finally, we use these sparse attention vectors to directly classify new inputs via majority vote (denoted by the "Classifier" block). + +steps: (i) extracting the attention vectors from all attention heads in the model, (ii) identifying a sparse set of vectors based on their ability to consistently return the correct label for some support set of examples, and (iii) using these sparse features to classify new queries. We begin with a formal description of the transformer decoder LLM and its attention mechanism, followed by the detailed methodology for sparse attention vector selection and classification. A detailed view of our method is shown in Figure 2. + +# 3.1. Preliminaries + +A transformer-based large language model (LLM) with $L$ layers and $H$ attention heads per layer processes input sequences through multi-head self-attention mechanisms. Each layer combines multiple attention heads to capture different aspects of the input sequence, followed by feedforward networks for further processing. + +Multi-Head Attention. Let $x = \{x_{1},x_{2},\ldots ,x_{T}\}$ represent a sequence of input tokens, where $x_{i}$ is the $i^{\text{th}}$ token. For each layer $l\in \{1,\dots ,L\}$ , the input sequence is projected into queries, keys, and values for each attention head $m\in \{1,\dots ,H\}$ . Each head performs the following scaled dot-product attention: + +$$ +\mathbf {h} _ {l} ^ {m} \left(x _ {i}\right) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {m}}}\right) V +$$ + +where $Q$ , $K$ , and $V$ are the query, key, and value matrices respectively, and the dimensionality of each head $d_{m}$ which is given by $\frac{d}{H}$ (the embedding dimension divided by the number of heads). We denote $\mathbf{h}_l^m(x_i)$ as an attention vector for head $m$ in layer $l$ . + +The outputs of all heads are concatenated and projected to form the layer output: + +$$ +\operatorname {M u l t i H e a d} \left(x _ {i}\right) = \operatorname {C o n c a t} \left(\mathbf {h} _ {l} ^ {1} \left(x _ {i}\right), \dots , \mathbf {h} _ {l} ^ {H} \left(x _ {i}\right)\right) W ^ {O} +$$ + +where $W^{O}$ is the output projection matrix. + +In our work, we look to leverage attention vectors for the purpose of vision-language classification tasks. Specifically, the attention vectors are used as latent representations of the inputs to both find attention heads in an LMM suited for a classification task and then perform downstream inference using those selected attention heads. We describe our method in detail in the sections that follow. + +# 3.2. Sparse Attention Vectors + +Our key insight is that within the many attention heads and transformer layers of an LMM, there exists a sparse subset + +that can serve as effective features for vision-language classification tasks. We present a three-step method to identify and utilize these features to build lightweight classifiers. + +Step 1: Extracting Attention Vectors. Given a frozen LMM and few-shot examples of sequence-label pairs $\{(x_1, y_1), (x_2, y_2), \ldots, (x_N, y_N)\}$ we first extract the attention vectors for each sequence $x_i$ . Specifically, we compute the attention vector $\mathbf{h}_l^m(x_i^T)$ for head $m$ from layer $l$ for the final token $x_i^T$ . This yields a set of attention vectors $\{\mathbf{h}_l^m(x_i^T) \mid i = 1, \dots, N\}$ for each head $m$ and layer $l$ . +Step 2: Identifying Relevant Vectors. The central question is how to identify which attention vectors are naturally suited for the classification task at hand. We evaluate each vector's ability as a local classifier by computing its performance under a nearest class centroid classifier. + +Specifically, for each class $c \in \mathcal{C}$ , compute its centroid (or mean) attention vector across the few shot examples: + +$$ +\mu_ {c} ^ {l, m} = \frac {1}{| N _ {c} |} \sum_ {j: y _ {j} = c} \mathbf {h} _ {l} ^ {m} (x _ {j} ^ {T}) +$$ + +where $N_{c} = \{j:y_{j} = c\}$ is the set of indices of examples with label $c$ . For each input $x_{i}$ , we compute its cosine similarity to each class centroid head: + +$$ +s _ {l, m} \left(x _ {i}, c\right) = \frac {\mathbf {h} _ {l} ^ {m} \left(x _ {i} ^ {T}\right) \cdot \mu_ {c} ^ {l , m}}{\left\| \mathbf {h} _ {l} ^ {m} \left(x _ {i} ^ {T}\right) \right\| \left\| \mu_ {c} ^ {l , m} \right\|}, \quad \forall c \in \mathcal {C} +$$ + +Next, we measure the classification accuracy of each head (i.e. local classifier): + +$$ +\operatorname {s c o r e} (l, m) = \sum_ {i = 1} ^ {N} \mathbf {1} [ \hat {y} = y _ {i} ] +$$ + +where the nearest class centroid label is given as $\hat{y} = \operatorname*{argmax}_{c\in \mathcal{C}}s_{l,m}(x_i,c)$ , and $\mathbf{1}[\cdot ]$ is the indicator function that + +evaluates to 1 when the condition is true (and 0 otherwise). + +We denote the set of $k$ top-scoring heads as $\mathcal{H}_{\mathrm{SAV}}$ : + +$\mathcal{H}_{\mathrm{SAV}} = \{(l,m)\mid \mathrm{score}(l,m)$ is among $k$ highest scores} + +# Step 3: Classification with Sparse Attention Vectors. + +Given a query sequence $Q$ to classify, we leverage our sparse set of heads $\mathcal{H}_{\mathrm{SAV}}$ for prediction. For each head $(l,m)\in \mathcal{H}_{\mathrm{SAV}}$ , we compute the class centroid $\mu_c^{l,m}$ closest to the query as follows: + +$$ +\hat{y}_{l,m} = \operatorname *{arg max}_{c\in \mathcal{C}}s_{l,m}(Q^{T},c) +$$ + +where $s_{l,m}(\cdot, \cdot)$ is defined as in Step 2. Our final class prediction is that of a global classifier that counts the majority vote across all local classifiers (heads) in $\mathcal{H}_{\mathrm{SAV}}$ : + +$$ +\underset {y \in \mathcal {C}} {\arg \max } \sum_ {(l, m) \in \mathcal {H} _ {\mathrm {S A V}}} \mathbf {1} [ \hat {y} _ {l, m} = y ] +$$ + +This approach reveals a surprising capability of LMMs: with just a few carefully selected attention heads ( $|\mathcal{H}_{\mathrm{SAV}}| \ll LH$ ), we can transform a generative language model into a lightweight vision-language classifier. This finding suggests that classification-relevant features naturally emerge within specific attention heads during model pretraining. In theory, one could replace our local and global classifiers with more complex variants (e.g. linear models, MLPs), something we explore in 3a. + +# 4. Evaluation + +We apply SAVs to two state-of-the-art LMMs—LLaVA-OneVision [45] and Qwen2-VL [98]. We also do a rigorous comparison of our method to strong few-shot and finetuning baselines on a variety of different image-text and image-only vision-language classification tasks covering safety, VQA, image-text retrieval, and simple classification. + +# 4.1. Implementation Details + +We implemented our approach in PyTorch [76]. We use the official implementations of each model, and all of our experiments can be run on a single NVIDIA A100 GPU. We use the default hyperparameters provided by the original models' (LLaVA-OV and Qwen2-VL) codebases. These hyperparameters are detailed in Table 8. More details of the implementation is included in the supplementary in Section 8. + +# 4.2. Benchmarks + +Safety. (1) LMM-Halucination [9] is a dataset which evaluates the hallucinations of the models when answering multimodal tasks. We report the raw classification accuracy of our method. Thus, the set of class labels for this task is given by $\mathcal{C} = \{\text{"hallucinating","not hallucinating"}\}$ . (2) VLGuard [114] is a dataset focusing on vision-language safety which identifies 4 main categories of harmful content: Privacy, Risky Behavior, Deception and Hateful Speech. VLGuard proposes Attack Success Rate (ASR) for evaluating unsafe inputs and Helpfulness for evaluating safe inputs. We simply reformat it as a classification task, where the set of class labels is given by $\mathcal{C} = \{\text{"safe","unsafe"}\}$ . + +VQA Datasets. In our work, we evaluate on VQA benchmarks, many of which can be formulated as a classification task. (1) BLINK [18] contains many tasks that are intuitive for humans but complicated for multimodal models such as multi-view reasoning, and visual similarity comparison. Since potential answers in BLINK are multiple choice, the class labels would be given as $\mathcal{C} = \{\text{"A", "B", "C", "D"}\}$ (note: the number of labels depends on the possible number of options allowed for a task). (2) NaturalBench-VQA [44] is a compositional dataset collected from natural image-text corpora but validated with human filtering. Each sample of the dataset contains two questions on compositional + +
ModelImage-Text TasksImage-Only Tasks
SafetyVQAI-T RetrievalClassification
MHaluVLGuardBLINKVizWizNB (T)NB (D)NB (G)NB-r (T)NB-r (D)NB-r (G)SCEuroSATPetsImagenet-IkFlowerCUB
CLIP-------41.845.023.235.364.088.196.192.897.8
SigLip-------54.554.931.242.763.998.397.695.898.4
GPT-4o†39.445.759.058.164.666.439.665.067.040.548.370.598.398.399.099.3
LLaVA-1.544.331.038.050.037.743.812.736.742.712.224.926.343.020.651.849.0
Instruct-BLIP44.018.538.734.520.224.24.019.521.31.14.720.456.011.018.233.8
LLaVA-OV-7B34.731.445.060.452.053.327.056.258.032.115.366.588.192.583.285.3
+4-shot-ICL25.035.038.947.847.650.422.248.249.631.516.147.163.949.063.860.6
+MTV37.332.944.561.156.258.030.758.159.533.428.765.588.595.983.285.6
+LoRA78.390.047.063.158.660.932.459.460.335.430.485.096.893.291.291.8
+SAVs80.894.351.866.160.362.335.172.773.053.137.686.797.097.599.697.5
+46.1+62.9+6.8+5.7+8.3+9.0+8.1+16.5+15.0+20.5+22.3+20.2+8.9+5.0+16.4+12.2
Qwen2-VL-7B24.026.943.368.353.856.628.560.261.935.624.954.792.684.493.793.2
+4-shot-ICL40.452.937.667.138.241.315.242.445.622.725.229.443.543.859.448.4
+MTV32.321.941.968.554.857.329.763.564.037.040.752.391.791.294.394.1
+LoRA84.887.746.370.855.357.428.865.266.140.438.472.998.496.497.195.0
+SAVs85.196.047.268.357.660.932.370.071.042.547.579.998.197.699.898.7
+61.1+69.1+3.9+0+3.8+4.3+3.8+9.8+9.1+13.8+22.6+25.2+5.5+13.2+6.1+5.5
+ +Table 1. Results evaluation on Safety, Visual Question Answering (VQA), Image-Text Retrieval (I-T Retrieval), and Image Classification benchmarks. The best result for each generative model is shown in bold and the second best in underline. We gray out additional baselines. We note that CLIP and SigLIP cannot be evaluating directly on tasks with interleaved image-text queries. † GPT-4o is a close-source model and as such, is shown just as an upperbound (since SAVs are not directly applicable). Key: NB - NaturalBench, SC - SUGARCREPE. + +differences between two similar images, making Natural-Bench especially challenging for any existing VL models. The class labels are $\mathcal{C} = \{\text{"A", "B"}\}$ . As suggested in the paper, we evaluate "question accuracy" which awards one point if a model correctly answers a question for both images, "image accuracy" which awards a point when a model answers both questions for an image, and finally "group accuracy" awards one point when a model correctly answers all four pairs. (3) Vizwiz [20] is VQA dataset designed to progress research in vision systems to assist blind and vision-impaired individuals. The dataset was collected by asking blind people to take pictures and record questions about the image. Although typically a generative task, we reformulate Vizwiz as first a classification task distinguishing answerable and unanswerable questions followed by standard response generation. The class labels are $\mathcal{C} = \{\text{"answerable", "unanswerable"}\}$ . + +Image-Text Retrieval. NaturalBench-Retrieval [44] and SUGARCREPE [28] both measure fine-grained semantic understanding in image-text pairs, with class labels $\mathcal{C} = \{\text{"Yes","No"}\}$ indicating whether an image-text pair is correctly matched. While SUGARCREPE presents one image with two captions (original and altered), NaturalBench-Retrieval adds complexity by using two similar images with two corresponding captions, effectively eliminating lan + +guage bias and requiring models to capture more nuanced visual-semantic relationships. + +Image Classification. Image classification tasks evaluate a model's ability to categorize images into predefined classes. We evaluate on several standard classification benchmarks: EuroSAT [22] (satellite imagery for land use classification), Oxford-IIIT-Pets [75] (pet breed identification with visually similar categories), Flowers [71] (flower species recognition), Caltech Birds (CUB) [96] (fine-grained bird species classification), and ImageNet-1k [10] (general object recognition). For each dataset except ImageNet, we formulate the task as a 4-way multiple choice question with labels $\mathcal{C} = \{\text{"Class 1", "Class 2", "Class 3", "Class 4"}\}$ . Due to model oversaturation in simpler formulations, ImageNet-1k is formulated as a 16-way classification problem. + +# 4.3. Baselines + +For our results, we utilized SAVs with 20 examples per label. We compared our method with multiple SOTA baselines, including GPT-4o [73]. As a closed-source model, GPT-4o is presented simply as a strong closed-source baseline. Furthermore, we present the results of classic LMMs, LLaVA-1.5 and InstructBLIP, which have been previously applied to image classification [111]. Zero-shot (ZS) baselines are implemented by querying the model directly and generating a response. For image-text retrieval, ZS uses the + +![](images/05605116cb7f6c8d344467a0426d347ec8e9163f545fc879c8a595883b12cdb0.jpg) + +![](images/b1c9470e44c6ce7d0af80d2f8e8caffe3444b5bfc10907139ff66cb2f83db399.jpg) +Hallucination Top Head Att. Vec. +Figure 3. Head Visualization. On the top row, we show the top-20 attention head locations for a given task. All the heads are indexed by their (head, layer) position, and the selected heads are highlighted. On the bottom row, we visualize the attention vector of few-shot examples for the top head of the given class with t-SNE clustering [26]. Each point represents the embedding of an input sample. + +![](images/4712e0fc0c0a1557f5d1850a96de892b23ef1e18a91749bd57466faabd0d933b.jpg) + +![](images/dd94dbdbfae19ace29169120ccef96730682d57d58d77d2113727445158b6e17.jpg) +Relative Depth Top Head Att. Vec. + +![](images/885ef9438ea4cda3ac22178c71e7606c33e55761b9be9b3d81f963995b6a0575.jpg) + +![](images/d007f3842e214eaaf127dc025ec5b879ba4526d7ef4ca0bf10d8ed961a8477ab.jpg) +Image Cls. Top Head Att. Vec. + +SOTA generative scoring method VQAScore [57]. In addition to ZS, we also compare to several test-time and finetuning few-shot methods (all with the exact same sample complexity as SAVs). For instance, we compare to the current SOTA multimodal few-shot method, MTV [30] as well as doing 4-shot ICL and LoRA [29] finetuning for each task. + +# 4.4. Results + +Results are shown in Table 1. An advantage of our method is its adaptability to any VL classification task that has image, text, or interleaved image-text inputs. Thus, we apply SAVs to a wide range of tasks in safety, VQA, imagetext retrieval and image classification. Furthermore, our approach even significantly improves over SOTA ZS, few-shot, and finetuning methods. One interesting observation is that few-shot ICL consistently deteriorates performance. We hypothesize that instructioning tuning of LMMs disrupts the few-shot prompting capabilities gained during pretraining in favor of structured instructions as shown + +in [14, 86]. On the other hand, SAVs overcome this by directly operating on the internal activations. Finally, not only is our method successful across a wide-range of tasks, but it also improves on challenging visual perception and compositional reasoning tasks (e.g., BLINK and NaturalBench) that all VL models struggle with. Please refer to our supplementary section in Section 7 for mor results. + +# 4.5. Ablations + +We perform a comprehensive ablation study of our method on MHaluBench, NaturalBench, and EuroSAT (see Table 2). For more ablations, please refer to Section 7.1 in the Supp. For all ablations, we use LLaVA-OneVision-7B. + +Varying number of examples. In Figure 4 (left), we examine the impact of varying the number of few-shot examples used in our method. Our primary results in Table 1 indicate that just 20 examples per label are necessary to yield state-of-the-art performance on a variety of VL classification tasks. This ablation shows that accuracy on these tasks + +![](images/4db5e75369ac156a903e916af66682b50f468d69d42a5bb067797b93b7fa1f98.jpg) +Figure 4. Scaling Property of SAVs. Performance of LLaVA-OneVision-7B + SAVs on varying number of few-shot examples per label (left). Performance of of LLaVA-OneVision-7B + SAVs on varying numbers of attention vectors used (right). + +![](images/5033e9e9949bf23a4d849a305d38d697674b415c51e7364dbd5cbd3df9cb5afa.jpg) + +can scale with increasing numbers of examples per label. + +Varying number of attention vectors. In order to extract attention vectors, we select a very sparse set of heads from hundreds. We vary the number of attention vectors selected in Step 2 of our method and demonstrate that just 20 vectors are enough to realize nearly all of the classification accuracy of our method. Results are shown in Figure 4 (right). + +SAVs are flexible to different evaluation strategies. In our method, we leverage class centroid classification as the evaluation method for selecting sparse features. We view this flexibility of the sparsification method to be a key feature of our work. As such, we compare our class centroid classification approach to k-nearest neighbors (KNN) and linear probing. For linear probing, we train a lightweight MLP module for 20 epochs using the top heads' features. All methods make use of the same 20 examples per label for consistency. Our results in Table 3a show that class centroid classification outperforms both KNN classification and is comparable with linear probing. + +Comparing heads vs. layers. Based on prior work and transformer-architecture intuition, we treat the attention vectors outputted by the heads as a viable set of features for classification tasks. We verify this intuition by comparing the performance of selecting 2 sparse layers to selecting sparse heads as feature maps for our tasks. As shown in Table 3b, head-based attention vectors outperform concatenated layer features on all three benchmarks. + +# 4.6. Additional Experiments + +In this subsection, we present experiments that demonstrate additional properties and capabilities of SAVs, beyond its use as features for VL classification tasks. Additional experiments can be found in Section 7.2 of the Supplementary. For all experiments, we use LLaVA-OneVision-7B. + +Visualizing SAVs. SAVs are both an efficient and interpretable method for leveraging generative LMMs for VL + +classification tasks. To emphasize this point, we show the selected heads for hallucination detection, relative depth, and image classification in the first row of Figure 3. The visualizations demonstrate both the sparsity and specificity of the SAVs that are used for each task. Unlike prompting and finetuning methods, our approach clearly outlines exactly where in the model's activation space informative attention vector features are extracted from. We furthermore show that the features extracted from these heads are useful for the given task in the second row of the figure, where we visualize the features outputted by the top selected head for each few-shot sample via t-SNE [26]. The clear clustering of examples of the same label indicates that even with a single head, high-quality features are being selected as SAVs. + +Evaluating SAVs generalizability. Here, we ask whether SAVs extracted from one task, can generalize to another similar task. We utilize SAV heads from VLGuard to evaluate on MhaluBench and vice versa. We do a similar approach with LoRA, by applying LoRA finetuned on VLGuard to MHaluBench. Interestingly, our results in Table 2a show that SAVs generalize between tasks, while LoRA weights, as expected, overfit to the finetuned task. + +Comparing SAVs to CLIP/SigLIP on interleaved imagetext tasks. As discussed in Section 1, SAVs are fully multimodal features able to represent inputs that are image-only, text-only, and even interleaved image-text. This is something that is not possible to directly replicate with CLIP and SigLIP models which have separate image and text encoders. Nevertheless, we compare our method to both CLIP and SigLIP on tasks that require interleaved imagetext inputs. While SAVs can do this natively, we enable this comparison by concatenating the separate image and text features of CLIP and SigLIP in order to evaluate on MHaluBench and NaturalBench. We find that our method vastly outperforms concatenated CLIP and SigLIP features on both benchmarks as shown in Figure 2b. This result + +(a) Generalization + +
MHBVLG
Zero-Shot34.731.4
VLG LoRA34.190.0
VLG SAV55.394.3
+ +(b) Interleaved Tasks + +
MHBNB
CLIP51.91.2
SigLIP48.61.2
SAVs80.835.1
+ +(c) SAVs vs Few-Shot SigLIP + +
NBES
SigLIP Few-shot Mean1.288.2
SigLIP Few-shot Cross-Modal6.886.4
SAVs35.186.7
+ +Table 2. Additional SAVs Experiments. We (a) demonstrate the generalization of SAV heads to similar tasks, (b) show the effectiveness of SAV for tasks with interleaved image-text inputs, and (c) compare SAVs with few-shot formulations of CLIP and SigLIP. +(a) Classification Methods + +
MHBNBES
Class Centroid80.835.186.7
KNN53.011.078.1
Linear Probe82.532.983.1
+ +(b) Sparse Configurations + +
MHBNBES
Sparse Heads80.835.186.7
Sparse Layers79.028.481.8
+ +Table 3. SAVs Ablations. We perform several ablations to identify the important aspects of our method that contribute to its effectiveness. In particular, we compare the effectiveness of (a) different classification methods and (b) head feature sparsification versus layer feature sparsification. Note: MHB represents MHaluBench, NB represents NaturalBench Group Score, and ES represents EuroSAT. For more ablations, please refer to Section 7.1 in the Supplementary. + +demonstrates the adaptability of our method to any VL classification regardless of the input's modality. + +Comparing SAVs to few-shot SigCLIP. As SAVs are extracted with few-shot examples, we compare our method to an analogous version of few-shot SigLIP. However, because SigLIP cannot be directly made few-shot, we adapt SigLIP as a few-shot class centroid classifier. One method used is the current SOTA few-shot classification method for encoder models Cross-Modal Adaptation [56]. We apply a second method where we aggregate CLIP/SigLIP embeddings into a mean embedding for each label. Then, just as in SAVs, we perform class centroid classification for each query using our set of mean SigLIP embeddings. We note that for image classification like EuroSAT, only image embeddings are needed, but for MHaluBench, multimodal embeddings are necessary. SAVs are inherently multimodal and so can be flexibly applied to both, but CLIP/SigLIP only have image-only or text-only embeddings. To overcome this, we use SigLIP image features for EuroSAT and concatenate image and text features for MHaluBench. Interestingly, while SiGLIP is comparable to SAVs on EuroSAT, our method vastly outperforms SigLIP in the few-shot setting for MHaluBench, suggesting generalizability of our method for a variety of multimodal tasks that CLIP-like struggle with. Our results are shown in Table 2c. + +# 5. Conclusion + +Our research demonstrates the effectiveness of extracting Sparse Attention Vectors (SAVs) from the heads of an LMM and utilizing them directly for vision-language classification. Our method stands out by using only few-shot exam + +plies per label and only less than $1\%$ of the heads to outperform zero-shot, few-shot, and fine-tuned baselines on a variety of image-text and image-only tasks. In addition, SAVs allows generative LMMs to close the gap with closed-source GPT-4o while also being an interpretable method that can generalize to similar tasks. Our ablations reveal the flexibility of using any classification method as a sparsification method for attention vectors and also shows that features are found as outputs of heads rather than layers. Overall, these results show that SAVs is a lightweight, performant, and generalizable method for extending generative LMMs' multimodal classification abilities. We are encouraged by the outcomes, and anticipate many directions for future work. In addition to methodological improvements, we look forward to the application of SAVs as features for multimodal retrieval, data compression, or more generally as a distilled representation for downstream models. + +# 6. Limitations + +Sparse Attention Vectors are a significant step in generalizing the capabilities of generative LMMs to classification tasks. Nevertheless, it is valuable to consider certain limitations of our approach. SAVs are a method that requires access to the model's internal architecture and so may not be directly applicable to closed-source models like GPT-4 [73] and Gemini [89, 90]. Additionally, some tasks like image-text retrieval [28, 91] can benefit from more fine-grained confidence metrics attached to each label than proportion of voting heads per label. These challenges prompt future work in these directions as well as exciting questions about how to use SAVs as feature embeddings for other tasks. + +# Acknowledgements. + +We would like to thank Abrar Anwar and Tyler Bonnen for helpful feedback and discussions. This project was supported in part by DoD, including PTG and/or LwLL programs, as well as BAIR's industrial alliance programs. + +# References + +[1] S. R. Bowman A. Williams, N. Nangia. A broad-coverage challenge corpus for sentence understanding through inference. ArXiv, 2017. 2 +[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1 +[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv, abs/2308.12966, 2023. 1 +[4] Pierre Baldi. Autoencoders, unsupervised learning, and deep architectures. In ICML 2011 Unsupervised and Transfer Learning Workshop, 2011. 2 +[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3 +[6] Matyas Bohacek and Michal Bravansky. When XGBoost outperforms GPT-4 on text classification: A case study. In Trustworthy Natural Language Processing (TrustNLP) 4th Workshop, pages 51–60, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 2 +[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. *ArXiv*, abs/2005.14165, 2020. 2 +[8] Martin Juan José E Bucher and Marco Martini. Finetuned 'small' llms (still) significantly outperform zero-shot generative ai models in text classification. ArXiv, abs/2406.08660, 2024. 1, 2 +[9] Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Jinjie Gu, and Huajun Chen. Unified hallucination detection for multimodal large language models. arXiv preprint arXiv:2402.03190, 2024. 4, 3 +[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5 +[11] Xiang Deng, Vasilisa Bashlovkina, Feng Han, Simon Baumgartner, and Michael Bendersky. Llms to the moon? + +reddit market sentiment analysis with large language models. Companion Proceedings of the ACM Web Conference 2023, 2023. 2 +[12] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11162-11173, 2021. 2 +[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805, 2019. 1, 2 +[14] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. ArXiv, abs/2403.12736, 2024. 6 +[15] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Herv'e J'egou. Training vision transformers for image retrieval. ArXiv, abs/2102.05644, 2021. 1, 2 +[16] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2 +[17] Evelina Fedorenko, Michael K. Behr, and Nancy Kanwisher. Functional specificity for high-level linguistic processing in the human brain. Proceedings of the National Academy of Sciences, 108(39):16428-16433, 2011. 2 +[18] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 2, 4, 3 +[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2 +[20] Danna Gurari, Qing Li, Abigale Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 5 +[21] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2 +[22] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5 +[23] Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors. ArXiv, abs/2310.15916, 2023. 2 +[24] Roee Hendel, Mor Geva, and Amir Globerson. Incontext learning creates task vectors. arXiv preprint arXiv:2310.15916, 2023. 2 +[25] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson. + +Incorporating structured representations into pretrained vision $\&$ language models using scene graphs. In 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2 +[26] L. Hinton G, van der Maaten. Visualizing data using t-sne. Journal of machine learning research, 2008. 6, 7 +[27] Alberto Hojel, Yutong Bai, Trevor Darrell, Amir Globerson, and Amir Bar. Finding visual task vectors. In European Conference on Computer Vision (ECCV), pages 257-273. Springer, 2025. 2 +[28] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. ArXiv, abs/2306.14610, 2023. 5, 8 +[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6 +[30] Brandon Huang, Chancharik Mitra, Assaf Arbelle, Leonid Karlinsky, Trevor Darrell, and Roei Herzig. Multimodal task vectors enable many-shot multimodal in-context learning. arXiv preprint arXiv:2406.15334, 2024. 2, 6, 3 +[31] Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2 +[32] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6693-6702, 2019. 1 +[33] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 2 +[34] Ting Jiang, Shaohan Huang, Zi qiang Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Liangjie Zhang, and Qi Zhang. Promptbert: Improving bert sentence embeddings with prompts. In Conference on Empirical Methods in Natural Language Processing, 2022. 2 +[35] Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. Scaling sentence embeddings with large language models. ArXiv, abs/2307.16645, 2023. 2 +[36] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. ArXiv, abs/2407.12580, 2024. 2 +[37] Nancy Kanwisher. Domain specificity in face perception. Nature neuroscience, 3(8):759-763, 2000. 2 +[38] Sahar Kazemzadeh, Vicente Ordonez, Marc andre Matten, and Tamara L. Berg. Referitgame: Referring to objects in photographs of natural scenes. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 1 +[39] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 2 + +[40] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2 +[41] Yibin Lei, Di Wu, Tianyi Zhou, Tao Shen, Yu Cao, Chongyang Tao, and Andrew Yates. Meta-task prompting elicits embeddings from large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 2 +[42] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2 +[43] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv, abs/2307.16125, 2023. 1 +[44] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Natural-bench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 2, 4, 5 +[45] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, abs/2408.03326, 2024. 4 +[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2 +[47] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2 +[48] Shiyu Li, Yang Tang, Shizhe Chen, and Xi Chen. Conan-embedding: General text embedding with more and better negative samples. ArXiv, abs/2408.15710, 2024. 2 +[49] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2 +[50] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23390-23400, 2022. 2 +[51] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2 +[52] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2 +[53] Jiacheng Lin, Kun Qian, Haoyu Han, Nurendra Choudhary, Tianxin Wei, Zhongruo Wang, Sahika Genc, Edward W + +Huang, Sheng Wang, Karthik Subbian, et al. Unleashing the power of lms as multi-modal encoders for text and graph-structured data. arXiv preprint arXiv:2410.11235, 2024. 2 +[54] Tsung-Yi Lin, M. Maire, Serge J. Belongie, James Hays, P. Perona, D. Ramanan, Piotr Dollar, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), 2014. 1, 3 +[55] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 2 +[56] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 8 +[57] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2024. 6 +[58] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2025. 2 +[59] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1 +[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3 +[61] Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera, Luca Zancato, and Stefan O Soatto. Meaning representations from trajectories in autoregressive models. ArXiv, abs/2310.18348, 2023. 2 +[62] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 2020. 2 +[63] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 33:5081–5092, 2020. 2 +[64] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. ArXiv, abs/2405.20797, 2024. 2 +[65] Junhua Mao, Jonathan Huang, Alexander Toshev, OanaMaria Camburu, Alan Loddon Yuille, and Kevin P. Murphy. Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2015. 1 +[66] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 2 +[67] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain of thought prompting + +for large multimodal models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[68] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. In Conference of the European Chapter of the Association for Computational Linguistics, 2022. 2 +[69] Jianmo Ni, Gustavo Hernández Abrego, Noah Constant, Ji Ma, Keith B. Hall, Daniel Matthew Cer, and Yinfei Yang. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. ArXiv, abs/2108.08877, 2021. 2 +[70] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith B. Hall, Ming-Wei Chang, and Yinfei Yang. Large dual encoders are generalizable retrievers. ArXiv, abs/2112.07899, 2021. 2 +[71] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5 +[72] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 2 +[73] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 1, 5, 8, 3 +[74] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 2 +[75] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5, 6 +[76] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 3 +[77] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2 +[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2 +[79] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a + +reward model. Advances in Neural Information Processing Systems, 36, 2024. 2 +[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2 +[81] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 1, 2 +[82] David E Rumelhart, James L McClelland, PDP Research Group, et al. Parallel distributed processing, volume 1: Explorations in the microstructure of cognition: Foundations. The MIT press, 1986. 2 +[83] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815-823, 2015. 1, 2 +[84] Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014. 2 +[85] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8309-8318, 2019. 3 +[86] Aaditya K. Singh, Stephanie C.Y. Chan, Ted Moskovitz, Erin Grant, Andrew M. Saxe, and Felix Hill. The transient nature of emergent in-context learning in transformers. In Advances in neural information processing systems (NeurIPS), 2023. 6 +[87] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2013. 2 +[88] Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. Text classification via large language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2 +[89] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. ArXiv, abs/2403.05530, 2024. 8 +[90] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 8 +[91] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 8 +[92] Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau. Function vectors in large language models. ArXiv, abs/2310.15213, 2023. 2 + +[93] Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. Function vectors in large language models. arXiv preprint arXiv:2310.15213, 2023. 2 +[94] Matthew A. Turk and Alex Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3:71-86, 1991. 1, 2 +[95] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. ArXiv, abs/1711.00937, 2017. 2 +[96] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5 +[97] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. ArXiv, abs/2401.00368, 2023. 2 +[98] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Ke-Yang Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. ArXiv, abs/2409.12191, 2024. 4 +[99] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. arXiv preprint arXiv:2401.11206, 2024. 2 +[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3 +[101] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. Internvideo2: Scaling video foundation models for multimodal video understanding. ArXiv, abs/2403.15377, 2024. 2 +[102] Zhiqiang Wang, Yiran Pang, and Yanbin Lin. Large language models are zero-shot text classifiers. ArXiv, abs/2312.01044, 2023. 2 +[103] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. ArXiv, abs/2109.01652, 2021. 2 +[104] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 2 +[105] Dean Wyatte, Fatemeh Tahmasbi, Ming Li, and Thomas Markovich. Scaling laws for discriminative classification in large language models. ArXiv, abs/2405.15765, 2024. 2 +[106] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng + +Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 3 +[107] Peter Young, Alice Lai, Micah Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1 +[108] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 2 +[109] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2 +[110] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and C. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine Learning in Health Care, 2020. 2 +[111] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 1, 2, 5 +[112] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. ArXiv, abs/2309.07915, 2023. 2 +[113] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3 +[114] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 2, 4 \ No newline at end of file diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/images.zip b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3634118f6f986270da19e211cc027588f82a534b --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a828b405018c60b4513bc4fb965f6bc1d98f4fa8efabf9149fa5b35bf04bc278 +size 491314 diff --git a/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/layout.json b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..c6952b64e897d472eae9027d897eff096bdd358c --- /dev/null +++ b/2025/Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features/layout.json @@ -0,0 +1,10600 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 541, + 138 + ], + "type": "text", + "content": "Enhancing Few-Shot Vision-Language Classification with Large Multimodal Model Features" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": "Chancharik Mitra" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Brandon Huang" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Tianning Chai" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Zhiqiu Lin" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Assaf Arbelle" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Rogerio Feris" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Leonid Karlinsky" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Trevor Darrell" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Deva Ramanan" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "text", + "content": " Roei Herzig" + }, + { + "bbox": [ + 83, + 160, + 526, + 191 + ], + "type": "inline_equation", + "content": "^{2,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "spans": [ + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "text", + "content": "Carnegie Mellon University " + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "text", + "content": "University of California, Berkeley " + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "text", + "content": "IBM Research " + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 143, + 202, + 468, + 232 + ], + "type": "text", + "content": "MIT-IBM Watson AI Lab" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 284, + 297, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 536 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 536 + ], + "type": "text", + "content": "Generative Large Multimodal Models (LMMs) like LLaVA and Qwen-VL excel at a wide variety of vision-language (VL) tasks. Despite strong performance, LMMs' generative outputs are not specialized for vision-language classification tasks (i.e., tasks with vision-language inputs and discrete labels) such as image classification and multiple-choice VQA. One key challenge in utilizing LMMs for these tasks is the extraction of useful features from generative LMMs. To overcome this, we propose an approach that leverages multimodal feature extraction from the LMM's latent space. Toward this end, we present Sparse Attention Vectors (SAVs)—a finetuning-free method that leverages sparse attention head activations (fewer than " + }, + { + "bbox": [ + 54, + 284, + 297, + 536 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 284, + 297, + 536 + ], + "type": "text", + "content": " of the heads) in LMMs as strong feature representations. With only few-shot examples, SAVs demonstrate state-of-the-art performance compared to a variety of few-shot and finetuned baselines on a collection of vision-language classification tasks. Our experiments also imply that SAVs can scale in performance with additional examples and generalize to similar tasks, establishing SAVs as both effective and robust multimodal feature representations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 296, + 714 + ], + "type": "text", + "content": "Generative Large Multimodal Models (LMMs) such as GPT-4V [73], LLaVA [59, 60], and QwenVL [3] demonstrate state-of-the-art performance on open-ended vision-language (VL) tasks like image captioning [54, 107], visual question answering [2, 32, 43], and language grounding [38, 65]. However, despite their remarkable performance on generative tasks, these models struggle on vision-language classification tasks, where responses are a discrete set of labels [8, 111]. Indeed, LMMs with billions of parameters and trained on trillions more tokens of data underperform smaller encoder VLMs like CLIP and SigLIP [8, 111]" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 258, + 432, + 466 + ], + "blocks": [ + { + "bbox": [ + 316, + 258, + 432, + 466 + ], + "lines": [ + { + "bbox": [ + 316, + 258, + 432, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 432, + 466 + ], + "type": "image", + "image_path": "666b706b0f887295d28038f80fd97824ec1f34ab370bb4485266bf6b575b1e24.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 487, + 555, + 554 + ], + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 554 + ], + "type": "text", + "content": "Figure 1. Sparse Attention Vectors (SAVs) Overview. We develop a method for extracting features from a generative LMM without finetuning. We first extract a sparse set of attention vectors for each task given a set of few-shot examples, and then, we utilize these attention vectors directly as features for downstream vision-language classification tasks." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 434, + 258, + 552, + 477 + ], + "blocks": [ + { + "bbox": [ + 434, + 258, + 552, + 477 + ], + "lines": [ + { + "bbox": [ + 434, + 258, + 552, + 477 + ], + "spans": [ + { + "bbox": [ + 434, + 258, + 552, + 477 + ], + "type": "image", + "image_path": "86cf6fc0bf9307c7557fbc168f79bbd5df228af602254180e66b6440ec931564.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 555, + 675 + ], + "type": "text", + "content": "and even classical machine learning methods [6] on image classification tasks. One reason that CLIP-like models excel at classification is the ease of extracting visual or textual features from modality-specific encoders. However, such models cannot process joint vision-language inputs, as LMMs can. Our goal is to extract compact features from generative LMMs, enabling downstream discrete labeling tasks such as classification and multiple-choice VQA." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Feature extraction is a well-explored field in both vision-only [15, 83, 94] and language-only encoder models [13, 81], but such is not the case with generative models. Most" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2760" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 192 + ], + "type": "text", + "content": "current methods for extracting features from generative models require carefully constructed prompts [41, 67, 104], specialized architectures [51], few-shot prompting [7, 11, 112], and finetuning [8, 64, 111]. Recent work [111] shows that prompting approaches fail to close the gap with encoder VLMs, while finetuning requires training-scale data for every unseen task, which is inefficient. However, as we observe in our work, generative models still offer the promise of more flexible, truly multimodal features as compared to modality-specific features from CLIP-like models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 198, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 413 + ], + "type": "text", + "content": "A source of inspiration for our method is long-standing work in neuroscience that suggests certain areas of the brain are reserved for specific tasks [17, 37] (i.e. functional specificity). Motivated by this idea, we refer to recent interpretability research that has focused on identifying specific heads in transformer-based models that correspond to particular tasks [72]. The most prominent of these methods is a line of work that looks to enhance vision-language capabilities using task vectors [24, 27, 30, 93], which are compact implicit representations of tasks encoded in the activations of a transformer model. While promising, these methods ultimately use these representations to augment a model's generative capabilities. On the other hand, we seek to use feature representations directly as classifiers. Nevertheless, this intuition from interpretability informs our work on Sparse Attention Vectors (SAVs), which are sparse features in an LMMs activation space that can be directly exploited for few-shot VL classification." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "text", + "content": "Our method has three steps: First, we extract features (i.e., attention vectors) from the output of each head of the LMM for some few-shot labeled examples (" + }, + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "inline_equation", + "content": "\\approx 20" + }, + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "text", + "content": " per label). Second, we average these attention vectors over the examples in each class and evaluate their accuracy as centroids in a class centroid classifier. We then select the top 20 heads by classification accuracy as our SAVs. In this way, we identify a very sparse set of attention vectors (less than " + }, + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 420, + 295, + 588 + ], + "type": "text", + "content": " of the total number of heads) that can be used for discriminative tasks. Finally, we perform inference on the given task by doing a majority vote across this sparse set of attention vectors for each new query. This approach requires only few-shot examples at test-time to extract effective multimodal embeddings. An overview is shown in Figure 1." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "We summarize our main contributions as follows: (i) We introduce a novel method that yields a sparse set of attention vectors to serve as highly effective features for individual VL classification tasks; (ii) Our method surpasses zero-shot, few-shot, and LoRA fine-tuned baselines across multiple tasks (" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "+7\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " improvement on average over LoRA on challenging benchmarks like BLINK [18], VLGuard [114], and NaturalBench [44]); (iii) We establish several advantageous properties of our approach, including strong generalization capabilities and favorable scaling characteristics." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 91, + 553, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 295 + ], + "type": "text", + "content": "Controllable Generation for Classification. Controllable text generation in LMMs guides model outputs to meet specific constraints. For classification tasks with generative LMMs, several approaches exist: test-time hard prompting [7, 104] uses prompt engineering or few-shot examples to elicit class label outputs [11, 88, 102, 105, 111]; direct probability analysis of generated class labels [55, 58] for image-text retrieval; and soft prompting methods that finetune learnable tokens [42, 49]. Other techniques include instruction finetuning [103] on labeled data [6, 111] and preference modeling like DPO and RLHF [16, 74, 79]. Our method, however, is finetuning-free and directly selects class labels without preference data. Most related work shows that internal transformer representations called task vectors [23, 27, 30, 92] can encapsulate ICL example tasks. Beyond previous approaches, we use a sparse set of attention vectors directly as VL classification task task features." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 296, + 553, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 553, + 415 + ], + "type": "text", + "content": "Vision-Language Features. Feature extraction seeks useful representations for diverse downstream tasks. Early embedding techniques including autoencoders [4, 39, 62, 63, 82], Word2Vec [66] and GloVe [77] transformed inputs into vector representations, followed by advances in NLP [13, 19, 68, 81] and vision [15, 83, 94]. Recent methods like CLIP and SigCLIP [46, 47, 50, 78, 108, 109] explore multimodal correlations through contrastive learning or sigmoid loss. These representations offer flexibility across tasks [12, 25, 33, 40, 80, 95] and domains [53, 101, 110]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 416, + 553, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 553, + 548 + ], + "type": "text", + "content": "Interestingly, extracting features from generative models poses unique challenges in identifying optimal extraction points. Some approaches finetune encoder VLMs on LLM-generated data [48, 97] or finetune LMMs directly on specific tasks [31, 111]. More efficient methods finetune encoder-decoder representations for better alignment [36, 64, 69, 70]. Finetuning-free approaches use distillation prompts to extract representations from model weights or activations [34, 35, 41, 61], while other methods employ mixture-of-experts [52], expert models [99], or embedding reranking [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 548, + 553, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 553, + 632 + ], + "type": "text", + "content": "Thus, current SOTA faces the following challenges: (1) modality-specific rather than multimodal features, (2) requiring finetuning, (3) limited flexibility due to specialized prompts, and (4) dependence on multiple models. Our approach provides effective multimodal embeddings without any gradient-based finetuning and flexibly apply to various VL classification tasks without additional models." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 645, + 374, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 645, + 374, + 657 + ], + "spans": [ + { + "bbox": [ + 314, + 645, + 374, + 657 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "In this section, we outline our approach for using sparse attention vectors from the activation space of a transformer-based large multimodal model (LMM) as features for any VL classification task. The method consists of three main" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 313, + 757 + ], + "type": "text", + "content": "2761" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 81, + 553, + 264 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 181, + 79 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 181, + 79 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 181, + 79 + ], + "type": "text", + "content": "1. Extract Attention Vectors" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 65, + 81, + 553, + 264 + ], + "lines": [ + { + "bbox": [ + 65, + 81, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 65, + 81, + 553, + 264 + ], + "type": "image", + "image_path": "2cd40e211ec3da687d4de37aa8df1d752ce1602a2582550ee7c4548b4ee2391b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 304, + 70, + 463, + 80 + ], + "lines": [ + { + "bbox": [ + 304, + 70, + 463, + 80 + ], + "spans": [ + { + "bbox": [ + 304, + 70, + 463, + 80 + ], + "type": "text", + "content": "2. Find Sparse Attention Vectors (SAVs):" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 304, + 89, + 517, + 99 + ], + "lines": [ + { + "bbox": [ + 304, + 89, + 517, + 99 + ], + "spans": [ + { + "bbox": [ + 304, + 89, + 517, + 99 + ], + "type": "text", + "content": "Select Sparse Attention Vectors" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 71, + 278, + 527, + 328 + ], + "blocks": [ + { + "bbox": [ + 236, + 266, + 367, + 277 + ], + "lines": [ + { + "bbox": [ + 236, + 266, + 367, + 277 + ], + "spans": [ + { + "bbox": [ + 236, + 266, + 367, + 277 + ], + "type": "text", + "content": "3. SAVs for Discriminative Tasks:" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 71, + 278, + 527, + 328 + ], + "lines": [ + { + "bbox": [ + 71, + 278, + 527, + 328 + ], + "spans": [ + { + "bbox": [ + 71, + 278, + 527, + 328 + ], + "type": "image", + "image_path": "4902bb64a69ee271e40c1203d2849b6ad4c020c4b2c3f22df58fca0e8bbbcaa5.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 339, + 555, + 395 + ], + "lines": [ + { + "bbox": [ + 54, + 339, + 555, + 395 + ], + "spans": [ + { + "bbox": [ + 54, + 339, + 555, + 395 + ], + "type": "text", + "content": "Figure 2. Sparse Attention Vectors (SAVs) Detailed View. Our method is broken into the following three parts: (1) Given a set of few shot examples to be classified by a frozen LMM, we extract attention vectors across all heads and layers. (2) These attention vectors are averaged across the set of examples for each class. For each head, we use these mean attention vectors as centroid classifiers, which are then used to select a sparse set of " + }, + { + "bbox": [ + 54, + 339, + 555, + 395 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 339, + 555, + 395 + ], + "type": "text", + "content": " heads with the highest classification accuracy. (3) Finally, we use these sparse attention vectors to directly classify new inputs via majority vote (denoted by the \"Classifier\" block)." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 294, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 294, + 523 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 294, + 523 + ], + "type": "text", + "content": "steps: (i) extracting the attention vectors from all attention heads in the model, (ii) identifying a sparse set of vectors based on their ability to consistently return the correct label for some support set of examples, and (iii) using these sparse features to classify new queries. We begin with a formal description of the transformer decoder LLM and its attention mechanism, followed by the detailed methodology for sparse attention vector selection and classification. A detailed view of our method is shown in Figure 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 141, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 141, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 141, + 555 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "text", + "content": "A transformer-based large language model (LLM) with " + }, + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "text", + "content": " layers and " + }, + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 565, + 295, + 638 + ], + "type": "text", + "content": " attention heads per layer processes input sequences through multi-head self-attention mechanisms. Each layer combines multiple attention heads to capture different aspects of the input sequence, followed by feedforward networks for further processing." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": "Multi-Head Attention. Let " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "x = \\{x_{1},x_{2},\\ldots ,x_{T}\\}" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " represent a sequence of input tokens, where " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "i^{\\text{th}}" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": " token. For each layer " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "l\\in \\{1,\\dots ,L\\}" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": ", the input sequence is projected into queries, keys, and values for each attention head " + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "inline_equation", + "content": "m\\in \\{1,\\dots ,H\\}" + }, + { + "bbox": [ + 54, + 641, + 295, + 714 + ], + "type": "text", + "content": ". Each head performs the following scaled dot-product attention:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 103, + 553, + 263 + ], + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 103, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 359, + 103, + 553, + 263 + ], + "type": "image", + "image_path": "6a415ec770099d3bc8758751273bab9fda41c11d81f9268f7f6b0c26fb658a42.jpg" + } + ] + } + ], + "index": 11, + "type": "text" + }, + { + "bbox": [ + 314, + 427, + 500, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 500, + 452 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 500, + 452 + ], + "type": "interline_equation", + "content": "\\mathbf {h} _ {l} ^ {m} \\left(x _ {i}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {m}}}\\right) V", + "image_path": "197c8ede5414294b5bc2aa296b681febc07ee7b3658fd2230f5c6af51c8821f2.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": " are the query, key, and value matrices respectively, and the dimensionality of each head " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "d_{m}" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": " which is given by " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\frac{d}{H}" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": " (the embedding dimension divided by the number of heads). We denote " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_l^m(x_i)" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": " as an attention vector for head " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": " in layer " + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 454, + 555, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 538 + ], + "type": "text", + "content": "The outputs of all heads are concatenated and projected to form the layer output:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 331, + 545, + 537, + 561 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 545, + 537, + 561 + ], + "spans": [ + { + "bbox": [ + 331, + 545, + 537, + 561 + ], + "type": "interline_equation", + "content": "\\operatorname {M u l t i H e a d} \\left(x _ {i}\\right) = \\operatorname {C o n c a t} \\left(\\mathbf {h} _ {l} ^ {1} \\left(x _ {i}\\right), \\dots , \\mathbf {h} _ {l} ^ {H} \\left(x _ {i}\\right)\\right) W ^ {O}", + "image_path": "82c1bf8de4d346bffff7487360fa892ec02acc5b4a17c38a3405972fced4265c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 568, + 487, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 487, + 580 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 487, + 580 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 568, + 487, + 580 + ], + "type": "inline_equation", + "content": "W^{O}" + }, + { + "bbox": [ + 313, + 568, + 487, + 580 + ], + "type": "text", + "content": " is the output projection matrix." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 664 + ], + "type": "text", + "content": "In our work, we look to leverage attention vectors for the purpose of vision-language classification tasks. Specifically, the attention vectors are used as latent representations of the inputs to both find attention heads in an LMM suited for a classification task and then perform downstream inference using those selected attention heads. We describe our method in detail in the sections that follow." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 454, + 685 + ], + "type": "text", + "content": "3.2. Sparse Attention Vectors" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Our key insight is that within the many attention heads and transformer layers of an LMM, there exists a sparse subset" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2762" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "that can serve as effective features for vision-language classification tasks. We present a three-step method to identify and utilize these features to build lightweight classifiers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 251 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": "Step 1: Extracting Attention Vectors. Given a frozen LMM and few-shot examples of sequence-label pairs " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "\\{(x_1, y_1), (x_2, y_2), \\ldots, (x_N, y_N)\\}" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " we first extract the attention vectors for each sequence " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "x_i" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": ". Specifically, we compute the attention vector " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_l^m(x_i^T)" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " for head " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " from layer " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " for the final token " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "x_i^T" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": ". This yields a set of attention vectors " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{h}_l^m(x_i^T) \\mid i = 1, \\dots, N\\}" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " for each head " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": " and layer " + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 108, + 295, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 251 + ], + "type": "text", + "content": "Step 2: Identifying Relevant Vectors. The central question is how to identify which attention vectors are naturally suited for the classification task at hand. We evaluate each vector's ability as a local classifier by computing its performance under a nearest class centroid classifier." + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 251, + 295, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 276 + ], + "type": "text", + "content": "Specifically, for each class " + }, + { + "bbox": [ + 55, + 251, + 295, + 276 + ], + "type": "inline_equation", + "content": "c \\in \\mathcal{C}" + }, + { + "bbox": [ + 55, + 251, + 295, + 276 + ], + "type": "text", + "content": ", compute its centroid (or mean) attention vector across the few shot examples:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 117, + 282, + 233, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 282, + 233, + 312 + ], + "spans": [ + { + "bbox": [ + 117, + 282, + 233, + 312 + ], + "type": "interline_equation", + "content": "\\mu_ {c} ^ {l, m} = \\frac {1}{| N _ {c} |} \\sum_ {j: y _ {j} = c} \\mathbf {h} _ {l} ^ {m} (x _ {j} ^ {T})", + "image_path": "d4660e16d03534e99c0f5b745777ccd583e225c9791fda8a1a4a77b87d2c7b6e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "inline_equation", + "content": "N_{c} = \\{j:y_{j} = c\\}" + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "text", + "content": " is the set of indices of examples with label " + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "text", + "content": ". For each input " + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 319, + 295, + 356 + ], + "type": "text", + "content": ", we compute its cosine similarity to each class centroid head:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 362, + 264, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 362, + 264, + 392 + ], + "spans": [ + { + "bbox": [ + 87, + 362, + 264, + 392 + ], + "type": "interline_equation", + "content": "s _ {l, m} \\left(x _ {i}, c\\right) = \\frac {\\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\cdot \\mu_ {c} ^ {l , m}}{\\left\\| \\mathbf {h} _ {l} ^ {m} \\left(x _ {i} ^ {T}\\right) \\right\\| \\left\\| \\mu_ {c} ^ {l , m} \\right\\|}, \\quad \\forall c \\in \\mathcal {C}", + "image_path": "b3270e0c5ebd16a1f689053b7962d9c98c8e8f1faba5a6de32228217d261ff22.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 295, + 422 + ], + "type": "text", + "content": "Next, we measure the classification accuracy of each head (i.e. local classifier):" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 429, + 233, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 429, + 233, + 461 + ], + "spans": [ + { + "bbox": [ + 117, + 429, + 233, + 461 + ], + "type": "interline_equation", + "content": "\\operatorname {s c o r e} (l, m) = \\sum_ {i = 1} ^ {N} \\mathbf {1} [ \\hat {y} = y _ {i} ]", + "image_path": "5ae97133faab3f46860b41148e768dd0329f02ba2757c260365f0455371bdc5a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "text", + "content": "where the nearest class centroid label is given as " + }, + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\operatorname*{argmax}_{c\\in \\mathcal{C}}s_{l,m}(x_i,c)" + }, + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\mathbf{1}[\\cdot ]" + }, + { + "bbox": [ + 55, + 468, + 295, + 499 + ], + "type": "text", + "content": " is the indicator function that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 499, + 294, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 499, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 294, + 510 + ], + "type": "text", + "content": "evaluates to 1 when the condition is true (and 0 otherwise)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "text", + "content": "We denote the set of " + }, + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "text", + "content": " top-scoring heads as " + }, + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{SAV}}" + }, + { + "bbox": [ + 55, + 510, + 259, + 522 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "spans": [ + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{SAV}} = \\{(l,m)\\mid \\mathrm{score}(l,m)" + }, + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "type": "text", + "content": " is among " + }, + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 59, + 529, + 290, + 544 + ], + "type": "text", + "content": " highest scores}" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 551, + 280, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 280, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 280, + 562 + ], + "type": "text", + "content": "Step 3: Classification with Sparse Attention Vectors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": "Given a query sequence " + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": " to classify, we leverage our sparse set of heads " + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{SAV}}" + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": " for prediction. For each head " + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "inline_equation", + "content": "(l,m)\\in \\mathcal{H}_{\\mathrm{SAV}}" + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": ", we compute the class centroid " + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "inline_equation", + "content": "\\mu_c^{l,m}" + }, + { + "bbox": [ + 55, + 563, + 295, + 611 + ], + "type": "text", + "content": " closest to the query as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 116, + 617, + 235, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 617, + 235, + 638 + ], + "spans": [ + { + "bbox": [ + 116, + 617, + 235, + 638 + ], + "type": "interline_equation", + "content": "\\hat{y}_{l,m} = \\operatorname *{arg max}_{c\\in \\mathcal{C}}s_{l,m}(Q^{T},c)", + "image_path": "47aa61172c28e5e503c7c94457373b0978d27642f0fb531ba7d333d78962b3ab.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "inline_equation", + "content": "s_{l,m}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "content": " is defined as in Step 2. Our final class prediction is that of a global classifier that counts the majority vote across all local classifiers (heads) in " + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{SAV}}" + }, + { + "bbox": [ + 55, + 645, + 295, + 681 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 111, + 689, + 240, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 689, + 240, + 715 + ], + "spans": [ + { + "bbox": [ + 111, + 689, + 240, + 715 + ], + "type": "interline_equation", + "content": "\\underset {y \\in \\mathcal {C}} {\\arg \\max } \\sum_ {(l, m) \\in \\mathcal {H} _ {\\mathrm {S A V}}} \\mathbf {1} [ \\hat {y} _ {l, m} = y ]", + "image_path": "42fc231fa1059581a241cecf72897991a0e59ca8dabc78ea1248e7cb673a5c41.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "This approach reveals a surprising capability of LMMs: with just a few carefully selected attention heads (" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "inline_equation", + "content": "|\\mathcal{H}_{\\mathrm{SAV}}| \\ll LH" + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "), we can transform a generative language model into a lightweight vision-language classifier. This finding suggests that classification-relevant features naturally emerge within specific attention heads during model pretraining. In theory, one could replace our local and global classifiers with more complex variants (e.g. linear models, MLPs), something we explore in 3a." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 190, + 386, + 202 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 190, + 386, + 202 + ], + "spans": [ + { + "bbox": [ + 314, + 190, + 386, + 202 + ], + "type": "text", + "content": "4. Evaluation" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 210, + 553, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 553, + 282 + ], + "type": "text", + "content": "We apply SAVs to two state-of-the-art LMMs—LLaVA-OneVision [45] and Qwen2-VL [98]. We also do a rigorous comparison of our method to strong few-shot and finetuning baselines on a variety of different image-text and image-only vision-language classification tasks covering safety, VQA, image-text retrieval, and simple classification." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 289, + 447, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 447, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 447, + 301 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 306, + 554, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 554, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 554, + 401 + ], + "type": "text", + "content": "We implemented our approach in PyTorch [76]. We use the official implementations of each model, and all of our experiments can be run on a single NVIDIA A100 GPU. We use the default hyperparameters provided by the original models' (LLaVA-OV and Qwen2-VL) codebases. These hyperparameters are detailed in Table 8. More details of the implementation is included in the supplementary in Section 8." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 409, + 395, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 409, + 395, + 420 + ], + "spans": [ + { + "bbox": [ + 314, + 409, + 395, + 420 + ], + "type": "text", + "content": "4.2. Benchmarks" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": "Safety. (1) LMM-Halucination [9] is a dataset which evaluates the hallucinations of the models when answering multimodal tasks. We report the raw classification accuracy of our method. Thus, the set of class labels for this task is given by " + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"hallucinating\",\"not hallucinating\"}\\}" + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": ". (2) VLGuard [114] is a dataset focusing on vision-language safety which identifies 4 main categories of harmful content: Privacy, Risky Behavior, Deception and Hateful Speech. VLGuard proposes Attack Success Rate (ASR) for evaluating unsafe inputs and Helpfulness for evaluating safe inputs. We simply reformat it as a classification task, where the set of class labels is given by " + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"safe\",\"unsafe\"}\\}" + }, + { + "bbox": [ + 313, + 426, + 554, + 569 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "VQA Datasets. In our work, we evaluate on VQA benchmarks, many of which can be formulated as a classification task. (1) BLINK [18] contains many tasks that are intuitive for humans but complicated for multimodal models such as multi-view reasoning, and visual similarity comparison. Since potential answers in BLINK are multiple choice, the class labels would be given as " + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"A\", \"B\", \"C\", \"D\"}\\}" + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": " (note: the number of labels depends on the possible number of options allowed for a task). (2) NaturalBench-VQA [44] is a compositional dataset collected from natural image-text corpora but validated with human filtering. Each sample of the dataset contains two questions on compositional" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2763" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 328 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 328 + ], + "type": "table", + "html": "
ModelImage-Text TasksImage-Only Tasks
SafetyVQAI-T RetrievalClassification
MHaluVLGuardBLINKVizWizNB (T)NB (D)NB (G)NB-r (T)NB-r (D)NB-r (G)SCEuroSATPetsImagenet-IkFlowerCUB
CLIP-------41.845.023.235.364.088.196.192.897.8
SigLip-------54.554.931.242.763.998.397.695.898.4
GPT-4o†39.445.759.058.164.666.439.665.067.040.548.370.598.398.399.099.3
LLaVA-1.544.331.038.050.037.743.812.736.742.712.224.926.343.020.651.849.0
Instruct-BLIP44.018.538.734.520.224.24.019.521.31.14.720.456.011.018.233.8
LLaVA-OV-7B34.731.445.060.452.053.327.056.258.032.115.366.588.192.583.285.3
+4-shot-ICL25.035.038.947.847.650.422.248.249.631.516.147.163.949.063.860.6
+MTV37.332.944.561.156.258.030.758.159.533.428.765.588.595.983.285.6
+LoRA78.390.047.063.158.660.932.459.460.335.430.485.096.893.291.291.8
+SAVs80.894.351.866.160.362.335.172.773.053.137.686.797.097.599.697.5
+46.1+62.9+6.8+5.7+8.3+9.0+8.1+16.5+15.0+20.5+22.3+20.2+8.9+5.0+16.4+12.2
Qwen2-VL-7B24.026.943.368.353.856.628.560.261.935.624.954.792.684.493.793.2
+4-shot-ICL40.452.937.667.138.241.315.242.445.622.725.229.443.543.859.448.4
+MTV32.321.941.968.554.857.329.763.564.037.040.752.391.791.294.394.1
+LoRA84.887.746.370.855.357.428.865.266.140.438.472.998.496.497.195.0
+SAVs85.196.047.268.357.660.932.370.071.042.547.579.998.197.699.898.7
+61.1+69.1+3.9+0+3.8+4.3+3.8+9.8+9.1+13.8+22.6+25.2+5.5+13.2+6.1+5.5
", + "image_path": "8cdfe9004740d5cd677fd2358c3698958316fd4f461d85805df5c7df17f8b785.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "lines": [ + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 555, + 381 + ], + "type": "text", + "content": "Table 1. Results evaluation on Safety, Visual Question Answering (VQA), Image-Text Retrieval (I-T Retrieval), and Image Classification benchmarks. The best result for each generative model is shown in bold and the second best in underline. We gray out additional baselines. We note that CLIP and SigLIP cannot be evaluating directly on tasks with interleaved image-text queries. † GPT-4o is a close-source model and as such, is shown just as an upperbound (since SAVs are not directly applicable). Key: NB - NaturalBench, SC - SUGARCREPE." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "text", + "content": "differences between two similar images, making Natural-Bench especially challenging for any existing VL models. The class labels are " + }, + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"A\", \"B\"}\\}" + }, + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "text", + "content": ". As suggested in the paper, we evaluate \"question accuracy\" which awards one point if a model correctly answers a question for both images, \"image accuracy\" which awards a point when a model answers both questions for an image, and finally \"group accuracy\" awards one point when a model correctly answers all four pairs. (3) Vizwiz [20] is VQA dataset designed to progress research in vision systems to assist blind and vision-impaired individuals. The dataset was collected by asking blind people to take pictures and record questions about the image. Although typically a generative task, we reformulate Vizwiz as first a classification task distinguishing answerable and unanswerable questions followed by standard response generation. The class labels are " + }, + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"answerable\", \"unanswerable\"}\\}" + }, + { + "bbox": [ + 54, + 402, + 295, + 606 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Image-Text Retrieval. NaturalBench-Retrieval [44] and SUGARCREPE [28] both measure fine-grained semantic understanding in image-text pairs, with class labels " + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"Yes\",\"No\"}\\}" + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": " indicating whether an image-text pair is correctly matched. While SUGARCREPE presents one image with two captions (original and altered), NaturalBench-Retrieval adds complexity by using two similar images with two corresponding captions, effectively eliminating lan" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 403, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 426 + ], + "type": "text", + "content": "guage bias and requiring models to capture more nuanced visual-semantic relationships." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 427, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 582 + ], + "type": "text", + "content": "Image Classification. Image classification tasks evaluate a model's ability to categorize images into predefined classes. We evaluate on several standard classification benchmarks: EuroSAT [22] (satellite imagery for land use classification), Oxford-IIIT-Pets [75] (pet breed identification with visually similar categories), Flowers [71] (flower species recognition), Caltech Birds (CUB) [96] (fine-grained bird species classification), and ImageNet-1k [10] (general object recognition). For each dataset except ImageNet, we formulate the task as a 4-way multiple choice question with labels " + }, + { + "bbox": [ + 313, + 427, + 555, + 582 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{\\text{\"Class 1\", \"Class 2\", \"Class 3\", \"Class 4\"}\\}" + }, + { + "bbox": [ + 313, + 427, + 555, + 582 + ], + "type": "text", + "content": ". Due to model oversaturation in simpler formulations, ImageNet-1k is formulated as a 16-way classification problem." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 588, + 381, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 381, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 381, + 600 + ], + "type": "text", + "content": "4.3. Baselines" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 715 + ], + "type": "text", + "content": "For our results, we utilized SAVs with 20 examples per label. We compared our method with multiple SOTA baselines, including GPT-4o [73]. As a closed-source model, GPT-4o is presented simply as a strong closed-source baseline. Furthermore, we present the results of classic LMMs, LLaVA-1.5 and InstructBLIP, which have been previously applied to image classification [111]. Zero-shot (ZS) baselines are implemented by querying the model directly and generating a response. For image-text retrieval, ZS uses the" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2764" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 217, + 249 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 217, + 249 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 217, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 217, + 249 + ], + "type": "image", + "image_path": "05605116cb7f6c8d344467a0426d347ec8e9163f545fc879c8a595883b12cdb0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 261, + 216, + 411 + ], + "blocks": [ + { + "bbox": [ + 78, + 250, + 212, + 261 + ], + "lines": [ + { + "bbox": [ + 78, + 250, + 212, + 261 + ], + "spans": [ + { + "bbox": [ + 78, + 250, + 212, + 261 + ], + "type": "text", + "content": "Hallucination Top Head Att. Vec." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 77, + 261, + 216, + 411 + ], + "lines": [ + { + "bbox": [ + 77, + 261, + 216, + 411 + ], + "spans": [ + { + "bbox": [ + 77, + 261, + 216, + 411 + ], + "type": "image", + "image_path": "b1c9470e44c6ce7d0af80d2f8e8caffe3444b5bfc10907139ff66cb2f83db399.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 421, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 54, + 421, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 421, + 555, + 456 + ], + "type": "text", + "content": "Figure 3. Head Visualization. On the top row, we show the top-20 attention head locations for a given task. All the heads are indexed by their (head, layer) position, and the selected heads are highlighted. On the bottom row, we visualize the attention vector of few-shot examples for the top head of the given class with t-SNE clustering [26]. Each point represents the embedding of an input sample." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 233, + 70, + 386, + 249 + ], + "blocks": [ + { + "bbox": [ + 233, + 70, + 386, + 249 + ], + "lines": [ + { + "bbox": [ + 233, + 70, + 386, + 249 + ], + "spans": [ + { + "bbox": [ + 233, + 70, + 386, + 249 + ], + "type": "image", + "image_path": "4712e0fc0c0a1557f5d1850a96de892b23ef1e18a91749bd57466faabd0d933b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 244, + 261, + 384, + 411 + ], + "blocks": [ + { + "bbox": [ + 245, + 250, + 383, + 261 + ], + "lines": [ + { + "bbox": [ + 245, + 250, + 383, + 261 + ], + "spans": [ + { + "bbox": [ + 245, + 250, + 383, + 261 + ], + "type": "text", + "content": "Relative Depth Top Head Att. Vec." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 261, + 384, + 411 + ], + "lines": [ + { + "bbox": [ + 244, + 261, + 384, + 411 + ], + "spans": [ + { + "bbox": [ + 244, + 261, + 384, + 411 + ], + "type": "image", + "image_path": "dd94dbdbfae19ace29169120ccef96730682d57d58d77d2113727445158b6e17.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 402, + 70, + 553, + 249 + ], + "blocks": [ + { + "bbox": [ + 402, + 70, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 402, + 70, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 402, + 70, + 553, + 249 + ], + "type": "image", + "image_path": "885ef9438ea4cda3ac22178c71e7606c33e55761b9be9b3d81f963995b6a0575.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 413, + 261, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 422, + 250, + 543, + 261 + ], + "lines": [ + { + "bbox": [ + 422, + 250, + 543, + 261 + ], + "spans": [ + { + "bbox": [ + 422, + 250, + 543, + 261 + ], + "type": "text", + "content": "Image Cls. Top Head Att. Vec." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 413, + 261, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 413, + 261, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 413, + 261, + 553, + 411 + ], + "type": "image", + "image_path": "d007f3842e214eaaf127dc025ec5b879ba4526d7ef4ca0bf10d8ed961a8477ab.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 475, + 295, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 475, + 295, + 548 + ], + "spans": [ + { + "bbox": [ + 54, + 475, + 295, + 548 + ], + "type": "text", + "content": "SOTA generative scoring method VQAScore [57]. In addition to ZS, we also compare to several test-time and finetuning few-shot methods (all with the exact same sample complexity as SAVs). For instance, we compare to the current SOTA multimodal few-shot method, MTV [30] as well as doing 4-shot ICL and LoRA [29] finetuning for each task." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 562, + 113, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 113, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 113, + 574 + ], + "type": "text", + "content": "4.4. Results" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 713 + ], + "type": "text", + "content": "Results are shown in Table 1. An advantage of our method is its adaptability to any VL classification task that has image, text, or interleaved image-text inputs. Thus, we apply SAVs to a wide range of tasks in safety, VQA, imagetext retrieval and image classification. Furthermore, our approach even significantly improves over SOTA ZS, few-shot, and finetuning methods. One interesting observation is that few-shot ICL consistently deteriorates performance. We hypothesize that instructioning tuning of LMMs disrupts the few-shot prompting capabilities gained during pretraining in favor of structured instructions as shown" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 475, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 559 + ], + "type": "text", + "content": "in [14, 86]. On the other hand, SAVs overcome this by directly operating on the internal activations. Finally, not only is our method successful across a wide-range of tasks, but it also improves on challenging visual perception and compositional reasoning tasks (e.g., BLINK and NaturalBench) that all VL models struggle with. Please refer to our supplementary section in Section 7 for mor results." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 572, + 382, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 572, + 382, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 572, + 382, + 583 + ], + "type": "text", + "content": "4.5. Ablations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 592, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 555, + 640 + ], + "type": "text", + "content": "We perform a comprehensive ablation study of our method on MHaluBench, NaturalBench, and EuroSAT (see Table 2). For more ablations, please refer to Section 7.1 in the Supp. For all ablations, we use LLaVA-OneVision-7B." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Varying number of examples. In Figure 4 (left), we examine the impact of varying the number of few-shot examples used in our method. Our primary results in Table 1 indicate that just 20 examples per label are necessary to yield state-of-the-art performance on a variety of VL classification tasks. This ablation shows that accuracy on these tasks" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2765" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 77, + 301, + 219 + ], + "blocks": [ + { + "bbox": [ + 65, + 77, + 301, + 219 + ], + "lines": [ + { + "bbox": [ + 65, + 77, + 301, + 219 + ], + "spans": [ + { + "bbox": [ + 65, + 77, + 301, + 219 + ], + "type": "image", + "image_path": "4db5e75369ac156a903e916af66682b50f468d69d42a5bb067797b93b7fa1f98.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 237, + 555, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 237, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 237, + 555, + 262 + ], + "type": "text", + "content": "Figure 4. Scaling Property of SAVs. Performance of LLaVA-OneVision-7B + SAVs on varying number of few-shot examples per label (left). Performance of of LLaVA-OneVision-7B + SAVs on varying numbers of attention vectors used (right)." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 78, + 545, + 218 + ], + "blocks": [ + { + "bbox": [ + 306, + 78, + 545, + 218 + ], + "lines": [ + { + "bbox": [ + 306, + 78, + 545, + 218 + ], + "spans": [ + { + "bbox": [ + 306, + 78, + 545, + 218 + ], + "type": "image", + "image_path": "5033e9e9949bf23a4d849a305d38d697674b415c51e7364dbd5cbd3df9cb5afa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 281, + 286, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 281, + 286, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 281, + 286, + 293 + ], + "type": "text", + "content": "can scale with increasing numbers of examples per label." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 295, + 365 + ], + "type": "text", + "content": "Varying number of attention vectors. In order to extract attention vectors, we select a very sparse set of heads from hundreds. We vary the number of attention vectors selected in Step 2 of our method and demonstrate that just 20 vectors are enough to realize nearly all of the classification accuracy of our method. Results are shown in Figure 4 (right)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 295, + 508 + ], + "type": "text", + "content": "SAVs are flexible to different evaluation strategies. In our method, we leverage class centroid classification as the evaluation method for selecting sparse features. We view this flexibility of the sparsification method to be a key feature of our work. As such, we compare our class centroid classification approach to k-nearest neighbors (KNN) and linear probing. For linear probing, we train a lightweight MLP module for 20 epochs using the top heads' features. All methods make use of the same 20 examples per label for consistency. Our results in Table 3a show that class centroid classification outperforms both KNN classification and is comparable with linear probing." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 604 + ], + "type": "text", + "content": "Comparing heads vs. layers. Based on prior work and transformer-architecture intuition, we treat the attention vectors outputted by the heads as a viable set of features for classification tasks. We verify this intuition by comparing the performance of selecting 2 sparse layers to selecting sparse heads as feature maps for our tasks. As shown in Table 3b, head-based attention vectors outperform concatenated layer features on all three benchmarks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 612, + 190, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 190, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 190, + 624 + ], + "type": "text", + "content": "4.6. Additional Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 689 + ], + "type": "text", + "content": "In this subsection, we present experiments that demonstrate additional properties and capabilities of SAVs, beyond its use as features for VL classification tasks. Additional experiments can be found in Section 7.2 of the Supplementary. For all experiments, we use LLaVA-OneVision-7B." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Visualizing SAVs. SAVs are both an efficient and interpretable method for leveraging generative LMMs for VL" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 281, + 553, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 449 + ], + "type": "text", + "content": "classification tasks. To emphasize this point, we show the selected heads for hallucination detection, relative depth, and image classification in the first row of Figure 3. The visualizations demonstrate both the sparsity and specificity of the SAVs that are used for each task. Unlike prompting and finetuning methods, our approach clearly outlines exactly where in the model's activation space informative attention vector features are extracted from. We furthermore show that the features extracted from these heads are useful for the given task in the second row of the figure, where we visualize the features outputted by the top selected head for each few-shot sample via t-SNE [26]. The clear clustering of examples of the same label indicates that even with a single head, high-quality features are being selected as SAVs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 449, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 553, + 544 + ], + "type": "text", + "content": "Evaluating SAVs generalizability. Here, we ask whether SAVs extracted from one task, can generalize to another similar task. We utilize SAV heads from VLGuard to evaluate on MhaluBench and vice versa. We do a similar approach with LoRA, by applying LoRA finetuned on VLGuard to MHaluBench. Interestingly, our results in Table 2a show that SAVs generalize between tasks, while LoRA weights, as expected, overfit to the finetuned task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 713 + ], + "type": "text", + "content": "Comparing SAVs to CLIP/SigLIP on interleaved imagetext tasks. As discussed in Section 1, SAVs are fully multimodal features able to represent inputs that are image-only, text-only, and even interleaved image-text. This is something that is not possible to directly replicate with CLIP and SigLIP models which have separate image and text encoders. Nevertheless, we compare our method to both CLIP and SigLIP on tasks that require interleaved imagetext inputs. While SAVs can do this natively, we enable this comparison by concatenating the separate image and text features of CLIP and SigLIP in order to evaluate on MHaluBench and NaturalBench. We find that our method vastly outperforms concatenated CLIP and SigLIP features on both benchmarks as shown in Figure 2b. This result" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2766" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 79, + 206, + 145 + ], + "blocks": [ + { + "bbox": [ + 102, + 69, + 162, + 78 + ], + "lines": [ + { + "bbox": [ + 102, + 69, + 162, + 78 + ], + "spans": [ + { + "bbox": [ + 102, + 69, + 162, + 78 + ], + "type": "text", + "content": "(a) Generalization" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 79, + 206, + 145 + ], + "lines": [ + { + "bbox": [ + 59, + 79, + 206, + 145 + ], + "spans": [ + { + "bbox": [ + 59, + 79, + 206, + 145 + ], + "type": "table", + "html": "
MHBVLG
Zero-Shot34.731.4
VLG LoRA34.190.0
VLG SAV55.394.3
", + "image_path": "e5fc9b2cae56711f480565a887a0833d39d3b34481188fa311460bb29776dd61.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 227, + 79, + 352, + 145 + ], + "blocks": [ + { + "bbox": [ + 255, + 69, + 323, + 78 + ], + "lines": [ + { + "bbox": [ + 255, + 69, + 323, + 78 + ], + "spans": [ + { + "bbox": [ + 255, + 69, + 323, + 78 + ], + "type": "text", + "content": "(b) Interleaved Tasks" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 227, + 79, + 352, + 145 + ], + "lines": [ + { + "bbox": [ + 227, + 79, + 352, + 145 + ], + "spans": [ + { + "bbox": [ + 227, + 79, + 352, + 145 + ], + "type": "table", + "html": "
MHBNB
CLIP51.91.2
SigLIP48.61.2
SAVs80.835.1
", + "image_path": "303d1f8de34fb67ccc42d68ed35051b4db279a26dbc9f5eacf0b16a74d4aeebd.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 373, + 79, + 554, + 145 + ], + "blocks": [ + { + "bbox": [ + 414, + 69, + 512, + 78 + ], + "lines": [ + { + "bbox": [ + 414, + 69, + 512, + 78 + ], + "spans": [ + { + "bbox": [ + 414, + 69, + 512, + 78 + ], + "type": "text", + "content": "(c) SAVs vs Few-Shot SigLIP" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 373, + 79, + 554, + 145 + ], + "lines": [ + { + "bbox": [ + 373, + 79, + 554, + 145 + ], + "spans": [ + { + "bbox": [ + 373, + 79, + 554, + 145 + ], + "type": "table", + "html": "
NBES
SigLIP Few-shot Mean1.288.2
SigLIP Few-shot Cross-Modal6.886.4
SAVs35.186.7
", + "image_path": "c6aa54d27ee23fa49161c55874ffaa1453f85f1e37cb8626cc3d4cb1191c0de0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 59, + 199, + 279, + 258 + ], + "blocks": [ + { + "bbox": [ + 55, + 152, + 555, + 175 + ], + "lines": [ + { + "bbox": [ + 55, + 152, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 555, + 175 + ], + "type": "text", + "content": "Table 2. Additional SAVs Experiments. We (a) demonstrate the generalization of SAV heads to similar tasks, (b) show the effectiveness of SAV for tasks with interleaved image-text inputs, and (c) compare SAVs with few-shot formulations of CLIP and SigLIP." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 126, + 190, + 212, + 198 + ], + "lines": [ + { + "bbox": [ + 126, + 190, + 212, + 198 + ], + "spans": [ + { + "bbox": [ + 126, + 190, + 212, + 198 + ], + "type": "text", + "content": "(a) Classification Methods" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 199, + 279, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 199, + 279, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 199, + 279, + 258 + ], + "type": "table", + "html": "
MHBNBES
Class Centroid80.835.186.7
KNN53.011.078.1
Linear Probe82.532.983.1
", + "image_path": "ab6d622f49bdd0496b14f6f4f56a01132e5f41d868d757a7890ae7a0b04266da.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 200, + 533, + 257 + ], + "blocks": [ + { + "bbox": [ + 381, + 190, + 466, + 199 + ], + "lines": [ + { + "bbox": [ + 381, + 190, + 466, + 199 + ], + "spans": [ + { + "bbox": [ + 381, + 190, + 466, + 199 + ], + "type": "text", + "content": "(b) Sparse Configurations" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 200, + 533, + 257 + ], + "lines": [ + { + "bbox": [ + 315, + 200, + 533, + 257 + ], + "spans": [ + { + "bbox": [ + 315, + 200, + 533, + 257 + ], + "type": "table", + "html": "
MHBNBES
Sparse Heads80.835.186.7
Sparse Layers79.028.481.8
", + "image_path": "08629bc2dd45de0c1d409823c6a452668f868b703631a723f94d2d2872350fab.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 265, + 555, + 310 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 555, + 310 + ], + "type": "text", + "content": "Table 3. SAVs Ablations. We perform several ablations to identify the important aspects of our method that contribute to its effectiveness. In particular, we compare the effectiveness of (a) different classification methods and (b) head feature sparsification versus layer feature sparsification. Note: MHB represents MHaluBench, NB represents NaturalBench Group Score, and ES represents EuroSAT. For more ablations, please refer to Section 7.1 in the Supplementary." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 331, + 294, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 294, + 354 + ], + "type": "text", + "content": "demonstrates the adaptability of our method to any VL classification regardless of the input's modality." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 356, + 296, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 296, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 296, + 631 + ], + "type": "text", + "content": "Comparing SAVs to few-shot SigCLIP. As SAVs are extracted with few-shot examples, we compare our method to an analogous version of few-shot SigLIP. However, because SigLIP cannot be directly made few-shot, we adapt SigLIP as a few-shot class centroid classifier. One method used is the current SOTA few-shot classification method for encoder models Cross-Modal Adaptation [56]. We apply a second method where we aggregate CLIP/SigLIP embeddings into a mean embedding for each label. Then, just as in SAVs, we perform class centroid classification for each query using our set of mean SigLIP embeddings. We note that for image classification like EuroSAT, only image embeddings are needed, but for MHaluBench, multimodal embeddings are necessary. SAVs are inherently multimodal and so can be flexibly applied to both, but CLIP/SigLIP only have image-only or text-only embeddings. To overcome this, we use SigLIP image features for EuroSAT and concatenate image and text features for MHaluBench. Interestingly, while SiGLIP is comparable to SAVs on EuroSAT, our method vastly outperforms SigLIP in the few-shot setting for MHaluBench, suggesting generalizability of our method for a variety of multimodal tasks that CLIP-like struggle with. Our results are shown in Table 2c." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 644, + 128, + 657 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "Our research demonstrates the effectiveness of extracting Sparse Attention Vectors (SAVs) from the heads of an LMM and utilizing them directly for vision-language classification. Our method stands out by using only few-shot exam" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 331, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 534 + ], + "type": "text", + "content": "plies per label and only less than " + }, + { + "bbox": [ + 313, + 331, + 555, + 534 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 534 + ], + "type": "text", + "content": " of the heads to outperform zero-shot, few-shot, and fine-tuned baselines on a variety of image-text and image-only tasks. In addition, SAVs allows generative LMMs to close the gap with closed-source GPT-4o while also being an interpretable method that can generalize to similar tasks. Our ablations reveal the flexibility of using any classification method as a sparsification method for attention vectors and also shows that features are found as outputs of heads rather than layers. Overall, these results show that SAVs is a lightweight, performant, and generalizable method for extending generative LMMs' multimodal classification abilities. We are encouraged by the outcomes, and anticipate many directions for future work. In addition to methodological improvements, we look forward to the application of SAVs as features for multimodal retrieval, data compression, or more generally as a distilled representation for downstream models." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 548, + 389, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 389, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 389, + 560 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Sparse Attention Vectors are a significant step in generalizing the capabilities of generative LMMs to classification tasks. Nevertheless, it is valuable to consider certain limitations of our approach. SAVs are a method that requires access to the model's internal architecture and so may not be directly applicable to closed-source models like GPT-4 [73] and Gemini [89, 90]. Additionally, some tasks like image-text retrieval [28, 91] can benefit from more fine-grained confidence metrics attached to each label than proportion of voting heads per label. These challenges prompt future work in these directions as well as exciting questions about how to use SAVs as feature embeddings for other tasks." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2767" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 144, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 144, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 144, + 85 + ], + "type": "text", + "content": "Acknowledgements." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 87, + 297, + 137 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 87, + 297, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 87, + 297, + 137 + ], + "type": "text", + "content": "We would like to thank Abrar Anwar and Tyler Bonnen for helpful feedback and discussions. This project was supported in part by DoD, including PTG and/or LwLL programs, as well as BAIR's industrial alliance programs." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 146, + 115, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 146, + 115, + 158 + ], + "spans": [ + { + "bbox": [ + 56, + 146, + 115, + 158 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 166, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 166, + 294, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 166, + 294, + 198 + ], + "spans": [ + { + "bbox": [ + 66, + 166, + 294, + 198 + ], + "type": "text", + "content": "[1] S. R. Bowman A. Williams, N. Nangia. A broad-coverage challenge corpus for sentence understanding through inference. ArXiv, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 200, + 295, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 200, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 66, + 200, + 295, + 253 + ], + "type": "text", + "content": "[2] Stanislaw Antol, Aishwarya Agrawal, Jiasen Lu, Margaret Mitchell, Dhruv Batra, C Lawrence Zitnick, and Devi Parikh. Vqa: Visual question answering. In IEEE International Conference on Computer Vision, pages 2425-2433, 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 255, + 294, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 255, + 294, + 298 + ], + "spans": [ + { + "bbox": [ + 66, + 255, + 294, + 298 + ], + "type": "text", + "content": "[3] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A frontier large vision-language model with versatile abilities. ArXiv, abs/2308.12966, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 300, + 294, + 333 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 300, + 294, + 333 + ], + "spans": [ + { + "bbox": [ + 66, + 300, + 294, + 333 + ], + "type": "text", + "content": "[4] Pierre Baldi. Autoencoders, unsupervised learning, and deep architectures. In ICML 2011 Unsupervised and Transfer Learning Workshop, 2011. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 335, + 294, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 335, + 294, + 366 + ], + "spans": [ + { + "bbox": [ + 66, + 335, + 294, + 366 + ], + "type": "text", + "content": "[5] Rohan Bavishi, Erich Elsen, Curtis Hawthorne, Maxwell Nye, Augustus Odena, Arushi Somani, and Săgnak Taşürlar. Introducing our multimodal models, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 369, + 294, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 369, + 294, + 422 + ], + "spans": [ + { + "bbox": [ + 66, + 369, + 294, + 422 + ], + "type": "text", + "content": "[6] Matyas Bohacek and Michal Bravansky. When XGBoost outperforms GPT-4 on text classification: A case study. In Trustworthy Natural Language Processing (TrustNLP) 4th Workshop, pages 51–60, Mexico City, Mexico, 2024. Association for Computational Linguistics. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 424, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 424, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 66, + 424, + 295, + 544 + ], + "type": "text", + "content": "[7] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeff Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, and Dario Amodei. Language models are few-shot learners. *ArXiv*, abs/2005.14165, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 546, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 546, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 66, + 546, + 294, + 589 + ], + "type": "text", + "content": "[8] Martin Juan José E Bucher and Marco Martini. Finetuned 'small' llms (still) significantly outperform zero-shot generative ai models in text classification. ArXiv, abs/2406.08660, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 591, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 591, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 66, + 591, + 294, + 644 + ], + "type": "text", + "content": "[9] Xiang Chen, Chenxi Wang, Yida Xue, Ningyu Zhang, Xiaoyan Yang, Qiang Li, Yue Shen, Jinjie Gu, and Huajun Chen. Unified hallucination detection for multimodal large language models. arXiv preprint arXiv:2402.03190, 2024. 4, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 689 + ], + "type": "text", + "content": "[10] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248–255. IEEE, 2009. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 713 + ], + "type": "text", + "content": "[11] Xiang Deng, Vasilisa Bashlovkina, Feng Han, Simon Baumgartner, and Michael Bendersky. Llms to the moon?" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 105 + ], + "type": "text", + "content": "reddit market sentiment analysis with large language models. Companion Proceedings of the ACM Web Conference 2023, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "text", + "content": "[12] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In 2021 IEEE/CVF conference on computer vision and pattern recognition (CVPR), pages 11162-11173, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 152, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 152, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 321, + 152, + 553, + 194 + ], + "type": "text", + "content": "[13] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. ArXiv, abs/1810.04805, 2019. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 321, + 198, + 553, + 240 + ], + "type": "text", + "content": "[14] Sivan Doveh, Shaked Perek, M Jehanzeb Mirza, Wei Lin, Amit Alfassy, Assaf Arbelle, Shimon Ullman, and Leonid Karlinsky. Towards multimodal in-context learning for vision & language models. ArXiv, abs/2403.12736, 2024. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 243, + 553, + 274 + ], + "type": "text", + "content": "[15] Alaaeldin El-Nouby, Natalia Neverova, Ivan Laptev, and Herv'e J'egou. Training vision transformers for image retrieval. ArXiv, abs/2102.05644, 2021. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "text", + "content": "[16] Kawin Ethayarajh, Winnie Xu, Niklas Muennighoff, Dan Jurafsky, and Douwe Kiela. Kto: Model alignment as prospect theoretic optimization. ArXiv, abs/2402.01306, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "type": "text", + "content": "[17] Evelina Fedorenko, Michael K. Behr, and Nancy Kanwisher. Functional specificity for high-level linguistic processing in the human brain. Proceedings of the National Academy of Sciences, 108(39):16428-16433, 2011. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 366, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 366, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 366, + 553, + 420 + ], + "type": "text", + "content": "[18] Xingyu Fu, Yushi Hu, Bangzheng Li, Yu Feng, Haoyu Wang, Xudong Lin, Dan Roth, Noah A Smith, Wei-Chiu Ma, and Ranjay Krishna. Blink: Multimodal large language models can see but not perceive. arXiv preprint arXiv:2404.12390, 2024. 2, 4, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 453 + ], + "type": "text", + "content": "[19] Tianyu Gao, Xingcheng Yao, and Danqi Chen. Simcse: Simple contrastive learning of sentence embeddings. ArXiv, abs/2104.08821, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 456, + 553, + 510 + ], + "type": "text", + "content": "[20] Danna Gurari, Qing Li, Abigale Stangl, Anhong Guo, Chi Lin, Kristen Grauman, Jiebo Luo, and Jeffrey P. Bigham. Vizwiz grand challenge: Answering visual questions from blind people. 2018 IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3608-3617, 2018. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 512, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 565 + ], + "type": "text", + "content": "[21] Donghoon Han, Eunhwan Park, Gisang Lee, Adam Lee, and Nojun Kwak. Merlin: Multimodal embedding refinement via llm-based iterative navigation for text-video retrieval-erank pipeline. arXiv preprint arXiv:2407.12508, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 567, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 567, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 567, + 553, + 621 + ], + "type": "text", + "content": "[22] Patrick Helber, Benjamin Bischke, Andreas Dengel, and Damian Borth. Eurosat: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 12(7):2217-2226, 2019. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 624, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 655 + ], + "type": "text", + "content": "[23] Roee Hendel, Mor Geva, and Amir Globerson. In-context learning creates task vectors. ArXiv, abs/2310.15916, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 321, + 658, + 553, + 689 + ], + "type": "text", + "content": "[24] Roee Hendel, Mor Geva, and Amir Globerson. Incontext learning creates task vectors. arXiv preprint arXiv:2310.15916, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 692, + 553, + 713 + ], + "type": "text", + "content": "[25] Roei Herzig, Alon Mendelson, Leonid Karlinsky, Assaf Arabelle, Rogerio Feris, Trevor Darrell, and Amir Globerson." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2768" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "text", + "content": "Incorporating structured representations into pretrained vision " + }, + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "inline_equation", + "content": "\\&" + }, + { + "bbox": [ + 80, + 73, + 294, + 116 + ], + "type": "text", + "content": " language models using scene graphs. In 2023 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 138 + ], + "type": "text", + "content": "[26] L. Hinton G, van der Maaten. Visualizing data using t-sne. Journal of machine learning research, 2008. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "text", + "content": "[27] Alberto Hojel, Yutong Bai, Trevor Darrell, Amir Globerson, and Amir Bar. Finding visual task vectors. In European Conference on Computer Vision (ECCV), pages 257-273. Springer, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 295, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 227 + ], + "type": "text", + "content": "[28] Cheng-Yu Hsieh, Jieyu Zhang, Zixian Ma, Aniruddha Kembhavi, and Ranjay Krishna. Sugarcrepe: Fixing hackable benchmarks for vision-language compositionality. ArXiv, abs/2306.14610, 2023. 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 271 + ], + "type": "text", + "content": "[29] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models. arXiv preprint arXiv:2106.09685, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 294, + 316 + ], + "type": "text", + "content": "[30] Brandon Huang, Chancharik Mitra, Assaf Arbelle, Leonid Karlinsky, Trevor Darrell, and Roei Herzig. Multimodal task vectors enable many-shot multimodal in-context learning. arXiv preprint arXiv:2406.15334, 2024. 2, 6, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 294, + 369 + ], + "type": "text", + "content": "[31] Weiquan Huang, Aoqi Wu, Yifan Yang, Xufang Luo, Yuqing Yang, Liang Hu, Qi Dai, Xiyang Dai, Dongdong Chen, Chong Luo, et al. Llm2clip: Powerful language model unlock richer visual representation. arXiv preprint arXiv:2411.04997, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 372, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 372, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 372, + 294, + 425 + ], + "type": "text", + "content": "[32] Drew A. Hudson and Christopher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6693-6702, 2019. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 426, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 469 + ], + "type": "text", + "content": "[33] Achiya Jerbi, Roei Herzig, Jonathan Berant, Gal Chechik, and Amir Globerson. Learning object detection from captions via textual scene attributes. ArXiv, abs/2009.14558, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 525 + ], + "type": "text", + "content": "[34] Ting Jiang, Shaohan Huang, Zi qiang Zhang, Deqing Wang, Fuzhen Zhuang, Furu Wei, Haizhen Huang, Liangjie Zhang, and Qi Zhang. Promptbert: Improving bert sentence embeddings with prompts. In Conference on Empirical Methods in Natural Language Processing, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 558 + ], + "type": "text", + "content": "[35] Ting Jiang, Shaohan Huang, Zhongzhi Luan, Deqing Wang, and Fuzhen Zhuang. Scaling sentence embeddings with large language models. ArXiv, abs/2307.16645, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "text", + "content": "[36] Ting Jiang, Minghui Song, Zihan Zhang, Haizhen Huang, Weiwei Deng, Feng Sun, Qi Zhang, Deqing Wang, and Fuzhen Zhuang. E5-v: Universal embeddings with multimodal large language models. ArXiv, abs/2407.12580, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 635 + ], + "type": "text", + "content": "[37] Nancy Kanwisher. Domain specificity in face perception. Nature neuroscience, 3(8):759-763, 2000. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "text", + "content": "[38] Sahar Kazemzadeh, Vicente Ordonez, Marc andre Matten, and Tamara L. Berg. Referitgame: Referring to objects in photographs of natural scenes. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2014. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[39] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. CoRR, abs/1312.6114, 2013. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 127 + ], + "type": "text", + "content": "[40] Doyup Lee, Chiheon Kim, Saehoon Kim, Minsu Cho, and Wook-Shin Han. Autoregressive image generation using residual quantization. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11513-11522, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 320, + 129, + 553, + 182 + ], + "type": "text", + "content": "[41] Yibin Lei, Di Wu, Tianyi Zhou, Tao Shen, Yu Cao, Chongyang Tao, and Andrew Yates. Meta-task prompting elicits embeddings from large language models. In Annual Meeting of the Association for Computational Linguistics, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 320, + 183, + 553, + 227 + ], + "type": "text", + "content": "[42] Brian Lester, Rami Al-Rfou, and Noah Constant. The power of scale for parameter-efficient prompt tuning. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 320, + 228, + 553, + 271 + ], + "type": "text", + "content": "[43] Bohao Li, Rui Wang, Guangzhi Wang, Yuying Ge, Yixiao Ge, and Ying Shan. Seed-bench: Benchmarking multimodal llms with generative comprehension. ArXiv, abs/2307.16125, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 272, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 272, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 320, + 272, + 553, + 336 + ], + "type": "text", + "content": "[44] Baiqi Li, Zhiqiu Lin, Wenxuan Peng, Jean de Dieu Nyandwi, Daniel Jiang, Zixian Ma, Simran Khanuja, Ranjay Krishna, Graham Neubig, and Deva Ramanan. Natural-bench: Evaluating vision-language models on natural adversarial samples. arXiv preprint arXiv:2410.14669, 2024. 2, 4, 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 320, + 338, + 553, + 380 + ], + "type": "text", + "content": "[45] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. ArXiv, abs/2408.03326, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 383, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 383, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 320, + 383, + 553, + 426 + ], + "type": "text", + "content": "[46] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. arXiv preprint arXiv:2201.12086, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 320, + 426, + 553, + 469 + ], + "type": "text", + "content": "[47] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 471, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 471, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 320, + 471, + 553, + 503 + ], + "type": "text", + "content": "[48] Shiyu Li, Yang Tang, Shizhe Chen, and Xi Chen. Conan-embedding: General text embedding with more and better negative samples. ArXiv, abs/2408.15710, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 504, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 504, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 320, + 504, + 553, + 568 + ], + "type": "text", + "content": "[49] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), abs/2101.00190, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 320, + 570, + 553, + 624 + ], + "type": "text", + "content": "[50] Yanghao Li, Haoqi Fan, Ronghang Hu, Christoph Feichtenhofer, and Kaiming He. Scaling language-image pretraining via masking. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 23390-23400, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 320, + 625, + 553, + 657 + ], + "type": "text", + "content": "[51] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 658, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 658, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 320, + 658, + 553, + 689 + ], + "type": "text", + "content": "[52] Ziyue Li and Tianyi Zhou. Your mixture-of-experts llm is secretly an embedding model for free. arXiv preprint arXiv:2410.10814, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 691, + 553, + 713 + ], + "type": "text", + "content": "[53] Jiacheng Lin, Kun Qian, Haoyu Han, Nurendra Choudhary, Tianxin Wei, Zhongruo Wang, Sahika Genc, Edward W" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2769" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "text", + "content": "Huang, Sheng Wang, Karthik Subbian, et al. Unleashing the power of lms as multi-modal encoders for text and graph-structured data. arXiv preprint arXiv:2410.11235, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 106, + 295, + 150 + ], + "type": "text", + "content": "[54] Tsung-Yi Lin, M. Maire, Serge J. Belongie, James Hays, P. Perona, D. Ramanan, Piotr Dollar, and C. L. Zitnick. Microsoft coco: Common objects in context. In European Conference on Computer Vision (ECCV), 2014. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 150, + 294, + 194 + ], + "type": "text", + "content": "[55] Zhiqiu Lin, Xinyue Chen, Deepak Pathak, Pengchuan Zhang, and Deva Ramanan. Revisiting the role of language priors in vision-language models. arXiv preprint arXiv:2306.01879, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 249 + ], + "type": "text", + "content": "[56] Zhiqiu Lin, Samuel Yu, Zhiyi Kuang, Deepak Pathak, and Deva Ramanan. Multimodality helps unimodality: Cross-modal few-shot learning with multimodal models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19325-19337, 2023. 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 295, + 304 + ], + "type": "text", + "content": "[57] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "type": "text", + "content": "[58] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with imageto-text generation. In European Conference on Computer Vision, pages 366-384. Springer, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 294, + 382 + ], + "type": "text", + "content": "[59] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 415 + ], + "type": "text", + "content": "[60] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Advances in Neural Information Processing Systems (NeurIPS), 2023. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 294, + 459 + ], + "type": "text", + "content": "[61] Tian Yu Liu, Matthew Trager, Alessandro Achille, Pramuditha Perera, Luca Zancato, and Stefan O Soatto. Meaning representations from trajectories in autoregressive models. ArXiv, abs/2310.18348, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 459, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 459, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 459, + 294, + 503 + ], + "type": "text", + "content": "[62] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 503, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 503, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 62, + 503, + 294, + 548 + ], + "type": "text", + "content": "[63] Romain Lopez, Pierre Boyeau, Nir Yosef, Michael Jordan, and Jeffrey Regier. Decision-making with auto-encoding variational bayes. Advances in Neural Information Processing Systems, 33:5081–5092, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 62, + 548, + 294, + 591 + ], + "type": "text", + "content": "[64] Shiyin Lu, Yang Li, Qing-Guo Chen, Zhao Xu, Weihua Luo, Kaifu Zhang, and Han-Jia Ye. Ovis: Structural embedding alignment for multimodal large language model. ArXiv, abs/2405.20797, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 592, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 294, + 647 + ], + "type": "text", + "content": "[65] Junhua Mao, Jonathan Huang, Alexander Toshev, OanaMaria Camburu, Alan Loddon Yuille, and Kevin P. Murphy. Generation and comprehension of unambiguous object descriptions. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 11-20, 2015. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 647, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 691 + ], + "type": "text", + "content": "[66] Tomas Mikolov, Ilya Sutskever, Kai Chen, Greg S Corrado, and Jeff Dean. Distributed representations of words and phrases and their compositionality. Advances in neural information processing systems, 26, 2013. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 714 + ], + "type": "text", + "content": "[67] Chancharik Mitra, Brandon Huang, Trevor Darrell, and Roei Herzig. Compositional chain of thought prompting" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 95 + ], + "type": "text", + "content": "for large multimodal models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 321, + 96, + 553, + 140 + ], + "type": "text", + "content": "[68] Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. Mteb: Massive text embedding benchmark. In Conference of the European Chapter of the Association for Computational Linguistics, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "spans": [ + { + "bbox": [ + 320, + 141, + 553, + 185 + ], + "type": "text", + "content": "[69] Jianmo Ni, Gustavo Hernández Abrego, Noah Constant, Ji Ma, Keith B. Hall, Daniel Matthew Cer, and Yinfei Yang. Sentence-t5: Scalable sentence encoders from pre-trained text-to-text models. ArXiv, abs/2108.08877, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 186, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 553, + 241 + ], + "type": "text", + "content": "[70] Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hernandez Abrego, Ji Ma, Vincent Zhao, Yi Luan, Keith B. Hall, Ming-Wei Chang, and Yinfei Yang. Large dual encoders are generalizable retrievers. ArXiv, abs/2112.07899, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 243, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 243, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 320, + 243, + 553, + 286 + ], + "type": "text", + "content": "[71] Maria-Elena Nilsback and Andrew Zisserman. Automated flower classification over a large number of classes. 2008 Sixth Indian Conference on Computer Vision, Graphics & Image Processing, pages 722-729, 2008. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 287, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 287, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 320, + 287, + 553, + 342 + ], + "type": "text", + "content": "[72] Catherine Olsson, Nelson Elhage, Neel Nanda, Nicholas Joseph, Nova DasSarma, Tom Henighan, Ben Mann, Amanda Askell, Yuntao Bai, Anna Chen, et al. Incontext learning and induction heads. arXiv preprint arXiv:2209.11895, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 343, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 343, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 320, + 343, + 553, + 365 + ], + "type": "text", + "content": "[73] OpenAI. Gpt-4 technical report. ArXiv, abs/2303.08774, 2023. 1, 5, 8, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 366, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 366, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 553, + 454 + ], + "type": "text", + "content": "[74] Long Ouyang, Jeff Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke E. Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul Francis Christiano, Jan Leike, and Ryan J. Lowe. Training language models to follow instructions with human feedback. ArXiv, abs/2203.02155, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 456, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 553, + 499 + ], + "type": "text", + "content": "[75] Omkar M Parkhi, Andrea Vedaldi, Andrew Zisserman, and CV Jawahar. Cats and dogs. In IEEE/CVF conference on computer vision and pattern recognition, pages 3498-3505. IEEE, 2012. 5, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 500, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 500, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 320, + 500, + 553, + 567 + ], + "type": "text", + "content": "[76] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 4, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 568, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 568, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 320, + 568, + 553, + 612 + ], + "type": "text", + "content": "[77] Jeffrey Pennington, Richard Socher, and Christopher D. Manning. Glove: Global vectors for word representation. In 2014 conference on empirical methods in natural language processing (EMNLP), pages 1532-1543, 2014. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 613, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 613, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 320, + 613, + 553, + 679 + ], + "type": "text", + "content": "[78] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), pages 8748-8763. PMLR, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 680, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 680, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 320, + 680, + 553, + 714 + ], + "type": "text", + "content": "[79] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2770" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 81, + 73, + 294, + 95 + ], + "type": "text", + "content": "reward model. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 95, + 295, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 95, + 295, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 95, + 295, + 138 + ], + "type": "text", + "content": "[80] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. ArXiv, abs/2204.06125, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 183 + ], + "type": "text", + "content": "[81] Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence embeddings using siamese bert-networks. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2019. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 295, + 228 + ], + "type": "text", + "content": "[82] David E Rumelhart, James L McClelland, PDP Research Group, et al. Parallel distributed processing, volume 1: Explorations in the microstructure of cognition: Foundations. The MIT press, 1986. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 295, + 271 + ], + "type": "text", + "content": "[83] Florian Schroff, Dmitry Kalenichenko, and James Philbin. Facenet: A unified embedding for face recognition and clustering. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 815-823, 2015. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "text", + "content": "[84] Shai Shalev-Shwartz and Shai Ben-David. Understanding machine learning: From theory to algorithms. Cambridge university press, 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 305, + 295, + 360 + ], + "type": "text", + "content": "[85] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8309-8318, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 361, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 295, + 415 + ], + "type": "text", + "content": "[86] Aaditya K. Singh, Stephanie C.Y. Chan, Ted Moskovitz, Erin Grant, Andrew M. Saxe, and Felix Hill. The transient nature of emergent in-context learning in transformers. In Advances in neural information processing systems (NeurIPS), 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 415, + 295, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 415, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 62, + 415, + 295, + 481 + ], + "type": "text", + "content": "[87] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), 2013. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 482, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 295, + 525 + ], + "type": "text", + "content": "[88] Xiaofei Sun, Xiaoya Li, Jiwei Li, Fei Wu, Shangwei Guo, Tianwei Zhang, and Guoyin Wang. Text classification via large language models. In Conference on Empirical Methods in Natural Language Processing (EMNLP), 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 526, + 295, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 295, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 295, + 558 + ], + "type": "text", + "content": "[89] Gemini Team. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. ArXiv, abs/2403.05530, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 559, + 295, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 295, + 613 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 295, + 613 + ], + "type": "text", + "content": "[90] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 614, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 295, + 679 + ], + "type": "text", + "content": "[91] Tristan Thrush, Ryan Jiang, Max Bartolo, Amanpreet Singh, Adina Williams, Douwe Kiela, and Candace Ross. Winoground: Probing vision and language models for visio-linguistic compositionality. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5238–5248, 2022. 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 680, + 295, + 713 + ], + "type": "text", + "content": "[92] Eric Todd, Millicent Li, Arnab Sen Sharma, Aaron Mueller, Byron C. Wallace, and David Bau. Function vectors in large language models. ArXiv, abs/2310.15213, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 554, + 116 + ], + "type": "text", + "content": "[93] Eric Todd, Millicent L Li, Arnab Sen Sharma, Aaron Mueller, Byron C Wallace, and David Bau. Function vectors in large language models. arXiv preprint arXiv:2310.15213, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 118, + 554, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 554, + 150 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 554, + 150 + ], + "type": "text", + "content": "[94] Matthew A. Turk and Alex Pentland. Eigenfaces for recognition. Journal of Cognitive Neuroscience, 3:71-86, 1991. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 554, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 554, + 185 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 554, + 185 + ], + "type": "text", + "content": "[95] Aäron van den Oord, Oriol Vinyals, and Koray Kavukcuoglu. Neural discrete representation learning. ArXiv, abs/1711.00937, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 186, + 554, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 186, + 554, + 218 + ], + "spans": [ + { + "bbox": [ + 320, + 186, + 554, + 218 + ], + "type": "text", + "content": "[96] Catherine Wah, Steve Branson, Peter Welinder, Pietro Perona, and Serge J. Belongie. The caltech-ucsd birds-200-2011 dataset. 2011. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 320, + 220, + 554, + 262 + ], + "type": "text", + "content": "[97] Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. Improving text embeddings with large language models. ArXiv, abs/2401.00368, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 264, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 264, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 320, + 264, + 554, + 341 + ], + "type": "text", + "content": "[98] Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Ke-Yang Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Yang Fan, Kai Dang, Mengfei Du, Xuancheng Ren, Rui Men, Dayiheng Liu, Chang Zhou, Jingren Zhou, and Junyang Lin. Qwen2-vl: Enhancing vision-language model's perception of the world at any resolution. ArXiv, abs/2409.12191, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 342, + 554, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 342, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 320, + 342, + 554, + 397 + ], + "type": "text", + "content": "[99] Pengyu Wang, Dong Zhang, Linyang Li, Chenkun Tan, Xinghao Wang, Ke Ren, Botian Jiang, and Xipeng Qiu. Inferaligner: Inference-time alignment for harmlessness through cross-model guidance. arXiv preprint arXiv:2401.11206, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 443 + ], + "type": "text", + "content": "[100] Xinlong Wang, Xiaosong Zhang, Zhengxiong Luo, Quan Sun, Yufeng Cui, Jinsheng Wang, Fan Zhang, Yueze Wang, Zhen Li, Qiying Yu, et al. Emu3: Next-token prediction is all you need. arXiv preprint arXiv:2409.18869, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 444, + 554, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 554, + 520 + ], + "type": "text", + "content": "[101] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, and Limin Wang. Internvideo2: Scaling video foundation models for multimodal video understanding. ArXiv, abs/2403.15377, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 521, + 554, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 554, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 554, + 554 + ], + "type": "text", + "content": "[102] Zhiqiang Wang, Yiran Pang, and Yanbin Lin. Large language models are zero-shot text classifiers. ArXiv, abs/2312.01044, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 554, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 599 + ], + "type": "text", + "content": "[103] Jason Wei, Maarten Bosma, Vincent Zhao, Kelvin Guu, Adams Wei Yu, Brian Lester, Nan Du, Andrew M. Dai, and Quoc V. Le. Finetuned language models are zero-shot learners. ArXiv, abs/2109.01652, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 600, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 554, + 645 + ], + "type": "text", + "content": "[104] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Ed Huai hsin Chi, F. Xia, Quoc Le, and Denny Zhou. Chain of thought prompting elicits reasoning in large language models. ArXiv, abs/2201.11903, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 679 + ], + "type": "text", + "content": "[105] Dean Wyatte, Fatemeh Tahmasbi, Ming Li, and Thomas Markovich. Scaling laws for discriminative classification in large language models. ArXiv, abs/2405.15765, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[106] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yi Zhou, Junyan Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, Chenliang Li, Yuanhong Xu, Hehong Chen, Junfeng" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2771" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 487 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 106 + ], + "type": "text", + "content": "Tian, Qiang Qi, Ji Zhang, and Feiyan Huang. mplug-owl: Modularization empowers large language models with multimodality. ArXiv, abs/2304.14178, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 162 + ], + "type": "text", + "content": "[107] Peter Young, Alice Lai, Micah Hodosh, and J. Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2:67-78, 2014. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 295, + 206 + ], + "type": "text", + "content": "[108] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 11941-11952, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[109] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In IEEE/CVF International Conference on Computer Vision, pages 11975-11986, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[110] Yuhao Zhang, Hang Jiang, Yasuhide Miura, Christopher D. Manning, and C. Langlotz. Contrastive learning of medical visual representations from paired images and text. In Machine Learning in Health Care, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[111] Yuhui Zhang, Alyssa Unell, Xiaohan Wang, Dhruba Ghosh, Yuchang Su, Ludwig Schmidt, and Serena Yeung-Levy. Why are visually-grounded language models bad at image classification? ArXiv, abs/2405.18415, 2024. 1, 2, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 396 + ], + "type": "text", + "content": "[112] Haozhe Zhao, Zefan Cai, Shuzheng Si, Xiaojian Ma, Kaikai An, Liang Chen, Zixuan Liu, Sheng Wang, Wenjuan Han, and Baobao Chang. Mmicl: Empowering vision-language model with multi-modal in-context learning. ArXiv, abs/2309.07915, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[113] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 487 + ], + "type": "text", + "content": "[114] Yongshuo Zong, Ondrej Bohdal, Tingyang Yu, Yongxin Yang, and Timothy Hospedales. Safety fine-tuning at (almost) no cost: A baseline for vision large language models. arXiv preprint arXiv:2402.02207, 2024. 2, 4" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2772" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_content_list.json b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..19a262a77903f323203edf6b45099a445011b0de --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_content_list.json @@ -0,0 +1,2154 @@ +[ + { + "type": "text", + "text": "Enhancing Image Restoration Transformer via Adaptive Translation Equivariance", + "text_level": 1, + "bbox": [ + 261, + 130, + 736, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "JiaKui Hu $^{1,2,3}$ , Zhengjian Yao $^{1,2,3}$ , Lujia Jin $^{4}$ , Hangzhou He $^{1,2,3}$ , Yanye Lu $^{1,2,3*}$ $^{1}$ Institute of Medical Technology, Peking University Health Science Center, Peking University \n $^{2}$ Biomedical Engineering Department, College of Future Technology, Peking University \n $^{3}$ National Biomedical Imaging Center, Peking University \n $^{4}$ China Mobile Research Institute", + "bbox": [ + 127, + 202, + 869, + 290 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jkhu29@stu.pku.edu.cn, yanye.lu@pu.edu.cn", + "bbox": [ + 308, + 295, + 683, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 344, + 326, + 359 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivalence Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization.", + "bbox": [ + 89, + 375, + 483, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 748, + 220, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration aims to reconstruct high-quality (HQ) images from their low-quality (LQ) counterparts, which are often affected by degradations resulting from the imaging system or environment. As a constituent of the imaging system, image restoration methods should satisfy the following fidelity property: the restoration results for the target region should remain equivariant under geometric transfor-", + "bbox": [ + 89, + 773, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mations of the input. This fidelity property is termed translation equivariance, which is one of the inherent inductive biases in image restoration.", + "bbox": [ + 511, + 345, + 906, + 391 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdfa586a084a50a3c2d482ef86180369c3b78e0a3c76f38347db32ba256bbb6e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 400, + 583, + 463 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee6cd240c1cf4f8aaff89f54d7d19f1c6cb10270ebb3df8f46fe4852cc26134a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 401, + 658, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/751f7cbbb4f62e1fdce47b7731e7cad2c80493e55ae46a1fa2e299befad9dadd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 401, + 733, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf0a4a3481b6cba5998820aa47ad0a7839d26b8a71ffd1fa410e6a48c6b105cf.jpg", + "image_caption": [ + "Figure 1. $\\mathcal{T}(\\cdot)$ means the translation function. \"SA\" and \"WA\" are commonly used for self-attention and window attention, respectively, but disrupt translation equivariance (TE) due to position encoding and feature shifting. \"TEA\" is our proposed translation equivariance adaptive attention, which satisfies TE. TE promotes faster convergence and better generalization." + ], + "image_footnote": [], + "bbox": [ + 517, + 463, + 584, + 525 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/da6343229e66edb0dd5ff2a569869cbaad7a89debfbf41c46a8a2df35698dbe0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 589, + 463, + 656, + 526 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8d13e64892278fc66fdd464a8c3aa86cd7b0ae570742405a579726918a8348b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 666, + 463, + 733, + 527 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/408acbc476c84a68cb551091348986afecde47974edc227cdc3008eded1723a6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 397, + 895, + 462 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c391195427018b5c444fad6e84337353ae4342c12105840af78cd4294df64b2d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 740, + 462, + 895, + 527 + ], + "page_idx": 0 + }, + { + "type": "table", + "img_path": "images/9df326fe6c5a3d9446564755d8e3ef1fb80dd564e033591237a240316873fbdf.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttentionsTEPerformanceNTK [25]SRGA [33]Complexity
PSNR↑/ SSIM↑Condition ↓Value ↓
SAX28.21 / 0.843511140.444.075O(N2)
WAX27.45 / 0.82541746.493.655O(N)
TEA (Ours)✓28.67 / 0.8489236.783.275O(N)
", + "bbox": [ + 517, + 619, + 913, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Table 1. In $4\\times$ image super-resolution, we study performance, NTK [8] convergence speed, SRGA [33] generalization, and computation complexities. We replace WA in SwinIR with SA and TEA for testing. TEA shows fast convergence, better generalization, and linear complexity. The results are tested on Urban100.", + "bbox": [ + 511, + 691, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In restoration networks [32, 62, 68, 71] based on convolution neural networks (CNNs), the paradigm of sliding information extraction inherent to convolutions [18] imparted a natural translation equivariance (TE) property in restoration networks. However, as shown in Figure 1 and Table 1,", + "bbox": [ + 511, + 766, + 906, + 842 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Translation equivariance and translation invariance are concepts that are frequently misconstrued. Our paper is dedicated to the analysis of equivariance. We have provided a clear delimitation and discussion concerning these concepts in the supplementary material.", + "bbox": [ + 511, + 851, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "footer", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16047", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the advent of Self Attention (SA) [51] and Window Attention (WA) [34] within Transformers has broken this fidelity property, leading to slow convergence and poor generalization in restoration transformers.", + "bbox": [ + 89, + 90, + 480, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose two fundamental strategies to bring TE back into the restoration transformers: (1) slide indexing and (2) component stacking. The slide indexing strategy employs a fixed container, e.g., a fixed rectangular window, to globally extract the input in a sliding fashion. This method requires the operators to calculate instantaneously, thereby maintaining the consistency of positional coordinates between the input and output of the model, which facilitates the integration of TE. The component stacking strategy involves stacking modules that satisfy the TE property in parallel or in series, thus enhancing architectures while ensuring TE in complex networks.", + "bbox": [ + 88, + 152, + 482, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, these foundational strategies present a persistent challenge in the design of restoration transformers. Selecting SA to achieve superior performance involves accepting its substantial computational complexity [63]; alternatively, opting for sliding window attention provides linear complexity but results in a fixed receptive field [48] and reduced performance. To break this dilemma, we replace the global key-value pair indexing mechanism in vanilla SA with an adaptive slide indexing mechanism to design the Translation Equivariance Adaptive (TEA) attention. To retain the extraction of global information, TEA incorporates a downsampling self-attention branch without position encoding, which reduces the key-value pair's resolution by aggregating them, thereby providing TEA with coarse, yet effective global information.", + "bbox": [ + 91, + 333, + 482, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions are threefold. First, we investigate TE in image restoration, analyzing its impact on convergence and generalization. The results indicate that the integration of TE is essential in the design of restoration networks. Second, we propose TEA, which uses adaptive slide indexing methods to incorporate the TE property while preserving linear complexity and ensuring a global receptive field, thereby enhancing restoration transformers. Third, we introduce TEA to the vanilla Vision Transformer to design TEAFormer. Experiments show that TEA achieves State-of-The-Art (SoTA) performance in various image restoration tasks, while having faster convergence and more robust generalization.", + "bbox": [ + 89, + 560, + 482, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 768, + 232, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Vision Transformers with inductive bias. Following the pioneering introduction of the Transformer [51] into vision tasks by ViT [16], SA has emerged as a key component in vision networks. However, ViT is challenging to train, exhibits slow convergence, and requires substantial data support [49]. Previous works, such as ViTAE [59, 70], attributed this to a lack of inductive bias and advocated in", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tegrating inductive bias with ViT to develop more efficient ones. AACNet [4] emphasized the importance of translation equivariance for visual tasks. The Swin [34] and CSwin [15] Transformers incorporated local processing into the vision Transformer, thus proposing window attention. CVT [56] and FasterViT [21] sequentially stack convolutions and attention layers, resulting in a hybrid architecture. PVT [52, 53] used average pooling for downsampling to achieve weak translation equivariance. Both DAT [57] and BiFormer [74] dynamically determined the windows accessible to each token. Recently, some studies have introduced translation equivariance into ViT by modifying position encoding [4, 58] and attention calculation logic [44, 46].", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent image restoration Transformers. Similar to the development of vision transformers, the design of image restoration models also focuses on how to better design local global information processors [9, 31, 61, 63, 73]. Since the resolution of the input image is generally high in the image restoration task, directly using SA will result in a huge computation cost. To address this issue, SwinIR [31] introduced window attention [34]. HAT [9] enlarged the receptive field of the restoration transformers with the squeeze-excitation module [22] and the overlapping mechanism. Restormer [63] and DAT [10] performed channel-wise attention, thereby facilitating more efficient feature re-weighting. GRL [61] extracted the feature on local, regional, and global scales by integrating the attention of the anchor window. IPG [48] exploited the flexibility of the network by proposing the degree-variant graph solution, breaking the rigidity of the previous methods.", + "bbox": [ + 511, + 287, + 903, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Translation Equivariance", + "text_level": 1, + "bbox": [ + 511, + 558, + 751, + 574 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we define the translation equivariance and propose two fundamental strategies for integrating translation equivariance into restoration networks.", + "bbox": [ + 511, + 583, + 903, + 627 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Definition 3.1. Translation Equivalence. We call function $\\Phi(\\cdot)$ is translation equivariant to translation operation $\\mathcal{T}(\\cdot)$ , if $\\Phi(\\mathcal{T}(x)) = \\mathcal{T}(\\Phi(x))$ , where $x$ is the input signal.", + "bbox": [ + 511, + 628, + 903, + 674 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In image restoration tasks, $x$ is an image or a feature in the latent space typically, $\\mathcal{T}(\\cdot)$ refers to a shift in pixels [20]. Based on this definition, two theorems are derived:", + "bbox": [ + 511, + 674, + 903, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 3.2. If $\\Phi(x)_i$ is transformed from $x_j = [i - b, i + b]$ , where $b$ is the sliding boundary, the function $\\Phi(\\cdot)$ fulfills Definition 3.1.", + "bbox": [ + 511, + 718, + 903, + 762 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Theorem 3.3. If the functions $\\Phi_1(\\cdot)$ and $\\Phi_2(\\cdot)$ fulfill Definition 3.1, the functions $\\Phi_2(\\Phi_1(\\cdot))$ and $\\Phi_1(\\cdot) + \\Phi_2(\\cdot)$ also fulfill Definition 3.1.", + "bbox": [ + 511, + 763, + 903, + 808 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reformulating the existing operators in line with Theorem 3.2 gives them the translation equivariance property. By stacking these operators, either in parallel or in serial, as specified in Theorem 3.3, more complex architectures are also capable of translation equivariance. Consequently, the model can be further enhanced via translation equivariance.", + "bbox": [ + 511, + 810, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16048", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d39d1908a1a44d458737227015e082d0cd52a96a1591ab640e3aa5928d0a8ea1.jpg", + "image_caption": [ + "Figure 2. The overall architecture of our proposed TEAFormer." + ], + "image_footnote": [], + "bbox": [ + 91, + 90, + 483, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 224, + 181, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we conform to Theorems 3.1 and 3.2 to devise a translation equivariance restoration transformer, termed Translation Equivariance trAnsFormer (TEAFormer). We first present the overall architecture of TEAFormer and then introduce its components in detail.", + "bbox": [ + 89, + 250, + 483, + 325 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Overall Architecture", + "text_level": 1, + "bbox": [ + 89, + 337, + 287, + 351 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall architecture of the proposed TEAFormer is illustrated in Figure 2 with image super-resolution (SR) presented as a demonstration. Given an input image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H \\times W$ denotes the resolution, TEAFormer first applies a convolution layer to obtain a shallow feature $F_{0} \\in \\mathbb{R}^{H \\times W \\times D}$ , where $D$ is the embedding dimension of the network. Next, following classical restoration networks [31], TEAFormer adopts a residual-in-residual structure to construct a deep feature extraction module comprised of $N_{g}$ Translation Equivariance Groups (TEGs). Each TEG consists of $N_{b}$ Translation Equivariance Blocks (TEBs) and a convolution layer. Each TEB, in turn, is composed of one Translation Equivariance Attention (TEA) and one feed-forward network. Subsequently, a convolutional layer is employed to extract the deep feature $F_{1} \\in \\mathbb{R}^{H \\times W \\times D}$ from the feature extracted by TEGs. After extraction through the deep feature extraction module, we use the restore module to generate the high-quality image $\\hat{I}$ from the feature $F_{r} = F_{0} + F_{1}$ . As for other image restoration tasks that do not involve changes in resolution, we build our TEAFormer following Restormer [63].", + "bbox": [ + 89, + 359, + 483, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Adaptive Translation Equivariance", + "text_level": 1, + "bbox": [ + 89, + 686, + 397, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Adaptive Translation Equivalence is integrated through the Adaptive Sliding Key-Value Self Attention (ASkvSA) and Downsampled Self Attention (DSA) as shown in Figure 3.", + "bbox": [ + 89, + 710, + 483, + 755 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Key-value indexing in self-attention. The nature of Self Attention (SA) without position bias lies in the indexing of key-value pairs, followed by feature re-weighting. Given a flattened input $X \\in \\mathbb{R}^{N \\times D}$ , where $N$ represents the number of tokens and $D$ represents the embedding dimension, self-attention can be formulated as following:", + "bbox": [ + 89, + 756, + 483, + 845 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} (X) = \\operatorname {S o f t m a x} \\left(\\frac {X W _ {q} \\left(X W _ {k}\\right) ^ {\\top}}{\\sqrt {D}}\\right) X W _ {v}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 854, + 482, + 887 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W_{q}, W_{k}$ , and $W_{v}$ denote the weights of the linear", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/e6c10ae9c6e10be075614c0c5d9a19113754516380df00a1cf2d3bcc1ac65d6a.jpg", + "image_caption": [ + "Figure 3. The inner structure of TEA with two components. (a) Adaptive Sliding key-value Self Attention (ASkvSA) and (b) Downsampled Self Attention (DSA)." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "projections utilized to compute the query $(Q)$ , key $(K)$ , and value $(V)$ from input $X$ .", + "bbox": [ + 511, + 364, + 905, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In Eq. 1, the softmax similarity function is utilized for the generation of attention maps between a query and a key, while the dot product is employed for re-weighting between the attention map and a value. Given that subscripting a matrix with $i$ returns the $i$ -th row as a vector, an alternative equation of self-attention can be written as follows.", + "bbox": [ + 511, + 397, + 905, + 488 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 502, + 903, + 546 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(Q_i, K_j) = \\exp \\left(\\frac{Q_i^\\top K_j}{\\sqrt{D}}\\right)$ .", + "bbox": [ + 511, + 554, + 751, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. 2 illustrates that self-attention indexes key-value pairs $(K_{j}$ and $V_{j}, j = 1, \\dots, N)$ for each query $(Q_{i})$ .", + "bbox": [ + 511, + 578, + 905, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sliding key-value Self Attention. In accordance with Theorem 3.2, we demonstrate that sliding window attention constitutes a generalized form of self-attention, which can be designated as Sliding Key-Value Self-Attention (SkvSA). SkvSA is formulated as follows.", + "bbox": [ + 511, + 608, + 905, + 684 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S k v S A} _ {i} (X) = \\frac {\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 691, + 903, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b_{1}$ and $b_{2}$ are the sliding boundary on the left and right when sliding on the one-dimensional vector $X$ . In our SkvSA, $b_{1} = b_{2} = \\frac{w\\cdot s}{2}$ , where $w$ is the size of the sliding window and $s$ is the stride size during sliding.", + "bbox": [ + 511, + 737, + 905, + 797 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Eq. 3 says that, for the $i$ -th query $Q_{i}$ , SkvSA systematically extracts the key-value pairs from a designated sliding window, extending from the index $i - \\frac{w \\cdot s}{2}$ to the index $i + \\frac{w \\cdot s}{2}$ with stride $s^2$ . Subsequently, SkvSA calculates the attention map and re-weights the feature in place.", + "bbox": [ + 511, + 797, + 905, + 875 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "2We set $w\\cdot s\\leq N$ in experiments thus enabling SkvSA to slide on $X$", + "bbox": [ + 529, + 886, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16049", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Boundary processing. Note that, in scenarios where $i - \\frac{w \\cdot s}{2} < 0$ or $i + \\frac{w \\cdot s}{2} > N$ occur, the slide indices of the key-value pairs $j$ will exceed the boundary of the input $X$ , leading to an indexing failure. To ameliorate the issue associated with slide indexing at sequence's boundaries, we propose a boundary blocking indexing method. Specifically, for queries located at the boundary of the sequence, the sliding indices of key-value pairs are \"blocked\" and converted into indexing within a predefined blocking window. In this study, the sizes of the predefined blocking window and the sliding window are identical. Accordingly, the sliding boundaries $b$ can be appropriately revised as follows:", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nb = \\left(b _ {1}, b _ {2}\\right) = \\left\\{ \\begin{array}{l l} \\left(0, \\frac {w \\cdot s}{2}\\right), & i - \\frac {w \\cdot s}{2} < 0, \\\\ \\left(\\frac {w \\cdot s}{2}, N\\right), & i + \\frac {w \\cdot s}{2} > N, \\\\ \\left(\\frac {w \\cdot s}{2}, \\frac {w \\cdot s}{2}\\right), & \\text {o t h e r s .} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 273, + 483, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The SkvSA and its boundary processing are capable of extracting and processing information via a fixed sliding window, adhering to the translation equivariance property as established in Theorem 3.2.", + "bbox": [ + 89, + 354, + 483, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive Sliding key-value Self Attention. However, the fixed window size constrains the receptive field, leading to the entrapment of restoration transformers within rigidity [48]. To mitigate this rigidity, it is anticipated that the SkvSA will ascertain the optimal window size and position specific to each query $Q_{i}$ . To achieve this endeavor, we devise the Adaptive Sliding key-value Self Attention (ASkvSA), which adaptively determines key-value indices and shuffles them within the fixed sliding window of $Q_{i}$ .", + "bbox": [ + 89, + 416, + 483, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 3 (a), with regard to Key $K$ and Value $V$ , ASkvSA initially reshapes $K$ and $V$ from one-dimensional (1d) vectors into two-dimensional (2d) matrices. Subsequently, it applies a 2d depth-wise convolution with kernel size $k$ to generate their adaptive indices $\\mathcal{F} \\in \\mathbb{R}^{H \\times W \\times 2}$ . $\\mathcal{F}[h, w]$ is a vector of length 2 in the $h-$ th row and $w-$ th column of $\\mathcal{F}$ , which can describe a coordinate. Following this, ASkvSA performs key-value indexing based on established indices $\\mathcal{F}_K$ and $\\mathcal{F}_V$ by shuffling the pixel located at the coordinate $\\mathcal{F}[h, w]$ of $K$ and $V$ to the pixel at the coordinate $(h, w)$ of $K$ and $V$ .", + "bbox": [ + 89, + 553, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A S k v S A} (X) = \\frac {\\sum_ {j = \\mathcal {F} (K) , l = \\mathcal {F} (V)} \\sin \\left(Q _ {i} , K _ {j}\\right) V _ {l}}{\\sum_ {j = \\mathcal {F} (K)} \\sin \\left(Q _ {i} , K _ {j}\\right)}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 727, + 482, + 765 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptive slide indexing does not compromise the translation equivariance of ASkvSA. ASkvSA stacks convolutions with SkvSA in serial to generate adaptive indices of key-value pairs. Given that convolution satisfies the property of translation equivariance, when $K_{h,w}$ is translated to $K_{h + \\delta_h, w + \\delta_w}$ , key's adaptive indices $\\mathcal{F}_K$ will shuffle the pixel located at $\\mathcal{F}[h + \\delta_h, w + \\delta_w]$ of $K$ to the pixel at $(h + \\delta_h, w + \\delta_w)$ of $K$ . According to Theorem 3.3, ASkvSA satisfies the translation equivariance property.", + "bbox": [ + 89, + 765, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Adaptively aggregated with global key-value pairs. Although ASkvSA can identify appropriate key-value pairs beyond a fixed sliding window, some remote but valid pixels still escape ASkvSA's indexing. Consequently, we propose downsampled self-attention (DSA), which adaptively combines global key value indexes through downsampling, providing the network with a coarse, yet efficient global indexing of key-value pairs. DSA can be formulated as:", + "bbox": [ + 511, + 90, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {D S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right) V _ {j} ^ {\\prime}}{\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 213, + 906, + 255 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $K'$ and $V'$ represents the downsampled key-value pair calculated in ASkvSA, and $N_d$ denotes the number of tokens associated with the downsampled key-value pair. We employ average pooling as the downsampling operator due to its efficiency and coarse fulfillment of TE.", + "bbox": [ + 511, + 256, + 906, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, according to Figure 3, the ASkvSA and DSA outputs are adaptively aggregated to obtain our TEA.", + "bbox": [ + 511, + 330, + 905, + 362 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {T E A} _ {i} (X) = \\alpha_ {s} \\operatorname {A S k v S A} (X) + \\alpha_ {d} \\operatorname {D S A} (X), \\quad (7) \\text {w h e r e} \\alpha_ {s} \\text {a n d} \\alpha_ {d} \\text {a r e l a r n a b l e p a r a m e t e r s . T E A s t a c k s D S A w i t h A S k v S A i n p a r a l l e l a r . A c c o r d i n g t o T h e o r e m 3 . 3 , T E A s a t s i s f i e s t h e t r a n s l a t i o n e q u i v a r i a n c e p r o p e r t y .}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 373, + 905, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.3. Computational Cost and Hyper-Parameters", + "text_level": 1, + "bbox": [ + 511, + 444, + 885, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We demonstrate that TEA can maintain a computational cost that is similar to that of window attention [34] under well-designed hyperparameter settings. We first list four integral hyperparameters of TEA:", + "bbox": [ + 511, + 465, + 905, + 527 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- $w$ : the sliding window size of ASkvSA;", + "- $s$ : the sliding stride size of ASkvSA;", + "- $k$ : the kernel size of adaptive indexes generation;", + "- $N_{d}$ : the number of tokens in DSA." + ], + "bbox": [ + 511, + 530, + 851, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "According to Section 4.2, TEA uses $3ND^{2}$ FLOPs for the Query-Key-Value linear projections, $2NDk^{2}$ FLOPs for the depth-wise convolution in adaptive indices generation, $Nw^{2}D$ FLOPs for attention map computation between query $Q \\in \\mathbb{R}^{N \\times D}$ and indexed key $K \\in \\mathbb{R}^{w^{2} \\times D}$ , $Nw^{2}D$ FLOPs for feature re-weighting between attention map and indexed value $V \\in \\mathbb{R}^{w^{2} \\times D}$ , and $2NN_{d}D$ FLOPs for DSA. The total FLOPs of TEA is $3ND^{2} + 2NDk^{2} + 2Nw^{2}D + 2NN_{d}D$ . With the hyper-parameters held constant, the FLOPs of TEA exhibit linear growth as a function of $N$ , thereby implying that TEA is $\\mathcal{O}(N)$ .", + "bbox": [ + 511, + 592, + 905, + 761 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Consequently, when the hyper-parameters of TEA are adjusted carefully, its computational cost becomes comparable to that of window attention. In our TEAFormer, the hyper-parameter $w$ has been set to 15, $k$ to 3, and $N_{d}$ to 16, thereby modulating the computational complexity of TEA to be even slightly lower than that of window attention with window size 16.", + "bbox": [ + 511, + 761, + 905, + 867 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "[45] proved that downsampling can meet strict TE through learnable polyphase (LP). We compare these downsamplers in Sec. 5.6", + "bbox": [ + 511, + 875, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16050", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/64b0051396129607c276140ed5a0c3c152fa7e7876e72dc22a88c18e541b2150.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelwindow sizeTranslation equivarianceParamsFLOPsLatencyPerformanceNTK [25]SRGA [33]
SkvSAASkvSADSA(M)(T)(ms)Urban100 [24] PSNR↑ / SSIM↑Condition ↓Value ↓
SwinIR [31]811.900.462130.027.45 / 0.82541746.493.655
SwinIR-Large1621.420.897214.527.94 / 0.83621554.653.610
TEAFformer15✓21.860.688230.128.31 / 0.8444243.753.206
TEAFformer15✓21.980.757284.428.47 / 0.8457283.993.298
TEAFformer15✓✓21.860.966340.928.49 / 0.8470203.063.261
TEAFformer15✓✓21.981.035386.728.67 / 0.8489236.783.275
", + "bbox": [ + 138, + 89, + 856, + 209 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/dbcd9de97f5c6c5a169c450276b15adf9fd59e7e6d66a3405239969a302eba04.jpg", + "table_caption": [ + "Table 2. Ablation study on TEA. Our TEA enhances performance, convergence speed, and generalization." + ], + "table_footnote": [], + "table_body": "
MethodScaleParams (M)FLOPs (T)Latency (ms)Set5 [5]Set14 [64]B100 [37]Urban100 [24]Manga109 [40]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
EDSR [32]40.76.006147.138.110.960233.920.919532.320.901332.930.935139.100.9773
RCAN [71]15.42.259155.538.270.961434.120.921632.410.902733.340.938439.440.9786
SwinIR-L [31]0.91156.4188.638.140.961133.860.920632.310.901232.760.934039.120.9783
SwinIR [31]11.81.848514.438.420.962334.460.925032.530.904133.810.942739.920.9797
EDT [30]11.51.965101038.450.962434.570.925832.520.904133.800.942539.930.9800
SRFormer [73]×210.31.790787.238.510.962734.440.925332.570.904634.090.944940.070.9802
HAT [9]20.63.662901.938.630.963034.860.927432.620.905334.450.946640.260.9809
GRL-B* [61]20.14.422262138.480.962734.640.926532.550.904533.970.943740.060.9804
DAT [10]14.72.155121538.580.962934.810.927232.610.905134.370.945840.330.9807
IPG [48]16.84.732270338.610.963234.730.927032.600.905234.480.946440.240.9810
TEAFormer-L (Ours)0.830.237254.038.270.861834.260.924032.430.902933.510.939639.560.9792
TEAFormer (Ours)21.84.133149338.640.963135.120.928432.720.906435.230.950940.320.9807
EDSR [32]43.11.85349.1332.460.896828.800.787627.710.742026.640.803331.020.9148
RCAN [71]15.60.58746.1532.630.900228.870.788927.770.743626.820.808731.220.9173
SwinIR-L [31]0.880.03948.7132.440.897628.770.785827.690.740626.470.798030.920.9151
SwinIR [31]11.90.462130.032.920.904429.090.795027.920.748927.450.825432.030.9260
EDT [30]11.60.514251.332.820.903129.090.793927.910.748327.460.824632.030.9254
SRFormer [73]×410.40.470189.232.930.904129.080.795327.940.750227.680.831132.440.9271
HAT [9]20.80.938227.733.040.905629.230.797328.000.751727.970.836832.480.9292
GRL-B* [61]20.21.129761.832.900.903929.140.795627.960.749727.530.827632.190.9266
DAT [10]14.80.561296.833.080.905529.230.797328.000.751527.870.834332.510.9291
IPG [48]17.01.206646.733.150.906229.240.797327.990.751928.130.839232.530.9300
TEAFormer-L (Ours)0.830.05969.6432.630.900228.960.789927.820.744227.040.810531.560.9197
TEAFormer (Ours)22.01.035386.733.230.906629.460.800928.110.754428.670.848932.990.9323
", + "bbox": [ + 94, + 237, + 919, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. Classical image SR results. Performance and complexity are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two scores in each column are highlighted in red and blue. Methods with “*” are replicated with standard training setting, following IPG [48]. The “-L” indicates that the given method is a lightweight version.", + "bbox": [ + 89, + 554, + 906, + 595 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6c7067faf5c2b46ff3b533a01d757f0cb5961cc2b2d1bc6dbf7286c7317ba969.jpg", + "image_caption": [ + "BSD100 $(4\\times)$ :img_86016" + ], + "image_footnote": [], + "bbox": [ + 112, + 609, + 372, + 720 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/840872a5dd68612db0b0565822aa3a0a03c1daf2f1af968502b703786197ec8d.jpg", + "image_caption": [ + "Bicubic" + ], + "image_footnote": [], + "bbox": [ + 377, + 609, + 506, + 657 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a074904d3c0310f9e029dab75edb38e89fd1dde70d61c8b307415fc32b51da01.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 609, + 635, + 657 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/307518198a438373c3d3f5fba824ef264c65e277d8b405e63f303427e51d2395.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 637, + 609, + 763, + 657 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c1d57f9f8802a91cb21a666fd6406dd726a5b0bcd49b17d4204fa05d06d09eeb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 609, + 893, + 657 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3118184afb85fe7f7e9c6c9b032b153a0502791a2113b5b827339d81f3e663b6.jpg", + "image_caption": [ + "Urban100 $(4\\times)$ : img_033" + ], + "image_footnote": [], + "bbox": [ + 112, + 736, + 372, + 847 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ec5b7b7d95ae8cd4d83052467bac870b16d7f4c6d3068194afa984a5310f9a22.jpg", + "image_caption": [ + "HQ" + ], + "image_footnote": [], + "bbox": [ + 377, + 672, + 506, + 720 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ad1d99216bd1f3a2bc3bc267b173c9b9eb2d86f04b3b223c383297d283b94b45.jpg", + "image_caption": [ + "RCAN [71]" + ], + "image_footnote": [], + "bbox": [ + 506, + 672, + 635, + 720 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e47ff34894efd7c4063afe023e1eb6e2c71c9c921e33ae87b966fdfe49565a52.jpg", + "image_caption": [ + "SwinIR [31]" + ], + "image_footnote": [], + "bbox": [ + 637, + 672, + 763, + 720 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1353974588e7314d074f12197e44c1956eba46dfb7ad889f5b78f9c5bfc9d1ad.jpg", + "image_caption": [ + "HAT [9]" + ], + "image_footnote": [], + "bbox": [ + 764, + 672, + 893, + 720 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ede9013da5424406bfd1cb581208ffe1035bff4ba5cb2cac8470effb764450be.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 379, + 736, + 506, + 784 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/69c463004db8e8a1705f04e74eda412ad0d36d07a2125ef0e0757e4099bcdb9a.jpg", + "image_caption": [ + "SRFormer [73]" + ], + "image_footnote": [], + "bbox": [ + 506, + 736, + 635, + 784 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/679fa1161e951e2f67436f92d5fc7aaf7c52e7a89e2ca1ea7a670cfccc162c55.jpg", + "image_caption": [ + "IPG [48]" + ], + "image_footnote": [], + "bbox": [ + 637, + 736, + 763, + 784 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/95586d9d0a06acbb543721aa1974282dfb39f2f47fa3cdc8eb04c1025e685100.jpg", + "image_caption": [ + "TEAFormer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 736, + 893, + 784 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/59051c683c0dd65276b62f2680811ce0bf7a606ef16ca7f15f2cb75e8b9b37c7.jpg", + "image_caption": [ + "Bicubic", + "HQ" + ], + "image_footnote": [], + "bbox": [ + 379, + 799, + 506, + 847 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f0d908e7afd0c624201203807a2cea98a13f5518991f55a8ef3a75c7c45578db.jpg", + "image_caption": [ + "RCAN [71]", + "SRFormer [73]", + "Figure 4. Qualitative comparison with recent SOTA methods on the image SR task." + ], + "image_footnote": [], + "bbox": [ + 506, + 799, + 635, + 847 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6fd2f6a1eddd722cd5726cbeeba4e00fcc09af5cd030d1f27335a21db28fd2ba.jpg", + "image_caption": [ + "SwinIR [31]", + "IPG [48]" + ], + "image_footnote": [], + "bbox": [ + 637, + 799, + 763, + 847 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0cc8778a261620189553d400b5e40ce540bda41f8bcdce79b24bbdddb39c0df5.jpg", + "image_caption": [ + "HAT [9]", + "TEAFormer (Ours)" + ], + "image_footnote": [], + "bbox": [ + 764, + 799, + 893, + 847 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16051", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/44397cee24ca27691a902e183c701519d89297e633ad9f3d3b76e56dbfbf1186.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsIndoor ScenesOutdoor ScenesCombined
(M)PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓
DPDNetS [1]32.326.540.8160.0310.23922.250.6820.0560.31324.340.7470.0440.277
KPACs [47]2.0627.970.8520.0260.18222.620.7010.0530.26925.220.7740.0400.227
IFANs [27]10.528.110.8610.0260.17922.760.7200.0520.25425.370.7890.0390.217
RestormerS [63]26.128.870.8820.0250.14523.240.7430.0500.20925.980.8110.0380.178
SFNetS [12]13.229.160.8780.0230.16823.450.7470.0490.24426.230.8110.0370.207
GRLS-B [61]19.929.060.8860.0240.13923.450.7610.0490.19626.180.8220.0370.168
IRNeXTS [13]13.229.220.8790.0240.16723.530.7520.0490.24426.300.8140.0370.206
TEAFormerS (Ours)15.429.500.8920.0230.15023.550.7670.0500.21126.450.8280.0370.181
DPDNetD [1]32.327.480.8490.0290.18922.900.7260.0520.25525.130.7860.0410.223
RDPD_D [2]8.2028.100.8430.0270.21022.820.7040.0530.29825.390.7720.0400.255
IFAN_D [27]10.528.660.8680.0250.17223.460.7430.0490.24025.990.8040.0370.207
Restormer_D [63]26.129.480.8950.0230.13423.970.7730.0470.17526.660.8330.0350.155
Uformer_D [54]20.628.230.8600.0260.19923.100.7280.0510.28525.650.7950.0390.243
GRL_D-B [61]19.929.830.9030.0220.11424.390.7950.0450.15027.040.8470.0340.133
TEAFormer_D (Ours)15.430.160.9050.0220.12824.450.7950.0450.17027.230.8490.0340.150
", + "bbox": [ + 99, + 80, + 893, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 4. Defocus deblurring comparisons on the DPDD [1] testset (containing 37 indoor and 39 outdoor scenes). S: single-image defocus deblurring. D: dual-pixel defocus deblurring. The best two results are highlighted in red and blue, respectively.", + "bbox": [ + 89, + 297, + 903, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 332, + 223, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Ablation study", + "text_level": 1, + "bbox": [ + 89, + 358, + 240, + 375 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We analyze the influence exerted by the TEA on both performance, computational complexity, generalization, and training convergence in Table 2. In this experiment, we disassemble TEA into several plug-and-play modules and gradually add them into the SwinIR baseline restoration network. To ensure comparable parameters and FLOPs among the models being evaluated, we adjusted the window size of SwinIR to 16 and increased the number of parameters to 21 M, which is termed SwinIR-Large. Following the incorporation of our SkvSA, there is a marked improvement in the convergence rate and generalization abilities. Subsequent to the integration of our ASkvSA and DSA, there is a progressive enhancement in the model's performance, without detrimental effects on convergence and generalization. The results of this experiment demonstrate the efficacy of each component within the TEA. Furthermore, it illustrates that TEA resolves the dilemma between self-attention and sliding window attention, thereby attaining high performance by the global receptive field and reducing computational complexity simultaneously.", + "bbox": [ + 89, + 380, + 483, + 683 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Image SR", + "text_level": 1, + "bbox": [ + 89, + 695, + 202, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We used the same training hyperparameters and evaluation methods as IPG [48]. All of our models were trained from scratch on the DF2K [3, 32] datasets with $64 \\times 64$ resolution. The optimizer was Adam [26], the learning rate was $2 \\times 10^{-4}$ while using the cosine scheduler. We used $\\mathrm{L}_1$ loss as a criterion.", + "bbox": [ + 89, + 718, + 482, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We first compare TEAFormer with various classical image SR methods, including EDSR [32], RCAN [71], SwinIR [31], EDT [30], HAT [9], SRFormer [73], GRL [61], DAT [10], and IPG [48]. We compare them on synthetic benchmarks [5, 24, 37, 40, 64] generated by bicubic downsampling. The FLOPs and latency reported were", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "calculated with one NVIDIA A800 GPU on 768 resolution.", + "bbox": [ + 511, + 334, + 903, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative comparison of the methods for classical image SR is shown in Table 3. The exceptional performance of TEAFormer is distinctly evident as it achieves the highest results across nearly all five datasets for various scale factors. Incorporating the attention module with translation equivalence introduces a more suitable and productive inductive bias to the restoration model, thereby enhancing its performance. For the $\\times 4$ SR, our TEAFormer achieves a $28.67\\mathrm{dB}$ PSNR score on the Urban100 dataset, which is $0.7\\mathrm{dB}$ higher than HAT [9] while both has $20\\mathrm{M}$ parameters. In comparison to the preceding image SR method IPG [48], TEAFormer demonstrates a gain of $0.54\\mathrm{dB}$ with comparable FLOPs and a $2\\times$ faster inference.", + "bbox": [ + 509, + 349, + 906, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further conduct TEAFormer-Light, which only comprises 829k parameters. In comparison to the CNN method HAN [42], TEAFormer-Light has been demonstrated to achieve superior performance with only its $1.3\\%$ parameters. In comparison to other lightweight transformer methods, TEAFormer also demonstrates an absolute performance advantage, as evidenced by the fact that TEAFormer-Light achieves a higher PSNR than SwinIR-L on the Urban100 [24] dataset by 0.57 dB.", + "bbox": [ + 511, + 547, + 908, + 684 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Image Defocus Deblurring", + "text_level": 1, + "bbox": [ + 511, + 695, + 754, + 710 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training settings. We employed the same training hyperparameters and evaluation as Restormer [63] in image deblurring tasks. Our TEAFormer were trained from scratch using progressive learning [63]. During the training phase, the optimizer adopted was AdamW [35] with $\\mathrm{L}_1$ loss.", + "bbox": [ + 511, + 718, + 903, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We compare our TEAFormer with single image defocus deblurring methods [1, 13, 14, 27, 47, 61, 63] and dual-pixel image defocus deblurring methods [1, 2, 27, 54, 61, 63] on DPDD [1] dataset in Table 4. Our TEAFormer significantly outperforms state-of-the-art schemes for the single-image and dual-pixel defocus deblurring tasks on PSNR and SSIM metrics. In particular in the category", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16052", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a50c42c38dc7fd6d4882f9cf2e7aae606bdf0310ed3247b4116b9e2848ca586.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 88, + 372, + 213 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e8d8386e85b8b68b24f1c3464dac06ca8cde59ef568637255ad28223227d1348.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with recent SOTA methods on the image defocus blur task. Please zoom in for better view." + ], + "image_footnote": [], + "bbox": [ + 377, + 88, + 504, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5535a2ec071383096f6e31e4a7fa1f9d438cef59cf2b65e394d3c73871c5cc37.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 506, + 88, + 633, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c0ee018cd4b939932d67af6dd9095e6507b1f7cba06de2478781c79b1cc0af70.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 88, + 761, + 214 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e93c9b4174e7f866d190746fde1dcb62a83bdfd4643e24cb48fc549b7a90bb62.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 764, + 88, + 892, + 214 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1b579d8d693be77c2ea08f772aa9ba036bf371f79825edb0124f23775de711df.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsColorGrayscale
(M)CBSD68 [39]Kodak24 [17]McMaster [69]Urban100 [24]Set12 [67]BSD68 [39]Urban100 [24]
σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50
DnCNN [67]0.5633.9031.2427.9534.6032.1428.9533.4531.5228.6232.9830.8127.5932.8630.4427.18
DRUNet [68]32.634.3031.6928.5135.3132.8929.8635.4033.1430.0834.8132.6029.6133.2530.9427.90
EDT-B [30]11.534.3931.7628.5635.3732.9429.8735.6133.3430.2535.2233.0730.16---
SwinIR* [31]11.734.3931.7828.5535.3332.8829.7735.5933.2030.1935.1132.8929.7733.3431.0027.88
Restormer [63]26.134.3931.7828.5935.4433.0230.0035.5533.3130.2935.0632.9130.0233.3531.0428.01
ART* [66]16.234.4031.7928.5935.4733.0430.0035.6133.3330.2735.1232.9529.9933.3531.0528.01
GRL-B* [61]19.834.4031.7828.5835.4233.0330.0235.6333.3630.2635.2233.0130.2433.3431.0227.99
TEAFormer (Ours)15.434.4031.8028.6235.5033.0830.0535.6833.4230.3835.2533.1830.3733.3731.0628.04
", + "bbox": [ + 94, + 232, + 906, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Color and grayscale image Gaussian denoising results in terms of PSNR $\\uparrow$ . Model complexity and prediction accuracy are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two results are highlighted in red and blue, respectively. Methods with \" $*$ \" are replicated with standard training setting, which learns a single model to handle various noise levels.", + "bbox": [ + 89, + 371, + 906, + 412 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "of single-pixel indoor scene, TEAFormer yields 0.44 dB improvements over the GRL previous best transformer method [61]. Figure 5 illustrates that our TEAFormer is more effective in removing spatially varying defocus blur than other restoration transformers.", + "bbox": [ + 89, + 414, + 482, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Image Denoising", + "text_level": 1, + "bbox": [ + 89, + 494, + 256, + 512 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training settings. The image denoising and defocus deblurring tasks used the same hyper-parameters for training.", + "bbox": [ + 89, + 518, + 482, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Gaussian Denoising Results. We perform denoising experiments on synthetic benchmark datasets generated with additive white Gaussian noise (Set12 [67], BSD68 [38], Urban100 [24], Kodak24 [17] and McMaster [69]). Following DRUNet [68], we learn a single model to handle various noise levels, including 15, 25, and 50. Our TEAFourner achieves state-of-the-art performance under both experimental settings on different datasets and noise levels. Specifically, for the challenging noise level 50 on color Urban100 [24] dataset, TEAFourner achieves 0.13 dB gain over the previous best restoration transformer GRL [61], as shown in Table 5. Similar performance gains can be observed for the Gaussian gray denoising.", + "bbox": [ + 89, + 549, + 482, + 746 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.5. All-in-one Image Restoration", + "text_level": 1, + "bbox": [ + 89, + 757, + 349, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training settings. Following DCPT [23], a 5D all-in-one image restoration and enhancement task was conducted with the objective of training the network to eliminate five distinct types of degradation: haze, rain, Gaussian noise, motion blur, and low light. The models were trained from scratch on datasets with $128 \\times 128$ resolution without progressive learning. During training, the optimizer was set as AdamW, the learning rate was set as $3 \\times 10^{-4}$ while using", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the cosine scheduler. We used $L_{1}$ loss as a criterion.", + "bbox": [ + 511, + 414, + 857, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. As illustrated in Table 6, TEAFormer surpasses the most recently proposed methods specifically designed for comprehensive all-in-one restoration tasks (IDR [65], PromptIR [43], and InstructIR [11]). The average PSNR of TEAFormer in all five tasks reached 31.22 dB, thereby exceeding the current state-of-the-art 5D all-in-one restoration method, InstructIR [11], by 1.67 dB. TEAFormer consistently outperforms prior PSNR in every indicator of each task, signifying its exceptional proficiency in image restoration under complex multi-task conditions.", + "bbox": [ + 511, + 430, + 905, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mixed degradation. As shown in Table 7, TEAFormer surpasses the most recently proposed methods under mixed degradation scenarios, specifically designed for universal restoration tasks [36, 72]). Within scenarios characterized by a combination of three degradation (low-light, haze, and rain), TEAFormer demonstrates superior performance over the current state-of-the-art universal restoration method, DiffUIR [72], by a margin of $0.42\\mathrm{dB}$ .", + "bbox": [ + 511, + 582, + 905, + 704 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.6. Discussion", + "text_level": 1, + "bbox": [ + 513, + 710, + 632, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameters analysis. The performance under various hyperparameters is reported in Table 8 $(w)$ , Table 9 $(s)$ , Table 10 $(k)$ and Table 11 $(N_d)$ . Our settings $(w = 15, s = 4, k = 3, N_d = 16)$ optimally balanced performance and computation cost. Note that changing some of the hyperparameters does not affect the computational complexity as discussed in Section 4.3. For $s$ in Table 9, parameters and FLOPs are omitted because $s$ does not affect them.", + "bbox": [ + 511, + 733, + 905, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Compare different downsampler in DSA. The incorporation of DSA into our TEA substantially enhances the performance of the model without adding undue complexity.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16053", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/906585e99da5fe0b28a6ecc870100a8fcf2339604726a6e09056c572bf8d4d15.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodParamsDehazing on SOTS [28]Deraining on Rain100L [60]Denoising on BSD68 [38]Deblurring on GoPro [41]Low-Light on LOL [55]Average
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.04 / 0.88432.98 / 0.95130.91 / 0.88224.35 / 0.78118.18 / 0.73525.49 / 0.846
IDR [65]15.325.24 / 0.94335.63 / 0.96531.60 / 0.88727.87 / 0.84621.34 / 0.82628.34 / 0.893
PromptIR [43]35.625.20 / 0.93135.94 / 0.96431.17 / 0.88227.32 / 0.84220.94 / 0.79928.11 / 0.883
InstructIR [11]35.127.10 / 0.95636.84 / 0.97331.40 / 0.88729.40 / 0.88623.00 / 0.83629.55 / 0.907
SwinIR [31]11.821.50 / 0.89130.78 / 0.92330.59 / 0.86824.52 / 0.77317.81 / 0.72325.04 / 0.835
NAFNet [7]67.925.23 / 0.93935.56 / 0.96731.02 / 0.88326.53 / 0.80820.49 / 0.80927.76 / 0.881
Restormer [63]26.124.09 / 0.92734.81 / 0.96231.49 / 0.88427.22 / 0.82920.41 / 0.80627.60 / 0.881
TEAFormer (Ours)15.431.57 / 0.98039.44 / 0.98631.54 / 0.89030.53 / 0.90823.06 / 0.85631.22 / 0.924
", + "bbox": [ + 127, + 88, + 866, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a4d8814858f9972493a82632b410fcd27208c6746a1d3afddb67a07a59ae5fb7.jpg", + "image_caption": [ + "LQ" + ], + "image_footnote": [], + "bbox": [ + 102, + 281, + 218, + 342 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b1f2300adaa7ed38d31071431e54bd8c0cd46762e74eaabb3a3173c529cc93f.jpg", + "image_caption": [ + "InstrucrIR [11]" + ], + "image_footnote": [], + "bbox": [ + 220, + 281, + 334, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71d0654490210f525dafda7e9d73ac836263dad432f0b7b25f8018f4823c5362.jpg", + "image_caption": [ + "OneRestore [19]" + ], + "image_footnote": [], + "bbox": [ + 336, + 281, + 452, + 340 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99e7714b36460186d16a8eedbb92ac31227c3593ab5eede043fa2d827c0d16a1.jpg", + "image_caption": [ + "HQ" + ], + "image_footnote": [], + "bbox": [ + 102, + 354, + 218, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/12c8e9e1664bf7e969153bac10b2f9fe3a715e5b58f0b44559a78b83ac274668.jpg", + "image_caption": [ + "DACLIP [36]" + ], + "image_footnote": [], + "bbox": [ + 220, + 354, + 334, + 412 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/56d5bc26ccd2281c8dd0451fbde398d2e161d806ef052f8c8ed7de148be47eb4.jpg", + "image_caption": [ + "TEAFormer (Ours)", + "Figure 6. Visual comparison on low-light+haze+rain mixed degradation. TEAFormer can restore the correct lightness." + ], + "image_footnote": [], + "bbox": [ + 336, + 354, + 452, + 412 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b1feee8dcbd171a697ec6745ecea3792e7507796b287e534d89e7a063522b7ee.jpg", + "table_caption": [ + "Table 6. 5D All-in-one image restoration results. Dash lines separate the results of methods designed specifically for all-in-one image restoration task and methods that only focus on architecture enhancement. The best two results are highlighted in red and blue, respectively. TEAFormer sets the new state-of-the-art performance in each task." + ], + "table_footnote": [], + "table_body": "
MethodParams low-light+haze+rain low-light+haze+snow
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.80 / 0.70822.23 / 0.725
TransWeather [50]31.021.55 / 0.67821.01 / 0.655
PromptIR [43]35.623.74 / 0.75223.33 / 0.747
DACLIP [36]17422.96 / 0.71223.32 / 0.751
DiffUIR-L [72]36.325.46 / 0.77925.89 / 0.788
OneRestore [19]5.9825.18 / 0.79525.28 / 0.797
InstructIR [11]35.124.84 / 0.77724.32 / 0.760
TEAFormer (Ours)15.425.88 / 0.79626.05 / 0.800
", + "bbox": [ + 470, + 287, + 901, + 435 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/330b0a532757ac0a12fd8e07503765531a9033b55be2779c9ae27e183922ad9d.jpg", + "table_caption": [ + "Table 7. Mixed degraded image restoration results on CDD [19]." + ], + "table_footnote": [ + "Table 8. Ablations of sliding window size $w$ ." + ], + "table_body": "
wFLOPs (T)Urban100 [24] PSNR ↑/SSIM ↑
70.82728.35 / 0.8441
151.03528.67 / 0.8489
311.86628.73 / 0.8493
", + "bbox": [ + 93, + 455, + 254, + 540 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cc3849e77d2774eae9cf70053bd5677780ece7d421b1088f3ac4a073ed69ac19.jpg", + "table_caption": [], + "table_footnote": [ + "Table 9. Ablations of sliding stride $s$" + ], + "table_body": "
sUrban100 [24]
PSNR ↑ / SSIM ↑
128.43 / 0.8431
228.51 / 0.8453
428.67 / 0.8489
", + "bbox": [ + 256, + 455, + 369, + 540 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/cdfa300054871c267f0fa5857091b820b3e317e8dfc07ba3c97d8062c6bbf6d4.jpg", + "table_caption": [], + "table_footnote": [ + "Table 10. Ablations of kernel size $k$ for adaptive in- in DSA. \"None\" stands for indices. \"None\" stands for no no DSA used. adaptive indices used.", + "Table 12. Ablations of down-samplers in DSA. The higher the S-Conv, the more strictly the model meets TE." + ], + "table_body": "
kFLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Nd(T)FLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Down (%)↑S-conv (%)↑Urban100 [24] PSNR ↑ / SSIM ↑
None0.96628.49 / 0.8470None0.75728.47 / 0.8457MaxPool87.328.61 / 0.8477
11.02828.53 / 0.8471161.03528.67 / 0.8489LPD91.728.64 / 0.8493
31.03528.67 / 0.8489N18.3628.80 / 0.8503AvgPool91.228.67 / 0.8489
", + "bbox": [ + 370, + 455, + 901, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "However, the average pooling employed in DSA does not strictly adhere to the property of TE. Using a comparative ablation of several downsampling modules in Table 12, we experimentally demonstrate that the use of average pooling does not compromise the model's performance or its TE.", + "bbox": [ + 89, + 599, + 483, + 675 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Does TEAFormer converge faster and generalize better in practice? Table 1 and Table 2 have illustrated that TEAFormer converges faster and generalizes better than other networks theoretically due to the incorporation of translation equivariance. Figure 7 (left) further demonstrates that TEAFormer reaches a lower loss faster than other restoration transformers [9, 31] in practice. Following MaskDenoise [6], we compare the statistics of the last layer feature in SwinIR [31] and TEAFormer under two different real-world degradations on the PIES [33] dataset (out of the training dataset) in Figure 7 (right). The results show that TEAFormer can produce a more consistent output when processing LQ images with different degradations. Therefore, TEAFormer generalizes better in practice.", + "bbox": [ + 89, + 695, + 483, + 906 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e486577d5539c649550783b4f33df6f06ed8f4e981024632d753d3b438cedb94.jpg", + "image_caption": [ + "Figure 7. Comparison on convergence (left) and generalization (right)." + ], + "image_footnote": [], + "bbox": [ + 521, + 599, + 673, + 674 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/314d73fb472fdca92b1593ae33e8790d037d0eea31edcf1e6bcddd02dbf5d820.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 669, + 606, + 781, + 672 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c275121ffca482548733363c9f7ee340838c565ac730280273df7cb45c6666c6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 784, + 607, + 892, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 696, + 632, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we start by examining the translation equivariance, which serves as an inherent inductive bias in image restoration. We then define the translation equivariance and offer two strategies for incorporating it into restoration networks: (1) slide indexing and (2) component stacking. To address the fixed receptive field provided by these strategies, we proposed TEA and TEAFormer via adaptive key-value indexing. The efficacy of TEAFormer has been demonstrated in various image restoration tasks, with ablation studies confirming that this efficacy arises from the incorporation of TEA. In the future, we intend to explore more equivariances into restoration networks.", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16054", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 257, + 187, + 273 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 6", + "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6", + "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6", + "[4] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 2", + "[5] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference. BMVA press, 2012. 5, 6", + "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 8", + "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision - ECCV 2022, pages 17-33, Cham, 2022. Springer Nature Switzerland. 8", + "[8] Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre-training or strong data augmentations. In International Conference on Learning Representations, 2022. 1", + "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2, 5, 6, 8", + "[10] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In ICCV, 2023. 2, 5, 6" + ], + "bbox": [ + 93, + 282, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported financially by the Natural Science Foundation of China (82371112, 623B2001), the Science Foundation of Peking University Cancer Hospital (JC202505), Natural Science Foundation of Beijing Municipality (Z210008) and the Clinical Medicine Plus X - Young Scholars Project of Peking University, the Fundamental Research Funds for the Central Universities (PKU2025PKULCXQ008).", + "[11] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 7, 8", + "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2022. 6", + "[13] Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, and Alois Knoll. Irnext: Rethinking convolutional network design for image restoration. In Proceedings of the 40th International Conference on Machine Learning, 2023. 6", + "[14] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 6, 7", + "[15] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 2", + "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2", + "[17] Rich Franzen. Kodak lossless true color image suite, 2013. 7", + "[18] Jonathan Gordon, Wessel P. Bruinsma, Andrew Y. K. Foong, James Requeima, Yann Dubois, and Richard E. Turner. Convolutional conditional neural processes. In International Conference on Learning Representations, 2020. 1", + "[19] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In European Conference on Computer Vision, 2024. 8", + "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185-6194, 2023. 2", + "[21] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2", + "[23] JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Universal image restoration pre-training via degradation classification. The Thirteenth International Conference on Learning Representations, 2025. 7" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16055", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7, 8", + "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 1, 5", + "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 6", + "[27] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6", + "[28] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 2018. 8", + "[29] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, 2022. 8", + "[30] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1089–1097. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 5, 6, 7", + "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 5, 6, 7, 8", + "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 5, 6", + "[33] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497-14513, 2023. 1, 5, 8", + "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 4", + "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6", + "[36] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for universal image restoration. In The Twelfth International Conference on Learning Representations, 2023. 7, 8", + "[37] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423 vol.2, 2001. 5, 6", + "[38] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 7, 8", + "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7", + "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools Appl., 76(20):21811-21838, 2017. 5, 6", + "[41] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 8", + "[42] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 191-207. Springer, 2020. 6", + "[43] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. NeurIPS, 2023. 7, 8", + "[44] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self-attention in vision models. Advances in neural information processing systems, 32, 2019. 2", + "[45] Renan A Rojas-Gomez, Teck-Yian Lim, Alex Schwing, Minh Do, and Raymond A Yeh. Learnable polyphase sampling for shift invariant and equivariant convolutional networks. Advances in Neural Information Processing Systems, 35:35755-35768, 2022. 4", + "[46] Renan A Rojas-Gomez, Teck-Yian Lim, Minh N Do, and Raymond A Yeh. Making vision transformers truly shift-equivariant. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5568-5577, 2024. 2", + "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proc. ICCV, 2021. 6", + "[48] Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang. Image processing gnn: Breaking rigidity in super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24108-24117, 2024. 2, 4, 5, 6", + "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16056", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[50] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 8", + "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[52] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2", + "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2", + "[54] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 6", + "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In British Machine Vision Conference, 2018. 8", + "[56] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 2", + "[57] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2", + "[58] Renjun Xu, Kaifan Yang, Ke Liu, and Fengxiang He. $e(2)$ -equivariant vision transformer. In Uncertainty in Artificial Intelligence, pages 2356-2366. PMLR, 2023. 2", + "[59] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vi-tae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34, 2021. 2", + "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. TPAMI, 2019. 8", + "[61] Xiaoyu Xiang Yawei Li, Yuchen Fan, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6, 7", + "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 1", + "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + ], + "bbox": [ + 91, + 92, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 3, 6, 7, 8", + "[64] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference, Avignon, France*, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 6", + "[65] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 7, 8", + "[66] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 7", + "[67] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 7", + "[68] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 1, 7", + "[69] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 7", + "[70] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. arXiv preprint arXiv:2202.10108, 2022. 2", + "[71] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 5, 6", + "[72] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7, 8", + "[73] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5, 6", + "[74] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 796 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16057", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_model.json b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_model.json new file mode 100644 index 0000000000000000000000000000000000000000..919551323888e68c64db3147f916ebb9106ea4d5 --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_model.json @@ -0,0 +1,3181 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.263, + 0.131, + 0.738, + 0.177 + ], + "angle": 0, + "content": "Enhancing Image Restoration Transformer via Adaptive Translation Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.128, + 0.203, + 0.87, + 0.291 + ], + "angle": 0, + "content": "JiaKui Hu\\(^{1,2,3}\\), Zhengjian Yao\\(^{1,2,3}\\), Lujia Jin\\(^{4}\\), Hangzhou He\\(^{1,2,3}\\), Yanye Lu\\(^{1,2,3*}\\) \n\\(^{1}\\)Institute of Medical Technology, Peking University Health Science Center, Peking University \n\\(^{2}\\)Biomedical Engineering Department, College of Future Technology, Peking University \n\\(^{3}\\)National Biomedical Imaging Center, Peking University \n\\(^{4}\\)China Mobile Research Institute" + }, + { + "type": "text", + "bbox": [ + 0.31, + 0.296, + 0.684, + 0.31 + ], + "angle": 0, + "content": "jkhu29@stu.pku.edu.cn, yanye.lu@pu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.345, + 0.327, + 0.361 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.376, + 0.485, + 0.739 + ], + "angle": 0, + "content": "Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivalence Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.749, + 0.222, + 0.765 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.774, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Image restoration aims to reconstruct high-quality (HQ) images from their low-quality (LQ) counterparts, which are often affected by degradations resulting from the imaging system or environment. As a constituent of the imaging system, image restoration methods should satisfy the following fidelity property: the restoration results for the target region should remain equivariant under geometric transfor-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.346, + 0.907, + 0.392 + ], + "angle": 0, + "content": "mations of the input. This fidelity property is termed translation equivariance, which is one of the inherent inductive biases in image restoration." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.401, + 0.584, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.59, + 0.402, + 0.659, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.402, + 0.735, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.464, + 0.585, + 0.526 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.591, + 0.464, + 0.658, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.668, + 0.464, + 0.734, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.398, + 0.897, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.741, + 0.463, + 0.897, + 0.529 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.532, + 0.907, + 0.616 + ], + "angle": 0, + "content": "Figure 1. \\(\\mathcal{T}(\\cdot)\\) means the translation function. \"SA\" and \"WA\" are commonly used for self-attention and window attention, respectively, but disrupt translation equivariance (TE) due to position encoding and feature shifting. \"TEA\" is our proposed translation equivariance adaptive attention, which satisfies TE. TE promotes faster convergence and better generalization." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.621, + 0.915, + 0.69 + ], + "angle": 0, + "content": "
AttentionsTEPerformanceNTK [25]SRGA [33]Complexity
PSNR↑/ SSIM↑Condition ↓Value ↓
SAX28.21 / 0.843511140.444.075O(N2)
WAX27.45 / 0.82541746.493.655O(N)
TEA (Ours)✓28.67 / 0.8489236.783.275O(N)
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.692, + 0.907, + 0.761 + ], + "angle": 0, + "content": "Table 1. In \\(4\\times\\) image super-resolution, we study performance, NTK [8] convergence speed, SRGA [33] generalization, and computation complexities. We replace WA in SwinIR with SA and TEA for testing. TEA shows fast convergence, better generalization, and linear complexity. The results are tested on Urban100." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.767, + 0.907, + 0.843 + ], + "angle": 0, + "content": "In restoration networks [32, 62, 68, 71] based on convolution neural networks (CNNs), the paradigm of sliding information extraction inherent to convolutions [18] imparted a natural translation equivariance (TE) property in restoration networks. However, as shown in Figure 1 and Table 1," + }, + { + "type": "page_footnote", + "bbox": [ + 0.512, + 0.852, + 0.909, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Translation equivariance and translation invariance are concepts that are frequently misconstrued. Our paper is dedicated to the analysis of equivariance. We have provided a clear delimitation and discussion concerning these concepts in the supplementary material." + }, + { + "type": "footer", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16047" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.151 + ], + "angle": 0, + "content": "the advent of Self Attention (SA) [51] and Window Attention (WA) [34] within Transformers has broken this fidelity property, leading to slow convergence and poor generalization in restoration transformers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.153, + 0.483, + 0.333 + ], + "angle": 0, + "content": "In this paper, we propose two fundamental strategies to bring TE back into the restoration transformers: (1) slide indexing and (2) component stacking. The slide indexing strategy employs a fixed container, e.g., a fixed rectangular window, to globally extract the input in a sliding fashion. This method requires the operators to calculate instantaneously, thereby maintaining the consistency of positional coordinates between the input and output of the model, which facilitates the integration of TE. The component stacking strategy involves stacking modules that satisfy the TE property in parallel or in series, thus enhancing architectures while ensuring TE in complex networks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.483, + 0.56 + ], + "angle": 0, + "content": "However, these foundational strategies present a persistent challenge in the design of restoration transformers. Selecting SA to achieve superior performance involves accepting its substantial computational complexity [63]; alternatively, opting for sliding window attention provides linear complexity but results in a fixed receptive field [48] and reduced performance. To break this dilemma, we replace the global key-value pair indexing mechanism in vanilla SA with an adaptive slide indexing mechanism to design the Translation Equivariance Adaptive (TEA) attention. To retain the extraction of global information, TEA incorporates a downsampling self-attention branch without position encoding, which reduces the key-value pair's resolution by aggregating them, thereby providing TEA with coarse, yet effective global information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.561, + 0.483, + 0.756 + ], + "angle": 0, + "content": "In summary, our contributions are threefold. First, we investigate TE in image restoration, analyzing its impact on convergence and generalization. The results indicate that the integration of TE is essential in the design of restoration networks. Second, we propose TEA, which uses adaptive slide indexing methods to incorporate the TE property while preserving linear complexity and ensuring a global receptive field, thereby enhancing restoration transformers. Third, we introduce TEA to the vanilla Vision Transformer to design TEAFormer. Experiments show that TEA achieves State-of-The-Art (SoTA) performance in various image restoration tasks, while having faster convergence and more robust generalization." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.77, + 0.233, + 0.787 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Vision Transformers with inductive bias. Following the pioneering introduction of the Transformer [51] into vision tasks by ViT [16], SA has emerged as a key component in vision networks. However, ViT is challenging to train, exhibits slow convergence, and requires substantial data support [49]. Previous works, such as ViTAE [59, 70], attributed this to a lack of inductive bias and advocated in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "tegrating inductive bias with ViT to develop more efficient ones. AACNet [4] emphasized the importance of translation equivariance for visual tasks. The Swin [34] and CSwin [15] Transformers incorporated local processing into the vision Transformer, thus proposing window attention. CVT [56] and FasterViT [21] sequentially stack convolutions and attention layers, resulting in a hybrid architecture. PVT [52, 53] used average pooling for downsampling to achieve weak translation equivariance. Both DAT [57] and BiFormer [74] dynamically determined the windows accessible to each token. Recently, some studies have introduced translation equivariance into ViT by modifying position encoding [4, 58] and attention calculation logic [44, 46]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.545 + ], + "angle": 0, + "content": "Recent image restoration Transformers. Similar to the development of vision transformers, the design of image restoration models also focuses on how to better design local global information processors [9, 31, 61, 63, 73]. Since the resolution of the input image is generally high in the image restoration task, directly using SA will result in a huge computation cost. To address this issue, SwinIR [31] introduced window attention [34]. HAT [9] enlarged the receptive field of the restoration transformers with the squeeze-excitation module [22] and the overlapping mechanism. Restormer [63] and DAT [10] performed channel-wise attention, thereby facilitating more efficient feature re-weighting. GRL [61] extracted the feature on local, regional, and global scales by integrating the attention of the anchor window. IPG [48] exploited the flexibility of the network by proposing the degree-variant graph solution, breaking the rigidity of the previous methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.559, + 0.753, + 0.575 + ], + "angle": 0, + "content": "3. Translation Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.628 + ], + "angle": 0, + "content": "In this section, we define the translation equivariance and propose two fundamental strategies for integrating translation equivariance into restoration networks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.629, + 0.905, + 0.675 + ], + "angle": 0, + "content": "Definition 3.1. Translation Equivalence. We call function \\(\\Phi(\\cdot)\\) is translation equivariant to translation operation \\(\\mathcal{T}(\\cdot)\\), if \\(\\Phi(\\mathcal{T}(x)) = \\mathcal{T}(\\Phi(x))\\), where \\(x\\) is the input signal." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.675, + 0.905, + 0.718 + ], + "angle": 0, + "content": "In image restoration tasks, \\( x \\) is an image or a feature in the latent space typically, \\( \\mathcal{T}(\\cdot) \\) refers to a shift in pixels [20]. Based on this definition, two theorems are derived:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.719, + 0.905, + 0.763 + ], + "angle": 0, + "content": "Theorem 3.2. If \\(\\Phi(x)_i\\) is transformed from \\(x_j = [i - b, i + b]\\), where \\(b\\) is the sliding boundary, the function \\(\\Phi(\\cdot)\\) fulfills Definition 3.1." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.765, + 0.905, + 0.809 + ], + "angle": 0, + "content": "Theorem 3.3. If the functions \\(\\Phi_1(\\cdot)\\) and \\(\\Phi_2(\\cdot)\\) fulfill Definition 3.1, the functions \\(\\Phi_2(\\Phi_1(\\cdot))\\) and \\(\\Phi_1(\\cdot) + \\Phi_2(\\cdot)\\) also fulfill Definition 3.1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Reformulating the existing operators in line with Theorem 3.2 gives them the translation equivariance property. By stacking these operators, either in parallel or in serial, as specified in Theorem 3.3, more complex architectures are also capable of translation equivariance. Consequently, the model can be further enhanced via translation equivariance." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "16048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.092, + 0.485, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.203, + 0.475, + 0.218 + ], + "angle": 0, + "content": "Figure 2. The overall architecture of our proposed TEAFormer." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.226, + 0.182, + 0.241 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.251, + 0.484, + 0.327 + ], + "angle": 0, + "content": "In this section, we conform to Theorems 3.1 and 3.2 to devise a translation equivariance restoration transformer, termed Translation Equivariance trAnsFormer (TEAFormer). We first present the overall architecture of TEAFormer and then introduce its components in detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.338, + 0.288, + 0.352 + ], + "angle": 0, + "content": "4.1. Overall Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.679 + ], + "angle": 0, + "content": "The overall architecture of the proposed TEAFormer is illustrated in Figure 2 with image super-resolution (SR) presented as a demonstration. Given an input image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), where \\( H \\times W \\) denotes the resolution, TEAFormer first applies a convolution layer to obtain a shallow feature \\( F_{0} \\in \\mathbb{R}^{H \\times W \\times D} \\), where \\( D \\) is the embedding dimension of the network. Next, following classical restoration networks [31], TEAFormer adopts a residual-in-residual structure to construct a deep feature extraction module comprised of \\( N_{g} \\) Translation Equivariance Groups (TEGs). Each TEG consists of \\( N_{b} \\) Translation Equivariance Blocks (TEBs) and a convolution layer. Each TEB, in turn, is composed of one Translation Equivariance Attention (TEA) and one feed-forward network. Subsequently, a convolutional layer is employed to extract the deep feature \\( F_{1} \\in \\mathbb{R}^{H \\times W \\times D} \\) from the feature extracted by TEGs. After extraction through the deep feature extraction module, we use the restore module to generate the high-quality image \\( \\hat{I} \\) from the feature \\( F_{r} = F_{0} + F_{1} \\). As for other image restoration tasks that do not involve changes in resolution, we build our TEAFormer following Restormer [63]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.688, + 0.398, + 0.705 + ], + "angle": 0, + "content": "4.2. Adaptive Translation Equivariance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.756 + ], + "angle": 0, + "content": "Adaptive Translation Equivalence is integrated through the Adaptive Sliding Key-Value Self Attention (ASkvSA) and Downsampled Self Attention (DSA) as shown in Figure 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.757, + 0.484, + 0.847 + ], + "angle": 0, + "content": "Key-value indexing in self-attention. The nature of Self Attention (SA) without position bias lies in the indexing of key-value pairs, followed by feature re-weighting. Given a flattened input \\( X \\in \\mathbb{R}^{N \\times D} \\), where \\( N \\) represents the number of tokens and \\( D \\) represents the embedding dimension, self-attention can be formulated as following:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.855, + 0.483, + 0.888 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} (X) = \\operatorname {S o f t m a x} \\left(\\frac {X W _ {q} \\left(X W _ {k}\\right) ^ {\\top}}{\\sqrt {D}}\\right) X W _ {v}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(W_{q}, W_{k}\\), and \\(W_{v}\\) denote the weights of the linear" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.091, + 0.905, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.317, + 0.906, + 0.359 + ], + "angle": 0, + "content": "Figure 3. The inner structure of TEA with two components. (a) Adaptive Sliding key-value Self Attention (ASkvSA) and (b) Downsampled Self Attention (DSA)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.906, + 0.397 + ], + "angle": 0, + "content": "projections utilized to compute the query \\((Q)\\), key \\((K)\\), and value \\((V)\\) from input \\(X\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.906, + 0.489 + ], + "angle": 0, + "content": "In Eq. 1, the softmax similarity function is utilized for the generation of attention maps between a query and a key, while the dot product is employed for re-weighting between the attention map and a value. Given that subscripting a matrix with \\( i \\) returns the \\( i \\)-th row as a vector, an alternative equation of self-attention can be written as follows." + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.503, + 0.905, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.555, + 0.753, + 0.579 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(Q_i, K_j) = \\exp \\left(\\frac{Q_i^\\top K_j}{\\sqrt{D}}\\right)\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.579, + 0.906, + 0.609 + ], + "angle": 0, + "content": "Eq. 2 illustrates that self-attention indexes key-value pairs \\((K_{j}\\) and \\(V_{j}, j = 1, \\dots, N)\\) for each query \\((Q_{i})\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.609, + 0.906, + 0.685 + ], + "angle": 0, + "content": "Sliding key-value Self Attention. In accordance with Theorem 3.2, we demonstrate that sliding window attention constitutes a generalized form of self-attention, which can be designated as Sliding Key-Value Self-Attention (SkvSA). SkvSA is formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.693, + 0.905, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S k v S A} _ {i} (X) = \\frac {\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.738, + 0.906, + 0.799 + ], + "angle": 0, + "content": "where \\( b_{1} \\) and \\( b_{2} \\) are the sliding boundary on the left and right when sliding on the one-dimensional vector \\( X \\). In our SkvSA, \\( b_{1} = b_{2} = \\frac{w\\cdot s}{2} \\), where \\( w \\) is the size of the sliding window and \\( s \\) is the stride size during sliding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.799, + 0.906, + 0.875 + ], + "angle": 0, + "content": "Eq. 3 says that, for the \\(i\\)-th query \\(Q_{i}\\), SkvSA systematically extracts the key-value pairs from a designated sliding window, extending from the index \\(i - \\frac{w \\cdot s}{2}\\) to the index \\(i + \\frac{w \\cdot s}{2}\\) with stride \\(s^2\\). Subsequently, SkvSA calculates the attention map and re-weights the feature in place." + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.905, + 0.901 + ], + "angle": 0, + "content": "2We set \\(w\\cdot s\\leq N\\) in experiments thus enabling SkvSA to slide on \\(X\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16049" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "Boundary processing. Note that, in scenarios where \\( i - \\frac{w \\cdot s}{2} < 0 \\) or \\( i + \\frac{w \\cdot s}{2} > N \\) occur, the slide indices of the key-value pairs \\( j \\) will exceed the boundary of the input \\( X \\), leading to an indexing failure. To ameliorate the issue associated with slide indexing at sequence's boundaries, we propose a boundary blocking indexing method. Specifically, for queries located at the boundary of the sequence, the sliding indices of key-value pairs are \"blocked\" and converted into indexing within a predefined blocking window. In this study, the sizes of the predefined blocking window and the sliding window are identical. Accordingly, the sliding boundaries \\( b \\) can be appropriately revised as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.275, + 0.484, + 0.356 + ], + "angle": 0, + "content": "\\[\nb = \\left(b _ {1}, b _ {2}\\right) = \\left\\{ \\begin{array}{l l} \\left(0, \\frac {w \\cdot s}{2}\\right), & i - \\frac {w \\cdot s}{2} < 0, \\\\ \\left(\\frac {w \\cdot s}{2}, N\\right), & i + \\frac {w \\cdot s}{2} > N, \\\\ \\left(\\frac {w \\cdot s}{2}, \\frac {w \\cdot s}{2}\\right), & \\text {o t h e r s .} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.356, + 0.484, + 0.415 + ], + "angle": 0, + "content": "The SkvSA and its boundary processing are capable of extracting and processing information via a fixed sliding window, adhering to the translation equivariance property as established in Theorem 3.2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.552 + ], + "angle": 0, + "content": "Adaptive Sliding key-value Self Attention. However, the fixed window size constrains the receptive field, leading to the entrapment of restoration transformers within rigidity [48]. To mitigate this rigidity, it is anticipated that the SkvSA will ascertain the optimal window size and position specific to each query \\( Q_{i} \\). To achieve this endeavor, we devise the Adaptive Sliding key-value Self Attention (ASkvSA), which adaptively determines key-value indices and shuffles them within the fixed sliding window of \\( Q_{i} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.72 + ], + "angle": 0, + "content": "As shown in Figure 3 (a), with regard to Key \\(K\\) and Value \\(V\\), ASkvSA initially reshapes \\(K\\) and \\(V\\) from one-dimensional (1d) vectors into two-dimensional (2d) matrices. Subsequently, it applies a 2d depth-wise convolution with kernel size \\(k\\) to generate their adaptive indices \\(\\mathcal{F} \\in \\mathbb{R}^{H \\times W \\times 2}\\). \\(\\mathcal{F}[h, w]\\) is a vector of length 2 in the \\(h-\\)th row and \\(w-\\)th column of \\(\\mathcal{F}\\), which can describe a coordinate. Following this, ASkvSA performs key-value indexing based on established indices \\(\\mathcal{F}_K\\) and \\(\\mathcal{F}_V\\) by shuffling the pixel located at the coordinate \\(\\mathcal{F}[h, w]\\) of \\(K\\) and \\(V\\) to the pixel at the coordinate \\((h, w)\\) of \\(K\\) and \\(V\\)." + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.728, + 0.483, + 0.766 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A S k v S A} (X) = \\frac {\\sum_ {j = \\mathcal {F} (K) , l = \\mathcal {F} (V)} \\sin \\left(Q _ {i} , K _ {j}\\right) V _ {l}}{\\sum_ {j = \\mathcal {F} (K)} \\sin \\left(Q _ {i} , K _ {j}\\right)}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Adaptive slide indexing does not compromise the translation equivariance of ASkvSA. ASkvSA stacks convolutions with SkvSA in serial to generate adaptive indices of key-value pairs. Given that convolution satisfies the property of translation equivariance, when \\( K_{h,w} \\) is translated to \\( K_{h + \\delta_h, w + \\delta_w} \\), key's adaptive indices \\( \\mathcal{F}_K \\) will shuffle the pixel located at \\( \\mathcal{F}[h + \\delta_h, w + \\delta_w] \\) of \\( K \\) to the pixel at \\( (h + \\delta_h, w + \\delta_w) \\) of \\( K \\). According to Theorem 3.3, ASkvSA satisfies the translation equivariance property." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.212 + ], + "angle": 0, + "content": "Adaptively aggregated with global key-value pairs. Although ASkvSA can identify appropriate key-value pairs beyond a fixed sliding window, some remote but valid pixels still escape ASkvSA's indexing. Consequently, we propose downsampled self-attention (DSA), which adaptively combines global key value indexes through downsampling, providing the network with a coarse, yet efficient global indexing of key-value pairs. DSA can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.214, + 0.907, + 0.256 + ], + "angle": 0, + "content": "\\[\n\\mathrm {D S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right) V _ {j} ^ {\\prime}}{\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.257, + 0.907, + 0.331 + ], + "angle": 0, + "content": "where \\( K' \\) and \\( V' \\) represents the downsampled key-value pair calculated in ASkvSA, and \\( N_d \\) denotes the number of tokens associated with the downsampled key-value pair. We employ average pooling as the downsampling operator due to its efficiency and coarse fulfillment of TE." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.332, + 0.906, + 0.363 + ], + "angle": 0, + "content": "Finally, according to Figure 3, the ASkvSA and DSA outputs are adaptively aggregated to obtain our TEA." + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.374, + 0.906, + 0.436 + ], + "angle": 0, + "content": "\\[\n\\mathrm {T E A} _ {i} (X) = \\alpha_ {s} \\operatorname {A S k v S A} (X) + \\alpha_ {d} \\operatorname {D S A} (X), \\quad (7) \\text {w h e r e} \\alpha_ {s} \\text {a n d} \\alpha_ {d} \\text {a r e l a r n a b l e p a r a m e t e r s . T E A s t a c k s D S A w i t h A S k v S A i n p a r a l l e l a r . A c c o r d i n g t o T h e o r e m 3 . 3 , T E A s a t s i s f i e s t h e t r a n s l a t i o n e q u i v a r i a n c e p r o p e r t y .}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.445, + 0.886, + 0.461 + ], + "angle": 0, + "content": "4.3. Computational Cost and Hyper-Parameters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.528 + ], + "angle": 0, + "content": "We demonstrate that TEA can maintain a computational cost that is similar to that of window attention [34] under well-designed hyperparameter settings. We first list four integral hyperparameters of TEA:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.531, + 0.793, + 0.545 + ], + "angle": 0, + "content": "- \\( w \\) : the sliding window size of ASkvSA;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.773, + 0.56 + ], + "angle": 0, + "content": "- \\(s\\): the sliding stride size of ASkvSA;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.561, + 0.852, + 0.575 + ], + "angle": 0, + "content": "- \\( k \\) : the kernel size of adaptive indexes generation;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.576, + 0.758, + 0.59 + ], + "angle": 0, + "content": "- \\(N_{d}\\): the number of tokens in DSA." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.531, + 0.852, + 0.59 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.593, + 0.906, + 0.762 + ], + "angle": 0, + "content": "According to Section 4.2, TEA uses \\(3ND^{2}\\) FLOPs for the Query-Key-Value linear projections, \\(2NDk^{2}\\) FLOPs for the depth-wise convolution in adaptive indices generation, \\(Nw^{2}D\\) FLOPs for attention map computation between query \\(Q \\in \\mathbb{R}^{N \\times D}\\) and indexed key \\(K \\in \\mathbb{R}^{w^{2} \\times D}\\), \\(Nw^{2}D\\) FLOPs for feature re-weighting between attention map and indexed value \\(V \\in \\mathbb{R}^{w^{2} \\times D}\\), and \\(2NN_{d}D\\) FLOPs for DSA. The total FLOPs of TEA is \\(3ND^{2} + 2NDk^{2} + 2Nw^{2}D + 2NN_{d}D\\). With the hyper-parameters held constant, the FLOPs of TEA exhibit linear growth as a function of \\(N\\), thereby implying that TEA is \\(\\mathcal{O}(N)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Consequently, when the hyper-parameters of TEA are adjusted carefully, its computational cost becomes comparable to that of window attention. In our TEAFormer, the hyper-parameter \\( w \\) has been set to 15, \\( k \\) to 3, and \\( N_{d} \\) to 16, thereby modulating the computational complexity of TEA to be even slightly lower than that of window attention with window size 16." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.875, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[45] proved that downsampling can meet strict TE through learnable polyphase (LP). We compare these downsamplers in Sec. 5.6" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16050" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.14, + 0.09, + 0.857, + 0.21 + ], + "angle": 0, + "content": "
Modelwindow sizeTranslation equivarianceParamsFLOPsLatencyPerformanceNTK [25]SRGA [33]
SkvSAASkvSADSA(M)(T)(ms)Urban100 [24] PSNR↑ / SSIM↑Condition ↓Value ↓
SwinIR [31]811.900.462130.027.45 / 0.82541746.493.655
SwinIR-Large1621.420.897214.527.94 / 0.83621554.653.610
TEAFformer15✓21.860.688230.128.31 / 0.8444243.753.206
TEAFformer15✓21.980.757284.428.47 / 0.8457283.993.298
TEAFformer15✓✓21.860.966340.928.49 / 0.8470203.063.261
TEAFformer15✓✓21.981.035386.728.67 / 0.8489236.783.275
" + }, + { + "type": "table_caption", + "bbox": [ + 0.185, + 0.211, + 0.812, + 0.224 + ], + "angle": 0, + "content": "Table 2. Ablation study on TEA. Our TEA enhances performance, convergence speed, and generalization." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.238, + 0.921, + 0.553 + ], + "angle": 0, + "content": "
MethodScaleParams (M)FLOPs (T)Latency (ms)Set5 [5]Set14 [64]B100 [37]Urban100 [24]Manga109 [40]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
EDSR [32]40.76.006147.138.110.960233.920.919532.320.901332.930.935139.100.9773
RCAN [71]15.42.259155.538.270.961434.120.921632.410.902733.340.938439.440.9786
SwinIR-L [31]0.91156.4188.638.140.961133.860.920632.310.901232.760.934039.120.9783
SwinIR [31]11.81.848514.438.420.962334.460.925032.530.904133.810.942739.920.9797
EDT [30]11.51.965101038.450.962434.570.925832.520.904133.800.942539.930.9800
SRFormer [73]×210.31.790787.238.510.962734.440.925332.570.904634.090.944940.070.9802
HAT [9]20.63.662901.938.630.963034.860.927432.620.905334.450.946640.260.9809
GRL-B* [61]20.14.422262138.480.962734.640.926532.550.904533.970.943740.060.9804
DAT [10]14.72.155121538.580.962934.810.927232.610.905134.370.945840.330.9807
IPG [48]16.84.732270338.610.963234.730.927032.600.905234.480.946440.240.9810
TEAFormer-L (Ours)0.830.237254.038.270.861834.260.924032.430.902933.510.939639.560.9792
TEAFormer (Ours)21.84.133149338.640.963135.120.928432.720.906435.230.950940.320.9807
EDSR [32]43.11.85349.1332.460.896828.800.787627.710.742026.640.803331.020.9148
RCAN [71]15.60.58746.1532.630.900228.870.788927.770.743626.820.808731.220.9173
SwinIR-L [31]0.880.03948.7132.440.897628.770.785827.690.740626.470.798030.920.9151
SwinIR [31]11.90.462130.032.920.904429.090.795027.920.748927.450.825432.030.9260
EDT [30]11.60.514251.332.820.903129.090.793927.910.748327.460.824632.030.9254
SRFormer [73]×410.40.470189.232.930.904129.080.795327.940.750227.680.831132.440.9271
HAT [9]20.80.938227.733.040.905629.230.797328.000.751727.970.836832.480.9292
GRL-B* [61]20.21.129761.832.900.903929.140.795627.960.749727.530.827632.190.9266
DAT [10]14.80.561296.833.080.905529.230.797328.000.751527.870.834332.510.9291
IPG [48]17.01.206646.733.150.906229.240.797327.990.751928.130.839232.530.9300
TEAFormer-L (Ours)0.830.05969.6432.630.900228.960.789927.820.744227.040.810531.560.9197
TEAFormer (Ours)22.01.035386.733.230.906629.460.800928.110.754428.670.848932.990.9323
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.555, + 0.908, + 0.597 + ], + "angle": 0, + "content": "Table 3. Classical image SR results. Performance and complexity are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two scores in each column are highlighted in red and blue. Methods with “*” are replicated with standard training setting, following IPG [48]. The “-L” indicates that the given method is a lightweight version." + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.61, + 0.373, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.724, + 0.32, + 0.736 + ], + "angle": 0, + "content": "BSD100 \\((4\\times)\\) :img_86016" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.61, + 0.507, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.661, + 0.466, + 0.672 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.61, + 0.636, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.538, + 0.661, + 0.605, + 0.673 + ], + "angle": 0, + "content": "RCAN [71]" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.61, + 0.764, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.665, + 0.661, + 0.736, + 0.673 + ], + "angle": 0, + "content": "SwinIR [31]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.611, + 0.894, + 0.659 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.661, + 0.854, + 0.673 + ], + "angle": 0, + "content": "HAT [9]" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.737, + 0.374, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.85, + 0.316, + 0.863 + ], + "angle": 0, + "content": "Urban100 \\((4\\times)\\) : img_033" + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.674, + 0.507, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.724, + 0.455, + 0.736 + ], + "angle": 0, + "content": "HQ" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.674, + 0.636, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.529, + 0.724, + 0.615, + 0.736 + ], + "angle": 0, + "content": "SRFormer [73]" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.674, + 0.764, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.724, + 0.726, + 0.736 + ], + "angle": 0, + "content": "IPG [48]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.674, + 0.894, + 0.722 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.724, + 0.89, + 0.736 + ], + "angle": 0, + "content": "TEAFormer (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.737, + 0.508, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.42, + 0.788, + 0.466, + 0.799 + ], + "angle": 0, + "content": "Bicubic" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.737, + 0.636, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.539, + 0.788, + 0.605, + 0.799 + ], + "angle": 0, + "content": "RCAN [71]" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.737, + 0.764, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.666, + 0.788, + 0.736, + 0.799 + ], + "angle": 0, + "content": "SwinIR [31]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.737, + 0.894, + 0.785 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.806, + 0.788, + 0.854, + 0.8 + ], + "angle": 0, + "content": "HAT [9]" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.8, + 0.508, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.851, + 0.455, + 0.862 + ], + "angle": 0, + "content": "HQ" + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.8, + 0.636, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.851, + 0.615, + 0.863 + ], + "angle": 0, + "content": "SRFormer [73]" + }, + { + "type": "image", + "bbox": [ + 0.638, + 0.8, + 0.764, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.851, + 0.726, + 0.863 + ], + "angle": 0, + "content": "IPG [48]" + }, + { + "type": "image", + "bbox": [ + 0.766, + 0.8, + 0.894, + 0.848 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.851, + 0.889, + 0.863 + ], + "angle": 0, + "content": "TEAFormer (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.25, + 0.864, + 0.744, + 0.877 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison with recent SOTA methods on the image SR task." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16051" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.101, + 0.082, + 0.894, + 0.297 + ], + "angle": 0, + "content": "
MethodParamsIndoor ScenesOutdoor ScenesCombined
(M)PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓
DPDNetS [1]32.326.540.8160.0310.23922.250.6820.0560.31324.340.7470.0440.277
KPACs [47]2.0627.970.8520.0260.18222.620.7010.0530.26925.220.7740.0400.227
IFANs [27]10.528.110.8610.0260.17922.760.7200.0520.25425.370.7890.0390.217
RestormerS [63]26.128.870.8820.0250.14523.240.7430.0500.20925.980.8110.0380.178
SFNetS [12]13.229.160.8780.0230.16823.450.7470.0490.24426.230.8110.0370.207
GRLS-B [61]19.929.060.8860.0240.13923.450.7610.0490.19626.180.8220.0370.168
IRNeXTS [13]13.229.220.8790.0240.16723.530.7520.0490.24426.300.8140.0370.206
TEAFormerS (Ours)15.429.500.8920.0230.15023.550.7670.0500.21126.450.8280.0370.181
DPDNetD [1]32.327.480.8490.0290.18922.900.7260.0520.25525.130.7860.0410.223
RDPD_D [2]8.2028.100.8430.0270.21022.820.7040.0530.29825.390.7720.0400.255
IFAN_D [27]10.528.660.8680.0250.17223.460.7430.0490.24025.990.8040.0370.207
Restormer_D [63]26.129.480.8950.0230.13423.970.7730.0470.17526.660.8330.0350.155
Uformer_D [54]20.628.230.8600.0260.19923.100.7280.0510.28525.650.7950.0390.243
GRL_D-B [61]19.929.830.9030.0220.11424.390.7950.0450.15027.040.8470.0340.133
TEAFormer_D (Ours)15.430.160.9050.0220.12824.450.7950.0450.17027.230.8490.0340.150
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.298, + 0.905, + 0.327 + ], + "angle": 0, + "content": "Table 4. Defocus deblurring comparisons on the DPDD [1] testset (containing 37 indoor and 39 outdoor scenes). S: single-image defocus deblurring. D: dual-pixel defocus deblurring. The best two results are highlighted in red and blue, respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.333, + 0.224, + 0.35 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.359, + 0.241, + 0.375 + ], + "angle": 0, + "content": "5.1. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.484, + 0.684 + ], + "angle": 0, + "content": "We analyze the influence exerted by the TEA on both performance, computational complexity, generalization, and training convergence in Table 2. In this experiment, we disassemble TEA into several plug-and-play modules and gradually add them into the SwinIR baseline restoration network. To ensure comparable parameters and FLOPs among the models being evaluated, we adjusted the window size of SwinIR to 16 and increased the number of parameters to 21 M, which is termed SwinIR-Large. Following the incorporation of our SkvSA, there is a marked improvement in the convergence rate and generalization abilities. Subsequent to the integration of our ASkvSA and DSA, there is a progressive enhancement in the model's performance, without detrimental effects on convergence and generalization. The results of this experiment demonstrate the efficacy of each component within the TEA. Furthermore, it illustrates that TEA resolves the dilemma between self-attention and sliding window attention, thereby attaining high performance by the global receptive field and reducing computational complexity simultaneously." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.696, + 0.204, + 0.712 + ], + "angle": 0, + "content": "5.2. Image SR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.483, + 0.809 + ], + "angle": 0, + "content": "Training settings. We used the same training hyperparameters and evaluation methods as IPG [48]. All of our models were trained from scratch on the DF2K [3, 32] datasets with \\(64 \\times 64\\) resolution. The optimizer was Adam [26], the learning rate was \\(2 \\times 10^{-4}\\) while using the cosine scheduler. We used \\(\\mathrm{L}_1\\) loss as a criterion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Results. We first compare TEAFormer with various classical image SR methods, including EDSR [32], RCAN [71], SwinIR [31], EDT [30], HAT [9], SRFormer [73], GRL [61], DAT [10], and IPG [48]. We compare them on synthetic benchmarks [5, 24, 37, 40, 64] generated by bicubic downsampling. The FLOPs and latency reported were" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.335, + 0.905, + 0.349 + ], + "angle": 0, + "content": "calculated with one NVIDIA A800 GPU on 768 resolution." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.351, + 0.907, + 0.547 + ], + "angle": 0, + "content": "The quantitative comparison of the methods for classical image SR is shown in Table 3. The exceptional performance of TEAFormer is distinctly evident as it achieves the highest results across nearly all five datasets for various scale factors. Incorporating the attention module with translation equivalence introduces a more suitable and productive inductive bias to the restoration model, thereby enhancing its performance. For the \\(\\times 4\\) SR, our TEAFormer achieves a \\(28.67\\mathrm{dB}\\) PSNR score on the Urban100 dataset, which is \\(0.7\\mathrm{dB}\\) higher than HAT [9] while both has \\(20\\mathrm{M}\\) parameters. In comparison to the preceding image SR method IPG [48], TEAFormer demonstrates a gain of \\(0.54\\mathrm{dB}\\) with comparable FLOPs and a \\(2\\times\\) faster inference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.685 + ], + "angle": 0, + "content": "We further conduct TEAFormer-Light, which only comprises 829k parameters. In comparison to the CNN method HAN [42], TEAFormer-Light has been demonstrated to achieve superior performance with only its \\(1.3\\%\\) parameters. In comparison to other lightweight transformer methods, TEAFormer also demonstrates an absolute performance advantage, as evidenced by the fact that TEAFormer-Light achieves a higher PSNR than SwinIR-L on the Urban100 [24] dataset by 0.57 dB." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.696, + 0.755, + 0.712 + ], + "angle": 0, + "content": "5.3. Image Defocus Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.719, + 0.905, + 0.794 + ], + "angle": 0, + "content": "Training settings. We employed the same training hyperparameters and evaluation as Restormer [63] in image deblurring tasks. Our TEAFormer were trained from scratch using progressive learning [63]. During the training phase, the optimizer adopted was AdamW [35] with \\( \\mathrm{L}_1 \\) loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Results. We compare our TEAFormer with single image defocus deblurring methods [1, 13, 14, 27, 47, 61, 63] and dual-pixel image defocus deblurring methods [1, 2, 27, 54, 61, 63] on DPDD [1] dataset in Table 4. Our TEAFormer significantly outperforms state-of-the-art schemes for the single-image and dual-pixel defocus deblurring tasks on PSNR and SSIM metrics. In particular in the category" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16052" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.373, + 0.214 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.379, + 0.089, + 0.506, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.508, + 0.089, + 0.635, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.637, + 0.089, + 0.762, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.765, + 0.089, + 0.893, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.215, + 0.866, + 0.228 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with recent SOTA methods on the image defocus blur task. Please zoom in for better view." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.233, + 0.907, + 0.369 + ], + "angle": 0, + "content": "
MethodParamsColorGrayscale
(M)CBSD68 [39]Kodak24 [17]McMaster [69]Urban100 [24]Set12 [67]BSD68 [39]Urban100 [24]
σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50
DnCNN [67]0.5633.9031.2427.9534.6032.1428.9533.4531.5228.6232.9830.8127.5932.8630.4427.18
DRUNet [68]32.634.3031.6928.5135.3132.8929.8635.4033.1430.0834.8132.6029.6133.2530.9427.90
EDT-B [30]11.534.3931.7628.5635.3732.9429.8735.6133.3430.2535.2233.0730.16---
SwinIR* [31]11.734.3931.7828.5535.3332.8829.7735.5933.2030.1935.1132.8929.7733.3431.0027.88
Restormer [63]26.134.3931.7828.5935.4433.0230.0035.5533.3130.2935.0632.9130.0233.3531.0428.01
ART* [66]16.234.4031.7928.5935.4733.0430.0035.6133.3330.2735.1232.9529.9933.3531.0528.01
GRL-B* [61]19.834.4031.7828.5835.4233.0330.0235.6333.3630.2635.2233.0130.2433.3431.0227.99
TEAFormer (Ours)15.434.4031.8028.6235.5033.0830.0535.6833.4230.3835.2533.1830.3733.3731.0628.04
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Table 5. Color and grayscale image Gaussian denoising results in terms of PSNR \\( \\uparrow \\) . Model complexity and prediction accuracy are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two results are highlighted in red and blue, respectively. Methods with \" \\( * \\) \" are replicated with standard training setting, which learns a single model to handle various noise levels." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.49 + ], + "angle": 0, + "content": "of single-pixel indoor scene, TEAFormer yields 0.44 dB improvements over the GRL previous best transformer method [61]. Figure 5 illustrates that our TEAFormer is more effective in removing spatially varying defocus blur than other restoration transformers." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.496, + 0.258, + 0.513 + ], + "angle": 0, + "content": "5.4. Image Denoising" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.519, + 0.483, + 0.549 + ], + "angle": 0, + "content": "Training settings. The image denoising and defocus deblurring tasks used the same hyper-parameters for training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.483, + 0.747 + ], + "angle": 0, + "content": "Gaussian Denoising Results. We perform denoising experiments on synthetic benchmark datasets generated with additive white Gaussian noise (Set12 [67], BSD68 [38], Urban100 [24], Kodak24 [17] and McMaster [69]). Following DRUNet [68], we learn a single model to handle various noise levels, including 15, 25, and 50. Our TEAFourner achieves state-of-the-art performance under both experimental settings on different datasets and noise levels. Specifically, for the challenging noise level 50 on color Urban100 [24] dataset, TEAFourner achieves 0.13 dB gain over the previous best restoration transformer GRL [61], as shown in Table 5. Similar performance gains can be observed for the Gaussian gray denoising." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.758, + 0.351, + 0.773 + ], + "angle": 0, + "content": "5.5. All-in-one Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Training settings. Following DCPT [23], a 5D all-in-one image restoration and enhancement task was conducted with the objective of training the network to eliminate five distinct types of degradation: haze, rain, Gaussian noise, motion blur, and low light. The models were trained from scratch on datasets with \\(128 \\times 128\\) resolution without progressive learning. During training, the optimizer was set as AdamW, the learning rate was set as \\(3 \\times 10^{-4}\\) while using" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.858, + 0.429 + ], + "angle": 0, + "content": "the cosine scheduler. We used \\( L_{1} \\) loss as a criterion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.581 + ], + "angle": 0, + "content": "Results. As illustrated in Table 6, TEAFormer surpasses the most recently proposed methods specifically designed for comprehensive all-in-one restoration tasks (IDR [65], PromptIR [43], and InstructIR [11]). The average PSNR of TEAFormer in all five tasks reached 31.22 dB, thereby exceeding the current state-of-the-art 5D all-in-one restoration method, InstructIR [11], by 1.67 dB. TEAFormer consistently outperforms prior PSNR in every indicator of each task, signifying its exceptional proficiency in image restoration under complex multi-task conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.705 + ], + "angle": 0, + "content": "Mixed degradation. As shown in Table 7, TEAFormer surpasses the most recently proposed methods under mixed degradation scenarios, specifically designed for universal restoration tasks [36, 72]). Within scenarios characterized by a combination of three degradation (low-light, haze, and rain), TEAFormer demonstrates superior performance over the current state-of-the-art universal restoration method, DiffUIR [72], by a margin of \\(0.42\\mathrm{dB}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.711, + 0.633, + 0.725 + ], + "angle": 0, + "content": "5.6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Hyperparameters analysis. The performance under various hyperparameters is reported in Table 8 \\((w)\\), Table 9 \\((s)\\), Table 10 \\((k)\\) and Table 11 \\((N_d)\\). Our settings \\((w = 15, s = 4, k = 3, N_d = 16)\\) optimally balanced performance and computation cost. Note that changing some of the hyperparameters does not affect the computational complexity as discussed in Section 4.3. For \\(s\\) in Table 9, parameters and FLOPs are omitted because \\(s\\) does not affect them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Compare different downsampler in DSA. The incorporation of DSA into our TEA substantially enhances the performance of the model without adding undue complexity." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16053" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.129, + 0.089, + 0.867, + 0.236 + ], + "angle": 0, + "content": "
MethodParamsDehazing on SOTS [28]Deraining on Rain100L [60]Denoising on BSD68 [38]Deblurring on GoPro [41]Low-Light on LOL [55]Average
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.04 / 0.88432.98 / 0.95130.91 / 0.88224.35 / 0.78118.18 / 0.73525.49 / 0.846
IDR [65]15.325.24 / 0.94335.63 / 0.96531.60 / 0.88727.87 / 0.84621.34 / 0.82628.34 / 0.893
PromptIR [43]35.625.20 / 0.93135.94 / 0.96431.17 / 0.88227.32 / 0.84220.94 / 0.79928.11 / 0.883
InstructIR [11]35.127.10 / 0.95636.84 / 0.97331.40 / 0.88729.40 / 0.88623.00 / 0.83629.55 / 0.907
SwinIR [31]11.821.50 / 0.89130.78 / 0.92330.59 / 0.86824.52 / 0.77317.81 / 0.72325.04 / 0.835
NAFNet [7]67.925.23 / 0.93935.56 / 0.96731.02 / 0.88326.53 / 0.80820.49 / 0.80927.76 / 0.881
Restormer [63]26.124.09 / 0.92734.81 / 0.96231.49 / 0.88427.22 / 0.82920.41 / 0.80627.60 / 0.881
TEAFormer (Ours)15.431.57 / 0.98039.44 / 0.98631.54 / 0.89030.53 / 0.90823.06 / 0.85631.22 / 0.924
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.237, + 0.907, + 0.28 + ], + "angle": 0, + "content": "Table 6. 5D All-in-one image restoration results. Dash lines separate the results of methods designed specifically for all-in-one image restoration task and methods that only focus on architecture enhancement. The best two results are highlighted in red and blue, respectively. TEAFormer sets the new state-of-the-art performance in each task." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.282, + 0.22, + 0.343 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.345, + 0.171, + 0.354 + ], + "angle": 0, + "content": "LQ" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.282, + 0.335, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.344, + 0.316, + 0.354 + ], + "angle": 0, + "content": "InstrucrIR [11]" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.282, + 0.453, + 0.342 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.356, + 0.344, + 0.436, + 0.354 + ], + "angle": 0, + "content": "OneRestore [19]" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.355, + 0.22, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.153, + 0.416, + 0.171, + 0.426 + ], + "angle": 0, + "content": "HQ" + }, + { + "type": "image", + "bbox": [ + 0.222, + 0.355, + 0.335, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.246, + 0.416, + 0.311, + 0.426 + ], + "angle": 0, + "content": "DACLIP [36]" + }, + { + "type": "image", + "bbox": [ + 0.337, + 0.355, + 0.453, + 0.414 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.416, + 0.446, + 0.426 + ], + "angle": 0, + "content": "TEAFormer (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.429, + 0.468, + 0.456 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on low-light+haze+rain mixed degradation. TEAFormer can restore the correct lightness." + }, + { + "type": "table", + "bbox": [ + 0.472, + 0.288, + 0.902, + 0.436 + ], + "angle": 0, + "content": "
MethodParams low-light+haze+rain low-light+haze+snow
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.80 / 0.70822.23 / 0.725
TransWeather [50]31.021.55 / 0.67821.01 / 0.655
PromptIR [43]35.623.74 / 0.75223.33 / 0.747
DACLIP [36]17422.96 / 0.71223.32 / 0.751
DiffUIR-L [72]36.325.46 / 0.77925.89 / 0.788
OneRestore [19]5.9825.18 / 0.79525.28 / 0.797
InstructIR [11]35.124.84 / 0.77724.32 / 0.760
TEAFormer (Ours)15.425.88 / 0.79626.05 / 0.800
" + }, + { + "type": "table_caption", + "bbox": [ + 0.491, + 0.438, + 0.882, + 0.451 + ], + "angle": 0, + "content": "Table 7. Mixed degraded image restoration results on CDD [19]." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.457, + 0.255, + 0.541 + ], + "angle": 0, + "content": "
wFLOPs (T)Urban100 [24] PSNR ↑/SSIM ↑
70.82728.35 / 0.8441
151.03528.67 / 0.8489
311.86628.73 / 0.8493
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.543, + 0.25, + 0.571 + ], + "angle": 0, + "content": "Table 8. Ablations of sliding window size \\( w \\) ." + }, + { + "type": "table", + "bbox": [ + 0.257, + 0.457, + 0.37, + 0.541 + ], + "angle": 0, + "content": "
sUrban100 [24]
PSNR ↑ / SSIM ↑
128.43 / 0.8431
228.51 / 0.8453
428.67 / 0.8489
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.254, + 0.543, + 0.364, + 0.583 + ], + "angle": 0, + "content": "Table 9. Ablations of sliding stride \\(s\\)" + }, + { + "type": "table", + "bbox": [ + 0.371, + 0.457, + 0.902, + 0.541 + ], + "angle": 0, + "content": "
kFLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Nd(T)FLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Down (%)↑S-conv (%)↑Urban100 [24] PSNR ↑ / SSIM ↑
None0.96628.49 / 0.8470None0.75728.47 / 0.8457MaxPool87.328.61 / 0.8477
11.02828.53 / 0.8471161.03528.67 / 0.8489LPD91.728.64 / 0.8493
31.03528.67 / 0.8489N18.3628.80 / 0.8503AvgPool91.228.67 / 0.8489
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.368, + 0.543, + 0.704, + 0.598 + ], + "angle": 0, + "content": "Table 10. Ablations of kernel size \\( k \\) for adaptive in- in DSA. \"None\" stands for indices. \"None\" stands for no no DSA used. adaptive indices used." + }, + { + "type": "table_footnote", + "bbox": [ + 0.708, + 0.543, + 0.89, + 0.598 + ], + "angle": 0, + "content": "Table 12. Ablations of down-samplers in DSA. The higher the S-Conv, the more strictly the model meets TE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.484, + 0.676 + ], + "angle": 0, + "content": "However, the average pooling employed in DSA does not strictly adhere to the property of TE. Using a comparative ablation of several downsampling modules in Table 12, we experimentally demonstrate that the use of average pooling does not compromise the model's performance or its TE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.696, + 0.484, + 0.907 + ], + "angle": 0, + "content": "Does TEAFormer converge faster and generalize better in practice? Table 1 and Table 2 have illustrated that TEAFormer converges faster and generalizes better than other networks theoretically due to the incorporation of translation equivariance. Figure 7 (left) further demonstrates that TEAFormer reaches a lower loss faster than other restoration transformers [9, 31] in practice. Following MaskDenoise [6], we compare the statistics of the last layer feature in SwinIR [31] and TEAFormer under two different real-world degradations on the PIES [33] dataset (out of the training dataset) in Figure 7 (right). The results show that TEAFormer can produce a more consistent output when processing LQ images with different degradations. Therefore, TEAFormer generalizes better in practice." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.601, + 0.674, + 0.675 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.671, + 0.607, + 0.782, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.785, + 0.608, + 0.893, + 0.674 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.677, + 0.898, + 0.69 + ], + "angle": 0, + "content": "Figure 7. Comparison on convergence (left) and generalization (right)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.698, + 0.633, + 0.713 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we start by examining the translation equivariance, which serves as an inherent inductive bias in image restoration. We then define the translation equivariance and offer two strategies for incorporating it into restoration networks: (1) slide indexing and (2) component stacking. To address the fixed receptive field provided by these strategies, we proposed TEA and TEAFormer via adaptive key-value indexing. The efficacy of TEAFormer has been demonstrated in various image restoration tasks, with ablation studies confirming that this efficacy arises from the incorporation of TEA. In the future, we intend to explore more equivariances into restoration networks." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16054" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.228 + ], + "angle": 0, + "content": "This work was supported financially by the Natural Science Foundation of China (82371112, 623B2001), the Science Foundation of Peking University Cancer Hospital (JC202505), Natural Science Foundation of Beijing Municipality (Z210008) and the Clinical Medicine Plus X - Young Scholars Project of Peking University, the Fundamental Research Funds for the Central Universities (PKU2025PKULCXQ008)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.258, + 0.188, + 0.274 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.284, + 0.483, + 0.326 + ], + "angle": 0, + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.328, + 0.484, + 0.397 + ], + "angle": 0, + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.4, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[4] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.515, + 0.484, + 0.583 + ], + "angle": 0, + "content": "[5] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference. BMVA press, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.587, + 0.485, + 0.668 + ], + "angle": 0, + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.483, + 0.727 + ], + "angle": 0, + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision - ECCV 2022, pages 17-33, Cham, 2022. Springer Nature Switzerland. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.73, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[8] Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre-training or strong data augmentations. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.788, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[10] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In ICCV, 2023. 2, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.284, + 0.485, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[11] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[13] Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, and Alois Knoll. Irnext: Rethinking convolutional network design for image restoration. In Proceedings of the 40th International Conference on Machine Learning, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.906, + 0.346 + ], + "angle": 0, + "content": "[14] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[15] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.433, + 0.906, + 0.529 + ], + "angle": 0, + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[17] Rich Franzen. Kodak lossless true color image suite, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.561, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[18] Jonathan Gordon, Wessel P. Bruinsma, Andrew Y. K. Foong, James Requeima, Yann Dubois, and Richard E. Turner. Convolutional conditional neural processes. In International Conference on Learning Representations, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.618, + 0.906, + 0.672 + ], + "angle": 0, + "content": "[19] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In European Conference on Computer Vision, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185-6194, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.732, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[21] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[23] JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Universal image restoration pre-training via degradation classification. The Thirteenth International Conference on Learning Representations, 2025. 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16055" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[24] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 1, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[27] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.331, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[28] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.373, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[29] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[30] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1089–1097. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.565 + ], + "angle": 0, + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.568, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[33] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497-14513, 2023. 1, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[36] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for universal image restoration. In The Twelfth International Conference on Learning Representations, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423 vol.2, 2001. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[38] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.286 + ], + "angle": 0, + "content": "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools Appl., 76(20):21811-21838, 2017. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[41] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[42] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 191-207. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[43] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. NeurIPS, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[44] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self-attention in vision models. Advances in neural information processing systems, 32, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[45] Renan A Rojas-Gomez, Teck-Yian Lim, Alex Schwing, Minh Do, and Raymond A Yeh. Learnable polyphase sampling for shift invariant and equivariant convolutional networks. Advances in Neural Information Processing Systems, 35:35755-35768, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[46] Renan A Rojas-Gomez, Teck-Yian Lim, Minh N Do, and Raymond A Yeh. Making vision transformers truly shift-equivariant. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5568-5577, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proc. ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[48] Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang. Image processing gnn: Breaking rigidity in super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24108-24117, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16056" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[50] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[52] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[54] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In British Machine Vision Conference, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[56] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[57] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[58] Renjun Xu, Kaifan Yang, Ke Liu, and Fengxiang He. \\(e(2)\\)-equivariant vision transformer. In Uncertainty in Artificial Intelligence, pages 2356-2366. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[59] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vi-tae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. TPAMI, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[61] Xiaoyu Xiang Yawei Li, Yuchen Fan, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[64] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference, Avignon, France*, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[65] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[66] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[67] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[68] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.432, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[69] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.488, + 0.905, + 0.543 + ], + "angle": 0, + "content": "[70] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. arXiv preprint arXiv:2202.10108, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.545, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[71] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.588, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[72] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[73] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.729, + 0.905, + 0.797 + ], + "angle": 0, + "content": "[74] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.797 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16057" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_origin.pdf b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a3362ae762bb65b1c7f4173d1af5691749920bb2 --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/315e18c5-4a71-4399-bd7d-013d4c4b0332_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dfe23a93c798b06e747304aff898bd882bba61191c7e3703aaa781879c7b3eac +size 17759660 diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/full.md b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/full.md new file mode 100644 index 0000000000000000000000000000000000000000..1a8674509c0e4ab02b9baec25ca96b15f790ae08 --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/full.md @@ -0,0 +1,436 @@ +# Enhancing Image Restoration Transformer via Adaptive Translation Equivariance + +JiaKui Hu $^{1,2,3}$ , Zhengjian Yao $^{1,2,3}$ , Lujia Jin $^{4}$ , Hangzhou He $^{1,2,3}$ , Yanye Lu $^{1,2,3*}$ $^{1}$ Institute of Medical Technology, Peking University Health Science Center, Peking University + $^{2}$ Biomedical Engineering Department, College of Future Technology, Peking University + $^{3}$ National Biomedical Imaging Center, Peking University + $^{4}$ China Mobile Research Institute + +jkhu29@stu.pku.edu.cn, yanye.lu@pu.edu.cn + +# Abstract + +Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivalence Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization. + +# 1. Introduction + +Image restoration aims to reconstruct high-quality (HQ) images from their low-quality (LQ) counterparts, which are often affected by degradations resulting from the imaging system or environment. As a constituent of the imaging system, image restoration methods should satisfy the following fidelity property: the restoration results for the target region should remain equivariant under geometric transfor- + +mations of the input. This fidelity property is termed translation equivariance, which is one of the inherent inductive biases in image restoration. + +![](images/cdfa586a084a50a3c2d482ef86180369c3b78e0a3c76f38347db32ba256bbb6e.jpg) + +![](images/ee6cd240c1cf4f8aaff89f54d7d19f1c6cb10270ebb3df8f46fe4852cc26134a.jpg) + +![](images/751f7cbbb4f62e1fdce47b7731e7cad2c80493e55ae46a1fa2e299befad9dadd.jpg) + +![](images/cf0a4a3481b6cba5998820aa47ad0a7839d26b8a71ffd1fa410e6a48c6b105cf.jpg) +Figure 1. $\mathcal{T}(\cdot)$ means the translation function. "SA" and "WA" are commonly used for self-attention and window attention, respectively, but disrupt translation equivariance (TE) due to position encoding and feature shifting. "TEA" is our proposed translation equivariance adaptive attention, which satisfies TE. TE promotes faster convergence and better generalization. + +![](images/da6343229e66edb0dd5ff2a569869cbaad7a89debfbf41c46a8a2df35698dbe0.jpg) + +![](images/a8d13e64892278fc66fdd464a8c3aa86cd7b0ae570742405a579726918a8348b.jpg) + +![](images/408acbc476c84a68cb551091348986afecde47974edc227cdc3008eded1723a6.jpg) + +![](images/c391195427018b5c444fad6e84337353ae4342c12105840af78cd4294df64b2d.jpg) + +
AttentionsTEPerformanceNTK [25]SRGA [33]Complexity
PSNR↑/ SSIM↑Condition ↓Value ↓
SAX28.21 / 0.843511140.444.075O(N2)
WAX27.45 / 0.82541746.493.655O(N)
TEA (Ours)✓28.67 / 0.8489236.783.275O(N)
+ +Table 1. In $4\times$ image super-resolution, we study performance, NTK [8] convergence speed, SRGA [33] generalization, and computation complexities. We replace WA in SwinIR with SA and TEA for testing. TEA shows fast convergence, better generalization, and linear complexity. The results are tested on Urban100. + +In restoration networks [32, 62, 68, 71] based on convolution neural networks (CNNs), the paradigm of sliding information extraction inherent to convolutions [18] imparted a natural translation equivariance (TE) property in restoration networks. However, as shown in Figure 1 and Table 1, + +the advent of Self Attention (SA) [51] and Window Attention (WA) [34] within Transformers has broken this fidelity property, leading to slow convergence and poor generalization in restoration transformers. + +In this paper, we propose two fundamental strategies to bring TE back into the restoration transformers: (1) slide indexing and (2) component stacking. The slide indexing strategy employs a fixed container, e.g., a fixed rectangular window, to globally extract the input in a sliding fashion. This method requires the operators to calculate instantaneously, thereby maintaining the consistency of positional coordinates between the input and output of the model, which facilitates the integration of TE. The component stacking strategy involves stacking modules that satisfy the TE property in parallel or in series, thus enhancing architectures while ensuring TE in complex networks. + +However, these foundational strategies present a persistent challenge in the design of restoration transformers. Selecting SA to achieve superior performance involves accepting its substantial computational complexity [63]; alternatively, opting for sliding window attention provides linear complexity but results in a fixed receptive field [48] and reduced performance. To break this dilemma, we replace the global key-value pair indexing mechanism in vanilla SA with an adaptive slide indexing mechanism to design the Translation Equivariance Adaptive (TEA) attention. To retain the extraction of global information, TEA incorporates a downsampling self-attention branch without position encoding, which reduces the key-value pair's resolution by aggregating them, thereby providing TEA with coarse, yet effective global information. + +In summary, our contributions are threefold. First, we investigate TE in image restoration, analyzing its impact on convergence and generalization. The results indicate that the integration of TE is essential in the design of restoration networks. Second, we propose TEA, which uses adaptive slide indexing methods to incorporate the TE property while preserving linear complexity and ensuring a global receptive field, thereby enhancing restoration transformers. Third, we introduce TEA to the vanilla Vision Transformer to design TEAFormer. Experiments show that TEA achieves State-of-The-Art (SoTA) performance in various image restoration tasks, while having faster convergence and more robust generalization. + +# 2. Related Work + +Vision Transformers with inductive bias. Following the pioneering introduction of the Transformer [51] into vision tasks by ViT [16], SA has emerged as a key component in vision networks. However, ViT is challenging to train, exhibits slow convergence, and requires substantial data support [49]. Previous works, such as ViTAE [59, 70], attributed this to a lack of inductive bias and advocated in + +tegrating inductive bias with ViT to develop more efficient ones. AACNet [4] emphasized the importance of translation equivariance for visual tasks. The Swin [34] and CSwin [15] Transformers incorporated local processing into the vision Transformer, thus proposing window attention. CVT [56] and FasterViT [21] sequentially stack convolutions and attention layers, resulting in a hybrid architecture. PVT [52, 53] used average pooling for downsampling to achieve weak translation equivariance. Both DAT [57] and BiFormer [74] dynamically determined the windows accessible to each token. Recently, some studies have introduced translation equivariance into ViT by modifying position encoding [4, 58] and attention calculation logic [44, 46]. + +Recent image restoration Transformers. Similar to the development of vision transformers, the design of image restoration models also focuses on how to better design local global information processors [9, 31, 61, 63, 73]. Since the resolution of the input image is generally high in the image restoration task, directly using SA will result in a huge computation cost. To address this issue, SwinIR [31] introduced window attention [34]. HAT [9] enlarged the receptive field of the restoration transformers with the squeeze-excitation module [22] and the overlapping mechanism. Restormer [63] and DAT [10] performed channel-wise attention, thereby facilitating more efficient feature re-weighting. GRL [61] extracted the feature on local, regional, and global scales by integrating the attention of the anchor window. IPG [48] exploited the flexibility of the network by proposing the degree-variant graph solution, breaking the rigidity of the previous methods. + +# 3. Translation Equivariance + +In this section, we define the translation equivariance and propose two fundamental strategies for integrating translation equivariance into restoration networks. + +Definition 3.1. Translation Equivalence. We call function $\Phi(\cdot)$ is translation equivariant to translation operation $\mathcal{T}(\cdot)$ , if $\Phi(\mathcal{T}(x)) = \mathcal{T}(\Phi(x))$ , where $x$ is the input signal. + +In image restoration tasks, $x$ is an image or a feature in the latent space typically, $\mathcal{T}(\cdot)$ refers to a shift in pixels [20]. Based on this definition, two theorems are derived: + +Theorem 3.2. If $\Phi(x)_i$ is transformed from $x_j = [i - b, i + b]$ , where $b$ is the sliding boundary, the function $\Phi(\cdot)$ fulfills Definition 3.1. + +Theorem 3.3. If the functions $\Phi_1(\cdot)$ and $\Phi_2(\cdot)$ fulfill Definition 3.1, the functions $\Phi_2(\Phi_1(\cdot))$ and $\Phi_1(\cdot) + \Phi_2(\cdot)$ also fulfill Definition 3.1. + +Reformulating the existing operators in line with Theorem 3.2 gives them the translation equivariance property. By stacking these operators, either in parallel or in serial, as specified in Theorem 3.3, more complex architectures are also capable of translation equivariance. Consequently, the model can be further enhanced via translation equivariance. + +![](images/d39d1908a1a44d458737227015e082d0cd52a96a1591ab640e3aa5928d0a8ea1.jpg) +Figure 2. The overall architecture of our proposed TEAFormer. + +# 4. Method + +In this section, we conform to Theorems 3.1 and 3.2 to devise a translation equivariance restoration transformer, termed Translation Equivariance trAnsFormer (TEAFormer). We first present the overall architecture of TEAFormer and then introduce its components in detail. + +# 4.1. Overall Architecture + +The overall architecture of the proposed TEAFormer is illustrated in Figure 2 with image super-resolution (SR) presented as a demonstration. Given an input image $I \in \mathbb{R}^{H \times W \times 3}$ , where $H \times W$ denotes the resolution, TEAFormer first applies a convolution layer to obtain a shallow feature $F_{0} \in \mathbb{R}^{H \times W \times D}$ , where $D$ is the embedding dimension of the network. Next, following classical restoration networks [31], TEAFormer adopts a residual-in-residual structure to construct a deep feature extraction module comprised of $N_{g}$ Translation Equivariance Groups (TEGs). Each TEG consists of $N_{b}$ Translation Equivariance Blocks (TEBs) and a convolution layer. Each TEB, in turn, is composed of one Translation Equivariance Attention (TEA) and one feed-forward network. Subsequently, a convolutional layer is employed to extract the deep feature $F_{1} \in \mathbb{R}^{H \times W \times D}$ from the feature extracted by TEGs. After extraction through the deep feature extraction module, we use the restore module to generate the high-quality image $\hat{I}$ from the feature $F_{r} = F_{0} + F_{1}$ . As for other image restoration tasks that do not involve changes in resolution, we build our TEAFormer following Restormer [63]. + +# 4.2. Adaptive Translation Equivariance + +Adaptive Translation Equivalence is integrated through the Adaptive Sliding Key-Value Self Attention (ASkvSA) and Downsampled Self Attention (DSA) as shown in Figure 3. + +Key-value indexing in self-attention. The nature of Self Attention (SA) without position bias lies in the indexing of key-value pairs, followed by feature re-weighting. Given a flattened input $X \in \mathbb{R}^{N \times D}$ , where $N$ represents the number of tokens and $D$ represents the embedding dimension, self-attention can be formulated as following: + +$$ +\operatorname {S A} (X) = \operatorname {S o f t m a x} \left(\frac {X W _ {q} \left(X W _ {k}\right) ^ {\top}}{\sqrt {D}}\right) X W _ {v}, \tag {1} +$$ + +where $W_{q}, W_{k}$ , and $W_{v}$ denote the weights of the linear + +![](images/e6c10ae9c6e10be075614c0c5d9a19113754516380df00a1cf2d3bcc1ac65d6a.jpg) +Figure 3. The inner structure of TEA with two components. (a) Adaptive Sliding key-value Self Attention (ASkvSA) and (b) Downsampled Self Attention (DSA). + +projections utilized to compute the query $(Q)$ , key $(K)$ , and value $(V)$ from input $X$ . + +In Eq. 1, the softmax similarity function is utilized for the generation of attention maps between a query and a key, while the dot product is employed for re-weighting between the attention map and a value. Given that subscripting a matrix with $i$ returns the $i$ -th row as a vector, an alternative equation of self-attention can be written as follows. + +$$ +\operatorname {S A} _ {i} (X) = \frac {\sum_ {j = 1} ^ {N} \operatorname {s i m} \left(Q _ {i} , K _ {j}\right) V _ {j}}{\sum_ {j = 1} ^ {N} \operatorname {s i m} \left(Q _ {i} , K _ {j}\right)}, \tag {2} +$$ + +where $\mathrm{sim}(Q_i, K_j) = \exp \left(\frac{Q_i^\top K_j}{\sqrt{D}}\right)$ . + +Eq. 2 illustrates that self-attention indexes key-value pairs $(K_{j}$ and $V_{j}, j = 1, \dots, N)$ for each query $(Q_{i})$ . + +Sliding key-value Self Attention. In accordance with Theorem 3.2, we demonstrate that sliding window attention constitutes a generalized form of self-attention, which can be designated as Sliding Key-Value Self-Attention (SkvSA). SkvSA is formulated as follows. + +$$ +\operatorname {S k v S A} _ {i} (X) = \frac {\sum_ {j = (i - b _ {1}) \bmod s = 0} ^ {(i + b _ {2}) \bmod s = 0} \operatorname {s i m} \left(Q _ {i} , K _ {j}\right) V _ {j}}{\sum_ {j = (i - b _ {1}) \bmod s = 0} ^ {(i + b _ {2}) \bmod s = 0} \operatorname {s i m} \left(Q _ {i} , K _ {j}\right)}, \tag {3} +$$ + +where $b_{1}$ and $b_{2}$ are the sliding boundary on the left and right when sliding on the one-dimensional vector $X$ . In our SkvSA, $b_{1} = b_{2} = \frac{w\cdot s}{2}$ , where $w$ is the size of the sliding window and $s$ is the stride size during sliding. + +Eq. 3 says that, for the $i$ -th query $Q_{i}$ , SkvSA systematically extracts the key-value pairs from a designated sliding window, extending from the index $i - \frac{w \cdot s}{2}$ to the index $i + \frac{w \cdot s}{2}$ with stride $s^2$ . Subsequently, SkvSA calculates the attention map and re-weights the feature in place. + +Boundary processing. Note that, in scenarios where $i - \frac{w \cdot s}{2} < 0$ or $i + \frac{w \cdot s}{2} > N$ occur, the slide indices of the key-value pairs $j$ will exceed the boundary of the input $X$ , leading to an indexing failure. To ameliorate the issue associated with slide indexing at sequence's boundaries, we propose a boundary blocking indexing method. Specifically, for queries located at the boundary of the sequence, the sliding indices of key-value pairs are "blocked" and converted into indexing within a predefined blocking window. In this study, the sizes of the predefined blocking window and the sliding window are identical. Accordingly, the sliding boundaries $b$ can be appropriately revised as follows: + +$$ +b = \left(b _ {1}, b _ {2}\right) = \left\{ \begin{array}{l l} \left(0, \frac {w \cdot s}{2}\right), & i - \frac {w \cdot s}{2} < 0, \\ \left(\frac {w \cdot s}{2}, N\right), & i + \frac {w \cdot s}{2} > N, \\ \left(\frac {w \cdot s}{2}, \frac {w \cdot s}{2}\right), & \text {o t h e r s .} \end{array} \right. \tag {4} +$$ + +The SkvSA and its boundary processing are capable of extracting and processing information via a fixed sliding window, adhering to the translation equivariance property as established in Theorem 3.2. + +Adaptive Sliding key-value Self Attention. However, the fixed window size constrains the receptive field, leading to the entrapment of restoration transformers within rigidity [48]. To mitigate this rigidity, it is anticipated that the SkvSA will ascertain the optimal window size and position specific to each query $Q_{i}$ . To achieve this endeavor, we devise the Adaptive Sliding key-value Self Attention (ASkvSA), which adaptively determines key-value indices and shuffles them within the fixed sliding window of $Q_{i}$ . + +As shown in Figure 3 (a), with regard to Key $K$ and Value $V$ , ASkvSA initially reshapes $K$ and $V$ from one-dimensional (1d) vectors into two-dimensional (2d) matrices. Subsequently, it applies a 2d depth-wise convolution with kernel size $k$ to generate their adaptive indices $\mathcal{F} \in \mathbb{R}^{H \times W \times 2}$ . $\mathcal{F}[h, w]$ is a vector of length 2 in the $h-$ th row and $w-$ th column of $\mathcal{F}$ , which can describe a coordinate. Following this, ASkvSA performs key-value indexing based on established indices $\mathcal{F}_K$ and $\mathcal{F}_V$ by shuffling the pixel located at the coordinate $\mathcal{F}[h, w]$ of $K$ and $V$ to the pixel at the coordinate $(h, w)$ of $K$ and $V$ . + +$$ +\operatorname {A S k v S A} (X) = \frac {\sum_ {j = \mathcal {F} (K) , l = \mathcal {F} (V)} \sin \left(Q _ {i} , K _ {j}\right) V _ {l}}{\sum_ {j = \mathcal {F} (K)} \sin \left(Q _ {i} , K _ {j}\right)}. \tag {5} +$$ + +Adaptive slide indexing does not compromise the translation equivariance of ASkvSA. ASkvSA stacks convolutions with SkvSA in serial to generate adaptive indices of key-value pairs. Given that convolution satisfies the property of translation equivariance, when $K_{h,w}$ is translated to $K_{h + \delta_h, w + \delta_w}$ , key's adaptive indices $\mathcal{F}_K$ will shuffle the pixel located at $\mathcal{F}[h + \delta_h, w + \delta_w]$ of $K$ to the pixel at $(h + \delta_h, w + \delta_w)$ of $K$ . According to Theorem 3.3, ASkvSA satisfies the translation equivariance property. + +Adaptively aggregated with global key-value pairs. Although ASkvSA can identify appropriate key-value pairs beyond a fixed sliding window, some remote but valid pixels still escape ASkvSA's indexing. Consequently, we propose downsampled self-attention (DSA), which adaptively combines global key value indexes through downsampling, providing the network with a coarse, yet efficient global indexing of key-value pairs. DSA can be formulated as: + +$$ +\mathrm {D S A} _ {i} (X) = \frac {\sum_ {j = 1} ^ {N _ {d}} \operatorname {s i m} \left(Q _ {i} , K _ {j} ^ {\prime}\right) V _ {j} ^ {\prime}}{\sum_ {j = 1} ^ {N _ {d}} \operatorname {s i m} \left(Q _ {i} , K _ {j} ^ {\prime}\right)}, \tag {6} +$$ + +where $K'$ and $V'$ represents the downsampled key-value pair calculated in ASkvSA, and $N_d$ denotes the number of tokens associated with the downsampled key-value pair. We employ average pooling as the downsampling operator due to its efficiency and coarse fulfillment of TE. + +Finally, according to Figure 3, the ASkvSA and DSA outputs are adaptively aggregated to obtain our TEA. + +$$ +\mathrm {T E A} _ {i} (X) = \alpha_ {s} \operatorname {A S k v S A} (X) + \alpha_ {d} \operatorname {D S A} (X), \quad (7) \text {w h e r e} \alpha_ {s} \text {a n d} \alpha_ {d} \text {a r e l a r n a b l e p a r a m e t e r s . T E A s t a c k s D S A w i t h A S k v S A i n p a r a l l e l a r . A c c o r d i n g t o T h e o r e m 3 . 3 , T E A s a t s i s f i e s t h e t r a n s l a t i o n e q u i v a r i a n c e p r o p e r t y .} +$$ + +# 4.3. Computational Cost and Hyper-Parameters + +We demonstrate that TEA can maintain a computational cost that is similar to that of window attention [34] under well-designed hyperparameter settings. We first list four integral hyperparameters of TEA: + +- $w$ : the sliding window size of ASkvSA; +- $s$ : the sliding stride size of ASkvSA; +- $k$ : the kernel size of adaptive indexes generation; +- $N_{d}$ : the number of tokens in DSA. + +According to Section 4.2, TEA uses $3ND^{2}$ FLOPs for the Query-Key-Value linear projections, $2NDk^{2}$ FLOPs for the depth-wise convolution in adaptive indices generation, $Nw^{2}D$ FLOPs for attention map computation between query $Q \in \mathbb{R}^{N \times D}$ and indexed key $K \in \mathbb{R}^{w^{2} \times D}$ , $Nw^{2}D$ FLOPs for feature re-weighting between attention map and indexed value $V \in \mathbb{R}^{w^{2} \times D}$ , and $2NN_{d}D$ FLOPs for DSA. The total FLOPs of TEA is $3ND^{2} + 2NDk^{2} + 2Nw^{2}D + 2NN_{d}D$ . With the hyper-parameters held constant, the FLOPs of TEA exhibit linear growth as a function of $N$ , thereby implying that TEA is $\mathcal{O}(N)$ . + +Consequently, when the hyper-parameters of TEA are adjusted carefully, its computational cost becomes comparable to that of window attention. In our TEAFormer, the hyper-parameter $w$ has been set to 15, $k$ to 3, and $N_{d}$ to 16, thereby modulating the computational complexity of TEA to be even slightly lower than that of window attention with window size 16. + +
Modelwindow sizeTranslation equivarianceParamsFLOPsLatencyPerformanceNTK [25]SRGA [33]
SkvSAASkvSADSA(M)(T)(ms)Urban100 [24] PSNR↑ / SSIM↑Condition ↓Value ↓
SwinIR [31]811.900.462130.027.45 / 0.82541746.493.655
SwinIR-Large1621.420.897214.527.94 / 0.83621554.653.610
TEAFformer15✓21.860.688230.128.31 / 0.8444243.753.206
TEAFformer15✓21.980.757284.428.47 / 0.8457283.993.298
TEAFformer15✓✓21.860.966340.928.49 / 0.8470203.063.261
TEAFformer15✓✓21.981.035386.728.67 / 0.8489236.783.275
+ +Table 2. Ablation study on TEA. Our TEA enhances performance, convergence speed, and generalization. + +
MethodScaleParams (M)FLOPs (T)Latency (ms)Set5 [5]Set14 [64]B100 [37]Urban100 [24]Manga109 [40]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
EDSR [32]40.76.006147.138.110.960233.920.919532.320.901332.930.935139.100.9773
RCAN [71]15.42.259155.538.270.961434.120.921632.410.902733.340.938439.440.9786
SwinIR-L [31]0.91156.4188.638.140.961133.860.920632.310.901232.760.934039.120.9783
SwinIR [31]11.81.848514.438.420.962334.460.925032.530.904133.810.942739.920.9797
EDT [30]11.51.965101038.450.962434.570.925832.520.904133.800.942539.930.9800
SRFormer [73]×210.31.790787.238.510.962734.440.925332.570.904634.090.944940.070.9802
HAT [9]20.63.662901.938.630.963034.860.927432.620.905334.450.946640.260.9809
GRL-B* [61]20.14.422262138.480.962734.640.926532.550.904533.970.943740.060.9804
DAT [10]14.72.155121538.580.962934.810.927232.610.905134.370.945840.330.9807
IPG [48]16.84.732270338.610.963234.730.927032.600.905234.480.946440.240.9810
TEAFormer-L (Ours)0.830.237254.038.270.861834.260.924032.430.902933.510.939639.560.9792
TEAFormer (Ours)21.84.133149338.640.963135.120.928432.720.906435.230.950940.320.9807
EDSR [32]43.11.85349.1332.460.896828.800.787627.710.742026.640.803331.020.9148
RCAN [71]15.60.58746.1532.630.900228.870.788927.770.743626.820.808731.220.9173
SwinIR-L [31]0.880.03948.7132.440.897628.770.785827.690.740626.470.798030.920.9151
SwinIR [31]11.90.462130.032.920.904429.090.795027.920.748927.450.825432.030.9260
EDT [30]11.60.514251.332.820.903129.090.793927.910.748327.460.824632.030.9254
SRFormer [73]×410.40.470189.232.930.904129.080.795327.940.750227.680.831132.440.9271
HAT [9]20.80.938227.733.040.905629.230.797328.000.751727.970.836832.480.9292
GRL-B* [61]20.21.129761.832.900.903929.140.795627.960.749727.530.827632.190.9266
DAT [10]14.80.561296.833.080.905529.230.797328.000.751527.870.834332.510.9291
IPG [48]17.01.206646.733.150.906229.240.797327.990.751928.130.839232.530.9300
TEAFormer-L (Ours)0.830.05969.6432.630.900228.960.789927.820.744227.040.810531.560.9197
TEAFormer (Ours)22.01.035386.733.230.906629.460.800928.110.754428.670.848932.990.9323
+ +Table 3. Classical image SR results. Performance and complexity are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two scores in each column are highlighted in red and blue. Methods with “*” are replicated with standard training setting, following IPG [48]. The “-L” indicates that the given method is a lightweight version. + +![](images/6c7067faf5c2b46ff3b533a01d757f0cb5961cc2b2d1bc6dbf7286c7317ba969.jpg) +BSD100 $(4\times)$ :img_86016 + +![](images/840872a5dd68612db0b0565822aa3a0a03c1daf2f1af968502b703786197ec8d.jpg) +Bicubic + +![](images/a074904d3c0310f9e029dab75edb38e89fd1dde70d61c8b307415fc32b51da01.jpg) + +![](images/307518198a438373c3d3f5fba824ef264c65e277d8b405e63f303427e51d2395.jpg) + +![](images/c1d57f9f8802a91cb21a666fd6406dd726a5b0bcd49b17d4204fa05d06d09eeb.jpg) + +![](images/3118184afb85fe7f7e9c6c9b032b153a0502791a2113b5b827339d81f3e663b6.jpg) +Urban100 $(4\times)$ : img_033 + +![](images/ec5b7b7d95ae8cd4d83052467bac870b16d7f4c6d3068194afa984a5310f9a22.jpg) +HQ + +![](images/ad1d99216bd1f3a2bc3bc267b173c9b9eb2d86f04b3b223c383297d283b94b45.jpg) +RCAN [71] + +![](images/e47ff34894efd7c4063afe023e1eb6e2c71c9c921e33ae87b966fdfe49565a52.jpg) +SwinIR [31] + +![](images/1353974588e7314d074f12197e44c1956eba46dfb7ad889f5b78f9c5bfc9d1ad.jpg) +HAT [9] + +![](images/ede9013da5424406bfd1cb581208ffe1035bff4ba5cb2cac8470effb764450be.jpg) + +![](images/69c463004db8e8a1705f04e74eda412ad0d36d07a2125ef0e0757e4099bcdb9a.jpg) +SRFormer [73] + +![](images/679fa1161e951e2f67436f92d5fc7aaf7c52e7a89e2ca1ea7a670cfccc162c55.jpg) +IPG [48] + +![](images/95586d9d0a06acbb543721aa1974282dfb39f2f47fa3cdc8eb04c1025e685100.jpg) +TEAFormer (Ours) + +![](images/59051c683c0dd65276b62f2680811ce0bf7a606ef16ca7f15f2cb75e8b9b37c7.jpg) +Bicubic +HQ + +![](images/f0d908e7afd0c624201203807a2cea98a13f5518991f55a8ef3a75c7c45578db.jpg) +RCAN [71] +SRFormer [73] +Figure 4. Qualitative comparison with recent SOTA methods on the image SR task. + +![](images/6fd2f6a1eddd722cd5726cbeeba4e00fcc09af5cd030d1f27335a21db28fd2ba.jpg) +SwinIR [31] +IPG [48] + +![](images/0cc8778a261620189553d400b5e40ce540bda41f8bcdce79b24bbdddb39c0df5.jpg) +HAT [9] +TEAFormer (Ours) + +
MethodParamsIndoor ScenesOutdoor ScenesCombined
(M)PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓
DPDNetS [1]32.326.540.8160.0310.23922.250.6820.0560.31324.340.7470.0440.277
KPACs [47]2.0627.970.8520.0260.18222.620.7010.0530.26925.220.7740.0400.227
IFANs [27]10.528.110.8610.0260.17922.760.7200.0520.25425.370.7890.0390.217
RestormerS [63]26.128.870.8820.0250.14523.240.7430.0500.20925.980.8110.0380.178
SFNetS [12]13.229.160.8780.0230.16823.450.7470.0490.24426.230.8110.0370.207
GRLS-B [61]19.929.060.8860.0240.13923.450.7610.0490.19626.180.8220.0370.168
IRNeXTS [13]13.229.220.8790.0240.16723.530.7520.0490.24426.300.8140.0370.206
TEAFormerS (Ours)15.429.500.8920.0230.15023.550.7670.0500.21126.450.8280.0370.181
DPDNetD [1]32.327.480.8490.0290.18922.900.7260.0520.25525.130.7860.0410.223
RDPD_D [2]8.2028.100.8430.0270.21022.820.7040.0530.29825.390.7720.0400.255
IFAN_D [27]10.528.660.8680.0250.17223.460.7430.0490.24025.990.8040.0370.207
Restormer_D [63]26.129.480.8950.0230.13423.970.7730.0470.17526.660.8330.0350.155
Uformer_D [54]20.628.230.8600.0260.19923.100.7280.0510.28525.650.7950.0390.243
GRL_D-B [61]19.929.830.9030.0220.11424.390.7950.0450.15027.040.8470.0340.133
TEAFormer_D (Ours)15.430.160.9050.0220.12824.450.7950.0450.17027.230.8490.0340.150
+ +Table 4. Defocus deblurring comparisons on the DPDD [1] testset (containing 37 indoor and 39 outdoor scenes). S: single-image defocus deblurring. D: dual-pixel defocus deblurring. The best two results are highlighted in red and blue, respectively. + +# 5. Experiments + +# 5.1. Ablation study + +We analyze the influence exerted by the TEA on both performance, computational complexity, generalization, and training convergence in Table 2. In this experiment, we disassemble TEA into several plug-and-play modules and gradually add them into the SwinIR baseline restoration network. To ensure comparable parameters and FLOPs among the models being evaluated, we adjusted the window size of SwinIR to 16 and increased the number of parameters to 21 M, which is termed SwinIR-Large. Following the incorporation of our SkvSA, there is a marked improvement in the convergence rate and generalization abilities. Subsequent to the integration of our ASkvSA and DSA, there is a progressive enhancement in the model's performance, without detrimental effects on convergence and generalization. The results of this experiment demonstrate the efficacy of each component within the TEA. Furthermore, it illustrates that TEA resolves the dilemma between self-attention and sliding window attention, thereby attaining high performance by the global receptive field and reducing computational complexity simultaneously. + +# 5.2. Image SR + +Training settings. We used the same training hyperparameters and evaluation methods as IPG [48]. All of our models were trained from scratch on the DF2K [3, 32] datasets with $64 \times 64$ resolution. The optimizer was Adam [26], the learning rate was $2 \times 10^{-4}$ while using the cosine scheduler. We used $\mathrm{L}_1$ loss as a criterion. + +Results. We first compare TEAFormer with various classical image SR methods, including EDSR [32], RCAN [71], SwinIR [31], EDT [30], HAT [9], SRFormer [73], GRL [61], DAT [10], and IPG [48]. We compare them on synthetic benchmarks [5, 24, 37, 40, 64] generated by bicubic downsampling. The FLOPs and latency reported were + +calculated with one NVIDIA A800 GPU on 768 resolution. + +The quantitative comparison of the methods for classical image SR is shown in Table 3. The exceptional performance of TEAFormer is distinctly evident as it achieves the highest results across nearly all five datasets for various scale factors. Incorporating the attention module with translation equivalence introduces a more suitable and productive inductive bias to the restoration model, thereby enhancing its performance. For the $\times 4$ SR, our TEAFormer achieves a $28.67\mathrm{dB}$ PSNR score on the Urban100 dataset, which is $0.7\mathrm{dB}$ higher than HAT [9] while both has $20\mathrm{M}$ parameters. In comparison to the preceding image SR method IPG [48], TEAFormer demonstrates a gain of $0.54\mathrm{dB}$ with comparable FLOPs and a $2\times$ faster inference. + +We further conduct TEAFormer-Light, which only comprises 829k parameters. In comparison to the CNN method HAN [42], TEAFormer-Light has been demonstrated to achieve superior performance with only its $1.3\%$ parameters. In comparison to other lightweight transformer methods, TEAFormer also demonstrates an absolute performance advantage, as evidenced by the fact that TEAFormer-Light achieves a higher PSNR than SwinIR-L on the Urban100 [24] dataset by 0.57 dB. + +# 5.3. Image Defocus Deblurring + +Training settings. We employed the same training hyperparameters and evaluation as Restormer [63] in image deblurring tasks. Our TEAFormer were trained from scratch using progressive learning [63]. During the training phase, the optimizer adopted was AdamW [35] with $\mathrm{L}_1$ loss. + +Results. We compare our TEAFormer with single image defocus deblurring methods [1, 13, 14, 27, 47, 61, 63] and dual-pixel image defocus deblurring methods [1, 2, 27, 54, 61, 63] on DPDD [1] dataset in Table 4. Our TEAFormer significantly outperforms state-of-the-art schemes for the single-image and dual-pixel defocus deblurring tasks on PSNR and SSIM metrics. In particular in the category + +![](images/3a50c42c38dc7fd6d4882f9cf2e7aae606bdf0310ed3247b4116b9e2848ca586.jpg) + +![](images/e8d8386e85b8b68b24f1c3464dac06ca8cde59ef568637255ad28223227d1348.jpg) +Figure 5. Qualitative comparison with recent SOTA methods on the image defocus blur task. Please zoom in for better view. + +![](images/5535a2ec071383096f6e31e4a7fa1f9d438cef59cf2b65e394d3c73871c5cc37.jpg) + +![](images/c0ee018cd4b939932d67af6dd9095e6507b1f7cba06de2478781c79b1cc0af70.jpg) + +![](images/e93c9b4174e7f866d190746fde1dcb62a83bdfd4643e24cb48fc549b7a90bb62.jpg) + +
MethodParamsColorGrayscale
(M)CBSD68 [39]Kodak24 [17]McMaster [69]Urban100 [24]Set12 [67]BSD68 [39]Urban100 [24]
σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50
DnCNN [67]0.5633.9031.2427.9534.6032.1428.9533.4531.5228.6232.9830.8127.5932.8630.4427.18
DRUNet [68]32.634.3031.6928.5135.3132.8929.8635.4033.1430.0834.8132.6029.6133.2530.9427.90
EDT-B [30]11.534.3931.7628.5635.3732.9429.8735.6133.3430.2535.2233.0730.16---
SwinIR* [31]11.734.3931.7828.5535.3332.8829.7735.5933.2030.1935.1132.8929.7733.3431.0027.88
Restormer [63]26.134.3931.7828.5935.4433.0230.0035.5533.3130.2935.0632.9130.0233.3531.0428.01
ART* [66]16.234.4031.7928.5935.4733.0430.0035.6133.3330.2735.1232.9529.9933.3531.0528.01
GRL-B* [61]19.834.4031.7828.5835.4233.0330.0235.6333.3630.2635.2233.0130.2433.3431.0227.99
TEAFormer (Ours)15.434.4031.8028.6235.5033.0830.0535.6833.4230.3835.2533.1830.3733.3731.0628.04
+ +Table 5. Color and grayscale image Gaussian denoising results in terms of PSNR $\uparrow$ . Model complexity and prediction accuracy are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two results are highlighted in red and blue, respectively. Methods with " $*$ " are replicated with standard training setting, which learns a single model to handle various noise levels. + +of single-pixel indoor scene, TEAFormer yields 0.44 dB improvements over the GRL previous best transformer method [61]. Figure 5 illustrates that our TEAFormer is more effective in removing spatially varying defocus blur than other restoration transformers. + +# 5.4. Image Denoising + +Training settings. The image denoising and defocus deblurring tasks used the same hyper-parameters for training. + +Gaussian Denoising Results. We perform denoising experiments on synthetic benchmark datasets generated with additive white Gaussian noise (Set12 [67], BSD68 [38], Urban100 [24], Kodak24 [17] and McMaster [69]). Following DRUNet [68], we learn a single model to handle various noise levels, including 15, 25, and 50. Our TEAFourner achieves state-of-the-art performance under both experimental settings on different datasets and noise levels. Specifically, for the challenging noise level 50 on color Urban100 [24] dataset, TEAFourner achieves 0.13 dB gain over the previous best restoration transformer GRL [61], as shown in Table 5. Similar performance gains can be observed for the Gaussian gray denoising. + +# 5.5. All-in-one Image Restoration + +Training settings. Following DCPT [23], a 5D all-in-one image restoration and enhancement task was conducted with the objective of training the network to eliminate five distinct types of degradation: haze, rain, Gaussian noise, motion blur, and low light. The models were trained from scratch on datasets with $128 \times 128$ resolution without progressive learning. During training, the optimizer was set as AdamW, the learning rate was set as $3 \times 10^{-4}$ while using + +the cosine scheduler. We used $L_{1}$ loss as a criterion. + +Results. As illustrated in Table 6, TEAFormer surpasses the most recently proposed methods specifically designed for comprehensive all-in-one restoration tasks (IDR [65], PromptIR [43], and InstructIR [11]). The average PSNR of TEAFormer in all five tasks reached 31.22 dB, thereby exceeding the current state-of-the-art 5D all-in-one restoration method, InstructIR [11], by 1.67 dB. TEAFormer consistently outperforms prior PSNR in every indicator of each task, signifying its exceptional proficiency in image restoration under complex multi-task conditions. + +Mixed degradation. As shown in Table 7, TEAFormer surpasses the most recently proposed methods under mixed degradation scenarios, specifically designed for universal restoration tasks [36, 72]). Within scenarios characterized by a combination of three degradation (low-light, haze, and rain), TEAFormer demonstrates superior performance over the current state-of-the-art universal restoration method, DiffUIR [72], by a margin of $0.42\mathrm{dB}$ . + +# 5.6. Discussion + +Hyperparameters analysis. The performance under various hyperparameters is reported in Table 8 $(w)$ , Table 9 $(s)$ , Table 10 $(k)$ and Table 11 $(N_d)$ . Our settings $(w = 15, s = 4, k = 3, N_d = 16)$ optimally balanced performance and computation cost. Note that changing some of the hyperparameters does not affect the computational complexity as discussed in Section 4.3. For $s$ in Table 9, parameters and FLOPs are omitted because $s$ does not affect them. + +Compare different downsampler in DSA. The incorporation of DSA into our TEA substantially enhances the performance of the model without adding undue complexity. + +
MethodParamsDehazing on SOTS [28]Deraining on Rain100L [60]Denoising on BSD68 [38]Deblurring on GoPro [41]Low-Light on LOL [55]Average
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.04 / 0.88432.98 / 0.95130.91 / 0.88224.35 / 0.78118.18 / 0.73525.49 / 0.846
IDR [65]15.325.24 / 0.94335.63 / 0.96531.60 / 0.88727.87 / 0.84621.34 / 0.82628.34 / 0.893
PromptIR [43]35.625.20 / 0.93135.94 / 0.96431.17 / 0.88227.32 / 0.84220.94 / 0.79928.11 / 0.883
InstructIR [11]35.127.10 / 0.95636.84 / 0.97331.40 / 0.88729.40 / 0.88623.00 / 0.83629.55 / 0.907
SwinIR [31]11.821.50 / 0.89130.78 / 0.92330.59 / 0.86824.52 / 0.77317.81 / 0.72325.04 / 0.835
NAFNet [7]67.925.23 / 0.93935.56 / 0.96731.02 / 0.88326.53 / 0.80820.49 / 0.80927.76 / 0.881
Restormer [63]26.124.09 / 0.92734.81 / 0.96231.49 / 0.88427.22 / 0.82920.41 / 0.80627.60 / 0.881
TEAFormer (Ours)15.431.57 / 0.98039.44 / 0.98631.54 / 0.89030.53 / 0.90823.06 / 0.85631.22 / 0.924
+ +![](images/a4d8814858f9972493a82632b410fcd27208c6746a1d3afddb67a07a59ae5fb7.jpg) +LQ + +![](images/8b1f2300adaa7ed38d31071431e54bd8c0cd46762e74eaabb3a3173c529cc93f.jpg) +InstrucrIR [11] + +![](images/71d0654490210f525dafda7e9d73ac836263dad432f0b7b25f8018f4823c5362.jpg) +OneRestore [19] + +![](images/99e7714b36460186d16a8eedbb92ac31227c3593ab5eede043fa2d827c0d16a1.jpg) +HQ + +![](images/12c8e9e1664bf7e969153bac10b2f9fe3a715e5b58f0b44559a78b83ac274668.jpg) +DACLIP [36] + +![](images/56d5bc26ccd2281c8dd0451fbde398d2e161d806ef052f8c8ed7de148be47eb4.jpg) +TEAFormer (Ours) +Figure 6. Visual comparison on low-light+haze+rain mixed degradation. TEAFormer can restore the correct lightness. + +Table 6. 5D All-in-one image restoration results. Dash lines separate the results of methods designed specifically for all-in-one image restoration task and methods that only focus on architecture enhancement. The best two results are highlighted in red and blue, respectively. TEAFormer sets the new state-of-the-art performance in each task. + +
MethodParams low-light+haze+rain low-light+haze+snow
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.80 / 0.70822.23 / 0.725
TransWeather [50]31.021.55 / 0.67821.01 / 0.655
PromptIR [43]35.623.74 / 0.75223.33 / 0.747
DACLIP [36]17422.96 / 0.71223.32 / 0.751
DiffUIR-L [72]36.325.46 / 0.77925.89 / 0.788
OneRestore [19]5.9825.18 / 0.79525.28 / 0.797
InstructIR [11]35.124.84 / 0.77724.32 / 0.760
TEAFormer (Ours)15.425.88 / 0.79626.05 / 0.800
+ +Table 7. Mixed degraded image restoration results on CDD [19]. + +
wFLOPs (T)Urban100 [24] PSNR ↑/SSIM ↑
70.82728.35 / 0.8441
151.03528.67 / 0.8489
311.86628.73 / 0.8493
+ +Table 8. Ablations of sliding window size $w$ . + +
sUrban100 [24]
PSNR ↑ / SSIM ↑
128.43 / 0.8431
228.51 / 0.8453
428.67 / 0.8489
+ +Table 9. Ablations of sliding stride $s$ + +
kFLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Nd(T)FLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Down (%)↑S-conv (%)↑Urban100 [24] PSNR ↑ / SSIM ↑
None0.96628.49 / 0.8470None0.75728.47 / 0.8457MaxPool87.328.61 / 0.8477
11.02828.53 / 0.8471161.03528.67 / 0.8489LPD91.728.64 / 0.8493
31.03528.67 / 0.8489N18.3628.80 / 0.8503AvgPool91.228.67 / 0.8489
+ +Table 10. Ablations of kernel size $k$ for adaptive in- in DSA. "None" stands for indices. "None" stands for no no DSA used. adaptive indices used. +Table 12. Ablations of down-samplers in DSA. The higher the S-Conv, the more strictly the model meets TE. + +However, the average pooling employed in DSA does not strictly adhere to the property of TE. Using a comparative ablation of several downsampling modules in Table 12, we experimentally demonstrate that the use of average pooling does not compromise the model's performance or its TE. + +Does TEAFormer converge faster and generalize better in practice? Table 1 and Table 2 have illustrated that TEAFormer converges faster and generalizes better than other networks theoretically due to the incorporation of translation equivariance. Figure 7 (left) further demonstrates that TEAFormer reaches a lower loss faster than other restoration transformers [9, 31] in practice. Following MaskDenoise [6], we compare the statistics of the last layer feature in SwinIR [31] and TEAFormer under two different real-world degradations on the PIES [33] dataset (out of the training dataset) in Figure 7 (right). The results show that TEAFormer can produce a more consistent output when processing LQ images with different degradations. Therefore, TEAFormer generalizes better in practice. + +![](images/e486577d5539c649550783b4f33df6f06ed8f4e981024632d753d3b438cedb94.jpg) +Figure 7. Comparison on convergence (left) and generalization (right). + +![](images/314d73fb472fdca92b1593ae33e8790d037d0eea31edcf1e6bcddd02dbf5d820.jpg) + +![](images/c275121ffca482548733363c9f7ee340838c565ac730280273df7cb45c6666c6.jpg) + +# 6. Conclusion + +In this paper, we start by examining the translation equivariance, which serves as an inherent inductive bias in image restoration. We then define the translation equivariance and offer two strategies for incorporating it into restoration networks: (1) slide indexing and (2) component stacking. To address the fixed receptive field provided by these strategies, we proposed TEA and TEAFormer via adaptive key-value indexing. The efficacy of TEAFormer has been demonstrated in various image restoration tasks, with ablation studies confirming that this efficacy arises from the incorporation of TEA. In the future, we intend to explore more equivariances into restoration networks. + +# Acknowledgement + +# References + +[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 6 +[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6 +[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6 +[4] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 2 +[5] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference. BMVA press, 2012. 5, 6 +[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 8 +[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision - ECCV 2022, pages 17-33, Cham, 2022. Springer Nature Switzerland. 8 +[8] Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre-training or strong data augmentations. In International Conference on Learning Representations, 2022. 1 +[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2, 5, 6, 8 +[10] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In ICCV, 2023. 2, 5, 6 + +This work was supported financially by the Natural Science Foundation of China (82371112, 623B2001), the Science Foundation of Peking University Cancer Hospital (JC202505), Natural Science Foundation of Beijing Municipality (Z210008) and the Clinical Medicine Plus X - Young Scholars Project of Peking University, the Fundamental Research Funds for the Central Universities (PKU2025PKULCXQ008). +[11] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 7, 8 +[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2022. 6 +[13] Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, and Alois Knoll. Irnext: Rethinking convolutional network design for image restoration. In Proceedings of the 40th International Conference on Machine Learning, 2023. 6 +[14] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 6, 7 +[15] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 2 +[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2 +[17] Rich Franzen. Kodak lossless true color image suite, 2013. 7 +[18] Jonathan Gordon, Wessel P. Bruinsma, Andrew Y. K. Foong, James Requeima, Yann Dubois, and Richard E. Turner. Convolutional conditional neural processes. In International Conference on Learning Representations, 2020. 1 +[19] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In European Conference on Computer Vision, 2024. 8 +[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185-6194, 2023. 2 +[21] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. In The Twelfth International Conference on Learning Representations, 2024. 2 +[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2 +[23] JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Universal image restoration pre-training via degradation classification. The Thirteenth International Conference on Learning Representations, 2025. 7 + +[24] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7, 8 +[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 1, 5 +[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 6 +[27] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6 +[28] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 2018. 8 +[29] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, 2022. 8 +[30] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1089–1097. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 5, 6, 7 +[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 5, 6, 7, 8 +[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 5, 6 +[33] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497-14513, 2023. 1, 5, 8 +[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 4 +[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6 +[36] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for universal image restoration. In The Twelfth International Conference on Learning Representations, 2023. 7, 8 +[37] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to + +evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423 vol.2, 2001. 5, 6 +[38] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 7, 8 +[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7 +[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools Appl., 76(20):21811-21838, 2017. 5, 6 +[41] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 8 +[42] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 191-207. Springer, 2020. 6 +[43] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. NeurIPS, 2023. 7, 8 +[44] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self-attention in vision models. Advances in neural information processing systems, 32, 2019. 2 +[45] Renan A Rojas-Gomez, Teck-Yian Lim, Alex Schwing, Minh Do, and Raymond A Yeh. Learnable polyphase sampling for shift invariant and equivariant convolutional networks. Advances in Neural Information Processing Systems, 35:35755-35768, 2022. 4 +[46] Renan A Rojas-Gomez, Teck-Yian Lim, Minh N Do, and Raymond A Yeh. Making vision transformers truly shift-equivariant. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5568-5577, 2024. 2 +[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proc. ICCV, 2021. 6 +[48] Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang. Image processing gnn: Breaking rigidity in super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24108-24117, 2024. 2, 4, 5, 6 +[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2 + +[50] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 8 +[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[52] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2 +[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2 +[54] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 6 +[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In British Machine Vision Conference, 2018. 8 +[56] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 2 +[57] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2 +[58] Renjun Xu, Kaifan Yang, Ke Liu, and Fengxiang He. $e(2)$ -equivariant vision transformer. In Uncertainty in Artificial Intelligence, pages 2356-2366. PMLR, 2023. 2 +[59] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vi-tae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34, 2021. 2 +[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. TPAMI, 2019. 8 +[61] Xiaoyu Xiang Yawei Li, Yuchen Fan, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6, 7 +[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 1 +[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. + +Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 3, 6, 7, 8 +[64] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference, Avignon, France*, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 6 +[65] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 7, 8 +[66] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 7 +[67] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 7 +[68] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 1, 7 +[69] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 7 +[70] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. arXiv preprint arXiv:2202.10108, 2022. 2 +[71] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 5, 6 +[72] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7, 8 +[73] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5, 6 +[74] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 2 \ No newline at end of file diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/images.zip b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc4f68f6627e80eb0e639ad5e89724269af02f99 --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6f2066a61b55e70ca6af747efc95866cc24949e55caeabbc3a576f17987e470 +size 1197232 diff --git a/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/layout.json b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b83a8eaf2835e76ce4f30d307f0b3bb633031a58 --- /dev/null +++ b/2025/Enhancing Image Restoration Transformer via Adaptive Translation Equivariance/layout.json @@ -0,0 +1,12986 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "spans": [ + { + "bbox": [ + 160, + 103, + 451, + 140 + ], + "type": "text", + "content": "Enhancing Image Restoration Transformer via Adaptive Translation Equivariance" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "spans": [ + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": "JiaKui Hu" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": ", Zhengjian Yao" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": ", Lujia Jin" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": ", Hangzhou He" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": ", Yanye Lu" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{1,2,3*}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": "Institute of Medical Technology, Peking University Health Science Center, Peking University \n" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": "Biomedical Engineering Department, College of Future Technology, Peking University \n" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": "National Biomedical Imaging Center, Peking University \n" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 78, + 160, + 532, + 230 + ], + "type": "text", + "content": "China Mobile Research Institute" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 189, + 234, + 418, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 189, + 234, + 418, + 245 + ], + "spans": [ + { + "bbox": [ + 189, + 234, + 418, + 245 + ], + "type": "text", + "content": "jkhu29@stu.pku.edu.cn, yanye.lu@pu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "spans": [ + { + "bbox": [ + 152, + 273, + 200, + 285 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 297, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 585 + ], + "type": "text", + "content": "Translation equivariance is a fundamental inductive bias in image restoration, ensuring that translated inputs produce translated outputs. Attention mechanisms in modern restoration transformers undermine this property, adversely impacting both training convergence and generalization. To alleviate this issue, we propose two key strategies for incorporating translation equivariance: slide indexing and component stacking. Slide indexing maintains operator responses at fixed positions, with sliding window attention being a notable example, while component stacking enables the arrangement of translation-equivariant operators in parallel or sequentially, thereby building complex architectures while preserving translation equivariance. However, these strategies still create a dilemma in model design between the high computational cost of self-attention and the fixed receptive field associated with sliding window attention. To address this, we develop an adaptive sliding indexing mechanism to efficiently select key-value pairs for each query, which are then concatenated in parallel with globally aggregated key-value pairs. The designed network, called the Translation Equivalence Adaptive Transformer (TEAFormer), is assessed across a variety of image restoration tasks. The results highlight its superiority in terms of effectiveness, training convergence, and generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 135, + 605 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 296, + 696 + ], + "type": "text", + "content": "Image restoration aims to reconstruct high-quality (HQ) images from their low-quality (LQ) counterparts, which are often affected by degradations resulting from the imaging system or environment. As a constituent of the imaging system, image restoration methods should satisfy the following fidelity property: the restoration results for the target region should remain equivariant under geometric transfor-" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 274, + 555, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 310 + ], + "type": "text", + "content": "mations of the input. This fidelity property is termed translation equivariance, which is one of the inherent inductive biases in image restoration." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 317, + 357, + 367 + ], + "blocks": [ + { + "bbox": [ + 317, + 317, + 357, + 367 + ], + "lines": [ + { + "bbox": [ + 317, + 317, + 357, + 367 + ], + "spans": [ + { + "bbox": [ + 317, + 317, + 357, + 367 + ], + "type": "image", + "image_path": "cdfa586a084a50a3c2d482ef86180369c3b78e0a3c76f38347db32ba256bbb6e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 361, + 318, + 403, + 366 + ], + "blocks": [ + { + "bbox": [ + 361, + 318, + 403, + 366 + ], + "lines": [ + { + "bbox": [ + 361, + 318, + 403, + 366 + ], + "spans": [ + { + "bbox": [ + 361, + 318, + 403, + 366 + ], + "type": "image", + "image_path": "ee6cd240c1cf4f8aaff89f54d7d19f1c6cb10270ebb3df8f46fe4852cc26134a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 408, + 318, + 449, + 366 + ], + "blocks": [ + { + "bbox": [ + 408, + 318, + 449, + 366 + ], + "lines": [ + { + "bbox": [ + 408, + 318, + 449, + 366 + ], + "spans": [ + { + "bbox": [ + 408, + 318, + 449, + 366 + ], + "type": "image", + "image_path": "751f7cbbb4f62e1fdce47b7731e7cad2c80493e55ae46a1fa2e299befad9dadd.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 367, + 358, + 416 + ], + "blocks": [ + { + "bbox": [ + 317, + 367, + 358, + 416 + ], + "lines": [ + { + "bbox": [ + 317, + 367, + 358, + 416 + ], + "spans": [ + { + "bbox": [ + 317, + 367, + 358, + 416 + ], + "type": "image", + "image_path": "cf0a4a3481b6cba5998820aa47ad0a7839d26b8a71ffd1fa410e6a48c6b105cf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "type": "text", + "content": "Figure 1. " + }, + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 421, + 555, + 487 + ], + "type": "text", + "content": " means the translation function. \"SA\" and \"WA\" are commonly used for self-attention and window attention, respectively, but disrupt translation equivariance (TE) due to position encoding and feature shifting. \"TEA\" is our proposed translation equivariance adaptive attention, which satisfies TE. TE promotes faster convergence and better generalization." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 361, + 367, + 402, + 417 + ], + "blocks": [ + { + "bbox": [ + 361, + 367, + 402, + 417 + ], + "lines": [ + { + "bbox": [ + 361, + 367, + 402, + 417 + ], + "spans": [ + { + "bbox": [ + 361, + 367, + 402, + 417 + ], + "type": "image", + "image_path": "da6343229e66edb0dd5ff2a569869cbaad7a89debfbf41c46a8a2df35698dbe0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 408, + 367, + 449, + 418 + ], + "blocks": [ + { + "bbox": [ + 408, + 367, + 449, + 418 + ], + "lines": [ + { + "bbox": [ + 408, + 367, + 449, + 418 + ], + "spans": [ + { + "bbox": [ + 408, + 367, + 449, + 418 + ], + "type": "image", + "image_path": "a8d13e64892278fc66fdd464a8c3aa86cd7b0ae570742405a579726918a8348b.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 453, + 315, + 548, + 366 + ], + "blocks": [ + { + "bbox": [ + 453, + 315, + 548, + 366 + ], + "lines": [ + { + "bbox": [ + 453, + 315, + 548, + 366 + ], + "spans": [ + { + "bbox": [ + 453, + 315, + 548, + 366 + ], + "type": "image", + "image_path": "408acbc476c84a68cb551091348986afecde47974edc227cdc3008eded1723a6.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 453, + 366, + 548, + 418 + ], + "blocks": [ + { + "bbox": [ + 453, + 366, + 548, + 418 + ], + "lines": [ + { + "bbox": [ + 453, + 366, + 548, + 418 + ], + "spans": [ + { + "bbox": [ + 453, + 366, + 548, + 418 + ], + "type": "image", + "image_path": "c391195427018b5c444fad6e84337353ae4342c12105840af78cd4294df64b2d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 317, + 491, + 559, + 546 + ], + "blocks": [ + { + "bbox": [ + 317, + 491, + 559, + 546 + ], + "lines": [ + { + "bbox": [ + 317, + 491, + 559, + 546 + ], + "spans": [ + { + "bbox": [ + 317, + 491, + 559, + 546 + ], + "type": "table", + "html": "
AttentionsTEPerformanceNTK [25]SRGA [33]Complexity
PSNR↑/ SSIM↑Condition ↓Value ↓
SAX28.21 / 0.843511140.444.075O(N2)
WAX27.45 / 0.82541746.493.655O(N)
TEA (Ours)✓28.67 / 0.8489236.783.275O(N)
", + "image_path": "9df326fe6c5a3d9446564755d8e3ef1fb80dd564e033591237a240316873fbdf.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 548, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 555, + 602 + ], + "type": "text", + "content": "Table 1. In " + }, + { + "bbox": [ + 313, + 548, + 555, + 602 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 313, + 548, + 555, + 602 + ], + "type": "text", + "content": " image super-resolution, we study performance, NTK [8] convergence speed, SRGA [33] generalization, and computation complexities. We replace WA in SwinIR with SA and TEA for testing. TEA shows fast convergence, better generalization, and linear complexity. The results are tested on Urban100." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 607, + 555, + 667 + ], + "type": "text", + "content": "In restoration networks [32, 62, 68, 71] based on convolution neural networks (CNNs), the paradigm of sliding information extraction inherent to convolutions [18] imparted a natural translation equivariance (TE) property in restoration networks. However, as shown in Figure 1 and Table 1," + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 313, + 674, + 556, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 556, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 674, + 556, + 713 + ], + "type": "text", + "content": "Translation equivariance and translation invariance are concepts that are frequently misconstrued. Our paper is dedicated to the analysis of equivariance. We have provided a clear delimitation and discussion concerning these concepts in the supplementary material." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "footer", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16047" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 119 + ], + "type": "text", + "content": "the advent of Self Attention (SA) [51] and Window Attention (WA) [34] within Transformers has broken this fidelity property, leading to slow convergence and poor generalization in restoration transformers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 263 + ], + "type": "text", + "content": "In this paper, we propose two fundamental strategies to bring TE back into the restoration transformers: (1) slide indexing and (2) component stacking. The slide indexing strategy employs a fixed container, e.g., a fixed rectangular window, to globally extract the input in a sliding fashion. This method requires the operators to calculate instantaneously, thereby maintaining the consistency of positional coordinates between the input and output of the model, which facilitates the integration of TE. The component stacking strategy involves stacking modules that satisfy the TE property in parallel or in series, thus enhancing architectures while ensuring TE in complex networks." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 264, + 295, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 295, + 443 + ], + "type": "text", + "content": "However, these foundational strategies present a persistent challenge in the design of restoration transformers. Selecting SA to achieve superior performance involves accepting its substantial computational complexity [63]; alternatively, opting for sliding window attention provides linear complexity but results in a fixed receptive field [48] and reduced performance. To break this dilemma, we replace the global key-value pair indexing mechanism in vanilla SA with an adaptive slide indexing mechanism to design the Translation Equivariance Adaptive (TEA) attention. To retain the extraction of global information, TEA incorporates a downsampling self-attention branch without position encoding, which reduces the key-value pair's resolution by aggregating them, thereby providing TEA with coarse, yet effective global information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 295, + 598 + ], + "type": "text", + "content": "In summary, our contributions are threefold. First, we investigate TE in image restoration, analyzing its impact on convergence and generalization. The results indicate that the integration of TE is essential in the design of restoration networks. Second, we propose TEA, which uses adaptive slide indexing methods to incorporate the TE property while preserving linear complexity and ensuring a global receptive field, thereby enhancing restoration transformers. Third, we introduce TEA to the vanilla Vision Transformer to design TEAFormer. Experiments show that TEA achieves State-of-The-Art (SoTA) performance in various image restoration tasks, while having faster convergence and more robust generalization." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 142, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 142, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 142, + 623 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Vision Transformers with inductive bias. Following the pioneering introduction of the Transformer [51] into vision tasks by ViT [16], SA has emerged as a key component in vision networks. However, ViT is challenging to train, exhibits slow convergence, and requires substantial data support [49]. Previous works, such as ViTAE [59, 70], attributed this to a lack of inductive bias and advocated in" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "tegrating inductive bias with ViT to develop more efficient ones. AACNet [4] emphasized the importance of translation equivariance for visual tasks. The Swin [34] and CSwin [15] Transformers incorporated local processing into the vision Transformer, thus proposing window attention. CVT [56] and FasterViT [21] sequentially stack convolutions and attention layers, resulting in a hybrid architecture. PVT [52, 53] used average pooling for downsampling to achieve weak translation equivariance. Both DAT [57] and BiFormer [74] dynamically determined the windows accessible to each token. Recently, some studies have introduced translation equivariance into ViT by modifying position encoding [4, 58] and attention calculation logic [44, 46]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 553, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 431 + ], + "type": "text", + "content": "Recent image restoration Transformers. Similar to the development of vision transformers, the design of image restoration models also focuses on how to better design local global information processors [9, 31, 61, 63, 73]. Since the resolution of the input image is generally high in the image restoration task, directly using SA will result in a huge computation cost. To address this issue, SwinIR [31] introduced window attention [34]. HAT [9] enlarged the receptive field of the restoration transformers with the squeeze-excitation module [22] and the overlapping mechanism. Restormer [63] and DAT [10] performed channel-wise attention, thereby facilitating more efficient feature re-weighting. GRL [61] extracted the feature on local, regional, and global scales by integrating the attention of the anchor window. IPG [48] exploited the flexibility of the network by proposing the degree-variant graph solution, breaking the rigidity of the previous methods." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 460, + 455 + ], + "type": "text", + "content": "3. Translation Equivariance" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 462, + 553, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 497 + ], + "type": "text", + "content": "In this section, we define the translation equivariance and propose two fundamental strategies for integrating translation equivariance into restoration networks." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": "Definition 3.1. Translation Equivalence. We call function " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": " is translation equivariant to translation operation " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "\\Phi(\\mathcal{T}(x)) = \\mathcal{T}(\\Phi(x))" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 498, + 553, + 534 + ], + "type": "text", + "content": " is the input signal." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "text", + "content": "In image restoration tasks, " + }, + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "text", + "content": " is an image or a feature in the latent space typically, " + }, + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(\\cdot)" + }, + { + "bbox": [ + 313, + 534, + 553, + 568 + ], + "type": "text", + "content": " refers to a shift in pixels [20]. Based on this definition, two theorems are derived:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "content": "Theorem 3.2. If " + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "inline_equation", + "content": "\\Phi(x)_i" + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "content": " is transformed from " + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "inline_equation", + "content": "x_j = [i - b, i + b]" + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "content": " is the sliding boundary, the function " + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "inline_equation", + "content": "\\Phi(\\cdot)" + }, + { + "bbox": [ + 313, + 569, + 553, + 604 + ], + "type": "text", + "content": " fulfills Definition 3.1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": "Theorem 3.3. If the functions " + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Phi_1(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Phi_2(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": " fulfill Definition 3.1, the functions " + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Phi_2(\\Phi_1(\\cdot))" + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Phi_1(\\cdot) + \\Phi_2(\\cdot)" + }, + { + "bbox": [ + 313, + 605, + 553, + 640 + ], + "type": "text", + "content": " also fulfill Definition 3.1." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 713 + ], + "type": "text", + "content": "Reformulating the existing operators in line with Theorem 3.2 gives them the translation equivariance property. By stacking these operators, either in parallel or in serial, as specified in Theorem 3.3, more complex architectures are also capable of translation equivariance. Consequently, the model can be further enhanced via translation equivariance." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16048" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 72, + 296, + 152 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 152 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 152 + ], + "type": "image", + "image_path": "d39d1908a1a44d458737227015e082d0cd52a96a1591ab640e3aa5928d0a8ea1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 160, + 290, + 172 + ], + "lines": [ + { + "bbox": [ + 59, + 160, + 290, + 172 + ], + "spans": [ + { + "bbox": [ + 59, + 160, + 290, + 172 + ], + "type": "text", + "content": "Figure 2. The overall architecture of our proposed TEAFormer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 111, + 190 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 258 + ], + "type": "text", + "content": "In this section, we conform to Theorems 3.1 and 3.2 to devise a translation equivariance restoration transformer, termed Translation Equivariance trAnsFormer (TEAFormer). We first present the overall architecture of TEAFormer and then introduce its components in detail." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 267, + 176, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 176, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 176, + 278 + ], + "type": "text", + "content": "4.1. Overall Architecture" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": "The overall architecture of the proposed TEAFormer is illustrated in Figure 2 with image super-resolution (SR) presented as a demonstration. Given an input image " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " denotes the resolution, TEAFormer first applies a convolution layer to obtain a shallow feature " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "F_{0} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " is the embedding dimension of the network. Next, following classical restoration networks [31], TEAFormer adopts a residual-in-residual structure to construct a deep feature extraction module comprised of " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "N_{g}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " Translation Equivariance Groups (TEGs). Each TEG consists of " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "N_{b}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " Translation Equivariance Blocks (TEBs) and a convolution layer. Each TEB, in turn, is composed of one Translation Equivariance Attention (TEA) and one feed-forward network. Subsequently, a convolutional layer is employed to extract the deep feature " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "F_{1} \\in \\mathbb{R}^{H \\times W \\times D}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " from the feature extracted by TEGs. After extraction through the deep feature extraction module, we use the restore module to generate the high-quality image " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": " from the feature " + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "inline_equation", + "content": "F_{r} = F_{0} + F_{1}" + }, + { + "bbox": [ + 55, + 285, + 296, + 537 + ], + "type": "text", + "content": ". As for other image restoration tasks that do not involve changes in resolution, we build our TEAFormer following Restormer [63]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 243, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 243, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 243, + 558 + ], + "type": "text", + "content": "4.2. Adaptive Translation Equivariance" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 563, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 598 + ], + "type": "text", + "content": "Adaptive Translation Equivalence is integrated through the Adaptive Sliding Key-Value Self Attention (ASkvSA) and Downsampled Self Attention (DSA) as shown in Figure 3." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "text", + "content": "Key-value indexing in self-attention. The nature of Self Attention (SA) without position bias lies in the indexing of key-value pairs, followed by feature re-weighting. Given a flattened input " + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "inline_equation", + "content": "X \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "text", + "content": " represents the number of tokens and " + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 599, + 296, + 670 + ], + "type": "text", + "content": " represents the embedding dimension, self-attention can be formulated as following:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 86, + 677, + 295, + 703 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 677, + 295, + 703 + ], + "spans": [ + { + "bbox": [ + 86, + 677, + 295, + 703 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} (X) = \\operatorname {S o f t m a x} \\left(\\frac {X W _ {q} \\left(X W _ {k}\\right) ^ {\\top}}{\\sqrt {D}}\\right) X W _ {v}, \\tag {1}", + "image_path": "44adabf194fdf817395c874247c63e6e3c07c94e6e4b65199dec0bd1c48077e0.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{q}, W_{k}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{v}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " denote the weights of the linear" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 553, + 247 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 553, + 247 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 553, + 247 + ], + "type": "image", + "image_path": "e6c10ae9c6e10be075614c0c5d9a19113754516380df00a1cf2d3bcc1ac65d6a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 554, + 284 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 284 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 284 + ], + "type": "text", + "content": "Figure 3. The inner structure of TEA with two components. (a) Adaptive Sliding key-value Self Attention (ASkvSA) and (b) Downsampled Self Attention (DSA)." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": "projections utilized to compute the query " + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "inline_equation", + "content": "(Q)" + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "inline_equation", + "content": "(K)" + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "inline_equation", + "content": "(V)" + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": " from input " + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 289, + 554, + 314 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "text", + "content": "In Eq. 1, the softmax similarity function is utilized for the generation of attention maps between a query and a key, while the dot product is employed for re-weighting between the attention map and a value. Given that subscripting a matrix with " + }, + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "text", + "content": " returns the " + }, + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 315, + 554, + 387 + ], + "type": "text", + "content": "-th row as a vector, an alternative equation of self-attention can be written as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 363, + 398, + 553, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 398, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 363, + 398, + 553, + 433 + ], + "type": "interline_equation", + "content": "\\operatorname {S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = 1} ^ {N} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {2}", + "image_path": "fa856399ae6943422c27bc218fd3feec904d6920e0ee91e3ac9e7276ea9d0398.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 439, + 460, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 439, + 460, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 439, + 460, + 458 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 439, + 460, + 458 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(Q_i, K_j) = \\exp \\left(\\frac{Q_i^\\top K_j}{\\sqrt{D}}\\right)" + }, + { + "bbox": [ + 313, + 439, + 460, + 458 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "text", + "content": "Eq. 2 illustrates that self-attention indexes key-value pairs " + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "inline_equation", + "content": "(K_{j}" + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "inline_equation", + "content": "V_{j}, j = 1, \\dots, N)" + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "text", + "content": " for each query " + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "inline_equation", + "content": "(Q_{i})" + }, + { + "bbox": [ + 313, + 458, + 554, + 482 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 482, + 554, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 482, + 554, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 482, + 554, + 542 + ], + "type": "text", + "content": "Sliding key-value Self Attention. In accordance with Theorem 3.2, we demonstrate that sliding window attention constitutes a generalized form of self-attention, which can be designated as Sliding Key-Value Self-Attention (SkvSA). SkvSA is formulated as follows." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 324, + 548, + 553, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 548, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 324, + 548, + 553, + 585 + ], + "type": "interline_equation", + "content": "\\operatorname {S k v S A} _ {i} (X) = \\frac {\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right) V _ {j}}{\\sum_ {j = (i - b _ {1}) \\bmod s = 0} ^ {(i + b _ {2}) \\bmod s = 0} \\operatorname {s i m} \\left(Q _ {i} , K _ {j}\\right)}, \\tag {3}", + "image_path": "a68f2ac9c45a85e9c0cf650d0171ca4c9218b4523104c83dd8a4858e69ea4ba0.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "b_{1}" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "b_{2}" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": " are the sliding boundary on the left and right when sliding on the one-dimensional vector " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": ". In our SkvSA, " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "b_{1} = b_{2} = \\frac{w\\cdot s}{2}" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": " is the size of the sliding window and " + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 584, + 554, + 632 + ], + "type": "text", + "content": " is the stride size during sliding." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": "Eq. 3 says that, for the " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": "-th query " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": ", SkvSA systematically extracts the key-value pairs from a designated sliding window, extending from the index " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "i - \\frac{w \\cdot s}{2}" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": " to the index " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "i + \\frac{w \\cdot s}{2}" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": " with stride " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "s^2" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": ". Subsequently, SkvSA calculates the attention map and re-weights the feature in place." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "type": "text", + "content": "2We set " + }, + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "type": "inline_equation", + "content": "w\\cdot s\\leq N" + }, + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "type": "text", + "content": " in experiments thus enabling SkvSA to slide on " + }, + { + "bbox": [ + 324, + 702, + 553, + 713 + ], + "type": "inline_equation", + "content": "X" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16049" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "Boundary processing. Note that, in scenarios where " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "i - \\frac{w \\cdot s}{2} < 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "i + \\frac{w \\cdot s}{2} > N" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " occur, the slide indices of the key-value pairs " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " will exceed the boundary of the input " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", leading to an indexing failure. To ameliorate the issue associated with slide indexing at sequence's boundaries, we propose a boundary blocking indexing method. Specifically, for queries located at the boundary of the sequence, the sliding indices of key-value pairs are \"blocked\" and converted into indexing within a predefined blocking window. In this study, the sizes of the predefined blocking window and the sliding window are identical. Accordingly, the sliding boundaries " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " can be appropriately revised as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 217, + 296, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 217, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 67, + 217, + 296, + 281 + ], + "type": "interline_equation", + "content": "b = \\left(b _ {1}, b _ {2}\\right) = \\left\\{ \\begin{array}{l l} \\left(0, \\frac {w \\cdot s}{2}\\right), & i - \\frac {w \\cdot s}{2} < 0, \\\\ \\left(\\frac {w \\cdot s}{2}, N\\right), & i + \\frac {w \\cdot s}{2} > N, \\\\ \\left(\\frac {w \\cdot s}{2}, \\frac {w \\cdot s}{2}\\right), & \\text {o t h e r s .} \\end{array} \\right. \\tag {4}", + "image_path": "0734fafc619dbf0f041da27d2093c3bc1e2b46ea1b5fd4ef039cd67fb969672c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "spans": [ + { + "bbox": [ + 55, + 281, + 296, + 328 + ], + "type": "text", + "content": "The SkvSA and its boundary processing are capable of extracting and processing information via a fixed sliding window, adhering to the translation equivariance property as established in Theorem 3.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "content": "Adaptive Sliding key-value Self Attention. However, the fixed window size constrains the receptive field, leading to the entrapment of restoration transformers within rigidity [48]. To mitigate this rigidity, it is anticipated that the SkvSA will ascertain the optimal window size and position specific to each query " + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "content": ". To achieve this endeavor, we devise the Adaptive Sliding key-value Self Attention (ASkvSA), which adaptively determines key-value indices and shuffles them within the fixed sliding window of " + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 330, + 296, + 437 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": "As shown in Figure 3 (a), with regard to Key " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " and Value " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": ", ASkvSA initially reshapes " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " from one-dimensional (1d) vectors into two-dimensional (2d) matrices. Subsequently, it applies a 2d depth-wise convolution with kernel size " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " to generate their adaptive indices " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F} \\in \\mathbb{R}^{H \\times W \\times 2}" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}[h, w]" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " is a vector of length 2 in the " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "h-" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": "th row and " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "w-" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": "th column of " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": ", which can describe a coordinate. Following this, ASkvSA performs key-value indexing based on established indices " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_K" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_V" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " by shuffling the pixel located at the coordinate " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "\\mathcal{F}[h, w]" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " to the pixel at the coordinate " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "(h, w)" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 438, + 296, + 570 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 576, + 295, + 606 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 576, + 295, + 606 + ], + "spans": [ + { + "bbox": [ + 67, + 576, + 295, + 606 + ], + "type": "interline_equation", + "content": "\\operatorname {A S k v S A} (X) = \\frac {\\sum_ {j = \\mathcal {F} (K) , l = \\mathcal {F} (V)} \\sin \\left(Q _ {i} , K _ {j}\\right) V _ {l}}{\\sum_ {j = \\mathcal {F} (K)} \\sin \\left(Q _ {i} , K _ {j}\\right)}. \\tag {5}", + "image_path": "268edb047b6feb82ab42ebcd87b56d0efa4b64e5bc27972199e23b861a9400c8.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": "Adaptive slide indexing does not compromise the translation equivariance of ASkvSA. ASkvSA stacks convolutions with SkvSA in serial to generate adaptive indices of key-value pairs. Given that convolution satisfies the property of translation equivariance, when " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{h,w}" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " is translated to " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "K_{h + \\delta_h, w + \\delta_w}" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": ", key's adaptive indices " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_K" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " will shuffle the pixel located at " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}[h + \\delta_h, w + \\delta_w]" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " to the pixel at " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "(h + \\delta_h, w + \\delta_w)" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 606, + 296, + 714 + ], + "type": "text", + "content": ". According to Theorem 3.3, ASkvSA satisfies the translation equivariance property." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 167 + ], + "type": "text", + "content": "Adaptively aggregated with global key-value pairs. Although ASkvSA can identify appropriate key-value pairs beyond a fixed sliding window, some remote but valid pixels still escape ASkvSA's indexing. Consequently, we propose downsampled self-attention (DSA), which adaptively combines global key value indexes through downsampling, providing the network with a coarse, yet efficient global indexing of key-value pairs. DSA can be formulated as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 359, + 169, + 555, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 169, + 555, + 202 + ], + "spans": [ + { + "bbox": [ + 359, + 169, + 555, + 202 + ], + "type": "interline_equation", + "content": "\\mathrm {D S A} _ {i} (X) = \\frac {\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right) V _ {j} ^ {\\prime}}{\\sum_ {j = 1} ^ {N _ {d}} \\operatorname {s i m} \\left(Q _ {i} , K _ {j} ^ {\\prime}\\right)}, \\tag {6}", + "image_path": "1f8dc95fb9d21000ad6c9b60f81dca3c5a4a4cef7ef84150af00fddaa96ea2a0.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "inline_equation", + "content": "K'" + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "inline_equation", + "content": "V'" + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "text", + "content": " represents the downsampled key-value pair calculated in ASkvSA, and " + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "inline_equation", + "content": "N_d" + }, + { + "bbox": [ + 313, + 203, + 555, + 262 + ], + "type": "text", + "content": " denotes the number of tokens associated with the downsampled key-value pair. We employ average pooling as the downsampling operator due to its efficiency and coarse fulfillment of TE." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 262, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 287 + ], + "type": "text", + "content": "Finally, according to Figure 3, the ASkvSA and DSA outputs are adaptively aggregated to obtain our TEA." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 296, + 554, + 345 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 296, + 554, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 296, + 554, + 345 + ], + "type": "interline_equation", + "content": "\\mathrm {T E A} _ {i} (X) = \\alpha_ {s} \\operatorname {A S k v S A} (X) + \\alpha_ {d} \\operatorname {D S A} (X), \\quad (7) \\text {w h e r e} \\alpha_ {s} \\text {a n d} \\alpha_ {d} \\text {a r e l a r n a b l e p a r a m e t e r s . T E A s t a c k s D S A w i t h A S k v S A i n p a r a l l e l a r . A c c o r d i n g t o T h e o r e m 3 . 3 , T E A s a t s i s f i e s t h e t r a n s l a t i o n e q u i v a r i a n c e p r o p e r t y .}", + "image_path": "688dbe499055362351d45434eb455cc986a2d51aeb07a804de9294c01561deea.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 352, + 542, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 542, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 542, + 365 + ], + "type": "text", + "content": "4.3. Computational Cost and Hyper-Parameters" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 418 + ], + "type": "text", + "content": "We demonstrate that TEA can maintain a computational cost that is similar to that of window attention [34] under well-designed hyperparameter settings. We first list four integral hyperparameters of TEA:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 420, + 521, + 467 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 420, + 485, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 485, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 485, + 431 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 313, + 420, + 485, + 431 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 420, + 485, + 431 + ], + "type": "text", + "content": " : the sliding window size of ASkvSA;" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 432, + 473, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 473, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 473, + 443 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 313, + 432, + 473, + 443 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 432, + 473, + 443 + ], + "type": "text", + "content": ": the sliding stride size of ASkvSA;" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 444, + 521, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 521, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 521, + 455 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 313, + 444, + 521, + 455 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 444, + 521, + 455 + ], + "type": "text", + "content": " : the kernel size of adaptive indexes generation;" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 456, + 463, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 463, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 463, + 467 + ], + "type": "text", + "content": "- " + }, + { + "bbox": [ + 313, + 456, + 463, + 467 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 313, + 456, + 463, + 467 + ], + "type": "text", + "content": ": the number of tokens in DSA." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "spans": [ + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": "According to Section 4.2, TEA uses " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "3ND^{2}" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " FLOPs for the Query-Key-Value linear projections, " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "2NDk^{2}" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " FLOPs for the depth-wise convolution in adaptive indices generation, " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "Nw^{2}D" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " FLOPs for attention map computation between query " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " and indexed key " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "K \\in \\mathbb{R}^{w^{2} \\times D}" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "Nw^{2}D" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " FLOPs for feature re-weighting between attention map and indexed value " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{w^{2} \\times D}" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "2NN_{d}D" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": " FLOPs for DSA. The total FLOPs of TEA is " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "3ND^{2} + 2NDk^{2} + 2Nw^{2}D + 2NN_{d}D" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": ". With the hyper-parameters held constant, the FLOPs of TEA exhibit linear growth as a function of " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": ", thereby implying that TEA is " + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(N)" + }, + { + "bbox": [ + 313, + 469, + 554, + 603 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "content": "Consequently, when the hyper-parameters of TEA are adjusted carefully, its computational cost becomes comparable to that of window attention. In our TEAFormer, the hyper-parameter " + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "content": " has been set to 15, " + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "content": " to 3, and " + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "inline_equation", + "content": "N_{d}" + }, + { + "bbox": [ + 313, + 603, + 554, + 687 + ], + "type": "text", + "content": " to 16, thereby modulating the computational complexity of TEA to be even slightly lower than that of window attention with window size 16." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 693, + 554, + 713 + ], + "type": "text", + "content": "[45] proved that downsampling can meet strict TE through learnable polyphase (LP). We compare these downsamplers in Sec. 5.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16050" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 71, + 524, + 166 + ], + "blocks": [ + { + "bbox": [ + 85, + 71, + 524, + 166 + ], + "lines": [ + { + "bbox": [ + 85, + 71, + 524, + 166 + ], + "spans": [ + { + "bbox": [ + 85, + 71, + 524, + 166 + ], + "type": "table", + "html": "
Modelwindow sizeTranslation equivarianceParamsFLOPsLatencyPerformanceNTK [25]SRGA [33]
SkvSAASkvSADSA(M)(T)(ms)Urban100 [24] PSNR↑ / SSIM↑Condition ↓Value ↓
SwinIR [31]811.900.462130.027.45 / 0.82541746.493.655
SwinIR-Large1621.420.897214.527.94 / 0.83621554.653.610
TEAFformer15✓21.860.688230.128.31 / 0.8444243.753.206
TEAFformer15✓21.980.757284.428.47 / 0.8457283.993.298
TEAFformer15✓✓21.860.966340.928.49 / 0.8470203.063.261
TEAFformer15✓✓21.981.035386.728.67 / 0.8489236.783.275
", + "image_path": "64b0051396129607c276140ed5a0c3c152fa7e7876e72dc22a88c18e541b2150.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 188, + 563, + 437 + ], + "blocks": [ + { + "bbox": [ + 113, + 167, + 496, + 177 + ], + "lines": [ + { + "bbox": [ + 113, + 167, + 496, + 177 + ], + "spans": [ + { + "bbox": [ + 113, + 167, + 496, + 177 + ], + "type": "text", + "content": "Table 2. Ablation study on TEA. Our TEA enhances performance, convergence speed, and generalization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 188, + 563, + 437 + ], + "lines": [ + { + "bbox": [ + 58, + 188, + 563, + 437 + ], + "spans": [ + { + "bbox": [ + 58, + 188, + 563, + 437 + ], + "type": "table", + "html": "
MethodScaleParams (M)FLOPs (T)Latency (ms)Set5 [5]Set14 [64]B100 [37]Urban100 [24]Manga109 [40]
PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑PSNR ↑SSIM ↑
EDSR [32]40.76.006147.138.110.960233.920.919532.320.901332.930.935139.100.9773
RCAN [71]15.42.259155.538.270.961434.120.921632.410.902733.340.938439.440.9786
SwinIR-L [31]0.91156.4188.638.140.961133.860.920632.310.901232.760.934039.120.9783
SwinIR [31]11.81.848514.438.420.962334.460.925032.530.904133.810.942739.920.9797
EDT [30]11.51.965101038.450.962434.570.925832.520.904133.800.942539.930.9800
SRFormer [73]×210.31.790787.238.510.962734.440.925332.570.904634.090.944940.070.9802
HAT [9]20.63.662901.938.630.963034.860.927432.620.905334.450.946640.260.9809
GRL-B* [61]20.14.422262138.480.962734.640.926532.550.904533.970.943740.060.9804
DAT [10]14.72.155121538.580.962934.810.927232.610.905134.370.945840.330.9807
IPG [48]16.84.732270338.610.963234.730.927032.600.905234.480.946440.240.9810
TEAFormer-L (Ours)0.830.237254.038.270.861834.260.924032.430.902933.510.939639.560.9792
TEAFormer (Ours)21.84.133149338.640.963135.120.928432.720.906435.230.950940.320.9807
EDSR [32]43.11.85349.1332.460.896828.800.787627.710.742026.640.803331.020.9148
RCAN [71]15.60.58746.1532.630.900228.870.788927.770.743626.820.808731.220.9173
SwinIR-L [31]0.880.03948.7132.440.897628.770.785827.690.740626.470.798030.920.9151
SwinIR [31]11.90.462130.032.920.904429.090.795027.920.748927.450.825432.030.9260
EDT [30]11.60.514251.332.820.903129.090.793927.910.748327.460.824632.030.9254
SRFormer [73]×410.40.470189.232.930.904129.080.795327.940.750227.680.831132.440.9271
HAT [9]20.80.938227.733.040.905629.230.797328.000.751727.970.836832.480.9292
GRL-B* [61]20.21.129761.832.900.903929.140.795627.960.749727.530.827632.190.9266
DAT [10]14.80.561296.833.080.905529.230.797328.000.751527.870.834332.510.9291
IPG [48]17.01.206646.733.150.906229.240.797327.990.751928.130.839232.530.9300
TEAFormer-L (Ours)0.830.05969.6432.630.900228.960.789927.820.744227.040.810531.560.9197
TEAFormer (Ours)22.01.035386.733.230.906629.460.800928.110.754428.670.848932.990.9323
", + "image_path": "dbcd9de97f5c6c5a169c450276b15adf9fd59e7e6d66a3405239969a302eba04.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 555, + 472 + ], + "lines": [ + { + "bbox": [ + 55, + 439, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 555, + 472 + ], + "type": "text", + "content": "Table 3. Classical image SR results. Performance and complexity are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two scores in each column are highlighted in red and blue. Methods with “*” are replicated with standard training setting, following IPG [48]. The “-L” indicates that the given method is a lightweight version." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 69, + 483, + 228, + 571 + ], + "blocks": [ + { + "bbox": [ + 69, + 483, + 228, + 571 + ], + "lines": [ + { + "bbox": [ + 69, + 483, + 228, + 571 + ], + "spans": [ + { + "bbox": [ + 69, + 483, + 228, + 571 + ], + "type": "image", + "image_path": "6c7067faf5c2b46ff3b533a01d757f0cb5961cc2b2d1bc6dbf7286c7317ba969.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 573, + 195, + 582 + ], + "lines": [ + { + "bbox": [ + 104, + 573, + 195, + 582 + ], + "spans": [ + { + "bbox": [ + 104, + 573, + 195, + 582 + ], + "type": "text", + "content": "BSD100 " + }, + { + "bbox": [ + 104, + 573, + 195, + 582 + ], + "type": "inline_equation", + "content": "(4\\times)" + }, + { + "bbox": [ + 104, + 573, + 195, + 582 + ], + "type": "text", + "content": " :img_86016" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 231, + 483, + 310, + 521 + ], + "blocks": [ + { + "bbox": [ + 231, + 483, + 310, + 521 + ], + "lines": [ + { + "bbox": [ + 231, + 483, + 310, + 521 + ], + "spans": [ + { + "bbox": [ + 231, + 483, + 310, + 521 + ], + "type": "image", + "image_path": "840872a5dd68612db0b0565822aa3a0a03c1daf2f1af968502b703786197ec8d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 257, + 523, + 285, + 532 + ], + "lines": [ + { + "bbox": [ + 257, + 523, + 285, + 532 + ], + "spans": [ + { + "bbox": [ + 257, + 523, + 285, + 532 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 310, + 483, + 389, + 521 + ], + "blocks": [ + { + "bbox": [ + 310, + 483, + 389, + 521 + ], + "lines": [ + { + "bbox": [ + 310, + 483, + 389, + 521 + ], + "spans": [ + { + "bbox": [ + 310, + 483, + 389, + 521 + ], + "type": "image", + "image_path": "a074904d3c0310f9e029dab75edb38e89fd1dde70d61c8b307415fc32b51da01.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 390, + 483, + 467, + 521 + ], + "blocks": [ + { + "bbox": [ + 390, + 483, + 467, + 521 + ], + "lines": [ + { + "bbox": [ + 390, + 483, + 467, + 521 + ], + "spans": [ + { + "bbox": [ + 390, + 483, + 467, + 521 + ], + "type": "image", + "image_path": "307518198a438373c3d3f5fba824ef264c65e277d8b405e63f303427e51d2395.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 468, + 483, + 547, + 521 + ], + "blocks": [ + { + "bbox": [ + 468, + 483, + 547, + 521 + ], + "lines": [ + { + "bbox": [ + 468, + 483, + 547, + 521 + ], + "spans": [ + { + "bbox": [ + 468, + 483, + 547, + 521 + ], + "type": "image", + "image_path": "c1d57f9f8802a91cb21a666fd6406dd726a5b0bcd49b17d4204fa05d06d09eeb.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 69, + 583, + 228, + 671 + ], + "blocks": [ + { + "bbox": [ + 69, + 583, + 228, + 671 + ], + "lines": [ + { + "bbox": [ + 69, + 583, + 228, + 671 + ], + "spans": [ + { + "bbox": [ + 69, + 583, + 228, + 671 + ], + "type": "image", + "image_path": "3118184afb85fe7f7e9c6c9b032b153a0502791a2113b5b827339d81f3e663b6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 673, + 193, + 683 + ], + "lines": [ + { + "bbox": [ + 105, + 673, + 193, + 683 + ], + "spans": [ + { + "bbox": [ + 105, + 673, + 193, + 683 + ], + "type": "text", + "content": "Urban100 " + }, + { + "bbox": [ + 105, + 673, + 193, + 683 + ], + "type": "inline_equation", + "content": "(4\\times)" + }, + { + "bbox": [ + 105, + 673, + 193, + 683 + ], + "type": "text", + "content": " : img_033" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 231, + 533, + 310, + 571 + ], + "blocks": [ + { + "bbox": [ + 231, + 533, + 310, + 571 + ], + "lines": [ + { + "bbox": [ + 231, + 533, + 310, + 571 + ], + "spans": [ + { + "bbox": [ + 231, + 533, + 310, + 571 + ], + "type": "image", + "image_path": "ec5b7b7d95ae8cd4d83052467bac870b16d7f4c6d3068194afa984a5310f9a22.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 573, + 278, + 582 + ], + "lines": [ + { + "bbox": [ + 264, + 573, + 278, + 582 + ], + "spans": [ + { + "bbox": [ + 264, + 573, + 278, + 582 + ], + "type": "text", + "content": "HQ" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 310, + 533, + 389, + 571 + ], + "blocks": [ + { + "bbox": [ + 329, + 523, + 370, + 533 + ], + "lines": [ + { + "bbox": [ + 329, + 523, + 370, + 533 + ], + "spans": [ + { + "bbox": [ + 329, + 523, + 370, + 533 + ], + "type": "text", + "content": "RCAN [71]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 533, + 389, + 571 + ], + "lines": [ + { + "bbox": [ + 310, + 533, + 389, + 571 + ], + "spans": [ + { + "bbox": [ + 310, + 533, + 389, + 571 + ], + "type": "image", + "image_path": "ad1d99216bd1f3a2bc3bc267b173c9b9eb2d86f04b3b223c383297d283b94b45.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 390, + 533, + 467, + 571 + ], + "blocks": [ + { + "bbox": [ + 406, + 523, + 450, + 533 + ], + "lines": [ + { + "bbox": [ + 406, + 523, + 450, + 533 + ], + "spans": [ + { + "bbox": [ + 406, + 523, + 450, + 533 + ], + "type": "text", + "content": "SwinIR [31]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 533, + 467, + 571 + ], + "lines": [ + { + "bbox": [ + 390, + 533, + 467, + 571 + ], + "spans": [ + { + "bbox": [ + 390, + 533, + 467, + 571 + ], + "type": "image", + "image_path": "e47ff34894efd7c4063afe023e1eb6e2c71c9c921e33ae87b966fdfe49565a52.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 468, + 533, + 547, + 571 + ], + "blocks": [ + { + "bbox": [ + 493, + 523, + 522, + 533 + ], + "lines": [ + { + "bbox": [ + 493, + 523, + 522, + 533 + ], + "spans": [ + { + "bbox": [ + 493, + 523, + 522, + 533 + ], + "type": "text", + "content": "HAT [9]" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 533, + 547, + 571 + ], + "lines": [ + { + "bbox": [ + 468, + 533, + 547, + 571 + ], + "spans": [ + { + "bbox": [ + 468, + 533, + 547, + 571 + ], + "type": "image", + "image_path": "1353974588e7314d074f12197e44c1956eba46dfb7ad889f5b78f9c5bfc9d1ad.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 232, + 583, + 310, + 621 + ], + "blocks": [ + { + "bbox": [ + 232, + 583, + 310, + 621 + ], + "lines": [ + { + "bbox": [ + 232, + 583, + 310, + 621 + ], + "spans": [ + { + "bbox": [ + 232, + 583, + 310, + 621 + ], + "type": "image", + "image_path": "ede9013da5424406bfd1cb581208ffe1035bff4ba5cb2cac8470effb764450be.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 310, + 583, + 389, + 621 + ], + "blocks": [ + { + "bbox": [ + 323, + 573, + 376, + 582 + ], + "lines": [ + { + "bbox": [ + 323, + 573, + 376, + 582 + ], + "spans": [ + { + "bbox": [ + 323, + 573, + 376, + 582 + ], + "type": "text", + "content": "SRFormer [73]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 583, + 389, + 621 + ], + "lines": [ + { + "bbox": [ + 310, + 583, + 389, + 621 + ], + "spans": [ + { + "bbox": [ + 310, + 583, + 389, + 621 + ], + "type": "image", + "image_path": "69c463004db8e8a1705f04e74eda412ad0d36d07a2125ef0e0757e4099bcdb9a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 390, + 583, + 467, + 621 + ], + "blocks": [ + { + "bbox": [ + 413, + 573, + 444, + 582 + ], + "lines": [ + { + "bbox": [ + 413, + 573, + 444, + 582 + ], + "spans": [ + { + "bbox": [ + 413, + 573, + 444, + 582 + ], + "type": "text", + "content": "IPG [48]" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 583, + 467, + 621 + ], + "lines": [ + { + "bbox": [ + 390, + 583, + 467, + 621 + ], + "spans": [ + { + "bbox": [ + 390, + 583, + 467, + 621 + ], + "type": "image", + "image_path": "679fa1161e951e2f67436f92d5fc7aaf7c52e7a89e2ca1ea7a670cfccc162c55.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 468, + 583, + 547, + 621 + ], + "blocks": [ + { + "bbox": [ + 471, + 573, + 544, + 582 + ], + "lines": [ + { + "bbox": [ + 471, + 573, + 544, + 582 + ], + "spans": [ + { + "bbox": [ + 471, + 573, + 544, + 582 + ], + "type": "text", + "content": "TEAFormer (Ours)" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 583, + 547, + 621 + ], + "lines": [ + { + "bbox": [ + 468, + 583, + 547, + 621 + ], + "spans": [ + { + "bbox": [ + 468, + 583, + 547, + 621 + ], + "type": "image", + "image_path": "95586d9d0a06acbb543721aa1974282dfb39f2f47fa3cdc8eb04c1025e685100.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 232, + 633, + 310, + 671 + ], + "blocks": [ + { + "bbox": [ + 257, + 624, + 285, + 632 + ], + "lines": [ + { + "bbox": [ + 257, + 624, + 285, + 632 + ], + "spans": [ + { + "bbox": [ + 257, + 624, + 285, + 632 + ], + "type": "text", + "content": "Bicubic" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 232, + 633, + 310, + 671 + ], + "lines": [ + { + "bbox": [ + 232, + 633, + 310, + 671 + ], + "spans": [ + { + "bbox": [ + 232, + 633, + 310, + 671 + ], + "type": "image", + "image_path": "59051c683c0dd65276b62f2680811ce0bf7a606ef16ca7f15f2cb75e8b9b37c7.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 673, + 278, + 682 + ], + "lines": [ + { + "bbox": [ + 264, + 673, + 278, + 682 + ], + "spans": [ + { + "bbox": [ + 264, + 673, + 278, + 682 + ], + "type": "text", + "content": "HQ" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 310, + 633, + 389, + 671 + ], + "blocks": [ + { + "bbox": [ + 329, + 624, + 370, + 632 + ], + "lines": [ + { + "bbox": [ + 329, + 624, + 370, + 632 + ], + "spans": [ + { + "bbox": [ + 329, + 624, + 370, + 632 + ], + "type": "text", + "content": "RCAN [71]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 310, + 633, + 389, + 671 + ], + "lines": [ + { + "bbox": [ + 310, + 633, + 389, + 671 + ], + "spans": [ + { + "bbox": [ + 310, + 633, + 389, + 671 + ], + "type": "image", + "image_path": "f0d908e7afd0c624201203807a2cea98a13f5518991f55a8ef3a75c7c45578db.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 673, + 376, + 683 + ], + "lines": [ + { + "bbox": [ + 323, + 673, + 376, + 683 + ], + "spans": [ + { + "bbox": [ + 323, + 673, + 376, + 683 + ], + "type": "text", + "content": "SRFormer [73]" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 153, + 684, + 455, + 694 + ], + "lines": [ + { + "bbox": [ + 153, + 684, + 455, + 694 + ], + "spans": [ + { + "bbox": [ + 153, + 684, + 455, + 694 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison with recent SOTA methods on the image SR task." + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 390, + 633, + 467, + 671 + ], + "blocks": [ + { + "bbox": [ + 407, + 624, + 450, + 632 + ], + "lines": [ + { + "bbox": [ + 407, + 624, + 450, + 632 + ], + "spans": [ + { + "bbox": [ + 407, + 624, + 450, + 632 + ], + "type": "text", + "content": "SwinIR [31]" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 390, + 633, + 467, + 671 + ], + "lines": [ + { + "bbox": [ + 390, + 633, + 467, + 671 + ], + "spans": [ + { + "bbox": [ + 390, + 633, + 467, + 671 + ], + "type": "image", + "image_path": "6fd2f6a1eddd722cd5726cbeeba4e00fcc09af5cd030d1f27335a21db28fd2ba.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 673, + 444, + 683 + ], + "lines": [ + { + "bbox": [ + 413, + 673, + 444, + 683 + ], + "spans": [ + { + "bbox": [ + 413, + 673, + 444, + 683 + ], + "type": "text", + "content": "IPG [48]" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 468, + 633, + 547, + 671 + ], + "blocks": [ + { + "bbox": [ + 493, + 624, + 522, + 633 + ], + "lines": [ + { + "bbox": [ + 493, + 624, + 522, + 633 + ], + "spans": [ + { + "bbox": [ + 493, + 624, + 522, + 633 + ], + "type": "text", + "content": "HAT [9]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 633, + 547, + 671 + ], + "lines": [ + { + "bbox": [ + 468, + 633, + 547, + 671 + ], + "spans": [ + { + "bbox": [ + 468, + 633, + 547, + 671 + ], + "type": "image", + "image_path": "0cc8778a261620189553d400b5e40ce540bda41f8bcdce79b24bbdddb39c0df5.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 673, + 544, + 683 + ], + "lines": [ + { + "bbox": [ + 471, + 673, + 544, + 683 + ], + "spans": [ + { + "bbox": [ + 471, + 673, + 544, + 683 + ], + "type": "text", + "content": "TEAFormer (Ours)" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_caption" + } + ], + "index": 38 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16051" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 64, + 547, + 235 + ], + "blocks": [ + { + "bbox": [ + 61, + 64, + 547, + 235 + ], + "lines": [ + { + "bbox": [ + 61, + 64, + 547, + 235 + ], + "spans": [ + { + "bbox": [ + 61, + 64, + 547, + 235 + ], + "type": "table", + "html": "
MethodParamsIndoor ScenesOutdoor ScenesCombined
(M)PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓PSNR ↑SSIM ↑MAE ↓LPIPS ↓
DPDNetS [1]32.326.540.8160.0310.23922.250.6820.0560.31324.340.7470.0440.277
KPACs [47]2.0627.970.8520.0260.18222.620.7010.0530.26925.220.7740.0400.227
IFANs [27]10.528.110.8610.0260.17922.760.7200.0520.25425.370.7890.0390.217
RestormerS [63]26.128.870.8820.0250.14523.240.7430.0500.20925.980.8110.0380.178
SFNetS [12]13.229.160.8780.0230.16823.450.7470.0490.24426.230.8110.0370.207
GRLS-B [61]19.929.060.8860.0240.13923.450.7610.0490.19626.180.8220.0370.168
IRNeXTS [13]13.229.220.8790.0240.16723.530.7520.0490.24426.300.8140.0370.206
TEAFormerS (Ours)15.429.500.8920.0230.15023.550.7670.0500.21126.450.8280.0370.181
DPDNetD [1]32.327.480.8490.0290.18922.900.7260.0520.25525.130.7860.0410.223
RDPD_D [2]8.2028.100.8430.0270.21022.820.7040.0530.29825.390.7720.0400.255
IFAN_D [27]10.528.660.8680.0250.17223.460.7430.0490.24025.990.8040.0370.207
Restormer_D [63]26.129.480.8950.0230.13423.970.7730.0470.17526.660.8330.0350.155
Uformer_D [54]20.628.230.8600.0260.19923.100.7280.0510.28525.650.7950.0390.243
GRL_D-B [61]19.929.830.9030.0220.11424.390.7950.0450.15027.040.8470.0340.133
TEAFormer_D (Ours)15.430.160.9050.0220.12824.450.7950.0450.17027.230.8490.0340.150
", + "image_path": "44397cee24ca27691a902e183c701519d89297e633ad9f3d3b76e56dbfbf1186.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 553, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 553, + 258 + ], + "type": "text", + "content": "Table 4. Defocus deblurring comparisons on the DPDD [1] testset (containing 37 indoor and 39 outdoor scenes). S: single-image defocus deblurring. D: dual-pixel defocus deblurring. The best two results are highlighted in red and blue, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 263, + 137, + 277 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 137, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 137, + 277 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 147, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 147, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 147, + 297 + ], + "type": "text", + "content": "5.1. Ablation study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 296, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 296, + 541 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 296, + 541 + ], + "type": "text", + "content": "We analyze the influence exerted by the TEA on both performance, computational complexity, generalization, and training convergence in Table 2. In this experiment, we disassemble TEA into several plug-and-play modules and gradually add them into the SwinIR baseline restoration network. To ensure comparable parameters and FLOPs among the models being evaluated, we adjusted the window size of SwinIR to 16 and increased the number of parameters to 21 M, which is termed SwinIR-Large. Following the incorporation of our SkvSA, there is a marked improvement in the convergence rate and generalization abilities. Subsequent to the integration of our ASkvSA and DSA, there is a progressive enhancement in the model's performance, without detrimental effects on convergence and generalization. The results of this experiment demonstrate the efficacy of each component within the TEA. Furthermore, it illustrates that TEA resolves the dilemma between self-attention and sliding window attention, thereby attaining high performance by the global receptive field and reducing computational complexity simultaneously." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 551, + 124, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 124, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 124, + 563 + ], + "type": "text", + "content": "5.2. Image SR" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": "Training settings. We used the same training hyperparameters and evaluation methods as IPG [48]. All of our models were trained from scratch on the DF2K [3, 32] datasets with " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " resolution. The optimizer was Adam [26], the learning rate was " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " while using the cosine scheduler. We used " + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 55, + 569, + 295, + 640 + ], + "type": "text", + "content": " loss as a criterion." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Results. We first compare TEAFormer with various classical image SR methods, including EDSR [32], RCAN [71], SwinIR [31], EDT [30], HAT [9], SRFormer [73], GRL [61], DAT [10], and IPG [48]. We compare them on synthetic benchmarks [5, 24, 37, 40, 64] generated by bicubic downsampling. The FLOPs and latency reported were" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 265, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 276 + ], + "type": "text", + "content": "calculated with one NVIDIA A800 GPU on 768 resolution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": "The quantitative comparison of the methods for classical image SR is shown in Table 3. The exceptional performance of TEAFormer is distinctly evident as it achieves the highest results across nearly all five datasets for various scale factors. Incorporating the attention module with translation equivalence introduces a more suitable and productive inductive bias to the restoration model, thereby enhancing its performance. For the " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\times 4" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " SR, our TEAFormer achieves a " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "28.67\\mathrm{dB}" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " PSNR score on the Urban100 dataset, which is " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "0.7\\mathrm{dB}" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " higher than HAT [9] while both has " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "20\\mathrm{M}" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " parameters. In comparison to the preceding image SR method IPG [48], TEAFormer demonstrates a gain of " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "0.54\\mathrm{dB}" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " with comparable FLOPs and a " + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "inline_equation", + "content": "2\\times" + }, + { + "bbox": [ + 312, + 277, + 555, + 433 + ], + "type": "text", + "content": " faster inference." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "text", + "content": "We further conduct TEAFormer-Light, which only comprises 829k parameters. In comparison to the CNN method HAN [42], TEAFormer-Light has been demonstrated to achieve superior performance with only its " + }, + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "inline_equation", + "content": "1.3\\%" + }, + { + "bbox": [ + 313, + 434, + 556, + 542 + ], + "type": "text", + "content": " parameters. In comparison to other lightweight transformer methods, TEAFormer also demonstrates an absolute performance advantage, as evidenced by the fact that TEAFormer-Light achieves a higher PSNR than SwinIR-L on the Urban100 [24] dataset by 0.57 dB." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 551, + 462, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 551, + 462, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 551, + 462, + 563 + ], + "type": "text", + "content": "5.3. Image Defocus Deblurring" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "text", + "content": "Training settings. We employed the same training hyperparameters and evaluation as Restormer [63] in image deblurring tasks. Our TEAFormer were trained from scratch using progressive learning [63]. During the training phase, the optimizer adopted was AdamW [35] with " + }, + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "inline_equation", + "content": "\\mathrm{L}_1" + }, + { + "bbox": [ + 313, + 569, + 553, + 628 + ], + "type": "text", + "content": " loss." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "Results. We compare our TEAFormer with single image defocus deblurring methods [1, 13, 14, 27, 47, 61, 63] and dual-pixel image defocus deblurring methods [1, 2, 27, 54, 61, 63] on DPDD [1] dataset in Table 4. Our TEAFormer significantly outperforms state-of-the-art schemes for the single-image and dual-pixel defocus deblurring tasks on PSNR and SSIM metrics. In particular in the category" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16052" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 228, + 169 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 228, + 169 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 228, + 169 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 228, + 169 + ], + "type": "image", + "image_path": "3a50c42c38dc7fd6d4882f9cf2e7aae606bdf0310ed3247b4116b9e2848ca586.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 231, + 70, + 309, + 170 + ], + "blocks": [ + { + "bbox": [ + 231, + 70, + 309, + 170 + ], + "lines": [ + { + "bbox": [ + 231, + 70, + 309, + 170 + ], + "spans": [ + { + "bbox": [ + 231, + 70, + 309, + 170 + ], + "type": "image", + "image_path": "e8d8386e85b8b68b24f1c3464dac06ca8cde59ef568637255ad28223227d1348.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 170, + 529, + 180 + ], + "lines": [ + { + "bbox": [ + 79, + 170, + 529, + 180 + ], + "spans": [ + { + "bbox": [ + 79, + 170, + 529, + 180 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with recent SOTA methods on the image defocus blur task. Please zoom in for better view." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 310, + 70, + 388, + 170 + ], + "blocks": [ + { + "bbox": [ + 310, + 70, + 388, + 170 + ], + "lines": [ + { + "bbox": [ + 310, + 70, + 388, + 170 + ], + "spans": [ + { + "bbox": [ + 310, + 70, + 388, + 170 + ], + "type": "image", + "image_path": "5535a2ec071383096f6e31e4a7fa1f9d438cef59cf2b65e394d3c73871c5cc37.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 70, + 466, + 170 + ], + "blocks": [ + { + "bbox": [ + 389, + 70, + 466, + 170 + ], + "lines": [ + { + "bbox": [ + 389, + 70, + 466, + 170 + ], + "spans": [ + { + "bbox": [ + 389, + 70, + 466, + 170 + ], + "type": "image", + "image_path": "c0ee018cd4b939932d67af6dd9095e6507b1f7cba06de2478781c79b1cc0af70.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 468, + 70, + 546, + 170 + ], + "blocks": [ + { + "bbox": [ + 468, + 70, + 546, + 170 + ], + "lines": [ + { + "bbox": [ + 468, + 70, + 546, + 170 + ], + "spans": [ + { + "bbox": [ + 468, + 70, + 546, + 170 + ], + "type": "image", + "image_path": "e93c9b4174e7f866d190746fde1dcb62a83bdfd4643e24cb48fc549b7a90bb62.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 184, + 555, + 292 + ], + "blocks": [ + { + "bbox": [ + 58, + 184, + 555, + 292 + ], + "lines": [ + { + "bbox": [ + 58, + 184, + 555, + 292 + ], + "spans": [ + { + "bbox": [ + 58, + 184, + 555, + 292 + ], + "type": "table", + "html": "
MethodParamsColorGrayscale
(M)CBSD68 [39]Kodak24 [17]McMaster [69]Urban100 [24]Set12 [67]BSD68 [39]Urban100 [24]
σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50σ=15 σ=25 σ=50
DnCNN [67]0.5633.9031.2427.9534.6032.1428.9533.4531.5228.6232.9830.8127.5932.8630.4427.18
DRUNet [68]32.634.3031.6928.5135.3132.8929.8635.4033.1430.0834.8132.6029.6133.2530.9427.90
EDT-B [30]11.534.3931.7628.5635.3732.9429.8735.6133.3430.2535.2233.0730.16---
SwinIR* [31]11.734.3931.7828.5535.3332.8829.7735.5933.2030.1935.1132.8929.7733.3431.0027.88
Restormer [63]26.134.3931.7828.5935.4433.0230.0035.5533.3130.2935.0632.9130.0233.3531.0428.01
ART* [66]16.234.4031.7928.5935.4733.0430.0035.6133.3330.2735.1232.9529.9933.3531.0528.01
GRL-B* [61]19.834.4031.7828.5835.4233.0330.0235.6333.3630.2635.2233.0130.2433.3431.0227.99
TEAFormer (Ours)15.434.4031.8028.6235.5033.0830.0535.6833.4230.3835.2533.1830.3733.3731.0628.04
", + "image_path": "1b579d8d693be77c2ea08f772aa9ba036bf371f79825edb0124f23775de711df.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "text", + "content": "Table 5. Color and grayscale image Gaussian denoising results in terms of PSNR " + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "inline_equation", + "content": "\\uparrow" + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "text", + "content": " . Model complexity and prediction accuracy are shown for better comparison. Dash lines separate the results of CNNs and Transformers. The best two results are highlighted in red and blue, respectively. Methods with \" " + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "inline_equation", + "content": "*" + }, + { + "bbox": [ + 55, + 294, + 555, + 327 + ], + "type": "text", + "content": " \" are replicated with standard training setting, which learns a single model to handle various noise levels." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 328, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 388 + ], + "type": "text", + "content": "of single-pixel indoor scene, TEAFormer yields 0.44 dB improvements over the GRL previous best transformer method [61]. Figure 5 illustrates that our TEAFormer is more effective in removing spatially varying defocus blur than other restoration transformers." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 392, + 157, + 406 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 157, + 406 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 157, + 406 + ], + "type": "text", + "content": "5.4. Image Denoising" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 434 + ], + "type": "text", + "content": "Training settings. The image denoising and defocus deblurring tasks used the same hyper-parameters for training." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 435, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 295, + 591 + ], + "type": "text", + "content": "Gaussian Denoising Results. We perform denoising experiments on synthetic benchmark datasets generated with additive white Gaussian noise (Set12 [67], BSD68 [38], Urban100 [24], Kodak24 [17] and McMaster [69]). Following DRUNet [68], we learn a single model to handle various noise levels, including 15, 25, and 50. Our TEAFourner achieves state-of-the-art performance under both experimental settings on different datasets and noise levels. Specifically, for the challenging noise level 50 on color Urban100 [24] dataset, TEAFourner achieves 0.13 dB gain over the previous best restoration transformer GRL [61], as shown in Table 5. Similar performance gains can be observed for the Gaussian gray denoising." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 600, + 214, + 612 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 214, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 214, + 612 + ], + "type": "text", + "content": "5.5. All-in-one Image Restoration" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Training settings. Following DCPT [23], a 5D all-in-one image restoration and enhancement task was conducted with the objective of training the network to eliminate five distinct types of degradation: haze, rain, Gaussian noise, motion blur, and low light. The models were trained from scratch on datasets with " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " resolution without progressive learning. During training, the optimizer was set as AdamW, the learning rate was set as " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " while using" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 328, + 525, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 525, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 525, + 339 + ], + "type": "text", + "content": "the cosine scheduler. We used " + }, + { + "bbox": [ + 313, + 328, + 525, + 339 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 313, + 328, + 525, + 339 + ], + "type": "text", + "content": " loss as a criterion." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 460 + ], + "type": "text", + "content": "Results. As illustrated in Table 6, TEAFormer surpasses the most recently proposed methods specifically designed for comprehensive all-in-one restoration tasks (IDR [65], PromptIR [43], and InstructIR [11]). The average PSNR of TEAFormer in all five tasks reached 31.22 dB, thereby exceeding the current state-of-the-art 5D all-in-one restoration method, InstructIR [11], by 1.67 dB. TEAFormer consistently outperforms prior PSNR in every indicator of each task, signifying its exceptional proficiency in image restoration under complex multi-task conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 558 + ], + "type": "text", + "content": "Mixed degradation. As shown in Table 7, TEAFormer surpasses the most recently proposed methods under mixed degradation scenarios, specifically designed for universal restoration tasks [36, 72]). Within scenarios characterized by a combination of three degradation (low-light, haze, and rain), TEAFormer demonstrates superior performance over the current state-of-the-art universal restoration method, DiffUIR [72], by a margin of " + }, + { + "bbox": [ + 313, + 461, + 554, + 558 + ], + "type": "inline_equation", + "content": "0.42\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 461, + 554, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 563, + 387, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 563, + 387, + 574 + ], + "spans": [ + { + "bbox": [ + 314, + 563, + 387, + 574 + ], + "type": "text", + "content": "5.6. Discussion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": "Hyperparameters analysis. The performance under various hyperparameters is reported in Table 8 " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "(w)" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": ", Table 9 " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "(s)" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": ", Table 10 " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "(k)" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": " and Table 11 " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "(N_d)" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": ". Our settings " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "(w = 15, s = 4, k = 3, N_d = 16)" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": " optimally balanced performance and computation cost. Note that changing some of the hyperparameters does not affect the computational complexity as discussed in Section 4.3. For " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": " in Table 9, parameters and FLOPs are omitted because " + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 581, + 554, + 676 + ], + "type": "text", + "content": " does not affect them." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Compare different downsampler in DSA. The incorporation of DSA into our TEA substantially enhances the performance of the model without adding undue complexity." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16053" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 70, + 530, + 186 + ], + "blocks": [ + { + "bbox": [ + 78, + 70, + 530, + 186 + ], + "lines": [ + { + "bbox": [ + 78, + 70, + 530, + 186 + ], + "spans": [ + { + "bbox": [ + 78, + 70, + 530, + 186 + ], + "type": "table", + "html": "
MethodParamsDehazing on SOTS [28]Deraining on Rain100L [60]Denoising on BSD68 [38]Deblurring on GoPro [41]Low-Light on LOL [55]Average
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.04 / 0.88432.98 / 0.95130.91 / 0.88224.35 / 0.78118.18 / 0.73525.49 / 0.846
IDR [65]15.325.24 / 0.94335.63 / 0.96531.60 / 0.88727.87 / 0.84621.34 / 0.82628.34 / 0.893
PromptIR [43]35.625.20 / 0.93135.94 / 0.96431.17 / 0.88227.32 / 0.84220.94 / 0.79928.11 / 0.883
InstructIR [11]35.127.10 / 0.95636.84 / 0.97331.40 / 0.88729.40 / 0.88623.00 / 0.83629.55 / 0.907
SwinIR [31]11.821.50 / 0.89130.78 / 0.92330.59 / 0.86824.52 / 0.77317.81 / 0.72325.04 / 0.835
NAFNet [7]67.925.23 / 0.93935.56 / 0.96731.02 / 0.88326.53 / 0.80820.49 / 0.80927.76 / 0.881
Restormer [63]26.124.09 / 0.92734.81 / 0.96231.49 / 0.88427.22 / 0.82920.41 / 0.80627.60 / 0.881
TEAFormer (Ours)15.431.57 / 0.98039.44 / 0.98631.54 / 0.89030.53 / 0.90823.06 / 0.85631.22 / 0.924
", + "image_path": "906585e99da5fe0b28a6ecc870100a8fcf2339604726a6e09056c572bf8d4d15.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 63, + 223, + 134, + 271 + ], + "blocks": [ + { + "bbox": [ + 63, + 223, + 134, + 271 + ], + "lines": [ + { + "bbox": [ + 63, + 223, + 134, + 271 + ], + "spans": [ + { + "bbox": [ + 63, + 223, + 134, + 271 + ], + "type": "image", + "image_path": "a4d8814858f9972493a82632b410fcd27208c6746a1d3afddb67a07a59ae5fb7.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 273, + 104, + 280 + ], + "lines": [ + { + "bbox": [ + 93, + 273, + 104, + 280 + ], + "spans": [ + { + "bbox": [ + 93, + 273, + 104, + 280 + ], + "type": "text", + "content": "LQ" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 135, + 223, + 205, + 270 + ], + "blocks": [ + { + "bbox": [ + 135, + 223, + 205, + 270 + ], + "lines": [ + { + "bbox": [ + 135, + 223, + 205, + 270 + ], + "spans": [ + { + "bbox": [ + 135, + 223, + 205, + 270 + ], + "type": "image", + "image_path": "8b1f2300adaa7ed38d31071431e54bd8c0cd46762e74eaabb3a3173c529cc93f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 272, + 193, + 280 + ], + "lines": [ + { + "bbox": [ + 148, + 272, + 193, + 280 + ], + "spans": [ + { + "bbox": [ + 148, + 272, + 193, + 280 + ], + "type": "text", + "content": "InstrucrIR [11]" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 206, + 223, + 277, + 270 + ], + "blocks": [ + { + "bbox": [ + 206, + 223, + 277, + 270 + ], + "lines": [ + { + "bbox": [ + 206, + 223, + 277, + 270 + ], + "spans": [ + { + "bbox": [ + 206, + 223, + 277, + 270 + ], + "type": "image", + "image_path": "71d0654490210f525dafda7e9d73ac836263dad432f0b7b25f8018f4823c5362.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 217, + 272, + 266, + 280 + ], + "lines": [ + { + "bbox": [ + 217, + 272, + 266, + 280 + ], + "spans": [ + { + "bbox": [ + 217, + 272, + 266, + 280 + ], + "type": "text", + "content": "OneRestore [19]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 63, + 281, + 134, + 327 + ], + "blocks": [ + { + "bbox": [ + 63, + 281, + 134, + 327 + ], + "lines": [ + { + "bbox": [ + 63, + 281, + 134, + 327 + ], + "spans": [ + { + "bbox": [ + 63, + 281, + 134, + 327 + ], + "type": "image", + "image_path": "99e7714b36460186d16a8eedbb92ac31227c3593ab5eede043fa2d827c0d16a1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 93, + 329, + 104, + 337 + ], + "lines": [ + { + "bbox": [ + 93, + 329, + 104, + 337 + ], + "spans": [ + { + "bbox": [ + 93, + 329, + 104, + 337 + ], + "type": "text", + "content": "HQ" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 135, + 281, + 205, + 327 + ], + "blocks": [ + { + "bbox": [ + 135, + 281, + 205, + 327 + ], + "lines": [ + { + "bbox": [ + 135, + 281, + 205, + 327 + ], + "spans": [ + { + "bbox": [ + 135, + 281, + 205, + 327 + ], + "type": "image", + "image_path": "12c8e9e1664bf7e969153bac10b2f9fe3a715e5b58f0b44559a78b83ac274668.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 150, + 329, + 190, + 337 + ], + "lines": [ + { + "bbox": [ + 150, + 329, + 190, + 337 + ], + "spans": [ + { + "bbox": [ + 150, + 329, + 190, + 337 + ], + "type": "text", + "content": "DACLIP [36]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 206, + 281, + 277, + 327 + ], + "blocks": [ + { + "bbox": [ + 206, + 281, + 277, + 327 + ], + "lines": [ + { + "bbox": [ + 206, + 281, + 277, + 327 + ], + "spans": [ + { + "bbox": [ + 206, + 281, + 277, + 327 + ], + "type": "image", + "image_path": "56d5bc26ccd2281c8dd0451fbde398d2e161d806ef052f8c8ed7de148be47eb4.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 329, + 272, + 337 + ], + "lines": [ + { + "bbox": [ + 211, + 329, + 272, + 337 + ], + "spans": [ + { + "bbox": [ + 211, + 329, + 272, + 337 + ], + "type": "text", + "content": "TEAFormer (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 339, + 286, + 361 + ], + "lines": [ + { + "bbox": [ + 55, + 339, + 286, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 286, + 361 + ], + "type": "text", + "content": "Figure 6. Visual comparison on low-light+haze+rain mixed degradation. TEAFormer can restore the correct lightness." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 288, + 228, + 552, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 187, + 555, + 221 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 555, + 221 + ], + "type": "text", + "content": "Table 6. 5D All-in-one image restoration results. Dash lines separate the results of methods designed specifically for all-in-one image restoration task and methods that only focus on architecture enhancement. The best two results are highlighted in red and blue, respectively. TEAFormer sets the new state-of-the-art performance in each task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 288, + 228, + 552, + 345 + ], + "lines": [ + { + "bbox": [ + 288, + 228, + 552, + 345 + ], + "spans": [ + { + "bbox": [ + 288, + 228, + 552, + 345 + ], + "type": "table", + "html": "
MethodParams low-light+haze+rain low-light+haze+snow
(M)PSNR↑ / SSIM↑PSNR↑ / SSIM↑
AirNet [29]8.9021.80 / 0.70822.23 / 0.725
TransWeather [50]31.021.55 / 0.67821.01 / 0.655
PromptIR [43]35.623.74 / 0.75223.33 / 0.747
DACLIP [36]17422.96 / 0.71223.32 / 0.751
DiffUIR-L [72]36.325.46 / 0.77925.89 / 0.788
OneRestore [19]5.9825.18 / 0.79525.28 / 0.797
InstructIR [11]35.124.84 / 0.77724.32 / 0.760
TEAFormer (Ours)15.425.88 / 0.79626.05 / 0.800
", + "image_path": "b1feee8dcbd171a697ec6745ecea3792e7507796b287e534d89e7a063522b7ee.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 57, + 361, + 156, + 428 + ], + "blocks": [ + { + "bbox": [ + 300, + 346, + 539, + 357 + ], + "lines": [ + { + "bbox": [ + 300, + 346, + 539, + 357 + ], + "spans": [ + { + "bbox": [ + 300, + 346, + 539, + 357 + ], + "type": "text", + "content": "Table 7. Mixed degraded image restoration results on CDD [19]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 361, + 156, + 428 + ], + "lines": [ + { + "bbox": [ + 57, + 361, + 156, + 428 + ], + "spans": [ + { + "bbox": [ + 57, + 361, + 156, + 428 + ], + "type": "table", + "html": "
wFLOPs (T)Urban100 [24] PSNR ↑/SSIM ↑
70.82728.35 / 0.8441
151.03528.67 / 0.8489
311.86628.73 / 0.8493
", + "image_path": "330b0a532757ac0a12fd8e07503765531a9033b55be2779c9ae27e183922ad9d.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 430, + 153, + 452 + ], + "lines": [ + { + "bbox": [ + 55, + 430, + 153, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 153, + 452 + ], + "type": "text", + "content": "Table 8. Ablations of sliding window size " + }, + { + "bbox": [ + 55, + 430, + 153, + 452 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 430, + 153, + 452 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 157, + 361, + 226, + 428 + ], + "blocks": [ + { + "bbox": [ + 157, + 361, + 226, + 428 + ], + "lines": [ + { + "bbox": [ + 157, + 361, + 226, + 428 + ], + "spans": [ + { + "bbox": [ + 157, + 361, + 226, + 428 + ], + "type": "table", + "html": "
sUrban100 [24]
PSNR ↑ / SSIM ↑
128.43 / 0.8431
228.51 / 0.8453
428.67 / 0.8489
", + "image_path": "cc3849e77d2774eae9cf70053bd5677780ece7d421b1088f3ac4a073ed69ac19.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 155, + 430, + 222, + 461 + ], + "lines": [ + { + "bbox": [ + 155, + 430, + 222, + 461 + ], + "spans": [ + { + "bbox": [ + 155, + 430, + 222, + 461 + ], + "type": "text", + "content": "Table 9. Ablations of sliding stride " + }, + { + "bbox": [ + 155, + 430, + 222, + 461 + ], + "type": "inline_equation", + "content": "s" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 19 + }, + { + "type": "table", + "bbox": [ + 227, + 361, + 552, + 428 + ], + "blocks": [ + { + "bbox": [ + 227, + 361, + 552, + 428 + ], + "lines": [ + { + "bbox": [ + 227, + 361, + 552, + 428 + ], + "spans": [ + { + "bbox": [ + 227, + 361, + 552, + 428 + ], + "type": "table", + "html": "
kFLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Nd(T)FLOPs (T)Urban100 [24] PSNR ↑ / SSIM ↑Down (%)↑S-conv (%)↑Urban100 [24] PSNR ↑ / SSIM ↑
None0.96628.49 / 0.8470None0.75728.47 / 0.8457MaxPool87.328.61 / 0.8477
11.02828.53 / 0.8471161.03528.67 / 0.8489LPD91.728.64 / 0.8493
31.03528.67 / 0.8489N18.3628.80 / 0.8503AvgPool91.228.67 / 0.8489
", + "image_path": "cdfa300054871c267f0fa5857091b820b3e317e8dfc07ba3c97d8062c6bbf6d4.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 225, + 430, + 430, + 473 + ], + "lines": [ + { + "bbox": [ + 225, + 430, + 430, + 473 + ], + "spans": [ + { + "bbox": [ + 225, + 430, + 430, + 473 + ], + "type": "text", + "content": "Table 10. Ablations of kernel size " + }, + { + "bbox": [ + 225, + 430, + 430, + 473 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 225, + 430, + 430, + 473 + ], + "type": "text", + "content": " for adaptive in- in DSA. \"None\" stands for indices. \"None\" stands for no no DSA used. adaptive indices used." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_footnote" + }, + { + "bbox": [ + 433, + 430, + 544, + 473 + ], + "lines": [ + { + "bbox": [ + 433, + 430, + 544, + 473 + ], + "spans": [ + { + "bbox": [ + 433, + 430, + 544, + 473 + ], + "type": "text", + "content": "Table 12. Ablations of down-samplers in DSA. The higher the S-Conv, the more strictly the model meets TE." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 21 + }, + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 296, + 535 + ], + "type": "text", + "content": "However, the average pooling employed in DSA does not strictly adhere to the property of TE. Using a comparative ablation of several downsampling modules in Table 12, we experimentally demonstrate that the use of average pooling does not compromise the model's performance or its TE." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 551, + 296, + 718 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 551, + 296, + 718 + ], + "spans": [ + { + "bbox": [ + 55, + 551, + 296, + 718 + ], + "type": "text", + "content": "Does TEAFormer converge faster and generalize better in practice? Table 1 and Table 2 have illustrated that TEAFormer converges faster and generalizes better than other networks theoretically due to the incorporation of translation equivariance. Figure 7 (left) further demonstrates that TEAFormer reaches a lower loss faster than other restoration transformers [9, 31] in practice. Following MaskDenoise [6], we compare the statistics of the last layer feature in SwinIR [31] and TEAFormer under two different real-world degradations on the PIES [33] dataset (out of the training dataset) in Figure 7 (right). The results show that TEAFormer can produce a more consistent output when processing LQ images with different degradations. Therefore, TEAFormer generalizes better in practice." + } + ] + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 319, + 475, + 412, + 534 + ], + "blocks": [ + { + "bbox": [ + 319, + 475, + 412, + 534 + ], + "lines": [ + { + "bbox": [ + 319, + 475, + 412, + 534 + ], + "spans": [ + { + "bbox": [ + 319, + 475, + 412, + 534 + ], + "type": "image", + "image_path": "e486577d5539c649550783b4f33df6f06ed8f4e981024632d753d3b438cedb94.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 536, + 549, + 546 + ], + "lines": [ + { + "bbox": [ + 318, + 536, + 549, + 546 + ], + "spans": [ + { + "bbox": [ + 318, + 536, + 549, + 546 + ], + "type": "text", + "content": "Figure 7. Comparison on convergence (left) and generalization (right)." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 410, + 480, + 478, + 533 + ], + "blocks": [ + { + "bbox": [ + 410, + 480, + 478, + 533 + ], + "lines": [ + { + "bbox": [ + 410, + 480, + 478, + 533 + ], + "spans": [ + { + "bbox": [ + 410, + 480, + 478, + 533 + ], + "type": "image", + "image_path": "314d73fb472fdca92b1593ae33e8790d037d0eea31edcf1e6bcddd02dbf5d820.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 480, + 481, + 546, + 533 + ], + "blocks": [ + { + "bbox": [ + 480, + 481, + 546, + 533 + ], + "lines": [ + { + "bbox": [ + 480, + 481, + 546, + 533 + ], + "spans": [ + { + "bbox": [ + 480, + 481, + 546, + 533 + ], + "type": "image", + "image_path": "c275121ffca482548733363c9f7ee340838c565ac730280273df7cb45c6666c6.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 387, + 564 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we start by examining the translation equivariance, which serves as an inherent inductive bias in image restoration. We then define the translation equivariance and offer two strategies for incorporating it into restoration networks: (1) slide indexing and (2) component stacking. To address the fixed receptive field provided by these strategies, we proposed TEA and TEAFormer via adaptive key-value indexing. The efficacy of TEAFormer has been demonstrated in various image restoration tasks, with ablation studies confirming that this efficacy arises from the incorporation of TEA. In the future, we intend to explore more equivariances into restoration networks." + } + ] + } + ], + "index": 31 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16054" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 204, + 115, + 217 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 224, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 224, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 224, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 61, + 224, + 295, + 258 + ], + "type": "text", + "content": "[1] Abdullah Abuolaim and Michael S Brown. Defocus deblurring using dual-pixel data. In European Conference on Computer Vision, pages 111-126. Springer, 2020. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 259, + 296, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 259, + 296, + 314 + ], + "spans": [ + { + "bbox": [ + 61, + 259, + 296, + 314 + ], + "type": "text", + "content": "[2] Abdullah Abuolaim, Mauricio Delbracio, Damien Kelly, Michael S Brown, and Peyman Milanfar. Learning to reduce defocus blur by realistically modeling dual-pixel data. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 295, + 360 + ], + "type": "text", + "content": "[3] Eirikur Agustsson and Radu Timofte. Ntire 2017 challenge on single image super-resolution: Dataset and study. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 361, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 295, + 407 + ], + "type": "text", + "content": "[4] Irwan Bello, Barret Zoph, Ashish Vaswani, Jonathon Shlens, and Quoc V Le. Attention augmented convolutional networks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3286-3295, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 407, + 296, + 461 + ], + "type": "text", + "content": "[5] Marco Bevilacqua, Aline Roumy, Christine Guillemot, and Marie Line Alberi-Morel. Low-complexity single-image super-resolution based on nonnegative neighbor embedding. In British Machine Vision Conference. BMVA press, 2012. 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 464, + 296, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 296, + 529 + ], + "type": "text", + "content": "[6] Haoyu Chen, Jinjin Gu, Yihao Liu, Salma Abdel Magid, Chao Dong, Qiong Wang, Hanspeter Pfister, and Lei Zhu. Masked image training for generalizable deep image denoising. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1692-1703, 2023. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 295, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 295, + 575 + ], + "type": "text", + "content": "[7] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In Computer Vision - ECCV 2022, pages 17-33, Cham, 2022. Springer Nature Switzerland. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 295, + 621 + ], + "type": "text", + "content": "[8] Xiangning Chen, Cho-Jui Hsieh, and Boqing Gong. When vision transformers outperform resnets without pre-training or strong data augmentations. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 624, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 624, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 624, + 295, + 678 + ], + "type": "text", + "content": "[9] Xiangyu Chen, Xintao Wang, Jiantao Zhou, Yu Qiao, and Chao Dong. Activating more pixels in image superresolution transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22367-22377, 2023. 2, 5, 6, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[10] Zheng Chen, Yulun Zhang, Jinjin Gu, Linghe Kong, Xiaokang Yang, and Fisher Yu. Dual aggregation transformer for image super-resolution. In ICCV, 2023. 2, 5, 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 180 + ], + "type": "text", + "content": "This work was supported financially by the Natural Science Foundation of China (82371112, 623B2001), the Science Foundation of Peking University Cancer Hospital (JC202505), Natural Science Foundation of Beijing Municipality (Z210008) and the Clinical Medicine Plus X - Young Scholars Project of Peking University, the Fundamental Research Funds for the Central Universities (PKU2025PKULCXQ008)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[11] Marcos V Conde, Gregor Geigle, and Radu Timofte. Instructir: High-quality image restoration following human instructions. In Proceedings of the European Conference on Computer Vision (ECCV), 2024. 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 173 + ], + "type": "text", + "content": "[12] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 217 + ], + "type": "text", + "content": "[13] Yuning Cui, Wenqi Ren, Sining Yang, Xiaochun Cao, and Alois Knoll. Irnext: Rethinking convolutional network design for image restoration. In Proceedings of the 40th International Conference on Machine Learning, 2023. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 554, + 274 + ], + "type": "text", + "content": "[14] Yuning Cui, Yi Tao, Zhenshan Bing, Wenqi Ren, Xinwei Gao, Xiaochun Cao, Kai Huang, and Alois Knoll. Selective frequency network for image restoration. In The Eleventh International Conference on Learning Representations, 2023. 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 341 + ], + "type": "text", + "content": "[15] Xiaoyi Dong, Jianmin Bao, Dongdong Chen, Weiming Zhang, Nenghai Yu, Lu Yuan, Dong Chen, and Baining Guo. Cswin transformer: A general vision transformer backbone with cross-shaped windows. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 12124-12134, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 342, + 554, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 342, + 554, + 418 + ], + "spans": [ + { + "bbox": [ + 317, + 342, + 554, + 418 + ], + "type": "text", + "content": "[16] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 421, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 553, + 441 + ], + "type": "text", + "content": "[17] Rich Franzen. Kodak lossless true color image suite, 2013. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 317, + 444, + 553, + 487 + ], + "type": "text", + "content": "[18] Jonathan Gordon, Wessel P. Bruinsma, Andrew Y. K. Foong, James Requeima, Yann Dubois, and Richard E. Turner. Convolutional conditional neural processes. In International Conference on Learning Representations, 2020. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 489, + 554, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 489, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 554, + 532 + ], + "type": "text", + "content": "[19] Yu Guo, Yuan Gao, Yuxu Lu, Ryan Wen Liu, and Shengfeng He. Onerestore: A universal restoration framework for composite degradation. In European Conference on Computer Vision, 2024. 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 554, + 578 + ], + "type": "text", + "content": "[20] Ali Hassani, Steven Walton, Jiachen Li, Shen Li, and Humphrey Shi. Neighborhood attention transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6185-6194, 2023. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 579, + 554, + 633 + ], + "type": "text", + "content": "[21] Ali Hatamizadeh, Greg Heinrich, Hongxu Yin, Andrew Tao, Jose M Alvarez, Jan Kautz, and Pavlo Molchanov. Fastervit: Fast vision transformers with hierarchical attention. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 554, + 669 + ], + "type": "text", + "content": "[22] Jie Hu, Li Shen, and Gang Sun. Squeeze-and-excitation networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7132-7141, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[23] JiaKui Hu, Lujia Jin, Zhengjian Yao, and Yanye Lu. Universal image restoration pre-training via degradation classification. The Thirteenth International Conference on Learning Representations, 2025. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16055" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[24] Jia-Bin Huang, Abhishek Singh, and Narendra Ahuja. Single image super-resolution from transformed self-exemplars. In 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 5197-5206, 2015. 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[25] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 1, 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 204 + ], + "type": "text", + "content": "[26] Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), San Diego, CA, USA, 2015. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 259 + ], + "type": "text", + "content": "[27] Junyong Lee, Hyeongseok Son, Jaesung Rim, Sunghyun Cho, and Seungyong Lee. Iterative filter adaptive network for single image defocus deblurring. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 262, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 294, + 293 + ], + "type": "text", + "content": "[28] Boyi Li, Wenqi Ren, Dengpan Fu, Dacheng Tao, Dan Feng, Wenjun Zeng, and Zhangyang Wang. Benchmarking single-image dehazing and beyond. TIP, 2018. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 295, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 295, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 295, + 294, + 326 + ], + "type": "text", + "content": "[29] Boyun Li, Xiao Liu, Peng Hu, Zhongqin Wu, Jiancheng Lv, and Xi Peng. All-in-one image restoration for unknown corruption. In CVPR, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[30] Wenbo Li, Xin Lu, Shengju Qian, and Jiangbo Lu. On efficient transformer-based image pre-training for low-level vision. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence, IJCAI-23, pages 1089–1097. International Joint Conferences on Artificial Intelligence Organization, 2023. Main Track. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 394, + 294, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 447 + ], + "type": "text", + "content": "[31] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. Swinir: Image restoration using swim transformer. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1833-1844, 2021. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[32] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In Proceedings of the IEEE conference on computer vision and pattern recognition workshops, pages 136-144, 2017. 1, 5, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[33] Yihao Liu, Hengyuan Zhao, Jinjin Gu, Yu Qiao, and Chao Dong. Evaluating the generalization ability of superresolution networks. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14497-14513, 2023. 1, 5, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 613 + ], + "type": "text", + "content": "[34] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 2, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "text", + "content": "[35] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 690 + ], + "type": "text", + "content": "[36] Ziwei Luo, Fredrik K Gustafsson, Zheng Zhao, Jens Sjolund, and Thomas B Schon. Controlling vision-language models for universal image restoration. In The Twelfth International Conference on Learning Representations, 2023. 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[37] D. Martin, C. Fowlkes, D. Tal, and J. Malik. A database of human segmented natural images and its application to" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423 vol.2, 2001. 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[38] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In ICCV, 2001. 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 226 + ], + "type": "text", + "content": "[39] David Martin, Charless Fowlkes, Doron Tal, and Jitendra Malik. A database of human segmented natural images and its application to evaluating segmentation algorithms and measuring ecological statistics. In Proceedings Eighth IEEE International Conference on Computer Vision. ICCV 2001, pages 416-423. IEEE, 2001. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[40] Yusuke Matsui, Kota Ito, Yuji Aramaki, Azuma Fujimoto, Toru Ogawa, Toshihiko Yamasaki, and Kiyoharu Aizawa. Sketch-based manga retrieval using manga109 dataset. Multimedia Tools Appl., 76(20):21811-21838, 2017. 5, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 304 + ], + "type": "text", + "content": "[41] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In CVPR, 2017. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 306, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 553, + 370 + ], + "type": "text", + "content": "[42] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XII 16, pages 191-207. Springer, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[43] Vaishnav Potlapalli, Syed Waqas Zamir, Salman H Khan, and Fahad Shahbaz Khan. Prompt: Prompting for all-in-one image restoration. NeurIPS, 2023. 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 448 + ], + "type": "text", + "content": "[44] Prajit Ramachandran, Niki Parmar, Ashish Vaswani, Irwan Bello, Anselm Levskaya, and Jon Shlens. Stand-alone self-attention in vision models. Advances in neural information processing systems, 32, 2019. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 502 + ], + "type": "text", + "content": "[45] Renan A Rojas-Gomez, Teck-Yian Lim, Alex Schwing, Minh Do, and Raymond A Yeh. Learnable polyphase sampling for shift invariant and equivariant convolutional networks. Advances in Neural Information Processing Systems, 35:35755-35768, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 557 + ], + "type": "text", + "content": "[46] Renan A Rojas-Gomez, Teck-Yian Lim, Minh N Do, and Raymond A Yeh. Making vision transformers truly shift-equivariant. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5568-5577, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 601 + ], + "type": "text", + "content": "[47] Hyeongseok Son, Junyong Lee, Sunghyun Cho, and Seungyong Lee. Single image defocus deblurring using kernel-sharing parallel atrous convolutions. In Proc. ICCV, 2021. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 656 + ], + "type": "text", + "content": "[48] Yuchuan Tian, Hanting Chen, Chao Xu, and Yunhe Wang. Image processing gnn: Breaking rigidity in super-resolution. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24108-24117, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[49] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16056" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[50] Jeya Maria Jose Valanarasu, Rajeev Yasarla, and Vishal M Patel. Transweather: Transformer-based restoration of images degraded by adverse weather conditions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2353-2363, 2022. 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 172 + ], + "type": "text", + "content": "[51] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 237 + ], + "type": "text", + "content": "[52] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In Proceedings of the IEEE/CVF international conference on computer vision, pages 568-578, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 282 + ], + "type": "text", + "content": "[53] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[54] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 370 + ], + "type": "text", + "content": "[55] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. In British Machine Vision Conference, 2018. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 426 + ], + "type": "text", + "content": "[56] Haiping Wu, Bin Xiao, Noel Codella, Mengchen Liu, Xiyang Dai, Lu Yuan, and Lei Zhang. Cvt: Introducing convolutions to vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 22-31, 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[57] Zhuofan Xia, Xuran Pan, Shiji Song, Li Erran Li, and Gao Huang. Vision transformer with deformable attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4794-4803, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "[58] Renjun Xu, Kaifan Yang, Ke Liu, and Fengxiang He. " + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "inline_equation", + "content": "e(2)" + }, + { + "bbox": [ + 56, + 471, + 294, + 503 + ], + "type": "text", + "content": "-equivariant vision transformer. In Uncertainty in Artificial Intelligence, pages 2356-2366. PMLR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 547 + ], + "type": "text", + "content": "[59] Yufei Xu, Qiming Zhang, Jing Zhang, and Dacheng Tao. Vi-tae: Vision transformer advanced by exploring intrinsic inductive bias. Advances in Neural Information Processing Systems, 34, 2021. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[60] Wenhan Yang, Robby T Tan, Jiashi Feng, Zongming Guo, Shuicheng Yan, and Jiaying Liu. Joint rain detection and removal from a single image with contextualized deep networks. TPAMI, 2019. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[61] Xiaoyu Xiang Yawei Li, Yuchen Fan, Denis Demandolx, Rakesh Ranjan, Radu Timofte, and Luc Van Gool. Efficient and explicit modelling of image hierarchies for image restoration. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[63] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 631 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "text", + "content": "[64] Roman Zeyde, Michael Elad, and Matan Protter. On single image scale-up using sparse-representations. In *Curves and Surfaces: 7th International Conference, Avignon, France*, June 24-30, 2010, Revised Selected Papers 7, pages 711-730. Springer, 2012. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 151, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 206 + ], + "type": "text", + "content": "[65] Jinghao Zhang, Jie Huang, Mingde Yao, Zizheng Yang, Hu Yu, Man Zhou, and Feng Zhao. Ingredient-oriented multidegradation learning for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5825-5835, 2023. 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 239 + ], + "type": "text", + "content": "[66] Jiale Zhang, Yulun Zhang, Jinjin Gu, Yongbing Zhang, Linghe Kong, and Xin Yuan. Accurate image restoration with attention retractable transformer. In ICLR, 2023. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 285 + ], + "type": "text", + "content": "[67] Kai Zhang, Wangmeng Zuo, Yunjin Chen, Deyu Meng, and Lei Zhang. Beyond a Gaussian denoiser: Residual learning of deep CNN for image denoising. IEEE Transactions on Image Processing, 26(7):3142-3155, 2017. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 341 + ], + "type": "text", + "content": "[68] Kai Zhang, Yawei Li, Wangmeng Zuo, Lei Zhang, Luc Van Gool, and Radu Timofte. Plug-and-play image restoration with deep denoiser prior. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10):6360-6376, 2021. 1, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 342, + 553, + 385 + ], + "type": "text", + "content": "[69] Lei Zhang, Xiaolin Wu, Antoni Buades, and Xin Li. Color demosaicking by local directional interpolation and nonlocal adaptive thresholding. Journal of Electronic imaging, 20(2): 023016-023016, 2011. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 386, + 553, + 430 + ], + "type": "text", + "content": "[70] Qiming Zhang, Yufei Xu, Jing Zhang, and Dacheng Tao. Vitaev2: Vision transformer advanced by exploring inductive bias for image recognition and beyond. arXiv preprint arXiv:2202.10108, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 431, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 431, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 431, + 553, + 464 + ], + "type": "text", + "content": "[71] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 1, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 465, + 553, + 520 + ], + "type": "text", + "content": "[72] Dian Zheng, Xiao-Ming Wu, Shuzhou Yang, Jian Zhang, Jian-Fang Hu, and Wei-shi Zheng. Selective hourglass mapping for universal image restoration based on diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 576 + ], + "type": "text", + "content": "[73] Yupeng Zhou, Zhen Li, Chun-Le Guo, Song Bai, Ming-Ming Cheng, and Qibin Hou. Srformer: Permuted self-attention for single image super-resolution. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12780–12791, 2023. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "spans": [ + { + "bbox": [ + 316, + 577, + 553, + 631 + ], + "type": "text", + "content": "[74] Lei Zhu, Xinjiang Wang, Zhanghan Ke, Wayne Zhang, and Rynson WH Lau. Bformer: Vision transformer with bi-level routing attention. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10323-10333, 2023. 2" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16057" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_content_list.json b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c9e012ea3d05a797516cb268505986753ee240ae --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_content_list.json @@ -0,0 +1,1463 @@ +[ + { + "type": "text", + "text": "Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction", + "text_level": 1, + "bbox": [ + 99, + 128, + 898, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mang Cao $^{1}$ Sanping Zhou $^{1*}$ Yizhe Li $^{1}$ Ye Deng $^{2}$ Wenli Huang $^{3}$ Le Wang $^{1}$", + "bbox": [ + 165, + 202, + 841, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University", + "bbox": [ + 189, + 227, + 820, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics $^{3}$ School of Electronic and Information Engineering, Ningbo University of Technology", + "bbox": [ + 99, + 281, + 906, + 316 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3298592685@stu.xjtu.edu.cn, spzhou@xjtu.edu.cn", + "bbox": [ + 299, + 325, + 709, + 340 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 376, + 326, + 392 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Sufficient cross-task interaction is crucial for success in multi-task dense prediction. However, sufficient interaction often results in high computational complexity, forcing existing methods to face the trade-off between interaction completeness and computational efficiency. To address this limitation, this work proposes a Bidirectional Interaction Mamba (BIM), which incorporates novel scanning mechanisms to adapt the Mamba modeling approach for multi-task dense prediction. On the one hand, we introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task-specific representations as bidirectional sequences during interaction. By integrating task-first and position-first scanning modes within a unified linear complexity architecture, BI-Scan efficiently preserves critical cross-task information. On the other hand, we employ a Multi-Scale Scan (MS-Scan) mechanism to achieve multi-granularity scene modeling. This design not only meets the diverse granularity requirements of various tasks but also enhances nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our BIM vs its state-of-the-art competitors. Codes are available online: https://github.com/mmm-cc/BIM_for_MTL.", + "bbox": [ + 88, + 409, + 485, + 773 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 803, + 220, + 818 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multi-task dense prediction is a critical visual task designed to simultaneously predict outputs for various pixel-level tasks, such as semantic segmentation, depth predic", + "bbox": [ + 89, + 829, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1b3bbfd3c79dbe5276a4fcfd18ed2c587642c24916d678b8a71e8a6003a3a93e.jpg", + "image_caption": [ + "Figure 1. Illustration of different cross-task interaction strategies. (a) Pair-Wise [8, 28, 31]: each task can access the complete information of other tasks, while the interaction complexity is $O(n^{2})$ . (b) Bridge Representation [1, 16, 24, 39]: the interaction complexity is $O(n)$ , yet, each task is limited to accessing only a partial subset of information from other tasks during interaction, which inevitably results in the loss of valuable details. (c) Ours Cross-Task Scan: It preserves the integrity of task information during interaction, with an interaction complexity of $O(n)$ ." + ], + "image_footnote": [], + "bbox": [ + 517, + 377, + 903, + 556 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion, surface normal estimation, and saliency detection. Existing approaches made mainly efforts [5, 30] to refine task-general representations, obtained from the encoder, into task-specific representations. In these approaches, cross-task interaction has been proven to be significant in multi-task dense prediction [26, 31, 35, 39], as it can enhance the robustness and effectiveness of task representations.", + "bbox": [ + 511, + 733, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Initially, multi-task dense prediction methods primarily utilized Convolutional Neural Networks (CNNs) with pairwise interactions, as illustrated in Fig. 1 (a). These networks [8, 28, 31] were carefully designed with specialized", + "bbox": [ + 511, + 839, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "18815", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "branches for each task, augmented by interaction modules that facilitated information exchanges between each pair of tasks. These pair-wise interaction approaches maintain the integrity of task information during interaction, promoting fine-grained cross-task interaction. While, the inherently limited receptive fields of CNN architectures often led to insufficient spatial interactions. In response to these challenges, transformer-style networks [35] demonstrate remarkable capability in modeling long-range dependencies. This capability substantially improves the interaction effectiveness of models in handling multi-task scenarios, leading to enhanced performance. However, these sufficient cross-task interactions usually introduce high interaction complexity, which typically increases quadratically with the number of tasks. The high interaction complexity limits their applicability in certain powerful models, particularly transformers.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enhance the efficiency of cross-task interactions, a crowd of methods [1, 16, 24, 39] with linear interaction complexity has been proposed, as depicted in Fig. 1 (b). Typically, these methods strategically select a representative task [1, 24] or extract representative information from all tasks [16, 39] to construct a compact bridge representation that aims to preserve essential interaction information. However, in these methods, the necessity of maintaining linear interaction complexity inherently limits the representational capacity of these bridge representations. This information compression inevitably leads to the degradation of task features, potentially discarding critical interaction details and ultimately resulting in suboptimal performance. Therefore, how to enhance sufficient interaction capability while maintaining linear interaction complexity remains an unresolved issue.", + "bbox": [ + 89, + 352, + 482, + 593 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In response to these challenges, based on Mamba architecture, we propose a simple yet effective Bidirectional Interaction Mamba (BIM) method that achieves comprehensive preservation of task information during interaction while maintaining linear interaction complexity. Specifically, we introduce a Bidirectional Interaction Scan (BI-Scan) mechanism to achieve sufficient cross-task interaction and a Multi-Scale Scan (MS-Scan) mechanism to deliver a comprehensive visual representation. The BI-Scan mechanism constructs each task representation into sequences and deploys a bidirectional scanning strategy for modeling cross-task interactions, efficiently mitigating information loss while preserving linear interaction complexity. Furthermore, BI-Scan enables multi-level cross-task interactions by integrating task-first and position-first scanning modes, facilitating more nuanced modeling of task relationships. In the MS-Scan, features are initially partitioned into multiple spaces, where the scene structure of images is modeled at various scales. Specific tasks subsequently integrate multi-scale scene structural information to", + "bbox": [ + 89, + 599, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enhance task-specific representations. During cross-task interaction, different task branches adaptively integrate interaction representations from multiple scales, addressing the varying granularity demands of each task for scene structural information.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows:", + "bbox": [ + 535, + 167, + 870, + 180 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a simple yet effective Bidricational Interaction Mamba (BIM) for multi-task dense prediction, enabling comprehensive cross-task interaction while preserving linear interaction complexity.", + "- We introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism to maintain the integrity of task information during interactions. The BI-Scan incorporates two scanning modes: task-first mode and position-first mode, facilitating multi-level cross-task interactions.", + "- We design a Multi-Scale Scan (MS-Scan) mechanism to capture scene structural information at multiple scales, thereby addressing the diverse granularity requirements of various tasks and enhancing more detailed cross-task interactions in multi-task dense prediction." + ], + "bbox": [ + 511, + 181, + 903, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 407, + 653, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multi-Task Learning. Most existing multi-task learning works focused on training optimization and network design. Optimization approaches balance tasks via gradient manipulation [13, 23, 34, 38] or loss-balancing [4, 14] to coordinate resource allocation for various tasks during training. The approaches of network design seek to enhance task representation learning by developing diverse mechanisms. Some CNN-based methods [29, 31] manually design interaction mechanisms to extract useful information across tasks. With the advancement of Transformer, current methods have gained improved global task interaction capabilities to enhance task representation efficiency. They enable superior pairwise [1, 24] or global [15, 32, 35, 36] task interactions. To mitigate computational complexity, MT-Mamba [16] introduce Mamba [9] to multi-task learning, which showcases effective long sequence modeling capabilities and achieving satisfactory performance. However, achieving both efficient and sufficient cross-task interaction remains a challenge.", + "bbox": [ + 511, + 431, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "State Space Models. In efficient long-range dependency modeling methods, state space models [11, 27] has become a striking alternative to Transformers. [10] proposed a Structured State Space Sequence (S4) model based on a new parameterization, which alleviates the computational and memory efficiency issues faced by SSM. Subsequently, numerous efforts [7, 21] are dedicated to bridging the performance disparity between SSMs and Transformers. Mamba [9] introduced an input-based parameterization method and hardware-aware algorithm, achieving performance on par with Transformers in natural language processing. This success inspired various efforts [43] towards", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "18816", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/8992b268b8f720d5883f299c660f667fa679b883a3df0ee9504c680d5b35e415.jpg", + "image_caption": [ + "(a) BIM for MTL" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 498, + 296 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a225d94f9ddce0b77bcf047531970919941810f93982158b2cd6cab243c28e6.jpg", + "image_caption": [ + "(b) MFR Block", + "Figure 2. Framework of the proposed BIM for multi-task dense prediction. (a) Overview of BIM, illustrating with depth estimation and surface normal estimation tasks. (b) Details of the MFR block, including T task-specific MSST blocks and a task-shared BCFR block." + ], + "image_footnote": [], + "bbox": [ + 501, + 90, + 903, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mamba's adaptation for visual tasks.", + "bbox": [ + 89, + 378, + 334, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The preservation of comprehensive image structural information poses a critical challenge in Mamba's sequential processing model, which has attracted considerable attention and effort [12, 17, 33, 41, 43]. VMamba [17] proposes the 2D Selective Scan (SS2D), a four-way scanning mechanism tailored for spatial domain traversal, aimed at enhancing Mamba's image modeling capabilities. Subsequent research studies have explored various scan patterns and combinations tailored to different tasks or scenarios [12, 33, 41]. However, by utilizing fixed token sizes, these methods overlook the importance of hierarchical spatial structural information in visual tasks.", + "bbox": [ + 89, + 393, + 483, + 574 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 585, + 181, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first outline the overall architecture of BIM for multitask dense prediction in Section 3.1, Subsequently, Section 3.2 focuses on the BIM Encoder, while Section 3.3 examines the BIM Decoder and its two scanning mechanisms: BI-Scan and MS-Scan. Finally, Section 3.4 discusses the optimization objectives.", + "bbox": [ + 89, + 611, + 483, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Framework Overview", + "text_level": 1, + "bbox": [ + 89, + 709, + 294, + 723 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown Fig. 2 (a), the proposed BIM framework consists of three main components: a task-shared encoder $\\Phi$ for extracting task-generic representations, a Mamba decoder $\\Theta$ for refining task features and T task-specific heads $\\{H_t\\}_{t=1}^T$ generating predictions $\\{\\hat{\\mathbf{Y}}_t\\}_{t=1}^T$ for individual tasks. This can be formulated as:", + "bbox": [ + 89, + 731, + 483, + 820 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\left\\{F _ {1}, F _ {2}, \\dots , F _ {T} \\right\\} = \\Theta (\\Phi (\\mathbf {I})), \\\\ \\hat {\\mathbf {Y}} _ {t} = H _ {t} \\left(F _ {t}\\right) = \\mathcal {P} _ {t} \\left(\\Psi \\left(F _ {t}\\right)\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 825, + 480, + 863 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}$ denotes the RGB input image, $F_{t}$ denotes the features generated by the decoder, corresponding", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to the distinct task $t$ . $\\hat{\\mathbf{Y}}_t$ represents the prediction for task $t$ , which has the same height $H$ and width $W$ as I. $\\Psi$ denotes a module designed to expand the task feature resolution to $H \\times W$ , which consists of a linear projection and a reshape operation. $\\mathcal{P}_t$ represents a linear layer that projects the feature channels to a specific number required by the corresponding task.", + "bbox": [ + 511, + 377, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. BIM Encoder", + "text_level": 1, + "bbox": [ + 511, + 492, + 655, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The encoder shares similarities with other methods [16]. We utilized a pretrained Swin Transformer [18] to extract task-generic features, which begins by dividing the input image $\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}$ into $\\frac{H}{w} \\times \\frac{W}{w}$ tokens of dimension $C$ through patch partitioning and linear layers, where $w$ denotes the partition size. These tokens are then processed through multiple stages involving alternating patch merging and Swin Transformer block processing, ultimately yielding hierarchical image representations:", + "bbox": [ + 511, + 513, + 906, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {G} = \\left\\{\\mathbf {G} ^ {1}, \\mathbf {G} ^ {2}, \\mathbf {G} ^ {3}, \\mathbf {G} ^ {4} \\right\\} = \\Phi (\\mathrm {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 659, + 903, + 676 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{G}$ represents the task-generic features extracted from the encoder $\\Phi$ . In our practical implementation, we utilize a partition size of 4, resulting in the following shapes for $\\mathbf{G}$ : $C \\times \\frac{H}{4} \\times \\frac{W}{4}$ , $2C \\times \\frac{H}{8} \\times \\frac{W}{8}$ , $4C \\times \\frac{H}{16} \\times \\frac{W}{16}$ and $8C \\times \\frac{H}{32} \\times \\frac{W}{32}$ , respectively.", + "bbox": [ + 511, + 686, + 905, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. BIM Decoder", + "text_level": 1, + "bbox": [ + 511, + 771, + 655, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The BIM Decoder consists of three Mamba Feature Refinement (MFR) blocks, as depicted in Fig. 2 (a). These MFRs are designed to bridge the gap between task-generic and task-specific representations. They generate task representations $\\{\\mathbf{F}_1^i,\\mathbf{F}_2^i,\\ldots ,\\mathbf{F}_T^i\\}_{i = 1}^3$ corresponding to the first three encoder stages, with dimensions of $4C\\times \\frac{H}{16}\\times \\frac{W}{16},2C\\times \\frac{H}{8}\\times \\frac{W}{8}$ , and $C\\times \\frac{H}{4}\\times \\frac{W}{4}$ , respectively. Finally, the last", + "bbox": [ + 511, + 792, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "18817", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "refined features $\\{\\mathbf{F}_t^3\\}_{t = 1}^T$ are input into the task heads as $\\{F_{t}\\}_{t = 1}^{T}$ to generate the final predictions $\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^T$", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We introduce two specialized Mamba blocks, Bidirectional Cross-Task Feature Refine (BCFR) and Multi-Scale Single-task (MSST) in each MFR block, as illustrated in Fig. 2 (b). The BCFR Block is designed to facilitate effective cross-task interaction while maintaining the integrity of task information. It begins by concatenating features $\\{\\mathcal{F}_t^s\\}_{t=1}^T$ from different task branches to construct task-shared features $\\mathcal{F}_{sh}^s$ using the BI-Scan mechanism. Subsequently, each task branch adaptively merges $\\mathcal{F}_{sh}^s$ by a selection value $\\mathcal{G}^s$ to obtain fine-tuned task representations $\\mathbf{F}^s = \\{\\mathbf{F}_t^s\\}_{t=1}^T$ :", + "bbox": [ + 89, + 121, + 483, + 289 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} _ {t} ^ {s} = \\mathcal {F} _ {t} ^ {s} + \\mathcal {G} _ {t} ^ {s} \\times \\mathcal {F} _ {s h} ^ {s} [ t ] + (1 - \\mathcal {G} _ {t} ^ {s}) \\times \\mathcal {F} _ {t, N} ^ {s}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 296, + 482, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{t,N}^{s}$ is derived from $\\mathcal{F}_t^s$ through layer normalization, and $\\mathcal{G}_t^s$ is obtained from $\\mathcal{F}_{t,N}^{s}$ via a linear layer followed by a Sigmoid function. $\\mathcal{F}_{sh}^{s}[t]$ is the $t$ -th component of $\\mathcal{F}_{sh}^{s}$ corresponding to the $t$ -th task. The MSST primarily aims to construct comprehensive representations through task-internal interactions. Its architecture is illustrated in Fig. 2 (b) and comprises two main branches: the scan branch and the gating branch. In the scan branch, we integrate a hierarchical global scene structure representation $\\mathcal{F}_h^s$ using a novel MS-Scan mechanism. Simultaneously, a gating signal $\\mathcal{G}^s$ from the gate branch regulates the flow of information. Subsequently, we adjust the channel dimensions of the multi-scale scene representation by applying a linear projection $\\mathcal{P}$ and establish a residual connection between $\\mathcal{F}_h^s$ and the input $\\mathcal{F}_{in}^{s}$ . This process is repeated twice to produce the final output $\\mathcal{F}^s$ :", + "bbox": [ + 89, + 321, + 483, + 564 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} ^ {s} = \\mathcal {F} _ {i n} ^ {s} + \\mathcal {P} \\left(\\mathcal {F} _ {h} ^ {s} \\times \\mathcal {G} ^ {s}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 573, + 482, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Bidirectional Interaction Scan. To maintain task information integrity during interaction with linear complexity, we propose a novel BI-Scan that serves as a cornerstone of BIM, as shown in Fig. 3. This module takes the representations of different tasks $\\{\\mathcal{F}_{t,N}^s\\}_{t = 1}^T$ as input, which is initially partitioned along the channel dimension into dual-branch representations, each with the dimensions $H\\times W\\times \\frac{C}{2}$ . The former undergoes processing through our Forward Scan module, while the latter is reversed along the task dimension before the Backward Scan processing. Specifically, the Forward Scan architecture implements hierarchical cross-task interaction through two scanning modes: Task-First mode and Position-First mode.", + "bbox": [ + 89, + 598, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task-First Mode: Following a similar configuration to SS2D in Vmamba [17], we implement multiple distinct Scan Patterns for each scanning mode to facilitate enhanced task interaction. Without loss of generality, we use the first scan pattern as an example. As shown in Fig. 3, in the task-first mode, the construction of global task information contains two processes: (1) Pattern-guided feature serialization", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9df8236e2bae9e178842067166da01f9f7fdc93fbf8de7b22d8f3e607bd8d9e0.jpg", + "image_caption": [ + "Figure 3. Framework of BI-Scan. It consists of two modules, Forward Scan and Backward Scan, each including Task-First and Position-First cross-task scanning modes." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 903, + 339 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$(\\mathcal{U}_1)$ : each task feature is serialized according to the scan pattern, resulting in $T$ subsequences, each of length $H \\times W$ . (2) Task order-guided subsequence aggregation $(\\mathcal{O}_1)$ : these subsequences are concatenated along the length dimension according to the specified task order, producing a global sequence $S$ of length $T \\times (H \\times W)$ . These processes can be expressed as:", + "bbox": [ + 511, + 420, + 906, + 526 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {1} = \\mathcal {O} _ {1} \\left(\\mathcal {U} _ {1} \\left(\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {2} ^ {s}, \\dots , \\mathcal {F} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(T \\times (H \\times W)\\right) \\times \\frac {C}{2}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 537, + 906, + 558 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This sequence undergoes global modeling through SSM [9], yielding task-level interaction information.", + "bbox": [ + 511, + 569, + 905, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position-First Mode: In the position-first mode, the output of Task-First mode is first restored from the sequence to the input task feature size, recorded as $\\{\\bar{F}_t^s\\}_{t=1}^T \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}$ . Subsequently, the construction of interaction representations also involves two steps: (1) Task order-guided feature serialization $(\\mathcal{O}_2)$ : Spatially aligned task representation tokens are constructed into $H \\times W$ subsequences, each of length $T$ , according to the specified task order. (2) Pattern-guided subsequence aggregation $(\\mathcal{U}_2)$ : These subsequences are concatenated along the length dimension according to the scan pattern, resulting in a sequence of length $(H \\times W) \\times T$ . The above processes can be expressed as:", + "bbox": [ + 511, + 599, + 906, + 796 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {2} = \\mathcal {U} _ {2} \\left(\\mathcal {O} _ {2} \\left(\\bar {\\mathcal {F}} _ {1} ^ {s}, \\bar {\\mathcal {F}} _ {2} ^ {s}, \\dots , \\bar {\\mathcal {F}} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(\\left(H \\times W\\right) \\times T\\right) \\times \\frac {C}{2}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 808, + 906, + 829 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This sequence is also processed through SSM, allowing for the derivation of task-level interaction insights.", + "bbox": [ + 511, + 839, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, the outputs corresponding to all patterns are fused via element-wise summation, yielding the output", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "18818", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/dca41e1f6e1d2fafed6c1200ece592ad6c004262a37659728b0f04e02b43d801.jpg", + "image_caption": [ + "Figure 4. Instructions for MS-Scan. It consists of three distinct operations, Input Handling, Multi-Scale Scanning, and Multi-Scale Fusion. In Multi-Scale Scanning, each branch $\\mathcal{B}_i$ has a distinct scanning size $s_i$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "of Forward Scan, denoted as $\\mathcal{F}_{for}^{s}$ . Finally, the output of the Backward Scan, $\\mathcal{F}_{back}^{s}$ , is reversed along the task dimension and concatenated with $\\mathcal{F}_{for}^{s}$ to yield the refined task representations $\\mathcal{F}_{sh}^{s}$ . In practical application, when employing BI-Scan independently, the selection of Scan Patterns aligns with those used in SS2D [17], specifically, the four directions shown in Fig. 3. When MS-Scan is introduced within the BI-Scan module, we incorporate Scan Patterns of varying scales without increasing computational complexity.", + "bbox": [ + 89, + 385, + 483, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Scale Scan. Mamba processes features by converting them into sequences of tokens. While a variety of tokenization strategies exist for text, image tokenization requires careful consideration of multiple approaches. Previous research has demonstrated that multi-scale processing is particularly effective for image data. To leverage this advantage, We propose the MS-Scan mechanism that requires minimal additional computational cost compared to single-scale scanning, as illustrated in Fig. 4 (a). In this framework, we employ multiple scanning scales, denoted as $\\{s_i\\}_{i=1}^N$ , in multiple branches $\\{\\mathcal{B}_i\\}_{i=1}^N$ , and transform the input image feature $x \\in \\mathbb{R}^{C \\times H \\times W}$ into token sequences of varying dimensions for Mamba modeling. For instance, the initial image feature can be tokenized into a sequence with a total length of $H \\times W$ , where each token has a dimension of $C$ . When applying a scanning scale of $s_i = 2$ , the image is divided into non-overlapping patches, resulting in a tokenized feature sequence with a dimensionality of $4C \\times \\left(\\frac{H}{2} \\times \\frac{W}{2}\\right)$ . Specifically, MS-Scan comprises three key components: input handling, multi-scale scanning, and multi-scale fusion.", + "bbox": [ + 89, + 521, + 483, + 838 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Input Handling: To construct inputs for $N$ different scanning branches $\\mathcal{B} = \\{\\mathcal{B}_i\\}_{i=1}^N$ , we perform two main operations. (1) Channel Split ( $\\mathcal{S}$ ): We begin by splitting the input representation $x$ into $N$ sub-features $\\{x_i\\}_{i=1}^N \\in$", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\mathbb{R}^{m\\times H\\times W}$ along the channel dimension, where $m = \\frac{C}{N}$ . (2) Window Tokenization $(\\mathcal{W}_i)$ : For the $i$ -th branch $\\mathcal{B}_i$ with scanning scale $s_i$ , we first divide $x_{i}$ into $\\frac{H}{s_i}\\times \\frac{W}{s_i}$ nonoverlapping patches, each of size $m\\times s_i\\times s_i$ . Subsequently, we concatenate the pixel feature values in each patch along the channel dimension, resulting in the scan input $\\bar{x}_i$ with a shape of $(m\\times s_i\\times s_i)\\times \\frac{H}{s_i}\\times \\frac{W}{s_i}$ . Ultimately, we obtain the input for all branches:", + "bbox": [ + 511, + 89, + 903, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\{\\bar {x} _ {1}, \\bar {x} _ {2}, \\dots , \\bar {x} _ {N} \\} = \\left\\{\\mathcal {W} _ {1}, \\mathcal {W} _ {2}, \\dots , \\mathcal {W} _ {N} \\right\\} (\\mathcal {S} (x)). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 219, + 903, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Scale Scanning: Following the input handling process, we employ distinct scanning scales to construct a multi-scale scene representation in each branch. For all branches $\\mathcal{B}$ , we utilize the four-way scanning method (SS2D) from VMamba [17] to generate scene features at the specified scale. This method creates four token sequences, each shaped as $C_i \\times (H_i \\times W_i)$ , by scanning the input features $\\bar{x}_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}$ in four directions. The resulting sequences are then processed by SSM [9] and combined to produce the output feature $\\bar{y} = \\{\\bar{y}_i\\}_{i=1}^N$ :", + "bbox": [ + 511, + 244, + 906, + 397 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {y} _ {i} = \\operatorname {S S 2 D} (\\bar {x} _ {i}) \\in \\mathbb {R} ^ {C _ {i} \\times H _ {i} \\times W _ {i}}, i \\in \\{1, 2, \\dots , N \\}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 402, + 903, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-Scale Fusion: In our approach to multi-scale feature fusion, we adopt a methodology that reverses the input handling process, consisting of two key steps. (1) Token Windowing $(\\mathcal{W}_i^{-1})$ : For each branch $\\mathcal{B}_i$ , we split each pixel feature into $s_i \\times s_i$ segments along the channel dimension, resulting in a set of feature segments $\\{\\bar{y}_{i,j}\\}_{j=1}^{s_i \\times s_i} \\in \\mathbb{R}^{m \\times 1 \\times 1}$ . These segments are then concatenated along the spatial dimensions (height and width) to form patches, which are subsequently combined to produce the output for $\\mathcal{B}_i$ . (2) Channel Concatenation $(S^{-1})$ : We concatenate the features from all branches along the channel dimension, yielding the final output feature $y \\in \\mathbb{R}^{C \\times H \\times W}$ :", + "bbox": [ + 511, + 428, + 906, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ny = \\mathcal {S} ^ {- 1} \\left(\\mathcal {W} _ {1} ^ {- 1} (\\bar {y} _ {1}), \\mathcal {W} _ {2} ^ {- 1} (\\bar {y} _ {2}), \\dots , \\mathcal {W} _ {N} ^ {- 1} (\\bar {y} _ {N})\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 632, + 903, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S^{-1}$ and $\\mathcal{W}_i^{-1}$ refer to the inverse operation of $S$ and $\\mathcal{W}_i$ respectively.", + "bbox": [ + 511, + 657, + 903, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Optimization Objective", + "text_level": 1, + "bbox": [ + 511, + 696, + 730, + 712 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We jointly train all tasks to optimize BIM encoder $\\Phi$ , BIM decoder $\\Theta$ , and task heads $\\{H\\}_{i=1}^{T}$ simultaneously. To maintain consistency with previous approaches, we use $L1$ loss for depth estimation and surface normal estimation tasks and the cross-entropy loss for other tasks, therefore, the optimization objective can be expressed as follows:", + "bbox": [ + 511, + 717, + 903, + 808 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\sum_ {t \\in \\mathcal {T}} \\lambda_ {t} \\mathcal {L} _ {t} \\left(\\hat {\\mathbf {Y}} _ {t}, \\mathbf {Y} _ {t}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 816, + 903, + 848 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{T}$ is the set of all tasks, $\\lambda_t$ , $\\mathcal{L}_t$ and $\\mathbf{Y}_t$ are the loss weight, loss function, and task label for image I in task $t$ , respectively.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "18819", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 89, + 215, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 114, + 279, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We performed experiments using the benchmark datasets NYUD-v2 [25] and PASCAL-Context [3]. NYUD-v2 primarily focuses on indoor scenes, with 795 and 654 RGB images for training and testing purposes. Tasks in NYUD-v2 include semantic segmentation, monocular depth estimation, surface normal estimation, and object boundary detection. PASCAL-Context encompasses indoor and outdoor scenes, offering pixel-level labels for tasks like semantic segmentation, human parsing, object boundary detection, surface normal estimation, and saliency detection tasks. The PASCAL-Context dataset contains 4,998 training images and 5,105 test images.", + "bbox": [ + 88, + 138, + 482, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We employ a pretrained SwinLarge Transformer [18] on ImageNet-22K [6] as our encoder. Our models are trained on the NYUD-v2 dataset for 50,000 iterations with a batch size of 4, and on the Pascal Context dataset for 75,000 iterations with a batch size of 6. Across all datasets, we use the Adam optimizer with a learning rate of $5 \\times 10^{-5}$ and a weight decay rate of $1 \\times 10^{-5}$ , alongside a polynomial learning rate scheduler. The preliminary decoder has an output channel number of 768. We follow common practice [16, 35] in resizing the input images and applying data augmentation techniques. Specifically, we resize the input images of NYUD-v2 and PASCAL-Context to $448 \\times 576$ and $512 \\times 512$ , respectively, and apply random color jittering, random cropping, random scaling, and random horizontal flipping.", + "bbox": [ + 89, + 320, + 483, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 559, + 462, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Main Results. Tab. 1 and Tab. 2 report a comparison of the proposed BIM against previous state-of-the-art methods, including MTmamba [16], MQTransformer [32], TaskPrompter [36], $\\mathrm{InvPT}++$ [37], $\\mathrm{InvPT}$ [35], ATRC [2], MTI-Net [29], PAD-Net [31], PSD [40], PAP [42], Cross-Stitch [22] and ASTMT [20] on NYUD-V2 and PASCAL-Context dataset respectively. Notably, the previous best method, i.e., MTmamba, and our BIM are built upon the Transformer-encoder Mamba-decoder architecture with the same backbone. On NYUD-v2, the performance of Semseg is clearly boosted from the previous best, i.e., 55.82 to 57.40 (+1.52). On Pascal-Context, we achieved superior performance on all tasks compared to MTMamba.", + "bbox": [ + 89, + 582, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons on Model Size and Computational Cost. In Tab. 3, we compare the FLOPs and parameter counts with other State-of-the-art methods. With the comparable computational cost, our BIM shows significant outperforms. our light-vision BIM (DBIM), which implements dilated sampling in MS-Scan, shows outperformance with reduced computational cost and parameter numbers, showcasing the advantage of these designs. Specifically, in MS", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/3f8722c62cc42a19babf7cc1b2449546489509c804a92f68afde7c893a74a762.jpg", + "table_caption": [ + "Table 1. Quantitative comparison with other state-of-the-art methods on NYUD-v2 dataset." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
CNN based
Cross-Stitch[22]36.340.629020.8876.38
PAP[42]36.720.617820.8276.42
PSD[40]36.690.624620.8776.42
PAD-Net[31]36.610.627020.8576.38
MTI-Net[29]45.940.536520.2777.86
ATRC[2]46.330.536320.1877.94
Transformer based
InvPT[35]53.660.518319.0478.10
TaskPrompter[36]55.300.515218.4778.20
InvPT++[37]53.850.509618.6778.10
MQTransformer[32]54.840.532519.6778.20
Mamba based
MTMamba[16]55.820.506618.6378.70
BIM(Ours)57.400.473318.5578.72
", + "bbox": [ + 529, + 128, + 890, + 363 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/feb10b2698bdac34ab17f3868588fcdb8d430f6a43d5b874f88991063878cbf7.jpg", + "table_caption": [ + "Table 2. Quantitative comparison with other state-of-the-art methods on Pascal-Context dataset." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF↑Normal mErr ↓Boundary odsF ↑
CNN based
PAD-Net[31]53.6059.6065.8015.3072.50
ASTMT[20]68.0061.1065.7014.7072.40
MTI-Net[29]61.7060.1884.7814.2370.80
ATRC[2]62.6959.4284.7014.2070.96
ATRC-ASPP[2]63.6060.2383.9114.3070.86
ATRC-BMTAS[2]67.6762.9382.2914.2472.42
Transformer based
InvPT[35]79.0367.6184.8114.1573.00
TaskPrompter[36]80.8968.8984.8313.7273.50
InvPT++[37]80.2269.1284.7413.7374.20
MQTransformer[32]78.9367.4183.5814.2173.90
Mamba based
MTmamba[16]81.1172.6284.1414.1478.80
BIM(Ours)81.2573.1584.1414.1379.05
", + "bbox": [ + 516, + 425, + 903, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Scan, we use dilated sampling in each scan branch $\\mathcal{B}$ to lighten the operation of MS-Scan. Specifically: (1) We only sample one token in a window when generating multi-scale sequences from image features, instead of all tokens. (2) When restoring the sequence to image features, we perform linear interpolation between these selected tokens. These operations do not introduce any parameters, and the computational burden is reduced because only a subset of tokens are used for modeling.", + "bbox": [ + 509, + 698, + 906, + 835 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. We qualitatively compared our proposed BIM with the previous best-performing method, as shown in Fig. 5. Our method shows clear improvements in detail, as highlighted in the circled regions.", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "18820", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e4bed02f21f0e9d7922c638ffa22122785dcd318e543231656ca2b51a64263f6.jpg", + "image_caption": [ + "Figure 5. Qualitative comparison with the best-performing method on Pascal-Context. Our method generates better details." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 315 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b0b70a5ca3519cb85bc3c99e67e0873ab745a566c950e12e81e612aa2be0550d.jpg", + "table_caption": [ + "Table 3. Complexity comparisons on NYUD-v2. †denotes the result is from [19]." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑FLOPs (G)↓# Params (M)↓
InvPT[35]53.660.518319.0478.10598402
TaskPrompter[36]55.300.515218.4778.20470392
InvPT++†[37]53.850.509618.6778.10-~402
MTMamba[16]55.820.506618.6378.70541308
DBIM (Ours)56.410.478918.6678.80495290
BIM (Ours)57.400.473318.5578.72547388
", + "bbox": [ + 96, + 405, + 478, + 508 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/297f220a630441873597088e4ac056b5b32573222b5e1d801ea58f24ee81b565.jpg", + "image_caption": [ + "Figure 6. Ablation study on Bidirectional Scan (left) and two scanning modes (right) in BI-Scan." + ], + "image_footnote": [], + "bbox": [ + 101, + 527, + 478, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Different Components. We performed ablation experiments on BI-Scan and MS-Scan using the NYUD-V2 dataset. These experiments all used Swin-Large Transformer as encoder. We use the same STL and MTL base model as in MTMamba. (1) The term \"STL\" indicates that each task is equipped with a separate network, utilizing two Swin Transformer blocks in each decoder stage, and (2) the \"MTL\" indicates that all tasks share a common encoder. (3) \"Baseline\" refers to MTMamba. (4) \"BI-Scan\" indicates that the MS-Scan is replaced with SS2D throughout the entire model. (5) \"MS-Scan\" refers to the implementation of", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/b7f570a6b01f6928d34c73c54e0ba4e33644c666c2940aa253f839fd67c9eba8.jpg", + "table_caption": [ + "Table 4. Effectiveness of different components on NYUD-v2." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑FLOPs # (G) ↓Params (M) ↓
STL54.320.516619.2177.30+0.001075889
MTL53.720.523919.9776.50-1.87466303
Baseline55.820.506618.6378.70+2.38541308
BI-Scan57.110.485618.6678.90+4.02547290
MS-Scan57.400.482918.7978.60+4.00541397
BIM57.400.473318.5578.72+4.83547388
", + "bbox": [ + 517, + 391, + 903, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the same cross-task interaction strategy as that used in the baseline. Lastly, (6) \"BIM\" represents the default method. As depicted in Tab. 4, The BI-Scan improved model performance with fewer parameters, emphasizing the necessity of retaining task-specific details in cross-task interactions. Meanwhile, the MS-Scan enabled multi-granularity scene modeling, further enhancing performance.", + "bbox": [ + 511, + 532, + 905, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Bidirectional Scan in BI-Scan. As shown in Fig. 6 (left), we compared the performance of unidirectional (Uni) and bidirectional (Bi) cross-task scanning. The results show that the introduction of bidirectional scanning significantly enhances the cross-task modeling capability, thereby improving the overall performance of the model with only a negligible increase in computational cost.", + "bbox": [ + 511, + 640, + 906, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Two Scan Modes in BI-Scan. As shown in Fig. 6 (right), we conducted ablation experiments on the Task-First Scan mode and Position-First Scan mode within BI-Scan. The results indicate that both Task-First (Task) and Position-First (Position) interaction modes achieve superior performance compared to the baseline (Base) at a lower computational cost. Furthermore, employing both modes together (All) results in further performance improvements. Ultimately, we incorporate both Task-First mode and Position-First mode as the final configuration of", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "18821", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "BI-Scan for BIM.", + "text_level": 1, + "bbox": [ + 93, + 90, + 210, + 104 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Task Order in BI-Scan. We experiment on NYUD-v2 to study the effect of the task order during scan interaction. In addition to fixed task orders, we also implemented a random task order modeling approach (Random), in which the order of tasks in BI-Scan is randomly shuffled during both training and testing. For clarity, we refer to the tasks Semseg, Depth, Normal, and Boundary as S, D, N, and B, respectively. As shown in Tab. 5, the overall performance of various fixed order models is relatively similar, while the Random setting exhibits a slight decrease in overall performance, which remains within an acceptable range. This indicates that our proposed BI-Scan demonstrates robustness to the task order. Ultimately, we adopt the first task order setting as the final configuration of BI-Scan for BIM.", + "bbox": [ + 93, + 107, + 480, + 316 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5c543644021481dd9c778aa7ca324a6ac58176cce840ea9799861ad9660406a1.jpg", + "table_caption": [ + "Table 5. Ablation on task order on NYUD-v2." + ], + "table_footnote": [], + "table_body": "
OrderSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
S-D-N-B57.110.485618.6678.90+4.02
N-S-D-B56.150.483218.6679.00+3.72
B-S-D-N56.970.483618.7378.90+3.96
S-N-D-B57.520.484318.7178.90+4.20
Random55.410.486118.6978.90+3.17
", + "bbox": [ + 94, + 357, + 480, + 463 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation on Scan Scale and Number of MS-Scan. We experimented with the impact of varying scanning scales and numbers on model performance, as shown in Tab. 6 and Tab. 7. Experiments were conducted on the NYUD-v2 dataset with Swin-Large Transformer as encoder. We experimented with three different settings for scale ablation: all MFRs use $\\{1,2\\}$ , $\\{1,3\\}$ , and $\\{1,4\\}$ two scanning scales, respectively. For number ablation, We compared three different settings, Type 1: all three MFRs use $\\{1\\}$ scale; Type 2: all three MFRs use $\\{1,4\\}$ two scales; Type 3: three MFRs use $\\{1,2\\}$ , $\\{1,2,4\\}$ and $\\{1,2,4,6\\}$ , respectively. The results showed that in multi-scale scanning, employing appropriate scanning scale partitions can effectively enhance overall performance. Significantly, all settings yield notable performance enhancements. Ultimately, we adopt $\\{1,4\\}$ scale setting for all MFRs as the final configuration.", + "bbox": [ + 93, + 479, + 480, + 720 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/df744a677ae74b7f539a91b34fbce08bfed44ef04e0e567e03ec957ea7ae3bf9.jpg", + "table_caption": [ + "Table 6. Different scanning scales of MS-Scan on NYUD-v2." + ], + "table_footnote": [], + "table_body": "
Scan ScaleSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
{1,2}57.390.475618.6078.80+4.68
{1,3}57.530.473918.6878.50+4.62
{1,4}57.400.473318.5578.72+4.83
", + "bbox": [ + 94, + 761, + 478, + 837 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance on Different Encoders. We evaluate the effect of model size on experimental performance, as presented in Tab. 8. All experiments were conducted on the", + "bbox": [ + 93, + 854, + 480, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5cbf69a1b10b62cfd697f6e3926cfcebc7fb6c11e4cd419d885422147c311bea.jpg", + "table_caption": [ + "Table 7. Different scanning number of MS-Scan on NYUD-v2." + ], + "table_footnote": [], + "table_body": "
Scan NumberSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
Type 157.110.485618.6678.90+4.02
Type 257.400.473318.5578.72+4.83
Type 356.760.475618.5978.70+4.35
", + "bbox": [ + 519, + 114, + 901, + 186 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "NYUD-v2 dataset. We compare our method with the previous best-performing model, MTMamba, using two different encoders: Swin-Base Transformer (denoted as '-Base') and Swin-Large Transformer (denoted as '-Large'). The results suggest that models with greater capacities typically exhibit superior performance. Furthermore, our approach has demonstrated superior performance across all encoder.", + "bbox": [ + 516, + 217, + 903, + 321 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a1bb0864b5ead5effa0e52b37ff71d4e3ff4f69c168ba3a9f2f7a5e16474904d.jpg", + "table_caption": [ + "Table 8. Performance on different encoders." + ], + "table_footnote": [], + "table_body": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
MTMamba-Base53.620.512619.2877.70
BIM-Base54.320.491518.8977.74
MTMamba-Large55.820.506618.6378.70
BIM-Large57.400.473318.5578.72
", + "bbox": [ + 521, + 364, + 893, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 516, + 488, + 630, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduce the Bidirectional Interaction Mamba (BIM) framework, a simple yet effective solution designed to address the inherent conflict between adequacy and complexity in cross-task interactions for multi-task dense prediction. We first propose a Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task information as bidirectional sequences to model cross-task interactions efficiently. By integrating task-first and position-first modes within linear complexity framework, BI-Scan preserves the integrity of task-specific information during interaction while maintaining computational efficiency. We further develop a Multi-Scale Scan (MS-Scan) mechanism, which captures scene structure information at multiple scales, alleviating the difficulty of feature learning and enhancing nuanced cross-task feature interactions. Extensive experiments show our method significantly enhances model performance.", + "bbox": [ + 516, + 515, + 903, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 516, + 776, + 691, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80, and the Public Welfare Research Program of Ningbo City under Grant 2024S063.", + "bbox": [ + 516, + 801, + 903, + 886 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "18822", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Deblina Bhattacharjee, Tong Zhang, Sabine Susstrunk, and Mathieu Salzmann. Mult: An end-to-end multitask learning transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12031-12041, 2022. 1, 2", + "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15869-15878, 2021. 6", + "[3] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6", + "[4] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 2", + "[5] Michael Crawshaw. Multi-task learning with deep neural networks: A survey. CoRR, abs/2009.09796, 2020. 1", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[7] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2", + "[8] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3205-3214, 2019. 1", + "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2024. 2, 4, 5", + "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. CoRR, abs/2111.00396, 2021. 2", + "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 2", + "[12] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3", + "[13] Woeseong Jeong and Kuk-Jin Yoon. Quantifying task priority for multi-task optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 363-372, 2024. 2" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 2", + "[15] Wei-Hong Li, Steven McDonagh, Ales Leonardis, and Hakan Bilen. Multi-task learning with 3d-aware regularization. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[16] Baijiong Lin, Weisen Jiang, Pengguang Chen, Yu Zhang, Shu Liu, and Ying-Cong Chen. MTMamba: Enhancing multi-task dense scene understanding by mamba-based decoders. In European Conference on Computer Vision, 2024. 1, 2, 3, 6, 7", + "[17] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. CoRR, abs/2401.10166, 2024. 3, 4, 5", + "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6", + "[19] Yuxiang Lu, Shengcao Cao, and Yu-Xiong Wang. Swiss army knife: Synergizing biases in knowledge from vision foundation models for multi-task learning. arXiv preprint arXiv:2410.14633, 2024. 7", + "[20] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1851-1860, 2019. 6", + "[21] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 2", + "[22] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994-4003, 2016. 6", + "[23] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 2", + "[24] Sara Shoouri, Mingyu Yang, Zichen Fan, and Hun-Seok Kim. Efficient computation sharing for multi-task visual scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17130-17141, 2023. 1, 2", + "[25] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6", + "[26] Dimitrios Sinodinos and Narges Armanfard. Ema-net: Efficient multitask affinity learning for dense scene predictions. arXiv preprint arXiv:2401.11124, 2024. 1" + ], + "bbox": [ + 93, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "18823", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Jimmy T. H. Smith, Andrew Warrington, and Scott W. Linderman. Simplified state space layers for sequence modeling. CoRR, abs/2208.04933, 2022. 2", + "[28] Guolei Sun, Thomas Probst, Danda Pani Paudel, Nikola Popovic, Menelaos Kanakis, Jagruti Patel, Dengxin Dai, and Luc Van Gool. Task switching network for multi-task learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8291-8300, 2021. 1", + "[29] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 527-543. Springer, 2020. 2, 6", + "[30] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 1", + "[31] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 675-684, 2018. 1, 2, 6", + "[32] Yangyang Xu, Xiangtai Li, Haobo Yuan, Yibo Yang, and Lefei Zhang. Multi-task learning with multi-query transformer for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6", + "[33] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. CoRR, abs/2403.17695, 2024. 3", + "[34] Feiyang Ye, Yueming Lyu, Xuehao Wang, Yu Zhang, and Ivor Tsang. Adaptive stochastic gradient algorithm for black-box multi-objective learning. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[35] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In European Conference on Computer Vision, pages 514-530. Springer, 2022. 1, 2, 6, 7", + "[36] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7", + "[37] Hanrong Ye and Dan Xu. Invpt++: Inverted pyramid multitask transformer for visual scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 6, 7", + "[38] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2", + "[39] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Rethinking of feature interaction for multi-task learning on dense prediction. arXiv preprint arXiv:2312.13514, 2023. 1, 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[40] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 6", + "[41] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024.3", + "[42] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4514–4523, 2020. 6", + "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 2, 3" + ], + "bbox": [ + 514, + 90, + 903, + 345 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "18824", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_model.json b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5dfcb57c24cc52426877686970f062544e69e4c8 --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_model.json @@ -0,0 +1,2013 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.1, + 0.13, + 0.9, + 0.175 + ], + "angle": 0, + "content": "Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction" + }, + { + "type": "text", + "bbox": [ + 0.166, + 0.203, + 0.842, + 0.224 + ], + "angle": 0, + "content": "Mang Cao\\(^{1}\\) Sanping Zhou\\(^{1*}\\) Yizhe Li\\(^{1}\\) Ye Deng\\(^{2}\\) Wenli Huang\\(^{3}\\) Le Wang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.19, + 0.228, + 0.821, + 0.282 + ], + "angle": 0, + "content": "\\(^{1}\\)National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.1, + 0.282, + 0.908, + 0.318 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics \\(^{3}\\) School of Electronic and Information Engineering, Ningbo University of Technology" + }, + { + "type": "text", + "bbox": [ + 0.3, + 0.327, + 0.71, + 0.341 + ], + "angle": 0, + "content": "3298592685@stu.xjtu.edu.cn, spzhou@xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.377, + 0.327, + 0.393 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.41, + 0.486, + 0.775 + ], + "angle": 0, + "content": "Sufficient cross-task interaction is crucial for success in multi-task dense prediction. However, sufficient interaction often results in high computational complexity, forcing existing methods to face the trade-off between interaction completeness and computational efficiency. To address this limitation, this work proposes a Bidirectional Interaction Mamba (BIM), which incorporates novel scanning mechanisms to adapt the Mamba modeling approach for multi-task dense prediction. On the one hand, we introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task-specific representations as bidirectional sequences during interaction. By integrating task-first and position-first scanning modes within a unified linear complexity architecture, BI-Scan efficiently preserves critical cross-task information. On the other hand, we employ a Multi-Scale Scan (MS-Scan) mechanism to achieve multi-granularity scene modeling. This design not only meets the diverse granularity requirements of various tasks but also enhances nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our BIM vs its state-of-the-art competitors. Codes are available online: https://github.com/mmm-cc/BIM_for_MTL." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.804, + 0.222, + 0.819 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.83, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Multi-task dense prediction is a critical visual task designed to simultaneously predict outputs for various pixel-level tasks, such as semantic segmentation, depth predic" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.378, + 0.905, + 0.558 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.579, + 0.907, + 0.705 + ], + "angle": 0, + "content": "Figure 1. Illustration of different cross-task interaction strategies. (a) Pair-Wise [8, 28, 31]: each task can access the complete information of other tasks, while the interaction complexity is \\( O(n^{2}) \\). (b) Bridge Representation [1, 16, 24, 39]: the interaction complexity is \\( O(n) \\), yet, each task is limited to accessing only a partial subset of information from other tasks during interaction, which inevitably results in the loss of valuable details. (c) Ours Cross-Task Scan: It preserves the integrity of task information during interaction, with an interaction complexity of \\( O(n) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.839 + ], + "angle": 0, + "content": "tion, surface normal estimation, and saliency detection. Existing approaches made mainly efforts [5, 30] to refine task-general representations, obtained from the encoder, into task-specific representations. In these approaches, cross-task interaction has been proven to be significant in multi-task dense prediction [26, 31, 35, 39], as it can enhance the robustness and effectiveness of task representations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Initially, multi-task dense prediction methods primarily utilized Convolutional Neural Networks (CNNs) with pairwise interactions, as illustrated in Fig. 1 (a). These networks [8, 28, 31] were carefully designed with specialized" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "branches for each task, augmented by interaction modules that facilitated information exchanges between each pair of tasks. These pair-wise interaction approaches maintain the integrity of task information during interaction, promoting fine-grained cross-task interaction. While, the inherently limited receptive fields of CNN architectures often led to insufficient spatial interactions. In response to these challenges, transformer-style networks [35] demonstrate remarkable capability in modeling long-range dependencies. This capability substantially improves the interaction effectiveness of models in handling multi-task scenarios, leading to enhanced performance. However, these sufficient cross-task interactions usually introduce high interaction complexity, which typically increases quadratically with the number of tasks. The high interaction complexity limits their applicability in certain powerful models, particularly transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.594 + ], + "angle": 0, + "content": "To enhance the efficiency of cross-task interactions, a crowd of methods [1, 16, 24, 39] with linear interaction complexity has been proposed, as depicted in Fig. 1 (b). Typically, these methods strategically select a representative task [1, 24] or extract representative information from all tasks [16, 39] to construct a compact bridge representation that aims to preserve essential interaction information. However, in these methods, the necessity of maintaining linear interaction complexity inherently limits the representational capacity of these bridge representations. This information compression inevitably leads to the degradation of task features, potentially discarding critical interaction details and ultimately resulting in suboptimal performance. Therefore, how to enhance sufficient interaction capability while maintaining linear interaction complexity remains an unresolved issue." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In response to these challenges, based on Mamba architecture, we propose a simple yet effective Bidirectional Interaction Mamba (BIM) method that achieves comprehensive preservation of task information during interaction while maintaining linear interaction complexity. Specifically, we introduce a Bidirectional Interaction Scan (BI-Scan) mechanism to achieve sufficient cross-task interaction and a Multi-Scale Scan (MS-Scan) mechanism to deliver a comprehensive visual representation. The BI-Scan mechanism constructs each task representation into sequences and deploys a bidirectional scanning strategy for modeling cross-task interactions, efficiently mitigating information loss while preserving linear interaction complexity. Furthermore, BI-Scan enables multi-level cross-task interactions by integrating task-first and position-first scanning modes, facilitating more nuanced modeling of task relationships. In the MS-Scan, features are initially partitioned into multiple spaces, where the scene structure of images is modeled at various scales. Specific tasks subsequently integrate multi-scale scene structural information to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "enhance task-specific representations. During cross-task interaction, different task branches adaptively integrate interaction representations from multiple scales, addressing the varying granularity demands of each task for scene structural information." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.168, + 0.872, + 0.181 + ], + "angle": 0, + "content": "The main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.183, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- We propose a simple yet effective Bidricational Interaction Mamba (BIM) for multi-task dense prediction, enabling comprehensive cross-task interaction while preserving linear interaction complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.243, + 0.905, + 0.317 + ], + "angle": 0, + "content": "- We introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism to maintain the integrity of task information during interactions. The BI-Scan incorporates two scanning modes: task-first mode and position-first mode, facilitating multi-level cross-task interactions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.318, + 0.905, + 0.393 + ], + "angle": 0, + "content": "- We design a Multi-Scale Scan (MS-Scan) mechanism to capture scene structural information at multiple scales, thereby addressing the diverse granularity requirements of various tasks and enhancing more detailed cross-task interactions in multi-task dense prediction." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.408, + 0.655, + 0.423 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Multi-Task Learning. Most existing multi-task learning works focused on training optimization and network design. Optimization approaches balance tasks via gradient manipulation [13, 23, 34, 38] or loss-balancing [4, 14] to coordinate resource allocation for various tasks during training. The approaches of network design seek to enhance task representation learning by developing diverse mechanisms. Some CNN-based methods [29, 31] manually design interaction mechanisms to extract useful information across tasks. With the advancement of Transformer, current methods have gained improved global task interaction capabilities to enhance task representation efficiency. They enable superior pairwise [1, 24] or global [15, 32, 35, 36] task interactions. To mitigate computational complexity, MT-Mamba [16] introduce Mamba [9] to multi-task learning, which showcases effective long sequence modeling capabilities and achieving satisfactory performance. However, achieving both efficient and sufficient cross-task interaction remains a challenge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "State Space Models. In efficient long-range dependency modeling methods, state space models [11, 27] has become a striking alternative to Transformers. [10] proposed a Structured State Space Sequence (S4) model based on a new parameterization, which alleviates the computational and memory efficiency issues faced by SSM. Subsequently, numerous efforts [7, 21] are dedicated to bridging the performance disparity between SSMs and Transformers. Mamba [9] introduced an input-based parameterization method and hardware-aware algorithm, achieving performance on par with Transformers in natural language processing. This success inspired various efforts [43] towards" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "18816" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.499, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.299, + 0.356, + 0.312 + ], + "angle": 0, + "content": "(a) BIM for MTL" + }, + { + "type": "image", + "bbox": [ + 0.502, + 0.091, + 0.905, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.653, + 0.299, + 0.741, + 0.312 + ], + "angle": 0, + "content": "(b) MFR Block" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 2. Framework of the proposed BIM for multi-task dense prediction. (a) Overview of BIM, illustrating with depth estimation and surface normal estimation tasks. (b) Details of the MFR block, including T task-specific MSST blocks and a task-shared BCFR block." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.379, + 0.336, + 0.393 + ], + "angle": 0, + "content": "Mamba's adaptation for visual tasks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.484, + 0.575 + ], + "angle": 0, + "content": "The preservation of comprehensive image structural information poses a critical challenge in Mamba's sequential processing model, which has attracted considerable attention and effort [12, 17, 33, 41, 43]. VMamba [17] proposes the 2D Selective Scan (SS2D), a four-way scanning mechanism tailored for spatial domain traversal, aimed at enhancing Mamba's image modeling capabilities. Subsequent research studies have explored various scan patterns and combinations tailored to different tasks or scenarios [12, 33, 41]. However, by utilizing fixed token sizes, these methods overlook the importance of hierarchical spatial structural information in visual tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.587, + 0.182, + 0.601 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We first outline the overall architecture of BIM for multitask dense prediction in Section 3.1, Subsequently, Section 3.2 focuses on the BIM Encoder, while Section 3.3 examines the BIM Decoder and its two scanning mechanisms: BI-Scan and MS-Scan. Finally, Section 3.4 discusses the optimization objectives." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.295, + 0.724 + ], + "angle": 0, + "content": "3.1. Framework Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.732, + 0.484, + 0.821 + ], + "angle": 0, + "content": "As shown Fig. 2 (a), the proposed BIM framework consists of three main components: a task-shared encoder \\(\\Phi\\) for extracting task-generic representations, a Mamba decoder \\(\\Theta\\) for refining task features and T task-specific heads \\(\\{H_t\\}_{t=1}^T\\) generating predictions \\(\\{\\hat{\\mathbf{Y}}_t\\}_{t=1}^T\\) for individual tasks. This can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.826, + 0.482, + 0.864 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\left\\{F _ {1}, F _ {2}, \\dots , F _ {T} \\right\\} = \\Theta (\\Phi (\\mathbf {I})), \\\\ \\hat {\\mathbf {Y}} _ {t} = H _ {t} \\left(F _ {t}\\right) = \\mathcal {P} _ {t} \\left(\\Psi \\left(F _ {t}\\right)\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}\\) denotes the RGB input image, \\(F_{t}\\) denotes the features generated by the decoder, corresponding" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.378, + 0.907, + 0.485 + ], + "angle": 0, + "content": "to the distinct task \\( t \\). \\( \\hat{\\mathbf{Y}}_t \\) represents the prediction for task \\( t \\), which has the same height \\( H \\) and width \\( W \\) as I. \\( \\Psi \\) denotes a module designed to expand the task feature resolution to \\( H \\times W \\), which consists of a linear projection and a reshape operation. \\( \\mathcal{P}_t \\) represents a linear layer that projects the feature channels to a specific number required by the corresponding task." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.493, + 0.656, + 0.507 + ], + "angle": 0, + "content": "3.2. BIM Encoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.907, + 0.651 + ], + "angle": 0, + "content": "The encoder shares similarities with other methods [16]. We utilized a pretrained Swin Transformer [18] to extract task-generic features, which begins by dividing the input image \\(\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}\\) into \\(\\frac{H}{w} \\times \\frac{W}{w}\\) tokens of dimension \\(C\\) through patch partitioning and linear layers, where \\(w\\) denotes the partition size. These tokens are then processed through multiple stages involving alternating patch merging and Swin Transformer block processing, ultimately yielding hierarchical image representations:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.66, + 0.905, + 0.678 + ], + "angle": 0, + "content": "\\[\n\\mathbf {G} = \\left\\{\\mathbf {G} ^ {1}, \\mathbf {G} ^ {2}, \\mathbf {G} ^ {3}, \\mathbf {G} ^ {4} \\right\\} = \\Phi (\\mathrm {I}), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.688, + 0.906, + 0.764 + ], + "angle": 0, + "content": "where \\(\\mathbf{G}\\) represents the task-generic features extracted from the encoder \\(\\Phi\\). In our practical implementation, we utilize a partition size of 4, resulting in the following shapes for \\(\\mathbf{G}\\): \\(C \\times \\frac{H}{4} \\times \\frac{W}{4}\\), \\(2C \\times \\frac{H}{8} \\times \\frac{W}{8}\\), \\(4C \\times \\frac{H}{16} \\times \\frac{W}{16}\\) and \\(8C \\times \\frac{H}{32} \\times \\frac{W}{32}\\), respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.772, + 0.656, + 0.786 + ], + "angle": 0, + "content": "3.3. BIM Decoder" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.907, + 0.904 + ], + "angle": 0, + "content": "The BIM Decoder consists of three Mamba Feature Refinement (MFR) blocks, as depicted in Fig. 2 (a). These MFRs are designed to bridge the gap between task-generic and task-specific representations. They generate task representations \\(\\{\\mathbf{F}_1^i,\\mathbf{F}_2^i,\\ldots ,\\mathbf{F}_T^i\\}_{i = 1}^3\\) corresponding to the first three encoder stages, with dimensions of \\(4C\\times \\frac{H}{16}\\times \\frac{W}{16},2C\\times \\frac{H}{8}\\times \\frac{W}{8}\\), and \\(C\\times \\frac{H}{4}\\times \\frac{W}{4}\\), respectively. Finally, the last" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.091, + 0.483, + 0.122 + ], + "angle": 0, + "content": "refined features \\(\\{\\mathbf{F}_t^3\\}_{t = 1}^T\\) are input into the task heads as \\(\\{F_{t}\\}_{t = 1}^{T}\\) to generate the final predictions \\(\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^T\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.29 + ], + "angle": 0, + "content": "We introduce two specialized Mamba blocks, Bidirectional Cross-Task Feature Refine (BCFR) and Multi-Scale Single-task (MSST) in each MFR block, as illustrated in Fig. 2 (b). The BCFR Block is designed to facilitate effective cross-task interaction while maintaining the integrity of task information. It begins by concatenating features \\(\\{\\mathcal{F}_t^s\\}_{t=1}^T\\) from different task branches to construct task-shared features \\(\\mathcal{F}_{sh}^s\\) using the BI-Scan mechanism. Subsequently, each task branch adaptively merges \\(\\mathcal{F}_{sh}^s\\) by a selection value \\(\\mathcal{G}^s\\) to obtain fine-tuned task representations \\(\\mathbf{F}^s = \\{\\mathbf{F}_t^s\\}_{t=1}^T\\):" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.297, + 0.483, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} _ {t} ^ {s} = \\mathcal {F} _ {t} ^ {s} + \\mathcal {G} _ {t} ^ {s} \\times \\mathcal {F} _ {s h} ^ {s} [ t ] + (1 - \\mathcal {G} _ {t} ^ {s}) \\times \\mathcal {F} _ {t, N} ^ {s}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.485, + 0.565 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{t,N}^{s}\\) is derived from \\(\\mathcal{F}_t^s\\) through layer normalization, and \\(\\mathcal{G}_t^s\\) is obtained from \\(\\mathcal{F}_{t,N}^{s}\\) via a linear layer followed by a Sigmoid function. \\(\\mathcal{F}_{sh}^{s}[t]\\) is the \\(t\\)-th component of \\(\\mathcal{F}_{sh}^{s}\\) corresponding to the \\(t\\)-th task. The MSST primarily aims to construct comprehensive representations through task-internal interactions. Its architecture is illustrated in Fig. 2 (b) and comprises two main branches: the scan branch and the gating branch. In the scan branch, we integrate a hierarchical global scene structure representation \\(\\mathcal{F}_h^s\\) using a novel MS-Scan mechanism. Simultaneously, a gating signal \\(\\mathcal{G}^s\\) from the gate branch regulates the flow of information. Subsequently, we adjust the channel dimensions of the multi-scale scene representation by applying a linear projection \\(\\mathcal{P}\\) and establish a residual connection between \\(\\mathcal{F}_h^s\\) and the input \\(\\mathcal{F}_{in}^{s}\\). This process is repeated twice to produce the final output \\(\\mathcal{F}^s\\):" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.574, + 0.483, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} ^ {s} = \\mathcal {F} _ {i n} ^ {s} + \\mathcal {P} \\left(\\mathcal {F} _ {h} ^ {s} \\times \\mathcal {G} ^ {s}\\right). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.794 + ], + "angle": 0, + "content": "Bidirectional Interaction Scan. To maintain task information integrity during interaction with linear complexity, we propose a novel BI-Scan that serves as a cornerstone of BIM, as shown in Fig. 3. This module takes the representations of different tasks \\(\\{\\mathcal{F}_{t,N}^s\\}_{t = 1}^T\\) as input, which is initially partitioned along the channel dimension into dual-branch representations, each with the dimensions \\(H\\times W\\times \\frac{C}{2}\\). The former undergoes processing through our Forward Scan module, while the latter is reversed along the task dimension before the Backward Scan processing. Specifically, the Forward Scan architecture implements hierarchical cross-task interaction through two scanning modes: Task-First mode and Position-First mode." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Task-First Mode: Following a similar configuration to SS2D in Vmamba [17], we implement multiple distinct Scan Patterns for each scanning mode to facilitate enhanced task interaction. Without loss of generality, we use the first scan pattern as an example. As shown in Fig. 3, in the task-first mode, the construction of global task information contains two processes: (1) Pattern-guided feature serialization" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.905, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.352, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Figure 3. Framework of BI-Scan. It consists of two modules, Forward Scan and Backward Scan, each including Task-First and Position-First cross-task scanning modes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.421, + 0.907, + 0.527 + ], + "angle": 0, + "content": "\\((\\mathcal{U}_1)\\): each task feature is serialized according to the scan pattern, resulting in \\(T\\) subsequences, each of length \\(H \\times W\\). (2) Task order-guided subsequence aggregation \\((\\mathcal{O}_1)\\): these subsequences are concatenated along the length dimension according to the specified task order, producing a global sequence \\(S\\) of length \\(T \\times (H \\times W)\\). These processes can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.538, + 0.907, + 0.559 + ], + "angle": 0, + "content": "\\[\nS _ {1} = \\mathcal {O} _ {1} \\left(\\mathcal {U} _ {1} \\left(\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {2} ^ {s}, \\dots , \\mathcal {F} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(T \\times (H \\times W)\\right) \\times \\frac {C}{2}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.599 + ], + "angle": 0, + "content": "This sequence undergoes global modeling through SSM [9], yielding task-level interaction information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.601, + 0.907, + 0.797 + ], + "angle": 0, + "content": "Position-First Mode: In the position-first mode, the output of Task-First mode is first restored from the sequence to the input task feature size, recorded as \\(\\{\\bar{F}_t^s\\}_{t=1}^T \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}\\). Subsequently, the construction of interaction representations also involves two steps: (1) Task order-guided feature serialization \\((\\mathcal{O}_2)\\): Spatially aligned task representation tokens are constructed into \\(H \\times W\\) subsequences, each of length \\(T\\), according to the specified task order. (2) Pattern-guided subsequence aggregation \\((\\mathcal{U}_2)\\): These subsequences are concatenated along the length dimension according to the scan pattern, resulting in a sequence of length \\((H \\times W) \\times T\\). The above processes can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.809, + 0.907, + 0.83 + ], + "angle": 0, + "content": "\\[\nS _ {2} = \\mathcal {U} _ {2} \\left(\\mathcal {O} _ {2} \\left(\\bar {\\mathcal {F}} _ {1} ^ {s}, \\bar {\\mathcal {F}} _ {2} ^ {s}, \\dots , \\bar {\\mathcal {F}} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(\\left(H \\times W\\right) \\times T\\right) \\times \\frac {C}{2}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.87 + ], + "angle": 0, + "content": "This sequence is also processed through SSM, allowing for the derivation of task-level interaction insights." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Subsequently, the outputs corresponding to all patterns are fused via element-wise summation, yielding the output" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "18818" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.359 + ], + "angle": 0, + "content": "Figure 4. Instructions for MS-Scan. It consists of three distinct operations, Input Handling, Multi-Scale Scanning, and Multi-Scale Fusion. In Multi-Scale Scanning, each branch \\(\\mathcal{B}_i\\) has a distinct scanning size \\(s_i\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.522 + ], + "angle": 0, + "content": "of Forward Scan, denoted as \\(\\mathcal{F}_{for}^{s}\\). Finally, the output of the Backward Scan, \\(\\mathcal{F}_{back}^{s}\\), is reversed along the task dimension and concatenated with \\(\\mathcal{F}_{for}^{s}\\) to yield the refined task representations \\(\\mathcal{F}_{sh}^{s}\\). In practical application, when employing BI-Scan independently, the selection of Scan Patterns aligns with those used in SS2D [17], specifically, the four directions shown in Fig. 3. When MS-Scan is introduced within the BI-Scan module, we incorporate Scan Patterns of varying scales without increasing computational complexity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Multi-Scale Scan. Mamba processes features by converting them into sequences of tokens. While a variety of tokenization strategies exist for text, image tokenization requires careful consideration of multiple approaches. Previous research has demonstrated that multi-scale processing is particularly effective for image data. To leverage this advantage, We propose the MS-Scan mechanism that requires minimal additional computational cost compared to single-scale scanning, as illustrated in Fig. 4 (a). In this framework, we employ multiple scanning scales, denoted as \\(\\{s_i\\}_{i=1}^N\\), in multiple branches \\(\\{\\mathcal{B}_i\\}_{i=1}^N\\), and transform the input image feature \\(x \\in \\mathbb{R}^{C \\times H \\times W}\\) into token sequences of varying dimensions for Mamba modeling. For instance, the initial image feature can be tokenized into a sequence with a total length of \\(H \\times W\\), where each token has a dimension of \\(C\\). When applying a scanning scale of \\(s_i = 2\\), the image is divided into non-overlapping patches, resulting in a tokenized feature sequence with a dimensionality of \\(4C \\times \\left(\\frac{H}{2} \\times \\frac{W}{2}\\right)\\). Specifically, MS-Scan comprises three key components: input handling, multi-scale scanning, and multi-scale fusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Input Handling: To construct inputs for \\(N\\) different scanning branches \\(\\mathcal{B} = \\{\\mathcal{B}_i\\}_{i=1}^N\\), we perform two main operations. (1) Channel Split (\\(\\mathcal{S}\\)): We begin by splitting the input representation \\(x\\) into \\(N\\) sub-features \\(\\{x_i\\}_{i=1}^N \\in\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.09, + 0.905, + 0.213 + ], + "angle": 0, + "content": "\\(\\mathbb{R}^{m\\times H\\times W}\\) along the channel dimension, where \\(m = \\frac{C}{N}\\). (2) Window Tokenization \\((\\mathcal{W}_i)\\): For the \\(i\\)-th branch \\(\\mathcal{B}_i\\) with scanning scale \\(s_i\\), we first divide \\(x_{i}\\) into \\(\\frac{H}{s_i}\\times \\frac{W}{s_i}\\) nonoverlapping patches, each of size \\(m\\times s_i\\times s_i\\). Subsequently, we concatenate the pixel feature values in each patch along the channel dimension, resulting in the scan input \\(\\bar{x}_i\\) with a shape of \\((m\\times s_i\\times s_i)\\times \\frac{H}{s_i}\\times \\frac{W}{s_i}\\). Ultimately, we obtain the input for all branches:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.22, + 0.905, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\{\\bar {x} _ {1}, \\bar {x} _ {2}, \\dots , \\bar {x} _ {N} \\} = \\left\\{\\mathcal {W} _ {1}, \\mathcal {W} _ {2}, \\dots , \\mathcal {W} _ {N} \\right\\} (\\mathcal {S} (x)). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.245, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Multi-Scale Scanning: Following the input handling process, we employ distinct scanning scales to construct a multi-scale scene representation in each branch. For all branches \\(\\mathcal{B}\\), we utilize the four-way scanning method (SS2D) from VMamba [17] to generate scene features at the specified scale. This method creates four token sequences, each shaped as \\(C_i \\times (H_i \\times W_i)\\), by scanning the input features \\(\\bar{x}_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}\\) in four directions. The resulting sequences are then processed by SSM [9] and combined to produce the output feature \\(\\bar{y} = \\{\\bar{y}_i\\}_{i=1}^N\\):" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.403, + 0.905, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\bar {y} _ {i} = \\operatorname {S S 2 D} (\\bar {x} _ {i}) \\in \\mathbb {R} ^ {C _ {i} \\times H _ {i} \\times W _ {i}}, i \\in \\{1, 2, \\dots , N \\}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Multi-Scale Fusion: In our approach to multi-scale feature fusion, we adopt a methodology that reverses the input handling process, consisting of two key steps. (1) Token Windowing \\((\\mathcal{W}_i^{-1})\\): For each branch \\(\\mathcal{B}_i\\), we split each pixel feature into \\(s_i \\times s_i\\) segments along the channel dimension, resulting in a set of feature segments \\(\\{\\bar{y}_{i,j}\\}_{j=1}^{s_i \\times s_i} \\in \\mathbb{R}^{m \\times 1 \\times 1}\\). These segments are then concatenated along the spatial dimensions (height and width) to form patches, which are subsequently combined to produce the output for \\(\\mathcal{B}_i\\). (2) Channel Concatenation \\((S^{-1})\\): We concatenate the features from all branches along the channel dimension, yielding the final output feature \\(y \\in \\mathbb{R}^{C \\times H \\times W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.633, + 0.905, + 0.653 + ], + "angle": 0, + "content": "\\[\ny = \\mathcal {S} ^ {- 1} \\left(\\mathcal {W} _ {1} ^ {- 1} (\\bar {y} _ {1}), \\mathcal {W} _ {2} ^ {- 1} (\\bar {y} _ {2}), \\dots , \\mathcal {W} _ {N} ^ {- 1} (\\bar {y} _ {N})\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.658, + 0.905, + 0.689 + ], + "angle": 0, + "content": "where \\(S^{-1}\\) and \\(\\mathcal{W}_i^{-1}\\) refer to the inverse operation of \\(S\\) and \\(\\mathcal{W}_i\\) respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.731, + 0.713 + ], + "angle": 0, + "content": "3.4. Optimization Objective" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.905, + 0.809 + ], + "angle": 0, + "content": "We jointly train all tasks to optimize BIM encoder \\(\\Phi\\), BIM decoder \\(\\Theta\\), and task heads \\(\\{H\\}_{i=1}^{T}\\) simultaneously. To maintain consistency with previous approaches, we use \\(L1\\) loss for depth estimation and surface normal estimation tasks and the cross-entropy loss for other tasks, therefore, the optimization objective can be expressed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.817, + 0.905, + 0.849 + ], + "angle": 0, + "content": "\\[\nL = \\sum_ {t \\in \\mathcal {T}} \\lambda_ {t} \\mathcal {L} _ {t} \\left(\\hat {\\mathbf {Y}} _ {t}, \\mathbf {Y} _ {t}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{T}\\) is the set of all tasks, \\(\\lambda_t\\), \\(\\mathcal{L}_t\\) and \\(\\mathbf{Y}_t\\) are the loss weight, loss function, and task label for image I in task \\(t\\), respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18819" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.216, + 0.108 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.281, + 0.133 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.139, + 0.483, + 0.32 + ], + "angle": 0, + "content": "Datasets. We performed experiments using the benchmark datasets NYUD-v2 [25] and PASCAL-Context [3]. NYUD-v2 primarily focuses on indoor scenes, with 795 and 654 RGB images for training and testing purposes. Tasks in NYUD-v2 include semantic segmentation, monocular depth estimation, surface normal estimation, and object boundary detection. PASCAL-Context encompasses indoor and outdoor scenes, offering pixel-level labels for tasks like semantic segmentation, human parsing, object boundary detection, surface normal estimation, and saliency detection tasks. The PASCAL-Context dataset contains 4,998 training images and 5,105 test images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.549 + ], + "angle": 0, + "content": "Implementation Details. We employ a pretrained SwinLarge Transformer [18] on ImageNet-22K [6] as our encoder. Our models are trained on the NYUD-v2 dataset for 50,000 iterations with a batch size of 4, and on the Pascal Context dataset for 75,000 iterations with a batch size of 6. Across all datasets, we use the Adam optimizer with a learning rate of \\(5 \\times 10^{-5}\\) and a weight decay rate of \\(1 \\times 10^{-5}\\), alongside a polynomial learning rate scheduler. The preliminary decoder has an output channel number of 768. We follow common practice [16, 35] in resizing the input images and applying data augmentation techniques. Specifically, we resize the input images of NYUD-v2 and PASCAL-Context to \\(448 \\times 576\\) and \\(512 \\times 512\\), respectively, and apply random color jittering, random cropping, random scaling, and random horizontal flipping." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.56, + 0.464, + 0.577 + ], + "angle": 0, + "content": "4.2. Comparisons with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.583, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Main Results. Tab. 1 and Tab. 2 report a comparison of the proposed BIM against previous state-of-the-art methods, including MTmamba [16], MQTransformer [32], TaskPrompter [36], \\(\\mathrm{InvPT}++\\) [37], \\(\\mathrm{InvPT}\\) [35], ATRC [2], MTI-Net [29], PAD-Net [31], PSD [40], PAP [42], Cross-Stitch [22] and ASTMT [20] on NYUD-V2 and PASCAL-Context dataset respectively. Notably, the previous best method, i.e., MTmamba, and our BIM are built upon the Transformer-encoder Mamba-decoder architecture with the same backbone. On NYUD-v2, the performance of Semseg is clearly boosted from the previous best, i.e., 55.82 to 57.40 (+1.52). On Pascal-Context, we achieved superior performance on all tasks compared to MTMamba." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Comparisons on Model Size and Computational Cost. In Tab. 3, we compare the FLOPs and parameter counts with other State-of-the-art methods. With the comparable computational cost, our BIM shows significant outperforms. our light-vision BIM (DBIM), which implements dilated sampling in MS-Scan, shows outperformance with reduced computational cost and parameter numbers, showcasing the advantage of these designs. Specifically, in MS" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with other state-of-the-art methods on NYUD-v2 dataset." + }, + { + "type": "table", + "bbox": [ + 0.53, + 0.129, + 0.892, + 0.364 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
CNN based
Cross-Stitch[22]36.340.629020.8876.38
PAP[42]36.720.617820.8276.42
PSD[40]36.690.624620.8776.42
PAD-Net[31]36.610.627020.8576.38
MTI-Net[29]45.940.536520.2777.86
ATRC[2]46.330.536320.1877.94
Transformer based
InvPT[35]53.660.518319.0478.10
TaskPrompter[36]55.300.515218.4778.20
InvPT++[37]53.850.509618.6778.10
MQTransformer[32]54.840.532519.6778.20
Mamba based
MTMamba[16]55.820.506618.6378.70
BIM(Ours)57.400.473318.5578.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.387, + 0.905, + 0.415 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with other state-of-the-art methods on Pascal-Context dataset." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.426, + 0.905, + 0.664 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF↑Normal mErr ↓Boundary odsF ↑
CNN based
PAD-Net[31]53.6059.6065.8015.3072.50
ASTMT[20]68.0061.1065.7014.7072.40
MTI-Net[29]61.7060.1884.7814.2370.80
ATRC[2]62.6959.4284.7014.2070.96
ATRC-ASPP[2]63.6060.2383.9114.3070.86
ATRC-BMTAS[2]67.6762.9382.2914.2472.42
Transformer based
InvPT[35]79.0367.6184.8114.1573.00
TaskPrompter[36]80.8968.8984.8313.7273.50
InvPT++[37]80.2269.1284.7413.7374.20
MQTransformer[32]78.9367.4183.5814.2173.90
Mamba based
MTmamba[16]81.1172.6284.1414.1478.80
BIM(Ours)81.2573.1584.1414.1379.05
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.699, + 0.907, + 0.837 + ], + "angle": 0, + "content": "Scan, we use dilated sampling in each scan branch \\(\\mathcal{B}\\) to lighten the operation of MS-Scan. Specifically: (1) We only sample one token in a window when generating multi-scale sequences from image features, instead of all tokens. (2) When restoring the sequence to image features, we perform linear interpolation between these selected tokens. These operations do not introduce any parameters, and the computational burden is reduced because only a subset of tokens are used for modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparisons. We qualitatively compared our proposed BIM with the previous best-performing method, as shown in Fig. 5. Our method shows clear improvements in detail, as highlighted in the circled regions." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "18820" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.134, + 0.328, + 0.861, + 0.343 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison with the best-performing method on Pascal-Context. Our method generates better details." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.367, + 0.483, + 0.395 + ], + "angle": 0, + "content": "Table 3. Complexity comparisons on NYUD-v2. †denotes the result is from [19]." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.406, + 0.48, + 0.51 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑FLOPs (G)↓# Params (M)↓
InvPT[35]53.660.518319.0478.10598402
TaskPrompter[36]55.300.515218.4778.20470392
InvPT++†[37]53.850.509618.6778.10-~402
MTMamba[16]55.820.506618.6378.70541308
DBIM (Ours)56.410.478918.6678.80495290
BIM (Ours)57.400.473318.5578.72547388
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.529, + 0.48, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.686 + ], + "angle": 0, + "content": "Figure 6. Ablation study on Bidirectional Scan (left) and two scanning modes (right) in BI-Scan." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.245, + 0.729 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Different Components. We performed ablation experiments on BI-Scan and MS-Scan using the NYUD-V2 dataset. These experiments all used Swin-Large Transformer as encoder. We use the same STL and MTL base model as in MTMamba. (1) The term \"STL\" indicates that each task is equipped with a separate network, utilizing two Swin Transformer blocks in each decoder stage, and (2) the \"MTL\" indicates that all tasks share a common encoder. (3) \"Baseline\" refers to MTMamba. (4) \"BI-Scan\" indicates that the MS-Scan is replaced with SS2D throughout the entire model. (5) \"MS-Scan\" refers to the implementation of" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.367, + 0.892, + 0.381 + ], + "angle": 0, + "content": "Table 4. Effectiveness of different components on NYUD-v2." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.503 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑FLOPs # (G) ↓Params (M) ↓
STL54.320.516619.2177.30+0.001075889
MTL53.720.523919.9776.50-1.87466303
Baseline55.820.506618.6378.70+2.38541308
BI-Scan57.110.485618.6678.90+4.02547290
MS-Scan57.400.482918.7978.60+4.00541397
BIM57.400.473318.5578.72+4.83547388
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.533, + 0.906, + 0.639 + ], + "angle": 0, + "content": "the same cross-task interaction strategy as that used in the baseline. Lastly, (6) \"BIM\" represents the default method. As depicted in Tab. 4, The BI-Scan improved model performance with fewer parameters, emphasizing the necessity of retaining task-specific details in cross-task interactions. Meanwhile, the MS-Scan enabled multi-granularity scene modeling, further enhancing performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.907, + 0.748 + ], + "angle": 0, + "content": "Effectiveness of Bidirectional Scan in BI-Scan. As shown in Fig. 6 (left), we compared the performance of unidirectional (Uni) and bidirectional (Bi) cross-task scanning. The results show that the introduction of bidirectional scanning significantly enhances the cross-task modeling capability, thereby improving the overall performance of the model with only a negligible increase in computational cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Two Scan Modes in BI-Scan. As shown in Fig. 6 (right), we conducted ablation experiments on the Task-First Scan mode and Position-First Scan mode within BI-Scan. The results indicate that both Task-First (Task) and Position-First (Position) interaction modes achieve superior performance compared to the baseline (Base) at a lower computational cost. Furthermore, employing both modes together (All) results in further performance improvements. Ultimately, we incorporate both Task-First mode and Position-First mode as the final configuration of" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18821" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.094, + 0.092, + 0.212, + 0.105 + ], + "angle": 0, + "content": "BI-Scan for BIM." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.108, + 0.482, + 0.318 + ], + "angle": 0, + "content": "Ablation on Task Order in BI-Scan. We experiment on NYUD-v2 to study the effect of the task order during scan interaction. In addition to fixed task orders, we also implemented a random task order modeling approach (Random), in which the order of tasks in BI-Scan is randomly shuffled during both training and testing. For clarity, we refer to the tasks Semseg, Depth, Normal, and Boundary as S, D, N, and B, respectively. As shown in Tab. 5, the overall performance of various fixed order models is relatively similar, while the Random setting exhibits a slight decrease in overall performance, which remains within an acceptable range. This indicates that our proposed BI-Scan demonstrates robustness to the task order. Ultimately, we adopt the first task order setting as the final configuration of BI-Scan for BIM." + }, + { + "type": "table_caption", + "bbox": [ + 0.152, + 0.333, + 0.424, + 0.346 + ], + "angle": 0, + "content": "Table 5. Ablation on task order on NYUD-v2." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.358, + 0.481, + 0.464 + ], + "angle": 0, + "content": "
OrderSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
S-D-N-B57.110.485618.6678.90+4.02
N-S-D-B56.150.483218.6679.00+3.72
B-S-D-N56.970.483618.7378.90+3.96
S-N-D-B57.520.484318.7178.90+4.20
Random55.410.486118.6978.90+3.17
" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.481, + 0.482, + 0.722 + ], + "angle": 0, + "content": "Ablation on Scan Scale and Number of MS-Scan. We experimented with the impact of varying scanning scales and numbers on model performance, as shown in Tab. 6 and Tab. 7. Experiments were conducted on the NYUD-v2 dataset with Swin-Large Transformer as encoder. We experimented with three different settings for scale ablation: all MFRs use \\(\\{1,2\\}\\), \\(\\{1,3\\}\\), and \\(\\{1,4\\}\\) two scanning scales, respectively. For number ablation, We compared three different settings, Type 1: all three MFRs use \\(\\{1\\}\\) scale; Type 2: all three MFRs use \\(\\{1,4\\}\\) two scales; Type 3: three MFRs use \\(\\{1,2\\}\\), \\(\\{1,2,4\\}\\) and \\(\\{1,2,4,6\\}\\), respectively. The results showed that in multi-scale scanning, employing appropriate scanning scale partitions can effectively enhance overall performance. Significantly, all settings yield notable performance enhancements. Ultimately, we adopt \\(\\{1,4\\}\\) scale setting for all MFRs as the final configuration." + }, + { + "type": "table_caption", + "bbox": [ + 0.104, + 0.736, + 0.47, + 0.75 + ], + "angle": 0, + "content": "Table 6. Different scanning scales of MS-Scan on NYUD-v2." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.762, + 0.48, + 0.838 + ], + "angle": 0, + "content": "
Scan ScaleSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
{1,2}57.390.475618.6078.80+4.68
{1,3}57.530.473918.6878.50+4.62
{1,4}57.400.473318.5578.72+4.83
" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.856, + 0.482, + 0.9 + ], + "angle": 0, + "content": "Performance on Different Encoders. We evaluate the effect of model size on experimental performance, as presented in Tab. 8. All experiments were conducted on the" + }, + { + "type": "table_caption", + "bbox": [ + 0.523, + 0.09, + 0.898, + 0.104 + ], + "angle": 0, + "content": "Table 7. Different scanning number of MS-Scan on NYUD-v2." + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.115, + 0.902, + 0.187 + ], + "angle": 0, + "content": "
Scan NumberSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
Type 157.110.485618.6678.90+4.02
Type 257.400.473318.5578.72+4.83
Type 356.760.475618.5978.70+4.35
" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.323 + ], + "angle": 0, + "content": "NYUD-v2 dataset. We compare our method with the previous best-performing model, MTMamba, using two different encoders: Swin-Base Transformer (denoted as '-Base') and Swin-Large Transformer (denoted as '-Large'). The results suggest that models with greater capacities typically exhibit superior performance. Furthermore, our approach has demonstrated superior performance across all encoder." + }, + { + "type": "table_caption", + "bbox": [ + 0.58, + 0.341, + 0.84, + 0.354 + ], + "angle": 0, + "content": "Table 8. Performance on different encoders." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.366, + 0.895, + 0.453 + ], + "angle": 0, + "content": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
MTMamba-Base53.620.512619.2877.70
BIM-Base54.320.491518.8977.74
MTMamba-Large55.820.506618.6378.70
BIM-Large57.400.473318.5578.72
" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.489, + 0.632, + 0.505 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.516, + 0.905, + 0.758 + ], + "angle": 0, + "content": "We introduce the Bidirectional Interaction Mamba (BIM) framework, a simple yet effective solution designed to address the inherent conflict between adequacy and complexity in cross-task interactions for multi-task dense prediction. We first propose a Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task information as bidirectional sequences to model cross-task interactions efficiently. By integrating task-first and position-first modes within linear complexity framework, BI-Scan preserves the integrity of task-specific information during interaction while maintaining computational efficiency. We further develop a Multi-Scale Scan (MS-Scan) mechanism, which captures scene structure information at multiple scales, alleviating the difficulty of feature learning and enhancing nuanced cross-task feature interactions. Extensive experiments show our method significantly enhances model performance." + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.777, + 0.692, + 0.794 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.887 + ], + "angle": 0, + "content": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80, and the Public Welfare Research Program of Ningbo City under Grant 2024S063." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "18822" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Deblina Bhattacharjee, Tong Zhang, Sabine Susstrunk, and Mathieu Salzmann. Mult: An end-to-end multitask learning transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12031-12041, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15869-15878, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[3] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.343, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[4] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.414, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[5] Michael Crawshaw. Multi-task learning with deep neural networks: A survey. CoRR, abs/2009.09796, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.57 + ], + "angle": 0, + "content": "[7] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.572, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[8] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3205-3214, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.672 + ], + "angle": 0, + "content": "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. CoRR, abs/2111.00396, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[12] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Woeseong Jeong and Kuk-Jin Yoon. Quantifying task priority for multi-task optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 363-372, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[14] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.906, + 0.22 + ], + "angle": 0, + "content": "[15] Wei-Hong Li, Steven McDonagh, Ales Leonardis, and Hakan Bilen. Multi-task learning with 3d-aware regularization. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.223, + 0.906, + 0.291 + ], + "angle": 0, + "content": "[16] Baijiong Lin, Weisen Jiang, Pengguang Chen, Yu Zhang, Shu Liu, and Ying-Cong Chen. MTMamba: Enhancing multi-task dense scene understanding by mamba-based decoders. In European Conference on Computer Vision, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.295, + 0.906, + 0.349 + ], + "angle": 0, + "content": "[17] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. CoRR, abs/2401.10166, 2024. 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.353, + 0.906, + 0.422 + ], + "angle": 0, + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.424, + 0.906, + 0.48 + ], + "angle": 0, + "content": "[19] Yuxiang Lu, Shengcao Cao, and Yu-Xiong Wang. Swiss army knife: Synergizing biases in knowledge from vision foundation models for multi-task learning. arXiv preprint arXiv:2410.14633, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.483, + 0.906, + 0.538 + ], + "angle": 0, + "content": "[20] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1851-1860, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.541, + 0.906, + 0.583 + ], + "angle": 0, + "content": "[21] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.585, + 0.906, + 0.641 + ], + "angle": 0, + "content": "[22] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994-4003, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.643, + 0.906, + 0.698 + ], + "angle": 0, + "content": "[23] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.701, + 0.906, + 0.77 + ], + "angle": 0, + "content": "[24] Sara Shoouri, Mingyu Yang, Zichen Fan, and Hun-Seok Kim. Efficient computation sharing for multi-task visual scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17130-17141, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.773, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[25] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[26] Dimitrios Sinodinos and Narges Armanfard. Ema-net: Efficient multitask affinity learning for dense scene predictions. arXiv preprint arXiv:2401.11124, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Jimmy T. H. Smith, Andrew Warrington, and Scott W. Linderman. Simplified state space layers for sequence modeling. CoRR, abs/2208.04933, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[28] Guolei Sun, Thomas Probst, Danda Pani Paudel, Nikola Popovic, Menelaos Kanakis, Jagruti Patel, Dengxin Dai, and Luc Van Gool. Task switching network for multi-task learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8291-8300, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[29] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 527-543. Springer, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[30] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[31] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 675-684, 2018. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[32] Yangyang Xu, Xiangtai Li, Haobo Yuan, Yibo Yang, and Lefei Zhang. Multi-task learning with multi-query transformer for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[33] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. CoRR, abs/2403.17695, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[34] Feiyang Ye, Yueming Lyu, Xuehao Wang, Yu Zhang, and Ivor Tsang. Adaptive stochastic gradient algorithm for black-box multi-objective learning. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[35] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In European Conference on Computer Vision, pages 514-530. Springer, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[36] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[37] Hanrong Ye and Dan Xu. Invpt++: Inverted pyramid multitask transformer for visual scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[38] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[39] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Rethinking of feature interaction for multi-task learning on dense prediction. arXiv preprint arXiv:2312.13514, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[40] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.164, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[41] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[42] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4514–4523, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.291, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.346 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "18824" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_origin.pdf b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..97ff2fb563272b4cf2e5b6fbbc1608c5b18795a2 --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/714687cb-b372-4c5e-8dc1-605cf43cbec1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4713b764748efc9b445e987129027aa57f415136362cbe4a62d8487b97691ea9 +size 989732 diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/full.md b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..a3c6b9223cd1d37a962e4a67de1718c295aaaecb --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/full.md @@ -0,0 +1,296 @@ +# Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction + +Mang Cao $^{1}$ Sanping Zhou $^{1*}$ Yizhe Li $^{1}$ Ye Deng $^{2}$ Wenli Huang $^{3}$ Le Wang $^{1}$ + +$^{1}$ National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University + +$^{2}$ School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics $^{3}$ School of Electronic and Information Engineering, Ningbo University of Technology + +3298592685@stu.xjtu.edu.cn, spzhou@xjtu.edu.cn + +# Abstract + +Sufficient cross-task interaction is crucial for success in multi-task dense prediction. However, sufficient interaction often results in high computational complexity, forcing existing methods to face the trade-off between interaction completeness and computational efficiency. To address this limitation, this work proposes a Bidirectional Interaction Mamba (BIM), which incorporates novel scanning mechanisms to adapt the Mamba modeling approach for multi-task dense prediction. On the one hand, we introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task-specific representations as bidirectional sequences during interaction. By integrating task-first and position-first scanning modes within a unified linear complexity architecture, BI-Scan efficiently preserves critical cross-task information. On the other hand, we employ a Multi-Scale Scan (MS-Scan) mechanism to achieve multi-granularity scene modeling. This design not only meets the diverse granularity requirements of various tasks but also enhances nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our BIM vs its state-of-the-art competitors. Codes are available online: https://github.com/mmm-cc/BIM_for_MTL. + +# 1. Introduction + +Multi-task dense prediction is a critical visual task designed to simultaneously predict outputs for various pixel-level tasks, such as semantic segmentation, depth predic + +![](images/1b3bbfd3c79dbe5276a4fcfd18ed2c587642c24916d678b8a71e8a6003a3a93e.jpg) +Figure 1. Illustration of different cross-task interaction strategies. (a) Pair-Wise [8, 28, 31]: each task can access the complete information of other tasks, while the interaction complexity is $O(n^{2})$ . (b) Bridge Representation [1, 16, 24, 39]: the interaction complexity is $O(n)$ , yet, each task is limited to accessing only a partial subset of information from other tasks during interaction, which inevitably results in the loss of valuable details. (c) Ours Cross-Task Scan: It preserves the integrity of task information during interaction, with an interaction complexity of $O(n)$ . + +tion, surface normal estimation, and saliency detection. Existing approaches made mainly efforts [5, 30] to refine task-general representations, obtained from the encoder, into task-specific representations. In these approaches, cross-task interaction has been proven to be significant in multi-task dense prediction [26, 31, 35, 39], as it can enhance the robustness and effectiveness of task representations. + +Initially, multi-task dense prediction methods primarily utilized Convolutional Neural Networks (CNNs) with pairwise interactions, as illustrated in Fig. 1 (a). These networks [8, 28, 31] were carefully designed with specialized + +branches for each task, augmented by interaction modules that facilitated information exchanges between each pair of tasks. These pair-wise interaction approaches maintain the integrity of task information during interaction, promoting fine-grained cross-task interaction. While, the inherently limited receptive fields of CNN architectures often led to insufficient spatial interactions. In response to these challenges, transformer-style networks [35] demonstrate remarkable capability in modeling long-range dependencies. This capability substantially improves the interaction effectiveness of models in handling multi-task scenarios, leading to enhanced performance. However, these sufficient cross-task interactions usually introduce high interaction complexity, which typically increases quadratically with the number of tasks. The high interaction complexity limits their applicability in certain powerful models, particularly transformers. + +To enhance the efficiency of cross-task interactions, a crowd of methods [1, 16, 24, 39] with linear interaction complexity has been proposed, as depicted in Fig. 1 (b). Typically, these methods strategically select a representative task [1, 24] or extract representative information from all tasks [16, 39] to construct a compact bridge representation that aims to preserve essential interaction information. However, in these methods, the necessity of maintaining linear interaction complexity inherently limits the representational capacity of these bridge representations. This information compression inevitably leads to the degradation of task features, potentially discarding critical interaction details and ultimately resulting in suboptimal performance. Therefore, how to enhance sufficient interaction capability while maintaining linear interaction complexity remains an unresolved issue. + +In response to these challenges, based on Mamba architecture, we propose a simple yet effective Bidirectional Interaction Mamba (BIM) method that achieves comprehensive preservation of task information during interaction while maintaining linear interaction complexity. Specifically, we introduce a Bidirectional Interaction Scan (BI-Scan) mechanism to achieve sufficient cross-task interaction and a Multi-Scale Scan (MS-Scan) mechanism to deliver a comprehensive visual representation. The BI-Scan mechanism constructs each task representation into sequences and deploys a bidirectional scanning strategy for modeling cross-task interactions, efficiently mitigating information loss while preserving linear interaction complexity. Furthermore, BI-Scan enables multi-level cross-task interactions by integrating task-first and position-first scanning modes, facilitating more nuanced modeling of task relationships. In the MS-Scan, features are initially partitioned into multiple spaces, where the scene structure of images is modeled at various scales. Specific tasks subsequently integrate multi-scale scene structural information to + +enhance task-specific representations. During cross-task interaction, different task branches adaptively integrate interaction representations from multiple scales, addressing the varying granularity demands of each task for scene structural information. + +The main contributions are summarized as follows: + +- We propose a simple yet effective Bidricational Interaction Mamba (BIM) for multi-task dense prediction, enabling comprehensive cross-task interaction while preserving linear interaction complexity. +- We introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism to maintain the integrity of task information during interactions. The BI-Scan incorporates two scanning modes: task-first mode and position-first mode, facilitating multi-level cross-task interactions. +- We design a Multi-Scale Scan (MS-Scan) mechanism to capture scene structural information at multiple scales, thereby addressing the diverse granularity requirements of various tasks and enhancing more detailed cross-task interactions in multi-task dense prediction. + +# 2. Related Work + +Multi-Task Learning. Most existing multi-task learning works focused on training optimization and network design. Optimization approaches balance tasks via gradient manipulation [13, 23, 34, 38] or loss-balancing [4, 14] to coordinate resource allocation for various tasks during training. The approaches of network design seek to enhance task representation learning by developing diverse mechanisms. Some CNN-based methods [29, 31] manually design interaction mechanisms to extract useful information across tasks. With the advancement of Transformer, current methods have gained improved global task interaction capabilities to enhance task representation efficiency. They enable superior pairwise [1, 24] or global [15, 32, 35, 36] task interactions. To mitigate computational complexity, MT-Mamba [16] introduce Mamba [9] to multi-task learning, which showcases effective long sequence modeling capabilities and achieving satisfactory performance. However, achieving both efficient and sufficient cross-task interaction remains a challenge. + +State Space Models. In efficient long-range dependency modeling methods, state space models [11, 27] has become a striking alternative to Transformers. [10] proposed a Structured State Space Sequence (S4) model based on a new parameterization, which alleviates the computational and memory efficiency issues faced by SSM. Subsequently, numerous efforts [7, 21] are dedicated to bridging the performance disparity between SSMs and Transformers. Mamba [9] introduced an input-based parameterization method and hardware-aware algorithm, achieving performance on par with Transformers in natural language processing. This success inspired various efforts [43] towards + +![](images/8992b268b8f720d5883f299c660f667fa679b883a3df0ee9504c680d5b35e415.jpg) +(a) BIM for MTL + +![](images/9a225d94f9ddce0b77bcf047531970919941810f93982158b2cd6cab243c28e6.jpg) +(b) MFR Block +Figure 2. Framework of the proposed BIM for multi-task dense prediction. (a) Overview of BIM, illustrating with depth estimation and surface normal estimation tasks. (b) Details of the MFR block, including T task-specific MSST blocks and a task-shared BCFR block. + +Mamba's adaptation for visual tasks. + +The preservation of comprehensive image structural information poses a critical challenge in Mamba's sequential processing model, which has attracted considerable attention and effort [12, 17, 33, 41, 43]. VMamba [17] proposes the 2D Selective Scan (SS2D), a four-way scanning mechanism tailored for spatial domain traversal, aimed at enhancing Mamba's image modeling capabilities. Subsequent research studies have explored various scan patterns and combinations tailored to different tasks or scenarios [12, 33, 41]. However, by utilizing fixed token sizes, these methods overlook the importance of hierarchical spatial structural information in visual tasks. + +# 3. Method + +We first outline the overall architecture of BIM for multitask dense prediction in Section 3.1, Subsequently, Section 3.2 focuses on the BIM Encoder, while Section 3.3 examines the BIM Decoder and its two scanning mechanisms: BI-Scan and MS-Scan. Finally, Section 3.4 discusses the optimization objectives. + +# 3.1. Framework Overview + +As shown Fig. 2 (a), the proposed BIM framework consists of three main components: a task-shared encoder $\Phi$ for extracting task-generic representations, a Mamba decoder $\Theta$ for refining task features and T task-specific heads $\{H_t\}_{t=1}^T$ generating predictions $\{\hat{\mathbf{Y}}_t\}_{t=1}^T$ for individual tasks. This can be formulated as: + +$$ +\begin{array}{l} \left\{F _ {1}, F _ {2}, \dots , F _ {T} \right\} = \Theta (\Phi (\mathbf {I})), \\ \hat {\mathbf {Y}} _ {t} = H _ {t} \left(F _ {t}\right) = \mathcal {P} _ {t} \left(\Psi \left(F _ {t}\right)\right), \tag {1} \\ \end{array} +$$ + +where $\mathbf{I} \in \mathbb{R}^{3 \times H \times W}$ denotes the RGB input image, $F_{t}$ denotes the features generated by the decoder, corresponding + +to the distinct task $t$ . $\hat{\mathbf{Y}}_t$ represents the prediction for task $t$ , which has the same height $H$ and width $W$ as I. $\Psi$ denotes a module designed to expand the task feature resolution to $H \times W$ , which consists of a linear projection and a reshape operation. $\mathcal{P}_t$ represents a linear layer that projects the feature channels to a specific number required by the corresponding task. + +# 3.2. BIM Encoder + +The encoder shares similarities with other methods [16]. We utilized a pretrained Swin Transformer [18] to extract task-generic features, which begins by dividing the input image $\mathbf{I} \in \mathbb{R}^{H \times W \times 3}$ into $\frac{H}{w} \times \frac{W}{w}$ tokens of dimension $C$ through patch partitioning and linear layers, where $w$ denotes the partition size. These tokens are then processed through multiple stages involving alternating patch merging and Swin Transformer block processing, ultimately yielding hierarchical image representations: + +$$ +\mathbf {G} = \left\{\mathbf {G} ^ {1}, \mathbf {G} ^ {2}, \mathbf {G} ^ {3}, \mathbf {G} ^ {4} \right\} = \Phi (\mathrm {I}), \tag {2} +$$ + +where $\mathbf{G}$ represents the task-generic features extracted from the encoder $\Phi$ . In our practical implementation, we utilize a partition size of 4, resulting in the following shapes for $\mathbf{G}$ : $C \times \frac{H}{4} \times \frac{W}{4}$ , $2C \times \frac{H}{8} \times \frac{W}{8}$ , $4C \times \frac{H}{16} \times \frac{W}{16}$ and $8C \times \frac{H}{32} \times \frac{W}{32}$ , respectively. + +# 3.3. BIM Decoder + +The BIM Decoder consists of three Mamba Feature Refinement (MFR) blocks, as depicted in Fig. 2 (a). These MFRs are designed to bridge the gap between task-generic and task-specific representations. They generate task representations $\{\mathbf{F}_1^i,\mathbf{F}_2^i,\ldots ,\mathbf{F}_T^i\}_{i = 1}^3$ corresponding to the first three encoder stages, with dimensions of $4C\times \frac{H}{16}\times \frac{W}{16},2C\times \frac{H}{8}\times \frac{W}{8}$ , and $C\times \frac{H}{4}\times \frac{W}{4}$ , respectively. Finally, the last + +refined features $\{\mathbf{F}_t^3\}_{t = 1}^T$ are input into the task heads as $\{F_{t}\}_{t = 1}^{T}$ to generate the final predictions $\{\hat{\mathbf{Y}}_t\}_{t = 1}^T$ + +We introduce two specialized Mamba blocks, Bidirectional Cross-Task Feature Refine (BCFR) and Multi-Scale Single-task (MSST) in each MFR block, as illustrated in Fig. 2 (b). The BCFR Block is designed to facilitate effective cross-task interaction while maintaining the integrity of task information. It begins by concatenating features $\{\mathcal{F}_t^s\}_{t=1}^T$ from different task branches to construct task-shared features $\mathcal{F}_{sh}^s$ using the BI-Scan mechanism. Subsequently, each task branch adaptively merges $\mathcal{F}_{sh}^s$ by a selection value $\mathcal{G}^s$ to obtain fine-tuned task representations $\mathbf{F}^s = \{\mathbf{F}_t^s\}_{t=1}^T$ : + +$$ +\mathbf {F} _ {t} ^ {s} = \mathcal {F} _ {t} ^ {s} + \mathcal {G} _ {t} ^ {s} \times \mathcal {F} _ {s h} ^ {s} [ t ] + (1 - \mathcal {G} _ {t} ^ {s}) \times \mathcal {F} _ {t, N} ^ {s}, \tag {3} +$$ + +where $\mathcal{F}_{t,N}^{s}$ is derived from $\mathcal{F}_t^s$ through layer normalization, and $\mathcal{G}_t^s$ is obtained from $\mathcal{F}_{t,N}^{s}$ via a linear layer followed by a Sigmoid function. $\mathcal{F}_{sh}^{s}[t]$ is the $t$ -th component of $\mathcal{F}_{sh}^{s}$ corresponding to the $t$ -th task. The MSST primarily aims to construct comprehensive representations through task-internal interactions. Its architecture is illustrated in Fig. 2 (b) and comprises two main branches: the scan branch and the gating branch. In the scan branch, we integrate a hierarchical global scene structure representation $\mathcal{F}_h^s$ using a novel MS-Scan mechanism. Simultaneously, a gating signal $\mathcal{G}^s$ from the gate branch regulates the flow of information. Subsequently, we adjust the channel dimensions of the multi-scale scene representation by applying a linear projection $\mathcal{P}$ and establish a residual connection between $\mathcal{F}_h^s$ and the input $\mathcal{F}_{in}^{s}$ . This process is repeated twice to produce the final output $\mathcal{F}^s$ : + +$$ +\mathcal {F} ^ {s} = \mathcal {F} _ {i n} ^ {s} + \mathcal {P} \left(\mathcal {F} _ {h} ^ {s} \times \mathcal {G} ^ {s}\right). \tag {4} +$$ + +Bidirectional Interaction Scan. To maintain task information integrity during interaction with linear complexity, we propose a novel BI-Scan that serves as a cornerstone of BIM, as shown in Fig. 3. This module takes the representations of different tasks $\{\mathcal{F}_{t,N}^s\}_{t = 1}^T$ as input, which is initially partitioned along the channel dimension into dual-branch representations, each with the dimensions $H\times W\times \frac{C}{2}$ . The former undergoes processing through our Forward Scan module, while the latter is reversed along the task dimension before the Backward Scan processing. Specifically, the Forward Scan architecture implements hierarchical cross-task interaction through two scanning modes: Task-First mode and Position-First mode. + +Task-First Mode: Following a similar configuration to SS2D in Vmamba [17], we implement multiple distinct Scan Patterns for each scanning mode to facilitate enhanced task interaction. Without loss of generality, we use the first scan pattern as an example. As shown in Fig. 3, in the task-first mode, the construction of global task information contains two processes: (1) Pattern-guided feature serialization + +![](images/9df8236e2bae9e178842067166da01f9f7fdc93fbf8de7b22d8f3e607bd8d9e0.jpg) +Figure 3. Framework of BI-Scan. It consists of two modules, Forward Scan and Backward Scan, each including Task-First and Position-First cross-task scanning modes. + +$(\mathcal{U}_1)$ : each task feature is serialized according to the scan pattern, resulting in $T$ subsequences, each of length $H \times W$ . (2) Task order-guided subsequence aggregation $(\mathcal{O}_1)$ : these subsequences are concatenated along the length dimension according to the specified task order, producing a global sequence $S$ of length $T \times (H \times W)$ . These processes can be expressed as: + +$$ +S _ {1} = \mathcal {O} _ {1} \left(\mathcal {U} _ {1} \left(\mathcal {F} _ {t} ^ {s}, \mathcal {F} _ {2} ^ {s}, \dots , \mathcal {F} _ {T} ^ {s}\right)\right) \in \mathbb {R} ^ {\left(T \times (H \times W)\right) \times \frac {C}{2}}. \tag {5} +$$ + +This sequence undergoes global modeling through SSM [9], yielding task-level interaction information. + +Position-First Mode: In the position-first mode, the output of Task-First mode is first restored from the sequence to the input task feature size, recorded as $\{\bar{F}_t^s\}_{t=1}^T \in \mathbb{R}^{H \times W \times \frac{C}{2}}$ . Subsequently, the construction of interaction representations also involves two steps: (1) Task order-guided feature serialization $(\mathcal{O}_2)$ : Spatially aligned task representation tokens are constructed into $H \times W$ subsequences, each of length $T$ , according to the specified task order. (2) Pattern-guided subsequence aggregation $(\mathcal{U}_2)$ : These subsequences are concatenated along the length dimension according to the scan pattern, resulting in a sequence of length $(H \times W) \times T$ . The above processes can be expressed as: + +$$ +S _ {2} = \mathcal {U} _ {2} \left(\mathcal {O} _ {2} \left(\bar {\mathcal {F}} _ {1} ^ {s}, \bar {\mathcal {F}} _ {2} ^ {s}, \dots , \bar {\mathcal {F}} _ {T} ^ {s}\right)\right) \in \mathbb {R} ^ {\left(\left(H \times W\right) \times T\right) \times \frac {C}{2}}. \tag {6} +$$ + +This sequence is also processed through SSM, allowing for the derivation of task-level interaction insights. + +Subsequently, the outputs corresponding to all patterns are fused via element-wise summation, yielding the output + +![](images/dca41e1f6e1d2fafed6c1200ece592ad6c004262a37659728b0f04e02b43d801.jpg) +Figure 4. Instructions for MS-Scan. It consists of three distinct operations, Input Handling, Multi-Scale Scanning, and Multi-Scale Fusion. In Multi-Scale Scanning, each branch $\mathcal{B}_i$ has a distinct scanning size $s_i$ . + +of Forward Scan, denoted as $\mathcal{F}_{for}^{s}$ . Finally, the output of the Backward Scan, $\mathcal{F}_{back}^{s}$ , is reversed along the task dimension and concatenated with $\mathcal{F}_{for}^{s}$ to yield the refined task representations $\mathcal{F}_{sh}^{s}$ . In practical application, when employing BI-Scan independently, the selection of Scan Patterns aligns with those used in SS2D [17], specifically, the four directions shown in Fig. 3. When MS-Scan is introduced within the BI-Scan module, we incorporate Scan Patterns of varying scales without increasing computational complexity. + +Multi-Scale Scan. Mamba processes features by converting them into sequences of tokens. While a variety of tokenization strategies exist for text, image tokenization requires careful consideration of multiple approaches. Previous research has demonstrated that multi-scale processing is particularly effective for image data. To leverage this advantage, We propose the MS-Scan mechanism that requires minimal additional computational cost compared to single-scale scanning, as illustrated in Fig. 4 (a). In this framework, we employ multiple scanning scales, denoted as $\{s_i\}_{i=1}^N$ , in multiple branches $\{\mathcal{B}_i\}_{i=1}^N$ , and transform the input image feature $x \in \mathbb{R}^{C \times H \times W}$ into token sequences of varying dimensions for Mamba modeling. For instance, the initial image feature can be tokenized into a sequence with a total length of $H \times W$ , where each token has a dimension of $C$ . When applying a scanning scale of $s_i = 2$ , the image is divided into non-overlapping patches, resulting in a tokenized feature sequence with a dimensionality of $4C \times \left(\frac{H}{2} \times \frac{W}{2}\right)$ . Specifically, MS-Scan comprises three key components: input handling, multi-scale scanning, and multi-scale fusion. + +Input Handling: To construct inputs for $N$ different scanning branches $\mathcal{B} = \{\mathcal{B}_i\}_{i=1}^N$ , we perform two main operations. (1) Channel Split ( $\mathcal{S}$ ): We begin by splitting the input representation $x$ into $N$ sub-features $\{x_i\}_{i=1}^N \in$ + +$\mathbb{R}^{m\times H\times W}$ along the channel dimension, where $m = \frac{C}{N}$ . (2) Window Tokenization $(\mathcal{W}_i)$ : For the $i$ -th branch $\mathcal{B}_i$ with scanning scale $s_i$ , we first divide $x_{i}$ into $\frac{H}{s_i}\times \frac{W}{s_i}$ nonoverlapping patches, each of size $m\times s_i\times s_i$ . Subsequently, we concatenate the pixel feature values in each patch along the channel dimension, resulting in the scan input $\bar{x}_i$ with a shape of $(m\times s_i\times s_i)\times \frac{H}{s_i}\times \frac{W}{s_i}$ . Ultimately, we obtain the input for all branches: + +$$ +\{\bar {x} _ {1}, \bar {x} _ {2}, \dots , \bar {x} _ {N} \} = \left\{\mathcal {W} _ {1}, \mathcal {W} _ {2}, \dots , \mathcal {W} _ {N} \right\} (\mathcal {S} (x)). \tag {7} +$$ + +Multi-Scale Scanning: Following the input handling process, we employ distinct scanning scales to construct a multi-scale scene representation in each branch. For all branches $\mathcal{B}$ , we utilize the four-way scanning method (SS2D) from VMamba [17] to generate scene features at the specified scale. This method creates four token sequences, each shaped as $C_i \times (H_i \times W_i)$ , by scanning the input features $\bar{x}_i \in \mathbb{R}^{C_i \times H_i \times W_i}$ in four directions. The resulting sequences are then processed by SSM [9] and combined to produce the output feature $\bar{y} = \{\bar{y}_i\}_{i=1}^N$ : + +$$ +\bar {y} _ {i} = \operatorname {S S 2 D} (\bar {x} _ {i}) \in \mathbb {R} ^ {C _ {i} \times H _ {i} \times W _ {i}}, i \in \{1, 2, \dots , N \}. \tag {8} +$$ + +Multi-Scale Fusion: In our approach to multi-scale feature fusion, we adopt a methodology that reverses the input handling process, consisting of two key steps. (1) Token Windowing $(\mathcal{W}_i^{-1})$ : For each branch $\mathcal{B}_i$ , we split each pixel feature into $s_i \times s_i$ segments along the channel dimension, resulting in a set of feature segments $\{\bar{y}_{i,j}\}_{j=1}^{s_i \times s_i} \in \mathbb{R}^{m \times 1 \times 1}$ . These segments are then concatenated along the spatial dimensions (height and width) to form patches, which are subsequently combined to produce the output for $\mathcal{B}_i$ . (2) Channel Concatenation $(S^{-1})$ : We concatenate the features from all branches along the channel dimension, yielding the final output feature $y \in \mathbb{R}^{C \times H \times W}$ : + +$$ +y = \mathcal {S} ^ {- 1} \left(\mathcal {W} _ {1} ^ {- 1} (\bar {y} _ {1}), \mathcal {W} _ {2} ^ {- 1} (\bar {y} _ {2}), \dots , \mathcal {W} _ {N} ^ {- 1} (\bar {y} _ {N})\right), \tag {9} +$$ + +where $S^{-1}$ and $\mathcal{W}_i^{-1}$ refer to the inverse operation of $S$ and $\mathcal{W}_i$ respectively. + +# 3.4. Optimization Objective + +We jointly train all tasks to optimize BIM encoder $\Phi$ , BIM decoder $\Theta$ , and task heads $\{H\}_{i=1}^{T}$ simultaneously. To maintain consistency with previous approaches, we use $L1$ loss for depth estimation and surface normal estimation tasks and the cross-entropy loss for other tasks, therefore, the optimization objective can be expressed as follows: + +$$ +L = \sum_ {t \in \mathcal {T}} \lambda_ {t} \mathcal {L} _ {t} \left(\hat {\mathbf {Y}} _ {t}, \mathbf {Y} _ {t}\right), \tag {10} +$$ + +where $\mathcal{T}$ is the set of all tasks, $\lambda_t$ , $\mathcal{L}_t$ and $\mathbf{Y}_t$ are the loss weight, loss function, and task label for image I in task $t$ , respectively. + +# 4. Experiment + +# 4.1. Experimental Setup + +Datasets. We performed experiments using the benchmark datasets NYUD-v2 [25] and PASCAL-Context [3]. NYUD-v2 primarily focuses on indoor scenes, with 795 and 654 RGB images for training and testing purposes. Tasks in NYUD-v2 include semantic segmentation, monocular depth estimation, surface normal estimation, and object boundary detection. PASCAL-Context encompasses indoor and outdoor scenes, offering pixel-level labels for tasks like semantic segmentation, human parsing, object boundary detection, surface normal estimation, and saliency detection tasks. The PASCAL-Context dataset contains 4,998 training images and 5,105 test images. + +Implementation Details. We employ a pretrained SwinLarge Transformer [18] on ImageNet-22K [6] as our encoder. Our models are trained on the NYUD-v2 dataset for 50,000 iterations with a batch size of 4, and on the Pascal Context dataset for 75,000 iterations with a batch size of 6. Across all datasets, we use the Adam optimizer with a learning rate of $5 \times 10^{-5}$ and a weight decay rate of $1 \times 10^{-5}$ , alongside a polynomial learning rate scheduler. The preliminary decoder has an output channel number of 768. We follow common practice [16, 35] in resizing the input images and applying data augmentation techniques. Specifically, we resize the input images of NYUD-v2 and PASCAL-Context to $448 \times 576$ and $512 \times 512$ , respectively, and apply random color jittering, random cropping, random scaling, and random horizontal flipping. + +# 4.2. Comparisons with State-of-the-art Methods + +Main Results. Tab. 1 and Tab. 2 report a comparison of the proposed BIM against previous state-of-the-art methods, including MTmamba [16], MQTransformer [32], TaskPrompter [36], $\mathrm{InvPT}++$ [37], $\mathrm{InvPT}$ [35], ATRC [2], MTI-Net [29], PAD-Net [31], PSD [40], PAP [42], Cross-Stitch [22] and ASTMT [20] on NYUD-V2 and PASCAL-Context dataset respectively. Notably, the previous best method, i.e., MTmamba, and our BIM are built upon the Transformer-encoder Mamba-decoder architecture with the same backbone. On NYUD-v2, the performance of Semseg is clearly boosted from the previous best, i.e., 55.82 to 57.40 (+1.52). On Pascal-Context, we achieved superior performance on all tasks compared to MTMamba. + +Comparisons on Model Size and Computational Cost. In Tab. 3, we compare the FLOPs and parameter counts with other State-of-the-art methods. With the comparable computational cost, our BIM shows significant outperforms. our light-vision BIM (DBIM), which implements dilated sampling in MS-Scan, shows outperformance with reduced computational cost and parameter numbers, showcasing the advantage of these designs. Specifically, in MS + +Table 1. Quantitative comparison with other state-of-the-art methods on NYUD-v2 dataset. + +
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
CNN based
Cross-Stitch[22]36.340.629020.8876.38
PAP[42]36.720.617820.8276.42
PSD[40]36.690.624620.8776.42
PAD-Net[31]36.610.627020.8576.38
MTI-Net[29]45.940.536520.2777.86
ATRC[2]46.330.536320.1877.94
Transformer based
InvPT[35]53.660.518319.0478.10
TaskPrompter[36]55.300.515218.4778.20
InvPT++[37]53.850.509618.6778.10
MQTransformer[32]54.840.532519.6778.20
Mamba based
MTMamba[16]55.820.506618.6378.70
BIM(Ours)57.400.473318.5578.72
+ +Table 2. Quantitative comparison with other state-of-the-art methods on Pascal-Context dataset. + +
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF↑Normal mErr ↓Boundary odsF ↑
CNN based
PAD-Net[31]53.6059.6065.8015.3072.50
ASTMT[20]68.0061.1065.7014.7072.40
MTI-Net[29]61.7060.1884.7814.2370.80
ATRC[2]62.6959.4284.7014.2070.96
ATRC-ASPP[2]63.6060.2383.9114.3070.86
ATRC-BMTAS[2]67.6762.9382.2914.2472.42
Transformer based
InvPT[35]79.0367.6184.8114.1573.00
TaskPrompter[36]80.8968.8984.8313.7273.50
InvPT++[37]80.2269.1284.7413.7374.20
MQTransformer[32]78.9367.4183.5814.2173.90
Mamba based
MTmamba[16]81.1172.6284.1414.1478.80
BIM(Ours)81.2573.1584.1414.1379.05
+ +Scan, we use dilated sampling in each scan branch $\mathcal{B}$ to lighten the operation of MS-Scan. Specifically: (1) We only sample one token in a window when generating multi-scale sequences from image features, instead of all tokens. (2) When restoring the sequence to image features, we perform linear interpolation between these selected tokens. These operations do not introduce any parameters, and the computational burden is reduced because only a subset of tokens are used for modeling. + +Qualitative Comparisons. We qualitatively compared our proposed BIM with the previous best-performing method, as shown in Fig. 5. Our method shows clear improvements in detail, as highlighted in the circled regions. + +![](images/e4bed02f21f0e9d7922c638ffa22122785dcd318e543231656ca2b51a64263f6.jpg) +Figure 5. Qualitative comparison with the best-performing method on Pascal-Context. Our method generates better details. + +Table 3. Complexity comparisons on NYUD-v2. †denotes the result is from [19]. + +
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑FLOPs (G)↓# Params (M)↓
InvPT[35]53.660.518319.0478.10598402
TaskPrompter[36]55.300.515218.4778.20470392
InvPT++†[37]53.850.509618.6778.10-~402
MTMamba[16]55.820.506618.6378.70541308
DBIM (Ours)56.410.478918.6678.80495290
BIM (Ours)57.400.473318.5578.72547388
+ +![](images/297f220a630441873597088e4ac056b5b32573222b5e1d801ea58f24ee81b565.jpg) +Figure 6. Ablation study on Bidirectional Scan (left) and two scanning modes (right) in BI-Scan. + +# 4.3. Ablation Study + +Effectiveness of Different Components. We performed ablation experiments on BI-Scan and MS-Scan using the NYUD-V2 dataset. These experiments all used Swin-Large Transformer as encoder. We use the same STL and MTL base model as in MTMamba. (1) The term "STL" indicates that each task is equipped with a separate network, utilizing two Swin Transformer blocks in each decoder stage, and (2) the "MTL" indicates that all tasks share a common encoder. (3) "Baseline" refers to MTMamba. (4) "BI-Scan" indicates that the MS-Scan is replaced with SS2D throughout the entire model. (5) "MS-Scan" refers to the implementation of + +Table 4. Effectiveness of different components on NYUD-v2. + +
ModelSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑FLOPs # (G) ↓Params (M) ↓
STL54.320.516619.2177.30+0.001075889
MTL53.720.523919.9776.50-1.87466303
Baseline55.820.506618.6378.70+2.38541308
BI-Scan57.110.485618.6678.90+4.02547290
MS-Scan57.400.482918.7978.60+4.00541397
BIM57.400.473318.5578.72+4.83547388
+ +the same cross-task interaction strategy as that used in the baseline. Lastly, (6) "BIM" represents the default method. As depicted in Tab. 4, The BI-Scan improved model performance with fewer parameters, emphasizing the necessity of retaining task-specific details in cross-task interactions. Meanwhile, the MS-Scan enabled multi-granularity scene modeling, further enhancing performance. + +Effectiveness of Bidirectional Scan in BI-Scan. As shown in Fig. 6 (left), we compared the performance of unidirectional (Uni) and bidirectional (Bi) cross-task scanning. The results show that the introduction of bidirectional scanning significantly enhances the cross-task modeling capability, thereby improving the overall performance of the model with only a negligible increase in computational cost. + +Effectiveness of Two Scan Modes in BI-Scan. As shown in Fig. 6 (right), we conducted ablation experiments on the Task-First Scan mode and Position-First Scan mode within BI-Scan. The results indicate that both Task-First (Task) and Position-First (Position) interaction modes achieve superior performance compared to the baseline (Base) at a lower computational cost. Furthermore, employing both modes together (All) results in further performance improvements. Ultimately, we incorporate both Task-First mode and Position-First mode as the final configuration of + +# BI-Scan for BIM. + +Ablation on Task Order in BI-Scan. We experiment on NYUD-v2 to study the effect of the task order during scan interaction. In addition to fixed task orders, we also implemented a random task order modeling approach (Random), in which the order of tasks in BI-Scan is randomly shuffled during both training and testing. For clarity, we refer to the tasks Semseg, Depth, Normal, and Boundary as S, D, N, and B, respectively. As shown in Tab. 5, the overall performance of various fixed order models is relatively similar, while the Random setting exhibits a slight decrease in overall performance, which remains within an acceptable range. This indicates that our proposed BI-Scan demonstrates robustness to the task order. Ultimately, we adopt the first task order setting as the final configuration of BI-Scan for BIM. + +Table 5. Ablation on task order on NYUD-v2. + +
OrderSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
S-D-N-B57.110.485618.6678.90+4.02
N-S-D-B56.150.483218.6679.00+3.72
B-S-D-N56.970.483618.7378.90+3.96
S-N-D-B57.520.484318.7178.90+4.20
Random55.410.486118.6978.90+3.17
+ +Ablation on Scan Scale and Number of MS-Scan. We experimented with the impact of varying scanning scales and numbers on model performance, as shown in Tab. 6 and Tab. 7. Experiments were conducted on the NYUD-v2 dataset with Swin-Large Transformer as encoder. We experimented with three different settings for scale ablation: all MFRs use $\{1,2\}$ , $\{1,3\}$ , and $\{1,4\}$ two scanning scales, respectively. For number ablation, We compared three different settings, Type 1: all three MFRs use $\{1\}$ scale; Type 2: all three MFRs use $\{1,4\}$ two scales; Type 3: three MFRs use $\{1,2\}$ , $\{1,2,4\}$ and $\{1,2,4,6\}$ , respectively. The results showed that in multi-scale scanning, employing appropriate scanning scale partitions can effectively enhance overall performance. Significantly, all settings yield notable performance enhancements. Ultimately, we adopt $\{1,4\}$ scale setting for all MFRs as the final configuration. + +Table 6. Different scanning scales of MS-Scan on NYUD-v2. + +
Scan ScaleSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
{1,2}57.390.475618.6078.80+4.68
{1,3}57.530.473918.6878.50+4.62
{1,4}57.400.473318.5578.72+4.83
+ +Performance on Different Encoders. We evaluate the effect of model size on experimental performance, as presented in Tab. 8. All experiments were conducted on the + +Table 7. Different scanning number of MS-Scan on NYUD-v2. + +
Scan NumberSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
Type 157.110.485618.6678.90+4.02
Type 257.400.473318.5578.72+4.83
Type 356.760.475618.5978.70+4.35
+ +NYUD-v2 dataset. We compare our method with the previous best-performing model, MTMamba, using two different encoders: Swin-Base Transformer (denoted as '-Base') and Swin-Large Transformer (denoted as '-Large'). The results suggest that models with greater capacities typically exhibit superior performance. Furthermore, our approach has demonstrated superior performance across all encoder. + +Table 8. Performance on different encoders. + +
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
MTMamba-Base53.620.512619.2877.70
BIM-Base54.320.491518.8977.74
MTMamba-Large55.820.506618.6378.70
BIM-Large57.400.473318.5578.72
+ +# 5. Conclusion + +We introduce the Bidirectional Interaction Mamba (BIM) framework, a simple yet effective solution designed to address the inherent conflict between adequacy and complexity in cross-task interactions for multi-task dense prediction. We first propose a Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task information as bidirectional sequences to model cross-task interactions efficiently. By integrating task-first and position-first modes within linear complexity framework, BI-Scan preserves the integrity of task-specific information during interaction while maintaining computational efficiency. We further develop a Multi-Scale Scan (MS-Scan) mechanism, which captures scene structure information at multiple scales, alleviating the difficulty of feature learning and enhancing nuanced cross-task feature interactions. Extensive experiments show our method significantly enhances model performance. + +# 6. Acknowledgement + +This work was supported in part by National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80, and the Public Welfare Research Program of Ningbo City under Grant 2024S063. + +# References + +[1] Deblina Bhattacharjee, Tong Zhang, Sabine Susstrunk, and Mathieu Salzmann. Mult: An end-to-end multitask learning transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12031-12041, 2022. 1, 2 +[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15869-15878, 2021. 6 +[3] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6 +[4] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 2 +[5] Michael Crawshaw. Multi-task learning with deep neural networks: A survey. CoRR, abs/2009.09796, 2020. 1 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[7] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2 +[8] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3205-3214, 2019. 1 +[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2024. 2, 4, 5 +[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. CoRR, abs/2111.00396, 2021. 2 +[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 2 +[12] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3 +[13] Woeseong Jeong and Kuk-Jin Yoon. Quantifying task priority for multi-task optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 363-372, 2024. 2 + +[14] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 2 +[15] Wei-Hong Li, Steven McDonagh, Ales Leonardis, and Hakan Bilen. Multi-task learning with 3d-aware regularization. In The Twelfth International Conference on Learning Representations, 2024. 2 +[16] Baijiong Lin, Weisen Jiang, Pengguang Chen, Yu Zhang, Shu Liu, and Ying-Cong Chen. MTMamba: Enhancing multi-task dense scene understanding by mamba-based decoders. In European Conference on Computer Vision, 2024. 1, 2, 3, 6, 7 +[17] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. CoRR, abs/2401.10166, 2024. 3, 4, 5 +[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6 +[19] Yuxiang Lu, Shengcao Cao, and Yu-Xiong Wang. Swiss army knife: Synergizing biases in knowledge from vision foundation models for multi-task learning. arXiv preprint arXiv:2410.14633, 2024. 7 +[20] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1851-1860, 2019. 6 +[21] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 2 +[22] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994-4003, 2016. 6 +[23] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 2 +[24] Sara Shoouri, Mingyu Yang, Zichen Fan, and Hun-Seok Kim. Efficient computation sharing for multi-task visual scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17130-17141, 2023. 1, 2 +[25] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6 +[26] Dimitrios Sinodinos and Narges Armanfard. Ema-net: Efficient multitask affinity learning for dense scene predictions. arXiv preprint arXiv:2401.11124, 2024. 1 + +[27] Jimmy T. H. Smith, Andrew Warrington, and Scott W. Linderman. Simplified state space layers for sequence modeling. CoRR, abs/2208.04933, 2022. 2 +[28] Guolei Sun, Thomas Probst, Danda Pani Paudel, Nikola Popovic, Menelaos Kanakis, Jagruti Patel, Dengxin Dai, and Luc Van Gool. Task switching network for multi-task learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8291-8300, 2021. 1 +[29] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 527-543. Springer, 2020. 2, 6 +[30] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 1 +[31] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 675-684, 2018. 1, 2, 6 +[32] Yangyang Xu, Xiangtai Li, Haobo Yuan, Yibo Yang, and Lefei Zhang. Multi-task learning with multi-query transformer for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6 +[33] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. CoRR, abs/2403.17695, 2024. 3 +[34] Feiyang Ye, Yueming Lyu, Xuehao Wang, Yu Zhang, and Ivor Tsang. Adaptive stochastic gradient algorithm for black-box multi-objective learning. In The Twelfth International Conference on Learning Representations, 2024. 2 +[35] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In European Conference on Computer Vision, pages 514-530. Springer, 2022. 1, 2, 6, 7 +[36] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7 +[37] Hanrong Ye and Dan Xu. Invpt++: Inverted pyramid multitask transformer for visual scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 6, 7 +[38] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2 +[39] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Rethinking of feature interaction for multi-task learning on dense prediction. arXiv preprint arXiv:2312.13514, 2023. 1, 2 + +[40] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 6 +[41] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024.3 +[42] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4514–4523, 2020. 6 +[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 2, 3 \ No newline at end of file diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/images.zip b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7f169a711ad111edd8750444995d9ca187a6422c --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:05f513ea9a96c55aa0585d9d64658009dbddad7d577443cb8b000339e7d7a3b4 +size 691314 diff --git a/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/layout.json b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..cd9793e7d40eb6bf6478de3048d99d7c712ce9a8 --- /dev/null +++ b/2025/Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction/layout.json @@ -0,0 +1,9042 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 550, + 138 + ], + "type": "text", + "content": "Enhancing Mamba Decoder with Bidirectional Interaction in Multi-Task Dense Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "spans": [ + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": "Mang Cao" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": " Sanping Zhou" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": " Yizhe Li" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": " Ye Deng" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": " Wenli Huang" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "text", + "content": " Le Wang" + }, + { + "bbox": [ + 101, + 160, + 515, + 177 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 180, + 502, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 180, + 502, + 223 + ], + "spans": [ + { + "bbox": [ + 116, + 180, + 502, + 223 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 116, + 180, + 502, + 223 + ], + "type": "text", + "content": "National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, National Engineering Research Center for Visual Information and Applications, Institute of Artificial Intelligence and Robotics, Xi'an Jiaotong University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "type": "text", + "content": " School of Computing and Artificial Intelligence, Southwestern University of Finance and Economics " + }, + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 223, + 555, + 251 + ], + "type": "text", + "content": " School of Electronic and Information Engineering, Ningbo University of Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 183, + 258, + 434, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 183, + 258, + 434, + 270 + ], + "spans": [ + { + "bbox": [ + 183, + 258, + 434, + 270 + ], + "type": "text", + "content": "3298592685@stu.xjtu.edu.cn, spzhou@xjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 298, + 200, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 298, + 200, + 311 + ], + "spans": [ + { + "bbox": [ + 152, + 298, + 200, + 311 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 324, + 297, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 324, + 297, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 324, + 297, + 613 + ], + "type": "text", + "content": "Sufficient cross-task interaction is crucial for success in multi-task dense prediction. However, sufficient interaction often results in high computational complexity, forcing existing methods to face the trade-off between interaction completeness and computational efficiency. To address this limitation, this work proposes a Bidirectional Interaction Mamba (BIM), which incorporates novel scanning mechanisms to adapt the Mamba modeling approach for multi-task dense prediction. On the one hand, we introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task-specific representations as bidirectional sequences during interaction. By integrating task-first and position-first scanning modes within a unified linear complexity architecture, BI-Scan efficiently preserves critical cross-task information. On the other hand, we employ a Multi-Scale Scan (MS-Scan) mechanism to achieve multi-granularity scene modeling. This design not only meets the diverse granularity requirements of various tasks but also enhances nuanced cross-task feature interactions. Extensive experiments on two challenging benchmarks, i.e., NYUD-V2 and PASCAL-Context, show the superiority of our BIM vs its state-of-the-art competitors. Codes are available online: https://github.com/mmm-cc/BIM_for_MTL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 636, + 135, + 648 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 636, + 135, + 648 + ], + "spans": [ + { + "bbox": [ + 55, + 636, + 135, + 648 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 296, + 693 + ], + "type": "text", + "content": "Multi-task dense prediction is a critical visual task designed to simultaneously predict outputs for various pixel-level tasks, such as semantic segmentation, depth predic" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 299, + 553, + 441 + ], + "blocks": [ + { + "bbox": [ + 317, + 299, + 553, + 441 + ], + "lines": [ + { + "bbox": [ + 317, + 299, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 553, + 441 + ], + "type": "image", + "image_path": "1b3bbfd3c79dbe5276a4fcfd18ed2c587642c24916d678b8a71e8a6003a3a93e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "text", + "content": "Figure 1. Illustration of different cross-task interaction strategies. (a) Pair-Wise [8, 28, 31]: each task can access the complete information of other tasks, while the interaction complexity is " + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "inline_equation", + "content": "O(n^{2})" + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "text", + "content": ". (b) Bridge Representation [1, 16, 24, 39]: the interaction complexity is " + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "text", + "content": ", yet, each task is limited to accessing only a partial subset of information from other tasks during interaction, which inevitably results in the loss of valuable details. (c) Ours Cross-Task Scan: It preserves the integrity of task information during interaction, with an interaction complexity of " + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "inline_equation", + "content": "O(n)" + }, + { + "bbox": [ + 313, + 458, + 555, + 558 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 581, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 664 + ], + "type": "text", + "content": "tion, surface normal estimation, and saliency detection. Existing approaches made mainly efforts [5, 30] to refine task-general representations, obtained from the encoder, into task-specific representations. In these approaches, cross-task interaction has been proven to be significant in multi-task dense prediction [26, 31, 35, 39], as it can enhance the robustness and effectiveness of task representations." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 714 + ], + "type": "text", + "content": "Initially, multi-task dense prediction methods primarily utilized Convolutional Neural Networks (CNNs) with pairwise interactions, as illustrated in Fig. 1 (a). These networks [8, 28, 31] were carefully designed with specialized" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18815" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "branches for each task, augmented by interaction modules that facilitated information exchanges between each pair of tasks. These pair-wise interaction approaches maintain the integrity of task information during interaction, promoting fine-grained cross-task interaction. While, the inherently limited receptive fields of CNN architectures often led to insufficient spatial interactions. In response to these challenges, transformer-style networks [35] demonstrate remarkable capability in modeling long-range dependencies. This capability substantially improves the interaction effectiveness of models in handling multi-task scenarios, leading to enhanced performance. However, these sufficient cross-task interactions usually introduce high interaction complexity, which typically increases quadratically with the number of tasks. The high interaction complexity limits their applicability in certain powerful models, particularly transformers." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 279, + 295, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 470 + ], + "type": "text", + "content": "To enhance the efficiency of cross-task interactions, a crowd of methods [1, 16, 24, 39] with linear interaction complexity has been proposed, as depicted in Fig. 1 (b). Typically, these methods strategically select a representative task [1, 24] or extract representative information from all tasks [16, 39] to construct a compact bridge representation that aims to preserve essential interaction information. However, in these methods, the necessity of maintaining linear interaction complexity inherently limits the representational capacity of these bridge representations. This information compression inevitably leads to the degradation of task features, potentially discarding critical interaction details and ultimately resulting in suboptimal performance. Therefore, how to enhance sufficient interaction capability while maintaining linear interaction complexity remains an unresolved issue." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 713 + ], + "type": "text", + "content": "In response to these challenges, based on Mamba architecture, we propose a simple yet effective Bidirectional Interaction Mamba (BIM) method that achieves comprehensive preservation of task information during interaction while maintaining linear interaction complexity. Specifically, we introduce a Bidirectional Interaction Scan (BI-Scan) mechanism to achieve sufficient cross-task interaction and a Multi-Scale Scan (MS-Scan) mechanism to deliver a comprehensive visual representation. The BI-Scan mechanism constructs each task representation into sequences and deploys a bidirectional scanning strategy for modeling cross-task interactions, efficiently mitigating information loss while preserving linear interaction complexity. Furthermore, BI-Scan enables multi-level cross-task interactions by integrating task-first and position-first scanning modes, facilitating more nuanced modeling of task relationships. In the MS-Scan, features are initially partitioned into multiple spaces, where the scene structure of images is modeled at various scales. Specific tasks subsequently integrate multi-scale scene structural information to" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "enhance task-specific representations. During cross-task interaction, different task branches adaptively integrate interaction representations from multiple scales, addressing the varying granularity demands of each task for scene structural information." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 328, + 133, + 533, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 133, + 533, + 143 + ], + "spans": [ + { + "bbox": [ + 328, + 133, + 533, + 143 + ], + "type": "text", + "content": "The main contributions are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 144, + 553, + 311 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 553, + 191 + ], + "type": "text", + "content": "- We propose a simple yet effective Bidricational Interaction Mamba (BIM) for multi-task dense prediction, enabling comprehensive cross-task interaction while preserving linear interaction complexity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 553, + 251 + ], + "type": "text", + "content": "- We introduce a novel Bidirectional Interaction Scan (BI-Scan) mechanism to maintain the integrity of task information during interactions. The BI-Scan incorporates two scanning modes: task-first mode and position-first mode, facilitating multi-level cross-task interactions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 251, + 553, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 553, + 311 + ], + "type": "text", + "content": "- We design a Multi-Scale Scan (MS-Scan) mechanism to capture scene structural information at multiple scales, thereby addressing the diverse granularity requirements of various tasks and enhancing more detailed cross-task interactions in multi-task dense prediction." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 323, + 400, + 335 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 569 + ], + "type": "text", + "content": "Multi-Task Learning. Most existing multi-task learning works focused on training optimization and network design. Optimization approaches balance tasks via gradient manipulation [13, 23, 34, 38] or loss-balancing [4, 14] to coordinate resource allocation for various tasks during training. The approaches of network design seek to enhance task representation learning by developing diverse mechanisms. Some CNN-based methods [29, 31] manually design interaction mechanisms to extract useful information across tasks. With the advancement of Transformer, current methods have gained improved global task interaction capabilities to enhance task representation efficiency. They enable superior pairwise [1, 24] or global [15, 32, 35, 36] task interactions. To mitigate computational complexity, MT-Mamba [16] introduce Mamba [9] to multi-task learning, which showcases effective long sequence modeling capabilities and achieving satisfactory performance. However, achieving both efficient and sufficient cross-task interaction remains a challenge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "State Space Models. In efficient long-range dependency modeling methods, state space models [11, 27] has become a striking alternative to Transformers. [10] proposed a Structured State Space Sequence (S4) model based on a new parameterization, which alleviates the computational and memory efficiency issues faced by SSM. Subsequently, numerous efforts [7, 21] are dedicated to bridging the performance disparity between SSMs and Transformers. Mamba [9] introduced an input-based parameterization method and hardware-aware algorithm, achieving performance on par with Transformers in natural language processing. This success inspired various efforts [43] towards" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "18816" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 305, + 235 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 305, + 235 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 305, + 235 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 305, + 235 + ], + "type": "image", + "image_path": "8992b268b8f720d5883f299c660f667fa679b883a3df0ee9504c680d5b35e415.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 236, + 217, + 247 + ], + "lines": [ + { + "bbox": [ + 155, + 236, + 217, + 247 + ], + "spans": [ + { + "bbox": [ + 155, + 236, + 217, + 247 + ], + "type": "text", + "content": "(a) BIM for MTL" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 307, + 72, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 307, + 72, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 307, + 72, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 307, + 72, + 553, + 235 + ], + "type": "image", + "image_path": "9a225d94f9ddce0b77bcf047531970919941810f93982158b2cd6cab243c28e6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 399, + 236, + 453, + 247 + ], + "lines": [ + { + "bbox": [ + 399, + 236, + 453, + 247 + ], + "spans": [ + { + "bbox": [ + 399, + 236, + 453, + 247 + ], + "type": "text", + "content": "(b) MFR Block" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 278 + ], + "type": "text", + "content": "Figure 2. Framework of the proposed BIM for multi-task dense prediction. (a) Overview of BIM, illustrating with depth estimation and surface normal estimation tasks. (b) Details of the MFR block, including T task-specific MSST blocks and a task-shared BCFR block." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 205, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 205, + 311 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 205, + 311 + ], + "type": "text", + "content": "Mamba's adaptation for visual tasks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 296, + 455 + ], + "type": "text", + "content": "The preservation of comprehensive image structural information poses a critical challenge in Mamba's sequential processing model, which has attracted considerable attention and effort [12, 17, 33, 41, 43]. VMamba [17] proposes the 2D Selective Scan (SS2D), a four-way scanning mechanism tailored for spatial domain traversal, aimed at enhancing Mamba's image modeling capabilities. Subsequent research studies have explored various scan patterns and combinations tailored to different tasks or scenarios [12, 33, 41]. However, by utilizing fixed token sizes, these methods overlook the importance of hierarchical spatial structural information in visual tasks." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 111, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 111, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 111, + 475 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 484, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 556 + ], + "type": "text", + "content": "We first outline the overall architecture of BIM for multitask dense prediction in Section 3.1, Subsequently, Section 3.2 focuses on the BIM Encoder, while Section 3.3 examines the BIM Decoder and its two scanning mechanisms: BI-Scan and MS-Scan. Finally, Section 3.4 discusses the optimization objectives." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 562, + 180, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 180, + 573 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 180, + 573 + ], + "type": "text", + "content": "3.1. Framework Overview" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "content": "As shown Fig. 2 (a), the proposed BIM framework consists of three main components: a task-shared encoder " + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "content": " for extracting task-generic representations, a Mamba decoder " + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "content": " for refining task features and T task-specific heads " + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\{H_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "content": " generating predictions " + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{Y}}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 579, + 296, + 650 + ], + "type": "text", + "content": " for individual tasks. This can be formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 654, + 294, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 654, + 294, + 684 + ], + "spans": [ + { + "bbox": [ + 113, + 654, + 294, + 684 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\left\\{F _ {1}, F _ {2}, \\dots , F _ {T} \\right\\} = \\Theta (\\Phi (\\mathbf {I})), \\\\ \\hat {\\mathbf {Y}} _ {t} = H _ {t} \\left(F _ {t}\\right) = \\mathcal {P} _ {t} \\left(\\Psi \\left(F _ {t}\\right)\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "abe5d3448e984cb7f1b69d6af8d6fbef2645f37b21c41251c8f110f31a673152.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " denotes the RGB input image, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " denotes the features generated by the decoder, corresponding" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": "to the distinct task " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Y}}_t" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": " represents the prediction for task " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": ", which has the same height " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": " as I. " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": " denotes a module designed to expand the task feature resolution to " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": ", which consists of a linear projection and a reshape operation. " + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_t" + }, + { + "bbox": [ + 313, + 299, + 555, + 384 + ], + "type": "text", + "content": " represents a linear layer that projects the feature channels to a specific number required by the corresponding task." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 390, + 401, + 401 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 401, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 401, + 401 + ], + "type": "text", + "content": "3.2. BIM Encoder" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "content": "The encoder shares similarities with other methods [16]. We utilized a pretrained Swin Transformer [18] to extract task-generic features, which begins by dividing the input image " + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{I} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "inline_equation", + "content": "\\frac{H}{w} \\times \\frac{W}{w}" + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "content": " tokens of dimension " + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "content": " through patch partitioning and linear layers, where " + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 407, + 555, + 515 + ], + "type": "text", + "content": " denotes the partition size. These tokens are then processed through multiple stages involving alternating patch merging and Swin Transformer block processing, ultimately yielding hierarchical image representations:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 367, + 522, + 553, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 522, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 367, + 522, + 553, + 536 + ], + "type": "interline_equation", + "content": "\\mathbf {G} = \\left\\{\\mathbf {G} ^ {1}, \\mathbf {G} ^ {2}, \\mathbf {G} ^ {3}, \\mathbf {G} ^ {4} \\right\\} = \\Phi (\\mathrm {I}), \\tag {2}", + "image_path": "2499144852f68dcb3c5809274b174c9c4dbb7a7d7b413371bb5b622edf18bd5b.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " represents the task-generic features extracted from the encoder " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ". In our practical implementation, we utilize a partition size of 4, resulting in the following shapes for " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "\\mathbf{G}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "C \\times \\frac{H}{4} \\times \\frac{W}{4}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "2C \\times \\frac{H}{8} \\times \\frac{W}{8}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "4C \\times \\frac{H}{16} \\times \\frac{W}{16}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "inline_equation", + "content": "8C \\times \\frac{H}{32} \\times \\frac{W}{32}" + }, + { + "bbox": [ + 313, + 544, + 554, + 605 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 611, + 401, + 622 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 611, + 401, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 611, + 401, + 622 + ], + "type": "text", + "content": "3.3. BIM Decoder" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "text", + "content": "The BIM Decoder consists of three Mamba Feature Refinement (MFR) blocks, as depicted in Fig. 2 (a). These MFRs are designed to bridge the gap between task-generic and task-specific representations. They generate task representations " + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_1^i,\\mathbf{F}_2^i,\\ldots ,\\mathbf{F}_T^i\\}_{i = 1}^3" + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "text", + "content": " corresponding to the first three encoder stages, with dimensions of " + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "inline_equation", + "content": "4C\\times \\frac{H}{16}\\times \\frac{W}{16},2C\\times \\frac{H}{8}\\times \\frac{W}{8}" + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "inline_equation", + "content": "C\\times \\frac{H}{4}\\times \\frac{W}{4}" + }, + { + "bbox": [ + 313, + 628, + 555, + 715 + ], + "type": "text", + "content": ", respectively. Finally, the last" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18817" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "refined features " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{F}_t^3\\}_{t = 1}^T" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " are input into the task heads as " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\{F_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " to generate the final predictions " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathbf{Y}}_t\\}_{t = 1}^T" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": "We introduce two specialized Mamba blocks, Bidirectional Cross-Task Feature Refine (BCFR) and Multi-Scale Single-task (MSST) in each MFR block, as illustrated in Fig. 2 (b). The BCFR Block is designed to facilitate effective cross-task interaction while maintaining the integrity of task information. It begins by concatenating features " + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_t^s\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": " from different task branches to construct task-shared features " + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sh}^s" + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": " using the BI-Scan mechanism. Subsequently, each task branch adaptively merges " + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sh}^s" + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": " by a selection value " + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^s" + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": " to obtain fine-tuned task representations " + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^s = \\{\\mathbf{F}_t^s\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 96, + 296, + 229 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 83, + 235, + 295, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 235, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 83, + 235, + 295, + 249 + ], + "type": "interline_equation", + "content": "\\mathbf {F} _ {t} ^ {s} = \\mathcal {F} _ {t} ^ {s} + \\mathcal {G} _ {t} ^ {s} \\times \\mathcal {F} _ {s h} ^ {s} [ t ] + (1 - \\mathcal {G} _ {t} ^ {s}) \\times \\mathcal {F} _ {t, N} ^ {s}, \\tag {3}", + "image_path": "ab1886b4e25fafc20654fd21456a0951484ac9d6d0740be3286d014b43ea4a74.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{t,N}^{s}" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " is derived from " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_t^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " through layer normalization, and " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " is obtained from " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{t,N}^{s}" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " via a linear layer followed by a Sigmoid function. " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sh}^{s}[t]" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": "-th component of " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sh}^{s}" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": "-th task. The MSST primarily aims to construct comprehensive representations through task-internal interactions. Its architecture is illustrated in Fig. 2 (b) and comprises two main branches: the scan branch and the gating branch. In the scan branch, we integrate a hierarchical global scene structure representation " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " using a novel MS-Scan mechanism. Simultaneously, a gating signal " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{G}^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " from the gate branch regulates the flow of information. Subsequently, we adjust the channel dimensions of the multi-scale scene representation by applying a linear projection " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " and establish a residual connection between " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": " and the input " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{in}^{s}" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": ". This process is repeated twice to produce the final output " + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^s" + }, + { + "bbox": [ + 55, + 255, + 296, + 447 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 454, + 295, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 454, + 295, + 468 + ], + "spans": [ + { + "bbox": [ + 120, + 454, + 295, + 468 + ], + "type": "interline_equation", + "content": "\\mathcal {F} ^ {s} = \\mathcal {F} _ {i n} ^ {s} + \\mathcal {P} \\left(\\mathcal {F} _ {h} ^ {s} \\times \\mathcal {G} ^ {s}\\right). \\tag {4}", + "image_path": "fa57e8a4449688db21b0abeb85efc3706dc61868a9296683c646929676976623.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "text", + "content": "Bidirectional Interaction Scan. To maintain task information integrity during interaction with linear complexity, we propose a novel BI-Scan that serves as a cornerstone of BIM, as shown in Fig. 3. This module takes the representations of different tasks " + }, + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}_{t,N}^s\\}_{t = 1}^T" + }, + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "text", + "content": " as input, which is initially partitioned along the channel dimension into dual-branch representations, each with the dimensions " + }, + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "inline_equation", + "content": "H\\times W\\times \\frac{C}{2}" + }, + { + "bbox": [ + 55, + 474, + 296, + 628 + ], + "type": "text", + "content": ". The former undergoes processing through our Forward Scan module, while the latter is reversed along the task dimension before the Backward Scan processing. Specifically, the Forward Scan architecture implements hierarchical cross-task interaction through two scanning modes: Task-First mode and Position-First mode." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "Task-First Mode: Following a similar configuration to SS2D in Vmamba [17], we implement multiple distinct Scan Patterns for each scanning mode to facilitate enhanced task interaction. Without loss of generality, we use the first scan pattern as an example. As shown in Fig. 3, in the task-first mode, the construction of global task information contains two processes: (1) Pattern-guided feature serialization" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 553, + 269 + ], + "type": "image", + "image_path": "9df8236e2bae9e178842067166da01f9f7fdc93fbf8de7b22d8f3e607bd8d9e0.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 312 + ], + "type": "text", + "content": "Figure 3. Framework of BI-Scan. It consists of two modules, Forward Scan and Backward Scan, each including Task-First and Position-First cross-task scanning modes." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "(\\mathcal{U}_1)" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": ": each task feature is serialized according to the scan pattern, resulting in " + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": " subsequences, each of length " + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": ". (2) Task order-guided subsequence aggregation " + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "(\\mathcal{O}_1)" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": ": these subsequences are concatenated along the length dimension according to the specified task order, producing a global sequence " + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": " of length " + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "inline_equation", + "content": "T \\times (H \\times W)" + }, + { + "bbox": [ + 313, + 333, + 555, + 417 + ], + "type": "text", + "content": ". These processes can be expressed as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 426, + 555, + 442 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 426, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 321, + 426, + 555, + 442 + ], + "type": "interline_equation", + "content": "S _ {1} = \\mathcal {O} _ {1} \\left(\\mathcal {U} _ {1} \\left(\\mathcal {F} _ {t} ^ {s}, \\mathcal {F} _ {2} ^ {s}, \\dots , \\mathcal {F} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(T \\times (H \\times W)\\right) \\times \\frac {C}{2}}. \\tag {5}", + "image_path": "03c84fa016e65a38b28711191b8f21401d7e43bd3c04d15444013dcc9359cfce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 451, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 474 + ], + "type": "text", + "content": "This sequence undergoes global modeling through SSM [9], yielding task-level interaction information." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": "Position-First Mode: In the position-first mode, the output of Task-First mode is first restored from the sequence to the input task feature size, recorded as " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "\\{\\bar{F}_t^s\\}_{t=1}^T \\in \\mathbb{R}^{H \\times W \\times \\frac{C}{2}}" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ". Subsequently, the construction of interaction representations also involves two steps: (1) Task order-guided feature serialization " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\mathcal{O}_2)" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ": Spatially aligned task representation tokens are constructed into " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": " subsequences, each of length " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ", according to the specified task order. (2) Pattern-guided subsequence aggregation " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "(\\mathcal{U}_2)" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ": These subsequences are concatenated along the length dimension according to the scan pattern, resulting in a sequence of length " + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "inline_equation", + "content": "(H \\times W) \\times T" + }, + { + "bbox": [ + 313, + 475, + 555, + 631 + ], + "type": "text", + "content": ". The above processes can be expressed as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 640, + 555, + 657 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 640, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 321, + 640, + 555, + 657 + ], + "type": "interline_equation", + "content": "S _ {2} = \\mathcal {U} _ {2} \\left(\\mathcal {O} _ {2} \\left(\\bar {\\mathcal {F}} _ {1} ^ {s}, \\bar {\\mathcal {F}} _ {2} ^ {s}, \\dots , \\bar {\\mathcal {F}} _ {T} ^ {s}\\right)\\right) \\in \\mathbb {R} ^ {\\left(\\left(H \\times W\\right) \\times T\\right) \\times \\frac {C}{2}}. \\tag {6}", + "image_path": "0cb65e99c96cd5ef4ef88f1a593c6856c8274eae32bc21ef7f923262c3a2c6a9.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 689 + ], + "type": "text", + "content": "This sequence is also processed through SSM, allowing for the derivation of task-level interaction insights." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Subsequently, the outputs corresponding to all patterns are fused via element-wise summation, yielding the output" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18818" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 231 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 231 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 231 + ], + "type": "image", + "image_path": "dca41e1f6e1d2fafed6c1200ece592ad6c004262a37659728b0f04e02b43d801.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "type": "text", + "content": "Figure 4. Instructions for MS-Scan. It consists of three distinct operations, Input Handling, Multi-Scale Scanning, and Multi-Scale Fusion. In Multi-Scale Scanning, each branch " + }, + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "type": "text", + "content": " has a distinct scanning size " + }, + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 55, + 239, + 296, + 284 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": "of Forward Scan, denoted as " + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{for}^{s}" + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": ". Finally, the output of the Backward Scan, " + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{back}^{s}" + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": ", is reversed along the task dimension and concatenated with " + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{for}^{s}" + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": " to yield the refined task representations " + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{sh}^{s}" + }, + { + "bbox": [ + 55, + 305, + 296, + 413 + ], + "type": "text", + "content": ". In practical application, when employing BI-Scan independently, the selection of Scan Patterns aligns with those used in SS2D [17], specifically, the four directions shown in Fig. 3. When MS-Scan is introduced within the BI-Scan module, we incorporate Scan Patterns of varying scales without increasing computational complexity." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": "Multi-Scale Scan. Mamba processes features by converting them into sequences of tokens. While a variety of tokenization strategies exist for text, image tokenization requires careful consideration of multiple approaches. Previous research has demonstrated that multi-scale processing is particularly effective for image data. To leverage this advantage, We propose the MS-Scan mechanism that requires minimal additional computational cost compared to single-scale scanning, as illustrated in Fig. 4 (a). In this framework, we employ multiple scanning scales, denoted as " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\{s_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ", in multiple branches " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{B}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ", and transform the input image feature " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": " into token sequences of varying dimensions for Mamba modeling. For instance, the initial image feature can be tokenized into a sequence with a total length of " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ", where each token has a dimension of " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ". When applying a scanning scale of " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "s_i = 2" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ", the image is divided into non-overlapping patches, resulting in a tokenized feature sequence with a dimensionality of " + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "inline_equation", + "content": "4C \\times \\left(\\frac{H}{2} \\times \\frac{W}{2}\\right)" + }, + { + "bbox": [ + 55, + 413, + 296, + 664 + ], + "type": "text", + "content": ". Specifically, MS-Scan comprises three key components: input handling, multi-scale scanning, and multi-scale fusion." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Input Handling: To construct inputs for " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " different scanning branches " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{B} = \\{\\mathcal{B}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", we perform two main operations. (1) Channel Split (" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "): We begin by splitting the input representation " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " sub-features " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{x_i\\}_{i=1}^N \\in" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{m\\times H\\times W}" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": " along the channel dimension, where " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "m = \\frac{C}{N}" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": ". (2) Window Tokenization " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "(\\mathcal{W}_i)" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": ": For the " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": "-th branch " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": " with scanning scale " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": ", we first divide " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\frac{H}{s_i}\\times \\frac{W}{s_i}" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": " nonoverlapping patches, each of size " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "m\\times s_i\\times s_i" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": ". Subsequently, we concatenate the pixel feature values in each patch along the channel dimension, resulting in the scan input " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "\\bar{x}_i" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": " with a shape of " + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "inline_equation", + "content": "(m\\times s_i\\times s_i)\\times \\frac{H}{s_i}\\times \\frac{W}{s_i}" + }, + { + "bbox": [ + 313, + 71, + 553, + 168 + ], + "type": "text", + "content": ". Ultimately, we obtain the input for all branches:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 328, + 174, + 553, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 174, + 553, + 188 + ], + "spans": [ + { + "bbox": [ + 328, + 174, + 553, + 188 + ], + "type": "interline_equation", + "content": "\\{\\bar {x} _ {1}, \\bar {x} _ {2}, \\dots , \\bar {x} _ {N} \\} = \\left\\{\\mathcal {W} _ {1}, \\mathcal {W} _ {2}, \\dots , \\mathcal {W} _ {N} \\right\\} (\\mathcal {S} (x)). \\tag {7}", + "image_path": "2bba252298d50b60fb5c19ec8959b7886865682afbb3bdaeb842d4d2f94540fd.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "content": "Multi-Scale Scanning: Following the input handling process, we employ distinct scanning scales to construct a multi-scale scene representation in each branch. For all branches " + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "content": ", we utilize the four-way scanning method (SS2D) from VMamba [17] to generate scene features at the specified scale. This method creates four token sequences, each shaped as " + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "inline_equation", + "content": "C_i \\times (H_i \\times W_i)" + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "content": ", by scanning the input features " + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\bar{x}_i \\in \\mathbb{R}^{C_i \\times H_i \\times W_i}" + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "content": " in four directions. The resulting sequences are then processed by SSM [9] and combined to produce the output feature " + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "inline_equation", + "content": "\\bar{y} = \\{\\bar{y}_i\\}_{i=1}^N" + }, + { + "bbox": [ + 313, + 194, + 555, + 315 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 319, + 553, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 319, + 553, + 333 + ], + "spans": [ + { + "bbox": [ + 325, + 319, + 553, + 333 + ], + "type": "interline_equation", + "content": "\\bar {y} _ {i} = \\operatorname {S S 2 D} (\\bar {x} _ {i}) \\in \\mathbb {R} ^ {C _ {i} \\times H _ {i} \\times W _ {i}}, i \\in \\{1, 2, \\dots , N \\}. \\tag {8}", + "image_path": "2428bbec6f7cd474f5293c8a7680ce0f1d0c8ae63382603e2951fb1aa55af39d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": "Multi-Scale Fusion: In our approach to multi-scale feature fusion, we adopt a methodology that reverses the input handling process, consisting of two key steps. (1) Token Windowing " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "(\\mathcal{W}_i^{-1})" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ": For each branch " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ", we split each pixel feature into " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "s_i \\times s_i" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": " segments along the channel dimension, resulting in a set of feature segments " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\{\\bar{y}_{i,j}\\}_{j=1}^{s_i \\times s_i} \\in \\mathbb{R}^{m \\times 1 \\times 1}" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ". These segments are then concatenated along the spatial dimensions (height and width) to form patches, which are subsequently combined to produce the output for " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{B}_i" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ". (2) Channel Concatenation " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "(S^{-1})" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ": We concatenate the features from all branches along the channel dimension, yielding the final output feature " + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 313, + 339, + 555, + 484 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 501, + 553, + 517 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 501, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 330, + 501, + 553, + 517 + ], + "type": "interline_equation", + "content": "y = \\mathcal {S} ^ {- 1} \\left(\\mathcal {W} _ {1} ^ {- 1} (\\bar {y} _ {1}), \\mathcal {W} _ {2} ^ {- 1} (\\bar {y} _ {2}), \\dots , \\mathcal {W} _ {N} ^ {- 1} (\\bar {y} _ {N})\\right), \\tag {9}", + "image_path": "b000f2e2906a3e220763efc1cdc4334926272f89b4d6dcabf62bf88785f13c1c.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "S^{-1}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_i^{-1}" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " refer to the inverse operation of " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{W}_i" + }, + { + "bbox": [ + 313, + 521, + 553, + 545 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 552, + 447, + 564 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 447, + 564 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 447, + 564 + ], + "type": "text", + "content": "3.4. Optimization Objective" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": "We jointly train all tasks to optimize BIM encoder " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": ", BIM decoder " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": ", and task heads " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "\\{H\\}_{i=1}^{T}" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": " simultaneously. To maintain consistency with previous approaches, we use " + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 313, + 568, + 553, + 640 + ], + "type": "text", + "content": " loss for depth estimation and surface normal estimation tasks and the cross-entropy loss for other tasks, therefore, the optimization objective can be expressed as follows:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 386, + 647, + 553, + 672 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 647, + 553, + 672 + ], + "spans": [ + { + "bbox": [ + 386, + 647, + 553, + 672 + ], + "type": "interline_equation", + "content": "L = \\sum_ {t \\in \\mathcal {T}} \\lambda_ {t} \\mathcal {L} _ {t} \\left(\\hat {\\mathbf {Y}} _ {t}, \\mathbf {Y} _ {t}\\right), \\tag {10}", + "image_path": "bf1eb784aca90c54be1f91269cbf3006112cb2e17eb698bf649026024661a449.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " is the set of all tasks, " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_t" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_t" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{Y}_t" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": " are the loss weight, loss function, and task label for image I in task " + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18819" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 132, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 132, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 132, + 85 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 171, + 105 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 110, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 110, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 110, + 295, + 253 + ], + "type": "text", + "content": "Datasets. We performed experiments using the benchmark datasets NYUD-v2 [25] and PASCAL-Context [3]. NYUD-v2 primarily focuses on indoor scenes, with 795 and 654 RGB images for training and testing purposes. Tasks in NYUD-v2 include semantic segmentation, monocular depth estimation, surface normal estimation, and object boundary detection. PASCAL-Context encompasses indoor and outdoor scenes, offering pixel-level labels for tasks like semantic segmentation, human parsing, object boundary detection, surface normal estimation, and saliency detection tasks. The PASCAL-Context dataset contains 4,998 training images and 5,105 test images." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "content": "Implementation Details. We employ a pretrained SwinLarge Transformer [18] on ImageNet-22K [6] as our encoder. Our models are trained on the NYUD-v2 dataset for 50,000 iterations with a batch size of 4, and on the Pascal Context dataset for 75,000 iterations with a batch size of 6. Across all datasets, we use the Adam optimizer with a learning rate of " + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "inline_equation", + "content": "5 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "content": " and a weight decay rate of " + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "content": ", alongside a polynomial learning rate scheduler. The preliminary decoder has an output channel number of 768. We follow common practice [16, 35] in resizing the input images and applying data augmentation techniques. Specifically, we resize the input images of NYUD-v2 and PASCAL-Context to " + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "inline_equation", + "content": "448 \\times 576" + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 55, + 254, + 296, + 434 + ], + "type": "text", + "content": ", respectively, and apply random color jittering, random cropping, random scaling, and random horizontal flipping." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 443, + 283, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 283, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 283, + 456 + ], + "type": "text", + "content": "4.2. Comparisons with State-of-the-art Methods" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "content": "Main Results. Tab. 1 and Tab. 2 report a comparison of the proposed BIM against previous state-of-the-art methods, including MTmamba [16], MQTransformer [32], TaskPrompter [36], " + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{InvPT}++" + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "content": " [37], " + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "inline_equation", + "content": "\\mathrm{InvPT}" + }, + { + "bbox": [ + 55, + 461, + 295, + 616 + ], + "type": "text", + "content": " [35], ATRC [2], MTI-Net [29], PAD-Net [31], PSD [40], PAP [42], Cross-Stitch [22] and ASTMT [20] on NYUD-V2 and PASCAL-Context dataset respectively. Notably, the previous best method, i.e., MTmamba, and our BIM are built upon the Transformer-encoder Mamba-decoder architecture with the same backbone. On NYUD-v2, the performance of Semseg is clearly boosted from the previous best, i.e., 55.82 to 57.40 (+1.52). On Pascal-Context, we achieved superior performance on all tasks compared to MTMamba." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Comparisons on Model Size and Computational Cost. In Tab. 3, we compare the FLOPs and parameter counts with other State-of-the-art methods. With the comparable computational cost, our BIM shows significant outperforms. our light-vision BIM (DBIM), which implements dilated sampling in MS-Scan, shows outperformance with reduced computational cost and parameter numbers, showcasing the advantage of these designs. Specifically, in MS" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 324, + 102, + 545, + 288 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with other state-of-the-art methods on NYUD-v2 dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 324, + 102, + 545, + 288 + ], + "lines": [ + { + "bbox": [ + 324, + 102, + 545, + 288 + ], + "spans": [ + { + "bbox": [ + 324, + 102, + 545, + 288 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
CNN based
Cross-Stitch[22]36.340.629020.8876.38
PAP[42]36.720.617820.8276.42
PSD[40]36.690.624620.8776.42
PAD-Net[31]36.610.627020.8576.38
MTI-Net[29]45.940.536520.2777.86
ATRC[2]46.330.536320.1877.94
Transformer based
InvPT[35]53.660.518319.0478.10
TaskPrompter[36]55.300.515218.4778.20
InvPT++[37]53.850.509618.6778.10
MQTransformer[32]54.840.532519.6778.20
Mamba based
MTMamba[16]55.820.506618.6378.70
BIM(Ours)57.400.473318.5578.72
", + "image_path": "3f8722c62cc42a19babf7cc1b2449546489509c804a92f68afde7c893a74a762.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 316, + 337, + 553, + 525 + ], + "blocks": [ + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 328 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with other state-of-the-art methods on Pascal-Context dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 337, + 553, + 525 + ], + "lines": [ + { + "bbox": [ + 316, + 337, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 337, + 553, + 525 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Parsing mIoU ↑Saliency maxF↑Normal mErr ↓Boundary odsF ↑
CNN based
PAD-Net[31]53.6059.6065.8015.3072.50
ASTMT[20]68.0061.1065.7014.7072.40
MTI-Net[29]61.7060.1884.7814.2370.80
ATRC[2]62.6959.4284.7014.2070.96
ATRC-ASPP[2]63.6060.2383.9114.3070.86
ATRC-BMTAS[2]67.6762.9382.2914.2472.42
Transformer based
InvPT[35]79.0367.6184.8114.1573.00
TaskPrompter[36]80.8968.8984.8313.7273.50
InvPT++[37]80.2269.1284.7413.7374.20
MQTransformer[32]78.9367.4183.5814.2173.90
Mamba based
MTmamba[16]81.1172.6284.1414.1478.80
BIM(Ours)81.2573.1584.1414.1379.05
", + "image_path": "feb10b2698bdac34ab17f3868588fcdb8d430f6a43d5b874f88991063878cbf7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 553, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 553, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 312, + 553, + 555, + 662 + ], + "type": "text", + "content": "Scan, we use dilated sampling in each scan branch " + }, + { + "bbox": [ + 312, + 553, + 555, + 662 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 312, + 553, + 555, + 662 + ], + "type": "text", + "content": " to lighten the operation of MS-Scan. Specifically: (1) We only sample one token in a window when generating multi-scale sequences from image features, instead of all tokens. (2) When restoring the sequence to image features, we perform linear interpolation between these selected tokens. These operations do not introduce any parameters, and the computational burden is reduced because only a subset of tokens are used for modeling." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Qualitative Comparisons. We qualitatively compared our proposed BIM with the previous best-performing method, as shown in Fig. 5. Our method shows clear improvements in detail, as highlighted in the circled regions." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "18820" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 250 + ], + "type": "image", + "image_path": "e4bed02f21f0e9d7922c638ffa22122785dcd318e543231656ca2b51a64263f6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 82, + 259, + 526, + 271 + ], + "lines": [ + { + "bbox": [ + 82, + 259, + 526, + 271 + ], + "spans": [ + { + "bbox": [ + 82, + 259, + 526, + 271 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison with the best-performing method on Pascal-Context. Our method generates better details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 321, + 293, + 403 + ], + "blocks": [ + { + "bbox": [ + 55, + 290, + 295, + 312 + ], + "lines": [ + { + "bbox": [ + 55, + 290, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 295, + 312 + ], + "type": "text", + "content": "Table 3. Complexity comparisons on NYUD-v2. †denotes the result is from [19]." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 321, + 293, + 403 + ], + "lines": [ + { + "bbox": [ + 59, + 321, + 293, + 403 + ], + "spans": [ + { + "bbox": [ + 59, + 321, + 293, + 403 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑FLOPs (G)↓# Params (M)↓
InvPT[35]53.660.518319.0478.10598402
TaskPrompter[36]55.300.515218.4778.20470392
InvPT++†[37]53.850.509618.6778.10-~402
MTMamba[16]55.820.506618.6378.70541308
DBIM (Ours)56.410.478918.6678.80495290
BIM (Ours)57.400.473318.5578.72547388
", + "image_path": "b0b70a5ca3519cb85bc3c99e67e0873ab745a566c950e12e81e612aa2be0550d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 62, + 418, + 293, + 506 + ], + "blocks": [ + { + "bbox": [ + 62, + 418, + 293, + 506 + ], + "lines": [ + { + "bbox": [ + 62, + 418, + 293, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 418, + 293, + 506 + ], + "type": "image", + "image_path": "297f220a630441873597088e4ac056b5b32573222b5e1d801ea58f24ee81b565.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 543 + ], + "type": "text", + "content": "Figure 6. Ablation study on Bidirectional Scan (left) and two scanning modes (right) in BI-Scan." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 149, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 577 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of Different Components. We performed ablation experiments on BI-Scan and MS-Scan using the NYUD-V2 dataset. These experiments all used Swin-Large Transformer as encoder. We use the same STL and MTL base model as in MTMamba. (1) The term \"STL\" indicates that each task is equipped with a separate network, utilizing two Swin Transformer blocks in each decoder stage, and (2) the \"MTL\" indicates that all tasks share a common encoder. (3) \"Baseline\" refers to MTMamba. (4) \"BI-Scan\" indicates that the MS-Scan is replaced with SS2D throughout the entire model. (5) \"MS-Scan\" refers to the implementation of" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 310, + 553, + 398 + ], + "blocks": [ + { + "bbox": [ + 321, + 290, + 545, + 301 + ], + "lines": [ + { + "bbox": [ + 321, + 290, + 545, + 301 + ], + "spans": [ + { + "bbox": [ + 321, + 290, + 545, + 301 + ], + "type": "text", + "content": "Table 4. Effectiveness of different components on NYUD-v2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 310, + 553, + 398 + ], + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 398 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑FLOPs # (G) ↓Params (M) ↓
STL54.320.516619.2177.30+0.001075889
MTL53.720.523919.9776.50-1.87466303
Baseline55.820.506618.6378.70+2.38541308
BI-Scan57.110.485618.6678.90+4.02547290
MS-Scan57.400.482918.7978.60+4.00541397
BIM57.400.473318.5578.72+4.83547388
", + "image_path": "b7f570a6b01f6928d34c73c54e0ba4e33644c666c2940aa253f839fd67c9eba8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 506 + ], + "type": "text", + "content": "the same cross-task interaction strategy as that used in the baseline. Lastly, (6) \"BIM\" represents the default method. As depicted in Tab. 4, The BI-Scan improved model performance with fewer parameters, emphasizing the necessity of retaining task-specific details in cross-task interactions. Meanwhile, the MS-Scan enabled multi-granularity scene modeling, further enhancing performance." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 555, + 592 + ], + "type": "text", + "content": "Effectiveness of Bidirectional Scan in BI-Scan. As shown in Fig. 6 (left), we compared the performance of unidirectional (Uni) and bidirectional (Bi) cross-task scanning. The results show that the introduction of bidirectional scanning significantly enhances the cross-task modeling capability, thereby improving the overall performance of the model with only a negligible increase in computational cost." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Effectiveness of Two Scan Modes in BI-Scan. As shown in Fig. 6 (right), we conducted ablation experiments on the Task-First Scan mode and Position-First Scan mode within BI-Scan. The results indicate that both Task-First (Task) and Position-First (Position) interaction modes achieve superior performance compared to the baseline (Base) at a lower computational cost. Furthermore, employing both modes together (All) results in further performance improvements. Ultimately, we incorporate both Task-First mode and Position-First mode as the final configuration of" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18821" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 129, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 129, + 83 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 129, + 83 + ], + "type": "text", + "content": "BI-Scan for BIM." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 85, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 85, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 57, + 85, + 294, + 251 + ], + "type": "text", + "content": "Ablation on Task Order in BI-Scan. We experiment on NYUD-v2 to study the effect of the task order during scan interaction. In addition to fixed task orders, we also implemented a random task order modeling approach (Random), in which the order of tasks in BI-Scan is randomly shuffled during both training and testing. For clarity, we refer to the tasks Semseg, Depth, Normal, and Boundary as S, D, N, and B, respectively. As shown in Tab. 5, the overall performance of various fixed order models is relatively similar, while the Random setting exhibits a slight decrease in overall performance, which remains within an acceptable range. This indicates that our proposed BI-Scan demonstrates robustness to the task order. Ultimately, we adopt the first task order setting as the final configuration of BI-Scan for BIM." + } + ] + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 283, + 294, + 367 + ], + "blocks": [ + { + "bbox": [ + 93, + 263, + 259, + 274 + ], + "lines": [ + { + "bbox": [ + 93, + 263, + 259, + 274 + ], + "spans": [ + { + "bbox": [ + 93, + 263, + 259, + 274 + ], + "type": "text", + "content": "Table 5. Ablation on task order on NYUD-v2." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 283, + 294, + 367 + ], + "lines": [ + { + "bbox": [ + 58, + 283, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 58, + 283, + 294, + 367 + ], + "type": "table", + "html": "
OrderSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
S-D-N-B57.110.485618.6678.90+4.02
N-S-D-B56.150.483218.6679.00+3.72
B-S-D-N56.970.483618.7378.90+3.96
S-N-D-B57.520.484318.7178.90+4.20
Random55.410.486118.6978.90+3.17
", + "image_path": "5c543644021481dd9c778aa7ca324a6ac58176cce840ea9799861ad9660406a1.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "spans": [ + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": "Ablation on Scan Scale and Number of MS-Scan. We experimented with the impact of varying scanning scales and numbers on model performance, as shown in Tab. 6 and Tab. 7. Experiments were conducted on the NYUD-v2 dataset with Swin-Large Transformer as encoder. We experimented with three different settings for scale ablation: all MFRs use " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,2\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,3\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,4\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": " two scanning scales, respectively. For number ablation, We compared three different settings, Type 1: all three MFRs use " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": " scale; Type 2: all three MFRs use " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,4\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": " two scales; Type 3: three MFRs use " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,2\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,2,4\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,2,4,6\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": ", respectively. The results showed that in multi-scale scanning, employing appropriate scanning scale partitions can effectively enhance overall performance. Significantly, all settings yield notable performance enhancements. Ultimately, we adopt " + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "inline_equation", + "content": "\\{1,4\\}" + }, + { + "bbox": [ + 57, + 380, + 294, + 571 + ], + "type": "text", + "content": " scale setting for all MFRs as the final configuration." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 603, + 293, + 663 + ], + "blocks": [ + { + "bbox": [ + 63, + 582, + 287, + 594 + ], + "lines": [ + { + "bbox": [ + 63, + 582, + 287, + 594 + ], + "spans": [ + { + "bbox": [ + 63, + 582, + 287, + 594 + ], + "type": "text", + "content": "Table 6. Different scanning scales of MS-Scan on NYUD-v2." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 603, + 293, + 663 + ], + "lines": [ + { + "bbox": [ + 58, + 603, + 293, + 663 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 293, + 663 + ], + "type": "table", + "html": "
Scan ScaleSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
{1,2}57.390.475618.6078.80+4.68
{1,3}57.530.473918.6878.50+4.62
{1,4}57.400.473318.5578.72+4.83
", + "image_path": "df744a677ae74b7f539a91b34fbce08bfed44ef04e0e567e03ec957ea7ae3bf9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 677, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 677, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 677, + 294, + 712 + ], + "type": "text", + "content": "Performance on Different Encoders. We evaluate the effect of model size on experimental performance, as presented in Tab. 8. All experiments were conducted on the" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 318, + 91, + 552, + 148 + ], + "blocks": [ + { + "bbox": [ + 320, + 71, + 549, + 82 + ], + "lines": [ + { + "bbox": [ + 320, + 71, + 549, + 82 + ], + "spans": [ + { + "bbox": [ + 320, + 71, + 549, + 82 + ], + "type": "text", + "content": "Table 7. Different scanning number of MS-Scan on NYUD-v2." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 91, + 552, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 91, + 552, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 91, + 552, + 148 + ], + "type": "table", + "html": "
Scan NumberSemseg mIoU ↑Depth RMSE ↓Normal mErr ↓Boundary odsF ↑MTL Gain Δm ↑
Type 157.110.485618.6678.90+4.02
Type 257.400.473318.5578.72+4.83
Type 356.760.475618.5978.70+4.35
", + "image_path": "5cbf69a1b10b62cfd697f6e3926cfcebc7fb6c11e4cd419d885422147c311bea.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 172, + 553, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 255 + ], + "type": "text", + "content": "NYUD-v2 dataset. We compare our method with the previous best-performing model, MTMamba, using two different encoders: Swin-Base Transformer (denoted as '-Base') and Swin-Large Transformer (denoted as '-Large'). The results suggest that models with greater capacities typically exhibit superior performance. Furthermore, our approach has demonstrated superior performance across all encoder." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 319, + 289, + 547, + 358 + ], + "blocks": [ + { + "bbox": [ + 354, + 270, + 514, + 280 + ], + "lines": [ + { + "bbox": [ + 354, + 270, + 514, + 280 + ], + "spans": [ + { + "bbox": [ + 354, + 270, + 514, + 280 + ], + "type": "text", + "content": "Table 8. Performance on different encoders." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 289, + 547, + 358 + ], + "lines": [ + { + "bbox": [ + 319, + 289, + 547, + 358 + ], + "spans": [ + { + "bbox": [ + 319, + 289, + 547, + 358 + ], + "type": "table", + "html": "
ModelSemseg mIoU ↑Depth RMSE↓Normal mErr ↓Boundary odsF ↑
MTMamba-Base53.620.512619.2877.70
BIM-Base54.320.491518.8977.74
MTMamba-Large55.820.506618.6378.70
BIM-Large57.400.473318.5578.72
", + "image_path": "a1bb0864b5ead5effa0e52b37ff71d4e3ff4f69c168ba3a9f2f7a5e16474904d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 387, + 386, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 386, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 386, + 399 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 408, + 553, + 600 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 553, + 600 + ], + "type": "text", + "content": "We introduce the Bidirectional Interaction Mamba (BIM) framework, a simple yet effective solution designed to address the inherent conflict between adequacy and complexity in cross-task interactions for multi-task dense prediction. We first propose a Bidirectional Interaction Scan (BI-Scan) mechanism, which constructs task information as bidirectional sequences to model cross-task interactions efficiently. By integrating task-first and position-first modes within linear complexity framework, BI-Scan preserves the integrity of task-specific information during interaction while maintaining computational efficiency. We further develop a Multi-Scale Scan (MS-Scan) mechanism, which captures scene structure information at multiple scales, alleviating the difficulty of feature learning and enhancing nuanced cross-task feature interactions. Extensive experiments show our method significantly enhances model performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 615, + 423, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 615, + 423, + 628 + ], + "spans": [ + { + "bbox": [ + 316, + 615, + 423, + 628 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 635, + 553, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 702 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 702 + ], + "type": "text", + "content": "This work was supported in part by National Natural Science Foundation of China under Grants 62088102, U24A20325 and 12326608, Key Research and Development Plan of Shaanxi Province under Grant 2024PT-ZCK-80, and the Public Welfare Research Program of Ningbo City under Grant 2024S063." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18822" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Deblina Bhattacharjee, Tong Zhang, Sabine Susstrunk, and Mathieu Salzmann. Mult: An end-to-end multitask learning transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12031-12041, 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] David Brüggemann, Menelaos Kanakis, Anton Obukhov, Stamatios Georgoulis, and Luc Van Gool. Exploring relational context for multi-task dense prediction. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15869-15878, 2021. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 269 + ], + "type": "text", + "content": "[3] Xianjie Chen, Roozbeh Mottaghi, Xiaobai Liu, Sanja Fidler, Raquel Urtasun, and Alan Yuille. Detect what you can: Detecting and representing objects using holistic models and body parts. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1971-1978, 2014. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 271, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 271, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 271, + 295, + 326 + ], + "type": "text", + "content": "[4] Zhao Chen, Vijay Badrinarayanan, Chen-Yu Lee, and Andrew Rabinovich. Gradnorm: Gradient normalization for adaptive loss balancing in deep multitask networks. In International conference on machine learning, pages 794-803. PMLR, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 327, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 350 + ], + "type": "text", + "content": "[5] Michael Crawshaw. Multi-task learning with deep neural networks: A survey. CoRR, abs/2009.09796, 2020. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 396 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 451 + ], + "type": "text", + "content": "[7] Daniel Y Fu, Tri Dao, Khaled Kamal Saab, Armin W Thomas, Atri Rudra, and Christopher Re. Hungry hungry hippos: Towards language modeling with state space models. In The Eleventh International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 453, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 453, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 453, + 295, + 509 + ], + "type": "text", + "content": "[8] Yuan Gao, Jiayi Ma, Mingbo Zhao, Wei Liu, and Alan L Yuille. Nddr-cnn: Layerwise feature fusing in multi-task cnns by neural discriminative dimensionality reduction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3205-3214, 2019. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 295, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 532 + ], + "type": "text", + "content": "[9] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces, 2024. 2, 4, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 295, + 566 + ], + "type": "text", + "content": "[10] Albert Gu, Karan Goel, and Christopher Ré. Efficiently modeling long sequences with structured state spaces. CoRR, abs/2111.00396, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 622 + ], + "type": "text", + "content": "[11] Albert Gu, Isys Johnson, Karan Goel, Khaled Saab, Tri Dao, Atri Rudra, and Christopher Ré. Combining recurrent, convolutional, and continuous-time models with linear state space layers. Advances in neural information processing systems, 34:572-585, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 667 + ], + "type": "text", + "content": "[12] Tao Huang, Xiaohuan Pei, Shan You, Fei Wang, Chen Qian, and Chang Xu. Localmamba: Visual state space model with windowed selective scan. arXiv preprint arXiv:2403.09338, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[13] Woeseong Jeong and Kuk-Jin Yoon. Quantifying task priority for multi-task optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 363-372, 2024. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[14] Alex Kendall, Yarin Gal, and Roberto Cipolla. Multi-task learning using uncertainty to weigh losses for scene geometry and semantics. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7482-7491, 2018. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 554, + 174 + ], + "type": "text", + "content": "[15] Wei-Hong Li, Steven McDonagh, Ales Leonardis, and Hakan Bilen. Multi-task learning with 3d-aware regularization. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 317, + 176, + 554, + 230 + ], + "type": "text", + "content": "[16] Baijiong Lin, Weisen Jiang, Pengguang Chen, Yu Zhang, Shu Liu, and Ying-Cong Chen. MTMamba: Enhancing multi-task dense scene understanding by mamba-based decoders. In European Conference on Computer Vision, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 233, + 554, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 233, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 233, + 554, + 276 + ], + "type": "text", + "content": "[17] Yue Liu, Yunjie Tian, Yuzhong Zhao, Hongtian Yu, Lingxi Xie, Yaowei Wang, Qixiang Ye, and Yunfan Liu. Vmamba: Visual state space model. CoRR, abs/2401.10166, 2024. 3, 4, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 279, + 554, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 279, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 279, + 554, + 334 + ], + "type": "text", + "content": "[18] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 335, + 554, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 335, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 335, + 554, + 380 + ], + "type": "text", + "content": "[19] Yuxiang Lu, Shengcao Cao, and Yu-Xiong Wang. Swiss army knife: Synergizing biases in knowledge from vision foundation models for multi-task learning. arXiv preprint arXiv:2410.14633, 2024. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 382, + 554, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 382, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 317, + 382, + 554, + 426 + ], + "type": "text", + "content": "[20] Kevis-Kokitsi Maninis, Ilija Radosavovic, and Iasonas Kokkinos. Attentive single-tasking of multiple tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1851-1860, 2019. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 428, + 554, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 428, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 554, + 461 + ], + "type": "text", + "content": "[21] Harsh Mehta, Ankit Gupta, Ashok Cutkosky, and Behnam Neyshabur. Long range language modeling via gated state spaces. arXiv preprint arXiv:2206.13947, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 463, + 554, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 463, + 554, + 507 + ], + "spans": [ + { + "bbox": [ + 317, + 463, + 554, + 507 + ], + "type": "text", + "content": "[22] Ishan Misra, Abhinav Shrivastava, Abhinav Gupta, and Martial Hebert. Cross-stitch networks for multi-task learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3994-4003, 2016. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 509, + 554, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 509, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 317, + 509, + 554, + 552 + ], + "type": "text", + "content": "[23] Aviv Navon, Aviv Shamsian, Idan Achituve, Haggai Maron, Kenji Kawaguchi, Gal Chechik, and Ethan Fetaya. Multitask learning as a bargaining game. arXiv preprint arXiv:2202.01017, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 554, + 609 + ], + "type": "text", + "content": "[24] Sara Shoouri, Mingyu Yang, Zichen Fan, and Hun-Seok Kim. Efficient computation sharing for multi-task visual scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17130-17141, 2023. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 612, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 612, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 612, + 554, + 678 + ], + "type": "text", + "content": "[25] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In Computer Vision-ECCV 2012: 12th European Conference on Computer Vision, Florence, Italy, October 7-13, 2012, Proceedings, Part V 12, pages 746-760. Springer, 2012. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[26] Dimitrios Sinodinos and Narges Armanfard. Ema-net: Efficient multitask affinity learning for dense scene predictions. arXiv preprint arXiv:2401.11124, 2024. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "18823" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[27] Jimmy T. H. Smith, Andrew Warrington, and Scott W. Linderman. Simplified state space layers for sequence modeling. CoRR, abs/2208.04933, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[28] Guolei Sun, Thomas Probst, Danda Pani Paudel, Nikola Popovic, Menelaos Kanakis, Jagruti Patel, Dengxin Dai, and Luc Van Gool. Task switching network for multi-task learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8291-8300, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "text", + "content": "[29] Simon Vandenhende, Stamatios Georgoulis, and Luc Van Gool. Mti-net: Multi-scale task interaction networks for multi-task learning. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 527-543. Springer, 2020. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[30] Simon Vandenhende, Stamatios Georgoulis, Wouter Van Gansbeke, Marc Proesmans, Dengxin Dai, and Luc Van Gool. Multi-task learning for dense prediction tasks: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(7):3614-3633, 2021. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[31] Dan Xu, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Pad-net: Multi-tasks guided prediction-and-distillation network for simultaneous depth estimation and scene parsing. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 675-684, 2018. 1, 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[32] Yangyang Xu, Xiangtai Li, Haobo Yuan, Yibo Yang, and Lefei Zhang. Multi-task learning with multi-query transformer for dense prediction. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 431 + ], + "type": "text", + "content": "[33] Chenhongyi Yang, Zehui Chen, Miguel Espinosa, Linus Ericsson, Zhenyu Wang, Jiaming Liu, and Elliot J. Crowley. Plainmamba: Improving non-hierarchical mamba in visual recognition. CoRR, abs/2403.17695, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 477 + ], + "type": "text", + "content": "[34] Feiyang Ye, Yueming Lyu, Xuehao Wang, Yu Zhang, and Ivor Tsang. Adaptive stochastic gradient algorithm for black-box multi-objective learning. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[35] Hanrong Ye and Dan Xu. Inverted pyramid multi-task transformer for dense scene understanding. In European Conference on Computer Vision, pages 514-530. Springer, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[36] Hanrong Ye and Dan Xu. Taskprompter: Spatial-channel multi-task prompting for dense scene understanding. In The Eleventh International Conference on Learning Representations, 2023. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[37] Hanrong Ye and Dan Xu. Invpt++: Inverted pyramid multitask transformer for visual scene understanding. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[38] Tianhe Yu, Saurabh Kumar, Abhishek Gupta, Sergey Levine, Karol Hausman, and Chelsea Finn. Gradient surgery for multi-task learning. Advances in Neural Information Processing Systems, 33:5824-5836, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[39] Jingdong Zhang, Jiayuan Fan, Peng Ye, Bo Zhang, Hancheng Ye, Baopu Li, Yancheng Cai, and Tao Chen. Rethinking of feature interaction for multi-task learning on dense prediction. arXiv preprint arXiv:2312.13514, 2023. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 274 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "text", + "content": "[40] Zhenyu Zhang, Zhen Cui, Chunyan Xu, Yan Yan, Nicu Sebe, and Jian Yang. Pattern-affinitive propagation across depth, surface normal and semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4106-4115, 2019. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 172 + ], + "type": "text", + "content": "[41] Sijie Zhao, Hao Chen, Xueliang Zhang, Pengfeng Xiao, Lei Bai, and Wanli Ouyang. Rs-mamba for large remote sensing image dense prediction. arXiv preprint arXiv:2404.02668, 2024.3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 174, + 553, + 228 + ], + "type": "text", + "content": "[42] Ling Zhou, Zhen Cui, Chunyan Xu, Zhenyu Zhang, Chaoqun Wang, Tong Zhang, and Jian Yang. Pattern-structure diffusion for multi-task learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4514–4523, 2020. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 315, + 230, + 553, + 274 + ], + "type": "text", + "content": "[43] Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, and Xinggang Wang. Vision mamba: Efficient visual representation learning with bidirectional state space model. In ICML, 2024. 2, 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "18824" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_content_list.json b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98cfeb286c3561f7e78f10b971b2c45edcb7eb70 --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_content_list.json @@ -0,0 +1,1491 @@ +[ + { + "type": "text", + "text": "Enhancing Numerical Prediction of MLLMs with Soft Labeling", + "text_level": 1, + "bbox": [ + 179, + 130, + 818, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pei Wang, Zhaowei Cai, Hao Yang, Davide Modolo, Ashwin Swaminathan \nAmazon AGI", + "bbox": [ + 202, + 180, + 794, + 218 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{pwwng,zhaoweic,haoyng,dmodolo,swashwin}@amazon.com", + "bbox": [ + 267, + 222, + 718, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 272, + 326, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The optimality of using the de facto cross-entropy loss with one-hot target distribution (hard labeling) is questioned when training (Multimodal) Large Language Models (LLMs/MLLMs). Although it is reasonable for language token prediction, which is a typical multi-class classification problem in discrete space, it is suboptimal for task like numerical prediction, which is a typical regression problem in continuous space. However, enabling regression in LLMs/MLLMs will complicate the training and next-token prediction paradigm at inference. Instead, to address this challenge, we propose a novel loss design, called soft labeling, which smooths the target probability distribution, enabling predictions to be penalized according to their distance to the target. This is similar to regression loss, which penalizes more on the further predictions in the continuous space, but will not change the model architecture and the next-token prediction paradigm of LLMs/MLLMs. We demonstrate the efficacy of soft labeling through extensive experiments on visual grounding, object counting, and chart understanding, achieving state-of-the-art performance on multiple benchmarks without bells and whistles. Soft labeling can be applied in any LLM/MLLM.", + "bbox": [ + 89, + 303, + 483, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 662, + 220, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, thanks to the success of Large Language Models (LLMs) [7, 52, 63], Multimodal Large Language Models (MLLMs) [2, 4, 31, 60] have achieved remarkable progresses across various tasks, such as image captioning, visual question answering, and many specialized domains like document analysis, chart interpretation, and visual grounding, etc. Their learning is to predict the next token, associated with the de facto cross-entropy loss across the token vocabulary, no matter the target is a word, digit, punctuation, or anything else. Although this unified mechanism simplifies the learning across various tasks and it becomes dominant, a natural question to ask is: is it optimal for any task? For example, the multi-class classification paradigm is reasonable for language word token predic", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e4191c460f4eab8707ef47d6a9b5c18dccd3e9d409a679b368e5c2cc30f92d06.jpg", + "image_caption": [ + "NNNN" + ], + "image_footnote": [ + "the prediction and target is considered." + ], + "bbox": [ + 519, + 276, + 612, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16c86ddf6c8b54bf4dd516b49f34b9e499c6fee6aa262857d560a10f61817c9f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 276, + 759, + 349 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0390ecc79235a4d0f700bea63ed5322eabf79df8b971a130149c4a122f6a7179.jpg", + "image_caption": [ + "Loss $\\mathrm{B} = 1.20$", + "Loss B=1.47" + ], + "image_footnote": [], + "bbox": [ + 790, + 276, + 893, + 351 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4b1075f62dc75a346207e86e800fed943b4ab01b483762a53988185adea3cfb3.jpg", + "image_caption": [ + "Loss $A = 1.20$" + ], + "image_footnote": [], + "bbox": [ + 526, + 354, + 633, + 416 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7ebf6f4fd75f87b3dfdd6d61b244e37897d7620f2a38adea9b404789795ec611.jpg", + "image_caption": [ + "Loss $A = 1.73$" + ], + "image_footnote": [], + "bbox": [ + 526, + 428, + 633, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Figure 1. Given an image with 4 trains, the MLLMs generate probabilistic predictions (A and B) for the count. In standard one-hot hard labeling, only the ground-truth numerical token has probability of 1, leading to equal loss values (1.20) for those two incorrect predictions (1 and 3 trains). In contrast, soft labeling assigns probabilities to numerical tokens according to their distances to the target. This results in different loss values (1.73 for prediction A of 1 train but 1.47 for prediction B of 3 trains), enabling higher penalty for further prediction.", + "bbox": [ + 511, + 489, + 906, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion because language word is in discrete space, but it is no longer optimal for objects that are in continuous space. In this paper, we are investigating the limits of standard cross-entropy loss in an important case, i.e., numerical prediction, which is in continuous space.", + "bbox": [ + 511, + 622, + 906, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Numerical prediction is a critical capability for LLMs and MLLMs, with many applications in object counting, object localization, chart/figure understanding, etc. For instance, answering the gross income from a tax report, or answering how many trains in a traffic scene as in Figure 1. Although the gross income value (e.g., 35421.7) and the number of trains (e.g., 4) are in continuous space, they are converted to strings (e.g., \"35421.7\" and \"4\") and then to discrete tokens to be processed by LLMs/MLLMs as a multi-class classification problem with cross-entropy loss. Cross-entropy loss measures the distance between the target and prediction distributions over the token vocabulary, and the target distribution is usually a one-hot vector where", + "bbox": [ + 511, + 703, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3424", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the probability on the target is 1 and the rest are 0. In this case, the wrong prediction to any non-target index can potentially have no difference. For example, in Figure 1, where there are 4 trains in a given image, predictions A and B have exactly the same cross-entropy loss because their probabilities on the target (i.e., 4) are the same (i.e., 0.3), although their overall distributions are quite different from each other. However, this is counter-intuitive. Prediction B of 3 trains should have lower loss than prediction A of 1 train, because the former is closer to the target. The cause of this issue is rooted in the one-hot representation of the target probability distribution, namely hard labeling, of the standard cross-entropy loss.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this problem, a new loss design is required, which can better reflect how wrong the prediction is for objects in continuous space. An ideal choice would be regression-style loss, e.g., L1 or L2 loss. However, asking LLMs/MLLMs to output scalar values is nontrivial without extra layers or heuristics. In addition, this will break the loss/architecture unification of LLMs/MLLMs and make the next-token prediction paradigm very complicated. Instead, we propose soft labeling for LLM/MLLM training, in contrast with hard labeling of the standard cross-entropy loss. In soft labeling, the target probability is no longer a one-hot probability distribution (Figure 1 middle), but smoothed where closer (further) classes have higher (lower) probabilities (Figure 1 bottom). Different from hard labeling, where only the target probability contributes to the loss, soft labeling enables losses coming from nearby classes, and considers the relative distance between the predicted and target tokens, similar to the regression loss. For example, in Figure 1, prediction A and B have the same loss using hard labeling, but the loss of B is lower than that of A for soft labeling. This is reasonable because prediction B of 3 trains is closer to the target of 4. In addition, the design of soft labeling will not modify the model architecture and next-token generation paradigm, which can be used in any LLM/MLLM.", + "bbox": [ + 91, + 291, + 482, + 666 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In our experiments, we demonstrate the effectiveness of the proposed soft labeling over hard labeling across three representative tasks of numerical prediction: visual grounding, object counting, and chart understanding on multiple benchmarks. Soft labeling consistently improves performance on all baselines. We also conduct comprehensive ablation studies on each component of soft labeling to gain deeper understanding of their impacts. By applying soft labeling, we achieve state-of-the-art or competitive performance on the RefCOCO+/g [25], ChartQA [40], and TallyQA [1] datasets, using only a standard LLaVA model [32] without any additional architectural modifications. This highlights the effectiveness of soft labeling in enhancing numerical prediction in MLLMs.", + "bbox": [ + 89, + 671, + 482, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Overall, this work makes three major contributions.", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "First, we question the optimality of using the de facto cross-entropy loss with one-hot target encoding (hard labeling) in any task of LLMs/MLLMs, and explain why standard cross-entropy loss leads to counter-intuitive loss assignment for a critical task, i.e., numerical prediction, in continuous space by nature. Second, we introduce soft labeling, where target probability distributions are smoothed such that closer (further) tokens to the target receive higher (lower) probabilities. This design better reflects the degree of error in predictions while maintaining compatibility with the standard next-token prediction paradigm of LLMs/MLLMs. Third, extensive experiments on visual grounding, object counting, and chart understanding demonstrate that soft labeling consistently outperforms hard labeling. The proposed approach achieves state-of-the-art or competitive results on benchmarks like RefCOCO, ChartQA, and TallyQA, without any architectural modification.", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 361, + 653, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Numerical prediction is a core challenge across a wide range of tasks in MLLMs and LLMs, spanning visual grounding [12, 48, 70], object counting [1], chart understanding [40, 43], visual question answering [18, 56], mathematical reasoning [38, 67], temporal reasoning [14, 16, 51] and document analysis [42, 59]. In visual grounding, models must quantify spatial location of objects with numerical bounding box coordinates [12, 48, 70]. In object counting, the model is required to predict precise numerical numbers of instances [1, 53]. Chart understanding relies on extracting and reasoning over numerical data in visual form [40, 43]. General VQA often involves answering questions that require numerical reasoning, such as counting, comparing quantities, or understanding scales [18, 56]. Mathematical reasoning, including arithmetic and algebraic problem-solving, tests models' ability to manipulate numerical values symbolically [38, 67]. Temporal reasoning in multimodal setups involves understanding and predicting time intervals or event sequences from textual and visual data [14, 16, 51]. Document analysis further highlights numerical prediction challenges, where models must extract and interpret numerical information from structured and unstructured data and make reasoning [42, 59]. These diverse applications underscore the pervasive role of numerical prediction in MLLMs and LLMs, making it a critical area for improvement.", + "bbox": [ + 511, + 387, + 903, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Ordinal regression (also called ordinal classification) is the problems where the target variable exhibits a relative ordering on an arbitrary scale, e.g. categories such as \"bad\", \"good\" and \"very good\". It has been proven useful in different research areas, including medical research [5, 35], monocular depth estimation [17], age estimation [46], pose estimation [21], etc. Ordinal regression has been well studies in conventional machine learn", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3425", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ing methods [19]. In the scenarios of deep learning, there are some typical methods. Class distance weighted loss relies on a coefficient for the loss of each class, which utilizes the distance to the ground-truth class and increases in relation to that distance [49]. The Earth Mover's Distance (EMD) loss uses the predicted probabilities of all classes and penalizes the miss-predictions according to a ground distance matrix that quantifies the dissimilarities between classes [9, 20]. Soft labeling converts ground truth data labels into soft probability distributions and the loss is computed on them [6, 15, 36, 65]. However, these methods have only shown success on simple multi category classification problem. For LLMs, the probability space of the output has tens to hundreds of thousands of tokens and they are mixture of numerical and non-numerical. The problem becomes difficulty. In this work, we will show how to apply ordinal regression to LLMs.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Soft labeling uses probability distributions over classes as labels instead of hard, one-hot labels. A typical type is label smoothing which has been found more robust to noisy annotations for classification problem [39, 61, 71] and has been applied to many fields such as know distillation [44, 73], network calibration [30], and some applications like image segmentation [66], text mining [34]. On the other hand, soft loss terms have been useful for domain and task transfer [64] in order to avoid dataset biases. Elaborate loss functions are defined in [68] to take into account the subjective scenicness of outdoor pictures, by trying to predict the same rating distribution of human annotations. Age estimation is another particular niche where soft labels have become popular [15, 58]. However, there is few research of applying soft labeling on numerical prediction of LLMs. In this paper, we will show that the label smoothing is just a specific format of our soft labeling and our proposal performs better.", + "bbox": [ + 89, + 347, + 482, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 618, + 181, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In traditional regression models, numerical prediction is typically handled using L1/L2-style loss functions, which penalize errors based on their distance to the target. Further distance incurs greater penalties and vice versa, encouraging the model to make predictions that are closer to the target. However, when training LLMs/MLLMs, the de facto cross-entropy loss with one-hot encoded target fails to account for the distance between predicted and target tokens properly. It does so in a limited and imprecise manner, a misclassification is a misclassification, regardless of how explicitly far the predicted token is from the target token. This limitation hinders the model's ability to predict accurate numerical values. To address this, a natural idea is to apply a regression-style loss to LLMs. However, in the context of LLMs, it is not trivial to directly convert the LLM's output into a scalar without introducing additional layers or heuristics. For this reason, we opt for a new loss de", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "sign which can be directly applied to the intrinsic output of LLM, similar to the standard cross-entropy loss.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by reviewing the LLM training loss. LLMs typically generate text token by token, treating the next-token prediction as a classification problem over a vocabulary of size $V$ . Given a model output logit $\\mathbf{z} \\in R^{V}$ from the last layer before the final prediction, the softmax function is applied to convert these logits into probability distribution $\\mathbf{p} \\in R^{V}$ such that $\\sum_{i=1}^{V} p_{i} = 1$ . This distribution is then compared to the target token using cross-entropy loss. Typically, the target token is encoded into a one-hot vector $\\mathbf{q} \\in R^{V}$ where it is a binary vector consisting of a value 1 at the correct token index, and 0s everywhere else. The cross-entropy loss is computed by", + "bbox": [ + 511, + 121, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {p}, \\mathbf {q}) = - \\sum_ {i = 1} ^ {V} q _ {i} \\log p _ {i}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 311, + 903, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since $\\mathbf{q}$ is a one-hot vector, the final loss simplifies to", + "bbox": [ + 511, + 363, + 867, + 378 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nL (\\mathbf {p}, \\mathbf {q}) = - \\log p _ {t}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 390, + 903, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t$ is the index of the target token. As shown, the loss function depends solely on the posterior probability $p_t$ of the target token. This ignores other rich information inherited in the full probability distribution $\\mathbf{p}$ .", + "bbox": [ + 511, + 417, + 903, + 477 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In fact, each probability $p_i$ reflects the model confidence on each possible token. When a prediction is incorrect, it is crucial to have a metric that precisely measures how \"wrong\" the prediction is by explicitly considering the model's confidence on all tokens, not just the target token. This becomes important when the target token is numerical and the numerical prediction is done by LLMs. Numerical tokens have relative distance between each other by nature, and when a prediction is incorrect, we should measure how wrong it is by assessing the distance from the target, rather than treating all wrong predictions equally as long as the confidence on the target token is the same, as is done in hard labeling. For example, in the scenario illustrated in Figure 1 to answer how many trains given an image, when we train an MLLM with cross-entropy loss and one-hot encoding, there are two possible incorrect predictions, A and B, during MLLM training. If we apply hard labeling to encode the supervision \"4\", both predictions incur the same cross-entropy loss, as both have a softmax probability of 0.3 for the target token \"4\". This approach fails to account for numerical proximity since prediction B, which assigns a maximum probability of 0.4 to the incorrect token \"3\", is closer to the correct answer \"4\" than prediction A, which assigns a maximum probability of 0.4 to the incorrect token \"1\". Intuitively, prediction B should receive a lower loss from the regression perspective. This limitation arises from the one-hot encoding in hard labeling which only retains the probability $p_t$ of the target token in the final loss computation", + "bbox": [ + 511, + 478, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3426", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "of (2). As a result, the final loss function does not consider the probabilities $p_i$ for other numerical tokens, which contain valuable information about inter-token relationship and numerical distance. The absence of this information makes the hard labeling loss suboptimal, and we need a new loss function that can better account for such relationships.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to address this issue, we propose soft labeling cross-entropy loss. We first reformulate the cross-entropy loss of (1) into a more general form,", + "bbox": [ + 89, + 183, + 480, + 228 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {C E} (\\mathbf {p}, \\mathbf {q} (t)) = \\sum_ {i = 1} ^ {V} q _ {i} (t) (- \\log p _ {i}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 241, + 482, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{q}(t)$ is the distribution in which the target token with index of $t$ is encoded, and $q_{i}(t) \\in [0,1]$ s.t. $\\sum_{i} q_{i}(t) = 1$ . In the standard one-hot encoding for cross-entropy, $\\mathbf{q}(t) = \\delta(t)$ is a Dirac delta distribution where $q_{i}(t)$ equals to 1 for $i = t$ , and 0 otherwise. In this case, (3) is simplified to (2), which corresponds to hard labeling. In fact, $\\mathbf{q}$ is not necessarily a one-hot distribution, but can be any distribution. The primary objective of $\\mathbf{q}$ is to guide the learning of the LLM model by minimizing the distance between two distributions. Instead of forcing the model to learn from a strict, hard label with harsh penalization, soft labeling relaxes this constraint by using a smooth distribution $\\mathbf{q}$ over all tokens.", + "bbox": [ + 89, + 295, + 482, + 476 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we focus on numerical prediction, where numerical values are typically tokenized digit-wise [26]. That is, each digit (0-9) is treated as an individual token, and multi-digit numbers are split accordingly. For example, in the LLaMA tokenizer, number \"123\" is tokenized into the token IDs \"29896\", \"29906\", and \"29941\". Therefore, we restrict our soft labeling discussion to the 10 digit tokens from \"0\" to \"9\", leaving the soft labeling of regular tokens for future work. For instance, we do not consider the soft labeling for text tokens such as \"two\", \"eight\", etc. Following [6, 65], we defines soft label for the 10 digit tokens from \"0\" to \"9\" as", + "bbox": [ + 89, + 477, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} ^ {S L} (t) = (1 - \\eta) \\boldsymbol {\\delta} (t) + \\eta \\psi (t) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 179, + 671, + 480, + 689 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $t$ is the index of the 10 digit tokens $\\{ \"0\", \"1\", ..., \"9\" \\}$ . For simplicity, we assume that these indices are consecutive and form a set $S$ ; if not, a straightforward mapping can be constructed to make them consecutive. $\\mathbf{q}^{SL}(t)$ is a soft distribution to represent the token with index of $t$ . $\\psi(t)$ is a probability distribution that has a support set on $S$ and typically centers at the token with index of $t$ . It can be Poisson, Binomial, Triangular, Uniform distributions, etc. $\\eta \\in [0,1]$ is a fixed hyperparameter that controls the mixture ratio of $\\delta(t)$ and $\\psi(t)$ .", + "bbox": [ + 89, + 703, + 482, + 868 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By smoothing the one-hot $\\mathbf{q}(t)$ of hard labeling, soft labeling allows $q_{i}^{SL}(t)$ to be greater than 0 for $i\\neq t$ and to", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "vary for different $i$ . This enables the assessment of relationship between the target token and other digit tokens. The relationship reflects the similarity between token indexed by $i$ and the token indexed by $t$ , with $q_{i}^{SL}(t)$ being higher when $i$ is closer to $t$ and lower otherwise.", + "bbox": [ + 511, + 90, + 903, + 165 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The final loss for each example is given by", + "bbox": [ + 532, + 167, + 816, + 181 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L = \\frac {1}{N _ {r} + N _ {n}} \\left(\\sum_ {\\left(x _ {r}, y _ {r}\\right) \\in \\left(\\mathbf {x} _ {r}, \\mathbf {y} _ {r}\\right)} L _ {C E} \\left(f \\left(x _ {r}\\right), \\mathbf {q} \\left(y _ {r}\\right)\\right) \\right. \\tag {5} \\\\ \\left. + \\lambda \\sum_ {\\left(x _ {n}, y _ {n}\\right) \\in \\left(\\mathbf {x} _ {n}, \\mathbf {y} _ {n}\\right)} L _ {C E} \\left(f \\left(x _ {n}\\right), \\mathbf {q} ^ {S L} \\left(y _ {n}\\right)\\right)\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 189, + 903, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{r}, N_{n}$ denote the numbers of regular tokens and numerical tokens within a sequence example, respectively. $\\mathbf{x}_r$ $(\\mathbf{x}_n)$ , $\\mathbf{y}_r$ $(\\mathbf{y}_n)$ represent the sets of regular (numerical) tokens where $\\mathbf{x}_r$ $(\\mathbf{x}_n)$ is the input set and $\\mathbf{y}_r$ $(\\mathbf{y}_n)$ is the index set of ground truth. $f(\\cdot)$ is the LLM with softmax output for each token prediction. $\\lambda$ controls the balance between the numerical token loss and the regular token loss in the final loss computation.", + "bbox": [ + 511, + 296, + 903, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the target is a digit token, compared to hard labeling (one-hot encoding) which solely rely on the probability of target token in the loss function, as indicated in (2), soft labeling incorporates the prediction probabilities of all incorrect digit tokens into loss computation. This allows the loss value to reflect the distance between the predicted token and target token. It therefore can be understood as an implicit way of introducing distance-based penalties into the classification loss, similar to how loss functions operate in regression. As illustrated in Figure 1, an incorrect prediction of \"1\" (Prediction A) has a greater distributional distance from the soft ground truth than an incorrect prediction of \"3\" (Prediction B). Consequently, Prediction A incurs a larger loss than Prediction B, aligning with the behavior of regression losses. This property arises naturally from the inter-digit token similarity encoded in soft labels where $q_{i}^{SL}(t)$ reflects the similarity ranking to the ground truth token. The closer a digit token is to the ground truth, the higher its value, enabling the model to account for the severity of prediction errors rather than treating all mistakes equally. In contrast, under hard labeling, predictions A and B receive the same loss due to the underlying assumption of one-hot encoding that all incorrect tokens are equally dissimilar to the ground truth.", + "bbox": [ + 511, + 417, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "From another perspective, in (3), $q_{i}(t)$ can be viewed as a weight for each component loss $-\\log p_{i}$ . During training, a soft label with $q_{i}^{SL}(t) > 0$ encourages the model to predict higher probabilities on tokens closer to the target token, and lower probabilities on further tokens. This also aligns with the goal of regression-style loss, which aims to make predictions as close as possible to the target in continuous space.", + "bbox": [ + 511, + 780, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3427", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/4992bd7474f23a003be48683b3a76729d58f3af7197753e4dacd547376a16694.jpg", + "table_caption": [ + "Table 1. The enhancement by soft labeling on visual grounding." + ], + "table_footnote": [], + "table_body": "
Pretrain modelModel typeFinetune dataLossRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
LLaVA-7BspecialistRefCOCO+/ghard soft81.787.374.473.682.063.379.078.8
85.0(+3.3)89.3(+2.0)78.8(+4.4)76.7(+3.1)83.4(+1.4)67.0(+3.7)82.0(+3.0)81.8(+3.0)
generalistLLaVA-Mix + RefCOCO+/ghard soft72.680.262.666.676.155.072.166.6
77.8(+5.2)83.7(+3.5)68.8(+6.2)72.0(+5.4)80.2(+4.1)60.4(+5.4)78.5(+6.4)71.5(+4.9)
LLaVA-13BspecialistRefCOCO+/ghard soft88.991.784.982.588.175.186.086.7
90.0(+1.1)92.7(+1.0)85.3(+0.4)83.5(+1.0)89.3(+1.2)76.7(+1.6)86.7(+0.7)87.4(+0.7)
generalistLLaVA-Mix + RefCOCO+/ghard soft79.385.071.474.782.964.878.073.0
81.0(+1.7)86.4(+1.4)72.6(+1.2)76.4(+1.7)83.3(+0.4)66.1(+1.3)79.2(+1.2)74.3(+1.3)
", + "bbox": [ + 94, + 113, + 903, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/afe2d20f7150c3212449e505288003a4d15a0f5e840b621be6453155900b719f.jpg", + "image_caption": [ + "(a) RefCOCOg Acc. v.s. $\\eta$ for Binomial distrib." + ], + "image_footnote": [], + "bbox": [ + 101, + 262, + 292, + 364 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb87aa59e99909317c5b29be7962ef69cf25869b65d34c15ee15164bce7775c4.jpg", + "image_caption": [ + "(b) RefCOCOg Acc. v.s. $\\eta$ for Poisson distrib." + ], + "image_footnote": [], + "bbox": [ + 303, + 262, + 493, + 366 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7b9876d45f43a5e58b733476308b46a7e43d137c3a2d67c8845335501698dcfe.jpg", + "image_caption": [ + "(c) RefCOCOg Acc. v.s. $\\eta$ for Triangular distrib." + ], + "image_footnote": [], + "bbox": [ + 503, + 262, + 694, + 366 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5c846cfc78de4dee874d63a472b664aad9b8a62e6cc59bf870e9cb6b0c71cafc.jpg", + "image_caption": [ + "(d) RefCOCOg Acc. v.s. $\\eta$ for Uniform distrib." + ], + "image_footnote": [], + "bbox": [ + 704, + 262, + 893, + 366 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/acad06ac982ac8545e94a6d063031aef1fee360b170dfb0e9513b11ab4284383.jpg", + "image_caption": [ + "(e) ChartQA Acc. v.s. $\\eta$ for Binomial distrib." + ], + "image_footnote": [], + "bbox": [ + 101, + 398, + 292, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5216aa1e2cefcdd523677e4c45a6cef60511a57dcaa16474207bf498b98a5c68.jpg", + "image_caption": [ + "(f) ChartQA Acc. v.s. $\\eta$ for Poisson distrib." + ], + "image_footnote": [], + "bbox": [ + 303, + 398, + 493, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9b99053548e53eb56fd22f72148891690c91c5c8b9bdcac754e9241f1686f5ee.jpg", + "image_caption": [ + "(g) ChartQA Acc. v.s. $\\eta$ for Triangular distrib." + ], + "image_footnote": [], + "bbox": [ + 504, + 398, + 694, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d12edcd12dc0b3f9fbd1eacbf6b920a3a45f3f9ffc13fccc873ccb8f3ac91032.jpg", + "image_caption": [ + "(h) ChartQA Acc. v.s. $\\eta$ for Uniform distrib." + ], + "image_footnote": [], + "bbox": [ + 704, + 398, + 893, + 501 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/cb29fc0a698c3888e7fccfc687f61548ee71b517506774f809f5a04f38657d4c.jpg", + "image_caption": [ + "(i) TallyQA RMSE v.s. $\\eta$ for Binomial distrib." + ], + "image_footnote": [], + "bbox": [ + 101, + 532, + 292, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f2a48044bde08e09ff808b886ed7e7ac00b34adc927a2c676e69c4b9cca44b19.jpg", + "image_caption": [ + "(j) TallyQA RMSE v.s. $\\eta$ for Poisson distrib." + ], + "image_footnote": [], + "bbox": [ + 303, + 532, + 493, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9c1ee120d8c086ea8b26c24295a6e246ac04227f8c6a191bac78bf9a00b070ef.jpg", + "image_caption": [ + "Figure 2. Ablation study of $\\eta$ of different distributions across three datasets, RefCOCOg val, ChartQA and TallyQA. On ChartQA, we evaluate on numerical subset. On TallyQA, we use the RMSE as metric and evaluate on simple balanced set." + ], + "image_footnote": [], + "bbox": [ + 504, + 532, + 694, + 636 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f33decacc9922b9eb88176a7ad4354eae961a544a9ddcd887c4c8cd3903f633f.jpg", + "image_caption": [ + "(k) TallyQA RMSE v.s. $\\eta$ for Triangular distrib.", + "(1) TallyQA RMSE v.s. $\\eta$ for Uniform distrib." + ], + "image_footnote": [], + "bbox": [ + 704, + 532, + 893, + 636 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/9b4a5962fef2ebbc4f996bd5ee4ca845b86088521aa3e67bae1c52d6ab82ae6f.jpg", + "table_caption": [ + "Table 2. The enhancement by soft labeling on chart understanding." + ], + "table_footnote": [], + "table_body": "
Pretrain modelModel typeFinetune dataLossChartQA
AllNumerical
LLaVA-7BspecialistChartQAhard soft67.966.3
69.6(+1.7)68.6(+2.3)
generalistLLaVA-Mix + ChartQAhard soft67.066.5
68.3(+1.3)68.6(+2.1)
LLaVA-13BspecialistChartQAhard soft72.872.0
73.4(+0.6)73.6(+1.6)
generalistLLaVA-Mix + ChartQAhard soft71.069.7
72.9 (+1.9)72.6(+2.9)
", + "bbox": [ + 91, + 724, + 491, + 857 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 700, + 643, + 717 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this work, we evaluate our approach on three representative numerical prediction tasks: visual grounding, object counting, and chart understanding. For visual grounding, we assess performance on three benchmarks RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), and RefCOCOg (val/test) [25] using the standard Acc@0.5 metric. For object counting, we evaluate on TallyQA's simple and complex test sets [1]. However, the original test sets are count-imbalanced, with ground truth counts ranging from 0 to 15 but the combined counts of 1 and 2 account for $78\\%$ of the simple testing set and $60\\%$ of the complex testing", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3428", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/f96e4c68b8fd4b17723e9096cc5efaecac8863f42e1fbfeeae8459f6efbfb22e.jpg", + "table_caption": [ + "Table 3. The enhancement by soft labeling on object counting." + ], + "table_footnote": [], + "table_body": "
Pretrain modelModel typeFinetune dataLossTallyQA simpleTallyQA complexTallyQA simple balancedTallyQA complex balanced
Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)
LLaVA-7BspecialistTallyQAhard78.51.0066.91.3038.53.8416.76.68
soft79.0(+0.5)0.97(-0.03)67.2(+0.3)1.29(-0.01)39.4(+0.9)3.20(-0.64)18.8(+2.1)6.32(-0.36)
generalistLLaVA-Mix + TallyQAhard79.81.4368.22.5439.42.3321.84.96
soft80.5(+0.7)1.35(-0.08)69.4(+1.2)1.28(-1.26)42.2(+2.8)2.24(-0.09)24.0(+2.2)4.41(-0.55)
LLaVA-13BspecialistTallyQAhard81.90.8567.21.2740.42.8422.95.98
soft82.7(+0.8)0.67(-0.18)69.3(+2.1)1.20(-0.07)42.2(+1.8)2.02(-0.82)27.0(+4.1)5.16(-0.82)
generalistLLaVA-Mix + TallyQAhard82.00.8170.91.3045.92.3723.94.87
soft82.5(+0.5)0.76(-0.05)71.5(+0.6)1.24(-0.06)47.8(+1.9)1.88(-0.49)29.1(+5.2)4.45(-0.42)
", + "bbox": [ + 91, + 112, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c51225a2da420922bd021d28afa931d794101c8bfb16079bcc5f66a94e189b3.jpg", + "image_caption": [ + "(a) RefCOCOg val" + ], + "image_footnote": [], + "bbox": [ + 96, + 263, + 339, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/eaa663f1a38b69465a29722f887c7e07dbd464a9ce8b0da6bee646e0c690ba95.jpg", + "image_caption": [ + "(b) ChartQA numerical subset" + ], + "image_footnote": [], + "bbox": [ + 375, + 262, + 620, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/40ba5169c74e5d603d0299885b79fc7dcb7a0db8afccf15630881fdcb9f6c837.jpg", + "image_caption": [ + "(c) TallyQA simple balanced", + "Figure 3. Ablation study of $\\lambda$ on three numerical prediction datasets." + ], + "image_footnote": [], + "bbox": [ + 660, + 263, + 901, + 392 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/09980095aa9fa6d9112082c63a04e899f876e32e813f98056e01e01d4681b968.jpg", + "table_caption": [ + "Table 4. The effect of different distributions. On TallyQA, RMSE is the metric." + ], + "table_footnote": [], + "table_body": "
DistributionRefCOCO valTallyQA balancedChartQA Numerical
**+*gSimpleComplex
Hard Label81.773.679.03.846.6866.3
Binomial84.776.881.93.136.2867.6
Posisson84.176.380.83.706.4967.0
Triangular85.076.782.03.206.3268.6
Uniform83.076.680.83.746.5666.8
", + "bbox": [ + 107, + 468, + 468, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "set. In addition, the testing sets also contain noisy labels. These make the evaluation results misleading. To address this, we also construct a balanced subset for reference by sampling a count-balanced simple and complex testing sets from TallyQA's test sets respectively, referred to as \"TallyQA simple balanced\" and \"TallyQA complex balanced\". The details are in the Supp. Accuracy and RMSE are the chosen metrics following [1]. For chart understanding, we use ChartQA benchmark [40], where approximately half of the questions have numerical answers. We report accuracy on both the full test set and a subset containing only numerical answers.", + "bbox": [ + 88, + 594, + 482, + 775 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following [4], we adopt a three-stage pipeline, where the first stage is to learn the vision-language (VL) adapter, the second stage is to pretrain LLM and VL adapter on a large set of pretraining data, and the third stage is to finetune the pretrained model on some small scale finetuning data. To evaluate the broad applicability of our approach, we conduct experiments on various pretraining and finetuning settings. For the pretraining of generalist models,", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "we use the publicly available LLaVA-7B architecture [32], pretrained on a sampled LAION-5M dataset [54], as our base model. For a stronger baseline, we use LLaVA-13B, pretrained on a 16M data mixture including LAION-5M for image captioning, GRIT-5M [47] for grounding, Objects365 [55] and OpenImages [27] of 3M in total for object counting, and 3M plot or chart related datasets such as PlotQA [43], UniChart [41], FigureQA [24] for chart understanding. For specialist model pretraining, we use three domain-specific data sources: a 12M mixture of grounding and detection data for visual grounding, a 6M counting data mixture with multiple tasks for object counting, and a 20M chart analysis data mixture for chart understanding, respectively. More details for pretraining data are given in Supp. After pretraining, the specialist models are finetuned on the task-specific training sets: RefCOCO, RefCOCO+, and RefCOCOg [25] for visual grounding; TallyQA [1] for object counting; and ChartQA [40] for chart understanding, and generalist models are finetuned on the same task-specific training sets but with the additional public 665K LLaVA-Mix tuning dataset [28]. The other settings are exactly the same as LLaVA's [32].", + "bbox": [ + 511, + 431, + 906, + 765 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. The Effectiveness of Soft Labeling", + "text_level": 1, + "bbox": [ + 511, + 772, + 812, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first compare soft labeling and one-hot hard labeling across various pretrained models, including LLaVA-7B (pretrained on LAION-5M) and LLaVA-13B (pretrained on a 16M multi-task mixture). They are fine-tuned as specialist and generalist models and evaluated on multiple tasks and benchmarks. Tables 1, 2 and 3 report the performance comparisons for visual grounding, chart understanding and", + "bbox": [ + 511, + 794, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3429", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/18c3056fa356b374cb5f156d38ecc83407d31bd5b3717607ba1193c42d4060ee.jpg", + "table_caption": [ + "Table 5. Performance comparison to the state-of-the-art models on visual grounding (Acc@0.5)." + ], + "table_footnote": [], + "table_body": "
ModelsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra-7B [12]87.090.680.281.687.472.182.382.2
Ferret-7B [70]87.591.482.580.887.473.183.984.8
Qwen-VL-7B [4]88.692.384.582.888.676.886.086.3
InternVL2-8B [13]87.191.180.779.887.971.482.782.7
VistaLLM [50]88.191.583.082.989.874.883.684.4
MiniGPT-v2 [11]88.791.785.380.085.174.584.484.7
Qwen2-VL-7B [4]91.793.687.385.890.579.587.387.8
Ferret-v2-7B [72]92.894.788.787.492.879.389.489.3
SoftLabeling-7B91.894.788.987.092.780.089.689.5
LION-12B [10]89.893.085.684.089.278.185.585.7
Shikra-13B [12]87.891.181.882.987.874.482.683.2
Ferret-13B [70]89.592.484.482.888.175.285.886.3
Ferret-v2-13B [72]92.695.088.987.492.181.489.490.0
SoftLabeling-13B92.795.089.087.692.782.389.890.0
", + "bbox": [ + 91, + 128, + 504, + 361 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "object counting tasks, respectively. Some observations are available. First, soft labeling consistently improves performance across all model and benchmark combinations, with most gains larger than 2 points, demonstrating its effectiveness and robustness. Second, the absolute gains are smaller for the stronger 13B model compared to the 7B model. This is expected due to performance saturation. Nonetheless, notable improvements are remained, e.g., a 2.9 points gain in accuracy on ChartQA numerical subset (generalist model) in Table 2 and 4.1/5.2 points accuracy gains on TallyQA's simple/complex balanced test sets in Table 3. Third, on ChartQA, the gain on numerical subset is higher than the full set. This aligns with the design of soft labeling which is proposed specifically for enhancing the numerical prediction ability. Finally, on object counting, gains are more pronounced on the more reasonable balanced test sets, particularly in RMSE, a preferable metric reflecting error distance to the ground truth, where the relative improvements exceed $10\\%$ in most cases. Notably, while enhancing numerical prediction, soft labeling does not degrade performance on generic tasks (please see details in Supp.)", + "bbox": [ + 89, + 378, + 485, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 710, + 243, + 726 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Soft Distribution The distribution $\\psi$ in (4) is crucial because it determines the degree of softness in soft labeling and the measurement of inter-digital token similarity. We investigate four widely used distributions, Binomial [6], Poisson [6], Triangular [65] and Uniform (label smoothing) [57]. The optimal $\\eta$ in (4) is exhaustively searched for each distribution and used for comparison. We assess performance using some representative benchmarks and metrics on the base LLaVA-7B specialist model, as reported in Table 4. Regardless of the chosen distribution, soft labeling consistently outperforms hard labeling, demonstrating", + "bbox": [ + 89, + 734, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88fe1b4422d6d100e5977c07aeca667f26d11a6729484c0d04da1676f104c103.jpg", + "table_caption": [ + "Table 6. Performance comparison to the state-of-the-art models on ChartQA." + ], + "table_footnote": [], + "table_body": "
ModelsChartQA
ChartPaLI-5B [8]77.3
ScreenAI 5B [3]76.7
MatCha4096 + LaMenDa [29]72.6
SMoLA-PaLI-X [69]74.6
Qwen-VL-Chat [4]66.3
Qwen-VL-Max [4]79.8
LLaVA-OV-7B [28]80.0
Cambrian-34B [62]73.8
Gemini Ultra [60]80.8
SoftLabeling81.5
", + "bbox": [ + 583, + 127, + 836, + 291 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/964173bed43c1607ed0d09d6a62fb25394a61653fc510420cb52465e7e8a6d8b.jpg", + "table_caption": [ + "Table 7. Performance comparison to the state-of-the-art models on TallyQA. * indicates that the numbers are copied from a third-party paper [37]." + ], + "table_footnote": [], + "table_body": "
ModelsTallyQA simple Acc.(↑) RMSE(↓)TallyQA complex Acc.(↑) RMSE(↓)
SMoLA-PaLI-X [69]86.3-77.1-
PaLI-X-VPD [22]86.2-76.6-
MoVie-ResNeXt [45]74.91.0056.81.43
RCN [1]71.81.1356.21.43
LLaVA-NEXT [33]79.8*0.70*67.9*1.76*
GPT-4V [2]73.6*0.86*62.6*1.58*
GPT-4o [23]81.5*0.60*71.7*1.21*
SoftLabeling86.60.5677.21.06
", + "bbox": [ + 532, + 358, + 887, + 507 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "its robustness. Among the distributions, Binomial and Triangular yield the best results, followed by Poisson, while Uniform performs the worst. This suggests that sharper distributions are more effective, as they provide more discriminative token similarities. In contrast, smoother distributions like Uniform treat all incorrect tokens as equally distant from the target, disregarding the relative distance information. Throughout this paper, we use Triangular distribution as default, though Binomial is comparable.", + "bbox": [ + 511, + 525, + 906, + 661 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Hyperparameter There are two primary hyperparameters influencing the performances, $\\eta$ of (4) and $\\lambda$ of (5). $\\eta$ controls the sharpness of the soft labeling. It is selected through exhaustive search based on the performance of the base LLaVA-7B specialist model on RefCOCOg val, ChartQA and TallyQA. For ChartQA, we evaluate accuracy on numerical subset. On TallyQA, we use RMSE as metric and evaluate on simple balanced set. Figure 2 shows the ablation study results. Our analysis shows that the model favors sharper soft labeling distributions with smaller $\\eta$ . However, if it is too small, the distribution is towards a hard one-hot encoding, leading to performance drop. On the other hand, a large $\\eta$ over smooths the distribution, reducing discrimination among numerical tokens. In our experiments, $\\eta$ is set to $0.05 / 0.02 / 0.08 / 0.05$ for Binomial/Poisson/Triangular/U", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3430", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7d950f7cd73796610b77251de0334a6a8fb628c4e36325fb1ae5a09667d58e69.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Image
Promptfirst sandwich on the left just beneath the forkthe man on the closest motorcycleWhat was the internet penetration rate in Africa in 2019?How many books did Blurb publish in 2018?
Hard labelingred boxred box2016000
Soft labelinggreen boxgreen box27.419053
Ground truthblue boxblue box28.619098
", + "bbox": [ + 91, + 88, + 908, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8. Qualitative examples of soft labeling improvement in visual grounding and chart understanding.", + "bbox": [ + 187, + 271, + 805, + 284 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "niform distribution, respectively, by default.", + "bbox": [ + 89, + 291, + 382, + 305 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As for $\\lambda$ of (5), it balances the contribution of numerical token loss and regular token loss in the final loss. We ablate its effect using the base 7B and stronger 13B specialist models across three datasets and present the results in Figure 3. It can be found that the models perform well and remain stable when $\\lambda$ ranges from 2 to 5, beyond which the performance degrades. In our experiments, we set $\\lambda$ as 2 by default.", + "bbox": [ + 88, + 306, + 483, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Comparison to the State-of-the-Art", + "text_level": 1, + "bbox": [ + 89, + 438, + 397, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate soft labeling and compare it with other methods across visual grounding, chart understanding, and object counting using public benchmarks, including RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), RefCOCOg (val/test), ChartQA (test), and TallyQA (simple/complex). The model is pretrained on scaled-up domain- and task-specific data and fine-tuned as specialist models (details in Section B of the Supp.). By default, our results are based on the 13B model, but for grounding, we follow standard practice and compare 7B and 13B models separately. As shown in Table 5, 6, and 7, our Soft-Labeling, without any architectural modifications, achieves the state-of-the-art performance on RefCOCOg among 7B scale models, RefCOCO, RefCOCO+, RefCOCOg among 13B scale models, ChartQA, and TallyQA. It also delivers competitive results on RefCOCO and RefCOCO+ in 7B scale model group. These results highlight the effectiveness of soft labeling.", + "bbox": [ + 91, + 459, + 483, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Visualization", + "text_level": 1, + "bbox": [ + 89, + 742, + 227, + 757 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 8 presents examples randomly selected from visual RefCOCOg and ChartQA test sets, demonstrating the improvement achieved by soft labeling. Compared to hard labeling, soft labeling produces more accurate numerical predictions for both bounding box coordinates and chart-based values, bringing them closer to the ground truth. For instance, in the motorcycle image, the bounding box predicted by soft labeling (green) better aligns with the ground truth (blue) in both position and scale. Similarly, in the chart-", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "based QA task, soft labeling provides more accurate numerical estimation. These examples highlight the effectiveness of soft labeling in improving numerical prediction across both spatial and quantitative tasks.", + "bbox": [ + 511, + 291, + 906, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation", + "text_level": 1, + "bbox": [ + 511, + 362, + 627, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we limit our scope to digit tokens. However, other tokens also exhibit semantic similarity, such as natural language numerals (\"one/two/three\") or categorical tokens (\"easy/medium/hard\"). Incorporating these into the loss computation is an avenue for future research. Another limitation is that our current loss function considers only single digits. Many numerical predictions involve multi-digit numbers, where digit-wise comparisons may not accurately reflect actual numerical distances. For instance, comparing \"21\" v.s. \"32\" digit-wise results in a distance of $|2 - 3| + |1 - 2| = 2$ which is smaller than $|2 - 1| + |2 - 9| = 8$ for \"21\" v.s. \"29\". However, their actual numerical distances are $|21 - 32| = 11$ and $|21 - 29| = 8$ , showing a discrepancy in digit-wise calculations. Addressing this limitation remains an open challenge.", + "bbox": [ + 511, + 388, + 906, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 619, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we have identified the limitations of cross-entropy loss with one-hot encoding for numerical prediction in LLMs/MLLMs, where treating numbers as discrete tokens fails to reflect their continuous nature. To address this, we have proposed soft labeling, which smooths target distributions to account for numerical proximity, ensuring that closer predictions incur lower loss. This method retains compatibility with the next-token prediction paradigm without requiring architectural modifications. Our experiments across visual grounding, object counting, and chart understanding have shown that soft labeling consistently improves numerical prediction performance. We have achieved state-of-the-art or competitive results on RefCOCO $+ / g$ , ChartQA, and TallyQA using a basic LLaVA model. Our findings highlight the importance of adapting appropriate loss functions for numerical tasks in LLMs/MLLMs.", + "bbox": [ + 511, + 643, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3431", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI, pages 8076-8084, 2019. 2, 5, 6, 7", + "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 7", + "[3] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 7", + "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arxiv 2023. arXiv preprint arXiv:2308.12966, 1(8), 2023. 1, 6, 7", + "[5] Javier Barbero-Gomez, Pedro-Antonio Gutierrez, VICTOR-Manuel Vargas, Juan-Antonio Vallejo-Casas, and Cesar Hervás-Martínez. An ordinal cnn approach for the assessment of neurological damage in parkinson's disease patients. Expert Systems with Applications, 182:115271, 2021. 2", + "[6] Christopher Beckham and Christopher Pal. Unimodal probability distributions for deep ordinal classification. In ICML, pages 411-419. PMLR, 2017. 3, 4, 7", + "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 1", + "[8] Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596, 2024. 7", + "[9] Alberto Castano, Pablo González, Jaime Alonso González, and Juan Jose Del Coz. Matching distributions algorithms based on the earth mover's distance for ordinal quantification. IEEE T-NNLS, 35(1):1050-1061, 2022. 3", + "[10] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In CVPR, pages 26540-26550, 2024. 7", + "[11] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7", + "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 7", + "[14] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Haotian Wang, Ming Liu, and Bing Qin. Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models. arXiv preprint arXiv:2311.17667, 2023. 2", + "[15] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In CVPR, pages 4738-4747, 2019. 3", + "[16] Bahare Fatemi, Mehran Kazemi, Anton Tsitsulin, Karishma Malkan, Jinyeong Yim, John Palowitch, Sungyong Seo, Jonathan Halcrow, and Bryan Perozzi. Test of time: A benchmark for evaluating llms on temporal reasoning. arXiv preprint arXiv:2406.09170, 2024. 2", + "[17] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 2", + "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 2", + "[19] Pedro Antonio Gutierrez, Maria Perez-Ortiz, Javier Sanchez-Monedero, Francisco Fernandez-Navarro, and Cesar Hervas-Martinez. Ordinal regression methods: survey and experimental study. IEEE Transactions on Knowledge and Data Engineering, 28(1):127-146, 2015. 3", + "[20] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared earth mover's distance-based loss for training deep neural networks. arXiv preprint arXiv:1611.05916, 2016. 3", + "[21] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035–1046, 2018. 2", + "[22] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 7", + "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7", + "[24] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 6", + "[25] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, pages 787-798, 2014. 2, 5, 6", + "[26] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3432", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6", + "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 7", + "[29] Zhuowan Li, Bhavan Jasani, Peng Tang, and Shabnam Ghadar. Synthesize step-by-step: Tools templates and llms as data generators for reasoning-based chart vqa. In CVPR, pages 13613-13623, 2024. 7", + "[30] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, pages 80-88, 2022. 3", + "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 1", + "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 2, 6", + "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.7", + "[34] Peiyang Liu, Xiangyu Xi, Wei Ye, and Shikun Zhang. Label smoothing for text mining. In ICCL, pages 2210-2219, 2022. 3", + "[35] Xiaofeng Liu, Yang Zou, Yuhang Song, Chao Yang, Jane You, and BV K Vijaya Kumar. Ordinal regression with neuron stick-breaking for medical diagnosis. In ECCV Workshops, pages 0-0, 2018. 2", + "[36] Xiaofeng Liu, Fangfang Fan, Lingsheng Kong, Zhihui Dao, Wanqing Xie, Jun Lu, and Jane You. Unimodal regularized neuron stick-breaking for ordinal classification. Neurocomputing, 388:34-44, 2020. 3", + "[37] Jian Lu, Shikhar Srivastava, Junyu Chen, Robik Shrestha, Manoj Acharya, Kushal Kafle, and Christopher Kanan. Revisiting multi-modal llm evaluation. arXiv preprint arXiv:2408.05334, 2024. 7", + "[38] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2", + "[39] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 3", + "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 6", + "[41] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 6" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 2", + "[43] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1527-1536, 2020. 2, 6", + "[44] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? NeurIPS, 32, 2019. 3", + "[45] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions for visual counting and beyond. arXiv preprint arXiv:2004.11883, 2020. 7", + "[46] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In CVPR, pages 4920-4928, 2016. 2", + "[47] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6", + "[48] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In ICLR, 2024. 2", + "[49] Gorkem Polat, Ilkay Ergenc, Haluk Tarik Kani, Yesim Ozen Alahdab, Ozlen Atug, and Alptekin Temizel. Class distance weighted cross-entropy loss for ulcerative colitis severity estimation. In Annual Conference on Medical Image Understanding and Analysis, pages 157–171. Springer, 2022. 3", + "[50] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In CVPR, pages 14076-14088, 2024. 7", + "[51] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2", + "[52] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1", + "[53] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 2", + "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 6", + "[55] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, pages 8430-8439, 2019. 6", + "[56] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 2" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3433", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7", + "[58] Zichang Tan, Shuai Zhou, Jun Wan, Zhen Lei, and Stan Z Li. Age estimation based on a single network with soft softmax of aging modeling. In ACCV, pages 203-216. Springer, 2016. 3", + "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 2", + "[60] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 7", + "[61] Christian Thiel. Classification on soft labels is robust against label noise. In International Conference on Knowledge-Based and Intelligent Information and Engineering Systems, pages 65-73. Springer, 2008. 3", + "[62] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 7", + "[63] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1", + "[64] Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In ICCV, pages 4068-4076, 2015. 3", + "[65] Víctor Manuel Vargas, Pedro Antonio Gutiérrez, Javier Barbero-Gómez, and César Hervás-Martínez. Soft labelling based on triangular distributions for ordinal classification. Information Fusion, 93:258-267, 2023. 3, 4, 7", + "[66] Sukesh Adiga Vasudeva, Jose Dolz, and Herve Lombaert. Geols: geodesic label smoothing for image segmentation. In Medical Imaging with Deep Learning, pages 468-478. PMLR, 2024. 3", + "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. NeurIPS, 37:95095-95169, 2024. 2", + "[68] Scott Workman, Richard Souvenir, and Nathan Jacobs. Understanding and mapping natural beauty. In ICCV, pages 5589-5598, 2017. 3", + "[69] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR, pages 14205-14215, 2024. 7", + "[70] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 7" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[71] Chang-Bin Zhang, Peng-Tao Jiang, Qibin Hou, Yunchao Wei, Qi Han, Zhen Li, and Ming-Ming Cheng. Delving deep into label smoothing. IEEE T-IP, 30:5984-5996, 2021. 3", + "[72] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 7", + "[73] Helong Zhou, Liangchen Song, Jiajie Chen, Ye Zhou, Guoli Wang, Junsong Yuan, and Qian Zhang. Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650, 2021. 3" + ], + "bbox": [ + 514, + 92, + 903, + 273 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3434", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_model.json b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_model.json new file mode 100644 index 0000000000000000000000000000000000000000..803d4ea8dec239e562c33acd5dbf1dec47ff7a36 --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.82, + 0.154 + ], + "angle": 0, + "content": "Enhancing Numerical Prediction of MLLMs with Soft Labeling" + }, + { + "type": "text", + "bbox": [ + 0.203, + 0.181, + 0.795, + 0.219 + ], + "angle": 0, + "content": "Pei Wang, Zhaowei Cai, Hao Yang, Davide Modolo, Ashwin Swaminathan \nAmazon AGI" + }, + { + "type": "text", + "bbox": [ + 0.269, + 0.223, + 0.72, + 0.238 + ], + "angle": 0, + "content": "{pwwng,zhaoweic,haoyng,dmodolo,swashwin}@amazon.com" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.273, + 0.327, + 0.289 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.304, + 0.485, + 0.637 + ], + "angle": 0, + "content": "The optimality of using the de facto cross-entropy loss with one-hot target distribution (hard labeling) is questioned when training (Multimodal) Large Language Models (LLMs/MLLMs). Although it is reasonable for language token prediction, which is a typical multi-class classification problem in discrete space, it is suboptimal for task like numerical prediction, which is a typical regression problem in continuous space. However, enabling regression in LLMs/MLLMs will complicate the training and next-token prediction paradigm at inference. Instead, to address this challenge, we propose a novel loss design, called soft labeling, which smooths the target probability distribution, enabling predictions to be penalized according to their distance to the target. This is similar to regression loss, which penalizes more on the further predictions in the continuous space, but will not change the model architecture and the next-token prediction paradigm of LLMs/MLLMs. We demonstrate the efficacy of soft labeling through extensive experiments on visual grounding, object counting, and chart understanding, achieving state-of-the-art performance on multiple benchmarks without bells and whistles. Soft labeling can be applied in any LLM/MLLM." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.664, + 0.222, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In recent years, thanks to the success of Large Language Models (LLMs) [7, 52, 63], Multimodal Large Language Models (MLLMs) [2, 4, 31, 60] have achieved remarkable progresses across various tasks, such as image captioning, visual question answering, and many specialized domains like document analysis, chart interpretation, and visual grounding, etc. Their learning is to predict the next token, associated with the de facto cross-entropy loss across the token vocabulary, no matter the target is a word, digit, punctuation, or anything else. Although this unified mechanism simplifies the learning across various tasks and it becomes dominant, a natural question to ask is: is it optimal for any task? For example, the multi-class classification paradigm is reasonable for language word token predic" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.277, + 0.614, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.626, + 0.277, + 0.644, + 0.349 + ], + "angle": 0, + "content": "NNNN" + }, + { + "type": "image", + "bbox": [ + 0.657, + 0.277, + 0.76, + 0.351 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.791, + 0.277, + 0.895, + 0.352 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.355, + 0.634, + 0.417 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.365, + 0.751, + 0.375 + ], + "angle": 0, + "content": "Loss \\(A = 1.20\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.365, + 0.879, + 0.375 + ], + "angle": 0, + "content": "Loss \\( \\mathrm{B} = 1.20 \\)" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.429, + 0.634, + 0.486 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.435, + 0.751, + 0.445 + ], + "angle": 0, + "content": "Loss \\(A = 1.73\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.435, + 0.879, + 0.444 + ], + "angle": 0, + "content": "Loss B=1.47" + }, + { + "type": "image_footnote", + "bbox": [ + 0.699, + 0.462, + 0.865, + 0.473 + ], + "angle": 0, + "content": "the prediction and target is considered." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Figure 1. Given an image with 4 trains, the MLLMs generate probabilistic predictions (A and B) for the count. In standard one-hot hard labeling, only the ground-truth numerical token has probability of 1, leading to equal loss values (1.20) for those two incorrect predictions (1 and 3 trains). In contrast, soft labeling assigns probabilities to numerical tokens according to their distances to the target. This results in different loss values (1.73 for prediction A of 1 train but 1.47 for prediction B of 3 trains), enabling higher penalty for further prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.699 + ], + "angle": 0, + "content": "tion because language word is in discrete space, but it is no longer optimal for objects that are in continuous space. In this paper, we are investigating the limits of standard cross-entropy loss in an important case, i.e., numerical prediction, which is in continuous space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Numerical prediction is a critical capability for LLMs and MLLMs, with many applications in object counting, object localization, chart/figure understanding, etc. For instance, answering the gross income from a tax report, or answering how many trains in a traffic scene as in Figure 1. Although the gross income value (e.g., 35421.7) and the number of trains (e.g., 4) are in continuous space, they are converted to strings (e.g., \"35421.7\" and \"4\") and then to discrete tokens to be processed by LLMs/MLLMs as a multi-class classification problem with cross-entropy loss. Cross-entropy loss measures the distance between the target and prediction distributions over the token vocabulary, and the target distribution is usually a one-hot vector where" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3424" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "the probability on the target is 1 and the rest are 0. In this case, the wrong prediction to any non-target index can potentially have no difference. For example, in Figure 1, where there are 4 trains in a given image, predictions A and B have exactly the same cross-entropy loss because their probabilities on the target (i.e., 4) are the same (i.e., 0.3), although their overall distributions are quite different from each other. However, this is counter-intuitive. Prediction B of 3 trains should have lower loss than prediction A of 1 train, because the former is closer to the target. The cause of this issue is rooted in the one-hot representation of the target probability distribution, namely hard labeling, of the standard cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.667 + ], + "angle": 0, + "content": "To address this problem, a new loss design is required, which can better reflect how wrong the prediction is for objects in continuous space. An ideal choice would be regression-style loss, e.g., L1 or L2 loss. However, asking LLMs/MLLMs to output scalar values is nontrivial without extra layers or heuristics. In addition, this will break the loss/architecture unification of LLMs/MLLMs and make the next-token prediction paradigm very complicated. Instead, we propose soft labeling for LLM/MLLM training, in contrast with hard labeling of the standard cross-entropy loss. In soft labeling, the target probability is no longer a one-hot probability distribution (Figure 1 middle), but smoothed where closer (further) classes have higher (lower) probabilities (Figure 1 bottom). Different from hard labeling, where only the target probability contributes to the loss, soft labeling enables losses coming from nearby classes, and considers the relative distance between the predicted and target tokens, similar to the regression loss. For example, in Figure 1, prediction A and B have the same loss using hard labeling, but the loss of B is lower than that of A for soft labeling. This is reasonable because prediction B of 3 trains is closer to the target of 4. In addition, the design of soft labeling will not modify the model architecture and next-token generation paradigm, which can be used in any LLM/MLLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.483, + 0.883 + ], + "angle": 0, + "content": "In our experiments, we demonstrate the effectiveness of the proposed soft labeling over hard labeling across three representative tasks of numerical prediction: visual grounding, object counting, and chart understanding on multiple benchmarks. Soft labeling consistently improves performance on all baselines. We also conduct comprehensive ablation studies on each component of soft labeling to gain deeper understanding of their impacts. By applying soft labeling, we achieve state-of-the-art or competitive performance on the RefCOCO+/g [25], ChartQA [40], and TallyQA [1] datasets, using only a standard LLaVA model [32] without any additional architectural modifications. This highlights the effectiveness of soft labeling in enhancing numerical prediction in MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Overall, this work makes three major contributions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "First, we question the optimality of using the de facto cross-entropy loss with one-hot target encoding (hard labeling) in any task of LLMs/MLLMs, and explain why standard cross-entropy loss leads to counter-intuitive loss assignment for a critical task, i.e., numerical prediction, in continuous space by nature. Second, we introduce soft labeling, where target probability distributions are smoothed such that closer (further) tokens to the target receive higher (lower) probabilities. This design better reflects the degree of error in predictions while maintaining compatibility with the standard next-token prediction paradigm of LLMs/MLLMs. Third, extensive experiments on visual grounding, object counting, and chart understanding demonstrate that soft labeling consistently outperforms hard labeling. The proposed approach achieves state-of-the-art or competitive results on benchmarks like RefCOCO, ChartQA, and TallyQA, without any architectural modification." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.362, + 0.655, + 0.379 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.388, + 0.905, + 0.78 + ], + "angle": 0, + "content": "Numerical prediction is a core challenge across a wide range of tasks in MLLMs and LLMs, spanning visual grounding [12, 48, 70], object counting [1], chart understanding [40, 43], visual question answering [18, 56], mathematical reasoning [38, 67], temporal reasoning [14, 16, 51] and document analysis [42, 59]. In visual grounding, models must quantify spatial location of objects with numerical bounding box coordinates [12, 48, 70]. In object counting, the model is required to predict precise numerical numbers of instances [1, 53]. Chart understanding relies on extracting and reasoning over numerical data in visual form [40, 43]. General VQA often involves answering questions that require numerical reasoning, such as counting, comparing quantities, or understanding scales [18, 56]. Mathematical reasoning, including arithmetic and algebraic problem-solving, tests models' ability to manipulate numerical values symbolically [38, 67]. Temporal reasoning in multimodal setups involves understanding and predicting time intervals or event sequences from textual and visual data [14, 16, 51]. Document analysis further highlights numerical prediction challenges, where models must extract and interpret numerical information from structured and unstructured data and make reasoning [42, 59]. These diverse applications underscore the pervasive role of numerical prediction in MLLMs and LLMs, making it a critical area for improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ordinal regression (also called ordinal classification) is the problems where the target variable exhibits a relative ordering on an arbitrary scale, e.g. categories such as \"bad\", \"good\" and \"very good\". It has been proven useful in different research areas, including medical research [5, 35], monocular depth estimation [17], age estimation [46], pose estimation [21], etc. Ordinal regression has been well studies in conventional machine learn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3425" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.348 + ], + "angle": 0, + "content": "ing methods [19]. In the scenarios of deep learning, there are some typical methods. Class distance weighted loss relies on a coefficient for the loss of each class, which utilizes the distance to the ground-truth class and increases in relation to that distance [49]. The Earth Mover's Distance (EMD) loss uses the predicted probabilities of all classes and penalizes the miss-predictions according to a ground distance matrix that quantifies the dissimilarities between classes [9, 20]. Soft labeling converts ground truth data labels into soft probability distributions and the loss is computed on them [6, 15, 36, 65]. However, these methods have only shown success on simple multi category classification problem. For LLMs, the probability space of the output has tens to hundreds of thousands of tokens and they are mixture of numerical and non-numerical. The problem becomes difficulty. In this work, we will show how to apply ordinal regression to LLMs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.483, + 0.606 + ], + "angle": 0, + "content": "Soft labeling uses probability distributions over classes as labels instead of hard, one-hot labels. A typical type is label smoothing which has been found more robust to noisy annotations for classification problem [39, 61, 71] and has been applied to many fields such as know distillation [44, 73], network calibration [30], and some applications like image segmentation [66], text mining [34]. On the other hand, soft loss terms have been useful for domain and task transfer [64] in order to avoid dataset biases. Elaborate loss functions are defined in [68] to take into account the subjective scenicness of outdoor pictures, by trying to predict the same rating distribution of human annotations. Age estimation is another particular niche where soft labels have become popular [15, 58]. However, there is few research of applying soft labeling on numerical prediction of LLMs. In this paper, we will show that the label smoothing is just a specific format of our soft labeling and our proposal performs better." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.182, + 0.635 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In traditional regression models, numerical prediction is typically handled using L1/L2-style loss functions, which penalize errors based on their distance to the target. Further distance incurs greater penalties and vice versa, encouraging the model to make predictions that are closer to the target. However, when training LLMs/MLLMs, the de facto cross-entropy loss with one-hot encoded target fails to account for the distance between predicted and target tokens properly. It does so in a limited and imprecise manner, a misclassification is a misclassification, regardless of how explicitly far the predicted token is from the target token. This limitation hinders the model's ability to predict accurate numerical values. To address this, a natural idea is to apply a regression-style loss to LLMs. However, in the context of LLMs, it is not trivial to directly convert the LLM's output into a scalar without introducing additional layers or heuristics. For this reason, we opt for a new loss de" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "sign which can be directly applied to the intrinsic output of LLM, similar to the standard cross-entropy loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.122, + 0.905, + 0.304 + ], + "angle": 0, + "content": "We begin by reviewing the LLM training loss. LLMs typically generate text token by token, treating the next-token prediction as a classification problem over a vocabulary of size \\( V \\). Given a model output logit \\( \\mathbf{z} \\in R^{V} \\) from the last layer before the final prediction, the softmax function is applied to convert these logits into probability distribution \\( \\mathbf{p} \\in R^{V} \\) such that \\( \\sum_{i=1}^{V} p_{i} = 1 \\). This distribution is then compared to the target token using cross-entropy loss. Typically, the target token is encoded into a one-hot vector \\( \\mathbf{q} \\in R^{V} \\) where it is a binary vector consisting of a value 1 at the correct token index, and 0s everywhere else. The cross-entropy loss is computed by" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.313, + 0.905, + 0.355 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {p}, \\mathbf {q}) = - \\sum_ {i = 1} ^ {V} q _ {i} \\log p _ {i}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.364, + 0.868, + 0.38 + ], + "angle": 0, + "content": "Since \\(\\mathbf{q}\\) is a one-hot vector, the final loss simplifies to" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.391, + 0.905, + 0.408 + ], + "angle": 0, + "content": "\\[\nL (\\mathbf {p}, \\mathbf {q}) = - \\log p _ {t}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.418, + 0.905, + 0.478 + ], + "angle": 0, + "content": "where \\( t \\) is the index of the target token. As shown, the loss function depends solely on the posterior probability \\( p_t \\) of the target token. This ignores other rich information inherited in the full probability distribution \\( \\mathbf{p} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.905, + 0.903 + ], + "angle": 0, + "content": "In fact, each probability \\( p_i \\) reflects the model confidence on each possible token. When a prediction is incorrect, it is crucial to have a metric that precisely measures how \"wrong\" the prediction is by explicitly considering the model's confidence on all tokens, not just the target token. This becomes important when the target token is numerical and the numerical prediction is done by LLMs. Numerical tokens have relative distance between each other by nature, and when a prediction is incorrect, we should measure how wrong it is by assessing the distance from the target, rather than treating all wrong predictions equally as long as the confidence on the target token is the same, as is done in hard labeling. For example, in the scenario illustrated in Figure 1 to answer how many trains given an image, when we train an MLLM with cross-entropy loss and one-hot encoding, there are two possible incorrect predictions, A and B, during MLLM training. If we apply hard labeling to encode the supervision \"4\", both predictions incur the same cross-entropy loss, as both have a softmax probability of 0.3 for the target token \"4\". This approach fails to account for numerical proximity since prediction B, which assigns a maximum probability of 0.4 to the incorrect token \"3\", is closer to the correct answer \"4\" than prediction A, which assigns a maximum probability of 0.4 to the incorrect token \"1\". Intuitively, prediction B should receive a lower loss from the regression perspective. This limitation arises from the one-hot encoding in hard labeling which only retains the probability \\( p_t \\) of the target token in the final loss computation" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "of (2). As a result, the final loss function does not consider the probabilities \\( p_i \\) for other numerical tokens, which contain valuable information about inter-token relationship and numerical distance. The absence of this information makes the hard labeling loss suboptimal, and we need a new loss function that can better account for such relationships." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.184, + 0.482, + 0.229 + ], + "angle": 0, + "content": "In order to address this issue, we propose soft labeling cross-entropy loss. We first reformulate the cross-entropy loss of (1) into a more general form," + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.242, + 0.483, + 0.282 + ], + "angle": 0, + "content": "\\[\nL _ {C E} (\\mathbf {p}, \\mathbf {q} (t)) = \\sum_ {i = 1} ^ {V} q _ {i} (t) (- \\log p _ {i}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.477 + ], + "angle": 0, + "content": "where \\(\\mathbf{q}(t)\\) is the distribution in which the target token with index of \\(t\\) is encoded, and \\(q_{i}(t) \\in [0,1]\\) s.t. \\(\\sum_{i} q_{i}(t) = 1\\). In the standard one-hot encoding for cross-entropy, \\(\\mathbf{q}(t) = \\delta(t)\\) is a Dirac delta distribution where \\(q_{i}(t)\\) equals to 1 for \\(i = t\\), and 0 otherwise. In this case, (3) is simplified to (2), which corresponds to hard labeling. In fact, \\(\\mathbf{q}\\) is not necessarily a one-hot distribution, but can be any distribution. The primary objective of \\(\\mathbf{q}\\) is to guide the learning of the LLM model by minimizing the distance between two distributions. Instead of forcing the model to learn from a strict, hard label with harsh penalization, soft labeling relaxes this constraint by using a smooth distribution \\(\\mathbf{q}\\) over all tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.483, + 0.658 + ], + "angle": 0, + "content": "In this work, we focus on numerical prediction, where numerical values are typically tokenized digit-wise [26]. That is, each digit (0-9) is treated as an individual token, and multi-digit numbers are split accordingly. For example, in the LLaMA tokenizer, number \"123\" is tokenized into the token IDs \"29896\", \"29906\", and \"29941\". Therefore, we restrict our soft labeling discussion to the 10 digit tokens from \"0\" to \"9\", leaving the soft labeling of regular tokens for future work. For instance, we do not consider the soft labeling for text tokens such as \"two\", \"eight\", etc. Following [6, 65], we defines soft label for the 10 digit tokens from \"0\" to \"9\" as" + }, + { + "type": "equation", + "bbox": [ + 0.181, + 0.672, + 0.482, + 0.69 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} ^ {S L} (t) = (1 - \\eta) \\boldsymbol {\\delta} (t) + \\eta \\psi (t) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "where \\( t \\) is the index of the 10 digit tokens \\( \\{ \"0\", \"1\", ..., \"9\" \\} \\). For simplicity, we assume that these indices are consecutive and form a set \\( S \\); if not, a straightforward mapping can be constructed to make them consecutive. \\( \\mathbf{q}^{SL}(t) \\) is a soft distribution to represent the token with index of \\( t \\). \\( \\psi(t) \\) is a probability distribution that has a support set on \\( S \\) and typically centers at the token with index of \\( t \\). It can be Poisson, Binomial, Triangular, Uniform distributions, etc. \\( \\eta \\in [0,1] \\) is a fixed hyperparameter that controls the mixture ratio of \\( \\delta(t) \\) and \\( \\psi(t) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "By smoothing the one-hot \\(\\mathbf{q}(t)\\) of hard labeling, soft labeling allows \\(q_{i}^{SL}(t)\\) to be greater than 0 for \\(i\\neq t\\) and to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.166 + ], + "angle": 0, + "content": "vary for different \\( i \\). This enables the assessment of relationship between the target token and other digit tokens. The relationship reflects the similarity between token indexed by \\( i \\) and the token indexed by \\( t \\), with \\( q_{i}^{SL}(t) \\) being higher when \\( i \\) is closer to \\( t \\) and lower otherwise." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.168, + 0.817, + 0.183 + ], + "angle": 0, + "content": "The final loss for each example is given by" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.19, + 0.905, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L = \\frac {1}{N _ {r} + N _ {n}} \\left(\\sum_ {\\left(x _ {r}, y _ {r}\\right) \\in \\left(\\mathbf {x} _ {r}, \\mathbf {y} _ {r}\\right)} L _ {C E} \\left(f \\left(x _ {r}\\right), \\mathbf {q} \\left(y _ {r}\\right)\\right) \\right. \\tag {5} \\\\ \\left. + \\lambda \\sum_ {\\left(x _ {n}, y _ {n}\\right) \\in \\left(\\mathbf {x} _ {n}, \\mathbf {y} _ {n}\\right)} L _ {C E} \\left(f \\left(x _ {n}\\right), \\mathbf {q} ^ {S L} \\left(y _ {n}\\right)\\right)\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.297, + 0.905, + 0.417 + ], + "angle": 0, + "content": "where \\(N_{r}, N_{n}\\) denote the numbers of regular tokens and numerical tokens within a sequence example, respectively. \\(\\mathbf{x}_r\\) \\((\\mathbf{x}_n)\\), \\(\\mathbf{y}_r\\) \\((\\mathbf{y}_n)\\) represent the sets of regular (numerical) tokens where \\(\\mathbf{x}_r\\) \\((\\mathbf{x}_n)\\) is the input set and \\(\\mathbf{y}_r\\) \\((\\mathbf{y}_n)\\) is the index set of ground truth. \\(f(\\cdot)\\) is the LLM with softmax output for each token prediction. \\(\\lambda\\) controls the balance between the numerical token loss and the regular token loss in the final loss computation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.905, + 0.78 + ], + "angle": 0, + "content": "When the target is a digit token, compared to hard labeling (one-hot encoding) which solely rely on the probability of target token in the loss function, as indicated in (2), soft labeling incorporates the prediction probabilities of all incorrect digit tokens into loss computation. This allows the loss value to reflect the distance between the predicted token and target token. It therefore can be understood as an implicit way of introducing distance-based penalties into the classification loss, similar to how loss functions operate in regression. As illustrated in Figure 1, an incorrect prediction of \"1\" (Prediction A) has a greater distributional distance from the soft ground truth than an incorrect prediction of \"3\" (Prediction B). Consequently, Prediction A incurs a larger loss than Prediction B, aligning with the behavior of regression losses. This property arises naturally from the inter-digit token similarity encoded in soft labels where \\( q_{i}^{SL}(t) \\) reflects the similarity ranking to the ground truth token. The closer a digit token is to the ground truth, the higher its value, enabling the model to account for the severity of prediction errors rather than treating all mistakes equally. In contrast, under hard labeling, predictions A and B receive the same loss due to the underlying assumption of one-hot encoding that all incorrect tokens are equally dissimilar to the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.906, + 0.901 + ], + "angle": 0, + "content": "From another perspective, in (3), \\( q_{i}(t) \\) can be viewed as a weight for each component loss \\( -\\log p_{i} \\). During training, a soft label with \\( q_{i}^{SL}(t) > 0 \\) encourages the model to predict higher probabilities on tokens closer to the target token, and lower probabilities on further tokens. This also aligns with the goal of regression-style loss, which aims to make predictions as close as possible to the target in continuous space." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3427" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.309, + 0.09, + 0.688, + 0.104 + ], + "angle": 0, + "content": "Table 1. The enhancement by soft labeling on visual grounding." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.114, + 0.904, + 0.246 + ], + "angle": 0, + "content": "
Pretrain modelModel typeFinetune dataLossRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
LLaVA-7BspecialistRefCOCO+/ghard soft81.787.374.473.682.063.379.078.8
85.0(+3.3)89.3(+2.0)78.8(+4.4)76.7(+3.1)83.4(+1.4)67.0(+3.7)82.0(+3.0)81.8(+3.0)
generalistLLaVA-Mix + RefCOCO+/ghard soft72.680.262.666.676.155.072.166.6
77.8(+5.2)83.7(+3.5)68.8(+6.2)72.0(+5.4)80.2(+4.1)60.4(+5.4)78.5(+6.4)71.5(+4.9)
LLaVA-13BspecialistRefCOCO+/ghard soft88.991.784.982.588.175.186.086.7
90.0(+1.1)92.7(+1.0)85.3(+0.4)83.5(+1.0)89.3(+1.2)76.7(+1.6)86.7(+0.7)87.4(+0.7)
generalistLLaVA-Mix + RefCOCO+/ghard soft79.385.071.474.782.964.878.073.0
81.0(+1.7)86.4(+1.4)72.6(+1.2)76.4(+1.7)83.3(+0.4)66.1(+1.3)79.2(+1.2)74.3(+1.3)
" + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.263, + 0.294, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.263, + 0.495, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.263, + 0.695, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.263, + 0.895, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.373, + 0.295, + 0.395 + ], + "angle": 0, + "content": "(a) RefCOCOg Acc. v.s. \\(\\eta\\) for Binomial distrib." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.399, + 0.294, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.373, + 0.496, + 0.395 + ], + "angle": 0, + "content": "(b) RefCOCOg Acc. v.s. \\(\\eta\\) for Poisson distrib." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.399, + 0.495, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.373, + 0.696, + 0.396 + ], + "angle": 0, + "content": "(c) RefCOCOg Acc. v.s. \\(\\eta\\) for Triangular distrib." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.399, + 0.695, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.373, + 0.898, + 0.395 + ], + "angle": 0, + "content": "(d) RefCOCOg Acc. v.s. \\(\\eta\\) for Uniform distrib." + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.399, + 0.895, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.508, + 0.296, + 0.53 + ], + "angle": 0, + "content": "(e) ChartQA Acc. v.s. \\(\\eta\\) for Binomial distrib." + }, + { + "type": "image", + "bbox": [ + 0.102, + 0.533, + 0.294, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.507, + 0.496, + 0.53 + ], + "angle": 0, + "content": "(f) ChartQA Acc. v.s. \\(\\eta\\) for Poisson distrib." + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.533, + 0.495, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.508, + 0.696, + 0.53 + ], + "angle": 0, + "content": "(g) ChartQA Acc. v.s. \\(\\eta\\) for Triangular distrib." + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.533, + 0.695, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.508, + 0.898, + 0.53 + ], + "angle": 0, + "content": "(h) ChartQA Acc. v.s. \\(\\eta\\) for Uniform distrib." + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.533, + 0.895, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.642, + 0.295, + 0.665 + ], + "angle": 0, + "content": "(i) TallyQA RMSE v.s. \\(\\eta\\) for Binomial distrib." + }, + { + "type": "image_caption", + "bbox": [ + 0.301, + 0.642, + 0.496, + 0.665 + ], + "angle": 0, + "content": "(j) TallyQA RMSE v.s. \\(\\eta\\) for Poisson distrib." + }, + { + "type": "image_caption", + "bbox": [ + 0.501, + 0.642, + 0.696, + 0.665 + ], + "angle": 0, + "content": "(k) TallyQA RMSE v.s. \\(\\eta\\) for Triangular distrib." + }, + { + "type": "image_caption", + "bbox": [ + 0.701, + 0.642, + 0.898, + 0.665 + ], + "angle": 0, + "content": "(1) TallyQA RMSE v.s. \\(\\eta\\) for Uniform distrib." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.668, + 0.905, + 0.696 + ], + "angle": 0, + "content": "Figure 2. Ablation study of \\(\\eta\\) of different distributions across three datasets, RefCOCOg val, ChartQA and TallyQA. On ChartQA, we evaluate on numerical subset. On TallyQA, we use the RMSE as metric and evaluate on simple balanced set." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.7, + 0.482, + 0.715 + ], + "angle": 0, + "content": "Table 2. The enhancement by soft labeling on chart understanding." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.726, + 0.493, + 0.858 + ], + "angle": 0, + "content": "
Pretrain modelModel typeFinetune dataLossChartQA
AllNumerical
LLaVA-7BspecialistChartQAhard soft67.966.3
69.6(+1.7)68.6(+2.3)
generalistLLaVA-Mix + ChartQAhard soft67.066.5
68.3(+1.3)68.6(+2.1)
LLaVA-13BspecialistChartQAhard soft72.872.0
73.4(+0.6)73.6(+1.6)
generalistLLaVA-Mix + ChartQAhard soft71.069.7
72.9 (+1.9)72.6(+2.9)
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.701, + 0.645, + 0.718 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this work, we evaluate our approach on three representative numerical prediction tasks: visual grounding, object counting, and chart understanding. For visual grounding, we assess performance on three benchmarks RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), and RefCOCOg (val/test) [25] using the standard Acc@0.5 metric. For object counting, we evaluate on TallyQA's simple and complex test sets [1]. However, the original test sets are count-imbalanced, with ground truth counts ranging from 0 to 15 but the combined counts of 1 and 2 account for \\(78\\%\\) of the simple testing set and \\(60\\%\\) of the complex testing" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3428" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.314, + 0.09, + 0.683, + 0.104 + ], + "angle": 0, + "content": "Table 3. The enhancement by soft labeling on object counting." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.113, + 0.905, + 0.248 + ], + "angle": 0, + "content": "
Pretrain modelModel typeFinetune dataLossTallyQA simpleTallyQA complexTallyQA simple balancedTallyQA complex balanced
Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)
LLaVA-7BspecialistTallyQAhard78.51.0066.91.3038.53.8416.76.68
soft79.0(+0.5)0.97(-0.03)67.2(+0.3)1.29(-0.01)39.4(+0.9)3.20(-0.64)18.8(+2.1)6.32(-0.36)
generalistLLaVA-Mix + TallyQAhard79.81.4368.22.5439.42.3321.84.96
soft80.5(+0.7)1.35(-0.08)69.4(+1.2)1.28(-1.26)42.2(+2.8)2.24(-0.09)24.0(+2.2)4.41(-0.55)
LLaVA-13BspecialistTallyQAhard81.90.8567.21.2740.42.8422.95.98
soft82.7(+0.8)0.67(-0.18)69.3(+2.1)1.20(-0.07)42.2(+1.8)2.02(-0.82)27.0(+4.1)5.16(-0.82)
generalistLLaVA-Mix + TallyQAhard82.00.8170.91.3045.92.3723.94.87
soft82.5(+0.5)0.76(-0.05)71.5(+0.6)1.24(-0.06)47.8(+1.9)1.88(-0.49)29.1(+5.2)4.45(-0.42)
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.264, + 0.341, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.166, + 0.4, + 0.266, + 0.412 + ], + "angle": 0, + "content": "(a) RefCOCOg val" + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.263, + 0.621, + 0.394 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.4, + 0.58, + 0.411 + ], + "angle": 0, + "content": "(b) ChartQA numerical subset" + }, + { + "type": "image", + "bbox": [ + 0.661, + 0.264, + 0.902, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.4, + 0.859, + 0.412 + ], + "angle": 0, + "content": "(c) TallyQA simple balanced" + }, + { + "type": "image_caption", + "bbox": [ + 0.293, + 0.413, + 0.701, + 0.427 + ], + "angle": 0, + "content": "Figure 3. Ablation study of \\(\\lambda\\) on three numerical prediction datasets." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.431, + 0.483, + 0.459 + ], + "angle": 0, + "content": "Table 4. The effect of different distributions. On TallyQA, RMSE is the metric." + }, + { + "type": "table", + "bbox": [ + 0.108, + 0.469, + 0.469, + 0.579 + ], + "angle": 0, + "content": "
DistributionRefCOCO valTallyQA balancedChartQA Numerical
**+*gSimpleComplex
Hard Label81.773.679.03.846.6866.3
Binomial84.776.881.93.136.2867.6
Posisson84.176.380.83.706.4967.0
Triangular85.076.782.03.206.3268.6
Uniform83.076.680.83.746.5666.8
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.595, + 0.483, + 0.776 + ], + "angle": 0, + "content": "set. In addition, the testing sets also contain noisy labels. These make the evaluation results misleading. To address this, we also construct a balanced subset for reference by sampling a count-balanced simple and complex testing sets from TallyQA's test sets respectively, referred to as \"TallyQA simple balanced\" and \"TallyQA complex balanced\". The details are in the Supp. Accuracy and RMSE are the chosen metrics following [1]. For chart understanding, we use ChartQA benchmark [40], where approximately half of the questions have numerical answers. We report accuracy on both the full test set and a subset containing only numerical answers." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Following [4], we adopt a three-stage pipeline, where the first stage is to learn the vision-language (VL) adapter, the second stage is to pretrain LLM and VL adapter on a large set of pretraining data, and the third stage is to finetune the pretrained model on some small scale finetuning data. To evaluate the broad applicability of our approach, we conduct experiments on various pretraining and finetuning settings. For the pretraining of generalist models," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.766 + ], + "angle": 0, + "content": "we use the publicly available LLaVA-7B architecture [32], pretrained on a sampled LAION-5M dataset [54], as our base model. For a stronger baseline, we use LLaVA-13B, pretrained on a 16M data mixture including LAION-5M for image captioning, GRIT-5M [47] for grounding, Objects365 [55] and OpenImages [27] of 3M in total for object counting, and 3M plot or chart related datasets such as PlotQA [43], UniChart [41], FigureQA [24] for chart understanding. For specialist model pretraining, we use three domain-specific data sources: a 12M mixture of grounding and detection data for visual grounding, a 6M counting data mixture with multiple tasks for object counting, and a 20M chart analysis data mixture for chart understanding, respectively. More details for pretraining data are given in Supp. After pretraining, the specialist models are finetuned on the task-specific training sets: RefCOCO, RefCOCO+, and RefCOCOg [25] for visual grounding; TallyQA [1] for object counting; and ChartQA [40] for chart understanding, and generalist models are finetuned on the same task-specific training sets but with the additional public 665K LLaVA-Mix tuning dataset [28]. The other settings are exactly the same as LLaVA's [32]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.813, + 0.79 + ], + "angle": 0, + "content": "4.1. The Effectiveness of Soft Labeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We first compare soft labeling and one-hot hard labeling across various pretrained models, including LLaVA-7B (pretrained on LAION-5M) and LLaVA-13B (pretrained on a 16M multi-task mixture). They are fine-tuned as specialist and generalist models and evaluated on multiple tasks and benchmarks. Tables 1, 2 and 3 report the performance comparisons for visual grounding, chart understanding and" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3429" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.089, + 0.486, + 0.119 + ], + "angle": 0, + "content": "Table 5. Performance comparison to the state-of-the-art models on visual grounding (Acc@0.5)." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.13, + 0.505, + 0.362 + ], + "angle": 0, + "content": "
ModelsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra-7B [12]87.090.680.281.687.472.182.382.2
Ferret-7B [70]87.591.482.580.887.473.183.984.8
Qwen-VL-7B [4]88.692.384.582.888.676.886.086.3
InternVL2-8B [13]87.191.180.779.887.971.482.782.7
VistaLLM [50]88.191.583.082.989.874.883.684.4
MiniGPT-v2 [11]88.791.785.380.085.174.584.484.7
Qwen2-VL-7B [4]91.793.687.385.890.579.587.387.8
Ferret-v2-7B [72]92.894.788.787.492.879.389.489.3
SoftLabeling-7B91.894.788.987.092.780.089.689.5
LION-12B [10]89.893.085.684.089.278.185.585.7
Shikra-13B [12]87.891.181.882.987.874.482.683.2
Ferret-13B [70]89.592.484.482.888.175.285.886.3
Ferret-v2-13B [72]92.695.088.987.492.181.489.490.0
SoftLabeling-13B92.795.089.087.692.782.389.890.0
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.486, + 0.697 + ], + "angle": 0, + "content": "object counting tasks, respectively. Some observations are available. First, soft labeling consistently improves performance across all model and benchmark combinations, with most gains larger than 2 points, demonstrating its effectiveness and robustness. Second, the absolute gains are smaller for the stronger 13B model compared to the 7B model. This is expected due to performance saturation. Nonetheless, notable improvements are remained, e.g., a 2.9 points gain in accuracy on ChartQA numerical subset (generalist model) in Table 2 and 4.1/5.2 points accuracy gains on TallyQA's simple/complex balanced test sets in Table 3. Third, on ChartQA, the gain on numerical subset is higher than the full set. This aligns with the design of soft labeling which is proposed specifically for enhancing the numerical prediction ability. Finally, on object counting, gains are more pronounced on the more reasonable balanced test sets, particularly in RMSE, a preferable metric reflecting error distance to the ground truth, where the relative improvements exceed \\(10\\%\\) in most cases. Notably, while enhancing numerical prediction, soft labeling does not degrade performance on generic tasks (please see details in Supp.)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.245, + 0.727 + ], + "angle": 0, + "content": "4.2. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Soft Distribution The distribution \\(\\psi\\) in (4) is crucial because it determines the degree of softness in soft labeling and the measurement of inter-digital token similarity. We investigate four widely used distributions, Binomial [6], Poisson [6], Triangular [65] and Uniform (label smoothing) [57]. The optimal \\(\\eta\\) in (4) is exhaustively searched for each distribution and used for comparison. We assess performance using some representative benchmarks and metrics on the base LLaVA-7B specialist model, as reported in Table 4. Regardless of the chosen distribution, soft labeling consistently outperforms hard labeling, demonstrating" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 6. Performance comparison to the state-of-the-art models on ChartQA." + }, + { + "type": "table", + "bbox": [ + 0.584, + 0.128, + 0.837, + 0.292 + ], + "angle": 0, + "content": "
ModelsChartQA
ChartPaLI-5B [8]77.3
ScreenAI 5B [3]76.7
MatCha4096 + LaMenDa [29]72.6
SMoLA-PaLI-X [69]74.6
Qwen-VL-Chat [4]66.3
Qwen-VL-Max [4]79.8
LLaVA-OV-7B [28]80.0
Cambrian-34B [62]73.8
Gemini Ultra [60]80.8
SoftLabeling81.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.307, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Table 7. Performance comparison to the state-of-the-art models on TallyQA. * indicates that the numbers are copied from a third-party paper [37]." + }, + { + "type": "table", + "bbox": [ + 0.533, + 0.359, + 0.888, + 0.508 + ], + "angle": 0, + "content": "
ModelsTallyQA simple Acc.(↑) RMSE(↓)TallyQA complex Acc.(↑) RMSE(↓)
SMoLA-PaLI-X [69]86.3-77.1-
PaLI-X-VPD [22]86.2-76.6-
MoVie-ResNeXt [45]74.91.0056.81.43
RCN [1]71.81.1356.21.43
LLaVA-NEXT [33]79.8*0.70*67.9*1.76*
GPT-4V [2]73.6*0.86*62.6*1.58*
GPT-4o [23]81.5*0.60*71.7*1.21*
SoftLabeling86.60.5677.21.06
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.662 + ], + "angle": 0, + "content": "its robustness. Among the distributions, Binomial and Triangular yield the best results, followed by Poisson, while Uniform performs the worst. This suggests that sharper distributions are more effective, as they provide more discriminative token similarities. In contrast, smoother distributions like Uniform treat all incorrect tokens as equally distant from the target, disregarding the relative distance information. Throughout this paper, we use Triangular distribution as default, though Binomial is comparable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Hyperparameter There are two primary hyperparameters influencing the performances, \\(\\eta\\) of (4) and \\(\\lambda\\) of (5). \\(\\eta\\) controls the sharpness of the soft labeling. It is selected through exhaustive search based on the performance of the base LLaVA-7B specialist model on RefCOCOg val, ChartQA and TallyQA. For ChartQA, we evaluate accuracy on numerical subset. On TallyQA, we use RMSE as metric and evaluate on simple balanced set. Figure 2 shows the ablation study results. Our analysis shows that the model favors sharper soft labeling distributions with smaller \\(\\eta\\). However, if it is too small, the distribution is towards a hard one-hot encoding, leading to performance drop. On the other hand, a large \\(\\eta\\) over smooths the distribution, reducing discrimination among numerical tokens. In our experiments, \\(\\eta\\) is set to \\(0.05 / 0.02 / 0.08 / 0.05\\) for Binomial/Poisson/Triangular/U" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "3430" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.91, + 0.272 + ], + "angle": 0, + "content": "
Image
Promptfirst sandwich on the left just beneath the forkthe man on the closest motorcycleWhat was the internet penetration rate in Africa in 2019?How many books did Blurb publish in 2018?
Hard labelingred boxred box2016000
Soft labelinggreen boxgreen box27.419053
Ground truthblue boxblue box28.619098
" + }, + { + "type": "table_caption", + "bbox": [ + 0.189, + 0.272, + 0.807, + 0.285 + ], + "angle": 0, + "content": "Table 8. Qualitative examples of soft labeling improvement in visual grounding and chart understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.383, + 0.306 + ], + "angle": 0, + "content": "niform distribution, respectively, by default." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.307, + 0.484, + 0.426 + ], + "angle": 0, + "content": "As for \\(\\lambda\\) of (5), it balances the contribution of numerical token loss and regular token loss in the final loss. We ablate its effect using the base 7B and stronger 13B specialist models across three datasets and present the results in Figure 3. It can be found that the models perform well and remain stable when \\(\\lambda\\) ranges from 2 to 5, beyond which the performance degrades. In our experiments, we set \\(\\lambda\\) as 2 by default." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.439, + 0.398, + 0.455 + ], + "angle": 0, + "content": "4.3. Comparison to the State-of-the-Art" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.46, + 0.485, + 0.733 + ], + "angle": 0, + "content": "We evaluate soft labeling and compare it with other methods across visual grounding, chart understanding, and object counting using public benchmarks, including RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), RefCOCOg (val/test), ChartQA (test), and TallyQA (simple/complex). The model is pretrained on scaled-up domain- and task-specific data and fine-tuned as specialist models (details in Section B of the Supp.). By default, our results are based on the 13B model, but for grounding, we follow standard practice and compare 7B and 13B models separately. As shown in Table 5, 6, and 7, our Soft-Labeling, without any architectural modifications, achieves the state-of-the-art performance on RefCOCOg among 7B scale models, RefCOCO, RefCOCO+, RefCOCOg among 13B scale models, ChartQA, and TallyQA. It also delivers competitive results on RefCOCO and RefCOCO+ in 7B scale model group. These results highlight the effectiveness of soft labeling." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.228, + 0.758 + ], + "angle": 0, + "content": "4.4. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Table 8 presents examples randomly selected from visual RefCOCOg and ChartQA test sets, demonstrating the improvement achieved by soft labeling. Compared to hard labeling, soft labeling produces more accurate numerical predictions for both bounding box coordinates and chart-based values, bringing them closer to the ground truth. For instance, in the motorcycle image, the bounding box predicted by soft labeling (green) better aligns with the ground truth (blue) in both position and scale. Similarly, in the chart-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.292, + 0.907, + 0.352 + ], + "angle": 0, + "content": "based QA task, soft labeling provides more accurate numerical estimation. These examples highlight the effectiveness of soft labeling in improving numerical prediction across both spatial and quantitative tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.363, + 0.628, + 0.379 + ], + "angle": 0, + "content": "5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.617 + ], + "angle": 0, + "content": "In this work, we limit our scope to digit tokens. However, other tokens also exhibit semantic similarity, such as natural language numerals (\"one/two/three\") or categorical tokens (\"easy/medium/hard\"). Incorporating these into the loss computation is an avenue for future research. Another limitation is that our current loss function considers only single digits. Many numerical predictions involve multi-digit numbers, where digit-wise comparisons may not accurately reflect actual numerical distances. For instance, comparing \"21\" v.s. \"32\" digit-wise results in a distance of \\( |2 - 3| + |1 - 2| = 2 \\) which is smaller than \\( |2 - 1| + |2 - 9| = 8 \\) for \"21\" v.s. \"29\". However, their actual numerical distances are \\( |21 - 32| = 11 \\) and \\( |21 - 29| = 8 \\), showing a discrepancy in digit-wise calculations. Addressing this limitation remains an open challenge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.634, + 0.635 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.9 + ], + "angle": 0, + "content": "In this paper, we have identified the limitations of cross-entropy loss with one-hot encoding for numerical prediction in LLMs/MLLMs, where treating numbers as discrete tokens fails to reflect their continuous nature. To address this, we have proposed soft labeling, which smooths target distributions to account for numerical proximity, ensuring that closer predictions incur lower loss. This method retains compatibility with the next-token prediction paradigm without requiring architectural modifications. Our experiments across visual grounding, object counting, and chart understanding have shown that soft labeling consistently improves numerical prediction performance. We have achieved state-of-the-art or competitive results on RefCOCO \\(+ / g\\), ChartQA, and TallyQA using a basic LLaVA model. Our findings highlight the importance of adapting appropriate loss functions for numerical tasks in LLMs/MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3431" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI, pages 8076-8084, 2019. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.383 + ], + "angle": 0, + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arxiv 2023. arXiv preprint arXiv:2308.12966, 1(8), 2023. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.457 + ], + "angle": 0, + "content": "[5] Javier Barbero-Gomez, Pedro-Antonio Gutierrez, VICTOR-Manuel Vargas, Juan-Antonio Vallejo-Casas, and Cesar Hervás-Martínez. An ordinal cnn approach for the assessment of neurological damage in parkinson's disease patients. Expert Systems with Applications, 182:115271, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.5 + ], + "angle": 0, + "content": "[6] Christopher Beckham and Christopher Pal. Unimodal probability distributions for deep ordinal classification. In ICML, pages 411-419. PMLR, 2017. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.502, + 0.483, + 0.571 + ], + "angle": 0, + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[8] Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[9] Alberto Castano, Pablo González, Jaime Alonso González, and Juan Jose Del Coz. Matching distributions algorithms based on the earth mover's distance for ordinal quantification. IEEE T-NNLS, 35(1):1050-1061, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[10] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In CVPR, pages 26540-26550, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[14] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Haotian Wang, Ming Liu, and Bing Qin. Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models. arXiv preprint arXiv:2311.17667, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[15] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In CVPR, pages 4738-4747, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[16] Bahare Fatemi, Mehran Kazemi, Anton Tsitsulin, Karishma Malkan, Jinyeong Yim, John Palowitch, Sungyong Seo, Jonathan Halcrow, and Bryan Perozzi. Test of time: A benchmark for evaluating llms on temporal reasoning. arXiv preprint arXiv:2406.09170, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[17] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.435, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[19] Pedro Antonio Gutierrez, Maria Perez-Ortiz, Javier Sanchez-Monedero, Francisco Fernandez-Navarro, and Cesar Hervas-Martinez. Ordinal regression methods: survey and experimental study. IEEE Transactions on Knowledge and Data Engineering, 28(1):127-146, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.505, + 0.905, + 0.546 + ], + "angle": 0, + "content": "[20] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared earth mover's distance-based loss for training deep neural networks. arXiv preprint arXiv:1611.05916, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[21] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035–1046, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.605, + 0.905, + 0.674 + ], + "angle": 0, + "content": "[22] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[24] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, pages 787-798, 2014. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[26] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3432" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[29] Zhuowan Li, Bhavan Jasani, Peng Tang, and Shabnam Ghadar. Synthesize step-by-step: Tools templates and llms as data generators for reasoning-based chart vqa. In CVPR, pages 13613-13623, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[30] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, pages 80-88, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[34] Peiyang Liu, Xiangyu Xi, Wei Ye, and Shikun Zhang. Label smoothing for text mining. In ICCL, pages 2210-2219, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[35] Xiaofeng Liu, Yang Zou, Yuhang Song, Chao Yang, Jane You, and BV K Vijaya Kumar. Ordinal regression with neuron stick-breaking for medical diagnosis. In ECCV Workshops, pages 0-0, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[36] Xiaofeng Liu, Fangfang Fan, Lingsheng Kong, Zhihui Dao, Wanqing Xie, Jun Lu, and Jane You. Unimodal regularized neuron stick-breaking for ordinal classification. Neurocomputing, 388:34-44, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[37] Jian Lu, Shikhar Srivastava, Junyu Chen, Robik Shrestha, Manoj Acharya, Kushal Kafle, and Christopher Kanan. Revisiting multi-modal llm evaluation. arXiv preprint arXiv:2408.05334, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[38] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[39] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[43] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1527-1536, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.904, + 0.207 + ], + "angle": 0, + "content": "[44] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? NeurIPS, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.249 + ], + "angle": 0, + "content": "[45] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions for visual counting and beyond. arXiv preprint arXiv:2004.11883, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.251, + 0.905, + 0.292 + ], + "angle": 0, + "content": "[46] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In CVPR, pages 4920-4928, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.905, + 0.349 + ], + "angle": 0, + "content": "[47] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.351, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[48] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[49] Gorkem Polat, Ilkay Ergenc, Haluk Tarik Kani, Yesim Ozen Alahdab, Ozlen Atug, and Alptekin Temizel. Class distance weighted cross-entropy loss for ulcerative colitis severity estimation. In Annual Conference on Medical Image Understanding and Analysis, pages 157–171. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.478, + 0.905, + 0.547 + ], + "angle": 0, + "content": "[50] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In CVPR, pages 14076-14088, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.615 + ], + "angle": 0, + "content": "[51] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.66 + ], + "angle": 0, + "content": "[52] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.662, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[53] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[55] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, pages 8430-8439, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3433" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[57] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.483, + 0.202 + ], + "angle": 0, + "content": "[58] Zichang Tan, Shuai Zhou, Jun Wan, Zhen Lei, and Stan Z Li. Age estimation based on a single network with soft softmax of aging modeling. In ACCV, pages 203-216. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[60] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[61] Christian Thiel. Classification on soft labels is robust against label noise. In International Conference on Knowledge-Based and Intelligent Information and Engineering Systems, pages 65-73. Springer, 2008. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[62] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[63] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[64] Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In ICCV, pages 4068-4076, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[65] Víctor Manuel Vargas, Pedro Antonio Gutiérrez, Javier Barbero-Gómez, and César Hervás-Martínez. Soft labelling based on triangular distributions for ordinal classification. Information Fusion, 93:258-267, 2023. 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[66] Sukesh Adiga Vasudeva, Jose Dolz, and Herve Lombaert. Geols: geodesic label smoothing for image segmentation. In Medical Imaging with Deep Learning, pages 468-478. PMLR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.733 + ], + "angle": 0, + "content": "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. NeurIPS, 37:95095-95169, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[68] Scott Workman, Richard Souvenir, and Nathan Jacobs. Understanding and mapping natural beauty. In ICCV, pages 5589-5598, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[69] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR, pages 14205-14215, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[70] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[71] Chang-Bin Zhang, Peng-Tao Jiang, Qibin Hou, Yunchao Wei, Qi Han, Zhen Li, and Ming-Ming Cheng. Delving deep into label smoothing. IEEE T-IP, 30:5984-5996, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[72] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[73] Helong Zhou, Liangchen Song, Jiajie Chen, Ye Zhou, Guoli Wang, Junsong Yuan, and Qian Zhang. Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650, 2021. 3" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3434" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_origin.pdf b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..581cdd9d408929a0a193ffc989ee772b5bacbc5c --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/fe57947e-0779-4b96-b592-99d87fccfb07_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3e5c4f8f0ad66f2adcfe27b326e7be710ae3aa7fec2bbaca8b0c0848213681fa +size 750686 diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/full.md b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f904fea2fc797b52fdbea5ccde387c63d180fda5 --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/full.md @@ -0,0 +1,318 @@ +# Enhancing Numerical Prediction of MLLMs with Soft Labeling + +Pei Wang, Zhaowei Cai, Hao Yang, Davide Modolo, Ashwin Swaminathan +Amazon AGI + +{pwwng,zhaoweic,haoyng,dmodolo,swashwin}@amazon.com + +# Abstract + +The optimality of using the de facto cross-entropy loss with one-hot target distribution (hard labeling) is questioned when training (Multimodal) Large Language Models (LLMs/MLLMs). Although it is reasonable for language token prediction, which is a typical multi-class classification problem in discrete space, it is suboptimal for task like numerical prediction, which is a typical regression problem in continuous space. However, enabling regression in LLMs/MLLMs will complicate the training and next-token prediction paradigm at inference. Instead, to address this challenge, we propose a novel loss design, called soft labeling, which smooths the target probability distribution, enabling predictions to be penalized according to their distance to the target. This is similar to regression loss, which penalizes more on the further predictions in the continuous space, but will not change the model architecture and the next-token prediction paradigm of LLMs/MLLMs. We demonstrate the efficacy of soft labeling through extensive experiments on visual grounding, object counting, and chart understanding, achieving state-of-the-art performance on multiple benchmarks without bells and whistles. Soft labeling can be applied in any LLM/MLLM. + +# 1. Introduction + +In recent years, thanks to the success of Large Language Models (LLMs) [7, 52, 63], Multimodal Large Language Models (MLLMs) [2, 4, 31, 60] have achieved remarkable progresses across various tasks, such as image captioning, visual question answering, and many specialized domains like document analysis, chart interpretation, and visual grounding, etc. Their learning is to predict the next token, associated with the de facto cross-entropy loss across the token vocabulary, no matter the target is a word, digit, punctuation, or anything else. Although this unified mechanism simplifies the learning across various tasks and it becomes dominant, a natural question to ask is: is it optimal for any task? For example, the multi-class classification paradigm is reasonable for language word token predic + +NNNN +![](images/e4191c460f4eab8707ef47d6a9b5c18dccd3e9d409a679b368e5c2cc30f92d06.jpg) +the prediction and target is considered. + +![](images/16c86ddf6c8b54bf4dd516b49f34b9e499c6fee6aa262857d560a10f61817c9f.jpg) + +![](images/0390ecc79235a4d0f700bea63ed5322eabf79df8b971a130149c4a122f6a7179.jpg) +Loss $\mathrm{B} = 1.20$ +Loss B=1.47 + +![](images/4b1075f62dc75a346207e86e800fed943b4ab01b483762a53988185adea3cfb3.jpg) +Loss $A = 1.20$ + +![](images/7ebf6f4fd75f87b3dfdd6d61b244e37897d7620f2a38adea9b404789795ec611.jpg) +Loss $A = 1.73$ + +Figure 1. Given an image with 4 trains, the MLLMs generate probabilistic predictions (A and B) for the count. In standard one-hot hard labeling, only the ground-truth numerical token has probability of 1, leading to equal loss values (1.20) for those two incorrect predictions (1 and 3 trains). In contrast, soft labeling assigns probabilities to numerical tokens according to their distances to the target. This results in different loss values (1.73 for prediction A of 1 train but 1.47 for prediction B of 3 trains), enabling higher penalty for further prediction. + +tion because language word is in discrete space, but it is no longer optimal for objects that are in continuous space. In this paper, we are investigating the limits of standard cross-entropy loss in an important case, i.e., numerical prediction, which is in continuous space. + +Numerical prediction is a critical capability for LLMs and MLLMs, with many applications in object counting, object localization, chart/figure understanding, etc. For instance, answering the gross income from a tax report, or answering how many trains in a traffic scene as in Figure 1. Although the gross income value (e.g., 35421.7) and the number of trains (e.g., 4) are in continuous space, they are converted to strings (e.g., "35421.7" and "4") and then to discrete tokens to be processed by LLMs/MLLMs as a multi-class classification problem with cross-entropy loss. Cross-entropy loss measures the distance between the target and prediction distributions over the token vocabulary, and the target distribution is usually a one-hot vector where + +the probability on the target is 1 and the rest are 0. In this case, the wrong prediction to any non-target index can potentially have no difference. For example, in Figure 1, where there are 4 trains in a given image, predictions A and B have exactly the same cross-entropy loss because their probabilities on the target (i.e., 4) are the same (i.e., 0.3), although their overall distributions are quite different from each other. However, this is counter-intuitive. Prediction B of 3 trains should have lower loss than prediction A of 1 train, because the former is closer to the target. The cause of this issue is rooted in the one-hot representation of the target probability distribution, namely hard labeling, of the standard cross-entropy loss. + +To address this problem, a new loss design is required, which can better reflect how wrong the prediction is for objects in continuous space. An ideal choice would be regression-style loss, e.g., L1 or L2 loss. However, asking LLMs/MLLMs to output scalar values is nontrivial without extra layers or heuristics. In addition, this will break the loss/architecture unification of LLMs/MLLMs and make the next-token prediction paradigm very complicated. Instead, we propose soft labeling for LLM/MLLM training, in contrast with hard labeling of the standard cross-entropy loss. In soft labeling, the target probability is no longer a one-hot probability distribution (Figure 1 middle), but smoothed where closer (further) classes have higher (lower) probabilities (Figure 1 bottom). Different from hard labeling, where only the target probability contributes to the loss, soft labeling enables losses coming from nearby classes, and considers the relative distance between the predicted and target tokens, similar to the regression loss. For example, in Figure 1, prediction A and B have the same loss using hard labeling, but the loss of B is lower than that of A for soft labeling. This is reasonable because prediction B of 3 trains is closer to the target of 4. In addition, the design of soft labeling will not modify the model architecture and next-token generation paradigm, which can be used in any LLM/MLLM. + +In our experiments, we demonstrate the effectiveness of the proposed soft labeling over hard labeling across three representative tasks of numerical prediction: visual grounding, object counting, and chart understanding on multiple benchmarks. Soft labeling consistently improves performance on all baselines. We also conduct comprehensive ablation studies on each component of soft labeling to gain deeper understanding of their impacts. By applying soft labeling, we achieve state-of-the-art or competitive performance on the RefCOCO+/g [25], ChartQA [40], and TallyQA [1] datasets, using only a standard LLaVA model [32] without any additional architectural modifications. This highlights the effectiveness of soft labeling in enhancing numerical prediction in MLLMs. + +Overall, this work makes three major contributions. + +First, we question the optimality of using the de facto cross-entropy loss with one-hot target encoding (hard labeling) in any task of LLMs/MLLMs, and explain why standard cross-entropy loss leads to counter-intuitive loss assignment for a critical task, i.e., numerical prediction, in continuous space by nature. Second, we introduce soft labeling, where target probability distributions are smoothed such that closer (further) tokens to the target receive higher (lower) probabilities. This design better reflects the degree of error in predictions while maintaining compatibility with the standard next-token prediction paradigm of LLMs/MLLMs. Third, extensive experiments on visual grounding, object counting, and chart understanding demonstrate that soft labeling consistently outperforms hard labeling. The proposed approach achieves state-of-the-art or competitive results on benchmarks like RefCOCO, ChartQA, and TallyQA, without any architectural modification. + +# 2. Related Work + +Numerical prediction is a core challenge across a wide range of tasks in MLLMs and LLMs, spanning visual grounding [12, 48, 70], object counting [1], chart understanding [40, 43], visual question answering [18, 56], mathematical reasoning [38, 67], temporal reasoning [14, 16, 51] and document analysis [42, 59]. In visual grounding, models must quantify spatial location of objects with numerical bounding box coordinates [12, 48, 70]. In object counting, the model is required to predict precise numerical numbers of instances [1, 53]. Chart understanding relies on extracting and reasoning over numerical data in visual form [40, 43]. General VQA often involves answering questions that require numerical reasoning, such as counting, comparing quantities, or understanding scales [18, 56]. Mathematical reasoning, including arithmetic and algebraic problem-solving, tests models' ability to manipulate numerical values symbolically [38, 67]. Temporal reasoning in multimodal setups involves understanding and predicting time intervals or event sequences from textual and visual data [14, 16, 51]. Document analysis further highlights numerical prediction challenges, where models must extract and interpret numerical information from structured and unstructured data and make reasoning [42, 59]. These diverse applications underscore the pervasive role of numerical prediction in MLLMs and LLMs, making it a critical area for improvement. + +Ordinal regression (also called ordinal classification) is the problems where the target variable exhibits a relative ordering on an arbitrary scale, e.g. categories such as "bad", "good" and "very good". It has been proven useful in different research areas, including medical research [5, 35], monocular depth estimation [17], age estimation [46], pose estimation [21], etc. Ordinal regression has been well studies in conventional machine learn + +ing methods [19]. In the scenarios of deep learning, there are some typical methods. Class distance weighted loss relies on a coefficient for the loss of each class, which utilizes the distance to the ground-truth class and increases in relation to that distance [49]. The Earth Mover's Distance (EMD) loss uses the predicted probabilities of all classes and penalizes the miss-predictions according to a ground distance matrix that quantifies the dissimilarities between classes [9, 20]. Soft labeling converts ground truth data labels into soft probability distributions and the loss is computed on them [6, 15, 36, 65]. However, these methods have only shown success on simple multi category classification problem. For LLMs, the probability space of the output has tens to hundreds of thousands of tokens and they are mixture of numerical and non-numerical. The problem becomes difficulty. In this work, we will show how to apply ordinal regression to LLMs. + +Soft labeling uses probability distributions over classes as labels instead of hard, one-hot labels. A typical type is label smoothing which has been found more robust to noisy annotations for classification problem [39, 61, 71] and has been applied to many fields such as know distillation [44, 73], network calibration [30], and some applications like image segmentation [66], text mining [34]. On the other hand, soft loss terms have been useful for domain and task transfer [64] in order to avoid dataset biases. Elaborate loss functions are defined in [68] to take into account the subjective scenicness of outdoor pictures, by trying to predict the same rating distribution of human annotations. Age estimation is another particular niche where soft labels have become popular [15, 58]. However, there is few research of applying soft labeling on numerical prediction of LLMs. In this paper, we will show that the label smoothing is just a specific format of our soft labeling and our proposal performs better. + +# 3. Method + +In traditional regression models, numerical prediction is typically handled using L1/L2-style loss functions, which penalize errors based on their distance to the target. Further distance incurs greater penalties and vice versa, encouraging the model to make predictions that are closer to the target. However, when training LLMs/MLLMs, the de facto cross-entropy loss with one-hot encoded target fails to account for the distance between predicted and target tokens properly. It does so in a limited and imprecise manner, a misclassification is a misclassification, regardless of how explicitly far the predicted token is from the target token. This limitation hinders the model's ability to predict accurate numerical values. To address this, a natural idea is to apply a regression-style loss to LLMs. However, in the context of LLMs, it is not trivial to directly convert the LLM's output into a scalar without introducing additional layers or heuristics. For this reason, we opt for a new loss de + +sign which can be directly applied to the intrinsic output of LLM, similar to the standard cross-entropy loss. + +We begin by reviewing the LLM training loss. LLMs typically generate text token by token, treating the next-token prediction as a classification problem over a vocabulary of size $V$ . Given a model output logit $\mathbf{z} \in R^{V}$ from the last layer before the final prediction, the softmax function is applied to convert these logits into probability distribution $\mathbf{p} \in R^{V}$ such that $\sum_{i=1}^{V} p_{i} = 1$ . This distribution is then compared to the target token using cross-entropy loss. Typically, the target token is encoded into a one-hot vector $\mathbf{q} \in R^{V}$ where it is a binary vector consisting of a value 1 at the correct token index, and 0s everywhere else. The cross-entropy loss is computed by + +$$ +L (\mathbf {p}, \mathbf {q}) = - \sum_ {i = 1} ^ {V} q _ {i} \log p _ {i}. \tag {1} +$$ + +Since $\mathbf{q}$ is a one-hot vector, the final loss simplifies to + +$$ +L (\mathbf {p}, \mathbf {q}) = - \log p _ {t}, \tag {2} +$$ + +where $t$ is the index of the target token. As shown, the loss function depends solely on the posterior probability $p_t$ of the target token. This ignores other rich information inherited in the full probability distribution $\mathbf{p}$ . + +In fact, each probability $p_i$ reflects the model confidence on each possible token. When a prediction is incorrect, it is crucial to have a metric that precisely measures how "wrong" the prediction is by explicitly considering the model's confidence on all tokens, not just the target token. This becomes important when the target token is numerical and the numerical prediction is done by LLMs. Numerical tokens have relative distance between each other by nature, and when a prediction is incorrect, we should measure how wrong it is by assessing the distance from the target, rather than treating all wrong predictions equally as long as the confidence on the target token is the same, as is done in hard labeling. For example, in the scenario illustrated in Figure 1 to answer how many trains given an image, when we train an MLLM with cross-entropy loss and one-hot encoding, there are two possible incorrect predictions, A and B, during MLLM training. If we apply hard labeling to encode the supervision "4", both predictions incur the same cross-entropy loss, as both have a softmax probability of 0.3 for the target token "4". This approach fails to account for numerical proximity since prediction B, which assigns a maximum probability of 0.4 to the incorrect token "3", is closer to the correct answer "4" than prediction A, which assigns a maximum probability of 0.4 to the incorrect token "1". Intuitively, prediction B should receive a lower loss from the regression perspective. This limitation arises from the one-hot encoding in hard labeling which only retains the probability $p_t$ of the target token in the final loss computation + +of (2). As a result, the final loss function does not consider the probabilities $p_i$ for other numerical tokens, which contain valuable information about inter-token relationship and numerical distance. The absence of this information makes the hard labeling loss suboptimal, and we need a new loss function that can better account for such relationships. + +In order to address this issue, we propose soft labeling cross-entropy loss. We first reformulate the cross-entropy loss of (1) into a more general form, + +$$ +L _ {C E} (\mathbf {p}, \mathbf {q} (t)) = \sum_ {i = 1} ^ {V} q _ {i} (t) (- \log p _ {i}), \tag {3} +$$ + +where $\mathbf{q}(t)$ is the distribution in which the target token with index of $t$ is encoded, and $q_{i}(t) \in [0,1]$ s.t. $\sum_{i} q_{i}(t) = 1$ . In the standard one-hot encoding for cross-entropy, $\mathbf{q}(t) = \delta(t)$ is a Dirac delta distribution where $q_{i}(t)$ equals to 1 for $i = t$ , and 0 otherwise. In this case, (3) is simplified to (2), which corresponds to hard labeling. In fact, $\mathbf{q}$ is not necessarily a one-hot distribution, but can be any distribution. The primary objective of $\mathbf{q}$ is to guide the learning of the LLM model by minimizing the distance between two distributions. Instead of forcing the model to learn from a strict, hard label with harsh penalization, soft labeling relaxes this constraint by using a smooth distribution $\mathbf{q}$ over all tokens. + +In this work, we focus on numerical prediction, where numerical values are typically tokenized digit-wise [26]. That is, each digit (0-9) is treated as an individual token, and multi-digit numbers are split accordingly. For example, in the LLaMA tokenizer, number "123" is tokenized into the token IDs "29896", "29906", and "29941". Therefore, we restrict our soft labeling discussion to the 10 digit tokens from "0" to "9", leaving the soft labeling of regular tokens for future work. For instance, we do not consider the soft labeling for text tokens such as "two", "eight", etc. Following [6, 65], we defines soft label for the 10 digit tokens from "0" to "9" as + +$$ +\mathbf {q} ^ {S L} (t) = (1 - \eta) \boldsymbol {\delta} (t) + \eta \psi (t) \tag {4} +$$ + +where $t$ is the index of the 10 digit tokens $\{ "0", "1", ..., "9" \}$ . For simplicity, we assume that these indices are consecutive and form a set $S$ ; if not, a straightforward mapping can be constructed to make them consecutive. $\mathbf{q}^{SL}(t)$ is a soft distribution to represent the token with index of $t$ . $\psi(t)$ is a probability distribution that has a support set on $S$ and typically centers at the token with index of $t$ . It can be Poisson, Binomial, Triangular, Uniform distributions, etc. $\eta \in [0,1]$ is a fixed hyperparameter that controls the mixture ratio of $\delta(t)$ and $\psi(t)$ . + +By smoothing the one-hot $\mathbf{q}(t)$ of hard labeling, soft labeling allows $q_{i}^{SL}(t)$ to be greater than 0 for $i\neq t$ and to + +vary for different $i$ . This enables the assessment of relationship between the target token and other digit tokens. The relationship reflects the similarity between token indexed by $i$ and the token indexed by $t$ , with $q_{i}^{SL}(t)$ being higher when $i$ is closer to $t$ and lower otherwise. + +The final loss for each example is given by + +$$ +\begin{array}{l} L = \frac {1}{N _ {r} + N _ {n}} \left(\sum_ {\left(x _ {r}, y _ {r}\right) \in \left(\mathbf {x} _ {r}, \mathbf {y} _ {r}\right)} L _ {C E} \left(f \left(x _ {r}\right), \mathbf {q} \left(y _ {r}\right)\right) \right. \tag {5} \\ \left. + \lambda \sum_ {\left(x _ {n}, y _ {n}\right) \in \left(\mathbf {x} _ {n}, \mathbf {y} _ {n}\right)} L _ {C E} \left(f \left(x _ {n}\right), \mathbf {q} ^ {S L} \left(y _ {n}\right)\right)\right), \\ \end{array} +$$ + +where $N_{r}, N_{n}$ denote the numbers of regular tokens and numerical tokens within a sequence example, respectively. $\mathbf{x}_r$ $(\mathbf{x}_n)$ , $\mathbf{y}_r$ $(\mathbf{y}_n)$ represent the sets of regular (numerical) tokens where $\mathbf{x}_r$ $(\mathbf{x}_n)$ is the input set and $\mathbf{y}_r$ $(\mathbf{y}_n)$ is the index set of ground truth. $f(\cdot)$ is the LLM with softmax output for each token prediction. $\lambda$ controls the balance between the numerical token loss and the regular token loss in the final loss computation. + +When the target is a digit token, compared to hard labeling (one-hot encoding) which solely rely on the probability of target token in the loss function, as indicated in (2), soft labeling incorporates the prediction probabilities of all incorrect digit tokens into loss computation. This allows the loss value to reflect the distance between the predicted token and target token. It therefore can be understood as an implicit way of introducing distance-based penalties into the classification loss, similar to how loss functions operate in regression. As illustrated in Figure 1, an incorrect prediction of "1" (Prediction A) has a greater distributional distance from the soft ground truth than an incorrect prediction of "3" (Prediction B). Consequently, Prediction A incurs a larger loss than Prediction B, aligning with the behavior of regression losses. This property arises naturally from the inter-digit token similarity encoded in soft labels where $q_{i}^{SL}(t)$ reflects the similarity ranking to the ground truth token. The closer a digit token is to the ground truth, the higher its value, enabling the model to account for the severity of prediction errors rather than treating all mistakes equally. In contrast, under hard labeling, predictions A and B receive the same loss due to the underlying assumption of one-hot encoding that all incorrect tokens are equally dissimilar to the ground truth. + +From another perspective, in (3), $q_{i}(t)$ can be viewed as a weight for each component loss $-\log p_{i}$ . During training, a soft label with $q_{i}^{SL}(t) > 0$ encourages the model to predict higher probabilities on tokens closer to the target token, and lower probabilities on further tokens. This also aligns with the goal of regression-style loss, which aims to make predictions as close as possible to the target in continuous space. + +Table 1. The enhancement by soft labeling on visual grounding. + +
Pretrain modelModel typeFinetune dataLossRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
LLaVA-7BspecialistRefCOCO+/ghard soft81.787.374.473.682.063.379.078.8
85.0(+3.3)89.3(+2.0)78.8(+4.4)76.7(+3.1)83.4(+1.4)67.0(+3.7)82.0(+3.0)81.8(+3.0)
generalistLLaVA-Mix + RefCOCO+/ghard soft72.680.262.666.676.155.072.166.6
77.8(+5.2)83.7(+3.5)68.8(+6.2)72.0(+5.4)80.2(+4.1)60.4(+5.4)78.5(+6.4)71.5(+4.9)
LLaVA-13BspecialistRefCOCO+/ghard soft88.991.784.982.588.175.186.086.7
90.0(+1.1)92.7(+1.0)85.3(+0.4)83.5(+1.0)89.3(+1.2)76.7(+1.6)86.7(+0.7)87.4(+0.7)
generalistLLaVA-Mix + RefCOCO+/ghard soft79.385.071.474.782.964.878.073.0
81.0(+1.7)86.4(+1.4)72.6(+1.2)76.4(+1.7)83.3(+0.4)66.1(+1.3)79.2(+1.2)74.3(+1.3)
+ +![](images/afe2d20f7150c3212449e505288003a4d15a0f5e840b621be6453155900b719f.jpg) +(a) RefCOCOg Acc. v.s. $\eta$ for Binomial distrib. + +![](images/cb87aa59e99909317c5b29be7962ef69cf25869b65d34c15ee15164bce7775c4.jpg) +(b) RefCOCOg Acc. v.s. $\eta$ for Poisson distrib. + +![](images/7b9876d45f43a5e58b733476308b46a7e43d137c3a2d67c8845335501698dcfe.jpg) +(c) RefCOCOg Acc. v.s. $\eta$ for Triangular distrib. + +![](images/5c846cfc78de4dee874d63a472b664aad9b8a62e6cc59bf870e9cb6b0c71cafc.jpg) +(d) RefCOCOg Acc. v.s. $\eta$ for Uniform distrib. + +![](images/acad06ac982ac8545e94a6d063031aef1fee360b170dfb0e9513b11ab4284383.jpg) +(e) ChartQA Acc. v.s. $\eta$ for Binomial distrib. + +![](images/5216aa1e2cefcdd523677e4c45a6cef60511a57dcaa16474207bf498b98a5c68.jpg) +(f) ChartQA Acc. v.s. $\eta$ for Poisson distrib. + +![](images/9b99053548e53eb56fd22f72148891690c91c5c8b9bdcac754e9241f1686f5ee.jpg) +(g) ChartQA Acc. v.s. $\eta$ for Triangular distrib. + +![](images/d12edcd12dc0b3f9fbd1eacbf6b920a3a45f3f9ffc13fccc873ccb8f3ac91032.jpg) +(h) ChartQA Acc. v.s. $\eta$ for Uniform distrib. + +![](images/cb29fc0a698c3888e7fccfc687f61548ee71b517506774f809f5a04f38657d4c.jpg) +(i) TallyQA RMSE v.s. $\eta$ for Binomial distrib. + +![](images/f2a48044bde08e09ff808b886ed7e7ac00b34adc927a2c676e69c4b9cca44b19.jpg) +(j) TallyQA RMSE v.s. $\eta$ for Poisson distrib. + +![](images/9c1ee120d8c086ea8b26c24295a6e246ac04227f8c6a191bac78bf9a00b070ef.jpg) +Figure 2. Ablation study of $\eta$ of different distributions across three datasets, RefCOCOg val, ChartQA and TallyQA. On ChartQA, we evaluate on numerical subset. On TallyQA, we use the RMSE as metric and evaluate on simple balanced set. + +![](images/f33decacc9922b9eb88176a7ad4354eae961a544a9ddcd887c4c8cd3903f633f.jpg) +(k) TallyQA RMSE v.s. $\eta$ for Triangular distrib. +(1) TallyQA RMSE v.s. $\eta$ for Uniform distrib. + +Table 2. The enhancement by soft labeling on chart understanding. + +
Pretrain modelModel typeFinetune dataLossChartQA
AllNumerical
LLaVA-7BspecialistChartQAhard soft67.966.3
69.6(+1.7)68.6(+2.3)
generalistLLaVA-Mix + ChartQAhard soft67.066.5
68.3(+1.3)68.6(+2.1)
LLaVA-13BspecialistChartQAhard soft72.872.0
73.4(+0.6)73.6(+1.6)
generalistLLaVA-Mix + ChartQAhard soft71.069.7
72.9 (+1.9)72.6(+2.9)
+ +# 4. Experiments + +In this work, we evaluate our approach on three representative numerical prediction tasks: visual grounding, object counting, and chart understanding. For visual grounding, we assess performance on three benchmarks RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), and RefCOCOg (val/test) [25] using the standard Acc@0.5 metric. For object counting, we evaluate on TallyQA's simple and complex test sets [1]. However, the original test sets are count-imbalanced, with ground truth counts ranging from 0 to 15 but the combined counts of 1 and 2 account for $78\%$ of the simple testing set and $60\%$ of the complex testing + +Table 3. The enhancement by soft labeling on object counting. + +
Pretrain modelModel typeFinetune dataLossTallyQA simpleTallyQA complexTallyQA simple balancedTallyQA complex balanced
Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)
LLaVA-7BspecialistTallyQAhard78.51.0066.91.3038.53.8416.76.68
soft79.0(+0.5)0.97(-0.03)67.2(+0.3)1.29(-0.01)39.4(+0.9)3.20(-0.64)18.8(+2.1)6.32(-0.36)
generalistLLaVA-Mix + TallyQAhard79.81.4368.22.5439.42.3321.84.96
soft80.5(+0.7)1.35(-0.08)69.4(+1.2)1.28(-1.26)42.2(+2.8)2.24(-0.09)24.0(+2.2)4.41(-0.55)
LLaVA-13BspecialistTallyQAhard81.90.8567.21.2740.42.8422.95.98
soft82.7(+0.8)0.67(-0.18)69.3(+2.1)1.20(-0.07)42.2(+1.8)2.02(-0.82)27.0(+4.1)5.16(-0.82)
generalistLLaVA-Mix + TallyQAhard82.00.8170.91.3045.92.3723.94.87
soft82.5(+0.5)0.76(-0.05)71.5(+0.6)1.24(-0.06)47.8(+1.9)1.88(-0.49)29.1(+5.2)4.45(-0.42)
+ +![](images/7c51225a2da420922bd021d28afa931d794101c8bfb16079bcc5f66a94e189b3.jpg) +(a) RefCOCOg val + +![](images/eaa663f1a38b69465a29722f887c7e07dbd464a9ce8b0da6bee646e0c690ba95.jpg) +(b) ChartQA numerical subset + +![](images/40ba5169c74e5d603d0299885b79fc7dcb7a0db8afccf15630881fdcb9f6c837.jpg) +(c) TallyQA simple balanced +Figure 3. Ablation study of $\lambda$ on three numerical prediction datasets. + +Table 4. The effect of different distributions. On TallyQA, RMSE is the metric. + +
DistributionRefCOCO valTallyQA balancedChartQA Numerical
**+*gSimpleComplex
Hard Label81.773.679.03.846.6866.3
Binomial84.776.881.93.136.2867.6
Posisson84.176.380.83.706.4967.0
Triangular85.076.782.03.206.3268.6
Uniform83.076.680.83.746.5666.8
+ +set. In addition, the testing sets also contain noisy labels. These make the evaluation results misleading. To address this, we also construct a balanced subset for reference by sampling a count-balanced simple and complex testing sets from TallyQA's test sets respectively, referred to as "TallyQA simple balanced" and "TallyQA complex balanced". The details are in the Supp. Accuracy and RMSE are the chosen metrics following [1]. For chart understanding, we use ChartQA benchmark [40], where approximately half of the questions have numerical answers. We report accuracy on both the full test set and a subset containing only numerical answers. + +Following [4], we adopt a three-stage pipeline, where the first stage is to learn the vision-language (VL) adapter, the second stage is to pretrain LLM and VL adapter on a large set of pretraining data, and the third stage is to finetune the pretrained model on some small scale finetuning data. To evaluate the broad applicability of our approach, we conduct experiments on various pretraining and finetuning settings. For the pretraining of generalist models, + +we use the publicly available LLaVA-7B architecture [32], pretrained on a sampled LAION-5M dataset [54], as our base model. For a stronger baseline, we use LLaVA-13B, pretrained on a 16M data mixture including LAION-5M for image captioning, GRIT-5M [47] for grounding, Objects365 [55] and OpenImages [27] of 3M in total for object counting, and 3M plot or chart related datasets such as PlotQA [43], UniChart [41], FigureQA [24] for chart understanding. For specialist model pretraining, we use three domain-specific data sources: a 12M mixture of grounding and detection data for visual grounding, a 6M counting data mixture with multiple tasks for object counting, and a 20M chart analysis data mixture for chart understanding, respectively. More details for pretraining data are given in Supp. After pretraining, the specialist models are finetuned on the task-specific training sets: RefCOCO, RefCOCO+, and RefCOCOg [25] for visual grounding; TallyQA [1] for object counting; and ChartQA [40] for chart understanding, and generalist models are finetuned on the same task-specific training sets but with the additional public 665K LLaVA-Mix tuning dataset [28]. The other settings are exactly the same as LLaVA's [32]. + +# 4.1. The Effectiveness of Soft Labeling + +We first compare soft labeling and one-hot hard labeling across various pretrained models, including LLaVA-7B (pretrained on LAION-5M) and LLaVA-13B (pretrained on a 16M multi-task mixture). They are fine-tuned as specialist and generalist models and evaluated on multiple tasks and benchmarks. Tables 1, 2 and 3 report the performance comparisons for visual grounding, chart understanding and + +Table 5. Performance comparison to the state-of-the-art models on visual grounding (Acc@0.5). + +
ModelsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra-7B [12]87.090.680.281.687.472.182.382.2
Ferret-7B [70]87.591.482.580.887.473.183.984.8
Qwen-VL-7B [4]88.692.384.582.888.676.886.086.3
InternVL2-8B [13]87.191.180.779.887.971.482.782.7
VistaLLM [50]88.191.583.082.989.874.883.684.4
MiniGPT-v2 [11]88.791.785.380.085.174.584.484.7
Qwen2-VL-7B [4]91.793.687.385.890.579.587.387.8
Ferret-v2-7B [72]92.894.788.787.492.879.389.489.3
SoftLabeling-7B91.894.788.987.092.780.089.689.5
LION-12B [10]89.893.085.684.089.278.185.585.7
Shikra-13B [12]87.891.181.882.987.874.482.683.2
Ferret-13B [70]89.592.484.482.888.175.285.886.3
Ferret-v2-13B [72]92.695.088.987.492.181.489.490.0
SoftLabeling-13B92.795.089.087.692.782.389.890.0
+ +object counting tasks, respectively. Some observations are available. First, soft labeling consistently improves performance across all model and benchmark combinations, with most gains larger than 2 points, demonstrating its effectiveness and robustness. Second, the absolute gains are smaller for the stronger 13B model compared to the 7B model. This is expected due to performance saturation. Nonetheless, notable improvements are remained, e.g., a 2.9 points gain in accuracy on ChartQA numerical subset (generalist model) in Table 2 and 4.1/5.2 points accuracy gains on TallyQA's simple/complex balanced test sets in Table 3. Third, on ChartQA, the gain on numerical subset is higher than the full set. This aligns with the design of soft labeling which is proposed specifically for enhancing the numerical prediction ability. Finally, on object counting, gains are more pronounced on the more reasonable balanced test sets, particularly in RMSE, a preferable metric reflecting error distance to the ground truth, where the relative improvements exceed $10\%$ in most cases. Notably, while enhancing numerical prediction, soft labeling does not degrade performance on generic tasks (please see details in Supp.) + +# 4.2. Ablation Study + +Soft Distribution The distribution $\psi$ in (4) is crucial because it determines the degree of softness in soft labeling and the measurement of inter-digital token similarity. We investigate four widely used distributions, Binomial [6], Poisson [6], Triangular [65] and Uniform (label smoothing) [57]. The optimal $\eta$ in (4) is exhaustively searched for each distribution and used for comparison. We assess performance using some representative benchmarks and metrics on the base LLaVA-7B specialist model, as reported in Table 4. Regardless of the chosen distribution, soft labeling consistently outperforms hard labeling, demonstrating + +Table 6. Performance comparison to the state-of-the-art models on ChartQA. + +
ModelsChartQA
ChartPaLI-5B [8]77.3
ScreenAI 5B [3]76.7
MatCha4096 + LaMenDa [29]72.6
SMoLA-PaLI-X [69]74.6
Qwen-VL-Chat [4]66.3
Qwen-VL-Max [4]79.8
LLaVA-OV-7B [28]80.0
Cambrian-34B [62]73.8
Gemini Ultra [60]80.8
SoftLabeling81.5
+ +Table 7. Performance comparison to the state-of-the-art models on TallyQA. * indicates that the numbers are copied from a third-party paper [37]. + +
ModelsTallyQA simple Acc.(↑) RMSE(↓)TallyQA complex Acc.(↑) RMSE(↓)
SMoLA-PaLI-X [69]86.3-77.1-
PaLI-X-VPD [22]86.2-76.6-
MoVie-ResNeXt [45]74.91.0056.81.43
RCN [1]71.81.1356.21.43
LLaVA-NEXT [33]79.8*0.70*67.9*1.76*
GPT-4V [2]73.6*0.86*62.6*1.58*
GPT-4o [23]81.5*0.60*71.7*1.21*
SoftLabeling86.60.5677.21.06
+ +its robustness. Among the distributions, Binomial and Triangular yield the best results, followed by Poisson, while Uniform performs the worst. This suggests that sharper distributions are more effective, as they provide more discriminative token similarities. In contrast, smoother distributions like Uniform treat all incorrect tokens as equally distant from the target, disregarding the relative distance information. Throughout this paper, we use Triangular distribution as default, though Binomial is comparable. + +Hyperparameter There are two primary hyperparameters influencing the performances, $\eta$ of (4) and $\lambda$ of (5). $\eta$ controls the sharpness of the soft labeling. It is selected through exhaustive search based on the performance of the base LLaVA-7B specialist model on RefCOCOg val, ChartQA and TallyQA. For ChartQA, we evaluate accuracy on numerical subset. On TallyQA, we use RMSE as metric and evaluate on simple balanced set. Figure 2 shows the ablation study results. Our analysis shows that the model favors sharper soft labeling distributions with smaller $\eta$ . However, if it is too small, the distribution is towards a hard one-hot encoding, leading to performance drop. On the other hand, a large $\eta$ over smooths the distribution, reducing discrimination among numerical tokens. In our experiments, $\eta$ is set to $0.05 / 0.02 / 0.08 / 0.05$ for Binomial/Poisson/Triangular/U + +
Image
Promptfirst sandwich on the left just beneath the forkthe man on the closest motorcycleWhat was the internet penetration rate in Africa in 2019?How many books did Blurb publish in 2018?
Hard labelingred boxred box2016000
Soft labelinggreen boxgreen box27.419053
Ground truthblue boxblue box28.619098
+ +Table 8. Qualitative examples of soft labeling improvement in visual grounding and chart understanding. + +niform distribution, respectively, by default. + +As for $\lambda$ of (5), it balances the contribution of numerical token loss and regular token loss in the final loss. We ablate its effect using the base 7B and stronger 13B specialist models across three datasets and present the results in Figure 3. It can be found that the models perform well and remain stable when $\lambda$ ranges from 2 to 5, beyond which the performance degrades. In our experiments, we set $\lambda$ as 2 by default. + +# 4.3. Comparison to the State-of-the-Art + +We evaluate soft labeling and compare it with other methods across visual grounding, chart understanding, and object counting using public benchmarks, including RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), RefCOCOg (val/test), ChartQA (test), and TallyQA (simple/complex). The model is pretrained on scaled-up domain- and task-specific data and fine-tuned as specialist models (details in Section B of the Supp.). By default, our results are based on the 13B model, but for grounding, we follow standard practice and compare 7B and 13B models separately. As shown in Table 5, 6, and 7, our Soft-Labeling, without any architectural modifications, achieves the state-of-the-art performance on RefCOCOg among 7B scale models, RefCOCO, RefCOCO+, RefCOCOg among 13B scale models, ChartQA, and TallyQA. It also delivers competitive results on RefCOCO and RefCOCO+ in 7B scale model group. These results highlight the effectiveness of soft labeling. + +# 4.4. Visualization + +Table 8 presents examples randomly selected from visual RefCOCOg and ChartQA test sets, demonstrating the improvement achieved by soft labeling. Compared to hard labeling, soft labeling produces more accurate numerical predictions for both bounding box coordinates and chart-based values, bringing them closer to the ground truth. For instance, in the motorcycle image, the bounding box predicted by soft labeling (green) better aligns with the ground truth (blue) in both position and scale. Similarly, in the chart- + +based QA task, soft labeling provides more accurate numerical estimation. These examples highlight the effectiveness of soft labeling in improving numerical prediction across both spatial and quantitative tasks. + +# 5. Limitation + +In this work, we limit our scope to digit tokens. However, other tokens also exhibit semantic similarity, such as natural language numerals ("one/two/three") or categorical tokens ("easy/medium/hard"). Incorporating these into the loss computation is an avenue for future research. Another limitation is that our current loss function considers only single digits. Many numerical predictions involve multi-digit numbers, where digit-wise comparisons may not accurately reflect actual numerical distances. For instance, comparing "21" v.s. "32" digit-wise results in a distance of $|2 - 3| + |1 - 2| = 2$ which is smaller than $|2 - 1| + |2 - 9| = 8$ for "21" v.s. "29". However, their actual numerical distances are $|21 - 32| = 11$ and $|21 - 29| = 8$ , showing a discrepancy in digit-wise calculations. Addressing this limitation remains an open challenge. + +# 6. Conclusion + +In this paper, we have identified the limitations of cross-entropy loss with one-hot encoding for numerical prediction in LLMs/MLLMs, where treating numbers as discrete tokens fails to reflect their continuous nature. To address this, we have proposed soft labeling, which smooths target distributions to account for numerical proximity, ensuring that closer predictions incur lower loss. This method retains compatibility with the next-token prediction paradigm without requiring architectural modifications. Our experiments across visual grounding, object counting, and chart understanding have shown that soft labeling consistently improves numerical prediction performance. We have achieved state-of-the-art or competitive results on RefCOCO $+ / g$ , ChartQA, and TallyQA using a basic LLaVA model. Our findings highlight the importance of adapting appropriate loss functions for numerical tasks in LLMs/MLLMs. + +# References + +[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI, pages 8076-8084, 2019. 2, 5, 6, 7 +[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 7 +[3] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 7 +[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arxiv 2023. arXiv preprint arXiv:2308.12966, 1(8), 2023. 1, 6, 7 +[5] Javier Barbero-Gomez, Pedro-Antonio Gutierrez, VICTOR-Manuel Vargas, Juan-Antonio Vallejo-Casas, and Cesar Hervás-Martínez. An ordinal cnn approach for the assessment of neurological damage in parkinson's disease patients. Expert Systems with Applications, 182:115271, 2021. 2 +[6] Christopher Beckham and Christopher Pal. Unimodal probability distributions for deep ordinal classification. In ICML, pages 411-419. PMLR, 2017. 3, 4, 7 +[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 1 +[8] Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596, 2024. 7 +[9] Alberto Castano, Pablo González, Jaime Alonso González, and Juan Jose Del Coz. Matching distributions algorithms based on the earth mover's distance for ordinal quantification. IEEE T-NNLS, 35(1):1050-1061, 2022. 3 +[10] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In CVPR, pages 26540-26550, 2024. 7 +[11] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7 +[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 7 + +[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 7 +[14] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Haotian Wang, Ming Liu, and Bing Qin. Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models. arXiv preprint arXiv:2311.17667, 2023. 2 +[15] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In CVPR, pages 4738-4747, 2019. 3 +[16] Bahare Fatemi, Mehran Kazemi, Anton Tsitsulin, Karishma Malkan, Jinyeong Yim, John Palowitch, Sungyong Seo, Jonathan Halcrow, and Bryan Perozzi. Test of time: A benchmark for evaluating llms on temporal reasoning. arXiv preprint arXiv:2406.09170, 2024. 2 +[17] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 2 +[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 2 +[19] Pedro Antonio Gutierrez, Maria Perez-Ortiz, Javier Sanchez-Monedero, Francisco Fernandez-Navarro, and Cesar Hervas-Martinez. Ordinal regression methods: survey and experimental study. IEEE Transactions on Knowledge and Data Engineering, 28(1):127-146, 2015. 3 +[20] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared earth mover's distance-based loss for training deep neural networks. arXiv preprint arXiv:1611.05916, 2016. 3 +[21] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035–1046, 2018. 2 +[22] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 7 +[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7 +[24] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 6 +[25] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, pages 787-798, 2014. 2, 5, 6 +[26] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4 + +[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6 +[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 7 +[29] Zhuowan Li, Bhavan Jasani, Peng Tang, and Shabnam Ghadar. Synthesize step-by-step: Tools templates and llms as data generators for reasoning-based chart vqa. In CVPR, pages 13613-13623, 2024. 7 +[30] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, pages 80-88, 2022. 3 +[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 1 +[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 2, 6 +[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.7 +[34] Peiyang Liu, Xiangyu Xi, Wei Ye, and Shikun Zhang. Label smoothing for text mining. In ICCL, pages 2210-2219, 2022. 3 +[35] Xiaofeng Liu, Yang Zou, Yuhang Song, Chao Yang, Jane You, and BV K Vijaya Kumar. Ordinal regression with neuron stick-breaking for medical diagnosis. In ECCV Workshops, pages 0-0, 2018. 2 +[36] Xiaofeng Liu, Fangfang Fan, Lingsheng Kong, Zhihui Dao, Wanqing Xie, Jun Lu, and Jane You. Unimodal regularized neuron stick-breaking for ordinal classification. Neurocomputing, 388:34-44, 2020. 3 +[37] Jian Lu, Shikhar Srivastava, Junyu Chen, Robik Shrestha, Manoj Acharya, Kushal Kafle, and Christopher Kanan. Revisiting multi-modal llm evaluation. arXiv preprint arXiv:2408.05334, 2024. 7 +[38] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2 +[39] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 3 +[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 6 +[41] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 6 + +[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 2 +[43] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1527-1536, 2020. 2, 6 +[44] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? NeurIPS, 32, 2019. 3 +[45] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions for visual counting and beyond. arXiv preprint arXiv:2004.11883, 2020. 7 +[46] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In CVPR, pages 4920-4928, 2016. 2 +[47] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6 +[48] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In ICLR, 2024. 2 +[49] Gorkem Polat, Ilkay Ergenc, Haluk Tarik Kani, Yesim Ozen Alahdab, Ozlen Atug, and Alptekin Temizel. Class distance weighted cross-entropy loss for ulcerative colitis severity estimation. In Annual Conference on Medical Image Understanding and Analysis, pages 157–171. Springer, 2022. 3 +[50] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In CVPR, pages 14076-14088, 2024. 7 +[51] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2 +[52] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1 +[53] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 2 +[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 6 +[55] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, pages 8430-8439, 2019. 6 +[56] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 2 + +[57] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7 +[58] Zichang Tan, Shuai Zhou, Jun Wan, Zhen Lei, and Stan Z Li. Age estimation based on a single network with soft softmax of aging modeling. In ACCV, pages 203-216. Springer, 2016. 3 +[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 2 +[60] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 7 +[61] Christian Thiel. Classification on soft labels is robust against label noise. In International Conference on Knowledge-Based and Intelligent Information and Engineering Systems, pages 65-73. Springer, 2008. 3 +[62] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 7 +[63] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1 +[64] Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In ICCV, pages 4068-4076, 2015. 3 +[65] Víctor Manuel Vargas, Pedro Antonio Gutiérrez, Javier Barbero-Gómez, and César Hervás-Martínez. Soft labelling based on triangular distributions for ordinal classification. Information Fusion, 93:258-267, 2023. 3, 4, 7 +[66] Sukesh Adiga Vasudeva, Jose Dolz, and Herve Lombaert. Geols: geodesic label smoothing for image segmentation. In Medical Imaging with Deep Learning, pages 468-478. PMLR, 2024. 3 +[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. NeurIPS, 37:95095-95169, 2024. 2 +[68] Scott Workman, Richard Souvenir, and Nathan Jacobs. Understanding and mapping natural beauty. In ICCV, pages 5589-5598, 2017. 3 +[69] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR, pages 14205-14215, 2024. 7 +[70] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 7 + +[71] Chang-Bin Zhang, Peng-Tao Jiang, Qibin Hou, Yunchao Wei, Qi Han, Zhen Li, and Ming-Ming Cheng. Delving deep into label smoothing. IEEE T-IP, 30:5984-5996, 2021. 3 +[72] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 7 +[73] Helong Zhou, Liangchen Song, Jiajie Chen, Ye Zhou, Guoli Wang, Junsong Yuan, and Qian Zhang. Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650, 2021. 3 \ No newline at end of file diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/images.zip b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..403448e1d32fde7ee17dea78b0f87d1699012b77 --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:926a7a604f516c077988c5fb37392977129ceeb781efe51e542f25e782be1eed +size 679665 diff --git a/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/layout.json b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8aac184b52118bdac574d63ce729a056c7028ec2 --- /dev/null +++ b/2025/Enhancing Numerical Prediction of MLLMs with Soft Labeling/layout.json @@ -0,0 +1,9366 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 501, + 121 + ], + "type": "text", + "content": "Enhancing Numerical Prediction of MLLMs with Soft Labeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 124, + 143, + 486, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 143, + 486, + 173 + ], + "spans": [ + { + "bbox": [ + 124, + 143, + 486, + 173 + ], + "type": "text", + "content": "Pei Wang, Zhaowei Cai, Hao Yang, Davide Modolo, Ashwin Swaminathan \nAmazon AGI" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 176, + 440, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 176, + 440, + 188 + ], + "spans": [ + { + "bbox": [ + 164, + 176, + 440, + 188 + ], + "type": "text", + "content": "{pwwng,zhaoweic,haoyng,dmodolo,swashwin}@amazon.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "spans": [ + { + "bbox": [ + 151, + 216, + 200, + 228 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 240, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 240, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 296, + 504 + ], + "type": "text", + "content": "The optimality of using the de facto cross-entropy loss with one-hot target distribution (hard labeling) is questioned when training (Multimodal) Large Language Models (LLMs/MLLMs). Although it is reasonable for language token prediction, which is a typical multi-class classification problem in discrete space, it is suboptimal for task like numerical prediction, which is a typical regression problem in continuous space. However, enabling regression in LLMs/MLLMs will complicate the training and next-token prediction paradigm at inference. Instead, to address this challenge, we propose a novel loss design, called soft labeling, which smooths the target probability distribution, enabling predictions to be penalized according to their distance to the target. This is similar to regression loss, which penalizes more on the further predictions in the continuous space, but will not change the model architecture and the next-token prediction paradigm of LLMs/MLLMs. We demonstrate the efficacy of soft labeling through extensive experiments on visual grounding, object counting, and chart understanding, achieving state-of-the-art performance on multiple benchmarks without bells and whistles. Soft labeling can be applied in any LLM/MLLM." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "In recent years, thanks to the success of Large Language Models (LLMs) [7, 52, 63], Multimodal Large Language Models (MLLMs) [2, 4, 31, 60] have achieved remarkable progresses across various tasks, such as image captioning, visual question answering, and many specialized domains like document analysis, chart interpretation, and visual grounding, etc. Their learning is to predict the next token, associated with the de facto cross-entropy loss across the token vocabulary, no matter the target is a word, digit, punctuation, or anything else. Although this unified mechanism simplifies the learning across various tasks and it becomes dominant, a natural question to ask is: is it optimal for any task? For example, the multi-class classification paradigm is reasonable for language word token predic" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 219, + 375, + 277 + ], + "blocks": [ + { + "bbox": [ + 318, + 219, + 375, + 277 + ], + "lines": [ + { + "bbox": [ + 318, + 219, + 375, + 277 + ], + "spans": [ + { + "bbox": [ + 318, + 219, + 375, + 277 + ], + "type": "image", + "image_path": "e4191c460f4eab8707ef47d6a9b5c18dccd3e9d409a679b368e5c2cc30f92d06.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 383, + 219, + 394, + 276 + ], + "lines": [ + { + "bbox": [ + 383, + 219, + 394, + 276 + ], + "spans": [ + { + "bbox": [ + 383, + 219, + 394, + 276 + ], + "type": "text", + "content": "NNNN" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 427, + 365, + 529, + 374 + ], + "lines": [ + { + "bbox": [ + 427, + 365, + 529, + 374 + ], + "spans": [ + { + "bbox": [ + 427, + 365, + 529, + 374 + ], + "type": "text", + "content": "the prediction and target is considered." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 402, + 219, + 465, + 277 + ], + "blocks": [ + { + "bbox": [ + 402, + 219, + 465, + 277 + ], + "lines": [ + { + "bbox": [ + 402, + 219, + 465, + 277 + ], + "spans": [ + { + "bbox": [ + 402, + 219, + 465, + 277 + ], + "type": "image", + "image_path": "16c86ddf6c8b54bf4dd516b49f34b9e499c6fee6aa262857d560a10f61817c9f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 484, + 219, + 547, + 278 + ], + "blocks": [ + { + "bbox": [ + 484, + 219, + 547, + 278 + ], + "lines": [ + { + "bbox": [ + 484, + 219, + 547, + 278 + ], + "spans": [ + { + "bbox": [ + 484, + 219, + 547, + 278 + ], + "type": "image", + "image_path": "0390ecc79235a4d0f700bea63ed5322eabf79df8b971a130149c4a122f6a7179.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 289, + 537, + 297 + ], + "lines": [ + { + "bbox": [ + 496, + 289, + 537, + 297 + ], + "spans": [ + { + "bbox": [ + 496, + 289, + 537, + 297 + ], + "type": "text", + "content": "Loss " + }, + { + "bbox": [ + 496, + 289, + 537, + 297 + ], + "type": "inline_equation", + "content": "\\mathrm{B} = 1.20" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 496, + 344, + 537, + 351 + ], + "lines": [ + { + "bbox": [ + 496, + 344, + 537, + 351 + ], + "spans": [ + { + "bbox": [ + 496, + 344, + 537, + 351 + ], + "type": "text", + "content": "Loss B=1.47" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 281, + 388, + 330 + ], + "blocks": [ + { + "bbox": [ + 322, + 281, + 388, + 330 + ], + "lines": [ + { + "bbox": [ + 322, + 281, + 388, + 330 + ], + "spans": [ + { + "bbox": [ + 322, + 281, + 388, + 330 + ], + "type": "image", + "image_path": "4b1075f62dc75a346207e86e800fed943b4ab01b483762a53988185adea3cfb3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 289, + 459, + 297 + ], + "lines": [ + { + "bbox": [ + 417, + 289, + 459, + 297 + ], + "spans": [ + { + "bbox": [ + 417, + 289, + 459, + 297 + ], + "type": "text", + "content": "Loss " + }, + { + "bbox": [ + 417, + 289, + 459, + 297 + ], + "type": "inline_equation", + "content": "A = 1.20" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 322, + 339, + 388, + 384 + ], + "blocks": [ + { + "bbox": [ + 322, + 339, + 388, + 384 + ], + "lines": [ + { + "bbox": [ + 322, + 339, + 388, + 384 + ], + "spans": [ + { + "bbox": [ + 322, + 339, + 388, + 384 + ], + "type": "image", + "image_path": "7ebf6f4fd75f87b3dfdd6d61b244e37897d7620f2a38adea9b404789795ec611.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 344, + 459, + 352 + ], + "lines": [ + { + "bbox": [ + 417, + 344, + 459, + 352 + ], + "spans": [ + { + "bbox": [ + 417, + 344, + 459, + 352 + ], + "type": "text", + "content": "Loss " + }, + { + "bbox": [ + 417, + 344, + 459, + 352 + ], + "type": "inline_equation", + "content": "A = 1.73" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 488 + ], + "type": "text", + "content": "Figure 1. Given an image with 4 trains, the MLLMs generate probabilistic predictions (A and B) for the count. In standard one-hot hard labeling, only the ground-truth numerical token has probability of 1, leading to equal loss values (1.20) for those two incorrect predictions (1 and 3 trains). In contrast, soft labeling assigns probabilities to numerical tokens according to their distances to the target. This results in different loss values (1.73 for prediction A of 1 train but 1.47 for prediction B of 3 trains), enabling higher penalty for further prediction." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 493, + 555, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 553 + ], + "type": "text", + "content": "tion because language word is in discrete space, but it is no longer optimal for objects that are in continuous space. In this paper, we are investigating the limits of standard cross-entropy loss in an important case, i.e., numerical prediction, which is in continuous space." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 713 + ], + "type": "text", + "content": "Numerical prediction is a critical capability for LLMs and MLLMs, with many applications in object counting, object localization, chart/figure understanding, etc. For instance, answering the gross income from a tax report, or answering how many trains in a traffic scene as in Figure 1. Although the gross income value (e.g., 35421.7) and the number of trains (e.g., 4) are in continuous space, they are converted to strings (e.g., \"35421.7\" and \"4\") and then to discrete tokens to be processed by LLMs/MLLMs as a multi-class classification problem with cross-entropy loss. Cross-entropy loss measures the distance between the target and prediction distributions over the token vocabulary, and the target distribution is usually a one-hot vector where" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3424" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "the probability on the target is 1 and the rest are 0. In this case, the wrong prediction to any non-target index can potentially have no difference. For example, in Figure 1, where there are 4 trains in a given image, predictions A and B have exactly the same cross-entropy loss because their probabilities on the target (i.e., 4) are the same (i.e., 0.3), although their overall distributions are quite different from each other. However, this is counter-intuitive. Prediction B of 3 trains should have lower loss than prediction A of 1 train, because the former is closer to the target. The cause of this issue is rooted in the one-hot representation of the target probability distribution, namely hard labeling, of the standard cross-entropy loss." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 231, + 295, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 528 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 528 + ], + "type": "text", + "content": "To address this problem, a new loss design is required, which can better reflect how wrong the prediction is for objects in continuous space. An ideal choice would be regression-style loss, e.g., L1 or L2 loss. However, asking LLMs/MLLMs to output scalar values is nontrivial without extra layers or heuristics. In addition, this will break the loss/architecture unification of LLMs/MLLMs and make the next-token prediction paradigm very complicated. Instead, we propose soft labeling for LLM/MLLM training, in contrast with hard labeling of the standard cross-entropy loss. In soft labeling, the target probability is no longer a one-hot probability distribution (Figure 1 middle), but smoothed where closer (further) classes have higher (lower) probabilities (Figure 1 bottom). Different from hard labeling, where only the target probability contributes to the loss, soft labeling enables losses coming from nearby classes, and considers the relative distance between the predicted and target tokens, similar to the regression loss. For example, in Figure 1, prediction A and B have the same loss using hard labeling, but the loss of B is lower than that of A for soft labeling. This is reasonable because prediction B of 3 trains is closer to the target of 4. In addition, the design of soft labeling will not modify the model architecture and next-token generation paradigm, which can be used in any LLM/MLLM." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 532, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 699 + ], + "type": "text", + "content": "In our experiments, we demonstrate the effectiveness of the proposed soft labeling over hard labeling across three representative tasks of numerical prediction: visual grounding, object counting, and chart understanding on multiple benchmarks. Soft labeling consistently improves performance on all baselines. We also conduct comprehensive ablation studies on each component of soft labeling to gain deeper understanding of their impacts. By applying soft labeling, we achieve state-of-the-art or competitive performance on the RefCOCO+/g [25], ChartQA [40], and TallyQA [1] datasets, using only a standard LLaVA model [32] without any additional architectural modifications. This highlights the effectiveness of soft labeling in enhancing numerical prediction in MLLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "Overall, this work makes three major contributions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "First, we question the optimality of using the de facto cross-entropy loss with one-hot target encoding (hard labeling) in any task of LLMs/MLLMs, and explain why standard cross-entropy loss leads to counter-intuitive loss assignment for a critical task, i.e., numerical prediction, in continuous space by nature. Second, we introduce soft labeling, where target probability distributions are smoothed such that closer (further) tokens to the target receive higher (lower) probabilities. This design better reflects the degree of error in predictions while maintaining compatibility with the standard next-token prediction paradigm of LLMs/MLLMs. Third, extensive experiments on visual grounding, object counting, and chart understanding demonstrate that soft labeling consistently outperforms hard labeling. The proposed approach achieves state-of-the-art or competitive results on benchmarks like RefCOCO, ChartQA, and TallyQA, without any architectural modification." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 286, + 400, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 286, + 400, + 300 + ], + "spans": [ + { + "bbox": [ + 314, + 286, + 400, + 300 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 307, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 307, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 307, + 553, + 617 + ], + "type": "text", + "content": "Numerical prediction is a core challenge across a wide range of tasks in MLLMs and LLMs, spanning visual grounding [12, 48, 70], object counting [1], chart understanding [40, 43], visual question answering [18, 56], mathematical reasoning [38, 67], temporal reasoning [14, 16, 51] and document analysis [42, 59]. In visual grounding, models must quantify spatial location of objects with numerical bounding box coordinates [12, 48, 70]. In object counting, the model is required to predict precise numerical numbers of instances [1, 53]. Chart understanding relies on extracting and reasoning over numerical data in visual form [40, 43]. General VQA often involves answering questions that require numerical reasoning, such as counting, comparing quantities, or understanding scales [18, 56]. Mathematical reasoning, including arithmetic and algebraic problem-solving, tests models' ability to manipulate numerical values symbolically [38, 67]. Temporal reasoning in multimodal setups involves understanding and predicting time intervals or event sequences from textual and visual data [14, 16, 51]. Document analysis further highlights numerical prediction challenges, where models must extract and interpret numerical information from structured and unstructured data and make reasoning [42, 59]. These diverse applications underscore the pervasive role of numerical prediction in MLLMs and LLMs, making it a critical area for improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Ordinal regression (also called ordinal classification) is the problems where the target variable exhibits a relative ordering on an arbitrary scale, e.g. categories such as \"bad\", \"good\" and \"very good\". It has been proven useful in different research areas, including medical research [5, 35], monocular depth estimation [17], age estimation [46], pose estimation [21], etc. Ordinal regression has been well studies in conventional machine learn" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3425" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "ing methods [19]. In the scenarios of deep learning, there are some typical methods. Class distance weighted loss relies on a coefficient for the loss of each class, which utilizes the distance to the ground-truth class and increases in relation to that distance [49]. The Earth Mover's Distance (EMD) loss uses the predicted probabilities of all classes and penalizes the miss-predictions according to a ground distance matrix that quantifies the dissimilarities between classes [9, 20]. Soft labeling converts ground truth data labels into soft probability distributions and the loss is computed on them [6, 15, 36, 65]. However, these methods have only shown success on simple multi category classification problem. For LLMs, the probability space of the output has tens to hundreds of thousands of tokens and they are mixture of numerical and non-numerical. The problem becomes difficulty. In this work, we will show how to apply ordinal regression to LLMs." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 295, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 295, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 295, + 479 + ], + "type": "text", + "content": "Soft labeling uses probability distributions over classes as labels instead of hard, one-hot labels. A typical type is label smoothing which has been found more robust to noisy annotations for classification problem [39, 61, 71] and has been applied to many fields such as know distillation [44, 73], network calibration [30], and some applications like image segmentation [66], text mining [34]. On the other hand, soft loss terms have been useful for domain and task transfer [64] in order to avoid dataset biases. Elaborate loss functions are defined in [68] to take into account the subjective scenicness of outdoor pictures, by trying to predict the same rating distribution of human annotations. Age estimation is another particular niche where soft labels have become popular [15, 58]. However, there is few research of applying soft labeling on numerical prediction of LLMs. In this paper, we will show that the label smoothing is just a specific format of our soft labeling and our proposal performs better." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 111, + 502 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "In traditional regression models, numerical prediction is typically handled using L1/L2-style loss functions, which penalize errors based on their distance to the target. Further distance incurs greater penalties and vice versa, encouraging the model to make predictions that are closer to the target. However, when training LLMs/MLLMs, the de facto cross-entropy loss with one-hot encoded target fails to account for the distance between predicted and target tokens properly. It does so in a limited and imprecise manner, a misclassification is a misclassification, regardless of how explicitly far the predicted token is from the target token. This limitation hinders the model's ability to predict accurate numerical values. To address this, a natural idea is to apply a regression-style loss to LLMs. However, in the context of LLMs, it is not trivial to directly convert the LLM's output into a scalar without introducing additional layers or heuristics. For this reason, we opt for a new loss de" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "sign which can be directly applied to the intrinsic output of LLM, similar to the standard cross-entropy loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": "We begin by reviewing the LLM training loss. LLMs typically generate text token by token, treating the next-token prediction as a classification problem over a vocabulary of size " + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": ". Given a model output logit " + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{z} \\in R^{V}" + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": " from the last layer before the final prediction, the softmax function is applied to convert these logits into probability distribution " + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{p} \\in R^{V}" + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\sum_{i=1}^{V} p_{i} = 1" + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": ". This distribution is then compared to the target token using cross-entropy loss. Typically, the target token is encoded into a one-hot vector " + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "inline_equation", + "content": "\\mathbf{q} \\in R^{V}" + }, + { + "bbox": [ + 313, + 96, + 553, + 240 + ], + "type": "text", + "content": " where it is a binary vector consisting of a value 1 at the correct token index, and 0s everywhere else. The cross-entropy loss is computed by" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 380, + 247, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 247, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 380, + 247, + 553, + 281 + ], + "type": "interline_equation", + "content": "L (\\mathbf {p}, \\mathbf {q}) = - \\sum_ {i = 1} ^ {V} q _ {i} \\log p _ {i}. \\tag {1}", + "image_path": "a1412ae93bbc54ecf2da9bbddb852c71893665d4b48a39e61bfe48b133126792.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 288, + 531, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 531, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 531, + 300 + ], + "type": "text", + "content": "Since " + }, + { + "bbox": [ + 313, + 288, + 531, + 300 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 313, + 288, + 531, + 300 + ], + "type": "text", + "content": " is a one-hot vector, the final loss simplifies to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 392, + 309, + 553, + 323 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 309, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 392, + 309, + 553, + 323 + ], + "type": "interline_equation", + "content": "L (\\mathbf {p}, \\mathbf {q}) = - \\log p _ {t}, \\tag {2}", + "image_path": "a40f37a74fd761989cb0fb0b03bd05fa5cc6bb70162714ab2b52bb045fc0f582.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "text", + "content": " is the index of the target token. As shown, the loss function depends solely on the posterior probability " + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "text", + "content": " of the target token. This ignores other rich information inherited in the full probability distribution " + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "inline_equation", + "content": "\\mathbf{p}" + }, + { + "bbox": [ + 313, + 331, + 553, + 378 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "text", + "content": "In fact, each probability " + }, + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "text", + "content": " reflects the model confidence on each possible token. When a prediction is incorrect, it is crucial to have a metric that precisely measures how \"wrong\" the prediction is by explicitly considering the model's confidence on all tokens, not just the target token. This becomes important when the target token is numerical and the numerical prediction is done by LLMs. Numerical tokens have relative distance between each other by nature, and when a prediction is incorrect, we should measure how wrong it is by assessing the distance from the target, rather than treating all wrong predictions equally as long as the confidence on the target token is the same, as is done in hard labeling. For example, in the scenario illustrated in Figure 1 to answer how many trains given an image, when we train an MLLM with cross-entropy loss and one-hot encoding, there are two possible incorrect predictions, A and B, during MLLM training. If we apply hard labeling to encode the supervision \"4\", both predictions incur the same cross-entropy loss, as both have a softmax probability of 0.3 for the target token \"4\". This approach fails to account for numerical proximity since prediction B, which assigns a maximum probability of 0.4 to the incorrect token \"3\", is closer to the correct answer \"4\" than prediction A, which assigns a maximum probability of 0.4 to the incorrect token \"1\". Intuitively, prediction B should receive a lower loss from the regression perspective. This limitation arises from the one-hot encoding in hard labeling which only retains the probability " + }, + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "inline_equation", + "content": "p_t" + }, + { + "bbox": [ + 313, + 379, + 553, + 715 + ], + "type": "text", + "content": " of the target token in the final loss computation" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3426" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "of (2). As a result, the final loss function does not consider the probabilities " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " for other numerical tokens, which contain valuable information about inter-token relationship and numerical distance. The absence of this information makes the hard labeling loss suboptimal, and we need a new loss function that can better account for such relationships." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 145, + 294, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 145, + 294, + 181 + ], + "spans": [ + { + "bbox": [ + 55, + 145, + 294, + 181 + ], + "type": "text", + "content": "In order to address this issue, we propose soft labeling cross-entropy loss. We first reformulate the cross-entropy loss of (1) into a more general form," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 100, + 191, + 295, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 191, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 100, + 191, + 295, + 223 + ], + "type": "interline_equation", + "content": "L _ {C E} (\\mathbf {p}, \\mathbf {q} (t)) = \\sum_ {i = 1} ^ {V} q _ {i} (t) (- \\log p _ {i}), \\tag {3}", + "image_path": "9bae33935e2b064e8f1b6cae77bc00944df9a2a576587f8136e1b62dcd4744df.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(t)" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " is the distribution in which the target token with index of " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " is encoded, and " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "q_{i}(t) \\in [0,1]" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " s.t. " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\sum_{i} q_{i}(t) = 1" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": ". In the standard one-hot encoding for cross-entropy, " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(t) = \\delta(t)" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " is a Dirac delta distribution where " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "q_{i}(t)" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " equals to 1 for " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "i = t" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": ", and 0 otherwise. In this case, (3) is simplified to (2), which corresponds to hard labeling. In fact, " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " is not necessarily a one-hot distribution, but can be any distribution. The primary objective of " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " is to guide the learning of the LLM model by minimizing the distance between two distributions. Instead of forcing the model to learn from a strict, hard label with harsh penalization, soft labeling relaxes this constraint by using a smooth distribution " + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 55, + 234, + 295, + 377 + ], + "type": "text", + "content": " over all tokens." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 378, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 295, + 521 + ], + "type": "text", + "content": "In this work, we focus on numerical prediction, where numerical values are typically tokenized digit-wise [26]. That is, each digit (0-9) is treated as an individual token, and multi-digit numbers are split accordingly. For example, in the LLaMA tokenizer, number \"123\" is tokenized into the token IDs \"29896\", \"29906\", and \"29941\". Therefore, we restrict our soft labeling discussion to the 10 digit tokens from \"0\" to \"9\", leaving the soft labeling of regular tokens for future work. For instance, we do not consider the soft labeling for text tokens such as \"two\", \"eight\", etc. Following [6, 65], we defines soft label for the 10 digit tokens from \"0\" to \"9\" as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 110, + 532, + 294, + 546 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 532, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 110, + 532, + 294, + 546 + ], + "type": "interline_equation", + "content": "\\mathbf {q} ^ {S L} (t) = (1 - \\eta) \\boldsymbol {\\delta} (t) + \\eta \\psi (t) \\tag {4}", + "image_path": "ce38c7d251d2c8451f1c760c8f62995c99ab55cde4bcb93cd900745b7479f027.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is the index of the 10 digit tokens " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\{ \"0\", \"1\", ..., \"9\" \\}" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": ". For simplicity, we assume that these indices are consecutive and form a set " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "; if not, a straightforward mapping can be constructed to make them consecutive. " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{q}^{SL}(t)" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is a soft distribution to represent the token with index of " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is a probability distribution that has a support set on " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and typically centers at the token with index of " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": ". It can be Poisson, Binomial, Triangular, Uniform distributions, etc. " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\eta \\in [0,1]" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " is a fixed hyperparameter that controls the mixture ratio of " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\delta(t)" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "\\psi(t)" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "By smoothing the one-hot " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}(t)" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " of hard labeling, soft labeling allows " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "q_{i}^{SL}(t)" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " to be greater than 0 for " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "i\\neq t" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " and to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": "vary for different " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": ". This enables the assessment of relationship between the target token and other digit tokens. The relationship reflects the similarity between token indexed by " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": " and the token indexed by " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "q_{i}^{SL}(t)" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": " being higher when " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": " is closer to " + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 553, + 131 + ], + "type": "text", + "content": " and lower otherwise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 133, + 500, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 133, + 500, + 144 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 500, + 144 + ], + "type": "text", + "content": "The final loss for each example is given by" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 150, + 553, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 150, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 330, + 150, + 553, + 228 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L = \\frac {1}{N _ {r} + N _ {n}} \\left(\\sum_ {\\left(x _ {r}, y _ {r}\\right) \\in \\left(\\mathbf {x} _ {r}, \\mathbf {y} _ {r}\\right)} L _ {C E} \\left(f \\left(x _ {r}\\right), \\mathbf {q} \\left(y _ {r}\\right)\\right) \\right. \\tag {5} \\\\ \\left. + \\lambda \\sum_ {\\left(x _ {n}, y _ {n}\\right) \\in \\left(\\mathbf {x} _ {n}, \\mathbf {y} _ {n}\\right)} L _ {C E} \\left(f \\left(x _ {n}\\right), \\mathbf {q} ^ {S L} \\left(y _ {n}\\right)\\right)\\right), \\\\ \\end{array}", + "image_path": "f50204a7de7cab57638e2112ab0bc4dd1503eda88a4d9f68eb150a5f6a4163e0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "N_{r}, N_{n}" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " denote the numbers of regular tokens and numerical tokens within a sequence example, respectively. " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_n)" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_r" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "(\\mathbf{y}_n)" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " represent the sets of regular (numerical) tokens where " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_r" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "(\\mathbf{x}_n)" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " is the input set and " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_r" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "(\\mathbf{y}_n)" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " is the index set of ground truth. " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " is the LLM with softmax output for each token prediction. " + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 235, + 553, + 330 + ], + "type": "text", + "content": " controls the balance between the numerical token loss and the regular token loss in the final loss computation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 331, + 553, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 617 + ], + "type": "text", + "content": "When the target is a digit token, compared to hard labeling (one-hot encoding) which solely rely on the probability of target token in the loss function, as indicated in (2), soft labeling incorporates the prediction probabilities of all incorrect digit tokens into loss computation. This allows the loss value to reflect the distance between the predicted token and target token. It therefore can be understood as an implicit way of introducing distance-based penalties into the classification loss, similar to how loss functions operate in regression. As illustrated in Figure 1, an incorrect prediction of \"1\" (Prediction A) has a greater distributional distance from the soft ground truth than an incorrect prediction of \"3\" (Prediction B). Consequently, Prediction A incurs a larger loss than Prediction B, aligning with the behavior of regression losses. This property arises naturally from the inter-digit token similarity encoded in soft labels where " + }, + { + "bbox": [ + 313, + 331, + 553, + 617 + ], + "type": "inline_equation", + "content": "q_{i}^{SL}(t)" + }, + { + "bbox": [ + 313, + 331, + 553, + 617 + ], + "type": "text", + "content": " reflects the similarity ranking to the ground truth token. The closer a digit token is to the ground truth, the higher its value, enabling the model to account for the severity of prediction errors rather than treating all mistakes equally. In contrast, under hard labeling, predictions A and B receive the same loss due to the underlying assumption of one-hot encoding that all incorrect tokens are equally dissimilar to the ground truth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": "From another perspective, in (3), " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "q_{i}(t)" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": " can be viewed as a weight for each component loss " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "-\\log p_{i}" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": ". During training, a soft label with " + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "inline_equation", + "content": "q_{i}^{SL}(t) > 0" + }, + { + "bbox": [ + 313, + 618, + 554, + 713 + ], + "type": "text", + "content": " encourages the model to predict higher probabilities on tokens closer to the target token, and lower probabilities on further tokens. This also aligns with the goal of regression-style loss, which aims to make predictions as close as possible to the target in continuous space." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3427" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 90, + 553, + 194 + ], + "blocks": [ + { + "bbox": [ + 189, + 71, + 421, + 82 + ], + "lines": [ + { + "bbox": [ + 189, + 71, + 421, + 82 + ], + "spans": [ + { + "bbox": [ + 189, + 71, + 421, + 82 + ], + "type": "text", + "content": "Table 1. The enhancement by soft labeling on visual grounding." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 90, + 553, + 194 + ], + "lines": [ + { + "bbox": [ + 58, + 90, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 58, + 90, + 553, + 194 + ], + "type": "table", + "html": "
Pretrain modelModel typeFinetune dataLossRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
LLaVA-7BspecialistRefCOCO+/ghard soft81.787.374.473.682.063.379.078.8
85.0(+3.3)89.3(+2.0)78.8(+4.4)76.7(+3.1)83.4(+1.4)67.0(+3.7)82.0(+3.0)81.8(+3.0)
generalistLLaVA-Mix + RefCOCO+/ghard soft72.680.262.666.676.155.072.166.6
77.8(+5.2)83.7(+3.5)68.8(+6.2)72.0(+5.4)80.2(+4.1)60.4(+5.4)78.5(+6.4)71.5(+4.9)
LLaVA-13BspecialistRefCOCO+/ghard soft88.991.784.982.588.175.186.086.7
90.0(+1.1)92.7(+1.0)85.3(+0.4)83.5(+1.0)89.3(+1.2)76.7(+1.6)86.7(+0.7)87.4(+0.7)
generalistLLaVA-Mix + RefCOCO+/ghard soft79.385.071.474.782.964.878.073.0
81.0(+1.7)86.4(+1.4)72.6(+1.2)76.4(+1.7)83.3(+0.4)66.1(+1.3)79.2(+1.2)74.3(+1.3)
", + "image_path": "4992bd7474f23a003be48683b3a76729d58f3af7197753e4dacd547376a16694.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 62, + 208, + 179, + 289 + ], + "blocks": [ + { + "bbox": [ + 62, + 208, + 179, + 289 + ], + "lines": [ + { + "bbox": [ + 62, + 208, + 179, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 208, + 179, + 289 + ], + "type": "image", + "image_path": "afe2d20f7150c3212449e505288003a4d15a0f5e840b621be6453155900b719f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 295, + 180, + 312 + ], + "lines": [ + { + "bbox": [ + 60, + 295, + 180, + 312 + ], + "spans": [ + { + "bbox": [ + 60, + 295, + 180, + 312 + ], + "type": "text", + "content": "(a) RefCOCOg Acc. v.s. " + }, + { + "bbox": [ + 60, + 295, + 180, + 312 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 60, + 295, + 180, + 312 + ], + "type": "text", + "content": " for Binomial distrib." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 186, + 208, + 302, + 290 + ], + "blocks": [ + { + "bbox": [ + 186, + 208, + 302, + 290 + ], + "lines": [ + { + "bbox": [ + 186, + 208, + 302, + 290 + ], + "spans": [ + { + "bbox": [ + 186, + 208, + 302, + 290 + ], + "type": "image", + "image_path": "cb87aa59e99909317c5b29be7962ef69cf25869b65d34c15ee15164bce7775c4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 295, + 303, + 312 + ], + "lines": [ + { + "bbox": [ + 184, + 295, + 303, + 312 + ], + "spans": [ + { + "bbox": [ + 184, + 295, + 303, + 312 + ], + "type": "text", + "content": "(b) RefCOCOg Acc. v.s. " + }, + { + "bbox": [ + 184, + 295, + 303, + 312 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 184, + 295, + 303, + 312 + ], + "type": "text", + "content": " for Poisson distrib." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 308, + 208, + 425, + 290 + ], + "blocks": [ + { + "bbox": [ + 308, + 208, + 425, + 290 + ], + "lines": [ + { + "bbox": [ + 308, + 208, + 425, + 290 + ], + "spans": [ + { + "bbox": [ + 308, + 208, + 425, + 290 + ], + "type": "image", + "image_path": "7b9876d45f43a5e58b733476308b46a7e43d137c3a2d67c8845335501698dcfe.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 295, + 425, + 313 + ], + "lines": [ + { + "bbox": [ + 306, + 295, + 425, + 313 + ], + "spans": [ + { + "bbox": [ + 306, + 295, + 425, + 313 + ], + "type": "text", + "content": "(c) RefCOCOg Acc. v.s. " + }, + { + "bbox": [ + 306, + 295, + 425, + 313 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 306, + 295, + 425, + 313 + ], + "type": "text", + "content": " for Triangular distrib." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 431, + 208, + 547, + 290 + ], + "blocks": [ + { + "bbox": [ + 431, + 208, + 547, + 290 + ], + "lines": [ + { + "bbox": [ + 431, + 208, + 547, + 290 + ], + "spans": [ + { + "bbox": [ + 431, + 208, + 547, + 290 + ], + "type": "image", + "image_path": "5c846cfc78de4dee874d63a472b664aad9b8a62e6cc59bf870e9cb6b0c71cafc.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 295, + 549, + 312 + ], + "lines": [ + { + "bbox": [ + 429, + 295, + 549, + 312 + ], + "spans": [ + { + "bbox": [ + 429, + 295, + 549, + 312 + ], + "type": "text", + "content": "(d) RefCOCOg Acc. v.s. " + }, + { + "bbox": [ + 429, + 295, + 549, + 312 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 429, + 295, + 549, + 312 + ], + "type": "text", + "content": " for Uniform distrib." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 62, + 316, + 179, + 397 + ], + "blocks": [ + { + "bbox": [ + 62, + 316, + 179, + 397 + ], + "lines": [ + { + "bbox": [ + 62, + 316, + 179, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 316, + 179, + 397 + ], + "type": "image", + "image_path": "acad06ac982ac8545e94a6d063031aef1fee360b170dfb0e9513b11ab4284383.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 402, + 181, + 419 + ], + "lines": [ + { + "bbox": [ + 60, + 402, + 181, + 419 + ], + "spans": [ + { + "bbox": [ + 60, + 402, + 181, + 419 + ], + "type": "text", + "content": "(e) ChartQA Acc. v.s. " + }, + { + "bbox": [ + 60, + 402, + 181, + 419 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 60, + 402, + 181, + 419 + ], + "type": "text", + "content": " for Binomial distrib." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 186, + 316, + 302, + 397 + ], + "blocks": [ + { + "bbox": [ + 186, + 316, + 302, + 397 + ], + "lines": [ + { + "bbox": [ + 186, + 316, + 302, + 397 + ], + "spans": [ + { + "bbox": [ + 186, + 316, + 302, + 397 + ], + "type": "image", + "image_path": "5216aa1e2cefcdd523677e4c45a6cef60511a57dcaa16474207bf498b98a5c68.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 401, + 303, + 419 + ], + "lines": [ + { + "bbox": [ + 184, + 401, + 303, + 419 + ], + "spans": [ + { + "bbox": [ + 184, + 401, + 303, + 419 + ], + "type": "text", + "content": "(f) ChartQA Acc. v.s. " + }, + { + "bbox": [ + 184, + 401, + 303, + 419 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 184, + 401, + 303, + 419 + ], + "type": "text", + "content": " for Poisson distrib." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 309, + 316, + 425, + 397 + ], + "blocks": [ + { + "bbox": [ + 309, + 316, + 425, + 397 + ], + "lines": [ + { + "bbox": [ + 309, + 316, + 425, + 397 + ], + "spans": [ + { + "bbox": [ + 309, + 316, + 425, + 397 + ], + "type": "image", + "image_path": "9b99053548e53eb56fd22f72148891690c91c5c8b9bdcac754e9241f1686f5ee.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 402, + 425, + 419 + ], + "lines": [ + { + "bbox": [ + 306, + 402, + 425, + 419 + ], + "spans": [ + { + "bbox": [ + 306, + 402, + 425, + 419 + ], + "type": "text", + "content": "(g) ChartQA Acc. v.s. " + }, + { + "bbox": [ + 306, + 402, + 425, + 419 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 306, + 402, + 425, + 419 + ], + "type": "text", + "content": " for Triangular distrib." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 431, + 316, + 547, + 397 + ], + "blocks": [ + { + "bbox": [ + 431, + 316, + 547, + 397 + ], + "lines": [ + { + "bbox": [ + 431, + 316, + 547, + 397 + ], + "spans": [ + { + "bbox": [ + 431, + 316, + 547, + 397 + ], + "type": "image", + "image_path": "d12edcd12dc0b3f9fbd1eacbf6b920a3a45f3f9ffc13fccc873ccb8f3ac91032.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 429, + 402, + 549, + 419 + ], + "lines": [ + { + "bbox": [ + 429, + 402, + 549, + 419 + ], + "spans": [ + { + "bbox": [ + 429, + 402, + 549, + 419 + ], + "type": "text", + "content": "(h) ChartQA Acc. v.s. " + }, + { + "bbox": [ + 429, + 402, + 549, + 419 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 429, + 402, + 549, + 419 + ], + "type": "text", + "content": " for Uniform distrib." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 62, + 422, + 179, + 504 + ], + "blocks": [ + { + "bbox": [ + 62, + 422, + 179, + 504 + ], + "lines": [ + { + "bbox": [ + 62, + 422, + 179, + 504 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 179, + 504 + ], + "type": "image", + "image_path": "cb29fc0a698c3888e7fccfc687f61548ee71b517506774f809f5a04f38657d4c.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 60, + 508, + 180, + 526 + ], + "lines": [ + { + "bbox": [ + 60, + 508, + 180, + 526 + ], + "spans": [ + { + "bbox": [ + 60, + 508, + 180, + 526 + ], + "type": "text", + "content": "(i) TallyQA RMSE v.s. " + }, + { + "bbox": [ + 60, + 508, + 180, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 60, + 508, + 180, + 526 + ], + "type": "text", + "content": " for Binomial distrib." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 186, + 422, + 302, + 504 + ], + "blocks": [ + { + "bbox": [ + 186, + 422, + 302, + 504 + ], + "lines": [ + { + "bbox": [ + 186, + 422, + 302, + 504 + ], + "spans": [ + { + "bbox": [ + 186, + 422, + 302, + 504 + ], + "type": "image", + "image_path": "f2a48044bde08e09ff808b886ed7e7ac00b34adc927a2c676e69c4b9cca44b19.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 508, + 303, + 526 + ], + "lines": [ + { + "bbox": [ + 184, + 508, + 303, + 526 + ], + "spans": [ + { + "bbox": [ + 184, + 508, + 303, + 526 + ], + "type": "text", + "content": "(j) TallyQA RMSE v.s. " + }, + { + "bbox": [ + 184, + 508, + 303, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 184, + 508, + 303, + 526 + ], + "type": "text", + "content": " for Poisson distrib." + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 309, + 422, + 425, + 504 + ], + "blocks": [ + { + "bbox": [ + 309, + 422, + 425, + 504 + ], + "lines": [ + { + "bbox": [ + 309, + 422, + 425, + 504 + ], + "spans": [ + { + "bbox": [ + 309, + 422, + 425, + 504 + ], + "type": "image", + "image_path": "9c1ee120d8c086ea8b26c24295a6e246ac04227f8c6a191bac78bf9a00b070ef.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 529, + 553, + 551 + ], + "lines": [ + { + "bbox": [ + 55, + 529, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 553, + 551 + ], + "type": "text", + "content": "Figure 2. Ablation study of " + }, + { + "bbox": [ + 55, + 529, + 553, + 551 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 529, + 553, + 551 + ], + "type": "text", + "content": " of different distributions across three datasets, RefCOCOg val, ChartQA and TallyQA. On ChartQA, we evaluate on numerical subset. On TallyQA, we use the RMSE as metric and evaluate on simple balanced set." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 431, + 422, + 547, + 504 + ], + "blocks": [ + { + "bbox": [ + 431, + 422, + 547, + 504 + ], + "lines": [ + { + "bbox": [ + 431, + 422, + 547, + 504 + ], + "spans": [ + { + "bbox": [ + 431, + 422, + 547, + 504 + ], + "type": "image", + "image_path": "f33decacc9922b9eb88176a7ad4354eae961a544a9ddcd887c4c8cd3903f633f.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 306, + 508, + 425, + 526 + ], + "lines": [ + { + "bbox": [ + 306, + 508, + 425, + 526 + ], + "spans": [ + { + "bbox": [ + 306, + 508, + 425, + 526 + ], + "type": "text", + "content": "(k) TallyQA RMSE v.s. " + }, + { + "bbox": [ + 306, + 508, + 425, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 306, + 508, + 425, + 526 + ], + "type": "text", + "content": " for Triangular distrib." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 429, + 508, + 549, + 526 + ], + "lines": [ + { + "bbox": [ + 429, + 508, + 549, + 526 + ], + "spans": [ + { + "bbox": [ + 429, + 508, + 549, + 526 + ], + "type": "text", + "content": "(1) TallyQA RMSE v.s. " + }, + { + "bbox": [ + 429, + 508, + 549, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 429, + 508, + 549, + 526 + ], + "type": "text", + "content": " for Uniform distrib." + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 56, + 574, + 301, + 679 + ], + "blocks": [ + { + "bbox": [ + 55, + 554, + 294, + 566 + ], + "lines": [ + { + "bbox": [ + 55, + 554, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 294, + 566 + ], + "type": "text", + "content": "Table 2. The enhancement by soft labeling on chart understanding." + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 574, + 301, + 679 + ], + "lines": [ + { + "bbox": [ + 56, + 574, + 301, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 574, + 301, + 679 + ], + "type": "table", + "html": "
Pretrain modelModel typeFinetune dataLossChartQA
AllNumerical
LLaVA-7BspecialistChartQAhard soft67.966.3
69.6(+1.7)68.6(+2.3)
generalistLLaVA-Mix + ChartQAhard soft67.066.5
68.3(+1.3)68.6(+2.1)
LLaVA-13BspecialistChartQAhard soft72.872.0
73.4(+0.6)73.6(+1.6)
generalistLLaVA-Mix + ChartQAhard soft71.069.7
72.9 (+1.9)72.6(+2.9)
", + "image_path": "9b4a5962fef2ebbc4f996bd5ee4ca845b86088521aa3e67bae1c52d6ab82ae6f.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "table_body" + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 555, + 394, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 555, + 394, + 568 + ], + "spans": [ + { + "bbox": [ + 314, + 555, + 394, + 568 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "In this work, we evaluate our approach on three representative numerical prediction tasks: visual grounding, object counting, and chart understanding. For visual grounding, we assess performance on three benchmarks RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), and RefCOCOg (val/test) [25] using the standard Acc@0.5 metric. For object counting, we evaluate on TallyQA's simple and complex test sets [1]. However, the original test sets are count-imbalanced, with ground truth counts ranging from 0 to 15 but the combined counts of 1 and 2 account for " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " of the simple testing set and " + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "inline_equation", + "content": "60\\%" + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": " of the complex testing" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3428" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 89, + 553, + 196 + ], + "blocks": [ + { + "bbox": [ + 192, + 71, + 417, + 82 + ], + "lines": [ + { + "bbox": [ + 192, + 71, + 417, + 82 + ], + "spans": [ + { + "bbox": [ + 192, + 71, + 417, + 82 + ], + "type": "text", + "content": "Table 3. The enhancement by soft labeling on object counting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 89, + 553, + 196 + ], + "lines": [ + { + "bbox": [ + 56, + 89, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 89, + 553, + 196 + ], + "type": "table", + "html": "
Pretrain modelModel typeFinetune dataLossTallyQA simpleTallyQA complexTallyQA simple balancedTallyQA complex balanced
Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)Acc. (↑)RMSE(↓)
LLaVA-7BspecialistTallyQAhard78.51.0066.91.3038.53.8416.76.68
soft79.0(+0.5)0.97(-0.03)67.2(+0.3)1.29(-0.01)39.4(+0.9)3.20(-0.64)18.8(+2.1)6.32(-0.36)
generalistLLaVA-Mix + TallyQAhard79.81.4368.22.5439.42.3321.84.96
soft80.5(+0.7)1.35(-0.08)69.4(+1.2)1.28(-1.26)42.2(+2.8)2.24(-0.09)24.0(+2.2)4.41(-0.55)
LLaVA-13BspecialistTallyQAhard81.90.8567.21.2740.42.8422.95.98
soft82.7(+0.8)0.67(-0.18)69.3(+2.1)1.20(-0.07)42.2(+1.8)2.02(-0.82)27.0(+4.1)5.16(-0.82)
generalistLLaVA-Mix + TallyQAhard82.00.8170.91.3045.92.3723.94.87
soft82.5(+0.5)0.76(-0.05)71.5(+0.6)1.24(-0.06)47.8(+1.9)1.88(-0.49)29.1(+5.2)4.45(-0.42)
", + "image_path": "f96e4c68b8fd4b17723e9096cc5efaecac8863f42e1fbfeeae8459f6efbfb22e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 209, + 208, + 312 + ], + "blocks": [ + { + "bbox": [ + 59, + 209, + 208, + 312 + ], + "lines": [ + { + "bbox": [ + 59, + 209, + 208, + 312 + ], + "spans": [ + { + "bbox": [ + 59, + 209, + 208, + 312 + ], + "type": "image", + "image_path": "7c51225a2da420922bd021d28afa931d794101c8bfb16079bcc5f66a94e189b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 101, + 316, + 162, + 326 + ], + "lines": [ + { + "bbox": [ + 101, + 316, + 162, + 326 + ], + "spans": [ + { + "bbox": [ + 101, + 316, + 162, + 326 + ], + "type": "text", + "content": "(a) RefCOCOg val" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 208, + 380, + 312 + ], + "blocks": [ + { + "bbox": [ + 230, + 208, + 380, + 312 + ], + "lines": [ + { + "bbox": [ + 230, + 208, + 380, + 312 + ], + "spans": [ + { + "bbox": [ + 230, + 208, + 380, + 312 + ], + "type": "image", + "image_path": "eaa663f1a38b69465a29722f887c7e07dbd464a9ce8b0da6bee646e0c690ba95.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 316, + 354, + 325 + ], + "lines": [ + { + "bbox": [ + 255, + 316, + 354, + 325 + ], + "spans": [ + { + "bbox": [ + 255, + 316, + 354, + 325 + ], + "type": "text", + "content": "(b) ChartQA numerical subset" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 404, + 209, + 552, + 311 + ], + "blocks": [ + { + "bbox": [ + 404, + 209, + 552, + 311 + ], + "lines": [ + { + "bbox": [ + 404, + 209, + 552, + 311 + ], + "spans": [ + { + "bbox": [ + 404, + 209, + 552, + 311 + ], + "type": "image", + "image_path": "40ba5169c74e5d603d0299885b79fc7dcb7a0db8afccf15630881fdcb9f6c837.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 316, + 525, + 326 + ], + "lines": [ + { + "bbox": [ + 432, + 316, + 525, + 326 + ], + "spans": [ + { + "bbox": [ + 432, + 316, + 525, + 326 + ], + "type": "text", + "content": "(c) TallyQA simple balanced" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 327, + 429, + 338 + ], + "lines": [ + { + "bbox": [ + 179, + 327, + 429, + 338 + ], + "spans": [ + { + "bbox": [ + 179, + 327, + 429, + 338 + ], + "type": "text", + "content": "Figure 3. Ablation study of " + }, + { + "bbox": [ + 179, + 327, + 429, + 338 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 179, + 327, + 429, + 338 + ], + "type": "text", + "content": " on three numerical prediction datasets." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 66, + 371, + 287, + 458 + ], + "blocks": [ + { + "bbox": [ + 55, + 341, + 295, + 363 + ], + "lines": [ + { + "bbox": [ + 55, + 341, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 295, + 363 + ], + "type": "text", + "content": "Table 4. The effect of different distributions. On TallyQA, RMSE is the metric." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 66, + 371, + 287, + 458 + ], + "lines": [ + { + "bbox": [ + 66, + 371, + 287, + 458 + ], + "spans": [ + { + "bbox": [ + 66, + 371, + 287, + 458 + ], + "type": "table", + "html": "
DistributionRefCOCO valTallyQA balancedChartQA Numerical
**+*gSimpleComplex
Hard Label81.773.679.03.846.6866.3
Binomial84.776.881.93.136.2867.6
Posisson84.176.380.83.706.4967.0
Triangular85.076.782.03.206.3268.6
Uniform83.076.680.83.746.5666.8
", + "image_path": "09980095aa9fa6d9112082c63a04e899f876e32e813f98056e01e01d4681b968.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 471, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 471, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 54, + 471, + 295, + 614 + ], + "type": "text", + "content": "set. In addition, the testing sets also contain noisy labels. These make the evaluation results misleading. To address this, we also construct a balanced subset for reference by sampling a count-balanced simple and complex testing sets from TallyQA's test sets respectively, referred to as \"TallyQA simple balanced\" and \"TallyQA complex balanced\". The details are in the Supp. Accuracy and RMSE are the chosen metrics following [1]. For chart understanding, we use ChartQA benchmark [40], where approximately half of the questions have numerical answers. We report accuracy on both the full test set and a subset containing only numerical answers." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Following [4], we adopt a three-stage pipeline, where the first stage is to learn the vision-language (VL) adapter, the second stage is to pretrain LLM and VL adapter on a large set of pretraining data, and the third stage is to finetune the pretrained model on some small scale finetuning data. To evaluate the broad applicability of our approach, we conduct experiments on various pretraining and finetuning settings. For the pretraining of generalist models," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 606 + ], + "type": "text", + "content": "we use the publicly available LLaVA-7B architecture [32], pretrained on a sampled LAION-5M dataset [54], as our base model. For a stronger baseline, we use LLaVA-13B, pretrained on a 16M data mixture including LAION-5M for image captioning, GRIT-5M [47] for grounding, Objects365 [55] and OpenImages [27] of 3M in total for object counting, and 3M plot or chart related datasets such as PlotQA [43], UniChart [41], FigureQA [24] for chart understanding. For specialist model pretraining, we use three domain-specific data sources: a 12M mixture of grounding and detection data for visual grounding, a 6M counting data mixture with multiple tasks for object counting, and a 20M chart analysis data mixture for chart understanding, respectively. More details for pretraining data are given in Supp. After pretraining, the specialist models are finetuned on the task-specific training sets: RefCOCO, RefCOCO+, and RefCOCOg [25] for visual grounding; TallyQA [1] for object counting; and ChartQA [40] for chart understanding, and generalist models are finetuned on the same task-specific training sets but with the additional public 665K LLaVA-Mix tuning dataset [28]. The other settings are exactly the same as LLaVA's [32]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 612, + 497, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 497, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 497, + 625 + ], + "type": "text", + "content": "4.1. The Effectiveness of Soft Labeling" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 715 + ], + "type": "text", + "content": "We first compare soft labeling and one-hot hard labeling across various pretrained models, including LLaVA-7B (pretrained on LAION-5M) and LLaVA-13B (pretrained on a 16M multi-task mixture). They are fine-tuned as specialist and generalist models and evaluated on multiple tasks and benchmarks. Tables 1, 2 and 3 report the performance comparisons for visual grounding, chart understanding and" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3429" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 102, + 309, + 286 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 297, + 94 + ], + "type": "text", + "content": "Table 5. Performance comparison to the state-of-the-art models on visual grounding (Acc@0.5)." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 102, + 309, + 286 + ], + "lines": [ + { + "bbox": [ + 56, + 102, + 309, + 286 + ], + "spans": [ + { + "bbox": [ + 56, + 102, + 309, + 286 + ], + "type": "table", + "html": "
ModelsRefCOCORefCOCO+RefCOCOg
valtestAtestBvaltestAtestBvaltest
Shikra-7B [12]87.090.680.281.687.472.182.382.2
Ferret-7B [70]87.591.482.580.887.473.183.984.8
Qwen-VL-7B [4]88.692.384.582.888.676.886.086.3
InternVL2-8B [13]87.191.180.779.887.971.482.782.7
VistaLLM [50]88.191.583.082.989.874.883.684.4
MiniGPT-v2 [11]88.791.785.380.085.174.584.484.7
Qwen2-VL-7B [4]91.793.687.385.890.579.587.387.8
Ferret-v2-7B [72]92.894.788.787.492.879.389.489.3
SoftLabeling-7B91.894.788.987.092.780.089.689.5
LION-12B [10]89.893.085.684.089.278.185.585.7
Shikra-13B [12]87.891.181.882.987.874.482.683.2
Ferret-13B [70]89.592.484.482.888.175.285.886.3
Ferret-v2-13B [72]92.695.088.987.492.181.489.490.0
SoftLabeling-13B92.795.089.087.692.782.389.890.0
", + "image_path": "18c3056fa356b374cb5f156d38ecc83407d31bd5b3717607ba1193c42d4060ee.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 297, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 297, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 297, + 552 + ], + "type": "text", + "content": "object counting tasks, respectively. Some observations are available. First, soft labeling consistently improves performance across all model and benchmark combinations, with most gains larger than 2 points, demonstrating its effectiveness and robustness. Second, the absolute gains are smaller for the stronger 13B model compared to the 7B model. This is expected due to performance saturation. Nonetheless, notable improvements are remained, e.g., a 2.9 points gain in accuracy on ChartQA numerical subset (generalist model) in Table 2 and 4.1/5.2 points accuracy gains on TallyQA's simple/complex balanced test sets in Table 3. Third, on ChartQA, the gain on numerical subset is higher than the full set. This aligns with the design of soft labeling which is proposed specifically for enhancing the numerical prediction ability. Finally, on object counting, gains are more pronounced on the more reasonable balanced test sets, particularly in RMSE, a preferable metric reflecting error distance to the ground truth, where the relative improvements exceed " + }, + { + "bbox": [ + 55, + 300, + 297, + 552 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 300, + 297, + 552 + ], + "type": "text", + "content": " in most cases. Notably, while enhancing numerical prediction, soft labeling does not degrade performance on generic tasks (please see details in Supp.)" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 149, + 575 + ], + "type": "text", + "content": "4.2. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": "Soft Distribution The distribution " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " in (4) is crucial because it determines the degree of softness in soft labeling and the measurement of inter-digital token similarity. We investigate four widely used distributions, Binomial [6], Poisson [6], Triangular [65] and Uniform (label smoothing) [57]. The optimal " + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 582, + 297, + 715 + ], + "type": "text", + "content": " in (4) is exhaustively searched for each distribution and used for comparison. We assess performance using some representative benchmarks and metrics on the base LLaVA-7B specialist model, as reported in Table 4. Regardless of the chosen distribution, soft labeling consistently outperforms hard labeling, demonstrating" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 357, + 101, + 512, + 231 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 6. Performance comparison to the state-of-the-art models on ChartQA." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 357, + 101, + 512, + 231 + ], + "lines": [ + { + "bbox": [ + 357, + 101, + 512, + 231 + ], + "spans": [ + { + "bbox": [ + 357, + 101, + 512, + 231 + ], + "type": "table", + "html": "
ModelsChartQA
ChartPaLI-5B [8]77.3
ScreenAI 5B [3]76.7
MatCha4096 + LaMenDa [29]72.6
SMoLA-PaLI-X [69]74.6
Qwen-VL-Chat [4]66.3
Qwen-VL-Max [4]79.8
LLaVA-OV-7B [28]80.0
Cambrian-34B [62]73.8
Gemini Ultra [60]80.8
SoftLabeling81.5
", + "image_path": "88fe1b4422d6d100e5977c07aeca667f26d11a6729484c0d04da1676f104c103.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 326, + 284, + 543, + 402 + ], + "blocks": [ + { + "bbox": [ + 313, + 243, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 243, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 555, + 277 + ], + "type": "text", + "content": "Table 7. Performance comparison to the state-of-the-art models on TallyQA. * indicates that the numbers are copied from a third-party paper [37]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 284, + 543, + 402 + ], + "lines": [ + { + "bbox": [ + 326, + 284, + 543, + 402 + ], + "spans": [ + { + "bbox": [ + 326, + 284, + 543, + 402 + ], + "type": "table", + "html": "
ModelsTallyQA simple Acc.(↑) RMSE(↓)TallyQA complex Acc.(↑) RMSE(↓)
SMoLA-PaLI-X [69]86.3-77.1-
PaLI-X-VPD [22]86.2-76.6-
MoVie-ResNeXt [45]74.91.0056.81.43
RCN [1]71.81.1356.21.43
LLaVA-NEXT [33]79.8*0.70*67.9*1.76*
GPT-4V [2]73.6*0.86*62.6*1.58*
GPT-4o [23]81.5*0.60*71.7*1.21*
SoftLabeling86.60.5677.21.06
", + "image_path": "964173bed43c1607ed0d09d6a62fb25394a61653fc510420cb52465e7e8a6d8b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 416, + 555, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 524 + ], + "type": "text", + "content": "its robustness. Among the distributions, Binomial and Triangular yield the best results, followed by Poisson, while Uniform performs the worst. This suggests that sharper distributions are more effective, as they provide more discriminative token similarities. In contrast, smoother distributions like Uniform treat all incorrect tokens as equally distant from the target, disregarding the relative distance information. Throughout this paper, we use Triangular distribution as default, though Binomial is comparable." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Hyperparameter There are two primary hyperparameters influencing the performances, " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " of (4) and " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " of (5). " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " controls the sharpness of the soft labeling. It is selected through exhaustive search based on the performance of the base LLaVA-7B specialist model on RefCOCOg val, ChartQA and TallyQA. For ChartQA, we evaluate accuracy on numerical subset. On TallyQA, we use RMSE as metric and evaluate on simple balanced set. Figure 2 shows the ablation study results. Our analysis shows that the model favors sharper soft labeling distributions with smaller " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": ". However, if it is too small, the distribution is towards a hard one-hot encoding, leading to performance drop. On the other hand, a large " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " over smooths the distribution, reducing discrimination among numerical tokens. In our experiments, " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " is set to " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.05 / 0.02 / 0.08 / 0.05" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " for Binomial/Poisson/Triangular/U" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3430" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 556, + 215 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 556, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 556, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 556, + 215 + ], + "type": "table", + "html": "
Image
Promptfirst sandwich on the left just beneath the forkthe man on the closest motorcycleWhat was the internet penetration rate in Africa in 2019?How many books did Blurb publish in 2018?
Hard labelingred boxred box2016000
Soft labelinggreen boxgreen box27.419053
Ground truthblue boxblue box28.619098
", + "image_path": "7d950f7cd73796610b77251de0334a6a8fb628c4e36325fb1ae5a09667d58e69.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 115, + 215, + 493, + 225 + ], + "lines": [ + { + "bbox": [ + 115, + 215, + 493, + 225 + ], + "spans": [ + { + "bbox": [ + 115, + 215, + 493, + 225 + ], + "type": "text", + "content": "Table 8. Qualitative examples of soft labeling improvement in visual grounding and chart understanding." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 231, + 234, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 234, + 242 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 234, + 242 + ], + "type": "text", + "content": "niform distribution, respectively, by default." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "text", + "content": "As for " + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "text", + "content": " of (5), it balances the contribution of numerical token loss and regular token loss in the final loss. We ablate its effect using the base 7B and stronger 13B specialist models across three datasets and present the results in Figure 3. It can be found that the models perform well and remain stable when " + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "text", + "content": " ranges from 2 to 5, beyond which the performance degrades. In our experiments, we set " + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 54, + 243, + 296, + 337 + ], + "type": "text", + "content": " as 2 by default." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 243, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 243, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 243, + 360 + ], + "type": "text", + "content": "4.3. Comparison to the State-of-the-Art" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 364, + 296, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 296, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 296, + 580 + ], + "type": "text", + "content": "We evaluate soft labeling and compare it with other methods across visual grounding, chart understanding, and object counting using public benchmarks, including RefCOCO (val/testA/testB), RefCOCO+ (val/testA/testB), RefCOCOg (val/test), ChartQA (test), and TallyQA (simple/complex). The model is pretrained on scaled-up domain- and task-specific data and fine-tuned as specialist models (details in Section B of the Supp.). By default, our results are based on the 13B model, but for grounding, we follow standard practice and compare 7B and 13B models separately. As shown in Table 5, 6, and 7, our Soft-Labeling, without any architectural modifications, achieves the state-of-the-art performance on RefCOCOg among 7B scale models, RefCOCO, RefCOCO+, RefCOCOg among 13B scale models, ChartQA, and TallyQA. It also delivers competitive results on RefCOCO and RefCOCO+ in 7B scale model group. These results highlight the effectiveness of soft labeling." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 588, + 139, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 139, + 600 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 139, + 600 + ], + "type": "text", + "content": "4.4. Visualization" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Table 8 presents examples randomly selected from visual RefCOCOg and ChartQA test sets, demonstrating the improvement achieved by soft labeling. Compared to hard labeling, soft labeling produces more accurate numerical predictions for both bounding box coordinates and chart-based values, bringing them closer to the ground truth. For instance, in the motorcycle image, the bounding box predicted by soft labeling (green) better aligns with the ground truth (blue) in both position and scale. Similarly, in the chart-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 278 + ], + "type": "text", + "content": "based QA task, soft labeling provides more accurate numerical estimation. These examples highlight the effectiveness of soft labeling in improving numerical prediction across both spatial and quantitative tasks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 287, + 384, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 384, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 384, + 300 + ], + "type": "text", + "content": "5. Limitation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": "In this work, we limit our scope to digit tokens. However, other tokens also exhibit semantic similarity, such as natural language numerals (\"one/two/three\") or categorical tokens (\"easy/medium/hard\"). Incorporating these into the loss computation is an avenue for future research. Another limitation is that our current loss function considers only single digits. Many numerical predictions involve multi-digit numbers, where digit-wise comparisons may not accurately reflect actual numerical distances. For instance, comparing \"21\" v.s. \"32\" digit-wise results in a distance of " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "|2 - 3| + |1 - 2| = 2" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " which is smaller than " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "|2 - 1| + |2 - 9| = 8" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " for \"21\" v.s. \"29\". However, their actual numerical distances are " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "|21 - 32| = 11" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "|21 - 29| = 8" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": ", showing a discrepancy in digit-wise calculations. Addressing this limitation remains an open challenge." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 388, + 502 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 712 + ], + "type": "text", + "content": "In this paper, we have identified the limitations of cross-entropy loss with one-hot encoding for numerical prediction in LLMs/MLLMs, where treating numbers as discrete tokens fails to reflect their continuous nature. To address this, we have proposed soft labeling, which smooths target distributions to account for numerical proximity, ensuring that closer predictions incur lower loss. This method retains compatibility with the next-token prediction paradigm without requiring architectural modifications. Our experiments across visual grounding, object counting, and chart understanding have shown that soft labeling consistently improves numerical prediction performance. We have achieved state-of-the-art or competitive results on RefCOCO " + }, + { + "bbox": [ + 313, + 510, + 556, + 712 + ], + "type": "inline_equation", + "content": "+ / g" + }, + { + "bbox": [ + 313, + 510, + 556, + 712 + ], + "type": "text", + "content": ", ChartQA, and TallyQA using a basic LLaVA model. Our findings highlight the importance of adapting appropriate loss functions for numerical tasks in LLMs/MLLMs." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3431" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Manoj Acharya, Kushal Kafle, and Christopher Kanan. Tallyqa: Answering complex counting questions. In AAAI, pages 8076-8084, 2019. 2, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 180 + ], + "type": "text", + "content": "[2] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 1, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 237 + ], + "type": "text", + "content": "[3] Gilles Baechler, Srinivas Sunkara, Maria Wang, Fedir Zubach, Hassan Mansoor, Vincent Etter, Victor Cärbune, Jason Lin, Jindong Chen, and Abhanshu Sharma. Screenai: A vision-language model for ui and infographics understanding. arXiv preprint arXiv:2402.04615, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 303 + ], + "type": "text", + "content": "[4] Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond. arxiv 2023. arXiv preprint arXiv:2308.12966, 1(8), 2023. 1, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 361 + ], + "type": "text", + "content": "[5] Javier Barbero-Gomez, Pedro-Antonio Gutierrez, VICTOR-Manuel Vargas, Juan-Antonio Vallejo-Casas, and Cesar Hervás-Martínez. An ordinal cnn approach for the assessment of neurological damage in parkinson's disease patients. Expert Systems with Applications, 182:115271, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 396 + ], + "type": "text", + "content": "[6] Christopher Beckham and Christopher Pal. Unimodal probability distributions for deep ordinal classification. In ICML, pages 411-419. PMLR, 2017. 3, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 397, + 295, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 397, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 62, + 397, + 295, + 452 + ], + "type": "text", + "content": "[7] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. NeurIPS, 33:1877-1901, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 507 + ], + "type": "text", + "content": "[8] Victor Carbune, Hassan Mansoor, Fangyu Liu, Rahul Aralikatte, Gilles Baechler, Jindong Chen, and Abhanshu Sharma. Chart-based reasoning: Transferring capabilities from llms to vlms. arXiv preprint arXiv:2403.12596, 2024. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 295, + 554 + ], + "type": "text", + "content": "[9] Alberto Castano, Pablo González, Jaime Alonso González, and Juan Jose Del Coz. Matching distributions algorithms based on the earth mover's distance for ordinal quantification. IEEE T-NNLS, 35(1):1050-1061, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 600 + ], + "type": "text", + "content": "[10] Gongwei Chen, Leyang Shen, Rui Shao, Xiang Deng, and Liqiang Nie. Lion: Empowering multimodal large language model with dual-level visual knowledge. In CVPR, pages 26540-26550, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 667 + ], + "type": "text", + "content": "[11] Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478, 2023. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 712 + ], + "type": "text", + "content": "[12] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In CVPR, pages 24185–24198, 2024. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 173 + ], + "type": "text", + "content": "[14] Zheng Chu, Jingchang Chen, Qianglong Chen, Weijiang Yu, Haotian Wang, Ming Liu, and Bing Qin. Timebench: A comprehensive evaluation of temporal reasoning abilities in large language models. arXiv preprint arXiv:2311.17667, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 196 + ], + "type": "text", + "content": "[15] Raul Diaz and Amit Marathe. Soft labels for ordinal regression. In CVPR, pages 4738-4747, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 252 + ], + "type": "text", + "content": "[16] Bahare Fatemi, Mehran Kazemi, Anton Tsitsulin, Karishma Malkan, Jinyeong Yim, John Palowitch, Sungyong Seo, Jonathan Halcrow, and Bryan Perozzi. Test of time: A benchmark for evaluating llms on temporal reasoning. arXiv preprint arXiv:2406.09170, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 297 + ], + "type": "text", + "content": "[17] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In CVPR, pages 2002-2011, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 298, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 553, + 342 + ], + "type": "text", + "content": "[18] Yash Goyal, Tejas Khot, Douglas Summers-Stay, Dhruv Batra, and Devi Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. In CVPR, pages 6904-6913, 2017. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 317, + 344, + 553, + 398 + ], + "type": "text", + "content": "[19] Pedro Antonio Gutierrez, Maria Perez-Ortiz, Javier Sanchez-Monedero, Francisco Fernandez-Navarro, and Cesar Hervas-Martinez. Ordinal regression methods: survey and experimental study. IEEE Transactions on Knowledge and Data Engineering, 28(1):127-146, 2015. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 399, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 399, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 399, + 553, + 432 + ], + "type": "text", + "content": "[20] Le Hou, Chen-Ping Yu, and Dimitris Samaras. Squared earth mover's distance-based loss for training deep neural networks. arXiv preprint arXiv:1611.05916, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 477 + ], + "type": "text", + "content": "[21] Heng-Wei Hsu, Tung-Yu Wu, Sheng Wan, Wing Hung Wong, and Chen-Yi Lee. Quatnet: Quaternion-based head pose estimation with multiregression loss. IEEE Transactions on Multimedia, 21(4):1035–1046, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 533 + ], + "type": "text", + "content": "[22] Yushi Hu, Otilia Stretcu, Chun-Ta Lu, Krishnamurthy Viswanathan, Kenji Hata, Enming Luo, Ranjay Krishna, and Ariel Fuxman. Visual program distillation: Distilling tools and programmatic reasoning into vision-language models. In CVPR, pages 9590-9601, 2024. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 578 + ], + "type": "text", + "content": "[23] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Weli-hinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 553, + 624 + ], + "type": "text", + "content": "[24] Samira Ebrahimi Kahou, Vincent Michalski, Adam Atkinson, Ákos Kádár, Adam Trischler, and Yoshua Bengio. Figureqa: An annotated figure dataset for visual reasoning. arXiv preprint arXiv:1710.07300, 2017. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 667 + ], + "type": "text", + "content": "[25] Sahar Kazemzadeh, Vicente Ordonez, Mark Matten, and Tamara Berg. Referitgame: Referring to objects in photographs of natural scenes. In EMNLP, pages 787-798, 2014. 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[26] Taku Kudo and John Richardson. Sentencepiece: A simple and language independent subword tokenizer and detokenizer for neural text processing. arXiv preprint arXiv:1808.06226, 2018. 4" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3432" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 138 + ], + "type": "text", + "content": "[27] Alina Kuznetsova, Hassan Rom, Neil Alldrin, Jasper Uijlings, Ivan Krasin, Jordi Pont-Tuset, Shahab Kamali, Stefan Popov, Matteo Malloci, Alexander Kolesnikov, et al. The open images dataset v4: Unified image classification, object detection, and visual relationship detection at scale. IJCV, 128(7):1956-1981, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[28] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Peiyuan Zhang, Yanwei Li, Ziwei Liu, et al. Llava-onevision: Easy visual task transfer. arXiv preprint arXiv:2408.03326, 2024. 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 228 + ], + "type": "text", + "content": "[29] Zhuowan Li, Bhavan Jasani, Peng Tang, and Shabnam Ghadar. Synthesize step-by-step: Tools templates and llms as data generators for reasoning-based chart vqa. In CVPR, pages 13613-13623, 2024. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 262 + ], + "type": "text", + "content": "[30] Bingyuan Liu, Ismail Ben Ayed, Adrian Galdran, and Jose Dolz. The devil is in the margin: Margin-based label smoothing for network calibration. In CVPR, pages 80-88, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 296 + ], + "type": "text", + "content": "[31] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. NeurIPS, 36:34892-34916, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 331 + ], + "type": "text", + "content": "[32] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, pages 26296-26306, 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 365 + ], + "type": "text", + "content": "[33] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 294, + 398 + ], + "type": "text", + "content": "[34] Peiyang Liu, Xiangyu Xi, Wei Ye, and Shikun Zhang. Label smoothing for text mining. In ICCL, pages 2210-2219, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[35] Xiaofeng Liu, Yang Zou, Yuhang Song, Chao Yang, Jane You, and BV K Vijaya Kumar. Ordinal regression with neuron stick-breaking for medical diagnosis. In ECCV Workshops, pages 0-0, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 488 + ], + "type": "text", + "content": "[36] Xiaofeng Liu, Fangfang Fan, Lingsheng Kong, Zhihui Dao, Wanqing Xie, Jun Lu, and Jane You. Unimodal regularized neuron stick-breaking for ordinal classification. Neurocomputing, 388:34-44, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[37] Jian Lu, Shikhar Srivastava, Junyu Chen, Robik Shrestha, Manoj Acharya, Kushal Kafle, and Christopher Kanan. Revisiting multi-modal llm evaluation. arXiv preprint arXiv:2408.05334, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 589 + ], + "type": "text", + "content": "[38] Pan Lu, Hritik Bansal, Tony Xia, Jiacheng Liu, Chunyuan Li, Hannaneh Hajishirzi, Hao Cheng, Kai-Wei Chang, Michel Galley, and Jianfeng Gao. Mathvista: Evaluating mathematical reasoning of foundation models in visual contexts. arXiv preprint arXiv:2310.02255, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 623 + ], + "type": "text", + "content": "[39] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 668 + ], + "type": "text", + "content": "[40] Ahmed Masry, Do Xuan Long, Jia Qing Tan, Shafiq Joty, and Enamul Hoque. Chartqa: A benchmark for question answering about charts with visual and logical reasoning. arXiv preprint arXiv:2203.10244, 2022. 2, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[41] Ahmed Masry, Parsa Kavehzadeh, Xuan Long Do, Enamul Hoque, and Shafiq Joty. Unichart: A universal vision-language pretrained model for chart comprehension and reasoning. arXiv preprint arXiv:2305.14761, 2023. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 106 + ], + "type": "text", + "content": "[42] Minesh Mathew, Dimosthenis Karatzas, and CV Jawahar. Docvqa: A dataset for vqa on document images. In WACV, pages 2200-2209, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[43] Nitesh Methani, Pritha Ganguly, Mitesh M Khapra, and Pratyush Kumar. Plotqa: Reasoning over scientific plots. In WACV, pages 1527-1536, 2020. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 163 + ], + "type": "text", + "content": "[44] Rafael Müller, Simon Kornblith, and Geoffrey E Hinton. When does label smoothing help? NeurIPS, 32, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 197 + ], + "type": "text", + "content": "[45] Duy-Kien Nguyen, Vedanuj Goswami, and Xinlei Chen. Movie: Revisiting modulated convolutions for visual counting and beyond. arXiv preprint arXiv:2004.11883, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 198, + 553, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 231 + ], + "type": "text", + "content": "[46] Zhenxing Niu, Mo Zhou, Le Wang, Xinbo Gao, and Gang Hua. Ordinal regression with multiple output cnn for age estimation. In CVPR, pages 4920-4928, 2016. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 553, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 276 + ], + "type": "text", + "content": "[47] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, and Furu Wei. Kosmos-2: Grounding multimodal large language models to the world. arXiv preprint arXiv:2306.14824, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 319 + ], + "type": "text", + "content": "[48] Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Qixiang Ye, and Furu Wei. Grounding multimodal large language models to the world. In ICLR, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 376 + ], + "type": "text", + "content": "[49] Gorkem Polat, Ilkay Ergenc, Haluk Tarik Kani, Yesim Ozen Alahdab, Ozlen Atug, and Alptekin Temizel. Class distance weighted cross-entropy loss for ulcerative colitis severity estimation. In Annual Conference on Medical Image Understanding and Analysis, pages 157–171. Springer, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "spans": [ + { + "bbox": [ + 316, + 378, + 553, + 433 + ], + "type": "text", + "content": "[50] Shraman Pramanick, Guangxing Han, Rui Hou, Sayan Nag, Ser-Nam Lim, Nicolas Ballas, Qifan Wang, Rama Chellappa, and Amjad Almahairi. Jack of all tasks master of many: Designing general-purpose coarse-to-fine vision-language model. In CVPR, pages 14076-14088, 2024. 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[51] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, TatSeng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. arXiv preprint arXiv:2402.11435, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 489, + 553, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 522 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 522 + ], + "type": "text", + "content": "[52] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 524, + 553, + 555 + ], + "type": "text", + "content": "[53] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 622 + ], + "type": "text", + "content": "[54] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. NeurIPS, 35:25278-25294, 2022. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[55] Shuai Shao, Zeming Li, Tianyuan Zhang, Chao Peng, Gang Yu, Xiangyu Zhang, Jing Li, and Jian Sun. Objects365: A large-scale, high-quality dataset for object detection. In ICCV, pages 8430-8439, 2019. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[56] Amanpreet Singh, Vivek Natarajan, Meet Shah, Yu Jiang, Xinlei Chen, Dhruv Batra, Devi Parikh, and Marcus Rohrbach. Towards vqa models that can read. In CVPR, pages 8317-8326, 2019. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3433" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[57] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In CVPR, pages 2818-2826, 2016. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 295, + 159 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 295, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 295, + 159 + ], + "type": "text", + "content": "[58] Zichang Tan, Shuai Zhou, Jun Wan, Zhen Lei, and Stan Z Li. Age estimation based on a single network with soft softmax of aging modeling. In ACCV, pages 203-216. Springer, 2016. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 194 + ], + "type": "text", + "content": "[59] Ryota Tanaka, Kyosuke Nishida, and Sen Yoshida. Visualmrc: Machine reading comprehension on document images. In AAAI, pages 13878-13888, 2021. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 248 + ], + "type": "text", + "content": "[60] Gemini Team, Rohan Anil, Sebastian Borgeaud, Jean-Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: a family of highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023. 1, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[61] Christian Thiel. Classification on soft labels is robust against label noise. In International Conference on Knowledge-Based and Intelligent Information and Engineering Systems, pages 65-73. Springer, 2008. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 358 + ], + "type": "text", + "content": "[62] Shengbang Tong, Ellis Brown, Penghao Wu, Sanghyun Woo, Manoj Middepogu, Sai Charitha Akula, Jihan Yang, Shusheng Yang, Adithya Iyer, Xichen Pan, Austin Wang, Rob Fergus, Yann LeCun, and Saining Xie. Cambrian-1: A fully open, vision-centric exploration of multimodal llms, 2024. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 415 + ], + "type": "text", + "content": "[63] Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 448 + ], + "type": "text", + "content": "[64] Eric Tzeng, Judy Hoffman, Trevor Darrell, and Kate Saenko. Simultaneous deep transfer across domains and tasks. In ICCV, pages 4068-4076, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 491 + ], + "type": "text", + "content": "[65] Víctor Manuel Vargas, Pedro Antonio Gutiérrez, Javier Barbero-Gómez, and César Hervás-Martínez. Soft labelling based on triangular distributions for ordinal classification. Information Fusion, 93:258-267, 2023. 3, 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 535 + ], + "type": "text", + "content": "[66] Sukesh Adiga Vasudeva, Jose Dolz, and Herve Lombaert. Geols: geodesic label smoothing for image segmentation. In Medical Imaging with Deep Learning, pages 468-478. PMLR, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 580 + ], + "type": "text", + "content": "[67] Ke Wang, Junting Pan, Weikang Shi, Zimu Lu, Houxing Ren, Aojun Zhou, Mingjie Zhan, and Hongsheng Li. Measuring multimodal mathematical reasoning with math-vision dataset. NeurIPS, 37:95095-95169, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 613 + ], + "type": "text", + "content": "[68] Scott Workman, Richard Souvenir, and Nathan Jacobs. Understanding and mapping natural beauty. In ICCV, pages 5589-5598, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 657 + ], + "type": "text", + "content": "[69] Jialin Wu, Xia Hu, Yaqing Wang, Bo Pang, and Radu Soricut. Omni-smola: Boosting generalist multimodal models with soft mixture of low-rank experts. In CVPR, pages 14205-14215, 2024. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[70] Haoxuan You, Haotian Zhang, Zhe Gan, Xianzhi Du, Bowen Zhang, Zirui Wang, Liangliang Cao, Shih-Fu Chang, and Yinfei Yang. Ferret: Refer and ground anything anywhere at any granularity. arXiv preprint arXiv:2310.07704, 2023. 2, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 217 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 106 + ], + "type": "text", + "content": "[71] Chang-Bin Zhang, Peng-Tao Jiang, Qibin Hou, Yunchao Wei, Qi Han, Zhen Li, and Ming-Ming Cheng. Delving deep into label smoothing. IEEE T-IP, 30:5984-5996, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[72] Haotian Zhang, Haoxuan You, Philipp Dufter, Bowen Zhang, Chen Chen, Hong-You Chen, Tsu-Jui Fu, William Yang Wang, Shih-Fu Chang, Zhe Gan, et al. Ferretv2: An improved baseline for referring and grounding with large language models. arXiv preprint arXiv:2404.07973, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 217 + ], + "type": "text", + "content": "[73] Helong Zhou, Liangchen Song, Jiajie Chen, Ye Zhou, Guoli Wang, Junsong Yuan, and Qian Zhang. Rethinking soft labels for knowledge distillation: A bias-variance tradeoff perspective. arXiv preprint arXiv:2102.00650, 2021. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3434" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_content_list.json b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3c5ab7bce41045273474d4aa2b993a250ef9fd2f --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_content_list.json @@ -0,0 +1,1980 @@ +[ + { + "type": "text", + "text": "Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning", + "text_level": 1, + "bbox": [ + 135, + 130, + 864, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Li $^{1*}$ , Jinpeng Wang $^{2*†}$ , Chaolei Tan $^{4}$ , Niu Lian $^{1}$ , Long Chen $^{4}$ , Yaowei Wang $^{1,3}$ , Min Zhang $^{1}$ , Shu-Tao Xia $^{2,3}$ , Bin Chen $^{1}$", + "bbox": [ + 230, + 179, + 764, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Harbin Institute of Technology, Shenzhen", + "bbox": [ + 370, + 218, + 625, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Tsinghua Shenzhen International Graduate School, Tsinghua University", + "bbox": [ + 281, + 234, + 715, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3Research Center of Artificial Intelligence, Peng Cheng Laboratory", + "bbox": [ + 297, + 253, + 699, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "4The Hong Kong University of Science and Technology", + "bbox": [ + 331, + 271, + 665, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "220110924@stu.hit.edu.cn $\\text{四}$ wjp20@mails.tsinghua.edu.cn", + "bbox": [ + 253, + 289, + 738, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 339, + 326, + 354 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFformer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFformer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce \"text $\\prec$ video\" hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFformer outperforms state-of-the-art methods. Code is released at https://github.com/lijun2005/ICCV25-HLFformer.", + "bbox": [ + 86, + 369, + 485, + 704 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 733, + 220, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-to-video retrieval (T2VR) [5, 11-13, 15, 18, 34, 37, 42] is a fundamental module in many search applications and a popular topic in multi-modal learning. While most T2VR models are developed for short clips or pre-trimmed video segments, they may face challenges where user queries describe only partial content in the video. This practical issue in real-world usage promotes a more challenging setting of", + "bbox": [ + 89, + 757, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2f6a4a5389cca2054b178cee763893f730f3b0be262e16d79f80f1f782230c13.jpg", + "image_caption": [ + "(a) Semantic Hierarchy in Untrimmed Videos", + "Figure 1. (a) Modeling the semantic hierarchy in untrimmed videos helps Partially Relevant Video Retrieval (PRVR). (b) Euclidean space is less effective in modeling semantic hierarchy due to the flat geometry. Data points with distant hierarchical relation may be close. (c) Hyperbolic space allows larger cardinals when approaching the edge, which is preferable to preserve the hierarchy." + ], + "image_footnote": [], + "bbox": [ + 526, + 337, + 720, + 516 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e73c642f342019f007b74c6fbe378b48667eaee066e9d0b296b2ed2216bba876.jpg", + "image_caption": [ + "(c) Hyperbolic Space" + ], + "image_footnote": [], + "bbox": [ + 722, + 338, + 893, + 516 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "partially relevant video retrieval (PRVR) [14], which aims to match each text query with the best untrimmed video.", + "bbox": [ + 509, + 655, + 906, + 685 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to unlabeled moment timestamps, PRVR requires solid abilities on (i) identifying key moments in videos for extracting informative features and (ii) learning robust cross-modal representations to match text queries and videos precisely. Prior arts have developed preliminary solutions on both aspects, while challenges remain. For (i), MS-SL [14] exhaustively enumerated consecutive frame combinations through multi-scale sliding windows, which inevitably engaged redundancy, noise, and a high computational complexity in extracting moment features. GMMFormer [56, 57] improved efficiency by leveraging Gaussian neighborhood priors to traverse each timestamp and discover potential key moments. However, it may still be hard to distinguish adjacent or semantically similar candidate moments. Though", + "bbox": [ + 509, + 689, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "23074", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "DL-DKD [16] neatly benefited from the pretrained CLIP [48] to enhance text-frame alignment, the temporal generalizability is bounded by the text-image teacher model. For (ii), most existing solutions inherited similar ideas from classic T2VR, e.g., ranking and contrastive learning, at a holistic level, but important characteristics of PRVR, e.g., partial relevance and semantic entailment, are still under-explored.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we take a hierarchical perspective to review the task, in the belief that videos naturally exhibit semantic hierarchy. As illustrated in Fig. 1(a), an untrimmed video can be regarded as a progression from frames to informative segments (e.g., Dunk), extended moments, and ultimately, the whole. Leveraging this intrinsic property is expected to benefit long video understanding. In particular for PRVR, the hierarchical prior provides positive guidance to arrange the moment features. Meanwhile, the supervisory signals from query-video matching can activate moment extraction more precisely through implicit bottom-up modeling. Exploring hierarchical features is never trivial. Unfortunately, existing PRVR approaches relying on Euclidean space are less effective in modeling the desired patterns in the flat geometry. We present Fig. 1(b) to exemplify this: two embeddings with a distant hierarchical relation may be spatially close to each other, as marked by the red arrows. Biased representation will increase the difficulty in disentangling informative moments from background, which limits the robustness in cross-modal matching considering partial relevance.", + "bbox": [ + 91, + 200, + 486, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the emerging success of hyperbolic learning [10, 17, 29, 31, 44], which takes advantage of exponentially expanding metric in non-Euclidean space to better capture hierarchical structure (Fig. 1(c)), we introduce HLFoRer, a sincere exploration of hyperbolic learning to enhance PRVR. On temporal modeling, we carefully design a dual-branch strategy to capture informative moment features comprehensively. Specifically, for the hyperbolic branch, we develop a Lorentz Attention Block (LAB) with the hyperbolic selfattention mechanism. With the implicit hierarchical prior through end-to-end matching optimization, LAB learns to activate informative moment features relevant to queries and distinguish them from noisy background in the hyperbolic space, compensating for the limitations of Euclidean attention in capturing hierarchical semantics. We integrate dual-branch moment features with a Mean-Guided Adaptive Interaction Module (MAIM), which is lightweight but effective. On cross-modal matching, drawing on the intrinsic \"text $\\prec$ video\" hierarchy in PRVR where textual queries are subordinate to their paired videos, we introduce a Partial Order Preservation (POP) loss that geometrically confines text embeddings within hyperbolic cone anchored by corresponding video representations in an auxiliary Lorentzian manifold. This hierarchical metric alignment ensures semantic consistency between localized text semantics and their parent video structure while preserving partial relevance.", + "bbox": [ + 91, + 508, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Empirical evaluations on three benchmark datasets: ActivityNet Captions [28], Charades-STA [22], and TVR [30] establish HLFoerner's state-of-the-art performance. Ablation studies confirm the necessity of hyperbolic geometry for hierarchical representation and the critical role of explicitly relational constraints in Partial Order Preservation Loss. Meanwhile, visual evidences further reveal that hyperbolic learning can enhance discriminative representation while maintaining video-text entailment, sharpening moment distinction and improving query alignment.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The primary contributions can be summarized as follows:", + "bbox": [ + 535, + 243, + 906, + 258 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose to enhance PRVR with hyperbolic learning, including a Lorentz attention block with hierarchical priors to enhance the moment feature extraction, which collaborates with Euclidean attention and hybrid-space fusion.", + "- We design a partial order preservation loss that geometrically enforces the \"text $\\prec$ video\" hierarchy through hyperbolic cone constraints, strengthening partial relevance.", + "- Extensive experiments on three benchmarks validate HLFformer's superiority, with analyses confirming the efficacy of hyperbolic modeling and geometric constraints." + ], + "bbox": [ + 511, + 258, + 906, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 426, + 663, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Partially Relevant Video Retrieval", + "text_level": 1, + "bbox": [ + 511, + 453, + 812, + 468 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "With the growth of video content [19, 35, 58], video retrieval has become a key research area. Given a text query, Text-to-Video Retrieval (T2VR) [5, 11, 15, 18, 34, 36, 37, 42, 54, 55] focuses on retrieving fully relevant videos from pre-trimmed short clips. Video Corpus Moment Retrieval (VCMR) [7, 30, 50, 51] aims to localize specific moments within videos from a large corpus. Partially Relevant Video Retrieval (PRVR) [8, 9, 14, 16, 26, 56, 57, 60], a more recent task introduced by Dong et al. [14], aims to retrieve partially relevant videos from large, untrimmed long video collections. Unlike T2VR, PRVR must address the challenge of partial relevance, where the query pertains to only a specific moment of the video. Though the first stage of VCMR is similar to PRVR, VCMR requires moment-level annotations, limiting scalability.", + "bbox": [ + 511, + 476, + 908, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methods enhance PRVR retrieval from various perspectives. MS-SL [14] defines the PRVR task as a Multi-instance Learning, providing a strong baseline with explicit redundant clip embeddings. GMMFormer [56, 57] and PEAN [26] propose implicit clip modeling to improve efficiency. DL-DKD [16] achieves great results through dynamic distillation of CLIP [48]. BGM-Net [60] exploits an instance-level matching scheme for pairing queries and videos. However, these methods predominantly rely on Euclidean space, which sometimes distort the hierarchical structures in untrimmed long videos. Consequently, they fail to fully exploit video hierarchy priors. To overcome this issue, we propose HLFormer to enhances PRVR by implicitly capturing hierarchical structures through hyperbolic learning.", + "bbox": [ + 511, + 688, + 910, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "23075", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Hyperbolic Learning", + "text_level": 1, + "bbox": [ + 89, + 90, + 289, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic learning has attracted significant attention for its effectiveness in modeling hierarchical structures in real-world datasets. Early studies in computer vision tasks explored hyperbolic image embeddings from image-label pairs [27, 44], while subsequent progress extended hyperbolic optimization to multi-modal learning. MERU [10] and HyCoCLIP [46] notably surpassed Euclidean counterparts like CLIP [48] via hyperbolic space adaptation. Applications span semantic segmentation [1, 4], recognition tasks (skin [61], action [39]), meta-learning [17], and detection frameworks (violence [31, 47], anomalies [33]). Recent advances in fully hyperbolic neural networks [6, 21, 24, 32, 52] further underscore their potential. Motivated by them, we present the first study to explore the potential of hyperbolic learning for PRVR. Unlike other methods such as DSRL [31] and HOVER [49], our approach utilizes hyperbolic space to compensate for the limitations of Euclidean space in capturing the hierarchical structure of untrimmed long videos. Furthermore, we introduce the Partial Order Preservation Loss to explicitly capture the partial relevance between video and text in hyperbolic space, improving retrieval performance.", + "bbox": [ + 89, + 112, + 485, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 443, + 181, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 468, + 230, + 483 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Hyperbolic Space Hyperbolic spaces are Riemannian manifolds with a constant negative curvature $K$ , contrasting with the zero-curvature (flat) geometry of Euclidean spaces. Among several isometrically equivalent hyperbolic models, we adopt the Lorentz model [45] for its numerical stability and computational efficiency, with $K$ set to -1 by default.", + "bbox": [ + 89, + 489, + 483, + 580 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lorentz Model Formally, an $n$ -dimensional Lorentz model is the Riemannian manifold $\\mathbb{L}^n = (\\mathcal{L}^n,\\mathfrak{g}_x)$ . $\\mathfrak{g}_x = \\mathrm{diag}(-1,1,\\dots ,1)$ is the Riemannian metric tensor. Each point in $\\mathbb{L}^n$ has the form $\\pmb {x} = [x_0,\\pmb {x}_s]\\in \\mathbb{R}^{n + 1},x_0 = \\sqrt{\\|\\pmb{x}_s\\|^2 + 1}\\in \\mathbb{R}$ . Following Chen et al. [6], we denote $x_0$ as time axis and $\\pmb{x}_{s}$ as spatial axes. $\\mathcal{L}^n$ is given by:", + "bbox": [ + 89, + 580, + 483, + 674 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {x} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {x}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = - 1, x _ {0} > 0 \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 684, + 483, + 702 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and the Lorentzian inner product given by:", + "bbox": [ + 89, + 713, + 375, + 729 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}} := - x _ {0} y _ {0} + \\boldsymbol {x} _ {s} ^ {\\top} \\boldsymbol {y} _ {s}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 192, + 739, + 483, + 757 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here $\\mathcal{L}^n$ is the upper sheet of hyperboloid in a $(n + 1)$ dimensional Minkowski space with the origin $o = (1,0,\\dots ,0)$ .", + "bbox": [ + 89, + 768, + 483, + 799 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Tangent Space The tangent space at $\\pmb{x} \\in \\mathbb{L}^n$ is a Euclidean space that is orthogonal to it, defined as:", + "bbox": [ + 89, + 799, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\boldsymbol {x}} \\mathbb {L} ^ {n} := \\left\\{\\boldsymbol {y} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {y}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = 0 \\right\\}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 840, + 483, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Where $\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}$ is a Euclidean subspace of $\\mathbb{R}^{n + 1}$ . In particular, the tangent space at the origin $\\pmb{o}$ is denoted as $\\mathcal{T}_o\\mathbb{L}^n$", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Logarithmic and Exponential Maps The mutual mapping between the hyperbolic space $\\mathbb{L}^n$ and the Euclidean subspace $\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}$ can be realized by logarithmic and exponential maps. The exponential map $\\exp_{\\pmb{x}}(z)$ can map any tangent vector $z\\in \\mathcal{T}_x\\mathbb{L}^n$ to $\\mathbb{L}^n$ , written as:", + "bbox": [ + 511, + 90, + 906, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\exp_ {\\boldsymbol {x}} (\\boldsymbol {z}) = \\cosh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\boldsymbol {x} + \\sinh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\frac {\\boldsymbol {z}}{\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 176, + 906, + 205 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\| z\\|_{\\mathcal{L}} = \\sqrt{\\langle z,z\\rangle_{\\mathcal{L}}}$ and the logarithmic map $\\log_{x}(\\pmb {y})$ plays an opposite role to map $\\pmb {y}\\in \\mathbb{L}^n$ to $\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}$ as follows:", + "bbox": [ + 511, + 217, + 906, + 250 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\log_ {\\boldsymbol {x}} (\\boldsymbol {y}) = \\frac {\\operatorname {a r c o s h} (- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}})}{\\sqrt {(- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) ^ {2} - 1}} (\\boldsymbol {y} + (\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) \\boldsymbol {x}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 258, + 906, + 294 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Lorentzian centroid The weighted centroid with respect to the squared Lorentzian distance, which solves $\\min_{\\mu \\in \\mathbb{L}^n}\\sum_{i = 1}^{m}\\nu_id^2_{\\mathcal{L}}(\\pmb {x}_i,\\mu)$ , with $\\pmb {x}_i\\in \\mathbb{L}^n$ and $\\nu_{i}\\geq 0,\\sum_{i = 1}^{m}\\nu_{i} > 0$ , is denoted as:", + "bbox": [ + 511, + 303, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mu = \\frac {\\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i}}{| | | \\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i} | | _ {\\mathcal {L}} |}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 375, + 906, + 411 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Problem Formulation and Overview", + "text_level": 1, + "bbox": [ + 511, + 419, + 826, + 433 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Partially Relevant Video Retrieval (PRVR) aims to retrieve videos containing a moment semantically relevant to a given text query, from a large corpus of untrimmed videos. In the PRVR database, each video has multiple moments and is associated with multiple text descriptions, with each text description corresponding to a specific moment of the related video. Critically, the temporal boundaries of these moments (i.e., start and end time points) are not annotated.", + "bbox": [ + 511, + 440, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this paper, we introduce HLFormer, the first hyperbolic modeling approach designed for PRVR. The proposed framework encompasses three key components: text query representation encoding, video representation encoding, and similarity computation, as illustrated in Fig. 2 (a).", + "bbox": [ + 511, + 561, + 906, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text Representation Given a text query of $N_{q}$ words, we first use a pre-trained RoBERTa [38] model to extract word-level features, which are then projected into a lower-dimensional space via a fully connected (FC) layer. A standard Transformer [53] layer is applied to obtain a sequence of $d$ -dimensional contextualized feature vectors, $\\pmb{Q} = \\{\\pmb{q}_i\\}_{i=1}^{N_q} \\in \\mathbb{R}^{N_q \\times d}$ . Finally, we utilize a simple attention mechanism to get the sentence embedding $\\pmb{q} \\in \\mathbb{R}^d$ :", + "bbox": [ + 511, + 637, + 908, + 760 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} = \\sum_ {i = 1} ^ {N _ {q}} \\boldsymbol {a} _ {i} ^ {q} \\times \\boldsymbol {q} _ {i}, \\quad \\boldsymbol {a} ^ {q} = \\operatorname {s o f t m a x} (\\boldsymbol {w} \\boldsymbol {Q} ^ {\\top}), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 771, + 906, + 813 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{w} \\in \\mathbb{R}^{1 \\times d}$ is a trainable vector, and $\\pmb{a}^q \\in \\mathbb{R}^{1 \\times N_q}$ represents the attention vector.", + "bbox": [ + 511, + 823, + 903, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video Representation Given an untrimmed video, we first extract embedding features using a pre-trained 2D or 3D CNN. Then we utilize the gaze branch and glance branch", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "23076", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c3e6a22ba70d5d2764040c9022422b886c27c2307b8912ffdaf9dcc57322a9ed.jpg", + "image_caption": [ + "(a) The overall framework of HLFormer" + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 694, + 316 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/651766bd5169af4403a9cd6563ecd8fc36a254a7ea4d66a9ca316711673281b8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 88, + 901, + 212 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b5da0877f37d3f9a127bd3ee0399aeb98ed138a9f9ed5aef29270ea1dc8edff9.jpg", + "image_caption": [ + "Figure 2. Overview of HLFoer. (a) The sentence embedding $\\pmb{q}$ is obtained via the query branch, while the gaze and glance branches encode the video, producing frame-level embedding $V_{f}$ and clip-level embedding $V_{c}$ and forming the video representation $V_{v}$ . $\\pmb{q}$ learns query diversity through $L_{div}$ and computes similarity scores $S_{f}$ and $S_{c}$ , while preserving partial order relations with $V_{v}$ using $L_{pop}$ . (b) HLFoer block combines parallel Lorentz and Euclidean attention blocks for multi-space encoding, with a Mean Guided Adaptive Interaction Module for dynamic aggregation. (c) Partial Order Preservation Loss ensures the text query embedding $t$ lies within the cone defined by the video embedding $v$ . The loss is zero if $t$ is inside the cone." + ], + "image_footnote": [], + "bbox": [ + 696, + 212, + 901, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "to capture frame-level and clip-level multi-granularity video representations, respectively. In the gaze branch, we densely sample $M_{f}$ frames, denoted as $\\pmb{F} \\in \\mathbb{R}^{M_f \\times D}$ , where $D$ is the frame feature dimension. The sampled frames are processed through a fully connected (FC) layer to reduce the dimensionality to $d$ , followed by the HLFoer block to obtain frame embeddings $\\pmb{V}_{\\pmb{f}} = \\{\\pmb{f}_{i}\\}_{i=1}^{M_{f}} \\in \\mathbb{R}^{M_{f} \\times d}$ , capturing semantically rich frame-level information for fine-grained relevance assessment to the query. The glance branch down-samples the input along the temporal dimension to aggregate frames into clips. Following MS-SL [14], a fixed number $M_{c}$ of clips is sparsely sampled by mean pooling over consecutive frames. A fully connected layer is applied to the pooled clip features, followed by the HLFoer block, generating clip embeddings $\\pmb{V}_{c} = \\{\\pmb{c}_{i}\\}_{i=1}^{M_{c}} \\in \\mathbb{R}^{M_{c} \\times d}$ . These embeddings capture adaptive clip-level information, enabling the model to perceive relevant moments at a coarser granularity.", + "bbox": [ + 88, + 455, + 485, + 715 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarity Computation To compute the similarity between a text-video pair $(\\mathcal{T},\\mathcal{V})$ , we first measure the abovementioned embeddings $q,V_{f}$ and $V_{c}$ . Then, we employ cosine similarity along with a max operation to calculate the frame-level and clip-level similarity scores:", + "bbox": [ + 89, + 715, + 483, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {f} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {M _ {f}}\\right) \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 805, + 483, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS _ {c} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {1}), \\dots , \\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {M _ {c}}) \\right\\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 825, + 442, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we compute the overall text-video pair similarity:", + "bbox": [ + 89, + 854, + 457, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathcal {T}, \\mathcal {V}) = \\alpha_ {f} S _ {f} (\\mathcal {T}, \\mathcal {V}) + \\alpha_ {c} S _ {c} (\\mathcal {T}, \\mathcal {V}), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 883, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{f},\\alpha_{c}\\in [0,1]$ are hyper-parameters satisfying $\\alpha_{f} + \\alpha_{c} = 1$ . Finally, we retrieve and rank partially relevant videos based on the computed similarity scores.", + "bbox": [ + 511, + 455, + 906, + 502 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. HlFormer Block", + "text_level": 1, + "bbox": [ + 511, + 512, + 681, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The HLFoer block constitutes the core of our method. As shown in Fig. 2 (b), it comprises three key modules: (i) Euclidean Attention Block, capturing fine-grained visual features in Euclidean space; (ii) Lorentz Attention Block, projecting video embeddings into hyperbolic Lorentz space for capturing the hierarchical structures of video; (iii) Mean-Guided Adaptive Interaction Module, dynamically fusing hybrid-space features. We describe the details below.", + "bbox": [ + 511, + 534, + 906, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Euclidean Attention Block Given M feature embeddings $\\pmb{x} \\in \\mathbb{R}^{M \\times d}$ , where $d$ is the feature dimension, the Euclidean Attention Block utilizes Euclidean Gaussian Attention [57] to capture multi-scale visual features, expressed as:", + "bbox": [ + 511, + 655, + 905, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {G A} (\\boldsymbol {x}) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {\\sigma} ^ {g} \\odot \\frac {\\boldsymbol {x} W ^ {q} \\left(\\boldsymbol {x} W ^ {k}\\right) ^ {\\top}}{\\sqrt {d _ {h}}}\\right) \\boldsymbol {x} W ^ {v}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 727, + 906, + 762 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{M}_{\\sigma}^{g}$ is the Gaussian matrix with elements $\\mathcal{M}_{\\sigma}^{g}(i,j) = \\frac{1}{2\\pi} e^{-\\frac{(j - i)^2}{\\sigma^2}}$ , and $\\sigma^2$ denotes the variance. By varying $\\sigma$ , feature interactions at different scales are modeled, generating video features with multiple receptive fields. $W^{q}, W^{k}, W^{v}$ are linear projections, while $d_{h}$ is the latent attention dimension, $\\odot$ denotes element-wise product. Finally, We replace the self-attention in Transformer block with Euclidean Gaussian attention to form the Euclidean Attention Block.", + "bbox": [ + 511, + 773, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "23077", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Lorentz Attention Block Given extracted Euclidean video embeddings $\\pmb{x}_{\\mathrm{in}}^{E} \\in \\mathbb{R}^{M \\times d}$ , we first project it to $\\mathbb{R}^{M \\times n}$ via a linear layer and apply scaling. Let $o := [1,0,\\dots,0]$ be the origin on the Lorentz manifold, satisfying $\\langle o, [0, x_{\\mathrm{in}}^{E}] \\rangle_{\\mathcal{L}} = 0$ . Thus, $[0, x_{\\mathrm{in}}^{E}]$ can be interpreted as a vector in the tangent space at $o$ . The Lorentz embedding is then obtained via the exponential map Eq. (4):", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {\\text {i n}} ^ {\\mathcal {L}} = \\exp_ {\\boldsymbol {o}} \\left(\\left[ 0, \\beta \\boldsymbol {x} _ {\\text {i n}} ^ {E} W _ {1} \\right]\\right) \\in \\mathbb {L} ^ {n}, \\mathbb {R} ^ {M \\times (n + 1)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 200, + 483, + 220 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{1}$ denotes the linear layer, $\\beta$ is a learnable scaling factor to prevent numerical overflow.", + "bbox": [ + 89, + 224, + 483, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Having obtained the Lorentz embedding $x_{\\mathrm{in}}^{\\mathcal{L}}$ , which inherently exhibits a prominent hierarchical structure due to the hyperbolic space properties, we next design a Lorentz linear transformation and Lorentz self-attention module to capture and fully leverage the hierarchical priors.", + "bbox": [ + 89, + 255, + 483, + 330 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Inspired by prior studies [6, 32], we redefine the Lorentz linear layer to learn a matrix $M = \\begin{bmatrix} \\boldsymbol{p}^{\\top} \\\\ \\boldsymbol{W} \\end{bmatrix}$ , where $\\boldsymbol{p} \\in \\mathbb{R}^{n+1}$ is a weight parameter and $\\boldsymbol{W} \\in \\mathbb{R}^{m \\times (n+1)}$ ensures that $\\forall \\boldsymbol{x} \\in \\mathbb{L}^n$ , $f_{\\boldsymbol{x}}(\\boldsymbol{M})\\boldsymbol{x} \\in \\mathbb{L}^m$ . Specifically, the transformation matrix $f_{\\boldsymbol{x}}(\\boldsymbol{M})$ is expressed as:", + "bbox": [ + 89, + 330, + 483, + 422 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\boldsymbol {x}} (\\boldsymbol {M}) = f _ {\\boldsymbol {x}} \\left(\\left[ \\begin{array}{c} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W} \\end{array} \\right]\\right) = \\left[ \\begin{array}{c} \\frac {\\sqrt {\\| W \\boldsymbol {x} \\| ^ {2} + 1}}{\\boldsymbol {p} ^ {\\top} \\boldsymbol {x}} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W}. \\end{array} \\right] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 428, + 483, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adding other components including normalization, the final definition of the Lorentz Linear layer becomes:", + "bbox": [ + 89, + 474, + 483, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} = \\operatorname {H L} (\\boldsymbol {x}) = \\left[ \\begin{array}{c} \\sqrt {\\| \\phi (\\boldsymbol {W} \\boldsymbol {x} , \\boldsymbol {p}) \\| ^ {2} + 1} \\\\ \\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) \\end{array} \\right], \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 508, + 483, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "with operation function:", + "bbox": [ + 89, + 541, + 253, + 555 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) = \\frac {\\lambda \\left(\\boldsymbol {p} ^ {\\top} \\boldsymbol {x} + b ^ {\\prime}\\right)}{\\| \\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b} \\|} (\\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b}), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 117, + 560, + 482, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{b}$ and $b'$ are bias terms, $\\lambda > 0$ regulates the scaling range. $h$ denotes the activation function.", + "bbox": [ + 89, + 599, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the Lorentz Linear Layer, we propose a Lorentz self-attention module that integrates Gaussian constraints into feature interactions, enabling multiscale and hierarchical video embeddings in hyperbolic space. Specifically, given a hyperbolic video embedding $\\pmb{x}_{\\mathrm{in}}^{\\mathcal{L}} \\in \\mathbb{L}^{n}$ , $\\mathbb{R}^{M \\times (n + 1)}$ , we first obtain the attention query $\\mathcal{Q}$ , key $\\mathcal{K}$ , and value $\\mathcal{V}$ using Eq. (13), all in the shape of $\\mathbb{R}^{M \\times (n + 1)}$ . We calculate attention scores based on Eq. (6) and apply a Gaussian matrix $\\mathcal{M}_{\\sigma}^{g} \\in \\mathbb{R}^{M \\times M}$ for element-wise multiplication with the score matrix to obtain a multi-scale receptive field. The output is defined as $\\pmb{x}_{\\mathrm{out}}^{\\mathcal{L}} = \\{\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_{|\\mathcal{Q}|}\\} \\in \\mathbb{R}^{M \\times (n + 1)}$ :", + "bbox": [ + 89, + 631, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS _ {i j} = \\frac {\\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , j)}{\\sqrt {(n + 1)}}\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {K} |} \\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {k}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , k)}{\\sqrt {(n + 1)}}\\right)}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 803, + 483, + 859 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\mu} _ {i} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {K} |} S _ {i j} \\boldsymbol {v} _ {j}}{\\left| \\left| \\sum_ {k = 1} ^ {| \\mathcal {K} |} S _ {i k} \\boldsymbol {v} _ {k} \\right| \\right| _ {\\mathcal {L}}} ,\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 863, + 331, + 905 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "the squared Lorentzian distance $d_{\\mathcal{L}}^2 (\\pmb {a},\\pmb {b}) = -2 - 2\\langle \\pmb {a},\\pmb {b}\\rangle_{\\mathcal{L}}$", + "bbox": [ + 511, + 90, + 906, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After computing $x_{\\mathrm{out}}^{\\mathcal{L}}$ , we apply the logarithmic map Eq. (5), while discarding the time axis, to obtain the Euclidean space embedding $x_{\\mathrm{mid}}^{E}$ . Then, the output $x_{\\mathrm{out}}^{E}$ is obtained through a Linear Layer followed by rescaling:", + "bbox": [ + 511, + 107, + 906, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {x} _ {\\text {m i d}} ^ {E} = \\operatorname {d r o p \\_ t i m e \\_ a x i s} \\left(\\log_ {\\boldsymbol {o}} \\left(\\boldsymbol {x} _ {\\text {o u t}} ^ {\\boldsymbol {\\mathcal {L}}}\\right)\\right) \\in \\mathbb {R} ^ {M \\times n}, \\\\ \\boldsymbol {x} _ {\\text {o u t}} ^ {E} = \\frac {\\boldsymbol {x} _ {\\text {m i d}} ^ {E} W _ {2}}{\\beta} \\in \\mathbb {R} ^ {M \\times d}, \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 178, + 906, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{2}\\in \\mathbb{R}^{n\\times d}$ , $\\beta$ is the scale factor in Eq. (11). Finally, We replace the self-attention in Transformer block with Lorentz attention to form the Lorentz Attention Block.", + "bbox": [ + 511, + 244, + 906, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mean-Guided Adaptive Interaction Module We arrange $N_{\\mathcal{L}}$ Lorentz and $N_{E}$ Euclidean Attention Blocks in parallel to construct $N_{O}$ Gaussian Attention Blocks for multi-scale hybrid-space video embeddings. To integrate these features, we introduce a Mean-Guided Adaptive Interaction Module, which utilizes globally pooled features to compute dynamic aggregation weights. Specifically, we first obtain the global query $\\varphi \\in \\mathbb{R}^{1\\times d}$ and compute aggregation weights via a Cross Attention Block consisting of a cross-attention layer (CA) followed by a fully connected layer (FC):", + "bbox": [ + 511, + 292, + 906, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {\\varphi} = \\operatorname {M e a n} \\left(\\boldsymbol {x} _ {\\sigma_ {1}}, \\boldsymbol {x} _ {\\sigma_ {2}},.., \\boldsymbol {x} _ {\\sigma_ {N _ {o}}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 457, + 774, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nw _ {i} = \\operatorname {F C} \\left(\\operatorname {C A} \\left(\\varphi , \\boldsymbol {x} _ {\\sigma_ {i}}, \\boldsymbol {x} _ {\\sigma_ {i}}\\right)\\right), i = 1, 2, \\dots , N _ {o},\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 474, + 859, + 492 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {w} _ {i, j} = \\frac {e ^ {w _ {i , j} / \\tau}}{\\sum_ {k = 1} ^ {N _ {o}} e ^ {w _ {k , j} / \\tau}}, j = 1, \\dots , M, \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 494, + 906, + 534 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {x}} _ {j} = \\sum_ {i = 1} ^ {N _ {o}} \\tilde {w} _ {i, j} \\boldsymbol {x} _ {\\sigma_ {i, j}}, j = 1, \\dots , M,\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 535, + 785, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {x} _ {\\text {M A I M}} = \\operatorname {C o n c a t} \\left(\\tilde {\\boldsymbol {x}} _ {1}, \\tilde {\\boldsymbol {x}} _ {2}, \\dots , \\tilde {\\boldsymbol {x}} _ {M}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 579, + 764, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{x}_{\\sigma_i} \\in \\mathbb{R}^{M \\times d}$ denotes the output of the $i$ -th Gaussian block and $M$ corresponds to the number of time points (i.e., clips or frames). $w_i \\in \\mathbb{R}^M$ represents the aggregation weights for the $i$ -th Gaussian block, and $\\tau$ is the temperature factor. $\\tilde{\\pmb{x}}_j \\in \\mathbb{R}^d$ denotes the aggregated feature at time point $j$ , while $\\pmb{x}_{\\mathrm{MAIM}}$ is the final output.", + "bbox": [ + 511, + 609, + 906, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Learning Objectives", + "text_level": 1, + "bbox": [ + 511, + 710, + 705, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the partial relevance in PRVR, where each video fully entails its corresponding text, a partial order relationship is established, with the text-query semantically subsumed by the video: text $\\prec$ video. Inspired by MERU [10], we propose the Partial Order Preservation Loss to enforce this relationship in Hyperbolic Space. Given $V_{f}$ and $V_{c}$ from Sec. 3.2, a simple attention module similar to Eq. (7) is applied, followed by mean pooling to get the unified video representation $V_{v}$ . The video and text representations are then mapped to Lorentz space via the exponential map, yielding $\\pmb{v},\\pmb{t}\\in \\mathbb{L}^n$ , as shown in Fig. 2(c). We define an entailment", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "23078", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cone for each $\\pmb{v}$ , which is characterized by the half-aperture:", + "bbox": [ + 89, + 90, + 485, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H A} (\\boldsymbol {v}) = \\arcsin \\left(\\frac {2 c}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\|}\\right). \\tag {18}\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 116, + 483, + 151 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$c = 0.1$ is used to define the boundary conditions near the origin. We measure the exterior angle $\\mathbf{EA}(\\boldsymbol{v},\\boldsymbol{t}) = \\pi - \\angle Ovt$ to penalize cases where $\\boldsymbol{t}$ falls outside the entailment cone:", + "bbox": [ + 89, + 160, + 483, + 205 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) = \\operatorname {a r c c o s} \\left(\\frac {t _ {0} + v _ {0} \\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\| \\sqrt {\\left(\\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {19}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 215, + 483, + 266 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The Loss for a single video-text pair is given by:", + "bbox": [ + 89, + 276, + 413, + 292 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p o p} (\\boldsymbol {v}, \\boldsymbol {t}) = \\max (0, \\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) - \\mathbf {H A} (\\boldsymbol {v})). \\tag {20}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 303, + 483, + 320 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Besides, following MS-SL [14], we use the standard similarity retrieval loss to train the model, denoted as $L_{sim}$ . Meanwhile, the query diversity [57] $L_{div}$ is used to improve retrieval performance. The aggregate loss is defined as:", + "bbox": [ + 89, + 330, + 485, + 391 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {a g g} = L _ {s i m} + \\lambda_ {1} L _ {d i v} + \\lambda_ {2} L _ {p o p}, \\tag {21}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 404, + 483, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "$\\lambda_{1}$ and $\\lambda_{2}$ are hyper-parameters that balance learning losses.", + "bbox": [ + 89, + 431, + 483, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 460, + 223, + 477 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 484, + 279, + 501 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets We conduct experiments on three benchmark datasets: (i) ActivityNet Captions [28], which comprises approximately 20K YouTube videos with an average duration of 118 seconds. Each video contains an average of 3.7 annotated moments with corresponding textual descriptions. (ii) TV show Retrieval (TVR) [30], consisting of 21.8K videos sourced from six TV shows. Each video is associated with five natural language descriptions covering different moments. (iii) Charades-STA [22], which includes 6,670 videos annotated with 16,128 sentence descriptions. On average, each video contains approximately 2.4 moments with corresponding textual queries. We adopt the same data split as used in prior studies[14, 57]. It is important to note that the moment annotations are unavailable in the PRVR task.", + "bbox": [ + 89, + 506, + 485, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Following previous works [14, 57], we adopt rank-based evaluation metrics, specifically $R@K$ ( $K = 1, 5, 10, 100$ ). The metric $R@K$ represents the proportion of queries for which the correct item appears within the top $K$ positions of the ranking list. All results are reported as percentages (\\%), where higher values indicate superior retrieval performance. To facilitate an overall comparison, we also report the Sum of all Recalls (SumR).", + "bbox": [ + 89, + 718, + 483, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 848, + 307, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data Processing For video representations on TVR, we employ the feature set provided by Lei et al. [30], which", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "comprises 3,072-dimensional visual features obtained by concatenating frame-level ResNet152 features [23] and segment-level I3D features [2]. For ActivityNet Captions and Charades-STA, we only utilize I3D features as provided by Zhang et al. [62] and Mun et al. [43], respectively. For sentence representations, we adopt the 768-dimensional RoBERTa features supplied by Lei et al. [30] for TVR. On ActivityNet Captions and Charades-STA, we employ 1,024-dimensional RoBERTa features extracted using MS-SL[14].", + "bbox": [ + 511, + 90, + 906, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Model Configurations The HLFormaler block consists of 8 Gaussian blocks $(N_O = 8)$ , 4 Lorentz Attention blocks $(N_{\\mathcal{L}} = 4)$ , with Gaussian variances ranging from $2^{1}$ to $2^{N_{\\mathcal{L}} - 1}$ and $\\infty$ , and 4 Euclidean Attention blocks $(N_E = 4)$ , with Gaussian variances ranging from $2^{1}$ to $2^{N_{E} - 1}$ and $\\infty$ . The latent dimension $d = 384$ with 4 attention heads.", + "bbox": [ + 511, + 227, + 908, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Configurations We employ the Adam optimizer with a mini-batch size of 128 and set the number of epochs to 100. The model is implemented using PyTorch and trained on one Nvidia RTX 3080 Ti GPU. We adopt a learning rate adjustment schedule similar to MS-SL.", + "bbox": [ + 511, + 319, + 906, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Comparison with State-of-the arts", + "text_level": 1, + "bbox": [ + 511, + 406, + 812, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines We select six representative PRVR baselines for comparison: MS-SL [14], PEAN [26], LH [20], BGM-Net [60], GMMFormer [57], and DL-DKD [16]. We also compare HLFomer with methods for T2VR and VCMR. For T2VR, we select six T2VR models: CE [37], HGR [5], DE++ [13], RIVRL [15], CLIP4Clip [40], Cap4Video [59], For VCMR, we consider four models: XML [30], ReLoCLNet [63], CONQUER [25] and JSG[7].", + "bbox": [ + 511, + 428, + 908, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Retrieval Performance Tab. 1 presents the retrieval performance of various models on three large-scale video datasets. As observed, T2VR models, designed to capture overall video-text relevance, underperform for PRVR. VCMR models, which focus on moment retrieval, achieve better results. PRVR methods perform best as they are specifically designed for this task. Attributed to hyperbolic space learning and effective utilization of video hierarchical structure priors, HLFformer consistently surpasses all baselines. It outperforms DL-DKD by $4.9\\%$ and $4.3\\%$ in SumR on ActivityNet Captions and TVR, respectively, and exceeds PEAN by $5.4\\%$ on Charades-STA.", + "bbox": [ + 511, + 549, + 908, + 731 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Model Analyses", + "text_level": 1, + "bbox": [ + 511, + 742, + 671, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Efficacy of Temporal Modeling Design We perform ablation studies to examine the effect of the attention block number $N_{o}$ and the attention mechanism ratio $N_{L} / N_{E}$ , with results shown in Fig. 3. Model performance improves as $N_{o}$ increases, then stabilizes or declines when $N_{o} \\geq 8$ . Even with only two attention blocks, HLFomer surpasses most competing methods. Furthermore, using solely Euclidean or Lorentz attention blocks results in suboptimal performance, whereas the hybrid attention block achieves the best results.", + "bbox": [ + 511, + 763, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "23079", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/7ab179770c422cbc6cd1ea48bd7ff400f23059a220a109d3ce568b30638f3186.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
T2VR
HGR [5]4.015.024.863.2107.01.23.87.333.445.71.74.98.335.250.1
RIVRL [15]5.218.028.266.4117.81.65.69.437.754.39.423.432.270.6135.6
DE++ [13]5.318.429.268.0121.01.75.69.637.154.18.821.930.267.4128.3
CE [37]5.519.129.971.1125.61.34.57.336.049.13.712.820.164.5101.1
CLIP4Clip [40]5.919.330.471.6127.31.86.510.944.263.49.924.334.372.5141.0
Cap4Video [59]6.320.430.972.6130.21.96.711.345.065.010.326.436.874.0147.5
VCMR
ReLoCLNet [63]5.718.930.072.0126.61.25.410.045.662.310.026.537.381.3155.1
XML [30]5.319.430.673.1128.41.66.010.146.964.610.728.138.180.3157.1
CONQUER [25]6.520.431.874.3133.11.86.310.347.566.011.028.939.681.3160.8
JSG [7]6.822.734.876.1140.52.47.712.849.872.7-----
PRVR
MS-SL [14]7.122.534.775.8140.11.87.111.847.768.413.532.143.483.4172.4
PEAN [26]7.423.035.575.9141.82.78.113.550.374.713.532.844.183.9174.2
LH [20]7.423.535.875.8142.42.17.512.950.172.713.233.244.485.5176.3
BGM-Net [60]7.223.836.076.9143.91.97.412.250.171.614.134.745.985.2179.9
GMMFormer [57]8.324.936.776.1146.02.17.812.550.672.913.933.344.584.9176.6
DL-DKD [16]8.025.037.577.1147.6-----14.434.945.884.9179.9
HLFormer (ours)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
", + "bbox": [ + 94, + 88, + 903, + 359 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/88811a4f4c6becef4ec94de9e1055bf5110304d6d8bd24ed4c8e95b9e5f8c8bf.jpg", + "table_caption": [ + "Table 1. Retrieval performance of HLFomer and other faithful methods on ActivityNet Captions, Charades-STA and TVR. State-of-the-art performance is highlighted in bold. “-” indicates that the corresponding results are unavailable." + ], + "table_footnote": [], + "table_body": "
IDModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
(0)HLFormer (full)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
Efficacy of Multi-scale Branches
(1)w/o gaze branch7.624.436.777.3146.11.88.013.950.874.513.934.045.285.3178.3
(2)w/o glance branch6.421.733.675.4137.21.67.713.148.470.811.430.541.882.4166.1
Efficacy of Different Loss Terms
(3)Lsim Only7.725.038.178.3149.12.08.113.252.075.315.136.247.886.0185.2
(4)w/o Ldiv8.526.639.678.8153.52.07.813.653.076.415.736.448.486.0186.5
(5)w/o Lpop8.626.939.778.8154.02.28.414.053.077.615.636.848.486.0186.8
Efficacy of various Aggregation Strategies
(6)w/ MP8.525.738.277.8150.22.08.013.252.175.315.236.547.486.0185.1
(7)w/ CL8.726.839.578.6153.62.08.213.952.076.115.336.948.486.0186.6
", + "bbox": [ + 94, + 412, + 903, + 584 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation Study of HLFomer. The best scores are marked in bold.", + "bbox": [ + 274, + 595, + 720, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This may be attributed to the differences in representational focus: Euclidean space emphasizes fine-grained local feature learning and sometimes overlooks global hierarchical structures, while hyperbolic space prioritizes global hierarchical relationships at the expense of local details. Moreover, hyperbolic space tends to be more sensitive to noise and numerically unstable. By integrating hybrid spaces, HLFormer achieves mutual compensation, enhancing representation learning and facilitating video semantic understanding.", + "bbox": [ + 88, + 636, + 482, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy of Hyperbolic Learning Hyperbolic learning demonstrates significant advantages in capturing the hierarchical structure of videos. As illustrated in Fig. 4(a), embeddings learned solely in Euclidean space exhibit indistinct cluster boundaries, with red and green points at the periphery closely interspersed. In contrast, Fig. 4(b) demonstrates that incorporating Lorentz attention facilitates the learning of more discriminative representations, while refining moment", + "bbox": [ + 88, + 779, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "cluster boundaries, increasing inter-moment separation, and compacting intra-moment frame distributions, revealing a more pronounced hierarchical structure.", + "bbox": [ + 511, + 636, + 903, + 680 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Efficacy of Multi-scale Branches To evaluate the effectiveness of the multi-scale branches, we conduct comparative experiments by removing either the glance clip-level branch or the gaze frame-level branch. As shown in Tab. 2, the absence of any branch leads to a noticeable performance degradation. These results not only validate the efficacy of the coarse-to-fine multi-granularity retrieval mechanism but also highlight the complementary nature of the two branches. Efficacies of Different Loss Terms To analyze the effectiveness of three loss terms (i.e. $L_{sim}$ , $L_{div}$ and $L_{pop}$ ) of HLFformer, we construct several HLFformer variants: (i) $L_{sim}$ Only: train the model with merely $L_{sim}$ . (ii) w/o $L_{div}$ : We train the model without query diverse learning. (iii) w/o $L_{pop}$ : HLFformer removes the partial order preservation task.", + "bbox": [ + 511, + 688, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "23080", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2bff62a8f9e2f8e23a6877da164dd58230f6a41f12d16ad61bba59a01564b888.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 102, + 89, + 281, + 157 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7e70021359ec2b0f54facb86ba45758fe6918dfcf734fcaba37433c427b018d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 292, + 90, + 470, + 156 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/44a9f3a7342add11f5b8893b13b4b9628917f888220e5ce513bf486665c6196d.jpg", + "image_caption": [ + "(a) ActivityNet Captions" + ], + "image_footnote": [], + "bbox": [ + 102, + 176, + 281, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/48c6f51ec532029942e43de228899d1de8cff8c74e4231bc338201247b35abf3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 294, + 176, + 470, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f3a4a1e670b44835a18105849a5dcc0124e2fdea302d1186cc924b7cb8cabe0e.jpg", + "image_caption": [ + "(b) Charades-STA" + ], + "image_footnote": [], + "bbox": [ + 102, + 263, + 281, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4be7380b2f802dd6b3975888c5fd189908fe19af24cad94e6ddf6231f09079fd.jpg", + "image_caption": [ + "(c) TVR" + ], + "image_footnote": [], + "bbox": [ + 294, + 263, + 470, + 328 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/49af380d9bb3f0b523b6cc8d740ff1b2ad2e194bce366a8d9212331623dba049.jpg", + "image_caption": [ + "Figure 3. The influence of different attention blocks, with default settings marked in bold.", + "(a) $w / o$ Lorentz Attention" + ], + "image_footnote": [], + "bbox": [ + 109, + 405, + 285, + 539 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/072ee8f433ac722281ac7aa476c465c1f75f70b5cebdf19a1f5e5f2a028acd94.jpg", + "image_caption": [ + "(b) $w/$ Lorentz Attention", + "Figure 4. The UMAP [41] visualization displays the learned frame embeddings from a video in TVR. Data points of the same color correspond to the same moment." + ], + "image_footnote": [], + "bbox": [ + 287, + 405, + 464, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Tab. 2, the worst performance occurs when only $L_{sim}$ is used. Comparing Variant (5) with Variant (3), adding $L_{div}$ increases the SumR, which can validate its necessity. Similarly, comparing Variant (4) with Variant (3) and Fig. 5, integrating $L_{pop}$ not only boosts retrieval accuracy but also ensures that the text query remains semantically embedded within the corresponding video, preserving partial relevance. Efficacy of Aggregation Strategy We compare three aggregation strategies: (i) $w/$ MP: mean pooling for static fusion. (ii) $w/$ CL: feature concatenation with linear layers. (iii) MAIM (default): mean-guided adaptive interaction module. As shown in Tab. 2, MP performs the worst due to its fixed static fusion, which limits semantic interaction. CL improves upon MP by leveraging linear layers for dynamic feature fusion. MAIM achieves the best performance by learning adaptive aggregation weights and dynamically selecting hyperbolic information under global guidance.", + "bbox": [ + 89, + 641, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e9ad8a8d7b285e1fe59545c9fa44c58a47e20dca76550d2c430c2e6e90a0c4d.jpg", + "image_caption": [ + "(a) $w / oL_{pop}$" + ], + "image_footnote": [], + "bbox": [ + 517, + 97, + 740, + 223 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a9d0c200c64728be273761a6d2756e584e898e07e438b6275623ba0445082acf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 746, + 88, + 903, + 222 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f23f2ef61f0a26d8c71cda61fecc309ebd0e86103a620bc4b41b3ce04330d01.jpg", + "image_caption": [ + "(b) $w / L_{pop}$" + ], + "image_footnote": [], + "bbox": [ + 517, + 247, + 741, + 375 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4f64357e5b574a934ff6fd896f9743c1fd144b39d82689d5cb707c6cc42abc31.jpg", + "image_caption": [ + "Figure 5. Visualization of the learned hyperbolic space. The closer to the origin, the higher semantic hierarchy and coarser granularity." + ], + "image_footnote": [], + "bbox": [ + 746, + 243, + 903, + 373 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Hyperbolic Space Inspired by HyCo-CLIP [46], we visualize the learned hyperbolic space by sampling 3K embeddings from the TVR training set. We analyze their norm distribution via histogram and reduce dimensionality using HoroPCA [3], as shown in Fig. 5. Glance branch embeddings are positioned closer to the origin than text query embeddings, indicating that clip-level video representations subsume textual queries. This phenomenon can be attributed to $L_{pop}$ , which enforces the partial order relationship between video and text representations. In contrast, without $L_{pop}$ , embeddings exhibit uncorrelated distributions. Moreover, text queries, being coarser in semantics, lie closer to the origin than fine-grained gaze-level embeddings, reflecting a clear hierarchical structure.", + "bbox": [ + 511, + 460, + 908, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 691, + 638, + 707 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose HLFormaler, a novel hyperbolic modeling framework tailored for PRVR. By leveraging the intrinsic geometric properties of hyperbolic space, HLFormaler effectively captures the hierarchical and multi-granular structure of untrimmed videos, thereby enhancing video-text retrieval accuracy. Furthermore, to ensure partial relevance between paired videos and text, a partial order preservation loss is introduced to enforce their semantic entailment. Extensive experiments indicate that HLFormaler consistently outperforms state-of-the-art methods. Our study offers a new perspective for PRVR with hyperbolic learning, which we hope will inspire further research in this direction.", + "bbox": [ + 511, + 719, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "23081", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. We sincerely thank the anonymous reviewers and chairs for their efforts and constructive suggestions, which have greatly helped us improve the manuscript. This work is supported in part by the National Natural Science Foundation of China under grant 624B2088, 62171248, 62301189, the PCNL KEY project (PCL2023AS6-1), and Shenzhen Science and Technology Program under Grant KJZD20240903103702004, JCYJ20220818101012025, GXWD20220811172936001. Long Chen was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), Huawei Gift Fund, and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04).", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 313, + 187, + 328 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 3", + "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6", + "[3] Ines Chami, Albert Gu, Dat Nguyen, and Christopher Ré. Horopca: Hyperbolic dimensionality reduction via horospherical projections, 2021. 8", + "[4] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275-1290, 2023. 3", + "[5] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 1, 2, 6, 7", + "[6] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. arXiv preprint arXiv:2105.14686, 2021. 3, 5", + "[7] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multi-granularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975–983, 2023. 2, 6, 7", + "[8] Dingxin Cheng, Shuhan Kong, Bin Jiang, and Qiang Guo. Transferable dual multi-granularity semantic excavating for partially relevant video retrieval. Image and Vision Computing, 149:105168, 2024. 2", + "[9] Cheol-Ho Cho, WonJun Moon, Woojin Jun, MinSeok Jung, and Jae-Pil Heo. Ambiguity-restrained text-video representation learning for partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2500-2508, 2025. 2", + "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic Image-" + ], + "bbox": [ + 93, + 338, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Text Representations. In Proceedings of the International Conference on Machine Learning, 2023. 2, 3, 5", + "[11] Jianfeng Dong, Xirong Li, and Cees GM Snoek. Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia, 20(12):3377-3388, 2018. 1, 2", + "[12] Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zeroexample video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2019.", + "[13] Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065-4080, 2021. 1, 6, 7", + "[14] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 246-257, 2022. 1, 2, 4, 6, 7", + "[15] Jianfeng Dong, Yabing Wang, Xianke Chen, Xiaoye Qu, Xirong Li, Yuan He, and Xun Wang. Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5680-5694, 2022. 1, 2, 6, 7", + "[16] Jianfeng Dong, Minsong Zhang, Zheng Zhang, Xianke Chen, Daizong Liu, Xiaoye Qu, Xun Wang, and Baolong Liu. Dual learning with dynamic knowledge distillation for partially relevant video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11302-11312, 2023. 2, 6, 7", + "[17] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399-7409, 2022. 2, 3", + "[18] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 1, 2", + "[19] Hao Fang, Changle Zhou, Jiawei Kong, Kuofeng Gao, Bin Chen, Tao Liang, Guojun Ma, and Shu-Tao Xia. Grounding language with vision: A conditional mutual information calibrated decoding strategy for reducing hallucinations in lvlms. arXiv preprint arXiv:2505.19678, 2025. 2", + "[20] Sheng Fang, Tiantian Dang, Shuhui Wang, and Qingming Huang. Linguistic hallucination for text-based video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 34(10):9692-9705, 2024. 6, 7", + "[21] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3", + "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 6", + "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "23082", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[24] Neil He, Menglin Yang, and Rex Ying. Lorentzian residual neural networks. arXiv preprint arXiv:2412.14695, 2024. 3", + "[25] Zhijian Hou, Chong-Wah Ngo, and Wing Kwong Chan. Conquer: Contextual query-aware ranking for video corpus moment retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3900-3908, 2021. 6, 7", + "[26] Xun Jiang, Zhiguo Chen, Xing Xu, Fumin Shen, Zuo Cao, and Xunliang Cai. Progressive event alignment network for partial relevant video retrieval. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1973-1978. IEEE, 2023. 2, 6, 7", + "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3", + "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 2, 6", + "[29] Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. In Proceedings of the 36th International Conference on Machine Learning, pages 3672-3681. PMLR, 2019. 2", + "[30] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 447-463. Springer, 2020. 2, 6, 7", + "[31] Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, and Xinbo Gao. Beyond euclidean: Dualspace representation learning for weakly supervised video violence detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3", + "[32] Keegan Lensink, Bas Peters, and Eldad Haber. Fully hyperbolic convolutional neural networks. Research in the Mathematical Sciences, 9(4):60, 2022. 3, 5", + "[33] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17511-17520, 2024. 3", + "[34] Xirong Li, Chaoxi Xu, Gang Yang, Zhineng Chen, and Jianfeng Dong. W2vv++ fully deep learning for ad-hoc video search. In Proceedings of the 27th ACM international conference on multimedia, pages 1786-1794, 2019. 1, 2", + "[35] Haitong Liu, Kuofeng Gao, Yang Bai, Jinmin Li, Jinxiao Shan, Tao Dai, and Shu-Tao Xia. Protecting your video content: Disrupting automated video-based lvm annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24056-24065, 2025. 2", + "[36] Peidong Liu, Dongliang Liao, Jinpeng Wang, Yangxin Wu, Gongfu Li, Shu-Tao Xia, and Jin Xu. Multi-task ranking with user behaviors for text-video search. In Companion Proceedings of the Web Conference 2022, pages 126-130, 2022. 2", + "[37] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2, 6, 7", + "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3", + "[39] Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1141-1150, 2020. 3", + "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 6, 7", + "[41] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 8", + "[42] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2", + "[43] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 6", + "[44] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3", + "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning, pages 3779-3788. PMLR, 2018. 3", + "[46] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In The Thirteenth International Conference on Learning Representations, 2025. 3, 8", + "[47] Xiaogang Peng, Hao Wen, Yikai Luo, Xiao Zhou, Keyang Yu, Yigang Wang, and Zizhao Wu. Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023. 3", + "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3", + "[49] Ruiqi Shi, Jun Wen, Wei Ji, Menglin Yang, Difei Gao, and Roger Zimmermann. HOVER: Hyperbolic video-text retrieval, 2024. 3", + "[50] Xue Song, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Spatial-temporal graphs for cross-modal text2video retrieval. IEEE Transactions on Multimedia, 24:2914-2923, 2021. 2", + "[51] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "23083", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "weakly-supervised video paragraph grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569-13580, 2024. 2", + "[52] Max Van Spengler, Erwin Berkhout, and Pascal Mettes. Poincaré resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 3", + "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3", + "[54] Jinpeng Wang, Bin Chen, Dongliang Liao, Ziyun Zeng, Gongfu Li, Shu-Tao Xia, and Jin Xu. Hybrid contrastive quantization for efficient cross-view video retrieval. In Proceedings of the ACM Web Conference 2022, pages 3020-3030, 2022. 2", + "[55] Jinpeng Wang, Ziyun Zeng, Bin Chen, Yuting Wang, Dongliang Liao, Gongfu Li, Yiru Wang, and Shu-Tao Xia. Hugs bring double benefits: Unsupervised cross-modal hashing with multi-granularity aligned transformers. International Journal of Computer Vision, 132(8):2765-2797, 2024. 2", + "[56] Yuting Wang, Jinpeng Wang, Bin Chen, Tao Dai, Ruisheng Luo, and Shu-Tao Xia. Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024. 1, 2", + "[57] Yuting Wang, Jinpeng Wang, Bin Chen, Ziyun Zeng, and Shu-Tao Xia. Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 4, 6, 7", + "[58] Zhihao Wang, Wenke Huang, Tian Chen, Zekun Shi, Guancheng Wan, Yu Qiao, Bin Yang, Jian Wang, Bing Li, and Mang Ye. An empirical study of federated prompt learning for vision language model. arXiv preprint arXiv:2505.23024, 2025. 2", + "[59] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704-10713, 2023. 6, 7", + "[60] Shukang Yin, Sirui Zhao, Hao Wang, Tong Xu, and Enhong Chen. Exploiting instance-level relationships in weakly supervised text-to-video retrieval. ACM Trans. Multim. Comput. Commun. Appl., 20(10):316:1-316:21, 2024. 2, 6, 7", + "[61] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3", + "[62] Bowen Zhang, Hexiang Hu, Joonseok Lee, Ming Zhao, Sheide Chammas, Vihan Jain, Eugene Ie, and Fei Sha. A hierarchical multi-modal encoder for moment localization in video corpus. arXiv preprint arXiv:2011.09046, 2020. 6", + "[63] Hao Zhang, Aixin Sun, Wei Jing, Guoshun Nan, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Video corpus moment retrieval with contrastive learning. In Proceedings of the 44th International ACM SIGIR Conference on Research" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "text", + "text": "and Development in Information Retrieval, pages 685-695, 2021. 6, 7", + "bbox": [ + 545, + 90, + 906, + 119 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "23084", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_model.json b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..990cd5bfe9eb36f2b62b340ffa8bd68fb776537f --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_model.json @@ -0,0 +1,2763 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.136, + 0.131, + 0.865, + 0.154 + ], + "angle": 0, + "content": "Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.18, + 0.765, + 0.217 + ], + "angle": 0, + "content": "Jun Li\\(^{1*}\\), Jinpeng Wang\\(^{2*†}\\), Chaolei Tan\\(^{4}\\), Niu Lian\\(^{1}\\), Long Chen\\(^{4}\\), Yaowei Wang\\(^{1,3}\\), Min Zhang\\(^{1}\\), Shu-Tao Xia\\(^{2,3}\\), Bin Chen\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.371, + 0.219, + 0.627, + 0.234 + ], + "angle": 0, + "content": "1Harbin Institute of Technology, Shenzhen" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.236, + 0.717, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)Tsinghua Shenzhen International Graduate School, Tsinghua University" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.254, + 0.7, + 0.27 + ], + "angle": 0, + "content": "3Research Center of Artificial Intelligence, Peng Cheng Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.333, + 0.272, + 0.666, + 0.288 + ], + "angle": 0, + "content": "4The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.254, + 0.29, + 0.739, + 0.305 + ], + "angle": 0, + "content": "220110924@stu.hit.edu.cn \\(\\text{四}\\) wjp20@mails.tsinghua.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.34, + 0.327, + 0.355 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.371, + 0.486, + 0.705 + ], + "angle": 0, + "content": "Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFformer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFformer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce \"text \\(\\prec\\) video\" hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFformer outperforms state-of-the-art methods. Code is released at https://github.com/lijun2005/ICCV25-HLFformer." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.734, + 0.222, + 0.749 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.758, + 0.485, + 0.866 + ], + "angle": 0, + "content": "Text-to-video retrieval (T2VR) [5, 11-13, 15, 18, 34, 37, 42] is a fundamental module in many search applications and a popular topic in multi-modal learning. While most T2VR models are developed for short clips or pre-trimmed video segments, they may face challenges where user queries describe only partial content in the video. This practical issue in real-world usage promotes a more challenging setting of" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.338, + 0.721, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.517, + 0.727, + 0.527 + ], + "angle": 0, + "content": "(a) Semantic Hierarchy in Untrimmed Videos" + }, + { + "type": "image", + "bbox": [ + 0.723, + 0.339, + 0.894, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.517, + 0.853, + 0.527 + ], + "angle": 0, + "content": "(c) Hyperbolic Space" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.539, + 0.908, + 0.622 + ], + "angle": 0, + "content": "Figure 1. (a) Modeling the semantic hierarchy in untrimmed videos helps Partially Relevant Video Retrieval (PRVR). (b) Euclidean space is less effective in modeling semantic hierarchy due to the flat geometry. Data points with distant hierarchical relation may be close. (c) Hyperbolic space allows larger cardinals when approaching the edge, which is preferable to preserve the hierarchy." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.656, + 0.907, + 0.686 + ], + "angle": 0, + "content": "partially relevant video retrieval (PRVR) [14], which aims to match each text query with the best untrimmed video." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.69, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Due to unlabeled moment timestamps, PRVR requires solid abilities on (i) identifying key moments in videos for extracting informative features and (ii) learning robust cross-modal representations to match text queries and videos precisely. Prior arts have developed preliminary solutions on both aspects, while challenges remain. For (i), MS-SL [14] exhaustively enumerated consecutive frame combinations through multi-scale sliding windows, which inevitably engaged redundancy, noise, and a high computational complexity in extracting moment features. GMMFormer [56, 57] improved efficiency by leveraging Gaussian neighborhood priors to traverse each timestamp and discover potential key moments. However, it may still be hard to distinguish adjacent or semantically similar candidate moments. Though" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23074" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "DL-DKD [16] neatly benefited from the pretrained CLIP [48] to enhance text-frame alignment, the temporal generalizability is bounded by the text-image teacher model. For (ii), most existing solutions inherited similar ideas from classic T2VR, e.g., ranking and contrastive learning, at a holistic level, but important characteristics of PRVR, e.g., partial relevance and semantic entailment, are still under-explored." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.202, + 0.487, + 0.504 + ], + "angle": 0, + "content": "In this paper, we take a hierarchical perspective to review the task, in the belief that videos naturally exhibit semantic hierarchy. As illustrated in Fig. 1(a), an untrimmed video can be regarded as a progression from frames to informative segments (e.g., Dunk), extended moments, and ultimately, the whole. Leveraging this intrinsic property is expected to benefit long video understanding. In particular for PRVR, the hierarchical prior provides positive guidance to arrange the moment features. Meanwhile, the supervisory signals from query-video matching can activate moment extraction more precisely through implicit bottom-up modeling. Exploring hierarchical features is never trivial. Unfortunately, existing PRVR approaches relying on Euclidean space are less effective in modeling the desired patterns in the flat geometry. We present Fig. 1(b) to exemplify this: two embeddings with a distant hierarchical relation may be spatially close to each other, as marked by the red arrows. Biased representation will increase the difficulty in disentangling informative moments from background, which limits the robustness in cross-modal matching considering partial relevance." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.509, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Inspired by the emerging success of hyperbolic learning [10, 17, 29, 31, 44], which takes advantage of exponentially expanding metric in non-Euclidean space to better capture hierarchical structure (Fig. 1(c)), we introduce HLFoRer, a sincere exploration of hyperbolic learning to enhance PRVR. On temporal modeling, we carefully design a dual-branch strategy to capture informative moment features comprehensively. Specifically, for the hyperbolic branch, we develop a Lorentz Attention Block (LAB) with the hyperbolic selfattention mechanism. With the implicit hierarchical prior through end-to-end matching optimization, LAB learns to activate informative moment features relevant to queries and distinguish them from noisy background in the hyperbolic space, compensating for the limitations of Euclidean attention in capturing hierarchical semantics. We integrate dual-branch moment features with a Mean-Guided Adaptive Interaction Module (MAIM), which is lightweight but effective. On cross-modal matching, drawing on the intrinsic \"text \\(\\prec\\) video\" hierarchy in PRVR where textual queries are subordinate to their paired videos, we introduce a Partial Order Preservation (POP) loss that geometrically confines text embeddings within hyperbolic cone anchored by corresponding video representations in an auxiliary Lorentzian manifold. This hierarchical metric alignment ensures semantic consistency between localized text semantics and their parent video structure while preserving partial relevance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.243 + ], + "angle": 0, + "content": "Empirical evaluations on three benchmark datasets: ActivityNet Captions [28], Charades-STA [22], and TVR [30] establish HLFoerner's state-of-the-art performance. Ablation studies confirm the necessity of hyperbolic geometry for hierarchical representation and the critical role of explicitly relational constraints in Partial Order Preservation Loss. Meanwhile, visual evidences further reveal that hyperbolic learning can enhance discriminative representation while maintaining video-text entailment, sharpening moment distinction and improving query alignment." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.244, + 0.908, + 0.259 + ], + "angle": 0, + "content": "The primary contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.32 + ], + "angle": 0, + "content": "- We propose to enhance PRVR with hyperbolic learning, including a Lorentz attention block with hierarchical priors to enhance the moment feature extraction, which collaborates with Euclidean attention and hybrid-space fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.321, + 0.907, + 0.365 + ], + "angle": 0, + "content": "- We design a partial order preservation loss that geometrically enforces the \"text \\( \\prec \\) video\" hierarchy through hyperbolic cone constraints, strengthening partial relevance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.907, + 0.412 + ], + "angle": 0, + "content": "- Extensive experiments on three benchmarks validate HLFformer's superiority, with analyses confirming the efficacy of hyperbolic modeling and geometric constraints." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.26, + 0.907, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.428, + 0.665, + 0.443 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.454, + 0.813, + 0.469 + ], + "angle": 0, + "content": "2.1. Partially Relevant Video Retrieval" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.909, + 0.689 + ], + "angle": 0, + "content": "With the growth of video content [19, 35, 58], video retrieval has become a key research area. Given a text query, Text-to-Video Retrieval (T2VR) [5, 11, 15, 18, 34, 36, 37, 42, 54, 55] focuses on retrieving fully relevant videos from pre-trimmed short clips. Video Corpus Moment Retrieval (VCMR) [7, 30, 50, 51] aims to localize specific moments within videos from a large corpus. Partially Relevant Video Retrieval (PRVR) [8, 9, 14, 16, 26, 56, 57, 60], a more recent task introduced by Dong et al. [14], aims to retrieve partially relevant videos from large, untrimmed long video collections. Unlike T2VR, PRVR must address the challenge of partial relevance, where the query pertains to only a specific moment of the video. Though the first stage of VCMR is similar to PRVR, VCMR requires moment-level annotations, limiting scalability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Existing methods enhance PRVR retrieval from various perspectives. MS-SL [14] defines the PRVR task as a Multi-instance Learning, providing a strong baseline with explicit redundant clip embeddings. GMMFormer [56, 57] and PEAN [26] propose implicit clip modeling to improve efficiency. DL-DKD [16] achieves great results through dynamic distillation of CLIP [48]. BGM-Net [60] exploits an instance-level matching scheme for pairing queries and videos. However, these methods predominantly rely on Euclidean space, which sometimes distort the hierarchical structures in untrimmed long videos. Consequently, they fail to fully exploit video hierarchy priors. To overcome this issue, we propose HLFormer to enhances PRVR by implicitly capturing hierarchical structures through hyperbolic learning." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "23075" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.29, + 0.108 + ], + "angle": 0, + "content": "2.2. Hyperbolic Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.486, + 0.431 + ], + "angle": 0, + "content": "Hyperbolic learning has attracted significant attention for its effectiveness in modeling hierarchical structures in real-world datasets. Early studies in computer vision tasks explored hyperbolic image embeddings from image-label pairs [27, 44], while subsequent progress extended hyperbolic optimization to multi-modal learning. MERU [10] and HyCoCLIP [46] notably surpassed Euclidean counterparts like CLIP [48] via hyperbolic space adaptation. Applications span semantic segmentation [1, 4], recognition tasks (skin [61], action [39]), meta-learning [17], and detection frameworks (violence [31, 47], anomalies [33]). Recent advances in fully hyperbolic neural networks [6, 21, 24, 32, 52] further underscore their potential. Motivated by them, we present the first study to explore the potential of hyperbolic learning for PRVR. Unlike other methods such as DSRL [31] and HOVER [49], our approach utilizes hyperbolic space to compensate for the limitations of Euclidean space in capturing the hierarchical structure of untrimmed long videos. Furthermore, we introduce the Partial Order Preservation Loss to explicitly capture the partial relevance between video and text in hyperbolic space, improving retrieval performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.444, + 0.182, + 0.459 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.232, + 0.484 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.582 + ], + "angle": 0, + "content": "Hyperbolic Space Hyperbolic spaces are Riemannian manifolds with a constant negative curvature \\(K\\), contrasting with the zero-curvature (flat) geometry of Euclidean spaces. Among several isometrically equivalent hyperbolic models, we adopt the Lorentz model [45] for its numerical stability and computational efficiency, with \\(K\\) set to -1 by default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.484, + 0.675 + ], + "angle": 0, + "content": "Lorentz Model Formally, an \\(n\\)-dimensional Lorentz model is the Riemannian manifold \\(\\mathbb{L}^n = (\\mathcal{L}^n,\\mathfrak{g}_x)\\). \\(\\mathfrak{g}_x = \\mathrm{diag}(-1,1,\\dots ,1)\\) is the Riemannian metric tensor. Each point in \\(\\mathbb{L}^n\\) has the form \\(\\pmb {x} = [x_0,\\pmb {x}_s]\\in \\mathbb{R}^{n + 1},x_0 = \\sqrt{\\|\\pmb{x}_s\\|^2 + 1}\\in \\mathbb{R}\\). Following Chen et al. [6], we denote \\(x_0\\) as time axis and \\(\\pmb{x}_{s}\\) as spatial axes. \\(\\mathcal{L}^n\\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.685, + 0.484, + 0.703 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {x} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {x}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = - 1, x _ {0} > 0 \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.714, + 0.376, + 0.73 + ], + "angle": 0, + "content": "and the Lorentzian inner product given by:" + }, + { + "type": "equation", + "bbox": [ + 0.194, + 0.74, + 0.484, + 0.758 + ], + "angle": 0, + "content": "\\[\n\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}} := - x _ {0} y _ {0} + \\boldsymbol {x} _ {s} ^ {\\top} \\boldsymbol {y} _ {s}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.485, + 0.8 + ], + "angle": 0, + "content": "Here \\(\\mathcal{L}^n\\) is the upper sheet of hyperboloid in a \\((n + 1)\\) dimensional Minkowski space with the origin \\(o = (1,0,\\dots ,0)\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.8, + 0.484, + 0.83 + ], + "angle": 0, + "content": "Tangent Space The tangent space at \\( \\pmb{x} \\in \\mathbb{L}^n \\) is a Euclidean space that is orthogonal to it, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.841, + 0.484, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\boldsymbol {x}} \\mathbb {L} ^ {n} := \\left\\{\\boldsymbol {y} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {y}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = 0 \\right\\}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Where \\(\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}\\) is a Euclidean subspace of \\(\\mathbb{R}^{n + 1}\\). In particular, the tangent space at the origin \\(\\pmb{o}\\) is denoted as \\(\\mathcal{T}_o\\mathbb{L}^n\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.168 + ], + "angle": 0, + "content": "Logarithmic and Exponential Maps The mutual mapping between the hyperbolic space \\(\\mathbb{L}^n\\) and the Euclidean subspace \\(\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}\\) can be realized by logarithmic and exponential maps. The exponential map \\(\\exp_{\\pmb{x}}(z)\\) can map any tangent vector \\(z\\in \\mathcal{T}_x\\mathbb{L}^n\\) to \\(\\mathbb{L}^n\\), written as:" + }, + { + "type": "equation", + "bbox": [ + 0.551, + 0.177, + 0.907, + 0.207 + ], + "angle": 0, + "content": "\\[\n\\exp_ {\\boldsymbol {x}} (\\boldsymbol {z}) = \\cosh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\boldsymbol {x} + \\sinh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\frac {\\boldsymbol {z}}{\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.218, + 0.907, + 0.25 + ], + "angle": 0, + "content": "where \\( \\| z\\|_{\\mathcal{L}} = \\sqrt{\\langle z,z\\rangle_{\\mathcal{L}}} \\) and the logarithmic map \\( \\log_{x}(\\pmb {y}) \\) plays an opposite role to map \\( \\pmb {y}\\in \\mathbb{L}^n \\) to \\( \\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n} \\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.259, + 0.907, + 0.295 + ], + "angle": 0, + "content": "\\[\n\\log_ {\\boldsymbol {x}} (\\boldsymbol {y}) = \\frac {\\operatorname {a r c o s h} (- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}})}{\\sqrt {(- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) ^ {2} - 1}} (\\boldsymbol {y} + (\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) \\boldsymbol {x}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.304, + 0.908, + 0.366 + ], + "angle": 0, + "content": "Lorentzian centroid The weighted centroid with respect to the squared Lorentzian distance, which solves \\(\\min_{\\mu \\in \\mathbb{L}^n}\\sum_{i = 1}^{m}\\nu_id^2_{\\mathcal{L}}(\\pmb {x}_i,\\mu)\\), with \\(\\pmb {x}_i\\in \\mathbb{L}^n\\) and \\(\\nu_{i}\\geq 0,\\sum_{i = 1}^{m}\\nu_{i} > 0\\), is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.375, + 0.907, + 0.412 + ], + "angle": 0, + "content": "\\[\n\\mu = \\frac {\\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i}}{| | | \\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i} | | _ {\\mathcal {L}} |}. \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.828, + 0.434 + ], + "angle": 0, + "content": "3.2. Problem Formulation and Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Partially Relevant Video Retrieval (PRVR) aims to retrieve videos containing a moment semantically relevant to a given text query, from a large corpus of untrimmed videos. In the PRVR database, each video has multiple moments and is associated with multiple text descriptions, with each text description corresponding to a specific moment of the related video. Critically, the temporal boundaries of these moments (i.e., start and end time points) are not annotated." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.638 + ], + "angle": 0, + "content": "In this paper, we introduce HLFormer, the first hyperbolic modeling approach designed for PRVR. The proposed framework encompasses three key components: text query representation encoding, video representation encoding, and similarity computation, as illustrated in Fig. 2 (a)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.638, + 0.909, + 0.761 + ], + "angle": 0, + "content": "Text Representation Given a text query of \\(N_{q}\\) words, we first use a pre-trained RoBERTa [38] model to extract word-level features, which are then projected into a lower-dimensional space via a fully connected (FC) layer. A standard Transformer [53] layer is applied to obtain a sequence of \\(d\\)-dimensional contextualized feature vectors, \\(\\pmb{Q} = \\{\\pmb{q}_i\\}_{i=1}^{N_q} \\in \\mathbb{R}^{N_q \\times d}\\). Finally, we utilize a simple attention mechanism to get the sentence embedding \\(\\pmb{q} \\in \\mathbb{R}^d\\):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.772, + 0.907, + 0.814 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} = \\sum_ {i = 1} ^ {N _ {q}} \\boldsymbol {a} _ {i} ^ {q} \\times \\boldsymbol {q} _ {i}, \\quad \\boldsymbol {a} ^ {q} = \\operatorname {s o f t m a x} (\\boldsymbol {w} \\boldsymbol {Q} ^ {\\top}), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.824, + 0.904, + 0.855 + ], + "angle": 0, + "content": "where \\(\\pmb{w} \\in \\mathbb{R}^{1 \\times d}\\) is a trainable vector, and \\(\\pmb{a}^q \\in \\mathbb{R}^{1 \\times N_q}\\) represents the attention vector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Video Representation Given an untrimmed video, we first extract embedding features using a pre-trained 2D or 3D CNN. Then we utilize the gaze branch and glance branch" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23076" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.695, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.322, + 0.471, + 0.335 + ], + "angle": 0, + "content": "(a) The overall framework of HLFormer" + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.089, + 0.903, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.697, + 0.213, + 0.903, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.347, + 0.907, + 0.43 + ], + "angle": 0, + "content": "Figure 2. Overview of HLFoer. (a) The sentence embedding \\( \\pmb{q} \\) is obtained via the query branch, while the gaze and glance branches encode the video, producing frame-level embedding \\( V_{f} \\) and clip-level embedding \\( V_{c} \\) and forming the video representation \\( V_{v} \\). \\( \\pmb{q} \\) learns query diversity through \\( L_{div} \\) and computes similarity scores \\( S_{f} \\) and \\( S_{c} \\), while preserving partial order relations with \\( V_{v} \\) using \\( L_{pop} \\). (b) HLFoer block combines parallel Lorentz and Euclidean attention blocks for multi-space encoding, with a Mean Guided Adaptive Interaction Module for dynamic aggregation. (c) Partial Order Preservation Loss ensures the text query embedding \\( t \\) lies within the cone defined by the video embedding \\( v \\). The loss is zero if \\( t \\) is inside the cone." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.457, + 0.486, + 0.716 + ], + "angle": 0, + "content": "to capture frame-level and clip-level multi-granularity video representations, respectively. In the gaze branch, we densely sample \\(M_{f}\\) frames, denoted as \\(\\pmb{F} \\in \\mathbb{R}^{M_f \\times D}\\), where \\(D\\) is the frame feature dimension. The sampled frames are processed through a fully connected (FC) layer to reduce the dimensionality to \\(d\\), followed by the HLFoer block to obtain frame embeddings \\(\\pmb{V}_{\\pmb{f}} = \\{\\pmb{f}_{i}\\}_{i=1}^{M_{f}} \\in \\mathbb{R}^{M_{f} \\times d}\\), capturing semantically rich frame-level information for fine-grained relevance assessment to the query. The glance branch down-samples the input along the temporal dimension to aggregate frames into clips. Following MS-SL [14], a fixed number \\(M_{c}\\) of clips is sparsely sampled by mean pooling over consecutive frames. A fully connected layer is applied to the pooled clip features, followed by the HLFoer block, generating clip embeddings \\(\\pmb{V}_{c} = \\{\\pmb{c}_{i}\\}_{i=1}^{M_{c}} \\in \\mathbb{R}^{M_{c} \\times d}\\). These embeddings capture adaptive clip-level information, enabling the model to perceive relevant moments at a coarser granularity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.485, + 0.793 + ], + "angle": 0, + "content": "Similarity Computation To compute the similarity between a text-video pair \\((\\mathcal{T},\\mathcal{V})\\) , we first measure the abovementioned embeddings \\(q,V_{f}\\) and \\(V_{c}\\) . Then, we employ cosine similarity along with a max operation to calculate the frame-level and clip-level similarity scores:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.806, + 0.484, + 0.829 + ], + "angle": 0, + "content": "\\[\nS _ {f} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {M _ {f}}\\right) \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.137, + 0.827, + 0.443, + 0.842 + ], + "angle": 0, + "content": "\\[\nS _ {c} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {1}), \\dots , \\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {M _ {c}}) \\right\\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.856, + 0.459, + 0.872 + ], + "angle": 0, + "content": "Next, we compute the overall text-video pair similarity:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.885, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\nS (\\mathcal {T}, \\mathcal {V}) = \\alpha_ {f} S _ {f} (\\mathcal {T}, \\mathcal {V}) + \\alpha_ {c} S _ {c} (\\mathcal {T}, \\mathcal {V}), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.457, + 0.907, + 0.503 + ], + "angle": 0, + "content": "where \\(\\alpha_{f},\\alpha_{c}\\in [0,1]\\) are hyper-parameters satisfying \\(\\alpha_{f} + \\alpha_{c} = 1\\). Finally, we retrieve and rank partially relevant videos based on the computed similarity scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.513, + 0.683, + 0.527 + ], + "angle": 0, + "content": "3.3. HlFormer Block" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.655 + ], + "angle": 0, + "content": "The HLFoer block constitutes the core of our method. As shown in Fig. 2 (b), it comprises three key modules: (i) Euclidean Attention Block, capturing fine-grained visual features in Euclidean space; (ii) Lorentz Attention Block, projecting video embeddings into hyperbolic Lorentz space for capturing the hierarchical structures of video; (iii) Mean-Guided Adaptive Interaction Module, dynamically fusing hybrid-space features. We describe the details below." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.656, + 0.906, + 0.717 + ], + "angle": 0, + "content": "Euclidean Attention Block Given M feature embeddings \\(\\pmb{x} \\in \\mathbb{R}^{M \\times d}\\), where \\(d\\) is the feature dimension, the Euclidean Attention Block utilizes Euclidean Gaussian Attention [57] to capture multi-scale visual features, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.728, + 0.907, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\operatorname {G A} (\\boldsymbol {x}) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {\\sigma} ^ {g} \\odot \\frac {\\boldsymbol {x} W ^ {q} \\left(\\boldsymbol {x} W ^ {k}\\right) ^ {\\top}}{\\sqrt {d _ {h}}}\\right) \\boldsymbol {x} W ^ {v}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.774, + 0.909, + 0.901 + ], + "angle": 0, + "content": "where \\(\\mathcal{M}_{\\sigma}^{g}\\) is the Gaussian matrix with elements \\(\\mathcal{M}_{\\sigma}^{g}(i,j) = \\frac{1}{2\\pi} e^{-\\frac{(j - i)^2}{\\sigma^2}}\\), and \\(\\sigma^2\\) denotes the variance. By varying \\(\\sigma\\), feature interactions at different scales are modeled, generating video features with multiple receptive fields. \\(W^{q}, W^{k}, W^{v}\\) are linear projections, while \\(d_{h}\\) is the latent attention dimension, \\(\\odot\\) denotes element-wise product. Finally, We replace the self-attention in Transformer block with Euclidean Gaussian attention to form the Euclidean Attention Block." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23077" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "Lorentz Attention Block Given extracted Euclidean video embeddings \\( \\pmb{x}_{\\mathrm{in}}^{E} \\in \\mathbb{R}^{M \\times d} \\), we first project it to \\( \\mathbb{R}^{M \\times n} \\) via a linear layer and apply scaling. Let \\( o := [1,0,\\dots,0] \\) be the origin on the Lorentz manifold, satisfying \\( \\langle o, [0, x_{\\mathrm{in}}^{E}] \\rangle_{\\mathcal{L}} = 0 \\). Thus, \\( [0, x_{\\mathrm{in}}^{E}] \\) can be interpreted as a vector in the tangent space at \\( o \\). The Lorentz embedding is then obtained via the exponential map Eq. (4):" + }, + { + "type": "equation", + "bbox": [ + 0.121, + 0.202, + 0.484, + 0.222 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {\\text {i n}} ^ {\\mathcal {L}} = \\exp_ {\\boldsymbol {o}} \\left(\\left[ 0, \\beta \\boldsymbol {x} _ {\\text {i n}} ^ {E} W _ {1} \\right]\\right) \\in \\mathbb {L} ^ {n}, \\mathbb {R} ^ {M \\times (n + 1)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.226, + 0.484, + 0.256 + ], + "angle": 0, + "content": "where \\( W_{1} \\) denotes the linear layer, \\( \\beta \\) is a learnable scaling factor to prevent numerical overflow." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.331 + ], + "angle": 0, + "content": "Having obtained the Lorentz embedding \\( x_{\\mathrm{in}}^{\\mathcal{L}} \\), which inherently exhibits a prominent hierarchical structure due to the hyperbolic space properties, we next design a Lorentz linear transformation and Lorentz self-attention module to capture and fully leverage the hierarchical priors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.424 + ], + "angle": 0, + "content": "Inspired by prior studies [6, 32], we redefine the Lorentz linear layer to learn a matrix \\( M = \\begin{bmatrix} \\boldsymbol{p}^{\\top} \\\\ \\boldsymbol{W} \\end{bmatrix} \\), where \\( \\boldsymbol{p} \\in \\mathbb{R}^{n+1} \\) is a weight parameter and \\( \\boldsymbol{W} \\in \\mathbb{R}^{m \\times (n+1)} \\) ensures that \\( \\forall \\boldsymbol{x} \\in \\mathbb{L}^n \\), \\( f_{\\boldsymbol{x}}(\\boldsymbol{M})\\boldsymbol{x} \\in \\mathbb{L}^m \\). Specifically, the transformation matrix \\( f_{\\boldsymbol{x}}(\\boldsymbol{M}) \\) is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.429, + 0.484, + 0.47 + ], + "angle": 0, + "content": "\\[\nf _ {\\boldsymbol {x}} (\\boldsymbol {M}) = f _ {\\boldsymbol {x}} \\left(\\left[ \\begin{array}{c} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W} \\end{array} \\right]\\right) = \\left[ \\begin{array}{c} \\frac {\\sqrt {\\| W \\boldsymbol {x} \\| ^ {2} + 1}}{\\boldsymbol {p} ^ {\\top} \\boldsymbol {x}} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W}. \\end{array} \\right] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.475, + 0.484, + 0.505 + ], + "angle": 0, + "content": "Adding other components including normalization, the final definition of the Lorentz Linear layer becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.51, + 0.484, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} = \\operatorname {H L} (\\boldsymbol {x}) = \\left[ \\begin{array}{c} \\sqrt {\\| \\phi (\\boldsymbol {W} \\boldsymbol {x} , \\boldsymbol {p}) \\| ^ {2} + 1} \\\\ \\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) \\end{array} \\right], \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.254, + 0.556 + ], + "angle": 0, + "content": "with operation function:" + }, + { + "type": "equation", + "bbox": [ + 0.118, + 0.561, + 0.483, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) = \\frac {\\lambda \\left(\\boldsymbol {p} ^ {\\top} \\boldsymbol {x} + b ^ {\\prime}\\right)}{\\| \\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b} \\|} (\\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b}), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.601, + 0.483, + 0.631 + ], + "angle": 0, + "content": "where \\( \\pmb{b} \\) and \\( b' \\) are bias terms, \\( \\lambda > 0 \\) regulates the scaling range. \\( h \\) denotes the activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.632, + 0.484, + 0.799 + ], + "angle": 0, + "content": "Based on the Lorentz Linear Layer, we propose a Lorentz self-attention module that integrates Gaussian constraints into feature interactions, enabling multiscale and hierarchical video embeddings in hyperbolic space. Specifically, given a hyperbolic video embedding \\(\\pmb{x}_{\\mathrm{in}}^{\\mathcal{L}} \\in \\mathbb{L}^{n}\\), \\(\\mathbb{R}^{M \\times (n + 1)}\\), we first obtain the attention query \\(\\mathcal{Q}\\), key \\(\\mathcal{K}\\), and value \\(\\mathcal{V}\\) using Eq. (13), all in the shape of \\(\\mathbb{R}^{M \\times (n + 1)}\\). We calculate attention scores based on Eq. (6) and apply a Gaussian matrix \\(\\mathcal{M}_{\\sigma}^{g} \\in \\mathbb{R}^{M \\times M}\\) for element-wise multiplication with the score matrix to obtain a multi-scale receptive field. The output is defined as \\(\\pmb{x}_{\\mathrm{out}}^{\\mathcal{L}} = \\{\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_{|\\mathcal{Q}|}\\} \\in \\mathbb{R}^{M \\times (n + 1)}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.804, + 0.484, + 0.861 + ], + "angle": 0, + "content": "\\[\nS _ {i j} = \\frac {\\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , j)}{\\sqrt {(n + 1)}}\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {K} |} \\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {k}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , k)}{\\sqrt {(n + 1)}}\\right)}, \\tag {15}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.864, + 0.332, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\mu} _ {i} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {K} |} S _ {i j} \\boldsymbol {v} _ {j}}{\\left| \\left| \\sum_ {k = 1} ^ {| \\mathcal {K} |} S _ {i k} \\boldsymbol {v} _ {k} \\right| \\right| _ {\\mathcal {L}}} ,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.908, + 0.108 + ], + "angle": 0, + "content": "the squared Lorentzian distance \\(d_{\\mathcal{L}}^2 (\\pmb {a},\\pmb {b}) = -2 - 2\\langle \\pmb {a},\\pmb {b}\\rangle_{\\mathcal{L}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.108, + 0.908, + 0.169 + ], + "angle": 0, + "content": "After computing \\( x_{\\mathrm{out}}^{\\mathcal{L}} \\), we apply the logarithmic map Eq. (5), while discarding the time axis, to obtain the Euclidean space embedding \\( x_{\\mathrm{mid}}^{E} \\). Then, the output \\( x_{\\mathrm{out}}^{E} \\) is obtained through a Linear Layer followed by rescaling:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.179, + 0.907, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {x} _ {\\text {m i d}} ^ {E} = \\operatorname {d r o p \\_ t i m e \\_ a x i s} \\left(\\log_ {\\boldsymbol {o}} \\left(\\boldsymbol {x} _ {\\text {o u t}} ^ {\\boldsymbol {\\mathcal {L}}}\\right)\\right) \\in \\mathbb {R} ^ {M \\times n}, \\\\ \\boldsymbol {x} _ {\\text {o u t}} ^ {E} = \\frac {\\boldsymbol {x} _ {\\text {m i d}} ^ {E} W _ {2}}{\\beta} \\in \\mathbb {R} ^ {M \\times d}, \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.908, + 0.292 + ], + "angle": 0, + "content": "where \\(W_{2}\\in \\mathbb{R}^{n\\times d}\\), \\(\\beta\\) is the scale factor in Eq. (11). Finally, We replace the self-attention in Transformer block with Lorentz attention to form the Lorentz Attention Block." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.293, + 0.908, + 0.445 + ], + "angle": 0, + "content": "Mean-Guided Adaptive Interaction Module We arrange \\(N_{\\mathcal{L}}\\) Lorentz and \\(N_{E}\\) Euclidean Attention Blocks in parallel to construct \\(N_{O}\\) Gaussian Attention Blocks for multi-scale hybrid-space video embeddings. To integrate these features, we introduce a Mean-Guided Adaptive Interaction Module, which utilizes globally pooled features to compute dynamic aggregation weights. Specifically, we first obtain the global query \\(\\varphi \\in \\mathbb{R}^{1\\times d}\\) and compute aggregation weights via a Cross Attention Block consisting of a cross-attention layer (CA) followed by a fully connected layer (FC):" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.458, + 0.776, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {\\varphi} = \\operatorname {M e a n} \\left(\\boldsymbol {x} _ {\\sigma_ {1}}, \\boldsymbol {x} _ {\\sigma_ {2}},.., \\boldsymbol {x} _ {\\sigma_ {N _ {o}}}\\right),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.476, + 0.86, + 0.493 + ], + "angle": 0, + "content": "\\[\nw _ {i} = \\operatorname {F C} \\left(\\operatorname {C A} \\left(\\varphi , \\boldsymbol {x} _ {\\sigma_ {i}}, \\boldsymbol {x} _ {\\sigma_ {i}}\\right)\\right), i = 1, 2, \\dots , N _ {o},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.495, + 0.907, + 0.535 + ], + "angle": 0, + "content": "\\[\n\\tilde {w} _ {i, j} = \\frac {e ^ {w _ {i , j} / \\tau}}{\\sum_ {k = 1} ^ {N _ {o}} e ^ {w _ {k , j} / \\tau}}, j = 1, \\dots , M, \\tag {17}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.536, + 0.787, + 0.576 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {x}} _ {j} = \\sum_ {i = 1} ^ {N _ {o}} \\tilde {w} _ {i, j} \\boldsymbol {x} _ {\\sigma_ {i, j}}, j = 1, \\dots , M,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.58, + 0.766, + 0.597 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {x} _ {\\text {M A I M}} = \\operatorname {C o n c a t} \\left(\\tilde {\\boldsymbol {x}} _ {1}, \\tilde {\\boldsymbol {x}} _ {2}, \\dots , \\tilde {\\boldsymbol {x}} _ {M}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.908, + 0.702 + ], + "angle": 0, + "content": "where \\(\\pmb{x}_{\\sigma_i} \\in \\mathbb{R}^{M \\times d}\\) denotes the output of the \\(i\\)-th Gaussian block and \\(M\\) corresponds to the number of time points (i.e., clips or frames). \\(w_i \\in \\mathbb{R}^M\\) represents the aggregation weights for the \\(i\\)-th Gaussian block, and \\(\\tau\\) is the temperature factor. \\(\\tilde{\\pmb{x}}_j \\in \\mathbb{R}^d\\) denotes the aggregated feature at time point \\(j\\), while \\(\\pmb{x}_{\\mathrm{MAIM}}\\) is the final output." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.706, + 0.729 + ], + "angle": 0, + "content": "3.4. Learning Objectives" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Given the partial relevance in PRVR, where each video fully entails its corresponding text, a partial order relationship is established, with the text-query semantically subsumed by the video: text \\(\\prec\\) video. Inspired by MERU [10], we propose the Partial Order Preservation Loss to enforce this relationship in Hyperbolic Space. Given \\(V_{f}\\) and \\(V_{c}\\) from Sec. 3.2, a simple attention module similar to Eq. (7) is applied, followed by mean pooling to get the unified video representation \\(V_{v}\\). The video and text representations are then mapped to Lorentz space via the exponential map, yielding \\(\\pmb{v},\\pmb{t}\\in \\mathbb{L}^n\\), as shown in Fig. 2(c). We define an entailment" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23078" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.108 + ], + "angle": 0, + "content": "cone for each \\(\\pmb{v}\\), which is characterized by the half-aperture:" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.117, + 0.484, + 0.152 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H A} (\\boldsymbol {v}) = \\arcsin \\left(\\frac {2 c}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\|}\\right). \\tag {18}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.161, + 0.485, + 0.207 + ], + "angle": 0, + "content": "\\(c = 0.1\\) is used to define the boundary conditions near the origin. We measure the exterior angle \\(\\mathbf{EA}(\\boldsymbol{v},\\boldsymbol{t}) = \\pi - \\angle Ovt\\) to penalize cases where \\(\\boldsymbol{t}\\) falls outside the entailment cone:" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.217, + 0.484, + 0.267 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) = \\operatorname {a r c c o s} \\left(\\frac {t _ {0} + v _ {0} \\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\| \\sqrt {\\left(\\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {19}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.277, + 0.414, + 0.293 + ], + "angle": 0, + "content": "The Loss for a single video-text pair is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.304, + 0.484, + 0.321 + ], + "angle": 0, + "content": "\\[\nL _ {p o p} (\\boldsymbol {v}, \\boldsymbol {t}) = \\max (0, \\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) - \\mathbf {H A} (\\boldsymbol {v})). \\tag {20}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.486, + 0.392 + ], + "angle": 0, + "content": "Besides, following MS-SL [14], we use the standard similarity retrieval loss to train the model, denoted as \\( L_{sim} \\). Meanwhile, the query diversity [57] \\( L_{div} \\) is used to improve retrieval performance. The aggregate loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.17, + 0.405, + 0.484, + 0.422 + ], + "angle": 0, + "content": "\\[\nL _ {a g g} = L _ {s i m} + \\lambda_ {1} L _ {d i v} + \\lambda_ {2} L _ {p o p}, \\tag {21}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.432, + 0.485, + 0.448 + ], + "angle": 0, + "content": "\\(\\lambda_{1}\\) and \\(\\lambda_{2}\\) are hyper-parameters that balance learning losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.461, + 0.224, + 0.478 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.281, + 0.502 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.486, + 0.717 + ], + "angle": 0, + "content": "Datasets We conduct experiments on three benchmark datasets: (i) ActivityNet Captions [28], which comprises approximately 20K YouTube videos with an average duration of 118 seconds. Each video contains an average of 3.7 annotated moments with corresponding textual descriptions. (ii) TV show Retrieval (TVR) [30], consisting of 21.8K videos sourced from six TV shows. Each video is associated with five natural language descriptions covering different moments. (iii) Charades-STA [22], which includes 6,670 videos annotated with 16,128 sentence descriptions. On average, each video contains approximately 2.4 moments with corresponding textual queries. We adopt the same data split as used in prior studies[14, 57]. It is important to note that the moment annotations are unavailable in the PRVR task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.719, + 0.485, + 0.84 + ], + "angle": 0, + "content": "Metrics Following previous works [14, 57], we adopt rank-based evaluation metrics, specifically \\( R@K \\) (\\( K = 1, 5, 10, 100 \\)). The metric \\( R@K \\) represents the proportion of queries for which the correct item appears within the top \\( K \\) positions of the ranking list. All results are reported as percentages (\\%), where higher values indicate superior retrieval performance. To facilitate an overall comparison, we also report the Sum of all Recalls (SumR)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.308, + 0.865 + ], + "angle": 0, + "content": "4.2. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Data Processing For video representations on TVR, we employ the feature set provided by Lei et al. [30], which" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.227 + ], + "angle": 0, + "content": "comprises 3,072-dimensional visual features obtained by concatenating frame-level ResNet152 features [23] and segment-level I3D features [2]. For ActivityNet Captions and Charades-STA, we only utilize I3D features as provided by Zhang et al. [62] and Mun et al. [43], respectively. For sentence representations, we adopt the 768-dimensional RoBERTa features supplied by Lei et al. [30] for TVR. On ActivityNet Captions and Charades-STA, we employ 1,024-dimensional RoBERTa features extracted using MS-SL[14]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.909, + 0.318 + ], + "angle": 0, + "content": "Model Configurations The HLFormaler block consists of 8 Gaussian blocks \\((N_O = 8)\\), 4 Lorentz Attention blocks \\((N_{\\mathcal{L}} = 4)\\), with Gaussian variances ranging from \\(2^{1}\\) to \\(2^{N_{\\mathcal{L}} - 1}\\) and \\(\\infty\\), and 4 Euclidean Attention blocks \\((N_E = 4)\\), with Gaussian variances ranging from \\(2^{1}\\) to \\(2^{N_{E} - 1}\\) and \\(\\infty\\). The latent dimension \\(d = 384\\) with 4 attention heads." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.32, + 0.908, + 0.395 + ], + "angle": 0, + "content": "Training Configurations We employ the Adam optimizer with a mini-batch size of 128 and set the number of epochs to 100. The model is implemented using PyTorch and trained on one Nvidia RTX 3080 Ti GPU. We adopt a learning rate adjustment schedule similar to MS-SL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.407, + 0.813, + 0.422 + ], + "angle": 0, + "content": "4.3. Comparison with State-of-the arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.429, + 0.909, + 0.549 + ], + "angle": 0, + "content": "Baselines We select six representative PRVR baselines for comparison: MS-SL [14], PEAN [26], LH [20], BGM-Net [60], GMMFormer [57], and DL-DKD [16]. We also compare HLFomer with methods for T2VR and VCMR. For T2VR, we select six T2VR models: CE [37], HGR [5], DE++ [13], RIVRL [15], CLIP4Clip [40], Cap4Video [59], For VCMR, we consider four models: XML [30], ReLoCLNet [63], CONQUER [25] and JSG[7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Retrieval Performance Tab. 1 presents the retrieval performance of various models on three large-scale video datasets. As observed, T2VR models, designed to capture overall video-text relevance, underperform for PRVR. VCMR models, which focus on moment retrieval, achieve better results. PRVR methods perform best as they are specifically designed for this task. Attributed to hyperbolic space learning and effective utilization of video hierarchical structure priors, HLFformer consistently surpasses all baselines. It outperforms DL-DKD by \\(4.9\\%\\) and \\(4.3\\%\\) in SumR on ActivityNet Captions and TVR, respectively, and exceeds PEAN by \\(5.4\\%\\) on Charades-STA." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.673, + 0.759 + ], + "angle": 0, + "content": "4.4. Model Analyses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Efficacy of Temporal Modeling Design We perform ablation studies to examine the effect of the attention block number \\( N_{o} \\) and the attention mechanism ratio \\( N_{L} / N_{E} \\), with results shown in Fig. 3. Model performance improves as \\( N_{o} \\) increases, then stabilizes or declines when \\( N_{o} \\geq 8 \\). Even with only two attention blocks, HLFomer surpasses most competing methods. Furthermore, using solely Euclidean or Lorentz attention blocks results in suboptimal performance, whereas the hybrid attention block achieves the best results." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23079" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.904, + 0.361 + ], + "angle": 0, + "content": "
ModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
T2VR
HGR [5]4.015.024.863.2107.01.23.87.333.445.71.74.98.335.250.1
RIVRL [15]5.218.028.266.4117.81.65.69.437.754.39.423.432.270.6135.6
DE++ [13]5.318.429.268.0121.01.75.69.637.154.18.821.930.267.4128.3
CE [37]5.519.129.971.1125.61.34.57.336.049.13.712.820.164.5101.1
CLIP4Clip [40]5.919.330.471.6127.31.86.510.944.263.49.924.334.372.5141.0
Cap4Video [59]6.320.430.972.6130.21.96.711.345.065.010.326.436.874.0147.5
VCMR
ReLoCLNet [63]5.718.930.072.0126.61.25.410.045.662.310.026.537.381.3155.1
XML [30]5.319.430.673.1128.41.66.010.146.964.610.728.138.180.3157.1
CONQUER [25]6.520.431.874.3133.11.86.310.347.566.011.028.939.681.3160.8
JSG [7]6.822.734.876.1140.52.47.712.849.872.7-----
PRVR
MS-SL [14]7.122.534.775.8140.11.87.111.847.768.413.532.143.483.4172.4
PEAN [26]7.423.035.575.9141.82.78.113.550.374.713.532.844.183.9174.2
LH [20]7.423.535.875.8142.42.17.512.950.172.713.233.244.485.5176.3
BGM-Net [60]7.223.836.076.9143.91.97.412.250.171.614.134.745.985.2179.9
GMMFormer [57]8.324.936.776.1146.02.17.812.550.672.913.933.344.584.9176.6
DL-DKD [16]8.025.037.577.1147.6-----14.434.945.884.9179.9
HLFormer (ours)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.372, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Table 1. Retrieval performance of HLFomer and other faithful methods on ActivityNet Captions, Charades-STA and TVR. State-of-the-art performance is highlighted in bold. “-” indicates that the corresponding results are unavailable." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.413, + 0.905, + 0.585 + ], + "angle": 0, + "content": "
IDModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
(0)HLFormer (full)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
Efficacy of Multi-scale Branches
(1)w/o gaze branch7.624.436.777.3146.11.88.013.950.874.513.934.045.285.3178.3
(2)w/o glance branch6.421.733.675.4137.21.67.713.148.470.811.430.541.882.4166.1
Efficacy of Different Loss Terms
(3)Lsim Only7.725.038.178.3149.12.08.113.252.075.315.136.247.886.0185.2
(4)w/o Ldiv8.526.639.678.8153.52.07.813.653.076.415.736.448.486.0186.5
(5)w/o Lpop8.626.939.778.8154.02.28.414.053.077.615.636.848.486.0186.8
Efficacy of various Aggregation Strategies
(6)w/ MP8.525.738.277.8150.22.08.013.252.175.315.236.547.486.0185.1
(7)w/ CL8.726.839.578.6153.62.08.213.952.076.115.336.948.486.0186.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.275, + 0.596, + 0.722, + 0.61 + ], + "angle": 0, + "content": "Table 2. Ablation Study of HLFomer. The best scores are marked in bold." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.637, + 0.483, + 0.773 + ], + "angle": 0, + "content": "This may be attributed to the differences in representational focus: Euclidean space emphasizes fine-grained local feature learning and sometimes overlooks global hierarchical structures, while hyperbolic space prioritizes global hierarchical relationships at the expense of local details. Moreover, hyperbolic space tends to be more sensitive to noise and numerically unstable. By integrating hybrid spaces, HLFormer achieves mutual compensation, enhancing representation learning and facilitating video semantic understanding." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Efficacy of Hyperbolic Learning Hyperbolic learning demonstrates significant advantages in capturing the hierarchical structure of videos. As illustrated in Fig. 4(a), embeddings learned solely in Euclidean space exhibit indistinct cluster boundaries, with red and green points at the periphery closely interspersed. In contrast, Fig. 4(b) demonstrates that incorporating Lorentz attention facilitates the learning of more discriminative representations, while refining moment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.905, + 0.681 + ], + "angle": 0, + "content": "cluster boundaries, increasing inter-moment separation, and compacting intra-moment frame distributions, revealing a more pronounced hierarchical structure." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Efficacy of Multi-scale Branches To evaluate the effectiveness of the multi-scale branches, we conduct comparative experiments by removing either the glance clip-level branch or the gaze frame-level branch. As shown in Tab. 2, the absence of any branch leads to a noticeable performance degradation. These results not only validate the efficacy of the coarse-to-fine multi-granularity retrieval mechanism but also highlight the complementary nature of the two branches. Efficacies of Different Loss Terms To analyze the effectiveness of three loss terms (i.e. \\( L_{sim} \\), \\( L_{div} \\) and \\( L_{pop} \\)) of HLFformer, we construct several HLFformer variants: (i) \\( L_{sim} \\) Only: train the model with merely \\( L_{sim} \\). (ii) w/o \\( L_{div} \\): We train the model without query diverse learning. (iii) w/o \\( L_{pop} \\): HLFformer removes the partial order preservation task." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "23080" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.09, + 0.282, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.293, + 0.092, + 0.471, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.228, + 0.165, + 0.347, + 0.176 + ], + "angle": 0, + "content": "(a) ActivityNet Captions" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.177, + 0.282, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.177, + 0.471, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.238, + 0.252, + 0.336, + 0.263 + ], + "angle": 0, + "content": "(b) Charades-STA" + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.264, + 0.282, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.295, + 0.264, + 0.471, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.264, + 0.336, + 0.311, + 0.348 + ], + "angle": 0, + "content": "(c) TVR" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.358, + 0.483, + 0.387 + ], + "angle": 0, + "content": "Figure 3. The influence of different attention blocks, with default settings marked in bold." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.406, + 0.287, + 0.54 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.121, + 0.544, + 0.276, + 0.557 + ], + "angle": 0, + "content": "(a) \\(w / o\\) Lorentz Attention" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.406, + 0.465, + 0.541 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.302, + 0.544, + 0.451, + 0.557 + ], + "angle": 0, + "content": "(b) \\(w/\\) Lorentz Attention" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.611 + ], + "angle": 0, + "content": "Figure 4. The UMAP [41] visualization displays the learned frame embeddings from a video in TVR. Data points of the same color correspond to the same moment." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.485, + 0.903 + ], + "angle": 0, + "content": "As shown in Tab. 2, the worst performance occurs when only \\( L_{sim} \\) is used. Comparing Variant (5) with Variant (3), adding \\( L_{div} \\) increases the SumR, which can validate its necessity. Similarly, comparing Variant (4) with Variant (3) and Fig. 5, integrating \\( L_{pop} \\) not only boosts retrieval accuracy but also ensures that the text query remains semantically embedded within the corresponding video, preserving partial relevance. Efficacy of Aggregation Strategy We compare three aggregation strategies: (i) \\( w/ \\) MP: mean pooling for static fusion. (ii) \\( w/ \\) CL: feature concatenation with linear layers. (iii) MAIM (default): mean-guided adaptive interaction module. As shown in Tab. 2, MP performs the worst due to its fixed static fusion, which limits semantic interaction. CL improves upon MP by leveraging linear layers for dynamic feature fusion. MAIM achieves the best performance by learning adaptive aggregation weights and dynamically selecting hyperbolic information under global guidance." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.098, + 0.741, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.674, + 0.228, + 0.747, + 0.242 + ], + "angle": 0, + "content": "(a) \\(w / oL_{pop}\\)" + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.089, + 0.905, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.248, + 0.742, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.747, + 0.244, + 0.904, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.676, + 0.379, + 0.744, + 0.394 + ], + "angle": 0, + "content": "(b) \\(w / L_{pop}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.404, + 0.907, + 0.432 + ], + "angle": 0, + "content": "Figure 5. Visualization of the learned hyperbolic space. The closer to the origin, the higher semantic hierarchy and coarser granularity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.462, + 0.909, + 0.675 + ], + "angle": 0, + "content": "Visualization of Hyperbolic Space Inspired by HyCo-CLIP [46], we visualize the learned hyperbolic space by sampling 3K embeddings from the TVR training set. We analyze their norm distribution via histogram and reduce dimensionality using HoroPCA [3], as shown in Fig. 5. Glance branch embeddings are positioned closer to the origin than text query embeddings, indicating that clip-level video representations subsume textual queries. This phenomenon can be attributed to \\( L_{pop} \\), which enforces the partial order relationship between video and text representations. In contrast, without \\( L_{pop} \\), embeddings exhibit uncorrelated distributions. Moreover, text queries, being coarser in semantics, lie closer to the origin than fine-grained gaze-level embeddings, reflecting a clear hierarchical structure." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.693, + 0.64, + 0.708 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose HLFormaler, a novel hyperbolic modeling framework tailored for PRVR. By leveraging the intrinsic geometric properties of hyperbolic space, HLFormaler effectively captures the hierarchical and multi-granular structure of untrimmed videos, thereby enhancing video-text retrieval accuracy. Furthermore, to ensure partial relevance between paired videos and text, a partial order preservation loss is introduced to enforce their semantic entailment. Extensive experiments indicate that HLFormaler consistently outperforms state-of-the-art methods. Our study offers a new perspective for PRVR with hyperbolic learning, which we hope will inspire further research in this direction." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "23081" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.288 + ], + "angle": 0, + "content": "Acknowledgments. We sincerely thank the anonymous reviewers and chairs for their efforts and constructive suggestions, which have greatly helped us improve the manuscript. This work is supported in part by the National Natural Science Foundation of China under grant 624B2088, 62171248, 62301189, the PCNL KEY project (PCL2023AS6-1), and Shenzhen Science and Technology Program under Grant KJZD20240903103702004, JCYJ20220818101012025, GXWD20220811172936001. Long Chen was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), Huawei Gift Fund, and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.314, + 0.188, + 0.329 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.339, + 0.486, + 0.407 + ], + "angle": 0, + "content": "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.409, + 0.484, + 0.464 + ], + "angle": 0, + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.484, + 0.506 + ], + "angle": 0, + "content": "[3] Ines Chami, Albert Gu, Dat Nguyen, and Christopher Ré. Horopca: Hyperbolic dimensionality reduction via horospherical projections, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.484, + 0.561 + ], + "angle": 0, + "content": "[4] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275-1290, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[5] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.675 + ], + "angle": 0, + "content": "[6] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. arXiv preprint arXiv:2105.14686, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[7] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multi-granularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975–983, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.747, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[8] Dingxin Cheng, Shuhan Kong, Bin Jiang, and Qiang Guo. Transferable dual multi-granularity semantic excavating for partially relevant video retrieval. Image and Vision Computing, 149:105168, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[9] Cheol-Ho Cho, WonJun Moon, Woojin Jun, MinSeok Jung, and Jae-Pil Heo. Ambiguity-restrained text-video representation learning for partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2500-2508, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic Image-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.339, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "Text Representations. In Proceedings of the International Conference on Machine Learning, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.176 + ], + "angle": 0, + "content": "[11] Jianfeng Dong, Xirong Li, and Cees GM Snoek. Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia, 20(12):3377-3388, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[12] Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zeroexample video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.906, + 0.305 + ], + "angle": 0, + "content": "[13] Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065-4080, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.907, + 0.373 + ], + "angle": 0, + "content": "[14] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 246-257, 2022. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.907, + 0.446 + ], + "angle": 0, + "content": "[15] Jianfeng Dong, Yabing Wang, Xianke Chen, Xiaoye Qu, Xirong Li, Yuan He, and Xun Wang. Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5680-5694, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.448, + 0.907, + 0.53 + ], + "angle": 0, + "content": "[16] Jianfeng Dong, Minsong Zhang, Zheng Zhang, Xianke Chen, Daizong Liu, Xiaoye Qu, Xun Wang, and Baolong Liu. Dual learning with dynamic knowledge distillation for partially relevant video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11302-11312, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.532, + 0.907, + 0.601 + ], + "angle": 0, + "content": "[17] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399-7409, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.603, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[18] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[19] Hao Fang, Changle Zhou, Jiawei Kong, Kuofeng Gao, Bin Chen, Tao Liang, Guojun Ma, and Shu-Tao Xia. Grounding language with vision: A conditional mutual information calibrated decoding strategy for reducing hallucinations in lvlms. arXiv preprint arXiv:2505.19678, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.907, + 0.771 + ], + "angle": 0, + "content": "[20] Sheng Fang, Tiantian Dang, Shuhui Wang, and Qingming Huang. Linguistic hallucination for text-based video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 34(10):9692-9705, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.907, + 0.815 + ], + "angle": 0, + "content": "[21] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23082" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.121, + 0.483, + 0.147 + ], + "angle": 0, + "content": "[24] Neil He, Menglin Yang, and Rex Ying. Lorentzian residual neural networks. arXiv preprint arXiv:2412.14695, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[25] Zhijian Hou, Chong-Wah Ngo, and Wing Kwong Chan. Conquer: Contextual query-aware ranking for video corpus moment retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3900-3908, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.205, + 0.484, + 0.272 + ], + "angle": 0, + "content": "[26] Xun Jiang, Zhiguo Chen, Xing Xu, Fumin Shen, Zuo Cao, and Xunliang Cai. Progressive event alignment network for partial relevant video retrieval. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1973-1978. IEEE, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.483, + 0.328 + ], + "angle": 0, + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[29] Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. In Proceedings of the 36th International Conference on Machine Learning, pages 3672-3681. PMLR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[30] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 447-463. Springer, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[31] Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, and Xinbo Gao. Beyond euclidean: Dualspace representation learning for weakly supervised video violence detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.621 + ], + "angle": 0, + "content": "[32] Keegan Lensink, Bas Peters, and Eldad Haber. Fully hyperbolic convolutional neural networks. Research in the Mathematical Sciences, 9(4):60, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.677 + ], + "angle": 0, + "content": "[33] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17511-17520, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[34] Xirong Li, Chaoxi Xu, Gang Yang, Zhineng Chen, and Jianfeng Dong. W2vv++ fully deep learning for ad-hoc video search. In Proceedings of the 27th ACM international conference on multimedia, pages 1786-1794, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.803 + ], + "angle": 0, + "content": "[35] Haitong Liu, Kuofeng Gao, Yang Bai, Jinmin Li, Jinxiao Shan, Tao Dai, and Shu-Tao Xia. Protecting your video content: Disrupting automated video-based lvm annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24056-24065, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[36] Peidong Liu, Dongliang Liao, Jinpeng Wang, Yangxin Wu, Gongfu Li, Shu-Tao Xia, and Jin Xu. Multi-task ranking with user behaviors for text-video search. In Companion Proceedings of the Web Conference 2022, pages 126-130, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.907, + 0.188 + ], + "angle": 0, + "content": "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.907, + 0.245 + ], + "angle": 0, + "content": "[39] Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1141-1150, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.907, + 0.301 + ], + "angle": 0, + "content": "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.907, + 0.344 + ], + "angle": 0, + "content": "[41] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.345, + 0.907, + 0.425 + ], + "angle": 0, + "content": "[42] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.907, + 0.482 + ], + "angle": 0, + "content": "[43] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.906, + 0.537 + ], + "angle": 0, + "content": "[44] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning, pages 3779-3788. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.907, + 0.663 + ], + "angle": 0, + "content": "[46] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In The Thirteenth International Conference on Learning Representations, 2025. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[47] Xiaogang Peng, Hao Wen, Yikai Luo, Xiao Zhou, Keyang Yu, Yigang Wang, and Zizhao Wu. Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.907, + 0.787 + ], + "angle": 0, + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.907, + 0.829 + ], + "angle": 0, + "content": "[49] Ruiqi Shi, Jun Wen, Wei Ji, Menglin Yang, Difei Gao, and Roger Zimmermann. HOVER: Hyperbolic video-text retrieval, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[50] Xue Song, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Spatial-temporal graphs for cross-modal text2video retrieval. IEEE Transactions on Multimedia, 24:2914-2923, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[51] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "23083" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "weakly-supervised video paragraph grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569-13580, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[52] Max Van Spengler, Erwin Berkhout, and Pascal Mettes. Poincaré resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.248 + ], + "angle": 0, + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[54] Jinpeng Wang, Bin Chen, Dongliang Liao, Ziyun Zeng, Gongfu Li, Shu-Tao Xia, and Jin Xu. Hybrid contrastive quantization for efficient cross-view video retrieval. In Proceedings of the ACM Web Conference 2022, pages 3020-3030, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.484, + 0.39 + ], + "angle": 0, + "content": "[55] Jinpeng Wang, Ziyun Zeng, Bin Chen, Yuting Wang, Dongliang Liao, Gongfu Li, Yiru Wang, and Shu-Tao Xia. Hugs bring double benefits: Unsupervised cross-modal hashing with multi-granularity aligned transformers. International Journal of Computer Vision, 132(8):2765-2797, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.484, + 0.433 + ], + "angle": 0, + "content": "[56] Yuting Wang, Jinpeng Wang, Bin Chen, Tao Dai, Ruisheng Luo, and Shu-Tao Xia. Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.484, + 0.503 + ], + "angle": 0, + "content": "[57] Yuting Wang, Jinpeng Wang, Bin Chen, Ziyun Zeng, and Shu-Tao Xia. Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.484, + 0.573 + ], + "angle": 0, + "content": "[58] Zhihao Wang, Wenke Huang, Tian Chen, Zekun Shi, Guancheng Wan, Yu Qiao, Bin Yang, Jian Wang, Bing Li, and Mang Ye. An empirical study of federated prompt learning for vision language model. arXiv preprint arXiv:2505.23024, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[59] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704-10713, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[60] Shukang Yin, Sirui Zhao, Hao Wang, Tong Xu, and Enhong Chen. Exploiting instance-level relationships in weakly supervised text-to-video retrieval. ACM Trans. Multim. Comput. Commun. Appl., 20(10):316:1-316:21, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[61] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[62] Bowen Zhang, Hexiang Hu, Joonseok Lee, Ming Zhao, Sheide Chammas, Vihan Jain, Eugene Ie, and Fei Sha. A hierarchical multi-modal encoder for moment localization in video corpus. arXiv preprint arXiv:2011.09046, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[63] Hao Zhang, Aixin Sun, Wei Jing, Guoshun Nan, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Video corpus moment retrieval with contrastive learning. In Proceedings of the 44th International ACM SIGIR Conference on Research" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.547, + 0.092, + 0.907, + 0.12 + ], + "angle": 0, + "content": "and Development in Information Retrieval, pages 685-695, 2021. 6, 7" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "23084" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_origin.pdf b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7bb043a14927e6e2aeccc9b050bcd875eb20a1ca --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/8267e0d4-8636-4d06-abe0-cf72edf7f96f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:32e2e59d0285100e9ca5ee79808ab18d0e37f68e9e4656328630a14b71e36356 +size 1420879 diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/full.md b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ae6b1577d58948f0f680460b532ee3e851fd4332 --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/full.md @@ -0,0 +1,428 @@ +# Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning + +Jun Li $^{1*}$ , Jinpeng Wang $^{2*†}$ , Chaolei Tan $^{4}$ , Niu Lian $^{1}$ , Long Chen $^{4}$ , Yaowei Wang $^{1,3}$ , Min Zhang $^{1}$ , Shu-Tao Xia $^{2,3}$ , Bin Chen $^{1}$ + +1Harbin Institute of Technology, Shenzhen + +$^{2}$ Tsinghua Shenzhen International Graduate School, Tsinghua University + +3Research Center of Artificial Intelligence, Peng Cheng Laboratory + +4The Hong Kong University of Science and Technology + +220110924@stu.hit.edu.cn $\text{四}$ wjp20@mails.tsinghua.edu.cn + +# Abstract + +Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFformer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFformer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce "text $\prec$ video" hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFformer outperforms state-of-the-art methods. Code is released at https://github.com/lijun2005/ICCV25-HLFformer. + +# 1. Introduction + +Text-to-video retrieval (T2VR) [5, 11-13, 15, 18, 34, 37, 42] is a fundamental module in many search applications and a popular topic in multi-modal learning. While most T2VR models are developed for short clips or pre-trimmed video segments, they may face challenges where user queries describe only partial content in the video. This practical issue in real-world usage promotes a more challenging setting of + +![](images/2f6a4a5389cca2054b178cee763893f730f3b0be262e16d79f80f1f782230c13.jpg) +(a) Semantic Hierarchy in Untrimmed Videos +Figure 1. (a) Modeling the semantic hierarchy in untrimmed videos helps Partially Relevant Video Retrieval (PRVR). (b) Euclidean space is less effective in modeling semantic hierarchy due to the flat geometry. Data points with distant hierarchical relation may be close. (c) Hyperbolic space allows larger cardinals when approaching the edge, which is preferable to preserve the hierarchy. + +![](images/e73c642f342019f007b74c6fbe378b48667eaee066e9d0b296b2ed2216bba876.jpg) +(c) Hyperbolic Space + +partially relevant video retrieval (PRVR) [14], which aims to match each text query with the best untrimmed video. + +Due to unlabeled moment timestamps, PRVR requires solid abilities on (i) identifying key moments in videos for extracting informative features and (ii) learning robust cross-modal representations to match text queries and videos precisely. Prior arts have developed preliminary solutions on both aspects, while challenges remain. For (i), MS-SL [14] exhaustively enumerated consecutive frame combinations through multi-scale sliding windows, which inevitably engaged redundancy, noise, and a high computational complexity in extracting moment features. GMMFormer [56, 57] improved efficiency by leveraging Gaussian neighborhood priors to traverse each timestamp and discover potential key moments. However, it may still be hard to distinguish adjacent or semantically similar candidate moments. Though + +DL-DKD [16] neatly benefited from the pretrained CLIP [48] to enhance text-frame alignment, the temporal generalizability is bounded by the text-image teacher model. For (ii), most existing solutions inherited similar ideas from classic T2VR, e.g., ranking and contrastive learning, at a holistic level, but important characteristics of PRVR, e.g., partial relevance and semantic entailment, are still under-explored. + +In this paper, we take a hierarchical perspective to review the task, in the belief that videos naturally exhibit semantic hierarchy. As illustrated in Fig. 1(a), an untrimmed video can be regarded as a progression from frames to informative segments (e.g., Dunk), extended moments, and ultimately, the whole. Leveraging this intrinsic property is expected to benefit long video understanding. In particular for PRVR, the hierarchical prior provides positive guidance to arrange the moment features. Meanwhile, the supervisory signals from query-video matching can activate moment extraction more precisely through implicit bottom-up modeling. Exploring hierarchical features is never trivial. Unfortunately, existing PRVR approaches relying on Euclidean space are less effective in modeling the desired patterns in the flat geometry. We present Fig. 1(b) to exemplify this: two embeddings with a distant hierarchical relation may be spatially close to each other, as marked by the red arrows. Biased representation will increase the difficulty in disentangling informative moments from background, which limits the robustness in cross-modal matching considering partial relevance. + +Inspired by the emerging success of hyperbolic learning [10, 17, 29, 31, 44], which takes advantage of exponentially expanding metric in non-Euclidean space to better capture hierarchical structure (Fig. 1(c)), we introduce HLFoRer, a sincere exploration of hyperbolic learning to enhance PRVR. On temporal modeling, we carefully design a dual-branch strategy to capture informative moment features comprehensively. Specifically, for the hyperbolic branch, we develop a Lorentz Attention Block (LAB) with the hyperbolic selfattention mechanism. With the implicit hierarchical prior through end-to-end matching optimization, LAB learns to activate informative moment features relevant to queries and distinguish them from noisy background in the hyperbolic space, compensating for the limitations of Euclidean attention in capturing hierarchical semantics. We integrate dual-branch moment features with a Mean-Guided Adaptive Interaction Module (MAIM), which is lightweight but effective. On cross-modal matching, drawing on the intrinsic "text $\prec$ video" hierarchy in PRVR where textual queries are subordinate to their paired videos, we introduce a Partial Order Preservation (POP) loss that geometrically confines text embeddings within hyperbolic cone anchored by corresponding video representations in an auxiliary Lorentzian manifold. This hierarchical metric alignment ensures semantic consistency between localized text semantics and their parent video structure while preserving partial relevance. + +Empirical evaluations on three benchmark datasets: ActivityNet Captions [28], Charades-STA [22], and TVR [30] establish HLFoerner's state-of-the-art performance. Ablation studies confirm the necessity of hyperbolic geometry for hierarchical representation and the critical role of explicitly relational constraints in Partial Order Preservation Loss. Meanwhile, visual evidences further reveal that hyperbolic learning can enhance discriminative representation while maintaining video-text entailment, sharpening moment distinction and improving query alignment. + +The primary contributions can be summarized as follows: + +- We propose to enhance PRVR with hyperbolic learning, including a Lorentz attention block with hierarchical priors to enhance the moment feature extraction, which collaborates with Euclidean attention and hybrid-space fusion. +- We design a partial order preservation loss that geometrically enforces the "text $\prec$ video" hierarchy through hyperbolic cone constraints, strengthening partial relevance. +- Extensive experiments on three benchmarks validate HLFformer's superiority, with analyses confirming the efficacy of hyperbolic modeling and geometric constraints. + +# 2. Related Works + +# 2.1. Partially Relevant Video Retrieval + +With the growth of video content [19, 35, 58], video retrieval has become a key research area. Given a text query, Text-to-Video Retrieval (T2VR) [5, 11, 15, 18, 34, 36, 37, 42, 54, 55] focuses on retrieving fully relevant videos from pre-trimmed short clips. Video Corpus Moment Retrieval (VCMR) [7, 30, 50, 51] aims to localize specific moments within videos from a large corpus. Partially Relevant Video Retrieval (PRVR) [8, 9, 14, 16, 26, 56, 57, 60], a more recent task introduced by Dong et al. [14], aims to retrieve partially relevant videos from large, untrimmed long video collections. Unlike T2VR, PRVR must address the challenge of partial relevance, where the query pertains to only a specific moment of the video. Though the first stage of VCMR is similar to PRVR, VCMR requires moment-level annotations, limiting scalability. + +Existing methods enhance PRVR retrieval from various perspectives. MS-SL [14] defines the PRVR task as a Multi-instance Learning, providing a strong baseline with explicit redundant clip embeddings. GMMFormer [56, 57] and PEAN [26] propose implicit clip modeling to improve efficiency. DL-DKD [16] achieves great results through dynamic distillation of CLIP [48]. BGM-Net [60] exploits an instance-level matching scheme for pairing queries and videos. However, these methods predominantly rely on Euclidean space, which sometimes distort the hierarchical structures in untrimmed long videos. Consequently, they fail to fully exploit video hierarchy priors. To overcome this issue, we propose HLFormer to enhances PRVR by implicitly capturing hierarchical structures through hyperbolic learning. + +# 2.2. Hyperbolic Learning + +Hyperbolic learning has attracted significant attention for its effectiveness in modeling hierarchical structures in real-world datasets. Early studies in computer vision tasks explored hyperbolic image embeddings from image-label pairs [27, 44], while subsequent progress extended hyperbolic optimization to multi-modal learning. MERU [10] and HyCoCLIP [46] notably surpassed Euclidean counterparts like CLIP [48] via hyperbolic space adaptation. Applications span semantic segmentation [1, 4], recognition tasks (skin [61], action [39]), meta-learning [17], and detection frameworks (violence [31, 47], anomalies [33]). Recent advances in fully hyperbolic neural networks [6, 21, 24, 32, 52] further underscore their potential. Motivated by them, we present the first study to explore the potential of hyperbolic learning for PRVR. Unlike other methods such as DSRL [31] and HOVER [49], our approach utilizes hyperbolic space to compensate for the limitations of Euclidean space in capturing the hierarchical structure of untrimmed long videos. Furthermore, we introduce the Partial Order Preservation Loss to explicitly capture the partial relevance between video and text in hyperbolic space, improving retrieval performance. + +# 3. Method + +# 3.1. Preliminaries + +Hyperbolic Space Hyperbolic spaces are Riemannian manifolds with a constant negative curvature $K$ , contrasting with the zero-curvature (flat) geometry of Euclidean spaces. Among several isometrically equivalent hyperbolic models, we adopt the Lorentz model [45] for its numerical stability and computational efficiency, with $K$ set to -1 by default. + +Lorentz Model Formally, an $n$ -dimensional Lorentz model is the Riemannian manifold $\mathbb{L}^n = (\mathcal{L}^n,\mathfrak{g}_x)$ . $\mathfrak{g}_x = \mathrm{diag}(-1,1,\dots ,1)$ is the Riemannian metric tensor. Each point in $\mathbb{L}^n$ has the form $\pmb {x} = [x_0,\pmb {x}_s]\in \mathbb{R}^{n + 1},x_0 = \sqrt{\|\pmb{x}_s\|^2 + 1}\in \mathbb{R}$ . Following Chen et al. [6], we denote $x_0$ as time axis and $\pmb{x}_{s}$ as spatial axes. $\mathcal{L}^n$ is given by: + +$$ +\mathcal {L} ^ {n} := \left\{\boldsymbol {x} \in \mathbb {R} ^ {n + 1} \mid \langle \boldsymbol {x}, \boldsymbol {x} \rangle_ {\mathcal {L}} = - 1, x _ {0} > 0 \right\}, \tag {1} +$$ + +and the Lorentzian inner product given by: + +$$ +\langle \boldsymbol {x}, \boldsymbol {y} \rangle_ {\mathcal {L}} := - x _ {0} y _ {0} + \boldsymbol {x} _ {s} ^ {\top} \boldsymbol {y} _ {s}. \tag {2} +$$ + +Here $\mathcal{L}^n$ is the upper sheet of hyperboloid in a $(n + 1)$ dimensional Minkowski space with the origin $o = (1,0,\dots ,0)$ . + +Tangent Space The tangent space at $\pmb{x} \in \mathbb{L}^n$ is a Euclidean space that is orthogonal to it, defined as: + +$$ +\mathcal {T} _ {\boldsymbol {x}} \mathbb {L} ^ {n} := \left\{\boldsymbol {y} \in \mathbb {R} ^ {n + 1} \mid \langle \boldsymbol {y}, \boldsymbol {x} \rangle_ {\mathcal {L}} = 0 \right\}. \tag {3} +$$ + +Where $\mathcal{T}_{\pmb{x}}\mathbb{L}^{n}$ is a Euclidean subspace of $\mathbb{R}^{n + 1}$ . In particular, the tangent space at the origin $\pmb{o}$ is denoted as $\mathcal{T}_o\mathbb{L}^n$ + +Logarithmic and Exponential Maps The mutual mapping between the hyperbolic space $\mathbb{L}^n$ and the Euclidean subspace $\mathcal{T}_{\pmb{x}}\mathbb{L}^{n}$ can be realized by logarithmic and exponential maps. The exponential map $\exp_{\pmb{x}}(z)$ can map any tangent vector $z\in \mathcal{T}_x\mathbb{L}^n$ to $\mathbb{L}^n$ , written as: + +$$ +\exp_ {\boldsymbol {x}} (\boldsymbol {z}) = \cosh (\| \boldsymbol {z} \| _ {\mathcal {L}}) \boldsymbol {x} + \sinh (\| \boldsymbol {z} \| _ {\mathcal {L}}) \frac {\boldsymbol {z}}{\| \boldsymbol {z} \| _ {\mathcal {L}}}, \tag {4} +$$ + +where $\| z\|_{\mathcal{L}} = \sqrt{\langle z,z\rangle_{\mathcal{L}}}$ and the logarithmic map $\log_{x}(\pmb {y})$ plays an opposite role to map $\pmb {y}\in \mathbb{L}^n$ to $\mathcal{T}_{\pmb{x}}\mathbb{L}^{n}$ as follows: + +$$ +\log_ {\boldsymbol {x}} (\boldsymbol {y}) = \frac {\operatorname {a r c o s h} (- \langle \boldsymbol {x} , \boldsymbol {y} \rangle_ {\mathcal {L}})}{\sqrt {(- \langle \boldsymbol {x} , \boldsymbol {y} \rangle_ {\mathcal {L}}) ^ {2} - 1}} (\boldsymbol {y} + (\langle \boldsymbol {x}, \boldsymbol {y} \rangle_ {\mathcal {L}}) \boldsymbol {x}). \tag {5} +$$ + +Lorentzian centroid The weighted centroid with respect to the squared Lorentzian distance, which solves $\min_{\mu \in \mathbb{L}^n}\sum_{i = 1}^{m}\nu_id^2_{\mathcal{L}}(\pmb {x}_i,\mu)$ , with $\pmb {x}_i\in \mathbb{L}^n$ and $\nu_{i}\geq 0,\sum_{i = 1}^{m}\nu_{i} > 0$ , is denoted as: + +$$ +\mu = \frac {\sum_ {i = 1} ^ {m} \nu_ {i} \boldsymbol {x} _ {i}}{| | | \sum_ {i = 1} ^ {m} \nu_ {i} \boldsymbol {x} _ {i} | | _ {\mathcal {L}} |}. \tag {6} +$$ + +# 3.2. Problem Formulation and Overview + +Partially Relevant Video Retrieval (PRVR) aims to retrieve videos containing a moment semantically relevant to a given text query, from a large corpus of untrimmed videos. In the PRVR database, each video has multiple moments and is associated with multiple text descriptions, with each text description corresponding to a specific moment of the related video. Critically, the temporal boundaries of these moments (i.e., start and end time points) are not annotated. + +In this paper, we introduce HLFormer, the first hyperbolic modeling approach designed for PRVR. The proposed framework encompasses three key components: text query representation encoding, video representation encoding, and similarity computation, as illustrated in Fig. 2 (a). + +Text Representation Given a text query of $N_{q}$ words, we first use a pre-trained RoBERTa [38] model to extract word-level features, which are then projected into a lower-dimensional space via a fully connected (FC) layer. A standard Transformer [53] layer is applied to obtain a sequence of $d$ -dimensional contextualized feature vectors, $\pmb{Q} = \{\pmb{q}_i\}_{i=1}^{N_q} \in \mathbb{R}^{N_q \times d}$ . Finally, we utilize a simple attention mechanism to get the sentence embedding $\pmb{q} \in \mathbb{R}^d$ : + +$$ +\boldsymbol {q} = \sum_ {i = 1} ^ {N _ {q}} \boldsymbol {a} _ {i} ^ {q} \times \boldsymbol {q} _ {i}, \quad \boldsymbol {a} ^ {q} = \operatorname {s o f t m a x} (\boldsymbol {w} \boldsymbol {Q} ^ {\top}), \tag {7} +$$ + +where $\pmb{w} \in \mathbb{R}^{1 \times d}$ is a trainable vector, and $\pmb{a}^q \in \mathbb{R}^{1 \times N_q}$ represents the attention vector. + +Video Representation Given an untrimmed video, we first extract embedding features using a pre-trained 2D or 3D CNN. Then we utilize the gaze branch and glance branch + +![](images/c3e6a22ba70d5d2764040c9022422b886c27c2307b8912ffdaf9dcc57322a9ed.jpg) +(a) The overall framework of HLFormer + +![](images/651766bd5169af4403a9cd6563ecd8fc36a254a7ea4d66a9ca316711673281b8.jpg) + +![](images/b5da0877f37d3f9a127bd3ee0399aeb98ed138a9f9ed5aef29270ea1dc8edff9.jpg) +Figure 2. Overview of HLFoer. (a) The sentence embedding $\pmb{q}$ is obtained via the query branch, while the gaze and glance branches encode the video, producing frame-level embedding $V_{f}$ and clip-level embedding $V_{c}$ and forming the video representation $V_{v}$ . $\pmb{q}$ learns query diversity through $L_{div}$ and computes similarity scores $S_{f}$ and $S_{c}$ , while preserving partial order relations with $V_{v}$ using $L_{pop}$ . (b) HLFoer block combines parallel Lorentz and Euclidean attention blocks for multi-space encoding, with a Mean Guided Adaptive Interaction Module for dynamic aggregation. (c) Partial Order Preservation Loss ensures the text query embedding $t$ lies within the cone defined by the video embedding $v$ . The loss is zero if $t$ is inside the cone. + +to capture frame-level and clip-level multi-granularity video representations, respectively. In the gaze branch, we densely sample $M_{f}$ frames, denoted as $\pmb{F} \in \mathbb{R}^{M_f \times D}$ , where $D$ is the frame feature dimension. The sampled frames are processed through a fully connected (FC) layer to reduce the dimensionality to $d$ , followed by the HLFoer block to obtain frame embeddings $\pmb{V}_{\pmb{f}} = \{\pmb{f}_{i}\}_{i=1}^{M_{f}} \in \mathbb{R}^{M_{f} \times d}$ , capturing semantically rich frame-level information for fine-grained relevance assessment to the query. The glance branch down-samples the input along the temporal dimension to aggregate frames into clips. Following MS-SL [14], a fixed number $M_{c}$ of clips is sparsely sampled by mean pooling over consecutive frames. A fully connected layer is applied to the pooled clip features, followed by the HLFoer block, generating clip embeddings $\pmb{V}_{c} = \{\pmb{c}_{i}\}_{i=1}^{M_{c}} \in \mathbb{R}^{M_{c} \times d}$ . These embeddings capture adaptive clip-level information, enabling the model to perceive relevant moments at a coarser granularity. + +Similarity Computation To compute the similarity between a text-video pair $(\mathcal{T},\mathcal{V})$ , we first measure the abovementioned embeddings $q,V_{f}$ and $V_{c}$ . Then, we employ cosine similarity along with a max operation to calculate the frame-level and clip-level similarity scores: + +$$ +S _ {f} (\mathcal {T}, \mathcal {V}) = \max \left\{\cos \left(\boldsymbol {q}, \boldsymbol {f} _ {1}\right), \dots , \cos \left(\boldsymbol {q}, \boldsymbol {f} _ {M _ {f}}\right) \right\}, \tag {8} +$$ + +$$ +S _ {c} (\mathcal {T}, \mathcal {V}) = \max \left\{\cos (\boldsymbol {q}, \boldsymbol {c} _ {1}), \dots , \cos (\boldsymbol {q}, \boldsymbol {c} _ {M _ {c}}) \right\}. +$$ + +Next, we compute the overall text-video pair similarity: + +$$ +S (\mathcal {T}, \mathcal {V}) = \alpha_ {f} S _ {f} (\mathcal {T}, \mathcal {V}) + \alpha_ {c} S _ {c} (\mathcal {T}, \mathcal {V}), \tag {9} +$$ + +where $\alpha_{f},\alpha_{c}\in [0,1]$ are hyper-parameters satisfying $\alpha_{f} + \alpha_{c} = 1$ . Finally, we retrieve and rank partially relevant videos based on the computed similarity scores. + +# 3.3. HlFormer Block + +The HLFoer block constitutes the core of our method. As shown in Fig. 2 (b), it comprises three key modules: (i) Euclidean Attention Block, capturing fine-grained visual features in Euclidean space; (ii) Lorentz Attention Block, projecting video embeddings into hyperbolic Lorentz space for capturing the hierarchical structures of video; (iii) Mean-Guided Adaptive Interaction Module, dynamically fusing hybrid-space features. We describe the details below. + +Euclidean Attention Block Given M feature embeddings $\pmb{x} \in \mathbb{R}^{M \times d}$ , where $d$ is the feature dimension, the Euclidean Attention Block utilizes Euclidean Gaussian Attention [57] to capture multi-scale visual features, expressed as: + +$$ +\operatorname {G A} (\boldsymbol {x}) = \operatorname {s o f t m a x} \left(\mathcal {M} _ {\sigma} ^ {g} \odot \frac {\boldsymbol {x} W ^ {q} \left(\boldsymbol {x} W ^ {k}\right) ^ {\top}}{\sqrt {d _ {h}}}\right) \boldsymbol {x} W ^ {v}, \tag {10} +$$ + +where $\mathcal{M}_{\sigma}^{g}$ is the Gaussian matrix with elements $\mathcal{M}_{\sigma}^{g}(i,j) = \frac{1}{2\pi} e^{-\frac{(j - i)^2}{\sigma^2}}$ , and $\sigma^2$ denotes the variance. By varying $\sigma$ , feature interactions at different scales are modeled, generating video features with multiple receptive fields. $W^{q}, W^{k}, W^{v}$ are linear projections, while $d_{h}$ is the latent attention dimension, $\odot$ denotes element-wise product. Finally, We replace the self-attention in Transformer block with Euclidean Gaussian attention to form the Euclidean Attention Block. + +Lorentz Attention Block Given extracted Euclidean video embeddings $\pmb{x}_{\mathrm{in}}^{E} \in \mathbb{R}^{M \times d}$ , we first project it to $\mathbb{R}^{M \times n}$ via a linear layer and apply scaling. Let $o := [1,0,\dots,0]$ be the origin on the Lorentz manifold, satisfying $\langle o, [0, x_{\mathrm{in}}^{E}] \rangle_{\mathcal{L}} = 0$ . Thus, $[0, x_{\mathrm{in}}^{E}]$ can be interpreted as a vector in the tangent space at $o$ . The Lorentz embedding is then obtained via the exponential map Eq. (4): + +$$ +\boldsymbol {x} _ {\text {i n}} ^ {\mathcal {L}} = \exp_ {\boldsymbol {o}} \left(\left[ 0, \beta \boldsymbol {x} _ {\text {i n}} ^ {E} W _ {1} \right]\right) \in \mathbb {L} ^ {n}, \mathbb {R} ^ {M \times (n + 1)}, \tag {11} +$$ + +where $W_{1}$ denotes the linear layer, $\beta$ is a learnable scaling factor to prevent numerical overflow. + +Having obtained the Lorentz embedding $x_{\mathrm{in}}^{\mathcal{L}}$ , which inherently exhibits a prominent hierarchical structure due to the hyperbolic space properties, we next design a Lorentz linear transformation and Lorentz self-attention module to capture and fully leverage the hierarchical priors. + +Inspired by prior studies [6, 32], we redefine the Lorentz linear layer to learn a matrix $M = \begin{bmatrix} \boldsymbol{p}^{\top} \\ \boldsymbol{W} \end{bmatrix}$ , where $\boldsymbol{p} \in \mathbb{R}^{n+1}$ is a weight parameter and $\boldsymbol{W} \in \mathbb{R}^{m \times (n+1)}$ ensures that $\forall \boldsymbol{x} \in \mathbb{L}^n$ , $f_{\boldsymbol{x}}(\boldsymbol{M})\boldsymbol{x} \in \mathbb{L}^m$ . Specifically, the transformation matrix $f_{\boldsymbol{x}}(\boldsymbol{M})$ is expressed as: + +$$ +f _ {\boldsymbol {x}} (\boldsymbol {M}) = f _ {\boldsymbol {x}} \left(\left[ \begin{array}{c} \boldsymbol {p} ^ {\top} \\ \boldsymbol {W} \end{array} \right]\right) = \left[ \begin{array}{c} \frac {\sqrt {\| W \boldsymbol {x} \| ^ {2} + 1}}{\boldsymbol {p} ^ {\top} \boldsymbol {x}} \boldsymbol {p} ^ {\top} \\ \boldsymbol {W}. \end{array} \right] \tag {12} +$$ + +Adding other components including normalization, the final definition of the Lorentz Linear layer becomes: + +$$ +\boldsymbol {y} = \operatorname {H L} (\boldsymbol {x}) = \left[ \begin{array}{c} \sqrt {\| \phi (\boldsymbol {W} \boldsymbol {x} , \boldsymbol {p}) \| ^ {2} + 1} \\ \phi (\boldsymbol {W} \boldsymbol {x}, \boldsymbol {p}) \end{array} \right], \tag {13} +$$ + +with operation function: + +$$ +\phi (\boldsymbol {W} \boldsymbol {x}, \boldsymbol {p}) = \frac {\lambda \left(\boldsymbol {p} ^ {\top} \boldsymbol {x} + b ^ {\prime}\right)}{\| \boldsymbol {W} h (\boldsymbol {x}) + \boldsymbol {b} \|} (\boldsymbol {W} h (\boldsymbol {x}) + \boldsymbol {b}), \tag {14} +$$ + +where $\pmb{b}$ and $b'$ are bias terms, $\lambda > 0$ regulates the scaling range. $h$ denotes the activation function. + +Based on the Lorentz Linear Layer, we propose a Lorentz self-attention module that integrates Gaussian constraints into feature interactions, enabling multiscale and hierarchical video embeddings in hyperbolic space. Specifically, given a hyperbolic video embedding $\pmb{x}_{\mathrm{in}}^{\mathcal{L}} \in \mathbb{L}^{n}$ , $\mathbb{R}^{M \times (n + 1)}$ , we first obtain the attention query $\mathcal{Q}$ , key $\mathcal{K}$ , and value $\mathcal{V}$ using Eq. (13), all in the shape of $\mathbb{R}^{M \times (n + 1)}$ . We calculate attention scores based on Eq. (6) and apply a Gaussian matrix $\mathcal{M}_{\sigma}^{g} \in \mathbb{R}^{M \times M}$ for element-wise multiplication with the score matrix to obtain a multi-scale receptive field. The output is defined as $\pmb{x}_{\mathrm{out}}^{\mathcal{L}} = \{\pmb{\mu}_1, \dots, \pmb{\mu}_{|\mathcal{Q}|}\} \in \mathbb{R}^{M \times (n + 1)}$ : + +$$ +S _ {i j} = \frac {\exp \left(\frac {- d _ {\mathcal {L}} ^ {2} \left(\boldsymbol {q} _ {i} , \boldsymbol {k} _ {j}\right) \odot \mathcal {M} _ {\sigma} ^ {g} (i , j)}{\sqrt {(n + 1)}}\right)}{\sum_ {k = 1} ^ {| \mathcal {K} |} \exp \left(\frac {- d _ {\mathcal {L}} ^ {2} \left(\boldsymbol {q} _ {i} , \boldsymbol {k} _ {k}\right) \odot \mathcal {M} _ {\sigma} ^ {g} (i , k)}{\sqrt {(n + 1)}}\right)}, \tag {15} +$$ + +$$ +\boldsymbol {\mu} _ {i} = \frac {\sum_ {j = 1} ^ {| \mathcal {K} |} S _ {i j} \boldsymbol {v} _ {j}}{\left| \left| \sum_ {k = 1} ^ {| \mathcal {K} |} S _ {i k} \boldsymbol {v} _ {k} \right| \right| _ {\mathcal {L}}} , +$$ + +the squared Lorentzian distance $d_{\mathcal{L}}^2 (\pmb {a},\pmb {b}) = -2 - 2\langle \pmb {a},\pmb {b}\rangle_{\mathcal{L}}$ + +After computing $x_{\mathrm{out}}^{\mathcal{L}}$ , we apply the logarithmic map Eq. (5), while discarding the time axis, to obtain the Euclidean space embedding $x_{\mathrm{mid}}^{E}$ . Then, the output $x_{\mathrm{out}}^{E}$ is obtained through a Linear Layer followed by rescaling: + +$$ +\begin{array}{l} \boldsymbol {x} _ {\text {m i d}} ^ {E} = \operatorname {d r o p \_ t i m e \_ a x i s} \left(\log_ {\boldsymbol {o}} \left(\boldsymbol {x} _ {\text {o u t}} ^ {\boldsymbol {\mathcal {L}}}\right)\right) \in \mathbb {R} ^ {M \times n}, \\ \boldsymbol {x} _ {\text {o u t}} ^ {E} = \frac {\boldsymbol {x} _ {\text {m i d}} ^ {E} W _ {2}}{\beta} \in \mathbb {R} ^ {M \times d}, \tag {16} \\ \end{array} +$$ + +where $W_{2}\in \mathbb{R}^{n\times d}$ , $\beta$ is the scale factor in Eq. (11). Finally, We replace the self-attention in Transformer block with Lorentz attention to form the Lorentz Attention Block. + +Mean-Guided Adaptive Interaction Module We arrange $N_{\mathcal{L}}$ Lorentz and $N_{E}$ Euclidean Attention Blocks in parallel to construct $N_{O}$ Gaussian Attention Blocks for multi-scale hybrid-space video embeddings. To integrate these features, we introduce a Mean-Guided Adaptive Interaction Module, which utilizes globally pooled features to compute dynamic aggregation weights. Specifically, we first obtain the global query $\varphi \in \mathbb{R}^{1\times d}$ and compute aggregation weights via a Cross Attention Block consisting of a cross-attention layer (CA) followed by a fully connected layer (FC): + +$$ +\boldsymbol {\varphi} = \operatorname {M e a n} \left(\boldsymbol {x} _ {\sigma_ {1}}, \boldsymbol {x} _ {\sigma_ {2}},.., \boldsymbol {x} _ {\sigma_ {N _ {o}}}\right), +$$ + +$$ +w _ {i} = \operatorname {F C} \left(\operatorname {C A} \left(\varphi , \boldsymbol {x} _ {\sigma_ {i}}, \boldsymbol {x} _ {\sigma_ {i}}\right)\right), i = 1, 2, \dots , N _ {o}, +$$ + +$$ +\tilde {w} _ {i, j} = \frac {e ^ {w _ {i , j} / \tau}}{\sum_ {k = 1} ^ {N _ {o}} e ^ {w _ {k , j} / \tau}}, j = 1, \dots , M, \tag {17} +$$ + +$$ +\tilde {\boldsymbol {x}} _ {j} = \sum_ {i = 1} ^ {N _ {o}} \tilde {w} _ {i, j} \boldsymbol {x} _ {\sigma_ {i, j}}, j = 1, \dots , M, +$$ + +$$ +\boldsymbol {x} _ {\text {M A I M}} = \operatorname {C o n c a t} \left(\tilde {\boldsymbol {x}} _ {1}, \tilde {\boldsymbol {x}} _ {2}, \dots , \tilde {\boldsymbol {x}} _ {M}\right), +$$ + +where $\pmb{x}_{\sigma_i} \in \mathbb{R}^{M \times d}$ denotes the output of the $i$ -th Gaussian block and $M$ corresponds to the number of time points (i.e., clips or frames). $w_i \in \mathbb{R}^M$ represents the aggregation weights for the $i$ -th Gaussian block, and $\tau$ is the temperature factor. $\tilde{\pmb{x}}_j \in \mathbb{R}^d$ denotes the aggregated feature at time point $j$ , while $\pmb{x}_{\mathrm{MAIM}}$ is the final output. + +# 3.4. Learning Objectives + +Given the partial relevance in PRVR, where each video fully entails its corresponding text, a partial order relationship is established, with the text-query semantically subsumed by the video: text $\prec$ video. Inspired by MERU [10], we propose the Partial Order Preservation Loss to enforce this relationship in Hyperbolic Space. Given $V_{f}$ and $V_{c}$ from Sec. 3.2, a simple attention module similar to Eq. (7) is applied, followed by mean pooling to get the unified video representation $V_{v}$ . The video and text representations are then mapped to Lorentz space via the exponential map, yielding $\pmb{v},\pmb{t}\in \mathbb{L}^n$ , as shown in Fig. 2(c). We define an entailment + +cone for each $\pmb{v}$ , which is characterized by the half-aperture: + +$$ +\mathbf {H A} (\boldsymbol {v}) = \arcsin \left(\frac {2 c}{\| \boldsymbol {v} _ {\boldsymbol {s}} \|}\right). \tag {18} +$$ + +$c = 0.1$ is used to define the boundary conditions near the origin. We measure the exterior angle $\mathbf{EA}(\boldsymbol{v},\boldsymbol{t}) = \pi - \angle Ovt$ to penalize cases where $\boldsymbol{t}$ falls outside the entailment cone: + +$$ +\mathbf {E A} (\boldsymbol {v}, \boldsymbol {t}) = \operatorname {a r c c o s} \left(\frac {t _ {0} + v _ {0} \langle \boldsymbol {v} , \boldsymbol {t} \rangle_ {\mathcal {L}}}{\| \boldsymbol {v} _ {\boldsymbol {s}} \| \sqrt {\left(\langle \boldsymbol {v} , \boldsymbol {t} \rangle_ {\mathcal {L}}\right) ^ {2} - 1}}\right). \tag {19} +$$ + +The Loss for a single video-text pair is given by: + +$$ +L _ {p o p} (\boldsymbol {v}, \boldsymbol {t}) = \max (0, \mathbf {E A} (\boldsymbol {v}, \boldsymbol {t}) - \mathbf {H A} (\boldsymbol {v})). \tag {20} +$$ + +Besides, following MS-SL [14], we use the standard similarity retrieval loss to train the model, denoted as $L_{sim}$ . Meanwhile, the query diversity [57] $L_{div}$ is used to improve retrieval performance. The aggregate loss is defined as: + +$$ +L _ {a g g} = L _ {s i m} + \lambda_ {1} L _ {d i v} + \lambda_ {2} L _ {p o p}, \tag {21} +$$ + +$\lambda_{1}$ and $\lambda_{2}$ are hyper-parameters that balance learning losses. + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets We conduct experiments on three benchmark datasets: (i) ActivityNet Captions [28], which comprises approximately 20K YouTube videos with an average duration of 118 seconds. Each video contains an average of 3.7 annotated moments with corresponding textual descriptions. (ii) TV show Retrieval (TVR) [30], consisting of 21.8K videos sourced from six TV shows. Each video is associated with five natural language descriptions covering different moments. (iii) Charades-STA [22], which includes 6,670 videos annotated with 16,128 sentence descriptions. On average, each video contains approximately 2.4 moments with corresponding textual queries. We adopt the same data split as used in prior studies[14, 57]. It is important to note that the moment annotations are unavailable in the PRVR task. + +Metrics Following previous works [14, 57], we adopt rank-based evaluation metrics, specifically $R@K$ ( $K = 1, 5, 10, 100$ ). The metric $R@K$ represents the proportion of queries for which the correct item appears within the top $K$ positions of the ranking list. All results are reported as percentages (\%), where higher values indicate superior retrieval performance. To facilitate an overall comparison, we also report the Sum of all Recalls (SumR). + +# 4.2. Implementation Details + +Data Processing For video representations on TVR, we employ the feature set provided by Lei et al. [30], which + +comprises 3,072-dimensional visual features obtained by concatenating frame-level ResNet152 features [23] and segment-level I3D features [2]. For ActivityNet Captions and Charades-STA, we only utilize I3D features as provided by Zhang et al. [62] and Mun et al. [43], respectively. For sentence representations, we adopt the 768-dimensional RoBERTa features supplied by Lei et al. [30] for TVR. On ActivityNet Captions and Charades-STA, we employ 1,024-dimensional RoBERTa features extracted using MS-SL[14]. + +Model Configurations The HLFormaler block consists of 8 Gaussian blocks $(N_O = 8)$ , 4 Lorentz Attention blocks $(N_{\mathcal{L}} = 4)$ , with Gaussian variances ranging from $2^{1}$ to $2^{N_{\mathcal{L}} - 1}$ and $\infty$ , and 4 Euclidean Attention blocks $(N_E = 4)$ , with Gaussian variances ranging from $2^{1}$ to $2^{N_{E} - 1}$ and $\infty$ . The latent dimension $d = 384$ with 4 attention heads. + +Training Configurations We employ the Adam optimizer with a mini-batch size of 128 and set the number of epochs to 100. The model is implemented using PyTorch and trained on one Nvidia RTX 3080 Ti GPU. We adopt a learning rate adjustment schedule similar to MS-SL. + +# 4.3. Comparison with State-of-the arts + +Baselines We select six representative PRVR baselines for comparison: MS-SL [14], PEAN [26], LH [20], BGM-Net [60], GMMFormer [57], and DL-DKD [16]. We also compare HLFomer with methods for T2VR and VCMR. For T2VR, we select six T2VR models: CE [37], HGR [5], DE++ [13], RIVRL [15], CLIP4Clip [40], Cap4Video [59], For VCMR, we consider four models: XML [30], ReLoCLNet [63], CONQUER [25] and JSG[7]. + +Retrieval Performance Tab. 1 presents the retrieval performance of various models on three large-scale video datasets. As observed, T2VR models, designed to capture overall video-text relevance, underperform for PRVR. VCMR models, which focus on moment retrieval, achieve better results. PRVR methods perform best as they are specifically designed for this task. Attributed to hyperbolic space learning and effective utilization of video hierarchical structure priors, HLFformer consistently surpasses all baselines. It outperforms DL-DKD by $4.9\%$ and $4.3\%$ in SumR on ActivityNet Captions and TVR, respectively, and exceeds PEAN by $5.4\%$ on Charades-STA. + +# 4.4. Model Analyses + +Efficacy of Temporal Modeling Design We perform ablation studies to examine the effect of the attention block number $N_{o}$ and the attention mechanism ratio $N_{L} / N_{E}$ , with results shown in Fig. 3. Model performance improves as $N_{o}$ increases, then stabilizes or declines when $N_{o} \geq 8$ . Even with only two attention blocks, HLFomer surpasses most competing methods. Furthermore, using solely Euclidean or Lorentz attention blocks results in suboptimal performance, whereas the hybrid attention block achieves the best results. + +
ModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
T2VR
HGR [5]4.015.024.863.2107.01.23.87.333.445.71.74.98.335.250.1
RIVRL [15]5.218.028.266.4117.81.65.69.437.754.39.423.432.270.6135.6
DE++ [13]5.318.429.268.0121.01.75.69.637.154.18.821.930.267.4128.3
CE [37]5.519.129.971.1125.61.34.57.336.049.13.712.820.164.5101.1
CLIP4Clip [40]5.919.330.471.6127.31.86.510.944.263.49.924.334.372.5141.0
Cap4Video [59]6.320.430.972.6130.21.96.711.345.065.010.326.436.874.0147.5
VCMR
ReLoCLNet [63]5.718.930.072.0126.61.25.410.045.662.310.026.537.381.3155.1
XML [30]5.319.430.673.1128.41.66.010.146.964.610.728.138.180.3157.1
CONQUER [25]6.520.431.874.3133.11.86.310.347.566.011.028.939.681.3160.8
JSG [7]6.822.734.876.1140.52.47.712.849.872.7-----
PRVR
MS-SL [14]7.122.534.775.8140.11.87.111.847.768.413.532.143.483.4172.4
PEAN [26]7.423.035.575.9141.82.78.113.550.374.713.532.844.183.9174.2
LH [20]7.423.535.875.8142.42.17.512.950.172.713.233.244.485.5176.3
BGM-Net [60]7.223.836.076.9143.91.97.412.250.171.614.134.745.985.2179.9
GMMFormer [57]8.324.936.776.1146.02.17.812.550.672.913.933.344.584.9176.6
DL-DKD [16]8.025.037.577.1147.6-----14.434.945.884.9179.9
HLFormer (ours)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
+ +Table 1. Retrieval performance of HLFomer and other faithful methods on ActivityNet Captions, Charades-STA and TVR. State-of-the-art performance is highlighted in bold. “-” indicates that the corresponding results are unavailable. + +
IDModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
(0)HLFormer (full)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
Efficacy of Multi-scale Branches
(1)w/o gaze branch7.624.436.777.3146.11.88.013.950.874.513.934.045.285.3178.3
(2)w/o glance branch6.421.733.675.4137.21.67.713.148.470.811.430.541.882.4166.1
Efficacy of Different Loss Terms
(3)Lsim Only7.725.038.178.3149.12.08.113.252.075.315.136.247.886.0185.2
(4)w/o Ldiv8.526.639.678.8153.52.07.813.653.076.415.736.448.486.0186.5
(5)w/o Lpop8.626.939.778.8154.02.28.414.053.077.615.636.848.486.0186.8
Efficacy of various Aggregation Strategies
(6)w/ MP8.525.738.277.8150.22.08.013.252.175.315.236.547.486.0185.1
(7)w/ CL8.726.839.578.6153.62.08.213.952.076.115.336.948.486.0186.6
+ +Table 2. Ablation Study of HLFomer. The best scores are marked in bold. + +This may be attributed to the differences in representational focus: Euclidean space emphasizes fine-grained local feature learning and sometimes overlooks global hierarchical structures, while hyperbolic space prioritizes global hierarchical relationships at the expense of local details. Moreover, hyperbolic space tends to be more sensitive to noise and numerically unstable. By integrating hybrid spaces, HLFormer achieves mutual compensation, enhancing representation learning and facilitating video semantic understanding. + +Efficacy of Hyperbolic Learning Hyperbolic learning demonstrates significant advantages in capturing the hierarchical structure of videos. As illustrated in Fig. 4(a), embeddings learned solely in Euclidean space exhibit indistinct cluster boundaries, with red and green points at the periphery closely interspersed. In contrast, Fig. 4(b) demonstrates that incorporating Lorentz attention facilitates the learning of more discriminative representations, while refining moment + +cluster boundaries, increasing inter-moment separation, and compacting intra-moment frame distributions, revealing a more pronounced hierarchical structure. + +Efficacy of Multi-scale Branches To evaluate the effectiveness of the multi-scale branches, we conduct comparative experiments by removing either the glance clip-level branch or the gaze frame-level branch. As shown in Tab. 2, the absence of any branch leads to a noticeable performance degradation. These results not only validate the efficacy of the coarse-to-fine multi-granularity retrieval mechanism but also highlight the complementary nature of the two branches. Efficacies of Different Loss Terms To analyze the effectiveness of three loss terms (i.e. $L_{sim}$ , $L_{div}$ and $L_{pop}$ ) of HLFformer, we construct several HLFformer variants: (i) $L_{sim}$ Only: train the model with merely $L_{sim}$ . (ii) w/o $L_{div}$ : We train the model without query diverse learning. (iii) w/o $L_{pop}$ : HLFformer removes the partial order preservation task. + +![](images/2bff62a8f9e2f8e23a6877da164dd58230f6a41f12d16ad61bba59a01564b888.jpg) + +![](images/7e70021359ec2b0f54facb86ba45758fe6918dfcf734fcaba37433c427b018d9.jpg) + +![](images/44a9f3a7342add11f5b8893b13b4b9628917f888220e5ce513bf486665c6196d.jpg) +(a) ActivityNet Captions + +![](images/48c6f51ec532029942e43de228899d1de8cff8c74e4231bc338201247b35abf3.jpg) + +![](images/f3a4a1e670b44835a18105849a5dcc0124e2fdea302d1186cc924b7cb8cabe0e.jpg) +(b) Charades-STA + +![](images/4be7380b2f802dd6b3975888c5fd189908fe19af24cad94e6ddf6231f09079fd.jpg) +(c) TVR + +![](images/49af380d9bb3f0b523b6cc8d740ff1b2ad2e194bce366a8d9212331623dba049.jpg) +Figure 3. The influence of different attention blocks, with default settings marked in bold. +(a) $w / o$ Lorentz Attention + +![](images/072ee8f433ac722281ac7aa476c465c1f75f70b5cebdf19a1f5e5f2a028acd94.jpg) +(b) $w/$ Lorentz Attention +Figure 4. The UMAP [41] visualization displays the learned frame embeddings from a video in TVR. Data points of the same color correspond to the same moment. + +As shown in Tab. 2, the worst performance occurs when only $L_{sim}$ is used. Comparing Variant (5) with Variant (3), adding $L_{div}$ increases the SumR, which can validate its necessity. Similarly, comparing Variant (4) with Variant (3) and Fig. 5, integrating $L_{pop}$ not only boosts retrieval accuracy but also ensures that the text query remains semantically embedded within the corresponding video, preserving partial relevance. Efficacy of Aggregation Strategy We compare three aggregation strategies: (i) $w/$ MP: mean pooling for static fusion. (ii) $w/$ CL: feature concatenation with linear layers. (iii) MAIM (default): mean-guided adaptive interaction module. As shown in Tab. 2, MP performs the worst due to its fixed static fusion, which limits semantic interaction. CL improves upon MP by leveraging linear layers for dynamic feature fusion. MAIM achieves the best performance by learning adaptive aggregation weights and dynamically selecting hyperbolic information under global guidance. + +![](images/3e9ad8a8d7b285e1fe59545c9fa44c58a47e20dca76550d2c430c2e6e90a0c4d.jpg) +(a) $w / oL_{pop}$ + +![](images/a9d0c200c64728be273761a6d2756e584e898e07e438b6275623ba0445082acf.jpg) + +![](images/6f23f2ef61f0a26d8c71cda61fecc309ebd0e86103a620bc4b41b3ce04330d01.jpg) +(b) $w / L_{pop}$ + +![](images/4f64357e5b574a934ff6fd896f9743c1fd144b39d82689d5cb707c6cc42abc31.jpg) +Figure 5. Visualization of the learned hyperbolic space. The closer to the origin, the higher semantic hierarchy and coarser granularity. + +Visualization of Hyperbolic Space Inspired by HyCo-CLIP [46], we visualize the learned hyperbolic space by sampling 3K embeddings from the TVR training set. We analyze their norm distribution via histogram and reduce dimensionality using HoroPCA [3], as shown in Fig. 5. Glance branch embeddings are positioned closer to the origin than text query embeddings, indicating that clip-level video representations subsume textual queries. This phenomenon can be attributed to $L_{pop}$ , which enforces the partial order relationship between video and text representations. In contrast, without $L_{pop}$ , embeddings exhibit uncorrelated distributions. Moreover, text queries, being coarser in semantics, lie closer to the origin than fine-grained gaze-level embeddings, reflecting a clear hierarchical structure. + +# 5. Conclusions + +In this paper, we propose HLFormaler, a novel hyperbolic modeling framework tailored for PRVR. By leveraging the intrinsic geometric properties of hyperbolic space, HLFormaler effectively captures the hierarchical and multi-granular structure of untrimmed videos, thereby enhancing video-text retrieval accuracy. Furthermore, to ensure partial relevance between paired videos and text, a partial order preservation loss is introduced to enforce their semantic entailment. Extensive experiments indicate that HLFormaler consistently outperforms state-of-the-art methods. Our study offers a new perspective for PRVR with hyperbolic learning, which we hope will inspire further research in this direction. + +Acknowledgments. We sincerely thank the anonymous reviewers and chairs for their efforts and constructive suggestions, which have greatly helped us improve the manuscript. This work is supported in part by the National Natural Science Foundation of China under grant 624B2088, 62171248, 62301189, the PCNL KEY project (PCL2023AS6-1), and Shenzhen Science and Technology Program under Grant KJZD20240903103702004, JCYJ20220818101012025, GXWD20220811172936001. Long Chen was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), Huawei Gift Fund, and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04). + +# References + +[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 3 +[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6 +[3] Ines Chami, Albert Gu, Dat Nguyen, and Christopher Ré. Horopca: Hyperbolic dimensionality reduction via horospherical projections, 2021. 8 +[4] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275-1290, 2023. 3 +[5] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 1, 2, 6, 7 +[6] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. arXiv preprint arXiv:2105.14686, 2021. 3, 5 +[7] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multi-granularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975–983, 2023. 2, 6, 7 +[8] Dingxin Cheng, Shuhan Kong, Bin Jiang, and Qiang Guo. Transferable dual multi-granularity semantic excavating for partially relevant video retrieval. Image and Vision Computing, 149:105168, 2024. 2 +[9] Cheol-Ho Cho, WonJun Moon, Woojin Jun, MinSeok Jung, and Jae-Pil Heo. Ambiguity-restrained text-video representation learning for partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2500-2508, 2025. 2 +[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic Image- + +Text Representations. In Proceedings of the International Conference on Machine Learning, 2023. 2, 3, 5 +[11] Jianfeng Dong, Xirong Li, and Cees GM Snoek. Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia, 20(12):3377-3388, 2018. 1, 2 +[12] Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zeroexample video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2019. +[13] Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065-4080, 2021. 1, 6, 7 +[14] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 246-257, 2022. 1, 2, 4, 6, 7 +[15] Jianfeng Dong, Yabing Wang, Xianke Chen, Xiaoye Qu, Xirong Li, Yuan He, and Xun Wang. Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5680-5694, 2022. 1, 2, 6, 7 +[16] Jianfeng Dong, Minsong Zhang, Zheng Zhang, Xianke Chen, Daizong Liu, Xiaoye Qu, Xun Wang, and Baolong Liu. Dual learning with dynamic knowledge distillation for partially relevant video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11302-11312, 2023. 2, 6, 7 +[17] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399-7409, 2022. 2, 3 +[18] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 1, 2 +[19] Hao Fang, Changle Zhou, Jiawei Kong, Kuofeng Gao, Bin Chen, Tao Liang, Guojun Ma, and Shu-Tao Xia. Grounding language with vision: A conditional mutual information calibrated decoding strategy for reducing hallucinations in lvlms. arXiv preprint arXiv:2505.19678, 2025. 2 +[20] Sheng Fang, Tiantian Dang, Shuhui Wang, and Qingming Huang. Linguistic hallucination for text-based video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 34(10):9692-9705, 2024. 6, 7 +[21] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3 +[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 6 +[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- + +ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[24] Neil He, Menglin Yang, and Rex Ying. Lorentzian residual neural networks. arXiv preprint arXiv:2412.14695, 2024. 3 +[25] Zhijian Hou, Chong-Wah Ngo, and Wing Kwong Chan. Conquer: Contextual query-aware ranking for video corpus moment retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3900-3908, 2021. 6, 7 +[26] Xun Jiang, Zhiguo Chen, Xing Xu, Fumin Shen, Zuo Cao, and Xunliang Cai. Progressive event alignment network for partial relevant video retrieval. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1973-1978. IEEE, 2023. 2, 6, 7 +[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3 +[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 2, 6 +[29] Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. In Proceedings of the 36th International Conference on Machine Learning, pages 3672-3681. PMLR, 2019. 2 +[30] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 447-463. Springer, 2020. 2, 6, 7 +[31] Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, and Xinbo Gao. Beyond euclidean: Dualspace representation learning for weakly supervised video violence detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3 +[32] Keegan Lensink, Bas Peters, and Eldad Haber. Fully hyperbolic convolutional neural networks. Research in the Mathematical Sciences, 9(4):60, 2022. 3, 5 +[33] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17511-17520, 2024. 3 +[34] Xirong Li, Chaoxi Xu, Gang Yang, Zhineng Chen, and Jianfeng Dong. W2vv++ fully deep learning for ad-hoc video search. In Proceedings of the 27th ACM international conference on multimedia, pages 1786-1794, 2019. 1, 2 +[35] Haitong Liu, Kuofeng Gao, Yang Bai, Jinmin Li, Jinxiao Shan, Tao Dai, and Shu-Tao Xia. Protecting your video content: Disrupting automated video-based lvm annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24056-24065, 2025. 2 +[36] Peidong Liu, Dongliang Liao, Jinpeng Wang, Yangxin Wu, Gongfu Li, Shu-Tao Xia, and Jin Xu. Multi-task ranking with user behaviors for text-video search. In Companion Proceedings of the Web Conference 2022, pages 126-130, 2022. 2 +[37] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using + +representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2, 6, 7 +[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3 +[39] Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1141-1150, 2020. 3 +[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 6, 7 +[41] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 8 +[42] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2 +[43] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 6 +[44] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3 +[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning, pages 3779-3788. PMLR, 2018. 3 +[46] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In The Thirteenth International Conference on Learning Representations, 2025. 3, 8 +[47] Xiaogang Peng, Hao Wen, Yikai Luo, Xiao Zhou, Keyang Yu, Yigang Wang, and Zizhao Wu. Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023. 3 +[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3 +[49] Ruiqi Shi, Jun Wen, Wei Ji, Menglin Yang, Difei Gao, and Roger Zimmermann. HOVER: Hyperbolic video-text retrieval, 2024. 3 +[50] Xue Song, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Spatial-temporal graphs for cross-modal text2video retrieval. IEEE Transactions on Multimedia, 24:2914-2923, 2021. 2 +[51] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for + +weakly-supervised video paragraph grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569-13580, 2024. 2 +[52] Max Van Spengler, Erwin Berkhout, and Pascal Mettes. Poincaré resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 3 +[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3 +[54] Jinpeng Wang, Bin Chen, Dongliang Liao, Ziyun Zeng, Gongfu Li, Shu-Tao Xia, and Jin Xu. Hybrid contrastive quantization for efficient cross-view video retrieval. In Proceedings of the ACM Web Conference 2022, pages 3020-3030, 2022. 2 +[55] Jinpeng Wang, Ziyun Zeng, Bin Chen, Yuting Wang, Dongliang Liao, Gongfu Li, Yiru Wang, and Shu-Tao Xia. Hugs bring double benefits: Unsupervised cross-modal hashing with multi-granularity aligned transformers. International Journal of Computer Vision, 132(8):2765-2797, 2024. 2 +[56] Yuting Wang, Jinpeng Wang, Bin Chen, Tao Dai, Ruisheng Luo, and Shu-Tao Xia. Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024. 1, 2 +[57] Yuting Wang, Jinpeng Wang, Bin Chen, Ziyun Zeng, and Shu-Tao Xia. Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 4, 6, 7 +[58] Zhihao Wang, Wenke Huang, Tian Chen, Zekun Shi, Guancheng Wan, Yu Qiao, Bin Yang, Jian Wang, Bing Li, and Mang Ye. An empirical study of federated prompt learning for vision language model. arXiv preprint arXiv:2505.23024, 2025. 2 +[59] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704-10713, 2023. 6, 7 +[60] Shukang Yin, Sirui Zhao, Hao Wang, Tong Xu, and Enhong Chen. Exploiting instance-level relationships in weakly supervised text-to-video retrieval. ACM Trans. Multim. Comput. Commun. Appl., 20(10):316:1-316:21, 2024. 2, 6, 7 +[61] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3 +[62] Bowen Zhang, Hexiang Hu, Joonseok Lee, Ming Zhao, Sheide Chammas, Vihan Jain, Eugene Ie, and Fei Sha. A hierarchical multi-modal encoder for moment localization in video corpus. arXiv preprint arXiv:2011.09046, 2020. 6 +[63] Hao Zhang, Aixin Sun, Wei Jing, Guoshun Nan, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Video corpus moment retrieval with contrastive learning. In Proceedings of the 44th International ACM SIGIR Conference on Research + +and Development in Information Retrieval, pages 685-695, 2021. 6, 7 \ No newline at end of file diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/images.zip b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c1cc5db019d57d1ed0087404895f0f821bcd103f --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0fa311831d6cd05a2b2ae5a67e9265d9377a89cab76aaf84e34a69ba3c2d1fe1 +size 658964 diff --git a/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/layout.json b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..904fdb761a9b82aa609de3bd1273f284d31277ac --- /dev/null +++ b/2025/Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning/layout.json @@ -0,0 +1,12014 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 529, + 121 + ], + "type": "text", + "content": "Enhancing Partially Relevant Video Retrieval with Hyperbolic Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "spans": [ + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": "Jun Li" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Jinpeng Wang" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2*†}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Chaolei Tan" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Niu Lian" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Long Chen" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Yaowei Wang" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1,3}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Min Zhang" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Shu-Tao Xia" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "text", + "content": ", Bin Chen" + }, + { + "bbox": [ + 141, + 142, + 468, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 227, + 173, + 383, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 227, + 173, + 383, + 185 + ], + "spans": [ + { + "bbox": [ + 227, + 173, + 383, + 185 + ], + "type": "text", + "content": "1Harbin Institute of Technology, Shenzhen" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 172, + 186, + 438, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 186, + 438, + 199 + ], + "spans": [ + { + "bbox": [ + 172, + 186, + 438, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 172, + 186, + 438, + 199 + ], + "type": "text", + "content": "Tsinghua Shenzhen International Graduate School, Tsinghua University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 182, + 201, + 428, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 201, + 428, + 213 + ], + "spans": [ + { + "bbox": [ + 182, + 201, + 428, + 213 + ], + "type": "text", + "content": "3Research Center of Artificial Intelligence, Peng Cheng Laboratory" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 203, + 215, + 407, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 203, + 215, + 407, + 228 + ], + "spans": [ + { + "bbox": [ + 203, + 215, + 407, + 228 + ], + "type": "text", + "content": "4The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 155, + 229, + 452, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 155, + 229, + 452, + 241 + ], + "spans": [ + { + "bbox": [ + 155, + 229, + 452, + 241 + ], + "type": "text", + "content": "220110924@stu.hit.edu.cn " + }, + { + "bbox": [ + 155, + 229, + 452, + 241 + ], + "type": "inline_equation", + "content": "\\text{四}" + }, + { + "bbox": [ + 155, + 229, + 452, + 241 + ], + "type": "text", + "content": " wjp20@mails.tsinghua.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 152, + 269, + 200, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 269, + 200, + 281 + ], + "spans": [ + { + "bbox": [ + 152, + 269, + 200, + 281 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 53, + 293, + 297, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 293, + 297, + 558 + ], + "spans": [ + { + "bbox": [ + 53, + 293, + 297, + 558 + ], + "type": "text", + "content": "Partially Relevant Video Retrieval (PRVR) addresses the critical challenge of matching untrimmed videos with text queries describing only partial content. Existing methods suffer from geometric distortion in Euclidean space that sometimes misrepresents the intrinsic hierarchical structure of videos and overlooks certain hierarchical semantics, ultimately leading to suboptimal temporal modeling. To address this issue, we propose the first hyperbolic modeling framework for PRVR, namely HLFformer, which leverages hyperbolic space learning to compensate for the suboptimal hierarchical modeling capabilities of Euclidean space. Specifically, HLFformer integrates the Lorentz Attention Block and Euclidean Attention Block to encode video embeddings in hybrid spaces, using the Mean-Guided Adaptive Interaction Module to dynamically fuse features. Additionally, we introduce a Partial Order Preservation Loss to enforce \"text " + }, + { + "bbox": [ + 53, + 293, + 297, + 558 + ], + "type": "inline_equation", + "content": "\\prec" + }, + { + "bbox": [ + 53, + 293, + 297, + 558 + ], + "type": "text", + "content": " video\" hierarchy through Lorentzian cone constraints. This approach further enhances cross-modal matching by reinforcing partial relevance between video content and text queries. Extensive experiments show that HLFformer outperforms state-of-the-art methods. Code is released at https://github.com/lijun2005/ICCV25-HLFformer." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 135, + 593 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 135, + 593 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 135, + 593 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 600, + 296, + 685 + ], + "type": "text", + "content": "Text-to-video retrieval (T2VR) [5, 11-13, 15, 18, 34, 37, 42] is a fundamental module in many search applications and a popular topic in multi-modal learning. While most T2VR models are developed for short clips or pre-trimmed video segments, they may face challenges where user queries describe only partial content in the video. This practical issue in real-world usage promotes a more challenging setting of" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 322, + 267, + 441, + 409 + ], + "blocks": [ + { + "bbox": [ + 322, + 267, + 441, + 409 + ], + "lines": [ + { + "bbox": [ + 322, + 267, + 441, + 409 + ], + "spans": [ + { + "bbox": [ + 322, + 267, + 441, + 409 + ], + "type": "image", + "image_path": "2f6a4a5389cca2054b178cee763893f730f3b0be262e16d79f80f1f782230c13.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 409, + 444, + 417 + ], + "lines": [ + { + "bbox": [ + 323, + 409, + 444, + 417 + ], + "spans": [ + { + "bbox": [ + 323, + 409, + 444, + 417 + ], + "type": "text", + "content": "(a) Semantic Hierarchy in Untrimmed Videos" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 426, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 492 + ], + "type": "text", + "content": "Figure 1. (a) Modeling the semantic hierarchy in untrimmed videos helps Partially Relevant Video Retrieval (PRVR). (b) Euclidean space is less effective in modeling semantic hierarchy due to the flat geometry. Data points with distant hierarchical relation may be close. (c) Hyperbolic space allows larger cardinals when approaching the edge, which is preferable to preserve the hierarchy." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 442, + 268, + 547, + 409 + ], + "blocks": [ + { + "bbox": [ + 442, + 268, + 547, + 409 + ], + "lines": [ + { + "bbox": [ + 442, + 268, + 547, + 409 + ], + "spans": [ + { + "bbox": [ + 442, + 268, + 547, + 409 + ], + "type": "image", + "image_path": "e73c642f342019f007b74c6fbe378b48667eaee066e9d0b296b2ed2216bba876.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 409, + 522, + 417 + ], + "lines": [ + { + "bbox": [ + 463, + 409, + 522, + 417 + ], + "spans": [ + { + "bbox": [ + 463, + 409, + 522, + 417 + ], + "type": "text", + "content": "(c) Hyperbolic Space" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 312, + 519, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 519, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 312, + 519, + 555, + 543 + ], + "type": "text", + "content": "partially relevant video retrieval (PRVR) [14], which aims to match each text query with the best untrimmed video." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 312, + 546, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 546, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 312, + 546, + 556, + 714 + ], + "type": "text", + "content": "Due to unlabeled moment timestamps, PRVR requires solid abilities on (i) identifying key moments in videos for extracting informative features and (ii) learning robust cross-modal representations to match text queries and videos precisely. Prior arts have developed preliminary solutions on both aspects, while challenges remain. For (i), MS-SL [14] exhaustively enumerated consecutive frame combinations through multi-scale sliding windows, which inevitably engaged redundancy, noise, and a high computational complexity in extracting moment features. GMMFormer [56, 57] improved efficiency by leveraging Gaussian neighborhood priors to traverse each timestamp and discover potential key moments. However, it may still be hard to distinguish adjacent or semantically similar candidate moments. Though" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23074" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "DL-DKD [16] neatly benefited from the pretrained CLIP [48] to enhance text-frame alignment, the temporal generalizability is bounded by the text-image teacher model. For (ii), most existing solutions inherited similar ideas from classic T2VR, e.g., ranking and contrastive learning, at a holistic level, but important characteristics of PRVR, e.g., partial relevance and semantic entailment, are still under-explored." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 159, + 298, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 159, + 298, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 159, + 298, + 399 + ], + "type": "text", + "content": "In this paper, we take a hierarchical perspective to review the task, in the belief that videos naturally exhibit semantic hierarchy. As illustrated in Fig. 1(a), an untrimmed video can be regarded as a progression from frames to informative segments (e.g., Dunk), extended moments, and ultimately, the whole. Leveraging this intrinsic property is expected to benefit long video understanding. In particular for PRVR, the hierarchical prior provides positive guidance to arrange the moment features. Meanwhile, the supervisory signals from query-video matching can activate moment extraction more precisely through implicit bottom-up modeling. Exploring hierarchical features is never trivial. Unfortunately, existing PRVR approaches relying on Euclidean space are less effective in modeling the desired patterns in the flat geometry. We present Fig. 1(b) to exemplify this: two embeddings with a distant hierarchical relation may be spatially close to each other, as marked by the red arrows. Biased representation will increase the difficulty in disentangling informative moments from background, which limits the robustness in cross-modal matching considering partial relevance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 403, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 403, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 403, + 298, + 714 + ], + "type": "text", + "content": "Inspired by the emerging success of hyperbolic learning [10, 17, 29, 31, 44], which takes advantage of exponentially expanding metric in non-Euclidean space to better capture hierarchical structure (Fig. 1(c)), we introduce HLFoRer, a sincere exploration of hyperbolic learning to enhance PRVR. On temporal modeling, we carefully design a dual-branch strategy to capture informative moment features comprehensively. Specifically, for the hyperbolic branch, we develop a Lorentz Attention Block (LAB) with the hyperbolic selfattention mechanism. With the implicit hierarchical prior through end-to-end matching optimization, LAB learns to activate informative moment features relevant to queries and distinguish them from noisy background in the hyperbolic space, compensating for the limitations of Euclidean attention in capturing hierarchical semantics. We integrate dual-branch moment features with a Mean-Guided Adaptive Interaction Module (MAIM), which is lightweight but effective. On cross-modal matching, drawing on the intrinsic \"text " + }, + { + "bbox": [ + 56, + 403, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\prec" + }, + { + "bbox": [ + 56, + 403, + 298, + 714 + ], + "type": "text", + "content": " video\" hierarchy in PRVR where textual queries are subordinate to their paired videos, we introduce a Partial Order Preservation (POP) loss that geometrically confines text embeddings within hyperbolic cone anchored by corresponding video representations in an auxiliary Lorentzian manifold. This hierarchical metric alignment ensures semantic consistency between localized text semantics and their parent video structure while preserving partial relevance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "Empirical evaluations on three benchmark datasets: ActivityNet Captions [28], Charades-STA [22], and TVR [30] establish HLFoerner's state-of-the-art performance. Ablation studies confirm the necessity of hyperbolic geometry for hierarchical representation and the critical role of explicitly relational constraints in Partial Order Preservation Loss. Meanwhile, visual evidences further reveal that hyperbolic learning can enhance discriminative representation while maintaining video-text entailment, sharpening moment distinction and improving query alignment." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 328, + 193, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 193, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 328, + 193, + 555, + 205 + ], + "type": "text", + "content": "The primary contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 205, + 555, + 326 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 313, + 205, + 555, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 555, + 253 + ], + "type": "text", + "content": "- We propose to enhance PRVR with hyperbolic learning, including a Lorentz attention block with hierarchical priors to enhance the moment feature extraction, which collaborates with Euclidean attention and hybrid-space fusion." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 254, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 254, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 254, + 555, + 289 + ], + "type": "text", + "content": "- We design a partial order preservation loss that geometrically enforces the \"text " + }, + { + "bbox": [ + 313, + 254, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\prec" + }, + { + "bbox": [ + 313, + 254, + 555, + 289 + ], + "type": "text", + "content": " video\" hierarchy through hyperbolic cone constraints, strengthening partial relevance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 326 + ], + "type": "text", + "content": "- Extensive experiments on three benchmarks validate HLFformer's superiority, with analyses confirming the efficacy of hyperbolic modeling and geometric constraints." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 406, + 350 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 359, + 497, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 359, + 497, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 497, + 371 + ], + "type": "text", + "content": "2.1. Partially Relevant Video Retrieval" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 377, + 556, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 545 + ], + "type": "text", + "content": "With the growth of video content [19, 35, 58], video retrieval has become a key research area. Given a text query, Text-to-Video Retrieval (T2VR) [5, 11, 15, 18, 34, 36, 37, 42, 54, 55] focuses on retrieving fully relevant videos from pre-trimmed short clips. Video Corpus Moment Retrieval (VCMR) [7, 30, 50, 51] aims to localize specific moments within videos from a large corpus. Partially Relevant Video Retrieval (PRVR) [8, 9, 14, 16, 26, 56, 57, 60], a more recent task introduced by Dong et al. [14], aims to retrieve partially relevant videos from large, untrimmed long video collections. Unlike T2VR, PRVR must address the challenge of partial relevance, where the query pertains to only a specific moment of the video. Though the first stage of VCMR is similar to PRVR, VCMR requires moment-level annotations, limiting scalability." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 557, + 714 + ], + "type": "text", + "content": "Existing methods enhance PRVR retrieval from various perspectives. MS-SL [14] defines the PRVR task as a Multi-instance Learning, providing a strong baseline with explicit redundant clip embeddings. GMMFormer [56, 57] and PEAN [26] propose implicit clip modeling to improve efficiency. DL-DKD [16] achieves great results through dynamic distillation of CLIP [48]. BGM-Net [60] exploits an instance-level matching scheme for pairing queries and videos. However, these methods predominantly rely on Euclidean space, which sometimes distort the hierarchical structures in untrimmed long videos. Consequently, they fail to fully exploit video hierarchy priors. To overcome this issue, we propose HLFormer to enhances PRVR by implicitly capturing hierarchical structures through hyperbolic learning." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "23075" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 177, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 177, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 177, + 85 + ], + "type": "text", + "content": "2.2. Hyperbolic Learning" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 297, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 297, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 297, + 341 + ], + "type": "text", + "content": "Hyperbolic learning has attracted significant attention for its effectiveness in modeling hierarchical structures in real-world datasets. Early studies in computer vision tasks explored hyperbolic image embeddings from image-label pairs [27, 44], while subsequent progress extended hyperbolic optimization to multi-modal learning. MERU [10] and HyCoCLIP [46] notably surpassed Euclidean counterparts like CLIP [48] via hyperbolic space adaptation. Applications span semantic segmentation [1, 4], recognition tasks (skin [61], action [39]), meta-learning [17], and detection frameworks (violence [31, 47], anomalies [33]). Recent advances in fully hyperbolic neural networks [6, 21, 24, 32, 52] further underscore their potential. Motivated by them, we present the first study to explore the potential of hyperbolic learning for PRVR. Unlike other methods such as DSRL [31] and HOVER [49], our approach utilizes hyperbolic space to compensate for the limitations of Euclidean space in capturing the hierarchical structure of untrimmed long videos. Furthermore, we introduce the Partial Order Preservation Loss to explicitly capture the partial relevance between video and text in hyperbolic space, improving retrieval performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 351, + 111, + 363 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 371, + 141, + 383 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 141, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 141, + 383 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": "Hyperbolic Space Hyperbolic spaces are Riemannian manifolds with a constant negative curvature " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": ", contrasting with the zero-curvature (flat) geometry of Euclidean spaces. Among several isometrically equivalent hyperbolic models, we adopt the Lorentz model [45] for its numerical stability and computational efficiency, with " + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 388, + 296, + 460 + ], + "type": "text", + "content": " set to -1 by default." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": "Lorentz Model Formally, an " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": "-dimensional Lorentz model is the Riemannian manifold " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n = (\\mathcal{L}^n,\\mathfrak{g}_x)" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathfrak{g}_x = \\mathrm{diag}(-1,1,\\dots ,1)" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " is the Riemannian metric tensor. Each point in " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " has the form " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\pmb {x} = [x_0,\\pmb {x}_s]\\in \\mathbb{R}^{n + 1},x_0 = \\sqrt{\\|\\pmb{x}_s\\|^2 + 1}\\in \\mathbb{R}" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": ". Following Chen et al. [6], we denote " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " as time axis and " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{s}" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " as spatial axes. " + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^n" + }, + { + "bbox": [ + 55, + 460, + 296, + 534 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 83, + 542, + 296, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 542, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 83, + 542, + 296, + 556 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {n} := \\left\\{\\boldsymbol {x} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {x}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = - 1, x _ {0} > 0 \\right\\}, \\tag {1}", + "image_path": "f69d90542e2e49085992a3c024c1a75a839b009944678354bb30a9d469b536fa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 565, + 230, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 230, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 230, + 578 + ], + "type": "text", + "content": "and the Lorentzian inner product given by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 118, + 586, + 296, + 600 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 586, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 118, + 586, + 296, + 600 + ], + "type": "interline_equation", + "content": "\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}} := - x _ {0} y _ {0} + \\boldsymbol {x} _ {s} ^ {\\top} \\boldsymbol {y} _ {s}. \\tag {2}", + "image_path": "017e5f5932f2a377f0f18ec893e9244ca5dc3eb10ff37960877d8c0549b9cb42.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "text", + "content": "Here " + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^n" + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "text", + "content": " is the upper sheet of hyperboloid in a " + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "inline_equation", + "content": "(n + 1)" + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "text", + "content": " dimensional Minkowski space with the origin " + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "inline_equation", + "content": "o = (1,0,\\dots ,0)" + }, + { + "bbox": [ + 55, + 609, + 296, + 633 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 633, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 296, + 657 + ], + "type": "text", + "content": "Tangent Space The tangent space at " + }, + { + "bbox": [ + 55, + 633, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 633, + 296, + 657 + ], + "type": "text", + "content": " is a Euclidean space that is orthogonal to it, defined as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 99, + 666, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 666, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 99, + 666, + 296, + 681 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\boldsymbol {x}} \\mathbb {L} ^ {n} := \\left\\{\\boldsymbol {y} \\in \\mathbb {R} ^ {n + 1} \\mid \\langle \\boldsymbol {y}, \\boldsymbol {x} \\rangle_ {\\mathcal {L}} = 0 \\right\\}. \\tag {3}", + "image_path": "9cf8c3cd14930542150f3192ea132a4287ff35a5e3aac1664c67666333e579fa.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a Euclidean subspace of " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{n + 1}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ". In particular, the tangent space at the origin " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\pmb{o}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is denoted as " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_o\\mathbb{L}^n" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": "Logarithmic and Exponential Maps The mutual mapping between the hyperbolic space " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " and the Euclidean subspace " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " can be realized by logarithmic and exponential maps. The exponential map " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\exp_{\\pmb{x}}(z)" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " can map any tangent vector " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "z\\in \\mathcal{T}_x\\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "inline_equation", + "content": "\\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 72, + 555, + 133 + ], + "type": "text", + "content": ", written as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 140, + 555, + 163 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 140, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 337, + 140, + 555, + 163 + ], + "type": "interline_equation", + "content": "\\exp_ {\\boldsymbol {x}} (\\boldsymbol {z}) = \\cosh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\boldsymbol {x} + \\sinh (\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}) \\frac {\\boldsymbol {z}}{\\| \\boldsymbol {z} \\| _ {\\mathcal {L}}}, \\tag {4}", + "image_path": "982b32a326231bed8c7d4df2ff813f6215f35dfa0197720b66ed1494ea77f083.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\| z\\|_{\\mathcal{L}} = \\sqrt{\\langle z,z\\rangle_{\\mathcal{L}}}" + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "content": " and the logarithmic map " + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\log_{x}(\\pmb {y})" + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "content": " plays an opposite role to map " + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\pmb {y}\\in \\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\pmb{x}}\\mathbb{L}^{n}" + }, + { + "bbox": [ + 313, + 172, + 555, + 198 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 205, + 555, + 233 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 205, + 555, + 233 + ], + "spans": [ + { + "bbox": [ + 326, + 205, + 555, + 233 + ], + "type": "interline_equation", + "content": "\\log_ {\\boldsymbol {x}} (\\boldsymbol {y}) = \\frac {\\operatorname {a r c o s h} (- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}})}{\\sqrt {(- \\langle \\boldsymbol {x} , \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) ^ {2} - 1}} (\\boldsymbol {y} + (\\langle \\boldsymbol {x}, \\boldsymbol {y} \\rangle_ {\\mathcal {L}}) \\boldsymbol {x}). \\tag {5}", + "image_path": "51d1910a7bf37bb52611c546c09d451bc7aa857d5c25c30bcc5786a893ac0d4c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "content": "Lorentzian centroid The weighted centroid with respect to the squared Lorentzian distance, which solves " + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\min_{\\mu \\in \\mathbb{L}^n}\\sum_{i = 1}^{m}\\nu_id^2_{\\mathcal{L}}(\\pmb {x}_i,\\mu)" + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\pmb {x}_i\\in \\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\nu_{i}\\geq 0,\\sum_{i = 1}^{m}\\nu_{i} > 0" + }, + { + "bbox": [ + 313, + 240, + 555, + 289 + ], + "type": "text", + "content": ", is denoted as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 297, + 555, + 326 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 297, + 555, + 326 + ], + "spans": [ + { + "bbox": [ + 386, + 297, + 555, + 326 + ], + "type": "interline_equation", + "content": "\\mu = \\frac {\\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i}}{| | | \\sum_ {i = 1} ^ {m} \\nu_ {i} \\boldsymbol {x} _ {i} | | _ {\\mathcal {L}} |}. \\tag {6}", + "image_path": "e2a17e4fb86a7792292e4f698610bb3e9924b2b3efc0fe32dc8e7d0dc703adf7.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 332, + 506, + 343 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 506, + 343 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 506, + 343 + ], + "type": "text", + "content": "3.2. Problem Formulation and Overview" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 349, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 445 + ], + "type": "text", + "content": "Partially Relevant Video Retrieval (PRVR) aims to retrieve videos containing a moment semantically relevant to a given text query, from a large corpus of untrimmed videos. In the PRVR database, each video has multiple moments and is associated with multiple text descriptions, with each text description corresponding to a specific moment of the related video. Critically, the temporal boundaries of these moments (i.e., start and end time points) are not annotated." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 445, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 505 + ], + "type": "text", + "content": "In this paper, we introduce HLFormer, the first hyperbolic modeling approach designed for PRVR. The proposed framework encompasses three key components: text query representation encoding, video representation encoding, and similarity computation, as illustrated in Fig. 2 (a)." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "content": "Text Representation Given a text query of " + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "content": " words, we first use a pre-trained RoBERTa [38] model to extract word-level features, which are then projected into a lower-dimensional space via a fully connected (FC) layer. A standard Transformer [53] layer is applied to obtain a sequence of " + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "content": "-dimensional contextualized feature vectors, " + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{Q} = \\{\\pmb{q}_i\\}_{i=1}^{N_q} \\in \\mathbb{R}^{N_q \\times d}" + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "content": ". Finally, we utilize a simple attention mechanism to get the sentence embedding " + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "inline_equation", + "content": "\\pmb{q} \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 505, + 556, + 602 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 345, + 611, + 555, + 644 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 611, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 345, + 611, + 555, + 644 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} = \\sum_ {i = 1} ^ {N _ {q}} \\boldsymbol {a} _ {i} ^ {q} \\times \\boldsymbol {q} _ {i}, \\quad \\boldsymbol {a} ^ {q} = \\operatorname {s o f t m a x} (\\boldsymbol {w} \\boldsymbol {Q} ^ {\\top}), \\tag {7}", + "image_path": "6f4e8795aaf91ba7b6f30faa3337f3b334b295d47c266d7ecf7cf933651383aa.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{w} \\in \\mathbb{R}^{1 \\times d}" + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "content": " is a trainable vector, and " + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "inline_equation", + "content": "\\pmb{a}^q \\in \\mathbb{R}^{1 \\times N_q}" + }, + { + "bbox": [ + 313, + 652, + 553, + 677 + ], + "type": "text", + "content": " represents the attention vector." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Video Representation Given an untrimmed video, we first extract embedding features using a pre-trained 2D or 3D CNN. Then we utilize the gaze branch and glance branch" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23076" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 425, + 251 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 425, + 251 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 425, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 425, + 251 + ], + "type": "image", + "image_path": "c3e6a22ba70d5d2764040c9022422b886c27c2307b8912ffdaf9dcc57322a9ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 152, + 255, + 288, + 265 + ], + "lines": [ + { + "bbox": [ + 152, + 255, + 288, + 265 + ], + "spans": [ + { + "bbox": [ + 152, + 255, + 288, + 265 + ], + "type": "text", + "content": "(a) The overall framework of HLFormer" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 425, + 70, + 552, + 168 + ], + "blocks": [ + { + "bbox": [ + 425, + 70, + 552, + 168 + ], + "lines": [ + { + "bbox": [ + 425, + 70, + 552, + 168 + ], + "spans": [ + { + "bbox": [ + 425, + 70, + 552, + 168 + ], + "type": "image", + "image_path": "651766bd5169af4403a9cd6563ecd8fc36a254a7ea4d66a9ca316711673281b8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 426, + 168, + 552, + 266 + ], + "blocks": [ + { + "bbox": [ + 426, + 168, + 552, + 266 + ], + "lines": [ + { + "bbox": [ + 426, + 168, + 552, + 266 + ], + "spans": [ + { + "bbox": [ + 426, + 168, + 552, + 266 + ], + "type": "image", + "image_path": "b5da0877f37d3f9a127bd3ee0399aeb98ed138a9f9ed5aef29270ea1dc8edff9.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": "Figure 2. Overview of HLFoer. (a) The sentence embedding " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " is obtained via the query branch, while the gaze and glance branches encode the video, producing frame-level embedding " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " and clip-level embedding " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " and forming the video representation " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "V_{v}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "\\pmb{q}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " learns query diversity through " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "L_{div}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " and computes similarity scores " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "S_{f}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "S_{c}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": ", while preserving partial order relations with " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "V_{v}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": ". (b) HLFoer block combines parallel Lorentz and Euclidean attention blocks for multi-space encoding, with a Mean Guided Adaptive Interaction Module for dynamic aggregation. (c) Partial Order Preservation Loss ensures the text query embedding " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " lies within the cone defined by the video embedding " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": ". The loss is zero if " + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 555, + 340 + ], + "type": "text", + "content": " is inside the cone." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": "to capture frame-level and clip-level multi-granularity video representations, respectively. In the gaze branch, we densely sample " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "M_{f}" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": " frames, denoted as " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{F} \\in \\mathbb{R}^{M_f \\times D}" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": " is the frame feature dimension. The sampled frames are processed through a fully connected (FC) layer to reduce the dimensionality to " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": ", followed by the HLFoer block to obtain frame embeddings " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{V}_{\\pmb{f}} = \\{\\pmb{f}_{i}\\}_{i=1}^{M_{f}} \\in \\mathbb{R}^{M_{f} \\times d}" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": ", capturing semantically rich frame-level information for fine-grained relevance assessment to the query. The glance branch down-samples the input along the temporal dimension to aggregate frames into clips. Following MS-SL [14], a fixed number " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "M_{c}" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": " of clips is sparsely sampled by mean pooling over consecutive frames. A fully connected layer is applied to the pooled clip features, followed by the HLFoer block, generating clip embeddings " + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{V}_{c} = \\{\\pmb{c}_{i}\\}_{i=1}^{M_{c}} \\in \\mathbb{R}^{M_{c} \\times d}" + }, + { + "bbox": [ + 54, + 361, + 297, + 567 + ], + "type": "text", + "content": ". These embeddings capture adaptive clip-level information, enabling the model to perceive relevant moments at a coarser granularity." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "text", + "content": "Similarity Computation To compute the similarity between a text-video pair " + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "inline_equation", + "content": "(\\mathcal{T},\\mathcal{V})" + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "text", + "content": " , we first measure the abovementioned embeddings " + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "inline_equation", + "content": "q,V_{f}" + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 55, + 567, + 296, + 628 + ], + "type": "text", + "content": " . Then, we employ cosine similarity along with a max operation to calculate the frame-level and clip-level similarity scores:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 79, + 638, + 296, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 638, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 79, + 638, + 296, + 656 + ], + "type": "interline_equation", + "content": "S _ {f} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {1}\\right), \\dots , \\cos \\left(\\boldsymbol {q}, \\boldsymbol {f} _ {M _ {f}}\\right) \\right\\}, \\tag {8}", + "image_path": "95f7336e7fc1e468c61d8834f9b00f92ed6d4aab571d456b07c802d6879fba67.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 83, + 654, + 271, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 654, + 271, + 666 + ], + "spans": [ + { + "bbox": [ + 83, + 654, + 271, + 666 + ], + "type": "interline_equation", + "content": "S _ {c} (\\mathcal {T}, \\mathcal {V}) = \\max \\left\\{\\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {1}), \\dots , \\cos (\\boldsymbol {q}, \\boldsymbol {c} _ {M _ {c}}) \\right\\}.", + "image_path": "69c0d219e4c9f3fd36c1a2414dc49ae383a3b6d92a46334f8d57b4bbf3c27c5b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 280, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 280, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 280, + 690 + ], + "type": "text", + "content": "Next, we compute the overall text-video pair similarity:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 94, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "S (\\mathcal {T}, \\mathcal {V}) = \\alpha_ {f} S _ {f} (\\mathcal {T}, \\mathcal {V}) + \\alpha_ {c} S _ {c} (\\mathcal {T}, \\mathcal {V}), \\tag {9}", + "image_path": "848864db5bad4343b30b043481c41bf216b4fe06af2b2352b6588b192464c2d7.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "inline_equation", + "content": "\\alpha_{f},\\alpha_{c}\\in [0,1]" + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "content": " are hyper-parameters satisfying " + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "inline_equation", + "content": "\\alpha_{f} + \\alpha_{c} = 1" + }, + { + "bbox": [ + 313, + 361, + 555, + 398 + ], + "type": "text", + "content": ". Finally, we retrieve and rank partially relevant videos based on the computed similarity scores." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 406, + 417, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 417, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 417, + 417 + ], + "type": "text", + "content": "3.3. HlFormer Block" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 423, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 518 + ], + "type": "text", + "content": "The HLFoer block constitutes the core of our method. As shown in Fig. 2 (b), it comprises three key modules: (i) Euclidean Attention Block, capturing fine-grained visual features in Euclidean space; (ii) Lorentz Attention Block, projecting video embeddings into hyperbolic Lorentz space for capturing the hierarchical structures of video; (iii) Mean-Guided Adaptive Interaction Module, dynamically fusing hybrid-space features. We describe the details below." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": "Euclidean Attention Block Given M feature embeddings " + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "inline_equation", + "content": "\\pmb{x} \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 519, + 554, + 567 + ], + "type": "text", + "content": " is the feature dimension, the Euclidean Attention Block utilizes Euclidean Gaussian Attention [57] to capture multi-scale visual features, expressed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 320, + 576, + 555, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 576, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 320, + 576, + 555, + 604 + ], + "type": "interline_equation", + "content": "\\operatorname {G A} (\\boldsymbol {x}) = \\operatorname {s o f t m a x} \\left(\\mathcal {M} _ {\\sigma} ^ {g} \\odot \\frac {\\boldsymbol {x} W ^ {q} \\left(\\boldsymbol {x} W ^ {k}\\right) ^ {\\top}}{\\sqrt {d _ {h}}}\\right) \\boldsymbol {x} W ^ {v}, \\tag {10}", + "image_path": "a3eefff1a92e75f15151a12763ea1a74b4d6d73e0cf0266a7dfcc18cc8046555.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\sigma}^{g}" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": " is the Gaussian matrix with elements " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\sigma}^{g}(i,j) = \\frac{1}{2\\pi} e^{-\\frac{(j - i)^2}{\\sigma^2}}" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": " denotes the variance. By varying " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": ", feature interactions at different scales are modeled, generating video features with multiple receptive fields. " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "W^{q}, W^{k}, W^{v}" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": " are linear projections, while " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "d_{h}" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": " is the latent attention dimension, " + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 613, + 556, + 713 + ], + "type": "text", + "content": " denotes element-wise product. Finally, We replace the self-attention in Transformer block with Euclidean Gaussian attention to form the Euclidean Attention Block." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23077" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "Lorentz Attention Block Given extracted Euclidean video embeddings " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}^{E} \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ", we first project it to " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M \\times n}" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " via a linear layer and apply scaling. Let " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "o := [1,0,\\dots,0]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " be the origin on the Lorentz manifold, satisfying " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "\\langle o, [0, x_{\\mathrm{in}}^{E}] \\rangle_{\\mathcal{L}} = 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". Thus, " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "[0, x_{\\mathrm{in}}^{E}]" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": " can be interpreted as a vector in the tangent space at " + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": ". The Lorentz embedding is then obtained via the exponential map Eq. (4):" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 74, + 159, + 296, + 175 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 159, + 296, + 175 + ], + "spans": [ + { + "bbox": [ + 74, + 159, + 296, + 175 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {\\text {i n}} ^ {\\mathcal {L}} = \\exp_ {\\boldsymbol {o}} \\left(\\left[ 0, \\beta \\boldsymbol {x} _ {\\text {i n}} ^ {E} W _ {1} \\right]\\right) \\in \\mathbb {L} ^ {n}, \\mathbb {R} ^ {M \\times (n + 1)}, \\tag {11}", + "image_path": "da7c4216a00289c3e0d89cac3809ecf3629ce996f859759e43e2ec83bedfd071.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "text", + "content": " denotes the linear layer, " + }, + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 178, + 296, + 202 + ], + "type": "text", + "content": " is a learnable scaling factor to prevent numerical overflow." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 202, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 262 + ], + "type": "text", + "content": "Having obtained the Lorentz embedding " + }, + { + "bbox": [ + 55, + 202, + 296, + 262 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{in}}^{\\mathcal{L}}" + }, + { + "bbox": [ + 55, + 202, + 296, + 262 + ], + "type": "text", + "content": ", which inherently exhibits a prominent hierarchical structure due to the hyperbolic space properties, we next design a Lorentz linear transformation and Lorentz self-attention module to capture and fully leverage the hierarchical priors." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": "Inspired by prior studies [6, 32], we redefine the Lorentz linear layer to learn a matrix " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "M = \\begin{bmatrix} \\boldsymbol{p}^{\\top} \\\\ \\boldsymbol{W} \\end{bmatrix}" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\boldsymbol{p} \\in \\mathbb{R}^{n+1}" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": " is a weight parameter and " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\boldsymbol{W} \\in \\mathbb{R}^{m \\times (n+1)}" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": " ensures that " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "\\forall \\boldsymbol{x} \\in \\mathbb{L}^n" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "f_{\\boldsymbol{x}}(\\boldsymbol{M})\\boldsymbol{x} \\in \\mathbb{L}^m" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": ". Specifically, the transformation matrix " + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "inline_equation", + "content": "f_{\\boldsymbol{x}}(\\boldsymbol{M})" + }, + { + "bbox": [ + 55, + 262, + 296, + 335 + ], + "type": "text", + "content": " is expressed as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 339, + 296, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 339, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 77, + 339, + 296, + 372 + ], + "type": "interline_equation", + "content": "f _ {\\boldsymbol {x}} (\\boldsymbol {M}) = f _ {\\boldsymbol {x}} \\left(\\left[ \\begin{array}{c} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W} \\end{array} \\right]\\right) = \\left[ \\begin{array}{c} \\frac {\\sqrt {\\| W \\boldsymbol {x} \\| ^ {2} + 1}}{\\boldsymbol {p} ^ {\\top} \\boldsymbol {x}} \\boldsymbol {p} ^ {\\top} \\\\ \\boldsymbol {W}. \\end{array} \\right] \\tag {12}", + "image_path": "a38485a664437421cb59eedc8f05b7c73954194e4f60b604d115bdea1a491041.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 376, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 296, + 399 + ], + "type": "text", + "content": "Adding other components including normalization, the final definition of the Lorentz Linear layer becomes:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 403, + 296, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 403, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 104, + 403, + 296, + 426 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} = \\operatorname {H L} (\\boldsymbol {x}) = \\left[ \\begin{array}{c} \\sqrt {\\| \\phi (\\boldsymbol {W} \\boldsymbol {x} , \\boldsymbol {p}) \\| ^ {2} + 1} \\\\ \\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) \\end{array} \\right], \\tag {13}", + "image_path": "8521f918f0dc3fcc26dccbf1ac599ed8c5fb619d0af879bfd31b2375e1b3c88d.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 429, + 155, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 155, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 155, + 440 + ], + "type": "text", + "content": "with operation function:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 72, + 444, + 295, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 444, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 72, + 444, + 295, + 472 + ], + "type": "interline_equation", + "content": "\\phi (\\boldsymbol {W} \\boldsymbol {x}, \\boldsymbol {p}) = \\frac {\\lambda \\left(\\boldsymbol {p} ^ {\\top} \\boldsymbol {x} + b ^ {\\prime}\\right)}{\\| \\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b} \\|} (\\boldsymbol {W} h (\\boldsymbol {x}) + \\boldsymbol {b}), \\tag {14}", + "image_path": "09cf4e53f940271556e114d5594dbb63389598e83cb987cf4444dc9ebb30bcb7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\pmb{b}" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "b'" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " are bias terms, " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "\\lambda > 0" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " regulates the scaling range. " + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 475, + 295, + 499 + ], + "type": "text", + "content": " denotes the activation function." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": "Based on the Lorentz Linear Layer, we propose a Lorentz self-attention module that integrates Gaussian constraints into feature interactions, enabling multiscale and hierarchical video embeddings in hyperbolic space. Specifically, given a hyperbolic video embedding " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{in}}^{\\mathcal{L}} \\in \\mathbb{L}^{n}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M \\times (n + 1)}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ", we first obtain the attention query " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{K}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": " using Eq. (13), all in the shape of " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{M \\times (n + 1)}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ". We calculate attention scores based on Eq. (6) and apply a Gaussian matrix " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{\\sigma}^{g} \\in \\mathbb{R}^{M \\times M}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": " for element-wise multiplication with the score matrix to obtain a multi-scale receptive field. The output is defined as " + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{out}}^{\\mathcal{L}} = \\{\\pmb{\\mu}_1, \\dots, \\pmb{\\mu}_{|\\mathcal{Q}|}\\} \\in \\mathbb{R}^{M \\times (n + 1)}" + }, + { + "bbox": [ + 55, + 500, + 296, + 632 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 93, + 636, + 296, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 636, + 296, + 681 + ], + "spans": [ + { + "bbox": [ + 93, + 636, + 296, + 681 + ], + "type": "interline_equation", + "content": "S _ {i j} = \\frac {\\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {j}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , j)}{\\sqrt {(n + 1)}}\\right)}{\\sum_ {k = 1} ^ {| \\mathcal {K} |} \\exp \\left(\\frac {- d _ {\\mathcal {L}} ^ {2} \\left(\\boldsymbol {q} _ {i} , \\boldsymbol {k} _ {k}\\right) \\odot \\mathcal {M} _ {\\sigma} ^ {g} (i , k)}{\\sqrt {(n + 1)}}\\right)}, \\tag {15}", + "image_path": "1680f253414b022cb631bb0eec0698e5992202fbc252eeaa0125da8183703129.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 96, + 684, + 203, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 684, + 203, + 717 + ], + "spans": [ + { + "bbox": [ + 96, + 684, + 203, + 717 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\mu} _ {i} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {K} |} S _ {i j} \\boldsymbol {v} _ {j}}{\\left| \\left| \\sum_ {k = 1} ^ {| \\mathcal {K} |} S _ {i k} \\boldsymbol {v} _ {k} \\right| \\right| _ {\\mathcal {L}}} ,", + "image_path": "5fb455467de3ce5b7290063a588d4abb105a9d7459e1329a3b6d6828c37729dc.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 72, + 555, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 85 + ], + "type": "text", + "content": "the squared Lorentzian distance " + }, + { + "bbox": [ + 313, + 72, + 555, + 85 + ], + "type": "inline_equation", + "content": "d_{\\mathcal{L}}^2 (\\pmb {a},\\pmb {b}) = -2 - 2\\langle \\pmb {a},\\pmb {b}\\rangle_{\\mathcal{L}}" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": "After computing " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{out}}^{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": ", we apply the logarithmic map Eq. (5), while discarding the time axis, to obtain the Euclidean space embedding " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{mid}}^{E}" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": ". Then, the output " + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "inline_equation", + "content": "x_{\\mathrm{out}}^{E}" + }, + { + "bbox": [ + 313, + 85, + 555, + 133 + ], + "type": "text", + "content": " is obtained through a Linear Layer followed by rescaling:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 323, + 141, + 555, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 141, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 323, + 141, + 555, + 184 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {x} _ {\\text {m i d}} ^ {E} = \\operatorname {d r o p \\_ t i m e \\_ a x i s} \\left(\\log_ {\\boldsymbol {o}} \\left(\\boldsymbol {x} _ {\\text {o u t}} ^ {\\boldsymbol {\\mathcal {L}}}\\right)\\right) \\in \\mathbb {R} ^ {M \\times n}, \\\\ \\boldsymbol {x} _ {\\text {o u t}} ^ {E} = \\frac {\\boldsymbol {x} _ {\\text {m i d}} ^ {E} W _ {2}}{\\beta} \\in \\mathbb {R} ^ {M \\times d}, \\tag {16} \\\\ \\end{array}", + "image_path": "6ede3a056e3819b07807b3ff3a62aa6c810878e14081c73f073465e1b6e0c844.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "inline_equation", + "content": "W_{2}\\in \\mathbb{R}^{n\\times d}" + }, + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 194, + 555, + 231 + ], + "type": "text", + "content": " is the scale factor in Eq. (11). Finally, We replace the self-attention in Transformer block with Lorentz attention to form the Lorentz Attention Block." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "content": "Mean-Guided Adaptive Interaction Module We arrange " + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "content": " Lorentz and " + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "inline_equation", + "content": "N_{E}" + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "content": " Euclidean Attention Blocks in parallel to construct " + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "inline_equation", + "content": "N_{O}" + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "content": " Gaussian Attention Blocks for multi-scale hybrid-space video embeddings. To integrate these features, we introduce a Mean-Guided Adaptive Interaction Module, which utilizes globally pooled features to compute dynamic aggregation weights. Specifically, we first obtain the global query " + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "inline_equation", + "content": "\\varphi \\in \\mathbb{R}^{1\\times d}" + }, + { + "bbox": [ + 313, + 232, + 555, + 352 + ], + "type": "text", + "content": " and compute aggregation weights via a Cross Attention Block consisting of a cross-attention layer (CA) followed by a fully connected layer (FC):" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 347, + 362, + 474, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 362, + 474, + 376 + ], + "spans": [ + { + "bbox": [ + 347, + 362, + 474, + 376 + ], + "type": "interline_equation", + "content": "\\boldsymbol {\\varphi} = \\operatorname {M e a n} \\left(\\boldsymbol {x} _ {\\sigma_ {1}}, \\boldsymbol {x} _ {\\sigma_ {2}},.., \\boldsymbol {x} _ {\\sigma_ {N _ {o}}}\\right),", + "image_path": "a924a3893b122824ce0f642980547083adc47c61fe6c8d656e91be1caa8635e2.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 345, + 376, + 526, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 376, + 526, + 390 + ], + "spans": [ + { + "bbox": [ + 345, + 376, + 526, + 390 + ], + "type": "interline_equation", + "content": "w _ {i} = \\operatorname {F C} \\left(\\operatorname {C A} \\left(\\varphi , \\boldsymbol {x} _ {\\sigma_ {i}}, \\boldsymbol {x} _ {\\sigma_ {i}}\\right)\\right), i = 1, 2, \\dots , N _ {o},", + "image_path": "7f2bf3655dc352a17de4573e1298f79c07129f8e80b23255ec49c5494f526f8b.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 338, + 392, + 555, + 423 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 392, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 338, + 392, + 555, + 423 + ], + "type": "interline_equation", + "content": "\\tilde {w} _ {i, j} = \\frac {e ^ {w _ {i , j} / \\tau}}{\\sum_ {k = 1} ^ {N _ {o}} e ^ {w _ {k , j} / \\tau}}, j = 1, \\dots , M, \\tag {17}", + "image_path": "7f13da151968ea67300dde6cda7b3db44d92a8895eec8a92b2732dd4755f4844.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 343, + 424, + 481, + 456 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 424, + 481, + 456 + ], + "spans": [ + { + "bbox": [ + 343, + 424, + 481, + 456 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {x}} _ {j} = \\sum_ {i = 1} ^ {N _ {o}} \\tilde {w} _ {i, j} \\boldsymbol {x} _ {\\sigma_ {i, j}}, j = 1, \\dots , M,", + "image_path": "f3b04a2859d4c591b0d33faf6dc9985557449c31db67c1ba49ff5af831a092f3.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 328, + 459, + 468, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 459, + 468, + 472 + ], + "spans": [ + { + "bbox": [ + 328, + 459, + 468, + 472 + ], + "type": "interline_equation", + "content": "\\boldsymbol {x} _ {\\text {M A I M}} = \\operatorname {C o n c a t} \\left(\\tilde {\\boldsymbol {x}} _ {1}, \\tilde {\\boldsymbol {x}} _ {2}, \\dots , \\tilde {\\boldsymbol {x}} _ {M}\\right),", + "image_path": "43787b863c04122ab80849df452e25f8a2c39bad06481c71a1b24171f3c2f33c.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\sigma_i} \\in \\mathbb{R}^{M \\times d}" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " denotes the output of the " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": "-th Gaussian block and " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " corresponds to the number of time points (i.e., clips or frames). " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "w_i \\in \\mathbb{R}^M" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " represents the aggregation weights for the " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": "-th Gaussian block, and " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " is the temperature factor. " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\tilde{\\pmb{x}}_j \\in \\mathbb{R}^d" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " denotes the aggregated feature at time point " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "inline_equation", + "content": "\\pmb{x}_{\\mathrm{MAIM}}" + }, + { + "bbox": [ + 313, + 483, + 555, + 555 + ], + "type": "text", + "content": " is the final output." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 563, + 432, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 432, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 432, + 577 + ], + "type": "text", + "content": "3.4. Learning Objectives" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Given the partial relevance in PRVR, where each video fully entails its corresponding text, a partial order relationship is established, with the text-query semantically subsumed by the video: text " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\prec" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " video. Inspired by MERU [10], we propose the Partial Order Preservation Loss to enforce this relationship in Hyperbolic Space. Given " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "V_{f}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "V_{c}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": " from Sec. 3.2, a simple attention module similar to Eq. (7) is applied, followed by mean pooling to get the unified video representation " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "V_{v}" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ". The video and text representations are then mapped to Lorentz space via the exponential map, yielding " + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "inline_equation", + "content": "\\pmb{v},\\pmb{t}\\in \\mathbb{L}^n" + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": ", as shown in Fig. 2(c). We define an entailment" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23078" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 85 + ], + "type": "text", + "content": "cone for each " + }, + { + "bbox": [ + 55, + 72, + 297, + 85 + ], + "type": "inline_equation", + "content": "\\pmb{v}" + }, + { + "bbox": [ + 55, + 72, + 297, + 85 + ], + "type": "text", + "content": ", which is characterized by the half-aperture:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 117, + 92, + 296, + 120 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 92, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 117, + 92, + 296, + 120 + ], + "type": "interline_equation", + "content": "\\mathbf {H A} (\\boldsymbol {v}) = \\arcsin \\left(\\frac {2 c}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\|}\\right). \\tag {18}", + "image_path": "4937436f2ec778e6898a1ee80234de07914e21f9c0fb429c11b0a2f0c91523dd.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "inline_equation", + "content": "c = 0.1" + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "text", + "content": " is used to define the boundary conditions near the origin. We measure the exterior angle " + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\mathbf{EA}(\\boldsymbol{v},\\boldsymbol{t}) = \\pi - \\angle Ovt" + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "text", + "content": " to penalize cases where " + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "inline_equation", + "content": "\\boldsymbol{t}" + }, + { + "bbox": [ + 55, + 127, + 296, + 163 + ], + "type": "text", + "content": " falls outside the entailment cone:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 70, + 171, + 296, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 171, + 296, + 211 + ], + "spans": [ + { + "bbox": [ + 70, + 171, + 296, + 211 + ], + "type": "interline_equation", + "content": "\\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) = \\operatorname {a r c c o s} \\left(\\frac {t _ {0} + v _ {0} \\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}}{\\| \\boldsymbol {v} _ {\\boldsymbol {s}} \\| \\sqrt {\\left(\\langle \\boldsymbol {v} , \\boldsymbol {t} \\rangle_ {\\mathcal {L}}\\right) ^ {2} - 1}}\\right). \\tag {19}", + "image_path": "372bf6ddadbe27824c3b63e2b2482716f20c51161e809f927615fe7d80ed92ec.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 219, + 253, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 219, + 253, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 219, + 253, + 232 + ], + "type": "text", + "content": "The Loss for a single video-text pair is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 87, + 240, + 296, + 254 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 240, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 87, + 240, + 296, + 254 + ], + "type": "interline_equation", + "content": "L _ {p o p} (\\boldsymbol {v}, \\boldsymbol {t}) = \\max (0, \\mathbf {E A} (\\boldsymbol {v}, \\boldsymbol {t}) - \\mathbf {H A} (\\boldsymbol {v})). \\tag {20}", + "image_path": "818e73d95eaa115b7e1d69524493276d313c15b1c32dbe6a8157e9b766946f28.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "text", + "content": "Besides, following MS-SL [14], we use the standard similarity retrieval loss to train the model, denoted as " + }, + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "inline_equation", + "content": "L_{sim}" + }, + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "text", + "content": ". Meanwhile, the query diversity [57] " + }, + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "inline_equation", + "content": "L_{div}" + }, + { + "bbox": [ + 55, + 262, + 297, + 310 + ], + "type": "text", + "content": " is used to improve retrieval performance. The aggregate loss is defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 104, + 320, + 296, + 334 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 320, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 104, + 320, + 296, + 334 + ], + "type": "interline_equation", + "content": "L _ {a g g} = L _ {s i m} + \\lambda_ {1} L _ {d i v} + \\lambda_ {2} L _ {p o p}, \\tag {21}", + "image_path": "9f0ab8953a864a032f851a3d99c07979d877c75fc1a26e97d36858d14bd15cef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{1}" + }, + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "type": "inline_equation", + "content": "\\lambda_{2}" + }, + { + "bbox": [ + 55, + 342, + 296, + 354 + ], + "type": "text", + "content": " are hyper-parameters that balance learning losses." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 365, + 137, + 378 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 137, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 137, + 378 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 384, + 171, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 171, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 171, + 397 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 401, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 297, + 567 + ], + "type": "text", + "content": "Datasets We conduct experiments on three benchmark datasets: (i) ActivityNet Captions [28], which comprises approximately 20K YouTube videos with an average duration of 118 seconds. Each video contains an average of 3.7 annotated moments with corresponding textual descriptions. (ii) TV show Retrieval (TVR) [30], consisting of 21.8K videos sourced from six TV shows. Each video is associated with five natural language descriptions covering different moments. (iii) Charades-STA [22], which includes 6,670 videos annotated with 16,128 sentence descriptions. On average, each video contains approximately 2.4 moments with corresponding textual queries. We adopt the same data split as used in prior studies[14, 57]. It is important to note that the moment annotations are unavailable in the PRVR task." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "content": "Metrics Following previous works [14, 57], we adopt rank-based evaluation metrics, specifically " + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "inline_equation", + "content": "R@K" + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "inline_equation", + "content": "K = 1, 5, 10, 100" + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "content": "). The metric " + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "inline_equation", + "content": "R@K" + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "content": " represents the proportion of queries for which the correct item appears within the top " + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 569, + 296, + 665 + ], + "type": "text", + "content": " positions of the ranking list. All results are reported as percentages (\\%), where higher values indicate superior retrieval performance. To facilitate an overall comparison, we also report the Sum of all Recalls (SumR)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 188, + 685 + ], + "type": "text", + "content": "4.2. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Data Processing For video representations on TVR, we employ the feature set provided by Lei et al. [30], which" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 179 + ], + "type": "text", + "content": "comprises 3,072-dimensional visual features obtained by concatenating frame-level ResNet152 features [23] and segment-level I3D features [2]. For ActivityNet Captions and Charades-STA, we only utilize I3D features as provided by Zhang et al. [62] and Mun et al. [43], respectively. For sentence representations, we adopt the 768-dimensional RoBERTa features supplied by Lei et al. [30] for TVR. On ActivityNet Captions and Charades-STA, we employ 1,024-dimensional RoBERTa features extracted using MS-SL[14]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": "Model Configurations The HLFormaler block consists of 8 Gaussian blocks " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "(N_O = 8)" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": ", 4 Lorentz Attention blocks " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "(N_{\\mathcal{L}} = 4)" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": ", with Gaussian variances ranging from " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "2^{1}" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "2^{N_{\\mathcal{L}} - 1}" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": ", and 4 Euclidean Attention blocks " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "(N_E = 4)" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": ", with Gaussian variances ranging from " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "2^{1}" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "2^{N_{E} - 1}" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "\\infty" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": ". The latent dimension " + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "inline_equation", + "content": "d = 384" + }, + { + "bbox": [ + 313, + 180, + 556, + 251 + ], + "type": "text", + "content": " with 4 attention heads." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 253, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 253, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 253, + 555, + 312 + ], + "type": "text", + "content": "Training Configurations We employ the Adam optimizer with a mini-batch size of 128 and set the number of epochs to 100. The model is implemented using PyTorch and trained on one Nvidia RTX 3080 Ti GPU. We adopt a learning rate adjustment schedule similar to MS-SL." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 322, + 497, + 334 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 497, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 497, + 334 + ], + "type": "text", + "content": "4.3. Comparison with State-of-the arts" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 339, + 556, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 556, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 556, + 434 + ], + "type": "text", + "content": "Baselines We select six representative PRVR baselines for comparison: MS-SL [14], PEAN [26], LH [20], BGM-Net [60], GMMFormer [57], and DL-DKD [16]. We also compare HLFomer with methods for T2VR and VCMR. For T2VR, we select six T2VR models: CE [37], HGR [5], DE++ [13], RIVRL [15], CLIP4Clip [40], Cap4Video [59], For VCMR, we consider four models: XML [30], ReLoCLNet [63], CONQUER [25] and JSG[7]." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "content": "Retrieval Performance Tab. 1 presents the retrieval performance of various models on three large-scale video datasets. As observed, T2VR models, designed to capture overall video-text relevance, underperform for PRVR. VCMR models, which focus on moment retrieval, achieve better results. PRVR methods perform best as they are specifically designed for this task. Attributed to hyperbolic space learning and effective utilization of video hierarchical structure priors, HLFformer consistently surpasses all baselines. It outperforms DL-DKD by " + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "inline_equation", + "content": "4.9\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "inline_equation", + "content": "4.3\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "content": " in SumR on ActivityNet Captions and TVR, respectively, and exceeds PEAN by " + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 435, + 556, + 579 + ], + "type": "text", + "content": " on Charades-STA." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 588, + 411, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 411, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 411, + 601 + ], + "type": "text", + "content": "4.4. Model Analyses" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": "Efficacy of Temporal Modeling Design We perform ablation studies to examine the effect of the attention block number " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " and the attention mechanism ratio " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "N_{L} / N_{E}" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": ", with results shown in Fig. 3. Model performance improves as " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "N_{o}" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": " increases, then stabilizes or declines when " + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "inline_equation", + "content": "N_{o} \\geq 8" + }, + { + "bbox": [ + 313, + 605, + 556, + 713 + ], + "type": "text", + "content": ". Even with only two attention blocks, HLFomer surpasses most competing methods. Furthermore, using solely Euclidean or Lorentz attention blocks results in suboptimal performance, whereas the hybrid attention block achieves the best results." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23079" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 553, + 285 + ], + "type": "table", + "html": "
ModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
T2VR
HGR [5]4.015.024.863.2107.01.23.87.333.445.71.74.98.335.250.1
RIVRL [15]5.218.028.266.4117.81.65.69.437.754.39.423.432.270.6135.6
DE++ [13]5.318.429.268.0121.01.75.69.637.154.18.821.930.267.4128.3
CE [37]5.519.129.971.1125.61.34.57.336.049.13.712.820.164.5101.1
CLIP4Clip [40]5.919.330.471.6127.31.86.510.944.263.49.924.334.372.5141.0
Cap4Video [59]6.320.430.972.6130.21.96.711.345.065.010.326.436.874.0147.5
VCMR
ReLoCLNet [63]5.718.930.072.0126.61.25.410.045.662.310.026.537.381.3155.1
XML [30]5.319.430.673.1128.41.66.010.146.964.610.728.138.180.3157.1
CONQUER [25]6.520.431.874.3133.11.86.310.347.566.011.028.939.681.3160.8
JSG [7]6.822.734.876.1140.52.47.712.849.872.7-----
PRVR
MS-SL [14]7.122.534.775.8140.11.87.111.847.768.413.532.143.483.4172.4
PEAN [26]7.423.035.575.9141.82.78.113.550.374.713.532.844.183.9174.2
LH [20]7.423.535.875.8142.42.17.512.950.172.713.233.244.485.5176.3
BGM-Net [60]7.223.836.076.9143.91.97.412.250.171.614.134.745.985.2179.9
GMMFormer [57]8.324.936.776.1146.02.17.812.550.672.913.933.344.584.9176.6
DL-DKD [16]8.025.037.577.1147.6-----14.434.945.884.9179.9
HLFormer (ours)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
", + "image_path": "7ab179770c422cbc6cd1ea48bd7ff400f23059a220a109d3ce568b30638f3186.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 327, + 553, + 463 + ], + "blocks": [ + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 555, + 316 + ], + "type": "text", + "content": "Table 1. Retrieval performance of HLFomer and other faithful methods on ActivityNet Captions, Charades-STA and TVR. State-of-the-art performance is highlighted in bold. “-” indicates that the corresponding results are unavailable." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 327, + 553, + 463 + ], + "lines": [ + { + "bbox": [ + 58, + 327, + 553, + 463 + ], + "spans": [ + { + "bbox": [ + 58, + 327, + 553, + 463 + ], + "type": "table", + "html": "
IDModelActivityNet CaptionsCharades-STATVR
R@1R@5R@10R@100SumRR@1R@5R@10R@100SumRR@1R@5R@10R@100SumR
(0)HLFormer (full)8.727.140.179.0154.92.68.513.754.078.715.737.148.586.4187.7
Efficacy of Multi-scale Branches
(1)w/o gaze branch7.624.436.777.3146.11.88.013.950.874.513.934.045.285.3178.3
(2)w/o glance branch6.421.733.675.4137.21.67.713.148.470.811.430.541.882.4166.1
Efficacy of Different Loss Terms
(3)Lsim Only7.725.038.178.3149.12.08.113.252.075.315.136.247.886.0185.2
(4)w/o Ldiv8.526.639.678.8153.52.07.813.653.076.415.736.448.486.0186.5
(5)w/o Lpop8.626.939.778.8154.02.28.414.053.077.615.636.848.486.0186.8
Efficacy of various Aggregation Strategies
(6)w/ MP8.525.738.277.8150.22.08.013.252.175.315.236.547.486.0185.1
(7)w/ CL8.726.839.578.6153.62.08.213.952.076.115.336.948.486.0186.6
", + "image_path": "88811a4f4c6becef4ec94de9e1055bf5110304d6d8bd24ed4c8e95b9e5f8c8bf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 168, + 472, + 441, + 483 + ], + "lines": [ + { + "bbox": [ + 168, + 472, + 441, + 483 + ], + "spans": [ + { + "bbox": [ + 168, + 472, + 441, + 483 + ], + "type": "text", + "content": "Table 2. Ablation Study of HLFomer. The best scores are marked in bold." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 504, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 504, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 54, + 504, + 295, + 612 + ], + "type": "text", + "content": "This may be attributed to the differences in representational focus: Euclidean space emphasizes fine-grained local feature learning and sometimes overlooks global hierarchical structures, while hyperbolic space prioritizes global hierarchical relationships at the expense of local details. Moreover, hyperbolic space tends to be more sensitive to noise and numerically unstable. By integrating hybrid spaces, HLFormer achieves mutual compensation, enhancing representation learning and facilitating video semantic understanding." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 295, + 714 + ], + "type": "text", + "content": "Efficacy of Hyperbolic Learning Hyperbolic learning demonstrates significant advantages in capturing the hierarchical structure of videos. As illustrated in Fig. 4(a), embeddings learned solely in Euclidean space exhibit indistinct cluster boundaries, with red and green points at the periphery closely interspersed. In contrast, Fig. 4(b) demonstrates that incorporating Lorentz attention facilitates the learning of more discriminative representations, while refining moment" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 504, + 553, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 553, + 539 + ], + "type": "text", + "content": "cluster boundaries, increasing inter-moment separation, and compacting intra-moment frame distributions, revealing a more pronounced hierarchical structure." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": "Efficacy of Multi-scale Branches To evaluate the effectiveness of the multi-scale branches, we conduct comparative experiments by removing either the glance clip-level branch or the gaze frame-level branch. As shown in Tab. 2, the absence of any branch leads to a noticeable performance degradation. These results not only validate the efficacy of the coarse-to-fine multi-granularity retrieval mechanism but also highlight the complementary nature of the two branches. Efficacies of Different Loss Terms To analyze the effectiveness of three loss terms (i.e. " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{sim}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{div}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ") of HLFformer, we construct several HLFformer variants: (i) " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{sim}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": " Only: train the model with merely " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{sim}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ". (ii) w/o " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{div}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ": We train the model without query diverse learning. (iii) w/o " + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 313, + 545, + 555, + 715 + ], + "type": "text", + "content": ": HLFformer removes the partial order preservation task." + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23080" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 71, + 172, + 125 + ], + "blocks": [ + { + "bbox": [ + 63, + 71, + 172, + 125 + ], + "lines": [ + { + "bbox": [ + 63, + 71, + 172, + 125 + ], + "spans": [ + { + "bbox": [ + 63, + 71, + 172, + 125 + ], + "type": "image", + "image_path": "2bff62a8f9e2f8e23a6877da164dd58230f6a41f12d16ad61bba59a01564b888.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 179, + 72, + 288, + 124 + ], + "blocks": [ + { + "bbox": [ + 179, + 72, + 288, + 124 + ], + "lines": [ + { + "bbox": [ + 179, + 72, + 288, + 124 + ], + "spans": [ + { + "bbox": [ + 179, + 72, + 288, + 124 + ], + "type": "image", + "image_path": "7e70021359ec2b0f54facb86ba45758fe6918dfcf734fcaba37433c427b018d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 63, + 140, + 172, + 194 + ], + "blocks": [ + { + "bbox": [ + 139, + 130, + 212, + 139 + ], + "lines": [ + { + "bbox": [ + 139, + 130, + 212, + 139 + ], + "spans": [ + { + "bbox": [ + 139, + 130, + 212, + 139 + ], + "type": "text", + "content": "(a) ActivityNet Captions" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 140, + 172, + 194 + ], + "lines": [ + { + "bbox": [ + 63, + 140, + 172, + 194 + ], + "spans": [ + { + "bbox": [ + 63, + 140, + 172, + 194 + ], + "type": "image", + "image_path": "44a9f3a7342add11f5b8893b13b4b9628917f888220e5ce513bf486665c6196d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 180, + 140, + 288, + 194 + ], + "blocks": [ + { + "bbox": [ + 180, + 140, + 288, + 194 + ], + "lines": [ + { + "bbox": [ + 180, + 140, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 180, + 140, + 288, + 194 + ], + "type": "image", + "image_path": "48c6f51ec532029942e43de228899d1de8cff8c74e4231bc338201247b35abf3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 63, + 209, + 172, + 260 + ], + "blocks": [ + { + "bbox": [ + 145, + 199, + 205, + 208 + ], + "lines": [ + { + "bbox": [ + 145, + 199, + 205, + 208 + ], + "spans": [ + { + "bbox": [ + 145, + 199, + 205, + 208 + ], + "type": "text", + "content": "(b) Charades-STA" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 63, + 209, + 172, + 260 + ], + "lines": [ + { + "bbox": [ + 63, + 209, + 172, + 260 + ], + "spans": [ + { + "bbox": [ + 63, + 209, + 172, + 260 + ], + "type": "image", + "image_path": "f3a4a1e670b44835a18105849a5dcc0124e2fdea302d1186cc924b7cb8cabe0e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 180, + 209, + 288, + 260 + ], + "blocks": [ + { + "bbox": [ + 180, + 209, + 288, + 260 + ], + "lines": [ + { + "bbox": [ + 180, + 209, + 288, + 260 + ], + "spans": [ + { + "bbox": [ + 180, + 209, + 288, + 260 + ], + "type": "image", + "image_path": "4be7380b2f802dd6b3975888c5fd189908fe19af24cad94e6ddf6231f09079fd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 161, + 266, + 190, + 275 + ], + "lines": [ + { + "bbox": [ + 161, + 266, + 190, + 275 + ], + "spans": [ + { + "bbox": [ + 161, + 266, + 190, + 275 + ], + "type": "text", + "content": "(c) TVR" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 67, + 321, + 175, + 427 + ], + "blocks": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "lines": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 295, + 306 + ], + "type": "text", + "content": "Figure 3. The influence of different attention blocks, with default settings marked in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 321, + 175, + 427 + ], + "lines": [ + { + "bbox": [ + 67, + 321, + 175, + 427 + ], + "spans": [ + { + "bbox": [ + 67, + 321, + 175, + 427 + ], + "type": "image", + "image_path": "49af380d9bb3f0b523b6cc8d740ff1b2ad2e194bce366a8d9212331623dba049.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 430, + 168, + 441 + ], + "lines": [ + { + "bbox": [ + 74, + 430, + 168, + 441 + ], + "spans": [ + { + "bbox": [ + 74, + 430, + 168, + 441 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 74, + 430, + 168, + 441 + ], + "type": "inline_equation", + "content": "w / o" + }, + { + "bbox": [ + 74, + 430, + 168, + 441 + ], + "type": "text", + "content": " Lorentz Attention" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 176, + 321, + 284, + 428 + ], + "blocks": [ + { + "bbox": [ + 176, + 321, + 284, + 428 + ], + "lines": [ + { + "bbox": [ + 176, + 321, + 284, + 428 + ], + "spans": [ + { + "bbox": [ + 176, + 321, + 284, + 428 + ], + "type": "image", + "image_path": "072ee8f433ac722281ac7aa476c465c1f75f70b5cebdf19a1f5e5f2a028acd94.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 184, + 430, + 276, + 441 + ], + "lines": [ + { + "bbox": [ + 184, + 430, + 276, + 441 + ], + "spans": [ + { + "bbox": [ + 184, + 430, + 276, + 441 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 184, + 430, + 276, + 441 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 184, + 430, + 276, + 441 + ], + "type": "text", + "content": " Lorentz Attention" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 450, + 296, + 483 + ], + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 483 + ], + "type": "text", + "content": "Figure 4. The UMAP [41] visualization displays the learned frame embeddings from a video in TVR. Data points of the same color correspond to the same moment." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": "As shown in Tab. 2, the worst performance occurs when only " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{sim}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " is used. Comparing Variant (5) with Variant (3), adding " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{div}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " increases the SumR, which can validate its necessity. Similarly, comparing Variant (4) with Variant (3) and Fig. 5, integrating " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " not only boosts retrieval accuracy but also ensures that the text query remains semantically embedded within the corresponding video, preserving partial relevance. Efficacy of Aggregation Strategy We compare three aggregation strategies: (i) " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " MP: mean pooling for static fusion. (ii) " + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "inline_equation", + "content": "w/" + }, + { + "bbox": [ + 55, + 508, + 296, + 715 + ], + "type": "text", + "content": " CL: feature concatenation with linear layers. (iii) MAIM (default): mean-guided adaptive interaction module. As shown in Tab. 2, MP performs the worst due to its fixed static fusion, which limits semantic interaction. CL improves upon MP by leveraging linear layers for dynamic feature fusion. MAIM achieves the best performance by learning adaptive aggregation weights and dynamically selecting hyperbolic information under global guidance." + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 317, + 77, + 453, + 177 + ], + "blocks": [ + { + "bbox": [ + 317, + 77, + 453, + 177 + ], + "lines": [ + { + "bbox": [ + 317, + 77, + 453, + 177 + ], + "spans": [ + { + "bbox": [ + 317, + 77, + 453, + 177 + ], + "type": "image", + "image_path": "3e9ad8a8d7b285e1fe59545c9fa44c58a47e20dca76550d2c430c2e6e90a0c4d.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 412, + 180, + 457, + 191 + ], + "lines": [ + { + "bbox": [ + 412, + 180, + 457, + 191 + ], + "spans": [ + { + "bbox": [ + 412, + 180, + 457, + 191 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 412, + 180, + 457, + 191 + ], + "type": "inline_equation", + "content": "w / oL_{pop}" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 457, + 70, + 553, + 176 + ], + "blocks": [ + { + "bbox": [ + 457, + 70, + 553, + 176 + ], + "lines": [ + { + "bbox": [ + 457, + 70, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 457, + 70, + 553, + 176 + ], + "type": "image", + "image_path": "a9d0c200c64728be273761a6d2756e584e898e07e438b6275623ba0445082acf.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 317, + 196, + 454, + 297 + ], + "blocks": [ + { + "bbox": [ + 317, + 196, + 454, + 297 + ], + "lines": [ + { + "bbox": [ + 317, + 196, + 454, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 196, + 454, + 297 + ], + "type": "image", + "image_path": "6f23f2ef61f0a26d8c71cda61fecc309ebd0e86103a620bc4b41b3ce04330d01.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 300, + 455, + 312 + ], + "lines": [ + { + "bbox": [ + 413, + 300, + 455, + 312 + ], + "spans": [ + { + "bbox": [ + 413, + 300, + 455, + 312 + ], + "type": "text", + "content": "(b) " + }, + { + "bbox": [ + 413, + 300, + 455, + 312 + ], + "type": "inline_equation", + "content": "w / L_{pop}" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 457, + 193, + 553, + 296 + ], + "blocks": [ + { + "bbox": [ + 457, + 193, + 553, + 296 + ], + "lines": [ + { + "bbox": [ + 457, + 193, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 457, + 193, + 553, + 296 + ], + "type": "image", + "image_path": "4f64357e5b574a934ff6fd896f9743c1fd144b39d82689d5cb707c6cc42abc31.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 342 + ], + "type": "text", + "content": "Figure 5. Visualization of the learned hyperbolic space. The closer to the origin, the higher semantic hierarchy and coarser granularity." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "text", + "content": "Visualization of Hyperbolic Space Inspired by HyCo-CLIP [46], we visualize the learned hyperbolic space by sampling 3K embeddings from the TVR training set. We analyze their norm distribution via histogram and reduce dimensionality using HoroPCA [3], as shown in Fig. 5. Glance branch embeddings are positioned closer to the origin than text query embeddings, indicating that clip-level video representations subsume textual queries. This phenomenon can be attributed to " + }, + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "text", + "content": ", which enforces the partial order relationship between video and text representations. In contrast, without " + }, + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "inline_equation", + "content": "L_{pop}" + }, + { + "bbox": [ + 313, + 365, + 556, + 534 + ], + "type": "text", + "content": ", embeddings exhibit uncorrelated distributions. Moreover, text queries, being coarser in semantics, lie closer to the origin than fine-grained gaze-level embeddings, reflecting a clear hierarchical structure." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 391, + 560 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 713 + ], + "type": "text", + "content": "In this paper, we propose HLFormaler, a novel hyperbolic modeling framework tailored for PRVR. By leveraging the intrinsic geometric properties of hyperbolic space, HLFormaler effectively captures the hierarchical and multi-granular structure of untrimmed videos, thereby enhancing video-text retrieval accuracy. Furthermore, to ensure partial relevance between paired videos and text, a partial order preservation loss is introduced to enforce their semantic entailment. Extensive experiments indicate that HLFormaler consistently outperforms state-of-the-art methods. Our study offers a new perspective for PRVR with hyperbolic learning, which we hope will inspire further research in this direction." + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23081" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "Acknowledgments. We sincerely thank the anonymous reviewers and chairs for their efforts and constructive suggestions, which have greatly helped us improve the manuscript. This work is supported in part by the National Natural Science Foundation of China under grant 624B2088, 62171248, 62301189, the PCNL KEY project (PCL2023AS6-1), and Shenzhen Science and Technology Program under Grant KJZD20240903103702004, JCYJ20220818101012025, GXWD20220811172936001. Long Chen was supported by the Hong Kong SAR RGC Early Career Scheme (26208924), the National Natural Science Foundation of China Young Scholar Fund (62402408), Huawei Gift Fund, and the HKUST Sports Science and Technology Research Grant (SSTRG24EG04)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 248, + 115, + 260 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 248, + 115, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 248, + 115, + 260 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 268, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 268, + 297, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 268, + 297, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 268, + 297, + 322 + ], + "type": "text", + "content": "[1] Mina Ghadimi Atigh, Julian Schoep, Erman Acar, Nanne Van Noord, and Pascal Mettes. Hyperbolic image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4453-4462, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 323, + 296, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 323, + 296, + 367 + ], + "spans": [ + { + "bbox": [ + 61, + 323, + 296, + 367 + ], + "type": "text", + "content": "[2] Joao Carreira and Andrew Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 6299-6308, 2017. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 369, + 296, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 296, + 400 + ], + "type": "text", + "content": "[3] Ines Chami, Albert Gu, Dat Nguyen, and Christopher Ré. Horopca: Hyperbolic dimensionality reduction via horospherical projections, 2021. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 402, + 296, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 296, + 444 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 296, + 444 + ], + "type": "text", + "content": "[4] Bike Chen, Wei Peng, Xiaofeng Cao, and Juha Röning. Hyperbolic uncertainty aware semantic segmentation. IEEE Transactions on Intelligent Transportation Systems, 25(2): 1275-1290, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 447, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 296, + 499 + ], + "type": "text", + "content": "[5] Shizhe Chen, Yida Zhao, Qin Jin, and Qi Wu. Fine-grained video-text retrieval with hierarchical graph reasoning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10638-10647, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 502, + 296, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 534 + ], + "type": "text", + "content": "[6] Weize Chen, Xu Han, Yankai Lin, Hexu Zhao, Zhiyuan Liu, Peng Li, Maosong Sun, and Jie Zhou. Fully hyperbolic neural networks. arXiv preprint arXiv:2105.14686, 2021. 3, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 296, + 590 + ], + "type": "text", + "content": "[7] Zhiguo Chen, Xun Jiang, Xing Xu, Zuo Cao, Yijun Mo, and Heng Tao Shen. Joint searching and grounding: Multi-granularity video content retrieval. In Proceedings of the 31st ACM International Conference on Multimedia, pages 975–983, 2023. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 591, + 296, + 635 + ], + "type": "text", + "content": "[8] Dingxin Cheng, Shuhan Kong, Bin Jiang, and Qiang Guo. Transferable dual multi-granularity semantic excavating for partially relevant video retrieval. Image and Vision Computing, 149:105168, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 296, + 689 + ], + "type": "text", + "content": "[9] Cheol-Ho Cho, WonJun Moon, Woojin Jun, MinSeok Jung, and Jae-Pil Heo. Ambiguity-restrained text-video representation learning for partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2500-2508, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[10] Karan Desai, Maximilian Nickel, Tanmay Rajpurohit, Justin Johnson, and Ramakrishna Vedantam. Hyperbolic Image-" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "Text Representations. In Proceedings of the International Conference on Machine Learning, 2023. 2, 3, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 97, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 139 + ], + "type": "text", + "content": "[11] Jianfeng Dong, Xirong Li, and Cees GM Snoek. Predicting visual features from text for image and video caption retrieval. IEEE Transactions on Multimedia, 20(12):3377-3388, 2018. 1, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 555, + 195 + ], + "type": "text", + "content": "[12] Jianfeng Dong, Xirong Li, Chaoxi Xu, Shouling Ji, Yuan He, Gang Yang, and Xun Wang. Dual encoding for zeroexample video retrieval. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9346-9355, 2019." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 554, + 241 + ], + "type": "text", + "content": "[13] Jianfeng Dong, Xirong Li, Chaoxi Xu, Xun Yang, Gang Yang, Xun Wang, and Meng Wang. Dual encoding for video retrieval by text. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8):4065-4080, 2021. 1, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 243, + 555, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 555, + 295 + ], + "type": "text", + "content": "[14] Jianfeng Dong, Xianke Chen, Minsong Zhang, Xun Yang, Shujie Chen, Xirong Li, and Xun Wang. Partially relevant video retrieval. In Proceedings of the 30th ACM International Conference on Multimedia, pages 246-257, 2022. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 555, + 353 + ], + "type": "text", + "content": "[15] Jianfeng Dong, Yabing Wang, Xianke Chen, Xiaoye Qu, Xirong Li, Yuan He, and Xun Wang. Reading-strategy inspired visual representation learning for text-to-video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5680-5694, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 354, + 555, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 354, + 555, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 354, + 555, + 419 + ], + "type": "text", + "content": "[16] Jianfeng Dong, Minsong Zhang, Zheng Zhang, Xianke Chen, Daizong Liu, Xiaoye Qu, Xun Wang, and Baolong Liu. Dual learning with dynamic knowledge distillation for partially relevant video retrieval. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11302-11312, 2023. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 421, + 555, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 421, + 555, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 421, + 555, + 475 + ], + "type": "text", + "content": "[17] Aleksandr Ermolov, Leyla Mirvakhabova, Valentin Khrulkov, Nicu Sebe, and Ivan Oseledets. Hyperbolic vision transformers: Combining improvements in metric learning. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7399-7409, 2022. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 477, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 477, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 477, + 554, + 510 + ], + "type": "text", + "content": "[18] Fartash Faghri, David J Fleet, Jamie Ryan Kiros, and Sanja Fidler. Vse++: Improving visual-semantic embeddings with hard negatives. arXiv preprint arXiv:1707.05612, 2017. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 555, + 566 + ], + "type": "text", + "content": "[19] Hao Fang, Changle Zhou, Jiawei Kong, Kuofeng Gao, Bin Chen, Tao Liang, Guojun Ma, and Shu-Tao Xia. Grounding language with vision: A conditional mutual information calibrated decoding strategy for reducing hallucinations in lvlms. arXiv preprint arXiv:2505.19678, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 567, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 555, + 610 + ], + "type": "text", + "content": "[20] Sheng Fang, Tiantian Dang, Shuhui Wang, and Qingming Huang. Linguistic hallucination for text-based video retrieval. IEEE Transactions on Circuits and Systems for Video Technology, 34(10):9692-9705, 2024. 6, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 555, + 645 + ], + "type": "text", + "content": "[21] Octavian Ganea, Gary Bécigneul, and Thomas Hofmann. Hyperbolic neural networks. Advances in neural information processing systems, 31, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 555, + 689 + ], + "type": "text", + "content": "[22] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In Proceedings of the IEEE international conference on computer vision, pages 5267-5275, 2017. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 555, + 713 + ], + "type": "text", + "content": "[23] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed-" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "23082" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "ings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 57, + 95, + 295, + 116 + ], + "type": "text", + "content": "[24] Neil He, Menglin Yang, and Rex Ying. Lorentzian residual neural networks. arXiv preprint arXiv:2412.14695, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 161 + ], + "type": "text", + "content": "[25] Zhijian Hou, Chong-Wah Ngo, and Wing Kwong Chan. Conquer: Contextual query-aware ranking for video corpus moment retrieval. In Proceedings of the 29th ACM International Conference on Multimedia, pages 3900-3908, 2021. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 162, + 296, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 162, + 296, + 215 + ], + "spans": [ + { + "bbox": [ + 57, + 162, + 296, + 215 + ], + "type": "text", + "content": "[26] Xun Jiang, Zhiguo Chen, Xing Xu, Fumin Shen, Zuo Cao, and Xunliang Cai. Progressive event alignment network for partial relevant video retrieval. In 2023 IEEE International Conference on Multimedia and Expo (ICME), pages 1973-1978. IEEE, 2023. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 295, + 259 + ], + "type": "text", + "content": "[27] Valentin Khrulkov, Leyla Mirvakhabova, Evgeniya Ustinova, Ivan Oseledets, and Victor Lempitsky. Hyperbolic image embeddings. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 304 + ], + "type": "text", + "content": "[28] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense-captioning events in videos. In Proceedings of the IEEE international conference on computer vision, pages 706–715, 2017. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 296, + 348 + ], + "type": "text", + "content": "[29] Marc Law, Renjie Liao, Jake Snell, and Richard Zemel. Lorentzian distance learning for hyperbolic representations. In Proceedings of the 36th International Conference on Machine Learning, pages 3672-3681. PMLR, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 403 + ], + "type": "text", + "content": "[30] Jie Lei, Licheng Yu, Tamara L Berg, and Mohit Bansal. Tvr: A large-scale dataset for video-subtitle moment retrieval. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXI 16, pages 447-463. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[31] Jiaxu Leng, Zhanjie Wu, Mingpi Tan, Yiran Liu, Ji Gan, Haosheng Chen, and Xinbo Gao. Beyond euclidean: Dualspace representation learning for weakly supervised video violence detection. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 491 + ], + "type": "text", + "content": "[32] Keegan Lensink, Bas Peters, and Eldad Haber. Fully hyperbolic convolutional neural networks. Research in the Mathematical Sciences, 9(4):60, 2022. 3, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "type": "text", + "content": "[33] Huimin Li, Zhentao Chen, Yunhao Xu, and Junlin Hu. Hyperbolic anomaly detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17511-17520, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "[34] Xirong Li, Chaoxi Xu, Gang Yang, Zhineng Chen, and Jianfeng Dong. W2vv++ fully deep learning for ad-hoc video search. In Proceedings of the 27th ACM international conference on multimedia, pages 1786-1794, 2019. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 635 + ], + "type": "text", + "content": "[35] Haitong Liu, Kuofeng Gao, Yang Bai, Jinmin Li, Jinxiao Shan, Tao Dai, and Shu-Tao Xia. Protecting your video content: Disrupting automated video-based lvm annotations. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 24056-24065, 2025. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 296, + 689 + ], + "type": "text", + "content": "[36] Peidong Liu, Dongliang Liao, Jinpeng Wang, Yangxin Wu, Gongfu Li, Shu-Tao Xia, and Jin Xu. Multi-task ranking with user behaviors for text-video search. In Companion Proceedings of the Web Conference 2022, pages 126-130, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[37] Yang Liu, Samuel Albanie, Arsha Nagrani, and Andrew Zisserman. Use what you have: Video retrieval using" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "representations from collaborative experts. arXiv preprint arXiv:1907.13487, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 148 + ], + "type": "text", + "content": "[38] Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach. arXiv preprint arXiv:1907.11692, 2019. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 555, + 194 + ], + "type": "text", + "content": "[39] Teng Long, Pascal Mettes, Heng Tao Shen, and Cees GM Snoek. Searching for actions on the hyperbole. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1141-1150, 2020. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 555, + 238 + ], + "type": "text", + "content": "[40] Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. Clip4clip: An empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing, 508:293-304, 2022. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 239, + 555, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 555, + 272 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 555, + 272 + ], + "type": "text", + "content": "[41] Leland McInnes, John Healy, and James Melville. Umap: Uniform manifold approximation and projection for dimension reduction. arXiv preprint arXiv:1802.03426, 2018. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 273, + 555, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 273, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 555, + 336 + ], + "type": "text", + "content": "[42] Antoine Miech, Dimitri Zhukov, Jean-Baptiste Alayrac, Makarand Tapaswi, Ivan Laptev, and Josef Sivic. Howto100m: Learning a text-video embedding by watching hundred million narrated video clips. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2630-2640, 2019. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 555, + 381 + ], + "type": "text", + "content": "[43] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10810-10819, 2020. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 554, + 425 + ], + "type": "text", + "content": "[44] Maximillian Nickel and Douwe Kiela. Poincaré embeddings for learning hierarchical representations. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 554, + 470 + ], + "type": "text", + "content": "[45] Maximillian Nickel and Douwe Kiela. Learning continuous hierarchies in the Lorentz model of hyperbolic geometry. In Proceedings of the 35th International Conference on Machine Learning, pages 3779-3788. PMLR, 2018. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 555, + 525 + ], + "type": "text", + "content": "[46] Avik Pal, Max van Spengler, Guido Maria D'Amely di Melendugno, Alessandro Flaborea, Fabio Galasso, and Pascal Mettes. Compositional entailment learning for hyperbolic vision-language models. In The Thirteenth International Conference on Learning Representations, 2025. 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 525, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 554, + 558 + ], + "type": "text", + "content": "[47] Xiaogang Peng, Hao Wen, Yikai Luo, Xiao Zhou, Keyang Yu, Yigang Wang, and Zizhao Wu. Learning weakly supervised audio-visual violence detection in hyperbolic space, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 623 + ], + "type": "text", + "content": "[48] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "type": "text", + "content": "[49] Ruiqi Shi, Jun Wen, Wei Ji, Menglin Yang, Difei Gao, and Roger Zimmermann. HOVER: Hyperbolic video-text retrieval, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 689 + ], + "type": "text", + "content": "[50] Xue Song, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Spatial-temporal graphs for cross-modal text2video retrieval. IEEE Transactions on Multimedia, 24:2914-2923, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 554, + 713 + ], + "type": "text", + "content": "[51] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "23083" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "weakly-supervised video paragraph grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13569-13580, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[52] Max Van Spengler, Erwin Berkhout, and Pascal Mettes. Poincaré resnet. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5419-5428, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 196 + ], + "type": "text", + "content": "[53] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 296, + 251 + ], + "type": "text", + "content": "[54] Jinpeng Wang, Bin Chen, Dongliang Liao, Ziyun Zeng, Gongfu Li, Shu-Tao Xia, and Jin Xu. Hybrid contrastive quantization for efficient cross-view video retrieval. In Proceedings of the ACM Web Conference 2022, pages 3020-3030, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 296, + 308 + ], + "type": "text", + "content": "[55] Jinpeng Wang, Ziyun Zeng, Bin Chen, Yuting Wang, Dongliang Liao, Gongfu Li, Yiru Wang, and Shu-Tao Xia. Hugs bring double benefits: Unsupervised cross-modal hashing with multi-granularity aligned transformers. International Journal of Computer Vision, 132(8):2765-2797, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 310, + 296, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 296, + 342 + ], + "type": "text", + "content": "[56] Yuting Wang, Jinpeng Wang, Bin Chen, Tao Dai, Ruisheng Luo, and Shu-Tao Xia. Gmmformer v2: An uncertainty-aware framework for partially relevant video retrieval, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 296, + 398 + ], + "type": "text", + "content": "[57] Yuting Wang, Jinpeng Wang, Bin Chen, Ziyun Zeng, and Shu-Tao Xia. Gmmformer: Gaussian-mixture-model based transformer for efficient partially relevant video retrieval. In Proceedings of the AAAI Conference on Artificial Intelligence, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 453 + ], + "type": "text", + "content": "[58] Zhihao Wang, Wenke Huang, Tian Chen, Zekun Shi, Guancheng Wan, Yu Qiao, Bin Yang, Jian Wang, Bing Li, and Mang Ye. An empirical study of federated prompt learning for vision language model. arXiv preprint arXiv:2505.23024, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 296, + 511 + ], + "type": "text", + "content": "[59] Wenhao Wu, Haipeng Luo, Bo Fang, Jingdong Wang, and Wanli Ouyang. Cap4video: What can auxiliary captions do for text-video retrieval? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10704-10713, 2023. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "text", + "content": "[60] Shukang Yin, Sirui Zhao, Hao Wang, Tong Xu, and Enhong Chen. Exploiting instance-level relationships in weakly supervised text-to-video retrieval. ACM Trans. Multim. Comput. Commun. Appl., 20(10):316:1-316:21, 2024. 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 296, + 624 + ], + "type": "text", + "content": "[61] Zhen Yu, Toan Nguyen, Yaniv Gal, Lie Ju, Shekhar S Chandra, Lei Zhang, Paul Bonnington, Victoria Mar, Zhiyong Wang, and Zongyuan Ge. Skin lesion recognition with class-hierarchy regularized hyperbolic embeddings. In International conference on medical image computing and computer-assisted intervention, pages 594-603. Springer, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 668 + ], + "type": "text", + "content": "[62] Bowen Zhang, Hexiang Hu, Joonseok Lee, Ming Zhao, Sheide Chammas, Vihan Jain, Eugene Ie, and Fei Sha. A hierarchical multi-modal encoder for moment localization in video corpus. arXiv preprint arXiv:2011.09046, 2020. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[63] Hao Zhang, Aixin Sun, Wei Jing, Guoshun Nan, Liangli Zhen, Joey Tianyi Zhou, and Rick Siow Mong Goh. Video corpus moment retrieval with contrastive learning. In Proceedings of the 44th International ACM SIGIR Conference on Research" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 334, + 72, + 555, + 95 + ], + "type": "text", + "content": "and Development in Information Retrieval, pages 685-695, 2021. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "23084" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_content_list.json b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..edd2ff4e4ca3f102fcc2cbae91d02ad1a98d056e --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_content_list.json @@ -0,0 +1,2441 @@ +[ + { + "type": "text", + "text": "Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection", + "text_level": 1, + "bbox": [ + 120, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xuehan Chen $^{1}$ Guangyu Ren $^{1*}$ Tianhong Dai $^{2}$ Tania Stathaki $^{2}$ Hengyan Liu $^{1*}$", + "bbox": [ + 158, + 202, + 844, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Xi'an Jiaotong-Liverpool University, China", + "bbox": [ + 323, + 220, + 674, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Imperial College London, United Kingdom", + "bbox": [ + 323, + 238, + 673, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xuehanchen83@gmail.com, {guangyu. ren, hengyan.liu}@xjtlu.edu.cn", + "bbox": [ + 218, + 258, + 777, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tianhong.dai15@alumni.imperial.ac.uk, t.stathaki@imperial.ac.uk", + "bbox": [ + 217, + 277, + 774, + 291 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundation models, such as Segment Anything Model (SAM), have exhibited remarkable performance in conventional segmentation tasks, primarily due to their training on large-scale datasets. Nonetheless, challenges remain in specific downstream tasks, such as Camouflaged Object Detection (COD). Existing research primarily aims to enhance performance by integrating additional multimodal information derived from other foundation models. However, directly leveraging the information generated by these models may introduce additional biases due to domain shifts. To address this issue, we propose an Adaptive Refinement Module (ARM), which efficiently processes multimodal information and simultaneously refining the mask prompt. Furthermore, we construct an auxiliary embedding that effectively exploits the intermediate information generated during ARM, providing SAM with richer feature representations. Experimental results indicate that our proposed architecture surpasses most state-of-the-art (SOTA) models in the COD task, particularly excelling in structured target segmentation.", + "bbox": [ + 88, + 358, + 483, + 661 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 693, + 220, + 709 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Camouflaged Object Detection (COD) is a challenging computer vision task, as it involves segmenting objects that are virtually identical to their immediate surroundings. It plays a crucial role in the field due to its wide range of practical applications, including object detection [34, 45], low-altitude economy [22], zoological research [31, 35, 37], and medical image segmentation [18, 52].", + "bbox": [ + 89, + 720, + 482, + 825 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current research [12, 14, 33, 41, 43] focuses on designing advanced architectures to extract and refine more detailed features exclusively from image data. However, the", + "bbox": [ + 89, + 827, + 483, + 873 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/50d4f60705a2b5a90993390a56097d82f8c3bd792a4f80336e65cd635d0f2f55.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 324, + 901, + 491 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/adc2cb2ca55e642b64beec235145ca67f12d5ecd5a73506f5d0ac261735ccba6.jpg", + "image_caption": [ + "(c) Ours", + "Figure 1. Comparison of our proposed Adaptive Refinement Module (ARM) with common COD frameworks. Unlike conventional methods that rely solely on images or manually annotated additional modalities, our approach not only leverages vision-language models (VLMs) to automatically generate multimodal data but also mitigates domain shift through the proposed ARM, providing high-quality prompts for SAM's decoder. Meanwhile, ARM can also fine-tune SAM's image encoder, further improving SAM's feature representation capability." + ], + "image_footnote": [], + "bbox": [ + 516, + 496, + 898, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "unique nature of the COD task, where the targets are visually similar to their surroundings, limits the effectiveness of image features. Additionally, the lack of data from other modalities further limits the models' ability to acquire meaningful features, making them highly susceptible to background, preventing them from accurately detecting camouflaged objects. Although some studies incorporate human-made auxiliary data, such as depth images [53], textual descriptions [61], or prompts [2, 3], to provide supple", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20672", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mentary information about targets, they are not ideal solutions as they are time-consuming and labor-intensive. As a result, despite significant progress in the COD task, two limitations remain: limited image features and laborious annotation processes.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, foundation models have emerged as a popular solution for tackling a range of tasks, including the Segment Anything Model (SAM) [27], which has proven highly effective in handling various tasks, such as medical image segmentation [44, 47] and semantic segmentation [4, 62]. Although SAM demonstrates impressive zero-shot generalization capabilities, Chen et al. [5] states that it still struggles to segment the complex camouflaged objects. Secondly, SAM can guide its segmentation by receiving prompts, and the most intuitive approach is to generate prompts through manual annotation, but as mentioned earlier, it is less practical due to the time and labor it demands.", + "bbox": [ + 88, + 169, + 480, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To mitigate the aforementioned limitations, we rethink the core issues that make the COD task challenging. In this task, there are no task-specific clues (e.g., category, size, location) of the targets. Meanwhile, the background's visual noise increases the task's complexity, resulting in ambiguous object boundaries and weakened feature learning. So, we propose a novel approach that uses textual information to complement visual features, enhancing the model's ability to localize target objects more accurately. In this work, we fully utilize the powerful pretrained vision-language models (VLMs) to automatically generate multimodal data without requiring additional training or fine-tuning. This data provides additional cues to assist in object localization. Subsequently, the VLMs further utilize the generated multimodal information to produce mask prompts that convey the object's approximate location to SAM.", + "bbox": [ + 89, + 352, + 482, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We also consider the biases introduced by directly utilizing generated multimodal information, which could lead to locality in the prompts, making them less comprehensive and potentially affecting the subsequent SAM segmentation performance. To mitigate this issue, this paper proposes an Adaptive Refinement Module (ARM), which incorporates simple adapters to refine the initially generated mask prompts. This module expands the regions in the initial prompts to create a more complete representation of the target objects, offering better guidance for downstream models. Furthermore, ARM effectively extracts vision information from the entire architecture, serving not only as a refinement mechanism for prompts but also as a fine-tuning module for SAM's image encoder, further enhancing SAM's performance.", + "bbox": [ + 89, + 595, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, this work proposes the ARM, which simultaneously refines the initial prompt generated by multimodal information and fine-tunes SAM's image encoder, enabling SAM to handle the COD task more effectively across multiple aspects. Our main contributions are as fol", + "bbox": [ + 89, + 825, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "lows:", + "bbox": [ + 513, + 92, + 552, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a cost-effective approach that automatically constructs mask prompts using multimodal information generated by BLIP, serving as additional cues without requiring extra training or fine-tuning. These prompts provide reliable guidance for SAM, thereby improving its segmentation performance.", + "- To mitigate the additional biases introduced by BLIP, we propose an Adaptive Refinement Module (ARM) to refine the initially generated prompts, enhancing their quality and providing more precise guidance for SAM. Meanwhile, this module can also fine-tune the SAM's image encoder, further enhancing SAM's segmentation accuracy and robustness.", + "- We further utilize the multimodal information generated by BLIP to construct an auxiliary embedding, which effectively exploits the intermediate information generated during the prompt generation process and provides SAM with richer features." + ], + "bbox": [ + 511, + 109, + 903, + 381 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 398, + 653, + 414 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single-Modality Based COD", + "text_level": 1, + "bbox": [ + 513, + 422, + 767, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image data is one of the most widely used modalities to tackle the COD task. Therefore, most researchers are currently exploring various approaches to extract information from single-modality image data in order to perform this task more effectively. CoFiNet [15], ZoomNext [43], ZoomNet [42] focus on extracting and fusing multiscale features from images. Although upgrading the model's backbone improves the accuracy of identifying objects, it also increases the consumption of computational resources.", + "bbox": [ + 511, + 446, + 903, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Besides directly extracting features from the images, CamoDiffusion [7], BiDiCOS [26], and diffCOD [6] transform the COD task into denoising tasks, gradually segmenting camouflaged objects through the denoising process. Moreover, EFNet [14], FSEL [50], FGSA-Net [63], and MAM-FFC [55] use spatial and frequency information from images to achieve a more precise segmentation. However, in the COD task, the background in images introduces significant noise. Even with those novel methods relying solely on images' intrinsic features, they still struggle to recognize objects in complex scenarios.", + "bbox": [ + 511, + 583, + 903, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, in addition to designing a new architecture, many works have also focused on improving the Segment Anything Model (SAM), taking advantage of its powerful zero-shot generalization ability and prompt strategy. SAM-COD [3] and WS-SAM [17], which combine sparse prompts with SAM to generate the final segmentation results, still require additional manual annotations for the sparse prompts, which remains time-consuming. MCA-SAM [65] and SAM-adapter [5] improve the model's performance in the COD task by injecting more image infor", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20673", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/94c915325f32d32b65f2b0f77d5581f8541f6026b44182b8af77d6fa67eca345.jpg", + "image_caption": [ + "Figure 2. Overview of our proposed architecture. BLIP automatically generates multimodal data from the input image and uses it to create the initial multimodal prompts and the auxiliary embedding. The Adaptive Refinement Module (ARM) is then employed to address the locality issue of the initial prompts. The ARM uses multiple adapters to extract visual information from both BLIP and SAM, which not only refines the initial multimodal prompts by integrating attention weights, resulting in more accurate mask prompts, but also fine-tunes SAM's image encoder. Finally, the SAM decoder receives image embedding, dense mask prompt embedding, and auxiliary sparse embedding to tackle the COD task." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 903, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "mation into SAM's image encoder using the adapter. They differ from our work, where the adapter serves a broader purpose in ARM. It is primarily used to refine the generated mask prompts while injecting additional information into SAM's image encoder, enabling the SAM to perform the COD task more effectively.", + "bbox": [ + 88, + 512, + 480, + 604 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Multi-Modality Based COD", + "text_level": 1, + "bbox": [ + 89, + 619, + 341, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the rapid growth of multimodal learning, different modalities have been increasingly applied to address the COD problem. HPGT [40] use the textual information to assist in completing the COD task. However, they rely on manually annotating text data for the COD dataset, which is a highly time-consuming process. Although ACU-MEN [61] utilizes the GPT-4 Vision model to generate textual descriptions and attributions for images, it still requires more than 30 volunteers to manually verify the accuracy of these generated texts, which is also a labor-intensive task. In contrast to these approaches, MMCPF [51] takes a different direction by interacting with a Multimodal Large Language Model (MLLM). They query the model using text to obtain the location coordinates of camouflaged objects, which supports the subsequent segmentation task. While this method reduces the need for text data annotation, it still faces the challenge of designing precise text prompts for querying the", + "bbox": [ + 89, + 643, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "MLLM. Ordinary prompts cannot let MLLM output the expected location coordinates of objects, which indicates that designing effective text prompts becomes an additional task that is time-consuming and labor-intensive as well. As a result, both visual and textual information have been employed to address the COD problem, but manual annotation and verification remain a limitation.", + "bbox": [ + 511, + 512, + 906, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In addition to textual information, depth features are utilized to provide valuable spatial information for models. Due to the fact that commonly used COD datasets only include RGB images without depth data, PopNet [56], Da-COD [54] and [57] employ depth estimation models to generate depth images from RGB-based COD datasets. Since the depth images are generated automatically by models, the final segmentation performance is highly reliant on the performance of those depth estimation methods. In contrast to the previous approaches, RISNet [53] uses a depth camera to manually create an RGB-D dataset for COD in agricultural scenes, which is also painstaking.", + "bbox": [ + 511, + 619, + 908, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 814, + 604, + 829 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a novel architecture, as illustrated in Fig. 2, which enables the fully automated generation of mask prompts for SAM and designs an innovative Adaptive Refinement Module (ARM) to enhance the quality of the", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20674", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "initially generated mask prompts. Specifically, a Vision-Language Model (VLM) first automatically generates textual data based on the input image. Then, by aligning multimodal information, the VLM generates the multimodal prompts to locate camouflaged objects locally. Subsequently, it is fed into the ARM for refinement, enabling prompts to capture more complete object structure information, thereby providing better guidance for SAM. During the prompt refinement process, ARM also fine-tunes SAM's image encoder. Meanwhile, the generated textual data is fully utilized, which is integrated with visual information to provide additional auxiliary embedding, further improving SAM's performance in identifying camouflaged objects.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the following subsections, we provide detailed explanations of each component of this architecture. First, we describe the multimodal prompt generation process in Sec. 3.1. Then we introduce the ARM in Sec. 3.2. Finally, Sec. 3.3 describes the different input embeddings fed into SAM's mask decoder to generate the final results.", + "bbox": [ + 89, + 287, + 483, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Multimodal Prompt Generation", + "text_level": 1, + "bbox": [ + 89, + 390, + 372, + 406 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.1. Multimodal Data Generation", + "text_level": 1, + "bbox": [ + 89, + 412, + 341, + 426 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As a robust zero-shot VLM, BILP [30] offers powerful image captioning capabilities that align with our goal of automatically generating supplementary textual data for subsequent prompt generation. Given an image $I \\in \\mathbb{R}^{H \\times W \\times 3}$ , the image embedding $Emb_{B_{img}}$ are extracted by the BLIP's image encoder $E_{B_{img}}$ . This embedding is then passed to BLIP's image-grounded text decoder, which generates a caption, as presented below:", + "bbox": [ + 89, + 431, + 483, + 553 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT = \\mathbf {D} _ {B _ {t e x t}} \\left(E m b _ {B _ {i m g}}\\right) = \\left\\{t _ {i j} \\right\\} _ {i = 1} ^ {L}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 571, + 482, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $T$ denotes the generated caption, $\\mathrm{D}_{B_{text}}$ is the BLIP's image-grounded text decoder, and $t_i$ represents the $i^{th}$ word in the caption, with a total of $L$ words. As exemplified by the potential caption generated in Fig. 2, it is clear that the captions generated by BLIP can correctly identify the camouflaged targets.", + "bbox": [ + 89, + 597, + 483, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, since the generated $T$ provides an overall description of the input image, it may include content that is irrelevant to the target objects, such as \"floor\" is the background information of the description \"There is a small dog that is lying on the floor\". Due to the complexity of the COD task, irrelevant information in the generated caption could act as noise, potentially distracting from generating a prompt that accurately indicates the target's location. Therefore, instead of utilizing the entire caption $T$ , we use the spaCy [21] library for part-of-speech tagging to capture the necessary content about the objects. Specifically, each $t_i$ in $T$ is tagged with its corresponding part of speech (e.g., noun, verb, adjective). Through observation, we found that most generated captions primarily use nouns to describe the", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "category of targets (e.g., man, bug, lion). Therefore, we extract the first noun $t_{noun}$ that appears in $T$ to provide more precise target category information, which helps the subsequent prompt generation more accurately locate the target object and reduces distractions caused by unrelated noise.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.2. Multimodal Prompt Generation", + "text_level": 1, + "bbox": [ + 511, + 174, + 782, + 189 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "BLIP has a powerful text-image alignment capability, which allows it to generate a multimodal prompt that can accurately locate the target object by aligning the text containing the target category information with the image. Inspired by [29], we use BLIP once again to handle the image-text retrieval task and generate the Gradient-weighted Class Activation Mapping (Grad-CAM) [46] to visualize the alignment process of BLIP with multimodal data, thereby generating the prompt.", + "bbox": [ + 511, + 193, + 905, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the $Emb_{B_{img}}$ and $t_{noun}$ are fed into BLIP's image-grounded text encoder $\\mathrm{E}_{B_{text}}$ , which then outputs a score reflecting how well the input image and text match, as follows:", + "bbox": [ + 511, + 330, + 905, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n(S, \\mathrm {A} _ {\\text {c r o s s}}) = \\mathrm {E} _ {B _ {\\text {t e x t}}} \\left(\\left(E m b _ {B i m g}, t _ {\\text {n o u n}}\\right), \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 406, + 903, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $S$ denotes the output score, and $A_{cross}$ is the attention weight from the cross attention layer of $\\mathbf{E}_{B_{text}}$ , which captures the relationship between the image and the text.", + "bbox": [ + 511, + 430, + 905, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate the Grad-CAM, the gradient is also important, so we conduct a backpropagation process to get the gradient $G_{cross}$ , and then generate the prompt, as presented below:", + "bbox": [ + 511, + 476, + 905, + 534 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} _ {\\text {i n i t}} = \\operatorname {G r a d C A M} = \\operatorname {M e a n} _ {\\text {t w i c e}} \\left(\\mathrm {G} _ {\\text {c r o s s}} \\odot \\mathrm {A} _ {\\text {c r o s s}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 549, + 903, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{P}_{init}$ is the initially generated multimodal prompt, $\\odot$ denotes the element-wise multiplication operation, and $\\mathrm{Mean}_{\\text{twice}}(\\cdot)$ performs two averaging steps, which average the influence of different text sequences on the image under each attention head, thereby obtaining the final $\\mathrm{P}_{init}$ .", + "bbox": [ + 511, + 577, + 905, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Adaptive Refinement Module", + "text_level": 1, + "bbox": [ + 511, + 662, + 774, + 679 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e0f946b2ef87161dd20044448373e14a5f38b4373779f4b11028f6af8180e855.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 516, + 700, + 599, + 782 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4c7666dc81fcc9eab190ffb92d9bad235a29da0abbfa4eba51da7dc2e55f8902.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 616, + 700, + 699, + 782 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/55c1714e7040f27fb6a07f99bd6bb57cbeefaf7474932e0ed15692599e1ac8d7.jpg", + "image_caption": [ + "Initial", + "Figure 3. Comparison between the initially generated prompt and the prompt refined using ARM." + ], + "image_footnote": [], + "bbox": [ + 715, + 700, + 800, + 782 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f6d202df113692182ce9599fac3445088ccb596b4c98756511065ecf06923f11.jpg", + "image_caption": [ + "Refined by ARM" + ], + "image_footnote": [], + "bbox": [ + 818, + 700, + 901, + 782 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 3, the initially generated multimodal prompt by BLIP only highlights the local regions of the camouflaged object rather than the entire object. As a", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20675", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "result, the mask prompt generated based on this multimodal prompt is incomplete, potentially negatively impacting SAM's performance. This locality issue is caused by domain shifts. When BLIP performs an image-text retrieval task, recognizing only the local region is sufficient to output the final matching score. Particularly in our work, the input text $t_{noun}$ consists of only a single noun representing the object category without any additional descriptive details. Therefore, the initially generated prompt can only roughly indicate the object's location but cannot fully capture its structure, resulting in a lower-quality mask prompt. To address this issue, we design an Adaptive Refinement Module (ARM) to refine the initial incomplete prompt, enabling it to cover the object more entirely and generate a higher-quality mask prompt.", + "bbox": [ + 89, + 90, + 483, + 318 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Refinement", + "text_level": 1, + "bbox": [ + 89, + 324, + 217, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first design adapters to extract features from the entire architecture. These adapters contain only two learnable layers, as shown below:", + "bbox": [ + 89, + 342, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {A d a p t e r} (x) = \\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {1} (x)\\right)\\right)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 396, + 482, + 414 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Adapter}(\\cdot)$ is an adapter module, $\\mathrm{Conv}_i(\\cdot)$ is the $i^{th}$ two-dimensional convolutional layer and $\\mathrm{GELU}(\\cdot)$ , the Gaussian Error Linear Unit [20], is an activation function.", + "bbox": [ + 89, + 421, + 483, + 467 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In ARM, there are $k$ adapters used to capture image features from the two distinct image encoders and fuse them to generate a more comprehensive representation:", + "bbox": [ + 89, + 468, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {A} ^ {k} = \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {B _ {i m g}}\\right) + \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {S _ {i m g}} ^ {k}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 518, + 482, + 550 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $Emb_A^k$ represents the output embedding of the $k^{th}$ adapter, and $Emb_{S_{img}}^k$ is the image embedding comes from $k^{th}$ transformer layer of SAM's image encoder. All adapters share the same weights.", + "bbox": [ + 89, + 551, + 483, + 614 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To better utilize the extracted image features, the output of the $k^{th}$ adapter is combined with that of the $k^{th}$ transformer layer in SAM's image encoder and fed into the $(k + 1)^{th}$ transformer layer, which could help fine-tune the SAM's image encoder:", + "bbox": [ + 89, + 614, + 483, + 691 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {S _ {i m g - i n}} ^ {k + 1} = E m b _ {A} ^ {k} + E m b _ {S _ {i m g - o u t}} ^ {k}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 699, + 482, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $Emb_{S_{img - in}}^{k + 1}$ denotes to the inputs of the $(k + 1)^{th}$ transformer layer in SAM's image encoder, and $Emb_{S_{img - out}}^k$ is the original output from $k^{th}$ transformer layer in SAM's image encoder.", + "bbox": [ + 89, + 727, + 483, + 795 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then, inspired by CLIP-ES [32], we extract the attention weight $A_{self}$ from $\\mathrm{E}_{B_{img}}$ instead of training an additional affinity network to refine the initial prompt. By leveraging the relationships between image patches encoded in the attention weights, the prompt can adjust itself to identify high-correlation regions and expand to cover a more complete representation of the object. However, considering the", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "complexity of the COD, the background introduces significant noise, which can distort the attention map to capture the relationships between image patches accurately. Therefore, ARM uses the output of the final adapter, which has absorbed all the image information, to adjust the attention weights.", + "bbox": [ + 511, + 90, + 905, + 180 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nA _ {s e l f} ^ {*} = A _ {s e l f} \\odot E m b _ {A} ^ {l a s t}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 617, + 181, + 905, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A_{\\text{self}}^{*}$ denotes the refined attention weights, and $Emb_{A}^{\\text{last}}$ represents the image embedding extracted from the final adapter module.", + "bbox": [ + 511, + 202, + 905, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, the masked attention weights are generated as affinity weights to refine the initial multimodal prompt:", + "bbox": [ + 511, + 250, + 905, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {P} _ {\\text {r e f i n e}} = \\left(A _ {\\text {s e l f}} ^ {*} \\odot \\operatorname {v e c} (M)\\right) \\otimes \\mathrm {P} _ {\\text {i n i t}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 304, + 905, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $vec(M)$ is the vectorization of mask $M$ , which are generated by extracting highlighted regions of the $\\mathbf{P}_{init}$ , and $\\otimes$ is the matrix multiplication. Finally, by setting a threshold, the highlighted target region from the $\\mathbf{P}_{refine}$ is extracted to generate a pseudo-mask, which serves as the final mask prompt $\\mathbf{P}_{mask}$ .", + "bbox": [ + 511, + 329, + 905, + 421 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Embedding Inputs for SAM", + "text_level": 1, + "bbox": [ + 511, + 428, + 764, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the mask prompt is generated, we will provide three inputs to SAM's mask decoder to support it in completing the COD task.", + "bbox": [ + 511, + 450, + 905, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, the primary generated mask prompt is encoded by SAM's prompt encoder $E_{S_{prompt}}$ , ultimately producing a dense embedding:", + "bbox": [ + 511, + 496, + 905, + 541 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {d e n s e} = \\mathrm {E} _ {S _ {p r o m p t}} \\left(\\mathrm {P} _ {m a s k}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 550, + 903, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Second, the image embedding produced by the final transformer layer of SAM's image encoder is combined with the visual features from the last adapter to generate the final image embedding:", + "bbox": [ + 511, + 575, + 905, + 636 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {S _ {i m g}} ^ {\\text {o u t}} = E m b _ {A} ^ {\\text {l a s t}} + E m b _ {S _ {i m g}} ^ {\\text {l a s t}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 643, + 903, + 664 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, to make the most of the generated multiple information, we encode $t_{noun}$ into the text embedding $Emb_{M_{text}}$ using Mamba [11], a sequence modeling architecture known for its advanced ability to process textual information. The generated $Emb_{M_{text}}$ clearly conveys the semantic information of the camouflaged target category and is concatenated with the $Emb_{B_{img}}$ extracted from BLIP to generate the auxiliary sparse embedding to provide richer features:", + "bbox": [ + 511, + 670, + 905, + 804 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nE m b _ {A u x} = \\operatorname {C o n c a t} \\left(E m b _ {M _ {t e x t}}, E m b _ {B _ {i m g}}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 815, + 903, + 832 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the $\\mathrm{Concat}(\\cdot)$ is the concatenate operation.", + "bbox": [ + 511, + 840, + 839, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, $Emb_{dense}$ , along with the $Emb_{Simg}^{out}$ and $Emb_{Aux}$ , are input into SAM's mask decoder to obtain the final segmentation result.", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20676", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/1a35930ad5dca829a62ecb0914e6ae3e441eaf5ad1302d38e9204368491b4256.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPub. YearCHAMELEONCAMOCOD10KNC4K
SmFβmaxEφMSmFβmaxEφMSmFβmaxEφMSmFβmaxEφM
ZoomNet [42]CVPR220.9010.8760.9470.0230.8200.8050.8890.0660.8370.7770.8960.0290.8520.8260.9030.044
SegMaR [25]CVPR220.9060.8880.9590.0250.8160.8030.8840.0710.8330.7740.9060.0340.8410.8260.9070.046
UEDG [39]TMM230.9110.8940.9680.0230.8630.8560.9290.0480.8580.8120.9340.0250.8790.8640.9350.035
TinyCOD [58]ICASSP230.8870.8610.9580.0300.8220.8070.8990.0660.8110.7420.9030.0360.8430.8170.9100.047
MSCAF-Net [36]TCSVT230.9120.9020.9700.0220.8730.8670.9370.0460.8650.8230.9360.0240.8870.8740.9420.032
FSPNet [24]CVPR230.9090.8900.9650.0230.8560.8460.9280.0500.8510.7940.9300.0260.8790.8590.9370.035
FEDER [16]CVPR230.8870.8680.9540.0300.8020.7890.8730.0710.8220.7680.9050.0320.8470.8330.9150.044
FSNet [49]TIP230.9050.8910.9750.0220.8800.8780.9410.0410.8700.8330.9480.0230.8910.8800.9480.031
CRNet [19]AAAI230.8180.7920.9090.0460.7350.7070.8300.0920.7330.6360.8450.049————
SDRNet [13]KBS240.9140.9010.9610.0240.8720.8670.9320.0490.8710.8280.9360.0230.8890.8760.9400.032
PRNet [23]TCSVT240.9150.9020.9730.0200.8720.8670.9300.0500.8730.8390.9430.0220.8910.8810.9420.031
CamoFormer-R [59]TPAMI240.8980.8800.9490.0260.8170.8010.8830.0680.8380.7860.9280.0290.8540.8290.9080.042
CamoFormer-P [59]TPAMI240.9100.8980.9660.0220.8720.8680.9380.0460.8690.8290.9390.0230.8920.8800.9460.030
ZoomNeXtpv2b3 [43]TPAMI240.9280.9190.9770.0170.8850.8860.9420.0420.8950.8640.9510.0180.9000.8910.9480.028
ZoomNeXtpv2b2 [43]TPAMI240.9220.9080.9690.0170.8740.8730.9310.0470.8870.8560.9450.0190.8920.8840.9410.030
DSAM [60]ACMM24————0.8320.8340.9200.0610.8450.8050.9300.0330.8720.8640.9420.040
MAMIFNet [55]IF250.9140.8990.9590.0210.8720.8700.9350.0450.8690.8260.9400.0230.8900.8780.9430.031
COMPrompter [64]SCIS250.8850.8640.9570.0300.8530.8560.9310.0540.8600.8260.9460.0270.8800.8760.9460.036
Ours—0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.0200.9060.8970.9470.033
", + "bbox": [ + 94, + 87, + 903, + 308 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02641099111cdf0262d4c159d7a66afaee13fa3e3b4b6965df5a4e203f9047db.jpg", + "image_caption": [ + "Table 1. Quantitative comparison with other SOTA methods across different datasets. The best results are highlighted in bold." + ], + "image_footnote": [], + "bbox": [ + 94, + 345, + 334, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2d4f7ae988551f9c6c24a05f35675d8e54c8d22b97738619bacd2879db67628.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 345, + 614, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6600128e05468b003601c1e71e6935e069f87abd042f645e922233193518a4df.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 345, + 898, + 494 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b76efdc460cc9b6c856481aba53e7db3c8b4d7f07dfb415e393fe5951a623dba.jpg", + "image_caption": [ + "Figure 4. $F_{\\beta}$ and PR curves of our work compared to other SOTA methods." + ], + "image_footnote": [], + "bbox": [ + 96, + 503, + 339, + 652 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3c6840d9d3725a06ddd38fc64291a34ede3bb546301309a6b94338575153432.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 375, + 503, + 620, + 651 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/591d62dc81d4ab271906060bf90b4a12a3787020f10ed73d2d47456959b48049.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 658, + 503, + 901, + 651 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 703, + 220, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 734, + 279, + 752 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. To verify the effectiveness of the proposed architecture in the COD task, we use four commonly used COD datasets for evaluation: CHAMELEON [48], CAMO [28], COD10K [10], and NC4K [38]. In detail, CHAMELEON consists of 76 images featuring highly camouflaged animals in natural environments, all used for testing. CAMO contains 1250 images, including both biological camouflage and artificially designed camouflage. Among them, 1000 images are used for training, while the rest are reserved", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "for testing. COD10K is more extensive, comprising 5066 images that include both artificial and natural camouflage, with 3040 images allocated for training and the remainder used for testing. While NC4K comprises 4,121 testing images specifically designed for the COD task. Notably, all four datasets provide only image data and do not include additional modalities, such as text.", + "bbox": [ + 511, + 704, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. In the subsequent experiments, we mainly use four metrics for evaluation: S-measure [8] $(Sm)$ , maximum F-measure [1] $(F_{\\beta}^{max})$ , maximum E-measure [9] $(E_{\\varphi}^{max})$ , and Mean Absolute Error $(M)$ . Among these, higher values of the $Sm$ , $F_{\\beta}^{max}$ , and $E_{\\varphi}^{max}$", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20677", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74c8575355ce31e2fdf9d320520eb450d5219d8f9cdbfce68cfb7cfb0cbac8fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 93, + 88, + 179, + 154 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a4f26e241fc3d48d7f1813d74b6e0dba4b906dbe58484f652853a7720606df79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 88, + 269, + 154 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ab8f8ed22efe07310571b887f51542350ce6721fc1a035a612e6253b69583d08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 88, + 359, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/69f387c467cc62a7c1f9aef7465865cc850abc79c9c98c2764bc0cb66ec09ca7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 88, + 450, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a029378a929356983f01b74b51ed45e4d41ff688125d94c9e3104db254f7657.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 88, + 540, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ecf46666e03eae29667354466666f7aa20b6bd79d76e4ec88add95ad99014380.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 89, + 630, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d9508d34edeac472c55f7551a6aa193a7f9d9debd28404c63a5dc7dc95ee8132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 89, + 720, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1ccda65e372f20b5395cd6e70c0856e85db95a257c774291fc02b05ca64fc11b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 89, + 812, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b6b4e9c96cd655459fcda4f1d870dc12ebab1ab493f15d9d079762373b442d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 89, + 903, + 152 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fb600f8f53b41ef36b1133e4b04a82c9e09213b5ef78d86dcf5cbcfb5d67f207.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 157, + 179, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b7682f308ab913ae53d398da451e5533d4868bddc380c37506a0c07f4fbe02ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 157, + 269, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/afe1446bf035a6836f665cc3832462da8af125919dd69e135d01a2a78d0b2753.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 157, + 359, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3bca2ccd10b949f19779e1d0c9bb679fa58013fb121c582cea93cbb79ccff60a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 157, + 450, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3dd6fc12da9b18382674a264b19bc070583d5c5615811afb35582f589b25a774.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 157, + 540, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c269fd63974c9c55200c7ee70b7bc571dc77782e884f7182b2f0d0d1a60c510b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 157, + 630, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c310ab2934aa67d9ff9a8f286fc0c15f5aacb3b6ac0f789656f41b5616b70952.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 157, + 720, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ee13013b871a79d6efe1c2729d938335821a2c99594bb6c3b824b99a0cc2a470.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 157, + 812, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b757697e323d173e790123f065a93d993420c70abd04844bcd2ad7af67c4f4d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 157, + 903, + 224 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/08f1ad7ec29e475d5c77c759b74f5be5237adeead3d5afac22a6ce13d33db15e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 229, + 179, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/652d58dd61d770cb3d8730c32dace42e26640522778f43d7294e5b655c76f78e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 183, + 229, + 269, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d01364f9fa35dad1aaa1a0c1f333c230d64f4a92e1b40f5d5ad338b8c4240321.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 229, + 359, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2c4d4bd31e2d112e685d56835932188124fafd11dfc5ed2be0a1b6bd6cee7722.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 366, + 229, + 450, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/18283c3e19110b5b6084c8c08ec2a9eb502254a49a42360b28b272d48383fcb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 455, + 229, + 540, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1d47931c7f5375bbc4f8b5522d1faa255d089ff85aba4d6b2295ae76cdd894c0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 547, + 229, + 630, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/903c58fd611f2580a6f225a83ef5938764e08c5769ff8913a161a9d432b9476f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 229, + 720, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7c832920d4c9c8215b03be14a922777c33aec278aa387bebc3a5450d654f122f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 727, + 229, + 812, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16f4d64288f2fd1070515fa0f02c0a1ba10fb14566e1002047277e97996d824d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 818, + 229, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b95ec0903d37511d43a43b3d122dbf783e444810395e5b0e175e424766787448.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 299, + 179, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/50c2e4853535d36f32580ae080100896f8f71a6a2db64bc24d93249392f5830c.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 183, + 300, + 269, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a065d746fd94323fd4e9743e6a3837638d3508839bed335401374a41bc7cdc0f.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 274, + 300, + 359, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1b46b78b13e1b9729d5e30566c9ec126ba07274705c5c0cf0dd4596502895108.jpg", + "image_caption": [ + "COMPrompter" + ], + "image_footnote": [], + "bbox": [ + 366, + 300, + 450, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e0f917b7ec1d27e6998821f8ab5a5505fdcf782b43a0bb663bd31459ca581031.jpg", + "image_caption": [ + "ZoomNeXt" + ], + "image_footnote": [], + "bbox": [ + 455, + 300, + 540, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/77938d9574bb33dc48fbc47b8490a06f14619d18f34b58f834a448db922588f7.jpg", + "image_caption": [ + "PRNet", + "Figure 5. Visual comparison of segmentation results with other SOTA methods on different samples." + ], + "image_footnote": [], + "bbox": [ + 547, + 300, + 630, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b1a80939311555b022867bfb6dd91248b884705d4077cd4ac9bba3036bca0fbf.jpg", + "image_caption": [ + "FSNet" + ], + "image_footnote": [], + "bbox": [ + 638, + 300, + 720, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d89ace3c2f3d616640d46d23530e6f8f6ae8d65de8fae9280af7ff96ae879a21.jpg", + "image_caption": [ + "FSPNet" + ], + "image_footnote": [], + "bbox": [ + 727, + 300, + 812, + 364 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71be723cb02203bee17bdbec6fcc278860063392b3ca9ce504da644b18ee5edd.jpg", + "image_caption": [ + "ZoomNet" + ], + "image_footnote": [], + "bbox": [ + 818, + 300, + 903, + 364 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indicate better model performance, whereas for $M$ , lower values indicate better performance. Additionally, the precision-recall curve (PR curve) and the $F_{\\beta}$ -threshold curve $(F_{\\beta}$ curve) are also considered for further evaluation.", + "bbox": [ + 89, + 430, + 483, + 492 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implement Details. Our proposed framework is primarily built upon the SAM-Vit Huge (ViT-H) [27] and implemented using PyTorch. The entire architecture is trained on an NVIDIA GeForce RTX 4090 GPU for 40 epochs with a batch size of 1. The main learning rate is set to 0.0002, while the learning rate for the ARM is 0.1 times that of the main learning rate, and the AdamW optimizer is used for optimization. Furthermore, both the VLM and LLM, used for multimodal prompt and auxiliary embedding generation, are pre-trained models, so no additional training or fine-tuning is required for these two.", + "bbox": [ + 89, + 498, + 483, + 666 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Comparisons With State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 674, + 465, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To effectively evaluate the performance of our work in the COD task, we first conducted a comparative analysis with the current state-of-the-art (SOTA) models.", + "bbox": [ + 89, + 696, + 483, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Evaluation. As shown in Tab. 1, which compares the evaluation results of our work and other SOTA models across four different datasets. Our work generally demonstrates exceptional performance in solving the COD task across various metrics on different datasets. In particular, our method excels in both the $Sm$ and $F_{\\beta}^{max}$ , showing superior performance in target localization and structure capture of the camouflaged objects in complex scenarios. For example, on the COD10K, our approach outperforms the second-best method by $1.6\\%$ in $Sm$ and $2.4\\%$ in", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$F_{\\beta}^{max}$ . Although our $F_{\\beta}^{max}$ is the second on the CAMO, the gap with the top-performing model is only $0.3\\%$ , which is negligible. Furthermore, from the comparison of the $F_{\\beta}$ and PR curves in Fig. 4, it is clear that our method (red solid curve) exhibits outstanding performance across all three datasets, with particularly remarkable results on the COD10K dataset.", + "bbox": [ + 511, + 430, + 906, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Evaluation Fig. 5 illustrates a comparison of the final segmentation maps of our work and other SOTA methods across different samples. It is clear that our approach achieves a more significant overlap with the ground truth (GT) than other SOTA models. Specifically, our method is more effective at accurately segmenting and capturing completely camouflaged objects. Even in more challenging scenarios, such as the third sample in Fig. 5, where multiple small camouflaged objects need to be identified, our approach segments more objects and provides more accurate segmentation, while the results from other models show noticeable omissions. These observations highlight the superior performance of our framework, not only in localizing camouflaged objects but also in maintaining the structural integrity of the camouflaged objects.", + "bbox": [ + 511, + 544, + 908, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 780, + 679, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 2, we conduct ablation studies on the improvements discussed in Sec. 3 to assess the effectiveness of each component.", + "bbox": [ + 511, + 801, + 906, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Auxiliary Embedding. Since the auxiliary embedding is generated alongside the multimodal data produced by BLIP, we first incorporate it into the SAM to", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20678", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c22e13e234176b3d8b920876a8d73098edac80d64ba1ab3a670daedf45226ea3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CHAMELEONCAMOCOD10K
SmFβmaxEφmaxMSmFβmaxEφmaxMSmFβmaxEφmaxM
SAM0.8500.8200.9110.0560.7970.7640.8640.1010.8100.7500.8810.054
SAM + EmbAUX0.8520.8160.9220.0470.8050.7780.8820.0890.8200.7510.8930.039
SAM + EmbAUX + Pcaptioninit0.8660.8360.9200.0460.8230.7950.8910.0810.8390.7760.9080.037
SAM + EmbAUX + Pnouninit0.8820.8590.9350.0400.8310.8070.8960.0800.8460.7890.9140.036
SAM + EmbAUX + PnounARM0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.020
", + "bbox": [ + 96, + 88, + 903, + 188 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Ablation studies on three different COD datasets. + denotes additional information provided to SAM, $Emb_{Aux}$ is the auxiliary sparse embedding, $P_{init}^{caption}$ is the initial mask prompt generated by the entire caption, $P_{init}^{noun}$ is the initial mask prompt generated by the extracted noun, $P_{ARM}^{noun}$ is the refined $P_{init}^{noun}$ by ARM.", + "bbox": [ + 89, + 198, + 906, + 241 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96f6bd4d5374a0023968bf3ec8648745c9f51c7a9e15aa2606c5af2d22986ebb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 265, + 187, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0ad3fdf6ce28e4641430ece9c327ddc13b8dbe9ec966323448d944d6ee69132.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 324, + 187, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9e02f4db4184054407802e9b3a946578375a2456bd82aa3048e9034244fde7e7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 382, + 187, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c9ace7879d2e460dcdedbdf5daad351fe0c6225bc00500908870bb0ff583dce.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 430, + 187, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f2d92d68d01c47cb3b7a17a7c6132d1fc87084f3e1fd3973f2a35f36cddaab1d.jpg", + "image_caption": [ + "Image" + ], + "image_footnote": [], + "bbox": [ + 94, + 479, + 187, + 531 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b3ae903b0735e5236a189dfd0d8b1bbb9f1e94c9a7d3819a4bd90e74d6faa3f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 191, + 266, + 285, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bc5237428dd668c32af7793d32dfac700b9a4bda0cc33ab45f1702f171f7f2a9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 324, + 285, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/accd0adc389c776c939aac9a94d508e4c4f8f81047dd0fb6545c27875463e72c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 382, + 285, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/96d9d34953ce2a0a72507e3c6af22086615e89e9daf4ee0db05a9ea3735dff13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 430, + 285, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a2b87cad6a2461aafa78aa5689b4e660d89c42549587ae483f3d7d74e1087a2.jpg", + "image_caption": [ + "GT" + ], + "image_footnote": [], + "bbox": [ + 191, + 479, + 285, + 530 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3ac3e62534fc3a5b73d9df592601fb972982eaef8548d0db14a84aa8ef79fdbd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 266, + 382, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/41bd62e19c8b1a689260eb1648bb12cc980e8b545346a9e5c26076b70d445761.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 324, + 382, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a2380f758b652612400f1a056636b16c03ab261c3405726bc6c531be1feda1b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 382, + 382, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/19d4394c6805bb428218f9cad76f9d7a2e5ded53e44ee5bbba46d1148a8b1062.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 430, + 382, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e0c709bf75816ce94a630106acba9aee7a84318a69e02f66e3608f0f99d9a350.jpg", + "image_caption": [ + "Initial" + ], + "image_footnote": [], + "bbox": [ + 287, + 479, + 382, + 530 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ae79fa60091721b506a01f02a062cc8cb0656809774d2fade069cc3461b4f91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 266, + 480, + 321 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28e7c04f602f25e9c4eb1a6182878d8d43b46f8f11338d0e65524bce331a003b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 324, + 480, + 378 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/34b5beeb6abfeb220152cb29ecb4ed28d614236100f638debc200dda9c4904ee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 382, + 480, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4999b7bbda47d2425a4db34b8e11a212f829ce1e7fed7b0f4f1e5121f0f077bd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 385, + 430, + 480, + 474 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7a57f11a0171448c428431c7c785cd336d6e22be18aaa1b79449f6dedec3f365.jpg", + "image_caption": [ + "Refined by ARM", + "Figure 6. Comparison between the initially generated prompt and the prompt refined using ARM." + ], + "image_footnote": [], + "bbox": [ + 385, + 479, + 480, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "evaluate its effectiveness. Comparing the $2^{nd}$ row with the $1^{st}$ row in Tab. 2, adding the auxiliary embedding leads to performance improvements across nearly all metrics on the three COD datasets. This indicates that the auxiliary embedding effectively enhances SAM by providing additional multimodal features.", + "bbox": [ + 89, + 616, + 483, + 705 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Multimodal Prompt. Our primary contribution is utilizing BLIP to generate multimodal prompts, providing SAM with more informative cues to handle the COD task. Experimental results ( $3^{rd}$ row of Tab. 2) show that prompts generated from full description already outperform SAM with auxiliary embedding. Furthermore, SAM's performance further improved when extracting the noun about the object category from the long description as textual information to generate prompts ( $4^{th}$ row of Tab. 2). This result not only proves the effectiveness of our generated prompts but also indicates that, although BLIP can produce descriptions closely related to the image, redun-", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "dant information introduces noise, preventing the generated prompt from accurately localizing the targets. In contrast, providing only category information helps generate more precise mask prompts.", + "bbox": [ + 511, + 267, + 905, + 329 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Refinement. As shown in the $5^{th}$ row results in Tab. 2, our proposed ARM significantly enhances SAM's performance on the COD task, achieving substantial improvements across all evaluation metrics on three datasets. This performance boost can be attributed to two key aspects: First, ARM refines the multimodal prompts initially generated by BLIP, as illustrated in Fig. 6. Due to domain shifts, these initial prompts often localize only partial regions of the target. However, after refining with ARM, the refined prompts achieve more complete target coverage. Consequently, the mask prompts generated from these refined prompts provide SAM with more precise cues. Second, ARM effectively integrates the image features extracted by the adapter into SAM, enabling fine-tuning of the image encoder. This process further enhances SAM's representation capability, leading to improved performance on the COD task.", + "bbox": [ + 511, + 349, + 906, + 604 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 641, + 633, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we propose a novel and cost-effective solution for the COD task. The main idea is to incorporate multimodal information to construct high-quality mask prompts, enhancing SAM's segmentation capability. Specifically, we use VLM to automatically generate multimodal data and prompts, laying a solid foundation for constructing the final high-quality mask prompts. Furthermore, we propose the Adaptive Refinement Module (ARM) to refine the initial multimodal prompts enabling more complete target representations and fine-tune SAM's image encoder, providing it with richer information. Additionally, experimental results demonstrate that our proposed architecture achieves outstanding performance across multiple COD datasets, particularly excelling in capturing the structural details of objects, thereby further validating our work's effectiveness.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20679", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 256, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work is supported by the XJTLU RDF project: RDF-22-02-118 & RDF-24-02-041", + "bbox": [ + 91, + 114, + 480, + 143 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 150, + 187, + 166 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 6", + "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 1", + "[3] Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. Sam-cod+: Sam-guided unified framework for weakly-supervised camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2", + "[4] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023. 2", + "[5] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?-sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2304.09148, 2023. 2", + "[6] Zhennan Chen, Rongrong Gao, Tian-Zhu Xiang, and Fan Lin. Diffusion model for camouflaged object detection. In ECAI 2023, pages 445-452. IOS Press, 2023. 2", + "[7] Zhongxi Chen, Ke Sun, and Xianming Lin. Camodiffusion: Camouflaged object detection via conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1280, 2024. 2", + "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 6", + "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6", + "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 6", + "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5", + "[12] Juwei Guan, Xiaolin Fang, Tongxin Zhu, Zhipeng Cai, Zhen Ling, Ming Yang, and Junzhou Luo. Idenet: Making neural network identify camouflaged objects like creatures. IEEE Transactions on Image Processing, 2024. 1", + "[13] Juwei Guan, Xiaolin Fang, Tongxin Zhu, and Weiqi Qian. Sdrnet: Camouflaged object detection with independent re" + ], + "bbox": [ + 93, + 175, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "construction of structure and detail. Knowledge-Based Systems, 299:112051, 2024. 6", + "[14] Juwei Guan, Weiqi Qian, Tongxin Zhu, and Xiaolin Fang. Promoting camouflaged object detection through novel edge-target interaction and frequency-spatial fusion. Neurocomputing, 617:129064, 2025. 1, 2", + "[15] Cunhan Guo and Heyan Huang. Cofinet: Unveiling camouflaged objects with multi-scale finesse. Neurocomputing, 614:128763, 2025. 2", + "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22046-22055, 2023. 6", + "[17] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36:30726-30737, 2023. 2", + "[18] Jiangping He, Yangjie Zhang, Maggie Chung, Michael Wang, Kun Wang, Yan Ma, Xiaoyang Ding, Qiang Li, and Yonglin Pu. Whole-body tumor segmentation from pet/ct images using a two-stage cascaded neural network with camouflaged object detection mechanisms. Medical Physics, 50 (10):6151-6162, 2023. 1", + "[19] Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI conference on artificial intelligence, pages 781-789, 2023. 6", + "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5", + "[21] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 4", + "[22] Bintao Hu, Hengyan Liu, Jianbo Du, Miguel López-Benitez, Celimuge Wu, Xiaoli Chu, and Dusit Niyato. Madqnenhanced computation offloading and resource allocation for 6g low-altitude economy vehicular networks. IEEE Transactions on Cognitive Communications and Networking, 2025. 1", + "[23] Xihang Hu, Xiaoli Zhang, Fasheng Wang, Jing Sun, and Fuming Sun. Efficient camouflaged object detection network based on global localization perception and local guidance refinement. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):5452-5465, 2024. 6", + "[24] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5557-5566, 2023. 6", + "[25] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4713-4722, 2022. 6" + ], + "bbox": [ + 517, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20680", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Xinhao Jiang, Wei Cai, Yao Ding, Xin Wang, Danfeng Hong, Xingyu Di, and Weijie Gao. Bidicos: Camouflaged object segmentation via bilateral diffusion model. Expert Systems with Applications, 255:124747, 2024. 2", + "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 7", + "[28] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and image understanding, 184:45-56, 2019. 6", + "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 4", + "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4", + "[31] Lin Li, Bo Dong, Eric Rigall, Tao Zhou, Junyu Dong, and Geng Chen. Marine animal segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(4): 2303-2314, 2021. 1", + "[32] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 5", + "[33] Hengyan Liu, Guangyu Ren, Tianhong Dai, Di Zhang, Pengjing Xu, Wenzhang Zhang, and Bintao Hu. Diversity-oriented contrastive learning for rgb-t scene parsing. In 2023 2nd International Conference on Sensing, Measurement, Communication and Internet of Things Technologies (SMC-IoT), pages 155-160. IEEE, 2023. 1", + "[34] Hengyan Liu, Wenzhang Zhang, Tianhong Dai, Longfei Yin, and Guangyu Ren. Multimodal frequency spectrum fusion schema for rgb-t image semantic segmentation. In 2024 33rd International Conference on Computer Communications and Networks (ICCNN), pages 1-6. IEEE, 2024. 1", + "[35] Limei Liu, Meng Liu, Kexin Meng, Linqiang Yang, Min Zhao, and Shuli Mei. Camouflaged locust segmentation based on pranet. Computers and Electronics in Agriculture, 198:107061, 2022. 1", + "[36] Yu Liu, Haihang Li, Juan Cheng, and Xun Chen. Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features. IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4934-4947, 2023. 6", + "[37] Yantong Liu, Sai Che, Liwei Ai, Chuanxiang Song, Zheyu Zhang, Yongkang Zhou, Xiao Yang, and Chen Xian. Camouflage detection: Optimization-based computer vision for alligator sinensis with low detectability in complex wild environments. Ecological Informatics, 83:102802, 2024. 1" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11591-11601, 2021. 6", + "[39] Yixuan Lyu, Hong Zhang, Yan Li, Hanyang Liu, Yifan Yang, and Ding Yuan. Uedg: Uncertainty-edge dual guided camouflage object detection. IEEE Transactions on Multimedia, 26:4050-4060, 2023. 6", + "[40] Tsui Qin Mok, Shuyong Gao, Haozhe Xing, Miaoyang He, Yan Wang, and Wenqiang Zhang. A holistically point-guided text framework for weakly-supervised camouflaged object detection. arXiv preprint arXiv:2501.06038, 2025. 3", + "[41] Yuzhen Niu, Lifen Yang, Rui Xu, Yuezhou Li, and Yuzhong Chen. Minet: Weakly-supervised camouflaged object detection through mutual interaction between region and edge cues. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6316-6325, 2024. 1", + "[42] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2160-2170, 2022. 2, 6", + "[43] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE transactions on pattern analysis and machine intelligence, 2024. 1, 2, 6", + "[44] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. arXiv preprint arXiv:2311.10121, 2023. 2", + "[45] Guangyu Ren, Yinxiao Yu, Hengyan Liu, and Tania Stathaki. Dynamic knowledge distillation with noise elimination for rgb-d salient object detection. Sensors, 22(16):6188, 2022. 1", + "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4", + "[47] Yiqing Shen, David Dreizin, Blanca Inigo, and Mathias Unberath. Protosam-3d: Interactive semantic segmentation in volumetric medical imaging via a segment anything model and mask-level prototypes. Computerized Medical Imaging and Graphics, page 102501, 2025. 2", + "[48] Przemysław Skurowski, Hassan Abdulameer, Jakub Blaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6", + "[49] Ze Song, Xudong Kang, Xiaohui Wei, Haibo Liu, Renwei Dian, and Shutao Li. Fsnet: Focus scanning network for camouflaged object detection. IEEE Transactions on Image Processing, 32:2267-2278, 2023. 6", + "[50] Yanguang Sun, Chunyan Xu, Jian Yang, Hanyu Xuan, and Lei Luo. Frequency-spatial entanglement learning for camouflaged object detection. In European Conference on Computer Vision, pages 343-360. Springer, 2024. 2" + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20681", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8805–8814, 2024. 3", + "[52] Huafeng Wang, Tianyu Hu, Yanan Zhang, Haodu Zhang, Yong Qi, Longzhen Wang, Jianhua Ma, and Minghua Du. Unveiling camouflaged and partially occluded colorectal polyps: Introducing cpsnet for accurate colon polyp segmentation. Computers in Biology and Medicine, 171:108186, 2024. 1", + "[53] Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. Depth-aware concealed crop detection in dense agricultural scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17201-17211, 2024. 1, 3", + "[54] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3297–3306, 2023. 3", + "[55] Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection. Information Fusion, 117:102871, 2025. 2, 6", + "[56] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1032-1042, 2023. 3", + "[57] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv preprint arXiv:2106.13217, 2021. 3", + "[58] Haozhe Xing, Shuyong Gao, Hao Tang, Tsui Qin Mok, Yanlan Kang, and Wenqiang Zhang. Tinycod: Tiny and effective model for camouflaged object detection. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 6", + "[59] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6", + "[60] Zhenni Yu, Xiaogin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. Exploring deeper! segment anything model with depth perception for camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4322-4330, 2024. 6", + "[61] Hong Zhang, Yixuan Lyu, Qian Yu, Hanyang Liu, Huimin Ma, Ding Yuan, and Yifan Yang. Unlocking attributes' contribution to successful camouflage: A combined textual and visual analysis strategy. In European Conference on Computer Vision, pages 315-331. Springer, 2024. 1, 3", + "[62] Jie Zhang, Xubing Yang, Rui Jiang, Wei Shao, and Li Zhang. Rsam-seg: A sam-based approach with prior knowledge in" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tegration for remote sensing image semantic segmentation. arXiv preprint arXiv:2402.19004, 2024. 2", + "[63] Shizhou Zhang, Dexuan Kong, Yinghui Xing, Yue Lu, Lingyan Ran, Guoqiang Liang, Hexu Wang, and Yanning Zhang. Frequency-guided spatial adaptation for camouflaged object detection. IEEE Transactions on Multimedia, 27:72-83, 2025. 2", + "[64] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 6", + "[65] Ke Zhou, Zhongwei Qiu, and Dongmei Fu. Multi-scale contrastive adaptor learning for segmenting anything in under-performed scenes. Neurocomputing, 606:128395, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 303 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20682", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_model.json b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..7d8b64a5944b7b6a9b5c483c04b38b461a6d1d0e --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_model.json @@ -0,0 +1,3203 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.16, + 0.203, + 0.845, + 0.222 + ], + "angle": 0, + "content": "Xuehan Chen\\(^{1}\\) Guangyu Ren\\(^{1*}\\) Tianhong Dai\\(^{2}\\) Tania Stathaki\\(^{2}\\) Hengyan Liu\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.222, + 0.676, + 0.239 + ], + "angle": 0, + "content": "\\(^{1}\\)Xi'an Jiaotong-Liverpool University, China" + }, + { + "type": "text", + "bbox": [ + 0.325, + 0.239, + 0.674, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Imperial College London, United Kingdom" + }, + { + "type": "text", + "bbox": [ + 0.22, + 0.26, + 0.778, + 0.275 + ], + "angle": 0, + "content": "xuehanchen83@gmail.com, {guangyu. ren, hengyan.liu}@xjtlu.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.278, + 0.776, + 0.292 + ], + "angle": 0, + "content": "tianhong.dai15@alumni.imperial.ac.uk, t.stathaki@imperial.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.485, + 0.662 + ], + "angle": 0, + "content": "Foundation models, such as Segment Anything Model (SAM), have exhibited remarkable performance in conventional segmentation tasks, primarily due to their training on large-scale datasets. Nonetheless, challenges remain in specific downstream tasks, such as Camouflaged Object Detection (COD). Existing research primarily aims to enhance performance by integrating additional multimodal information derived from other foundation models. However, directly leveraging the information generated by these models may introduce additional biases due to domain shifts. To address this issue, we propose an Adaptive Refinement Module (ARM), which efficiently processes multimodal information and simultaneously refining the mask prompt. Furthermore, we construct an auxiliary embedding that effectively exploits the intermediate information generated during ARM, providing SAM with richer feature representations. Experimental results indicate that our proposed architecture surpasses most state-of-the-art (SOTA) models in the COD task, particularly excelling in structured target segmentation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.694, + 0.222, + 0.71 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.721, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Camouflaged Object Detection (COD) is a challenging computer vision task, as it involves segmenting objects that are virtually identical to their immediate surroundings. It plays a crucial role in the field due to its wide range of practical applications, including object detection [34, 45], low-altitude economy [22], zoological research [31, 35, 37], and medical image segmentation [18, 52]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.828, + 0.484, + 0.874 + ], + "angle": 0, + "content": "Current research [12, 14, 33, 41, 43] focuses on designing advanced architectures to extract and refine more detailed features exclusively from image data. However, the" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.325, + 0.903, + 0.492 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.497, + 0.9, + 0.573 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.659, + 0.581, + 0.706, + 0.594 + ], + "angle": 0, + "content": "(c) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Figure 1. Comparison of our proposed Adaptive Refinement Module (ARM) with common COD frameworks. Unlike conventional methods that rely solely on images or manually annotated additional modalities, our approach not only leverages vision-language models (VLMs) to automatically generate multimodal data but also mitigates domain shift through the proposed ARM, providing high-quality prompts for SAM's decoder. Meanwhile, ARM can also fine-tune SAM's image encoder, further improving SAM's feature representation capability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "unique nature of the COD task, where the targets are visually similar to their surroundings, limits the effectiveness of image features. Additionally, the lack of data from other modalities further limits the models' ability to acquire meaningful features, making them highly susceptible to background, preventing them from accurately detecting camouflaged objects. Although some studies incorporate human-made auxiliary data, such as depth images [53], textual descriptions [61], or prompts [2, 3], to provide supple" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20672" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "mentary information about targets, they are not ideal solutions as they are time-consuming and labor-intensive. As a result, despite significant progress in the COD task, two limitations remain: limited image features and laborious annotation processes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.17, + 0.482, + 0.351 + ], + "angle": 0, + "content": "Recently, foundation models have emerged as a popular solution for tackling a range of tasks, including the Segment Anything Model (SAM) [27], which has proven highly effective in handling various tasks, such as medical image segmentation [44, 47] and semantic segmentation [4, 62]. Although SAM demonstrates impressive zero-shot generalization capabilities, Chen et al. [5] states that it still struggles to segment the complex camouflaged objects. Secondly, SAM can guide its segmentation by receiving prompts, and the most intuitive approach is to generate prompts through manual annotation, but as mentioned earlier, it is less practical due to the time and labor it demands." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.353, + 0.483, + 0.595 + ], + "angle": 0, + "content": "To mitigate the aforementioned limitations, we rethink the core issues that make the COD task challenging. In this task, there are no task-specific clues (e.g., category, size, location) of the targets. Meanwhile, the background's visual noise increases the task's complexity, resulting in ambiguous object boundaries and weakened feature learning. So, we propose a novel approach that uses textual information to complement visual features, enhancing the model's ability to localize target objects more accurately. In this work, we fully utilize the powerful pretrained vision-language models (VLMs) to automatically generate multimodal data without requiring additional training or fine-tuning. This data provides additional cues to assist in object localization. Subsequently, the VLMs further utilize the generated multimodal information to produce mask prompts that convey the object's approximate location to SAM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.822 + ], + "angle": 0, + "content": "We also consider the biases introduced by directly utilizing generated multimodal information, which could lead to locality in the prompts, making them less comprehensive and potentially affecting the subsequent SAM segmentation performance. To mitigate this issue, this paper proposes an Adaptive Refinement Module (ARM), which incorporates simple adapters to refine the initially generated mask prompts. This module expands the regions in the initial prompts to create a more complete representation of the target objects, offering better guidance for downstream models. Furthermore, ARM effectively extracts vision information from the entire architecture, serving not only as a refinement mechanism for prompts but also as a fine-tuning module for SAM's image encoder, further enhancing SAM's performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, this work proposes the ARM, which simultaneously refines the initial prompt generated by multimodal information and fine-tunes SAM's image encoder, enabling SAM to handle the COD task more effectively across multiple aspects. Our main contributions are as fol" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.093, + 0.553, + 0.105 + ], + "angle": 0, + "content": "lows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.111, + 0.905, + 0.201 + ], + "angle": 0, + "content": "- We propose a cost-effective approach that automatically constructs mask prompts using multimodal information generated by BLIP, serving as additional cues without requiring extra training or fine-tuning. These prompts provide reliable guidance for SAM, thereby improving its segmentation performance." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.202, + 0.905, + 0.306 + ], + "angle": 0, + "content": "- To mitigate the additional biases introduced by BLIP, we propose an Adaptive Refinement Module (ARM) to refine the initially generated prompts, enhancing their quality and providing more precise guidance for SAM. Meanwhile, this module can also fine-tune the SAM's image encoder, further enhancing SAM's segmentation accuracy and robustness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.307, + 0.905, + 0.382 + ], + "angle": 0, + "content": "- We further utilize the multimodal information generated by BLIP to construct an auxiliary embedding, which effectively exploits the intermediate information generated during the prompt generation process and provides SAM with richer features." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.111, + 0.905, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.399, + 0.655, + 0.415 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.768, + 0.44 + ], + "angle": 0, + "content": "2.1. Single-Modality Based COD" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.905, + 0.583 + ], + "angle": 0, + "content": "Image data is one of the most widely used modalities to tackle the COD task. Therefore, most researchers are currently exploring various approaches to extract information from single-modality image data in order to perform this task more effectively. CoFiNet [15], ZoomNext [43], ZoomNet [42] focus on extracting and fusing multiscale features from images. Although upgrading the model's backbone improves the accuracy of identifying objects, it also increases the consumption of computational resources." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Besides directly extracting features from the images, CamoDiffusion [7], BiDiCOS [26], and diffCOD [6] transform the COD task into denoising tasks, gradually segmenting camouflaged objects through the denoising process. Moreover, EFNet [14], FSEL [50], FGSA-Net [63], and MAM-FFC [55] use spatial and frequency information from images to achieve a more precise segmentation. However, in the COD task, the background in images introduces significant noise. Even with those novel methods relying solely on images' intrinsic features, they still struggle to recognize objects in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Furthermore, in addition to designing a new architecture, many works have also focused on improving the Segment Anything Model (SAM), taking advantage of its powerful zero-shot generalization ability and prompt strategy. SAM-COD [3] and WS-SAM [17], which combine sparse prompts with SAM to generate the final segmentation results, still require additional manual annotations for the sparse prompts, which remains time-consuming. MCA-SAM [65] and SAM-adapter [5] improve the model's performance in the COD task by injecting more image infor" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20673" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.393 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.403, + 0.907, + 0.489 + ], + "angle": 0, + "content": "Figure 2. Overview of our proposed architecture. BLIP automatically generates multimodal data from the input image and uses it to create the initial multimodal prompts and the auxiliary embedding. The Adaptive Refinement Module (ARM) is then employed to address the locality issue of the initial prompts. The ARM uses multiple adapters to extract visual information from both BLIP and SAM, which not only refines the initial multimodal prompts by integrating attention weights, resulting in more accurate mask prompts, but also fine-tunes SAM's image encoder. Finally, the SAM decoder receives image embedding, dense mask prompt embedding, and auxiliary sparse embedding to tackle the COD task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.513, + 0.482, + 0.605 + ], + "angle": 0, + "content": "mation into SAM's image encoder using the adapter. They differ from our work, where the adapter serves a broader purpose in ARM. It is primarily used to refine the generated mask prompts while injecting additional information into SAM's image encoder, enabling the SAM to perform the COD task more effectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.342, + 0.637 + ], + "angle": 0, + "content": "2.2. Multi-Modality Based COD" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.902 + ], + "angle": 0, + "content": "With the rapid growth of multimodal learning, different modalities have been increasingly applied to address the COD problem. HPGT [40] use the textual information to assist in completing the COD task. However, they rely on manually annotating text data for the COD dataset, which is a highly time-consuming process. Although ACU-MEN [61] utilizes the GPT-4 Vision model to generate textual descriptions and attributions for images, it still requires more than 30 volunteers to manually verify the accuracy of these generated texts, which is also a labor-intensive task. In contrast to these approaches, MMCPF [51] takes a different direction by interacting with a Multimodal Large Language Model (MLLM). They query the model using text to obtain the location coordinates of camouflaged objects, which supports the subsequent segmentation task. While this method reduces the need for text data annotation, it still faces the challenge of designing precise text prompts for querying the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.907, + 0.619 + ], + "angle": 0, + "content": "MLLM. Ordinary prompts cannot let MLLM output the expected location coordinates of objects, which indicates that designing effective text prompts becomes an additional task that is time-consuming and labor-intensive as well. As a result, both visual and textual information have been employed to address the COD problem, but manual annotation and verification remain a limitation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.62, + 0.909, + 0.801 + ], + "angle": 0, + "content": "In addition to textual information, depth features are utilized to provide valuable spatial information for models. Due to the fact that commonly used COD datasets only include RGB images without depth data, PopNet [56], Da-COD [54] and [57] employ depth estimation models to generate depth images from RGB-based COD datasets. Since the depth images are generated automatically by models, the final segmentation performance is highly reliant on the performance of those depth estimation methods. In contrast to the previous approaches, RISNet [53] uses a depth camera to manually create an RGB-D dataset for COD in agricultural scenes, which is also painstaking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.605, + 0.83 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We propose a novel architecture, as illustrated in Fig. 2, which enables the fully automated generation of mask prompts for SAM and designs an innovative Adaptive Refinement Module (ARM) to enhance the quality of the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20674" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.288 + ], + "angle": 0, + "content": "initially generated mask prompts. Specifically, a Vision-Language Model (VLM) first automatically generates textual data based on the input image. Then, by aligning multimodal information, the VLM generates the multimodal prompts to locate camouflaged objects locally. Subsequently, it is fed into the ARM for refinement, enabling prompts to capture more complete object structure information, thereby providing better guidance for SAM. During the prompt refinement process, ARM also fine-tunes SAM's image encoder. Meanwhile, the generated textual data is fully utilized, which is integrated with visual information to provide additional auxiliary embedding, further improving SAM's performance in identifying camouflaged objects." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.289, + 0.484, + 0.38 + ], + "angle": 0, + "content": "In the following subsections, we provide detailed explanations of each component of this architecture. First, we describe the multimodal prompt generation process in Sec. 3.1. Then we introduce the ARM in Sec. 3.2. Finally, Sec. 3.3 describes the different input embeddings fed into SAM's mask decoder to generate the final results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.391, + 0.373, + 0.407 + ], + "angle": 0, + "content": "3.1. Multimodal Prompt Generation" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.414, + 0.342, + 0.428 + ], + "angle": 0, + "content": "3.1.1. Multimodal Data Generation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.484, + 0.554 + ], + "angle": 0, + "content": "As a robust zero-shot VLM, BILP [30] offers powerful image captioning capabilities that align with our goal of automatically generating supplementary textual data for subsequent prompt generation. Given an image \\( I \\in \\mathbb{R}^{H \\times W \\times 3} \\), the image embedding \\( Emb_{B_{img}} \\) are extracted by the BLIP's image encoder \\( E_{B_{img}} \\). This embedding is then passed to BLIP's image-grounded text decoder, which generates a caption, as presented below:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.572, + 0.483, + 0.59 + ], + "angle": 0, + "content": "\\[\nT = \\mathbf {D} _ {B _ {t e x t}} \\left(E m b _ {B _ {i m g}}\\right) = \\left\\{t _ {i j} \\right\\} _ {i = 1} ^ {L}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.688 + ], + "angle": 0, + "content": "where \\(T\\) denotes the generated caption, \\(\\mathrm{D}_{B_{text}}\\) is the BLIP's image-grounded text decoder, and \\(t_i\\) represents the \\(i^{th}\\) word in the caption, with a total of \\(L\\) words. As exemplified by the potential caption generated in Fig. 2, it is clear that the captions generated by BLIP can correctly identify the camouflaged targets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "However, since the generated \\( T \\) provides an overall description of the input image, it may include content that is irrelevant to the target objects, such as \"floor\" is the background information of the description \"There is a small dog that is lying on the floor\". Due to the complexity of the COD task, irrelevant information in the generated caption could act as noise, potentially distracting from generating a prompt that accurately indicates the target's location. Therefore, instead of utilizing the entire caption \\( T \\), we use the spaCy [21] library for part-of-speech tagging to capture the necessary content about the objects. Specifically, each \\( t_i \\) in \\( T \\) is tagged with its corresponding part of speech (e.g., noun, verb, adjective). Through observation, we found that most generated captions primarily use nouns to describe the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "category of targets (e.g., man, bug, lion). Therefore, we extract the first noun \\( t_{noun} \\) that appears in \\( T \\) to provide more precise target category information, which helps the subsequent prompt generation more accurately locate the target object and reduces distractions caused by unrelated noise." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.175, + 0.783, + 0.19 + ], + "angle": 0, + "content": "3.1.2. Multimodal Prompt Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.194, + 0.906, + 0.331 + ], + "angle": 0, + "content": "BLIP has a powerful text-image alignment capability, which allows it to generate a multimodal prompt that can accurately locate the target object by aligning the text containing the target category information with the image. Inspired by [29], we use BLIP once again to handle the image-text retrieval task and generate the Gradient-weighted Class Activation Mapping (Grad-CAM) [46] to visualize the alignment process of BLIP with multimodal data, thereby generating the prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.331, + 0.906, + 0.39 + ], + "angle": 0, + "content": "Specifically, the \\(Emb_{B_{img}}\\) and \\(t_{noun}\\) are fed into BLIP's image-grounded text encoder \\(\\mathrm{E}_{B_{text}}\\), which then outputs a score reflecting how well the input image and text match, as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.407, + 0.905, + 0.425 + ], + "angle": 0, + "content": "\\[\n(S, \\mathrm {A} _ {\\text {c r o s s}}) = \\mathrm {E} _ {B _ {\\text {t e x t}}} \\left(\\left(E m b _ {B i m g}, t _ {\\text {n o u n}}\\right), \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.476 + ], + "angle": 0, + "content": "where \\(S\\) denotes the output score, and \\(A_{cross}\\) is the attention weight from the cross attention layer of \\(\\mathbf{E}_{B_{text}}\\), which captures the relationship between the image and the text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.535 + ], + "angle": 0, + "content": "To generate the Grad-CAM, the gradient is also important, so we conduct a backpropagation process to get the gradient \\( G_{cross} \\), and then generate the prompt, as presented below:" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.55, + 0.905, + 0.567 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} _ {\\text {i n i t}} = \\operatorname {G r a d C A M} = \\operatorname {M e a n} _ {\\text {t w i c e}} \\left(\\mathrm {G} _ {\\text {c r o s s}} \\odot \\mathrm {A} _ {\\text {c r o s s}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.906, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathrm{P}_{init}\\) is the initially generated multimodal prompt, \\(\\odot\\) denotes the element-wise multiplication operation, and \\(\\mathrm{Mean}_{\\text{twice}}(\\cdot)\\) performs two averaging steps, which average the influence of different text sequences on the image under each attention head, thereby obtaining the final \\(\\mathrm{P}_{init}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.776, + 0.68 + ], + "angle": 0, + "content": "3.2. Adaptive Refinement Module" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.701, + 0.6, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.54, + 0.787, + 0.574, + 0.799 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.701, + 0.7, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.787, + 0.667, + 0.797 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.716, + 0.701, + 0.801, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.743, + 0.787, + 0.775, + 0.797 + ], + "angle": 0, + "content": "Initial" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.702, + 0.902, + 0.784 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.787, + 0.905, + 0.798 + ], + "angle": 0, + "content": "Refined by ARM" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Figure 3. Comparison between the initially generated prompt and the prompt refined using ARM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Fig. 3, the initially generated multimodal prompt by BLIP only highlights the local regions of the camouflaged object rather than the entire object. As a" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20675" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.319 + ], + "angle": 0, + "content": "result, the mask prompt generated based on this multimodal prompt is incomplete, potentially negatively impacting SAM's performance. This locality issue is caused by domain shifts. When BLIP performs an image-text retrieval task, recognizing only the local region is sufficient to output the final matching score. Particularly in our work, the input text \\( t_{noun} \\) consists of only a single noun representing the object category without any additional descriptive details. Therefore, the initially generated prompt can only roughly indicate the object's location but cannot fully capture its structure, resulting in a lower-quality mask prompt. To address this issue, we design an Adaptive Refinement Module (ARM) to refine the initial incomplete prompt, enabling it to cover the object more entirely and generate a higher-quality mask prompt." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.218, + 0.338 + ], + "angle": 0, + "content": "3.2.1. Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.389 + ], + "angle": 0, + "content": "We first design adapters to extract features from the entire architecture. These adapters contain only two learnable layers, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.397, + 0.483, + 0.415 + ], + "angle": 0, + "content": "\\[\n\\operatorname {A d a p t e r} (x) = \\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {1} (x)\\right)\\right)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.468 + ], + "angle": 0, + "content": "where \\( \\mathrm{Adapter}(\\cdot) \\) is an adapter module, \\( \\mathrm{Conv}_i(\\cdot) \\) is the \\( i^{th} \\) two-dimensional convolutional layer and \\( \\mathrm{GELU}(\\cdot) \\), the Gaussian Error Linear Unit [20], is an activation function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.514 + ], + "angle": 0, + "content": "In ARM, there are \\(k\\) adapters used to capture image features from the two distinct image encoders and fuse them to generate a more comprehensive representation:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.52, + 0.483, + 0.551 + ], + "angle": 0, + "content": "\\[\nE m b _ {A} ^ {k} = \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {B _ {i m g}}\\right) + \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {S _ {i m g}} ^ {k}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.616 + ], + "angle": 0, + "content": "where the \\(Emb_A^k\\) represents the output embedding of the \\(k^{th}\\) adapter, and \\(Emb_{S_{img}}^k\\) is the image embedding comes from \\(k^{th}\\) transformer layer of SAM's image encoder. All adapters share the same weights." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.616, + 0.484, + 0.692 + ], + "angle": 0, + "content": "To better utilize the extracted image features, the output of the \\(k^{th}\\) adapter is combined with that of the \\(k^{th}\\) transformer layer in SAM's image encoder and fed into the \\((k + 1)^{th}\\) transformer layer, which could help fine-tune the SAM's image encoder:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.7, + 0.483, + 0.721 + ], + "angle": 0, + "content": "\\[\nE m b _ {S _ {i m g - i n}} ^ {k + 1} = E m b _ {A} ^ {k} + E m b _ {S _ {i m g - o u t}} ^ {k}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.484, + 0.796 + ], + "angle": 0, + "content": "where \\(Emb_{S_{img - in}}^{k + 1}\\) denotes to the inputs of the \\((k + 1)^{th}\\) transformer layer in SAM's image encoder, and \\(Emb_{S_{img - out}}^k\\) is the original output from \\(k^{th}\\) transformer layer in SAM's image encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Then, inspired by CLIP-ES [32], we extract the attention weight \\( A_{self} \\) from \\( \\mathrm{E}_{B_{img}} \\) instead of training an additional affinity network to refine the initial prompt. By leveraging the relationships between image patches encoded in the attention weights, the prompt can adjust itself to identify high-correlation regions and expand to cover a more complete representation of the object. However, considering the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.181 + ], + "angle": 0, + "content": "complexity of the COD, the background introduces significant noise, which can distort the attention map to capture the relationships between image patches accurately. Therefore, ARM uses the output of the final adapter, which has absorbed all the image information, to adjust the attention weights." + }, + { + "type": "equation", + "bbox": [ + 0.618, + 0.182, + 0.906, + 0.2 + ], + "angle": 0, + "content": "\\[\nA _ {s e l f} ^ {*} = A _ {s e l f} \\odot E m b _ {A} ^ {l a s t}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.203, + 0.906, + 0.25 + ], + "angle": 0, + "content": "where \\( A_{\\text{self}}^{*} \\) denotes the refined attention weights, and \\( Emb_{A}^{\\text{last}} \\) represents the image embedding extracted from the final adapter module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.251, + 0.906, + 0.296 + ], + "angle": 0, + "content": "Subsequently, the masked attention weights are generated as affinity weights to refine the initial multimodal prompt:" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.305, + 0.906, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\mathrm {P} _ {\\text {r e f i n e}} = \\left(A _ {\\text {s e l f}} ^ {*} \\odot \\operatorname {v e c} (M)\\right) \\otimes \\mathrm {P} _ {\\text {i n i t}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.906, + 0.422 + ], + "angle": 0, + "content": "where \\(vec(M)\\) is the vectorization of mask \\(M\\), which are generated by extracting highlighted regions of the \\(\\mathbf{P}_{init}\\), and \\(\\otimes\\) is the matrix multiplication. Finally, by setting a threshold, the highlighted target region from the \\(\\mathbf{P}_{refine}\\) is extracted to generate a pseudo-mask, which serves as the final mask prompt \\(\\mathbf{P}_{mask}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.429, + 0.766, + 0.445 + ], + "angle": 0, + "content": "3.3. Embedding Inputs for SAM" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.906, + 0.495 + ], + "angle": 0, + "content": "After the mask prompt is generated, we will provide three inputs to SAM's mask decoder to support it in completing the COD task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.497, + 0.906, + 0.542 + ], + "angle": 0, + "content": "First, the primary generated mask prompt is encoded by SAM's prompt encoder \\( E_{S_{prompt}} \\), ultimately producing a dense embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.604, + 0.551, + 0.905, + 0.568 + ], + "angle": 0, + "content": "\\[\nE m b _ {d e n s e} = \\mathrm {E} _ {S _ {p r o m p t}} \\left(\\mathrm {P} _ {m a s k}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.576, + 0.906, + 0.637 + ], + "angle": 0, + "content": "Second, the image embedding produced by the final transformer layer of SAM's image encoder is combined with the visual features from the last adapter to generate the final image embedding:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.644, + 0.905, + 0.665 + ], + "angle": 0, + "content": "\\[\nE m b _ {S _ {i m g}} ^ {\\text {o u t}} = E m b _ {A} ^ {\\text {l a s t}} + E m b _ {S _ {i m g}} ^ {\\text {l a s t}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.906, + 0.805 + ], + "angle": 0, + "content": "Additionally, to make the most of the generated multiple information, we encode \\( t_{noun} \\) into the text embedding \\( Emb_{M_{text}} \\) using Mamba [11], a sequence modeling architecture known for its advanced ability to process textual information. The generated \\( Emb_{M_{text}} \\) clearly conveys the semantic information of the camouflaged target category and is concatenated with the \\( Emb_{B_{img}} \\) extracted from BLIP to generate the auxiliary sparse embedding to provide richer features:" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.816, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\nE m b _ {A u x} = \\operatorname {C o n c a t} \\left(E m b _ {M _ {t e x t}}, E m b _ {B _ {i m g}}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.84, + 0.856 + ], + "angle": 0, + "content": "where the \\(\\mathrm{Concat}(\\cdot)\\) is the concatenate operation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Finally, \\(Emb_{dense}\\), along with the \\(Emb_{Simg}^{out}\\) and \\(Emb_{Aux}\\), are input into SAM's mask decoder to obtain the final segmentation result." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20676" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.088, + 0.905, + 0.309 + ], + "angle": 0, + "content": "
MethodPub. YearCHAMELEONCAMOCOD10KNC4K
SmFβmaxEφMSmFβmaxEφMSmFβmaxEφMSmFβmaxEφM
ZoomNet [42]CVPR220.9010.8760.9470.0230.8200.8050.8890.0660.8370.7770.8960.0290.8520.8260.9030.044
SegMaR [25]CVPR220.9060.8880.9590.0250.8160.8030.8840.0710.8330.7740.9060.0340.8410.8260.9070.046
UEDG [39]TMM230.9110.8940.9680.0230.8630.8560.9290.0480.8580.8120.9340.0250.8790.8640.9350.035
TinyCOD [58]ICASSP230.8870.8610.9580.0300.8220.8070.8990.0660.8110.7420.9030.0360.8430.8170.9100.047
MSCAF-Net [36]TCSVT230.9120.9020.9700.0220.8730.8670.9370.0460.8650.8230.9360.0240.8870.8740.9420.032
FSPNet [24]CVPR230.9090.8900.9650.0230.8560.8460.9280.0500.8510.7940.9300.0260.8790.8590.9370.035
FEDER [16]CVPR230.8870.8680.9540.0300.8020.7890.8730.0710.8220.7680.9050.0320.8470.8330.9150.044
FSNet [49]TIP230.9050.8910.9750.0220.8800.8780.9410.0410.8700.8330.9480.0230.8910.8800.9480.031
CRNet [19]AAAI230.8180.7920.9090.0460.7350.7070.8300.0920.7330.6360.8450.049————
SDRNet [13]KBS240.9140.9010.9610.0240.8720.8670.9320.0490.8710.8280.9360.0230.8890.8760.9400.032
PRNet [23]TCSVT240.9150.9020.9730.0200.8720.8670.9300.0500.8730.8390.9430.0220.8910.8810.9420.031
CamoFormer-R [59]TPAMI240.8980.8800.9490.0260.8170.8010.8830.0680.8380.7860.9280.0290.8540.8290.9080.042
CamoFormer-P [59]TPAMI240.9100.8980.9660.0220.8720.8680.9380.0460.8690.8290.9390.0230.8920.8800.9460.030
ZoomNeXtpv2b3 [43]TPAMI240.9280.9190.9770.0170.8850.8860.9420.0420.8950.8640.9510.0180.9000.8910.9480.028
ZoomNeXtpv2b2 [43]TPAMI240.9220.9080.9690.0170.8740.8730.9310.0470.8870.8560.9450.0190.8920.8840.9410.030
DSAM [60]ACMM24————0.8320.8340.9200.0610.8450.8050.9300.0330.8720.8640.9420.040
MAMIFNet [55]IF250.9140.8990.9590.0210.8720.8700.9350.0450.8690.8260.9400.0230.8900.8780.9430.031
COMPrompter [64]SCIS250.8850.8640.9570.0300.8530.8560.9310.0540.8600.8260.9460.0270.8800.8760.9460.036
Ours—0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.0200.9060.8970.9470.033
" + }, + { + "type": "image_caption", + "bbox": [ + 0.11, + 0.318, + 0.885, + 0.334 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison with other SOTA methods across different datasets. The best results are highlighted in bold." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.347, + 0.335, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.347, + 0.616, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.347, + 0.9, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.504, + 0.341, + 0.653 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.377, + 0.505, + 0.622, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.659, + 0.505, + 0.902, + 0.652 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.258, + 0.664, + 0.738, + 0.679 + ], + "angle": 0, + "content": "Figure 4. \\( F_{\\beta} \\) and PR curves of our work compared to other SOTA methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.704, + 0.222, + 0.721 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.281, + 0.753 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Datasets. To verify the effectiveness of the proposed architecture in the COD task, we use four commonly used COD datasets for evaluation: CHAMELEON [48], CAMO [28], COD10K [10], and NC4K [38]. In detail, CHAMELEON consists of 76 images featuring highly camouflaged animals in natural environments, all used for testing. CAMO contains 1250 images, including both biological camouflage and artificially designed camouflage. Among them, 1000 images are used for training, while the rest are reserved" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.81 + ], + "angle": 0, + "content": "for testing. COD10K is more extensive, comprising 5066 images that include both artificial and natural camouflage, with 3040 images allocated for training and the remainder used for testing. While NC4K comprises 4,121 testing images specifically designed for the COD task. Notably, all four datasets provide only image data and do not include additional modalities, such as text." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.904 + ], + "angle": 0, + "content": "Evaluation Metrics. In the subsequent experiments, we mainly use four metrics for evaluation: S-measure [8] \\((Sm)\\), maximum F-measure [1] \\((F_{\\beta}^{max})\\), maximum E-measure [9] \\((E_{\\varphi}^{max})\\), and Mean Absolute Error \\((M)\\). Among these, higher values of the \\(Sm\\), \\(F_{\\beta}^{max}\\), and \\(E_{\\varphi}^{max}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20677" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.18, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.089, + 0.27, + 0.155 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.089, + 0.361, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.089, + 0.451, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.089, + 0.541, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.09, + 0.632, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.09, + 0.722, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.09, + 0.813, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.09, + 0.904, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.159, + 0.18, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.159, + 0.27, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.159, + 0.361, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.159, + 0.451, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.159, + 0.541, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.159, + 0.632, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.159, + 0.722, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.159, + 0.813, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.159, + 0.904, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.231, + 0.18, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.231, + 0.27, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.231, + 0.361, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.231, + 0.451, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.231, + 0.541, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.231, + 0.632, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.231, + 0.722, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.231, + 0.813, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.231, + 0.904, + 0.297 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.3, + 0.18, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.368, + 0.153, + 0.379 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.184, + 0.301, + 0.27, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.219, + 0.368, + 0.237, + 0.378 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.276, + 0.301, + 0.361, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.304, + 0.368, + 0.329, + 0.378 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.301, + 0.451, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.368, + 0.445, + 0.379 + ], + "angle": 0, + "content": "COMPrompter" + }, + { + "type": "image", + "bbox": [ + 0.456, + 0.301, + 0.541, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.472, + 0.368, + 0.526, + 0.378 + ], + "angle": 0, + "content": "ZoomNeXt" + }, + { + "type": "image", + "bbox": [ + 0.548, + 0.301, + 0.632, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.368, + 0.602, + 0.378 + ], + "angle": 0, + "content": "PRNet" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.301, + 0.722, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.658, + 0.368, + 0.69, + 0.378 + ], + "angle": 0, + "content": "FSNet" + }, + { + "type": "image", + "bbox": [ + 0.728, + 0.301, + 0.813, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.747, + 0.368, + 0.785, + 0.378 + ], + "angle": 0, + "content": "FSPNet" + }, + { + "type": "image", + "bbox": [ + 0.82, + 0.301, + 0.904, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.837, + 0.368, + 0.884, + 0.378 + ], + "angle": 0, + "content": "ZoomNet" + }, + { + "type": "image_caption", + "bbox": [ + 0.184, + 0.39, + 0.812, + 0.405 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of segmentation results with other SOTA methods on different samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.493 + ], + "angle": 0, + "content": "indicate better model performance, whereas for \\(M\\), lower values indicate better performance. Additionally, the precision-recall curve (PR curve) and the \\(F_{\\beta}\\)-threshold curve \\((F_{\\beta}\\) curve) are also considered for further evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.499, + 0.484, + 0.667 + ], + "angle": 0, + "content": "Implement Details. Our proposed framework is primarily built upon the SAM-Vit Huge (ViT-H) [27] and implemented using PyTorch. The entire architecture is trained on an NVIDIA GeForce RTX 4090 GPU for 40 epochs with a batch size of 1. The main learning rate is set to 0.0002, while the learning rate for the ARM is 0.1 times that of the main learning rate, and the AdamW optimizer is used for optimization. Furthermore, both the VLM and LLM, used for multimodal prompt and auxiliary embedding generation, are pre-trained models, so no additional training or fine-tuning is required for these two." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.675, + 0.467, + 0.692 + ], + "angle": 0, + "content": "4.2. Comparisons With State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.697, + 0.484, + 0.743 + ], + "angle": 0, + "content": "To effectively evaluate the performance of our work in the COD task, we first conducted a comparative analysis with the current state-of-the-art (SOTA) models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Evaluation. As shown in Tab. 1, which compares the evaluation results of our work and other SOTA models across four different datasets. Our work generally demonstrates exceptional performance in solving the COD task across various metrics on different datasets. In particular, our method excels in both the \\( Sm \\) and \\( F_{\\beta}^{max} \\), showing superior performance in target localization and structure capture of the camouflaged objects in complex scenarios. For example, on the COD10K, our approach outperforms the second-best method by \\( 1.6\\% \\) in \\( Sm \\) and \\( 2.4\\% \\) in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.537 + ], + "angle": 0, + "content": "\\(F_{\\beta}^{max}\\). Although our \\(F_{\\beta}^{max}\\) is the second on the CAMO, the gap with the top-performing model is only \\(0.3\\%\\), which is negligible. Furthermore, from the comparison of the \\(F_{\\beta}\\) and PR curves in Fig. 4, it is clear that our method (red solid curve) exhibits outstanding performance across all three datasets, with particularly remarkable results on the COD10K dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Qualitative Evaluation Fig. 5 illustrates a comparison of the final segmentation maps of our work and other SOTA methods across different samples. It is clear that our approach achieves a more significant overlap with the ground truth (GT) than other SOTA models. Specifically, our method is more effective at accurately segmenting and capturing completely camouflaged objects. Even in more challenging scenarios, such as the third sample in Fig. 5, where multiple small camouflaged objects need to be identified, our approach segments more objects and provides more accurate segmentation, while the results from other models show noticeable omissions. These observations highlight the superior performance of our framework, not only in localizing camouflaged objects but also in maintaining the structural integrity of the camouflaged objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.781, + 0.681, + 0.795 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.803, + 0.907, + 0.849 + ], + "angle": 0, + "content": "As shown in Tab. 2, we conduct ablation studies on the improvements discussed in Sec. 3 to assess the effectiveness of each component." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Auxiliary Embedding. Since the auxiliary embedding is generated alongside the multimodal data produced by BLIP, we first incorporate it into the SAM to" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20678" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.904, + 0.189 + ], + "angle": 0, + "content": "
CHAMELEONCAMOCOD10K
SmFβmaxEφmaxMSmFβmaxEφmaxMSmFβmaxEφmaxM
SAM0.8500.8200.9110.0560.7970.7640.8640.1010.8100.7500.8810.054
SAM + EmbAUX0.8520.8160.9220.0470.8050.7780.8820.0890.8200.7510.8930.039
SAM + EmbAUX + Pcaptioninit0.8660.8360.9200.0460.8230.7950.8910.0810.8390.7760.9080.037
SAM + EmbAUX + Pnouninit0.8820.8590.9350.0400.8310.8070.8960.0800.8460.7890.9140.036
SAM + EmbAUX + PnounARM0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.020
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.199, + 0.908, + 0.242 + ], + "angle": 0, + "content": "Table 2. Ablation studies on three different COD datasets. + denotes additional information provided to SAM, \\(Emb_{Aux}\\) is the auxiliary sparse embedding, \\(P_{init}^{caption}\\) is the initial mask prompt generated by the entire caption, \\(P_{init}^{noun}\\) is the initial mask prompt generated by the extracted noun, \\(P_{ARM}^{noun}\\) is the refined \\(P_{init}^{noun}\\) by ARM." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.266, + 0.189, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.325, + 0.188, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.383, + 0.188, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.431, + 0.188, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.48, + 0.188, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.536, + 0.157, + 0.546 + ], + "angle": 0, + "content": "Image" + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.267, + 0.286, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.325, + 0.286, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.383, + 0.286, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.193, + 0.431, + 0.286, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.192, + 0.48, + 0.286, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.231, + 0.536, + 0.249, + 0.545 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.267, + 0.383, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.325, + 0.383, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.383, + 0.383, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.431, + 0.383, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.48, + 0.383, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.323, + 0.536, + 0.351, + 0.544 + ], + "angle": 0, + "content": "Initial" + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.267, + 0.481, + 0.322 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.325, + 0.481, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.383, + 0.481, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.431, + 0.481, + 0.476 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.387, + 0.48, + 0.481, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.394, + 0.536, + 0.471, + 0.545 + ], + "angle": 0, + "content": "Refined by ARM" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.558, + 0.483, + 0.587 + ], + "angle": 0, + "content": "Figure 6. Comparison between the initially generated prompt and the prompt refined using ARM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.617, + 0.484, + 0.707 + ], + "angle": 0, + "content": "evaluate its effectiveness. Comparing the \\(2^{nd}\\) row with the \\(1^{st}\\) row in Tab. 2, adding the auxiliary embedding leads to performance improvements across nearly all metrics on the three COD datasets. This indicates that the auxiliary embedding effectively enhances SAM by providing additional multimodal features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Multimodal Prompt. Our primary contribution is utilizing BLIP to generate multimodal prompts, providing SAM with more informative cues to handle the COD task. Experimental results (\\(3^{rd}\\) row of Tab. 2) show that prompts generated from full description already outperform SAM with auxiliary embedding. Furthermore, SAM's performance further improved when extracting the noun about the object category from the long description as textual information to generate prompts (\\(4^{th}\\) row of Tab. 2). This result not only proves the effectiveness of our generated prompts but also indicates that, although BLIP can produce descriptions closely related to the image, redun-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.906, + 0.33 + ], + "angle": 0, + "content": "dant information introduces noise, preventing the generated prompt from accurately localizing the targets. In contrast, providing only category information helps generate more precise mask prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.35, + 0.907, + 0.606 + ], + "angle": 0, + "content": "Effectiveness of Refinement. As shown in the \\(5^{th}\\) row results in Tab. 2, our proposed ARM significantly enhances SAM's performance on the COD task, achieving substantial improvements across all evaluation metrics on three datasets. This performance boost can be attributed to two key aspects: First, ARM refines the multimodal prompts initially generated by BLIP, as illustrated in Fig. 6. Due to domain shifts, these initial prompts often localize only partial regions of the target. However, after refining with ARM, the refined prompts achieve more complete target coverage. Consequently, the mask prompts generated from these refined prompts provide SAM with more precise cues. Second, ARM effectively integrates the image features extracted by the adapter into SAM, enabling fine-tuning of the image encoder. This process further enhances SAM's representation capability, leading to improved performance on the COD task." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.642, + 0.634, + 0.658 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel and cost-effective solution for the COD task. The main idea is to incorporate multimodal information to construct high-quality mask prompts, enhancing SAM's segmentation capability. Specifically, we use VLM to automatically generate multimodal data and prompts, laying a solid foundation for constructing the final high-quality mask prompts. Furthermore, we propose the Adaptive Refinement Module (ARM) to refine the initial multimodal prompts enabling more complete target representations and fine-tune SAM's image encoder, providing it with richer information. Additionally, experimental results demonstrate that our proposed architecture achieves outstanding performance across multiple COD datasets, particularly excelling in capturing the structural details of objects, thereby further validating our work's effectiveness." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "20679" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.258, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.116, + 0.482, + 0.145 + ], + "angle": 0, + "content": "The work is supported by the XJTLU RDF project: RDF-22-02-118 & RDF-24-02-041" + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.151, + 0.188, + 0.167 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.176, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.233, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.29, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[3] Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. Sam-cod+: Sam-guided unified framework for weakly-supervised camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.347, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[4] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.404, + 0.483, + 0.487 + ], + "angle": 0, + "content": "[5] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?-sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2304.09148, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.483, + 0.53 + ], + "angle": 0, + "content": "[6] Zhennan Chen, Rongrong Gao, Tian-Zhu Xiang, and Fan Lin. Diffusion model for camouflaged object detection. In ECAI 2023, pages 445-452. IOS Press, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.532, + 0.483, + 0.587 + ], + "angle": 0, + "content": "[7] Zhongxi Chen, Ke Sun, and Xianming Lin. Camodiffusion: Camouflaged object detection via conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1280, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.646, + 0.483, + 0.7 + ], + "angle": 0, + "content": "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[12] Juwei Guan, Xiaolin Fang, Tongxin Zhu, Zhipeng Cai, Zhen Ling, Ming Yang, and Junzhou Luo. Idenet: Making neural network identify camouflaged objects like creatures. IEEE Transactions on Image Processing, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Juwei Guan, Xiaolin Fang, Tongxin Zhu, and Weiqi Qian. Sdrnet: Camouflaged object detection with independent re" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.176, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "construction of structure and detail. Knowledge-Based Systems, 299:112051, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[14] Juwei Guan, Weiqi Qian, Tongxin Zhu, and Xiaolin Fang. Promoting camouflaged object detection through novel edge-target interaction and frequency-spatial fusion. Neurocomputing, 617:129064, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[15] Cunhan Guo and Heyan Huang. Cofinet: Unveiling camouflaged objects with multi-scale finesse. Neurocomputing, 614:128763, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22046-22055, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[17] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36:30726-30737, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[18] Jiangping He, Yangjie Zhang, Maggie Chung, Michael Wang, Kun Wang, Yan Ma, Xiaoyang Ding, Qiang Li, and Yonglin Pu. Whole-body tumor segmentation from pet/ct images using a two-stage cascaded neural network with camouflaged object detection mechanisms. Medical Physics, 50 (10):6151-6162, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.476, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI conference on artificial intelligence, pages 781-789, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.534, + 0.905, + 0.561 + ], + "angle": 0, + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[21] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.606, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[22] Bintao Hu, Hengyan Liu, Jianbo Du, Miguel López-Benitez, Celimuge Wu, Xiaoli Chu, and Dusit Niyato. Madqnenhanced computation offloading and resource allocation for 6g low-altitude economy vehicular networks. IEEE Transactions on Cognitive Communications and Networking, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[23] Xihang Hu, Xiaoli Zhang, Fasheng Wang, Jing Sun, and Fuming Sun. Efficient camouflaged object detection network based on global localization perception and local guidance refinement. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):5452-5465, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.761, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[24] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5557-5566, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[25] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4713-4722, 2022. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20680" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[26] Xinhao Jiang, Wei Cai, Yao Ding, Xin Wang, Danfeng Hong, Xingyu Di, and Weijie Gao. Bidicos: Camouflaged object segmentation via bilateral diffusion model. Expert Systems with Applications, 255:124747, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[28] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and image understanding, 184:45-56, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.343 + ], + "angle": 0, + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[31] Lin Li, Bo Dong, Eric Rigall, Tao Zhou, Junyu Dong, and Geng Chen. Marine animal segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(4): 2303-2314, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[32] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[33] Hengyan Liu, Guangyu Ren, Tianhong Dai, Di Zhang, Pengjing Xu, Wenzhang Zhang, and Bintao Hu. Diversity-oriented contrastive learning for rgb-t scene parsing. In 2023 2nd International Conference on Sensing, Measurement, Communication and Internet of Things Technologies (SMC-IoT), pages 155-160. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[34] Hengyan Liu, Wenzhang Zhang, Tianhong Dai, Longfei Yin, and Guangyu Ren. Multimodal frequency spectrum fusion schema for rgb-t image semantic segmentation. In 2024 33rd International Conference on Computer Communications and Networks (ICCNN), pages 1-6. IEEE, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[35] Limei Liu, Meng Liu, Kexin Meng, Linqiang Yang, Min Zhao, and Shuli Mei. Camouflaged locust segmentation based on pranet. Computers and Electronics in Agriculture, 198:107061, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[36] Yu Liu, Haihang Li, Juan Cheng, and Xun Chen. Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features. IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4934-4947, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Yantong Liu, Sai Che, Liwei Ai, Chuanxiang Song, Zheyu Zhang, Yongkang Zhou, Xiao Yang, and Chen Xian. Camouflage detection: Optimization-based computer vision for alligator sinensis with low detectability in complex wild environments. Ecological Informatics, 83:102802, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[38] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11591-11601, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[39] Yixuan Lyu, Hong Zhang, Yan Li, Hanyang Liu, Yifan Yang, and Ding Yuan. Uedg: Uncertainty-edge dual guided camouflage object detection. IEEE Transactions on Multimedia, 26:4050-4060, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[40] Tsui Qin Mok, Shuyong Gao, Haozhe Xing, Miaoyang He, Yan Wang, and Wenqiang Zhang. A holistically point-guided text framework for weakly-supervised camouflaged object detection. arXiv preprint arXiv:2501.06038, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[41] Yuzhen Niu, Lifen Yang, Rui Xu, Yuezhou Li, and Yuzhong Chen. Minet: Weakly-supervised camouflaged object detection through mutual interaction between region and edge cues. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6316-6325, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.907, + 0.417 + ], + "angle": 0, + "content": "[42] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2160-2170, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[43] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE transactions on pattern analysis and machine intelligence, 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[44] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. arXiv preprint arXiv:2311.10121, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[45] Guangyu Ren, Yinxiao Yu, Hengyan Liu, and Tania Stathaki. Dynamic knowledge distillation with noise elimination for rgb-d salient object detection. Sensors, 22(16):6188, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[47] Yiqing Shen, David Dreizin, Blanca Inigo, and Mathias Unberath. Protosam-3d: Interactive semantic segmentation in volumetric medical imaging via a segment anything model and mask-level prototypes. Computerized Medical Imaging and Graphics, page 102501, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.731, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[48] Przemysław Skurowski, Hassan Abdulameer, Jakub Blaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[49] Ze Song, Xudong Kang, Xiaohui Wei, Haibo Liu, Renwei Dian, and Shutao Li. Fsnet: Focus scanning network for camouflaged object detection. IEEE Transactions on Image Processing, 32:2267-2278, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[50] Yanguang Sun, Chunyan Xu, Jian Yang, Hanyu Xuan, and Lei Luo. Frequency-spatial entanglement learning for camouflaged object detection. In European Conference on Computer Vision, pages 343-360. Springer, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20681" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[51] Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8805–8814, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[52] Huafeng Wang, Tianyu Hu, Yanan Zhang, Haodu Zhang, Yong Qi, Longzhen Wang, Jianhua Ma, and Minghua Du. Unveiling camouflaged and partially occluded colorectal polyps: Introducing cpsnet for accurate colon polyp segmentation. Computers in Biology and Medicine, 171:108186, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.262, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[53] Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. Depth-aware concealed crop detection in dense agricultural scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17201-17211, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[54] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3297–3306, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[55] Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection. Information Fusion, 117:102871, 2025. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.531 + ], + "angle": 0, + "content": "[56] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1032-1042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.533, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[57] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv preprint arXiv:2106.13217, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.659 + ], + "angle": 0, + "content": "[58] Haozhe Xing, Shuyong Gao, Hao Tang, Tsui Qin Mok, Yanlan Kang, and Wenqiang Zhang. Tinycod: Tiny and effective model for camouflaged object detection. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[59] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[60] Zhenni Yu, Xiaogin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. Exploring deeper! segment anything model with depth perception for camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4322-4330, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[61] Hong Zhang, Yixuan Lyu, Qian Yu, Hanyang Liu, Huimin Ma, Ding Yuan, and Yifan Yang. Unlocking attributes' contribution to successful camouflage: A combined textual and visual analysis strategy. In European Conference on Computer Vision, pages 315-331. Springer, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[62] Jie Zhang, Xubing Yang, Rui Jiang, Wei Shao, and Li Zhang. Rsam-seg: A sam-based approach with prior knowledge in" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tegration for remote sensing image semantic segmentation. arXiv preprint arXiv:2402.19004, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[63] Shizhou Zhang, Dexuan Kong, Yinghui Xing, Yue Lu, Lingyan Ran, Guoqiang Liang, Hexu Wang, and Yanning Zhang. Frequency-guided spatial adaptation for camouflaged object detection. IEEE Transactions on Multimedia, 27:72-83, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[64] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[65] Ke Zhou, Zhongwei Qiu, and Dongmei Fu. Multi-scale contrastive adaptor learning for segmenting anything in under-performed scenes. Neurocomputing, 606:128395, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20682" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_origin.pdf b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c9df1c28af2fb9bb975d38b2e5c97a1d81b92068 --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/31897efc-f494-43de-9990-b3b0b058743c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7943321ff74702666fa2d4a891350201fcf443ccab526229c0d7cb0261f2e6d6 +size 5396066 diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/full.md b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f586d69b2c0662e97d0f68ce34218d8025925d7c --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/full.md @@ -0,0 +1,469 @@ +# Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection + +Xuehan Chen $^{1}$ Guangyu Ren $^{1*}$ Tianhong Dai $^{2}$ Tania Stathaki $^{2}$ Hengyan Liu $^{1*}$ + +$^{1}$ Xi'an Jiaotong-Liverpool University, China + +$^{2}$ Imperial College London, United Kingdom + +xuehanchen83@gmail.com, {guangyu. ren, hengyan.liu}@xjtlu.edu.cn + +tianhong.dai15@alumni.imperial.ac.uk, t.stathaki@imperial.ac.uk + +# Abstract + +Foundation models, such as Segment Anything Model (SAM), have exhibited remarkable performance in conventional segmentation tasks, primarily due to their training on large-scale datasets. Nonetheless, challenges remain in specific downstream tasks, such as Camouflaged Object Detection (COD). Existing research primarily aims to enhance performance by integrating additional multimodal information derived from other foundation models. However, directly leveraging the information generated by these models may introduce additional biases due to domain shifts. To address this issue, we propose an Adaptive Refinement Module (ARM), which efficiently processes multimodal information and simultaneously refining the mask prompt. Furthermore, we construct an auxiliary embedding that effectively exploits the intermediate information generated during ARM, providing SAM with richer feature representations. Experimental results indicate that our proposed architecture surpasses most state-of-the-art (SOTA) models in the COD task, particularly excelling in structured target segmentation. + +# 1. Introduction + +Camouflaged Object Detection (COD) is a challenging computer vision task, as it involves segmenting objects that are virtually identical to their immediate surroundings. It plays a crucial role in the field due to its wide range of practical applications, including object detection [34, 45], low-altitude economy [22], zoological research [31, 35, 37], and medical image segmentation [18, 52]. + +Current research [12, 14, 33, 41, 43] focuses on designing advanced architectures to extract and refine more detailed features exclusively from image data. However, the + +![](images/50d4f60705a2b5a90993390a56097d82f8c3bd792a4f80336e65cd635d0f2f55.jpg) + +![](images/adc2cb2ca55e642b64beec235145ca67f12d5ecd5a73506f5d0ac261735ccba6.jpg) +(c) Ours +Figure 1. Comparison of our proposed Adaptive Refinement Module (ARM) with common COD frameworks. Unlike conventional methods that rely solely on images or manually annotated additional modalities, our approach not only leverages vision-language models (VLMs) to automatically generate multimodal data but also mitigates domain shift through the proposed ARM, providing high-quality prompts for SAM's decoder. Meanwhile, ARM can also fine-tune SAM's image encoder, further improving SAM's feature representation capability. + +unique nature of the COD task, where the targets are visually similar to their surroundings, limits the effectiveness of image features. Additionally, the lack of data from other modalities further limits the models' ability to acquire meaningful features, making them highly susceptible to background, preventing them from accurately detecting camouflaged objects. Although some studies incorporate human-made auxiliary data, such as depth images [53], textual descriptions [61], or prompts [2, 3], to provide supple + +mentary information about targets, they are not ideal solutions as they are time-consuming and labor-intensive. As a result, despite significant progress in the COD task, two limitations remain: limited image features and laborious annotation processes. + +Recently, foundation models have emerged as a popular solution for tackling a range of tasks, including the Segment Anything Model (SAM) [27], which has proven highly effective in handling various tasks, such as medical image segmentation [44, 47] and semantic segmentation [4, 62]. Although SAM demonstrates impressive zero-shot generalization capabilities, Chen et al. [5] states that it still struggles to segment the complex camouflaged objects. Secondly, SAM can guide its segmentation by receiving prompts, and the most intuitive approach is to generate prompts through manual annotation, but as mentioned earlier, it is less practical due to the time and labor it demands. + +To mitigate the aforementioned limitations, we rethink the core issues that make the COD task challenging. In this task, there are no task-specific clues (e.g., category, size, location) of the targets. Meanwhile, the background's visual noise increases the task's complexity, resulting in ambiguous object boundaries and weakened feature learning. So, we propose a novel approach that uses textual information to complement visual features, enhancing the model's ability to localize target objects more accurately. In this work, we fully utilize the powerful pretrained vision-language models (VLMs) to automatically generate multimodal data without requiring additional training or fine-tuning. This data provides additional cues to assist in object localization. Subsequently, the VLMs further utilize the generated multimodal information to produce mask prompts that convey the object's approximate location to SAM. + +We also consider the biases introduced by directly utilizing generated multimodal information, which could lead to locality in the prompts, making them less comprehensive and potentially affecting the subsequent SAM segmentation performance. To mitigate this issue, this paper proposes an Adaptive Refinement Module (ARM), which incorporates simple adapters to refine the initially generated mask prompts. This module expands the regions in the initial prompts to create a more complete representation of the target objects, offering better guidance for downstream models. Furthermore, ARM effectively extracts vision information from the entire architecture, serving not only as a refinement mechanism for prompts but also as a fine-tuning module for SAM's image encoder, further enhancing SAM's performance. + +In summary, this work proposes the ARM, which simultaneously refines the initial prompt generated by multimodal information and fine-tunes SAM's image encoder, enabling SAM to handle the COD task more effectively across multiple aspects. Our main contributions are as fol + +lows: + +- We propose a cost-effective approach that automatically constructs mask prompts using multimodal information generated by BLIP, serving as additional cues without requiring extra training or fine-tuning. These prompts provide reliable guidance for SAM, thereby improving its segmentation performance. +- To mitigate the additional biases introduced by BLIP, we propose an Adaptive Refinement Module (ARM) to refine the initially generated prompts, enhancing their quality and providing more precise guidance for SAM. Meanwhile, this module can also fine-tune the SAM's image encoder, further enhancing SAM's segmentation accuracy and robustness. +- We further utilize the multimodal information generated by BLIP to construct an auxiliary embedding, which effectively exploits the intermediate information generated during the prompt generation process and provides SAM with richer features. + +# 2. Related Work + +# 2.1. Single-Modality Based COD + +Image data is one of the most widely used modalities to tackle the COD task. Therefore, most researchers are currently exploring various approaches to extract information from single-modality image data in order to perform this task more effectively. CoFiNet [15], ZoomNext [43], ZoomNet [42] focus on extracting and fusing multiscale features from images. Although upgrading the model's backbone improves the accuracy of identifying objects, it also increases the consumption of computational resources. + +Besides directly extracting features from the images, CamoDiffusion [7], BiDiCOS [26], and diffCOD [6] transform the COD task into denoising tasks, gradually segmenting camouflaged objects through the denoising process. Moreover, EFNet [14], FSEL [50], FGSA-Net [63], and MAM-FFC [55] use spatial and frequency information from images to achieve a more precise segmentation. However, in the COD task, the background in images introduces significant noise. Even with those novel methods relying solely on images' intrinsic features, they still struggle to recognize objects in complex scenarios. + +Furthermore, in addition to designing a new architecture, many works have also focused on improving the Segment Anything Model (SAM), taking advantage of its powerful zero-shot generalization ability and prompt strategy. SAM-COD [3] and WS-SAM [17], which combine sparse prompts with SAM to generate the final segmentation results, still require additional manual annotations for the sparse prompts, which remains time-consuming. MCA-SAM [65] and SAM-adapter [5] improve the model's performance in the COD task by injecting more image infor + +![](images/94c915325f32d32b65f2b0f77d5581f8541f6026b44182b8af77d6fa67eca345.jpg) +Figure 2. Overview of our proposed architecture. BLIP automatically generates multimodal data from the input image and uses it to create the initial multimodal prompts and the auxiliary embedding. The Adaptive Refinement Module (ARM) is then employed to address the locality issue of the initial prompts. The ARM uses multiple adapters to extract visual information from both BLIP and SAM, which not only refines the initial multimodal prompts by integrating attention weights, resulting in more accurate mask prompts, but also fine-tunes SAM's image encoder. Finally, the SAM decoder receives image embedding, dense mask prompt embedding, and auxiliary sparse embedding to tackle the COD task. + +mation into SAM's image encoder using the adapter. They differ from our work, where the adapter serves a broader purpose in ARM. It is primarily used to refine the generated mask prompts while injecting additional information into SAM's image encoder, enabling the SAM to perform the COD task more effectively. + +# 2.2. Multi-Modality Based COD + +With the rapid growth of multimodal learning, different modalities have been increasingly applied to address the COD problem. HPGT [40] use the textual information to assist in completing the COD task. However, they rely on manually annotating text data for the COD dataset, which is a highly time-consuming process. Although ACU-MEN [61] utilizes the GPT-4 Vision model to generate textual descriptions and attributions for images, it still requires more than 30 volunteers to manually verify the accuracy of these generated texts, which is also a labor-intensive task. In contrast to these approaches, MMCPF [51] takes a different direction by interacting with a Multimodal Large Language Model (MLLM). They query the model using text to obtain the location coordinates of camouflaged objects, which supports the subsequent segmentation task. While this method reduces the need for text data annotation, it still faces the challenge of designing precise text prompts for querying the + +MLLM. Ordinary prompts cannot let MLLM output the expected location coordinates of objects, which indicates that designing effective text prompts becomes an additional task that is time-consuming and labor-intensive as well. As a result, both visual and textual information have been employed to address the COD problem, but manual annotation and verification remain a limitation. + +In addition to textual information, depth features are utilized to provide valuable spatial information for models. Due to the fact that commonly used COD datasets only include RGB images without depth data, PopNet [56], Da-COD [54] and [57] employ depth estimation models to generate depth images from RGB-based COD datasets. Since the depth images are generated automatically by models, the final segmentation performance is highly reliant on the performance of those depth estimation methods. In contrast to the previous approaches, RISNet [53] uses a depth camera to manually create an RGB-D dataset for COD in agricultural scenes, which is also painstaking. + +# 3. Method + +We propose a novel architecture, as illustrated in Fig. 2, which enables the fully automated generation of mask prompts for SAM and designs an innovative Adaptive Refinement Module (ARM) to enhance the quality of the + +initially generated mask prompts. Specifically, a Vision-Language Model (VLM) first automatically generates textual data based on the input image. Then, by aligning multimodal information, the VLM generates the multimodal prompts to locate camouflaged objects locally. Subsequently, it is fed into the ARM for refinement, enabling prompts to capture more complete object structure information, thereby providing better guidance for SAM. During the prompt refinement process, ARM also fine-tunes SAM's image encoder. Meanwhile, the generated textual data is fully utilized, which is integrated with visual information to provide additional auxiliary embedding, further improving SAM's performance in identifying camouflaged objects. + +In the following subsections, we provide detailed explanations of each component of this architecture. First, we describe the multimodal prompt generation process in Sec. 3.1. Then we introduce the ARM in Sec. 3.2. Finally, Sec. 3.3 describes the different input embeddings fed into SAM's mask decoder to generate the final results. + +# 3.1. Multimodal Prompt Generation + +# 3.1.1. Multimodal Data Generation + +As a robust zero-shot VLM, BILP [30] offers powerful image captioning capabilities that align with our goal of automatically generating supplementary textual data for subsequent prompt generation. Given an image $I \in \mathbb{R}^{H \times W \times 3}$ , the image embedding $Emb_{B_{img}}$ are extracted by the BLIP's image encoder $E_{B_{img}}$ . This embedding is then passed to BLIP's image-grounded text decoder, which generates a caption, as presented below: + +$$ +T = \mathbf {D} _ {B _ {t e x t}} \left(E m b _ {B _ {i m g}}\right) = \left\{t _ {i j} \right\} _ {i = 1} ^ {L}, \tag {1} +$$ + +where $T$ denotes the generated caption, $\mathrm{D}_{B_{text}}$ is the BLIP's image-grounded text decoder, and $t_i$ represents the $i^{th}$ word in the caption, with a total of $L$ words. As exemplified by the potential caption generated in Fig. 2, it is clear that the captions generated by BLIP can correctly identify the camouflaged targets. + +However, since the generated $T$ provides an overall description of the input image, it may include content that is irrelevant to the target objects, such as "floor" is the background information of the description "There is a small dog that is lying on the floor". Due to the complexity of the COD task, irrelevant information in the generated caption could act as noise, potentially distracting from generating a prompt that accurately indicates the target's location. Therefore, instead of utilizing the entire caption $T$ , we use the spaCy [21] library for part-of-speech tagging to capture the necessary content about the objects. Specifically, each $t_i$ in $T$ is tagged with its corresponding part of speech (e.g., noun, verb, adjective). Through observation, we found that most generated captions primarily use nouns to describe the + +category of targets (e.g., man, bug, lion). Therefore, we extract the first noun $t_{noun}$ that appears in $T$ to provide more precise target category information, which helps the subsequent prompt generation more accurately locate the target object and reduces distractions caused by unrelated noise. + +# 3.1.2. Multimodal Prompt Generation + +BLIP has a powerful text-image alignment capability, which allows it to generate a multimodal prompt that can accurately locate the target object by aligning the text containing the target category information with the image. Inspired by [29], we use BLIP once again to handle the image-text retrieval task and generate the Gradient-weighted Class Activation Mapping (Grad-CAM) [46] to visualize the alignment process of BLIP with multimodal data, thereby generating the prompt. + +Specifically, the $Emb_{B_{img}}$ and $t_{noun}$ are fed into BLIP's image-grounded text encoder $\mathrm{E}_{B_{text}}$ , which then outputs a score reflecting how well the input image and text match, as follows: + +$$ +(S, \mathrm {A} _ {\text {c r o s s}}) = \mathrm {E} _ {B _ {\text {t e x t}}} \left(\left(E m b _ {B i m g}, t _ {\text {n o u n}}\right), \right. \tag {2} +$$ + +where $S$ denotes the output score, and $A_{cross}$ is the attention weight from the cross attention layer of $\mathbf{E}_{B_{text}}$ , which captures the relationship between the image and the text. + +To generate the Grad-CAM, the gradient is also important, so we conduct a backpropagation process to get the gradient $G_{cross}$ , and then generate the prompt, as presented below: + +$$ +\mathrm {P} _ {\text {i n i t}} = \operatorname {G r a d C A M} = \operatorname {M e a n} _ {\text {t w i c e}} \left(\mathrm {G} _ {\text {c r o s s}} \odot \mathrm {A} _ {\text {c r o s s}}\right), \tag {3} +$$ + +where $\mathrm{P}_{init}$ is the initially generated multimodal prompt, $\odot$ denotes the element-wise multiplication operation, and $\mathrm{Mean}_{\text{twice}}(\cdot)$ performs two averaging steps, which average the influence of different text sequences on the image under each attention head, thereby obtaining the final $\mathrm{P}_{init}$ . + +# 3.2. Adaptive Refinement Module + +![](images/e0f946b2ef87161dd20044448373e14a5f38b4373779f4b11028f6af8180e855.jpg) +Image + +![](images/4c7666dc81fcc9eab190ffb92d9bad235a29da0abbfa4eba51da7dc2e55f8902.jpg) +GT + +![](images/55c1714e7040f27fb6a07f99bd6bb57cbeefaf7474932e0ed15692599e1ac8d7.jpg) +Initial +Figure 3. Comparison between the initially generated prompt and the prompt refined using ARM. + +![](images/f6d202df113692182ce9599fac3445088ccb596b4c98756511065ecf06923f11.jpg) +Refined by ARM + +As shown in Fig. 3, the initially generated multimodal prompt by BLIP only highlights the local regions of the camouflaged object rather than the entire object. As a + +result, the mask prompt generated based on this multimodal prompt is incomplete, potentially negatively impacting SAM's performance. This locality issue is caused by domain shifts. When BLIP performs an image-text retrieval task, recognizing only the local region is sufficient to output the final matching score. Particularly in our work, the input text $t_{noun}$ consists of only a single noun representing the object category without any additional descriptive details. Therefore, the initially generated prompt can only roughly indicate the object's location but cannot fully capture its structure, resulting in a lower-quality mask prompt. To address this issue, we design an Adaptive Refinement Module (ARM) to refine the initial incomplete prompt, enabling it to cover the object more entirely and generate a higher-quality mask prompt. + +# 3.2.1. Refinement + +We first design adapters to extract features from the entire architecture. These adapters contain only two learnable layers, as shown below: + +$$ +\operatorname {A d a p t e r} (x) = \operatorname {G E L U} \left(\operatorname {C o n v} _ {2} \left(\operatorname {G E L U} \left(\operatorname {C o n v} _ {1} (x)\right)\right)\right), \tag {4} +$$ + +where $\mathrm{Adapter}(\cdot)$ is an adapter module, $\mathrm{Conv}_i(\cdot)$ is the $i^{th}$ two-dimensional convolutional layer and $\mathrm{GELU}(\cdot)$ , the Gaussian Error Linear Unit [20], is an activation function. + +In ARM, there are $k$ adapters used to capture image features from the two distinct image encoders and fuse them to generate a more comprehensive representation: + +$$ +E m b _ {A} ^ {k} = \operatorname {A d a p t e r} ^ {k} \left(E m b _ {B _ {i m g}}\right) + \operatorname {A d a p t e r} ^ {k} \left(E m b _ {S _ {i m g}} ^ {k}\right), \tag {5} +$$ + +where the $Emb_A^k$ represents the output embedding of the $k^{th}$ adapter, and $Emb_{S_{img}}^k$ is the image embedding comes from $k^{th}$ transformer layer of SAM's image encoder. All adapters share the same weights. + +To better utilize the extracted image features, the output of the $k^{th}$ adapter is combined with that of the $k^{th}$ transformer layer in SAM's image encoder and fed into the $(k + 1)^{th}$ transformer layer, which could help fine-tune the SAM's image encoder: + +$$ +E m b _ {S _ {i m g - i n}} ^ {k + 1} = E m b _ {A} ^ {k} + E m b _ {S _ {i m g - o u t}} ^ {k}, \tag {6} +$$ + +where $Emb_{S_{img - in}}^{k + 1}$ denotes to the inputs of the $(k + 1)^{th}$ transformer layer in SAM's image encoder, and $Emb_{S_{img - out}}^k$ is the original output from $k^{th}$ transformer layer in SAM's image encoder. + +Then, inspired by CLIP-ES [32], we extract the attention weight $A_{self}$ from $\mathrm{E}_{B_{img}}$ instead of training an additional affinity network to refine the initial prompt. By leveraging the relationships between image patches encoded in the attention weights, the prompt can adjust itself to identify high-correlation regions and expand to cover a more complete representation of the object. However, considering the + +complexity of the COD, the background introduces significant noise, which can distort the attention map to capture the relationships between image patches accurately. Therefore, ARM uses the output of the final adapter, which has absorbed all the image information, to adjust the attention weights. + +$$ +A _ {s e l f} ^ {*} = A _ {s e l f} \odot E m b _ {A} ^ {l a s t}, \tag {7} +$$ + +where $A_{\text{self}}^{*}$ denotes the refined attention weights, and $Emb_{A}^{\text{last}}$ represents the image embedding extracted from the final adapter module. + +Subsequently, the masked attention weights are generated as affinity weights to refine the initial multimodal prompt: + +$$ +\mathrm {P} _ {\text {r e f i n e}} = \left(A _ {\text {s e l f}} ^ {*} \odot \operatorname {v e c} (M)\right) \otimes \mathrm {P} _ {\text {i n i t}}, \tag {8} +$$ + +where $vec(M)$ is the vectorization of mask $M$ , which are generated by extracting highlighted regions of the $\mathbf{P}_{init}$ , and $\otimes$ is the matrix multiplication. Finally, by setting a threshold, the highlighted target region from the $\mathbf{P}_{refine}$ is extracted to generate a pseudo-mask, which serves as the final mask prompt $\mathbf{P}_{mask}$ . + +# 3.3. Embedding Inputs for SAM + +After the mask prompt is generated, we will provide three inputs to SAM's mask decoder to support it in completing the COD task. + +First, the primary generated mask prompt is encoded by SAM's prompt encoder $E_{S_{prompt}}$ , ultimately producing a dense embedding: + +$$ +E m b _ {d e n s e} = \mathrm {E} _ {S _ {p r o m p t}} \left(\mathrm {P} _ {m a s k}\right). \tag {9} +$$ + +Second, the image embedding produced by the final transformer layer of SAM's image encoder is combined with the visual features from the last adapter to generate the final image embedding: + +$$ +E m b _ {S _ {i m g}} ^ {\text {o u t}} = E m b _ {A} ^ {\text {l a s t}} + E m b _ {S _ {i m g}} ^ {\text {l a s t}}. \tag {10} +$$ + +Additionally, to make the most of the generated multiple information, we encode $t_{noun}$ into the text embedding $Emb_{M_{text}}$ using Mamba [11], a sequence modeling architecture known for its advanced ability to process textual information. The generated $Emb_{M_{text}}$ clearly conveys the semantic information of the camouflaged target category and is concatenated with the $Emb_{B_{img}}$ extracted from BLIP to generate the auxiliary sparse embedding to provide richer features: + +$$ +E m b _ {A u x} = \operatorname {C o n c a t} \left(E m b _ {M _ {t e x t}}, E m b _ {B _ {i m g}}\right), \tag {11} +$$ + +where the $\mathrm{Concat}(\cdot)$ is the concatenate operation. + +Finally, $Emb_{dense}$ , along with the $Emb_{Simg}^{out}$ and $Emb_{Aux}$ , are input into SAM's mask decoder to obtain the final segmentation result. + +
MethodPub. YearCHAMELEONCAMOCOD10KNC4K
SmFβmaxEφMSmFβmaxEφMSmFβmaxEφMSmFβmaxEφM
ZoomNet [42]CVPR220.9010.8760.9470.0230.8200.8050.8890.0660.8370.7770.8960.0290.8520.8260.9030.044
SegMaR [25]CVPR220.9060.8880.9590.0250.8160.8030.8840.0710.8330.7740.9060.0340.8410.8260.9070.046
UEDG [39]TMM230.9110.8940.9680.0230.8630.8560.9290.0480.8580.8120.9340.0250.8790.8640.9350.035
TinyCOD [58]ICASSP230.8870.8610.9580.0300.8220.8070.8990.0660.8110.7420.9030.0360.8430.8170.9100.047
MSCAF-Net [36]TCSVT230.9120.9020.9700.0220.8730.8670.9370.0460.8650.8230.9360.0240.8870.8740.9420.032
FSPNet [24]CVPR230.9090.8900.9650.0230.8560.8460.9280.0500.8510.7940.9300.0260.8790.8590.9370.035
FEDER [16]CVPR230.8870.8680.9540.0300.8020.7890.8730.0710.8220.7680.9050.0320.8470.8330.9150.044
FSNet [49]TIP230.9050.8910.9750.0220.8800.8780.9410.0410.8700.8330.9480.0230.8910.8800.9480.031
CRNet [19]AAAI230.8180.7920.9090.0460.7350.7070.8300.0920.7330.6360.8450.049————
SDRNet [13]KBS240.9140.9010.9610.0240.8720.8670.9320.0490.8710.8280.9360.0230.8890.8760.9400.032
PRNet [23]TCSVT240.9150.9020.9730.0200.8720.8670.9300.0500.8730.8390.9430.0220.8910.8810.9420.031
CamoFormer-R [59]TPAMI240.8980.8800.9490.0260.8170.8010.8830.0680.8380.7860.9280.0290.8540.8290.9080.042
CamoFormer-P [59]TPAMI240.9100.8980.9660.0220.8720.8680.9380.0460.8690.8290.9390.0230.8920.8800.9460.030
ZoomNeXtpv2b3 [43]TPAMI240.9280.9190.9770.0170.8850.8860.9420.0420.8950.8640.9510.0180.9000.8910.9480.028
ZoomNeXtpv2b2 [43]TPAMI240.9220.9080.9690.0170.8740.8730.9310.0470.8870.8560.9450.0190.8920.8840.9410.030
DSAM [60]ACMM24————0.8320.8340.9200.0610.8450.8050.9300.0330.8720.8640.9420.040
MAMIFNet [55]IF250.9140.8990.9590.0210.8720.8700.9350.0450.8690.8260.9400.0230.8900.8780.9430.031
COMPrompter [64]SCIS250.8850.8640.9570.0300.8530.8560.9310.0540.8600.8260.9460.0270.8800.8760.9460.036
Ours—0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.0200.9060.8970.9470.033
+ +![](images/02641099111cdf0262d4c159d7a66afaee13fa3e3b4b6965df5a4e203f9047db.jpg) +Table 1. Quantitative comparison with other SOTA methods across different datasets. The best results are highlighted in bold. + +![](images/d2d4f7ae988551f9c6c24a05f35675d8e54c8d22b97738619bacd2879db67628.jpg) + +![](images/6600128e05468b003601c1e71e6935e069f87abd042f645e922233193518a4df.jpg) + +![](images/b76efdc460cc9b6c856481aba53e7db3c8b4d7f07dfb415e393fe5951a623dba.jpg) +Figure 4. $F_{\beta}$ and PR curves of our work compared to other SOTA methods. + +![](images/f3c6840d9d3725a06ddd38fc64291a34ede3bb546301309a6b94338575153432.jpg) + +![](images/591d62dc81d4ab271906060bf90b4a12a3787020f10ed73d2d47456959b48049.jpg) + +# 4. Experiments + +# 4.1. Experimental Setup + +Datasets. To verify the effectiveness of the proposed architecture in the COD task, we use four commonly used COD datasets for evaluation: CHAMELEON [48], CAMO [28], COD10K [10], and NC4K [38]. In detail, CHAMELEON consists of 76 images featuring highly camouflaged animals in natural environments, all used for testing. CAMO contains 1250 images, including both biological camouflage and artificially designed camouflage. Among them, 1000 images are used for training, while the rest are reserved + +for testing. COD10K is more extensive, comprising 5066 images that include both artificial and natural camouflage, with 3040 images allocated for training and the remainder used for testing. While NC4K comprises 4,121 testing images specifically designed for the COD task. Notably, all four datasets provide only image data and do not include additional modalities, such as text. + +Evaluation Metrics. In the subsequent experiments, we mainly use four metrics for evaluation: S-measure [8] $(Sm)$ , maximum F-measure [1] $(F_{\beta}^{max})$ , maximum E-measure [9] $(E_{\varphi}^{max})$ , and Mean Absolute Error $(M)$ . Among these, higher values of the $Sm$ , $F_{\beta}^{max}$ , and $E_{\varphi}^{max}$ + +![](images/74c8575355ce31e2fdf9d320520eb450d5219d8f9cdbfce68cfb7cfb0cbac8fd.jpg) + +![](images/a4f26e241fc3d48d7f1813d74b6e0dba4b906dbe58484f652853a7720606df79.jpg) + +![](images/ab8f8ed22efe07310571b887f51542350ce6721fc1a035a612e6253b69583d08.jpg) + +![](images/69f387c467cc62a7c1f9aef7465865cc850abc79c9c98c2764bc0cb66ec09ca7.jpg) + +![](images/5a029378a929356983f01b74b51ed45e4d41ff688125d94c9e3104db254f7657.jpg) + +![](images/ecf46666e03eae29667354466666f7aa20b6bd79d76e4ec88add95ad99014380.jpg) + +![](images/d9508d34edeac472c55f7551a6aa193a7f9d9debd28404c63a5dc7dc95ee8132.jpg) + +![](images/1ccda65e372f20b5395cd6e70c0856e85db95a257c774291fc02b05ca64fc11b.jpg) + +![](images/0b6b4e9c96cd655459fcda4f1d870dc12ebab1ab493f15d9d079762373b442d5.jpg) + +![](images/fb600f8f53b41ef36b1133e4b04a82c9e09213b5ef78d86dcf5cbcfb5d67f207.jpg) + +![](images/b7682f308ab913ae53d398da451e5533d4868bddc380c37506a0c07f4fbe02ee.jpg) + +![](images/afe1446bf035a6836f665cc3832462da8af125919dd69e135d01a2a78d0b2753.jpg) + +![](images/3bca2ccd10b949f19779e1d0c9bb679fa58013fb121c582cea93cbb79ccff60a.jpg) + +![](images/3dd6fc12da9b18382674a264b19bc070583d5c5615811afb35582f589b25a774.jpg) + +![](images/c269fd63974c9c55200c7ee70b7bc571dc77782e884f7182b2f0d0d1a60c510b.jpg) + +![](images/c310ab2934aa67d9ff9a8f286fc0c15f5aacb3b6ac0f789656f41b5616b70952.jpg) + +![](images/ee13013b871a79d6efe1c2729d938335821a2c99594bb6c3b824b99a0cc2a470.jpg) + +![](images/b757697e323d173e790123f065a93d993420c70abd04844bcd2ad7af67c4f4d7.jpg) + +![](images/08f1ad7ec29e475d5c77c759b74f5be5237adeead3d5afac22a6ce13d33db15e.jpg) + +![](images/652d58dd61d770cb3d8730c32dace42e26640522778f43d7294e5b655c76f78e.jpg) + +![](images/d01364f9fa35dad1aaa1a0c1f333c230d64f4a92e1b40f5d5ad338b8c4240321.jpg) + +![](images/2c4d4bd31e2d112e685d56835932188124fafd11dfc5ed2be0a1b6bd6cee7722.jpg) + +![](images/18283c3e19110b5b6084c8c08ec2a9eb502254a49a42360b28b272d48383fcb5.jpg) + +![](images/1d47931c7f5375bbc4f8b5522d1faa255d089ff85aba4d6b2295ae76cdd894c0.jpg) + +![](images/903c58fd611f2580a6f225a83ef5938764e08c5769ff8913a161a9d432b9476f.jpg) + +![](images/7c832920d4c9c8215b03be14a922777c33aec278aa387bebc3a5450d654f122f.jpg) + +![](images/16f4d64288f2fd1070515fa0f02c0a1ba10fb14566e1002047277e97996d824d.jpg) + +![](images/b95ec0903d37511d43a43b3d122dbf783e444810395e5b0e175e424766787448.jpg) +Image + +![](images/50c2e4853535d36f32580ae080100896f8f71a6a2db64bc24d93249392f5830c.jpg) +GT + +![](images/a065d746fd94323fd4e9743e6a3837638d3508839bed335401374a41bc7cdc0f.jpg) +Ours + +![](images/1b46b78b13e1b9729d5e30566c9ec126ba07274705c5c0cf0dd4596502895108.jpg) +COMPrompter + +![](images/e0f917b7ec1d27e6998821f8ab5a5505fdcf782b43a0bb663bd31459ca581031.jpg) +ZoomNeXt + +![](images/77938d9574bb33dc48fbc47b8490a06f14619d18f34b58f834a448db922588f7.jpg) +PRNet +Figure 5. Visual comparison of segmentation results with other SOTA methods on different samples. + +![](images/b1a80939311555b022867bfb6dd91248b884705d4077cd4ac9bba3036bca0fbf.jpg) +FSNet + +![](images/d89ace3c2f3d616640d46d23530e6f8f6ae8d65de8fae9280af7ff96ae879a21.jpg) +FSPNet + +![](images/71be723cb02203bee17bdbec6fcc278860063392b3ca9ce504da644b18ee5edd.jpg) +ZoomNet + +indicate better model performance, whereas for $M$ , lower values indicate better performance. Additionally, the precision-recall curve (PR curve) and the $F_{\beta}$ -threshold curve $(F_{\beta}$ curve) are also considered for further evaluation. + +Implement Details. Our proposed framework is primarily built upon the SAM-Vit Huge (ViT-H) [27] and implemented using PyTorch. The entire architecture is trained on an NVIDIA GeForce RTX 4090 GPU for 40 epochs with a batch size of 1. The main learning rate is set to 0.0002, while the learning rate for the ARM is 0.1 times that of the main learning rate, and the AdamW optimizer is used for optimization. Furthermore, both the VLM and LLM, used for multimodal prompt and auxiliary embedding generation, are pre-trained models, so no additional training or fine-tuning is required for these two. + +# 4.2. Comparisons With State-of-the-art Methods + +To effectively evaluate the performance of our work in the COD task, we first conducted a comparative analysis with the current state-of-the-art (SOTA) models. + +Quantitative Evaluation. As shown in Tab. 1, which compares the evaluation results of our work and other SOTA models across four different datasets. Our work generally demonstrates exceptional performance in solving the COD task across various metrics on different datasets. In particular, our method excels in both the $Sm$ and $F_{\beta}^{max}$ , showing superior performance in target localization and structure capture of the camouflaged objects in complex scenarios. For example, on the COD10K, our approach outperforms the second-best method by $1.6\%$ in $Sm$ and $2.4\%$ in + +$F_{\beta}^{max}$ . Although our $F_{\beta}^{max}$ is the second on the CAMO, the gap with the top-performing model is only $0.3\%$ , which is negligible. Furthermore, from the comparison of the $F_{\beta}$ and PR curves in Fig. 4, it is clear that our method (red solid curve) exhibits outstanding performance across all three datasets, with particularly remarkable results on the COD10K dataset. + +Qualitative Evaluation Fig. 5 illustrates a comparison of the final segmentation maps of our work and other SOTA methods across different samples. It is clear that our approach achieves a more significant overlap with the ground truth (GT) than other SOTA models. Specifically, our method is more effective at accurately segmenting and capturing completely camouflaged objects. Even in more challenging scenarios, such as the third sample in Fig. 5, where multiple small camouflaged objects need to be identified, our approach segments more objects and provides more accurate segmentation, while the results from other models show noticeable omissions. These observations highlight the superior performance of our framework, not only in localizing camouflaged objects but also in maintaining the structural integrity of the camouflaged objects. + +# 4.3. Ablation Studies + +As shown in Tab. 2, we conduct ablation studies on the improvements discussed in Sec. 3 to assess the effectiveness of each component. + +Effectiveness of Auxiliary Embedding. Since the auxiliary embedding is generated alongside the multimodal data produced by BLIP, we first incorporate it into the SAM to + +
CHAMELEONCAMOCOD10K
SmFβmaxEφmaxMSmFβmaxEφmaxMSmFβmaxEφmaxM
SAM0.8500.8200.9110.0560.7970.7640.8640.1010.8100.7500.8810.054
SAM + EmbAUX0.8520.8160.9220.0470.8050.7780.8820.0890.8200.7510.8930.039
SAM + EmbAUX + Pcaptioninit0.8660.8360.9200.0460.8230.7950.8910.0810.8390.7760.9080.037
SAM + EmbAUX + Pnouninit0.8820.8590.9350.0400.8310.8070.8960.0800.8460.7890.9140.036
SAM + EmbAUX + PnounARM0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.020
+ +Table 2. Ablation studies on three different COD datasets. + denotes additional information provided to SAM, $Emb_{Aux}$ is the auxiliary sparse embedding, $P_{init}^{caption}$ is the initial mask prompt generated by the entire caption, $P_{init}^{noun}$ is the initial mask prompt generated by the extracted noun, $P_{ARM}^{noun}$ is the refined $P_{init}^{noun}$ by ARM. + +![](images/96f6bd4d5374a0023968bf3ec8648745c9f51c7a9e15aa2606c5af2d22986ebb.jpg) + +![](images/e0ad3fdf6ce28e4641430ece9c327ddc13b8dbe9ec966323448d944d6ee69132.jpg) + +![](images/9e02f4db4184054407802e9b3a946578375a2456bd82aa3048e9034244fde7e7.jpg) + +![](images/3c9ace7879d2e460dcdedbdf5daad351fe0c6225bc00500908870bb0ff583dce.jpg) + +![](images/f2d92d68d01c47cb3b7a17a7c6132d1fc87084f3e1fd3973f2a35f36cddaab1d.jpg) +Image + +![](images/b3ae903b0735e5236a189dfd0d8b1bbb9f1e94c9a7d3819a4bd90e74d6faa3f7.jpg) + +![](images/bc5237428dd668c32af7793d32dfac700b9a4bda0cc33ab45f1702f171f7f2a9.jpg) + +![](images/accd0adc389c776c939aac9a94d508e4c4f8f81047dd0fb6545c27875463e72c.jpg) + +![](images/96d9d34953ce2a0a72507e3c6af22086615e89e9daf4ee0db05a9ea3735dff13.jpg) + +![](images/9a2b87cad6a2461aafa78aa5689b4e660d89c42549587ae483f3d7d74e1087a2.jpg) +GT + +![](images/3ac3e62534fc3a5b73d9df592601fb972982eaef8548d0db14a84aa8ef79fdbd.jpg) + +![](images/41bd62e19c8b1a689260eb1648bb12cc980e8b545346a9e5c26076b70d445761.jpg) + +![](images/7a2380f758b652612400f1a056636b16c03ab261c3405726bc6c531be1feda1b.jpg) + +![](images/19d4394c6805bb428218f9cad76f9d7a2e5ded53e44ee5bbba46d1148a8b1062.jpg) + +![](images/e0c709bf75816ce94a630106acba9aee7a84318a69e02f66e3608f0f99d9a350.jpg) +Initial + +![](images/4ae79fa60091721b506a01f02a062cc8cb0656809774d2fade069cc3461b4f91.jpg) + +![](images/28e7c04f602f25e9c4eb1a6182878d8d43b46f8f11338d0e65524bce331a003b.jpg) + +![](images/34b5beeb6abfeb220152cb29ecb4ed28d614236100f638debc200dda9c4904ee.jpg) + +![](images/4999b7bbda47d2425a4db34b8e11a212f829ce1e7fed7b0f4f1e5121f0f077bd.jpg) + +![](images/7a57f11a0171448c428431c7c785cd336d6e22be18aaa1b79449f6dedec3f365.jpg) +Refined by ARM +Figure 6. Comparison between the initially generated prompt and the prompt refined using ARM. + +evaluate its effectiveness. Comparing the $2^{nd}$ row with the $1^{st}$ row in Tab. 2, adding the auxiliary embedding leads to performance improvements across nearly all metrics on the three COD datasets. This indicates that the auxiliary embedding effectively enhances SAM by providing additional multimodal features. + +Effectiveness of Multimodal Prompt. Our primary contribution is utilizing BLIP to generate multimodal prompts, providing SAM with more informative cues to handle the COD task. Experimental results ( $3^{rd}$ row of Tab. 2) show that prompts generated from full description already outperform SAM with auxiliary embedding. Furthermore, SAM's performance further improved when extracting the noun about the object category from the long description as textual information to generate prompts ( $4^{th}$ row of Tab. 2). This result not only proves the effectiveness of our generated prompts but also indicates that, although BLIP can produce descriptions closely related to the image, redun- + +dant information introduces noise, preventing the generated prompt from accurately localizing the targets. In contrast, providing only category information helps generate more precise mask prompts. + +Effectiveness of Refinement. As shown in the $5^{th}$ row results in Tab. 2, our proposed ARM significantly enhances SAM's performance on the COD task, achieving substantial improvements across all evaluation metrics on three datasets. This performance boost can be attributed to two key aspects: First, ARM refines the multimodal prompts initially generated by BLIP, as illustrated in Fig. 6. Due to domain shifts, these initial prompts often localize only partial regions of the target. However, after refining with ARM, the refined prompts achieve more complete target coverage. Consequently, the mask prompts generated from these refined prompts provide SAM with more precise cues. Second, ARM effectively integrates the image features extracted by the adapter into SAM, enabling fine-tuning of the image encoder. This process further enhances SAM's representation capability, leading to improved performance on the COD task. + +# 5. Conclusion + +In this work, we propose a novel and cost-effective solution for the COD task. The main idea is to incorporate multimodal information to construct high-quality mask prompts, enhancing SAM's segmentation capability. Specifically, we use VLM to automatically generate multimodal data and prompts, laying a solid foundation for constructing the final high-quality mask prompts. Furthermore, we propose the Adaptive Refinement Module (ARM) to refine the initial multimodal prompts enabling more complete target representations and fine-tune SAM's image encoder, providing it with richer information. Additionally, experimental results demonstrate that our proposed architecture achieves outstanding performance across multiple COD datasets, particularly excelling in capturing the structural details of objects, thereby further validating our work's effectiveness. + +# Acknowledgements + +The work is supported by the XJTLU RDF project: RDF-22-02-118 & RDF-24-02-041 + +# References + +[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 6 +[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 1 +[3] Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. Sam-cod+: Sam-guided unified framework for weakly-supervised camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2 +[4] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023. 2 +[5] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?-sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2304.09148, 2023. 2 +[6] Zhennan Chen, Rongrong Gao, Tian-Zhu Xiang, and Fan Lin. Diffusion model for camouflaged object detection. In ECAI 2023, pages 445-452. IOS Press, 2023. 2 +[7] Zhongxi Chen, Ke Sun, and Xianming Lin. Camodiffusion: Camouflaged object detection via conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1280, 2024. 2 +[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 6 +[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6 +[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 6 +[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5 +[12] Juwei Guan, Xiaolin Fang, Tongxin Zhu, Zhipeng Cai, Zhen Ling, Ming Yang, and Junzhou Luo. Idenet: Making neural network identify camouflaged objects like creatures. IEEE Transactions on Image Processing, 2024. 1 +[13] Juwei Guan, Xiaolin Fang, Tongxin Zhu, and Weiqi Qian. Sdrnet: Camouflaged object detection with independent re + +construction of structure and detail. Knowledge-Based Systems, 299:112051, 2024. 6 +[14] Juwei Guan, Weiqi Qian, Tongxin Zhu, and Xiaolin Fang. Promoting camouflaged object detection through novel edge-target interaction and frequency-spatial fusion. Neurocomputing, 617:129064, 2025. 1, 2 +[15] Cunhan Guo and Heyan Huang. Cofinet: Unveiling camouflaged objects with multi-scale finesse. Neurocomputing, 614:128763, 2025. 2 +[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22046-22055, 2023. 6 +[17] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36:30726-30737, 2023. 2 +[18] Jiangping He, Yangjie Zhang, Maggie Chung, Michael Wang, Kun Wang, Yan Ma, Xiaoyang Ding, Qiang Li, and Yonglin Pu. Whole-body tumor segmentation from pet/ct images using a two-stage cascaded neural network with camouflaged object detection mechanisms. Medical Physics, 50 (10):6151-6162, 2023. 1 +[19] Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI conference on artificial intelligence, pages 781-789, 2023. 6 +[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5 +[21] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 4 +[22] Bintao Hu, Hengyan Liu, Jianbo Du, Miguel López-Benitez, Celimuge Wu, Xiaoli Chu, and Dusit Niyato. Madqnenhanced computation offloading and resource allocation for 6g low-altitude economy vehicular networks. IEEE Transactions on Cognitive Communications and Networking, 2025. 1 +[23] Xihang Hu, Xiaoli Zhang, Fasheng Wang, Jing Sun, and Fuming Sun. Efficient camouflaged object detection network based on global localization perception and local guidance refinement. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):5452-5465, 2024. 6 +[24] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5557-5566, 2023. 6 +[25] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4713-4722, 2022. 6 + +[26] Xinhao Jiang, Wei Cai, Yao Ding, Xin Wang, Danfeng Hong, Xingyu Di, and Weijie Gao. Bidicos: Camouflaged object segmentation via bilateral diffusion model. Expert Systems with Applications, 255:124747, 2024. 2 +[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 7 +[28] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and image understanding, 184:45-56, 2019. 6 +[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 4 +[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4 +[31] Lin Li, Bo Dong, Eric Rigall, Tao Zhou, Junyu Dong, and Geng Chen. Marine animal segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(4): 2303-2314, 2021. 1 +[32] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 5 +[33] Hengyan Liu, Guangyu Ren, Tianhong Dai, Di Zhang, Pengjing Xu, Wenzhang Zhang, and Bintao Hu. Diversity-oriented contrastive learning for rgb-t scene parsing. In 2023 2nd International Conference on Sensing, Measurement, Communication and Internet of Things Technologies (SMC-IoT), pages 155-160. IEEE, 2023. 1 +[34] Hengyan Liu, Wenzhang Zhang, Tianhong Dai, Longfei Yin, and Guangyu Ren. Multimodal frequency spectrum fusion schema for rgb-t image semantic segmentation. In 2024 33rd International Conference on Computer Communications and Networks (ICCNN), pages 1-6. IEEE, 2024. 1 +[35] Limei Liu, Meng Liu, Kexin Meng, Linqiang Yang, Min Zhao, and Shuli Mei. Camouflaged locust segmentation based on pranet. Computers and Electronics in Agriculture, 198:107061, 2022. 1 +[36] Yu Liu, Haihang Li, Juan Cheng, and Xun Chen. Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features. IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4934-4947, 2023. 6 +[37] Yantong Liu, Sai Che, Liwei Ai, Chuanxiang Song, Zheyu Zhang, Yongkang Zhou, Xiao Yang, and Chen Xian. Camouflage detection: Optimization-based computer vision for alligator sinensis with low detectability in complex wild environments. Ecological Informatics, 83:102802, 2024. 1 + +[38] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11591-11601, 2021. 6 +[39] Yixuan Lyu, Hong Zhang, Yan Li, Hanyang Liu, Yifan Yang, and Ding Yuan. Uedg: Uncertainty-edge dual guided camouflage object detection. IEEE Transactions on Multimedia, 26:4050-4060, 2023. 6 +[40] Tsui Qin Mok, Shuyong Gao, Haozhe Xing, Miaoyang He, Yan Wang, and Wenqiang Zhang. A holistically point-guided text framework for weakly-supervised camouflaged object detection. arXiv preprint arXiv:2501.06038, 2025. 3 +[41] Yuzhen Niu, Lifen Yang, Rui Xu, Yuezhou Li, and Yuzhong Chen. Minet: Weakly-supervised camouflaged object detection through mutual interaction between region and edge cues. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6316-6325, 2024. 1 +[42] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2160-2170, 2022. 2, 6 +[43] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE transactions on pattern analysis and machine intelligence, 2024. 1, 2, 6 +[44] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. arXiv preprint arXiv:2311.10121, 2023. 2 +[45] Guangyu Ren, Yinxiao Yu, Hengyan Liu, and Tania Stathaki. Dynamic knowledge distillation with noise elimination for rgb-d salient object detection. Sensors, 22(16):6188, 2022. 1 +[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4 +[47] Yiqing Shen, David Dreizin, Blanca Inigo, and Mathias Unberath. Protosam-3d: Interactive semantic segmentation in volumetric medical imaging via a segment anything model and mask-level prototypes. Computerized Medical Imaging and Graphics, page 102501, 2025. 2 +[48] Przemysław Skurowski, Hassan Abdulameer, Jakub Blaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6 +[49] Ze Song, Xudong Kang, Xiaohui Wei, Haibo Liu, Renwei Dian, and Shutao Li. Fsnet: Focus scanning network for camouflaged object detection. IEEE Transactions on Image Processing, 32:2267-2278, 2023. 6 +[50] Yanguang Sun, Chunyan Xu, Jian Yang, Hanyu Xuan, and Lei Luo. Frequency-spatial entanglement learning for camouflaged object detection. In European Conference on Computer Vision, pages 343-360. Springer, 2024. 2 + +[51] Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8805–8814, 2024. 3 +[52] Huafeng Wang, Tianyu Hu, Yanan Zhang, Haodu Zhang, Yong Qi, Longzhen Wang, Jianhua Ma, and Minghua Du. Unveiling camouflaged and partially occluded colorectal polyps: Introducing cpsnet for accurate colon polyp segmentation. Computers in Biology and Medicine, 171:108186, 2024. 1 +[53] Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. Depth-aware concealed crop detection in dense agricultural scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17201-17211, 2024. 1, 3 +[54] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3297–3306, 2023. 3 +[55] Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection. Information Fusion, 117:102871, 2025. 2, 6 +[56] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1032-1042, 2023. 3 +[57] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv preprint arXiv:2106.13217, 2021. 3 +[58] Haozhe Xing, Shuyong Gao, Hao Tang, Tsui Qin Mok, Yanlan Kang, and Wenqiang Zhang. Tinycod: Tiny and effective model for camouflaged object detection. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 6 +[59] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6 +[60] Zhenni Yu, Xiaogin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. Exploring deeper! segment anything model with depth perception for camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4322-4330, 2024. 6 +[61] Hong Zhang, Yixuan Lyu, Qian Yu, Hanyang Liu, Huimin Ma, Ding Yuan, and Yifan Yang. Unlocking attributes' contribution to successful camouflage: A combined textual and visual analysis strategy. In European Conference on Computer Vision, pages 315-331. Springer, 2024. 1, 3 +[62] Jie Zhang, Xubing Yang, Rui Jiang, Wei Shao, and Li Zhang. Rsam-seg: A sam-based approach with prior knowledge in + +tegration for remote sensing image semantic segmentation. arXiv preprint arXiv:2402.19004, 2024. 2 +[63] Shizhou Zhang, Dexuan Kong, Yinghui Xing, Yue Lu, Lingyan Ran, Guoqiang Liang, Hexu Wang, and Yanning Zhang. Frequency-guided spatial adaptation for camouflaged object detection. IEEE Transactions on Multimedia, 27:72-83, 2025. 2 +[64] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 6 +[65] Ke Zhou, Zhongwei Qiu, and Dongmei Fu. Multi-scale contrastive adaptor learning for segmenting anything in under-performed scenes. Neurocomputing, 606:128395, 2024. 2 \ No newline at end of file diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/images.zip b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..67b4e29d1390394e21e899d935ad09645f6367ec --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a95ddc2a19ca3e29f6806fc801f258a5261d7db34cc2037a97cdf72b4571015a +size 797796 diff --git a/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/layout.json b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f9e98069569e6d75edc860d5300eab367aa2ab27 --- /dev/null +++ b/2025/Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection/layout.json @@ -0,0 +1,12306 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 74, + 102, + 536, + 140 + ], + "type": "text", + "content": "Enhancing Prompt Generation with Adaptive Refinement for Camouflaged Object Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "spans": [ + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "content": "Xuehan Chen" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "content": " Guangyu Ren" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "content": " Tianhong Dai" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "content": " Tania Stathaki" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "text", + "content": " Hengyan Liu" + }, + { + "bbox": [ + 97, + 160, + 517, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 175, + 413, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 175, + 413, + 189 + ], + "spans": [ + { + "bbox": [ + 198, + 175, + 413, + 189 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 198, + 175, + 413, + 189 + ], + "type": "text", + "content": "Xi'an Jiaotong-Liverpool University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 198, + 189, + 412, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 189, + 412, + 203 + ], + "spans": [ + { + "bbox": [ + 198, + 189, + 412, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 198, + 189, + 412, + 203 + ], + "type": "text", + "content": "Imperial College London, United Kingdom" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 134, + 205, + 476, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 205, + 476, + 217 + ], + "spans": [ + { + "bbox": [ + 134, + 205, + 476, + 217 + ], + "type": "text", + "content": "xuehanchen83@gmail.com, {guangyu. ren, hengyan.liu}@xjtlu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 220, + 474, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 220, + 474, + 231 + ], + "spans": [ + { + "bbox": [ + 133, + 220, + 474, + 231 + ], + "type": "text", + "content": "tianhong.dai15@alumni.imperial.ac.uk, t.stathaki@imperial.ac.uk" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 284, + 296, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 296, + 524 + ], + "type": "text", + "content": "Foundation models, such as Segment Anything Model (SAM), have exhibited remarkable performance in conventional segmentation tasks, primarily due to their training on large-scale datasets. Nonetheless, challenges remain in specific downstream tasks, such as Camouflaged Object Detection (COD). Existing research primarily aims to enhance performance by integrating additional multimodal information derived from other foundation models. However, directly leveraging the information generated by these models may introduce additional biases due to domain shifts. To address this issue, we propose an Adaptive Refinement Module (ARM), which efficiently processes multimodal information and simultaneously refining the mask prompt. Furthermore, we construct an auxiliary embedding that effectively exploits the intermediate information generated during ARM, providing SAM with richer feature representations. Experimental results indicate that our proposed architecture surpasses most state-of-the-art (SOTA) models in the COD task, particularly excelling in structured target segmentation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 135, + 562 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 571, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 654 + ], + "type": "text", + "content": "Camouflaged Object Detection (COD) is a challenging computer vision task, as it involves segmenting objects that are virtually identical to their immediate surroundings. It plays a crucial role in the field due to its wide range of practical applications, including object detection [34, 45], low-altitude economy [22], zoological research [31, 35, 37], and medical image segmentation [18, 52]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 655, + 296, + 692 + ], + "type": "text", + "content": "Current research [12, 14, 33, 41, 43] focuses on designing advanced architectures to extract and refine more detailed features exclusively from image data. However, the" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 316, + 257, + 552, + 389 + ], + "blocks": [ + { + "bbox": [ + 316, + 257, + 552, + 389 + ], + "lines": [ + { + "bbox": [ + 316, + 257, + 552, + 389 + ], + "spans": [ + { + "bbox": [ + 316, + 257, + 552, + 389 + ], + "type": "image", + "image_path": "50d4f60705a2b5a90993390a56097d82f8c3bd792a4f80336e65cd635d0f2f55.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 316, + 393, + 550, + 453 + ], + "blocks": [ + { + "bbox": [ + 316, + 393, + 550, + 453 + ], + "lines": [ + { + "bbox": [ + 316, + 393, + 550, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 550, + 453 + ], + "type": "image", + "image_path": "adc2cb2ca55e642b64beec235145ca67f12d5ecd5a73506f5d0ac261735ccba6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 403, + 460, + 432, + 470 + ], + "lines": [ + { + "bbox": [ + 403, + 460, + 432, + 470 + ], + "spans": [ + { + "bbox": [ + 403, + 460, + 432, + 470 + ], + "type": "text", + "content": "(c) Ours" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 479, + 555, + 579 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 579 + ], + "type": "text", + "content": "Figure 1. Comparison of our proposed Adaptive Refinement Module (ARM) with common COD frameworks. Unlike conventional methods that rely solely on images or manually annotated additional modalities, our approach not only leverages vision-language models (VLMs) to automatically generate multimodal data but also mitigates domain shift through the proposed ARM, providing high-quality prompts for SAM's decoder. Meanwhile, ARM can also fine-tune SAM's image encoder, further improving SAM's feature representation capability." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "unique nature of the COD task, where the targets are visually similar to their surroundings, limits the effectiveness of image features. Additionally, the lack of data from other modalities further limits the models' ability to acquire meaningful features, making them highly susceptible to background, preventing them from accurately detecting camouflaged objects. Although some studies incorporate human-made auxiliary data, such as depth images [53], textual descriptions [61], or prompts [2, 3], to provide supple" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20672" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "mentary information about targets, they are not ideal solutions as they are time-consuming and labor-intensive. As a result, despite significant progress in the COD task, two limitations remain: limited image features and laborious annotation processes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 134, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 134, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 134, + 294, + 277 + ], + "type": "text", + "content": "Recently, foundation models have emerged as a popular solution for tackling a range of tasks, including the Segment Anything Model (SAM) [27], which has proven highly effective in handling various tasks, such as medical image segmentation [44, 47] and semantic segmentation [4, 62]. Although SAM demonstrates impressive zero-shot generalization capabilities, Chen et al. [5] states that it still struggles to segment the complex camouflaged objects. Secondly, SAM can guide its segmentation by receiving prompts, and the most intuitive approach is to generate prompts through manual annotation, but as mentioned earlier, it is less practical due to the time and labor it demands." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 279, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 279, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 279, + 295, + 471 + ], + "type": "text", + "content": "To mitigate the aforementioned limitations, we rethink the core issues that make the COD task challenging. In this task, there are no task-specific clues (e.g., category, size, location) of the targets. Meanwhile, the background's visual noise increases the task's complexity, resulting in ambiguous object boundaries and weakened feature learning. So, we propose a novel approach that uses textual information to complement visual features, enhancing the model's ability to localize target objects more accurately. In this work, we fully utilize the powerful pretrained vision-language models (VLMs) to automatically generate multimodal data without requiring additional training or fine-tuning. This data provides additional cues to assist in object localization. Subsequently, the VLMs further utilize the generated multimodal information to produce mask prompts that convey the object's approximate location to SAM." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 651 + ], + "type": "text", + "content": "We also consider the biases introduced by directly utilizing generated multimodal information, which could lead to locality in the prompts, making them less comprehensive and potentially affecting the subsequent SAM segmentation performance. To mitigate this issue, this paper proposes an Adaptive Refinement Module (ARM), which incorporates simple adapters to refine the initially generated mask prompts. This module expands the regions in the initial prompts to create a more complete representation of the target objects, offering better guidance for downstream models. Furthermore, ARM effectively extracts vision information from the entire architecture, serving not only as a refinement mechanism for prompts but also as a fine-tuning module for SAM's image encoder, further enhancing SAM's performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 713 + ], + "type": "text", + "content": "In summary, this work proposes the ARM, which simultaneously refines the initial prompt generated by multimodal information and fine-tunes SAM's image encoder, enabling SAM to handle the COD task more effectively across multiple aspects. Our main contributions are as fol" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 73, + 338, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 73, + 338, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 73, + 338, + 83 + ], + "type": "text", + "content": "lows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 87, + 553, + 302 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 87, + 553, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 87, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 553, + 159 + ], + "type": "text", + "content": "- We propose a cost-effective approach that automatically constructs mask prompts using multimodal information generated by BLIP, serving as additional cues without requiring extra training or fine-tuning. These prompts provide reliable guidance for SAM, thereby improving its segmentation performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 159, + 553, + 242 + ], + "type": "text", + "content": "- To mitigate the additional biases introduced by BLIP, we propose an Adaptive Refinement Module (ARM) to refine the initially generated prompts, enhancing their quality and providing more precise guidance for SAM. Meanwhile, this module can also fine-tune the SAM's image encoder, further enhancing SAM's segmentation accuracy and robustness." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 243, + 553, + 302 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 302 + ], + "type": "text", + "content": "- We further utilize the multimodal information generated by BLIP to construct an auxiliary embedding, which effectively exploits the intermediate information generated during the prompt generation process and provides SAM with richer features." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 400, + 328 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 335, + 470, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 470, + 348 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 470, + 348 + ], + "type": "text", + "content": "2.1. Single-Modality Based COD" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 553, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 553, + 461 + ], + "type": "text", + "content": "Image data is one of the most widely used modalities to tackle the COD task. Therefore, most researchers are currently exploring various approaches to extract information from single-modality image data in order to perform this task more effectively. CoFiNet [15], ZoomNext [43], ZoomNet [42] focus on extracting and fusing multiscale features from images. Although upgrading the model's backbone improves the accuracy of identifying objects, it also increases the consumption of computational resources." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 593 + ], + "type": "text", + "content": "Besides directly extracting features from the images, CamoDiffusion [7], BiDiCOS [26], and diffCOD [6] transform the COD task into denoising tasks, gradually segmenting camouflaged objects through the denoising process. Moreover, EFNet [14], FSEL [50], FGSA-Net [63], and MAM-FFC [55] use spatial and frequency information from images to achieve a more precise segmentation. However, in the COD task, the background in images introduces significant noise. Even with those novel methods relying solely on images' intrinsic features, they still struggle to recognize objects in complex scenarios." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "Furthermore, in addition to designing a new architecture, many works have also focused on improving the Segment Anything Model (SAM), taking advantage of its powerful zero-shot generalization ability and prompt strategy. SAM-COD [3] and WS-SAM [17], which combine sparse prompts with SAM to generate the final segmentation results, still require additional manual annotations for the sparse prompts, which remains time-consuming. MCA-SAM [65] and SAM-adapter [5] improve the model's performance in the COD task by injecting more image infor" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20673" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 553, + 311 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 311 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 311 + ], + "type": "image", + "image_path": "94c915325f32d32b65f2b0f77d5581f8541f6026b44182b8af77d6fa67eca345.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 319, + 555, + 387 + ], + "lines": [ + { + "bbox": [ + 54, + 319, + 555, + 387 + ], + "spans": [ + { + "bbox": [ + 54, + 319, + 555, + 387 + ], + "type": "text", + "content": "Figure 2. Overview of our proposed architecture. BLIP automatically generates multimodal data from the input image and uses it to create the initial multimodal prompts and the auxiliary embedding. The Adaptive Refinement Module (ARM) is then employed to address the locality issue of the initial prompts. The ARM uses multiple adapters to extract visual information from both BLIP and SAM, which not only refines the initial multimodal prompts by integrating attention weights, resulting in more accurate mask prompts, but also fine-tunes SAM's image encoder. Finally, the SAM decoder receives image embedding, dense mask prompt embedding, and auxiliary sparse embedding to tackle the COD task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 406, + 294, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 406, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 406, + 294, + 479 + ], + "type": "text", + "content": "mation into SAM's image encoder using the adapter. They differ from our work, where the adapter serves a broader purpose in ARM. It is primarily used to refine the generated mask prompts while injecting additional information into SAM's image encoder, enabling the SAM to perform the COD task more effectively." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 491, + 209, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 209, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 209, + 504 + ], + "type": "text", + "content": "2.2. Multi-Modality Based COD" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 714 + ], + "type": "text", + "content": "With the rapid growth of multimodal learning, different modalities have been increasingly applied to address the COD problem. HPGT [40] use the textual information to assist in completing the COD task. However, they rely on manually annotating text data for the COD dataset, which is a highly time-consuming process. Although ACU-MEN [61] utilizes the GPT-4 Vision model to generate textual descriptions and attributions for images, it still requires more than 30 volunteers to manually verify the accuracy of these generated texts, which is also a labor-intensive task. In contrast to these approaches, MMCPF [51] takes a different direction by interacting with a Multimodal Large Language Model (MLLM). They query the model using text to obtain the location coordinates of camouflaged objects, which supports the subsequent segmentation task. While this method reduces the need for text data annotation, it still faces the challenge of designing precise text prompts for querying the" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 406, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 555, + 490 + ], + "type": "text", + "content": "MLLM. Ordinary prompts cannot let MLLM output the expected location coordinates of objects, which indicates that designing effective text prompts becomes an additional task that is time-consuming and labor-intensive as well. As a result, both visual and textual information have been employed to address the COD problem, but manual annotation and verification remain a limitation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 491, + 556, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 556, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 556, + 634 + ], + "type": "text", + "content": "In addition to textual information, depth features are utilized to provide valuable spatial information for models. Due to the fact that commonly used COD datasets only include RGB images without depth data, PopNet [56], Da-COD [54] and [57] employ depth estimation models to generate depth images from RGB-based COD datasets. Since the depth images are generated automatically by models, the final segmentation performance is highly reliant on the performance of those depth estimation methods. In contrast to the previous approaches, RISNet [53] uses a depth camera to manually create an RGB-D dataset for COD in agricultural scenes, which is also painstaking." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 645, + 370, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 370, + 657 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 370, + 657 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We propose a novel architecture, as illustrated in Fig. 2, which enables the fully automated generation of mask prompts for SAM and designs an innovative Adaptive Refinement Module (ARM) to enhance the quality of the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20674" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "initially generated mask prompts. Specifically, a Vision-Language Model (VLM) first automatically generates textual data based on the input image. Then, by aligning multimodal information, the VLM generates the multimodal prompts to locate camouflaged objects locally. Subsequently, it is fed into the ARM for refinement, enabling prompts to capture more complete object structure information, thereby providing better guidance for SAM. During the prompt refinement process, ARM also fine-tunes SAM's image encoder. Meanwhile, the generated textual data is fully utilized, which is integrated with visual information to provide additional auxiliary embedding, further improving SAM's performance in identifying camouflaged objects." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 296, + 300 + ], + "type": "text", + "content": "In the following subsections, we provide detailed explanations of each component of this architecture. First, we describe the multimodal prompt generation process in Sec. 3.1. Then we introduce the ARM in Sec. 3.2. Finally, Sec. 3.3 describes the different input embeddings fed into SAM's mask decoder to generate the final results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 309, + 228, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 228, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 228, + 322 + ], + "type": "text", + "content": "3.1. Multimodal Prompt Generation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 327, + 209, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 209, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 209, + 338 + ], + "type": "text", + "content": "3.1.1. Multimodal Data Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "content": "As a robust zero-shot VLM, BILP [30] offers powerful image captioning capabilities that align with our goal of automatically generating supplementary textual data for subsequent prompt generation. Given an image " + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "inline_equation", + "content": "I \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "content": ", the image embedding " + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "inline_equation", + "content": "Emb_{B_{img}}" + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "content": " are extracted by the BLIP's image encoder " + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "inline_equation", + "content": "E_{B_{img}}" + }, + { + "bbox": [ + 55, + 342, + 296, + 438 + ], + "type": "text", + "content": ". This embedding is then passed to BLIP's image-grounded text decoder, which generates a caption, as presented below:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 102, + 453, + 295, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 453, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 102, + 453, + 295, + 467 + ], + "type": "interline_equation", + "content": "T = \\mathbf {D} _ {B _ {t e x t}} \\left(E m b _ {B _ {i m g}}\\right) = \\left\\{t _ {i j} \\right\\} _ {i = 1} ^ {L}, \\tag {1}", + "image_path": "85e246acf5b791029f04d9311c0eed390a4fa56dc9db196890d876f026b6ebf4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": " denotes the generated caption, " + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\mathrm{D}_{B_{text}}" + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": " is the BLIP's image-grounded text decoder, and " + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": " word in the caption, with a total of " + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 473, + 296, + 544 + ], + "type": "text", + "content": " words. As exemplified by the potential caption generated in Fig. 2, it is clear that the captions generated by BLIP can correctly identify the camouflaged targets." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "However, since the generated " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " provides an overall description of the input image, it may include content that is irrelevant to the target objects, such as \"floor\" is the background information of the description \"There is a small dog that is lying on the floor\". Due to the complexity of the COD task, irrelevant information in the generated caption could act as noise, potentially distracting from generating a prompt that accurately indicates the target's location. Therefore, instead of utilizing the entire caption " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ", we use the spaCy [21] library for part-of-speech tagging to capture the necessary content about the objects. Specifically, each " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is tagged with its corresponding part of speech (e.g., noun, verb, adjective). Through observation, we found that most generated captions primarily use nouns to describe the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "category of targets (e.g., man, bug, lion). Therefore, we extract the first noun " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "t_{noun}" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " that appears in " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " to provide more precise target category information, which helps the subsequent prompt generation more accurately locate the target object and reduces distractions caused by unrelated noise." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 138, + 479, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 138, + 479, + 150 + ], + "spans": [ + { + "bbox": [ + 313, + 138, + 479, + 150 + ], + "type": "text", + "content": "3.1.2. Multimodal Prompt Generation" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 153, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 554, + 262 + ], + "type": "text", + "content": "BLIP has a powerful text-image alignment capability, which allows it to generate a multimodal prompt that can accurately locate the target object by aligning the text containing the target category information with the image. Inspired by [29], we use BLIP once again to handle the image-text retrieval task and generate the Gradient-weighted Class Activation Mapping (Grad-CAM) [46] to visualize the alignment process of BLIP with multimodal data, thereby generating the prompt." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "text", + "content": "Specifically, the " + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "inline_equation", + "content": "Emb_{B_{img}}" + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "inline_equation", + "content": "t_{noun}" + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "text", + "content": " are fed into BLIP's image-grounded text encoder " + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{B_{text}}" + }, + { + "bbox": [ + 313, + 262, + 554, + 308 + ], + "type": "text", + "content": ", which then outputs a score reflecting how well the input image and text match, as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 349, + 322, + 553, + 336 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 322, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 349, + 322, + 553, + 336 + ], + "type": "interline_equation", + "content": "(S, \\mathrm {A} _ {\\text {c r o s s}}) = \\mathrm {E} _ {B _ {\\text {t e x t}}} \\left(\\left(E m b _ {B i m g}, t _ {\\text {n o u n}}\\right), \\right. \\tag {2}", + "image_path": "c2ab272c593c8b6ddb0d6253a042ae0d139ba2e760b286f23d260be2ea7a3e59.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": " denotes the output score, and " + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "inline_equation", + "content": "A_{cross}" + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": " is the attention weight from the cross attention layer of " + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{B_{text}}" + }, + { + "bbox": [ + 313, + 341, + 554, + 376 + ], + "type": "text", + "content": ", which captures the relationship between the image and the text." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 377, + 554, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 423 + ], + "type": "text", + "content": "To generate the Grad-CAM, the gradient is also important, so we conduct a backpropagation process to get the gradient " + }, + { + "bbox": [ + 313, + 377, + 554, + 423 + ], + "type": "inline_equation", + "content": "G_{cross}" + }, + { + "bbox": [ + 313, + 377, + 554, + 423 + ], + "type": "text", + "content": ", and then generate the prompt, as presented below:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 323, + 435, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 435, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 323, + 435, + 553, + 449 + ], + "type": "interline_equation", + "content": "\\mathrm {P} _ {\\text {i n i t}} = \\operatorname {G r a d C A M} = \\operatorname {M e a n} _ {\\text {t w i c e}} \\left(\\mathrm {G} _ {\\text {c r o s s}} \\odot \\mathrm {A} _ {\\text {c r o s s}}\\right), \\tag {3}", + "image_path": "495146849ab807df22bfadbff2fc664225c8e82589852798a10f44b9cafad619.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{init}" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " is the initially generated multimodal prompt, " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " denotes the element-wise multiplication operation, and " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{Mean}_{\\text{twice}}(\\cdot)" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": " performs two averaging steps, which average the influence of different text sequences on the image under each attention head, thereby obtaining the final " + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{P}_{init}" + }, + { + "bbox": [ + 313, + 457, + 554, + 518 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 474, + 538 + ], + "type": "text", + "content": "3.2. Adaptive Refinement Module" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 316, + 555, + 367, + 620 + ], + "blocks": [ + { + "bbox": [ + 316, + 555, + 367, + 620 + ], + "lines": [ + { + "bbox": [ + 316, + 555, + 367, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 367, + 620 + ], + "type": "image", + "image_path": "e0f946b2ef87161dd20044448373e14a5f38b4373779f4b11028f6af8180e855.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 330, + 623, + 351, + 632 + ], + "lines": [ + { + "bbox": [ + 330, + 623, + 351, + 632 + ], + "spans": [ + { + "bbox": [ + 330, + 623, + 351, + 632 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 377, + 555, + 428, + 620 + ], + "blocks": [ + { + "bbox": [ + 377, + 555, + 428, + 620 + ], + "lines": [ + { + "bbox": [ + 377, + 555, + 428, + 620 + ], + "spans": [ + { + "bbox": [ + 377, + 555, + 428, + 620 + ], + "type": "image", + "image_path": "4c7666dc81fcc9eab190ffb92d9bad235a29da0abbfa4eba51da7dc2e55f8902.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 623, + 408, + 631 + ], + "lines": [ + { + "bbox": [ + 396, + 623, + 408, + 631 + ], + "spans": [ + { + "bbox": [ + 396, + 623, + 408, + 631 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 438, + 555, + 490, + 620 + ], + "blocks": [ + { + "bbox": [ + 438, + 555, + 490, + 620 + ], + "lines": [ + { + "bbox": [ + 438, + 555, + 490, + 620 + ], + "spans": [ + { + "bbox": [ + 438, + 555, + 490, + 620 + ], + "type": "image", + "image_path": "55c1714e7040f27fb6a07f99bd6bb57cbeefaf7474932e0ed15692599e1ac8d7.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 454, + 623, + 474, + 631 + ], + "lines": [ + { + "bbox": [ + 454, + 623, + 474, + 631 + ], + "spans": [ + { + "bbox": [ + 454, + 623, + 474, + 631 + ], + "type": "text", + "content": "Initial" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 642, + 554, + 664 + ], + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 664 + ], + "type": "text", + "content": "Figure 3. Comparison between the initially generated prompt and the prompt refined using ARM." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 501, + 555, + 552, + 620 + ], + "blocks": [ + { + "bbox": [ + 501, + 555, + 552, + 620 + ], + "lines": [ + { + "bbox": [ + 501, + 555, + 552, + 620 + ], + "spans": [ + { + "bbox": [ + 501, + 555, + 552, + 620 + ], + "type": "image", + "image_path": "f6d202df113692182ce9599fac3445088ccb596b4c98756511065ecf06923f11.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 623, + 553, + 632 + ], + "lines": [ + { + "bbox": [ + 500, + 623, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 500, + 623, + 553, + 632 + ], + "type": "text", + "content": "Refined by ARM" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "As shown in Fig. 3, the initially generated multimodal prompt by BLIP only highlights the local regions of the camouflaged object rather than the entire object. As a" + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20675" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": "result, the mask prompt generated based on this multimodal prompt is incomplete, potentially negatively impacting SAM's performance. This locality issue is caused by domain shifts. When BLIP performs an image-text retrieval task, recognizing only the local region is sufficient to output the final matching score. Particularly in our work, the input text " + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "inline_equation", + "content": "t_{noun}" + }, + { + "bbox": [ + 55, + 72, + 296, + 252 + ], + "type": "text", + "content": " consists of only a single noun representing the object category without any additional descriptive details. Therefore, the initially generated prompt can only roughly indicate the object's location but cannot fully capture its structure, resulting in a lower-quality mask prompt. To address this issue, we design an Adaptive Refinement Module (ARM) to refine the initial incomplete prompt, enabling it to cover the object more entirely and generate a higher-quality mask prompt." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 257, + 133, + 267 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 133, + 267 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 133, + 267 + ], + "type": "text", + "content": "3.2.1. Refinement" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "content": "We first design adapters to extract features from the entire architecture. These adapters contain only two learnable layers, as shown below:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 314, + 295, + 328 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 314, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 67, + 314, + 295, + 328 + ], + "type": "interline_equation", + "content": "\\operatorname {A d a p t e r} (x) = \\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {2} \\left(\\operatorname {G E L U} \\left(\\operatorname {C o n v} _ {1} (x)\\right)\\right)\\right), \\tag {4}", + "image_path": "986b9aea084a8161dbc080810f4ba549eec6d3bac0749381687b7ef819778e7b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{Adapter}(\\cdot)" + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "content": " is an adapter module, " + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{Conv}_i(\\cdot)" + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "content": " two-dimensional convolutional layer and " + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "inline_equation", + "content": "\\mathrm{GELU}(\\cdot)" + }, + { + "bbox": [ + 55, + 334, + 296, + 370 + ], + "type": "text", + "content": ", the Gaussian Error Linear Unit [20], is an activation function." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "content": "In ARM, there are " + }, + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 371, + 296, + 407 + ], + "type": "text", + "content": " adapters used to capture image features from the two distinct image encoders and fuse them to generate a more comprehensive representation:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 411, + 295, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 411, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 61, + 411, + 295, + 436 + ], + "type": "interline_equation", + "content": "E m b _ {A} ^ {k} = \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {B _ {i m g}}\\right) + \\operatorname {A d a p t e r} ^ {k} \\left(E m b _ {S _ {i m g}} ^ {k}\\right), \\tag {5}", + "image_path": "2619c35d31ebeb0b9601a4ed7c3e0066920c780f9e87182275bdc2226a57d1ae.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "inline_equation", + "content": "Emb_A^k" + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "content": " represents the output embedding of the " + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "content": " adapter, and " + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "inline_equation", + "content": "Emb_{S_{img}}^k" + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "content": " is the image embedding comes from " + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 437, + 296, + 487 + ], + "type": "text", + "content": " transformer layer of SAM's image encoder. All adapters share the same weights." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "text", + "content": "To better utilize the extracted image features, the output of the " + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "text", + "content": " adapter is combined with that of the " + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "text", + "content": " transformer layer in SAM's image encoder and fed into the " + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "inline_equation", + "content": "(k + 1)^{th}" + }, + { + "bbox": [ + 55, + 487, + 296, + 548 + ], + "type": "text", + "content": " transformer layer, which could help fine-tune the SAM's image encoder:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 93, + 554, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 554, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 93, + 554, + 295, + 571 + ], + "type": "interline_equation", + "content": "E m b _ {S _ {i m g - i n}} ^ {k + 1} = E m b _ {A} ^ {k} + E m b _ {S _ {i m g - o u t}} ^ {k}, \\tag {6}", + "image_path": "b26e2204641b90ddac45543040de7b47f2f9df5cc2066a566c1a4daaee34a74f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "inline_equation", + "content": "Emb_{S_{img - in}}^{k + 1}" + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "content": " denotes to the inputs of the " + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "inline_equation", + "content": "(k + 1)^{th}" + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "content": " transformer layer in SAM's image encoder, and " + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "inline_equation", + "content": "Emb_{S_{img - out}}^k" + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "content": " is the original output from " + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "inline_equation", + "content": "k^{th}" + }, + { + "bbox": [ + 55, + 576, + 296, + 630 + ], + "type": "text", + "content": " transformer layer in SAM's image encoder." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Then, inspired by CLIP-ES [32], we extract the attention weight " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "A_{self}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}_{B_{img}}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " instead of training an additional affinity network to refine the initial prompt. By leveraging the relationships between image patches encoded in the attention weights, the prompt can adjust itself to identify high-correlation regions and expand to cover a more complete representation of the object. However, considering the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 143 + ], + "type": "text", + "content": "complexity of the COD, the background introduces significant noise, which can distort the attention map to capture the relationships between image patches accurately. Therefore, ARM uses the output of the final adapter, which has absorbed all the image information, to adjust the attention weights." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 378, + 144, + 554, + 158 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 378, + 144, + 554, + 158 + ], + "spans": [ + { + "bbox": [ + 378, + 144, + 554, + 158 + ], + "type": "interline_equation", + "content": "A _ {s e l f} ^ {*} = A _ {s e l f} \\odot E m b _ {A} ^ {l a s t}, \\tag {7}", + "image_path": "2447cd4f6b52f1834c66ee63e48a10b18cd763ccde87785f9a51aabdb550621d.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "inline_equation", + "content": "A_{\\text{self}}^{*}" + }, + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "text", + "content": " denotes the refined attention weights, and " + }, + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "inline_equation", + "content": "Emb_{A}^{\\text{last}}" + }, + { + "bbox": [ + 313, + 160, + 554, + 198 + ], + "type": "text", + "content": " represents the image embedding extracted from the final adapter module." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 234 + ], + "type": "text", + "content": "Subsequently, the masked attention weights are generated as affinity weights to refine the initial multimodal prompt:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 355, + 241, + 554, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 241, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 355, + 241, + 554, + 255 + ], + "type": "interline_equation", + "content": "\\mathrm {P} _ {\\text {r e f i n e}} = \\left(A _ {\\text {s e l f}} ^ {*} \\odot \\operatorname {v e c} (M)\\right) \\otimes \\mathrm {P} _ {\\text {i n i t}}, \\tag {8}", + "image_path": "f9276fdcbf20063dd551714f504d5b749b6e19c1067522129442513e9e65f092.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "vec(M)" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": " is the vectorization of mask " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": ", which are generated by extracting highlighted regions of the " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{init}" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": " is the matrix multiplication. Finally, by setting a threshold, the highlighted target region from the " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{refine}" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": " is extracted to generate a pseudo-mask, which serves as the final mask prompt " + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "inline_equation", + "content": "\\mathbf{P}_{mask}" + }, + { + "bbox": [ + 313, + 261, + 554, + 334 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 339, + 468, + 352 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 468, + 352 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 468, + 352 + ], + "type": "text", + "content": "3.3. Embedding Inputs for SAM" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 392 + ], + "type": "text", + "content": "After the mask prompt is generated, we will provide three inputs to SAM's mask decoder to support it in completing the COD task." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 393, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 393, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 393, + 554, + 429 + ], + "type": "text", + "content": "First, the primary generated mask prompt is encoded by SAM's prompt encoder " + }, + { + "bbox": [ + 313, + 393, + 554, + 429 + ], + "type": "inline_equation", + "content": "E_{S_{prompt}}" + }, + { + "bbox": [ + 313, + 393, + 554, + 429 + ], + "type": "text", + "content": ", ultimately producing a dense embedding:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 369, + 436, + 553, + 449 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 436, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 369, + 436, + 553, + 449 + ], + "type": "interline_equation", + "content": "E m b _ {d e n s e} = \\mathrm {E} _ {S _ {p r o m p t}} \\left(\\mathrm {P} _ {m a s k}\\right). \\tag {9}", + "image_path": "d590a6e4a3297885606a42e33efdbd8cc89b7619d2397fed276e3d0241602dc0.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 456, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 554, + 504 + ], + "type": "text", + "content": "Second, the image embedding produced by the final transformer layer of SAM's image encoder is combined with the visual features from the last adapter to generate the final image embedding:" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 361, + 510, + 553, + 526 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 510, + 553, + 526 + ], + "spans": [ + { + "bbox": [ + 361, + 510, + 553, + 526 + ], + "type": "interline_equation", + "content": "E m b _ {S _ {i m g}} ^ {\\text {o u t}} = E m b _ {A} ^ {\\text {l a s t}} + E m b _ {S _ {i m g}} ^ {\\text {l a s t}}. \\tag {10}", + "image_path": "4e3ea6ff60c82f4f2ce1790674c81ae1e0036f23d3269a198f21e81bb130577d.jpg" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "content": "Additionally, to make the most of the generated multiple information, we encode " + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "inline_equation", + "content": "t_{noun}" + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "content": " into the text embedding " + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "inline_equation", + "content": "Emb_{M_{text}}" + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "content": " using Mamba [11], a sequence modeling architecture known for its advanced ability to process textual information. The generated " + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "inline_equation", + "content": "Emb_{M_{text}}" + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "content": " clearly conveys the semantic information of the camouflaged target category and is concatenated with the " + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "inline_equation", + "content": "Emb_{B_{img}}" + }, + { + "bbox": [ + 313, + 531, + 554, + 637 + ], + "type": "text", + "content": " extracted from BLIP to generate the auxiliary sparse embedding to provide richer features:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 336, + 646, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 646, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 336, + 646, + 553, + 659 + ], + "type": "interline_equation", + "content": "E m b _ {A u x} = \\operatorname {C o n c a t} \\left(E m b _ {M _ {t e x t}}, E m b _ {B _ {i m g}}\\right), \\tag {11}", + "image_path": "1f1ec02e6960a63d3a8ca0ed274c0238966b1851800a64b25235a9049199bfef.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 666, + 514, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 514, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 514, + 677 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 313, + 666, + 514, + 677 + ], + "type": "inline_equation", + "content": "\\mathrm{Concat}(\\cdot)" + }, + { + "bbox": [ + 313, + 666, + 514, + 677 + ], + "type": "text", + "content": " is the concatenate operation." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Finally, " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "Emb_{dense}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", along with the " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "Emb_{Simg}^{out}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "Emb_{Aux}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", are input into SAM's mask decoder to obtain the final segmentation result." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20676" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 69, + 553, + 244 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 553, + 244 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 553, + 244 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 553, + 244 + ], + "type": "table", + "html": "
MethodPub. YearCHAMELEONCAMOCOD10KNC4K
SmFβmaxEφMSmFβmaxEφMSmFβmaxEφMSmFβmaxEφM
ZoomNet [42]CVPR220.9010.8760.9470.0230.8200.8050.8890.0660.8370.7770.8960.0290.8520.8260.9030.044
SegMaR [25]CVPR220.9060.8880.9590.0250.8160.8030.8840.0710.8330.7740.9060.0340.8410.8260.9070.046
UEDG [39]TMM230.9110.8940.9680.0230.8630.8560.9290.0480.8580.8120.9340.0250.8790.8640.9350.035
TinyCOD [58]ICASSP230.8870.8610.9580.0300.8220.8070.8990.0660.8110.7420.9030.0360.8430.8170.9100.047
MSCAF-Net [36]TCSVT230.9120.9020.9700.0220.8730.8670.9370.0460.8650.8230.9360.0240.8870.8740.9420.032
FSPNet [24]CVPR230.9090.8900.9650.0230.8560.8460.9280.0500.8510.7940.9300.0260.8790.8590.9370.035
FEDER [16]CVPR230.8870.8680.9540.0300.8020.7890.8730.0710.8220.7680.9050.0320.8470.8330.9150.044
FSNet [49]TIP230.9050.8910.9750.0220.8800.8780.9410.0410.8700.8330.9480.0230.8910.8800.9480.031
CRNet [19]AAAI230.8180.7920.9090.0460.7350.7070.8300.0920.7330.6360.8450.049————
SDRNet [13]KBS240.9140.9010.9610.0240.8720.8670.9320.0490.8710.8280.9360.0230.8890.8760.9400.032
PRNet [23]TCSVT240.9150.9020.9730.0200.8720.8670.9300.0500.8730.8390.9430.0220.8910.8810.9420.031
CamoFormer-R [59]TPAMI240.8980.8800.9490.0260.8170.8010.8830.0680.8380.7860.9280.0290.8540.8290.9080.042
CamoFormer-P [59]TPAMI240.9100.8980.9660.0220.8720.8680.9380.0460.8690.8290.9390.0230.8920.8800.9460.030
ZoomNeXtpv2b3 [43]TPAMI240.9280.9190.9770.0170.8850.8860.9420.0420.8950.8640.9510.0180.9000.8910.9480.028
ZoomNeXtpv2b2 [43]TPAMI240.9220.9080.9690.0170.8740.8730.9310.0470.8870.8560.9450.0190.8920.8840.9410.030
DSAM [60]ACMM24————0.8320.8340.9200.0610.8450.8050.9300.0330.8720.8640.9420.040
MAMIFNet [55]IF250.9140.8990.9590.0210.8720.8700.9350.0450.8690.8260.9400.0230.8900.8780.9430.031
COMPrompter [64]SCIS250.8850.8640.9570.0300.8530.8560.9310.0540.8600.8260.9460.0270.8800.8760.9460.036
Ours—0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.0200.9060.8970.9470.033
", + "image_path": "1a35930ad5dca829a62ecb0914e6ae3e441eaf5ad1302d38e9204368491b4256.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 274, + 205, + 392 + ], + "blocks": [ + { + "bbox": [ + 67, + 251, + 541, + 264 + ], + "lines": [ + { + "bbox": [ + 67, + 251, + 541, + 264 + ], + "spans": [ + { + "bbox": [ + 67, + 251, + 541, + 264 + ], + "type": "text", + "content": "Table 1. Quantitative comparison with other SOTA methods across different datasets. The best results are highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 274, + 205, + 392 + ], + "lines": [ + { + "bbox": [ + 58, + 274, + 205, + 392 + ], + "spans": [ + { + "bbox": [ + 58, + 274, + 205, + 392 + ], + "type": "image", + "image_path": "02641099111cdf0262d4c159d7a66afaee13fa3e3b4b6965df5a4e203f9047db.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 230, + 274, + 376, + 392 + ], + "blocks": [ + { + "bbox": [ + 230, + 274, + 376, + 392 + ], + "lines": [ + { + "bbox": [ + 230, + 274, + 376, + 392 + ], + "spans": [ + { + "bbox": [ + 230, + 274, + 376, + 392 + ], + "type": "image", + "image_path": "d2d4f7ae988551f9c6c24a05f35675d8e54c8d22b97738619bacd2879db67628.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 402, + 274, + 550, + 392 + ], + "blocks": [ + { + "bbox": [ + 402, + 274, + 550, + 392 + ], + "lines": [ + { + "bbox": [ + 402, + 274, + 550, + 392 + ], + "spans": [ + { + "bbox": [ + 402, + 274, + 550, + 392 + ], + "type": "image", + "image_path": "6600128e05468b003601c1e71e6935e069f87abd042f645e922233193518a4df.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 399, + 208, + 517 + ], + "blocks": [ + { + "bbox": [ + 59, + 399, + 208, + 517 + ], + "lines": [ + { + "bbox": [ + 59, + 399, + 208, + 517 + ], + "spans": [ + { + "bbox": [ + 59, + 399, + 208, + 517 + ], + "type": "image", + "image_path": "b76efdc460cc9b6c856481aba53e7db3c8b4d7f07dfb415e393fe5951a623dba.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 157, + 525, + 451, + 537 + ], + "lines": [ + { + "bbox": [ + 157, + 525, + 451, + 537 + ], + "spans": [ + { + "bbox": [ + 157, + 525, + 451, + 537 + ], + "type": "text", + "content": "Figure 4. " + }, + { + "bbox": [ + 157, + 525, + 451, + 537 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 157, + 525, + 451, + 537 + ], + "type": "text", + "content": " and PR curves of our work compared to other SOTA methods." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 230, + 399, + 380, + 516 + ], + "blocks": [ + { + "bbox": [ + 230, + 399, + 380, + 516 + ], + "lines": [ + { + "bbox": [ + 230, + 399, + 380, + 516 + ], + "spans": [ + { + "bbox": [ + 230, + 399, + 380, + 516 + ], + "type": "image", + "image_path": "f3c6840d9d3725a06ddd38fc64291a34ede3bb546301309a6b94338575153432.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 403, + 399, + 552, + 516 + ], + "blocks": [ + { + "bbox": [ + 403, + 399, + 552, + 516 + ], + "lines": [ + { + "bbox": [ + 403, + 399, + 552, + 516 + ], + "spans": [ + { + "bbox": [ + 403, + 399, + 552, + 516 + ], + "type": "image", + "image_path": "591d62dc81d4ab271906060bf90b4a12a3787020f10ed73d2d47456959b48049.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 135, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 135, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 135, + 571 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 171, + 596 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 171, + 596 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 171, + 596 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "Datasets. To verify the effectiveness of the proposed architecture in the COD task, we use four commonly used COD datasets for evaluation: CHAMELEON [48], CAMO [28], COD10K [10], and NC4K [38]. In detail, CHAMELEON consists of 76 images featuring highly camouflaged animals in natural environments, all used for testing. CAMO contains 1250 images, including both biological camouflage and artificially designed camouflage. Among them, 1000 images are used for training, while the rest are reserved" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 641 + ], + "type": "text", + "content": "for testing. COD10K is more extensive, comprising 5066 images that include both artificial and natural camouflage, with 3040 images allocated for training and the remainder used for testing. While NC4K comprises 4,121 testing images specifically designed for the COD task. Notably, all four datasets provide only image data and do not include additional modalities, such as text." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. In the subsequent experiments, we mainly use four metrics for evaluation: S-measure [8] " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "(Sm)" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", maximum F-measure [1] " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "(F_{\\beta}^{max})" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", maximum E-measure [9] " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "(E_{\\varphi}^{max})" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", and Mean Absolute Error " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "(M)" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ". Among these, higher values of the " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "Sm" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "E_{\\varphi}^{max}" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20677" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 110, + 122 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 110, + 122 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 110, + 122 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 110, + 122 + ], + "type": "image", + "image_path": "74c8575355ce31e2fdf9d320520eb450d5219d8f9cdbfce68cfb7cfb0cbac8fd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 112, + 70, + 165, + 122 + ], + "blocks": [ + { + "bbox": [ + 112, + 70, + 165, + 122 + ], + "lines": [ + { + "bbox": [ + 112, + 70, + 165, + 122 + ], + "spans": [ + { + "bbox": [ + 112, + 70, + 165, + 122 + ], + "type": "image", + "image_path": "a4f26e241fc3d48d7f1813d74b6e0dba4b906dbe58484f652853a7720606df79.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 168, + 70, + 220, + 121 + ], + "blocks": [ + { + "bbox": [ + 168, + 70, + 220, + 121 + ], + "lines": [ + { + "bbox": [ + 168, + 70, + 220, + 121 + ], + "spans": [ + { + "bbox": [ + 168, + 70, + 220, + 121 + ], + "type": "image", + "image_path": "ab8f8ed22efe07310571b887f51542350ce6721fc1a035a612e6253b69583d08.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 224, + 70, + 276, + 121 + ], + "blocks": [ + { + "bbox": [ + 224, + 70, + 276, + 121 + ], + "lines": [ + { + "bbox": [ + 224, + 70, + 276, + 121 + ], + "spans": [ + { + "bbox": [ + 224, + 70, + 276, + 121 + ], + "type": "image", + "image_path": "69f387c467cc62a7c1f9aef7465865cc850abc79c9c98c2764bc0cb66ec09ca7.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 279, + 70, + 331, + 121 + ], + "blocks": [ + { + "bbox": [ + 279, + 70, + 331, + 121 + ], + "lines": [ + { + "bbox": [ + 279, + 70, + 331, + 121 + ], + "spans": [ + { + "bbox": [ + 279, + 70, + 331, + 121 + ], + "type": "image", + "image_path": "5a029378a929356983f01b74b51ed45e4d41ff688125d94c9e3104db254f7657.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 335, + 71, + 386, + 121 + ], + "blocks": [ + { + "bbox": [ + 335, + 71, + 386, + 121 + ], + "lines": [ + { + "bbox": [ + 335, + 71, + 386, + 121 + ], + "spans": [ + { + "bbox": [ + 335, + 71, + 386, + 121 + ], + "type": "image", + "image_path": "ecf46666e03eae29667354466666f7aa20b6bd79d76e4ec88add95ad99014380.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 391, + 71, + 441, + 121 + ], + "blocks": [ + { + "bbox": [ + 391, + 71, + 441, + 121 + ], + "lines": [ + { + "bbox": [ + 391, + 71, + 441, + 121 + ], + "spans": [ + { + "bbox": [ + 391, + 71, + 441, + 121 + ], + "type": "image", + "image_path": "d9508d34edeac472c55f7551a6aa193a7f9d9debd28404c63a5dc7dc95ee8132.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 445, + 71, + 497, + 121 + ], + "blocks": [ + { + "bbox": [ + 445, + 71, + 497, + 121 + ], + "lines": [ + { + "bbox": [ + 445, + 71, + 497, + 121 + ], + "spans": [ + { + "bbox": [ + 445, + 71, + 497, + 121 + ], + "type": "image", + "image_path": "1ccda65e372f20b5395cd6e70c0856e85db95a257c774291fc02b05ca64fc11b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 501, + 71, + 553, + 121 + ], + "blocks": [ + { + "bbox": [ + 501, + 71, + 553, + 121 + ], + "lines": [ + { + "bbox": [ + 501, + 71, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 501, + 71, + 553, + 121 + ], + "type": "image", + "image_path": "0b6b4e9c96cd655459fcda4f1d870dc12ebab1ab493f15d9d079762373b442d5.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 125, + 110, + 178 + ], + "blocks": [ + { + "bbox": [ + 58, + 125, + 110, + 178 + ], + "lines": [ + { + "bbox": [ + 58, + 125, + 110, + 178 + ], + "spans": [ + { + "bbox": [ + 58, + 125, + 110, + 178 + ], + "type": "image", + "image_path": "fb600f8f53b41ef36b1133e4b04a82c9e09213b5ef78d86dcf5cbcfb5d67f207.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 112, + 125, + 165, + 178 + ], + "blocks": [ + { + "bbox": [ + 112, + 125, + 165, + 178 + ], + "lines": [ + { + "bbox": [ + 112, + 125, + 165, + 178 + ], + "spans": [ + { + "bbox": [ + 112, + 125, + 165, + 178 + ], + "type": "image", + "image_path": "b7682f308ab913ae53d398da451e5533d4868bddc380c37506a0c07f4fbe02ee.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 168, + 125, + 220, + 178 + ], + "blocks": [ + { + "bbox": [ + 168, + 125, + 220, + 178 + ], + "lines": [ + { + "bbox": [ + 168, + 125, + 220, + 178 + ], + "spans": [ + { + "bbox": [ + 168, + 125, + 220, + 178 + ], + "type": "image", + "image_path": "afe1446bf035a6836f665cc3832462da8af125919dd69e135d01a2a78d0b2753.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 224, + 125, + 276, + 178 + ], + "blocks": [ + { + "bbox": [ + 224, + 125, + 276, + 178 + ], + "lines": [ + { + "bbox": [ + 224, + 125, + 276, + 178 + ], + "spans": [ + { + "bbox": [ + 224, + 125, + 276, + 178 + ], + "type": "image", + "image_path": "3bca2ccd10b949f19779e1d0c9bb679fa58013fb121c582cea93cbb79ccff60a.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 279, + 125, + 331, + 178 + ], + "blocks": [ + { + "bbox": [ + 279, + 125, + 331, + 178 + ], + "lines": [ + { + "bbox": [ + 279, + 125, + 331, + 178 + ], + "spans": [ + { + "bbox": [ + 279, + 125, + 331, + 178 + ], + "type": "image", + "image_path": "3dd6fc12da9b18382674a264b19bc070583d5c5615811afb35582f589b25a774.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 335, + 125, + 386, + 178 + ], + "blocks": [ + { + "bbox": [ + 335, + 125, + 386, + 178 + ], + "lines": [ + { + "bbox": [ + 335, + 125, + 386, + 178 + ], + "spans": [ + { + "bbox": [ + 335, + 125, + 386, + 178 + ], + "type": "image", + "image_path": "c269fd63974c9c55200c7ee70b7bc571dc77782e884f7182b2f0d0d1a60c510b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 391, + 125, + 441, + 178 + ], + "blocks": [ + { + "bbox": [ + 391, + 125, + 441, + 178 + ], + "lines": [ + { + "bbox": [ + 391, + 125, + 441, + 178 + ], + "spans": [ + { + "bbox": [ + 391, + 125, + 441, + 178 + ], + "type": "image", + "image_path": "c310ab2934aa67d9ff9a8f286fc0c15f5aacb3b6ac0f789656f41b5616b70952.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 445, + 125, + 497, + 178 + ], + "blocks": [ + { + "bbox": [ + 445, + 125, + 497, + 178 + ], + "lines": [ + { + "bbox": [ + 445, + 125, + 497, + 178 + ], + "spans": [ + { + "bbox": [ + 445, + 125, + 497, + 178 + ], + "type": "image", + "image_path": "ee13013b871a79d6efe1c2729d938335821a2c99594bb6c3b824b99a0cc2a470.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 501, + 125, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 501, + 125, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 501, + 125, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 501, + 125, + 553, + 178 + ], + "type": "image", + "image_path": "b757697e323d173e790123f065a93d993420c70abd04844bcd2ad7af67c4f4d7.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 58, + 182, + 110, + 235 + ], + "blocks": [ + { + "bbox": [ + 58, + 182, + 110, + 235 + ], + "lines": [ + { + "bbox": [ + 58, + 182, + 110, + 235 + ], + "spans": [ + { + "bbox": [ + 58, + 182, + 110, + 235 + ], + "type": "image", + "image_path": "08f1ad7ec29e475d5c77c759b74f5be5237adeead3d5afac22a6ce13d33db15e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 112, + 182, + 165, + 235 + ], + "blocks": [ + { + "bbox": [ + 112, + 182, + 165, + 235 + ], + "lines": [ + { + "bbox": [ + 112, + 182, + 165, + 235 + ], + "spans": [ + { + "bbox": [ + 112, + 182, + 165, + 235 + ], + "type": "image", + "image_path": "652d58dd61d770cb3d8730c32dace42e26640522778f43d7294e5b655c76f78e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 168, + 182, + 220, + 235 + ], + "blocks": [ + { + "bbox": [ + 168, + 182, + 220, + 235 + ], + "lines": [ + { + "bbox": [ + 168, + 182, + 220, + 235 + ], + "spans": [ + { + "bbox": [ + 168, + 182, + 220, + 235 + ], + "type": "image", + "image_path": "d01364f9fa35dad1aaa1a0c1f333c230d64f4a92e1b40f5d5ad338b8c4240321.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 224, + 182, + 276, + 235 + ], + "blocks": [ + { + "bbox": [ + 224, + 182, + 276, + 235 + ], + "lines": [ + { + "bbox": [ + 224, + 182, + 276, + 235 + ], + "spans": [ + { + "bbox": [ + 224, + 182, + 276, + 235 + ], + "type": "image", + "image_path": "2c4d4bd31e2d112e685d56835932188124fafd11dfc5ed2be0a1b6bd6cee7722.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 279, + 182, + 331, + 235 + ], + "blocks": [ + { + "bbox": [ + 279, + 182, + 331, + 235 + ], + "lines": [ + { + "bbox": [ + 279, + 182, + 331, + 235 + ], + "spans": [ + { + "bbox": [ + 279, + 182, + 331, + 235 + ], + "type": "image", + "image_path": "18283c3e19110b5b6084c8c08ec2a9eb502254a49a42360b28b272d48383fcb5.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 335, + 182, + 386, + 235 + ], + "blocks": [ + { + "bbox": [ + 335, + 182, + 386, + 235 + ], + "lines": [ + { + "bbox": [ + 335, + 182, + 386, + 235 + ], + "spans": [ + { + "bbox": [ + 335, + 182, + 386, + 235 + ], + "type": "image", + "image_path": "1d47931c7f5375bbc4f8b5522d1faa255d089ff85aba4d6b2295ae76cdd894c0.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 391, + 182, + 441, + 235 + ], + "blocks": [ + { + "bbox": [ + 391, + 182, + 441, + 235 + ], + "lines": [ + { + "bbox": [ + 391, + 182, + 441, + 235 + ], + "spans": [ + { + "bbox": [ + 391, + 182, + 441, + 235 + ], + "type": "image", + "image_path": "903c58fd611f2580a6f225a83ef5938764e08c5769ff8913a161a9d432b9476f.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 445, + 182, + 497, + 235 + ], + "blocks": [ + { + "bbox": [ + 445, + 182, + 497, + 235 + ], + "lines": [ + { + "bbox": [ + 445, + 182, + 497, + 235 + ], + "spans": [ + { + "bbox": [ + 445, + 182, + 497, + 235 + ], + "type": "image", + "image_path": "7c832920d4c9c8215b03be14a922777c33aec278aa387bebc3a5450d654f122f.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 501, + 182, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 501, + 182, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 501, + 182, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 501, + 182, + 553, + 235 + ], + "type": "image", + "image_path": "16f4d64288f2fd1070515fa0f02c0a1ba10fb14566e1002047277e97996d824d.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 58, + 237, + 110, + 289 + ], + "blocks": [ + { + "bbox": [ + 58, + 237, + 110, + 289 + ], + "lines": [ + { + "bbox": [ + 58, + 237, + 110, + 289 + ], + "spans": [ + { + "bbox": [ + 58, + 237, + 110, + 289 + ], + "type": "image", + "image_path": "b95ec0903d37511d43a43b3d122dbf783e444810395e5b0e175e424766787448.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 291, + 93, + 300 + ], + "lines": [ + { + "bbox": [ + 74, + 291, + 93, + 300 + ], + "spans": [ + { + "bbox": [ + 74, + 291, + 93, + 300 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 112, + 238, + 165, + 289 + ], + "blocks": [ + { + "bbox": [ + 112, + 238, + 165, + 289 + ], + "lines": [ + { + "bbox": [ + 112, + 238, + 165, + 289 + ], + "spans": [ + { + "bbox": [ + 112, + 238, + 165, + 289 + ], + "type": "image", + "image_path": "50c2e4853535d36f32580ae080100896f8f71a6a2db64bc24d93249392f5830c.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 134, + 291, + 145, + 299 + ], + "lines": [ + { + "bbox": [ + 134, + 291, + 145, + 299 + ], + "spans": [ + { + "bbox": [ + 134, + 291, + 145, + 299 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 168, + 238, + 220, + 289 + ], + "blocks": [ + { + "bbox": [ + 168, + 238, + 220, + 289 + ], + "lines": [ + { + "bbox": [ + 168, + 238, + 220, + 289 + ], + "spans": [ + { + "bbox": [ + 168, + 238, + 220, + 289 + ], + "type": "image", + "image_path": "a065d746fd94323fd4e9743e6a3837638d3508839bed335401374a41bc7cdc0f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 186, + 291, + 201, + 299 + ], + "lines": [ + { + "bbox": [ + 186, + 291, + 201, + 299 + ], + "spans": [ + { + "bbox": [ + 186, + 291, + 201, + 299 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 224, + 238, + 276, + 289 + ], + "blocks": [ + { + "bbox": [ + 224, + 238, + 276, + 289 + ], + "lines": [ + { + "bbox": [ + 224, + 238, + 276, + 289 + ], + "spans": [ + { + "bbox": [ + 224, + 238, + 276, + 289 + ], + "type": "image", + "image_path": "1b46b78b13e1b9729d5e30566c9ec126ba07274705c5c0cf0dd4596502895108.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 291, + 272, + 300 + ], + "lines": [ + { + "bbox": [ + 228, + 291, + 272, + 300 + ], + "spans": [ + { + "bbox": [ + 228, + 291, + 272, + 300 + ], + "type": "text", + "content": "COMPrompter" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 279, + 238, + 331, + 289 + ], + "blocks": [ + { + "bbox": [ + 279, + 238, + 331, + 289 + ], + "lines": [ + { + "bbox": [ + 279, + 238, + 331, + 289 + ], + "spans": [ + { + "bbox": [ + 279, + 238, + 331, + 289 + ], + "type": "image", + "image_path": "e0f917b7ec1d27e6998821f8ab5a5505fdcf782b43a0bb663bd31459ca581031.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 288, + 291, + 321, + 299 + ], + "lines": [ + { + "bbox": [ + 288, + 291, + 321, + 299 + ], + "spans": [ + { + "bbox": [ + 288, + 291, + 321, + 299 + ], + "type": "text", + "content": "ZoomNeXt" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_caption" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 335, + 238, + 386, + 289 + ], + "blocks": [ + { + "bbox": [ + 335, + 238, + 386, + 289 + ], + "lines": [ + { + "bbox": [ + 335, + 238, + 386, + 289 + ], + "spans": [ + { + "bbox": [ + 335, + 238, + 386, + 289 + ], + "type": "image", + "image_path": "77938d9574bb33dc48fbc47b8490a06f14619d18f34b58f834a448db922588f7.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 291, + 368, + 299 + ], + "lines": [ + { + "bbox": [ + 347, + 291, + 368, + 299 + ], + "spans": [ + { + "bbox": [ + 347, + 291, + 368, + 299 + ], + "type": "text", + "content": "PRNet" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 112, + 308, + 496, + 320 + ], + "lines": [ + { + "bbox": [ + 112, + 308, + 496, + 320 + ], + "spans": [ + { + "bbox": [ + 112, + 308, + 496, + 320 + ], + "type": "text", + "content": "Figure 5. Visual comparison of segmentation results with other SOTA methods on different samples." + } + ] + } + ], + "index": 45, + "angle": 0, + "type": "image_caption" + } + ], + "index": 37 + }, + { + "type": "image", + "bbox": [ + 391, + 238, + 441, + 289 + ], + "blocks": [ + { + "bbox": [ + 391, + 238, + 441, + 289 + ], + "lines": [ + { + "bbox": [ + 391, + 238, + 441, + 289 + ], + "spans": [ + { + "bbox": [ + 391, + 238, + 441, + 289 + ], + "type": "image", + "image_path": "b1a80939311555b022867bfb6dd91248b884705d4077cd4ac9bba3036bca0fbf.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 402, + 291, + 422, + 299 + ], + "lines": [ + { + "bbox": [ + 402, + 291, + 422, + 299 + ], + "spans": [ + { + "bbox": [ + 402, + 291, + 422, + 299 + ], + "type": "text", + "content": "FSNet" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "image", + "bbox": [ + 445, + 238, + 497, + 289 + ], + "blocks": [ + { + "bbox": [ + 445, + 238, + 497, + 289 + ], + "lines": [ + { + "bbox": [ + 445, + 238, + 497, + 289 + ], + "spans": [ + { + "bbox": [ + 445, + 238, + 497, + 289 + ], + "type": "image", + "image_path": "d89ace3c2f3d616640d46d23530e6f8f6ae8d65de8fae9280af7ff96ae879a21.jpg" + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 457, + 291, + 480, + 299 + ], + "lines": [ + { + "bbox": [ + 457, + 291, + 480, + 299 + ], + "spans": [ + { + "bbox": [ + 457, + 291, + 480, + 299 + ], + "type": "text", + "content": "FSPNet" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "image_caption" + } + ], + "index": 41 + }, + { + "type": "image", + "bbox": [ + 501, + 238, + 553, + 289 + ], + "blocks": [ + { + "bbox": [ + 501, + 238, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 501, + 238, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 501, + 238, + 553, + 289 + ], + "type": "image", + "image_path": "71be723cb02203bee17bdbec6fcc278860063392b3ca9ce504da644b18ee5edd.jpg" + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 291, + 541, + 299 + ], + "lines": [ + { + "bbox": [ + 512, + 291, + 541, + 299 + ], + "spans": [ + { + "bbox": [ + 512, + 291, + 541, + 299 + ], + "type": "text", + "content": "ZoomNet" + } + ] + } + ], + "index": 44, + "angle": 0, + "type": "image_caption" + } + ], + "index": 43 + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "text", + "content": "indicate better model performance, whereas for " + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "text", + "content": ", lower values indicate better performance. Additionally, the precision-recall curve (PR curve) and the " + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "text", + "content": "-threshold curve " + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "inline_equation", + "content": "(F_{\\beta}" + }, + { + "bbox": [ + 55, + 341, + 296, + 390 + ], + "type": "text", + "content": " curve) are also considered for further evaluation." + } + ] + } + ], + "index": 46 + }, + { + "bbox": [ + 55, + 395, + 296, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 296, + 528 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 296, + 528 + ], + "type": "text", + "content": "Implement Details. Our proposed framework is primarily built upon the SAM-Vit Huge (ViT-H) [27] and implemented using PyTorch. The entire architecture is trained on an NVIDIA GeForce RTX 4090 GPU for 40 epochs with a batch size of 1. The main learning rate is set to 0.0002, while the learning rate for the ARM is 0.1 times that of the main learning rate, and the AdamW optimizer is used for optimization. Furthermore, both the VLM and LLM, used for multimodal prompt and auxiliary embedding generation, are pre-trained models, so no additional training or fine-tuning is required for these two." + } + ] + } + ], + "index": 47 + }, + { + "bbox": [ + 55, + 534, + 285, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 285, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 285, + 548 + ], + "type": "text", + "content": "4.2. Comparisons With State-of-the-art Methods" + } + ] + } + ], + "index": 48 + }, + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 296, + 588 + ], + "type": "text", + "content": "To effectively evaluate the performance of our work in the COD task, we first conducted a comparative analysis with the current state-of-the-art (SOTA) models." + } + ] + } + ], + "index": 49 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Evaluation. As shown in Tab. 1, which compares the evaluation results of our work and other SOTA models across four different datasets. Our work generally demonstrates exceptional performance in solving the COD task across various metrics on different datasets. In particular, our method excels in both the " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "Sm" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", showing superior performance in target localization and structure capture of the camouflaged objects in complex scenarios. For example, on the COD10K, our approach outperforms the second-best method by " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "Sm" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "2.4\\%" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " in" + } + ] + } + ], + "index": 50 + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "text", + "content": ". Although our " + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "inline_equation", + "content": "F_{\\beta}^{max}" + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "text", + "content": " is the second on the CAMO, the gap with the top-performing model is only " + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "inline_equation", + "content": "0.3\\%" + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "text", + "content": ", which is negligible. Furthermore, from the comparison of the " + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "inline_equation", + "content": "F_{\\beta}" + }, + { + "bbox": [ + 313, + 341, + 555, + 425 + ], + "type": "text", + "content": " and PR curves in Fig. 4, it is clear that our method (red solid curve) exhibits outstanding performance across all three datasets, with particularly remarkable results on the COD10K dataset." + } + ] + } + ], + "index": 51 + }, + { + "bbox": [ + 313, + 431, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 556, + 612 + ], + "type": "text", + "content": "Qualitative Evaluation Fig. 5 illustrates a comparison of the final segmentation maps of our work and other SOTA methods across different samples. It is clear that our approach achieves a more significant overlap with the ground truth (GT) than other SOTA models. Specifically, our method is more effective at accurately segmenting and capturing completely camouflaged objects. Even in more challenging scenarios, such as the third sample in Fig. 5, where multiple small camouflaged objects need to be identified, our approach segments more objects and provides more accurate segmentation, while the results from other models show noticeable omissions. These observations highlight the superior performance of our framework, not only in localizing camouflaged objects but also in maintaining the structural integrity of the camouflaged objects." + } + ] + } + ], + "index": 52 + }, + { + "bbox": [ + 313, + 618, + 416, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 416, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 416, + 629 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 53 + }, + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 555, + 672 + ], + "type": "text", + "content": "As shown in Tab. 2, we conduct ablation studies on the improvements discussed in Sec. 3 to assess the effectiveness of each component." + } + ] + } + ], + "index": 54 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Effectiveness of Auxiliary Embedding. Since the auxiliary embedding is generated alongside the multimodal data produced by BLIP, we first incorporate it into the SAM to" + } + ] + } + ], + "index": 55 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20678" + } + ] + } + ], + "index": 56 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 149 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 149 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 149 + ], + "type": "table", + "html": "
CHAMELEONCAMOCOD10K
SmFβmaxEφmaxMSmFβmaxEφmaxMSmFβmaxEφmaxM
SAM0.8500.8200.9110.0560.7970.7640.8640.1010.8100.7500.8810.054
SAM + EmbAUX0.8520.8160.9220.0470.8050.7780.8820.0890.8200.7510.8930.039
SAM + EmbAUX + Pcaptioninit0.8660.8360.9200.0460.8230.7950.8910.0810.8390.7760.9080.037
SAM + EmbAUX + Pnouninit0.8820.8590.9350.0400.8310.8070.8960.0800.8460.7890.9140.036
SAM + EmbAUX + PnounARM0.9320.9220.9640.0230.8870.8830.9350.0460.9090.8850.9550.020
", + "image_path": "c22e13e234176b3d8b920876a8d73098edac80d64ba1ab3a670daedf45226ea3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "lines": [ + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": "Table 2. Ablation studies on three different COD datasets. + denotes additional information provided to SAM, " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "Emb_{Aux}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " is the auxiliary sparse embedding, " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "P_{init}^{caption}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " is the initial mask prompt generated by the entire caption, " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "P_{init}^{noun}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " is the initial mask prompt generated by the extracted noun, " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "P_{ARM}^{noun}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " is the refined " + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "inline_equation", + "content": "P_{init}^{noun}" + }, + { + "bbox": [ + 55, + 157, + 555, + 191 + ], + "type": "text", + "content": " by ARM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 58, + 210, + 115, + 255 + ], + "blocks": [ + { + "bbox": [ + 58, + 210, + 115, + 255 + ], + "lines": [ + { + "bbox": [ + 58, + 210, + 115, + 255 + ], + "spans": [ + { + "bbox": [ + 58, + 210, + 115, + 255 + ], + "type": "image", + "image_path": "96f6bd4d5374a0023968bf3ec8648745c9f51c7a9e15aa2606c5af2d22986ebb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 257, + 115, + 300 + ], + "blocks": [ + { + "bbox": [ + 58, + 257, + 115, + 300 + ], + "lines": [ + { + "bbox": [ + 58, + 257, + 115, + 300 + ], + "spans": [ + { + "bbox": [ + 58, + 257, + 115, + 300 + ], + "type": "image", + "image_path": "e0ad3fdf6ce28e4641430ece9c327ddc13b8dbe9ec966323448d944d6ee69132.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 58, + 303, + 115, + 338 + ], + "blocks": [ + { + "bbox": [ + 58, + 303, + 115, + 338 + ], + "lines": [ + { + "bbox": [ + 58, + 303, + 115, + 338 + ], + "spans": [ + { + "bbox": [ + 58, + 303, + 115, + 338 + ], + "type": "image", + "image_path": "9e02f4db4184054407802e9b3a946578375a2456bd82aa3048e9034244fde7e7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 341, + 115, + 376 + ], + "blocks": [ + { + "bbox": [ + 58, + 341, + 115, + 376 + ], + "lines": [ + { + "bbox": [ + 58, + 341, + 115, + 376 + ], + "spans": [ + { + "bbox": [ + 58, + 341, + 115, + 376 + ], + "type": "image", + "image_path": "3c9ace7879d2e460dcdedbdf5daad351fe0c6225bc00500908870bb0ff583dce.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 380, + 115, + 421 + ], + "blocks": [ + { + "bbox": [ + 58, + 380, + 115, + 421 + ], + "lines": [ + { + "bbox": [ + 58, + 380, + 115, + 421 + ], + "spans": [ + { + "bbox": [ + 58, + 380, + 115, + 421 + ], + "type": "image", + "image_path": "f2d92d68d01c47cb3b7a17a7c6132d1fc87084f3e1fd3973f2a35f36cddaab1d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 78, + 424, + 96, + 432 + ], + "lines": [ + { + "bbox": [ + 78, + 424, + 96, + 432 + ], + "spans": [ + { + "bbox": [ + 78, + 424, + 96, + 432 + ], + "type": "text", + "content": "Image" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 117, + 211, + 175, + 255 + ], + "blocks": [ + { + "bbox": [ + 117, + 211, + 175, + 255 + ], + "lines": [ + { + "bbox": [ + 117, + 211, + 175, + 255 + ], + "spans": [ + { + "bbox": [ + 117, + 211, + 175, + 255 + ], + "type": "image", + "image_path": "b3ae903b0735e5236a189dfd0d8b1bbb9f1e94c9a7d3819a4bd90e74d6faa3f7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 118, + 257, + 175, + 300 + ], + "blocks": [ + { + "bbox": [ + 118, + 257, + 175, + 300 + ], + "lines": [ + { + "bbox": [ + 118, + 257, + 175, + 300 + ], + "spans": [ + { + "bbox": [ + 118, + 257, + 175, + 300 + ], + "type": "image", + "image_path": "bc5237428dd668c32af7793d32dfac700b9a4bda0cc33ab45f1702f171f7f2a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 118, + 303, + 175, + 338 + ], + "blocks": [ + { + "bbox": [ + 118, + 303, + 175, + 338 + ], + "lines": [ + { + "bbox": [ + 118, + 303, + 175, + 338 + ], + "spans": [ + { + "bbox": [ + 118, + 303, + 175, + 338 + ], + "type": "image", + "image_path": "accd0adc389c776c939aac9a94d508e4c4f8f81047dd0fb6545c27875463e72c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 118, + 341, + 175, + 376 + ], + "blocks": [ + { + "bbox": [ + 118, + 341, + 175, + 376 + ], + "lines": [ + { + "bbox": [ + 118, + 341, + 175, + 376 + ], + "spans": [ + { + "bbox": [ + 118, + 341, + 175, + 376 + ], + "type": "image", + "image_path": "96d9d34953ce2a0a72507e3c6af22086615e89e9daf4ee0db05a9ea3735dff13.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 117, + 380, + 175, + 420 + ], + "blocks": [ + { + "bbox": [ + 117, + 380, + 175, + 420 + ], + "lines": [ + { + "bbox": [ + 117, + 380, + 175, + 420 + ], + "spans": [ + { + "bbox": [ + 117, + 380, + 175, + 420 + ], + "type": "image", + "image_path": "9a2b87cad6a2461aafa78aa5689b4e660d89c42549587ae483f3d7d74e1087a2.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 141, + 424, + 152, + 431 + ], + "lines": [ + { + "bbox": [ + 141, + 424, + 152, + 431 + ], + "spans": [ + { + "bbox": [ + 141, + 424, + 152, + 431 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 176, + 211, + 234, + 255 + ], + "blocks": [ + { + "bbox": [ + 176, + 211, + 234, + 255 + ], + "lines": [ + { + "bbox": [ + 176, + 211, + 234, + 255 + ], + "spans": [ + { + "bbox": [ + 176, + 211, + 234, + 255 + ], + "type": "image", + "image_path": "3ac3e62534fc3a5b73d9df592601fb972982eaef8548d0db14a84aa8ef79fdbd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 176, + 257, + 234, + 300 + ], + "blocks": [ + { + "bbox": [ + 176, + 257, + 234, + 300 + ], + "lines": [ + { + "bbox": [ + 176, + 257, + 234, + 300 + ], + "spans": [ + { + "bbox": [ + 176, + 257, + 234, + 300 + ], + "type": "image", + "image_path": "41bd62e19c8b1a689260eb1648bb12cc980e8b545346a9e5c26076b70d445761.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 176, + 303, + 234, + 338 + ], + "blocks": [ + { + "bbox": [ + 176, + 303, + 234, + 338 + ], + "lines": [ + { + "bbox": [ + 176, + 303, + 234, + 338 + ], + "spans": [ + { + "bbox": [ + 176, + 303, + 234, + 338 + ], + "type": "image", + "image_path": "7a2380f758b652612400f1a056636b16c03ab261c3405726bc6c531be1feda1b.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 176, + 341, + 234, + 376 + ], + "blocks": [ + { + "bbox": [ + 176, + 341, + 234, + 376 + ], + "lines": [ + { + "bbox": [ + 176, + 341, + 234, + 376 + ], + "spans": [ + { + "bbox": [ + 176, + 341, + 234, + 376 + ], + "type": "image", + "image_path": "19d4394c6805bb428218f9cad76f9d7a2e5ded53e44ee5bbba46d1148a8b1062.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 176, + 380, + 234, + 420 + ], + "blocks": [ + { + "bbox": [ + 176, + 380, + 234, + 420 + ], + "lines": [ + { + "bbox": [ + 176, + 380, + 234, + 420 + ], + "spans": [ + { + "bbox": [ + 176, + 380, + 234, + 420 + ], + "type": "image", + "image_path": "e0c709bf75816ce94a630106acba9aee7a84318a69e02f66e3608f0f99d9a350.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 197, + 424, + 214, + 430 + ], + "lines": [ + { + "bbox": [ + 197, + 424, + 214, + 430 + ], + "spans": [ + { + "bbox": [ + 197, + 424, + 214, + 430 + ], + "type": "text", + "content": "Initial" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 236, + 211, + 294, + 255 + ], + "blocks": [ + { + "bbox": [ + 236, + 211, + 294, + 255 + ], + "lines": [ + { + "bbox": [ + 236, + 211, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 236, + 211, + 294, + 255 + ], + "type": "image", + "image_path": "4ae79fa60091721b506a01f02a062cc8cb0656809774d2fade069cc3461b4f91.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 236, + 257, + 294, + 300 + ], + "blocks": [ + { + "bbox": [ + 236, + 257, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 236, + 257, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 236, + 257, + 294, + 300 + ], + "type": "image", + "image_path": "28e7c04f602f25e9c4eb1a6182878d8d43b46f8f11338d0e65524bce331a003b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 236, + 303, + 294, + 338 + ], + "blocks": [ + { + "bbox": [ + 236, + 303, + 294, + 338 + ], + "lines": [ + { + "bbox": [ + 236, + 303, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 236, + 303, + 294, + 338 + ], + "type": "image", + "image_path": "34b5beeb6abfeb220152cb29ecb4ed28d614236100f638debc200dda9c4904ee.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 236, + 341, + 294, + 376 + ], + "blocks": [ + { + "bbox": [ + 236, + 341, + 294, + 376 + ], + "lines": [ + { + "bbox": [ + 236, + 341, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 236, + 341, + 294, + 376 + ], + "type": "image", + "image_path": "4999b7bbda47d2425a4db34b8e11a212f829ce1e7fed7b0f4f1e5121f0f077bd.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 236, + 380, + 294, + 420 + ], + "blocks": [ + { + "bbox": [ + 236, + 380, + 294, + 420 + ], + "lines": [ + { + "bbox": [ + 236, + 380, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 236, + 380, + 294, + 420 + ], + "type": "image", + "image_path": "7a57f11a0171448c428431c7c785cd336d6e22be18aaa1b79449f6dedec3f365.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 241, + 424, + 288, + 431 + ], + "lines": [ + { + "bbox": [ + 241, + 424, + 288, + 431 + ], + "spans": [ + { + "bbox": [ + 241, + 424, + 288, + 431 + ], + "type": "text", + "content": "Refined by ARM" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 295, + 464 + ], + "type": "text", + "content": "Figure 6. Comparison between the initially generated prompt and the prompt refined using ARM." + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "text", + "content": "evaluate its effectiveness. Comparing the " + }, + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "text", + "content": " row with the " + }, + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "inline_equation", + "content": "1^{st}" + }, + { + "bbox": [ + 55, + 488, + 296, + 559 + ], + "type": "text", + "content": " row in Tab. 2, adding the auxiliary embedding leads to performance improvements across nearly all metrics on the three COD datasets. This indicates that the auxiliary embedding effectively enhances SAM by providing additional multimodal features." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "Effectiveness of Multimodal Prompt. Our primary contribution is utilizing BLIP to generate multimodal prompts, providing SAM with more informative cues to handle the COD task. Experimental results (" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " row of Tab. 2) show that prompts generated from full description already outperform SAM with auxiliary embedding. Furthermore, SAM's performance further improved when extracting the noun about the object category from the long description as textual information to generate prompts (" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "4^{th}" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " row of Tab. 2). This result not only proves the effectiveness of our generated prompts but also indicates that, although BLIP can produce descriptions closely related to the image, redun-" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 212, + 554, + 261 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 554, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 554, + 261 + ], + "type": "text", + "content": "dant information introduces noise, preventing the generated prompt from accurately localizing the targets. In contrast, providing only category information helps generate more precise mask prompts." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 277, + 555, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 555, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 555, + 479 + ], + "type": "text", + "content": "Effectiveness of Refinement. As shown in the " + }, + { + "bbox": [ + 313, + 277, + 555, + 479 + ], + "type": "inline_equation", + "content": "5^{th}" + }, + { + "bbox": [ + 313, + 277, + 555, + 479 + ], + "type": "text", + "content": " row results in Tab. 2, our proposed ARM significantly enhances SAM's performance on the COD task, achieving substantial improvements across all evaluation metrics on three datasets. This performance boost can be attributed to two key aspects: First, ARM refines the multimodal prompts initially generated by BLIP, as illustrated in Fig. 6. Due to domain shifts, these initial prompts often localize only partial regions of the target. However, after refining with ARM, the refined prompts achieve more complete target coverage. Consequently, the mask prompts generated from these refined prompts provide SAM with more precise cues. Second, ARM effectively integrates the image features extracted by the adapter into SAM, enabling fine-tuning of the image encoder. This process further enhances SAM's representation capability, leading to improved performance on the COD task." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "spans": [ + { + "bbox": [ + 314, + 508, + 388, + 521 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel and cost-effective solution for the COD task. The main idea is to incorporate multimodal information to construct high-quality mask prompts, enhancing SAM's segmentation capability. Specifically, we use VLM to automatically generate multimodal data and prompts, laying a solid foundation for constructing the final high-quality mask prompts. Furthermore, we propose the Adaptive Refinement Module (ARM) to refine the initial multimodal prompts enabling more complete target representations and fine-tune SAM's image encoder, providing it with richer information. Additionally, experimental results demonstrate that our proposed architecture achieves outstanding performance across multiple COD datasets, particularly excelling in capturing the structural details of objects, thereby further validating our work's effectiveness." + } + ] + } + ], + "index": 32 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20679" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 157, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "spans": [ + { + "bbox": [ + 56, + 91, + 294, + 114 + ], + "type": "text", + "content": "The work is supported by the XJTLU RDF project: RDF-22-02-118 & RDF-24-02-041" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 115, + 132 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 139, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 139, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 139, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 61, + 139, + 295, + 183 + ], + "type": "text", + "content": "[1] Radhakrishna Achanta, Sheila Hemami, Francisco Estrada, and Sabine Susstrunk. Frequency-tuned salient region detection. In 2009 IEEE conference on computer vision and pattern recognition, pages 1597-1604. IEEE, 2009. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 184, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 184, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 184, + 296, + 228 + ], + "type": "text", + "content": "[2] Huafeng Chen, Dian Shao, Guangqian Guo, and Shan Gao. Just a hint: Point-supervised camouflaged object detection. In European Conference on Computer Vision, pages 332-348. Springer, 2024. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 229, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 229, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 229, + 295, + 274 + ], + "type": "text", + "content": "[3] Huafeng Chen, Pengxu Wei, Guangqian Guo, and Shan Gao. Sam-cod+: Sam-guided unified framework for weakly-supervised camouflaged object detection. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 274, + 295, + 317 + ], + "type": "text", + "content": "[4] Tianle Chen, Zheda Mai, Ruiwen Li, and Wei-lun Chao. Segment anything model (sam) enhanced pseudo labels for weakly supervised semantic segmentation. arXiv preprint arXiv:2305.05803, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 295, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 385 + ], + "type": "text", + "content": "[5] Tianrun Chen, Lanyun Zhu, Chaotao Ding, Runlong Cao, Yan Wang, Zejian Li, Lingyun Sun, Papa Mao, and Ying Zang. Sam fails to segment anything?-sam-adapter: Adapting sam in underperformed scenes: Camouflage, shadow, medical image segmentation, and more. arXiv preprint arXiv:2304.09148, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 295, + 419 + ], + "type": "text", + "content": "[6] Zhennan Chen, Rongrong Gao, Tian-Zhu Xiang, and Fan Lin. Diffusion model for camouflaged object detection. In ECAI 2023, pages 445-452. IOS Press, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 421, + 295, + 464 + ], + "type": "text", + "content": "[7] Zhongxi Chen, Ke Sun, and Xianming Lin. Camodiffusion: Camouflaged object detection via conditional diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1272-1280, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 295, + 509 + ], + "type": "text", + "content": "[8] Deng-Ping Fan, Ming-Ming Cheng, Yun Liu, Tao Li, and Ali Borji. Structure-measure: A new way to evaluate foreground maps. In Proceedings of the IEEE international conference on computer vision, pages 4548-4557, 2017. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 511, + 295, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 511, + 295, + 554 + ], + "spans": [ + { + "bbox": [ + 62, + 511, + 295, + 554 + ], + "type": "text", + "content": "[9] Deng-Ping Fan, Cheng Gong, Yang Cao, Bo Ren, MingMing Cheng, and Ali Borji. Enhanced-alignment measure for binary foreground map evaluation. arXiv preprint arXiv:1805.10421, 2018. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 556, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 295, + 609 + ], + "type": "text", + "content": "[10] Deng-Ping Fan, Ge-Peng Ji, Guolei Sun, Ming-Ming Cheng, Jianbing Shen, and Ling Shao. Camouflaged object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2777-2787, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 295, + 644 + ], + "type": "text", + "content": "[11] Albert Gu and Tri Dao. Mamba: Linear-time sequence modeling with selective state spaces. arXiv preprint arXiv:2312.00752, 2023. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 689 + ], + "type": "text", + "content": "[12] Juwei Guan, Xiaolin Fang, Tongxin Zhu, Zhipeng Cai, Zhen Ling, Ming Yang, and Junzhou Luo. Idenet: Making neural network identify camouflaged objects like creatures. IEEE Transactions on Image Processing, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 295, + 713 + ], + "type": "text", + "content": "[13] Juwei Guan, Xiaolin Fang, Tongxin Zhu, and Weiqi Qian. Sdrnet: Camouflaged object detection with independent re" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "construction of structure and detail. Knowledge-Based Systems, 299:112051, 2024. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[14] Juwei Guan, Weiqi Qian, Tongxin Zhu, and Xiaolin Fang. Promoting camouflaged object detection through novel edge-target interaction and frequency-spatial fusion. Neurocomputing, 617:129064, 2025. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 173 + ], + "type": "text", + "content": "[15] Cunhan Guo and Heyan Huang. Cofinet: Unveiling camouflaged objects with multi-scale finesse. Neurocomputing, 614:128763, 2025. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 175, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 240 + ], + "type": "text", + "content": "[16] Chunming He, Kai Li, Yachao Zhang, Longxiang Tang, Yu-lun Zhang, Zhenhua Guo, and Xiu Li. Camouflaged object detection with feature decomposition and edge reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22046-22055, 2023. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 307 + ], + "type": "text", + "content": "[17] Chunming He, Kai Li, Yachao Zhang, Guoxia Xu, Longxiang Tang, Yulun Zhang, Zhenhua Guo, and Xiu Li. Weakly-supervised concealed object segmentation with sam-based pseudo labeling and multi-scale feature grouping. Advances in Neural Information Processing Systems, 36:30726-30737, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 374 + ], + "type": "text", + "content": "[18] Jiangping He, Yangjie Zhang, Maggie Chung, Michael Wang, Kun Wang, Yan Ma, Xiaoyang Ding, Qiang Li, and Yonglin Pu. Whole-body tumor segmentation from pet/ct images using a two-stage cascaded neural network with camouflaged object detection mechanisms. Medical Physics, 50 (10):6151-6162, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 376, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 376, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 317, + 376, + 553, + 421 + ], + "type": "text", + "content": "[19] Ruozhen He, Qihua Dong, Jiaying Lin, and Rynson WH Lau. Weakly-supervised camouflaged object detection with scribble annotations. In Proceedings of the AAAI conference on artificial intelligence, pages 781-789, 2023. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 444 + ], + "type": "text", + "content": "[20] Dan Hendrycks and Kevin Gimpel. Gaussian error linear units (gelus). arXiv preprint arXiv:1606.08415, 2016. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 445, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 478 + ], + "type": "text", + "content": "[21] Matthew Honnibal, Ines Montani, Sofie Van Landeghem, and Adriane Boyd. spaCy: Industrial-strength Natural Language Processing in Python. 2020. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[22] Bintao Hu, Hengyan Liu, Jianbo Du, Miguel López-Benitez, Celimuge Wu, Xiaoli Chu, and Dusit Niyato. Madqnenhanced computation offloading and resource allocation for 6g low-altitude economy vehicular networks. IEEE Transactions on Cognitive Communications and Networking, 2025. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 546, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 601 + ], + "type": "text", + "content": "[23] Xihang Hu, Xiaoli Zhang, Fasheng Wang, Jing Sun, and Fuming Sun. Efficient camouflaged object detection network based on global localization perception and local guidance refinement. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):5452-5465, 2024. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 602, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 602, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 602, + 553, + 657 + ], + "type": "text", + "content": "[24] Zhou Huang, Hang Dai, Tian-Zhu Xiang, Shuo Wang, Huai-Xin Chen, Jie Qin, and Huan Xiong. Feature shrinkage pyramid for camouflaged object detection with transformers. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5557-5566, 2023. 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 555, + 713 + ], + "type": "text", + "content": "[25] Qi Jia, Shuilian Yao, Yu Liu, Xin Fan, Risheng Liu, and Zhongxuan Luo. Segment, magnify and reiterate: Detecting camouflaged objects the hard way. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4713-4722, 2022. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20680" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[26] Xinhao Jiang, Wei Cai, Yao Ding, Xin Wang, Danfeng Hong, Xingyu Di, and Weijie Gao. Bidicos: Camouflaged object segmentation via bilateral diffusion model. Expert Systems with Applications, 255:124747, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 172 + ], + "type": "text", + "content": "[27] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[28] Trung-Nghia Le, Tam V Nguyen, Zhongliang Nie, Minh-Triet Tran, and Akihiro Sugimoto. Anabranch network for camouflaged object segmentation. Computer vision and image understanding, 184:45-56, 2019. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 271 + ], + "type": "text", + "content": "[29] Junnan Li, Ramprasaath Selvaraju, Akhilesh Gotmare, Shafiq Joty, Caiming Xiong, and Steven Chu Hong Hoi. Align before fuse: Vision and language representation learning with momentum distillation. Advances in neural information processing systems, 34:9694-9705, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 326 + ], + "type": "text", + "content": "[30] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International conference on machine learning, pages 12888-12900. PMLR, 2022. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[31] Lin Li, Bo Dong, Eric Rigall, Tao Zhou, Junyu Dong, and Geng Chen. Marine animal segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(4): 2303-2314, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 437 + ], + "type": "text", + "content": "[32] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 503 + ], + "type": "text", + "content": "[33] Hengyan Liu, Guangyu Ren, Tianhong Dai, Di Zhang, Pengjing Xu, Wenzhang Zhang, and Bintao Hu. Diversity-oriented contrastive learning for rgb-t scene parsing. In 2023 2nd International Conference on Sensing, Measurement, Communication and Internet of Things Technologies (SMC-IoT), pages 155-160. IEEE, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 558 + ], + "type": "text", + "content": "[34] Hengyan Liu, Wenzhang Zhang, Tianhong Dai, Longfei Yin, and Guangyu Ren. Multimodal frequency spectrum fusion schema for rgb-t image semantic segmentation. In 2024 33rd International Conference on Computer Communications and Networks (ICCNN), pages 1-6. IEEE, 2024. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[35] Limei Liu, Meng Liu, Kexin Meng, Linqiang Yang, Min Zhao, and Shuli Mei. Camouflaged locust segmentation based on pranet. Computers and Electronics in Agriculture, 198:107061, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 294, + 657 + ], + "type": "text", + "content": "[36] Yu Liu, Haihang Li, Juan Cheng, and Xun Chen. Mscaf-net: A general framework for camouflaged object detection via learning multi-scale context-aware features. IEEE Transactions on Circuits and Systems for Video Technology, 33(9): 4934-4947, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 713 + ], + "type": "text", + "content": "[37] Yantong Liu, Sai Che, Liwei Ai, Chuanxiang Song, Zheyu Zhang, Yongkang Zhou, Xiao Yang, and Chen Xian. Camouflage detection: Optimization-based computer vision for alligator sinensis with low detectability in complex wild environments. Ecological Informatics, 83:102802, 2024. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[38] Yunqiu Lv, Jing Zhang, Yuchao Dai, Aixuan Li, Bowen Liu, Nick Barnes, and Deng-Ping Fan. Simultaneously localize, segment and rank the camouflaged objects. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11591-11601, 2021. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[39] Yixuan Lyu, Hong Zhang, Yan Li, Hanyang Liu, Yifan Yang, and Ding Yuan. Uedg: Uncertainty-edge dual guided camouflage object detection. IEEE Transactions on Multimedia, 26:4050-4060, 2023. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 218 + ], + "type": "text", + "content": "[40] Tsui Qin Mok, Shuyong Gao, Haozhe Xing, Miaoyang He, Yan Wang, and Wenqiang Zhang. A holistically point-guided text framework for weakly-supervised camouflaged object detection. arXiv preprint arXiv:2501.06038, 2025. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 274 + ], + "type": "text", + "content": "[41] Yuzhen Niu, Lifen Yang, Rui Xu, Yuezhou Li, and Yuzhong Chen. Minet: Weakly-supervised camouflaged object detection through mutual interaction between region and edge cues. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 6316-6325, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 330 + ], + "type": "text", + "content": "[42] Youwei Pang, Xiaqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoom in and out: A mixed-scale triplet network for camouflaged object detection. In Proceedings of the IEEE/CVF Conference on computer vision and pattern recognition, pages 2160-2170, 2022. 2, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 385 + ], + "type": "text", + "content": "[43] Youwei Pang, Xiaoqi Zhao, Tian-Zhu Xiang, Lihe Zhang, and Huchuan Lu. Zoomnext: A unified collaborative pyramid network for camouflaged object detection. IEEE transactions on pattern analysis and machine intelligence, 2024. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 387, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 421 + ], + "type": "text", + "content": "[44] Quan Quan, Fenghe Tang, Zikang Xu, Heqin Zhu, and S Kevin Zhou. Slide-sam: Medical sam meets sliding window. arXiv preprint arXiv:2311.10121, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 455 + ], + "type": "text", + "content": "[45] Guangyu Ren, Yinxiao Yu, Hengyan Liu, and Tania Stathaki. Dynamic knowledge distillation with noise elimination for rgb-d salient object detection. Sensors, 22(16):6188, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 456, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 520 + ], + "type": "text", + "content": "[46] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE international conference on computer vision, pages 618-626, 2017. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 553, + 578 + ], + "type": "text", + "content": "[47] Yiqing Shen, David Dreizin, Blanca Inigo, and Mathias Unberath. Protosam-3d: Interactive semantic segmentation in volumetric medical imaging via a segment anything model and mask-level prototypes. Computerized Medical Imaging and Graphics, page 102501, 2025. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 623 + ], + "type": "text", + "content": "[48] Przemysław Skurowski, Hassan Abdulameer, Jakub Blaszczyk, Tomasz Depta, Adam Kornacki, and Przemysław Koziel. Animal camouflage analysis: Chameleon database. Unpublished manuscript, 2(6):7, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 668 + ], + "type": "text", + "content": "[49] Ze Song, Xudong Kang, Xiaohui Wei, Haibo Liu, Renwei Dian, and Shutao Li. Fsnet: Focus scanning network for camouflaged object detection. IEEE Transactions on Image Processing, 32:2267-2278, 2023. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[50] Yanguang Sun, Chunyan Xu, Jian Yang, Hanyu Xuan, and Lei Luo. Frequency-spatial entanglement learning for camouflaged object detection. In European Conference on Computer Vision, pages 343-360. Springer, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20681" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[51] Lv Tang, Peng-Tao Jiang, Zhi-Hao Shen, Hao Zhang, Jin-Wei Chen, and Bo Li. Chain of visual perception: Harnessing multimodal large language models for zero-shot camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8805–8814, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 205 + ], + "type": "text", + "content": "[52] Huafeng Wang, Tianyu Hu, Yanan Zhang, Haodu Zhang, Yong Qi, Longzhen Wang, Jianhua Ma, and Minghua Du. Unveiling camouflaged and partially occluded colorectal polyps: Introducing cpsnet for accurate colon polyp segmentation. Computers in Biology and Medicine, 171:108186, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 207, + 294, + 262 + ], + "type": "text", + "content": "[53] Liqiong Wang, Jinyu Yang, Yanfu Zhang, Fangyi Wang, and Feng Zheng. Depth-aware concealed crop detection in dense agricultural scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17201-17211, 2024. 1, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 308 + ], + "type": "text", + "content": "[54] Qingwei Wang, Jinyu Yang, Xiaosheng Yu, Fangyi Wang, Peng Chen, and Feng Zheng. Depth-aided camouflaged object detection. In Proceedings of the 31st ACM International Conference on Multimedia, pages 3297–3306, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 363 + ], + "type": "text", + "content": "[55] Tingran Wang, Zaiyang Yu, Jianwei Fang, Jinlong Xie, Feng Yang, Huang Zhang, Liping Zhang, Minghua Du, Lusi Li, and Xin Ning. Multidimensional fusion of frequency and spatial domain information for enhanced camouflaged object detection. Information Fusion, 117:102871, 2025. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 420 + ], + "type": "text", + "content": "[56] Zongwei Wu, Danda Pani Paudel, Deng-Ping Fan, Jingjing Wang, Shuo Wang, Cédric Demonceaux, Radu Timofte, and Luc Van Gool. Source-free depth for object pop-out. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1032-1042, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[57] Mochu Xiang, Jing Zhang, Yunqiu Lv, Aixuan Li, Yiran Zhong, and Yuchao Dai. Exploring depth contribution for camouflaged object detection. arXiv preprint arXiv:2106.13217, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 521 + ], + "type": "text", + "content": "[58] Haozhe Xing, Shuyong Gao, Hao Tang, Tsui Qin Mok, Yanlan Kang, and Wenqiang Zhang. Tinycod: Tiny and effective model for camouflaged object detection. In ICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2023. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 577 + ], + "type": "text", + "content": "[59] Bowen Yin, Xuying Zhang, Deng-Ping Fan, Shaohui Jiao, Ming-Ming Cheng, Luc Van Gool, and Qibin Hou. Camoformer: Masked separable attention for camouflaged object detection. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 294, + 633 + ], + "type": "text", + "content": "[60] Zhenni Yu, Xiaogin Zhang, Li Zhao, Yi Bin, and Guobao Xiao. Exploring deeper! segment anything model with depth perception for camouflaged object detection. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 4322-4330, 2024. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 294, + 689 + ], + "type": "text", + "content": "[61] Hong Zhang, Yixuan Lyu, Qian Yu, Hanyang Liu, Huimin Ma, Ding Yuan, and Yifan Yang. Unlocking attributes' contribution to successful camouflage: A combined textual and visual analysis strategy. In European Conference on Computer Vision, pages 315-331. Springer, 2024. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[62] Jie Zhang, Xubing Yang, Rui Jiang, Wei Shao, and Li Zhang. Rsam-seg: A sam-based approach with prior knowledge in" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 240 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tegration for remote sensing image semantic segmentation. arXiv preprint arXiv:2402.19004, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 149 + ], + "type": "text", + "content": "[63] Shizhou Zhang, Dexuan Kong, Yinghui Xing, Yue Lu, Lingyan Ran, Guoqiang Liang, Hexu Wang, and Yanning Zhang. Frequency-guided spatial adaptation for camouflaged object detection. IEEE Transactions on Multimedia, 27:72-83, 2025. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 205 + ], + "type": "text", + "content": "[64] Xiaoqin Zhang, Zhenni Yu, Li Zhao, Deng-Ping Fan, and Guobao Xiao. Comprompter: reconceptualized segment anything model with multiprompt network for camouflaged object detection. Science China Information Sciences, 68(1): 112104, 2025. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 240 + ], + "type": "text", + "content": "[65] Ke Zhou, Zhongwei Qiu, and Dongmei Fu. Multi-scale contrastive adaptor learning for segmenting anything in under-performed scenes. Neurocomputing, 606:128395, 2024. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20682" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_content_list.json b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e510b8969338cda9fc5b9f013c6d48bb17d2e591 --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_content_list.json @@ -0,0 +1,1573 @@ +[ + { + "type": "text", + "text": "Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment", + "text_level": 1, + "bbox": [ + 133, + 128, + 864, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ying Ba $^{1,2,3}$ , Tianyu Zhang $^{4}$ , Yalong Bai $^{4}$ , Wenyi Mo $^{1,2,3}$ , Tao Liang $^{4}$ , Bing Su $^{1,2,3\\dagger}$ , Ji-Rong Wen $^{1,2,3}$ $^{1}$ Gaoling School of Artificial Intelligence Renmin University of China Beijing, China \n $^{2}$ Beijing Key Laboratory of Research on Large Models and Intelligent Governance \n $^{3}$ Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE $^{4}$ iN2X", + "bbox": [ + 101, + 208, + 941, + 297 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1,2,3}$ yingba88@outlook.com, {mowenyi00, subingats}@gmail.com, jrwen@ruc.edu.cn \n $^{4}$ {tianyu1949, liang0305tao}@gmail.com, ylbai@outlook.com", + "bbox": [ + 228, + 300, + 815, + 334 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 243, + 369, + 330, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Contemporary image generation systems have achieved high fidelity and superior aesthetic quality beyond basic text-image alignment. However, existing evaluation frameworks have failed to evolve in parallel. This study reveals that human preference reward models fine-tuned based on CLIP and BLIP architectures have inherent flaws: they inappropriately assign low scores to images with rich details and high aesthetic value, creating a significant discrepancy with actual human aesthetic preferences. To address this issue, we design a novel evaluation score, ICT (Image-Contained-Text) score, that achieves and surpasses the objectives of text-image alignment by assessing the degree to which images represent textual content. Building upon this foundation, we further train an HP (High-Preference) score model using solely the image modality to enhance image aesthetics and detail quality while maintaining text-image alignment. Experiments demonstrate that the proposed evaluation model improves scoring accuracy by over $10\\%$ compared to existing methods, and achieves significant results in optimizing state-of-the-art text-to-image models. This research provides theoretical and empirical support for evolving image generation technology toward higher-order human aesthetic preferences. Code is available at https://github.com/BarretBa/ICTHP.", + "bbox": [ + 89, + 400, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 796, + 241, + 811 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, generation models have witnessed rapid development and found widespread applications across diverse domains [11, 17, 23, 29, 32, 42, 44, 47]. Modern", + "bbox": [ + 89, + 820, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a219871a8298e64b70e94219a657bcf91893f5662a7ca8cfbc77799f7facf632.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 368, + 705, + 468 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f1812da3ac72273ccac1fb5cfbeed2215436aec9baf9176f4dcb63209992b88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 369, + 903, + 469 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/3decd7529e2444420bc6235cf1c582a1d9652f72c3ee1b9ef8f9086daddf28d0.jpg", + "image_caption": [ + "Figure 1. Reward Model Scoring Paradox Across Multiple Models Based on CLIP [26] (CLIP Score, PickScore) and BLIP [15] (ITM Score, ImageReward). Refine images are generated with refined prompts by LLMs." + ], + "image_footnote": [], + "bbox": [ + 516, + 469, + 705, + 566 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f8b7a98ace20886f9dc4a0cb7e86f5c9ddd00f151b9f42586f9226e98502e6c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 705, + 469, + 900, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "high-performance diffusion models can not only accurately align texts and images but also generate visually impressive results. As the technology matures, the optimization focus of diffusion models has shifted from text-image alignment and prompt fidelity to generating high-quality images that better meet human aesthetic expectations [9, 19, 20, 37, 39]. In terms of optimization strategies, traditional data-driven training methods [3, 25, 30] are gradually being supplemented and expanded by reward model-based gradient backpropagation and reinforcement learning algorithms [34, 46]. Therefore, constructing a reward function that can precisely evaluate image quality is crucial for further improving the performance of diffusion models.", + "bbox": [ + 509, + 642, + 906, + 853 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "For text-to-image evaluation, text-image alignment (i.e., instruction compliance) serves as the fundamental requirement, while human preference determines the", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*This work was done at iN2X.", + "bbox": [ + 107, + 875, + 297, + 886 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 256, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19022", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "upper limit of image quality. Current reward models exhibit a critical flaw in measuring text-image alignment: they presume potential equivalence between the information content of images and text, whereas in reality, textual descriptions can hardly comprehensively capture all visual information in images. A direct manifestation is that for the same prompt, a basic image merely covering textual descriptions typically receives higher scores than a higher-quality image that contains richer details and visual elements extending beyond the literal description of the prompt. This phenomenon persists not only in foundational models [15, 26] but also affects fine-tuned variants with human preferences like ImageReward [40] and PickScore [13] (as shown in Fig. 1). Thus, previous reward models may effectively enhance Stable Diffusion-1.5/2.0 [29] (where text-image alignment remains imperfect) but degrade state-of-the-art large models like Stable Diffusion-3.5 [1], steering them to generate visually sparse and aesthetically deficient images.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the above issue and continue leveraging human preference data to train superior reward models, we propose a novel optimization objective that transcends text-image alignment, i.e., the image-contained-text score (ICT score). ICT score quantifies the extent to which an image contains textual information to evaluate the quality of generated images. By focusing on how well the image encapsulates textual semantics rather than enforcing bidirectional equivalence, ICT score prevents penalizing images that naturally contain richer contextual details beyond literal prompt descriptions, which aligns with human perception of high-quality outputs. The ICT score removes the ceiling constraints of human preference models, enabling us to incorporate ultra-high-quality preference data to train an advanced human preference model, achieving a more comprehensive evaluation system.", + "bbox": [ + 89, + 396, + 482, + 652 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this research can be summarized in the following three points:", + "bbox": [ + 89, + 655, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel image generation evaluation framework, including ICT score (quantifying the extent to which an image contains textual information) and HP score (evaluating how well an image meets human aesthetic preferences), fundamentally solving the problem of existing reward models underscoring high-quality images.", + "- A large-scale high-quality dataset (Pick-High) is constructed using chain-of-thought reasoning from large language models, providing a comprehensive benchmark for image generation quality assessment.", + "- We successfully apply the proposed reward model to optimize a state-of-the-art diffusion model (Stable Diffusion-3.5-turbo), demonstrating the effic" + ], + "bbox": [ + 93, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tiveness and superiority of this method in practical generation tasks.", + "bbox": [ + 532, + 90, + 903, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 136, + 678, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image quality evaluation for generative models. The evaluation of generated images primarily falls into two categories: human assessment and automated metrics. While the former serves as the gold standard, its high cost and subjective biases have led to the widespread application of the latter in large-scale evaluations. Image-generation models have seen significant quality improvements, evolving from GAN-based approaches such as AttnGAN [41] and DM-GAN [48] to diffusion-based models including DALL·E 2 [27], Stable Diffusion [7, 29], Midjourney [18], and FLUX [14]. Traditional metrics like FID [10] and IS [31] focus solely on distributional statistical properties and fail to capture human subjective preferences. Addressing this limitation, researchers have developed specialized human preference reward models trained on human preference data, such as PickScore [13], HPSv2 [38], and ImageReward [40]. However, we have observed that these reward models exhibit significant biases when evaluating high-quality images generated through prompt optimization, indicating they have not yet effectively guided diffusion models in generating images that align with human preferences for detail and content richness.", + "bbox": [ + 511, + 164, + 903, + 525 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image quality evaluation datasets. Addressing the limitations of existing human preference models, high-quality evaluation datasets are essential for generative model assessment. As generative models rapidly evolved, researchers shifted toward constructing datasets based on human preferences, including the PickaPic_v2 dataset [13], ImageRewardDB [40] developed by the ImageReward team, and the HPDv2 dataset [38] employed by HPS. Despite their scale, these datasets predominantly contain lower-quality images inadequate for evaluating high-quality outputs from modern diffusion models and lack diversity across quality levels. To address these issues, we constructed a 600GB high-quality dataset called Pick-High, incorporating portions of the Pick-a-Pic dataset during training to ensure a comprehensive quality distribution.", + "bbox": [ + 511, + 527, + 903, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Limitations of Text-image Alignment", + "text_level": 1, + "bbox": [ + 511, + 782, + 898, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal models such as CLIP employ discriminative contrastive loss during pre-training to achieve text-image alignment. According to the information bottleneck theory, their encoders extract minimal image discriminative information to maximize mutual information with text. This training mechanism causes CLIP", + "bbox": [ + 511, + 809, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19023", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/55f402f6f98c871a998306ef094d3a6ffdc267bcc5776b1cc4ce4e56da716d89.jpg", + "image_caption": [ + "Figure 2. Text-Image Feature Interaction Venn Diagram: Increasing Image Information Under Fixed Text Prompts." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 480, + 188 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "to exhibit bias when evaluating the same prompt: it tends to assign higher scores to simple images that merely faithfully reflect textual content, while giving lower scores to high-quality images rich in details, contradicting human aesthetic preferences.", + "bbox": [ + 89, + 244, + 483, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the principle of information decomposition, the information content of a generated image can be divided into:", + "bbox": [ + 89, + 321, + 483, + 366 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI (v) = I (v; t) + I (v | t), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 204, + 378, + 480, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I(v; t)$ represents the mutual information between the image $v$ and text $t$ (the image-text alignment component), and $I(v|t)$ represents the additional visual information contained in the image (aesthetics, texture, atmosphere, etc.).", + "bbox": [ + 89, + 407, + 483, + 482 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The CLIP scoring mechanism, based on cosine similarity, can be expressed as:", + "bbox": [ + 89, + 483, + 483, + 513 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {C L I P} (v, t) \\approx \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot I (v)}} = \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot (I (v ; t) + I (v | t))}}.\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 523, + 483, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, as the capability of generative models improves, the information content of images increases. In the transition from image-text misalignment to alignment, $I(v; t)$ increases faster than $I(v)$ , gradually raising CLIP scores. However, when high-quality models generate images rich in visual details, although $I(v; t)$ increases, $I(v|t)$ grows even faster, causing the denominator $\\sqrt{I(v)}$ to increase more than the numerator $I(v; t)$ , resulting in an overall decrease in scores, creating a contradiction with human preferences. This motivates us to design a novel objective to train foundation models from the perspective of reward models, which will be elaborated in the next section.", + "bbox": [ + 89, + 569, + 483, + 765 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 779, + 197, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section presents our proposed methodological framework. In Section 4.1, we introduce Pick-High dataset, a large-scale high-quality dataset constructed using refined prompts, which addresses the evaluation bias in existing reward models when assessing information-rich images. In Section 4.2, to resolve the inconsistency between the image-text alignment", + "bbox": [ + 89, + 795, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/53796012433464b4102cd66c8d504ce0bd93fa49da5780e7393321998fc2b7ac.jpg", + "image_caption": [ + "Figure 3. The Image-Contained-Text (ICT) Scoring Framework." + ], + "image_footnote": [], + "bbox": [ + 521, + 90, + 901, + 237 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "objective of foundation models and evaluation goals of reward models, we propose the Image-Contained-Text (ICT) optimization objective. This enhancement adapts the CLIP model for better evaluation of generative models. We further develop an innovative HP model that relies solely on image modality for human preference assessment. Finally, in Section 4.3, we detail how we utilize the improved reward model to optimize diffusion models, thus enhancing the overall quality of generated images.", + "bbox": [ + 511, + 277, + 906, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Pick-High Dataset", + "text_level": 1, + "bbox": [ + 511, + 433, + 691, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Existing reward models (including PickScore, HPSv2, and ImgReward) inherit the limitations of foundational multimodal models, i.e., assigning unreasonably low scores when evaluating complex images paired with simple prompts, which significantly deviates from human preferences. To address this deficiency, we curated 360,000 prompts from the PickAPic_v2 dataset and leveraged large language models' chain-of-thought capabilities to meticulously design a more refined set of prompts that closely align with human preferences. We input these optimized prompts into state-of-the-art image generation models to create 360,000 high-quality images, forming the Pick-High dataset. The detailed construction processes of Pick-High dataset can be seen in Appendix A.", + "bbox": [ + 511, + 452, + 906, + 678 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "PickAPic_v2 contains pairwise preferences of images. We denote the non-preferred low-quality image as $I_{1}$ and the preferred image as $I_{2}$ . The Pick-High dataset extends pairwise preferences to triplet rankings by introducing a high-quality image $I_{3}$ generated with refined prompts. Based on these quality levels, we establish a hierarchical scoring mechanism for both basic prompts $P_{\\mathrm{easy}}$ and refined prompts $P_{\\mathrm{ref}}$ , to better align human preferences.", + "bbox": [ + 511, + 680, + 906, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2. Reward Model Optimization Objective", + "text_level": 1, + "bbox": [ + 511, + 827, + 848, + 844 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.2.1. ICT Scores Design", + "text_level": 1, + "bbox": [ + 511, + 849, + 689, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In text-to-image model evaluation, existing foundation models (e.g., CLIP, BLIP) employ instance-level", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19024", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "image-text contrastive loss, which presents a fundamental issue: they tend to penalize high-quality images containing additional details. This evaluation mechanism significantly deviates from human preferences and reward model objectives. Based on these observations, we propose the Image-Contained-Text (ICT) scoring mechanism to quantify the extent to which an image expresses textual content.", + "bbox": [ + 88, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overview of our ICT scoring mechanism is shown in Fig. 3. We assign different scores to images $(I_1, I_2, I_3)$ based on their alignment quality with prompts. For high-quality images $I_3$ , we assign perfect scores for both basic prompts and refined prompts. To calculate ICT scores of $I_2$ and $I_1$ with basic prompts, we first develop a threshold-based mechanism instead of directly adopting CLIP similarity function:", + "bbox": [ + 88, + 210, + 480, + 332 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (I, P) = \\min (\\frac {\\operatorname {C L I P} (I , P)}{\\theta}, 1), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 338, + 480, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{CLIP}(\\mathrm{I},\\mathrm{P})$ denotes the CLIP similarity function of an image $I$ and its corresponding prompt $P$ and $\\theta$ represents the alignment threshold. This approach leverages the effectiveness of the CLIP model for assessment of low- and medium-quality images while circumventing the bias in CLIP scores of high-quality images. The basic prompt ICT scores $(E_1,E_2,E_3)$ of $I_{1}$ , $I_{2}$ , $I_{3}$ containing $P_{\\mathrm{easy}}$ are defined as follows:", + "bbox": [ + 89, + 372, + 482, + 494 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {3} = 1, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 502, + 480, + 516 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {2} = \\mathcal {C} \\left(I _ {2}, P _ {\\text {e a s y}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 520, + 480, + 536 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {1} = \\min \\left(\\mathcal {C} \\left(I _ {1}, P _ {\\text {e a s y}}\\right), E _ {2}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 540, + 480, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As for refined prompt ICT scores, based on basic prompt ICT scores, we further consider the CLIP similarities of $P_{\\mathrm{easy}}$ and $P_{\\mathrm{ref}}$ , defining the scores of $I_2$ and $I_1$ containing $P_{\\mathrm{ref}}$ as the product of $E_2$ , $E_1$ and the textual similarity, as shown below:", + "bbox": [ + 89, + 561, + 482, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {3} = 1, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 659, + 480, + 674 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {2} = E _ {2} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 678, + 480, + 694 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nR _ {1} = E _ {1} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 696, + 480, + 713 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Unlike previous discriminative training objectives focused on achieving binary yes/no text-image alignment, our proposed ICT scores categorize the alignment quality between images and texts into distinct levels. This novel objective not only encourages models to achieve image-text alignment but also avoids penalizing images containing additional details beyond textual descriptions. It represents a superior training objective that goes beyond conventional text-image alignment metrics. In the following sections, we will elaborate on implementing these ICT scores as innovative labeling mechanisms for training multimodal models.", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.2. ICT Model", + "text_level": 1, + "bbox": [ + 513, + 90, + 637, + 104 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We finetune the CLIP model on Pick-High and Pickapic_v2 datasets to obtain our ICT model. Each training sample consists of an image triplet with corresponding ICT labels $(E_1, E_2, E_3)$ for the basic prompt and $(R_1, R_2, R_3)$ for the refined prompt.", + "bbox": [ + 511, + 109, + 903, + 185 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model computes ICT scores between each image and both prompts: $y_{i,e} = s(P_{\\mathrm{easy}},I_i)$ and $y_{i,r} = s(P_{\\mathrm{ref}},I_i)$ , where $i \\in \\{1,2,3\\}$ , and the scoring function $s$ is defined as $s(P,I) = \\mathrm{Enc}_{\\mathrm{xt}}(P) \\cdot \\mathrm{Enc}_{\\mathrm{img}}(I) / \\tau$ , where $\\tau$ is the temperature parameter. We optimize the model by minimizing the mean squared error between predicted scores and ICT scores:", + "bbox": [ + 511, + 185, + 905, + 291 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {I C T}} = \\sum_ {i = 1} ^ {3} (E _ {i} - y _ {i, e}) ^ {2} + \\sum_ {i = 1} ^ {3} (R _ {i} - y _ {i, r}) ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 565, + 314, + 903, + 354 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To enhance the discriminative capability of the model, we introduce an in-batch negative sample learning strategy. For any combination of images and unpaired prompts within a batch, we set their ICT labels to 0 and compute the negative sample loss $\\mathcal{L}_{\\mathrm{neg}} = \\sum_{j\\neq i}[(y_{j,e}^i)^2 +(y_{j,r}^i)^2 ]$ where $y_{j,e}^{i}$ and $y_{j,r}^{i}$ represent the ICT scores between the image of the i-th sample and the basic/Refined prompts of the j-th sample, respectively. Considering the potential existence of false negative samples within batches, we adopt a smooth weighting strategy based on the sigmoid function for negative sample mining:", + "bbox": [ + 511, + 363, + 906, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw (y) = \\frac {1}{1 + e ^ {\\alpha (| y | - \\beta)}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 556, + 903, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ is the smoothing coefficient controlling the steepness of the weight curve, $\\beta$ is the threshold parameter, and $y$ is the predicted score of the sample. This function assigns lower weights to potential false negative samples with high predicted scores, thereby mitigating their negative impact on model training. The final negative sample loss function is:", + "bbox": [ + 511, + 589, + 905, + 695 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {n e g}} = \\sum_ {j \\neq i} \\left(w \\left(y _ {j, e} ^ {i}\\right) \\cdot y _ {j, e} ^ {i} ^ {2} + w \\left(y _ {j, r} ^ {i}\\right) \\cdot y _ {j, r} ^ {i} ^ {2}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 545, + 718, + 903, + 752 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The complete loss function is formulated as:", + "bbox": [ + 532, + 761, + 852, + 776 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {I C T}} + \\lambda \\mathcal {L} _ {\\text {n e g}}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 627, + 787, + 903, + 804 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ is a balancing factor that controls the contribution of negative sample learning.", + "bbox": [ + 511, + 814, + 903, + 844 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2.3. HP Model", + "text_level": 1, + "bbox": [ + 511, + 851, + 630, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "When the image fully expresses the semantic content of the prompt, the ICT score reaches its upper limit. To", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19025", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/769d00a99dfb7dae167baef5d09db605791fa3caab2293e8c0641f493a42e408.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAverage↑Pairwise Comparison↑
I2>I1I3>I2I3>I1
Random50.0050.0050.0050.00
ICT-groundtruth87.5562.8099.8799.98
CLIP [26]60.3064.2952.8063.79
ITC [15]57.3663.3348.4260.31
ITM [15]63.3963.8858.2867.99
Aesthetic Score168.1950.9077.8175.83
ImageReward [40]63.8164.5858.0268.84
HPS_v2 [38]72.8866.9871.6380.04
PickScore [13]79.0474.8075.3786.94
ICT87.5864.65100100
HP88.4764.97100100
ICT-HP88.8466.42100100
", + "bbox": [ + 91, + 88, + 483, + 276 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "further evaluate the quality of images, we focus solely on the image modality. For each preference triplet, we employ margin ranking loss to fine-tune both the CLIP model and its subsequent MLP layers. Specifically, given a triplet $\\{I_1,I_2,I_3\\}$ , where $I_{3}$ is preferred over $I_{2}$ , and $I_{2}$ is preferred over $I_{1}$ , we enforce the preference margins as:", + "bbox": [ + 89, + 337, + 483, + 444 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m a r g i n}} = \\sum [ \\max (0, - \\Delta (I _ {2}, I _ {1}) + m) + \\\\ \\left. \\max \\left(0, - \\Delta \\left(I _ {3}, I _ {2}\\right) + m\\right) \\right], \\tag {13} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 457, + 480, + 500 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Delta(I_i, I_j)$ represents the score difference between images $I_i$ and $I_j$ , $m$ is the margin threshold.", + "bbox": [ + 89, + 510, + 482, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Diffusion Model Optimization", + "text_level": 1, + "bbox": [ + 89, + 551, + 357, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We adopt DRaFT-K [5] as our fine-tuning method to optimize Stable-Diffusion-3.5-Large-Turbo [2] by directly maximizing the differentiable reward functions.", + "bbox": [ + 89, + 574, + 482, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our proposed ICT model and HP model can each be individually utilized as reward functions. Additionally, we multiply the outputs of both models to create a more comprehensive integrated reward function, which we term the ICT-HP model.", + "bbox": [ + 89, + 619, + 483, + 695 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 710, + 241, + 727 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 736, + 295, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Two-Stage Training Pipeline. We adopt a two-stage training method for the ICT model and HP model. In the first stage, we fine-tune the CLIP model using our constructed ICT labels as training targets. In the second stage, we freeze the ICT model parameters and continue training the HP model. During the model optimization phase, we comparatively evaluate", + "bbox": [ + 89, + 758, + 482, + 864 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/de558b0f3e99443579f62c6203e901f92eb902e900349d047002c0945f5c12d8.jpg", + "table_caption": [ + "Table 1. Preference prediction accuracy on test sets of Pick-High and Pickapic_v2. Bold and underlined values represent optimal and second-best performance." + ], + "table_footnote": [], + "table_body": "
ModelJPEG Compressibility↑Aesthetic Score↑
SD3.5-Large [1]374.806.307
FLUX.1-dev [14]270.586.436
SD3.5-Turbo [2]313.106.293
+ CLIP [26](crash)279.985.175
+ ICT (Ours)321.396.285
+ ICT-HP (Ours)330.236.300
+ HP (Ours)334.866.448
", + "bbox": [ + 514, + 88, + 905, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. JPEG Compressibility and Aesthetic Scores on Parti-Prompts. Bold and underlined values represent optimal and second-best performance.", + "bbox": [ + 511, + 205, + 903, + 250 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "three different reward models for optimizing the diffusion model LoRA adapters: ICT model, HP model, and the more comprehensive ICT-HP model.", + "bbox": [ + 511, + 258, + 903, + 304 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Datasets. Our training dataset for the ICT model consists of the Pick-High dataset and a subset of Pickapic_v2, comprising a total of 360,000 image triplets. For the diffusion model optimization process, we extract 32,000 non-repetitive text prompts from the Pick-High dataset to serve as our prompt dataset.", + "bbox": [ + 511, + 305, + 903, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For detailed training hyperparameter configurations and experimental specifications, please refer to the supplementary materials.", + "bbox": [ + 511, + 397, + 903, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.2. Reward Model Performance", + "text_level": 1, + "bbox": [ + 511, + 454, + 767, + 468 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluated the prediction accuracy of human preference models on the Pick-High and Pickapic_v2 test sets. As shown in Table 1, the image-text contrastive (ITC) classification heads of foundation multimodal pre-trained models CLIP and BLIP perform only comparable to random selection. This result indicates that the instance-level image-text alignment objective adopted by foundation multimodal models is not suitable for evaluating the quality of generative models. In terms of overall average accuracy, our proposed ICT-HP model achieves nearly $10\\%$ performance improvement compared to Pickscore. This significant improvement stems from the ability of our reward model to measure the extent to which images contain text, and on this basis, continue to measure image quality, thus more accurately identifying high-quality images and assigning them correspondingly high scores.", + "bbox": [ + 511, + 477, + 906, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Qualitative Results on SD3.5-Turbo Optimization. In Figure 4, we present generation results using real user prompts. All images are generated with the same initial noise to ensure fair comparison. The experimental results show that: the HPS_v2 model only enhances image tone with minimal detail improvement; the PickScore model, while increasing details, causes distortions in style and color fidelity; the ImageReward model shows almost no substantial improvement over the original model outputs. In contrast, our proposed reward models achieve performance improvements in", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/christophschuhmann/improved-aesthetic-predictor", + "bbox": [ + 89, + 875, + 434, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19026", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fb789d500c33a463c66876b371e3c225b69d9ff18327c384cfe2b6bd517904ba.jpg", + "image_caption": [ + "Figure 4. Qualitative Comparison of Optimization Results Across Reward Models Using Real User Prompts." + ], + "image_footnote": [], + "bbox": [ + 117, + 88, + 883, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "both textural details and tonal aesthetics. Specifically, the ICT model does not penalize increases in image information content, allowing diffusion models to continue optimizing toward richer visual content after achieving text-image alignment; meanwhile, the HP model effectively utilizes pure image features to further optimize when the ICT model reaches its performance ceiling. Compared to our approach, existing reward models tend to produce negative feedback on high-quality image features, thereby limiting their ability to enhance the performance of advanced diffusion models.", + "bbox": [ + 88, + 465, + 482, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Evaluation of Visual Information Density and Aesthetic Fidelity. To quantitatively evaluate the enhancement effect of our reward model on the original model in terms of generated image information content and aesthetic quality, we employed two objective metrics, JPEG Compressibility [12] and Aesthetic Score, for assessment on the Parti-Prompts dataset [43]. As shown in Table 2, among all tested models, the HP model achieved the most significant improvements in both metrics, followed by the ICT-HP model. Notably, the SD3.5-Turbo optimized by the HP model not only demonstrated substantial improvement in aesthetic metrics, but even slightly surpassed more comprehensive baseline models such as SD3.5-Large and FLUX.1-dev.", + "bbox": [ + 89, + 638, + 482, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparison of Simple Elements Generation Results. We conducted qualitative ex", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "periments on the generation of simple elements, with results shown in Figure 5. The ICT model can stably handle simple geometric shapes and effectively penalize complex elements that do not meet simplicity requirements when prompts include descriptors like \"simple\".", + "bbox": [ + 511, + 467, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comprehensive Evaluation on GenEval Benchmark. Table 3 presents the quantitative performance of various image generation models within the multidimensional evaluation framework of GenEval [8]. Using SD3.5-Turbo as the baseline, our ICT-HP and ICT models both demonstrate significant advantages. The ICT-HP model excels in multi-object generation and color fidelity; while the ICT model achieves the highest mean score among all models and delivers optimal performance on the Counting metric, showcasing its superior capability in precise object quantity control. In contrast, the CLIP model suffers comprehensive performance degradation due to training collapse, and PickScore exhibits weaker performance in the Colors metric due to limitations in tonal representation. These comparisons validate our methods' significant advantages in terms of text-image consistency.", + "bbox": [ + 511, + 547, + 908, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative and Qualitative Performance Comparison of Multi-metric Reward Models. We conducted quantitative analyses of optimization results across various reward models, as shown in Table 4. Experiments reveal that existing reward models including HPS, PickScore, and ImageReward ex", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19027", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/eda28ee21486d63104d22e13e64b73bccddae1a98245aa1f07b29f8f418d1782.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelMean↑Single↑Two↑Counting↑Colors↑Position↑Color Attribution↑
SDXL [22]0.550.980.740.390.850.150.23
DALL-E 2 [28]0.520.940.660.490.770.100.19
DALL-E 3 [21]0.670.960.870.470.830.430.45
SD3 [7]0.680.980.840.660.740.400.43
SD3.5-Turbo [2]0.690.990.870.690.800.250.55
+ ImageReward0.700.990.870.680.800.280.59
+ HPS_v20.690.980.860.680.810.270.57
+ PickScore0.660.990.850.670.740.240.48
+ ICT-HP (Ours)0.700.990.880.680.790.280.58
+ CLIP(crash)0.130.380.040.060.260.010.03
+ ICT (Ours)0.710.980.880.700.810.310.56
", + "bbox": [ + 91, + 88, + 903, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Quantitative Evaluation of Optimization Results Across Reward Models on GenEval Benchmark. Bold and underlined values represent optimal and second-best performance.", + "bbox": [ + 89, + 294, + 903, + 324 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6e0eb177eca2e6cb4a406f468ad4f5cdf2adc35e1a02a9aab8862e8dc6a3cd32.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison of Simple Elements Generation Results." + ], + "image_footnote": [], + "bbox": [ + 102, + 345, + 470, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "hibit performance degradation across multiple metrics. In contrast, our ICT-HP and HP models achieved improvements across all metrics, demonstrating the stability advantage of our method. Although PickScore obtained the highest aesthetic score, qualitative analysis and GenEval experiments indicated significant distortion in artistic style and tone, with color-related metrics significantly lower than other models. Figure 6 compares the original image, refined image, and our model-optimized results. ICT-HP and HP models not only generate intricate backgrounds that are difficult to describe directly through language but also demonstrate a more precise and comprehensive understanding of prompts, thus producing higher-quality images.", + "bbox": [ + 88, + 671, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Transfer Validation. As illustrated in Fig-", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f3181dc1e47d569e7cdcfcd197df7d8eac66a5a6424610e137315e570ffe4fc7.jpg", + "image_caption": [ + "Figure 6. Qualitative Comparison: Original Images, Refined Images, and Generations Optimized by Our Reward Models (HP/ICT-HP)." + ], + "image_footnote": [], + "bbox": [ + 535, + 345, + 885, + 614 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ure 7, transplanting the Text-Encoder trained by our ICT model into the SD2.1-base model significantly improves image quality. This demonstrates that our encoder integrates fine-grained visual details, thus transforming textual descriptions into more precise visual guidance signals. Figure 8 presents results of style transfer with InstantStyle [35]. The CLIP image encoder is replaced by our ICT image encoder. Our fine-tuned encoder not only precisely injects reference styles but also balances text descriptions, overcoming the style-content inconsistency. These experiments validate the effectiveness and advancement of ICT score training objectives.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19028", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/78f0560f985d054c10f15320301641460b8c72714c3fe4730d7deacc11e87068.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelCLIP↑ITM↑ImgRwd↑HPS↑Pick↑Aes↑ICT↑HP↑ICT-HP↑
SD3.5-Te0.3400.8310.9550.27722.0346.5550.9100.7770.717
SD3.5-Tr0.326↓0.757↓0.914↓0.276↓22.012↓6.780↑0.923↑0.781↑0.718↑
+HPS_v20.343↑0.835↑1.065↑0.282↑22.000↓6.590↑0.929↑0.777-0.722↑
+PickScore0.332↓0.816↓1.106↑0.277-22.719↑6.924↑0.929↑0.779↑0.723↑
+ImageReward0.340-0.841↑1.056↑0.279↑22.022↓6.569↑0.865↓0.763↓0.663↓
+ICT(Ours)0.344↑0.841↑1.014↑0.277-22.117↑6.561↑0.937↑0.776↓0.728↑
+ICT-HP(Ours)0.345↑0.839↑1.011↑0.277-22.147↑6.558↑0.938↑0.778↑0.730↑
+HP(Ours)0.342↑0.836↑1.069↑0.278↑22.231↑6.668↑0.937↑0.779↑0.730↑
", + "bbox": [ + 91, + 88, + 906, + 215 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3b97bb62704a1e0aaf875b026c5007287de72600af6aded4e4ee031fa36d014a.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison of ICT Text-Encoder Direct Migration to SD2.1-Base Model." + ], + "image_footnote": [], + "bbox": [ + 96, + 279, + 480, + 547 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f3b834f185eb05f1fc7e60f3dbd696be3f4f0412657e423a0aae325d6c2a90ea.jpg", + "table_caption": [ + "Table 4. Quantitative Results. SD3.5-Te/Tr denote images generated using easy/reflect prompts on SD-3.5-Turbo. Arrows (↑, ↓, -) indicate changes vs. SD3.5-Te. Bold and underlined values represent optimal and second-best performance." + ], + "table_footnote": [], + "table_body": "
ComparisonModelsWin Rate
vs. BaseICT model0.562
HP model0.709
ICT-HP model0.673
vs. PickScoreHP model0.617
ICT-HP model0.714
", + "bbox": [ + 112, + 597, + 460, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Human evaluation on 300 randomly selected samples from Parti-Prompts [43].", + "bbox": [ + 89, + 720, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. User-Study", + "text_level": 1, + "bbox": [ + 89, + 753, + 215, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a user study with ten annotators on 300 randomly selected prompts from Parti-Prompts [43]. Image pairs (ours vs. baseline) were randomly ordered, and annotators evaluated prompt expression and visual appeal. As shown in Table 5, our methods achieve higher win rates against both the SD3.5-large-Turbo base model and the PickScore-optimized model, confirming our approach's effectiveness.", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0dbadcdcc17bba6b5739de4ef09d8d300860308a3c11b4e6c10078c5777ef3cd.jpg", + "image_caption": [ + "Figure 8. Qualitative Comparison of HP/ICT ImageEncoder Performance in Style Injection Tasks." + ], + "image_footnote": [], + "bbox": [ + 517, + 280, + 903, + 545 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 601, + 650, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reveal the reward-scoring paradox of high-quality images and show that the inherent flaw in current text-to-image generation reward models lies in classic text-image alignment paradigms. We propose a novel training objective for reward models, i.e., ICT (Image-Contained-Text) score. This objective encourages models to achieve text-image alignment and avoid penalizing high-quality images containing extra details, that better match human preferences. Based on the ICT model, we further train a preference model that relies solely on image modality for human preference assessment. Our experiments show that the proposed approach improves scoring accuracy by over $10\\%$ compared to existing methods and markedly enhances state-of-the-art models like SD3.5-Turbo. Through both theoretical foundation and empirical validation, our work advances image generation technology toward better alignment with human preferences.", + "bbox": [ + 509, + 628, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19029", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China No. 62376277, the Public Computing Cloud of Renmin University of China and the Fund for Building World-Class Universities (Disciplines) of Renmin University of China.", + "bbox": [ + 89, + 113, + 485, + 185 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 199, + 226 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.2, 5", + "[2] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 5, 7", + "[3] Yalong Bai, Mohan Zhou, and Qing Yang. Styleinject: Parameter efficient tuning of text-to-image diffusion models. ArXiv, abs/2401.13942, 2024. 1", + "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 2818-2829. IEEE, 2023. 11", + "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards, 2024. 5, 11, 13", + "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022. 11", + "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2, 7", + "[8] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. CoRR, abs/2310.11513, 2023. 6", + "[9] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1", + "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 2", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1", + "[12] ISO/IEC. Information technology - digital compression and coding of continuous-tone still images - requirements and guidelines. International Organization for Standardization, 1992. ISO/IEC 10918-1:1992. 6" + ], + "bbox": [ + 93, + 236, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. 2, 5", + "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 5", + "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 1, 2, 5", + "[16] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 11, 13", + "[17] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Biye Li, Huaian Chen, and Yi Jin. Star: Scale-wise text-conditioned autoregressive image generation. arXiv preprint arXiv:2406.10797, 2024. 1", + "[18] Midjourney. Ai-powered image generation. https://www.midjourney.com/, 2022. 2", + "[19] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, JiRong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26617-26626, 2024. 1", + "[20] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, and Ji-Rong Wen. Uniform attention maps: Boosting image fidelity in reconstruction and editing. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pages 4420-4429. IEEE, 2025. 1", + "[21] OpenAI. Dall-e 3. https://openai.com/dall-e-3, 2023.7", + "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 7", + "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, 2024. 1", + "[24] Wenwen Qiang, Jingyao Wang, Zeen Song, Jiangmeng Li, and Changwen Zheng. On the out-of-distribution generalization of self-supervised learning. arXiv preprint arXiv:2505.16675, 2025. 11", + "[25] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1", + "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on ma" + ], + "bbox": [ + 517, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19030", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "chine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5", + "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2", + "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 7", + "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2", + "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2022. 1", + "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 2", + "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1", + "[33] Zeen Song, Siyu Zhao, Xingyu Zhang, Jiangmeng Li, Changwen Zheng, and Wenwen Qiang. Learning invariant causal mechanism from vision-language models. arXiv preprint arXiv:2405.15289, 2024. 11", + "[34] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 1", + "[35] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 7", + "[36] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models, 2023. 11", + "[37] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 1", + "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 2, 5" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 1", + "[40] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 2, 5", + "[41] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1316-1324. Computer Vision Foundation / IEEE Computer Society, 2018. 2", + "[42] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 3954-3964. ACM, 2023. 1", + "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6, 8", + "[44] Guiwei Zhang, Tianyu Zhang, Mohan Zhou, Yalong Bai, and Biye Li. V2flow: Unifying visual tokenization and large language model vocabularies for autoregressive image generation. arXiv preprint arXiv:2503.07493, 2025. 1", + "[45] Yanan Zhang, Jiangmeng Li, Lixiang Liu, and Wenwen Qiang. Rethinking misalignment in vision-language model adaptation from a causal perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 11", + "[46] Ziyi Zhang, Li Shen, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Bo Du, and Dacheng Tao. Aligning few-step diffusion models with dense reward difference learning. arXiv preprint arXiv:2411.11727, 2024. 1", + "[47] Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Llada 1.5: Variance-reduced preference optimization for large language diffusion models, 2025. 1", + "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. DM-GAN: dynamic memory generative adversarial networks for text-to-image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5802-5810. Computer Vision Foundation / IEEE, 2019. 2" + ], + "bbox": [ + 516, + 90, + 903, + 892 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19031", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_model.json b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e15cc8fe2c24a5f51555bcdfd53576d98a81f0b1 --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_model.json @@ -0,0 +1,2134 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.134, + 0.13, + 0.866, + 0.177 + ], + "angle": 0, + "content": "Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.209, + 0.942, + 0.298 + ], + "angle": 0, + "content": "Ying Ba\\(^{1,2,3}\\), Tianyu Zhang\\(^{4}\\), Yalong Bai\\(^{4}\\), Wenyi Mo\\(^{1,2,3}\\), Tao Liang\\(^{4}\\), Bing Su\\(^{1,2,3\\dagger}\\), Ji-Rong Wen\\(^{1,2,3}\\) \n\\(^{1}\\)Gaoling School of Artificial Intelligence Renmin University of China Beijing, China \n\\(^{2}\\)Beijing Key Laboratory of Research on Large Models and Intelligent Governance \n\\(^{3}\\)Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE \\(^{4}\\)iN2X" + }, + { + "type": "text", + "bbox": [ + 0.229, + 0.301, + 0.816, + 0.335 + ], + "angle": 0, + "content": "\\(^{1,2,3}\\)yingba88@outlook.com, {mowenyi00, subingats}@gmail.com, jrwen@ruc.edu.cn \n\\(^{4}\\) {tianyu1949, liang0305tao}@gmail.com, ylbai@outlook.com" + }, + { + "type": "title", + "bbox": [ + 0.244, + 0.37, + 0.331, + 0.385 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.401, + 0.486, + 0.793 + ], + "angle": 0, + "content": "Contemporary image generation systems have achieved high fidelity and superior aesthetic quality beyond basic text-image alignment. However, existing evaluation frameworks have failed to evolve in parallel. This study reveals that human preference reward models fine-tuned based on CLIP and BLIP architectures have inherent flaws: they inappropriately assign low scores to images with rich details and high aesthetic value, creating a significant discrepancy with actual human aesthetic preferences. To address this issue, we design a novel evaluation score, ICT (Image-Contained-Text) score, that achieves and surpasses the objectives of text-image alignment by assessing the degree to which images represent textual content. Building upon this foundation, we further train an HP (High-Preference) score model using solely the image modality to enhance image aesthetics and detail quality while maintaining text-image alignment. Experiments demonstrate that the proposed evaluation model improves scoring accuracy by over \\(10\\%\\) compared to existing methods, and achieves significant results in optimizing state-of-the-art text-to-image models. This research provides theoretical and empirical support for evolving image generation technology toward higher-order human aesthetic preferences. Code is available at https://github.com/BarretBa/ICTHP." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.797, + 0.242, + 0.812 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.821, + 0.485, + 0.868 + ], + "angle": 0, + "content": "In recent years, generation models have witnessed rapid development and found widespread applications across diverse domains [11, 17, 23, 29, 32, 42, 44, 47]. Modern" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.369, + 0.706, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.37, + 0.904, + 0.47 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.47, + 0.706, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.706, + 0.47, + 0.901, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.638 + ], + "angle": 0, + "content": "Figure 1. Reward Model Scoring Paradox Across Multiple Models Based on CLIP [26] (CLIP Score, PickScore) and BLIP [15] (ITM Score, ImageReward). Refine images are generated with refined prompts by LLMs." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.907, + 0.854 + ], + "angle": 0, + "content": "high-performance diffusion models can not only accurately align texts and images but also generate visually impressive results. As the technology matures, the optimization focus of diffusion models has shifted from text-image alignment and prompt fidelity to generating high-quality images that better meet human aesthetic expectations [9, 19, 20, 37, 39]. In terms of optimization strategies, traditional data-driven training methods [3, 25, 30] are gradually being supplemented and expanded by reward model-based gradient backpropagation and reinforcement learning algorithms [34, 46]. Therefore, constructing a reward function that can precisely evaluate image quality is crucial for further improving the performance of diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "For text-to-image evaluation, text-image alignment (i.e., instruction compliance) serves as the fundamental requirement, while human preference determines the" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.298, + 0.887 + ], + "angle": 0, + "content": "*This work was done at iN2X." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.258, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.298, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19022" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "upper limit of image quality. Current reward models exhibit a critical flaw in measuring text-image alignment: they presume potential equivalence between the information content of images and text, whereas in reality, textual descriptions can hardly comprehensively capture all visual information in images. A direct manifestation is that for the same prompt, a basic image merely covering textual descriptions typically receives higher scores than a higher-quality image that contains richer details and visual elements extending beyond the literal description of the prompt. This phenomenon persists not only in foundational models [15, 26] but also affects fine-tuned variants with human preferences like ImageReward [40] and PickScore [13] (as shown in Fig. 1). Thus, previous reward models may effectively enhance Stable Diffusion-1.5/2.0 [29] (where text-image alignment remains imperfect) but degrade state-of-the-art large models like Stable Diffusion-3.5 [1], steering them to generate visually sparse and aesthetically deficient images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.483, + 0.653 + ], + "angle": 0, + "content": "To address the above issue and continue leveraging human preference data to train superior reward models, we propose a novel optimization objective that transcends text-image alignment, i.e., the image-contained-text score (ICT score). ICT score quantifies the extent to which an image contains textual information to evaluate the quality of generated images. By focusing on how well the image encapsulates textual semantics rather than enforcing bidirectional equivalence, ICT score prevents penalizing images that naturally contain richer contextual details beyond literal prompt descriptions, which aligns with human perception of high-quality outputs. The ICT score removes the ceiling constraints of human preference models, enabling us to incorporate ultra-high-quality preference data to train an advanced human preference model, achieving a more comprehensive evaluation system." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.656, + 0.483, + 0.687 + ], + "angle": 0, + "content": "The main contributions of this research can be summarized in the following three points:" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.69, + 0.483, + 0.795 + ], + "angle": 0, + "content": "- We propose a novel image generation evaluation framework, including ICT score (quantifying the extent to which an image contains textual information) and HP score (evaluating how well an image meets human aesthetic preferences), fundamentally solving the problem of existing reward models underscoring high-quality images." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.796, + 0.483, + 0.856 + ], + "angle": 0, + "content": "- A large-scale high-quality dataset (Pick-High) is constructed using chain-of-thought reasoning from large language models, providing a comprehensive benchmark for image generation quality assessment." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.857, + 0.483, + 0.901 + ], + "angle": 0, + "content": "- We successfully apply the proposed reward model to optimize a state-of-the-art diffusion model (Stable Diffusion-3.5-turbo), demonstrating the effic" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.092, + 0.905, + 0.122 + ], + "angle": 0, + "content": "tiveness and superiority of this method in practical generation tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.137, + 0.679, + 0.154 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.165, + 0.905, + 0.526 + ], + "angle": 0, + "content": "Image quality evaluation for generative models. The evaluation of generated images primarily falls into two categories: human assessment and automated metrics. While the former serves as the gold standard, its high cost and subjective biases have led to the widespread application of the latter in large-scale evaluations. Image-generation models have seen significant quality improvements, evolving from GAN-based approaches such as AttnGAN [41] and DM-GAN [48] to diffusion-based models including DALL·E 2 [27], Stable Diffusion [7, 29], Midjourney [18], and FLUX [14]. Traditional metrics like FID [10] and IS [31] focus solely on distributional statistical properties and fail to capture human subjective preferences. Addressing this limitation, researchers have developed specialized human preference reward models trained on human preference data, such as PickScore [13], HPSv2 [38], and ImageReward [40]. However, we have observed that these reward models exhibit significant biases when evaluating high-quality images generated through prompt optimization, indicating they have not yet effectively guided diffusion models in generating images that align with human preferences for detail and content richness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.528, + 0.905, + 0.77 + ], + "angle": 0, + "content": "Image quality evaluation datasets. Addressing the limitations of existing human preference models, high-quality evaluation datasets are essential for generative model assessment. As generative models rapidly evolved, researchers shifted toward constructing datasets based on human preferences, including the PickaPic_v2 dataset [13], ImageRewardDB [40] developed by the ImageReward team, and the HPDv2 dataset [38] employed by HPS. Despite their scale, these datasets predominantly contain lower-quality images inadequate for evaluating high-quality outputs from modern diffusion models and lack diversity across quality levels. To address these issues, we constructed a 600GB high-quality dataset called Pick-High, incorporating portions of the Pick-a-Pic dataset during training to ensure a comprehensive quality distribution." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.784, + 0.9, + 0.802 + ], + "angle": 0, + "content": "3. Limitations of Text-image Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Multimodal models such as CLIP employ discriminative contrastive loss during pre-training to achieve text-image alignment. According to the information bottleneck theory, their encoders extract minimal image discriminative information to maximize mutual information with text. This training mechanism causes CLIP" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19023" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.482, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.202, + 0.483, + 0.231 + ], + "angle": 0, + "content": "Figure 2. Text-Image Feature Interaction Venn Diagram: Increasing Image Information Under Fixed Text Prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.246, + 0.484, + 0.321 + ], + "angle": 0, + "content": "to exhibit bias when evaluating the same prompt: it tends to assign higher scores to simple images that merely faithfully reflect textual content, while giving lower scores to high-quality images rich in details, contradicting human aesthetic preferences." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.367 + ], + "angle": 0, + "content": "According to the principle of information decomposition, the information content of a generated image can be divided into:" + }, + { + "type": "equation", + "bbox": [ + 0.205, + 0.38, + 0.482, + 0.397 + ], + "angle": 0, + "content": "\\[\nI (v) = I (v; t) + I (v | t), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.483 + ], + "angle": 0, + "content": "where \\( I(v; t) \\) represents the mutual information between the image \\( v \\) and text \\( t \\) (the image-text alignment component), and \\( I(v|t) \\) represents the additional visual information contained in the image (aesthetics, texture, atmosphere, etc.)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.515 + ], + "angle": 0, + "content": "The CLIP scoring mechanism, based on cosine similarity, can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.524, + 0.484, + 0.562 + ], + "angle": 0, + "content": "\\[\n\\mathrm {C L I P} (v, t) \\approx \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot I (v)}} = \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot (I (v ; t) + I (v | t))}}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.57, + 0.484, + 0.766 + ], + "angle": 0, + "content": "As shown in Figure 2, as the capability of generative models improves, the information content of images increases. In the transition from image-text misalignment to alignment, \\( I(v; t) \\) increases faster than \\( I(v) \\), gradually raising CLIP scores. However, when high-quality models generate images rich in visual details, although \\( I(v; t) \\) increases, \\( I(v|t) \\) grows even faster, causing the denominator \\( \\sqrt{I(v)} \\) to increase more than the numerator \\( I(v; t) \\), resulting in an overall decrease in scores, creating a contradiction with human preferences. This motivates us to design a novel objective to train foundation models from the perspective of reward models, which will be elaborated in the next section." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.78, + 0.198, + 0.795 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.901 + ], + "angle": 0, + "content": "This section presents our proposed methodological framework. In Section 4.1, we introduce Pick-High dataset, a large-scale high-quality dataset constructed using refined prompts, which addresses the evaluation bias in existing reward models when assessing information-rich images. In Section 4.2, to resolve the inconsistency between the image-text alignment" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.092, + 0.903, + 0.238 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.25, + 0.905, + 0.276 + ], + "angle": 0, + "content": "Figure 3. The Image-Contained-Text (ICT) Scoring Framework." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.278, + 0.907, + 0.43 + ], + "angle": 0, + "content": "objective of foundation models and evaluation goals of reward models, we propose the Image-Contained-Text (ICT) optimization objective. This enhancement adapts the CLIP model for better evaluation of generative models. We further develop an innovative HP model that relies solely on image modality for human preference assessment. Finally, in Section 4.3, we detail how we utilize the improved reward model to optimize diffusion models, thus enhancing the overall quality of generated images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.434, + 0.692, + 0.45 + ], + "angle": 0, + "content": "4.1. Pick-High Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.679 + ], + "angle": 0, + "content": "Existing reward models (including PickScore, HPSv2, and ImgReward) inherit the limitations of foundational multimodal models, i.e., assigning unreasonably low scores when evaluating complex images paired with simple prompts, which significantly deviates from human preferences. To address this deficiency, we curated 360,000 prompts from the PickAPic_v2 dataset and leveraged large language models' chain-of-thought capabilities to meticulously design a more refined set of prompts that closely align with human preferences. We input these optimized prompts into state-of-the-art image generation models to create 360,000 high-quality images, forming the Pick-High dataset. The detailed construction processes of Pick-High dataset can be seen in Appendix A." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.907, + 0.817 + ], + "angle": 0, + "content": "PickAPic_v2 contains pairwise preferences of images. We denote the non-preferred low-quality image as \\( I_{1} \\) and the preferred image as \\( I_{2} \\). The Pick-High dataset extends pairwise preferences to triplet rankings by introducing a high-quality image \\( I_{3} \\) generated with refined prompts. Based on these quality levels, we establish a hierarchical scoring mechanism for both basic prompts \\( P_{\\mathrm{easy}} \\) and refined prompts \\( P_{\\mathrm{ref}} \\), to better align human preferences." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.828, + 0.849, + 0.845 + ], + "angle": 0, + "content": "4.2. Reward Model Optimization Objective" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.851, + 0.691, + 0.866 + ], + "angle": 0, + "content": "4.2.1. ICT Scores Design" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In text-to-image model evaluation, existing foundation models (e.g., CLIP, BLIP) employ instance-level" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19024" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.089, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "image-text contrastive loss, which presents a fundamental issue: they tend to penalize high-quality images containing additional details. This evaluation mechanism significantly deviates from human preferences and reward model objectives. Based on these observations, we propose the Image-Contained-Text (ICT) scoring mechanism to quantify the extent to which an image expresses textual content." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.212, + 0.482, + 0.333 + ], + "angle": 0, + "content": "The overview of our ICT scoring mechanism is shown in Fig. 3. We assign different scores to images \\((I_1, I_2, I_3)\\) based on their alignment quality with prompts. For high-quality images \\(I_3\\), we assign perfect scores for both basic prompts and refined prompts. To calculate ICT scores of \\(I_2\\) and \\(I_1\\) with basic prompts, we first develop a threshold-based mechanism instead of directly adopting CLIP similarity function:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.339, + 0.482, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (I, P) = \\min (\\frac {\\operatorname {C L I P} (I , P)}{\\theta}, 1), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.373, + 0.483, + 0.495 + ], + "angle": 0, + "content": "where \\(\\mathrm{CLIP}(\\mathrm{I},\\mathrm{P})\\) denotes the CLIP similarity function of an image \\(I\\) and its corresponding prompt \\(P\\) and \\(\\theta\\) represents the alignment threshold. This approach leverages the effectiveness of the CLIP model for assessment of low- and medium-quality images while circumventing the bias in CLIP scores of high-quality images. The basic prompt ICT scores \\((E_1,E_2,E_3)\\) of \\(I_{1}\\), \\(I_{2}\\), \\(I_{3}\\) containing \\(P_{\\mathrm{easy}}\\) are defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.503, + 0.481, + 0.517 + ], + "angle": 0, + "content": "\\[\nE _ {3} = 1, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.521, + 0.481, + 0.537 + ], + "angle": 0, + "content": "\\[\nE _ {2} = \\mathcal {C} \\left(I _ {2}, P _ {\\text {e a s y}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.191, + 0.541, + 0.481, + 0.556 + ], + "angle": 0, + "content": "\\[\nE _ {1} = \\min \\left(\\mathcal {C} \\left(I _ {1}, P _ {\\text {e a s y}}\\right), E _ {2}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.562, + 0.483, + 0.638 + ], + "angle": 0, + "content": "As for refined prompt ICT scores, based on basic prompt ICT scores, we further consider the CLIP similarities of \\( P_{\\mathrm{easy}} \\) and \\( P_{\\mathrm{ref}} \\), defining the scores of \\( I_2 \\) and \\( I_1 \\) containing \\( P_{\\mathrm{ref}} \\) as the product of \\( E_2 \\), \\( E_1 \\) and the textual similarity, as shown below:" + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.66, + 0.481, + 0.675 + ], + "angle": 0, + "content": "\\[\nR _ {3} = 1, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.679, + 0.481, + 0.695 + ], + "angle": 0, + "content": "\\[\nR _ {2} = E _ {2} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right), \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.698, + 0.481, + 0.714 + ], + "angle": 0, + "content": "\\[\nR _ {1} = E _ {1} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Unlike previous discriminative training objectives focused on achieving binary yes/no text-image alignment, our proposed ICT scores categorize the alignment quality between images and texts into distinct levels. This novel objective not only encourages models to achieve image-text alignment but also avoids penalizing images containing additional details beyond textual descriptions. It represents a superior training objective that goes beyond conventional text-image alignment metrics. In the following sections, we will elaborate on implementing these ICT scores as innovative labeling mechanisms for training multimodal models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.092, + 0.638, + 0.105 + ], + "angle": 0, + "content": "4.2.2. ICT Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.11, + 0.905, + 0.186 + ], + "angle": 0, + "content": "We finetune the CLIP model on Pick-High and Pickapic_v2 datasets to obtain our ICT model. Each training sample consists of an image triplet with corresponding ICT labels \\((E_1, E_2, E_3)\\) for the basic prompt and \\((R_1, R_2, R_3)\\) for the refined prompt." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.186, + 0.906, + 0.292 + ], + "angle": 0, + "content": "The model computes ICT scores between each image and both prompts: \\( y_{i,e} = s(P_{\\mathrm{easy}},I_i) \\) and \\( y_{i,r} = s(P_{\\mathrm{ref}},I_i) \\), where \\( i \\in \\{1,2,3\\} \\), and the scoring function \\( s \\) is defined as \\( s(P,I) = \\mathrm{Enc}_{\\mathrm{xt}}(P) \\cdot \\mathrm{Enc}_{\\mathrm{img}}(I) / \\tau \\), where \\( \\tau \\) is the temperature parameter. We optimize the model by minimizing the mean squared error between predicted scores and ICT scores:" + }, + { + "type": "equation", + "bbox": [ + 0.566, + 0.315, + 0.905, + 0.356 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {I C T}} = \\sum_ {i = 1} ^ {3} (E _ {i} - y _ {i, e}) ^ {2} + \\sum_ {i = 1} ^ {3} (R _ {i} - y _ {i, r}) ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.907, + 0.547 + ], + "angle": 0, + "content": "To enhance the discriminative capability of the model, we introduce an in-batch negative sample learning strategy. For any combination of images and unpaired prompts within a batch, we set their ICT labels to 0 and compute the negative sample loss \\(\\mathcal{L}_{\\mathrm{neg}} = \\sum_{j\\neq i}[(y_{j,e}^i)^2 +(y_{j,r}^i)^2 ]\\) where \\(y_{j,e}^{i}\\) and \\(y_{j,r}^{i}\\) represent the ICT scores between the image of the i-th sample and the basic/Refined prompts of the j-th sample, respectively. Considering the potential existence of false negative samples within batches, we adopt a smooth weighting strategy based on the sigmoid function for negative sample mining:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.557, + 0.905, + 0.587 + ], + "angle": 0, + "content": "\\[\nw (y) = \\frac {1}{1 + e ^ {\\alpha (| y | - \\beta)}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.696 + ], + "angle": 0, + "content": "where \\(\\alpha\\) is the smoothing coefficient controlling the steepness of the weight curve, \\(\\beta\\) is the threshold parameter, and \\(y\\) is the predicted score of the sample. This function assigns lower weights to potential false negative samples with high predicted scores, thereby mitigating their negative impact on model training. The final negative sample loss function is:" + }, + { + "type": "equation", + "bbox": [ + 0.547, + 0.719, + 0.905, + 0.753 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {n e g}} = \\sum_ {j \\neq i} \\left(w \\left(y _ {j, e} ^ {i}\\right) \\cdot y _ {j, e} ^ {i} ^ {2} + w \\left(y _ {j, r} ^ {i}\\right) \\cdot y _ {j, r} ^ {i} ^ {2}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.762, + 0.853, + 0.777 + ], + "angle": 0, + "content": "The complete loss function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.629, + 0.789, + 0.905, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {I C T}} + \\lambda \\mathcal {L} _ {\\text {n e g}}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.815, + 0.905, + 0.845 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a balancing factor that controls the contribution of negative sample learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.632, + 0.866 + ], + "angle": 0, + "content": "4.2.3. HP Model" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "When the image fully expresses the semantic content of the prompt, the ICT score reaches its upper limit. To" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19025" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.485, + 0.277 + ], + "angle": 0, + "content": "
ModelAverage↑Pairwise Comparison↑
I2>I1I3>I2I3>I1
Random50.0050.0050.0050.00
ICT-groundtruth87.5562.8099.8799.98
CLIP [26]60.3064.2952.8063.79
ITC [15]57.3663.3348.4260.31
ITM [15]63.3963.8858.2867.99
Aesthetic Score168.1950.9077.8175.83
ImageReward [40]63.8164.5858.0268.84
HPS_v2 [38]72.8866.9871.6380.04
PickScore [13]79.0474.8075.3786.94
ICT87.5864.65100100
HP88.4764.97100100
ICT-HP88.8466.42100100
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.287, + 0.483, + 0.33 + ], + "angle": 0, + "content": "Table 1. Preference prediction accuracy on test sets of Pick-High and Pickapic_v2. Bold and underlined values represent optimal and second-best performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.338, + 0.484, + 0.445 + ], + "angle": 0, + "content": "further evaluate the quality of images, we focus solely on the image modality. For each preference triplet, we employ margin ranking loss to fine-tune both the CLIP model and its subsequent MLP layers. Specifically, given a triplet \\(\\{I_1,I_2,I_3\\}\\), where \\(I_{3}\\) is preferred over \\(I_{2}\\), and \\(I_{2}\\) is preferred over \\(I_{1}\\), we enforce the preference margins as:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.458, + 0.481, + 0.5 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m a r g i n}} = \\sum [ \\max (0, - \\Delta (I _ {2}, I _ {1}) + m) + \\\\ \\left. \\max \\left(0, - \\Delta \\left(I _ {3}, I _ {2}\\right) + m\\right) \\right], \\tag {13} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.511, + 0.483, + 0.543 + ], + "angle": 0, + "content": "where \\(\\Delta(I_i, I_j)\\) represents the score difference between images \\(I_i\\) and \\(I_j\\), \\(m\\) is the margin threshold." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.358, + 0.568 + ], + "angle": 0, + "content": "4.3. Diffusion Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.483, + 0.62 + ], + "angle": 0, + "content": "We adopt DRaFT-K [5] as our fine-tuning method to optimize Stable-Diffusion-3.5-Large-Turbo [2] by directly maximizing the differentiable reward functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.696 + ], + "angle": 0, + "content": "Our proposed ICT model and HP model can each be individually utilized as reward functions. Additionally, we multiply the outputs of both models to create a more comprehensive integrated reward function, which we term the ICT-HP model." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.242, + 0.728 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.737, + 0.297, + 0.754 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.865 + ], + "angle": 0, + "content": "Two-Stage Training Pipeline. We adopt a two-stage training method for the ICT model and HP model. In the first stage, we fine-tune the CLIP model using our constructed ICT labels as training targets. In the second stage, we freeze the ICT model parameters and continue training the HP model. During the model optimization phase, we comparatively evaluate" + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.089, + 0.906, + 0.198 + ], + "angle": 0, + "content": "
ModelJPEG Compressibility↑Aesthetic Score↑
SD3.5-Large [1]374.806.307
FLUX.1-dev [14]270.586.436
SD3.5-Turbo [2]313.106.293
+ CLIP [26](crash)279.985.175
+ ICT (Ours)321.396.285
+ ICT-HP (Ours)330.236.300
+ HP (Ours)334.866.448
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.207, + 0.905, + 0.25 + ], + "angle": 0, + "content": "Table 2. JPEG Compressibility and Aesthetic Scores on Parti-Prompts. Bold and underlined values represent optimal and second-best performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.259, + 0.905, + 0.305 + ], + "angle": 0, + "content": "three different reward models for optimizing the diffusion model LoRA adapters: ICT model, HP model, and the more comprehensive ICT-HP model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Training Datasets. Our training dataset for the ICT model consists of the Pick-High dataset and a subset of Pickapic_v2, comprising a total of 360,000 image triplets. For the diffusion model optimization process, we extract 32,000 non-repetitive text prompts from the Pick-High dataset to serve as our prompt dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.905, + 0.444 + ], + "angle": 0, + "content": "For detailed training hyperparameter configurations and experimental specifications, please refer to the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.455, + 0.768, + 0.469 + ], + "angle": 0, + "content": "5.2. Reward Model Performance" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.734 + ], + "angle": 0, + "content": "We evaluated the prediction accuracy of human preference models on the Pick-High and Pickapic_v2 test sets. As shown in Table 1, the image-text contrastive (ITC) classification heads of foundation multimodal pre-trained models CLIP and BLIP perform only comparable to random selection. This result indicates that the instance-level image-text alignment objective adopted by foundation multimodal models is not suitable for evaluating the quality of generative models. In terms of overall average accuracy, our proposed ICT-HP model achieves nearly \\(10\\%\\) performance improvement compared to Pickscore. This significant improvement stems from the ability of our reward model to measure the extent to which images contain text, and on this basis, continue to measure image quality, thus more accurately identifying high-quality images and assigning them correspondingly high scores." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results on SD3.5-Turbo Optimization. In Figure 4, we present generation results using real user prompts. All images are generated with the same initial noise to ensure fair comparison. The experimental results show that: the HPS_v2 model only enhances image tone with minimal detail improvement; the PickScore model, while increasing details, causes distortions in style and color fidelity; the ImageReward model shows almost no substantial improvement over the original model outputs. In contrast, our proposed reward models achieve performance improvements in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.875, + 0.436, + 0.901 + ], + "angle": 0, + "content": "1https://github.com/christophschuhmann/improved-aesthetic-predictor" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19026" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.089, + 0.885, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.136, + 0.437, + 0.86, + 0.452 + ], + "angle": 0, + "content": "Figure 4. Qualitative Comparison of Optimization Results Across Reward Models Using Real User Prompts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.467, + 0.483, + 0.634 + ], + "angle": 0, + "content": "both textural details and tonal aesthetics. Specifically, the ICT model does not penalize increases in image information content, allowing diffusion models to continue optimizing toward richer visual content after achieving text-image alignment; meanwhile, the HP model effectively utilizes pure image features to further optimize when the ICT model reaches its performance ceiling. Compared to our approach, existing reward models tend to produce negative feedback on high-quality image features, thereby limiting their ability to enhance the performance of advanced diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.639, + 0.483, + 0.866 + ], + "angle": 0, + "content": "Quantitative Evaluation of Visual Information Density and Aesthetic Fidelity. To quantitatively evaluate the enhancement effect of our reward model on the original model in terms of generated image information content and aesthetic quality, we employed two objective metrics, JPEG Compressibility [12] and Aesthetic Score, for assessment on the Parti-Prompts dataset [43]. As shown in Table 2, among all tested models, the HP model achieved the most significant improvements in both metrics, followed by the ICT-HP model. Notably, the SD3.5-Turbo optimized by the HP model not only demonstrated substantial improvement in aesthetic metrics, but even slightly surpassed more comprehensive baseline models such as SD3.5-Large and FLUX.1-dev." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Qualitative Comparison of Simple Elements Generation Results. We conducted qualitative ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.468, + 0.907, + 0.544 + ], + "angle": 0, + "content": "periments on the generation of simple elements, with results shown in Figure 5. The ICT model can stably handle simple geometric shapes and effectively penalize complex elements that do not meet simplicity requirements when prompts include descriptors like \"simple\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.806 + ], + "angle": 0, + "content": "Comprehensive Evaluation on GenEval Benchmark. Table 3 presents the quantitative performance of various image generation models within the multidimensional evaluation framework of GenEval [8]. Using SD3.5-Turbo as the baseline, our ICT-HP and ICT models both demonstrate significant advantages. The ICT-HP model excels in multi-object generation and color fidelity; while the ICT model achieves the highest mean score among all models and delivers optimal performance on the Counting metric, showcasing its superior capability in precise object quantity control. In contrast, the CLIP model suffers comprehensive performance degradation due to training collapse, and PickScore exhibits weaker performance in the Colors metric due to limitations in tonal representation. These comparisons validate our methods' significant advantages in terms of text-image consistency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Quantitative and Qualitative Performance Comparison of Multi-metric Reward Models. We conducted quantitative analyses of optimization results across various reward models, as shown in Table 4. Experiments reveal that existing reward models including HPS, PickScore, and ImageReward ex" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19027" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.284 + ], + "angle": 0, + "content": "
ModelMean↑Single↑Two↑Counting↑Colors↑Position↑Color Attribution↑
SDXL [22]0.550.980.740.390.850.150.23
DALL-E 2 [28]0.520.940.660.490.770.100.19
DALL-E 3 [21]0.670.960.870.470.830.430.45
SD3 [7]0.680.980.840.660.740.400.43
SD3.5-Turbo [2]0.690.990.870.690.800.250.55
+ ImageReward0.700.990.870.680.800.280.59
+ HPS_v20.690.980.860.680.810.270.57
+ PickScore0.660.990.850.670.740.240.48
+ ICT-HP (Ours)0.700.990.880.680.790.280.58
+ CLIP(crash)0.130.380.040.060.260.010.03
+ ICT (Ours)0.710.980.880.700.810.310.56
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.295, + 0.905, + 0.325 + ], + "angle": 0, + "content": "Table 3. Quantitative Evaluation of Optimization Results Across Reward Models on GenEval Benchmark. Bold and underlined values represent optimal and second-best performance." + }, + { + "type": "image", + "bbox": [ + 0.104, + 0.347, + 0.472, + 0.611 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.621, + 0.483, + 0.649 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison of Simple Elements Generation Results." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.672, + 0.483, + 0.884 + ], + "angle": 0, + "content": "hibit performance degradation across multiple metrics. In contrast, our ICT-HP and HP models achieved improvements across all metrics, demonstrating the stability advantage of our method. Although PickScore obtained the highest aesthetic score, qualitative analysis and GenEval experiments indicated significant distortion in artistic style and tone, with color-related metrics significantly lower than other models. Figure 6 compares the original image, refined image, and our model-optimized results. ICT-HP and HP models not only generate intricate backgrounds that are difficult to describe directly through language but also demonstrate a more precise and comprehensive understanding of prompts, thus producing higher-quality images." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Feature Transfer Validation. As illustrated in Fig-" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.347, + 0.887, + 0.615 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.628, + 0.908, + 0.671 + ], + "angle": 0, + "content": "Figure 6. Qualitative Comparison: Original Images, Refined Images, and Generations Optimized by Our Reward Models (HP/ICT-HP)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ure 7, transplanting the Text-Encoder trained by our ICT model into the SD2.1-base model significantly improves image quality. This demonstrates that our encoder integrates fine-grained visual details, thus transforming textual descriptions into more precise visual guidance signals. Figure 8 presents results of style transfer with InstantStyle [35]. The CLIP image encoder is replaced by our ICT image encoder. Our fine-tuned encoder not only precisely injects reference styles but also balances text descriptions, overcoming the style-content inconsistency. These experiments validate the effectiveness and advancement of ICT score training objectives." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19028" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.216 + ], + "angle": 0, + "content": "
ModelCLIP↑ITM↑ImgRwd↑HPS↑Pick↑Aes↑ICT↑HP↑ICT-HP↑
SD3.5-Te0.3400.8310.9550.27722.0346.5550.9100.7770.717
SD3.5-Tr0.326↓0.757↓0.914↓0.276↓22.012↓6.780↑0.923↑0.781↑0.718↑
+HPS_v20.343↑0.835↑1.065↑0.282↑22.000↓6.590↑0.929↑0.777-0.722↑
+PickScore0.332↓0.816↓1.106↑0.277-22.719↑6.924↑0.929↑0.779↑0.723↑
+ImageReward0.340-0.841↑1.056↑0.279↑22.022↓6.569↑0.865↓0.763↓0.663↓
+ICT(Ours)0.344↑0.841↑1.014↑0.277-22.117↑6.561↑0.937↑0.776↓0.728↑
+ICT-HP(Ours)0.345↑0.839↑1.011↑0.277-22.147↑6.558↑0.938↑0.778↑0.730↑
+HP(Ours)0.342↑0.836↑1.069↑0.278↑22.231↑6.668↑0.937↑0.779↑0.730↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.226, + 0.907, + 0.257 + ], + "angle": 0, + "content": "Table 4. Quantitative Results. SD3.5-Te/Tr denote images generated using easy/reflect prompts on SD-3.5-Turbo. Arrows (↑, ↓, -) indicate changes vs. SD3.5-Te. Bold and underlined values represent optimal and second-best performance." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.28, + 0.481, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.56, + 0.483, + 0.589 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison of ICT Text-Encoder Direct Migration to SD2.1-Base Model." + }, + { + "type": "table", + "bbox": [ + 0.114, + 0.598, + 0.462, + 0.713 + ], + "angle": 0, + "content": "
ComparisonModelsWin Rate
vs. BaseICT model0.562
HP model0.709
ICT-HP model0.673
vs. PickScoreHP model0.617
ICT-HP model0.714
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.722, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Table 5. Human evaluation on 300 randomly selected samples from Parti-Prompts [43]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.216, + 0.772 + ], + "angle": 0, + "content": "5.3. User-Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We conduct a user study with ten annotators on 300 randomly selected prompts from Parti-Prompts [43]. Image pairs (ours vs. baseline) were randomly ordered, and annotators evaluated prompt expression and visual appeal. As shown in Table 5, our methods achieve higher win rates against both the SD3.5-large-Turbo base model and the PickScore-optimized model, confirming our approach's effectiveness." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.281, + 0.905, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.558, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 8. Qualitative Comparison of HP/ICT ImageEncoder Performance in Style Injection Tasks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.602, + 0.651, + 0.617 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.629, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we reveal the reward-scoring paradox of high-quality images and show that the inherent flaw in current text-to-image generation reward models lies in classic text-image alignment paradigms. We propose a novel training objective for reward models, i.e., ICT (Image-Contained-Text) score. This objective encourages models to achieve text-image alignment and avoid penalizing high-quality images containing extra details, that better match human preferences. Based on the ICT model, we further train a preference model that relies solely on image modality for human preference assessment. Our experiments show that the proposed approach improves scoring accuracy by over \\(10\\%\\) compared to existing methods and markedly enhances state-of-the-art models like SD3.5-Turbo. Through both theoretical foundation and empirical validation, our work advances image generation technology toward better alignment with human preferences." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19029" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.486, + 0.186 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China No. 62376277, the Public Computing Cloud of Renmin University of China and the Fund for Building World-Class Universities (Disciplines) of Renmin University of China." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.2, + 0.227 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.237, + 0.482, + 0.278 + ], + "angle": 0, + "content": "[1] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.281, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[2] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.324, + 0.482, + 0.363 + ], + "angle": 0, + "content": "[3] Yalong Bai, Mohan Zhou, and Qing Yang. Styleinject: Parameter efficient tuning of text-to-image diffusion models. ArXiv, abs/2401.13942, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.366, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 2818-2829. IEEE, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.464, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards, 2024. 5, 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.507, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.549, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.634, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[8] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. CoRR, abs/2310.11513, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.103, + 0.69, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[9] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[12] ISO/IEC. Information technology - digital compression and coding of continuous-tone still images - requirements and guidelines. International Organization for Standardization, 1992. ISO/IEC 10918-1:1992. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.237, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.904, + 0.262 + ], + "angle": 0, + "content": "[16] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 11, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.265, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Biye Li, Huaian Chen, and Yi Jin. Star: Scale-wise text-conditioned autoregressive image generation. arXiv preprint arXiv:2406.10797, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.322, + 0.902, + 0.349 + ], + "angle": 0, + "content": "[18] Midjourney. Ai-powered image generation. https://www.midjourney.com/, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.351, + 0.906, + 0.42 + ], + "angle": 0, + "content": "[19] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, JiRong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26617-26626, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.422, + 0.906, + 0.516 + ], + "angle": 0, + "content": "[20] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, and Ji-Rong Wen. Uniform attention maps: Boosting image fidelity in reconstruction and editing. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pages 4420-4429. IEEE, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.904, + 0.546 + ], + "angle": 0, + "content": "[21] OpenAI. Dall-e 3. https://openai.com/dall-e-3, 2023.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.549, + 0.905, + 0.604 + ], + "angle": 0, + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.606, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[24] Wenwen Qiang, Jingyao Wang, Zeen Song, Jiangmeng Li, and Changwen Zheng. On the out-of-distribution generalization of self-supervised learning. arXiv preprint arXiv:2505.16675, 2025. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.747, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on ma" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19030" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "chine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.181, + 0.482, + 0.223 + ], + "angle": 0, + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.225, + 0.483, + 0.309 + ], + "angle": 0, + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.311, + 0.483, + 0.395 + ], + "angle": 0, + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.397, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.483, + 0.484 + ], + "angle": 0, + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.485, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[33] Zeen Song, Siyu Zhao, Xingyu Zhang, Jiangmeng Li, Changwen Zheng, and Wenwen Qiang. Learning invariant causal mechanism from vision-language models. arXiv preprint arXiv:2405.15289, 2024. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.543, + 0.483, + 0.641 + ], + "angle": 0, + "content": "[34] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.643, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[35] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.701, + 0.483, + 0.758 + ], + "angle": 0, + "content": "[36] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models, 2023. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[37] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[39] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[40] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[41] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1316-1324. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.331, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[42] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 3954-3964. ACM, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.416, + 0.905, + 0.527 + ], + "angle": 0, + "content": "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.528, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[44] Guiwei Zhang, Tianyu Zhang, Mohan Zhou, Yalong Bai, and Biye Li. V2flow: Unifying visual tokenization and large language model vocabularies for autoregressive image generation. arXiv preprint arXiv:2503.07493, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.598, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[45] Yanan Zhang, Jiangmeng Li, Lixiang Liu, and Wenwen Qiang. Rethinking misalignment in vision-language model adaptation from a causal perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 11" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.669, + 0.905, + 0.724 + ], + "angle": 0, + "content": "[46] Ziyi Zhang, Li Shen, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Bo Du, and Dacheng Tao. Aligning few-step diffusion models with dense reward difference learning. arXiv preprint arXiv:2411.11727, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.725, + 0.905, + 0.794 + ], + "angle": 0, + "content": "[47] Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Llada 1.5: Variance-reduced preference optimization for large language diffusion models, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.796, + 0.905, + 0.893 + ], + "angle": 0, + "content": "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. DM-GAN: dynamic memory generative adversarial networks for text-to-image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5802-5810. Computer Vision Foundation / IEEE, 2019. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.893 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19031" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_origin.pdf b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c6aa96e612a0c61c9fbbf61a0e2a4ce0fc56274d --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/4fbc7473-ddd0-4038-9d67-401857bfa6dc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cb61e496885eade1a9300aa6c7cdfa004c26a26f8cde3cc923b72597a9c3a114 +size 6933681 diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/full.md b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c040f9b64cc6f7977f3abedb64fe8645a3baf0ed --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/full.md @@ -0,0 +1,317 @@ +# Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment + +Ying Ba $^{1,2,3}$ , Tianyu Zhang $^{4}$ , Yalong Bai $^{4}$ , Wenyi Mo $^{1,2,3}$ , Tao Liang $^{4}$ , Bing Su $^{1,2,3\dagger}$ , Ji-Rong Wen $^{1,2,3}$ $^{1}$ Gaoling School of Artificial Intelligence Renmin University of China Beijing, China + $^{2}$ Beijing Key Laboratory of Research on Large Models and Intelligent Governance + $^{3}$ Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE $^{4}$ iN2X + +$^{1,2,3}$ yingba88@outlook.com, {mowenyi00, subingats}@gmail.com, jrwen@ruc.edu.cn + $^{4}$ {tianyu1949, liang0305tao}@gmail.com, ylbai@outlook.com + +# Abstract + +Contemporary image generation systems have achieved high fidelity and superior aesthetic quality beyond basic text-image alignment. However, existing evaluation frameworks have failed to evolve in parallel. This study reveals that human preference reward models fine-tuned based on CLIP and BLIP architectures have inherent flaws: they inappropriately assign low scores to images with rich details and high aesthetic value, creating a significant discrepancy with actual human aesthetic preferences. To address this issue, we design a novel evaluation score, ICT (Image-Contained-Text) score, that achieves and surpasses the objectives of text-image alignment by assessing the degree to which images represent textual content. Building upon this foundation, we further train an HP (High-Preference) score model using solely the image modality to enhance image aesthetics and detail quality while maintaining text-image alignment. Experiments demonstrate that the proposed evaluation model improves scoring accuracy by over $10\%$ compared to existing methods, and achieves significant results in optimizing state-of-the-art text-to-image models. This research provides theoretical and empirical support for evolving image generation technology toward higher-order human aesthetic preferences. Code is available at https://github.com/BarretBa/ICTHP. + +# 1. Introduction + +In recent years, generation models have witnessed rapid development and found widespread applications across diverse domains [11, 17, 23, 29, 32, 42, 44, 47]. Modern + +![](images/a219871a8298e64b70e94219a657bcf91893f5662a7ca8cfbc77799f7facf632.jpg) + +![](images/1f1812da3ac72273ccac1fb5cfbeed2215436aec9baf9176f4dcb63209992b88.jpg) + +![](images/3decd7529e2444420bc6235cf1c582a1d9652f72c3ee1b9ef8f9086daddf28d0.jpg) +Figure 1. Reward Model Scoring Paradox Across Multiple Models Based on CLIP [26] (CLIP Score, PickScore) and BLIP [15] (ITM Score, ImageReward). Refine images are generated with refined prompts by LLMs. + +![](images/f8b7a98ace20886f9dc4a0cb7e86f5c9ddd00f151b9f42586f9226e98502e6c3.jpg) + +high-performance diffusion models can not only accurately align texts and images but also generate visually impressive results. As the technology matures, the optimization focus of diffusion models has shifted from text-image alignment and prompt fidelity to generating high-quality images that better meet human aesthetic expectations [9, 19, 20, 37, 39]. In terms of optimization strategies, traditional data-driven training methods [3, 25, 30] are gradually being supplemented and expanded by reward model-based gradient backpropagation and reinforcement learning algorithms [34, 46]. Therefore, constructing a reward function that can precisely evaluate image quality is crucial for further improving the performance of diffusion models. + +For text-to-image evaluation, text-image alignment (i.e., instruction compliance) serves as the fundamental requirement, while human preference determines the + +upper limit of image quality. Current reward models exhibit a critical flaw in measuring text-image alignment: they presume potential equivalence between the information content of images and text, whereas in reality, textual descriptions can hardly comprehensively capture all visual information in images. A direct manifestation is that for the same prompt, a basic image merely covering textual descriptions typically receives higher scores than a higher-quality image that contains richer details and visual elements extending beyond the literal description of the prompt. This phenomenon persists not only in foundational models [15, 26] but also affects fine-tuned variants with human preferences like ImageReward [40] and PickScore [13] (as shown in Fig. 1). Thus, previous reward models may effectively enhance Stable Diffusion-1.5/2.0 [29] (where text-image alignment remains imperfect) but degrade state-of-the-art large models like Stable Diffusion-3.5 [1], steering them to generate visually sparse and aesthetically deficient images. + +To address the above issue and continue leveraging human preference data to train superior reward models, we propose a novel optimization objective that transcends text-image alignment, i.e., the image-contained-text score (ICT score). ICT score quantifies the extent to which an image contains textual information to evaluate the quality of generated images. By focusing on how well the image encapsulates textual semantics rather than enforcing bidirectional equivalence, ICT score prevents penalizing images that naturally contain richer contextual details beyond literal prompt descriptions, which aligns with human perception of high-quality outputs. The ICT score removes the ceiling constraints of human preference models, enabling us to incorporate ultra-high-quality preference data to train an advanced human preference model, achieving a more comprehensive evaluation system. + +The main contributions of this research can be summarized in the following three points: + +- We propose a novel image generation evaluation framework, including ICT score (quantifying the extent to which an image contains textual information) and HP score (evaluating how well an image meets human aesthetic preferences), fundamentally solving the problem of existing reward models underscoring high-quality images. +- A large-scale high-quality dataset (Pick-High) is constructed using chain-of-thought reasoning from large language models, providing a comprehensive benchmark for image generation quality assessment. +- We successfully apply the proposed reward model to optimize a state-of-the-art diffusion model (Stable Diffusion-3.5-turbo), demonstrating the effic + +tiveness and superiority of this method in practical generation tasks. + +# 2. Related Work + +Image quality evaluation for generative models. The evaluation of generated images primarily falls into two categories: human assessment and automated metrics. While the former serves as the gold standard, its high cost and subjective biases have led to the widespread application of the latter in large-scale evaluations. Image-generation models have seen significant quality improvements, evolving from GAN-based approaches such as AttnGAN [41] and DM-GAN [48] to diffusion-based models including DALL·E 2 [27], Stable Diffusion [7, 29], Midjourney [18], and FLUX [14]. Traditional metrics like FID [10] and IS [31] focus solely on distributional statistical properties and fail to capture human subjective preferences. Addressing this limitation, researchers have developed specialized human preference reward models trained on human preference data, such as PickScore [13], HPSv2 [38], and ImageReward [40]. However, we have observed that these reward models exhibit significant biases when evaluating high-quality images generated through prompt optimization, indicating they have not yet effectively guided diffusion models in generating images that align with human preferences for detail and content richness. + +Image quality evaluation datasets. Addressing the limitations of existing human preference models, high-quality evaluation datasets are essential for generative model assessment. As generative models rapidly evolved, researchers shifted toward constructing datasets based on human preferences, including the PickaPic_v2 dataset [13], ImageRewardDB [40] developed by the ImageReward team, and the HPDv2 dataset [38] employed by HPS. Despite their scale, these datasets predominantly contain lower-quality images inadequate for evaluating high-quality outputs from modern diffusion models and lack diversity across quality levels. To address these issues, we constructed a 600GB high-quality dataset called Pick-High, incorporating portions of the Pick-a-Pic dataset during training to ensure a comprehensive quality distribution. + +# 3. Limitations of Text-image Alignment + +Multimodal models such as CLIP employ discriminative contrastive loss during pre-training to achieve text-image alignment. According to the information bottleneck theory, their encoders extract minimal image discriminative information to maximize mutual information with text. This training mechanism causes CLIP + +![](images/55f402f6f98c871a998306ef094d3a6ffdc267bcc5776b1cc4ce4e56da716d89.jpg) +Figure 2. Text-Image Feature Interaction Venn Diagram: Increasing Image Information Under Fixed Text Prompts. + +to exhibit bias when evaluating the same prompt: it tends to assign higher scores to simple images that merely faithfully reflect textual content, while giving lower scores to high-quality images rich in details, contradicting human aesthetic preferences. + +According to the principle of information decomposition, the information content of a generated image can be divided into: + +$$ +I (v) = I (v; t) + I (v | t), \tag {1} +$$ + +where $I(v; t)$ represents the mutual information between the image $v$ and text $t$ (the image-text alignment component), and $I(v|t)$ represents the additional visual information contained in the image (aesthetics, texture, atmosphere, etc.). + +The CLIP scoring mechanism, based on cosine similarity, can be expressed as: + +$$ +\mathrm {C L I P} (v, t) \approx \frac {I (v ; t)}{\sqrt {I (t) \cdot I (v)}} = \frac {I (v ; t)}{\sqrt {I (t) \cdot (I (v ; t) + I (v | t))}}. +$$ + +As shown in Figure 2, as the capability of generative models improves, the information content of images increases. In the transition from image-text misalignment to alignment, $I(v; t)$ increases faster than $I(v)$ , gradually raising CLIP scores. However, when high-quality models generate images rich in visual details, although $I(v; t)$ increases, $I(v|t)$ grows even faster, causing the denominator $\sqrt{I(v)}$ to increase more than the numerator $I(v; t)$ , resulting in an overall decrease in scores, creating a contradiction with human preferences. This motivates us to design a novel objective to train foundation models from the perspective of reward models, which will be elaborated in the next section. + +# 4. Method + +This section presents our proposed methodological framework. In Section 4.1, we introduce Pick-High dataset, a large-scale high-quality dataset constructed using refined prompts, which addresses the evaluation bias in existing reward models when assessing information-rich images. In Section 4.2, to resolve the inconsistency between the image-text alignment + +![](images/53796012433464b4102cd66c8d504ce0bd93fa49da5780e7393321998fc2b7ac.jpg) +Figure 3. The Image-Contained-Text (ICT) Scoring Framework. + +objective of foundation models and evaluation goals of reward models, we propose the Image-Contained-Text (ICT) optimization objective. This enhancement adapts the CLIP model for better evaluation of generative models. We further develop an innovative HP model that relies solely on image modality for human preference assessment. Finally, in Section 4.3, we detail how we utilize the improved reward model to optimize diffusion models, thus enhancing the overall quality of generated images. + +# 4.1. Pick-High Dataset + +Existing reward models (including PickScore, HPSv2, and ImgReward) inherit the limitations of foundational multimodal models, i.e., assigning unreasonably low scores when evaluating complex images paired with simple prompts, which significantly deviates from human preferences. To address this deficiency, we curated 360,000 prompts from the PickAPic_v2 dataset and leveraged large language models' chain-of-thought capabilities to meticulously design a more refined set of prompts that closely align with human preferences. We input these optimized prompts into state-of-the-art image generation models to create 360,000 high-quality images, forming the Pick-High dataset. The detailed construction processes of Pick-High dataset can be seen in Appendix A. + +PickAPic_v2 contains pairwise preferences of images. We denote the non-preferred low-quality image as $I_{1}$ and the preferred image as $I_{2}$ . The Pick-High dataset extends pairwise preferences to triplet rankings by introducing a high-quality image $I_{3}$ generated with refined prompts. Based on these quality levels, we establish a hierarchical scoring mechanism for both basic prompts $P_{\mathrm{easy}}$ and refined prompts $P_{\mathrm{ref}}$ , to better align human preferences. + +# 4.2. Reward Model Optimization Objective + +# 4.2.1. ICT Scores Design + +In text-to-image model evaluation, existing foundation models (e.g., CLIP, BLIP) employ instance-level + +image-text contrastive loss, which presents a fundamental issue: they tend to penalize high-quality images containing additional details. This evaluation mechanism significantly deviates from human preferences and reward model objectives. Based on these observations, we propose the Image-Contained-Text (ICT) scoring mechanism to quantify the extent to which an image expresses textual content. + +The overview of our ICT scoring mechanism is shown in Fig. 3. We assign different scores to images $(I_1, I_2, I_3)$ based on their alignment quality with prompts. For high-quality images $I_3$ , we assign perfect scores for both basic prompts and refined prompts. To calculate ICT scores of $I_2$ and $I_1$ with basic prompts, we first develop a threshold-based mechanism instead of directly adopting CLIP similarity function: + +$$ +\mathcal {C} (I, P) = \min (\frac {\operatorname {C L I P} (I , P)}{\theta}, 1), \tag {2} +$$ + +where $\mathrm{CLIP}(\mathrm{I},\mathrm{P})$ denotes the CLIP similarity function of an image $I$ and its corresponding prompt $P$ and $\theta$ represents the alignment threshold. This approach leverages the effectiveness of the CLIP model for assessment of low- and medium-quality images while circumventing the bias in CLIP scores of high-quality images. The basic prompt ICT scores $(E_1,E_2,E_3)$ of $I_{1}$ , $I_{2}$ , $I_{3}$ containing $P_{\mathrm{easy}}$ are defined as follows: + +$$ +E _ {3} = 1, \tag {3} +$$ + +$$ +E _ {2} = \mathcal {C} \left(I _ {2}, P _ {\text {e a s y}}\right), \tag {4} +$$ + +$$ +E _ {1} = \min \left(\mathcal {C} \left(I _ {1}, P _ {\text {e a s y}}\right), E _ {2}\right). \tag {5} +$$ + +As for refined prompt ICT scores, based on basic prompt ICT scores, we further consider the CLIP similarities of $P_{\mathrm{easy}}$ and $P_{\mathrm{ref}}$ , defining the scores of $I_2$ and $I_1$ containing $P_{\mathrm{ref}}$ as the product of $E_2$ , $E_1$ and the textual similarity, as shown below: + +$$ +R _ {3} = 1, \tag {6} +$$ + +$$ +R _ {2} = E _ {2} \times \operatorname {C L I P} \left(P _ {\text {e a s y}}, P _ {\text {r e f}}\right), \tag {7} +$$ + +$$ +R _ {1} = E _ {1} \times \operatorname {C L I P} \left(P _ {\text {e a s y}}, P _ {\text {r e f}}\right). \tag {8} +$$ + +Unlike previous discriminative training objectives focused on achieving binary yes/no text-image alignment, our proposed ICT scores categorize the alignment quality between images and texts into distinct levels. This novel objective not only encourages models to achieve image-text alignment but also avoids penalizing images containing additional details beyond textual descriptions. It represents a superior training objective that goes beyond conventional text-image alignment metrics. In the following sections, we will elaborate on implementing these ICT scores as innovative labeling mechanisms for training multimodal models. + +# 4.2.2. ICT Model + +We finetune the CLIP model on Pick-High and Pickapic_v2 datasets to obtain our ICT model. Each training sample consists of an image triplet with corresponding ICT labels $(E_1, E_2, E_3)$ for the basic prompt and $(R_1, R_2, R_3)$ for the refined prompt. + +The model computes ICT scores between each image and both prompts: $y_{i,e} = s(P_{\mathrm{easy}},I_i)$ and $y_{i,r} = s(P_{\mathrm{ref}},I_i)$ , where $i \in \{1,2,3\}$ , and the scoring function $s$ is defined as $s(P,I) = \mathrm{Enc}_{\mathrm{xt}}(P) \cdot \mathrm{Enc}_{\mathrm{img}}(I) / \tau$ , where $\tau$ is the temperature parameter. We optimize the model by minimizing the mean squared error between predicted scores and ICT scores: + +$$ +\mathcal {L} _ {\mathrm {I C T}} = \sum_ {i = 1} ^ {3} (E _ {i} - y _ {i, e}) ^ {2} + \sum_ {i = 1} ^ {3} (R _ {i} - y _ {i, r}) ^ {2}. \tag {9} +$$ + +To enhance the discriminative capability of the model, we introduce an in-batch negative sample learning strategy. For any combination of images and unpaired prompts within a batch, we set their ICT labels to 0 and compute the negative sample loss $\mathcal{L}_{\mathrm{neg}} = \sum_{j\neq i}[(y_{j,e}^i)^2 +(y_{j,r}^i)^2 ]$ where $y_{j,e}^{i}$ and $y_{j,r}^{i}$ represent the ICT scores between the image of the i-th sample and the basic/Refined prompts of the j-th sample, respectively. Considering the potential existence of false negative samples within batches, we adopt a smooth weighting strategy based on the sigmoid function for negative sample mining: + +$$ +w (y) = \frac {1}{1 + e ^ {\alpha (| y | - \beta)}}, \tag {10} +$$ + +where $\alpha$ is the smoothing coefficient controlling the steepness of the weight curve, $\beta$ is the threshold parameter, and $y$ is the predicted score of the sample. This function assigns lower weights to potential false negative samples with high predicted scores, thereby mitigating their negative impact on model training. The final negative sample loss function is: + +$$ +\mathcal {L} _ {\mathrm {n e g}} = \sum_ {j \neq i} \left(w \left(y _ {j, e} ^ {i}\right) \cdot y _ {j, e} ^ {i} ^ {2} + w \left(y _ {j, r} ^ {i}\right) \cdot y _ {j, r} ^ {i} ^ {2}\right). \tag {11} +$$ + +The complete loss function is formulated as: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\mathrm {I C T}} + \lambda \mathcal {L} _ {\text {n e g}}, \tag {12} +$$ + +where $\lambda$ is a balancing factor that controls the contribution of negative sample learning. + +# 4.2.3. HP Model + +When the image fully expresses the semantic content of the prompt, the ICT score reaches its upper limit. To + +
ModelAverage↑Pairwise Comparison↑
I2>I1I3>I2I3>I1
Random50.0050.0050.0050.00
ICT-groundtruth87.5562.8099.8799.98
CLIP [26]60.3064.2952.8063.79
ITC [15]57.3663.3348.4260.31
ITM [15]63.3963.8858.2867.99
Aesthetic Score168.1950.9077.8175.83
ImageReward [40]63.8164.5858.0268.84
HPS_v2 [38]72.8866.9871.6380.04
PickScore [13]79.0474.8075.3786.94
ICT87.5864.65100100
HP88.4764.97100100
ICT-HP88.8466.42100100
+ +further evaluate the quality of images, we focus solely on the image modality. For each preference triplet, we employ margin ranking loss to fine-tune both the CLIP model and its subsequent MLP layers. Specifically, given a triplet $\{I_1,I_2,I_3\}$ , where $I_{3}$ is preferred over $I_{2}$ , and $I_{2}$ is preferred over $I_{1}$ , we enforce the preference margins as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {m a r g i n}} = \sum [ \max (0, - \Delta (I _ {2}, I _ {1}) + m) + \\ \left. \max \left(0, - \Delta \left(I _ {3}, I _ {2}\right) + m\right) \right], \tag {13} \\ \end{array} +$$ + +where $\Delta(I_i, I_j)$ represents the score difference between images $I_i$ and $I_j$ , $m$ is the margin threshold. + +# 4.3. Diffusion Model Optimization + +We adopt DRaFT-K [5] as our fine-tuning method to optimize Stable-Diffusion-3.5-Large-Turbo [2] by directly maximizing the differentiable reward functions. + +Our proposed ICT model and HP model can each be individually utilized as reward functions. Additionally, we multiply the outputs of both models to create a more comprehensive integrated reward function, which we term the ICT-HP model. + +# 5. Experiments + +# 5.1. Experimental Settings + +Two-Stage Training Pipeline. We adopt a two-stage training method for the ICT model and HP model. In the first stage, we fine-tune the CLIP model using our constructed ICT labels as training targets. In the second stage, we freeze the ICT model parameters and continue training the HP model. During the model optimization phase, we comparatively evaluate + +Table 1. Preference prediction accuracy on test sets of Pick-High and Pickapic_v2. Bold and underlined values represent optimal and second-best performance. + +
ModelJPEG Compressibility↑Aesthetic Score↑
SD3.5-Large [1]374.806.307
FLUX.1-dev [14]270.586.436
SD3.5-Turbo [2]313.106.293
+ CLIP [26](crash)279.985.175
+ ICT (Ours)321.396.285
+ ICT-HP (Ours)330.236.300
+ HP (Ours)334.866.448
+ +Table 2. JPEG Compressibility and Aesthetic Scores on Parti-Prompts. Bold and underlined values represent optimal and second-best performance. + +three different reward models for optimizing the diffusion model LoRA adapters: ICT model, HP model, and the more comprehensive ICT-HP model. + +Training Datasets. Our training dataset for the ICT model consists of the Pick-High dataset and a subset of Pickapic_v2, comprising a total of 360,000 image triplets. For the diffusion model optimization process, we extract 32,000 non-repetitive text prompts from the Pick-High dataset to serve as our prompt dataset. + +For detailed training hyperparameter configurations and experimental specifications, please refer to the supplementary materials. + +# 5.2. Reward Model Performance + +We evaluated the prediction accuracy of human preference models on the Pick-High and Pickapic_v2 test sets. As shown in Table 1, the image-text contrastive (ITC) classification heads of foundation multimodal pre-trained models CLIP and BLIP perform only comparable to random selection. This result indicates that the instance-level image-text alignment objective adopted by foundation multimodal models is not suitable for evaluating the quality of generative models. In terms of overall average accuracy, our proposed ICT-HP model achieves nearly $10\%$ performance improvement compared to Pickscore. This significant improvement stems from the ability of our reward model to measure the extent to which images contain text, and on this basis, continue to measure image quality, thus more accurately identifying high-quality images and assigning them correspondingly high scores. + +Qualitative Results on SD3.5-Turbo Optimization. In Figure 4, we present generation results using real user prompts. All images are generated with the same initial noise to ensure fair comparison. The experimental results show that: the HPS_v2 model only enhances image tone with minimal detail improvement; the PickScore model, while increasing details, causes distortions in style and color fidelity; the ImageReward model shows almost no substantial improvement over the original model outputs. In contrast, our proposed reward models achieve performance improvements in + +![](images/fb789d500c33a463c66876b371e3c225b69d9ff18327c384cfe2b6bd517904ba.jpg) +Figure 4. Qualitative Comparison of Optimization Results Across Reward Models Using Real User Prompts. + +both textural details and tonal aesthetics. Specifically, the ICT model does not penalize increases in image information content, allowing diffusion models to continue optimizing toward richer visual content after achieving text-image alignment; meanwhile, the HP model effectively utilizes pure image features to further optimize when the ICT model reaches its performance ceiling. Compared to our approach, existing reward models tend to produce negative feedback on high-quality image features, thereby limiting their ability to enhance the performance of advanced diffusion models. + +Quantitative Evaluation of Visual Information Density and Aesthetic Fidelity. To quantitatively evaluate the enhancement effect of our reward model on the original model in terms of generated image information content and aesthetic quality, we employed two objective metrics, JPEG Compressibility [12] and Aesthetic Score, for assessment on the Parti-Prompts dataset [43]. As shown in Table 2, among all tested models, the HP model achieved the most significant improvements in both metrics, followed by the ICT-HP model. Notably, the SD3.5-Turbo optimized by the HP model not only demonstrated substantial improvement in aesthetic metrics, but even slightly surpassed more comprehensive baseline models such as SD3.5-Large and FLUX.1-dev. + +Qualitative Comparison of Simple Elements Generation Results. We conducted qualitative ex + +periments on the generation of simple elements, with results shown in Figure 5. The ICT model can stably handle simple geometric shapes and effectively penalize complex elements that do not meet simplicity requirements when prompts include descriptors like "simple". + +Comprehensive Evaluation on GenEval Benchmark. Table 3 presents the quantitative performance of various image generation models within the multidimensional evaluation framework of GenEval [8]. Using SD3.5-Turbo as the baseline, our ICT-HP and ICT models both demonstrate significant advantages. The ICT-HP model excels in multi-object generation and color fidelity; while the ICT model achieves the highest mean score among all models and delivers optimal performance on the Counting metric, showcasing its superior capability in precise object quantity control. In contrast, the CLIP model suffers comprehensive performance degradation due to training collapse, and PickScore exhibits weaker performance in the Colors metric due to limitations in tonal representation. These comparisons validate our methods' significant advantages in terms of text-image consistency. + +Quantitative and Qualitative Performance Comparison of Multi-metric Reward Models. We conducted quantitative analyses of optimization results across various reward models, as shown in Table 4. Experiments reveal that existing reward models including HPS, PickScore, and ImageReward ex + +
ModelMean↑Single↑Two↑Counting↑Colors↑Position↑Color Attribution↑
SDXL [22]0.550.980.740.390.850.150.23
DALL-E 2 [28]0.520.940.660.490.770.100.19
DALL-E 3 [21]0.670.960.870.470.830.430.45
SD3 [7]0.680.980.840.660.740.400.43
SD3.5-Turbo [2]0.690.990.870.690.800.250.55
+ ImageReward0.700.990.870.680.800.280.59
+ HPS_v20.690.980.860.680.810.270.57
+ PickScore0.660.990.850.670.740.240.48
+ ICT-HP (Ours)0.700.990.880.680.790.280.58
+ CLIP(crash)0.130.380.040.060.260.010.03
+ ICT (Ours)0.710.980.880.700.810.310.56
+ +Table 3. Quantitative Evaluation of Optimization Results Across Reward Models on GenEval Benchmark. Bold and underlined values represent optimal and second-best performance. + +![](images/6e0eb177eca2e6cb4a406f468ad4f5cdf2adc35e1a02a9aab8862e8dc6a3cd32.jpg) +Figure 5. Qualitative Comparison of Simple Elements Generation Results. + +hibit performance degradation across multiple metrics. In contrast, our ICT-HP and HP models achieved improvements across all metrics, demonstrating the stability advantage of our method. Although PickScore obtained the highest aesthetic score, qualitative analysis and GenEval experiments indicated significant distortion in artistic style and tone, with color-related metrics significantly lower than other models. Figure 6 compares the original image, refined image, and our model-optimized results. ICT-HP and HP models not only generate intricate backgrounds that are difficult to describe directly through language but also demonstrate a more precise and comprehensive understanding of prompts, thus producing higher-quality images. + +Feature Transfer Validation. As illustrated in Fig- + +![](images/f3181dc1e47d569e7cdcfcd197df7d8eac66a5a6424610e137315e570ffe4fc7.jpg) +Figure 6. Qualitative Comparison: Original Images, Refined Images, and Generations Optimized by Our Reward Models (HP/ICT-HP). + +ure 7, transplanting the Text-Encoder trained by our ICT model into the SD2.1-base model significantly improves image quality. This demonstrates that our encoder integrates fine-grained visual details, thus transforming textual descriptions into more precise visual guidance signals. Figure 8 presents results of style transfer with InstantStyle [35]. The CLIP image encoder is replaced by our ICT image encoder. Our fine-tuned encoder not only precisely injects reference styles but also balances text descriptions, overcoming the style-content inconsistency. These experiments validate the effectiveness and advancement of ICT score training objectives. + +
ModelCLIP↑ITM↑ImgRwd↑HPS↑Pick↑Aes↑ICT↑HP↑ICT-HP↑
SD3.5-Te0.3400.8310.9550.27722.0346.5550.9100.7770.717
SD3.5-Tr0.326↓0.757↓0.914↓0.276↓22.012↓6.780↑0.923↑0.781↑0.718↑
+HPS_v20.343↑0.835↑1.065↑0.282↑22.000↓6.590↑0.929↑0.777-0.722↑
+PickScore0.332↓0.816↓1.106↑0.277-22.719↑6.924↑0.929↑0.779↑0.723↑
+ImageReward0.340-0.841↑1.056↑0.279↑22.022↓6.569↑0.865↓0.763↓0.663↓
+ICT(Ours)0.344↑0.841↑1.014↑0.277-22.117↑6.561↑0.937↑0.776↓0.728↑
+ICT-HP(Ours)0.345↑0.839↑1.011↑0.277-22.147↑6.558↑0.938↑0.778↑0.730↑
+HP(Ours)0.342↑0.836↑1.069↑0.278↑22.231↑6.668↑0.937↑0.779↑0.730↑
+ +![](images/3b97bb62704a1e0aaf875b026c5007287de72600af6aded4e4ee031fa36d014a.jpg) +Figure 7. Qualitative Comparison of ICT Text-Encoder Direct Migration to SD2.1-Base Model. + +Table 4. Quantitative Results. SD3.5-Te/Tr denote images generated using easy/reflect prompts on SD-3.5-Turbo. Arrows (↑, ↓, -) indicate changes vs. SD3.5-Te. Bold and underlined values represent optimal and second-best performance. + +
ComparisonModelsWin Rate
vs. BaseICT model0.562
HP model0.709
ICT-HP model0.673
vs. PickScoreHP model0.617
ICT-HP model0.714
+ +Table 5. Human evaluation on 300 randomly selected samples from Parti-Prompts [43]. + +# 5.3. User-Study + +We conduct a user study with ten annotators on 300 randomly selected prompts from Parti-Prompts [43]. Image pairs (ours vs. baseline) were randomly ordered, and annotators evaluated prompt expression and visual appeal. As shown in Table 5, our methods achieve higher win rates against both the SD3.5-large-Turbo base model and the PickScore-optimized model, confirming our approach's effectiveness. + +![](images/0dbadcdcc17bba6b5739de4ef09d8d300860308a3c11b4e6c10078c5777ef3cd.jpg) +Figure 8. Qualitative Comparison of HP/ICT ImageEncoder Performance in Style Injection Tasks. + +# 6. Conclusion + +In this paper, we reveal the reward-scoring paradox of high-quality images and show that the inherent flaw in current text-to-image generation reward models lies in classic text-image alignment paradigms. We propose a novel training objective for reward models, i.e., ICT (Image-Contained-Text) score. This objective encourages models to achieve text-image alignment and avoid penalizing high-quality images containing extra details, that better match human preferences. Based on the ICT model, we further train a preference model that relies solely on image modality for human preference assessment. Our experiments show that the proposed approach improves scoring accuracy by over $10\%$ compared to existing methods and markedly enhances state-of-the-art models like SD3.5-Turbo. Through both theoretical foundation and empirical validation, our work advances image generation technology toward better alignment with human preferences. + +# Acknowledgments + +This work was supported in part by the National Natural Science Foundation of China No. 62376277, the Public Computing Cloud of Renmin University of China and the Fund for Building World-Class Universities (Disciplines) of Renmin University of China. + +# References + +[1] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.2, 5 +[2] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 5, 7 +[3] Yalong Bai, Mohan Zhou, and Qing Yang. Styleinject: Parameter efficient tuning of text-to-image diffusion models. ArXiv, abs/2401.13942, 2024. 1 +[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 2818-2829. IEEE, 2023. 11 +[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards, 2024. 5, 11, 13 +[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022. 11 +[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2, 7 +[8] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. CoRR, abs/2310.11513, 2023. 6 +[9] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1 +[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 2 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1 +[12] ISO/IEC. Information technology - digital compression and coding of continuous-tone still images - requirements and guidelines. International Organization for Standardization, 1992. ISO/IEC 10918-1:1992. 6 + +[13] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. 2, 5 +[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 5 +[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 1, 2, 5 +[16] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 11, 13 +[17] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Biye Li, Huaian Chen, and Yi Jin. Star: Scale-wise text-conditioned autoregressive image generation. arXiv preprint arXiv:2406.10797, 2024. 1 +[18] Midjourney. Ai-powered image generation. https://www.midjourney.com/, 2022. 2 +[19] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, JiRong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26617-26626, 2024. 1 +[20] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, and Ji-Rong Wen. Uniform attention maps: Boosting image fidelity in reconstruction and editing. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pages 4420-4429. IEEE, 2025. 1 +[21] OpenAI. Dall-e 3. https://openai.com/dall-e-3, 2023.7 +[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 7 +[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, 2024. 1 +[24] Wenwen Qiang, Jingyao Wang, Zeen Song, Jiangmeng Li, and Changwen Zheng. On the out-of-distribution generalization of self-supervised learning. arXiv preprint arXiv:2505.16675, 2025. 11 +[25] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1 +[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on ma + +chine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5 +[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2 +[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 7 +[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2 +[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2022. 1 +[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 2 +[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1 +[33] Zeen Song, Siyu Zhao, Xingyu Zhang, Jiangmeng Li, Changwen Zheng, and Wenwen Qiang. Learning invariant causal mechanism from vision-language models. arXiv preprint arXiv:2405.15289, 2024. 11 +[34] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 1 +[35] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 7 +[36] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models, 2023. 11 +[37] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 1 +[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 2, 5 + +[39] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 1 +[40] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 2, 5 +[41] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1316-1324. Computer Vision Foundation / IEEE Computer Society, 2018. 2 +[42] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 3954-3964. ACM, 2023. 1 +[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6, 8 +[44] Guiwei Zhang, Tianyu Zhang, Mohan Zhou, Yalong Bai, and Biye Li. V2flow: Unifying visual tokenization and large language model vocabularies for autoregressive image generation. arXiv preprint arXiv:2503.07493, 2025. 1 +[45] Yanan Zhang, Jiangmeng Li, Lixiang Liu, and Wenwen Qiang. Rethinking misalignment in vision-language model adaptation from a causal perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 11 +[46] Ziyi Zhang, Li Shen, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Bo Du, and Dacheng Tao. Aligning few-step diffusion models with dense reward difference learning. arXiv preprint arXiv:2411.11727, 2024. 1 +[47] Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Llada 1.5: Variance-reduced preference optimization for large language diffusion models, 2025. 1 +[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. DM-GAN: dynamic memory generative adversarial networks for text-to-image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5802-5810. Computer Vision Foundation / IEEE, 2019. 2 \ No newline at end of file diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/images.zip b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f4dfda8e703d70faffab66ed153408f55839a805 --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe948053fc8ccb361d4e71079d86b6573e68bc8bb0cccf3db7ee149320d60aad +size 922683 diff --git a/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/layout.json b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..91fe3a4b4561ac3fe1388e4b542bee6a61ab8392 --- /dev/null +++ b/2025/Enhancing Reward Models for High-quality Image Generation_ Beyond Text-Image Alignment/layout.json @@ -0,0 +1,8015 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 102, + 529, + 140 + ], + "type": "text", + "content": "Enhancing Reward Models for High-quality Image Generation: Beyond Text-Image Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": "Ying Ba" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Tianyu Zhang" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Yalong Bai" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Wenyi Mo" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Tao Liang" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Bing Su" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": ", Ji-Rong Wen" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": "Gaoling School of Artificial Intelligence Renmin University of China Beijing, China \n" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": "Beijing Key Laboratory of Research on Large Models and Intelligent Governance \n" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": "Engineering Research Center of Next-Generation Intelligent Search and Recommendation, MOE " + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 62, + 165, + 576, + 236 + ], + "type": "text", + "content": "iN2X" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "spans": [ + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + }, + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "type": "text", + "content": "yingba88@outlook.com, {mowenyi00, subingats}@gmail.com, jrwen@ruc.edu.cn \n" + }, + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 238, + 499, + 265 + ], + "type": "text", + "content": " {tianyu1949, liang0305tao}@gmail.com, ylbai@outlook.com" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 149, + 293, + 202, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 293, + 202, + 304 + ], + "spans": [ + { + "bbox": [ + 149, + 293, + 202, + 304 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 317, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 317, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 317, + 297, + 628 + ], + "type": "text", + "content": "Contemporary image generation systems have achieved high fidelity and superior aesthetic quality beyond basic text-image alignment. However, existing evaluation frameworks have failed to evolve in parallel. This study reveals that human preference reward models fine-tuned based on CLIP and BLIP architectures have inherent flaws: they inappropriately assign low scores to images with rich details and high aesthetic value, creating a significant discrepancy with actual human aesthetic preferences. To address this issue, we design a novel evaluation score, ICT (Image-Contained-Text) score, that achieves and surpasses the objectives of text-image alignment by assessing the degree to which images represent textual content. Building upon this foundation, we further train an HP (High-Preference) score model using solely the image modality to enhance image aesthetics and detail quality while maintaining text-image alignment. Experiments demonstrate that the proposed evaluation model improves scoring accuracy by over " + }, + { + "bbox": [ + 55, + 317, + 297, + 628 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 317, + 297, + 628 + ], + "type": "text", + "content": " compared to existing methods, and achieves significant results in optimizing state-of-the-art text-to-image models. This research provides theoretical and empirical support for evolving image generation technology toward higher-order human aesthetic preferences. Code is available at https://github.com/BarretBa/ICTHP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 631, + 148, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 631, + 148, + 643 + ], + "spans": [ + { + "bbox": [ + 56, + 631, + 148, + 643 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 650, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 296, + 687 + ], + "type": "text", + "content": "In recent years, generation models have witnessed rapid development and found widespread applications across diverse domains [11, 17, 23, 29, 32, 42, 44, 47]. Modern" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 292, + 432, + 371 + ], + "blocks": [ + { + "bbox": [ + 316, + 292, + 432, + 371 + ], + "lines": [ + { + "bbox": [ + 316, + 292, + 432, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 432, + 371 + ], + "type": "image", + "image_path": "a219871a8298e64b70e94219a657bcf91893f5662a7ca8cfbc77799f7facf632.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 432, + 293, + 553, + 372 + ], + "blocks": [ + { + "bbox": [ + 432, + 293, + 553, + 372 + ], + "lines": [ + { + "bbox": [ + 432, + 293, + 553, + 372 + ], + "spans": [ + { + "bbox": [ + 432, + 293, + 553, + 372 + ], + "type": "image", + "image_path": "1f1812da3ac72273ccac1fb5cfbeed2215436aec9baf9176f4dcb63209992b88.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 372, + 432, + 449 + ], + "blocks": [ + { + "bbox": [ + 316, + 372, + 432, + 449 + ], + "lines": [ + { + "bbox": [ + 316, + 372, + 432, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 432, + 449 + ], + "type": "image", + "image_path": "3decd7529e2444420bc6235cf1c582a1d9652f72c3ee1b9ef8f9086daddf28d0.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 460, + 555, + 505 + ], + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 505 + ], + "type": "text", + "content": "Figure 1. Reward Model Scoring Paradox Across Multiple Models Based on CLIP [26] (CLIP Score, PickScore) and BLIP [15] (ITM Score, ImageReward). Refine images are generated with refined prompts by LLMs." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 432, + 372, + 551, + 449 + ], + "blocks": [ + { + "bbox": [ + 432, + 372, + 551, + 449 + ], + "lines": [ + { + "bbox": [ + 432, + 372, + 551, + 449 + ], + "spans": [ + { + "bbox": [ + 432, + 372, + 551, + 449 + ], + "type": "image", + "image_path": "f8b7a98ace20886f9dc4a0cb7e86f5c9ddd00f151b9f42586f9226e98502e6c3.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 312, + 509, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 555, + 676 + ], + "type": "text", + "content": "high-performance diffusion models can not only accurately align texts and images but also generate visually impressive results. As the technology matures, the optimization focus of diffusion models has shifted from text-image alignment and prompt fidelity to generating high-quality images that better meet human aesthetic expectations [9, 19, 20, 37, 39]. In terms of optimization strategies, traditional data-driven training methods [3, 25, 30] are gradually being supplemented and expanded by reward model-based gradient backpropagation and reinforcement learning algorithms [34, 46]. Therefore, constructing a reward function that can precisely evaluate image quality is crucial for further improving the performance of diffusion models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "For text-to-image evaluation, text-image alignment (i.e., instruction compliance) serves as the fundamental requirement, while human preference determines the" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 182, + 702 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 182, + 702 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 182, + 702 + ], + "type": "text", + "content": "*This work was done at iN2X." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 157, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19022" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "upper limit of image quality. Current reward models exhibit a critical flaw in measuring text-image alignment: they presume potential equivalence between the information content of images and text, whereas in reality, textual descriptions can hardly comprehensively capture all visual information in images. A direct manifestation is that for the same prompt, a basic image merely covering textual descriptions typically receives higher scores than a higher-quality image that contains richer details and visual elements extending beyond the literal description of the prompt. This phenomenon persists not only in foundational models [15, 26] but also affects fine-tuned variants with human preferences like ImageReward [40] and PickScore [13] (as shown in Fig. 1). Thus, previous reward models may effectively enhance Stable Diffusion-1.5/2.0 [29] (where text-image alignment remains imperfect) but degrade state-of-the-art large models like Stable Diffusion-3.5 [1], steering them to generate visually sparse and aesthetically deficient images." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 295, + 517 + ], + "type": "text", + "content": "To address the above issue and continue leveraging human preference data to train superior reward models, we propose a novel optimization objective that transcends text-image alignment, i.e., the image-contained-text score (ICT score). ICT score quantifies the extent to which an image contains textual information to evaluate the quality of generated images. By focusing on how well the image encapsulates textual semantics rather than enforcing bidirectional equivalence, ICT score prevents penalizing images that naturally contain richer contextual details beyond literal prompt descriptions, which aligns with human perception of high-quality outputs. The ICT score removes the ceiling constraints of human preference models, enabling us to incorporate ultra-high-quality preference data to train an advanced human preference model, achieving a more comprehensive evaluation system." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 519, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 544 + ], + "type": "text", + "content": "The main contributions of this research can be summarized in the following three points:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 546, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 57, + 546, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 546, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 57, + 546, + 295, + 629 + ], + "type": "text", + "content": "- We propose a novel image generation evaluation framework, including ICT score (quantifying the extent to which an image contains textual information) and HP score (evaluating how well an image meets human aesthetic preferences), fundamentally solving the problem of existing reward models underscoring high-quality images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 630, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 630, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 57, + 630, + 295, + 677 + ], + "type": "text", + "content": "- A large-scale high-quality dataset (Pick-High) is constructed using chain-of-thought reasoning from large language models, providing a comprehensive benchmark for image generation quality assessment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 678, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 678, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 678, + 295, + 713 + ], + "type": "text", + "content": "- We successfully apply the proposed reward model to optimize a state-of-the-art diffusion model (Stable Diffusion-3.5-turbo), demonstrating the effic" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 326, + 72, + 553, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 72, + 553, + 96 + ], + "spans": [ + { + "bbox": [ + 326, + 72, + 553, + 96 + ], + "type": "text", + "content": "tiveness and superiority of this method in practical generation tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 108, + 415, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 108, + 415, + 121 + ], + "spans": [ + { + "bbox": [ + 314, + 108, + 415, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 130, + 553, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 130, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 130, + 553, + 416 + ], + "type": "text", + "content": "Image quality evaluation for generative models. The evaluation of generated images primarily falls into two categories: human assessment and automated metrics. While the former serves as the gold standard, its high cost and subjective biases have led to the widespread application of the latter in large-scale evaluations. Image-generation models have seen significant quality improvements, evolving from GAN-based approaches such as AttnGAN [41] and DM-GAN [48] to diffusion-based models including DALL·E 2 [27], Stable Diffusion [7, 29], Midjourney [18], and FLUX [14]. Traditional metrics like FID [10] and IS [31] focus solely on distributional statistical properties and fail to capture human subjective preferences. Addressing this limitation, researchers have developed specialized human preference reward models trained on human preference data, such as PickScore [13], HPSv2 [38], and ImageReward [40]. However, we have observed that these reward models exhibit significant biases when evaluating high-quality images generated through prompt optimization, indicating they have not yet effectively guided diffusion models in generating images that align with human preferences for detail and content richness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 418, + 553, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 553, + 609 + ], + "type": "text", + "content": "Image quality evaluation datasets. Addressing the limitations of existing human preference models, high-quality evaluation datasets are essential for generative model assessment. As generative models rapidly evolved, researchers shifted toward constructing datasets based on human preferences, including the PickaPic_v2 dataset [13], ImageRewardDB [40] developed by the ImageReward team, and the HPDv2 dataset [38] employed by HPS. Despite their scale, these datasets predominantly contain lower-quality images inadequate for evaluating high-quality outputs from modern diffusion models and lack diversity across quality levels. To address these issues, we constructed a 600GB high-quality dataset called Pick-High, incorporating portions of the Pick-a-Pic dataset during training to ensure a comprehensive quality distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 620, + 550, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 620, + 550, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 620, + 550, + 635 + ], + "type": "text", + "content": "3. Limitations of Text-image Alignment" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 553, + 713 + ], + "type": "text", + "content": "Multimodal models such as CLIP employ discriminative contrastive loss during pre-training to achieve text-image alignment. According to the information bottleneck theory, their encoders extract minimal image discriminative information to maximize mutual information with text. This training mechanism causes CLIP" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19023" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 294, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 294, + 149 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 294, + 149 + ], + "type": "image", + "image_path": "55f402f6f98c871a998306ef094d3a6ffdc267bcc5776b1cc4ce4e56da716d89.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 159, + 295, + 182 + ], + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 182 + ], + "type": "text", + "content": "Figure 2. Text-Image Feature Interaction Venn Diagram: Increasing Image Information Under Fixed Text Prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 194, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 254 + ], + "type": "text", + "content": "to exhibit bias when evaluating the same prompt: it tends to assign higher scores to simple images that merely faithfully reflect textual content, while giving lower scores to high-quality images rich in details, contradicting human aesthetic preferences." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 290 + ], + "type": "text", + "content": "According to the principle of information decomposition, the information content of a generated image can be divided into:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 125, + 300, + 294, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 125, + 300, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 125, + 300, + 294, + 314 + ], + "type": "interline_equation", + "content": "I (v) = I (v; t) + I (v | t), \\tag {1}", + "image_path": "180fd294af97cfa3c2028697bd8dac407517c97deeef7a60e8938375e1530024.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "I(v; t)" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " represents the mutual information between the image " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " (the image-text alignment component), and " + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "inline_equation", + "content": "I(v|t)" + }, + { + "bbox": [ + 55, + 323, + 296, + 382 + ], + "type": "text", + "content": " represents the additional visual information contained in the image (aesthetics, texture, atmosphere, etc.)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": "The CLIP scoring mechanism, based on cosine similarity, can be expressed as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 415, + 296, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 296, + 445 + ], + "type": "interline_equation", + "content": "\\mathrm {C L I P} (v, t) \\approx \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot I (v)}} = \\frac {I (v ; t)}{\\sqrt {I (t) \\cdot (I (v ; t) + I (v | t))}}.", + "image_path": "46c6b752536aec6b302c44ba732cede4752c407f5d8af3b7b78b59d7a38f3ed5.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "spans": [ + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": "As shown in Figure 2, as the capability of generative models improves, the information content of images increases. In the transition from image-text misalignment to alignment, " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "I(v; t)" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": " increases faster than " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "I(v)" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": ", gradually raising CLIP scores. However, when high-quality models generate images rich in visual details, although " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "I(v; t)" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": " increases, " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "I(v|t)" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": " grows even faster, causing the denominator " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "\\sqrt{I(v)}" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": " to increase more than the numerator " + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "inline_equation", + "content": "I(v; t)" + }, + { + "bbox": [ + 55, + 451, + 296, + 606 + ], + "type": "text", + "content": ", resulting in an overall decrease in scores, creating a contradiction with human preferences. This motivates us to design a novel objective to train foundation models from the perspective of reward models, which will be elaborated in the next section." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 617, + 121, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 121, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 121, + 629 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 713 + ], + "type": "text", + "content": "This section presents our proposed methodological framework. In Section 4.1, we introduce Pick-High dataset, a large-scale high-quality dataset constructed using refined prompts, which addresses the evaluation bias in existing reward models when assessing information-rich images. In Section 4.2, to resolve the inconsistency between the image-text alignment" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 72, + 552, + 188 + ], + "blocks": [ + { + "bbox": [ + 319, + 72, + 552, + 188 + ], + "lines": [ + { + "bbox": [ + 319, + 72, + 552, + 188 + ], + "spans": [ + { + "bbox": [ + 319, + 72, + 552, + 188 + ], + "type": "image", + "image_path": "53796012433464b4102cd66c8d504ce0bd93fa49da5780e7393321998fc2b7ac.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 553, + 218 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 218 + ], + "type": "text", + "content": "Figure 3. The Image-Contained-Text (ICT) Scoring Framework." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 220, + 555, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 340 + ], + "type": "text", + "content": "objective of foundation models and evaluation goals of reward models, we propose the Image-Contained-Text (ICT) optimization objective. This enhancement adapts the CLIP model for better evaluation of generative models. We further develop an innovative HP model that relies solely on image modality for human preference assessment. Finally, in Section 4.3, we detail how we utilize the improved reward model to optimize diffusion models, thus enhancing the overall quality of generated images." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 343, + 423, + 356 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 423, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 423, + 356 + ], + "type": "text", + "content": "4.1. Pick-High Dataset" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 358, + 555, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 537 + ], + "type": "text", + "content": "Existing reward models (including PickScore, HPSv2, and ImgReward) inherit the limitations of foundational multimodal models, i.e., assigning unreasonably low scores when evaluating complex images paired with simple prompts, which significantly deviates from human preferences. To address this deficiency, we curated 360,000 prompts from the PickAPic_v2 dataset and leveraged large language models' chain-of-thought capabilities to meticulously design a more refined set of prompts that closely align with human preferences. We input these optimized prompts into state-of-the-art image generation models to create 360,000 high-quality images, forming the Pick-High dataset. The detailed construction processes of Pick-High dataset can be seen in Appendix A." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": "PickAPic_v2 contains pairwise preferences of images. We denote the non-preferred low-quality image as " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": " and the preferred image as " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": ". The Pick-High dataset extends pairwise preferences to triplet rankings by introducing a high-quality image " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "I_{3}" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": " generated with refined prompts. Based on these quality levels, we establish a hierarchical scoring mechanism for both basic prompts " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{easy}}" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": " and refined prompts " + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ref}}" + }, + { + "bbox": [ + 313, + 539, + 555, + 647 + ], + "type": "text", + "content": ", to better align human preferences." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 655, + 519, + 669 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 655, + 519, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 655, + 519, + 669 + ], + "type": "text", + "content": "4.2. Reward Model Optimization Objective" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 673, + 422, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 422, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 422, + 685 + ], + "type": "text", + "content": "4.2.1. ICT Scores Design" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "In text-to-image model evaluation, existing foundation models (e.g., CLIP, BLIP) employ instance-level" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19024" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 54, + 72, + 294, + 167 + ], + "type": "text", + "content": "image-text contrastive loss, which presents a fundamental issue: they tend to penalize high-quality images containing additional details. This evaluation mechanism significantly deviates from human preferences and reward model objectives. Based on these observations, we propose the Image-Contained-Text (ICT) scoring mechanism to quantify the extent to which an image expresses textual content." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "content": "The overview of our ICT scoring mechanism is shown in Fig. 3. We assign different scores to images " + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "inline_equation", + "content": "(I_1, I_2, I_3)" + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "content": " based on their alignment quality with prompts. For high-quality images " + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "inline_equation", + "content": "I_3" + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "content": ", we assign perfect scores for both basic prompts and refined prompts. To calculate ICT scores of " + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 54, + 167, + 294, + 263 + ], + "type": "text", + "content": " with basic prompts, we first develop a threshold-based mechanism instead of directly adopting CLIP similarity function:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 109, + 268, + 294, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 268, + 294, + 291 + ], + "spans": [ + { + "bbox": [ + 109, + 268, + 294, + 291 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (I, P) = \\min (\\frac {\\operatorname {C L I P} (I , P)}{\\theta}, 1), \\tag {2}", + "image_path": "4f046d8a9a61b1fa1fdf2fb3b9d325d87d165e735e59acf0039241ab724c5ef8.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}(\\mathrm{I},\\mathrm{P})" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " denotes the CLIP similarity function of an image " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " and its corresponding prompt " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " represents the alignment threshold. This approach leverages the effectiveness of the CLIP model for assessment of low- and medium-quality images while circumventing the bias in CLIP scores of high-quality images. The basic prompt ICT scores " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "(E_1,E_2,E_3)" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "I_{3}" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{easy}}" + }, + { + "bbox": [ + 55, + 295, + 295, + 392 + ], + "type": "text", + "content": " are defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 116, + 398, + 294, + 409 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 398, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 116, + 398, + 294, + 409 + ], + "type": "interline_equation", + "content": "E _ {3} = 1, \\tag {3}", + "image_path": "959b700d441d55c4e20d34707c01209448bf1fea6cafed505cf9d474e0a276c2.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 412, + 294, + 425 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 412, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 116, + 412, + 294, + 425 + ], + "type": "interline_equation", + "content": "E _ {2} = \\mathcal {C} \\left(I _ {2}, P _ {\\text {e a s y}}\\right), \\tag {4}", + "image_path": "33505d861bbdab99a5436437fd29e415dd1e267cb9051a15b8bc54f7a017540c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 428, + 294, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 428, + 294, + 440 + ], + "spans": [ + { + "bbox": [ + 116, + 428, + 294, + 440 + ], + "type": "interline_equation", + "content": "E _ {1} = \\min \\left(\\mathcal {C} \\left(I _ {1}, P _ {\\text {e a s y}}\\right), E _ {2}\\right). \\tag {5}", + "image_path": "ad929b742cd25d25a1f0ead1319c2e62a4b9cd945ad16f8bc514bd84fcf128f8.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": "As for refined prompt ICT scores, based on basic prompt ICT scores, we further consider the CLIP similarities of " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{easy}}" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": ", defining the scores of " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "I_2" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "I_1" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "P_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": " as the product of " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "E_2" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "inline_equation", + "content": "E_1" + }, + { + "bbox": [ + 55, + 445, + 295, + 505 + ], + "type": "text", + "content": " and the textual similarity, as shown below:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 113, + 522, + 294, + 534 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 522, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 113, + 522, + 294, + 534 + ], + "type": "interline_equation", + "content": "R _ {3} = 1, \\tag {6}", + "image_path": "99d9421d9c439c222bb0d46a580b3a892eb168504dffeb7ce856b41865d49ddd.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 113, + 537, + 294, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 537, + 294, + 550 + ], + "spans": [ + { + "bbox": [ + 113, + 537, + 294, + 550 + ], + "type": "interline_equation", + "content": "R _ {2} = E _ {2} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right), \\tag {7}", + "image_path": "d7f4de74b6ba11d6fec26c8209e5e94df7942b3ac02045da11de4fd2e038e7cb.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 113, + 552, + 294, + 565 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 552, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 113, + 552, + 294, + 565 + ], + "type": "interline_equation", + "content": "R _ {1} = E _ {1} \\times \\operatorname {C L I P} \\left(P _ {\\text {e a s y}}, P _ {\\text {r e f}}\\right). \\tag {8}", + "image_path": "9ba3f5eeaad365d675aa01ba28b355f68a3b447585a0e1581cfe89754015d884.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Unlike previous discriminative training objectives focused on achieving binary yes/no text-image alignment, our proposed ICT scores categorize the alignment quality between images and texts into distinct levels. This novel objective not only encourages models to achieve image-text alignment but also avoids penalizing images containing additional details beyond textual descriptions. It represents a superior training objective that goes beyond conventional text-image alignment metrics. In the following sections, we will elaborate on implementing these ICT scores as innovative labeling mechanisms for training multimodal models." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 390, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 390, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 390, + 83 + ], + "type": "text", + "content": "4.2.2. ICT Model" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "text", + "content": "We finetune the CLIP model on Pick-High and Pickapic_v2 datasets to obtain our ICT model. Each training sample consists of an image triplet with corresponding ICT labels " + }, + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "(E_1, E_2, E_3)" + }, + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "text", + "content": " for the basic prompt and " + }, + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "inline_equation", + "content": "(R_1, R_2, R_3)" + }, + { + "bbox": [ + 313, + 87, + 553, + 147 + ], + "type": "text", + "content": " for the refined prompt." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "spans": [ + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": "The model computes ICT scores between each image and both prompts: " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "y_{i,e} = s(P_{\\mathrm{easy}},I_i)" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "y_{i,r} = s(P_{\\mathrm{ref}},I_i)" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "i \\in \\{1,2,3\\}" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": ", and the scoring function " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": " is defined as " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "s(P,I) = \\mathrm{Enc}_{\\mathrm{xt}}(P) \\cdot \\mathrm{Enc}_{\\mathrm{img}}(I) / \\tau" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 147, + 554, + 231 + ], + "type": "text", + "content": " is the temperature parameter. We optimize the model by minimizing the mean squared error between predicted scores and ICT scores:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 346, + 249, + 553, + 281 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 249, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 346, + 249, + 553, + 281 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {I C T}} = \\sum_ {i = 1} ^ {3} (E _ {i} - y _ {i, e}) ^ {2} + \\sum_ {i = 1} ^ {3} (R _ {i} - y _ {i, r}) ^ {2}. \\tag {9}", + "image_path": "c7cc65d74c666a710a283622005e676699b5b18eb81ad0a2177a796f5d69e887.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "text", + "content": "To enhance the discriminative capability of the model, we introduce an in-batch negative sample learning strategy. For any combination of images and unpaired prompts within a batch, we set their ICT labels to 0 and compute the negative sample loss " + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{neg}} = \\sum_{j\\neq i}[(y_{j,e}^i)^2 +(y_{j,r}^i)^2 ]" + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "inline_equation", + "content": "y_{j,e}^{i}" + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "inline_equation", + "content": "y_{j,r}^{i}" + }, + { + "bbox": [ + 313, + 288, + 555, + 433 + ], + "type": "text", + "content": " represent the ICT scores between the image of the i-th sample and the basic/Refined prompts of the j-th sample, respectively. Considering the potential existence of false negative samples within batches, we adopt a smooth weighting strategy based on the sigmoid function for negative sample mining:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 441, + 553, + 464 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 441, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 388, + 441, + 553, + 464 + ], + "type": "interline_equation", + "content": "w (y) = \\frac {1}{1 + e ^ {\\alpha (| y | - \\beta)}}, \\tag {10}", + "image_path": "64ed70d4ff0642d8dc3da76b11e1eaa06812d229d0078cc6c2448de1174d0eda.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "text", + "content": " is the smoothing coefficient controlling the steepness of the weight curve, " + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "text", + "content": " is the threshold parameter, and " + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 467, + 554, + 551 + ], + "type": "text", + "content": " is the predicted score of the sample. This function assigns lower weights to potential false negative samples with high predicted scores, thereby mitigating their negative impact on model training. The final negative sample loss function is:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 334, + 569, + 553, + 596 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 334, + 569, + 553, + 596 + ], + "spans": [ + { + "bbox": [ + 334, + 569, + 553, + 596 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {n e g}} = \\sum_ {j \\neq i} \\left(w \\left(y _ {j, e} ^ {i}\\right) \\cdot y _ {j, e} ^ {i} ^ {2} + w \\left(y _ {j, r} ^ {i}\\right) \\cdot y _ {j, r} ^ {i} ^ {2}\\right). \\tag {11}", + "image_path": "067c29c642fe63178bfb32720f1302e145e3e308f66ad98a189f750082469d67.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 326, + 603, + 522, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 603, + 522, + 615 + ], + "spans": [ + { + "bbox": [ + 326, + 603, + 522, + 615 + ], + "type": "text", + "content": "The complete loss function is formulated as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "spans": [ + { + "bbox": [ + 384, + 624, + 553, + 637 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\mathrm {I C T}} + \\lambda \\mathcal {L} _ {\\text {n e g}}, \\tag {12}", + "image_path": "7d9662bf93cfda3e37821385f17b9d64e2c455a196c11f0b2eab3aa321b0ebde.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 645, + 553, + 669 + ], + "type": "text", + "content": " is a balancing factor that controls the contribution of negative sample learning." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 674, + 386, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 386, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 386, + 685 + ], + "type": "text", + "content": "4.2.3. HP Model" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "When the image fully expresses the semantic content of the prompt, the ICT score reaches its upper limit. To" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19025" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 296, + 219 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 296, + 219 + ], + "type": "table", + "html": "
ModelAverage↑Pairwise Comparison↑
I2>I1I3>I2I3>I1
Random50.0050.0050.0050.00
ICT-groundtruth87.5562.8099.8799.98
CLIP [26]60.3064.2952.8063.79
ITC [15]57.3663.3348.4260.31
ITM [15]63.3963.8858.2867.99
Aesthetic Score168.1950.9077.8175.83
ImageReward [40]63.8164.5858.0268.84
HPS_v2 [38]72.8866.9871.6380.04
PickScore [13]79.0474.8075.3786.94
ICT87.5864.65100100
HP88.4764.97100100
ICT-HP88.8466.42100100
", + "image_path": "769d00a99dfb7dae167baef5d09db605791fa3caab2293e8c0641f493a42e408.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": "further evaluate the quality of images, we focus solely on the image modality. For each preference triplet, we employ margin ranking loss to fine-tune both the CLIP model and its subsequent MLP layers. Specifically, given a triplet " + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "inline_equation", + "content": "\\{I_1,I_2,I_3\\}" + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "inline_equation", + "content": "I_{3}" + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": " is preferred over " + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "inline_equation", + "content": "I_{2}" + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": " is preferred over " + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "inline_equation", + "content": "I_{1}" + }, + { + "bbox": [ + 55, + 267, + 296, + 352 + ], + "type": "text", + "content": ", we enforce the preference margins as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 81, + 362, + 294, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 362, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 81, + 362, + 294, + 396 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {m a r g i n}} = \\sum [ \\max (0, - \\Delta (I _ {2}, I _ {1}) + m) + \\\\ \\left. \\max \\left(0, - \\Delta \\left(I _ {3}, I _ {2}\\right) + m\\right) \\right], \\tag {13} \\\\ \\end{array}", + "image_path": "ee053b8ccdf6cdf9efd1841d5e490729b43cf5dfacb1cf350e2fddd121afb107.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "inline_equation", + "content": "\\Delta(I_i, I_j)" + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "content": " represents the score difference between images " + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "inline_equation", + "content": "I_i" + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "inline_equation", + "content": "I_j" + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 404, + 295, + 430 + ], + "type": "text", + "content": " is the margin threshold." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 219, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 219, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 219, + 449 + ], + "type": "text", + "content": "4.3. Diffusion Model Optimization" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 295, + 491 + ], + "type": "text", + "content": "We adopt DRaFT-K [5] as our fine-tuning method to optimize Stable-Diffusion-3.5-Large-Turbo [2] by directly maximizing the differentiable reward functions." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 551 + ], + "type": "text", + "content": "Our proposed ICT model and HP model can each be individually utilized as reward functions. Additionally, we multiply the outputs of both models to create a more comprehensive integrated reward function, which we term the ICT-HP model." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 148, + 576 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 583, + 181, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 583, + 181, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 583, + 181, + 597 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 685 + ], + "type": "text", + "content": "Two-Stage Training Pipeline. We adopt a two-stage training method for the ICT model and HP model. In the first stage, we fine-tune the CLIP model using our constructed ICT labels as training targets. In the second stage, we freeze the ICT model parameters and continue training the HP model. During the model optimization phase, we comparatively evaluate" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 156 + ], + "blocks": [ + { + "bbox": [ + 55, + 227, + 295, + 261 + ], + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 261 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 261 + ], + "type": "text", + "content": "Table 1. Preference prediction accuracy on test sets of Pick-High and Pickapic_v2. Bold and underlined values represent optimal and second-best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 156 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 156 + ], + "type": "table", + "html": "
ModelJPEG Compressibility↑Aesthetic Score↑
SD3.5-Large [1]374.806.307
FLUX.1-dev [14]270.586.436
SD3.5-Turbo [2]313.106.293
+ CLIP [26](crash)279.985.175
+ ICT (Ours)321.396.285
+ ICT-HP (Ours)330.236.300
+ HP (Ours)334.866.448
", + "image_path": "de558b0f3e99443579f62c6203e901f92eb902e900349d047002c0945f5c12d8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 198 + ], + "type": "text", + "content": "Table 2. JPEG Compressibility and Aesthetic Scores on Parti-Prompts. Bold and underlined values represent optimal and second-best performance." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 241 + ], + "type": "text", + "content": "three different reward models for optimizing the diffusion model LoRA adapters: ICT model, HP model, and the more comprehensive ICT-HP model." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 553, + 314 + ], + "type": "text", + "content": "Training Datasets. Our training dataset for the ICT model consists of the Pick-High dataset and a subset of Pickapic_v2, comprising a total of 360,000 image triplets. For the diffusion model optimization process, we extract 32,000 non-repetitive text prompts from the Pick-High dataset to serve as our prompt dataset." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 315, + 553, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 553, + 351 + ], + "type": "text", + "content": "For detailed training hyperparameter configurations and experimental specifications, please refer to the supplementary materials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 360, + 470, + 371 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 470, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 470, + 371 + ], + "type": "text", + "content": "5.2. Reward Model Performance" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 378, + 555, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 581 + ], + "type": "text", + "content": "We evaluated the prediction accuracy of human preference models on the Pick-High and Pickapic_v2 test sets. As shown in Table 1, the image-text contrastive (ITC) classification heads of foundation multimodal pre-trained models CLIP and BLIP perform only comparable to random selection. This result indicates that the instance-level image-text alignment objective adopted by foundation multimodal models is not suitable for evaluating the quality of generative models. In terms of overall average accuracy, our proposed ICT-HP model achieves nearly " + }, + { + "bbox": [ + 313, + 378, + 555, + 581 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 378, + 555, + 581 + ], + "type": "text", + "content": " performance improvement compared to Pickscore. This significant improvement stems from the ability of our reward model to measure the extent to which images contain text, and on this basis, continue to measure image quality, thus more accurately identifying high-quality images and assigning them correspondingly high scores." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Qualitative Results on SD3.5-Turbo Optimization. In Figure 4, we present generation results using real user prompts. All images are generated with the same initial noise to ensure fair comparison. The experimental results show that: the HPS_v2 model only enhances image tone with minimal detail improvement; the PickScore model, while increasing details, causes distortions in style and color fidelity; the ImageReward model shows almost no substantial improvement over the original model outputs. In contrast, our proposed reward models achieve performance improvements in" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 693, + 266, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 266, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 266, + 713 + ], + "type": "text", + "content": "1https://github.com/christophschuhmann/improved-aesthetic-predictor" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19026" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 70, + 541, + 336 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 541, + 336 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 541, + 336 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 541, + 336 + ], + "type": "image", + "image_path": "fb789d500c33a463c66876b371e3c225b69d9ff18327c384cfe2b6bd517904ba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 83, + 346, + 526, + 357 + ], + "lines": [ + { + "bbox": [ + 83, + 346, + 526, + 357 + ], + "spans": [ + { + "bbox": [ + 83, + 346, + 526, + 357 + ], + "type": "text", + "content": "Figure 4. Qualitative Comparison of Optimization Results Across Reward Models Using Real User Prompts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 369, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 369, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 369, + 295, + 502 + ], + "type": "text", + "content": "both textural details and tonal aesthetics. Specifically, the ICT model does not penalize increases in image information content, allowing diffusion models to continue optimizing toward richer visual content after achieving text-image alignment; meanwhile, the HP model effectively utilizes pure image features to further optimize when the ICT model reaches its performance ceiling. Compared to our approach, existing reward models tend to produce negative feedback on high-quality image features, thereby limiting their ability to enhance the performance of advanced diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 506, + 295, + 685 + ], + "type": "text", + "content": "Quantitative Evaluation of Visual Information Density and Aesthetic Fidelity. To quantitatively evaluate the enhancement effect of our reward model on the original model in terms of generated image information content and aesthetic quality, we employed two objective metrics, JPEG Compressibility [12] and Aesthetic Score, for assessment on the Parti-Prompts dataset [43]. As shown in Table 2, among all tested models, the HP model achieved the most significant improvements in both metrics, followed by the ICT-HP model. Notably, the SD3.5-Turbo optimized by the HP model not only demonstrated substantial improvement in aesthetic metrics, but even slightly surpassed more comprehensive baseline models such as SD3.5-Large and FLUX.1-dev." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "Qualitative Comparison of Simple Elements Generation Results. We conducted qualitative ex" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 370, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 370, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 370, + 555, + 430 + ], + "type": "text", + "content": "periments on the generation of simple elements, with results shown in Figure 5. The ICT model can stably handle simple geometric shapes and effectively penalize complex elements that do not meet simplicity requirements when prompts include descriptors like \"simple\"." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 434, + 556, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 638 + ], + "type": "text", + "content": "Comprehensive Evaluation on GenEval Benchmark. Table 3 presents the quantitative performance of various image generation models within the multidimensional evaluation framework of GenEval [8]. Using SD3.5-Turbo as the baseline, our ICT-HP and ICT models both demonstrate significant advantages. The ICT-HP model excels in multi-object generation and color fidelity; while the ICT model achieves the highest mean score among all models and delivers optimal performance on the Counting metric, showcasing its superior capability in precise object quantity control. In contrast, the CLIP model suffers comprehensive performance degradation due to training collapse, and PickScore exhibits weaker performance in the Colors metric due to limitations in tonal representation. These comparisons validate our methods' significant advantages in terms of text-image consistency." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Quantitative and Qualitative Performance Comparison of Multi-metric Reward Models. We conducted quantitative analyses of optimization results across various reward models, as shown in Table 4. Experiments reveal that existing reward models including HPS, PickScore, and ImageReward ex" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19027" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 224 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 224 + ], + "type": "table", + "html": "
ModelMean↑Single↑Two↑Counting↑Colors↑Position↑Color Attribution↑
SDXL [22]0.550.980.740.390.850.150.23
DALL-E 2 [28]0.520.940.660.490.770.100.19
DALL-E 3 [21]0.670.960.870.470.830.430.45
SD3 [7]0.680.980.840.660.740.400.43
SD3.5-Turbo [2]0.690.990.870.690.800.250.55
+ ImageReward0.700.990.870.680.800.280.59
+ HPS_v20.690.980.860.680.810.270.57
+ PickScore0.660.990.850.670.740.240.48
+ ICT-HP (Ours)0.700.990.880.680.790.280.58
+ CLIP(crash)0.130.380.040.060.260.010.03
+ ICT (Ours)0.710.980.880.700.810.310.56
", + "image_path": "eda28ee21486d63104d22e13e64b73bccddae1a98245aa1f07b29f8f418d1782.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 233, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 55, + 233, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 55, + 233, + 553, + 257 + ], + "type": "text", + "content": "Table 3. Quantitative Evaluation of Optimization Results Across Reward Models on GenEval Benchmark. Bold and underlined values represent optimal and second-best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 63, + 274, + 288, + 483 + ], + "blocks": [ + { + "bbox": [ + 63, + 274, + 288, + 483 + ], + "lines": [ + { + "bbox": [ + 63, + 274, + 288, + 483 + ], + "spans": [ + { + "bbox": [ + 63, + 274, + 288, + 483 + ], + "type": "image", + "image_path": "6e0eb177eca2e6cb4a406f468ad4f5cdf2adc35e1a02a9aab8862e8dc6a3cd32.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 491, + 295, + 514 + ], + "lines": [ + { + "bbox": [ + 55, + 491, + 295, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 295, + 514 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison of Simple Elements Generation Results." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 532, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 532, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 532, + 295, + 700 + ], + "type": "text", + "content": "hibit performance degradation across multiple metrics. In contrast, our ICT-HP and HP models achieved improvements across all metrics, demonstrating the stability advantage of our method. Although PickScore obtained the highest aesthetic score, qualitative analysis and GenEval experiments indicated significant distortion in artistic style and tone, with color-related metrics significantly lower than other models. Figure 6 compares the original image, refined image, and our model-optimized results. ICT-HP and HP models not only generate intricate backgrounds that are difficult to describe directly through language but also demonstrate a more precise and comprehensive understanding of prompts, thus producing higher-quality images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Feature Transfer Validation. As illustrated in Fig-" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 328, + 274, + 542, + 487 + ], + "blocks": [ + { + "bbox": [ + 328, + 274, + 542, + 487 + ], + "lines": [ + { + "bbox": [ + 328, + 274, + 542, + 487 + ], + "spans": [ + { + "bbox": [ + 328, + 274, + 542, + 487 + ], + "type": "image", + "image_path": "f3181dc1e47d569e7cdcfcd197df7d8eac66a5a6424610e137315e570ffe4fc7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 531 + ], + "type": "text", + "content": "Figure 6. Qualitative Comparison: Original Images, Refined Images, and Generations Optimized by Our Reward Models (HP/ICT-HP)." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "ure 7, transplanting the Text-Encoder trained by our ICT model into the SD2.1-base model significantly improves image quality. This demonstrates that our encoder integrates fine-grained visual details, thus transforming textual descriptions into more precise visual guidance signals. Figure 8 presents results of style transfer with InstantStyle [35]. The CLIP image encoder is replaced by our ICT image encoder. Our fine-tuned encoder not only precisely injects reference styles but also balances text descriptions, overcoming the style-content inconsistency. These experiments validate the effectiveness and advancement of ICT score training objectives." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19028" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 555, + 171 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 171 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 171 + ], + "type": "table", + "html": "
ModelCLIP↑ITM↑ImgRwd↑HPS↑Pick↑Aes↑ICT↑HP↑ICT-HP↑
SD3.5-Te0.3400.8310.9550.27722.0346.5550.9100.7770.717
SD3.5-Tr0.326↓0.757↓0.914↓0.276↓22.012↓6.780↑0.923↑0.781↑0.718↑
+HPS_v20.343↑0.835↑1.065↑0.282↑22.000↓6.590↑0.929↑0.777-0.722↑
+PickScore0.332↓0.816↓1.106↑0.277-22.719↑6.924↑0.929↑0.779↑0.723↑
+ImageReward0.340-0.841↑1.056↑0.279↑22.022↓6.569↑0.865↓0.763↓0.663↓
+ICT(Ours)0.344↑0.841↑1.014↑0.277-22.117↑6.561↑0.937↑0.776↓0.728↑
+ICT-HP(Ours)0.345↑0.839↑1.011↑0.277-22.147↑6.558↑0.938↑0.778↑0.730↑
+HP(Ours)0.342↑0.836↑1.069↑0.278↑22.231↑6.668↑0.937↑0.779↑0.730↑
", + "image_path": "78f0560f985d054c10f15320301641460b8c72714c3fe4730d7deacc11e87068.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 221, + 294, + 434 + ], + "blocks": [ + { + "bbox": [ + 59, + 221, + 294, + 434 + ], + "lines": [ + { + "bbox": [ + 59, + 221, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 59, + 221, + 294, + 434 + ], + "type": "image", + "image_path": "3b97bb62704a1e0aaf875b026c5007287de72600af6aded4e4ee031fa36d014a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 466 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison of ICT Text-Encoder Direct Migration to SD2.1-Base Model." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 69, + 473, + 282, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 178, + 555, + 203 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 203 + ], + "type": "text", + "content": "Table 4. Quantitative Results. SD3.5-Te/Tr denote images generated using easy/reflect prompts on SD-3.5-Turbo. Arrows (↑, ↓, -) indicate changes vs. SD3.5-Te. Bold and underlined values represent optimal and second-best performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 69, + 473, + 282, + 564 + ], + "lines": [ + { + "bbox": [ + 69, + 473, + 282, + 564 + ], + "spans": [ + { + "bbox": [ + 69, + 473, + 282, + 564 + ], + "type": "table", + "html": "
ComparisonModelsWin Rate
vs. BaseICT model0.562
HP model0.709
ICT-HP model0.673
vs. PickScoreHP model0.617
ICT-HP model0.714
", + "image_path": "f3b834f185eb05f1fc7e60f3dbd696be3f4f0412657e423a0aae325d6c2a90ea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "lines": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 571, + 295, + 594 + ], + "type": "text", + "content": "Table 5. Human evaluation on 300 randomly selected samples from Parti-Prompts [43]." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 597, + 132, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 132, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 132, + 611 + ], + "type": "text", + "content": "5.3. User-Study" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "We conduct a user study with ten annotators on 300 randomly selected prompts from Parti-Prompts [43]. Image pairs (ours vs. baseline) were randomly ordered, and annotators evaluated prompt expression and visual appeal. As shown in Table 5, our methods achieve higher win rates against both the SD3.5-large-Turbo base model and the PickScore-optimized model, confirming our approach's effectiveness." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 222, + 553, + 432 + ], + "blocks": [ + { + "bbox": [ + 317, + 222, + 553, + 432 + ], + "lines": [ + { + "bbox": [ + 317, + 222, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 222, + 553, + 432 + ], + "type": "image", + "image_path": "0dbadcdcc17bba6b5739de4ef09d8d300860308a3c11b4e6c10078c5777ef3cd.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 464 + ], + "type": "text", + "content": "Figure 8. Qualitative Comparison of HP/ICT ImageEncoder Performance in Style Injection Tasks." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 476, + 398, + 488 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 398, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 398, + 488 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we reveal the reward-scoring paradox of high-quality images and show that the inherent flaw in current text-to-image generation reward models lies in classic text-image alignment paradigms. We propose a novel training objective for reward models, i.e., ICT (Image-Contained-Text) score. This objective encourages models to achieve text-image alignment and avoid penalizing high-quality images containing extra details, that better match human preferences. Based on the ICT model, we further train a preference model that relies solely on image modality for human preference assessment. Our experiments show that the proposed approach improves scoring accuracy by over " + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 312, + 498, + 555, + 713 + ], + "type": "text", + "content": " compared to existing methods and markedly enhances state-of-the-art models like SD3.5-Turbo. Through both theoretical foundation and empirical validation, our work advances image generation technology toward better alignment with human preferences." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19029" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 147 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China No. 62376277, the Public Computing Cloud of Renmin University of China and the Fund for Building World-Class Universities (Disciplines) of Renmin University of China." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 122, + 179 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 122, + 179 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 122, + 179 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 187, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 62, + 187, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 187, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 62, + 187, + 294, + 220 + ], + "type": "text", + "content": "[1] Stability AI. Stable diffusion 3.5 large. https://huggingface.co/stabilityai/stable-diffusion-3.5-large, 2024.2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 222, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 222, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 222, + 294, + 255 + ], + "type": "text", + "content": "[2] Stability AI. Stable diffusion 3.5 large turbo. https://huggingface.co/stabilityai/stable-diffusion-3.5-large-turbo, 2024. 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 256, + 294, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 256, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 63, + 256, + 294, + 287 + ], + "type": "text", + "content": "[3] Yalong Bai, Mohan Zhou, and Qing Yang. Styleinject: Parameter efficient tuning of text-to-image diffusion models. ArXiv, abs/2401.13942, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 289, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 289, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 289, + 294, + 365 + ], + "type": "text", + "content": "[4] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuhmann, Ludwig Schmidt, and Jenia Jitsev. Reproducible scaling laws for contrastive language-image learning. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), page 2818-2829. IEEE, 2023. 11" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 63, + 367, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 367, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 63, + 367, + 294, + 399 + ], + "type": "text", + "content": "[5] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards, 2024. 5, 11, 13" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 63, + 401, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 401, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 63, + 401, + 294, + 432 + ], + "type": "text", + "content": "[6] Ming Ding, Wendi Zheng, Wenyi Hong, and Jie Tang. Cogview2: Faster and better text-to-image generation via hierarchical transformers, 2022. 11" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 434, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 434, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 63, + 434, + 294, + 499 + ], + "type": "text", + "content": "[7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. arXiv preprint arXiv:2403.03206, 2024. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 63, + 502, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 502, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 63, + 502, + 294, + 544 + ], + "type": "text", + "content": "[8] Dhruba Ghosh, Hanna Hajishirzi, and Ludwig Schmidt. Geneval: An object-focused framework for evaluating text-to-image alignment. CoRR, abs/2310.11513, 2023. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 63, + 546, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 546, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 63, + 546, + 294, + 590 + ], + "type": "text", + "content": "[9] Yaru Hao, Zewen Chi, Li Dong, and Furu Wei. Optimizing prompts for text-to-image generation. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 294, + 634 + ], + "type": "text", + "content": "[10] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 294, + 668 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[12] ISO/IEC. Information technology - digital compression and coding of continuous-tone still images - requirements and guidelines. International Organization for Standardization, 1992. ISO/IEC 10918-1:1992. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 317, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 317, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation, 2023. 2, 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 139 + ], + "type": "text", + "content": "[14] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[15] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 207 + ], + "type": "text", + "content": "[16] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization, 2019. 11, 13" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 252 + ], + "type": "text", + "content": "[17] Xiaoxiao Ma, Mohan Zhou, Tao Liang, Yalong Bai, Tiejun Zhao, Biye Li, Huaian Chen, and Yi Jin. Star: Scale-wise text-conditioned autoregressive image generation. arXiv preprint arXiv:2406.10797, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 255, + 552, + 276 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 552, + 276 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 552, + 276 + ], + "type": "text", + "content": "[18] Midjourney. Ai-powered image generation. https://www.midjourney.com/, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 277, + 554, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 554, + 332 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 554, + 332 + ], + "type": "text", + "content": "[19] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, JiRong Wen, and Qing Yang. Dynamic prompt optimizing for text-to-image generation. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26617-26626, 2024. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 334, + 554, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 334, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 334, + 554, + 408 + ], + "type": "text", + "content": "[20] Wenyi Mo, Tianyu Zhang, Yalong Bai, Bing Su, and Ji-Rong Wen. Uniform attention maps: Boosting image fidelity in reconstruction and editing. In IEEE/CVF Winter Conference on Applications of Computer Vision, WACV 2025, Tucson, AZ, USA, February 26 - March 6, 2025, pages 4420-4429. IEEE, 2025. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 411, + 553, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 432 + ], + "type": "text", + "content": "[21] OpenAI. Dall-e 3. https://openai.com/dall-e-3, 2023.7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 434, + 553, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 434, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 317, + 434, + 553, + 478 + ], + "type": "text", + "content": "[22] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. Sdxl: Improving latent diffusion models for high-resolution image synthesis, 2023. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 544 + ], + "type": "text", + "content": "[23] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In *The Twelfth International Conference on Learning Representations*, 2024. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "text", + "content": "[24] Wenwen Qiang, Jingyao Wang, Zeen Song, Jiangmeng Li, and Changwen Zheng. On the out-of-distribution generalization of self-supervised learning. arXiv preprint arXiv:2505.16675, 2025. 11" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 317, + 591, + 553, + 656 + ], + "type": "text", + "content": "[25] Zeju Qiu, Weiyang Liu, Haiwen Feng, Yuxuan Xue, Yao Feng, Zhen Liu, Dan Zhang, Adrian Weller, and Bernhard Scholkopf. Controlling text-to-image diffusion by orthogonal finetuning. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on ma" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19030" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "chine learning, pages 8748-8763. PmLR, 2021. 1, 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[27] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with CLIP latents. CoRR, abs/2204.06125, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "spans": [ + { + "bbox": [ + 56, + 143, + 294, + 176 + ], + "type": "text", + "content": "[28] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents, 2022. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 178, + 295, + 244 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 295, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 295, + 244 + ], + "type": "text", + "content": "[29] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 246, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 246, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 295, + 312 + ], + "type": "text", + "content": "[30] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. Dreambooth: Fine tuning text-to-image diffusion models for subject-driven generation. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 22500-22510, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 314, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 314, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 314, + 294, + 347 + ], + "type": "text", + "content": "[31] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 349, + 295, + 383 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 295, + 383 + ], + "type": "text", + "content": "[32] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In International Conference on Learning Representations, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 384, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 384, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 56, + 384, + 294, + 429 + ], + "type": "text", + "content": "[33] Zeen Song, Siyu Zhao, Xingyu Zhang, Jiangmeng Li, Changwen Zheng, and Wenwen Qiang. Learning invariant causal mechanism from vision-language models. arXiv preprint arXiv:2405.15289, 2024. 11" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 430, + 295, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 430, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 56, + 430, + 295, + 507 + ], + "type": "text", + "content": "[34] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq R. Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8228-8238, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 509, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 509, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 295, + 553 + ], + "type": "text", + "content": "[35] Haofan Wang, Qixun Wang, Xu Bai, Zekui Qin, and Anthony Chen. Instantstyle: Free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 555, + 295, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 295, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 295, + 600 + ], + "type": "text", + "content": "[36] Zijie J. Wang, Evan Montoya, David Munechika, Haoyang Yang, Benjamin Hoover, and Duen Horng Chau. Diffusiondb: A large-scale prompt gallery dataset for text-to-image generative models, 2023. 11" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 656 + ], + "type": "text", + "content": "[37] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 712 + ], + "type": "text", + "content": "[38] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis, 2023. 2, 5" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 707 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[39] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[40] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation, 2023. 2, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 174, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 261 + ], + "type": "text", + "content": "[41] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In 2018 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2018, Salt Lake City, UT, USA, June 18-22, 2018, pages 1316-1324. Computer Vision Foundation / IEEE Computer Society, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 262, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 262, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 262, + 553, + 328 + ], + "type": "text", + "content": "[42] Zhao Yang, Bing Su, and Ji-Rong Wen. Synthesizing long-term human motions with diffusion models via coherent sampling. In Proceedings of the 31st ACM International Conference on Multimedia, MM 2023, Ottawa, ON, Canada, 29 October 2023-3 November 2023, pages 3954-3964. ACM, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 329, + 553, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 329, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 316, + 329, + 553, + 417 + ], + "type": "text", + "content": "[43] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, Ben Hutchinson, Wei Han, Zarana Parekh, Xin Li, Han Zhang, Jason Baldridge, and Yonghui Wu. Scaling autoregressive models for content-rich text-to-image generation. Transactions on Machine Learning Research, 2022. Featured Certification. 6, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 418, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 418, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 418, + 553, + 472 + ], + "type": "text", + "content": "[44] Guiwei Zhang, Tianyu Zhang, Mohan Zhou, Yalong Bai, and Biye Li. V2flow: Unifying visual tokenization and large language model vocabularies for autoregressive image generation. arXiv preprint arXiv:2503.07493, 2025. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 473, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 473, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 473, + 553, + 529 + ], + "type": "text", + "content": "[45] Yanan Zhang, Jiangmeng Li, Lixiang Liu, and Wenwen Qiang. Rethinking misalignment in vision-language model adaptation from a causal perspective. In The Thirty-eighth Annual Conference on Neural Information Processing Systems. 11" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 529, + 553, + 573 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 529, + 553, + 573 + ], + "spans": [ + { + "bbox": [ + 316, + 529, + 553, + 573 + ], + "type": "text", + "content": "[46] Ziyi Zhang, Li Shen, Sen Zhang, Deheng Ye, Yong Luo, Miaojing Shi, Bo Du, and Dacheng Tao. Aligning few-step diffusion models with dense reward difference learning. arXiv preprint arXiv:2411.11727, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 574, + 553, + 628 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 574, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 316, + 574, + 553, + 628 + ], + "type": "text", + "content": "[47] Fengqi Zhu, Rongzhen Wang, Shen Nie, Xiaolu Zhang, Chunwei Wu, Jun Hu, Jun Zhou, Jianfei Chen, Yankai Lin, Ji-Rong Wen, and Chongxuan Li. Llada 1.5: Variance-reduced preference optimization for large language diffusion models, 2025. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 630, + 553, + 707 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 630, + 553, + 707 + ], + "spans": [ + { + "bbox": [ + 316, + 630, + 553, + 707 + ], + "type": "text", + "content": "[48] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. DM-GAN: dynamic memory generative adversarial networks for text-to-image synthesis. In IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2019, Long Beach, CA, USA, June 16-20, 2019, pages 5802-5810. Computer Vision Foundation / IEEE, 2019. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19031" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_content_list.json b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b6b91b3bc51f07ea950a4d950b5d57b4af3f16f --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_content_list.json @@ -0,0 +1,1882 @@ +[ + { + "type": "text", + "text": "Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation", + "text_level": 1, + "bbox": [ + 107, + 128, + 890, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhenhua Ning $^{1,2,3\\dagger*}$ Zhuotao Tian $^{1*}$ Shaoshuai Shi $^{4}$ Guangming Lu $^{1}$ Daojing He $^{1}$ Wenjie Pei $^{1,2\\boxtimes}$ Li Jiang $^{3\\boxtimes}$ $^{1}$ Harbin Institute of Technology, Shenzhen $^{2}$ Pengcheng Laboratory \n $^{3}$ The Chinese University of Hong Kong, Shenzhen $^{4}$ Voyager Research, Didi Chuxing wenjiecoder@outlook.com jiangli@cuhk.edu.cn", + "bbox": [ + 135, + 200, + 857, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 343 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation $(R^2 S)$ , a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the $R^2 S$ and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work.", + "bbox": [ + 89, + 357, + 485, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 734, + 220, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, multi-modal large language models (MLLMs) [1, 25, 34, 37, 38, 43] have managed to connect language and vision, bringing in a new era of artificial intelligence. These models have shown unprecedented abilities, enabling various fields to understand and create textual and visual content, such as human-computer interaction, multimedia understanding, embodied intelligence, and more.", + "bbox": [ + 89, + 758, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building upon these foundations, recent developments in MLLMs have led to the emergence of 3D MLLMs [6, 16, 17, 33, 35], which demonstrate remarkable capabilities in 3D point cloud perception through enhanced vision-language alignment. Despite these achievements, current 3D MLLM-based approaches exhibit limitations in processing instructions that require sophisticated spatial reasoning. As shown in Fig. 1, the baseline model fails to accurately localize the target object. This observation raises an important question: while 3D point clouds inherently contain richer spatial information (e.g., precise measurements of size and position) compared to 2D images, why do models still struggle to effectively leverage these geometric properties? We posit that the development of robust spatial reasoning capabilities requires two critical components: an explicit reasoning mechanism and a dataset that encompasses complex reasoning scenarios.", + "bbox": [ + 511, + 328, + 906, + 584 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we introduce the Relevant Reasoning Segmentation $(\\mathbb{R}^2\\mathbb{S})$ , a reasoning-based segmentation framework that enhances the spatial reasoning capabilities of 3D MLLMs through two key steps: 1) Reasoning Prior Learning and 2) Prior-guided Refinement. In the first step, the model is trained to generate spatial priors by identifying target-relevant objects, providing essential contextual information. In the second step, the model leverages these reasoning priors to refine hidden representations, facilitating precise target localization. In other words, the two steps emulate human reasoning behavior, i.e., initially identifying relevant elements broadly and then scrutinizing them closely to pinpoint the target.", + "bbox": [ + 511, + 590, + 908, + 787 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Analysis of existing datasets reveals that instructions requiring reasoning are often oversimplified and ambiguous, as illustrated in Fig. 5. This limitation constrains the training and evaluation processes for developing reasoning capabilities in 3D MLLMs. To address this limitation, we present 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes,", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Research Assistant at CUHKSZ. *Equal contribution to this work. Corresponding author.", + "bbox": [ + 107, + 875, + 423, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7851", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4fd67059117f7dafb1e80a13a50e9aedfcd6caacc2b1e29fcbe9f768e2b88f3.jpg", + "image_caption": [ + "Figure 1. In contrast to the baseline model in Sec. 3.1, we introduce Relevant Reasoning Segmentation to explicitly guide the model's reasoning by emphasizing target-relevant objects. In this instance, the object identified as the nearest resting place to the open door should be the chair marked by the red rectangle. The first to third columns in the image represent the prediction from the base method, the outcomes of reasoning prior learning, and the final prediction following prior guided refinement." + ], + "image_footnote": [], + "bbox": [ + 114, + 87, + 369, + 281 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d9366d63eda1eaa8336f1d160f7a54f60dda448b913bb9d9d87fedb6a0bc8590.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 388, + 87, + 883, + 281 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and spatial relationships. The dataset comprises 25,185 training samples and 3,966 validation samples, with quality assurance maintained through manual verification processes. This dataset facilitates enhancement and comprehensive assessment of spatial reasoning capabilities.", + "bbox": [ + 88, + 375, + 480, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Both our proposed $\\mathrm{R}^2\\mathrm{S}$ framework and the 3D ReasonSeg dataset contribute to substantial performance improvements on established benchmarks, demonstrating their complementary effects in enhancing spatial reasoning capabilities. In summary, our main contributions are:", + "bbox": [ + 88, + 450, + 482, + 527 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose the Relevant Segmentation $(\\mathbb{R}^2\\mathbb{S})$ , a reasoning-based segmentation framework that enhances spatial reasoning capabilities of 3D MLLMs by emulating human reasoning processes through two key components: Reasoning Prior Learning and Prior-guided Refinement.", + "- We propose 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes, and spatial relationships, designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs.", + "- Extensive experimental results demonstrate that our $\\mathrm{R}^2\\mathrm{S}$ framework, combined with the 3D ReasonSeg dataset, achieves significant improvements in spatial reasoning across multiple benchmarks." + ], + "bbox": [ + 107, + 531, + 482, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 775, + 233, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 3D Scene Understanding", + "text_level": 1, + "bbox": [ + 89, + 801, + 316, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D scene understanding requires models to detect objects and comprehend their characteristics in 3D space [3, 7, 8, 22, 29, 31]. Among these tasks, 3D segmentation presents unique challenges as it demands point-wise classification within point clouds. Recent studies demonstrate", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "that language integration enhances model performance, as evidenced in 3D Question Answering and 3D Referring tasks [2, 4, 19, 28, 40]. These language-guided tasks, which require models to respond to instructions or identify objects based on descriptions, underscore the importance of language comprehension and spatial reasoning capabilities.", + "bbox": [ + 511, + 375, + 906, + 465 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The integration of language has supported models in transitioning from closed-set to open-set scenarios, such as in 3D Open Vocabulary Segmentation [4, 11, 12, 30, 32]. This step is critical as it allows the model to break free from being constrained by the finite set of categories in the training dataset. While most task-oriented studies focus on specific objectives, recent advancements in 3D multi-modal large language models have enabled addressing multiple 3D downstream tasks using these models [6, 18]. Furthermore, the assistance of 3D multi-modal large language models has empowered models to tackle more intricate object localization tasks [9, 42].", + "bbox": [ + 511, + 465, + 908, + 648 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "These prior studies have not adequately addressed the complex spatial reasoning capabilities required for comprehensive understanding of 3D scenes. In this work, we introduce Relevant Reasoning Segmentation and 3D ReasonSeg to address complex spatial reasoning challenges.", + "bbox": [ + 511, + 648, + 908, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. 3D Multi-modal Large Language Model", + "text_level": 1, + "bbox": [ + 511, + 739, + 856, + 755 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Significant progress has been made in the realm of 3D multi-modal large language models (3D MLLMs). Initial endeavors have predominantly concentrated on point clouds at the object level [14, 15, 26, 35], posing challenges in tackling complex tasks at the scene level. Recently, researchers have shifted attention to understanding entire scene point clouds [6, 16, 17, 33], significantly advancing 3D scene understanding.", + "bbox": [ + 511, + 762, + 906, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Drawing from these advances, we present an efficient", + "bbox": [ + 531, + 885, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7852", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "framework that handles various scene understanding tasks, with particular strength in segmentation. While existing 3D MLLMs can process basic spatial relationships, they struggle with complex spatial reasoning tasks. Our model, however, demonstrates superior performance in handling sophisticated spatial reasoning challenges.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 194, + 181, + 209 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Efforts such as those outlined in [6, 16, 17] were dedicated to leveraging 3D MLLMs for a deeper comprehension of 3D point clouds. Nevertheless, it is evident that current methods encounter challenges in effectively managing intricate spatial relationships. Specifically, as illustrated in Fig. 2, while the model can accurately identify \"What is the object that is used for sitting?\", it struggles with queries such as \"What is the object that is used for sitting and located near the desk with a monitor on it?\"", + "bbox": [ + 89, + 219, + 483, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To tackle this issue, in this work, we instruct the model to generate reasoning priors by paying attention to objects relevant to the target. Subsequently, the model examines the scene and these relevant objects more closely to locate the final target. We refer to this process as Relevant Reasoning Segmentation $(\\mathbb{R}^2\\mathbb{S})$", + "bbox": [ + 89, + 356, + 483, + 446 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The remainder of this section is organized as follows: Sec. 3.1 introduces our baseline model, which demonstrates strong dense perception capabilities on 3D point cloud data. Sec. 3.2 elaborates on our proposed Relevant Reasoning Segmentation framework. Sec. 3.3 details the construction of our 3D ReasonSeg dataset, and Sec. 3.4 describes the comprehensive training pipeline.", + "bbox": [ + 89, + 446, + 483, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Baseline Model", + "text_level": 1, + "bbox": [ + 89, + 561, + 243, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our baseline architecture consists of a multi-modal large language model (MLLM) and a Segmentation Head. Following [23], we extend the MLLM's vocabulary with a special token [SEG] that instructs the Segmentation Head to generate corresponding segmentation mask. This design enables the MLLM to effectively perform fine-grained segmentation tasks while maintaining its core language understanding capabilities.", + "bbox": [ + 89, + 583, + 482, + 703 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-modal Large Language Model. There are three key components in the adopted MLLM for feature extraction: a visual encoder, a Q-Former, and a large language model (LLM). Initially, the point cloud $X_{p}$ is processed through the visual encoder to extract dense features. To optimize computational efficiency, we employ super-point [24] within the visual backbone, aggregating the dense visual features into super-point features $f_{p}$ to reduce the following computational burden. Subsequently, the Q-Former [25] compresses the scene's information into several latent queries $q_{l}$ based on the super-point features $f_{p}$ and text embedding $w$ of the instruction. After that, the latent queries $\\hat{q}_{l}$ will serve as the visual input, along with", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the text embedding $\\mathbf{w}$ , to be processed by the large language model $\\mathcal{F}_{LLM}$ . This process is formulated as:", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {f} _ {p} = \\mathcal {F} _ {e n c} (\\pmb {X} _ {p}), \\quad \\hat {\\pmb {q}} _ {l} = \\mathcal {F} _ {Q} (\\pmb {q} _ {l}, \\pmb {w}, \\pmb {f} _ {p}). \\qquad (1)\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 136, + 906, + 156 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t x t} = \\mathcal {F} _ {L L M} (\\hat {\\boldsymbol {q}} _ {l}, \\boldsymbol {w}). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 174, + 906, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inside the Q-Former $\\mathcal{F}_Q$ , $\\pmb{q}_l$ is concatenated with $\\pmb{w}$ to be the query in both the self-attention and cross-attention layers of $\\mathcal{F}_Q$ , while $\\pmb{f}_p$ serves as the key and value in the cross-attention layer:", + "bbox": [ + 511, + 196, + 906, + 258 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\text {S e l f A t t n} (\\text {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w})), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 271, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\operatorname {C r o s s A t t n} (\\operatorname {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w}), \\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 289, + 905, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Segmentation Head. If the output of LLM, i.e., $\\mathbf{y}_{\\text{txt}}$ , contains [SEG] tokens, it means that the current query requires segmentation predictions. Then, the corresponding hidden features of the [SEG] token, i.e., $\\mathbf{h}_{seg}$ , are processed by the segmentation head $\\mathcal{F}_{dec}$ that follows the decoder structure of [21] to yield the requisite segmentation mask $M$ , based on the reasoning between $\\mathbf{h}_{seg}$ and $\\mathbf{f}_p$ . The process can be formulated as follows.", + "bbox": [ + 511, + 316, + 905, + 438 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {h} _ {\\text {s e g}}, \\boldsymbol {f} _ {p}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 453, + 906, + 470 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The aforementioned framework seamlessly incorporates dense visual perception abilities into the output of the language model, allowing for the fusion of linguistic and visual information processing.", + "bbox": [ + 511, + 476, + 905, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Relevant Reasoning Segmentation", + "text_level": 1, + "bbox": [ + 511, + 546, + 810, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The baseline model can decently handle point cloud perception tasks. However, in our observations, the direct instruction tuning performed in [23] did not consider the spatial relationships among objects in the scene, hindering complex spatial reasoning, as illustrated in Fig. 8. Hence, to address this issue, we suggest that the model initially identifies the objects related to the target and their spatial relationships, serving as the reasoning prior to guide the target localization. This process is thus termed as Relevant Reasoning Segmentation $(\\mathrm{R}^2\\mathrm{S})$ , and it contains two steps as follows.", + "bbox": [ + 511, + 568, + 905, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Step 1: Reasoning Prior Learning. While the model may encounter challenges in understanding intricate interobject relationships, it can recognize objects relevant to the target. As shown in Fig. 2, despite the model's limitation in accurately segmenting the chair beside the desk supporting a monitor, it demonstrates the ability to detect desks and chairs. Given this capability, we may instruct the model to initially segment approximate question-related objects, thus establishing the visual reasoning prior as the hint to help identify the final target.", + "bbox": [ + 511, + 719, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Following the baseline model's process, we first use an instruction structured as: \"Given the 3D", + "bbox": [ + 511, + 869, + 905, + 898 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7853", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/95a33270b54c03ace6259e2db950f676b97fd22a9df65e3af635b0ce522fea87.jpg", + "image_caption": [ + "[QUESTION]: What is the object that is used for sitting?" + ], + "image_footnote": [], + "bbox": [ + 112, + 88, + 431, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/082b1e41b76fab545242a15a66576dfb70f3631be0bfc67072883afb742b74fe.jpg", + "image_caption": [ + "[QUESTION]: What is the object that is used for sitting and located near the desk with a monitor on it?" + ], + "image_footnote": [], + "bbox": [ + 454, + 88, + 883, + 242 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/7bbc94e07301c79120a30b4d64e4c74fb0f544b9d6bb6600d46210f52cfa6a37.jpg", + "image_caption": [ + "Figure 2. The image on the left depicts the LLM's response to the question \"What is the object that is used for sitting?\" Conversely, the right side of the image illustrates the LLM's response to the question \"What is the object that is used for sitting and located near the desk with a monitor on it.\" It is noteworthy that the LLM provides an incorrect answer in the latter case. The accurate response should identify the chair within the red rectangle.", + "Figure 3. The baseline model consists of a visual backbone for point feature encoding, a Q-Former that connects visual-language representations, a LLM for instruction processing, and a mask decoder that converts [SEG] tokens into segmentation masks." + ], + "image_footnote": [], + "bbox": [ + 109, + 352, + 890, + 526 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "scene, [QUESTION]. Please segment the question-related objects that may help answer the question.\" *. The corresponding instruction embedding $\\mathbf{w}$ will be processed by Eq. (1) and Eq. (2) to obtain the text output $\\mathbf{y}_{txt}$ from which we can obtain the [SEG] tokens representing these target-relevant objects. Subsequently, the segmentation head decodes the hidden features $\\mathbf{h}_{seg}$ of [SEG] tokens into masks $M_r$ for these target-relevant objects.", + "bbox": [ + 88, + 598, + 482, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Next, we adopt mask-pooling to aggregate the superpoint features $\\pmb{f}_p$ with masks $M_r$ , yielding the representations $\\pmb{f}_r$ for the target-relevant objects. The features $\\pmb{f}_r$ serve as the reasoning priors in the subsequent step. The above process can be formally expressed as:", + "bbox": [ + 88, + 734, + 482, + 809 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} _ {r} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {h} _ {\\text {s e g}}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 823, + 482, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {f} _ {r} = \\boldsymbol {f} _ {p} \\times \\boldsymbol {M} _ {r} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 859, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Step 2: Prior-guided Refinement. In the absence of explicit instructions, the latent queries $\\hat{\\pmb{q}}_l$ derived from Eq. (1) may exhibit a bias towards irrelevant areas, potentially impacting the reasoning process in Eq. (2). Besides, the mask-pooling operation in Eq. (7) may cause the loss of essential spatial details. Therefore, we propose the prior-guided refinement which commences by feeding the initial latent queries $\\pmb{q}_l$ , text outputs $y_{txt}$ and target-relevant features $\\pmb{f}_r$ from step 1 into the Q-Former $\\mathcal{F}_Q$ to yield the refined query $\\pmb{q}_l'$ and target-relevant features $\\pmb{f}_r'$ .", + "bbox": [ + 511, + 597, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To make use of the obtained reasoning pri- ors, we employ a revised instruction like: \"Given the 3D scene, the question is [QUESTION]. Please answer the question and output segmentation mask. The question-related objects are $\\{f_r^1,f_r^2$ $\\dots ,f_r^n\\}$ , and you may need to pay attention to them\" $^\\dagger$ where the target-relevant", + "bbox": [ + 511, + 750, + 908, + 872 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "*More templates are shown in the supplementary file", + "bbox": [ + 107, + 886, + 388, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "†More templates are shown in the supplementary file.", + "bbox": [ + 531, + 886, + 815, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7854", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0d1dc52e37d91fdca0acf1369377737c6d5e2c3fa49cd61a6e7b462ad655c2bb.jpg", + "image_caption": [ + "Figure 4. The proposed Relevant Reasoning Segmentation framework. Stage 1 segments question-related objects as priors $(f_1', \\dots, f_6')$ . Stage 2 refines hidden representations of instructions based on these priors, enabling the LLM to generate more accurate responses." + ], + "image_footnote": [], + "bbox": [ + 99, + 85, + 888, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "features $\\pmb{f}_r$ of $n$ potentially relevant objects have been inserted to form the new instruction whose text embedding is therefore denoted as $\\pmb{w}' \\oplus \\pmb{f}_r$ . To this end, the prior-guided refinement regarding the latent query and the target-relevant features can be formulated as:", + "bbox": [ + 89, + 297, + 485, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {f} _ {r} ^ {\\prime} = \\mathcal {F} _ {Q} \\left(\\boldsymbol {q} _ {l}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r}, \\boldsymbol {f} _ {p}\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 383, + 482, + 402 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\pmb{w}^{\\prime}$ and $\\pmb{f}_{p}$ are the text embeddings and super-point features, respectively. Unlike Eq. (3), during the prior-guided refinement, $\\pmb{q}_{l}$ is concatenated with the instruction embedding $\\pmb{w}^{\\prime} \\oplus \\pmb{f}_{r}$ to serve as the queries for both self-attention and cross-attention layers in the Q-Former $\\mathcal{F}_Q$ . Meanwhile, $\\pmb{f}_{p}$ continues to function as the key and value in the cross-attention layer. In Eq. (8), the enhanced latent query $\\pmb{q}_{l}^{\\prime}$ has been enriched with target-related clues from $\\pmb{f}_{r}$ . Simultaneously, utilizing the point-cloud features $\\pmb{f}_{p}$ , the target-relevant features $\\pmb{f}_{r}$ are refined with essential spatial intricacies to produce $\\pmb{f}_{r}^{\\prime}$ . Next, the refined query $\\pmb{q}_{l}^{\\prime}$ and target-relevant features $\\pmb{f}_{r}^{\\prime}$ are adopted for generating the new text output $\\pmb{y}_{txt}^{\\prime}$ with the LLM $\\mathcal{F}_{LLM}$ :", + "bbox": [ + 89, + 412, + 483, + 609 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {y} _ {t x t} ^ {\\prime} = \\mathcal {F} _ {L L M} \\left(\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r} ^ {\\prime}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 619, + 482, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the instruction embedding $\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}$ originally used in Eq. (8) has been accordingly updated to $\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}^{\\prime}$ . The hidden features $h_{seg}^{\\prime}$ of the [SEG] tokens are extracted from the new text output $\\pmb{y}_{txt}^{\\prime}$ and processed by the decoder $\\mathcal{F}_{dec}$ to generate the refined final mask prediction $M^{\\prime}$ :", + "bbox": [ + 89, + 648, + 483, + 724 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {M} ^ {\\prime} = \\mathcal {F} _ {d e c} \\left(\\boldsymbol {h} _ {s e g} ^ {\\prime}, \\boldsymbol {f} _ {p}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 734, + 482, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. 3D ReasonSeg", + "text_level": 1, + "bbox": [ + 89, + 772, + 236, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Analysis of existing datasets, as demonstrated in Fig. 5, reveals that spatial relations such as \"in front of\" often lack complexity and exhibit potential ambiguity. Using such datasets for training and evaluation may inadequately address the spatial reasoning requirements of 3D MLLMs. To facilitate enhancement and assessment of spatial reasoning capabilities in 3D MLLMs, we present 3D ReasonSeg, a", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5de7a2200b20097d670dec606552502345bac8b4ac36a4b61ff7b51653e3f773.jpg", + "image_caption": [ + "Figure 5. This illustrates a data sample from ScanRefer, where the target object is described as \"A brown chair placed in front of a plant.\" The spatial relationship in this description shows ambiguity, as there are multiple chairs in front of the plant." + ], + "image_footnote": [], + "bbox": [ + 545, + 295, + 875, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "dataset specifically designed for reasoning tasks in 3D point cloud environments. Empirical results in Tab. 2 indicate that models trained on 3D ReasonSeg achieve substantial performance improvements over baseline implementations.", + "bbox": [ + 511, + 474, + 905, + 535 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the proposed 3D ReasonSeg requires the model to identify the target object based on a question that does not explicitly mention it. The model needs to understand the object functions, visual attributes, and spatial relationships of the objects described in the query and reason to segment the target object. To implement this task, we use a template structured as: \"USER: Given the 3D scene, please answer the question and output segmentation mask: [QUESTION]. ASSISTANT: [ANSWER]\".", + "bbox": [ + 511, + 536, + 906, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We construct the 3D ReasonSeg dataset by integrating SceneVerse's [20] detailed object descriptions with LLama 3.1 [13]. Based on objects' 3D spatial information (positions and sizes) from point clouds and the object descriptions from SceneVerse, LLama 3.1 generates question-reasoning-answer pairs for reasoning tasks.", + "bbox": [ + 511, + 688, + 905, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After collecting numerous question-reasoning-answer pairs, we apply rule-based filtering (e.g., calculating distances between objects and removing outliers) to eliminate instances with unclear or incorrect spatial relationships. The resulting 3D ReasonSeg dataset contains 29,151 high-quality data samples. Examples are shown in Fig. 6, with details provided in the supplementary material.", + "bbox": [ + 511, + 779, + 905, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct a statistical analysis of ScanQA, ScanRe", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7855", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "fer, and 3D ReasonSeg datasets, examining their training (#Train) and validation (#Val) set sizes, average word counts per sample (Avg. Words), and average number of relevant objects (Avg. Objects). The detailed statistics are summarized in Tab. 1, showing that our proposed 3D ReasonSeg dataset contains more complex object relationships.", + "bbox": [ + 89, + 90, + 483, + 183 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/64b6eaeab5d8b41e74d125d9346d53622f62825fff9ae89e9a064c12fba43c6b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Train#ValAvg. WordsAvg. Objects
ScanQA26,5634,6758.81.5
ScanRefer36,6659,50817.81.8
3D ReasonSeg25,1853,96619.65.4
", + "bbox": [ + 104, + 193, + 472, + 253 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Comparative statistics across ScanQA, ScanRefer, and 3D ReasonSeg datasets.", + "bbox": [ + 89, + 263, + 482, + 292 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[QUESTION]: What is the nearest resting place to the bicycle?", + "bbox": [ + 127, + 330, + 316, + 354 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[ANSWER]: The couch [32].", + "bbox": [ + 127, + 380, + 284, + 393 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/36b436c34ab671c1bf579ca79358f853bb7617d314655ec72a7c58a3465fd107.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 330, + 324, + 447, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[QUESTION]: What is the object that is used for storing books and is located near the chair and the door?", + "bbox": [ + 127, + 410, + 320, + 445 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[ANSWER]: The shelf [6].", + "bbox": [ + 127, + 460, + 272, + 474 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a31427c8e0700260019421d7f59901ed18076f43c7cbe19017abcacc322a985.jpg", + "image_caption": [ + "Figure 6. Visualization of 3D ReasonSeg samples, where instance ID colors correspond to their respective object colors in the image." + ], + "image_footnote": [], + "bbox": [ + 331, + 407, + 447, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Model Training", + "text_level": 1, + "bbox": [ + 89, + 550, + 248, + 566 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training Datasets. Following [6], our training data includes established 3D datasets such as ScanQA [2], ScanRefer [4], ScanNet200 [10], the ScanNet subset of 3D-LLM [16], and SceneVerse [20], along with our proposed 3D ReasonSeg dataset.", + "bbox": [ + 89, + 571, + 482, + 647 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For Relevant Reasoning Segmentation outlined in Sec. 3.2, we extend the ScanQA, Scanrefer, and 3D ReasonSeg to incorporate target-relevant object identification. This process leverages scene object annotations and question-answer pairs from the original datasets, with LLama 3.1 [13] serving as the backbone for target-relevant object identification. Representative examples are presented in Fig. 7, with detailed prompts and implementation procedures in the supplementary material.", + "bbox": [ + 89, + 648, + 482, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Loss Function. During training, the model is optimized using the cross-entropy loss $\\mathcal{L}_{\\text {txt }}$ for text generation, along with the binary cross-entropy (BCE) and DICE losses ( $\\mathcal{L}_{\\text {bce }}$ and $\\mathcal{L}_{\\text {dice }}$ ) for segmentation. Following [22], we assign equal weights of 1 to the BCE and DICE loss. Thus, the overall training objective $\\mathcal{L}$ can be expressed as:", + "bbox": [ + 89, + 784, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{txt}$ is the weight for text generation loss.", + "bbox": [ + 511, + 90, + 834, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relevant Objects Augmentation. Our relevant segmentation leverages mask predictions of relevant objects as priors to enhance latent query and text embedding. However, direct implementation leads to overfitting during training, as the ground truth relevant objects remain fixed during training but vary during inference, creating a train-test discrepancy. To address this issue, we implement random omission and addition of relevant objects during training, simulating diverse inference scenarios with incomplete priors.", + "bbox": [ + 511, + 107, + 903, + 243 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[QUESTION]: How many pictures are on the wall?", + "bbox": [ + 535, + 266, + 717, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[ANSWER]: 4", + "bbox": [ + 535, + 290, + 612, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relevant Objects: picture [23][24][25][26]", + "bbox": [ + 535, + 306, + 653, + 330 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7dfdb1ef97733c6182b81dda58a2a35e0891a3460b818bcf537dd615893ed5da.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 751, + 258, + 888, + 334 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[DESCRIPTION]: this is the black kitchen counter with the sink in its. above that kitchen counter is a paper towel dispenser.", + "bbox": [ + 535, + 343, + 746, + 377 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Relevant Objects: kitchen counter [1], sink [11] and paper towel dispenser [15]", + "bbox": [ + 535, + 380, + 684, + 412 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f147fe0b62d136125a486422cfd1030cc48ee882a79343a9b65c629f2ce0afc9.jpg", + "image_caption": [ + "Figure 7. Examples of target-relevant data. The first row shows an instance from ScanQA, and the second row is related to ScanRefer. The visualizations of target-relevant objects are displayed on the right side of each row." + ], + "image_footnote": [], + "bbox": [ + 751, + 340, + 888, + 414 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 513, + 520, + 645, + 537 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 545, + 714, + 561 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the encoder of a pretrained Oneformer3D as visual backbone and the decoder of it as the mask decoder, and a pre-trained OPT 1.3B [39] as the language model. To preserve the learned knowledge of the pre-trained language model, we completely freeze the parameters of the language model while fully training the parameters of all other modules. Note that, we found it is necessary to also fine-tune the visual backbone. We adopt the AdamW [27] optimizer with a weight decay of 0.1 and a learning rate decaying from 1e-4 to 1e-6 with a cosine annealing scheduler. The text generation loss weight $\\lambda_{txt}$ is set to 0.5. The number of learnable queries in Q-Former is set to 32. We randomly sample 300k points from each 3D scene as the 3D input. All the experiments are conducted on eight Nvidia H20 GPUs.", + "bbox": [ + 511, + 566, + 906, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark and Metrics. In this study, we assess our methods on ScanQA [2], ScanRefer [4], and our proposed 3D ReasonSeg. ScanQA is a dataset for the 3D question answering task, evaluating visual and spatial comprehension of 3D environments. The evaluation metrics of ScanQA include BLEU-4, CIDEr, METEOR, and Rouge-L. ScanRefer evaluates the capability to localize 3D object using a natural", + "bbox": [ + 511, + 795, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7856", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/1003638d72d3556cfb37f9ea3fe07ab5c0bf54ec7419ee34d249d5ce4c5b69b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C ↑M ↑R ↑gIoU
ScanRefer [4]37.324.3-----
MVT [19]40.833.3-----
3DVG-Trans [41]45.934.5-----
ViL3DRel [5]47.937.7-----
M3DRef-CLIP [40]51.944.7-----
ScanQA [2]--10.164.913.133.3-
3D-VisTA [44]50.645.813.172.9---
LLM-Grounder [36]17.15.3-----
3D-LLM [16]30.3-12.069.414.535.7-
Chat-3D [33]--6.453.211.928.5-
Chat-3D v2 [17]35.930.47.377.116.140.1-
LL3DA [6]--13.576.815.937.3-
ScanReason [42]53.141.1-----
Grounded 3D-LLM [9]47.944.113.472.7--25.6
baseline54.638.712.870.513.834.6-
+ 3D ReasonSeg58.943.613.375.215.536.729.2
+ R2S59.548.713.977.316.438.133.1
", + "bbox": [ + 155, + 88, + 841, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "language description. Prediction accuracy is measured by the percentage of predictions with Intersection over Union (IoU) above 0.25 and 0.5 compared to ground truth masks. For 3D ReasonSeg, the model is required to focus on object attributes and relationships to reason, answering questions and localizing target objects. Following [23], the general Intersection over Union (gIoU) metric is used, defined as the average of all per-scene Intersection-over-Unions (IoUs). Detailed formulations of these tasks can be found in the supplementary material.", + "bbox": [ + 89, + 453, + 483, + 604 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 618, + 230, + 633 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We evaluate our model's effectiveness through comprehensive comparisons with state-of-the-art (SOTA) methods, encompassing both task-specific specialist models and multitask generalist models. For a fair comparison, we finetune Grounded 3D-LLM [9], the previous SOTA model on ScanRefer, on our proposed 3D ReasonSeg dataset and assess its performance on the 3D ReasonSeg benchmark.", + "bbox": [ + 89, + 641, + 483, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 2, our baseline model is comparable to existing approaches on both ScanRefer and ScanQA. While our baseline model demonstrates robust performance, the incorporation of 3D ReasonSeg and $\\mathbb{R}^2\\mathbb{S}$ yields consistent substantial improvements across all metrics, validating the effectiveness of our approach. Notably, $\\mathbb{R}^2\\mathbb{S}$ significantly enhances the Acc@50 metric on ScanRefer and the gIoU metric on 3D ReasonSeg, tasks that require robust spatial reasoning, aligning with our expectations.", + "bbox": [ + 89, + 748, + 482, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Prior-guided Refinement. To verify the", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/e873547c3d60edbec52bccd7423d3d68b452f94e51b2fe090671aba649601dbd.jpg", + "table_caption": [ + "Table 2. Results among our methods and previous related works. Entries in gray denote models specialized for specific datasets." + ], + "table_footnote": [], + "table_body": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
wo PR58.241.37.766.113.433.227.5
text-based58.443.513.775.515.636.929.8
R2S59.548.713.977.316.438.133.1
", + "bbox": [ + 517, + 450, + 903, + 511 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c6bf2ab6144141d22718abf9a6175a041a42ef0df10dfb5dc5044a0635b0d549.jpg", + "table_caption": [ + "Table 3. Ablation study on the Prior-guided Refinement." + ], + "table_footnote": [], + "table_body": "
number of latent queriesScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C↑M↑R↑gloU
1657.240.514.075.315.236.828.9
3258.943.613.375.215.536.729.2
4858.741.614.574.315.136.128.4
", + "bbox": [ + 526, + 546, + 893, + 601 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4c48f426fc816808cc474feb9c67d46bf2a16c85c9fe7b894781fa3ffdd89514.jpg", + "table_caption": [ + "Table 4. Ablation study on the number of latent queries. The results are based on our base method." + ], + "table_footnote": [], + "table_body": "
λtxtScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
0.161.343.912.368.214.334.028.9
0.558.943.613.375.215.536.729.2
1.058.741.413.272.615.235.628.6
", + "bbox": [ + 517, + 651, + 903, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Ablation study on the loss weight configurations, where $\\lambda_{txt}$ means the weight of $\\mathcal{L}_{txt}$ .", + "bbox": [ + 511, + 724, + 906, + 753 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "necessity of Prior-guided Refinement (PR), we investigate the approach without PR (wo PR) and directly representing target-relevant objects as textual information (text-based). The results are shown in Tab. 3.", + "bbox": [ + 511, + 777, + 906, + 837 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "When Prior-guided Refinement is removed, the model simultaneously predict relevant objects and process instructions. Results indicate that the approach without Prior-guided Refinement not only fails to improve performance", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7857", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/74f26e80c719c52dcc76bd81e6a495e31cfde3bea504bf008023514cd70a69f7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 97, + 305, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What is under the desk in front of the doorway? Please answer the question.", + "bbox": [ + 312, + 107, + 485, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e55245836a1e8f3e32d6354542a37206f14af2071e6e8915a34906b13ffbbf5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 140, + 171, + 307, + 244 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2137490cb9d3d511307f8439be23c0a346ddb12e6cab2b4c042c2b8d762f030b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 313, + 150, + 473, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What is under the desk in front of the doorway? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them.", + "bbox": [ + 310, + 176, + 485, + 228 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASSISTANT: office chair", + "bbox": [ + 313, + 232, + 454, + 243 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7b0424f19987811f88c8f53912f87c5240be412a4c15184aaf8eb6803a1c8c14.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 138, + 253, + 308, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What is in the center of the room? Please answer the question.", + "bbox": [ + 313, + 268, + 478, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/72f9be3e22b9dddc56d7bc895567e67f5cb557720cdae4f6bb5486aa93086ccf.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 315, + 306, + 424, + 319 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What is in the center of the room? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them.", + "bbox": [ + 315, + 333, + 482, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASSISTANT: kitchen island", + "bbox": [ + 315, + 393, + 465, + 405 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask.", + "bbox": [ + 513, + 95, + 805, + 128 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASSISTANT: the towel [SEG]", + "bbox": [ + 514, + 128, + 663, + 138 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dcd364a6105a3588229286a0f8dc234dd69a9a9479c261a25b0b62eda5252206.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 138, + 807, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask. The question-related objects are [SEG][SEG][SEG], and you may need to pay attention to them.", + "bbox": [ + 514, + 250, + 813, + 294 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ASSISTANT: the towel [SEG]", + "bbox": [ + 516, + 296, + 647, + 306 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ed56bd369f5edc2548b1494cccad8d39a27307da0db61915f174b6e218cc4e9f.jpg", + "image_caption": [ + "Figure 8. Visual qualitative results on ScanQA and 3D ReasonSeg." + ], + "image_footnote": [], + "bbox": [ + 514, + 306, + 807, + 410 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "but potentially degrades it. This degradation likely stems from the increased computational complexity of concurrent object prediction and instruction processing without supplementary contextual information.", + "bbox": [ + 89, + 470, + 482, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The text-based method represents an ablation variant of $\\mathbf{R}^2\\mathbf{S}$ where segmentation is omitted, providing only text prompts such as \"The question-related objects are chairs, and you may need to pay attention to them\". Although text-based approach demonstrates modest improvements, our method consistently achieves superior results. This performance advantage can be attributed to our approach's ability to extract precise instance-level visual features through segmentation masks. In contrast, text-based method fails to capture spatial information and demonstrates inherent limitations in differentiating multiple instances of identical categories within a scene (e.g., multiple chairs).", + "bbox": [ + 89, + 549, + 483, + 746 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Number of Latent Queries. We investigate the number of Q-Former latent queries. As demonstrated in Tab. 4, experiments with 16, 32, and 48 latent queries reveal that 32 queries yield optimal performance.", + "bbox": [ + 89, + 762, + 483, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Loss Weight. We conduct an ablation study on the text loss weight $\\lambda_{txt}$ using values of 0.1, 0.5, and 1.0. Our experiments demonstrate that setting $\\lambda_{txt}$ to 1.0 achieves optimal balance in performance, as evidenced in Tab. 5.", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 470, + 694, + 486 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As illustrated in Fig. 8, we present a comparison between the baseline and $\\mathbf{R}^2\\mathbf{S}$ . The left portion of the figure compares baseline predictions (upper row) with $\\mathbf{R}^2\\mathbf{S}$ predictions (lower row) for ScanQA. The lower row shows segmentation results for target-relevant objects. Similarly, the right portion compares baseline and $\\mathbf{R}^2\\mathbf{S}$ predictions for 3D ReasonSeg. Additional qualitative results are provided in the supplementary material.", + "bbox": [ + 511, + 493, + 906, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Limitation", + "text_level": 1, + "bbox": [ + 511, + 626, + 627, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Due to the use of large language models for data generation instead of human annotation, some noise remains in our dataset even after data cleaning procedures, potentially impacting model performance. We will further optimize our dataset in future work.", + "bbox": [ + 511, + 651, + 906, + 726 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 739, + 640, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce the Relevant Reasoning Segmentation framework, a reasoning-based segmentation framework that enhances spatial reasoning through Reasoning Prior Learning and Prior-guided Refinement. Furthermore, we propose 3D ReasonSeg, a reasoning-based segmentation dataset designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs. We hope our work provides new insights and directions for exploring the spatial reasoning capabilities in 3D MLLMs.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7858", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62372133, 62376074), the National Key R&D Program of China (Grant No. 2024YFE0215300), the Shenzhen Fundamental Research Program (Grant No. JCYJ20220818102415032), the Shenzhen Science and Technology Program (Grant No. JSGGKQTD20221101115655027), the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2025A1515011434), and the Guangdong Provincial Key Laboratory of Mathematical Foundations for Artificial Intelligence (Grant No. 2023B1212010001).", + "bbox": [ + 89, + 113, + 485, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 292, + 187, + 308 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1", + "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 6, 7", + "[3] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2021. 2", + "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 6, 7", + "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 7", + "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 1, 2, 3, 6, 7", + "[7] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2", + "[8] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18527-18536, 2023. 2" + ], + "bbox": [ + 99, + 316, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 2, 7", + "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 6", + "[11] Anurag Das, Xinting Hu, Li Jiang, and Bernt Schiele. Mtaclip: Language-guided semantic segmentation with mask-text alignment. arXiv preprint arXiv:2407.21654, 2024. 2", + "[12] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2", + "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, and Akhil Mathur et.al. The llama 3 herd of models, 2024. 5, 6", + "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2", + "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2", + "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 2, 3, 6, 7", + "[17] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 1, 2, 3, 7", + "[18] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2", + "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 2, 7", + "[20] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 5, 6", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7859", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3", + "[22] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 2, 6", + "[23] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 3, 7", + "[24] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 3", + "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 3", + "[26] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 2", + "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2", + "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 2", + "[30] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2", + "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2", + "[32] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2", + "[33] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 2, 7", + "[34] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 1" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 1, 2", + "[36] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 7", + "[37] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 1", + "[38] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1", + "[39] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6", + "[40] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 2, 7", + "[41] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 7", + "[42] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 2, 7", + "[43] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1", + "[44] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 7" + ], + "bbox": [ + 516, + 92, + 903, + 738 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7860", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_model.json b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_model.json new file mode 100644 index 0000000000000000000000000000000000000000..14678b6183a3318d9626d93322fd4f83e04afcbc --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_model.json @@ -0,0 +1,2409 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.13, + 0.892, + 0.177 + ], + "angle": 0, + "content": "Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.137, + 0.202, + 0.858, + 0.294 + ], + "angle": 0, + "content": "Zhenhua Ning\\(^{1,2,3\\dagger*}\\) Zhuotao Tian\\(^{1*}\\) Shaoshuai Shi\\(^{4}\\) Guangming Lu\\(^{1}\\) Daojing He\\(^{1}\\) Wenjie Pei\\(^{1,2\\boxtimes}\\) Li Jiang\\(^{3\\boxtimes}\\) \n\\(^{1}\\)Harbin Institute of Technology, Shenzhen \\(^{2}\\)Pengcheng Laboratory \n\\(^{3}\\)The Chinese University of Hong Kong, Shenzhen \\(^{4}\\)Voyager Research, Didi Chuxing wenjiecoder@outlook.com jiangli@cuhk.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.344 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.358, + 0.486, + 0.707 + ], + "angle": 0, + "content": "Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation \\((R^2 S)\\), a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the \\( R^2 S \\) and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.735, + 0.222, + 0.751 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.76, + 0.485, + 0.866 + ], + "angle": 0, + "content": "In recent years, multi-modal large language models (MLLMs) [1, 25, 34, 37, 38, 43] have managed to connect language and vision, bringing in a new era of artificial intelligence. These models have shown unprecedented abilities, enabling various fields to understand and create textual and visual content, such as human-computer interaction, multimedia understanding, embodied intelligence, and more." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.908, + 0.585 + ], + "angle": 0, + "content": "Building upon these foundations, recent developments in MLLMs have led to the emergence of 3D MLLMs [6, 16, 17, 33, 35], which demonstrate remarkable capabilities in 3D point cloud perception through enhanced vision-language alignment. Despite these achievements, current 3D MLLM-based approaches exhibit limitations in processing instructions that require sophisticated spatial reasoning. As shown in Fig. 1, the baseline model fails to accurately localize the target object. This observation raises an important question: while 3D point clouds inherently contain richer spatial information (e.g., precise measurements of size and position) compared to 2D images, why do models still struggle to effectively leverage these geometric properties? We posit that the development of robust spatial reasoning capabilities requires two critical components: an explicit reasoning mechanism and a dataset that encompasses complex reasoning scenarios." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.909, + 0.789 + ], + "angle": 0, + "content": "In this work, we introduce the Relevant Reasoning Segmentation \\((\\mathbb{R}^2\\mathbb{S})\\), a reasoning-based segmentation framework that enhances the spatial reasoning capabilities of 3D MLLMs through two key steps: 1) Reasoning Prior Learning and 2) Prior-guided Refinement. In the first step, the model is trained to generate spatial priors by identifying target-relevant objects, providing essential contextual information. In the second step, the model leverages these reasoning priors to refine hidden representations, facilitating precise target localization. In other words, the two steps emulate human reasoning behavior, i.e., initially identifying relevant elements broadly and then scrutinizing them closely to pinpoint the target." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Analysis of existing datasets reveals that instructions requiring reasoning are often oversimplified and ambiguous, as illustrated in Fig. 5. This limitation constrains the training and evaluation processes for developing reasoning capabilities in 3D MLLMs. To address this limitation, we present 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes," + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.424, + 0.901 + ], + "angle": 0, + "content": "Research Assistant at CUHKSZ. *Equal contribution to this work. Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7851" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.088, + 0.37, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.389, + 0.088, + 0.885, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.293, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 1. In contrast to the baseline model in Sec. 3.1, we introduce Relevant Reasoning Segmentation to explicitly guide the model's reasoning by emphasizing target-relevant objects. In this instance, the object identified as the nearest resting place to the open door should be the chair marked by the red rectangle. The first to third columns in the image represent the prediction from the base method, the outcomes of reasoning prior learning, and the final prediction following prior guided refinement." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.375, + 0.482, + 0.451 + ], + "angle": 0, + "content": "and spatial relationships. The dataset comprises 25,185 training samples and 3,966 validation samples, with quality assurance maintained through manual verification processes. This dataset facilitates enhancement and comprehensive assessment of spatial reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.452, + 0.483, + 0.528 + ], + "angle": 0, + "content": "Both our proposed \\(\\mathrm{R}^2\\mathrm{S}\\) framework and the 3D ReasonSeg dataset contribute to substantial performance improvements on established benchmarks, demonstrating their complementary effects in enhancing spatial reasoning capabilities. In summary, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.532, + 0.483, + 0.622 + ], + "angle": 0, + "content": "- We propose the Relevant Segmentation \\((\\mathbb{R}^2\\mathbb{S})\\), a reasoning-based segmentation framework that enhances spatial reasoning capabilities of 3D MLLMs by emulating human reasoning processes through two key components: Reasoning Prior Learning and Prior-guided Refinement." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.623, + 0.483, + 0.698 + ], + "angle": 0, + "content": "- We propose 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes, and spatial relationships, designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.699, + 0.483, + 0.759 + ], + "angle": 0, + "content": "- Extensive experimental results demonstrate that our \\(\\mathrm{R}^2\\mathrm{S}\\) framework, combined with the 3D ReasonSeg dataset, achieves significant improvements in spatial reasoning across multiple benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.532, + 0.483, + 0.759 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.776, + 0.234, + 0.792 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.318, + 0.818 + ], + "angle": 0, + "content": "2.1. 3D Scene Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "3D scene understanding requires models to detect objects and comprehend their characteristics in 3D space [3, 7, 8, 22, 29, 31]. Among these tasks, 3D segmentation presents unique challenges as it demands point-wise classification within point clouds. Recent studies demonstrate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.466 + ], + "angle": 0, + "content": "that language integration enhances model performance, as evidenced in 3D Question Answering and 3D Referring tasks [2, 4, 19, 28, 40]. These language-guided tasks, which require models to respond to instructions or identify objects based on descriptions, underscore the importance of language comprehension and spatial reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.649 + ], + "angle": 0, + "content": "The integration of language has supported models in transitioning from closed-set to open-set scenarios, such as in 3D Open Vocabulary Segmentation [4, 11, 12, 30, 32]. This step is critical as it allows the model to break free from being constrained by the finite set of categories in the training dataset. While most task-oriented studies focus on specific objectives, recent advancements in 3D multi-modal large language models have enabled addressing multiple 3D downstream tasks using these models [6, 18]. Furthermore, the assistance of 3D multi-modal large language models has empowered models to tackle more intricate object localization tasks [9, 42]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.65, + 0.909, + 0.727 + ], + "angle": 0, + "content": "These prior studies have not adequately addressed the complex spatial reasoning capabilities required for comprehensive understanding of 3D scenes. In this work, we introduce Relevant Reasoning Segmentation and 3D ReasonSeg to address complex spatial reasoning challenges." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.857, + 0.756 + ], + "angle": 0, + "content": "2.2. 3D Multi-modal Large Language Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Significant progress has been made in the realm of 3D multi-modal large language models (3D MLLMs). Initial endeavors have predominantly concentrated on point clouds at the object level [14, 15, 26, 35], posing challenges in tackling complex tasks at the scene level. Recently, researchers have shifted attention to understanding entire scene point clouds [6, 16, 17, 33], significantly advancing 3D scene understanding." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Drawing from these advances, we present an efficient" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "framework that handles various scene understanding tasks, with particular strength in segmentation. While existing 3D MLLMs can process basic spatial relationships, they struggle with complex spatial reasoning tasks. Our model, however, demonstrates superior performance in handling sophisticated spatial reasoning challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.195, + 0.182, + 0.21 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.356 + ], + "angle": 0, + "content": "Efforts such as those outlined in [6, 16, 17] were dedicated to leveraging 3D MLLMs for a deeper comprehension of 3D point clouds. Nevertheless, it is evident that current methods encounter challenges in effectively managing intricate spatial relationships. Specifically, as illustrated in Fig. 2, while the model can accurately identify \"What is the object that is used for sitting?\", it struggles with queries such as \"What is the object that is used for sitting and located near the desk with a monitor on it?\"" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.447 + ], + "angle": 0, + "content": "To tackle this issue, in this work, we instruct the model to generate reasoning priors by paying attention to objects relevant to the target. Subsequently, the model examines the scene and these relevant objects more closely to locate the final target. We refer to this process as Relevant Reasoning Segmentation \\((\\mathbb{R}^2\\mathbb{S})\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.553 + ], + "angle": 0, + "content": "The remainder of this section is organized as follows: Sec. 3.1 introduces our baseline model, which demonstrates strong dense perception capabilities on 3D point cloud data. Sec. 3.2 elaborates on our proposed Relevant Reasoning Segmentation framework. Sec. 3.3 details the construction of our 3D ReasonSeg dataset, and Sec. 3.4 describes the comprehensive training pipeline." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.562, + 0.245, + 0.576 + ], + "angle": 0, + "content": "3.1. Baseline Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Our baseline architecture consists of a multi-modal large language model (MLLM) and a Segmentation Head. Following [23], we extend the MLLM's vocabulary with a special token [SEG] that instructs the Segmentation Head to generate corresponding segmentation mask. This design enables the MLLM to effectively perform fine-grained segmentation tasks while maintaining its core language understanding capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Multi-modal Large Language Model. There are three key components in the adopted MLLM for feature extraction: a visual encoder, a Q-Former, and a large language model (LLM). Initially, the point cloud \\( X_{p} \\) is processed through the visual encoder to extract dense features. To optimize computational efficiency, we employ super-point [24] within the visual backbone, aggregating the dense visual features into super-point features \\( f_{p} \\) to reduce the following computational burden. Subsequently, the Q-Former [25] compresses the scene's information into several latent queries \\( q_{l} \\) based on the super-point features \\( f_{p} \\) and text embedding \\( w \\) of the instruction. After that, the latent queries \\( \\hat{q}_{l} \\) will serve as the visual input, along with" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "the text embedding \\( \\mathbf{w} \\), to be processed by the large language model \\( \\mathcal{F}_{LLM} \\). This process is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.137, + 0.907, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\pmb {f} _ {p} = \\mathcal {F} _ {e n c} (\\pmb {X} _ {p}), \\quad \\hat {\\pmb {q}} _ {l} = \\mathcal {F} _ {Q} (\\pmb {q} _ {l}, \\pmb {w}, \\pmb {f} _ {p}). \\qquad (1)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.175, + 0.907, + 0.192 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t x t} = \\mathcal {F} _ {L L M} (\\hat {\\boldsymbol {q}} _ {l}, \\boldsymbol {w}). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.259 + ], + "angle": 0, + "content": "Inside the Q-Former \\(\\mathcal{F}_Q\\), \\(\\pmb{q}_l\\) is concatenated with \\(\\pmb{w}\\) to be the query in both the self-attention and cross-attention layers of \\(\\mathcal{F}_Q\\), while \\(\\pmb{f}_p\\) serves as the key and value in the cross-attention layer:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.272, + 0.907, + 0.288 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\text {S e l f A t t n} (\\text {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w})), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.29, + 0.906, + 0.308 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\operatorname {C r o s s A t t n} (\\operatorname {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w}), \\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p}). \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.906, + 0.439 + ], + "angle": 0, + "content": "Segmentation Head. If the output of LLM, i.e., \\( \\mathbf{y}_{\\text{txt}} \\), contains [SEG] tokens, it means that the current query requires segmentation predictions. Then, the corresponding hidden features of the [SEG] token, i.e., \\( \\mathbf{h}_{seg} \\), are processed by the segmentation head \\( \\mathcal{F}_{dec} \\) that follows the decoder structure of [21] to yield the requisite segmentation mask \\( M \\), based on the reasoning between \\( \\mathbf{h}_{seg} \\) and \\( \\mathbf{f}_p \\). The process can be formulated as follows." + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.454, + 0.907, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {h} _ {\\text {s e g}}, \\boldsymbol {f} _ {p}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.538 + ], + "angle": 0, + "content": "The aforementioned framework seamlessly incorporates dense visual perception abilities into the output of the language model, allowing for the fusion of linguistic and visual information processing." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.547, + 0.812, + 0.563 + ], + "angle": 0, + "content": "3.2. Relevant Reasoning Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.719 + ], + "angle": 0, + "content": "The baseline model can decently handle point cloud perception tasks. However, in our observations, the direct instruction tuning performed in [23] did not consider the spatial relationships among objects in the scene, hindering complex spatial reasoning, as illustrated in Fig. 8. Hence, to address this issue, we suggest that the model initially identifies the objects related to the target and their spatial relationships, serving as the reasoning prior to guide the target localization. This process is thus termed as Relevant Reasoning Segmentation \\((\\mathrm{R}^2\\mathrm{S})\\), and it contains two steps as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Step 1: Reasoning Prior Learning. While the model may encounter challenges in understanding intricate interobject relationships, it can recognize objects relevant to the target. As shown in Fig. 2, despite the model's limitation in accurately segmenting the chair beside the desk supporting a monitor, it demonstrates the ability to detect desks and chairs. Given this capability, we may instruct the model to initially segment approximate question-related objects, thus establishing the visual reasoning prior as the hint to help identify the final target." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Following the baseline model's process, we first use an instruction structured as: \"Given the 3D" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7853" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.089, + 0.432, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.159, + 0.246, + 0.396, + 0.273 + ], + "angle": 0, + "content": "[QUESTION]: What is the object that is used for sitting?" + }, + { + "type": "image", + "bbox": [ + 0.455, + 0.089, + 0.885, + 0.243 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.49, + 0.246, + 0.866, + 0.272 + ], + "angle": 0, + "content": "[QUESTION]: What is the object that is used for sitting and located near the desk with a monitor on it?" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.283, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Figure 2. The image on the left depicts the LLM's response to the question \"What is the object that is used for sitting?\" Conversely, the right side of the image illustrates the LLM's response to the question \"What is the object that is used for sitting and located near the desk with a monitor on it.\" It is noteworthy that the LLM provides an incorrect answer in the latter case. The accurate response should identify the chair within the red rectangle." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.353, + 0.891, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.544, + 0.908, + 0.573 + ], + "angle": 0, + "content": "Figure 3. The baseline model consists of a visual backbone for point feature encoding, a Q-Former that connects visual-language representations, a LLM for instruction processing, and a mask decoder that converts [SEG] tokens into segmentation masks." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.599, + 0.483, + 0.734 + ], + "angle": 0, + "content": "scene, [QUESTION]. Please segment the question-related objects that may help answer the question.\" *. The corresponding instruction embedding \\( \\mathbf{w} \\) will be processed by Eq. (1) and Eq. (2) to obtain the text output \\( \\mathbf{y}_{txt} \\) from which we can obtain the [SEG] tokens representing these target-relevant objects. Subsequently, the segmentation head decodes the hidden features \\( \\mathbf{h}_{seg} \\) of [SEG] tokens into masks \\( M_r \\) for these target-relevant objects." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.735, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Next, we adopt mask-pooling to aggregate the superpoint features \\( \\pmb{f}_p \\) with masks \\( M_r \\), yielding the representations \\( \\pmb{f}_r \\) for the target-relevant objects. The features \\( \\pmb{f}_r \\) serve as the reasoning priors in the subsequent step. The above process can be formally expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.824, + 0.483, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} _ {r} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {h} _ {\\text {s e g}}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.231, + 0.861, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {f} _ {r} = \\boldsymbol {f} _ {p} \\times \\boldsymbol {M} _ {r} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.75 + ], + "angle": 0, + "content": "Step 2: Prior-guided Refinement. In the absence of explicit instructions, the latent queries \\(\\hat{\\pmb{q}}_l\\) derived from Eq. (1) may exhibit a bias towards irrelevant areas, potentially impacting the reasoning process in Eq. (2). Besides, the mask-pooling operation in Eq. (7) may cause the loss of essential spatial details. Therefore, we propose the prior-guided refinement which commences by feeding the initial latent queries \\(\\pmb{q}_l\\), text outputs \\(y_{txt}\\) and target-relevant features \\(\\pmb{f}_r\\) from step 1 into the Q-Former \\(\\mathcal{F}_Q\\) to yield the refined query \\(\\pmb{q}_l'\\) and target-relevant features \\(\\pmb{f}_r'\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.909, + 0.873 + ], + "angle": 0, + "content": "To make use of the obtained reasoning pri- ors, we employ a revised instruction like: \"Given the 3D scene, the question is [QUESTION]. Please answer the question and output segmentation mask. The question-related objects are \\(\\{f_r^1,f_r^2\\) \\(\\dots ,f_r^n\\}\\) , and you may need to pay attention to them\" \\(^\\dagger\\) where the target-relevant" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.39, + 0.901 + ], + "angle": 0, + "content": "*More templates are shown in the supplementary file" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.816, + 0.901 + ], + "angle": 0, + "content": "†More templates are shown in the supplementary file." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7854" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.086, + 0.89, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.908, + 0.274 + ], + "angle": 0, + "content": "Figure 4. The proposed Relevant Reasoning Segmentation framework. Stage 1 segments question-related objects as priors \\((f_1', \\dots, f_6')\\). Stage 2 refines hidden representations of instructions based on these priors, enabling the LLM to generate more accurate responses." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.486, + 0.375 + ], + "angle": 0, + "content": "features \\( \\pmb{f}_r \\) of \\( n \\) potentially relevant objects have been inserted to form the new instruction whose text embedding is therefore denoted as \\( \\pmb{w}' \\oplus \\pmb{f}_r \\). To this end, the prior-guided refinement regarding the latent query and the target-relevant features can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.183, + 0.385, + 0.483, + 0.404 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {f} _ {r} ^ {\\prime} = \\mathcal {F} _ {Q} \\left(\\boldsymbol {q} _ {l}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r}, \\boldsymbol {f} _ {p}\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.611 + ], + "angle": 0, + "content": "where \\(\\pmb{w}^{\\prime}\\) and \\(\\pmb{f}_{p}\\) are the text embeddings and super-point features, respectively. Unlike Eq. (3), during the prior-guided refinement, \\(\\pmb{q}_{l}\\) is concatenated with the instruction embedding \\(\\pmb{w}^{\\prime} \\oplus \\pmb{f}_{r}\\) to serve as the queries for both self-attention and cross-attention layers in the Q-Former \\(\\mathcal{F}_Q\\). Meanwhile, \\(\\pmb{f}_{p}\\) continues to function as the key and value in the cross-attention layer. In Eq. (8), the enhanced latent query \\(\\pmb{q}_{l}^{\\prime}\\) has been enriched with target-related clues from \\(\\pmb{f}_{r}\\). Simultaneously, utilizing the point-cloud features \\(\\pmb{f}_{p}\\), the target-relevant features \\(\\pmb{f}_{r}\\) are refined with essential spatial intricacies to produce \\(\\pmb{f}_{r}^{\\prime}\\). Next, the refined query \\(\\pmb{q}_{l}^{\\prime}\\) and target-relevant features \\(\\pmb{f}_{r}^{\\prime}\\) are adopted for generating the new text output \\(\\pmb{y}_{txt}^{\\prime}\\) with the LLM \\(\\mathcal{F}_{LLM}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.189, + 0.62, + 0.483, + 0.641 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {y} _ {t x t} ^ {\\prime} = \\mathcal {F} _ {L L M} \\left(\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r} ^ {\\prime}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.484, + 0.725 + ], + "angle": 0, + "content": "where the instruction embedding \\(\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}\\) originally used in Eq. (8) has been accordingly updated to \\(\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}^{\\prime}\\). The hidden features \\(h_{seg}^{\\prime}\\) of the [SEG] tokens are extracted from the new text output \\(\\pmb{y}_{txt}^{\\prime}\\) and processed by the decoder \\(\\mathcal{F}_{dec}\\) to generate the refined final mask prediction \\(M^{\\prime}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.736, + 0.483, + 0.755 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {M} ^ {\\prime} = \\mathcal {F} _ {d e c} \\left(\\boldsymbol {h} _ {s e g} ^ {\\prime}, \\boldsymbol {f} _ {p}\\right) \\tag {10}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.773, + 0.238, + 0.79 + ], + "angle": 0, + "content": "3.3. 3D ReasonSeg" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Analysis of existing datasets, as demonstrated in Fig. 5, reveals that spatial relations such as \"in front of\" often lack complexity and exhibit potential ambiguity. Using such datasets for training and evaluation may inadequately address the spatial reasoning requirements of 3D MLLMs. To facilitate enhancement and assessment of spatial reasoning capabilities in 3D MLLMs, we present 3D ReasonSeg, a" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.296, + 0.877, + 0.382 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.392, + 0.908, + 0.45 + ], + "angle": 0, + "content": "Figure 5. This illustrates a data sample from ScanRefer, where the target object is described as \"A brown chair placed in front of a plant.\" The spatial relationship in this description shows ambiguity, as there are multiple chairs in front of the plant." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.906, + 0.536 + ], + "angle": 0, + "content": "dataset specifically designed for reasoning tasks in 3D point cloud environments. Empirical results in Tab. 2 indicate that models trained on 3D ReasonSeg achieve substantial performance improvements over baseline implementations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.688 + ], + "angle": 0, + "content": "Specifically, the proposed 3D ReasonSeg requires the model to identify the target object based on a question that does not explicitly mention it. The model needs to understand the object functions, visual attributes, and spatial relationships of the objects described in the query and reason to segment the target object. To implement this task, we use a template structured as: \"USER: Given the 3D scene, please answer the question and output segmentation mask: [QUESTION]. ASSISTANT: [ANSWER]\"." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.779 + ], + "angle": 0, + "content": "We construct the 3D ReasonSeg dataset by integrating SceneVerse's [20] detailed object descriptions with LLama 3.1 [13]. Based on objects' 3D spatial information (positions and sizes) from point clouds and the object descriptions from SceneVerse, LLama 3.1 generates question-reasoning-answer pairs for reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.906, + 0.885 + ], + "angle": 0, + "content": "After collecting numerous question-reasoning-answer pairs, we apply rule-based filtering (e.g., calculating distances between objects and removing outliers) to eliminate instances with unclear or incorrect spatial relationships. The resulting 3D ReasonSeg dataset contains 29,151 high-quality data samples. Examples are shown in Fig. 6, with details provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We conduct a statistical analysis of ScanQA, ScanRe" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7855" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.184 + ], + "angle": 0, + "content": "fer, and 3D ReasonSeg datasets, examining their training (#Train) and validation (#Val) set sizes, average word counts per sample (Avg. Words), and average number of relevant objects (Avg. Objects). The detailed statistics are summarized in Tab. 1, showing that our proposed 3D ReasonSeg dataset contains more complex object relationships." + }, + { + "type": "table", + "bbox": [ + 0.106, + 0.194, + 0.473, + 0.254 + ], + "angle": 0, + "content": "
Dataset#Train#ValAvg. WordsAvg. Objects
ScanQA26,5634,6758.81.5
ScanRefer36,6659,50817.81.8
3D ReasonSeg25,1853,96619.65.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.264, + 0.483, + 0.293 + ], + "angle": 0, + "content": "Table 1. Comparative statistics across ScanQA, ScanRefer, and 3D ReasonSeg datasets." + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.331, + 0.318, + 0.356 + ], + "angle": 0, + "content": "[QUESTION]: What is the nearest resting place to the bicycle?" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.381, + 0.285, + 0.395 + ], + "angle": 0, + "content": "[ANSWER]: The couch [32]." + }, + { + "type": "image", + "bbox": [ + 0.331, + 0.325, + 0.449, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.411, + 0.321, + 0.446 + ], + "angle": 0, + "content": "[QUESTION]: What is the object that is used for storing books and is located near the chair and the door?" + }, + { + "type": "text", + "bbox": [ + 0.129, + 0.462, + 0.273, + 0.476 + ], + "angle": 0, + "content": "[ANSWER]: The shelf [6]." + }, + { + "type": "image", + "bbox": [ + 0.332, + 0.409, + 0.449, + 0.48 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.498, + 0.483, + 0.527 + ], + "angle": 0, + "content": "Figure 6. Visualization of 3D ReasonSeg samples, where instance ID colors correspond to their respective object colors in the image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.551, + 0.249, + 0.567 + ], + "angle": 0, + "content": "3.4. Model Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.573, + 0.483, + 0.648 + ], + "angle": 0, + "content": "Training Datasets. Following [6], our training data includes established 3D datasets such as ScanQA [2], ScanRefer [4], ScanNet200 [10], the ScanNet subset of 3D-LLM [16], and SceneVerse [20], along with our proposed 3D ReasonSeg dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.649, + 0.483, + 0.784 + ], + "angle": 0, + "content": "For Relevant Reasoning Segmentation outlined in Sec. 3.2, we extend the ScanQA, Scanrefer, and 3D ReasonSeg to incorporate target-relevant object identification. This process leverages scene object annotations and question-answer pairs from the original datasets, with LLama 3.1 [13] serving as the backbone for target-relevant object identification. Representative examples are presented in Fig. 7, with detailed prompts and implementation procedures in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.785, + 0.483, + 0.875 + ], + "angle": 0, + "content": "Loss Function. During training, the model is optimized using the cross-entropy loss \\(\\mathcal{L}_{\\text {txt }}\\) for text generation, along with the binary cross-entropy (BCE) and DICE losses (\\(\\mathcal{L}_{\\text {bce }}\\) and \\(\\mathcal{L}_{\\text {dice }}\\)) for segmentation. Following [22], we assign equal weights of 1 to the BCE and DICE loss. Thus, the overall training objective \\(\\mathcal{L}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.835, + 0.107 + ], + "angle": 0, + "content": "where \\(\\lambda_{txt}\\) is the weight for text generation loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.108, + 0.905, + 0.244 + ], + "angle": 0, + "content": "Relevant Objects Augmentation. Our relevant segmentation leverages mask predictions of relevant objects as priors to enhance latent query and text embedding. However, direct implementation leads to overfitting during training, as the ground truth relevant objects remain fixed during training but vary during inference, creating a train-test discrepancy. To address this issue, we implement random omission and addition of relevant objects during training, simulating diverse inference scenarios with incomplete priors." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.267, + 0.718, + 0.289 + ], + "angle": 0, + "content": "[QUESTION]: How many pictures are on the wall?" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.291, + 0.613, + 0.304 + ], + "angle": 0, + "content": "[ANSWER]: 4" + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.308, + 0.654, + 0.332 + ], + "angle": 0, + "content": "Relevant Objects: picture [23][24][25][26]" + }, + { + "type": "image", + "bbox": [ + 0.752, + 0.26, + 0.89, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.344, + 0.748, + 0.378 + ], + "angle": 0, + "content": "[DESCRIPTION]: this is the black kitchen counter with the sink in its. above that kitchen counter is a paper towel dispenser." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.381, + 0.685, + 0.414 + ], + "angle": 0, + "content": "Relevant Objects: kitchen counter [1], sink [11] and paper towel dispenser [15]" + }, + { + "type": "image", + "bbox": [ + 0.753, + 0.341, + 0.89, + 0.415 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.433, + 0.906, + 0.488 + ], + "angle": 0, + "content": "Figure 7. Examples of target-relevant data. The first row shows an instance from ScanQA, and the second row is related to ScanRefer. The visualizations of target-relevant objects are displayed on the right side of each row." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.521, + 0.646, + 0.538 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.546, + 0.715, + 0.563 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.794 + ], + "angle": 0, + "content": "Implementation Details. We adopt the encoder of a pretrained Oneformer3D as visual backbone and the decoder of it as the mask decoder, and a pre-trained OPT 1.3B [39] as the language model. To preserve the learned knowledge of the pre-trained language model, we completely freeze the parameters of the language model while fully training the parameters of all other modules. Note that, we found it is necessary to also fine-tune the visual backbone. We adopt the AdamW [27] optimizer with a weight decay of 0.1 and a learning rate decaying from 1e-4 to 1e-6 with a cosine annealing scheduler. The text generation loss weight \\(\\lambda_{txt}\\) is set to 0.5. The number of learnable queries in Q-Former is set to 32. We randomly sample 300k points from each 3D scene as the 3D input. All the experiments are conducted on eight Nvidia H20 GPUs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Benchmark and Metrics. In this study, we assess our methods on ScanQA [2], ScanRefer [4], and our proposed 3D ReasonSeg. ScanQA is a dataset for the 3D question answering task, evaluating visual and spatial comprehension of 3D environments. The evaluation metrics of ScanQA include BLEU-4, CIDEr, METEOR, and Rouge-L. ScanRefer evaluates the capability to localize 3D object using a natural" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7856" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.156, + 0.089, + 0.842, + 0.403 + ], + "angle": 0, + "content": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C ↑M ↑R ↑gIoU
ScanRefer [4]37.324.3-----
MVT [19]40.833.3-----
3DVG-Trans [41]45.934.5-----
ViL3DRel [5]47.937.7-----
M3DRef-CLIP [40]51.944.7-----
ScanQA [2]--10.164.913.133.3-
3D-VisTA [44]50.645.813.172.9---
LLM-Grounder [36]17.15.3-----
3D-LLM [16]30.3-12.069.414.535.7-
Chat-3D [33]--6.453.211.928.5-
Chat-3D v2 [17]35.930.47.377.116.140.1-
LL3DA [6]--13.576.815.937.3-
ScanReason [42]53.141.1-----
Grounded 3D-LLM [9]47.944.113.472.7--25.6
baseline54.638.712.870.513.834.6-
+ 3D ReasonSeg58.943.613.375.215.536.729.2
+ R2S59.548.713.977.316.438.133.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.12, + 0.413, + 0.877, + 0.428 + ], + "angle": 0, + "content": "Table 2. Results among our methods and previous related works. Entries in gray denote models specialized for specific datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.454, + 0.484, + 0.606 + ], + "angle": 0, + "content": "language description. Prediction accuracy is measured by the percentage of predictions with Intersection over Union (IoU) above 0.25 and 0.5 compared to ground truth masks. For 3D ReasonSeg, the model is required to focus on object attributes and relationships to reason, answering questions and localizing target objects. Following [23], the general Intersection over Union (gIoU) metric is used, defined as the average of all per-scene Intersection-over-Unions (IoUs). Detailed formulations of these tasks can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.619, + 0.231, + 0.634 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.484, + 0.748 + ], + "angle": 0, + "content": "We evaluate our model's effectiveness through comprehensive comparisons with state-of-the-art (SOTA) methods, encompassing both task-specific specialist models and multitask generalist models. For a fair comparison, we finetune Grounded 3D-LLM [9], the previous SOTA model on ScanRefer, on our proposed 3D ReasonSeg dataset and assess its performance on the 3D ReasonSeg benchmark." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.884 + ], + "angle": 0, + "content": "As shown in Tab. 2, our baseline model is comparable to existing approaches on both ScanRefer and ScanQA. While our baseline model demonstrates robust performance, the incorporation of 3D ReasonSeg and \\(\\mathbb{R}^2\\mathbb{S}\\) yields consistent substantial improvements across all metrics, validating the effectiveness of our approach. Notably, \\(\\mathbb{R}^2\\mathbb{S}\\) significantly enhances the Acc@50 metric on ScanRefer and the gIoU metric on 3D ReasonSeg, tasks that require robust spatial reasoning, aligning with our expectations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Prior-guided Refinement. To verify the" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.452, + 0.905, + 0.512 + ], + "angle": 0, + "content": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
wo PR58.241.37.766.113.433.227.5
text-based58.443.513.775.515.636.929.8
R2S59.548.713.977.316.438.133.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.542, + 0.522, + 0.878, + 0.536 + ], + "angle": 0, + "content": "Table 3. Ablation study on the Prior-guided Refinement." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.547, + 0.895, + 0.602 + ], + "angle": 0, + "content": "
number of latent queriesScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C↑M↑R↑gloU
1657.240.514.075.315.236.828.9
3258.943.613.375.215.536.729.2
4858.741.614.574.315.136.128.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.613, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Table 4. Ablation study on the number of latent queries. The results are based on our base method." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.652, + 0.905, + 0.716 + ], + "angle": 0, + "content": "
λtxtScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
0.161.343.912.368.214.334.028.9
0.558.943.613.375.215.536.729.2
1.058.741.413.272.615.235.628.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.726, + 0.907, + 0.755 + ], + "angle": 0, + "content": "Table 5. Ablation study on the loss weight configurations, where \\(\\lambda_{txt}\\) means the weight of \\(\\mathcal{L}_{txt}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.907, + 0.838 + ], + "angle": 0, + "content": "necessity of Prior-guided Refinement (PR), we investigate the approach without PR (wo PR) and directly representing target-relevant objects as textual information (text-based). The results are shown in Tab. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "When Prior-guided Refinement is removed, the model simultaneously predict relevant objects and process instructions. Results indicate that the approach without Prior-guided Refinement not only fails to improve performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7857" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.141, + 0.098, + 0.307, + 0.171 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.313, + 0.108, + 0.486, + 0.14 + ], + "angle": 0, + "content": "USER: What is under the desk in front of the doorway? Please answer the question." + }, + { + "type": "image", + "bbox": [ + 0.141, + 0.172, + 0.308, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.314, + 0.151, + 0.474, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.177, + 0.486, + 0.229 + ], + "angle": 0, + "content": "USER: What is under the desk in front of the doorway? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them." + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.233, + 0.455, + 0.244 + ], + "angle": 0, + "content": "ASSISTANT: office chair" + }, + { + "type": "image", + "bbox": [ + 0.14, + 0.255, + 0.309, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.315, + 0.269, + 0.48, + 0.292 + ], + "angle": 0, + "content": "USER: What is in the center of the room? Please answer the question." + }, + { + "type": "image", + "bbox": [ + 0.316, + 0.307, + 0.426, + 0.32 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.334, + 0.483, + 0.387 + ], + "angle": 0, + "content": "USER: What is in the center of the room? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them." + }, + { + "type": "text", + "bbox": [ + 0.316, + 0.394, + 0.466, + 0.406 + ], + "angle": 0, + "content": "ASSISTANT: kitchen island" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.097, + 0.807, + 0.129 + ], + "angle": 0, + "content": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask." + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.129, + 0.664, + 0.14 + ], + "angle": 0, + "content": "ASSISTANT: the towel [SEG]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.14, + 0.808, + 0.246 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.251, + 0.815, + 0.295 + ], + "angle": 0, + "content": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask. The question-related objects are [SEG][SEG][SEG], and you may need to pay attention to them." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.297, + 0.648, + 0.308 + ], + "angle": 0, + "content": "ASSISTANT: the towel [SEG]" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.308, + 0.808, + 0.411 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.3, + 0.431, + 0.696, + 0.446 + ], + "angle": 0, + "content": "Figure 8. Visual qualitative results on ScanQA and 3D ReasonSeg." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.472, + 0.483, + 0.533 + ], + "angle": 0, + "content": "but potentially degrades it. This degradation likely stems from the increased computational complexity of concurrent object prediction and instruction processing without supplementary contextual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.747 + ], + "angle": 0, + "content": "The text-based method represents an ablation variant of \\(\\mathbf{R}^2\\mathbf{S}\\) where segmentation is omitted, providing only text prompts such as \"The question-related objects are chairs, and you may need to pay attention to them\". Although text-based approach demonstrates modest improvements, our method consistently achieves superior results. This performance advantage can be attributed to our approach's ability to extract precise instance-level visual features through segmentation masks. In contrast, text-based method fails to capture spatial information and demonstrates inherent limitations in differentiating multiple instances of identical categories within a scene (e.g., multiple chairs)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.484, + 0.825 + ], + "angle": 0, + "content": "The Number of Latent Queries. We investigate the number of Q-Former latent queries. As demonstrated in Tab. 4, experiments with 16, 32, and 48 latent queries reveal that 32 queries yield optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Loss Weight. We conduct an ablation study on the text loss weight \\(\\lambda_{txt}\\) using values of 0.1, 0.5, and 1.0. Our experiments demonstrate that setting \\(\\lambda_{txt}\\) to 1.0 achieves optimal balance in performance, as evidenced in Tab. 5." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.472, + 0.696, + 0.487 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.907, + 0.614 + ], + "angle": 0, + "content": "As illustrated in Fig. 8, we present a comparison between the baseline and \\(\\mathbf{R}^2\\mathbf{S}\\). The left portion of the figure compares baseline predictions (upper row) with \\(\\mathbf{R}^2\\mathbf{S}\\) predictions (lower row) for ScanQA. The lower row shows segmentation results for target-relevant objects. Similarly, the right portion compares baseline and \\(\\mathbf{R}^2\\mathbf{S}\\) predictions for 3D ReasonSeg. Additional qualitative results are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.628, + 0.642 + ], + "angle": 0, + "content": "5. Limitation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.907, + 0.727 + ], + "angle": 0, + "content": "Due to the use of large language models for data generation instead of human annotation, some noise remains in our dataset even after data cleaning procedures, potentially impacting model performance. We will further optimize our dataset in future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.74, + 0.641, + 0.755 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduce the Relevant Reasoning Segmentation framework, a reasoning-based segmentation framework that enhances spatial reasoning through Reasoning Prior Learning and Prior-guided Refinement. Furthermore, we propose 3D ReasonSeg, a reasoning-based segmentation dataset designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs. We hope our work provides new insights and directions for exploring the spatial reasoning capabilities in 3D MLLMs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7858" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.269 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62372133, 62376074), the National Key R&D Program of China (Grant No. 2024YFE0215300), the Shenzhen Fundamental Research Program (Grant No. JCYJ20220818102415032), the Shenzhen Science and Technology Program (Grant No. JSGGKQTD20221101115655027), the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2025A1515011434), and the Guangdong Provincial Key Laboratory of Mathematical Foundations for Artificial Intelligence (Grant No. 2023B1212010001)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.293, + 0.188, + 0.309 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.318, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.402, + 0.482, + 0.47 + ], + "angle": 0, + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.482, + 0.54 + ], + "angle": 0, + "content": "[3] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.482, + 0.748 + ], + "angle": 0, + "content": "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[7] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[8] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18527-18536, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.318, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[11] Anurag Das, Xinting Hu, Li Jiang, and Bernt Schiele. Mtaclip: Language-guided semantic segmentation with mask-text alignment. arXiv preprint arXiv:2407.21654, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[12] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, and Akhil Mathur et.al. The llama 3 herd of models, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.661 + ], + "angle": 0, + "content": "[17] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[18] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[20] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[22] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[23] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[24] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.317, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.482, + 0.454 + ], + "angle": 0, + "content": "[26] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[30] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[32] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[33] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[34] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[36] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[37] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[38] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.374, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[39] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.445, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[40] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.515, + 0.905, + 0.571 + ], + "angle": 0, + "content": "[41] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.572, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[42] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.614, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[43] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.671, + 0.905, + 0.739 + ], + "angle": 0, + "content": "[44] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.739 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7860" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_origin.pdf b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a25cea118743009ebfb8a46ad3e61da0d2fe92e --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/dda84173-1ef6-4cdd-867f-e3fcb580b499_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d4bbfcc140e7c9d6278ed98f3cd9a06f22535efa7b987cf5187ec0b1c55d2a10 +size 4762262 diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/full.md b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..661c87fb79205ec067b60a68145d57a85814d2f1 --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/full.md @@ -0,0 +1,363 @@ +# Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation + +Zhenhua Ning $^{1,2,3\dagger*}$ Zhuotao Tian $^{1*}$ Shaoshuai Shi $^{4}$ Guangming Lu $^{1}$ Daojing He $^{1}$ Wenjie Pei $^{1,2\boxtimes}$ Li Jiang $^{3\boxtimes}$ $^{1}$ Harbin Institute of Technology, Shenzhen $^{2}$ Pengcheng Laboratory + $^{3}$ The Chinese University of Hong Kong, Shenzhen $^{4}$ Voyager Research, Didi Chuxing wenjiecoder@outlook.com jiangli@cuhk.edu.cn + +# Abstract + +Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation $(R^2 S)$ , a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the $R^2 S$ and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work. + +# 1. Introduction + +In recent years, multi-modal large language models (MLLMs) [1, 25, 34, 37, 38, 43] have managed to connect language and vision, bringing in a new era of artificial intelligence. These models have shown unprecedented abilities, enabling various fields to understand and create textual and visual content, such as human-computer interaction, multimedia understanding, embodied intelligence, and more. + +Building upon these foundations, recent developments in MLLMs have led to the emergence of 3D MLLMs [6, 16, 17, 33, 35], which demonstrate remarkable capabilities in 3D point cloud perception through enhanced vision-language alignment. Despite these achievements, current 3D MLLM-based approaches exhibit limitations in processing instructions that require sophisticated spatial reasoning. As shown in Fig. 1, the baseline model fails to accurately localize the target object. This observation raises an important question: while 3D point clouds inherently contain richer spatial information (e.g., precise measurements of size and position) compared to 2D images, why do models still struggle to effectively leverage these geometric properties? We posit that the development of robust spatial reasoning capabilities requires two critical components: an explicit reasoning mechanism and a dataset that encompasses complex reasoning scenarios. + +In this work, we introduce the Relevant Reasoning Segmentation $(\mathbb{R}^2\mathbb{S})$ , a reasoning-based segmentation framework that enhances the spatial reasoning capabilities of 3D MLLMs through two key steps: 1) Reasoning Prior Learning and 2) Prior-guided Refinement. In the first step, the model is trained to generate spatial priors by identifying target-relevant objects, providing essential contextual information. In the second step, the model leverages these reasoning priors to refine hidden representations, facilitating precise target localization. In other words, the two steps emulate human reasoning behavior, i.e., initially identifying relevant elements broadly and then scrutinizing them closely to pinpoint the target. + +Analysis of existing datasets reveals that instructions requiring reasoning are often oversimplified and ambiguous, as illustrated in Fig. 5. This limitation constrains the training and evaluation processes for developing reasoning capabilities in 3D MLLMs. To address this limitation, we present 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes, + +![](images/d4fd67059117f7dafb1e80a13a50e9aedfcd6caacc2b1e29fcbe9f768e2b88f3.jpg) +Figure 1. In contrast to the baseline model in Sec. 3.1, we introduce Relevant Reasoning Segmentation to explicitly guide the model's reasoning by emphasizing target-relevant objects. In this instance, the object identified as the nearest resting place to the open door should be the chair marked by the red rectangle. The first to third columns in the image represent the prediction from the base method, the outcomes of reasoning prior learning, and the final prediction following prior guided refinement. + +![](images/d9366d63eda1eaa8336f1d160f7a54f60dda448b913bb9d9d87fedb6a0bc8590.jpg) + +and spatial relationships. The dataset comprises 25,185 training samples and 3,966 validation samples, with quality assurance maintained through manual verification processes. This dataset facilitates enhancement and comprehensive assessment of spatial reasoning capabilities. + +Both our proposed $\mathrm{R}^2\mathrm{S}$ framework and the 3D ReasonSeg dataset contribute to substantial performance improvements on established benchmarks, demonstrating their complementary effects in enhancing spatial reasoning capabilities. In summary, our main contributions are: + +- We propose the Relevant Segmentation $(\mathbb{R}^2\mathbb{S})$ , a reasoning-based segmentation framework that enhances spatial reasoning capabilities of 3D MLLMs by emulating human reasoning processes through two key components: Reasoning Prior Learning and Prior-guided Refinement. +- We propose 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes, and spatial relationships, designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs. +- Extensive experimental results demonstrate that our $\mathrm{R}^2\mathrm{S}$ framework, combined with the 3D ReasonSeg dataset, achieves significant improvements in spatial reasoning across multiple benchmarks. + +# 2. Related Work + +# 2.1. 3D Scene Understanding + +3D scene understanding requires models to detect objects and comprehend their characteristics in 3D space [3, 7, 8, 22, 29, 31]. Among these tasks, 3D segmentation presents unique challenges as it demands point-wise classification within point clouds. Recent studies demonstrate + +that language integration enhances model performance, as evidenced in 3D Question Answering and 3D Referring tasks [2, 4, 19, 28, 40]. These language-guided tasks, which require models to respond to instructions or identify objects based on descriptions, underscore the importance of language comprehension and spatial reasoning capabilities. + +The integration of language has supported models in transitioning from closed-set to open-set scenarios, such as in 3D Open Vocabulary Segmentation [4, 11, 12, 30, 32]. This step is critical as it allows the model to break free from being constrained by the finite set of categories in the training dataset. While most task-oriented studies focus on specific objectives, recent advancements in 3D multi-modal large language models have enabled addressing multiple 3D downstream tasks using these models [6, 18]. Furthermore, the assistance of 3D multi-modal large language models has empowered models to tackle more intricate object localization tasks [9, 42]. + +These prior studies have not adequately addressed the complex spatial reasoning capabilities required for comprehensive understanding of 3D scenes. In this work, we introduce Relevant Reasoning Segmentation and 3D ReasonSeg to address complex spatial reasoning challenges. + +# 2.2. 3D Multi-modal Large Language Model + +Significant progress has been made in the realm of 3D multi-modal large language models (3D MLLMs). Initial endeavors have predominantly concentrated on point clouds at the object level [14, 15, 26, 35], posing challenges in tackling complex tasks at the scene level. Recently, researchers have shifted attention to understanding entire scene point clouds [6, 16, 17, 33], significantly advancing 3D scene understanding. + +Drawing from these advances, we present an efficient + +framework that handles various scene understanding tasks, with particular strength in segmentation. While existing 3D MLLMs can process basic spatial relationships, they struggle with complex spatial reasoning tasks. Our model, however, demonstrates superior performance in handling sophisticated spatial reasoning challenges. + +# 3. Method + +Efforts such as those outlined in [6, 16, 17] were dedicated to leveraging 3D MLLMs for a deeper comprehension of 3D point clouds. Nevertheless, it is evident that current methods encounter challenges in effectively managing intricate spatial relationships. Specifically, as illustrated in Fig. 2, while the model can accurately identify "What is the object that is used for sitting?", it struggles with queries such as "What is the object that is used for sitting and located near the desk with a monitor on it?" + +To tackle this issue, in this work, we instruct the model to generate reasoning priors by paying attention to objects relevant to the target. Subsequently, the model examines the scene and these relevant objects more closely to locate the final target. We refer to this process as Relevant Reasoning Segmentation $(\mathbb{R}^2\mathbb{S})$ + +The remainder of this section is organized as follows: Sec. 3.1 introduces our baseline model, which demonstrates strong dense perception capabilities on 3D point cloud data. Sec. 3.2 elaborates on our proposed Relevant Reasoning Segmentation framework. Sec. 3.3 details the construction of our 3D ReasonSeg dataset, and Sec. 3.4 describes the comprehensive training pipeline. + +# 3.1. Baseline Model + +Our baseline architecture consists of a multi-modal large language model (MLLM) and a Segmentation Head. Following [23], we extend the MLLM's vocabulary with a special token [SEG] that instructs the Segmentation Head to generate corresponding segmentation mask. This design enables the MLLM to effectively perform fine-grained segmentation tasks while maintaining its core language understanding capabilities. + +Multi-modal Large Language Model. There are three key components in the adopted MLLM for feature extraction: a visual encoder, a Q-Former, and a large language model (LLM). Initially, the point cloud $X_{p}$ is processed through the visual encoder to extract dense features. To optimize computational efficiency, we employ super-point [24] within the visual backbone, aggregating the dense visual features into super-point features $f_{p}$ to reduce the following computational burden. Subsequently, the Q-Former [25] compresses the scene's information into several latent queries $q_{l}$ based on the super-point features $f_{p}$ and text embedding $w$ of the instruction. After that, the latent queries $\hat{q}_{l}$ will serve as the visual input, along with + +the text embedding $\mathbf{w}$ , to be processed by the large language model $\mathcal{F}_{LLM}$ . This process is formulated as: + +$$ +\pmb {f} _ {p} = \mathcal {F} _ {e n c} (\pmb {X} _ {p}), \quad \hat {\pmb {q}} _ {l} = \mathcal {F} _ {Q} (\pmb {q} _ {l}, \pmb {w}, \pmb {f} _ {p}). \qquad (1) +$$ + +$$ +\boldsymbol {y} _ {t x t} = \mathcal {F} _ {L L M} (\hat {\boldsymbol {q}} _ {l}, \boldsymbol {w}). \tag {2} +$$ + +Inside the Q-Former $\mathcal{F}_Q$ , $\pmb{q}_l$ is concatenated with $\pmb{w}$ to be the query in both the self-attention and cross-attention layers of $\mathcal{F}_Q$ , while $\pmb{f}_p$ serves as the key and value in the cross-attention layer: + +$$ +\boldsymbol {q} _ {l}, \boldsymbol {w} = \text {S e l f A t t n} (\text {C o n c a t} (\boldsymbol {q} _ {l}, \boldsymbol {w})), \tag {3} +$$ + +$$ +\boldsymbol {q} _ {l}, \boldsymbol {w} = \operatorname {C r o s s A t t n} (\operatorname {C o n c a t} (\boldsymbol {q} _ {l}, \boldsymbol {w}), \boldsymbol {f} _ {p}, \boldsymbol {f} _ {p}). \tag {4} +$$ + +Segmentation Head. If the output of LLM, i.e., $\mathbf{y}_{\text{txt}}$ , contains [SEG] tokens, it means that the current query requires segmentation predictions. Then, the corresponding hidden features of the [SEG] token, i.e., $\mathbf{h}_{seg}$ , are processed by the segmentation head $\mathcal{F}_{dec}$ that follows the decoder structure of [21] to yield the requisite segmentation mask $M$ , based on the reasoning between $\mathbf{h}_{seg}$ and $\mathbf{f}_p$ . The process can be formulated as follows. + +$$ +\boldsymbol {M} = \mathcal {F} _ {\text {d e c}} \left(\boldsymbol {h} _ {\text {s e g}}, \boldsymbol {f} _ {p}\right). \tag {5} +$$ + +The aforementioned framework seamlessly incorporates dense visual perception abilities into the output of the language model, allowing for the fusion of linguistic and visual information processing. + +# 3.2. Relevant Reasoning Segmentation + +The baseline model can decently handle point cloud perception tasks. However, in our observations, the direct instruction tuning performed in [23] did not consider the spatial relationships among objects in the scene, hindering complex spatial reasoning, as illustrated in Fig. 8. Hence, to address this issue, we suggest that the model initially identifies the objects related to the target and their spatial relationships, serving as the reasoning prior to guide the target localization. This process is thus termed as Relevant Reasoning Segmentation $(\mathrm{R}^2\mathrm{S})$ , and it contains two steps as follows. + +Step 1: Reasoning Prior Learning. While the model may encounter challenges in understanding intricate interobject relationships, it can recognize objects relevant to the target. As shown in Fig. 2, despite the model's limitation in accurately segmenting the chair beside the desk supporting a monitor, it demonstrates the ability to detect desks and chairs. Given this capability, we may instruct the model to initially segment approximate question-related objects, thus establishing the visual reasoning prior as the hint to help identify the final target. + +Following the baseline model's process, we first use an instruction structured as: "Given the 3D + +![](images/95a33270b54c03ace6259e2db950f676b97fd22a9df65e3af635b0ce522fea87.jpg) +[QUESTION]: What is the object that is used for sitting? + +![](images/082b1e41b76fab545242a15a66576dfb70f3631be0bfc67072883afb742b74fe.jpg) +[QUESTION]: What is the object that is used for sitting and located near the desk with a monitor on it? + +![](images/7bbc94e07301c79120a30b4d64e4c74fb0f544b9d6bb6600d46210f52cfa6a37.jpg) +Figure 2. The image on the left depicts the LLM's response to the question "What is the object that is used for sitting?" Conversely, the right side of the image illustrates the LLM's response to the question "What is the object that is used for sitting and located near the desk with a monitor on it." It is noteworthy that the LLM provides an incorrect answer in the latter case. The accurate response should identify the chair within the red rectangle. +Figure 3. The baseline model consists of a visual backbone for point feature encoding, a Q-Former that connects visual-language representations, a LLM for instruction processing, and a mask decoder that converts [SEG] tokens into segmentation masks. + +scene, [QUESTION]. Please segment the question-related objects that may help answer the question." *. The corresponding instruction embedding $\mathbf{w}$ will be processed by Eq. (1) and Eq. (2) to obtain the text output $\mathbf{y}_{txt}$ from which we can obtain the [SEG] tokens representing these target-relevant objects. Subsequently, the segmentation head decodes the hidden features $\mathbf{h}_{seg}$ of [SEG] tokens into masks $M_r$ for these target-relevant objects. + +Next, we adopt mask-pooling to aggregate the superpoint features $\pmb{f}_p$ with masks $M_r$ , yielding the representations $\pmb{f}_r$ for the target-relevant objects. The features $\pmb{f}_r$ serve as the reasoning priors in the subsequent step. The above process can be formally expressed as: + +$$ +\boldsymbol {M} _ {r} = \mathcal {F} _ {\text {d e c}} \left(\boldsymbol {f} _ {p}, \boldsymbol {h} _ {\text {s e g}}\right), \tag {6} +$$ + +$$ +\boldsymbol {f} _ {r} = \boldsymbol {f} _ {p} \times \boldsymbol {M} _ {r} \tag {7} +$$ + +Step 2: Prior-guided Refinement. In the absence of explicit instructions, the latent queries $\hat{\pmb{q}}_l$ derived from Eq. (1) may exhibit a bias towards irrelevant areas, potentially impacting the reasoning process in Eq. (2). Besides, the mask-pooling operation in Eq. (7) may cause the loss of essential spatial details. Therefore, we propose the prior-guided refinement which commences by feeding the initial latent queries $\pmb{q}_l$ , text outputs $y_{txt}$ and target-relevant features $\pmb{f}_r$ from step 1 into the Q-Former $\mathcal{F}_Q$ to yield the refined query $\pmb{q}_l'$ and target-relevant features $\pmb{f}_r'$ . + +To make use of the obtained reasoning pri- ors, we employ a revised instruction like: "Given the 3D scene, the question is [QUESTION]. Please answer the question and output segmentation mask. The question-related objects are $\{f_r^1,f_r^2$ $\dots ,f_r^n\}$ , and you may need to pay attention to them" $^\dagger$ where the target-relevant + +![](images/0d1dc52e37d91fdca0acf1369377737c6d5e2c3fa49cd61a6e7b462ad655c2bb.jpg) +Figure 4. The proposed Relevant Reasoning Segmentation framework. Stage 1 segments question-related objects as priors $(f_1', \dots, f_6')$ . Stage 2 refines hidden representations of instructions based on these priors, enabling the LLM to generate more accurate responses. + +features $\pmb{f}_r$ of $n$ potentially relevant objects have been inserted to form the new instruction whose text embedding is therefore denoted as $\pmb{w}' \oplus \pmb{f}_r$ . To this end, the prior-guided refinement regarding the latent query and the target-relevant features can be formulated as: + +$$ +\boldsymbol {q} _ {l} ^ {\prime}, \boldsymbol {f} _ {r} ^ {\prime} = \mathcal {F} _ {Q} \left(\boldsymbol {q} _ {l}, \boldsymbol {w} ^ {\prime} \oplus \boldsymbol {f} _ {r}, \boldsymbol {f} _ {p}\right) \tag {8} +$$ + +where $\pmb{w}^{\prime}$ and $\pmb{f}_{p}$ are the text embeddings and super-point features, respectively. Unlike Eq. (3), during the prior-guided refinement, $\pmb{q}_{l}$ is concatenated with the instruction embedding $\pmb{w}^{\prime} \oplus \pmb{f}_{r}$ to serve as the queries for both self-attention and cross-attention layers in the Q-Former $\mathcal{F}_Q$ . Meanwhile, $\pmb{f}_{p}$ continues to function as the key and value in the cross-attention layer. In Eq. (8), the enhanced latent query $\pmb{q}_{l}^{\prime}$ has been enriched with target-related clues from $\pmb{f}_{r}$ . Simultaneously, utilizing the point-cloud features $\pmb{f}_{p}$ , the target-relevant features $\pmb{f}_{r}$ are refined with essential spatial intricacies to produce $\pmb{f}_{r}^{\prime}$ . Next, the refined query $\pmb{q}_{l}^{\prime}$ and target-relevant features $\pmb{f}_{r}^{\prime}$ are adopted for generating the new text output $\pmb{y}_{txt}^{\prime}$ with the LLM $\mathcal{F}_{LLM}$ : + +$$ +\boldsymbol {y} _ {t x t} ^ {\prime} = \mathcal {F} _ {L L M} \left(\boldsymbol {q} _ {l} ^ {\prime}, \boldsymbol {w} ^ {\prime} \oplus \boldsymbol {f} _ {r} ^ {\prime}\right), \tag {9} +$$ + +where the instruction embedding $\pmb{w}^{\prime}\oplus \pmb{f}_{r}$ originally used in Eq. (8) has been accordingly updated to $\pmb{w}^{\prime}\oplus \pmb{f}_{r}^{\prime}$ . The hidden features $h_{seg}^{\prime}$ of the [SEG] tokens are extracted from the new text output $\pmb{y}_{txt}^{\prime}$ and processed by the decoder $\mathcal{F}_{dec}$ to generate the refined final mask prediction $M^{\prime}$ : + +$$ +\boldsymbol {M} ^ {\prime} = \mathcal {F} _ {d e c} \left(\boldsymbol {h} _ {s e g} ^ {\prime}, \boldsymbol {f} _ {p}\right) \tag {10} +$$ + +# 3.3. 3D ReasonSeg + +Analysis of existing datasets, as demonstrated in Fig. 5, reveals that spatial relations such as "in front of" often lack complexity and exhibit potential ambiguity. Using such datasets for training and evaluation may inadequately address the spatial reasoning requirements of 3D MLLMs. To facilitate enhancement and assessment of spatial reasoning capabilities in 3D MLLMs, we present 3D ReasonSeg, a + +![](images/5de7a2200b20097d670dec606552502345bac8b4ac36a4b61ff7b51653e3f773.jpg) +Figure 5. This illustrates a data sample from ScanRefer, where the target object is described as "A brown chair placed in front of a plant." The spatial relationship in this description shows ambiguity, as there are multiple chairs in front of the plant. + +dataset specifically designed for reasoning tasks in 3D point cloud environments. Empirical results in Tab. 2 indicate that models trained on 3D ReasonSeg achieve substantial performance improvements over baseline implementations. + +Specifically, the proposed 3D ReasonSeg requires the model to identify the target object based on a question that does not explicitly mention it. The model needs to understand the object functions, visual attributes, and spatial relationships of the objects described in the query and reason to segment the target object. To implement this task, we use a template structured as: "USER: Given the 3D scene, please answer the question and output segmentation mask: [QUESTION]. ASSISTANT: [ANSWER]". + +We construct the 3D ReasonSeg dataset by integrating SceneVerse's [20] detailed object descriptions with LLama 3.1 [13]. Based on objects' 3D spatial information (positions and sizes) from point clouds and the object descriptions from SceneVerse, LLama 3.1 generates question-reasoning-answer pairs for reasoning tasks. + +After collecting numerous question-reasoning-answer pairs, we apply rule-based filtering (e.g., calculating distances between objects and removing outliers) to eliminate instances with unclear or incorrect spatial relationships. The resulting 3D ReasonSeg dataset contains 29,151 high-quality data samples. Examples are shown in Fig. 6, with details provided in the supplementary material. + +We conduct a statistical analysis of ScanQA, ScanRe + +fer, and 3D ReasonSeg datasets, examining their training (#Train) and validation (#Val) set sizes, average word counts per sample (Avg. Words), and average number of relevant objects (Avg. Objects). The detailed statistics are summarized in Tab. 1, showing that our proposed 3D ReasonSeg dataset contains more complex object relationships. + +
Dataset#Train#ValAvg. WordsAvg. Objects
ScanQA26,5634,6758.81.5
ScanRefer36,6659,50817.81.8
3D ReasonSeg25,1853,96619.65.4
+ +Table 1. Comparative statistics across ScanQA, ScanRefer, and 3D ReasonSeg datasets. + +[QUESTION]: What is the nearest resting place to the bicycle? + +[ANSWER]: The couch [32]. + +![](images/36b436c34ab671c1bf579ca79358f853bb7617d314655ec72a7c58a3465fd107.jpg) + +[QUESTION]: What is the object that is used for storing books and is located near the chair and the door? + +[ANSWER]: The shelf [6]. + +![](images/3a31427c8e0700260019421d7f59901ed18076f43c7cbe19017abcacc322a985.jpg) +Figure 6. Visualization of 3D ReasonSeg samples, where instance ID colors correspond to their respective object colors in the image. + +# 3.4. Model Training + +Training Datasets. Following [6], our training data includes established 3D datasets such as ScanQA [2], ScanRefer [4], ScanNet200 [10], the ScanNet subset of 3D-LLM [16], and SceneVerse [20], along with our proposed 3D ReasonSeg dataset. + +For Relevant Reasoning Segmentation outlined in Sec. 3.2, we extend the ScanQA, Scanrefer, and 3D ReasonSeg to incorporate target-relevant object identification. This process leverages scene object annotations and question-answer pairs from the original datasets, with LLama 3.1 [13] serving as the backbone for target-relevant object identification. Representative examples are presented in Fig. 7, with detailed prompts and implementation procedures in the supplementary material. + +Loss Function. During training, the model is optimized using the cross-entropy loss $\mathcal{L}_{\text {txt }}$ for text generation, along with the binary cross-entropy (BCE) and DICE losses ( $\mathcal{L}_{\text {bce }}$ and $\mathcal{L}_{\text {dice }}$ ) for segmentation. Following [22], we assign equal weights of 1 to the BCE and DICE loss. Thus, the overall training objective $\mathcal{L}$ can be expressed as: + +$$ +\mathcal {L} = \lambda_ {t x t} \mathcal {L} _ {t x t} + \mathcal {L} _ {b c e} + \mathcal {L} _ {d i c e} \tag {11} +$$ + +where $\lambda_{txt}$ is the weight for text generation loss. + +Relevant Objects Augmentation. Our relevant segmentation leverages mask predictions of relevant objects as priors to enhance latent query and text embedding. However, direct implementation leads to overfitting during training, as the ground truth relevant objects remain fixed during training but vary during inference, creating a train-test discrepancy. To address this issue, we implement random omission and addition of relevant objects during training, simulating diverse inference scenarios with incomplete priors. + +[QUESTION]: How many pictures are on the wall? + +[ANSWER]: 4 + +Relevant Objects: picture [23][24][25][26] + +![](images/7dfdb1ef97733c6182b81dda58a2a35e0891a3460b818bcf537dd615893ed5da.jpg) + +[DESCRIPTION]: this is the black kitchen counter with the sink in its. above that kitchen counter is a paper towel dispenser. + +Relevant Objects: kitchen counter [1], sink [11] and paper towel dispenser [15] + +![](images/f147fe0b62d136125a486422cfd1030cc48ee882a79343a9b65c629f2ce0afc9.jpg) +Figure 7. Examples of target-relevant data. The first row shows an instance from ScanQA, and the second row is related to ScanRefer. The visualizations of target-relevant objects are displayed on the right side of each row. + +# 4. Experiments + +# 4.1. Experimental Setting + +Implementation Details. We adopt the encoder of a pretrained Oneformer3D as visual backbone and the decoder of it as the mask decoder, and a pre-trained OPT 1.3B [39] as the language model. To preserve the learned knowledge of the pre-trained language model, we completely freeze the parameters of the language model while fully training the parameters of all other modules. Note that, we found it is necessary to also fine-tune the visual backbone. We adopt the AdamW [27] optimizer with a weight decay of 0.1 and a learning rate decaying from 1e-4 to 1e-6 with a cosine annealing scheduler. The text generation loss weight $\lambda_{txt}$ is set to 0.5. The number of learnable queries in Q-Former is set to 32. We randomly sample 300k points from each 3D scene as the 3D input. All the experiments are conducted on eight Nvidia H20 GPUs. + +Benchmark and Metrics. In this study, we assess our methods on ScanQA [2], ScanRefer [4], and our proposed 3D ReasonSeg. ScanQA is a dataset for the 3D question answering task, evaluating visual and spatial comprehension of 3D environments. The evaluation metrics of ScanQA include BLEU-4, CIDEr, METEOR, and Rouge-L. ScanRefer evaluates the capability to localize 3D object using a natural + +
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C ↑M ↑R ↑gIoU
ScanRefer [4]37.324.3-----
MVT [19]40.833.3-----
3DVG-Trans [41]45.934.5-----
ViL3DRel [5]47.937.7-----
M3DRef-CLIP [40]51.944.7-----
ScanQA [2]--10.164.913.133.3-
3D-VisTA [44]50.645.813.172.9---
LLM-Grounder [36]17.15.3-----
3D-LLM [16]30.3-12.069.414.535.7-
Chat-3D [33]--6.453.211.928.5-
Chat-3D v2 [17]35.930.47.377.116.140.1-
LL3DA [6]--13.576.815.937.3-
ScanReason [42]53.141.1-----
Grounded 3D-LLM [9]47.944.113.472.7--25.6
baseline54.638.712.870.513.834.6-
+ 3D ReasonSeg58.943.613.375.215.536.729.2
+ R2S59.548.713.977.316.438.133.1
+ +language description. Prediction accuracy is measured by the percentage of predictions with Intersection over Union (IoU) above 0.25 and 0.5 compared to ground truth masks. For 3D ReasonSeg, the model is required to focus on object attributes and relationships to reason, answering questions and localizing target objects. Following [23], the general Intersection over Union (gIoU) metric is used, defined as the average of all per-scene Intersection-over-Unions (IoUs). Detailed formulations of these tasks can be found in the supplementary material. + +# 4.2. Main Results + +We evaluate our model's effectiveness through comprehensive comparisons with state-of-the-art (SOTA) methods, encompassing both task-specific specialist models and multitask generalist models. For a fair comparison, we finetune Grounded 3D-LLM [9], the previous SOTA model on ScanRefer, on our proposed 3D ReasonSeg dataset and assess its performance on the 3D ReasonSeg benchmark. + +As shown in Tab. 2, our baseline model is comparable to existing approaches on both ScanRefer and ScanQA. While our baseline model demonstrates robust performance, the incorporation of 3D ReasonSeg and $\mathbb{R}^2\mathbb{S}$ yields consistent substantial improvements across all metrics, validating the effectiveness of our approach. Notably, $\mathbb{R}^2\mathbb{S}$ significantly enhances the Acc@50 metric on ScanRefer and the gIoU metric on 3D ReasonSeg, tasks that require robust spatial reasoning, aligning with our expectations. + +Effectiveness of Prior-guided Refinement. To verify the + +Table 2. Results among our methods and previous related works. Entries in gray denote models specialized for specific datasets. + +
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
wo PR58.241.37.766.113.433.227.5
text-based58.443.513.775.515.636.929.8
R2S59.548.713.977.316.438.133.1
+ +Table 3. Ablation study on the Prior-guided Refinement. + +
number of latent queriesScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C↑M↑R↑gloU
1657.240.514.075.315.236.828.9
3258.943.613.375.215.536.729.2
4858.741.614.574.315.136.128.4
+ +Table 4. Ablation study on the number of latent queries. The results are based on our base method. + +
λtxtScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
0.161.343.912.368.214.334.028.9
0.558.943.613.375.215.536.729.2
1.058.741.413.272.615.235.628.6
+ +Table 5. Ablation study on the loss weight configurations, where $\lambda_{txt}$ means the weight of $\mathcal{L}_{txt}$ . + +necessity of Prior-guided Refinement (PR), we investigate the approach without PR (wo PR) and directly representing target-relevant objects as textual information (text-based). The results are shown in Tab. 3. + +When Prior-guided Refinement is removed, the model simultaneously predict relevant objects and process instructions. Results indicate that the approach without Prior-guided Refinement not only fails to improve performance + +![](images/74f26e80c719c52dcc76bd81e6a495e31cfde3bea504bf008023514cd70a69f7.jpg) + +USER: What is under the desk in front of the doorway? Please answer the question. + +![](images/e55245836a1e8f3e32d6354542a37206f14af2071e6e8915a34906b13ffbbf5d.jpg) + +![](images/2137490cb9d3d511307f8439be23c0a346ddb12e6cab2b4c042c2b8d762f030b.jpg) + +USER: What is under the desk in front of the doorway? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them. + +ASSISTANT: office chair + +![](images/7b0424f19987811f88c8f53912f87c5240be412a4c15184aaf8eb6803a1c8c14.jpg) + +USER: What is in the center of the room? Please answer the question. + +![](images/72f9be3e22b9dddc56d7bc895567e67f5cb557720cdae4f6bb5486aa93086ccf.jpg) + +USER: What is in the center of the room? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them. + +ASSISTANT: kitchen island + +USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask. + +ASSISTANT: the towel [SEG] + +![](images/dcd364a6105a3588229286a0f8dc234dd69a9a9479c261a25b0b62eda5252206.jpg) + +USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask. The question-related objects are [SEG][SEG][SEG], and you may need to pay attention to them. + +ASSISTANT: the towel [SEG] + +![](images/ed56bd369f5edc2548b1494cccad8d39a27307da0db61915f174b6e218cc4e9f.jpg) +Figure 8. Visual qualitative results on ScanQA and 3D ReasonSeg. + +but potentially degrades it. This degradation likely stems from the increased computational complexity of concurrent object prediction and instruction processing without supplementary contextual information. + +The text-based method represents an ablation variant of $\mathbf{R}^2\mathbf{S}$ where segmentation is omitted, providing only text prompts such as "The question-related objects are chairs, and you may need to pay attention to them". Although text-based approach demonstrates modest improvements, our method consistently achieves superior results. This performance advantage can be attributed to our approach's ability to extract precise instance-level visual features through segmentation masks. In contrast, text-based method fails to capture spatial information and demonstrates inherent limitations in differentiating multiple instances of identical categories within a scene (e.g., multiple chairs). + +The Number of Latent Queries. We investigate the number of Q-Former latent queries. As demonstrated in Tab. 4, experiments with 16, 32, and 48 latent queries reveal that 32 queries yield optimal performance. + +Loss Weight. We conduct an ablation study on the text loss weight $\lambda_{txt}$ using values of 0.1, 0.5, and 1.0. Our experiments demonstrate that setting $\lambda_{txt}$ to 1.0 achieves optimal balance in performance, as evidenced in Tab. 5. + +# 4.3. Qualitative Results + +As illustrated in Fig. 8, we present a comparison between the baseline and $\mathbf{R}^2\mathbf{S}$ . The left portion of the figure compares baseline predictions (upper row) with $\mathbf{R}^2\mathbf{S}$ predictions (lower row) for ScanQA. The lower row shows segmentation results for target-relevant objects. Similarly, the right portion compares baseline and $\mathbf{R}^2\mathbf{S}$ predictions for 3D ReasonSeg. Additional qualitative results are provided in the supplementary material. + +# 5. Limitation + +Due to the use of large language models for data generation instead of human annotation, some noise remains in our dataset even after data cleaning procedures, potentially impacting model performance. We will further optimize our dataset in future work. + +# 6. Conclusions + +In this paper, we introduce the Relevant Reasoning Segmentation framework, a reasoning-based segmentation framework that enhances spatial reasoning through Reasoning Prior Learning and Prior-guided Refinement. Furthermore, we propose 3D ReasonSeg, a reasoning-based segmentation dataset designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs. We hope our work provides new insights and directions for exploring the spatial reasoning capabilities in 3D MLLMs. + +# Acknowledgments + +This work was supported by the National Natural Science Foundation of China (Grant Nos. 62372133, 62376074), the National Key R&D Program of China (Grant No. 2024YFE0215300), the Shenzhen Fundamental Research Program (Grant No. JCYJ20220818102415032), the Shenzhen Science and Technology Program (Grant No. JSGGKQTD20221101115655027), the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2025A1515011434), and the Guangdong Provincial Key Laboratory of Mathematical Foundations for Artificial Intelligence (Grant No. 2023B1212010001). + +# References + +[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1 +[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 6, 7 +[3] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2021. 2 +[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 6, 7 +[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 7 +[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 1, 2, 3, 6, 7 +[7] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2 +[8] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18527-18536, 2023. 2 + +[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 2, 7 +[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 6 +[11] Anurag Das, Xinting Hu, Li Jiang, and Bernt Schiele. Mtaclip: Language-guided semantic segmentation with mask-text alignment. arXiv preprint arXiv:2407.21654, 2024. 2 +[12] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2 +[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, and Akhil Mathur et.al. The llama 3 herd of models, 2024. 5, 6 +[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2 +[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2 +[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 2, 3, 6, 7 +[17] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 1, 2, 3, 7 +[18] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2 +[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 2, 7 +[20] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 5, 6 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and + +Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3 +[22] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 2, 6 +[23] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 3, 7 +[24] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 3 +[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 3 +[26] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 2 +[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2 +[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 2 +[30] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2 +[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2 +[32] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2 +[33] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 2, 7 +[34] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 1 + +[35] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 1, 2 +[36] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 7 +[37] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 1 +[38] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1 +[39] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6 +[40] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 2, 7 +[41] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 7 +[42] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 2, 7 +[43] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1 +[44] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 7 \ No newline at end of file diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/images.zip b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0247a5366fd07b8294335853bac0590d58b5fdd9 --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:39324b7578e5c0ea41dead0ed6e915ca759e6566e9a603e3e76067d45b69a55f +size 619372 diff --git a/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/layout.json b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..f364e233893a1214fa43f773d09f93e44de993f3 --- /dev/null +++ b/2025/Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation/layout.json @@ -0,0 +1,9384 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "spans": [ + { + "bbox": [ + 66, + 102, + 545, + 140 + ], + "type": "text", + "content": "Enhancing Spatial Reasoning in Multimodal Large Language Models through Reasoning-based Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "spans": [ + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": "Zhenhua Ning" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1,2,3\\dagger*}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Zhuotao Tian" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Shaoshuai Shi" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Guangming Lu" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Daojing He" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Wenjie Pei" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1,2\\boxtimes}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": " Li Jiang" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{3\\boxtimes}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": "Harbin Institute of Technology, Shenzhen " + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": "Pengcheng Laboratory \n" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": "The Chinese University of Hong Kong, Shenzhen " + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 83, + 159, + 525, + 232 + ], + "type": "text", + "content": "Voyager Research, Didi Chuxing wenjiecoder@outlook.com jiangli@cuhk.edu.cn" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 272 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "text", + "content": "Recent advances in point cloud perception have demonstrated remarkable progress in scene understanding through vision-language alignment leveraging large language models (LLMs). However, existing methods may still encounter challenges in handling complex instructions that require accurate spatial reasoning, even if the 3D point cloud data provides detailed spatial cues such as size and position for identifying the targets. To tackle this issue, we propose Relevant Reasoning Segmentation " + }, + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "inline_equation", + "content": "(R^2 S)" + }, + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "text", + "content": ", a reasoning-based segmentation framework. The framework emulates human cognitive processes by decomposing spatial reasoning into two sequential stages: first identifying relevant elements, then processing instructions guided by their associated visual priors. Furthermore, acknowledging the inadequacy of existing datasets in complex reasoning tasks, we introduce 3D ReasonSeg, a reasoning-based segmentation dataset comprising 25,185 training samples and 3,966 validation samples with precise annotations. Both quantitative and qualitative experiments demonstrate that the " + }, + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "inline_equation", + "content": "R^2 S" + }, + { + "bbox": [ + 55, + 283, + 297, + 559 + ], + "type": "text", + "content": " and 3D ReasonSeg effectively endow 3D point cloud perception with stronger spatial reasoning capabilities, and we hope that they can serve as a new baseline and benchmark for future work." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 135, + 594 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 296, + 685 + ], + "type": "text", + "content": "In recent years, multi-modal large language models (MLLMs) [1, 25, 34, 37, 38, 43] have managed to connect language and vision, bringing in a new era of artificial intelligence. These models have shown unprecedented abilities, enabling various fields to understand and create textual and visual content, such as human-computer interaction, multimedia understanding, embodied intelligence, and more." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 260, + 555, + 463 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 463 + ], + "type": "text", + "content": "Building upon these foundations, recent developments in MLLMs have led to the emergence of 3D MLLMs [6, 16, 17, 33, 35], which demonstrate remarkable capabilities in 3D point cloud perception through enhanced vision-language alignment. Despite these achievements, current 3D MLLM-based approaches exhibit limitations in processing instructions that require sophisticated spatial reasoning. As shown in Fig. 1, the baseline model fails to accurately localize the target object. This observation raises an important question: while 3D point clouds inherently contain richer spatial information (e.g., precise measurements of size and position) compared to 2D images, why do models still struggle to effectively leverage these geometric properties? We posit that the development of robust spatial reasoning capabilities requires two critical components: an explicit reasoning mechanism and a dataset that encompasses complex reasoning scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "text", + "content": "In this work, we introduce the Relevant Reasoning Segmentation " + }, + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2\\mathbb{S})" + }, + { + "bbox": [ + 313, + 468, + 556, + 624 + ], + "type": "text", + "content": ", a reasoning-based segmentation framework that enhances the spatial reasoning capabilities of 3D MLLMs through two key steps: 1) Reasoning Prior Learning and 2) Prior-guided Refinement. In the first step, the model is trained to generate spatial priors by identifying target-relevant objects, providing essential contextual information. In the second step, the model leverages these reasoning priors to refine hidden representations, facilitating precise target localization. In other words, the two steps emulate human reasoning behavior, i.e., initially identifying relevant elements broadly and then scrutinizing them closely to pinpoint the target." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "Analysis of existing datasets reveals that instructions requiring reasoning are often oversimplified and ambiguous, as illustrated in Fig. 5. This limitation constrains the training and evaluation processes for developing reasoning capabilities in 3D MLLMs. To address this limitation, we present 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes," + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 259, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 259, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 259, + 713 + ], + "type": "text", + "content": "Research Assistant at CUHKSZ. *Equal contribution to this work. Corresponding author." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7851" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 69, + 226, + 223 + ], + "blocks": [ + { + "bbox": [ + 70, + 69, + 226, + 223 + ], + "lines": [ + { + "bbox": [ + 70, + 69, + 226, + 223 + ], + "spans": [ + { + "bbox": [ + 70, + 69, + 226, + 223 + ], + "type": "image", + "image_path": "d4fd67059117f7dafb1e80a13a50e9aedfcd6caacc2b1e29fcbe9f768e2b88f3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 232, + 555, + 277 + ], + "type": "text", + "content": "Figure 1. In contrast to the baseline model in Sec. 3.1, we introduce Relevant Reasoning Segmentation to explicitly guide the model's reasoning by emphasizing target-relevant objects. In this instance, the object identified as the nearest resting place to the open door should be the chair marked by the red rectangle. The first to third columns in the image represent the prediction from the base method, the outcomes of reasoning prior learning, and the final prediction following prior guided refinement." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 69, + 541, + 223 + ], + "blocks": [ + { + "bbox": [ + 238, + 69, + 541, + 223 + ], + "lines": [ + { + "bbox": [ + 238, + 69, + 541, + 223 + ], + "spans": [ + { + "bbox": [ + 238, + 69, + 541, + 223 + ], + "type": "image", + "image_path": "d9366d63eda1eaa8336f1d160f7a54f60dda448b913bb9d9d87fedb6a0bc8590.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 297, + 294, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 297, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 294, + 357 + ], + "type": "text", + "content": "and spatial relationships. The dataset comprises 25,185 training samples and 3,966 validation samples, with quality assurance maintained through manual verification processes. This dataset facilitates enhancement and comprehensive assessment of spatial reasoning capabilities." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 357, + 295, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 357, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 295, + 418 + ], + "type": "text", + "content": "Both our proposed " + }, + { + "bbox": [ + 54, + 357, + 295, + 418 + ], + "type": "inline_equation", + "content": "\\mathrm{R}^2\\mathrm{S}" + }, + { + "bbox": [ + 54, + 357, + 295, + 418 + ], + "type": "text", + "content": " framework and the 3D ReasonSeg dataset contribute to substantial performance improvements on established benchmarks, demonstrating their complementary effects in enhancing spatial reasoning capabilities. In summary, our main contributions are:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 421, + 295, + 601 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 66, + 421, + 295, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 421, + 295, + 492 + ], + "spans": [ + { + "bbox": [ + 66, + 421, + 295, + 492 + ], + "type": "text", + "content": "- We propose the Relevant Segmentation " + }, + { + "bbox": [ + 66, + 421, + 295, + 492 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2\\mathbb{S})" + }, + { + "bbox": [ + 66, + 421, + 295, + 492 + ], + "type": "text", + "content": ", a reasoning-based segmentation framework that enhances spatial reasoning capabilities of 3D MLLMs by emulating human reasoning processes through two key components: Reasoning Prior Learning and Prior-guided Refinement." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 493, + 295, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 493, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 66, + 493, + 295, + 552 + ], + "type": "text", + "content": "- We propose 3D ReasonSeg, a reasoning-based segmentation dataset that focuses on object functionality, visual attributes, and spatial relationships, designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 553, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 553, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 66, + 553, + 295, + 601 + ], + "type": "text", + "content": "- Extensive experimental results demonstrate that our " + }, + { + "bbox": [ + 66, + 553, + 295, + 601 + ], + "type": "inline_equation", + "content": "\\mathrm{R}^2\\mathrm{S}" + }, + { + "bbox": [ + 66, + 553, + 295, + 601 + ], + "type": "text", + "content": " framework, combined with the 3D ReasonSeg dataset, achieves significant improvements in spatial reasoning across multiple benchmarks." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 614, + 143, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 614, + 143, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 614, + 143, + 627 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "type": "text", + "content": "2.1. 3D Scene Understanding" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "3D scene understanding requires models to detect objects and comprehend their characteristics in 3D space [3, 7, 8, 22, 29, 31]. Among these tasks, 3D segmentation presents unique challenges as it demands point-wise classification within point clouds. Recent studies demonstrate" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 369 + ], + "type": "text", + "content": "that language integration enhances model performance, as evidenced in 3D Question Answering and 3D Referring tasks [2, 4, 19, 28, 40]. These language-guided tasks, which require models to respond to instructions or identify objects based on descriptions, underscore the importance of language comprehension and spatial reasoning capabilities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 369, + 556, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 514 + ], + "type": "text", + "content": "The integration of language has supported models in transitioning from closed-set to open-set scenarios, such as in 3D Open Vocabulary Segmentation [4, 11, 12, 30, 32]. This step is critical as it allows the model to break free from being constrained by the finite set of categories in the training dataset. While most task-oriented studies focus on specific objectives, recent advancements in 3D multi-modal large language models have enabled addressing multiple 3D downstream tasks using these models [6, 18]. Furthermore, the assistance of 3D multi-modal large language models has empowered models to tackle more intricate object localization tasks [9, 42]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 556, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 575 + ], + "type": "text", + "content": "These prior studies have not adequately addressed the complex spatial reasoning capabilities required for comprehensive understanding of 3D scenes. In this work, we introduce Relevant Reasoning Segmentation and 3D ReasonSeg to address complex spatial reasoning challenges." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 586, + 524, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 524, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 524, + 598 + ], + "type": "text", + "content": "2.2. 3D Multi-modal Large Language Model" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 700 + ], + "type": "text", + "content": "Significant progress has been made in the realm of 3D multi-modal large language models (3D MLLMs). Initial endeavors have predominantly concentrated on point clouds at the object level [14, 15, 26, 35], posing challenges in tackling complex tasks at the scene level. Recently, researchers have shifted attention to understanding entire scene point clouds [6, 16, 17, 33], significantly advancing 3D scene understanding." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 713 + ], + "type": "text", + "content": "Drawing from these advances, we present an efficient" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7852" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "framework that handles various scene understanding tasks, with particular strength in segmentation. While existing 3D MLLMs can process basic spatial relationships, they struggle with complex spatial reasoning tasks. Our model, however, demonstrates superior performance in handling sophisticated spatial reasoning challenges." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 154, + 111, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 154, + 111, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 154, + 111, + 166 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 174, + 296, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 281 + ], + "type": "text", + "content": "Efforts such as those outlined in [6, 16, 17] were dedicated to leveraging 3D MLLMs for a deeper comprehension of 3D point clouds. Nevertheless, it is evident that current methods encounter challenges in effectively managing intricate spatial relationships. Specifically, as illustrated in Fig. 2, while the model can accurately identify \"What is the object that is used for sitting?\", it struggles with queries such as \"What is the object that is used for sitting and located near the desk with a monitor on it?\"" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "type": "text", + "content": "To tackle this issue, in this work, we instruct the model to generate reasoning priors by paying attention to objects relevant to the target. Subsequently, the model examines the scene and these relevant objects more closely to locate the final target. We refer to this process as Relevant Reasoning Segmentation " + }, + { + "bbox": [ + 55, + 282, + 296, + 354 + ], + "type": "inline_equation", + "content": "(\\mathbb{R}^2\\mathbb{S})" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 437 + ], + "type": "text", + "content": "The remainder of this section is organized as follows: Sec. 3.1 introduces our baseline model, which demonstrates strong dense perception capabilities on 3D point cloud data. Sec. 3.2 elaborates on our proposed Relevant Reasoning Segmentation framework. Sec. 3.3 details the construction of our 3D ReasonSeg dataset, and Sec. 3.4 describes the comprehensive training pipeline." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 149, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 149, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 149, + 456 + ], + "type": "text", + "content": "3.1. Baseline Model" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 557 + ], + "type": "text", + "content": "Our baseline architecture consists of a multi-modal large language model (MLLM) and a Segmentation Head. Following [23], we extend the MLLM's vocabulary with a special token [SEG] that instructs the Segmentation Head to generate corresponding segmentation mask. This design enables the MLLM to effectively perform fine-grained segmentation tasks while maintaining its core language understanding capabilities." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "Multi-modal Large Language Model. There are three key components in the adopted MLLM for feature extraction: a visual encoder, a Q-Former, and a large language model (LLM). Initially, the point cloud " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "X_{p}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is processed through the visual encoder to extract dense features. To optimize computational efficiency, we employ super-point [24] within the visual backbone, aggregating the dense visual features into super-point features " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " to reduce the following computational burden. Subsequently, the Q-Former [25] compresses the scene's information into several latent queries " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "q_{l}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " based on the super-point features " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "f_{p}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " and text embedding " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " of the instruction. After that, the latent queries " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{q}_{l}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " will serve as the visual input, along with" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "the text embedding " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ", to be processed by the large language model " + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{LLM}" + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": ". This process is formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 350, + 108, + 555, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 108, + 555, + 124 + ], + "spans": [ + { + "bbox": [ + 350, + 108, + 555, + 124 + ], + "type": "interline_equation", + "content": "\\pmb {f} _ {p} = \\mathcal {F} _ {e n c} (\\pmb {X} _ {p}), \\quad \\hat {\\pmb {q}} _ {l} = \\mathcal {F} _ {Q} (\\pmb {q} _ {l}, \\pmb {w}, \\pmb {f} _ {p}). \\qquad (1)", + "image_path": "3bf54bd6d98be8da6732c7c823da53781422785681d2bfe1aad4b2b9e4a03ec8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 388, + 138, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 138, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 388, + 138, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t x t} = \\mathcal {F} _ {L L M} (\\hat {\\boldsymbol {q}} _ {l}, \\boldsymbol {w}). \\tag {2}", + "image_path": "d69ff2a2e10cc8b924a738ee84dc87d411cf8cb6d91e9b272d79311b0c16d184.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": "Inside the Q-Former " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{q}_l" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " is concatenated with " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{w}" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " to be the query in both the self-attention and cross-attention layers of " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": ", while " + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 313, + 156, + 555, + 205 + ], + "type": "text", + "content": " serves as the key and value in the cross-attention layer:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 353, + 215, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 215, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 353, + 215, + 555, + 228 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\text {S e l f A t t n} (\\text {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w})), \\tag {3}", + "image_path": "b6921b4c4805f8fb523341dc29ea0ff629bb4bed616bc9016410a8f51ea4e86e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 328, + 229, + 554, + 243 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 229, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 328, + 229, + 554, + 243 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {l}, \\boldsymbol {w} = \\operatorname {C r o s s A t t n} (\\operatorname {C o n c a t} (\\boldsymbol {q} _ {l}, \\boldsymbol {w}), \\boldsymbol {f} _ {p}, \\boldsymbol {f} _ {p}). \\tag {4}", + "image_path": "503343a03959883a58e9c9dd8ee81b52eb7d34cf7114aa379094163b207f839c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": "Segmentation Head. If the output of LLM, i.e., " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{\\text{txt}}" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": ", contains [SEG] tokens, it means that the current query requires segmentation predictions. Then, the corresponding hidden features of the [SEG] token, i.e., " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{seg}" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": ", are processed by the segmentation head " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dec}" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": " that follows the decoder structure of [21] to yield the requisite segmentation mask " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": ", based on the reasoning between " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{seg}" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_p" + }, + { + "bbox": [ + 313, + 251, + 554, + 347 + ], + "type": "text", + "content": ". The process can be formulated as follows." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 388, + 359, + 555, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 359, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 388, + 359, + 555, + 373 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {h} _ {\\text {s e g}}, \\boldsymbol {f} _ {p}\\right). \\tag {5}", + "image_path": "c75a7bf63d573ddc8e982060da7c0102e393c9759fa8fd3537ce2ee77872f085.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 426 + ], + "type": "text", + "content": "The aforementioned framework seamlessly incorporates dense visual perception abilities into the output of the language model, allowing for the fusion of linguistic and visual information processing." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 433, + 496, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 496, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 496, + 445 + ], + "type": "text", + "content": "3.2. Relevant Reasoning Segmentation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": "The baseline model can decently handle point cloud perception tasks. However, in our observations, the direct instruction tuning performed in [23] did not consider the spatial relationships among objects in the scene, hindering complex spatial reasoning, as illustrated in Fig. 8. Hence, to address this issue, we suggest that the model initially identifies the objects related to the target and their spatial relationships, serving as the reasoning prior to guide the target localization. This process is thus termed as Relevant Reasoning Segmentation " + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "inline_equation", + "content": "(\\mathrm{R}^2\\mathrm{S})" + }, + { + "bbox": [ + 313, + 450, + 554, + 569 + ], + "type": "text", + "content": ", and it contains two steps as follows." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 689 + ], + "type": "text", + "content": "Step 1: Reasoning Prior Learning. While the model may encounter challenges in understanding intricate interobject relationships, it can recognize objects relevant to the target. As shown in Fig. 2, despite the model's limitation in accurately segmenting the chair beside the desk supporting a monitor, it demonstrates the ability to detect desks and chairs. Given this capability, we may instruct the model to initially segment approximate question-related objects, thus establishing the visual reasoning prior as the hint to help identify the final target." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 712 + ], + "type": "text", + "content": "Following the baseline model's process, we first use an instruction structured as: \"Given the 3D" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7853" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 70, + 264, + 192 + ], + "blocks": [ + { + "bbox": [ + 69, + 70, + 264, + 192 + ], + "lines": [ + { + "bbox": [ + 69, + 70, + 264, + 192 + ], + "spans": [ + { + "bbox": [ + 69, + 70, + 264, + 192 + ], + "type": "image", + "image_path": "95a33270b54c03ace6259e2db950f676b97fd22a9df65e3af635b0ce522fea87.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 97, + 194, + 242, + 216 + ], + "lines": [ + { + "bbox": [ + 97, + 194, + 242, + 216 + ], + "spans": [ + { + "bbox": [ + 97, + 194, + 242, + 216 + ], + "type": "text", + "content": "[QUESTION]: What is the object that is used for sitting?" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 278, + 70, + 541, + 192 + ], + "blocks": [ + { + "bbox": [ + 278, + 70, + 541, + 192 + ], + "lines": [ + { + "bbox": [ + 278, + 70, + 541, + 192 + ], + "spans": [ + { + "bbox": [ + 278, + 70, + 541, + 192 + ], + "type": "image", + "image_path": "082b1e41b76fab545242a15a66576dfb70f3631be0bfc67072883afb742b74fe.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 299, + 194, + 529, + 215 + ], + "lines": [ + { + "bbox": [ + 299, + 194, + 529, + 215 + ], + "spans": [ + { + "bbox": [ + 299, + 194, + 529, + 215 + ], + "type": "text", + "content": "[QUESTION]: What is the object that is used for sitting and located near the desk with a monitor on it?" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 279, + 545, + 417 + ], + "blocks": [ + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 269 + ], + "type": "text", + "content": "Figure 2. The image on the left depicts the LLM's response to the question \"What is the object that is used for sitting?\" Conversely, the right side of the image illustrates the LLM's response to the question \"What is the object that is used for sitting and located near the desk with a monitor on it.\" It is noteworthy that the LLM provides an incorrect answer in the latter case. The accurate response should identify the chair within the red rectangle." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 67, + 279, + 545, + 417 + ], + "lines": [ + { + "bbox": [ + 67, + 279, + 545, + 417 + ], + "spans": [ + { + "bbox": [ + 67, + 279, + 545, + 417 + ], + "type": "image", + "image_path": "7bbc94e07301c79120a30b4d64e4c74fb0f544b9d6bb6600d46210f52cfa6a37.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 430, + 555, + 453 + ], + "lines": [ + { + "bbox": [ + 54, + 430, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 54, + 430, + 555, + 453 + ], + "type": "text", + "content": "Figure 3. The baseline model consists of a visual backbone for point feature encoding, a Q-Former that connects visual-language representations, a LLM for instruction processing, and a mask decoder that converts [SEG] tokens into segmentation masks." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": "scene, [QUESTION]. Please segment the question-related objects that may help answer the question.\" *. The corresponding instruction embedding " + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": " will be processed by Eq. (1) and Eq. (2) to obtain the text output " + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{y}_{txt}" + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": " from which we can obtain the [SEG] tokens representing these target-relevant objects. Subsequently, the segmentation head decodes the hidden features " + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{seg}" + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": " of [SEG] tokens into masks " + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 54, + 474, + 295, + 581 + ], + "type": "text", + "content": " for these target-relevant objects." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "content": "Next, we adopt mask-pooling to aggregate the superpoint features " + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{f}_p" + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "content": " with masks " + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "inline_equation", + "content": "M_r" + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "content": ", yielding the representations " + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{f}_r" + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "content": " for the target-relevant objects. The features " + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "inline_equation", + "content": "\\pmb{f}_r" + }, + { + "bbox": [ + 54, + 582, + 295, + 641 + ], + "type": "text", + "content": " serve as the reasoning priors in the subsequent step. The above process can be formally expressed as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 126, + 652, + 295, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 652, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 126, + 652, + 295, + 667 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} _ {r} = \\mathcal {F} _ {\\text {d e c}} \\left(\\boldsymbol {f} _ {p}, \\boldsymbol {h} _ {\\text {s e g}}\\right), \\tag {6}", + "image_path": "d11a2c33d1bc75e52540c638bd1b1331ba2b6f8925052ea01bb92fde59be0980.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 141, + 681, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 681, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 141, + 681, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\boldsymbol {f} _ {r} = \\boldsymbol {f} _ {p} \\times \\boldsymbol {M} _ {r} \\tag {7}", + "image_path": "2a63cb12bc0225cd6352ca26a24a521087bcf961b4c979572371e798aa1daa8a.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": "Step 2: Prior-guided Refinement. In the absence of explicit instructions, the latent queries " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{q}}_l" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": " derived from Eq. (1) may exhibit a bias towards irrelevant areas, potentially impacting the reasoning process in Eq. (2). Besides, the mask-pooling operation in Eq. (7) may cause the loss of essential spatial details. Therefore, we propose the prior-guided refinement which commences by feeding the initial latent queries " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{q}_l" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": ", text outputs " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "y_{txt}" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": " and target-relevant features " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}_r" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": " from step 1 into the Q-Former " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": " to yield the refined query " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{q}_l'" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": " and target-relevant features " + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "inline_equation", + "content": "\\pmb{f}_r'" + }, + { + "bbox": [ + 313, + 473, + 555, + 594 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "text", + "content": "To make use of the obtained reasoning pri- ors, we employ a revised instruction like: \"Given the 3D scene, the question is [QUESTION]. Please answer the question and output segmentation mask. The question-related objects are " + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "inline_equation", + "content": "\\{f_r^1,f_r^2" + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "inline_equation", + "content": "\\dots ,f_r^n\\}" + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "text", + "content": " , and you may need to pay attention to them\" " + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 313, + 594, + 556, + 691 + ], + "type": "text", + "content": " where the target-relevant" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 238, + 713 + ], + "type": "text", + "content": "*More templates are shown in the supplementary file" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 325, + 702, + 499, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 499, + 713 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 499, + 713 + ], + "type": "text", + "content": "†More templates are shown in the supplementary file." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7854" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 68, + 544, + 184 + ], + "blocks": [ + { + "bbox": [ + 61, + 68, + 544, + 184 + ], + "lines": [ + { + "bbox": [ + 61, + 68, + 544, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 68, + 544, + 184 + ], + "type": "image", + "image_path": "0d1dc52e37d91fdca0acf1369377737c6d5e2c3fa49cd61a6e7b462ad655c2bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 217 + ], + "type": "text", + "content": "Figure 4. The proposed Relevant Reasoning Segmentation framework. Stage 1 segments question-related objects as priors " + }, + { + "bbox": [ + 55, + 191, + 555, + 217 + ], + "type": "inline_equation", + "content": "(f_1', \\dots, f_6')" + }, + { + "bbox": [ + 55, + 191, + 555, + 217 + ], + "type": "text", + "content": ". Stage 2 refines hidden representations of instructions based on these priors, enabling the LLM to generate more accurate responses." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "text", + "content": "features " + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{f}_r" + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "text", + "content": " potentially relevant objects have been inserted to form the new instruction whose text embedding is therefore denoted as " + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "inline_equation", + "content": "\\pmb{w}' \\oplus \\pmb{f}_r" + }, + { + "bbox": [ + 55, + 236, + 297, + 297 + ], + "type": "text", + "content": ". To this end, the prior-guided refinement regarding the latent query and the target-relevant features can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 111, + 304, + 295, + 319 + ], + "type": "interline_equation", + "content": "\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {f} _ {r} ^ {\\prime} = \\mathcal {F} _ {Q} \\left(\\boldsymbol {q} _ {l}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r}, \\boldsymbol {f} _ {p}\\right) \\tag {8}", + "image_path": "981c94a4ffa4df9f1bf01ac165feba4b4657b01ba96407b0fd641e6681213fda.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " are the text embeddings and super-point features, respectively. Unlike Eq. (3), during the prior-guided refinement, " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{l}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " is concatenated with the instruction embedding " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{\\prime} \\oplus \\pmb{f}_{r}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " to serve as the queries for both self-attention and cross-attention layers in the Q-Former " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_Q" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": ". Meanwhile, " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " continues to function as the key and value in the cross-attention layer. In Eq. (8), the enhanced latent query " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " has been enriched with target-related clues from " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{r}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": ". Simultaneously, utilizing the point-cloud features " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{p}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": ", the target-relevant features " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{r}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " are refined with essential spatial intricacies to produce " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{r}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": ". Next, the refined query " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{q}_{l}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " and target-relevant features " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{r}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " are adopted for generating the new text output " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{txt}^{\\prime}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": " with the LLM " + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{LLM}" + }, + { + "bbox": [ + 55, + 327, + 296, + 483 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 115, + 491, + 295, + 507 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 491, + 295, + 507 + ], + "spans": [ + { + "bbox": [ + 115, + 491, + 295, + 507 + ], + "type": "interline_equation", + "content": "\\boldsymbol {y} _ {t x t} ^ {\\prime} = \\mathcal {F} _ {L L M} \\left(\\boldsymbol {q} _ {l} ^ {\\prime}, \\boldsymbol {w} ^ {\\prime} \\oplus \\boldsymbol {f} _ {r} ^ {\\prime}\\right), \\tag {9}", + "image_path": "89f33d49b222627b7db9c49af731e7ada20d548cf72ff9344ed035230266af61.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": "where the instruction embedding " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " originally used in Eq. (8) has been accordingly updated to " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{w}^{\\prime}\\oplus \\pmb{f}_{r}^{\\prime}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": ". The hidden features " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "h_{seg}^{\\prime}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " of the [SEG] tokens are extracted from the new text output " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\pmb{y}_{txt}^{\\prime}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " and processed by the decoder " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{dec}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": " to generate the refined final mask prediction " + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "inline_equation", + "content": "M^{\\prime}" + }, + { + "bbox": [ + 55, + 514, + 296, + 574 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 129, + 582, + 295, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 582, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 129, + 582, + 295, + 597 + ], + "type": "interline_equation", + "content": "\\boldsymbol {M} ^ {\\prime} = \\mathcal {F} _ {d e c} \\left(\\boldsymbol {h} _ {s e g} ^ {\\prime}, \\boldsymbol {f} _ {p}\\right) \\tag {10}", + "image_path": "85eaaa713074a69f88c367212600995dac9f902001a557f2dd35a252de9bea64.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 612, + 145, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 145, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 145, + 625 + ], + "type": "text", + "content": "3.3. 3D ReasonSeg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Analysis of existing datasets, as demonstrated in Fig. 5, reveals that spatial relations such as \"in front of\" often lack complexity and exhibit potential ambiguity. Using such datasets for training and evaluation may inadequately address the spatial reasoning requirements of 3D MLLMs. To facilitate enhancement and assessment of spatial reasoning capabilities in 3D MLLMs, we present 3D ReasonSeg, a" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 334, + 234, + 536, + 302 + ], + "blocks": [ + { + "bbox": [ + 334, + 234, + 536, + 302 + ], + "lines": [ + { + "bbox": [ + 334, + 234, + 536, + 302 + ], + "spans": [ + { + "bbox": [ + 334, + 234, + 536, + 302 + ], + "type": "image", + "image_path": "5de7a2200b20097d670dec606552502345bac8b4ac36a4b61ff7b51653e3f773.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 310, + 555, + 356 + ], + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 356 + ], + "type": "text", + "content": "Figure 5. This illustrates a data sample from ScanRefer, where the target object is described as \"A brown chair placed in front of a plant.\" The spatial relationship in this description shows ambiguity, as there are multiple chairs in front of the plant." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 424 + ], + "type": "text", + "content": "dataset specifically designed for reasoning tasks in 3D point cloud environments. Empirical results in Tab. 2 indicate that models trained on 3D ReasonSeg achieve substantial performance improvements over baseline implementations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 544 + ], + "type": "text", + "content": "Specifically, the proposed 3D ReasonSeg requires the model to identify the target object based on a question that does not explicitly mention it. The model needs to understand the object functions, visual attributes, and spatial relationships of the objects described in the query and reason to segment the target object. To implement this task, we use a template structured as: \"USER: Given the 3D scene, please answer the question and output segmentation mask: [QUESTION]. ASSISTANT: [ANSWER]\"." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 616 + ], + "type": "text", + "content": "We construct the 3D ReasonSeg dataset by integrating SceneVerse's [20] detailed object descriptions with LLama 3.1 [13]. Based on objects' 3D spatial information (positions and sizes) from point clouds and the object descriptions from SceneVerse, LLama 3.1 generates question-reasoning-answer pairs for reasoning tasks." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 554, + 700 + ], + "type": "text", + "content": "After collecting numerous question-reasoning-answer pairs, we apply rule-based filtering (e.g., calculating distances between objects and removing outliers) to eliminate instances with unclear or incorrect spatial relationships. The resulting 3D ReasonSeg dataset contains 29,151 high-quality data samples. Examples are shown in Fig. 6, with details provided in the supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "We conduct a statistical analysis of ScanQA, ScanRe" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7855" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 145 + ], + "type": "text", + "content": "fer, and 3D ReasonSeg datasets, examining their training (#Train) and validation (#Val) set sizes, average word counts per sample (Avg. Words), and average number of relevant objects (Avg. Objects). The detailed statistics are summarized in Tab. 1, showing that our proposed 3D ReasonSeg dataset contains more complex object relationships." + } + ] + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 64, + 153, + 289, + 201 + ], + "blocks": [ + { + "bbox": [ + 64, + 153, + 289, + 201 + ], + "lines": [ + { + "bbox": [ + 64, + 153, + 289, + 201 + ], + "spans": [ + { + "bbox": [ + 64, + 153, + 289, + 201 + ], + "type": "table", + "html": "
Dataset#Train#ValAvg. WordsAvg. Objects
ScanQA26,5634,6758.81.5
ScanRefer36,6659,50817.81.8
3D ReasonSeg25,1853,96619.65.4
", + "image_path": "64b6eaeab5d8b41e74d125d9346d53622f62825fff9ae89e9a064c12fba43c6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 295, + 232 + ], + "type": "text", + "content": "Table 1. Comparative statistics across ScanQA, ScanRefer, and 3D ReasonSeg datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 78, + 262, + 194, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 262, + 194, + 281 + ], + "spans": [ + { + "bbox": [ + 78, + 262, + 194, + 281 + ], + "type": "text", + "content": "[QUESTION]: What is the nearest resting place to the bicycle?" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 78, + 301, + 174, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 301, + 174, + 312 + ], + "spans": [ + { + "bbox": [ + 78, + 301, + 174, + 312 + ], + "type": "text", + "content": "[ANSWER]: The couch [32]." + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 202, + 257, + 274, + 315 + ], + "blocks": [ + { + "bbox": [ + 202, + 257, + 274, + 315 + ], + "lines": [ + { + "bbox": [ + 202, + 257, + 274, + 315 + ], + "spans": [ + { + "bbox": [ + 202, + 257, + 274, + 315 + ], + "type": "image", + "image_path": "36b436c34ab671c1bf579ca79358f853bb7617d314655ec72a7c58a3465fd107.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 78, + 325, + 196, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 325, + 196, + 353 + ], + "spans": [ + { + "bbox": [ + 78, + 325, + 196, + 353 + ], + "type": "text", + "content": "[QUESTION]: What is the object that is used for storing books and is located near the chair and the door?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 78, + 365, + 167, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 365, + 167, + 376 + ], + "spans": [ + { + "bbox": [ + 78, + 365, + 167, + 376 + ], + "type": "text", + "content": "[ANSWER]: The shelf [6]." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 203, + 323, + 274, + 380 + ], + "blocks": [ + { + "bbox": [ + 203, + 323, + 274, + 380 + ], + "lines": [ + { + "bbox": [ + 203, + 323, + 274, + 380 + ], + "spans": [ + { + "bbox": [ + 203, + 323, + 274, + 380 + ], + "type": "image", + "image_path": "3a31427c8e0700260019421d7f59901ed18076f43c7cbe19017abcacc322a985.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "lines": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 295, + 417 + ], + "type": "text", + "content": "Figure 6. Visualization of 3D ReasonSeg samples, where instance ID colors correspond to their respective object colors in the image." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 436, + 152, + 449 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 436, + 152, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 436, + 152, + 449 + ], + "type": "text", + "content": "3.4. Model Training" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 453, + 295, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 295, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 295, + 513 + ], + "type": "text", + "content": "Training Datasets. Following [6], our training data includes established 3D datasets such as ScanQA [2], ScanRefer [4], ScanNet200 [10], the ScanNet subset of 3D-LLM [16], and SceneVerse [20], along with our proposed 3D ReasonSeg dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 514, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 295, + 620 + ], + "type": "text", + "content": "For Relevant Reasoning Segmentation outlined in Sec. 3.2, we extend the ScanQA, Scanrefer, and 3D ReasonSeg to incorporate target-relevant object identification. This process leverages scene object annotations and question-answer pairs from the original datasets, with LLama 3.1 [13] serving as the backbone for target-relevant object identification. Representative examples are presented in Fig. 7, with detailed prompts and implementation procedures in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": "Loss Function. During training, the model is optimized using the cross-entropy loss " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {txt }}" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": " for text generation, along with the binary cross-entropy (BCE) and DICE losses (" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {bce }}" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text {dice }}" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": ") for segmentation. Following [22], we assign equal weights of 1 to the BCE and DICE loss. Thus, the overall training objective " + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 621, + 295, + 693 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 114, + 701, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 114, + 701, + 295, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {t x t} \\mathcal {L} _ {t x t} + \\mathcal {L} _ {b c e} + \\mathcal {L} _ {d i c e} \\tag {11}", + "image_path": "6936e7084f29c137eb3a396d2affac449b96844d5e7dc0c6a6375030ad44c82c.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 511, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 511, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 511, + 84 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 511, + 84 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 313, + 72, + 511, + 84 + ], + "type": "text", + "content": " is the weight for text generation loss." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 85, + 553, + 193 + ], + "type": "text", + "content": "Relevant Objects Augmentation. Our relevant segmentation leverages mask predictions of relevant objects as priors to enhance latent query and text embedding. However, direct implementation leads to overfitting during training, as the ground truth relevant objects remain fixed during training but vary during inference, creating a train-test discrepancy. To address this issue, we implement random omission and addition of relevant objects during training, simulating diverse inference scenarios with incomplete priors." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 328, + 211, + 439, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 211, + 439, + 228 + ], + "spans": [ + { + "bbox": [ + 328, + 211, + 439, + 228 + ], + "type": "text", + "content": "[QUESTION]: How many pictures are on the wall?" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 230, + 375, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 230, + 375, + 240 + ], + "spans": [ + { + "bbox": [ + 328, + 230, + 375, + 240 + ], + "type": "text", + "content": "[ANSWER]: 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 328, + 243, + 400, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 243, + 400, + 262 + ], + "spans": [ + { + "bbox": [ + 328, + 243, + 400, + 262 + ], + "type": "text", + "content": "Relevant Objects: picture [23][24][25][26]" + } + ] + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 460, + 205, + 544, + 265 + ], + "blocks": [ + { + "bbox": [ + 460, + 205, + 544, + 265 + ], + "lines": [ + { + "bbox": [ + 460, + 205, + 544, + 265 + ], + "spans": [ + { + "bbox": [ + 460, + 205, + 544, + 265 + ], + "type": "image", + "image_path": "7dfdb1ef97733c6182b81dda58a2a35e0891a3460b818bcf537dd615893ed5da.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 328, + 272, + 457, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 272, + 457, + 299 + ], + "spans": [ + { + "bbox": [ + 328, + 272, + 457, + 299 + ], + "type": "text", + "content": "[DESCRIPTION]: this is the black kitchen counter with the sink in its. above that kitchen counter is a paper towel dispenser." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 328, + 301, + 419, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 301, + 419, + 327 + ], + "spans": [ + { + "bbox": [ + 328, + 301, + 419, + 327 + ], + "type": "text", + "content": "Relevant Objects: kitchen counter [1], sink [11] and paper towel dispenser [15]" + } + ] + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 460, + 270, + 544, + 328 + ], + "blocks": [ + { + "bbox": [ + 460, + 270, + 544, + 328 + ], + "lines": [ + { + "bbox": [ + 460, + 270, + 544, + 328 + ], + "spans": [ + { + "bbox": [ + 460, + 270, + 544, + 328 + ], + "type": "image", + "image_path": "f147fe0b62d136125a486422cfd1030cc48ee882a79343a9b65c629f2ce0afc9.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 554, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 386 + ], + "type": "text", + "content": "Figure 7. Examples of target-relevant data. The first row shows an instance from ScanQA, and the second row is related to ScanRefer. The visualizations of target-relevant objects are displayed on the right side of each row." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 412, + 395, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 412, + 395, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 412, + 395, + 426 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 432, + 437, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 437, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 437, + 445 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": "Implementation Details. We adopt the encoder of a pretrained Oneformer3D as visual backbone and the decoder of it as the mask decoder, and a pre-trained OPT 1.3B [39] as the language model. To preserve the learned knowledge of the pre-trained language model, we completely freeze the parameters of the language model while fully training the parameters of all other modules. Note that, we found it is necessary to also fine-tune the visual backbone. We adopt the AdamW [27] optimizer with a weight decay of 0.1 and a learning rate decaying from 1e-4 to 1e-6 with a cosine annealing scheduler. The text generation loss weight " + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 313, + 449, + 555, + 628 + ], + "type": "text", + "content": " is set to 0.5. The number of learnable queries in Q-Former is set to 32. We randomly sample 300k points from each 3D scene as the 3D input. All the experiments are conducted on eight Nvidia H20 GPUs." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 713 + ], + "type": "text", + "content": "Benchmark and Metrics. In this study, we assess our methods on ScanQA [2], ScanRefer [4], and our proposed 3D ReasonSeg. ScanQA is a dataset for the 3D question answering task, evaluating visual and spatial comprehension of 3D environments. The evaluation metrics of ScanQA include BLEU-4, CIDEr, METEOR, and Rouge-L. ScanRefer evaluates the capability to localize 3D object using a natural" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7856" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 95, + 70, + 515, + 319 + ], + "blocks": [ + { + "bbox": [ + 95, + 70, + 515, + 319 + ], + "lines": [ + { + "bbox": [ + 95, + 70, + 515, + 319 + ], + "spans": [ + { + "bbox": [ + 95, + 70, + 515, + 319 + ], + "type": "table", + "html": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C ↑M ↑R ↑gIoU
ScanRefer [4]37.324.3-----
MVT [19]40.833.3-----
3DVG-Trans [41]45.934.5-----
ViL3DRel [5]47.937.7-----
M3DRef-CLIP [40]51.944.7-----
ScanQA [2]--10.164.913.133.3-
3D-VisTA [44]50.645.813.172.9---
LLM-Grounder [36]17.15.3-----
3D-LLM [16]30.3-12.069.414.535.7-
Chat-3D [33]--6.453.211.928.5-
Chat-3D v2 [17]35.930.47.377.116.140.1-
LL3DA [6]--13.576.815.937.3-
ScanReason [42]53.141.1-----
Grounded 3D-LLM [9]47.944.113.472.7--25.6
baseline54.638.712.870.513.834.6-
+ 3D ReasonSeg58.943.613.375.215.536.729.2
+ R2S59.548.713.977.316.438.133.1
", + "image_path": "1003638d72d3556cfb37f9ea3fe07ab5c0bf54ec7419ee34d249d5ce4c5b69b7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 359, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 359, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 359, + 296, + 479 + ], + "type": "text", + "content": "language description. Prediction accuracy is measured by the percentage of predictions with Intersection over Union (IoU) above 0.25 and 0.5 compared to ground truth masks. For 3D ReasonSeg, the model is required to focus on object attributes and relationships to reason, answering questions and localizing target objects. Following [23], the general Intersection over Union (gIoU) metric is used, defined as the average of all per-scene Intersection-over-Unions (IoUs). Detailed formulations of these tasks can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 490, + 141, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 141, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 141, + 502 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 508, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 296, + 592 + ], + "type": "text", + "content": "We evaluate our model's effectiveness through comprehensive comparisons with state-of-the-art (SOTA) methods, encompassing both task-specific specialist models and multitask generalist models. For a fair comparison, we finetune Grounded 3D-LLM [9], the previous SOTA model on ScanRefer, on our proposed 3D ReasonSeg dataset and assess its performance on the 3D ReasonSeg benchmark." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "content": "As shown in Tab. 2, our baseline model is comparable to existing approaches on both ScanRefer and ScanQA. While our baseline model demonstrates robust performance, the incorporation of 3D ReasonSeg and " + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2\\mathbb{S}" + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "content": " yields consistent substantial improvements across all metrics, validating the effectiveness of our approach. Notably, " + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^2\\mathbb{S}" + }, + { + "bbox": [ + 55, + 593, + 295, + 700 + ], + "type": "text", + "content": " significantly enhances the Acc@50 metric on ScanRefer and the gIoU metric on 3D ReasonSeg, tasks that require robust spatial reasoning, aligning with our expectations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Effectiveness of Prior-guided Refinement. To verify the" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 357, + 553, + 405 + ], + "blocks": [ + { + "bbox": [ + 73, + 327, + 536, + 338 + ], + "lines": [ + { + "bbox": [ + 73, + 327, + 536, + 338 + ], + "spans": [ + { + "bbox": [ + 73, + 327, + 536, + 338 + ], + "type": "text", + "content": "Table 2. Results among our methods and previous related works. Entries in gray denote models specialized for specific datasets." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 357, + 553, + 405 + ], + "lines": [ + { + "bbox": [ + 317, + 357, + 553, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 357, + 553, + 405 + ], + "type": "table", + "html": "
MethodScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
wo PR58.241.37.766.113.433.227.5
text-based58.443.513.775.515.636.929.8
R2S59.548.713.977.316.438.133.1
", + "image_path": "e873547c3d60edbec52bccd7423d3d68b452f94e51b2fe090671aba649601dbd.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 433, + 547, + 476 + ], + "blocks": [ + { + "bbox": [ + 331, + 413, + 537, + 424 + ], + "lines": [ + { + "bbox": [ + 331, + 413, + 537, + 424 + ], + "spans": [ + { + "bbox": [ + 331, + 413, + 537, + 424 + ], + "type": "text", + "content": "Table 3. Ablation study on the Prior-guided Refinement." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 433, + 547, + 476 + ], + "lines": [ + { + "bbox": [ + 322, + 433, + 547, + 476 + ], + "spans": [ + { + "bbox": [ + 322, + 433, + 547, + 476 + ], + "type": "table", + "html": "
number of latent queriesScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4 ↑C↑M↑R↑gloU
1657.240.514.075.315.236.828.9
3258.943.613.375.215.536.729.2
4858.741.614.574.315.136.128.4
", + "image_path": "c6bf2ab6144141d22718abf9a6175a041a42ef0df10dfb5dc5044a0635b0d549.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 516, + 553, + 567 + ], + "blocks": [ + { + "bbox": [ + 313, + 485, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 506 + ], + "type": "text", + "content": "Table 4. Ablation study on the number of latent queries. The results are based on our base method." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 516, + 553, + 567 + ], + "lines": [ + { + "bbox": [ + 317, + 516, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 516, + 553, + 567 + ], + "type": "table", + "html": "
λtxtScanReferScanQA3D ReasonSeg
Acc@25Acc@50B-4↑C↑M↑R↑gIoU
0.161.343.912.368.214.334.028.9
0.558.943.613.375.215.536.729.2
1.058.741.413.272.615.235.628.6
", + "image_path": "4c48f426fc816808cc474feb9c67d46bf2a16c85c9fe7b894781fa3ffdd89514.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "lines": [ + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "type": "text", + "content": "Table 5. Ablation study on the loss weight configurations, where " + }, + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "type": "text", + "content": " means the weight of " + }, + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{txt}" + }, + { + "bbox": [ + 313, + 574, + 555, + 597 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 616, + 555, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 555, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 555, + 663 + ], + "type": "text", + "content": "necessity of Prior-guided Refinement (PR), we investigate the approach without PR (wo PR) and directly representing target-relevant objects as textual information (text-based). The results are shown in Tab. 3." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "When Prior-guided Refinement is removed, the model simultaneously predict relevant objects and process instructions. Results indicate that the approach without Prior-guided Refinement not only fails to improve performance" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7857" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 86, + 77, + 187, + 135 + ], + "blocks": [ + { + "bbox": [ + 86, + 77, + 187, + 135 + ], + "lines": [ + { + "bbox": [ + 86, + 77, + 187, + 135 + ], + "spans": [ + { + "bbox": [ + 86, + 77, + 187, + 135 + ], + "type": "image", + "image_path": "74f26e80c719c52dcc76bd81e6a495e31cfde3bea504bf008023514cd70a69f7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 191, + 85, + 297, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 191, + 85, + 297, + 110 + ], + "spans": [ + { + "bbox": [ + 191, + 85, + 297, + 110 + ], + "type": "text", + "content": "USER: What is under the desk in front of the doorway? Please answer the question." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 86, + 136, + 188, + 194 + ], + "blocks": [ + { + "bbox": [ + 86, + 136, + 188, + 194 + ], + "lines": [ + { + "bbox": [ + 86, + 136, + 188, + 194 + ], + "spans": [ + { + "bbox": [ + 86, + 136, + 188, + 194 + ], + "type": "image", + "image_path": "e55245836a1e8f3e32d6354542a37206f14af2071e6e8915a34906b13ffbbf5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 192, + 119, + 290, + 129 + ], + "blocks": [ + { + "bbox": [ + 192, + 119, + 290, + 129 + ], + "lines": [ + { + "bbox": [ + 192, + 119, + 290, + 129 + ], + "spans": [ + { + "bbox": [ + 192, + 119, + 290, + 129 + ], + "type": "image", + "image_path": "2137490cb9d3d511307f8439be23c0a346ddb12e6cab2b4c042c2b8d762f030b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 190, + 140, + 297, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 140, + 297, + 181 + ], + "spans": [ + { + "bbox": [ + 190, + 140, + 297, + 181 + ], + "type": "text", + "content": "USER: What is under the desk in front of the doorway? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 192, + 184, + 278, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 184, + 278, + 193 + ], + "spans": [ + { + "bbox": [ + 192, + 184, + 278, + 193 + ], + "type": "text", + "content": "ASSISTANT: office chair" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 85, + 201, + 189, + 262 + ], + "blocks": [ + { + "bbox": [ + 85, + 201, + 189, + 262 + ], + "lines": [ + { + "bbox": [ + 85, + 201, + 189, + 262 + ], + "spans": [ + { + "bbox": [ + 85, + 201, + 189, + 262 + ], + "type": "image", + "image_path": "7b0424f19987811f88c8f53912f87c5240be412a4c15184aaf8eb6803a1c8c14.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 192, + 213, + 293, + 231 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 192, + 213, + 293, + 231 + ], + "spans": [ + { + "bbox": [ + 192, + 213, + 293, + 231 + ], + "type": "text", + "content": "USER: What is in the center of the room? Please answer the question." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 193, + 243, + 260, + 253 + ], + "blocks": [ + { + "bbox": [ + 193, + 243, + 260, + 253 + ], + "lines": [ + { + "bbox": [ + 193, + 243, + 260, + 253 + ], + "spans": [ + { + "bbox": [ + 193, + 243, + 260, + 253 + ], + "type": "image", + "image_path": "72f9be3e22b9dddc56d7bc895567e67f5cb557720cdae4f6bb5486aa93086ccf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 193, + 264, + 295, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 264, + 295, + 306 + ], + "spans": [ + { + "bbox": [ + 193, + 264, + 295, + 306 + ], + "type": "text", + "content": "USER: What is in the center of the room? Please answer the question. The question-related objects are [SEG], and you may need to pay attention to them." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 193, + 312, + 285, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 193, + 312, + 285, + 321 + ], + "spans": [ + { + "bbox": [ + 193, + 312, + 285, + 321 + ], + "type": "text", + "content": "ASSISTANT: kitchen island" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 76, + 493, + 102 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 76, + 493, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 76, + 493, + 102 + ], + "type": "text", + "content": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 102, + 406, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 102, + 406, + 110 + ], + "spans": [ + { + "bbox": [ + 315, + 102, + 406, + 110 + ], + "type": "text", + "content": "ASSISTANT: the towel [SEG]" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 110, + 494, + 194 + ], + "blocks": [ + { + "bbox": [ + 315, + 110, + 494, + 194 + ], + "lines": [ + { + "bbox": [ + 315, + 110, + 494, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 110, + 494, + 194 + ], + "type": "image", + "image_path": "dcd364a6105a3588229286a0f8dc234dd69a9a9479c261a25b0b62eda5252206.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 198, + 498, + 233 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 198, + 498, + 233 + ], + "spans": [ + { + "bbox": [ + 315, + 198, + 498, + 233 + ], + "type": "text", + "content": "USER: What object is used for drying the body and is located on the towel rack near the bathtub? Please answer the question with segmentation mask. The question-related objects are [SEG][SEG][SEG], and you may need to pay attention to them." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 235, + 396, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 235, + 396, + 243 + ], + "spans": [ + { + "bbox": [ + 316, + 235, + 396, + 243 + ], + "type": "text", + "content": "ASSISTANT: the towel [SEG]" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 243, + 494, + 325 + ], + "blocks": [ + { + "bbox": [ + 315, + 243, + 494, + 325 + ], + "lines": [ + { + "bbox": [ + 315, + 243, + 494, + 325 + ], + "spans": [ + { + "bbox": [ + 315, + 243, + 494, + 325 + ], + "type": "image", + "image_path": "ed56bd369f5edc2548b1494cccad8d39a27307da0db61915f174b6e218cc4e9f.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 183, + 341, + 425, + 353 + ], + "lines": [ + { + "bbox": [ + 183, + 341, + 425, + 353 + ], + "spans": [ + { + "bbox": [ + 183, + 341, + 425, + 353 + ], + "type": "text", + "content": "Figure 8. Visual qualitative results on ScanQA and 3D ReasonSeg." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 295, + 422 + ], + "type": "text", + "content": "but potentially degrades it. This degradation likely stems from the increased computational complexity of concurrent object prediction and instruction processing without supplementary contextual information." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 435, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 591 + ], + "type": "text", + "content": "The text-based method represents an ablation variant of " + }, + { + "bbox": [ + 55, + 435, + 296, + 591 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2\\mathbf{S}" + }, + { + "bbox": [ + 55, + 435, + 296, + 591 + ], + "type": "text", + "content": " where segmentation is omitted, providing only text prompts such as \"The question-related objects are chairs, and you may need to pay attention to them\". Although text-based approach demonstrates modest improvements, our method consistently achieves superior results. This performance advantage can be attributed to our approach's ability to extract precise instance-level visual features through segmentation masks. In contrast, text-based method fails to capture spatial information and demonstrates inherent limitations in differentiating multiple instances of identical categories within a scene (e.g., multiple chairs)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 653 + ], + "type": "text", + "content": "The Number of Latent Queries. We investigate the number of Q-Former latent queries. As demonstrated in Tab. 4, experiments with 16, 32, and 48 latent queries reveal that 32 queries yield optimal performance." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Loss Weight. We conduct an ablation study on the text loss weight " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " using values of 0.1, 0.5, and 1.0. Our experiments demonstrate that setting " + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{txt}" + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": " to 1.0 achieves optimal balance in performance, as evidenced in Tab. 5." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 425, + 385 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "text", + "content": "As illustrated in Fig. 8, we present a comparison between the baseline and " + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2\\mathbf{S}" + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "text", + "content": ". The left portion of the figure compares baseline predictions (upper row) with " + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2\\mathbf{S}" + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "text", + "content": " predictions (lower row) for ScanQA. The lower row shows segmentation results for target-relevant objects. Similarly, the right portion compares baseline and " + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "inline_equation", + "content": "\\mathbf{R}^2\\mathbf{S}" + }, + { + "bbox": [ + 313, + 391, + 555, + 486 + ], + "type": "text", + "content": " predictions for 3D ReasonSeg. Additional qualitative results are provided in the supplementary material." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 496, + 384, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 384, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 384, + 508 + ], + "type": "text", + "content": "5. Limitation" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 516, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 555, + 575 + ], + "type": "text", + "content": "Due to the use of large language models for data generation instead of human annotation, some noise remains in our dataset even after data cleaning procedures, potentially impacting model performance. We will further optimize our dataset in future work." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 392, + 597 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduce the Relevant Reasoning Segmentation framework, a reasoning-based segmentation framework that enhances spatial reasoning through Reasoning Prior Learning and Prior-guided Refinement. Furthermore, we propose 3D ReasonSeg, a reasoning-based segmentation dataset designed to enhance and comprehensively assess the spatial reasoning capabilities in 3D MLLMs. We hope our work provides new insights and directions for exploring the spatial reasoning capabilities in 3D MLLMs." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7858" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 213 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (Grant Nos. 62372133, 62376074), the National Key R&D Program of China (Grant No. 2024YFE0215300), the Shenzhen Fundamental Research Program (Grant No. JCYJ20220818102415032), the Shenzhen Science and Technology Program (Grant No. JSGGKQTD20221101115655027), the Guangdong Basic and Applied Basic Research Foundation (Grant No. 2025A1515011434), and the Guangdong Provincial Key Laboratory of Mathematical Foundations for Artificial Intelligence (Grant No. 2023B1212010001)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 115, + 244 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 251, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 251, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 316 + ], + "type": "text", + "content": "[1] Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Mensch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning. Advances in neural information processing systems, 35:23716-23736, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 318, + 294, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 318, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 61, + 318, + 294, + 372 + ], + "type": "text", + "content": "[2] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 294, + 427 + ], + "type": "text", + "content": "[3] Yingjie Cai, Xuesong Chen, Chao Zhang, Kwan-Yee Lin, Xiaogang Wang, and Hongsheng Li. Semantic scene completion via integrating instances and scene in-the-loop. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 324-333, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "text", + "content": "[4] Dave Zhenyu Chen, Angel X Chang, and Matthias Nießner. Scanrefer: 3d object localization in rgb-d scans using natural language. In European conference on computer vision, pages 202-221. Springer, 2020. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 472, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 294, + 525 + ], + "type": "text", + "content": "[5] Shizhe Chen, Pierre-Louis Guhur, Makarand Tapaswi, Cordelia Schmid, and Ivan Laptev. Language conditioned spatial relation reasoning for 3d object grounding. Advances in neural information processing systems, 35:20522-20535, 2022. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 526, + 294, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 294, + 592 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 294, + 592 + ], + "type": "text", + "content": "[6] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 294, + 647 + ], + "type": "text", + "content": "[7] Xuesong Chen, Shaoshuai Shi, Benjin Zhu, Ka Chun Cheung, Hang Xu, and Hongsheng Li. Mppnet: Multi-frame feature intertwining with proxy points for 3d temporal object detection. In European Conference on Computer Vision, pages 680-697. Springer, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 647, + 294, + 712 + ], + "type": "text", + "content": "[8] Xuesong Chen, Shaoshuai Shi, Chao Zhang, Benjamin Zhu, Qiang Wang, Ka Chun Cheung, Simon See, and Hongsheng Li. Trajectoryformer: 3d object tracking transformer with predictive trajectory hypotheses. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18527-18536, 2023. 2" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[9] Yilun Chen, Shuai Yang, Haifeng Huang, Tai Wang, Ruiyuan Lyu, Runsen Xu, Dahua Lin, and Jiangmiao Pang. Grounded 3d-llm with referent tokens. arXiv preprint arXiv:2405.10370, 2024. 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 172 + ], + "type": "text", + "content": "[10] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828–5839, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 205 + ], + "type": "text", + "content": "[11] Anurag Das, Xinting Hu, Li Jiang, and Bernt Schiele. Mtaclip: Language-guided semantic segmentation with mask-text alignment. arXiv preprint arXiv:2407.21654, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 259 + ], + "type": "text", + "content": "[12] Runyu Ding, Jihan Yang, Chuhui Xue, Wenqing Zhang, Song Bai, and Xiaojuan Qi. Pla: Language-driven open-vocabulary 3d scene understanding. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7010-7019, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 303 + ], + "type": "text", + "content": "[13] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, and Akhil Mathur et.al. The llama 3 herd of models, 2024. 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 370 + ], + "type": "text", + "content": "[14] Ziyu Guo, Renrui Zhang, Xiangyang Zhu, Yiwen Tang, Xianzheng Ma, Jiaming Han, Kexin Chen, Peng Gao, Xianzhi Li, Hongsheng Li, et al. Point-bind & point-llm: Aligning point cloud with multi-modality for 3d understanding, generation, and instruction following. arXiv preprint arXiv:2309.00615, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[15] Jiaming Han, Renrui Zhang, Wenqi Shao, Peng Gao, Peng Xu, Han Xiao, Kaipeng Zhang, Chris Liu, Song Wen, Ziyu Guo, et al. Imagebind-llm: Multi-modality instruction tuning. arXiv preprint arXiv:2309.03905, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 469 + ], + "type": "text", + "content": "[16] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 471, + 553, + 523 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 523 + ], + "type": "text", + "content": "[17] Haifeng Huang, Zehan Wang, Rongjie Huang, Luping Liu, Xize Cheng, Yang Zhao, Tao Jin, and Zhou Zhao. Chat-3d v2: Bridging 3d scene and large language models with object identifiers. arXiv preprint arXiv:2312.08168, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 526, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 578 + ], + "type": "text", + "content": "[18] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 624 + ], + "type": "text", + "content": "[19] Shijia Huang, Yilun Chen, Jiaya Jia, and Liwei Wang. Multiview transformer for 3d visual grounding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15524-15533, 2022. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 679 + ], + "type": "text", + "content": "[20] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. arXiv preprint arXiv:2401.09340, 2024. 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollar, and" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7859" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 150 + ], + "type": "text", + "content": "[22] Maxim Kolodiazhnyi, Anna Vorontsova, Anton Konushin, and Danila Rukhovich. Oneformer3d: One transformer for unified point cloud segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20943-20953, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 205 + ], + "type": "text", + "content": "[23] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9579-9589, 2024. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 294, + 250 + ], + "type": "text", + "content": "[24] Loic Landrieu and Martin Simonovsky. Large-scale point cloud semantic segmentation with superpoint graphs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4558-4567, 2018. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 294, + 304 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 294, + 359 + ], + "type": "text", + "content": "[26] Dingning Liu, Xiaoshui Huang, Yuenan Hou, Zhihui Wang, Zhenfei Yin, Yongshun Gong, Peng Gao, and Wanli Ouyang. Uni3d-llm: Unifying point cloud perception, generation and editing with large language models. arXiv preprint arXiv:2402.03327, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "text", + "content": "[27] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 426 + ], + "type": "text", + "content": "[28] Xiaojian Ma, Silong Yong, Zilong Zheng, Qing Li, Yitao Liang, Song-Chun Zhu, and Siyuan Huang. Sqa3d: Situated question answering in 3d scenes. arXiv preprint arXiv:2210.07474, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 470 + ], + "type": "text", + "content": "[29] Ishan Misra, Rohit Girdhar, and Armand Joulin. An end-to-end transformer model for 3d object detection. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2906-2917, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[30] Songyou Peng, Kyle Genova, Chiyu Jiang, Andrea Tagliasacchi, Marc Pollefeys, Thomas Funkhouser, et al. Openscene: 3d scene understanding with open vocabularies. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 815-824, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 526, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 579 + ], + "type": "text", + "content": "[31] Shaoshuai Shi, Xiaogang Wang, and Hongsheng Li. Pointrcnn: 3d object proposal generation and detection from point cloud. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 770-779, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[32] Ayca Takmaz, Elisabetta Fedele, Robert W Sumner, Marc Pollefeys, Federico Tombari, and Francis Engelmann. Openmask3d: Open-vocabulary 3d instance segmentation. arXiv preprint arXiv:2306.13631, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 668 + ], + "type": "text", + "content": "[33] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 1, 2, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[34] Chenfei Wu, Shengming Yin, Weizhen Qi, Xiaodong Wang, Zecheng Tang, and Nan Duan. Visual chatgpt: Talking, drawing and editing with visual foundation models. arXiv preprint arXiv:2303.04671, 2023. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 585 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[35] Runsen Xu, Xiaolong Wang, Tai Wang, Yilun Chen, Jiangmiao Pang, and Dahua Lin. Pointllm: Empowering large language models to understand point clouds. arXiv preprint arXiv:2308.16911, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 182 + ], + "type": "text", + "content": "[36] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 239 + ], + "type": "text", + "content": "[37] Zhengyuan Yang, Linjie Li, Jianfeng Wang, Kevin Lin, Ehsan Azarnasab, Faisal Ahmed, Zicheng Liu, Ce Liu, Michael Zeng, and Lijuan Wang. Mm-react: Prompting chatgpt for multimodal reasoning and action. arXiv preprint arXiv:2303.11381, 2023. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 295 + ], + "type": "text", + "content": "[38] Qinghao Ye, Haiyang Xu, Guohai Xu, Jiabo Ye, Ming Yan, Yiyang Zhou, Junyang Wang, Anwen Hu, Pengcheng Shi, Yaya Shi, et al. MPLug-owl: Modularization empowers large language models with multimodality. arXiv preprint arXiv:2304.14178, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 296, + 553, + 350 + ], + "type": "text", + "content": "[39] Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen, Christopher Dewan, Mona Diab, Xian Li, Xi Victoria Lin, et al. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 352, + 553, + 406 + ], + "type": "text", + "content": "[40] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 2, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 316, + 407, + 553, + 452 + ], + "type": "text", + "content": "[41] Lichen Zhao, Daigang Cai, Lu Sheng, and Dong Xu. 3dvg-transformer: Relation modeling for visual grounding on point clouds. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2928-2937, 2021. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 485 + ], + "type": "text", + "content": "[42] Chenming Zhu, Tai Wang, Wenwei Zhang, Kai Chen, and Xihui Liu. Empowering 3d visual grounding with reasoning capabilities. arXiv preprint arXiv:2407.01525, 2024. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 486, + 553, + 529 + ], + "type": "text", + "content": "[43] Deyao Zhu, Jun Chen, Xiaogian Shen, Xiang Li, and Mohamed Elhoseiny. Minigpt-4: Enhancing vision-language understanding with advanced large language models. arXiv preprint arXiv:2304.10592, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 531, + 553, + 585 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 531, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 316, + 531, + 553, + 585 + ], + "type": "text", + "content": "[44] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 7" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7860" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_content_list.json b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..935de1e1899d8c3247aa0a71a727536041179bc3 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_content_list.json @@ -0,0 +1,1633 @@ +[ + { + "type": "text", + "text": "Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching", + "text_level": 1, + "bbox": [ + 98, + 128, + 898, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhankai Li $^{1}$ , Weiping Wang $^{1}$ , Jie Li $^{1}$ , Shigeng Zhang $^{1}$ , Yunan Hu $^{1}$ , Song Guo $^{2}$", + "bbox": [ + 184, + 202, + 805, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Central South University", + "bbox": [ + 397, + 227, + 601, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "2The Hong Kong University of Science and Technology", + "bbox": [ + 276, + 244, + 718, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "janlzk7@gmail.com, {wpwang, lijie55, sgzhang, hyn318}@csu.edu.cn, songguo@ust.hk", + "bbox": [ + 145, + 266, + 849, + 281 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 316, + 326, + 330 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the field of AI security, deep neural networks (DNNs) are highly sensitive to adversarial examples (AEs), which can cause incorrect predictions with minimal input perturbations. Although AEs exhibit transferability across models, targeted attack success rates (TASRs) are low due to differences in feature dimensions and decision boundaries. To enhance targeted AE transferability, we propose a novel approach using Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS) in the optimization process. Specifically, we employ a siamese-network-like framework to generate both non-targeted and targeted AEs. The ITC mechanism applies non-targeted adversarial gradients each epoch to impede the optimization of targeted perturbations, thereby improving robustness. Additionally, a top-k SDS strategy guides AEs to penetrate target class regions in the latent space while distancing from nontargeted regions, achieving optimal transferability. Compared to state-of-the-art competition-based attacks, our method significantly improves transferable TASRs by $16.1\\%$ and $21.4\\%$ on mainstream CNNs and ViTs, respectively, and demonstrates superior defense-breaking capabilities. Our code is available here.", + "bbox": [ + 89, + 347, + 485, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 709, + 220, + 724 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Adversarial attacks involve adding small, often imperceptible perturbations to input data, which can lead otherwise high-performing models to produce incorrect outputs. The concept of adversarial attacks was first introduced by [23] and further developed by [7], who demonstrated the vulnerability of deep neural networks (DNNs) to these perturbations. With the rapid advancement of DNNs in recent years, the security risks revealed by adversarial attacks have become a critical issue. Adversarial attacks are now widely used to assess model security, improve model robustness, and drive research in defense technologies.", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, the significance of adversarial attacks extends beyond these applications, as they have inadvertently become a widely used defense mechanism in real-world scenarios. A prime example is Google's login verification system, reCAPTCHA, which aims to distinguish real users from bots to prevent abuse. Adversarial examples (AEs) are one of the primary image verification methods employed, as illustrated in Figure 1. This raises an important question: is it secure enough? Since we do not know the bot's model architecture or training domain, the transferability of AEs becomes especially crucial. Most current research on adversarial transferability focuses on non-targeted attacks [14, 40], even extending to cross-architecture and cross-domain attacks [15, 37], achieving notable performance. In contrast, targeted attacks face a significant increase in difficulty due to the substantial changes in decision boundaries, and as a result, they still maintain a relatively low success rate. However, for login verification, given that the output category of non-targeted attacks is uncontrollable, and even the original verification images without perturbation may have \"low-quality\" issues (leading to misidentification by bots), in such cases, targeted AEs are often safer and more reliable than non-targeted AEs that may cause the deployed bots to misjudge into the ground truth category.", + "bbox": [ + 511, + 316, + 906, + 680 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, effective targeted attack methods are primarily based on transformation [1, 18, 29, 34], achieving transferability through input augmentation. However, the relentless pursuit and stacking of a greater variety of transformations, although enhancing transferability to some extent, does not guarantee robust performance in white-box attacks. Moreover, this stacking approach based on affine and perspective transformations offers limited insight into the complexity and vulnerability of DNNs. Recently, competition-based attacks have been proposed [2, 30], offering a unique perspective. Their advantage lies in compelling AEs to consider the more complex and diverse feature representations within DNNs during their generation, simulating various noise interferences and attack strategies,", + "bbox": [ + 511, + 689, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3716", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c4522ccb122a1c16a16b7d176ecd530a786125a1646d71d9284388dcab84ce61.jpg", + "image_caption": [ + "Figure 1. Example of Google's web login verification system, reCAPTCHA. It determines whether a user is human or a bot based on whether they can correctly match text to corresponding images." + ], + "image_footnote": [], + "bbox": [ + 119, + 88, + 452, + 262 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and promoting the learning of more robust feature combinations. The competitive mechanism, by delving into the model's decision boundaries, helps to uncover the latent vulnerabilities of DNNs and provides a new perspective for understanding and improving the model's generalization capabilities. However, no matter what the attack method is, it cannot conceal the widespread insufficiency in the transferability performance of current targeted adversarial attacks.", + "bbox": [ + 88, + 332, + 482, + 450 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this challenge, we propose a new targeted adversarial attack method called ITDS, which consists of two components: Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS), aimed at enhancing the transferability of targeted AEs. In the first part, we adopt a siamese-network-like framework, innovatively leveraging the ground truth gradients of multiple target class non-targeted AEs at intermediate stages to exert inverse \"competitiveness\" on non-target class targeted AEs during the optimization process. In the second part, to tackle the multi-dimensional complexity of decision boundaries in targeted attacks, we introduce a top- $k$ SDS strategy to seek stable perturbation update directions. This guides AEs to penetrate target class regions within the latent multidimensional space while globally distancing from several closest non-target region boundaries. The experimental results demonstrate that the proposed ITDS method has achieved significant advantages in enhancing the transferability and defense-breaking capabilities of targeted AEs, surpassing the existing state-of-the-art (SOTA) methods.", + "bbox": [ + 91, + 454, + 483, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 768, + 233, + 784 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we will briefly overview the background of adversarial attacks. Then focus on various targeted attack techniques that demonstrate effective transferability.", + "bbox": [ + 89, + 792, + 483, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Adversarial Attacks", + "text_level": 1, + "bbox": [ + 89, + 848, + 282, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Goodfellow et.al first designed a simple yet effective method called Fast Gradient Sign Method (FGSM) [7]. This", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "method generates AEs by making small adjustments to the input $x$ (original sample) along the direction of the gradient, i.e., $x^{adv} = x + \\delta$ . In the case of targeted attacks, FGSM updates the input $x$ in the direction that minimizes the classification loss for a given target class $y$ , with the objective function as follows:", + "bbox": [ + 511, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {x ^ {a d v}} \\mathcal {J} (x ^ {a d v}, y; \\theta) \\text {s . t .} \\| x - x ^ {a d v} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 189, + 903, + 214 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where $\\theta$ is the model parameters, $\\mathcal{J}$ denotes the Cross-Entropy (CE) loss function, and the $\\ell_{\\infty}$ norm is used to constrain the perturbation within the range $\\epsilon$ .", + "bbox": [ + 511, + 220, + 903, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "I-FGSM [13] is an iterative extension of FGSM, which gradually refines adversarial perturbation with small step sizes. To address gradient vanishing issues in targeted attacks, [39] proposes using Logits loss instead of CE loss, significantly enhancing targeted attack effectiveness. Based on this, [33] further proposes downscaling logit calibration with a temperature factor and an adaptive margin. While these methods provide a foundation for enhancing targeted attack effectiveness, their transferability remains limited.", + "bbox": [ + 511, + 266, + 906, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Transferable Adversarial Attacks", + "text_level": 1, + "bbox": [ + 511, + 410, + 807, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we introduce advanced transferable attacks that can successfully deceive other models using AEs generated on a single surrogate model without additional training. We categorize these methods into transformation-based and competitive-based attacks, focusing primarily on transferable targeted attacks, along with certain non-targeted attacks that can also be adapted for targeted attacks.", + "bbox": [ + 511, + 431, + 905, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Transformation-Based Attacks. DIM [34] is a representative transformation-based attack, which resizes and pads images with a certain probability at each step of the multi-step attack process. SSM [18] applies frequency domain augmentations through spectrum transformations on the input. ODIM [1] method significantly improves the targeted attack transferability by mapping input images to random 3D objects and applying varied rendering techniques. SIA [31] generates structurally diverse AEs by applying various random transformations to the images. Both ODIM and SIA integrate DIM into their implementation principles and share a similar lineage in transformation strategies. BSR [29] randomly shuffles and rotates the blocks within images, calculates the ensemble gradients of this new set of images, reduces attention heatmap differences across various models, and significantly enhances transferability.", + "bbox": [ + 511, + 537, + 906, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Competition-based Attacks. Admix [30] combines features from different categories in the input domain, creating multiple images for gradient computation, and enhancing transferability without degrading white-box performance. FIA [32] employs aggregated gradients to disrupt essential object-aware features while suppressing model-specific features, boosting AE transferability. RPA [38] adopts a similar approach to FIA, introducing random patch transforma", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3717", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions to benign images. To alleviate the overfitting dilemma common in an AE crafted by simple iterative attacks, FFT [35] encourages features conducive to the target class while discouraging those associated with the original class in an intermediate layer of the source model. CFM [2] introduces a competitive strategy, crafting targeted adversarial perturbations with diverse features by introducing two types of competitor noises: adverse perturbations toward different target classes and friendly perturbations towards the correct class, achieving SOTA results.", + "bbox": [ + 89, + 90, + 483, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. The Design of ITDS", + "text_level": 1, + "bbox": [ + 89, + 255, + 282, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section consists of two parts: the Inverse Target Gradient Competition mechanism and the Spatial Distance Stretching strategy. In each part, we first explain the motivation and then propose the corresponding method.", + "bbox": [ + 89, + 280, + 483, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Inverse Target Gradient Competition (ITC)", + "text_level": 1, + "bbox": [ + 89, + 349, + 460, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Motivation of ITC", + "text_level": 1, + "bbox": [ + 89, + 372, + 264, + 386 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The motivation for the first part of our research arises from a question: What is a competition mechanism? From existing work related to competition mechanisms, we summarize it as a way to enhance the transferability of AEs by promoting diversified attack strategies through simulating various noise interferences. Since noise interference is involved, the essence of competitiveness can be interpreted as an inverse force. For targeted attacks, could we identify the most competitive noise? We believe this is the key issue. To address this critical question, we turn to target class samples and consider the gradients of the target label for these samples. These gradients essentially represent the optimal force linking target class samples with the target label, which can be used in both forward and inverse directions. Therefore, we have a bold hypothesis: when using this gradient in inverse, it not only effectively generates non-targeted AEs but also serves as the optimal competitiveness for targeted attacks.", + "bbox": [ + 89, + 391, + 483, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Method of ITC", + "text_level": 1, + "bbox": [ + 89, + 655, + 243, + 669 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Figure 2 shows the overall framework of our Inverse Target Gradient Competition mechanism, which is jointly driven by both non-targeted and targeted attacks within a siamese-network-like architecture. Our strategy is to utilize the inverse target gradient to introduce resistance during the optimization process of targeted AEs. This inverse target gradient can be generated for any samples; but which inverse target gradient should we choose? As mentioned earlier, we seek the strongest resistance, which is the inverse gradient of the target label with respect to the target class sample. The approach of generating the inverse target gradient is illustrated in the lower part of Figure 2, where we use 'cock' as the target label as an example. In each sub-epochs (epochssub), multiple target class images that are correctly classified by the siamese classifier are selected to generate", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "non-targeted AEs. The optimization goal is to minimize the logit value of the target label. We set $\\mathrm{epoch}_{sub}$ to be one-tenth of the total epochs, as we believe that target class samples need a certain number of iterations to transform into robust non-targeted AEs. Accordingly, the non-targeted AEs generated in each $\\mathrm{epoch}_{sub}$ contain different levels and increasingly deeper latent features of target class. Therefore, we consider the inverse target gradient perturbations generated in each $\\mathrm{epoch}_{sub}$ to also have diverse potential competitive characteristics, and during the generation of nontargeted AEs in each epoch, multiple inverse target perturbations are generated to enrich competitive characteristics. The formulas are as follows:", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {g} _ {t} ^ {z} = \\nabla_ {z _ {t} ^ {a d v}} \\mathcal {L} (z _ {t} ^ {a d v}, y; \\theta), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 297, + 906, + 319 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t + 1} ^ {z} = \\mu \\cdot g _ {t} ^ {z} + \\tilde {g} _ {t} ^ {z}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 342, + 906, + 359 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {z, \\epsilon} \\{z _ {t} ^ {a d v} - \\alpha \\cdot \\operatorname {s i g n} (g _ {t + 1} ^ {z}) \\}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 378, + 906, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $z$ is the target class sample, $\\tilde{g}_t^z$ represents the inverse target gradient, and we follow [1, 2, 39] to use the Logit loss $\\mathcal{L}$ (opposite to CE loss optimization). $g_{t + 1}^{z}$ represents the momentum gradient without normalization for $z$ , and $z_{t + 1}^{adv}$ is the non-targeted AE generated in each round.", + "bbox": [ + 511, + 402, + 906, + 479 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a377ae4ff4cfa262b59dbfa1e310b1f38b511cb76d59b64e1f06b23f8c1d6607.jpg", + "image_caption": [ + "Figure 2. The overall framework of the ITC, e.g., combined with DIM, utilizes a siamese-network-like architecture to generate targeted and non-targeted AEs simultaneously. The upper part focuses on targeted attacks against non-target class samples, while the lower part addresses non-targeted attacks on target class samples. Note that for non-targeted attacks, the optimization condition utilizes the logits value of the target label, whereas the goal of the targeted attack is to stretch the spatial distance proposed in SDS." + ], + "image_footnote": [], + "bbox": [ + 514, + 492, + 906, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The process of generating targeted AEs is shown in the upper part of Figure 2, where the input is a non-target image, and the optimization goal is to maximize the spatial distance, which we introduce in the SDS of the next section. Each round of generated targeted AEs competes with", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3718", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "a certain proportion of inverse target gradients from different stages. The formulas are as follows:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {L} \\left(\\delta_ {i} \\cdot \\left(x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}\\right), y; \\theta\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 128, + 480, + 184 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ng _ {t + 1} = \\mu \\cdot g _ {t} + \\bar {g} _ {t} ^ {\\text {c p t}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 196, + 482, + 215 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {x, \\epsilon} \\left\\{x _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t + 1}\\right) \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 232, + 482, + 251 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n$ represents the number of target class samples in per epochs $_{sub}$ , $\\delta_{i}$ and $m$ are the scaling factor and level, $\\beta$ is the competition ratio, and each round of targeted AEs competes with the corresponding inverse target gradient.", + "bbox": [ + 89, + 253, + 483, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "There are two notable points. The first point (P1) is whether, for targeted or non-targeted attacks, the calculation of momentum gradients differs from the traditional MIFGSM [3], and we have not employed norm normalization (N-norm). One of the reasons is the absence of gradient explosion in our experimental evaluations, and more importantly, our experiments have shown that using N-norm would limit our attack performance. We believe that N-norm might restrict the magnitude of gradient updates, resulting in less aggressive generation of AEs. Without N-norm, gradient updates could become more flexible, allowing for more effective exploration of the model's decision boundary, thereby generating more aggressive AEs.", + "bbox": [ + 89, + 315, + 483, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The second point (P2) is that we did not use the momentum gradient as the competitive gradient. The reason is that the momentum gradient is essentially a cumulative gradient. It weakens the strength of the inverse target features in the corresponding round. Since we want to maintain the integrity of the inverse target features, using it without any normalization operation in conjunction with P1 would cause excessive damage to the input features. This would be catastrophic for targeted attacks. These two points will also be discussed in subsequent ablation experiments.", + "bbox": [ + 89, + 511, + 483, + 662 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5883614a28bfcc96c9c66418acb147b6bac45409622873337a6338d6a8f4bb9a.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 94, + 676, + 181, + 739 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/0c50cd2ec667da5ca912afa440004cf14cdee048fe3fdc36ac2dc304bd1083f5.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 189, + 676, + 274, + 739 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5928543575ec402b6c0440fb86cb17e6ed7c8e24720e1f3fa575d6875ab41d3a.jpg", + "image_caption": [ + "(c)", + "Figure 3. Comparison of targeted AEs with strong and weak transferability. Panels (a) and (c) show AEs with weak transferability, while (b) and (d) show AEs with strong transferability." + ], + "image_footnote": [], + "bbox": [ + 300, + 676, + 385, + 739 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6fbab54b3fc182edb5b5b4a1347815fbee0ab12b2732a16a8abaabee283e3fba.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 395, + 676, + 480, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Spatial Distance Stretching (SDS)", + "text_level": 1, + "bbox": [ + 89, + 828, + 383, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Motivation of SDS", + "text_level": 1, + "bbox": [ + 89, + 851, + 266, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The motivation for the second part of our study arises from an interesting phenomenon commonly observed in highly", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/3b364741ee6f638259da75a58d739f5a79f68750b315741ae8f710378eb1dd3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
LabelFigure12345
motor_scooter(a)42.0%5.5%4.3%3.6%3.2%
(b)83.3%4.9%2.1%1.1%0.3%
siamese_cat(c)49.6%5.6%5.3%5.1%5.0%
(d)90.5%2.1%1.3%0.7%0.2%
", + "bbox": [ + 522, + 88, + 901, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. The average top-5 confidence on black-box models successfully attacked by targeted AEs as shown in Figure 3.", + "bbox": [ + 511, + 165, + 903, + 195 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Algorithm 1 The ITDS Attack Algorithm", + "text_level": 1, + "bbox": [ + 514, + 210, + 792, + 227 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: Siamese classifier $f$ with parameters $\\theta$ , Logit loss $\\mathcal{L}$ , our Distance loss $\\mathcal{D}$ , and an original sample $x$ with target label $y$", + "bbox": [ + 511, + 241, + 890, + 286 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: The maximum perturbation $\\epsilon$ , the step size $\\alpha$ , the number of epochs $T$ and $\\mathrm{epochs}_{sub} T_{sub}$", + "bbox": [ + 511, + 287, + 887, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: The decay factor $\\mu$ , scaling level $m$ , and $n$ random target class samples $Z$ ( $z \\in Z$ )", + "bbox": [ + 511, + 316, + 895, + 345 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Input: The competition ratio $\\beta$ , and topk classifications $k$", + "bbox": [ + 511, + 347, + 897, + 363 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Output: An adversarial example $x^{adv}$ .", + "bbox": [ + 513, + 363, + 772, + 377 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1: $g_0^z = 0$ ; $g_0 = 0$ ; $z_0^{adv} = z$ ; $x_0^{adv} = x$", + "2: for $t$ in range $(T)$ do", + "3: if $t\\% T_{sub} == 0$ then", + "4: Reselect $n$ random target class samples $Z$", + "5: end if", + "6: Get the inverse target gradient for each $\\tilde{g}_t^z$ of $n$ random target class samples by Eq.(2)", + "7: Get the momentum inverse target gradient for each $g_{t+1}^{z}$ by Eq.(3)", + "8: Update each $z_{t+1}^{adv}$ by Eq.(4)", + "9: Get the set $Q$ of top- $k$ classifications by $f(x_{t}^{adv})$", + "10: Calculate the integrate gradient of competition examples by:" + ], + "bbox": [ + 519, + 388, + 903, + 585 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {D} \\big (\\delta_ {i} \\cdot (x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}), y \\big)\n$$\n", + "text_format": "latex", + "bbox": [ + 555, + 598, + 903, + 638 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "11: Get the momentum gradient by Eq.(6)", + "12: Update the adversarial example $x_{t + 1}^{adv}$ by Eq.(7)", + "13: end for", + "14: return $x^{adv} = x_T^{adv}$ ." + ], + "bbox": [ + 517, + 656, + 867, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "transferable targeted AEs. As shown in Figure 3, we present two 'cock'-targeted examples as representatives: targeted AEs generated using ResNet50 with strong and weak transferability, from left to right. In (a) and (c), we have AEs that can only fool fewer than three black-box models into predicting the target class, while (b) and (d) show AEs capable of misleading over twelve black-box models to the target class. Table 1 presents the top-5 average confidence scores for each AE on the successfully misled black-box models. It can be seen that the average confidence for the target la", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3719", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "bel (top-1) in highly transferable AEs is significantly higher than that in low-transferability AEs. Notably, highly transferable AEs not only exhibit higher top-1 confidence but also have lower confidence scores for subsequent rankings (rank-2 and beyond) compared to low-transferability AEs, with each subsequent rank showing an even larger gap.", + "bbox": [ + 89, + 90, + 480, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Method of SDS", + "text_level": 1, + "bbox": [ + 89, + 186, + 243, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on this observation, we naturally propose a hypothesis by inverse inference: increasing the gap between the target class and other classes under a top- $k$ setting can enhance adversarial transferability. Therefore, the key question becomes how to widen this gap. To address this, we devised two strategies from different perspectives, as shown in the following formulas:", + "bbox": [ + 89, + 207, + 483, + 313 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {u n t g} (x) = \\frac {1}{k - 1} \\sum_ {q \\in Q} f _ {q} (x) \\quad \\text {s . t .} \\quad y \\notin Q, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 320, + 482, + 359 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {S 1} (x _ {t} ^ {a d v}, y) = f _ {y} (x _ {t} ^ {a d v}) - f _ {u n t g} (x _ {t} ^ {a d v}), \\qquad (9)\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 369, + 483, + 392 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} _ {S 2} \\left(x _ {t} ^ {a d v}, y\\right) = \\frac {f _ {y} \\left(x _ {t} ^ {a d v}\\right) - f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right)}{\\left\\| \\nabla_ {x _ {t} ^ {a d v}} f _ {y} \\left(x _ {t} ^ {a d v}\\right) - \\nabla_ {x _ {t} ^ {a d v}} f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right) \\right\\| _ {2}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 412, + 483, + 469 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The first strategy (S1) maximizes the target class logit value relative to the aggregated logits of other classes within the top- $k$ setting, enhancing target class features in the latent feature space and diminishing those of similar other classes, thus widening the latent feature gap, as shown in Equation (9). The second strategy (S2), inspired by DeepFool's approach to calculating point-to-hyperplane distances [19], guides AEs to deeply infiltrate the target class region in multi-dimensional geometric space while maintaining a global distance from the aggregated boundaries of the closest non-target regions, as detailed in Equation (10). In these equations, $f(\\cdot)$ represents the classifier, $Q$ denotes the set of top- $k$ classes (excluding $y$ ), and $f_{q}(x)$ and $f_{y}(x)$ represent the output values of $f(x)$ corresponding to $q$ and $y$ , respectively. The algorithm of ITDS is shown in Algorithm 1.", + "bbox": [ + 89, + 470, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 709, + 223, + 726 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first outline the experimental setup, then evaluate the targeted-attack performance of both competition-based and transformation-based methods separately. Afterward, we evaluate the ability of our ITDS to break through defenses and conduct ablation studies.", + "bbox": [ + 89, + 734, + 483, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 818, + 281, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset. To ensure rigorous evaluation, we randomly selected 2,000 high-quality images from 1,000 classes in the ILSVRC2012 validation set [21]. These images were correctly identified by all models tested, including CNNs, ViTs", + "bbox": [ + 89, + 840, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and adversarial training models, using intersection extraction, with more than $90\\%$ top-1 average confidence. The target class images of our method were randomly drawn from the corresponding category in the ILSVRC2012 dataset.", + "bbox": [ + 511, + 90, + 903, + 151 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. We choose four normally pre-trained CNNs, i.e., ResNet50 (RN50) [9], VGG16 [22], MobileNet-v3-large (MN-v3) [11], RegNet-y-32gf (RegN) [20] as surrogate (white-box) models. Concurrently, we choose four white-box independent normally pre-trained CNNs, i.e., Inception-v3 (Inc-v3) [24], ResNet101 (RN101), DenseNet161 (DN161) [12], EfficientNet-b7 (ENet) [25], and four ViTs, i.e., VisionTransformer-b-16 (ViT) [5], DeiT-base (DeiT) [26], ConViT-base (ConViT) [6] and PiT-base (PiT) [10] to serve as target black-box models.", + "bbox": [ + 511, + 152, + 905, + 303 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We adopt six challenging adversarial attacks as our baselines, which have released the code and provide the necessary parameters for reproduction, i.e., Competition-based: Admix [30], CFM [2], Transformation-Based: DIM [34], ODIM [1], SIA [31], and BSR [29]. All the baselines are combined with MI-FGSM.", + "bbox": [ + 511, + 304, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Defenses. To further demonstrate the effectiveness of ITDS, we consider four additional advanced defense methods, i.e., JPEG [8], FD [17], GDMP [28], Score-Opt (SO) [36], and three adversarial training defense models [27], i.e., $\\mathrm{Def}_{\\text{lin}f4}$ ( $\\mathrm{Def}_4$ ), $\\mathrm{Def}_{\\text{lin}f8}$ ( $\\mathrm{Def}_8$ ) and $\\mathrm{Def}_{l2-3}$ ( $\\mathrm{Def}_{l2}$ ) from ResNet50, which are proven to be robust to adversarial attacks on ImageNet datasets.", + "bbox": [ + 511, + 396, + 905, + 501 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Attack setting. For all methods, we set the maximum perturbation of $\\epsilon = 16$ , the number of epochs $T = 100$ , the step size $\\alpha = 1.6$ . We adopt the decay factor $\\mu = 1.0$ for all the methods that employ the MI-FGSM, and a transformation probability of $p = 0.7$ for DIM and ODIM. Furthermore, all parameters for Admix, SIA, CFM, and BSR are set as described in their respective published papers. For our ITDS, we set the number of epochs $_{sub}$ to $T / 10$ , scaling level $m = 5$ with $\\delta_{i} = 1 / 2^{i}$ , the number of random target class samples $n = 3$ , the competition ratio $\\beta = 1.5$ , and $k = 10$ of the top- $k$ . For each AE of the trial, the target label is randomly chosen from the other 999 classes excluding its true label, and the final results are averaged over three trials.", + "bbox": [ + 511, + 502, + 905, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Evaluation on Competition-based Attacks", + "text_level": 1, + "bbox": [ + 511, + 710, + 870, + 728 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first evaluate the attack performance of various competition-based attacks. Since the FIA and RPA introduced in related work are only applicable to non-targeted attacks, and FFT must be combined with existing methods to be usable, we ultimately chose Admix and CFM as competitors for ITDS. To demonstrate the versatility of our approach, the surrogate models we use to craft AEs are structurally independent of each other. The targeted attack success rates (TASRs) are shown in Table 2, where the upper row indicates the targeted model under attack, and the left column lists the tested surrogate models. To test the feasi-", + "bbox": [ + 511, + 734, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3720", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ae51110f14759bc3b8f12b0fce6a8707d1cf60214ef10e5c46ec6eac4ef8abe3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50Admix100*7.92.14.91.638.619.80.60.30.10.20.16.9
CFM100*45.017.735.59.587.760.75.62.01.51.53.924.6
ITDSfs100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDSgs100*46.235.049.328.487.680.417.910.04.14.46.633.6
VGG16Admix4.388.4*1.47.70.70.94.10.50.10.10.10.51.8
CFM3.986.1*1.28.40.81.02.80.60.10.20.10.31.7
ITDSfs18.8100*12.020.68.07.618.45.61.00.60.51.58.6
ITDSgs18.7100*12.019.99.08.218.55.71.40.50.81.98.7
MN-v3Admix0.60.499.3*0.30.60.30.40.40.20.10.20.10.3
CFM10.48.9100*10.47.49.67.714.37.22.74.03.17.8
ITDSfs16.49.6100*15.211.816.015.219.814.53.84.24.411.9
ITDSgs16.79.9100*14.911.715.915.616.814.43.24.83.811.6
RegNAdmix1.41.11.695.8*0.40.51.60.50.30.20.20.40.7
CFM21.423.314.192.3*4.412.119.69.01.41.21.15.910.3
ITDSfs38.216.025.0100*17.824.840.623.911.17.67.214.020.5
ITDSgs41.217.726.899.9*19.727.743.125.013.39.68.417.422.7
", + "bbox": [ + 106, + 88, + 893, + 310 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5448009a179afc5106ea3a0154deb085daaebbc39b9e3a8e359315a955701e30.jpg", + "table_caption": [ + "Table 2. TASRs (%) on twelve pre-trained models with various competition-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks, and BAvg means average black-box TASR. ITDS $_{fs}$ and ITDS $_{gs}$ represent the integration of S1 and S2 in SDS, respectively, and the boldface represents their results." + ], + "table_footnote": [], + "table_body": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50DIM100*33.614.131.915.365.059.113.53.11.92.63.822.2
ODIM99.9*52.833.249.837.570.265.830.114.59.210.415.535.3
SIA100*86.856.386.540.499.093.534.330.621.724.040.355.7
BSR100*87.648.787.231.097.692.827.818.116.916.241.551.4
Admix-DIM100*51.925.553.629.183.877.228.88.04.35.110.134.3
CFM-DIM99.7*68.447.966.049.288.479.541.918.012.514.920.046.0
ITDS-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
VGG16DIM12.985.8*3.418.22.93.510.63.50.20.30.21.45.2
ODIM33.578.8*10.028.211.213.125.113.62.61.41.34.113.1
SIA50.481.7*13.555.67.320.543.77.72.11.21.69.219.3
BSR24.384.7*3.834.61.77.525.11.40.40.30.32.19.2
Admix-DIM20.288.0*6.823.95.75.416.08.00.80.50.52.18.2
CFM-DIM19.286.3*4.123.73.95.914.06.10.40.60.41.67.2
ITDS-DIM42.0100*24.834.825.719.835.324.98.52.83.27.820.9
MN-v3DIM6.32.599.2*4.96.94.55.516.26.01.82.52.35.4
ODIM24.917.399.4*21.222.122.823.236.120.09.611.611.220.0
SIA32.721.498.7*29.214.126.125.424.219.17.910.610.320.1
BSR22.012.598.9*18.38.714.413.015.911.04.46.57.812.2
Admix-DIM18.06.399.4*11.914.812.113.032.012.53.75.54.47.5
CFM-DIM34.724.999.8*31.629.433.427.844.228.111.814.211.626.5
ITDS-DIM49.527.3100*39.939.244.941.658.046.718.117.013.736.0
RegNDIM24.815.212.492.9*9.312.331.421.43.63.63.013.013.6
ODIM51.343.331.290.3*29.235.656.949.315.712.911.626.133.0
SIA71.272.138.695.7*15.648.072.033.119.717.918.647.741.3
BSR49.063.022.194.1*6.124.947.715.67.08.88.938.026.4
Admix-DIM39.425.220.692.6*14.823.147.733.17.67.17.322.922.6
CFM-DIM71.561.950.893.0*38.361.072.361.622.521.722.143.347.9
ITDS-DIM82.250.568.999.9*64.673.184.975.656.543.040.854.763.2
", + "bbox": [ + 96, + 378, + 906, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. TASRs (%) on twelve pre-trained models with various transformation-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks. ITDS-DIM represents the combination of ITDS, which integrates S1, with DIM, and the boldface represents the results of this combination.", + "bbox": [ + 89, + 761, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "bility of the two strategies in SDS, ITDS is implemented in two forms: $\\mathrm{ITDS}_{fs}$ for S1 (feature space) and $\\mathrm{ITDS}_{gs}$ for S2 (geometric space).", + "bbox": [ + 89, + 824, + 480, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It can be seen that, compared to other competition-based attacks, ITDS demonstrates an overwhelming advantage.", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Taking $\\mathrm{ITDS}_{gs}$ as an example, its average black-box TASR (BAvg.) is $9.0\\%$ , $7.0\\%$ , $3.8\\%$ , and $12.4\\%$ higher than the existing SOTA method CFM on RN50, VGG16, MN-v3, and RegN, respectively. From the perspective of cross-architecture models, the largest gaps in average black-box", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3721", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/f0571c4951201106970db8f09ee8c8318777e8034fd63c4a854fe845df1627b0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
ITDS1100*41.524.340.418.984.670.211.55.52.92.74.227.9
ITDS1+S100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDS1-DIM99.9*71.158.673.356.892.587.148.435.118.819.123.653.1
ITDS1+S-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
", + "bbox": [ + 116, + 87, + 885, + 154 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Evaluation of the effectiveness of two modules in ITDS, where 'I' represents ITC and 'S' represents SDS. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. Boldface represents the best results. Note that single SDS requires combination with existing attack methods, which we discuss in the appendix.", + "bbox": [ + 88, + 162, + 906, + 205 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "TASR between it and CFM on CNNs and ViTs are $13.9\\%$ and $9.7\\%$ on RegN, respectively. It should be noted that the white-box attack performance of our ITDS is also significantly superior, being the best in terms of white-box TASR in any surrogate model, especially on VGG16 and RegN, where it is $13.9\\%$ and $7.6\\%$ higher than CFM, respectively. In the comparison between $\\mathrm{ITDS}_{fs}$ and $\\mathrm{ITDS}_{gs}$ , the results seem to be evenly matched, with each showing their own merits in the three trials. Therefore, we believe that both strategies are viable, which precisely proves the effectiveness and diversity of the proposed SDS principle.", + "bbox": [ + 89, + 224, + 485, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Transformation-based Attacks", + "text_level": 1, + "bbox": [ + 89, + 406, + 473, + 422 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Currently, transformation-based methods are the main approach to adversarial attacks and are the most effective way to improve the transferability of attacks. Numerous studies [2, 4, 16] have shown that non-transformation-based methods combined with DIM can further enhance the transferability of crafted AEs. This combination approach has become a common standard in industry, frequently compared to transformation-based methods for fairness. In this section, we evaluate the TASR of transformation-based and DIM-combined competition-based adversarial attacks. The former includes DIM, ODIM, SIA, and BSR, while the latter includes Admix-DIM, CFM-DIM, and ITDS-DIM. It should be noted that, due to the similar performance of S1 and S2 in SDS, in all subsequent experiments involving ITDS-DIM, we selected $\\mathrm{ITDS}_{fs}$ as the representative to combine with DIM. The results are shown in Table 3.", + "bbox": [ + 91, + 430, + 483, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "It can be observed that all competition-based methods show a significant improvement in attack performance after being combined with DIM. Compared to the strongest competition-based attack CFM-DIM, ITDS-DIM achieved average black-box TASR improvements of $15.7\\%$ , $13.7\\%$ , $9.5\\%$ , and $15.3\\%$ on RN50, VGG16, MN-v3, and RegN, respectively. The maximum average black-box TASR gap between the two on CNNs and ViTs is $16.1\\%$ on RN50 and $21.4\\%$ on RegN. Even when compared with the SOTA transformation-based attacks, our method is not inferior, and even surpasses them in all aspects. Compared to the strongest transformation-based attack SIA, ITDS-DIM achieved average black-box TASR improvements of $15.9\\%$ and $21.9\\%$ on MN-v3 and RegN, respectively. The maximum average black-box TASR gap between the two on", + "bbox": [ + 89, + 674, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d8110ee373581f5fa85dee9e4c5b891eaa13c9a4d31273436d689de330d6072c.jpg", + "image_caption": [ + "CNNs and ViTs is $21.3\\%$ and $22.8\\%$ on RegN, respectively." + ], + "image_footnote": [], + "bbox": [ + 514, + 253, + 702, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/222e52068242b59fa606527f05786f70506e1e28f7d2617921a261c2309ad180.jpg", + "image_caption": [ + "(b)" + ], + "image_footnote": [], + "bbox": [ + 718, + 253, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0e7e1b35e618e3529ce6f106e2176493d2789d024f353fb11df5f42fa526ae6f.jpg", + "image_caption": [ + "(a)", + "(c)", + "Figure 4. Ablation studies on the RN50 model. (a) - (d): TASRs $(\\%)$ on the other eleven models with the adversaries crafted by ITDS-DIM, where the default value for $n$ , $\\beta$ , $k$ and $T_{sub}$ are set to 3, 1.5, 10 and 10 respectively, when test parameters for each other." + ], + "image_footnote": [], + "bbox": [ + 516, + 467, + 702, + 662 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d06bb72df22cfc1b7e323325dfdc2b5bbd9605c872e77ef16058a9ac1299619.jpg", + "image_caption": [ + "(d)" + ], + "image_footnote": [], + "bbox": [ + 715, + 467, + 901, + 662 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 772, + 679, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first present a series of ablation studies conducted on RN50 to assess the impact of each parameter in ITDS on attack performance, including the number of random target class samples $n$ , the competitive ratio of the inverse target gradient $\\beta$ , the value of $k$ in top- $k$ , and the number of epochs $t_{sub}$ for non-targeted attack $T_{sub}$ . In all ablation experiments, unless otherwise specified, all parameters other", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3722", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/624889832acda06da70912c29c414fc4d5850ceea0d67c6b6c4a3aa42cfd56ef.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
AblationTarget Model
Norm (P1)MI (P2)RN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
✓7.1*0.51.00.40.40.80.30.20.20.20.10.10.4
✓100*72.866.175.865.190.188.254.737.216.917.423.755.3
✓✓100*59.741.366.152.890.991.345.127.814.516.221.147.9
100*79.370.382.869.695.694.662.044.423.624.731.661.7
", + "bbox": [ + 96, + 88, + 903, + 170 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/af8a74639bfa0ec6a84b2862576b8c3de3eec5592eea996d0810466ca00aa4c8.jpg", + "table_caption": [ + "Table 5. TASRs (%) of ITDS-DIM by ablating inner modules of the ITC. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. In which Norm and MI correspond to P1 and P2 in Section 3.1.2, respectively. Boldface represents the best results." + ], + "table_footnote": [ + "Table 6. TASRs (\\%) of breaking-through four defense methods and three adversarial training defense models. The results for defense methods are based on the average performance of AEs crafted on RN50, VGG16, MN-v3, and RegN respectively, while the results for defense models are based on the performance of AEs crafted by the corresponding models. Bolder for the best." + ], + "table_body": "
AttackJPEGFDSOGDMPDef4Def8Defl2
SIA83.289.844.31.539.014.054.7
Admix-DIM80.092.947.71.934.814.553.1
CFM-DIM82.592.651.54.125.28.745.5
ITDS-DIM99.499.991.331.250.225.666.8
", + "bbox": [ + 98, + 227, + 478, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "than the test parameters are fixed according to the experimental setup.", + "bbox": [ + 89, + 410, + 482, + 440 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As shown in Figure 4a, TASR increases with $n$ , as the diversity of target class images directly affects attack performance. However, the curve levels off after $n = 3$ , and considering computational costs, we selected $n = 3$ . In Figure 4b, most models achieve optimal performance at $\\beta = 1.5$ with a few exceptions (e.g., MN-v3). We believe a moderate competitive gradient provides an \"exercise\" effect, while an overly strong gradient disrupts input features, reducing attack performance. In Figure 4c, optimal performance generally occurs at $k = 5$ or $k = 10$ . We posit that an overly small $k$ under-covers classes and fails to enlarge global margins, whereas an overly large one blurs the optimization. At $k = 10$ , cross-architecture transferability is even better. In Figure 4d, evaluating epochsub $(T_{sub})$ for non-targeted attacks, we found that larger values do not always lead to better performance. As discussed in the ITC section, inverse target gradient perturbations generated at each $T_{sub}$ have diverse potential competitive features, boosting attack performance. At $T_{sub} = 1$ , though it involves more target class samples, it lacks deeper latent features of the target class, limiting attack performance. Beyond a certain $T_{sub}$ value, missing target class features reduce the diversity of inverse gradients, causing performance to regress.", + "bbox": [ + 91, + 444, + 482, + 790 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Furthermore, to verify the effectiveness of the ITC and SDS components in ITDS, we conducted experimental evaluations as shown in Table 4. The results indicate that ITDS using only ITC has already surpassed the attack performance of CFM in Table 2. Regardless of whether DIM is combined or not, SDS can bring additional performance improvements to ITDS. These findings collectively support the", + "bbox": [ + 89, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "key roles of ITC and SDS in ITDS.", + "bbox": [ + 511, + 229, + 746, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify the impact of internal modules within the ITC on attack performance, we conducted ablation experiments as shown in Table 5. It can be observed that the use of N-norm may limit the scale of gradient updates, which could result in generated AEs lacking sufficient aggressiveness. On the contrary, if we do not use N-norm (P1), gradient updates will gain greater freedom, which helps to explore the model's decision boundary more deeply, and may thus produce more aggressive AEs. Simultaneously, momentum gradients, as a technique for accumulating previous gradients, will weaken the intensity of inverse target features in the current iteration. Since we wish to maintain the integrity of inverse target features and avoid any normalization, the inverse target momentum gradients in each round will compete with the AE when combined with P1, leading to excessive deformation of AE features, which is disastrous for implementing precisely targeted attacks.", + "bbox": [ + 511, + 244, + 906, + 503 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Evaluation on Advanced Defenses", + "text_level": 1, + "bbox": [ + 511, + 508, + 808, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we evaluate the capabilities of various attacks to break through different defense mechanisms. As shown in Table 6, our method exhibits exceptional robustness against a wide range of defense strategies and models. For example, when the Score-OPT method is employed, the average TASR of our approach is nearly $40\\%$ higher than that of CFM-DIM. Furthermore, our method achieves the best performance in all three defense models evaluated.", + "bbox": [ + 511, + 531, + 905, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces a novel method named ITDS, which significantly improves the transferability of targeted AEs by incorporating the ITC mechanism and the SDS strategy. Experimental results demonstrate that ITDS achieves a substantial increase in average transferable TASRs across various mainstream CNN and ViT models, outperforming SOTA competition-based and transformation-based methods. Our ablation studies further confirmed the key roles of the ITC and SDS components in enhancing attack performance. Moreover, ITDS exhibits superior effectiveness in countering multiple defense methods and models. These findings confirm the effectiveness and practicality of ITDS in enhancing the transferability of targeted AEs and provide a new perspective for the field of adversarial research.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3723", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by funding from the National Natural Science Foundation of China (NSFC) under Grants 62272486, 62172154, and 62372473, the Natural Science Foundation of Hunan Province under Grant 2023JJ70016, the Hong Kong Research Grants Council (RGC) General Research Fund under Grants 152244/21E, 152169/22E, 152228/23E, and 162161/24E, the Research Impact Fund under Grant R5011-23F, the Collaborative Research Fund under Grant C1042-23GF, the NSFC/RGC Collaborative Research Scheme under Grant CRS_HKUST602/24, the Areas of Excellence Scheme under Grant AoE/E-601/22-R, and the InnoHK initiative (HKGAI). Professor Shigeng Zhang is the corresponding author of this paper.", + "bbox": [ + 89, + 116, + 485, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 330, + 187, + 345 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2022. 1, 2, 3, 5", + "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 1, 3, 5, 7", + "[3] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 4", + "[4] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 7", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5", + "[6] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference on Machine Learning, pages 2286-2296. PMLR, 2021. 5", + "[7] Goodfellow et al. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2", + "[8] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2018. 5" + ], + "bbox": [ + 99, + 354, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5", + "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 5", + "[11] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5", + "[12] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 5", + "[13] Kurakin et al. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2", + "[14] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Foolmix: Strengthen the transferability of adversarial examples by dual-blending and direction update strategy. IEEE Transactions on Information Forensics and Security, 2024. 1", + "[15] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Ucg: A universal cross-domain generator for transferable adversarial examples. IEEE Transactions on Information Forensics and Security, 2024. 1", + "[16] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 7", + "[17] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature distillation: Dnn-oriented jpeg compression against adversarial examples. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 860-868. IEEE, 2019. 5", + "[18] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549-566. Springer, 2022. 1, 2", + "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 5", + "[20] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 5", + "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3724", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5", + "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[23] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1", + "[24] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5", + "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 5", + "[26] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5", + "[27] Francisco Utrera, Evan Kravitz, N Benjamin Erickson, Rajiv Khanna, and Michael W Mahoney. Adversarily-trained deep nets transfer better: Illustration on image classification. arXiv preprint arXiv:2007.05869, 2021. 5", + "[28] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 5", + "[29] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 1, 2, 5", + "[30] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 1, 2, 5", + "[31] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 5", + "[32] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2", + "[33] Juanjuan Weng, Zhiming Luo, Shaozi Li, Nicu Sebe, and Zhun Zhong. Logit margin matters: Improving transferable targeted adversarial attack by logit calibration. IEEE Transactions on Information Forensics and Security, 18:3561-3574, 2023. 2", + "[34] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 1, 2, 5" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[35] Zeng et al. Enhancing targeted transferability via feature space fine-tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4475-4479. IEEE, 2024. 3", + "[36] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 36:51810-51829, 2023. 5", + "[37] Qilong Zhang, Xiaodan Li, YueFeng Chen, Jingkuan Song, Lianli Gao, Yuan He, et al. Beyond imagenet attack: Towards crafting adversarial examples for black-box domains. In International Conference on Learning Representations, 2022. 1", + "[38] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quaxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, pages 1672-1678, 2022. 2", + "[39] Zhao et al. On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems, 34:6115-6128, 2021. 2, 3", + "[40] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 1" + ], + "bbox": [ + 516, + 92, + 903, + 444 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3725", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_model.json b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..da02caa60ee886806f291086d481b1b976f81770 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_model.json @@ -0,0 +1,2266 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.099, + 0.13, + 0.9, + 0.177 + ], + "angle": 0, + "content": "Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.806, + 0.223 + ], + "angle": 0, + "content": "Zhankai Li\\(^{1}\\), Weiping Wang\\(^{1}\\), Jie Li\\(^{1}\\), Shigeng Zhang\\(^{1}\\), Yunan Hu\\(^{1}\\), Song Guo\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.228, + 0.602, + 0.246 + ], + "angle": 0, + "content": "1Central South University" + }, + { + "type": "text", + "bbox": [ + 0.277, + 0.246, + 0.719, + 0.265 + ], + "angle": 0, + "content": "2The Hong Kong University of Science and Technology" + }, + { + "type": "text", + "bbox": [ + 0.146, + 0.267, + 0.851, + 0.282 + ], + "angle": 0, + "content": "janlzk7@gmail.com, {wpwang, lijie55, sgzhang, hyn318}@csu.edu.cn, songguo@ust.hk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.317, + 0.327, + 0.332 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.486, + 0.681 + ], + "angle": 0, + "content": "In the field of AI security, deep neural networks (DNNs) are highly sensitive to adversarial examples (AEs), which can cause incorrect predictions with minimal input perturbations. Although AEs exhibit transferability across models, targeted attack success rates (TASRs) are low due to differences in feature dimensions and decision boundaries. To enhance targeted AE transferability, we propose a novel approach using Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS) in the optimization process. Specifically, we employ a siamese-network-like framework to generate both non-targeted and targeted AEs. The ITC mechanism applies non-targeted adversarial gradients each epoch to impede the optimization of targeted perturbations, thereby improving robustness. Additionally, a top-k SDS strategy guides AEs to penetrate target class regions in the latent space while distancing from nontargeted regions, achieving optimal transferability. Compared to state-of-the-art competition-based attacks, our method significantly improves transferable TASRs by \\(16.1\\%\\) and \\(21.4\\%\\) on mainstream CNNs and ViTs, respectively, and demonstrates superior defense-breaking capabilities. Our code is available here." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.222, + 0.725 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Adversarial attacks involve adding small, often imperceptible perturbations to input data, which can lead otherwise high-performing models to produce incorrect outputs. The concept of adversarial attacks was first introduced by [23] and further developed by [7], who demonstrated the vulnerability of deep neural networks (DNNs) to these perturbations. With the rapid advancement of DNNs in recent years, the security risks revealed by adversarial attacks have become a critical issue. Adversarial attacks are now widely used to assess model security, improve model robustness, and drive research in defense technologies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.318, + 0.908, + 0.681 + ], + "angle": 0, + "content": "However, the significance of adversarial attacks extends beyond these applications, as they have inadvertently become a widely used defense mechanism in real-world scenarios. A prime example is Google's login verification system, reCAPTCHA, which aims to distinguish real users from bots to prevent abuse. Adversarial examples (AEs) are one of the primary image verification methods employed, as illustrated in Figure 1. This raises an important question: is it secure enough? Since we do not know the bot's model architecture or training domain, the transferability of AEs becomes especially crucial. Most current research on adversarial transferability focuses on non-targeted attacks [14, 40], even extending to cross-architecture and cross-domain attacks [15, 37], achieving notable performance. In contrast, targeted attacks face a significant increase in difficulty due to the substantial changes in decision boundaries, and as a result, they still maintain a relatively low success rate. However, for login verification, given that the output category of non-targeted attacks is uncontrollable, and even the original verification images without perturbation may have \"low-quality\" issues (leading to misidentification by bots), in such cases, targeted AEs are often safer and more reliable than non-targeted AEs that may cause the deployed bots to misjudge into the ground truth category." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Currently, effective targeted attack methods are primarily based on transformation [1, 18, 29, 34], achieving transferability through input augmentation. However, the relentless pursuit and stacking of a greater variety of transformations, although enhancing transferability to some extent, does not guarantee robust performance in white-box attacks. Moreover, this stacking approach based on affine and perspective transformations offers limited insight into the complexity and vulnerability of DNNs. Recently, competition-based attacks have been proposed [2, 30], offering a unique perspective. Their advantage lies in compelling AEs to consider the more complex and diverse feature representations within DNNs during their generation, simulating various noise interferences and attack strategies," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3716" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.12, + 0.089, + 0.454, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.486, + 0.32 + ], + "angle": 0, + "content": "Figure 1. Example of Google's web login verification system, reCAPTCHA. It determines whether a user is human or a bot based on whether they can correctly match text to corresponding images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.483, + 0.452 + ], + "angle": 0, + "content": "and promoting the learning of more robust feature combinations. The competitive mechanism, by delving into the model's decision boundaries, helps to uncover the latent vulnerabilities of DNNs and provides a new perspective for understanding and improving the model's generalization capabilities. However, no matter what the attack method is, it cannot conceal the widespread insufficiency in the transferability performance of current targeted adversarial attacks." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.455, + 0.485, + 0.759 + ], + "angle": 0, + "content": "To address this challenge, we propose a new targeted adversarial attack method called ITDS, which consists of two components: Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS), aimed at enhancing the transferability of targeted AEs. In the first part, we adopt a siamese-network-like framework, innovatively leveraging the ground truth gradients of multiple target class non-targeted AEs at intermediate stages to exert inverse \"competitiveness\" on non-target class targeted AEs during the optimization process. In the second part, to tackle the multi-dimensional complexity of decision boundaries in targeted attacks, we introduce a top-\\(k\\) SDS strategy to seek stable perturbation update directions. This guides AEs to penetrate target class regions within the latent multidimensional space while globally distancing from several closest non-target region boundaries. The experimental results demonstrate that the proposed ITDS method has achieved significant advantages in enhancing the transferability and defense-breaking capabilities of targeted AEs, surpassing the existing state-of-the-art (SOTA) methods." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.769, + 0.235, + 0.785 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.794, + 0.485, + 0.84 + ], + "angle": 0, + "content": "In this section, we will briefly overview the background of adversarial attacks. Then focus on various targeted attack techniques that demonstrate effective transferability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.283, + 0.863 + ], + "angle": 0, + "content": "2.1. Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Goodfellow et.al first designed a simple yet effective method called Fast Gradient Sign Method (FGSM) [7]. This" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "method generates AEs by making small adjustments to the input \\( x \\) (original sample) along the direction of the gradient, i.e., \\( x^{adv} = x + \\delta \\). In the case of targeted attacks, FGSM updates the input \\( x \\) in the direction that minimizes the classification loss for a given target class \\( y \\), with the objective function as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.19, + 0.905, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {x ^ {a d v}} \\mathcal {J} (x ^ {a d v}, y; \\theta) \\text {s . t .} \\| x - x ^ {a d v} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.222, + 0.905, + 0.266 + ], + "angle": 0, + "content": "where \\(\\theta\\) is the model parameters, \\(\\mathcal{J}\\) denotes the Cross-Entropy (CE) loss function, and the \\(\\ell_{\\infty}\\) norm is used to constrain the perturbation within the range \\(\\epsilon\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.907, + 0.403 + ], + "angle": 0, + "content": "I-FGSM [13] is an iterative extension of FGSM, which gradually refines adversarial perturbation with small step sizes. To address gradient vanishing issues in targeted attacks, [39] proposes using Logits loss instead of CE loss, significantly enhancing targeted attack effectiveness. Based on this, [33] further proposes downscaling logit calibration with a temperature factor and an adaptive margin. While these methods provide a foundation for enhancing targeted attack effectiveness, their transferability remains limited." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.411, + 0.808, + 0.426 + ], + "angle": 0, + "content": "2.2. Transferable Adversarial Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.538 + ], + "angle": 0, + "content": "In this section, we introduce advanced transferable attacks that can successfully deceive other models using AEs generated on a single surrogate model without additional training. We categorize these methods into transformation-based and competitive-based attacks, focusing primarily on transferable targeted attacks, along with certain non-targeted attacks that can also be adapted for targeted attacks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Transformation-Based Attacks. DIM [34] is a representative transformation-based attack, which resizes and pads images with a certain probability at each step of the multi-step attack process. SSM [18] applies frequency domain augmentations through spectrum transformations on the input. ODIM [1] method significantly improves the targeted attack transferability by mapping input images to random 3D objects and applying varied rendering techniques. SIA [31] generates structurally diverse AEs by applying various random transformations to the images. Both ODIM and SIA integrate DIM into their implementation principles and share a similar lineage in transformation strategies. BSR [29] randomly shuffles and rotates the blocks within images, calculates the ensemble gradients of this new set of images, reduces attention heatmap differences across various models, and significantly enhances transferability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Competition-based Attacks. Admix [30] combines features from different categories in the input domain, creating multiple images for gradient computation, and enhancing transferability without degrading white-box performance. FIA [32] employs aggregated gradients to disrupt essential object-aware features while suppressing model-specific features, boosting AE transferability. RPA [38] adopts a similar approach to FIA, introducing random patch transforma" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3717" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.243 + ], + "angle": 0, + "content": "tions to benign images. To alleviate the overfitting dilemma common in an AE crafted by simple iterative attacks, FFT [35] encourages features conducive to the target class while discouraging those associated with the original class in an intermediate layer of the source model. CFM [2] introduces a competitive strategy, crafting targeted adversarial perturbations with diverse features by introducing two types of competitor noises: adverse perturbations toward different target classes and friendly perturbations towards the correct class, achieving SOTA results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.256, + 0.283, + 0.273 + ], + "angle": 0, + "content": "3. The Design of ITDS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.484, + 0.343 + ], + "angle": 0, + "content": "This section consists of two parts: the Inverse Target Gradient Competition mechanism and the Spatial Distance Stretching strategy. In each part, we first explain the motivation and then propose the corresponding method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.351, + 0.462, + 0.367 + ], + "angle": 0, + "content": "3.1. Inverse Target Gradient Competition (ITC)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.373, + 0.266, + 0.387 + ], + "angle": 0, + "content": "3.1.1. Motivation of ITC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.485, + 0.65 + ], + "angle": 0, + "content": "The motivation for the first part of our research arises from a question: What is a competition mechanism? From existing work related to competition mechanisms, we summarize it as a way to enhance the transferability of AEs by promoting diversified attack strategies through simulating various noise interferences. Since noise interference is involved, the essence of competitiveness can be interpreted as an inverse force. For targeted attacks, could we identify the most competitive noise? We believe this is the key issue. To address this critical question, we turn to target class samples and consider the gradients of the target label for these samples. These gradients essentially represent the optimal force linking target class samples with the target label, which can be used in both forward and inverse directions. Therefore, we have a bold hypothesis: when using this gradient in inverse, it not only effectively generates non-targeted AEs but also serves as the optimal competitiveness for targeted attacks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.245, + 0.67 + ], + "angle": 0, + "content": "3.1.2. Method of ITC" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Figure 2 shows the overall framework of our Inverse Target Gradient Competition mechanism, which is jointly driven by both non-targeted and targeted attacks within a siamese-network-like architecture. Our strategy is to utilize the inverse target gradient to introduce resistance during the optimization process of targeted AEs. This inverse target gradient can be generated for any samples; but which inverse target gradient should we choose? As mentioned earlier, we seek the strongest resistance, which is the inverse gradient of the target label with respect to the target class sample. The approach of generating the inverse target gradient is illustrated in the lower part of Figure 2, where we use 'cock' as the target label as an example. In each sub-epochs (epochssub), multiple target class images that are correctly classified by the siamese classifier are selected to generate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.288 + ], + "angle": 0, + "content": "non-targeted AEs. The optimization goal is to minimize the logit value of the target label. We set \\(\\mathrm{epoch}_{sub}\\) to be one-tenth of the total epochs, as we believe that target class samples need a certain number of iterations to transform into robust non-targeted AEs. Accordingly, the non-targeted AEs generated in each \\(\\mathrm{epoch}_{sub}\\) contain different levels and increasingly deeper latent features of target class. Therefore, we consider the inverse target gradient perturbations generated in each \\(\\mathrm{epoch}_{sub}\\) to also have diverse potential competitive characteristics, and during the generation of nontargeted AEs in each epoch, multiple inverse target perturbations are generated to enrich competitive characteristics. The formulas are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.622, + 0.299, + 0.907, + 0.32 + ], + "angle": 0, + "content": "\\[\n\\tilde {g} _ {t} ^ {z} = \\nabla_ {z _ {t} ^ {a d v}} \\mathcal {L} (z _ {t} ^ {a d v}, y; \\theta), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.343, + 0.907, + 0.361 + ], + "angle": 0, + "content": "\\[\ng _ {t + 1} ^ {z} = \\mu \\cdot g _ {t} ^ {z} + \\tilde {g} _ {t} ^ {z}, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.379, + 0.907, + 0.4 + ], + "angle": 0, + "content": "\\[\nz _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {z, \\epsilon} \\{z _ {t} ^ {a d v} - \\alpha \\cdot \\operatorname {s i g n} (g _ {t + 1} ^ {z}) \\}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.404, + 0.907, + 0.481 + ], + "angle": 0, + "content": "where \\(z\\) is the target class sample, \\(\\tilde{g}_t^z\\) represents the inverse target gradient, and we follow [1, 2, 39] to use the Logit loss \\(\\mathcal{L}\\) (opposite to CE loss optimization). \\(g_{t + 1}^{z}\\) represents the momentum gradient without normalization for \\(z\\), and \\(z_{t + 1}^{adv}\\) is the non-targeted AE generated in each round." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.493, + 0.907, + 0.691 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.7, + 0.909, + 0.812 + ], + "angle": 0, + "content": "Figure 2. The overall framework of the ITC, e.g., combined with DIM, utilizes a siamese-network-like architecture to generate targeted and non-targeted AEs simultaneously. The upper part focuses on targeted attacks against non-target class samples, while the lower part addresses non-targeted attacks on target class samples. Note that for non-targeted attacks, the optimization condition utilizes the logits value of the target label, whereas the goal of the targeted attack is to stretch the spatial distance proposed in SDS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The process of generating targeted AEs is shown in the upper part of Figure 2, where the input is a non-target image, and the optimization goal is to maximize the spatial distance, which we introduce in the SDS of the next section. Each round of generated targeted AEs competes with" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3718" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "a certain proportion of inverse target gradients from different stages. The formulas are as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.13, + 0.482, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {L} \\left(\\delta_ {i} \\cdot \\left(x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}\\right), y; \\theta\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.216, + 0.198, + 0.483, + 0.216 + ], + "angle": 0, + "content": "\\[\ng _ {t + 1} = \\mu \\cdot g _ {t} + \\bar {g} _ {t} ^ {\\text {c p t}}, \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.233, + 0.483, + 0.252 + ], + "angle": 0, + "content": "\\[\nx _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {x, \\epsilon} \\left\\{x _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t + 1}\\right) \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.255, + 0.484, + 0.316 + ], + "angle": 0, + "content": "where \\( n \\) represents the number of target class samples in per epochs \\( _{sub} \\), \\( \\delta_{i} \\) and \\( m \\) are the scaling factor and level, \\( \\beta \\) is the competition ratio, and each round of targeted AEs competes with the corresponding inverse target gradient." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.512 + ], + "angle": 0, + "content": "There are two notable points. The first point (P1) is whether, for targeted or non-targeted attacks, the calculation of momentum gradients differs from the traditional MIFGSM [3], and we have not employed norm normalization (N-norm). One of the reasons is the absence of gradient explosion in our experimental evaluations, and more importantly, our experiments have shown that using N-norm would limit our attack performance. We believe that N-norm might restrict the magnitude of gradient updates, resulting in less aggressive generation of AEs. Without N-norm, gradient updates could become more flexible, allowing for more effective exploration of the model's decision boundary, thereby generating more aggressive AEs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.512, + 0.484, + 0.664 + ], + "angle": 0, + "content": "The second point (P2) is that we did not use the momentum gradient as the competitive gradient. The reason is that the momentum gradient is essentially a cumulative gradient. It weakens the strength of the inverse target features in the corresponding round. Since we want to maintain the integrity of the inverse target features, using it without any normalization operation in conjunction with P1 would cause excessive damage to the input features. This would be catastrophic for targeted attacks. These two points will also be discussed in subsequent ablation experiments." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.677, + 0.182, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.742, + 0.149, + 0.754 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.19, + 0.677, + 0.276, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.226, + 0.742, + 0.245, + 0.754 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.301, + 0.677, + 0.387, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.338, + 0.742, + 0.356, + 0.753 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.396, + 0.677, + 0.481, + 0.741 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.432, + 0.742, + 0.45, + 0.753 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.809 + ], + "angle": 0, + "content": "Figure 3. Comparison of targeted AEs with strong and weak transferability. Panels (a) and (c) show AEs with weak transferability, while (b) and (d) show AEs with strong transferability." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.829, + 0.385, + 0.847 + ], + "angle": 0, + "content": "3.2. Spatial Distance Stretching (SDS)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.852, + 0.267, + 0.866 + ], + "angle": 0, + "content": "3.2.1. Motivation of SDS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The motivation for the second part of our study arises from an interesting phenomenon commonly observed in highly" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.089, + 0.902, + 0.157 + ], + "angle": 0, + "content": "
LabelFigure12345
motor_scooter(a)42.0%5.5%4.3%3.6%3.2%
(b)83.3%4.9%2.1%1.1%0.3%
siamese_cat(c)49.6%5.6%5.3%5.1%5.0%
(d)90.5%2.1%1.3%0.7%0.2%
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.166, + 0.905, + 0.196 + ], + "angle": 0, + "content": "Table 1. The average top-5 confidence on black-box models successfully attacked by targeted AEs as shown in Figure 3." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.212, + 0.794, + 0.228 + ], + "angle": 0, + "content": "Algorithm 1 The ITDS Attack Algorithm" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.242, + 0.892, + 0.287 + ], + "angle": 0, + "content": "Input: Siamese classifier \\( f \\) with parameters \\( \\theta \\), Logit loss \\( \\mathcal{L} \\), our Distance loss \\( \\mathcal{D} \\), and an original sample \\( x \\) with target label \\( y \\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.888, + 0.317 + ], + "angle": 0, + "content": "Input: The maximum perturbation \\(\\epsilon\\), the step size \\(\\alpha\\), the number of epochs \\(T\\) and \\(\\mathrm{epochs}_{sub} T_{sub}\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.897, + 0.347 + ], + "angle": 0, + "content": "Input: The decay factor \\(\\mu\\), scaling level \\(m\\), and \\(n\\) random target class samples \\(Z\\) (\\(z \\in Z\\))" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.348, + 0.898, + 0.364 + ], + "angle": 0, + "content": "Input: The competition ratio \\(\\beta\\), and topk classifications \\(k\\)" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.364, + 0.773, + 0.378 + ], + "angle": 0, + "content": "Output: An adversarial example \\( x^{adv} \\)." + }, + { + "type": "text", + "bbox": [ + 0.523, + 0.389, + 0.79, + 0.407 + ], + "angle": 0, + "content": "1: \\( g_0^z = 0 \\); \\( g_0 = 0 \\); \\( z_0^{adv} = z \\); \\( x_0^{adv} = x \\)" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.407, + 0.687, + 0.421 + ], + "angle": 0, + "content": "2: for \\(t\\) in range \\((T)\\) do" + }, + { + "type": "text", + "bbox": [ + 0.526, + 0.421, + 0.704, + 0.435 + ], + "angle": 0, + "content": "3: if \\( t\\% T_{sub} == 0 \\) then" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.436, + 0.852, + 0.451 + ], + "angle": 0, + "content": "4: Reselect \\( n \\) random target class samples \\( Z \\)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.452, + 0.604, + 0.464 + ], + "angle": 0, + "content": "5: end if" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.466, + 0.905, + 0.496 + ], + "angle": 0, + "content": "6: Get the inverse target gradient for each \\(\\tilde{g}_t^z\\) of \\(n\\) random target class samples by Eq.(2)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.497, + 0.905, + 0.527 + ], + "angle": 0, + "content": "7: Get the momentum inverse target gradient for each \\( g_{t+1}^{z} \\) by Eq.(3)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.527, + 0.744, + 0.542 + ], + "angle": 0, + "content": "8: Update each \\( z_{t+1}^{adv} \\) by Eq.(4)" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.542, + 0.879, + 0.557 + ], + "angle": 0, + "content": "9: Get the set \\(Q\\) of top- \\(k\\) classifications by \\(f(x_{t}^{adv})\\)" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.558, + 0.905, + 0.587 + ], + "angle": 0, + "content": "10: Calculate the integrate gradient of competition examples by:" + }, + { + "type": "list", + "bbox": [ + 0.52, + 0.389, + 0.905, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "equation", + "bbox": [ + 0.557, + 0.599, + 0.905, + 0.64 + ], + "angle": 0, + "content": "\\[\n\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {D} \\big (\\delta_ {i} \\cdot (x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}), y \\big)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.657, + 0.812, + 0.672 + ], + "angle": 0, + "content": "11: Get the momentum gradient by Eq.(6)" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.672, + 0.869, + 0.688 + ], + "angle": 0, + "content": "12: Update the adversarial example \\( x_{t + 1}^{adv} \\) by Eq.(7)" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.688, + 0.597, + 0.7 + ], + "angle": 0, + "content": "13: end for" + }, + { + "type": "text", + "bbox": [ + 0.519, + 0.701, + 0.686, + 0.718 + ], + "angle": 0, + "content": "14: return \\( x^{adv} = x_T^{adv} \\)." + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.657, + 0.869, + 0.718 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "transferable targeted AEs. As shown in Figure 3, we present two 'cock'-targeted examples as representatives: targeted AEs generated using ResNet50 with strong and weak transferability, from left to right. In (a) and (c), we have AEs that can only fool fewer than three black-box models into predicting the target class, while (b) and (d) show AEs capable of misleading over twelve black-box models to the target class. Table 1 presents the top-5 average confidence scores for each AE on the successfully misled black-box models. It can be seen that the average confidence for the target la" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3719" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.184 + ], + "angle": 0, + "content": "bel (top-1) in highly transferable AEs is significantly higher than that in low-transferability AEs. Notably, highly transferable AEs not only exhibit higher top-1 confidence but also have lower confidence scores for subsequent rankings (rank-2 and beyond) compared to low-transferability AEs, with each subsequent rank showing an even larger gap." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.188, + 0.245, + 0.203 + ], + "angle": 0, + "content": "3.2.2. Method of SDS" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.208, + 0.484, + 0.314 + ], + "angle": 0, + "content": "Based on this observation, we naturally propose a hypothesis by inverse inference: increasing the gap between the target class and other classes under a top-\\(k\\) setting can enhance adversarial transferability. Therefore, the key question becomes how to widen this gap. To address this, we devised two strategies from different perspectives, as shown in the following formulas:" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.321, + 0.483, + 0.36 + ], + "angle": 0, + "content": "\\[\nf _ {u n t g} (x) = \\frac {1}{k - 1} \\sum_ {q \\in Q} f _ {q} (x) \\quad \\text {s . t .} \\quad y \\notin Q, \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.37, + 0.484, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {S 1} (x _ {t} ^ {a d v}, y) = f _ {y} (x _ {t} ^ {a d v}) - f _ {u n t g} (x _ {t} ^ {a d v}), \\qquad (9)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.414, + 0.484, + 0.47 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} _ {S 2} \\left(x _ {t} ^ {a d v}, y\\right) = \\frac {f _ {y} \\left(x _ {t} ^ {a d v}\\right) - f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right)}{\\left\\| \\nabla_ {x _ {t} ^ {a d v}} f _ {y} \\left(x _ {t} ^ {a d v}\\right) - \\nabla_ {x _ {t} ^ {a d v}} f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right) \\right\\| _ {2}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.698 + ], + "angle": 0, + "content": "The first strategy (S1) maximizes the target class logit value relative to the aggregated logits of other classes within the top-\\(k\\) setting, enhancing target class features in the latent feature space and diminishing those of similar other classes, thus widening the latent feature gap, as shown in Equation (9). The second strategy (S2), inspired by DeepFool's approach to calculating point-to-hyperplane distances [19], guides AEs to deeply infiltrate the target class region in multi-dimensional geometric space while maintaining a global distance from the aggregated boundaries of the closest non-target regions, as detailed in Equation (10). In these equations, \\(f(\\cdot)\\) represents the classifier, \\(Q\\) denotes the set of top-\\(k\\) classes (excluding \\(y\\)), and \\(f_{q}(x)\\) and \\(f_{y}(x)\\) represent the output values of \\(f(x)\\) corresponding to \\(q\\) and \\(y\\), respectively. The algorithm of ITDS is shown in Algorithm 1." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.224, + 0.727 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.81 + ], + "angle": 0, + "content": "We first outline the experimental setup, then evaluate the targeted-attack performance of both competition-based and transformation-based methods separately. Afterward, we evaluate the ability of our ITDS to break through defenses and conduct ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.282, + 0.835 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dataset. To ensure rigorous evaluation, we randomly selected 2,000 high-quality images from 1,000 classes in the ILSVRC2012 validation set [21]. These images were correctly identified by all models tested, including CNNs, ViTs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.152 + ], + "angle": 0, + "content": "and adversarial training models, using intersection extraction, with more than \\(90\\%\\) top-1 average confidence. The target class images of our method were randomly drawn from the corresponding category in the ILSVRC2012 dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.906, + 0.304 + ], + "angle": 0, + "content": "Models. We choose four normally pre-trained CNNs, i.e., ResNet50 (RN50) [9], VGG16 [22], MobileNet-v3-large (MN-v3) [11], RegNet-y-32gf (RegN) [20] as surrogate (white-box) models. Concurrently, we choose four white-box independent normally pre-trained CNNs, i.e., Inception-v3 (Inc-v3) [24], ResNet101 (RN101), DenseNet161 (DN161) [12], EfficientNet-b7 (ENet) [25], and four ViTs, i.e., VisionTransformer-b-16 (ViT) [5], DeiT-base (DeiT) [26], ConViT-base (ConViT) [6] and PiT-base (PiT) [10] to serve as target black-box models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.305, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Baselines. We adopt six challenging adversarial attacks as our baselines, which have released the code and provide the necessary parameters for reproduction, i.e., Competition-based: Admix [30], CFM [2], Transformation-Based: DIM [34], ODIM [1], SIA [31], and BSR [29]. All the baselines are combined with MI-FGSM." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.397, + 0.906, + 0.502 + ], + "angle": 0, + "content": "Defenses. To further demonstrate the effectiveness of ITDS, we consider four additional advanced defense methods, i.e., JPEG [8], FD [17], GDMP [28], Score-Opt (SO) [36], and three adversarial training defense models [27], i.e., \\(\\mathrm{Def}_{\\text{lin}f4}\\) (\\(\\mathrm{Def}_4\\)), \\(\\mathrm{Def}_{\\text{lin}f8}\\) (\\(\\mathrm{Def}_8\\)) and \\(\\mathrm{Def}_{l2-3}\\) (\\(\\mathrm{Def}_{l2}\\)) from ResNet50, which are proven to be robust to adversarial attacks on ImageNet datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Attack setting. For all methods, we set the maximum perturbation of \\(\\epsilon = 16\\), the number of epochs \\(T = 100\\), the step size \\(\\alpha = 1.6\\). We adopt the decay factor \\(\\mu = 1.0\\) for all the methods that employ the MI-FGSM, and a transformation probability of \\(p = 0.7\\) for DIM and ODIM. Furthermore, all parameters for Admix, SIA, CFM, and BSR are set as described in their respective published papers. For our ITDS, we set the number of epochs \\(_{sub}\\) to \\(T / 10\\), scaling level \\(m = 5\\) with \\(\\delta_{i} = 1 / 2^{i}\\), the number of random target class samples \\(n = 3\\), the competition ratio \\(\\beta = 1.5\\), and \\(k = 10\\) of the top-\\(k\\). For each AE of the trial, the target label is randomly chosen from the other 999 classes excluding its true label, and the final results are averaged over three trials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.871, + 0.729 + ], + "angle": 0, + "content": "4.2. Evaluation on Competition-based Attacks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We first evaluate the attack performance of various competition-based attacks. Since the FIA and RPA introduced in related work are only applicable to non-targeted attacks, and FFT must be combined with existing methods to be usable, we ultimately chose Admix and CFM as competitors for ITDS. To demonstrate the versatility of our approach, the surrogate models we use to craft AEs are structurally independent of each other. The targeted attack success rates (TASRs) are shown in Table 2, where the upper row indicates the targeted model under attack, and the left column lists the tested surrogate models. To test the feasi-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3720" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.895, + 0.311 + ], + "angle": 0, + "content": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50Admix100*7.92.14.91.638.619.80.60.30.10.20.16.9
CFM100*45.017.735.59.587.760.75.62.01.51.53.924.6
ITDSfs100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDSgs100*46.235.049.328.487.680.417.910.04.14.46.633.6
VGG16Admix4.388.4*1.47.70.70.94.10.50.10.10.10.51.8
CFM3.986.1*1.28.40.81.02.80.60.10.20.10.31.7
ITDSfs18.8100*12.020.68.07.618.45.61.00.60.51.58.6
ITDSgs18.7100*12.019.99.08.218.55.71.40.50.81.98.7
MN-v3Admix0.60.499.3*0.30.60.30.40.40.20.10.20.10.3
CFM10.48.9100*10.47.49.67.714.37.22.74.03.17.8
ITDSfs16.49.6100*15.211.816.015.219.814.53.84.24.411.9
ITDSgs16.79.9100*14.911.715.915.616.814.43.24.83.811.6
RegNAdmix1.41.11.695.8*0.40.51.60.50.30.20.20.40.7
CFM21.423.314.192.3*4.412.119.69.01.41.21.15.910.3
ITDSfs38.216.025.0100*17.824.840.623.911.17.67.214.020.5
ITDSgs41.217.726.899.9*19.727.743.125.013.39.68.417.422.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.323, + 0.905, + 0.365 + ], + "angle": 0, + "content": "Table 2. TASRs (%) on twelve pre-trained models with various competition-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks, and BAvg means average black-box TASR. ITDS\\(_{fs}\\) and ITDS\\(_{gs}\\) represent the integration of S1 and S2 in SDS, respectively, and the boldface represents their results." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.379, + 0.907, + 0.751 + ], + "angle": 0, + "content": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50DIM100*33.614.131.915.365.059.113.53.11.92.63.822.2
ODIM99.9*52.833.249.837.570.265.830.114.59.210.415.535.3
SIA100*86.856.386.540.499.093.534.330.621.724.040.355.7
BSR100*87.648.787.231.097.692.827.818.116.916.241.551.4
Admix-DIM100*51.925.553.629.183.877.228.88.04.35.110.134.3
CFM-DIM99.7*68.447.966.049.288.479.541.918.012.514.920.046.0
ITDS-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
VGG16DIM12.985.8*3.418.22.93.510.63.50.20.30.21.45.2
ODIM33.578.8*10.028.211.213.125.113.62.61.41.34.113.1
SIA50.481.7*13.555.67.320.543.77.72.11.21.69.219.3
BSR24.384.7*3.834.61.77.525.11.40.40.30.32.19.2
Admix-DIM20.288.0*6.823.95.75.416.08.00.80.50.52.18.2
CFM-DIM19.286.3*4.123.73.95.914.06.10.40.60.41.67.2
ITDS-DIM42.0100*24.834.825.719.835.324.98.52.83.27.820.9
MN-v3DIM6.32.599.2*4.96.94.55.516.26.01.82.52.35.4
ODIM24.917.399.4*21.222.122.823.236.120.09.611.611.220.0
SIA32.721.498.7*29.214.126.125.424.219.17.910.610.320.1
BSR22.012.598.9*18.38.714.413.015.911.04.46.57.812.2
Admix-DIM18.06.399.4*11.914.812.113.032.012.53.75.54.47.5
CFM-DIM34.724.999.8*31.629.433.427.844.228.111.814.211.626.5
ITDS-DIM49.527.3100*39.939.244.941.658.046.718.117.013.736.0
RegNDIM24.815.212.492.9*9.312.331.421.43.63.63.013.013.6
ODIM51.343.331.290.3*29.235.656.949.315.712.911.626.133.0
SIA71.272.138.695.7*15.648.072.033.119.717.918.647.741.3
BSR49.063.022.194.1*6.124.947.715.67.08.88.938.026.4
Admix-DIM39.425.220.692.6*14.823.147.733.17.67.17.322.922.6
CFM-DIM71.561.950.893.0*38.361.072.361.622.521.722.143.347.9
ITDS-DIM82.250.568.999.9*64.673.184.975.656.543.040.854.763.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.762, + 0.908, + 0.806 + ], + "angle": 0, + "content": "Table 3. TASRs (%) on twelve pre-trained models with various transformation-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks. ITDS-DIM represents the combination of ITDS, which integrates S1, with DIM, and the boldface represents the results of this combination." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.87 + ], + "angle": 0, + "content": "bility of the two strategies in SDS, ITDS is implemented in two forms: \\(\\mathrm{ITDS}_{fs}\\) for S1 (feature space) and \\(\\mathrm{ITDS}_{gs}\\) for S2 (geometric space)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "It can be seen that, compared to other competition-based attacks, ITDS demonstrates an overwhelming advantage." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Taking \\(\\mathrm{ITDS}_{gs}\\) as an example, its average black-box TASR (BAvg.) is \\(9.0\\%\\), \\(7.0\\%\\), \\(3.8\\%\\), and \\(12.4\\%\\) higher than the existing SOTA method CFM on RN50, VGG16, MN-v3, and RegN, respectively. From the perspective of cross-architecture models, the largest gaps in average black-box" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "3721" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.117, + 0.088, + 0.887, + 0.155 + ], + "angle": 0, + "content": "
AttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
ITDS1100*41.524.340.418.984.670.211.55.52.92.74.227.9
ITDS1+S100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDS1-DIM99.9*71.158.673.356.892.587.148.435.118.819.123.653.1
ITDS1+S-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.163, + 0.908, + 0.207 + ], + "angle": 0, + "content": "Table 4. Evaluation of the effectiveness of two modules in ITDS, where 'I' represents ITC and 'S' represents SDS. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. Boldface represents the best results. Note that single SDS requires combination with existing attack methods, which we discuss in the appendix." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.225, + 0.486, + 0.392 + ], + "angle": 0, + "content": "TASR between it and CFM on CNNs and ViTs are \\(13.9\\%\\) and \\(9.7\\%\\) on RegN, respectively. It should be noted that the white-box attack performance of our ITDS is also significantly superior, being the best in terms of white-box TASR in any surrogate model, especially on VGG16 and RegN, where it is \\(13.9\\%\\) and \\(7.6\\%\\) higher than CFM, respectively. In the comparison between \\(\\mathrm{ITDS}_{fs}\\) and \\(\\mathrm{ITDS}_{gs}\\), the results seem to be evenly matched, with each showing their own merits in the three trials. Therefore, we believe that both strategies are viable, which precisely proves the effectiveness and diversity of the proposed SDS principle." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.407, + 0.475, + 0.423 + ], + "angle": 0, + "content": "4.3. Evaluation on Transformation-based Attacks" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.431, + 0.485, + 0.672 + ], + "angle": 0, + "content": "Currently, transformation-based methods are the main approach to adversarial attacks and are the most effective way to improve the transferability of attacks. Numerous studies [2, 4, 16] have shown that non-transformation-based methods combined with DIM can further enhance the transferability of crafted AEs. This combination approach has become a common standard in industry, frequently compared to transformation-based methods for fairness. In this section, we evaluate the TASR of transformation-based and DIM-combined competition-based adversarial attacks. The former includes DIM, ODIM, SIA, and BSR, while the latter includes Admix-DIM, CFM-DIM, and ITDS-DIM. It should be noted that, due to the similar performance of S1 and S2 in SDS, in all subsequent experiments involving ITDS-DIM, we selected \\(\\mathrm{ITDS}_{fs}\\) as the representative to combine with DIM. The results are shown in Table 3." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.486, + 0.902 + ], + "angle": 0, + "content": "It can be observed that all competition-based methods show a significant improvement in attack performance after being combined with DIM. Compared to the strongest competition-based attack CFM-DIM, ITDS-DIM achieved average black-box TASR improvements of \\(15.7\\%\\), \\(13.7\\%\\), \\(9.5\\%\\), and \\(15.3\\%\\) on RN50, VGG16, MN-v3, and RegN, respectively. The maximum average black-box TASR gap between the two on CNNs and ViTs is \\(16.1\\%\\) on RN50 and \\(21.4\\%\\) on RegN. Even when compared with the SOTA transformation-based attacks, our method is not inferior, and even surpasses them in all aspects. Compared to the strongest transformation-based attack SIA, ITDS-DIM achieved average black-box TASR improvements of \\(15.9\\%\\) and \\(21.9\\%\\) on MN-v3 and RegN, respectively. The maximum average black-box TASR gap between the two on" + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.225, + 0.907, + 0.241 + ], + "angle": 0, + "content": "CNNs and ViTs is \\(21.3\\%\\) and \\(22.8\\%\\) on RegN, respectively." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.255, + 0.703, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.599, + 0.454, + 0.617, + 0.466 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.255, + 0.907, + 0.452 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.805, + 0.454, + 0.822, + 0.466 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.468, + 0.703, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.6, + 0.668, + 0.617, + 0.679 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.468, + 0.902, + 0.664 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.668, + 0.818, + 0.679 + ], + "angle": 0, + "content": "(d)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.691, + 0.908, + 0.749 + ], + "angle": 0, + "content": "Figure 4. Ablation studies on the RN50 model. (a) - (d): TASRs \\((\\%)\\) on the other eleven models with the adversaries crafted by ITDS-DIM, where the default value for \\(n\\), \\(\\beta\\), \\(k\\) and \\(T_{sub}\\) are set to 3, 1.5, 10 and 10 respectively, when test parameters for each other." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.681, + 0.789 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "We first present a series of ablation studies conducted on RN50 to assess the impact of each parameter in ITDS on attack performance, including the number of random target class samples \\( n \\), the competitive ratio of the inverse target gradient \\( \\beta \\), the value of \\( k \\) in top- \\( k \\), and the number of epochs \\( t_{sub} \\) for non-targeted attack \\( T_{sub} \\). In all ablation experiments, unless otherwise specified, all parameters other" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3722" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.097, + 0.089, + 0.905, + 0.171 + ], + "angle": 0, + "content": "
AblationTarget Model
Norm (P1)MI (P2)RN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
✓7.1*0.51.00.40.40.80.30.20.20.20.10.10.4
✓100*72.866.175.865.190.188.254.737.216.917.423.755.3
✓✓100*59.741.366.152.890.991.345.127.814.516.221.147.9
100*79.370.382.869.695.694.662.044.423.624.731.661.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.907, + 0.21 + ], + "angle": 0, + "content": "Table 5. TASRs (%) of ITDS-DIM by ablating inner modules of the ITC. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. In which Norm and MI correspond to P1 and P2 in Section 3.1.2, respectively. Boldface represents the best results." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.228, + 0.48, + 0.293 + ], + "angle": 0, + "content": "
AttackJPEGFDSOGDMPDef4Def8Defl2
SIA83.289.844.31.539.014.054.7
Admix-DIM80.092.947.71.934.814.553.1
CFM-DIM82.592.651.54.125.28.745.5
ITDS-DIM99.499.991.331.250.225.666.8
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.302, + 0.483, + 0.386 + ], + "angle": 0, + "content": "Table 6. TASRs (\\%) of breaking-through four defense methods and three adversarial training defense models. The results for defense methods are based on the average performance of AEs crafted on RN50, VGG16, MN-v3, and RegN respectively, while the results for defense models are based on the performance of AEs crafted by the corresponding models. Bolder for the best." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.411, + 0.483, + 0.441 + ], + "angle": 0, + "content": "than the test parameters are fixed according to the experimental setup." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.445, + 0.483, + 0.791 + ], + "angle": 0, + "content": "As shown in Figure 4a, TASR increases with \\( n \\), as the diversity of target class images directly affects attack performance. However, the curve levels off after \\( n = 3 \\), and considering computational costs, we selected \\( n = 3 \\). In Figure 4b, most models achieve optimal performance at \\( \\beta = 1.5 \\) with a few exceptions (e.g., MN-v3). We believe a moderate competitive gradient provides an \"exercise\" effect, while an overly strong gradient disrupts input features, reducing attack performance. In Figure 4c, optimal performance generally occurs at \\( k = 5 \\) or \\( k = 10 \\). We posit that an overly small \\( k \\) under-covers classes and fails to enlarge global margins, whereas an overly large one blurs the optimization. At \\( k = 10 \\), cross-architecture transferability is even better. In Figure 4d, evaluating epochsub \\( (T_{sub}) \\) for non-targeted attacks, we found that larger values do not always lead to better performance. As discussed in the ITC section, inverse target gradient perturbations generated at each \\( T_{sub} \\) have diverse potential competitive features, boosting attack performance. At \\( T_{sub} = 1 \\), though it involves more target class samples, it lacks deeper latent features of the target class, limiting attack performance. Beyond a certain \\( T_{sub} \\) value, missing target class features reduce the diversity of inverse gradients, causing performance to regress." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Furthermore, to verify the effectiveness of the ITC and SDS components in ITDS, we conducted experimental evaluations as shown in Table 4. The results indicate that ITDS using only ITC has already surpassed the attack performance of CFM in Table 2. Regardless of whether DIM is combined or not, SDS can bring additional performance improvements to ITDS. These findings collectively support the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.747, + 0.245 + ], + "angle": 0, + "content": "key roles of ITC and SDS in ITDS." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.907, + 0.504 + ], + "angle": 0, + "content": "To verify the impact of internal modules within the ITC on attack performance, we conducted ablation experiments as shown in Table 5. It can be observed that the use of N-norm may limit the scale of gradient updates, which could result in generated AEs lacking sufficient aggressiveness. On the contrary, if we do not use N-norm (P1), gradient updates will gain greater freedom, which helps to explore the model's decision boundary more deeply, and may thus produce more aggressive AEs. Simultaneously, momentum gradients, as a technique for accumulating previous gradients, will weaken the intensity of inverse target features in the current iteration. Since we wish to maintain the integrity of inverse target features and avoid any normalization, the inverse target momentum gradients in each round will compete with the AE when combined with P1, leading to excessive deformation of AE features, which is disastrous for implementing precisely targeted attacks." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.81, + 0.525 + ], + "angle": 0, + "content": "4.5. Evaluation on Advanced Defenses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.653 + ], + "angle": 0, + "content": "In this section, we evaluate the capabilities of various attacks to break through different defense mechanisms. As shown in Table 6, our method exhibits exceptional robustness against a wide range of defense strategies and models. For example, when the Score-OPT method is employed, the average TASR of our approach is nearly \\(40\\%\\) higher than that of CFM-DIM. Furthermore, our method achieves the best performance in all three defense models evaluated." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper introduces a novel method named ITDS, which significantly improves the transferability of targeted AEs by incorporating the ITC mechanism and the SDS strategy. Experimental results demonstrate that ITDS achieves a substantial increase in average transferable TASRs across various mainstream CNN and ViT models, outperforming SOTA competition-based and transformation-based methods. Our ablation studies further confirmed the key roles of the ITC and SDS components in enhancing attack performance. Moreover, ITDS exhibits superior effectiveness in countering multiple defense methods and models. These findings confirm the effectiveness and practicality of ITDS in enhancing the transferability of targeted AEs and provide a new perspective for the field of adversarial research." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3723" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.486, + 0.314 + ], + "angle": 0, + "content": "This work was supported by funding from the National Natural Science Foundation of China (NSFC) under Grants 62272486, 62172154, and 62372473, the Natural Science Foundation of Hunan Province under Grant 2023JJ70016, the Hong Kong Research Grants Council (RGC) General Research Fund under Grants 152244/21E, 152169/22E, 152228/23E, and 162161/24E, the Research Impact Fund under Grant R5011-23F, the Collaborative Research Fund under Grant C1042-23GF, the NSFC/RGC Collaborative Research Scheme under Grant CRS_HKUST602/24, the Areas of Excellence Scheme under Grant AoE/E-601/22-R, and the InnoHK initiative (HKGAI). Professor Shigeng Zhang is the corresponding author of this paper." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.331, + 0.188, + 0.347 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.356, + 0.483, + 0.439 + ], + "angle": 0, + "content": "[1] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2022. 1, 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.442, + 0.483, + 0.524 + ], + "angle": 0, + "content": "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 1, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.528, + 0.483, + 0.596 + ], + "angle": 0, + "content": "[3] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.6, + 0.483, + 0.669 + ], + "angle": 0, + "content": "[4] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.672, + 0.483, + 0.754 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.757, + 0.483, + 0.826 + ], + "angle": 0, + "content": "[6] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference on Machine Learning, pages 2286-2296. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.829, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[7] Goodfellow et al. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[8] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2018. 5" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.356, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.907, + 0.148 + ], + "angle": 0, + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.908, + 0.219 + ], + "angle": 0, + "content": "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.907, + 0.29 + ], + "angle": 0, + "content": "[11] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.292, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[12] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[13] Kurakin et al. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[14] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Foolmix: Strengthen the transferability of adversarial examples by dual-blending and direction update strategy. IEEE Transactions on Information Forensics and Security, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[15] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Ucg: A universal cross-domain generator for transferable adversarial examples. IEEE Transactions on Information Forensics and Security, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[16] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[17] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature distillation: Dnn-oriented jpeg compression against adversarial examples. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 860-868. IEEE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[18] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549-566. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[20] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3724" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.484, + 0.164 + ], + "angle": 0, + "content": "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.167, + 0.483, + 0.195 + ], + "angle": 0, + "content": "[23] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.198, + 0.483, + 0.265 + ], + "angle": 0, + "content": "[24] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.269, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.327, + 0.483, + 0.396 + ], + "angle": 0, + "content": "[26] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.399, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[27] Francisco Utrera, Evan Kravitz, N Benjamin Erickson, Rajiv Khanna, and Michael W Mahoney. Adversarily-trained deep nets transfer better: Illustration on image classification. arXiv preprint arXiv:2007.05869, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.483, + 0.497 + ], + "angle": 0, + "content": "[28] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.501, + 0.483, + 0.569 + ], + "angle": 0, + "content": "[29] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.573, + 0.483, + 0.627 + ], + "angle": 0, + "content": "[30] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.631, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[31] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[32] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[33] Juanjuan Weng, Zhiming Luo, Shaozi Li, Nicu Sebe, and Zhun Zhong. Logit margin matters: Improving transferable targeted adversarial attack by logit calibration. IEEE Transactions on Information Forensics and Security, 18:3561-3574, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 1, 2, 5" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[35] Zeng et al. Enhancing targeted transferability via feature space fine-tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4475-4479. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[36] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 36:51810-51829, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[37] Qilong Zhang, Xiaodan Li, YueFeng Chen, Jingkuan Song, Lianli Gao, Yuan He, et al. Beyond imagenet attack: Towards crafting adversarial examples for black-box domains. In International Conference on Learning Representations, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.277, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[38] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quaxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, pages 1672-1678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.374 + ], + "angle": 0, + "content": "[39] Zhao et al. On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems, 34:6115-6128, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[40] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3725" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_origin.pdf b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8576e0113cf6d7c22c1aea47575b828e77ea1196 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/b533dcf6-58e9-4bab-bdf5-cd55686734f9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f79cb559f1c39fbddd11040c12c15d94c90273bc8ea85b8ea78b0f06c6e60728 +size 863382 diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/full.md b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/full.md new file mode 100644 index 0000000000000000000000000000000000000000..159a1e6d83584616610f0fa8fd3e84824d1f7062 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/full.md @@ -0,0 +1,331 @@ +# Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching + +Zhankai Li $^{1}$ , Weiping Wang $^{1}$ , Jie Li $^{1}$ , Shigeng Zhang $^{1}$ , Yunan Hu $^{1}$ , Song Guo $^{2}$ + +1Central South University + +2The Hong Kong University of Science and Technology + +janlzk7@gmail.com, {wpwang, lijie55, sgzhang, hyn318}@csu.edu.cn, songguo@ust.hk + +# Abstract + +In the field of AI security, deep neural networks (DNNs) are highly sensitive to adversarial examples (AEs), which can cause incorrect predictions with minimal input perturbations. Although AEs exhibit transferability across models, targeted attack success rates (TASRs) are low due to differences in feature dimensions and decision boundaries. To enhance targeted AE transferability, we propose a novel approach using Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS) in the optimization process. Specifically, we employ a siamese-network-like framework to generate both non-targeted and targeted AEs. The ITC mechanism applies non-targeted adversarial gradients each epoch to impede the optimization of targeted perturbations, thereby improving robustness. Additionally, a top-k SDS strategy guides AEs to penetrate target class regions in the latent space while distancing from nontargeted regions, achieving optimal transferability. Compared to state-of-the-art competition-based attacks, our method significantly improves transferable TASRs by $16.1\%$ and $21.4\%$ on mainstream CNNs and ViTs, respectively, and demonstrates superior defense-breaking capabilities. Our code is available here. + +# 1. Introduction + +Adversarial attacks involve adding small, often imperceptible perturbations to input data, which can lead otherwise high-performing models to produce incorrect outputs. The concept of adversarial attacks was first introduced by [23] and further developed by [7], who demonstrated the vulnerability of deep neural networks (DNNs) to these perturbations. With the rapid advancement of DNNs in recent years, the security risks revealed by adversarial attacks have become a critical issue. Adversarial attacks are now widely used to assess model security, improve model robustness, and drive research in defense technologies. + +However, the significance of adversarial attacks extends beyond these applications, as they have inadvertently become a widely used defense mechanism in real-world scenarios. A prime example is Google's login verification system, reCAPTCHA, which aims to distinguish real users from bots to prevent abuse. Adversarial examples (AEs) are one of the primary image verification methods employed, as illustrated in Figure 1. This raises an important question: is it secure enough? Since we do not know the bot's model architecture or training domain, the transferability of AEs becomes especially crucial. Most current research on adversarial transferability focuses on non-targeted attacks [14, 40], even extending to cross-architecture and cross-domain attacks [15, 37], achieving notable performance. In contrast, targeted attacks face a significant increase in difficulty due to the substantial changes in decision boundaries, and as a result, they still maintain a relatively low success rate. However, for login verification, given that the output category of non-targeted attacks is uncontrollable, and even the original verification images without perturbation may have "low-quality" issues (leading to misidentification by bots), in such cases, targeted AEs are often safer and more reliable than non-targeted AEs that may cause the deployed bots to misjudge into the ground truth category. + +Currently, effective targeted attack methods are primarily based on transformation [1, 18, 29, 34], achieving transferability through input augmentation. However, the relentless pursuit and stacking of a greater variety of transformations, although enhancing transferability to some extent, does not guarantee robust performance in white-box attacks. Moreover, this stacking approach based on affine and perspective transformations offers limited insight into the complexity and vulnerability of DNNs. Recently, competition-based attacks have been proposed [2, 30], offering a unique perspective. Their advantage lies in compelling AEs to consider the more complex and diverse feature representations within DNNs during their generation, simulating various noise interferences and attack strategies, + +![](images/c4522ccb122a1c16a16b7d176ecd530a786125a1646d71d9284388dcab84ce61.jpg) +Figure 1. Example of Google's web login verification system, reCAPTCHA. It determines whether a user is human or a bot based on whether they can correctly match text to corresponding images. + +and promoting the learning of more robust feature combinations. The competitive mechanism, by delving into the model's decision boundaries, helps to uncover the latent vulnerabilities of DNNs and provides a new perspective for understanding and improving the model's generalization capabilities. However, no matter what the attack method is, it cannot conceal the widespread insufficiency in the transferability performance of current targeted adversarial attacks. + +To address this challenge, we propose a new targeted adversarial attack method called ITDS, which consists of two components: Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS), aimed at enhancing the transferability of targeted AEs. In the first part, we adopt a siamese-network-like framework, innovatively leveraging the ground truth gradients of multiple target class non-targeted AEs at intermediate stages to exert inverse "competitiveness" on non-target class targeted AEs during the optimization process. In the second part, to tackle the multi-dimensional complexity of decision boundaries in targeted attacks, we introduce a top- $k$ SDS strategy to seek stable perturbation update directions. This guides AEs to penetrate target class regions within the latent multidimensional space while globally distancing from several closest non-target region boundaries. The experimental results demonstrate that the proposed ITDS method has achieved significant advantages in enhancing the transferability and defense-breaking capabilities of targeted AEs, surpassing the existing state-of-the-art (SOTA) methods. + +# 2. Related Work + +In this section, we will briefly overview the background of adversarial attacks. Then focus on various targeted attack techniques that demonstrate effective transferability. + +# 2.1. Adversarial Attacks + +Goodfellow et.al first designed a simple yet effective method called Fast Gradient Sign Method (FGSM) [7]. This + +method generates AEs by making small adjustments to the input $x$ (original sample) along the direction of the gradient, i.e., $x^{adv} = x + \delta$ . In the case of targeted attacks, FGSM updates the input $x$ in the direction that minimizes the classification loss for a given target class $y$ , with the objective function as follows: + +$$ +\arg \min _ {x ^ {a d v}} \mathcal {J} (x ^ {a d v}, y; \theta) \text {s . t .} \| x - x ^ {a d v} \| _ {\infty} \leq \epsilon , \tag {1} +$$ + +where $\theta$ is the model parameters, $\mathcal{J}$ denotes the Cross-Entropy (CE) loss function, and the $\ell_{\infty}$ norm is used to constrain the perturbation within the range $\epsilon$ . + +I-FGSM [13] is an iterative extension of FGSM, which gradually refines adversarial perturbation with small step sizes. To address gradient vanishing issues in targeted attacks, [39] proposes using Logits loss instead of CE loss, significantly enhancing targeted attack effectiveness. Based on this, [33] further proposes downscaling logit calibration with a temperature factor and an adaptive margin. While these methods provide a foundation for enhancing targeted attack effectiveness, their transferability remains limited. + +# 2.2. Transferable Adversarial Attacks + +In this section, we introduce advanced transferable attacks that can successfully deceive other models using AEs generated on a single surrogate model without additional training. We categorize these methods into transformation-based and competitive-based attacks, focusing primarily on transferable targeted attacks, along with certain non-targeted attacks that can also be adapted for targeted attacks. + +Transformation-Based Attacks. DIM [34] is a representative transformation-based attack, which resizes and pads images with a certain probability at each step of the multi-step attack process. SSM [18] applies frequency domain augmentations through spectrum transformations on the input. ODIM [1] method significantly improves the targeted attack transferability by mapping input images to random 3D objects and applying varied rendering techniques. SIA [31] generates structurally diverse AEs by applying various random transformations to the images. Both ODIM and SIA integrate DIM into their implementation principles and share a similar lineage in transformation strategies. BSR [29] randomly shuffles and rotates the blocks within images, calculates the ensemble gradients of this new set of images, reduces attention heatmap differences across various models, and significantly enhances transferability. + +Competition-based Attacks. Admix [30] combines features from different categories in the input domain, creating multiple images for gradient computation, and enhancing transferability without degrading white-box performance. FIA [32] employs aggregated gradients to disrupt essential object-aware features while suppressing model-specific features, boosting AE transferability. RPA [38] adopts a similar approach to FIA, introducing random patch transforma + +tions to benign images. To alleviate the overfitting dilemma common in an AE crafted by simple iterative attacks, FFT [35] encourages features conducive to the target class while discouraging those associated with the original class in an intermediate layer of the source model. CFM [2] introduces a competitive strategy, crafting targeted adversarial perturbations with diverse features by introducing two types of competitor noises: adverse perturbations toward different target classes and friendly perturbations towards the correct class, achieving SOTA results. + +# 3. The Design of ITDS + +This section consists of two parts: the Inverse Target Gradient Competition mechanism and the Spatial Distance Stretching strategy. In each part, we first explain the motivation and then propose the corresponding method. + +# 3.1. Inverse Target Gradient Competition (ITC) + +# 3.1.1. Motivation of ITC + +The motivation for the first part of our research arises from a question: What is a competition mechanism? From existing work related to competition mechanisms, we summarize it as a way to enhance the transferability of AEs by promoting diversified attack strategies through simulating various noise interferences. Since noise interference is involved, the essence of competitiveness can be interpreted as an inverse force. For targeted attacks, could we identify the most competitive noise? We believe this is the key issue. To address this critical question, we turn to target class samples and consider the gradients of the target label for these samples. These gradients essentially represent the optimal force linking target class samples with the target label, which can be used in both forward and inverse directions. Therefore, we have a bold hypothesis: when using this gradient in inverse, it not only effectively generates non-targeted AEs but also serves as the optimal competitiveness for targeted attacks. + +# 3.1.2. Method of ITC + +Figure 2 shows the overall framework of our Inverse Target Gradient Competition mechanism, which is jointly driven by both non-targeted and targeted attacks within a siamese-network-like architecture. Our strategy is to utilize the inverse target gradient to introduce resistance during the optimization process of targeted AEs. This inverse target gradient can be generated for any samples; but which inverse target gradient should we choose? As mentioned earlier, we seek the strongest resistance, which is the inverse gradient of the target label with respect to the target class sample. The approach of generating the inverse target gradient is illustrated in the lower part of Figure 2, where we use 'cock' as the target label as an example. In each sub-epochs (epochssub), multiple target class images that are correctly classified by the siamese classifier are selected to generate + +non-targeted AEs. The optimization goal is to minimize the logit value of the target label. We set $\mathrm{epoch}_{sub}$ to be one-tenth of the total epochs, as we believe that target class samples need a certain number of iterations to transform into robust non-targeted AEs. Accordingly, the non-targeted AEs generated in each $\mathrm{epoch}_{sub}$ contain different levels and increasingly deeper latent features of target class. Therefore, we consider the inverse target gradient perturbations generated in each $\mathrm{epoch}_{sub}$ to also have diverse potential competitive characteristics, and during the generation of nontargeted AEs in each epoch, multiple inverse target perturbations are generated to enrich competitive characteristics. The formulas are as follows: + +$$ +\tilde {g} _ {t} ^ {z} = \nabla_ {z _ {t} ^ {a d v}} \mathcal {L} (z _ {t} ^ {a d v}, y; \theta), \tag {2} +$$ + +$$ +g _ {t + 1} ^ {z} = \mu \cdot g _ {t} ^ {z} + \tilde {g} _ {t} ^ {z}, \tag {3} +$$ + +$$ +z _ {t + 1} ^ {a d v} = \operatorname {C l i p} _ {z, \epsilon} \{z _ {t} ^ {a d v} - \alpha \cdot \operatorname {s i g n} (g _ {t + 1} ^ {z}) \}, \tag {4} +$$ + +where $z$ is the target class sample, $\tilde{g}_t^z$ represents the inverse target gradient, and we follow [1, 2, 39] to use the Logit loss $\mathcal{L}$ (opposite to CE loss optimization). $g_{t + 1}^{z}$ represents the momentum gradient without normalization for $z$ , and $z_{t + 1}^{adv}$ is the non-targeted AE generated in each round. + +![](images/a377ae4ff4cfa262b59dbfa1e310b1f38b511cb76d59b64e1f06b23f8c1d6607.jpg) +Figure 2. The overall framework of the ITC, e.g., combined with DIM, utilizes a siamese-network-like architecture to generate targeted and non-targeted AEs simultaneously. The upper part focuses on targeted attacks against non-target class samples, while the lower part addresses non-targeted attacks on target class samples. Note that for non-targeted attacks, the optimization condition utilizes the logits value of the target label, whereas the goal of the targeted attack is to stretch the spatial distance proposed in SDS. + +The process of generating targeted AEs is shown in the upper part of Figure 2, where the input is a non-target image, and the optimization goal is to maximize the spatial distance, which we introduce in the SDS of the next section. Each round of generated targeted AEs competes with + +a certain proportion of inverse target gradients from different stages. The formulas are as follows: + +$$ +\bar {g} _ {t} ^ {c p t} = \frac {1}{n \cdot m} \sum_ {z \in Z} \sum_ {i = 0} ^ {m - 1} \nabla_ {x _ {t} ^ {a d v}} \mathcal {L} \left(\delta_ {i} \cdot \left(x _ {t} ^ {a d v} - \beta \cdot \tilde {g} _ {t} ^ {z}\right), y; \theta\right), \tag {5} +$$ + +$$ +g _ {t + 1} = \mu \cdot g _ {t} + \bar {g} _ {t} ^ {\text {c p t}}, \tag {6} +$$ + +$$ +x _ {t + 1} ^ {a d v} = \operatorname {C l i p} _ {x, \epsilon} \left\{x _ {t} ^ {a d v} + \alpha \cdot \operatorname {s i g n} \left(g _ {t + 1}\right) \right\}, \tag {7} +$$ + +where $n$ represents the number of target class samples in per epochs $_{sub}$ , $\delta_{i}$ and $m$ are the scaling factor and level, $\beta$ is the competition ratio, and each round of targeted AEs competes with the corresponding inverse target gradient. + +There are two notable points. The first point (P1) is whether, for targeted or non-targeted attacks, the calculation of momentum gradients differs from the traditional MIFGSM [3], and we have not employed norm normalization (N-norm). One of the reasons is the absence of gradient explosion in our experimental evaluations, and more importantly, our experiments have shown that using N-norm would limit our attack performance. We believe that N-norm might restrict the magnitude of gradient updates, resulting in less aggressive generation of AEs. Without N-norm, gradient updates could become more flexible, allowing for more effective exploration of the model's decision boundary, thereby generating more aggressive AEs. + +The second point (P2) is that we did not use the momentum gradient as the competitive gradient. The reason is that the momentum gradient is essentially a cumulative gradient. It weakens the strength of the inverse target features in the corresponding round. Since we want to maintain the integrity of the inverse target features, using it without any normalization operation in conjunction with P1 would cause excessive damage to the input features. This would be catastrophic for targeted attacks. These two points will also be discussed in subsequent ablation experiments. + +![](images/5883614a28bfcc96c9c66418acb147b6bac45409622873337a6338d6a8f4bb9a.jpg) +(a) + +![](images/0c50cd2ec667da5ca912afa440004cf14cdee048fe3fdc36ac2dc304bd1083f5.jpg) +(b) + +![](images/5928543575ec402b6c0440fb86cb17e6ed7c8e24720e1f3fa575d6875ab41d3a.jpg) +(c) +Figure 3. Comparison of targeted AEs with strong and weak transferability. Panels (a) and (c) show AEs with weak transferability, while (b) and (d) show AEs with strong transferability. + +![](images/6fbab54b3fc182edb5b5b4a1347815fbee0ab12b2732a16a8abaabee283e3fba.jpg) +(d) + +# 3.2. Spatial Distance Stretching (SDS) + +# 3.2.1. Motivation of SDS + +The motivation for the second part of our study arises from an interesting phenomenon commonly observed in highly + +
LabelFigure12345
motor_scooter(a)42.0%5.5%4.3%3.6%3.2%
(b)83.3%4.9%2.1%1.1%0.3%
siamese_cat(c)49.6%5.6%5.3%5.1%5.0%
(d)90.5%2.1%1.3%0.7%0.2%
+ +Table 1. The average top-5 confidence on black-box models successfully attacked by targeted AEs as shown in Figure 3. + +# Algorithm 1 The ITDS Attack Algorithm + +Input: Siamese classifier $f$ with parameters $\theta$ , Logit loss $\mathcal{L}$ , our Distance loss $\mathcal{D}$ , and an original sample $x$ with target label $y$ + +Input: The maximum perturbation $\epsilon$ , the step size $\alpha$ , the number of epochs $T$ and $\mathrm{epochs}_{sub} T_{sub}$ + +Input: The decay factor $\mu$ , scaling level $m$ , and $n$ random target class samples $Z$ ( $z \in Z$ ) + +Input: The competition ratio $\beta$ , and topk classifications $k$ + +Output: An adversarial example $x^{adv}$ . + +1: $g_0^z = 0$ ; $g_0 = 0$ ; $z_0^{adv} = z$ ; $x_0^{adv} = x$ +2: for $t$ in range $(T)$ do +3: if $t\% T_{sub} == 0$ then +4: Reselect $n$ random target class samples $Z$ +5: end if +6: Get the inverse target gradient for each $\tilde{g}_t^z$ of $n$ random target class samples by Eq.(2) +7: Get the momentum inverse target gradient for each $g_{t+1}^{z}$ by Eq.(3) +8: Update each $z_{t+1}^{adv}$ by Eq.(4) +9: Get the set $Q$ of top- $k$ classifications by $f(x_{t}^{adv})$ +10: Calculate the integrate gradient of competition examples by: + +$$ +\bar {g} _ {t} ^ {c p t} = \frac {1}{n \cdot m} \sum_ {z \in Z} \sum_ {i = 0} ^ {m - 1} \nabla_ {x _ {t} ^ {a d v}} \mathcal {D} \big (\delta_ {i} \cdot (x _ {t} ^ {a d v} - \beta \cdot \tilde {g} _ {t} ^ {z}), y \big) +$$ + +11: Get the momentum gradient by Eq.(6) +12: Update the adversarial example $x_{t + 1}^{adv}$ by Eq.(7) +13: end for +14: return $x^{adv} = x_T^{adv}$ . + +transferable targeted AEs. As shown in Figure 3, we present two 'cock'-targeted examples as representatives: targeted AEs generated using ResNet50 with strong and weak transferability, from left to right. In (a) and (c), we have AEs that can only fool fewer than three black-box models into predicting the target class, while (b) and (d) show AEs capable of misleading over twelve black-box models to the target class. Table 1 presents the top-5 average confidence scores for each AE on the successfully misled black-box models. It can be seen that the average confidence for the target la + +bel (top-1) in highly transferable AEs is significantly higher than that in low-transferability AEs. Notably, highly transferable AEs not only exhibit higher top-1 confidence but also have lower confidence scores for subsequent rankings (rank-2 and beyond) compared to low-transferability AEs, with each subsequent rank showing an even larger gap. + +# 3.2.2. Method of SDS + +Based on this observation, we naturally propose a hypothesis by inverse inference: increasing the gap between the target class and other classes under a top- $k$ setting can enhance adversarial transferability. Therefore, the key question becomes how to widen this gap. To address this, we devised two strategies from different perspectives, as shown in the following formulas: + +$$ +f _ {u n t g} (x) = \frac {1}{k - 1} \sum_ {q \in Q} f _ {q} (x) \quad \text {s . t .} \quad y \notin Q, \tag {8} +$$ + +$$ +\mathcal {D} _ {S 1} (x _ {t} ^ {a d v}, y) = f _ {y} (x _ {t} ^ {a d v}) - f _ {u n t g} (x _ {t} ^ {a d v}), \qquad (9) +$$ + +$$ +\mathcal {D} _ {S 2} \left(x _ {t} ^ {a d v}, y\right) = \frac {f _ {y} \left(x _ {t} ^ {a d v}\right) - f _ {u n t g} \left(x _ {t} ^ {a d v}\right)}{\left\| \nabla_ {x _ {t} ^ {a d v}} f _ {y} \left(x _ {t} ^ {a d v}\right) - \nabla_ {x _ {t} ^ {a d v}} f _ {u n t g} \left(x _ {t} ^ {a d v}\right) \right\| _ {2}}. \tag {10} +$$ + +The first strategy (S1) maximizes the target class logit value relative to the aggregated logits of other classes within the top- $k$ setting, enhancing target class features in the latent feature space and diminishing those of similar other classes, thus widening the latent feature gap, as shown in Equation (9). The second strategy (S2), inspired by DeepFool's approach to calculating point-to-hyperplane distances [19], guides AEs to deeply infiltrate the target class region in multi-dimensional geometric space while maintaining a global distance from the aggregated boundaries of the closest non-target regions, as detailed in Equation (10). In these equations, $f(\cdot)$ represents the classifier, $Q$ denotes the set of top- $k$ classes (excluding $y$ ), and $f_{q}(x)$ and $f_{y}(x)$ represent the output values of $f(x)$ corresponding to $q$ and $y$ , respectively. The algorithm of ITDS is shown in Algorithm 1. + +# 4. Experiments + +We first outline the experimental setup, then evaluate the targeted-attack performance of both competition-based and transformation-based methods separately. Afterward, we evaluate the ability of our ITDS to break through defenses and conduct ablation studies. + +# 4.1. Experimental Setup + +Dataset. To ensure rigorous evaluation, we randomly selected 2,000 high-quality images from 1,000 classes in the ILSVRC2012 validation set [21]. These images were correctly identified by all models tested, including CNNs, ViTs + +and adversarial training models, using intersection extraction, with more than $90\%$ top-1 average confidence. The target class images of our method were randomly drawn from the corresponding category in the ILSVRC2012 dataset. + +Models. We choose four normally pre-trained CNNs, i.e., ResNet50 (RN50) [9], VGG16 [22], MobileNet-v3-large (MN-v3) [11], RegNet-y-32gf (RegN) [20] as surrogate (white-box) models. Concurrently, we choose four white-box independent normally pre-trained CNNs, i.e., Inception-v3 (Inc-v3) [24], ResNet101 (RN101), DenseNet161 (DN161) [12], EfficientNet-b7 (ENet) [25], and four ViTs, i.e., VisionTransformer-b-16 (ViT) [5], DeiT-base (DeiT) [26], ConViT-base (ConViT) [6] and PiT-base (PiT) [10] to serve as target black-box models. + +Baselines. We adopt six challenging adversarial attacks as our baselines, which have released the code and provide the necessary parameters for reproduction, i.e., Competition-based: Admix [30], CFM [2], Transformation-Based: DIM [34], ODIM [1], SIA [31], and BSR [29]. All the baselines are combined with MI-FGSM. + +Defenses. To further demonstrate the effectiveness of ITDS, we consider four additional advanced defense methods, i.e., JPEG [8], FD [17], GDMP [28], Score-Opt (SO) [36], and three adversarial training defense models [27], i.e., $\mathrm{Def}_{\text{lin}f4}$ ( $\mathrm{Def}_4$ ), $\mathrm{Def}_{\text{lin}f8}$ ( $\mathrm{Def}_8$ ) and $\mathrm{Def}_{l2-3}$ ( $\mathrm{Def}_{l2}$ ) from ResNet50, which are proven to be robust to adversarial attacks on ImageNet datasets. + +Attack setting. For all methods, we set the maximum perturbation of $\epsilon = 16$ , the number of epochs $T = 100$ , the step size $\alpha = 1.6$ . We adopt the decay factor $\mu = 1.0$ for all the methods that employ the MI-FGSM, and a transformation probability of $p = 0.7$ for DIM and ODIM. Furthermore, all parameters for Admix, SIA, CFM, and BSR are set as described in their respective published papers. For our ITDS, we set the number of epochs $_{sub}$ to $T / 10$ , scaling level $m = 5$ with $\delta_{i} = 1 / 2^{i}$ , the number of random target class samples $n = 3$ , the competition ratio $\beta = 1.5$ , and $k = 10$ of the top- $k$ . For each AE of the trial, the target label is randomly chosen from the other 999 classes excluding its true label, and the final results are averaged over three trials. + +# 4.2. Evaluation on Competition-based Attacks + +We first evaluate the attack performance of various competition-based attacks. Since the FIA and RPA introduced in related work are only applicable to non-targeted attacks, and FFT must be combined with existing methods to be usable, we ultimately chose Admix and CFM as competitors for ITDS. To demonstrate the versatility of our approach, the surrogate models we use to craft AEs are structurally independent of each other. The targeted attack success rates (TASRs) are shown in Table 2, where the upper row indicates the targeted model under attack, and the left column lists the tested surrogate models. To test the feasi- + +
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50Admix100*7.92.14.91.638.619.80.60.30.10.20.16.9
CFM100*45.017.735.59.587.760.75.62.01.51.53.924.6
ITDSfs100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDSgs100*46.235.049.328.487.680.417.910.04.14.46.633.6
VGG16Admix4.388.4*1.47.70.70.94.10.50.10.10.10.51.8
CFM3.986.1*1.28.40.81.02.80.60.10.20.10.31.7
ITDSfs18.8100*12.020.68.07.618.45.61.00.60.51.58.6
ITDSgs18.7100*12.019.99.08.218.55.71.40.50.81.98.7
MN-v3Admix0.60.499.3*0.30.60.30.40.40.20.10.20.10.3
CFM10.48.9100*10.47.49.67.714.37.22.74.03.17.8
ITDSfs16.49.6100*15.211.816.015.219.814.53.84.24.411.9
ITDSgs16.79.9100*14.911.715.915.616.814.43.24.83.811.6
RegNAdmix1.41.11.695.8*0.40.51.60.50.30.20.20.40.7
CFM21.423.314.192.3*4.412.119.69.01.41.21.15.910.3
ITDSfs38.216.025.0100*17.824.840.623.911.17.67.214.020.5
ITDSgs41.217.726.899.9*19.727.743.125.013.39.68.417.422.7
+ +Table 2. TASRs (%) on twelve pre-trained models with various competition-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks, and BAvg means average black-box TASR. ITDS $_{fs}$ and ITDS $_{gs}$ represent the integration of S1 and S2 in SDS, respectively, and the boldface represents their results. + +
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50DIM100*33.614.131.915.365.059.113.53.11.92.63.822.2
ODIM99.9*52.833.249.837.570.265.830.114.59.210.415.535.3
SIA100*86.856.386.540.499.093.534.330.621.724.040.355.7
BSR100*87.648.787.231.097.692.827.818.116.916.241.551.4
Admix-DIM100*51.925.553.629.183.877.228.88.04.35.110.134.3
CFM-DIM99.7*68.447.966.049.288.479.541.918.012.514.920.046.0
ITDS-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
VGG16DIM12.985.8*3.418.22.93.510.63.50.20.30.21.45.2
ODIM33.578.8*10.028.211.213.125.113.62.61.41.34.113.1
SIA50.481.7*13.555.67.320.543.77.72.11.21.69.219.3
BSR24.384.7*3.834.61.77.525.11.40.40.30.32.19.2
Admix-DIM20.288.0*6.823.95.75.416.08.00.80.50.52.18.2
CFM-DIM19.286.3*4.123.73.95.914.06.10.40.60.41.67.2
ITDS-DIM42.0100*24.834.825.719.835.324.98.52.83.27.820.9
MN-v3DIM6.32.599.2*4.96.94.55.516.26.01.82.52.35.4
ODIM24.917.399.4*21.222.122.823.236.120.09.611.611.220.0
SIA32.721.498.7*29.214.126.125.424.219.17.910.610.320.1
BSR22.012.598.9*18.38.714.413.015.911.04.46.57.812.2
Admix-DIM18.06.399.4*11.914.812.113.032.012.53.75.54.47.5
CFM-DIM34.724.999.8*31.629.433.427.844.228.111.814.211.626.5
ITDS-DIM49.527.3100*39.939.244.941.658.046.718.117.013.736.0
RegNDIM24.815.212.492.9*9.312.331.421.43.63.63.013.013.6
ODIM51.343.331.290.3*29.235.656.949.315.712.911.626.133.0
SIA71.272.138.695.7*15.648.072.033.119.717.918.647.741.3
BSR49.063.022.194.1*6.124.947.715.67.08.88.938.026.4
Admix-DIM39.425.220.692.6*14.823.147.733.17.67.17.322.922.6
CFM-DIM71.561.950.893.0*38.361.072.361.622.521.722.143.347.9
ITDS-DIM82.250.568.999.9*64.673.184.975.656.543.040.854.763.2
+ +Table 3. TASRs (%) on twelve pre-trained models with various transformation-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks. ITDS-DIM represents the combination of ITDS, which integrates S1, with DIM, and the boldface represents the results of this combination. + +bility of the two strategies in SDS, ITDS is implemented in two forms: $\mathrm{ITDS}_{fs}$ for S1 (feature space) and $\mathrm{ITDS}_{gs}$ for S2 (geometric space). + +It can be seen that, compared to other competition-based attacks, ITDS demonstrates an overwhelming advantage. + +Taking $\mathrm{ITDS}_{gs}$ as an example, its average black-box TASR (BAvg.) is $9.0\%$ , $7.0\%$ , $3.8\%$ , and $12.4\%$ higher than the existing SOTA method CFM on RN50, VGG16, MN-v3, and RegN, respectively. From the perspective of cross-architecture models, the largest gaps in average black-box + +
AttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
ITDS1100*41.524.340.418.984.670.211.55.52.92.74.227.9
ITDS1+S100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDS1-DIM99.9*71.158.673.356.892.587.148.435.118.819.123.653.1
ITDS1+S-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
+ +Table 4. Evaluation of the effectiveness of two modules in ITDS, where 'I' represents ITC and 'S' represents SDS. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. Boldface represents the best results. Note that single SDS requires combination with existing attack methods, which we discuss in the appendix. + +TASR between it and CFM on CNNs and ViTs are $13.9\%$ and $9.7\%$ on RegN, respectively. It should be noted that the white-box attack performance of our ITDS is also significantly superior, being the best in terms of white-box TASR in any surrogate model, especially on VGG16 and RegN, where it is $13.9\%$ and $7.6\%$ higher than CFM, respectively. In the comparison between $\mathrm{ITDS}_{fs}$ and $\mathrm{ITDS}_{gs}$ , the results seem to be evenly matched, with each showing their own merits in the three trials. Therefore, we believe that both strategies are viable, which precisely proves the effectiveness and diversity of the proposed SDS principle. + +# 4.3. Evaluation on Transformation-based Attacks + +Currently, transformation-based methods are the main approach to adversarial attacks and are the most effective way to improve the transferability of attacks. Numerous studies [2, 4, 16] have shown that non-transformation-based methods combined with DIM can further enhance the transferability of crafted AEs. This combination approach has become a common standard in industry, frequently compared to transformation-based methods for fairness. In this section, we evaluate the TASR of transformation-based and DIM-combined competition-based adversarial attacks. The former includes DIM, ODIM, SIA, and BSR, while the latter includes Admix-DIM, CFM-DIM, and ITDS-DIM. It should be noted that, due to the similar performance of S1 and S2 in SDS, in all subsequent experiments involving ITDS-DIM, we selected $\mathrm{ITDS}_{fs}$ as the representative to combine with DIM. The results are shown in Table 3. + +It can be observed that all competition-based methods show a significant improvement in attack performance after being combined with DIM. Compared to the strongest competition-based attack CFM-DIM, ITDS-DIM achieved average black-box TASR improvements of $15.7\%$ , $13.7\%$ , $9.5\%$ , and $15.3\%$ on RN50, VGG16, MN-v3, and RegN, respectively. The maximum average black-box TASR gap between the two on CNNs and ViTs is $16.1\%$ on RN50 and $21.4\%$ on RegN. Even when compared with the SOTA transformation-based attacks, our method is not inferior, and even surpasses them in all aspects. Compared to the strongest transformation-based attack SIA, ITDS-DIM achieved average black-box TASR improvements of $15.9\%$ and $21.9\%$ on MN-v3 and RegN, respectively. The maximum average black-box TASR gap between the two on + +![](images/d8110ee373581f5fa85dee9e4c5b891eaa13c9a4d31273436d689de330d6072c.jpg) +CNNs and ViTs is $21.3\%$ and $22.8\%$ on RegN, respectively. + +![](images/222e52068242b59fa606527f05786f70506e1e28f7d2617921a261c2309ad180.jpg) +(b) + +![](images/0e7e1b35e618e3529ce6f106e2176493d2789d024f353fb11df5f42fa526ae6f.jpg) +(a) +(c) +Figure 4. Ablation studies on the RN50 model. (a) - (d): TASRs $(\%)$ on the other eleven models with the adversaries crafted by ITDS-DIM, where the default value for $n$ , $\beta$ , $k$ and $T_{sub}$ are set to 3, 1.5, 10 and 10 respectively, when test parameters for each other. + +![](images/7d06bb72df22cfc1b7e323325dfdc2b5bbd9605c872e77ef16058a9ac1299619.jpg) +(d) + +# 4.4. Ablation Studies + +We first present a series of ablation studies conducted on RN50 to assess the impact of each parameter in ITDS on attack performance, including the number of random target class samples $n$ , the competitive ratio of the inverse target gradient $\beta$ , the value of $k$ in top- $k$ , and the number of epochs $t_{sub}$ for non-targeted attack $T_{sub}$ . In all ablation experiments, unless otherwise specified, all parameters other + +
AblationTarget Model
Norm (P1)MI (P2)RN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
✓7.1*0.51.00.40.40.80.30.20.20.20.10.10.4
✓100*72.866.175.865.190.188.254.737.216.917.423.755.3
✓✓100*59.741.366.152.890.991.345.127.814.516.221.147.9
100*79.370.382.869.695.694.662.044.423.624.731.661.7
+ +Table 5. TASRs (%) of ITDS-DIM by ablating inner modules of the ITC. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. In which Norm and MI correspond to P1 and P2 in Section 3.1.2, respectively. Boldface represents the best results. + +
AttackJPEGFDSOGDMPDef4Def8Defl2
SIA83.289.844.31.539.014.054.7
Admix-DIM80.092.947.71.934.814.553.1
CFM-DIM82.592.651.54.125.28.745.5
ITDS-DIM99.499.991.331.250.225.666.8
+ +Table 6. TASRs (\%) of breaking-through four defense methods and three adversarial training defense models. The results for defense methods are based on the average performance of AEs crafted on RN50, VGG16, MN-v3, and RegN respectively, while the results for defense models are based on the performance of AEs crafted by the corresponding models. Bolder for the best. + +than the test parameters are fixed according to the experimental setup. + +As shown in Figure 4a, TASR increases with $n$ , as the diversity of target class images directly affects attack performance. However, the curve levels off after $n = 3$ , and considering computational costs, we selected $n = 3$ . In Figure 4b, most models achieve optimal performance at $\beta = 1.5$ with a few exceptions (e.g., MN-v3). We believe a moderate competitive gradient provides an "exercise" effect, while an overly strong gradient disrupts input features, reducing attack performance. In Figure 4c, optimal performance generally occurs at $k = 5$ or $k = 10$ . We posit that an overly small $k$ under-covers classes and fails to enlarge global margins, whereas an overly large one blurs the optimization. At $k = 10$ , cross-architecture transferability is even better. In Figure 4d, evaluating epochsub $(T_{sub})$ for non-targeted attacks, we found that larger values do not always lead to better performance. As discussed in the ITC section, inverse target gradient perturbations generated at each $T_{sub}$ have diverse potential competitive features, boosting attack performance. At $T_{sub} = 1$ , though it involves more target class samples, it lacks deeper latent features of the target class, limiting attack performance. Beyond a certain $T_{sub}$ value, missing target class features reduce the diversity of inverse gradients, causing performance to regress. + +Furthermore, to verify the effectiveness of the ITC and SDS components in ITDS, we conducted experimental evaluations as shown in Table 4. The results indicate that ITDS using only ITC has already surpassed the attack performance of CFM in Table 2. Regardless of whether DIM is combined or not, SDS can bring additional performance improvements to ITDS. These findings collectively support the + +key roles of ITC and SDS in ITDS. + +To verify the impact of internal modules within the ITC on attack performance, we conducted ablation experiments as shown in Table 5. It can be observed that the use of N-norm may limit the scale of gradient updates, which could result in generated AEs lacking sufficient aggressiveness. On the contrary, if we do not use N-norm (P1), gradient updates will gain greater freedom, which helps to explore the model's decision boundary more deeply, and may thus produce more aggressive AEs. Simultaneously, momentum gradients, as a technique for accumulating previous gradients, will weaken the intensity of inverse target features in the current iteration. Since we wish to maintain the integrity of inverse target features and avoid any normalization, the inverse target momentum gradients in each round will compete with the AE when combined with P1, leading to excessive deformation of AE features, which is disastrous for implementing precisely targeted attacks. + +# 4.5. Evaluation on Advanced Defenses + +In this section, we evaluate the capabilities of various attacks to break through different defense mechanisms. As shown in Table 6, our method exhibits exceptional robustness against a wide range of defense strategies and models. For example, when the Score-OPT method is employed, the average TASR of our approach is nearly $40\%$ higher than that of CFM-DIM. Furthermore, our method achieves the best performance in all three defense models evaluated. + +# 5. Conclusion + +This paper introduces a novel method named ITDS, which significantly improves the transferability of targeted AEs by incorporating the ITC mechanism and the SDS strategy. Experimental results demonstrate that ITDS achieves a substantial increase in average transferable TASRs across various mainstream CNN and ViT models, outperforming SOTA competition-based and transformation-based methods. Our ablation studies further confirmed the key roles of the ITC and SDS components in enhancing attack performance. Moreover, ITDS exhibits superior effectiveness in countering multiple defense methods and models. These findings confirm the effectiveness and practicality of ITDS in enhancing the transferability of targeted AEs and provide a new perspective for the field of adversarial research. + +# Acknowledgments + +This work was supported by funding from the National Natural Science Foundation of China (NSFC) under Grants 62272486, 62172154, and 62372473, the Natural Science Foundation of Hunan Province under Grant 2023JJ70016, the Hong Kong Research Grants Council (RGC) General Research Fund under Grants 152244/21E, 152169/22E, 152228/23E, and 162161/24E, the Research Impact Fund under Grant R5011-23F, the Collaborative Research Fund under Grant C1042-23GF, the NSFC/RGC Collaborative Research Scheme under Grant CRS_HKUST602/24, the Areas of Excellence Scheme under Grant AoE/E-601/22-R, and the InnoHK initiative (HKGAI). Professor Shigeng Zhang is the corresponding author of this paper. + +# References + +[1] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2022. 1, 2, 3, 5 +[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 1, 3, 5, 7 +[3] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 4 +[4] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 7 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5 +[6] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference on Machine Learning, pages 2286-2296. PMLR, 2021. 5 +[7] Goodfellow et al. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2 +[8] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2018. 5 + +[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5 +[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 5 +[11] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5 +[12] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 5 +[13] Kurakin et al. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2 +[14] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Foolmix: Strengthen the transferability of adversarial examples by dual-blending and direction update strategy. IEEE Transactions on Information Forensics and Security, 2024. 1 +[15] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Ucg: A universal cross-domain generator for transferable adversarial examples. IEEE Transactions on Information Forensics and Security, 2024. 1 +[16] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 7 +[17] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature distillation: Dnn-oriented jpeg compression against adversarial examples. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 860-868. IEEE, 2019. 5 +[18] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549-566. Springer, 2022. 1, 2 +[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 5 +[20] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 5 +[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large + +scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5 +[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[23] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1 +[24] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5 +[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 5 +[26] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5 +[27] Francisco Utrera, Evan Kravitz, N Benjamin Erickson, Rajiv Khanna, and Michael W Mahoney. Adversarily-trained deep nets transfer better: Illustration on image classification. arXiv preprint arXiv:2007.05869, 2021. 5 +[28] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 5 +[29] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 1, 2, 5 +[30] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 1, 2, 5 +[31] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 5 +[32] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2 +[33] Juanjuan Weng, Zhiming Luo, Shaozi Li, Nicu Sebe, and Zhun Zhong. Logit margin matters: Improving transferable targeted adversarial attack by logit calibration. IEEE Transactions on Information Forensics and Security, 18:3561-3574, 2023. 2 +[34] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 1, 2, 5 + +[35] Zeng et al. Enhancing targeted transferability via feature space fine-tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4475-4479. IEEE, 2024. 3 +[36] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 36:51810-51829, 2023. 5 +[37] Qilong Zhang, Xiaodan Li, YueFeng Chen, Jingkuan Song, Lianli Gao, Yuan He, et al. Beyond imagenet attack: Towards crafting adversarial examples for black-box domains. In International Conference on Learning Representations, 2022. 1 +[38] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quaxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, pages 1672-1678, 2022. 2 +[39] Zhao et al. On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems, 34:6115-6128, 2021. 2, 3 +[40] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 1 \ No newline at end of file diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/images.zip b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dc7feb7cd899b91e1c009ef6c2ffb2ad991b5db4 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:146ba1ae0868e9f24d40185609ced7857221d48a7a2ba31c020388b49624f80e +size 691568 diff --git a/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/layout.json b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e62a582bc7260e94825a94b722bbb6c2ff0d3c87 --- /dev/null +++ b/2025/Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching/layout.json @@ -0,0 +1,9833 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 550, + 140 + ], + "type": "text", + "content": "Enhancing Transferability of Targeted Adversarial Examples via Inverse Target Gradient Competition and Spatial Distance Stretching" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": "Zhankai Li" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": ", Weiping Wang" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": ", Jie Li" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": ", Shigeng Zhang" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": ", Yunan Hu" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "text", + "content": ", Song Guo" + }, + { + "bbox": [ + 113, + 160, + 493, + 176 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 180, + 368, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 180, + 368, + 194 + ], + "spans": [ + { + "bbox": [ + 243, + 180, + 368, + 194 + ], + "type": "text", + "content": "1Central South University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 169, + 194, + 440, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 169, + 194, + 440, + 209 + ], + "spans": [ + { + "bbox": [ + 169, + 194, + 440, + 209 + ], + "type": "text", + "content": "2The Hong Kong University of Science and Technology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 89, + 211, + 520, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 211, + 520, + 223 + ], + "spans": [ + { + "bbox": [ + 89, + 211, + 520, + 223 + ], + "type": "text", + "content": "janlzk7@gmail.com, {wpwang, lijie55, sgzhang, hyn318}@csu.edu.cn, songguo@ust.hk" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 251, + 200, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 251, + 200, + 262 + ], + "spans": [ + { + "bbox": [ + 151, + 251, + 200, + 262 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "text", + "content": "In the field of AI security, deep neural networks (DNNs) are highly sensitive to adversarial examples (AEs), which can cause incorrect predictions with minimal input perturbations. Although AEs exhibit transferability across models, targeted attack success rates (TASRs) are low due to differences in feature dimensions and decision boundaries. To enhance targeted AE transferability, we propose a novel approach using Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS) in the optimization process. Specifically, we employ a siamese-network-like framework to generate both non-targeted and targeted AEs. The ITC mechanism applies non-targeted adversarial gradients each epoch to impede the optimization of targeted perturbations, thereby improving robustness. Additionally, a top-k SDS strategy guides AEs to penetrate target class regions in the latent space while distancing from nontargeted regions, achieving optimal transferability. Compared to state-of-the-art competition-based attacks, our method significantly improves transferable TASRs by " + }, + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "inline_equation", + "content": "16.1\\%" + }, + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "inline_equation", + "content": "21.4\\%" + }, + { + "bbox": [ + 55, + 275, + 297, + 539 + ], + "type": "text", + "content": " on mainstream CNNs and ViTs, respectively, and demonstrates superior defense-breaking capabilities. Our code is available here." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 135, + 574 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "Adversarial attacks involve adding small, often imperceptible perturbations to input data, which can lead otherwise high-performing models to produce incorrect outputs. The concept of adversarial attacks was first introduced by [23] and further developed by [7], who demonstrated the vulnerability of deep neural networks (DNNs) to these perturbations. With the rapid advancement of DNNs in recent years, the security risks revealed by adversarial attacks have become a critical issue. Adversarial attacks are now widely used to assess model security, improve model robustness, and drive research in defense technologies." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 251, + 555, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 539 + ], + "type": "text", + "content": "However, the significance of adversarial attacks extends beyond these applications, as they have inadvertently become a widely used defense mechanism in real-world scenarios. A prime example is Google's login verification system, reCAPTCHA, which aims to distinguish real users from bots to prevent abuse. Adversarial examples (AEs) are one of the primary image verification methods employed, as illustrated in Figure 1. This raises an important question: is it secure enough? Since we do not know the bot's model architecture or training domain, the transferability of AEs becomes especially crucial. Most current research on adversarial transferability focuses on non-targeted attacks [14, 40], even extending to cross-architecture and cross-domain attacks [15, 37], achieving notable performance. In contrast, targeted attacks face a significant increase in difficulty due to the substantial changes in decision boundaries, and as a result, they still maintain a relatively low success rate. However, for login verification, given that the output category of non-targeted attacks is uncontrollable, and even the original verification images without perturbation may have \"low-quality\" issues (leading to misidentification by bots), in such cases, targeted AEs are often safer and more reliable than non-targeted AEs that may cause the deployed bots to misjudge into the ground truth category." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 715 + ], + "type": "text", + "content": "Currently, effective targeted attack methods are primarily based on transformation [1, 18, 29, 34], achieving transferability through input augmentation. However, the relentless pursuit and stacking of a greater variety of transformations, although enhancing transferability to some extent, does not guarantee robust performance in white-box attacks. Moreover, this stacking approach based on affine and perspective transformations offers limited insight into the complexity and vulnerability of DNNs. Recently, competition-based attacks have been proposed [2, 30], offering a unique perspective. Their advantage lies in compelling AEs to consider the more complex and diverse feature representations within DNNs during their generation, simulating various noise interferences and attack strategies," + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3716" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 73, + 70, + 277, + 208 + ], + "blocks": [ + { + "bbox": [ + 73, + 70, + 277, + 208 + ], + "lines": [ + { + "bbox": [ + 73, + 70, + 277, + 208 + ], + "spans": [ + { + "bbox": [ + 73, + 70, + 277, + 208 + ], + "type": "image", + "image_path": "c4522ccb122a1c16a16b7d176ecd530a786125a1646d71d9284388dcab84ce61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 218, + 297, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 297, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 297, + 253 + ], + "type": "text", + "content": "Figure 1. Example of Google's web login verification system, reCAPTCHA. It determines whether a user is human or a bot based on whether they can correctly match text to corresponding images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 263, + 295, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 295, + 357 + ], + "type": "text", + "content": "and promoting the learning of more robust feature combinations. The competitive mechanism, by delving into the model's decision boundaries, helps to uncover the latent vulnerabilities of DNNs and provides a new perspective for understanding and improving the model's generalization capabilities. However, no matter what the attack method is, it cannot conceal the widespread insufficiency in the transferability performance of current targeted adversarial attacks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 360, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 296, + 601 + ], + "type": "text", + "content": "To address this challenge, we propose a new targeted adversarial attack method called ITDS, which consists of two components: Inverse Target Gradient Competition (ITC) and Spatial Distance Stretching (SDS), aimed at enhancing the transferability of targeted AEs. In the first part, we adopt a siamese-network-like framework, innovatively leveraging the ground truth gradients of multiple target class non-targeted AEs at intermediate stages to exert inverse \"competitiveness\" on non-target class targeted AEs during the optimization process. In the second part, to tackle the multi-dimensional complexity of decision boundaries in targeted attacks, we introduce a top-" + }, + { + "bbox": [ + 56, + 360, + 296, + 601 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 360, + 296, + 601 + ], + "type": "text", + "content": " SDS strategy to seek stable perturbation update directions. This guides AEs to penetrate target class regions within the latent multidimensional space while globally distancing from several closest non-target region boundaries. The experimental results demonstrate that the proposed ITDS method has achieved significant advantages in enhancing the transferability and defense-breaking capabilities of targeted AEs, surpassing the existing state-of-the-art (SOTA) methods." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 609, + 143, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 143, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 143, + 621 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 296, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 296, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 296, + 665 + ], + "type": "text", + "content": "In this section, we will briefly overview the background of adversarial attacks. Then focus on various targeted attack techniques that demonstrate effective transferability." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 672, + 173, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 173, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 173, + 683 + ], + "type": "text", + "content": "2.1. Adversarial Attacks" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Goodfellow et.al first designed a simple yet effective method called Fast Gradient Sign Method (FGSM) [7]. This" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": "method generates AEs by making small adjustments to the input " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " (original sample) along the direction of the gradient, i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "x^{adv} = x + \\delta" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ". In the case of targeted attacks, FGSM updates the input " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": " in the direction that minimizes the classification loss for a given target class " + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 72, + 553, + 144 + ], + "type": "text", + "content": ", with the objective function as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 339, + 150, + 553, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 150, + 553, + 170 + ], + "spans": [ + { + "bbox": [ + 339, + 150, + 553, + 170 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {x ^ {a d v}} \\mathcal {J} (x ^ {a d v}, y; \\theta) \\text {s . t .} \\| x - x ^ {a d v} \\| _ {\\infty} \\leq \\epsilon , \\tag {1}", + "image_path": "a3781e78e41fec88e2427c5744c5b50922176fd8fec4400811b5cf27f755113a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "content": " is the model parameters, " + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "inline_equation", + "content": "\\mathcal{J}" + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "content": " denotes the Cross-Entropy (CE) loss function, and the " + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "inline_equation", + "content": "\\ell_{\\infty}" + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "content": " norm is used to constrain the perturbation within the range " + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 175, + 553, + 210 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 319 + ], + "type": "text", + "content": "I-FGSM [13] is an iterative extension of FGSM, which gradually refines adversarial perturbation with small step sizes. To address gradient vanishing issues in targeted attacks, [39] proposes using Logits loss instead of CE loss, significantly enhancing targeted attack effectiveness. Based on this, [33] further proposes downscaling logit calibration with a temperature factor and an adaptive margin. While these methods provide a foundation for enhancing targeted attack effectiveness, their transferability remains limited." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 325, + 494, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 494, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 494, + 337 + ], + "type": "text", + "content": "2.2. Transferable Adversarial Attacks" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": "In this section, we introduce advanced transferable attacks that can successfully deceive other models using AEs generated on a single surrogate model without additional training. We categorize these methods into transformation-based and competitive-based attacks, focusing primarily on transferable targeted attacks, along with certain non-targeted attacks that can also be adapted for targeted attacks." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 617 + ], + "type": "text", + "content": "Transformation-Based Attacks. DIM [34] is a representative transformation-based attack, which resizes and pads images with a certain probability at each step of the multi-step attack process. SSM [18] applies frequency domain augmentations through spectrum transformations on the input. ODIM [1] method significantly improves the targeted attack transferability by mapping input images to random 3D objects and applying varied rendering techniques. SIA [31] generates structurally diverse AEs by applying various random transformations to the images. Both ODIM and SIA integrate DIM into their implementation principles and share a similar lineage in transformation strategies. BSR [29] randomly shuffles and rotates the blocks within images, calculates the ensemble gradients of this new set of images, reduces attention heatmap differences across various models, and significantly enhances transferability." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Competition-based Attacks. Admix [30] combines features from different categories in the input domain, creating multiple images for gradient computation, and enhancing transferability without degrading white-box performance. FIA [32] employs aggregated gradients to disrupt essential object-aware features while suppressing model-specific features, boosting AE transferability. RPA [38] adopts a similar approach to FIA, introducing random patch transforma" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3717" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 192 + ], + "type": "text", + "content": "tions to benign images. To alleviate the overfitting dilemma common in an AE crafted by simple iterative attacks, FFT [35] encourages features conducive to the target class while discouraging those associated with the original class in an intermediate layer of the source model. CFM [2] introduces a competitive strategy, crafting targeted adversarial perturbations with diverse features by introducing two types of competitor noises: adverse perturbations toward different target classes and friendly perturbations towards the correct class, achieving SOTA results." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 173, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 173, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 173, + 216 + ], + "type": "text", + "content": "3. The Design of ITDS" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 296, + 271 + ], + "type": "text", + "content": "This section consists of two parts: the Inverse Target Gradient Competition mechanism and the Spatial Distance Stretching strategy. In each part, we first explain the motivation and then propose the corresponding method." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 282, + 290 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 282, + 290 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 282, + 290 + ], + "type": "text", + "content": "3.1. Inverse Target Gradient Competition (ITC)" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 295, + 162, + 306 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 295, + 162, + 306 + ], + "spans": [ + { + "bbox": [ + 55, + 295, + 162, + 306 + ], + "type": "text", + "content": "3.1.1. Motivation of ITC" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 310, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 514 + ], + "type": "text", + "content": "The motivation for the first part of our research arises from a question: What is a competition mechanism? From existing work related to competition mechanisms, we summarize it as a way to enhance the transferability of AEs by promoting diversified attack strategies through simulating various noise interferences. Since noise interference is involved, the essence of competitiveness can be interpreted as an inverse force. For targeted attacks, could we identify the most competitive noise? We believe this is the key issue. To address this critical question, we turn to target class samples and consider the gradients of the target label for these samples. These gradients essentially represent the optimal force linking target class samples with the target label, which can be used in both forward and inverse directions. Therefore, we have a bold hypothesis: when using this gradient in inverse, it not only effectively generates non-targeted AEs but also serves as the optimal competitiveness for targeted attacks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 519, + 149, + 530 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 149, + 530 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 149, + 530 + ], + "type": "text", + "content": "3.1.2. Method of ITC" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Figure 2 shows the overall framework of our Inverse Target Gradient Competition mechanism, which is jointly driven by both non-targeted and targeted attacks within a siamese-network-like architecture. Our strategy is to utilize the inverse target gradient to introduce resistance during the optimization process of targeted AEs. This inverse target gradient can be generated for any samples; but which inverse target gradient should we choose? As mentioned earlier, we seek the strongest resistance, which is the inverse gradient of the target label with respect to the target class sample. The approach of generating the inverse target gradient is illustrated in the lower part of Figure 2, where we use 'cock' as the target label as an example. In each sub-epochs (epochssub), multiple target class images that are correctly classified by the siamese classifier are selected to generate" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "non-targeted AEs. The optimization goal is to minimize the logit value of the target label. We set " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{epoch}_{sub}" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " to be one-tenth of the total epochs, as we believe that target class samples need a certain number of iterations to transform into robust non-targeted AEs. Accordingly, the non-targeted AEs generated in each " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{epoch}_{sub}" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " contain different levels and increasingly deeper latent features of target class. Therefore, we consider the inverse target gradient perturbations generated in each " + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{epoch}_{sub}" + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": " to also have diverse potential competitive characteristics, and during the generation of nontargeted AEs in each epoch, multiple inverse target perturbations are generated to enrich competitive characteristics. The formulas are as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 380, + 236, + 555, + 253 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 236, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 380, + 236, + 555, + 253 + ], + "type": "interline_equation", + "content": "\\tilde {g} _ {t} ^ {z} = \\nabla_ {z _ {t} ^ {a d v}} \\mathcal {L} (z _ {t} ^ {a d v}, y; \\theta), \\tag {2}", + "image_path": "190b31460b27f4dfe9982fc24d3019b7c469a3c1c52ad12987a2cae2ccd29efe.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 392, + 271, + 555, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 271, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 392, + 271, + 555, + 285 + ], + "type": "interline_equation", + "content": "g _ {t + 1} ^ {z} = \\mu \\cdot g _ {t} ^ {z} + \\tilde {g} _ {t} ^ {z}, \\tag {3}", + "image_path": "84080f61d57f75fa7e03211a6e0cdedaa96489311ecf8e2ba6c660297c5d2351.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 351, + 300, + 555, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 300, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 351, + 300, + 555, + 316 + ], + "type": "interline_equation", + "content": "z _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {z, \\epsilon} \\{z _ {t} ^ {a d v} - \\alpha \\cdot \\operatorname {s i g n} (g _ {t + 1} ^ {z}) \\}, \\tag {4}", + "image_path": "274125f1e0766c777555cb96c4be7204b1971a6765de1091902f0b4fc0d05ac1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": " is the target class sample, " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t^z" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": " represents the inverse target gradient, and we follow [1, 2, 39] to use the Logit loss " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": " (opposite to CE loss optimization). " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "g_{t + 1}^{z}" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": " represents the momentum gradient without normalization for " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "inline_equation", + "content": "z_{t + 1}^{adv}" + }, + { + "bbox": [ + 313, + 319, + 555, + 380 + ], + "type": "text", + "content": " is the non-targeted AE generated in each round." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 390, + 555, + 547 + ], + "blocks": [ + { + "bbox": [ + 315, + 390, + 555, + 547 + ], + "lines": [ + { + "bbox": [ + 315, + 390, + 555, + 547 + ], + "spans": [ + { + "bbox": [ + 315, + 390, + 555, + 547 + ], + "type": "image", + "image_path": "a377ae4ff4cfa262b59dbfa1e310b1f38b511cb76d59b64e1f06b23f8c1d6607.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 554, + 556, + 643 + ], + "lines": [ + { + "bbox": [ + 313, + 554, + 556, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 556, + 643 + ], + "type": "text", + "content": "Figure 2. The overall framework of the ITC, e.g., combined with DIM, utilizes a siamese-network-like architecture to generate targeted and non-targeted AEs simultaneously. The upper part focuses on targeted attacks against non-target class samples, while the lower part addresses non-targeted attacks on target class samples. Note that for non-targeted attacks, the optimization condition utilizes the logits value of the target label, whereas the goal of the targeted attack is to stretch the spatial distance proposed in SDS." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "The process of generating targeted AEs is shown in the upper part of Figure 2, where the input is a non-target image, and the optimization goal is to maximize the spatial distance, which we introduce in the SDS of the next section. Each round of generated targeted AEs competes with" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3718" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "a certain proportion of inverse target gradients from different stages. The formulas are as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 58, + 102, + 294, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 102, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 294, + 146 + ], + "type": "interline_equation", + "content": "\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {L} \\left(\\delta_ {i} \\cdot \\left(x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}\\right), y; \\theta\\right), \\tag {5}", + "image_path": "1cb354b9f129de754b1d4c8fc5ea8578a1d72244b8f7934bf7b9595cab46abe4.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 132, + 156, + 295, + 171 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 156, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 132, + 156, + 295, + 171 + ], + "type": "interline_equation", + "content": "g _ {t + 1} = \\mu \\cdot g _ {t} + \\bar {g} _ {t} ^ {\\text {c p t}}, \\tag {6}", + "image_path": "381c26f1a96450fcb79637aa473b7220e9cbef7685ba3b6ce940a754d1b561cc.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 184, + 295, + 199 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 184, + 295, + 199 + ], + "spans": [ + { + "bbox": [ + 93, + 184, + 295, + 199 + ], + "type": "interline_equation", + "content": "x _ {t + 1} ^ {a d v} = \\operatorname {C l i p} _ {x, \\epsilon} \\left\\{x _ {t} ^ {a d v} + \\alpha \\cdot \\operatorname {s i g n} \\left(g _ {t + 1}\\right) \\right\\}, \\tag {7}", + "image_path": "37e6159f9f02a1dcf8dd6a8fbe6fc543398483aed5486acddb991722afa66a2c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": " represents the number of target class samples in per epochs " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "_{sub}" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\delta_{i}" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": " are the scaling factor and level, " + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 201, + 296, + 250 + ], + "type": "text", + "content": " is the competition ratio, and each round of targeted AEs competes with the corresponding inverse target gradient." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 250, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 405 + ], + "type": "text", + "content": "There are two notable points. The first point (P1) is whether, for targeted or non-targeted attacks, the calculation of momentum gradients differs from the traditional MIFGSM [3], and we have not employed norm normalization (N-norm). One of the reasons is the absence of gradient explosion in our experimental evaluations, and more importantly, our experiments have shown that using N-norm would limit our attack performance. We believe that N-norm might restrict the magnitude of gradient updates, resulting in less aggressive generation of AEs. Without N-norm, gradient updates could become more flexible, allowing for more effective exploration of the model's decision boundary, thereby generating more aggressive AEs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 405, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 405, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 405, + 296, + 525 + ], + "type": "text", + "content": "The second point (P2) is that we did not use the momentum gradient as the competitive gradient. The reason is that the momentum gradient is essentially a cumulative gradient. It weakens the strength of the inverse target features in the corresponding round. Since we want to maintain the integrity of the inverse target features, using it without any normalization operation in conjunction with P1 would cause excessive damage to the input features. This would be catastrophic for targeted attacks. These two points will also be discussed in subsequent ablation experiments." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 58, + 536, + 111, + 586 + ], + "blocks": [ + { + "bbox": [ + 58, + 536, + 111, + 586 + ], + "lines": [ + { + "bbox": [ + 58, + 536, + 111, + 586 + ], + "spans": [ + { + "bbox": [ + 58, + 536, + 111, + 586 + ], + "type": "image", + "image_path": "5883614a28bfcc96c9c66418acb147b6bac45409622873337a6338d6a8f4bb9a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 587, + 91, + 597 + ], + "lines": [ + { + "bbox": [ + 80, + 587, + 91, + 597 + ], + "spans": [ + { + "bbox": [ + 80, + 587, + 91, + 597 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 116, + 536, + 168, + 586 + ], + "blocks": [ + { + "bbox": [ + 116, + 536, + 168, + 586 + ], + "lines": [ + { + "bbox": [ + 116, + 536, + 168, + 586 + ], + "spans": [ + { + "bbox": [ + 116, + 536, + 168, + 586 + ], + "type": "image", + "image_path": "0c50cd2ec667da5ca912afa440004cf14cdee048fe3fdc36ac2dc304bd1083f5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 138, + 587, + 149, + 597 + ], + "lines": [ + { + "bbox": [ + 138, + 587, + 149, + 597 + ], + "spans": [ + { + "bbox": [ + 138, + 587, + 149, + 597 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 184, + 536, + 236, + 586 + ], + "blocks": [ + { + "bbox": [ + 184, + 536, + 236, + 586 + ], + "lines": [ + { + "bbox": [ + 184, + 536, + 236, + 586 + ], + "spans": [ + { + "bbox": [ + 184, + 536, + 236, + 586 + ], + "type": "image", + "image_path": "5928543575ec402b6c0440fb86cb17e6ed7c8e24720e1f3fa575d6875ab41d3a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 206, + 587, + 217, + 596 + ], + "lines": [ + { + "bbox": [ + 206, + 587, + 217, + 596 + ], + "spans": [ + { + "bbox": [ + 206, + 587, + 217, + 596 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 605, + 296, + 640 + ], + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 640 + ], + "type": "text", + "content": "Figure 3. Comparison of targeted AEs with strong and weak transferability. Panels (a) and (c) show AEs with weak transferability, while (b) and (d) show AEs with strong transferability." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 242, + 536, + 294, + 586 + ], + "blocks": [ + { + "bbox": [ + 242, + 536, + 294, + 586 + ], + "lines": [ + { + "bbox": [ + 242, + 536, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 242, + 536, + 294, + 586 + ], + "type": "image", + "image_path": "6fbab54b3fc182edb5b5b4a1347815fbee0ab12b2732a16a8abaabee283e3fba.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 264, + 587, + 275, + 596 + ], + "lines": [ + { + "bbox": [ + 264, + 587, + 275, + 596 + ], + "spans": [ + { + "bbox": [ + 264, + 587, + 275, + 596 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 656, + 235, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 235, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 235, + 670 + ], + "type": "text", + "content": "3.2. Spatial Distance Stretching (SDS)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 674, + 163, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 674, + 163, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 674, + 163, + 685 + ], + "type": "text", + "content": "3.2.1. Motivation of SDS" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "The motivation for the second part of our study arises from an interesting phenomenon commonly observed in highly" + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 552, + 124 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 552, + 124 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 552, + 124 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 552, + 124 + ], + "type": "table", + "html": "
LabelFigure12345
motor_scooter(a)42.0%5.5%4.3%3.6%3.2%
(b)83.3%4.9%2.1%1.1%0.3%
siamese_cat(c)49.6%5.6%5.3%5.1%5.0%
(d)90.5%2.1%1.3%0.7%0.2%
", + "image_path": "3b364741ee6f638259da75a58d739f5a79f68750b315741ae8f710378eb1dd3e.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_body" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 131, + 553, + 155 + ], + "lines": [ + { + "bbox": [ + 313, + 131, + 553, + 155 + ], + "spans": [ + { + "bbox": [ + 313, + 131, + 553, + 155 + ], + "type": "text", + "content": "Table 1. The average top-5 confidence on black-box models successfully attacked by targeted AEs as shown in Figure 3." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 315, + 167, + 485, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 167, + 485, + 180 + ], + "spans": [ + { + "bbox": [ + 315, + 167, + 485, + 180 + ], + "type": "text", + "content": "Algorithm 1 The ITDS Attack Algorithm" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": "Input: Siamese classifier " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": " with parameters " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": ", Logit loss " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": ", our Distance loss " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": ", and an original sample " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "text", + "content": " with target label " + }, + { + "bbox": [ + 313, + 191, + 545, + 227 + ], + "type": "inline_equation", + "content": "y" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "text", + "content": "Input: The maximum perturbation " + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "text", + "content": ", the step size " + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "text", + "content": ", the number of epochs " + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 543, + 251 + ], + "type": "inline_equation", + "content": "\\mathrm{epochs}_{sub} T_{sub}" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": "Input: The decay factor " + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": ", scaling level " + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": " random target class samples " + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "inline_equation", + "content": "z \\in Z" + }, + { + "bbox": [ + 313, + 251, + 548, + 274 + ], + "type": "text", + "content": ")" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "type": "text", + "content": "Input: The competition ratio " + }, + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "type": "text", + "content": ", and topk classifications " + }, + { + "bbox": [ + 313, + 275, + 549, + 288 + ], + "type": "inline_equation", + "content": "k" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 288, + 473, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 288, + 473, + 299 + ], + "spans": [ + { + "bbox": [ + 314, + 288, + 473, + 299 + ], + "type": "text", + "content": "Output: An adversarial example " + }, + { + "bbox": [ + 314, + 288, + 473, + 299 + ], + "type": "inline_equation", + "content": "x^{adv}" + }, + { + "bbox": [ + 314, + 288, + 473, + 299 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 318, + 308, + 553, + 464 + ], + "type": "list", + "angle": 0, + "index": 37, + "blocks": [ + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "spans": [ + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "text", + "content": "1: " + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "inline_equation", + "content": "g_0^z = 0" + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "inline_equation", + "content": "g_0 = 0" + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "inline_equation", + "content": "z_0^{adv} = z" + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 320, + 308, + 483, + 322 + ], + "type": "inline_equation", + "content": "x_0^{adv} = x" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "spans": [ + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "text", + "content": "2: for " + }, + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "text", + "content": " in range " + }, + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "inline_equation", + "content": "(T)" + }, + { + "bbox": [ + 321, + 322, + 420, + 333 + ], + "type": "text", + "content": " do" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 333, + 430, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 333, + 430, + 344 + ], + "spans": [ + { + "bbox": [ + 321, + 333, + 430, + 344 + ], + "type": "text", + "content": "3: if " + }, + { + "bbox": [ + 321, + 333, + 430, + 344 + ], + "type": "inline_equation", + "content": "t\\% T_{sub} == 0" + }, + { + "bbox": [ + 321, + 333, + 430, + 344 + ], + "type": "text", + "content": " then" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "spans": [ + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "type": "text", + "content": "4: Reselect " + }, + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "type": "text", + "content": " random target class samples " + }, + { + "bbox": [ + 321, + 345, + 521, + 357 + ], + "type": "inline_equation", + "content": "Z" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 321, + 357, + 369, + 367 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 357, + 369, + 367 + ], + "spans": [ + { + "bbox": [ + 321, + 357, + 369, + 367 + ], + "type": "text", + "content": "5: end if" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "text", + "content": "6: Get the inverse target gradient for each " + }, + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "inline_equation", + "content": "\\tilde{g}_t^z" + }, + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 321, + 369, + 553, + 392 + ], + "type": "text", + "content": " random target class samples by Eq.(2)" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 321, + 393, + 553, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 393, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 321, + 393, + 553, + 417 + ], + "type": "text", + "content": "7: Get the momentum inverse target gradient for each " + }, + { + "bbox": [ + 321, + 393, + 553, + 417 + ], + "type": "inline_equation", + "content": "g_{t+1}^{z}" + }, + { + "bbox": [ + 321, + 393, + 553, + 417 + ], + "type": "text", + "content": " by Eq.(3)" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 321, + 417, + 455, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 417, + 455, + 429 + ], + "spans": [ + { + "bbox": [ + 321, + 417, + 455, + 429 + ], + "type": "text", + "content": "8: Update each " + }, + { + "bbox": [ + 321, + 417, + 455, + 429 + ], + "type": "inline_equation", + "content": "z_{t+1}^{adv}" + }, + { + "bbox": [ + 321, + 417, + 455, + 429 + ], + "type": "text", + "content": " by Eq.(4)" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "spans": [ + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "text", + "content": "9: Get the set " + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "text", + "content": " of top- " + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "text", + "content": " classifications by " + }, + { + "bbox": [ + 321, + 429, + 537, + 441 + ], + "type": "inline_equation", + "content": "f(x_{t}^{adv})" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 318, + 441, + 553, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 441, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 318, + 441, + 553, + 464 + ], + "type": "text", + "content": "10: Calculate the integrate gradient of competition examples by:" + } + ] + } + ], + "index": 36 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 340, + 474, + 553, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 340, + 474, + 553, + 506 + ], + "spans": [ + { + "bbox": [ + 340, + 474, + 553, + 506 + ], + "type": "interline_equation", + "content": "\\bar {g} _ {t} ^ {c p t} = \\frac {1}{n \\cdot m} \\sum_ {z \\in Z} \\sum_ {i = 0} ^ {m - 1} \\nabla_ {x _ {t} ^ {a d v}} \\mathcal {D} \\big (\\delta_ {i} \\cdot (x _ {t} ^ {a d v} - \\beta \\cdot \\tilde {g} _ {t} ^ {z}), y \\big)", + "image_path": "f89ca770463bdf9fd23f1e8a44dab378c4a0904e45e16e6eab06f22c246cb398.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 317, + 520, + 531, + 568 + ], + "type": "list", + "angle": 0, + "index": 43, + "blocks": [ + { + "bbox": [ + 317, + 520, + 496, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 496, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 496, + 532 + ], + "type": "text", + "content": "11: Get the momentum gradient by Eq.(6)" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 317, + 532, + 531, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 532, + 531, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 532, + 531, + 544 + ], + "type": "text", + "content": "12: Update the adversarial example " + }, + { + "bbox": [ + 317, + 532, + 531, + 544 + ], + "type": "inline_equation", + "content": "x_{t + 1}^{adv}" + }, + { + "bbox": [ + 317, + 532, + 531, + 544 + ], + "type": "text", + "content": " by Eq.(7)" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 317, + 544, + 365, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 544, + 365, + 554 + ], + "spans": [ + { + "bbox": [ + 317, + 544, + 365, + 554 + ], + "type": "text", + "content": "13: end for" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 317, + 555, + 419, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 419, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 419, + 568 + ], + "type": "text", + "content": "14: return " + }, + { + "bbox": [ + 317, + 555, + 419, + 568 + ], + "type": "inline_equation", + "content": "x^{adv} = x_T^{adv}" + }, + { + "bbox": [ + 317, + 555, + 419, + 568 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 42 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "transferable targeted AEs. As shown in Figure 3, we present two 'cock'-targeted examples as representatives: targeted AEs generated using ResNet50 with strong and weak transferability, from left to right. In (a) and (c), we have AEs that can only fool fewer than three black-box models into predicting the target class, while (b) and (d) show AEs capable of misleading over twelve black-box models to the target class. Table 1 presents the top-5 average confidence scores for each AE on the successfully misled black-box models. It can be seen that the average confidence for the target la" + } + ] + } + ], + "index": 44 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3719" + } + ] + } + ], + "index": 45 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 145 + ], + "type": "text", + "content": "bel (top-1) in highly transferable AEs is significantly higher than that in low-transferability AEs. Notably, highly transferable AEs not only exhibit higher top-1 confidence but also have lower confidence scores for subsequent rankings (rank-2 and beyond) compared to low-transferability AEs, with each subsequent rank showing an even larger gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 148, + 149, + 160 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 148, + 149, + 160 + ], + "spans": [ + { + "bbox": [ + 55, + 148, + 149, + 160 + ], + "type": "text", + "content": "3.2.2. Method of SDS" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 164, + 296, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 248 + ], + "type": "text", + "content": "Based on this observation, we naturally propose a hypothesis by inverse inference: increasing the gap between the target class and other classes under a top-" + }, + { + "bbox": [ + 55, + 164, + 296, + 248 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 164, + 296, + 248 + ], + "type": "text", + "content": " setting can enhance adversarial transferability. Therefore, the key question becomes how to widen this gap. To address this, we devised two strategies from different perspectives, as shown in the following formulas:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 85, + 254, + 295, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 254, + 295, + 285 + ], + "spans": [ + { + "bbox": [ + 85, + 254, + 295, + 285 + ], + "type": "interline_equation", + "content": "f _ {u n t g} (x) = \\frac {1}{k - 1} \\sum_ {q \\in Q} f _ {q} (x) \\quad \\text {s . t .} \\quad y \\notin Q, \\tag {8}", + "image_path": "b456569ec553b9fad84becaf971c70dca3a63143a68fbfdef8554867aa489029.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 90, + 293, + 296, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 293, + 296, + 311 + ], + "spans": [ + { + "bbox": [ + 90, + 293, + 296, + 311 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {S 1} (x _ {t} ^ {a d v}, y) = f _ {y} (x _ {t} ^ {a d v}) - f _ {u n t g} (x _ {t} ^ {a d v}), \\qquad (9)", + "image_path": "90ceaae726c666014e214a3c0e1206875f8c6b4b63d013fd088f8b5d32c23f00.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 296, + 372 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 296, + 372 + ], + "type": "interline_equation", + "content": "\\mathcal {D} _ {S 2} \\left(x _ {t} ^ {a d v}, y\\right) = \\frac {f _ {y} \\left(x _ {t} ^ {a d v}\\right) - f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right)}{\\left\\| \\nabla_ {x _ {t} ^ {a d v}} f _ {y} \\left(x _ {t} ^ {a d v}\\right) - \\nabla_ {x _ {t} ^ {a d v}} f _ {u n t g} \\left(x _ {t} ^ {a d v}\\right) \\right\\| _ {2}}. \\tag {10}", + "image_path": "403fbc575ad50a14006a3d2f6396ca7160f71deb250a73a6d9d3768721c82baa.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": "The first strategy (S1) maximizes the target class logit value relative to the aggregated logits of other classes within the top-" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " setting, enhancing target class features in the latent feature space and diminishing those of similar other classes, thus widening the latent feature gap, as shown in Equation (9). The second strategy (S2), inspired by DeepFool's approach to calculating point-to-hyperplane distances [19], guides AEs to deeply infiltrate the target class region in multi-dimensional geometric space while maintaining a global distance from the aggregated boundaries of the closest non-target regions, as detailed in Equation (10). In these equations, " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "f(\\cdot)" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " represents the classifier, " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " denotes the set of top-" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " classes (excluding " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": "), and " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "f_{q}(x)" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "f_{y}(x)" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " represent the output values of " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "f(x)" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " corresponding to " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 373, + 296, + 552 + ], + "type": "text", + "content": ", respectively. The algorithm of ITDS is shown in Algorithm 1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 137, + 575 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 641 + ], + "type": "text", + "content": "We first outline the experimental setup, then evaluate the targeted-attack performance of both competition-based and transformation-based methods separately. Afterward, we evaluate the ability of our ITDS to break through defenses and conduct ablation studies." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 648, + 172, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 172, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 172, + 661 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 296, + 714 + ], + "type": "text", + "content": "Dataset. To ensure rigorous evaluation, we randomly selected 2,000 high-quality images from 1,000 classes in the ILSVRC2012 validation set [21]. These images were correctly identified by all models tested, including CNNs, ViTs" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": "and adversarial training models, using intersection extraction, with more than " + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 120 + ], + "type": "text", + "content": " top-1 average confidence. The target class images of our method were randomly drawn from the corresponding category in the ILSVRC2012 dataset." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 121, + 554, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 554, + 240 + ], + "type": "text", + "content": "Models. We choose four normally pre-trained CNNs, i.e., ResNet50 (RN50) [9], VGG16 [22], MobileNet-v3-large (MN-v3) [11], RegNet-y-32gf (RegN) [20] as surrogate (white-box) models. Concurrently, we choose four white-box independent normally pre-trained CNNs, i.e., Inception-v3 (Inc-v3) [24], ResNet101 (RN101), DenseNet161 (DN161) [12], EfficientNet-b7 (ENet) [25], and four ViTs, i.e., VisionTransformer-b-16 (ViT) [5], DeiT-base (DeiT) [26], ConViT-base (ConViT) [6] and PiT-base (PiT) [10] to serve as target black-box models." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 241, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 553, + 312 + ], + "type": "text", + "content": "Baselines. We adopt six challenging adversarial attacks as our baselines, which have released the code and provide the necessary parameters for reproduction, i.e., Competition-based: Admix [30], CFM [2], Transformation-Based: DIM [34], ODIM [1], SIA [31], and BSR [29]. All the baselines are combined with MI-FGSM." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": "Defenses. To further demonstrate the effectiveness of ITDS, we consider four additional advanced defense methods, i.e., JPEG [8], FD [17], GDMP [28], Score-Opt (SO) [36], and three adversarial training defense models [27], i.e., " + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_{\\text{lin}f4}" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_4" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_{\\text{lin}f8}" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_8" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_{l2-3}" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "inline_equation", + "content": "\\mathrm{Def}_{l2}" + }, + { + "bbox": [ + 313, + 314, + 554, + 397 + ], + "type": "text", + "content": ") from ResNet50, which are proven to be robust to adversarial attacks on ImageNet datasets." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": "Attack setting. For all methods, we set the maximum perturbation of " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\epsilon = 16" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", the number of epochs " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "T = 100" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", the step size " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\alpha = 1.6" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ". We adopt the decay factor " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mu = 1.0" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": " for all the methods that employ the MI-FGSM, and a transformation probability of " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "p = 0.7" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": " for DIM and ODIM. Furthermore, all parameters for Admix, SIA, CFM, and BSR are set as described in their respective published papers. For our ITDS, we set the number of epochs " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "_{sub}" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "T / 10" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", scaling level " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "m = 5" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\delta_{i} = 1 / 2^{i}" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", the number of random target class samples " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", the competition ratio " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\beta = 1.5" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": " of the top-" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 398, + 554, + 555 + ], + "type": "text", + "content": ". For each AE of the trial, the target label is randomly chosen from the other 999 classes excluding its true label, and the final results are averaged over three trials." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 533, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 533, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 533, + 577 + ], + "type": "text", + "content": "4.2. Evaluation on Competition-based Attacks" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 713 + ], + "type": "text", + "content": "We first evaluate the attack performance of various competition-based attacks. Since the FIA and RPA introduced in related work are only applicable to non-targeted attacks, and FFT must be combined with existing methods to be usable, we ultimately chose Admix and CFM as competitors for ITDS. To demonstrate the versatility of our approach, the surrogate models we use to craft AEs are structurally independent of each other. The targeted attack success rates (TASRs) are shown in Table 2, where the upper row indicates the targeted model under attack, and the left column lists the tested surrogate models. To test the feasi-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3720" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 547, + 246 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 547, + 246 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 547, + 246 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 547, + 246 + ], + "type": "table", + "html": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50Admix100*7.92.14.91.638.619.80.60.30.10.20.16.9
CFM100*45.017.735.59.587.760.75.62.01.51.53.924.6
ITDSfs100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDSgs100*46.235.049.328.487.680.417.910.04.14.46.633.6
VGG16Admix4.388.4*1.47.70.70.94.10.50.10.10.10.51.8
CFM3.986.1*1.28.40.81.02.80.60.10.20.10.31.7
ITDSfs18.8100*12.020.68.07.618.45.61.00.60.51.58.6
ITDSgs18.7100*12.019.99.08.218.55.71.40.50.81.98.7
MN-v3Admix0.60.499.3*0.30.60.30.40.40.20.10.20.10.3
CFM10.48.9100*10.47.49.67.714.37.22.74.03.17.8
ITDSfs16.49.6100*15.211.816.015.219.814.53.84.24.411.9
ITDSgs16.79.9100*14.911.715.915.616.814.43.24.83.811.6
RegNAdmix1.41.11.695.8*0.40.51.60.50.30.20.20.40.7
CFM21.423.314.192.3*4.412.119.69.01.41.21.15.910.3
ITDSfs38.216.025.0100*17.824.840.623.911.17.67.214.020.5
ITDSgs41.217.726.899.9*19.727.743.125.013.39.68.417.422.7
", + "image_path": "ae51110f14759bc3b8f12b0fce6a8707d1cf60214ef10e5c46ec6eac4ef8abe3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 300, + 555, + 594 + ], + "blocks": [ + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "type": "text", + "content": "Table 2. TASRs (%) on twelve pre-trained models with various competition-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks, and BAvg means average black-box TASR. ITDS" + }, + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "type": "inline_equation", + "content": "_{fs}" + }, + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "type": "text", + "content": " and ITDS" + }, + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "type": "inline_equation", + "content": "_{gs}" + }, + { + "bbox": [ + 55, + 255, + 553, + 289 + ], + "type": "text", + "content": " represent the integration of S1 and S2 in SDS, respectively, and the boldface represents their results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 300, + 555, + 594 + ], + "lines": [ + { + "bbox": [ + 59, + 300, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 59, + 300, + 555, + 594 + ], + "type": "table", + "html": "
ModelAttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
RN50DIM100*33.614.131.915.365.059.113.53.11.92.63.822.2
ODIM99.9*52.833.249.837.570.265.830.114.59.210.415.535.3
SIA100*86.856.386.540.499.093.534.330.621.724.040.355.7
BSR100*87.648.787.231.097.692.827.818.116.916.241.551.4
Admix-DIM100*51.925.553.629.183.877.228.88.04.35.110.134.3
CFM-DIM99.7*68.447.966.049.288.479.541.918.012.514.920.046.0
ITDS-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
VGG16DIM12.985.8*3.418.22.93.510.63.50.20.30.21.45.2
ODIM33.578.8*10.028.211.213.125.113.62.61.41.34.113.1
SIA50.481.7*13.555.67.320.543.77.72.11.21.69.219.3
BSR24.384.7*3.834.61.77.525.11.40.40.30.32.19.2
Admix-DIM20.288.0*6.823.95.75.416.08.00.80.50.52.18.2
CFM-DIM19.286.3*4.123.73.95.914.06.10.40.60.41.67.2
ITDS-DIM42.0100*24.834.825.719.835.324.98.52.83.27.820.9
MN-v3DIM6.32.599.2*4.96.94.55.516.26.01.82.52.35.4
ODIM24.917.399.4*21.222.122.823.236.120.09.611.611.220.0
SIA32.721.498.7*29.214.126.125.424.219.17.910.610.320.1
BSR22.012.598.9*18.38.714.413.015.911.04.46.57.812.2
Admix-DIM18.06.399.4*11.914.812.113.032.012.53.75.54.47.5
CFM-DIM34.724.999.8*31.629.433.427.844.228.111.814.211.626.5
ITDS-DIM49.527.3100*39.939.244.941.658.046.718.117.013.736.0
RegNDIM24.815.212.492.9*9.312.331.421.43.63.63.013.013.6
ODIM51.343.331.290.3*29.235.656.949.315.712.911.626.133.0
SIA71.272.138.695.7*15.648.072.033.119.717.918.647.741.3
BSR49.063.022.194.1*6.124.947.715.67.08.88.938.026.4
Admix-DIM39.425.220.692.6*14.823.147.733.17.67.17.322.922.6
CFM-DIM71.561.950.893.0*38.361.072.361.622.521.722.143.347.9
ITDS-DIM82.250.568.999.9*64.673.184.975.656.543.040.854.763.2
", + "image_path": "5448009a179afc5106ea3a0154deb085daaebbc39b9e3a8e359315a955701e30.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 603, + 555, + 638 + ], + "lines": [ + { + "bbox": [ + 55, + 603, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 555, + 638 + ], + "type": "text", + "content": "Table 3. TASRs (%) on twelve pre-trained models with various transformation-based attacks. The AEs are crafted on the RN50, VGG16, MN-v3, and RegN models, respectively. An asterisk (*) indicates white-box attacks. ITDS-DIM represents the combination of ITDS, which integrates S1, with DIM, and the boldface represents the results of this combination." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "text", + "content": "bility of the two strategies in SDS, ITDS is implemented in two forms: " + }, + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{fs}" + }, + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "text", + "content": " for S1 (feature space) and " + }, + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{gs}" + }, + { + "bbox": [ + 55, + 653, + 294, + 689 + ], + "type": "text", + "content": " for S2 (geometric space)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "It can be seen that, compared to other competition-based attacks, ITDS demonstrates an overwhelming advantage." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Taking " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{gs}" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " as an example, its average black-box TASR (BAvg.) is " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "9.0\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "7.0\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "3.8\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "inline_equation", + "content": "12.4\\%" + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": " higher than the existing SOTA method CFM on RN50, VGG16, MN-v3, and RegN, respectively. From the perspective of cross-architecture models, the largest gaps in average black-box" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3721" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 71, + 69, + 542, + 122 + ], + "blocks": [ + { + "bbox": [ + 71, + 69, + 542, + 122 + ], + "lines": [ + { + "bbox": [ + 71, + 69, + 542, + 122 + ], + "spans": [ + { + "bbox": [ + 71, + 69, + 542, + 122 + ], + "type": "table", + "html": "
AttackRN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
ITDS1100*41.524.340.418.984.670.211.55.52.92.74.227.9
ITDS1+S100*47.634.850.826.990.181.515.49.33.74.76.433.7
ITDS1-DIM99.9*71.158.673.356.892.587.148.435.118.819.123.653.1
ITDS1+S-DIM100*79.370.382.869.695.694.662.044.423.624.731.661.7
", + "image_path": "f0571c4951201106970db8f09ee8c8318777e8034fd63c4a854fe845df1627b0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 129, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 54, + 129, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 54, + 129, + 555, + 163 + ], + "type": "text", + "content": "Table 4. Evaluation of the effectiveness of two modules in ITDS, where 'I' represents ITC and 'S' represents SDS. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. Boldface represents the best results. Note that single SDS requires combination with existing attack methods, which we discuss in the appendix." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": "TASR between it and CFM on CNNs and ViTs are " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "9.7\\%" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": " on RegN, respectively. It should be noted that the white-box attack performance of our ITDS is also significantly superior, being the best in terms of white-box TASR in any surrogate model, especially on VGG16 and RegN, where it is " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "13.9\\%" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "7.6\\%" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": " higher than CFM, respectively. In the comparison between " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{fs}" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{gs}" + }, + { + "bbox": [ + 55, + 178, + 297, + 310 + ], + "type": "text", + "content": ", the results seem to be evenly matched, with each showing their own merits in the three trials. Therefore, we believe that both strategies are viable, which precisely proves the effectiveness and diversity of the proposed SDS principle." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 322, + 290, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 290, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 290, + 335 + ], + "type": "text", + "content": "4.3. Evaluation on Transformation-based Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 341, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 296, + 532 + ], + "type": "text", + "content": "Currently, transformation-based methods are the main approach to adversarial attacks and are the most effective way to improve the transferability of attacks. Numerous studies [2, 4, 16] have shown that non-transformation-based methods combined with DIM can further enhance the transferability of crafted AEs. This combination approach has become a common standard in industry, frequently compared to transformation-based methods for fairness. In this section, we evaluate the TASR of transformation-based and DIM-combined competition-based adversarial attacks. The former includes DIM, ODIM, SIA, and BSR, while the latter includes Admix-DIM, CFM-DIM, and ITDS-DIM. It should be noted that, due to the similar performance of S1 and S2 in SDS, in all subsequent experiments involving ITDS-DIM, we selected " + }, + { + "bbox": [ + 56, + 341, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathrm{ITDS}_{fs}" + }, + { + "bbox": [ + 56, + 341, + 296, + 532 + ], + "type": "text", + "content": " as the representative to combine with DIM. The results are shown in Table 3." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": "It can be observed that all competition-based methods show a significant improvement in attack performance after being combined with DIM. Compared to the strongest competition-based attack CFM-DIM, ITDS-DIM achieved average black-box TASR improvements of " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "15.7\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "13.7\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "9.5\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "15.3\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": " on RN50, VGG16, MN-v3, and RegN, respectively. The maximum average black-box TASR gap between the two on CNNs and ViTs is " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "16.1\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": " on RN50 and " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "21.4\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": " on RegN. Even when compared with the SOTA transformation-based attacks, our method is not inferior, and even surpasses them in all aspects. Compared to the strongest transformation-based attack SIA, ITDS-DIM achieved average black-box TASR improvements of " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "15.9\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "inline_equation", + "content": "21.9\\%" + }, + { + "bbox": [ + 55, + 534, + 297, + 714 + ], + "type": "text", + "content": " on MN-v3 and RegN, respectively. The maximum average black-box TASR gap between the two on" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 201, + 430, + 357 + ], + "blocks": [ + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "lines": [ + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "spans": [ + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "type": "text", + "content": "CNNs and ViTs is " + }, + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "type": "inline_equation", + "content": "21.3\\%" + }, + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "type": "inline_equation", + "content": "22.8\\%" + }, + { + "bbox": [ + 314, + 178, + 555, + 190 + ], + "type": "text", + "content": " on RegN, respectively." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 201, + 430, + 357 + ], + "lines": [ + { + "bbox": [ + 315, + 201, + 430, + 357 + ], + "spans": [ + { + "bbox": [ + 315, + 201, + 430, + 357 + ], + "type": "image", + "image_path": "d8110ee373581f5fa85dee9e4c5b891eaa13c9a4d31273436d689de330d6072c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 201, + 555, + 357 + ], + "blocks": [ + { + "bbox": [ + 440, + 201, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 440, + 201, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 440, + 201, + 555, + 357 + ], + "type": "image", + "image_path": "222e52068242b59fa606527f05786f70506e1e28f7d2617921a261c2309ad180.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 492, + 359, + 503, + 369 + ], + "lines": [ + { + "bbox": [ + 492, + 359, + 503, + 369 + ], + "spans": [ + { + "bbox": [ + 492, + 359, + 503, + 369 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 370, + 430, + 525 + ], + "blocks": [ + { + "bbox": [ + 366, + 359, + 377, + 369 + ], + "lines": [ + { + "bbox": [ + 366, + 359, + 377, + 369 + ], + "spans": [ + { + "bbox": [ + 366, + 359, + 377, + 369 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 316, + 370, + 430, + 525 + ], + "lines": [ + { + "bbox": [ + 316, + 370, + 430, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 370, + 430, + 525 + ], + "type": "image", + "image_path": "0e7e1b35e618e3529ce6f106e2176493d2789d024f353fb11df5f42fa526ae6f.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 367, + 529, + 377, + 537 + ], + "lines": [ + { + "bbox": [ + 367, + 529, + 377, + 537 + ], + "spans": [ + { + "bbox": [ + 367, + 529, + 377, + 537 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "lines": [ + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": "Figure 4. Ablation studies on the RN50 model. (a) - (d): TASRs " + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": " on the other eleven models with the adversaries crafted by ITDS-DIM, where the default value for " + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "inline_equation", + "content": "T_{sub}" + }, + { + "bbox": [ + 313, + 547, + 555, + 593 + ], + "type": "text", + "content": " are set to 3, 1.5, 10 and 10 respectively, when test parameters for each other." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 438, + 370, + 552, + 525 + ], + "blocks": [ + { + "bbox": [ + 438, + 370, + 552, + 525 + ], + "lines": [ + { + "bbox": [ + 438, + 370, + 552, + 525 + ], + "spans": [ + { + "bbox": [ + 438, + 370, + 552, + 525 + ], + "type": "image", + "image_path": "7d06bb72df22cfc1b7e323325dfdc2b5bbd9605c872e77ef16058a9ac1299619.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 529, + 500, + 537 + ], + "lines": [ + { + "bbox": [ + 489, + 529, + 500, + 537 + ], + "spans": [ + { + "bbox": [ + 489, + 529, + 500, + 537 + ], + "type": "text", + "content": "(d)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 612, + 416, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 416, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 416, + 624 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "We first present a series of ablation studies conducted on RN50 to assess the impact of each parameter in ITDS on attack performance, including the number of random target class samples " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", the competitive ratio of the inverse target gradient " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", the value of " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " in top- " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ", and the number of epochs " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "t_{sub}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": " for non-targeted attack " + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "inline_equation", + "content": "T_{sub}" + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": ". In all ablation experiments, unless otherwise specified, all parameters other" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3722" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 553, + 135 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 553, + 135 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 553, + 135 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 553, + 135 + ], + "type": "table", + "html": "
AblationTarget Model
Norm (P1)MI (P2)RN50VGG16MN-v3RegNInc-v3RN101DN161EffNViTDeiTConViTPiTBAvg.
✓7.1*0.51.00.40.40.80.30.20.20.20.10.10.4
✓100*72.866.175.865.190.188.254.737.216.917.423.755.3
✓✓100*59.741.366.152.890.991.345.127.814.516.221.147.9
100*79.370.382.869.695.694.662.044.423.624.731.661.7
", + "image_path": "624889832acda06da70912c29c414fc4d5850ceea0d67c6b6c4a3aa42cfd56ef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 60, + 180, + 293, + 232 + ], + "blocks": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 555, + 166 + ], + "type": "text", + "content": "Table 5. TASRs (%) of ITDS-DIM by ablating inner modules of the ITC. The AEs are crafted in RN50. An asterisk (*) indicates white-box attacks. In which Norm and MI correspond to P1 and P2 in Section 3.1.2, respectively. Boldface represents the best results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 180, + 293, + 232 + ], + "lines": [ + { + "bbox": [ + 60, + 180, + 293, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 180, + 293, + 232 + ], + "type": "table", + "html": "
AttackJPEGFDSOGDMPDef4Def8Defl2
SIA83.289.844.31.539.014.054.7
Admix-DIM80.092.947.71.934.814.553.1
CFM-DIM82.592.651.54.125.28.745.5
ITDS-DIM99.499.991.331.250.225.666.8
", + "image_path": "af8a74639bfa0ec6a84b2862576b8c3de3eec5592eea996d0810466ca00aa4c8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 239, + 295, + 305 + ], + "lines": [ + { + "bbox": [ + 55, + 239, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 295, + 305 + ], + "type": "text", + "content": "Table 6. TASRs (\\%) of breaking-through four defense methods and three adversarial training defense models. The results for defense methods are based on the average performance of AEs crafted on RN50, VGG16, MN-v3, and RegN respectively, while the results for defense models are based on the performance of AEs crafted by the corresponding models. Bolder for the best." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 325, + 295, + 349 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 295, + 349 + ], + "type": "text", + "content": "than the test parameters are fixed according to the experimental setup." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "spans": [ + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": "As shown in Figure 4a, TASR increases with " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ", as the diversity of target class images directly affects attack performance. However, the curve levels off after " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ", and considering computational costs, we selected " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "n = 3" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ". In Figure 4b, most models achieve optimal performance at " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "\\beta = 1.5" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " with a few exceptions (e.g., MN-v3). We believe a moderate competitive gradient provides an \"exercise\" effect, while an overly strong gradient disrupts input features, reducing attack performance. In Figure 4c, optimal performance generally occurs at " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ". We posit that an overly small " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " under-covers classes and fails to enlarge global margins, whereas an overly large one blurs the optimization. At " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "k = 10" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ", cross-architecture transferability is even better. In Figure 4d, evaluating epochsub " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "(T_{sub})" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " for non-targeted attacks, we found that larger values do not always lead to better performance. As discussed in the ITC section, inverse target gradient perturbations generated at each " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "T_{sub}" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " have diverse potential competitive features, boosting attack performance. At " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "T_{sub} = 1" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": ", though it involves more target class samples, it lacks deeper latent features of the target class, limiting attack performance. Beyond a certain " + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "inline_equation", + "content": "T_{sub}" + }, + { + "bbox": [ + 56, + 352, + 295, + 626 + ], + "type": "text", + "content": " value, missing target class features reduce the diversity of inverse gradients, causing performance to regress." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 715 + ], + "type": "text", + "content": "Furthermore, to verify the effectiveness of the ITC and SDS components in ITDS, we conducted experimental evaluations as shown in Table 4. The results indicate that ITDS using only ITC has already surpassed the attack performance of CFM in Table 2. Regardless of whether DIM is combined or not, SDS can bring additional performance improvements to ITDS. These findings collectively support the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 182, + 457, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 457, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 457, + 194 + ], + "type": "text", + "content": "key roles of ITC and SDS in ITDS." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 194, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 555, + 399 + ], + "type": "text", + "content": "To verify the impact of internal modules within the ITC on attack performance, we conducted ablation experiments as shown in Table 5. It can be observed that the use of N-norm may limit the scale of gradient updates, which could result in generated AEs lacking sufficient aggressiveness. On the contrary, if we do not use N-norm (P1), gradient updates will gain greater freedom, which helps to explore the model's decision boundary more deeply, and may thus produce more aggressive AEs. Simultaneously, momentum gradients, as a technique for accumulating previous gradients, will weaken the intensity of inverse target features in the current iteration. Since we wish to maintain the integrity of inverse target features and avoid any normalization, the inverse target momentum gradients in each round will compete with the AE when combined with P1, leading to excessive deformation of AE features, which is disastrous for implementing precisely targeted attacks." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 403, + 495, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 495, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 495, + 415 + ], + "type": "text", + "content": "4.5. Evaluation on Advanced Defenses" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 421, + 554, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 517 + ], + "type": "text", + "content": "In this section, we evaluate the capabilities of various attacks to break through different defense mechanisms. As shown in Table 6, our method exhibits exceptional robustness against a wide range of defense strategies and models. For example, when the Score-OPT method is employed, the average TASR of our approach is nearly " + }, + { + "bbox": [ + 313, + 421, + 554, + 517 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 313, + 421, + 554, + 517 + ], + "type": "text", + "content": " higher than that of CFM-DIM. Furthermore, our method achieves the best performance in all three defense models evaluated." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "This paper introduces a novel method named ITDS, which significantly improves the transferability of targeted AEs by incorporating the ITC mechanism and the SDS strategy. Experimental results demonstrate that ITDS achieves a substantial increase in average transferable TASRs across various mainstream CNN and ViT models, outperforming SOTA competition-based and transformation-based methods. Our ablation studies further confirmed the key roles of the ITC and SDS components in enhancing attack performance. Moreover, ITDS exhibits superior effectiveness in countering multiple defense methods and models. These findings confirm the effectiveness and practicality of ITDS in enhancing the transferability of targeted AEs and provide a new perspective for the field of adversarial research." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3723" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 297, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 297, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 297, + 248 + ], + "type": "text", + "content": "This work was supported by funding from the National Natural Science Foundation of China (NSFC) under Grants 62272486, 62172154, and 62372473, the Natural Science Foundation of Hunan Province under Grant 2023JJ70016, the Hong Kong Research Grants Council (RGC) General Research Fund under Grants 152244/21E, 152169/22E, 152228/23E, and 162161/24E, the Research Impact Fund under Grant R5011-23F, the Collaborative Research Fund under Grant C1042-23GF, the NSFC/RGC Collaborative Research Scheme under Grant CRS_HKUST602/24, the Areas of Excellence Scheme under Grant AoE/E-601/22-R, and the InnoHK initiative (HKGAI). Professor Shigeng Zhang is the corresponding author of this paper." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 262, + 115, + 274 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 281, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 61, + 281, + 295, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 281, + 295, + 347 + ], + "spans": [ + { + "bbox": [ + 61, + 281, + 295, + 347 + ], + "type": "text", + "content": "[1] Junyoung Byun, Seungju Cho, Myung-Joon Kwon, Hee-Seon Kim, and Changick Kim. Improving the transferability of targeted adversarial examples through object-based diverse input. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15244-15253, 2022. 1, 2, 3, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 350, + 295, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 350, + 295, + 415 + ], + "spans": [ + { + "bbox": [ + 61, + 350, + 295, + 415 + ], + "type": "text", + "content": "[2] Junyoung Byun, Myung-Joon Kwon, Seungju Cho, Yoonji Kim, and Changick Kim. Introducing competition to boost the transferability of targeted adversarial examples through clean feature mixup. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24648-24657, 2023. 1, 3, 5, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 418, + 295, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 418, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 61, + 418, + 295, + 472 + ], + "type": "text", + "content": "[3] Yinpeng Dong, Fangzhou Liao, Tianyu Pang, Hang Su, Jun Zhu, Xiaolin Hu, and Jianguo Li. Boosting adversarial attacks with momentum. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 9185-9193, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 62, + 475, + 295, + 529 + ], + "type": "text", + "content": "[4] Yinpeng Dong, Tianyu Pang, Hang Su, and Jun Zhu. Evading defenses to transferable adversarial examples by translation-invariant attacks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4312-4321, 2019. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 532, + 295, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 532, + 295, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 532, + 295, + 597 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 599, + 295, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 599, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 62, + 599, + 295, + 654 + ], + "type": "text", + "content": "[6] Stéphane d'Ascoli, Hugo Touvron, Matthew L Leavitt, Ari S Morcos, Giulio Biroli, and Levent Sagun. Convit: Improving vision transformers with soft convolutional inductive biases. In International Conference on Machine Learning, pages 2286-2296. PMLR, 2021. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 62, + 656, + 295, + 678 + ], + "type": "text", + "content": "[7] Goodfellow et al. Explaining and harnessing adversarial examples. arXiv preprint arXiv:1412.6572, 2014. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "text", + "content": "[8] Chuan Guo, Mayank Rana, Moustapha Cisse, and Laurens Van Der Maaten. Countering adversarial images using input transformations. arXiv preprint arXiv:1711.00117, 2018. 5" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 555, + 117 + ], + "type": "text", + "content": "[9] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 555, + 173 + ], + "type": "text", + "content": "[10] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11936-11945, 2021. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 555, + 229 + ], + "type": "text", + "content": "[11] Andrew Howard, Mark Sandler, Grace Chu, Liang-Chieh Chen, Bo Chen, Mingxing Tan, Weijun Wang, Yukun Zhu, Ruoming Pang, Vijay Vasudevan, et al. Searching for mobilenetv3. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1314-1324, 2019. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 231, + 554, + 274 + ], + "type": "text", + "content": "[12] Gao Huang, Zhuang Liu, Laurens Van Der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4700-4708, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[13] Kurakin et al. Adversarial examples in the physical world. In Artificial intelligence safety and security, pages 99-112. Chapman and Hall/CRC, 2018. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 310, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 363 + ], + "type": "text", + "content": "[14] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Foolmix: Strengthen the transferability of adversarial examples by dual-blending and direction update strategy. IEEE Transactions on Information Forensics and Security, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 409 + ], + "type": "text", + "content": "[15] Zhankai Li, Weiping Wang, Jie Li, Kai Chen, and Shigeng Zhang. Ucg: A universal cross-domain generator for transferable adversarial examples. IEEE Transactions on Information Forensics and Security, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 454 + ], + "type": "text", + "content": "[16] Jiadong Lin, Chuanbiao Song, Kun He, Liwei Wang, and John E Hopcroft. Nesterov accelerated gradient and scale invariance for adversarial attacks. In International Conference on Learning Representations, 2020. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[17] Zihao Liu, Qi Liu, Tao Liu, Nuo Xu, Xue Lin, Yanzhi Wang, and Wujie Wen. Feature distillation: Dnn-oriented jpeg compression against adversarial examples. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 860-868. IEEE, 2019. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[18] Yuyang Long, Qilong Zhang, Boheng Zeng, Lianli Gao, Xianglong Liu, Jian Zhang, and Jingkuan Song. Frequency domain model augmentation for adversarial attack. In European Conference on Computer Vision, pages 549-566. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 621 + ], + "type": "text", + "content": "[19] Seyed-Mohsen Moosavi-Dezfooli, Alhussein Fawzi, and Pascal Frossard. Deepfool: a simple and accurate method to fool deep neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2574–2582, 2016. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 677 + ], + "type": "text", + "content": "[20] Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, and Piotr Dólar. Designing network design spaces. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10428-10436, 2020. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[21] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3724" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 296, + 95 + ], + "type": "text", + "content": "scale visual recognition challenge. International journal of computer vision, 115:211-252, 2015. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 296, + 129 + ], + "type": "text", + "content": "[22] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 132, + 295, + 154 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 132, + 295, + 154 + ], + "spans": [ + { + "bbox": [ + 56, + 132, + 295, + 154 + ], + "type": "text", + "content": "[23] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 156, + 295, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 156, + 295, + 209 + ], + "spans": [ + { + "bbox": [ + 56, + 156, + 295, + 209 + ], + "type": "text", + "content": "[24] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2818-2826, 2016. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 213, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 213, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 213, + 295, + 255 + ], + "type": "text", + "content": "[25] Mingxing Tan and Quoc Le. Efficientnet: Rethinking model scaling for convolutional neural networks. In International conference on machine learning, pages 6105-6114. PMLR, 2019. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 258, + 295, + 313 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 258, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 258, + 295, + 313 + ], + "type": "text", + "content": "[26] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 358 + ], + "type": "text", + "content": "[27] Francisco Utrera, Evan Kravitz, N Benjamin Erickson, Rajiv Khanna, and Michael W Mahoney. Adversarily-trained deep nets transfer better: Illustration on image classification. arXiv preprint arXiv:2007.05869, 2021. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 393 + ], + "type": "text", + "content": "[28] Jinyi Wang, Zhaoyang Lyu, Dahua Lin, Bo Dai, and Hongfei Fu. Guided diffusion model for adversarial purification. arXiv preprint arXiv:2205.14969, 2022. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 396, + 295, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 295, + 450 + ], + "type": "text", + "content": "[29] Kunyu Wang, Xuanran He, Wenxuan Wang, and Xiaosen Wang. Boosting adversarial transferability by block shuffle and rotation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24336-24346, 2024. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 453, + 295, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 295, + 496 + ], + "type": "text", + "content": "[30] Xiaosen Wang, Xuanran He, Jingdong Wang, and Kun He. Admix: Enhancing the transferability of adversarial attacks. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16158-16167, 2021. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 499, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 499, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 56, + 499, + 295, + 543 + ], + "type": "text", + "content": "[31] Xiaosen Wang, Zeliang Zhang, and Jianping Zhang. Structure invariant transformation for better adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4607-4619, 2023. 2, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 295, + 598 + ], + "type": "text", + "content": "[32] Zhibo Wang, Hengchang Guo, Zhifei Zhang, Wenxin Liu, Zhan Qin, and Kui Ren. Feature importance-aware transferable adversarial attacks. In Proceedings of the IEEE/CVF international conference on computer vision, pages 7639-7648, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 655 + ], + "type": "text", + "content": "[33] Juanjuan Weng, Zhiming Luo, Shaozi Li, Nicu Sebe, and Zhun Zhong. Logit margin matters: Improving transferable targeted adversarial attack by logit calibration. IEEE Transactions on Information Forensics and Security, 18:3561-3574, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[34] Cihang Xie, Zhishuai Zhang, Yuyin Zhou, Song Bai, Jianyu Wang, Zhou Ren, and Alan L Yuille. Improving transferability of adversarial examples with input diversity. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2730-2739, 2019. 1, 2, 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 352 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[35] Zeng et al. Enhancing targeted transferability via feature space fine-tuning. In ICASSP 2024-2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 4475-4479. IEEE, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[36] Boya Zhang, Weijian Luo, and Zhihua Zhang. Enhancing adversarial robustness via score-based optimization. Advances in Neural Information Processing Systems, 36:51810-51829, 2023. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 163, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 216 + ], + "type": "text", + "content": "[37] Qilong Zhang, Xiaodan Li, YueFeng Chen, Jingkuan Song, Lianli Gao, Yuan He, et al. Beyond imagenet attack: Towards crafting adversarial examples for black-box domains. In International Conference on Learning Representations, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 219, + 553, + 261 + ], + "type": "text", + "content": "[38] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quaxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, pages 1672-1678, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 296 + ], + "type": "text", + "content": "[39] Zhao et al. On success and simplicity: A second look at transferable targeted attacks. Advances in Neural Information Processing Systems, 34:6115-6128, 2021. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 352 + ], + "type": "text", + "content": "[40] Rongyi Zhu, Zeliang Zhang, Susan Liang, Zhuo Liu, and Chenliang Xu. Learning to transform dynamically for better adversarial transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24273-24283, 2024. 1" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3725" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_content_list.json b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..d28e25e9d01e4cbac7be13b77abc5b37ec0ce06c --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_content_list.json @@ -0,0 +1,1619 @@ +[ + { + "type": "text", + "text": "Enhancing Transformers Through Conditioned Embedded Tokens", + "text_level": 1, + "bbox": [ + 163, + 130, + 833, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hemanth Saratchandran \nAustralian Institute for Machine Learning \nAdelaide University", + "bbox": [ + 145, + 181, + 478, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hemanth.saratchandran@adelaide.edu.au", + "bbox": [ + 145, + 237, + 477, + 250 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Simon Lucey \nAustralian Institute for Machine Learning \nAdelaide University", + "bbox": [ + 519, + 181, + 849, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "simon.lucey@adelaide.edu.au", + "bbox": [ + 563, + 237, + 803, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/60619f460852316481314c0f675e2b59c233c40ff87cbb2b669a8ea26eb4c00e.jpg", + "image_caption": [ + "Figure 1. Schematic representation of conditioned embeddings for a vision transformer: An image is divided into $N$ patches, with each patch then embedded as a high-dimensional vector $X_{i}\\in \\mathbb{R}^{d}$ for $1\\leq i\\leq N$ . These vectors are then concatenated to form the embedded token matrix $X = [X_{1}\\dots X_{N}]^{T}$ . To improve the conditioning of the self-attention mechanism, a correction term $C = [C_1\\dots C_N]^T$ is added to $X$ , reducing its condition number. The modified matrix is then fed into the first layer of the transformer (positional encoding not shown). Its effect on the self-attention equation is illustrated in the green equation (for simplicity, layer normalization has been omitted)." + ], + "image_footnote": [], + "bbox": [ + 130, + 286, + 851, + 640 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 733, + 326, + 750 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we de", + "bbox": [ + 88, + 779, + 485, + 900 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "velop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image", + "bbox": [ + 511, + 734, + 908, + 887 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4786", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 165, + 222, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The transformer architecture [23] has become a cornerstone of modern machine learning, driving breakthroughs across various domains including computer vision [3, 5, 15, 22], natural language processing [23, 27, 28], and robotics [16, 18]. Its success is largely attributed to its ability to model complex relationships in data without relying on traditional recurrent or convolutional structures. At the heart of the transformer is the self-attention mechanism, which dynamically assigns importance to different parts of the input, capturing both local and global dependencies. Unlike traditional architectures that process inputs in a fixed order, transformers utilize self-attention to integrate global context at each layer, making them highly effective for tasks that require contextual understanding.", + "bbox": [ + 89, + 190, + 483, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In general, transformers process input data by first converting it into discrete tokens, which are then mapped to a high-dimensional vector space through an embedding layer. For text, these tokens represent words or sub-words, while for images, image patches are treated as tokens. The resulting embedded tokens form a matrix that is passed into the first transformer layer for self-attention computations. This process allows the model to effectively learn relationships between tokens in a high-dimensional space.", + "bbox": [ + 89, + 402, + 483, + 537 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore the conditioning of the self-attention matrix in transformer architectures—an essential yet often overlooked factor influencing optimization dynamics. The conditioning of a matrix is measured by its condition number, defined as the ratio of its largest singular value to its smallest. A high condition number indicates ill-conditioning, which is a well-known challenge for the convergence of gradient-based optimization methods [17].", + "bbox": [ + 89, + 537, + 483, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous work has addressed conditioning in feedforward neural networks. In [20], a methodology was introduced to condition network weights, leading to improved optimization. Similarly, [14] examined the condition of the neural tangent kernel (NTK), proposing a framework for reducing its condition number and demonstrating enhanced convergence for gradient descent. While these studies have focused on conditioning in feedforward architectures, its role in self-attention mechanisms remains largely unexplored. This gap motivates our investigation into how conditioning affects optimization in transformers.", + "bbox": [ + 89, + 659, + 483, + 824 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We develop a theoretical framework to analyze the condition number of the self-attention matrix within a transformer's attention layer. Our analysis establishes a direct relationship between the condition number of the first-layer self-attention matrix and that of the embedded tokens.", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Leveraging this insight, we introduce conditioned embedded tokens, a method that applies a structured correction term to the embedded token matrix, effectively reducing its condition number and, in turn, improving the conditioning of the self-attention matrix. While a full theoretical proof linking this improvement to better convergence in gradient-based optimization would require analyzing the NTK of a transformer—a notoriously difficult task—we provide strong empirical evidence that conditioned embedded tokens enhances performance across a range of transformer applications.", + "bbox": [ + 511, + 90, + 903, + 256 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Figure 1 illustrates how conditioned embedded tokens are integrated into a vision transformer. An image is first divided into tokens via patches, which are vectorized and passed through a learnable embedding layer, producing embedded tokens in a high-dimensional vector space. These tokens form the matrix $X = [X_{1}\\dots X_{N}]^{T}$ . A correction term $C = [C_1\\cdot \\cdot \\cdot C_N]^T$ is added to $X$ , reducing its condition number before the modified embedding matrix $X + C$ is fed into the first transformer layer (positional encoding not shown). This correction not only lowers the condition number of the self-attention matrix in the first layer but also has a cascading effect, improving conditioning in subsequent layers. In general, positional encodings are added to the tokens before they are fed into the first transformer layer; however, they are omitted in Figure 1 for simplicity. Our theory will show that the optimal correction term can be derived from the singular value decomposition (SVD) of $X$ .", + "bbox": [ + 511, + 257, + 903, + 513 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our framework across various applications, including image classification, object detection, instance segmentation, and natural language processing. A key advantage of conditioned embedded tokens is their flexibility—they can be used with a wide range of attention mechanisms, making them compatible with modern transformer architectures. As we demonstrate, our methodology can serve as a drop-in enhancement for advanced attention mechanisms proposed in recent works [2, 4, 15, 22, 25, 26], offering a simple yet effective improvement to existing models. Our main contributions include:", + "bbox": [ + 511, + 513, + 905, + 679 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A theoretical framework introducing conditioned embedded tokens for improving the conditioning of the self-attention matrix within the first transformer layer yielding a more stable transformer architecture.", + "2. An empirical evaluation of conditioned embedded tokens across various transformer architectures and applications, including image classification, object detection, instance segmentation, and natural language processing, showing in each case its superior performance." + ], + "bbox": [ + 506, + 680, + 903, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 829, + 653, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Attention: Various strategies have been introduced to enhance the efficiency and effectiveness of transformers, particularly by addressing their computational overhead and re", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4787", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "thinking attention mechanisms. The Data-Efficient Image Transformer (DeiT) [22] improves training efficiency in vision transformers by leveraging distillation tokens, allowing for competitive performance without requiring massive datasets. The Cross-Covariance Image Transformer (XCiT) [2] redefines attention by operating on cross-covariances of spatial features, enabling more efficient spatial interactions while reducing computational demands. The Nyströmformer [25], approximates standard self-attention using the Nyström method, reducing quadratic complexity to near-linear. This technique has proven particularly effective for long-range sequence modeling. In this paper, we demonstrate that our conditioning framework can be integrated as a drop-in replacement into these advanced attention mechanisms, yielding superior performance across multiple architectures.", + "bbox": [ + 89, + 90, + 483, + 330 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Conditioning: In [20], the authors investigate the condition number of weight matrices in feedforward neural networks and demonstrate that lower condition numbers lead to improved accuracy across various applications. They propose a preconditioner matrix that multiplies each layer's weight matrix, effectively reducing its condition number and enhancing the training of dense weights. In [14], neural network optimization is explored through the lens of the neural tangent kernel. The authors introduce the PL* condition, a variant of the Polyak-Lojasiewicz condition, and show that this approach improves the conditioning of the neural tangent kernel, which, in the infinite-width limit, governs the dynamics of gradient descent [10]. Finally, [1] demonstrates that increasing the depth of feedforward neural networks improves their conditioning leading to more efficient optimization. See [11, 12, 19, 20] for further results.", + "bbox": [ + 89, + 332, + 483, + 588 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Notation", + "text_level": 1, + "bbox": [ + 89, + 602, + 191, + 618 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we formally define the transformer architecture by describing the structure of a transformer layer, along with establishing notation for various mathematical elements that will be referenced in subsequent sections. For further foundational details on transformers, readers may consult Vaswani et al. [23].", + "bbox": [ + 89, + 628, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A layer in a transformer can be represented as a mapping", + "bbox": [ + 109, + 719, + 482, + 734 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T}: \\mathbb {R} ^ {N \\times D} \\rightarrow \\mathbb {R} ^ {N \\times D} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 743, + 482, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "defined by", + "bbox": [ + 89, + 772, + 165, + 787 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} (X) = \\mathbf {F} (\\mathbf {A} (X) + X), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 787, + 482, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{F}$ denotes a feed forward network with a residual connection, and $\\mathbf{A}$ represents an attention mechanism. In general, layer normalization is added however for simplicity we omit layer normalization for this discussion.", + "bbox": [ + 89, + 809, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To begin with input data $x$ is tokenized into $N$ tokens yielding a tokenized representation $(x_{1},\\dots ,x_{N})$ where", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$x_{i}\\in \\mathbb{R}^{t\\times 1}$ . Each $x_{i}$ is then mapped to a high dimensional space $\\mathbb{R}^{d\\times 1}$ using a learnable embedding layer $E\\in \\mathbb{R}^{d\\times t}$ so that", + "bbox": [ + 511, + 90, + 903, + 133 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE x _ {i} \\in \\mathbb {R} ^ {d \\times 1}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 663, + 135, + 903, + 152 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The matrix $[Ex_1 \\cdots Ex_n]^T \\in \\mathbb{R}^{N \\times d}$ represents the embedded tokens associated to the input data. To capture the order of the tokens, a positional encoding matrix $P \\in \\mathbb{R}^{N \\times d}$ is added forming $X = [Ex_1 \\cdots Ex_n]^T + P \\in \\mathbb{R}^{N \\times d}$ . The matrix $X$ is then input into the first layer of the transformer. For the theoretical Sec. 4 we will often use the notation $X$ to denote embedded tokens often leaving out explicit mention of the positional encoding part for ease of notation.", + "bbox": [ + 511, + 157, + 906, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The key component of the transformer is self-attention. This is composed of three learnable matrices, a query $(W_{Q})$ , key $(W_{K})$ , and value $(W_{V})$ , where $W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times d}$ and $W_{V} \\in \\mathbb{R}^{D \\times d}$ . The output of the attention head $\\mathbf{A}(X)$ is then given by", + "bbox": [ + 511, + 279, + 905, + 354 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} (X) = \\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right) X W _ {V}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 362, + 903, + 381 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where softmax is the softmax activation that acts row-wise on the matrix $XW_{Q}W_{k}^{T}X^{T}$ . Note that then $A(X) \\in \\mathbb{R}^{N \\times d}$ . In the case no activation is used we have the definition of linear self-attention which we will denote by", + "bbox": [ + 511, + 390, + 905, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {L A} (X) := X W _ {Q} W _ {k} ^ {T} X ^ {T} X W _ {V}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 459, + 903, + 479 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In general, multiple attention heads $\\mathbf{A}_i$ for $1\\leq i\\leq h$ are utilized, each of dimension $N\\times d_{i} = N\\times \\frac{d}{h}$ . These are then concatenated to produce a multi-head attention output,", + "bbox": [ + 511, + 487, + 905, + 532 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n[ \\mathbf {A} _ {1}, \\dots , \\mathbf {A} _ {h} ], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 542, + 903, + 560 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "which is subsequently fed into a feedforward layer. Similarly, a multi-head linear attention can be defined in the same way. The full transformer architecture is obtained by sequentially stacking several such transformer layers, as defined in Eq. (2).", + "bbox": [ + 511, + 569, + 905, + 645 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Theoretical Framework", + "text_level": 1, + "bbox": [ + 511, + 657, + 736, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Conditioning of self-attention", + "text_level": 1, + "bbox": [ + 511, + 681, + 774, + 698 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we analyze the conditioning of the self-attention matrix in a transformer. Our objective is to demonstrate that the condition number of the self-attention matrix in the first layer is influenced by the condition number of the embedded tokens. This insight motivates our key finding: reducing the condition number of the embedded tokens leads to a better-conditioned self-attention matrix in the first layer, ultimately improving the overall conditioning of the transformer. In this section, we omit positional encoding and layer normalization to keep the core theoretical insight clear and accessible. In Sec. 5, we will show our insights go through for a variety of practical transformers that use positional encoding and layer normalization.", + "bbox": [ + 511, + 703, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4788", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Definition 4.1. Let $A$ be an $N \\times d$ matrix of full rank. The condition number of $A$ , denoted by $\\kappa$ , is defined as", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (A) = \\frac {\\sigma_ {1}}{\\sigma_ {k}} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 126, + 482, + 156 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $k = \\min \\{N, d\\}$ , and $\\sigma_1$ denotes the largest singular value of $A$ and $\\sigma_k$ the smallest singular value of $A$ . Note that as $A$ is assumed full rank $\\sigma_k \\neq 0$ and the condition number is well-defined. Furthermore, observe that $\\kappa(A) \\geq 1$ since $\\sigma_1 \\geq \\sigma_k$ .", + "bbox": [ + 89, + 161, + 483, + 237 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We have the following estimate for the condition number of linear self-attention and softmax self-attention.", + "bbox": [ + 89, + 243, + 483, + 273 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Proposition 4.2. Let $X$ denote an input for an attention layer as defined in Sec. 3. Assume $\\mathbf{L}\\mathbf{A}(X)$ and $\\mathbf{A}(X)$ have full rank. We then have", + "bbox": [ + 88, + 279, + 483, + 325 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\kappa (\\mathbf {L A} (X)) \\leq \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3} (8) \\\\ \\kappa (\\mathbf {A} (X)) \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\kappa (X) \\kappa \\left(W _ {V}\\right). (9) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 330, + 482, + 383 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Proposition 4.2 is given in Sec. A of the Supp. material.", + "bbox": [ + 89, + 391, + 483, + 421 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 4.3. The statement of Proposition 4.2 assumes the attention matrices are full rank. Since the layers of a transformer are randomly initialized we have that with probability 1 the attention matrices will be full rank at initialization.", + "bbox": [ + 89, + 424, + 483, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We point out that although Eq. (8) and Eq. (9) only provide an upper bound for the condition number, we found that the this upper bound provides a useful measure for estimating the condition number of the self-attention matrix. Therefore, for the self-attention matrices $\\mathbf{A}(X)$ and $\\mathbf{L}\\mathbf{A}(X)$ we make the following definition", + "bbox": [ + 89, + 488, + 483, + 579 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu (\\mathbf {L A} (X)) := \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3}. (10) \\\\ \\mu (\\mathbf {A} (X)) := \\kappa \\left(\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right)\\right) \\kappa (X) \\kappa \\left(W _ {V}\\right). (11) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 585, + 482, + 637 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Conditioned embedded tokens", + "text_level": 1, + "bbox": [ + 89, + 643, + 359, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we present our main theorem, which provides a method for reducing the condition number of embedded tokens derived from a dataset. This reduction subsequently decreases the measure $\\mu (\\mathbf{A}(X))$ , a key metric for assessing the condition number of the self-attention matrix in the first transformer layer.", + "bbox": [ + 89, + 666, + 483, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 4.4. Let $(x_{1},\\dots ,x_{N})$ denote $N$ tokens associated to an input dataset with $x_{i}\\in \\mathbb{R}^{t\\times 1}$ . Let $E\\in \\mathbb{R}^{d\\times t}$ denote an embedding matrix so that $Ex_{i}\\in \\mathbb{R}^{d\\times 1}$ . Let $X = [Ex_1\\dots Ex_N]^T\\in \\mathbb{R}^{N\\times d}$ denote the matrix of embedded tokens. Assume that $\\kappa (X) > 2$ . Then there exists a matrix $C\\in \\mathbb{R}^{N\\times d}$ such that", + "bbox": [ + 89, + 763, + 483, + 853 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\kappa (X + C) \\leq 2 \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 862, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and hence that $\\kappa(X + C) < \\kappa(X)$ .", + "bbox": [ + 89, + 885, + 326, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In particular, Theorem 4.4 shows that if the condition number of the embedded tokens $X$ is extremely large. Adding a correction term given by $C$ will bring it down significantly. As we shall see in the experiments Sec. 5 the condition number $\\kappa(X)$ is empirically extremely large. The proof of Theorem 4.4 is given in Sec. A of the Supp. material.", + "bbox": [ + 511, + 90, + 905, + 196 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 4.5. Let $X$ denote a collection of embedded tokens, as defined in the statement of Theorem 4.4, and let $C$ denote the matrix given by Theorem 4.4. We then have", + "bbox": [ + 511, + 207, + 906, + 252 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu (\\mathbf {L A} (X + C)) \\leq \\mu (\\mathbf {L A} (X)). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 263, + 906, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assume that", + "bbox": [ + 511, + 292, + 599, + 306 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\kappa (\\text {s o f t m a x} ((X + C) W _ {Q} W _ {K} ^ {T} (X ^ {T} + C ^ {T}))) \\\\ \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\tag {14} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 318, + 905, + 357 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "then", + "bbox": [ + 513, + 369, + 545, + 382 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mu (\\mathbf {A} (X + C)) \\leq \\mu (\\mathbf {A} (X)) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 383, + 903, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu (\\mathbf{L}\\mathbf{A}(X))$ is defined by Eq. (10) and $\\mu (\\mathbf{A}(X))$ by Eq. (11).", + "bbox": [ + 511, + 406, + 905, + 436 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 4.5 is given in Sec. A of the Supp. material. A natural question arises: can self-attention be conditioned directly through queries, keys, and values, independent of the embedded tokens? Unfortunately, our approach of introducing a correction term does not extend to these components. This is because our correction modifies the singular values of the embedded tokens, whereas altering the singular values of queries, keys, or values could disrupt how self-attention captures spatial relationships between tokens.", + "bbox": [ + 511, + 446, + 905, + 597 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 4.6. We note that in the case of self-attention with a softmax activation Theorem 4.5 requires an assumption on the condition number of the non-linear self-attention probability matrix term softmax $((X + C)W_{Q}W_{K}^{T}(X^{T} + C^{T}))$ . Although this may be seen as a limitation of the theory we will see in Sec. 5 that this framework still provides a useful methodology to obtain good performance on a variety of transformer applications.", + "bbox": [ + 511, + 603, + 905, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 4.7. Theorem 4.5 demonstrates that conditioning embedded tokens can improve the condition number of the self-attention matrix in the first layer. However, it does not address how this affects the condition number in subsequent layers. In Sec. 5, we will show that this effect propagates through the network, leading to a reduced condition number in deeper self-attention layers.", + "bbox": [ + 511, + 729, + 905, + 835 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Remark 4.8. Our theory, as presented in Theorem 4.4 and Theorem 4.5, demonstrates that conditioning embedded tokens reduces the condition number of the self-attention matrix in the first layer. However, the framework does not", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4789", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/e5d71b57ac5437af76b6431fa5f87af0415b8638e9fffdb19c312c55f01e876b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Models
ViT-baseDeiT-baseSwin-baseXciT-mediumDaViT-base
Original embedded tokens80.381.683.182.283.6
Conditioned embedded tokens81.382.583.982.984.6
", + "bbox": [ + 158, + 90, + 839, + 172 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of vision transformers with their original embedded tokens verse one with a conditioned conditioned embedded tokens pre-trained on the ImageNet-1k dataset. We report the classification top-1% accuracy. In each case we see the transformers employing conditioned embedded tokens (ours) outperforms the original ones.", + "bbox": [ + 88, + 186, + 906, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "theoretically establish how this improvement translates to better optimization and, consequently, higher accuracy in applications. Nonetheless, in Sec. 5, we provide empirical evidence showing that our approach consistently enhances accuracy across various applications, highlighting its practical effectiveness.", + "bbox": [ + 88, + 255, + 480, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 359, + 225, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we evaluate our insight from Sec. 4 on a variety of transformer applications. For a detailed analysis of how we implemented conditioned embedded tokens see Sec. A in the Supp. material.", + "bbox": [ + 89, + 385, + 483, + 446 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Image Classification", + "text_level": 1, + "bbox": [ + 89, + 455, + 282, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we apply our theoretical insights from Sec. 4 to vision transformers on an image classification task. We will train all vision transformers from scratch on the ImageNet-1k dataset. We will be testing on the following vision transformers. In each case we will compare the original architecture with one that uses conditioned embedded tokens.", + "bbox": [ + 89, + 477, + 483, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ViT [5]: ViT is a pioneering architecture that applies transformer-based processing to images by treating them as sequences of non-overlapping patches. In our study, we utilize ViT-Base (ViT-B), which is configured with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. The model employs a standard self-attention mechanism to capture dependencies across patches.", + "bbox": [ + 89, + 583, + 483, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DeiT [22]: DeiT builds upon ViT but is optimized for data efficiency. Unlike standard ViT, it employs a data-starvation training strategy [22], enabling faster convergence while maintaining strong performance. In our setup, we use DeiT-Base (DeiT-B) with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. Like ViT, DeiT utilizes a self-attention mechanism for feature extraction.", + "bbox": [ + 89, + 688, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Swin Transformer [15]: Swin Transformer introduces a hierarchical architecture and a novel shifted window-based self-attention mechanism, significantly improving efficiency and effectiveness in vision tasks. In our experiments, we adopt the Swin-Base (Swin-B) variant, which features 128 channels in the hidden layers of the first stage.", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4ada79ab080496b73870e69728c9d70f45572144b0a9e3e91857f5ec62f1cbeb.jpg", + "image_caption": [ + "Figure 2. Condition number comparison for ViT-B on ImageNet-1k. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over training epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + ], + "image_footnote": [], + "bbox": [ + 517, + 253, + 890, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The attention operates over non-overlapping windows of size $M = 7$ , while the query dimension per head is set to $d = 32$ . The network follows a hierarchical structure with layers arranged as $\\{2, 2, 18, 2\\}$ across different stages.", + "bbox": [ + 511, + 575, + 906, + 638 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "XCiT [2]: XCiT diverges from standard ViT architectures by incorporating two key innovations. First, it integrates Local Patch Interaction within each block, using depth-wise $3 \\times 3$ convolutions, Batch Normalization, GELU activation, and an additional depth-wise convolution to enhance local feature representation. Second, it employs Cross-Covariance Attention, where attention maps are derived from the cross-covariance matrix of the key and query projections. For our experiments, we utilize the XCiT-M24 (XCiT-M) model with a patch size of 16, an embedding dimension of 512, 8 attention heads, and 24 layers.", + "bbox": [ + 511, + 640, + 908, + 806 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "DaViT [4]: DaViT extends traditional vision transformers by incorporating both spatial and channel attention mechanisms, enhancing feature extraction across multiple dimensions. While conventional transformers rely solely on spatial self-attention, DaViT introduces channel attention to capture dependencies across feature channels, complement", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4790", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8f74f7fd647159277849ad27cb07b2f9afdcd6c96901d7904d49115109d8ea25.jpg", + "table_caption": [], + "table_footnote": [ + "Table 2. Performance evaluation of object detection and instance segmentation on the COCO mini-val set. The backbone networks are pretrained on ImageNet-1k and integrated into the Mask R-CNN framework. The reported metrics include $AP^b$ (Average Precision for bounding box predictions), $AP_{50/75}^b$ (Average Precision at IoU thresholds of 0.50 and 0.75 for bounding boxes), $AP^m$ (Average Precision for mask predictions), and $AP_{50/75}^m$ (Average Precision at IoU thresholds of 0.50 and 0.75 for mask predictions). In each case we see the XCiT architecture employing conditioned embedded tokens (ours) out performs the original architecture." + ], + "table_body": "
ModelAPbAPb50APb75APmAPm50APm75
XCiT-S + original embedded tokens44.966.148.940.163.142.8
XCiT-S + conditioned embedded tokens45.766.649.740.463.543.4
XCiT-M + original embedded tokens45.766.849.640.863.643.3
XCiT-M + conditioned embedded tokens46.267.449.841.463.844.2
", + "bbox": [ + 94, + 89, + 903, + 200 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing spatial interactions. This hybrid approach allows DaViT to effectively model both local and global relationships. In our experiments, we use the DaViT-Base (DaViT-B) model, configured with a patch size of 4, embedding dimensions of (128, 256, 512, 1024), attention heads of (4, 8, 16, 32), and a layer distribution of (1, 1, 9, 1).", + "bbox": [ + 89, + 311, + 480, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: Table 1 presents a comparison of the top- $1\\%$ test accuracy of five vision transformers trained from scratch on the ImageNet-1k dataset. Each model is evaluated using both the original embedded tokens and our conditioned embedded tokens, following the approach outlined in Sec. 4.2. While ViT-Base [5] and DeiT-Base [22] utilize standard self-attention mechanisms, Swin-Base [15], XCiT-Medium [2], and DaViT-Base [4] incorporate more advanced attention mechanisms. We include these latter models to illustrate that, although our theoretical analysis focuses on self-attention, conditioned embedded tokens can be easily integrated into modern transformers with more sophisticated architectures. As shown in the table, models using conditioned embedded tokens consistently achieve higher accuracy. Full implementation details and hyperparameter settings can be found in Sec. B.2 of the Supp. material.", + "bbox": [ + 89, + 429, + 482, + 671 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis: Figure 2 illustrates the condition numbers of the embedded tokens, the first-layer self-attention matrix (averaged across all heads), and the overall self-attention matrix (averaged across all layers) for the ViT-B architecture. We compare these metrics for models using the original embedded tokens and our conditioned embedded tokens. These values were computed at each training epoch and then averaged over all epochs to generate the plots. The figure clearly shows that the ViT-B architecture with conditioned embedded tokens maintains on average a significantly lower condition number throughout training.", + "bbox": [ + 89, + 672, + 482, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation and hardware: For full details on the implementation of each vision transformer, the training hyperparameters and the hardware used please see Sec. B.2 in the Supp. material.", + "bbox": [ + 89, + 839, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Object Detection and Image Segmentation", + "text_level": 1, + "bbox": [ + 511, + 310, + 872, + 327 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we extend our approach from Sec. 4.2 to object detection and instance segmentation tasks. To evaluate its effectiveness in fine-tuning scenarios, we adapt a pretrained XCiT model, originally trained on ImageNet-1k, for these tasks. Our experiments leverage the COCO 2017 dataset [13], which consists of 118K training images and 5K validation images spanning 80 object categories. For model architecture, we employ XCiT as the backbone of the Mask R-CNN framework [9], incorporating a Feature Pyramid Network (FPN) to enhance multi-scale feature representation. To integrate XCiT with FPN, we modify its inherently columnar structure by extracting features from multiple layers. We do this by using an XCiT-small that has 12 layers (XCiT-S) and a XCiT-medium (XCiT-M) that has 24 layers. These features, originally computed at a fixed stride of 16, are adjusted to strides of [4, 8, 16, 32] to align with the FPN hierarchy. Max pooling is used for downsampling, while a single transposed convolution layer facilitates upsampling. The model is then fine-tuned for 36 epochs using the AdamW optimizer, with a learning rate of $10^{-4}$ , a weight decay of 0.05, and a batch size of 16. This training strategy effectively demonstrates how XCiT's learned representations can be adapted for downstream tasks.", + "bbox": [ + 511, + 333, + 903, + 680 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results: We pretrained a total of four models on the ImageNet-1k dataset, namely an XCiT-S and XCiT-M both using the original embedded tokens following the methodology from Sec. 4.2. The results of the experiment are shown in Tab. 2. As can be seen from the table in both cases the transformer using conditioned embedded tokens out performs the original.", + "bbox": [ + 511, + 681, + 903, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used see Sec. B.3 in the Supp. material.", + "bbox": [ + 511, + 789, + 903, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Language Models", + "text_level": 1, + "bbox": [ + 511, + 847, + 686, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we evaluate our insights from Sec. 4.2 on two different language models.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4791", + "bbox": [ + 482, + 945, + 511, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a568153e1b4a87e5b9091aabb5766111d2c418b2310b50af60c346e337bf7dcc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MNLISST-2STSBRTEQNLIQQPMRPCCoLAGLUE
Crammed BERT (original)83.892.386.355.190.187.385.048.978.6
Crammed BERT (ours)84.292.586.555.691.187.486.353.779.7
", + "bbox": [ + 135, + 88, + 864, + 157 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Crammed BERT: In the first experiment we applied our insights from Sec. 4.2 to a Crammed BERT language model [8], trained entirely from scratch using masked language modeling with conditioned embedded tokens. The model consists of 110 million parameters, with 12 transformer layers and 12 attention heads per layer. We train two versions from scratch: the original Crammed BERT [8] and a variant incorporating conditioned embedded tokens using the insight from Sec. 4.2. Both models are trained on The Pile dataset [7], a large-scale corpus designed for language model training following the pretraining regime in [8]. After pretraining, we evaluate the performance on the GLUE benchmark [24] following the evaluation methodology outlined in [8].", + "bbox": [ + 88, + 222, + 485, + 434 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results: The evaluation results are shown in Tab. 3 with each number representing the accuracy. We see from the table that in each task the model employing conditioned embedded tokens outperforms the original Crammed Bert model on every task within the GLUE benchmark.", + "bbox": [ + 89, + 434, + 483, + 511 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GPT-2: We conducted a second experiment training a GPT-2 model with a decoder-only transformer architecture, consisting of 12 layers, 12 attention heads per layer, and an embedding dimension of 512. The model was trained using masked self-attention to predict the next token, leveraging the TinyStories dataset [6] for language generation. Two versions of the model were trained from scratch: one using the original embedded tokens as in [6] and another incorporating the conditioned embedded tokens from Sec. 4.2. Both models were evaluated using perplexity loss (PPL), a widely used metric for assessing language models on datasets like TinyStories. Perplexity measures the model's confidence in its predictions by computing the exponential of the average negative log-likelihood of the predicted tokens. A lower perplexity score indicates that the model assigns higher probabilities to correct tokens, leading to better fluency and coherence in text generation.", + "bbox": [ + 91, + 536, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results and Analysis: Tab. 4 presents the results of our experiment, demonstrating that the model trained with conditioned embedded tokens consistently achieves lower validation loss and lower perplexity compared to the original GPT-2 model. These results indicate that conditioning the embedded tokens improves model performance, leading to more accurate language generation. Fig. 3 plots the con", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cbd38a5e03f0b5f2bf5eb7d42e2d355ea5c761e69a73dd137f6bd1f603d639c7.jpg", + "table_caption": [ + "Table 3. Evaluation of a pre-trained Crammed BERT on the GLUE benchmark with the original embedded tokens (original) and our conditioned embedded tokens (ours). As can be seen from the table our methodology out performs the original in every task." + ], + "table_footnote": [], + "table_body": "
Val loss
GPT-2 original2.41
GPT-2 conditioned (ours)2.36
", + "bbox": [ + 578, + 219, + 841, + 270 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. GPT-2 model trained on the TinyStories dataset. We compare two models, the original model and one using conditioned embedded tokens (conditioned). We report the validation loss; in both cases, our model yields a lower loss, indicating better performance.", + "bbox": [ + 511, + 279, + 906, + 349 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/387be26c555014682eeea3fa4309c09c84fde3390f2c2c79465566c317baac9d.jpg", + "image_caption": [ + "Figure 3. Condition number comparison for GPT-2 training on TinyStories. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over all epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + ], + "image_footnote": [], + "bbox": [ + 516, + 363, + 890, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dition numbers of the embedded tokens, self-attention in first layer averaged over all heads and self-attention averaged over all layers. Each of these are evaluated over each training epoch and then averaged over all epochs.", + "bbox": [ + 511, + 681, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.4 in the Supp. material.", + "bbox": [ + 511, + 762, + 906, + 808 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Nyströmformer for Long Range Sequences", + "text_level": 1, + "bbox": [ + 511, + 818, + 879, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We applied our methodology to the Nyströmformer [25], a transformer architecture designed for efficient long-range dependency modeling on the Long-Range Arena (LRA) benchmark [21]. We trained two Nyströmformer models:", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4792", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/7c03fa035e6ed6b2fb009bdc4336915ef7d161fb421e39ce09ff3563d78bb73e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelListOpsTextRetrievalImagePathfinder
Nyströmformer (original)37.163.879.839.972.9
Nyströmformer conditioned (ours)37.964.980.940.180.0
", + "bbox": [ + 143, + 89, + 857, + 159 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison of a Nyströmformer using its original embedded tokens (original) with a Nyströmformer using conditioned embedded tokens (ours) on the LRA benchmark. We report the evaluation accuracy (\\%). As can be seen from the table our methodology yields a higher accuracy on each task.", + "bbox": [ + 88, + 171, + 906, + 218 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "one utilizing conditioned embedded tokens and the original Nyströmformer for comparison. The Long-Range Arena (LRA) benchmark consists of five distinct tasks:", + "bbox": [ + 89, + 241, + 480, + 285 + ], + "page_idx": 7 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. ListOps: A synthetic task that tests hierarchical reasoning by evaluating nested arithmetic operations.", + "2. Text Classification: Character-level text classification, requiring models to process long sequences of text.", + "3. Retrieval: A document retrieval task that measures the model's ability to identify relevant patterns within long textual inputs.", + "4. Image Classification: Modeling image data as a sequence of flattened pixel values to assess a model's capability in visual recognition.", + "5. Pathfinder: A visual reasoning challenge that requires detecting connected paths in an image, emphasizing spatial relationship modeling." + ], + "bbox": [ + 89, + 286, + 482, + 483 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results and Analysis: Tab. 5 shows the results of the experiment. As can be seen from the table the Nyströmformer that employed conditioned embedded tokens outperformed the original Nyströmformer from [25] in every task in the LRA benchmark. We computed the condition numbers of the original Nyströmformer and one employing conditioned embedded tokens throughout training for the embedded tokens, the self-attention within the first layer averaged over all heads and the self-attention averaged over all layers. We then took the average for each task over all training epochs and averaged over the five tasks within the LRA benchmark. The plots are shown in Fig. 4. As can be seen from the figure, the Nyströmformer employing conditioned embedded tokens has lower condition number in all cases.", + "bbox": [ + 88, + 500, + 482, + 710 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.5 in the Supp. material.", + "bbox": [ + 89, + 710, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Limitations", + "text_level": 1, + "bbox": [ + 89, + 768, + 212, + 786 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our work explored the conditioning of transformer attention blocks through embedded tokens. While we showed that conditioned embedded tokens reduce the first-layer self-attention condition number, their impact on optimization remains theoretically unproven. Prior studies [1, 10, 14] have linked NTK conditioning to optimization in feedforward networks, raising an important question: how does", + "bbox": [ + 88, + 794, + 482, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/509eeaef0fd0de7359dd3a4183512e40b1e73296a74c1eeb13a9f0437d909d39.jpg", + "image_caption": [ + "Figure 4. Condition number comparison for Nyströmformer on the LRA benchmark. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, averaged over training epochs and tasks. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + ], + "image_footnote": [], + "bbox": [ + 517, + 239, + 890, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "conditioning embedded tokens affect the NTK in transformers? Investigating this could offer deeper insights into how embedded token structure influences transformer optimization.", + "bbox": [ + 511, + 570, + 906, + 630 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 647, + 633, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a framework for analyzing the conditioning of the self-attention matrix in the first layer of a transformer, highlighting its dependence on the embedded tokens. Our analysis demonstrated that by adding a carefully designed correction term—which we refer to as conditioned embedded tokens—we could significantly reduce the condition number of self-attention in the first layer, leading to a better-conditioned attention mechanism. We empirically validated the effectiveness of conditioned embedded tokens across a diverse set of tasks, including image classification, object detection, instance segmentation, language modeling, and long-range sequence modeling. In all cases, our approach proved to be a simple, drop-in replacement for existing embedding methods, consistently improving model performance across these applications.", + "bbox": [ + 509, + 674, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4793", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Naman Agarwal, Pranjal Awasthi, and Satyen Kale. A deep conditioning treatment of neural networks. In Algorithmic Learning Theory, pages 249-305. PMLR, 2021. 3, 8", + "[2] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 3, 5, 6", + "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2", + "[4] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In European conference on computer vision, pages 74-92. Springer, 2022. 2, 5, 6", + "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5, 6", + "[6] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759, 2023. 7", + "[7] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Aadi Thite, Eric Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2021. 7", + "[8] Jonas Geiping and Tom Goldstein. Cramming: Training a language model on a singlegpu in one day. In International Conference on Machine Learning, pages 11117-11143. PMLR, 2023. 7", + "[9] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6", + "[10] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3, 8", + "[11] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. arXiv preprint arXiv:2403.19243, 2024. 3", + "[12] Yiping Ji, Hemanth Saratchandran, Peyman Moghaddam, and Simon Lucey. Always skip attention. arXiv preprint arXiv:2505.01996, 2025.3", + "[13] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Chaoyue Liu, Libin Zhu, and Mikhail Belkin. Loss landscapes and optimization in over-parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59:85-116, 2022. 2, 3, 8", + "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6", + "[16] Abhisek Maiti, Sander Oude Elberink, and George Vosselman. Transfusion: Multi-modal fusion network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6536-6546, 2023. 2", + "[17] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 2", + "[18] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Multi-agent generative trajectory forecasting with heterogeneous data for control. arXiv preprint arXiv:2001.03093, 2, 2020. 2", + "[19] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking attention: Polynomial alternatives to softmax in transformers. arXiv preprint arXiv:2410.18613, 2024. 3", + "[20] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2025. 2, 3", + "[21] Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. Long range arena: A benchmark for efficient transformers. arXiv preprint arXiv:2011.04006, 2020. 7", + "[22] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6", + "[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3", + "[24] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 7", + "[25] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14138–14148, 2021. 2, 3, 7, 8", + "[26] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 2", + "[27] Q Zhen, W Sun, H Deng, D Li, Y Wei, B Lv, J Yan, L Kong, and Y Zhong. cosformer: rethinking softmax in attention." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4794", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In International Conference on Learning Representations, 2022. 2", + "[28] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized bert pre-training approach with post-training. In Proceedings of the 20th chinese national conference on computational linguistics, pages 1218-1227, 2021. 2" + ], + "bbox": [ + 91, + 90, + 482, + 176 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4795", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_model.json b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..3bd97b640adc1259e910fb36d1e36ec11cd0fb23 --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_model.json @@ -0,0 +1,1991 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.165, + 0.131, + 0.834, + 0.154 + ], + "angle": 0, + "content": "Enhancing Transformers Through Conditioned Embedded Tokens" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.182, + 0.479, + 0.234 + ], + "angle": 0, + "content": "Hemanth Saratchandran \nAustralian Institute for Machine Learning \nAdelaide University" + }, + { + "type": "text", + "bbox": [ + 0.147, + 0.238, + 0.478, + 0.251 + ], + "angle": 0, + "content": "hemanth.saratchandran@adelaide.edu.au" + }, + { + "type": "text", + "bbox": [ + 0.52, + 0.182, + 0.851, + 0.235 + ], + "angle": 0, + "content": "Simon Lucey \nAustralian Institute for Machine Learning \nAdelaide University" + }, + { + "type": "text", + "bbox": [ + 0.564, + 0.238, + 0.804, + 0.251 + ], + "angle": 0, + "content": "simon.lucey@adelaide.edu.au" + }, + { + "type": "image", + "bbox": [ + 0.131, + 0.287, + 0.852, + 0.641 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.651, + 0.908, + 0.723 + ], + "angle": 0, + "content": "Figure 1. Schematic representation of conditioned embeddings for a vision transformer: An image is divided into \\(N\\) patches, with each patch then embedded as a high-dimensional vector \\(X_{i}\\in \\mathbb{R}^{d}\\) for \\(1\\leq i\\leq N\\). These vectors are then concatenated to form the embedded token matrix \\(X = [X_{1}\\dots X_{N}]^{T}\\). To improve the conditioning of the self-attention mechanism, a correction term \\(C = [C_1\\dots C_N]^T\\) is added to \\(X\\), reducing its condition number. The modified matrix is then fed into the first layer of the transformer (positional encoding not shown). Its effect on the self-attention equation is illustrated in the green equation (for simplicity, layer normalization has been omitted)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.734, + 0.327, + 0.751 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we de" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.888 + ], + "angle": 0, + "content": "velop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4786" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.166, + 0.223, + 0.182 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.191, + 0.485, + 0.403 + ], + "angle": 0, + "content": "The transformer architecture [23] has become a cornerstone of modern machine learning, driving breakthroughs across various domains including computer vision [3, 5, 15, 22], natural language processing [23, 27, 28], and robotics [16, 18]. Its success is largely attributed to its ability to model complex relationships in data without relying on traditional recurrent or convolutional structures. At the heart of the transformer is the self-attention mechanism, which dynamically assigns importance to different parts of the input, capturing both local and global dependencies. Unlike traditional architectures that process inputs in a fixed order, transformers utilize self-attention to integrate global context at each layer, making them highly effective for tasks that require contextual understanding." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.538 + ], + "angle": 0, + "content": "In general, transformers process input data by first converting it into discrete tokens, which are then mapped to a high-dimensional vector space through an embedding layer. For text, these tokens represent words or sub-words, while for images, image patches are treated as tokens. The resulting embedded tokens form a matrix that is passed into the first transformer layer for self-attention computations. This process allows the model to effectively learn relationships between tokens in a high-dimensional space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.484, + 0.659 + ], + "angle": 0, + "content": "In this paper, we explore the conditioning of the self-attention matrix in transformer architectures—an essential yet often overlooked factor influencing optimization dynamics. The conditioning of a matrix is measured by its condition number, defined as the ratio of its largest singular value to its smallest. A high condition number indicates ill-conditioning, which is a well-known challenge for the convergence of gradient-based optimization methods [17]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Previous work has addressed conditioning in feedforward neural networks. In [20], a methodology was introduced to condition network weights, leading to improved optimization. Similarly, [14] examined the condition of the neural tangent kernel (NTK), proposing a framework for reducing its condition number and demonstrating enhanced convergence for gradient descent. While these studies have focused on conditioning in feedforward architectures, its role in self-attention mechanisms remains largely unexplored. This gap motivates our investigation into how conditioning affects optimization in transformers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We develop a theoretical framework to analyze the condition number of the self-attention matrix within a transformer's attention layer. Our analysis establishes a direct relationship between the condition number of the first-layer self-attention matrix and that of the embedded tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.257 + ], + "angle": 0, + "content": "Leveraging this insight, we introduce conditioned embedded tokens, a method that applies a structured correction term to the embedded token matrix, effectively reducing its condition number and, in turn, improving the conditioning of the self-attention matrix. While a full theoretical proof linking this improvement to better convergence in gradient-based optimization would require analyzing the NTK of a transformer—a notoriously difficult task—we provide strong empirical evidence that conditioned embedded tokens enhances performance across a range of transformer applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.514 + ], + "angle": 0, + "content": "Figure 1 illustrates how conditioned embedded tokens are integrated into a vision transformer. An image is first divided into tokens via patches, which are vectorized and passed through a learnable embedding layer, producing embedded tokens in a high-dimensional vector space. These tokens form the matrix \\( X = [X_{1}\\dots X_{N}]^{T} \\). A correction term \\( C = [C_1\\cdot \\cdot \\cdot C_N]^T \\) is added to \\( X \\), reducing its condition number before the modified embedding matrix \\( X + C \\) is fed into the first transformer layer (positional encoding not shown). This correction not only lowers the condition number of the self-attention matrix in the first layer but also has a cascading effect, improving conditioning in subsequent layers. In general, positional encodings are added to the tokens before they are fed into the first transformer layer; however, they are omitted in Figure 1 for simplicity. Our theory will show that the optimal correction term can be derived from the singular value decomposition (SVD) of \\( X \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.515, + 0.906, + 0.68 + ], + "angle": 0, + "content": "We validate our framework across various applications, including image classification, object detection, instance segmentation, and natural language processing. A key advantage of conditioned embedded tokens is their flexibility—they can be used with a wide range of attention mechanisms, making them compatible with modern transformer architectures. As we demonstrate, our methodology can serve as a drop-in enhancement for advanced attention mechanisms proposed in recent works [2, 4, 15, 22, 25, 26], offering a simple yet effective improvement to existing models. Our main contributions include:" + }, + { + "type": "text", + "bbox": [ + 0.509, + 0.681, + 0.905, + 0.739 + ], + "angle": 0, + "content": "1. A theoretical framework introducing conditioned embedded tokens for improving the conditioning of the self-attention matrix within the first transformer layer yielding a more stable transformer architecture." + }, + { + "type": "text", + "bbox": [ + 0.508, + 0.741, + 0.905, + 0.815 + ], + "angle": 0, + "content": "2. An empirical evaluation of conditioned embedded tokens across various transformer architectures and applications, including image classification, object detection, instance segmentation, and natural language processing, showing in each case its superior performance." + }, + { + "type": "list", + "bbox": [ + 0.508, + 0.681, + 0.905, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.83, + 0.655, + 0.846 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Attention: Various strategies have been introduced to enhance the efficiency and effectiveness of transformers, particularly by addressing their computational overhead and re" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.332 + ], + "angle": 0, + "content": "thinking attention mechanisms. The Data-Efficient Image Transformer (DeiT) [22] improves training efficiency in vision transformers by leveraging distillation tokens, allowing for competitive performance without requiring massive datasets. The Cross-Covariance Image Transformer (XCiT) [2] redefines attention by operating on cross-covariances of spatial features, enabling more efficient spatial interactions while reducing computational demands. The Nyströmformer [25], approximates standard self-attention using the Nyström method, reducing quadratic complexity to near-linear. This technique has proven particularly effective for long-range sequence modeling. In this paper, we demonstrate that our conditioning framework can be integrated as a drop-in replacement into these advanced attention mechanisms, yielding superior performance across multiple architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.589 + ], + "angle": 0, + "content": "Conditioning: In [20], the authors investigate the condition number of weight matrices in feedforward neural networks and demonstrate that lower condition numbers lead to improved accuracy across various applications. They propose a preconditioner matrix that multiplies each layer's weight matrix, effectively reducing its condition number and enhancing the training of dense weights. In [14], neural network optimization is explored through the lens of the neural tangent kernel. The authors introduce the PL* condition, a variant of the Polyak-Lojasiewicz condition, and show that this approach improves the conditioning of the neural tangent kernel, which, in the infinite-width limit, governs the dynamics of gradient descent [10]. Finally, [1] demonstrates that increasing the depth of feedforward neural networks improves their conditioning leading to more efficient optimization. See [11, 12, 19, 20] for further results." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.603, + 0.192, + 0.619 + ], + "angle": 0, + "content": "3. Notation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.718 + ], + "angle": 0, + "content": "In this section, we formally define the transformer architecture by describing the structure of a transformer layer, along with establishing notation for various mathematical elements that will be referenced in subsequent sections. For further foundational details on transformers, readers may consult Vaswani et al. [23]." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.72, + 0.483, + 0.735 + ], + "angle": 0, + "content": "A layer in a transformer can be represented as a mapping" + }, + { + "type": "equation", + "bbox": [ + 0.213, + 0.744, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T}: \\mathbb {R} ^ {N \\times D} \\rightarrow \\mathbb {R} ^ {N \\times D} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.773, + 0.166, + 0.788 + ], + "angle": 0, + "content": "defined by" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.788, + 0.483, + 0.804 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} (X) = \\mathbf {F} (\\mathbf {A} (X) + X), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathbf{F}\\) denotes a feed forward network with a residual connection, and \\(\\mathbf{A}\\) represents an attention mechanism. In general, layer normalization is added however for simplicity we omit layer normalization for this discussion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "To begin with input data \\( x \\) is tokenized into \\( N \\) tokens yielding a tokenized representation \\( (x_{1},\\dots ,x_{N}) \\) where" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.135 + ], + "angle": 0, + "content": "\\(x_{i}\\in \\mathbb{R}^{t\\times 1}\\). Each \\(x_{i}\\) is then mapped to a high dimensional space \\(\\mathbb{R}^{d\\times 1}\\) using a learnable embedding layer \\(E\\in \\mathbb{R}^{d\\times t}\\) so that" + }, + { + "type": "equation", + "bbox": [ + 0.664, + 0.136, + 0.905, + 0.153 + ], + "angle": 0, + "content": "\\[\nE x _ {i} \\in \\mathbb {R} ^ {d \\times 1}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.158, + 0.907, + 0.279 + ], + "angle": 0, + "content": "The matrix \\([Ex_1 \\cdots Ex_n]^T \\in \\mathbb{R}^{N \\times d}\\) represents the embedded tokens associated to the input data. To capture the order of the tokens, a positional encoding matrix \\(P \\in \\mathbb{R}^{N \\times d}\\) is added forming \\(X = [Ex_1 \\cdots Ex_n]^T + P \\in \\mathbb{R}^{N \\times d}\\). The matrix \\(X\\) is then input into the first layer of the transformer. For the theoretical Sec. 4 we will often use the notation \\(X\\) to denote embedded tokens often leaving out explicit mention of the positional encoding part for ease of notation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.906, + 0.355 + ], + "angle": 0, + "content": "The key component of the transformer is self-attention. This is composed of three learnable matrices, a query \\((W_{Q})\\), key \\((W_{K})\\), and value \\((W_{V})\\), where \\(W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times d}\\) and \\(W_{V} \\in \\mathbb{R}^{D \\times d}\\). The output of the attention head \\(\\mathbf{A}(X)\\) is then given by" + }, + { + "type": "equation", + "bbox": [ + 0.567, + 0.363, + 0.905, + 0.382 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} (X) = \\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right) X W _ {V}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.391, + 0.906, + 0.452 + ], + "angle": 0, + "content": "where softmax is the softmax activation that acts row-wise on the matrix \\( XW_{Q}W_{k}^{T}X^{T} \\). Note that then \\( A(X) \\in \\mathbb{R}^{N \\times d} \\). In the case no activation is used we have the definition of linear self-attention which we will denote by" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.46, + 0.905, + 0.48 + ], + "angle": 0, + "content": "\\[\n\\mathbf {L A} (X) := X W _ {Q} W _ {k} ^ {T} X ^ {T} X W _ {V}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.488, + 0.906, + 0.534 + ], + "angle": 0, + "content": "In general, multiple attention heads \\(\\mathbf{A}_i\\) for \\(1\\leq i\\leq h\\) are utilized, each of dimension \\(N\\times d_{i} = N\\times \\frac{d}{h}\\). These are then concatenated to produce a multi-head attention output," + }, + { + "type": "equation", + "bbox": [ + 0.659, + 0.544, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\n[ \\mathbf {A} _ {1}, \\dots , \\mathbf {A} _ {h} ], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.57, + 0.906, + 0.646 + ], + "angle": 0, + "content": "which is subsequently fed into a feedforward layer. Similarly, a multi-head linear attention can be defined in the same way. The full transformer architecture is obtained by sequentially stacking several such transformer layers, as defined in Eq. (2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.738, + 0.673 + ], + "angle": 0, + "content": "4. Theoretical Framework" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.776, + 0.699 + ], + "angle": 0, + "content": "4.1. Conditioning of self-attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this section, we analyze the conditioning of the self-attention matrix in a transformer. Our objective is to demonstrate that the condition number of the self-attention matrix in the first layer is influenced by the condition number of the embedded tokens. This insight motivates our key finding: reducing the condition number of the embedded tokens leads to a better-conditioned self-attention matrix in the first layer, ultimately improving the overall conditioning of the transformer. In this section, we omit positional encoding and layer normalization to keep the core theoretical insight clear and accessible. In Sec. 5, we will show our insights go through for a variety of practical transformers that use positional encoding and layer normalization." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4788" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "Definition 4.1. Let \\( A \\) be an \\( N \\times d \\) matrix of full rank. The condition number of \\( A \\), denoted by \\( \\kappa \\), is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.127, + 0.483, + 0.157 + ], + "angle": 0, + "content": "\\[\n\\kappa (A) = \\frac {\\sigma_ {1}}{\\sigma_ {k}} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.162, + 0.484, + 0.238 + ], + "angle": 0, + "content": "where \\( k = \\min \\{N, d\\} \\), and \\( \\sigma_1 \\) denotes the largest singular value of \\( A \\) and \\( \\sigma_k \\) the smallest singular value of \\( A \\). Note that as \\( A \\) is assumed full rank \\( \\sigma_k \\neq 0 \\) and the condition number is well-defined. Furthermore, observe that \\( \\kappa(A) \\geq 1 \\) since \\( \\sigma_1 \\geq \\sigma_k \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.244, + 0.484, + 0.274 + ], + "angle": 0, + "content": "We have the following estimate for the condition number of linear self-attention and softmax self-attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.28, + 0.484, + 0.326 + ], + "angle": 0, + "content": "Proposition 4.2. Let \\( X \\) denote an input for an attention layer as defined in Sec. 3. Assume \\( \\mathbf{L}\\mathbf{A}(X) \\) and \\( \\mathbf{A}(X) \\) have full rank. We then have" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.332, + 0.483, + 0.385 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\kappa (\\mathbf {L A} (X)) \\leq \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3} (8) \\\\ \\kappa (\\mathbf {A} (X)) \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\kappa (X) \\kappa \\left(W _ {V}\\right). (9) \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.484, + 0.422 + ], + "angle": 0, + "content": "The proof of Proposition 4.2 is given in Sec. A of the Supp. material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.485 + ], + "angle": 0, + "content": "Remark 4.3. The statement of Proposition 4.2 assumes the attention matrices are full rank. Since the layers of a transformer are randomly initialized we have that with probability 1 the attention matrices will be full rank at initialization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.484, + 0.58 + ], + "angle": 0, + "content": "We point out that although Eq. (8) and Eq. (9) only provide an upper bound for the condition number, we found that the this upper bound provides a useful measure for estimating the condition number of the self-attention matrix. Therefore, for the self-attention matrices \\(\\mathbf{A}(X)\\) and \\(\\mathbf{L}\\mathbf{A}(X)\\) we make the following definition" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.586, + 0.483, + 0.638 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu (\\mathbf {L A} (X)) := \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3}. (10) \\\\ \\mu (\\mathbf {A} (X)) := \\kappa \\left(\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right)\\right) \\kappa (X) \\kappa \\left(W _ {V}\\right). (11) \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.645, + 0.361, + 0.66 + ], + "angle": 0, + "content": "4.2. Conditioned embedded tokens" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.667, + 0.484, + 0.759 + ], + "angle": 0, + "content": "In this section, we present our main theorem, which provides a method for reducing the condition number of embedded tokens derived from a dataset. This reduction subsequently decreases the measure \\(\\mu (\\mathbf{A}(X))\\), a key metric for assessing the condition number of the self-attention matrix in the first transformer layer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Theorem 4.4. Let \\((x_{1},\\dots ,x_{N})\\) denote \\(N\\) tokens associated to an input dataset with \\(x_{i}\\in \\mathbb{R}^{t\\times 1}\\). Let \\(E\\in \\mathbb{R}^{d\\times t}\\) denote an embedding matrix so that \\(Ex_{i}\\in \\mathbb{R}^{d\\times 1}\\). Let \\(X = [Ex_1\\dots Ex_N]^T\\in \\mathbb{R}^{N\\times d}\\) denote the matrix of embedded tokens. Assume that \\(\\kappa (X) > 2\\). Then there exists a matrix \\(C\\in \\mathbb{R}^{N\\times d}\\) such that" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.863, + 0.483, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\kappa (X + C) \\leq 2 \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.886, + 0.328, + 0.902 + ], + "angle": 0, + "content": "and hence that \\(\\kappa(X + C) < \\kappa(X)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.198 + ], + "angle": 0, + "content": "In particular, Theorem 4.4 shows that if the condition number of the embedded tokens \\( X \\) is extremely large. Adding a correction term given by \\( C \\) will bring it down significantly. As we shall see in the experiments Sec. 5 the condition number \\( \\kappa(X) \\) is empirically extremely large. The proof of Theorem 4.4 is given in Sec. A of the Supp. material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.208, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Theorem 4.5. Let \\( X \\) denote a collection of embedded tokens, as defined in the statement of Theorem 4.4, and let \\( C \\) denote the matrix given by Theorem 4.4. We then have" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.265, + 0.907, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\mu (\\mathbf {L A} (X + C)) \\leq \\mu (\\mathbf {L A} (X)). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.294, + 0.6, + 0.307 + ], + "angle": 0, + "content": "Assume that" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.319, + 0.906, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\kappa (\\text {s o f t m a x} ((X + C) W _ {Q} W _ {K} ^ {T} (X ^ {T} + C ^ {T}))) \\\\ \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\tag {14} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.37, + 0.547, + 0.383 + ], + "angle": 0, + "content": "then" + }, + { + "type": "equation", + "bbox": [ + 0.616, + 0.384, + 0.905, + 0.401 + ], + "angle": 0, + "content": "\\[\n\\mu (\\mathbf {A} (X + C)) \\leq \\mu (\\mathbf {A} (X)) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.407, + 0.906, + 0.438 + ], + "angle": 0, + "content": "where \\(\\mu (\\mathbf{L}\\mathbf{A}(X))\\) is defined by Eq. (10) and \\(\\mu (\\mathbf{A}(X))\\) by Eq. (11)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.598 + ], + "angle": 0, + "content": "The proof of Theorem 4.5 is given in Sec. A of the Supp. material. A natural question arises: can self-attention be conditioned directly through queries, keys, and values, independent of the embedded tokens? Unfortunately, our approach of introducing a correction term does not extend to these components. This is because our correction modifies the singular values of the embedded tokens, whereas altering the singular values of queries, keys, or values could disrupt how self-attention captures spatial relationships between tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.604, + 0.906, + 0.726 + ], + "angle": 0, + "content": "Remark 4.6. We note that in the case of self-attention with a softmax activation Theorem 4.5 requires an assumption on the condition number of the non-linear self-attention probability matrix term softmax \\(((X + C)W_{Q}W_{K}^{T}(X^{T} + C^{T}))\\). Although this may be seen as a limitation of the theory we will see in Sec. 5 that this framework still provides a useful methodology to obtain good performance on a variety of transformer applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.73, + 0.906, + 0.836 + ], + "angle": 0, + "content": "Remark 4.7. Theorem 4.5 demonstrates that conditioning embedded tokens can improve the condition number of the self-attention matrix in the first layer. However, it does not address how this affects the condition number in subsequent layers. In Sec. 5, we will show that this effect propagates through the network, leading to a reduced condition number in deeper self-attention layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Remark 4.8. Our theory, as presented in Theorem 4.4 and Theorem 4.5, demonstrates that conditioning embedded tokens reduces the condition number of the self-attention matrix in the first layer. However, the framework does not" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4789" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.159, + 0.091, + 0.84, + 0.174 + ], + "angle": 0, + "content": "
Models
ViT-baseDeiT-baseSwin-baseXciT-mediumDaViT-base
Original embedded tokens80.381.683.182.283.6
Conditioned embedded tokens81.382.583.982.984.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.187, + 0.908, + 0.232 + ], + "angle": 0, + "content": "Table 1. Comparison of vision transformers with their original embedded tokens verse one with a conditioned conditioned embedded tokens pre-trained on the ImageNet-1k dataset. We report the classification top-1% accuracy. In each case we see the transformers employing conditioned embedded tokens (ours) outperforms the original ones." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.256, + 0.482, + 0.348 + ], + "angle": 0, + "content": "theoretically establish how this improvement translates to better optimization and, consequently, higher accuracy in applications. Nonetheless, in Sec. 5, we provide empirical evidence showing that our approach consistently enhances accuracy across various applications, highlighting its practical effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.361, + 0.226, + 0.379 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.386, + 0.484, + 0.448 + ], + "angle": 0, + "content": "In this section, we evaluate our insight from Sec. 4 on a variety of transformer applications. For a detailed analysis of how we implemented conditioned embedded tokens see Sec. A in the Supp. material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.456, + 0.284, + 0.472 + ], + "angle": 0, + "content": "5.1. Image Classification" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.478, + 0.484, + 0.583 + ], + "angle": 0, + "content": "In this section, we apply our theoretical insights from Sec. 4 to vision transformers on an image classification task. We will train all vision transformers from scratch on the ImageNet-1k dataset. We will be testing on the following vision transformers. In each case we will compare the original architecture with one that uses conditioned embedded tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.689 + ], + "angle": 0, + "content": "ViT [5]: ViT is a pioneering architecture that applies transformer-based processing to images by treating them as sequences of non-overlapping patches. In our study, we utilize ViT-Base (ViT-B), which is configured with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. The model employs a standard self-attention mechanism to capture dependencies across patches." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.809 + ], + "angle": 0, + "content": "DeiT [22]: DeiT builds upon ViT but is optimized for data efficiency. Unlike standard ViT, it employs a data-starvation training strategy [22], enabling faster convergence while maintaining strong performance. In our setup, we use DeiT-Base (DeiT-B) with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. Like ViT, DeiT utilizes a self-attention mechanism for feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Swin Transformer [15]: Swin Transformer introduces a hierarchical architecture and a novel shifted window-based self-attention mechanism, significantly improving efficiency and effectiveness in vision tasks. In our experiments, we adopt the Swin-Base (Swin-B) variant, which features 128 channels in the hidden layers of the first stage." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.254, + 0.892, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.46, + 0.908, + 0.545 + ], + "angle": 0, + "content": "Figure 2. Condition number comparison for ViT-B on ImageNet-1k. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over training epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.577, + 0.907, + 0.639 + ], + "angle": 0, + "content": "The attention operates over non-overlapping windows of size \\( M = 7 \\), while the query dimension per head is set to \\( d = 32 \\). The network follows a hierarchical structure with layers arranged as \\( \\{2, 2, 18, 2\\} \\) across different stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.641, + 0.909, + 0.807 + ], + "angle": 0, + "content": "XCiT [2]: XCiT diverges from standard ViT architectures by incorporating two key innovations. First, it integrates Local Patch Interaction within each block, using depth-wise \\(3 \\times 3\\) convolutions, Batch Normalization, GELU activation, and an additional depth-wise convolution to enhance local feature representation. Second, it employs Cross-Covariance Attention, where attention maps are derived from the cross-covariance matrix of the key and query projections. For our experiments, we utilize the XCiT-M24 (XCiT-M) model with a patch size of 16, an embedding dimension of 512, 8 attention heads, and 24 layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "DaViT [4]: DaViT extends traditional vision transformers by incorporating both spatial and channel attention mechanisms, enhancing feature extraction across multiple dimensions. While conventional transformers rely solely on spatial self-attention, DaViT introduces channel attention to capture dependencies across feature channels, complement" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4790" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.904, + 0.202 + ], + "angle": 0, + "content": "
ModelAPbAPb50APb75APmAPm50APm75
XCiT-S + original embedded tokens44.966.148.940.163.142.8
XCiT-S + conditioned embedded tokens45.766.649.740.463.543.4
XCiT-M + original embedded tokens45.766.849.640.863.643.3
XCiT-M + conditioned embedded tokens46.267.449.841.463.844.2
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.216, + 0.907, + 0.286 + ], + "angle": 0, + "content": "Table 2. Performance evaluation of object detection and instance segmentation on the COCO mini-val set. The backbone networks are pretrained on ImageNet-1k and integrated into the Mask R-CNN framework. The reported metrics include \\(AP^b\\) (Average Precision for bounding box predictions), \\(AP_{50/75}^b\\) (Average Precision at IoU thresholds of 0.50 and 0.75 for bounding boxes), \\(AP^m\\) (Average Precision for mask predictions), and \\(AP_{50/75}^m\\) (Average Precision at IoU thresholds of 0.50 and 0.75 for mask predictions). In each case we see the XCiT architecture employing conditioned embedded tokens (ours) out performs the original architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.482, + 0.404 + ], + "angle": 0, + "content": "ing spatial interactions. This hybrid approach allows DaViT to effectively model both local and global relationships. In our experiments, we use the DaViT-Base (DaViT-B) model, configured with a patch size of 4, embedding dimensions of (128, 256, 512, 1024), attention heads of (4, 8, 16, 32), and a layer distribution of (1, 1, 9, 1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.672 + ], + "angle": 0, + "content": "Results: Table 1 presents a comparison of the top- \\(1\\%\\) test accuracy of five vision transformers trained from scratch on the ImageNet-1k dataset. Each model is evaluated using both the original embedded tokens and our conditioned embedded tokens, following the approach outlined in Sec. 4.2. While ViT-Base [5] and DeiT-Base [22] utilize standard self-attention mechanisms, Swin-Base [15], XCiT-Medium [2], and DaViT-Base [4] incorporate more advanced attention mechanisms. We include these latter models to illustrate that, although our theoretical analysis focuses on self-attention, conditioned embedded tokens can be easily integrated into modern transformers with more sophisticated architectures. As shown in the table, models using conditioned embedded tokens consistently achieve higher accuracy. Full implementation details and hyperparameter settings can be found in Sec. B.2 of the Supp. material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.839 + ], + "angle": 0, + "content": "Analysis: Figure 2 illustrates the condition numbers of the embedded tokens, the first-layer self-attention matrix (averaged across all heads), and the overall self-attention matrix (averaged across all layers) for the ViT-B architecture. We compare these metrics for models using the original embedded tokens and our conditioned embedded tokens. These values were computed at each training epoch and then averaged over all epochs to generate the plots. The figure clearly shows that the ViT-B architecture with conditioned embedded tokens maintains on average a significantly lower condition number throughout training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Implementation and hardware: For full details on the implementation of each vision transformer, the training hyperparameters and the hardware used please see Sec. B.2 in the Supp. material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.311, + 0.874, + 0.328 + ], + "angle": 0, + "content": "5.2. Object Detection and Image Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.334, + 0.905, + 0.681 + ], + "angle": 0, + "content": "In this section, we extend our approach from Sec. 4.2 to object detection and instance segmentation tasks. To evaluate its effectiveness in fine-tuning scenarios, we adapt a pretrained XCiT model, originally trained on ImageNet-1k, for these tasks. Our experiments leverage the COCO 2017 dataset [13], which consists of 118K training images and 5K validation images spanning 80 object categories. For model architecture, we employ XCiT as the backbone of the Mask R-CNN framework [9], incorporating a Feature Pyramid Network (FPN) to enhance multi-scale feature representation. To integrate XCiT with FPN, we modify its inherently columnar structure by extracting features from multiple layers. We do this by using an XCiT-small that has 12 layers (XCiT-S) and a XCiT-medium (XCiT-M) that has 24 layers. These features, originally computed at a fixed stride of 16, are adjusted to strides of [4, 8, 16, 32] to align with the FPN hierarchy. Max pooling is used for downsampling, while a single transposed convolution layer facilitates upsampling. The model is then fine-tuned for 36 epochs using the AdamW optimizer, with a learning rate of \\(10^{-4}\\), a weight decay of 0.05, and a batch size of 16. This training strategy effectively demonstrates how XCiT's learned representations can be adapted for downstream tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.683, + 0.905, + 0.789 + ], + "angle": 0, + "content": "Results: We pretrained a total of four models on the ImageNet-1k dataset, namely an XCiT-S and XCiT-M both using the original embedded tokens following the methodology from Sec. 4.2. The results of the experiment are shown in Tab. 2. As can be seen from the table in both cases the transformer using conditioned embedded tokens out performs the original." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.79, + 0.905, + 0.835 + ], + "angle": 0, + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used see Sec. B.3 in the Supp. material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.687, + 0.864 + ], + "angle": 0, + "content": "5.3. Language Models" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In this section, we evaluate our insights from Sec. 4.2 on two different language models." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.513, + 0.957 + ], + "angle": 0, + "content": "4791" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.136, + 0.089, + 0.865, + 0.158 + ], + "angle": 0, + "content": "
MNLISST-2STSBRTEQNLIQQPMRPCCoLAGLUE
Crammed BERT (original)83.892.386.355.190.187.385.048.978.6
Crammed BERT (ours)84.292.586.555.691.187.486.353.779.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.168, + 0.908, + 0.199 + ], + "angle": 0, + "content": "Table 3. Evaluation of a pre-trained Crammed BERT on the GLUE benchmark with the original embedded tokens (original) and our conditioned embedded tokens (ours). As can be seen from the table our methodology out performs the original in every task." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.223, + 0.486, + 0.435 + ], + "angle": 0, + "content": "Crammed BERT: In the first experiment we applied our insights from Sec. 4.2 to a Crammed BERT language model [8], trained entirely from scratch using masked language modeling with conditioned embedded tokens. The model consists of 110 million parameters, with 12 transformer layers and 12 attention heads per layer. We train two versions from scratch: the original Crammed BERT [8] and a variant incorporating conditioned embedded tokens using the insight from Sec. 4.2. Both models are trained on The Pile dataset [7], a large-scale corpus designed for language model training following the pretraining regime in [8]. After pretraining, we evaluate the performance on the GLUE benchmark [24] following the evaluation methodology outlined in [8]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.485, + 0.512 + ], + "angle": 0, + "content": "Results: The evaluation results are shown in Tab. 3 with each number representing the accuracy. We see from the table that in each task the model employing conditioned embedded tokens outperforms the original Crammed Bert model on every task within the GLUE benchmark." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.537, + 0.483, + 0.794 + ], + "angle": 0, + "content": "GPT-2: We conducted a second experiment training a GPT-2 model with a decoder-only transformer architecture, consisting of 12 layers, 12 attention heads per layer, and an embedding dimension of 512. The model was trained using masked self-attention to predict the next token, leveraging the TinyStories dataset [6] for language generation. Two versions of the model were trained from scratch: one using the original embedded tokens as in [6] and another incorporating the conditioned embedded tokens from Sec. 4.2. Both models were evaluated using perplexity loss (PPL), a widely used metric for assessing language models on datasets like TinyStories. Perplexity measures the model's confidence in its predictions by computing the exponential of the average negative log-likelihood of the predicted tokens. A lower perplexity score indicates that the model assigns higher probabilities to correct tokens, leading to better fluency and coherence in text generation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.485, + 0.904 + ], + "angle": 0, + "content": "Results and Analysis: Tab. 4 presents the results of our experiment, demonstrating that the model trained with conditioned embedded tokens consistently achieves lower validation loss and lower perplexity compared to the original GPT-2 model. These results indicate that conditioning the embedded tokens improves model performance, leading to more accurate language generation. Fig. 3 plots the con" + }, + { + "type": "table", + "bbox": [ + 0.58, + 0.22, + 0.842, + 0.271 + ], + "angle": 0, + "content": "
Val loss
GPT-2 original2.41
GPT-2 conditioned (ours)2.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.28, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Table 4. GPT-2 model trained on the TinyStories dataset. We compare two models, the original model and one using conditioned embedded tokens (conditioned). We report the validation loss; in both cases, our model yields a lower loss, indicating better performance." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.364, + 0.892, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.571, + 0.909, + 0.656 + ], + "angle": 0, + "content": "Figure 3. Condition number comparison for GPT-2 training on TinyStories. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over all epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.682, + 0.908, + 0.744 + ], + "angle": 0, + "content": "dition numbers of the embedded tokens, self-attention in first layer averaged over all heads and self-attention averaged over all layers. Each of these are evaluated over each training epoch and then averaged over all epochs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.908, + 0.809 + ], + "angle": 0, + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.4 in the Supp. material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.88, + 0.836 + ], + "angle": 0, + "content": "5.4. Nyströmformer for Long Range Sequences" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.908, + 0.903 + ], + "angle": 0, + "content": "We applied our methodology to the Nyströmformer [25], a transformer architecture designed for efficient long-range dependency modeling on the Long-Range Arena (LRA) benchmark [21]. We trained two Nyströmformer models:" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "4792" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.144, + 0.09, + 0.858, + 0.16 + ], + "angle": 0, + "content": "
ModelListOpsTextRetrievalImagePathfinder
Nyströmformer (original)37.163.879.839.972.9
Nyströmformer conditioned (ours)37.964.980.940.180.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.172, + 0.908, + 0.219 + ], + "angle": 0, + "content": "Table 5. Comparison of a Nyströmformer using its original embedded tokens (original) with a Nyströmformer using conditioned embedded tokens (ours) on the LRA benchmark. We report the evaluation accuracy (\\%). As can be seen from the table our methodology yields a higher accuracy on each task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.482, + 0.286 + ], + "angle": 0, + "content": "one utilizing conditioned embedded tokens and the original Nyströmformer for comparison. The Long-Range Arena (LRA) benchmark consists of five distinct tasks:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.287, + 0.483, + 0.317 + ], + "angle": 0, + "content": "1. ListOps: A synthetic task that tests hierarchical reasoning by evaluating nested arithmetic operations." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.317, + 0.482, + 0.347 + ], + "angle": 0, + "content": "2. Text Classification: Character-level text classification, requiring models to process long sequences of text." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.347, + 0.483, + 0.392 + ], + "angle": 0, + "content": "3. Retrieval: A document retrieval task that measures the model's ability to identify relevant patterns within long textual inputs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.392, + 0.483, + 0.437 + ], + "angle": 0, + "content": "4. Image Classification: Modeling image data as a sequence of flattened pixel values to assess a model's capability in visual recognition." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.438, + 0.483, + 0.484 + ], + "angle": 0, + "content": "5. Pathfinder: A visual reasoning challenge that requires detecting connected paths in an image, emphasizing spatial relationship modeling." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.287, + 0.483, + 0.484 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.501, + 0.483, + 0.712 + ], + "angle": 0, + "content": "Results and Analysis: Tab. 5 shows the results of the experiment. As can be seen from the table the Nyströmformer that employed conditioned embedded tokens outperformed the original Nyströmformer from [25] in every task in the LRA benchmark. We computed the condition numbers of the original Nyströmformer and one employing conditioned embedded tokens throughout training for the embedded tokens, the self-attention within the first layer averaged over all heads and the self-attention averaged over all layers. We then took the average for each task over all training epochs and averaged over the five tasks within the LRA benchmark. The plots are shown in Fig. 4. As can be seen from the figure, the Nyströmformer employing conditioned embedded tokens has lower condition number in all cases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.483, + 0.759 + ], + "angle": 0, + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.5 in the Supp. material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.77, + 0.214, + 0.787 + ], + "angle": 0, + "content": "6. Limitations" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.795, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our work explored the conditioning of transformer attention blocks through embedded tokens. While we showed that conditioned embedded tokens reduce the first-layer self-attention condition number, their impact on optimization remains theoretically unproven. Prior studies [1, 10, 14] have linked NTK conditioning to optimization in feedforward networks, raising an important question: how does" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.24, + 0.892, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.445, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 4. Condition number comparison for Nyströmformer on the LRA benchmark. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, averaged over training epochs and tasks. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.907, + 0.631 + ], + "angle": 0, + "content": "conditioning embedded tokens affect the NTK in transformers? Investigating this could offer deeper insights into how embedded token structure influences transformer optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.634, + 0.664 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.675, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we introduced a framework for analyzing the conditioning of the self-attention matrix in the first layer of a transformer, highlighting its dependence on the embedded tokens. Our analysis demonstrated that by adding a carefully designed correction term—which we refer to as conditioned embedded tokens—we could significantly reduce the condition number of self-attention in the first layer, leading to a better-conditioned attention mechanism. We empirically validated the effectiveness of conditioned embedded tokens across a diverse set of tasks, including image classification, object detection, instance segmentation, language modeling, and long-range sequence modeling. In all cases, our approach proved to be a simple, drop-in replacement for existing embedding methods, consistently improving model performance across these applications." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4793" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Naman Agarwal, Pranjal Awasthi, and Satyen Kale. A deep conditioning treatment of neural networks. In Algorithmic Learning Theory, pages 249-305. PMLR, 2021. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.302, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[4] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In European conference on computer vision, pages 74-92. Springer, 2022. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.442 + ], + "angle": 0, + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.444, + 0.483, + 0.486 + ], + "angle": 0, + "content": "[6] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.483, + 0.557 + ], + "angle": 0, + "content": "[7] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Aadi Thite, Eric Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.559, + 0.483, + 0.613 + ], + "angle": 0, + "content": "[8] Jonas Geiping and Tom Goldstein. Cramming: Training a language model on a singlegpu in one day. In International Conference on Machine Learning, pages 11117-11143. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.616, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.483, + 0.715 + ], + "angle": 0, + "content": "[10] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.483, + 0.771 + ], + "angle": 0, + "content": "[11] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. arXiv preprint arXiv:2403.19243, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.815 + ], + "angle": 0, + "content": "[12] Yiping Ji, Hemanth Saratchandran, Peyman Moghaddam, and Simon Lucey. Always skip attention. arXiv preprint arXiv:2505.01996, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Chaoyue Liu, Libin Zhu, and Mikhail Belkin. Loss landscapes and optimization in over-parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59:85-116, 2022. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.221, + 0.906, + 0.288 + ], + "angle": 0, + "content": "[16] Abhisek Maiti, Sander Oude Elberink, and George Vosselman. Transfusion: Multi-modal fusion network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6536-6546, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[17] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[18] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Multi-agent generative trajectory forecasting with heterogeneous data for control. arXiv preprint arXiv:2001.03093, 2, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.378, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[19] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking attention: Polynomial alternatives to softmax in transformers. arXiv preprint arXiv:2410.18613, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[20] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.491, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[21] Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. Long range arena: A benchmark for efficient transformers. arXiv preprint arXiv:2011.04006, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.906, + 0.63 + ], + "angle": 0, + "content": "[22] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.633, + 0.906, + 0.687 + ], + "angle": 0, + "content": "[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.906, + 0.744 + ], + "angle": 0, + "content": "[24] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.906, + 0.815 + ], + "angle": 0, + "content": "[25] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14138–14148, 2021. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[26] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Q Zhen, W Sun, H Deng, D Li, Y Wei, B Lv, J Yan, L Kong, and Y Zhong. cosformer: rethinking softmax in attention." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4794" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "In International Conference on Learning Representations, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.178 + ], + "angle": 0, + "content": "[28] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized bert pre-training approach with post-training. In Proceedings of the 20th chinese national conference on computational linguistics, pages 1218-1227, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "4795" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_origin.pdf b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b9581fe2ef69dd59881896deaad785a95fcf0983 --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/e4867dd3-615a-4802-a597-48b7c4a2f52c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88fb17ff836e1fb4d1c99c6efc099451ea021d2a3c5d98c79a50e99a5e7c92b6 +size 476059 diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/full.md b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..90415552ae4c051ea52befabb97466b3f276ff14 --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/full.md @@ -0,0 +1,318 @@ +# Enhancing Transformers Through Conditioned Embedded Tokens + +Hemanth Saratchandran +Australian Institute for Machine Learning +Adelaide University + +hemanth.saratchandran@adelaide.edu.au + +Simon Lucey +Australian Institute for Machine Learning +Adelaide University + +simon.lucey@adelaide.edu.au + +![](images/60619f460852316481314c0f675e2b59c233c40ff87cbb2b669a8ea26eb4c00e.jpg) +Figure 1. Schematic representation of conditioned embeddings for a vision transformer: An image is divided into $N$ patches, with each patch then embedded as a high-dimensional vector $X_{i}\in \mathbb{R}^{d}$ for $1\leq i\leq N$ . These vectors are then concatenated to form the embedded token matrix $X = [X_{1}\dots X_{N}]^{T}$ . To improve the conditioning of the self-attention mechanism, a correction term $C = [C_1\dots C_N]^T$ is added to $X$ , reducing its condition number. The modified matrix is then fed into the first layer of the transformer (positional encoding not shown). Its effect on the self-attention equation is illustrated in the green equation (for simplicity, layer normalization has been omitted). + +# Abstract + +Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we de + +velop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image + +classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness. + +# 1. Introduction + +The transformer architecture [23] has become a cornerstone of modern machine learning, driving breakthroughs across various domains including computer vision [3, 5, 15, 22], natural language processing [23, 27, 28], and robotics [16, 18]. Its success is largely attributed to its ability to model complex relationships in data without relying on traditional recurrent or convolutional structures. At the heart of the transformer is the self-attention mechanism, which dynamically assigns importance to different parts of the input, capturing both local and global dependencies. Unlike traditional architectures that process inputs in a fixed order, transformers utilize self-attention to integrate global context at each layer, making them highly effective for tasks that require contextual understanding. + +In general, transformers process input data by first converting it into discrete tokens, which are then mapped to a high-dimensional vector space through an embedding layer. For text, these tokens represent words or sub-words, while for images, image patches are treated as tokens. The resulting embedded tokens form a matrix that is passed into the first transformer layer for self-attention computations. This process allows the model to effectively learn relationships between tokens in a high-dimensional space. + +In this paper, we explore the conditioning of the self-attention matrix in transformer architectures—an essential yet often overlooked factor influencing optimization dynamics. The conditioning of a matrix is measured by its condition number, defined as the ratio of its largest singular value to its smallest. A high condition number indicates ill-conditioning, which is a well-known challenge for the convergence of gradient-based optimization methods [17]. + +Previous work has addressed conditioning in feedforward neural networks. In [20], a methodology was introduced to condition network weights, leading to improved optimization. Similarly, [14] examined the condition of the neural tangent kernel (NTK), proposing a framework for reducing its condition number and demonstrating enhanced convergence for gradient descent. While these studies have focused on conditioning in feedforward architectures, its role in self-attention mechanisms remains largely unexplored. This gap motivates our investigation into how conditioning affects optimization in transformers. + +We develop a theoretical framework to analyze the condition number of the self-attention matrix within a transformer's attention layer. Our analysis establishes a direct relationship between the condition number of the first-layer self-attention matrix and that of the embedded tokens. + +Leveraging this insight, we introduce conditioned embedded tokens, a method that applies a structured correction term to the embedded token matrix, effectively reducing its condition number and, in turn, improving the conditioning of the self-attention matrix. While a full theoretical proof linking this improvement to better convergence in gradient-based optimization would require analyzing the NTK of a transformer—a notoriously difficult task—we provide strong empirical evidence that conditioned embedded tokens enhances performance across a range of transformer applications. + +Figure 1 illustrates how conditioned embedded tokens are integrated into a vision transformer. An image is first divided into tokens via patches, which are vectorized and passed through a learnable embedding layer, producing embedded tokens in a high-dimensional vector space. These tokens form the matrix $X = [X_{1}\dots X_{N}]^{T}$ . A correction term $C = [C_1\cdot \cdot \cdot C_N]^T$ is added to $X$ , reducing its condition number before the modified embedding matrix $X + C$ is fed into the first transformer layer (positional encoding not shown). This correction not only lowers the condition number of the self-attention matrix in the first layer but also has a cascading effect, improving conditioning in subsequent layers. In general, positional encodings are added to the tokens before they are fed into the first transformer layer; however, they are omitted in Figure 1 for simplicity. Our theory will show that the optimal correction term can be derived from the singular value decomposition (SVD) of $X$ . + +We validate our framework across various applications, including image classification, object detection, instance segmentation, and natural language processing. A key advantage of conditioned embedded tokens is their flexibility—they can be used with a wide range of attention mechanisms, making them compatible with modern transformer architectures. As we demonstrate, our methodology can serve as a drop-in enhancement for advanced attention mechanisms proposed in recent works [2, 4, 15, 22, 25, 26], offering a simple yet effective improvement to existing models. Our main contributions include: + +1. A theoretical framework introducing conditioned embedded tokens for improving the conditioning of the self-attention matrix within the first transformer layer yielding a more stable transformer architecture. +2. An empirical evaluation of conditioned embedded tokens across various transformer architectures and applications, including image classification, object detection, instance segmentation, and natural language processing, showing in each case its superior performance. + +# 2. Related Work + +Attention: Various strategies have been introduced to enhance the efficiency and effectiveness of transformers, particularly by addressing their computational overhead and re + +thinking attention mechanisms. The Data-Efficient Image Transformer (DeiT) [22] improves training efficiency in vision transformers by leveraging distillation tokens, allowing for competitive performance without requiring massive datasets. The Cross-Covariance Image Transformer (XCiT) [2] redefines attention by operating on cross-covariances of spatial features, enabling more efficient spatial interactions while reducing computational demands. The Nyströmformer [25], approximates standard self-attention using the Nyström method, reducing quadratic complexity to near-linear. This technique has proven particularly effective for long-range sequence modeling. In this paper, we demonstrate that our conditioning framework can be integrated as a drop-in replacement into these advanced attention mechanisms, yielding superior performance across multiple architectures. + +Conditioning: In [20], the authors investigate the condition number of weight matrices in feedforward neural networks and demonstrate that lower condition numbers lead to improved accuracy across various applications. They propose a preconditioner matrix that multiplies each layer's weight matrix, effectively reducing its condition number and enhancing the training of dense weights. In [14], neural network optimization is explored through the lens of the neural tangent kernel. The authors introduce the PL* condition, a variant of the Polyak-Lojasiewicz condition, and show that this approach improves the conditioning of the neural tangent kernel, which, in the infinite-width limit, governs the dynamics of gradient descent [10]. Finally, [1] demonstrates that increasing the depth of feedforward neural networks improves their conditioning leading to more efficient optimization. See [11, 12, 19, 20] for further results. + +# 3. Notation + +In this section, we formally define the transformer architecture by describing the structure of a transformer layer, along with establishing notation for various mathematical elements that will be referenced in subsequent sections. For further foundational details on transformers, readers may consult Vaswani et al. [23]. + +A layer in a transformer can be represented as a mapping + +$$ +\mathbf {T}: \mathbb {R} ^ {N \times D} \rightarrow \mathbb {R} ^ {N \times D} \tag {1} +$$ + +defined by + +$$ +\mathbf {T} (X) = \mathbf {F} (\mathbf {A} (X) + X), \tag {2} +$$ + +where $\mathbf{F}$ denotes a feed forward network with a residual connection, and $\mathbf{A}$ represents an attention mechanism. In general, layer normalization is added however for simplicity we omit layer normalization for this discussion. + +To begin with input data $x$ is tokenized into $N$ tokens yielding a tokenized representation $(x_{1},\dots ,x_{N})$ where + +$x_{i}\in \mathbb{R}^{t\times 1}$ . Each $x_{i}$ is then mapped to a high dimensional space $\mathbb{R}^{d\times 1}$ using a learnable embedding layer $E\in \mathbb{R}^{d\times t}$ so that + +$$ +E x _ {i} \in \mathbb {R} ^ {d \times 1}. \tag {3} +$$ + +The matrix $[Ex_1 \cdots Ex_n]^T \in \mathbb{R}^{N \times d}$ represents the embedded tokens associated to the input data. To capture the order of the tokens, a positional encoding matrix $P \in \mathbb{R}^{N \times d}$ is added forming $X = [Ex_1 \cdots Ex_n]^T + P \in \mathbb{R}^{N \times d}$ . The matrix $X$ is then input into the first layer of the transformer. For the theoretical Sec. 4 we will often use the notation $X$ to denote embedded tokens often leaving out explicit mention of the positional encoding part for ease of notation. + +The key component of the transformer is self-attention. This is composed of three learnable matrices, a query $(W_{Q})$ , key $(W_{K})$ , and value $(W_{V})$ , where $W_{Q}, W_{K} \in \mathbb{R}^{D \times d}$ and $W_{V} \in \mathbb{R}^{D \times d}$ . The output of the attention head $\mathbf{A}(X)$ is then given by + +$$ +\mathbf {A} (X) = \operatorname {s o f t m a x} \left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\right) X W _ {V}, \tag {4} +$$ + +where softmax is the softmax activation that acts row-wise on the matrix $XW_{Q}W_{k}^{T}X^{T}$ . Note that then $A(X) \in \mathbb{R}^{N \times d}$ . In the case no activation is used we have the definition of linear self-attention which we will denote by + +$$ +\mathbf {L A} (X) := X W _ {Q} W _ {k} ^ {T} X ^ {T} X W _ {V}. \tag {5} +$$ + +In general, multiple attention heads $\mathbf{A}_i$ for $1\leq i\leq h$ are utilized, each of dimension $N\times d_{i} = N\times \frac{d}{h}$ . These are then concatenated to produce a multi-head attention output, + +$$ +[ \mathbf {A} _ {1}, \dots , \mathbf {A} _ {h} ], \tag {6} +$$ + +which is subsequently fed into a feedforward layer. Similarly, a multi-head linear attention can be defined in the same way. The full transformer architecture is obtained by sequentially stacking several such transformer layers, as defined in Eq. (2). + +# 4. Theoretical Framework + +# 4.1. Conditioning of self-attention + +In this section, we analyze the conditioning of the self-attention matrix in a transformer. Our objective is to demonstrate that the condition number of the self-attention matrix in the first layer is influenced by the condition number of the embedded tokens. This insight motivates our key finding: reducing the condition number of the embedded tokens leads to a better-conditioned self-attention matrix in the first layer, ultimately improving the overall conditioning of the transformer. In this section, we omit positional encoding and layer normalization to keep the core theoretical insight clear and accessible. In Sec. 5, we will show our insights go through for a variety of practical transformers that use positional encoding and layer normalization. + +Definition 4.1. Let $A$ be an $N \times d$ matrix of full rank. The condition number of $A$ , denoted by $\kappa$ , is defined as + +$$ +\kappa (A) = \frac {\sigma_ {1}}{\sigma_ {k}} \tag {7} +$$ + +where $k = \min \{N, d\}$ , and $\sigma_1$ denotes the largest singular value of $A$ and $\sigma_k$ the smallest singular value of $A$ . Note that as $A$ is assumed full rank $\sigma_k \neq 0$ and the condition number is well-defined. Furthermore, observe that $\kappa(A) \geq 1$ since $\sigma_1 \geq \sigma_k$ . + +We have the following estimate for the condition number of linear self-attention and softmax self-attention. + +Proposition 4.2. Let $X$ denote an input for an attention layer as defined in Sec. 3. Assume $\mathbf{L}\mathbf{A}(X)$ and $\mathbf{A}(X)$ have full rank. We then have + +$$ +\begin{array}{l} \kappa (\mathbf {L A} (X)) \leq \kappa \left(W _ {Q}\right) \kappa \left(W _ {K}\right) \kappa \left(W _ {V}\right) \kappa (X) ^ {3} (8) \\ \kappa (\mathbf {A} (X)) \leq \kappa (\operatorname {s o f t m a x} \left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\right)) \kappa (X) \kappa \left(W _ {V}\right). (9) \\ \end{array} +$$ + +The proof of Proposition 4.2 is given in Sec. A of the Supp. material. + +Remark 4.3. The statement of Proposition 4.2 assumes the attention matrices are full rank. Since the layers of a transformer are randomly initialized we have that with probability 1 the attention matrices will be full rank at initialization. + +We point out that although Eq. (8) and Eq. (9) only provide an upper bound for the condition number, we found that the this upper bound provides a useful measure for estimating the condition number of the self-attention matrix. Therefore, for the self-attention matrices $\mathbf{A}(X)$ and $\mathbf{L}\mathbf{A}(X)$ we make the following definition + +$$ +\begin{array}{l} \mu (\mathbf {L A} (X)) := \kappa \left(W _ {Q}\right) \kappa \left(W _ {K}\right) \kappa \left(W _ {V}\right) \kappa (X) ^ {3}. (10) \\ \mu (\mathbf {A} (X)) := \kappa \left(\operatorname {s o f t m a x} \left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\right)\right) \kappa (X) \kappa \left(W _ {V}\right). (11) \\ \end{array} +$$ + +# 4.2. Conditioned embedded tokens + +In this section, we present our main theorem, which provides a method for reducing the condition number of embedded tokens derived from a dataset. This reduction subsequently decreases the measure $\mu (\mathbf{A}(X))$ , a key metric for assessing the condition number of the self-attention matrix in the first transformer layer. + +Theorem 4.4. Let $(x_{1},\dots ,x_{N})$ denote $N$ tokens associated to an input dataset with $x_{i}\in \mathbb{R}^{t\times 1}$ . Let $E\in \mathbb{R}^{d\times t}$ denote an embedding matrix so that $Ex_{i}\in \mathbb{R}^{d\times 1}$ . Let $X = [Ex_1\dots Ex_N]^T\in \mathbb{R}^{N\times d}$ denote the matrix of embedded tokens. Assume that $\kappa (X) > 2$ . Then there exists a matrix $C\in \mathbb{R}^{N\times d}$ such that + +$$ +\kappa (X + C) \leq 2 \tag {12} +$$ + +and hence that $\kappa(X + C) < \kappa(X)$ . + +In particular, Theorem 4.4 shows that if the condition number of the embedded tokens $X$ is extremely large. Adding a correction term given by $C$ will bring it down significantly. As we shall see in the experiments Sec. 5 the condition number $\kappa(X)$ is empirically extremely large. The proof of Theorem 4.4 is given in Sec. A of the Supp. material. + +Theorem 4.5. Let $X$ denote a collection of embedded tokens, as defined in the statement of Theorem 4.4, and let $C$ denote the matrix given by Theorem 4.4. We then have + +$$ +\mu (\mathbf {L A} (X + C)) \leq \mu (\mathbf {L A} (X)). \tag {13} +$$ + +Assume that + +$$ +\begin{array}{l} \kappa (\text {s o f t m a x} ((X + C) W _ {Q} W _ {K} ^ {T} (X ^ {T} + C ^ {T}))) \\ \leq \kappa (\operatorname {s o f t m a x} \left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\right)) \tag {14} \\ \end{array} +$$ + +then + +$$ +\mu (\mathbf {A} (X + C)) \leq \mu (\mathbf {A} (X)) \tag {15} +$$ + +where $\mu (\mathbf{L}\mathbf{A}(X))$ is defined by Eq. (10) and $\mu (\mathbf{A}(X))$ by Eq. (11). + +The proof of Theorem 4.5 is given in Sec. A of the Supp. material. A natural question arises: can self-attention be conditioned directly through queries, keys, and values, independent of the embedded tokens? Unfortunately, our approach of introducing a correction term does not extend to these components. This is because our correction modifies the singular values of the embedded tokens, whereas altering the singular values of queries, keys, or values could disrupt how self-attention captures spatial relationships between tokens. + +Remark 4.6. We note that in the case of self-attention with a softmax activation Theorem 4.5 requires an assumption on the condition number of the non-linear self-attention probability matrix term softmax $((X + C)W_{Q}W_{K}^{T}(X^{T} + C^{T}))$ . Although this may be seen as a limitation of the theory we will see in Sec. 5 that this framework still provides a useful methodology to obtain good performance on a variety of transformer applications. + +Remark 4.7. Theorem 4.5 demonstrates that conditioning embedded tokens can improve the condition number of the self-attention matrix in the first layer. However, it does not address how this affects the condition number in subsequent layers. In Sec. 5, we will show that this effect propagates through the network, leading to a reduced condition number in deeper self-attention layers. + +Remark 4.8. Our theory, as presented in Theorem 4.4 and Theorem 4.5, demonstrates that conditioning embedded tokens reduces the condition number of the self-attention matrix in the first layer. However, the framework does not + +
Models
ViT-baseDeiT-baseSwin-baseXciT-mediumDaViT-base
Original embedded tokens80.381.683.182.283.6
Conditioned embedded tokens81.382.583.982.984.6
+ +Table 1. Comparison of vision transformers with their original embedded tokens verse one with a conditioned conditioned embedded tokens pre-trained on the ImageNet-1k dataset. We report the classification top-1% accuracy. In each case we see the transformers employing conditioned embedded tokens (ours) outperforms the original ones. + +theoretically establish how this improvement translates to better optimization and, consequently, higher accuracy in applications. Nonetheless, in Sec. 5, we provide empirical evidence showing that our approach consistently enhances accuracy across various applications, highlighting its practical effectiveness. + +# 5. Experiments + +In this section, we evaluate our insight from Sec. 4 on a variety of transformer applications. For a detailed analysis of how we implemented conditioned embedded tokens see Sec. A in the Supp. material. + +# 5.1. Image Classification + +In this section, we apply our theoretical insights from Sec. 4 to vision transformers on an image classification task. We will train all vision transformers from scratch on the ImageNet-1k dataset. We will be testing on the following vision transformers. In each case we will compare the original architecture with one that uses conditioned embedded tokens. + +ViT [5]: ViT is a pioneering architecture that applies transformer-based processing to images by treating them as sequences of non-overlapping patches. In our study, we utilize ViT-Base (ViT-B), which is configured with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. The model employs a standard self-attention mechanism to capture dependencies across patches. + +DeiT [22]: DeiT builds upon ViT but is optimized for data efficiency. Unlike standard ViT, it employs a data-starvation training strategy [22], enabling faster convergence while maintaining strong performance. In our setup, we use DeiT-Base (DeiT-B) with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. Like ViT, DeiT utilizes a self-attention mechanism for feature extraction. + +Swin Transformer [15]: Swin Transformer introduces a hierarchical architecture and a novel shifted window-based self-attention mechanism, significantly improving efficiency and effectiveness in vision tasks. In our experiments, we adopt the Swin-Base (Swin-B) variant, which features 128 channels in the hidden layers of the first stage. + +![](images/4ada79ab080496b73870e69728c9d70f45572144b0a9e3e91857f5ec62f1cbeb.jpg) +Figure 2. Condition number comparison for ViT-B on ImageNet-1k. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over training epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original. + +The attention operates over non-overlapping windows of size $M = 7$ , while the query dimension per head is set to $d = 32$ . The network follows a hierarchical structure with layers arranged as $\{2, 2, 18, 2\}$ across different stages. + +XCiT [2]: XCiT diverges from standard ViT architectures by incorporating two key innovations. First, it integrates Local Patch Interaction within each block, using depth-wise $3 \times 3$ convolutions, Batch Normalization, GELU activation, and an additional depth-wise convolution to enhance local feature representation. Second, it employs Cross-Covariance Attention, where attention maps are derived from the cross-covariance matrix of the key and query projections. For our experiments, we utilize the XCiT-M24 (XCiT-M) model with a patch size of 16, an embedding dimension of 512, 8 attention heads, and 24 layers. + +DaViT [4]: DaViT extends traditional vision transformers by incorporating both spatial and channel attention mechanisms, enhancing feature extraction across multiple dimensions. While conventional transformers rely solely on spatial self-attention, DaViT introduces channel attention to capture dependencies across feature channels, complement + +
ModelAPbAPb50APb75APmAPm50APm75
XCiT-S + original embedded tokens44.966.148.940.163.142.8
XCiT-S + conditioned embedded tokens45.766.649.740.463.543.4
XCiT-M + original embedded tokens45.766.849.640.863.643.3
XCiT-M + conditioned embedded tokens46.267.449.841.463.844.2
+ +Table 2. Performance evaluation of object detection and instance segmentation on the COCO mini-val set. The backbone networks are pretrained on ImageNet-1k and integrated into the Mask R-CNN framework. The reported metrics include $AP^b$ (Average Precision for bounding box predictions), $AP_{50/75}^b$ (Average Precision at IoU thresholds of 0.50 and 0.75 for bounding boxes), $AP^m$ (Average Precision for mask predictions), and $AP_{50/75}^m$ (Average Precision at IoU thresholds of 0.50 and 0.75 for mask predictions). In each case we see the XCiT architecture employing conditioned embedded tokens (ours) out performs the original architecture. + +ing spatial interactions. This hybrid approach allows DaViT to effectively model both local and global relationships. In our experiments, we use the DaViT-Base (DaViT-B) model, configured with a patch size of 4, embedding dimensions of (128, 256, 512, 1024), attention heads of (4, 8, 16, 32), and a layer distribution of (1, 1, 9, 1). + +Results: Table 1 presents a comparison of the top- $1\%$ test accuracy of five vision transformers trained from scratch on the ImageNet-1k dataset. Each model is evaluated using both the original embedded tokens and our conditioned embedded tokens, following the approach outlined in Sec. 4.2. While ViT-Base [5] and DeiT-Base [22] utilize standard self-attention mechanisms, Swin-Base [15], XCiT-Medium [2], and DaViT-Base [4] incorporate more advanced attention mechanisms. We include these latter models to illustrate that, although our theoretical analysis focuses on self-attention, conditioned embedded tokens can be easily integrated into modern transformers with more sophisticated architectures. As shown in the table, models using conditioned embedded tokens consistently achieve higher accuracy. Full implementation details and hyperparameter settings can be found in Sec. B.2 of the Supp. material. + +Analysis: Figure 2 illustrates the condition numbers of the embedded tokens, the first-layer self-attention matrix (averaged across all heads), and the overall self-attention matrix (averaged across all layers) for the ViT-B architecture. We compare these metrics for models using the original embedded tokens and our conditioned embedded tokens. These values were computed at each training epoch and then averaged over all epochs to generate the plots. The figure clearly shows that the ViT-B architecture with conditioned embedded tokens maintains on average a significantly lower condition number throughout training. + +Implementation and hardware: For full details on the implementation of each vision transformer, the training hyperparameters and the hardware used please see Sec. B.2 in the Supp. material. + +# 5.2. Object Detection and Image Segmentation + +In this section, we extend our approach from Sec. 4.2 to object detection and instance segmentation tasks. To evaluate its effectiveness in fine-tuning scenarios, we adapt a pretrained XCiT model, originally trained on ImageNet-1k, for these tasks. Our experiments leverage the COCO 2017 dataset [13], which consists of 118K training images and 5K validation images spanning 80 object categories. For model architecture, we employ XCiT as the backbone of the Mask R-CNN framework [9], incorporating a Feature Pyramid Network (FPN) to enhance multi-scale feature representation. To integrate XCiT with FPN, we modify its inherently columnar structure by extracting features from multiple layers. We do this by using an XCiT-small that has 12 layers (XCiT-S) and a XCiT-medium (XCiT-M) that has 24 layers. These features, originally computed at a fixed stride of 16, are adjusted to strides of [4, 8, 16, 32] to align with the FPN hierarchy. Max pooling is used for downsampling, while a single transposed convolution layer facilitates upsampling. The model is then fine-tuned for 36 epochs using the AdamW optimizer, with a learning rate of $10^{-4}$ , a weight decay of 0.05, and a batch size of 16. This training strategy effectively demonstrates how XCiT's learned representations can be adapted for downstream tasks. + +Results: We pretrained a total of four models on the ImageNet-1k dataset, namely an XCiT-S and XCiT-M both using the original embedded tokens following the methodology from Sec. 4.2. The results of the experiment are shown in Tab. 2. As can be seen from the table in both cases the transformer using conditioned embedded tokens out performs the original. + +Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used see Sec. B.3 in the Supp. material. + +# 5.3. Language Models + +In this section, we evaluate our insights from Sec. 4.2 on two different language models. + +
MNLISST-2STSBRTEQNLIQQPMRPCCoLAGLUE
Crammed BERT (original)83.892.386.355.190.187.385.048.978.6
Crammed BERT (ours)84.292.586.555.691.187.486.353.779.7
+ +Crammed BERT: In the first experiment we applied our insights from Sec. 4.2 to a Crammed BERT language model [8], trained entirely from scratch using masked language modeling with conditioned embedded tokens. The model consists of 110 million parameters, with 12 transformer layers and 12 attention heads per layer. We train two versions from scratch: the original Crammed BERT [8] and a variant incorporating conditioned embedded tokens using the insight from Sec. 4.2. Both models are trained on The Pile dataset [7], a large-scale corpus designed for language model training following the pretraining regime in [8]. After pretraining, we evaluate the performance on the GLUE benchmark [24] following the evaluation methodology outlined in [8]. + +Results: The evaluation results are shown in Tab. 3 with each number representing the accuracy. We see from the table that in each task the model employing conditioned embedded tokens outperforms the original Crammed Bert model on every task within the GLUE benchmark. + +GPT-2: We conducted a second experiment training a GPT-2 model with a decoder-only transformer architecture, consisting of 12 layers, 12 attention heads per layer, and an embedding dimension of 512. The model was trained using masked self-attention to predict the next token, leveraging the TinyStories dataset [6] for language generation. Two versions of the model were trained from scratch: one using the original embedded tokens as in [6] and another incorporating the conditioned embedded tokens from Sec. 4.2. Both models were evaluated using perplexity loss (PPL), a widely used metric for assessing language models on datasets like TinyStories. Perplexity measures the model's confidence in its predictions by computing the exponential of the average negative log-likelihood of the predicted tokens. A lower perplexity score indicates that the model assigns higher probabilities to correct tokens, leading to better fluency and coherence in text generation. + +Results and Analysis: Tab. 4 presents the results of our experiment, demonstrating that the model trained with conditioned embedded tokens consistently achieves lower validation loss and lower perplexity compared to the original GPT-2 model. These results indicate that conditioning the embedded tokens improves model performance, leading to more accurate language generation. Fig. 3 plots the con + +Table 3. Evaluation of a pre-trained Crammed BERT on the GLUE benchmark with the original embedded tokens (original) and our conditioned embedded tokens (ours). As can be seen from the table our methodology out performs the original in every task. + +
Val loss
GPT-2 original2.41
GPT-2 conditioned (ours)2.36
+ +Table 4. GPT-2 model trained on the TinyStories dataset. We compare two models, the original model and one using conditioned embedded tokens (conditioned). We report the validation loss; in both cases, our model yields a lower loss, indicating better performance. + +![](images/387be26c555014682eeea3fa4309c09c84fde3390f2c2c79465566c317baac9d.jpg) +Figure 3. Condition number comparison for GPT-2 training on TinyStories. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over all epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original. + +dition numbers of the embedded tokens, self-attention in first layer averaged over all heads and self-attention averaged over all layers. Each of these are evaluated over each training epoch and then averaged over all epochs. + +Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.4 in the Supp. material. + +# 5.4. Nyströmformer for Long Range Sequences + +We applied our methodology to the Nyströmformer [25], a transformer architecture designed for efficient long-range dependency modeling on the Long-Range Arena (LRA) benchmark [21]. We trained two Nyströmformer models: + +
ModelListOpsTextRetrievalImagePathfinder
Nyströmformer (original)37.163.879.839.972.9
Nyströmformer conditioned (ours)37.964.980.940.180.0
+ +Table 5. Comparison of a Nyströmformer using its original embedded tokens (original) with a Nyströmformer using conditioned embedded tokens (ours) on the LRA benchmark. We report the evaluation accuracy (\%). As can be seen from the table our methodology yields a higher accuracy on each task. + +one utilizing conditioned embedded tokens and the original Nyströmformer for comparison. The Long-Range Arena (LRA) benchmark consists of five distinct tasks: + +1. ListOps: A synthetic task that tests hierarchical reasoning by evaluating nested arithmetic operations. +2. Text Classification: Character-level text classification, requiring models to process long sequences of text. +3. Retrieval: A document retrieval task that measures the model's ability to identify relevant patterns within long textual inputs. +4. Image Classification: Modeling image data as a sequence of flattened pixel values to assess a model's capability in visual recognition. +5. Pathfinder: A visual reasoning challenge that requires detecting connected paths in an image, emphasizing spatial relationship modeling. + +Results and Analysis: Tab. 5 shows the results of the experiment. As can be seen from the table the Nyströmformer that employed conditioned embedded tokens outperformed the original Nyströmformer from [25] in every task in the LRA benchmark. We computed the condition numbers of the original Nyströmformer and one employing conditioned embedded tokens throughout training for the embedded tokens, the self-attention within the first layer averaged over all heads and the self-attention averaged over all layers. We then took the average for each task over all training epochs and averaged over the five tasks within the LRA benchmark. The plots are shown in Fig. 4. As can be seen from the figure, the Nyströmformer employing conditioned embedded tokens has lower condition number in all cases. + +Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.5 in the Supp. material. + +# 6. Limitations + +Our work explored the conditioning of transformer attention blocks through embedded tokens. While we showed that conditioned embedded tokens reduce the first-layer self-attention condition number, their impact on optimization remains theoretically unproven. Prior studies [1, 10, 14] have linked NTK conditioning to optimization in feedforward networks, raising an important question: how does + +![](images/509eeaef0fd0de7359dd3a4183512e40b1e73296a74c1eeb13a9f0437d909d39.jpg) +Figure 4. Condition number comparison for Nyströmformer on the LRA benchmark. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, averaged over training epochs and tasks. Our model (conditioned) consistently achieves a significantly lower condition number than the original. + +conditioning embedded tokens affect the NTK in transformers? Investigating this could offer deeper insights into how embedded token structure influences transformer optimization. + +# 7. Conclusion + +In this paper, we introduced a framework for analyzing the conditioning of the self-attention matrix in the first layer of a transformer, highlighting its dependence on the embedded tokens. Our analysis demonstrated that by adding a carefully designed correction term—which we refer to as conditioned embedded tokens—we could significantly reduce the condition number of self-attention in the first layer, leading to a better-conditioned attention mechanism. We empirically validated the effectiveness of conditioned embedded tokens across a diverse set of tasks, including image classification, object detection, instance segmentation, language modeling, and long-range sequence modeling. In all cases, our approach proved to be a simple, drop-in replacement for existing embedding methods, consistently improving model performance across these applications. + +# References + +[1] Naman Agarwal, Pranjal Awasthi, and Satyen Kale. A deep conditioning treatment of neural networks. In Algorithmic Learning Theory, pages 249-305. PMLR, 2021. 3, 8 +[2] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 3, 5, 6 +[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2 +[4] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In European conference on computer vision, pages 74-92. Springer, 2022. 2, 5, 6 +[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5, 6 +[6] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759, 2023. 7 +[7] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Aadi Thite, Eric Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2021. 7 +[8] Jonas Geiping and Tom Goldstein. Cramming: Training a language model on a singlegpu in one day. In International Conference on Machine Learning, pages 11117-11143. PMLR, 2023. 7 +[9] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6 +[10] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3, 8 +[11] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. arXiv preprint arXiv:2403.19243, 2024. 3 +[12] Yiping Ji, Hemanth Saratchandran, Peyman Moghaddam, and Simon Lucey. Always skip attention. arXiv preprint arXiv:2505.01996, 2025.3 +[13] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6 + +[14] Chaoyue Liu, Libin Zhu, and Mikhail Belkin. Loss landscapes and optimization in over-parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59:85-116, 2022. 2, 3, 8 +[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6 +[16] Abhisek Maiti, Sander Oude Elberink, and George Vosselman. Transfusion: Multi-modal fusion network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6536-6546, 2023. 2 +[17] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 2 +[18] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Multi-agent generative trajectory forecasting with heterogeneous data for control. arXiv preprint arXiv:2001.03093, 2, 2020. 2 +[19] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking attention: Polynomial alternatives to softmax in transformers. arXiv preprint arXiv:2410.18613, 2024. 3 +[20] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2025. 2, 3 +[21] Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. Long range arena: A benchmark for efficient transformers. arXiv preprint arXiv:2011.04006, 2020. 7 +[22] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6 +[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3 +[24] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 7 +[25] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14138–14148, 2021. 2, 3, 7, 8 +[26] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 2 +[27] Q Zhen, W Sun, H Deng, D Li, Y Wei, B Lv, J Yan, L Kong, and Y Zhong. cosformer: rethinking softmax in attention. + +In International Conference on Learning Representations, 2022. 2 +[28] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized bert pre-training approach with post-training. In Proceedings of the 20th chinese national conference on computational linguistics, pages 1218-1227, 2021. 2 \ No newline at end of file diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/images.zip b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..c8eaf05e7ac1c97e104f6c7c2f9b02875e227ac0 --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ffcd56c6d155c16645ee10db1a72a4729e673376523c0ec1d42695ad513118ef +size 304278 diff --git a/2025/Enhancing Transformers Through Conditioned Embedded Tokens/layout.json b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..968fbe9f2b10a03a300a4f258206af6bc99bce62 --- /dev/null +++ b/2025/Enhancing Transformers Through Conditioned Embedded Tokens/layout.json @@ -0,0 +1,7684 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "spans": [ + { + "bbox": [ + 100, + 103, + 510, + 121 + ], + "type": "text", + "content": "Enhancing Transformers Through Conditioned Embedded Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 89, + 144, + 293, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 144, + 293, + 185 + ], + "spans": [ + { + "bbox": [ + 89, + 144, + 293, + 185 + ], + "type": "text", + "content": "Hemanth Saratchandran \nAustralian Institute for Machine Learning \nAdelaide University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 188, + 292, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 188, + 292, + 198 + ], + "spans": [ + { + "bbox": [ + 89, + 188, + 292, + 198 + ], + "type": "text", + "content": "hemanth.saratchandran@adelaide.edu.au" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 318, + 144, + 520, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 144, + 520, + 186 + ], + "spans": [ + { + "bbox": [ + 318, + 144, + 520, + 186 + ], + "type": "text", + "content": "Simon Lucey \nAustralian Institute for Machine Learning \nAdelaide University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 345, + 188, + 492, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 188, + 492, + 198 + ], + "spans": [ + { + "bbox": [ + 345, + 188, + 492, + 198 + ], + "type": "text", + "content": "simon.lucey@adelaide.edu.au" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 80, + 227, + 521, + 507 + ], + "blocks": [ + { + "bbox": [ + 80, + 227, + 521, + 507 + ], + "lines": [ + { + "bbox": [ + 80, + 227, + 521, + 507 + ], + "spans": [ + { + "bbox": [ + 80, + 227, + 521, + 507 + ], + "type": "image", + "image_path": "60619f460852316481314c0f675e2b59c233c40ff87cbb2b669a8ea26eb4c00e.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "lines": [ + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "spans": [ + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": "Figure 1. Schematic representation of conditioned embeddings for a vision transformer: An image is divided into " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": " patches, with each patch then embedded as a high-dimensional vector " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "X_{i}\\in \\mathbb{R}^{d}" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "1\\leq i\\leq N" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": ". These vectors are then concatenated to form the embedded token matrix " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "X = [X_{1}\\dots X_{N}]^{T}" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": ". To improve the conditioning of the self-attention mechanism, a correction term " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "C = [C_1\\dots C_N]^T" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 515, + 555, + 572 + ], + "type": "text", + "content": ", reducing its condition number. The modified matrix is then fed into the first layer of the transformer (positional encoding not shown). Its effect on the self-attention equation is illustrated in the green equation (for simplicity, layer normalization has been omitted)." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 581, + 200, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 581, + 200, + 594 + ], + "spans": [ + { + "bbox": [ + 151, + 581, + 200, + 594 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 617, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 297, + 713 + ], + "type": "text", + "content": "Transformers have transformed modern machine learning, driving breakthroughs in computer vision, natural language processing, and robotics. At the core of their success lies the attention mechanism, which enables the modeling of global dependencies among input tokens. However, we reveal that the attention block in transformers suffers from inherent ill-conditioning, which hampers gradient-based optimization and leads to inefficient training. To address this, we de" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 556, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 703 + ], + "type": "text", + "content": "velop a theoretical framework that establishes a direct relationship between the conditioning of the attention block and that of the embedded tokenized data. Building on this insight, we introduce conditioned embedded tokens, a method that systematically modifies the embedded tokens to improve the conditioning of the attention mechanism. Our analysis demonstrates that this approach significantly mitigates ill-conditioning, leading to more stable and efficient training. We validate our methodology across various transformer architectures, achieving consistent improvements in image" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4786" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "classification, object detection, instance segmentation, and natural language processing, highlighting its broad applicability and effectiveness." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 131, + 136, + 144 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 151, + 296, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 296, + 319 + ], + "type": "text", + "content": "The transformer architecture [23] has become a cornerstone of modern machine learning, driving breakthroughs across various domains including computer vision [3, 5, 15, 22], natural language processing [23, 27, 28], and robotics [16, 18]. Its success is largely attributed to its ability to model complex relationships in data without relying on traditional recurrent or convolutional structures. At the heart of the transformer is the self-attention mechanism, which dynamically assigns importance to different parts of the input, capturing both local and global dependencies. Unlike traditional architectures that process inputs in a fixed order, transformers utilize self-attention to integrate global context at each layer, making them highly effective for tasks that require contextual understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 426 + ], + "type": "text", + "content": "In general, transformers process input data by first converting it into discrete tokens, which are then mapped to a high-dimensional vector space through an embedding layer. For text, these tokens represent words or sub-words, while for images, image patches are treated as tokens. The resulting embedded tokens form a matrix that is passed into the first transformer layer for self-attention computations. This process allows the model to effectively learn relationships between tokens in a high-dimensional space." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 296, + 521 + ], + "type": "text", + "content": "In this paper, we explore the conditioning of the self-attention matrix in transformer architectures—an essential yet often overlooked factor influencing optimization dynamics. The conditioning of a matrix is measured by its condition number, defined as the ratio of its largest singular value to its smallest. A high condition number indicates ill-conditioning, which is a well-known challenge for the convergence of gradient-based optimization methods [17]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 522, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 653 + ], + "type": "text", + "content": "Previous work has addressed conditioning in feedforward neural networks. In [20], a methodology was introduced to condition network weights, leading to improved optimization. Similarly, [14] examined the condition of the neural tangent kernel (NTK), proposing a framework for reducing its condition number and demonstrating enhanced convergence for gradient descent. While these studies have focused on conditioning in feedforward architectures, its role in self-attention mechanisms remains largely unexplored. This gap motivates our investigation into how conditioning affects optimization in transformers." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "We develop a theoretical framework to analyze the condition number of the self-attention matrix within a transformer's attention layer. Our analysis establishes a direct relationship between the condition number of the first-layer self-attention matrix and that of the embedded tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 203 + ], + "type": "text", + "content": "Leveraging this insight, we introduce conditioned embedded tokens, a method that applies a structured correction term to the embedded token matrix, effectively reducing its condition number and, in turn, improving the conditioning of the self-attention matrix. While a full theoretical proof linking this improvement to better convergence in gradient-based optimization would require analyzing the NTK of a transformer—a notoriously difficult task—we provide strong empirical evidence that conditioned embedded tokens enhances performance across a range of transformer applications." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": "Figure 1 illustrates how conditioned embedded tokens are integrated into a vision transformer. An image is first divided into tokens via patches, which are vectorized and passed through a learnable embedding layer, producing embedded tokens in a high-dimensional vector space. These tokens form the matrix " + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "inline_equation", + "content": "X = [X_{1}\\dots X_{N}]^{T}" + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": ". A correction term " + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "inline_equation", + "content": "C = [C_1\\cdot \\cdot \\cdot C_N]^T" + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": " is added to " + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": ", reducing its condition number before the modified embedding matrix " + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "inline_equation", + "content": "X + C" + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": " is fed into the first transformer layer (positional encoding not shown). This correction not only lowers the condition number of the self-attention matrix in the first layer but also has a cascading effect, improving conditioning in subsequent layers. In general, positional encodings are added to the tokens before they are fed into the first transformer layer; however, they are omitted in Figure 1 for simplicity. Our theory will show that the optimal correction term can be derived from the singular value decomposition (SVD) of " + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 204, + 553, + 407 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 407, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 538 + ], + "type": "text", + "content": "We validate our framework across various applications, including image classification, object detection, instance segmentation, and natural language processing. A key advantage of conditioned embedded tokens is their flexibility—they can be used with a wide range of attention mechanisms, making them compatible with modern transformer architectures. As we demonstrate, our methodology can serve as a drop-in enhancement for advanced attention mechanisms proposed in recent works [2, 4, 15, 22, 25, 26], offering a simple yet effective improvement to existing models. Our main contributions include:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 310, + 539, + 553, + 645 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 311, + 539, + 553, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 311, + 539, + 553, + 585 + ], + "spans": [ + { + "bbox": [ + 311, + 539, + 553, + 585 + ], + "type": "text", + "content": "1. A theoretical framework introducing conditioned embedded tokens for improving the conditioning of the self-attention matrix within the first transformer layer yielding a more stable transformer architecture." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 310, + 586, + 553, + 645 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 310, + 586, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 310, + 586, + 553, + 645 + ], + "type": "text", + "content": "2. An empirical evaluation of conditioned embedded tokens across various transformer architectures and applications, including image classification, object detection, instance segmentation, and natural language processing, showing in each case its superior performance." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 400, + 670 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Attention: Various strategies have been introduced to enhance the efficiency and effectiveness of transformers, particularly by addressing their computational overhead and re" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4787" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 262 + ], + "type": "text", + "content": "thinking attention mechanisms. The Data-Efficient Image Transformer (DeiT) [22] improves training efficiency in vision transformers by leveraging distillation tokens, allowing for competitive performance without requiring massive datasets. The Cross-Covariance Image Transformer (XCiT) [2] redefines attention by operating on cross-covariances of spatial features, enabling more efficient spatial interactions while reducing computational demands. The Nyströmformer [25], approximates standard self-attention using the Nyström method, reducing quadratic complexity to near-linear. This technique has proven particularly effective for long-range sequence modeling. In this paper, we demonstrate that our conditioning framework can be integrated as a drop-in replacement into these advanced attention mechanisms, yielding superior performance across multiple architectures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 263, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 466 + ], + "type": "text", + "content": "Conditioning: In [20], the authors investigate the condition number of weight matrices in feedforward neural networks and demonstrate that lower condition numbers lead to improved accuracy across various applications. They propose a preconditioner matrix that multiplies each layer's weight matrix, effectively reducing its condition number and enhancing the training of dense weights. In [14], neural network optimization is explored through the lens of the neural tangent kernel. The authors introduce the PL* condition, a variant of the Polyak-Lojasiewicz condition, and show that this approach improves the conditioning of the neural tangent kernel, which, in the infinite-width limit, governs the dynamics of gradient descent [10]. Finally, [1] demonstrates that increasing the depth of feedforward neural networks improves their conditioning leading to more efficient optimization. See [11, 12, 19, 20] for further results." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 477, + 117, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 117, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 117, + 490 + ], + "type": "text", + "content": "3. Notation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 568 + ], + "type": "text", + "content": "In this section, we formally define the transformer architecture by describing the structure of a transformer layer, along with establishing notation for various mathematical elements that will be referenced in subsequent sections. For further foundational details on transformers, readers may consult Vaswani et al. [23]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 570, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 570, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 67, + 570, + 295, + 582 + ], + "type": "text", + "content": "A layer in a transformer can be represented as a mapping" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 130, + 589, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 589, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 130, + 589, + 295, + 602 + ], + "type": "interline_equation", + "content": "\\mathbf {T}: \\mathbb {R} ^ {N \\times D} \\rightarrow \\mathbb {R} ^ {N \\times D} \\tag {1}", + "image_path": "f039962353674b118ebf4cb5b53789773f9053307d48fc88b63fe453feb36058.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 612, + 101, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 612, + 101, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 612, + 101, + 624 + ], + "type": "text", + "content": "defined by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 624, + 295, + 636 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 624, + 295, + 636 + ], + "spans": [ + { + "bbox": [ + 123, + 624, + 295, + 636 + ], + "type": "interline_equation", + "content": "\\mathbf {T} (X) = \\mathbf {F} (\\mathbf {A} (X) + X), \\tag {2}", + "image_path": "cbe4023c28a798ef4120c6488ae751ecc18ef4a5b9a76d8954265dedfdeef234.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " denotes a feed forward network with a residual connection, and " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{A}" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " represents an attention mechanism. In general, layer normalization is added however for simplicity we omit layer normalization for this discussion." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "To begin with input data " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is tokenized into " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " tokens yielding a tokenized representation " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "(x_{1},\\dots ,x_{N})" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " where" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathbb{R}^{t\\times 1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "content": ". Each " + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "content": " is then mapped to a high dimensional space " + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{d\\times 1}" + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "content": " using a learnable embedding layer " + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "inline_equation", + "content": "E\\in \\mathbb{R}^{d\\times t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 106 + ], + "type": "text", + "content": " so that" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 406, + 107, + 553, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 406, + 107, + 553, + 121 + ], + "spans": [ + { + "bbox": [ + 406, + 107, + 553, + 121 + ], + "type": "interline_equation", + "content": "E x _ {i} \\in \\mathbb {R} ^ {d \\times 1}. \\tag {3}", + "image_path": "4924e590e0112e7f40659d14b2ad4cb4d515b51d570e2b1ec29896bf53ad4060.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": "The matrix " + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "inline_equation", + "content": "[Ex_1 \\cdots Ex_n]^T \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": " represents the embedded tokens associated to the input data. To capture the order of the tokens, a positional encoding matrix " + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "inline_equation", + "content": "P \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": " is added forming " + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "inline_equation", + "content": "X = [Ex_1 \\cdots Ex_n]^T + P \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": ". The matrix " + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": " is then input into the first layer of the transformer. For the theoretical Sec. 4 we will often use the notation " + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 125, + 555, + 220 + ], + "type": "text", + "content": " to denote embedded tokens often leaving out explicit mention of the positional encoding part for ease of notation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": "The key component of the transformer is self-attention. This is composed of three learnable matrices, a query " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "(W_{Q})" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": ", key " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "(W_{K})" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": ", and value " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "(W_{V})" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K} \\in \\mathbb{R}^{D \\times d}" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "W_{V} \\in \\mathbb{R}^{D \\times d}" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": ". The output of the attention head " + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(X)" + }, + { + "bbox": [ + 313, + 221, + 554, + 281 + ], + "type": "text", + "content": " is then given by" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 347, + 287, + 553, + 302 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 287, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 347, + 287, + 553, + 302 + ], + "type": "interline_equation", + "content": "\\mathbf {A} (X) = \\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right) X W _ {V}, \\tag {4}", + "image_path": "c66ebee9114a19950cfa5d239946261dee24cd91330f5d33aa0ce4cf667d63c7.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "text", + "content": "where softmax is the softmax activation that acts row-wise on the matrix " + }, + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "inline_equation", + "content": "XW_{Q}W_{k}^{T}X^{T}" + }, + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "text", + "content": ". Note that then " + }, + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "inline_equation", + "content": "A(X) \\in \\mathbb{R}^{N \\times d}" + }, + { + "bbox": [ + 313, + 309, + 554, + 357 + ], + "type": "text", + "content": ". In the case no activation is used we have the definition of linear self-attention which we will denote by" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 365, + 364, + 553, + 380 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 364, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 365, + 364, + 553, + 380 + ], + "type": "interline_equation", + "content": "\\mathbf {L A} (X) := X W _ {Q} W _ {k} ^ {T} X ^ {T} X W _ {V}. \\tag {5}", + "image_path": "384d17564aa0ca797c1fdeb217c4fcbd2485a6404902d26548f74614ca64437e.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "text", + "content": "In general, multiple attention heads " + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "inline_equation", + "content": "\\mathbf{A}_i" + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "inline_equation", + "content": "1\\leq i\\leq h" + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "text", + "content": " are utilized, each of dimension " + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "inline_equation", + "content": "N\\times d_{i} = N\\times \\frac{d}{h}" + }, + { + "bbox": [ + 313, + 386, + 554, + 422 + ], + "type": "text", + "content": ". These are then concatenated to produce a multi-head attention output," + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 403, + 430, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 430, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 403, + 430, + 553, + 444 + ], + "type": "interline_equation", + "content": "[ \\mathbf {A} _ {1}, \\dots , \\mathbf {A} _ {h} ], \\tag {6}", + "image_path": "9cf3596fb7e8db665483223f847c96c251f9419ca01169e99a2f859d9cfdcab9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 451, + 554, + 511 + ], + "type": "text", + "content": "which is subsequently fed into a feedforward layer. Similarly, a multi-head linear attention can be defined in the same way. The full transformer architecture is obtained by sequentially stacking several such transformer layers, as defined in Eq. (2)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 521, + 451, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 451, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 451, + 533 + ], + "type": "text", + "content": "4. Theoretical Framework" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 540, + 474, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 474, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 474, + 553 + ], + "type": "text", + "content": "4.1. Conditioning of self-attention" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 713 + ], + "type": "text", + "content": "In this section, we analyze the conditioning of the self-attention matrix in a transformer. Our objective is to demonstrate that the condition number of the self-attention matrix in the first layer is influenced by the condition number of the embedded tokens. This insight motivates our key finding: reducing the condition number of the embedded tokens leads to a better-conditioned self-attention matrix in the first layer, ultimately improving the overall conditioning of the transformer. In this section, we omit positional encoding and layer normalization to keep the core theoretical insight clear and accessible. In Sec. 5, we will show our insights go through for a variety of practical transformers that use positional encoding and layer normalization." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4788" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "Definition 4.1. Let " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " be an " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "N \\times d" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": " matrix of full rank. The condition number of " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", denoted by " + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "inline_equation", + "content": "\\kappa" + }, + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": ", is defined as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 100, + 295, + 124 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 100, + 295, + 124 + ], + "spans": [ + { + "bbox": [ + 149, + 100, + 295, + 124 + ], + "type": "interline_equation", + "content": "\\kappa (A) = \\frac {\\sigma_ {1}}{\\sigma_ {k}} \\tag {7}", + "image_path": "ebac83ec5086743bef787479fdc600e8366ac679af3b264c1091265e75d6100e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "spans": [ + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "k = \\min \\{N, d\\}" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\sigma_1" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " denotes the largest singular value of " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\sigma_k" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " the smallest singular value of " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": ". Note that as " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " is assumed full rank " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\sigma_k \\neq 0" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " and the condition number is well-defined. Furthermore, observe that " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\kappa(A) \\geq 1" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": " since " + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "inline_equation", + "content": "\\sigma_1 \\geq \\sigma_k" + }, + { + "bbox": [ + 55, + 128, + 296, + 188 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 193, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 217 + ], + "type": "text", + "content": "We have the following estimate for the condition number of linear self-attention and softmax self-attention." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "text", + "content": "Proposition 4.2. Let " + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "text", + "content": " denote an input for an attention layer as defined in Sec. 3. Assume " + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{L}\\mathbf{A}(X)" + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(X)" + }, + { + "bbox": [ + 54, + 221, + 296, + 258 + ], + "type": "text", + "content": " have full rank. We then have" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 262, + 295, + 304 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 262, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 58, + 262, + 295, + 304 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\kappa (\\mathbf {L A} (X)) \\leq \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3} (8) \\\\ \\kappa (\\mathbf {A} (X)) \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\kappa (X) \\kappa \\left(W _ {V}\\right). (9) \\\\ \\end{array}", + "image_path": "931e755fe77344d2f785fcc9081e66b64687115c0eede5b468796a5d4386ea14.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 296, + 334 + ], + "type": "text", + "content": "The proof of Proposition 4.2 is given in Sec. A of the Supp. material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 384 + ], + "type": "text", + "content": "Remark 4.3. The statement of Proposition 4.2 assumes the attention matrices are full rank. Since the layers of a transformer are randomly initialized we have that with probability 1 the attention matrices will be full rank at initialization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "text", + "content": "We point out that although Eq. (8) and Eq. (9) only provide an upper bound for the condition number, we found that the this upper bound provides a useful measure for estimating the condition number of the self-attention matrix. Therefore, for the self-attention matrices " + }, + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{A}(X)" + }, + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "inline_equation", + "content": "\\mathbf{L}\\mathbf{A}(X)" + }, + { + "bbox": [ + 55, + 387, + 296, + 459 + ], + "type": "text", + "content": " we make the following definition" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 58, + 464, + 295, + 505 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 464, + 295, + 505 + ], + "spans": [ + { + "bbox": [ + 58, + 464, + 295, + 505 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu (\\mathbf {L A} (X)) := \\kappa \\left(W _ {Q}\\right) \\kappa \\left(W _ {K}\\right) \\kappa \\left(W _ {V}\\right) \\kappa (X) ^ {3}. (10) \\\\ \\mu (\\mathbf {A} (X)) := \\kappa \\left(\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {k} ^ {T} X ^ {T}\\right)\\right) \\kappa (X) \\kappa \\left(W _ {V}\\right). (11) \\\\ \\end{array}", + "image_path": "922b36b67e08cbf5f85cddefe80137da514ee2d0adc300117260b51a5970a9b9.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 510, + 220, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 220, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 220, + 522 + ], + "type": "text", + "content": "4.2. Conditioned embedded tokens" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 528, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 296, + 601 + ], + "type": "text", + "content": "In this section, we present our main theorem, which provides a method for reducing the condition number of embedded tokens derived from a dataset. This reduction subsequently decreases the measure " + }, + { + "bbox": [ + 55, + 528, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{A}(X))" + }, + { + "bbox": [ + 55, + 528, + 296, + 601 + ], + "type": "text", + "content": ", a key metric for assessing the condition number of the self-attention matrix in the first transformer layer." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": "Theorem 4.4. Let " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "(x_{1},\\dots ,x_{N})" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": " denote " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": " tokens associated to an input dataset with " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathbb{R}^{t\\times 1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "E\\in \\mathbb{R}^{d\\times t}" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": " denote an embedding matrix so that " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "Ex_{i}\\in \\mathbb{R}^{d\\times 1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "X = [Ex_1\\dots Ex_N]^T\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": " denote the matrix of embedded tokens. Assume that " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "\\kappa (X) > 2" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": ". Then there exists a matrix " + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{N\\times d}" + }, + { + "bbox": [ + 55, + 605, + 296, + 676 + ], + "type": "text", + "content": " such that" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 143, + 683, + 295, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 683, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 143, + 683, + 295, + 696 + ], + "type": "interline_equation", + "content": "\\kappa (X + C) \\leq 2 \\tag {12}", + "image_path": "bc1d04e32d189728eaec77bcec1791978177cb8cefb49409698d6664515aacb3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 701, + 200, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 200, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 200, + 714 + ], + "type": "text", + "content": "and hence that " + }, + { + "bbox": [ + 55, + 701, + 200, + 714 + ], + "type": "inline_equation", + "content": "\\kappa(X + C) < \\kappa(X)" + }, + { + "bbox": [ + 55, + 701, + 200, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": "In particular, Theorem 4.4 shows that if the condition number of the embedded tokens " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " is extremely large. Adding a correction term given by " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " will bring it down significantly. As we shall see in the experiments Sec. 5 the condition number " + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "inline_equation", + "content": "\\kappa(X)" + }, + { + "bbox": [ + 313, + 72, + 554, + 156 + ], + "type": "text", + "content": " is empirically extremely large. The proof of Theorem 4.4 is given in Sec. A of the Supp. material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "text", + "content": "Theorem 4.5. Let " + }, + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "text", + "content": " denote a collection of embedded tokens, as defined in the statement of Theorem 4.4, and let " + }, + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 164, + 555, + 200 + ], + "type": "text", + "content": " denote the matrix given by Theorem 4.4. We then have" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 209, + 555, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 209, + 555, + 223 + ], + "spans": [ + { + "bbox": [ + 367, + 209, + 555, + 223 + ], + "type": "interline_equation", + "content": "\\mu (\\mathbf {L A} (X + C)) \\leq \\mu (\\mathbf {L A} (X)). \\tag {13}", + "image_path": "c718dddf1595441a9a2e7f2a0eb59a079dca0ada8715d826fb08a2bb4c3efd70.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 232, + 367, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 367, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 367, + 243 + ], + "type": "text", + "content": "Assume that" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 343, + 252, + 554, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 252, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 343, + 252, + 554, + 283 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\kappa (\\text {s o f t m a x} ((X + C) W _ {Q} W _ {K} ^ {T} (X ^ {T} + C ^ {T}))) \\\\ \\leq \\kappa (\\operatorname {s o f t m a x} \\left(X W _ {Q} W _ {K} ^ {T} X ^ {T}\\right)) \\tag {14} \\\\ \\end{array}", + "image_path": "90f8d8b5fed52590ba0462cc5ae3d9c92765f9d40687701ea08cede7ec762d49.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 293, + 334, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 293, + 334, + 303 + ], + "spans": [ + { + "bbox": [ + 314, + 293, + 334, + 303 + ], + "type": "text", + "content": "then" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 376, + 304, + 553, + 317 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 304, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 376, + 304, + 553, + 317 + ], + "type": "interline_equation", + "content": "\\mu (\\mathbf {A} (X + C)) \\leq \\mu (\\mathbf {A} (X)) \\tag {15}", + "image_path": "1666bf68fff76e31b103c520945629c2b4d625dfb4ba63013096d0b1931b41af.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{L}\\mathbf{A}(X))" + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": " is defined by Eq. (10) and " + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "inline_equation", + "content": "\\mu (\\mathbf{A}(X))" + }, + { + "bbox": [ + 313, + 322, + 554, + 346 + ], + "type": "text", + "content": " by Eq. (11)." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 354, + 554, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 473 + ], + "type": "text", + "content": "The proof of Theorem 4.5 is given in Sec. A of the Supp. material. A natural question arises: can self-attention be conditioned directly through queries, keys, and values, independent of the embedded tokens? Unfortunately, our approach of introducing a correction term does not extend to these components. This is because our correction modifies the singular values of the embedded tokens, whereas altering the singular values of queries, keys, or values could disrupt how self-attention captures spatial relationships between tokens." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 478, + 554, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 478, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 478, + 554, + 574 + ], + "type": "text", + "content": "Remark 4.6. We note that in the case of self-attention with a softmax activation Theorem 4.5 requires an assumption on the condition number of the non-linear self-attention probability matrix term softmax " + }, + { + "bbox": [ + 313, + 478, + 554, + 574 + ], + "type": "inline_equation", + "content": "((X + C)W_{Q}W_{K}^{T}(X^{T} + C^{T}))" + }, + { + "bbox": [ + 313, + 478, + 554, + 574 + ], + "type": "text", + "content": ". Although this may be seen as a limitation of the theory we will see in Sec. 5 that this framework still provides a useful methodology to obtain good performance on a variety of transformer applications." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 554, + 662 + ], + "type": "text", + "content": "Remark 4.7. Theorem 4.5 demonstrates that conditioning embedded tokens can improve the condition number of the self-attention matrix in the first layer. However, it does not address how this affects the condition number in subsequent layers. In Sec. 5, we will show that this effect propagates through the network, leading to a reduced condition number in deeper self-attention layers." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Remark 4.8. Our theory, as presented in Theorem 4.4 and Theorem 4.5, demonstrates that conditioning embedded tokens reduces the condition number of the self-attention matrix in the first layer. However, the framework does not" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4789" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 97, + 72, + 514, + 137 + ], + "blocks": [ + { + "bbox": [ + 97, + 72, + 514, + 137 + ], + "lines": [ + { + "bbox": [ + 97, + 72, + 514, + 137 + ], + "spans": [ + { + "bbox": [ + 97, + 72, + 514, + 137 + ], + "type": "table", + "html": "
Models
ViT-baseDeiT-baseSwin-baseXciT-mediumDaViT-base
Original embedded tokens80.381.683.182.283.6
Conditioned embedded tokens81.382.583.982.984.6
", + "image_path": "e5d71b57ac5437af76b6431fa5f87af0415b8638e9fffdb19c312c55f01e876b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "lines": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "spans": [ + { + "bbox": [ + 54, + 148, + 555, + 183 + ], + "type": "text", + "content": "Table 1. Comparison of vision transformers with their original embedded tokens verse one with a conditioned conditioned embedded tokens pre-trained on the ImageNet-1k dataset. We report the classification top-1% accuracy. In each case we see the transformers employing conditioned embedded tokens (ours) outperforms the original ones." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 202, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 202, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 294, + 275 + ], + "type": "text", + "content": "theoretically establish how this improvement translates to better optimization and, consequently, higher accuracy in applications. Nonetheless, in Sec. 5, we provide empirical evidence showing that our approach consistently enhances accuracy across various applications, highlighting its practical effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 138, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 138, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 138, + 300 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 305, + 296, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 305, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 305, + 296, + 354 + ], + "type": "text", + "content": "In this section, we evaluate our insight from Sec. 4 on a variety of transformer applications. For a detailed analysis of how we implemented conditioned embedded tokens see Sec. A in the Supp. material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 361, + 173, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 173, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 173, + 373 + ], + "type": "text", + "content": "5.1. Image Classification" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 378, + 296, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 378, + 296, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 378, + 296, + 461 + ], + "type": "text", + "content": "In this section, we apply our theoretical insights from Sec. 4 to vision transformers on an image classification task. We will train all vision transformers from scratch on the ImageNet-1k dataset. We will be testing on the following vision transformers. In each case we will compare the original architecture with one that uses conditioned embedded tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 545 + ], + "type": "text", + "content": "ViT [5]: ViT is a pioneering architecture that applies transformer-based processing to images by treating them as sequences of non-overlapping patches. In our study, we utilize ViT-Base (ViT-B), which is configured with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. The model employs a standard self-attention mechanism to capture dependencies across patches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 640 + ], + "type": "text", + "content": "DeiT [22]: DeiT builds upon ViT but is optimized for data efficiency. Unlike standard ViT, it employs a data-starvation training strategy [22], enabling faster convergence while maintaining strong performance. In our setup, we use DeiT-Base (DeiT-B) with a patch size of 16, an embedding dimension of 768, 12 attention heads, and 12 layers. Like ViT, DeiT utilizes a self-attention mechanism for feature extraction." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Swin Transformer [15]: Swin Transformer introduces a hierarchical architecture and a novel shifted window-based self-attention mechanism, significantly improving efficiency and effectiveness in vision tasks. In our experiments, we adopt the Swin-Base (Swin-B) variant, which features 128 channels in the hidden layers of the first stage." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 201, + 545, + 356 + ], + "blocks": [ + { + "bbox": [ + 317, + 201, + 545, + 356 + ], + "lines": [ + { + "bbox": [ + 317, + 201, + 545, + 356 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 545, + 356 + ], + "type": "image", + "image_path": "4ada79ab080496b73870e69728c9d70f45572144b0a9e3e91857f5ec62f1cbeb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 364, + 555, + 431 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 431 + ], + "type": "text", + "content": "Figure 2. Condition number comparison for ViT-B on ImageNet-1k. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over training epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "text", + "content": "The attention operates over non-overlapping windows of size " + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "inline_equation", + "content": "M = 7" + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "text", + "content": ", while the query dimension per head is set to " + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "inline_equation", + "content": "d = 32" + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "text", + "content": ". The network follows a hierarchical structure with layers arranged as " + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "inline_equation", + "content": "\\{2, 2, 18, 2\\}" + }, + { + "bbox": [ + 313, + 456, + 555, + 506 + ], + "type": "text", + "content": " across different stages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "text", + "content": "XCiT [2]: XCiT diverges from standard ViT architectures by incorporating two key innovations. First, it integrates Local Patch Interaction within each block, using depth-wise " + }, + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 507, + 556, + 639 + ], + "type": "text", + "content": " convolutions, Batch Normalization, GELU activation, and an additional depth-wise convolution to enhance local feature representation. Second, it employs Cross-Covariance Attention, where attention maps are derived from the cross-covariance matrix of the key and query projections. For our experiments, we utilize the XCiT-M24 (XCiT-M) model with a patch size of 16, an embedding dimension of 512, 8 attention heads, and 24 layers." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "DaViT [4]: DaViT extends traditional vision transformers by incorporating both spatial and channel attention mechanisms, enhancing feature extraction across multiple dimensions. While conventional transformers rely solely on spatial self-attention, DaViT introduces channel attention to capture dependencies across feature channels, complement" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4790" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 159 + ], + "type": "table", + "html": "
ModelAPbAPb50APb75APmAPm50APm75
XCiT-S + original embedded tokens44.966.148.940.163.142.8
XCiT-S + conditioned embedded tokens45.766.649.740.463.543.4
XCiT-M + original embedded tokens45.766.849.640.863.643.3
XCiT-M + conditioned embedded tokens46.267.449.841.463.844.2
", + "image_path": "8f74f7fd647159277849ad27cb07b2f9afdcd6c96901d7904d49115109d8ea25.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "text", + "content": "Table 2. Performance evaluation of object detection and instance segmentation on the COCO mini-val set. The backbone networks are pretrained on ImageNet-1k and integrated into the Mask R-CNN framework. The reported metrics include " + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "inline_equation", + "content": "AP^b" + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "text", + "content": " (Average Precision for bounding box predictions), " + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "inline_equation", + "content": "AP_{50/75}^b" + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "text", + "content": " (Average Precision at IoU thresholds of 0.50 and 0.75 for bounding boxes), " + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "inline_equation", + "content": "AP^m" + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "text", + "content": " (Average Precision for mask predictions), and " + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "inline_equation", + "content": "AP_{50/75}^m" + }, + { + "bbox": [ + 55, + 171, + 555, + 226 + ], + "type": "text", + "content": " (Average Precision at IoU thresholds of 0.50 and 0.75 for mask predictions). In each case we see the XCiT architecture employing conditioned embedded tokens (ours) out performs the original architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 247, + 294, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 294, + 319 + ], + "type": "text", + "content": "ing spatial interactions. This hybrid approach allows DaViT to effectively model both local and global relationships. In our experiments, we use the DaViT-Base (DaViT-B) model, configured with a patch size of 4, embedding dimensions of (128, 256, 512, 1024), attention heads of (4, 8, 16, 32), and a layer distribution of (1, 1, 9, 1)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 340, + 295, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 532 + ], + "type": "text", + "content": "Results: Table 1 presents a comparison of the top- " + }, + { + "bbox": [ + 55, + 340, + 295, + 532 + ], + "type": "inline_equation", + "content": "1\\%" + }, + { + "bbox": [ + 55, + 340, + 295, + 532 + ], + "type": "text", + "content": " test accuracy of five vision transformers trained from scratch on the ImageNet-1k dataset. Each model is evaluated using both the original embedded tokens and our conditioned embedded tokens, following the approach outlined in Sec. 4.2. While ViT-Base [5] and DeiT-Base [22] utilize standard self-attention mechanisms, Swin-Base [15], XCiT-Medium [2], and DaViT-Base [4] incorporate more advanced attention mechanisms. We include these latter models to illustrate that, although our theoretical analysis focuses on self-attention, conditioned embedded tokens can be easily integrated into modern transformers with more sophisticated architectures. As shown in the table, models using conditioned embedded tokens consistently achieve higher accuracy. Full implementation details and hyperparameter settings can be found in Sec. B.2 of the Supp. material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 664 + ], + "type": "text", + "content": "Analysis: Figure 2 illustrates the condition numbers of the embedded tokens, the first-layer self-attention matrix (averaged across all heads), and the overall self-attention matrix (averaged across all layers) for the ViT-B architecture. We compare these metrics for models using the original embedded tokens and our conditioned embedded tokens. These values were computed at each training epoch and then averaged over all epochs to generate the plots. The figure clearly shows that the ViT-B architecture with conditioned embedded tokens maintains on average a significantly lower condition number throughout training." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 295, + 714 + ], + "type": "text", + "content": "Implementation and hardware: For full details on the implementation of each vision transformer, the training hyperparameters and the hardware used please see Sec. B.2 in the Supp. material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 246, + 534, + 259 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 534, + 259 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 534, + 259 + ], + "type": "text", + "content": "5.2. Object Detection and Image Segmentation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 264, + 553, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 264, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 553, + 539 + ], + "type": "text", + "content": "In this section, we extend our approach from Sec. 4.2 to object detection and instance segmentation tasks. To evaluate its effectiveness in fine-tuning scenarios, we adapt a pretrained XCiT model, originally trained on ImageNet-1k, for these tasks. Our experiments leverage the COCO 2017 dataset [13], which consists of 118K training images and 5K validation images spanning 80 object categories. For model architecture, we employ XCiT as the backbone of the Mask R-CNN framework [9], incorporating a Feature Pyramid Network (FPN) to enhance multi-scale feature representation. To integrate XCiT with FPN, we modify its inherently columnar structure by extracting features from multiple layers. We do this by using an XCiT-small that has 12 layers (XCiT-S) and a XCiT-medium (XCiT-M) that has 24 layers. These features, originally computed at a fixed stride of 16, are adjusted to strides of [4, 8, 16, 32] to align with the FPN hierarchy. Max pooling is used for downsampling, while a single transposed convolution layer facilitates upsampling. The model is then fine-tuned for 36 epochs using the AdamW optimizer, with a learning rate of " + }, + { + "bbox": [ + 313, + 264, + 553, + 539 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 313, + 264, + 553, + 539 + ], + "type": "text", + "content": ", a weight decay of 0.05, and a batch size of 16. This training strategy effectively demonstrates how XCiT's learned representations can be adapted for downstream tasks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 540, + 553, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 553, + 624 + ], + "type": "text", + "content": "Results: We pretrained a total of four models on the ImageNet-1k dataset, namely an XCiT-S and XCiT-M both using the original embedded tokens following the methodology from Sec. 4.2. The results of the experiment are shown in Tab. 2. As can be seen from the table in both cases the transformer using conditioned embedded tokens out performs the original." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 625, + 553, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 625, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 625, + 553, + 661 + ], + "type": "text", + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used see Sec. B.3 in the Supp. material." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 671, + 420, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 420, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 420, + 684 + ], + "type": "text", + "content": "5.3. Language Models" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "In this section, we evaluate our insights from Sec. 4.2 on two different language models." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 313, + 757 + ], + "type": "text", + "content": "4791" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 83, + 70, + 529, + 125 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 125 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 125 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 125 + ], + "type": "table", + "html": "
MNLISST-2STSBRTEQNLIQQPMRPCCoLAGLUE
Crammed BERT (original)83.892.386.355.190.187.385.048.978.6
Crammed BERT (ours)84.292.586.555.691.187.486.353.779.7
", + "image_path": "a568153e1b4a87e5b9091aabb5766111d2c418b2310b50af60c346e337bf7dcc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 176, + 297, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 176, + 297, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 176, + 297, + 344 + ], + "type": "text", + "content": "Crammed BERT: In the first experiment we applied our insights from Sec. 4.2 to a Crammed BERT language model [8], trained entirely from scratch using masked language modeling with conditioned embedded tokens. The model consists of 110 million parameters, with 12 transformer layers and 12 attention heads per layer. We train two versions from scratch: the original Crammed BERT [8] and a variant incorporating conditioned embedded tokens using the insight from Sec. 4.2. Both models are trained on The Pile dataset [7], a large-scale corpus designed for language model training following the pretraining regime in [8]. After pretraining, we evaluate the performance on the GLUE benchmark [24] following the evaluation methodology outlined in [8]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 405 + ], + "type": "text", + "content": "Results: The evaluation results are shown in Tab. 3 with each number representing the accuracy. We see from the table that in each task the model employing conditioned embedded tokens outperforms the original Crammed Bert model on every task within the GLUE benchmark." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 425, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 425, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 56, + 425, + 295, + 628 + ], + "type": "text", + "content": "GPT-2: We conducted a second experiment training a GPT-2 model with a decoder-only transformer architecture, consisting of 12 layers, 12 attention heads per layer, and an embedding dimension of 512. The model was trained using masked self-attention to predict the next token, leveraging the TinyStories dataset [6] for language generation. Two versions of the model were trained from scratch: one using the original embedded tokens as in [6] and another incorporating the conditioned embedded tokens from Sec. 4.2. Both models were evaluated using perplexity loss (PPL), a widely used metric for assessing language models on datasets like TinyStories. Perplexity measures the model's confidence in its predictions by computing the exponential of the average negative log-likelihood of the predicted tokens. A lower perplexity score indicates that the model assigns higher probabilities to correct tokens, leading to better fluency and coherence in text generation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "Results and Analysis: Tab. 4 presents the results of our experiment, demonstrating that the model trained with conditioned embedded tokens consistently achieves lower validation loss and lower perplexity compared to the original GPT-2 model. These results indicate that conditioning the embedded tokens improves model performance, leading to more accurate language generation. Fig. 3 plots the con" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 354, + 174, + 515, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 133, + 555, + 157 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 555, + 157 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 555, + 157 + ], + "type": "text", + "content": "Table 3. Evaluation of a pre-trained Crammed BERT on the GLUE benchmark with the original embedded tokens (original) and our conditioned embedded tokens (ours). As can be seen from the table our methodology out performs the original in every task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 354, + 174, + 515, + 214 + ], + "lines": [ + { + "bbox": [ + 354, + 174, + 515, + 214 + ], + "spans": [ + { + "bbox": [ + 354, + 174, + 515, + 214 + ], + "type": "table", + "html": "
Val loss
GPT-2 original2.41
GPT-2 conditioned (ours)2.36
", + "image_path": "cbd38a5e03f0b5f2bf5eb7d42e2d355ea5c761e69a73dd137f6bd1f603d639c7.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 555, + 277 + ], + "type": "text", + "content": "Table 4. GPT-2 model trained on the TinyStories dataset. We compare two models, the original model and one using conditioned embedded tokens (conditioned). We report the validation loss; in both cases, our model yields a lower loss, indicating better performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 288, + 545, + 445 + ], + "blocks": [ + { + "bbox": [ + 316, + 288, + 545, + 445 + ], + "lines": [ + { + "bbox": [ + 316, + 288, + 545, + 445 + ], + "spans": [ + { + "bbox": [ + 316, + 288, + 545, + 445 + ], + "type": "image", + "image_path": "387be26c555014682eeea3fa4309c09c84fde3390f2c2c79465566c317baac9d.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 452, + 556, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 519 + ], + "type": "text", + "content": "Figure 3. Condition number comparison for GPT-2 training on TinyStories. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, respectively, averaged over all epochs. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 540, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 589 + ], + "type": "text", + "content": "dition numbers of the embedded tokens, self-attention in first layer averaged over all heads and self-attention averaged over all layers. Each of these are evaluated over each training epoch and then averaged over all epochs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 604, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 555, + 640 + ], + "type": "text", + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.4 in the Supp. material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 648, + 538, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 538, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 538, + 662 + ], + "type": "text", + "content": "5.4. Nyströmformer for Long Range Sequences" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "We applied our methodology to the Nyströmformer [25], a transformer architecture designed for efficient long-range dependency modeling on the Long-Range Arena (LRA) benchmark [21]. We trained two Nyströmformer models:" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "4792" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 71, + 525, + 126 + ], + "blocks": [ + { + "bbox": [ + 88, + 71, + 525, + 126 + ], + "lines": [ + { + "bbox": [ + 88, + 71, + 525, + 126 + ], + "spans": [ + { + "bbox": [ + 88, + 71, + 525, + 126 + ], + "type": "table", + "html": "
ModelListOpsTextRetrievalImagePathfinder
Nyströmformer (original)37.163.879.839.972.9
Nyströmformer conditioned (ours)37.964.980.940.180.0
", + "image_path": "7c03fa035e6ed6b2fb009bdc4336915ef7d161fb421e39ce09ff3563d78bb73e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 136, + 555, + 173 + ], + "lines": [ + { + "bbox": [ + 54, + 136, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 54, + 136, + 555, + 173 + ], + "type": "text", + "content": "Table 5. Comparison of a Nyströmformer using its original embedded tokens (original) with a Nyströmformer using conditioned embedded tokens (ours) on the LRA benchmark. We report the evaluation accuracy (\\%). As can be seen from the table our methodology yields a higher accuracy on each task." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 191, + 294, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 294, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 294, + 226 + ], + "type": "text", + "content": "one utilizing conditioned embedded tokens and the original Nyströmformer for comparison. The Long-Range Arena (LRA) benchmark consists of five distinct tasks:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 295, + 383 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 295, + 251 + ], + "type": "text", + "content": "1. ListOps: A synthetic task that tests hierarchical reasoning by evaluating nested arithmetic operations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 251, + 294, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 294, + 274 + ], + "type": "text", + "content": "2. Text Classification: Character-level text classification, requiring models to process long sequences of text." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 295, + 310 + ], + "type": "text", + "content": "3. Retrieval: A document retrieval task that measures the model's ability to identify relevant patterns within long textual inputs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 310, + 295, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 346 + ], + "type": "text", + "content": "4. Image Classification: Modeling image data as a sequence of flattened pixel values to assess a model's capability in visual recognition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 346, + 295, + 383 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 295, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 295, + 383 + ], + "type": "text", + "content": "5. Pathfinder: A visual reasoning challenge that requires detecting connected paths in an image, emphasizing spatial relationship modeling." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 54, + 396, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 396, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 54, + 396, + 295, + 563 + ], + "type": "text", + "content": "Results and Analysis: Tab. 5 shows the results of the experiment. As can be seen from the table the Nyströmformer that employed conditioned embedded tokens outperformed the original Nyströmformer from [25] in every task in the LRA benchmark. We computed the condition numbers of the original Nyströmformer and one employing conditioned embedded tokens throughout training for the embedded tokens, the self-attention within the first layer averaged over all heads and the self-attention averaged over all layers. We then took the average for each task over all training epochs and averaged over the five tasks within the LRA benchmark. The plots are shown in Fig. 4. As can be seen from the figure, the Nyströmformer employing conditioned embedded tokens has lower condition number in all cases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 563, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 295, + 601 + ], + "type": "text", + "content": "Implementation and hardware: For full details on the implementation, the training hyperparameters and the hardware used please see Sec. B.5 in the Supp. material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 130, + 623 + ], + "type": "text", + "content": "6. Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 629, + 295, + 715 + ], + "type": "text", + "content": "Our work explored the conditioning of transformer attention blocks through embedded tokens. While we showed that conditioned embedded tokens reduce the first-layer self-attention condition number, their impact on optimization remains theoretically unproven. Prior studies [1, 10, 14] have linked NTK conditioning to optimization in feedforward networks, raising an important question: how does" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 190, + 545, + 344 + ], + "blocks": [ + { + "bbox": [ + 317, + 190, + 545, + 344 + ], + "lines": [ + { + "bbox": [ + 317, + 190, + 545, + 344 + ], + "spans": [ + { + "bbox": [ + 317, + 190, + 545, + 344 + ], + "type": "image", + "image_path": "509eeaef0fd0de7359dd3a4183512e40b1e73296a74c1eeb13a9f0437d909d39.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 555, + 430 + ], + "type": "text", + "content": "Figure 4. Condition number comparison for Nyströmformer on the LRA benchmark. The left, middle, and right bars show the condition number of embedded tokens, first-layer self-attention (averaged across heads), and self-attention across all layers, averaged over training epochs and tasks. Our model (conditioned) consistently achieves a significantly lower condition number than the original." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 452, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 555, + 499 + ], + "type": "text", + "content": "conditioning embedded tokens affect the NTK in transformers? Investigating this could offer deeper insights into how embedded token structure influences transformer optimization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 388, + 525 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 534, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we introduced a framework for analyzing the conditioning of the self-attention matrix in the first layer of a transformer, highlighting its dependence on the embedded tokens. Our analysis demonstrated that by adding a carefully designed correction term—which we refer to as conditioned embedded tokens—we could significantly reduce the condition number of self-attention in the first layer, leading to a better-conditioned attention mechanism. We empirically validated the effectiveness of conditioned embedded tokens across a diverse set of tasks, including image classification, object detection, instance segmentation, language modeling, and long-range sequence modeling. In all cases, our approach proved to be a simple, drop-in replacement for existing embedding methods, consistently improving model performance across these applications." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4793" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Naman Agarwal, Pranjal Awasthi, and Satyen Kale. A deep conditioning treatment of neural networks. In Algorithmic Learning Theory, pages 249-305. PMLR, 2021. 3, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 190 + ], + "type": "text", + "content": "[2] Alaaeldin Ali, Hugo Touvron, Mathilde Caron, Piotr Bojanowski, Matthijs Douze, Armand Joulin, Ivan Laptev, Natalia Neverova, Gabriel Synnaeve, Jakob Verbeek, et al. Xcit: Cross-covariance image transformers. Advances in neural information processing systems, 34:20014-20027, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 237 + ], + "type": "text", + "content": "[3] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In European conference on computer vision, pages 213-229. Springer, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[4] Mingyu Ding, Bin Xiao, Noel Codella, Ping Luo, Jingdong Wang, and Lu Yuan. Davit: Dual attention vision transformers. In European conference on computer vision, pages 74-92. Springer, 2022. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 350 + ], + "type": "text", + "content": "[5] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 351, + 295, + 384 + ], + "type": "text", + "content": "[6] Ronen Eldan and Yuanzhi Li. Tinystories: How small can language models be and still speak coherent english? arXiv preprint arXiv:2305.07759, 2023. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 295, + 441 + ], + "type": "text", + "content": "[7] Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Aadi Thite, Eric Nabeshima, et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2021. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 442, + 295, + 485 + ], + "type": "text", + "content": "[8] Jonas Geiping and Tom Goldstein. Cramming: Training a language model on a singlegpu in one day. In International Conference on Machine Learning, pages 11117-11143. PMLR, 2023. 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 487, + 295, + 520 + ], + "type": "text", + "content": "[9] Kaiming He, Georgia Gkioxari, Piotr Dólár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 295, + 566 + ], + "type": "text", + "content": "[10] Arthur Jacot, Franck Gabriel, and Clément Hongler. Neural tangent kernel: Convergence and generalization in neural networks. Advances in neural information processing systems, 31, 2018. 3, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 610 + ], + "type": "text", + "content": "[11] Yiping Ji, Hemanth Saratchandran, Cameron Gordon, Zeyu Zhang, and Simon Lucey. Efficient learning with sine-activated low-rank matrices. arXiv preprint arXiv:2403.19243, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 645 + ], + "type": "text", + "content": "[12] Yiping Ji, Hemanth Saratchandran, Peyman Moghaddam, and Simon Lucey. Always skip attention. arXiv preprint arXiv:2505.01996, 2025.3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 714 + ], + "type": "text", + "content": "[13] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Chaoyue Liu, Libin Zhu, and Mikhail Belkin. Loss landscapes and optimization in over-parameterized non-linear systems and neural networks. Applied and Computational Harmonic Analysis, 59:85-116, 2022. 2, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 173 + ], + "type": "text", + "content": "[15] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10012-10022, 2021. 2, 5, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 554, + 228 + ], + "type": "text", + "content": "[16] Abhisek Maiti, Sander Oude Elberink, and George Vosselman. Transfusion: Multi-modal fusion network for semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6536-6546, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 252 + ], + "type": "text", + "content": "[17] Jorge Nocedal and Stephen J Wright. Numerical optimization. Springer, 1999. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 554, + 297 + ], + "type": "text", + "content": "[18] Tim Salzmann, Boris Ivanovic, Punarjay Chakravarty, and Marco Pavone. Trajectory: Multi-agent generative trajectory forecasting with heterogeneous data for control. arXiv preprint arXiv:2001.03093, 2, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 299, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 299, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 299, + 554, + 341 + ], + "type": "text", + "content": "[19] Hemanth Saratchandran, Jianqiao Zheng, Yiping Ji, Wenbo Zhang, and Simon Lucey. Rethinking attention: Polynomial alternatives to softmax in transformers. arXiv preprint arXiv:2410.18613, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 387 + ], + "type": "text", + "content": "[20] Hemanth Saratchandran, Thomas X Wang, and Simon Lucey. Weight conditioning for smooth optimization of neural networks. In European Conference on Computer Vision, pages 310-325. Springer, 2025. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 442 + ], + "type": "text", + "content": "[21] Yi Tay, Mostafa Dehghani, Samira Abnar, Yikang Shen, Dara Bahri, Philip Pham, Jinfeng Rao, Liu Yang, Sebastian Ruder, and Donald Metzler. Long range arena: A benchmark for efficient transformers. arXiv preprint arXiv:2011.04006, 2020. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 554, + 498 + ], + "type": "text", + "content": "[22] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In International conference on machine learning, pages 10347-10357. PMLR, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 501, + 554, + 544 + ], + "type": "text", + "content": "[23] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 554, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 554, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 554, + 589 + ], + "type": "text", + "content": "[24] Alex Wang, Amanpreet Singh, Julian Michael, Felix Hill, Omer Levy, and Samuel R Bowman. Glue: A multi-task benchmark and analysis platform for natural language understanding. arXiv preprint arXiv:1804.07461, 2018. 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 554, + 645 + ], + "type": "text", + "content": "[25] Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, and Vikas Singh. Nystromformer: A nystrom-based algorithm for approximating self-attention. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 14138–14148, 2021. 2, 3, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 689 + ], + "type": "text", + "content": "[26] Li Yuan, Qibin Hou, Zihang Jiang, Jiashi Feng, and Shuicheng Yan. Volo: Vision observer for visual recognition. IEEE transactions on pattern analysis and machine intelligence, 45(5):6575-6586, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[27] Q Zhen, W Sun, H Deng, D Li, Y Wei, B Lv, J Yan, L Kong, and Y Zhong. cosformer: rethinking softmax in attention." + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4794" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 94 + ], + "type": "text", + "content": "In International Conference on Learning Representations, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[28] Liu Zhuang, Lin Wayne, Shi Ya, and Zhao Jun. A robustly optimized bert pre-training approach with post-training. In Proceedings of the 20th chinese national conference on computational linguistics, pages 1218-1227, 2021. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4795" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_content_list.json b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..88a9baa241a7f38c1fc1cc5d6e6905d79cf38b3b --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_content_list.json @@ -0,0 +1,1529 @@ +[ + { + "type": "text", + "text": "Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shiwei Zhang $^{1,\\dagger}$ , Qi Zhou $^{1,\\dagger}$ , Wei Ke $^{1,2,*}$", + "bbox": [ + 336, + 202, + 658, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Software Engineering, Xi'an Jiaotong University, China", + "bbox": [ + 233, + 220, + 763, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Pengcheng Laboratory, Shenzhen, China", + "bbox": [ + 333, + 238, + 663, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangshiwei96,zhouqq}@stu.xjtu.edu.cn,wei.ke@mail.xjtu.edu.cn", + "bbox": [ + 217, + 258, + 774, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Text-guided zero-shot object counting leverages vision-language models (VLMs) to count objects of an arbitrary class given by a text prompt. Existing approaches for this challenging task only utilize local patch-level features to fuse with text feature, ignoring the important influence of the global image-level feature. In this paper, we propose a universal strategy that can exploit both local patch-level features and global image-level feature simultaneously. Specifically, to improve the localization ability of VLMs, we propose Text-guided Local Ranking. Depending on the prior knowledge that foreground patches have higher similarity with the text prompt, a new local-text rank loss is designed to increase the differences between the similarity scores of foreground and background patches which push foreground and background patches apart. To enhance the counting ability of VLMs, Number-evoked Global Attention is introduced to first align global image-level feature with multiple number-conditioned text prompts. Then, the one with the highest similarity is selected to compute cross-attention with the global image-level feature. Through extensive experiments on widely used datasets and methods, the proposed approach has demonstrated superior advancements in performance, generalization, and scalability. Furthermore, to better evaluate text-guided zero-shot object counting methods, we propose a dataset named ZSC-8K, which is larger and more challenging, to establish a new benchmark. Codes and dataset are released at https://github.com/zaqai/LGCount.", + "bbox": [ + 89, + 340, + 483, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 792, + 220, + 806 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, with the development of deep learning techniques [10, 37, 42], category-specific counting (CSC) methods [30, 44] have gradually matured. However, existing CSC methods face notable limitations in practical applica", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cb563e8f39c3719aa40742e7e1d44073cbdd7c66949540d53cf6a26a6fa4a1e5.jpg", + "image_caption": [ + "Figure 1. Comparison of different pipelines for text-guided zero-shot object counting. (a) Only fusing single text prompt and local patch-level features. (b) Instead of overlooking the global image-level feature, our universal strategy leverages the global image-level feature to align with multiple number-conditioned text prompts and select the one with the highest similarity to calculate cross-attention with image-level feature. At last, the feature obtained by cross-attention is added to the local patch-level features." + ], + "image_footnote": [], + "bbox": [ + 517, + 309, + 903, + 512 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tions. These approaches require training dedicated models on extensive labeled datasets for each object category, so they fundamentally lack the capacity to handle unseen objects in a human-like manner. Therefore, counting methods that are not limited by fixed categories have gradually attracted the attention of the academic community, as they offer greater flexibility and broader application potential.", + "bbox": [ + 509, + 652, + 906, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many exemplar-guided class-agnostic object counting (CAC) methods [38, 40, 52] have been presented in this context. These methods take several human-annotated bounding boxes of objects as exemplars to specify the objects-of-interest and train a unified network that can estimate the number of arbitrary objects-of-interest. However, previous CAC methods heavily rely on the assumption that accurate bounding boxes of objects can be easily obtained. In practice, this requirement is unfriendly to end-users lacking pro", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Equal contribution. * Corresponding author.", + "bbox": [ + 114, + 886, + 359, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21097", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "fessional knowledge and is even impossible to fulfill for a fully automatic system (e.g. a wildlife monitoring system).", + "bbox": [ + 89, + 90, + 480, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To provide a more flexible object counting approach, text-guided zero-shot object counting (ZSC) has been proposed by [49] and further developed by [8, 18] which leverage vision-language models (VLMs) [27, 37]. Under this setting, the bounding boxes of exemplars are replaced by the class name. During inference, users only need to specify the class name of the object that they want to count, and the model can automatically output the number of objects of this class. Even if the class is completely unseen during the training stage, the counter should still adapt to this unseen class without any additional data. Nevertheless, as Fig. 1(a) shows, existing ZSC methods only utilize local patch-level features from vision encoder to fuse with text feature and then use the fused features to make predictions. The global image-level feature, which is also derived from the vision encoder, is discarded, and its significance is completely overlooked by previous methods.", + "bbox": [ + 88, + 125, + 482, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose a universal strategy which exploits both local patch-level features and global image-level feature to enhance text-guided zero-shot object counting as shown in Fig. 1(b). It consists of two parts: Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). In TLR, we reveal a critical limitation in prior method [18]: its utilization of local patch-level features inadvertently reduces the discriminative gap between foreground and background similarity scores, which fundamentally undermines VLMs' localization performance. To alleviate this limitation, we leverage the prior knowledge that the similarity scores of foreground patches must be greater than those of background features to design a novel local-text rank loss. This loss is designed to increase the differences between the foreground and background similarity scores which push the foreground and background patches apart to strengthen the localization ability of VLMs. In NGA, we exploit the global image-level feature and propose multiple number-conditioned text prompts. Unlike previous single text prompt that merely specify the class, our multiple number-conditioned text prompts not only indicate the class but also have predefined number ranges. And a similarity loss is proposed to align the global image-level feature with the prompt that has the closest number range. Then, to find a more accurate range, a coarse-to-fine global-number alignment scheme is proposed. In the end, the number-conditioned text prompt with the highest similarity is selected to calculate cross-attention with global image-level feature. Considering that the text prompt used in VLMs' pre-training stage usually lacks numeric information, by interacting number-conditioned text prompt with global image-level feature, our proposed NGA significantly improves the counting ability of VLMs.", + "bbox": [ + 91, + 383, + 482, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To verify the effectiveness of our proposed strategy, we", + "bbox": [ + 109, + 885, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "conducted experiments on three prevailing datasets: FSC-147 [38], CARPK [15] and ShanghaiTech [55]. Moreover, to perform a more comprehensive evaluation of ZSC methods, we create a larger and more challenging dataset, named ZSC-8K, which increases the proportion of common classes and moves long-tailed categories in FSC-147 to the test set of ZSC-8K to evaluate the performance on long-tailed classes. The results demonstrate that our method significantly surpasses previous text-guided zero-shot object counting methods. Additionally, various backbones including CLIP [37] and GroundingDino [27] are also employed in our experiments to confirm the scalability of our strategy.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, the main contributions of the proposed approach can be summarized as follows:", + "bbox": [ + 511, + 277, + 903, + 306 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We develop a universal strategy consists of Text-guided Local Ranking and Number-evoked Global Attention. This strategy is capable of exploiting both local patch-level features and global image-level feature for text-guided zero-shot object counting.", + "- Text-guided Local Ranking is proposed to improve the localization ability of VLMs by using a local-text rank loss to push the foreground and background patches apart. Number-evoked Global Attention is presented to enhance the counting ability of VLMs by computing crossattention between global image-level feature and number-conditioned text prompt with highest similarity.", + "- We introduce ZSC-8K, a scaled-up dataset consists of 8,361 images and 114 categories for comprehensive evaluation, which can benefit future researchers in this area.", + "- The proposed strategy achieves state-of-the-art performance on ZSC-8K dataset and three other popular datasets. Its effectiveness, scalability and cross-dataset generalizability are verified by experiments." + ], + "bbox": [ + 513, + 311, + 903, + 598 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 627, + 661, + 642 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Category-specific object counting", + "text_level": 1, + "bbox": [ + 511, + 657, + 805, + 672 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-specific object counting (CSC) concentrates on counting objects of one specific category, for instance, humans [6, 19, 21, 24, 26, 30, 44, 55], penguins [5], birds [1], cells [48], or cars [15, 31, 32, 53]. Generally, existing CSC approaches can be classified into two groups: point-based methods and density-based methods. Point-based methods [6, 26, 44, 54] apply an object detector to the image and count the number of objects based on the detected points. Density-based methods [22, 39, 55] predict a density map for each input image, and the final result is obtained by summing up the pixel values.", + "bbox": [ + 511, + 684, + 903, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, since existing CSC methods need to train a specific network with extensive labeled data for each class of object, they are deemed time-consuming and laborious.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21098", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Exemplar-guided class-agnostic counting", + "text_level": 1, + "bbox": [ + 89, + 90, + 442, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To ease the burden of humans and enable the counting model to count arbitrary categories, researchers propose exemplar-guided class-agnostic counting (CAC) [4, 12, 16, 25, 29, 33, 40, 51, 52] which leverages human annotated images or patches as exemplars of a category to train a unified network that can count arbitrary categories.", + "bbox": [ + 89, + 112, + 483, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[12] indicates that color, shape and scale diversity are crucial for CAC performance and proposes an EFA module to handle diversity from all aspects. [16] proposes an interactive class-agnostic object counting framework. FamNet [38] presents an adaptation strategy to adapt network to any novel visual category at test time. [33] introduces an uncertainty-aware object detector. BMNet [40] proposes a similarity-aware CAC framework. CFOCNet [51] uses reference images as exemplars to reduce the huge cost of data collection. [52] encourages the model to inspect the query image by focusing more on the regions akin to the support images. CountGD [3] leverages both visual exemplar and text prompt to boost the model. [45] extracts multiple occluded features of exemplars for comparison and utilizes the position of instances for more accurate counting results.", + "bbox": [ + 89, + 203, + 483, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CAC methods usually assume that high-quality bounding boxes of objects can be obtained easily. However, obtaining these exemplars is not as easy as assumed. In practice, the requirement of providing exemplars is unfriendly to users who lack professional knowledge and will be an obstacle when we want to build a fully automatic system.", + "bbox": [ + 89, + 430, + 483, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Text-guided zero-shot object counting", + "text_level": 1, + "bbox": [ + 89, + 531, + 416, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Considering that high-quality exemplars are not always easily obtained and aiming to construct a user-friendly system, text-guided zero-shot object counting (ZSC) [2, 8, 18, 20, 41, 49] is proposed to provide a more flexible and user-friendly approach.", + "bbox": [ + 89, + 551, + 482, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[49] first introduces ZSC and trains a VAE model to generate the class prototype. The generated prototype is used to select class-relevant patches as exemplars to train a counting network. Later, with the application of vision-language models (VLMs) [27, 37], even automatically obtained exemplars are no longer needed. [2, 18, 20, 34] leverage CLIP [37] to solve ZSC task. [18] proposes an end-to-end pipeline to estimate density maps for a given class, [2] introduces a feature interaction module to enhance the feature and [20] incorporates SPT, LAT, and SaSC modules to tailor VLBase for object counting. ABC123 [14] proposed a blind counter but can only count images with fewer than 300 objects. Furthermore, [8] employs GroundingDino [27] as its backbone, and once again, the boundary of ZSC is advanced. VA-Count [56] then uses GroundingDino to generate good exemplars to boost the model.", + "bbox": [ + 89, + 628, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nonetheless, previous ZSC approaches only leverage local patch-level features from the vision encoder to interact", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/894ce1e8f5528f8387c77ffcc39724c28db43ce4ac2ef396832bbdfd4abdec9f.jpg", + "image_caption": [ + "Figure 2. Distribution of each category in FSC-147 [38] (left) and the proposed ZSC-8K (right)." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 220 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/ff42230f125cf18df62e2eda154aa6203081be489b0de5ac92d22dea9e09e998.jpg", + "table_caption": [], + "table_footnote": [ + "Table 1. Statistics of each set in ZSC-8K, including the number of point annotations (max, min, avg, total)." + ], + "table_body": "
#Images#ClassesMaxMinAvgTotal
Overall83611141766146.23386498
Train5476301766149.72272243
Val112729485135.1639623
Testnormal139739953137.1051829
Testlongtail175857953142.4574632
", + "bbox": [ + 514, + 261, + 921, + 364 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with text feature. The global image-level feature is discarded, and its significance is largely overlooked. It still lacks a systematic method to utilize both local and global features to enhance ZSC approaches.", + "bbox": [ + 511, + 407, + 906, + 468 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. ZSC-8K Dataset", + "text_level": 1, + "bbox": [ + 511, + 481, + 676, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "FSC-147 [38] dataset is the most widely-used dataset for text-guided zero-shot object counting (ZSC). However, it demonstrates an unrealistic class distribution characterized by excessive long-tailed categories (polka dots, cereals, shallots, etc.) and insufficient samples for common classes. This distributional bias deviates markedly from real-world occurrence patterns. Therefore, to conduct a more comprehensive evaluation of ZSC methods, we propose ZSC-8K dataset, inspired by REC dataset [8]. As shown in Fig. 2, compared to FSC-147, ZSC-8K is larger and more challenging which means it increases the proportion of common classes (e.g. cars, people) and moves long-tailed classes from FSC-147 to test set to evaluate the performance of ZSC methods on long-tailed classes.", + "bbox": [ + 511, + 507, + 906, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Data collection. Images in the ZSC-8K dataset are collected from multiple existing datasets, including FSC147 [38], JHU-Crowd [43], NWPU-Crowd [46], VisDrone [57], DETRAC [47], CARPK [15], MALL [7] and Crowd Surveillance [50]. The proportions of each dataset are depicted in the supplementary material. These datasets belong to different domains, including crowd counting, car counting and animal counting, etc.", + "bbox": [ + 511, + 719, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Dataset statistics. ZSC-8K contains 8,361 images with a total of 386,498 point annotations. The number of point annotations in an image ranges from 1 to 1,766. A histogram of point annotations per image is provided in the", + "bbox": [ + 511, + 839, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21099", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/005ba00e8bffe892c82abd4887d803ed6364bfb55f2fd1170f84c2684fadae84.jpg", + "image_caption": [ + "Figure 3. Illustration of our training pipeline. In Text-guided Local Ranking, depending on the prior knowledge of similarity score ranking, local-text rank loss is proposed to push the foreground and background patches apart which further improves the localization ability of VLMs. In Number-evoked Global Attention, a coarse-to-fine global-number alignment scheme is proposed to align global image-level feature with multiple coarse number-conditioned text prompts firstly, and with multiple fine number-conditioned text prompts secondly. The fine number-conditioned text prompt with the highest similarity score is selected to calculate cross-attention with image-level feature. And the feature obtained by cross-attention is added to the local patch-level features. In inference stage, the test images and their corresponding number-conditioned text prompts are forwarded to decoder to make predictions." + ], + "image_footnote": [ + "Ft Text feature F1 Local patch-level features Fg Global image-level feature Fct Feature with highest similarity Pct Prompt of Fct" + ], + "bbox": [ + 98, + 89, + 537, + 327 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/54f288501a5899167ac38658c8d651f31d459ebe7b5bae8648d2a09d27e20fa2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 539, + 89, + 901, + 327 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "supplementary material. The dataset can be divided into 5 major categories and 114 subcategories. The proportions of each category are shown in Fig. 2. As it depicts, people are the category with the highest proportion and cars are the second which are significantly higher than that in FSC-147 dataset [38]. We split ZSC-8K dataset into train, validation and test sets. Specific statistics for each set are detailed in Tab. 1. It is worth noting that two versions of test sets are proposed: $\\text{test}_{\\text{normal}}$ and $\\text{test}_{\\text{longtail}}$ . In $\\text{test}_{\\text{normal}}$ , images of normal classes (1397 images & 39 classes) are included, while $\\text{test}_{\\text{longtail}}$ is obtained by merging the long-tailed classes (361 images & 20 classes) from FSC-147 into $\\text{test}_{\\text{normal}}$ . To meet the zero-shot setting, the categories in train set do not overlap with validation and test sets.", + "bbox": [ + 88, + 470, + 485, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Methodology", + "text_level": 1, + "bbox": [ + 89, + 700, + 225, + 717 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 726, + 200, + 739 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of ZSC is to count anything using only text guidance. Pretrained vision-language models (VLMs) are commonly used for this purpose. Denote $I \\in R^{H \\times W \\times 3}$ as an input image containing an arbitrary object class, and a text prompt $t$ is given to specify the object class of this image. In previous methods, $I$ is initially fed into the image encoder of pretrained VLMs to obtain local patch-level features and $t$ is then encoded by the text encoder to obtain text feature. Subsequently, a fuse module is designed to combine the two modalities, and the fused feature is used to predict the num", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ber of the specified category. In this paper, we argue that previous methods only leverage local patch-level features to fuse with the text feature, thereby fully overlooking the importance of the global image-level feature. Therefore, to exploit local patch-level features and global image-level simultaneously, we propose a strategy consists of Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). The overall framework is shown in Fig. 3.", + "bbox": [ + 511, + 470, + 906, + 592 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, the local patch-level features and text feature are used to compute a novel local-text rank loss that can increase the differences between foreground and background similarity scores in TLR. The localization ability of VLMs is improved by TLR. Then, we leverage the global image-level feature in NGA. Multiple number-conditioned text prompts are designed and a coarse-to-fine global-number alignment scheme is proposed to select the number-conditioned text prompt with highest similarity. The selected prompt is considered to have the closest number range and its feature is used to calculate cross-attention with the global image-level feature to enhance the counting ability. Finally, the feature obtained by cross-attention is aggregated with local patch-level features through element-wise summation. This combined representations is then fed into decoder to make predictions. TLR and NGA are detailed in Sec. 4.2 and Sec. 4.3, respectively.", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21100", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Text-guided Local Ranking", + "text_level": 1, + "bbox": [ + 89, + 90, + 336, + 107 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "VLMs have shown impressive performance in estimating text-image similarity by aligning text and image representations. However, as demonstrated by ClipCount [18], pretrained VLMs intrinsically lack localization ability, making it non-trivial to transfer this ability to object counting. Although ClipCount [18] utilized the local patch-level features $F_{l}$ and the text feature $F_{t}$ to construct a contrastive loss to enhance localization ability by pushing foreground and background patches apart, this loss cannot effectively increase the discrepancy between foreground and background similarity scores, which in turn adversely affects localization ability. As shown in Fig. 4(a), the differences between foreground and background similarity scores in ClipCount [18] are small, with most differences are close to zero. To alleviate this issue, we propose Text-guided Local Ranking (TLR) in which a novel local-text rank loss is designed to improve the localization ability by increasing the differences between foreground and background similarity scores. As Fig. 4(b) shows, with TLR, the differences become much larger than those of ClipCount [18].", + "bbox": [ + 89, + 116, + 483, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Fig. 3, $F_{l}$ and $F_{t}$ are also used in the proposed TLR, a linear projection maps them to the same channels. To determine which patch contains the object, max-pooling is used to transfer the ground truth density map to a patch-level binary mask. Patches whose value is equal to 1 are foreground patches which have the object and the value of background patches is 0. We denote the foreground patches as $\\mathcal{P}$ and the background patches as $\\mathcal{B}$ , the relation of $\\mathcal{P}$ and $\\mathcal{B}$ can be summarized as:", + "bbox": [ + 89, + 422, + 483, + 559 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(F _ {l} ^ {p}, F _ {t}\\right) > s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\forall p \\in \\mathcal {P}, \\forall q \\in \\mathcal {B}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 150, + 580, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s(,)$ denotes the cosine similarity. And Formula 1 indicates the prior knowledge that the similarity score between any foreground patch feature and the text feature is greater than that of any background patch feature. Depending on the prior knowledge, we construct the local-text rank loss. The similarity score sets are derived as:", + "bbox": [ + 89, + 618, + 483, + 709 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathcal {P}) = \\{\\dots , s \\left(F _ {l} ^ {p}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {P} | = l _ {1}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 731, + 482, + 750 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS (\\mathcal {B}) = \\{\\dots , s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {B} | = l _ {2}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 797, + 482, + 815 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $S(\\mathcal{P})$ and $S(\\mathcal{B})$ are similarity score sets of foreground and background patches respectively. $l_{1}$ and $l_{2}$ are the lengths of the two sets. $\\tau_{1}$ is a temperature parameter. Next, we denote $d_{pq} = (s(F_l^p,F_t) - s(F_l^q,F_t)) / \\tau_1$ , and the", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bd21b25f9c9718d218ab757a4a4f6f2285321c40a52ec15b800eeaacd0cdd3fb.jpg", + "image_caption": [ + "Figure 4. The differences between foreground and background similarity scores. (a) shows the differences between foreground and background scores in ClipCount [18] are small, most of the differences are close to 0. (b) shows the differences with our TLR become larger. Hence, the features of foreground and background patches are pushed apart." + ], + "image_footnote": [], + "bbox": [ + 517, + 88, + 903, + 199 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "differences between the two sets can be computed by:", + "bbox": [ + 511, + 294, + 870, + 310 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} M _ {d i f} = S (\\mathcal {P}) \\stackrel {\\text {b r o a d c a s t - s u b t r a c t}} {\\longrightarrow} S (\\mathcal {B}) \\\\ = \\left[ \\begin{array}{c c c c} d _ {1 1} & d _ {1 2} & \\dots & d _ {1 l _ {2}} \\\\ d _ {2 1} & d _ {2 2} & \\dots & d _ {2 l _ {2}} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ d _ {l _ {1} 1} & d _ {l _ {1} 2} & \\dots & d _ {l _ {1} l _ {2}} \\end{array} \\right] \\quad , \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 319, + 906, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $M_{dif}$ is the result of subtracting $S(\\mathcal{B})$ from $S(\\mathcal{P})$ using the broadcasting mechanism. Finally, with the prior knowledge, the local-text rank loss can be formulated as:", + "bbox": [ + 511, + 422, + 905, + 469 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {r a n k}} = \\frac {\\sum \\operatorname {c l a m p} \\left(\\lambda - M _ {\\text {d i f}} , \\min = 0\\right)}{l _ {1} l _ {2}}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 477, + 905, + 510 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ is a hyperparameter that controls the rank loss.", + "bbox": [ + 511, + 518, + 872, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With the proposed local-text rank loss, the differences are increased by ensuring the similarity scores of foreground patches are bigger than background patches. Thus the features of foreground and background patches are pushed apart and the localization ability is improved.", + "bbox": [ + 511, + 534, + 905, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Number-Evoked Global Attention", + "text_level": 1, + "bbox": [ + 511, + 617, + 810, + 632 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Because most of the text prompts used in pre-training stage of VLMs lack descriptions of the number of objects, the performance of VLMs in object counting is largely limited. To mitigate this limitation, we propose Number-Evoked Global Attention (NGA) to enhance the counting ability of VLMs by calculating the cross-attention between the global image-level feature $F_{g}$ and the feature of the number-conditioned text prompt with highest similarity $F_{ct}^{h}$ .", + "bbox": [ + 511, + 638, + 905, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, $F_{g}$ that is ignored by pervious ZSC methods is leveraged to select $P_{ct}^{h}$ , the number-conditioned text prompt with highest similarity, as shown in Fig. 3. Specifically, we first propose a global-number alignment scheme to identify a coarse number range for an image. To achieve this, we replace the single text prompt by multiple number-conditioned text prompts. Different from previous single text prompt which only specify the object class, every prompt of our proposed multiple number-conditioned text prompts has an object class and a predefined number range.", + "bbox": [ + 511, + 760, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21101", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8503dddfdfc2f3bf94de412a48e3218ebd1a2be6d143a48cf9f3a8a2724f834f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneVenueVal setTest set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ZSC [49]ResNet+VAECVPR'2326.9388.6322.09115.17
TFOC [41]SAMWACV'24--20.10132.83
CountTX* [2]CLIPBMCV'2317.7665.2116.70105.21
VA-Count [56]CLIPECCV'2417.8773.2217.88129.31
GeCo [36]CLIPNeurIPS'2414.8164.9513.30108.72
DAVEprm [35]CLIPCVPR'2415.4852.5714.90103.42
GroundingREC* [8]GDinoCVPR'2412.3761.8611.27108.21
VLCounter [20]CLIPAAAI'2418.0665.1317.05106.16
VLCounter+OursCLIP-17.88(-0.18)63.88(-1.25)16.56(-0.49)107.47(+1.31)
ClipCount [18]CLIPMM'2318.9366.1017.45104.88
ClipCount+OursCLIP-17.52(-1.41)58.39(-7.71)16.42(-1.03)100.63(-4.25)
CountGD [3]GDinoNeurIPS'2412.1447.5114.76120.42
CountGD+OursGDino-11.62(-0.52)46.37(-1.14)12.84(-1.92)115.61(-4.81)
GDino [27](w/ finetune)*GDinoECCV'2412.0762.1311.76104.98
GDino(w/ finetune)+OursGDino-11.09(-0.98)60.48(-1.65)10.74(-1.02)106.5(+1.52)
", + "bbox": [ + 91, + 88, + 919, + 347 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison with state-of-the-art methods on FSC-147 dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes.", + "bbox": [ + 89, + 347, + 906, + 376 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initially design 5 distinct number ranges and leverage these ranges to generate 5 different number-conditioned text prompts such as \"There are between {number} and {number} {class}\" . Then, a global-number similarity loss is proposed to maximize the similarity between the global image-level feature and the positive number-conditioned text prompt which has the closest number range. The global-number similarity loss can be formulated as:", + "bbox": [ + 89, + 387, + 483, + 500 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s i m} = - \\log \\frac {\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right)}{\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right) + \\sum_ {j \\in \\mathcal {N}} \\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {j}\\right) / \\tau_ {2}\\right)}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 505, + 486, + 549 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $s(\\cdot)$ denotes the cosine similarity, $\\tau_{2}$ is a temperature parameter and $F_{ct}^{i}$ is the positive number-conditioned text prompt embedding. $\\mathcal{N}$ denotes the set of negative number-conditioned text prompt embeddings. For CLIP-based methods, we take the global `cls_token` as our $F_{g}$ , while for GroundingDino[27], we obtain the $F_{g}$ by using average pooling on the last layer features of Swin-T.", + "bbox": [ + 89, + 551, + 483, + 655 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Through above global-number alignment scheme, model can learn to find a coarse number range for an image. But this coarse range is still too wide to serve as an accurate prompt. Hence, as shown in Fig. 3, we further propose a coarse-to-fine global-number alignment scheme to generate more accurate number-conditioned text prompt. The number-conditioned text prompts are designed to be hierarchical, which means that the coarse range with highest similarity is further divided into 5 ranges, and the global-number similarity loss is adopted again to maximize the similarity of the global image-level feature and the positive fine number-conditioned text prompt.", + "bbox": [ + 88, + 657, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Overall, the loss of our NGA can be formulated as:", + "bbox": [ + 109, + 838, + 447, + 852 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {N G A} = \\delta L _ {s i m} ^ {\\text {c o a r s e}} + (1 - \\delta) L _ {s i m} ^ {\\text {f i n e}}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 859, + 482, + 880 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{sim}^{coarse}$ is the global-number similarity loss of coarse", + "bbox": [ + 89, + 885, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "number-conditioned text prompts and $L_{sim}^{fine}$ is for fine number-conditioned text prompts. $\\delta$ is a weight term to balance the effect of coarse and fine similarity loss.", + "bbox": [ + 511, + 386, + 906, + 431 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With our proposed scheme, $P_{ct}^{h}$ is selected and considered to have the closest number range. The feature of it $F_{ct}^{h}$ is used to compute cross-attention with $F_{g}$ in NGA, thereby the counting ability of VLMs is enhanced.", + "bbox": [ + 511, + 431, + 906, + 494 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 508, + 645, + 526 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Firstly, the experimental setting, including implementation details and evaluation metrics, is given in Sec. 5.1. Then, we apply our strategy to three different methods and compare their performances with state-of-the-art text-guided zero-shot object counting approaches in Sec. 5.2. Through ablation studies in Sec. 5.3, we quantitatively evaluate the effectiveness of each component in our strategy. Next, we conduct experiments to discuss the impact of the coarse-to-fine global-number alignment scheme, the effect of our strategy on cross-dataset generalization and the influence of $\\delta$ in Sec. 5.4. Finally, visualizations in Sec. 5.5 qualitatively verify the efficacy of our strategy.", + "bbox": [ + 511, + 534, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 511, + 727, + 714, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. To evaluate our strategy comprehensively, we apply our strategy to two CLIP-based methods [18, 20] and one GroundingDino(GDino)-based method [27]. For CLIP-based methods, the vision encoder is VIT-B/16 [11]. VPT [17] is used with 20 visual prompts in each layer. The maxpooling in TLR adopts a kernel size of $16 \\times 16$ , stride=16 and padding=0. For the density decoder, we use $3 \\times 3$ convolution kernel with stride=1 and padding=1 to keep resolution. The batch size is 32 and the overall training epochs are 440. For GDino-based method,", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21102", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5d15b7a34bc40a56b652c5f6293f505209dc0a7a65b9eac5341b13d7a18497ee.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneVal setTestnormal setTestlongtail set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ClipCount* [18]CLIP17.5330.4218.6838.6122.6553.79
ClipCount+OursCLIP16.65(-0.88)29.82(-0.6)17.24(-1.44)36.92(-1.69)21.90(-0.75)53.66(-0.13)
GroundingREC* [8]GDino14.3425.0113.7828.9616.6541.80
GDino* [27](w/ finetune)GDino13.9221.6414.2329.6816.5738.85
GDino(w/ finetune)+OursGDino11.17(-2.75)17.30(-4.34)12.39(-1.84)27.75(-1.93)14.68((-1.89)38.10((-0.75)
", + "bbox": [ + 91, + 88, + 903, + 200 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/41debf986d6be280bcc0af8202ea5e7988580e1744044f0af6c94cd37109aa49.jpg", + "table_caption": [ + "Table 3. Comparison with state-of-the-art methods on our ZSC-8K dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes." + ], + "table_footnote": [], + "table_body": "
MethodVal setTest set
MAERMSEMAERMSE
None18.9366.1017.45104.88
TLR18.4260.7516.71100.08
NGA18.2962.9116.81103.21
TLR+NGA17.5258.3916.42100.63
", + "bbox": [ + 91, + 237, + 486, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Ablation study on proposed NGA and TLR.", + "bbox": [ + 129, + 339, + 442, + 353 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the text encoder is BERT [9] and visual encoder is SwinT [28]. The overall training epochs are 30. And the other hyperparameters are same with [8]. We use AdamW optimizer to optimize our model, for CLIP-based methods, the learning rate is 1e-4 and for GDino-based method, the learning rate is 1e-5. For all methods, $\\delta$ is set as 0.5. $\\lambda$ is set as 10. Following [18], $\\tau_{1}$ and $\\tau_{2}$ are set as 0.07.", + "bbox": [ + 89, + 369, + 482, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation metrics. Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are widely used evaluation metrics in text-guided zero-shot counting tasks which is also applied in this paper. MAE = $\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|$ and RMSE = $\\sqrt{\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|^2}$ , where $N$ is the total number of testing images, $C_m^{pred}$ and $C_m^{gt}$ are the prediction and the ground truth count of the $m$ -th image.", + "bbox": [ + 89, + 476, + 482, + 592 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Compare with State-of-the-art Methods", + "text_level": 1, + "bbox": [ + 89, + 604, + 431, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we apply the proposed strategy to three existing approaches and compare the performances with state-of-the-art text-guided zero-shot object counting methods on two datasets: FSC-147 [38] and ZSC-8K, to evaluate the effectiveness of our strategy.", + "bbox": [ + 89, + 627, + 482, + 702 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results on FSC-147 are shown in Tab. 2. As it depicts, our approach outperforms all previous state-of-the-art methods, regardless of the backbone. Compared with the best performances of previous methods, our approach reduces the MAE by $8.1\\%$ and $4.7\\%$ on the validation set and test set. And our method decreases the RMSE by $2.5\\%$ and $2.7\\%$ on the validation set and test set respectively. Our method also shows its scalability by comparisons with the baselines, the performances of VLCounter+Ours, ClipCount+Ours, CountGD+Ours and GDino(w/ finetune)+Ours are all better than their baselines.", + "bbox": [ + 89, + 703, + 482, + 868 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Then, we compare the performance of our method with previous ZSC methods on our proposed dataset, ZSC-8K.", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results in Tab. 3 demonstrate that our method obtains the best performance on ZSC-8K dataset. Specifically, compared to the previous best results, the MAE is reduced by $19.8\\%$ and $10.1\\%$ on validation set and testnormal set. And every baseline is improved by our strategy. Furthermore, we evaluate ZSC methods on the testlongtail set in the last two columns of Tab. 3. Although our method still achieves optimal performance, the results indicate a significant performance gap between normal and long-tailed classes, highlighting the limitation of ZSC methods.", + "bbox": [ + 511, + 239, + 903, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The remarkable results on the FSC-147 and ZSC-8K datasets illustrate the effectiveness of our strategy. In particular, whether CLIP or GDino is used as the backbone, our method significantly outperforms the baselines, which strongly validates the scalability of our strategy.", + "bbox": [ + 511, + 398, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 502, + 666, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct ablation experiments to verify the effectiveness of the proposed Text-guided Local Ranking(TLR) and Number-Evoked Global Attention(NGA) in Tab. 4.", + "bbox": [ + 511, + 531, + 903, + 577 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments are conducted with ClipCount [18] on FSC-147 [38]. We start with a baseline where no modules are applied. It achieves 18.93 MAE and 66.10 RMSE on validation set, 17.45 MAE and 104.88 RMSE on test set. Then, the single use of TLR and NGA is analyzed. By leveraging the proposed local-text rank loss, the foreground and background patches are pushed apart and the localization ability of VLMs is improved by TLR. The effectiveness is verified by the results shown in Tab. 4. Compared to baseline, TLR improves the MAE by $2.7\\%$ on validation set and $4.2\\%$ on test set. Next, the proposed NGA is adopted to study its effectiveness. With NGA which calculates crossattention between the number-conditioned text prompt and global image-level feature, its performance significantly improved, the error reduces by $3.7\\%$ and $1.6\\%$ for MAE and RMSE on test set, respectively. Hence, the results verify that the proposed NGA can enhance the counting ability of VLMs. Finally, the combination of TLR and NGA enhances the overall performance even further, improving the MAE by $7.4\\%$ on validation set and $5.9\\%$ on test set, suggesting the proposed TLR and NGA are complementary.", + "bbox": [ + 511, + 583, + 903, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21103", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/65c9bf6b490f1daa83d949f7a62007f4a43bba38f03edf3975c37ad1d6825a45.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVal setTest set
MAERMSEMAERMSE
w/ coarse18.1761.7317.34100.79
w/ coarse-to-fine17.5258.3916.42100.63
", + "bbox": [ + 91, + 88, + 486, + 160 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/437dcb114426ed314d493750b4570b0ffd8d6d254c7842c09824a84975f8dbef.jpg", + "table_caption": [ + "Table 5. The effect of the proposed coarse-to-fine global-number alignment scheme." + ], + "table_footnote": [ + "Table 6. Cross-dataset evaluation on ShanghaiTech dataset. \"SHA\", \"SHB\" and \"FSC\" denote ShangTech Part A, ShanghaiTech Part B and FSC-147 datasets respectively." + ], + "table_body": "
MethodTypeTraining→TestingMAERMSE
MCNN [55]SpecificSHB→SHA221.4357.8
CrowdClip [23]SpecificSHB→SHA217.0322.7
RCC [13]GenericFSC→SHA240.1366.9
ClipCount [18]GenericFSC→SHA192.6308.4
ClipCount+OursGenericFSC→SHA176.24279.20
", + "bbox": [ + 91, + 191, + 491, + 292 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Discussions", + "text_level": 1, + "bbox": [ + 89, + 348, + 215, + 363 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of coarse-to-fine global-number alignment scheme. A coarse-to-fine global-number alignment scheme is designed to select accurate number-conditioned text prompt to enhance the counting ability of VLMs in NGA. We conduct experiments to study the effect of this scheme. \"w/ coarse\" means we only use coarse number-conditioned text prompts. All experiments are performed on the FSC-147 dataset with ClipCount [18].", + "bbox": [ + 88, + 371, + 482, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "As Tab. 5 shows, with the proposed scheme, it achieves 17.52 MAE on validation set and 16.42 MAE on test set which are better than only applying coarse number-conditioned text prompts. The results prove the effectiveness of our coarse-to-fine global-number alignment scheme. The generalizing ability across datasets. To evaluate the cross-dataset generalization ability, we compare our approach with ZSC methods on the ShanghaiTech [55] crowd counting dataset. On ShanghaiTech dataset, we compare with specific methods which are category-specific and generic methods that are ZSC methods. The results are summarized in Tab. 6. Our proposed generic object counting approach outperforms both category-specific counting methods and ZSC methods. It surpasses CrowdClip [23] by 40.76 MAE and 42.8 RMSE without training on crowd counting dataset. Moreover, compared to the second best ZSC method, it gains 16.36 MAE improvement and 29.2 RMSE improvement.", + "bbox": [ + 89, + 492, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, above results demonstrate that our strategy can significantly improve the cross-dataset generalizing ability. The effect of $\\delta$ . $\\delta$ is used to balance the effect of coarse and fine similarity losses. Experiments are conducted on FSC-147 dataset with ClipCount[18] to analyze it. As Fig. 5 shows, $\\delta = 0.5$ achieves the best performance on both val and test sets. It is also indicated that a small value of $\\delta$ (0.1) may be inadequate to find a accurate coarse range, while a large value (0.9) would lose fine-grained information.", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab6934e464cf5976a0a0ba2feeadb048f93a60aa63a700748f60248a362b7223.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 88, + 903, + 236 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/1e738f1126779289492e43953de7de0b2bb63be4a318543018b51710012f3bc7.jpg", + "image_caption": [ + "Figure 6. Qualitative results on FSC147 and ZSC-8K. Ground-truth numbers are shown at the top-right corner." + ], + "image_footnote": [], + "bbox": [ + 511, + 237, + 903, + 415 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Visualization", + "text_level": 1, + "bbox": [ + 513, + 444, + 650, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To examine the effect of our approach qualitatively, we visualize some results predicted by ClipCount [18] and our approach in Fig. 6. As the Fig. 6 shows, the results of our method are closer to the ground-truth number and our method can detect the objects that are missed by ClipCount. These visualizations further verify that our approach can improve the counting ability and localization ability of VLMs. More qualitative results are shown in supplement.", + "bbox": [ + 511, + 465, + 906, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 594, + 633, + 609 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a universal strategy for text-guided zero-shot object counting. We introduce Text-guided Local Ranking to improve the localization ability of VLMs by a local-text rank loss, which effectively push the features of foreground and background patches apart. Additionally, we propose Number-evoked Global Attention, which leverages the global image-level feature to align with multiple number-conditioned text prompts, and then selects the prompt with highest similarity to calculate crossattention with the global image-level feature, thereby enhancing the counting ability. Furthermore, we propose a dataset named ZSC-8K for comprehensive evaluation. Extensive experiments conducted on the ZSC-8K dataset and three other public datasets demonstrate the effectiveness, scalability, and cross-dataset generalizability of our method. Although effective, the results on the long-tailed test set reveal substantial performance gaps when handling long-tailed classes. This persistent challenge across all ZSC methods highlights a critical direction for future research.", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21104", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 240, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Nature Science Foundation of China (NSFC) under Grant 62376209; and in part by the Fundamental Research Funds for the Central Universities under Grant xtr072024017.", + "bbox": [ + 89, + 114, + 482, + 175 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 190, + 186, + 205 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hüseyin Gökhan Akçay, Bekir Kabasakal, Duygugül Aksu, Nusret Demir, Melih Öz, and Ali Erdoğan. Automated bird counting with deep learning for regional bird distribution mapping. Animals, 10(7):1207, 2020. 2", + "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. BMVC, 2023. 3, 6", + "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. NeurIPS, 2024. 3, 6, 2", + "[4] Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. Interactive object counting. In ECCV, pages 504-518. Springer, 2014. 3", + "[5] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. In ECCV, pages 483-498. Springer, 2016. 2", + "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Ming-Hsuan Yang, SyYen Kuo, et al. Improving point-based crowd counting and localization based on auxiliary point guidance. ECCV, 2024. 2", + "[7] Ke Chen, Shaogang Gong, Tao Xiang, and Chen Change Loy. Cumulative attribute space for age and crowd density estimation. In CVPR, pages 2467-2474, 2013. 3, 1", + "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In CVPR, pages 16985-16995, 2024. 2, 3, 6, 7", + "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 7", + "[10] SongLin Dong, ChengLi Tan, ZhenTao Zuo, YuHang He, YiHong Gong, TianGang Zhou, JunMin Liu, and JiangShe Zhang. Brain-inspired dual-pathway neural network architecture and its generalization analysis. Science China Technological Sciences, 67(8):2319-2330, 2024. 1", + "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6", + "[12] Shenjian Gong, Shanshan Zhang, Jian Yang, Dengxin Dai, and Bernt Schiele. Class-agnostic object counting robust to intraclass diversity. In ECCV, pages 388–403. Springer, 2022. 3", + "[13] Michael Hobley and Victor Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. arXiv preprint arXiv:2205.10203, 2022. 8, 2", + "[14] M. Hobley and V. Prisacariu. Abc easy as 123: A blind counter for exemplar-free multi-class class-agnostic counting. In ECCV, 2024. 3" + ], + "bbox": [ + 93, + 215, + 482, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In ICCV, pages 4145-4153, 2017. 2, 3, 1", + "[16] Yifeng Huang, Viresh Ranjan, and Minh Hoai. Interactive class-agnostic object counting. In ICCV, pages 22312–22322, 2023. 3", + "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 6", + "[18] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, pages 4535-4545, 2023. 2, 3, 5, 6, 7, 8", + "[19] Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In CVPR, pages 4705-4714, 2020. 2", + "[20] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI, pages 2714-2722, 2024. 3, 6, 2", + "[21] Victor S. Lempitsky and Andrew Zisserman. Learning to count objects in images. In NIPS, pages 1324-1332, 2010. 2", + "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes. In CVPR, pages 1091-1100, 2018. 2", + "[23] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. In CVPR, pages 2893-2903, 2023. 8", + "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, pages 19596-19605, 2022. 2", + "[25] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. BMVC, 2022. 3", + "[26] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In ICCV, pages 1676-1685, 2023. 2", + "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 6, 7", + "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 7", + "[29] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In ACCV, pages 669–684. Springer, 2019. 3", + "[30] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In ICCV, pages 6141-6150, 2019. 1, 2", + "[31] Thomas Moranduzzo and Farid Melgani. Automatic car counting method for unmanned aerial vehicle images. IEEE Transactions on Geoscience and remote sensing, 52(3): 1635-1647, 2013. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21105", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[32] T Nathan Mundhenk, Goran Konjevod, Wesam A Sakla, and Kofi Boakye. A large contextual dataset for classification, detection and counting of cars with deep learning. In ECCV, pages 785-800. Springer, 2016. 2", + "[33] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In ECCV, pages 348-365. Springer, 2022. 3", + "[34] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 3", + "[35] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. CVPR, 2024. 6", + "[36] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems, 37:66260-66282, 2024. 6", + "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3", + "[38] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 1, 2, 3, 4, 7", + "[39] Deepak Babu Sam, Shiv Surya, and R. Venkatesh Babu. Switching convolutional neural network for crowd counting. In CVPR, pages 4031-4039, 2017. 2", + "[40] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In CVPR, pages 9529–9538, 2022. 1, 3", + "[41] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In WACV, pages 323-331, 2024. 3, 6", + "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1", + "[43] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. JHU-CROWD++: large-scale crowd counting dataset and A benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 3, 1", + "[44] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In ICCV, pages 3365-3374, 2021. 1, 2", + "[45] Yuejiao Su, Yi Wang, Lei Yao, and Lap-Pui Chau. Few-shot class-agnostic counting with occlusion augmentation and localization. In ISCAS, pages 1-5. IEEE, 2024. 3", + "[46] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. NWPU-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2141-2149, 2021. 3, 1", + "[47] Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Yang, and Siwei Lyu. Ua-detrac: A new benchmark and protocol for multi-object detection and tracking. Computer Vision and Image Understanding, 193:102907, 2020. 3, 1", + "[48] Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional regression networks. Computer methods in biomechanics and biomedical engineering: Imaging & Visualization, 6(3): 283-292, 2018. 2", + "[49] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In CVPR, pages 15548-15557, 2023. 2, 3, 6", + "[50] Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo, Xiao Tan, Yezhen Wang, Shilei Wen, and Errui Ding. Perspective-guided convolution networks for crowd counting. In ICCV, pages 952-961, 2019. 3, 1", + "[51] Shuo-Diao Yang, Hung-Ting Su, Winston H. Hsu, and Wen-Chin Chen. Class-agnostic few-shot object counting. In WACV, pages 870–878, 2021. 3", + "[52] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. In WACV, pages 6315-6324, 2023. 1, 3", + "[53] Shanghang Zhang, Guanhang Wu, Joao P Costeira, and José MF Moura. Fcn-rlstm: Deep spatio-temporal neural networks for vehicle counting in city cameras. In ICCV, pages 3667-3676, 2017. 2", + "[54] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In ACM Multimedia, 2024. 2", + "[55] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In CVPR, pages 589-597, 2016. 2, 8", + "[56] H.L Zhu, J.L Yuan, Z.W Yang, Y. Guo, Z. Wang, X. Zhong, and S.F He. Zero-shot object counting with good exemplars. In ECCV, 2024. 3, 6", + "[57] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 3, 1" + ], + "bbox": [ + 516, + 90, + 905, + 670 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21106", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_model.json b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_model.json new file mode 100644 index 0000000000000000000000000000000000000000..a00cecf3938ce4f636aefc2877ac51d8b9040638 --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.873, + 0.175 + ], + "angle": 0, + "content": "Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention" + }, + { + "type": "text", + "bbox": [ + 0.338, + 0.203, + 0.66, + 0.221 + ], + "angle": 0, + "content": "Shiwei Zhang\\(^{1,\\dagger}\\), Qi Zhou\\(^{1,\\dagger}\\), Wei Ke\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.235, + 0.221, + 0.764, + 0.24 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Software Engineering, Xi'an Jiaotong University, China" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.239, + 0.664, + 0.257 + ], + "angle": 0, + "content": "\\(^{2}\\)Pengcheng Laboratory, Shenzhen, China" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.259, + 0.775, + 0.275 + ], + "angle": 0, + "content": "{zhangshiwei96,zhouqq}@stu.xjtu.edu.cn,wei.ke@mail.xjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.765 + ], + "angle": 0, + "content": "Text-guided zero-shot object counting leverages vision-language models (VLMs) to count objects of an arbitrary class given by a text prompt. Existing approaches for this challenging task only utilize local patch-level features to fuse with text feature, ignoring the important influence of the global image-level feature. In this paper, we propose a universal strategy that can exploit both local patch-level features and global image-level feature simultaneously. Specifically, to improve the localization ability of VLMs, we propose Text-guided Local Ranking. Depending on the prior knowledge that foreground patches have higher similarity with the text prompt, a new local-text rank loss is designed to increase the differences between the similarity scores of foreground and background patches which push foreground and background patches apart. To enhance the counting ability of VLMs, Number-evoked Global Attention is introduced to first align global image-level feature with multiple number-conditioned text prompts. Then, the one with the highest similarity is selected to compute cross-attention with the global image-level feature. Through extensive experiments on widely used datasets and methods, the proposed approach has demonstrated superior advancements in performance, generalization, and scalability. Furthermore, to better evaluate text-guided zero-shot object counting methods, we propose a dataset named ZSC-8K, which is larger and more challenging, to establish a new benchmark. Codes and dataset are released at https://github.com/zaqai/LGCount." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.793, + 0.222, + 0.808 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.879 + ], + "angle": 0, + "content": "In recent years, with the development of deep learning techniques [10, 37, 42], category-specific counting (CSC) methods [30, 44] have gradually matured. However, existing CSC methods face notable limitations in practical applica" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.31, + 0.905, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Figure 1. Comparison of different pipelines for text-guided zero-shot object counting. (a) Only fusing single text prompt and local patch-level features. (b) Instead of overlooking the global image-level feature, our universal strategy leverages the global image-level feature to align with multiple number-conditioned text prompts and select the one with the highest similarity to calculate cross-attention with image-level feature. At last, the feature obtained by cross-attention is added to the local patch-level features." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.654, + 0.907, + 0.761 + ], + "angle": 0, + "content": "tions. These approaches require training dedicated models on extensive labeled datasets for each object category, so they fundamentally lack the capacity to handle unseen objects in a human-like manner. Therefore, counting methods that are not limited by fixed categories have gradually attracted the attention of the academic community, as they offer greater flexibility and broader application potential." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Many exemplar-guided class-agnostic object counting (CAC) methods [38, 40, 52] have been presented in this context. These methods take several human-annotated bounding boxes of objects as exemplars to specify the objects-of-interest and train a unified network that can estimate the number of arbitrary objects-of-interest. However, previous CAC methods heavily rely on the assumption that accurate bounding boxes of objects can be easily obtained. In practice, this requirement is unfriendly to end-users lacking pro" + }, + { + "type": "page_footnote", + "bbox": [ + 0.115, + 0.887, + 0.361, + 0.901 + ], + "angle": 0, + "content": "† Equal contribution. * Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21097" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.122 + ], + "angle": 0, + "content": "fessional knowledge and is even impossible to fulfill for a fully automatic system (e.g. a wildlife monitoring system)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.125, + 0.483, + 0.381 + ], + "angle": 0, + "content": "To provide a more flexible object counting approach, text-guided zero-shot object counting (ZSC) has been proposed by [49] and further developed by [8, 18] which leverage vision-language models (VLMs) [27, 37]. Under this setting, the bounding boxes of exemplars are replaced by the class name. During inference, users only need to specify the class name of the object that they want to count, and the model can automatically output the number of objects of this class. Even if the class is completely unseen during the training stage, the counter should still adapt to this unseen class without any additional data. Nevertheless, as Fig. 1(a) shows, existing ZSC methods only utilize local patch-level features from vision encoder to fuse with text feature and then use the fused features to make predictions. The global image-level feature, which is also derived from the vision encoder, is discarded, and its significance is completely overlooked by previous methods." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.385, + 0.483, + 0.883 + ], + "angle": 0, + "content": "In this paper, we propose a universal strategy which exploits both local patch-level features and global image-level feature to enhance text-guided zero-shot object counting as shown in Fig. 1(b). It consists of two parts: Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). In TLR, we reveal a critical limitation in prior method [18]: its utilization of local patch-level features inadvertently reduces the discriminative gap between foreground and background similarity scores, which fundamentally undermines VLMs' localization performance. To alleviate this limitation, we leverage the prior knowledge that the similarity scores of foreground patches must be greater than those of background features to design a novel local-text rank loss. This loss is designed to increase the differences between the foreground and background similarity scores which push the foreground and background patches apart to strengthen the localization ability of VLMs. In NGA, we exploit the global image-level feature and propose multiple number-conditioned text prompts. Unlike previous single text prompt that merely specify the class, our multiple number-conditioned text prompts not only indicate the class but also have predefined number ranges. And a similarity loss is proposed to align the global image-level feature with the prompt that has the closest number range. Then, to find a more accurate range, a coarse-to-fine global-number alignment scheme is proposed. In the end, the number-conditioned text prompt with the highest similarity is selected to calculate cross-attention with global image-level feature. Considering that the text prompt used in VLMs' pre-training stage usually lacks numeric information, by interacting number-conditioned text prompt with global image-level feature, our proposed NGA significantly improves the counting ability of VLMs." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To verify the effectiveness of our proposed strategy, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "conducted experiments on three prevailing datasets: FSC-147 [38], CARPK [15] and ShanghaiTech [55]. Moreover, to perform a more comprehensive evaluation of ZSC methods, we create a larger and more challenging dataset, named ZSC-8K, which increases the proportion of common classes and moves long-tailed categories in FSC-147 to the test set of ZSC-8K to evaluate the performance on long-tailed classes. The results demonstrate that our method significantly surpasses previous text-guided zero-shot object counting methods. Additionally, various backbones including CLIP [37] and GroundingDino [27] are also employed in our experiments to confirm the scalability of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.278, + 0.905, + 0.308 + ], + "angle": 0, + "content": "In summary, the main contributions of the proposed approach can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.313, + 0.904, + 0.388 + ], + "angle": 0, + "content": "- We develop a universal strategy consists of Text-guided Local Ranking and Number-evoked Global Attention. This strategy is capable of exploiting both local patch-level features and global image-level feature for text-guided zero-shot object counting." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.389, + 0.905, + 0.493 + ], + "angle": 0, + "content": "- Text-guided Local Ranking is proposed to improve the localization ability of VLMs by using a local-text rank loss to push the foreground and background patches apart. Number-evoked Global Attention is presented to enhance the counting ability of VLMs by computing crossattention between global image-level feature and number-conditioned text prompt with highest similarity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.495, + 0.905, + 0.538 + ], + "angle": 0, + "content": "- We introduce ZSC-8K, a scaled-up dataset consists of 8,361 images and 114 categories for comprehensive evaluation, which can benefit future researchers in this area." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.54, + 0.905, + 0.599 + ], + "angle": 0, + "content": "- The proposed strategy achieves state-of-the-art performance on ZSC-8K dataset and three other popular datasets. Its effectiveness, scalability and cross-dataset generalizability are verified by experiments." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.313, + 0.905, + 0.599 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.628, + 0.663, + 0.643 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.806, + 0.674 + ], + "angle": 0, + "content": "2.1. Category-specific object counting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.905, + 0.851 + ], + "angle": 0, + "content": "Class-specific object counting (CSC) concentrates on counting objects of one specific category, for instance, humans [6, 19, 21, 24, 26, 30, 44, 55], penguins [5], birds [1], cells [48], or cars [15, 31, 32, 53]. Generally, existing CSC approaches can be classified into two groups: point-based methods and density-based methods. Point-based methods [6, 26, 44, 54] apply an object detector to the image and count the number of objects based on the detected points. Density-based methods [22, 39, 55] predict a density map for each input image, and the final result is obtained by summing up the pixel values." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "However, since existing CSC methods need to train a specific network with extensive labeled data for each class of object, they are deemed time-consuming and laborious." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21098" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.443, + 0.108 + ], + "angle": 0, + "content": "2.2. Exemplar-guided class-agnostic counting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.113, + 0.484, + 0.204 + ], + "angle": 0, + "content": "To ease the burden of humans and enable the counting model to count arbitrary categories, researchers propose exemplar-guided class-agnostic counting (CAC) [4, 12, 16, 25, 29, 33, 40, 51, 52] which leverages human annotated images or patches as exemplars of a category to train a unified network that can count arbitrary categories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.204, + 0.484, + 0.431 + ], + "angle": 0, + "content": "[12] indicates that color, shape and scale diversity are crucial for CAC performance and proposes an EFA module to handle diversity from all aspects. [16] proposes an interactive class-agnostic object counting framework. FamNet [38] presents an adaptation strategy to adapt network to any novel visual category at test time. [33] introduces an uncertainty-aware object detector. BMNet [40] proposes a similarity-aware CAC framework. CFOCNet [51] uses reference images as exemplars to reduce the huge cost of data collection. [52] encourages the model to inspect the query image by focusing more on the regions akin to the support images. CountGD [3] leverages both visual exemplar and text prompt to boost the model. [45] extracts multiple occluded features of exemplars for comparison and utilizes the position of instances for more accurate counting results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.431, + 0.484, + 0.522 + ], + "angle": 0, + "content": "CAC methods usually assume that high-quality bounding boxes of objects can be obtained easily. However, obtaining these exemplars is not as easy as assumed. In practice, the requirement of providing exemplars is unfriendly to users who lack professional knowledge and will be an obstacle when we want to build a fully automatic system." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.532, + 0.418, + 0.548 + ], + "angle": 0, + "content": "2.3. Text-guided zero-shot object counting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.629 + ], + "angle": 0, + "content": "Considering that high-quality exemplars are not always easily obtained and aiming to construct a user-friendly system, text-guided zero-shot object counting (ZSC) [2, 8, 18, 20, 41, 49] is proposed to provide a more flexible and user-friendly approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.483, + 0.87 + ], + "angle": 0, + "content": "[49] first introduces ZSC and trains a VAE model to generate the class prototype. The generated prototype is used to select class-relevant patches as exemplars to train a counting network. Later, with the application of vision-language models (VLMs) [27, 37], even automatically obtained exemplars are no longer needed. [2, 18, 20, 34] leverage CLIP [37] to solve ZSC task. [18] proposes an end-to-end pipeline to estimate density maps for a given class, [2] introduces a feature interaction module to enhance the feature and [20] incorporates SPT, LAT, and SaSC modules to tailor VLBase for object counting. ABC123 [14] proposed a blind counter but can only count images with fewer than 300 objects. Furthermore, [8] employs GroundingDino [27] as its backbone, and once again, the boundary of ZSC is advanced. VA-Count [56] then uses GroundingDino to generate good exemplars to boost the model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Nonetheless, previous ZSC approaches only leverage local patch-level features from the vision encoder to interact" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.231, + 0.907, + 0.26 + ], + "angle": 0, + "content": "Figure 2. Distribution of each category in FSC-147 [38] (left) and the proposed ZSC-8K (right)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.262, + 0.922, + 0.365 + ], + "angle": 0, + "content": "
#Images#ClassesMaxMinAvgTotal
Overall83611141766146.23386498
Train5476301766149.72272243
Val112729485135.1639623
Testnormal139739953137.1051829
Testlongtail175857953142.4574632
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.513, + 0.365, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Table 1. Statistics of each set in ZSC-8K, including the number of point annotations (max, min, avg, total)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.408, + 0.907, + 0.469 + ], + "angle": 0, + "content": "with text feature. The global image-level feature is discarded, and its significance is largely overlooked. It still lacks a systematic method to utilize both local and global features to enhance ZSC approaches." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.677, + 0.497 + ], + "angle": 0, + "content": "3. ZSC-8K Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.719 + ], + "angle": 0, + "content": "FSC-147 [38] dataset is the most widely-used dataset for text-guided zero-shot object counting (ZSC). However, it demonstrates an unrealistic class distribution characterized by excessive long-tailed categories (polka dots, cereals, shallots, etc.) and insufficient samples for common classes. This distributional bias deviates markedly from real-world occurrence patterns. Therefore, to conduct a more comprehensive evaluation of ZSC methods, we propose ZSC-8K dataset, inspired by REC dataset [8]. As shown in Fig. 2, compared to FSC-147, ZSC-8K is larger and more challenging which means it increases the proportion of common classes (e.g. cars, people) and moves long-tailed classes from FSC-147 to test set to evaluate the performance of ZSC methods on long-tailed classes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Data collection. Images in the ZSC-8K dataset are collected from multiple existing datasets, including FSC147 [38], JHU-Crowd [43], NWPU-Crowd [46], VisDrone [57], DETRAC [47], CARPK [15], MALL [7] and Crowd Surveillance [50]. The proportions of each dataset are depicted in the supplementary material. These datasets belong to different domains, including crowd counting, car counting and animal counting, etc." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Dataset statistics. ZSC-8K contains 8,361 images with a total of 386,498 point annotations. The number of point annotations in an image ranges from 1 to 1,766. A histogram of point annotations per image is provided in the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21099" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.538, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.54, + 0.09, + 0.902, + 0.328 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.095, + 0.33, + 0.899, + 0.351 + ], + "angle": 0, + "content": "Ft Text feature F1 Local patch-level features Fg Global image-level feature Fct Feature with highest similarity Pct Prompt of Fct" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.36, + 0.907, + 0.457 + ], + "angle": 0, + "content": "Figure 3. Illustration of our training pipeline. In Text-guided Local Ranking, depending on the prior knowledge of similarity score ranking, local-text rank loss is proposed to push the foreground and background patches apart which further improves the localization ability of VLMs. In Number-evoked Global Attention, a coarse-to-fine global-number alignment scheme is proposed to align global image-level feature with multiple coarse number-conditioned text prompts firstly, and with multiple fine number-conditioned text prompts secondly. The fine number-conditioned text prompt with the highest similarity score is selected to calculate cross-attention with image-level feature. And the feature obtained by cross-attention is added to the local patch-level features. In inference stage, the test images and their corresponding number-conditioned text prompts are forwarded to decoder to make predictions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.471, + 0.486, + 0.684 + ], + "angle": 0, + "content": "supplementary material. The dataset can be divided into 5 major categories and 114 subcategories. The proportions of each category are shown in Fig. 2. As it depicts, people are the category with the highest proportion and cars are the second which are significantly higher than that in FSC-147 dataset [38]. We split ZSC-8K dataset into train, validation and test sets. Specific statistics for each set are detailed in Tab. 1. It is worth noting that two versions of test sets are proposed: \\( \\text{test}_{\\text{normal}} \\) and \\( \\text{test}_{\\text{longtail}} \\). In \\( \\text{test}_{\\text{normal}} \\), images of normal classes (1397 images & 39 classes) are included, while \\( \\text{test}_{\\text{longtail}} \\) is obtained by merging the long-tailed classes (361 images & 20 classes) from FSC-147 into \\( \\text{test}_{\\text{normal}} \\). To meet the zero-shot setting, the categories in train set do not overlap with validation and test sets." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.701, + 0.226, + 0.718 + ], + "angle": 0, + "content": "4. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.202, + 0.741 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.485, + 0.901 + ], + "angle": 0, + "content": "The goal of ZSC is to count anything using only text guidance. Pretrained vision-language models (VLMs) are commonly used for this purpose. Denote \\( I \\in R^{H \\times W \\times 3} \\) as an input image containing an arbitrary object class, and a text prompt \\( t \\) is given to specify the object class of this image. In previous methods, \\( I \\) is initially fed into the image encoder of pretrained VLMs to obtain local patch-level features and \\( t \\) is then encoded by the text encoder to obtain text feature. Subsequently, a fuse module is designed to combine the two modalities, and the fused feature is used to predict the num" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.593 + ], + "angle": 0, + "content": "ber of the specified category. In this paper, we argue that previous methods only leverage local patch-level features to fuse with the text feature, thereby fully overlooking the importance of the global image-level feature. Therefore, to exploit local patch-level features and global image-level simultaneously, we propose a strategy consists of Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). The overall framework is shown in Fig. 3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Specifically, the local patch-level features and text feature are used to compute a novel local-text rank loss that can increase the differences between foreground and background similarity scores in TLR. The localization ability of VLMs is improved by TLR. Then, we leverage the global image-level feature in NGA. Multiple number-conditioned text prompts are designed and a coarse-to-fine global-number alignment scheme is proposed to select the number-conditioned text prompt with highest similarity. The selected prompt is considered to have the closest number range and its feature is used to calculate cross-attention with the global image-level feature to enhance the counting ability. Finally, the feature obtained by cross-attention is aggregated with local patch-level features through element-wise summation. This combined representations is then fed into decoder to make predictions. TLR and NGA are detailed in Sec. 4.2 and Sec. 4.3, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21100" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.338, + 0.108 + ], + "angle": 0, + "content": "4.2. Text-guided Local Ranking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.117, + 0.484, + 0.42 + ], + "angle": 0, + "content": "VLMs have shown impressive performance in estimating text-image similarity by aligning text and image representations. However, as demonstrated by ClipCount [18], pretrained VLMs intrinsically lack localization ability, making it non-trivial to transfer this ability to object counting. Although ClipCount [18] utilized the local patch-level features \\( F_{l} \\) and the text feature \\( F_{t} \\) to construct a contrastive loss to enhance localization ability by pushing foreground and background patches apart, this loss cannot effectively increase the discrepancy between foreground and background similarity scores, which in turn adversely affects localization ability. As shown in Fig. 4(a), the differences between foreground and background similarity scores in ClipCount [18] are small, with most differences are close to zero. To alleviate this issue, we propose Text-guided Local Ranking (TLR) in which a novel local-text rank loss is designed to improve the localization ability by increasing the differences between foreground and background similarity scores. As Fig. 4(b) shows, with TLR, the differences become much larger than those of ClipCount [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.424, + 0.484, + 0.56 + ], + "angle": 0, + "content": "As shown in Fig. 3, \\( F_{l} \\) and \\( F_{t} \\) are also used in the proposed TLR, a linear projection maps them to the same channels. To determine which patch contains the object, max-pooling is used to transfer the ground truth density map to a patch-level binary mask. Patches whose value is equal to 1 are foreground patches which have the object and the value of background patches is 0. We denote the foreground patches as \\( \\mathcal{P} \\) and the background patches as \\( \\mathcal{B} \\), the relation of \\( \\mathcal{P} \\) and \\( \\mathcal{B} \\) can be summarized as:" + }, + { + "type": "equation", + "bbox": [ + 0.151, + 0.582, + 0.483, + 0.6 + ], + "angle": 0, + "content": "\\[\ns \\left(F _ {l} ^ {p}, F _ {t}\\right) > s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\forall p \\in \\mathcal {P}, \\forall q \\in \\mathcal {B}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.619, + 0.484, + 0.71 + ], + "angle": 0, + "content": "where \\( s(,) \\) denotes the cosine similarity. And Formula 1 indicates the prior knowledge that the similarity score between any foreground patch feature and the text feature is greater than that of any background patch feature. Depending on the prior knowledge, we construct the local-text rank loss. The similarity score sets are derived as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.732, + 0.483, + 0.75 + ], + "angle": 0, + "content": "\\[\nS (\\mathcal {P}) = \\{\\dots , s \\left(F _ {l} ^ {p}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {P} | = l _ {1}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.798, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\nS (\\mathcal {B}) = \\{\\dots , s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {B} | = l _ {2}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( S(\\mathcal{P}) \\) and \\( S(\\mathcal{B}) \\) are similarity score sets of foreground and background patches respectively. \\( l_{1} \\) and \\( l_{2} \\) are the lengths of the two sets. \\( \\tau_{1} \\) is a temperature parameter. Next, we denote \\( d_{pq} = (s(F_l^p,F_t) - s(F_l^q,F_t)) / \\tau_1 \\), and the" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.204, + 0.907, + 0.288 + ], + "angle": 0, + "content": "Figure 4. The differences between foreground and background similarity scores. (a) shows the differences between foreground and background scores in ClipCount [18] are small, most of the differences are close to 0. (b) shows the differences with our TLR become larger. Hence, the features of foreground and background patches are pushed apart." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.295, + 0.871, + 0.311 + ], + "angle": 0, + "content": "differences between the two sets can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.32, + 0.907, + 0.416 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} M _ {d i f} = S (\\mathcal {P}) \\stackrel {\\text {b r o a d c a s t - s u b t r a c t}} {\\longrightarrow} S (\\mathcal {B}) \\\\ = \\left[ \\begin{array}{c c c c} d _ {1 1} & d _ {1 2} & \\dots & d _ {1 l _ {2}} \\\\ d _ {2 1} & d _ {2 2} & \\dots & d _ {2 l _ {2}} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ d _ {l _ {1} 1} & d _ {l _ {1} 2} & \\dots & d _ {l _ {1} l _ {2}} \\end{array} \\right] \\quad , \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.424, + 0.906, + 0.47 + ], + "angle": 0, + "content": "where \\( M_{dif} \\) is the result of subtracting \\( S(\\mathcal{B}) \\) from \\( S(\\mathcal{P}) \\) using the broadcasting mechanism. Finally, with the prior knowledge, the local-text rank loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.478, + 0.906, + 0.511 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {r a n k}} = \\frac {\\sum \\operatorname {c l a m p} \\left(\\lambda - M _ {\\text {d i f}} , \\min = 0\\right)}{l _ {1} l _ {2}}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.519, + 0.874, + 0.534 + ], + "angle": 0, + "content": "where \\(\\lambda\\) is a hyperparameter that controls the rank loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.906, + 0.61 + ], + "angle": 0, + "content": "With the proposed local-text rank loss, the differences are increased by ensuring the similarity scores of foreground patches are bigger than background patches. Thus the features of foreground and background patches are pushed apart and the localization ability is improved." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.812, + 0.633 + ], + "angle": 0, + "content": "4.3. Number-Evoked Global Attention" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.761 + ], + "angle": 0, + "content": "Because most of the text prompts used in pre-training stage of VLMs lack descriptions of the number of objects, the performance of VLMs in object counting is largely limited. To mitigate this limitation, we propose Number-Evoked Global Attention (NGA) to enhance the counting ability of VLMs by calculating the cross-attention between the global image-level feature \\( F_{g} \\) and the feature of the number-conditioned text prompt with highest similarity \\( F_{ct}^{h} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.761, + 0.907, + 0.902 + ], + "angle": 0, + "content": "First, \\( F_{g} \\) that is ignored by pervious ZSC methods is leveraged to select \\( P_{ct}^{h} \\), the number-conditioned text prompt with highest similarity, as shown in Fig. 3. Specifically, we first propose a global-number alignment scheme to identify a coarse number range for an image. To achieve this, we replace the single text prompt by multiple number-conditioned text prompts. Different from previous single text prompt which only specify the object class, every prompt of our proposed multiple number-conditioned text prompts has an object class and a predefined number range." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21101" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.92, + 0.348 + ], + "angle": 0, + "content": "
MethodBackboneVenueVal setTest set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ZSC [49]ResNet+VAECVPR'2326.9388.6322.09115.17
TFOC [41]SAMWACV'24--20.10132.83
CountTX* [2]CLIPBMCV'2317.7665.2116.70105.21
VA-Count [56]CLIPECCV'2417.8773.2217.88129.31
GeCo [36]CLIPNeurIPS'2414.8164.9513.30108.72
DAVEprm [35]CLIPCVPR'2415.4852.5714.90103.42
GroundingREC* [8]GDinoCVPR'2412.3761.8611.27108.21
VLCounter [20]CLIPAAAI'2418.0665.1317.05106.16
VLCounter+OursCLIP-17.88(-0.18)63.88(-1.25)16.56(-0.49)107.47(+1.31)
ClipCount [18]CLIPMM'2318.9366.1017.45104.88
ClipCount+OursCLIP-17.52(-1.41)58.39(-7.71)16.42(-1.03)100.63(-4.25)
CountGD [3]GDinoNeurIPS'2412.1447.5114.76120.42
CountGD+OursGDino-11.62(-0.52)46.37(-1.14)12.84(-1.92)115.61(-4.81)
GDino [27](w/ finetune)*GDinoECCV'2412.0762.1311.76104.98
GDino(w/ finetune)+OursGDino-11.09(-0.98)60.48(-1.65)10.74(-1.02)106.5(+1.52)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.348, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Table 2. Comparison with state-of-the-art methods on FSC-147 dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.388, + 0.484, + 0.501 + ], + "angle": 0, + "content": "We initially design 5 distinct number ranges and leverage these ranges to generate 5 different number-conditioned text prompts such as \"There are between {number} and {number} {class}\" . Then, a global-number similarity loss is proposed to maximize the similarity between the global image-level feature and the positive number-conditioned text prompt which has the closest number range. The global-number similarity loss can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.506, + 0.487, + 0.55 + ], + "angle": 0, + "content": "\\[\nL _ {s i m} = - \\log \\frac {\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right)}{\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right) + \\sum_ {j \\in \\mathcal {N}} \\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {j}\\right) / \\tau_ {2}\\right)}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.552, + 0.484, + 0.656 + ], + "angle": 0, + "content": "where \\(s(\\cdot)\\) denotes the cosine similarity, \\(\\tau_{2}\\) is a temperature parameter and \\(F_{ct}^{i}\\) is the positive number-conditioned text prompt embedding. \\(\\mathcal{N}\\) denotes the set of negative number-conditioned text prompt embeddings. For CLIP-based methods, we take the global `cls_token` as our \\(F_{g}\\), while for GroundingDino[27], we obtain the \\(F_{g}\\) by using average pooling on the last layer features of Swin-T." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.658, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Through above global-number alignment scheme, model can learn to find a coarse number range for an image. But this coarse range is still too wide to serve as an accurate prompt. Hence, as shown in Fig. 3, we further propose a coarse-to-fine global-number alignment scheme to generate more accurate number-conditioned text prompt. The number-conditioned text prompts are designed to be hierarchical, which means that the coarse range with highest similarity is further divided into 5 ranges, and the global-number similarity loss is adopted again to maximize the similarity of the global image-level feature and the positive fine number-conditioned text prompt." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.839, + 0.449, + 0.853 + ], + "angle": 0, + "content": "Overall, the loss of our NGA can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.167, + 0.86, + 0.483, + 0.881 + ], + "angle": 0, + "content": "\\[\nL _ {N G A} = \\delta L _ {s i m} ^ {\\text {c o a r s e}} + (1 - \\delta) L _ {s i m} ^ {\\text {f i n e}}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( L_{sim}^{coarse} \\) is the global-number similarity loss of coarse" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.907, + 0.433 + ], + "angle": 0, + "content": "number-conditioned text prompts and \\( L_{sim}^{fine} \\) is for fine number-conditioned text prompts. \\( \\delta \\) is a weight term to balance the effect of coarse and fine similarity loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.907, + 0.495 + ], + "angle": 0, + "content": "With our proposed scheme, \\( P_{ct}^{h} \\) is selected and considered to have the closest number range. The feature of it \\( F_{ct}^{h} \\) is used to compute cross-attention with \\( F_{g} \\) in NGA, thereby the counting ability of VLMs is enhanced." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.509, + 0.646, + 0.527 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Firstly, the experimental setting, including implementation details and evaluation metrics, is given in Sec. 5.1. Then, we apply our strategy to three different methods and compare their performances with state-of-the-art text-guided zero-shot object counting approaches in Sec. 5.2. Through ablation studies in Sec. 5.3, we quantitatively evaluate the effectiveness of each component in our strategy. Next, we conduct experiments to discuss the impact of the coarse-to-fine global-number alignment scheme, the effect of our strategy on cross-dataset generalization and the influence of \\(\\delta\\) in Sec. 5.4. Finally, visualizations in Sec. 5.5 qualitatively verify the efficacy of our strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.728, + 0.715, + 0.744 + ], + "angle": 0, + "content": "5.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation details. To evaluate our strategy comprehensively, we apply our strategy to two CLIP-based methods [18, 20] and one GroundingDino(GDino)-based method [27]. For CLIP-based methods, the vision encoder is VIT-B/16 [11]. VPT [17] is used with 20 visual prompts in each layer. The maxpooling in TLR adopts a kernel size of \\(16 \\times 16\\), stride=16 and padding=0. For the density decoder, we use \\(3 \\times 3\\) convolution kernel with stride=1 and padding=1 to keep resolution. The batch size is 32 and the overall training epochs are 440. For GDino-based method," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21102" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.202 + ], + "angle": 0, + "content": "
MethodBackboneVal setTestnormal setTestlongtail set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ClipCount* [18]CLIP17.5330.4218.6838.6122.6553.79
ClipCount+OursCLIP16.65(-0.88)29.82(-0.6)17.24(-1.44)36.92(-1.69)21.90(-0.75)53.66(-0.13)
GroundingREC* [8]GDino14.3425.0113.7828.9616.6541.80
GDino* [27](w/ finetune)GDino13.9221.6414.2329.6816.5738.85
GDino(w/ finetune)+OursGDino11.17(-2.75)17.30(-4.34)12.39(-1.84)27.75(-1.93)14.68((-1.89)38.10((-0.75)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.204, + 0.905, + 0.233 + ], + "angle": 0, + "content": "Table 3. Comparison with state-of-the-art methods on our ZSC-8K dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.238, + 0.488, + 0.34 + ], + "angle": 0, + "content": "
MethodVal setTest set
MAERMSEMAERMSE
None18.9366.1017.45104.88
TLR18.4260.7516.71100.08
NGA18.2962.9116.81103.21
TLR+NGA17.5258.3916.42100.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.13, + 0.34, + 0.443, + 0.354 + ], + "angle": 0, + "content": "Table 4. Ablation study on proposed NGA and TLR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.476 + ], + "angle": 0, + "content": "the text encoder is BERT [9] and visual encoder is SwinT [28]. The overall training epochs are 30. And the other hyperparameters are same with [8]. We use AdamW optimizer to optimize our model, for CLIP-based methods, the learning rate is 1e-4 and for GDino-based method, the learning rate is 1e-5. For all methods, \\(\\delta\\) is set as 0.5. \\(\\lambda\\) is set as 10. Following [18], \\(\\tau_{1}\\) and \\(\\tau_{2}\\) are set as 0.07." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.483, + 0.593 + ], + "angle": 0, + "content": "Evaluation metrics. Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are widely used evaluation metrics in text-guided zero-shot counting tasks which is also applied in this paper. MAE = \\(\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|\\) and RMSE = \\(\\sqrt{\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|^2}\\), where \\(N\\) is the total number of testing images, \\(C_m^{pred}\\) and \\(C_m^{gt}\\) are the prediction and the ground truth count of the \\(m\\)-th image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.605, + 0.433, + 0.621 + ], + "angle": 0, + "content": "5.2. Compare with State-of-the-art Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.703 + ], + "angle": 0, + "content": "In this section, we apply the proposed strategy to three existing approaches and compare the performances with state-of-the-art text-guided zero-shot object counting methods on two datasets: FSC-147 [38] and ZSC-8K, to evaluate the effectiveness of our strategy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.869 + ], + "angle": 0, + "content": "The results on FSC-147 are shown in Tab. 2. As it depicts, our approach outperforms all previous state-of-the-art methods, regardless of the backbone. Compared with the best performances of previous methods, our approach reduces the MAE by \\(8.1\\%\\) and \\(4.7\\%\\) on the validation set and test set. And our method decreases the RMSE by \\(2.5\\%\\) and \\(2.7\\%\\) on the validation set and test set respectively. Our method also shows its scalability by comparisons with the baselines, the performances of VLCounter+Ours, ClipCount+Ours, CountGD+Ours and GDino(w/ finetune)+Ours are all better than their baselines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Then, we compare the performance of our method with previous ZSC methods on our proposed dataset, ZSC-8K." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.241, + 0.905, + 0.392 + ], + "angle": 0, + "content": "The results in Tab. 3 demonstrate that our method obtains the best performance on ZSC-8K dataset. Specifically, compared to the previous best results, the MAE is reduced by \\(19.8\\%\\) and \\(10.1\\%\\) on validation set and testnormal set. And every baseline is improved by our strategy. Furthermore, we evaluate ZSC methods on the testlongtail set in the last two columns of Tab. 3. Although our method still achieves optimal performance, the results indicate a significant performance gap between normal and long-tailed classes, highlighting the limitation of ZSC methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.476 + ], + "angle": 0, + "content": "The remarkable results on the FSC-147 and ZSC-8K datasets illustrate the effectiveness of our strategy. In particular, whether CLIP or GDino is used as the backbone, our method significantly outperforms the baselines, which strongly validates the scalability of our strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.503, + 0.667, + 0.52 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.905, + 0.578 + ], + "angle": 0, + "content": "We conduct ablation experiments to verify the effectiveness of the proposed Text-guided Local Ranking(TLR) and Number-Evoked Global Attention(NGA) in Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.905, + 0.903 + ], + "angle": 0, + "content": "All experiments are conducted with ClipCount [18] on FSC-147 [38]. We start with a baseline where no modules are applied. It achieves 18.93 MAE and 66.10 RMSE on validation set, 17.45 MAE and 104.88 RMSE on test set. Then, the single use of TLR and NGA is analyzed. By leveraging the proposed local-text rank loss, the foreground and background patches are pushed apart and the localization ability of VLMs is improved by TLR. The effectiveness is verified by the results shown in Tab. 4. Compared to baseline, TLR improves the MAE by \\(2.7\\%\\) on validation set and \\(4.2\\%\\) on test set. Next, the proposed NGA is adopted to study its effectiveness. With NGA which calculates crossattention between the number-conditioned text prompt and global image-level feature, its performance significantly improved, the error reduces by \\(3.7\\%\\) and \\(1.6\\%\\) for MAE and RMSE on test set, respectively. Hence, the results verify that the proposed NGA can enhance the counting ability of VLMs. Finally, the combination of TLR and NGA enhances the overall performance even further, improving the MAE by \\(7.4\\%\\) on validation set and \\(5.9\\%\\) on test set, suggesting the proposed TLR and NGA are complementary." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21103" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.488, + 0.161 + ], + "angle": 0, + "content": "
MethodVal setTest set
MAERMSEMAERMSE
w/ coarse18.1761.7317.34100.79
w/ coarse-to-fine17.5258.3916.42100.63
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.162, + 0.483, + 0.189 + ], + "angle": 0, + "content": "Table 5. The effect of the proposed coarse-to-fine global-number alignment scheme." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.193, + 0.493, + 0.294 + ], + "angle": 0, + "content": "
MethodTypeTraining→TestingMAERMSE
MCNN [55]SpecificSHB→SHA221.4357.8
CrowdClip [23]SpecificSHB→SHA217.0322.7
RCC [13]GenericFSC→SHA240.1366.9
ClipCount [18]GenericFSC→SHA192.6308.4
ClipCount+OursGenericFSC→SHA176.24279.20
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.091, + 0.294, + 0.483, + 0.336 + ], + "angle": 0, + "content": "Table 6. Cross-dataset evaluation on ShanghaiTech dataset. \"SHA\", \"SHB\" and \"FSC\" denote ShangTech Part A, ShanghaiTech Part B and FSC-147 datasets respectively." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.349, + 0.216, + 0.364 + ], + "angle": 0, + "content": "5.4. Discussions" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.483, + 0.492 + ], + "angle": 0, + "content": "The effect of coarse-to-fine global-number alignment scheme. A coarse-to-fine global-number alignment scheme is designed to select accurate number-conditioned text prompt to enhance the counting ability of VLMs in NGA. We conduct experiments to study the effect of this scheme. \"w/ coarse\" means we only use coarse number-conditioned text prompts. All experiments are performed on the FSC-147 dataset with ClipCount [18]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.765 + ], + "angle": 0, + "content": "As Tab. 5 shows, with the proposed scheme, it achieves 17.52 MAE on validation set and 16.42 MAE on test set which are better than only applying coarse number-conditioned text prompts. The results prove the effectiveness of our coarse-to-fine global-number alignment scheme. The generalizing ability across datasets. To evaluate the cross-dataset generalization ability, we compare our approach with ZSC methods on the ShanghaiTech [55] crowd counting dataset. On ShanghaiTech dataset, we compare with specific methods which are category-specific and generic methods that are ZSC methods. The results are summarized in Tab. 6. Our proposed generic object counting approach outperforms both category-specific counting methods and ZSC methods. It surpasses CrowdClip [23] by 40.76 MAE and 42.8 RMSE without training on crowd counting dataset. Moreover, compared to the second best ZSC method, it gains 16.36 MAE improvement and 29.2 RMSE improvement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Overall, above results demonstrate that our strategy can significantly improve the cross-dataset generalizing ability. The effect of \\(\\delta\\). \\(\\delta\\) is used to balance the effect of coarse and fine similarity losses. Experiments are conducted on FSC-147 dataset with ClipCount[18] to analyze it. As Fig. 5 shows, \\(\\delta = 0.5\\) achieves the best performance on both val and test sets. It is also indicated that a small value of \\(\\delta\\) (0.1) may be inadequate to find a accurate coarse range, while a large value (0.9) would lose fine-grained information." + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.089, + 0.905, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.238, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.417, + 0.905, + 0.444 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on FSC147 and ZSC-8K. Ground-truth numbers are shown at the top-right corner." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.445, + 0.651, + 0.459 + ], + "angle": 0, + "content": "5.5. Visualization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.907, + 0.589 + ], + "angle": 0, + "content": "To examine the effect of our approach qualitatively, we visualize some results predicted by ClipCount [18] and our approach in Fig. 6. As the Fig. 6 shows, the results of our method are closer to the ground-truth number and our method can detect the objects that are missed by ClipCount. These visualizations further verify that our approach can improve the counting ability and localization ability of VLMs. More qualitative results are shown in supplement." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.595, + 0.634, + 0.61 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present a universal strategy for text-guided zero-shot object counting. We introduce Text-guided Local Ranking to improve the localization ability of VLMs by a local-text rank loss, which effectively push the features of foreground and background patches apart. Additionally, we propose Number-evoked Global Attention, which leverages the global image-level feature to align with multiple number-conditioned text prompts, and then selects the prompt with highest similarity to calculate crossattention with the global image-level feature, thereby enhancing the counting ability. Furthermore, we propose a dataset named ZSC-8K for comprehensive evaluation. Extensive experiments conducted on the ZSC-8K dataset and three other public datasets demonstrate the effectiveness, scalability, and cross-dataset generalizability of our method. Although effective, the results on the long-tailed test set reveal substantial performance gaps when handling long-tailed classes. This persistent challenge across all ZSC methods highlights a critical direction for future research." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21104" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.241, + 0.108 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.116, + 0.483, + 0.176 + ], + "angle": 0, + "content": "This work was supported in part by the National Nature Science Foundation of China (NSFC) under Grant 62376209; and in part by the Fundamental Research Funds for the Central Universities under Grant xtr072024017." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.191, + 0.187, + 0.207 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.216, + 0.482, + 0.271 + ], + "angle": 0, + "content": "[1] Hüseyin Gökhan Akçay, Bekir Kabasakal, Duygugül Aksu, Nusret Demir, Melih Öz, and Ali Erdoğan. Automated bird counting with deep learning for regional bird distribution mapping. Animals, 10(7):1207, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. BMVC, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. NeurIPS, 2024. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[4] Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. Interactive object counting. In ECCV, pages 504-518. Springer, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[5] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. In ECCV, pages 483-498. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.446, + 0.482, + 0.499 + ], + "angle": 0, + "content": "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Ming-Hsuan Yang, SyYen Kuo, et al. Improving point-based crowd counting and localization based on auxiliary point guidance. ECCV, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.503, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[7] Ke Chen, Shaogang Gong, Tao Xiang, and Chen Change Loy. Cumulative attribute space for age and crowd density estimation. In CVPR, pages 2467-2474, 2013. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.547, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In CVPR, pages 16985-16995, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.589, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.646, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[10] SongLin Dong, ChengLi Tan, ZhenTao Zuo, YuHang He, YiHong Gong, TianGang Zhou, JunMin Liu, and JiangShe Zhang. Brain-inspired dual-pathway neural network architecture and its generalization analysis. Science China Technological Sciences, 67(8):2319-2330, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[12] Shenjian Gong, Shanshan Zhang, Jian Yang, Dengxin Dai, and Bernt Schiele. Class-agnostic object counting robust to intraclass diversity. In ECCV, pages 388–403. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Michael Hobley and Victor Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. arXiv preprint arXiv:2205.10203, 2022. 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] M. Hobley and V. Prisacariu. Abc easy as 123: A blind counter for exemplar-free multi-class class-agnostic counting. In ECCV, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.216, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In ICCV, pages 4145-4153, 2017. 2, 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[16] Yifeng Huang, Viresh Ranjan, and Minh Hoai. Interactive class-agnostic object counting. In ICCV, pages 22312–22322, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.519, + 0.178, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[18] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, pages 4535-4545, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.904, + 0.327 + ], + "angle": 0, + "content": "[19] Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In CVPR, pages 4705-4714, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.904, + 0.382 + ], + "angle": 0, + "content": "[20] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI, pages 2714-2722, 2024. 3, 6, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.413 + ], + "angle": 0, + "content": "[21] Victor S. Lempitsky and Andrew Zisserman. Learning to count objects in images. In NIPS, pages 1324-1332, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.415, + 0.904, + 0.466 + ], + "angle": 0, + "content": "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes. In CVPR, pages 1091-1100, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[23] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. In CVPR, pages 2893-2903, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.904, + 0.566 + ], + "angle": 0, + "content": "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, pages 19596-19605, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.904, + 0.607 + ], + "angle": 0, + "content": "[25] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. BMVC, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.904, + 0.649 + ], + "angle": 0, + "content": "[26] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In ICCV, pages 1676-1685, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.904, + 0.803 + ], + "angle": 0, + "content": "[29] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In ACCV, pages 669–684. Springer, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.805, + 0.904, + 0.845 + ], + "angle": 0, + "content": "[30] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In ICCV, pages 6141-6150, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[31] Thomas Moranduzzo and Farid Melgani. Automatic car counting method for unmanned aerial vehicle images. IEEE Transactions on Geoscience and remote sensing, 52(3): 1635-1647, 2013. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21105" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[32] T Nathan Mundhenk, Goran Konjevod, Wesam A Sakla, and Kofi Boakye. A large contextual dataset for classification, detection and counting of cars with deep learning. In ECCV, pages 785-800. Springer, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[33] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In ECCV, pages 348-365. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.191, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[34] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[35] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. CVPR, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[36] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems, 37:66260-66282, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[38] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[39] Deepak Babu Sam, Shiv Surya, and R. Venkatesh Babu. Switching convolutional neural network for crowd counting. In CVPR, pages 4031-4039, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[40] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In CVPR, pages 9529–9538, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[41] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In WACV, pages 323-331, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.482, + 0.644 + ], + "angle": 0, + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[43] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. JHU-CROWD++: large-scale crowd counting dataset and A benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[44] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In ICCV, pages 3365-3374, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[45] Yuejiao Su, Yi Wang, Lei Yao, and Lap-Pui Chau. Few-shot class-agnostic counting with occlusion augmentation and localization. In ISCAS, pages 1-5. IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[46] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. NWPU-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2141-2149, 2021. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[47] Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.906, + 0.134 + ], + "angle": 0, + "content": "Yang, and Siwei Lyu. Ua-detrac: A new benchmark and protocol for multi-object detection and tracking. Computer Vision and Image Understanding, 193:102907, 2020. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.204 + ], + "angle": 0, + "content": "[48] Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional regression networks. Computer methods in biomechanics and biomedical engineering: Imaging & Visualization, 6(3): 283-292, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.206, + 0.906, + 0.247 + ], + "angle": 0, + "content": "[49] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In CVPR, pages 15548-15557, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[50] Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo, Xiao Tan, Yezhen Wang, Shilei Wen, and Errui Ding. Perspective-guided convolution networks for crowd counting. In ICCV, pages 952-961, 2019. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[51] Shuo-Diao Yang, Hung-Ting Su, Winston H. Hsu, and Wen-Chin Chen. Class-agnostic few-shot object counting. In WACV, pages 870–878, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.906, + 0.401 + ], + "angle": 0, + "content": "[52] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. In WACV, pages 6315-6324, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.906, + 0.46 + ], + "angle": 0, + "content": "[53] Shanghang Zhang, Guanhang Wu, Joao P Costeira, and José MF Moura. Fcn-rlstm: Deep spatio-temporal neural networks for vehicle counting in city cameras. In ICCV, pages 3667-3676, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[54] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In ACM Multimedia, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.559 + ], + "angle": 0, + "content": "[55] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In CVPR, pages 589-597, 2016. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.56, + 0.906, + 0.601 + ], + "angle": 0, + "content": "[56] H.L Zhu, J.L Yuan, Z.W Yang, Y. Guo, Z. Wang, X. Zhong, and S.F He. Zero-shot object counting with good exemplars. In ECCV, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.906, + 0.671 + ], + "angle": 0, + "content": "[57] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 3, 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.671 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21106" + } + ] +] \ No newline at end of file diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_origin.pdf b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5179b98f6a497d55bb2259384c26bb4612e846fd --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/42dbced4-3afc-4fc3-a1f8-6ab3c4e3bb96_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:18ede53091ee2294ca7bdd882d6559ef1d0e3c157cda46a274b6397a7c8b4376 +size 589598 diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/full.md b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9e78e63dc16a5bb16a16e03b059123987d8667e3 --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/full.md @@ -0,0 +1,308 @@ +# Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention + +Shiwei Zhang $^{1,\dagger}$ , Qi Zhou $^{1,\dagger}$ , Wei Ke $^{1,2,*}$ + +$^{1}$ School of Software Engineering, Xi'an Jiaotong University, China + +$^{2}$ Pengcheng Laboratory, Shenzhen, China + +{zhangshiwei96,zhouqq}@stu.xjtu.edu.cn,wei.ke@mail.xjtu.edu.cn + +# Abstract + +Text-guided zero-shot object counting leverages vision-language models (VLMs) to count objects of an arbitrary class given by a text prompt. Existing approaches for this challenging task only utilize local patch-level features to fuse with text feature, ignoring the important influence of the global image-level feature. In this paper, we propose a universal strategy that can exploit both local patch-level features and global image-level feature simultaneously. Specifically, to improve the localization ability of VLMs, we propose Text-guided Local Ranking. Depending on the prior knowledge that foreground patches have higher similarity with the text prompt, a new local-text rank loss is designed to increase the differences between the similarity scores of foreground and background patches which push foreground and background patches apart. To enhance the counting ability of VLMs, Number-evoked Global Attention is introduced to first align global image-level feature with multiple number-conditioned text prompts. Then, the one with the highest similarity is selected to compute cross-attention with the global image-level feature. Through extensive experiments on widely used datasets and methods, the proposed approach has demonstrated superior advancements in performance, generalization, and scalability. Furthermore, to better evaluate text-guided zero-shot object counting methods, we propose a dataset named ZSC-8K, which is larger and more challenging, to establish a new benchmark. Codes and dataset are released at https://github.com/zaqai/LGCount. + +# 1. Introduction + +In recent years, with the development of deep learning techniques [10, 37, 42], category-specific counting (CSC) methods [30, 44] have gradually matured. However, existing CSC methods face notable limitations in practical applica + +![](images/cb563e8f39c3719aa40742e7e1d44073cbdd7c66949540d53cf6a26a6fa4a1e5.jpg) +Figure 1. Comparison of different pipelines for text-guided zero-shot object counting. (a) Only fusing single text prompt and local patch-level features. (b) Instead of overlooking the global image-level feature, our universal strategy leverages the global image-level feature to align with multiple number-conditioned text prompts and select the one with the highest similarity to calculate cross-attention with image-level feature. At last, the feature obtained by cross-attention is added to the local patch-level features. + +tions. These approaches require training dedicated models on extensive labeled datasets for each object category, so they fundamentally lack the capacity to handle unseen objects in a human-like manner. Therefore, counting methods that are not limited by fixed categories have gradually attracted the attention of the academic community, as they offer greater flexibility and broader application potential. + +Many exemplar-guided class-agnostic object counting (CAC) methods [38, 40, 52] have been presented in this context. These methods take several human-annotated bounding boxes of objects as exemplars to specify the objects-of-interest and train a unified network that can estimate the number of arbitrary objects-of-interest. However, previous CAC methods heavily rely on the assumption that accurate bounding boxes of objects can be easily obtained. In practice, this requirement is unfriendly to end-users lacking pro + +fessional knowledge and is even impossible to fulfill for a fully automatic system (e.g. a wildlife monitoring system). + +To provide a more flexible object counting approach, text-guided zero-shot object counting (ZSC) has been proposed by [49] and further developed by [8, 18] which leverage vision-language models (VLMs) [27, 37]. Under this setting, the bounding boxes of exemplars are replaced by the class name. During inference, users only need to specify the class name of the object that they want to count, and the model can automatically output the number of objects of this class. Even if the class is completely unseen during the training stage, the counter should still adapt to this unseen class without any additional data. Nevertheless, as Fig. 1(a) shows, existing ZSC methods only utilize local patch-level features from vision encoder to fuse with text feature and then use the fused features to make predictions. The global image-level feature, which is also derived from the vision encoder, is discarded, and its significance is completely overlooked by previous methods. + +In this paper, we propose a universal strategy which exploits both local patch-level features and global image-level feature to enhance text-guided zero-shot object counting as shown in Fig. 1(b). It consists of two parts: Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). In TLR, we reveal a critical limitation in prior method [18]: its utilization of local patch-level features inadvertently reduces the discriminative gap between foreground and background similarity scores, which fundamentally undermines VLMs' localization performance. To alleviate this limitation, we leverage the prior knowledge that the similarity scores of foreground patches must be greater than those of background features to design a novel local-text rank loss. This loss is designed to increase the differences between the foreground and background similarity scores which push the foreground and background patches apart to strengthen the localization ability of VLMs. In NGA, we exploit the global image-level feature and propose multiple number-conditioned text prompts. Unlike previous single text prompt that merely specify the class, our multiple number-conditioned text prompts not only indicate the class but also have predefined number ranges. And a similarity loss is proposed to align the global image-level feature with the prompt that has the closest number range. Then, to find a more accurate range, a coarse-to-fine global-number alignment scheme is proposed. In the end, the number-conditioned text prompt with the highest similarity is selected to calculate cross-attention with global image-level feature. Considering that the text prompt used in VLMs' pre-training stage usually lacks numeric information, by interacting number-conditioned text prompt with global image-level feature, our proposed NGA significantly improves the counting ability of VLMs. + +To verify the effectiveness of our proposed strategy, we + +conducted experiments on three prevailing datasets: FSC-147 [38], CARPK [15] and ShanghaiTech [55]. Moreover, to perform a more comprehensive evaluation of ZSC methods, we create a larger and more challenging dataset, named ZSC-8K, which increases the proportion of common classes and moves long-tailed categories in FSC-147 to the test set of ZSC-8K to evaluate the performance on long-tailed classes. The results demonstrate that our method significantly surpasses previous text-guided zero-shot object counting methods. Additionally, various backbones including CLIP [37] and GroundingDino [27] are also employed in our experiments to confirm the scalability of our strategy. + +In summary, the main contributions of the proposed approach can be summarized as follows: + +- We develop a universal strategy consists of Text-guided Local Ranking and Number-evoked Global Attention. This strategy is capable of exploiting both local patch-level features and global image-level feature for text-guided zero-shot object counting. +- Text-guided Local Ranking is proposed to improve the localization ability of VLMs by using a local-text rank loss to push the foreground and background patches apart. Number-evoked Global Attention is presented to enhance the counting ability of VLMs by computing crossattention between global image-level feature and number-conditioned text prompt with highest similarity. +- We introduce ZSC-8K, a scaled-up dataset consists of 8,361 images and 114 categories for comprehensive evaluation, which can benefit future researchers in this area. +- The proposed strategy achieves state-of-the-art performance on ZSC-8K dataset and three other popular datasets. Its effectiveness, scalability and cross-dataset generalizability are verified by experiments. + +# 2. Related Works + +# 2.1. Category-specific object counting + +Class-specific object counting (CSC) concentrates on counting objects of one specific category, for instance, humans [6, 19, 21, 24, 26, 30, 44, 55], penguins [5], birds [1], cells [48], or cars [15, 31, 32, 53]. Generally, existing CSC approaches can be classified into two groups: point-based methods and density-based methods. Point-based methods [6, 26, 44, 54] apply an object detector to the image and count the number of objects based on the detected points. Density-based methods [22, 39, 55] predict a density map for each input image, and the final result is obtained by summing up the pixel values. + +However, since existing CSC methods need to train a specific network with extensive labeled data for each class of object, they are deemed time-consuming and laborious. + +# 2.2. Exemplar-guided class-agnostic counting + +To ease the burden of humans and enable the counting model to count arbitrary categories, researchers propose exemplar-guided class-agnostic counting (CAC) [4, 12, 16, 25, 29, 33, 40, 51, 52] which leverages human annotated images or patches as exemplars of a category to train a unified network that can count arbitrary categories. + +[12] indicates that color, shape and scale diversity are crucial for CAC performance and proposes an EFA module to handle diversity from all aspects. [16] proposes an interactive class-agnostic object counting framework. FamNet [38] presents an adaptation strategy to adapt network to any novel visual category at test time. [33] introduces an uncertainty-aware object detector. BMNet [40] proposes a similarity-aware CAC framework. CFOCNet [51] uses reference images as exemplars to reduce the huge cost of data collection. [52] encourages the model to inspect the query image by focusing more on the regions akin to the support images. CountGD [3] leverages both visual exemplar and text prompt to boost the model. [45] extracts multiple occluded features of exemplars for comparison and utilizes the position of instances for more accurate counting results. + +CAC methods usually assume that high-quality bounding boxes of objects can be obtained easily. However, obtaining these exemplars is not as easy as assumed. In practice, the requirement of providing exemplars is unfriendly to users who lack professional knowledge and will be an obstacle when we want to build a fully automatic system. + +# 2.3. Text-guided zero-shot object counting + +Considering that high-quality exemplars are not always easily obtained and aiming to construct a user-friendly system, text-guided zero-shot object counting (ZSC) [2, 8, 18, 20, 41, 49] is proposed to provide a more flexible and user-friendly approach. + +[49] first introduces ZSC and trains a VAE model to generate the class prototype. The generated prototype is used to select class-relevant patches as exemplars to train a counting network. Later, with the application of vision-language models (VLMs) [27, 37], even automatically obtained exemplars are no longer needed. [2, 18, 20, 34] leverage CLIP [37] to solve ZSC task. [18] proposes an end-to-end pipeline to estimate density maps for a given class, [2] introduces a feature interaction module to enhance the feature and [20] incorporates SPT, LAT, and SaSC modules to tailor VLBase for object counting. ABC123 [14] proposed a blind counter but can only count images with fewer than 300 objects. Furthermore, [8] employs GroundingDino [27] as its backbone, and once again, the boundary of ZSC is advanced. VA-Count [56] then uses GroundingDino to generate good exemplars to boost the model. + +Nonetheless, previous ZSC approaches only leverage local patch-level features from the vision encoder to interact + +![](images/894ce1e8f5528f8387c77ffcc39724c28db43ce4ac2ef396832bbdfd4abdec9f.jpg) +Figure 2. Distribution of each category in FSC-147 [38] (left) and the proposed ZSC-8K (right). + +
#Images#ClassesMaxMinAvgTotal
Overall83611141766146.23386498
Train5476301766149.72272243
Val112729485135.1639623
Testnormal139739953137.1051829
Testlongtail175857953142.4574632
+ +Table 1. Statistics of each set in ZSC-8K, including the number of point annotations (max, min, avg, total). + +with text feature. The global image-level feature is discarded, and its significance is largely overlooked. It still lacks a systematic method to utilize both local and global features to enhance ZSC approaches. + +# 3. ZSC-8K Dataset + +FSC-147 [38] dataset is the most widely-used dataset for text-guided zero-shot object counting (ZSC). However, it demonstrates an unrealistic class distribution characterized by excessive long-tailed categories (polka dots, cereals, shallots, etc.) and insufficient samples for common classes. This distributional bias deviates markedly from real-world occurrence patterns. Therefore, to conduct a more comprehensive evaluation of ZSC methods, we propose ZSC-8K dataset, inspired by REC dataset [8]. As shown in Fig. 2, compared to FSC-147, ZSC-8K is larger and more challenging which means it increases the proportion of common classes (e.g. cars, people) and moves long-tailed classes from FSC-147 to test set to evaluate the performance of ZSC methods on long-tailed classes. + +Data collection. Images in the ZSC-8K dataset are collected from multiple existing datasets, including FSC147 [38], JHU-Crowd [43], NWPU-Crowd [46], VisDrone [57], DETRAC [47], CARPK [15], MALL [7] and Crowd Surveillance [50]. The proportions of each dataset are depicted in the supplementary material. These datasets belong to different domains, including crowd counting, car counting and animal counting, etc. + +Dataset statistics. ZSC-8K contains 8,361 images with a total of 386,498 point annotations. The number of point annotations in an image ranges from 1 to 1,766. A histogram of point annotations per image is provided in the + +Figure 3. Illustration of our training pipeline. In Text-guided Local Ranking, depending on the prior knowledge of similarity score ranking, local-text rank loss is proposed to push the foreground and background patches apart which further improves the localization ability of VLMs. In Number-evoked Global Attention, a coarse-to-fine global-number alignment scheme is proposed to align global image-level feature with multiple coarse number-conditioned text prompts firstly, and with multiple fine number-conditioned text prompts secondly. The fine number-conditioned text prompt with the highest similarity score is selected to calculate cross-attention with image-level feature. And the feature obtained by cross-attention is added to the local patch-level features. In inference stage, the test images and their corresponding number-conditioned text prompts are forwarded to decoder to make predictions. +![](images/005ba00e8bffe892c82abd4887d803ed6364bfb55f2fd1170f84c2684fadae84.jpg) +Ft Text feature F1 Local patch-level features Fg Global image-level feature Fct Feature with highest similarity Pct Prompt of Fct + +![](images/54f288501a5899167ac38658c8d651f31d459ebe7b5bae8648d2a09d27e20fa2.jpg) + +supplementary material. The dataset can be divided into 5 major categories and 114 subcategories. The proportions of each category are shown in Fig. 2. As it depicts, people are the category with the highest proportion and cars are the second which are significantly higher than that in FSC-147 dataset [38]. We split ZSC-8K dataset into train, validation and test sets. Specific statistics for each set are detailed in Tab. 1. It is worth noting that two versions of test sets are proposed: $\text{test}_{\text{normal}}$ and $\text{test}_{\text{longtail}}$ . In $\text{test}_{\text{normal}}$ , images of normal classes (1397 images & 39 classes) are included, while $\text{test}_{\text{longtail}}$ is obtained by merging the long-tailed classes (361 images & 20 classes) from FSC-147 into $\text{test}_{\text{normal}}$ . To meet the zero-shot setting, the categories in train set do not overlap with validation and test sets. + +# 4. Methodology + +# 4.1. Overview + +The goal of ZSC is to count anything using only text guidance. Pretrained vision-language models (VLMs) are commonly used for this purpose. Denote $I \in R^{H \times W \times 3}$ as an input image containing an arbitrary object class, and a text prompt $t$ is given to specify the object class of this image. In previous methods, $I$ is initially fed into the image encoder of pretrained VLMs to obtain local patch-level features and $t$ is then encoded by the text encoder to obtain text feature. Subsequently, a fuse module is designed to combine the two modalities, and the fused feature is used to predict the num + +ber of the specified category. In this paper, we argue that previous methods only leverage local patch-level features to fuse with the text feature, thereby fully overlooking the importance of the global image-level feature. Therefore, to exploit local patch-level features and global image-level simultaneously, we propose a strategy consists of Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). The overall framework is shown in Fig. 3. + +Specifically, the local patch-level features and text feature are used to compute a novel local-text rank loss that can increase the differences between foreground and background similarity scores in TLR. The localization ability of VLMs is improved by TLR. Then, we leverage the global image-level feature in NGA. Multiple number-conditioned text prompts are designed and a coarse-to-fine global-number alignment scheme is proposed to select the number-conditioned text prompt with highest similarity. The selected prompt is considered to have the closest number range and its feature is used to calculate cross-attention with the global image-level feature to enhance the counting ability. Finally, the feature obtained by cross-attention is aggregated with local patch-level features through element-wise summation. This combined representations is then fed into decoder to make predictions. TLR and NGA are detailed in Sec. 4.2 and Sec. 4.3, respectively. + +# 4.2. Text-guided Local Ranking + +VLMs have shown impressive performance in estimating text-image similarity by aligning text and image representations. However, as demonstrated by ClipCount [18], pretrained VLMs intrinsically lack localization ability, making it non-trivial to transfer this ability to object counting. Although ClipCount [18] utilized the local patch-level features $F_{l}$ and the text feature $F_{t}$ to construct a contrastive loss to enhance localization ability by pushing foreground and background patches apart, this loss cannot effectively increase the discrepancy between foreground and background similarity scores, which in turn adversely affects localization ability. As shown in Fig. 4(a), the differences between foreground and background similarity scores in ClipCount [18] are small, with most differences are close to zero. To alleviate this issue, we propose Text-guided Local Ranking (TLR) in which a novel local-text rank loss is designed to improve the localization ability by increasing the differences between foreground and background similarity scores. As Fig. 4(b) shows, with TLR, the differences become much larger than those of ClipCount [18]. + +As shown in Fig. 3, $F_{l}$ and $F_{t}$ are also used in the proposed TLR, a linear projection maps them to the same channels. To determine which patch contains the object, max-pooling is used to transfer the ground truth density map to a patch-level binary mask. Patches whose value is equal to 1 are foreground patches which have the object and the value of background patches is 0. We denote the foreground patches as $\mathcal{P}$ and the background patches as $\mathcal{B}$ , the relation of $\mathcal{P}$ and $\mathcal{B}$ can be summarized as: + +$$ +s \left(F _ {l} ^ {p}, F _ {t}\right) > s \left(F _ {l} ^ {q}, F _ {t}\right), \forall p \in \mathcal {P}, \forall q \in \mathcal {B}, \tag {1} +$$ + +where $s(,)$ denotes the cosine similarity. And Formula 1 indicates the prior knowledge that the similarity score between any foreground patch feature and the text feature is greater than that of any background patch feature. Depending on the prior knowledge, we construct the local-text rank loss. The similarity score sets are derived as: + +$$ +S (\mathcal {P}) = \{\dots , s \left(F _ {l} ^ {p}, F _ {t}\right), \dots \} / \tau_ {1}, | \mathcal {P} | = l _ {1}, \tag {2} +$$ + +$$ +S (\mathcal {B}) = \{\dots , s \left(F _ {l} ^ {q}, F _ {t}\right), \dots \} / \tau_ {1}, | \mathcal {B} | = l _ {2}, \tag {3} +$$ + +where $S(\mathcal{P})$ and $S(\mathcal{B})$ are similarity score sets of foreground and background patches respectively. $l_{1}$ and $l_{2}$ are the lengths of the two sets. $\tau_{1}$ is a temperature parameter. Next, we denote $d_{pq} = (s(F_l^p,F_t) - s(F_l^q,F_t)) / \tau_1$ , and the + +![](images/bd21b25f9c9718d218ab757a4a4f6f2285321c40a52ec15b800eeaacd0cdd3fb.jpg) +Figure 4. The differences between foreground and background similarity scores. (a) shows the differences between foreground and background scores in ClipCount [18] are small, most of the differences are close to 0. (b) shows the differences with our TLR become larger. Hence, the features of foreground and background patches are pushed apart. + +differences between the two sets can be computed by: + +$$ +\begin{array}{l} M _ {d i f} = S (\mathcal {P}) \stackrel {\text {b r o a d c a s t - s u b t r a c t}} {\longrightarrow} S (\mathcal {B}) \\ = \left[ \begin{array}{c c c c} d _ {1 1} & d _ {1 2} & \dots & d _ {1 l _ {2}} \\ d _ {2 1} & d _ {2 2} & \dots & d _ {2 l _ {2}} \\ \vdots & \vdots & \ddots & \vdots \\ d _ {l _ {1} 1} & d _ {l _ {1} 2} & \dots & d _ {l _ {1} l _ {2}} \end{array} \right] \quad , \tag {4} \\ \end{array} +$$ + +where $M_{dif}$ is the result of subtracting $S(\mathcal{B})$ from $S(\mathcal{P})$ using the broadcasting mechanism. Finally, with the prior knowledge, the local-text rank loss can be formulated as: + +$$ +L _ {\text {r a n k}} = \frac {\sum \operatorname {c l a m p} \left(\lambda - M _ {\text {d i f}} , \min = 0\right)}{l _ {1} l _ {2}}, \tag {5} +$$ + +where $\lambda$ is a hyperparameter that controls the rank loss. + +With the proposed local-text rank loss, the differences are increased by ensuring the similarity scores of foreground patches are bigger than background patches. Thus the features of foreground and background patches are pushed apart and the localization ability is improved. + +# 4.3. Number-Evoked Global Attention + +Because most of the text prompts used in pre-training stage of VLMs lack descriptions of the number of objects, the performance of VLMs in object counting is largely limited. To mitigate this limitation, we propose Number-Evoked Global Attention (NGA) to enhance the counting ability of VLMs by calculating the cross-attention between the global image-level feature $F_{g}$ and the feature of the number-conditioned text prompt with highest similarity $F_{ct}^{h}$ . + +First, $F_{g}$ that is ignored by pervious ZSC methods is leveraged to select $P_{ct}^{h}$ , the number-conditioned text prompt with highest similarity, as shown in Fig. 3. Specifically, we first propose a global-number alignment scheme to identify a coarse number range for an image. To achieve this, we replace the single text prompt by multiple number-conditioned text prompts. Different from previous single text prompt which only specify the object class, every prompt of our proposed multiple number-conditioned text prompts has an object class and a predefined number range. + +
MethodBackboneVenueVal setTest set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ZSC [49]ResNet+VAECVPR'2326.9388.6322.09115.17
TFOC [41]SAMWACV'24--20.10132.83
CountTX* [2]CLIPBMCV'2317.7665.2116.70105.21
VA-Count [56]CLIPECCV'2417.8773.2217.88129.31
GeCo [36]CLIPNeurIPS'2414.8164.9513.30108.72
DAVEprm [35]CLIPCVPR'2415.4852.5714.90103.42
GroundingREC* [8]GDinoCVPR'2412.3761.8611.27108.21
VLCounter [20]CLIPAAAI'2418.0665.1317.05106.16
VLCounter+OursCLIP-17.88(-0.18)63.88(-1.25)16.56(-0.49)107.47(+1.31)
ClipCount [18]CLIPMM'2318.9366.1017.45104.88
ClipCount+OursCLIP-17.52(-1.41)58.39(-7.71)16.42(-1.03)100.63(-4.25)
CountGD [3]GDinoNeurIPS'2412.1447.5114.76120.42
CountGD+OursGDino-11.62(-0.52)46.37(-1.14)12.84(-1.92)115.61(-4.81)
GDino [27](w/ finetune)*GDinoECCV'2412.0762.1311.76104.98
GDino(w/ finetune)+OursGDino-11.09(-0.98)60.48(-1.65)10.74(-1.02)106.5(+1.52)
+ +Table 2. Comparison with state-of-the-art methods on FSC-147 dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes. + +We initially design 5 distinct number ranges and leverage these ranges to generate 5 different number-conditioned text prompts such as "There are between {number} and {number} {class}" . Then, a global-number similarity loss is proposed to maximize the similarity between the global image-level feature and the positive number-conditioned text prompt which has the closest number range. The global-number similarity loss can be formulated as: + +$$ +L _ {s i m} = - \log \frac {\exp \left(s \left(F _ {g} , F _ {c t} ^ {i}\right) / \tau_ {2}\right)}{\exp \left(s \left(F _ {g} , F _ {c t} ^ {i}\right) / \tau_ {2}\right) + \sum_ {j \in \mathcal {N}} \exp \left(s \left(F _ {g} , F _ {c t} ^ {j}\right) / \tau_ {2}\right)}, \tag {6} +$$ + +where $s(\cdot)$ denotes the cosine similarity, $\tau_{2}$ is a temperature parameter and $F_{ct}^{i}$ is the positive number-conditioned text prompt embedding. $\mathcal{N}$ denotes the set of negative number-conditioned text prompt embeddings. For CLIP-based methods, we take the global `cls_token` as our $F_{g}$ , while for GroundingDino[27], we obtain the $F_{g}$ by using average pooling on the last layer features of Swin-T. + +Through above global-number alignment scheme, model can learn to find a coarse number range for an image. But this coarse range is still too wide to serve as an accurate prompt. Hence, as shown in Fig. 3, we further propose a coarse-to-fine global-number alignment scheme to generate more accurate number-conditioned text prompt. The number-conditioned text prompts are designed to be hierarchical, which means that the coarse range with highest similarity is further divided into 5 ranges, and the global-number similarity loss is adopted again to maximize the similarity of the global image-level feature and the positive fine number-conditioned text prompt. + +Overall, the loss of our NGA can be formulated as: + +$$ +L _ {N G A} = \delta L _ {s i m} ^ {\text {c o a r s e}} + (1 - \delta) L _ {s i m} ^ {\text {f i n e}}, \tag {7} +$$ + +where $L_{sim}^{coarse}$ is the global-number similarity loss of coarse + +number-conditioned text prompts and $L_{sim}^{fine}$ is for fine number-conditioned text prompts. $\delta$ is a weight term to balance the effect of coarse and fine similarity loss. + +With our proposed scheme, $P_{ct}^{h}$ is selected and considered to have the closest number range. The feature of it $F_{ct}^{h}$ is used to compute cross-attention with $F_{g}$ in NGA, thereby the counting ability of VLMs is enhanced. + +# 5. Experiments + +Firstly, the experimental setting, including implementation details and evaluation metrics, is given in Sec. 5.1. Then, we apply our strategy to three different methods and compare their performances with state-of-the-art text-guided zero-shot object counting approaches in Sec. 5.2. Through ablation studies in Sec. 5.3, we quantitatively evaluate the effectiveness of each component in our strategy. Next, we conduct experiments to discuss the impact of the coarse-to-fine global-number alignment scheme, the effect of our strategy on cross-dataset generalization and the influence of $\delta$ in Sec. 5.4. Finally, visualizations in Sec. 5.5 qualitatively verify the efficacy of our strategy. + +# 5.1. Experimental Setting + +Implementation details. To evaluate our strategy comprehensively, we apply our strategy to two CLIP-based methods [18, 20] and one GroundingDino(GDino)-based method [27]. For CLIP-based methods, the vision encoder is VIT-B/16 [11]. VPT [17] is used with 20 visual prompts in each layer. The maxpooling in TLR adopts a kernel size of $16 \times 16$ , stride=16 and padding=0. For the density decoder, we use $3 \times 3$ convolution kernel with stride=1 and padding=1 to keep resolution. The batch size is 32 and the overall training epochs are 440. For GDino-based method, + +
MethodBackboneVal setTestnormal setTestlongtail set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ClipCount* [18]CLIP17.5330.4218.6838.6122.6553.79
ClipCount+OursCLIP16.65(-0.88)29.82(-0.6)17.24(-1.44)36.92(-1.69)21.90(-0.75)53.66(-0.13)
GroundingREC* [8]GDino14.3425.0113.7828.9616.6541.80
GDino* [27](w/ finetune)GDino13.9221.6414.2329.6816.5738.85
GDino(w/ finetune)+OursGDino11.17(-2.75)17.30(-4.34)12.39(-1.84)27.75(-1.93)14.68((-1.89)38.10((-0.75)
+ +Table 3. Comparison with state-of-the-art methods on our ZSC-8K dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes. + +
MethodVal setTest set
MAERMSEMAERMSE
None18.9366.1017.45104.88
TLR18.4260.7516.71100.08
NGA18.2962.9116.81103.21
TLR+NGA17.5258.3916.42100.63
+ +Table 4. Ablation study on proposed NGA and TLR. + +the text encoder is BERT [9] and visual encoder is SwinT [28]. The overall training epochs are 30. And the other hyperparameters are same with [8]. We use AdamW optimizer to optimize our model, for CLIP-based methods, the learning rate is 1e-4 and for GDino-based method, the learning rate is 1e-5. For all methods, $\delta$ is set as 0.5. $\lambda$ is set as 10. Following [18], $\tau_{1}$ and $\tau_{2}$ are set as 0.07. + +Evaluation metrics. Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are widely used evaluation metrics in text-guided zero-shot counting tasks which is also applied in this paper. MAE = $\frac{1}{N}\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|$ and RMSE = $\sqrt{\frac{1}{N}\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|^2}$ , where $N$ is the total number of testing images, $C_m^{pred}$ and $C_m^{gt}$ are the prediction and the ground truth count of the $m$ -th image. + +# 5.2. Compare with State-of-the-art Methods + +In this section, we apply the proposed strategy to three existing approaches and compare the performances with state-of-the-art text-guided zero-shot object counting methods on two datasets: FSC-147 [38] and ZSC-8K, to evaluate the effectiveness of our strategy. + +The results on FSC-147 are shown in Tab. 2. As it depicts, our approach outperforms all previous state-of-the-art methods, regardless of the backbone. Compared with the best performances of previous methods, our approach reduces the MAE by $8.1\%$ and $4.7\%$ on the validation set and test set. And our method decreases the RMSE by $2.5\%$ and $2.7\%$ on the validation set and test set respectively. Our method also shows its scalability by comparisons with the baselines, the performances of VLCounter+Ours, ClipCount+Ours, CountGD+Ours and GDino(w/ finetune)+Ours are all better than their baselines. + +Then, we compare the performance of our method with previous ZSC methods on our proposed dataset, ZSC-8K. + +The results in Tab. 3 demonstrate that our method obtains the best performance on ZSC-8K dataset. Specifically, compared to the previous best results, the MAE is reduced by $19.8\%$ and $10.1\%$ on validation set and testnormal set. And every baseline is improved by our strategy. Furthermore, we evaluate ZSC methods on the testlongtail set in the last two columns of Tab. 3. Although our method still achieves optimal performance, the results indicate a significant performance gap between normal and long-tailed classes, highlighting the limitation of ZSC methods. + +The remarkable results on the FSC-147 and ZSC-8K datasets illustrate the effectiveness of our strategy. In particular, whether CLIP or GDino is used as the backbone, our method significantly outperforms the baselines, which strongly validates the scalability of our strategy. + +# 5.3. Ablation Study + +We conduct ablation experiments to verify the effectiveness of the proposed Text-guided Local Ranking(TLR) and Number-Evoked Global Attention(NGA) in Tab. 4. + +All experiments are conducted with ClipCount [18] on FSC-147 [38]. We start with a baseline where no modules are applied. It achieves 18.93 MAE and 66.10 RMSE on validation set, 17.45 MAE and 104.88 RMSE on test set. Then, the single use of TLR and NGA is analyzed. By leveraging the proposed local-text rank loss, the foreground and background patches are pushed apart and the localization ability of VLMs is improved by TLR. The effectiveness is verified by the results shown in Tab. 4. Compared to baseline, TLR improves the MAE by $2.7\%$ on validation set and $4.2\%$ on test set. Next, the proposed NGA is adopted to study its effectiveness. With NGA which calculates crossattention between the number-conditioned text prompt and global image-level feature, its performance significantly improved, the error reduces by $3.7\%$ and $1.6\%$ for MAE and RMSE on test set, respectively. Hence, the results verify that the proposed NGA can enhance the counting ability of VLMs. Finally, the combination of TLR and NGA enhances the overall performance even further, improving the MAE by $7.4\%$ on validation set and $5.9\%$ on test set, suggesting the proposed TLR and NGA are complementary. + +
MethodVal setTest set
MAERMSEMAERMSE
w/ coarse18.1761.7317.34100.79
w/ coarse-to-fine17.5258.3916.42100.63
+ +Table 5. The effect of the proposed coarse-to-fine global-number alignment scheme. + +
MethodTypeTraining→TestingMAERMSE
MCNN [55]SpecificSHB→SHA221.4357.8
CrowdClip [23]SpecificSHB→SHA217.0322.7
RCC [13]GenericFSC→SHA240.1366.9
ClipCount [18]GenericFSC→SHA192.6308.4
ClipCount+OursGenericFSC→SHA176.24279.20
+ +Table 6. Cross-dataset evaluation on ShanghaiTech dataset. "SHA", "SHB" and "FSC" denote ShangTech Part A, ShanghaiTech Part B and FSC-147 datasets respectively. + +# 5.4. Discussions + +The effect of coarse-to-fine global-number alignment scheme. A coarse-to-fine global-number alignment scheme is designed to select accurate number-conditioned text prompt to enhance the counting ability of VLMs in NGA. We conduct experiments to study the effect of this scheme. "w/ coarse" means we only use coarse number-conditioned text prompts. All experiments are performed on the FSC-147 dataset with ClipCount [18]. + +As Tab. 5 shows, with the proposed scheme, it achieves 17.52 MAE on validation set and 16.42 MAE on test set which are better than only applying coarse number-conditioned text prompts. The results prove the effectiveness of our coarse-to-fine global-number alignment scheme. The generalizing ability across datasets. To evaluate the cross-dataset generalization ability, we compare our approach with ZSC methods on the ShanghaiTech [55] crowd counting dataset. On ShanghaiTech dataset, we compare with specific methods which are category-specific and generic methods that are ZSC methods. The results are summarized in Tab. 6. Our proposed generic object counting approach outperforms both category-specific counting methods and ZSC methods. It surpasses CrowdClip [23] by 40.76 MAE and 42.8 RMSE without training on crowd counting dataset. Moreover, compared to the second best ZSC method, it gains 16.36 MAE improvement and 29.2 RMSE improvement. + +Overall, above results demonstrate that our strategy can significantly improve the cross-dataset generalizing ability. The effect of $\delta$ . $\delta$ is used to balance the effect of coarse and fine similarity losses. Experiments are conducted on FSC-147 dataset with ClipCount[18] to analyze it. As Fig. 5 shows, $\delta = 0.5$ achieves the best performance on both val and test sets. It is also indicated that a small value of $\delta$ (0.1) may be inadequate to find a accurate coarse range, while a large value (0.9) would lose fine-grained information. + +![](images/ab6934e464cf5976a0a0ba2feeadb048f93a60aa63a700748f60248a362b7223.jpg) + +![](images/1e738f1126779289492e43953de7de0b2bb63be4a318543018b51710012f3bc7.jpg) +Figure 6. Qualitative results on FSC147 and ZSC-8K. Ground-truth numbers are shown at the top-right corner. + +# 5.5. Visualization + +To examine the effect of our approach qualitatively, we visualize some results predicted by ClipCount [18] and our approach in Fig. 6. As the Fig. 6 shows, the results of our method are closer to the ground-truth number and our method can detect the objects that are missed by ClipCount. These visualizations further verify that our approach can improve the counting ability and localization ability of VLMs. More qualitative results are shown in supplement. + +# 6. Conclusion + +In this paper, we present a universal strategy for text-guided zero-shot object counting. We introduce Text-guided Local Ranking to improve the localization ability of VLMs by a local-text rank loss, which effectively push the features of foreground and background patches apart. Additionally, we propose Number-evoked Global Attention, which leverages the global image-level feature to align with multiple number-conditioned text prompts, and then selects the prompt with highest similarity to calculate crossattention with the global image-level feature, thereby enhancing the counting ability. Furthermore, we propose a dataset named ZSC-8K for comprehensive evaluation. Extensive experiments conducted on the ZSC-8K dataset and three other public datasets demonstrate the effectiveness, scalability, and cross-dataset generalizability of our method. Although effective, the results on the long-tailed test set reveal substantial performance gaps when handling long-tailed classes. This persistent challenge across all ZSC methods highlights a critical direction for future research. + +# Acknowledgment + +This work was supported in part by the National Nature Science Foundation of China (NSFC) under Grant 62376209; and in part by the Fundamental Research Funds for the Central Universities under Grant xtr072024017. + +# References + +[1] Hüseyin Gökhan Akçay, Bekir Kabasakal, Duygugül Aksu, Nusret Demir, Melih Öz, and Ali Erdoğan. Automated bird counting with deep learning for regional bird distribution mapping. Animals, 10(7):1207, 2020. 2 +[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. BMVC, 2023. 3, 6 +[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. NeurIPS, 2024. 3, 6, 2 +[4] Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. Interactive object counting. In ECCV, pages 504-518. Springer, 2014. 3 +[5] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. In ECCV, pages 483-498. Springer, 2016. 2 +[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Ming-Hsuan Yang, SyYen Kuo, et al. Improving point-based crowd counting and localization based on auxiliary point guidance. ECCV, 2024. 2 +[7] Ke Chen, Shaogang Gong, Tao Xiang, and Chen Change Loy. Cumulative attribute space for age and crowd density estimation. In CVPR, pages 2467-2474, 2013. 3, 1 +[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In CVPR, pages 16985-16995, 2024. 2, 3, 6, 7 +[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 7 +[10] SongLin Dong, ChengLi Tan, ZhenTao Zuo, YuHang He, YiHong Gong, TianGang Zhou, JunMin Liu, and JiangShe Zhang. Brain-inspired dual-pathway neural network architecture and its generalization analysis. Science China Technological Sciences, 67(8):2319-2330, 2024. 1 +[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6 +[12] Shenjian Gong, Shanshan Zhang, Jian Yang, Dengxin Dai, and Bernt Schiele. Class-agnostic object counting robust to intraclass diversity. In ECCV, pages 388–403. Springer, 2022. 3 +[13] Michael Hobley and Victor Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. arXiv preprint arXiv:2205.10203, 2022. 8, 2 +[14] M. Hobley and V. Prisacariu. Abc easy as 123: A blind counter for exemplar-free multi-class class-agnostic counting. In ECCV, 2024. 3 + +[15] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In ICCV, pages 4145-4153, 2017. 2, 3, 1 +[16] Yifeng Huang, Viresh Ranjan, and Minh Hoai. Interactive class-agnostic object counting. In ICCV, pages 22312–22322, 2023. 3 +[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 6 +[18] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, pages 4535-4545, 2023. 2, 3, 5, 6, 7, 8 +[19] Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In CVPR, pages 4705-4714, 2020. 2 +[20] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI, pages 2714-2722, 2024. 3, 6, 2 +[21] Victor S. Lempitsky and Andrew Zisserman. Learning to count objects in images. In NIPS, pages 1324-1332, 2010. 2 +[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes. In CVPR, pages 1091-1100, 2018. 2 +[23] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. In CVPR, pages 2893-2903, 2023. 8 +[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, pages 19596-19605, 2022. 2 +[25] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. BMVC, 2022. 3 +[26] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In ICCV, pages 1676-1685, 2023. 2 +[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 6, 7 +[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 7 +[29] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In ACCV, pages 669–684. Springer, 2019. 3 +[30] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In ICCV, pages 6141-6150, 2019. 1, 2 +[31] Thomas Moranduzzo and Farid Melgani. Automatic car counting method for unmanned aerial vehicle images. IEEE Transactions on Geoscience and remote sensing, 52(3): 1635-1647, 2013. 2 + +[32] T Nathan Mundhenk, Goran Konjevod, Wesam A Sakla, and Kofi Boakye. A large contextual dataset for classification, detection and counting of cars with deep learning. In ECCV, pages 785-800. Springer, 2016. 2 +[33] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In ECCV, pages 348-365. Springer, 2022. 3 +[34] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 3 +[35] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. CVPR, 2024. 6 +[36] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems, 37:66260-66282, 2024. 6 +[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3 +[38] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 1, 2, 3, 4, 7 +[39] Deepak Babu Sam, Shiv Surya, and R. Venkatesh Babu. Switching convolutional neural network for crowd counting. In CVPR, pages 4031-4039, 2017. 2 +[40] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In CVPR, pages 9529–9538, 2022. 1, 3 +[41] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In WACV, pages 323-331, 2024. 3, 6 +[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1 +[43] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. JHU-CROWD++: large-scale crowd counting dataset and A benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 3, 1 +[44] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In ICCV, pages 3365-3374, 2021. 1, 2 +[45] Yuejiao Su, Yi Wang, Lei Yao, and Lap-Pui Chau. Few-shot class-agnostic counting with occlusion augmentation and localization. In ISCAS, pages 1-5. IEEE, 2024. 3 +[46] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. NWPU-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2141-2149, 2021. 3, 1 +[47] Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan + +Yang, and Siwei Lyu. Ua-detrac: A new benchmark and protocol for multi-object detection and tracking. Computer Vision and Image Understanding, 193:102907, 2020. 3, 1 +[48] Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional regression networks. Computer methods in biomechanics and biomedical engineering: Imaging & Visualization, 6(3): 283-292, 2018. 2 +[49] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In CVPR, pages 15548-15557, 2023. 2, 3, 6 +[50] Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo, Xiao Tan, Yezhen Wang, Shilei Wen, and Errui Ding. Perspective-guided convolution networks for crowd counting. In ICCV, pages 952-961, 2019. 3, 1 +[51] Shuo-Diao Yang, Hung-Ting Su, Winston H. Hsu, and Wen-Chin Chen. Class-agnostic few-shot object counting. In WACV, pages 870–878, 2021. 3 +[52] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. In WACV, pages 6315-6324, 2023. 1, 3 +[53] Shanghang Zhang, Guanhang Wu, Joao P Costeira, and José MF Moura. Fcn-rlstm: Deep spatio-temporal neural networks for vehicle counting in city cameras. In ICCV, pages 3667-3676, 2017. 2 +[54] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In ACM Multimedia, 2024. 2 +[55] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In CVPR, pages 589-597, 2016. 2, 8 +[56] H.L Zhu, J.L Yuan, Z.W Yang, Y. Guo, Z. Wang, X. Zhong, and S.F He. Zero-shot object counting with good exemplars. In ECCV, 2024. 3, 6 +[57] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 3, 1 \ No newline at end of file diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/images.zip b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a514a69fad0cabfcfccaf631f9891c565b882c3a --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6d47a0063caa4534619554b72dbdfa79be0704d34081ca2194a9c98f5b7ab17 +size 526194 diff --git a/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/layout.json b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1eab80f5c12d4d3a62d16bfb8b9d5dba2b5c65de --- /dev/null +++ b/2025/Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention/layout.json @@ -0,0 +1,8221 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "text", + "content": "Enhancing Zero-shot Object Counting via Text-guided Local Ranking and Number-evoked Global Attention" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "spans": [ + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "text", + "content": "Shiwei Zhang" + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "text", + "content": ", Qi Zhou" + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "text", + "content": ", Wei Ke" + }, + { + "bbox": [ + 206, + 160, + 403, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 143, + 175, + 467, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 175, + 467, + 190 + ], + "spans": [ + { + "bbox": [ + 143, + 175, + 467, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 143, + 175, + 467, + 190 + ], + "type": "text", + "content": "School of Software Engineering, Xi'an Jiaotong University, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 204, + 189, + 406, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 189, + 406, + 203 + ], + "spans": [ + { + "bbox": [ + 204, + 189, + 406, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 204, + 189, + 406, + 203 + ], + "type": "text", + "content": "Pengcheng Laboratory, Shenzhen, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "type": "text", + "content": "{zhangshiwei96,zhouqq}@stu.xjtu.edu.cn,wei.ke@mail.xjtu.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 605 + ], + "type": "text", + "content": "Text-guided zero-shot object counting leverages vision-language models (VLMs) to count objects of an arbitrary class given by a text prompt. Existing approaches for this challenging task only utilize local patch-level features to fuse with text feature, ignoring the important influence of the global image-level feature. In this paper, we propose a universal strategy that can exploit both local patch-level features and global image-level feature simultaneously. Specifically, to improve the localization ability of VLMs, we propose Text-guided Local Ranking. Depending on the prior knowledge that foreground patches have higher similarity with the text prompt, a new local-text rank loss is designed to increase the differences between the similarity scores of foreground and background patches which push foreground and background patches apart. To enhance the counting ability of VLMs, Number-evoked Global Attention is introduced to first align global image-level feature with multiple number-conditioned text prompts. Then, the one with the highest similarity is selected to compute cross-attention with the global image-level feature. Through extensive experiments on widely used datasets and methods, the proposed approach has demonstrated superior advancements in performance, generalization, and scalability. Furthermore, to better evaluate text-guided zero-shot object counting methods, we propose a dataset named ZSC-8K, which is larger and more challenging, to establish a new benchmark. Codes and dataset are released at https://github.com/zaqai/LGCount." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 628, + 135, + 639 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "In recent years, with the development of deep learning techniques [10, 37, 42], category-specific counting (CSC) methods [30, 44] have gradually matured. However, existing CSC methods face notable limitations in practical applica" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 317, + 245, + 553, + 406 + ], + "blocks": [ + { + "bbox": [ + 317, + 245, + 553, + 406 + ], + "lines": [ + { + "bbox": [ + 317, + 245, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 317, + 245, + 553, + 406 + ], + "type": "image", + "image_path": "cb563e8f39c3719aa40742e7e1d44073cbdd7c66949540d53cf6a26a6fa4a1e5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 411, + 555, + 499 + ], + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 499 + ], + "type": "text", + "content": "Figure 1. Comparison of different pipelines for text-guided zero-shot object counting. (a) Only fusing single text prompt and local patch-level features. (b) Instead of overlooking the global image-level feature, our universal strategy leverages the global image-level feature to align with multiple number-conditioned text prompts and select the one with the highest similarity to calculate cross-attention with image-level feature. At last, the feature obtained by cross-attention is added to the local patch-level features." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 517, + 555, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 517, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 312, + 517, + 555, + 602 + ], + "type": "text", + "content": "tions. These approaches require training dedicated models on extensive labeled datasets for each object category, so they fundamentally lack the capacity to handle unseen objects in a human-like manner. Therefore, counting methods that are not limited by fixed categories have gradually attracted the attention of the academic community, as they offer greater flexibility and broader application potential." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Many exemplar-guided class-agnostic object counting (CAC) methods [38, 40, 52] have been presented in this context. These methods take several human-annotated bounding boxes of objects as exemplars to specify the objects-of-interest and train a unified network that can estimate the number of arbitrary objects-of-interest. However, previous CAC methods heavily rely on the assumption that accurate bounding boxes of objects can be easily obtained. In practice, this requirement is unfriendly to end-users lacking pro" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 220, + 713 + ], + "type": "text", + "content": "† Equal contribution. * Corresponding author." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21097" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 96 + ], + "type": "text", + "content": "fessional knowledge and is even impossible to fulfill for a fully automatic system (e.g. a wildlife monitoring system)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 99, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 99, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 54, + 99, + 295, + 301 + ], + "type": "text", + "content": "To provide a more flexible object counting approach, text-guided zero-shot object counting (ZSC) has been proposed by [49] and further developed by [8, 18] which leverage vision-language models (VLMs) [27, 37]. Under this setting, the bounding boxes of exemplars are replaced by the class name. During inference, users only need to specify the class name of the object that they want to count, and the model can automatically output the number of objects of this class. Even if the class is completely unseen during the training stage, the counter should still adapt to this unseen class without any additional data. Nevertheless, as Fig. 1(a) shows, existing ZSC methods only utilize local patch-level features from vision encoder to fuse with text feature and then use the fused features to make predictions. The global image-level feature, which is also derived from the vision encoder, is discarded, and its significance is completely overlooked by previous methods." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 304, + 295, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 295, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 295, + 699 + ], + "type": "text", + "content": "In this paper, we propose a universal strategy which exploits both local patch-level features and global image-level feature to enhance text-guided zero-shot object counting as shown in Fig. 1(b). It consists of two parts: Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). In TLR, we reveal a critical limitation in prior method [18]: its utilization of local patch-level features inadvertently reduces the discriminative gap between foreground and background similarity scores, which fundamentally undermines VLMs' localization performance. To alleviate this limitation, we leverage the prior knowledge that the similarity scores of foreground patches must be greater than those of background features to design a novel local-text rank loss. This loss is designed to increase the differences between the foreground and background similarity scores which push the foreground and background patches apart to strengthen the localization ability of VLMs. In NGA, we exploit the global image-level feature and propose multiple number-conditioned text prompts. Unlike previous single text prompt that merely specify the class, our multiple number-conditioned text prompts not only indicate the class but also have predefined number ranges. And a similarity loss is proposed to align the global image-level feature with the prompt that has the closest number range. Then, to find a more accurate range, a coarse-to-fine global-number alignment scheme is proposed. In the end, the number-conditioned text prompt with the highest similarity is selected to calculate cross-attention with global image-level feature. Considering that the text prompt used in VLMs' pre-training stage usually lacks numeric information, by interacting number-conditioned text prompt with global image-level feature, our proposed NGA significantly improves the counting ability of VLMs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 294, + 713 + ], + "type": "text", + "content": "To verify the effectiveness of our proposed strategy, we" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "conducted experiments on three prevailing datasets: FSC-147 [38], CARPK [15] and ShanghaiTech [55]. Moreover, to perform a more comprehensive evaluation of ZSC methods, we create a larger and more challenging dataset, named ZSC-8K, which increases the proportion of common classes and moves long-tailed categories in FSC-147 to the test set of ZSC-8K to evaluate the performance on long-tailed classes. The results demonstrate that our method significantly surpasses previous text-guided zero-shot object counting methods. Additionally, various backbones including CLIP [37] and GroundingDino [27] are also employed in our experiments to confirm the scalability of our strategy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 553, + 243 + ], + "type": "text", + "content": "In summary, the main contributions of the proposed approach can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 247, + 553, + 474 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 247, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 247, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 314, + 247, + 553, + 307 + ], + "type": "text", + "content": "- We develop a universal strategy consists of Text-guided Local Ranking and Number-evoked Global Attention. This strategy is capable of exploiting both local patch-level features and global image-level feature for text-guided zero-shot object counting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 308, + 553, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 553, + 390 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 553, + 390 + ], + "type": "text", + "content": "- Text-guided Local Ranking is proposed to improve the localization ability of VLMs by using a local-text rank loss to push the foreground and background patches apart. Number-evoked Global Attention is presented to enhance the counting ability of VLMs by computing crossattention between global image-level feature and number-conditioned text prompt with highest similarity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 392, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 392, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 314, + 392, + 553, + 426 + ], + "type": "text", + "content": "- We introduce ZSC-8K, a scaled-up dataset consists of 8,361 images and 114 categories for comprehensive evaluation, which can benefit future researchers in this area." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 427, + 553, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 553, + 474 + ], + "type": "text", + "content": "- The proposed strategy achieves state-of-the-art performance on ZSC-8K dataset and three other popular datasets. Its effectiveness, scalability and cross-dataset generalizability are verified by experiments." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 497, + 405, + 509 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 497, + 405, + 509 + ], + "spans": [ + { + "bbox": [ + 314, + 497, + 405, + 509 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 493, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 493, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 493, + 533 + ], + "type": "text", + "content": "2.1. Category-specific object counting" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 542, + 553, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 553, + 673 + ], + "type": "text", + "content": "Class-specific object counting (CSC) concentrates on counting objects of one specific category, for instance, humans [6, 19, 21, 24, 26, 30, 44, 55], penguins [5], birds [1], cells [48], or cars [15, 31, 32, 53]. Generally, existing CSC approaches can be classified into two groups: point-based methods and density-based methods. Point-based methods [6, 26, 44, 54] apply an object detector to the image and count the number of objects based on the detected points. Density-based methods [22, 39, 55] predict a density map for each input image, and the final result is obtained by summing up the pixel values." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "However, since existing CSC methods need to train a specific network with extensive labeled data for each class of object, they are deemed time-consuming and laborious." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "21098" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 271, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 271, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 271, + 85 + ], + "type": "text", + "content": "2.2. Exemplar-guided class-agnostic counting" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 55, + 89, + 296, + 161 + ], + "type": "text", + "content": "To ease the burden of humans and enable the counting model to count arbitrary categories, researchers propose exemplar-guided class-agnostic counting (CAC) [4, 12, 16, 25, 29, 33, 40, 51, 52] which leverages human annotated images or patches as exemplars of a category to train a unified network that can count arbitrary categories." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 161, + 296, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 161, + 296, + 341 + ], + "spans": [ + { + "bbox": [ + 55, + 161, + 296, + 341 + ], + "type": "text", + "content": "[12] indicates that color, shape and scale diversity are crucial for CAC performance and proposes an EFA module to handle diversity from all aspects. [16] proposes an interactive class-agnostic object counting framework. FamNet [38] presents an adaptation strategy to adapt network to any novel visual category at test time. [33] introduces an uncertainty-aware object detector. BMNet [40] proposes a similarity-aware CAC framework. CFOCNet [51] uses reference images as exemplars to reduce the huge cost of data collection. [52] encourages the model to inspect the query image by focusing more on the regions akin to the support images. CountGD [3] leverages both visual exemplar and text prompt to boost the model. [45] extracts multiple occluded features of exemplars for comparison and utilizes the position of instances for more accurate counting results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 341, + 296, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 296, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 296, + 413 + ], + "type": "text", + "content": "CAC methods usually assume that high-quality bounding boxes of objects can be obtained easily. However, obtaining these exemplars is not as easy as assumed. In practice, the requirement of providing exemplars is unfriendly to users who lack professional knowledge and will be an obstacle when we want to build a fully automatic system." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 255, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 255, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 255, + 434 + ], + "type": "text", + "content": "2.3. Text-guided zero-shot object counting" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 498 + ], + "type": "text", + "content": "Considering that high-quality exemplars are not always easily obtained and aiming to construct a user-friendly system, text-guided zero-shot object counting (ZSC) [2, 8, 18, 20, 41, 49] is proposed to provide a more flexible and user-friendly approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 689 + ], + "type": "text", + "content": "[49] first introduces ZSC and trains a VAE model to generate the class prototype. The generated prototype is used to select class-relevant patches as exemplars to train a counting network. Later, with the application of vision-language models (VLMs) [27, 37], even automatically obtained exemplars are no longer needed. [2, 18, 20, 34] leverage CLIP [37] to solve ZSC task. [18] proposes an end-to-end pipeline to estimate density maps for a given class, [2] introduces a feature interaction module to enhance the feature and [20] incorporates SPT, LAT, and SaSC modules to tailor VLBase for object counting. ABC123 [14] proposed a blind counter but can only count images with fewer than 300 objects. Furthermore, [8] employs GroundingDino [27] as its backbone, and once again, the boundary of ZSC is advanced. VA-Count [56] then uses GroundingDino to generate good exemplars to boost the model." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Nonetheless, previous ZSC approaches only leverage local patch-level features from the vision encoder to interact" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 175 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 175 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 175 + ], + "type": "image", + "image_path": "894ce1e8f5528f8387c77ffcc39724c28db43ce4ac2ef396832bbdfd4abdec9f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 182, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 313, + 182, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 555, + 205 + ], + "type": "text", + "content": "Figure 2. Distribution of each category in FSC-147 [38] (left) and the proposed ZSC-8K (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 207, + 564, + 289 + ], + "blocks": [ + { + "bbox": [ + 315, + 207, + 564, + 289 + ], + "lines": [ + { + "bbox": [ + 315, + 207, + 564, + 289 + ], + "spans": [ + { + "bbox": [ + 315, + 207, + 564, + 289 + ], + "type": "table", + "html": "
#Images#ClassesMaxMinAvgTotal
Overall83611141766146.23386498
Train5476301766149.72272243
Val112729485135.1639623
Testnormal139739953137.1051829
Testlongtail175857953142.4574632
", + "image_path": "ff42230f125cf18df62e2eda154aa6203081be489b0de5ac92d22dea9e09e998.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 555, + 312 + ], + "type": "text", + "content": "Table 1. Statistics of each set in ZSC-8K, including the number of point annotations (max, min, avg, total)." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 323, + 555, + 371 + ], + "type": "text", + "content": "with text feature. The global image-level feature is discarded, and its significance is largely overlooked. It still lacks a systematic method to utilize both local and global features to enhance ZSC approaches." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 381, + 414, + 393 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 414, + 393 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 414, + 393 + ], + "type": "text", + "content": "3. ZSC-8K Dataset" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 402, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 569 + ], + "type": "text", + "content": "FSC-147 [38] dataset is the most widely-used dataset for text-guided zero-shot object counting (ZSC). However, it demonstrates an unrealistic class distribution characterized by excessive long-tailed categories (polka dots, cereals, shallots, etc.) and insufficient samples for common classes. This distributional bias deviates markedly from real-world occurrence patterns. Therefore, to conduct a more comprehensive evaluation of ZSC methods, we propose ZSC-8K dataset, inspired by REC dataset [8]. As shown in Fig. 2, compared to FSC-147, ZSC-8K is larger and more challenging which means it increases the proportion of common classes (e.g. cars, people) and moves long-tailed classes from FSC-147 to test set to evaluate the performance of ZSC methods on long-tailed classes." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 665 + ], + "type": "text", + "content": "Data collection. Images in the ZSC-8K dataset are collected from multiple existing datasets, including FSC147 [38], JHU-Crowd [43], NWPU-Crowd [46], VisDrone [57], DETRAC [47], CARPK [15], MALL [7] and Crowd Surveillance [50]. The proportions of each dataset are depicted in the supplementary material. These datasets belong to different domains, including crowd counting, car counting and animal counting, etc." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 713 + ], + "type": "text", + "content": "Dataset statistics. ZSC-8K contains 8,361 images with a total of 386,498 point annotations. The number of point annotations in an image ranges from 1 to 1,766. A histogram of point annotations per image is provided in the" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21099" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 329, + 259 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 329, + 259 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 329, + 259 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 329, + 259 + ], + "type": "image", + "image_path": "005ba00e8bffe892c82abd4887d803ed6364bfb55f2fd1170f84c2684fadae84.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 261, + 550, + 277 + ], + "lines": [ + { + "bbox": [ + 58, + 261, + 550, + 277 + ], + "spans": [ + { + "bbox": [ + 58, + 261, + 550, + 277 + ], + "type": "text", + "content": "Ft Text feature F1 Local patch-level features Fg Global image-level feature Fct Feature with highest similarity Pct Prompt of Fct" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 54, + 285, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 285, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 285, + 555, + 361 + ], + "type": "text", + "content": "Figure 3. Illustration of our training pipeline. In Text-guided Local Ranking, depending on the prior knowledge of similarity score ranking, local-text rank loss is proposed to push the foreground and background patches apart which further improves the localization ability of VLMs. In Number-evoked Global Attention, a coarse-to-fine global-number alignment scheme is proposed to align global image-level feature with multiple coarse number-conditioned text prompts firstly, and with multiple fine number-conditioned text prompts secondly. The fine number-conditioned text prompt with the highest similarity score is selected to calculate cross-attention with image-level feature. And the feature obtained by cross-attention is added to the local patch-level features. In inference stage, the test images and their corresponding number-conditioned text prompts are forwarded to decoder to make predictions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 330, + 71, + 552, + 259 + ], + "blocks": [ + { + "bbox": [ + 330, + 71, + 552, + 259 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 552, + 259 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 552, + 259 + ], + "type": "image", + "image_path": "54f288501a5899167ac38658c8d651f31d459ebe7b5bae8648d2a09d27e20fa2.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "spans": [ + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": "supplementary material. The dataset can be divided into 5 major categories and 114 subcategories. The proportions of each category are shown in Fig. 2. As it depicts, people are the category with the highest proportion and cars are the second which are significantly higher than that in FSC-147 dataset [38]. We split ZSC-8K dataset into train, validation and test sets. Specific statistics for each set are detailed in Tab. 1. It is worth noting that two versions of test sets are proposed: " + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\text{test}_{\\text{normal}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\text{test}_{\\text{longtail}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": ". In " + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\text{test}_{\\text{normal}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": ", images of normal classes (1397 images & 39 classes) are included, while " + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\text{test}_{\\text{longtail}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": " is obtained by merging the long-tailed classes (361 images & 20 classes) from FSC-147 into " + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "inline_equation", + "content": "\\text{test}_{\\text{normal}}" + }, + { + "bbox": [ + 54, + 373, + 297, + 541 + ], + "type": "text", + "content": ". To meet the zero-shot setting, the categories in train set do not overlap with validation and test sets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 138, + 568 + ], + "type": "text", + "content": "4. Methodology" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 123, + 586 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "The goal of ZSC is to count anything using only text guidance. Pretrained vision-language models (VLMs) are commonly used for this purpose. Denote " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "I \\in R^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " as an input image containing an arbitrary object class, and a text prompt " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is given to specify the object class of this image. In previous methods, " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is initially fed into the image encoder of pretrained VLMs to obtain local patch-level features and " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " is then encoded by the text encoder to obtain text feature. Subsequently, a fuse module is designed to combine the two modalities, and the fused feature is used to predict the num" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 469 + ], + "type": "text", + "content": "ber of the specified category. In this paper, we argue that previous methods only leverage local patch-level features to fuse with the text feature, thereby fully overlooking the importance of the global image-level feature. Therefore, to exploit local patch-level features and global image-level simultaneously, we propose a strategy consists of Text-guided Local Ranking (TLR) and Number-evoked Global Attention (NGA). The overall framework is shown in Fig. 3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "Specifically, the local patch-level features and text feature are used to compute a novel local-text rank loss that can increase the differences between foreground and background similarity scores in TLR. The localization ability of VLMs is improved by TLR. Then, we leverage the global image-level feature in NGA. Multiple number-conditioned text prompts are designed and a coarse-to-fine global-number alignment scheme is proposed to select the number-conditioned text prompt with highest similarity. The selected prompt is considered to have the closest number range and its feature is used to calculate cross-attention with the global image-level feature to enhance the counting ability. Finally, the feature obtained by cross-attention is aggregated with local patch-level features through element-wise summation. This combined representations is then fed into decoder to make predictions. TLR and NGA are detailed in Sec. 4.2 and Sec. 4.3, respectively." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21100" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 206, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 206, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 206, + 85 + ], + "type": "text", + "content": "4.2. Text-guided Local Ranking" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "text", + "content": "VLMs have shown impressive performance in estimating text-image similarity by aligning text and image representations. However, as demonstrated by ClipCount [18], pretrained VLMs intrinsically lack localization ability, making it non-trivial to transfer this ability to object counting. Although ClipCount [18] utilized the local patch-level features " + }, + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "text", + "content": " and the text feature " + }, + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 92, + 296, + 332 + ], + "type": "text", + "content": " to construct a contrastive loss to enhance localization ability by pushing foreground and background patches apart, this loss cannot effectively increase the discrepancy between foreground and background similarity scores, which in turn adversely affects localization ability. As shown in Fig. 4(a), the differences between foreground and background similarity scores in ClipCount [18] are small, with most differences are close to zero. To alleviate this issue, we propose Text-guided Local Ranking (TLR) in which a novel local-text rank loss is designed to improve the localization ability by increasing the differences between foreground and background similarity scores. As Fig. 4(b) shows, with TLR, the differences become much larger than those of ClipCount [18]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": "As shown in Fig. 3, " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "F_{l}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "F_{t}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " are also used in the proposed TLR, a linear projection maps them to the same channels. To determine which patch contains the object, max-pooling is used to transfer the ground truth density map to a patch-level binary mask. Patches whose value is equal to 1 are foreground patches which have the object and the value of background patches is 0. We denote the foreground patches as " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " and the background patches as " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": ", the relation of " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 55, + 335, + 296, + 443 + ], + "type": "text", + "content": " can be summarized as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 92, + 460, + 295, + 475 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 460, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 92, + 460, + 295, + 475 + ], + "type": "interline_equation", + "content": "s \\left(F _ {l} ^ {p}, F _ {t}\\right) > s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\forall p \\in \\mathcal {P}, \\forall q \\in \\mathcal {B}, \\tag {1}", + "image_path": "925aa5accb866cee354934db3a3138490ceab5ffffe72263f097feeb97fa7b6b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 490, + 296, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 490, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 296, + 562 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 490, + 296, + 562 + ], + "type": "inline_equation", + "content": "s(,)" + }, + { + "bbox": [ + 55, + 490, + 296, + 562 + ], + "type": "text", + "content": " denotes the cosine similarity. And Formula 1 indicates the prior knowledge that the similarity score between any foreground patch feature and the text feature is greater than that of any background patch feature. Depending on the prior knowledge, we construct the local-text rank loss. The similarity score sets are derived as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 579, + 295, + 594 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 579, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 86, + 579, + 295, + 594 + ], + "type": "interline_equation", + "content": "S (\\mathcal {P}) = \\{\\dots , s \\left(F _ {l} ^ {p}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {P} | = l _ {1}, \\tag {2}", + "image_path": "a87e95775e517f74112f32173da32dc81bc68b9847db20c706be59e7b7282b04.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 87, + 632, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 632, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 87, + 632, + 295, + 646 + ], + "type": "interline_equation", + "content": "S (\\mathcal {B}) = \\{\\dots , s \\left(F _ {l} ^ {q}, F _ {t}\\right), \\dots \\} / \\tau_ {1}, | \\mathcal {B} | = l _ {2}, \\tag {3}", + "image_path": "600c5be8c5f98ce186e606263d223da6d0f76134b484c3a1e41e9c191fecad2e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "S(\\mathcal{P})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "S(\\mathcal{B})" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " are similarity score sets of foreground and background patches respectively. " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "l_{2}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " are the lengths of the two sets. " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " is a temperature parameter. Next, we denote " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "d_{pq} = (s(F_l^p,F_t) - s(F_l^q,F_t)) / \\tau_1" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", and the" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 70, + 553, + 158 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 158 + ], + "type": "image", + "image_path": "bd21b25f9c9718d218ab757a4a4f6f2285321c40a52ec15b800eeaacd0cdd3fb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 161, + 555, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 161, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 555, + 228 + ], + "type": "text", + "content": "Figure 4. The differences between foreground and background similarity scores. (a) shows the differences between foreground and background scores in ClipCount [18] are small, most of the differences are close to 0. (b) shows the differences with our TLR become larger. Hence, the features of foreground and background patches are pushed apart." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 233, + 533, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 533, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 533, + 246 + ], + "type": "text", + "content": "differences between the two sets can be computed by:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 353, + 253, + 555, + 329 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 253, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 353, + 253, + 555, + 329 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} M _ {d i f} = S (\\mathcal {P}) \\stackrel {\\text {b r o a d c a s t - s u b t r a c t}} {\\longrightarrow} S (\\mathcal {B}) \\\\ = \\left[ \\begin{array}{c c c c} d _ {1 1} & d _ {1 2} & \\dots & d _ {1 l _ {2}} \\\\ d _ {2 1} & d _ {2 2} & \\dots & d _ {2 l _ {2}} \\\\ \\vdots & \\vdots & \\ddots & \\vdots \\\\ d _ {l _ {1} 1} & d _ {l _ {1} 2} & \\dots & d _ {l _ {1} l _ {2}} \\end{array} \\right] \\quad , \\tag {4} \\\\ \\end{array}", + "image_path": "77a6416aadc575f2ce235761a7bc969393c68b4e9a9aa102db7ad67a2f8e19d6.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "inline_equation", + "content": "M_{dif}" + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "text", + "content": " is the result of subtracting " + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "inline_equation", + "content": "S(\\mathcal{B})" + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "inline_equation", + "content": "S(\\mathcal{P})" + }, + { + "bbox": [ + 313, + 335, + 554, + 372 + ], + "type": "text", + "content": " using the broadcasting mechanism. Finally, with the prior knowledge, the local-text rank loss can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 347, + 378, + 554, + 404 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 378, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 347, + 378, + 554, + 404 + ], + "type": "interline_equation", + "content": "L _ {\\text {r a n k}} = \\frac {\\sum \\operatorname {c l a m p} \\left(\\lambda - M _ {\\text {d i f}} , \\min = 0\\right)}{l _ {1} l _ {2}}, \\tag {5}", + "image_path": "b5a04432fb1d1ad02a7b776eab724ae5cd3ba99c0d8ff91b4dc0cc705ac0c644.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 411, + 534, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 534, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 534, + 422 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 411, + 534, + 422 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 411, + 534, + 422 + ], + "type": "text", + "content": " is a hyperparameter that controls the rank loss." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 423, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 483 + ], + "type": "text", + "content": "With the proposed local-text rank loss, the differences are increased by ensuring the similarity scores of foreground patches are bigger than background patches. Thus the features of foreground and background patches are pushed apart and the localization ability is improved." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 489, + 496, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 496, + 501 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 496, + 501 + ], + "type": "text", + "content": "4.3. Number-Evoked Global Attention" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "content": "Because most of the text prompts used in pre-training stage of VLMs lack descriptions of the number of objects, the performance of VLMs in object counting is largely limited. To mitigate this limitation, we propose Number-Evoked Global Attention (NGA) to enhance the counting ability of VLMs by calculating the cross-attention between the global image-level feature " + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "content": " and the feature of the number-conditioned text prompt with highest similarity " + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "inline_equation", + "content": "F_{ct}^{h}" + }, + { + "bbox": [ + 313, + 506, + 554, + 602 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "text", + "content": "First, " + }, + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "text", + "content": " that is ignored by pervious ZSC methods is leveraged to select " + }, + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{ct}^{h}" + }, + { + "bbox": [ + 313, + 602, + 555, + 714 + ], + "type": "text", + "content": ", the number-conditioned text prompt with highest similarity, as shown in Fig. 3. Specifically, we first propose a global-number alignment scheme to identify a coarse number range for an image. To achieve this, we replace the single text prompt by multiple number-conditioned text prompts. Different from previous single text prompt which only specify the object class, every prompt of our proposed multiple number-conditioned text prompts has an object class and a predefined number range." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21101" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 563, + 275 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 563, + 275 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 563, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 563, + 275 + ], + "type": "table", + "html": "
MethodBackboneVenueVal setTest set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ZSC [49]ResNet+VAECVPR'2326.9388.6322.09115.17
TFOC [41]SAMWACV'24--20.10132.83
CountTX* [2]CLIPBMCV'2317.7665.2116.70105.21
VA-Count [56]CLIPECCV'2417.8773.2217.88129.31
GeCo [36]CLIPNeurIPS'2414.8164.9513.30108.72
DAVEprm [35]CLIPCVPR'2415.4852.5714.90103.42
GroundingREC* [8]GDinoCVPR'2412.3761.8611.27108.21
VLCounter [20]CLIPAAAI'2418.0665.1317.05106.16
VLCounter+OursCLIP-17.88(-0.18)63.88(-1.25)16.56(-0.49)107.47(+1.31)
ClipCount [18]CLIPMM'2318.9366.1017.45104.88
ClipCount+OursCLIP-17.52(-1.41)58.39(-7.71)16.42(-1.03)100.63(-4.25)
CountGD [3]GDinoNeurIPS'2412.1447.5114.76120.42
CountGD+OursGDino-11.62(-0.52)46.37(-1.14)12.84(-1.92)115.61(-4.81)
GDino [27](w/ finetune)*GDinoECCV'2412.0762.1311.76104.98
GDino(w/ finetune)+OursGDino-11.09(-0.98)60.48(-1.65)10.74(-1.02)106.5(+1.52)
", + "image_path": "8503dddfdfc2f3bf94de412a48e3218ebd1a2be6d143a48cf9f3a8a2724f834f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 275, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 55, + 275, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 555, + 298 + ], + "type": "text", + "content": "Table 2. Comparison with state-of-the-art methods on FSC-147 dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 307, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 307, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 296, + 396 + ], + "type": "text", + "content": "We initially design 5 distinct number ranges and leverage these ranges to generate 5 different number-conditioned text prompts such as \"There are between {number} and {number} {class}\" . Then, a global-number similarity loss is proposed to maximize the similarity between the global image-level feature and the positive number-conditioned text prompt which has the closest number range. The global-number similarity loss can be formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 400, + 298, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 298, + 435 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 298, + 435 + ], + "type": "interline_equation", + "content": "L _ {s i m} = - \\log \\frac {\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right)}{\\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {i}\\right) / \\tau_ {2}\\right) + \\sum_ {j \\in \\mathcal {N}} \\exp \\left(s \\left(F _ {g} , F _ {c t} ^ {j}\\right) / \\tau_ {2}\\right)}, \\tag {6}", + "image_path": "1364f33413609a08dfe83b13469c615d5319ce15848a036ae67c29775ebc131f.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "s(\\cdot)" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": " denotes the cosine similarity, " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": " is a temperature parameter and " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "F_{ct}^{i}" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": " is the positive number-conditioned text prompt embedding. " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "\\mathcal{N}" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": " denotes the set of negative number-conditioned text prompt embeddings. For CLIP-based methods, we take the global `cls_token` as our " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": ", while for GroundingDino[27], we obtain the " + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 55, + 437, + 296, + 519 + ], + "type": "text", + "content": " by using average pooling on the last layer features of Swin-T." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 521, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 663 + ], + "type": "text", + "content": "Through above global-number alignment scheme, model can learn to find a coarse number range for an image. But this coarse range is still too wide to serve as an accurate prompt. Hence, as shown in Fig. 3, we further propose a coarse-to-fine global-number alignment scheme to generate more accurate number-conditioned text prompt. The number-conditioned text prompts are designed to be hierarchical, which means that the coarse range with highest similarity is further divided into 5 ranges, and the global-number similarity loss is adopted again to maximize the similarity of the global image-level feature and the positive fine number-conditioned text prompt." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 664, + 274, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 664, + 274, + 675 + ], + "spans": [ + { + "bbox": [ + 67, + 664, + 274, + 675 + ], + "type": "text", + "content": "Overall, the loss of our NGA can be formulated as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 681, + 295, + 697 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 681, + 295, + 697 + ], + "spans": [ + { + "bbox": [ + 102, + 681, + 295, + 697 + ], + "type": "interline_equation", + "content": "L _ {N G A} = \\delta L _ {s i m} ^ {\\text {c o a r s e}} + (1 - \\delta) L _ {s i m} ^ {\\text {f i n e}}, \\tag {7}", + "image_path": "140a24d56b52697dacee5b42a7aff38f9cb8422f56ed6818911b757677def098.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{sim}^{coarse}" + }, + { + "bbox": [ + 55, + 701, + 296, + 715 + ], + "type": "text", + "content": " is the global-number similarity loss of coarse" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": "number-conditioned text prompts and " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "L_{sim}^{fine}" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": " is for fine number-conditioned text prompts. " + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 306, + 555, + 342 + ], + "type": "text", + "content": " is a weight term to balance the effect of coarse and fine similarity loss." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "text", + "content": "With our proposed scheme, " + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "inline_equation", + "content": "P_{ct}^{h}" + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "text", + "content": " is selected and considered to have the closest number range. The feature of it " + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "inline_equation", + "content": "F_{ct}^{h}" + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "text", + "content": " is used to compute cross-attention with " + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "inline_equation", + "content": "F_{g}" + }, + { + "bbox": [ + 313, + 342, + 555, + 392 + ], + "type": "text", + "content": " in NGA, thereby the counting ability of VLMs is enhanced." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 395, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 395, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 395, + 417 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 423, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 568 + ], + "type": "text", + "content": "Firstly, the experimental setting, including implementation details and evaluation metrics, is given in Sec. 5.1. Then, we apply our strategy to three different methods and compare their performances with state-of-the-art text-guided zero-shot object counting approaches in Sec. 5.2. Through ablation studies in Sec. 5.3, we quantitatively evaluate the effectiveness of each component in our strategy. Next, we conduct experiments to discuss the impact of the coarse-to-fine global-number alignment scheme, the effect of our strategy on cross-dataset generalization and the influence of " + }, + { + "bbox": [ + 313, + 423, + 555, + 568 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 423, + 555, + 568 + ], + "type": "text", + "content": " in Sec. 5.4. Finally, visualizations in Sec. 5.5 qualitatively verify the efficacy of our strategy." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 576, + 437, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 437, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 437, + 589 + ], + "type": "text", + "content": "5.1. Experimental Setting" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Implementation details. To evaluate our strategy comprehensively, we apply our strategy to two CLIP-based methods [18, 20] and one GroundingDino(GDino)-based method [27]. For CLIP-based methods, the vision encoder is VIT-B/16 [11]. VPT [17] is used with 20 visual prompts in each layer. The maxpooling in TLR adopts a kernel size of " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", stride=16 and padding=0. For the density decoder, we use " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " convolution kernel with stride=1 and padding=1 to keep resolution. The batch size is 32 and the overall training epochs are 440. For GDino-based method," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21102" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 159 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 159 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 159 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 159 + ], + "type": "table", + "html": "
MethodBackboneVal setTestnormal setTestlongtail set
MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)MAE (↓)RMSE (↓)
ClipCount* [18]CLIP17.5330.4218.6838.6122.6553.79
ClipCount+OursCLIP16.65(-0.88)29.82(-0.6)17.24(-1.44)36.92(-1.69)21.90(-0.75)53.66(-0.13)
GroundingREC* [8]GDino14.3425.0113.7828.9616.6541.80
GDino* [27](w/ finetune)GDino13.9221.6414.2329.6816.5738.85
GDino(w/ finetune)+OursGDino11.17(-2.75)17.30(-4.34)12.39(-1.84)27.75(-1.93)14.68((-1.89)38.10((-0.75)
", + "image_path": "5d15b7a34bc40a56b652c5f6293f505209dc0a7a65b9eac5341b13d7a18497ee.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 188, + 298, + 269 + ], + "blocks": [ + { + "bbox": [ + 56, + 161, + 553, + 184 + ], + "lines": [ + { + "bbox": [ + 56, + 161, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 553, + 184 + ], + "type": "text", + "content": "Table 3. Comparison with state-of-the-art methods on our ZSC-8K dataset. Bold results are the best performance. Performance improvements by our strategy are indicated in parentheses. * means the results are reproduced by open-source codes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 188, + 298, + 269 + ], + "lines": [ + { + "bbox": [ + 56, + 188, + 298, + 269 + ], + "spans": [ + { + "bbox": [ + 56, + 188, + 298, + 269 + ], + "type": "table", + "html": "
MethodVal setTest set
MAERMSEMAERMSE
None18.9366.1017.45104.88
TLR18.4260.7516.71100.08
NGA18.2962.9116.81103.21
TLR+NGA17.5258.3916.42100.63
", + "image_path": "41debf986d6be280bcc0af8202ea5e7988580e1744044f0af6c94cd37109aa49.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 269, + 271, + 280 + ], + "lines": [ + { + "bbox": [ + 79, + 269, + 271, + 280 + ], + "spans": [ + { + "bbox": [ + 79, + 269, + 271, + 280 + ], + "type": "text", + "content": "Table 4. Ablation study on proposed NGA and TLR." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "content": "the text encoder is BERT [9] and visual encoder is SwinT [28]. The overall training epochs are 30. And the other hyperparameters are same with [8]. We use AdamW optimizer to optimize our model, for CLIP-based methods, the learning rate is 1e-4 and for GDino-based method, the learning rate is 1e-5. For all methods, " + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "content": " is set as 0.5. " + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "content": " is set as 10. Following [18], " + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\tau_{1}" + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "inline_equation", + "content": "\\tau_{2}" + }, + { + "bbox": [ + 55, + 293, + 295, + 376 + ], + "type": "text", + "content": " are set as 0.07." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": "Evaluation metrics. Mean Absolute Error (MAE) and Root Mean Squared Error (RMSE) are widely used evaluation metrics in text-guided zero-shot counting tasks which is also applied in this paper. MAE = " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": " and RMSE = " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\sqrt{\\frac{1}{N}\\sum_{m=1}^{N}|C_m^{pred} - C_m^{gt}|^2}" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": " is the total number of testing images, " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "C_m^{pred}" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "C_m^{gt}" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": " are the prediction and the ground truth count of the " + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 55, + 377, + 295, + 469 + ], + "type": "text", + "content": "-th image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 479, + 264, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 264, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 264, + 491 + ], + "type": "text", + "content": "5.2. Compare with State-of-the-art Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 556 + ], + "type": "text", + "content": "In this section, we apply the proposed strategy to three existing approaches and compare the performances with state-of-the-art text-guided zero-shot object counting methods on two datasets: FSC-147 [38] and ZSC-8K, to evaluate the effectiveness of our strategy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": "The results on FSC-147 are shown in Tab. 2. As it depicts, our approach outperforms all previous state-of-the-art methods, regardless of the backbone. Compared with the best performances of previous methods, our approach reduces the MAE by " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "8.1\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "4.7\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " on the validation set and test set. And our method decreases the RMSE by " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "2.5\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 688 + ], + "type": "text", + "content": " on the validation set and test set respectively. Our method also shows its scalability by comparisons with the baselines, the performances of VLCounter+Ours, ClipCount+Ours, CountGD+Ours and GDino(w/ finetune)+Ours are all better than their baselines." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Then, we compare the performance of our method with previous ZSC methods on our proposed dataset, ZSC-8K." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "text", + "content": "The results in Tab. 3 demonstrate that our method obtains the best performance on ZSC-8K dataset. Specifically, compared to the previous best results, the MAE is reduced by " + }, + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "inline_equation", + "content": "19.8\\%" + }, + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "inline_equation", + "content": "10.1\\%" + }, + { + "bbox": [ + 313, + 190, + 553, + 310 + ], + "type": "text", + "content": " on validation set and testnormal set. And every baseline is improved by our strategy. Furthermore, we evaluate ZSC methods on the testlongtail set in the last two columns of Tab. 3. Although our method still achieves optimal performance, the results indicate a significant performance gap between normal and long-tailed classes, highlighting the limitation of ZSC methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 376 + ], + "type": "text", + "content": "The remarkable results on the FSC-147 and ZSC-8K datasets illustrate the effectiveness of our strategy. In particular, whether CLIP or GDino is used as the backbone, our method significantly outperforms the baselines, which strongly validates the scalability of our strategy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 398, + 408, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 408, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 408, + 411 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 421, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 457 + ], + "type": "text", + "content": "We conduct ablation experiments to verify the effectiveness of the proposed Text-guided Local Ranking(TLR) and Number-Evoked Global Attention(NGA) in Tab. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": "All experiments are conducted with ClipCount [18] on FSC-147 [38]. We start with a baseline where no modules are applied. It achieves 18.93 MAE and 66.10 RMSE on validation set, 17.45 MAE and 104.88 RMSE on test set. Then, the single use of TLR and NGA is analyzed. By leveraging the proposed local-text rank loss, the foreground and background patches are pushed apart and the localization ability of VLMs is improved by TLR. The effectiveness is verified by the results shown in Tab. 4. Compared to baseline, TLR improves the MAE by " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " on validation set and " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "4.2\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " on test set. Next, the proposed NGA is adopted to study its effectiveness. With NGA which calculates crossattention between the number-conditioned text prompt and global image-level feature, its performance significantly improved, the error reduces by " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "3.7\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " for MAE and RMSE on test set, respectively. Hence, the results verify that the proposed NGA can enhance the counting ability of VLMs. Finally, the combination of TLR and NGA enhances the overall performance even further, improving the MAE by " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " on validation set and " + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 462, + 553, + 715 + ], + "type": "text", + "content": " on test set, suggesting the proposed TLR and NGA are complementary." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21103" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 298, + 127 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 298, + 127 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 298, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 298, + 127 + ], + "type": "table", + "html": "
MethodVal setTest set
MAERMSEMAERMSE
w/ coarse18.1761.7317.34100.79
w/ coarse-to-fine17.5258.3916.42100.63
", + "image_path": "65c9bf6b490f1daa83d949f7a62007f4a43bba38f03edf3975c37ad1d6825a45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 152, + 301, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 128, + 295, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 128, + 295, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 295, + 149 + ], + "type": "text", + "content": "Table 5. The effect of the proposed coarse-to-fine global-number alignment scheme." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 152, + 301, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 152, + 301, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 301, + 232 + ], + "type": "table", + "html": "
MethodTypeTraining→TestingMAERMSE
MCNN [55]SpecificSHB→SHA221.4357.8
CrowdClip [23]SpecificSHB→SHA217.0322.7
RCC [13]GenericFSC→SHA240.1366.9
ClipCount [18]GenericFSC→SHA192.6308.4
ClipCount+OursGenericFSC→SHA176.24279.20
", + "image_path": "437dcb114426ed314d493750b4570b0ffd8d6d254c7842c09824a84975f8dbef.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "lines": [ + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 295, + 266 + ], + "type": "text", + "content": "Table 6. Cross-dataset evaluation on ShanghaiTech dataset. \"SHA\", \"SHB\" and \"FSC\" denote ShangTech Part A, ShanghaiTech Part B and FSC-147 datasets respectively." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 276, + 132, + 288 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 132, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 132, + 288 + ], + "type": "text", + "content": "5.4. Discussions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 294, + 295, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 295, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 295, + 389 + ], + "type": "text", + "content": "The effect of coarse-to-fine global-number alignment scheme. A coarse-to-fine global-number alignment scheme is designed to select accurate number-conditioned text prompt to enhance the counting ability of VLMs in NGA. We conduct experiments to study the effect of this scheme. \"w/ coarse\" means we only use coarse number-conditioned text prompts. All experiments are performed on the FSC-147 dataset with ClipCount [18]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 605 + ], + "type": "text", + "content": "As Tab. 5 shows, with the proposed scheme, it achieves 17.52 MAE on validation set and 16.42 MAE on test set which are better than only applying coarse number-conditioned text prompts. The results prove the effectiveness of our coarse-to-fine global-number alignment scheme. The generalizing ability across datasets. To evaluate the cross-dataset generalization ability, we compare our approach with ZSC methods on the ShanghaiTech [55] crowd counting dataset. On ShanghaiTech dataset, we compare with specific methods which are category-specific and generic methods that are ZSC methods. The results are summarized in Tab. 6. Our proposed generic object counting approach outperforms both category-specific counting methods and ZSC methods. It surpasses CrowdClip [23] by 40.76 MAE and 42.8 RMSE without training on crowd counting dataset. Moreover, compared to the second best ZSC method, it gains 16.36 MAE improvement and 29.2 RMSE improvement." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Overall, above results demonstrate that our strategy can significantly improve the cross-dataset generalizing ability. The effect of " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " is used to balance the effect of coarse and fine similarity losses. Experiments are conducted on FSC-147 dataset with ClipCount[18] to analyze it. As Fig. 5 shows, " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta = 0.5" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " achieves the best performance on both val and test sets. It is also indicated that a small value of " + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": " (0.1) may be inadequate to find a accurate coarse range, while a large value (0.9) would lose fine-grained information." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 553, + 187 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 553, + 187 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 553, + 187 + ], + "type": "image", + "image_path": "ab6934e464cf5976a0a0ba2feeadb048f93a60aa63a700748f60248a362b7223.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 313, + 188, + 553, + 329 + ], + "blocks": [ + { + "bbox": [ + 313, + 188, + 553, + 329 + ], + "lines": [ + { + "bbox": [ + 313, + 188, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 188, + 553, + 329 + ], + "type": "image", + "image_path": "1e738f1126779289492e43953de7de0b2bb63be4a318543018b51710012f3bc7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 330, + 553, + 351 + ], + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 351 + ], + "type": "text", + "content": "Figure 6. Qualitative results on FSC147 and ZSC-8K. Ground-truth numbers are shown at the top-right corner." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 352, + 398, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 352, + 398, + 363 + ], + "spans": [ + { + "bbox": [ + 314, + 352, + 398, + 363 + ], + "type": "text", + "content": "5.5. Visualization" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 555, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 555, + 466 + ], + "type": "text", + "content": "To examine the effect of our approach qualitatively, we visualize some results predicted by ClipCount [18] and our approach in Fig. 6. As the Fig. 6 shows, the results of our method are closer to the ground-truth number and our method can detect the objects that are missed by ClipCount. These visualizations further verify that our approach can improve the counting ability and localization ability of VLMs. More qualitative results are shown in supplement." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 388, + 483 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we present a universal strategy for text-guided zero-shot object counting. We introduce Text-guided Local Ranking to improve the localization ability of VLMs by a local-text rank loss, which effectively push the features of foreground and background patches apart. Additionally, we propose Number-evoked Global Attention, which leverages the global image-level feature to align with multiple number-conditioned text prompts, and then selects the prompt with highest similarity to calculate crossattention with the global image-level feature, thereby enhancing the counting ability. Furthermore, we propose a dataset named ZSC-8K for comprehensive evaluation. Extensive experiments conducted on the ZSC-8K dataset and three other public datasets demonstrate the effectiveness, scalability, and cross-dataset generalizability of our method. Although effective, the results on the long-tailed test set reveal substantial performance gaps when handling long-tailed classes. This persistent challenge across all ZSC methods highlights a critical direction for future research." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21104" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 147, + 85 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 139 + ], + "type": "text", + "content": "This work was supported in part by the National Nature Science Foundation of China (NSFC) under Grant 62376209; and in part by the Fundamental Research Funds for the Central Universities under Grant xtr072024017." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 114, + 163 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 171, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 171, + 294, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 294, + 214 + ], + "type": "text", + "content": "[1] Hüseyin Gökhan Akçay, Bekir Kabasakal, Duygugül Aksu, Nusret Demir, Melih Öz, and Ali Erdoğan. Automated bird counting with deep learning for regional bird distribution mapping. Animals, 10(7):1207, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 216, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 295, + 248 + ], + "type": "text", + "content": "[2] Niki Amini-Naeni, Kiana Amini-Naeni, Tengda Han, and Andrew Zisserman. Open-world text-specified object counting. BMVC, 2023. 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 251, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 281 + ], + "type": "text", + "content": "[3] Niki Amini-Naieni, Tengda Han, and Andrew Zisserman. Countgd: Multi-modal open-world counting. NeurIPS, 2024. 3, 6, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 317 + ], + "type": "text", + "content": "[4] Carlos Arteta, Victor Lempitsky, J Alison Noble, and Andrew Zisserman. Interactive object counting. In ECCV, pages 504-518. Springer, 2014. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 350 + ], + "type": "text", + "content": "[5] Carlos Arteta, Victor Lempitsky, and Andrew Zisserman. Counting in the wild. In ECCV, pages 483-498. Springer, 2016. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 353, + 294, + 395 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 353, + 294, + 395 + ], + "spans": [ + { + "bbox": [ + 62, + 353, + 294, + 395 + ], + "type": "text", + "content": "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Ming-Hsuan Yang, SyYen Kuo, et al. Improving point-based crowd counting and localization based on auxiliary point guidance. ECCV, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 398, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 398, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 398, + 294, + 430 + ], + "type": "text", + "content": "[7] Ke Chen, Shaogang Gong, Tao Xiang, and Chen Change Loy. Cumulative attribute space for age and crowd density estimation. In CVPR, pages 2467-2474, 2013. 3, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 433, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 433, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 62, + 433, + 294, + 463 + ], + "type": "text", + "content": "[8] Siyang Dai, Jun Liu, and Ngai-Man Cheung. Referring expression counting. In CVPR, pages 16985-16995, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 466, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 466, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 466, + 294, + 509 + ], + "type": "text", + "content": "[9] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In naacL-HLT, page 2. Minneapolis, Minnesota, 2019. 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 511, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 511, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 511, + 294, + 565 + ], + "type": "text", + "content": "[10] SongLin Dong, ChengLi Tan, ZhenTao Zuo, YuHang He, YiHong Gong, TianGang Zhou, JunMin Liu, and JiangShe Zhang. Brain-inspired dual-pathway neural network architecture and its generalization analysis. Science China Technological Sciences, 67(8):2319-2330, 2024. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. arXiv preprint arXiv:2010.11929, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "text", + "content": "[12] Shenjian Gong, Shanshan Zhang, Jian Yang, Dengxin Dai, and Bernt Schiele. Class-agnostic object counting robust to intraclass diversity. In ECCV, pages 388–403. Springer, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Michael Hobley and Victor Prisacariu. Learning to count anything: Reference-less class-agnostic counting with weak supervision. arXiv preprint arXiv:2205.10203, 2022. 8, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[14] M. Hobley and V. Prisacariu. Abc easy as 123: A blind counter for exemplar-free multi-class class-agnostic counting. In ECCV, 2024. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Meng-Ru Hsieh, Yen-Liang Lin, and Winston H Hsu. Drone-based object counting by spatially regularized regional proposal network. In ICCV, pages 4145-4153, 2017. 2, 3, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 137 + ], + "type": "text", + "content": "[16] Yifeng Huang, Viresh Ranjan, and Minh Hoai. Interactive class-agnostic object counting. In ICCV, pages 22312–22322, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 182 + ], + "type": "text", + "content": "[17] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727. Springer, 2022. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 184, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 216 + ], + "type": "text", + "content": "[18] Ruixiang Jiang, Lingbo Liu, and Changwen Chen. Clipcount: Towards text-guided zero-shot object counting. In ACM Multimedia, pages 4535-4545, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 258 + ], + "type": "text", + "content": "[19] Xiaoheng Jiang, Li Zhang, Mingliang Xu, Tianzhu Zhang, Pei Lv, Bing Zhou, Xin Yang, and Yanwei Pang. Attention scaling for crowd counting. In CVPR, pages 4705-4714, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 261, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 302 + ], + "type": "text", + "content": "[20] Seunggu Kang, WonJun Moon, Euiyeon Kim, and Jae-Pil Heo. Vlcounter: Text-aware visual representation for zero-shot object counting. In AAAI, pages 2714-2722, 2024. 3, 6, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 327 + ], + "type": "text", + "content": "[21] Victor S. Lempitsky and Andrew Zisserman. Learning to count objects in images. In NIPS, pages 1324-1332, 2010. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 328, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 328, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 317, + 328, + 553, + 369 + ], + "type": "text", + "content": "[22] Yuhong Li, Xiaofan Zhang, and Deming Chen. CSRNet: Dilated convolutional neural networks for understanding the highly congested scenes. In CVPR, pages 1091-1100, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 414 + ], + "type": "text", + "content": "[23] Dingkang Liang, Jiahao Xie, Zhikang Zou, Xiaqing Ye, Wei Xu, and Xiang Bai. Crowdclip: Unsupervised crowd counting via vision-language model. In CVPR, pages 2893-2903, 2023. 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 448 + ], + "type": "text", + "content": "[24] Hui Lin, Zhiheng Ma, Rongrong Ji, Yaowei Wang, and Xiaopeng Hong. Boosting crowd counting via multifaceted attention. In CVPR, pages 19596-19605, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 480 + ], + "type": "text", + "content": "[25] Chang Liu, Yujie Zhong, Andrew Zisserman, and Weidi Xie. Countr: Transformer-based generalised visual counting. BMVC, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 483, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 514 + ], + "type": "text", + "content": "[26] Chengxin Liu, Hao Lu, Zhiguo Cao, and Tongliang Liu. Point-query quadtree for crowd counting, localization, and more. In ICCV, pages 1676-1685, 2023. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 568 + ], + "type": "text", + "content": "[27] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 613 + ], + "type": "text", + "content": "[28] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, pages 10012-10022, 2021. 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 615, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 553, + 635 + ], + "type": "text", + "content": "[29] Erika Lu, Weidi Xie, and Andrew Zisserman. Class-agnostic counting. In ACCV, pages 669–684. Springer, 2019. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 637, + 553, + 669 + ], + "type": "text", + "content": "[30] Zhiheng Ma, Xing Wei, Xiaopeng Hong, and Yihong Gong. Bayesian loss for crowd count estimation with point supervision. In ICCV, pages 6141-6150, 2019. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[31] Thomas Moranduzzo and Farid Melgani. Automatic car counting method for unmanned aerial vehicle images. IEEE Transactions on Geoscience and remote sensing, 52(3): 1635-1647, 2013. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21105" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 117 + ], + "type": "text", + "content": "[32] T Nathan Mundhenk, Goran Konjevod, Wesam A Sakla, and Kofi Boakye. A large contextual dataset for classification, detection and counting of cars with deep learning. In ECCV, pages 785-800. Springer, 2016. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[33] Thanh Nguyen, Chau Pham, Khoi Nguyen, and Minh Hoai. Few-shot object counting and detection. In ECCV, pages 348-365. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 151, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 151, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 151, + 294, + 196 + ], + "type": "text", + "content": "[34] Roni Paiss, Ariel Ephrat, Omer Tov, Shiran Zada, Inbar Mosseri, Michal Irani, and Tali Dekel. Teaching clip to count to ten. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 3170-3180, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 229 + ], + "type": "text", + "content": "[35] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. Dave - a detect-and-verify paradigm for low-shot counting. CVPR, 2024. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 274 + ], + "type": "text", + "content": "[36] Jer Pelhan, Alan Lukezic, Vitjan Zavrtanik, and Matej Kristan. A novel unified architecture for low-shot counting by detection and segmentation. Advances in Neural Information Processing Systems, 37:66260-66282, 2024. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 330 + ], + "type": "text", + "content": "[37] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[38] Viresh Ranjan, Udbhav Sharma, Thu Nguyen, and Minh Hoai. Learning to count everything. In CVPR, pages 3394-3403, 2021. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[39] Deepak Babu Sam, Shiv Surya, and R. Venkatesh Babu. Switching convolutional neural network for crowd counting. In CVPR, pages 4031-4039, 2017. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 442 + ], + "type": "text", + "content": "[40] Min Shi, Hao Lu, Chen Feng, Chengxin Liu, and Zhiguo Cao. Represent, compare, and learn: A similarity-aware framework for class-agnostic counting. In CVPR, pages 9529–9538, 2022. 1, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[41] Zenglin Shi, Ying Sun, and Mengmi Zhang. Training-free object counting with prompts. In WACV, pages 323-331, 2024. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 294, + 510 + ], + "type": "text", + "content": "[42] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 555 + ], + "type": "text", + "content": "[43] Vishwanath A. Sindagi, Rajeev Yasarla, and Vishal M. Patel. JHU-CROWD++: large-scale crowd counting dataset and A benchmark method. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(5):2594-2609, 2022. 3, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 294, + 609 + ], + "type": "text", + "content": "[44] Qingyu Song, Changan Wang, Zhengkai Jiang, Yabiao Wang, Ying Tai, Chengjie Wang, Jilin Li, Feiyue Huang, and Yang Wu. Rethinking counting and localization in crowds: A purely point-based framework. In ICCV, pages 3365-3374, 2021. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 612, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 294, + 645 + ], + "type": "text", + "content": "[45] Yuejiao Su, Yi Wang, Lei Yao, and Lap-Pui Chau. Few-shot class-agnostic counting with occlusion augmentation and localization. In ISCAS, pages 1-5. IEEE, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[46] Qi Wang, Junyu Gao, Wei Lin, and Xuelong Li. NWPU-crowd: A large-scale benchmark for crowd counting and localization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(6):2141-2149, 2021. 3, 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 294, + 714 + ], + "type": "text", + "content": "[47] Longyin Wen, Dawei Du, Zhaowei Cai, Zhen Lei, Ming-Ching Chang, Honggang Qi, Jongwoo Lim, Ming-Hsuan" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 554, + 531 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 72, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 554, + 106 + ], + "type": "text", + "content": "Yang, and Siwei Lyu. Ua-detrac: A new benchmark and protocol for multi-object detection and tracking. Computer Vision and Image Understanding, 193:102907, 2020. 3, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 161 + ], + "type": "text", + "content": "[48] Weidi Xie, J Alison Noble, and Andrew Zisserman. Microscopy cell counting and detection with fully convolutional regression networks. Computer methods in biomechanics and biomedical engineering: Imaging & Visualization, 6(3): 283-292, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 554, + 195 + ], + "type": "text", + "content": "[49] Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, and Dimitris Samaras. Zero-shot object counting. In CVPR, pages 15548-15557, 2023. 2, 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 240 + ], + "type": "text", + "content": "[50] Zhaoyi Yan, Yuchen Yuan, Wangmeng Zuo, Xiao Tan, Yezhen Wang, Shilei Wen, and Errui Ding. Perspective-guided convolution networks for crowd counting. In ICCV, pages 952-961, 2019. 3, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "text", + "content": "[51] Shuo-Diao Yang, Hung-Ting Su, Winston H. Hsu, and Wen-Chin Chen. Class-agnostic few-shot object counting. In WACV, pages 870–878, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 274, + 554, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 554, + 317 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 554, + 317 + ], + "type": "text", + "content": "[52] Zhiyuan You, Kai Yang, Wenhan Luo, Xin Lu, Lei Cui, and Xinyi Le. Few-shot object counting with similarity-aware feature enhancement. In WACV, pages 6315-6324, 2023. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 554, + 364 + ], + "type": "text", + "content": "[53] Shanghang Zhang, Guanhang Wu, Joao P Costeira, and José MF Moura. Fcn-rlstm: Deep spatio-temporal neural networks for vehicle counting in city cameras. In ICCV, pages 3667-3676, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 398 + ], + "type": "text", + "content": "[54] Shiwei Zhang, Wei Ke, Shuai Liu, Xiaopeng Hong, and Tong Zhang. Boosting semi-supervised crowd counting with scale-based active learning. In ACM Multimedia, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 442 + ], + "type": "text", + "content": "[55] Yingying Zhang, Desen Zhou, Siqin Chen, Shenghua Gao, and Yi Ma. Single-image crowd counting via multi-column convolutional neural network. In CVPR, pages 589-597, 2016. 2, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 443, + 554, + 475 + ], + "type": "text", + "content": "[56] H.L Zhu, J.L Yuan, Z.W Yang, Y. Guo, Z. Wang, X. Zhong, and S.F He. Zero-shot object counting with good exemplars. In ECCV, 2024. 3, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 554, + 531 + ], + "type": "text", + "content": "[57] Pengfei Zhu, Longyin Wen, Dawei Du, Xiao Bian, Heng Fan, Qinghua Hu, and Haibin Ling. Detection and tracking meet drones challenge. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7380-7399, 2021. 3, 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21106" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_content_list.json b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..96ed8b77a25cdefe1e8c3c449b4f570eff4eed25 --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_content_list.json @@ -0,0 +1,1555 @@ +[ + { + "type": "text", + "text": "Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need", + "text_level": 1, + "bbox": [ + 179, + 130, + 821, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yongchuan Cui $^{1,2}$ Peng Liu $^{1,2,*}$ Hui Zhang $^{3,*}$", + "bbox": [ + 272, + 202, + 725, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Aerospace Information Research Institute, Chinese Academy of Sciences, 100094 Beijing, China", + "bbox": [ + 109, + 220, + 887, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, 101408 Beijing, China", + "bbox": [ + 230, + 238, + 764, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ School of Engineering Medicine, Beihang University, 100191 Beijing, China", + "bbox": [ + 187, + 273, + 807, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "yongchuancui@gmail.com, liupeng202303@aircas.ac.cn, hui.zhang@buaa.edu.cn", + "bbox": [ + 171, + 295, + 818, + 309 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 329, + 326, + 344 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing deep learning-based models for remote sensing pansharpening exhibit exceptional performance on training datasets. However, due to sensor-specific characteristics and varying imaging conditions, these models suffer from substantial performance degradation when applied to unseen satellite data, lacking generalizability and thus limiting their applicability. We argue that the performance drops stem primarily from distributional discrepancies from different sources and the key to addressing this challenge lies in bridging the gap between training and testing distributions. To validate the idea and further achieve a \"train once, deploy forever\" capability, this paper introduces a novel and intuitive approach to empower any pansharpening models with generalizability by employing a unified distribution strategy (UniPAN). Specifically, we construct a distribution transformation function that normalizes the pixels sampled from different sources to conform to an identical distribution. The deep models are trained on the transformed domain, and during testing on new datasets, the new data are also transformed to match the training distribution. UniPAN aims to train and test the model on a unified and consistent distribution, thereby enhancing its generalizability. Extensive experiments validate the efficacy of UniPAN, demonstrating its potential to significantly enhance the performance of deep pansharpening models across diverse satellite sensors.", + "bbox": [ + 86, + 361, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Codes: https://github.com/yc-cui/UniPAN.", + "bbox": [ + 89, + 739, + 465, + 753 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 786, + 220, + 801 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Due to the inherent limitations of sensor imaging technology, advanced remote sensing satellites such as IKONOS, QuickBird, GaoFen-2, and WorldView-3, necessitate a compromise between the acquisition of extensive spectral infor", + "bbox": [ + 89, + 811, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/16bf4c0c34df7e504d3392c3eea04942c9c8ef0216e1716cf0f703efef000caf.jpg", + "image_caption": [ + "Figure 1. Comparison of generalization metrics w/ and w/o UniPAN trained on WorldView-3 (WV3) using FusionNet [19]." + ], + "image_footnote": [], + "bbox": [ + 517, + 335, + 635, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/940d3e0128c46e96e9a5ee907ff658059987a3ffb9243dadf7b3555aebc93195.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 647, + 335, + 764, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cfed76fb5cdad7a1cb0c3d41934aac9950234a70a1df91b45dc495f5dcc52f03.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 777, + 335, + 893, + 494 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/863f9367fcf43cf5ac9450f61a583ec0381744051b44eaeca8811c755dfcd24c.jpg", + "image_caption": [ + "Figure 2. Comparison of tSNE visualization w/ and w/o UniPAN." + ], + "image_footnote": [], + "bbox": [ + 514, + 553, + 895, + 705 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mation and maintaining fine spatial details in their captured imagery [20, 49, 68]. Consequently, these satellites typically capture either high spatial resolution single-band panchromatic (PAN) images or low spatial resolution multispectral (LRMS) images. Pansharpening is a crucial technique in remote sensing image processing, aiming to combine PAN and LRMS images to generate fused images with high spatial resolution and abundant spectra (HRMS).", + "bbox": [ + 511, + 746, + 906, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional pansharpening methods have been extensively studied over the past few decades. However, these", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 94, + 887, + 217, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11850", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "methods may introduce spectral distortions, require significant computational resources, and are sensitive to the quality of input data [20, 42, 49, 68]. In contrast, deep learning (DL)-based methods have shown great potential in pansharpening. Notably, various innovative architectures have emerged, starting from early convolutional models [19, 47, 82, 83], progressing to Transformer-based approaches [8, 21, 22, 24, 50, 62, 65], then generative methods [43, 46, 48, 58, 78], and continuing with the emergence of the latest novel designs [9, 23, 31, 53, 86]. The networks are evolving towards larger parameters capacity, stronger representation ability, and increasingly complex and diverse structures.", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, as deep pansharpening thrives, the overly powerful fitting capabilities of these networks have concurrently presented several ongoing concerns, e.g. generalizability, interpretability, over-reliance on high-quality data, etc. This paper specifically delves into the widely criticized issue of generalizability. Specifically, although deep models achieve excellent performance on training datasets, they often fail to generalize well to new data due to distribution shift such as different noise levels, varying spectral signatures and sensor conditioning characteristics. While models can achieve satisfactory performances on new datasets through retraining or fine-tuning, the process is time-consuming and fine-tuning on every new dataset is impractical and cumbersome. The ideal scenario would involve developing a pansharpening model that can be trained once and deployed effectively across diverse datasets and sensors, thereby providing a train once, deploy forever capability.", + "bbox": [ + 91, + 292, + 486, + 549 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, this paper introduces a novel approach that aims to empower any pansharpening models with generalizability by employing a unified distribution strategy (dubbed as UniPAN). The proposed method aims to bridge the gap between training and testing distributions, thereby improving the model's ability to generalize to unseen data. The core idea is to normalize the distribution of spatial pixels from the training dataset to conform to a predefined distribution (e.g. Gaussian distribution) and deep models are then trained on the transformed domain. This process is model-agnostic and requires no additional trainable parameters. When generalizing to new data, testing datasets are also transformed to match the training distribution, ensuring that the models are evaluated under the same distributional assumptions. Fig. 1 presents the cross-satellite generalization performance of a well-known pansharpening model FusionNet [19] on NBUPansharpRSData [49] trained w/ and w/o UniPAN. The model utilizing UniPAN demonstrates a significant improvement on full-resolution metrics, which validates the strong generalization ability for real-world scenes. To further elucidate the efficacy of UniPAN, Fig. 2 presents tSNE [64] dimensionality reduction clustering of multispectral images from diverse satellites using NBUPansharpRSData [49] w/", + "bbox": [ + 91, + 553, + 486, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "and w/o UniPAN. The pre-UniPAN visualization reveals distinct clustering patterns, with data points from different satellites forming separate groups, indicative of substantial inter-satellite distributional disparities. In contrast, following the application of UniPAN, a notable transformation is observed: the previously separated clusters become substantially intermingled, demonstrating a significant reduction in distributional discrepancies across different satellite datasets (e.g. the purple scatters of QuickBird). This visual evidence underscores UniPAN's capability to harmonize heterogeneous satellite data, effectively creating a more cohesive and unified representation.", + "bbox": [ + 511, + 90, + 908, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The contributions of this paper could be summarized as follows:", + "bbox": [ + 511, + 273, + 906, + 301 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose UniPAN, a model-agnostic approach that employs a unified distribution strategy to empower any deep pansharpening models with generalizability, achieving a train once, deploy forever capability. This strategy provides two advantages: 1) Architectural independence: Compatible with any existing or future network architecture without structural modifications. 2) Training efficiency: Eliminates the need for retraining or fine-tuning when adapting to new sensors.", + "- UniPAN is implemented via inverse transform sampling, and we prove its equivalence to the optimal solution of optimal transport in the one-dimensional case. UniPAN requires no additional trainable parameters and only needs to be fitted once before training begins, introducing almost no additional computational overhead. It serves as a viable alternative to vanilla data normalization techniques such as min-max and $z$ -score normalization, offering a more robust and unified approach to data distribution alignment.", + "- We conduct extensive experiments using a variety of deep pansharpening models across numerous satellites. The simple yet intuitive and effective strategy provides a universal solution to the generalization challenge in DL-based pansharpening and experimental results demonstrate its effectiveness, showcasing significant improvements in cross-dataset generalization across a wide range of models and satellite sensors." + ], + "bbox": [ + 513, + 306, + 908, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 715, + 656, + 732 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Pansharpening Methods", + "text_level": 1, + "bbox": [ + 511, + 741, + 738, + 758 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional Methods. Traditional pansharpening methods can be broadly categorized into three paradigms: component substitution (CS), multi-resolution analysis (MRA), and variational optimization (VO) [20, 42, 49, 68]. CS-based methods [4, 10, 27, 38, 55, 61] operate by substituting the spatial component of the LRMS image with the PAN image. Representative algorithms include Gram-Schmidt (GS) [4, 38], principal component analysis (PCA) [10, 61], and intensity-hue-saturation (IHS) [27, 55]. While computationally effi", + "bbox": [ + 511, + 763, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11851", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "cient, CS methods often introduce spectral distortions due to mismatches between PAN and substituted components. MRA-based methods [1-3, 13, 26, 41] employ multi-scale decomposition tools (e.g. wavelets [26], contourlets [13]) to inject high-frequency details from PAN into LRMS. Typical implementations include MTF-GLP [1-3] and SFIM [41]. These methods generally preserve spectral fidelity better than CS approaches but may exhibit limited spatial enhancement. VO-based approaches [7, 52, 63, 67, 74, 75] formulate pansharpening as an optimization problem with handcrafted priors, such as the variational wavelet-based methods [7] and Bayesian approaches [52]. Despite their mathematical elegance, traditional methods often require manual parameter tuning and struggle with complex spatial-spectral relationships [20, 68].", + "bbox": [ + 89, + 90, + 483, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Deep Learning-Based Methods. The advent of deep learning (DL) have revolutionized pansharpening by enabling end-to-end learning of spatial-spectral relationships. Early pioneering convolutional neural networks (CNNs) such as PNN [47], PanNet [82], and MSDCNN [83] established the feasibility of learning fusion rules from a mount of simulated data. Subsequent innovations introduced multi-scale architectures [33, 72, 76], attention mechanisms [39, 73], and kernel customizing [12, 18, 25, 34, 85], etc. The CNN-based methods demonstrated superior performance over traditional approaches but faced limitations in modeling long-range dependencies. In contrast, Transformer [24, 45, 65] addressed this limitation with self-attention mechanisms. Although various Transformer architectures specifically designed for pansharpening [8, 21, 22, 50, 62] have achieved impressive performance, they face criticism for quadratic computational complexity w.r.t. sequence length [16]. To address this limitation, recent studies have explored linear-complexity models [16, 31, 53, 86] that maintain performance while improving efficiency. To further enhance fidelity preservation, generative models like generative adversarial networks (GANs) [17, 43, 44, 46, 79] and diffusion models [48, 58, 77, 78, 88] have emerged as promising solutions. Beyond conventional configurations, contemporary research has expanded into more challenging frontiers to address practical constraints, including zero-shot learning paradigms [9, 23, 57, 71], frequency domain integration [30, 40, 84, 89, 90], and interpretable fusion mechanisms [32, 70, 80, 81].", + "bbox": [ + 91, + 319, + 486, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Generalization in DL-Based Pansharpening", + "text_level": 1, + "bbox": [ + 89, + 771, + 464, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite remarkable progress in DL-based pansharpening, generalization capability remains a persistent challenge. The issue manifests conspicuously in two critical scenarios: 1) Models trained under the Wald protocol [69] using spatially degraded data often fail to maintain performance when applied to full-resolution inference, frequently exhibiting severe spectral distortions. 2) Networks optimized for specific", + "bbox": [ + 89, + 794, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "satellite sensors demonstrate substantial performance degradation when deployed on unseen satellite platforms due to distribution shifts in spectral signatures and sensor characteristics. Existing approaches to address these challenges primarily focus on unsupervised learning [14, 15, 46], generalized module design [77], or advanced data augmentation [11]. However, such methods typically require either meticulous architectural redesign or computationally intensive retraining/fine-tuning procedures. These requirements pose significant practical limitations for real-world deployment.", + "bbox": [ + 511, + 90, + 906, + 256 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In contrast to model-centric solutions, our approach adopts a fundamentally different perspective by addressing the generalization challenge through data-space intervention. The proposed UniPAN framework introduces a model-agnostic preprocessing technique that operates on raw sensor data. Compared to traditional linear data normalization such as min-max scaling or $z$ -score standardization, our method employs a non-linear mapping to project the input data onto a unified distribution, effectively harmonizing the distribution mismatches from all sensors.", + "bbox": [ + 511, + 257, + 908, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 419, + 604, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Distribution Transformation", + "text_level": 1, + "bbox": [ + 511, + 443, + 769, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The core challenge in cross-domain pansharpening stems from the distribution discrepancy between source (training) and target (testing) domains. Let $\\mathcal{X}_S$ denote the source domain (training data) and $\\mathcal{X}_T$ the target domain (testing data). Our objective is to learn a transformation $\\mathcal{T}:\\mathcal{X}\\to \\mathcal{Z}$ that maps both domains to a shared space $\\mathcal{Z}$ with unified distribution $\\rho_t$ , e.g. standard normal distribution.", + "bbox": [ + 511, + 465, + 908, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Inverse Transform Sampling. Given a univariate r.v. $X \\sim \\rho_s$ , we seek a transformation $\\mathcal{T} : \\mathbb{R} \\to \\mathbb{R}$ such that $Z = \\mathcal{T}(X) \\sim \\rho_t$ . The inverse transform sampling theorem provides a theoretically grounded solution:", + "bbox": [ + 511, + 571, + 908, + 633 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 636, + 906, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $F_{S}(x) = \\rho_{s}(-\\infty, x]$ is the cumulative distribution function (CDF) of $X$ , and $F_{T}^{-1}(u) = \\inf \\{z \\in \\mathbb{R} : F_{T}(z) \\geq u\\}$ is the inverse CDF of the target distribution. By the universality of the uniform, $U = F_{S}(X)$ follows a uniform distribution, and $Z = F_{T}^{-1}(U)$ inherits $\\rho_{t}$ exactly when $F_{S}$ is continuous. To provide more intuitively analyze of the algorithm's operational process and efficiency, we establish a fundamental connection between inverse transform sampling and the well-studied optimal transport theory [51, 54].", + "bbox": [ + 511, + 660, + 908, + 796 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Optimal Transport Perspective. Given two probability distributions $\\rho_{s}$ (source domain) and $\\rho_{t}$ (target domain) defined on $\\mathbb{R}^n$ and $\\mathbb{R}^m$ respectively, we consider the Monge problem [51, 54]:", + "bbox": [ + 511, + 796, + 908, + 856 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} (\\rho_ {s}, \\rho_ {t}) \\triangleq \\min _ { \\begin{array}{c} T: \\mathbb {R} ^ {n} \\to \\mathbb {R} ^ {m} \\\\ T _ {\\sharp} \\rho_ {s} = \\rho_ {t} \\end{array} } \\int_ {\\mathbb {R} ^ {n}} c (x, T (x)) \\rho_ {s} (x) d x, \\quad (2)\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 861, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11852", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c:\\Omega_s\\times \\Omega_t\\to \\mathbb{R}^+$ is a cost function and for any measurable Borel set $E\\in \\Omega_s$ $T:\\Omega_s\\rightarrow \\Omega_t$ transfers measures from one space $\\Omega_{s}$ to another space $\\Omega_t$", + "bbox": [ + 89, + 90, + 483, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT _ {\\sharp} \\rho_ {s} (E) = \\rho_ {s} \\left(T ^ {- 1} (E)\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 148, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "here $T_{\\sharp}$ is the pushforward operator. The Monge problem seeks a transport plan $T$ from $\\rho_s$ to $\\rho_t$ that minimizes the transportation cost. The optimal $T^{*}$ is also known as the Monge map of Eq. (2).", + "bbox": [ + 89, + 179, + 483, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $F_{S}(x) = \\rho_{s}(-\\infty, x]$ and $F_{T}(z) = \\rho_{t}(-\\infty, z]$ denote the CDFs of the source and target distributions respectively. In the univariate setting ( $n = m = 1, \\Omega_{s} = \\Omega_{t} = \\mathbb{R}$ ), when the source measure $\\rho_{s}$ is continuous and the cost function takes the form $c(x, z) = h(|x - z|)$ with $h$ being strictly convex, by cyclical monotonicity [66], the Monge map $T^{*}$ satisfies:", + "bbox": [ + 89, + 241, + 485, + 344 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nF _ {S} (x) = F _ {T} \\left(T ^ {*} (x)\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 348, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "yielding the explicit expression:", + "bbox": [ + 89, + 375, + 303, + 388 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nT ^ {*} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 400, + 483, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This remarkable result coincides exactly with the inverse transform sampling formula Eq. (1), i.e. $T^{*} = \\mathcal{T}$ . The transport cost in this setting [37, 56, 59] can be computed as", + "bbox": [ + 89, + 431, + 485, + 478 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {C} \\left(\\rho_ {s}, \\rho_ {t}\\right) = \\int_ {0} ^ {1} c \\left(F _ {S} ^ {- 1} (u), F _ {T} ^ {- 1} (u)\\right) d u. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 489, + 483, + 525 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For the $L_{1}$ cost function $c(x,z) = |x - z|$ , this corresponds to the $W_{1}$ Wasserstein distance [36, 59] between the distributions. In the one-dimensional (1D) scenario, UniPAN operates independently on spatial pixels within each spectral band. This implementation can be achieved through interpolation-based inverse transform sampling without framing it as an optimization problem. However, given the inherent correlations between spectral bands and interdependencies among spatial patches, multidimensional extension of this framework may be promising, albeit with increased implementation complexity. The theoretical derivation establishing connections with optimal transport theory serves to facilitate future generalization of UniPAN to multidimensional configurations. Building upon the theoretical foundation of distribution transformation through optimal transport, in the following, we present a practical implementation of 1D UniPAN.", + "bbox": [ + 89, + 532, + 485, + 789 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. UniPAN Implementation", + "text_level": 1, + "bbox": [ + 89, + 801, + 316, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We present a practical implementation for 1D finite discrete distributions of UniPAN through a quantile-based distribution transformation algorithm.", + "bbox": [ + 89, + 824, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finite Discrete Distribution Transformation. Given pixel values $X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}$ sampled from a spectral band,", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 UniPAN Training and Testing Procedure" + ], + "code_body": "Training Phase \nInput: Training dataset $\\mathcal{D}_{\\mathrm{train}} = \\{L_i, P_i, H_i\\}_{i=1}^N$ , inverse CDF $F_T^{-1}$ of target distribution $\\rho_t$ , model $\\mathcal{M}$ \nOutput: Trained model $M$ \n1: Initialize transformed LRMS $\\tilde{L}$ and PAN $\\tilde{P}$ \n2: for each band $b$ of LRMS $\\{L_i\\}_{i=1}^N$ & PAN $\\{P_i\\}_{i=1}^N$ do \n3: Sample $n$ pixels $X_b \\gets \\{x_1, \\ldots, x_n\\}$ from band $b$ \n4: Compute $m$ quantiles $Q = \\{q_0, \\ldots, q_{m-1}\\} \\triangleright$ Eq. (7) \n5: Linspace uniform CDF $U = \\{0, \\frac{1}{m-1}, \\ldots, 1\\}$ \n6: for each pixel $x$ in band $b$ do \n7: Forward and backward interp by $u_\\uparrow$ and $u_\\downarrow \\triangleright$ Eq. (9) \n8: Compute eCDF value $u(x) \\gets (u_\\uparrow + u_\\downarrow) / 2 \\triangleright$ Eq. (8) \n9: Transform $x$ to $z \\gets F_T^{-1}(u(x))$ and assign $z$ to corresponding band of $\\tilde{L}$ and $\\tilde{P}$ \n10: end for \n11: end for \n12: while not converged do \n13: Train $\\mathcal{M}$ with $\\mathcal{L} = \\|H - \\mathcal{M}(\\tilde{L}, \\tilde{P}) + L \\uparrow\\|_1$ \n14: end while \n15: return Trained model $\\mathcal{M}$", + "bbox": [ + 513, + 109, + 903, + 390 + ], + "page_idx": 3 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Testing Phase" + ], + "code_body": "Input: Testing dataset $\\mathcal{D}_{\\mathrm{test}} = \\{L_i,P_i\\}_{i = 1}^N$ , inverse CDF $F_{T}^{-1}$ of target distribution $\\rho_{t}$ , trained model $\\mathcal{M}$ \nOutput: Pansharpening result $\\hat{H}^{\\mathrm{test}}$ \n16: for each band $b$ of LRMS & PAN do \n17: Transform $L,P$ to $\\tilde{L},\\tilde{P}$ using the training transformation \n18: end for \n19: Generate $\\hat{H}^{\\mathrm{test}}\\gets \\mathcal{M}(\\tilde{L},\\tilde{P}) + L\\uparrow$ \n20: return Fused result $\\hat{H}^{\\mathrm{test}}$", + "bbox": [ + 514, + 409, + 905, + 523 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we first compute $m$ empirical quantiles $Q = \\{q_{0},\\dots,q_{m - 1}\\}$ where:", + "bbox": [ + 511, + 549, + 905, + 578 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq _ {i} = \\inf \\left\\{x \\in X \\mid P (X \\leq x) \\geq \\frac {i}{m - 1} \\right\\}, i = \\{0, 1, \\dots , m - 1 \\}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 511, + 585, + 906, + 628 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These quantiles discretize the empirical cumulative distribution function (eCDF) of the source data. The number of quantiles $m$ controls the approximation fidelity.", + "bbox": [ + 511, + 628, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given any data point $x$ to be transformed, the next step involves determining its value in the eCDF through interpolation. However, when the source distribution contains repeated values or sparse regions, quantile estimation may yield non-unique or overlapping quantile boundaries. This can result in abrupt jumps in the eCDF, thereby violating the properties of monotonicity and smoothness. Therefore, we use bidirectional interpolation to estimate a robust eCDF value of $x$ :", + "bbox": [ + 511, + 674, + 908, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} u (x) = u _ {\\uparrow} (x) / 2 + u _ {\\downarrow} (x) / 2 \\\\ = \\operatorname {i n t e r p} (x, Q, U) / 2 + \\tag {8} \\\\ (1 - \\text {i n t e r p} (- x, - Q ^ {\\text {r e v}}, - U ^ {\\text {r e v}})) / 2, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 816, + 906, + 872 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $U = \\{0,\\frac{1}{m - 1},\\dots,1\\}$ , $Q^{\\mathrm{rev}}$ is the reversed quantiles.", + "bbox": [ + 511, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11853", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/28d2467620b790e9cf13a464cb50438888716a5e9fd946cba3d998ef02912e8e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
QNRTρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.980(+0.005)0.946(-0.010)0.971(+0.000)0.963(-0.003)0.964(-0.004)0.974(-0.000)0.985(+0.010)0.961(+0.004)0.979(+0.007)0.984(+0.017)0.982(+0.013)0.987(+0.013)
IK0.950(+0.005)0.912(+0.039)0.915(+0.045)0.924(+0.011)0.916(+0.022)0.958(+0.011)0.956(+0.011)0.919(+0.045)0.940(+0.070)0.950(+0.037)0.945(+0.051)0.968(+0.021)
QB0.957(+0.006)0.936(+0.016)0.931(+0.047)0.953(+0.017)0.944(+0.014)0.970(+0.016)0.965(+0.014)0.944(+0.024)0.966(+0.082)0.969(+0.032)0.969(+0.038)0.970(+0.017)
WV20.880(+0.026)0.811(+0.029)0.832(+0.062)0.946(+0.008)0.916(+0.012)0.951(+0.012)0.872(+0.018)0.775(-0.006)0.825(+0.055)0.944(+0.006)0.914(+0.011)0.940(+0.001)
WV30.862(-0.006)0.804(+0.035)0.795(+0.067)0.921(-0.011)0.919(+0.011)0.950(+0.001)0.880(+0.011)0.805(+0.036)0.813(+0.084)0.943(+0.010)0.924(+0.016)0.952(+0.003)
WV40.924(+0.009)0.898(+0.045)0.875(+0.036)0.948(-0.000)0.929(+0.003)0.972(+0.006)0.939(+0.024)0.889(+0.035)0.903(+0.064)0.952(+0.002)0.935(+0.009)0.973(+0.007)
Dλ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.007(-0.003)0.022(+0.008)0.005(-0.000)0.017(+0.004)0.015(+0.005)0.011(+0.000)0.004(-0.005)0.012(-0.001)0.004(-0.001)0.004(-0.008)0.004(-0.005)0.004(-0.005)
IK0.011(-0.000)0.020(-0.015)0.018(-0.011)0.031(+0.004)0.022(-0.003)0.014(-0.003)0.008(-0.003)0.020(-0.015)0.013(-0.016)0.013(-0.014)0.009(-0.016)0.010(-0.008)
QB0.013(-0.000)0.011(-0.004)0.015(-0.010)0.016(-0.004)0.014(+0.000)0.010(-0.005)0.009(-0.003)0.009(-0.006)0.007(-0.017)0.008(-0.013)0.006(-0.007)0.012(-0.003)
WV20.042(-0.006)0.066(-0.007)0.037(-0.019)0.008(-0.000)0.014(-0.000)0.016(-0.002)0.046(-0.002)0.086(+0.012)0.048(-0.008)0.008(+0.000)0.015(+0.001)0.022(+0.002)
WV30.050(+0.010)0.067(-0.011)0.049(-0.024)0.027(+0.011)0.014(+0.000)0.016(+0.002)0.042(+0.002)0.072(-0.006)0.051(-0.022)0.013(-0.002)0.011(-0.002)0.014(+0.000)
WV40.024(-0.000)0.031(-0.012)0.029(-0.009)0.017(+0.008)0.018(+0.008)0.005(-0.001)0.016(-0.008)0.032(-0.011)0.022(-0.016)0.014(+0.004)0.015(+0.005)0.004(-0.002)
DS↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.012(-0.003)0.032(+0.003)0.023(+0.000)0.020(+0.000)0.021(+0.000)0.015(-0.000)0.010(-0.005)0.026(-0.002)0.016(-0.006)0.010(-0.009)0.013(-0.007)0.007(-0.008)
IK0.039(-0.005)0.068(-0.026)0.068(-0.036)0.046(-0.014)0.063(-0.019)0.027(-0.007)0.035(-0.008)0.062(-0.033)0.046(-0.058)0.037(-0.024)0.045(-0.036)0.022(-0.013)
QB0.030(-0.005)0.053(-0.012)0.054(-0.039)0.031(-0.012)0.042(-0.014)0.020(-0.012)0.025(-0.011)0.046(-0.018)0.026(-0.067)0.023(-0.020)0.024(-0.031)0.020(-0.012)
WV20.081(-0.022)0.134(-0.024)0.137(-0.049)0.045(-0.008)0.071(-0.012)0.032(-0.010)0.085(-0.018)0.153(-0.005)0.133(-0.053)0.047(-0.006)0.071(-0.012)0.039(-0.003)
WV30.096(-0.000)0.141(-0.027)0.167(-0.050)0.054(+0.002)0.067(-0.011)0.033(-0.004)0.083(-0.013)0.134(-0.033)0.146(-0.072)0.043(-0.008)0.064(-0.014)0.033(-0.003)
WV40.053(-0.009)0.073(-0.036)0.099(-0.029)0.035(-0.006)0.053(-0.011)0.021(-0.005)0.045(-0.018)0.082(-0.027)0.076(-0.052)0.036(-0.004)0.052(-0.012)0.021(-0.005)
Dρ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.043(+0.003)0.110(+0.030)0.026(-0.008)0.079(+0.032)0.062(+0.015)0.041(-0.002)0.019(-0.020)0.059(-0.021)0.022(-0.013)0.015(-0.032)0.017(-0.029)0.017(-0.026)
IK0.082(-0.006)0.095(-0.038)0.090(-0.058)0.126(+0.024)0.110(-0.010)0.061(-0.008)0.064(-0.024)0.092(-0.040)0.066(-0.083)0.051(-0.049)0.061(-0.059)0.045(-0.024)
QB0.061(-0.003)0.154(-0.020)0.179(-0.059)0.135(+0.011)-0.133(-0.045)-0.093(-0.026)-0.164(-0.029)-0.235(-0.002)-0.200(-0.058)-0.100(-0.013)-0.126(-0.048)-0.112(-0.007)
WV2-1.166(-0.27)-1.198(-0.339)-1.199(-0.539)-1.155(-0.459)-1.133(-0.455)-1.137(-0.451)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)
WV3-2.205(+0.18)-2.214(-0.41)-2.234(-0.533)-2.231(-0.419)-2.173(+0.11)-2.173(-0.415)-2.176(-0.419)-2.226(-0.429)-2.219(-0.468)-2.117(-0.434)-1.131(-0.439)-1.131(-0.439)
WV4-3.155(-0.55)-3.112(-0.42)-3.115(-0.44)-3.146(+0.88)-3.169(+0.77)-3.169(-0.477)-3.173(-0.477)-3.176(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)
", + "bbox": [ + 94, + 78, + 906, + 410 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Average results of 12 models. Rows correspond to training satellites, and columns correspond to testing satellites. Green indicates performance improvement, while red indicates performance degradation. The values in parentheses represent the average delta compared to the baseline without UniPAN.", + "bbox": [ + 89, + 420, + 906, + 462 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The function $\\text{interp}$ represents a linear interpolation between the quantile point $Q$ and the reference point $U$ , with the interpolation formula given by:", + "bbox": [ + 89, + 474, + 485, + 521 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {i n t e r p} (x, Q, U) = \\left\\{ \\begin{array}{l l} U [ 0 ], & x \\leq Q [ 0 ] \\\\ f (x, i, Q, U), & Q [ i ] < x \\leq Q [ i + 1 ] \\\\ U [ m - 1 ], & x > Q [ m - 1 ] \\end{array} \\right. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 527, + 493, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f$ computes the $Q$ to $U$ mapping,", + "bbox": [ + 89, + 595, + 354, + 611 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (x, i, Q, U) = U [ i ] + \\frac {x - Q [ i ]}{Q [ i + 1 ] - Q [ i ]} (U [ i + 1 ] - U [ i ]) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 616, + 483, + 661 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The last step is to apply inverse CDF of target distribution,", + "bbox": [ + 109, + 662, + 483, + 678 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} (x) = F _ {T} ^ {- 1} (u (x)) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 690, + 483, + 709 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and Testing Procedure. The training and testing procedures of UniPAN are described in Algorithm 1. Here, $L \\uparrow$ denotes the upsampled LRMS image using bicubic interpolation to match the size of the PAN image. The model $\\mathcal{M}$ is trained to predict the residual of LRMS with an $l_{1}$ loss function that minimizes the discrepancy between the predicted HRMS image and the ground truth.", + "bbox": [ + 89, + 710, + 483, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As for the target distribution $\\rho_{t}$ , we test two widely used distributions, i.e. standard normal distribution $\\mathcal{N}(0,1)$ and standard uniform distribution $\\mathcal{U}(0,1)$ , the transformation $\\mathcal{T}(x)$ becomes:", + "bbox": [ + 89, + 816, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} _ {\\mathrm {n}} (x) = \\Phi^ {- 1} \\left(F _ {S} (x)\\right), \\quad \\mathcal {T} _ {\\mathrm {u}} (x) = F _ {S} (x), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 883, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\Phi^{-1}$ denotes the inverse CDF of standard normal distribution. Since the inverse CDF of standard uniform distribution is an identical mapping, $\\mathcal{T}_{\\mathrm{u}}$ becomes the estimated empirical distribution itself. The standard normal distribution transformation non-linearly warps the input space to match Gaussian statistics, while the uniform transformation directly maps pixel intensities to their empirical quantiles.", + "bbox": [ + 511, + 474, + 906, + 582 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 513, + 595, + 637, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Setups", + "text_level": 1, + "bbox": [ + 513, + 619, + 601, + 636 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We conducted experiments utilizing the Red, Green, Blue, and NIR bands of the NBUPansharpRSDataset dataset [49]. The variations in resolution among different satellites, coupled with significant differences in land covers and land uses, render this dataset particularly suitable for assessing the generalization capabilities of pansharpening models. The dataset comprises data from 6 satellites. For each satellite, we partitioned the data into training, validation, and testing sets in a ratio of $7:1:2$ .", + "bbox": [ + 511, + 643, + 906, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. To validate the effectiveness and generalization capability of UniPAN, we conducted comparative experiments between baseline (w/o UniPAN) and enhanced (w/ UniPAN) configurations using diverse commonly-recognized models, including FusionNet [19], PanNet [82], MSDDN [28, 29], GPPNN [80], MSDCNN [83], FeINFN [40], PNN [47], LAGConv [34], SFIIN [89, 90], PreMix [18], UAPN [87], MDCUN [81].", + "bbox": [ + 511, + 779, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11854", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3d5e7231ab00420a762b9cb2b8063163b09d86c7345ce42bb2458470f4ae8ebf.jpg", + "image_caption": [ + "Figure 3. Generalization testing on GaoFen-1, the networks are trained on WorldView-3. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the $D_{\\lambda} / D_{S}$ value. In most cases, the use of UniPAN yields superior spectral and spatial preservation. For instance, MDCUN [81] exhibits a highly close alignment with the LRMS in spectral characteristics with UniPAN of $\\rho_t = \\mathcal{N}(0,1)$ ." + ], + "image_footnote": [], + "bbox": [ + 109, + 80, + 888, + 462 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "All implementations followed identical experimental procedures: a batch size of 8 (except for MDCUN [81] and FeINFN [40] which used 4), the Adam [35] optimizer $(\\beta_{1} = 0.9, \\beta_{2} = 0.999)$ with an initial learning rate of $10^{-3}$ , and StepLR scheduling which reduces the learning rate by a factor of 0.8 every 100 epochs. Training was performed on an NVIDIA GeForce RTX 4090 GPU for 600 epochs using the Wald protocol [69] with $l_{1}$ loss, with best model selected based on optimal validation performance. Note that, to prevent data leakage of testing set, the transformation fitting process is conducted on the training set of each satellite.", + "bbox": [ + 88, + 546, + 485, + 715 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. The full-reference metrics based on spatial degradation (e.g. PSNR, SSIM) can only reflect performance in simulated reduced-resolution experiments and fail to demonstrate effectiveness in real-world full-resolution scenarios [6, 16, 46]. In contrast, no-reference metrics like QNR explicitly quantify model performance in both spectral preservation and spatial enhancement, enabling direct evaluation of fusion results at native resolution while avoiding theoretical biases introduced by downsampling. To better align with practical applications where reference images are unavailable, this paper adopts widely-used no-reference spectral and spatial metrics, i.e. QNR [5], $D_{\\lambda}$ [5], $D_{S}$ [5],", + "bbox": [ + 88, + 719, + 486, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and $D_{\\rho}$ [60], to comprehensively assess cross-resolution and cross-satellite generalization capabilities.", + "bbox": [ + 511, + 546, + 906, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.Quantitative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 589, + 735, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 presents the average generalization performance of various models enhanced with UniPAN under different target distributions. Green indicate metric improvements, while red denote performance degradation. Slight yet observable differences exist between the two target distributions. Overall, the uniform distribution achieves better performance than the normal distribution. The normal distribution demonstrated a modest decline in performance across the WV2 and WV3 datasets. This phenomenon may stem from the peaked characteristics of the normal distribution, which could hinder model training, whereas the uniform distribution provides smoother normalization (see Fig. 5 in Sec. 4.3).", + "bbox": [ + 509, + 612, + 908, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Notably, UniPAN with $\\rho_{t} = \\mathcal{U}(0,1)$ achieves significant improvements across most metrics, with only marginal declines in rare cases. For instance, when generalizing from WV3 to QB, UniPAN improves the QNR metric by over 0.084 (from 0.729 to 0.813) averaged on all 12 models, underscoring its effectiveness. Beyond QNR, UniPAN also enhances spectral preservation $(D_{\\lambda})$ and spatial consistency", + "bbox": [ + 509, + 795, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11855", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7094b3dd5c36ccab6dfc0ce1cfc6fec32a0db01927e72b11d194d9ccb4062908.jpg", + "image_caption": [ + "Figure 4. Full-resolution generalization testing on QuickBird. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the $D_{\\lambda} / D_{S}$ value." + ], + "image_footnote": [], + "bbox": [ + 109, + 80, + 888, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$(D_S, D_\\rho)$ , further validating the efficacy of the unified distribution strategy. These results confirm that aligning heterogeneous satellite data to a unified distribution effectively mitigates domain shifts, enabling robust cross-sensor generalization. However, while UniPAN improves spatial fidelity greatly, it may cause slight spectral distortion, as evidenced by the minor performance degradation observed in the $D_{\\lambda}$ metric. This is reasonable since the transformation applied to each spectral band maintains the relative spatial structure but might result in minimal loss of spectral information. Nevertheless, such information loss is extremely limited, as the overall enhancement in performance remains predominant.", + "bbox": [ + 88, + 518, + 485, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Visual Comparison", + "text_level": 1, + "bbox": [ + 89, + 712, + 274, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Fusion Results. To illustrate the practical impact of UniPAN, Figs. 3 and 4 compare fusion results w/ and w/o UniPAN. In most cases, UniPAN achieves highly favorable results in both spectral and spatial preservation. Furthermore, the remarkably high QNR values demonstrate its superior capability in balancing spectral and spatial fidelity, e.g. in Fig. 3, MDCUN [81] applying UniPAN with normal distribution enhanced the QNR by 0.1514 (from 0.7157 to 0.8671) and aligned fused results closely with the ground truth spectra. In the full-resolution generalization test depicted in Fig. 4, even with minimal visual differences,", + "bbox": [ + 89, + 734, + 485, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UniPAN achieves metrically significant improvements in spectral and spatial preservation. This capability is critical for downstream applications requiring accurate spectral information. Additional visual comparison results are included in the supplementary materials.", + "bbox": [ + 511, + 518, + 906, + 594 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison of Data Distribution. Fig. 5 further analyzes data distribution alignment. The kernel density estimation (KDE) curves (top right area and diagonals) reveal that UniPAN successfully transforms the original complex and heterogeneous distributions into a unified target distribution. The quantile-quantile (Q-Q) plots (bottom left area) demonstrate harmonized distributions across satellites after applying UniPAN (The greater the deviation from the $y = x$ line, the more significant the disparity between the two distributions). For example, the original Q-Q plots between WV4 and QB show significant dispersion, whereas post-UniPAN points close tightly around $y = x$ . This visual evidence corroborates UniPAN's ability to bridge distribution gaps.", + "bbox": [ + 511, + 595, + 908, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Parameter Analysis", + "text_level": 1, + "bbox": [ + 511, + 801, + 702, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "UniPAN introduces two hyperparameters: the number of sampled pixels $n$ and the number of quantiles $m$ . Figs. 6 and 7 investigate their impact on performance of FusionNet [19] trained on WV2 and tested on GF1. As $n$ and $m$ vary, the metrics stabilize. Generally, $\\mathcal{U}(0,1)$ demonstrates", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11856", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8c87594f3788d04fa0663c5f741b7a96fcc3f177ea1bdfaf0fd4bddbde54473a.jpg", + "image_caption": [ + "(a) w/o UniPAN" + ], + "image_footnote": [], + "bbox": [ + 89, + 89, + 354, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2d479758e8371c95ac4f981afd8545de0e50873d93bc8c37d60e998c1b472acf.jpg", + "image_caption": [ + "(b) w/ UniPAN, $\\rho_t = \\mathcal{N}(0,1)$" + ], + "image_footnote": [], + "bbox": [ + 359, + 89, + 625, + 291 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/524d7929265bed16ce9a843829177ce385f263449f98df7282245d38ecc97d0f.jpg", + "image_caption": [ + "(c) $\\mathrm{w} / \\mathrm{UniPAN}$ $\\rho_t = \\mathcal{U}(0,1)$" + ], + "image_footnote": [], + "bbox": [ + 638, + 90, + 898, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/956fc6ef24fd91914e2bd1eb223b55b7d72695991efa19372dd43972da283cb7.jpg", + "image_caption": [ + "Figure 5. Comparison of data distributions before and after applying UniPAN. The upper-right and diagonals of each subplot display the kernel density estimation (KDE) curves of Green band for each satellite. The scatter plots in the lower-left areas of each subplots represent quantile-quantile (Q-Q) plots of green band, with the dashed line indicating $y = x$ . The greater the deviation from the $y = x$ line, the more significant the disparity between the two distributions. Zoom in for better view.", + "Figure 6. Parameter analysis of number of quantiles, with number of sampled pixels fixed to 10000." + ], + "image_footnote": [], + "bbox": [ + 84, + 376, + 472, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "superior performance compared to $\\mathcal{N}(0,1)$ . Both configurations achieve statistically significant improvements over baseline methods without applying UniPAN. This parametric robustness demonstrates that optimal performance can be attained without exhaustive hyperparameter tuning, significantly reducing computational overhead.", + "bbox": [ + 89, + 659, + 486, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 89, + 767, + 339, + 784 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We proposed UniPAN, a unified distribution strategy addressing generalization challenges in deep learning-based pansharpening. UniPAN eliminates spectral discrepancies by normalizing multi-source remote sensing imagery into predefined target distributions, effectively bridging the gap between training and testing distributions. Extensive quantitative and visual comparisons across 6 satellite sensors and", + "bbox": [ + 89, + 794, + 485, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a8b78f92931d5d2121022107a706625190cffd080196318790c7203579d9b664.jpg", + "image_caption": [ + "Figure 7. Parameter analysis of number of sampled pixels, with number of quantiles fixed to 1000." + ], + "image_footnote": [], + "bbox": [ + 509, + 376, + 890, + 611 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "12 models demonstrate that UniPAN significantly enhances cross-sensor generalization. Notably, this enhancement is achieved without introducing any additional trainable parameters or the need for retraining or fine-tuning the models.", + "bbox": [ + 511, + 657, + 906, + 718 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While this work aligns theoretically with 1D optimal transport, considering interdependencies of spectral bands or spatial patches, future research could explore extending the method to the multi-dimensional case with optimal transport theories. Additionally, the experimental results indicate that the choice of target distribution has a notable impact on performance, thus it is worthwhile investigating adaptive target domain selection. To further validate the generalizability of the proposed unified distribution strategy, the ideas presented in this work could be extended to other remote sensing image processing tasks, such as multi-modal fusion and hyperspectral pansharpening, etc.", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11857", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment. This research was partially supported by the National Natural Science Foundation of China under Grant 32441114, 41971397, 61731022, and the project Y1H103101A.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 165, + 187, + 181 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] B. Aiazzi, L. Alparone, S. Baronti, and A. Garzelli. Context-driven fusion of high spatial and spectral resolution images based on oversampled multiresolution analysis. IEEE TGRS, 40(10):2300-2312, 2002. 3", + "[2] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva. An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas. In 2003 2nd GRSS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, pages 90-94, 2003.", + "[3] Bruno Aiazzi, L. Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. MTF-tailored multiscale fusion of high-resolution MS and Pan imagery. *Photogramm. Eng. Remote Sens.*, 72:591-596, 2006. 3", + "[4] Bruno Aiazzi, Stefano Baronti, and Massimo Selva. Improving component substitution pansharpening through multivariate regression of MS+Pan data. IEEE TGRS, 45(10): 3230-3239, 2007. 2", + "[5] Luciano Alparone, Bruno Aiazzi, Stefano Baronti, Andrea Garzelli, Filippo Nencini, and Massimo Selva. Multispectral and panchromatic data fusion assessment without reference. *Photogramm. Eng. Remote Sens.*, 74(2):193-200, 2008. 6", + "[6] Alberto Arienzo, Gemine Vivone, Andrea Garzelli, Luciano Alparone, and Jocelyn Chanussot. Full-resolution quality assessment of pansharpening: Theoretical and hands-on approaches. IEEE GRSM, 10(3):168-201, 2022. 6", + "[7] Coloma Ballester, Vicent Caselles, and Laura Igual. A variational model for $\\mathrm{P + XS}$ image fusion. IJCV, 69(1):43-58, 2006. 3", + "[8] Wele Gedara Chaminda Bandara and Vishal M Patel. Hypertransformer: A textural and spectral feature fusion Transformer for pansharpening. In CVPR, pages 1767-1777, 2022, 2, 3", + "[9] Qi Cao, Liang-Jian Deng, Wu Wang, Junming Hou, and Gemine Vivone. Zero-shot semi-supervised learning for pansharpening. Inf. Fusion., 101:102001, 2024. 2, 3", + "[10] Pat S. Chavez and Andy Yaw Kwarteng. Extracting spectral contrast in landsat thematic mapper image data using selective principal component analysis. Photogramm. Eng. Remote Sens., 55:339-348, 1989. 2", + "[11] Lihui Chen, Gemine Vivone, Zihao Nie, Jocelyn Chanussot, and Xiaomin Yang. Spatial data augmentation: Improving the generalization of neural networks for pansharpening. IEEE TGRS, 61:1-11, 2023. 3", + "[12] Zhi-Xuan Chen, Cheng Jin, Tian-Jing Zhang, Xiao Wu, and Liang-Jian Deng. SpanConv: A new convolution via spanning kernel space for lightweight pansharpening. In IJCAI, pages 841-847, 2022. Main Track. 3", + "[13] Minhyeok Choi, Rae Young Kim, and Myeong-Ryong Nam. Fusion of multispectral and panchromatic satellite images" + ], + "bbox": [ + 93, + 191, + 485, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "using the curvelet transform. IEEE GRSL, 2(2):136-140, 2005. 3", + "[14] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE TGRS, pages 1-1, 2022. 3", + "[15] Matteo Ciotola, Giovanni Poggi, and Giuseppe Scarpa. Unsupervised deep learning-based pansharpening with jointly enhanced spectral and spatial fidelity. IEEE TGRS, 61:1-17, 2023. 3", + "[16] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pixel-wise ensembled masked autoencoder for multispectral pansharpening. IEEE TGRS, 62:1-22, 2024. 3, 6", + "[17] Yongchuan Cui, Peng Liu, Bingze Song, Lingjun Zhao, Yan Ma, and Lajiao Chen. Reconstruction of large-scale missing data in remote sensing images using Extend-GAN. IEEE GRSL, 21:1-5, 2024. 3", + "[18] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pansharpening via predictive filtering with element-wise feature mixing. ISPRS J. Photogramm. Remote Sens., 219:22-37, 2025. 3, 5", + "[19] Liang-Jian Deng, Gemine Vivone, Cheng Jin, and Jocelyn Chanussot. Detail injection-based deep convolutional neural networks for pansharpening. IEEE TGRS, 59(8):6995-7010, 2021. 1, 2, 5, 7", + "[20] Liang-jian Deng, Gemine Vivone, Mercedes E. Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pansharpening: A benchmark, from shallow to deep networks. IEEE GRSM, 10(3):279-315, 2022. 1, 2, 3", + "[21] Shangqi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, and Rui Wen. Bidirectional dilation Transformer for multispectral and hyperspectral image fusion. In IJCAI, 2023. 2, 3", + "[22] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. PSRT: Pyramid shuffle-and-reshuffle Transformer for multispectral and hyperspectral image fusion. IEEE TGRS, 61:1-15, 2023. 2, 3", + "[23] Renwei Dian, Anjing Guo, and Shutao Li. Zero-shot hyperspectral sharpening. IEEE TPAMI, 45(10):12650-12666, 2023. 2, 3", + "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3", + "[25] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In CVPR, pages 27738-27747, 2024. 3", + "[26] Andrea Garzelli and Filippo Nencini. Wavelet-based fusion of optical and SAR image data in remote sensing applications. Int. J. Remote Sens., 23(23):5025-5039, 2002. 3", + "[27] Morteza Ghahremani and Hassan Ghassemian. Nonlinear IHS: A promising method for pan-sharpening. IEEE GRSL, 13(11):1606-1610, 2016. 2" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11858", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pyramid dual domain injection network for pan-sharpening. In ICCV, pages 12862-12871, 2023. 5", + "[29] Xuanhua He, Keyu Yan, Jie Zhang, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Multiscale dual-domain guidance network for pan-sharpening. IEEE TGRS, 61:1-13, 2023. 5", + "[30] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. AAAI, 38(3):2121-2129, 2024. 3", + "[31] Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. PanMamba: Effective pan-sharpening with state space model. Inf.Fusion., 115:102779, 2025. 2, 3", + "[32] Junming Hou, Xinyang Liu, Chenxu Wu, Xiaofeng Cong, Chentong Huang, Liang-Jian Deng, and Jian Wei You. Bidomain uncertainty gated recursive network for pan-sharpening. Inf. Fusion., 118:102938, 2025. 3", + "[33] Cheng Jin, Liang-Jian Deng, Ting-Zhu Huang, and Gemine Vivone. Laplacian pyramid networks: A new approach for multispectral pansharpening. Inf. Fusion., 78:158-170, 2022. 3", + "[34] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. LAGConv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. AAAI, 36(1):1113-1121, 2022. 3, 5", + "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[36] Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepcev, and Gustavo K. Rohde. Optimal mass transport: Signal processing and machine-learning applications. IEEE SPM, 34 (4):43-59, 2017. 4", + "[37] Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo K. Rohde. Generalized sliced wasserstein distances, 2019. 4", + "[38] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. 2000. US Patent 6,011,875. 2", + "[39] Zhixuan Li, Jinjiang Li, Fan Zhang, and Linwei Fan. CADUI: Cross-attention-based depth unfolding iteration network for pansharpening remote sensing images. IEEE TGRS, 61:1–20, 2023. 3", + "[40] YuJie Liang, ZiHan Cao, Shangqi Deng, Hong-Xia Dou, and Liang-Jian Deng. Fourier-enhanced implicit neural fusion network for multispectral and hyperspectral image fusion. In NeurIPS, pages 63441-63465. Curran Associates, Inc., 2024. 3, 5, 6", + "[41] J. G. Liu. Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details. Int. J. Remote Sens., 21(18):3461-3472, 2000. 3", + "[42] Peng Liu, Jun Li, Lizhe Wang, and Guojin He. Remote sensing data fusion with generative adversarial networks: State-of-the-art methods and future research directions. IEEE GRSM, 10(2):295-328, 2022. 2", + "[43] Qingjie Liu, Huanyu Zhou, Qizhi Xu, Xiangyu Liu, and Yun-hong Wang. PSGAN: A generative adversarial network for" + ], + "bbox": [ + 91, + 90, + 485, + 901 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "remote sensing image pan-sharpening. IEEE TGRS, 59(12): 10227-10242, 2020. 2, 3", + "[44] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-GAN: a multiscale unsupervised network for remote sensing image pansharpening. IEEE TGRS, 61:1-18, 2023. 3", + "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In ICCV, 2021. 3", + "[46] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-GAN: An unsupervised pan-sharpening method for remote sensing image fusion. Inf. Fusion., 62: 110-120, 2020. 2, 3, 6", + "[47] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. In Remote Sens., page 594, 2016. 2, 3, 5", + "[48] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. PanDiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE TGRS, 61:1-17, 2023. 2, 3", + "[49] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE GRSM, 9(1):18-52, 2021. 1, 2, 5", + "[50] Xiangchao Meng, Nan Wang, Feng Shao, and Shutao Li. Vision Transformer for pansharpening. IEEE TGRS, 60:1-11, 2022. 2, 3", + "[51] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pages 666-704, 1781. 3", + "[52] Fjola Palsson, Johannes R. Sveinsson, and Magnus O. Ulfarsson. A new pansharpening algorithm based on total variation. IEEE GRSL, 11(1):318-322, 2014. 3", + "[53] Siran Peng, Xiangyu Zhu, Haoyu Deng, Liang-Jian Deng, and Zhen Lei. FusionMamba: Efficient remote sensing image fusion with state space model. IEEE TGRS, 2024. 2, 3", + "[54] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Found. Trends Mach. Learn., 11(5-6):355-607, 2019. 3", + "[55] Sheila Rahmani, Melissa Strait, Daria Merkurjev, Michael Moeller, and Todd Wittman. An adaptive IHS pan-sharpening method. IEEE GRSL, 7(4):746-750, 2010. 2", + "[56] Aaditya Ramdas, Nicolas Garcia, and Marco Cuturi. On Wasserstein two sample testing and related families of nonparametric tests, 2015. 4", + "[57] Xiangyu Rui, Xiangyong Cao, Yining Li, and Deyu Meng. Variational zero-shot multispectral pansharpening. IEEE TGRS, 62:1-16, 2024. 3", + "[58] Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, and Deyu Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion., 107:102325, 2024. 2, 3", + "[59] Filippo Santambrogio. One-dimensional issues, pages 59-85. Springer International Publishing, Cham, 2015. 4" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11859", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Giuseppe Scarpa and Matteo Ciotola. Full-resolution quality assessment for pansharpening. Remote Sens., 14(8), 2022. 6", + "[61] Vijay P. Shah, Nicolas H. Younan, and Roger L. King. An efficient pan-sharpening method via a combined adaptive PCA approach and contourlets. IEEE TGRS, 46(5):1323-1335, 2008. 2", + "[62] Xunyang Su, Jinjiang Li, and Zhen Hua. Transformer-based regression network for pansharpening remote sensing images. IEEE TGRS, 60:1-23, 2022. 2, 3", + "[63] Weiwei Sun, Jun Zhou, Xiangchao Meng, Gang Yang, Kai Ren, and Jiangtao Peng. Coupled temporal variation information estimation and resolution enhancement for remote sensing spatial-temporal-spectral fusion. IEEE TGRS, 61: 1-18, 2023. 3", + "[64] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008. 2", + "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3", + "[66] Cédric Villani et al. Optimal transport: old and new. Springer, 2008. 4", + "[67] Gemine Vivone, Miguel Simões, Mauro Dalla Mura, Rocco Restaino, José M. Bioucas-Dias, Giorgio A. Licciardi, and Jocelyn Chanussot. Pansharpening based on semiblind deconvolution. IEEE TGRS, 53(4):1997-2010, 2015. 3", + "[68] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O. Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE GRSM, 9(1):53-81, 2021. 1, 2, 3", + "[69] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. *Photogramm. Eng. Remote Sens.*, 63:691-699, 1997. 3, 6", + "[70] Hebaixu Wang, Meiqi Gong, Xiaoguang Mei, Hao Zhang, and Jiayi Ma. Deep unfolded network with intrinsic supervision for pan-sharpening. AAAI, 38(6):5419-5426, 2024. 3", + "[71] Hebaixu Wang, Hao Zhang, Xin Tian, and Jiayi Ma. Zero-Sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation. Inf. Fusion., 101:102003, 2024. 3", + "[72] Jun-Da Wang, Liang-Jian Deng, Chen-Yu Zhao, Xiao Wu, Hong-Ming Chen, and Gemine Vivone. Cascadic multireceptive learning for multispectral pansharpening. IEEE TGRS, 61:1-16, 2023. 3", + "[73] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE TGRS, 62:1-16, 2024. 3", + "[74] Zhong-Cheng Wu, Ting-Zhu Huang, Liang-Jian Deng, JinFan Hu, and Gemine Vivone. $\\mathrm{VO + }$ Net: An adaptive approach using variational optimization and deep learning for panchromatic sharpening. IEEE TGRS, 60:1-16, 2022. 3", + "[75] Jin-Liang Xiao, Ting-Zhu Huang, Liang-Jian Deng, Zhong-Cheng Wu, Xiao Wu, and Gemine Vivone. Variational pan-sharpening based on coefficient estimation with nonlocal regression. IEEE TGRS, 61:1-15, 2023. 3" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Guangxu Xie, Rencan Nie, Jinde Cao, He Li, and Jintao Li. A deep multiresolution representation framework for pansharpening. IEEE TGRS, 62:1-16, 2024. 3", + "[77] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xi-wei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE TGRS, 2024. 3", + "[78] Yinghui Xing, Litao Qu, Shizhou Zhang, Kai Zhang, Yanning Zhang, and Lorenzo Bruzzone. CrossDiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE TIP, 2024. 2, 3", + "[79] Qizhi Xu, Yuan Li, Jinyan Nie, Qingjie Liu, and Mengyao Guo. UPanGAN: Unsupervised pansharpening based on the spectral and spatial loss constrained generative adversarial network. Inf. Fusion., 91:31-46, 2023. 3", + "[80] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In CVPR, pages 1366-1375, 2021. 3, 5", + "[81] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pansharpening. In CVPR, pages 1778-1787, 2022. 3, 5, 6, 7", + "[82] Jie Yang, Xiao-Yang Fu, Yan Hu, Yue Huang, Xing Ding, and John Paisley. PanNet: A deep network architecture for pan-sharpening. ICCV, pages 1753-1761, 2017. 2, 3, 5", + "[83] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE JSTARS, 11(3):978-989, 2018. 2, 3, 5", + "[84] Jie Zhang, Xuanhua He, Keyu Yan, Ke Cao, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Pan-sharpening with wavelet-enhanced high-frequency information. IEEE TGRS, 62:1-14, 2024. 3", + "[85] Chen-Yu Zhao, Tian-Jing Zhang, Ran Ran, Zhi-Xuan Chen, and Liang-Jian Deng. LGPCConv: Learnable gaussian perturbation convolution for lightweight pansharpening. In IJCAI, pages 4647-4655, 2023. Main Track. 3", + "[86] Mengjiao Zhao, Mengting Ma, Xiangdong Li, Xiaowen Ma, Wei Zhang, and Siyang Song. Supervised detail-guided multiscale state space model for pan-sharpening. IEEE TGRS, 2024. 2, 3", + "[87] Kaiwen Zheng, Jie Huang, Man Zhou, Danfeng Hong, and Feng Zhao. Deep adaptive pansharpening via uncertainty-aware image fusion. IEEE TGRS, 61:1-15, 2023. 5", + "[88] Yu Zhong, Xiao Wu, Liang-Jian Deng, ZiHan Cao, and Hong-Xia Dou. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pansharpening. In NeurIPS, pages 77962-77986. Curran Associates, Inc., 2024. 3", + "[89] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In ECCV, pages 274-291, Cham, 2022. Springer Nature Switzerland. 3, 5", + "[90] Man Zhou, Jie Huang, Keyu Yan, Danfeng Hong, Xiuping Jia, Jocelyn Chanussot, and Chongyi Li. A general spatial-frequency learning framework for multimodal image fusion. IEEE TPAMI, pages 1-18, 2024. 3, 5" + ], + "bbox": [ + 516, + 90, + 906, + 885 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11860", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_model.json b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_model.json new file mode 100644 index 0000000000000000000000000000000000000000..05ac227578def4513e7d23737dbf0b5e0bdc090a --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_model.json @@ -0,0 +1,2565 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.18, + 0.131, + 0.823, + 0.175 + ], + "angle": 0, + "content": "Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need" + }, + { + "type": "text", + "bbox": [ + 0.274, + 0.203, + 0.726, + 0.222 + ], + "angle": 0, + "content": "Yongchuan Cui \\(^{1,2}\\) Peng Liu \\(^{1,2,*}\\) Hui Zhang \\(^{3,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.222, + 0.888, + 0.24 + ], + "angle": 0, + "content": "1Aerospace Information Research Institute, Chinese Academy of Sciences, 100094 Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.239, + 0.766, + 0.275 + ], + "angle": 0, + "content": "\\(^{2}\\) School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, 101408 Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.275, + 0.808, + 0.293 + ], + "angle": 0, + "content": "\\(^{3}\\) School of Engineering Medicine, Beihang University, 100191 Beijing, China" + }, + { + "type": "text", + "bbox": [ + 0.173, + 0.296, + 0.82, + 0.31 + ], + "angle": 0, + "content": "yongchuancui@gmail.com, liupeng202303@aircas.ac.cn, hui.zhang@buaa.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.33, + 0.327, + 0.345 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.362, + 0.486, + 0.74 + ], + "angle": 0, + "content": "Existing deep learning-based models for remote sensing pansharpening exhibit exceptional performance on training datasets. However, due to sensor-specific characteristics and varying imaging conditions, these models suffer from substantial performance degradation when applied to unseen satellite data, lacking generalizability and thus limiting their applicability. We argue that the performance drops stem primarily from distributional discrepancies from different sources and the key to addressing this challenge lies in bridging the gap between training and testing distributions. To validate the idea and further achieve a \"train once, deploy forever\" capability, this paper introduces a novel and intuitive approach to empower any pansharpening models with generalizability by employing a unified distribution strategy (UniPAN). Specifically, we construct a distribution transformation function that normalizes the pixels sampled from different sources to conform to an identical distribution. The deep models are trained on the transformed domain, and during testing on new datasets, the new data are also transformed to match the training distribution. UniPAN aims to train and test the model on a unified and consistent distribution, thereby enhancing its generalizability. Extensive experiments validate the efficacy of UniPAN, demonstrating its potential to significantly enhance the performance of deep pansharpening models across diverse satellite sensors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.466, + 0.755 + ], + "angle": 0, + "content": "Codes: https://github.com/yc-cui/UniPAN." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.787, + 0.222, + 0.803 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.813, + 0.486, + 0.875 + ], + "angle": 0, + "content": "Due to the inherent limitations of sensor imaging technology, advanced remote sensing satellites such as IKONOS, QuickBird, GaoFen-2, and WorldView-3, necessitate a compromise between the acquisition of extensive spectral infor" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.336, + 0.637, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.648, + 0.337, + 0.765, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.779, + 0.337, + 0.895, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Comparison of generalization metrics w/ and w/o UniPAN trained on WorldView-3 (WV3) using FusionNet [19]." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.554, + 0.896, + 0.706 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.514, + 0.718, + 0.904, + 0.733 + ], + "angle": 0, + "content": "Figure 2. Comparison of tSNE visualization w/ and w/o UniPAN." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.907, + 0.868 + ], + "angle": 0, + "content": "mation and maintaining fine spatial details in their captured imagery [20, 49, 68]. Consequently, these satellites typically capture either high spatial resolution single-band panchromatic (PAN) images or low spatial resolution multispectral (LRMS) images. Pansharpening is a crucial technique in remote sensing image processing, aiming to combine PAN and LRMS images to generate fused images with high spatial resolution and abundant spectra (HRMS)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Traditional pansharpening methods have been extensively studied over the past few decades. However, these" + }, + { + "type": "page_footnote", + "bbox": [ + 0.095, + 0.888, + 0.218, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11850" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.288 + ], + "angle": 0, + "content": "methods may introduce spectral distortions, require significant computational resources, and are sensitive to the quality of input data [20, 42, 49, 68]. In contrast, deep learning (DL)-based methods have shown great potential in pansharpening. Notably, various innovative architectures have emerged, starting from early convolutional models [19, 47, 82, 83], progressing to Transformer-based approaches [8, 21, 22, 24, 50, 62, 65], then generative methods [43, 46, 48, 58, 78], and continuing with the emergence of the latest novel designs [9, 23, 31, 53, 86]. The networks are evolving towards larger parameters capacity, stronger representation ability, and increasingly complex and diverse structures." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.293, + 0.487, + 0.55 + ], + "angle": 0, + "content": "However, as deep pansharpening thrives, the overly powerful fitting capabilities of these networks have concurrently presented several ongoing concerns, e.g. generalizability, interpretability, over-reliance on high-quality data, etc. This paper specifically delves into the widely criticized issue of generalizability. Specifically, although deep models achieve excellent performance on training datasets, they often fail to generalize well to new data due to distribution shift such as different noise levels, varying spectral signatures and sensor conditioning characteristics. While models can achieve satisfactory performances on new datasets through retraining or fine-tuning, the process is time-consuming and fine-tuning on every new dataset is impractical and cumbersome. The ideal scenario would involve developing a pansharpening model that can be trained once and deployed effectively across diverse datasets and sensors, thereby providing a train once, deploy forever capability." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.554, + 0.487, + 0.903 + ], + "angle": 0, + "content": "To this end, this paper introduces a novel approach that aims to empower any pansharpening models with generalizability by employing a unified distribution strategy (dubbed as UniPAN). The proposed method aims to bridge the gap between training and testing distributions, thereby improving the model's ability to generalize to unseen data. The core idea is to normalize the distribution of spatial pixels from the training dataset to conform to a predefined distribution (e.g. Gaussian distribution) and deep models are then trained on the transformed domain. This process is model-agnostic and requires no additional trainable parameters. When generalizing to new data, testing datasets are also transformed to match the training distribution, ensuring that the models are evaluated under the same distributional assumptions. Fig. 1 presents the cross-satellite generalization performance of a well-known pansharpening model FusionNet [19] on NBUPansharpRSData [49] trained w/ and w/o UniPAN. The model utilizing UniPAN demonstrates a significant improvement on full-resolution metrics, which validates the strong generalization ability for real-world scenes. To further elucidate the efficacy of UniPAN, Fig. 2 presents tSNE [64] dimensionality reduction clustering of multispectral images from diverse satellites using NBUPansharpRSData [49] w/" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.272 + ], + "angle": 0, + "content": "and w/o UniPAN. The pre-UniPAN visualization reveals distinct clustering patterns, with data points from different satellites forming separate groups, indicative of substantial inter-satellite distributional disparities. In contrast, following the application of UniPAN, a notable transformation is observed: the previously separated clusters become substantially intermingled, demonstrating a significant reduction in distributional discrepancies across different satellite datasets (e.g. the purple scatters of QuickBird). This visual evidence underscores UniPAN's capability to harmonize heterogeneous satellite data, effectively creating a more cohesive and unified representation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.908, + 0.303 + ], + "angle": 0, + "content": "The contributions of this paper could be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.308, + 0.909, + 0.444 + ], + "angle": 0, + "content": "- We propose UniPAN, a model-agnostic approach that employs a unified distribution strategy to empower any deep pansharpening models with generalizability, achieving a train once, deploy forever capability. This strategy provides two advantages: 1) Architectural independence: Compatible with any existing or future network architecture without structural modifications. 2) Training efficiency: Eliminates the need for retraining or fine-tuning when adapting to new sensors." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.444, + 0.909, + 0.58 + ], + "angle": 0, + "content": "- UniPAN is implemented via inverse transform sampling, and we prove its equivalence to the optimal solution of optimal transport in the one-dimensional case. UniPAN requires no additional trainable parameters and only needs to be fitted once before training begins, introducing almost no additional computational overhead. It serves as a viable alternative to vanilla data normalization techniques such as min-max and \\(z\\)-score normalization, offering a more robust and unified approach to data distribution alignment." + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.58, + 0.909, + 0.701 + ], + "angle": 0, + "content": "- We conduct extensive experiments using a variety of deep pansharpening models across numerous satellites. The simple yet intuitive and effective strategy provides a universal solution to the generalization challenge in DL-based pansharpening and experimental results demonstrate its effectiveness, showcasing significant improvements in cross-dataset generalization across a wide range of models and satellite sensors." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.308, + 0.909, + 0.701 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.717, + 0.658, + 0.733 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.74, + 0.759 + ], + "angle": 0, + "content": "2.1. Pansharpening Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Traditional Methods. Traditional pansharpening methods can be broadly categorized into three paradigms: component substitution (CS), multi-resolution analysis (MRA), and variational optimization (VO) [20, 42, 49, 68]. CS-based methods [4, 10, 27, 38, 55, 61] operate by substituting the spatial component of the LRMS image with the PAN image. Representative algorithms include Gram-Schmidt (GS) [4, 38], principal component analysis (PCA) [10, 61], and intensity-hue-saturation (IHS) [27, 55]. While computationally effi" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11851" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.318 + ], + "angle": 0, + "content": "cient, CS methods often introduce spectral distortions due to mismatches between PAN and substituted components. MRA-based methods [1-3, 13, 26, 41] employ multi-scale decomposition tools (e.g. wavelets [26], contourlets [13]) to inject high-frequency details from PAN into LRMS. Typical implementations include MTF-GLP [1-3] and SFIM [41]. These methods generally preserve spectral fidelity better than CS approaches but may exhibit limited spatial enhancement. VO-based approaches [7, 52, 63, 67, 74, 75] formulate pansharpening as an optimization problem with handcrafted priors, such as the variational wavelet-based methods [7] and Bayesian approaches [52]. Despite their mathematical elegance, traditional methods often require manual parameter tuning and struggle with complex spatial-spectral relationships [20, 68]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.32, + 0.487, + 0.76 + ], + "angle": 0, + "content": "Deep Learning-Based Methods. The advent of deep learning (DL) have revolutionized pansharpening by enabling end-to-end learning of spatial-spectral relationships. Early pioneering convolutional neural networks (CNNs) such as PNN [47], PanNet [82], and MSDCNN [83] established the feasibility of learning fusion rules from a mount of simulated data. Subsequent innovations introduced multi-scale architectures [33, 72, 76], attention mechanisms [39, 73], and kernel customizing [12, 18, 25, 34, 85], etc. The CNN-based methods demonstrated superior performance over traditional approaches but faced limitations in modeling long-range dependencies. In contrast, Transformer [24, 45, 65] addressed this limitation with self-attention mechanisms. Although various Transformer architectures specifically designed for pansharpening [8, 21, 22, 50, 62] have achieved impressive performance, they face criticism for quadratic computational complexity w.r.t. sequence length [16]. To address this limitation, recent studies have explored linear-complexity models [16, 31, 53, 86] that maintain performance while improving efficiency. To further enhance fidelity preservation, generative models like generative adversarial networks (GANs) [17, 43, 44, 46, 79] and diffusion models [48, 58, 77, 78, 88] have emerged as promising solutions. Beyond conventional configurations, contemporary research has expanded into more challenging frontiers to address practical constraints, including zero-shot learning paradigms [9, 23, 57, 71], frequency domain integration [30, 40, 84, 89, 90], and interpretable fusion mechanisms [32, 70, 80, 81]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.465, + 0.789 + ], + "angle": 0, + "content": "2.2. Generalization in DL-Based Pansharpening" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Despite remarkable progress in DL-based pansharpening, generalization capability remains a persistent challenge. The issue manifests conspicuously in two critical scenarios: 1) Models trained under the Wald protocol [69] using spatially degraded data often fail to maintain performance when applied to full-resolution inference, frequently exhibiting severe spectral distortions. 2) Networks optimized for specific" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.257 + ], + "angle": 0, + "content": "satellite sensors demonstrate substantial performance degradation when deployed on unseen satellite platforms due to distribution shifts in spectral signatures and sensor characteristics. Existing approaches to address these challenges primarily focus on unsupervised learning [14, 15, 46], generalized module design [77], or advanced data augmentation [11]. However, such methods typically require either meticulous architectural redesign or computationally intensive retraining/fine-tuning procedures. These requirements pose significant practical limitations for real-world deployment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.909, + 0.409 + ], + "angle": 0, + "content": "In contrast to model-centric solutions, our approach adopts a fundamentally different perspective by addressing the generalization challenge through data-space intervention. The proposed UniPAN framework introduces a model-agnostic preprocessing technique that operates on raw sensor data. Compared to traditional linear data normalization such as min-max scaling or \\(z\\)-score standardization, our method employs a non-linear mapping to project the input data onto a unified distribution, effectively harmonizing the distribution mismatches from all sensors." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.42, + 0.606, + 0.435 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.771, + 0.459 + ], + "angle": 0, + "content": "3.1. Distribution Transformation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.572 + ], + "angle": 0, + "content": "The core challenge in cross-domain pansharpening stems from the distribution discrepancy between source (training) and target (testing) domains. Let \\(\\mathcal{X}_S\\) denote the source domain (training data) and \\(\\mathcal{X}_T\\) the target domain (testing data). Our objective is to learn a transformation \\(\\mathcal{T}:\\mathcal{X}\\to \\mathcal{Z}\\) that maps both domains to a shared space \\(\\mathcal{Z}\\) with unified distribution \\(\\rho_t\\), e.g. standard normal distribution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.573, + 0.909, + 0.634 + ], + "angle": 0, + "content": "Inverse Transform Sampling. Given a univariate r.v. \\( X \\sim \\rho_s \\), we seek a transformation \\( \\mathcal{T} : \\mathbb{R} \\to \\mathbb{R} \\) such that \\( Z = \\mathcal{T}(X) \\sim \\rho_t \\). The inverse transform sampling theorem provides a theoretically grounded solution:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.637, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.909, + 0.797 + ], + "angle": 0, + "content": "where \\( F_{S}(x) = \\rho_{s}(-\\infty, x] \\) is the cumulative distribution function (CDF) of \\( X \\), and \\( F_{T}^{-1}(u) = \\inf \\{z \\in \\mathbb{R} : F_{T}(z) \\geq u\\} \\) is the inverse CDF of the target distribution. By the universality of the uniform, \\( U = F_{S}(X) \\) follows a uniform distribution, and \\( Z = F_{T}^{-1}(U) \\) inherits \\( \\rho_{t} \\) exactly when \\( F_{S} \\) is continuous. To provide more intuitively analyze of the algorithm's operational process and efficiency, we establish a fundamental connection between inverse transform sampling and the well-studied optimal transport theory [51, 54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.797, + 0.909, + 0.857 + ], + "angle": 0, + "content": "Optimal Transport Perspective. Given two probability distributions \\(\\rho_{s}\\) (source domain) and \\(\\rho_{t}\\) (target domain) defined on \\(\\mathbb{R}^n\\) and \\(\\mathbb{R}^m\\) respectively, we consider the Monge problem [51, 54]:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.862, + 0.909, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} (\\rho_ {s}, \\rho_ {t}) \\triangleq \\min _ { \\begin{array}{c} T: \\mathbb {R} ^ {n} \\to \\mathbb {R} ^ {m} \\\\ T _ {\\sharp} \\rho_ {s} = \\rho_ {t} \\end{array} } \\int_ {\\mathbb {R} ^ {n}} c (x, T (x)) \\rho_ {s} (x) d x, \\quad (2)\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11852" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.138 + ], + "angle": 0, + "content": "where \\(c:\\Omega_s\\times \\Omega_t\\to \\mathbb{R}^+\\) is a cost function and for any measurable Borel set \\(E\\in \\Omega_s\\) \\(T:\\Omega_s\\rightarrow \\Omega_t\\) transfers measures from one space \\(\\Omega_{s}\\) to another space \\(\\Omega_t\\)" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.15, + 0.484, + 0.168 + ], + "angle": 0, + "content": "\\[\nT _ {\\sharp} \\rho_ {s} (E) = \\rho_ {s} \\left(T ^ {- 1} (E)\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.484, + 0.24 + ], + "angle": 0, + "content": "here \\( T_{\\sharp} \\) is the pushforward operator. The Monge problem seeks a transport plan \\( T \\) from \\( \\rho_s \\) to \\( \\rho_t \\) that minimizes the transportation cost. The optimal \\( T^{*} \\) is also known as the Monge map of Eq. (2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.242, + 0.486, + 0.345 + ], + "angle": 0, + "content": "Let \\( F_{S}(x) = \\rho_{s}(-\\infty, x] \\) and \\( F_{T}(z) = \\rho_{t}(-\\infty, z] \\) denote the CDFs of the source and target distributions respectively. In the univariate setting (\\( n = m = 1, \\Omega_{s} = \\Omega_{t} = \\mathbb{R} \\)), when the source measure \\( \\rho_{s} \\) is continuous and the cost function takes the form \\( c(x, z) = h(|x - z|) \\) with \\( h \\) being strictly convex, by cyclical monotonicity [66], the Monge map \\( T^{*} \\) satisfies:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.349, + 0.484, + 0.365 + ], + "angle": 0, + "content": "\\[\nF _ {S} (x) = F _ {T} \\left(T ^ {*} (x)\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.305, + 0.39 + ], + "angle": 0, + "content": "yielding the explicit expression:" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.401, + 0.484, + 0.42 + ], + "angle": 0, + "content": "\\[\nT ^ {*} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.486, + 0.479 + ], + "angle": 0, + "content": "This remarkable result coincides exactly with the inverse transform sampling formula Eq. (1), i.e. \\( T^{*} = \\mathcal{T} \\). The transport cost in this setting [37, 56, 59] can be computed as" + }, + { + "type": "equation", + "bbox": [ + 0.153, + 0.491, + 0.484, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {C} \\left(\\rho_ {s}, \\rho_ {t}\\right) = \\int_ {0} ^ {1} c \\left(F _ {S} ^ {- 1} (u), F _ {T} ^ {- 1} (u)\\right) d u. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.486, + 0.79 + ], + "angle": 0, + "content": "For the \\( L_{1} \\) cost function \\( c(x,z) = |x - z| \\), this corresponds to the \\( W_{1} \\) Wasserstein distance [36, 59] between the distributions. In the one-dimensional (1D) scenario, UniPAN operates independently on spatial pixels within each spectral band. This implementation can be achieved through interpolation-based inverse transform sampling without framing it as an optimization problem. However, given the inherent correlations between spectral bands and interdependencies among spatial patches, multidimensional extension of this framework may be promising, albeit with increased implementation complexity. The theoretical derivation establishing connections with optimal transport theory serves to facilitate future generalization of UniPAN to multidimensional configurations. Building upon the theoretical foundation of distribution transformation through optimal transport, in the following, we present a practical implementation of 1D UniPAN." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.318, + 0.818 + ], + "angle": 0, + "content": "3.2. UniPAN Implementation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.87 + ], + "angle": 0, + "content": "We present a practical implementation for 1D finite discrete distributions of UniPAN through a quantile-based distribution transformation algorithm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Finite Discrete Distribution Transformation. Given pixel values \\( X = \\{x_{1}, x_{2}, \\dots, x_{n}\\} \\) sampled from a spectral band," + }, + { + "type": "code_caption", + "bbox": [ + 0.515, + 0.091, + 0.873, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 UniPAN Training and Testing Procedure" + }, + { + "type": "algorithm", + "bbox": [ + 0.514, + 0.111, + 0.905, + 0.391 + ], + "angle": 0, + "content": "Training Phase \nInput: Training dataset \\(\\mathcal{D}_{\\mathrm{train}} = \\{L_i, P_i, H_i\\}_{i=1}^N\\), inverse CDF \\(F_T^{-1}\\) of target distribution \\(\\rho_t\\), model \\(\\mathcal{M}\\) \nOutput: Trained model \\(M\\) \n1: Initialize transformed LRMS \\(\\tilde{L}\\) and PAN \\(\\tilde{P}\\) \n2: for each band \\(b\\) of LRMS \\(\\{L_i\\}_{i=1}^N\\) & PAN \\(\\{P_i\\}_{i=1}^N\\) do \n3: Sample \\(n\\) pixels \\(X_b \\gets \\{x_1, \\ldots, x_n\\}\\) from band \\(b\\) \n4: Compute \\(m\\) quantiles \\(Q = \\{q_0, \\ldots, q_{m-1}\\} \\triangleright\\) Eq. (7) \n5: Linspace uniform CDF \\(U = \\{0, \\frac{1}{m-1}, \\ldots, 1\\}\\) \n6: for each pixel \\(x\\) in band \\(b\\) do \n7: Forward and backward interp by \\(u_\\uparrow\\) and \\(u_\\downarrow \\triangleright\\) Eq. (9) \n8: Compute eCDF value \\(u(x) \\gets (u_\\uparrow + u_\\downarrow) / 2 \\triangleright\\) Eq. (8) \n9: Transform \\(x\\) to \\(z \\gets F_T^{-1}(u(x))\\) and assign \\(z\\) to corresponding band of \\(\\tilde{L}\\) and \\(\\tilde{P}\\) \n10: end for \n11: end for \n12: while not converged do \n13: Train \\(\\mathcal{M}\\) with \\(\\mathcal{L} = \\|H - \\mathcal{M}(\\tilde{L}, \\tilde{P}) + L \\uparrow\\|_1\\) \n14: end while \n15: return Trained model \\(\\mathcal{M}\\)" + }, + { + "type": "code_caption", + "bbox": [ + 0.516, + 0.398, + 0.605, + 0.41 + ], + "angle": 0, + "content": "Testing Phase" + }, + { + "type": "algorithm", + "bbox": [ + 0.516, + 0.41, + 0.906, + 0.524 + ], + "angle": 0, + "content": "Input: Testing dataset \\(\\mathcal{D}_{\\mathrm{test}} = \\{L_i,P_i\\}_{i = 1}^N\\) , inverse CDF \\(F_{T}^{-1}\\) of target distribution \\(\\rho_{t}\\) , trained model \\(\\mathcal{M}\\) \nOutput: Pansharpening result \\(\\hat{H}^{\\mathrm{test}}\\) \n16: for each band \\(b\\) of LRMS & PAN do \n17: Transform \\(L,P\\) to \\(\\tilde{L},\\tilde{P}\\) using the training transformation \n18: end for \n19: Generate \\(\\hat{H}^{\\mathrm{test}}\\gets \\mathcal{M}(\\tilde{L},\\tilde{P}) + L\\uparrow\\) \n20: return Fused result \\(\\hat{H}^{\\mathrm{test}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.906, + 0.579 + ], + "angle": 0, + "content": "we first compute \\(m\\) empirical quantiles \\(Q = \\{q_{0},\\dots,q_{m - 1}\\}\\) where:" + }, + { + "type": "equation", + "bbox": [ + 0.513, + 0.587, + 0.908, + 0.629 + ], + "angle": 0, + "content": "\\[\nq _ {i} = \\inf \\left\\{x \\in X \\mid P (X \\leq x) \\geq \\frac {i}{m - 1} \\right\\}, i = \\{0, 1, \\dots , m - 1 \\}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.908, + 0.674 + ], + "angle": 0, + "content": "These quantiles discretize the empirical cumulative distribution function (eCDF) of the source data. The number of quantiles \\( m \\) controls the approximation fidelity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.809 + ], + "angle": 0, + "content": "Given any data point \\( x \\) to be transformed, the next step involves determining its value in the eCDF through interpolation. However, when the source distribution contains repeated values or sparse regions, quantile estimation may yield non-unique or overlapping quantile boundaries. This can result in abrupt jumps in the eCDF, thereby violating the properties of monotonicity and smoothness. Therefore, we use bidirectional interpolation to estimate a robust eCDF value of \\( x \\):" + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.818, + 0.907, + 0.873 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} u (x) = u _ {\\uparrow} (x) / 2 + u _ {\\downarrow} (x) / 2 \\\\ = \\operatorname {i n t e r p} (x, Q, U) / 2 + \\tag {8} \\\\ (1 - \\text {i n t e r p} (- x, - Q ^ {\\text {r e v}}, - U ^ {\\text {r e v}})) / 2, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.885, + 0.908, + 0.904 + ], + "angle": 0, + "content": "where \\(U = \\{0,\\frac{1}{m - 1},\\dots,1\\}\\), \\(Q^{\\mathrm{rev}}\\) is the reversed quantiles." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11853" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.079, + 0.907, + 0.411 + ], + "angle": 0, + "content": "
QNRTρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.980(+0.005)0.946(-0.010)0.971(+0.000)0.963(-0.003)0.964(-0.004)0.974(-0.000)0.985(+0.010)0.961(+0.004)0.979(+0.007)0.984(+0.017)0.982(+0.013)0.987(+0.013)
IK0.950(+0.005)0.912(+0.039)0.915(+0.045)0.924(+0.011)0.916(+0.022)0.958(+0.011)0.956(+0.011)0.919(+0.045)0.940(+0.070)0.950(+0.037)0.945(+0.051)0.968(+0.021)
QB0.957(+0.006)0.936(+0.016)0.931(+0.047)0.953(+0.017)0.944(+0.014)0.970(+0.016)0.965(+0.014)0.944(+0.024)0.966(+0.082)0.969(+0.032)0.969(+0.038)0.970(+0.017)
WV20.880(+0.026)0.811(+0.029)0.832(+0.062)0.946(+0.008)0.916(+0.012)0.951(+0.012)0.872(+0.018)0.775(-0.006)0.825(+0.055)0.944(+0.006)0.914(+0.011)0.940(+0.001)
WV30.862(-0.006)0.804(+0.035)0.795(+0.067)0.921(-0.011)0.919(+0.011)0.950(+0.001)0.880(+0.011)0.805(+0.036)0.813(+0.084)0.943(+0.010)0.924(+0.016)0.952(+0.003)
WV40.924(+0.009)0.898(+0.045)0.875(+0.036)0.948(-0.000)0.929(+0.003)0.972(+0.006)0.939(+0.024)0.889(+0.035)0.903(+0.064)0.952(+0.002)0.935(+0.009)0.973(+0.007)
Dλ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.007(-0.003)0.022(+0.008)0.005(-0.000)0.017(+0.004)0.015(+0.005)0.011(+0.000)0.004(-0.005)0.012(-0.001)0.004(-0.001)0.004(-0.008)0.004(-0.005)0.004(-0.005)
IK0.011(-0.000)0.020(-0.015)0.018(-0.011)0.031(+0.004)0.022(-0.003)0.014(-0.003)0.008(-0.003)0.020(-0.015)0.013(-0.016)0.013(-0.014)0.009(-0.016)0.010(-0.008)
QB0.013(-0.000)0.011(-0.004)0.015(-0.010)0.016(-0.004)0.014(+0.000)0.010(-0.005)0.009(-0.003)0.009(-0.006)0.007(-0.017)0.008(-0.013)0.006(-0.007)0.012(-0.003)
WV20.042(-0.006)0.066(-0.007)0.037(-0.019)0.008(-0.000)0.014(-0.000)0.016(-0.002)0.046(-0.002)0.086(+0.012)0.048(-0.008)0.008(+0.000)0.015(+0.001)0.022(+0.002)
WV30.050(+0.010)0.067(-0.011)0.049(-0.024)0.027(+0.011)0.014(+0.000)0.016(+0.002)0.042(+0.002)0.072(-0.006)0.051(-0.022)0.013(-0.002)0.011(-0.002)0.014(+0.000)
WV40.024(-0.000)0.031(-0.012)0.029(-0.009)0.017(+0.008)0.018(+0.008)0.005(-0.001)0.016(-0.008)0.032(-0.011)0.022(-0.016)0.014(+0.004)0.015(+0.005)0.004(-0.002)
DS↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.012(-0.003)0.032(+0.003)0.023(+0.000)0.020(+0.000)0.021(+0.000)0.015(-0.000)0.010(-0.005)0.026(-0.002)0.016(-0.006)0.010(-0.009)0.013(-0.007)0.007(-0.008)
IK0.039(-0.005)0.068(-0.026)0.068(-0.036)0.046(-0.014)0.063(-0.019)0.027(-0.007)0.035(-0.008)0.062(-0.033)0.046(-0.058)0.037(-0.024)0.045(-0.036)0.022(-0.013)
QB0.030(-0.005)0.053(-0.012)0.054(-0.039)0.031(-0.012)0.042(-0.014)0.020(-0.012)0.025(-0.011)0.046(-0.018)0.026(-0.067)0.023(-0.020)0.024(-0.031)0.020(-0.012)
WV20.081(-0.022)0.134(-0.024)0.137(-0.049)0.045(-0.008)0.071(-0.012)0.032(-0.010)0.085(-0.018)0.153(-0.005)0.133(-0.053)0.047(-0.006)0.071(-0.012)0.039(-0.003)
WV30.096(-0.000)0.141(-0.027)0.167(-0.050)0.054(+0.002)0.067(-0.011)0.033(-0.004)0.083(-0.013)0.134(-0.033)0.146(-0.072)0.043(-0.008)0.064(-0.014)0.033(-0.003)
WV40.053(-0.009)0.073(-0.036)0.099(-0.029)0.035(-0.006)0.053(-0.011)0.021(-0.005)0.045(-0.018)0.082(-0.027)0.076(-0.052)0.036(-0.004)0.052(-0.012)0.021(-0.005)
Dρ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.043(+0.003)0.110(+0.030)0.026(-0.008)0.079(+0.032)0.062(+0.015)0.041(-0.002)0.019(-0.020)0.059(-0.021)0.022(-0.013)0.015(-0.032)0.017(-0.029)0.017(-0.026)
IK0.082(-0.006)0.095(-0.038)0.090(-0.058)0.126(+0.024)0.110(-0.010)0.061(-0.008)0.064(-0.024)0.092(-0.040)0.066(-0.083)0.051(-0.049)0.061(-0.059)0.045(-0.024)
QB0.061(-0.003)0.154(-0.020)0.179(-0.059)0.135(+0.011)-0.133(-0.045)-0.093(-0.026)-0.164(-0.029)-0.235(-0.002)-0.200(-0.058)-0.100(-0.013)-0.126(-0.048)-0.112(-0.007)
WV2-1.166(-0.27)-1.198(-0.339)-1.199(-0.539)-1.155(-0.459)-1.133(-0.455)-1.137(-0.451)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)
WV3-2.205(+0.18)-2.214(-0.41)-2.234(-0.533)-2.231(-0.419)-2.173(+0.11)-2.173(-0.415)-2.176(-0.419)-2.226(-0.429)-2.219(-0.468)-2.117(-0.434)-1.131(-0.439)-1.131(-0.439)
WV4-3.155(-0.55)-3.112(-0.42)-3.115(-0.44)-3.146(+0.88)-3.169(+0.77)-3.169(-0.477)-3.173(-0.477)-3.176(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.421, + 0.907, + 0.463 + ], + "angle": 0, + "content": "Table 1. Average results of 12 models. Rows correspond to training satellites, and columns correspond to testing satellites. Green indicates performance improvement, while red indicates performance degradation. The values in parentheses represent the average delta compared to the baseline without UniPAN." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.486, + 0.522 + ], + "angle": 0, + "content": "The function \\( \\text{interp} \\) represents a linear interpolation between the quantile point \\( Q \\) and the reference point \\( U \\), with the interpolation formula given by:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.528, + 0.495, + 0.596 + ], + "angle": 0, + "content": "\\[\n\\operatorname {i n t e r p} (x, Q, U) = \\left\\{ \\begin{array}{l l} U [ 0 ], & x \\leq Q [ 0 ] \\\\ f (x, i, Q, U), & Q [ i ] < x \\leq Q [ i + 1 ] \\\\ U [ m - 1 ], & x > Q [ m - 1 ] \\end{array} \\right. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.597, + 0.356, + 0.612 + ], + "angle": 0, + "content": "where \\(f\\) computes the \\(Q\\) to \\(U\\) mapping," + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.617, + 0.484, + 0.662 + ], + "angle": 0, + "content": "\\[\nf (x, i, Q, U) = U [ i ] + \\frac {x - Q [ i ]}{Q [ i + 1 ] - Q [ i ]} (U [ i + 1 ] - U [ i ]) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.663, + 0.485, + 0.679 + ], + "angle": 0, + "content": "The last step is to apply inverse CDF of target distribution," + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.691, + 0.484, + 0.71 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} (x) = F _ {T} ^ {- 1} (u (x)) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.818 + ], + "angle": 0, + "content": "Training and Testing Procedure. The training and testing procedures of UniPAN are described in Algorithm 1. Here, \\( L \\uparrow \\) denotes the upsampled LRMS image using bicubic interpolation to match the size of the PAN image. The model \\( \\mathcal{M} \\) is trained to predict the residual of LRMS with an \\( l_{1} \\) loss function that minimizes the discrepancy between the predicted HRMS image and the ground truth." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.818, + 0.484, + 0.879 + ], + "angle": 0, + "content": "As for the target distribution \\(\\rho_{t}\\), we test two widely used distributions, i.e. standard normal distribution \\(\\mathcal{N}(0,1)\\) and standard uniform distribution \\(\\mathcal{U}(0,1)\\), the transformation \\(\\mathcal{T}(x)\\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.884, + 0.484, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} _ {\\mathrm {n}} (x) = \\Phi^ {- 1} \\left(F _ {S} (x)\\right), \\quad \\mathcal {T} _ {\\mathrm {u}} (x) = F _ {S} (x), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.907, + 0.583 + ], + "angle": 0, + "content": "where \\(\\Phi^{-1}\\) denotes the inverse CDF of standard normal distribution. Since the inverse CDF of standard uniform distribution is an identical mapping, \\(\\mathcal{T}_{\\mathrm{u}}\\) becomes the estimated empirical distribution itself. The standard normal distribution transformation non-linearly warps the input space to match Gaussian statistics, while the uniform transformation directly maps pixel intensities to their empirical quantiles." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.596, + 0.638, + 0.614 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.621, + 0.602, + 0.637 + ], + "angle": 0, + "content": "4.1. Setups" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Datasets. We conducted experiments utilizing the Red, Green, Blue, and NIR bands of the NBUPansharpRSDataset dataset [49]. The variations in resolution among different satellites, coupled with significant differences in land covers and land uses, render this dataset particularly suitable for assessing the generalization capabilities of pansharpening models. The dataset comprises data from 6 satellites. For each satellite, we partitioned the data into training, validation, and testing sets in a ratio of \\(7:1:2\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Implementation Details. To validate the effectiveness and generalization capability of UniPAN, we conducted comparative experiments between baseline (w/o UniPAN) and enhanced (w/ UniPAN) configurations using diverse commonly-recognized models, including FusionNet [19], PanNet [82], MSDDN [28, 29], GPPNN [80], MSDCNN [83], FeINFN [40], PNN [47], LAGConv [34], SFIIN [89, 90], PreMix [18], UAPN [87], MDCUN [81]." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11854" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.082, + 0.889, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.476, + 0.908, + 0.532 + ], + "angle": 0, + "content": "Figure 3. Generalization testing on GaoFen-1, the networks are trained on WorldView-3. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the \\( D_{\\lambda} / D_{S} \\) value. In most cases, the use of UniPAN yields superior spectral and spatial preservation. For instance, MDCUN [81] exhibits a highly close alignment with the LRMS in spectral characteristics with UniPAN of \\( \\rho_t = \\mathcal{N}(0,1) \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.547, + 0.486, + 0.716 + ], + "angle": 0, + "content": "All implementations followed identical experimental procedures: a batch size of 8 (except for MDCUN [81] and FeINFN [40] which used 4), the Adam [35] optimizer \\((\\beta_{1} = 0.9, \\beta_{2} = 0.999)\\) with an initial learning rate of \\(10^{-3}\\), and StepLR scheduling which reduces the learning rate by a factor of 0.8 every 100 epochs. Training was performed on an NVIDIA GeForce RTX 4090 GPU for 600 epochs using the Wald protocol [69] with \\(l_{1}\\) loss, with best model selected based on optimal validation performance. Note that, to prevent data leakage of testing set, the transformation fitting process is conducted on the training set of each satellite." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.487, + 0.904 + ], + "angle": 0, + "content": "Evaluation Metrics. The full-reference metrics based on spatial degradation (e.g. PSNR, SSIM) can only reflect performance in simulated reduced-resolution experiments and fail to demonstrate effectiveness in real-world full-resolution scenarios [6, 16, 46]. In contrast, no-reference metrics like QNR explicitly quantify model performance in both spectral preservation and spatial enhancement, enabling direct evaluation of fusion results at native resolution while avoiding theoretical biases introduced by downsampling. To better align with practical applications where reference images are unavailable, this paper adopts widely-used no-reference spectral and spatial metrics, i.e. QNR [5], \\(D_{\\lambda}\\) [5], \\(D_{S}\\) [5]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.579 + ], + "angle": 0, + "content": "and \\(D_{\\rho}\\) [60], to comprehensively assess cross-resolution and cross-satellite generalization capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.59, + 0.736, + 0.607 + ], + "angle": 0, + "content": "4.2.Quantitative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.613, + 0.909, + 0.794 + ], + "angle": 0, + "content": "Tab. 1 presents the average generalization performance of various models enhanced with UniPAN under different target distributions. Green indicate metric improvements, while red denote performance degradation. Slight yet observable differences exist between the two target distributions. Overall, the uniform distribution achieves better performance than the normal distribution. The normal distribution demonstrated a modest decline in performance across the WV2 and WV3 datasets. This phenomenon may stem from the peaked characteristics of the normal distribution, which could hinder model training, whereas the uniform distribution provides smoother normalization (see Fig. 5 in Sec. 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.796, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Notably, UniPAN with \\(\\rho_{t} = \\mathcal{U}(0,1)\\) achieves significant improvements across most metrics, with only marginal declines in rare cases. For instance, when generalizing from WV3 to QB, UniPAN improves the QNR metric by over 0.084 (from 0.729 to 0.813) averaged on all 12 models, underscoring its effectiveness. Beyond QNR, UniPAN also enhances spectral preservation \\((D_{\\lambda})\\) and spatial consistency" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11855" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.082, + 0.889, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.476, + 0.908, + 0.505 + ], + "angle": 0, + "content": "Figure 4. Full-resolution generalization testing on QuickBird. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the \\( D_{\\lambda} / D_{S} \\) value." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.52, + 0.486, + 0.702 + ], + "angle": 0, + "content": "\\((D_S, D_\\rho)\\), further validating the efficacy of the unified distribution strategy. These results confirm that aligning heterogeneous satellite data to a unified distribution effectively mitigates domain shifts, enabling robust cross-sensor generalization. However, while UniPAN improves spatial fidelity greatly, it may cause slight spectral distortion, as evidenced by the minor performance degradation observed in the \\(D_{\\lambda}\\) metric. This is reasonable since the transformation applied to each spectral band maintains the relative spatial structure but might result in minimal loss of spectral information. Nevertheless, such information loss is extremely limited, as the overall enhancement in performance remains predominant." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.276, + 0.729 + ], + "angle": 0, + "content": "4.3. Visual Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.901 + ], + "angle": 0, + "content": "Comparison of Fusion Results. To illustrate the practical impact of UniPAN, Figs. 3 and 4 compare fusion results w/ and w/o UniPAN. In most cases, UniPAN achieves highly favorable results in both spectral and spatial preservation. Furthermore, the remarkably high QNR values demonstrate its superior capability in balancing spectral and spatial fidelity, e.g. in Fig. 3, MDCUN [81] applying UniPAN with normal distribution enhanced the QNR by 0.1514 (from 0.7157 to 0.8671) and aligned fused results closely with the ground truth spectra. In the full-resolution generalization test depicted in Fig. 4, even with minimal visual differences," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.52, + 0.907, + 0.595 + ], + "angle": 0, + "content": "UniPAN achieves metrically significant improvements in spectral and spatial preservation. This capability is critical for downstream applications requiring accurate spectral information. Additional visual comparison results are included in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.909, + 0.793 + ], + "angle": 0, + "content": "Comparison of Data Distribution. Fig. 5 further analyzes data distribution alignment. The kernel density estimation (KDE) curves (top right area and diagonals) reveal that UniPAN successfully transforms the original complex and heterogeneous distributions into a unified target distribution. The quantile-quantile (Q-Q) plots (bottom left area) demonstrate harmonized distributions across satellites after applying UniPAN (The greater the deviation from the \\( y = x \\) line, the more significant the disparity between the two distributions). For example, the original Q-Q plots between WV4 and QB show significant dispersion, whereas post-UniPAN points close tightly around \\( y = x \\). This visual evidence corroborates UniPAN's ability to bridge distribution gaps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.704, + 0.82 + ], + "angle": 0, + "content": "4.4. Parameter Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "UniPAN introduces two hyperparameters: the number of sampled pixels \\( n \\) and the number of quantiles \\( m \\). Figs. 6 and 7 investigate their impact on performance of FusionNet [19] trained on WV2 and tested on GF1. As \\( n \\) and \\( m \\) vary, the metrics stabilize. Generally, \\( \\mathcal{U}(0,1) \\) demonstrates" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11856" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.09, + 0.355, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.294, + 0.268, + 0.305 + ], + "angle": 0, + "content": "(a) w/o UniPAN" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.09, + 0.627, + 0.292 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.415, + 0.294, + 0.578, + 0.306 + ], + "angle": 0, + "content": "(b) w/ UniPAN, \\(\\rho_t = \\mathcal{N}(0,1)\\)" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.091, + 0.9, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.691, + 0.294, + 0.848, + 0.306 + ], + "angle": 0, + "content": "(c) \\(\\mathrm{w} / \\mathrm{UniPAN}\\) \\(\\rho_t = \\mathcal{U}(0,1)\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.311, + 0.907, + 0.368 + ], + "angle": 0, + "content": "Figure 5. Comparison of data distributions before and after applying UniPAN. The upper-right and diagonals of each subplot display the kernel density estimation (KDE) curves of Green band for each satellite. The scatter plots in the lower-left areas of each subplots represent quantile-quantile (Q-Q) plots of green band, with the dashed line indicating \\( y = x \\). The greater the deviation from the \\( y = x \\) line, the more significant the disparity between the two distributions. Zoom in for better view." + }, + { + "type": "image", + "bbox": [ + 0.086, + 0.377, + 0.473, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.617, + 0.483, + 0.646 + ], + "angle": 0, + "content": "Figure 6. Parameter analysis of number of quantiles, with number of sampled pixels fixed to 10000." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.487, + 0.751 + ], + "angle": 0, + "content": "superior performance compared to \\(\\mathcal{N}(0,1)\\). Both configurations achieve statistically significant improvements over baseline methods without applying UniPAN. This parametric robustness demonstrates that optimal performance can be attained without exhaustive hyperparameter tuning, significantly reducing computational overhead." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.768, + 0.34, + 0.785 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.901 + ], + "angle": 0, + "content": "We proposed UniPAN, a unified distribution strategy addressing generalization challenges in deep learning-based pansharpening. UniPAN eliminates spectral discrepancies by normalizing multi-source remote sensing imagery into predefined target distributions, effectively bridging the gap between training and testing distributions. Extensive quantitative and visual comparisons across 6 satellite sensors and" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.377, + 0.892, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.617, + 0.907, + 0.646 + ], + "angle": 0, + "content": "Figure 7. Parameter analysis of number of sampled pixels, with number of quantiles fixed to 1000." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.719 + ], + "angle": 0, + "content": "12 models demonstrate that UniPAN significantly enhances cross-sensor generalization. Notably, this enhancement is achieved without introducing any additional trainable parameters or the need for retraining or fine-tuning the models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "While this work aligns theoretically with 1D optimal transport, considering interdependencies of spectral bands or spatial patches, future research could explore extending the method to the multi-dimensional case with optimal transport theories. Additionally, the experimental results indicate that the choice of target distribution has a notable impact on performance, thus it is worthwhile investigating adaptive target domain selection. To further validate the generalizability of the proposed unified distribution strategy, the ideas presented in this work could be extended to other remote sensing image processing tasks, such as multi-modal fusion and hyperspectral pansharpening, etc." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11857" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "Acknowledgment. This research was partially supported by the National Natural Science Foundation of China under Grant 32441114, 41971397, 61731022, and the project Y1H103101A." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.166, + 0.188, + 0.182 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.486, + 0.246 + ], + "angle": 0, + "content": "[1] B. Aiazzi, L. Alparone, S. Baronti, and A. Garzelli. Context-driven fusion of high spatial and spectral resolution images based on oversampled multiresolution analysis. IEEE TGRS, 40(10):2300-2312, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.249, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[2] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva. An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas. In 2003 2nd GRSS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, pages 90-94, 2003." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.333, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[3] Bruno Aiazzi, L. Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. MTF-tailored multiscale fusion of high-resolution MS and Pan imagery. *Photogramm. Eng. Remote Sens.*, 72:591-596, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.39, + 0.484, + 0.444 + ], + "angle": 0, + "content": "[4] Bruno Aiazzi, Stefano Baronti, and Massimo Selva. Improving component substitution pansharpening through multivariate regression of MS+Pan data. IEEE TGRS, 45(10): 3230-3239, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[5] Luciano Alparone, Bruno Aiazzi, Stefano Baronti, Andrea Garzelli, Filippo Nencini, and Massimo Selva. Multispectral and panchromatic data fusion assessment without reference. *Photogramm. Eng. Remote Sens.*, 74(2):193-200, 2008. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] Alberto Arienzo, Gemine Vivone, Andrea Garzelli, Luciano Alparone, and Jocelyn Chanussot. Full-resolution quality assessment of pansharpening: Theoretical and hands-on approaches. IEEE GRSM, 10(3):168-201, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.6 + ], + "angle": 0, + "content": "[7] Coloma Ballester, Vicent Caselles, and Laura Igual. A variational model for \\(\\mathrm{P + XS}\\) image fusion. IJCV, 69(1):43-58, 2006. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.603, + 0.484, + 0.657 + ], + "angle": 0, + "content": "[8] Wele Gedara Chaminda Bandara and Vishal M Patel. Hypertransformer: A textural and spectral feature fusion Transformer for pansharpening. In CVPR, pages 1767-1777, 2022, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.66, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[9] Qi Cao, Liang-Jian Deng, Wu Wang, Junming Hou, and Gemine Vivone. Zero-shot semi-supervised learning for pansharpening. Inf. Fusion., 101:102001, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.704, + 0.484, + 0.758 + ], + "angle": 0, + "content": "[10] Pat S. Chavez and Andy Yaw Kwarteng. Extracting spectral contrast in landsat thematic mapper image data using selective principal component analysis. Photogramm. Eng. Remote Sens., 55:339-348, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.484, + 0.814 + ], + "angle": 0, + "content": "[11] Lihui Chen, Gemine Vivone, Zihao Nie, Jocelyn Chanussot, and Xiaomin Yang. Spatial data augmentation: Improving the generalization of neural networks for pansharpening. IEEE TGRS, 61:1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[12] Zhi-Xuan Chen, Cheng Jin, Tian-Jing Zhang, Xiao Wu, and Liang-Jian Deng. SpanConv: A new convolution via spanning kernel space for lightweight pansharpening. In IJCAI, pages 841-847, 2022. Main Track. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[13] Minhyeok Choi, Rae Young Kim, and Myeong-Ryong Nam. Fusion of multispectral and panchromatic satellite images" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.192, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "using the curvelet transform. IEEE GRSL, 2(2):136-140, 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.908, + 0.178 + ], + "angle": 0, + "content": "[14] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE TGRS, pages 1-1, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.908, + 0.234 + ], + "angle": 0, + "content": "[15] Matteo Ciotola, Giovanni Poggi, and Giuseppe Scarpa. Unsupervised deep learning-based pansharpening with jointly enhanced spectral and spatial fidelity. IEEE TGRS, 61:1-17, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.238, + 0.908, + 0.292 + ], + "angle": 0, + "content": "[16] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pixel-wise ensembled masked autoencoder for multispectral pansharpening. IEEE TGRS, 62:1-22, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.907, + 0.351 + ], + "angle": 0, + "content": "[17] Yongchuan Cui, Peng Liu, Bingze Song, Lingjun Zhao, Yan Ma, and Lajiao Chen. Reconstruction of large-scale missing data in remote sensing images using Extend-GAN. IEEE GRSL, 21:1-5, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.354, + 0.907, + 0.408 + ], + "angle": 0, + "content": "[18] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pansharpening via predictive filtering with element-wise feature mixing. ISPRS J. Photogramm. Remote Sens., 219:22-37, 2025. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.907, + 0.466 + ], + "angle": 0, + "content": "[19] Liang-Jian Deng, Gemine Vivone, Cheng Jin, and Jocelyn Chanussot. Detail injection-based deep convolutional neural networks for pansharpening. IEEE TGRS, 59(8):6995-7010, 2021. 1, 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.908, + 0.537 + ], + "angle": 0, + "content": "[20] Liang-jian Deng, Gemine Vivone, Mercedes E. Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pansharpening: A benchmark, from shallow to deep networks. IEEE GRSM, 10(3):279-315, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.906, + 0.583 + ], + "angle": 0, + "content": "[21] Shangqi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, and Rui Wen. Bidirectional dilation Transformer for multispectral and hyperspectral image fusion. In IJCAI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.585, + 0.908, + 0.64 + ], + "angle": 0, + "content": "[22] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. PSRT: Pyramid shuffle-and-reshuffle Transformer for multispectral and hyperspectral image fusion. IEEE TGRS, 61:1-15, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.908, + 0.683 + ], + "angle": 0, + "content": "[23] Renwei Dian, Anjing Guo, and Shutao Li. Zero-shot hyperspectral sharpening. IEEE TPAMI, 45(10):12650-12666, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.686, + 0.908, + 0.769 + ], + "angle": 0, + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.908, + 0.814 + ], + "angle": 0, + "content": "[25] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In CVPR, pages 27738-27747, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.816, + 0.908, + 0.857 + ], + "angle": 0, + "content": "[26] Andrea Garzelli and Filippo Nencini. Wavelet-based fusion of optical and SAR image data in remote sensing applications. Int. J. Remote Sens., 23(23):5025-5039, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[27] Morteza Ghahremani and Hassan Ghassemian. Nonlinear IHS: A promising method for pan-sharpening. IEEE GRSL, 13(11):1606-1610, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11858" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.135 + ], + "angle": 0, + "content": "[28] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pyramid dual domain injection network for pan-sharpening. In ICCV, pages 12862-12871, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.485, + 0.19 + ], + "angle": 0, + "content": "[29] Xuanhua He, Keyu Yan, Jie Zhang, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Multiscale dual-domain guidance network for pan-sharpening. IEEE TGRS, 61:1-13, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[30] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. AAAI, 38(3):2121-2129, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.29 + ], + "angle": 0, + "content": "[31] Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. PanMamba: Effective pan-sharpening with state space model. Inf.Fusion., 115:102779, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.347 + ], + "angle": 0, + "content": "[32] Junming Hou, Xinyang Liu, Chenxu Wu, Xiaofeng Cong, Chentong Huang, Liang-Jian Deng, and Jian Wei You. Bidomain uncertainty gated recursive network for pan-sharpening. Inf. Fusion., 118:102938, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[33] Cheng Jin, Liang-Jian Deng, Ting-Zhu Huang, and Gemine Vivone. Laplacian pyramid networks: A new approach for multispectral pansharpening. Inf. Fusion., 78:158-170, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.404, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[34] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. LAGConv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. AAAI, 36(1):1113-1121, 2022. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[36] Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepcev, and Gustavo K. Rohde. Optimal mass transport: Signal processing and machine-learning applications. IEEE SPM, 34 (4):43-59, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[37] Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo K. Rohde. Generalized sliced wasserstein distances, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.603, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[38] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. 2000. US Patent 6,011,875. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[39] Zhixuan Li, Jinjiang Li, Fan Zhang, and Linwei Fan. CADUI: Cross-attention-based depth unfolding iteration network for pansharpening remote sensing images. IEEE TGRS, 61:1–20, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.484, + 0.771 + ], + "angle": 0, + "content": "[40] YuJie Liang, ZiHan Cao, Shangqi Deng, Hong-Xia Dou, and Liang-Jian Deng. Fourier-enhanced implicit neural fusion network for multispectral and hyperspectral image fusion. In NeurIPS, pages 63441-63465. Curran Associates, Inc., 2024. 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.773, + 0.484, + 0.815 + ], + "angle": 0, + "content": "[41] J. G. Liu. Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details. Int. J. Remote Sens., 21(18):3461-3472, 2000. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[42] Peng Liu, Jun Li, Lizhe Wang, and Guojin He. Remote sensing data fusion with generative adversarial networks: State-of-the-art methods and future research directions. IEEE GRSM, 10(2):295-328, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.872, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[43] Qingjie Liu, Huanyu Zhou, Qizhi Xu, Xiangyu Liu, and Yun-hong Wang. PSGAN: A generative adversarial network for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.12 + ], + "angle": 0, + "content": "remote sensing image pan-sharpening. IEEE TGRS, 59(12): 10227-10242, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.178 + ], + "angle": 0, + "content": "[44] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-GAN: a multiscale unsupervised network for remote sensing image pansharpening. IEEE TGRS, 61:1-18, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.18, + 0.907, + 0.235 + ], + "angle": 0, + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In ICCV, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.292 + ], + "angle": 0, + "content": "[46] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-GAN: An unsupervised pan-sharpening method for remote sensing image fusion. Inf. Fusion., 62: 110-120, 2020. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.907, + 0.337 + ], + "angle": 0, + "content": "[47] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. In Remote Sens., page 594, 2016. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.338, + 0.907, + 0.393 + ], + "angle": 0, + "content": "[48] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. PanDiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE TGRS, 61:1-17, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.395, + 0.907, + 0.465 + ], + "angle": 0, + "content": "[49] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE GRSM, 9(1):18-52, 2021. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.467, + 0.907, + 0.507 + ], + "angle": 0, + "content": "[50] Xiangchao Meng, Nan Wang, Feng Shao, and Shutao Li. Vision Transformer for pansharpening. IEEE TGRS, 60:1-11, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.509, + 0.907, + 0.551 + ], + "angle": 0, + "content": "[51] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pages 666-704, 1781. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.907, + 0.595 + ], + "angle": 0, + "content": "[52] Fjola Palsson, Johannes R. Sveinsson, and Magnus O. Ulfarsson. A new pansharpening algorithm based on total variation. IEEE GRSL, 11(1):318-322, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.597, + 0.907, + 0.64 + ], + "angle": 0, + "content": "[53] Siran Peng, Xiangyu Zhu, Haoyu Deng, Liang-Jian Deng, and Zhen Lei. FusionMamba: Efficient remote sensing image fusion with state space model. IEEE TGRS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.641, + 0.907, + 0.682 + ], + "angle": 0, + "content": "[54] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Found. Trends Mach. Learn., 11(5-6):355-607, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.684, + 0.907, + 0.726 + ], + "angle": 0, + "content": "[55] Sheila Rahmani, Melissa Strait, Daria Merkurjev, Michael Moeller, and Todd Wittman. An adaptive IHS pan-sharpening method. IEEE GRSL, 7(4):746-750, 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.728, + 0.907, + 0.77 + ], + "angle": 0, + "content": "[56] Aaditya Ramdas, Nicolas Garcia, and Marco Cuturi. On Wasserstein two sample testing and related families of nonparametric tests, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.772, + 0.907, + 0.813 + ], + "angle": 0, + "content": "[57] Xiangyu Rui, Xiangyong Cao, Yining Li, and Deyu Meng. Variational zero-shot multispectral pansharpening. IEEE TGRS, 62:1-16, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.814, + 0.907, + 0.87 + ], + "angle": 0, + "content": "[58] Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, and Deyu Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion., 107:102325, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.907, + 0.902 + ], + "angle": 0, + "content": "[59] Filippo Santambrogio. One-dimensional issues, pages 59-85. Springer International Publishing, Cham, 2015. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11859" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[60] Giuseppe Scarpa and Matteo Ciotola. Full-resolution quality assessment for pansharpening. Remote Sens., 14(8), 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[61] Vijay P. Shah, Nicolas H. Younan, and Roger L. King. An efficient pan-sharpening method via a combined adaptive PCA approach and contourlets. IEEE TGRS, 46(5):1323-1335, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.217 + ], + "angle": 0, + "content": "[62] Xunyang Su, Jinjiang Li, and Zhen Hua. Transformer-based regression network for pansharpening remote sensing images. IEEE TGRS, 60:1-23, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[63] Weiwei Sun, Jun Zhou, Xiangchao Meng, Gang Yang, Kai Ren, and Jiangtao Peng. Coupled temporal variation information estimation and resolution enhancement for remote sensing spatial-temporal-spectral fusion. IEEE TGRS, 61: 1-18, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.484, + 0.328 + ], + "angle": 0, + "content": "[64] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[66] Cédric Villani et al. Optimal transport: old and new. Springer, 2008. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[67] Gemine Vivone, Miguel Simões, Mauro Dalla Mura, Rocco Restaino, José M. Bioucas-Dias, Giorgio A. Licciardi, and Jocelyn Chanussot. Pansharpening based on semiblind deconvolution. IEEE TGRS, 53(4):1997-2010, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.457, + 0.484, + 0.538 + ], + "angle": 0, + "content": "[68] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O. Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE GRSM, 9(1):53-81, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.484, + 0.594 + ], + "angle": 0, + "content": "[69] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. *Photogramm. Eng. Remote Sens.*, 63:691-699, 1997. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.484, + 0.636 + ], + "angle": 0, + "content": "[70] Hebaixu Wang, Meiqi Gong, Xiaoguang Mei, Hao Zhang, and Jiayi Ma. Deep unfolded network with intrinsic supervision for pan-sharpening. AAAI, 38(6):5419-5426, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.484, + 0.691 + ], + "angle": 0, + "content": "[71] Hebaixu Wang, Hao Zhang, Xin Tian, and Jiayi Ma. Zero-Sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation. Inf. Fusion., 101:102003, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.484, + 0.747 + ], + "angle": 0, + "content": "[72] Jun-Da Wang, Liang-Jian Deng, Chen-Yu Zhao, Xiao Wu, Hong-Ming Chen, and Gemine Vivone. Cascadic multireceptive learning for multispectral pansharpening. IEEE TGRS, 61:1-16, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[73] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE TGRS, 62:1-16, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.845 + ], + "angle": 0, + "content": "[74] Zhong-Cheng Wu, Ting-Zhu Huang, Liang-Jian Deng, JinFan Hu, and Gemine Vivone. \\(\\mathrm{VO + }\\) Net: An adaptive approach using variational optimization and deep learning for panchromatic sharpening. IEEE TGRS, 60:1-16, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[75] Jin-Liang Xiao, Ting-Zhu Huang, Liang-Jian Deng, Zhong-Cheng Wu, Xiao Wu, and Gemine Vivone. Variational pan-sharpening based on coefficient estimation with nonlocal regression. IEEE TGRS, 61:1-15, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.134 + ], + "angle": 0, + "content": "[76] Guangxu Xie, Rencan Nie, Jinde Cao, He Li, and Jintao Li. A deep multiresolution representation framework for pansharpening. IEEE TGRS, 62:1-16, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[77] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xi-wei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE TGRS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[78] Yinghui Xing, Litao Qu, Shizhou Zhang, Kai Zhang, Yanning Zhang, and Lorenzo Bruzzone. CrossDiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE TIP, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.907, + 0.303 + ], + "angle": 0, + "content": "[79] Qizhi Xu, Yuan Li, Jinyan Nie, Qingjie Liu, and Mengyao Guo. UPanGAN: Unsupervised pansharpening based on the spectral and spatial loss constrained generative adversarial network. Inf. Fusion., 91:31-46, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.906, + 0.347 + ], + "angle": 0, + "content": "[80] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In CVPR, pages 1366-1375, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[81] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pansharpening. In CVPR, pages 1778-1787, 2022. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.907, + 0.447 + ], + "angle": 0, + "content": "[82] Jie Yang, Xiao-Yang Fu, Yan Hu, Yue Huang, Xing Ding, and John Paisley. PanNet: A deep network architecture for pan-sharpening. ICCV, pages 1753-1761, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[83] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE JSTARS, 11(3):978-989, 2018. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.907, + 0.559 + ], + "angle": 0, + "content": "[84] Jie Zhang, Xuanhua He, Keyu Yan, Ke Cao, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Pan-sharpening with wavelet-enhanced high-frequency information. IEEE TGRS, 62:1-14, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.907, + 0.616 + ], + "angle": 0, + "content": "[85] Chen-Yu Zhao, Tian-Jing Zhang, Ran Ran, Zhi-Xuan Chen, and Liang-Jian Deng. LGPCConv: Learnable gaussian perturbation convolution for lightweight pansharpening. In IJCAI, pages 4647-4655, 2023. Main Track. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.907, + 0.671 + ], + "angle": 0, + "content": "[86] Mengjiao Zhao, Mengting Ma, Xiangdong Li, Xiaowen Ma, Wei Zhang, and Siyang Song. Supervised detail-guided multiscale state space model for pan-sharpening. IEEE TGRS, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.907, + 0.715 + ], + "angle": 0, + "content": "[87] Kaiwen Zheng, Jie Huang, Man Zhou, Danfeng Hong, and Feng Zhao. Deep adaptive pansharpening via uncertainty-aware image fusion. IEEE TGRS, 61:1-15, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.907, + 0.771 + ], + "angle": 0, + "content": "[88] Yu Zhong, Xiao Wu, Liang-Jian Deng, ZiHan Cao, and Hong-Xia Dou. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pansharpening. In NeurIPS, pages 77962-77986. Curran Associates, Inc., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.906, + 0.829 + ], + "angle": 0, + "content": "[89] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In ECCV, pages 274-291, Cham, 2022. Springer Nature Switzerland. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.907, + 0.886 + ], + "angle": 0, + "content": "[90] Man Zhou, Jie Huang, Keyu Yan, Danfeng Hong, Xiuping Jia, Jocelyn Chanussot, and Chongyi Li. A general spatial-frequency learning framework for multimodal image fusion. IEEE TPAMI, pages 1-18, 2024. 3, 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.886 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11860" + } + ] +] \ No newline at end of file diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_origin.pdf b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6719d0bf5ca157c27d17a9899791bf4f3bad63c5 --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/84a6c0dd-500e-40a7-b080-833571baa669_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3db67a8cbd4f66ac347abdcb8a03e3fe0809238c28aebd4f3f9e27ea17eaf47b +size 22178259 diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/full.md b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/full.md new file mode 100644 index 0000000000000000000000000000000000000000..648bd8041e6576f4da55ba56308778a8e3df0172 --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/full.md @@ -0,0 +1,366 @@ +# Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need + +Yongchuan Cui $^{1,2}$ Peng Liu $^{1,2,*}$ Hui Zhang $^{3,*}$ + +1Aerospace Information Research Institute, Chinese Academy of Sciences, 100094 Beijing, China + +$^{2}$ School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, 101408 Beijing, China + +$^{3}$ School of Engineering Medicine, Beihang University, 100191 Beijing, China + +yongchuancui@gmail.com, liupeng202303@aircas.ac.cn, hui.zhang@buaa.edu.cn + +# Abstract + +Existing deep learning-based models for remote sensing pansharpening exhibit exceptional performance on training datasets. However, due to sensor-specific characteristics and varying imaging conditions, these models suffer from substantial performance degradation when applied to unseen satellite data, lacking generalizability and thus limiting their applicability. We argue that the performance drops stem primarily from distributional discrepancies from different sources and the key to addressing this challenge lies in bridging the gap between training and testing distributions. To validate the idea and further achieve a "train once, deploy forever" capability, this paper introduces a novel and intuitive approach to empower any pansharpening models with generalizability by employing a unified distribution strategy (UniPAN). Specifically, we construct a distribution transformation function that normalizes the pixels sampled from different sources to conform to an identical distribution. The deep models are trained on the transformed domain, and during testing on new datasets, the new data are also transformed to match the training distribution. UniPAN aims to train and test the model on a unified and consistent distribution, thereby enhancing its generalizability. Extensive experiments validate the efficacy of UniPAN, demonstrating its potential to significantly enhance the performance of deep pansharpening models across diverse satellite sensors. + +Codes: https://github.com/yc-cui/UniPAN. + +# 1. Introduction + +Due to the inherent limitations of sensor imaging technology, advanced remote sensing satellites such as IKONOS, QuickBird, GaoFen-2, and WorldView-3, necessitate a compromise between the acquisition of extensive spectral infor + +![](images/16bf4c0c34df7e504d3392c3eea04942c9c8ef0216e1716cf0f703efef000caf.jpg) +Figure 1. Comparison of generalization metrics w/ and w/o UniPAN trained on WorldView-3 (WV3) using FusionNet [19]. + +![](images/940d3e0128c46e96e9a5ee907ff658059987a3ffb9243dadf7b3555aebc93195.jpg) + +![](images/cfed76fb5cdad7a1cb0c3d41934aac9950234a70a1df91b45dc495f5dcc52f03.jpg) + +![](images/863f9367fcf43cf5ac9450f61a583ec0381744051b44eaeca8811c755dfcd24c.jpg) +Figure 2. Comparison of tSNE visualization w/ and w/o UniPAN. + +mation and maintaining fine spatial details in their captured imagery [20, 49, 68]. Consequently, these satellites typically capture either high spatial resolution single-band panchromatic (PAN) images or low spatial resolution multispectral (LRMS) images. Pansharpening is a crucial technique in remote sensing image processing, aiming to combine PAN and LRMS images to generate fused images with high spatial resolution and abundant spectra (HRMS). + +Traditional pansharpening methods have been extensively studied over the past few decades. However, these + +methods may introduce spectral distortions, require significant computational resources, and are sensitive to the quality of input data [20, 42, 49, 68]. In contrast, deep learning (DL)-based methods have shown great potential in pansharpening. Notably, various innovative architectures have emerged, starting from early convolutional models [19, 47, 82, 83], progressing to Transformer-based approaches [8, 21, 22, 24, 50, 62, 65], then generative methods [43, 46, 48, 58, 78], and continuing with the emergence of the latest novel designs [9, 23, 31, 53, 86]. The networks are evolving towards larger parameters capacity, stronger representation ability, and increasingly complex and diverse structures. + +However, as deep pansharpening thrives, the overly powerful fitting capabilities of these networks have concurrently presented several ongoing concerns, e.g. generalizability, interpretability, over-reliance on high-quality data, etc. This paper specifically delves into the widely criticized issue of generalizability. Specifically, although deep models achieve excellent performance on training datasets, they often fail to generalize well to new data due to distribution shift such as different noise levels, varying spectral signatures and sensor conditioning characteristics. While models can achieve satisfactory performances on new datasets through retraining or fine-tuning, the process is time-consuming and fine-tuning on every new dataset is impractical and cumbersome. The ideal scenario would involve developing a pansharpening model that can be trained once and deployed effectively across diverse datasets and sensors, thereby providing a train once, deploy forever capability. + +To this end, this paper introduces a novel approach that aims to empower any pansharpening models with generalizability by employing a unified distribution strategy (dubbed as UniPAN). The proposed method aims to bridge the gap between training and testing distributions, thereby improving the model's ability to generalize to unseen data. The core idea is to normalize the distribution of spatial pixels from the training dataset to conform to a predefined distribution (e.g. Gaussian distribution) and deep models are then trained on the transformed domain. This process is model-agnostic and requires no additional trainable parameters. When generalizing to new data, testing datasets are also transformed to match the training distribution, ensuring that the models are evaluated under the same distributional assumptions. Fig. 1 presents the cross-satellite generalization performance of a well-known pansharpening model FusionNet [19] on NBUPansharpRSData [49] trained w/ and w/o UniPAN. The model utilizing UniPAN demonstrates a significant improvement on full-resolution metrics, which validates the strong generalization ability for real-world scenes. To further elucidate the efficacy of UniPAN, Fig. 2 presents tSNE [64] dimensionality reduction clustering of multispectral images from diverse satellites using NBUPansharpRSData [49] w/ + +and w/o UniPAN. The pre-UniPAN visualization reveals distinct clustering patterns, with data points from different satellites forming separate groups, indicative of substantial inter-satellite distributional disparities. In contrast, following the application of UniPAN, a notable transformation is observed: the previously separated clusters become substantially intermingled, demonstrating a significant reduction in distributional discrepancies across different satellite datasets (e.g. the purple scatters of QuickBird). This visual evidence underscores UniPAN's capability to harmonize heterogeneous satellite data, effectively creating a more cohesive and unified representation. + +The contributions of this paper could be summarized as follows: + +- We propose UniPAN, a model-agnostic approach that employs a unified distribution strategy to empower any deep pansharpening models with generalizability, achieving a train once, deploy forever capability. This strategy provides two advantages: 1) Architectural independence: Compatible with any existing or future network architecture without structural modifications. 2) Training efficiency: Eliminates the need for retraining or fine-tuning when adapting to new sensors. +- UniPAN is implemented via inverse transform sampling, and we prove its equivalence to the optimal solution of optimal transport in the one-dimensional case. UniPAN requires no additional trainable parameters and only needs to be fitted once before training begins, introducing almost no additional computational overhead. It serves as a viable alternative to vanilla data normalization techniques such as min-max and $z$ -score normalization, offering a more robust and unified approach to data distribution alignment. +- We conduct extensive experiments using a variety of deep pansharpening models across numerous satellites. The simple yet intuitive and effective strategy provides a universal solution to the generalization challenge in DL-based pansharpening and experimental results demonstrate its effectiveness, showcasing significant improvements in cross-dataset generalization across a wide range of models and satellite sensors. + +# 2. Related Work + +# 2.1. Pansharpening Methods + +Traditional Methods. Traditional pansharpening methods can be broadly categorized into three paradigms: component substitution (CS), multi-resolution analysis (MRA), and variational optimization (VO) [20, 42, 49, 68]. CS-based methods [4, 10, 27, 38, 55, 61] operate by substituting the spatial component of the LRMS image with the PAN image. Representative algorithms include Gram-Schmidt (GS) [4, 38], principal component analysis (PCA) [10, 61], and intensity-hue-saturation (IHS) [27, 55]. While computationally effi + +cient, CS methods often introduce spectral distortions due to mismatches between PAN and substituted components. MRA-based methods [1-3, 13, 26, 41] employ multi-scale decomposition tools (e.g. wavelets [26], contourlets [13]) to inject high-frequency details from PAN into LRMS. Typical implementations include MTF-GLP [1-3] and SFIM [41]. These methods generally preserve spectral fidelity better than CS approaches but may exhibit limited spatial enhancement. VO-based approaches [7, 52, 63, 67, 74, 75] formulate pansharpening as an optimization problem with handcrafted priors, such as the variational wavelet-based methods [7] and Bayesian approaches [52]. Despite their mathematical elegance, traditional methods often require manual parameter tuning and struggle with complex spatial-spectral relationships [20, 68]. + +Deep Learning-Based Methods. The advent of deep learning (DL) have revolutionized pansharpening by enabling end-to-end learning of spatial-spectral relationships. Early pioneering convolutional neural networks (CNNs) such as PNN [47], PanNet [82], and MSDCNN [83] established the feasibility of learning fusion rules from a mount of simulated data. Subsequent innovations introduced multi-scale architectures [33, 72, 76], attention mechanisms [39, 73], and kernel customizing [12, 18, 25, 34, 85], etc. The CNN-based methods demonstrated superior performance over traditional approaches but faced limitations in modeling long-range dependencies. In contrast, Transformer [24, 45, 65] addressed this limitation with self-attention mechanisms. Although various Transformer architectures specifically designed for pansharpening [8, 21, 22, 50, 62] have achieved impressive performance, they face criticism for quadratic computational complexity w.r.t. sequence length [16]. To address this limitation, recent studies have explored linear-complexity models [16, 31, 53, 86] that maintain performance while improving efficiency. To further enhance fidelity preservation, generative models like generative adversarial networks (GANs) [17, 43, 44, 46, 79] and diffusion models [48, 58, 77, 78, 88] have emerged as promising solutions. Beyond conventional configurations, contemporary research has expanded into more challenging frontiers to address practical constraints, including zero-shot learning paradigms [9, 23, 57, 71], frequency domain integration [30, 40, 84, 89, 90], and interpretable fusion mechanisms [32, 70, 80, 81]. + +# 2.2. Generalization in DL-Based Pansharpening + +Despite remarkable progress in DL-based pansharpening, generalization capability remains a persistent challenge. The issue manifests conspicuously in two critical scenarios: 1) Models trained under the Wald protocol [69] using spatially degraded data often fail to maintain performance when applied to full-resolution inference, frequently exhibiting severe spectral distortions. 2) Networks optimized for specific + +satellite sensors demonstrate substantial performance degradation when deployed on unseen satellite platforms due to distribution shifts in spectral signatures and sensor characteristics. Existing approaches to address these challenges primarily focus on unsupervised learning [14, 15, 46], generalized module design [77], or advanced data augmentation [11]. However, such methods typically require either meticulous architectural redesign or computationally intensive retraining/fine-tuning procedures. These requirements pose significant practical limitations for real-world deployment. + +In contrast to model-centric solutions, our approach adopts a fundamentally different perspective by addressing the generalization challenge through data-space intervention. The proposed UniPAN framework introduces a model-agnostic preprocessing technique that operates on raw sensor data. Compared to traditional linear data normalization such as min-max scaling or $z$ -score standardization, our method employs a non-linear mapping to project the input data onto a unified distribution, effectively harmonizing the distribution mismatches from all sensors. + +# 3. Method + +# 3.1. Distribution Transformation + +The core challenge in cross-domain pansharpening stems from the distribution discrepancy between source (training) and target (testing) domains. Let $\mathcal{X}_S$ denote the source domain (training data) and $\mathcal{X}_T$ the target domain (testing data). Our objective is to learn a transformation $\mathcal{T}:\mathcal{X}\to \mathcal{Z}$ that maps both domains to a shared space $\mathcal{Z}$ with unified distribution $\rho_t$ , e.g. standard normal distribution. + +Inverse Transform Sampling. Given a univariate r.v. $X \sim \rho_s$ , we seek a transformation $\mathcal{T} : \mathbb{R} \to \mathbb{R}$ such that $Z = \mathcal{T}(X) \sim \rho_t$ . The inverse transform sampling theorem provides a theoretically grounded solution: + +$$ +\mathcal {T} (x) = F _ {T} ^ {- 1} \left(F _ {S} (x)\right), \tag {1} +$$ + +where $F_{S}(x) = \rho_{s}(-\infty, x]$ is the cumulative distribution function (CDF) of $X$ , and $F_{T}^{-1}(u) = \inf \{z \in \mathbb{R} : F_{T}(z) \geq u\}$ is the inverse CDF of the target distribution. By the universality of the uniform, $U = F_{S}(X)$ follows a uniform distribution, and $Z = F_{T}^{-1}(U)$ inherits $\rho_{t}$ exactly when $F_{S}$ is continuous. To provide more intuitively analyze of the algorithm's operational process and efficiency, we establish a fundamental connection between inverse transform sampling and the well-studied optimal transport theory [51, 54]. + +Optimal Transport Perspective. Given two probability distributions $\rho_{s}$ (source domain) and $\rho_{t}$ (target domain) defined on $\mathbb{R}^n$ and $\mathbb{R}^m$ respectively, we consider the Monge problem [51, 54]: + +$$ +\mathcal {C} (\rho_ {s}, \rho_ {t}) \triangleq \min _ { \begin{array}{c} T: \mathbb {R} ^ {n} \to \mathbb {R} ^ {m} \\ T _ {\sharp} \rho_ {s} = \rho_ {t} \end{array} } \int_ {\mathbb {R} ^ {n}} c (x, T (x)) \rho_ {s} (x) d x, \quad (2) +$$ + +where $c:\Omega_s\times \Omega_t\to \mathbb{R}^+$ is a cost function and for any measurable Borel set $E\in \Omega_s$ $T:\Omega_s\rightarrow \Omega_t$ transfers measures from one space $\Omega_{s}$ to another space $\Omega_t$ + +$$ +T _ {\sharp} \rho_ {s} (E) = \rho_ {s} \left(T ^ {- 1} (E)\right), \tag {3} +$$ + +here $T_{\sharp}$ is the pushforward operator. The Monge problem seeks a transport plan $T$ from $\rho_s$ to $\rho_t$ that minimizes the transportation cost. The optimal $T^{*}$ is also known as the Monge map of Eq. (2). + +Let $F_{S}(x) = \rho_{s}(-\infty, x]$ and $F_{T}(z) = \rho_{t}(-\infty, z]$ denote the CDFs of the source and target distributions respectively. In the univariate setting ( $n = m = 1, \Omega_{s} = \Omega_{t} = \mathbb{R}$ ), when the source measure $\rho_{s}$ is continuous and the cost function takes the form $c(x, z) = h(|x - z|)$ with $h$ being strictly convex, by cyclical monotonicity [66], the Monge map $T^{*}$ satisfies: + +$$ +F _ {S} (x) = F _ {T} \left(T ^ {*} (x)\right), \tag {4} +$$ + +yielding the explicit expression: + +$$ +T ^ {*} (x) = F _ {T} ^ {- 1} \left(F _ {S} (x)\right). \tag {5} +$$ + +This remarkable result coincides exactly with the inverse transform sampling formula Eq. (1), i.e. $T^{*} = \mathcal{T}$ . The transport cost in this setting [37, 56, 59] can be computed as + +$$ +\mathcal {C} \left(\rho_ {s}, \rho_ {t}\right) = \int_ {0} ^ {1} c \left(F _ {S} ^ {- 1} (u), F _ {T} ^ {- 1} (u)\right) d u. \tag {6} +$$ + +For the $L_{1}$ cost function $c(x,z) = |x - z|$ , this corresponds to the $W_{1}$ Wasserstein distance [36, 59] between the distributions. In the one-dimensional (1D) scenario, UniPAN operates independently on spatial pixels within each spectral band. This implementation can be achieved through interpolation-based inverse transform sampling without framing it as an optimization problem. However, given the inherent correlations between spectral bands and interdependencies among spatial patches, multidimensional extension of this framework may be promising, albeit with increased implementation complexity. The theoretical derivation establishing connections with optimal transport theory serves to facilitate future generalization of UniPAN to multidimensional configurations. Building upon the theoretical foundation of distribution transformation through optimal transport, in the following, we present a practical implementation of 1D UniPAN. + +# 3.2. UniPAN Implementation + +We present a practical implementation for 1D finite discrete distributions of UniPAN through a quantile-based distribution transformation algorithm. + +Finite Discrete Distribution Transformation. Given pixel values $X = \{x_{1}, x_{2}, \dots, x_{n}\}$ sampled from a spectral band, + +Algorithm 1 UniPAN Training and Testing Procedure +Training Phase +Input: Training dataset $\mathcal{D}_{\mathrm{train}} = \{L_i, P_i, H_i\}_{i=1}^N$ , inverse CDF $F_T^{-1}$ of target distribution $\rho_t$ , model $\mathcal{M}$ +Output: Trained model $M$ +1: Initialize transformed LRMS $\tilde{L}$ and PAN $\tilde{P}$ +2: for each band $b$ of LRMS $\{L_i\}_{i=1}^N$ & PAN $\{P_i\}_{i=1}^N$ do +3: Sample $n$ pixels $X_b \gets \{x_1, \ldots, x_n\}$ from band $b$ +4: Compute $m$ quantiles $Q = \{q_0, \ldots, q_{m-1}\} \triangleright$ Eq. (7) +5: Linspace uniform CDF $U = \{0, \frac{1}{m-1}, \ldots, 1\}$ +6: for each pixel $x$ in band $b$ do +7: Forward and backward interp by $u_\uparrow$ and $u_\downarrow \triangleright$ Eq. (9) +8: Compute eCDF value $u(x) \gets (u_\uparrow + u_\downarrow) / 2 \triangleright$ Eq. (8) +9: Transform $x$ to $z \gets F_T^{-1}(u(x))$ and assign $z$ to corresponding band of $\tilde{L}$ and $\tilde{P}$ +10: end for +11: end for +12: while not converged do +13: Train $\mathcal{M}$ with $\mathcal{L} = \|H - \mathcal{M}(\tilde{L}, \tilde{P}) + L \uparrow\|_1$ +14: end while +15: return Trained model $\mathcal{M}$ + +Testing Phase +Input: Testing dataset $\mathcal{D}_{\mathrm{test}} = \{L_i,P_i\}_{i = 1}^N$ , inverse CDF $F_{T}^{-1}$ of target distribution $\rho_{t}$ , trained model $\mathcal{M}$ +Output: Pansharpening result $\hat{H}^{\mathrm{test}}$ +16: for each band $b$ of LRMS & PAN do +17: Transform $L,P$ to $\tilde{L},\tilde{P}$ using the training transformation +18: end for +19: Generate $\hat{H}^{\mathrm{test}}\gets \mathcal{M}(\tilde{L},\tilde{P}) + L\uparrow$ +20: return Fused result $\hat{H}^{\mathrm{test}}$ + +we first compute $m$ empirical quantiles $Q = \{q_{0},\dots,q_{m - 1}\}$ where: + +$$ +q _ {i} = \inf \left\{x \in X \mid P (X \leq x) \geq \frac {i}{m - 1} \right\}, i = \{0, 1, \dots , m - 1 \}. \tag {7} +$$ + +These quantiles discretize the empirical cumulative distribution function (eCDF) of the source data. The number of quantiles $m$ controls the approximation fidelity. + +Given any data point $x$ to be transformed, the next step involves determining its value in the eCDF through interpolation. However, when the source distribution contains repeated values or sparse regions, quantile estimation may yield non-unique or overlapping quantile boundaries. This can result in abrupt jumps in the eCDF, thereby violating the properties of monotonicity and smoothness. Therefore, we use bidirectional interpolation to estimate a robust eCDF value of $x$ : + +$$ +\begin{array}{l} u (x) = u _ {\uparrow} (x) / 2 + u _ {\downarrow} (x) / 2 \\ = \operatorname {i n t e r p} (x, Q, U) / 2 + \tag {8} \\ (1 - \text {i n t e r p} (- x, - Q ^ {\text {r e v}}, - U ^ {\text {r e v}})) / 2, \\ \end{array} +$$ + +where $U = \{0,\frac{1}{m - 1},\dots,1\}$ , $Q^{\mathrm{rev}}$ is the reversed quantiles. + +
QNRTρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.980(+0.005)0.946(-0.010)0.971(+0.000)0.963(-0.003)0.964(-0.004)0.974(-0.000)0.985(+0.010)0.961(+0.004)0.979(+0.007)0.984(+0.017)0.982(+0.013)0.987(+0.013)
IK0.950(+0.005)0.912(+0.039)0.915(+0.045)0.924(+0.011)0.916(+0.022)0.958(+0.011)0.956(+0.011)0.919(+0.045)0.940(+0.070)0.950(+0.037)0.945(+0.051)0.968(+0.021)
QB0.957(+0.006)0.936(+0.016)0.931(+0.047)0.953(+0.017)0.944(+0.014)0.970(+0.016)0.965(+0.014)0.944(+0.024)0.966(+0.082)0.969(+0.032)0.969(+0.038)0.970(+0.017)
WV20.880(+0.026)0.811(+0.029)0.832(+0.062)0.946(+0.008)0.916(+0.012)0.951(+0.012)0.872(+0.018)0.775(-0.006)0.825(+0.055)0.944(+0.006)0.914(+0.011)0.940(+0.001)
WV30.862(-0.006)0.804(+0.035)0.795(+0.067)0.921(-0.011)0.919(+0.011)0.950(+0.001)0.880(+0.011)0.805(+0.036)0.813(+0.084)0.943(+0.010)0.924(+0.016)0.952(+0.003)
WV40.924(+0.009)0.898(+0.045)0.875(+0.036)0.948(-0.000)0.929(+0.003)0.972(+0.006)0.939(+0.024)0.889(+0.035)0.903(+0.064)0.952(+0.002)0.935(+0.009)0.973(+0.007)
Dλ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.007(-0.003)0.022(+0.008)0.005(-0.000)0.017(+0.004)0.015(+0.005)0.011(+0.000)0.004(-0.005)0.012(-0.001)0.004(-0.001)0.004(-0.008)0.004(-0.005)0.004(-0.005)
IK0.011(-0.000)0.020(-0.015)0.018(-0.011)0.031(+0.004)0.022(-0.003)0.014(-0.003)0.008(-0.003)0.020(-0.015)0.013(-0.016)0.013(-0.014)0.009(-0.016)0.010(-0.008)
QB0.013(-0.000)0.011(-0.004)0.015(-0.010)0.016(-0.004)0.014(+0.000)0.010(-0.005)0.009(-0.003)0.009(-0.006)0.007(-0.017)0.008(-0.013)0.006(-0.007)0.012(-0.003)
WV20.042(-0.006)0.066(-0.007)0.037(-0.019)0.008(-0.000)0.014(-0.000)0.016(-0.002)0.046(-0.002)0.086(+0.012)0.048(-0.008)0.008(+0.000)0.015(+0.001)0.022(+0.002)
WV30.050(+0.010)0.067(-0.011)0.049(-0.024)0.027(+0.011)0.014(+0.000)0.016(+0.002)0.042(+0.002)0.072(-0.006)0.051(-0.022)0.013(-0.002)0.011(-0.002)0.014(+0.000)
WV40.024(-0.000)0.031(-0.012)0.029(-0.009)0.017(+0.008)0.018(+0.008)0.005(-0.001)0.016(-0.008)0.032(-0.011)0.022(-0.016)0.014(+0.004)0.015(+0.005)0.004(-0.002)
DS↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.012(-0.003)0.032(+0.003)0.023(+0.000)0.020(+0.000)0.021(+0.000)0.015(-0.000)0.010(-0.005)0.026(-0.002)0.016(-0.006)0.010(-0.009)0.013(-0.007)0.007(-0.008)
IK0.039(-0.005)0.068(-0.026)0.068(-0.036)0.046(-0.014)0.063(-0.019)0.027(-0.007)0.035(-0.008)0.062(-0.033)0.046(-0.058)0.037(-0.024)0.045(-0.036)0.022(-0.013)
QB0.030(-0.005)0.053(-0.012)0.054(-0.039)0.031(-0.012)0.042(-0.014)0.020(-0.012)0.025(-0.011)0.046(-0.018)0.026(-0.067)0.023(-0.020)0.024(-0.031)0.020(-0.012)
WV20.081(-0.022)0.134(-0.024)0.137(-0.049)0.045(-0.008)0.071(-0.012)0.032(-0.010)0.085(-0.018)0.153(-0.005)0.133(-0.053)0.047(-0.006)0.071(-0.012)0.039(-0.003)
WV30.096(-0.000)0.141(-0.027)0.167(-0.050)0.054(+0.002)0.067(-0.011)0.033(-0.004)0.083(-0.013)0.134(-0.033)0.146(-0.072)0.043(-0.008)0.064(-0.014)0.033(-0.003)
WV40.053(-0.009)0.073(-0.036)0.099(-0.029)0.035(-0.006)0.053(-0.011)0.021(-0.005)0.045(-0.018)0.082(-0.027)0.076(-0.052)0.036(-0.004)0.052(-0.012)0.021(-0.005)
Dρ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.043(+0.003)0.110(+0.030)0.026(-0.008)0.079(+0.032)0.062(+0.015)0.041(-0.002)0.019(-0.020)0.059(-0.021)0.022(-0.013)0.015(-0.032)0.017(-0.029)0.017(-0.026)
IK0.082(-0.006)0.095(-0.038)0.090(-0.058)0.126(+0.024)0.110(-0.010)0.061(-0.008)0.064(-0.024)0.092(-0.040)0.066(-0.083)0.051(-0.049)0.061(-0.059)0.045(-0.024)
QB0.061(-0.003)0.154(-0.020)0.179(-0.059)0.135(+0.011)-0.133(-0.045)-0.093(-0.026)-0.164(-0.029)-0.235(-0.002)-0.200(-0.058)-0.100(-0.013)-0.126(-0.048)-0.112(-0.007)
WV2-1.166(-0.27)-1.198(-0.339)-1.199(-0.539)-1.155(-0.459)-1.133(-0.455)-1.137(-0.451)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)
WV3-2.205(+0.18)-2.214(-0.41)-2.234(-0.533)-2.231(-0.419)-2.173(+0.11)-2.173(-0.415)-2.176(-0.419)-2.226(-0.429)-2.219(-0.468)-2.117(-0.434)-1.131(-0.439)-1.131(-0.439)
WV4-3.155(-0.55)-3.112(-0.42)-3.115(-0.44)-3.146(+0.88)-3.169(+0.77)-3.169(-0.477)-3.173(-0.477)-3.176(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)
+ +Table 1. Average results of 12 models. Rows correspond to training satellites, and columns correspond to testing satellites. Green indicates performance improvement, while red indicates performance degradation. The values in parentheses represent the average delta compared to the baseline without UniPAN. + +The function $\text{interp}$ represents a linear interpolation between the quantile point $Q$ and the reference point $U$ , with the interpolation formula given by: + +$$ +\operatorname {i n t e r p} (x, Q, U) = \left\{ \begin{array}{l l} U [ 0 ], & x \leq Q [ 0 ] \\ f (x, i, Q, U), & Q [ i ] < x \leq Q [ i + 1 ] \\ U [ m - 1 ], & x > Q [ m - 1 ] \end{array} \right. \tag {9} +$$ + +where $f$ computes the $Q$ to $U$ mapping, + +$$ +f (x, i, Q, U) = U [ i ] + \frac {x - Q [ i ]}{Q [ i + 1 ] - Q [ i ]} (U [ i + 1 ] - U [ i ]) \tag {10} +$$ + +The last step is to apply inverse CDF of target distribution, + +$$ +\mathcal {T} (x) = F _ {T} ^ {- 1} (u (x)) \tag {11} +$$ + +Training and Testing Procedure. The training and testing procedures of UniPAN are described in Algorithm 1. Here, $L \uparrow$ denotes the upsampled LRMS image using bicubic interpolation to match the size of the PAN image. The model $\mathcal{M}$ is trained to predict the residual of LRMS with an $l_{1}$ loss function that minimizes the discrepancy between the predicted HRMS image and the ground truth. + +As for the target distribution $\rho_{t}$ , we test two widely used distributions, i.e. standard normal distribution $\mathcal{N}(0,1)$ and standard uniform distribution $\mathcal{U}(0,1)$ , the transformation $\mathcal{T}(x)$ becomes: + +$$ +\mathcal {T} _ {\mathrm {n}} (x) = \Phi^ {- 1} \left(F _ {S} (x)\right), \quad \mathcal {T} _ {\mathrm {u}} (x) = F _ {S} (x), \tag {12} +$$ + +where $\Phi^{-1}$ denotes the inverse CDF of standard normal distribution. Since the inverse CDF of standard uniform distribution is an identical mapping, $\mathcal{T}_{\mathrm{u}}$ becomes the estimated empirical distribution itself. The standard normal distribution transformation non-linearly warps the input space to match Gaussian statistics, while the uniform transformation directly maps pixel intensities to their empirical quantiles. + +# 4. Experiment + +# 4.1. Setups + +Datasets. We conducted experiments utilizing the Red, Green, Blue, and NIR bands of the NBUPansharpRSDataset dataset [49]. The variations in resolution among different satellites, coupled with significant differences in land covers and land uses, render this dataset particularly suitable for assessing the generalization capabilities of pansharpening models. The dataset comprises data from 6 satellites. For each satellite, we partitioned the data into training, validation, and testing sets in a ratio of $7:1:2$ . + +Implementation Details. To validate the effectiveness and generalization capability of UniPAN, we conducted comparative experiments between baseline (w/o UniPAN) and enhanced (w/ UniPAN) configurations using diverse commonly-recognized models, including FusionNet [19], PanNet [82], MSDDN [28, 29], GPPNN [80], MSDCNN [83], FeINFN [40], PNN [47], LAGConv [34], SFIIN [89, 90], PreMix [18], UAPN [87], MDCUN [81]. + +![](images/3d5e7231ab00420a762b9cb2b8063163b09d86c7345ce42bb2458470f4ae8ebf.jpg) +Figure 3. Generalization testing on GaoFen-1, the networks are trained on WorldView-3. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the $D_{\lambda} / D_{S}$ value. In most cases, the use of UniPAN yields superior spectral and spatial preservation. For instance, MDCUN [81] exhibits a highly close alignment with the LRMS in spectral characteristics with UniPAN of $\rho_t = \mathcal{N}(0,1)$ . + +All implementations followed identical experimental procedures: a batch size of 8 (except for MDCUN [81] and FeINFN [40] which used 4), the Adam [35] optimizer $(\beta_{1} = 0.9, \beta_{2} = 0.999)$ with an initial learning rate of $10^{-3}$ , and StepLR scheduling which reduces the learning rate by a factor of 0.8 every 100 epochs. Training was performed on an NVIDIA GeForce RTX 4090 GPU for 600 epochs using the Wald protocol [69] with $l_{1}$ loss, with best model selected based on optimal validation performance. Note that, to prevent data leakage of testing set, the transformation fitting process is conducted on the training set of each satellite. + +Evaluation Metrics. The full-reference metrics based on spatial degradation (e.g. PSNR, SSIM) can only reflect performance in simulated reduced-resolution experiments and fail to demonstrate effectiveness in real-world full-resolution scenarios [6, 16, 46]. In contrast, no-reference metrics like QNR explicitly quantify model performance in both spectral preservation and spatial enhancement, enabling direct evaluation of fusion results at native resolution while avoiding theoretical biases introduced by downsampling. To better align with practical applications where reference images are unavailable, this paper adopts widely-used no-reference spectral and spatial metrics, i.e. QNR [5], $D_{\lambda}$ [5], $D_{S}$ [5], + +and $D_{\rho}$ [60], to comprehensively assess cross-resolution and cross-satellite generalization capabilities. + +# 4.2.Quantitative Evaluation + +Tab. 1 presents the average generalization performance of various models enhanced with UniPAN under different target distributions. Green indicate metric improvements, while red denote performance degradation. Slight yet observable differences exist between the two target distributions. Overall, the uniform distribution achieves better performance than the normal distribution. The normal distribution demonstrated a modest decline in performance across the WV2 and WV3 datasets. This phenomenon may stem from the peaked characteristics of the normal distribution, which could hinder model training, whereas the uniform distribution provides smoother normalization (see Fig. 5 in Sec. 4.3). + +Notably, UniPAN with $\rho_{t} = \mathcal{U}(0,1)$ achieves significant improvements across most metrics, with only marginal declines in rare cases. For instance, when generalizing from WV3 to QB, UniPAN improves the QNR metric by over 0.084 (from 0.729 to 0.813) averaged on all 12 models, underscoring its effectiveness. Beyond QNR, UniPAN also enhances spectral preservation $(D_{\lambda})$ and spatial consistency + +![](images/7094b3dd5c36ccab6dfc0ce1cfc6fec32a0db01927e72b11d194d9ccb4062908.jpg) +Figure 4. Full-resolution generalization testing on QuickBird. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the $D_{\lambda} / D_{S}$ value. + +$(D_S, D_\rho)$ , further validating the efficacy of the unified distribution strategy. These results confirm that aligning heterogeneous satellite data to a unified distribution effectively mitigates domain shifts, enabling robust cross-sensor generalization. However, while UniPAN improves spatial fidelity greatly, it may cause slight spectral distortion, as evidenced by the minor performance degradation observed in the $D_{\lambda}$ metric. This is reasonable since the transformation applied to each spectral band maintains the relative spatial structure but might result in minimal loss of spectral information. Nevertheless, such information loss is extremely limited, as the overall enhancement in performance remains predominant. + +# 4.3. Visual Comparison + +Comparison of Fusion Results. To illustrate the practical impact of UniPAN, Figs. 3 and 4 compare fusion results w/ and w/o UniPAN. In most cases, UniPAN achieves highly favorable results in both spectral and spatial preservation. Furthermore, the remarkably high QNR values demonstrate its superior capability in balancing spectral and spatial fidelity, e.g. in Fig. 3, MDCUN [81] applying UniPAN with normal distribution enhanced the QNR by 0.1514 (from 0.7157 to 0.8671) and aligned fused results closely with the ground truth spectra. In the full-resolution generalization test depicted in Fig. 4, even with minimal visual differences, + +UniPAN achieves metrically significant improvements in spectral and spatial preservation. This capability is critical for downstream applications requiring accurate spectral information. Additional visual comparison results are included in the supplementary materials. + +Comparison of Data Distribution. Fig. 5 further analyzes data distribution alignment. The kernel density estimation (KDE) curves (top right area and diagonals) reveal that UniPAN successfully transforms the original complex and heterogeneous distributions into a unified target distribution. The quantile-quantile (Q-Q) plots (bottom left area) demonstrate harmonized distributions across satellites after applying UniPAN (The greater the deviation from the $y = x$ line, the more significant the disparity between the two distributions). For example, the original Q-Q plots between WV4 and QB show significant dispersion, whereas post-UniPAN points close tightly around $y = x$ . This visual evidence corroborates UniPAN's ability to bridge distribution gaps. + +# 4.4. Parameter Analysis + +UniPAN introduces two hyperparameters: the number of sampled pixels $n$ and the number of quantiles $m$ . Figs. 6 and 7 investigate their impact on performance of FusionNet [19] trained on WV2 and tested on GF1. As $n$ and $m$ vary, the metrics stabilize. Generally, $\mathcal{U}(0,1)$ demonstrates + +![](images/8c87594f3788d04fa0663c5f741b7a96fcc3f177ea1bdfaf0fd4bddbde54473a.jpg) +(a) w/o UniPAN + +![](images/2d479758e8371c95ac4f981afd8545de0e50873d93bc8c37d60e998c1b472acf.jpg) +(b) w/ UniPAN, $\rho_t = \mathcal{N}(0,1)$ + +![](images/524d7929265bed16ce9a843829177ce385f263449f98df7282245d38ecc97d0f.jpg) +(c) $\mathrm{w} / \mathrm{UniPAN}$ $\rho_t = \mathcal{U}(0,1)$ + +![](images/956fc6ef24fd91914e2bd1eb223b55b7d72695991efa19372dd43972da283cb7.jpg) +Figure 5. Comparison of data distributions before and after applying UniPAN. The upper-right and diagonals of each subplot display the kernel density estimation (KDE) curves of Green band for each satellite. The scatter plots in the lower-left areas of each subplots represent quantile-quantile (Q-Q) plots of green band, with the dashed line indicating $y = x$ . The greater the deviation from the $y = x$ line, the more significant the disparity between the two distributions. Zoom in for better view. +Figure 6. Parameter analysis of number of quantiles, with number of sampled pixels fixed to 10000. + +superior performance compared to $\mathcal{N}(0,1)$ . Both configurations achieve statistically significant improvements over baseline methods without applying UniPAN. This parametric robustness demonstrates that optimal performance can be attained without exhaustive hyperparameter tuning, significantly reducing computational overhead. + +# 5. Conclusion and Discussion + +We proposed UniPAN, a unified distribution strategy addressing generalization challenges in deep learning-based pansharpening. UniPAN eliminates spectral discrepancies by normalizing multi-source remote sensing imagery into predefined target distributions, effectively bridging the gap between training and testing distributions. Extensive quantitative and visual comparisons across 6 satellite sensors and + +![](images/a8b78f92931d5d2121022107a706625190cffd080196318790c7203579d9b664.jpg) +Figure 7. Parameter analysis of number of sampled pixels, with number of quantiles fixed to 1000. + +12 models demonstrate that UniPAN significantly enhances cross-sensor generalization. Notably, this enhancement is achieved without introducing any additional trainable parameters or the need for retraining or fine-tuning the models. + +While this work aligns theoretically with 1D optimal transport, considering interdependencies of spectral bands or spatial patches, future research could explore extending the method to the multi-dimensional case with optimal transport theories. Additionally, the experimental results indicate that the choice of target distribution has a notable impact on performance, thus it is worthwhile investigating adaptive target domain selection. To further validate the generalizability of the proposed unified distribution strategy, the ideas presented in this work could be extended to other remote sensing image processing tasks, such as multi-modal fusion and hyperspectral pansharpening, etc. + +Acknowledgment. This research was partially supported by the National Natural Science Foundation of China under Grant 32441114, 41971397, 61731022, and the project Y1H103101A. + +# References + +[1] B. Aiazzi, L. Alparone, S. Baronti, and A. Garzelli. Context-driven fusion of high spatial and spectral resolution images based on oversampled multiresolution analysis. IEEE TGRS, 40(10):2300-2312, 2002. 3 +[2] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva. An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas. In 2003 2nd GRSS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, pages 90-94, 2003. +[3] Bruno Aiazzi, L. Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. MTF-tailored multiscale fusion of high-resolution MS and Pan imagery. *Photogramm. Eng. Remote Sens.*, 72:591-596, 2006. 3 +[4] Bruno Aiazzi, Stefano Baronti, and Massimo Selva. Improving component substitution pansharpening through multivariate regression of MS+Pan data. IEEE TGRS, 45(10): 3230-3239, 2007. 2 +[5] Luciano Alparone, Bruno Aiazzi, Stefano Baronti, Andrea Garzelli, Filippo Nencini, and Massimo Selva. Multispectral and panchromatic data fusion assessment without reference. *Photogramm. Eng. Remote Sens.*, 74(2):193-200, 2008. 6 +[6] Alberto Arienzo, Gemine Vivone, Andrea Garzelli, Luciano Alparone, and Jocelyn Chanussot. Full-resolution quality assessment of pansharpening: Theoretical and hands-on approaches. IEEE GRSM, 10(3):168-201, 2022. 6 +[7] Coloma Ballester, Vicent Caselles, and Laura Igual. A variational model for $\mathrm{P + XS}$ image fusion. IJCV, 69(1):43-58, 2006. 3 +[8] Wele Gedara Chaminda Bandara and Vishal M Patel. Hypertransformer: A textural and spectral feature fusion Transformer for pansharpening. In CVPR, pages 1767-1777, 2022, 2, 3 +[9] Qi Cao, Liang-Jian Deng, Wu Wang, Junming Hou, and Gemine Vivone. Zero-shot semi-supervised learning for pansharpening. Inf. Fusion., 101:102001, 2024. 2, 3 +[10] Pat S. Chavez and Andy Yaw Kwarteng. Extracting spectral contrast in landsat thematic mapper image data using selective principal component analysis. Photogramm. Eng. Remote Sens., 55:339-348, 1989. 2 +[11] Lihui Chen, Gemine Vivone, Zihao Nie, Jocelyn Chanussot, and Xiaomin Yang. Spatial data augmentation: Improving the generalization of neural networks for pansharpening. IEEE TGRS, 61:1-11, 2023. 3 +[12] Zhi-Xuan Chen, Cheng Jin, Tian-Jing Zhang, Xiao Wu, and Liang-Jian Deng. SpanConv: A new convolution via spanning kernel space for lightweight pansharpening. In IJCAI, pages 841-847, 2022. Main Track. 3 +[13] Minhyeok Choi, Rae Young Kim, and Myeong-Ryong Nam. Fusion of multispectral and panchromatic satellite images + +using the curvelet transform. IEEE GRSL, 2(2):136-140, 2005. 3 +[14] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE TGRS, pages 1-1, 2022. 3 +[15] Matteo Ciotola, Giovanni Poggi, and Giuseppe Scarpa. Unsupervised deep learning-based pansharpening with jointly enhanced spectral and spatial fidelity. IEEE TGRS, 61:1-17, 2023. 3 +[16] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pixel-wise ensembled masked autoencoder for multispectral pansharpening. IEEE TGRS, 62:1-22, 2024. 3, 6 +[17] Yongchuan Cui, Peng Liu, Bingze Song, Lingjun Zhao, Yan Ma, and Lajiao Chen. Reconstruction of large-scale missing data in remote sensing images using Extend-GAN. IEEE GRSL, 21:1-5, 2024. 3 +[18] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pansharpening via predictive filtering with element-wise feature mixing. ISPRS J. Photogramm. Remote Sens., 219:22-37, 2025. 3, 5 +[19] Liang-Jian Deng, Gemine Vivone, Cheng Jin, and Jocelyn Chanussot. Detail injection-based deep convolutional neural networks for pansharpening. IEEE TGRS, 59(8):6995-7010, 2021. 1, 2, 5, 7 +[20] Liang-jian Deng, Gemine Vivone, Mercedes E. Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pansharpening: A benchmark, from shallow to deep networks. IEEE GRSM, 10(3):279-315, 2022. 1, 2, 3 +[21] Shangqi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, and Rui Wen. Bidirectional dilation Transformer for multispectral and hyperspectral image fusion. In IJCAI, 2023. 2, 3 +[22] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. PSRT: Pyramid shuffle-and-reshuffle Transformer for multispectral and hyperspectral image fusion. IEEE TGRS, 61:1-15, 2023. 2, 3 +[23] Renwei Dian, Anjing Guo, and Shutao Li. Zero-shot hyperspectral sharpening. IEEE TPAMI, 45(10):12650-12666, 2023. 2, 3 +[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3 +[25] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In CVPR, pages 27738-27747, 2024. 3 +[26] Andrea Garzelli and Filippo Nencini. Wavelet-based fusion of optical and SAR image data in remote sensing applications. Int. J. Remote Sens., 23(23):5025-5039, 2002. 3 +[27] Morteza Ghahremani and Hassan Ghassemian. Nonlinear IHS: A promising method for pan-sharpening. IEEE GRSL, 13(11):1606-1610, 2016. 2 + +[28] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pyramid dual domain injection network for pan-sharpening. In ICCV, pages 12862-12871, 2023. 5 +[29] Xuanhua He, Keyu Yan, Jie Zhang, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Multiscale dual-domain guidance network for pan-sharpening. IEEE TGRS, 61:1-13, 2023. 5 +[30] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. AAAI, 38(3):2121-2129, 2024. 3 +[31] Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. PanMamba: Effective pan-sharpening with state space model. Inf.Fusion., 115:102779, 2025. 2, 3 +[32] Junming Hou, Xinyang Liu, Chenxu Wu, Xiaofeng Cong, Chentong Huang, Liang-Jian Deng, and Jian Wei You. Bidomain uncertainty gated recursive network for pan-sharpening. Inf. Fusion., 118:102938, 2025. 3 +[33] Cheng Jin, Liang-Jian Deng, Ting-Zhu Huang, and Gemine Vivone. Laplacian pyramid networks: A new approach for multispectral pansharpening. Inf. Fusion., 78:158-170, 2022. 3 +[34] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. LAGConv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. AAAI, 36(1):1113-1121, 2022. 3, 5 +[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[36] Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepcev, and Gustavo K. Rohde. Optimal mass transport: Signal processing and machine-learning applications. IEEE SPM, 34 (4):43-59, 2017. 4 +[37] Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo K. Rohde. Generalized sliced wasserstein distances, 2019. 4 +[38] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. 2000. US Patent 6,011,875. 2 +[39] Zhixuan Li, Jinjiang Li, Fan Zhang, and Linwei Fan. CADUI: Cross-attention-based depth unfolding iteration network for pansharpening remote sensing images. IEEE TGRS, 61:1–20, 2023. 3 +[40] YuJie Liang, ZiHan Cao, Shangqi Deng, Hong-Xia Dou, and Liang-Jian Deng. Fourier-enhanced implicit neural fusion network for multispectral and hyperspectral image fusion. In NeurIPS, pages 63441-63465. Curran Associates, Inc., 2024. 3, 5, 6 +[41] J. G. Liu. Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details. Int. J. Remote Sens., 21(18):3461-3472, 2000. 3 +[42] Peng Liu, Jun Li, Lizhe Wang, and Guojin He. Remote sensing data fusion with generative adversarial networks: State-of-the-art methods and future research directions. IEEE GRSM, 10(2):295-328, 2022. 2 +[43] Qingjie Liu, Huanyu Zhou, Qizhi Xu, Xiangyu Liu, and Yun-hong Wang. PSGAN: A generative adversarial network for + +remote sensing image pan-sharpening. IEEE TGRS, 59(12): 10227-10242, 2020. 2, 3 +[44] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-GAN: a multiscale unsupervised network for remote sensing image pansharpening. IEEE TGRS, 61:1-18, 2023. 3 +[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In ICCV, 2021. 3 +[46] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-GAN: An unsupervised pan-sharpening method for remote sensing image fusion. Inf. Fusion., 62: 110-120, 2020. 2, 3, 6 +[47] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. In Remote Sens., page 594, 2016. 2, 3, 5 +[48] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. PanDiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE TGRS, 61:1-17, 2023. 2, 3 +[49] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE GRSM, 9(1):18-52, 2021. 1, 2, 5 +[50] Xiangchao Meng, Nan Wang, Feng Shao, and Shutao Li. Vision Transformer for pansharpening. IEEE TGRS, 60:1-11, 2022. 2, 3 +[51] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pages 666-704, 1781. 3 +[52] Fjola Palsson, Johannes R. Sveinsson, and Magnus O. Ulfarsson. A new pansharpening algorithm based on total variation. IEEE GRSL, 11(1):318-322, 2014. 3 +[53] Siran Peng, Xiangyu Zhu, Haoyu Deng, Liang-Jian Deng, and Zhen Lei. FusionMamba: Efficient remote sensing image fusion with state space model. IEEE TGRS, 2024. 2, 3 +[54] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Found. Trends Mach. Learn., 11(5-6):355-607, 2019. 3 +[55] Sheila Rahmani, Melissa Strait, Daria Merkurjev, Michael Moeller, and Todd Wittman. An adaptive IHS pan-sharpening method. IEEE GRSL, 7(4):746-750, 2010. 2 +[56] Aaditya Ramdas, Nicolas Garcia, and Marco Cuturi. On Wasserstein two sample testing and related families of nonparametric tests, 2015. 4 +[57] Xiangyu Rui, Xiangyong Cao, Yining Li, and Deyu Meng. Variational zero-shot multispectral pansharpening. IEEE TGRS, 62:1-16, 2024. 3 +[58] Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, and Deyu Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion., 107:102325, 2024. 2, 3 +[59] Filippo Santambrogio. One-dimensional issues, pages 59-85. Springer International Publishing, Cham, 2015. 4 + +[60] Giuseppe Scarpa and Matteo Ciotola. Full-resolution quality assessment for pansharpening. Remote Sens., 14(8), 2022. 6 +[61] Vijay P. Shah, Nicolas H. Younan, and Roger L. King. An efficient pan-sharpening method via a combined adaptive PCA approach and contourlets. IEEE TGRS, 46(5):1323-1335, 2008. 2 +[62] Xunyang Su, Jinjiang Li, and Zhen Hua. Transformer-based regression network for pansharpening remote sensing images. IEEE TGRS, 60:1-23, 2022. 2, 3 +[63] Weiwei Sun, Jun Zhou, Xiangchao Meng, Gang Yang, Kai Ren, and Jiangtao Peng. Coupled temporal variation information estimation and resolution enhancement for remote sensing spatial-temporal-spectral fusion. IEEE TGRS, 61: 1-18, 2023. 3 +[64] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008. 2 +[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3 +[66] Cédric Villani et al. Optimal transport: old and new. Springer, 2008. 4 +[67] Gemine Vivone, Miguel Simões, Mauro Dalla Mura, Rocco Restaino, José M. Bioucas-Dias, Giorgio A. Licciardi, and Jocelyn Chanussot. Pansharpening based on semiblind deconvolution. IEEE TGRS, 53(4):1997-2010, 2015. 3 +[68] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O. Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE GRSM, 9(1):53-81, 2021. 1, 2, 3 +[69] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. *Photogramm. Eng. Remote Sens.*, 63:691-699, 1997. 3, 6 +[70] Hebaixu Wang, Meiqi Gong, Xiaoguang Mei, Hao Zhang, and Jiayi Ma. Deep unfolded network with intrinsic supervision for pan-sharpening. AAAI, 38(6):5419-5426, 2024. 3 +[71] Hebaixu Wang, Hao Zhang, Xin Tian, and Jiayi Ma. Zero-Sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation. Inf. Fusion., 101:102003, 2024. 3 +[72] Jun-Da Wang, Liang-Jian Deng, Chen-Yu Zhao, Xiao Wu, Hong-Ming Chen, and Gemine Vivone. Cascadic multireceptive learning for multispectral pansharpening. IEEE TGRS, 61:1-16, 2023. 3 +[73] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE TGRS, 62:1-16, 2024. 3 +[74] Zhong-Cheng Wu, Ting-Zhu Huang, Liang-Jian Deng, JinFan Hu, and Gemine Vivone. $\mathrm{VO + }$ Net: An adaptive approach using variational optimization and deep learning for panchromatic sharpening. IEEE TGRS, 60:1-16, 2022. 3 +[75] Jin-Liang Xiao, Ting-Zhu Huang, Liang-Jian Deng, Zhong-Cheng Wu, Xiao Wu, and Gemine Vivone. Variational pan-sharpening based on coefficient estimation with nonlocal regression. IEEE TGRS, 61:1-15, 2023. 3 + +[76] Guangxu Xie, Rencan Nie, Jinde Cao, He Li, and Jintao Li. A deep multiresolution representation framework for pansharpening. IEEE TGRS, 62:1-16, 2024. 3 +[77] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xi-wei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE TGRS, 2024. 3 +[78] Yinghui Xing, Litao Qu, Shizhou Zhang, Kai Zhang, Yanning Zhang, and Lorenzo Bruzzone. CrossDiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE TIP, 2024. 2, 3 +[79] Qizhi Xu, Yuan Li, Jinyan Nie, Qingjie Liu, and Mengyao Guo. UPanGAN: Unsupervised pansharpening based on the spectral and spatial loss constrained generative adversarial network. Inf. Fusion., 91:31-46, 2023. 3 +[80] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In CVPR, pages 1366-1375, 2021. 3, 5 +[81] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pansharpening. In CVPR, pages 1778-1787, 2022. 3, 5, 6, 7 +[82] Jie Yang, Xiao-Yang Fu, Yan Hu, Yue Huang, Xing Ding, and John Paisley. PanNet: A deep network architecture for pan-sharpening. ICCV, pages 1753-1761, 2017. 2, 3, 5 +[83] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE JSTARS, 11(3):978-989, 2018. 2, 3, 5 +[84] Jie Zhang, Xuanhua He, Keyu Yan, Ke Cao, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Pan-sharpening with wavelet-enhanced high-frequency information. IEEE TGRS, 62:1-14, 2024. 3 +[85] Chen-Yu Zhao, Tian-Jing Zhang, Ran Ran, Zhi-Xuan Chen, and Liang-Jian Deng. LGPCConv: Learnable gaussian perturbation convolution for lightweight pansharpening. In IJCAI, pages 4647-4655, 2023. Main Track. 3 +[86] Mengjiao Zhao, Mengting Ma, Xiangdong Li, Xiaowen Ma, Wei Zhang, and Siyang Song. Supervised detail-guided multiscale state space model for pan-sharpening. IEEE TGRS, 2024. 2, 3 +[87] Kaiwen Zheng, Jie Huang, Man Zhou, Danfeng Hong, and Feng Zhao. Deep adaptive pansharpening via uncertainty-aware image fusion. IEEE TGRS, 61:1-15, 2023. 5 +[88] Yu Zhong, Xiao Wu, Liang-Jian Deng, ZiHan Cao, and Hong-Xia Dou. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pansharpening. In NeurIPS, pages 77962-77986. Curran Associates, Inc., 2024. 3 +[89] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In ECCV, pages 274-291, Cham, 2022. Springer Nature Switzerland. 3, 5 +[90] Man Zhou, Jie Huang, Keyu Yan, Danfeng Hong, Xiuping Jia, Jocelyn Chanussot, and Chongyi Li. A general spatial-frequency learning framework for multimodal image fusion. IEEE TPAMI, pages 1-18, 2024. 3, 5 \ No newline at end of file diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/images.zip b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f150a80addc09571fb24e3a4bbb52cfffce70c11 --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e854fd7c7703350e93dfab3008d6c9a43a9c3476dde9ffd59045e16111e4358b +size 1192345 diff --git a/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/layout.json b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a99cdfb4f8a68d401847607e939f8f34532e0503 --- /dev/null +++ b/2025/Enpowering Your Pansharpening Models with Generalizability_ Unified Distribution is All You Need/layout.json @@ -0,0 +1,10450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 110, + 103, + 503, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 110, + 103, + 503, + 138 + ], + "spans": [ + { + "bbox": [ + 110, + 103, + 503, + 138 + ], + "type": "text", + "content": "Enpowering Your Pansharpening Models with Generalizability: Unified Distribution is All You Need" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "spans": [ + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "text", + "content": "Yongchuan Cui " + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "text", + "content": " Peng Liu " + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "text", + "content": " Hui Zhang " + }, + { + "bbox": [ + 167, + 160, + 444, + 175 + ], + "type": "inline_equation", + "content": "^{3,*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 175, + 543, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 175, + 543, + 190 + ], + "spans": [ + { + "bbox": [ + 67, + 175, + 543, + 190 + ], + "type": "text", + "content": "1Aerospace Information Research Institute, Chinese Academy of Sciences, 100094 Beijing, China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 189, + 468, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 189, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 189, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 189, + 468, + 217 + ], + "type": "text", + "content": " School of Electronic, Electrical and Communication Engineering, University of Chinese Academy of Sciences, 101408 Beijing, China" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 217, + 494, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 217, + 494, + 232 + ], + "spans": [ + { + "bbox": [ + 115, + 217, + 494, + 232 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 115, + 217, + 494, + 232 + ], + "type": "text", + "content": " School of Engineering Medicine, Beihang University, 100191 Beijing, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 105, + 234, + 501, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 234, + 501, + 245 + ], + "spans": [ + { + "bbox": [ + 105, + 234, + 501, + 245 + ], + "type": "text", + "content": "yongchuancui@gmail.com, liupeng202303@aircas.ac.cn, hui.zhang@buaa.edu.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 261, + 200, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 261, + 200, + 273 + ], + "spans": [ + { + "bbox": [ + 151, + 261, + 200, + 273 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 53, + 286, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 286, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 53, + 286, + 297, + 586 + ], + "type": "text", + "content": "Existing deep learning-based models for remote sensing pansharpening exhibit exceptional performance on training datasets. However, due to sensor-specific characteristics and varying imaging conditions, these models suffer from substantial performance degradation when applied to unseen satellite data, lacking generalizability and thus limiting their applicability. We argue that the performance drops stem primarily from distributional discrepancies from different sources and the key to addressing this challenge lies in bridging the gap between training and testing distributions. To validate the idea and further achieve a \"train once, deploy forever\" capability, this paper introduces a novel and intuitive approach to empower any pansharpening models with generalizability by employing a unified distribution strategy (UniPAN). Specifically, we construct a distribution transformation function that normalizes the pixels sampled from different sources to conform to an identical distribution. The deep models are trained on the transformed domain, and during testing on new datasets, the new data are also transformed to match the training distribution. UniPAN aims to train and test the model on a unified and consistent distribution, thereby enhancing its generalizability. Extensive experiments validate the efficacy of UniPAN, demonstrating its potential to significantly enhance the performance of deep pansharpening models across diverse satellite sensors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 285, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 285, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 285, + 597 + ], + "type": "text", + "content": "Codes: https://github.com/yc-cui/UniPAN." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 135, + 635 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 643, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 643, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 643, + 297, + 693 + ], + "type": "text", + "content": "Due to the inherent limitations of sensor imaging technology, advanced remote sensing satellites such as IKONOS, QuickBird, GaoFen-2, and WorldView-3, necessitate a compromise between the acquisition of extensive spectral infor" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 266, + 389, + 392 + ], + "blocks": [ + { + "bbox": [ + 317, + 266, + 389, + 392 + ], + "lines": [ + { + "bbox": [ + 317, + 266, + 389, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 389, + 392 + ], + "type": "image", + "image_path": "16bf4c0c34df7e504d3392c3eea04942c9c8ef0216e1716cf0f703efef000caf.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 427 + ], + "type": "text", + "content": "Figure 1. Comparison of generalization metrics w/ and w/o UniPAN trained on WorldView-3 (WV3) using FusionNet [19]." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 396, + 266, + 468, + 392 + ], + "blocks": [ + { + "bbox": [ + 396, + 266, + 468, + 392 + ], + "lines": [ + { + "bbox": [ + 396, + 266, + 468, + 392 + ], + "spans": [ + { + "bbox": [ + 396, + 266, + 468, + 392 + ], + "type": "image", + "image_path": "940d3e0128c46e96e9a5ee907ff658059987a3ffb9243dadf7b3555aebc93195.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 476, + 266, + 547, + 392 + ], + "blocks": [ + { + "bbox": [ + 476, + 266, + 547, + 392 + ], + "lines": [ + { + "bbox": [ + 476, + 266, + 547, + 392 + ], + "spans": [ + { + "bbox": [ + 476, + 266, + 547, + 392 + ], + "type": "image", + "image_path": "cfed76fb5cdad7a1cb0c3d41934aac9950234a70a1df91b45dc495f5dcc52f03.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 315, + 438, + 548, + 559 + ], + "blocks": [ + { + "bbox": [ + 315, + 438, + 548, + 559 + ], + "lines": [ + { + "bbox": [ + 315, + 438, + 548, + 559 + ], + "spans": [ + { + "bbox": [ + 315, + 438, + 548, + 559 + ], + "type": "image", + "image_path": "863f9367fcf43cf5ac9450f61a583ec0381744051b44eaeca8811c755dfcd24c.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 314, + 568, + 553, + 580 + ], + "lines": [ + { + "bbox": [ + 314, + 568, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 568, + 553, + 580 + ], + "type": "text", + "content": "Figure 2. Comparison of tSNE visualization w/ and w/o UniPAN." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 687 + ], + "type": "text", + "content": "mation and maintaining fine spatial details in their captured imagery [20, 49, 68]. Consequently, these satellites typically capture either high spatial resolution single-band panchromatic (PAN) images or low spatial resolution multispectral (LRMS) images. Pansharpening is a crucial technique in remote sensing image processing, aiming to combine PAN and LRMS images to generate fused images with high spatial resolution and abundant spectra (HRMS)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Traditional pansharpening methods have been extensively studied over the past few decades. However, these" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "spans": [ + { + "bbox": [ + 58, + 703, + 133, + 713 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11850" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "methods may introduce spectral distortions, require significant computational resources, and are sensitive to the quality of input data [20, 42, 49, 68]. In contrast, deep learning (DL)-based methods have shown great potential in pansharpening. Notably, various innovative architectures have emerged, starting from early convolutional models [19, 47, 82, 83], progressing to Transformer-based approaches [8, 21, 22, 24, 50, 62, 65], then generative methods [43, 46, 48, 58, 78], and continuing with the emergence of the latest novel designs [9, 23, 31, 53, 86]. The networks are evolving towards larger parameters capacity, stronger representation ability, and increasingly complex and diverse structures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 232, + 298, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 298, + 435 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 298, + 435 + ], + "type": "text", + "content": "However, as deep pansharpening thrives, the overly powerful fitting capabilities of these networks have concurrently presented several ongoing concerns, e.g. generalizability, interpretability, over-reliance on high-quality data, etc. This paper specifically delves into the widely criticized issue of generalizability. Specifically, although deep models achieve excellent performance on training datasets, they often fail to generalize well to new data due to distribution shift such as different noise levels, varying spectral signatures and sensor conditioning characteristics. While models can achieve satisfactory performances on new datasets through retraining or fine-tuning, the process is time-consuming and fine-tuning on every new dataset is impractical and cumbersome. The ideal scenario would involve developing a pansharpening model that can be trained once and deployed effectively across diverse datasets and sensors, thereby providing a train once, deploy forever capability." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 438, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 438, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 56, + 438, + 298, + 715 + ], + "type": "text", + "content": "To this end, this paper introduces a novel approach that aims to empower any pansharpening models with generalizability by employing a unified distribution strategy (dubbed as UniPAN). The proposed method aims to bridge the gap between training and testing distributions, thereby improving the model's ability to generalize to unseen data. The core idea is to normalize the distribution of spatial pixels from the training dataset to conform to a predefined distribution (e.g. Gaussian distribution) and deep models are then trained on the transformed domain. This process is model-agnostic and requires no additional trainable parameters. When generalizing to new data, testing datasets are also transformed to match the training distribution, ensuring that the models are evaluated under the same distributional assumptions. Fig. 1 presents the cross-satellite generalization performance of a well-known pansharpening model FusionNet [19] on NBUPansharpRSData [49] trained w/ and w/o UniPAN. The model utilizing UniPAN demonstrates a significant improvement on full-resolution metrics, which validates the strong generalization ability for real-world scenes. To further elucidate the efficacy of UniPAN, Fig. 2 presents tSNE [64] dimensionality reduction clustering of multispectral images from diverse satellites using NBUPansharpRSData [49] w/" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 556, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 215 + ], + "type": "text", + "content": "and w/o UniPAN. The pre-UniPAN visualization reveals distinct clustering patterns, with data points from different satellites forming separate groups, indicative of substantial inter-satellite distributional disparities. In contrast, following the application of UniPAN, a notable transformation is observed: the previously separated clusters become substantially intermingled, demonstrating a significant reduction in distributional discrepancies across different satellite datasets (e.g. the purple scatters of QuickBird). This visual evidence underscores UniPAN's capability to harmonize heterogeneous satellite data, effectively creating a more cohesive and unified representation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 239 + ], + "type": "text", + "content": "The contributions of this paper could be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 243, + 556, + 555 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 243, + 556, + 351 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 243, + 556, + 351 + ], + "spans": [ + { + "bbox": [ + 314, + 243, + 556, + 351 + ], + "type": "text", + "content": "- We propose UniPAN, a model-agnostic approach that employs a unified distribution strategy to empower any deep pansharpening models with generalizability, achieving a train once, deploy forever capability. This strategy provides two advantages: 1) Architectural independence: Compatible with any existing or future network architecture without structural modifications. 2) Training efficiency: Eliminates the need for retraining or fine-tuning when adapting to new sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 315, + 351, + 556, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 351, + 556, + 459 + ], + "spans": [ + { + "bbox": [ + 315, + 351, + 556, + 459 + ], + "type": "text", + "content": "- UniPAN is implemented via inverse transform sampling, and we prove its equivalence to the optimal solution of optimal transport in the one-dimensional case. UniPAN requires no additional trainable parameters and only needs to be fitted once before training begins, introducing almost no additional computational overhead. It serves as a viable alternative to vanilla data normalization techniques such as min-max and " + }, + { + "bbox": [ + 315, + 351, + 556, + 459 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 315, + 351, + 556, + 459 + ], + "type": "text", + "content": "-score normalization, offering a more robust and unified approach to data distribution alignment." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 315, + 459, + 556, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 459, + 556, + 555 + ], + "spans": [ + { + "bbox": [ + 315, + 459, + 556, + 555 + ], + "type": "text", + "content": "- We conduct extensive experiments using a variety of deep pansharpening models across numerous satellites. The simple yet intuitive and effective strategy provides a universal solution to the generalization challenge in DL-based pansharpening and experimental results demonstrate its effectiveness, showcasing significant improvements in cross-dataset generalization across a wide range of models and satellite sensors." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 567, + 402, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 567, + 402, + 580 + ], + "spans": [ + { + "bbox": [ + 314, + 567, + 402, + 580 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 587, + 452, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 452, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 452, + 601 + ], + "type": "text", + "content": "2.1. Pansharpening Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 557, + 715 + ], + "type": "text", + "content": "Traditional Methods. Traditional pansharpening methods can be broadly categorized into three paradigms: component substitution (CS), multi-resolution analysis (MRA), and variational optimization (VO) [20, 42, 49, 68]. CS-based methods [4, 10, 27, 38, 55, 61] operate by substituting the spatial component of the LRMS image with the PAN image. Representative algorithms include Gram-Schmidt (GS) [4, 38], principal component analysis (PCA) [10, 61], and intensity-hue-saturation (IHS) [27, 55]. While computationally effi" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11851" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 251 + ], + "type": "text", + "content": "cient, CS methods often introduce spectral distortions due to mismatches between PAN and substituted components. MRA-based methods [1-3, 13, 26, 41] employ multi-scale decomposition tools (e.g. wavelets [26], contourlets [13]) to inject high-frequency details from PAN into LRMS. Typical implementations include MTF-GLP [1-3] and SFIM [41]. These methods generally preserve spectral fidelity better than CS approaches but may exhibit limited spatial enhancement. VO-based approaches [7, 52, 63, 67, 74, 75] formulate pansharpening as an optimization problem with handcrafted priors, such as the variational wavelet-based methods [7] and Bayesian approaches [52]. Despite their mathematical elegance, traditional methods often require manual parameter tuning and struggle with complex spatial-spectral relationships [20, 68]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 253, + 298, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 298, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 298, + 601 + ], + "type": "text", + "content": "Deep Learning-Based Methods. The advent of deep learning (DL) have revolutionized pansharpening by enabling end-to-end learning of spatial-spectral relationships. Early pioneering convolutional neural networks (CNNs) such as PNN [47], PanNet [82], and MSDCNN [83] established the feasibility of learning fusion rules from a mount of simulated data. Subsequent innovations introduced multi-scale architectures [33, 72, 76], attention mechanisms [39, 73], and kernel customizing [12, 18, 25, 34, 85], etc. The CNN-based methods demonstrated superior performance over traditional approaches but faced limitations in modeling long-range dependencies. In contrast, Transformer [24, 45, 65] addressed this limitation with self-attention mechanisms. Although various Transformer architectures specifically designed for pansharpening [8, 21, 22, 50, 62] have achieved impressive performance, they face criticism for quadratic computational complexity w.r.t. sequence length [16]. To address this limitation, recent studies have explored linear-complexity models [16, 31, 53, 86] that maintain performance while improving efficiency. To further enhance fidelity preservation, generative models like generative adversarial networks (GANs) [17, 43, 44, 46, 79] and diffusion models [48, 58, 77, 78, 88] have emerged as promising solutions. Beyond conventional configurations, contemporary research has expanded into more challenging frontiers to address practical constraints, including zero-shot learning paradigms [9, 23, 57, 71], frequency domain integration [30, 40, 84, 89, 90], and interpretable fusion mechanisms [32, 70, 80, 81]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 611, + 284, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 284, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 284, + 624 + ], + "type": "text", + "content": "2.2. Generalization in DL-Based Pansharpening" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 714 + ], + "type": "text", + "content": "Despite remarkable progress in DL-based pansharpening, generalization capability remains a persistent challenge. The issue manifests conspicuously in two critical scenarios: 1) Models trained under the Wald protocol [69] using spatially degraded data often fail to maintain performance when applied to full-resolution inference, frequently exhibiting severe spectral distortions. 2) Networks optimized for specific" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 203 + ], + "type": "text", + "content": "satellite sensors demonstrate substantial performance degradation when deployed on unseen satellite platforms due to distribution shifts in spectral signatures and sensor characteristics. Existing approaches to address these challenges primarily focus on unsupervised learning [14, 15, 46], generalized module design [77], or advanced data augmentation [11]. However, such methods typically require either meticulous architectural redesign or computationally intensive retraining/fine-tuning procedures. These requirements pose significant practical limitations for real-world deployment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 204, + 556, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 556, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 556, + 323 + ], + "type": "text", + "content": "In contrast to model-centric solutions, our approach adopts a fundamentally different perspective by addressing the generalization challenge through data-space intervention. The proposed UniPAN framework introduces a model-agnostic preprocessing technique that operates on raw sensor data. Compared to traditional linear data normalization such as min-max scaling or " + }, + { + "bbox": [ + 313, + 204, + 556, + 323 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 204, + 556, + 323 + ], + "type": "text", + "content": "-score standardization, our method employs a non-linear mapping to project the input data onto a unified distribution, effectively harmonizing the distribution mismatches from all sensors." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 332, + 370, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 332, + 370, + 344 + ], + "spans": [ + { + "bbox": [ + 314, + 332, + 370, + 344 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 351, + 471, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 471, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 471, + 363 + ], + "type": "text", + "content": "3.1. Distribution Transformation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": "The core challenge in cross-domain pansharpening stems from the distribution discrepancy between source (training) and target (testing) domains. Let " + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_S" + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": " denote the source domain (training data) and " + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_T" + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": " the target domain (testing data). Our objective is to learn a transformation " + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{T}:\\mathcal{X}\\to \\mathcal{Z}" + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": " that maps both domains to a shared space " + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": " with unified distribution " + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "inline_equation", + "content": "\\rho_t" + }, + { + "bbox": [ + 313, + 369, + 556, + 453 + ], + "type": "text", + "content": ", e.g. standard normal distribution." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "text", + "content": "Inverse Transform Sampling. Given a univariate r.v. " + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "inline_equation", + "content": "X \\sim \\rho_s" + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "text", + "content": ", we seek a transformation " + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{T} : \\mathbb{R} \\to \\mathbb{R}" + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "inline_equation", + "content": "Z = \\mathcal{T}(X) \\sim \\rho_t" + }, + { + "bbox": [ + 313, + 453, + 556, + 502 + ], + "type": "text", + "content": ". The inverse transform sampling theorem provides a theoretically grounded solution:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 388, + 504, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 504, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 388, + 504, + 555, + 519 + ], + "type": "interline_equation", + "content": "\\mathcal {T} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right), \\tag {1}", + "image_path": "bd03c54d2938c7b17e10e46cf10aacc9afcb2e17d7e2d0bc37c6f90b0264422d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "F_{S}(x) = \\rho_{s}(-\\infty, x]" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " is the cumulative distribution function (CDF) of " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "F_{T}^{-1}(u) = \\inf \\{z \\in \\mathbb{R} : F_{T}(z) \\geq u\\}" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " is the inverse CDF of the target distribution. By the universality of the uniform, " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "U = F_{S}(X)" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " follows a uniform distribution, and " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "Z = F_{T}^{-1}(U)" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " inherits " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "\\rho_{t}" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " exactly when " + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "inline_equation", + "content": "F_{S}" + }, + { + "bbox": [ + 313, + 523, + 556, + 631 + ], + "type": "text", + "content": " is continuous. To provide more intuitively analyze of the algorithm's operational process and efficiency, we establish a fundamental connection between inverse transform sampling and the well-studied optimal transport theory [51, 54]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "spans": [ + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "content": "Optimal Transport Perspective. Given two probability distributions " + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "inline_equation", + "content": "\\rho_{s}" + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "content": " (source domain) and " + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "inline_equation", + "content": "\\rho_{t}" + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "content": " (target domain) defined on " + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^m" + }, + { + "bbox": [ + 313, + 631, + 556, + 678 + ], + "type": "text", + "content": " respectively, we consider the Monge problem [51, 54]:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 322, + 682, + 556, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 682, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 322, + 682, + 556, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {C} (\\rho_ {s}, \\rho_ {t}) \\triangleq \\min _ { \\begin{array}{c} T: \\mathbb {R} ^ {n} \\to \\mathbb {R} ^ {m} \\\\ T _ {\\sharp} \\rho_ {s} = \\rho_ {t} \\end{array} } \\int_ {\\mathbb {R} ^ {n}} c (x, T (x)) \\rho_ {s} (x) d x, \\quad (2)", + "image_path": "dfd6a976cd8defc052848b43d314a137fed8c97031b0549905c3956dccb9438c.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11852" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "c:\\Omega_s\\times \\Omega_t\\to \\mathbb{R}^+" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " is a cost function and for any measurable Borel set " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "E\\in \\Omega_s" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "T:\\Omega_s\\rightarrow \\Omega_t" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " transfers measures from one space " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\Omega_{s}" + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "text", + "content": " to another space " + }, + { + "bbox": [ + 55, + 72, + 296, + 109 + ], + "type": "inline_equation", + "content": "\\Omega_t" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 118, + 296, + 133 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 118, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 123, + 118, + 296, + 133 + ], + "type": "interline_equation", + "content": "T _ {\\sharp} \\rho_ {s} (E) = \\rho_ {s} \\left(T ^ {- 1} (E)\\right), \\tag {3}", + "image_path": "9e49073095aef22808e35f9eafee199edfda58ca506c49c2cdb900147c7aec80.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": "here " + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "inline_equation", + "content": "T_{\\sharp}" + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": " is the pushforward operator. The Monge problem seeks a transport plan " + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\rho_s" + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "inline_equation", + "content": "\\rho_t" + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": " that minimizes the transportation cost. The optimal " + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": " is also known as the Monge map of Eq. (2)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "F_{S}(x) = \\rho_{s}(-\\infty, x]" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "F_{T}(z) = \\rho_{t}(-\\infty, z]" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " denote the CDFs of the source and target distributions respectively. In the univariate setting (" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "n = m = 1, \\Omega_{s} = \\Omega_{t} = \\mathbb{R}" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": "), when the source measure " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "\\rho_{s}" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " is continuous and the cost function takes the form " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "c(x, z) = h(|x - z|)" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " being strictly convex, by cyclical monotonicity [66], the Monge map " + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "inline_equation", + "content": "T^{*}" + }, + { + "bbox": [ + 55, + 191, + 297, + 273 + ], + "type": "text", + "content": " satisfies:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 276, + 296, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 276, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 130, + 276, + 296, + 289 + ], + "type": "interline_equation", + "content": "F _ {S} (x) = F _ {T} \\left(T ^ {*} (x)\\right), \\tag {4}", + "image_path": "593411f55d7c4ae041fe7535256d0383ffc6b3dc21f76fe6f874256cb21a5f29.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 186, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 186, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 186, + 308 + ], + "type": "text", + "content": "yielding the explicit expression:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 127, + 317, + 296, + 332 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 317, + 296, + 332 + ], + "spans": [ + { + "bbox": [ + 127, + 317, + 296, + 332 + ], + "type": "interline_equation", + "content": "T ^ {*} (x) = F _ {T} ^ {- 1} \\left(F _ {S} (x)\\right). \\tag {5}", + "image_path": "fdc61d518ae5ff16a9c93d6bb5a05539711a3c6c627eed93be700005b0cefac9.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 342, + 297, + 379 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 297, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 297, + 379 + ], + "type": "text", + "content": "This remarkable result coincides exactly with the inverse transform sampling formula Eq. (1), i.e. " + }, + { + "bbox": [ + 55, + 342, + 297, + 379 + ], + "type": "inline_equation", + "content": "T^{*} = \\mathcal{T}" + }, + { + "bbox": [ + 55, + 342, + 297, + 379 + ], + "type": "text", + "content": ". The transport cost in this setting [37, 56, 59] can be computed as" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 93, + 388, + 296, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 388, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 93, + 388, + 296, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {C} \\left(\\rho_ {s}, \\rho_ {t}\\right) = \\int_ {0} ^ {1} c \\left(F _ {S} ^ {- 1} (u), F _ {T} ^ {- 1} (u)\\right) d u. \\tag {6}", + "image_path": "37d0dbc67d5ca2ea77e0e980aa65792e0945d7c0fde3071e166edd6aa26e1070.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "text", + "content": "For the " + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "text", + "content": " cost function " + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "inline_equation", + "content": "c(x,z) = |x - z|" + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "text", + "content": ", this corresponds to the " + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 55, + 422, + 297, + 625 + ], + "type": "text", + "content": " Wasserstein distance [36, 59] between the distributions. In the one-dimensional (1D) scenario, UniPAN operates independently on spatial pixels within each spectral band. This implementation can be achieved through interpolation-based inverse transform sampling without framing it as an optimization problem. However, given the inherent correlations between spectral bands and interdependencies among spatial patches, multidimensional extension of this framework may be promising, albeit with increased implementation complexity. The theoretical derivation establishing connections with optimal transport theory serves to facilitate future generalization of UniPAN to multidimensional configurations. Building upon the theoretical foundation of distribution transformation through optimal transport, in the following, we present a practical implementation of 1D UniPAN." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 194, + 647 + ], + "type": "text", + "content": "3.2. UniPAN Implementation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 689 + ], + "type": "text", + "content": "We present a practical implementation for 1D finite discrete distributions of UniPAN through a quantile-based distribution transformation algorithm." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Finite Discrete Distribution Transformation. Given pixel values " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "X = \\{x_{1}, x_{2}, \\dots, x_{n}\\}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " sampled from a spectral band," + } + ] + } + ], + "index": 12 + }, + { + "type": "code", + "bbox": [ + 314, + 87, + 553, + 309 + ], + "blocks": [ + { + "bbox": [ + 315, + 72, + 534, + 84 + ], + "lines": [ + { + "bbox": [ + 315, + 72, + 534, + 84 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 534, + 84 + ], + "type": "text", + "content": "Algorithm 1 UniPAN Training and Testing Procedure" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": "Training Phase \nInput: Training dataset " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{train}} = \\{L_i, P_i, H_i\\}_{i=1}^N" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": ", inverse CDF " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "F_T^{-1}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " of target distribution " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\rho_t" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": ", model " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \nOutput: Trained model " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n1: Initialize transformed LRMS " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\tilde{L}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " and PAN " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n2: for each band " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " of LRMS " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\{L_i\\}_{i=1}^N" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " & PAN " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\{P_i\\}_{i=1}^N" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " do \n3: Sample " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " pixels " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "X_b \\gets \\{x_1, \\ldots, x_n\\}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " from band " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n4: Compute " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " quantiles " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "Q = \\{q_0, \\ldots, q_{m-1}\\} \\triangleright" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " Eq. (7) \n5: Linspace uniform CDF " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "U = \\{0, \\frac{1}{m-1}, \\ldots, 1\\}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n6: for each pixel " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " in band " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " do \n7: Forward and backward interp by " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "u_\\uparrow" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "u_\\downarrow \\triangleright" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " Eq. (9) \n8: Compute eCDF value " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "u(x) \\gets (u_\\uparrow + u_\\downarrow) / 2 \\triangleright" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " Eq. (8) \n9: Transform " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "z \\gets F_T^{-1}(u(x))" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " and assign " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " to corresponding band of " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\tilde{L}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\tilde{P}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n10: end for \n11: end for \n12: while not converged do \n13: Train " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{L} = \\|H - \\mathcal{M}(\\tilde{L}, \\tilde{P}) + L \\uparrow\\|_1" + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "text", + "content": " \n14: end while \n15: return Trained model " + }, + { + "bbox": [ + 314, + 87, + 553, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "code_body" + } + ], + "index": 14, + "sub_type": "algorithm" + }, + { + "type": "code", + "bbox": [ + 315, + 324, + 554, + 415 + ], + "blocks": [ + { + "bbox": [ + 315, + 315, + 370, + 324 + ], + "lines": [ + { + "bbox": [ + 315, + 315, + 370, + 324 + ], + "spans": [ + { + "bbox": [ + 315, + 315, + 370, + 324 + ], + "type": "text", + "content": "Testing Phase" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "lines": [ + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": "Input: Testing dataset " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_{\\mathrm{test}} = \\{L_i,P_i\\}_{i = 1}^N" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " , inverse CDF " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "F_{T}^{-1}" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " of target distribution " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\rho_{t}" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " , trained model " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " \nOutput: Pansharpening result " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\hat{H}^{\\mathrm{test}}" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " \n16: for each band " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " of LRMS & PAN do \n17: Transform " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "L,P" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\tilde{L},\\tilde{P}" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " using the training transformation \n18: end for \n19: Generate " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\hat{H}^{\\mathrm{test}}\\gets \\mathcal{M}(\\tilde{L},\\tilde{P}) + L\\uparrow" + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "text", + "content": " \n20: return Fused result " + }, + { + "bbox": [ + 315, + 324, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\hat{H}^{\\mathrm{test}}" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "code_body" + } + ], + "index": 16, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "text", + "content": "we first compute " + }, + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "text", + "content": " empirical quantiles " + }, + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "inline_equation", + "content": "Q = \\{q_{0},\\dots,q_{m - 1}\\}" + }, + { + "bbox": [ + 313, + 435, + 554, + 458 + ], + "type": "text", + "content": " where:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 498 + ], + "type": "interline_equation", + "content": "q _ {i} = \\inf \\left\\{x \\in X \\mid P (X \\leq x) \\geq \\frac {i}{m - 1} \\right\\}, i = \\{0, 1, \\dots , m - 1 \\}. \\tag {7}", + "image_path": "7393c2aff332b2025f6c43995240096fbc7309d4569b955c2d9609fb60b71439.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "content": "These quantiles discretize the empirical cumulative distribution function (eCDF) of the source data. The number of quantiles " + }, + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 498, + 555, + 533 + ], + "type": "text", + "content": " controls the approximation fidelity." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "text", + "content": "Given any data point " + }, + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "text", + "content": " to be transformed, the next step involves determining its value in the eCDF through interpolation. However, when the source distribution contains repeated values or sparse regions, quantile estimation may yield non-unique or overlapping quantile boundaries. This can result in abrupt jumps in the eCDF, thereby violating the properties of monotonicity and smoothness. Therefore, we use bidirectional interpolation to estimate a robust eCDF value of " + }, + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 534, + 556, + 640 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 336, + 647, + 555, + 691 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 647, + 555, + 691 + ], + "spans": [ + { + "bbox": [ + 336, + 647, + 555, + 691 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} u (x) = u _ {\\uparrow} (x) / 2 + u _ {\\downarrow} (x) / 2 \\\\ = \\operatorname {i n t e r p} (x, Q, U) / 2 + \\tag {8} \\\\ (1 - \\text {i n t e r p} (- x, - Q ^ {\\text {r e v}}, - U ^ {\\text {r e v}})) / 2, \\\\ \\end{array}", + "image_path": "728cc546a813d10d96c2d93741185bdfb1c2607dba09950999742a67531c48dd.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "inline_equation", + "content": "U = \\{0,\\frac{1}{m - 1},\\dots,1\\}" + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "inline_equation", + "content": "Q^{\\mathrm{rev}}" + }, + { + "bbox": [ + 313, + 700, + 555, + 715 + ], + "type": "text", + "content": " is the reversed quantiles." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11853" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 62, + 555, + 325 + ], + "blocks": [ + { + "bbox": [ + 58, + 62, + 555, + 325 + ], + "lines": [ + { + "bbox": [ + 58, + 62, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 58, + 62, + 555, + 325 + ], + "type": "table", + "html": "
QNRTρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.980(+0.005)0.946(-0.010)0.971(+0.000)0.963(-0.003)0.964(-0.004)0.974(-0.000)0.985(+0.010)0.961(+0.004)0.979(+0.007)0.984(+0.017)0.982(+0.013)0.987(+0.013)
IK0.950(+0.005)0.912(+0.039)0.915(+0.045)0.924(+0.011)0.916(+0.022)0.958(+0.011)0.956(+0.011)0.919(+0.045)0.940(+0.070)0.950(+0.037)0.945(+0.051)0.968(+0.021)
QB0.957(+0.006)0.936(+0.016)0.931(+0.047)0.953(+0.017)0.944(+0.014)0.970(+0.016)0.965(+0.014)0.944(+0.024)0.966(+0.082)0.969(+0.032)0.969(+0.038)0.970(+0.017)
WV20.880(+0.026)0.811(+0.029)0.832(+0.062)0.946(+0.008)0.916(+0.012)0.951(+0.012)0.872(+0.018)0.775(-0.006)0.825(+0.055)0.944(+0.006)0.914(+0.011)0.940(+0.001)
WV30.862(-0.006)0.804(+0.035)0.795(+0.067)0.921(-0.011)0.919(+0.011)0.950(+0.001)0.880(+0.011)0.805(+0.036)0.813(+0.084)0.943(+0.010)0.924(+0.016)0.952(+0.003)
WV40.924(+0.009)0.898(+0.045)0.875(+0.036)0.948(-0.000)0.929(+0.003)0.972(+0.006)0.939(+0.024)0.889(+0.035)0.903(+0.064)0.952(+0.002)0.935(+0.009)0.973(+0.007)
Dλ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.007(-0.003)0.022(+0.008)0.005(-0.000)0.017(+0.004)0.015(+0.005)0.011(+0.000)0.004(-0.005)0.012(-0.001)0.004(-0.001)0.004(-0.008)0.004(-0.005)0.004(-0.005)
IK0.011(-0.000)0.020(-0.015)0.018(-0.011)0.031(+0.004)0.022(-0.003)0.014(-0.003)0.008(-0.003)0.020(-0.015)0.013(-0.016)0.013(-0.014)0.009(-0.016)0.010(-0.008)
QB0.013(-0.000)0.011(-0.004)0.015(-0.010)0.016(-0.004)0.014(+0.000)0.010(-0.005)0.009(-0.003)0.009(-0.006)0.007(-0.017)0.008(-0.013)0.006(-0.007)0.012(-0.003)
WV20.042(-0.006)0.066(-0.007)0.037(-0.019)0.008(-0.000)0.014(-0.000)0.016(-0.002)0.046(-0.002)0.086(+0.012)0.048(-0.008)0.008(+0.000)0.015(+0.001)0.022(+0.002)
WV30.050(+0.010)0.067(-0.011)0.049(-0.024)0.027(+0.011)0.014(+0.000)0.016(+0.002)0.042(+0.002)0.072(-0.006)0.051(-0.022)0.013(-0.002)0.011(-0.002)0.014(+0.000)
WV40.024(-0.000)0.031(-0.012)0.029(-0.009)0.017(+0.008)0.018(+0.008)0.005(-0.001)0.016(-0.008)0.032(-0.011)0.022(-0.016)0.014(+0.004)0.015(+0.005)0.004(-0.002)
DS↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.012(-0.003)0.032(+0.003)0.023(+0.000)0.020(+0.000)0.021(+0.000)0.015(-0.000)0.010(-0.005)0.026(-0.002)0.016(-0.006)0.010(-0.009)0.013(-0.007)0.007(-0.008)
IK0.039(-0.005)0.068(-0.026)0.068(-0.036)0.046(-0.014)0.063(-0.019)0.027(-0.007)0.035(-0.008)0.062(-0.033)0.046(-0.058)0.037(-0.024)0.045(-0.036)0.022(-0.013)
QB0.030(-0.005)0.053(-0.012)0.054(-0.039)0.031(-0.012)0.042(-0.014)0.020(-0.012)0.025(-0.011)0.046(-0.018)0.026(-0.067)0.023(-0.020)0.024(-0.031)0.020(-0.012)
WV20.081(-0.022)0.134(-0.024)0.137(-0.049)0.045(-0.008)0.071(-0.012)0.032(-0.010)0.085(-0.018)0.153(-0.005)0.133(-0.053)0.047(-0.006)0.071(-0.012)0.039(-0.003)
WV30.096(-0.000)0.141(-0.027)0.167(-0.050)0.054(+0.002)0.067(-0.011)0.033(-0.004)0.083(-0.013)0.134(-0.033)0.146(-0.072)0.043(-0.008)0.064(-0.014)0.033(-0.003)
WV40.053(-0.009)0.073(-0.036)0.099(-0.029)0.035(-0.006)0.053(-0.011)0.021(-0.005)0.045(-0.018)0.082(-0.027)0.076(-0.052)0.036(-0.004)0.052(-0.012)0.021(-0.005)
Dρ↓ρt = N(0,1)ρt = U(0,1)
GF1IKQBWV2WV3WV4GF1IKQBWV2WV3WV4
GF10.043(+0.003)0.110(+0.030)0.026(-0.008)0.079(+0.032)0.062(+0.015)0.041(-0.002)0.019(-0.020)0.059(-0.021)0.022(-0.013)0.015(-0.032)0.017(-0.029)0.017(-0.026)
IK0.082(-0.006)0.095(-0.038)0.090(-0.058)0.126(+0.024)0.110(-0.010)0.061(-0.008)0.064(-0.024)0.092(-0.040)0.066(-0.083)0.051(-0.049)0.061(-0.059)0.045(-0.024)
QB0.061(-0.003)0.154(-0.020)0.179(-0.059)0.135(+0.011)-0.133(-0.045)-0.093(-0.026)-0.164(-0.029)-0.235(-0.002)-0.200(-0.058)-0.100(-0.013)-0.126(-0.048)-0.112(-0.007)
WV2-1.166(-0.27)-1.198(-0.339)-1.199(-0.539)-1.155(-0.459)-1.133(-0.455)-1.137(-0.451)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)-1.173(-0.477)-1.166(-0.477)-1.166(-0.477)
WV3-2.205(+0.18)-2.214(-0.41)-2.234(-0.533)-2.231(-0.419)-2.173(+0.11)-2.173(-0.415)-2.176(-0.419)-2.226(-0.429)-2.219(-0.468)-2.117(-0.434)-1.131(-0.439)-1.131(-0.439)
WV4-3.155(-0.55)-3.112(-0.42)-3.115(-0.44)-3.146(+0.88)-3.169(+0.77)-3.169(-0.477)-3.173(-0.477)-3.176(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)-3.173(-0.477)
", + "image_path": "28d2467620b790e9cf13a464cb50438888716a5e9fd946cba3d998ef02912e8e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 333, + 555, + 366 + ], + "lines": [ + { + "bbox": [ + 55, + 333, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 333, + 555, + 366 + ], + "type": "text", + "content": "Table 1. Average results of 12 models. Rows correspond to training satellites, and columns correspond to testing satellites. Green indicates performance improvement, while red indicates performance degradation. The values in parentheses represent the average delta compared to the baseline without UniPAN." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": "The function " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "\\text{interp}" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": " represents a linear interpolation between the quantile point " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": " and the reference point " + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 376, + 297, + 413 + ], + "type": "text", + "content": ", with the interpolation formula given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 418, + 302, + 472 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 418, + 302, + 472 + ], + "spans": [ + { + "bbox": [ + 56, + 418, + 302, + 472 + ], + "type": "interline_equation", + "content": "\\operatorname {i n t e r p} (x, Q, U) = \\left\\{ \\begin{array}{l l} U [ 0 ], & x \\leq Q [ 0 ] \\\\ f (x, i, Q, U), & Q [ i ] < x \\leq Q [ i + 1 ] \\\\ U [ m - 1 ], & x > Q [ m - 1 ] \\end{array} \\right. \\tag {9}", + "image_path": "2dd4b20241128589e2a618e794a62dd910eae8b518e9846138318f4ff1ef7781.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "text", + "content": " computes the " + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "inline_equation", + "content": "U" + }, + { + "bbox": [ + 55, + 472, + 217, + 484 + ], + "type": "text", + "content": " mapping," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 488, + 296, + 524 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 488, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 58, + 488, + 296, + 524 + ], + "type": "interline_equation", + "content": "f (x, i, Q, U) = U [ i ] + \\frac {x - Q [ i ]}{Q [ i + 1 ] - Q [ i ]} (U [ i + 1 ] - U [ i ]) \\tag {10}", + "image_path": "55ab71db73cc559b6c20208ee3ff0a445e48e80fca1b0d344468a01a1a85b770.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 525, + 296, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 525, + 296, + 537 + ], + "spans": [ + { + "bbox": [ + 67, + 525, + 296, + 537 + ], + "type": "text", + "content": "The last step is to apply inverse CDF of target distribution," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 133, + 547, + 296, + 562 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 547, + 296, + 562 + ], + "spans": [ + { + "bbox": [ + 133, + 547, + 296, + 562 + ], + "type": "interline_equation", + "content": "\\mathcal {T} (x) = F _ {T} ^ {- 1} (u (x)) \\tag {11}", + "image_path": "10501abd913ea06dd5748a4646e809c1fdb9e0c602aeef7b85d65d21eff3fc13.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "text", + "content": "Training and Testing Procedure. The training and testing procedures of UniPAN are described in Algorithm 1. Here, " + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "inline_equation", + "content": "L \\uparrow" + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "text", + "content": " denotes the upsampled LRMS image using bicubic interpolation to match the size of the PAN image. The model " + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "text", + "content": " is trained to predict the residual of LRMS with an " + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 55, + 563, + 296, + 647 + ], + "type": "text", + "content": " loss function that minimizes the discrepancy between the predicted HRMS image and the ground truth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": "As for the target distribution " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\rho_{t}" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": ", we test two widely used distributions, i.e. standard normal distribution " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": " and standard uniform distribution " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,1)" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": ", the transformation " + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{T}(x)" + }, + { + "bbox": [ + 55, + 647, + 296, + 696 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 90, + 700, + 296, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 700, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 90, + 700, + 296, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {T} _ {\\mathrm {n}} (x) = \\Phi^ {- 1} \\left(F _ {S} (x)\\right), \\quad \\mathcal {T} _ {\\mathrm {u}} (x) = F _ {S} (x), \\tag {12}", + "image_path": "be50aa8becae2e2209d0fd5b448c68cda95ee3eb91fa3025b71c94db8cb585cd.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\Phi^{-1}" + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": " denotes the inverse CDF of standard normal distribution. Since the inverse CDF of standard uniform distribution is an identical mapping, " + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{u}}" + }, + { + "bbox": [ + 313, + 376, + 555, + 461 + ], + "type": "text", + "content": " becomes the estimated empirical distribution itself. The standard normal distribution transformation non-linearly warps the input space to match Gaussian statistics, while the uniform transformation directly maps pixel intensities to their empirical quantiles." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 472, + 390, + 486 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 472, + 390, + 486 + ], + "spans": [ + { + "bbox": [ + 314, + 472, + 390, + 486 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 491, + 368, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 491, + 368, + 504 + ], + "spans": [ + { + "bbox": [ + 314, + 491, + 368, + 504 + ], + "type": "text", + "content": "4.1. Setups" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "text", + "content": "Datasets. We conducted experiments utilizing the Red, Green, Blue, and NIR bands of the NBUPansharpRSDataset dataset [49]. The variations in resolution among different satellites, coupled with significant differences in land covers and land uses, render this dataset particularly suitable for assessing the generalization capabilities of pansharpening models. The dataset comprises data from 6 satellites. For each satellite, we partitioned the data into training, validation, and testing sets in a ratio of " + }, + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "inline_equation", + "content": "7:1:2" + }, + { + "bbox": [ + 313, + 510, + 555, + 616 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 714 + ], + "type": "text", + "content": "Implementation Details. To validate the effectiveness and generalization capability of UniPAN, we conducted comparative experiments between baseline (w/o UniPAN) and enhanced (w/ UniPAN) configurations using diverse commonly-recognized models, including FusionNet [19], PanNet [82], MSDDN [28, 29], GPPNN [80], MSDCNN [83], FeINFN [40], PNN [47], LAGConv [34], SFIIN [89, 90], PreMix [18], UAPN [87], MDCUN [81]." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "11854" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 64, + 544, + 366 + ], + "blocks": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "lines": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "type": "image", + "image_path": "3d5e7231ab00420a762b9cb2b8063163b09d86c7345ce42bb2458470f4ae8ebf.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "text", + "content": "Figure 3. Generalization testing on GaoFen-1, the networks are trained on WorldView-3. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the " + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "inline_equation", + "content": "D_{\\lambda} / D_{S}" + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "text", + "content": " value. In most cases, the use of UniPAN yields superior spectral and spatial preservation. For instance, MDCUN [81] exhibits a highly close alignment with the LRMS in spectral characteristics with UniPAN of " + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "inline_equation", + "content": "\\rho_t = \\mathcal{N}(0,1)" + }, + { + "bbox": [ + 54, + 376, + 555, + 421 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "spans": [ + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "text", + "content": "All implementations followed identical experimental procedures: a batch size of 8 (except for MDCUN [81] and FeINFN [40] which used 4), the Adam [35] optimizer " + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "inline_equation", + "content": "(\\beta_{1} = 0.9, \\beta_{2} = 0.999)" + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "text", + "content": " with an initial learning rate of " + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "inline_equation", + "content": "10^{-3}" + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "text", + "content": ", and StepLR scheduling which reduces the learning rate by a factor of 0.8 every 100 epochs. Training was performed on an NVIDIA GeForce RTX 4090 GPU for 600 epochs using the Wald protocol [69] with " + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 54, + 433, + 297, + 567 + ], + "type": "text", + "content": " loss, with best model selected based on optimal validation performance. Note that, to prevent data leakage of testing set, the transformation fitting process is conducted on the training set of each satellite." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "text", + "content": "Evaluation Metrics. The full-reference metrics based on spatial degradation (e.g. PSNR, SSIM) can only reflect performance in simulated reduced-resolution experiments and fail to demonstrate effectiveness in real-world full-resolution scenarios [6, 16, 46]. In contrast, no-reference metrics like QNR explicitly quantify model performance in both spectral preservation and spatial enhancement, enabling direct evaluation of fusion results at native resolution while avoiding theoretical biases introduced by downsampling. To better align with practical applications where reference images are unavailable, this paper adopts widely-used no-reference spectral and spatial metrics, i.e. QNR [5], " + }, + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "inline_equation", + "content": "D_{\\lambda}" + }, + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "text", + "content": " [5], " + }, + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "inline_equation", + "content": "D_{S}" + }, + { + "bbox": [ + 54, + 570, + 298, + 715 + ], + "type": "text", + "content": " [5]," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 433, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 458 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 313, + 433, + 555, + 458 + ], + "type": "inline_equation", + "content": "D_{\\rho}" + }, + { + "bbox": [ + 313, + 433, + 555, + 458 + ], + "type": "text", + "content": " [60], to comprehensively assess cross-resolution and cross-satellite generalization capabilities." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 467, + 450, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 450, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 450, + 480 + ], + "type": "text", + "content": "4.2.Quantitative Evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 312, + 485, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 485, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 312, + 485, + 556, + 628 + ], + "type": "text", + "content": "Tab. 1 presents the average generalization performance of various models enhanced with UniPAN under different target distributions. Green indicate metric improvements, while red denote performance degradation. Slight yet observable differences exist between the two target distributions. Overall, the uniform distribution achieves better performance than the normal distribution. The normal distribution demonstrated a modest decline in performance across the WV2 and WV3 datasets. This phenomenon may stem from the peaked characteristics of the normal distribution, which could hinder model training, whereas the uniform distribution provides smoother normalization (see Fig. 5 in Sec. 4.3)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "text", + "content": "Notably, UniPAN with " + }, + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\rho_{t} = \\mathcal{U}(0,1)" + }, + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "text", + "content": " achieves significant improvements across most metrics, with only marginal declines in rare cases. For instance, when generalizing from WV3 to QB, UniPAN improves the QNR metric by over 0.084 (from 0.729 to 0.813) averaged on all 12 models, underscoring its effectiveness. Beyond QNR, UniPAN also enhances spectral preservation " + }, + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "inline_equation", + "content": "(D_{\\lambda})" + }, + { + "bbox": [ + 312, + 630, + 557, + 715 + ], + "type": "text", + "content": " and spatial consistency" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11855" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 64, + 544, + 366 + ], + "blocks": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "lines": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "spans": [ + { + "bbox": [ + 67, + 64, + 544, + 366 + ], + "type": "image", + "image_path": "7094b3dd5c36ccab6dfc0ce1cfc6fec32a0db01927e72b11d194d9ccb4062908.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 376, + 555, + 399 + ], + "lines": [ + { + "bbox": [ + 54, + 376, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 555, + 399 + ], + "type": "text", + "content": "Figure 4. Full-resolution generalization testing on QuickBird. The bright green box indicates the zoomed-in area. Below the zoomed-in image, the first line is the QNR value, and the second line is the " + }, + { + "bbox": [ + 54, + 376, + 555, + 399 + ], + "type": "inline_equation", + "content": "D_{\\lambda} / D_{S}" + }, + { + "bbox": [ + 54, + 376, + 555, + 399 + ], + "type": "text", + "content": " value." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "(D_S, D_\\rho)" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": ", further validating the efficacy of the unified distribution strategy. These results confirm that aligning heterogeneous satellite data to a unified distribution effectively mitigates domain shifts, enabling robust cross-sensor generalization. However, while UniPAN improves spatial fidelity greatly, it may cause slight spectral distortion, as evidenced by the minor performance degradation observed in the " + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "inline_equation", + "content": "D_{\\lambda}" + }, + { + "bbox": [ + 54, + 411, + 297, + 555 + ], + "type": "text", + "content": " metric. This is reasonable since the transformation applied to each spectral band maintains the relative spatial structure but might result in minimal loss of spectral information. Nevertheless, such information loss is extremely limited, as the overall enhancement in performance remains predominant." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 564, + 168, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 168, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 168, + 577 + ], + "type": "text", + "content": "4.3. Visual Comparison" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 713 + ], + "type": "text", + "content": "Comparison of Fusion Results. To illustrate the practical impact of UniPAN, Figs. 3 and 4 compare fusion results w/ and w/o UniPAN. In most cases, UniPAN achieves highly favorable results in both spectral and spatial preservation. Furthermore, the remarkably high QNR values demonstrate its superior capability in balancing spectral and spatial fidelity, e.g. in Fig. 3, MDCUN [81] applying UniPAN with normal distribution enhanced the QNR by 0.1514 (from 0.7157 to 0.8671) and aligned fused results closely with the ground truth spectra. In the full-resolution generalization test depicted in Fig. 4, even with minimal visual differences," + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 555, + 471 + ], + "type": "text", + "content": "UniPAN achieves metrically significant improvements in spectral and spatial preservation. This capability is critical for downstream applications requiring accurate spectral information. Additional visual comparison results are included in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "content": "Comparison of Data Distribution. Fig. 5 further analyzes data distribution alignment. The kernel density estimation (KDE) curves (top right area and diagonals) reveal that UniPAN successfully transforms the original complex and heterogeneous distributions into a unified target distribution. The quantile-quantile (Q-Q) plots (bottom left area) demonstrate harmonized distributions across satellites after applying UniPAN (The greater the deviation from the " + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "content": " line, the more significant the disparity between the two distributions). For example, the original Q-Q plots between WV4 and QB show significant dispersion, whereas post-UniPAN points close tightly around " + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 313, + 472, + 556, + 628 + ], + "type": "text", + "content": ". This visual evidence corroborates UniPAN's ability to bridge distribution gaps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 635, + 430, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 430, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 430, + 649 + ], + "type": "text", + "content": "4.4. Parameter Analysis" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "UniPAN introduces two hyperparameters: the number of sampled pixels " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " and the number of quantiles " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": ". Figs. 6 and 7 investigate their impact on performance of FusionNet [19] trained on WV2 and tested on GF1. As " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " vary, the metrics stabilize. Generally, " + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{U}(0,1)" + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": " demonstrates" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11856" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 71, + 217, + 231 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 217, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 217, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 217, + 231 + ], + "type": "image", + "image_path": "8c87594f3788d04fa0663c5f741b7a96fcc3f177ea1bdfaf0fd4bddbde54473a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 232, + 164, + 241 + ], + "lines": [ + { + "bbox": [ + 109, + 232, + 164, + 241 + ], + "spans": [ + { + "bbox": [ + 109, + 232, + 164, + 241 + ], + "type": "text", + "content": "(a) w/o UniPAN" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 71, + 383, + 231 + ], + "blocks": [ + { + "bbox": [ + 220, + 71, + 383, + 231 + ], + "lines": [ + { + "bbox": [ + 220, + 71, + 383, + 231 + ], + "spans": [ + { + "bbox": [ + 220, + 71, + 383, + 231 + ], + "type": "image", + "image_path": "2d479758e8371c95ac4f981afd8545de0e50873d93bc8c37d60e998c1b472acf.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 253, + 232, + 353, + 242 + ], + "lines": [ + { + "bbox": [ + 253, + 232, + 353, + 242 + ], + "spans": [ + { + "bbox": [ + 253, + 232, + 353, + 242 + ], + "type": "text", + "content": "(b) w/ UniPAN, " + }, + { + "bbox": [ + 253, + 232, + 353, + 242 + ], + "type": "inline_equation", + "content": "\\rho_t = \\mathcal{N}(0,1)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 72, + 550, + 230 + ], + "blocks": [ + { + "bbox": [ + 391, + 72, + 550, + 230 + ], + "lines": [ + { + "bbox": [ + 391, + 72, + 550, + 230 + ], + "spans": [ + { + "bbox": [ + 391, + 72, + 550, + 230 + ], + "type": "image", + "image_path": "524d7929265bed16ce9a843829177ce385f263449f98df7282245d38ecc97d0f.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 422, + 232, + 518, + 242 + ], + "lines": [ + { + "bbox": [ + 422, + 232, + 518, + 242 + ], + "spans": [ + { + "bbox": [ + 422, + 232, + 518, + 242 + ], + "type": "text", + "content": "(c) " + }, + { + "bbox": [ + 422, + 232, + 518, + 242 + ], + "type": "inline_equation", + "content": "\\mathrm{w} / \\mathrm{UniPAN}" + }, + { + "bbox": [ + 422, + 232, + 518, + 242 + ], + "type": "inline_equation", + "content": "\\rho_t = \\mathcal{U}(0,1)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 52, + 298, + 289, + 484 + ], + "blocks": [ + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "lines": [ + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": "Figure 5. Comparison of data distributions before and after applying UniPAN. The upper-right and diagonals of each subplot display the kernel density estimation (KDE) curves of Green band for each satellite. The scatter plots in the lower-left areas of each subplots represent quantile-quantile (Q-Q) plots of green band, with the dashed line indicating " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": ". The greater the deviation from the " + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "inline_equation", + "content": "y = x" + }, + { + "bbox": [ + 54, + 246, + 555, + 291 + ], + "type": "text", + "content": " line, the more significant the disparity between the two distributions. Zoom in for better view." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 52, + 298, + 289, + 484 + ], + "lines": [ + { + "bbox": [ + 52, + 298, + 289, + 484 + ], + "spans": [ + { + "bbox": [ + 52, + 298, + 289, + 484 + ], + "type": "image", + "image_path": "956fc6ef24fd91914e2bd1eb223b55b7d72695991efa19372dd43972da283cb7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 488, + 295, + 511 + ], + "lines": [ + { + "bbox": [ + 55, + 488, + 295, + 511 + ], + "spans": [ + { + "bbox": [ + 55, + 488, + 295, + 511 + ], + "type": "text", + "content": "Figure 6. Parameter analysis of number of quantiles, with number of sampled pixels fixed to 10000." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 298, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 298, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 298, + 594 + ], + "type": "text", + "content": "superior performance compared to " + }, + { + "bbox": [ + 55, + 522, + 298, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 522, + 298, + 594 + ], + "type": "text", + "content": ". Both configurations achieve statistically significant improvements over baseline methods without applying UniPAN. This parametric robustness demonstrates that optimal performance can be attained without exhaustive hyperparameter tuning, significantly reducing computational overhead." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 608, + 208, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 608, + 208, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 608, + 208, + 621 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "content": "We proposed UniPAN, a unified distribution strategy addressing generalization challenges in deep learning-based pansharpening. UniPAN eliminates spectral discrepancies by normalizing multi-source remote sensing imagery into predefined target distributions, effectively bridging the gap between training and testing distributions. Extensive quantitative and visual comparisons across 6 satellite sensors and" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 312, + 298, + 545, + 484 + ], + "blocks": [ + { + "bbox": [ + 312, + 298, + 545, + 484 + ], + "lines": [ + { + "bbox": [ + 312, + 298, + 545, + 484 + ], + "spans": [ + { + "bbox": [ + 312, + 298, + 545, + 484 + ], + "type": "image", + "image_path": "a8b78f92931d5d2121022107a706625190cffd080196318790c7203579d9b664.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 488, + 555, + 511 + ], + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 511 + ], + "type": "text", + "content": "Figure 7. Parameter analysis of number of sampled pixels, with number of quantiles fixed to 1000." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 569 + ], + "type": "text", + "content": "12 models demonstrate that UniPAN significantly enhances cross-sensor generalization. Notably, this enhancement is achieved without introducing any additional trainable parameters or the need for retraining or fine-tuning the models." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "While this work aligns theoretically with 1D optimal transport, considering interdependencies of spectral bands or spatial patches, future research could explore extending the method to the multi-dimensional case with optimal transport theories. Additionally, the experimental results indicate that the choice of target distribution has a notable impact on performance, thus it is worthwhile investigating adaptive target domain selection. To further validate the generalizability of the proposed unified distribution strategy, the ideas presented in this work could be extended to other remote sensing image processing tasks, such as multi-modal fusion and hyperspectral pansharpening, etc." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11857" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "Acknowledgment. This research was partially supported by the National Natural Science Foundation of China under Grant 32441114, 41971397, 61731022, and the project Y1H103101A." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "spans": [ + { + "bbox": [ + 56, + 131, + 115, + 144 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 152, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 152, + 297, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 297, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 297, + 194 + ], + "type": "text", + "content": "[1] B. Aiazzi, L. Alparone, S. Baronti, and A. Garzelli. Context-driven fusion of high spatial and spectral resolution images based on oversampled multiresolution analysis. IEEE TGRS, 40(10):2300-2312, 2002. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 197, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 197, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 296, + 262 + ], + "type": "text", + "content": "[2] B. Aiazzi, L. Alparone, S. Baronti, A. Garzelli, and M. Selva. An MTF-based spectral distortion minimizing model for pan-sharpening of very high resolution multispectral images of urban areas. In 2003 2nd GRSS/ISPRS Joint Workshop on Remote Sensing and Data Fusion over Urban Areas, pages 90-94, 2003." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 62, + 263, + 296, + 306 + ], + "type": "text", + "content": "[3] Bruno Aiazzi, L. Alparone, Stefano Baronti, Andrea Garzelli, and Massimo Selva. MTF-tailored multiscale fusion of high-resolution MS and Pan imagery. *Photogramm. Eng. Remote Sens.*, 72:591-596, 2006. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 308, + 296, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 296, + 351 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 296, + 351 + ], + "type": "text", + "content": "[4] Bruno Aiazzi, Stefano Baronti, and Massimo Selva. Improving component substitution pansharpening through multivariate regression of MS+Pan data. IEEE TGRS, 45(10): 3230-3239, 2007. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 296, + 397 + ], + "type": "text", + "content": "[5] Luciano Alparone, Bruno Aiazzi, Stefano Baronti, Andrea Garzelli, Filippo Nencini, and Massimo Selva. Multispectral and panchromatic data fusion assessment without reference. *Photogramm. Eng. Remote Sens.*, 74(2):193-200, 2008. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "text", + "content": "[6] Alberto Arienzo, Gemine Vivone, Andrea Garzelli, Luciano Alparone, and Jocelyn Chanussot. Full-resolution quality assessment of pansharpening: Theoretical and hands-on approaches. IEEE GRSM, 10(3):168-201, 2022. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 444, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 475 + ], + "type": "text", + "content": "[7] Coloma Ballester, Vicent Caselles, and Laura Igual. A variational model for " + }, + { + "bbox": [ + 62, + 444, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{P + XS}" + }, + { + "bbox": [ + 62, + 444, + 296, + 475 + ], + "type": "text", + "content": " image fusion. IJCV, 69(1):43-58, 2006. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 477, + 296, + 520 + ], + "type": "text", + "content": "[8] Wele Gedara Chaminda Bandara and Vishal M Patel. Hypertransformer: A textural and spectral feature fusion Transformer for pansharpening. In CVPR, pages 1767-1777, 2022, 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 522, + 296, + 555 + ], + "type": "text", + "content": "[9] Qi Cao, Liang-Jian Deng, Wu Wang, Junming Hou, and Gemine Vivone. Zero-shot semi-supervised learning for pansharpening. Inf. Fusion., 101:102001, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "spans": [ + { + "bbox": [ + 57, + 557, + 296, + 600 + ], + "type": "text", + "content": "[10] Pat S. Chavez and Andy Yaw Kwarteng. Extracting spectral contrast in landsat thematic mapper image data using selective principal component analysis. Photogramm. Eng. Remote Sens., 55:339-348, 1989. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 296, + 644 + ], + "type": "text", + "content": "[11] Lihui Chen, Gemine Vivone, Zihao Nie, Jocelyn Chanussot, and Xiaomin Yang. Spatial data augmentation: Improving the generalization of neural networks for pansharpening. IEEE TGRS, 61:1-11, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 296, + 689 + ], + "type": "text", + "content": "[12] Zhi-Xuan Chen, Cheng Jin, Tian-Jing Zhang, Xiao Wu, and Liang-Jian Deng. SpanConv: A new convolution via spanning kernel space for lightweight pansharpening. In IJCAI, pages 841-847, 2022. Main Track. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 714 + ], + "type": "text", + "content": "[13] Minhyeok Choi, Rae Young Kim, and Myeong-Ryong Nam. Fusion of multispectral and panchromatic satellite images" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "using the curvelet transform. IEEE GRSL, 2(2):136-140, 2005. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 140 + ], + "type": "text", + "content": "[14] Matteo Ciotola, Sergio Vitale, Antonio Mazza, Giovanni Poggi, and Giuseppe Scarpa. Pansharpening by convolutional neural networks in the full resolution framework. IEEE TGRS, pages 1-1, 2022. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 143, + 555, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 555, + 185 + ], + "type": "text", + "content": "[15] Matteo Ciotola, Giovanni Poggi, and Giuseppe Scarpa. Unsupervised deep learning-based pansharpening with jointly enhanced spectral and spatial fidelity. IEEE TGRS, 61:1-17, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 188, + 555, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 188, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 188, + 555, + 231 + ], + "type": "text", + "content": "[16] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pixel-wise ensembled masked autoencoder for multispectral pansharpening. IEEE TGRS, 62:1-22, 2024. 3, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 234, + 555, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 555, + 277 + ], + "type": "text", + "content": "[17] Yongchuan Cui, Peng Liu, Bingze Song, Lingjun Zhao, Yan Ma, and Lajiao Chen. Reconstruction of large-scale missing data in remote sensing images using Extend-GAN. IEEE GRSL, 21:1-5, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 280, + 555, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 280, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 280, + 555, + 323 + ], + "type": "text", + "content": "[18] Yongchuan Cui, Peng Liu, Yan Ma, Lajiao Chen, Mengzhen Xu, and Xingyan Guo. Pansharpening via predictive filtering with element-wise feature mixing. ISPRS J. Photogramm. Remote Sens., 219:22-37, 2025. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 326, + 555, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 555, + 369 + ], + "type": "text", + "content": "[19] Liang-Jian Deng, Gemine Vivone, Cheng Jin, and Jocelyn Chanussot. Detail injection-based deep convolutional neural networks for pansharpening. IEEE TGRS, 59(8):6995-7010, 2021. 1, 2, 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 372, + 555, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 555, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 555, + 425 + ], + "type": "text", + "content": "[20] Liang-jian Deng, Gemine Vivone, Mercedes E. Paoletti, Giuseppe Scarpa, Jiang He, Yongjun Zhang, Jocelyn Chanussot, and Antonio Plaza. Machine learning in pansharpening: A benchmark, from shallow to deep networks. IEEE GRSM, 10(3):279-315, 2022. 1, 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 428, + 554, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 554, + 461 + ], + "type": "text", + "content": "[21] Shangqi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, and Rui Wen. Bidirectional dilation Transformer for multispectral and hyperspectral image fusion. In IJCAI, 2023. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 463, + 555, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 463, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 463, + 555, + 506 + ], + "type": "text", + "content": "[22] Shang-Qi Deng, Liang-Jian Deng, Xiao Wu, Ran Ran, Danfeng Hong, and Gemine Vivone. PSRT: Pyramid shuffle-and-reshuffle Transformer for multispectral and hyperspectral image fusion. IEEE TGRS, 61:1-15, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 509, + 555, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 555, + 540 + ], + "type": "text", + "content": "[23] Renwei Dian, Anjing Guo, and Shutao Li. Zero-shot hyperspectral sharpening. IEEE TPAMI, 45(10):12650-12666, 2023. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 543, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 543, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 543, + 555, + 609 + ], + "type": "text", + "content": "[24] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 611, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 555, + 644 + ], + "type": "text", + "content": "[25] Yule Duan, Xiao Wu, Haoyu Deng, and Liang-Jian Deng. Content-adaptive non-local convolution for remote sensing pansharpening. In CVPR, pages 27738-27747, 2024. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 646, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 646, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 646, + 555, + 678 + ], + "type": "text", + "content": "[26] Andrea Garzelli and Filippo Nencini. Wavelet-based fusion of optical and SAR image data in remote sensing applications. Int. J. Remote Sens., 23(23):5025-5039, 2002. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 555, + 713 + ], + "type": "text", + "content": "[27] Morteza Ghahremani and Hassan Ghassemian. Nonlinear IHS: A promising method for pan-sharpening. IEEE GRSL, 13(11):1606-1610, 2016. 2" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11858" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[28] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Pyramid dual domain injection network for pan-sharpening. In ICCV, pages 12862-12871, 2023. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[29] Xuanhua He, Keyu Yan, Jie Zhang, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Multiscale dual-domain guidance network for pan-sharpening. IEEE TGRS, 61:1-13, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 185 + ], + "type": "text", + "content": "[30] Xuanhua He, Keyu Yan, Rui Li, Chengjun Xie, Jie Zhang, and Man Zhou. Frequency-adaptive pan-sharpening with mixture of experts. AAAI, 38(3):2121-2129, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 229 + ], + "type": "text", + "content": "[31] Xuanhua He, Ke Cao, Jie Zhang, Keyu Yan, Yingying Wang, Rui Li, Chengjun Xie, Danfeng Hong, and Man Zhou. PanMamba: Effective pan-sharpening with state space model. Inf.Fusion., 115:102779, 2025. 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 274 + ], + "type": "text", + "content": "[32] Junming Hou, Xinyang Liu, Chenxu Wu, Xiaofeng Cong, Chentong Huang, Liang-Jian Deng, and Jian Wei You. Bidomain uncertainty gated recursive network for pan-sharpening. Inf. Fusion., 118:102938, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 296, + 319 + ], + "type": "text", + "content": "[33] Cheng Jin, Liang-Jian Deng, Ting-Zhu Huang, and Gemine Vivone. Laplacian pyramid networks: A new approach for multispectral pansharpening. Inf. Fusion., 78:158-170, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 296, + 364 + ], + "type": "text", + "content": "[34] Zi-Rong Jin, Tian-Jing Zhang, Tai-Xiang Jiang, Gemine Vivone, and Liang-Jian Deng. LAGConv: Local-context adaptive convolution kernels with global harmonic bias for pan-sharpening. AAAI, 36(1):1113-1121, 2022. 3, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 397 + ], + "type": "text", + "content": "[35] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 296, + 443 + ], + "type": "text", + "content": "[36] Soheil Kolouri, Se Rim Park, Matthew Thorpe, Dejan Slepcev, and Gustavo K. Rohde. Optimal mass transport: Signal processing and machine-learning applications. IEEE SPM, 34 (4):43-59, 2017. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 296, + 476 + ], + "type": "text", + "content": "[37] Soheil Kolouri, Kimia Nadjahi, Umut Simsekli, Roland Badeau, and Gustavo K. Rohde. Generalized sliced wasserstein distances, 2019. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 477, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 477, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 477, + 296, + 510 + ], + "type": "text", + "content": "[38] Craig A Laben and Bernard V Brower. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. 2000. US Patent 6,011,875. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 296, + 555 + ], + "type": "text", + "content": "[39] Zhixuan Li, Jinjiang Li, Fan Zhang, and Linwei Fan. CADUI: Cross-attention-based depth unfolding iteration network for pansharpening remote sensing images. IEEE TGRS, 61:1–20, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 296, + 610 + ], + "type": "text", + "content": "[40] YuJie Liang, ZiHan Cao, Shangqi Deng, Hong-Xia Dou, and Liang-Jian Deng. Fourier-enhanced implicit neural fusion network for multispectral and hyperspectral image fusion. In NeurIPS, pages 63441-63465. Curran Associates, Inc., 2024. 3, 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 612, + 296, + 645 + ], + "type": "text", + "content": "[41] J. G. Liu. Smoothing filter-based intensity modulation: A spectral preserve image fusion technique for improving spatial details. Int. J. Remote Sens., 21(18):3461-3472, 2000. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 296, + 689 + ], + "type": "text", + "content": "[42] Peng Liu, Jun Li, Lizhe Wang, and Guojin He. Remote sensing data fusion with generative adversarial networks: State-of-the-art methods and future research directions. IEEE GRSM, 10(2):295-328, 2022. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 690, + 296, + 714 + ], + "type": "text", + "content": "[43] Qingjie Liu, Huanyu Zhou, Qizhi Xu, Xiangyu Liu, and Yun-hong Wang. PSGAN: A generative adversarial network for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "remote sensing image pan-sharpening. IEEE TGRS, 59(12): 10227-10242, 2020. 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 140 + ], + "type": "text", + "content": "[44] Xiaobo Liu, Xiang Liu, Haoran Dai, Xudong Kang, Antonio Plaza, and Wenjie Zu. Mun-GAN: a multiscale unsupervised network for remote sensing image pansharpening. IEEE TGRS, 61:1-18, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 142, + 555, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 142, + 555, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 142, + 555, + 186 + ], + "type": "text", + "content": "[45] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin Transformer: Hierarchical vision Transformer using shifted windows. In ICCV, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 186, + 555, + 231 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 231 + ], + "type": "text", + "content": "[46] Jiayi Ma, Wei Yu, Chen Chen, Pengwei Liang, Xiaojie Guo, and Junjun Jiang. Pan-GAN: An unsupervised pan-sharpening method for remote sensing image fusion. Inf. Fusion., 62: 110-120, 2020. 2, 3, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 232, + 555, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 555, + 266 + ], + "type": "text", + "content": "[47] Giuseppe Masi, Davide Cozzolino, Luisa Verdoliva, and Giuseppe Scarpa. Pansharpening by convolutional neural networks. In Remote Sens., page 594, 2016. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 267, + 555, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 267, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 267, + 555, + 311 + ], + "type": "text", + "content": "[48] Qingyan Meng, Wenxu Shi, Sijia Li, and Linlin Zhang. PanDiff: A novel pansharpening method based on denoising diffusion probabilistic model. IEEE TGRS, 61:1-17, 2023. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 312, + 555, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 555, + 368 + ], + "type": "text", + "content": "[49] Xiangchao Meng, Yiming Xiong, Feng Shao, Huanfeng Shen, Weiwei Sun, Gang Yang, Qiangqiang Yuan, Randi Fu, and Hongyan Zhang. A large-scale benchmark data set for evaluating pansharpening performance: Overview and implementation. IEEE GRSM, 9(1):18-52, 2021. 1, 2, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 369, + 555, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 369, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 316, + 369, + 555, + 401 + ], + "type": "text", + "content": "[50] Xiangchao Meng, Nan Wang, Feng Shao, and Shutao Li. Vision Transformer for pansharpening. IEEE TGRS, 60:1-11, 2022. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 403, + 555, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 403, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 316, + 403, + 555, + 436 + ], + "type": "text", + "content": "[51] Gaspard Monge. Mémoire sur la théorie des déblais et des remblais. Mem. Math. Phys. Acad. Royale Sci., pages 666-704, 1781. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 437, + 555, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 555, + 471 + ], + "type": "text", + "content": "[52] Fjola Palsson, Johannes R. Sveinsson, and Magnus O. Ulfarsson. A new pansharpening algorithm based on total variation. IEEE GRSL, 11(1):318-322, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 472, + 555, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 555, + 506 + ], + "type": "text", + "content": "[53] Siran Peng, Xiangyu Zhu, Haoyu Deng, Liang-Jian Deng, and Zhen Lei. FusionMamba: Efficient remote sensing image fusion with state space model. IEEE TGRS, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 507, + 555, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 507, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 507, + 555, + 540 + ], + "type": "text", + "content": "[54] Gabriel Peyre and Marco Cuturi. Computational optimal transport. Found. Trends Mach. Learn., 11(5-6):355-607, 2019. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 541, + 555, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 541, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 316, + 541, + 555, + 574 + ], + "type": "text", + "content": "[55] Sheila Rahmani, Melissa Strait, Daria Merkurjev, Michael Moeller, and Todd Wittman. An adaptive IHS pan-sharpening method. IEEE GRSL, 7(4):746-750, 2010. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 576, + 555, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 576, + 555, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 576, + 555, + 609 + ], + "type": "text", + "content": "[56] Aaditya Ramdas, Nicolas Garcia, and Marco Cuturi. On Wasserstein two sample testing and related families of nonparametric tests, 2015. 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 611, + 555, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 611, + 555, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 611, + 555, + 643 + ], + "type": "text", + "content": "[57] Xiangyu Rui, Xiangyong Cao, Yining Li, and Deyu Meng. Variational zero-shot multispectral pansharpening. IEEE TGRS, 62:1-16, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 644, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 644, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 644, + 555, + 689 + ], + "type": "text", + "content": "[58] Xiangyu Rui, Xiangyong Cao, Li Pang, Zeyu Zhu, Zongsheng Yue, and Deyu Meng. Unsupervised hyperspectral pansharpening via low-rank diffusion model. Inf. Fusion., 107:102325, 2024. 2, 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 555, + 714 + ], + "type": "text", + "content": "[59] Filippo Santambrogio. One-dimensional issues, pages 59-85. Springer International Publishing, Cham, 2015. 4" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11859" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "text", + "content": "[60] Giuseppe Scarpa and Matteo Ciotola. Full-resolution quality assessment for pansharpening. Remote Sens., 14(8), 2022. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "type": "text", + "content": "[61] Vijay P. Shah, Nicolas H. Younan, and Roger L. King. An efficient pan-sharpening method via a combined adaptive PCA approach and contourlets. IEEE TGRS, 46(5):1323-1335, 2008. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 171 + ], + "type": "text", + "content": "[62] Xunyang Su, Jinjiang Li, and Zhen Hua. Transformer-based regression network for pansharpening remote sensing images. IEEE TGRS, 60:1-23, 2022. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 296, + 226 + ], + "type": "text", + "content": "[63] Weiwei Sun, Jun Zhou, Xiangchao Meng, Gang Yang, Kai Ren, and Jiangtao Peng. Coupled temporal variation information estimation and resolution enhancement for remote sensing spatial-temporal-spectral fusion. IEEE TGRS, 61: 1-18, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 259 + ], + "type": "text", + "content": "[64] Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE. J. Mach. Learn. Res., 9(86):2579-2605, 2008. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 296, + 293 + ], + "type": "text", + "content": "[65] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In NeurIPS, 2017. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 315 + ], + "type": "text", + "content": "[66] Cédric Villani et al. Optimal transport: old and new. Springer, 2008. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 360 + ], + "type": "text", + "content": "[67] Gemine Vivone, Miguel Simões, Mauro Dalla Mura, Rocco Restaino, José M. Bioucas-Dias, Giorgio A. Licciardi, and Jocelyn Chanussot. Pansharpening based on semiblind deconvolution. IEEE TGRS, 53(4):1997-2010, 2015. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 361, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 296, + 426 + ], + "type": "text", + "content": "[68] Gemine Vivone, Mauro Dalla Mura, Andrea Garzelli, Rocco Restaino, Giuseppe Scarpa, Magnus O. Ulfarsson, Luciano Alparone, and Jocelyn Chanussot. A new benchmark based on recent advances in multispectral pansharpening: Revisiting pansharpening with classical and emerging pansharpening methods. IEEE GRSM, 9(1):53-81, 2021. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 296, + 470 + ], + "type": "text", + "content": "[69] Lucien Wald, Thierry Ranchin, and Marc Mangolini. Fusion of satellite images of different spatial resolutions: Assessing the quality of resulting images. *Photogramm. Eng. Remote Sens.*, 63:691-699, 1997. 3, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 296, + 503 + ], + "type": "text", + "content": "[70] Hebaixu Wang, Meiqi Gong, Xiaoguang Mei, Hao Zhang, and Jiayi Ma. Deep unfolded network with intrinsic supervision for pan-sharpening. AAAI, 38(6):5419-5426, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 296, + 547 + ], + "type": "text", + "content": "[71] Hebaixu Wang, Hao Zhang, Xin Tian, and Jiayi Ma. Zero-Sharpen: A universal pansharpening method across satellites for reducing scale-variance gap via zero-shot variation. Inf. Fusion., 101:102003, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 296, + 591 + ], + "type": "text", + "content": "[72] Jun-Da Wang, Liang-Jian Deng, Chen-Yu Zhao, Xiao Wu, Hong-Ming Chen, and Gemine Vivone. Cascadic multireceptive learning for multispectral pansharpening. IEEE TGRS, 61:1-16, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "text", + "content": "[73] Yingying Wang, Xuanhua He, Yuhang Dong, Yunlong Lin, Yue Huang, and Xinghao Ding. Cross-modality interaction network for pan-sharpening. IEEE TGRS, 62:1-16, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": "[74] Zhong-Cheng Wu, Ting-Zhu Huang, Liang-Jian Deng, JinFan Hu, and Gemine Vivone. " + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "inline_equation", + "content": "\\mathrm{VO + }" + }, + { + "bbox": [ + 56, + 625, + 296, + 669 + ], + "type": "text", + "content": " Net: An adaptive approach using variational optimization and deep learning for panchromatic sharpening. IEEE TGRS, 60:1-16, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 296, + 713 + ], + "type": "text", + "content": "[75] Jin-Liang Xiao, Ting-Zhu Huang, Liang-Jian Deng, Zhong-Cheng Wu, Xiao Wu, and Gemine Vivone. Variational pan-sharpening based on coefficient estimation with nonlocal regression. IEEE TGRS, 61:1-15, 2023. 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 701 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "text", + "content": "[76] Guangxu Xie, Rencan Nie, Jinde Cao, He Li, and Jintao Li. A deep multiresolution representation framework for pansharpening. IEEE TGRS, 62:1-16, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[77] Yinghui Xing, Litao Qu, Shizhou Zhang, Jiapeng Feng, Xi-wei Zhang, and Yanning Zhang. Empower generalizability for pansharpening through text-modulated diffusion model. IEEE TGRS, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[78] Yinghui Xing, Litao Qu, Shizhou Zhang, Kai Zhang, Yanning Zhang, and Lorenzo Bruzzone. CrossDiff: Exploring self-supervised representation of pansharpening via cross-predictive diffusion model. IEEE TIP, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 555, + 239 + ], + "type": "text", + "content": "[79] Qizhi Xu, Yuan Li, Jinyan Nie, Qingjie Liu, and Mengyao Guo. UPanGAN: Unsupervised pansharpening based on the spectral and spatial loss constrained generative adversarial network. Inf. Fusion., 91:31-46, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 554, + 274 + ], + "type": "text", + "content": "[80] Shuang Xu, Jiangshe Zhang, Zixiang Zhao, Kai Sun, Junmin Liu, and Chunxia Zhang. Deep gradient projection networks for pan-sharpening. In CVPR, pages 1366-1375, 2021. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 555, + 319 + ], + "type": "text", + "content": "[81] Gang Yang, Man Zhou, Keyu Yan, Aiping Liu, Xueyang Fu, and Fan Wang. Memory-augmented deep conditional unfolding network for pansharpening. In CVPR, pages 1778-1787, 2022. 3, 5, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 320, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 555, + 354 + ], + "type": "text", + "content": "[82] Jie Yang, Xiao-Yang Fu, Yan Hu, Yue Huang, Xing Ding, and John Paisley. PanNet: A deep network architecture for pan-sharpening. ICCV, pages 1753-1761, 2017. 2, 3, 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[83] Qiangqiang Yuan, Yancong Wei, Xiangchao Meng, Huanfeng Shen, and Liangpei Zhang. A multiscale and multidepth convolutional neural network for remote sensing imagery pan-sharpening. IEEE JSTARS, 11(3):978-989, 2018. 2, 3, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 442 + ], + "type": "text", + "content": "[84] Jie Zhang, Xuanhua He, Keyu Yan, Ke Cao, Rui Li, Chengjun Xie, Man Zhou, and Danfeng Hong. Pan-sharpening with wavelet-enhanced high-frequency information. IEEE TGRS, 62:1-14, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 444, + 555, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 555, + 487 + ], + "type": "text", + "content": "[85] Chen-Yu Zhao, Tian-Jing Zhang, Ran Ran, Zhi-Xuan Chen, and Liang-Jian Deng. LGPCConv: Learnable gaussian perturbation convolution for lightweight pansharpening. In IJCAI, pages 4647-4655, 2023. Main Track. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 555, + 531 + ], + "type": "text", + "content": "[86] Mengjiao Zhao, Mengting Ma, Xiangdong Li, Xiaowen Ma, Wei Zhang, and Siyang Song. Supervised detail-guided multiscale state space model for pan-sharpening. IEEE TGRS, 2024. 2, 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 533, + 555, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 555, + 566 + ], + "type": "text", + "content": "[87] Kaiwen Zheng, Jie Huang, Man Zhou, Danfeng Hong, and Feng Zhao. Deep adaptive pansharpening via uncertainty-aware image fusion. IEEE TGRS, 61:1-15, 2023. 5" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 610 + ], + "type": "text", + "content": "[88] Yu Zhong, Xiao Wu, Liang-Jian Deng, ZiHan Cao, and Hong-Xia Dou. SSDiff: Spatial-spectral integrated diffusion model for remote sensing pansharpening. In NeurIPS, pages 77962-77986. Curran Associates, Inc., 2024. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 612, + 554, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 554, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 554, + 656 + ], + "type": "text", + "content": "[89] Man Zhou, Jie Huang, Keyu Yan, Hu Yu, Xueyang Fu, Aiping Liu, Xian Wei, and Feng Zhao. Spatial-frequency domain information integration for pan-sharpening. In ECCV, pages 274-291, Cham, 2022. Springer Nature Switzerland. 3, 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 657, + 555, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 555, + 701 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 555, + 701 + ], + "type": "text", + "content": "[90] Man Zhou, Jie Huang, Keyu Yan, Danfeng Hong, Xiuping Jia, Jocelyn Chanussot, and Chongyi Li. A general spatial-frequency learning framework for multimodal image fusion. IEEE TPAMI, pages 1-18, 2024. 3, 5" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11860" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_content_list.json b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..75635a6f59cff0c1398629b745dcaadf784c0996 --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_content_list.json @@ -0,0 +1,1736 @@ +[ + { + "type": "text", + "text": "Enrich and Detect: Video Temporal Grounding with Multimodal LLMs", + "text_level": 1, + "bbox": [ + 259, + 130, + 738, + 172 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shraman Pramanick $^{\\dagger,2}$ Effrosyni Mavroudi $^{1}$ Yale Song $^{1}$ Rama Chellappa $^{2}$ Lorenzo Torresani $^{\\diamond 1}$ Triantafyllos Afouras $^{1}$", + "bbox": [ + 166, + 176, + 828, + 213 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ FAIR, Meta $^{2}$ Johns Hopkins University $^{3}$ Northeastern University", + "bbox": [ + 225, + 213, + 764, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 252, + 326, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios.", + "bbox": [ + 88, + 289, + 483, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 648, + 220, + 665 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Video temporal grounding [32, 40, 53, 80, 102] aims to identify temporal intervals in a video that correspond to a set of provided language queries. The task is essential for applications such as video editing and content retrieval. Conversely, video captioning [1, 59, 78, 79, 104] entails generating a natural language description for a given video segment, effectively translating visual content into text. The two tasks are in fact dual, as the outputs of one task are the inputs to the other, and vice versa. Intuitively, there is significant potential in exploiting this synergy, however it has largely remained unexplored, with previous works typically specializing in one of the two tasks [8, 32, 53, 104] or solv-", + "bbox": [ + 89, + 667, + 482, + 849 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f87cd08c5ac3e0806cb62d1dcb85321e4020c190ba7d7613f840bab5a47ea8e8.jpg", + "image_caption": [ + "Figure 1. Our proposed system. ED-VTG performs video temporal grounding as a two-stage process: since user queries are often incomplete or coarse, the first stage involves producing an enriched query that adds additional details to the original, making it easier to ground. Meanwhile, a contextualized embedding is generated, containing all the information about the interval to be predicted. In the second stage, an interval decoder translates these embeddings into precise temporal boundaries." + ], + "image_footnote": [], + "bbox": [ + 517, + 252, + 903, + 428 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing them in multi-task setting [72, 85] without investigating how each task can benefit the other.", + "bbox": [ + 511, + 551, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we exploit this duality by leveraging captioning to enhance grounding. Our key observation is that natural language queries often lack the completeness or detail necessary for effective temporal localization; indeed, existing grounding datasets frequently contain poorly worded, coarse, and potentially incomplete queries. The quality and completeness of these queries however is crucial for the precision of the temporal grounding. A natural hypothesis, therefore, is that more detailed queries, which could be obtained via conditional captioning, can significantly enhance grounding accuracy. For example, as illustrated in Figure 1, a vague query like 'Man starts surfing' can be refined into a more detailed description such as 'The man with a yellow surfboard slowly runs to start surfing,' resulting in grounding with more accurate temporal boundaries. In other cases, refining an abstract concept may involve breaking down a complex query into simpler components that are more directly groundable.", + "bbox": [ + 509, + 583, + 906, + 854 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This key idea of query enrichment forms the basis of our proposed approach. Concretely, we transform grounding into a two-stage reasoning process using a multi-modal", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Work done during an internship at FAIR.", + "bbox": [ + 112, + 862, + 339, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Work done while at FAIR.", + "bbox": [ + 112, + 875, + 263, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "spraman3@meta.com, afourast@meta.com", + "bbox": [ + 112, + 887, + 401, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24297", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "LLM: first the model enriches the input query into a more detailed description by adding missing details based on the video content, and then temporally localizes the resulting enriched query in the video.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To effectively perform the temporal localization, we introduce a lightweight perception decoder that specializes in generating precise temporal boundaries, conditioned on a contextual representation, allowing the LLM to focus on language outputs, where it excels. The perception decoder allows us to leverage purposefully crafted training objectives for temporal grounding, building on prior knowledge and task-specific characteristics developed from years of research in object detection [17, 19, 30, 67, 71, 77] and temporal localization [32, 40, 53, 61, 102, 105].", + "bbox": [ + 88, + 154, + 480, + 305 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning to jointly enrich and detect requires high-quality enriched query labels. We obtain those by using a strong external captioning model which we condition on the original queries and the video content of the target temporal boundaries. However powerful, these models are prone to hallucinations and there is no guarantee that the enriched queries will always be easier to ground than the original ones. At the same time, annotating the ground truth to determine which query - original or enriched - is a better candidate, is extremely expensive. To address this issue, we propose training in a multiple-instance learning (MIL) fashion, that enables the model to autonomously determine which query is better suited for the task during training.", + "bbox": [ + 89, + 306, + 482, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we note that our proposed method is not equivalent to a data augmentation approach which simply preprocesses the training set to generate enhanced queries that are directly used for training. While this simpler alternative offers some of the same benefits, it suffers from the limitation that extracting enriched queries during training requires knowledge of the ground-truth temporal segments, i.e., the grounding targets. Since during inference these segments are unknown, the original queries must be used as input, which, as we will show experimentally, is suboptimal. Our method overcomes this limitation by learning to jointly enrich and detect, demonstrating superior performance.", + "bbox": [ + 89, + 506, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, our contributions are as follows: (i) we introduce a cascaded approach to temporal grounding, where the model first enriches the provided language query based on the video context and then proceeds to localize it; (ii) we enable multi-modal LLMs to accurately localize text queries using a lightweight decoder which allows for training with detection objectives tailored to the task; (iii) we propose a multiple-instance learning paradigm that enables the model to dynamically select the query that leads to better temporal localization; (iv) we achieve state-of-the-art results on several temporal grounding benchmarks, for both single query grounding and paragraph grounding, demonstrating, for the first time, an LLM-based model that surpasses or performs comparably to specialist models.", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 89, + 663, + 106 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLM-based temporal grounding. Prior works have explored using LLMs for grounding natural language sentences in videos, either using raw text tokens to represent timestamps [21, 36, 39, 48, 72, 99] or by adding hundreds of special tokens to the LLM's vocabulary to represent video frames [22, 63, 85]. Our approach differs from these methods in that by utilizing a lightweight interval decoder we can apply detection losses such as L1 and gIoU with minimal added complexity; we additionally exploit the LLM's potential to describe video content in detail.", + "bbox": [ + 511, + 114, + 903, + 265 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specialist models. There is a rich variety of specialist models in the literature that are tailored to specific variants of temporal grounding, e.g. single query temporal grounding [102] and video paragraph grounding [82], and as such, achieve strong performance [6, 26, 41, 46, 75, 82, 98, 105, 106]. Modern methods typically employ a multi-modal transformer [102] that fuses dense video features with text embeddings of the language query, followed by a specialized detector head for performing detections [16, 32, 53, 74-76]. However, because these models are often trained on limited datasets for such a narrowly defined task, they struggle with generalization. Indeed, the zero-shot performance of these methods is limited [13, 89, 90, 92, 94]. In this work, we aim to address the shortcomings of previous methods by fully combining the generalization abilities of multi-modal LLMs with the advantages of specialist models.", + "bbox": [ + 511, + 266, + 903, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dense captioning. Our method is related to dense video captioning [29], where the objective is to segment a given video into multiple parts and simultaneously provide descriptive captions for each segment. Traditional approaches tackle this task either by first determining the segments and then providing descriptions [6, 21, 25, 38, 111, 113], or jointly learning both tasks [10, 38, 93, 112, 116]. Recent advances have demonstrated video-conditioned LLMs to excel in this task [88, 115, 117]. While there are similarities between our enrich-and-detect paradigm and dense captioning, we solve a different task, namely video temporal grounding, where the input query is given and constrains the problem.", + "bbox": [ + 511, + 523, + 903, + 718 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt augmentation with LLMs. Off-the-shelf LLMs have recently been used to augment input prompts for tasks such as image retrieval and image classification by providing additional, clarifying descriptions. These augmented descriptions can aid generalization in various vision and NLP tasks such as visual question answering [12, 62, 84], dialog generation [15], and visual classification [51]. Our approach goes beyond using an off-the-shelf LLM for offline query enrichment, rather learning to dynamically enrich queries during inference time towards better grounding.", + "bbox": [ + 511, + 719, + 903, + 869 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prior work has also investigated the use of captioners and language models for data augmentation by rephras", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24298", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/13e74b12896c30c0d81ec6251001ef8b5548091ae7ff3f03979c3fbe2181b8f2.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed Enrich and Detect framework: (left) ED-VTG pipeline: Given an untrimmed video and a query $Q$ to be grounded, the inputs are first tokenized into video tokens $T_V$ and text tokens $T_Q$ . The tokens are then fed into an LLM, which first generates an enriched query by, e.g., filling in any missing details and then emitting an interval token $<\\mathrm{INT}>$ . The embedding of this special token is finally decoded into the predicted temporal interval via a lightweight interval decoder. In the example shown here, the vague input query is enriched into a more detailed one by our model which can be subsequently grounded more easily. (right) Training: ED-VTG is trained using ground-truth temporal intervals and pseudo-labels of enriched queries, generated by an external off-the-shelf captioning model [33], which – unlike our model at inference time – has access to the ground-truth intervals. For every sample during training, the proposed multiple instance learning (MIL) framework allows ED-VTG to assess both the original or the enriched queries and generate two sets of predictions, $\\hat{I}^{enr}$ , the interval predicted using the enriched query, and $\\hat{I}^{dir}$ using the original query. Next, the model backpropagates using the better prediction (i.e., lower grounding loss). Hence, during training, ED-VTG dynamically learns to decide for which sample enrichment is necessary and, based on that, performs detection." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 604, + 335 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4bf8cc78fe8deee8d7277a221578c7477b596e6b7051152da2e6d902c5fa8399.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 617, + 89, + 908, + 335 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing existing descriptions or providing new improved ones. LaViLa[109] uses rephraser and narrator models to improve the quality of the training data for video-text alignment. Augmentation during training is sufficient for that task, as the goal is to learn joint representations; for our grounding task however simply augmenting the training set, without enriching queries during inference is not as effective, as we empirically demonstrate.", + "bbox": [ + 88, + 506, + 482, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multiple instance learning. Multiple instance learning (MIL)[11] is a technique commonly used in weakly-supervised vision problems, when a collection of potential solutions is available but exact annotations are not. It has been successfully applied to a range of tasks, including classification [3], weakly supervised object detection[7, 9] and temporal action localization[57, 60].", + "bbox": [ + 88, + 627, + 482, + 733 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 742, + 181, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given an untrimmed video $\\mathcal{V}$ and a set of $N$ associated textual queries $\\mathcal{Q} = \\{Q_1, \\ldots, Q_N\\}$ , temporal grounding aims to identify the corresponding temporal interval for each query. Formally, the output is a set of temporal intervals $\\mathcal{I} = \\{I_1, \\ldots, I_N\\}$ . For $N = 1$ , the task takes the form of single-query temporal grounding, which we will simply refer to as STG. For clarity, we describe our approach in the STG setting, without loss of generality; the pipeline however readily extends to $N > 1$ .", + "bbox": [ + 88, + 763, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our approach aims to tackle temporal grounding by leveraging a multimodal LLM that $(i)$ transforms input queries into intermediate, enriched queries by adding missing details based on the video input, $(ii)$ generates contextualized embeddings of each latent segment to enable temporal interval prediction, $(iii)$ decodes these embeddings into concrete temporal boundaries.", + "bbox": [ + 511, + 507, + 906, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the following, we formally introduce our proposed ED-VTG model in Section 3.1, and proceed to describe how to train it with enriched query pseudo-labels within a MIL framework (Section 3.2).", + "bbox": [ + 511, + 612, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Model", + "text_level": 1, + "bbox": [ + 511, + 680, + 599, + 694 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our model (Figure 2) consists of three key modules: a vision encoder that extracts video representations, a LLM that jointly processes video and language, and a lightweight interval decoder that generates precise temporal boundaries.", + "bbox": [ + 511, + 702, + 905, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Enrich. Given a single input query $Q$ about a video $V$ with $T$ frames, the vision encoder represents the video as a sequence of $R$ visual tokens $\\mathbf{T}_{\\mathrm{V}} \\in \\mathbb{R}^{R \\times D}$ , where $D$ is the token dimension. The tokenized video features $\\mathbf{T}_{\\mathrm{V}}$ are fed along with the tokenized query $\\mathbf{T}_{\\mathcal{Q}}$ to the LLM, which generates an enriched query $\\hat{Q}^{enr} = \\{\\hat{y}_1, \\dots, \\hat{y}_l, \\dots, \\hat{y}_{L^{enr}}\\}$ one token at a time:", + "bbox": [ + 511, + 762, + 905, + 867 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {l} = \\mathcal {F} _ {\\mathrm {L L M}} \\left(\\hat {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 880, + 903, + 897 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24299", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When the model is ready to ground the query, the LLM emits a new, special token $<\\mathrm{INT}>$ to trigger interval prediction. In other words, the text prediction takes the form:", + "bbox": [ + 89, + 90, + 482, + 137 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {y}} = \\text {\" T h e q u e y} \\hat {Q} ^ {e n r} \\text {o c c u r s a t} < I N T > \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 147, + 143, + 482, + 160 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detect. To detect the temporal interval corresponding to the enriched query $\\hat{Q}^{enr}$ , we introduce an interval decoder $\\mathcal{F}_{\\mathrm{dec}}$ that predicts an interval $\\hat{\\mathbf{I}}$ parametrized by the center $\\hat{c}$ and width $\\hat{w}$ of the predicted interval. We selected this parameterization for its advantage in decoupling position from scale, as supported by literature in object detection [42, 68, 69, 114]. The interval decoder takes the form", + "bbox": [ + 89, + 176, + 483, + 282 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {I}} = (\\hat {c}, \\hat {w}) = \\mathcal {F} _ {\\mathrm {d e c}} \\left(\\mathcal {G} \\left(\\mathbf {h} _ {\\text {i n t}}\\right), \\mathbf {T} _ {\\mathrm {V}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 169, + 295, + 482, + 313 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{h}_{int} \\in \\mathbb{R}^D$ is the hidden state of the LLM corresponding to the token and $\\mathcal{G}$ is a linear projection layer. The decoder, functioning as the regression component of a temporal detector, consists of two transformer layers followed by a multi-layer perceptron (MLP), which processes a concatenation of its two inputs, $(\\mathcal{G}(\\mathbf{h}_{\\mathrm{int}}), \\mathbf{T}_{\\mathrm{V}})$ . It finally outputs the predicted interval $\\hat{\\mathbf{I}}$ , grounding the input query $\\mathcal{Q}$ .", + "bbox": [ + 89, + 325, + 483, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notice that our formulation allows the quality of the enriched query $\\hat{Q}^{enr}$ to directly influence the accuracy of the predicted interval $\\hat{\\mathbf{I}}$ . This is because the generation of the interval token $<\\mathrm{INT}>$ (and consequently its hidden state $\\mathbf{h}_{\\mathrm{int}}$ ) is conditioned on the enriched query prediction. Establishing this cascaded dependency chain, i.e., $(V, Q) \\to \\hat{Q}^{enr}$ and $(V, \\hat{Q}^{enr}) \\to \\hat{I}$ , is the key idea of our approach.", + "bbox": [ + 89, + 450, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Training", + "text_level": 1, + "bbox": [ + 89, + 566, + 194, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The model is trained end-to-end using two primary loss functions: a language modeling loss $\\mathcal{L}_{\\mathrm{LM}}$ and a temporal grounding loss $\\mathcal{L}_{\\mathrm{grnd}}$ that help supervise the \"enrich\" and \"detect\" aspects of our model, respectively.", + "bbox": [ + 89, + 590, + 483, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enrich. Given the target output text $\\mathbf{y} = \\{y_1, y_2, \\dots, y_L\\}$ , $\\mathcal{L}_{\\mathrm{LM}}$ is calculated as the cross-entropy loss that evaluates the likelihood of $\\mathbf{y}$ under the predicted probability distribution generated by the model:", + "bbox": [ + 89, + 664, + 483, + 724 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {L M}} = - \\sum_ {t = 1} ^ {T} \\log P \\left(y _ {l} \\mid \\mathbf {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 726, + 482, + 767 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To provide a proper supervisory signal for query enrichment, we need the ground-truth pair of $(Q, Q^{enr})$ . However, such datasets do not exist, nor is it practical to annotate a dataset solely for this purpose. Here, we capitalize on the tremendous amount of progress made in the video captioning literature [1, 78, 79], and use an off-the-shelf captioning model [33] to generate pseudo ground-truth $Q^{enr}$ by refining the original query $Q$ given its video V (see more details in Section 4.1).", + "bbox": [ + 89, + 776, + 483, + 912 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/446eddf0a68baa2db656a78bfb5ad2f8454fd85df7b059fdcb12adbb3973d816.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DatasetDomainTasksCorpusEval. Protocol# Train SamplesAvg Vid Length (s)Avg Span Length (s)
DiDeMo [4]OpenSTGPT-32.8K54.576.49 (11.9%)
QuerYD [58]CookingSTGPT-13.6K158.787.68 (4.8%)
COIN [83]OpenVPGPT-7.5K143.7115.06 (10.5%)
HiREST [96]OpenSTG, VPGPT-0.8K208.0344.50 (21.4%)
ViTT† [23]OpenVPGPT-4.9K287.17-
YTTemporal [97]OpenVPGPT-28.8K327.364.0 (1.2%)
CrossTask [118]ProceduralAGPT-2.7K297.09.61 (3.2%)
VideoCC [55]OpenSTGPT-45.0K415.899.88 (2.3%)
Charades-STA [14]IndoorSTGFT, EvalZS, FT12.4K31.178.29 (26.6%)
Charades-CD-OOD [95]IndoorVPGFT, EvalFT4.5K30.607.90 (25.8%)
ANet-Captions [29]OpenSTG, VPGFT, EvalZS, FT9.5K117.6335.61 (30.3%)
TACoS [70]CookingSTG, VPGFT, EvalZS, FT9.8K224.3423.33 (10.4%)
YouCook2 [110]CookingVPGFT, EvalFT1.2K311.4120.07 (6.4%)
NEX-T-GQA° [90]OpenQGEvalZS-39.606.69 (16.9%)
HT-Step [2]CookingAGFT, EvalFT17.4K393.8914.88 (3.7%)
", + "bbox": [ + 513, + 89, + 905, + 241 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1. Dataset statistics, corresponding tasks, and evaluation protocol. The upper side of the table represents datasets used for pre-training, resulting in a total of 136K samples. The lower side represents datasets used for fine-tuning and evaluation. We cover four different video grounding tasks: single-query temporal grounding (STG), video paragraph grounding (VPG), question grounding (QG), and article grounding (AG). QG is used only for evaluation to assess the model's generalization capability. Average interval lengths compared to the corresponding video durations are shown in brown, denoting the annotation granularity. $^\\dagger$ VITT contains single timestamp annotation instead of intervals. $^\\circ$ NExTGQA contains only evaluation split.", + "bbox": [ + 511, + 247, + 906, + 414 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Detect. Given a target temporal interval $I = (c, w)$ , the grounding loss $\\mathcal{L}_{\\mathrm{ground}}$ is computed as a combination of the L1 loss and the generalized Intersection over Union (gIoU) loss [73], applied on the predicted temporal interval $(\\hat{c}, \\hat{w})$ :", + "bbox": [ + 511, + 424, + 905, + 486 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g r n d}} = \\lambda_ {\\mathrm {L 1}} \\left(\\left| (\\hat {c} - c | + | \\hat {w} - w |\\right) \\right. \\\\ + \\lambda_ {\\mathrm {g I o U}} \\mathrm {g I o U} ((\\hat {c}, \\hat {w}), (c, w)) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 494, + 903, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. MIL framework", + "text_level": 1, + "bbox": [ + 513, + 534, + 673, + 547 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A caveat with the pseudo-labeled enriched queries is that they can be noisy and include hallucinations; as a result, some of them may lead to a temporal interval prediction inferior to the original query. To address this, we propose considering multiple options for the target query (2 in the case of a single query scenario, $Q$ and $Q^{enr}$ ), by adopting a multiple instance learning (MIL) framework [11]. This approach allows the model to choose between the enriched and the original input queries during training, depending on which version leads to a better temporal interval prediction. During inference, this means that our ED-VTG model can choose to either enrich the query if it misses important details, or decide to \"carry over\" the original query into $\\hat{Q}^{enr}$ when it is concrete enough (in which case $Q = \\hat{Q}^{enr}$ ).", + "bbox": [ + 511, + 551, + 906, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, to perform MIL, we collect two temporal interval predictions by running two forward passes with different LLM inputs (in a teacher-forcing fashion), i.e. $\\mathbf{y}^{\\mathrm{dir}}$ that is formed using the original query $\\mathcal{Q}$ , and $\\mathbf{y}^{\\mathrm{enr}}$ using the pseudo-labeled enriched query $\\mathcal{Q}^{\\mathrm{enr}}$ . We illustrate this schematically in Figure 2 (right). Recall that through the contextualized interval representation $\\mathbf{h}_{\\mathrm{int}}$ , the predicted interval depends on the LLM teacher-forced input. Hence, these inputs produce respective interval predictions, $\\hat{\\mathbf{I}}^{\\mathrm{dir}}$ and", + "bbox": [ + 511, + 765, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24300", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/92f6118e80ce571788b3d2f5edb71a6bce4b364ccfac399510dd523f3d41afec.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
UniVTG [40]X4.2MZS44.125.210.027.1----5.21.30.34.4
SeViLA [94]X129MZS----31.619.010.123.0----
PSVL [56]X-ZS46.231.314.231.244.730.114.729.6----
LT-ZVG [28]X-ZS52.937.219.336.047.632.615.431.8----
Video-LLaMA° [103]✓2.7MZS25.210.63.416.821.910.84.916.55.11.20.83.4
Video-ChatGPT° [49]✓100KZS27.26.21.919.719.510.64.814.26.31.71.04.3
Valley [47]✓100KZS28.41.80.321.430.613.78.121.9----
VideoChat2 [36]✓2MZS38.014.33.824.640.827.89.327.9----
Momenter [63]✓10MZS42.626.611.628.542.923.012.429.3----
VTimeLLM° [21]✓170KZS51.027.511.431.244.027.814.330.47.01.80.84.5
TimeChat° [72]✓125KZS-32.213.4-----6.82.10.84.7
HawkEye [87]✓715KZS50.631.414.533.749.129.310.732.7----
ChatVTG [66]✓100KZS52.733.015.934.940.722.59.427.28.13.71.35.5
ED-VTG✓136KZS59.539.319.840.252.133.116.035.214.56.02.312.7
ΔOurs - HawkEye--ZS8.9↑7.9↑5.3↑6.5↑3.0↑3.8↑5.3↑2.5↑----
ΔOurs - ChatVTG--ZS--------6.4↑2.3↑1.0↑7.2↑
", + "bbox": [ + 98, + 88, + 897, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 2. Zero-shot STG results on Charades, ActivityNet, and TACoS test splits. For all three datasets, ED-VTG gains significant improvement over all existing methods, including task-specific, non-generalist models. We use boldface for the best and underline the second-best result for each metric, among the generalist models. $^\\circ$ Official checkpoints are used for TACoS evaluation.", + "bbox": [ + 89, + 316, + 906, + 359 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\hat{\\mathbf{I}}^{\\mathrm{enr}}$ . We select the query version that results in the smallest grounding loss and use it to compute our overall training objective:", + "bbox": [ + 89, + 367, + 482, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {d i r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} & \\text {i f} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} < \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} \\\\ \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {e n r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 414, + 482, + 454 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{dir}}$ and $\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{enr}}$ are the language losses for targets $\\mathbf{y}^{\\mathrm{dir}}$ and $\\mathbf{y}^{\\mathrm{enr}}$ respectively, and $\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{dir}}$ and $\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{enr}}$ the grounding losses for predictions $\\hat{\\mathbf{I}}^{\\mathrm{dir}}$ and $\\hat{\\mathbf{I}}^{\\mathrm{enr}}$ respectively. The hyperparameters $\\lambda_{\\mathrm{LM}}$ and $\\lambda_{\\mathrm{grnd}}$ are the relative weights of the language modeling and grounding losses.", + "bbox": [ + 89, + 460, + 483, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 547, + 225, + 563 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We design experiments to study three key questions related to our architecture and training framework: Q1) How does ED-VTG perform on various video grounding tasks in comparison to the current state-of-the-art? Q2) How beneficial is our query enrichment approach, within the MIL paradigm, as opposed to directly grounding the original queries? Q3) Does utilizing an interval decoder with specifically tailored grounding objectives offer advantages over predicting timestamps as raw text or special tokens?", + "bbox": [ + 89, + 565, + 483, + 702 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 705, + 192, + 720 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 summarizes the datasets that we used during the pre-training and fine-tuning stages. During pre-training, we use a total of 136K medium-to-long duration videos from 8 public datasets annotated with text queries and the corresponding intervals. As a preprocessing step, we collect pseudo-labels for enriched queries using an external captioning model [33]. In short, we take each video segment defined by the ground-truth intervals and prompt the captioning model to enrich the original query while preserving its meaning given the video segment. We provide full details of this process with the exact prompt used in the supplementary material.", + "bbox": [ + 89, + 727, + 482, + 909 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Tasks", + "text_level": 1, + "bbox": [ + 511, + 367, + 594, + 381 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 also summarizes the tasks for which we use each dataset. Here we briefly describe these tasks, and highlight how our approach is applied to solve them.", + "bbox": [ + 511, + 385, + 906, + 430 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Single-Query Temporal Grounding (STG) involves identifying a single time window in response to a single input language query $(N = 1)$ .", + "bbox": [ + 511, + 430, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Video Paragraph Grounding (VPG) involves grounding $N > 1$ sentences to $N$ corresponding time windows. Our ED-VTG model can be trained to predict multiple enriched queries (one per original input query) interleaved with $<\\mathrm{INT}>$ tokens that get separately decoded into intervals. Since there are multiple queries in the input, we run multiple forward passes through the LLM to perform MIL, wherein each pass selects a random number of queries to be enriched.", + "bbox": [ + 511, + 476, + 905, + 611 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Question Grounding (QG) involves retrieving evidence intervals to answer questions, facilitating explainable video QA. As in STG, the input is a single query, and the output a single interval.", + "bbox": [ + 511, + 612, + 905, + 671 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Article grounding (AG) is an extension of VPG where the model is given multiple queries as input, some of which may not be groundable in the video. Hence, the model must 1) identify which queries are groundable and 2) predict intervals for the groundable queries.", + "bbox": [ + 511, + 672, + 905, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that, while we train our model jointly on multiple tasks, they are unified as a single task in the form of (Video, Text) $\\rightarrow$ (Time intervals). Besides the differences in training data sources, we use the identical training objective (Eq. 6) during pre-training and fine-tuning stages.", + "bbox": [ + 511, + 748, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Evaluation Protocol", + "text_level": 1, + "bbox": [ + 511, + 832, + 702, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the common practices in the literature [21, 72, 87], we evaluate ED-VTG in two primary evaluation protocols: (i) Zero-shot (ZS), where the pre-trained model", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24301", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/11fc7bf06138baa15a9a368f0a07cf9b0cf83077539f83b22bd78289da581d16.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
VSLNet (C3D) [102]X-FT64.347.330.245.263.243.226.243.229.624.320.024.1
MS-2D-TAN (I3D) [107]X-FT-56.636.2-62.145.528.3-42.033.622.1-
Moment-DETR [32]X236KFT65.852.130.645.5----38.024.712.025.5
UnLoc-B [91]X650KFT-58.135.4--48.029.7-----
MomentDiff [37]X-FT-55.632.4-----46.628.912.430.4
LGI [54]X-FT73.059.535.551.458.541.523.141.1----
BAM-DETR [31]X-FT72.960.039.452.3----56.741.526.839.3
InternVideo2* + CG-DETR [86]X2.1MFT79.770.048.958.8--------
SG-DETR [18]X-FT-71.152.860.7-----46.433.942.4
EMB (ELA) [24]X-FT79.769.251.462.273.758.740.756.263.352.537.048.4
BLIP-2 (frames only) [34]✓129MFT-43.332.6--25.89.7-----
VideoChat2 [36]✓2MFT----55.534.717.738.9----
TimeChat [72]✓125KFT-46.723.7-----27.715.16.418.4
HawkEye [87]✓715KFT72.558.328.849.355.934.717.939.1----
VtimeLLM [21]✓170KFT--------26.814.46.118.0
ED-VTG✓136KFT78.262.135.052.667.645.122.744.946.031.515.832.4
ΔOurs - HawkEye--FT5.7↑3.8↑6.2↑3.3↑11.7↑10.4↑4.8↑5.8↑----
ΔOurs - VTimeLLM--FT--------19.2↑17.1↑9.7↑14.4↑
", + "bbox": [ + 98, + 88, + 897, + 319 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Fine-tuned STG results on Charades, ActivityNet, and TACoS test splits. For all datasets, ED-VTG achieves strong improvements over previous generalist models, and performs comparably to task-specific expert models. *Though InterVideo2 is a generalist model, it fine-tunes CG-DETR [52] head for grounding tasks, using the LLM as a feature extractor. For completeness, we report all competitive existing works, including task-specific SOTA specialist models (shown in gray), but directly compare ED-VTG only with generalist frameworks - using boldface for the best and underlining the second best results among generalist models (see discussion in Section 4.3).", + "bbox": [ + 89, + 321, + 903, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "is assessed directly without any fine-tuning on downstream datasets, and (ii) Fine-tuned (FT), where the model undergoes additional training on specific tasks and datasets. We also present results without pre-training to assess its importance. For the STG and VPG tasks, following existing works [6, 21, 72, 76, 87], we report the mean intersection over union (mIoU) and Recall@1 for IoU $\\geq m$ ( $R@m$ ), with $m \\in \\{0.3, 0.5, 0.7\\}$ . For the QG task, we report intersection over prediction (IoP) in addition to IoU, following the NeXT-GQA [90] evaluation protocol. For AG, we adhere to the HT-Step [2] protocol and report article-grounding mAP scores across various IoU thresholds.", + "bbox": [ + 88, + 398, + 480, + 578 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison to non-generalist models. In the results section, we report all existing SOTA and competitive methods for a complete comparison, however we divide them into generalist and non-generalist (specialist) models. We note that specialist methods are heavily tailored to the task and in practice often overfit specific datasets which limits their transferability (as is evidenced by the zero-shot results in Table 2). We therefore focus the discussion around our method's comparison to other generalist models.", + "bbox": [ + 88, + 580, + 482, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 726, + 307, + 742 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We initialize the video encoder and LLM with the VideoLLaMA-7B [103] checkpoint, which is a similarly sized backbone used in existing LLM-based video grounding models [21, 63, 72, 87]. Video-LLaMA is trained on video captioning tasks with WebVid [5] and VideoChat [35]. The video encoder includes a ViT-G/14 from EVA-CLIP [81] as the frame feature extractor, followed by image and video QFormers. We initialize the decoder with random weights. We keep the ViT frozen, apply LoRA [20] with rank 32 to the LLM and fully tune the Q-Formers, decoder, and linear", + "bbox": [ + 88, + 750, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "layers. We use LLaVA OneVision (OV) 72B [33] as the external captioner for obtaining enriched queries in the training sets. For the VPG task, we run four forward passes to perform MIL. We pre-train our model for 40 epochs with a batch size 256, using AdamW [45] with a peak learning rate of 5e-5 and a cosine scheduler [44] with a linear warmup for the first $20\\%$ steps. Pre-training takes 2 days on 16 V100 nodes (8 cards with 32G GPU memory each). Additional details on pre-training, fine-tuning, and task-specific instructions are provided in the supplementary material.", + "bbox": [ + 511, + 398, + 903, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.5. Results", + "text_level": 1, + "bbox": [ + 513, + 550, + 607, + 564 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Single-Query Temporal Grounding (STG). We start by comparing ED-VTG against the state-of-the-art across three different STG benchmarks, namely Charades-STA, ActivityNet-Captions, and TACoS, in a zero-shot evaluation setting. We show the results in Table 2. On Charades, ED-VTG achieves ZS scores of 59.5, 39.3, and 19.8 for R@0.3, R@0.5, and R@0.7, respectively, significantly outperforming all baseline models. Notably, ED-VTG surpasses Momenter [63] and HawkEye [87] by 11.4 and 6.2 absolute mIoU points despite these models being pretrained with 100x and 6x more segment-level data, respectively. A similar trend is observed on ActivityNet and TACoS where our model achieves improvements of 2.5 and 7.2 mIoU points over the nearest LLM-based models in zero-shot setting. Notably, the TACoS dataset contains short, under-specified queries, longer input videos, and fine-grained interval annotations, which pose challenges for LLMs with a fixed number of input frames. The enriched query descriptions that ED-VTG generates enable it to more accurately align them to video frames and precisely retrieve the correct intervals.", + "bbox": [ + 511, + 568, + 903, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "On the fine-tuned STG evaluation, as shown in Table", + "bbox": [ + 531, + 885, + 903, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24302", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b4bb572f5e2d0d1dfa044c9dcd514a4ca9ee6d4a098d89e98c596daa18bc7c3e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodGeneralist ModelCharades-CD-OOD
R@0.3R@0.5mIoU
DepNet [6]✗45.627.629.3
DRN [98]✗40.530.4-
STLG [46]✗48.330.4-
SVPTR [26]✗50.328.532.1
SiamGTR [82]✗59.135.538.9
VTimeLLM1[21]✓53.234.035.1
TimeChat1[72]✓60.536.138.3
ED-VTG70.747.345.0
ΔOurs - TimeChat-10.2↑11.2↑6.7↑
", + "bbox": [ + 91, + 88, + 289, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/06d8348d6087d4fa696190975f3a1b5baf9d121addde293b4ab435c60e71d363.jpg", + "table_caption": [ + "(a) Results on Charades-CD-OOD." + ], + "table_footnote": [], + "table_body": "
MethodGeneralist ModelANet-Captions
R@0.3R@0.5mIoU
CBLN [41]X66.348.127.6
2D-TAN [105]X59.544.5-
3D-TPN [106]X67.651.5-
DepNet [6]X72.855.9-
SVPTR [26]X78.161.755.9
VTimeLLM1[21]✓66.150.345.6
TimeChat1[72]✓67.951.547.0
ED-VTG✓74.158.053.7
ΔOurs - TimeChat-6.2↑6.5↑6.7↑
", + "bbox": [ + 295, + 88, + 493, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c7980896280dc00b58b6fc979c7af2c5aa27dbd31bfd0c1ec4d97d94a9d1b99a.jpg", + "table_caption": [ + "(b) Results on ActivityNet." + ], + "table_footnote": [], + "table_body": "
MethodGeneralist ModelTACoS
R@0.3R@0.5mIoU
CMIN [108]✗24.618.1-
2D-TAN [105]✗37.325.3-
3D-TPN [106]✗40.326.5-
DepNet [6]✗41.327.2-
SVPTR [26]✗47.928.231.4
VTimeLLM†[21]✓40.225.627.9
TimeChat†[72]✓39.525.627.8
ED-VTG✓46.227.830.7
ΔOurs - TimeChat-6.7↑2.2↑2.9↑
", + "bbox": [ + 500, + 88, + 697, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/98d83c44578ce41e167dbce255cb90e2d17300cf9682c5bd9e01371d3c6e327c.jpg", + "table_caption": [ + "(c) Results on TACoS." + ], + "table_footnote": [], + "table_body": "
MethodGeneralist ModelYouCook2
R@0.3R@0.5mIoU
DORi [75]✗43.430.530.5
DORi* [75]✗42.329.929.9
LocFormer [76]✗46.831.330.9
ExCL [16]✗26.616.218.9
TMLGA [74]✗34.823.124.4
VTimeLLM1[21]✓41.318.524.3
TimeChat1[72]✓40.919.026.6
ED-VTG✓48.128.031.5
ΔOurs - TimeChat-7.2↑9.0↑4.9↑
", + "bbox": [ + 702, + 88, + 901, + 188 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/0edae7f604909a333aadb9a48d53480f864f1ec98cd4abb107fbecf4536f8f2f.jpg", + "table_caption": [ + "(d) Results on YouCook2.", + "Table 4. Performance on VPG task on four different benchmarks: Charades-CD-OOD, ActivityNet-Captions, TACoS, YouCook2. ED-VTG significantly improves over previous LLM-based models, and performs comparably to state-of-the-art specialist models. ${}^{ \\dagger }$ We fine-tune VTimeLLM and TimeChat checkpoints for the VPG task. Dori ${}^{ \\star }$ represents frozen text (BERT) encoder during fine-tuning." + ], + "table_footnote": [], + "table_body": "
MethodGeneralist ModelNExT-GQA
mIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VGT [89]✘24.726.024.63.04.21.4
VIOLETv2 [13]✘23.625.123.33.14.31.3
Temp[CLIP] NG+ [90]✘25.731.425.512.117.58.9
FrozenBiLM NG+ [92]✘24.228.523.79.613.56.1
SeViLA [94]✘29.534.722.921.729.213.8
\\( LLoVi7B^{\\circ} \\) [100]✓20.7-20.58.7-6.0
VideoStreaming° [64]✓32.2-31.019.3-13.3
LongRepo 7B° [27]✓20.3-20.08.7-6.0
DeVi [65]✓33.8-32.220.717.4-
HawkEye [87]✓---25.737.019.5
ED-VTG✓34.745.133.526.639.519.8
\\( \\Delta_{Ours - SeViLA} \\)-5.2↑10.4↑10.6↑4.9↑10.3↑6.0↑
\\( \\Delta_{Ours - HawkEye} \\)----0.9↑2.5↑0.3↑
", + "bbox": [ + 91, + 252, + 483, + 405 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/357e4b7130a1da6f2b622e23c924dc1b7de6ccd26ab1ffd2ab0b4829bc2dfe7a.jpg", + "table_caption": [ + "Table 5. Performance on QG task on NeXT-GQA test split. ED-VTG consistently achieves consistent improvements over the existing models across all metrics. Results of LLoVi, LongRepo and VideoStreaming are from [64]." + ], + "table_footnote": [], + "table_body": "
MethodSeenUnseen
↑mAP@IoU↑mAP@IoU
@0.3@0.5@0.7@[0.3-0.7]@0.3@0.5@0.7@[0.3-0.7]
UMT [43]15.78.73.29.19.44.91.75.3
MT+BCE [2, 50]46.229.912.929.831.618.77.719.3
ActionFormer-T [101]41.230.818.330.229.720.310.720.4
Timechat† [72]45.329.014.429.030.717.87.518.7
ED-VTG48.931.518.032.533.021.211.121.6
ΔOurs - TimeChat3.6 ↑2.5 ↑3.6 ↑3.5 ↑2.3 ↑3.4 ↑3.6 ↑2.9 ↑
", + "bbox": [ + 91, + 468, + 483, + 570 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 6. Performance on AG task on HT-Step seen and unseen val split. ED-VTG is the first LLM-based model to report results on for video grounding in the presence of negative, non-groundable queries and sets a new state of the art in terms of average mAP score across various IoU thresholds. We fine-tune the official TimeChat checkpoint for AG task.", + "bbox": [ + 89, + 574, + 483, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "3. ED-VTG demonstrates an impressive gain of 5.7 and 11.7 points in R@0.3 over HawkEye on Charades and ActivityNet, respectively. Similarly on TACoS, ED-VTG surpasses all existing MLLMs by a considerable margin, e.g. 14.0 and $14.4\\mathrm{mIoU}$ points over TimeChat and VTimeLLM. Furthermore, ED-VTG also beats many existing task-specific specialist models in all three benchmarks, significantly reducing the gap between specialist models and MLLMs for STG.", + "bbox": [ + 89, + 664, + 482, + 800 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Video Paragraph Grounding (VPG). Next, we fine-tune ED-VTG for the VPG task, where the model processes multiple input queries in a temporal sequence. While specialist models have reported results in the past, no LLM-based models have previously addressed this challenging task. To establish a baseline for comparison, we fine-tune the", + "bbox": [ + 89, + 809, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "officially released VTimeLLM and TimeChat pre-trained checkpoints. As shown in Table 4a, ED-VTG significantly outperforms both LLMs on the Charades-CD-OOD dataset, achieving an absolute gain of 9.9 and $6.7\\mathrm{mIoU}$ points over VTimeLLM and TimeChat, respectively. Additionally, ED-VTG surpasses all specialist models on this dataset by a substantial margin, setting a new state-of-the-art.", + "bbox": [ + 511, + 255, + 903, + 359 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We also evaluate VPG performance on three other benchmarks: ActivityNet-Captions, TACoS, and YouCook2, as presented in Tables 4b, 4c, and 4d. Consistent with the results on Charades, ED-VTG outperforms LLM-based models across all metrics on these datasets, with mIoU gains of 6.7, 2.9, and 4.9 over TimeChat on ActivityNet-Captions, TACoS, and YouCook2, respectively. ED-VTG also exceeds many existing specialist VPG models, demonstrating the effectiveness of our enriched queries and cascaded interval decoder.", + "bbox": [ + 511, + 362, + 905, + 513 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Question Grounding (QG). To further assess the model's generalization capabilities, we conduct zero-shot evaluation on the held-out QG task. Our results on the NeXT-GQA test set are presented in Table 5. Notably, ED-VTG achieves state-of-the-art performance across all metrics, outperforming both specialist and LLM-based models by a significant margin. Specifically, ED-VTG surpasses the existing best baseline, HawkEye, by 2.5 points in terms of IoU@0.3 score, demonstrating its strong generalizability.", + "bbox": [ + 511, + 518, + 905, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Article Grounding (AG). We assess ED-VTG on the AG on the HT-Step benchmark. Table 6 shows the fine-tuned performance on the AG task, where ED-VTG shows significant improvements over existing baselines. Notably, on the challenging unseen split, our model achieves the best results across all metrics, surpassing both the LLM and specialist models by a decent margin. The ability to handle non-groundable queries underscores ED-VTG's real-world applicability, as it does not always assume that the query is occurring in the input video.", + "bbox": [ + 511, + 659, + 905, + 810 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.6. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 816, + 666, + 832 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Enriched Queries. We examine the impact of query enrichment through a step-by-step ablation, as shown in Tables 7 and 8. Initially, we compare the effect of enrichment without MIL paradigm against direct grounding. As", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24303", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1e9534a438d01b1066b72f497c90405064b432e3e3ba753aca80688399119c6a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 87, + 883, + 141 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fc6f1677a7f4aa25e9c93618e186c22c2a6ec1544e1ea72eae13b5239fdd7c86.jpg", + "image_caption": [ + "Figure 3. Example of query enrichment and detection made by ED-VTG on video paragraph grounding (VPG) task from the ActivityNet-Captions [29] dataset. In this specific sample, three different queries are enriched and localized together." + ], + "image_footnote": [], + "bbox": [ + 109, + 146, + 887, + 250 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e9d1602c659b14e3e038f4fd7fbe97b2522cc610f71fe1d36801c69e54dbc076.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Training ParadigmCharades-STA STGActivityNet-Captions STG
ZSFT w/o PTFTZSFT w/o PTFT
R@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect48.130.631.051.431.533.268.949.045.846.326.029.650.330.134.061.138.039.2
Enrich & Detect58.137.337.760.137.038.475.156.649.750.729.533.456.335.537.865.543.443.8
Enrich & Detect w/ MIL59.539.340.262.838.440.378.262.152.652.133.135.257.536.238.667.645.144.9
", + "bbox": [ + 91, + 286, + 906, + 359 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/ea1aa702e83a397a7d3d0753e06e994e26c068953250f43a6cb738cf52449b81.jpg", + "table_caption": [ + "Table 7. Ablation on the effect of enriched queries. Our proposed enrich & detect framework significantly gains over directly grounding the input queries across different evaluation settings. Introducing the MIL framework further improves the performance. FT w/o PT refers directly fine-tuning on respective datasets, without performing pre-training." + ], + "table_footnote": [], + "table_body": "
Training ParadigmCharades-STAANet-Captions
R@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect51.431.533.250.330.134.0
Offline Enrich + Detect51.731.533.449.829.933.7
Enrich & Detect60.137.038.456.335.537.8
", + "bbox": [ + 91, + 412, + 480, + 483 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0348abf1f16ccbda6ba3fdedb0a2b51b1f06eeb76d8bd3e43bc5959754d12604.jpg", + "table_caption": [ + "Table 8. Ablation on enrichment as a training pre-processing step. The two step enrich & detect framework is more helpful since the trained model learn to perform autonomous enrichment during evaluation. Reported results are in FT w/o PT setting." + ], + "table_footnote": [], + "table_body": "
DecoderObjectivesCharades-STA STGANet-Captions
LML1gIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
-✓--54.233.234.151.031.635.5
✓✓✓-58.536.037.055.834.437.1
✓✓-✓58.936.237.155.834.737.3
✓✓✓✓60.137.038.456.335.537.8
", + "bbox": [ + 91, + 549, + 480, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 9. Ablation study on different training objectives. Training the model only using the LM loss (without the decoder) leads to significant performance drop. Results are in FT w/o PT setting.", + "bbox": [ + 89, + 633, + 482, + 676 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "indicated in the first two rows of Table 7, enriched queries lead to significant improvements on the Charades-STA and ActivityNet-Captions STG benchmarks. In the zero-shot (ZS) setting, enrichment results in gains of 6.7 and 3.8 mIoU points on these datasets, respectively. Introducing the MIL paradigm further enhances performance, adding 2.5 and $1.8\\mathrm{mIoU}$ points. Similar improvements are observed in other evaluation settings, highlighting the substantial effectiveness of query enrichment within the MIL framework.", + "bbox": [ + 89, + 686, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Additionally, we investigate the effect of offline enrichment in Table 8. In this scenario, instead of using a two-step grounding process, we enrich the queries as a training enrichment step, and the model is then directly provided with these enriched queries as input and asked to directly per", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "form grounding. However, we find that offline enrichment is not advantageous, primarily due to the lack of enrichment during evaluation. In contrast, our two-step grounding approach allows the trained model to learn how to enrich queries and improve them autonomously when necessary, resulting in significant performance gains.", + "bbox": [ + 511, + 414, + 906, + 505 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Training Objectives. We ablate different training objectives on the Charades and ActivityNet STG benchmarks, as shown in Table 9. The decoder achieves optimal performance when both L1 and gIoU objectives are used together; omitting either one slightly reduces the scores.", + "bbox": [ + 511, + 505, + 905, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.7. Qualitative Results and Error Analysis", + "text_level": 1, + "bbox": [ + 511, + 585, + 846, + 602 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Figure 3 visualizes a VPG sample from the ActivityNet-Captions dataset where ED-VTG meaningfully enriches the three input queries, then proceeds to precisely ground them. Please refer to supplementary for more qualitative results, including intuitive demonstrations of the flexibility in choosing between enrichment or direct detection, and comparison with TimeChat baseline. We also present there some interesting failure cases, such as the ones involving small and obscured objects in long input videos.", + "bbox": [ + 511, + 604, + 906, + 741 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 746, + 633, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented ED-VTG, a novel method for fine-grained video temporal grounding using multi-modal LLMs. By enhancing queries with additional details, utilizing a lightweight decoder and trained in a multiple-instance framework, ED-VTG accurately locates temporal boundaries in videos. Our experiments show that our method outperforms existing LLM-based works and is competitive with specialized models, especially in zero-shot settings, setting a new, strong benchmark for video grounding tasks.", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24304", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgement", + "text_level": 1, + "bbox": [ + 89, + 90, + 272, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This codebase for this project is built on the TimeChat [72] and VTimeLLM [21] repository. We would like to thank the respective authors for their contributions. We gratefully acknowledge the following colleagues at FAIR for valuable discussions and support of our project: Tushar Nagarajan, Huiyu Wang, Yujie Lu, and Arjun Somayazulu.", + "bbox": [ + 89, + 108, + 485, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 210, + 187, + 227 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Moloud Abdar, Meenakshi Kollati, Swaraja Kuraparthi, Farhad Pourpanah, Daniel McDuff, Mohammad Ghavamzadeh, Shuicheng Yan, Abduallah Mohamed, Abbas Khosravi, Erik Cambria, et al. A review of deep learning for video captioning. arXiv preprint arXiv:2304.11431, 2023. 1, 4", + "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In NeurIPS, 2024. 4, 6, 7", + "[3] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. In NeurIPS, pages 561-568, 2003. 3", + "[4] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 4", + "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6", + "[6] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, pages 920-928, 2021. 2, 6, 7", + "[7] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, pages 2846-2854, 2016. 3", + "[8] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. NeurIPS, 34:28442-28453, 2021. 1", + "[9] Ramazan Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multi-fold multiple instance learning. IEEE TPAMI, 39, 2015. 3", + "[10] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2", + "[11] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1):31-71, 1997. 3, 4", + "[12] Li Dong, Jonathan Mallinson, Siva Reddy, and Mirella Lapata. Learning to paraphrase for question answering. In EMNLP, 2017. 2", + "[13] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, pages 22898-22909, 2023. 2, 7", + "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 4" + ], + "bbox": [ + 101, + 236, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Silin Gao, Yichi Zhang, Zhijian Ou, and Zhou Yu. Paraphrase augmented task-oriented dialog generation. In ACL, pages 639-649, 2020. 2", + "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander G Hauptmann. Excl: Extractive clip localization using natural language descriptions. In NAACL, pages 1984-1990, 2019. 2, 7", + "[17] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE TPAMI, 38(1): 142-158, 2015. 2", + "[18] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 6", + "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2", + "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6", + "[21] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, pages 14271-14280, 2024. 2, 5, 6, 7, 9", + "[22] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, pages 202-218. Springer, 2025. 2", + "[23] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL, pages 470-490, 2020. 4", + "[24] Jiabo Huang, Hailin Jin, Shaogang Gong, and Yang Liu. Video activity localisation with uncertainties in temporal boundary. In ECCV, pages 724-740. Springer, 2022. 6", + "[25] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops, 2020. 2", + "[26] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, pages 2466-2475, 2022. 2, 7", + "[27] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024. 7", + "[28] Dahye Kim, Jungin Park, Jiyoung Lee, Seongheon Park, and Kwanghoon Sohn. Language-free training for zero-shot video grounding. In WACV, pages 2539-2548, 2023. 5", + "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In ICCV, pages 706-715, 2017. 2, 4, 8", + "[30] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 2" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24305", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 6", + "[32] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 1, 2, 6", + "[33] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. TMLR, 2025. 3, 4, 5, 6", + "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 6", + "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6", + "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6", + "[37] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36, 2024. 6", + "[38] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. CVPR, pages 7492-7500, 2018. 2", + "[39] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, pages 6657-6678, 2024. 2", + "[40] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 5", + "[41] Daizong Liu, Xiaoye Qu, Jianfeng Dong, Pan Zhou, Yu Cheng, Wei Wei, Zichuan Xu, and Yulai Xie. Context-aware biaffine localizing network for temporal sentence grounding. In CVPR, pages 11235-11244, 2021. 2, 7", + "[42] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 4", + "[43] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 7", + "[44] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6", + "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + ], + "bbox": [ + 99, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. IEEE Transactions on Multimedia, 25:7747-7757, 2022. 2, 7", + "[47] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5", + "[48] Kajing Ma, Xianghao Zang, Zerun Feng, Han Fang, Chao Ban, Yuhan Wei, Zhongjiang He, Yongxiang Li, and Hao Sun. Llavilo: Boosting video moment retrieval via adapter-based multimodal modeling. In ICCV Workshops, pages 2790-2795. IEEE, 2023. 2", + "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 5", + "[50] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In ICCV, pages 15201-15213, 2023. 7", + "[51] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 2", + "[52] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 6", + "[53] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2", + "[54] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, pages 10810-10819, 2020. 6", + "[55] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, pages 407-426. Springer, 2022. 4", + "[56] Jinwoo Nam, Daechul Ahn, Dongyeop Kang, Seong Jong Ha, and Jonghyun Choi. Zero-shot natural language video localization. In ICCV, pages 1470-1479, 2021. 5", + "[57] Phuc Xuan Nguyen, Tianzhu Liu, Gaurav Prasad, Hung Hai Bui, Binh Pham, and Svetha Venkatesh. Weakly supervised action localization by sparse temporal pooling network. In CVPR, pages 6752-6761, 2019. 3", + "[58] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 4", + "[59] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, pages 10870-10879, 2020. 1", + "[60] Sujoy Paul, Sourya Roy, and Amit K Roy-Chowdhury. W-talc: Weakly-supervised temporal activity localization and classification. In ECCV, pages 563-579, 2018. 3" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24306", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In ICCV, pages 5285-5297, 2023. 2", + "[62] Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Rephrase, augment, reason: Visual grounding of questions for vision-language models. In ICLR, 2024. 2", + "[63] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 2, 5, 6", + "[64] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. NeurIPS, 37:119336-119360, 2024. 7", + "[65] Hangyu Qin, Junbin Xiao, and Angela Yao. Question-answering dense video events. arXiv preprint arXiv:2409.04388, 2024.7", + "[66] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 5", + "[67] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, pages 13009-13018, 2024. 2", + "[68] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In CVPR, pages 6517-6525, 2017. 4", + "[69] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 4", + "[70] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 1:25-36, 2013. 4", + "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 2", + "[72] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 1, 2, 5, 6, 7, 9", + "[73] Seyed Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. CVPR, pages 658-666, 2019. 4", + "[74] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, pages 2464-2473, 2020. 2, 7", + "[75] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hongdong Li, and Stephen Gould. Dori: Discovering object relationships for moment localization of a natural language query in a video. In WACV, pages 1079–1088, 2021. 2, 7" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[76] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hiroya Takamura, and Qi Wu. Memory-efficient temporal moment localization in long videos. In EACL, pages 1909–1924, 2023. 2, 6, 7", + "[77] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 2", + "[78] Himanshu Sharma, Manmohan Agrahari, Sujeet Kumar Singh, Mohd Firoj, and Ravi Kumar Mishra. Image captioning: a comprehensive survey. In 2020 International Conference on Power Electronics & IoT Applications in Renewable Energy and its Control (PARC), pages 325-328. IEEE, 2020. 1, 4", + "[79] Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Silvia Cascianelli, Giuseppe Fiameni, and Rita Cucchiara. From show to tell: A survey on deep learning-based image captioning. IEEE TPAMI, 45(1):539-559, 2022. 1, 4", + "[80] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In AAAI, pages 4998-5007, 2024. 1", + "[81] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6", + "[82] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. In CVPR, pages 13569-13580, 2024. 2, 7", + "[83] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 4", + "[84] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In ICCV, pages 1397-1407, 2021. 2", + "[85] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 1, 2", + "[86] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6", + "[87] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024.5,6,7", + "[88] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, pages 18699-18708, 2024. 2", + "[89] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. In ECCV, pages 39-58. Springer, 2022. 2, 7", + "[90] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, pages 13204-13214, 2024. 2, 4, 6, 7" + ], + "bbox": [ + 524, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24307", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[91] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In CVPR, pages 13623-13633, 2023. 6", + "[92] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, pages 124-141, 2022. 2, 7", + "[93] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2", + "[94] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2024. 2, 5, 7", + "[95] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In Proceedings of the 2nd international workshop on human-centric multimedia analysis, pages 13-21, 2021. 4", + "[96] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 4", + "[97] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 4", + "[98] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020. 2, 7", + "[99] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR, 2025. 2", + "[100] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. In EMNLP, pages 21715–21737, 2024. 7", + "[101] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, pages 492-510. Springer, 2022. 7", + "[102] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, pages 6543-6554, 2020. 1, 2, 6", + "[103] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 5, 6", + "[104] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In CVPR, pages 8327–8336, 2019. 1", + "[105] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment" + ], + "bbox": [ + 93, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7", + "[106] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7", + "[107] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. IEEE TPAMI, 44(12):9073-9087, 2021. 6", + "[108] Zhu Zhang, Zhijie Lin, Zhou Zhao, and Zhenxin Xiao. Cross-modal interaction networks for query-based moment retrieval in videos. In ACM SIGIR, pages 655-664, 2019. 7", + "[109] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 3", + "[110] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4", + "[111] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2", + "[112] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. CVPR, pages 8739-8748, 2018. 2", + "[113] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2", + "[114] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4", + "[115] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, pages 18243-18252, 2024. 2", + "[116] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. ArXiv, abs/2204.08121, 2022. 2", + "[117] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 5651-5665, Gyeongju, Republic of Korea, 2022. International Committee on Computational Linguistics. 2", + "[118] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 4" + ], + "bbox": [ + 516, + 92, + 903, + 814 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "24308", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_model.json b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_model.json new file mode 100644 index 0000000000000000000000000000000000000000..d7af18ea8138d564f3518b28581cc980eaf5cad3 --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_model.json @@ -0,0 +1,2974 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.26, + 0.131, + 0.74, + 0.174 + ], + "angle": 0, + "content": "Enrich and Detect: Video Temporal Grounding with Multimodal LLMs" + }, + { + "type": "text", + "bbox": [ + 0.168, + 0.177, + 0.83, + 0.214 + ], + "angle": 0, + "content": "Shraman Pramanick\\(^{\\dagger,2}\\) Effrosyni Mavroudi\\(^{1}\\) Yale Song\\(^{1}\\) Rama Chellappa\\(^{2}\\) Lorenzo Torresani\\(^{\\diamond 1}\\) Triantafyllos Afouras\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.227, + 0.214, + 0.766, + 0.233 + ], + "angle": 0, + "content": "\\(^{1}\\)FAIR, Meta \\(^{2}\\)Johns Hopkins University \\(^{3}\\)Northeastern University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.253, + 0.327, + 0.269 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.29, + 0.485, + 0.637 + ], + "angle": 0, + "content": "We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.222, + 0.666 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.669, + 0.483, + 0.851 + ], + "angle": 0, + "content": "Video temporal grounding [32, 40, 53, 80, 102] aims to identify temporal intervals in a video that correspond to a set of provided language queries. The task is essential for applications such as video editing and content retrieval. Conversely, video captioning [1, 59, 78, 79, 104] entails generating a natural language description for a given video segment, effectively translating visual content into text. The two tasks are in fact dual, as the outputs of one task are the inputs to the other, and vice versa. Intuitively, there is significant potential in exploiting this synergy, however it has largely remained unexplored, with previous works typically specializing in one of the two tasks [8, 32, 53, 104] or solv-" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.253, + 0.905, + 0.429 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.433, + 0.908, + 0.544 + ], + "angle": 0, + "content": "Figure 1. Our proposed system. ED-VTG performs video temporal grounding as a two-stage process: since user queries are often incomplete or coarse, the first stage involves producing an enriched query that adds additional details to the original, making it easier to ground. Meanwhile, a contextualized embedding is generated, containing all the information about the interval to be predicted. In the second stage, an interval decoder translates these embeddings into precise temporal boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.583 + ], + "angle": 0, + "content": "ing them in multi-task setting [72, 85] without investigating how each task can benefit the other." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.584, + 0.907, + 0.856 + ], + "angle": 0, + "content": "In this work, we exploit this duality by leveraging captioning to enhance grounding. Our key observation is that natural language queries often lack the completeness or detail necessary for effective temporal localization; indeed, existing grounding datasets frequently contain poorly worded, coarse, and potentially incomplete queries. The quality and completeness of these queries however is crucial for the precision of the temporal grounding. A natural hypothesis, therefore, is that more detailed queries, which could be obtained via conditional captioning, can significantly enhance grounding accuracy. For example, as illustrated in Figure 1, a vague query like 'Man starts surfing' can be refined into a more detailed description such as 'The man with a yellow surfboard slowly runs to start surfing,' resulting in grounding with more accurate temporal boundaries. In other cases, refining an abstract concept may involve breaking down a complex query into simpler components that are more directly groundable." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This key idea of query enrichment forms the basis of our proposed approach. Concretely, we transform grounding into a two-stage reasoning process using a multi-modal" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.863, + 0.341, + 0.876 + ], + "angle": 0, + "content": "\\(\\dagger\\) Work done during an internship at FAIR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.876, + 0.264, + 0.888 + ], + "angle": 0, + "content": "Work done while at FAIR." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.403, + 0.9 + ], + "angle": 0, + "content": "spraman3@meta.com, afourast@meta.com" + }, + { + "type": "list", + "bbox": [ + 0.114, + 0.863, + 0.403, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24297" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "LLM: first the model enriches the input query into a more detailed description by adding missing details based on the video content, and then temporally localizes the resulting enriched query in the video." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.155, + 0.482, + 0.306 + ], + "angle": 0, + "content": "To effectively perform the temporal localization, we introduce a lightweight perception decoder that specializes in generating precise temporal boundaries, conditioned on a contextual representation, allowing the LLM to focus on language outputs, where it excels. The perception decoder allows us to leverage purposefully crafted training objectives for temporal grounding, building on prior knowledge and task-specific characteristics developed from years of research in object detection [17, 19, 30, 67, 71, 77] and temporal localization [32, 40, 53, 61, 102, 105]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.483, + 0.504 + ], + "angle": 0, + "content": "Learning to jointly enrich and detect requires high-quality enriched query labels. We obtain those by using a strong external captioning model which we condition on the original queries and the video content of the target temporal boundaries. However powerful, these models are prone to hallucinations and there is no guarantee that the enriched queries will always be easier to ground than the original ones. At the same time, annotating the ground truth to determine which query - original or enriched - is a better candidate, is extremely expensive. To address this issue, we propose training in a multiple-instance learning (MIL) fashion, that enables the model to autonomously determine which query is better suited for the task during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.507, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Finally, we note that our proposed method is not equivalent to a data augmentation approach which simply preprocesses the training set to generate enhanced queries that are directly used for training. While this simpler alternative offers some of the same benefits, it suffers from the limitation that extracting enriched queries during training requires knowledge of the ground-truth temporal segments, i.e., the grounding targets. Since during inference these segments are unknown, the original queries must be used as input, which, as we will show experimentally, is suboptimal. Our method overcomes this limitation by learning to jointly enrich and detect, demonstrating superior performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To summarize, our contributions are as follows: (i) we introduce a cascaded approach to temporal grounding, where the model first enriches the provided language query based on the video context and then proceeds to localize it; (ii) we enable multi-modal LLMs to accurately localize text queries using a lightweight decoder which allows for training with detection objectives tailored to the task; (iii) we propose a multiple-instance learning paradigm that enables the model to dynamically select the query that leads to better temporal localization; (iv) we achieve state-of-the-art results on several temporal grounding benchmarks, for both single query grounding and paragraph grounding, demonstrating, for the first time, an LLM-based model that surpasses or performs comparably to specialist models." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.665, + 0.107 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.266 + ], + "angle": 0, + "content": "LLM-based temporal grounding. Prior works have explored using LLMs for grounding natural language sentences in videos, either using raw text tokens to represent timestamps [21, 36, 39, 48, 72, 99] or by adding hundreds of special tokens to the LLM's vocabulary to represent video frames [22, 63, 85]. Our approach differs from these methods in that by utilizing a lightweight interval decoder we can apply detection losses such as L1 and gIoU with minimal added complexity; we additionally exploit the LLM's potential to describe video content in detail." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.523 + ], + "angle": 0, + "content": "Specialist models. There is a rich variety of specialist models in the literature that are tailored to specific variants of temporal grounding, e.g. single query temporal grounding [102] and video paragraph grounding [82], and as such, achieve strong performance [6, 26, 41, 46, 75, 82, 98, 105, 106]. Modern methods typically employ a multi-modal transformer [102] that fuses dense video features with text embeddings of the language query, followed by a specialized detector head for performing detections [16, 32, 53, 74-76]. However, because these models are often trained on limited datasets for such a narrowly defined task, they struggle with generalization. Indeed, the zero-shot performance of these methods is limited [13, 89, 90, 92, 94]. In this work, we aim to address the shortcomings of previous methods by fully combining the generalization abilities of multi-modal LLMs with the advantages of specialist models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.905, + 0.719 + ], + "angle": 0, + "content": "Dense captioning. Our method is related to dense video captioning [29], where the objective is to segment a given video into multiple parts and simultaneously provide descriptive captions for each segment. Traditional approaches tackle this task either by first determining the segments and then providing descriptions [6, 21, 25, 38, 111, 113], or jointly learning both tasks [10, 38, 93, 112, 116]. Recent advances have demonstrated video-conditioned LLMs to excel in this task [88, 115, 117]. While there are similarities between our enrich-and-detect paradigm and dense captioning, we solve a different task, namely video temporal grounding, where the input query is given and constrains the problem." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.87 + ], + "angle": 0, + "content": "Prompt augmentation with LLMs. Off-the-shelf LLMs have recently been used to augment input prompts for tasks such as image retrieval and image classification by providing additional, clarifying descriptions. These augmented descriptions can aid generalization in various vision and NLP tasks such as visual question answering [12, 62, 84], dialog generation [15], and visual classification [51]. Our approach goes beyond using an off-the-shelf LLM for offline query enrichment, rather learning to dynamically enrich queries during inference time towards better grounding." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Prior work has also investigated the use of captioners and language models for data augmentation by rephras" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24298" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.605, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.619, + 0.09, + 0.91, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.342, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed Enrich and Detect framework: (left) ED-VTG pipeline: Given an untrimmed video and a query \\( Q \\) to be grounded, the inputs are first tokenized into video tokens \\( T_V \\) and text tokens \\( T_Q \\). The tokens are then fed into an LLM, which first generates an enriched query by, e.g., filling in any missing details and then emitting an interval token \\( <\\mathrm{INT}> \\). The embedding of this special token is finally decoded into the predicted temporal interval via a lightweight interval decoder. In the example shown here, the vague input query is enriched into a more detailed one by our model which can be subsequently grounded more easily. (right) Training: ED-VTG is trained using ground-truth temporal intervals and pseudo-labels of enriched queries, generated by an external off-the-shelf captioning model [33], which – unlike our model at inference time – has access to the ground-truth intervals. For every sample during training, the proposed multiple instance learning (MIL) framework allows ED-VTG to assess both the original or the enriched queries and generate two sets of predictions, \\( \\hat{I}^{enr} \\), the interval predicted using the enriched query, and \\( \\hat{I}^{dir} \\) using the original query. Next, the model backpropagates using the better prediction (i.e., lower grounding loss). Hence, during training, ED-VTG dynamically learns to decide for which sample enrichment is necessary and, based on that, performs detection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.627 + ], + "angle": 0, + "content": "ing existing descriptions or providing new improved ones. LaViLa[109] uses rephraser and narrator models to improve the quality of the training data for video-text alignment. Augmentation during training is sufficient for that task, as the goal is to learn joint representations; for our grounding task however simply augmenting the training set, without enriching queries during inference is not as effective, as we empirically demonstrate." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.628, + 0.483, + 0.734 + ], + "angle": 0, + "content": "Multiple instance learning. Multiple instance learning (MIL)[11] is a technique commonly used in weakly-supervised vision problems, when a collection of potential solutions is available but exact annotations are not. It has been successfully applied to a range of tasks, including classification [3], weakly supervised object detection[7, 9] and temporal action localization[57, 60]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.182, + 0.759 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Given an untrimmed video \\(\\mathcal{V}\\) and a set of \\(N\\) associated textual queries \\(\\mathcal{Q} = \\{Q_1, \\ldots, Q_N\\}\\), temporal grounding aims to identify the corresponding temporal interval for each query. Formally, the output is a set of temporal intervals \\(\\mathcal{I} = \\{I_1, \\ldots, I_N\\}\\). For \\(N = 1\\), the task takes the form of single-query temporal grounding, which we will simply refer to as STG. For clarity, we describe our approach in the STG setting, without loss of generality; the pipeline however readily extends to \\(N > 1\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.612 + ], + "angle": 0, + "content": "Our approach aims to tackle temporal grounding by leveraging a multimodal LLM that \\((i)\\) transforms input queries into intermediate, enriched queries by adding missing details based on the video input, \\((ii)\\) generates contextualized embeddings of each latent segment to enable temporal interval prediction, \\((iii)\\) decodes these embeddings into concrete temporal boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.673 + ], + "angle": 0, + "content": "In the following, we formally introduce our proposed ED-VTG model in Section 3.1, and proceed to describe how to train it with enriched query pseudo-labels within a MIL framework (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.681, + 0.6, + 0.695 + ], + "angle": 0, + "content": "3.1. Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.703, + 0.906, + 0.763 + ], + "angle": 0, + "content": "Our model (Figure 2) consists of three key modules: a vision encoder that extracts video representations, a LLM that jointly processes video and language, and a lightweight interval decoder that generates precise temporal boundaries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.868 + ], + "angle": 0, + "content": "Enrich. Given a single input query \\(Q\\) about a video \\(V\\) with \\(T\\) frames, the vision encoder represents the video as a sequence of \\(R\\) visual tokens \\(\\mathbf{T}_{\\mathrm{V}} \\in \\mathbb{R}^{R \\times D}\\), where \\(D\\) is the token dimension. The tokenized video features \\(\\mathbf{T}_{\\mathrm{V}}\\) are fed along with the tokenized query \\(\\mathbf{T}_{\\mathcal{Q}}\\) to the LLM, which generates an enriched query \\(\\hat{Q}^{enr} = \\{\\hat{y}_1, \\dots, \\hat{y}_l, \\dots, \\hat{y}_{L^{enr}}\\}\\) one token at a time:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.881, + 0.905, + 0.898 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {l} = \\mathcal {F} _ {\\mathrm {L L M}} \\left(\\hat {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right). \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24299" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.138 + ], + "angle": 0, + "content": "When the model is ready to ground the query, the LLM emits a new, special token \\(<\\mathrm{INT}>\\) to trigger interval prediction. In other words, the text prediction takes the form:" + }, + { + "type": "equation", + "bbox": [ + 0.148, + 0.144, + 0.483, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {y}} = \\text {\" T h e q u e y} \\hat {Q} ^ {e n r} \\text {o c c u r s a t} < I N T > \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.177, + 0.484, + 0.283 + ], + "angle": 0, + "content": "Detect. To detect the temporal interval corresponding to the enriched query \\(\\hat{Q}^{enr}\\), we introduce an interval decoder \\(\\mathcal{F}_{\\mathrm{dec}}\\) that predicts an interval \\(\\hat{\\mathbf{I}}\\) parametrized by the center \\(\\hat{c}\\) and width \\(\\hat{w}\\) of the predicted interval. We selected this parameterization for its advantage in decoupling position from scale, as supported by literature in object detection [42, 68, 69, 114]. The interval decoder takes the form" + }, + { + "type": "equation", + "bbox": [ + 0.171, + 0.296, + 0.483, + 0.314 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {I}} = (\\hat {c}, \\hat {w}) = \\mathcal {F} _ {\\mathrm {d e c}} \\left(\\mathcal {G} \\left(\\mathbf {h} _ {\\text {i n t}}\\right), \\mathbf {T} _ {\\mathrm {V}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.45 + ], + "angle": 0, + "content": "where \\(\\mathbf{h}_{int} \\in \\mathbb{R}^D\\) is the hidden state of the LLM corresponding to the token and \\(\\mathcal{G}\\) is a linear projection layer. The decoder, functioning as the regression component of a temporal detector, consists of two transformer layers followed by a multi-layer perceptron (MLP), which processes a concatenation of its two inputs, \\((\\mathcal{G}(\\mathbf{h}_{\\mathrm{int}}), \\mathbf{T}_{\\mathrm{V}})\\). It finally outputs the predicted interval \\(\\hat{\\mathbf{I}}\\), grounding the input query \\(\\mathcal{Q}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.557 + ], + "angle": 0, + "content": "Notice that our formulation allows the quality of the enriched query \\(\\hat{Q}^{enr}\\) to directly influence the accuracy of the predicted interval \\(\\hat{\\mathbf{I}}\\). This is because the generation of the interval token \\(<\\mathrm{INT}>\\) (and consequently its hidden state \\(\\mathbf{h}_{\\mathrm{int}}\\)) is conditioned on the enriched query prediction. Establishing this cascaded dependency chain, i.e., \\((V, Q) \\to \\hat{Q}^{enr}\\) and \\((V, \\hat{Q}^{enr}) \\to \\hat{I}\\), is the key idea of our approach." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.195, + 0.584 + ], + "angle": 0, + "content": "3.2. Training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.653 + ], + "angle": 0, + "content": "The model is trained end-to-end using two primary loss functions: a language modeling loss \\(\\mathcal{L}_{\\mathrm{LM}}\\) and a temporal grounding loss \\(\\mathcal{L}_{\\mathrm{grnd}}\\) that help supervise the \"enrich\" and \"detect\" aspects of our model, respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.484, + 0.725 + ], + "angle": 0, + "content": "Enrich. Given the target output text \\(\\mathbf{y} = \\{y_1, y_2, \\dots, y_L\\}\\), \\(\\mathcal{L}_{\\mathrm{LM}}\\) is calculated as the cross-entropy loss that evaluates the likelihood of \\(\\mathbf{y}\\) under the predicted probability distribution generated by the model:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.727, + 0.483, + 0.768 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {L M}} = - \\sum_ {t = 1} ^ {T} \\log P \\left(y _ {l} \\mid \\mathbf {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.777, + 0.484, + 0.913 + ], + "angle": 0, + "content": "To provide a proper supervisory signal for query enrichment, we need the ground-truth pair of \\((Q, Q^{enr})\\). However, such datasets do not exist, nor is it practical to annotate a dataset solely for this purpose. Here, we capitalize on the tremendous amount of progress made in the video captioning literature [1, 78, 79], and use an off-the-shelf captioning model [33] to generate pseudo ground-truth \\(Q^{enr}\\) by refining the original query \\(Q\\) given its video V (see more details in Section 4.1)." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.09, + 0.906, + 0.242 + ], + "angle": 0, + "content": "
DatasetDomainTasksCorpusEval. Protocol# Train SamplesAvg Vid Length (s)Avg Span Length (s)
DiDeMo [4]OpenSTGPT-32.8K54.576.49 (11.9%)
QuerYD [58]CookingSTGPT-13.6K158.787.68 (4.8%)
COIN [83]OpenVPGPT-7.5K143.7115.06 (10.5%)
HiREST [96]OpenSTG, VPGPT-0.8K208.0344.50 (21.4%)
ViTT† [23]OpenVPGPT-4.9K287.17-
YTTemporal [97]OpenVPGPT-28.8K327.364.0 (1.2%)
CrossTask [118]ProceduralAGPT-2.7K297.09.61 (3.2%)
VideoCC [55]OpenSTGPT-45.0K415.899.88 (2.3%)
Charades-STA [14]IndoorSTGFT, EvalZS, FT12.4K31.178.29 (26.6%)
Charades-CD-OOD [95]IndoorVPGFT, EvalFT4.5K30.607.90 (25.8%)
ANet-Captions [29]OpenSTG, VPGFT, EvalZS, FT9.5K117.6335.61 (30.3%)
TACoS [70]CookingSTG, VPGFT, EvalZS, FT9.8K224.3423.33 (10.4%)
YouCook2 [110]CookingVPGFT, EvalFT1.2K311.4120.07 (6.4%)
NEX-T-GQA° [90]OpenQGEvalZS-39.606.69 (16.9%)
HT-Step [2]CookingAGFT, EvalFT17.4K393.8914.88 (3.7%)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.248, + 0.907, + 0.415 + ], + "angle": 0, + "content": "Table 1. Dataset statistics, corresponding tasks, and evaluation protocol. The upper side of the table represents datasets used for pre-training, resulting in a total of 136K samples. The lower side represents datasets used for fine-tuning and evaluation. We cover four different video grounding tasks: single-query temporal grounding (STG), video paragraph grounding (VPG), question grounding (QG), and article grounding (AG). QG is used only for evaluation to assess the model's generalization capability. Average interval lengths compared to the corresponding video durations are shown in brown, denoting the annotation granularity. \\(^\\dagger\\)VITT contains single timestamp annotation instead of intervals. \\(^\\circ\\)NExTGQA contains only evaluation split." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.425, + 0.906, + 0.487 + ], + "angle": 0, + "content": "Detect. Given a target temporal interval \\( I = (c, w) \\), the grounding loss \\( \\mathcal{L}_{\\mathrm{ground}} \\) is computed as a combination of the L1 loss and the generalized Intersection over Union (gIoU) loss [73], applied on the predicted temporal interval \\( (\\hat{c}, \\hat{w}) \\):" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.496, + 0.905, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g r n d}} = \\lambda_ {\\mathrm {L 1}} \\left(\\left| (\\hat {c} - c | + | \\hat {w} - w |\\right) \\right. \\\\ + \\lambda_ {\\mathrm {g I o U}} \\mathrm {g I o U} ((\\hat {c}, \\hat {w}), (c, w)) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.535, + 0.674, + 0.549 + ], + "angle": 0, + "content": "3.2.1. MIL framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.765 + ], + "angle": 0, + "content": "A caveat with the pseudo-labeled enriched queries is that they can be noisy and include hallucinations; as a result, some of them may lead to a temporal interval prediction inferior to the original query. To address this, we propose considering multiple options for the target query (2 in the case of a single query scenario, \\( Q \\) and \\( Q^{enr} \\)), by adopting a multiple instance learning (MIL) framework [11]. This approach allows the model to choose between the enriched and the original input queries during training, depending on which version leads to a better temporal interval prediction. During inference, this means that our ED-VTG model can choose to either enrich the query if it misses important details, or decide to \"carry over\" the original query into \\( \\hat{Q}^{enr} \\) when it is concrete enough (in which case \\( Q = \\hat{Q}^{enr} \\))." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Formally, to perform MIL, we collect two temporal interval predictions by running two forward passes with different LLM inputs (in a teacher-forcing fashion), i.e. \\(\\mathbf{y}^{\\mathrm{dir}}\\) that is formed using the original query \\(\\mathcal{Q}\\), and \\(\\mathbf{y}^{\\mathrm{enr}}\\) using the pseudo-labeled enriched query \\(\\mathcal{Q}^{\\mathrm{enr}}\\). We illustrate this schematically in Figure 2 (right). Recall that through the contextualized interval representation \\(\\mathbf{h}_{\\mathrm{int}}\\), the predicted interval depends on the LLM teacher-forced input. Hence, these inputs produce respective interval predictions, \\(\\hat{\\mathbf{I}}^{\\mathrm{dir}}\\) and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24300" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.898, + 0.313 + ], + "angle": 0, + "content": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
UniVTG [40]X4.2MZS44.125.210.027.1----5.21.30.34.4
SeViLA [94]X129MZS----31.619.010.123.0----
PSVL [56]X-ZS46.231.314.231.244.730.114.729.6----
LT-ZVG [28]X-ZS52.937.219.336.047.632.615.431.8----
Video-LLaMA° [103]✓2.7MZS25.210.63.416.821.910.84.916.55.11.20.83.4
Video-ChatGPT° [49]✓100KZS27.26.21.919.719.510.64.814.26.31.71.04.3
Valley [47]✓100KZS28.41.80.321.430.613.78.121.9----
VideoChat2 [36]✓2MZS38.014.33.824.640.827.89.327.9----
Momenter [63]✓10MZS42.626.611.628.542.923.012.429.3----
VTimeLLM° [21]✓170KZS51.027.511.431.244.027.814.330.47.01.80.84.5
TimeChat° [72]✓125KZS-32.213.4-----6.82.10.84.7
HawkEye [87]✓715KZS50.631.414.533.749.129.310.732.7----
ChatVTG [66]✓100KZS52.733.015.934.940.722.59.427.28.13.71.35.5
ED-VTG✓136KZS59.539.319.840.252.133.116.035.214.56.02.312.7
ΔOurs - HawkEye--ZS8.9↑7.9↑5.3↑6.5↑3.0↑3.8↑5.3↑2.5↑----
ΔOurs - ChatVTG--ZS--------6.4↑2.3↑1.0↑7.2↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.317, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Table 2. Zero-shot STG results on Charades, ActivityNet, and TACoS test splits. For all three datasets, ED-VTG gains significant improvement over all existing methods, including task-specific, non-generalist models. We use boldface for the best and underline the second-best result for each metric, among the generalist models. \\(^\\circ\\)Official checkpoints are used for TACoS evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.413 + ], + "angle": 0, + "content": "\\(\\hat{\\mathbf{I}}^{\\mathrm{enr}}\\). We select the query version that results in the smallest grounding loss and use it to compute our overall training objective:" + }, + { + "type": "equation", + "bbox": [ + 0.106, + 0.415, + 0.483, + 0.455 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {d i r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} & \\text {i f} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} < \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} \\\\ \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {e n r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.543 + ], + "angle": 0, + "content": "where \\(\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{dir}}\\) and \\(\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{enr}}\\) are the language losses for targets \\(\\mathbf{y}^{\\mathrm{dir}}\\) and \\(\\mathbf{y}^{\\mathrm{enr}}\\) respectively, and \\(\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{dir}}\\) and \\(\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{enr}}\\) the grounding losses for predictions \\(\\hat{\\mathbf{I}}^{\\mathrm{dir}}\\) and \\(\\hat{\\mathbf{I}}^{\\mathrm{enr}}\\) respectively. The hyperparameters \\(\\lambda_{\\mathrm{LM}}\\) and \\(\\lambda_{\\mathrm{grnd}}\\) are the relative weights of the language modeling and grounding losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.548, + 0.226, + 0.564 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.703 + ], + "angle": 0, + "content": "We design experiments to study three key questions related to our architecture and training framework: Q1) How does ED-VTG perform on various video grounding tasks in comparison to the current state-of-the-art? Q2) How beneficial is our query enrichment approach, within the MIL paradigm, as opposed to directly grounding the original queries? Q3) Does utilizing an interval decoder with specifically tailored grounding objectives offer advantages over predicting timestamps as raw text or special tokens?" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.706, + 0.194, + 0.721 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.728, + 0.483, + 0.91 + ], + "angle": 0, + "content": "Table 1 summarizes the datasets that we used during the pre-training and fine-tuning stages. During pre-training, we use a total of 136K medium-to-long duration videos from 8 public datasets annotated with text queries and the corresponding intervals. As a preprocessing step, we collect pseudo-labels for enriched queries using an external captioning model [33]. In short, we take each video segment defined by the ground-truth intervals and prompt the captioning model to enrich the original query while preserving its meaning given the video segment. We provide full details of this process with the exact prompt used in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.368, + 0.596, + 0.382 + ], + "angle": 0, + "content": "4.2. Tasks" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.386, + 0.907, + 0.431 + ], + "angle": 0, + "content": "Table 1 also summarizes the tasks for which we use each dataset. Here we briefly describe these tasks, and highlight how our approach is applied to solve them." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Single-Query Temporal Grounding (STG) involves identifying a single time window in response to a single input language query \\((N = 1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Video Paragraph Grounding (VPG) involves grounding \\( N > 1 \\) sentences to \\( N \\) corresponding time windows. Our ED-VTG model can be trained to predict multiple enriched queries (one per original input query) interleaved with \\( <\\mathrm{INT}> \\) tokens that get separately decoded into intervals. Since there are multiple queries in the input, we run multiple forward passes through the LLM to perform MIL, wherein each pass selects a random number of queries to be enriched." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.672 + ], + "angle": 0, + "content": "Question Grounding (QG) involves retrieving evidence intervals to answer questions, facilitating explainable video QA. As in STG, the input is a single query, and the output a single interval." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Article grounding (AG) is an extension of VPG where the model is given multiple queries as input, some of which may not be groundable in the video. Hence, the model must 1) identify which queries are groundable and 2) predict intervals for the groundable queries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Note that, while we train our model jointly on multiple tasks, they are unified as a single task in the form of (Video, Text) \\(\\rightarrow\\) (Time intervals). Besides the differences in training data sources, we use the identical training objective (Eq. 6) during pre-training and fine-tuning stages." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.704, + 0.848 + ], + "angle": 0, + "content": "4.3. Evaluation Protocol" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Following the common practices in the literature [21, 72, 87], we evaluate ED-VTG in two primary evaluation protocols: (i) Zero-shot (ZS), where the pre-trained model" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24301" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.099, + 0.089, + 0.898, + 0.32 + ], + "angle": 0, + "content": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
VSLNet (C3D) [102]X-FT64.347.330.245.263.243.226.243.229.624.320.024.1
MS-2D-TAN (I3D) [107]X-FT-56.636.2-62.145.528.3-42.033.622.1-
Moment-DETR [32]X236KFT65.852.130.645.5----38.024.712.025.5
UnLoc-B [91]X650KFT-58.135.4--48.029.7-----
MomentDiff [37]X-FT-55.632.4-----46.628.912.430.4
LGI [54]X-FT73.059.535.551.458.541.523.141.1----
BAM-DETR [31]X-FT72.960.039.452.3----56.741.526.839.3
InternVideo2* + CG-DETR [86]X2.1MFT79.770.048.958.8--------
SG-DETR [18]X-FT-71.152.860.7-----46.433.942.4
EMB (ELA) [24]X-FT79.769.251.462.273.758.740.756.263.352.537.048.4
BLIP-2 (frames only) [34]✓129MFT-43.332.6--25.89.7-----
VideoChat2 [36]✓2MFT----55.534.717.738.9----
TimeChat [72]✓125KFT-46.723.7-----27.715.16.418.4
HawkEye [87]✓715KFT72.558.328.849.355.934.717.939.1----
VtimeLLM [21]✓170KFT--------26.814.46.118.0
ED-VTG✓136KFT78.262.135.052.667.645.122.744.946.031.515.832.4
ΔOurs - HawkEye--FT5.7↑3.8↑6.2↑3.3↑11.7↑10.4↑4.8↑5.8↑----
ΔOurs - VTimeLLM--FT--------19.2↑17.1↑9.7↑14.4↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.322, + 0.904, + 0.391 + ], + "angle": 0, + "content": "Table 3. Fine-tuned STG results on Charades, ActivityNet, and TACoS test splits. For all datasets, ED-VTG achieves strong improvements over previous generalist models, and performs comparably to task-specific expert models. *Though InterVideo2 is a generalist model, it fine-tunes CG-DETR [52] head for grounding tasks, using the LLM as a feature extractor. For completeness, we report all competitive existing works, including task-specific SOTA specialist models (shown in gray), but directly compare ED-VTG only with generalist frameworks - using boldface for the best and underlining the second best results among generalist models (see discussion in Section 4.3)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.482, + 0.579 + ], + "angle": 0, + "content": "is assessed directly without any fine-tuning on downstream datasets, and (ii) Fine-tuned (FT), where the model undergoes additional training on specific tasks and datasets. We also present results without pre-training to assess its importance. For the STG and VPG tasks, following existing works [6, 21, 72, 76, 87], we report the mean intersection over union (mIoU) and Recall@1 for IoU \\(\\geq m\\) (\\(R@m\\)), with \\(m \\in \\{0.3, 0.5, 0.7\\}\\). For the QG task, we report intersection over prediction (IoP) in addition to IoU, following the NeXT-GQA [90] evaluation protocol. For AG, we adhere to the HT-Step [2] protocol and report article-grounding mAP scores across various IoU thresholds." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Comparison to non-generalist models. In the results section, we report all existing SOTA and competitive methods for a complete comparison, however we divide them into generalist and non-generalist (specialist) models. We note that specialist methods are heavily tailored to the task and in practice often overfit specific datasets which limits their transferability (as is evidenced by the zero-shot results in Table 2). We therefore focus the discussion around our method's comparison to other generalist models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.727, + 0.308, + 0.743 + ], + "angle": 0, + "content": "4.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We initialize the video encoder and LLM with the VideoLLaMA-7B [103] checkpoint, which is a similarly sized backbone used in existing LLM-based video grounding models [21, 63, 72, 87]. Video-LLaMA is trained on video captioning tasks with WebVid [5] and VideoChat [35]. The video encoder includes a ViT-G/14 from EVA-CLIP [81] as the frame feature extractor, followed by image and video QFormers. We initialize the decoder with random weights. We keep the ViT frozen, apply LoRA [20] with rank 32 to the LLM and fully tune the Q-Formers, decoder, and linear" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.905, + 0.549 + ], + "angle": 0, + "content": "layers. We use LLaVA OneVision (OV) 72B [33] as the external captioner for obtaining enriched queries in the training sets. For the VPG task, we run four forward passes to perform MIL. We pre-train our model for 40 epochs with a batch size 256, using AdamW [45] with a peak learning rate of 5e-5 and a cosine scheduler [44] with a linear warmup for the first \\(20\\%\\) steps. Pre-training takes 2 days on 16 V100 nodes (8 cards with 32G GPU memory each). Additional details on pre-training, fine-tuning, and task-specific instructions are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.551, + 0.608, + 0.565 + ], + "angle": 0, + "content": "4.5. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.884 + ], + "angle": 0, + "content": "Single-Query Temporal Grounding (STG). We start by comparing ED-VTG against the state-of-the-art across three different STG benchmarks, namely Charades-STA, ActivityNet-Captions, and TACoS, in a zero-shot evaluation setting. We show the results in Table 2. On Charades, ED-VTG achieves ZS scores of 59.5, 39.3, and 19.8 for R@0.3, R@0.5, and R@0.7, respectively, significantly outperforming all baseline models. Notably, ED-VTG surpasses Momenter [63] and HawkEye [87] by 11.4 and 6.2 absolute mIoU points despite these models being pretrained with 100x and 6x more segment-level data, respectively. A similar trend is observed on ActivityNet and TACoS where our model achieves improvements of 2.5 and 7.2 mIoU points over the nearest LLM-based models in zero-shot setting. Notably, the TACoS dataset contains short, under-specified queries, longer input videos, and fine-grained interval annotations, which pose challenges for LLMs with a fixed number of input frames. The enriched query descriptions that ED-VTG generates enable it to more accurately align them to video frames and precisely retrieve the correct intervals." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.905, + 0.9 + ], + "angle": 0, + "content": "On the fine-tuned STG evaluation, as shown in Table" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24302" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.29, + 0.189 + ], + "angle": 0, + "content": "
MethodGeneralist ModelCharades-CD-OOD
R@0.3R@0.5mIoU
DepNet [6]✗45.627.629.3
DRN [98]✗40.530.4-
STLG [46]✗48.330.4-
SVPTR [26]✗50.328.532.1
SiamGTR [82]✗59.135.538.9
VTimeLLM1[21]✓53.234.035.1
TimeChat1[72]✓60.536.138.3
ED-VTG70.747.345.0
ΔOurs - TimeChat-10.2↑11.2↑6.7↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.192, + 0.287, + 0.203 + ], + "angle": 0, + "content": "(a) Results on Charades-CD-OOD." + }, + { + "type": "table", + "bbox": [ + 0.297, + 0.089, + 0.495, + 0.189 + ], + "angle": 0, + "content": "
MethodGeneralist ModelANet-Captions
R@0.3R@0.5mIoU
CBLN [41]X66.348.127.6
2D-TAN [105]X59.544.5-
3D-TPN [106]X67.651.5-
DepNet [6]X72.855.9-
SVPTR [26]X78.161.755.9
VTimeLLM1[21]✓66.150.345.6
TimeChat1[72]✓67.951.547.0
ED-VTG✓74.158.053.7
ΔOurs - TimeChat-6.2↑6.5↑6.7↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.323, + 0.192, + 0.468, + 0.204 + ], + "angle": 0, + "content": "(b) Results on ActivityNet." + }, + { + "type": "table", + "bbox": [ + 0.501, + 0.089, + 0.698, + 0.189 + ], + "angle": 0, + "content": "
MethodGeneralist ModelTACoS
R@0.3R@0.5mIoU
CMIN [108]✗24.618.1-
2D-TAN [105]✗37.325.3-
3D-TPN [106]✗40.326.5-
DepNet [6]✗41.327.2-
SVPTR [26]✗47.928.231.4
VTimeLLM†[21]✓40.225.627.9
TimeChat†[72]✓39.525.627.8
ED-VTG✓46.227.830.7
ΔOurs - TimeChat-6.7↑2.2↑2.9↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.539, + 0.192, + 0.658, + 0.203 + ], + "angle": 0, + "content": "(c) Results on TACoS." + }, + { + "type": "table", + "bbox": [ + 0.704, + 0.089, + 0.902, + 0.189 + ], + "angle": 0, + "content": "
MethodGeneralist ModelYouCook2
R@0.3R@0.5mIoU
DORi [75]✗43.430.530.5
DORi* [75]✗42.329.929.9
LocFormer [76]✗46.831.330.9
ExCL [16]✗26.616.218.9
TMLGA [74]✗34.823.124.4
VTimeLLM1[21]✓41.318.524.3
TimeChat1[72]✓40.919.026.6
ED-VTG✓48.128.031.5
ΔOurs - TimeChat-7.2↑9.0↑4.9↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.734, + 0.192, + 0.872, + 0.203 + ], + "angle": 0, + "content": "(d) Results on YouCook2." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.205, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Table 4. Performance on VPG task on four different benchmarks: Charades-CD-OOD, ActivityNet-Captions, TACoS, YouCook2. ED-VTG significantly improves over previous LLM-based models, and performs comparably to state-of-the-art specialist models. \\( {}^{ \\dagger } \\) We fine-tune VTimeLLM and TimeChat checkpoints for the VPG task. Dori \\( {}^{ \\star } \\) represents frozen text (BERT) encoder during fine-tuning." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.253, + 0.485, + 0.406 + ], + "angle": 0, + "content": "
MethodGeneralist ModelNExT-GQA
mIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VGT [89]✘24.726.024.63.04.21.4
VIOLETv2 [13]✘23.625.123.33.14.31.3
Temp[CLIP] NG+ [90]✘25.731.425.512.117.58.9
FrozenBiLM NG+ [92]✘24.228.523.79.613.56.1
SeViLA [94]✘29.534.722.921.729.213.8
\\( LLoVi7B^{\\circ} \\) [100]✓20.7-20.58.7-6.0
VideoStreaming° [64]✓32.2-31.019.3-13.3
LongRepo 7B° [27]✓20.3-20.08.7-6.0
DeVi [65]✓33.8-32.220.717.4-
HawkEye [87]✓---25.737.019.5
ED-VTG✓34.745.133.526.639.519.8
\\( \\Delta_{Ours - SeViLA} \\)-5.2↑10.4↑10.6↑4.9↑10.3↑6.0↑
\\( \\Delta_{Ours - HawkEye} \\)----0.9↑2.5↑0.3↑
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.411, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Table 5. Performance on QG task on NeXT-GQA test split. ED-VTG consistently achieves consistent improvements over the existing models across all metrics. Results of LLoVi, LongRepo and VideoStreaming are from [64]." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.469, + 0.484, + 0.571 + ], + "angle": 0, + "content": "
MethodSeenUnseen
↑mAP@IoU↑mAP@IoU
@0.3@0.5@0.7@[0.3-0.7]@0.3@0.5@0.7@[0.3-0.7]
UMT [43]15.78.73.29.19.44.91.75.3
MT+BCE [2, 50]46.229.912.929.831.618.77.719.3
ActionFormer-T [101]41.230.818.330.229.720.310.720.4
Timechat† [72]45.329.014.429.030.717.87.518.7
ED-VTG48.931.518.032.533.021.211.121.6
ΔOurs - TimeChat3.6 ↑2.5 ↑3.6 ↑3.5 ↑2.3 ↑3.4 ↑3.6 ↑2.9 ↑
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.575, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Table 6. Performance on AG task on HT-Step seen and unseen val split. ED-VTG is the first LLM-based model to report results on for video grounding in the presence of negative, non-groundable queries and sets a new state of the art in terms of average mAP score across various IoU thresholds. We fine-tune the official TimeChat checkpoint for AG task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.801 + ], + "angle": 0, + "content": "3. ED-VTG demonstrates an impressive gain of 5.7 and 11.7 points in R@0.3 over HawkEye on Charades and ActivityNet, respectively. Similarly on TACoS, ED-VTG surpasses all existing MLLMs by a considerable margin, e.g. 14.0 and \\(14.4\\mathrm{mIoU}\\) points over TimeChat and VTimeLLM. Furthermore, ED-VTG also beats many existing task-specific specialist models in all three benchmarks, significantly reducing the gap between specialist models and MLLMs for STG." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Video Paragraph Grounding (VPG). Next, we fine-tune ED-VTG for the VPG task, where the model processes multiple input queries in a temporal sequence. While specialist models have reported results in the past, no LLM-based models have previously addressed this challenging task. To establish a baseline for comparison, we fine-tune the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.256, + 0.905, + 0.361 + ], + "angle": 0, + "content": "officially released VTimeLLM and TimeChat pre-trained checkpoints. As shown in Table 4a, ED-VTG significantly outperforms both LLMs on the Charades-CD-OOD dataset, achieving an absolute gain of 9.9 and \\(6.7\\mathrm{mIoU}\\) points over VTimeLLM and TimeChat, respectively. Additionally, ED-VTG surpasses all specialist models on this dataset by a substantial margin, setting a new state-of-the-art." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.514 + ], + "angle": 0, + "content": "We also evaluate VPG performance on three other benchmarks: ActivityNet-Captions, TACoS, and YouCook2, as presented in Tables 4b, 4c, and 4d. Consistent with the results on Charades, ED-VTG outperforms LLM-based models across all metrics on these datasets, with mIoU gains of 6.7, 2.9, and 4.9 over TimeChat on ActivityNet-Captions, TACoS, and YouCook2, respectively. ED-VTG also exceeds many existing specialist VPG models, demonstrating the effectiveness of our enriched queries and cascaded interval decoder." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Question Grounding (QG). To further assess the model's generalization capabilities, we conduct zero-shot evaluation on the held-out QG task. Our results on the NeXT-GQA test set are presented in Table 5. Notably, ED-VTG achieves state-of-the-art performance across all metrics, outperforming both specialist and LLM-based models by a significant margin. Specifically, ED-VTG surpasses the existing best baseline, HawkEye, by 2.5 points in terms of IoU@0.3 score, demonstrating its strong generalizability." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.811 + ], + "angle": 0, + "content": "Article Grounding (AG). We assess ED-VTG on the AG on the HT-Step benchmark. Table 6 shows the fine-tuned performance on the AG task, where ED-VTG shows significant improvements over existing baselines. Notably, on the challenging unseen split, our model achieves the best results across all metrics, surpassing both the LLM and specialist models by a decent margin. The ability to handle non-groundable queries underscores ED-VTG's real-world applicability, as it does not always assume that the query is occurring in the input video." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.817, + 0.667, + 0.833 + ], + "angle": 0, + "content": "4.6. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Effect of Enriched Queries. We examine the impact of query enrichment through a step-by-step ablation, as shown in Tables 7 and 8. Initially, we compare the effect of enrichment without MIL paradigm against direct grounding. As" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24303" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.088, + 0.885, + 0.142 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.147, + 0.888, + 0.251 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.907, + 0.283 + ], + "angle": 0, + "content": "Figure 3. Example of query enrichment and detection made by ED-VTG on video paragraph grounding (VPG) task from the ActivityNet-Captions [29] dataset. In this specific sample, three different queries are enriched and localized together." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.287, + 0.907, + 0.361 + ], + "angle": 0, + "content": "
Training ParadigmCharades-STA STGActivityNet-Captions STG
ZSFT w/o PTFTZSFT w/o PTFT
R@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect48.130.631.051.431.533.268.949.045.846.326.029.650.330.134.061.138.039.2
Enrich & Detect58.137.337.760.137.038.475.156.649.750.729.533.456.335.537.865.543.443.8
Enrich & Detect w/ MIL59.539.340.262.838.440.378.262.152.652.133.135.257.536.238.667.645.144.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.364, + 0.907, + 0.409 + ], + "angle": 0, + "content": "Table 7. Ablation on the effect of enriched queries. Our proposed enrich & detect framework significantly gains over directly grounding the input queries across different evaluation settings. Introducing the MIL framework further improves the performance. FT w/o PT refers directly fine-tuning on respective datasets, without performing pre-training." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.484 + ], + "angle": 0, + "content": "
Training ParadigmCharades-STAANet-Captions
R@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect51.431.533.250.330.134.0
Offline Enrich + Detect51.731.533.449.829.933.7
Enrich & Detect60.137.038.456.335.537.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.486, + 0.483, + 0.543 + ], + "angle": 0, + "content": "Table 8. Ablation on enrichment as a training pre-processing step. The two step enrich & detect framework is more helpful since the trained model learn to perform autonomous enrichment during evaluation. Reported results are in FT w/o PT setting." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.55, + 0.482, + 0.63 + ], + "angle": 0, + "content": "
DecoderObjectivesCharades-STA STGANet-Captions
LML1gIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
-✓--54.233.234.151.031.635.5
✓✓✓-58.536.037.055.834.437.1
✓✓-✓58.936.237.155.834.737.3
✓✓✓✓60.137.038.456.335.537.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.678 + ], + "angle": 0, + "content": "Table 9. Ablation study on different training objectives. Training the model only using the LM loss (without the decoder) leads to significant performance drop. Results are in FT w/o PT setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.822 + ], + "angle": 0, + "content": "indicated in the first two rows of Table 7, enriched queries lead to significant improvements on the Charades-STA and ActivityNet-Captions STG benchmarks. In the zero-shot (ZS) setting, enrichment results in gains of 6.7 and 3.8 mIoU points on these datasets, respectively. Introducing the MIL paradigm further enhances performance, adding 2.5 and \\(1.8\\mathrm{mIoU}\\) points. Similar improvements are observed in other evaluation settings, highlighting the substantial effectiveness of query enrichment within the MIL framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Additionally, we investigate the effect of offline enrichment in Table 8. In this scenario, instead of using a two-step grounding process, we enrich the queries as a training enrichment step, and the model is then directly provided with these enriched queries as input and asked to directly per" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.415, + 0.907, + 0.506 + ], + "angle": 0, + "content": "form grounding. However, we find that offline enrichment is not advantageous, primarily due to the lack of enrichment during evaluation. In contrast, our two-step grounding approach allows the trained model to learn how to enrich queries and improve them autonomously when necessary, resulting in significant performance gains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.582 + ], + "angle": 0, + "content": "Training Objectives. We ablate different training objectives on the Charades and ActivityNet STG benchmarks, as shown in Table 9. The decoder achieves optimal performance when both L1 and gIoU objectives are used together; omitting either one slightly reduces the scores." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.587, + 0.848, + 0.603 + ], + "angle": 0, + "content": "4.7. Qualitative Results and Error Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.605, + 0.907, + 0.742 + ], + "angle": 0, + "content": "Figure 3 visualizes a VPG sample from the ActivityNet-Captions dataset where ED-VTG meaningfully enriches the three input queries, then proceeds to precisely ground them. Please refer to supplementary for more qualitative results, including intuitive demonstrations of the flexibility in choosing between enrichment or direct detection, and comparison with TimeChat baseline. We also present there some interesting failure cases, such as the ones involving small and obscured objects in long input videos." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.747, + 0.634, + 0.761 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this paper, we presented ED-VTG, a novel method for fine-grained video temporal grounding using multi-modal LLMs. By enhancing queries with additional details, utilizing a lightweight decoder and trained in a multiple-instance framework, ED-VTG accurately locates temporal boundaries in videos. Our experiments show that our method outperforms existing LLM-based works and is competitive with specialized models, especially in zero-shot settings, setting a new, strong benchmark for video grounding tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24304" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.273, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.109, + 0.486, + 0.201 + ], + "angle": 0, + "content": "This codebase for this project is built on the TimeChat [72] and VTimeLLM [21] repository. We would like to thank the respective authors for their contributions. We gratefully acknowledge the following colleagues at FAIR for valuable discussions and support of our project: Tushar Nagarajan, Huiyu Wang, Yujie Lu, and Arjun Somayazulu." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.212, + 0.188, + 0.228 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.237, + 0.486, + 0.32 + ], + "angle": 0, + "content": "[1] Moloud Abdar, Meenakshi Kollati, Swaraja Kuraparthi, Farhad Pourpanah, Daniel McDuff, Mohammad Ghavamzadeh, Shuicheng Yan, Abduallah Mohamed, Abbas Khosravi, Erik Cambria, et al. A review of deep learning for video captioning. arXiv preprint arXiv:2304.11431, 2023. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.322, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In NeurIPS, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.379, + 0.483, + 0.419 + ], + "angle": 0, + "content": "[3] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. In NeurIPS, pages 561-568, 2003. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.421, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[4] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.477, + 0.484, + 0.518 + ], + "angle": 0, + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.52, + 0.483, + 0.547 + ], + "angle": 0, + "content": "[6] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, pages 920-928, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.549, + 0.483, + 0.575 + ], + "angle": 0, + "content": "[7] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, pages 2846-2854, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.577, + 0.483, + 0.617 + ], + "angle": 0, + "content": "[8] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. NeurIPS, 34:28442-28453, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.62, + 0.483, + 0.66 + ], + "angle": 0, + "content": "[9] Ramazan Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multi-fold multiple instance learning. IEEE TPAMI, 39, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.483, + 0.703 + ], + "angle": 0, + "content": "[10] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.705, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[11] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1):31-71, 1997. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.761, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[12] Li Dong, Jonathan Mallinson, Siva Reddy, and Mirella Lapata. Learning to paraphrase for question answering. In EMNLP, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[13] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, pages 22898-22909, 2023. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 4" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.237, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[15] Silin Gao, Yichi Zhang, Zhijian Ou, and Zhou Yu. Paraphrase augmented task-oriented dialog generation. In ACL, pages 639-649, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.137, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander G Hauptmann. Excl: Extractive clip localization using natural language descriptions. In NAACL, pages 1984-1990, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.195, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[17] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE TPAMI, 38(1): 142-158, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.252, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[18] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.31, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.34, + 0.905, + 0.394 + ], + "angle": 0, + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.397, + 0.905, + 0.45 + ], + "angle": 0, + "content": "[21] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, pages 14271-14280, 2024. 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[22] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, pages 202-218. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[23] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL, pages 470-490, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.556, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[24] Jiabo Huang, Hailin Jin, Shaogang Gong, and Yang Liu. Video activity localisation with uncertainties in temporal boundary. In ECCV, pages 724-740. Springer, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.6, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[25] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.63, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[26] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, pages 2466-2475, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.687, + 0.905, + 0.741 + ], + "angle": 0, + "content": "[27] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.745, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[28] Dahye Kim, Jungin Park, Jiyoung Lee, Seongheon Park, and Kwanghoon Sohn. Language-free training for zero-shot video grounding. In WACV, pages 2539-2548, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In ICCV, pages 706-715, 2017. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.523, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[30] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.523, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24305" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.151, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[32] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.194, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[33] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. TMLR, 2025. 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.252, + 0.482, + 0.307 + ], + "angle": 0, + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.31, + 0.482, + 0.364 + ], + "angle": 0, + "content": "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.435 + ], + "angle": 0, + "content": "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.44, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[37] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[38] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. CVPR, pages 7492-7500, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[39] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, pages 6657-6678, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.599, + 0.482, + 0.666 + ], + "angle": 0, + "content": "[40] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.671, + 0.482, + 0.725 + ], + "angle": 0, + "content": "[41] Daizong Liu, Xiaoye Qu, Jianfeng Dong, Pan Zhou, Yu Cheng, Wei Wei, Zichuan Xu, and Yulai Xie. Context-aware biaffine localizing network for temporal sentence grounding. In CVPR, pages 11235-11244, 2021. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.728, + 0.482, + 0.782 + ], + "angle": 0, + "content": "[42] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.786, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[43] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.844, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[44] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[46] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. IEEE Transactions on Multimedia, 25:7747-7757, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.15, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[47] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[48] Kajing Ma, Xianghao Zang, Zerun Feng, Han Fang, Chao Ban, Yuhan Wei, Zhongjiang He, Yongxiang Li, and Hao Sun. Llavilo: Boosting video moment retrieval via adapter-based multimodal modeling. In ICCV Workshops, pages 2790-2795. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.291, + 0.905, + 0.345 + ], + "angle": 0, + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.348, + 0.905, + 0.401 + ], + "angle": 0, + "content": "[50] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In ICCV, pages 15201-15213, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.405, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[51] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.434, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[52] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[53] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.548, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[54] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, pages 10810-10819, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[55] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, pages 407-426. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.648, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[56] Jinwoo Nam, Daechul Ahn, Dongyeop Kang, Seong Jong Ha, and Jonghyun Choi. Zero-shot natural language video localization. In ICCV, pages 1470-1479, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.69, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[57] Phuc Xuan Nguyen, Tianzhu Liu, Gaurav Prasad, Hung Hai Bui, Binh Pham, and Svetha Venkatesh. Weakly supervised action localization by sparse temporal pooling network. In CVPR, pages 6752-6761, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.747, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[58] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.803, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[59] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, pages 10870-10879, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.861, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[60] Sujoy Paul, Sourya Roy, and Amit K Roy-Chowdhury. W-talc: Weakly-supervised temporal activity localization and classification. In ECCV, pages 563-579, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24306" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[61] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In ICCV, pages 5285-5297, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.162, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[62] Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Rephrase, augment, reason: Visual grounding of questions for vision-language models. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.204, + 0.483, + 0.259 + ], + "angle": 0, + "content": "[63] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.26, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[64] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. NeurIPS, 37:119336-119360, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.482, + 0.356 + ], + "angle": 0, + "content": "[65] Hangyu Qin, Junbin Xiao, and Angela Yao. Question-answering dense video events. arXiv preprint arXiv:2409.04388, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[66] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[67] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, pages 13009-13018, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.484, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[68] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In CVPR, pages 6517-6525, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[69] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.553, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[70] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 1:25-36, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.595, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.651, + 0.482, + 0.705 + ], + "angle": 0, + "content": "[72] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 1, 2, 5, 6, 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.762 + ], + "angle": 0, + "content": "[73] Seyed Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. CVPR, pages 658-666, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.763, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[74] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, pages 2464-2473, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[75] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hongdong Li, and Stephen Gould. Dori: Discovering object relationships for moment localization of a natural language query in a video. In WACV, pages 1079–1088, 2021. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[76] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hiroya Takamura, and Qi Wu. Memory-efficient temporal moment localization in long videos. In EACL, pages 1909–1924, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.149, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[77] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.178, + 0.905, + 0.258 + ], + "angle": 0, + "content": "[78] Himanshu Sharma, Manmohan Agrahari, Sujeet Kumar Singh, Mohd Firoj, and Ravi Kumar Mishra. Image captioning: a comprehensive survey. In 2020 International Conference on Power Electronics & IoT Applications in Renewable Energy and its Control (PARC), pages 325-328. IEEE, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.261, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[79] Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Silvia Cascianelli, Giuseppe Fiameni, and Rita Cucchiara. From show to tell: A survey on deep learning-based image captioning. IEEE TPAMI, 45(1):539-559, 2022. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.316, + 0.905, + 0.369 + ], + "angle": 0, + "content": "[80] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In AAAI, pages 4998-5007, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.372, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[81] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[82] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. In CVPR, pages 13569-13580, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.47, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[83] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.526, + 0.905, + 0.566 + ], + "angle": 0, + "content": "[84] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In ICCV, pages 1397-1407, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.568, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[85] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.637, + 0.905, + 0.705 + ], + "angle": 0, + "content": "[86] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.707, + 0.905, + 0.76 + ], + "angle": 0, + "content": "[87] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024.5,6,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.762, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[88] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, pages 18699-18708, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[89] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. In ECCV, pages 39-58. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[90] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, pages 13204-13214, 2024. 2, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24307" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[91] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In CVPR, pages 13623-13633, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[92] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, pages 124-141, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.207, + 0.483, + 0.275 + ], + "angle": 0, + "content": "[93] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[94] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2024. 2, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.32, + 0.483, + 0.388 + ], + "angle": 0, + "content": "[95] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In Proceedings of the 2nd international workshop on human-centric multimedia analysis, pages 13-21, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[96] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[97] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[98] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.575, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[99] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[100] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. In EMNLP, pages 21715–21737, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.702, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[101] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, pages 492-510. Springer, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[102] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, pages 6543-6554, 2020. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.788, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[103] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[104] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In CVPR, pages 8327–8336, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[105] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[106] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[107] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. IEEE TPAMI, 44(12):9073-9087, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[108] Zhu Zhang, Zhijie Lin, Zhou Zhao, and Zhenxin Xiao. Cross-modal interaction networks for query-based moment retrieval in videos. In ACM SIGIR, pages 655-664, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.319 + ], + "angle": 0, + "content": "[109] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.321, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[110] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.904, + 0.403 + ], + "angle": 0, + "content": "[111] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[112] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. CVPR, pages 8739-8748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[113] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[114] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[115] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, pages 18243-18252, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[116] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. ArXiv, abs/2204.08121, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.758 + ], + "angle": 0, + "content": "[117] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 5651-5665, Gyeongju, Republic of Korea, 2022. International Committee on Computational Linguistics. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[118] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.815 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24308" + } + ] +] \ No newline at end of file diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_origin.pdf b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..5ad34560f5405150b11e176c90a74c1627bf56f1 --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/6520b7cd-57d9-4f71-bc18-4af7c56a0a7f_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:88844206c74a02d5233d24639ac430eb4354a625f0369808f24ff3e583fd4da6 +size 3376576 diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/full.md b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2f253adc1005a78e9a09b34418c68d512fa8ecf2 --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/full.md @@ -0,0 +1,380 @@ +# Enrich and Detect: Video Temporal Grounding with Multimodal LLMs + +Shraman Pramanick $^{\dagger,2}$ Effrosyni Mavroudi $^{1}$ Yale Song $^{1}$ Rama Chellappa $^{2}$ Lorenzo Torresani $^{\diamond 1}$ Triantafyllos Afouras $^{1}$ + +$^{1}$ FAIR, Meta $^{2}$ Johns Hopkins University $^{3}$ Northeastern University + +# Abstract + +We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios. + +# 1. Introduction + +Video temporal grounding [32, 40, 53, 80, 102] aims to identify temporal intervals in a video that correspond to a set of provided language queries. The task is essential for applications such as video editing and content retrieval. Conversely, video captioning [1, 59, 78, 79, 104] entails generating a natural language description for a given video segment, effectively translating visual content into text. The two tasks are in fact dual, as the outputs of one task are the inputs to the other, and vice versa. Intuitively, there is significant potential in exploiting this synergy, however it has largely remained unexplored, with previous works typically specializing in one of the two tasks [8, 32, 53, 104] or solv- + +![](images/f87cd08c5ac3e0806cb62d1dcb85321e4020c190ba7d7613f840bab5a47ea8e8.jpg) +Figure 1. Our proposed system. ED-VTG performs video temporal grounding as a two-stage process: since user queries are often incomplete or coarse, the first stage involves producing an enriched query that adds additional details to the original, making it easier to ground. Meanwhile, a contextualized embedding is generated, containing all the information about the interval to be predicted. In the second stage, an interval decoder translates these embeddings into precise temporal boundaries. + +ing them in multi-task setting [72, 85] without investigating how each task can benefit the other. + +In this work, we exploit this duality by leveraging captioning to enhance grounding. Our key observation is that natural language queries often lack the completeness or detail necessary for effective temporal localization; indeed, existing grounding datasets frequently contain poorly worded, coarse, and potentially incomplete queries. The quality and completeness of these queries however is crucial for the precision of the temporal grounding. A natural hypothesis, therefore, is that more detailed queries, which could be obtained via conditional captioning, can significantly enhance grounding accuracy. For example, as illustrated in Figure 1, a vague query like 'Man starts surfing' can be refined into a more detailed description such as 'The man with a yellow surfboard slowly runs to start surfing,' resulting in grounding with more accurate temporal boundaries. In other cases, refining an abstract concept may involve breaking down a complex query into simpler components that are more directly groundable. + +This key idea of query enrichment forms the basis of our proposed approach. Concretely, we transform grounding into a two-stage reasoning process using a multi-modal + +LLM: first the model enriches the input query into a more detailed description by adding missing details based on the video content, and then temporally localizes the resulting enriched query in the video. + +To effectively perform the temporal localization, we introduce a lightweight perception decoder that specializes in generating precise temporal boundaries, conditioned on a contextual representation, allowing the LLM to focus on language outputs, where it excels. The perception decoder allows us to leverage purposefully crafted training objectives for temporal grounding, building on prior knowledge and task-specific characteristics developed from years of research in object detection [17, 19, 30, 67, 71, 77] and temporal localization [32, 40, 53, 61, 102, 105]. + +Learning to jointly enrich and detect requires high-quality enriched query labels. We obtain those by using a strong external captioning model which we condition on the original queries and the video content of the target temporal boundaries. However powerful, these models are prone to hallucinations and there is no guarantee that the enriched queries will always be easier to ground than the original ones. At the same time, annotating the ground truth to determine which query - original or enriched - is a better candidate, is extremely expensive. To address this issue, we propose training in a multiple-instance learning (MIL) fashion, that enables the model to autonomously determine which query is better suited for the task during training. + +Finally, we note that our proposed method is not equivalent to a data augmentation approach which simply preprocesses the training set to generate enhanced queries that are directly used for training. While this simpler alternative offers some of the same benefits, it suffers from the limitation that extracting enriched queries during training requires knowledge of the ground-truth temporal segments, i.e., the grounding targets. Since during inference these segments are unknown, the original queries must be used as input, which, as we will show experimentally, is suboptimal. Our method overcomes this limitation by learning to jointly enrich and detect, demonstrating superior performance. + +To summarize, our contributions are as follows: (i) we introduce a cascaded approach to temporal grounding, where the model first enriches the provided language query based on the video context and then proceeds to localize it; (ii) we enable multi-modal LLMs to accurately localize text queries using a lightweight decoder which allows for training with detection objectives tailored to the task; (iii) we propose a multiple-instance learning paradigm that enables the model to dynamically select the query that leads to better temporal localization; (iv) we achieve state-of-the-art results on several temporal grounding benchmarks, for both single query grounding and paragraph grounding, demonstrating, for the first time, an LLM-based model that surpasses or performs comparably to specialist models. + +# 2. Related Works + +LLM-based temporal grounding. Prior works have explored using LLMs for grounding natural language sentences in videos, either using raw text tokens to represent timestamps [21, 36, 39, 48, 72, 99] or by adding hundreds of special tokens to the LLM's vocabulary to represent video frames [22, 63, 85]. Our approach differs from these methods in that by utilizing a lightweight interval decoder we can apply detection losses such as L1 and gIoU with minimal added complexity; we additionally exploit the LLM's potential to describe video content in detail. + +Specialist models. There is a rich variety of specialist models in the literature that are tailored to specific variants of temporal grounding, e.g. single query temporal grounding [102] and video paragraph grounding [82], and as such, achieve strong performance [6, 26, 41, 46, 75, 82, 98, 105, 106]. Modern methods typically employ a multi-modal transformer [102] that fuses dense video features with text embeddings of the language query, followed by a specialized detector head for performing detections [16, 32, 53, 74-76]. However, because these models are often trained on limited datasets for such a narrowly defined task, they struggle with generalization. Indeed, the zero-shot performance of these methods is limited [13, 89, 90, 92, 94]. In this work, we aim to address the shortcomings of previous methods by fully combining the generalization abilities of multi-modal LLMs with the advantages of specialist models. + +Dense captioning. Our method is related to dense video captioning [29], where the objective is to segment a given video into multiple parts and simultaneously provide descriptive captions for each segment. Traditional approaches tackle this task either by first determining the segments and then providing descriptions [6, 21, 25, 38, 111, 113], or jointly learning both tasks [10, 38, 93, 112, 116]. Recent advances have demonstrated video-conditioned LLMs to excel in this task [88, 115, 117]. While there are similarities between our enrich-and-detect paradigm and dense captioning, we solve a different task, namely video temporal grounding, where the input query is given and constrains the problem. + +Prompt augmentation with LLMs. Off-the-shelf LLMs have recently been used to augment input prompts for tasks such as image retrieval and image classification by providing additional, clarifying descriptions. These augmented descriptions can aid generalization in various vision and NLP tasks such as visual question answering [12, 62, 84], dialog generation [15], and visual classification [51]. Our approach goes beyond using an off-the-shelf LLM for offline query enrichment, rather learning to dynamically enrich queries during inference time towards better grounding. + +Prior work has also investigated the use of captioners and language models for data augmentation by rephras + +![](images/13e74b12896c30c0d81ec6251001ef8b5548091ae7ff3f03979c3fbe2181b8f2.jpg) +Figure 2. Overview of the proposed Enrich and Detect framework: (left) ED-VTG pipeline: Given an untrimmed video and a query $Q$ to be grounded, the inputs are first tokenized into video tokens $T_V$ and text tokens $T_Q$ . The tokens are then fed into an LLM, which first generates an enriched query by, e.g., filling in any missing details and then emitting an interval token $<\mathrm{INT}>$ . The embedding of this special token is finally decoded into the predicted temporal interval via a lightweight interval decoder. In the example shown here, the vague input query is enriched into a more detailed one by our model which can be subsequently grounded more easily. (right) Training: ED-VTG is trained using ground-truth temporal intervals and pseudo-labels of enriched queries, generated by an external off-the-shelf captioning model [33], which – unlike our model at inference time – has access to the ground-truth intervals. For every sample during training, the proposed multiple instance learning (MIL) framework allows ED-VTG to assess both the original or the enriched queries and generate two sets of predictions, $\hat{I}^{enr}$ , the interval predicted using the enriched query, and $\hat{I}^{dir}$ using the original query. Next, the model backpropagates using the better prediction (i.e., lower grounding loss). Hence, during training, ED-VTG dynamically learns to decide for which sample enrichment is necessary and, based on that, performs detection. + +![](images/4bf8cc78fe8deee8d7277a221578c7477b596e6b7051152da2e6d902c5fa8399.jpg) + +ing existing descriptions or providing new improved ones. LaViLa[109] uses rephraser and narrator models to improve the quality of the training data for video-text alignment. Augmentation during training is sufficient for that task, as the goal is to learn joint representations; for our grounding task however simply augmenting the training set, without enriching queries during inference is not as effective, as we empirically demonstrate. + +Multiple instance learning. Multiple instance learning (MIL)[11] is a technique commonly used in weakly-supervised vision problems, when a collection of potential solutions is available but exact annotations are not. It has been successfully applied to a range of tasks, including classification [3], weakly supervised object detection[7, 9] and temporal action localization[57, 60]. + +# 3. Method + +Given an untrimmed video $\mathcal{V}$ and a set of $N$ associated textual queries $\mathcal{Q} = \{Q_1, \ldots, Q_N\}$ , temporal grounding aims to identify the corresponding temporal interval for each query. Formally, the output is a set of temporal intervals $\mathcal{I} = \{I_1, \ldots, I_N\}$ . For $N = 1$ , the task takes the form of single-query temporal grounding, which we will simply refer to as STG. For clarity, we describe our approach in the STG setting, without loss of generality; the pipeline however readily extends to $N > 1$ . + +Our approach aims to tackle temporal grounding by leveraging a multimodal LLM that $(i)$ transforms input queries into intermediate, enriched queries by adding missing details based on the video input, $(ii)$ generates contextualized embeddings of each latent segment to enable temporal interval prediction, $(iii)$ decodes these embeddings into concrete temporal boundaries. + +In the following, we formally introduce our proposed ED-VTG model in Section 3.1, and proceed to describe how to train it with enriched query pseudo-labels within a MIL framework (Section 3.2). + +# 3.1. Model + +Our model (Figure 2) consists of three key modules: a vision encoder that extracts video representations, a LLM that jointly processes video and language, and a lightweight interval decoder that generates precise temporal boundaries. + +Enrich. Given a single input query $Q$ about a video $V$ with $T$ frames, the vision encoder represents the video as a sequence of $R$ visual tokens $\mathbf{T}_{\mathrm{V}} \in \mathbb{R}^{R \times D}$ , where $D$ is the token dimension. The tokenized video features $\mathbf{T}_{\mathrm{V}}$ are fed along with the tokenized query $\mathbf{T}_{\mathcal{Q}}$ to the LLM, which generates an enriched query $\hat{Q}^{enr} = \{\hat{y}_1, \dots, \hat{y}_l, \dots, \hat{y}_{L^{enr}}\}$ one token at a time: + +$$ +\hat {y} _ {l} = \mathcal {F} _ {\mathrm {L L M}} \left(\hat {y} _ {< l}, \mathbf {T} _ {\mathrm {V}}, \mathbf {T} _ {\mathcal {Q}}\right). \tag {1} +$$ + +When the model is ready to ground the query, the LLM emits a new, special token $<\mathrm{INT}>$ to trigger interval prediction. In other words, the text prediction takes the form: + +$$ +\hat {\mathbf {y}} = \text {" T h e q u e y} \hat {Q} ^ {e n r} \text {o c c u r s a t} < I N T > \tag {2} +$$ + +Detect. To detect the temporal interval corresponding to the enriched query $\hat{Q}^{enr}$ , we introduce an interval decoder $\mathcal{F}_{\mathrm{dec}}$ that predicts an interval $\hat{\mathbf{I}}$ parametrized by the center $\hat{c}$ and width $\hat{w}$ of the predicted interval. We selected this parameterization for its advantage in decoupling position from scale, as supported by literature in object detection [42, 68, 69, 114]. The interval decoder takes the form + +$$ +\tilde {\mathbf {I}} = (\hat {c}, \hat {w}) = \mathcal {F} _ {\mathrm {d e c}} \left(\mathcal {G} \left(\mathbf {h} _ {\text {i n t}}\right), \mathbf {T} _ {\mathrm {V}}\right), \tag {3} +$$ + +where $\mathbf{h}_{int} \in \mathbb{R}^D$ is the hidden state of the LLM corresponding to the token and $\mathcal{G}$ is a linear projection layer. The decoder, functioning as the regression component of a temporal detector, consists of two transformer layers followed by a multi-layer perceptron (MLP), which processes a concatenation of its two inputs, $(\mathcal{G}(\mathbf{h}_{\mathrm{int}}), \mathbf{T}_{\mathrm{V}})$ . It finally outputs the predicted interval $\hat{\mathbf{I}}$ , grounding the input query $\mathcal{Q}$ . + +Notice that our formulation allows the quality of the enriched query $\hat{Q}^{enr}$ to directly influence the accuracy of the predicted interval $\hat{\mathbf{I}}$ . This is because the generation of the interval token $<\mathrm{INT}>$ (and consequently its hidden state $\mathbf{h}_{\mathrm{int}}$ ) is conditioned on the enriched query prediction. Establishing this cascaded dependency chain, i.e., $(V, Q) \to \hat{Q}^{enr}$ and $(V, \hat{Q}^{enr}) \to \hat{I}$ , is the key idea of our approach. + +# 3.2. Training + +The model is trained end-to-end using two primary loss functions: a language modeling loss $\mathcal{L}_{\mathrm{LM}}$ and a temporal grounding loss $\mathcal{L}_{\mathrm{grnd}}$ that help supervise the "enrich" and "detect" aspects of our model, respectively. + +Enrich. Given the target output text $\mathbf{y} = \{y_1, y_2, \dots, y_L\}$ , $\mathcal{L}_{\mathrm{LM}}$ is calculated as the cross-entropy loss that evaluates the likelihood of $\mathbf{y}$ under the predicted probability distribution generated by the model: + +$$ +\mathcal {L} _ {\mathrm {L M}} = - \sum_ {t = 1} ^ {T} \log P \left(y _ {l} \mid \mathbf {y} _ {< l}, \mathbf {T} _ {\mathrm {V}}, \mathbf {T} _ {\mathcal {Q}}\right) \tag {4} +$$ + +To provide a proper supervisory signal for query enrichment, we need the ground-truth pair of $(Q, Q^{enr})$ . However, such datasets do not exist, nor is it practical to annotate a dataset solely for this purpose. Here, we capitalize on the tremendous amount of progress made in the video captioning literature [1, 78, 79], and use an off-the-shelf captioning model [33] to generate pseudo ground-truth $Q^{enr}$ by refining the original query $Q$ given its video V (see more details in Section 4.1). + +
DatasetDomainTasksCorpusEval. Protocol# Train SamplesAvg Vid Length (s)Avg Span Length (s)
DiDeMo [4]OpenSTGPT-32.8K54.576.49 (11.9%)
QuerYD [58]CookingSTGPT-13.6K158.787.68 (4.8%)
COIN [83]OpenVPGPT-7.5K143.7115.06 (10.5%)
HiREST [96]OpenSTG, VPGPT-0.8K208.0344.50 (21.4%)
ViTT† [23]OpenVPGPT-4.9K287.17-
YTTemporal [97]OpenVPGPT-28.8K327.364.0 (1.2%)
CrossTask [118]ProceduralAGPT-2.7K297.09.61 (3.2%)
VideoCC [55]OpenSTGPT-45.0K415.899.88 (2.3%)
Charades-STA [14]IndoorSTGFT, EvalZS, FT12.4K31.178.29 (26.6%)
Charades-CD-OOD [95]IndoorVPGFT, EvalFT4.5K30.607.90 (25.8%)
ANet-Captions [29]OpenSTG, VPGFT, EvalZS, FT9.5K117.6335.61 (30.3%)
TACoS [70]CookingSTG, VPGFT, EvalZS, FT9.8K224.3423.33 (10.4%)
YouCook2 [110]CookingVPGFT, EvalFT1.2K311.4120.07 (6.4%)
NEX-T-GQA° [90]OpenQGEvalZS-39.606.69 (16.9%)
HT-Step [2]CookingAGFT, EvalFT17.4K393.8914.88 (3.7%)
+ +Table 1. Dataset statistics, corresponding tasks, and evaluation protocol. The upper side of the table represents datasets used for pre-training, resulting in a total of 136K samples. The lower side represents datasets used for fine-tuning and evaluation. We cover four different video grounding tasks: single-query temporal grounding (STG), video paragraph grounding (VPG), question grounding (QG), and article grounding (AG). QG is used only for evaluation to assess the model's generalization capability. Average interval lengths compared to the corresponding video durations are shown in brown, denoting the annotation granularity. $^\dagger$ VITT contains single timestamp annotation instead of intervals. $^\circ$ NExTGQA contains only evaluation split. + +Detect. Given a target temporal interval $I = (c, w)$ , the grounding loss $\mathcal{L}_{\mathrm{ground}}$ is computed as a combination of the L1 loss and the generalized Intersection over Union (gIoU) loss [73], applied on the predicted temporal interval $(\hat{c}, \hat{w})$ : + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {g r n d}} = \lambda_ {\mathrm {L 1}} \left(\left| (\hat {c} - c | + | \hat {w} - w |\right) \right. \\ + \lambda_ {\mathrm {g I o U}} \mathrm {g I o U} ((\hat {c}, \hat {w}), (c, w)) \tag {5} \\ \end{array} +$$ + +# 3.2.1. MIL framework + +A caveat with the pseudo-labeled enriched queries is that they can be noisy and include hallucinations; as a result, some of them may lead to a temporal interval prediction inferior to the original query. To address this, we propose considering multiple options for the target query (2 in the case of a single query scenario, $Q$ and $Q^{enr}$ ), by adopting a multiple instance learning (MIL) framework [11]. This approach allows the model to choose between the enriched and the original input queries during training, depending on which version leads to a better temporal interval prediction. During inference, this means that our ED-VTG model can choose to either enrich the query if it misses important details, or decide to "carry over" the original query into $\hat{Q}^{enr}$ when it is concrete enough (in which case $Q = \hat{Q}^{enr}$ ). + +Formally, to perform MIL, we collect two temporal interval predictions by running two forward passes with different LLM inputs (in a teacher-forcing fashion), i.e. $\mathbf{y}^{\mathrm{dir}}$ that is formed using the original query $\mathcal{Q}$ , and $\mathbf{y}^{\mathrm{enr}}$ using the pseudo-labeled enriched query $\mathcal{Q}^{\mathrm{enr}}$ . We illustrate this schematically in Figure 2 (right). Recall that through the contextualized interval representation $\mathbf{h}_{\mathrm{int}}$ , the predicted interval depends on the LLM teacher-forced input. Hence, these inputs produce respective interval predictions, $\hat{\mathbf{I}}^{\mathrm{dir}}$ and + +
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
UniVTG [40]X4.2MZS44.125.210.027.1----5.21.30.34.4
SeViLA [94]X129MZS----31.619.010.123.0----
PSVL [56]X-ZS46.231.314.231.244.730.114.729.6----
LT-ZVG [28]X-ZS52.937.219.336.047.632.615.431.8----
Video-LLaMA° [103]✓2.7MZS25.210.63.416.821.910.84.916.55.11.20.83.4
Video-ChatGPT° [49]✓100KZS27.26.21.919.719.510.64.814.26.31.71.04.3
Valley [47]✓100KZS28.41.80.321.430.613.78.121.9----
VideoChat2 [36]✓2MZS38.014.33.824.640.827.89.327.9----
Momenter [63]✓10MZS42.626.611.628.542.923.012.429.3----
VTimeLLM° [21]✓170KZS51.027.511.431.244.027.814.330.47.01.80.84.5
TimeChat° [72]✓125KZS-32.213.4-----6.82.10.84.7
HawkEye [87]✓715KZS50.631.414.533.749.129.310.732.7----
ChatVTG [66]✓100KZS52.733.015.934.940.722.59.427.28.13.71.35.5
ED-VTG✓136KZS59.539.319.840.252.133.116.035.214.56.02.312.7
ΔOurs - HawkEye--ZS8.9↑7.9↑5.3↑6.5↑3.0↑3.8↑5.3↑2.5↑----
ΔOurs - ChatVTG--ZS--------6.4↑2.3↑1.0↑7.2↑
+ +Table 2. Zero-shot STG results on Charades, ActivityNet, and TACoS test splits. For all three datasets, ED-VTG gains significant improvement over all existing methods, including task-specific, non-generalist models. We use boldface for the best and underline the second-best result for each metric, among the generalist models. $^\circ$ Official checkpoints are used for TACoS evaluation. + +$\hat{\mathbf{I}}^{\mathrm{enr}}$ . We select the query version that results in the smallest grounding loss and use it to compute our overall training objective: + +$$ +\mathcal {L} = \left\{ \begin{array}{l l} \lambda_ {\mathrm {L M}} \mathcal {L} _ {\mathrm {L M}} ^ {\mathrm {d i r}} + \lambda_ {\mathrm {g r n d}} \mathcal {L} _ {\mathrm {g r n d}} ^ {\mathrm {d i r}} & \text {i f} \mathcal {L} _ {\mathrm {g r n d}} ^ {\mathrm {d i r}} < \mathcal {L} _ {\mathrm {g r n d}} ^ {\mathrm {e n r}} \\ \lambda_ {\mathrm {L M}} \mathcal {L} _ {\mathrm {L M}} ^ {\mathrm {e n r}} + \lambda_ {\mathrm {g r n d}} \mathcal {L} _ {\mathrm {g r n d}} ^ {\mathrm {e n r}} & \text {o t h e r w i s e} \end{array} \right. \tag {6} +$$ + +where $\mathcal{L}_{\mathrm{LM}}^{\mathrm{dir}}$ and $\mathcal{L}_{\mathrm{LM}}^{\mathrm{enr}}$ are the language losses for targets $\mathbf{y}^{\mathrm{dir}}$ and $\mathbf{y}^{\mathrm{enr}}$ respectively, and $\mathcal{L}_{\mathrm{grnd}}^{\mathrm{dir}}$ and $\mathcal{L}_{\mathrm{grnd}}^{\mathrm{enr}}$ the grounding losses for predictions $\hat{\mathbf{I}}^{\mathrm{dir}}$ and $\hat{\mathbf{I}}^{\mathrm{enr}}$ respectively. The hyperparameters $\lambda_{\mathrm{LM}}$ and $\lambda_{\mathrm{grnd}}$ are the relative weights of the language modeling and grounding losses. + +# 4. Experiments + +We design experiments to study three key questions related to our architecture and training framework: Q1) How does ED-VTG perform on various video grounding tasks in comparison to the current state-of-the-art? Q2) How beneficial is our query enrichment approach, within the MIL paradigm, as opposed to directly grounding the original queries? Q3) Does utilizing an interval decoder with specifically tailored grounding objectives offer advantages over predicting timestamps as raw text or special tokens? + +# 4.1. Datasets + +Table 1 summarizes the datasets that we used during the pre-training and fine-tuning stages. During pre-training, we use a total of 136K medium-to-long duration videos from 8 public datasets annotated with text queries and the corresponding intervals. As a preprocessing step, we collect pseudo-labels for enriched queries using an external captioning model [33]. In short, we take each video segment defined by the ground-truth intervals and prompt the captioning model to enrich the original query while preserving its meaning given the video segment. We provide full details of this process with the exact prompt used in the supplementary material. + +# 4.2. Tasks + +Table 1 also summarizes the tasks for which we use each dataset. Here we briefly describe these tasks, and highlight how our approach is applied to solve them. + +Single-Query Temporal Grounding (STG) involves identifying a single time window in response to a single input language query $(N = 1)$ . + +Video Paragraph Grounding (VPG) involves grounding $N > 1$ sentences to $N$ corresponding time windows. Our ED-VTG model can be trained to predict multiple enriched queries (one per original input query) interleaved with $<\mathrm{INT}>$ tokens that get separately decoded into intervals. Since there are multiple queries in the input, we run multiple forward passes through the LLM to perform MIL, wherein each pass selects a random number of queries to be enriched. + +Question Grounding (QG) involves retrieving evidence intervals to answer questions, facilitating explainable video QA. As in STG, the input is a single query, and the output a single interval. + +Article grounding (AG) is an extension of VPG where the model is given multiple queries as input, some of which may not be groundable in the video. Hence, the model must 1) identify which queries are groundable and 2) predict intervals for the groundable queries. + +Note that, while we train our model jointly on multiple tasks, they are unified as a single task in the form of (Video, Text) $\rightarrow$ (Time intervals). Besides the differences in training data sources, we use the identical training objective (Eq. 6) during pre-training and fine-tuning stages. + +# 4.3. Evaluation Protocol + +Following the common practices in the literature [21, 72, 87], we evaluate ED-VTG in two primary evaluation protocols: (i) Zero-shot (ZS), where the pre-trained model + +
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
VSLNet (C3D) [102]X-FT64.347.330.245.263.243.226.243.229.624.320.024.1
MS-2D-TAN (I3D) [107]X-FT-56.636.2-62.145.528.3-42.033.622.1-
Moment-DETR [32]X236KFT65.852.130.645.5----38.024.712.025.5
UnLoc-B [91]X650KFT-58.135.4--48.029.7-----
MomentDiff [37]X-FT-55.632.4-----46.628.912.430.4
LGI [54]X-FT73.059.535.551.458.541.523.141.1----
BAM-DETR [31]X-FT72.960.039.452.3----56.741.526.839.3
InternVideo2* + CG-DETR [86]X2.1MFT79.770.048.958.8--------
SG-DETR [18]X-FT-71.152.860.7-----46.433.942.4
EMB (ELA) [24]X-FT79.769.251.462.273.758.740.756.263.352.537.048.4
BLIP-2 (frames only) [34]✓129MFT-43.332.6--25.89.7-----
VideoChat2 [36]✓2MFT----55.534.717.738.9----
TimeChat [72]✓125KFT-46.723.7-----27.715.16.418.4
HawkEye [87]✓715KFT72.558.328.849.355.934.717.939.1----
VtimeLLM [21]✓170KFT--------26.814.46.118.0
ED-VTG✓136KFT78.262.135.052.667.645.122.744.946.031.515.832.4
ΔOurs - HawkEye--FT5.7↑3.8↑6.2↑3.3↑11.7↑10.4↑4.8↑5.8↑----
ΔOurs - VTimeLLM--FT--------19.2↑17.1↑9.7↑14.4↑
+ +Table 3. Fine-tuned STG results on Charades, ActivityNet, and TACoS test splits. For all datasets, ED-VTG achieves strong improvements over previous generalist models, and performs comparably to task-specific expert models. *Though InterVideo2 is a generalist model, it fine-tunes CG-DETR [52] head for grounding tasks, using the LLM as a feature extractor. For completeness, we report all competitive existing works, including task-specific SOTA specialist models (shown in gray), but directly compare ED-VTG only with generalist frameworks - using boldface for the best and underlining the second best results among generalist models (see discussion in Section 4.3). + +is assessed directly without any fine-tuning on downstream datasets, and (ii) Fine-tuned (FT), where the model undergoes additional training on specific tasks and datasets. We also present results without pre-training to assess its importance. For the STG and VPG tasks, following existing works [6, 21, 72, 76, 87], we report the mean intersection over union (mIoU) and Recall@1 for IoU $\geq m$ ( $R@m$ ), with $m \in \{0.3, 0.5, 0.7\}$ . For the QG task, we report intersection over prediction (IoP) in addition to IoU, following the NeXT-GQA [90] evaluation protocol. For AG, we adhere to the HT-Step [2] protocol and report article-grounding mAP scores across various IoU thresholds. + +Comparison to non-generalist models. In the results section, we report all existing SOTA and competitive methods for a complete comparison, however we divide them into generalist and non-generalist (specialist) models. We note that specialist methods are heavily tailored to the task and in practice often overfit specific datasets which limits their transferability (as is evidenced by the zero-shot results in Table 2). We therefore focus the discussion around our method's comparison to other generalist models. + +# 4.4. Implementation Details + +We initialize the video encoder and LLM with the VideoLLaMA-7B [103] checkpoint, which is a similarly sized backbone used in existing LLM-based video grounding models [21, 63, 72, 87]. Video-LLaMA is trained on video captioning tasks with WebVid [5] and VideoChat [35]. The video encoder includes a ViT-G/14 from EVA-CLIP [81] as the frame feature extractor, followed by image and video QFormers. We initialize the decoder with random weights. We keep the ViT frozen, apply LoRA [20] with rank 32 to the LLM and fully tune the Q-Formers, decoder, and linear + +layers. We use LLaVA OneVision (OV) 72B [33] as the external captioner for obtaining enriched queries in the training sets. For the VPG task, we run four forward passes to perform MIL. We pre-train our model for 40 epochs with a batch size 256, using AdamW [45] with a peak learning rate of 5e-5 and a cosine scheduler [44] with a linear warmup for the first $20\%$ steps. Pre-training takes 2 days on 16 V100 nodes (8 cards with 32G GPU memory each). Additional details on pre-training, fine-tuning, and task-specific instructions are provided in the supplementary material. + +# 4.5. Results + +Single-Query Temporal Grounding (STG). We start by comparing ED-VTG against the state-of-the-art across three different STG benchmarks, namely Charades-STA, ActivityNet-Captions, and TACoS, in a zero-shot evaluation setting. We show the results in Table 2. On Charades, ED-VTG achieves ZS scores of 59.5, 39.3, and 19.8 for R@0.3, R@0.5, and R@0.7, respectively, significantly outperforming all baseline models. Notably, ED-VTG surpasses Momenter [63] and HawkEye [87] by 11.4 and 6.2 absolute mIoU points despite these models being pretrained with 100x and 6x more segment-level data, respectively. A similar trend is observed on ActivityNet and TACoS where our model achieves improvements of 2.5 and 7.2 mIoU points over the nearest LLM-based models in zero-shot setting. Notably, the TACoS dataset contains short, under-specified queries, longer input videos, and fine-grained interval annotations, which pose challenges for LLMs with a fixed number of input frames. The enriched query descriptions that ED-VTG generates enable it to more accurately align them to video frames and precisely retrieve the correct intervals. + +On the fine-tuned STG evaluation, as shown in Table + +
MethodGeneralist ModelCharades-CD-OOD
R@0.3R@0.5mIoU
DepNet [6]✗45.627.629.3
DRN [98]✗40.530.4-
STLG [46]✗48.330.4-
SVPTR [26]✗50.328.532.1
SiamGTR [82]✗59.135.538.9
VTimeLLM1[21]✓53.234.035.1
TimeChat1[72]✓60.536.138.3
ED-VTG70.747.345.0
ΔOurs - TimeChat-10.2↑11.2↑6.7↑
+ +(a) Results on Charades-CD-OOD. + +
MethodGeneralist ModelANet-Captions
R@0.3R@0.5mIoU
CBLN [41]X66.348.127.6
2D-TAN [105]X59.544.5-
3D-TPN [106]X67.651.5-
DepNet [6]X72.855.9-
SVPTR [26]X78.161.755.9
VTimeLLM1[21]✓66.150.345.6
TimeChat1[72]✓67.951.547.0
ED-VTG✓74.158.053.7
ΔOurs - TimeChat-6.2↑6.5↑6.7↑
+ +(b) Results on ActivityNet. + +
MethodGeneralist ModelTACoS
R@0.3R@0.5mIoU
CMIN [108]✗24.618.1-
2D-TAN [105]✗37.325.3-
3D-TPN [106]✗40.326.5-
DepNet [6]✗41.327.2-
SVPTR [26]✗47.928.231.4
VTimeLLM†[21]✓40.225.627.9
TimeChat†[72]✓39.525.627.8
ED-VTG✓46.227.830.7
ΔOurs - TimeChat-6.7↑2.2↑2.9↑
+ +(c) Results on TACoS. + +
MethodGeneralist ModelYouCook2
R@0.3R@0.5mIoU
DORi [75]✗43.430.530.5
DORi* [75]✗42.329.929.9
LocFormer [76]✗46.831.330.9
ExCL [16]✗26.616.218.9
TMLGA [74]✗34.823.124.4
VTimeLLM1[21]✓41.318.524.3
TimeChat1[72]✓40.919.026.6
ED-VTG✓48.128.031.5
ΔOurs - TimeChat-7.2↑9.0↑4.9↑
+ +(d) Results on YouCook2. +Table 4. Performance on VPG task on four different benchmarks: Charades-CD-OOD, ActivityNet-Captions, TACoS, YouCook2. ED-VTG significantly improves over previous LLM-based models, and performs comparably to state-of-the-art specialist models. ${}^{ \dagger }$ We fine-tune VTimeLLM and TimeChat checkpoints for the VPG task. Dori ${}^{ \star }$ represents frozen text (BERT) encoder during fine-tuning. + +
MethodGeneralist ModelNExT-GQA
mIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VGT [89]✘24.726.024.63.04.21.4
VIOLETv2 [13]✘23.625.123.33.14.31.3
Temp[CLIP] NG+ [90]✘25.731.425.512.117.58.9
FrozenBiLM NG+ [92]✘24.228.523.79.613.56.1
SeViLA [94]✘29.534.722.921.729.213.8
\( LLoVi7B^{\circ} \) [100]✓20.7-20.58.7-6.0
VideoStreaming° [64]✓32.2-31.019.3-13.3
LongRepo 7B° [27]✓20.3-20.08.7-6.0
DeVi [65]✓33.8-32.220.717.4-
HawkEye [87]✓---25.737.019.5
ED-VTG✓34.745.133.526.639.519.8
\( \Delta_{Ours - SeViLA} \)-5.2↑10.4↑10.6↑4.9↑10.3↑6.0↑
\( \Delta_{Ours - HawkEye} \)----0.9↑2.5↑0.3↑
+ +Table 5. Performance on QG task on NeXT-GQA test split. ED-VTG consistently achieves consistent improvements over the existing models across all metrics. Results of LLoVi, LongRepo and VideoStreaming are from [64]. + +
MethodSeenUnseen
↑mAP@IoU↑mAP@IoU
@0.3@0.5@0.7@[0.3-0.7]@0.3@0.5@0.7@[0.3-0.7]
UMT [43]15.78.73.29.19.44.91.75.3
MT+BCE [2, 50]46.229.912.929.831.618.77.719.3
ActionFormer-T [101]41.230.818.330.229.720.310.720.4
Timechat† [72]45.329.014.429.030.717.87.518.7
ED-VTG48.931.518.032.533.021.211.121.6
ΔOurs - TimeChat3.6 ↑2.5 ↑3.6 ↑3.5 ↑2.3 ↑3.4 ↑3.6 ↑2.9 ↑
+ +Table 6. Performance on AG task on HT-Step seen and unseen val split. ED-VTG is the first LLM-based model to report results on for video grounding in the presence of negative, non-groundable queries and sets a new state of the art in terms of average mAP score across various IoU thresholds. We fine-tune the official TimeChat checkpoint for AG task. + +3. ED-VTG demonstrates an impressive gain of 5.7 and 11.7 points in R@0.3 over HawkEye on Charades and ActivityNet, respectively. Similarly on TACoS, ED-VTG surpasses all existing MLLMs by a considerable margin, e.g. 14.0 and $14.4\mathrm{mIoU}$ points over TimeChat and VTimeLLM. Furthermore, ED-VTG also beats many existing task-specific specialist models in all three benchmarks, significantly reducing the gap between specialist models and MLLMs for STG. + +Video Paragraph Grounding (VPG). Next, we fine-tune ED-VTG for the VPG task, where the model processes multiple input queries in a temporal sequence. While specialist models have reported results in the past, no LLM-based models have previously addressed this challenging task. To establish a baseline for comparison, we fine-tune the + +officially released VTimeLLM and TimeChat pre-trained checkpoints. As shown in Table 4a, ED-VTG significantly outperforms both LLMs on the Charades-CD-OOD dataset, achieving an absolute gain of 9.9 and $6.7\mathrm{mIoU}$ points over VTimeLLM and TimeChat, respectively. Additionally, ED-VTG surpasses all specialist models on this dataset by a substantial margin, setting a new state-of-the-art. + +We also evaluate VPG performance on three other benchmarks: ActivityNet-Captions, TACoS, and YouCook2, as presented in Tables 4b, 4c, and 4d. Consistent with the results on Charades, ED-VTG outperforms LLM-based models across all metrics on these datasets, with mIoU gains of 6.7, 2.9, and 4.9 over TimeChat on ActivityNet-Captions, TACoS, and YouCook2, respectively. ED-VTG also exceeds many existing specialist VPG models, demonstrating the effectiveness of our enriched queries and cascaded interval decoder. + +Question Grounding (QG). To further assess the model's generalization capabilities, we conduct zero-shot evaluation on the held-out QG task. Our results on the NeXT-GQA test set are presented in Table 5. Notably, ED-VTG achieves state-of-the-art performance across all metrics, outperforming both specialist and LLM-based models by a significant margin. Specifically, ED-VTG surpasses the existing best baseline, HawkEye, by 2.5 points in terms of IoU@0.3 score, demonstrating its strong generalizability. + +Article Grounding (AG). We assess ED-VTG on the AG on the HT-Step benchmark. Table 6 shows the fine-tuned performance on the AG task, where ED-VTG shows significant improvements over existing baselines. Notably, on the challenging unseen split, our model achieves the best results across all metrics, surpassing both the LLM and specialist models by a decent margin. The ability to handle non-groundable queries underscores ED-VTG's real-world applicability, as it does not always assume that the query is occurring in the input video. + +# 4.6. Ablation Study + +Effect of Enriched Queries. We examine the impact of query enrichment through a step-by-step ablation, as shown in Tables 7 and 8. Initially, we compare the effect of enrichment without MIL paradigm against direct grounding. As + +![](images/1e9534a438d01b1066b72f497c90405064b432e3e3ba753aca80688399119c6a.jpg) + +![](images/fc6f1677a7f4aa25e9c93618e186c22c2a6ec1544e1ea72eae13b5239fdd7c86.jpg) +Figure 3. Example of query enrichment and detection made by ED-VTG on video paragraph grounding (VPG) task from the ActivityNet-Captions [29] dataset. In this specific sample, three different queries are enriched and localized together. + +
Training ParadigmCharades-STA STGActivityNet-Captions STG
ZSFT w/o PTFTZSFT w/o PTFT
R@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect48.130.631.051.431.533.268.949.045.846.326.029.650.330.134.061.138.039.2
Enrich & Detect58.137.337.760.137.038.475.156.649.750.729.533.456.335.537.865.543.443.8
Enrich & Detect w/ MIL59.539.340.262.838.440.378.262.152.652.133.135.257.536.238.667.645.144.9
+ +Table 7. Ablation on the effect of enriched queries. Our proposed enrich & detect framework significantly gains over directly grounding the input queries across different evaluation settings. Introducing the MIL framework further improves the performance. FT w/o PT refers directly fine-tuning on respective datasets, without performing pre-training. + +
Training ParadigmCharades-STAANet-Captions
R@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect51.431.533.250.330.134.0
Offline Enrich + Detect51.731.533.449.829.933.7
Enrich & Detect60.137.038.456.335.537.8
+ +Table 8. Ablation on enrichment as a training pre-processing step. The two step enrich & detect framework is more helpful since the trained model learn to perform autonomous enrichment during evaluation. Reported results are in FT w/o PT setting. + +
DecoderObjectivesCharades-STA STGANet-Captions
LML1gIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
-✓--54.233.234.151.031.635.5
✓✓✓-58.536.037.055.834.437.1
✓✓-✓58.936.237.155.834.737.3
✓✓✓✓60.137.038.456.335.537.8
+ +Table 9. Ablation study on different training objectives. Training the model only using the LM loss (without the decoder) leads to significant performance drop. Results are in FT w/o PT setting. + +indicated in the first two rows of Table 7, enriched queries lead to significant improvements on the Charades-STA and ActivityNet-Captions STG benchmarks. In the zero-shot (ZS) setting, enrichment results in gains of 6.7 and 3.8 mIoU points on these datasets, respectively. Introducing the MIL paradigm further enhances performance, adding 2.5 and $1.8\mathrm{mIoU}$ points. Similar improvements are observed in other evaluation settings, highlighting the substantial effectiveness of query enrichment within the MIL framework. + +Additionally, we investigate the effect of offline enrichment in Table 8. In this scenario, instead of using a two-step grounding process, we enrich the queries as a training enrichment step, and the model is then directly provided with these enriched queries as input and asked to directly per + +form grounding. However, we find that offline enrichment is not advantageous, primarily due to the lack of enrichment during evaluation. In contrast, our two-step grounding approach allows the trained model to learn how to enrich queries and improve them autonomously when necessary, resulting in significant performance gains. + +Training Objectives. We ablate different training objectives on the Charades and ActivityNet STG benchmarks, as shown in Table 9. The decoder achieves optimal performance when both L1 and gIoU objectives are used together; omitting either one slightly reduces the scores. + +# 4.7. Qualitative Results and Error Analysis + +Figure 3 visualizes a VPG sample from the ActivityNet-Captions dataset where ED-VTG meaningfully enriches the three input queries, then proceeds to precisely ground them. Please refer to supplementary for more qualitative results, including intuitive demonstrations of the flexibility in choosing between enrichment or direct detection, and comparison with TimeChat baseline. We also present there some interesting failure cases, such as the ones involving small and obscured objects in long input videos. + +# 5. Conclusion + +In this paper, we presented ED-VTG, a novel method for fine-grained video temporal grounding using multi-modal LLMs. By enhancing queries with additional details, utilizing a lightweight decoder and trained in a multiple-instance framework, ED-VTG accurately locates temporal boundaries in videos. Our experiments show that our method outperforms existing LLM-based works and is competitive with specialized models, especially in zero-shot settings, setting a new, strong benchmark for video grounding tasks. + +# 6. Acknowledgement + +This codebase for this project is built on the TimeChat [72] and VTimeLLM [21] repository. We would like to thank the respective authors for their contributions. We gratefully acknowledge the following colleagues at FAIR for valuable discussions and support of our project: Tushar Nagarajan, Huiyu Wang, Yujie Lu, and Arjun Somayazulu. + +# References + +[1] Moloud Abdar, Meenakshi Kollati, Swaraja Kuraparthi, Farhad Pourpanah, Daniel McDuff, Mohammad Ghavamzadeh, Shuicheng Yan, Abduallah Mohamed, Abbas Khosravi, Erik Cambria, et al. A review of deep learning for video captioning. arXiv preprint arXiv:2304.11431, 2023. 1, 4 +[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In NeurIPS, 2024. 4, 6, 7 +[3] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. In NeurIPS, pages 561-568, 2003. 3 +[4] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 4 +[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6 +[6] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, pages 920-928, 2021. 2, 6, 7 +[7] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, pages 2846-2854, 2016. 3 +[8] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. NeurIPS, 34:28442-28453, 2021. 1 +[9] Ramazan Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multi-fold multiple instance learning. IEEE TPAMI, 39, 2015. 3 +[10] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2 +[11] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1):31-71, 1997. 3, 4 +[12] Li Dong, Jonathan Mallinson, Siva Reddy, and Mirella Lapata. Learning to paraphrase for question answering. In EMNLP, 2017. 2 +[13] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, pages 22898-22909, 2023. 2, 7 +[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 4 + +[15] Silin Gao, Yichi Zhang, Zhijian Ou, and Zhou Yu. Paraphrase augmented task-oriented dialog generation. In ACL, pages 639-649, 2020. 2 +[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander G Hauptmann. Excl: Extractive clip localization using natural language descriptions. In NAACL, pages 1984-1990, 2019. 2, 7 +[17] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE TPAMI, 38(1): 142-158, 2015. 2 +[18] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 6 +[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2 +[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6 +[21] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, pages 14271-14280, 2024. 2, 5, 6, 7, 9 +[22] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, pages 202-218. Springer, 2025. 2 +[23] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL, pages 470-490, 2020. 4 +[24] Jiabo Huang, Hailin Jin, Shaogang Gong, and Yang Liu. Video activity localisation with uncertainties in temporal boundary. In ECCV, pages 724-740. Springer, 2022. 6 +[25] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops, 2020. 2 +[26] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, pages 2466-2475, 2022. 2, 7 +[27] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024. 7 +[28] Dahye Kim, Jungin Park, Jiyoung Lee, Seongheon Park, and Kwanghoon Sohn. Language-free training for zero-shot video grounding. In WACV, pages 2539-2548, 2023. 5 +[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In ICCV, pages 706-715, 2017. 2, 4, 8 +[30] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 2 + +[31] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 6 +[32] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 1, 2, 6 +[33] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. TMLR, 2025. 3, 4, 5, 6 +[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 6 +[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6 +[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6 +[37] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36, 2024. 6 +[38] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. CVPR, pages 7492-7500, 2018. 2 +[39] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, pages 6657-6678, 2024. 2 +[40] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 5 +[41] Daizong Liu, Xiaoye Qu, Jianfeng Dong, Pan Zhou, Yu Cheng, Wei Wei, Zichuan Xu, and Yulai Xie. Context-aware biaffine localizing network for temporal sentence grounding. In CVPR, pages 11235-11244, 2021. 2, 7 +[42] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 4 +[43] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 7 +[44] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6 +[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6 + +[46] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. IEEE Transactions on Multimedia, 25:7747-7757, 2022. 2, 7 +[47] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5 +[48] Kajing Ma, Xianghao Zang, Zerun Feng, Han Fang, Chao Ban, Yuhan Wei, Zhongjiang He, Yongxiang Li, and Hao Sun. Llavilo: Boosting video moment retrieval via adapter-based multimodal modeling. In ICCV Workshops, pages 2790-2795. IEEE, 2023. 2 +[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 5 +[50] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In ICCV, pages 15201-15213, 2023. 7 +[51] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 2 +[52] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 6 +[53] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2 +[54] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, pages 10810-10819, 2020. 6 +[55] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, pages 407-426. Springer, 2022. 4 +[56] Jinwoo Nam, Daechul Ahn, Dongyeop Kang, Seong Jong Ha, and Jonghyun Choi. Zero-shot natural language video localization. In ICCV, pages 1470-1479, 2021. 5 +[57] Phuc Xuan Nguyen, Tianzhu Liu, Gaurav Prasad, Hung Hai Bui, Binh Pham, and Svetha Venkatesh. Weakly supervised action localization by sparse temporal pooling network. In CVPR, pages 6752-6761, 2019. 3 +[58] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 4 +[59] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, pages 10870-10879, 2020. 1 +[60] Sujoy Paul, Sourya Roy, and Amit K Roy-Chowdhury. W-talc: Weakly-supervised temporal activity localization and classification. In ECCV, pages 563-579, 2018. 3 + +[61] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In ICCV, pages 5285-5297, 2023. 2 +[62] Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Rephrase, augment, reason: Visual grounding of questions for vision-language models. In ICLR, 2024. 2 +[63] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 2, 5, 6 +[64] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. NeurIPS, 37:119336-119360, 2024. 7 +[65] Hangyu Qin, Junbin Xiao, and Angela Yao. Question-answering dense video events. arXiv preprint arXiv:2409.04388, 2024.7 +[66] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 5 +[67] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, pages 13009-13018, 2024. 2 +[68] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In CVPR, pages 6517-6525, 2017. 4 +[69] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 4 +[70] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 1:25-36, 2013. 4 +[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 2 +[72] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 1, 2, 5, 6, 7, 9 +[73] Seyed Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. CVPR, pages 658-666, 2019. 4 +[74] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, pages 2464-2473, 2020. 2, 7 +[75] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hongdong Li, and Stephen Gould. Dori: Discovering object relationships for moment localization of a natural language query in a video. In WACV, pages 1079–1088, 2021. 2, 7 + +[76] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hiroya Takamura, and Qi Wu. Memory-efficient temporal moment localization in long videos. In EACL, pages 1909–1924, 2023. 2, 6, 7 +[77] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 2 +[78] Himanshu Sharma, Manmohan Agrahari, Sujeet Kumar Singh, Mohd Firoj, and Ravi Kumar Mishra. Image captioning: a comprehensive survey. In 2020 International Conference on Power Electronics & IoT Applications in Renewable Energy and its Control (PARC), pages 325-328. IEEE, 2020. 1, 4 +[79] Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Silvia Cascianelli, Giuseppe Fiameni, and Rita Cucchiara. From show to tell: A survey on deep learning-based image captioning. IEEE TPAMI, 45(1):539-559, 2022. 1, 4 +[80] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In AAAI, pages 4998-5007, 2024. 1 +[81] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6 +[82] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. In CVPR, pages 13569-13580, 2024. 2, 7 +[83] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 4 +[84] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In ICCV, pages 1397-1407, 2021. 2 +[85] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 1, 2 +[86] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6 +[87] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024.5,6,7 +[88] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, pages 18699-18708, 2024. 2 +[89] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. In ECCV, pages 39-58. Springer, 2022. 2, 7 +[90] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, pages 13204-13214, 2024. 2, 4, 6, 7 + +[91] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In CVPR, pages 13623-13633, 2023. 6 +[92] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, pages 124-141, 2022. 2, 7 +[93] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2 +[94] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2024. 2, 5, 7 +[95] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In Proceedings of the 2nd international workshop on human-centric multimedia analysis, pages 13-21, 2021. 4 +[96] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 4 +[97] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 4 +[98] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020. 2, 7 +[99] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR, 2025. 2 +[100] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. In EMNLP, pages 21715–21737, 2024. 7 +[101] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, pages 492-510. Springer, 2022. 7 +[102] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, pages 6543-6554, 2020. 1, 2, 6 +[103] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 5, 6 +[104] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In CVPR, pages 8327–8336, 2019. 1 +[105] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment + +localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7 +[106] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7 +[107] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. IEEE TPAMI, 44(12):9073-9087, 2021. 6 +[108] Zhu Zhang, Zhijie Lin, Zhou Zhao, and Zhenxin Xiao. Cross-modal interaction networks for query-based moment retrieval in videos. In ACM SIGIR, pages 655-664, 2019. 7 +[109] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 3 +[110] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4 +[111] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2 +[112] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. CVPR, pages 8739-8748, 2018. 2 +[113] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2 +[114] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4 +[115] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, pages 18243-18252, 2024. 2 +[116] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. ArXiv, abs/2204.08121, 2022. 2 +[117] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 5651-5665, Gyeongju, Republic of Korea, 2022. International Committee on Computational Linguistics. 2 +[118] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 4 \ No newline at end of file diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/images.zip b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..00c84daf42bafd68a03f21f9b440670004efd98f --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62058265d65eadf39c264fcd28166a11349e2aac477c8655f0792b4c9503bfc4 +size 702506 diff --git a/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/layout.json b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b362843d951c39a084a8f0190597c76ba30a1df --- /dev/null +++ b/2025/Enrich and Detect_ Video Temporal Grounding with Multimodal LLMs/layout.json @@ -0,0 +1,10979 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 159, + 103, + 452, + 137 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 159, + 103, + 452, + 137 + ], + "spans": [ + { + "bbox": [ + 159, + 103, + 452, + 137 + ], + "type": "text", + "content": "Enrich and Detect: Video Temporal Grounding with Multimodal LLMs" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "spans": [ + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": "Shraman Pramanick" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{\\dagger,2}" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": " Effrosyni Mavroudi" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": " Yale Song" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": " Rama Chellappa" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": " Lorenzo Torresani" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{\\diamond 1}" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "text", + "content": " Triantafyllos Afouras" + }, + { + "bbox": [ + 102, + 140, + 507, + 169 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "spans": [ + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "text", + "content": "FAIR, Meta " + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "text", + "content": "Johns Hopkins University " + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 138, + 169, + 468, + 184 + ], + "type": "text", + "content": "Northeastern University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "spans": [ + { + "bbox": [ + 151, + 200, + 200, + 213 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 229, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 229, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 229, + 296, + 504 + ], + "type": "text", + "content": "We introduce ED-VTG, a method for fine-grained video temporal grounding utilizing multi-modal large language models. Our approach harnesses the capabilities of multimodal LLMs to jointly process text and video, in order to effectively localize natural language queries in videos through a two-stage process. Rather than being directly grounded, language queries are initially transformed into enriched sentences that incorporate missing details and cues to aid in grounding. In the second stage, these enriched queries are grounded, using a lightweight decoder, which specializes at predicting accurate boundaries conditioned on contextualized representations of the enriched queries. To mitigate noise and reduce the impact of hallucinations, our model is trained with a multiple-instance-learning objective that dynamically selects the optimal version of the query for each training sample. We demonstrate state-of-the-art results across various benchmarks in temporal video grounding and paragraph grounding settings. Experiments reveal that our method significantly outperforms all previously proposed LLM-based temporal grounding approaches and is either superior or comparable to specialized models, while maintaining a clear advantage against them in zero-shot evaluation scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 514, + 135, + 527 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 135, + 527 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 135, + 527 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 529, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 295, + 673 + ], + "type": "text", + "content": "Video temporal grounding [32, 40, 53, 80, 102] aims to identify temporal intervals in a video that correspond to a set of provided language queries. The task is essential for applications such as video editing and content retrieval. Conversely, video captioning [1, 59, 78, 79, 104] entails generating a natural language description for a given video segment, effectively translating visual content into text. The two tasks are in fact dual, as the outputs of one task are the inputs to the other, and vice versa. Intuitively, there is significant potential in exploiting this synergy, however it has largely remained unexplored, with previous works typically specializing in one of the two tasks [8, 32, 53, 104] or solv-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 200, + 553, + 339 + ], + "blocks": [ + { + "bbox": [ + 317, + 200, + 553, + 339 + ], + "lines": [ + { + "bbox": [ + 317, + 200, + 553, + 339 + ], + "spans": [ + { + "bbox": [ + 317, + 200, + 553, + 339 + ], + "type": "image", + "image_path": "f87cd08c5ac3e0806cb62d1dcb85321e4020c190ba7d7613f840bab5a47ea8e8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Our proposed system. ED-VTG performs video temporal grounding as a two-stage process: since user queries are often incomplete or coarse, the first stage involves producing an enriched query that adds additional details to the original, making it easier to ground. Meanwhile, a contextualized embedding is generated, containing all the information about the interval to be predicted. In the second stage, an interval decoder translates these embeddings into precise temporal boundaries." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 461 + ], + "type": "text", + "content": "ing them in multi-task setting [72, 85] without investigating how each task can benefit the other." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 462, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 462, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 312, + 462, + 555, + 677 + ], + "type": "text", + "content": "In this work, we exploit this duality by leveraging captioning to enhance grounding. Our key observation is that natural language queries often lack the completeness or detail necessary for effective temporal localization; indeed, existing grounding datasets frequently contain poorly worded, coarse, and potentially incomplete queries. The quality and completeness of these queries however is crucial for the precision of the temporal grounding. A natural hypothesis, therefore, is that more detailed queries, which could be obtained via conditional captioning, can significantly enhance grounding accuracy. For example, as illustrated in Figure 1, a vague query like 'Man starts surfing' can be refined into a more detailed description such as 'The man with a yellow surfboard slowly runs to start surfing,' resulting in grounding with more accurate temporal boundaries. In other cases, refining an abstract concept may involve breaking down a complex query into simpler components that are more directly groundable." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "This key idea of query enrichment forms the basis of our proposed approach. Concretely, we transform grounding into a two-stage reasoning process using a multi-modal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 683, + 208, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 683, + 208, + 693 + ], + "spans": [ + { + "bbox": [ + 69, + 683, + 208, + 693 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 69, + 683, + 208, + 693 + ], + "type": "text", + "content": " Work done during an internship at FAIR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 69, + 693, + 161, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 693, + 161, + 703 + ], + "spans": [ + { + "bbox": [ + 69, + 693, + 161, + 703 + ], + "type": "text", + "content": "Work done while at FAIR." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 69, + 703, + 246, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 246, + 712 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 246, + 712 + ], + "type": "text", + "content": "spraman3@meta.com, afourast@meta.com" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24297" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "LLM: first the model enriches the input query into a more detailed description by adding missing details based on the video content, and then temporally localizes the resulting enriched query in the video." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 122, + 294, + 242 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 122, + 294, + 242 + ], + "spans": [ + { + "bbox": [ + 54, + 122, + 294, + 242 + ], + "type": "text", + "content": "To effectively perform the temporal localization, we introduce a lightweight perception decoder that specializes in generating precise temporal boundaries, conditioned on a contextual representation, allowing the LLM to focus on language outputs, where it excels. The perception decoder allows us to leverage purposefully crafted training objectives for temporal grounding, building on prior knowledge and task-specific characteristics developed from years of research in object detection [17, 19, 30, 67, 71, 77] and temporal localization [32, 40, 53, 61, 102, 105]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 243, + 295, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 295, + 399 + ], + "type": "text", + "content": "Learning to jointly enrich and detect requires high-quality enriched query labels. We obtain those by using a strong external captioning model which we condition on the original queries and the video content of the target temporal boundaries. However powerful, these models are prone to hallucinations and there is no guarantee that the enriched queries will always be easier to ground than the original ones. At the same time, annotating the ground truth to determine which query - original or enriched - is a better candidate, is extremely expensive. To address this issue, we propose training in a multiple-instance learning (MIL) fashion, that enables the model to autonomously determine which query is better suited for the task during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 401, + 295, + 544 + ], + "type": "text", + "content": "Finally, we note that our proposed method is not equivalent to a data augmentation approach which simply preprocesses the training set to generate enhanced queries that are directly used for training. While this simpler alternative offers some of the same benefits, it suffers from the limitation that extracting enriched queries during training requires knowledge of the ground-truth temporal segments, i.e., the grounding targets. Since during inference these segments are unknown, the original queries must be used as input, which, as we will show experimentally, is suboptimal. Our method overcomes this limitation by learning to jointly enrich and detect, demonstrating superior performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "To summarize, our contributions are as follows: (i) we introduce a cascaded approach to temporal grounding, where the model first enriches the provided language query based on the video context and then proceeds to localize it; (ii) we enable multi-modal LLMs to accurately localize text queries using a lightweight decoder which allows for training with detection objectives tailored to the task; (iii) we propose a multiple-instance learning paradigm that enables the model to dynamically select the query that leads to better temporal localization; (iv) we achieve state-of-the-art results on several temporal grounding benchmarks, for both single query grounding and paragraph grounding, demonstrating, for the first time, an LLM-based model that surpasses or performs comparably to specialist models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 406, + 84 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 210 + ], + "type": "text", + "content": "LLM-based temporal grounding. Prior works have explored using LLMs for grounding natural language sentences in videos, either using raw text tokens to represent timestamps [21, 36, 39, 48, 72, 99] or by adding hundreds of special tokens to the LLM's vocabulary to represent video frames [22, 63, 85]. Our approach differs from these methods in that by utilizing a lightweight interval decoder we can apply detection losses such as L1 and gIoU with minimal added complexity; we additionally exploit the LLM's potential to describe video content in detail." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 211, + 553, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 414 + ], + "type": "text", + "content": "Specialist models. There is a rich variety of specialist models in the literature that are tailored to specific variants of temporal grounding, e.g. single query temporal grounding [102] and video paragraph grounding [82], and as such, achieve strong performance [6, 26, 41, 46, 75, 82, 98, 105, 106]. Modern methods typically employ a multi-modal transformer [102] that fuses dense video features with text embeddings of the language query, followed by a specialized detector head for performing detections [16, 32, 53, 74-76]. However, because these models are often trained on limited datasets for such a narrowly defined task, they struggle with generalization. Indeed, the zero-shot performance of these methods is limited [13, 89, 90, 92, 94]. In this work, we aim to address the shortcomings of previous methods by fully combining the generalization abilities of multi-modal LLMs with the advantages of specialist models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 569 + ], + "type": "text", + "content": "Dense captioning. Our method is related to dense video captioning [29], where the objective is to segment a given video into multiple parts and simultaneously provide descriptive captions for each segment. Traditional approaches tackle this task either by first determining the segments and then providing descriptions [6, 21, 25, 38, 111, 113], or jointly learning both tasks [10, 38, 93, 112, 116]. Recent advances have demonstrated video-conditioned LLMs to excel in this task [88, 115, 117]. While there are similarities between our enrich-and-detect paradigm and dense captioning, we solve a different task, namely video temporal grounding, where the input query is given and constrains the problem." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 570, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 689 + ], + "type": "text", + "content": "Prompt augmentation with LLMs. Off-the-shelf LLMs have recently been used to augment input prompts for tasks such as image retrieval and image classification by providing additional, clarifying descriptions. These augmented descriptions can aid generalization in various vision and NLP tasks such as visual question answering [12, 62, 84], dialog generation [15], and visual classification [51]. Our approach goes beyond using an off-the-shelf LLM for offline query enrichment, rather learning to dynamically enrich queries during inference time towards better grounding." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Prior work has also investigated the use of captioners and language models for data augmentation by rephras" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24298" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 370, + 266 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 370, + 266 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 370, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 370, + 266 + ], + "type": "image", + "image_path": "13e74b12896c30c0d81ec6251001ef8b5548091ae7ff3f03979c3fbe2181b8f2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed Enrich and Detect framework: (left) ED-VTG pipeline: Given an untrimmed video and a query " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": " to be grounded, the inputs are first tokenized into video tokens " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "T_V" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": " and text tokens " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "T_Q" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": ". The tokens are then fed into an LLM, which first generates an enriched query by, e.g., filling in any missing details and then emitting an interval token " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "<\\mathrm{INT}>" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": ". The embedding of this special token is finally decoded into the predicted temporal interval via a lightweight interval decoder. In the example shown here, the vague input query is enriched into a more detailed one by our model which can be subsequently grounded more easily. (right) Training: ED-VTG is trained using ground-truth temporal intervals and pseudo-labels of enriched queries, generated by an external off-the-shelf captioning model [33], which – unlike our model at inference time – has access to the ground-truth intervals. For every sample during training, the proposed multiple instance learning (MIL) framework allows ED-VTG to assess both the original or the enriched queries and generate two sets of predictions, " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\hat{I}^{enr}" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": ", the interval predicted using the enriched query, and " + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\hat{I}^{dir}" + }, + { + "bbox": [ + 54, + 270, + 555, + 392 + ], + "type": "text", + "content": " using the original query. Next, the model backpropagates using the better prediction (i.e., lower grounding loss). Hence, during training, ED-VTG dynamically learns to decide for which sample enrichment is necessary and, based on that, performs detection." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 378, + 71, + 556, + 266 + ], + "blocks": [ + { + "bbox": [ + 378, + 71, + 556, + 266 + ], + "lines": [ + { + "bbox": [ + 378, + 71, + 556, + 266 + ], + "spans": [ + { + "bbox": [ + 378, + 71, + 556, + 266 + ], + "type": "image", + "image_path": "4bf8cc78fe8deee8d7277a221578c7477b596e6b7051152da2e6d902c5fa8399.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 401, + 295, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 496 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 496 + ], + "type": "text", + "content": "ing existing descriptions or providing new improved ones. LaViLa[109] uses rephraser and narrator models to improve the quality of the training data for video-text alignment. Augmentation during training is sufficient for that task, as the goal is to learn joint representations; for our grounding task however simply augmenting the training set, without enriching queries during inference is not as effective, as we empirically demonstrate." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 497, + 295, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 497, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 54, + 497, + 295, + 581 + ], + "type": "text", + "content": "Multiple instance learning. Multiple instance learning (MIL)[11] is a technique commonly used in weakly-supervised vision problems, when a collection of potential solutions is available but exact annotations are not. It has been successfully applied to a range of tasks, including classification [3], weakly supervised object detection[7, 9] and temporal action localization[57, 60]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 588, + 111, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 111, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 111, + 601 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "Given an untrimmed video " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": " and a set of " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": " associated textual queries " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{Q} = \\{Q_1, \\ldots, Q_N\\}" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": ", temporal grounding aims to identify the corresponding temporal interval for each query. Formally, the output is a set of temporal intervals " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{I} = \\{I_1, \\ldots, I_N\\}" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": ". For " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "N = 1" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": ", the task takes the form of single-query temporal grounding, which we will simply refer to as STG. For clarity, we describe our approach in the STG setting, without loss of generality; the pipeline however readily extends to " + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "N > 1" + }, + { + "bbox": [ + 54, + 605, + 295, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "text", + "content": "Our approach aims to tackle temporal grounding by leveraging a multimodal LLM that " + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "inline_equation", + "content": "(i)" + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "text", + "content": " transforms input queries into intermediate, enriched queries by adding missing details based on the video input, " + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "inline_equation", + "content": "(ii)" + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "text", + "content": " generates contextualized embeddings of each latent segment to enable temporal interval prediction, " + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "inline_equation", + "content": "(iii)" + }, + { + "bbox": [ + 313, + 402, + 555, + 484 + ], + "type": "text", + "content": " decodes these embeddings into concrete temporal boundaries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 485, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 533 + ], + "type": "text", + "content": "In the following, we formally introduce our proposed ED-VTG model in Section 3.1, and proceed to describe how to train it with enriched query pseudo-labels within a MIL framework (Section 3.2)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 539, + 367, + 550 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 367, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 367, + 550 + ], + "type": "text", + "content": "3.1. Model" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 604 + ], + "type": "text", + "content": "Our model (Figure 2) consists of three key modules: a vision encoder that extracts video representations, a LLM that jointly processes video and language, and a lightweight interval decoder that generates precise temporal boundaries." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": "Enrich. Given a single input query " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " about a video " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " frames, the vision encoder represents the video as a sequence of " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " visual tokens " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{V}} \\in \\mathbb{R}^{R \\times D}" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " is the token dimension. The tokenized video features " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathrm{V}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " are fed along with the tokenized query " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{\\mathcal{Q}}" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " to the LLM, which generates an enriched query " + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "inline_equation", + "content": "\\hat{Q}^{enr} = \\{\\hat{y}_1, \\dots, \\hat{y}_l, \\dots, \\hat{y}_{L^{enr}}\\}" + }, + { + "bbox": [ + 313, + 604, + 554, + 687 + ], + "type": "text", + "content": " one token at a time:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 377, + 697, + 553, + 711 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 697, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 377, + 697, + 553, + 711 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {l} = \\mathcal {F} _ {\\mathrm {L L M}} \\left(\\hat {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right). \\tag {1}", + "image_path": "745130ad96ab01673fca02e66c9ed33b38d0002c6c3f0d090826fb51becb0b77.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24299" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": "When the model is ready to ground the query, the LLM emits a new, special token " + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "inline_equation", + "content": "<\\mathrm{INT}>" + }, + { + "bbox": [ + 55, + 72, + 295, + 109 + ], + "type": "text", + "content": " to trigger interval prediction. In other words, the text prediction takes the form:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 90, + 114, + 295, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 114, + 295, + 127 + ], + "spans": [ + { + "bbox": [ + 90, + 114, + 295, + 127 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {y}} = \\text {\" T h e q u e y} \\hat {Q} ^ {e n r} \\text {o c c u r s a t} < I N T > \\tag {2}", + "image_path": "5fea0af847806609fc90a9a6299845afd65c566f77f841f79a3bd07e0857c381.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": "Detect. To detect the temporal interval corresponding to the enriched query " + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "inline_equation", + "content": "\\hat{Q}^{enr}" + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": ", we introduce an interval decoder " + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{dec}}" + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": " that predicts an interval " + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": " parametrized by the center " + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "inline_equation", + "content": "\\hat{c}" + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "inline_equation", + "content": "\\hat{w}" + }, + { + "bbox": [ + 55, + 140, + 296, + 224 + ], + "type": "text", + "content": " of the predicted interval. We selected this parameterization for its advantage in decoupling position from scale, as supported by literature in object detection [42, 68, 69, 114]. The interval decoder takes the form" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 234, + 295, + 248 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 234, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 104, + 234, + 295, + 248 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {I}} = (\\hat {c}, \\hat {w}) = \\mathcal {F} _ {\\mathrm {d e c}} \\left(\\mathcal {G} \\left(\\mathbf {h} _ {\\text {i n t}}\\right), \\mathbf {T} _ {\\mathrm {V}}\\right), \\tag {3}", + "image_path": "62a8a5d528a4a53c5b45daab14d0d6054e063329e6c089e77a739b0692b7abb9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{int} \\in \\mathbb{R}^D" + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": " is the hidden state of the LLM corresponding to the token and " + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": " is a linear projection layer. The decoder, functioning as the regression component of a temporal detector, consists of two transformer layers followed by a multi-layer perceptron (MLP), which processes a concatenation of its two inputs, " + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "inline_equation", + "content": "(\\mathcal{G}(\\mathbf{h}_{\\mathrm{int}}), \\mathbf{T}_{\\mathrm{V}})" + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": ". It finally outputs the predicted interval " + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": ", grounding the input query " + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 55, + 258, + 296, + 356 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": "Notice that our formulation allows the quality of the enriched query " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\hat{Q}^{enr}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " to directly influence the accuracy of the predicted interval " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": ". This is because the generation of the interval token " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "<\\mathrm{INT}>" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " (and consequently its hidden state " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": ") is conditioned on the enriched query prediction. Establishing this cascaded dependency chain, i.e., " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "(V, Q) \\to \\hat{Q}^{enr}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "inline_equation", + "content": "(V, \\hat{Q}^{enr}) \\to \\hat{I}" + }, + { + "bbox": [ + 55, + 357, + 296, + 441 + ], + "type": "text", + "content": ", is the key idea of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 119, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 119, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 119, + 462 + ], + "type": "text", + "content": "3.2. Training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": "The model is trained end-to-end using two primary loss functions: a language modeling loss " + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": " and a temporal grounding loss " + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{grnd}}" + }, + { + "bbox": [ + 55, + 468, + 296, + 517 + ], + "type": "text", + "content": " that help supervise the \"enrich\" and \"detect\" aspects of our model, respectively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "text", + "content": "Enrich. Given the target output text " + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{y} = \\{y_1, y_2, \\dots, y_L\\}" + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}}" + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "text", + "content": " is calculated as the cross-entropy loss that evaluates the likelihood of " + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "inline_equation", + "content": "\\mathbf{y}" + }, + { + "bbox": [ + 55, + 526, + 296, + 574 + ], + "type": "text", + "content": " under the predicted probability distribution generated by the model:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 94, + 575, + 295, + 608 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 575, + 295, + 608 + ], + "spans": [ + { + "bbox": [ + 94, + 575, + 295, + 608 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {L M}} = - \\sum_ {t = 1} ^ {T} \\log P \\left(y _ {l} \\mid \\mathbf {y} _ {< l}, \\mathbf {T} _ {\\mathrm {V}}, \\mathbf {T} _ {\\mathcal {Q}}\\right) \\tag {4}", + "image_path": "4dced866d299268f7c8ef59d691eb5a3a6993f438fba57342beeb7252744273f.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "text", + "content": "To provide a proper supervisory signal for query enrichment, we need the ground-truth pair of " + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "inline_equation", + "content": "(Q, Q^{enr})" + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "text", + "content": ". However, such datasets do not exist, nor is it practical to annotate a dataset solely for this purpose. Here, we capitalize on the tremendous amount of progress made in the video captioning literature [1, 78, 79], and use an off-the-shelf captioning model [33] to generate pseudo ground-truth " + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "inline_equation", + "content": "Q^{enr}" + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "text", + "content": " by refining the original query " + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 615, + 296, + 723 + ], + "type": "text", + "content": " given its video V (see more details in Section 4.1)." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 314, + 71, + 554, + 191 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 554, + 191 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 554, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 554, + 191 + ], + "type": "table", + "html": "
DatasetDomainTasksCorpusEval. Protocol# Train SamplesAvg Vid Length (s)Avg Span Length (s)
DiDeMo [4]OpenSTGPT-32.8K54.576.49 (11.9%)
QuerYD [58]CookingSTGPT-13.6K158.787.68 (4.8%)
COIN [83]OpenVPGPT-7.5K143.7115.06 (10.5%)
HiREST [96]OpenSTG, VPGPT-0.8K208.0344.50 (21.4%)
ViTT† [23]OpenVPGPT-4.9K287.17-
YTTemporal [97]OpenVPGPT-28.8K327.364.0 (1.2%)
CrossTask [118]ProceduralAGPT-2.7K297.09.61 (3.2%)
VideoCC [55]OpenSTGPT-45.0K415.899.88 (2.3%)
Charades-STA [14]IndoorSTGFT, EvalZS, FT12.4K31.178.29 (26.6%)
Charades-CD-OOD [95]IndoorVPGFT, EvalFT4.5K30.607.90 (25.8%)
ANet-Captions [29]OpenSTG, VPGFT, EvalZS, FT9.5K117.6335.61 (30.3%)
TACoS [70]CookingSTG, VPGFT, EvalZS, FT9.8K224.3423.33 (10.4%)
YouCook2 [110]CookingVPGFT, EvalFT1.2K311.4120.07 (6.4%)
NEX-T-GQA° [90]OpenQGEvalZS-39.606.69 (16.9%)
HT-Step [2]CookingAGFT, EvalFT17.4K393.8914.88 (3.7%)
", + "image_path": "446eddf0a68baa2db656a78bfb5ad2f8454fd85df7b059fdcb12adbb3973d816.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "lines": [ + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "type": "text", + "content": "Table 1. Dataset statistics, corresponding tasks, and evaluation protocol. The upper side of the table represents datasets used for pre-training, resulting in a total of 136K samples. The lower side represents datasets used for fine-tuning and evaluation. We cover four different video grounding tasks: single-query temporal grounding (STG), video paragraph grounding (VPG), question grounding (QG), and article grounding (AG). QG is used only for evaluation to assess the model's generalization capability. Average interval lengths compared to the corresponding video durations are shown in brown, denoting the annotation granularity. " + }, + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "type": "text", + "content": "VITT contains single timestamp annotation instead of intervals. " + }, + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "type": "inline_equation", + "content": "^\\circ" + }, + { + "bbox": [ + 313, + 196, + 555, + 328 + ], + "type": "text", + "content": "NExTGQA contains only evaluation split." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "text", + "content": "Detect. Given a target temporal interval " + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "inline_equation", + "content": "I = (c, w)" + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "text", + "content": ", the grounding loss " + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{ground}}" + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "text", + "content": " is computed as a combination of the L1 loss and the generalized Intersection over Union (gIoU) loss [73], applied on the predicted temporal interval " + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "inline_equation", + "content": "(\\hat{c}, \\hat{w})" + }, + { + "bbox": [ + 313, + 336, + 554, + 385 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 361, + 392, + 553, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 392, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 361, + 392, + 553, + 421 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {g r n d}} = \\lambda_ {\\mathrm {L 1}} \\left(\\left| (\\hat {c} - c | + | \\hat {w} - w |\\right) \\right. \\\\ + \\lambda_ {\\mathrm {g I o U}} \\mathrm {g I o U} ((\\hat {c}, \\hat {w}), (c, w)) \\tag {5} \\\\ \\end{array}", + "image_path": "5cfde3e703e7c8db4f44231e6e469205c1733cbec10af2471d79df2cc690efc9.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 423, + 412, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 423, + 412, + 434 + ], + "spans": [ + { + "bbox": [ + 314, + 423, + 412, + 434 + ], + "type": "text", + "content": "3.2.1. MIL framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": "A caveat with the pseudo-labeled enriched queries is that they can be noisy and include hallucinations; as a result, some of them may lead to a temporal interval prediction inferior to the original query. To address this, we propose considering multiple options for the target query (2 in the case of a single query scenario, " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "Q^{enr}" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": "), by adopting a multiple instance learning (MIL) framework [11]. This approach allows the model to choose between the enriched and the original input queries during training, depending on which version leads to a better temporal interval prediction. During inference, this means that our ED-VTG model can choose to either enrich the query if it misses important details, or decide to \"carry over\" the original query into " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "\\hat{Q}^{enr}" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": " when it is concrete enough (in which case " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "Q = \\hat{Q}^{enr}" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": "Formally, to perform MIL, we collect two temporal interval predictions by running two forward passes with different LLM inputs (in a teacher-forcing fashion), i.e. " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " that is formed using the original query " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " using the pseudo-labeled enriched query " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{Q}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ". We illustrate this schematically in Figure 2 (right). Recall that through the contextualized interval representation " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{\\mathrm{int}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": ", the predicted interval depends on the LLM teacher-forced input. Hence, these inputs produce respective interval predictions, " + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 313, + 606, + 556, + 715 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24300" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 549, + 247 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 549, + 247 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 549, + 247 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 549, + 247 + ], + "type": "table", + "html": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
UniVTG [40]X4.2MZS44.125.210.027.1----5.21.30.34.4
SeViLA [94]X129MZS----31.619.010.123.0----
PSVL [56]X-ZS46.231.314.231.244.730.114.729.6----
LT-ZVG [28]X-ZS52.937.219.336.047.632.615.431.8----
Video-LLaMA° [103]✓2.7MZS25.210.63.416.821.910.84.916.55.11.20.83.4
Video-ChatGPT° [49]✓100KZS27.26.21.919.719.510.64.814.26.31.71.04.3
Valley [47]✓100KZS28.41.80.321.430.613.78.121.9----
VideoChat2 [36]✓2MZS38.014.33.824.640.827.89.327.9----
Momenter [63]✓10MZS42.626.611.628.542.923.012.429.3----
VTimeLLM° [21]✓170KZS51.027.511.431.244.027.814.330.47.01.80.84.5
TimeChat° [72]✓125KZS-32.213.4-----6.82.10.84.7
HawkEye [87]✓715KZS50.631.414.533.749.129.310.732.7----
ChatVTG [66]✓100KZS52.733.015.934.940.722.59.427.28.13.71.35.5
ED-VTG✓136KZS59.539.319.840.252.133.116.035.214.56.02.312.7
ΔOurs - HawkEye--ZS8.9↑7.9↑5.3↑6.5↑3.0↑3.8↑5.3↑2.5↑----
ΔOurs - ChatVTG--ZS--------6.4↑2.3↑1.0↑7.2↑
", + "image_path": "92f6118e80ce571788b3d2f5edb71a6bce4b364ccfac399510dd523f3d41afec.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": "Table 2. Zero-shot STG results on Charades, ActivityNet, and TACoS test splits. For all three datasets, ED-VTG gains significant improvement over all existing methods, including task-specific, non-generalist models. We use boldface for the best and underline the second-best result for each metric, among the generalist models. " + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "inline_equation", + "content": "^\\circ" + }, + { + "bbox": [ + 55, + 251, + 555, + 285 + ], + "type": "text", + "content": "Official checkpoints are used for TACoS evaluation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 55, + 291, + 295, + 327 + ], + "type": "text", + "content": ". We select the query version that results in the smallest grounding loss and use it to compute our overall training objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 64, + 328, + 295, + 360 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 328, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 64, + 328, + 295, + 360 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\left\\{ \\begin{array}{l l} \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {d i r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} & \\text {i f} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {d i r}} < \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} \\\\ \\lambda_ {\\mathrm {L M}} \\mathcal {L} _ {\\mathrm {L M}} ^ {\\mathrm {e n r}} + \\lambda_ {\\mathrm {g r n d}} \\mathcal {L} _ {\\mathrm {g r n d}} ^ {\\mathrm {e n r}} & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {6}", + "image_path": "788ad87e798d33150eaebeeb4e98deda38b890f9afa9ca7f634bf531ca7f0837.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{LM}}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " are the language losses for targets " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathbf{y}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " respectively, and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{grnd}}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " the grounding losses for predictions " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{dir}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{I}}^{\\mathrm{enr}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " respectively. The hyperparameters " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{LM}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{grnd}}" + }, + { + "bbox": [ + 55, + 365, + 296, + 430 + ], + "type": "text", + "content": " are the relative weights of the language modeling and grounding losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 434, + 138, + 446 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 138, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 138, + 446 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "content": "We design experiments to study three key questions related to our architecture and training framework: Q1) How does ED-VTG perform on various video grounding tasks in comparison to the current state-of-the-art? Q2) How beneficial is our query enrichment approach, within the MIL paradigm, as opposed to directly grounding the original queries? Q3) Does utilizing an interval decoder with specifically tailored grounding objectives offer advantages over predicting timestamps as raw text or special tokens?" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 559, + 118, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 559, + 118, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 559, + 118, + 571 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 295, + 720 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 295, + 720 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 295, + 720 + ], + "type": "text", + "content": "Table 1 summarizes the datasets that we used during the pre-training and fine-tuning stages. During pre-training, we use a total of 136K medium-to-long duration videos from 8 public datasets annotated with text queries and the corresponding intervals. As a preprocessing step, we collect pseudo-labels for enriched queries using an external captioning model [33]. In short, we take each video segment defined by the ground-truth intervals and prompt the captioning model to enrich the original query while preserving its meaning given the video segment. We provide full details of this process with the exact prompt used in the supplementary material." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 364, + 302 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 364, + 302 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 364, + 302 + ], + "type": "text", + "content": "4.2. Tasks" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 305, + 555, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 555, + 341 + ], + "type": "text", + "content": "Table 1 also summarizes the tasks for which we use each dataset. Here we briefly describe these tasks, and highlight how our approach is applied to solve them." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 341, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 555, + 376 + ], + "type": "text", + "content": "Single-Query Temporal Grounding (STG) involves identifying a single time window in response to a single input language query " + }, + { + "bbox": [ + 313, + 341, + 555, + 376 + ], + "type": "inline_equation", + "content": "(N = 1)" + }, + { + "bbox": [ + 313, + 341, + 555, + 376 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": "Video Paragraph Grounding (VPG) involves grounding " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "N > 1" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": " sentences to " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": " corresponding time windows. Our ED-VTG model can be trained to predict multiple enriched queries (one per original input query) interleaved with " + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "inline_equation", + "content": "<\\mathrm{INT}>" + }, + { + "bbox": [ + 313, + 377, + 554, + 484 + ], + "type": "text", + "content": " tokens that get separately decoded into intervals. Since there are multiple queries in the input, we run multiple forward passes through the LLM to perform MIL, wherein each pass selects a random number of queries to be enriched." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 485, + 554, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 532 + ], + "type": "text", + "content": "Question Grounding (QG) involves retrieving evidence intervals to answer questions, facilitating explainable video QA. As in STG, the input is a single query, and the output a single interval." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 554, + 592 + ], + "type": "text", + "content": "Article grounding (AG) is an extension of VPG where the model is given multiple queries as input, some of which may not be groundable in the video. Hence, the model must 1) identify which queries are groundable and 2) predict intervals for the groundable queries." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "content": "Note that, while we train our model jointly on multiple tasks, they are unified as a single task in the form of (Video, Text) " + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 593, + 554, + 653 + ], + "type": "text", + "content": " (Time intervals). Besides the differences in training data sources, we use the identical training objective (Eq. 6) during pre-training and fine-tuning stages." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 659, + 430, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 430, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 430, + 671 + ], + "type": "text", + "content": "4.3. Evaluation Protocol" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Following the common practices in the literature [21, 72, 87], we evaluate ED-VTG in two primary evaluation protocols: (i) Zero-shot (ZS), where the pre-trained model" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24301" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 70, + 549, + 253 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 549, + 253 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 549, + 253 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 549, + 253 + ], + "type": "table", + "html": "
MethodGeneralist Model# Train SamplesEval.Charades-STAActivityNet-CaptionsTACoS
R@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoUR@0.3R@0.5R@0.7mIoU
VSLNet (C3D) [102]X-FT64.347.330.245.263.243.226.243.229.624.320.024.1
MS-2D-TAN (I3D) [107]X-FT-56.636.2-62.145.528.3-42.033.622.1-
Moment-DETR [32]X236KFT65.852.130.645.5----38.024.712.025.5
UnLoc-B [91]X650KFT-58.135.4--48.029.7-----
MomentDiff [37]X-FT-55.632.4-----46.628.912.430.4
LGI [54]X-FT73.059.535.551.458.541.523.141.1----
BAM-DETR [31]X-FT72.960.039.452.3----56.741.526.839.3
InternVideo2* + CG-DETR [86]X2.1MFT79.770.048.958.8--------
SG-DETR [18]X-FT-71.152.860.7-----46.433.942.4
EMB (ELA) [24]X-FT79.769.251.462.273.758.740.756.263.352.537.048.4
BLIP-2 (frames only) [34]✓129MFT-43.332.6--25.89.7-----
VideoChat2 [36]✓2MFT----55.534.717.738.9----
TimeChat [72]✓125KFT-46.723.7-----27.715.16.418.4
HawkEye [87]✓715KFT72.558.328.849.355.934.717.939.1----
VtimeLLM [21]✓170KFT--------26.814.46.118.0
ED-VTG✓136KFT78.262.135.052.667.645.122.744.946.031.515.832.4
ΔOurs - HawkEye--FT5.7↑3.8↑6.2↑3.3↑11.7↑10.4↑4.8↑5.8↑----
ΔOurs - VTimeLLM--FT--------19.2↑17.1↑9.7↑14.4↑
", + "image_path": "11fc7bf06138baa15a9a368f0a07cf9b0cf83077539f83b22bd78289da581d16.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 553, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 255, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 553, + 309 + ], + "type": "text", + "content": "Table 3. Fine-tuned STG results on Charades, ActivityNet, and TACoS test splits. For all datasets, ED-VTG achieves strong improvements over previous generalist models, and performs comparably to task-specific expert models. *Though InterVideo2 is a generalist model, it fine-tunes CG-DETR [52] head for grounding tasks, using the LLM as a feature extractor. For completeness, we report all competitive existing works, including task-specific SOTA specialist models (shown in gray), but directly compare ED-VTG only with generalist frameworks - using boldface for the best and underlining the second best results among generalist models (see discussion in Section 4.3)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "text", + "content": "is assessed directly without any fine-tuning on downstream datasets, and (ii) Fine-tuned (FT), where the model undergoes additional training on specific tasks and datasets. We also present results without pre-training to assess its importance. For the STG and VPG tasks, following existing works [6, 21, 72, 76, 87], we report the mean intersection over union (mIoU) and Recall@1 for IoU " + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "inline_equation", + "content": "\\geq m" + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "inline_equation", + "content": "R@m" + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "text", + "content": "), with " + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "inline_equation", + "content": "m \\in \\{0.3, 0.5, 0.7\\}" + }, + { + "bbox": [ + 54, + 316, + 294, + 458 + ], + "type": "text", + "content": ". For the QG task, we report intersection over prediction (IoP) in addition to IoU, following the NeXT-GQA [90] evaluation protocol. For AG, we adhere to the HT-Step [2] protocol and report article-grounding mAP scores across various IoU thresholds." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 460, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 568 + ], + "type": "text", + "content": "Comparison to non-generalist models. In the results section, we report all existing SOTA and competitive methods for a complete comparison, however we divide them into generalist and non-generalist (specialist) models. We note that specialist methods are heavily tailored to the task and in practice often overfit specific datasets which limits their transferability (as is evidenced by the zero-shot results in Table 2). We therefore focus the discussion around our method's comparison to other generalist models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 575, + 188, + 588 + ], + "type": "text", + "content": "4.4. Implementation Details" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 713 + ], + "type": "text", + "content": "We initialize the video encoder and LLM with the VideoLLaMA-7B [103] checkpoint, which is a similarly sized backbone used in existing LLM-based video grounding models [21, 63, 72, 87]. Video-LLaMA is trained on video captioning tasks with WebVid [5] and VideoChat [35]. The video encoder includes a ViT-G/14 from EVA-CLIP [81] as the frame feature extractor, followed by image and video QFormers. We initialize the decoder with random weights. We keep the ViT frozen, apply LoRA [20] with rank 32 to the LLM and fully tune the Q-Formers, decoder, and linear" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 316, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 434 + ], + "type": "text", + "content": "layers. We use LLaVA OneVision (OV) 72B [33] as the external captioner for obtaining enriched queries in the training sets. For the VPG task, we run four forward passes to perform MIL. We pre-train our model for 40 epochs with a batch size 256, using AdamW [45] with a peak learning rate of 5e-5 and a cosine scheduler [44] with a linear warmup for the first " + }, + { + "bbox": [ + 313, + 316, + 553, + 434 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 316, + 553, + 434 + ], + "type": "text", + "content": " steps. Pre-training takes 2 days on 16 V100 nodes (8 cards with 32G GPU memory each). Additional details on pre-training, fine-tuning, and task-specific instructions are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 436, + 372, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 436, + 372, + 447 + ], + "spans": [ + { + "bbox": [ + 314, + 436, + 372, + 447 + ], + "type": "text", + "content": "4.5. Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 450, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 700 + ], + "type": "text", + "content": "Single-Query Temporal Grounding (STG). We start by comparing ED-VTG against the state-of-the-art across three different STG benchmarks, namely Charades-STA, ActivityNet-Captions, and TACoS, in a zero-shot evaluation setting. We show the results in Table 2. On Charades, ED-VTG achieves ZS scores of 59.5, 39.3, and 19.8 for R@0.3, R@0.5, and R@0.7, respectively, significantly outperforming all baseline models. Notably, ED-VTG surpasses Momenter [63] and HawkEye [87] by 11.4 and 6.2 absolute mIoU points despite these models being pretrained with 100x and 6x more segment-level data, respectively. A similar trend is observed on ActivityNet and TACoS where our model achieves improvements of 2.5 and 7.2 mIoU points over the nearest LLM-based models in zero-shot setting. Notably, the TACoS dataset contains short, under-specified queries, longer input videos, and fine-grained interval annotations, which pose challenges for LLMs with a fixed number of input frames. The enriched query descriptions that ED-VTG generates enable it to more accurately align them to video frames and precisely retrieve the correct intervals." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 553, + 712 + ], + "type": "text", + "content": "On the fine-tuned STG evaluation, as shown in Table" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24302" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 177, + 149 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 177, + 149 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 177, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 177, + 149 + ], + "type": "table", + "html": "
MethodGeneralist ModelCharades-CD-OOD
R@0.3R@0.5mIoU
DepNet [6]✗45.627.629.3
DRN [98]✗40.530.4-
STLG [46]✗48.330.4-
SVPTR [26]✗50.328.532.1
SiamGTR [82]✗59.135.538.9
VTimeLLM1[21]✓53.234.035.1
TimeChat1[72]✓60.536.138.3
ED-VTG70.747.345.0
ΔOurs - TimeChat-10.2↑11.2↑6.7↑
", + "image_path": "b4bb572f5e2d0d1dfa044c9dcd514a4ca9ee6d4a098d89e98c596daa18bc7c3e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 181, + 70, + 302, + 149 + ], + "blocks": [ + { + "bbox": [ + 58, + 152, + 175, + 160 + ], + "lines": [ + { + "bbox": [ + 58, + 152, + 175, + 160 + ], + "spans": [ + { + "bbox": [ + 58, + 152, + 175, + 160 + ], + "type": "text", + "content": "(a) Results on Charades-CD-OOD." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 181, + 70, + 302, + 149 + ], + "lines": [ + { + "bbox": [ + 181, + 70, + 302, + 149 + ], + "spans": [ + { + "bbox": [ + 181, + 70, + 302, + 149 + ], + "type": "table", + "html": "
MethodGeneralist ModelANet-Captions
R@0.3R@0.5mIoU
CBLN [41]X66.348.127.6
2D-TAN [105]X59.544.5-
3D-TPN [106]X67.651.5-
DepNet [6]X72.855.9-
SVPTR [26]X78.161.755.9
VTimeLLM1[21]✓66.150.345.6
TimeChat1[72]✓67.951.547.0
ED-VTG✓74.158.053.7
ΔOurs - TimeChat-6.2↑6.5↑6.7↑
", + "image_path": "06d8348d6087d4fa696190975f3a1b5baf9d121addde293b4ab435c60e71d363.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 306, + 70, + 427, + 149 + ], + "blocks": [ + { + "bbox": [ + 197, + 152, + 286, + 161 + ], + "lines": [ + { + "bbox": [ + 197, + 152, + 286, + 161 + ], + "spans": [ + { + "bbox": [ + 197, + 152, + 286, + 161 + ], + "type": "text", + "content": "(b) Results on ActivityNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 306, + 70, + 427, + 149 + ], + "lines": [ + { + "bbox": [ + 306, + 70, + 427, + 149 + ], + "spans": [ + { + "bbox": [ + 306, + 70, + 427, + 149 + ], + "type": "table", + "html": "
MethodGeneralist ModelTACoS
R@0.3R@0.5mIoU
CMIN [108]✗24.618.1-
2D-TAN [105]✗37.325.3-
3D-TPN [106]✗40.326.5-
DepNet [6]✗41.327.2-
SVPTR [26]✗47.928.231.4
VTimeLLM†[21]✓40.225.627.9
TimeChat†[72]✓39.525.627.8
ED-VTG✓46.227.830.7
ΔOurs - TimeChat-6.7↑2.2↑2.9↑
", + "image_path": "c7980896280dc00b58b6fc979c7af2c5aa27dbd31bfd0c1ec4d97d94a9d1b99a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 430, + 70, + 552, + 149 + ], + "blocks": [ + { + "bbox": [ + 329, + 152, + 402, + 160 + ], + "lines": [ + { + "bbox": [ + 329, + 152, + 402, + 160 + ], + "spans": [ + { + "bbox": [ + 329, + 152, + 402, + 160 + ], + "type": "text", + "content": "(c) Results on TACoS." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 430, + 70, + 552, + 149 + ], + "lines": [ + { + "bbox": [ + 430, + 70, + 552, + 149 + ], + "spans": [ + { + "bbox": [ + 430, + 70, + 552, + 149 + ], + "type": "table", + "html": "
MethodGeneralist ModelYouCook2
R@0.3R@0.5mIoU
DORi [75]✗43.430.530.5
DORi* [75]✗42.329.929.9
LocFormer [76]✗46.831.330.9
ExCL [16]✗26.616.218.9
TMLGA [74]✗34.823.124.4
VTimeLLM1[21]✓41.318.524.3
TimeChat1[72]✓40.919.026.6
ED-VTG✓48.128.031.5
ΔOurs - TimeChat-7.2↑9.0↑4.9↑
", + "image_path": "98d83c44578ce41e167dbce255cb90e2d17300cf9682c5bd9e01371d3c6e327c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 56, + 200, + 296, + 321 + ], + "blocks": [ + { + "bbox": [ + 449, + 152, + 533, + 160 + ], + "lines": [ + { + "bbox": [ + 449, + 152, + 533, + 160 + ], + "spans": [ + { + "bbox": [ + 449, + 152, + 533, + 160 + ], + "type": "text", + "content": "(d) Results on YouCook2." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "type": "text", + "content": "Table 4. Performance on VPG task on four different benchmarks: Charades-CD-OOD, ActivityNet-Captions, TACoS, YouCook2. ED-VTG significantly improves over previous LLM-based models, and performs comparably to state-of-the-art specialist models. " + }, + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "type": "inline_equation", + "content": "{}^{ \\dagger }" + }, + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "type": "text", + "content": " We fine-tune VTimeLLM and TimeChat checkpoints for the VPG task. Dori " + }, + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "type": "inline_equation", + "content": "{}^{ \\star }" + }, + { + "bbox": [ + 55, + 162, + 555, + 196 + ], + "type": "text", + "content": " represents frozen text (BERT) encoder during fine-tuning." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 200, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 56, + 200, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 56, + 200, + 296, + 321 + ], + "type": "table", + "html": "
MethodGeneralist ModelNExT-GQA
mIoPIoP@0.3IoP@0.5mIoUIoU@0.3IoU@0.5
VGT [89]✘24.726.024.63.04.21.4
VIOLETv2 [13]✘23.625.123.33.14.31.3
Temp[CLIP] NG+ [90]✘25.731.425.512.117.58.9
FrozenBiLM NG+ [92]✘24.228.523.79.613.56.1
SeViLA [94]✘29.534.722.921.729.213.8
\\( LLoVi7B^{\\circ} \\) [100]✓20.7-20.58.7-6.0
VideoStreaming° [64]✓32.2-31.019.3-13.3
LongRepo 7B° [27]✓20.3-20.08.7-6.0
DeVi [65]✓33.8-32.220.717.4-
HawkEye [87]✓---25.737.019.5
ED-VTG✓34.745.133.526.639.519.8
\\( \\Delta_{Ours - SeViLA} \\)-5.2↑10.4↑10.6↑4.9↑10.3↑6.0↑
\\( \\Delta_{Ours - HawkEye} \\)----0.9↑2.5↑0.3↑
", + "image_path": "0edae7f604909a333aadb9a48d53480f864f1ec98cd4abb107fbecf4536f8f2f.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 56, + 371, + 296, + 452 + ], + "blocks": [ + { + "bbox": [ + 55, + 325, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 325, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 296, + 369 + ], + "type": "text", + "content": "Table 5. Performance on QG task on NeXT-GQA test split. ED-VTG consistently achieves consistent improvements over the existing models across all metrics. Results of LLoVi, LongRepo and VideoStreaming are from [64]." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 371, + 296, + 452 + ], + "lines": [ + { + "bbox": [ + 56, + 371, + 296, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 296, + 452 + ], + "type": "table", + "html": "
MethodSeenUnseen
↑mAP@IoU↑mAP@IoU
@0.3@0.5@0.7@[0.3-0.7]@0.3@0.5@0.7@[0.3-0.7]
UMT [43]15.78.73.29.19.44.91.75.3
MT+BCE [2, 50]46.229.912.929.831.618.77.719.3
ActionFormer-T [101]41.230.818.330.229.720.310.720.4
Timechat† [72]45.329.014.429.030.717.87.518.7
ED-VTG48.931.518.032.533.021.211.121.6
ΔOurs - TimeChat3.6 ↑2.5 ↑3.6 ↑3.5 ↑2.3 ↑3.4 ↑3.6 ↑2.9 ↑
", + "image_path": "357e4b7130a1da6f2b622e23c924dc1b7de6ccd26ab1ffd2ab0b4829bc2dfe7a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 455, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 296, + 521 + ], + "type": "text", + "content": "Table 6. Performance on AG task on HT-Step seen and unseen val split. ED-VTG is the first LLM-based model to report results on for video grounding in the presence of negative, non-groundable queries and sets a new state of the art in terms of average mAP score across various IoU thresholds. We fine-tune the official TimeChat checkpoint for AG task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 526, + 295, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 634 + ], + "type": "text", + "content": "3. ED-VTG demonstrates an impressive gain of 5.7 and 11.7 points in R@0.3 over HawkEye on Charades and ActivityNet, respectively. Similarly on TACoS, ED-VTG surpasses all existing MLLMs by a considerable margin, e.g. 14.0 and " + }, + { + "bbox": [ + 55, + 526, + 295, + 634 + ], + "type": "inline_equation", + "content": "14.4\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 526, + 295, + 634 + ], + "type": "text", + "content": " points over TimeChat and VTimeLLM. Furthermore, ED-VTG also beats many existing task-specific specialist models in all three benchmarks, significantly reducing the gap between specialist models and MLLMs for STG." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 713 + ], + "type": "text", + "content": "Video Paragraph Grounding (VPG). Next, we fine-tune ED-VTG for the VPG task, where the model processes multiple input queries in a temporal sequence. While specialist models have reported results in the past, no LLM-based models have previously addressed this challenging task. To establish a baseline for comparison, we fine-tune the" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 202, + 553, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 202, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 202, + 553, + 285 + ], + "type": "text", + "content": "officially released VTimeLLM and TimeChat pre-trained checkpoints. As shown in Table 4a, ED-VTG significantly outperforms both LLMs on the Charades-CD-OOD dataset, achieving an absolute gain of 9.9 and " + }, + { + "bbox": [ + 313, + 202, + 553, + 285 + ], + "type": "inline_equation", + "content": "6.7\\mathrm{mIoU}" + }, + { + "bbox": [ + 313, + 202, + 553, + 285 + ], + "type": "text", + "content": " points over VTimeLLM and TimeChat, respectively. Additionally, ED-VTG surpasses all specialist models on this dataset by a substantial margin, setting a new state-of-the-art." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 287, + 554, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 407 + ], + "type": "text", + "content": "We also evaluate VPG performance on three other benchmarks: ActivityNet-Captions, TACoS, and YouCook2, as presented in Tables 4b, 4c, and 4d. Consistent with the results on Charades, ED-VTG outperforms LLM-based models across all metrics on these datasets, with mIoU gains of 6.7, 2.9, and 4.9 over TimeChat on ActivityNet-Captions, TACoS, and YouCook2, respectively. ED-VTG also exceeds many existing specialist VPG models, demonstrating the effectiveness of our enriched queries and cascaded interval decoder." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 554, + 519 + ], + "type": "text", + "content": "Question Grounding (QG). To further assess the model's generalization capabilities, we conduct zero-shot evaluation on the held-out QG task. Our results on the NeXT-GQA test set are presented in Table 5. Notably, ED-VTG achieves state-of-the-art performance across all metrics, outperforming both specialist and LLM-based models by a significant margin. Specifically, ED-VTG surpasses the existing best baseline, HawkEye, by 2.5 points in terms of IoU@0.3 score, demonstrating its strong generalizability." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 522, + 554, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 642 + ], + "type": "text", + "content": "Article Grounding (AG). We assess ED-VTG on the AG on the HT-Step benchmark. Table 6 shows the fine-tuned performance on the AG task, where ED-VTG shows significant improvements over existing baselines. Notably, on the challenging unseen split, our model achieves the best results across all metrics, surpassing both the LLM and specialist models by a decent margin. The ability to handle non-groundable queries underscores ED-VTG's real-world applicability, as it does not always assume that the query is occurring in the input video." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 408, + 659 + ], + "type": "text", + "content": "4.6. Ablation Study" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Effect of Enriched Queries. We examine the impact of query enrichment through a step-by-step ablation, as shown in Tables 7 and 8. Initially, we compare the effect of enrichment without MIL paradigm against direct grounding. As" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24303" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 69, + 541, + 112 + ], + "blocks": [ + { + "bbox": [ + 69, + 69, + 541, + 112 + ], + "lines": [ + { + "bbox": [ + 69, + 69, + 541, + 112 + ], + "spans": [ + { + "bbox": [ + 69, + 69, + 541, + 112 + ], + "type": "image", + "image_path": "1e9534a438d01b1066b72f497c90405064b432e3e3ba753aca80688399119c6a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 67, + 116, + 543, + 198 + ], + "blocks": [ + { + "bbox": [ + 67, + 116, + 543, + 198 + ], + "lines": [ + { + "bbox": [ + 67, + 116, + 543, + 198 + ], + "spans": [ + { + "bbox": [ + 67, + 116, + 543, + 198 + ], + "type": "image", + "image_path": "fc6f1677a7f4aa25e9c93618e186c22c2a6ec1544e1ea72eae13b5239fdd7c86.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 555, + 224 + ], + "type": "text", + "content": "Figure 3. Example of query enrichment and detection made by ED-VTG on video paragraph grounding (VPG) task from the ActivityNet-Captions [29] dataset. In this specific sample, three different queries are enriched and localized together." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 227, + 555, + 285 + ], + "blocks": [ + { + "bbox": [ + 56, + 227, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 56, + 227, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 227, + 555, + 285 + ], + "type": "table", + "html": "
Training ParadigmCharades-STA STGActivityNet-Captions STG
ZSFT w/o PTFTZSFT w/o PTFT
R@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect48.130.631.051.431.533.268.949.045.846.326.029.650.330.134.061.138.039.2
Enrich & Detect58.137.337.760.137.038.475.156.649.750.729.533.456.335.537.865.543.443.8
Enrich & Detect w/ MIL59.539.340.262.838.440.378.262.152.652.133.135.257.536.238.667.645.144.9
", + "image_path": "e9d1602c659b14e3e038f4fd7fbe97b2522cc610f71fe1d36801c69e54dbc076.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 327, + 294, + 383 + ], + "blocks": [ + { + "bbox": [ + 55, + 288, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 55, + 288, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 288, + 555, + 323 + ], + "type": "text", + "content": "Table 7. Ablation on the effect of enriched queries. Our proposed enrich & detect framework significantly gains over directly grounding the input queries across different evaluation settings. Introducing the MIL framework further improves the performance. FT w/o PT refers directly fine-tuning on respective datasets, without performing pre-training." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 327, + 294, + 383 + ], + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 383 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 383 + ], + "type": "table", + "html": "
Training ParadigmCharades-STAANet-Captions
R@0.3R@0.5mIoUR@0.3R@0.5mIoU
Detect51.431.533.250.330.134.0
Offline Enrich + Detect51.731.533.449.829.933.7
Enrich & Detect60.137.038.456.335.537.8
", + "image_path": "ea1aa702e83a397a7d3d0753e06e994e26c068953250f43a6cb738cf52449b81.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 56, + 435, + 294, + 498 + ], + "blocks": [ + { + "bbox": [ + 55, + 384, + 295, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 384, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 295, + 430 + ], + "type": "text", + "content": "Table 8. Ablation on enrichment as a training pre-processing step. The two step enrich & detect framework is more helpful since the trained model learn to perform autonomous enrichment during evaluation. Reported results are in FT w/o PT setting." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 435, + 294, + 498 + ], + "lines": [ + { + "bbox": [ + 56, + 435, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 435, + 294, + 498 + ], + "type": "table", + "html": "
DecoderObjectivesCharades-STA STGANet-Captions
LML1gIoUR@0.3R@0.5mIoUR@0.3R@0.5mIoU
-✓--54.233.234.151.031.635.5
✓✓✓-58.536.037.055.834.437.1
✓✓-✓58.936.237.155.834.737.3
✓✓✓✓60.137.038.456.335.537.8
", + "image_path": "0348abf1f16ccbda6ba3fdedb0a2b51b1f06eeb76d8bd3e43bc5959754d12604.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 536 + ], + "type": "text", + "content": "Table 9. Ablation study on different training objectives. Training the model only using the LM loss (without the decoder) leads to significant performance drop. Results are in FT w/o PT setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 544, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 651 + ], + "type": "text", + "content": "indicated in the first two rows of Table 7, enriched queries lead to significant improvements on the Charades-STA and ActivityNet-Captions STG benchmarks. In the zero-shot (ZS) setting, enrichment results in gains of 6.7 and 3.8 mIoU points on these datasets, respectively. Introducing the MIL paradigm further enhances performance, adding 2.5 and " + }, + { + "bbox": [ + 55, + 544, + 295, + 651 + ], + "type": "inline_equation", + "content": "1.8\\mathrm{mIoU}" + }, + { + "bbox": [ + 55, + 544, + 295, + 651 + ], + "type": "text", + "content": " points. Similar improvements are observed in other evaluation settings, highlighting the substantial effectiveness of query enrichment within the MIL framework." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Additionally, we investigate the effect of offline enrichment in Table 8. In this scenario, instead of using a two-step grounding process, we enrich the queries as a training enrichment step, and the model is then directly provided with these enriched queries as input and asked to directly per" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 328, + 555, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 555, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 555, + 400 + ], + "type": "text", + "content": "form grounding. However, we find that offline enrichment is not advantageous, primarily due to the lack of enrichment during evaluation. In contrast, our two-step grounding approach allows the trained model to learn how to enrich queries and improve them autonomously when necessary, resulting in significant performance gains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 460 + ], + "type": "text", + "content": "Training Objectives. We ablate different training objectives on the Charades and ActivityNet STG benchmarks, as shown in Table 9. The decoder achieves optimal performance when both L1 and gIoU objectives are used together; omitting either one slightly reduces the scores." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 464, + 518, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 518, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 518, + 477 + ], + "type": "text", + "content": "4.7. Qualitative Results and Error Analysis" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 587 + ], + "type": "text", + "content": "Figure 3 visualizes a VPG sample from the ActivityNet-Captions dataset where ED-VTG meaningfully enriches the three input queries, then proceeds to precisely ground them. Please refer to supplementary for more qualitative results, including intuitive demonstrations of the flexibility in choosing between enrichment or direct detection, and comparison with TimeChat baseline. We also present there some interesting failure cases, such as the ones involving small and obscured objects in long input videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 388, + 602 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "In this paper, we presented ED-VTG, a novel method for fine-grained video temporal grounding using multi-modal LLMs. By enhancing queries with additional details, utilizing a lightweight decoder and trained in a multiple-instance framework, ED-VTG accurately locates temporal boundaries in videos. Our experiments show that our method outperforms existing LLM-based works and is competitive with specialized models, especially in zero-shot settings, setting a new, strong benchmark for video grounding tasks." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24304" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 167, + 85 + ], + "type": "text", + "content": "6. Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 86, + 297, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 86, + 297, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 86, + 297, + 159 + ], + "type": "text", + "content": "This codebase for this project is built on the TimeChat [72] and VTimeLLM [21] repository. We would like to thank the respective authors for their contributions. We gratefully acknowledge the following colleagues at FAIR for valuable discussions and support of our project: Tushar Nagarajan, Huiyu Wang, Yujie Lu, and Arjun Somayazulu." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "spans": [ + { + "bbox": [ + 56, + 167, + 115, + 180 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 187, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 66, + 187, + 297, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 187, + 297, + 253 + ], + "spans": [ + { + "bbox": [ + 66, + 187, + 297, + 253 + ], + "type": "text", + "content": "[1] Moloud Abdar, Meenakshi Kollati, Swaraja Kuraparthi, Farhad Pourpanah, Daniel McDuff, Mohammad Ghavamzadeh, Shuicheng Yan, Abduallah Mohamed, Abbas Khosravi, Erik Cambria, et al. A review of deep learning for video captioning. arXiv preprint arXiv:2304.11431, 2023. 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 255, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 255, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 66, + 255, + 296, + 297 + ], + "type": "text", + "content": "[2] Triantafyllos Afouras, Effrosyni Mavroudi, Tushar Nagarajan, Huiyu Wang, and Lorenzo Torresani. Ht-step: Aligning instructional articles with how-to videos. In NeurIPS, 2024. 4, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 300, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 300, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 66, + 300, + 295, + 331 + ], + "type": "text", + "content": "[3] Stuart Andrews, Ioannis Tsochantaridis, and Thomas Hofmann. Support vector machines for multiple-instance learning. In NeurIPS, pages 561-568, 2003. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 333, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 333, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 66, + 333, + 296, + 376 + ], + "type": "text", + "content": "[4] Lisa Anne Hendricks, Oliver Wang, Eli Shechtman, Josef Sivic, Trevor Darrell, and Bryan Russell. Localizing moments in video with natural language. In ICCV, pages 5803-5812, 2017. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 377, + 296, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 377, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 66, + 377, + 296, + 410 + ], + "type": "text", + "content": "[5] Max Bain, Arsha Nagrani, Gül Varol, and Andrew Zisserman. Frozen in time: A joint video and image encoder for end-to-end retrieval. In ICCV, pages 1728-1738, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 411, + 295, + 433 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 411, + 295, + 433 + ], + "spans": [ + { + "bbox": [ + 66, + 411, + 295, + 433 + ], + "type": "text", + "content": "[6] Peijun Bao, Qian Zheng, and Yadong Mu. Dense events grounding in video. In AAAI, pages 920-928, 2021. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 434, + 295, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 434, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 66, + 434, + 295, + 455 + ], + "type": "text", + "content": "[7] Hakan Bilen and Andrea Vedaldi. Weakly supervised deep detection networks. In CVPR, pages 2846-2854, 2016. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 456, + 295, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 456, + 295, + 488 + ], + "spans": [ + { + "bbox": [ + 66, + 456, + 295, + 488 + ], + "type": "text", + "content": "[8] Yi-Wen Chen, Yi-Hsuan Tsai, and Ming-Hsuan Yang. End-to-end multi-modal video temporal grounding. NeurIPS, 34:28442-28453, 2021. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 66, + 491, + 295, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 491, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 66, + 491, + 295, + 522 + ], + "type": "text", + "content": "[9] Ramazan Cinbis, Jakob Verbeek, and Cordelia Schmid. Weakly supervised object localization with multi-fold multiple instance learning. IEEE TPAMI, 39, 2015. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 524, + 295, + 556 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 295, + 556 + ], + "type": "text", + "content": "[10] Chaorui Deng, Shizhe Chen, Da Chen, Yuan He, and Qi Wu. Sketch, ground, and refine: Top-down dense video captioning. In CVPR, pages 234-243, 2021. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 558, + 295, + 601 + ], + "type": "text", + "content": "[11] Thomas G. Dietterich, Richard H. Lathrop, and Tomás Lozano-Pérez. Solving the multiple instance problem with axis-parallel rectangles. Artificial Intelligence, 89(1):31-71, 1997. 3, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 602, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 602, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 602, + 295, + 634 + ], + "type": "text", + "content": "[12] Li Dong, Jonathan Mallinson, Siva Reddy, and Mirella Lapata. Learning to paraphrase for question answering. In EMNLP, 2017. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 62, + 636, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 679 + ], + "type": "text", + "content": "[13] Tsu-Jui Fu, Linjie Li, Zhe Gan, Kevin Lin, William Yang Wang, Lijuan Wang, and Zicheng Liu. An empirical study of end-to-end video-language transformers with masked visual modeling. In CVPR, pages 22898-22909, 2023. 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 681, + 295, + 713 + ], + "type": "text", + "content": "[14] Jiyang Gao, Chen Sun, Zhenheng Yang, and Ram Nevatia. Tall: Temporal activity localization via language query. In ICCV, pages 5267-5275, 2017. 4" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 106 + ], + "type": "text", + "content": "[15] Silin Gao, Yichi Zhang, Zhijian Ou, and Zhou Yu. Paraphrase augmented task-oriented dialog generation. In ACL, pages 639-649, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 108, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 108, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 320, + 108, + 553, + 151 + ], + "type": "text", + "content": "[16] Soham Ghosh, Anuva Agarwal, Zarana Parekh, and Alexander G Hauptmann. Excl: Extractive clip localization using natural language descriptions. In NAACL, pages 1984-1990, 2019. 2, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 154, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 154, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 154, + 553, + 196 + ], + "type": "text", + "content": "[17] Ross Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Region-based convolutional networks for accurate object detection and segmentation. IEEE TPAMI, 38(1): 142-158, 2015. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 199, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 199, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 320, + 199, + 553, + 242 + ], + "type": "text", + "content": "[18] Aleksandr Gordeev, Vladimir Dokholyan, Irina Tolstykh, and Maksim Kuprashevich. Saliency-guided detr for moment retrieval and highlight detection. arXiv preprint arXiv:2410.01615, 2024. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 245, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 245, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 320, + 245, + 553, + 266 + ], + "type": "text", + "content": "[19] Kaiming He, Georgia Gkioxari, Piotr Dólar, and Ross Girshick. Mask r-cnn. In ICCV, pages 2961-2969, 2017. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 269, + 553, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 269, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 320, + 269, + 553, + 312 + ], + "type": "text", + "content": "[20] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In ICLR, 2022. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 314, + 553, + 356 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 314, + 553, + 356 + ], + "spans": [ + { + "bbox": [ + 320, + 314, + 553, + 356 + ], + "type": "text", + "content": "[21] Bin Huang, Xin Wang, Hong Chen, Zihan Song, and Wenwu Zhu. Vtimellm: Empower llm to grasp video moments. In CVPR, pages 14271-14280, 2024. 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 320, + 360, + 553, + 403 + ], + "type": "text", + "content": "[22] De-An Huang, Shijia Liao, Subhashree Radhakrishnan, Hongxu Yin, Pavlo Molchanov, Zhiding Yu, and Jan Kautz. Lita: Language instructed temporal-localization assistant. In ECCV, pages 202-218. Springer, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 320, + 405, + 553, + 437 + ], + "type": "text", + "content": "[23] Gabriel Huang, Bo Pang, Zhenhai Zhu, Clara Rivera, and Radu Soricut. Multimodal pretraining for dense video captioning. In AACL, pages 470-490, 2020. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 440, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 440, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 320, + 440, + 553, + 473 + ], + "type": "text", + "content": "[24] Jiabo Huang, Hailin Jin, Shaogang Gong, and Yang Liu. Video activity localisation with uncertainties in temporal boundary. In ECCV, pages 724-740. Springer, 2022. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 320, + 475, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 475, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 320, + 475, + 553, + 496 + ], + "type": "text", + "content": "[25] Vladimir Iashin and Esa Rahtu. Multi-modal dense video captioning. In CVPR Workshops, 2020. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 320, + 498, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 498, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 320, + 498, + 553, + 541 + ], + "type": "text", + "content": "[26] Xun Jiang, Xing Xu, Jingran Zhang, Fumin Shen, Zuo Cao, and Heng Tao Shen. Semi-supervised video paragraph grounding with contrastive encoder. In CVPR, pages 2466-2475, 2022. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 320, + 544, + 553, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 544, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 320, + 544, + 553, + 586 + ], + "type": "text", + "content": "[27] Kumara Kahatapitiya, Kanchana Ranasinghe, Jongwoo Park, and Michael S Ryoo. Language repository for long video understanding. arXiv preprint arXiv:2403.14622, 2024. 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 320, + 590, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 590, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 320, + 590, + 553, + 632 + ], + "type": "text", + "content": "[28] Dahye Kim, Jungin Park, Jiyoung Lee, Seongheon Park, and Kwanghoon Sohn. Language-free training for zero-shot video grounding. In WACV, pages 2539-2548, 2023. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 320, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 320, + 635, + 553, + 667 + ], + "type": "text", + "content": "[29] Ranjay Krishna, Kenji Hata, Frederic Ren, Li Fei-Fei, and Juan Carlos Niebles. Dense captioning events in videos. In ICCV, pages 706-715, 2017. 2, 4, 8" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 712 + ], + "type": "text", + "content": "[30] Xin Lai, Zhuotao Tian, Yukang Chen, Yanwei Li, Yuhui Yuan, Shu Liu, and Jiaya Jia. Lisa: Reasoning segmentation via large language model. In CVPR, pages 9579-9589, 2024. 2" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24305" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[31] Pilhyeon Lee and Hyeran Byun. Bam-detr: Boundary-aligned moment detection transformer for temporal sentence grounding in videos. In ECCV, pages 220-238. Springer, 2024. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 61, + 119, + 294, + 151 + ], + "type": "text", + "content": "[32] Jie Lei, Tamara L Berg, and Mohit Bansal. Detecting moments and highlights in videos via natural language queries. NeurIPS, 34:11846-11858, 2021. 1, 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 153, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 153, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 62, + 153, + 294, + 196 + ], + "type": "text", + "content": "[33] Bo Li, Yuanhan Zhang, Dong Guo, Renrui Zhang, Feng Li, Hao Zhang, Kaichen Zhang, Yanwei Li, Ziwei Liu, and Chunyuan Li. Llava-onevision: Easy visual task transfer. TMLR, 2025. 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 62, + 199, + 294, + 243 + ], + "type": "text", + "content": "[34] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742. PMLR, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 245, + 294, + 288 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 245, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 62, + 245, + 294, + 288 + ], + "type": "text", + "content": "[35] KunChang Li, Yinan He, Yi Wang, Yizhuo Li, Henhai Wang, Ping Luo, Yali Wang, Limin Wang, and Yu Qiao. Videochat: Chat-centric video understanding. arXiv preprint arXiv:2305.06355, 2023. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 344 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 344 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 344 + ], + "type": "text", + "content": "[36] Kunchang Li, Yali Wang, Yinan He, Yizhuo Li, Yi Wang, Yi Liu, Zun Wang, Jilan Xu, Guo Chen, Ping Luo, et al. Mvbench: A comprehensive multi-modal video understanding benchmark. In CVPR, pages 22195-22206, 2024. 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 348, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 348, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 348, + 294, + 390 + ], + "type": "text", + "content": "[37] Pandeng Li, Chen-Wei Xie, Hongtao Xie, Liming Zhao, Lei Zhang, Yun Zheng, Deli Zhao, and Yongdong Zhang. Momentdiff: Generative video moment retrieval from random to real. NeurIPS, 36, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 294, + 426 + ], + "type": "text", + "content": "[38] Yehao Li, Ting Yao, Yingwei Pan, Hongyang Chao, and Tao Mei. Jointly localizing and describing events for dense video captioning. CVPR, pages 7492-7500, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 294, + 471 + ], + "type": "text", + "content": "[39] Zhaowei Li, Qi Xu, Dong Zhang, Hang Song, Yiqing Cai, Qi Qi, Ran Zhou, Junting Pan, Zefeng Li, Vu Tu, et al. Groundinggpt: Language enhanced multi-modal grounding model. In ACL, pages 6657-6678, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 474, + 294, + 527 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 474, + 294, + 527 + ], + "spans": [ + { + "bbox": [ + 62, + 474, + 294, + 527 + ], + "type": "text", + "content": "[40] Kevin Qinghong Lin, Pengchuan Zhang, Joya Chen, Shraman Pramanick, Difei Gao, Alex Jinpeng Wang, Rui Yan, and Mike Zheng Shou. Univtg: Towards unified video-language temporal grounding. In ICCV, pages 2794-2804, 2023. 1, 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 531, + 294, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 531, + 294, + 574 + ], + "spans": [ + { + "bbox": [ + 62, + 531, + 294, + 574 + ], + "type": "text", + "content": "[41] Daizong Liu, Xiaoye Qu, Jianfeng Dong, Pan Zhou, Yu Cheng, Wei Wei, Zichuan Xu, and Yulai Xie. Context-aware biaffine localizing network for temporal sentence grounding. In CVPR, pages 11235-11244, 2021. 2, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 576, + 294, + 619 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 576, + 294, + 619 + ], + "spans": [ + { + "bbox": [ + 62, + 576, + 294, + 619 + ], + "type": "text", + "content": "[42] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37, 2016. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 622, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 622, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 62, + 622, + 294, + 666 + ], + "type": "text", + "content": "[43] Ye Liu, Siyuan Li, Yang Wu, Chang-Wen Chen, Ying Shan, and Xiaohu Qie. Umt: Unified multi-modal transformers for joint video moment retrieval and highlight detection. In CVPR, pages 3042-3051, 2022. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 668, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 668, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 668, + 294, + 689 + ], + "type": "text", + "content": "[44] Ilya Loshchilov and Frank Hutter. Sgdr: Stochastic gradient descent with warm restarts. In *ICLR*, 2017. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 692, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 294, + 712 + ], + "type": "text", + "content": "[45] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[46] Fan Luo, Shaoxiang Chen, Jingjing Chen, Zuxuan Wu, and Yu-Gang Jiang. Self-supervised learning for semi-supervised temporal language grounding. IEEE Transactions on Multimedia, 25:7747-7757, 2022. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 171 + ], + "type": "text", + "content": "[47] Ruipu Luo, Ziwang Zhao, Min Yang, Junwei Dong, Da Li, Pengcheng Lu, Tao Wang, Linmei Hu, Minghui Qiu, and Zhongyu Wei. Valley: Video assistant with large language model enhanced ability. arXiv preprint arXiv:2306.07207, 2023.5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 553, + 228 + ], + "type": "text", + "content": "[48] Kajing Ma, Xianghao Zang, Zerun Feng, Han Fang, Chao Ban, Yuhan Wei, Zhongjiang He, Yongxiang Li, and Hao Sun. Llavilo: Boosting video moment retrieval via adapter-based multimodal modeling. In ICCV Workshops, pages 2790-2795. IEEE, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 230, + 553, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 230, + 553, + 273 + ], + "spans": [ + { + "bbox": [ + 321, + 230, + 553, + 273 + ], + "type": "text", + "content": "[49] Muhammad Maaz, Hanoona Rasheed, Salman Khan, and Fahad Shahbaz Khan. Video-chatgpt: Towards detailed video understanding via large vision and language models. In ACL, 2024. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 275, + 553, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 275, + 553, + 317 + ], + "spans": [ + { + "bbox": [ + 321, + 275, + 553, + 317 + ], + "type": "text", + "content": "[50] Effrosyni Mavroudi, Triantafyllos Afouras, and Lorenzo Torresani. Learning to ground instructional articles in videos through narrations. In ICCV, pages 15201-15213, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 320, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 320, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 321, + 320, + 553, + 341 + ], + "type": "text", + "content": "[51] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. *ICLR*, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 343, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 343, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 321, + 343, + 553, + 386 + ], + "type": "text", + "content": "[52] WonJun Moon, Sangeek Hyun, SuBeen Lee, and Jae-Pil Heo. Correlation-guided query-dependency calibration for video temporal grounding. arXiv preprint arXiv:2311.08835, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 553, + 431 + ], + "type": "text", + "content": "[53] WonJun Moon, Sangeek Hyun, SangUk Park, Dongchan Park, and Jae-Pil Heo. Query-dependent video representation for moment retrieval and highlight detection. In CVPR, pages 23023-23033, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 434, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 434, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 553, + 464 + ], + "type": "text", + "content": "[54] Jonghwan Mun, Minsu Cho, and Bohyung Han. Local-global video-text interactions for temporal grounding. In CVPR, pages 10810-10819, 2020. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 468, + 553, + 510 + ], + "type": "text", + "content": "[55] Arsha Nagrani, Paul Hongsuck Seo, Bryan Seybold, Anja Hauth, Santiago Manen, Chen Sun, and Cordelia Schmid. Learning audio-video modalities from image captions. In ECCV, pages 407-426. Springer, 2022. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 513, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 513, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 321, + 513, + 553, + 544 + ], + "type": "text", + "content": "[56] Jinwoo Nam, Daechul Ahn, Dongyeop Kang, Seong Jong Ha, and Jonghyun Choi. Zero-shot natural language video localization. In ICCV, pages 1470-1479, 2021. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 546, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 546, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 321, + 546, + 553, + 589 + ], + "type": "text", + "content": "[57] Phuc Xuan Nguyen, Tianzhu Liu, Gaurav Prasad, Hung Hai Bui, Binh Pham, and Svetha Venkatesh. Weakly supervised action localization by sparse temporal pooling network. In CVPR, pages 6752-6761, 2019. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 591, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 591, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 321, + 591, + 553, + 633 + ], + "type": "text", + "content": "[58] Andreea-Maria Oncescu, Joao F Henriques, Yang Liu, Andrew Zisserman, and Samuel Albanie. Queryd: A video dataset with high-quality text and audio narrations. In ICASSP, pages 2265-2269. IEEE, 2021. 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 635, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 635, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 321, + 635, + 553, + 679 + ], + "type": "text", + "content": "[59] Boxiao Pan, Haoye Cai, De-An Huang, Kuan-Hui Lee, Adrien Gaidon, Ehsan Adeli, and Juan Carlos Niebles. Spatio-temporal graph for video captioning with knowledge distillation. In CVPR, pages 10870-10879, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "text", + "content": "[60] Sujoy Paul, Sourya Roy, and Amit K Roy-Chowdhury. W-talc: Weakly-supervised temporal activity localization and classification. In ECCV, pages 563-579, 2018. 3" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "24306" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[61] Shraman Pramanick, Yale Song, Sayan Nag, Kevin Qinghong Lin, Hardik Shah, Mike Zheng Shou, Rama Chellappa, and Pengchuan Zhang. Egovlpv2: Egocentric video-language pre-training with fusion in the backbone. In ICCV, pages 5285-5297, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 128, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 128, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 61, + 128, + 294, + 160 + ], + "type": "text", + "content": "[62] Archiki Prasad, Elias Stengel-Eskin, and Mohit Bansal. Rephrase, augment, reason: Visual grounding of questions for vision-language models. In ICLR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 161, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 161, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 62, + 161, + 295, + 205 + ], + "type": "text", + "content": "[63] Long Qian, Juncheng Li, Yu Wu, Yaobo Ye, Hao Fei, Tat-Seng Chua, Yueting Zhuang, and Siliang Tang. Momentor: Advancing video large language model with fine-grained temporal reasoning. In ICML, 2024. 2, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 248 + ], + "type": "text", + "content": "[64] Rui Qian, Xiaoyi Dong, Pan Zhang, Yuhang Zang, Shuanggrui Ding, Dahua Lin, and Jiaqi Wang. Streaming long video understanding with large language models. NeurIPS, 37:119336-119360, 2024. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 294, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 294, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 294, + 281 + ], + "type": "text", + "content": "[65] Hangyu Qin, Junbin Xiao, and Angela Yao. Question-answering dense video events. arXiv preprint arXiv:2409.04388, 2024.7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 294, + 326 + ], + "type": "text", + "content": "[66] Mengxue Qu, Xiaodong Chen, Wu Liu, Alicia Li, and Yao Zhao. Chatvtg: Video temporal grounding via chat with video dialogue large language models. In CVPR, pages 1847-1856, 2024. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 62, + 327, + 294, + 381 + ], + "type": "text", + "content": "[67] Hanoona Rasheed, Muhammad Maaz, Sahal Shaji, Abdelrahman Shaker, Salman Khan, Hisham Cholakkal, Rao M Anwer, Eric Xing, Ming-Hsuan Yang, and Fahad S Khan. Glamm: Pixel grounding large multimodal model. In CVPR, pages 13009-13018, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 62, + 383, + 294, + 403 + ], + "type": "text", + "content": "[68] Joseph Redmon and Ali Farhadi. Yolo9000: Better, faster, stronger. In CVPR, pages 6517-6525, 2017. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 62, + 404, + 294, + 437 + ], + "type": "text", + "content": "[69] Joseph Redmon, Santosh Divvala, Ross Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In CVPR, pages 779-788, 2016. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 62, + 437, + 294, + 470 + ], + "type": "text", + "content": "[70] Michaela Regneri, Marcus Rohrbach, Dominikus Wetzel, Stefan Thater, Bernt Schiele, and Manfred Pinkal. Grounding action descriptions in videos. TACL, 1:25-36, 2013. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 62, + 471, + 294, + 514 + ], + "type": "text", + "content": "[71] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE TPAMI, 39(6):1137-1149, 2016. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 515, + 294, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 515, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 62, + 515, + 294, + 558 + ], + "type": "text", + "content": "[72] Shuhuai Ren, Linli Yao, Shicheng Li, Xu Sun, and Lu Hou. Timechat: A time-sensitive multimodal large language model for long video understanding. In CVPR, pages 14313-14323, 2024. 1, 2, 5, 6, 7, 9" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 559, + 294, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 603 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 603 + ], + "type": "text", + "content": "[73] Seyed Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian D. Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. CVPR, pages 658-666, 2019. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 604, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 604, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 604, + 294, + 657 + ], + "type": "text", + "content": "[74] Cristian Rodriguez, Edison Marrese-Taylor, Fatemeh Sadat Saleh, Hongdong Li, and Stephen Gould. Proposal-free temporal moment localization of a natural-language query in video using guided attention. In WACV, pages 2464-2473, 2020. 2, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[75] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hongdong Li, and Stephen Gould. Dori: Discovering object relationships for moment localization of a natural language query in a video. In WACV, pages 1079–1088, 2021. 2, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 321, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[76] Cristian Rodriguez-Opazo, Edison Marrese-Taylor, Basura Fernando, Hiroya Takamura, and Qi Wu. Memory-efficient temporal moment localization in long videos. In EACL, pages 1909–1924, 2023. 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 118, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 118, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 321, + 118, + 553, + 138 + ], + "type": "text", + "content": "[77] T-YLPG Ross and GKHP Dólar. Focal loss for dense object detection. In CVPR, pages 2980-2988, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 140, + 553, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 140, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 321, + 140, + 553, + 204 + ], + "type": "text", + "content": "[78] Himanshu Sharma, Manmohan Agrahari, Sujeet Kumar Singh, Mohd Firoj, and Ravi Kumar Mishra. Image captioning: a comprehensive survey. In 2020 International Conference on Power Electronics & IoT Applications in Renewable Energy and its Control (PARC), pages 325-328. IEEE, 2020. 1, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 206, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 206, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 321, + 206, + 553, + 248 + ], + "type": "text", + "content": "[79] Matteo Stefanini, Marcella Cornia, Lorenzo Baraldi, Silvia Cascianelli, Giuseppe Fiameni, and Rita Cucchiara. From show to tell: A survey on deep learning-based image captioning. IEEE TPAMI, 45(1):539-559, 2022. 1, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 250, + 553, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 250, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 321, + 250, + 553, + 292 + ], + "type": "text", + "content": "[80] Hao Sun, Mingyao Zhou, Wenjing Chen, and Wei Xie. Trdetr: Task-reciprocal transformer for joint moment retrieval and highlight detection. In AAAI, pages 4998-5007, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 294, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 294, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 321, + 294, + 553, + 326 + ], + "type": "text", + "content": "[81] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 321, + 327, + 553, + 370 + ], + "type": "text", + "content": "[82] Chaolei Tan, Jianhuang Lai, Wei-Shi Zheng, and Jian-Fang Hu. Siamese learning with joint alignment and regression for weakly-supervised video paragraph grounding. In CVPR, pages 13569-13580, 2024. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 321, + 372, + 553, + 415 + ], + "type": "text", + "content": "[83] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. In CVPR, pages 1207-1216, 2019. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 416, + 553, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 416, + 553, + 448 + ], + "spans": [ + { + "bbox": [ + 321, + 416, + 553, + 448 + ], + "type": "text", + "content": "[84] Damien Teney, Ehsan Abbasnejad, and Anton van den Hengel. Unshuffling data for improved generalization in visual question answering. In ICCV, pages 1397-1407, 2021. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 449, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 449, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 321, + 449, + 553, + 502 + ], + "type": "text", + "content": "[85] Haibo Wang, Zhiyang Xu, Yu Cheng, Shizhe Diao, Yufan Zhou, Yixin Cao, Qifan Wang, Weifeng Ge, and Lifu Huang. Grounded-videollm: Sharpening fine-grained temporal grounding in video large language models. arXiv preprint arXiv:2410.03290, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 504, + 553, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 504, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 321, + 504, + 553, + 558 + ], + "type": "text", + "content": "[86] Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Rongkun Zheng, Jilan Xu, Zun Wang, et al. Internvideo2: Scaling video foundation models for multimodal video understanding. arXiv preprint arXiv:2403.15377, 2024. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 559, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 559, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 321, + 559, + 553, + 601 + ], + "type": "text", + "content": "[87] Yueqian Wang, Xiaojun Meng, Jianxin Liang, Yuxuan Wang, Qun Liu, and Dongyan Zhao. Hawkeye: Training video-text llms for grounding text in videos. arXiv preprint arXiv:2403.10228, 2024.5,6,7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 603, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 603, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 321, + 603, + 553, + 645 + ], + "type": "text", + "content": "[88] Hao Wu, Huabin Liu, Yu Qiao, and Xiao Sun. Dibs: Enhancing dense video captioning with unlabeled videos via pseudo boundary enrichment and online refinement. In CVPR, pages 18699-18708, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 680 + ], + "type": "text", + "content": "[89] Junbin Xiao, Pan Zhou, Tat-Seng Chua, and Shuicheng Yan. Video graph transformer for video question answering. In ECCV, pages 39-58. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "text", + "content": "[90] Junbin Xiao, Angela Yao, Yicong Li, and Tat-Seng Chua. Can i trust your answer? visually grounded video question answering. In CVPR, pages 13204-13214, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24307" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 116 + ], + "type": "text", + "content": "[91] Shen Yan, Xuehan Xiong, Arsha Nagrani, Anurag Arnab, Zhonghao Wang, Weina Ge, David Ross, and Cordelia Schmid. Unloc: A unified framework for video localization tasks. In CVPR, pages 13623-13633, 2023. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 61, + 118, + 294, + 161 + ], + "type": "text", + "content": "[92] Antoine Yang, Antoine Miech, Josef Sivic, Ivan Laptev, and Cordelia Schmid. Zero-shot video question answering via frozen bidirectional language models. In NeurIPS, pages 124-141, 2022. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 163, + 295, + 217 + ], + "type": "text", + "content": "[93] Antoine Yang, Arsha Nagrani, Paul Hongsuck Seo, Antoine Miech, Jordi Pont-Tuset, Ivan Laptev, Josef Sivic, and Cordelia Schmid. Vid2seq: Large-scale pretraining of a visual language model for dense video captioning. In CVPR, pages 10714-10726, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 62, + 220, + 294, + 251 + ], + "type": "text", + "content": "[94] Shoubin Yu, Jaemin Cho, Prateek Yadav, and Mohit Bansal. Self-chained image-language model for video localization and question answering. NeurIPS, 36, 2024. 2, 5, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 253, + 295, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 253, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 62, + 253, + 295, + 307 + ], + "type": "text", + "content": "[95] Yitian Yuan, Xiaohan Lan, Xin Wang, Long Chen, Zhi Wang, and Wenwu Zhu. A closer look at temporal sentence grounding in videos: Dataset and metric. In Proceedings of the 2nd international workshop on human-centric multimedia analysis, pages 13-21, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 309, + 294, + 352 + ], + "type": "text", + "content": "[96] Abhay Zala, Jaemin Cho, Satwik Kottur, Xilun Chen, Barlas Oguz, Yashar Mehdad, and Mohit Bansal. Hierarchical video-moment retrieval and step-captioning. In CVPR, pages 23056-23065, 2023. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 408 + ], + "type": "text", + "content": "[97] Rowan Zellers, Jiasen Lu, Ximing Lu, Youngjae Yu, Yanpeng Zhao, Mohammadreza Salehi, Aditya Kusupati, Jack Hessel, Ali Farhadi, and Yejin Choi. Merlot reserve: Neural script knowledge through vision and language and sound. In CVPR, pages 16375-16387, 2022. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 294, + 453 + ], + "type": "text", + "content": "[98] Runhao Zeng, Haoming Xu, Wenbing Huang, Peihao Chen, Mingkui Tan, and Chuang Gan. Dense regression network for video grounding. In CVPR, pages 10287-10296, 2020. 2, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 455, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 455, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 62, + 455, + 294, + 508 + ], + "type": "text", + "content": "[99] Xiangyu Zeng, Kunchang Li, Chenting Wang, Xinhao Li, Tianxiang Jiang, Ziang Yan, Songze Li, Yansong Shi, Zhengrong Yue, Yi Wang, et al. Timesuite: Improving mllms for long video understanding via grounded tuning. In ICLR, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 510, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 554 + ], + "type": "text", + "content": "[100] Ce Zhang, Taixi Lu, Md Mohaiminul Islam, Ziyang Wang, Shoubin Yu, Mohit Bansal, and Gedas Bertasius. A simple lmm framework for long-range video question-answering. In EMNLP, pages 21715–21737, 2024. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 555, + 294, + 588 + ], + "type": "text", + "content": "[101] Chen-Lin Zhang, Jianxin Wu, and Yin Li. Actionformer: Localizing moments of actions with transformers. In ECCV, pages 492-510. Springer, 2022. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 622 + ], + "type": "text", + "content": "[102] Hao Zhang, Aixin Sun, Wei Jing, and Joey Tianyi Zhou. Span-based localizing network for natural language video localization. In ACL, pages 6543-6554, 2020. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 294, + 656 + ], + "type": "text", + "content": "[103] Hang Zhang, Xin Li, and Lidong Bing. Video-llama: An instruction-tuned audio-visual language model for video understanding. In EMNLP, pages 543–553, 2023. 5, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 689 + ], + "type": "text", + "content": "[104] Junchao Zhang and Yuxin Peng. Object-aware aggregation with bidirectional temporal graph for video captioning. In CVPR, pages 8327–8336, 2019. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[105] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 645 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 94 + ], + "type": "text", + "content": "localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 139 + ], + "type": "text", + "content": "[106] Songyang Zhang, Houwen Peng, Jianlong Fu, and Jiebo Luo. Learning 2d temporal adjacent networks for moment localization with natural language. In AAAI, pages 12870-12877, 2020. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 184 + ], + "type": "text", + "content": "[107] Songyang Zhang, Houwen Peng, Jianlong Fu, Yijuan Lu, and Jiebo Luo. Multi-scale 2d temporal adjacency networks for moment localization with natural language. IEEE TPAMI, 44(12):9073-9087, 2021. 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 218 + ], + "type": "text", + "content": "[108] Zhu Zhang, Zhijie Lin, Zhou Zhao, and Zhenxin Xiao. Cross-modal interaction networks for query-based moment retrieval in videos. In ACM SIGIR, pages 655-664, 2019. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 252 + ], + "type": "text", + "content": "[109] Yue Zhao, Ishan Misra, Philipp Krahenbuhl, and Rohit Girdhar. Learning video representations from large language models. In CVPR, pages 6586-6597, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 317, + 254, + 553, + 285 + ], + "type": "text", + "content": "[110] Luowei Zhou, Chenliang Xu, and Jason Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[111] Luowei Zhou, Chenliang Xu, and Jason J Corso. Towards automatic learning of procedures from web instructional videos. In AAAI, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 321, + 553, + 364 + ], + "type": "text", + "content": "[112] Luowei Zhou, Yingbo Zhou, Jason J. Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. CVPR, pages 8739-8748, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 366, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 408 + ], + "type": "text", + "content": "[113] Luowei Zhou, Yingbo Zhou, Jason J Corso, Richard Socher, and Caiming Xiong. End-to-end dense video captioning with masked transformer. In CVPR, pages 8739-8748, 2018. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 411, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 441 + ], + "type": "text", + "content": "[114] Xingyi Zhou, Dequan Wang, and Philipp Krahenbuhl. Objects as points. In arXiv preprint arXiv:1904.07850, 2019. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 488 + ], + "type": "text", + "content": "[115] Xingyi Zhou, Anurag Arnab, Shyamal Buch, Shen Yan, Austin Myers, Xuehan Xiong, Arsha Nagrani, and Cordelia Schmid. Streaming dense video captioning. In CVPR, pages 18243-18252, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 490, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 532 + ], + "type": "text", + "content": "[116] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. ArXiv, abs/2204.08121, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 534, + 553, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 600 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 600 + ], + "type": "text", + "content": "[117] Wanrong Zhu, Bo Pang, Ashish V. Thapliyal, William Yang Wang, and Radu Soricut. End-to-end dense video captioning as sequence generation. In Proceedings of the 29th International Conference on Computational Linguistics, pages 5651-5665, Gyeongju, Republic of Korea, 2022. International Committee on Computational Linguistics. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 645 + ], + "type": "text", + "content": "[118] Dimitri Zhukov, Jean-Baptiste Alayrac, Ramazan Gokberk Cinbis, David Fouhey, Ivan Laptev, and Josef Sivic. Cross-task weakly supervised learning from instructional videos. In CVPR, pages 3537-3545, 2019. 4" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24308" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_content_list.json b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..3e9f89a4bb07ae1b2f43c72555ff30f56321106e --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_content_list.json @@ -0,0 +1,1593 @@ +[ + { + "type": "text", + "text": "Ensemble Foreground Management for Unsupervised Object Discovery*", + "text_level": 1, + "bbox": [ + 138, + 130, + 866, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziling Wu, $^{\\dagger}$ Armaghan Moemeni, Praminda Caleb-Solly University of Nottingham, UK", + "bbox": [ + 277, + 180, + 717, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{ziling.wu, armaghan.moemeni, praminda.caleb-solly}@nottingham.ac.uk", + "bbox": [ + 196, + 219, + 795, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted annotations. Recent progress in self-supervised representation learning [9, 66] has led to some success in UOD algorithms [21, 53, 69]. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or background, and 2) knowing how many objects remain undiscovered. To address these two problems, previous solutions rely on foreground priors [53, 59, 67, 69] to distinguish if the discovered region is foreground, and conduct one or fixed iterations of discovery. However, the existing foreground priors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper introduces UnionCut, a robust and well-grounded foreground prior based on min-cut [2] and ensemble methods [18] that detects the union of foreground areas of an image, allowing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the foreground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut.", + "bbox": [ + 89, + 301, + 483, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 768, + 220, + 786 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Instance segmentation is a fundamental computer vision task requiring the algorithm to recognize and segment each instance in an image, which has been widely applied in autonomous driving [14], medical image analysis [48], re", + "bbox": [ + 89, + 796, + 483, + 857 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7256056cee5522836befa7ab7fdc71a9c2b46db91fa09351ac94329987b6e63.jpg", + "image_caption": [ + "Figure 1. Examples of fixing MaskCut [69]'s errors and making its discovery stop at appropriate time using UnionCut and UnionSeg." + ], + "image_footnote": [], + "bbox": [ + 517, + 268, + 903, + 406 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "mote sensing [79], etc. However, pixel-level annotations are required to train an instance segmentation model, which are expensive and time-consuming to obtain. To cut down the training cost of instance segmentation models, strategies employing semi-supervision [12, 20, 29, 32, 43, 55, 56, 71], weak supervision [13, 33, 34, 39, 44, 57, 68, 72], and self-supervision [21, 35, 53, 67, 69] have been proposed to reduce the usage of pixel-level annotations.", + "bbox": [ + 511, + 474, + 906, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Self-supervision aims to train an instance segmentation model without any annotations. It utilizes an unsupervised object discovery (UOD) algorithm to detect and segment objects in images without handcrafted annotations, following which, the detection result is used as the ground truth to train a segmentation model [15, 16, 21, 26, 35, 67]. Recent state-of-the-art UOD algorithms [21, 35, 53, 67, 69] are based on a vision transformer (ViT) [19] pre-trained by self-supervised representation learning [9]. These draw on a phenomenon where the attention map of the last layer of the ViT indicates regions of objects in the image. However, without the foreground knowledge given by the ground truth, existing UOD methods cannot robustly judge whether the explored region belongs to the foreground or not, and as such, the algorithm could return the background area of the image as a discovered object. As a remedy, existing approaches introduce foreground priors (e.g. the location and size of the discovered region) to help them distinguish if an area in the image is foreground. However, we have found that existing foreground priors are not robust because they", + "bbox": [ + 511, + 599, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "\\*Author's Accepted Manuscript, Released under the Creative Commons license: Attribution 4.0 International (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ $\\dagger$ Corresponding author", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20268", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "are based on simple rules, assumptions, and heuristic design. Besides, the number of objects in the image varies and is unknown to UOD approaches, resulting in a dilemma of when to stop the discovery. Thus, existing UOD algorithms either segment only one region as the object or continue discovery for a predetermined number of iterations, leading to either missing objects or mistakenly identifying background as foreground due to unreliable foreground priors and excessive exploration. For example, in Fig. 1, MaskCut [69], a UOD method, returns a background area (the 1st and 3rd row), and misses the car on the left of the 2nd image.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the problem, we propose an ensemble method named UnionCut. It detects the union of foreground regions in an image, referred to as \"foreground union\" in this paper. With this foreground union as the foreground prior, UOD methods can accurately judge if a discovered region is the foreground by checking if the foreground union covers it, and stop further exploration when the majority of the foreground union has been discovered, as shown in Fig. 1. Unlike previous assumption-based foreground priors, UnionCut's robustness and reliability are demonstrated with mathematical and statistical ensemble theories. Besides, we propose UnionSeg, a ViT distilled from UnionCut to increase the efficiency and accuracy of foreground union detection.", + "bbox": [ + 89, + 258, + 480, + 455 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our work makes the following contributions: Insights: Previous UOD studies focused mainly on the design of UOD algorithms, overlooking foreground priors' importance. We reveal that UOD's performance can also be increased by robust foreground priors. Methods: We propose UnionCut based on ensemble and UnionSeg using distillation to detect foreground union in the image, serving as foreground priors for UOD methods to verify results and decide when to stop discovery. Unlike previous heuristic and assumption-based foreground priors, UnionCut is mathematically and statistically robust and reliable. Results: After being integrated with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks, indicating that UnionCut and UnionSeg enable UOD algorithms to give more accurate predictions on the existence of objects in the image. Impacts: UnionCut and UnionSeg can be used as future UOD methods' default foreground priors, taking the place of existing assumption-based, heuristic, and unreliable foreground priors, and enhancing UOD performance.", + "bbox": [ + 89, + 455, + 482, + 787 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 805, + 232, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Self-supervised Representation Learning", + "text_level": 1, + "bbox": [ + 89, + 832, + 439, + 849 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-supervised representation learning aims to train the model to generate distinguishable embeddings for images and output similar embeddings of images with the same se", + "bbox": [ + 89, + 854, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "mantic information without using handcrafted annotations. Most existing approaches [8-11, 25, 28, 42, 66, 76] use a student-teacher architecture, where each model processes one of two differently augmented versions (e.g. horizontal flipping, randomly cropping, colour jitter, etc.) of the same image. The optimization goal is to make the models produce similar outputs for augmentations of the same image and distinct outputs for augmentations from different images. DINO [9] is a representative method of these approaches, where the attention map of the last layer of a ViT pre-trained by DINO indicates the location of objects in the image. Such a property is employed in our approach.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Unsupervised Object Discovery", + "text_level": 1, + "bbox": [ + 511, + 295, + 790, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "As for UOD, Wang et al. [67] proposed FreeMask to discover objects by comparing the similarity between features from a self-supervised CNN [66] at each pixel of the image. Instead of using CNN, in [21, 53, 59, 69], embeddings of patches output by a ViT pre-trained by DINO[9] are utilized to generate pseudo-masks. In [53, 59], a patch is selected as the seed, and the pseudo mask is given by combining patches sharing similar features with the seed. To increase the pseudo-mask quality, Normalized Cut (NCut) [50] is leveraged by Wang et al. [21] to segment the feature map output by DINO for pseudo-mask generation. In [69], NCut is expanded to segment multiple objects in an image by iteratively removing a new segmented region from the image. However, the absence of ground truth makes it hard to determine if any discovered region is actually part of the foreground. Our approach provides a solution to this.", + "bbox": [ + 511, + 321, + 903, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Foreground Priors", + "text_level": 1, + "bbox": [ + 511, + 587, + 692, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "UOD algorithms leverage foreground priors to judge if the discovered region belongs to the foreground. Specifically, Simeoni et al. [53] assumes that the area occupied by the foreground should be smaller than the background. FreeMask [67] removes discovered regions whose widths are longer than $95\\%$ of the width of the image or are located in the top $20\\%$ region of the image. Gansbeke et al. [59] and Simeoni et al. [54] proposed that the semantic information of the foreground should be close to the global semantic information of the entire image. In [21, 69], the region where the point with the highest absolute confidence value is located is made as the foreground. Wang et al. [69] suggested that the foreground is unlikely to cover three or four corners of the image. Li and Shin [35] proposed that a background mask is likely to occupy at least two edges of the image. Although these priors achieved limited success, they are all heuristic and assumption-based. In contrast, the robustness and reliability of our proposed methods are demonstrated by mathematical and statistical ensemble theories.", + "bbox": [ + 511, + 613, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20269", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/380f4eb0f2e6ed120a1ba9d121c52c6e65b88b222142cad94d0ae4726457445e.jpg", + "image_caption": [ + "Figure 2. Framework of UnionCut. UnionCut takes an image as input with features extracted from a frozen DINO, and applies Unit Voters (UVs) with each using one patch as its seed. Red rectangles refer to the seed patch, and blue areas indicate the UV's output of the region similar to its seed patch. Aggregating all UV's outputs yields a heat map of the background, which is then inverted, thresholded, and rectified with a corner prior to produce a binary mask of the image's foreground union." + ], + "image_footnote": [], + "bbox": [ + 125, + 89, + 872, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Ensemble Methods", + "text_level": 1, + "bbox": [ + 89, + 373, + 274, + 387 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Ensemble methods are learning algorithms that construct a set of classifiers and then classify new data points by taking a (weighted) vote of their predictions [18]. Most ensemble methods can be classified into two types: 1) Parallel a set of weak classifiers run in parallel and form a strong classifier by performing a plurality vote, e.g. Pasting [5], Bagging [4], and Random Forest [6]; and 2) Sequential running batches of weak classifiers iteratively with each batch focusing on hard cases of the last one, e.g. [60] and AdaBoost [23]. UnionCut is of the former: a set of weak classifiers is created for an image, with each in parallel voting every patch $(8 \\times 8$ pixels) of the image as foreground or background, which forms a strong classifier by aggregating votes from them to detect the foreground union of an image.", + "bbox": [ + 89, + 395, + 483, + 608 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 619, + 225, + 636 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UnionCut is an ensemble method, enabling UOD methods to judge if a detected region is foreground and stop discovery when most objects have been discovered to avoid over-segmentation. In the context of ensemble methods, UnionCut serves as the strong classifier, based on many parallel weak classifiers named \"unit voter\" (UV) here. Fig. 2 illustrates the frameworks of the proposed UnionCut. Unlike previous assumption-based foreground priors mentioned in Sec. 2.3, UnionCut is reliable and robust with ensemble theories. First, we introduce the details of UV.", + "bbox": [ + 89, + 643, + 483, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Unit Voter", + "text_level": 1, + "bbox": [ + 89, + 801, + 209, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UV is based on features output by a ViT pretrained by DINO [9]. In our implementation, an image is resized into $224 \\times 224$ and serialized into $28 \\times 28$ (i.e. 784) patches of $8 \\times 8$ pixels before feature extraction of each patch with the ViT. The set of patches of an image is defined as $P =$", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/b4eb1719381731e08dc1656ddb5f8503c0d2a65289a6c27500513d98804d5ee1.jpg", + "image_caption": [ + "Figure 3. Overview of Unit Voter (UV). For clarity, the image is divided into $3 \\times 3$ patches. With the seed patch and anti-seed patches of the UV initialized, the image is modeled as a graph with each patch and two dummy terminals being nodes and connected by edges whose thickness reflects their weights. After applying min-cut to the graph, thin edges are removed to divide the graph into two disjoint parts containing one terminal node. The part with the Source node is the region output by the UV, indicating the area of features similar to the seed patch." + ], + "image_footnote": [], + "bbox": [ + 598, + 372, + 830, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\{p_1,p_2,\\dots ,p_{784}\\}$ with the corresponding L2-normalized Key vectors as their features output by the last layer of the model, which is defined as $K = \\{k_{1},k_{2},\\ldots ,k_{784}\\}$ . Then, the patches and features are leveraged by each UV.", + "bbox": [ + 511, + 732, + 903, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As required by ensemble methods, the number of weak classifiers (i.e. UV) to form a strong classifier (i.e. Union-Cut) should be significant, and each UV should be different [18]. To achieve these goals, 784 independent UVs can be built based on the 784 patches of an image, with each UV corresponding to a unique patch $p_f \\in P$ . The patch corresponding to a UV is called that UV's \"seed patch\".", + "bbox": [ + 511, + 795, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20270", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Taking a UV built from a seed patch $p_f \\in P$ as an example, its pipeline is illustrated in Fig. 3. Specifically, given a UV's seed patch $p_f \\in P$ with feature $k_f \\in K$ , the set of patches dissimilar to $p_f$ , defined as the UV's anti-seed patches set $B_f = \\{p_b | p_b \\in P, b \\neq f, k_b^T k_f < 0\\}$ , can be obtained. With $p_f, B_f$ , all $28 \\times 28$ patches $P$ and features $K$ , a directed graph is constructed, as shown in Fig. 3, where each patch is a node and two dummy terminal nodes, named \"Source\" and \"Target\", are added. In the graph, nodes of adjacent patches (i.e. eight-neighborhood patches) are connected by bidirectional edges (named \"n-links\") with weights representing their similarity. Directed edges named \"t-links\" connect the Source to each patch node or each patch node to the Target, with weights based on the likelihood of a patch being more similar to the UV's seed patch $p_f$ or its anti-seed patches set $B_f$ . The terms \"t-link\" and \"n-link\" originate from early image segmentation methods that modeled images as graphs [3, 49]. The purpose of building this graph is to split the image by removing edges, forming two disconnected subgraphs with each either containing the Source or the Target, while minimizing the total weight of the removed edges. The subgraph with the Source is expected to contain patches similar to the UV's seed patch, while the one with the Target includes patches similar to the anti-seed patches $B_f$ . This bipartition process is also known as the min-cut problem, which is solved by UV using an efficient solution [2]. Next, we introduce the algorithms for assigning weights to n-links and t-links.", + "bbox": [ + 89, + 90, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "n-links The weight $w(p_i, p_j)$ of the n-link between two adjacent patches $p_i$ and $p_j$ is given by Eq. (1).", + "bbox": [ + 89, + 513, + 483, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(p _ {i}, p _ {j}\\right) = \\frac {e ^ {- \\frac {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}{\\beta}}}{\\sqrt {\\left(i _ {x} - j _ {x}\\right) ^ {2} + \\left(i _ {y} - j _ {y}\\right) ^ {2}}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 564, + 480, + 606 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\beta = \\frac {2 \\sum_ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} (p _ {i} , p _ {j}) (k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}{C}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 606, + 439, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(i_x,i_y)$ and $(j_{x},j_{y})$ are the indices of the two patches under the scale of the feature map (i.e. $28\\times 28$ ) output by the ViT. $\\mathbf{1}(p_i,p_j)$ is an indicator function which returns 1 if the $p_i$ and $p_j$ are adjacent, otherwise 0. Finally, $C$ is the number of n-links in the graph.", + "bbox": [ + 89, + 643, + 482, + 718 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "t-links An UV's graph has three kinds of patches: seed patch $p_f$ , anti-seed patches set $B_f$ , and the other patches $P - B_f \\cup \\{p_f\\}$ , whose weight assignment strategies for t-links are different. Suppose the Source and Target are defined as $S$ and $T$ . After min-cut, the t-link from $S$ to $p_f$ should be kept while the t-link between $p_f$ and $T$ should be cut. To this end, a large weight $W$ is given to the t-link from $S$ to $p_f$ so that it is too large to be cut, and the weight between $p_f$ and $T$ is 0. Following early graph cut methods [3, 49], $W$ can be decided by Eq. (2).", + "bbox": [ + 89, + 719, + 483, + 869 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW = 1 + \\max _ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} \\left(p _ {i}, p _ {j}\\right) w \\left(p _ {i}, p _ {j}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 885, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similarly, anti-seed patches must stay connected with $T$ after min-cut. Therefore, t-links from $\\forall p_{b}\\in B_{f}$ to $T$ must have a large value, which can also be given by Eq. (2), while the weights of t-links from $S$ to $\\forall p_{b}\\in B_{f}$ are 0.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weights of other t-links between $S$ or $T$ and $p_i \\notin B_f \\bigcup \\{p_f\\}$ are decided by Eq. (3).", + "bbox": [ + 513, + 151, + 906, + 183 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw \\left(p _ {i}, S\\right) = - \\log \\frac {e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}}}{e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}}} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 200, + 903, + 239 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nw (p _ {i}, T) = - \\log \\frac {e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}{e ^ {\\sqrt {(k _ {i} - k _ {f}) ^ {T} (k _ {i} - k _ {f})}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 236, + 874, + 266 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Eq. (3) explains that if a patch $p_i \\notin B_f \\cup \\{p_f\\}$ is more similar to anti-seed patches set $B_f$ than the seed patch $p_f$ , $w(p_i, T)$ should exceed $w(p_i, S)$ , making the t-link from $S$ to $p_i$ more likely to be cut, and vice versa.", + "bbox": [ + 511, + 272, + 905, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the graph initialized using the method introduced above, the UV executes a min-cut and returns a binary mask, where \"1\" indicates patches remaining connected to the Source, i.e. regions sharing features with the seed patch $p_f$ . Therefore, the function of each UV is to return the region that is similar to its seed patch.", + "bbox": [ + 511, + 332, + 905, + 422 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. UnionCut", + "text_level": 1, + "bbox": [ + 513, + 431, + 625, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "UnionCut is a strong classifier composed of 784 UVs as weak classifiers, with each UV using a unique patch from the image's $28 \\times 28$ patches as its seed patch. Considering the function of UV introduced above, there are two types of UVs in UnionCut: 1) background UV: if its seed patch is background, the UV returns a binary mask indicating background regions; and 2) foreground UV: if its seed patch is foreground, the UV outputs a binary mask of foreground regions. UnionCut's output is determined by voting from the 784 UVs. Given the set of binary masks $M = \\{m_f | \\forall p_f \\in P\\}$ returned by all UVs, a heat map $A$ is generated by aggregating all UVs' output using Eq. (4).", + "bbox": [ + 511, + 452, + 906, + 635 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nA (x, y) = \\sum_ {m _ {f} \\in M} m _ {f} (x, y) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 641, + 903, + 660 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x \\in [1, 28]$ and $y \\in [1, 28]$ indicate the 2D coordinate of a patch's response on $A$ .", + "bbox": [ + 511, + 666, + 905, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For easier understanding, intuitively, the foreground usually occupies a smaller area than the background in the image [53]. Thus, we can assume that there are more background UVs than foreground UVs in UnionCut, causing background areas to have higher intensities on the aggregated heat map $A$ . An example can be seen in Fig. 2. To make it intuitive, the brightness of $A$ is inverted so that foreground regions have higher responses than the background, producing an inverted heat map $H$ using Eq. (5). By thresholding $H$ , most objects in the image are segmented as the foreground union.", + "bbox": [ + 511, + 696, + 906, + 862 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH (x, y) = 2 5 5 - 2 5 5 \\times \\frac {A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)}{\\operatorname* {m a x} _ {x , y} A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 882, + 903, + 905 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20271", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that the assumption introduced above is for easier understanding of our approach. UnionCut is mathematically and statistically robust, with an explanation in Appendix 6. Moreover, UnionCut remains effective even when the foreground area is larger than the background, with examples in Appendix 7. The mathematical explanation also supports this claim. Furthermore, compared to more straightforward methods detecting patches similar to a given patch, e.g. matching with cosine similarity [53, 54], UV is more sophisticated. Appendix 8 provides a detailed analysis of UV's irreplaceable advantages and necessity.", + "bbox": [ + 89, + 90, + 483, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.1. Thresholding", + "text_level": 1, + "bbox": [ + 89, + 263, + 230, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A threshold for the bipartition of the heat map is required to estimate the foreground union. UnionCut automatically determines a threshold for the heat map $H$ by clustering the response values on it with Mean-Shift [24]. Since the bright areas on the heat map represent the foreground, the resulting clusters are sorted by their centres in descending order, with the top half retained and identified as the union of the foreground in the image. Finally, a $28 \\times 28$ binary mask $U_{cut}$ with \"1\" as the foreground union is obtained.", + "bbox": [ + 89, + 281, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Corner Prior", + "text_level": 1, + "bbox": [ + 89, + 424, + 228, + 436 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Similar to [69], a corner prior is used by UnionCut to check the reliability of the foreground union mask. Specifically, if $U_{cut}$ occupies four corners of the image, it must be inverted (i.e. $U_{cut} := 1 - U_{cut}$ ) because the foreground union is unlikely to occupy four corners simultaneously. The reliability of this corner prior is discussed in Appendix 9.", + "bbox": [ + 89, + 441, + 483, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. UnionSeg", + "text_level": 1, + "bbox": [ + 89, + 541, + 202, + 558 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "While UnionCut effectively detects an image's foreground union, running 784 UVs per image is computationally expensive. To improve efficiency, we introduce UnionSeg, a ViT-based model distilled from UnionCut as a surrogate model. Unlike UnionCut, UnionSeg is an end-to-end model that directly predicts the foreground union from an input image. To minimize learnable parameters, inspired by FOUND [54], UnionSeg uses a frozen ViT trained by DINO [9] as its backbone, followed by a learnable $1 \\times 1$ convolutional layer and a sigmoid function, compressing each ViT patch feature into a scalar confidence score. Patches with scores above 0.5 are classified as the foreground union. Appendix 10 details the differences between UnionSeg and FOUND. Suppose the soft output of UnionSeg of an image is $U_{seg}^{s}$ , with its thresholded hard output defined as $U_{seg}^{h}$ . Given UnionCut's output $U_{cut}$ of an image, the label $L$ of an image used to train UnionSeg is defined by Eq. (6).", + "bbox": [ + 89, + 563, + 483, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL = \\left\\{ \\begin{array}{l l} U _ {c u t}, & \\text {i f I o U} \\left(U _ {s e g} ^ {h}, U _ {c u t}\\right) < 0. 5, \\\\ U _ {s e g} ^ {h}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 828, + 482, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $IoU(\\cdot)$ returns the intersection over union between two binary masks. The motivation behind Eq. (6) is that", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UnionCut's output may contain errors, so UnionSeg does not rely entirely on UnionCut's output as labels. This enables UnionSeg to predict the foreground union more efficiently and accurately than UnionCut. The loss function used to train UnionSeg is shown in Eq. (7).", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL o s s = L \\log U _ {s e g} ^ {s} + (1 - L) \\log \\left(1 - U _ {s e g} ^ {s}\\right) + 1 (\\text {i t e r} < 1 0 0) \\Lambda \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 191, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\Lambda = U _ {c u t} \\log U _ {s e g} ^ {s} + (1 - U _ {c u t}) \\log (1 - U _ {s e g} ^ {s})\n$$\n", + "text_format": "latex", + "bbox": [ + 534, + 208, + 779, + 222 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}(iter < 100)$ returns 1 if the training iterations are fewer than 100; otherwise 0. This ensures that at the start of training, when UnionSeg is not yet well trained, it is forced to use both UnionCut's output $U_{cut}$ and $L$ as labels to aid learning.", + "bbox": [ + 511, + 234, + 906, + 310 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 325, + 645, + 343 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Several experiments are conducted to evaluate the effectiveness of UnionCut/UnionSeg, comparing the performance of state-of-the-art UOD methods before and after integrating UnionCut or UnionSeg on several tasks: unsupervised single object discovery (Sec. 4.1), unsupervised saliency detection (Sec. 4.2), and self-supervised instance segmentation (Sec. 4.3). We then discuss the properties of UnionCut and UnionSeg in Sec. 4.4, and evaluate the impact of different components of them in Sec. 4.5. Additionally, we provide more qualitative comparisons in Appendix 15.", + "bbox": [ + 511, + 351, + 906, + 502 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baseline UOD methods Our experiments use existing state-of-the-art UOD methods for single or multiple object discovery as baselines, i.e. LOST [53], TokenCut [21], FOUND [54] and CutLER [69]. To enhance their performance, we integrate them with UnionCut and UnionSeg by replacing their standard foreground priors with UnionCut/UnionSeg, with details provided in Appendix 11.", + "bbox": [ + 511, + 503, + 905, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details We use a ViT-S/8 [19] pretrained by DINO [9] as the image feature extractor for UnionCut and UnionSeg. Following [52, 54], we distill UnionSeg from UnionCut using DUTS-TR[64], a saliency detection dataset with 10,553 images. Training is done with a batch size of 50, an initial learning rate of 0.05 that decays to $95\\%$ every 50 iterations, and AdamW [38] for parameter updates over 600 iterations. Implementation details for all tested methods, i.e. LOST+UnionCut/UnionSeg, TokenCut+UnionCut/UnionSeg, FOUND+UnionCut/UnionSeg, and CutLER+UnionSeg, are provided in Appendix 11.", + "bbox": [ + 511, + 609, + 905, + 776 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Unsupervised Single Object Discovery", + "text_level": 1, + "bbox": [ + 511, + 786, + 841, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this experiment, we combine baseline UOD methods with UnionCut and UnionSeg to see if their performance improves in unsupervised single object discovery.", + "bbox": [ + 511, + 809, + 905, + 854 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Benchmark Following [21, 53, 54], the trainval sets of PASCAL VOC07 (5011 images), VOC12 (11540 images) [22] and COCO20K [37, 61] (19817 randomly sam", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20272", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/34e5c2cfde20f9176fb09061c9c2d37162cc8019e9b3693f58dde9285d7b32aa.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodVOC07VOC12COCO20K
- No learning -
Selective Search [58]18.820.916.0
EdgeBoxes [80]31.131.628.8
Kim et al. [30]43.946.435.1
Zhang et al. [75]46.250.534.8
DDT+ [70]50.253.138.2
rOSD [61]54.555.348.5
LOD [62]53.655.148.5
DINO-seg [9, 53](ViT-S/16 [9])45.846.242.0
DSS [40](ViT-S/16 [9])62.766.452.2
LOST [53](ViT-S/16 [9])61.964.050.7
LOST(ViT-S/16 [9])+UnionCut62.4(0.5↑)65.3(1.3↑)51.9(1.2↑)
LOST(ViT-S/16 [9])+UnionSeg63.2(1.3↑)67.1(3.1↑)54.9(4.2↑)
TokenCut [21](ViT-S/16 [9])68.872.158.8
TokenCut(ViT-S/16 [9])+UnionCut69.2(0.4↑)72.3(0.2↑)62.1(3.3↑)
TokenCut(ViT-S/16 [9])+UnionSeg69.7(0.9↑)72.7(0.6↑)62.6(3.8↑)
- With learning -
FreeSOLO [67, 69]44.049.735.2
LOD+CAD [21, 53]56.361.652.7
rOSD+CAD [21, 53]58.362.353.0
LOST+CAD [53](ViT-S/16 [9])65.770.457.5
SelfMask [52]72.375.362.7
FOUND [54](ViT-S/16 [9])72.576.162.9
FOUND(ViT-S/16 [9])‡71.475.762.9
FOUND(ViT-S/16 [9])+UnionCut72.6(1.2↑)76.1(0.4↑)63.0(0.1↑)
FOUND(ViT-S/16 [9])+UnionSeg72.3(0.9↑)75.9(0.2↑)63.1(0.2↑)
CutLER [69](ViT-B/8 [9])†73.369.570.7
CutLER(ViT-B/8 [9])+UnionSeg73.8(0.5↑)71.2(1.7↑)72.4(1.7↑)
", + "bbox": [ + 91, + 88, + 496, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performances of UOD methods on single object discovery with red indicates the highest score. CAD: a class-agnostic detector trained with the output of no-learning UOD methods as the ground truth. †: results of official implementations and checkpoints by our measurement. ‡: results of our replication with official implementation.", + "bbox": [ + 89, + 493, + 483, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pled images from COCO2014 [37]) are selected as the benchmark for evaluation.", + "bbox": [ + 89, + 608, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Similar to [21, 53, 54, 61, 62], we use CorLoc (Correct Localization) to evaluate UOD algorithms' performance on single object discovery. CorLoc scores an image as 1 if the intersection over union (IoU) between any ground truth bounding box and the algorithm's predicted bounding box exceeds 0.5, otherwise it scores 0. In our experiments, since CutLER can detect multiple objects, we select the bounding box with the highest confidence score as its single object prediction for the image.", + "bbox": [ + 89, + 641, + 482, + 777 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 shows that after integrating UnionCut or UnionSeg, our selected baseline UOD methods improve across all three benchmarks, achieving state-of-the-art results, indicating that UnionCut/UnionSeg as a robust foreground prior enhances UOD algorithms' performance in single object discovery. In Appendix 12, we also evaluate these UOD methods from the perspective of the performance upper boundary proposed by Rambhatla et al. [45].", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Unsupervised Saliency Detection", + "text_level": 1, + "bbox": [ + 511, + 90, + 802, + 107 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the performance of UOD methods before and after integrating UnionCut/UnionSeg on saliency detection, which involves segmenting salient objects in images. Since CutLER predicts multiple object masks per image, we use the union of these masks as its saliency detection result.", + "bbox": [ + 511, + 112, + 903, + 186 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Similar to [21, 40, 53, 54, 63], ECSSD [51] (1000 real-world images), DUTS-TE [64] (5019 images sampled from ImageNet test), and DUT-OMRON [74] (5168 natural images) are chosen as the benchmark.", + "bbox": [ + 511, + 186, + 903, + 247 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics Following [21, 53, 54], we report results with three metrics: Accuracy, IoU, and $\\max F_{\\beta}$ . Accuracy measures the percentage of correctly classified pixels compared to ground truth. IoU is the ratio of the intersection between the UOD output and ground truth to their union. $\\max F_{\\beta}$ is calculated as $\\frac{(1 + \\beta^2)\\text{Precision} \\times \\text{Recall}}{\\beta^2\\text{Precision} + \\text{Recall}}$ where Precision is the percentage of predicted pixels matching ground truth, and Recall is the proportion of ground truth pixels covered by the prediction. We set $\\beta$ to 0.3, as in [21, 53, 54].", + "bbox": [ + 511, + 248, + 905, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 2 illustrates that the selected baseline methods improve when integrated with UnionCut/UnionSeg. TokenCut, with UnionSeg, performs best among no-learning approaches. And the performance of FOUND, the previous state-of-the-art method, is also increased by UnionCut and UnionSeg on most benchmarks and metrics.", + "bbox": [ + 511, + 388, + 903, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Self-supervised Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 486, + 844, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Some UOD methods target self-supervised instance segmentation, where models are trained using pseudo-annotations given by UOD algorithms, such as MaskCut for CutLER and FreeMask for FreeSOLO [67]. In the experiment, we explore whether UnionSeg improves CutLER's performance on instance segmentation. Similar to [69], COCO20k [37, 61] and COCO [37](5000 images from COCO2017's validation set) are used for evaluation, with average precision (AP) as the metric. Tab. 3 illustrates that by using UnionSeg to improve the pseudo-labels given by MaskCut, CutLER performs better on most metrics.", + "bbox": [ + 511, + 508, + 905, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since UnionCut with MaskCut takes much time to generate the foreground union and UOD results of all images in ImageNet (7 weeks in total, with 4 weeks for UnionCut and 3 weeks for MaskCut with 18 subprocesses running in parallel limited by our available computational resources), we only consider UnionSeg in this section. However, the effectiveness of UnionCut on self-supervised instance segmentation is also of interest. Therefore, we conduct another experiment that trains CutLER on a much smaller dataset (the subset of VOC12 containing 2913 images with pixel-level annotations). Details can be found in Appendix 13.", + "bbox": [ + 511, + 674, + 905, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Properties of UnionCut and UnionSeg", + "text_level": 1, + "bbox": [ + 511, + 848, + 841, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We explore the properties of UnionCut and UnionSeg. First, their accuracy in foreground union detection is evaluated.", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20273", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/c6bda0a9d5e27c7d32a84618f57f85d60a368695e21cb015d6aad2b036d211fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodDUT-OMRONDUTS-TEECSSD
Acc.IoUmaxFβAcc.IoUmaxFβAcc.IoUmaxFβ
- No learning -
HS [73]84.343.356.182.636.950.484.750.867.3
wCtr [77]83.841.654.183.539.252.286.251.768.4
WSC [36]86.538.752.386.238.452.885.249.868.3
DeepUSPS [41]77.930.541.477.330.542.579.544.058.4
BigBiGAN [63]85.645.354.987.849.860.889.967.278.2
E-BigBiGAN [63]86.046.456.388.251.162.490.668.479.7
LOST [53]79.741.047.387.151.861.189.565.475.8
DSS [40]-56.7--51.4--73.3-
MOST [45]8747.557.089.753.866.689.063.179.1
TokenCut [21, 54](ViT-S/16 [9])89.761.869.791.462.475.593.477.287.4
TokenCut(ViT-S/16 [9])+UnionCut89.662.4(0.6↑)71.0(1.3↑)91.6(0.2↑)63.2(0.8↑)80.2(4.7↑)93.8(0.4↑)77.3(0.1↑)91.0(3.6↑)
TokenCut(ViT-S/16 [9])+UnionSeg90.6(0.9↑)62.1(0.3↑)71.2(1.5↑)92.3(0.9↑)63.4(1.0↑)80.6(5.1↑)94.1(0.7↑)77.291.0(3.6↑)
- With learning -
LOST [53]81.848.957.888.757.269.791.672.383.7
SelfMask [52]91.965.5-93.366.0-95.581.8-
UnSeGArmaNet [46]----60.3--77.8-
CutLER [69](ViT-B/8 [9])†74.343.848.178.146.554.287.269.376.3
CutLER(ViT-B/8 [9])+UnionSeg87.7(13.4↑)57.8(14.0↑)64.1(16.0↑)90.1(12.0↑)61.4(14.9↑)72.3(18.1↑)94.0(6.8↑)79.6(10.3↑)87.8(11.5↑)
FOUND [54](ViT-S/16 [9])92.261.370.894.266.376.395.181.393.5
FOUND(ViT-S/16 [9])‡91.860.869.494.066.075.495.081.093.0
FOUND(ViT-S/16 [9])+UnionCut92.4(0.6↑)61.8(1.0↑)71.2(1.8↑)94.4(0.4↑)66.7(0.7↑)76.7(1.3↑)94.980.693.3(0.3↑)
FOUND(ViT-S/16 [9])+UnionSeg92.5(0.7↑)61.7(0.9↑)71.7(2.3↑)94.4(0.4↑)66.6(0.6↑)77.1(1.7↑)94.780.293.4(0.4↑)
", + "bbox": [ + 89, + 88, + 911, + 359 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/08de503d4bfad9deadc3f7c7396e34934edaa6d52dad5e6393c4ef40d850da99.jpg", + "table_caption": [ + "Table 2. Performances of UOD methods on saliency detection with red indicates the highest score. †: results of official codes and checkpoints by our measurement. ‡: results of our replication with official implementation." + ], + "table_footnote": [], + "table_body": "
MethodsPretrainDetectorInit.COCO20KCOCO val2017
\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)
LOST [53]IN+COCOFRCNN [47]DINO---2.41.01.1------
MaskDistill [59]IN+COCOMRCNN [27]MoCo [28]---6.82.12.9------
FreeSOLO [67, 69]IN+COCOSOLOv2 [65]DenseCL [66]9.73.24.19.73.44.39.63.14.29.43.34.3
DETReg [1]INDDETR [78]SwAV [8]------3.10.61.08.81.93.3
DINO [9]IN-DINO1.70.10.3---------
ProMerge [35]IN-DINO-----3.0-----2.4
ProMerge+ [35]INCascade [7]DINO-----8.9-----9.0
CutLER [69]INCascade [7]DINO22.412.511.919.610.09.221.911.812.318.99.79.2
CutLER+UnionSegINCascade [7]DINO24.112.212.920.39.110.023.712.212.919.79.09.9
", + "bbox": [ + 91, + 410, + 908, + 542 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Performances of self-supervised instance segmentation. The dataset (Pretrain) used to train each model (Detector) and each model's initialization (Init) are also listed. IN denotes ImageNet.", + "bbox": [ + 89, + 551, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Baseline Li et al. [35] proposes ProMerge, which contains a foreground union detection method. FOUND [54] is a UOD method; however, considering that it includes an algorithm for detecting background regions in an image, we extract this algorithm separately as a baseline and treat the inversion of its output as the foreground union detection result. For convenience, we still refer to their foreground union detection algorithms as ProMerge and FOUND.", + "bbox": [ + 89, + 606, + 482, + 727 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Benchmark The subset of VOC12 with 2913 images of pixel-level annotations is used as the benchmark, and the ground truth mask of each instance is combined as the ground truth of the foreground union in an image.", + "bbox": [ + 89, + 729, + 482, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metrics IoU, precision, and recall are used to measure the performance. Besides, since representative datasets [22, 37] label objects of a given class list, there may be cases where some objects in the image are not labelled and thus are not considered part of the foreground union by the ground truth (Examples can be seen in Appendix 14). Therefore, directly using these metrics and dataset-provided", + "bbox": [ + 89, + 794, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "labels to measure the performance of foreground union detection is inaccurate. To address this, we propose a fuzzy evaluation metric similar to CorLoc, named CorUnion. Specifically, we compare the foreground union output by UnionCut or UnionSeg with the ground truth for a given image and compute a specific metric (e.g. IoU). If this metric exceeds a certain threshold, we consider the foreground union detection on that image successful. The success rate across the dataset is then used as the CorUnion value. This approach also allows us to study the properties of UnionCut and UnionSeg by using different metrics or adjusting the threshold for successful detection.", + "bbox": [ + 511, + 606, + 906, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Foreground Union Accuracy Tab. 4 shows the performance of UnionCut and UnionSeg, measured by precision, recall, and IoU. We also report CorUnion using IoU as the metric to judge an image's foreground union detection success, with increasing IoU threshold from 0.5 to 0.9. As shown in Fig. 4, both UnionCut and UnionSeg are more accurate than the baselines at foreground union detection.", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20274", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/6fad0d611a35740c56412ae32614045610d9f9b0f4aa6722cf113fed6a838cf6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPrecisionRecallIoU
FOUND69.281.157.9
ProMerge-73.4-78.4-59.9
UnionCut-70.7-84.6-60.9
UnionSeg80.978.665.7
", + "bbox": [ + 125, + 88, + 450, + 150 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/063578444ed6b0d4ebf754a8379dfd52974d81605a6a7404dd16912c2677489d.jpg", + "image_caption": [ + "Figure 4. Foreground union detection measured by CorUnion." + ], + "image_footnote": [], + "bbox": [ + 122, + 189, + 450, + 276 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Instances Count In addition, we assessed UnionCut and UnionSeg's performance across varying instance counts in VOC12, categorized into groups based on the number of instances per image. As shown in Fig. 5a, UnionSeg, a training-based method, outperforms training-free approaches in precision, supporting that self-training corrects initial errors made by training-free methods. This finding aligns with previous research [21, 53, 54]. As instance numbers increase, UnionSeg maintains its precision, while other methods decline. Furthermore, Fig. 5b shows that UnionCut surpasses other methods in recall of the foreground union, remaining stable as the number of instances rises, unlike the decrease seen with other methods. This demonstrates the functional distinction between UnionCut and other training-free methods: as the number of instances per image increases, UnionCut consistently predicts the foreground union covering most instances, though precision inevitably decreases. This consistency highlights UnionCut's effectiveness as a robust prior for UOD algorithms to cease exploration optimally.", + "bbox": [ + 88, + 329, + 482, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Overall, UnionSeg outperforms UnionCut and other baseline methods (Fig. 4) with higher precision (Fig. 5a), making it better suited for identifying if a detected area is foreground. UnionCut, with its higher recall (Fig. 5b), is better for determining if the discovery is complete.", + "bbox": [ + 89, + 632, + 482, + 708 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e5632f2f1a552857653fe6a8a559047ee5c9cadffe68d96fc7f51687ef9e88b4.jpg", + "image_caption": [ + "Figure 5. Performance of foreground union detection on images with varying instance numbers, measured by CorUnion with precision $>0.8$ (a) and recall $>0.8$ (b) as success criteria." + ], + "image_footnote": [], + "bbox": [ + 96, + 720, + 478, + 843 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 90, + 666, + 107 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "UnionCut consists of two elements: aggregated UVs and corner prior, while UnionSeg builds on UnionCut with distillation. To evaluate each element's contribution, TokenCut [21] is used as the baseline, with each element integrated for single object discovery. Benchmarks and metrics are inherited from Sec. 4.1. As shown in Tab. 5, each element contributes to the performance increase of UOD, demonstrating their effectiveness.", + "bbox": [ + 511, + 112, + 906, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/11b88d7a025da918a8a4e142630aec93e5382382d64722dac29127a14b09493d.jpg", + "table_caption": [ + "Table 4. Average performance on foreground union detection." + ], + "table_footnote": [], + "table_body": "
MethodaUcpdVOC07VOC12COCO20K
TokenCut [21]68.872.158.8
TokenCut+aU✓69.072.362.0
TokenCut+UnionCut✓✓69.272.362.1
TokenCut+UnionSeg✓✓✓69.772.762.6
", + "bbox": [ + 527, + 247, + 893, + 323 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further evaluate the inference speed of UnionCut and UnionSeg. We include FOUND and ProMerge in Sec. 4.4 as the baseline. Our implementation is based on Python 3.11.9, Pytorch 2.3.1 and CUDA 12.1. As shown in Tab. 6, UnionSeg significantly increases the efficiency of foreground union detection compared with UnionCut.", + "bbox": [ + 511, + 392, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2a8206581eb91e6f9abc327bc23eaa5d24f66e5ae18f2dd226ac273a529d1f12.jpg", + "table_caption": [ + "Table 5. Contribution of each component of UnionCut and UnionSeg. aU, cp, and d denote aggregated UVs, corner prior and distillation, respectively." + ], + "table_footnote": [], + "table_body": "
MethodCPUGPUFPS
FOUNDIntel 14700KFNVIDIA RTX4070Ti Super2
ProMergeIntel 14700KFNVIDIA RTX4070Ti Super3
UnionCutIntel 14700KFNVIDIA RTX4070Ti Super0.1
UnionSegIntel 14700KFNVIDIA RTX4070Ti Super125
", + "bbox": [ + 526, + 496, + 893, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Inference speed of UnionSeg and UnionCut. FPS (frames per second) is evaluated throughout the whole pipeline from the image input to the output of the foreground union.", + "bbox": [ + 511, + 578, + 906, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 652, + 633, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces UnionCut, an unsupervised ensemble method to robustly predict foreground unions in images. Additionally, we propose UnionSeg distilled from UnionCut to detect foreground unions more accurately and efficiently. Both of them help UOD algorithms determine if discovered areas are in the foreground and when to stop discovery on an image, enhancing UOD's accuracy. While both are effective, each has unique advantages, and the choice between them depends on the specific UOD requirements for precision, recall, and efficiency. Our experiments show that integrating UnionCut or UnionSeg improves the performance of state-of-the-art UOD methods in single object discovery, saliency detection, and instance segmentation, demonstrating their effectiveness as robust foreground priors for unsupervised object discovery.", + "bbox": [ + 511, + 676, + 908, + 888 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20275", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Amir Bar, Xin Wang, Vadim Kantorov, Colorado J Reed, Roei Herzig, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Detreg: Unsupervised pretraining with region priors for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14605-14615, 2022. 7", + "[2] Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE transactions on pattern analysis and machine intelligence, 26(9):1124-1137, 2004. 1, 4", + "[3] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 4", + "[4] Leo Breiman. Bagging predictors. Machine learning, 24: 123-140, 1996. 3", + "[5] Leo Breiman. Pasting small votes for classification in large databases and on-line. Machine learning, 36:85-103, 1999. 3", + "[6] Leo Breiman. Random forests. Machine learning, 45:5-32, 2001. 3", + "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7, 6", + "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 7", + "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6, 7", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020.", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2", + "[12] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 1", + "[13] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969–978, 2022. 1", + "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1", + "[15] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning from video to detect foreground objects in single images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4335-4343, 2017. 1", + "[16] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning of foreground object segmentation. International Journal of Computer Vision, 127:1279-1302, 2019. 1", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6", + "[18] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 1, 3, 4", + "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 5, 6", + "[20] Rosana El Jurdi, Caroline Petitjean, Paul Honeine, and Fahed Abdallah. Bb-unet: U-net with bounding box prior. IEEE Journal of Selected Topics in Signal Processing, 14(6):1189-1198, 2020. 1", + "[21] Y. Wang et al. Self-supervised transformers for unsupervised object discovery using normalized cut. In CVPR, 2022. 1, 2, 5, 6, 7, 8", + "[22] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 7", + "[23] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1): 119-139, 1997. 3", + "[24] Keinosuke Fukunaga and Larry Hostetler. The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Transactions on information theory, 21(1):32-40, 1975. 5", + "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2", + "[26] Emanuela Haller, Adina Magda Florea, and Marius Leordeanu. Iterative knowledge exchange between deep learning and space-time spectral clustering for unsupervised segmentation in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7638-7656, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20276", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7", + "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 7", + "[29] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. 1", + "[30] Gunhee Kim and Antonio Torralba. Unsupervised detection of regions of interest using iterative link analysis. Advances in neural information processing systems, 22, 2009. 6, 7", + "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7", + "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1", + "[33] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In European Conference on Computer Vision, pages 290-308. Springer, 2020. 1", + "[34] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1", + "[35] Dylan Li and Gyungin Shin. Promerge: Prompt and merge for unsupervised instance segmentation. In ECCV, 2024. 1, 2, 7", + "[36] Nianyi Li, Bilin Sun, and Jingyi Yu. A weighted sparse coding framework for saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5223, 2015. 7", + "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 1, 2, 3, 4", + "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[39] R Austin McEver and BS Manjunath. Pcams: Weakly supervised semantic segmentation using point supervision. arXiv preprint arXiv:2007.05615, 2020. 1", + "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 7", + "[42] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 2", + "[43] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 1", + "[44] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8843-8850, 2019. 1", + "[45] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. Most: Multiple object localization with self-supervised transformers for object discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15823-15834, 2023. 6, 7", + "[46] Kovvuri Sai Gopal Reddy, Bodduluri Saran, A. Mudit Adityaja, Saurabh J. Shigwan, Nitin Kumar, and Snehasis Mukherjee. Unsegarmanet: Unsupervised image segmentation using graph neural networks with convolutionalarma filters, 2024.7", + "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 7", + "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1", + "[49] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 4", + "[50] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2", + "[51] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4): 717-729, 2015. 6, 1, 2, 3, 4", + "[52] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6, 7" + ], + "bbox": [ + 514, + 90, + 903, + 897 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20277", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC 2021-32nd British Machine Vision Conference, 2021. 1, 2, 4, 5, 6, 7, 8, 3", + "[54] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3176-3186, 2023. 1, 2, 5, 6, 7, 8, 3, 4", + "[55] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017. 1", + "[56] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1", + "[57] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1", + "[58] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104: 154-171, 2013. 6, 7", + "[59] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022. 1, 2, 7, 3", + "[60] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, pages I-I. IEEE, 2001. 3", + "[61] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 779-795. Springer, 2020. 5, 6, 7", + "[62] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 6, 7", + "[63] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6, 7", + "[64] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 136-145, 2017. 5, 6, 1, 2, 3, 4", + "[65] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmenta" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tion. Advances in Neural information processing systems, 33:17721-17732, 2020. 7", + "[66] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 1, 2, 7", + "[67] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 1, 2, 6, 7", + "[68] Xuehui Wang, Kai Zhao, Ruixin Zhang, Shouhong Ding, Yan Wang, and Wei Shen. Contrastmask: Contrastive learning to segment every thing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11604-11613, 2022. 1", + "[69] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 5, 6, 7, 8", + "[70] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 6, 7", + "[71] Yulin Wu, Chang Liu, Lei Chen, Dong Zhao, Qinghe Zheng, and Hongchao Zhou. Perturbation consistency and mutual information regularization for semi-supervised semantic segmentation. Multimedia Systems, 29(2):511-523, 2023. 1", + "[72] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, Ferdous Sohel, and Dan Xu. Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6984-6993, 2021. 1", + "[73] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 7", + "[74] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3166-3173, 2013. 6, 4", + "[75] Runsheng Zhang, Yaping Huang, Mengyang Pu, Jian Zhang, Qingji Guan, Qi Zou, and Haibin Ling. Object discovery from a single unlabeled image by mining frequent itemsets with multi-scale features. IEEE Transactions on Image Processing, 29:8606-8621, 2020. 6, 7", + "[76] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021. 2", + "[77] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20278", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 7", + "[79] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 1", + "[80] C Lawrence Zitnick and Piotr Dólar. Edge boxes: Locating object proposals from edges. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 391-405. Springer, 2014. 6, 7" + ], + "bbox": [ + 91, + 92, + 480, + 287 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "20279", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_model.json b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0eb3c6ea420910340149c45287dd985da987b87c --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_model.json @@ -0,0 +1,2435 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.867, + 0.154 + ], + "angle": 0, + "content": "Ensemble Foreground Management for Unsupervised Object Discovery*" + }, + { + "type": "text", + "bbox": [ + 0.278, + 0.181, + 0.718, + 0.217 + ], + "angle": 0, + "content": "Ziling Wu,\\(^{\\dagger}\\) Armaghan Moemeni, Praminda Caleb-Solly University of Nottingham, UK" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.22, + 0.796, + 0.234 + ], + "angle": 0, + "content": "{ziling.wu, armaghan.moemeni, praminda.caleb-solly}@nottingham.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.302, + 0.485, + 0.739 + ], + "angle": 0, + "content": "Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted annotations. Recent progress in self-supervised representation learning [9, 66] has led to some success in UOD algorithms [21, 53, 69]. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or background, and 2) knowing how many objects remain undiscovered. To address these two problems, previous solutions rely on foreground priors [53, 59, 67, 69] to distinguish if the discovered region is foreground, and conduct one or fixed iterations of discovery. However, the existing foreground priors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper introduces UnionCut, a robust and well-grounded foreground prior based on min-cut [2] and ensemble methods [18] that detects the union of foreground areas of an image, allowing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the foreground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.77, + 0.222, + 0.787 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.797, + 0.484, + 0.858 + ], + "angle": 0, + "content": "Instance segmentation is a fundamental computer vision task requiring the algorithm to recognize and segment each instance in an image, which has been widely applied in autonomous driving [14], medical image analysis [48], re" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.27, + 0.905, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.418, + 0.907, + 0.447 + ], + "angle": 0, + "content": "Figure 1. Examples of fixing MaskCut [69]'s errors and making its discovery stop at appropriate time using UnionCut and UnionSeg." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.597 + ], + "angle": 0, + "content": "mote sensing [79], etc. However, pixel-level annotations are required to train an instance segmentation model, which are expensive and time-consuming to obtain. To cut down the training cost of instance segmentation models, strategies employing semi-supervision [12, 20, 29, 32, 43, 55, 56, 71], weak supervision [13, 33, 34, 39, 44, 57, 68, 72], and self-supervision [21, 35, 53, 67, 69] have been proposed to reduce the usage of pixel-level annotations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.6, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Self-supervision aims to train an instance segmentation model without any annotations. It utilizes an unsupervised object discovery (UOD) algorithm to detect and segment objects in images without handcrafted annotations, following which, the detection result is used as the ground truth to train a segmentation model [15, 16, 21, 26, 35, 67]. Recent state-of-the-art UOD algorithms [21, 35, 53, 67, 69] are based on a vision transformer (ViT) [19] pre-trained by self-supervised representation learning [9]. These draw on a phenomenon where the attention map of the last layer of the ViT indicates regions of objects in the image. However, without the foreground knowledge given by the ground truth, existing UOD methods cannot robustly judge whether the explored region belongs to the foreground or not, and as such, the algorithm could return the background area of the image as a discovered object. As a remedy, existing approaches introduce foreground priors (e.g. the location and size of the discovered region) to help them distinguish if an area in the image is foreground. However, we have found that existing foreground priors are not robust because they" + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.87, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\*Author's Accepted Manuscript, Released under the Creative Commons license: Attribution 4.0 International (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ \\(\\dagger\\) Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20268" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "are based on simple rules, assumptions, and heuristic design. Besides, the number of objects in the image varies and is unknown to UOD approaches, resulting in a dilemma of when to stop the discovery. Thus, existing UOD algorithms either segment only one region as the object or continue discovery for a predetermined number of iterations, leading to either missing objects or mistakenly identifying background as foreground due to unreliable foreground priors and excessive exploration. For example, in Fig. 1, MaskCut [69], a UOD method, returns a background area (the 1st and 3rd row), and misses the car on the left of the 2nd image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.259, + 0.482, + 0.456 + ], + "angle": 0, + "content": "To address the problem, we propose an ensemble method named UnionCut. It detects the union of foreground regions in an image, referred to as \"foreground union\" in this paper. With this foreground union as the foreground prior, UOD methods can accurately judge if a discovered region is the foreground by checking if the foreground union covers it, and stop further exploration when the majority of the foreground union has been discovered, as shown in Fig. 1. Unlike previous assumption-based foreground priors, UnionCut's robustness and reliability are demonstrated with mathematical and statistical ensemble theories. Besides, we propose UnionSeg, a ViT distilled from UnionCut to increase the efficiency and accuracy of foreground union detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.789 + ], + "angle": 0, + "content": "Our work makes the following contributions: Insights: Previous UOD studies focused mainly on the design of UOD algorithms, overlooking foreground priors' importance. We reveal that UOD's performance can also be increased by robust foreground priors. Methods: We propose UnionCut based on ensemble and UnionSeg using distillation to detect foreground union in the image, serving as foreground priors for UOD methods to verify results and decide when to stop discovery. Unlike previous heuristic and assumption-based foreground priors, UnionCut is mathematically and statistically robust and reliable. Results: After being integrated with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks, indicating that UnionCut and UnionSeg enable UOD algorithms to give more accurate predictions on the existence of objects in the image. Impacts: UnionCut and UnionSeg can be used as future UOD methods' default foreground priors, taking the place of existing assumption-based, heuristic, and unreliable foreground priors, and enhancing UOD performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.806, + 0.233, + 0.822 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.441, + 0.85 + ], + "angle": 0, + "content": "2.1. Self-supervised Representation Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Self-supervised representation learning aims to train the model to generate distinguishable embeddings for images and output similar embeddings of images with the same se" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "mantic information without using handcrafted annotations. Most existing approaches [8-11, 25, 28, 42, 66, 76] use a student-teacher architecture, where each model processes one of two differently augmented versions (e.g. horizontal flipping, randomly cropping, colour jitter, etc.) of the same image. The optimization goal is to make the models produce similar outputs for augmentations of the same image and distinct outputs for augmentations from different images. DINO [9] is a representative method of these approaches, where the attention map of the last layer of a ViT pre-trained by DINO indicates the location of objects in the image. Such a property is employed in our approach." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.296, + 0.792, + 0.312 + ], + "angle": 0, + "content": "2.2. Unsupervised Object Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.905, + 0.565 + ], + "angle": 0, + "content": "As for UOD, Wang et al. [67] proposed FreeMask to discover objects by comparing the similarity between features from a self-supervised CNN [66] at each pixel of the image. Instead of using CNN, in [21, 53, 59, 69], embeddings of patches output by a ViT pre-trained by DINO[9] are utilized to generate pseudo-masks. In [53, 59], a patch is selected as the seed, and the pseudo mask is given by combining patches sharing similar features with the seed. To increase the pseudo-mask quality, Normalized Cut (NCut) [50] is leveraged by Wang et al. [21] to segment the feature map output by DINO for pseudo-mask generation. In [69], NCut is expanded to segment multiple objects in an image by iteratively removing a new segmented region from the image. However, the absence of ground truth makes it hard to determine if any discovered region is actually part of the foreground. Our approach provides a solution to this." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.694, + 0.603 + ], + "angle": 0, + "content": "2.3. Foreground Priors" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.901 + ], + "angle": 0, + "content": "UOD algorithms leverage foreground priors to judge if the discovered region belongs to the foreground. Specifically, Simeoni et al. [53] assumes that the area occupied by the foreground should be smaller than the background. FreeMask [67] removes discovered regions whose widths are longer than \\(95\\%\\) of the width of the image or are located in the top \\(20\\%\\) region of the image. Gansbeke et al. [59] and Simeoni et al. [54] proposed that the semantic information of the foreground should be close to the global semantic information of the entire image. In [21, 69], the region where the point with the highest absolute confidence value is located is made as the foreground. Wang et al. [69] suggested that the foreground is unlikely to cover three or four corners of the image. Li and Shin [35] proposed that a background mask is likely to occupy at least two edges of the image. Although these priors achieved limited success, they are all heuristic and assumption-based. In contrast, the robustness and reliability of our proposed methods are demonstrated by mathematical and statistical ensemble theories." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20269" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.127, + 0.09, + 0.873, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.291, + 0.908, + 0.351 + ], + "angle": 0, + "content": "Figure 2. Framework of UnionCut. UnionCut takes an image as input with features extracted from a frozen DINO, and applies Unit Voters (UVs) with each using one patch as its seed. Red rectangles refer to the seed patch, and blue areas indicate the UV's output of the region similar to its seed patch. Aggregating all UV's outputs yields a heat map of the background, which is then inverted, thresholded, and rectified with a corner prior to produce a binary mask of the image's foreground union." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.374, + 0.276, + 0.388 + ], + "angle": 0, + "content": "2.4. Ensemble Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.396, + 0.484, + 0.609 + ], + "angle": 0, + "content": "Ensemble methods are learning algorithms that construct a set of classifiers and then classify new data points by taking a (weighted) vote of their predictions [18]. Most ensemble methods can be classified into two types: 1) Parallel a set of weak classifiers run in parallel and form a strong classifier by performing a plurality vote, e.g. Pasting [5], Bagging [4], and Random Forest [6]; and 2) Sequential running batches of weak classifiers iteratively with each batch focusing on hard cases of the last one, e.g. [60] and AdaBoost [23]. UnionCut is of the former: a set of weak classifiers is created for an image, with each in parallel voting every patch \\((8 \\times 8\\) pixels) of the image as foreground or background, which forms a strong classifier by aggregating votes from them to detect the foreground union of an image." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.62, + 0.226, + 0.637 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.795 + ], + "angle": 0, + "content": "UnionCut is an ensemble method, enabling UOD methods to judge if a detected region is foreground and stop discovery when most objects have been discovered to avoid over-segmentation. In the context of ensemble methods, UnionCut serves as the strong classifier, based on many parallel weak classifiers named \"unit voter\" (UV) here. Fig. 2 illustrates the frameworks of the proposed UnionCut. Unlike previous assumption-based foreground priors mentioned in Sec. 2.3, UnionCut is reliable and robust with ensemble theories. First, we introduce the details of UV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.21, + 0.818 + ], + "angle": 0, + "content": "3.1. Unit Voter" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "UV is based on features output by a ViT pretrained by DINO [9]. In our implementation, an image is resized into \\(224 \\times 224\\) and serialized into \\(28 \\times 28\\) (i.e. 784) patches of \\(8 \\times 8\\) pixels before feature extraction of each patch with the ViT. The set of patches of an image is defined as \\(P =\\)" + }, + { + "type": "image", + "bbox": [ + 0.599, + 0.373, + 0.831, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Figure 3. Overview of Unit Voter (UV). For clarity, the image is divided into \\(3 \\times 3\\) patches. With the seed patch and anti-seed patches of the UV initialized, the image is modeled as a graph with each patch and two dummy terminals being nodes and connected by edges whose thickness reflects their weights. After applying min-cut to the graph, thin edges are removed to divide the graph into two disjoint parts containing one terminal node. The part with the Source node is the region output by the UV, indicating the area of features similar to the seed patch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.905, + 0.794 + ], + "angle": 0, + "content": "\\(\\{p_1,p_2,\\dots ,p_{784}\\}\\) with the corresponding L2-normalized Key vectors as their features output by the last layer of the model, which is defined as \\(K = \\{k_{1},k_{2},\\ldots ,k_{784}\\}\\). Then, the patches and features are leveraged by each UV." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As required by ensemble methods, the number of weak classifiers (i.e. UV) to form a strong classifier (i.e. Union-Cut) should be significant, and each UV should be different [18]. To achieve these goals, 784 independent UVs can be built based on the 784 patches of an image, with each UV corresponding to a unique patch \\( p_f \\in P \\). The patch corresponding to a UV is called that UV's \"seed patch\"." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20270" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.514 + ], + "angle": 0, + "content": "Taking a UV built from a seed patch \\( p_f \\in P \\) as an example, its pipeline is illustrated in Fig. 3. Specifically, given a UV's seed patch \\( p_f \\in P \\) with feature \\( k_f \\in K \\), the set of patches dissimilar to \\( p_f \\), defined as the UV's anti-seed patches set \\( B_f = \\{p_b | p_b \\in P, b \\neq f, k_b^T k_f < 0\\} \\), can be obtained. With \\( p_f, B_f \\), all \\( 28 \\times 28 \\) patches \\( P \\) and features \\( K \\), a directed graph is constructed, as shown in Fig. 3, where each patch is a node and two dummy terminal nodes, named \"Source\" and \"Target\", are added. In the graph, nodes of adjacent patches (i.e. eight-neighborhood patches) are connected by bidirectional edges (named \"n-links\") with weights representing their similarity. Directed edges named \"t-links\" connect the Source to each patch node or each patch node to the Target, with weights based on the likelihood of a patch being more similar to the UV's seed patch \\( p_f \\) or its anti-seed patches set \\( B_f \\). The terms \"t-link\" and \"n-link\" originate from early image segmentation methods that modeled images as graphs [3, 49]. The purpose of building this graph is to split the image by removing edges, forming two disconnected subgraphs with each either containing the Source or the Target, while minimizing the total weight of the removed edges. The subgraph with the Source is expected to contain patches similar to the UV's seed patch, while the one with the Target includes patches similar to the anti-seed patches \\( B_f \\). This bipartition process is also known as the min-cut problem, which is solved by UV using an efficient solution [2]. Next, we introduce the algorithms for assigning weights to n-links and t-links." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.547 + ], + "angle": 0, + "content": "n-links The weight \\( w(p_i, p_j) \\) of the n-link between two adjacent patches \\( p_i \\) and \\( p_j \\) is given by Eq. (1)." + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.565, + 0.482, + 0.607 + ], + "angle": 0, + "content": "\\[\nw \\left(p _ {i}, p _ {j}\\right) = \\frac {e ^ {- \\frac {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}{\\beta}}}{\\sqrt {\\left(i _ {x} - j _ {x}\\right) ^ {2} + \\left(i _ {y} - j _ {y}\\right) ^ {2}}} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.607, + 0.44, + 0.637 + ], + "angle": 0, + "content": "\\[\n\\beta = \\frac {2 \\sum_ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} (p _ {i} , p _ {j}) (k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}{C}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.719 + ], + "angle": 0, + "content": "where \\((i_x,i_y)\\) and \\((j_{x},j_{y})\\) are the indices of the two patches under the scale of the feature map (i.e. \\(28\\times 28\\)) output by the ViT. \\(\\mathbf{1}(p_i,p_j)\\) is an indicator function which returns 1 if the \\(p_i\\) and \\(p_j\\) are adjacent, otherwise 0. Finally, \\(C\\) is the number of n-links in the graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.871 + ], + "angle": 0, + "content": "t-links An UV's graph has three kinds of patches: seed patch \\( p_f \\), anti-seed patches set \\( B_f \\), and the other patches \\( P - B_f \\cup \\{p_f\\} \\), whose weight assignment strategies for t-links are different. Suppose the Source and Target are defined as \\( S \\) and \\( T \\). After min-cut, the t-link from \\( S \\) to \\( p_f \\) should be kept while the t-link between \\( p_f \\) and \\( T \\) should be cut. To this end, a large weight \\( W \\) is given to the t-link from \\( S \\) to \\( p_f \\) so that it is too large to be cut, and the weight between \\( p_f \\) and \\( T \\) is 0. Following early graph cut methods [3, 49], \\( W \\) can be decided by Eq. (2)." + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.886, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nW = 1 + \\max _ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} \\left(p _ {i}, p _ {j}\\right) w \\left(p _ {i}, p _ {j}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "Similarly, anti-seed patches must stay connected with \\(T\\) after min-cut. Therefore, t-links from \\(\\forall p_{b}\\in B_{f}\\) to \\(T\\) must have a large value, which can also be given by Eq. (2), while the weights of t-links from \\(S\\) to \\(\\forall p_{b}\\in B_{f}\\) are 0." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.152, + 0.907, + 0.184 + ], + "angle": 0, + "content": "Weights of other t-links between \\( S \\) or \\( T \\) and \\( p_i \\notin B_f \\bigcup \\{p_f\\} \\) are decided by Eq. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.523, + 0.201, + 0.905, + 0.24 + ], + "angle": 0, + "content": "\\[\nw \\left(p _ {i}, S\\right) = - \\log \\frac {e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}}}{e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}}} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.237, + 0.875, + 0.267 + ], + "angle": 0, + "content": "\\[\nw (p _ {i}, T) = - \\log \\frac {e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}{e ^ {\\sqrt {(k _ {i} - k _ {f}) ^ {T} (k _ {i} - k _ {f})}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.906, + 0.332 + ], + "angle": 0, + "content": "Eq. (3) explains that if a patch \\( p_i \\notin B_f \\cup \\{p_f\\} \\) is more similar to anti-seed patches set \\( B_f \\) than the seed patch \\( p_f \\), \\( w(p_i, T) \\) should exceed \\( w(p_i, S) \\), making the t-link from \\( S \\) to \\( p_i \\) more likely to be cut, and vice versa." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.424 + ], + "angle": 0, + "content": "With the graph initialized using the method introduced above, the UV executes a min-cut and returns a binary mask, where \"1\" indicates patches remaining connected to the Source, i.e. regions sharing features with the seed patch \\( p_f \\). Therefore, the function of each UV is to return the region that is similar to its seed patch." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.432, + 0.627, + 0.446 + ], + "angle": 0, + "content": "3.2. UnionCut" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.907, + 0.636 + ], + "angle": 0, + "content": "UnionCut is a strong classifier composed of 784 UVs as weak classifiers, with each UV using a unique patch from the image's \\(28 \\times 28\\) patches as its seed patch. Considering the function of UV introduced above, there are two types of UVs in UnionCut: 1) background UV: if its seed patch is background, the UV returns a binary mask indicating background regions; and 2) foreground UV: if its seed patch is foreground, the UV outputs a binary mask of foreground regions. UnionCut's output is determined by voting from the 784 UVs. Given the set of binary masks \\(M = \\{m_f | \\forall p_f \\in P\\}\\) returned by all UVs, a heat map \\(A\\) is generated by aggregating all UVs' output using Eq. (4)." + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.642, + 0.905, + 0.661 + ], + "angle": 0, + "content": "\\[\nA (x, y) = \\sum_ {m _ {f} \\in M} m _ {f} (x, y) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.906, + 0.697 + ], + "angle": 0, + "content": "where \\( x \\in [1, 28] \\) and \\( y \\in [1, 28] \\) indicate the 2D coordinate of a patch's response on \\( A \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.697, + 0.907, + 0.863 + ], + "angle": 0, + "content": "For easier understanding, intuitively, the foreground usually occupies a smaller area than the background in the image [53]. Thus, we can assume that there are more background UVs than foreground UVs in UnionCut, causing background areas to have higher intensities on the aggregated heat map \\( A \\). An example can be seen in Fig. 2. To make it intuitive, the brightness of \\( A \\) is inverted so that foreground regions have higher responses than the background, producing an inverted heat map \\( H \\) using Eq. (5). By thresholding \\( H \\), most objects in the image are segmented as the foreground union." + }, + { + "type": "equation", + "bbox": [ + 0.536, + 0.883, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\nH (x, y) = 2 5 5 - 2 5 5 \\times \\frac {A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)}{\\operatorname* {m a x} _ {x , y} A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)} \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20271" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.259 + ], + "angle": 0, + "content": "Note that the assumption introduced above is for easier understanding of our approach. UnionCut is mathematically and statistically robust, with an explanation in Appendix 6. Moreover, UnionCut remains effective even when the foreground area is larger than the background, with examples in Appendix 7. The mathematical explanation also supports this claim. Furthermore, compared to more straightforward methods detecting patches similar to a given patch, e.g. matching with cosine similarity [53, 54], UV is more sophisticated. Appendix 8 provides a detailed analysis of UV's irreplaceable advantages and necessity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.232, + 0.279 + ], + "angle": 0, + "content": "3.2.1. Thresholding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.419 + ], + "angle": 0, + "content": "A threshold for the bipartition of the heat map is required to estimate the foreground union. UnionCut automatically determines a threshold for the heat map \\( H \\) by clustering the response values on it with Mean-Shift [24]. Since the bright areas on the heat map represent the foreground, the resulting clusters are sorted by their centres in descending order, with the top half retained and identified as the union of the foreground in the image. Finally, a \\( 28 \\times 28 \\) binary mask \\( U_{cut} \\) with \"1\" as the foreground union is obtained." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.23, + 0.438 + ], + "angle": 0, + "content": "3.2.2. Corner Prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Similar to [69], a corner prior is used by UnionCut to check the reliability of the foreground union mask. Specifically, if \\( U_{cut} \\) occupies four corners of the image, it must be inverted (i.e. \\( U_{cut} := 1 - U_{cut} \\)) because the foreground union is unlikely to occupy four corners simultaneously. The reliability of this corner prior is discussed in Appendix 9." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.542, + 0.203, + 0.559 + ], + "angle": 0, + "content": "3.3. UnionSeg" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.822 + ], + "angle": 0, + "content": "While UnionCut effectively detects an image's foreground union, running 784 UVs per image is computationally expensive. To improve efficiency, we introduce UnionSeg, a ViT-based model distilled from UnionCut as a surrogate model. Unlike UnionCut, UnionSeg is an end-to-end model that directly predicts the foreground union from an input image. To minimize learnable parameters, inspired by FOUND [54], UnionSeg uses a frozen ViT trained by DINO [9] as its backbone, followed by a learnable \\(1 \\times 1\\) convolutional layer and a sigmoid function, compressing each ViT patch feature into a scalar confidence score. Patches with scores above 0.5 are classified as the foreground union. Appendix 10 details the differences between UnionSeg and FOUND. Suppose the soft output of UnionSeg of an image is \\(U_{seg}^{s}\\), with its thresholded hard output defined as \\(U_{seg}^{h}\\). Given UnionCut's output \\(U_{cut}\\) of an image, the label \\(L\\) of an image used to train UnionSeg is defined by Eq. (6)." + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.829, + 0.483, + 0.864 + ], + "angle": 0, + "content": "\\[\nL = \\left\\{ \\begin{array}{l l} U _ {c u t}, & \\text {i f I o U} \\left(U _ {s e g} ^ {h}, U _ {c u t}\\right) < 0. 5, \\\\ U _ {s e g} ^ {h}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\( IoU(\\cdot) \\) returns the intersection over union between two binary masks. The motivation behind Eq. (6) is that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "UnionCut's output may contain errors, so UnionSeg does not rely entirely on UnionCut's output as labels. This enables UnionSeg to predict the foreground union more efficiently and accurately than UnionCut. The loss function used to train UnionSeg is shown in Eq. (7)." + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.192, + 0.906, + 0.213 + ], + "angle": 0, + "content": "\\[\nL o s s = L \\log U _ {s e g} ^ {s} + (1 - L) \\log \\left(1 - U _ {s e g} ^ {s}\\right) + 1 (\\text {i t e r} < 1 0 0) \\Lambda \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.535, + 0.209, + 0.78, + 0.223 + ], + "angle": 0, + "content": "\\[\n\\Lambda = U _ {c u t} \\log U _ {s e g} ^ {s} + (1 - U _ {c u t}) \\log (1 - U _ {s e g} ^ {s})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.235, + 0.907, + 0.311 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}(iter < 100)\\) returns 1 if the training iterations are fewer than 100; otherwise 0. This ensures that at the start of training, when UnionSeg is not yet well trained, it is forced to use both UnionCut's output \\(U_{cut}\\) and \\(L\\) as labels to aid learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.326, + 0.646, + 0.344 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.352, + 0.907, + 0.503 + ], + "angle": 0, + "content": "Several experiments are conducted to evaluate the effectiveness of UnionCut/UnionSeg, comparing the performance of state-of-the-art UOD methods before and after integrating UnionCut or UnionSeg on several tasks: unsupervised single object discovery (Sec. 4.1), unsupervised saliency detection (Sec. 4.2), and self-supervised instance segmentation (Sec. 4.3). We then discuss the properties of UnionCut and UnionSeg in Sec. 4.4, and evaluate the impact of different components of them in Sec. 4.5. Additionally, we provide more qualitative comparisons in Appendix 15." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.504, + 0.906, + 0.61 + ], + "angle": 0, + "content": "Baseline UOD methods Our experiments use existing state-of-the-art UOD methods for single or multiple object discovery as baselines, i.e. LOST [53], TokenCut [21], FOUND [54] and CutLER [69]. To enhance their performance, we integrate them with UnionCut and UnionSeg by replacing their standard foreground priors with UnionCut/UnionSeg, with details provided in Appendix 11." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.906, + 0.777 + ], + "angle": 0, + "content": "Implementation Details We use a ViT-S/8 [19] pretrained by DINO [9] as the image feature extractor for UnionCut and UnionSeg. Following [52, 54], we distill UnionSeg from UnionCut using DUTS-TR[64], a saliency detection dataset with 10,553 images. Training is done with a batch size of 50, an initial learning rate of 0.05 that decays to \\(95\\%\\) every 50 iterations, and AdamW [38] for parameter updates over 600 iterations. Implementation details for all tested methods, i.e. LOST+UnionCut/UnionSeg, TokenCut+UnionCut/UnionSeg, FOUND+UnionCut/UnionSeg, and CutLER+UnionSeg, are provided in Appendix 11." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.787, + 0.843, + 0.804 + ], + "angle": 0, + "content": "4.1. Unsupervised Single Object Discovery" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.856 + ], + "angle": 0, + "content": "In this experiment, we combine baseline UOD methods with UnionCut and UnionSeg to see if their performance improves in unsupervised single object discovery." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Benchmark Following [21, 53, 54], the trainval sets of PASCAL VOC07 (5011 images), VOC12 (11540 images) [22] and COCO20K [37, 61] (19817 randomly sam" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20272" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.498, + 0.482 + ], + "angle": 0, + "content": "
MethodVOC07VOC12COCO20K
- No learning -
Selective Search [58]18.820.916.0
EdgeBoxes [80]31.131.628.8
Kim et al. [30]43.946.435.1
Zhang et al. [75]46.250.534.8
DDT+ [70]50.253.138.2
rOSD [61]54.555.348.5
LOD [62]53.655.148.5
DINO-seg [9, 53](ViT-S/16 [9])45.846.242.0
DSS [40](ViT-S/16 [9])62.766.452.2
LOST [53](ViT-S/16 [9])61.964.050.7
LOST(ViT-S/16 [9])+UnionCut62.4(0.5↑)65.3(1.3↑)51.9(1.2↑)
LOST(ViT-S/16 [9])+UnionSeg63.2(1.3↑)67.1(3.1↑)54.9(4.2↑)
TokenCut [21](ViT-S/16 [9])68.872.158.8
TokenCut(ViT-S/16 [9])+UnionCut69.2(0.4↑)72.3(0.2↑)62.1(3.3↑)
TokenCut(ViT-S/16 [9])+UnionSeg69.7(0.9↑)72.7(0.6↑)62.6(3.8↑)
- With learning -
FreeSOLO [67, 69]44.049.735.2
LOD+CAD [21, 53]56.361.652.7
rOSD+CAD [21, 53]58.362.353.0
LOST+CAD [53](ViT-S/16 [9])65.770.457.5
SelfMask [52]72.375.362.7
FOUND [54](ViT-S/16 [9])72.576.162.9
FOUND(ViT-S/16 [9])‡71.475.762.9
FOUND(ViT-S/16 [9])+UnionCut72.6(1.2↑)76.1(0.4↑)63.0(0.1↑)
FOUND(ViT-S/16 [9])+UnionSeg72.3(0.9↑)75.9(0.2↑)63.1(0.2↑)
CutLER [69](ViT-B/8 [9])†73.369.570.7
CutLER(ViT-B/8 [9])+UnionSeg73.8(0.5↑)71.2(1.7↑)72.4(1.7↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Table 1. Performances of UOD methods on single object discovery with red indicates the highest score. CAD: a class-agnostic detector trained with the output of no-learning UOD methods as the ground truth. †: results of official implementations and checkpoints by our measurement. ‡: results of our replication with official implementation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.639 + ], + "angle": 0, + "content": "pled images from COCO2014 [37]) are selected as the benchmark for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.642, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Metrics Similar to [21, 53, 54, 61, 62], we use CorLoc (Correct Localization) to evaluate UOD algorithms' performance on single object discovery. CorLoc scores an image as 1 if the intersection over union (IoU) between any ground truth bounding box and the algorithm's predicted bounding box exceeds 0.5, otherwise it scores 0. In our experiments, since CutLER can detect multiple objects, we select the bounding box with the highest confidence score as its single object prediction for the image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Tab. 1 shows that after integrating UnionCut or UnionSeg, our selected baseline UOD methods improve across all three benchmarks, achieving state-of-the-art results, indicating that UnionCut/UnionSeg as a robust foreground prior enhances UOD algorithms' performance in single object discovery. In Appendix 12, we also evaluate these UOD methods from the perspective of the performance upper boundary proposed by Rambhatla et al. [45]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.091, + 0.803, + 0.108 + ], + "angle": 0, + "content": "4.2. Unsupervised Saliency Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.905, + 0.188 + ], + "angle": 0, + "content": "We compare the performance of UOD methods before and after integrating UnionCut/UnionSeg on saliency detection, which involves segmenting salient objects in images. Since CutLER predicts multiple object masks per image, we use the union of these masks as its saliency detection result." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.188, + 0.905, + 0.248 + ], + "angle": 0, + "content": "Benchmark Similar to [21, 40, 53, 54, 63], ECSSD [51] (1000 real-world images), DUTS-TE [64] (5019 images sampled from ImageNet test), and DUT-OMRON [74] (5168 natural images) are chosen as the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.249, + 0.906, + 0.388 + ], + "angle": 0, + "content": "Metrics Following [21, 53, 54], we report results with three metrics: Accuracy, IoU, and \\(\\max F_{\\beta}\\). Accuracy measures the percentage of correctly classified pixels compared to ground truth. IoU is the ratio of the intersection between the UOD output and ground truth to their union. \\(\\max F_{\\beta}\\) is calculated as \\(\\frac{(1 + \\beta^2)\\text{Precision} \\times \\text{Recall}}{\\beta^2\\text{Precision} + \\text{Recall}}\\) where Precision is the percentage of predicted pixels matching ground truth, and Recall is the proportion of ground truth pixels covered by the prediction. We set \\(\\beta\\) to 0.3, as in [21, 53, 54]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.905, + 0.48 + ], + "angle": 0, + "content": "Tab. 2 illustrates that the selected baseline methods improve when integrated with UnionCut/UnionSeg. TokenCut, with UnionSeg, performs best among no-learning approaches. And the performance of FOUND, the previous state-of-the-art method, is also increased by UnionCut and UnionSeg on most benchmarks and metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.487, + 0.845, + 0.504 + ], + "angle": 0, + "content": "4.3. Self-supervised Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.906, + 0.675 + ], + "angle": 0, + "content": "Some UOD methods target self-supervised instance segmentation, where models are trained using pseudo-annotations given by UOD algorithms, such as MaskCut for CutLER and FreeMask for FreeSOLO [67]. In the experiment, we explore whether UnionSeg improves CutLER's performance on instance segmentation. Similar to [69], COCO20k [37, 61] and COCO [37](5000 images from COCO2017's validation set) are used for evaluation, with average precision (AP) as the metric. Tab. 3 illustrates that by using UnionSeg to improve the pseudo-labels given by MaskCut, CutLER performs better on most metrics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.906, + 0.842 + ], + "angle": 0, + "content": "Since UnionCut with MaskCut takes much time to generate the foreground union and UOD results of all images in ImageNet (7 weeks in total, with 4 weeks for UnionCut and 3 weeks for MaskCut with 18 subprocesses running in parallel limited by our available computational resources), we only consider UnionSeg in this section. However, the effectiveness of UnionCut on self-supervised instance segmentation is also of interest. Therefore, we conduct another experiment that trains CutLER on a much smaller dataset (the subset of VOC12 containing 2913 images with pixel-level annotations). Details can be found in Appendix 13." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.842, + 0.866 + ], + "angle": 0, + "content": "4.4. Properties of UnionCut and UnionSeg" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We explore the properties of UnionCut and UnionSeg. First, their accuracy in foreground union detection is evaluated." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20273" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.091, + 0.089, + 0.912, + 0.361 + ], + "angle": 0, + "content": "
MethodDUT-OMRONDUTS-TEECSSD
Acc.IoUmaxFβAcc.IoUmaxFβAcc.IoUmaxFβ
- No learning -
HS [73]84.343.356.182.636.950.484.750.867.3
wCtr [77]83.841.654.183.539.252.286.251.768.4
WSC [36]86.538.752.386.238.452.885.249.868.3
DeepUSPS [41]77.930.541.477.330.542.579.544.058.4
BigBiGAN [63]85.645.354.987.849.860.889.967.278.2
E-BigBiGAN [63]86.046.456.388.251.162.490.668.479.7
LOST [53]79.741.047.387.151.861.189.565.475.8
DSS [40]-56.7--51.4--73.3-
MOST [45]8747.557.089.753.866.689.063.179.1
TokenCut [21, 54](ViT-S/16 [9])89.761.869.791.462.475.593.477.287.4
TokenCut(ViT-S/16 [9])+UnionCut89.662.4(0.6↑)71.0(1.3↑)91.6(0.2↑)63.2(0.8↑)80.2(4.7↑)93.8(0.4↑)77.3(0.1↑)91.0(3.6↑)
TokenCut(ViT-S/16 [9])+UnionSeg90.6(0.9↑)62.1(0.3↑)71.2(1.5↑)92.3(0.9↑)63.4(1.0↑)80.6(5.1↑)94.1(0.7↑)77.291.0(3.6↑)
- With learning -
LOST [53]81.848.957.888.757.269.791.672.383.7
SelfMask [52]91.965.5-93.366.0-95.581.8-
UnSeGArmaNet [46]----60.3--77.8-
CutLER [69](ViT-B/8 [9])†74.343.848.178.146.554.287.269.376.3
CutLER(ViT-B/8 [9])+UnionSeg87.7(13.4↑)57.8(14.0↑)64.1(16.0↑)90.1(12.0↑)61.4(14.9↑)72.3(18.1↑)94.0(6.8↑)79.6(10.3↑)87.8(11.5↑)
FOUND [54](ViT-S/16 [9])92.261.370.894.266.376.395.181.393.5
FOUND(ViT-S/16 [9])‡91.860.869.494.066.075.495.081.093.0
FOUND(ViT-S/16 [9])+UnionCut92.4(0.6↑)61.8(1.0↑)71.2(1.8↑)94.4(0.4↑)66.7(0.7↑)76.7(1.3↑)94.980.693.3(0.3↑)
FOUND(ViT-S/16 [9])+UnionSeg92.5(0.7↑)61.7(0.9↑)71.7(2.3↑)94.4(0.4↑)66.6(0.6↑)77.1(1.7↑)94.780.293.4(0.4↑)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.37, + 0.908, + 0.399 + ], + "angle": 0, + "content": "Table 2. Performances of UOD methods on saliency detection with red indicates the highest score. †: results of official codes and checkpoints by our measurement. ‡: results of our replication with official implementation." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.411, + 0.909, + 0.543 + ], + "angle": 0, + "content": "
MethodsPretrainDetectorInit.COCO20KCOCO val2017
\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)
LOST [53]IN+COCOFRCNN [47]DINO---2.41.01.1------
MaskDistill [59]IN+COCOMRCNN [27]MoCo [28]---6.82.12.9------
FreeSOLO [67, 69]IN+COCOSOLOv2 [65]DenseCL [66]9.73.24.19.73.44.39.63.14.29.43.34.3
DETReg [1]INDDETR [78]SwAV [8]------3.10.61.08.81.93.3
DINO [9]IN-DINO1.70.10.3---------
ProMerge [35]IN-DINO-----3.0-----2.4
ProMerge+ [35]INCascade [7]DINO-----8.9-----9.0
CutLER [69]INCascade [7]DINO22.412.511.919.610.09.221.911.812.318.99.79.2
CutLER+UnionSegINCascade [7]DINO24.112.212.920.39.110.023.712.212.919.79.09.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.552, + 0.908, + 0.581 + ], + "angle": 0, + "content": "Table 3. Performances of self-supervised instance segmentation. The dataset (Pretrain) used to train each model (Detector) and each model's initialization (Init) are also listed. IN denotes ImageNet." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.728 + ], + "angle": 0, + "content": "Baseline Li et al. [35] proposes ProMerge, which contains a foreground union detection method. FOUND [54] is a UOD method; however, considering that it includes an algorithm for detecting background regions in an image, we extract this algorithm separately as a baseline and treat the inversion of its output as the foreground union detection result. For convenience, we still refer to their foreground union detection algorithms as ProMerge and FOUND." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.731, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Benchmark The subset of VOC12 with 2913 images of pixel-level annotations is used as the benchmark, and the ground truth mask of each instance is combined as the ground truth of the foreground union in an image." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Metrics IoU, precision, and recall are used to measure the performance. Besides, since representative datasets [22, 37] label objects of a given class list, there may be cases where some objects in the image are not labelled and thus are not considered part of the foreground union by the ground truth (Examples can be seen in Appendix 14). Therefore, directly using these metrics and dataset-provided" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.907, + 0.789 + ], + "angle": 0, + "content": "labels to measure the performance of foreground union detection is inaccurate. To address this, we propose a fuzzy evaluation metric similar to CorLoc, named CorUnion. Specifically, we compare the foreground union output by UnionCut or UnionSeg with the ground truth for a given image and compute a specific metric (e.g. IoU). If this metric exceeds a certain threshold, we consider the foreground union detection on that image successful. The success rate across the dataset is then used as the CorUnion value. This approach also allows us to study the properties of UnionCut and UnionSeg by using different metrics or adjusting the threshold for successful detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Foreground Union Accuracy Tab. 4 shows the performance of UnionCut and UnionSeg, measured by precision, recall, and IoU. We also report CorUnion using IoU as the metric to judge an image's foreground union detection success, with increasing IoU threshold from 0.5 to 0.9. As shown in Fig. 4, both UnionCut and UnionSeg are more accurate than the baselines at foreground union detection." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20274" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.126, + 0.089, + 0.451, + 0.151 + ], + "angle": 0, + "content": "
MethodPrecisionRecallIoU
FOUND69.281.157.9
ProMerge-73.4-78.4-59.9
UnionCut-70.7-84.6-60.9
UnionSeg80.978.665.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.161, + 0.482, + 0.177 + ], + "angle": 0, + "content": "Table 4. Average performance on foreground union detection." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.19, + 0.451, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.289, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Foreground union detection measured by CorUnion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.33, + 0.483, + 0.633 + ], + "angle": 0, + "content": "Instances Count In addition, we assessed UnionCut and UnionSeg's performance across varying instance counts in VOC12, categorized into groups based on the number of instances per image. As shown in Fig. 5a, UnionSeg, a training-based method, outperforms training-free approaches in precision, supporting that self-training corrects initial errors made by training-free methods. This finding aligns with previous research [21, 53, 54]. As instance numbers increase, UnionSeg maintains its precision, while other methods decline. Furthermore, Fig. 5b shows that UnionCut surpasses other methods in recall of the foreground union, remaining stable as the number of instances rises, unlike the decrease seen with other methods. This demonstrates the functional distinction between UnionCut and other training-free methods: as the number of instances per image increases, UnionCut consistently predicts the foreground union covering most instances, though precision inevitably decreases. This consistency highlights UnionCut's effectiveness as a robust prior for UOD algorithms to cease exploration optimally." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.709 + ], + "angle": 0, + "content": "Overall, UnionSeg outperforms UnionCut and other baseline methods (Fig. 4) with higher precision (Fig. 5a), making it better suited for identifying if a detected area is foreground. UnionCut, with its higher recall (Fig. 5b), is better for determining if the discovery is complete." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.721, + 0.48, + 0.844 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.897 + ], + "angle": 0, + "content": "Figure 5. Performance of foreground union detection on images with varying instance numbers, measured by CorUnion with precision \\(>0.8\\) (a) and recall \\(>0.8\\) (b) as success criteria." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.667, + 0.108 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.234 + ], + "angle": 0, + "content": "UnionCut consists of two elements: aggregated UVs and corner prior, while UnionSeg builds on UnionCut with distillation. To evaluate each element's contribution, TokenCut [21] is used as the baseline, with each element integrated for single object discovery. Benchmarks and metrics are inherited from Sec. 4.1. As shown in Tab. 5, each element contributes to the performance increase of UOD, demonstrating their effectiveness." + }, + { + "type": "table", + "bbox": [ + 0.529, + 0.248, + 0.895, + 0.324 + ], + "angle": 0, + "content": "
MethodaUcpdVOC07VOC12COCO20K
TokenCut [21]68.872.158.8
TokenCut+aU✓69.072.362.0
TokenCut+UnionCut✓✓69.272.362.1
TokenCut+UnionSeg✓✓✓69.772.762.6
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.334, + 0.907, + 0.377 + ], + "angle": 0, + "content": "Table 5. Contribution of each component of UnionCut and UnionSeg. aU, cp, and d denote aggregated UVs, corner prior and distillation, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.393, + 0.907, + 0.485 + ], + "angle": 0, + "content": "We further evaluate the inference speed of UnionCut and UnionSeg. We include FOUND and ProMerge in Sec. 4.4 as the baseline. Our implementation is based on Python 3.11.9, Pytorch 2.3.1 and CUDA 12.1. As shown in Tab. 6, UnionSeg significantly increases the efficiency of foreground union detection compared with UnionCut." + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.497, + 0.895, + 0.57 + ], + "angle": 0, + "content": "
MethodCPUGPUFPS
FOUNDIntel 14700KFNVIDIA RTX4070Ti Super2
ProMergeIntel 14700KFNVIDIA RTX4070Ti Super3
UnionCutIntel 14700KFNVIDIA RTX4070Ti Super0.1
UnionSegIntel 14700KFNVIDIA RTX4070Ti Super125
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.579, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Table 6. Inference speed of UnionSeg and UnionCut. FPS (frames per second) is evaluated throughout the whole pipeline from the image input to the output of the foreground union." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.654, + 0.634, + 0.669 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.909, + 0.89 + ], + "angle": 0, + "content": "This paper introduces UnionCut, an unsupervised ensemble method to robustly predict foreground unions in images. Additionally, we propose UnionSeg distilled from UnionCut to detect foreground unions more accurately and efficiently. Both of them help UOD algorithms determine if discovered areas are in the foreground and when to stop discovery on an image, enhancing UOD's accuracy. While both are effective, each has unique advantages, and the choice between them depends on the specific UOD requirements for precision, recall, and efficiency. Our experiments show that integrating UnionCut or UnionSeg improves the performance of state-of-the-art UOD methods in single object discovery, saliency detection, and instance segmentation, demonstrating their effectiveness as robust foreground priors for unsupervised object discovery." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20275" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.198 + ], + "angle": 0, + "content": "[1] Amir Bar, Xin Wang, Vadim Kantorov, Colorado J Reed, Roei Herzig, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Detreg: Unsupervised pretraining with region priors for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14605-14615, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE transactions on pattern analysis and machine intelligence, 26(9):1124-1137, 2004. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.256, + 0.483, + 0.323 + ], + "angle": 0, + "content": "[3] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.326, + 0.482, + 0.352 + ], + "angle": 0, + "content": "[4] Leo Breiman. Bagging predictors. Machine learning, 24: 123-140, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.354, + 0.483, + 0.394 + ], + "angle": 0, + "content": "[5] Leo Breiman. Pasting small votes for classification in large databases and on-line. Machine learning, 36:85-103, 1999. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.397, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[6] Leo Breiman. Random forests. Machine learning, 45:5-32, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.549 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.551, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.483, + 0.676 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.677, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[12] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[13] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969–978, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.148 + ], + "angle": 0, + "content": "Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[15] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning from video to detect foreground objects in single images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4335-4343, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.223, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[16] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning of foreground object segmentation. International Journal of Computer Vision, 127:1279-1302, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.336 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.38 + ], + "angle": 0, + "content": "[18] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[20] Rosana El Jurdi, Caroline Petitjean, Paul Honeine, and Fahed Abdallah. Bb-unet: U-net with bounding box prior. IEEE Journal of Selected Topics in Signal Processing, 14(6):1189-1198, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[21] Y. Wang et al. Self-supervised transformers for unsupervised object discovery using normalized cut. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[22] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[23] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1): 119-139, 1997. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.688, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[24] Keinosuke Fukunaga and Larry Hostetler. The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Transactions on information theory, 21(1):32-40, 1975. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.746, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Emanuela Haller, Adina Magda Florea, and Marius Leordeanu. Iterative knowledge exchange between deep learning and space-time spectral clustering for unsupervised segmentation in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7638-7656, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20276" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[27] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.203, + 0.482, + 0.259 + ], + "angle": 0, + "content": "[29] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[30] Gunhee Kim and Antonio Torralba. Unsupervised detection of regions of interest using iterative link analysis. Advances in neural information processing systems, 22, 2009. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.483 + ], + "angle": 0, + "content": "[33] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In European Conference on Computer Vision, pages 290-308. Springer, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[34] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.593 + ], + "angle": 0, + "content": "[35] Dylan Li and Gyungin Shin. Promerge: Prompt and merge for unsupervised instance segmentation. In ECCV, 2024. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[36] Nianyi Li, Bilin Sun, and Jingyi Yu. A weighted sparse coding framework for saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5223, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.651, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.817 + ], + "angle": 0, + "content": "[39] R Austin McEver and BS Manjunath. Pcams: Weakly supervised semantic segmentation using point supervision. arXiv preprint arXiv:2007.05615, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[41] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.163, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[42] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.248, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[43] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.319, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[44] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8843-8850, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[45] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. Most: Multiple object localization with self-supervised transformers for object discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15823-15834, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.461, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[46] Kovvuri Sai Gopal Reddy, Bodduluri Saran, A. Mudit Adityaja, Saurabh J. Shigwan, Nitin Kumar, and Snehasis Mukherjee. Unsegarmanet: Unsupervised image segmentation using graph neural networks with convolutionalarma filters, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.532, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.589, + 0.905, + 0.672 + ], + "angle": 0, + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.674, + 0.905, + 0.729 + ], + "angle": 0, + "content": "[49] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.731, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[50] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[51] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4): 717-729, 2015. 6, 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.831, + 0.905, + 0.898 + ], + "angle": 0, + "content": "[52] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.898 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20277" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[53] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC 2021-32nd British Machine Vision Conference, 2021. 1, 2, 4, 5, 6, 7, 8, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[54] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3176-3186, 2023. 1, 2, 5, 6, 7, 8, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.246, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[55] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[56] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[57] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[58] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104: 154-171, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[59] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022. 1, 2, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[60] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, pages I-I. IEEE, 2001. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[61] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 779-795. Springer, 2020. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[62] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[63] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.873 + ], + "angle": 0, + "content": "[64] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 136-145, 2017. 5, 6, 1, 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmenta" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "tion. Advances in Neural information processing systems, 33:17721-17732, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[66] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.19, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[67] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.273, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[68] Xuehui Wang, Kai Zhao, Ruixin Zhang, Shouhong Ding, Yan Wang, and Wei Shen. Contrastmask: Contrastive learning to segment every thing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11604-11613, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[69] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[70] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[71] Yulin Wu, Chang Liu, Lei Chen, Dong Zhao, Qinghe Zheng, and Hongchao Zhou. Perturbation consistency and mutual information regularization for semi-supervised semantic segmentation. Multimedia Systems, 29(2):511-523, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[72] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, Ferdous Sohel, and Dan Xu. Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6984-6993, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.649 + ], + "angle": 0, + "content": "[73] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[74] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3166-3173, 2013. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[75] Runsheng Zhang, Yaping Huang, Mengyang Pu, Jian Zhang, Qingji Guan, Qi Zou, and Haibin Ling. Object discovery from a single unlabeled image by mining frequent itemsets with multi-scale features. IEEE Transactions on Image Processing, 29:8606-8621, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[76] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[77] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20278" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[79] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[80] C Lawrence Zitnick and Piotr Dólar. Edge boxes: Locating object proposals from edges. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 391-405. Springer, 2014. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.288 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20279" + } + ] +] \ No newline at end of file diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_origin.pdf b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..8da492c95749e00241372dff90c0ca0394904550 --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/8437cd25-5fba-48bf-86af-e65bf6a1e6c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:bd11914a1a234cf2c9c2ca9c67e98a34237a5483ecb2202f44687ec15e0210cf +size 6836771 diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/full.md b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/full.md new file mode 100644 index 0000000000000000000000000000000000000000..c112f9dc6a0be17a18944d73e037eb2cd04ceed0 --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/full.md @@ -0,0 +1,333 @@ +# Ensemble Foreground Management for Unsupervised Object Discovery* + +Ziling Wu, $^{\dagger}$ Armaghan Moemeni, Praminda Caleb-Solly University of Nottingham, UK + +{ziling.wu, armaghan.moemeni, praminda.caleb-solly}@nottingham.ac.uk + +# Abstract + +Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted annotations. Recent progress in self-supervised representation learning [9, 66] has led to some success in UOD algorithms [21, 53, 69]. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or background, and 2) knowing how many objects remain undiscovered. To address these two problems, previous solutions rely on foreground priors [53, 59, 67, 69] to distinguish if the discovered region is foreground, and conduct one or fixed iterations of discovery. However, the existing foreground priors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper introduces UnionCut, a robust and well-grounded foreground prior based on min-cut [2] and ensemble methods [18] that detects the union of foreground areas of an image, allowing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the foreground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut. + +# 1. Introduction + +Instance segmentation is a fundamental computer vision task requiring the algorithm to recognize and segment each instance in an image, which has been widely applied in autonomous driving [14], medical image analysis [48], re + +![](images/a7256056cee5522836befa7ab7fdc71a9c2b46db91fa09351ac94329987b6e63.jpg) +Figure 1. Examples of fixing MaskCut [69]'s errors and making its discovery stop at appropriate time using UnionCut and UnionSeg. + +mote sensing [79], etc. However, pixel-level annotations are required to train an instance segmentation model, which are expensive and time-consuming to obtain. To cut down the training cost of instance segmentation models, strategies employing semi-supervision [12, 20, 29, 32, 43, 55, 56, 71], weak supervision [13, 33, 34, 39, 44, 57, 68, 72], and self-supervision [21, 35, 53, 67, 69] have been proposed to reduce the usage of pixel-level annotations. + +Self-supervision aims to train an instance segmentation model without any annotations. It utilizes an unsupervised object discovery (UOD) algorithm to detect and segment objects in images without handcrafted annotations, following which, the detection result is used as the ground truth to train a segmentation model [15, 16, 21, 26, 35, 67]. Recent state-of-the-art UOD algorithms [21, 35, 53, 67, 69] are based on a vision transformer (ViT) [19] pre-trained by self-supervised representation learning [9]. These draw on a phenomenon where the attention map of the last layer of the ViT indicates regions of objects in the image. However, without the foreground knowledge given by the ground truth, existing UOD methods cannot robustly judge whether the explored region belongs to the foreground or not, and as such, the algorithm could return the background area of the image as a discovered object. As a remedy, existing approaches introduce foreground priors (e.g. the location and size of the discovered region) to help them distinguish if an area in the image is foreground. However, we have found that existing foreground priors are not robust because they + +are based on simple rules, assumptions, and heuristic design. Besides, the number of objects in the image varies and is unknown to UOD approaches, resulting in a dilemma of when to stop the discovery. Thus, existing UOD algorithms either segment only one region as the object or continue discovery for a predetermined number of iterations, leading to either missing objects or mistakenly identifying background as foreground due to unreliable foreground priors and excessive exploration. For example, in Fig. 1, MaskCut [69], a UOD method, returns a background area (the 1st and 3rd row), and misses the car on the left of the 2nd image. + +To address the problem, we propose an ensemble method named UnionCut. It detects the union of foreground regions in an image, referred to as "foreground union" in this paper. With this foreground union as the foreground prior, UOD methods can accurately judge if a discovered region is the foreground by checking if the foreground union covers it, and stop further exploration when the majority of the foreground union has been discovered, as shown in Fig. 1. Unlike previous assumption-based foreground priors, UnionCut's robustness and reliability are demonstrated with mathematical and statistical ensemble theories. Besides, we propose UnionSeg, a ViT distilled from UnionCut to increase the efficiency and accuracy of foreground union detection. + +Our work makes the following contributions: Insights: Previous UOD studies focused mainly on the design of UOD algorithms, overlooking foreground priors' importance. We reveal that UOD's performance can also be increased by robust foreground priors. Methods: We propose UnionCut based on ensemble and UnionSeg using distillation to detect foreground union in the image, serving as foreground priors for UOD methods to verify results and decide when to stop discovery. Unlike previous heuristic and assumption-based foreground priors, UnionCut is mathematically and statistically robust and reliable. Results: After being integrated with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks, indicating that UnionCut and UnionSeg enable UOD algorithms to give more accurate predictions on the existence of objects in the image. Impacts: UnionCut and UnionSeg can be used as future UOD methods' default foreground priors, taking the place of existing assumption-based, heuristic, and unreliable foreground priors, and enhancing UOD performance. + +# 2. Related Work + +# 2.1. Self-supervised Representation Learning + +Self-supervised representation learning aims to train the model to generate distinguishable embeddings for images and output similar embeddings of images with the same se + +mantic information without using handcrafted annotations. Most existing approaches [8-11, 25, 28, 42, 66, 76] use a student-teacher architecture, where each model processes one of two differently augmented versions (e.g. horizontal flipping, randomly cropping, colour jitter, etc.) of the same image. The optimization goal is to make the models produce similar outputs for augmentations of the same image and distinct outputs for augmentations from different images. DINO [9] is a representative method of these approaches, where the attention map of the last layer of a ViT pre-trained by DINO indicates the location of objects in the image. Such a property is employed in our approach. + +# 2.2. Unsupervised Object Discovery + +As for UOD, Wang et al. [67] proposed FreeMask to discover objects by comparing the similarity between features from a self-supervised CNN [66] at each pixel of the image. Instead of using CNN, in [21, 53, 59, 69], embeddings of patches output by a ViT pre-trained by DINO[9] are utilized to generate pseudo-masks. In [53, 59], a patch is selected as the seed, and the pseudo mask is given by combining patches sharing similar features with the seed. To increase the pseudo-mask quality, Normalized Cut (NCut) [50] is leveraged by Wang et al. [21] to segment the feature map output by DINO for pseudo-mask generation. In [69], NCut is expanded to segment multiple objects in an image by iteratively removing a new segmented region from the image. However, the absence of ground truth makes it hard to determine if any discovered region is actually part of the foreground. Our approach provides a solution to this. + +# 2.3. Foreground Priors + +UOD algorithms leverage foreground priors to judge if the discovered region belongs to the foreground. Specifically, Simeoni et al. [53] assumes that the area occupied by the foreground should be smaller than the background. FreeMask [67] removes discovered regions whose widths are longer than $95\%$ of the width of the image or are located in the top $20\%$ region of the image. Gansbeke et al. [59] and Simeoni et al. [54] proposed that the semantic information of the foreground should be close to the global semantic information of the entire image. In [21, 69], the region where the point with the highest absolute confidence value is located is made as the foreground. Wang et al. [69] suggested that the foreground is unlikely to cover three or four corners of the image. Li and Shin [35] proposed that a background mask is likely to occupy at least two edges of the image. Although these priors achieved limited success, they are all heuristic and assumption-based. In contrast, the robustness and reliability of our proposed methods are demonstrated by mathematical and statistical ensemble theories. + +![](images/380f4eb0f2e6ed120a1ba9d121c52c6e65b88b222142cad94d0ae4726457445e.jpg) +Figure 2. Framework of UnionCut. UnionCut takes an image as input with features extracted from a frozen DINO, and applies Unit Voters (UVs) with each using one patch as its seed. Red rectangles refer to the seed patch, and blue areas indicate the UV's output of the region similar to its seed patch. Aggregating all UV's outputs yields a heat map of the background, which is then inverted, thresholded, and rectified with a corner prior to produce a binary mask of the image's foreground union. + +# 2.4. Ensemble Methods + +Ensemble methods are learning algorithms that construct a set of classifiers and then classify new data points by taking a (weighted) vote of their predictions [18]. Most ensemble methods can be classified into two types: 1) Parallel a set of weak classifiers run in parallel and form a strong classifier by performing a plurality vote, e.g. Pasting [5], Bagging [4], and Random Forest [6]; and 2) Sequential running batches of weak classifiers iteratively with each batch focusing on hard cases of the last one, e.g. [60] and AdaBoost [23]. UnionCut is of the former: a set of weak classifiers is created for an image, with each in parallel voting every patch $(8 \times 8$ pixels) of the image as foreground or background, which forms a strong classifier by aggregating votes from them to detect the foreground union of an image. + +# 3. Methodology + +UnionCut is an ensemble method, enabling UOD methods to judge if a detected region is foreground and stop discovery when most objects have been discovered to avoid over-segmentation. In the context of ensemble methods, UnionCut serves as the strong classifier, based on many parallel weak classifiers named "unit voter" (UV) here. Fig. 2 illustrates the frameworks of the proposed UnionCut. Unlike previous assumption-based foreground priors mentioned in Sec. 2.3, UnionCut is reliable and robust with ensemble theories. First, we introduce the details of UV. + +# 3.1. Unit Voter + +UV is based on features output by a ViT pretrained by DINO [9]. In our implementation, an image is resized into $224 \times 224$ and serialized into $28 \times 28$ (i.e. 784) patches of $8 \times 8$ pixels before feature extraction of each patch with the ViT. The set of patches of an image is defined as $P =$ + +![](images/b4eb1719381731e08dc1656ddb5f8503c0d2a65289a6c27500513d98804d5ee1.jpg) +Figure 3. Overview of Unit Voter (UV). For clarity, the image is divided into $3 \times 3$ patches. With the seed patch and anti-seed patches of the UV initialized, the image is modeled as a graph with each patch and two dummy terminals being nodes and connected by edges whose thickness reflects their weights. After applying min-cut to the graph, thin edges are removed to divide the graph into two disjoint parts containing one terminal node. The part with the Source node is the region output by the UV, indicating the area of features similar to the seed patch. + +$\{p_1,p_2,\dots ,p_{784}\}$ with the corresponding L2-normalized Key vectors as their features output by the last layer of the model, which is defined as $K = \{k_{1},k_{2},\ldots ,k_{784}\}$ . Then, the patches and features are leveraged by each UV. + +As required by ensemble methods, the number of weak classifiers (i.e. UV) to form a strong classifier (i.e. Union-Cut) should be significant, and each UV should be different [18]. To achieve these goals, 784 independent UVs can be built based on the 784 patches of an image, with each UV corresponding to a unique patch $p_f \in P$ . The patch corresponding to a UV is called that UV's "seed patch". + +Taking a UV built from a seed patch $p_f \in P$ as an example, its pipeline is illustrated in Fig. 3. Specifically, given a UV's seed patch $p_f \in P$ with feature $k_f \in K$ , the set of patches dissimilar to $p_f$ , defined as the UV's anti-seed patches set $B_f = \{p_b | p_b \in P, b \neq f, k_b^T k_f < 0\}$ , can be obtained. With $p_f, B_f$ , all $28 \times 28$ patches $P$ and features $K$ , a directed graph is constructed, as shown in Fig. 3, where each patch is a node and two dummy terminal nodes, named "Source" and "Target", are added. In the graph, nodes of adjacent patches (i.e. eight-neighborhood patches) are connected by bidirectional edges (named "n-links") with weights representing their similarity. Directed edges named "t-links" connect the Source to each patch node or each patch node to the Target, with weights based on the likelihood of a patch being more similar to the UV's seed patch $p_f$ or its anti-seed patches set $B_f$ . The terms "t-link" and "n-link" originate from early image segmentation methods that modeled images as graphs [3, 49]. The purpose of building this graph is to split the image by removing edges, forming two disconnected subgraphs with each either containing the Source or the Target, while minimizing the total weight of the removed edges. The subgraph with the Source is expected to contain patches similar to the UV's seed patch, while the one with the Target includes patches similar to the anti-seed patches $B_f$ . This bipartition process is also known as the min-cut problem, which is solved by UV using an efficient solution [2]. Next, we introduce the algorithms for assigning weights to n-links and t-links. + +n-links The weight $w(p_i, p_j)$ of the n-link between two adjacent patches $p_i$ and $p_j$ is given by Eq. (1). + +$$ +w \left(p _ {i}, p _ {j}\right) = \frac {e ^ {- \frac {\left(k _ {i} - k _ {j}\right) ^ {T} \left(k _ {i} - k _ {j}\right)}{\beta}}}{\sqrt {\left(i _ {x} - j _ {x}\right) ^ {2} + \left(i _ {y} - j _ {y}\right) ^ {2}}} \tag {1} +$$ + +$$ +\beta = \frac {2 \sum_ {p _ {i} \in P} \sum_ {p _ {j} \in P} \mathbf {1} (p _ {i} , p _ {j}) (k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}{C} +$$ + +where $(i_x,i_y)$ and $(j_{x},j_{y})$ are the indices of the two patches under the scale of the feature map (i.e. $28\times 28$ ) output by the ViT. $\mathbf{1}(p_i,p_j)$ is an indicator function which returns 1 if the $p_i$ and $p_j$ are adjacent, otherwise 0. Finally, $C$ is the number of n-links in the graph. + +t-links An UV's graph has three kinds of patches: seed patch $p_f$ , anti-seed patches set $B_f$ , and the other patches $P - B_f \cup \{p_f\}$ , whose weight assignment strategies for t-links are different. Suppose the Source and Target are defined as $S$ and $T$ . After min-cut, the t-link from $S$ to $p_f$ should be kept while the t-link between $p_f$ and $T$ should be cut. To this end, a large weight $W$ is given to the t-link from $S$ to $p_f$ so that it is too large to be cut, and the weight between $p_f$ and $T$ is 0. Following early graph cut methods [3, 49], $W$ can be decided by Eq. (2). + +$$ +W = 1 + \max _ {p _ {i} \in P} \sum_ {p _ {j} \in P} \mathbf {1} \left(p _ {i}, p _ {j}\right) w \left(p _ {i}, p _ {j}\right) \tag {2} +$$ + +Similarly, anti-seed patches must stay connected with $T$ after min-cut. Therefore, t-links from $\forall p_{b}\in B_{f}$ to $T$ must have a large value, which can also be given by Eq. (2), while the weights of t-links from $S$ to $\forall p_{b}\in B_{f}$ are 0. + +Weights of other t-links between $S$ or $T$ and $p_i \notin B_f \bigcup \{p_f\}$ are decided by Eq. (3). + +$$ +w \left(p _ {i}, S\right) = - \log \frac {e ^ {\sqrt {\left(k _ {i} - k _ {f}\right) ^ {T} \left(k _ {i} - k _ {f}\right)}}}{e ^ {\sqrt {\left(k _ {i} - k _ {f}\right) ^ {T} \left(k _ {i} - k _ {f}\right)}} + e ^ {\min _ {p _ {j} \in B _ {f}} \sqrt {\left(k _ {i} - k _ {j}\right) ^ {T} \left(k _ {i} - k _ {j}\right)}}} \tag {3} +$$ + +$$ +w (p _ {i}, T) = - \log \frac {e ^ {\min _ {p _ {j} \in B _ {f}} \sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}{e ^ {\sqrt {(k _ {i} - k _ {f}) ^ {T} (k _ {i} - k _ {f})}} + e ^ {\min _ {p _ {j} \in B _ {f}} \sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}} +$$ + +Eq. (3) explains that if a patch $p_i \notin B_f \cup \{p_f\}$ is more similar to anti-seed patches set $B_f$ than the seed patch $p_f$ , $w(p_i, T)$ should exceed $w(p_i, S)$ , making the t-link from $S$ to $p_i$ more likely to be cut, and vice versa. + +With the graph initialized using the method introduced above, the UV executes a min-cut and returns a binary mask, where "1" indicates patches remaining connected to the Source, i.e. regions sharing features with the seed patch $p_f$ . Therefore, the function of each UV is to return the region that is similar to its seed patch. + +# 3.2. UnionCut + +UnionCut is a strong classifier composed of 784 UVs as weak classifiers, with each UV using a unique patch from the image's $28 \times 28$ patches as its seed patch. Considering the function of UV introduced above, there are two types of UVs in UnionCut: 1) background UV: if its seed patch is background, the UV returns a binary mask indicating background regions; and 2) foreground UV: if its seed patch is foreground, the UV outputs a binary mask of foreground regions. UnionCut's output is determined by voting from the 784 UVs. Given the set of binary masks $M = \{m_f | \forall p_f \in P\}$ returned by all UVs, a heat map $A$ is generated by aggregating all UVs' output using Eq. (4). + +$$ +A (x, y) = \sum_ {m _ {f} \in M} m _ {f} (x, y) \tag {4} +$$ + +where $x \in [1, 28]$ and $y \in [1, 28]$ indicate the 2D coordinate of a patch's response on $A$ . + +For easier understanding, intuitively, the foreground usually occupies a smaller area than the background in the image [53]. Thus, we can assume that there are more background UVs than foreground UVs in UnionCut, causing background areas to have higher intensities on the aggregated heat map $A$ . An example can be seen in Fig. 2. To make it intuitive, the brightness of $A$ is inverted so that foreground regions have higher responses than the background, producing an inverted heat map $H$ using Eq. (5). By thresholding $H$ , most objects in the image are segmented as the foreground union. + +$$ +H (x, y) = 2 5 5 - 2 5 5 \times \frac {A (x , y) - \operatorname* {m i n} _ {x , y} A (x , y)}{\operatorname* {m a x} _ {x , y} A (x , y) - \operatorname* {m i n} _ {x , y} A (x , y)} \tag {5} +$$ + +Note that the assumption introduced above is for easier understanding of our approach. UnionCut is mathematically and statistically robust, with an explanation in Appendix 6. Moreover, UnionCut remains effective even when the foreground area is larger than the background, with examples in Appendix 7. The mathematical explanation also supports this claim. Furthermore, compared to more straightforward methods detecting patches similar to a given patch, e.g. matching with cosine similarity [53, 54], UV is more sophisticated. Appendix 8 provides a detailed analysis of UV's irreplaceable advantages and necessity. + +# 3.2.1. Thresholding + +A threshold for the bipartition of the heat map is required to estimate the foreground union. UnionCut automatically determines a threshold for the heat map $H$ by clustering the response values on it with Mean-Shift [24]. Since the bright areas on the heat map represent the foreground, the resulting clusters are sorted by their centres in descending order, with the top half retained and identified as the union of the foreground in the image. Finally, a $28 \times 28$ binary mask $U_{cut}$ with "1" as the foreground union is obtained. + +# 3.2.2. Corner Prior + +Similar to [69], a corner prior is used by UnionCut to check the reliability of the foreground union mask. Specifically, if $U_{cut}$ occupies four corners of the image, it must be inverted (i.e. $U_{cut} := 1 - U_{cut}$ ) because the foreground union is unlikely to occupy four corners simultaneously. The reliability of this corner prior is discussed in Appendix 9. + +# 3.3. UnionSeg + +While UnionCut effectively detects an image's foreground union, running 784 UVs per image is computationally expensive. To improve efficiency, we introduce UnionSeg, a ViT-based model distilled from UnionCut as a surrogate model. Unlike UnionCut, UnionSeg is an end-to-end model that directly predicts the foreground union from an input image. To minimize learnable parameters, inspired by FOUND [54], UnionSeg uses a frozen ViT trained by DINO [9] as its backbone, followed by a learnable $1 \times 1$ convolutional layer and a sigmoid function, compressing each ViT patch feature into a scalar confidence score. Patches with scores above 0.5 are classified as the foreground union. Appendix 10 details the differences between UnionSeg and FOUND. Suppose the soft output of UnionSeg of an image is $U_{seg}^{s}$ , with its thresholded hard output defined as $U_{seg}^{h}$ . Given UnionCut's output $U_{cut}$ of an image, the label $L$ of an image used to train UnionSeg is defined by Eq. (6). + +$$ +L = \left\{ \begin{array}{l l} U _ {c u t}, & \text {i f I o U} \left(U _ {s e g} ^ {h}, U _ {c u t}\right) < 0. 5, \\ U _ {s e g} ^ {h}, & \text {o t h e r w i s e .} \end{array} \right. \tag {6} +$$ + +where $IoU(\cdot)$ returns the intersection over union between two binary masks. The motivation behind Eq. (6) is that + +UnionCut's output may contain errors, so UnionSeg does not rely entirely on UnionCut's output as labels. This enables UnionSeg to predict the foreground union more efficiently and accurately than UnionCut. The loss function used to train UnionSeg is shown in Eq. (7). + +$$ +L o s s = L \log U _ {s e g} ^ {s} + (1 - L) \log \left(1 - U _ {s e g} ^ {s}\right) + 1 (\text {i t e r} < 1 0 0) \Lambda \tag {7} +$$ + +$$ +\Lambda = U _ {c u t} \log U _ {s e g} ^ {s} + (1 - U _ {c u t}) \log (1 - U _ {s e g} ^ {s}) +$$ + +where $\mathbf{1}(iter < 100)$ returns 1 if the training iterations are fewer than 100; otherwise 0. This ensures that at the start of training, when UnionSeg is not yet well trained, it is forced to use both UnionCut's output $U_{cut}$ and $L$ as labels to aid learning. + +# 4. Experiments + +Several experiments are conducted to evaluate the effectiveness of UnionCut/UnionSeg, comparing the performance of state-of-the-art UOD methods before and after integrating UnionCut or UnionSeg on several tasks: unsupervised single object discovery (Sec. 4.1), unsupervised saliency detection (Sec. 4.2), and self-supervised instance segmentation (Sec. 4.3). We then discuss the properties of UnionCut and UnionSeg in Sec. 4.4, and evaluate the impact of different components of them in Sec. 4.5. Additionally, we provide more qualitative comparisons in Appendix 15. + +Baseline UOD methods Our experiments use existing state-of-the-art UOD methods for single or multiple object discovery as baselines, i.e. LOST [53], TokenCut [21], FOUND [54] and CutLER [69]. To enhance their performance, we integrate them with UnionCut and UnionSeg by replacing their standard foreground priors with UnionCut/UnionSeg, with details provided in Appendix 11. + +Implementation Details We use a ViT-S/8 [19] pretrained by DINO [9] as the image feature extractor for UnionCut and UnionSeg. Following [52, 54], we distill UnionSeg from UnionCut using DUTS-TR[64], a saliency detection dataset with 10,553 images. Training is done with a batch size of 50, an initial learning rate of 0.05 that decays to $95\%$ every 50 iterations, and AdamW [38] for parameter updates over 600 iterations. Implementation details for all tested methods, i.e. LOST+UnionCut/UnionSeg, TokenCut+UnionCut/UnionSeg, FOUND+UnionCut/UnionSeg, and CutLER+UnionSeg, are provided in Appendix 11. + +# 4.1. Unsupervised Single Object Discovery + +In this experiment, we combine baseline UOD methods with UnionCut and UnionSeg to see if their performance improves in unsupervised single object discovery. + +Benchmark Following [21, 53, 54], the trainval sets of PASCAL VOC07 (5011 images), VOC12 (11540 images) [22] and COCO20K [37, 61] (19817 randomly sam + +
MethodVOC07VOC12COCO20K
- No learning -
Selective Search [58]18.820.916.0
EdgeBoxes [80]31.131.628.8
Kim et al. [30]43.946.435.1
Zhang et al. [75]46.250.534.8
DDT+ [70]50.253.138.2
rOSD [61]54.555.348.5
LOD [62]53.655.148.5
DINO-seg [9, 53](ViT-S/16 [9])45.846.242.0
DSS [40](ViT-S/16 [9])62.766.452.2
LOST [53](ViT-S/16 [9])61.964.050.7
LOST(ViT-S/16 [9])+UnionCut62.4(0.5↑)65.3(1.3↑)51.9(1.2↑)
LOST(ViT-S/16 [9])+UnionSeg63.2(1.3↑)67.1(3.1↑)54.9(4.2↑)
TokenCut [21](ViT-S/16 [9])68.872.158.8
TokenCut(ViT-S/16 [9])+UnionCut69.2(0.4↑)72.3(0.2↑)62.1(3.3↑)
TokenCut(ViT-S/16 [9])+UnionSeg69.7(0.9↑)72.7(0.6↑)62.6(3.8↑)
- With learning -
FreeSOLO [67, 69]44.049.735.2
LOD+CAD [21, 53]56.361.652.7
rOSD+CAD [21, 53]58.362.353.0
LOST+CAD [53](ViT-S/16 [9])65.770.457.5
SelfMask [52]72.375.362.7
FOUND [54](ViT-S/16 [9])72.576.162.9
FOUND(ViT-S/16 [9])‡71.475.762.9
FOUND(ViT-S/16 [9])+UnionCut72.6(1.2↑)76.1(0.4↑)63.0(0.1↑)
FOUND(ViT-S/16 [9])+UnionSeg72.3(0.9↑)75.9(0.2↑)63.1(0.2↑)
CutLER [69](ViT-B/8 [9])†73.369.570.7
CutLER(ViT-B/8 [9])+UnionSeg73.8(0.5↑)71.2(1.7↑)72.4(1.7↑)
+ +Table 1. Performances of UOD methods on single object discovery with red indicates the highest score. CAD: a class-agnostic detector trained with the output of no-learning UOD methods as the ground truth. †: results of official implementations and checkpoints by our measurement. ‡: results of our replication with official implementation. + +pled images from COCO2014 [37]) are selected as the benchmark for evaluation. + +Metrics Similar to [21, 53, 54, 61, 62], we use CorLoc (Correct Localization) to evaluate UOD algorithms' performance on single object discovery. CorLoc scores an image as 1 if the intersection over union (IoU) between any ground truth bounding box and the algorithm's predicted bounding box exceeds 0.5, otherwise it scores 0. In our experiments, since CutLER can detect multiple objects, we select the bounding box with the highest confidence score as its single object prediction for the image. + +Tab. 1 shows that after integrating UnionCut or UnionSeg, our selected baseline UOD methods improve across all three benchmarks, achieving state-of-the-art results, indicating that UnionCut/UnionSeg as a robust foreground prior enhances UOD algorithms' performance in single object discovery. In Appendix 12, we also evaluate these UOD methods from the perspective of the performance upper boundary proposed by Rambhatla et al. [45]. + +# 4.2. Unsupervised Saliency Detection + +We compare the performance of UOD methods before and after integrating UnionCut/UnionSeg on saliency detection, which involves segmenting salient objects in images. Since CutLER predicts multiple object masks per image, we use the union of these masks as its saliency detection result. + +Benchmark Similar to [21, 40, 53, 54, 63], ECSSD [51] (1000 real-world images), DUTS-TE [64] (5019 images sampled from ImageNet test), and DUT-OMRON [74] (5168 natural images) are chosen as the benchmark. + +Metrics Following [21, 53, 54], we report results with three metrics: Accuracy, IoU, and $\max F_{\beta}$ . Accuracy measures the percentage of correctly classified pixels compared to ground truth. IoU is the ratio of the intersection between the UOD output and ground truth to their union. $\max F_{\beta}$ is calculated as $\frac{(1 + \beta^2)\text{Precision} \times \text{Recall}}{\beta^2\text{Precision} + \text{Recall}}$ where Precision is the percentage of predicted pixels matching ground truth, and Recall is the proportion of ground truth pixels covered by the prediction. We set $\beta$ to 0.3, as in [21, 53, 54]. + +Tab. 2 illustrates that the selected baseline methods improve when integrated with UnionCut/UnionSeg. TokenCut, with UnionSeg, performs best among no-learning approaches. And the performance of FOUND, the previous state-of-the-art method, is also increased by UnionCut and UnionSeg on most benchmarks and metrics. + +# 4.3. Self-supervised Instance Segmentation + +Some UOD methods target self-supervised instance segmentation, where models are trained using pseudo-annotations given by UOD algorithms, such as MaskCut for CutLER and FreeMask for FreeSOLO [67]. In the experiment, we explore whether UnionSeg improves CutLER's performance on instance segmentation. Similar to [69], COCO20k [37, 61] and COCO [37](5000 images from COCO2017's validation set) are used for evaluation, with average precision (AP) as the metric. Tab. 3 illustrates that by using UnionSeg to improve the pseudo-labels given by MaskCut, CutLER performs better on most metrics. + +Since UnionCut with MaskCut takes much time to generate the foreground union and UOD results of all images in ImageNet (7 weeks in total, with 4 weeks for UnionCut and 3 weeks for MaskCut with 18 subprocesses running in parallel limited by our available computational resources), we only consider UnionSeg in this section. However, the effectiveness of UnionCut on self-supervised instance segmentation is also of interest. Therefore, we conduct another experiment that trains CutLER on a much smaller dataset (the subset of VOC12 containing 2913 images with pixel-level annotations). Details can be found in Appendix 13. + +# 4.4. Properties of UnionCut and UnionSeg + +We explore the properties of UnionCut and UnionSeg. First, their accuracy in foreground union detection is evaluated. + +
MethodDUT-OMRONDUTS-TEECSSD
Acc.IoUmaxFβAcc.IoUmaxFβAcc.IoUmaxFβ
- No learning -
HS [73]84.343.356.182.636.950.484.750.867.3
wCtr [77]83.841.654.183.539.252.286.251.768.4
WSC [36]86.538.752.386.238.452.885.249.868.3
DeepUSPS [41]77.930.541.477.330.542.579.544.058.4
BigBiGAN [63]85.645.354.987.849.860.889.967.278.2
E-BigBiGAN [63]86.046.456.388.251.162.490.668.479.7
LOST [53]79.741.047.387.151.861.189.565.475.8
DSS [40]-56.7--51.4--73.3-
MOST [45]8747.557.089.753.866.689.063.179.1
TokenCut [21, 54](ViT-S/16 [9])89.761.869.791.462.475.593.477.287.4
TokenCut(ViT-S/16 [9])+UnionCut89.662.4(0.6↑)71.0(1.3↑)91.6(0.2↑)63.2(0.8↑)80.2(4.7↑)93.8(0.4↑)77.3(0.1↑)91.0(3.6↑)
TokenCut(ViT-S/16 [9])+UnionSeg90.6(0.9↑)62.1(0.3↑)71.2(1.5↑)92.3(0.9↑)63.4(1.0↑)80.6(5.1↑)94.1(0.7↑)77.291.0(3.6↑)
- With learning -
LOST [53]81.848.957.888.757.269.791.672.383.7
SelfMask [52]91.965.5-93.366.0-95.581.8-
UnSeGArmaNet [46]----60.3--77.8-
CutLER [69](ViT-B/8 [9])†74.343.848.178.146.554.287.269.376.3
CutLER(ViT-B/8 [9])+UnionSeg87.7(13.4↑)57.8(14.0↑)64.1(16.0↑)90.1(12.0↑)61.4(14.9↑)72.3(18.1↑)94.0(6.8↑)79.6(10.3↑)87.8(11.5↑)
FOUND [54](ViT-S/16 [9])92.261.370.894.266.376.395.181.393.5
FOUND(ViT-S/16 [9])‡91.860.869.494.066.075.495.081.093.0
FOUND(ViT-S/16 [9])+UnionCut92.4(0.6↑)61.8(1.0↑)71.2(1.8↑)94.4(0.4↑)66.7(0.7↑)76.7(1.3↑)94.980.693.3(0.3↑)
FOUND(ViT-S/16 [9])+UnionSeg92.5(0.7↑)61.7(0.9↑)71.7(2.3↑)94.4(0.4↑)66.6(0.6↑)77.1(1.7↑)94.780.293.4(0.4↑)
+ +Table 2. Performances of UOD methods on saliency detection with red indicates the highest score. †: results of official codes and checkpoints by our measurement. ‡: results of our replication with official implementation. + +
MethodsPretrainDetectorInit.COCO20KCOCO val2017
\(AP_{50}^{box}\)\(AP_{75}^{box}\)\(AP^{box}\)\(AP_{50}^{mask}\)\(AP_{75}^{mask}\)\(AP^{mask}\)\(AP_{50}^{box}\)\(AP_{75}^{box}\)\(AP^{box}\)\(AP_{50}^{mask}\)\(AP_{75}^{mask}\)\(AP^{mask}\)
LOST [53]IN+COCOFRCNN [47]DINO---2.41.01.1------
MaskDistill [59]IN+COCOMRCNN [27]MoCo [28]---6.82.12.9------
FreeSOLO [67, 69]IN+COCOSOLOv2 [65]DenseCL [66]9.73.24.19.73.44.39.63.14.29.43.34.3
DETReg [1]INDDETR [78]SwAV [8]------3.10.61.08.81.93.3
DINO [9]IN-DINO1.70.10.3---------
ProMerge [35]IN-DINO-----3.0-----2.4
ProMerge+ [35]INCascade [7]DINO-----8.9-----9.0
CutLER [69]INCascade [7]DINO22.412.511.919.610.09.221.911.812.318.99.79.2
CutLER+UnionSegINCascade [7]DINO24.112.212.920.39.110.023.712.212.919.79.09.9
+ +Table 3. Performances of self-supervised instance segmentation. The dataset (Pretrain) used to train each model (Detector) and each model's initialization (Init) are also listed. IN denotes ImageNet. + +Baseline Li et al. [35] proposes ProMerge, which contains a foreground union detection method. FOUND [54] is a UOD method; however, considering that it includes an algorithm for detecting background regions in an image, we extract this algorithm separately as a baseline and treat the inversion of its output as the foreground union detection result. For convenience, we still refer to their foreground union detection algorithms as ProMerge and FOUND. + +Benchmark The subset of VOC12 with 2913 images of pixel-level annotations is used as the benchmark, and the ground truth mask of each instance is combined as the ground truth of the foreground union in an image. + +Metrics IoU, precision, and recall are used to measure the performance. Besides, since representative datasets [22, 37] label objects of a given class list, there may be cases where some objects in the image are not labelled and thus are not considered part of the foreground union by the ground truth (Examples can be seen in Appendix 14). Therefore, directly using these metrics and dataset-provided + +labels to measure the performance of foreground union detection is inaccurate. To address this, we propose a fuzzy evaluation metric similar to CorLoc, named CorUnion. Specifically, we compare the foreground union output by UnionCut or UnionSeg with the ground truth for a given image and compute a specific metric (e.g. IoU). If this metric exceeds a certain threshold, we consider the foreground union detection on that image successful. The success rate across the dataset is then used as the CorUnion value. This approach also allows us to study the properties of UnionCut and UnionSeg by using different metrics or adjusting the threshold for successful detection. + +Foreground Union Accuracy Tab. 4 shows the performance of UnionCut and UnionSeg, measured by precision, recall, and IoU. We also report CorUnion using IoU as the metric to judge an image's foreground union detection success, with increasing IoU threshold from 0.5 to 0.9. As shown in Fig. 4, both UnionCut and UnionSeg are more accurate than the baselines at foreground union detection. + +
MethodPrecisionRecallIoU
FOUND69.281.157.9
ProMerge-73.4-78.4-59.9
UnionCut-70.7-84.6-60.9
UnionSeg80.978.665.7
+ +![](images/063578444ed6b0d4ebf754a8379dfd52974d81605a6a7404dd16912c2677489d.jpg) +Figure 4. Foreground union detection measured by CorUnion. + +Instances Count In addition, we assessed UnionCut and UnionSeg's performance across varying instance counts in VOC12, categorized into groups based on the number of instances per image. As shown in Fig. 5a, UnionSeg, a training-based method, outperforms training-free approaches in precision, supporting that self-training corrects initial errors made by training-free methods. This finding aligns with previous research [21, 53, 54]. As instance numbers increase, UnionSeg maintains its precision, while other methods decline. Furthermore, Fig. 5b shows that UnionCut surpasses other methods in recall of the foreground union, remaining stable as the number of instances rises, unlike the decrease seen with other methods. This demonstrates the functional distinction between UnionCut and other training-free methods: as the number of instances per image increases, UnionCut consistently predicts the foreground union covering most instances, though precision inevitably decreases. This consistency highlights UnionCut's effectiveness as a robust prior for UOD algorithms to cease exploration optimally. + +Overall, UnionSeg outperforms UnionCut and other baseline methods (Fig. 4) with higher precision (Fig. 5a), making it better suited for identifying if a detected area is foreground. UnionCut, with its higher recall (Fig. 5b), is better for determining if the discovery is complete. + +![](images/e5632f2f1a552857653fe6a8a559047ee5c9cadffe68d96fc7f51687ef9e88b4.jpg) +Figure 5. Performance of foreground union detection on images with varying instance numbers, measured by CorUnion with precision $>0.8$ (a) and recall $>0.8$ (b) as success criteria. + +# 4.5. Ablation Study + +UnionCut consists of two elements: aggregated UVs and corner prior, while UnionSeg builds on UnionCut with distillation. To evaluate each element's contribution, TokenCut [21] is used as the baseline, with each element integrated for single object discovery. Benchmarks and metrics are inherited from Sec. 4.1. As shown in Tab. 5, each element contributes to the performance increase of UOD, demonstrating their effectiveness. + +Table 4. Average performance on foreground union detection. + +
MethodaUcpdVOC07VOC12COCO20K
TokenCut [21]68.872.158.8
TokenCut+aU✓69.072.362.0
TokenCut+UnionCut✓✓69.272.362.1
TokenCut+UnionSeg✓✓✓69.772.762.6
+ +We further evaluate the inference speed of UnionCut and UnionSeg. We include FOUND and ProMerge in Sec. 4.4 as the baseline. Our implementation is based on Python 3.11.9, Pytorch 2.3.1 and CUDA 12.1. As shown in Tab. 6, UnionSeg significantly increases the efficiency of foreground union detection compared with UnionCut. + +Table 5. Contribution of each component of UnionCut and UnionSeg. aU, cp, and d denote aggregated UVs, corner prior and distillation, respectively. + +
MethodCPUGPUFPS
FOUNDIntel 14700KFNVIDIA RTX4070Ti Super2
ProMergeIntel 14700KFNVIDIA RTX4070Ti Super3
UnionCutIntel 14700KFNVIDIA RTX4070Ti Super0.1
UnionSegIntel 14700KFNVIDIA RTX4070Ti Super125
+ +Table 6. Inference speed of UnionSeg and UnionCut. FPS (frames per second) is evaluated throughout the whole pipeline from the image input to the output of the foreground union. + +# 5. Conclusion + +This paper introduces UnionCut, an unsupervised ensemble method to robustly predict foreground unions in images. Additionally, we propose UnionSeg distilled from UnionCut to detect foreground unions more accurately and efficiently. Both of them help UOD algorithms determine if discovered areas are in the foreground and when to stop discovery on an image, enhancing UOD's accuracy. While both are effective, each has unique advantages, and the choice between them depends on the specific UOD requirements for precision, recall, and efficiency. Our experiments show that integrating UnionCut or UnionSeg improves the performance of state-of-the-art UOD methods in single object discovery, saliency detection, and instance segmentation, demonstrating their effectiveness as robust foreground priors for unsupervised object discovery. + +# References + +[1] Amir Bar, Xin Wang, Vadim Kantorov, Colorado J Reed, Roei Herzig, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Detreg: Unsupervised pretraining with region priors for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14605-14615, 2022. 7 +[2] Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE transactions on pattern analysis and machine intelligence, 26(9):1124-1137, 2004. 1, 4 +[3] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 4 +[4] Leo Breiman. Bagging predictors. Machine learning, 24: 123-140, 1996. 3 +[5] Leo Breiman. Pasting small votes for classification in large databases and on-line. Machine learning, 36:85-103, 1999. 3 +[6] Leo Breiman. Random forests. Machine learning, 45:5-32, 2001. 3 +[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7, 6 +[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 7 +[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6, 7 +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020. +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2 +[12] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 1 +[13] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969–978, 2022. 1 +[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe + +Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1 +[15] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning from video to detect foreground objects in single images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4335-4343, 2017. 1 +[16] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning of foreground object segmentation. International Journal of Computer Vision, 127:1279-1302, 2019. 1 +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6 +[18] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 1, 3, 4 +[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 5, 6 +[20] Rosana El Jurdi, Caroline Petitjean, Paul Honeine, and Fahed Abdallah. Bb-unet: U-net with bounding box prior. IEEE Journal of Selected Topics in Signal Processing, 14(6):1189-1198, 2020. 1 +[21] Y. Wang et al. Self-supervised transformers for unsupervised object discovery using normalized cut. In CVPR, 2022. 1, 2, 5, 6, 7, 8 +[22] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 7 +[23] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1): 119-139, 1997. 3 +[24] Keinosuke Fukunaga and Larry Hostetler. The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Transactions on information theory, 21(1):32-40, 1975. 5 +[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2 +[26] Emanuela Haller, Adina Magda Florea, and Marius Leordeanu. Iterative knowledge exchange between deep learning and space-time spectral clustering for unsupervised segmentation in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7638-7656, 2021. 1 + +[27] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7 +[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 7 +[29] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. 1 +[30] Gunhee Kim and Antonio Torralba. Unsupervised detection of regions of interest using iterative link analysis. Advances in neural information processing systems, 22, 2009. 6, 7 +[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7 +[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1 +[33] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In European Conference on Computer Vision, pages 290-308. Springer, 2020. 1 +[34] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1 +[35] Dylan Li and Gyungin Shin. Promerge: Prompt and merge for unsupervised instance segmentation. In ECCV, 2024. 1, 2, 7 +[36] Nianyi Li, Bilin Sun, and Jingyi Yu. A weighted sparse coding framework for saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5223, 2015. 7 +[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 1, 2, 3, 4 +[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[39] R Austin McEver and BS Manjunath. Pcams: Weakly supervised semantic segmentation using point supervision. arXiv preprint arXiv:2007.05615, 2020. 1 +[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 6, 7 + +[41] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 7 +[42] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 2 +[43] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 1 +[44] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8843-8850, 2019. 1 +[45] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. Most: Multiple object localization with self-supervised transformers for object discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15823-15834, 2023. 6, 7 +[46] Kovvuri Sai Gopal Reddy, Bodduluri Saran, A. Mudit Adityaja, Saurabh J. Shigwan, Nitin Kumar, and Snehasis Mukherjee. Unsegarmanet: Unsupervised image segmentation using graph neural networks with convolutionalarma filters, 2024.7 +[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 7 +[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1 +[49] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. "grabcut" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 4 +[50] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2 +[51] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4): 717-729, 2015. 6, 1, 2, 3, 4 +[52] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6, 7 + +[53] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC 2021-32nd British Machine Vision Conference, 2021. 1, 2, 4, 5, 6, 7, 8, 3 +[54] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3176-3186, 2023. 1, 2, 5, 6, 7, 8, 3, 4 +[55] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017. 1 +[56] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1 +[57] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1 +[58] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104: 154-171, 2013. 6, 7 +[59] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022. 1, 2, 7, 3 +[60] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, pages I-I. IEEE, 2001. 3 +[61] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 779-795. Springer, 2020. 5, 6, 7 +[62] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 6, 7 +[63] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6, 7 +[64] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 136-145, 2017. 5, 6, 1, 2, 3, 4 +[65] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmenta + +tion. Advances in Neural information processing systems, 33:17721-17732, 2020. 7 +[66] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 1, 2, 7 +[67] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 1, 2, 6, 7 +[68] Xuehui Wang, Kai Zhao, Ruixin Zhang, Shouhong Ding, Yan Wang, and Wei Shen. Contrastmask: Contrastive learning to segment every thing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11604-11613, 2022. 1 +[69] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 5, 6, 7, 8 +[70] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 6, 7 +[71] Yulin Wu, Chang Liu, Lei Chen, Dong Zhao, Qinghe Zheng, and Hongchao Zhou. Perturbation consistency and mutual information regularization for semi-supervised semantic segmentation. Multimedia Systems, 29(2):511-523, 2023. 1 +[72] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, Ferdous Sohel, and Dan Xu. Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6984-6993, 2021. 1 +[73] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 7 +[74] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3166-3173, 2013. 6, 4 +[75] Runsheng Zhang, Yaping Huang, Mengyang Pu, Jian Zhang, Qingji Guan, Qi Zou, and Haibin Ling. Object discovery from a single unlabeled image by mining frequent itemsets with multi-scale features. IEEE Transactions on Image Processing, 29:8606-8621, 2020. 6, 7 +[76] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021. 2 +[77] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 7 + +[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 7 +[79] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 1 +[80] C Lawrence Zitnick and Piotr Dólar. Edge boxes: Locating object proposals from edges. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 391-405. Springer, 2014. 6, 7 \ No newline at end of file diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/images.zip b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..128bca1b4ad90cf3cf786b0c333d2ff474f3fd2e --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b51d16dc0448c28b3451c11915b4ba885fdad304c632ae1c2720d62c71a3d3a +size 574455 diff --git a/2025/Ensemble Foreground Management for Unsupervised Object Discovery/layout.json b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e369271c7de2637cc4b7e884281e45875bc46215 --- /dev/null +++ b/2025/Ensemble Foreground Management for Unsupervised Object Discovery/layout.json @@ -0,0 +1,9300 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 530, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 530, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 530, + 121 + ], + "type": "text", + "content": "Ensemble Foreground Management for Unsupervised Object Discovery*" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 170, + 143, + 439, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 170, + 143, + 439, + 171 + ], + "spans": [ + { + "bbox": [ + 170, + 143, + 439, + 171 + ], + "type": "text", + "content": "Ziling Wu," + }, + { + "bbox": [ + 170, + 143, + 439, + 171 + ], + "type": "inline_equation", + "content": "^{\\dagger}" + }, + { + "bbox": [ + 170, + 143, + 439, + 171 + ], + "type": "text", + "content": " Armaghan Moemeni, Praminda Caleb-Solly University of Nottingham, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "spans": [ + { + "bbox": [ + 120, + 174, + 487, + 185 + ], + "type": "text", + "content": "{ziling.wu, armaghan.moemeni, praminda.caleb-solly}@nottingham.ac.uk" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 239, + 296, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 239, + 296, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 239, + 296, + 585 + ], + "type": "text", + "content": "Unsupervised object discovery (UOD) aims to detect and segment objects in 2D images without handcrafted annotations. Recent progress in self-supervised representation learning [9, 66] has led to some success in UOD algorithms [21, 53, 69]. However, the absence of ground truth provides existing UOD methods with two challenges: 1) determining if a discovered region is foreground or background, and 2) knowing how many objects remain undiscovered. To address these two problems, previous solutions rely on foreground priors [53, 59, 67, 69] to distinguish if the discovered region is foreground, and conduct one or fixed iterations of discovery. However, the existing foreground priors are heuristic and not always robust, and a fixed number of discoveries leads to under or over-segmentation, since the number of objects in images varies. This paper introduces UnionCut, a robust and well-grounded foreground prior based on min-cut [2] and ensemble methods [18] that detects the union of foreground areas of an image, allowing UOD algorithms to identify foreground objects and stop discovery once the majority of the foreground union in the image is segmented. In addition, we propose UnionSeg, a distilled transformer of UnionCut that outputs the foreground union more efficiently and accurately. Our experiments show that by combining with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks. The code is available at https://github.com/YFaris/UnionCut." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 609, + 135, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 135, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 135, + 623 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 631, + 296, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 631, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 55, + 631, + 296, + 679 + ], + "type": "text", + "content": "Instance segmentation is a fundamental computer vision task requiring the algorithm to recognize and segment each instance in an image, which has been widely applied in autonomous driving [14], medical image analysis [48], re" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 213, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 317, + 213, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 317, + 213, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 317, + 213, + 553, + 322 + ], + "type": "image", + "image_path": "a7256056cee5522836befa7ab7fdc71a9c2b46db91fa09351ac94329987b6e63.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 354 + ], + "type": "text", + "content": "Figure 1. Examples of fixing MaskCut [69]'s errors and making its discovery stop at appropriate time using UnionCut and UnionSeg." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 472 + ], + "type": "text", + "content": "mote sensing [79], etc. However, pixel-level annotations are required to train an instance segmentation model, which are expensive and time-consuming to obtain. To cut down the training cost of instance segmentation models, strategies employing semi-supervision [12, 20, 29, 32, 43, 55, 56, 71], weak supervision [13, 33, 34, 39, 44, 57, 68, 72], and self-supervision [21, 35, 53, 67, 69] have been proposed to reduce the usage of pixel-level annotations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 475, + 556, + 714 + ], + "type": "text", + "content": "Self-supervision aims to train an instance segmentation model without any annotations. It utilizes an unsupervised object discovery (UOD) algorithm to detect and segment objects in images without handcrafted annotations, following which, the detection result is used as the ground truth to train a segmentation model [15, 16, 21, 26, 35, 67]. Recent state-of-the-art UOD algorithms [21, 35, 53, 67, 69] are based on a vision transformer (ViT) [19] pre-trained by self-supervised representation learning [9]. These draw on a phenomenon where the attention map of the last layer of the ViT indicates regions of objects in the image. However, without the foreground knowledge given by the ground truth, existing UOD methods cannot robustly judge whether the explored region belongs to the foreground or not, and as such, the algorithm could return the background area of the image as a discovered object. As a remedy, existing approaches introduce foreground priors (e.g. the location and size of the discovered region) to help them distinguish if an area in the image is foreground. However, we have found that existing foreground priors are not robust because they" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "\\*Author's Accepted Manuscript, Released under the Creative Commons license: Attribution 4.0 International (CC BY 4.0) https://creativecommons.org/licenses/by/4.0/ " + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": " Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20268" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "are based on simple rules, assumptions, and heuristic design. Besides, the number of objects in the image varies and is unknown to UOD approaches, resulting in a dilemma of when to stop the discovery. Thus, existing UOD algorithms either segment only one region as the object or continue discovery for a predetermined number of iterations, leading to either missing objects or mistakenly identifying background as foreground due to unreliable foreground priors and excessive exploration. For example, in Fig. 1, MaskCut [69], a UOD method, returns a background area (the 1st and 3rd row), and misses the car on the left of the 2nd image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 294, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 361 + ], + "type": "text", + "content": "To address the problem, we propose an ensemble method named UnionCut. It detects the union of foreground regions in an image, referred to as \"foreground union\" in this paper. With this foreground union as the foreground prior, UOD methods can accurately judge if a discovered region is the foreground by checking if the foreground union covers it, and stop further exploration when the majority of the foreground union has been discovered, as shown in Fig. 1. Unlike previous assumption-based foreground priors, UnionCut's robustness and reliability are demonstrated with mathematical and statistical ensemble theories. Besides, we propose UnionSeg, a ViT distilled from UnionCut to increase the efficiency and accuracy of foreground union detection." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 361, + 295, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 624 + ], + "type": "text", + "content": "Our work makes the following contributions: Insights: Previous UOD studies focused mainly on the design of UOD algorithms, overlooking foreground priors' importance. We reveal that UOD's performance can also be increased by robust foreground priors. Methods: We propose UnionCut based on ensemble and UnionSeg using distillation to detect foreground union in the image, serving as foreground priors for UOD methods to verify results and decide when to stop discovery. Unlike previous heuristic and assumption-based foreground priors, UnionCut is mathematically and statistically robust and reliable. Results: After being integrated with UnionCut or UnionSeg, previous state-of-the-art UOD methods [21, 53, 54, 69] witness an increase in the performance of single object discovery, saliency detection and self-supervised instance segmentation on various benchmarks, indicating that UnionCut and UnionSeg enable UOD algorithms to give more accurate predictions on the existence of objects in the image. Impacts: UnionCut and UnionSeg can be used as future UOD methods' default foreground priors, taking the place of existing assumption-based, heuristic, and unreliable foreground priors, and enhancing UOD performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 638, + 142, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 638, + 142, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 638, + 142, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 659, + 269, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 269, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 269, + 673 + ], + "type": "text", + "content": "2.1. Self-supervised Representation Learning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 294, + 713 + ], + "type": "text", + "content": "Self-supervised representation learning aims to train the model to generate distinguishable embeddings for images and output similar embeddings of images with the same se" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "mantic information without using handcrafted annotations. Most existing approaches [8-11, 25, 28, 42, 66, 76] use a student-teacher architecture, where each model processes one of two differently augmented versions (e.g. horizontal flipping, randomly cropping, colour jitter, etc.) of the same image. The optimization goal is to make the models produce similar outputs for augmentations of the same image and distinct outputs for augmentations from different images. DINO [9] is a representative method of these approaches, where the attention map of the last layer of a ViT pre-trained by DINO indicates the location of objects in the image. Such a property is employed in our approach." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 234, + 484, + 247 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 484, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 484, + 247 + ], + "type": "text", + "content": "2.2. Unsupervised Object Discovery" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 255, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 553, + 447 + ], + "type": "text", + "content": "As for UOD, Wang et al. [67] proposed FreeMask to discover objects by comparing the similarity between features from a self-supervised CNN [66] at each pixel of the image. Instead of using CNN, in [21, 53, 59, 69], embeddings of patches output by a ViT pre-trained by DINO[9] are utilized to generate pseudo-masks. In [53, 59], a patch is selected as the seed, and the pseudo mask is given by combining patches sharing similar features with the seed. To increase the pseudo-mask quality, Normalized Cut (NCut) [50] is leveraged by Wang et al. [21] to segment the feature map output by DINO for pseudo-mask generation. In [69], NCut is expanded to segment multiple objects in an image by iteratively removing a new segmented region from the image. However, the absence of ground truth makes it hard to determine if any discovered region is actually part of the foreground. Our approach provides a solution to this." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 465, + 424, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 424, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 424, + 477 + ], + "type": "text", + "content": "2.3. Foreground Priors" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": "UOD algorithms leverage foreground priors to judge if the discovered region belongs to the foreground. Specifically, Simeoni et al. [53] assumes that the area occupied by the foreground should be smaller than the background. FreeMask [67] removes discovered regions whose widths are longer than " + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": " of the width of the image or are located in the top " + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 486, + 553, + 713 + ], + "type": "text", + "content": " region of the image. Gansbeke et al. [59] and Simeoni et al. [54] proposed that the semantic information of the foreground should be close to the global semantic information of the entire image. In [21, 69], the region where the point with the highest absolute confidence value is located is made as the foreground. Wang et al. [69] suggested that the foreground is unlikely to cover three or four corners of the image. Li and Shin [35] proposed that a background mask is likely to occupy at least two edges of the image. Although these priors achieved limited success, they are all heuristic and assumption-based. In contrast, the robustness and reliability of our proposed methods are demonstrated by mathematical and statistical ensemble theories." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20269" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 77, + 71, + 534, + 220 + ], + "blocks": [ + { + "bbox": [ + 77, + 71, + 534, + 220 + ], + "lines": [ + { + "bbox": [ + 77, + 71, + 534, + 220 + ], + "spans": [ + { + "bbox": [ + 77, + 71, + 534, + 220 + ], + "type": "image", + "image_path": "380f4eb0f2e6ed120a1ba9d121c52c6e65b88b222142cad94d0ae4726457445e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 54, + 230, + 555, + 277 + ], + "type": "text", + "content": "Figure 2. Framework of UnionCut. UnionCut takes an image as input with features extracted from a frozen DINO, and applies Unit Voters (UVs) with each using one patch as its seed. Red rectangles refer to the seed patch, and blue areas indicate the UV's output of the region similar to its seed patch. Aggregating all UV's outputs yields a heat map of the background, which is then inverted, thresholded, and rectified with a corner prior to produce a binary mask of the image's foreground union." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 296, + 168, + 307 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 296, + 168, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 168, + 307 + ], + "type": "text", + "content": "2.4. Ensemble Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 296, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 296, + 482 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 296, + 482 + ], + "type": "text", + "content": "Ensemble methods are learning algorithms that construct a set of classifiers and then classify new data points by taking a (weighted) vote of their predictions [18]. Most ensemble methods can be classified into two types: 1) Parallel a set of weak classifiers run in parallel and form a strong classifier by performing a plurality vote, e.g. Pasting [5], Bagging [4], and Random Forest [6]; and 2) Sequential running batches of weak classifiers iteratively with each batch focusing on hard cases of the last one, e.g. [60] and AdaBoost [23]. UnionCut is of the former: a set of weak classifiers is created for an image, with each in parallel voting every patch " + }, + { + "bbox": [ + 55, + 313, + 296, + 482 + ], + "type": "inline_equation", + "content": "(8 \\times 8" + }, + { + "bbox": [ + 55, + 313, + 296, + 482 + ], + "type": "text", + "content": " pixels) of the image as foreground or background, which forms a strong classifier by aggregating votes from them to detect the foreground union of an image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 491, + 138, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 138, + 504 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 138, + 504 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 629 + ], + "type": "text", + "content": "UnionCut is an ensemble method, enabling UOD methods to judge if a detected region is foreground and stop discovery when most objects have been discovered to avoid over-segmentation. In the context of ensemble methods, UnionCut serves as the strong classifier, based on many parallel weak classifiers named \"unit voter\" (UV) here. Fig. 2 illustrates the frameworks of the proposed UnionCut. Unlike previous assumption-based foreground priors mentioned in Sec. 2.3, UnionCut is reliable and robust with ensemble theories. First, we introduce the details of UV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 128, + 647 + ], + "type": "text", + "content": "3.1. Unit Voter" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "UV is based on features output by a ViT pretrained by DINO [9]. In our implementation, an image is resized into " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "224 \\times 224" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " and serialized into " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " (i.e. 784) patches of " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": " pixels before feature extraction of each patch with the ViT. The set of patches of an image is defined as " + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "inline_equation", + "content": "P =" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 366, + 295, + 508, + 449 + ], + "blocks": [ + { + "bbox": [ + 366, + 295, + 508, + 449 + ], + "lines": [ + { + "bbox": [ + 366, + 295, + 508, + 449 + ], + "spans": [ + { + "bbox": [ + 366, + 295, + 508, + 449 + ], + "type": "image", + "image_path": "b4eb1719381731e08dc1656ddb5f8503c0d2a65289a6c27500513d98804d5ee1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 457, + 555, + 557 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 557 + ], + "type": "text", + "content": "Figure 3. Overview of Unit Voter (UV). For clarity, the image is divided into " + }, + { + "bbox": [ + 313, + 457, + 555, + 557 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 457, + 555, + 557 + ], + "type": "text", + "content": " patches. With the seed patch and anti-seed patches of the UV initialized, the image is modeled as a graph with each patch and two dummy terminals being nodes and connected by edges whose thickness reflects their weights. After applying min-cut to the graph, thin edges are removed to divide the graph into two disjoint parts containing one terminal node. The part with the Source node is the region output by the UV, indicating the area of features similar to the seed patch." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "inline_equation", + "content": "\\{p_1,p_2,\\dots ,p_{784}\\}" + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "content": " with the corresponding L2-normalized Key vectors as their features output by the last layer of the model, which is defined as " + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "inline_equation", + "content": "K = \\{k_{1},k_{2},\\ldots ,k_{784}\\}" + }, + { + "bbox": [ + 313, + 580, + 553, + 628 + ], + "type": "text", + "content": ". Then, the patches and features are leveraged by each UV." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": "As required by ensemble methods, the number of weak classifiers (i.e. UV) to form a strong classifier (i.e. Union-Cut) should be significant, and each UV should be different [18]. To achieve these goals, 784 independent UVs can be built based on the 784 patches of an image, with each UV corresponding to a unique patch " + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "inline_equation", + "content": "p_f \\in P" + }, + { + "bbox": [ + 313, + 630, + 555, + 714 + ], + "type": "text", + "content": ". The patch corresponding to a UV is called that UV's \"seed patch\"." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20270" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": "Taking a UV built from a seed patch " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_f \\in P" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": " as an example, its pipeline is illustrated in Fig. 3. Specifically, given a UV's seed patch " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_f \\in P" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": " with feature " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "k_f \\in K" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ", the set of patches dissimilar to " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ", defined as the UV's anti-seed patches set " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "B_f = \\{p_b | p_b \\in P, b \\neq f, k_b^T k_f < 0\\}" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ", can be obtained. With " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_f, B_f" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ", all " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": " patches " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "P" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": " and features " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ", a directed graph is constructed, as shown in Fig. 3, where each patch is a node and two dummy terminal nodes, named \"Source\" and \"Target\", are added. In the graph, nodes of adjacent patches (i.e. eight-neighborhood patches) are connected by bidirectional edges (named \"n-links\") with weights representing their similarity. Directed edges named \"t-links\" connect the Source to each patch node or each patch node to the Target, with weights based on the likelihood of a patch being more similar to the UV's seed patch " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": " or its anti-seed patches set " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "B_f" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ". The terms \"t-link\" and \"n-link\" originate from early image segmentation methods that modeled images as graphs [3, 49]. The purpose of building this graph is to split the image by removing edges, forming two disconnected subgraphs with each either containing the Source or the Target, while minimizing the total weight of the removed edges. The subgraph with the Source is expected to contain patches similar to the UV's seed patch, while the one with the Target includes patches similar to the anti-seed patches " + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "inline_equation", + "content": "B_f" + }, + { + "bbox": [ + 55, + 72, + 296, + 407 + ], + "type": "text", + "content": ". This bipartition process is also known as the min-cut problem, which is solved by UV using an efficient solution [2]. Next, we introduce the algorithms for assigning weights to n-links and t-links." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "content": "n-links The weight " + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "inline_equation", + "content": "w(p_i, p_j)" + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "content": " of the n-link between two adjacent patches " + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 407, + 296, + 433 + ], + "type": "text", + "content": " is given by Eq. (1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 81, + 447, + 294, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 447, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 81, + 447, + 294, + 480 + ], + "type": "interline_equation", + "content": "w \\left(p _ {i}, p _ {j}\\right) = \\frac {e ^ {- \\frac {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}{\\beta}}}{\\sqrt {\\left(i _ {x} - j _ {x}\\right) ^ {2} + \\left(i _ {y} - j _ {y}\\right) ^ {2}}} \\tag {1}", + "image_path": "382a926fbbbf32fa9f116d4b1f0da8bcd1267a33e3f9213acb3879e1d5905801.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 480, + 269, + 504 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 480, + 269, + 504 + ], + "spans": [ + { + "bbox": [ + 82, + 480, + 269, + 504 + ], + "type": "interline_equation", + "content": "\\beta = \\frac {2 \\sum_ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} (p _ {i} , p _ {j}) (k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}{C}", + "image_path": "21eeb6f43dedfc0d2ddad35208eef4e20ce25730454f6217d1541680ead66342.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "(i_x,i_y)" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "(j_{x},j_{y})" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " are the indices of the two patches under the scale of the feature map (i.e. " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "28\\times 28" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": ") output by the ViT. " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(p_i,p_j)" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " is an indicator function which returns 1 if the " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "p_j" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " are adjacent, otherwise 0. Finally, " + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": " is the number of n-links in the graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": "t-links An UV's graph has three kinds of patches: seed patch " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": ", anti-seed patches set " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "B_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": ", and the other patches " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "P - B_f \\cup \\{p_f\\}" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": ", whose weight assignment strategies for t-links are different. Suppose the Source and Target are defined as " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": ". After min-cut, the t-link from " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " should be kept while the t-link between " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " should be cut. To this end, a large weight " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " is given to the t-link from " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " so that it is too large to be cut, and the weight between " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " is 0. Following early graph cut methods [3, 49], " + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 570, + 296, + 689 + ], + "type": "text", + "content": " can be decided by Eq. (2)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 701, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 701, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 88, + 701, + 295, + 715 + ], + "type": "interline_equation", + "content": "W = 1 + \\max _ {p _ {i} \\in P} \\sum_ {p _ {j} \\in P} \\mathbf {1} \\left(p _ {i}, p _ {j}\\right) w \\left(p _ {i}, p _ {j}\\right) \\tag {2}", + "image_path": "fc2e3640f5b378f820f85f3c3b3b86d51f90a8032d20e7d3ef992e4fd9ffc32c.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "Similarly, anti-seed patches must stay connected with " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " after min-cut. Therefore, t-links from " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\forall p_{b}\\in B_{f}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " must have a large value, which can also be given by Eq. (2), while the weights of t-links from " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "inline_equation", + "content": "\\forall p_{b}\\in B_{f}" + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": " are 0." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "text", + "content": "Weights of other t-links between " + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "inline_equation", + "content": "p_i \\notin B_f \\bigcup \\{p_f\\}" + }, + { + "bbox": [ + 314, + 120, + 555, + 145 + ], + "type": "text", + "content": " are decided by Eq. (3)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 159, + 553, + 190 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 159, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 320, + 159, + 553, + 190 + ], + "type": "interline_equation", + "content": "w \\left(p _ {i}, S\\right) = - \\log \\frac {e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}}}{e ^ {\\sqrt {\\left(k _ {i} - k _ {f}\\right) ^ {T} \\left(k _ {i} - k _ {f}\\right)}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {\\left(k _ {i} - k _ {j}\\right) ^ {T} \\left(k _ {i} - k _ {j}\\right)}}} \\tag {3}", + "image_path": "184c9dd5eb961181f321cd530eaeb6f80716a7d71ab51f5014316a886890449b.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 321, + 187, + 535, + 211 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 187, + 535, + 211 + ], + "spans": [ + { + "bbox": [ + 321, + 187, + 535, + 211 + ], + "type": "interline_equation", + "content": "w (p _ {i}, T) = - \\log \\frac {e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}{e ^ {\\sqrt {(k _ {i} - k _ {f}) ^ {T} (k _ {i} - k _ {f})}} + e ^ {\\min _ {p _ {j} \\in B _ {f}} \\sqrt {(k _ {i} - k _ {j}) ^ {T} (k _ {i} - k _ {j})}}}", + "image_path": "8e331ee188042c809164ed5976f8b866da4796bf4af664ce596e48b28fc9d06b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": "Eq. (3) explains that if a patch " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "p_i \\notin B_f \\cup \\{p_f\\}" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " is more similar to anti-seed patches set " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "B_f" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " than the seed patch " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "w(p_i, T)" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " should exceed " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "w(p_i, S)" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": ", making the t-link from " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 216, + 554, + 262 + ], + "type": "text", + "content": " more likely to be cut, and vice versa." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 263, + 554, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 335 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 335 + ], + "type": "text", + "content": "With the graph initialized using the method introduced above, the UV executes a min-cut and returns a binary mask, where \"1\" indicates patches remaining connected to the Source, i.e. regions sharing features with the seed patch " + }, + { + "bbox": [ + 313, + 263, + 554, + 335 + ], + "type": "inline_equation", + "content": "p_f" + }, + { + "bbox": [ + 313, + 263, + 554, + 335 + ], + "type": "text", + "content": ". Therefore, the function of each UV is to return the region that is similar to its seed patch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 342, + 383, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 342, + 383, + 353 + ], + "spans": [ + { + "bbox": [ + 314, + 342, + 383, + 353 + ], + "type": "text", + "content": "3.2. UnionCut" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "text", + "content": "UnionCut is a strong classifier composed of 784 UVs as weak classifiers, with each UV using a unique patch from the image's " + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "text", + "content": " patches as its seed patch. Considering the function of UV introduced above, there are two types of UVs in UnionCut: 1) background UV: if its seed patch is background, the UV returns a binary mask indicating background regions; and 2) foreground UV: if its seed patch is foreground, the UV outputs a binary mask of foreground regions. UnionCut's output is determined by voting from the 784 UVs. Given the set of binary masks " + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "inline_equation", + "content": "M = \\{m_f | \\forall p_f \\in P\\}" + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "text", + "content": " returned by all UVs, a heat map " + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 358, + 555, + 503 + ], + "type": "text", + "content": " is generated by aggregating all UVs' output using Eq. (4)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 376, + 508, + 553, + 523 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 508, + 553, + 523 + ], + "spans": [ + { + "bbox": [ + 376, + 508, + 553, + 523 + ], + "type": "interline_equation", + "content": "A (x, y) = \\sum_ {m _ {f} \\in M} m _ {f} (x, y) \\tag {4}", + "image_path": "a16be1a27d200a567e322d275bc36ce1483d8cc65cafe9d488441f1868fd6934.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "inline_equation", + "content": "x \\in [1, 28]" + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "inline_equation", + "content": "y \\in [1, 28]" + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "text", + "content": " indicate the 2D coordinate of a patch's response on " + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 528, + 554, + 552 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "content": "For easier understanding, intuitively, the foreground usually occupies a smaller area than the background in the image [53]. Thus, we can assume that there are more background UVs than foreground UVs in UnionCut, causing background areas to have higher intensities on the aggregated heat map " + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "content": ". An example can be seen in Fig. 2. To make it intuitive, the brightness of " + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "content": " is inverted so that foreground regions have higher responses than the background, producing an inverted heat map " + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "content": " using Eq. (5). By thresholding " + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 552, + 555, + 683 + ], + "type": "text", + "content": ", most objects in the image are segmented as the foreground union." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 328, + 699, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 699, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 328, + 699, + 553, + 717 + ], + "type": "interline_equation", + "content": "H (x, y) = 2 5 5 - 2 5 5 \\times \\frac {A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)}{\\operatorname* {m a x} _ {x , y} A (x , y) - \\operatorname* {m i n} _ {x , y} A (x , y)} \\tag {5}", + "image_path": "8da041e83a31cee49070e67ac941fdfda24be4b0999479c34d10a456e85207f0.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20271" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 205 + ], + "type": "text", + "content": "Note that the assumption introduced above is for easier understanding of our approach. UnionCut is mathematically and statistically robust, with an explanation in Appendix 6. Moreover, UnionCut remains effective even when the foreground area is larger than the background, with examples in Appendix 7. The mathematical explanation also supports this claim. Furthermore, compared to more straightforward methods detecting patches similar to a given patch, e.g. matching with cosine similarity [53, 54], UV is more sophisticated. Appendix 8 provides a detailed analysis of UV's irreplaceable advantages and necessity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 141, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 141, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 141, + 220 + ], + "type": "text", + "content": "3.2.1. Thresholding" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "text", + "content": "A threshold for the bipartition of the heat map is required to estimate the foreground union. UnionCut automatically determines a threshold for the heat map " + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "text", + "content": " by clustering the response values on it with Mean-Shift [24]. Since the bright areas on the heat map represent the foreground, the resulting clusters are sorted by their centres in descending order, with the top half retained and identified as the union of the foreground in the image. Finally, a " + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "inline_equation", + "content": "28 \\times 28" + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "text", + "content": " binary mask " + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "inline_equation", + "content": "U_{cut}" + }, + { + "bbox": [ + 55, + 223, + 296, + 331 + ], + "type": "text", + "content": " with \"1\" as the foreground union is obtained." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 336, + 140, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 140, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 140, + 346 + ], + "type": "text", + "content": "3.2.2. Corner Prior" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "text", + "content": "Similar to [69], a corner prior is used by UnionCut to check the reliability of the foreground union mask. Specifically, if " + }, + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "inline_equation", + "content": "U_{cut}" + }, + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "text", + "content": " occupies four corners of the image, it must be inverted (i.e. " + }, + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "inline_equation", + "content": "U_{cut} := 1 - U_{cut}" + }, + { + "bbox": [ + 55, + 350, + 296, + 422 + ], + "type": "text", + "content": ") because the foreground union is unlikely to occupy four corners simultaneously. The reliability of this corner prior is discussed in Appendix 9." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 429, + 124, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 124, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 124, + 442 + ], + "type": "text", + "content": "3.3. UnionSeg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": "While UnionCut effectively detects an image's foreground union, running 784 UVs per image is computationally expensive. To improve efficiency, we introduce UnionSeg, a ViT-based model distilled from UnionCut as a surrogate model. Unlike UnionCut, UnionSeg is an end-to-end model that directly predicts the foreground union from an input image. To minimize learnable parameters, inspired by FOUND [54], UnionSeg uses a frozen ViT trained by DINO [9] as its backbone, followed by a learnable " + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": " convolutional layer and a sigmoid function, compressing each ViT patch feature into a scalar confidence score. Patches with scores above 0.5 are classified as the foreground union. Appendix 10 details the differences between UnionSeg and FOUND. Suppose the soft output of UnionSeg of an image is " + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "inline_equation", + "content": "U_{seg}^{s}" + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": ", with its thresholded hard output defined as " + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "inline_equation", + "content": "U_{seg}^{h}" + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": ". Given UnionCut's output " + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "inline_equation", + "content": "U_{cut}" + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": " of an image, the label " + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 446, + 296, + 651 + ], + "type": "text", + "content": " of an image used to train UnionSeg is defined by Eq. (6)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 656, + 295, + 684 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 656, + 295, + 684 + ], + "spans": [ + { + "bbox": [ + 102, + 656, + 295, + 684 + ], + "type": "interline_equation", + "content": "L = \\left\\{ \\begin{array}{l l} U _ {c u t}, & \\text {i f I o U} \\left(U _ {s e g} ^ {h}, U _ {c u t}\\right) < 0. 5, \\\\ U _ {s e g} ^ {h}, & \\text {o t h e r w i s e .} \\end{array} \\right. \\tag {6}", + "image_path": "0c9319c78376cbb4d900d048f1ca53eab6a1c32c57c4a00085116770e9c314a1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "IoU(\\cdot)" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " returns the intersection over union between two binary masks. The motivation behind Eq. (6) is that" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "UnionCut's output may contain errors, so UnionSeg does not rely entirely on UnionCut's output as labels. This enables UnionSeg to predict the foreground union more efficiently and accurately than UnionCut. The loss function used to train UnionSeg is shown in Eq. (7)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 327, + 152, + 554, + 168 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 152, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 327, + 152, + 554, + 168 + ], + "type": "interline_equation", + "content": "L o s s = L \\log U _ {s e g} ^ {s} + (1 - L) \\log \\left(1 - U _ {s e g} ^ {s}\\right) + 1 (\\text {i t e r} < 1 0 0) \\Lambda \\tag {7}", + "image_path": "90409e74e19aa9032c6c456485c1974cfab07a8797357758778e59ca9d5e0921.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 327, + 165, + 477, + 176 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 327, + 165, + 477, + 176 + ], + "spans": [ + { + "bbox": [ + 327, + 165, + 477, + 176 + ], + "type": "interline_equation", + "content": "\\Lambda = U _ {c u t} \\log U _ {s e g} ^ {s} + (1 - U _ {c u t}) \\log (1 - U _ {s e g} ^ {s})", + "image_path": "714631064e091efb3870787c174161288b3470ad1f2a4c6047840d2c096041e0.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "inline_equation", + "content": "\\mathbf{1}(iter < 100)" + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "content": " returns 1 if the training iterations are fewer than 100; otherwise 0. This ensures that at the start of training, when UnionSeg is not yet well trained, it is forced to use both UnionCut's output " + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "inline_equation", + "content": "U_{cut}" + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 186, + 555, + 246 + ], + "type": "text", + "content": " as labels to aid learning." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 258, + 395, + 272 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 395, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 395, + 272 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 278, + 555, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 278, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 313, + 278, + 555, + 398 + ], + "type": "text", + "content": "Several experiments are conducted to evaluate the effectiveness of UnionCut/UnionSeg, comparing the performance of state-of-the-art UOD methods before and after integrating UnionCut or UnionSeg on several tasks: unsupervised single object discovery (Sec. 4.1), unsupervised saliency detection (Sec. 4.2), and self-supervised instance segmentation (Sec. 4.3). We then discuss the properties of UnionCut and UnionSeg in Sec. 4.4, and evaluate the impact of different components of them in Sec. 4.5. Additionally, we provide more qualitative comparisons in Appendix 15." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 399, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 554, + 483 + ], + "type": "text", + "content": "Baseline UOD methods Our experiments use existing state-of-the-art UOD methods for single or multiple object discovery as baselines, i.e. LOST [53], TokenCut [21], FOUND [54] and CutLER [69]. To enhance their performance, we integrate them with UnionCut and UnionSeg by replacing their standard foreground priors with UnionCut/UnionSeg, with details provided in Appendix 11." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "type": "text", + "content": "Implementation Details We use a ViT-S/8 [19] pretrained by DINO [9] as the image feature extractor for UnionCut and UnionSeg. Following [52, 54], we distill UnionSeg from UnionCut using DUTS-TR[64], a saliency detection dataset with 10,553 images. Training is done with a batch size of 50, an initial learning rate of 0.05 that decays to " + }, + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 483, + 554, + 615 + ], + "type": "text", + "content": " every 50 iterations, and AdamW [38] for parameter updates over 600 iterations. Implementation details for all tested methods, i.e. LOST+UnionCut/UnionSeg, TokenCut+UnionCut/UnionSeg, FOUND+UnionCut/UnionSeg, and CutLER+UnionSeg, are provided in Appendix 11." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 623, + 515, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 623, + 515, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 623, + 515, + 636 + ], + "type": "text", + "content": "4.1. Unsupervised Single Object Discovery" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 677 + ], + "type": "text", + "content": "In this experiment, we combine baseline UOD methods with UnionCut and UnionSeg to see if their performance improves in unsupervised single object discovery." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Benchmark Following [21, 53, 54], the trainval sets of PASCAL VOC07 (5011 images), VOC12 (11540 images) [22] and COCO20K [37, 61] (19817 randomly sam" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20272" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 304, + 381 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 304, + 381 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 304, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 304, + 381 + ], + "type": "table", + "html": "
MethodVOC07VOC12COCO20K
- No learning -
Selective Search [58]18.820.916.0
EdgeBoxes [80]31.131.628.8
Kim et al. [30]43.946.435.1
Zhang et al. [75]46.250.534.8
DDT+ [70]50.253.138.2
rOSD [61]54.555.348.5
LOD [62]53.655.148.5
DINO-seg [9, 53](ViT-S/16 [9])45.846.242.0
DSS [40](ViT-S/16 [9])62.766.452.2
LOST [53](ViT-S/16 [9])61.964.050.7
LOST(ViT-S/16 [9])+UnionCut62.4(0.5↑)65.3(1.3↑)51.9(1.2↑)
LOST(ViT-S/16 [9])+UnionSeg63.2(1.3↑)67.1(3.1↑)54.9(4.2↑)
TokenCut [21](ViT-S/16 [9])68.872.158.8
TokenCut(ViT-S/16 [9])+UnionCut69.2(0.4↑)72.3(0.2↑)62.1(3.3↑)
TokenCut(ViT-S/16 [9])+UnionSeg69.7(0.9↑)72.7(0.6↑)62.6(3.8↑)
- With learning -
FreeSOLO [67, 69]44.049.735.2
LOD+CAD [21, 53]56.361.652.7
rOSD+CAD [21, 53]58.362.353.0
LOST+CAD [53](ViT-S/16 [9])65.770.457.5
SelfMask [52]72.375.362.7
FOUND [54](ViT-S/16 [9])72.576.162.9
FOUND(ViT-S/16 [9])‡71.475.762.9
FOUND(ViT-S/16 [9])+UnionCut72.6(1.2↑)76.1(0.4↑)63.0(0.1↑)
FOUND(ViT-S/16 [9])+UnionSeg72.3(0.9↑)75.9(0.2↑)63.1(0.2↑)
CutLER [69](ViT-B/8 [9])†73.369.570.7
CutLER(ViT-B/8 [9])+UnionSeg73.8(0.5↑)71.2(1.7↑)72.4(1.7↑)
", + "image_path": "34e5c2cfde20f9176fb09061c9c2d37162cc8019e9b3693f58dde9285d7b32aa.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 456 + ], + "type": "text", + "content": "Table 1. Performances of UOD methods on single object discovery with red indicates the highest score. CAD: a class-agnostic detector trained with the output of no-learning UOD methods as the ground truth. †: results of official implementations and checkpoints by our measurement. ‡: results of our replication with official implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 482, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 506 + ], + "type": "text", + "content": "pled images from COCO2014 [37]) are selected as the benchmark for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 508, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 508, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 508, + 295, + 616 + ], + "type": "text", + "content": "Metrics Similar to [21, 53, 54, 61, 62], we use CorLoc (Correct Localization) to evaluate UOD algorithms' performance on single object discovery. CorLoc scores an image as 1 if the intersection over union (IoU) between any ground truth bounding box and the algorithm's predicted bounding box exceeds 0.5, otherwise it scores 0. In our experiments, since CutLER can detect multiple objects, we select the bounding box with the highest confidence score as its single object prediction for the image." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Tab. 1 shows that after integrating UnionCut or UnionSeg, our selected baseline UOD methods improve across all three benchmarks, achieving state-of-the-art results, indicating that UnionCut/UnionSeg as a robust foreground prior enhances UOD algorithms' performance in single object discovery. In Appendix 12, we also evaluate these UOD methods from the perspective of the performance upper boundary proposed by Rambhatla et al. [45]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 491, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 491, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 491, + 85 + ], + "type": "text", + "content": "4.2. Unsupervised Saliency Detection" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 89, + 553, + 148 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 553, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 553, + 148 + ], + "type": "text", + "content": "We compare the performance of UOD methods before and after integrating UnionCut/UnionSeg on saliency detection, which involves segmenting salient objects in images. Since CutLER predicts multiple object masks per image, we use the union of these masks as its saliency detection result." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 148, + 553, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 553, + 196 + ], + "type": "text", + "content": "Benchmark Similar to [21, 40, 53, 54, 63], ECSSD [51] (1000 real-world images), DUTS-TE [64] (5019 images sampled from ImageNet test), and DUT-OMRON [74] (5168 natural images) are chosen as the benchmark." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "content": "Metrics Following [21, 53, 54], we report results with three metrics: Accuracy, IoU, and " + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\max F_{\\beta}" + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "content": ". Accuracy measures the percentage of correctly classified pixels compared to ground truth. IoU is the ratio of the intersection between the UOD output and ground truth to their union. " + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\max F_{\\beta}" + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "content": " is calculated as " + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\frac{(1 + \\beta^2)\\text{Precision} \\times \\text{Recall}}{\\beta^2\\text{Precision} + \\text{Recall}}" + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "content": " where Precision is the percentage of predicted pixels matching ground truth, and Recall is the proportion of ground truth pixels covered by the prediction. We set " + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 197, + 554, + 307 + ], + "type": "text", + "content": " to 0.3, as in [21, 53, 54]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 308, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 553, + 380 + ], + "type": "text", + "content": "Tab. 2 illustrates that the selected baseline methods improve when integrated with UnionCut/UnionSeg. TokenCut, with UnionSeg, performs best among no-learning approaches. And the performance of FOUND, the previous state-of-the-art method, is also increased by UnionCut and UnionSeg on most benchmarks and metrics." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 385, + 517, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 517, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 517, + 399 + ], + "type": "text", + "content": "4.3. Self-supervised Instance Segmentation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 554, + 534 + ], + "type": "text", + "content": "Some UOD methods target self-supervised instance segmentation, where models are trained using pseudo-annotations given by UOD algorithms, such as MaskCut for CutLER and FreeMask for FreeSOLO [67]. In the experiment, we explore whether UnionSeg improves CutLER's performance on instance segmentation. Similar to [69], COCO20k [37, 61] and COCO [37](5000 images from COCO2017's validation set) are used for evaluation, with average precision (AP) as the metric. Tab. 3 illustrates that by using UnionSeg to improve the pseudo-labels given by MaskCut, CutLER performs better on most metrics." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 534, + 554, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 554, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 554, + 666 + ], + "type": "text", + "content": "Since UnionCut with MaskCut takes much time to generate the foreground union and UOD results of all images in ImageNet (7 weeks in total, with 4 weeks for UnionCut and 3 weeks for MaskCut with 18 subprocesses running in parallel limited by our available computational resources), we only consider UnionSeg in this section. However, the effectiveness of UnionCut on self-supervised instance segmentation is also of interest. Therefore, we conduct another experiment that trains CutLER on a much smaller dataset (the subset of VOC12 containing 2913 images with pixel-level annotations). Details can be found in Appendix 13." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 515, + 685 + ], + "type": "text", + "content": "4.4. Properties of UnionCut and UnionSeg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "We explore the properties of UnionCut and UnionSeg. First, their accuracy in foreground union detection is evaluated." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20273" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 55, + 70, + 558, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 558, + 285 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 558, + 285 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 558, + 285 + ], + "type": "table", + "html": "
MethodDUT-OMRONDUTS-TEECSSD
Acc.IoUmaxFβAcc.IoUmaxFβAcc.IoUmaxFβ
- No learning -
HS [73]84.343.356.182.636.950.484.750.867.3
wCtr [77]83.841.654.183.539.252.286.251.768.4
WSC [36]86.538.752.386.238.452.885.249.868.3
DeepUSPS [41]77.930.541.477.330.542.579.544.058.4
BigBiGAN [63]85.645.354.987.849.860.889.967.278.2
E-BigBiGAN [63]86.046.456.388.251.162.490.668.479.7
LOST [53]79.741.047.387.151.861.189.565.475.8
DSS [40]-56.7--51.4--73.3-
MOST [45]8747.557.089.753.866.689.063.179.1
TokenCut [21, 54](ViT-S/16 [9])89.761.869.791.462.475.593.477.287.4
TokenCut(ViT-S/16 [9])+UnionCut89.662.4(0.6↑)71.0(1.3↑)91.6(0.2↑)63.2(0.8↑)80.2(4.7↑)93.8(0.4↑)77.3(0.1↑)91.0(3.6↑)
TokenCut(ViT-S/16 [9])+UnionSeg90.6(0.9↑)62.1(0.3↑)71.2(1.5↑)92.3(0.9↑)63.4(1.0↑)80.6(5.1↑)94.1(0.7↑)77.291.0(3.6↑)
- With learning -
LOST [53]81.848.957.888.757.269.791.672.383.7
SelfMask [52]91.965.5-93.366.0-95.581.8-
UnSeGArmaNet [46]----60.3--77.8-
CutLER [69](ViT-B/8 [9])†74.343.848.178.146.554.287.269.376.3
CutLER(ViT-B/8 [9])+UnionSeg87.7(13.4↑)57.8(14.0↑)64.1(16.0↑)90.1(12.0↑)61.4(14.9↑)72.3(18.1↑)94.0(6.8↑)79.6(10.3↑)87.8(11.5↑)
FOUND [54](ViT-S/16 [9])92.261.370.894.266.376.395.181.393.5
FOUND(ViT-S/16 [9])‡91.860.869.494.066.075.495.081.093.0
FOUND(ViT-S/16 [9])+UnionCut92.4(0.6↑)61.8(1.0↑)71.2(1.8↑)94.4(0.4↑)66.7(0.7↑)76.7(1.3↑)94.980.693.3(0.3↑)
FOUND(ViT-S/16 [9])+UnionSeg92.5(0.7↑)61.7(0.9↑)71.7(2.3↑)94.4(0.4↑)66.6(0.6↑)77.1(1.7↑)94.780.293.4(0.4↑)
", + "image_path": "c6bda0a9d5e27c7d32a84618f57f85d60a368695e21cb015d6aad2b036d211fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 325, + 556, + 430 + ], + "blocks": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Table 2. Performances of UOD methods on saliency detection with red indicates the highest score. †: results of official codes and checkpoints by our measurement. ‡: results of our replication with official implementation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 325, + 556, + 430 + ], + "lines": [ + { + "bbox": [ + 56, + 325, + 556, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 325, + 556, + 430 + ], + "type": "table", + "html": "
MethodsPretrainDetectorInit.COCO20KCOCO val2017
\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)\\(AP_{50}^{box}\\)\\(AP_{75}^{box}\\)\\(AP^{box}\\)\\(AP_{50}^{mask}\\)\\(AP_{75}^{mask}\\)\\(AP^{mask}\\)
LOST [53]IN+COCOFRCNN [47]DINO---2.41.01.1------
MaskDistill [59]IN+COCOMRCNN [27]MoCo [28]---6.82.12.9------
FreeSOLO [67, 69]IN+COCOSOLOv2 [65]DenseCL [66]9.73.24.19.73.44.39.63.14.29.43.34.3
DETReg [1]INDDETR [78]SwAV [8]------3.10.61.08.81.93.3
DINO [9]IN-DINO1.70.10.3---------
ProMerge [35]IN-DINO-----3.0-----2.4
ProMerge+ [35]INCascade [7]DINO-----8.9-----9.0
CutLER [69]INCascade [7]DINO22.412.511.919.610.09.221.911.812.318.99.79.2
CutLER+UnionSegINCascade [7]DINO24.112.212.920.39.110.023.712.212.919.79.09.9
", + "image_path": "08de503d4bfad9deadc3f7c7396e34934edaa6d52dad5e6393c4ef40d850da99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 555, + 460 + ], + "type": "text", + "content": "Table 3. Performances of self-supervised instance segmentation. The dataset (Pretrain) used to train each model (Detector) and each model's initialization (Init) are also listed. IN denotes ImageNet." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 480, + 295, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 576 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 576 + ], + "type": "text", + "content": "Baseline Li et al. [35] proposes ProMerge, which contains a foreground union detection method. FOUND [54] is a UOD method; however, considering that it includes an algorithm for detecting background regions in an image, we extract this algorithm separately as a baseline and treat the inversion of its output as the foreground union detection result. For convenience, we still refer to their foreground union detection algorithms as ProMerge and FOUND." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 578, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 295, + 628 + ], + "type": "text", + "content": "Benchmark The subset of VOC12 with 2913 images of pixel-level annotations is used as the benchmark, and the ground truth mask of each instance is combined as the ground truth of the foreground union in an image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 714 + ], + "type": "text", + "content": "Metrics IoU, precision, and recall are used to measure the performance. Besides, since representative datasets [22, 37] label objects of a given class list, there may be cases where some objects in the image are not labelled and thus are not considered part of the foreground union by the ground truth (Examples can be seen in Appendix 14). Therefore, directly using these metrics and dataset-provided" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 555, + 624 + ], + "type": "text", + "content": "labels to measure the performance of foreground union detection is inaccurate. To address this, we propose a fuzzy evaluation metric similar to CorLoc, named CorUnion. Specifically, we compare the foreground union output by UnionCut or UnionSeg with the ground truth for a given image and compute a specific metric (e.g. IoU). If this metric exceeds a certain threshold, we consider the foreground union detection on that image successful. The success rate across the dataset is then used as the CorUnion value. This approach also allows us to study the properties of UnionCut and UnionSeg by using different metrics or adjusting the threshold for successful detection." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Foreground Union Accuracy Tab. 4 shows the performance of UnionCut and UnionSeg, measured by precision, recall, and IoU. We also report CorUnion using IoU as the metric to judge an image's foreground union detection success, with increasing IoU threshold from 0.5 to 0.9. As shown in Fig. 4, both UnionCut and UnionSeg are more accurate than the baselines at foreground union detection." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20274" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 70, + 276, + 119 + ], + "blocks": [ + { + "bbox": [ + 77, + 70, + 276, + 119 + ], + "lines": [ + { + "bbox": [ + 77, + 70, + 276, + 119 + ], + "spans": [ + { + "bbox": [ + 77, + 70, + 276, + 119 + ], + "type": "table", + "html": "
MethodPrecisionRecallIoU
FOUND69.281.157.9
ProMerge-73.4-78.4-59.9
UnionCut-70.7-84.6-60.9
UnionSeg80.978.665.7
", + "image_path": "6fad0d611a35740c56412ae32614045610d9f9b0f4aa6722cf113fed6a838cf6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 75, + 150, + 276, + 219 + ], + "blocks": [ + { + "bbox": [ + 75, + 150, + 276, + 219 + ], + "lines": [ + { + "bbox": [ + 75, + 150, + 276, + 219 + ], + "spans": [ + { + "bbox": [ + 75, + 150, + 276, + 219 + ], + "type": "image", + "image_path": "063578444ed6b0d4ebf754a8379dfd52974d81605a6a7404dd16912c2677489d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 228, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 228, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 295, + 240 + ], + "type": "text", + "content": "Figure 4. Foreground union detection measured by CorUnion." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 261, + 295, + 501 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 261, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 295, + 501 + ], + "type": "text", + "content": "Instances Count In addition, we assessed UnionCut and UnionSeg's performance across varying instance counts in VOC12, categorized into groups based on the number of instances per image. As shown in Fig. 5a, UnionSeg, a training-based method, outperforms training-free approaches in precision, supporting that self-training corrects initial errors made by training-free methods. This finding aligns with previous research [21, 53, 54]. As instance numbers increase, UnionSeg maintains its precision, while other methods decline. Furthermore, Fig. 5b shows that UnionCut surpasses other methods in recall of the foreground union, remaining stable as the number of instances rises, unlike the decrease seen with other methods. This demonstrates the functional distinction between UnionCut and other training-free methods: as the number of instances per image increases, UnionCut consistently predicts the foreground union covering most instances, though precision inevitably decreases. This consistency highlights UnionCut's effectiveness as a robust prior for UOD algorithms to cease exploration optimally." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 501, + 295, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 561 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 561 + ], + "type": "text", + "content": "Overall, UnionSeg outperforms UnionCut and other baseline methods (Fig. 4) with higher precision (Fig. 5a), making it better suited for identifying if a detected area is foreground. UnionCut, with its higher recall (Fig. 5b), is better for determining if the discovery is complete." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 571, + 293, + 668 + ], + "blocks": [ + { + "bbox": [ + 59, + 571, + 293, + 668 + ], + "lines": [ + { + "bbox": [ + 59, + 571, + 293, + 668 + ], + "spans": [ + { + "bbox": [ + 59, + 571, + 293, + 668 + ], + "type": "image", + "image_path": "e5632f2f1a552857653fe6a8a559047ee5c9cadffe68d96fc7f51687ef9e88b4.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "text", + "content": "Figure 5. Performance of foreground union detection on images with varying instance numbers, measured by CorUnion with precision " + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "inline_equation", + "content": ">0.8" + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "text", + "content": " (a) and recall " + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "inline_equation", + "content": ">0.8" + }, + { + "bbox": [ + 55, + 677, + 295, + 710 + ], + "type": "text", + "content": " (b) as success criteria." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 408, + 85 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 185 + ], + "type": "text", + "content": "UnionCut consists of two elements: aggregated UVs and corner prior, while UnionSeg builds on UnionCut with distillation. To evaluate each element's contribution, TokenCut [21] is used as the baseline, with each element integrated for single object discovery. Benchmarks and metrics are inherited from Sec. 4.1. As shown in Tab. 5, each element contributes to the performance increase of UOD, demonstrating their effectiveness." + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 323, + 196, + 547, + 256 + ], + "blocks": [ + { + "bbox": [ + 56, + 127, + 294, + 140 + ], + "lines": [ + { + "bbox": [ + 56, + 127, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 127, + 294, + 140 + ], + "type": "text", + "content": "Table 4. Average performance on foreground union detection." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 323, + 196, + 547, + 256 + ], + "lines": [ + { + "bbox": [ + 323, + 196, + 547, + 256 + ], + "spans": [ + { + "bbox": [ + 323, + 196, + 547, + 256 + ], + "type": "table", + "html": "
MethodaUcpdVOC07VOC12COCO20K
TokenCut [21]68.872.158.8
TokenCut+aU✓69.072.362.0
TokenCut+UnionCut✓✓69.272.362.1
TokenCut+UnionSeg✓✓✓69.772.762.6
", + "image_path": "11b88d7a025da918a8a4e142630aec93e5382382d64722dac29127a14b09493d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 311, + 555, + 384 + ], + "type": "text", + "content": "We further evaluate the inference speed of UnionCut and UnionSeg. We include FOUND and ProMerge in Sec. 4.4 as the baseline. Our implementation is based on Python 3.11.9, Pytorch 2.3.1 and CUDA 12.1. As shown in Tab. 6, UnionSeg significantly increases the efficiency of foreground union detection compared with UnionCut." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 322, + 393, + 547, + 451 + ], + "blocks": [ + { + "bbox": [ + 313, + 264, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 313, + 264, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 264, + 555, + 298 + ], + "type": "text", + "content": "Table 5. Contribution of each component of UnionCut and UnionSeg. aU, cp, and d denote aggregated UVs, corner prior and distillation, respectively." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 322, + 393, + 547, + 451 + ], + "lines": [ + { + "bbox": [ + 322, + 393, + 547, + 451 + ], + "spans": [ + { + "bbox": [ + 322, + 393, + 547, + 451 + ], + "type": "table", + "html": "
MethodCPUGPUFPS
FOUNDIntel 14700KFNVIDIA RTX4070Ti Super2
ProMergeIntel 14700KFNVIDIA RTX4070Ti Super3
UnionCutIntel 14700KFNVIDIA RTX4070Ti Super0.1
UnionSegIntel 14700KFNVIDIA RTX4070Ti Super125
", + "image_path": "2a8206581eb91e6f9abc327bc23eaa5d24f66e5ae18f2dd226ac273a529d1f12.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 458, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 313, + 458, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 555, + 492 + ], + "type": "text", + "content": "Table 6. Inference speed of UnionSeg and UnionCut. FPS (frames per second) is evaluated throughout the whole pipeline from the image input to the output of the foreground union." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 517, + 388, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 388, + 529 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 388, + 529 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 536, + 556, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 556, + 704 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 556, + 704 + ], + "type": "text", + "content": "This paper introduces UnionCut, an unsupervised ensemble method to robustly predict foreground unions in images. Additionally, we propose UnionSeg distilled from UnionCut to detect foreground unions more accurately and efficiently. Both of them help UOD algorithms determine if discovered areas are in the foreground and when to stop discovery on an image, enhancing UOD's accuracy. While both are effective, each has unique advantages, and the choice between them depends on the specific UOD requirements for precision, recall, and efficiency. Our experiments show that integrating UnionCut or UnionSeg improves the performance of state-of-the-art UOD methods in single object discovery, saliency detection, and instance segmentation, demonstrating their effectiveness as robust foreground priors for unsupervised object discovery." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20275" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 156 + ], + "type": "text", + "content": "[1] Amir Bar, Xin Wang, Vadim Kantorov, Colorado J Reed, Roei Herzig, Gal Chechik, Anna Rohrbach, Trevor Darrell, and Amir Globerson. Detreg: Unsupervised pretraining with region priors for object detection. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14605-14615, 2022. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 201 + ], + "type": "text", + "content": "[2] Yuri Boykov and Vladimir Kolmogorov. An experimental comparison of min-cut/max-flow algorithms for energy minimization in vision. IEEE transactions on pattern analysis and machine intelligence, 26(9):1124-1137, 2004. 1, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 62, + 202, + 295, + 255 + ], + "type": "text", + "content": "[3] Yuri Y Boykov and M-P Jolly. Interactive graph cuts for optimal boundary & region segmentation of objects in nd images. In Proceedings eighth IEEE international conference on computer vision. ICCV 2001, pages 105-112. IEEE, 2001. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 294, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 278 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 278 + ], + "type": "text", + "content": "[4] Leo Breiman. Bagging predictors. Machine learning, 24: 123-140, 1996. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 280, + 295, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 280, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 62, + 280, + 295, + 312 + ], + "type": "text", + "content": "[5] Leo Breiman. Pasting small votes for classification in large databases and on-line. Machine learning, 36:85-103, 1999. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 314, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 314, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 314, + 294, + 334 + ], + "type": "text", + "content": "[6] Leo Breiman. Random forests. Machine learning, 45:5-32, 2001. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 295, + 380 + ], + "type": "text", + "content": "[7] Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delving into high quality object detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6154-6162, 2018. 7, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 434 + ], + "type": "text", + "content": "[8] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33:9912-9924, 2020. 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 436, + 295, + 491 + ], + "type": "text", + "content": "[9] Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9650-9660, 2021. 1, 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 295, + 535 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pages 1597-1607. PMLR, 2020." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 536, + 295, + 579 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "text", + "content": "[12] Xiaokang Chen, Yuhui Yuan, Gang Zeng, and Jingdong Wang. Semi-supervised semantic segmentation with cross pseudo supervision. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2613-2622, 2021. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 690 + ], + "type": "text", + "content": "[13] Zhaozheng Chen, Tan Wang, Xiongwei Wu, Xian-Sheng Hua, Hanwang Zhang, and Qianru Sun. Class re-activation maps for weakly-supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 969–978, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[14] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 117 + ], + "type": "text", + "content": "Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proc. of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 119, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 174 + ], + "type": "text", + "content": "[15] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning from video to detect foreground objects in single images. In Proceedings of the IEEE International Conference on Computer Vision, pages 4335-4343, 2017. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 176, + 553, + 219 + ], + "type": "text", + "content": "[16] Ioana Croitoru, Simion-Vlad Bogolin, and Marius Leordeanu. Unsupervised learning of foreground object segmentation. International Journal of Computer Vision, 127:1279-1302, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 300 + ], + "type": "text", + "content": "[18] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 1, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 303, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 369 + ], + "type": "text", + "content": "[19] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2020. 1, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[20] Rosana El Jurdi, Caroline Petitjean, Paul Honeine, and Fahed Abdallah. Bb-unet: U-net with bounding box prior. IEEE Journal of Selected Topics in Signal Processing, 14(6):1189-1198, 2020. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "text", + "content": "[21] Y. Wang et al. Self-supervised transformers for unsupervised object discovery using normalized cut. In CVPR, 2022. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "text", + "content": "[22] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. International journal of computer vision, 88(2):303-338, 2010. 5, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 541 + ], + "type": "text", + "content": "[23] Yoav Freund and Robert E Schapire. A decision-theoretic generalization of on-line learning and an application to boosting. Journal of computer and system sciences, 55(1): 119-139, 1997. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 587 + ], + "type": "text", + "content": "[24] Keinosuke Fukunaga and Larry Hostetler. The estimation of the gradient of a density function, with applications in pattern recognition. IEEE Transactions on information theory, 21(1):32-40, 1975. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 656 + ], + "type": "text", + "content": "[25] Jean-Bastien Grill, Florian Strub, Florent Alché, Corentin Tallec, Pierre Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Avila Pires, Zhaohan Guo, Mohammad Gheshlaghi Azar, et al. Bootstrap your own latent-a new approach to self-supervised learning. Advances in neural information processing systems, 33:21271-21284, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[26] Emanuela Haller, Adina Magda Florea, and Marius Leordeanu. Iterative knowledge exchange between deep learning and space-time spectral clustering for unsupervised segmentation in videos. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):7638-7656, 2021. 1" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20276" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[27] Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask r-cnn. In Proceedings of the IEEE international conference on computer vision, pages 2961-2969, 2017. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 160 + ], + "type": "text", + "content": "[28] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9729-9738, 2020. 2, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 160, + 294, + 205 + ], + "type": "text", + "content": "[29] Wei-Chih Hung, Yi-Hsuan Tsai, Yan-Ting Liou, Yen-Yu Lin, and Ming-Hsuan Yang. Adversarial learning for semi-supervised semantic segmentation. arXiv preprint arXiv:1802.07934, 2018. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 239 + ], + "type": "text", + "content": "[30] Gunhee Kim and Antonio Torralba. Unsupervised detection of regions of interest using iterative link analysis. Advances in neural information processing systems, 22, 2009. 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 270 + ], + "type": "text", + "content": "[31] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 294, + 326 + ], + "type": "text", + "content": "[32] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4015-4026, 2023. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 382 + ], + "type": "text", + "content": "[33] Viveka Kulharia, Siddhartha Chandra, Amit Agrawal, Philip Torr, and Ambrish Tyagi. Box2seg: Attention weighted loss and discriminative feature learning for weakly supervised segmentation. In European Conference on Computer Vision, pages 290-308. Springer, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[34] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2643-2652, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 469 + ], + "type": "text", + "content": "[35] Dylan Li and Gyungin Shin. Promerge: Prompt and merge for unsupervised instance segmentation. In ECCV, 2024. 1, 2, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 514 + ], + "type": "text", + "content": "[36] Nianyi Li, Bilin Sun, and Jingyi Yu. A weighted sparse coding framework for saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5216-5223, 2015. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 515, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 515, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 515, + 294, + 590 + ], + "type": "text", + "content": "[37] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 740-755. Springer, 2014. 5, 6, 7, 1, 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 613 + ], + "type": "text", + "content": "[38] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 647 + ], + "type": "text", + "content": "[39] R Austin McEver and BS Manjunath. Pcams: Weakly supervised semantic segmentation using point supervision. arXiv preprint arXiv:2007.05615, 2020. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 712 + ], + "type": "text", + "content": "[40] Luke Melas-Kyriazi, Christian Rupprecht, Iro Laina, and Andrea Vedaldi. Deep spectral methods: A surprisingly strong baseline for unsupervised semantic segmentation and localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8364-8375, 2022. 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 711 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 128 + ], + "type": "text", + "content": "[41] Tam Nguyen, Maximilian Dax, Chaithanya Kumar Mummadi, Nhung Ngo, Thi Hoai Phuong Nguyen, Zhongyu Lou, and Thomas Brox. Deepusps: Deep robust unsupervised saliency prediction via self-supervision. Advances in Neural Information Processing Systems, 32, 2019. 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 315, + 129, + 553, + 194 + ], + "type": "text", + "content": "[42] Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy V Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 196, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 196, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 553, + 251 + ], + "type": "text", + "content": "[43] Yassine Ouali, Céline Hudelot, and Myriam Tami. Semi-supervised semantic segmentation with cross-consistency training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12674-12684, 2020. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 252, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 252, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 315, + 252, + 553, + 306 + ], + "type": "text", + "content": "[44] Rui Qian, Yunchao Wei, Honghui Shi, Jiachen Li, Jiaying Liu, and Thomas Huang. Weakly supervised scene parsing with point-based distance metric learning. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 8843-8850, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 315, + 308, + 553, + 363 + ], + "type": "text", + "content": "[45] Sai Saketh Rambhatla, Ishan Misra, Rama Chellappa, and Abhinav Shrivastava. Most: Multiple object localization with self-supervised transformers for object discovery. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15823-15834, 2023. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 365, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 365, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 315, + 365, + 553, + 419 + ], + "type": "text", + "content": "[46] Kovvuri Sai Gopal Reddy, Bodduluri Saran, A. Mudit Adityaja, Saurabh J. Shigwan, Nitin Kumar, and Snehasis Mukherjee. Unsegarmanet: Unsupervised image segmentation using graph neural networks with convolutionalarma filters, 2024.7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 421, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 421, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 315, + 421, + 553, + 464 + ], + "type": "text", + "content": "[47] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 466, + 553, + 532 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 466, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 315, + 466, + 553, + 532 + ], + "type": "text", + "content": "[48] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical Image Computing and Computer-Assisted Intervention-MICCAI 2015: 18th International Conference, Munich, Germany, October 5-9, 2015, Proceedings, Part III 18, pages 234-241. Springer, 2015. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 533, + 553, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 533, + 553, + 577 + ], + "spans": [ + { + "bbox": [ + 315, + 533, + 553, + 577 + ], + "type": "text", + "content": "[49] Carsten Rother, Vladimir Kolmogorov, and Andrew Blake. \"grabcut\" interactive foreground extraction using iterated graph cuts. ACM transactions on graphics (TOG), 23(3): 309-314, 2004. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 578, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 578, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 315, + 578, + 553, + 611 + ], + "type": "text", + "content": "[50] Jianbo Shi and Jitendra Malik. Normalized cuts and image segmentation. IEEE Transactions on pattern analysis and machine intelligence, 22(8):888-905, 2000. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 315, + 613, + 553, + 656 + ], + "type": "text", + "content": "[51] Jianping Shi, Qiong Yan, Li Xu, and Jiaya Jia. Hierarchical image saliency detection on extended cssd. IEEE transactions on pattern analysis and machine intelligence, 38(4): 717-729, 2015. 6, 1, 2, 3, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 658, + 553, + 711 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 658, + 553, + 711 + ], + "spans": [ + { + "bbox": [ + 315, + 658, + 553, + 711 + ], + "type": "text", + "content": "[52] Gyungin Shin, Samuel Albanie, and Weidi Xie. Unsupervised salient object detection with spectral cluster voting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3971-3980, 2022. 5, 6, 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20277" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[53] Oriane Simeoni, Gilles Puy, Huy V Vo, Simon Roburin, Spyros Gidaris, Andrei Bursuc, Patrick Pérez, Renaud Marlet, and Jean Ponce. Localizing objects with self-supervised transformers and no labels. In BMVC 2021-32nd British Machine Vision Conference, 2021. 1, 2, 4, 5, 6, 7, 8, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 194 + ], + "type": "text", + "content": "[54] Oriane Simeoni, Chloe Sekkat, Gilles Puy, Antonín Vobecký, Éloi Zablocki, and Patrick Pérez. Unsupervised object localization: Observing the background to discover objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3176-3186, 2023. 1, 2, 5, 6, 7, 8, 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 194, + 294, + 239 + ], + "type": "text", + "content": "[55] Nasim Souly, Concetto Spampinato, and Mubarak Shah. Semi supervised semantic segmentation using generative adversarial network. In Proceedings of the IEEE international conference on computer vision, pages 5688-5696, 2017. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 282 + ], + "type": "text", + "content": "[56] Antti Tarvainen and Harri Valpola. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Advances in neural information processing systems, 30, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[57] Zhi Tian, Chunhua Shen, Xinlong Wang, and Hao Chen. Boxinst: High-performance instance segmentation with box annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5443-5452, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 381 + ], + "type": "text", + "content": "[58] Jasper RR Uijlings, Koen EA Van De Sande, Theo Gevers, and Arnold WM Smeulders. Selective search for object recognition. International journal of computer vision, 104: 154-171, 2013. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 294, + 426 + ], + "type": "text", + "content": "[59] Wouter Van Gansbeke, Simon Vandenhende, and Luc Van Gool. Discovering object masks with transformers for unsupervised semantic segmentation. arXiv preprint arXiv:2206.06363, 2022. 1, 2, 7, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 480 + ], + "type": "text", + "content": "[60] Paul Viola and Michael Jones. Rapid object detection using a boosted cascade of simple features. In Proceedings of the 2001 IEEE computer society conference on computer vision and pattern recognition. CVPR 2001, pages I-I. IEEE, 2001. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 481, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 547 + ], + "type": "text", + "content": "[61] Huy V Vo, Patrick Pérez, and Jean Ponce. Toward unsupervised, multi-object discovery in large-scale image collections. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXIII 16, pages 779-795. Springer, 2020. 5, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 591 + ], + "type": "text", + "content": "[62] Van Huy Vo, Elena Sizikova, Cordelia Schmid, Patrick Pérez, and Jean Ponce. Large-scale unsupervised object discovery. Advances in Neural Information Processing Systems, 34:16764-16778, 2021. 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[63] Andrey Voynov, Stanislav Morozov, and Artem Babenko. Object segmentation without labels with large-scale generative models. In International Conference on Machine Learning, pages 10596-10606. PMLR, 2021. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 691 + ], + "type": "text", + "content": "[64] Lijun Wang, Huchuan Lu, Yifan Wang, Mengyang Feng, Dong Wang, Baocai Yin, and Xiang Ruan. Learning to detect salient objects with image-level supervision. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 136-145, 2017. 5, 6, 1, 2, 3, 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[65] Xinlong Wang, Rufeng Zhang, Tao Kong, Lei Li, and Chunhua Shen. Solov2: Dynamic and fast instance segmenta" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tion. Advances in Neural information processing systems, 33:17721-17732, 2020. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 149 + ], + "type": "text", + "content": "[66] Xinlong Wang, Rufeng Zhang, Chunhua Shen, Tao Kong, and Lei Li. Dense contrastive learning for self-supervised visual pre-training. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3024-3033, 2021. 1, 2, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 150, + 553, + 215 + ], + "type": "text", + "content": "[67] Xinlong Wang, Zhiding Yu, Shalini De Mello, Jan Kautz, Anima Anandkumar, Chunhua Shen, and Jose M Alvarez. Freesolo: Learning to segment objects without annotations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14176-14186, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 216, + 553, + 271 + ], + "type": "text", + "content": "[68] Xuehui Wang, Kai Zhao, Ruixin Zhang, Shouhong Ding, Yan Wang, and Wei Shen. Contrastmask: Contrastive learning to segment every thing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11604-11613, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 327 + ], + "type": "text", + "content": "[69] Xudong Wang, Rohit Girdhar, Stella X Yu, and Ishan Misra. Cut and learn for unsupervised object detection and instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3124-3134, 2023. 1, 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 371 + ], + "type": "text", + "content": "[70] Xiu-Shen Wei, Chen-Lin Zhang, Jianxin Wu, Chunhua Shen, and Zhi-Hua Zhou. Unsupervised object discovery and co-localization by deep descriptor transformation. Pattern Recognition, 88:113-126, 2019. 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 415 + ], + "type": "text", + "content": "[71] Yulin Wu, Chang Liu, Lei Chen, Dong Zhao, Qinghe Zheng, and Hongchao Zhou. Perturbation consistency and mutual information regularization for semi-supervised semantic segmentation. Multimedia Systems, 29(2):511-523, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 470 + ], + "type": "text", + "content": "[72] Lian Xu, Wanli Ouyang, Mohammed Bennamoun, Farid Boussaid, Ferdous Sohel, and Dan Xu. Leveraging auxiliary tasks with affinity learning for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 6984-6993, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 514 + ], + "type": "text", + "content": "[73] Qiong Yan, Li Xu, Jianping Shi, and Jiaya Jia. Hierarchical saliency detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1155-1162, 2013. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 568 + ], + "type": "text", + "content": "[74] Chuan Yang, Lihe Zhang, Huchuan Lu, Xiang Ruan, and Ming-Hsuan Yang. Saliency detection via graph-based manifold ranking. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3166-3173, 2013. 6, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[75] Runsheng Zhang, Yaping Huang, Mengyang Pu, Jian Zhang, Qingji Guan, Qi Zou, and Haibin Ling. Object discovery from a single unlabeled image by mining frequent itemsets with multi-scale features. IEEE Transactions on Image Processing, 29:8606-8621, 2020. 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 669 + ], + "type": "text", + "content": "[76] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. Image bert pre-training with online tokenizer. In International Conference on Learning Representations, 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[77] Wangjiang Zhu, Shuang Liang, Yichen Wei, and Jian Sun. Saliency optimization from robust background detection. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2814-2821, 2014. 7" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20278" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[78] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 171 + ], + "type": "text", + "content": "[79] Xiao Xiang Zhu, Devis Tuia, Lichao Mou, Gui-Song Xia, Liangpei Zhang, Feng Xu, and Friedrich Fraundorfer. Deep learning in remote sensing: A comprehensive review and list of resources. IEEE geoscience and remote sensing magazine, 5(4):8-36, 2017. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[80] C Lawrence Zitnick and Piotr Dólar. Edge boxes: Locating object proposals from edges. In Computer Vision-ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13, pages 391-405. Springer, 2014. 6, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "20279" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_content_list.json b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7deb8347041d581cf34199e004d13e5a175d52a9 --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_content_list.json @@ -0,0 +1,1771 @@ +[ + { + "type": "text", + "text": "Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment", + "text_level": 1, + "bbox": [ + 99, + 130, + 898, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Renye Yan $^{1*}$ , Jikang Cheng $^{1*}$ , Yaozhong Gan $^{4}$ , Shikun Sun $^{3}$ , You Wu $^{4}$ , Yunfan Yang $^{1}$ , Ling Liang $^{1}$ , Jinlong Lin $^{1}$ , Yeshuang Zhu $^{2\\dagger}$ , Jie Zhou $^{2}$ , Jinchao Zhang $^{2}$ , Junliang Xing $^{3}$ , Yimao Cai $^{1\\dagger}$ , Ru Huang $^{1}$", + "bbox": [ + 106, + 179, + 887, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Peking University $^{1}$ , Tencent Inc $^{2}$ , Tsinghua University $^{3}$ , Nanjing University $^{4}$ .", + "bbox": [ + 267, + 218, + 728, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "victory@stu.pku.edu.cn, jikangcheng99@gmail.com, yshzhu@tencent.com, caiyimao@pu.edu.cn", + "bbox": [ + 104, + 237, + 883, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While fine-tuning diffusion models with reinforcement learning (RL) has demonstrated effectiveness in directly optimizing downstream objectives, existing RL frameworks are prone to overfitting the rewards, leading to outputs that deviate from the true data distribution and exhibit reduced diversity. To address this issue, we introduce entropy as a quantitative measure to enhance the exploratory capacity of diffusion models' denoising policies. We propose an adaptive mechanism that dynamically adjusts the application and magnitude of entropy and regularization, guided by real-time quality estimation of intermediate noised states. Theoretically, we prove the convergence of our entropy-enhanced policy optimization and establish two critical properties: 1) global entropy increases through training, ensuring robust exploration capabilities, and 2) entropy systematically decreases during the denoising process, enabling a phase transition from early-stage diversity promotion to late-stage distributional fidelity. Building on this foundation, we propose a plug-and-play RL module that adaptively controls entropy and optimizes denoising steps. Extensive evaluations demonstrate theoretical soundness and empirical robustness of our method, achieving state-of-the-art quality-diversity trade-offs across benchmarks. Notably, our framework significantly improves the rewards and reduces denoising steps in training by up to $40\\%$ .", + "bbox": [ + 88, + 320, + 485, + 698 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 729, + 220, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, diffusion models [12, 23, 29-31, 36] have demonstrated powerful capabilities in text-to-image generation. However, these models are designed to match the distribution of training data, lacking the ability to directly optimize downstream objectives, such as aesthetics and preference scores. To this end, reinforcement learning (RL) has been explored and shown promising potential in fine", + "bbox": [ + 89, + 755, + 483, + 863 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tuning diffusion models with various reward functions or human feedback (i.e., RLAIF and RLHF) [1, 5, 49]. In these methods, policy optimization is guided by a global reward computed on the final generated image, which is either differentiable or non-differentiable, making RL a flexible and promising solution to overcoming the limitations of supervised fine-tuning for diffusion models.", + "bbox": [ + 511, + 287, + 905, + 393 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, current RL methods rely solely on rewards from the final generated image to update policies, overlooking the importance of exploration in promoting diversity during denoising. This makes fine-tuned models susceptible to reward hacking [14], where generated images achieve high scores in reward models but significantly deviates from the true distribution, exhibiting deteriorated quality and low diversity. This phenomenon has been partially attributed to sparse and delayed rewards and addressed with early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51]. However, the problem remains to be effectively solved.", + "bbox": [ + 511, + 396, + 905, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper proposes a novel RL method with adaptive entropy regulation to address the reward hacking issue in fine-tuning diffusion models. We formally introduce entropy as a quantitative indicator to assess the exploratory capability of denoising strategies in diffusion models, integrating it into the RL optimization objectives. Theoretical analysis establishes the convergence guarantees of the proposed method for newly derived policies. Furthermore, to overcome the limitations of fixed parameterization in conventional KL regulation, we derive an adaptive mechanism that dynamically balances exploration diversity and convergence stability throughout the denoising process. Such adaptive regulation enables progressive entropy adjustment, facilitating a smooth transition from initial exploration diversity to stable convergence toward the target distribution in later denoising stages.", + "bbox": [ + 511, + 580, + 905, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building upon this theoretical framework, we formalize AdaEnt as a plug-and-play module compatible with existing RL fine-tuning architectures for diffusion models. AdaEnt addresses two critical challenges in current RL frameworks: 1) sparse reward signals caused by insuffi", + "bbox": [ + 511, + 824, + 905, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal Contribution", + "bbox": [ + 107, + 875, + 220, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding Author", + "bbox": [ + 109, + 887, + 236, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1924", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cient optimization guidance during intermediate denoising steps and 2) diversity loss resulting from KL constraints. By incorporating entropy exploration terms at each denoising step and leveraging the $\\eta$ parameter flexibility in DDIM, our method regulates entropy variations for principled exploration of denoising polices. The adaptive mechanism achieves dynamic balance between exploration and convergence through real-time quality assessment of intermediate outputs. Specifically, we introduce an image perception model trained alongside with the policy model to predict step-wise weights between diversity exploration and convergence stability in the optimization objective. Notably, the adaptive factors also serve as indicators of generation quality, which can be naturally used to reduce the total number of denoising steps, enabling efficient computation and lightweight storage.", + "bbox": [ + 89, + 90, + 480, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper highlights three main contributions:", + "bbox": [ + 109, + 333, + 419, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We theoretically derive an entropy-based exploration strategy and present an effective implementation for diffusion models.", + "- We propose an adaptive mechanism to balance the entropy-guided exploration and the regularization of real data distribution, which captures different characteristics in the early and late denoising process with a theoretical guarantee.", + "- We propose a dynamic step alignment technique based on the entropy-guided exploration to automatically determine when to stop denoising, making our method computationally efficient for both training and inference." + ], + "bbox": [ + 91, + 352, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Experiments show that AdaEnt effectively mitigates reward hacking and achieves the highest rewards while preserving the diversity of images. The results generalize to various pre-trained diffusion models, reward functions, and RL frameworks. Furthermore, through visualization analysis, we confirm the useful design of entropy and self-adaptive mechanism, which seeks diversity at the early stage and gradually switches to maintain faithfulness in the late stage. AdaEnt works as a plug-and-play module that is broadly applicable in practice.", + "bbox": [ + 89, + 536, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 702, + 232, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. RL for Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 727, + 313, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conventional training of diffusion models relies on likelihood-based objectives [12], yet these often misalign with perceptual quality metrics, resulting in degraded visual outputs [1, 16, 24]. Reinforcement learning (RL) [34] addresses this limitation by reframing the diffusion denoising process as a Markov Decision Process (MDP), enabling direct optimization of task-specific rewards such as aesthetic quality, compression efficiency, or human preference alignment. Building on pre-trained text-to-image models, RL demonstrates an effective and flexible framework to fur", + "bbox": [ + 89, + 750, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ther improve the diffusion model, where recent successful techniques for large language models, such as PPO [34] and DPO [28], can be applied to improve training stability and efficiency. Notable contributions include reward model-driven approaches like DDPO [1], SLCD [25] and DPOK [5], as well as human preference-aligned methods such as D3PO [49], Diffusion-DPO [42], DEPO [39], and HPS [45]. This paper follows this line of work, aiming to improve reward-driven performance while systematically mitigating the inherited drawbacks of over-optimization brought by RL.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Reward Hacking in Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 267, + 834, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reward hacking describes the phenomenon that the model maximizes the reward score but produces results that lack diversity or authenticity [8, 21, 35, 51]. For example, when the model overly prioritizes specific reward signals, such as image aesthetics or text alignment, it can lead to structural distortions, loss of diversity, or artifacts in generated images, ultimately compromising generation quality. This phenomenon has been partially attributed to sparse and delayed rewards, which affects reward optimization in both RL and direct back-propagation methods [4, 26, 41]. Methods that rely on direct access to the gradient of differentiable reward model pose higher risks of reward hacking [8]. While techniques like early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51] have been applied in RL methods to address this issue, the problem remains to be effectively solved [6, 7]. In contrast, this paper focuses on mitigating reward hacking in RL by adapting the policy optimization process and removing unnecessary steps to reduce computation costs.", + "bbox": [ + 511, + 290, + 906, + 578 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Entropy-based RL", + "text_level": 1, + "bbox": [ + 511, + 588, + 692, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Entropy-based reinforcement learning (RL) integrates information-theoretic concepts into policy optimization to encourage exploration. By leveraging entropy regularization [9, 22, 40, 48], these methods encourage stochasticity in action selection and enhance learning stability and adaptability. However, aimlessly increase entropy for exploration will ultimately hurt the policy [11, 47]. For a better trade-off between exploration and exploitation, SAC [10] automates entropy coefficient tuning by jointly maximizing expected returns and entropy. To our knowledge, this paper is the first to leverage entropy as a theoretically guaranteed exploration metric for optimizing diffusion models and adaptively balances the effect of entropy and distribution regularization.", + "bbox": [ + 511, + 609, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.4. Dynamic-Step Adaptation for Diffusion Models", + "text_level": 1, + "bbox": [ + 511, + 816, + 903, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing methodologies for acceleration can be divided into reducing the number of steps [15] or computation cost per step [18, 19, 43]. Still, they often use a uniform approach without accounting for each generated image's quality at", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1925", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "each denoising step. Complementary to the above methods, recent work [50] applies different denoising steps for each prompt by training a step prediction policy network. This paper proposes to dynamically early stop the denoising process to mitigate reward hacking, with an additional benefit of accelerating the RL training of diffusion models, which need not a dedicated model for step number determination.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 210, + 181, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This paper introduces the adaptive exploration of diffusion policy optimization through the lens of entropy. In this section, we first prove the convergence of policy optimization with our new entropy-enhanced training objective in diffusion policy optimization. Coping with the trade-off between entropy-based exploration of RL and the inherent distribution modeling of the diffusion model, we then propose an adaptive mechanism to automatically balance the trade-off between exploration and the original objective, which is theoretically guaranteed to adjust entropy to allow a smooth transition from early-stage diversity exploration to late-stage data distribution maintenance.", + "bbox": [ + 89, + 234, + 483, + 417 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Diffusion as Markov Decision Processes", + "text_level": 1, + "bbox": [ + 89, + 425, + 431, + 440 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In diffusion models [12, 36], data distribution $p(\\mathbf{x}_0)$ is modeled through a Markov forward process $q(\\mathbf{x}_t|\\mathbf{x}_{t-1})$ that gradually adds noise to the data and a reverse process $p_\\theta(\\mathbf{x}_{t-1}|\\mathbf{x}_t)$ for generation, defined as:", + "bbox": [ + 89, + 446, + 483, + 508 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\sqrt {1 - \\alpha_ {t}} \\mathbf {I}), \\\\ p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 520, + 483, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\sqrt{1 - \\alpha_t}$ denotes the noise scale at timestep $t$ , $\\mu_{\\theta}(\\mathbf{x}_t,t)$ is the predicted mean by the model, and $\\sigma_t^2$ is a time-dependent variance.", + "bbox": [ + 89, + 568, + 483, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current RL fine-tuning methods treat the generation process of diffusion models as a Markov Decision Process (MDP), represented by the five-tuple $\\langle S, \\mathcal{A}, P, \\rho_0, R \\rangle$ :", + "bbox": [ + 89, + 614, + 483, + 659 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\mathbf {z}, t, \\mathbf {x} _ {T - t}) \\in S, \\\\ \\pi_ {\\theta} \\left(\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}\\right) \\triangleq p _ {\\theta} \\left(\\mathbf {x} _ {T - t - 1} \\mid \\mathbf {x} _ {T - t}, \\mathbf {z}\\right), \\\\ P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\mathbf {z}}, \\delta_ {T - t - 1}, \\delta_ {\\mathbf {x} _ {T - t - 1}}\\right), \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - t - 1}, \\tag {2} \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\mathbf {z}), \\delta_ {T}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ R \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left\\{ \\begin{array}{l l} r \\left(\\mathbf {x} _ {0}, \\mathbf {z}\\right) & \\text {i f} t = T, \\\\ 0 & \\text {o t h e r w i s e}. \\end{array} \\right. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 671, + 480, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathcal{S}$ denotes the state space, containing all possible image states $s_t$ in the diffusion model's generation process; $\\mathcal{A}$ represents the action space, corresponding to the denoising operations $a_t$ ; $P(s_{t + 1}|s_t,a_t)$ is the state transition probability, defining the probability of transitioning from state $s_t$ to", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$s_{t + 1}$ after taking action $a_t$ ; $R(s_t, a_t)$ represents the reward function, which typically evaluates the final image quality upon completion (i.e., at $t = T$ ); $\\rho_0$ is the distribution of initial states.", + "bbox": [ + 511, + 90, + 906, + 150 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An important feature of diffusion models is that the sampling procedure has a wide design space after the score-matching training process, allowing a variance schedule ranging from larger to smaller. A popular uniform formulation of this sampling process is given in [37], where the generation process is defined as follows:", + "bbox": [ + 511, + 152, + 906, + 244 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mu_ {\\theta} (\\mathbf {x} _ {t - 1}) = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t})}{\\sqrt {\\alpha_ {t}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\eta^ {2} \\frac {(1 - \\alpha_ {t - 1})}{(1 - \\alpha_ {t})} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t}) \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 258, + 888, + 342 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {t - 1} = \\eta \\sqrt {\\frac {\\left(1 - \\alpha_ {t - 1}\\right)}{\\left(1 - \\alpha_ {t}\\right)} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 345, + 903, + 398 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta}(\\mathbf{x}_{t - 1})$ and $\\sigma_{t - 1}$ represent the mean and variance of the current denoising step, $\\eta$ determines the variance of each generation step. The $\\epsilon_{\\theta}^{(t)}(\\mathbf{x}_t)$ is the neural network for predicting noise. Note that most algorithms only model the generation process for the case where $\\eta = 1$ , which restricts the policy search to a narrower space.", + "bbox": [ + 511, + 398, + 906, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Provable Exploration Entropy for Diffusion", + "text_level": 1, + "bbox": [ + 511, + 506, + 883, + 522 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In reinforcement learning, the entropy of policy $H(\\pi)$ represents the diversity of action choice on the current state. To increase the diversity with a more extensive exploration region, an entropy-related term is added to the original loss of RL fine-tuning. A common goal in fine-tuning the diffusion models is to maximize the expected reward of the generated images given the prompt distribution. Furthermore, the equation below holds true.", + "bbox": [ + 511, + 530, + 905, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ] \\\\ = \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0: T} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 667, + 821, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "From [5], we know the following expression holds:", + "bbox": [ + 511, + 727, + 852, + 742 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) | | p _ {p r e} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 537, + 758, + 882, + 816 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, we conclude that", + "bbox": [ + 513, + 832, + 697, + 847 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right)\\right) = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} \\left(x _ {t} \\mid z\\right)} \\left[ H \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right)\\right) \\right],\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 862, + 880, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1926", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7339c0a9e3530cce54424893818b20253255f20627615c26720d313e79c50e39.jpg", + "image_caption": [ + "Figure 1. AdaEnt's framework. The entropy-guided exploration expands policy's search area at each denoising step, regularized by the KL divergence of the current policy and reference policy. The trade-off between the two effects is coordinated by a perception model that evaluates the real-time noised images at each step. The prediction of the adaptive module also indicates unnecessary denoising steps and dynamically stops the denoising process to reduce the computation cost in training." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 893, + 282 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and", + "bbox": [ + 89, + 380, + 120, + 393 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) \\mid \\mid p _ {\\theta_ {k - 1}} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {\\theta_ {k - 1}} (x _ {t - 1}, z)) ]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 402, + 464, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, using the KL divergence and entropy, we propose the following objective for regularized training at the $t$ -th step:", + "bbox": [ + 89, + 467, + 482, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (z)} \\mathbb {E} _ {p _ {\\theta} (x _ {0: T} | z)} [ r (x _ {0}, z) + \\sum_ {t = 1} ^ {T} \\alpha_ {1} H (p _ {\\theta} (x _ {t - 1} | x _ {t}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {2} K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {3} K L \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right) \\mid p _ {\\theta_ {k - 1}} \\left(x _ {t - 1}, z\\right)\\right), \\tag {4} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 518, + 480, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha_{1} \\geq 0$ , $\\alpha_{2}$ and $\\alpha_{3} \\geq 0$ . A significant concern about including entropy in the loss function of policy optimization is that it may lead to unstable policies. To address this issue, we theoretically analyze the new loss function and prove that policy optimization is convergent.", + "bbox": [ + 89, + 660, + 483, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.1. Given the objective of Eq. (4), the optimal policy $p_{\\theta}^{*}(x_{t - 1}|x_{t},z)$ has the following expression:", + "bbox": [ + 89, + 741, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} ^ {*} (x _ {t - 1} | x _ {t}, z) \\propto \\exp \\{\\frac {Q ^ {*} (x _ {t - 1} , z) + \\alpha_ {2} \\log p _ {p r e} (x _ {t - 1} | x _ {t} , z)}{\\alpha_ {1} + \\alpha_ {2}} \\},\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 777, + 468, + 811 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Q^{*}(x_{t - 1},z) = r(s_{0},z)I_{\\{t - 1 = 0\\}} + (\\alpha_{1} + \\alpha_{2})$ .", + "bbox": [ + 89, + 815, + 480, + 830 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$\\log \\sum_{x_{t - 2}}\\exp \\{\\frac{Q^{*}(x_{t - 2},z) + \\alpha_{2}\\log p_{pre}(x_{t - 2}|x_{t - 1},z)}{\\alpha_{1} + \\alpha_{2}}\\} .$", + "bbox": [ + 91, + 832, + 429, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 3.1 is detailed in Appendix A.1. This theorem guarantees that with entropy added to training objective, the policy will ultimately converge.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.2. When $\\hat{\\alpha}_1 \\leq \\alpha_1$ , we have $H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\hat{\\alpha}_1,\\alpha_2) \\leq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2)$ . Specially, compared to not including an entropy term, that is $\\alpha_1 = 0$ , in the objective function of Eq. (4), adding an entropy term results in a higher entropy after the algorithm converges, that is $H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2) \\geq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);0,\\alpha_2)$ .", + "bbox": [ + 511, + 378, + 906, + 470 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 3.2 is detailed in Appendix A.2. This theorem demonstrates that adding an entropy term to the loss function increases the policy entropy, thereby enhancing the algorithm's exploratory behavior. Through greater exploration capability, the agent can discover better actions and achieve higher returns.", + "bbox": [ + 511, + 474, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Theorem 3.3. If $C = \\min_{x_{t-1} \\neq y_{t-1}} [\\log p_{pre}(y_{t-1}|x_t, z) - \\log p_{pre}(x_{t-1}|x_t, z)] / [Q^*(x_{t-1}, z) - Q^*(y_{t-1}, z)] > 0$ , the objective function of Eq. (4) contains only a KL term and no entropy term, that is $\\alpha_1 = 0$ , if $1/C \\leq \\hat{\\alpha}_2 \\leq \\alpha_2$ , we have $H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\hat{\\alpha}_2) \\geq H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\alpha_2)$ .", + "bbox": [ + 511, + 571, + 905, + 648 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proof of Theorem 3.3 is detailed in Appendix A.3. This theorem indicates that adding a KL term to the loss function may reduce the exploration capability of the policy if the parameters are not properly tuned, thereby affecting the algorithm's performance.", + "bbox": [ + 511, + 652, + 905, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. AdaEnt: A Practical Instantiation", + "text_level": 1, + "bbox": [ + 511, + 739, + 823, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on the above theoretical foundation, we propose AdaEnt, a plug-and-play module that adaptatively controls entropy and a number of steps to combat reward hacking in diffusion policy optimization. The architecture of AdaEnt is illustrated in Fig. 1.", + "bbox": [ + 511, + 765, + 905, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Entropy-Driven Exploration", + "text_level": 1, + "bbox": [ + 511, + 848, + 769, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To cope with the diversity degradation brought by reward hacking issue, we introduce the entropy of denoising policy", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1927", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$p_{\\theta}(x_{t - 1}|x_t,z)$ to enhance exploration. In a normal distribution, the entropy is related to its variance. From Eq. (3), we define a new variance:", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sigma_ {T - t} = \\eta_ {\\theta , \\phi} (\\mathbf {x} _ {T - t}, T - t) \\sqrt {\\frac {1 - \\alpha_ {T - t - 1}}{(1 - \\alpha_ {T - t})} \\cdot \\left(1 - \\frac {\\alpha_ {T - t}}{\\alpha_ {T - t - 1}}\\right)},\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 143, + 480, + 184 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\eta_{\\theta ,\\phi}(\\mathbf{x}_{T - t},T - t)$ is a dynamic, lightweight head with parameters $\\phi$ , built upon the neural network $\\mu_{\\theta}$ ,designed to provide a latent and step-aware dynamic variance.", + "bbox": [ + 89, + 191, + 482, + 238 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Trade-Off between Exploration and Regularization", + "text_level": 1, + "bbox": [ + 89, + 246, + 482, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In diffusion models, different denoising steps play various roles in image generation. Previous research has shown that earlier steps significantly impact the semantic content and global layout of the image, while later steps influence the texture and details [3]. Motivated by this phenomenon, we propose an adaptive regularization scheme.", + "bbox": [ + 89, + 282, + 483, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We introduce an adaptive balance factor $\\gamma_{t}$ to dynamically adjust the balance between exploration diversity and denoising convergence across the denoising process. For the sake of simplicity, we introduce the PPO algorithm [34]. Therefore, the objective to optimize (Eq. (4)) is updated:", + "bbox": [ + 89, + 375, + 483, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {J} = \\max _ {\\theta} \\mathbb {E} [ \\min (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)} A, c l i p (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)}, 1 - \\epsilon , 1 + \\epsilon) A) \\\\ + \\sum_ {t = 1} ^ {T} \\left(1 - \\gamma_ {t}\\right) H \\left(p _ {\\theta} (\\cdot)\\right) - \\sum_ {t = 1} ^ {T} \\gamma_ {t} K L \\left(p _ {\\theta} (\\cdot)\\right) \\left\\| p _ {p r e} (\\cdot)\\right), \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 94, + 458, + 486, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $A$ is an advantage function. According to Eq. (5), as the generation progresses, the regularization term will dominate the optimization and seek directions toward the distribution of real images within the high-reward region.", + "bbox": [ + 89, + 547, + 482, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A prediction model $f(\\mathbf{x}_t)$ is trained on each intermediate denoised (latent) image $\\mathbf{x}_t$ , with intermediate denoised images close to the final generated image $\\mathbf{x}_0$ labeled as 1, and those close to initial noise $\\mathbf{x}_T$ labeled as 0. The binary classification loss function is defined as follows:", + "bbox": [ + 89, + 607, + 483, + 681 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {p r e d} = - \\sum_ {i} \\left(y _ {i} \\log \\left(f \\left(x _ {i}\\right)\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - f \\left(x _ {i}\\right)\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 91, + 689, + 482, + 733 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $y_{i} \\in \\{0,1\\}$ is the label indicating whether the image is noiseless, and $f(x_{i})$ is the model's predicted probability that image $x_{i}$ is noiseless. After training, this network can evaluate the denoising level of any intermediate image during the denoising process, and we let $\\gamma_{t} = f(\\mathbf{x}_{t})$ as a measure of the denoising process indicating how well an intermediate image aligns with real image distribution.", + "bbox": [ + 89, + 734, + 482, + 839 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Dynamic Step Decision", + "text_level": 1, + "bbox": [ + 89, + 848, + 303, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "One limitation of diffusion models is the reliance on many iterative denoising steps. The drawback will be exacerbated", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "when performing RL fine-tuning, leading to a high computation demand and exacerbation of reward hacking. As later denoising steps contribute less to the overall appearance of the generated image, applying an early stop on the denoising process will make training and inference more efficient and mediately alleviate reward hacking. The criterion of early stop is dynamically determined by how well the current image aligns with true image distribution. According to Sec. 4.2, $\\gamma_{t} = 1$ means the current state of image satisfies the quality. Hence we apply early stop in this scenario, which is completely dynamic according to the complexity of generation and current image quality, without any predefined hyperparameters.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Note that in existing diffusion step adaptation methods, the total number of denoising steps have to be determined before hand, which is impossible for the adaptive module to make a decision based on sufficient information about the image quality. In contrast, our adaptation mechanism monitors the generated image at each step and determines the total number of steps based on real-time quality of the current state. Recent study on denoising diffusion models reveals the temporal information may be not necessary [38], which supports our design of dynamically cutting off the iterative denoising process. The algorithmic procedure of AdaEnt can be found in Appendix B.", + "bbox": [ + 511, + 289, + 906, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 488, + 645, + 505 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 513, + 702, + 531 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To compare with previous work, we use the same prompt set consisting of 45 different animal classes for training (Simple-animals), which has been widely adopted for reward fine-tuning [1, 49, 51]. For testing, we generate 8 different noises for each animal class, totaling 360 test samples. For generalization evaluation, larger-scale prompt sets are further considered in testing as done in [1], including 398 animal classes from ImageNet (ImageNet-animals) and the above 45 animals each concatenated with an activity (Activity-animals).", + "bbox": [ + 511, + 537, + 905, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rewards and Metrics. In this paper, we first deploy three metrics as the reward functions, that is, Aesthetic Score [33], Compressibility, and Incompressibility. To quantitatively evaluate performance in face of reward hacking, we then introduce cross-reward generalization and well-acknowledged metrics as suggested in [51]: ImageReward [46], Inception Score (IS) [32], PickScore [17], and ClipScore [27]. Details of each metric can be found in Appendix C.", + "bbox": [ + 511, + 689, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. Following [1], we take Stable Diffusion v1.4 (SDv14) as our main test bed for reward finetuning. We also consider Stable Diffusion v1.5 (SDv15), v2.1-turbo (SDv21), and XL1.0 (XL) to verify the robustness of our method with different backbones. All models", + "bbox": [ + 511, + 825, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1928", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7e2716bbf81b9c217f24a7fcebaec1166ef2d6ff4bf556ca9c34c34be0d5161c.jpg", + "image_caption": [ + "Figure 2. Reward curves of RL fine-tuning SDv14 with three different reward functions. Our proposed method consistently outperforms baselines in all three experimental scenarios, in terms of both peak achieved rewards and the sample efficiency." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 356, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/51d2b5b7fd95b548b53b906bb30906d7d916fff85b2c7171978946bab1220fb1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 88, + 635, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b225500f3bae1a5e4e23b453d44263071fcab49de0b4345acf9ce77a9ac5f5d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 645, + 88, + 906, + 251 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are fine-tuned with low-rank adaptation (LoRA) [13] on 8 NVIDIA Tesla A100 GPUs. More training details can be found in Appendix C.", + "bbox": [ + 89, + 315, + 483, + 361 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 375, + 230, + 390 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We benchmark our methods against state-of-the-art RL-based diffusion fine-tuning frameworks, including DDPO [1], DPOK [5], D3PO [49], and TDPO [51]. As shown in Fig. 2, our method demonstrates consistent superiority in learning all three reward functions, validating its effectiveness for optimizing downstream objectives. Crucially, our balanced state exploration and adaptive regularization mechanism achieves promote the rate of reward growth by a large margin compared to DPOK's fixed-weight regularization strategy. Unlike dense-reward methods requiring evaluations of noised states' long-term impacts, our approach implements stepwise adaptive entropy to guide denoising space exploration, enabling sample-efficient optimization. Consequently, we attain rewards slightly higher than the latest dense-reward method TDPO but using much fewer training samples.", + "bbox": [ + 88, + 398, + 482, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Moreover, to systematically investigate reward hacking mitigation, we test the models optimized for Aesthetic Score with different aesthetic scores and general metrics for image quality and text-to-image alignment. We report the min, mean, and max values of each metric across the entire training process. As shown in Tab. 1, our generated images achieves overall higher scores on the diverse evaluation perspectives, effectively circumventing reward hacking phenomena even under elevated reward conditions. The results indicate that the proposed exploratory entropy effectively contributes to improved diversity and visual fidelity of generated images, along with less overfitting to the learning reward function.", + "bbox": [ + 88, + 642, + 482, + 837 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In summary, through this balanced exploration-regularization paradigm, our method demonstrates superior capability in discovering latent high-reward policies, thereby overcoming conventional performance plateaus", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8f8012150ab19ae67b1081b44089d27fff32128eeea6c6b6b6ba787b5d8ecc38.jpg", + "image_caption": [ + "Figure 3. Distribution visualizations of our intermediate variables. (a) Distribution of predicted $\\gamma$ during training. The predictor progressively learns to effectively distinguish latent codes of different stages. (b) Distribution of denoising step number during training. Our adaptive step mechanism effectively reduces the denoising steps in training." + ], + "image_footnote": [], + "bbox": [ + 514, + 311, + 707, + 388 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6eb55d24e75359497c7695246717d0f014863d5daf831675e6155064fdd1171.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 714, + 311, + 903, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and achieving measurable enhancements in both generation quality and quantitative reward metrics.", + "bbox": [ + 511, + 511, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bda6cce7f839434189d3c9a6863c2c1372c14ae19b52441cb13816aa6fcca061.jpg", + "image_caption": [ + "Figure 4. Left: Entropy can be optimized in AdaEnt under different scaling factors. Right: Excessive application of either entropy or KL regularization impedes reward learning and may ultimately induce catastrophic collapse." + ], + "image_footnote": [], + "bbox": [ + 517, + 556, + 694, + 660 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/de7bc767c5bc6589fb6cc7fe488d553f65ae6c0181d984855e5f61cc458f6322.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 556, + 905, + 660 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 513, + 756, + 668, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effect of $\\gamma$ Predictor. The coefficient $\\gamma$ is introduced to adaptively balance the exploratory entropy and KL regularization based on the current denoising situation. In Fig. 3 (a), we show the distributions of $\\gamma$ with 50-step denoising during training. It can be observed that the predictor progressively learns to effectively differentiate latent codes that are more distinctive (i.e., those at the beginning and end of denoising), whereas the results in the intermediate steps are", + "bbox": [ + 511, + 779, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1929", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/55553df9e29d3772418448c72c5fce8fc6db81c0931c37efe8c2f8f4348aa0ec.jpg", + "table_caption": [ + "Table 1. Quality assessment from multiple unseen metric perspectives. All results are evaluated across training on the aesthetic score and sampled with the same training iteration. * refers to the Inception Score being sampled with the same aesthetic score of different methods. Since D3PO performs limited in learning rewards (see Fig. 2), the range of aesthetic score data is not enough for inception score collection. #Top1 indicates the instances when each method achieves the best performance across all metrics." + ], + "table_footnote": [], + "table_body": "
MethodPickScoreInceptionScore*ImageRewardClipScore#Top1
MinMeanMaxMinMeanMaxMinMeanMaxMinMeanMax
DDPO [1]21.2821.3721.5321.1122.9523.670.89491.07071.24930.29660.30190.30831
DPOK [5]21.2721.4021.5320.4822.2923.460.90601.04961.13400.30080.30460.30831
D3PO [49]21.2821.2921.31---0.86230.89620.93170.30000.30120.30221
TDPO [51]21.0621.3821.7119.9421.8823.950.00480.84531.13540.27540.30100.31132
AdaEnt21.2421.4121.6422.1124.3426.190.90781.12121.31230.30000.30500.30988
", + "bbox": [ + 91, + 155, + 913, + 282 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/60f9876634bb3d6e177ec10f86f563bfebf7294cea22e42ae7146848da21fc06.jpg", + "image_caption": [ + "Figure 5. Comparison of aesthetic reward curves and visual quality on SDv15 and SDv21. Our method consistently achieves higher rewards without over-optimizing across different backbones, while baseline methods compromise either visual quality or text-image alignment to attain aesthetic rewards." + ], + "image_footnote": [], + "bbox": [ + 94, + 303, + 480, + 486 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "predicted with non-binary confidence. This adaptive behavior dynamically balances entropy and the KL divergence as the denoising process progresses, facilitating a shift from exploration to regularization during each denoising process. Balance between Entropy and KL Divergence. In Fig. 4, we conduct an in-depth investigation into the effects of introducing entropy in reward learning. Specifically, with a fixed KL, we design ablation variants with entropy scaled by different factors (i.e., $\\mu_{e}$ ). The baselines, DDPO and DPOK, have a fixed schedule of entropy (according to Eq. (1)). In our proposed framework, it is observed that entropy can be optimized to increase under different scaling factors, and corresponding with a better learning of reward. However, when entropy's order of magnitude overwhelms KL divergence (i.e., $\\mu_{e} = 1.0$ ), over-exploration may lead to the collapse of training. Therefore, the result demonstrates the necessity of introducing the trade-off between entropy and KL, which ensures a well-balanced training process for improved generating performance.", + "bbox": [ + 88, + 613, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ebf51ade7304065f957aa981e79c5d6bd4443cb6b88efc62bf91fc64526adab3.jpg", + "table_caption": [ + "Table 2. Generalization results of SDv15 and SDv21 on unseen metrics. The metrics are abbreviated and the same in Tab. 1." + ], + "table_footnote": [], + "table_body": "
MethodSDv15SDv21#Top1
PSISIRCSPSISIRCS
DDPO21.1716.900.820.3020.347.89-0.210.260
DPOK21.1818.680.840.3120.5310.280.040.272
D3PO21.0122.060.800.3318.843.14-1.580.252
TDPO19.682.40-1.570.2418.521.56-2.090.240
AdaEnt21.2118.990.850.3020.598.140.100.275
", + "bbox": [ + 516, + 344, + 903, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d80e1b8ff98d2572201721ee303d4809174f4fafaad5ba9d4be4c76f190a9a2a.jpg", + "image_caption": [ + "Figure 6. Comparison of aesthetic reward curves and visual quality on SD-XL. Notably, we observe that our method is the only one that successfully works on SD-XL." + ], + "image_footnote": [], + "bbox": [ + 516, + 476, + 697, + 580 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f21f33e9edb90bf43b74dc454e1bb849107943946d6c077059b20f42550e77f4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 477, + 903, + 566 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Results on Different Backbones.", + "text_level": 1, + "bbox": [ + 511, + 662, + 794, + 678 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to verify the generalizability of our method across different models for practical applications, we hereby extend our experiment to SDv15 and SDv21. In Fig. 5, our method also achieves superior effectiveness in reward learning. For generalization evaluation, we also evaluate the generated image quality with more metric perspectives for SDv15 and SDv21. As shown in Tab. 2, AdaEnt achieves superior overall quality, and despite TDPO shows comparable aesthetic score with our method, it is demonstrated to be failed in prompt-image alignments.", + "bbox": [ + 509, + 686, + 906, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Moreover, considering that SD-XL is one of the most advanced and extensive pre-trained model, we provide a concise reward-learning experiment on SD-XL in Fig. 6. The results demonstrate that AdaEnt can still bring reward gains", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1930", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a76a196830567cde79441aba76233adfd078f17a133e4f5465dde1d1d86fb51.jpg", + "image_caption": [ + "Figure 7. Comparison of generalization ability to larger scale prompt sets with unseen prompts. The experiments are conducted on SDv14 and trained with the small Simple-animals set." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 210 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9a31224dd9cb1057704a0cbf0235c6f8cd05aeb9bb6f4647ef0a9368e1645e1e.jpg", + "image_caption": [ + "Figure 8. Comparison of generalization ability to more complex prompts with multiple categories. All methods are implemented in SD-XL. The performance of our method is highlighted in darker colors, demonstrating AdaEnt's superior performance in both image quality and prompt-image alignment." + ], + "image_footnote": [], + "bbox": [ + 91, + 279, + 480, + 421 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "for larger and more powerful models.", + "bbox": [ + 89, + 532, + 339, + 547 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Generalization to Unseen Prompts.", + "text_level": 1, + "bbox": [ + 89, + 560, + 393, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To investigate whether the performance of RL fine-tuned models is limited to the prompts seen in training, which is an ordinary case of reward hacking, we hereby conduct prompts generalization experiments from two perspectives.", + "bbox": [ + 89, + 583, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to larger-scale prompt sets. To evaluate the generalization ability to unseen prompts, we evaluate the aesthetic score on ImageNet-animals and Activity-animals. The experiments are conducted on SDv14, using prompts from Simple-animals as the training set. As presented in Fig. 7, our method achieves promising aesthetic scores on the unseen datasets, which demonstrates the effectiveness of our method.", + "bbox": [ + 89, + 643, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization to more complex prompts. Additionally, we assess the performance on more complex prompts. To be specific, we further conduct experiments on SD-XL with a subset of the HPSv2 dataset [44], which contains complex prompts that are not limited to animals. We identify several categories in the prompt set and measure the prompt-image consistency using clip-score. We also provide the aesthetic score to assess the quality of generated images. As presented in Fig. 8, our method exhibits the best perfor", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "mance in the aesthetic score and the maintenance of image-prompts alignment, which suggests our superiority.", + "bbox": [ + 511, + 90, + 903, + 122 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.6. Computation Efficiency of AdaEnt", + "text_level": 1, + "bbox": [ + 511, + 131, + 815, + 147 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Since the adaptive mechanism introduced in our work allows for the early termination of the denoising process once denoising is essentially complete, it is expected to improve the training efficiency of diffusion models. Firstly, we provide the visualized distributions of denoising step number during training in Fig. 3 (b), which verifies the capability of adaptively reducing the step number. Quantitatively, we take the aesthetic score $= 7$ as the training terminals of each method for evaluation. In Tab. 3, we provide GFLOPs and time for single batch (size=12) sampling, along with the total sampling step and time for the overall training process. The results show that the proposed AdaEnt can significantly accelerate the training process with almost no increase in computational cost (i.e., GFLOPS).", + "bbox": [ + 511, + 152, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/a610126ae975f8c24cac1332b69cd506a893a2a30ea9499ddf5eeda5bbee89cb.jpg", + "table_caption": [ + "Table 3. Computation Efficiency of integrating our proposed dynamic early stop mechanism with existing RL frameworks. TDS refers to total denoising steps. Results of both overall training and single-batch inference are reported." + ], + "table_footnote": [], + "table_body": "
MethodOverallSingle batch
TDSTimeGFLOPSTime
DDPO1.843+e611.98h1.034e+069.56s
DPOK2.015+e614.76h9.69s
DDPO+Ours0.242+e65.52h1.044e+066.01s
DPOK+Ours0.294+e67.16h7.12s
", + "bbox": [ + 514, + 445, + 888, + 560 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 588, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work addresses the critical challenge of reward hacking in diffusion model optimization through reinforcement learning, where conventional RL frameworks tend to produce overfitted outputs that compromise distributional fidelity and generation diversity. Our theoretical analyses establish entropy as a proxy to enhance the exploratory capacity of denoising policies. The theoretical basis is instantiated as AdaEnt, a framework that achieves adaptive entropy control and step-wise optimization through three interconnected mechanisms: entropy-aware policy updates, distributional consistency regularization, and noise-level dependent exploration scheduling. Extensive experiments demonstrate that our method effectively mitigates reward hacking while preserving the diffusion model's inherent generation capabilities, establishing new state-of-the-art performance in quality-diversity trade-off across multiple diffusion-based RL benchmarks. The adaptive entropy proposed in this paper offers new perspectives for future RL fine-tuning of diffusion models.", + "bbox": [ + 511, + 612, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1931", + "bbox": [ + 483, + 945, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Acknowledgment", + "text_level": 1, + "bbox": [ + 91, + 90, + 263, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 250, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 2, 5, 6, 7", + "[2] Bo Dai, Albert E. Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, and Le Song. SBEED: convergent reinforcement learning with nonlinear function approximation. In Proceedings of International Conference on Machine Learning., 2018. 12", + "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5", + "[4] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2024. 2", + "[5] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 6, 7", + "[6] Yaozhong Gan, Renye Yan, Xiaoyang Tan, Zhe Wu, and Jun-liang Xing. Transductive off-policy proximal policy optimization. arXiv preprint arXiv:2406.03894, 2024. 2", + "[7] Yaozhong Gan, Renye Yan, Zhe Wu, and Junliang Xing. Reflective policy optimization. arXiv preprint arXiv:2406.03678, 2024. 2", + "[8] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835-10866. PMLR, 2023. 2", + "[9] Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International conference on machine learning, pages 1352-1361. PMLR, 2017. 2", + "[10] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, pages 1861-1870. Pmlr, 2018. 2", + "[11] Seungyul Han and Youngchul Sung. A max-min entropy framework for reinforcement learning. Advances in Neural Information Processing Systems, 34:25732-25745, 2021. 2" + ], + "bbox": [ + 93, + 275, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This paper is supported by the National Science Foundation of China under Grant No.62025401, 62341407, U24B200170, 62406008; Beijing Municipal Science and Technology Program under Grant Z241100004224015; Beijing Natural Science Foundation under Grant No.L223004; '111' Project under Grant No.B18001. This work is also supported in part by the Natural Science Foundation of China under Grant No. 62222606.", + "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3", + "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6", + "[14] Rohit Jena, Ali Taghibakhshi, Sahil Jain, Gerald Shen, Nima Tajbakhsh, and Arash Vahdat. Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models. arXiv preprint arXiv:2409.06493, 2024. 1", + "[15] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2", + "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 2", + "[17] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15", + "[18] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. CoRR, 2023. 2", + "[19] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 2", + "[20] Jincheng Mei, Chenjun Xiao, Ruitong Huang, Dale Schuurmans, and Martin Müller. On principled entropy exploration in policy optimization. In Proceedings of International Joint Conference on Artificial Intelligence, 2019. 12", + "[21] Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, and Dacheng Tao. Inform: Mitigating reward hacking in rlhf via information-theoretic reward modeling. Advances in Neural Information Processing Systems, 37: 134387-134429, 2025. 2", + "[22] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016. 2", + "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1", + "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + ], + "bbox": [ + 93, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1932", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Owen Oertell, Shikun Sun, Yiding Chen, Jin Peng Zhou, Zhiyong Wang, and Wen Sun. Efficient controllable diffusion via optimal classifier guidance. arXiv preprint arXiv:2505.21666, 2025. 2", + "[26] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2", + "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR. 5, 15", + "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2", + "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1", + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022.", + "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1", + "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5, 15", + "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 2022. 5", + "[34] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5", + "[35] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming. Advances in Neural Information Processing Systems, 35:9460-9471, 2022. 2", + "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3", + "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon." + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3", + "[38] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv preprint arXiv:2502.13129, 2025. 5", + "[39] Shikun Sun, Chengrui Wang, Min Zhou, Zixuan Wang, Xiaoyu Qin, Tiezheng Ge, Bo Zheng, and Jia Jia. Depo: Enhancing e-commerce image background generation with short trajectory direct expected preference optimization. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2", + "[40] Marc Toussaint. Robot trajectory optimization using approximate inference. In Proceedings of the 26th annual international conference on machine learning, pages 1049-1056, 2009. 2", + "[41] Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Finetuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024. 2", + "[42] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2", + "[43] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 2", + "[44] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023.8", + "[45] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2", + "[46] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15", + "[47] Renye Yan, Zhe Wu, Yuan Zhan, Pin Tao, Zongwei Wang, Yimao Cai, and Junliang Xing. Mnemonic dictionary learning for intrinsic motivation in reinforcement learning. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2023. 2", + "[48] Renye Yan, Yaozhong Gan, You Wu, Ling Liang, Junliang Xing, Yimao Cai, and Ru Huang. The exploration-exploitation dilemma revisited: An entropy perspective. arXiv preprint arXiv:2408.09974, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1933", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 1, 2, 5, 6, 7", + "[50] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 3", + "[51] Ziyi Zhang, Sen Zhang, Yibing Zhan, Yong Luo, Yonggang Wen, and Dacheng Tao. Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. In International Conference on Machine Learning, pages 60396-60413. PMLR, 2024. 1, 2, 5, 6, 7" + ], + "bbox": [ + 91, + 90, + 480, + 287 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1934", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_model.json b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_model.json new file mode 100644 index 0000000000000000000000000000000000000000..698f6a7a0f695cb432273d9e68b649c71adad685 --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.101, + 0.131, + 0.899, + 0.154 + ], + "angle": 0, + "content": "Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.18, + 0.888, + 0.217 + ], + "angle": 0, + "content": "Renye Yan\\(^{1*}\\), Jikang Cheng\\(^{1*}\\), Yaozhong Gan\\(^{4}\\), Shikun Sun\\(^{3}\\), You Wu\\(^{4}\\), Yunfan Yang\\(^{1}\\), Ling Liang\\(^{1}\\), Jinlong Lin\\(^{1}\\), Yeshuang Zhu\\(^{2\\dagger}\\), Jie Zhou\\(^{2}\\), Jinchao Zhang\\(^{2}\\), Junliang Xing\\(^{3}\\), Yimao Cai\\(^{1\\dagger}\\), Ru Huang\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.268, + 0.219, + 0.729, + 0.235 + ], + "angle": 0, + "content": "Peking University\\(^{1}\\), Tencent Inc\\(^{2}\\), Tsinghua University\\(^{3}\\), Nanjing University\\(^{4}\\)." + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.238, + 0.885, + 0.253 + ], + "angle": 0, + "content": "victory@stu.pku.edu.cn, jikangcheng99@gmail.com, yshzhu@tencent.com, caiyimao@pu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.321, + 0.486, + 0.699 + ], + "angle": 0, + "content": "While fine-tuning diffusion models with reinforcement learning (RL) has demonstrated effectiveness in directly optimizing downstream objectives, existing RL frameworks are prone to overfitting the rewards, leading to outputs that deviate from the true data distribution and exhibit reduced diversity. To address this issue, we introduce entropy as a quantitative measure to enhance the exploratory capacity of diffusion models' denoising policies. We propose an adaptive mechanism that dynamically adjusts the application and magnitude of entropy and regularization, guided by real-time quality estimation of intermediate noised states. Theoretically, we prove the convergence of our entropy-enhanced policy optimization and establish two critical properties: 1) global entropy increases through training, ensuring robust exploration capabilities, and 2) entropy systematically decreases during the denoising process, enabling a phase transition from early-stage diversity promotion to late-stage distributional fidelity. Building on this foundation, we propose a plug-and-play RL module that adaptively controls entropy and optimizes denoising steps. Extensive evaluations demonstrate theoretical soundness and empirical robustness of our method, achieving state-of-the-art quality-diversity trade-offs across benchmarks. Notably, our framework significantly improves the rewards and reduces denoising steps in training by up to \\(40\\%\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.73, + 0.222, + 0.745 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.484, + 0.864 + ], + "angle": 0, + "content": "In recent years, diffusion models [12, 23, 29-31, 36] have demonstrated powerful capabilities in text-to-image generation. However, these models are designed to match the distribution of training data, lacking the ability to directly optimize downstream objectives, such as aesthetics and preference scores. To this end, reinforcement learning (RL) has been explored and shown promising potential in fine" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.906, + 0.394 + ], + "angle": 0, + "content": "tuning diffusion models with various reward functions or human feedback (i.e., RLAIF and RLHF) [1, 5, 49]. In these methods, policy optimization is guided by a global reward computed on the final generated image, which is either differentiable or non-differentiable, making RL a flexible and promising solution to overcoming the limitations of supervised fine-tuning for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.906, + 0.579 + ], + "angle": 0, + "content": "However, current RL methods rely solely on rewards from the final generated image to update policies, overlooking the importance of exploration in promoting diversity during denoising. This makes fine-tuned models susceptible to reward hacking [14], where generated images achieve high scores in reward models but significantly deviates from the true distribution, exhibiting deteriorated quality and low diversity. This phenomenon has been partially attributed to sparse and delayed rewards and addressed with early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51]. However, the problem remains to be effectively solved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.824 + ], + "angle": 0, + "content": "This paper proposes a novel RL method with adaptive entropy regulation to address the reward hacking issue in fine-tuning diffusion models. We formally introduce entropy as a quantitative indicator to assess the exploratory capability of denoising strategies in diffusion models, integrating it into the RL optimization objectives. Theoretical analysis establishes the convergence guarantees of the proposed method for newly derived policies. Furthermore, to overcome the limitations of fixed parameterization in conventional KL regulation, we derive an adaptive mechanism that dynamically balances exploration diversity and convergence stability throughout the denoising process. Such adaptive regulation enables progressive entropy adjustment, facilitating a smooth transition from initial exploration diversity to stable convergence toward the target distribution in later denoising stages." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Building upon this theoretical framework, we formalize AdaEnt as a plug-and-play module compatible with existing RL fine-tuning architectures for diffusion models. AdaEnt addresses two critical challenges in current RL frameworks: 1) sparse reward signals caused by insuffi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.221, + 0.888 + ], + "angle": 0, + "content": "*Equal Contribution" + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.238, + 0.9 + ], + "angle": 0, + "content": "† Corresponding Author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1924" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.332 + ], + "angle": 0, + "content": "cient optimization guidance during intermediate denoising steps and 2) diversity loss resulting from KL constraints. By incorporating entropy exploration terms at each denoising step and leveraging the \\(\\eta\\) parameter flexibility in DDIM, our method regulates entropy variations for principled exploration of denoising polices. The adaptive mechanism achieves dynamic balance between exploration and convergence through real-time quality assessment of intermediate outputs. Specifically, we introduce an image perception model trained alongside with the policy model to predict step-wise weights between diversity exploration and convergence stability in the optimization objective. Notably, the adaptive factors also serve as indicators of generation quality, which can be naturally used to reduce the total number of denoising steps, enabling efficient computation and lightweight storage." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.334, + 0.421, + 0.348 + ], + "angle": 0, + "content": "This paper highlights three main contributions:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.353, + 0.483, + 0.396 + ], + "angle": 0, + "content": "- We theoretically derive an entropy-based exploration strategy and present an effective implementation for diffusion models." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.398, + 0.483, + 0.471 + ], + "angle": 0, + "content": "- We propose an adaptive mechanism to balance the entropy-guided exploration and the regularization of real data distribution, which captures different characteristics in the early and late denoising process with a theoretical guarantee." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.473, + 0.483, + 0.533 + ], + "angle": 0, + "content": "- We propose a dynamic step alignment technique based on the entropy-guided exploration to automatically determine when to stop denoising, making our method computationally efficient for both training and inference." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.353, + 0.483, + 0.533 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.688 + ], + "angle": 0, + "content": "Experiments show that AdaEnt effectively mitigates reward hacking and achieves the highest rewards while preserving the diversity of images. The results generalize to various pre-trained diffusion models, reward functions, and RL frameworks. Furthermore, through visualization analysis, we confirm the useful design of entropy and self-adaptive mechanism, which seeks diversity at the early stage and gradually switches to maintain faithfulness in the late stage. AdaEnt works as a plug-and-play module that is broadly applicable in practice." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.703, + 0.233, + 0.718 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.728, + 0.315, + 0.743 + ], + "angle": 0, + "content": "2.1. RL for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Conventional training of diffusion models relies on likelihood-based objectives [12], yet these often misalign with perceptual quality metrics, resulting in degraded visual outputs [1, 16, 24]. Reinforcement learning (RL) [34] addresses this limitation by reframing the diffusion denoising process as a Markov Decision Process (MDP), enabling direct optimization of task-specific rewards such as aesthetic quality, compression efficiency, or human preference alignment. Building on pre-trained text-to-image models, RL demonstrates an effective and flexible framework to fur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "ther improve the diffusion model, where recent successful techniques for large language models, such as PPO [34] and DPO [28], can be applied to improve training stability and efficiency. Notable contributions include reward model-driven approaches like DDPO [1], SLCD [25] and DPOK [5], as well as human preference-aligned methods such as D3PO [49], Diffusion-DPO [42], DEPO [39], and HPS [45]. This paper follows this line of work, aiming to improve reward-driven performance while systematically mitigating the inherited drawbacks of over-optimization brought by RL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.268, + 0.836, + 0.284 + ], + "angle": 0, + "content": "2.2. Reward Hacking in Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.579 + ], + "angle": 0, + "content": "Reward hacking describes the phenomenon that the model maximizes the reward score but produces results that lack diversity or authenticity [8, 21, 35, 51]. For example, when the model overly prioritizes specific reward signals, such as image aesthetics or text alignment, it can lead to structural distortions, loss of diversity, or artifacts in generated images, ultimately compromising generation quality. This phenomenon has been partially attributed to sparse and delayed rewards, which affects reward optimization in both RL and direct back-propagation methods [4, 26, 41]. Methods that rely on direct access to the gradient of differentiable reward model pose higher risks of reward hacking [8]. While techniques like early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51] have been applied in RL methods to address this issue, the problem remains to be effectively solved [6, 7]. In contrast, this paper focuses on mitigating reward hacking in RL by adapting the policy optimization process and removing unnecessary steps to reduce computation costs." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.694, + 0.605 + ], + "angle": 0, + "content": "2.3. Entropy-based RL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.808 + ], + "angle": 0, + "content": "Entropy-based reinforcement learning (RL) integrates information-theoretic concepts into policy optimization to encourage exploration. By leveraging entropy regularization [9, 22, 40, 48], these methods encourage stochasticity in action selection and enhance learning stability and adaptability. However, aimlessly increase entropy for exploration will ultimately hurt the policy [11, 47]. For a better trade-off between exploration and exploitation, SAC [10] automates entropy coefficient tuning by jointly maximizing expected returns and entropy. To our knowledge, this paper is the first to leverage entropy as a theoretically guaranteed exploration metric for optimizing diffusion models and adaptively balances the effect of entropy and distribution regularization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.905, + 0.835 + ], + "angle": 0, + "content": "2.4. Dynamic-Step Adaptation for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Existing methodologies for acceleration can be divided into reducing the number of steps [15] or computation cost per step [18, 19, 43]. Still, they often use a uniform approach without accounting for each generated image's quality at" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1925" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "each denoising step. Complementary to the above methods, recent work [50] applies different denoising steps for each prompt by training a step prediction policy network. This paper proposes to dynamically early stop the denoising process to mitigate reward hacking, with an additional benefit of accelerating the RL training of diffusion models, which need not a dedicated model for step number determination." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.211, + 0.182, + 0.226 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.236, + 0.485, + 0.418 + ], + "angle": 0, + "content": "This paper introduces the adaptive exploration of diffusion policy optimization through the lens of entropy. In this section, we first prove the convergence of policy optimization with our new entropy-enhanced training objective in diffusion policy optimization. Coping with the trade-off between entropy-based exploration of RL and the inherent distribution modeling of the diffusion model, we then propose an adaptive mechanism to automatically balance the trade-off between exploration and the original objective, which is theoretically guaranteed to adjust entropy to allow a smooth transition from early-stage diversity exploration to late-stage data distribution maintenance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.426, + 0.432, + 0.441 + ], + "angle": 0, + "content": "3.1. Diffusion as Markov Decision Processes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.448, + 0.484, + 0.51 + ], + "angle": 0, + "content": "In diffusion models [12, 36], data distribution \\( p(\\mathbf{x}_0) \\) is modeled through a Markov forward process \\( q(\\mathbf{x}_t|\\mathbf{x}_{t-1}) \\) that gradually adds noise to the data and a reverse process \\( p_\\theta(\\mathbf{x}_{t-1}|\\mathbf{x}_t) \\) for generation, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.521, + 0.484, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\sqrt {1 - \\alpha_ {t}} \\mathbf {I}), \\\\ p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.569, + 0.484, + 0.614 + ], + "angle": 0, + "content": "where \\(\\sqrt{1 - \\alpha_t}\\) denotes the noise scale at timestep \\(t\\), \\(\\mu_{\\theta}(\\mathbf{x}_t,t)\\) is the predicted mean by the model, and \\(\\sigma_t^2\\) is a time-dependent variance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.484, + 0.66 + ], + "angle": 0, + "content": "Current RL fine-tuning methods treat the generation process of diffusion models as a Markov Decision Process (MDP), represented by the five-tuple \\(\\langle S, \\mathcal{A}, P, \\rho_0, R \\rangle\\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.672, + 0.482, + 0.817 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\mathbf {z}, t, \\mathbf {x} _ {T - t}) \\in S, \\\\ \\pi_ {\\theta} \\left(\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}\\right) \\triangleq p _ {\\theta} \\left(\\mathbf {x} _ {T - t - 1} \\mid \\mathbf {x} _ {T - t}, \\mathbf {z}\\right), \\\\ P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\mathbf {z}}, \\delta_ {T - t - 1}, \\delta_ {\\mathbf {x} _ {T - t - 1}}\\right), \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - t - 1}, \\tag {2} \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\mathbf {z}), \\delta_ {T}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ R \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left\\{ \\begin{array}{l l} r \\left(\\mathbf {x} _ {0}, \\mathbf {z}\\right) & \\text {i f} t = T, \\\\ 0 & \\text {o t h e r w i s e}. \\end{array} \\right. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{S}\\) denotes the state space, containing all possible image states \\(s_t\\) in the diffusion model's generation process; \\(\\mathcal{A}\\) represents the action space, corresponding to the denoising operations \\(a_t\\); \\(P(s_{t + 1}|s_t,a_t)\\) is the state transition probability, defining the probability of transitioning from state \\(s_t\\) to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.151 + ], + "angle": 0, + "content": "\\(s_{t + 1}\\) after taking action \\(a_t\\); \\(R(s_t, a_t)\\) represents the reward function, which typically evaluates the final image quality upon completion (i.e., at \\(t = T\\)); \\(\\rho_0\\) is the distribution of initial states." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.245 + ], + "angle": 0, + "content": "An important feature of diffusion models is that the sampling procedure has a wide design space after the score-matching training process, allowing a variance schedule ranging from larger to smaller. A popular uniform formulation of this sampling process is given in [37], where the generation process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.26, + 0.889, + 0.343 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mu_ {\\theta} (\\mathbf {x} _ {t - 1}) = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t})}{\\sqrt {\\alpha_ {t}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\eta^ {2} \\frac {(1 - \\alpha_ {t - 1})}{(1 - \\alpha_ {t})} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t}) \\\\ \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.346, + 0.905, + 0.399 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {t - 1} = \\eta \\sqrt {\\frac {\\left(1 - \\alpha_ {t - 1}\\right)}{\\left(1 - \\alpha_ {t}\\right)} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.4, + 0.907, + 0.492 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta}(\\mathbf{x}_{t - 1})\\) and \\(\\sigma_{t - 1}\\) represent the mean and variance of the current denoising step, \\(\\eta\\) determines the variance of each generation step. The \\(\\epsilon_{\\theta}^{(t)}(\\mathbf{x}_t)\\) is the neural network for predicting noise. Note that most algorithms only model the generation process for the case where \\(\\eta = 1\\), which restricts the policy search to a narrower space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.885, + 0.523 + ], + "angle": 0, + "content": "3.2. Provable Exploration Entropy for Diffusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.906, + 0.651 + ], + "angle": 0, + "content": "In reinforcement learning, the entropy of policy \\( H(\\pi) \\) represents the diversity of action choice on the current state. To increase the diversity with a more extensive exploration region, an entropy-related term is added to the original loss of RL fine-tuning. A common goal in fine-tuning the diffusion models is to maximize the expected reward of the generated images given the prompt distribution. Furthermore, the equation below holds true." + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.668, + 0.822, + 0.713 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ] \\\\ = \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0: T} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.728, + 0.853, + 0.743 + ], + "angle": 0, + "content": "From [5], we know the following expression holds:" + }, + { + "type": "equation", + "bbox": [ + 0.539, + 0.759, + 0.883, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) | | p _ {p r e} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.833, + 0.698, + 0.848 + ], + "angle": 0, + "content": "Similarly, we conclude that" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.863, + 0.882, + 0.905 + ], + "angle": 0, + "content": "\\[\nH \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right)\\right) = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} \\left(x _ {t} \\mid z\\right)} \\left[ H \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right)\\right) \\right],\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1926" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.895, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.297, + 0.908, + 0.355 + ], + "angle": 0, + "content": "Figure 1. AdaEnt's framework. The entropy-guided exploration expands policy's search area at each denoising step, regularized by the KL divergence of the current policy and reference policy. The trade-off between the two effects is coordinated by a perception model that evaluates the real-time noised images at each step. The prediction of the adaptive module also indicates unnecessary denoising steps and dynamically stops the denoising process to reduce the computation cost in training." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.381, + 0.122, + 0.394 + ], + "angle": 0, + "content": "and" + }, + { + "type": "equation", + "bbox": [ + 0.111, + 0.403, + 0.465, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) \\mid \\mid p _ {\\theta_ {k - 1}} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {\\theta_ {k - 1}} (x _ {t - 1}, z)) ]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.483, + 0.515 + ], + "angle": 0, + "content": "Finally, using the KL divergence and entropy, we propose the following objective for regularized training at the \\(t\\)-th step:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.52, + 0.482, + 0.66 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (z)} \\mathbb {E} _ {p _ {\\theta} (x _ {0: T} | z)} [ r (x _ {0}, z) + \\sum_ {t = 1} ^ {T} \\alpha_ {1} H (p _ {\\theta} (x _ {t - 1} | x _ {t}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {2} K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {3} K L \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right) \\mid p _ {\\theta_ {k - 1}} \\left(x _ {t - 1}, z\\right)\\right), \\tag {4} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.661, + 0.484, + 0.737 + ], + "angle": 0, + "content": "where \\(\\alpha_{1} \\geq 0\\), \\(\\alpha_{2}\\) and \\(\\alpha_{3} \\geq 0\\). A significant concern about including entropy in the loss function of policy optimization is that it may lead to unstable policies. To address this issue, we theoretically analyze the new loss function and prove that policy optimization is convergent." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.742, + 0.484, + 0.774 + ], + "angle": 0, + "content": "Theorem 3.1. Given the objective of Eq. (4), the optimal policy \\( p_{\\theta}^{*}(x_{t - 1}|x_{t},z) \\) has the following expression:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.779, + 0.47, + 0.812 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} ^ {*} (x _ {t - 1} | x _ {t}, z) \\propto \\exp \\{\\frac {Q ^ {*} (x _ {t - 1} , z) + \\alpha_ {2} \\log p _ {p r e} (x _ {t - 1} | x _ {t} , z)}{\\alpha_ {1} + \\alpha_ {2}} \\},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.482, + 0.832 + ], + "angle": 0, + "content": "where \\( Q^{*}(x_{t - 1},z) = r(s_{0},z)I_{\\{t - 1 = 0\\}} + (\\alpha_{1} + \\alpha_{2}) \\)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.833, + 0.431, + 0.854 + ], + "angle": 0, + "content": "\\(\\log \\sum_{x_{t - 2}}\\exp \\{\\frac{Q^{*}(x_{t - 2},z) + \\alpha_{2}\\log p_{pre}(x_{t - 2}|x_{t - 1},z)}{\\alpha_{1} + \\alpha_{2}}\\} .\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The proof of Theorem 3.1 is detailed in Appendix A.1. This theorem guarantees that with entropy added to training objective, the policy will ultimately converge." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.38, + 0.907, + 0.472 + ], + "angle": 0, + "content": "Theorem 3.2. When \\(\\hat{\\alpha}_1 \\leq \\alpha_1\\), we have \\(H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\hat{\\alpha}_1,\\alpha_2) \\leq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2)\\). Specially, compared to not including an entropy term, that is \\(\\alpha_1 = 0\\), in the objective function of Eq. (4), adding an entropy term results in a higher entropy after the algorithm converges, that is \\(H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2) \\geq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);0,\\alpha_2)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.568 + ], + "angle": 0, + "content": "The proof of Theorem 3.2 is detailed in Appendix A.2. This theorem demonstrates that adding an entropy term to the loss function increases the policy entropy, thereby enhancing the algorithm's exploratory behavior. Through greater exploration capability, the agent can discover better actions and achieve higher returns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.572, + 0.906, + 0.65 + ], + "angle": 0, + "content": "Theorem 3.3. If \\( C = \\min_{x_{t-1} \\neq y_{t-1}} [\\log p_{pre}(y_{t-1}|x_t, z) - \\log p_{pre}(x_{t-1}|x_t, z)] / [Q^*(x_{t-1}, z) - Q^*(y_{t-1}, z)] > 0 \\), the objective function of Eq. (4) contains only a KL term and no entropy term, that is \\( \\alpha_1 = 0 \\), if \\( 1/C \\leq \\hat{\\alpha}_2 \\leq \\alpha_2 \\), we have \\( H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\hat{\\alpha}_2) \\geq H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\alpha_2) \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.654, + 0.906, + 0.73 + ], + "angle": 0, + "content": "The proof of Theorem 3.3 is detailed in Appendix A.3. This theorem indicates that adding a KL term to the loss function may reduce the exploration capability of the policy if the parameters are not properly tuned, thereby affecting the algorithm's performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.741, + 0.824, + 0.756 + ], + "angle": 0, + "content": "4. AdaEnt: A Practical Instantiation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.906, + 0.843 + ], + "angle": 0, + "content": "Building on the above theoretical foundation, we propose AdaEnt, a plug-and-play module that adaptatively controls entropy and a number of steps to combat reward hacking in diffusion policy optimization. The architecture of AdaEnt is illustrated in Fig. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.771, + 0.866 + ], + "angle": 0, + "content": "4.1. Entropy-Driven Exploration" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To cope with the diversity degradation brought by reward hacking issue, we introduce the entropy of denoising policy" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1927" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "\\(p_{\\theta}(x_{t - 1}|x_t,z)\\) to enhance exploration. In a normal distribution, the entropy is related to its variance. From Eq. (3), we define a new variance:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.144, + 0.482, + 0.185 + ], + "angle": 0, + "content": "\\[\n\\sigma_ {T - t} = \\eta_ {\\theta , \\phi} (\\mathbf {x} _ {T - t}, T - t) \\sqrt {\\frac {1 - \\alpha_ {T - t - 1}}{(1 - \\alpha_ {T - t})} \\cdot \\left(1 - \\frac {\\alpha_ {T - t}}{\\alpha_ {T - t - 1}}\\right)},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.483, + 0.239 + ], + "angle": 0, + "content": "where \\(\\eta_{\\theta ,\\phi}(\\mathbf{x}_{T - t},T - t)\\) is a dynamic, lightweight head with parameters \\(\\phi\\) , built upon the neural network \\(\\mu_{\\theta}\\) ,designed to provide a latent and step-aware dynamic variance." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.247, + 0.483, + 0.276 + ], + "angle": 0, + "content": "4.2. Trade-Off between Exploration and Regularization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.484, + 0.374 + ], + "angle": 0, + "content": "In diffusion models, different denoising steps play various roles in image generation. Previous research has shown that earlier steps significantly impact the semantic content and global layout of the image, while later steps influence the texture and details [3]. Motivated by this phenomenon, we propose an adaptive regularization scheme." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.451 + ], + "angle": 0, + "content": "We introduce an adaptive balance factor \\(\\gamma_{t}\\) to dynamically adjust the balance between exploration diversity and denoising convergence across the denoising process. For the sake of simplicity, we introduce the PPO algorithm [34]. Therefore, the objective to optimize (Eq. (4)) is updated:" + }, + { + "type": "equation", + "bbox": [ + 0.095, + 0.459, + 0.487, + 0.547 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {J} = \\max _ {\\theta} \\mathbb {E} [ \\min (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)} A, c l i p (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)}, 1 - \\epsilon , 1 + \\epsilon) A) \\\\ + \\sum_ {t = 1} ^ {T} \\left(1 - \\gamma_ {t}\\right) H \\left(p _ {\\theta} (\\cdot)\\right) - \\sum_ {t = 1} ^ {T} \\gamma_ {t} K L \\left(p _ {\\theta} (\\cdot)\\right) \\left\\| p _ {p r e} (\\cdot)\\right), \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.483, + 0.608 + ], + "angle": 0, + "content": "where \\( A \\) is an advantage function. According to Eq. (5), as the generation progresses, the regularization term will dominate the optimization and seek directions toward the distribution of real images within the high-reward region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.683 + ], + "angle": 0, + "content": "A prediction model \\( f(\\mathbf{x}_t) \\) is trained on each intermediate denoised (latent) image \\( \\mathbf{x}_t \\), with intermediate denoised images close to the final generated image \\( \\mathbf{x}_0 \\) labeled as 1, and those close to initial noise \\( \\mathbf{x}_T \\) labeled as 0. The binary classification loss function is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.092, + 0.69, + 0.483, + 0.734 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {p r e d} = - \\sum_ {i} \\left(y _ {i} \\log \\left(f \\left(x _ {i}\\right)\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - f \\left(x _ {i}\\right)\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.84 + ], + "angle": 0, + "content": "where \\( y_{i} \\in \\{0,1\\} \\) is the label indicating whether the image is noiseless, and \\( f(x_{i}) \\) is the model's predicted probability that image \\( x_{i} \\) is noiseless. After training, this network can evaluate the denoising level of any intermediate image during the denoising process, and we let \\( \\gamma_{t} = f(\\mathbf{x}_{t}) \\) as a measure of the denoising process indicating how well an intermediate image aligns with real image distribution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.305, + 0.866 + ], + "angle": 0, + "content": "4.3. Dynamic Step Decision" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "One limitation of diffusion models is the reliance on many iterative denoising steps. The drawback will be exacerbated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "when performing RL fine-tuning, leading to a high computation demand and exacerbation of reward hacking. As later denoising steps contribute less to the overall appearance of the generated image, applying an early stop on the denoising process will make training and inference more efficient and mediately alleviate reward hacking. The criterion of early stop is dynamically determined by how well the current image aligns with true image distribution. According to Sec. 4.2, \\(\\gamma_{t} = 1\\) means the current state of image satisfies the quality. Hence we apply early stop in this scenario, which is completely dynamic according to the complexity of generation and current image quality, without any predefined hyperparameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.29, + 0.907, + 0.471 + ], + "angle": 0, + "content": "Note that in existing diffusion step adaptation methods, the total number of denoising steps have to be determined before hand, which is impossible for the adaptive module to make a decision based on sufficient information about the image quality. In contrast, our adaptation mechanism monitors the generated image at each step and determines the total number of steps based on real-time quality of the current state. Recent study on denoising diffusion models reveals the temporal information may be not necessary [38], which supports our design of dynamically cutting off the iterative denoising process. The algorithmic procedure of AdaEnt can be found in Appendix B." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.489, + 0.646, + 0.506 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.704, + 0.532 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.689 + ], + "angle": 0, + "content": "Datasets. To compare with previous work, we use the same prompt set consisting of 45 different animal classes for training (Simple-animals), which has been widely adopted for reward fine-tuning [1, 49, 51]. For testing, we generate 8 different noises for each animal class, totaling 360 test samples. For generalization evaluation, larger-scale prompt sets are further considered in testing as done in [1], including 398 animal classes from ImageNet (ImageNet-animals) and the above 45 animals each concatenated with an activity (Activity-animals)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Rewards and Metrics. In this paper, we first deploy three metrics as the reward functions, that is, Aesthetic Score [33], Compressibility, and Incompressibility. To quantitatively evaluate performance in face of reward hacking, we then introduce cross-reward generalization and well-acknowledged metrics as suggested in [51]: ImageReward [46], Inception Score (IS) [32], PickScore [17], and ClipScore [27]. Details of each metric can be found in Appendix C." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Implementation Details. Following [1], we take Stable Diffusion v1.4 (SDv14) as our main test bed for reward finetuning. We also consider Stable Diffusion v1.5 (SDv15), v2.1-turbo (SDv21), and XL1.0 (XL) to verify the robustness of our method with different backbones. All models" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1928" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.357, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.089, + 0.637, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.089, + 0.907, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.261, + 0.908, + 0.292 + ], + "angle": 0, + "content": "Figure 2. Reward curves of RL fine-tuning SDv14 with three different reward functions. Our proposed method consistently outperforms baselines in all three experimental scenarios, in terms of both peak achieved rewards and the sample efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.362 + ], + "angle": 0, + "content": "are fine-tuned with low-rank adaptation (LoRA) [13] on 8 NVIDIA Tesla A100 GPUs. More training details can be found in Appendix C." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.376, + 0.231, + 0.391 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.399, + 0.483, + 0.642 + ], + "angle": 0, + "content": "We benchmark our methods against state-of-the-art RL-based diffusion fine-tuning frameworks, including DDPO [1], DPOK [5], D3PO [49], and TDPO [51]. As shown in Fig. 2, our method demonstrates consistent superiority in learning all three reward functions, validating its effectiveness for optimizing downstream objectives. Crucially, our balanced state exploration and adaptive regularization mechanism achieves promote the rate of reward growth by a large margin compared to DPOK's fixed-weight regularization strategy. Unlike dense-reward methods requiring evaluations of noised states' long-term impacts, our approach implements stepwise adaptive entropy to guide denoising space exploration, enabling sample-efficient optimization. Consequently, we attain rewards slightly higher than the latest dense-reward method TDPO but using much fewer training samples." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.643, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Moreover, to systematically investigate reward hacking mitigation, we test the models optimized for Aesthetic Score with different aesthetic scores and general metrics for image quality and text-to-image alignment. We report the min, mean, and max values of each metric across the entire training process. As shown in Tab. 1, our generated images achieves overall higher scores on the diverse evaluation perspectives, effectively circumventing reward hacking phenomena even under elevated reward conditions. The results indicate that the proposed exploratory entropy effectively contributes to improved diversity and visual fidelity of generated images, along with less overfitting to the learning reward function." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "In summary, through this balanced exploration-regularization paradigm, our method demonstrates superior capability in discovering latent high-reward policies, thereby overcoming conventional performance plateaus" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.312, + 0.708, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.313, + 0.904, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.399, + 0.907, + 0.485 + ], + "angle": 0, + "content": "Figure 3. Distribution visualizations of our intermediate variables. (a) Distribution of predicted \\(\\gamma\\) during training. The predictor progressively learns to effectively distinguish latent codes of different stages. (b) Distribution of denoising step number during training. Our adaptive step mechanism effectively reduces the denoising steps in training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.512, + 0.907, + 0.543 + ], + "angle": 0, + "content": "and achieving measurable enhancements in both generation quality and quantitative reward metrics." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.557, + 0.695, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.695, + 0.557, + 0.906, + 0.661 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.673, + 0.907, + 0.729 + ], + "angle": 0, + "content": "Figure 4. Left: Entropy can be optimized in AdaEnt under different scaling factors. Right: Excessive application of either entropy or KL regularization impedes reward learning and may ultimately induce catastrophic collapse." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.757, + 0.669, + 0.774 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Effect of \\(\\gamma\\) Predictor. The coefficient \\(\\gamma\\) is introduced to adaptively balance the exploratory entropy and KL regularization based on the current denoising situation. In Fig. 3 (a), we show the distributions of \\(\\gamma\\) with 50-step denoising during training. It can be observed that the predictor progressively learns to effectively differentiate latent codes that are more distinctive (i.e., those at the beginning and end of denoising), whereas the results in the intermediate steps are" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1929" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Quality assessment from multiple unseen metric perspectives. All results are evaluated across training on the aesthetic score and sampled with the same training iteration. * refers to the Inception Score being sampled with the same aesthetic score of different methods. Since D3PO performs limited in learning rewards (see Fig. 2), the range of aesthetic score data is not enough for inception score collection. #Top1 indicates the instances when each method achieves the best performance across all metrics." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.156, + 0.915, + 0.284 + ], + "angle": 0, + "content": "
MethodPickScoreInceptionScore*ImageRewardClipScore#Top1
MinMeanMaxMinMeanMaxMinMeanMaxMinMeanMax
DDPO [1]21.2821.3721.5321.1122.9523.670.89491.07071.24930.29660.30190.30831
DPOK [5]21.2721.4021.5320.4822.2923.460.90601.04961.13400.30080.30460.30831
D3PO [49]21.2821.2921.31---0.86230.89620.93170.30000.30120.30221
TDPO [51]21.0621.3821.7119.9421.8823.950.00480.84531.13540.27540.30100.31132
AdaEnt21.2421.4121.6422.1124.3426.190.90781.12121.31230.30000.30500.30988
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.304, + 0.482, + 0.487 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.499, + 0.483, + 0.569 + ], + "angle": 0, + "content": "Figure 5. Comparison of aesthetic reward curves and visual quality on SDv15 and SDv21. Our method consistently achieves higher rewards without over-optimizing across different backbones, while baseline methods compromise either visual quality or text-image alignment to attain aesthetic rewards." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.614, + 0.483, + 0.902 + ], + "angle": 0, + "content": "predicted with non-binary confidence. This adaptive behavior dynamically balances entropy and the KL divergence as the denoising process progresses, facilitating a shift from exploration to regularization during each denoising process. Balance between Entropy and KL Divergence. In Fig. 4, we conduct an in-depth investigation into the effects of introducing entropy in reward learning. Specifically, with a fixed KL, we design ablation variants with entropy scaled by different factors (i.e., \\(\\mu_{e}\\)). The baselines, DDPO and DPOK, have a fixed schedule of entropy (according to Eq. (1)). In our proposed framework, it is observed that entropy can be optimized to increase under different scaling factors, and corresponding with a better learning of reward. However, when entropy's order of magnitude overwhelms KL divergence (i.e., \\(\\mu_{e} = 1.0\\)), over-exploration may lead to the collapse of training. Therefore, the result demonstrates the necessity of introducing the trade-off between entropy and KL, which ensures a well-balanced training process for improved generating performance." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.306, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 2. Generalization results of SDv15 and SDv21 on unseen metrics. The metrics are abbreviated and the same in Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.345, + 0.905, + 0.459 + ], + "angle": 0, + "content": "
MethodSDv15SDv21#Top1
PSISIRCSPSISIRCS
DDPO21.1716.900.820.3020.347.89-0.210.260
DPOK21.1818.680.840.3120.5310.280.040.272
D3PO21.0122.060.800.3318.843.14-1.580.252
TDPO19.682.40-1.570.2418.521.56-2.090.240
AdaEnt21.2118.990.850.3020.598.140.100.275
" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.477, + 0.698, + 0.581 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.478, + 0.904, + 0.568 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.634 + ], + "angle": 0, + "content": "Figure 6. Comparison of aesthetic reward curves and visual quality on SD-XL. Notably, we observe that our method is the only one that successfully works on SD-XL." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.795, + 0.679 + ], + "angle": 0, + "content": "5.4. Results on Different Backbones." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.688, + 0.907, + 0.839 + ], + "angle": 0, + "content": "In order to verify the generalizability of our method across different models for practical applications, we hereby extend our experiment to SDv15 and SDv21. In Fig. 5, our method also achieves superior effectiveness in reward learning. For generalization evaluation, we also evaluate the generated image quality with more metric perspectives for SDv15 and SDv21. As shown in Tab. 2, AdaEnt achieves superior overall quality, and despite TDPO shows comparable aesthetic score with our method, it is demonstrated to be failed in prompt-image alignments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Moreover, considering that SD-XL is one of the most advanced and extensive pre-trained model, we provide a concise reward-learning experiment on SD-XL in Fig. 6. The results demonstrate that AdaEnt can still bring reward gains" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1930" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.483, + 0.266 + ], + "angle": 0, + "content": "Figure 7. Comparison of generalization ability to larger scale prompt sets with unseen prompts. The experiments are conducted on SDv14 and trained with the small Simple-animals set." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.28, + 0.481, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.434, + 0.484, + 0.504 + ], + "angle": 0, + "content": "Figure 8. Comparison of generalization ability to more complex prompts with multiple categories. All methods are implemented in SD-XL. The performance of our method is highlighted in darker colors, demonstrating AdaEnt's superior performance in both image quality and prompt-image alignment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.533, + 0.341, + 0.548 + ], + "angle": 0, + "content": "for larger and more powerful models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.561, + 0.395, + 0.577 + ], + "angle": 0, + "content": "5.5. Generalization to Unseen Prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.645 + ], + "angle": 0, + "content": "To investigate whether the performance of RL fine-tuned models is limited to the prompts seen in training, which is an ordinary case of reward hacking, we hereby conduct prompts generalization experiments from two perspectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.764 + ], + "angle": 0, + "content": "Generalization to larger-scale prompt sets. To evaluate the generalization ability to unseen prompts, we evaluate the aesthetic score on ImageNet-animals and Activity-animals. The experiments are conducted on SDv14, using prompts from Simple-animals as the training set. As presented in Fig. 7, our method achieves promising aesthetic scores on the unseen datasets, which demonstrates the effectiveness of our method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Generalization to more complex prompts. Additionally, we assess the performance on more complex prompts. To be specific, we further conduct experiments on SD-XL with a subset of the HPSv2 dataset [44], which contains complex prompts that are not limited to animals. We identify several categories in the prompt set and measure the prompt-image consistency using clip-score. We also provide the aesthetic score to assess the quality of generated images. As presented in Fig. 8, our method exhibits the best perfor" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.123 + ], + "angle": 0, + "content": "mance in the aesthetic score and the maintenance of image-prompts alignment, which suggests our superiority." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.132, + 0.816, + 0.148 + ], + "angle": 0, + "content": "5.6. Computation Efficiency of AdaEnt" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Since the adaptive mechanism introduced in our work allows for the early termination of the denoising process once denoising is essentially complete, it is expected to improve the training efficiency of diffusion models. Firstly, we provide the visualized distributions of denoising step number during training in Fig. 3 (b), which verifies the capability of adaptively reducing the step number. Quantitatively, we take the aesthetic score \\(= 7\\) as the training terminals of each method for evaluation. In Tab. 3, we provide GFLOPs and time for single batch (size=12) sampling, along with the total sampling step and time for the overall training process. The results show that the proposed AdaEnt can significantly accelerate the training process with almost no increase in computational cost (i.e., GFLOPS)." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.379, + 0.907, + 0.435 + ], + "angle": 0, + "content": "Table 3. Computation Efficiency of integrating our proposed dynamic early stop mechanism with existing RL frameworks. TDS refers to total denoising steps. Results of both overall training and single-batch inference are reported." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.446, + 0.89, + 0.561 + ], + "angle": 0, + "content": "
MethodOverallSingle batch
TDSTimeGFLOPSTime
DDPO1.843+e611.98h1.034e+069.56s
DPOK2.015+e614.76h9.69s
DDPO+Ours0.242+e65.52h1.044e+066.01s
DPOK+Ours0.294+e67.16h7.12s
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.589, + 0.634, + 0.604 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work addresses the critical challenge of reward hacking in diffusion model optimization through reinforcement learning, where conventional RL frameworks tend to produce overfitted outputs that compromise distributional fidelity and generation diversity. Our theoretical analyses establish entropy as a proxy to enhance the exploratory capacity of denoising policies. The theoretical basis is instantiated as AdaEnt, a framework that achieves adaptive entropy control and step-wise optimization through three interconnected mechanisms: entropy-aware policy updates, distributional consistency regularization, and noise-level dependent exploration scheduling. Extensive experiments demonstrate that our method effectively mitigates reward hacking while preserving the diffusion model's inherent generation capabilities, establishing new state-of-the-art performance in quality-diversity trade-off across multiple diffusion-based RL benchmarks. The adaptive entropy proposed in this paper offers new perspectives for future RL fine-tuning of diffusion models." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1931" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.264, + 0.108 + ], + "angle": 0, + "content": "7. Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.237 + ], + "angle": 0, + "content": "This paper is supported by the National Science Foundation of China under Grant No.62025401, 62341407, U24B200170, 62406008; Beijing Municipal Science and Technology Program under Grant Z241100004224015; Beijing Natural Science Foundation under Grant No.L223004; '111' Project under Grant No.B18001. This work is also supported in part by the Natural Science Foundation of China under Grant No. 62222606." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.251, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.276, + 0.484, + 0.331 + ], + "angle": 0, + "content": "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.333, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[2] Bo Dai, Albert E. Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, and Le Song. SBEED: convergent reinforcement learning with nonlinear function approximation. In Proceedings of International Conference on Machine Learning., 2018. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.403, + 0.484, + 0.445 + ], + "angle": 0, + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.447, + 0.484, + 0.516 + ], + "angle": 0, + "content": "[4] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.518, + 0.484, + 0.601 + ], + "angle": 0, + "content": "[5] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.484, + 0.644 + ], + "angle": 0, + "content": "[6] Yaozhong Gan, Renye Yan, Xiaoyang Tan, Zhe Wu, and Jun-liang Xing. Transductive off-policy proximal policy optimization. arXiv preprint arXiv:2406.03894, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.645, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[7] Yaozhong Gan, Renye Yan, Zhe Wu, and Junliang Xing. Reflective policy optimization. arXiv preprint arXiv:2406.03678, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.689, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[8] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835-10866. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.731, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[9] Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International conference on machine learning, pages 1352-1361. PMLR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, pages 1861-1870. Pmlr, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[11] Seungyul Han and Youngchul Sung. A max-min entropy framework for reinforcement learning. Advances in Neural Information Processing Systems, 34:25732-25745, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.276, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.264 + ], + "angle": 0, + "content": "[14] Rohit Jena, Ali Taghibakhshi, Sahil Jain, Gerald Shen, Nima Tajbakhsh, and Arash Vahdat. Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models. arXiv preprint arXiv:2409.06493, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.267, + 0.905, + 0.335 + ], + "angle": 0, + "content": "[15] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.905, + 0.381 + ], + "angle": 0, + "content": "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.383, + 0.905, + 0.451 + ], + "angle": 0, + "content": "[17] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[18] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. CoRR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.513, + 0.908, + 0.569 + ], + "angle": 0, + "content": "[19] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[20] Jincheng Mei, Chenjun Xiao, Ruitong Huang, Dale Schuurmans, and Martin Müller. On principled entropy exploration in policy optimization. In Proceedings of International Joint Conference on Artificial Intelligence, 2019. 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[21] Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, and Dacheng Tao. Inform: Mitigating reward hacking in rlhf via information-theoretic reward modeling. Advances in Neural Information Processing Systems, 37: 134387-134429, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[22] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.773, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.908, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1932" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[25] Owen Oertell, Shikun Sun, Yiding Chen, Jin Peng Zhou, Zhiyong Wang, and Wen Sun. Efficient controllable diffusion via optimal classifier guidance. arXiv preprint arXiv:2505.21666, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[34] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[35] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming. Advances in Neural Information Processing Systems, 35:9460-9471, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.816, + 0.482, + 0.884 + ], + "angle": 0, + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.887, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[38] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv preprint arXiv:2502.13129, 2025. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[39] Shikun Sun, Chengrui Wang, Min Zhou, Zixuan Wang, Xiaoyu Qin, Tiezheng Ge, Bo Zheng, and Jia Jia. Depo: Enhancing e-commerce image background generation with short trajectory direct expected preference optimization. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.304 + ], + "angle": 0, + "content": "[40] Marc Toussaint. Robot trajectory optimization using approximate inference. In Proceedings of the 26th annual international conference on machine learning, pages 1049-1056, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[41] Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Finetuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.476 + ], + "angle": 0, + "content": "[42] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.477, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[43] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[44] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[45] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.704, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[46] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[47] Renye Yan, Zhe Wu, Yuan Zhan, Pin Tao, Zongwei Wang, Yimao Cai, and Junliang Xing. Mnemonic dictionary learning for intrinsic motivation in reinforcement learning. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[48] Renye Yan, Yaozhong Gan, You Wu, Ling Liang, Junliang Xing, Yimao Cai, and Ru Huang. The exploration-exploitation dilemma revisited: An entropy perspective. arXiv preprint arXiv:2408.09974, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1933" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[49] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.22, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[51] Ziyi Zhang, Sen Zhang, Yibing Zhan, Yong Luo, Yonggang Wen, and Dacheng Tao. Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. In International Conference on Machine Learning, pages 60396-60413. PMLR, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1934" + } + ] +] \ No newline at end of file diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_origin.pdf b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..bba195282b358268596390cbad0b6263c8270d80 --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/c4a55a51-0565-4ccd-a910-9c67e543b847_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e00c5b91824ae2e79d4433d86abe718f52952034cd84abc0248d33fb51cae033 +size 4461346 diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/full.md b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f8ee4ae318624fe3668840bb3c2f6a6cce06b554 --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/full.md @@ -0,0 +1,351 @@ +# Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment + +Renye Yan $^{1*}$ , Jikang Cheng $^{1*}$ , Yaozhong Gan $^{4}$ , Shikun Sun $^{3}$ , You Wu $^{4}$ , Yunfan Yang $^{1}$ , Ling Liang $^{1}$ , Jinlong Lin $^{1}$ , Yeshuang Zhu $^{2\dagger}$ , Jie Zhou $^{2}$ , Jinchao Zhang $^{2}$ , Junliang Xing $^{3}$ , Yimao Cai $^{1\dagger}$ , Ru Huang $^{1}$ + +Peking University $^{1}$ , Tencent Inc $^{2}$ , Tsinghua University $^{3}$ , Nanjing University $^{4}$ . + +victory@stu.pku.edu.cn, jikangcheng99@gmail.com, yshzhu@tencent.com, caiyimao@pu.edu.cn + +# Abstract + +While fine-tuning diffusion models with reinforcement learning (RL) has demonstrated effectiveness in directly optimizing downstream objectives, existing RL frameworks are prone to overfitting the rewards, leading to outputs that deviate from the true data distribution and exhibit reduced diversity. To address this issue, we introduce entropy as a quantitative measure to enhance the exploratory capacity of diffusion models' denoising policies. We propose an adaptive mechanism that dynamically adjusts the application and magnitude of entropy and regularization, guided by real-time quality estimation of intermediate noised states. Theoretically, we prove the convergence of our entropy-enhanced policy optimization and establish two critical properties: 1) global entropy increases through training, ensuring robust exploration capabilities, and 2) entropy systematically decreases during the denoising process, enabling a phase transition from early-stage diversity promotion to late-stage distributional fidelity. Building on this foundation, we propose a plug-and-play RL module that adaptively controls entropy and optimizes denoising steps. Extensive evaluations demonstrate theoretical soundness and empirical robustness of our method, achieving state-of-the-art quality-diversity trade-offs across benchmarks. Notably, our framework significantly improves the rewards and reduces denoising steps in training by up to $40\%$ . + +# 1. Introduction + +In recent years, diffusion models [12, 23, 29-31, 36] have demonstrated powerful capabilities in text-to-image generation. However, these models are designed to match the distribution of training data, lacking the ability to directly optimize downstream objectives, such as aesthetics and preference scores. To this end, reinforcement learning (RL) has been explored and shown promising potential in fine + +tuning diffusion models with various reward functions or human feedback (i.e., RLAIF and RLHF) [1, 5, 49]. In these methods, policy optimization is guided by a global reward computed on the final generated image, which is either differentiable or non-differentiable, making RL a flexible and promising solution to overcoming the limitations of supervised fine-tuning for diffusion models. + +However, current RL methods rely solely on rewards from the final generated image to update policies, overlooking the importance of exploration in promoting diversity during denoising. This makes fine-tuned models susceptible to reward hacking [14], where generated images achieve high scores in reward models but significantly deviates from the true distribution, exhibiting deteriorated quality and low diversity. This phenomenon has been partially attributed to sparse and delayed rewards and addressed with early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51]. However, the problem remains to be effectively solved. + +This paper proposes a novel RL method with adaptive entropy regulation to address the reward hacking issue in fine-tuning diffusion models. We formally introduce entropy as a quantitative indicator to assess the exploratory capability of denoising strategies in diffusion models, integrating it into the RL optimization objectives. Theoretical analysis establishes the convergence guarantees of the proposed method for newly derived policies. Furthermore, to overcome the limitations of fixed parameterization in conventional KL regulation, we derive an adaptive mechanism that dynamically balances exploration diversity and convergence stability throughout the denoising process. Such adaptive regulation enables progressive entropy adjustment, facilitating a smooth transition from initial exploration diversity to stable convergence toward the target distribution in later denoising stages. + +Building upon this theoretical framework, we formalize AdaEnt as a plug-and-play module compatible with existing RL fine-tuning architectures for diffusion models. AdaEnt addresses two critical challenges in current RL frameworks: 1) sparse reward signals caused by insuffi + +cient optimization guidance during intermediate denoising steps and 2) diversity loss resulting from KL constraints. By incorporating entropy exploration terms at each denoising step and leveraging the $\eta$ parameter flexibility in DDIM, our method regulates entropy variations for principled exploration of denoising polices. The adaptive mechanism achieves dynamic balance between exploration and convergence through real-time quality assessment of intermediate outputs. Specifically, we introduce an image perception model trained alongside with the policy model to predict step-wise weights between diversity exploration and convergence stability in the optimization objective. Notably, the adaptive factors also serve as indicators of generation quality, which can be naturally used to reduce the total number of denoising steps, enabling efficient computation and lightweight storage. + +This paper highlights three main contributions: + +- We theoretically derive an entropy-based exploration strategy and present an effective implementation for diffusion models. +- We propose an adaptive mechanism to balance the entropy-guided exploration and the regularization of real data distribution, which captures different characteristics in the early and late denoising process with a theoretical guarantee. +- We propose a dynamic step alignment technique based on the entropy-guided exploration to automatically determine when to stop denoising, making our method computationally efficient for both training and inference. + +Experiments show that AdaEnt effectively mitigates reward hacking and achieves the highest rewards while preserving the diversity of images. The results generalize to various pre-trained diffusion models, reward functions, and RL frameworks. Furthermore, through visualization analysis, we confirm the useful design of entropy and self-adaptive mechanism, which seeks diversity at the early stage and gradually switches to maintain faithfulness in the late stage. AdaEnt works as a plug-and-play module that is broadly applicable in practice. + +# 2. Related Work + +# 2.1. RL for Diffusion Models + +Conventional training of diffusion models relies on likelihood-based objectives [12], yet these often misalign with perceptual quality metrics, resulting in degraded visual outputs [1, 16, 24]. Reinforcement learning (RL) [34] addresses this limitation by reframing the diffusion denoising process as a Markov Decision Process (MDP), enabling direct optimization of task-specific rewards such as aesthetic quality, compression efficiency, or human preference alignment. Building on pre-trained text-to-image models, RL demonstrates an effective and flexible framework to fur + +ther improve the diffusion model, where recent successful techniques for large language models, such as PPO [34] and DPO [28], can be applied to improve training stability and efficiency. Notable contributions include reward model-driven approaches like DDPO [1], SLCD [25] and DPOK [5], as well as human preference-aligned methods such as D3PO [49], Diffusion-DPO [42], DEPO [39], and HPS [45]. This paper follows this line of work, aiming to improve reward-driven performance while systematically mitigating the inherited drawbacks of over-optimization brought by RL. + +# 2.2. Reward Hacking in Diffusion Models + +Reward hacking describes the phenomenon that the model maximizes the reward score but produces results that lack diversity or authenticity [8, 21, 35, 51]. For example, when the model overly prioritizes specific reward signals, such as image aesthetics or text alignment, it can lead to structural distortions, loss of diversity, or artifacts in generated images, ultimately compromising generation quality. This phenomenon has been partially attributed to sparse and delayed rewards, which affects reward optimization in both RL and direct back-propagation methods [4, 26, 41]. Methods that rely on direct access to the gradient of differentiable reward model pose higher risks of reward hacking [8]. While techniques like early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51] have been applied in RL methods to address this issue, the problem remains to be effectively solved [6, 7]. In contrast, this paper focuses on mitigating reward hacking in RL by adapting the policy optimization process and removing unnecessary steps to reduce computation costs. + +# 2.3. Entropy-based RL + +Entropy-based reinforcement learning (RL) integrates information-theoretic concepts into policy optimization to encourage exploration. By leveraging entropy regularization [9, 22, 40, 48], these methods encourage stochasticity in action selection and enhance learning stability and adaptability. However, aimlessly increase entropy for exploration will ultimately hurt the policy [11, 47]. For a better trade-off between exploration and exploitation, SAC [10] automates entropy coefficient tuning by jointly maximizing expected returns and entropy. To our knowledge, this paper is the first to leverage entropy as a theoretically guaranteed exploration metric for optimizing diffusion models and adaptively balances the effect of entropy and distribution regularization. + +# 2.4. Dynamic-Step Adaptation for Diffusion Models + +Existing methodologies for acceleration can be divided into reducing the number of steps [15] or computation cost per step [18, 19, 43]. Still, they often use a uniform approach without accounting for each generated image's quality at + +each denoising step. Complementary to the above methods, recent work [50] applies different denoising steps for each prompt by training a step prediction policy network. This paper proposes to dynamically early stop the denoising process to mitigate reward hacking, with an additional benefit of accelerating the RL training of diffusion models, which need not a dedicated model for step number determination. + +# 3. Method + +This paper introduces the adaptive exploration of diffusion policy optimization through the lens of entropy. In this section, we first prove the convergence of policy optimization with our new entropy-enhanced training objective in diffusion policy optimization. Coping with the trade-off between entropy-based exploration of RL and the inherent distribution modeling of the diffusion model, we then propose an adaptive mechanism to automatically balance the trade-off between exploration and the original objective, which is theoretically guaranteed to adjust entropy to allow a smooth transition from early-stage diversity exploration to late-stage data distribution maintenance. + +# 3.1. Diffusion as Markov Decision Processes + +In diffusion models [12, 36], data distribution $p(\mathbf{x}_0)$ is modeled through a Markov forward process $q(\mathbf{x}_t|\mathbf{x}_{t-1})$ that gradually adds noise to the data and a reverse process $p_\theta(\mathbf{x}_{t-1}|\mathbf{x}_t)$ for generation, defined as: + +$$ +\begin{array}{l} q (\mathbf {x} _ {t} | \mathbf {x} _ {t - 1}) = \mathcal {N} (\mathbf {x} _ {t}; \sqrt {\alpha_ {t}} \mathbf {x} _ {t - 1}, \sqrt {1 - \alpha_ {t}} \mathbf {I}), \\ p _ {\theta} \left(\mathbf {x} _ {t - 1} \mid \mathbf {x} _ {t}\right) = \mathcal {N} \left(\mathbf {x} _ {t - 1}; \mu_ {\theta} \left(\mathbf {x} _ {t}, t\right), \sigma_ {t} ^ {2} \mathbf {I}\right), \tag {1} \\ \end{array} +$$ + +where $\sqrt{1 - \alpha_t}$ denotes the noise scale at timestep $t$ , $\mu_{\theta}(\mathbf{x}_t,t)$ is the predicted mean by the model, and $\sigma_t^2$ is a time-dependent variance. + +Current RL fine-tuning methods treat the generation process of diffusion models as a Markov Decision Process (MDP), represented by the five-tuple $\langle S, \mathcal{A}, P, \rho_0, R \rangle$ : + +$$ +\begin{array}{l} \mathbf {s} _ {t} \triangleq (\mathbf {z}, t, \mathbf {x} _ {T - t}) \in S, \\ \pi_ {\theta} \left(\mathbf {a} _ {t} \mid \mathbf {s} _ {t}\right) \triangleq p _ {\theta} \left(\mathbf {x} _ {T - t - 1} \mid \mathbf {x} _ {T - t}, \mathbf {z}\right), \\ P \left(\mathbf {s} _ {t + 1} \mid \mathbf {s} _ {t}, \mathbf {a} _ {t}\right) \triangleq \left(\delta_ {\mathbf {z}}, \delta_ {T - t - 1}, \delta_ {\mathbf {x} _ {T - t - 1}}\right), \\ \mathbf {a} _ {t} \triangleq \mathbf {x} _ {T - t - 1}, \tag {2} \\ \rho_ {0} \left(\mathbf {s} _ {0}\right) \triangleq (p (\mathbf {z}), \delta_ {T}, \mathcal {N} (\mathbf {0}, \mathbf {I})) \\ R \left(\mathbf {s} _ {t}, \mathbf {a} _ {t}\right) \triangleq \left\{ \begin{array}{l l} r \left(\mathbf {x} _ {0}, \mathbf {z}\right) & \text {i f} t = T, \\ 0 & \text {o t h e r w i s e}. \end{array} \right. \\ \end{array} +$$ + +where $\mathcal{S}$ denotes the state space, containing all possible image states $s_t$ in the diffusion model's generation process; $\mathcal{A}$ represents the action space, corresponding to the denoising operations $a_t$ ; $P(s_{t + 1}|s_t,a_t)$ is the state transition probability, defining the probability of transitioning from state $s_t$ to + +$s_{t + 1}$ after taking action $a_t$ ; $R(s_t, a_t)$ represents the reward function, which typically evaluates the final image quality upon completion (i.e., at $t = T$ ); $\rho_0$ is the distribution of initial states. + +An important feature of diffusion models is that the sampling procedure has a wide design space after the score-matching training process, allowing a variance schedule ranging from larger to smaller. A popular uniform formulation of this sampling process is given in [37], where the generation process is defined as follows: + +$$ +\begin{array}{l} \mu_ {\theta} (\mathbf {x} _ {t - 1}) = \sqrt {\alpha_ {t - 1}} \left(\frac {\mathbf {x} _ {t} - \sqrt {1 - \alpha_ {t}} \epsilon_ {\theta} ^ {(t)} (\mathbf {x} _ {t})}{\sqrt {\alpha_ {t}}}\right) \\ + \sqrt {1 - \alpha_ {t - 1} - \eta^ {2} \frac {(1 - \alpha_ {t - 1})}{(1 - \alpha_ {t})} \left(1 - \frac {\alpha_ {t}}{\alpha_ {t - 1}}\right)} \epsilon_ {\theta} ^ {(t)} (\mathbf {x} _ {t}) \\ \end{array} +$$ + +$$ +\sigma_ {t - 1} = \eta \sqrt {\frac {\left(1 - \alpha_ {t - 1}\right)}{\left(1 - \alpha_ {t}\right)} \left(1 - \frac {\alpha_ {t}}{\alpha_ {t - 1}}\right)}, \tag {3} +$$ + +where $\mu_{\theta}(\mathbf{x}_{t - 1})$ and $\sigma_{t - 1}$ represent the mean and variance of the current denoising step, $\eta$ determines the variance of each generation step. The $\epsilon_{\theta}^{(t)}(\mathbf{x}_t)$ is the neural network for predicting noise. Note that most algorithms only model the generation process for the case where $\eta = 1$ , which restricts the policy search to a narrower space. + +# 3.2. Provable Exploration Entropy for Diffusion + +In reinforcement learning, the entropy of policy $H(\pi)$ represents the diversity of action choice on the current state. To increase the diversity with a more extensive exploration region, an entropy-related term is added to the original loss of RL fine-tuning. A common goal in fine-tuning the diffusion models is to maximize the expected reward of the generated images given the prompt distribution. Furthermore, the equation below holds true. + +$$ +\begin{array}{l} \max _ {\theta} \mathbb {E} _ {p (\mathbf {z})} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {0} | \mathbf {z})} [ r (\mathbf {x} _ {0}, \mathbf {z}) ] \\ = \max _ {\theta} \mathbb {E} _ {p (\mathbf {z})} \mathbb {E} _ {p _ {\theta} (\mathbf {x} _ {0: T} | \mathbf {z})} [ r (\mathbf {x} _ {0}, \mathbf {z}) ]. \\ \end{array} +$$ + +From [5], we know the following expression holds: + +$$ +\begin{array}{l} K L \left(p _ {\theta} \left(x _ {0: T} | z\right) | | p _ {p r e} \left(x _ {0: T} | z\right)\right) \\ = \sum_ {t = 1} ^ {T} \mathbb {E} _ {p _ {\theta} (x _ {t} | z)} [ K L (p _ {\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) ]. \\ \end{array} +$$ + +Similarly, we conclude that + +$$ +H \left(p _ {\theta} \left(x _ {0: T} | z\right)\right) = \sum_ {t = 1} ^ {T} \mathbb {E} _ {p _ {\theta} \left(x _ {t} \mid z\right)} \left[ H \left(p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, z\right)\right) \right], +$$ + +![](images/7339c0a9e3530cce54424893818b20253255f20627615c26720d313e79c50e39.jpg) +Figure 1. AdaEnt's framework. The entropy-guided exploration expands policy's search area at each denoising step, regularized by the KL divergence of the current policy and reference policy. The trade-off between the two effects is coordinated by a perception model that evaluates the real-time noised images at each step. The prediction of the adaptive module also indicates unnecessary denoising steps and dynamically stops the denoising process to reduce the computation cost in training. + +and + +$$ +\begin{array}{l} K L \left(p _ {\theta} \left(x _ {0: T} | z\right) \mid \mid p _ {\theta_ {k - 1}} \left(x _ {0: T} | z\right)\right) \\ = \sum_ {t = 1} ^ {T} \mathbb {E} _ {p _ {\theta} (x _ {t} | z)} [ K L (p _ {\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {\theta_ {k - 1}} (x _ {t - 1}, z)) ]. \\ \end{array} +$$ + +Finally, using the KL divergence and entropy, we propose the following objective for regularized training at the $t$ -th step: + +$$ +\begin{array}{l} \max _ {\theta} \mathbb {E} _ {p (z)} \mathbb {E} _ {p _ {\theta} (x _ {0: T} | z)} [ r (x _ {0}, z) + \sum_ {t = 1} ^ {T} \alpha_ {1} H (p _ {\theta} (x _ {t - 1} | x _ {t}, z)) \\ - \sum_ {t = 1} ^ {T} \alpha_ {2} K L (p _ {\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) \\ - \sum_ {t = 1} ^ {T} \alpha_ {3} K L \left(p _ {\theta} \left(x _ {t - 1} \mid x _ {t}, z\right) \mid p _ {\theta_ {k - 1}} \left(x _ {t - 1}, z\right)\right), \tag {4} \\ \end{array} +$$ + +where $\alpha_{1} \geq 0$ , $\alpha_{2}$ and $\alpha_{3} \geq 0$ . A significant concern about including entropy in the loss function of policy optimization is that it may lead to unstable policies. To address this issue, we theoretically analyze the new loss function and prove that policy optimization is convergent. + +Theorem 3.1. Given the objective of Eq. (4), the optimal policy $p_{\theta}^{*}(x_{t - 1}|x_{t},z)$ has the following expression: + +$$ +p _ {\theta} ^ {*} (x _ {t - 1} | x _ {t}, z) \propto \exp \{\frac {Q ^ {*} (x _ {t - 1} , z) + \alpha_ {2} \log p _ {p r e} (x _ {t - 1} | x _ {t} , z)}{\alpha_ {1} + \alpha_ {2}} \}, +$$ + +where $Q^{*}(x_{t - 1},z) = r(s_{0},z)I_{\{t - 1 = 0\}} + (\alpha_{1} + \alpha_{2})$ . + +$\log \sum_{x_{t - 2}}\exp \{\frac{Q^{*}(x_{t - 2},z) + \alpha_{2}\log p_{pre}(x_{t - 2}|x_{t - 1},z)}{\alpha_{1} + \alpha_{2}}\} .$ + +The proof of Theorem 3.1 is detailed in Appendix A.1. This theorem guarantees that with entropy added to training objective, the policy will ultimately converge. + +Theorem 3.2. When $\hat{\alpha}_1 \leq \alpha_1$ , we have $H(p_{\theta}^{*}(x_{t-1}|x_t,z);\hat{\alpha}_1,\alpha_2) \leq H(p_{\theta}^{*}(x_{t-1}|x_t,z);\alpha_1,\alpha_2)$ . Specially, compared to not including an entropy term, that is $\alpha_1 = 0$ , in the objective function of Eq. (4), adding an entropy term results in a higher entropy after the algorithm converges, that is $H(p_{\theta}^{*}(x_{t-1}|x_t,z);\alpha_1,\alpha_2) \geq H(p_{\theta}^{*}(x_{t-1}|x_t,z);0,\alpha_2)$ . + +The proof of Theorem 3.2 is detailed in Appendix A.2. This theorem demonstrates that adding an entropy term to the loss function increases the policy entropy, thereby enhancing the algorithm's exploratory behavior. Through greater exploration capability, the agent can discover better actions and achieve higher returns. + +Theorem 3.3. If $C = \min_{x_{t-1} \neq y_{t-1}} [\log p_{pre}(y_{t-1}|x_t, z) - \log p_{pre}(x_{t-1}|x_t, z)] / [Q^*(x_{t-1}, z) - Q^*(y_{t-1}, z)] > 0$ , the objective function of Eq. (4) contains only a KL term and no entropy term, that is $\alpha_1 = 0$ , if $1/C \leq \hat{\alpha}_2 \leq \alpha_2$ , we have $H(p_\theta^*(x_{t-1}|x_t, z); 0, \hat{\alpha}_2) \geq H(p_\theta^*(x_{t-1}|x_t, z); 0, \alpha_2)$ . + +The proof of Theorem 3.3 is detailed in Appendix A.3. This theorem indicates that adding a KL term to the loss function may reduce the exploration capability of the policy if the parameters are not properly tuned, thereby affecting the algorithm's performance. + +# 4. AdaEnt: A Practical Instantiation + +Building on the above theoretical foundation, we propose AdaEnt, a plug-and-play module that adaptatively controls entropy and a number of steps to combat reward hacking in diffusion policy optimization. The architecture of AdaEnt is illustrated in Fig. 1. + +# 4.1. Entropy-Driven Exploration + +To cope with the diversity degradation brought by reward hacking issue, we introduce the entropy of denoising policy + +$p_{\theta}(x_{t - 1}|x_t,z)$ to enhance exploration. In a normal distribution, the entropy is related to its variance. From Eq. (3), we define a new variance: + +$$ +\sigma_ {T - t} = \eta_ {\theta , \phi} (\mathbf {x} _ {T - t}, T - t) \sqrt {\frac {1 - \alpha_ {T - t - 1}}{(1 - \alpha_ {T - t})} \cdot \left(1 - \frac {\alpha_ {T - t}}{\alpha_ {T - t - 1}}\right)}, +$$ + +where $\eta_{\theta ,\phi}(\mathbf{x}_{T - t},T - t)$ is a dynamic, lightweight head with parameters $\phi$ , built upon the neural network $\mu_{\theta}$ ,designed to provide a latent and step-aware dynamic variance. + +# 4.2. Trade-Off between Exploration and Regularization + +In diffusion models, different denoising steps play various roles in image generation. Previous research has shown that earlier steps significantly impact the semantic content and global layout of the image, while later steps influence the texture and details [3]. Motivated by this phenomenon, we propose an adaptive regularization scheme. + +We introduce an adaptive balance factor $\gamma_{t}$ to dynamically adjust the balance between exploration diversity and denoising convergence across the denoising process. For the sake of simplicity, we introduce the PPO algorithm [34]. Therefore, the objective to optimize (Eq. (4)) is updated: + +$$ +\begin{array}{l} \mathcal {J} = \max _ {\theta} \mathbb {E} [ \min (\frac {p _ {\theta} (\cdot)}{p _ {\theta_ {t}} (\cdot)} A, c l i p (\frac {p _ {\theta} (\cdot)}{p _ {\theta_ {t}} (\cdot)}, 1 - \epsilon , 1 + \epsilon) A) \\ + \sum_ {t = 1} ^ {T} \left(1 - \gamma_ {t}\right) H \left(p _ {\theta} (\cdot)\right) - \sum_ {t = 1} ^ {T} \gamma_ {t} K L \left(p _ {\theta} (\cdot)\right) \left\| p _ {p r e} (\cdot)\right), \tag {5} \\ \end{array} +$$ + +where $A$ is an advantage function. According to Eq. (5), as the generation progresses, the regularization term will dominate the optimization and seek directions toward the distribution of real images within the high-reward region. + +A prediction model $f(\mathbf{x}_t)$ is trained on each intermediate denoised (latent) image $\mathbf{x}_t$ , with intermediate denoised images close to the final generated image $\mathbf{x}_0$ labeled as 1, and those close to initial noise $\mathbf{x}_T$ labeled as 0. The binary classification loss function is defined as follows: + +$$ +\mathcal {L} _ {p r e d} = - \sum_ {i} \left(y _ {i} \log \left(f \left(x _ {i}\right)\right) + \left(1 - y _ {i}\right) \log \left(1 - f \left(x _ {i}\right)\right)\right), \tag {6} +$$ + +where $y_{i} \in \{0,1\}$ is the label indicating whether the image is noiseless, and $f(x_{i})$ is the model's predicted probability that image $x_{i}$ is noiseless. After training, this network can evaluate the denoising level of any intermediate image during the denoising process, and we let $\gamma_{t} = f(\mathbf{x}_{t})$ as a measure of the denoising process indicating how well an intermediate image aligns with real image distribution. + +# 4.3. Dynamic Step Decision + +One limitation of diffusion models is the reliance on many iterative denoising steps. The drawback will be exacerbated + +when performing RL fine-tuning, leading to a high computation demand and exacerbation of reward hacking. As later denoising steps contribute less to the overall appearance of the generated image, applying an early stop on the denoising process will make training and inference more efficient and mediately alleviate reward hacking. The criterion of early stop is dynamically determined by how well the current image aligns with true image distribution. According to Sec. 4.2, $\gamma_{t} = 1$ means the current state of image satisfies the quality. Hence we apply early stop in this scenario, which is completely dynamic according to the complexity of generation and current image quality, without any predefined hyperparameters. + +Note that in existing diffusion step adaptation methods, the total number of denoising steps have to be determined before hand, which is impossible for the adaptive module to make a decision based on sufficient information about the image quality. In contrast, our adaptation mechanism monitors the generated image at each step and determines the total number of steps based on real-time quality of the current state. Recent study on denoising diffusion models reveals the temporal information may be not necessary [38], which supports our design of dynamically cutting off the iterative denoising process. The algorithmic procedure of AdaEnt can be found in Appendix B. + +# 5. Experiments + +# 5.1. Experimental Setup + +Datasets. To compare with previous work, we use the same prompt set consisting of 45 different animal classes for training (Simple-animals), which has been widely adopted for reward fine-tuning [1, 49, 51]. For testing, we generate 8 different noises for each animal class, totaling 360 test samples. For generalization evaluation, larger-scale prompt sets are further considered in testing as done in [1], including 398 animal classes from ImageNet (ImageNet-animals) and the above 45 animals each concatenated with an activity (Activity-animals). + +Rewards and Metrics. In this paper, we first deploy three metrics as the reward functions, that is, Aesthetic Score [33], Compressibility, and Incompressibility. To quantitatively evaluate performance in face of reward hacking, we then introduce cross-reward generalization and well-acknowledged metrics as suggested in [51]: ImageReward [46], Inception Score (IS) [32], PickScore [17], and ClipScore [27]. Details of each metric can be found in Appendix C. + +Implementation Details. Following [1], we take Stable Diffusion v1.4 (SDv14) as our main test bed for reward finetuning. We also consider Stable Diffusion v1.5 (SDv15), v2.1-turbo (SDv21), and XL1.0 (XL) to verify the robustness of our method with different backbones. All models + +![](images/7e2716bbf81b9c217f24a7fcebaec1166ef2d6ff4bf556ca9c34c34be0d5161c.jpg) +Figure 2. Reward curves of RL fine-tuning SDv14 with three different reward functions. Our proposed method consistently outperforms baselines in all three experimental scenarios, in terms of both peak achieved rewards and the sample efficiency. + +![](images/51d2b5b7fd95b548b53b906bb30906d7d916fff85b2c7171978946bab1220fb1.jpg) + +![](images/4b225500f3bae1a5e4e23b453d44263071fcab49de0b4345acf9ce77a9ac5f5d.jpg) + +are fine-tuned with low-rank adaptation (LoRA) [13] on 8 NVIDIA Tesla A100 GPUs. More training details can be found in Appendix C. + +# 5.2. Main Results + +We benchmark our methods against state-of-the-art RL-based diffusion fine-tuning frameworks, including DDPO [1], DPOK [5], D3PO [49], and TDPO [51]. As shown in Fig. 2, our method demonstrates consistent superiority in learning all three reward functions, validating its effectiveness for optimizing downstream objectives. Crucially, our balanced state exploration and adaptive regularization mechanism achieves promote the rate of reward growth by a large margin compared to DPOK's fixed-weight regularization strategy. Unlike dense-reward methods requiring evaluations of noised states' long-term impacts, our approach implements stepwise adaptive entropy to guide denoising space exploration, enabling sample-efficient optimization. Consequently, we attain rewards slightly higher than the latest dense-reward method TDPO but using much fewer training samples. + +Moreover, to systematically investigate reward hacking mitigation, we test the models optimized for Aesthetic Score with different aesthetic scores and general metrics for image quality and text-to-image alignment. We report the min, mean, and max values of each metric across the entire training process. As shown in Tab. 1, our generated images achieves overall higher scores on the diverse evaluation perspectives, effectively circumventing reward hacking phenomena even under elevated reward conditions. The results indicate that the proposed exploratory entropy effectively contributes to improved diversity and visual fidelity of generated images, along with less overfitting to the learning reward function. + +In summary, through this balanced exploration-regularization paradigm, our method demonstrates superior capability in discovering latent high-reward policies, thereby overcoming conventional performance plateaus + +![](images/8f8012150ab19ae67b1081b44089d27fff32128eeea6c6b6b6ba787b5d8ecc38.jpg) +Figure 3. Distribution visualizations of our intermediate variables. (a) Distribution of predicted $\gamma$ during training. The predictor progressively learns to effectively distinguish latent codes of different stages. (b) Distribution of denoising step number during training. Our adaptive step mechanism effectively reduces the denoising steps in training. + +![](images/d6eb55d24e75359497c7695246717d0f014863d5daf831675e6155064fdd1171.jpg) + +and achieving measurable enhancements in both generation quality and quantitative reward metrics. + +![](images/bda6cce7f839434189d3c9a6863c2c1372c14ae19b52441cb13816aa6fcca061.jpg) +Figure 4. Left: Entropy can be optimized in AdaEnt under different scaling factors. Right: Excessive application of either entropy or KL regularization impedes reward learning and may ultimately induce catastrophic collapse. + +![](images/de7bc767c5bc6589fb6cc7fe488d553f65ae6c0181d984855e5f61cc458f6322.jpg) + +# 5.3. Ablation Study + +Effect of $\gamma$ Predictor. The coefficient $\gamma$ is introduced to adaptively balance the exploratory entropy and KL regularization based on the current denoising situation. In Fig. 3 (a), we show the distributions of $\gamma$ with 50-step denoising during training. It can be observed that the predictor progressively learns to effectively differentiate latent codes that are more distinctive (i.e., those at the beginning and end of denoising), whereas the results in the intermediate steps are + +Table 1. Quality assessment from multiple unseen metric perspectives. All results are evaluated across training on the aesthetic score and sampled with the same training iteration. * refers to the Inception Score being sampled with the same aesthetic score of different methods. Since D3PO performs limited in learning rewards (see Fig. 2), the range of aesthetic score data is not enough for inception score collection. #Top1 indicates the instances when each method achieves the best performance across all metrics. + +
MethodPickScoreInceptionScore*ImageRewardClipScore#Top1
MinMeanMaxMinMeanMaxMinMeanMaxMinMeanMax
DDPO [1]21.2821.3721.5321.1122.9523.670.89491.07071.24930.29660.30190.30831
DPOK [5]21.2721.4021.5320.4822.2923.460.90601.04961.13400.30080.30460.30831
D3PO [49]21.2821.2921.31---0.86230.89620.93170.30000.30120.30221
TDPO [51]21.0621.3821.7119.9421.8823.950.00480.84531.13540.27540.30100.31132
AdaEnt21.2421.4121.6422.1124.3426.190.90781.12121.31230.30000.30500.30988
+ +![](images/60f9876634bb3d6e177ec10f86f563bfebf7294cea22e42ae7146848da21fc06.jpg) +Figure 5. Comparison of aesthetic reward curves and visual quality on SDv15 and SDv21. Our method consistently achieves higher rewards without over-optimizing across different backbones, while baseline methods compromise either visual quality or text-image alignment to attain aesthetic rewards. + +predicted with non-binary confidence. This adaptive behavior dynamically balances entropy and the KL divergence as the denoising process progresses, facilitating a shift from exploration to regularization during each denoising process. Balance between Entropy and KL Divergence. In Fig. 4, we conduct an in-depth investigation into the effects of introducing entropy in reward learning. Specifically, with a fixed KL, we design ablation variants with entropy scaled by different factors (i.e., $\mu_{e}$ ). The baselines, DDPO and DPOK, have a fixed schedule of entropy (according to Eq. (1)). In our proposed framework, it is observed that entropy can be optimized to increase under different scaling factors, and corresponding with a better learning of reward. However, when entropy's order of magnitude overwhelms KL divergence (i.e., $\mu_{e} = 1.0$ ), over-exploration may lead to the collapse of training. Therefore, the result demonstrates the necessity of introducing the trade-off between entropy and KL, which ensures a well-balanced training process for improved generating performance. + +Table 2. Generalization results of SDv15 and SDv21 on unseen metrics. The metrics are abbreviated and the same in Tab. 1. + +
MethodSDv15SDv21#Top1
PSISIRCSPSISIRCS
DDPO21.1716.900.820.3020.347.89-0.210.260
DPOK21.1818.680.840.3120.5310.280.040.272
D3PO21.0122.060.800.3318.843.14-1.580.252
TDPO19.682.40-1.570.2418.521.56-2.090.240
AdaEnt21.2118.990.850.3020.598.140.100.275
+ +![](images/d80e1b8ff98d2572201721ee303d4809174f4fafaad5ba9d4be4c76f190a9a2a.jpg) +Figure 6. Comparison of aesthetic reward curves and visual quality on SD-XL. Notably, we observe that our method is the only one that successfully works on SD-XL. + +![](images/f21f33e9edb90bf43b74dc454e1bb849107943946d6c077059b20f42550e77f4.jpg) + +# 5.4. Results on Different Backbones. + +In order to verify the generalizability of our method across different models for practical applications, we hereby extend our experiment to SDv15 and SDv21. In Fig. 5, our method also achieves superior effectiveness in reward learning. For generalization evaluation, we also evaluate the generated image quality with more metric perspectives for SDv15 and SDv21. As shown in Tab. 2, AdaEnt achieves superior overall quality, and despite TDPO shows comparable aesthetic score with our method, it is demonstrated to be failed in prompt-image alignments. + +Moreover, considering that SD-XL is one of the most advanced and extensive pre-trained model, we provide a concise reward-learning experiment on SD-XL in Fig. 6. The results demonstrate that AdaEnt can still bring reward gains + +![](images/0a76a196830567cde79441aba76233adfd078f17a133e4f5465dde1d1d86fb51.jpg) +Figure 7. Comparison of generalization ability to larger scale prompt sets with unseen prompts. The experiments are conducted on SDv14 and trained with the small Simple-animals set. + +![](images/9a31224dd9cb1057704a0cbf0235c6f8cd05aeb9bb6f4647ef0a9368e1645e1e.jpg) +Figure 8. Comparison of generalization ability to more complex prompts with multiple categories. All methods are implemented in SD-XL. The performance of our method is highlighted in darker colors, demonstrating AdaEnt's superior performance in both image quality and prompt-image alignment. + +for larger and more powerful models. + +# 5.5. Generalization to Unseen Prompts. + +To investigate whether the performance of RL fine-tuned models is limited to the prompts seen in training, which is an ordinary case of reward hacking, we hereby conduct prompts generalization experiments from two perspectives. + +Generalization to larger-scale prompt sets. To evaluate the generalization ability to unseen prompts, we evaluate the aesthetic score on ImageNet-animals and Activity-animals. The experiments are conducted on SDv14, using prompts from Simple-animals as the training set. As presented in Fig. 7, our method achieves promising aesthetic scores on the unseen datasets, which demonstrates the effectiveness of our method. + +Generalization to more complex prompts. Additionally, we assess the performance on more complex prompts. To be specific, we further conduct experiments on SD-XL with a subset of the HPSv2 dataset [44], which contains complex prompts that are not limited to animals. We identify several categories in the prompt set and measure the prompt-image consistency using clip-score. We also provide the aesthetic score to assess the quality of generated images. As presented in Fig. 8, our method exhibits the best perfor + +mance in the aesthetic score and the maintenance of image-prompts alignment, which suggests our superiority. + +# 5.6. Computation Efficiency of AdaEnt + +Since the adaptive mechanism introduced in our work allows for the early termination of the denoising process once denoising is essentially complete, it is expected to improve the training efficiency of diffusion models. Firstly, we provide the visualized distributions of denoising step number during training in Fig. 3 (b), which verifies the capability of adaptively reducing the step number. Quantitatively, we take the aesthetic score $= 7$ as the training terminals of each method for evaluation. In Tab. 3, we provide GFLOPs and time for single batch (size=12) sampling, along with the total sampling step and time for the overall training process. The results show that the proposed AdaEnt can significantly accelerate the training process with almost no increase in computational cost (i.e., GFLOPS). + +Table 3. Computation Efficiency of integrating our proposed dynamic early stop mechanism with existing RL frameworks. TDS refers to total denoising steps. Results of both overall training and single-batch inference are reported. + +
MethodOverallSingle batch
TDSTimeGFLOPSTime
DDPO1.843+e611.98h1.034e+069.56s
DPOK2.015+e614.76h9.69s
DDPO+Ours0.242+e65.52h1.044e+066.01s
DPOK+Ours0.294+e67.16h7.12s
+ +# 6. Conclusion + +This work addresses the critical challenge of reward hacking in diffusion model optimization through reinforcement learning, where conventional RL frameworks tend to produce overfitted outputs that compromise distributional fidelity and generation diversity. Our theoretical analyses establish entropy as a proxy to enhance the exploratory capacity of denoising policies. The theoretical basis is instantiated as AdaEnt, a framework that achieves adaptive entropy control and step-wise optimization through three interconnected mechanisms: entropy-aware policy updates, distributional consistency regularization, and noise-level dependent exploration scheduling. Extensive experiments demonstrate that our method effectively mitigates reward hacking while preserving the diffusion model's inherent generation capabilities, establishing new state-of-the-art performance in quality-diversity trade-off across multiple diffusion-based RL benchmarks. The adaptive entropy proposed in this paper offers new perspectives for future RL fine-tuning of diffusion models. + +# 7. Acknowledgment + +# References + +[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 2, 5, 6, 7 +[2] Bo Dai, Albert E. Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, and Le Song. SBEED: convergent reinforcement learning with nonlinear function approximation. In Proceedings of International Conference on Machine Learning., 2018. 12 +[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5 +[4] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2024. 2 +[5] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 6, 7 +[6] Yaozhong Gan, Renye Yan, Xiaoyang Tan, Zhe Wu, and Jun-liang Xing. Transductive off-policy proximal policy optimization. arXiv preprint arXiv:2406.03894, 2024. 2 +[7] Yaozhong Gan, Renye Yan, Zhe Wu, and Junliang Xing. Reflective policy optimization. arXiv preprint arXiv:2406.03678, 2024. 2 +[8] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835-10866. PMLR, 2023. 2 +[9] Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International conference on machine learning, pages 1352-1361. PMLR, 2017. 2 +[10] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, pages 1861-1870. Pmlr, 2018. 2 +[11] Seungyul Han and Youngchul Sung. A max-min entropy framework for reinforcement learning. Advances in Neural Information Processing Systems, 34:25732-25745, 2021. 2 + +This paper is supported by the National Science Foundation of China under Grant No.62025401, 62341407, U24B200170, 62406008; Beijing Municipal Science and Technology Program under Grant Z241100004224015; Beijing Natural Science Foundation under Grant No.L223004; '111' Project under Grant No.B18001. This work is also supported in part by the Natural Science Foundation of China under Grant No. 62222606. +[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3 +[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6 +[14] Rohit Jena, Ali Taghibakhshi, Sahil Jain, Gerald Shen, Nima Tajbakhsh, and Arash Vahdat. Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models. arXiv preprint arXiv:2409.06493, 2024. 1 +[15] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2 +[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 2 +[17] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15 +[18] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. CoRR, 2023. 2 +[19] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 2 +[20] Jincheng Mei, Chenjun Xiao, Ruitong Huang, Dale Schuurmans, and Martin Müller. On principled entropy exploration in policy optimization. In Proceedings of International Joint Conference on Artificial Intelligence, 2019. 12 +[21] Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, and Dacheng Tao. Inform: Mitigating reward hacking in rlhf via information-theoretic reward modeling. Advances in Neural Information Processing Systems, 37: 134387-134429, 2025. 2 +[22] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016. 2 +[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1 +[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2 + +[25] Owen Oertell, Shikun Sun, Yiding Chen, Jin Peng Zhou, Zhiyong Wang, and Wen Sun. Efficient controllable diffusion via optimal classifier guidance. arXiv preprint arXiv:2505.21666, 2025. 2 +[26] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2 +[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR. 5, 15 +[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2 +[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. +[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1 +[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5, 15 +[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 2022. 5 +[34] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5 +[35] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming. Advances in Neural Information Processing Systems, 35:9460-9471, 2022. 2 +[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3 +[37] Jiaming Song, Chenlin Meng, and Stefano Ermon. + +Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3 +[38] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv preprint arXiv:2502.13129, 2025. 5 +[39] Shikun Sun, Chengrui Wang, Min Zhou, Zixuan Wang, Xiaoyu Qin, Tiezheng Ge, Bo Zheng, and Jia Jia. Depo: Enhancing e-commerce image background generation with short trajectory direct expected preference optimization. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2 +[40] Marc Toussaint. Robot trajectory optimization using approximate inference. In Proceedings of the 26th annual international conference on machine learning, pages 1049-1056, 2009. 2 +[41] Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Finetuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024. 2 +[42] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2 +[43] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 2 +[44] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023.8 +[45] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2 +[46] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15 +[47] Renye Yan, Zhe Wu, Yuan Zhan, Pin Tao, Zongwei Wang, Yimao Cai, and Junliang Xing. Mnemonic dictionary learning for intrinsic motivation in reinforcement learning. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2023. 2 +[48] Renye Yan, Yaozhong Gan, You Wu, Ling Liang, Junliang Xing, Yimao Cai, and Ru Huang. The exploration-exploitation dilemma revisited: An entropy perspective. arXiv preprint arXiv:2408.09974, 2024. 2 + +[49] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 1, 2, 5, 6, 7 +[50] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 3 +[51] Ziyi Zhang, Sen Zhang, Yibing Zhan, Yong Luo, Yonggang Wen, and Dacheng Tao. Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. In International Conference on Machine Learning, pages 60396-60413. PMLR, 2024. 1, 2, 5, 6, 7 \ No newline at end of file diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/images.zip b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..437f1570d152cc14a58260dbc9df5084ee44e69a --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c14491b0991d5136ef7436bc3cb268842bf39cf9cb8d96c5601235323f3ee00a +size 628262 diff --git a/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/layout.json b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..bf9dafb13529ba6c413d15bb1a9d45db72aa2e6c --- /dev/null +++ b/2025/Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment/layout.json @@ -0,0 +1,8840 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "spans": [ + { + "bbox": [ + 61, + 103, + 550, + 121 + ], + "type": "text", + "content": "Entropy-Adaptive Diffusion Policy Optimization with Dynamic Step Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "spans": [ + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": "Renye Yan" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Jikang Cheng" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Yaozhong Gan" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Shikun Sun" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", You Wu" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Yunfan Yang" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Ling Liang" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Jinlong Lin" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Yeshuang Zhu" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Jie Zhou" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Jinchao Zhang" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Junliang Xing" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Yimao Cai" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "text", + "content": ", Ru Huang" + }, + { + "bbox": [ + 65, + 142, + 543, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "spans": [ + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "content": "Peking University" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "content": ", Tencent Inc" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "content": ", Tsinghua University" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "content": ", Nanjing University" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 164, + 173, + 446, + 186 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 64, + 188, + 541, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 188, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 64, + 188, + 541, + 200 + ], + "type": "text", + "content": "victory@stu.pku.edu.cn, jikangcheng99@gmail.com, yshzhu@tencent.com, caiyimao@pu.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 254, + 297, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 254, + 297, + 553 + ], + "spans": [ + { + "bbox": [ + 54, + 254, + 297, + 553 + ], + "type": "text", + "content": "While fine-tuning diffusion models with reinforcement learning (RL) has demonstrated effectiveness in directly optimizing downstream objectives, existing RL frameworks are prone to overfitting the rewards, leading to outputs that deviate from the true data distribution and exhibit reduced diversity. To address this issue, we introduce entropy as a quantitative measure to enhance the exploratory capacity of diffusion models' denoising policies. We propose an adaptive mechanism that dynamically adjusts the application and magnitude of entropy and regularization, guided by real-time quality estimation of intermediate noised states. Theoretically, we prove the convergence of our entropy-enhanced policy optimization and establish two critical properties: 1) global entropy increases through training, ensuring robust exploration capabilities, and 2) entropy systematically decreases during the denoising process, enabling a phase transition from early-stage diversity promotion to late-stage distributional fidelity. Building on this foundation, we propose a plug-and-play RL module that adaptively controls entropy and optimizes denoising steps. Extensive evaluations demonstrate theoretical soundness and empirical robustness of our method, achieving state-of-the-art quality-diversity trade-offs across benchmarks. Notably, our framework significantly improves the rewards and reduces denoising steps in training by up to " + }, + { + "bbox": [ + 54, + 254, + 297, + 553 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 54, + 254, + 297, + 553 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 135, + 590 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 598, + 296, + 684 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 684 + ], + "type": "text", + "content": "In recent years, diffusion models [12, 23, 29-31, 36] have demonstrated powerful capabilities in text-to-image generation. However, these models are designed to match the distribution of training data, lacking the ability to directly optimize downstream objectives, such as aesthetics and preference scores. To this end, reinforcement learning (RL) has been explored and shown promising potential in fine" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 312 + ], + "type": "text", + "content": "tuning diffusion models with various reward functions or human feedback (i.e., RLAIF and RLHF) [1, 5, 49]. In these methods, policy optimization is guided by a global reward computed on the final generated image, which is either differentiable or non-differentiable, making RL a flexible and promising solution to overcoming the limitations of supervised fine-tuning for diffusion models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 314, + 554, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 554, + 458 + ], + "type": "text", + "content": "However, current RL methods rely solely on rewards from the final generated image to update policies, overlooking the importance of exploration in promoting diversity during denoising. This makes fine-tuned models susceptible to reward hacking [14], where generated images achieve high scores in reward models but significantly deviates from the true distribution, exhibiting deteriorated quality and low diversity. This phenomenon has been partially attributed to sparse and delayed rewards and addressed with early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51]. However, the problem remains to be effectively solved." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 554, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 652 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 652 + ], + "type": "text", + "content": "This paper proposes a novel RL method with adaptive entropy regulation to address the reward hacking issue in fine-tuning diffusion models. We formally introduce entropy as a quantitative indicator to assess the exploratory capability of denoising strategies in diffusion models, integrating it into the RL optimization objectives. Theoretical analysis establishes the convergence guarantees of the proposed method for newly derived policies. Furthermore, to overcome the limitations of fixed parameterization in conventional KL regulation, we derive an adaptive mechanism that dynamically balances exploration diversity and convergence stability throughout the denoising process. Such adaptive regulation enables progressive entropy adjustment, facilitating a smooth transition from initial exploration diversity to stable convergence toward the target distribution in later denoising stages." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 715 + ], + "type": "text", + "content": "Building upon this theoretical framework, we formalize AdaEnt as a plug-and-play module compatible with existing RL fine-tuning architectures for diffusion models. AdaEnt addresses two critical challenges in current RL frameworks: 1) sparse reward signals caused by insuffi" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 135, + 703 + ], + "type": "text", + "content": "*Equal Contribution" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 145, + 712 + ], + "type": "text", + "content": "† Corresponding Author" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1924" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": "cient optimization guidance during intermediate denoising steps and 2) diversity loss resulting from KL constraints. By incorporating entropy exploration terms at each denoising step and leveraging the " + }, + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 72, + 294, + 262 + ], + "type": "text", + "content": " parameter flexibility in DDIM, our method regulates entropy variations for principled exploration of denoising polices. The adaptive mechanism achieves dynamic balance between exploration and convergence through real-time quality assessment of intermediate outputs. Specifically, we introduce an image perception model trained alongside with the policy model to predict step-wise weights between diversity exploration and convergence stability in the optimization objective. Notably, the adaptive factors also serve as indicators of generation quality, which can be naturally used to reduce the total number of denoising steps, enabling efficient computation and lightweight storage." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 67, + 264, + 257, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 264, + 257, + 275 + ], + "spans": [ + { + "bbox": [ + 67, + 264, + 257, + 275 + ], + "type": "text", + "content": "This paper highlights three main contributions:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 279, + 295, + 422 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 56, + 279, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 279, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 56, + 279, + 295, + 313 + ], + "type": "text", + "content": "- We theoretically derive an entropy-based exploration strategy and present an effective implementation for diffusion models." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 315, + 295, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 295, + 373 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 295, + 373 + ], + "type": "text", + "content": "- We propose an adaptive mechanism to balance the entropy-guided exploration and the regularization of real data distribution, which captures different characteristics in the early and late denoising process with a theoretical guarantee." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 374, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 295, + 422 + ], + "type": "text", + "content": "- We propose a dynamic step alignment technique based on the entropy-guided exploration to automatically determine when to stop denoising, making our method computationally efficient for both training and inference." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 544 + ], + "type": "text", + "content": "Experiments show that AdaEnt effectively mitigates reward hacking and achieves the highest rewards while preserving the diversity of images. The results generalize to various pre-trained diffusion models, reward functions, and RL frameworks. Furthermore, through visualization analysis, we confirm the useful design of entropy and self-adaptive mechanism, which seeks diversity at the early stage and gradually switches to maintain faithfulness in the late stage. AdaEnt works as a plug-and-play module that is broadly applicable in practice." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 556, + 142, + 568 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 142, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 142, + 568 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "spans": [ + { + "bbox": [ + 55, + 576, + 192, + 588 + ], + "type": "text", + "content": "2.1. RL for Diffusion Models" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 713 + ], + "type": "text", + "content": "Conventional training of diffusion models relies on likelihood-based objectives [12], yet these often misalign with perceptual quality metrics, resulting in degraded visual outputs [1, 16, 24]. Reinforcement learning (RL) [34] addresses this limitation by reframing the diffusion denoising process as a Markov Decision Process (MDP), enabling direct optimization of task-specific rewards such as aesthetic quality, compression efficiency, or human preference alignment. Building on pre-trained text-to-image models, RL demonstrates an effective and flexible framework to fur" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "ther improve the diffusion model, where recent successful techniques for large language models, such as PPO [34] and DPO [28], can be applied to improve training stability and efficiency. Notable contributions include reward model-driven approaches like DDPO [1], SLCD [25] and DPOK [5], as well as human preference-aligned methods such as D3PO [49], Diffusion-DPO [42], DEPO [39], and HPS [45]. This paper follows this line of work, aiming to improve reward-driven performance while systematically mitigating the inherited drawbacks of over-optimization brought by RL." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 212, + 511, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 511, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 511, + 224 + ], + "type": "text", + "content": "2.2. Reward Hacking in Diffusion Models" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 230, + 555, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 458 + ], + "type": "text", + "content": "Reward hacking describes the phenomenon that the model maximizes the reward score but produces results that lack diversity or authenticity [8, 21, 35, 51]. For example, when the model overly prioritizes specific reward signals, such as image aesthetics or text alignment, it can lead to structural distortions, loss of diversity, or artifacts in generated images, ultimately compromising generation quality. This phenomenon has been partially attributed to sparse and delayed rewards, which affects reward optimization in both RL and direct back-propagation methods [4, 26, 41]. Methods that rely on direct access to the gradient of differentiable reward model pose higher risks of reward hacking [8]. While techniques like early stopping [1], Kullback-Leibler (KL)-regularization [5], and dense temporal rewards [51] have been applied in RL methods to address this issue, the problem remains to be effectively solved [6, 7]. In contrast, this paper focuses on mitigating reward hacking in RL by adapting the policy optimization process and removing unnecessary steps to reduce computation costs." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 466, + 424, + 479 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 424, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 424, + 479 + ], + "type": "text", + "content": "2.3. Entropy-based RL" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 639 + ], + "type": "text", + "content": "Entropy-based reinforcement learning (RL) integrates information-theoretic concepts into policy optimization to encourage exploration. By leveraging entropy regularization [9, 22, 40, 48], these methods encourage stochasticity in action selection and enhance learning stability and adaptability. However, aimlessly increase entropy for exploration will ultimately hurt the policy [11, 47]. For a better trade-off between exploration and exploitation, SAC [10] automates entropy coefficient tuning by jointly maximizing expected returns and entropy. To our knowledge, this paper is the first to leverage entropy as a theoretically guaranteed exploration metric for optimizing diffusion models and adaptively balances the effect of entropy and distribution regularization." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 647, + 553, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 553, + 661 + ], + "type": "text", + "content": "2.4. Dynamic-Step Adaptation for Diffusion Models" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Existing methodologies for acceleration can be divided into reducing the number of steps [15] or computation cost per step [18, 19, 43]. Still, they often use a uniform approach without accounting for each generated image's quality at" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1925" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "each denoising step. Complementary to the above methods, recent work [50] applies different denoising steps for each prompt by training a step prediction policy network. This paper proposes to dynamically early stop the denoising process to mitigate reward hacking, with an additional benefit of accelerating the RL training of diffusion models, which need not a dedicated model for step number determination." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 111, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 111, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 111, + 178 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 186, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 186, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 296, + 331 + ], + "type": "text", + "content": "This paper introduces the adaptive exploration of diffusion policy optimization through the lens of entropy. In this section, we first prove the convergence of policy optimization with our new entropy-enhanced training objective in diffusion policy optimization. Coping with the trade-off between entropy-based exploration of RL and the inherent distribution modeling of the diffusion model, we then propose an adaptive mechanism to automatically balance the trade-off between exploration and the original objective, which is theoretically guaranteed to adjust entropy to allow a smooth transition from early-stage diversity exploration to late-stage data distribution maintenance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 264, + 349 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 264, + 349 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 264, + 349 + ], + "type": "text", + "content": "3.1. Diffusion as Markov Decision Processes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": "In diffusion models [12, 36], data distribution " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "p(\\mathbf{x}_0)" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " is modeled through a Markov forward process " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "q(\\mathbf{x}_t|\\mathbf{x}_{t-1})" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " that gradually adds noise to the data and a reverse process " + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "inline_equation", + "content": "p_\\theta(\\mathbf{x}_{t-1}|\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 354, + 296, + 403 + ], + "type": "text", + "content": " for generation, defined as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 412, + 296, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 412, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 86, + 412, + 296, + 441 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q (\\mathbf {x} _ {t} | \\mathbf {x} _ {t - 1}) = \\mathcal {N} (\\mathbf {x} _ {t}; \\sqrt {\\alpha_ {t}} \\mathbf {x} _ {t - 1}, \\sqrt {1 - \\alpha_ {t}} \\mathbf {I}), \\\\ p _ {\\theta} \\left(\\mathbf {x} _ {t - 1} \\mid \\mathbf {x} _ {t}\\right) = \\mathcal {N} \\left(\\mathbf {x} _ {t - 1}; \\mu_ {\\theta} \\left(\\mathbf {x} _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathbf {I}\\right), \\tag {1} \\\\ \\end{array}", + "image_path": "a7d9f8000b11682dd74a87e7726aa0053c96ecd2d7fcedb906ead95b7bcca46f.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\sqrt{1 - \\alpha_t}" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " denotes the noise scale at timestep " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{x}_t,t)" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " is the predicted mean by the model, and " + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "inline_equation", + "content": "\\sigma_t^2" + }, + { + "bbox": [ + 55, + 450, + 296, + 486 + ], + "type": "text", + "content": " is a time-dependent variance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 487, + 296, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 296, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 296, + 522 + ], + "type": "text", + "content": "Current RL fine-tuning methods treat the generation process of diffusion models as a Markov Decision Process (MDP), represented by the five-tuple " + }, + { + "bbox": [ + 55, + 487, + 296, + 522 + ], + "type": "inline_equation", + "content": "\\langle S, \\mathcal{A}, P, \\rho_0, R \\rangle" + }, + { + "bbox": [ + 55, + 487, + 296, + 522 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 89, + 532, + 294, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 532, + 294, + 647 + ], + "spans": [ + { + "bbox": [ + 89, + 532, + 294, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {s} _ {t} \\triangleq (\\mathbf {z}, t, \\mathbf {x} _ {T - t}) \\in S, \\\\ \\pi_ {\\theta} \\left(\\mathbf {a} _ {t} \\mid \\mathbf {s} _ {t}\\right) \\triangleq p _ {\\theta} \\left(\\mathbf {x} _ {T - t - 1} \\mid \\mathbf {x} _ {T - t}, \\mathbf {z}\\right), \\\\ P \\left(\\mathbf {s} _ {t + 1} \\mid \\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left(\\delta_ {\\mathbf {z}}, \\delta_ {T - t - 1}, \\delta_ {\\mathbf {x} _ {T - t - 1}}\\right), \\\\ \\mathbf {a} _ {t} \\triangleq \\mathbf {x} _ {T - t - 1}, \\tag {2} \\\\ \\rho_ {0} \\left(\\mathbf {s} _ {0}\\right) \\triangleq (p (\\mathbf {z}), \\delta_ {T}, \\mathcal {N} (\\mathbf {0}, \\mathbf {I})) \\\\ R \\left(\\mathbf {s} _ {t}, \\mathbf {a} _ {t}\\right) \\triangleq \\left\\{ \\begin{array}{l l} r \\left(\\mathbf {x} _ {0}, \\mathbf {z}\\right) & \\text {i f} t = T, \\\\ 0 & \\text {o t h e r w i s e}. \\end{array} \\right. \\\\ \\end{array}", + "image_path": "1391bcff2cf6d19269b7dbad8ac313dee30cb3321aa95fdf6326fec0a8fa522f.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{S}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " denotes the state space, containing all possible image states " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " in the diffusion model's generation process; " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " represents the action space, corresponding to the denoising operations " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "P(s_{t + 1}|s_t,a_t)" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " is the state transition probability, defining the probability of transitioning from state " + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "inline_equation", + "content": "s_t" + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": " to" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "s_{t + 1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " after taking action " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "a_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "R(s_t, a_t)" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " represents the reward function, which typically evaluates the final image quality upon completion (i.e., at " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "t = T" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": "); " + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "inline_equation", + "content": "\\rho_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 119 + ], + "type": "text", + "content": " is the distribution of initial states." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 555, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 194 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 194 + ], + "type": "text", + "content": "An important feature of diffusion models is that the sampling procedure has a wide design space after the score-matching training process, allowing a variance schedule ranging from larger to smaller. A popular uniform formulation of this sampling process is given in [37], where the generation process is defined as follows:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 205, + 544, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 205, + 544, + 271 + ], + "spans": [ + { + "bbox": [ + 326, + 205, + 544, + 271 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mu_ {\\theta} (\\mathbf {x} _ {t - 1}) = \\sqrt {\\alpha_ {t - 1}} \\left(\\frac {\\mathbf {x} _ {t} - \\sqrt {1 - \\alpha_ {t}} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t})}{\\sqrt {\\alpha_ {t}}}\\right) \\\\ + \\sqrt {1 - \\alpha_ {t - 1} - \\eta^ {2} \\frac {(1 - \\alpha_ {t - 1})}{(1 - \\alpha_ {t})} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)} \\epsilon_ {\\theta} ^ {(t)} (\\mathbf {x} _ {t}) \\\\ \\end{array}", + "image_path": "17def67e5da575c0ccb46e7a09ae217f81783db16dea24656498ae9d34c0f0ce.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 274, + 553, + 316 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 274, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 325, + 274, + 553, + 316 + ], + "type": "interline_equation", + "content": "\\sigma_ {t - 1} = \\eta \\sqrt {\\frac {\\left(1 - \\alpha_ {t - 1}\\right)}{\\left(1 - \\alpha_ {t}\\right)} \\left(1 - \\frac {\\alpha_ {t}}{\\alpha_ {t - 1}}\\right)}, \\tag {3}", + "image_path": "41a0a78238e62f2bd7e0b246aad09aad1f96b40f9f4f000b3efaee736f91c6f6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}(\\mathbf{x}_{t - 1})" + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\sigma_{t - 1}" + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": " represent the mean and variance of the current denoising step, " + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": " determines the variance of each generation step. The " + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}^{(t)}(\\mathbf{x}_t)" + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": " is the neural network for predicting noise. Note that most algorithms only model the generation process for the case where " + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "inline_equation", + "content": "\\eta = 1" + }, + { + "bbox": [ + 313, + 316, + 555, + 389 + ], + "type": "text", + "content": ", which restricts the policy search to a narrower space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 401, + 541, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 541, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 541, + 414 + ], + "type": "text", + "content": "3.2. Provable Exploration Entropy for Diffusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 420, + 554, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 554, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 554, + 515 + ], + "type": "text", + "content": "In reinforcement learning, the entropy of policy " + }, + { + "bbox": [ + 313, + 420, + 554, + 515 + ], + "type": "inline_equation", + "content": "H(\\pi)" + }, + { + "bbox": [ + 313, + 420, + 554, + 515 + ], + "type": "text", + "content": " represents the diversity of action choice on the current state. To increase the diversity with a more extensive exploration region, an entropy-related term is added to the original loss of RL fine-tuning. A common goal in fine-tuning the diffusion models is to maximize the expected reward of the generated images given the prompt distribution. Furthermore, the equation below holds true." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 367, + 529, + 503, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 529, + 503, + 564 + ], + "spans": [ + { + "bbox": [ + 367, + 529, + 503, + 564 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ] \\\\ = \\max _ {\\theta} \\mathbb {E} _ {p (\\mathbf {z})} \\mathbb {E} _ {p _ {\\theta} (\\mathbf {x} _ {0: T} | \\mathbf {z})} [ r (\\mathbf {x} _ {0}, \\mathbf {z}) ]. \\\\ \\end{array}", + "image_path": "d76a30ba4bfff68d6bae682b210616eb42e34257c66c6c493cac110ad2cfc768.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 576, + 522, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 522, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 522, + 588 + ], + "type": "text", + "content": "From [5], we know the following expression holds:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 329, + 601, + 540, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 601, + 540, + 647 + ], + "spans": [ + { + "bbox": [ + 329, + 601, + 540, + 647 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) | | p _ {p r e} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) ]. \\\\ \\end{array}", + "image_path": "eebf88a5ba43eac7f4197ae760e80e2b6c207c6f9d0574a9a3bd3b0a4871bf69.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 314, + 659, + 427, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 427, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 427, + 671 + ], + "type": "text", + "content": "Similarly, we conclude that" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 328, + 683, + 539, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 683, + 539, + 716 + ], + "spans": [ + { + "bbox": [ + 328, + 683, + 539, + 716 + ], + "type": "interline_equation", + "content": "H \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right)\\right) = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} \\left(x _ {t} \\mid z\\right)} \\left[ H \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right)\\right) \\right],", + "image_path": "445abf37b48c3b62787cfdc8a5b8990db5764affa205acb0ec570af47386c124.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1926" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 547, + 224 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 547, + 224 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 547, + 224 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 547, + 224 + ], + "type": "image", + "image_path": "7339c0a9e3530cce54424893818b20253255f20627615c26720d313e79c50e39.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 555, + 281 + ], + "type": "text", + "content": "Figure 1. AdaEnt's framework. The entropy-guided exploration expands policy's search area at each denoising step, regularized by the KL divergence of the current policy and reference policy. The trade-off between the two effects is coordinated by a perception model that evaluates the real-time noised images at each step. The prediction of the adaptive module also indicates unnecessary denoising steps and dynamically stops the denoising process to reduce the computation cost in training." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 301, + 74, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 74, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 74, + 312 + ], + "type": "text", + "content": "and" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 319, + 284, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 319, + 284, + 365 + ], + "spans": [ + { + "bbox": [ + 67, + 319, + 284, + 365 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} K L \\left(p _ {\\theta} \\left(x _ {0: T} | z\\right) \\mid \\mid p _ {\\theta_ {k - 1}} \\left(x _ {0: T} | z\\right)\\right) \\\\ = \\sum_ {t = 1} ^ {T} \\mathbb {E} _ {p _ {\\theta} (x _ {t} | z)} [ K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {\\theta_ {k - 1}} (x _ {t - 1}, z)) ]. \\\\ \\end{array}", + "image_path": "ca96f5cd2dc559388b41a583a8f934e732d340a6ded0e7534ba76ac22d34d853.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "text", + "content": "Finally, using the KL divergence and entropy, we propose the following objective for regularized training at the " + }, + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 370, + 295, + 407 + ], + "type": "text", + "content": "-th step:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 411, + 294, + 522 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 411, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 67, + 411, + 294, + 522 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max _ {\\theta} \\mathbb {E} _ {p (z)} \\mathbb {E} _ {p _ {\\theta} (x _ {0: T} | z)} [ r (x _ {0}, z) + \\sum_ {t = 1} ^ {T} \\alpha_ {1} H (p _ {\\theta} (x _ {t - 1} | x _ {t}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {2} K L (p _ {\\theta} (x _ {t - 1} | x _ {t}, z) | | p _ {p r e} (x _ {t - 1}, z)) \\\\ - \\sum_ {t = 1} ^ {T} \\alpha_ {3} K L \\left(p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}, z\\right) \\mid p _ {\\theta_ {k - 1}} \\left(x _ {t - 1}, z\\right)\\right), \\tag {4} \\\\ \\end{array}", + "image_path": "6d41db42f044ea627a78f80cf43a0a773cc21411ad0890c83b6a4bb4879353b4.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\alpha_{1} \\geq 0" + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\alpha_{2}" + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "inline_equation", + "content": "\\alpha_{3} \\geq 0" + }, + { + "bbox": [ + 55, + 523, + 296, + 583 + ], + "type": "text", + "content": ". A significant concern about including entropy in the loss function of policy optimization is that it may lead to unstable policies. To address this issue, we theoretically analyze the new loss function and prove that policy optimization is convergent." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 587, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 587, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 587, + 296, + 613 + ], + "type": "text", + "content": "Theorem 3.1. Given the objective of Eq. (4), the optimal policy " + }, + { + "bbox": [ + 55, + 587, + 296, + 613 + ], + "type": "inline_equation", + "content": "p_{\\theta}^{*}(x_{t - 1}|x_{t},z)" + }, + { + "bbox": [ + 55, + 587, + 296, + 613 + ], + "type": "text", + "content": " has the following expression:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 616, + 287, + 643 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 616, + 287, + 643 + ], + "spans": [ + { + "bbox": [ + 62, + 616, + 287, + 643 + ], + "type": "interline_equation", + "content": "p _ {\\theta} ^ {*} (x _ {t - 1} | x _ {t}, z) \\propto \\exp \\{\\frac {Q ^ {*} (x _ {t - 1} , z) + \\alpha_ {2} \\log p _ {p r e} (x _ {t - 1} | x _ {t} , z)}{\\alpha_ {1} + \\alpha_ {2}} \\},", + "image_path": "ad31e2f1956aef7de72968e8b46c8a3f46e69f28013ae9458b7c7718af276780.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 646, + 294, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 294, + 658 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 646, + 294, + 658 + ], + "type": "inline_equation", + "content": "Q^{*}(x_{t - 1},z) = r(s_{0},z)I_{\\{t - 1 = 0\\}} + (\\alpha_{1} + \\alpha_{2})" + }, + { + "bbox": [ + 55, + 646, + 294, + 658 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 659, + 263, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 659, + 263, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 659, + 263, + 676 + ], + "type": "inline_equation", + "content": "\\log \\sum_{x_{t - 2}}\\exp \\{\\frac{Q^{*}(x_{t - 2},z) + \\alpha_{2}\\log p_{pre}(x_{t - 2}|x_{t - 1},z)}{\\alpha_{1} + \\alpha_{2}}\\} ." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "The proof of Theorem 3.1 is detailed in Appendix A.1. This theorem guarantees that with entropy added to training objective, the policy will ultimately converge." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "content": "Theorem 3.2. When " + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\hat{\\alpha}_1 \\leq \\alpha_1" + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "inline_equation", + "content": "H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\hat{\\alpha}_1,\\alpha_2) \\leq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2)" + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "content": ". Specially, compared to not including an entropy term, that is " + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "inline_equation", + "content": "\\alpha_1 = 0" + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "content": ", in the objective function of Eq. (4), adding an entropy term results in a higher entropy after the algorithm converges, that is " + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "inline_equation", + "content": "H(p_{\\theta}^{*}(x_{t-1}|x_t,z);\\alpha_1,\\alpha_2) \\geq H(p_{\\theta}^{*}(x_{t-1}|x_t,z);0,\\alpha_2)" + }, + { + "bbox": [ + 313, + 300, + 555, + 373 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 376, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 449 + ], + "type": "text", + "content": "The proof of Theorem 3.2 is detailed in Appendix A.2. This theorem demonstrates that adding an entropy term to the loss function increases the policy entropy, thereby enhancing the algorithm's exploratory behavior. Through greater exploration capability, the agent can discover better actions and achieve higher returns." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "content": "Theorem 3.3. If " + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "inline_equation", + "content": "C = \\min_{x_{t-1} \\neq y_{t-1}} [\\log p_{pre}(y_{t-1}|x_t, z) - \\log p_{pre}(x_{t-1}|x_t, z)] / [Q^*(x_{t-1}, z) - Q^*(y_{t-1}, z)] > 0" + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "content": ", the objective function of Eq. (4) contains only a KL term and no entropy term, that is " + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "inline_equation", + "content": "\\alpha_1 = 0" + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "inline_equation", + "content": "1/C \\leq \\hat{\\alpha}_2 \\leq \\alpha_2" + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "content": ", we have " + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "inline_equation", + "content": "H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\hat{\\alpha}_2) \\geq H(p_\\theta^*(x_{t-1}|x_t, z); 0, \\alpha_2)" + }, + { + "bbox": [ + 313, + 453, + 554, + 514 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 517, + 554, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 554, + 578 + ], + "type": "text", + "content": "The proof of Theorem 3.3 is detailed in Appendix A.3. This theorem indicates that adding a KL term to the loss function may reduce the exploration capability of the policy if the parameters are not properly tuned, thereby affecting the algorithm's performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 586, + 504, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 504, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 504, + 598 + ], + "type": "text", + "content": "4. AdaEnt: A Practical Instantiation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 554, + 667 + ], + "type": "text", + "content": "Building on the above theoretical foundation, we propose AdaEnt, a plug-and-play module that adaptatively controls entropy and a number of steps to combat reward hacking in diffusion policy optimization. The architecture of AdaEnt is illustrated in Fig. 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 471, + 685 + ], + "type": "text", + "content": "4.1. Entropy-Driven Exploration" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "To cope with the diversity degradation brought by reward hacking issue, we introduce the entropy of denoising policy" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1927" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "inline_equation", + "content": "p_{\\theta}(x_{t - 1}|x_t,z)" + }, + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": " to enhance exploration. In a normal distribution, the entropy is related to its variance. From Eq. (3), we define a new variance:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 114, + 294, + 146 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 114, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 56, + 114, + 294, + 146 + ], + "type": "interline_equation", + "content": "\\sigma_ {T - t} = \\eta_ {\\theta , \\phi} (\\mathbf {x} _ {T - t}, T - t) \\sqrt {\\frac {1 - \\alpha_ {T - t - 1}}{(1 - \\alpha_ {T - t})} \\cdot \\left(1 - \\frac {\\alpha_ {T - t}}{\\alpha_ {T - t - 1}}\\right)},", + "image_path": "e564aa06ef6a23ac11dbeb4d26190f3bc296aa09a627c8681227ba546c954737.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "inline_equation", + "content": "\\eta_{\\theta ,\\phi}(\\mathbf{x}_{T - t},T - t)" + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "text", + "content": " is a dynamic, lightweight head with parameters " + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "text", + "content": " , built upon the neural network " + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta}" + }, + { + "bbox": [ + 55, + 152, + 295, + 189 + ], + "type": "text", + "content": " ,designed to provide a latent and step-aware dynamic variance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 195, + 295, + 218 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 295, + 218 + ], + "type": "text", + "content": "4.2. Trade-Off between Exploration and Regularization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 224, + 296, + 296 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 296 + ], + "type": "text", + "content": "In diffusion models, different denoising steps play various roles in image generation. Previous research has shown that earlier steps significantly impact the semantic content and global layout of the image, while later steps influence the texture and details [3]. Motivated by this phenomenon, we propose an adaptive regularization scheme." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "content": "We introduce an adaptive balance factor " + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "inline_equation", + "content": "\\gamma_{t}" + }, + { + "bbox": [ + 55, + 297, + 296, + 357 + ], + "type": "text", + "content": " to dynamically adjust the balance between exploration diversity and denoising convergence across the denoising process. For the sake of simplicity, we introduce the PPO algorithm [34]. Therefore, the objective to optimize (Eq. (4)) is updated:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 58, + 363, + 298, + 433 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 363, + 298, + 433 + ], + "spans": [ + { + "bbox": [ + 58, + 363, + 298, + 433 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {J} = \\max _ {\\theta} \\mathbb {E} [ \\min (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)} A, c l i p (\\frac {p _ {\\theta} (\\cdot)}{p _ {\\theta_ {t}} (\\cdot)}, 1 - \\epsilon , 1 + \\epsilon) A) \\\\ + \\sum_ {t = 1} ^ {T} \\left(1 - \\gamma_ {t}\\right) H \\left(p _ {\\theta} (\\cdot)\\right) - \\sum_ {t = 1} ^ {T} \\gamma_ {t} K L \\left(p _ {\\theta} (\\cdot)\\right) \\left\\| p _ {p r e} (\\cdot)\\right), \\tag {5} \\\\ \\end{array}", + "image_path": "a9e9250a015cde5bd64435a0026247ef4fa011080e5f18f91c1224f7802ca6e6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 434, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 295, + 481 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 434, + 295, + 481 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 434, + 295, + 481 + ], + "type": "text", + "content": " is an advantage function. According to Eq. (5), as the generation progresses, the regularization term will dominate the optimization and seek directions toward the distribution of real images within the high-reward region." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": "A prediction model " + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}_t)" + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": " is trained on each intermediate denoised (latent) image " + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_t" + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": ", with intermediate denoised images close to the final generated image " + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_0" + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": " labeled as 1, and those close to initial noise " + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_T" + }, + { + "bbox": [ + 55, + 481, + 296, + 540 + ], + "type": "text", + "content": " labeled as 0. The binary classification loss function is defined as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 546, + 295, + 581 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 581 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {p r e d} = - \\sum_ {i} \\left(y _ {i} \\log \\left(f \\left(x _ {i}\\right)\\right) + \\left(1 - y _ {i}\\right) \\log \\left(1 - f \\left(x _ {i}\\right)\\right)\\right), \\tag {6}", + "image_path": "1e27ca99b293d685b696afef37bb04a8355ac669bb5f28e7a99458c8538e736a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "inline_equation", + "content": "y_{i} \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": " is the label indicating whether the image is noiseless, and " + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "inline_equation", + "content": "f(x_{i})" + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": " is the model's predicted probability that image " + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": " is noiseless. After training, this network can evaluate the denoising level of any intermediate image during the denoising process, and we let " + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "inline_equation", + "content": "\\gamma_{t} = f(\\mathbf{x}_{t})" + }, + { + "bbox": [ + 55, + 582, + 295, + 665 + ], + "type": "text", + "content": " as a measure of the denoising process indicating how well an intermediate image aligns with real image distribution." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 186, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 186, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 186, + 685 + ], + "type": "text", + "content": "4.3. Dynamic Step Decision" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "One limitation of diffusion models is the reliance on many iterative denoising steps. The drawback will be exacerbated" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "when performing RL fine-tuning, leading to a high computation demand and exacerbation of reward hacking. As later denoising steps contribute less to the overall appearance of the generated image, applying an early stop on the denoising process will make training and inference more efficient and mediately alleviate reward hacking. The criterion of early stop is dynamically determined by how well the current image aligns with true image distribution. According to Sec. 4.2, " + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "inline_equation", + "content": "\\gamma_{t} = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": " means the current state of image satisfies the quality. Hence we apply early stop in this scenario, which is completely dynamic according to the complexity of generation and current image quality, without any predefined hyperparameters." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 229, + 555, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 229, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 229, + 555, + 373 + ], + "type": "text", + "content": "Note that in existing diffusion step adaptation methods, the total number of denoising steps have to be determined before hand, which is impossible for the adaptive module to make a decision based on sufficient information about the image quality. In contrast, our adaptation mechanism monitors the generated image at each step and determines the total number of steps based on real-time quality of the current state. Recent study on denoising diffusion models reveals the temporal information may be not necessary [38], which supports our design of dynamically cutting off the iterative denoising process. The algorithmic procedure of AdaEnt can be found in Appendix B." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 387, + 395, + 400 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 387, + 395, + 400 + ], + "spans": [ + { + "bbox": [ + 314, + 387, + 395, + 400 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 430, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 430, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 430, + 421 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 545 + ], + "type": "text", + "content": "Datasets. To compare with previous work, we use the same prompt set consisting of 45 different animal classes for training (Simple-animals), which has been widely adopted for reward fine-tuning [1, 49, 51]. For testing, we generate 8 different noises for each animal class, totaling 360 test samples. For generalization evaluation, larger-scale prompt sets are further considered in testing as done in [1], including 398 animal classes from ImageNet (ImageNet-animals) and the above 45 animals each concatenated with an activity (Activity-animals)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 653 + ], + "type": "text", + "content": "Rewards and Metrics. In this paper, we first deploy three metrics as the reward functions, that is, Aesthetic Score [33], Compressibility, and Incompressibility. To quantitatively evaluate performance in face of reward hacking, we then introduce cross-reward generalization and well-acknowledged metrics as suggested in [51]: ImageReward [46], Inception Score (IS) [32], PickScore [17], and ClipScore [27]. Details of each metric can be found in Appendix C." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 712 + ], + "type": "text", + "content": "Implementation Details. Following [1], we take Stable Diffusion v1.4 (SDv14) as our main test bed for reward finetuning. We also consider Stable Diffusion v1.5 (SDv15), v2.1-turbo (SDv21), and XL1.0 (XL) to verify the robustness of our method with different backbones. All models" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1928" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 218, + 199 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 218, + 199 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 218, + 199 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 218, + 199 + ], + "type": "image", + "image_path": "7e2716bbf81b9c217f24a7fcebaec1166ef2d6ff4bf556ca9c34c34be0d5161c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 555, + 231 + ], + "type": "text", + "content": "Figure 2. Reward curves of RL fine-tuning SDv14 with three different reward functions. Our proposed method consistently outperforms baselines in all three experimental scenarios, in terms of both peak achieved rewards and the sample efficiency." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 223, + 70, + 389, + 199 + ], + "blocks": [ + { + "bbox": [ + 223, + 70, + 389, + 199 + ], + "lines": [ + { + "bbox": [ + 223, + 70, + 389, + 199 + ], + "spans": [ + { + "bbox": [ + 223, + 70, + 389, + 199 + ], + "type": "image", + "image_path": "51d2b5b7fd95b548b53b906bb30906d7d916fff85b2c7171978946bab1220fb1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 395, + 70, + 555, + 199 + ], + "blocks": [ + { + "bbox": [ + 395, + 70, + 555, + 199 + ], + "lines": [ + { + "bbox": [ + 395, + 70, + 555, + 199 + ], + "spans": [ + { + "bbox": [ + 395, + 70, + 555, + 199 + ], + "type": "image", + "image_path": "4b225500f3bae1a5e4e23b453d44263071fcab49de0b4345acf9ce77a9ac5f5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 250, + 296, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 286 + ], + "type": "text", + "content": "are fine-tuned with low-rank adaptation (LoRA) [13] on 8 NVIDIA Tesla A100 GPUs. More training details can be found in Appendix C." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 297, + 141, + 309 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 141, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 141, + 309 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 295, + 508 + ], + "type": "text", + "content": "We benchmark our methods against state-of-the-art RL-based diffusion fine-tuning frameworks, including DDPO [1], DPOK [5], D3PO [49], and TDPO [51]. As shown in Fig. 2, our method demonstrates consistent superiority in learning all three reward functions, validating its effectiveness for optimizing downstream objectives. Crucially, our balanced state exploration and adaptive regularization mechanism achieves promote the rate of reward growth by a large margin compared to DPOK's fixed-weight regularization strategy. Unlike dense-reward methods requiring evaluations of noised states' long-term impacts, our approach implements stepwise adaptive entropy to guide denoising space exploration, enabling sample-efficient optimization. Consequently, we attain rewards slightly higher than the latest dense-reward method TDPO but using much fewer training samples." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 509, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 509, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 54, + 509, + 295, + 663 + ], + "type": "text", + "content": "Moreover, to systematically investigate reward hacking mitigation, we test the models optimized for Aesthetic Score with different aesthetic scores and general metrics for image quality and text-to-image alignment. We report the min, mean, and max values of each metric across the entire training process. As shown in Tab. 1, our generated images achieves overall higher scores on the diverse evaluation perspectives, effectively circumventing reward hacking phenomena even under elevated reward conditions. The results indicate that the proposed exploratory entropy effectively contributes to improved diversity and visual fidelity of generated images, along with less overfitting to the learning reward function." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "In summary, through this balanced exploration-regularization paradigm, our method demonstrates superior capability in discovering latent high-reward policies, thereby overcoming conventional performance plateaus" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 315, + 247, + 433, + 308 + ], + "blocks": [ + { + "bbox": [ + 315, + 247, + 433, + 308 + ], + "lines": [ + { + "bbox": [ + 315, + 247, + 433, + 308 + ], + "spans": [ + { + "bbox": [ + 315, + 247, + 433, + 308 + ], + "type": "image", + "image_path": "8f8012150ab19ae67b1081b44089d27fff32128eeea6c6b6b6ba787b5d8ecc38.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 316, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 384 + ], + "type": "text", + "content": "Figure 3. Distribution visualizations of our intermediate variables. (a) Distribution of predicted " + }, + { + "bbox": [ + 313, + 316, + 555, + 384 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 316, + 555, + 384 + ], + "type": "text", + "content": " during training. The predictor progressively learns to effectively distinguish latent codes of different stages. (b) Distribution of denoising step number during training. Our adaptive step mechanism effectively reduces the denoising steps in training." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 437, + 247, + 553, + 308 + ], + "blocks": [ + { + "bbox": [ + 437, + 247, + 553, + 308 + ], + "lines": [ + { + "bbox": [ + 437, + 247, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 437, + 247, + 553, + 308 + ], + "type": "image", + "image_path": "d6eb55d24e75359497c7695246717d0f014863d5daf831675e6155064fdd1171.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 405, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 430 + ], + "type": "text", + "content": "and achieving measurable enhancements in both generation quality and quantitative reward metrics." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 441, + 425, + 523 + ], + "blocks": [ + { + "bbox": [ + 317, + 441, + 425, + 523 + ], + "lines": [ + { + "bbox": [ + 317, + 441, + 425, + 523 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 425, + 523 + ], + "type": "image", + "image_path": "bda6cce7f839434189d3c9a6863c2c1372c14ae19b52441cb13816aa6fcca061.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 533, + 555, + 577 + ], + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 577 + ], + "type": "text", + "content": "Figure 4. Left: Entropy can be optimized in AdaEnt under different scaling factors. Right: Excessive application of either entropy or KL regularization impedes reward learning and may ultimately induce catastrophic collapse." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 425, + 441, + 554, + 523 + ], + "blocks": [ + { + "bbox": [ + 425, + 441, + 554, + 523 + ], + "lines": [ + { + "bbox": [ + 425, + 441, + 554, + 523 + ], + "spans": [ + { + "bbox": [ + 425, + 441, + 554, + 523 + ], + "type": "image", + "image_path": "de7bc767c5bc6589fb6cc7fe488d553f65ae6c0181d984855e5f61cc458f6322.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 599, + 409, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 599, + 409, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 599, + 409, + 613 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": "Effect of " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " Predictor. The coefficient " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " is introduced to adaptively balance the exploratory entropy and KL regularization based on the current denoising situation. In Fig. 3 (a), we show the distributions of " + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 617, + 555, + 715 + ], + "type": "text", + "content": " with 50-step denoising during training. It can be observed that the predictor progressively learns to effectively differentiate latent codes that are more distinctive (i.e., those at the beginning and end of denoising), whereas the results in the intermediate steps are" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1929" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 123, + 559, + 224 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Quality assessment from multiple unseen metric perspectives. All results are evaluated across training on the aesthetic score and sampled with the same training iteration. * refers to the Inception Score being sampled with the same aesthetic score of different methods. Since D3PO performs limited in learning rewards (see Fig. 2), the range of aesthetic score data is not enough for inception score collection. #Top1 indicates the instances when each method achieves the best performance across all metrics." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 123, + 559, + 224 + ], + "lines": [ + { + "bbox": [ + 56, + 123, + 559, + 224 + ], + "spans": [ + { + "bbox": [ + 56, + 123, + 559, + 224 + ], + "type": "table", + "html": "
MethodPickScoreInceptionScore*ImageRewardClipScore#Top1
MinMeanMaxMinMeanMaxMinMeanMaxMinMeanMax
DDPO [1]21.2821.3721.5321.1122.9523.670.89491.07071.24930.29660.30190.30831
DPOK [5]21.2721.4021.5320.4822.2923.460.90601.04961.13400.30080.30460.30831
D3PO [49]21.2821.2921.31---0.86230.89620.93170.30000.30120.30221
TDPO [51]21.0621.3821.7119.9421.8823.950.00480.84531.13540.27540.30100.31132
AdaEnt21.2421.4121.6422.1124.3426.190.90781.12121.31230.30000.30500.30988
", + "image_path": "55553df9e29d3772418448c72c5fce8fc6db81c0931c37efe8c2f8f4348aa0ec.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 240, + 294, + 385 + ], + "blocks": [ + { + "bbox": [ + 58, + 240, + 294, + 385 + ], + "lines": [ + { + "bbox": [ + 58, + 240, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 58, + 240, + 294, + 385 + ], + "type": "image", + "image_path": "60f9876634bb3d6e177ec10f86f563bfebf7294cea22e42ae7146848da21fc06.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 395, + 295, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 395, + 295, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 295, + 450 + ], + "type": "text", + "content": "Figure 5. Comparison of aesthetic reward curves and visual quality on SDv15 and SDv21. Our method consistently achieves higher rewards without over-optimizing across different backbones, while baseline methods compromise either visual quality or text-image alignment to attain aesthetic rewards." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "text", + "content": "predicted with non-binary confidence. This adaptive behavior dynamically balances entropy and the KL divergence as the denoising process progresses, facilitating a shift from exploration to regularization during each denoising process. Balance between Entropy and KL Divergence. In Fig. 4, we conduct an in-depth investigation into the effects of introducing entropy in reward learning. Specifically, with a fixed KL, we design ablation variants with entropy scaled by different factors (i.e., " + }, + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{e}" + }, + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "text", + "content": "). The baselines, DDPO and DPOK, have a fixed schedule of entropy (according to Eq. (1)). In our proposed framework, it is observed that entropy can be optimized to increase under different scaling factors, and corresponding with a better learning of reward. However, when entropy's order of magnitude overwhelms KL divergence (i.e., " + }, + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mu_{e} = 1.0" + }, + { + "bbox": [ + 54, + 486, + 295, + 714 + ], + "type": "text", + "content": "), over-exploration may lead to the collapse of training. Therefore, the result demonstrates the necessity of introducing the trade-off between entropy and KL, which ensures a well-balanced training process for improved generating performance." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 316, + 273, + 553, + 363 + ], + "blocks": [ + { + "bbox": [ + 313, + 242, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 264 + ], + "type": "text", + "content": "Table 2. Generalization results of SDv15 and SDv21 on unseen metrics. The metrics are abbreviated and the same in Tab. 1." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 273, + 553, + 363 + ], + "lines": [ + { + "bbox": [ + 316, + 273, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 273, + 553, + 363 + ], + "type": "table", + "html": "
MethodSDv15SDv21#Top1
PSISIRCSPSISIRCS
DDPO21.1716.900.820.3020.347.89-0.210.260
DPOK21.1818.680.840.3120.5310.280.040.272
D3PO21.0122.060.800.3318.843.14-1.580.252
TDPO19.682.40-1.570.2418.521.56-2.090.240
AdaEnt21.2118.990.850.3020.598.140.100.275
", + "image_path": "ebf51ade7304065f957aa981e79c5d6bd4443cb6b88efc62bf91fc64526adab3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 377, + 427, + 460 + ], + "blocks": [ + { + "bbox": [ + 316, + 377, + 427, + 460 + ], + "lines": [ + { + "bbox": [ + 316, + 377, + 427, + 460 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 427, + 460 + ], + "type": "image", + "image_path": "d80e1b8ff98d2572201721ee303d4809174f4fafaad5ba9d4be4c76f190a9a2a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 502 + ], + "type": "text", + "content": "Figure 6. Comparison of aesthetic reward curves and visual quality on SD-XL. Notably, we observe that our method is the only one that successfully works on SD-XL." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 430, + 378, + 553, + 449 + ], + "blocks": [ + { + "bbox": [ + 430, + 378, + 553, + 449 + ], + "lines": [ + { + "bbox": [ + 430, + 378, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 430, + 378, + 553, + 449 + ], + "type": "image", + "image_path": "f21f33e9edb90bf43b74dc454e1bb849107943946d6c077059b20f42550e77f4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 525, + 486, + 537 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 486, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 486, + 537 + ], + "type": "text", + "content": "5.4. Results on Different Backbones." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "type": "text", + "content": "In order to verify the generalizability of our method across different models for practical applications, we hereby extend our experiment to SDv15 and SDv21. In Fig. 5, our method also achieves superior effectiveness in reward learning. For generalization evaluation, we also evaluate the generated image quality with more metric perspectives for SDv15 and SDv21. As shown in Tab. 2, AdaEnt achieves superior overall quality, and despite TDPO shows comparable aesthetic score with our method, it is demonstrated to be failed in prompt-image alignments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "Moreover, considering that SD-XL is one of the most advanced and extensive pre-trained model, we provide a concise reward-learning experiment on SD-XL in Fig. 6. The results demonstrate that AdaEnt can still bring reward gains" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1930" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 167 + ], + "type": "image", + "image_path": "0a76a196830567cde79441aba76233adfd078f17a133e4f5465dde1d1d86fb51.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 295, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 295, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 295, + 210 + ], + "type": "text", + "content": "Figure 7. Comparison of generalization ability to larger scale prompt sets with unseen prompts. The experiments are conducted on SDv14 and trained with the small Simple-animals set." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 221, + 294, + 334 + ], + "blocks": [ + { + "bbox": [ + 56, + 221, + 294, + 334 + ], + "lines": [ + { + "bbox": [ + 56, + 221, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 56, + 221, + 294, + 334 + ], + "type": "image", + "image_path": "9a31224dd9cb1057704a0cbf0235c6f8cd05aeb9bb6f4647ef0a9368e1645e1e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 343, + 296, + 399 + ], + "lines": [ + { + "bbox": [ + 55, + 343, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 296, + 399 + ], + "type": "text", + "content": "Figure 8. Comparison of generalization ability to more complex prompts with multiple categories. All methods are implemented in SD-XL. The performance of our method is highlighted in darker colors, demonstrating AdaEnt's superior performance in both image quality and prompt-image alignment." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 208, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 208, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 208, + 434 + ], + "type": "text", + "content": "for larger and more powerful models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 444, + 241, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 444, + 241, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 241, + 456 + ], + "type": "text", + "content": "5.5. Generalization to Unseen Prompts." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 510 + ], + "type": "text", + "content": "To investigate whether the performance of RL fine-tuned models is limited to the prompts seen in training, which is an ordinary case of reward hacking, we hereby conduct prompts generalization experiments from two perspectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 605 + ], + "type": "text", + "content": "Generalization to larger-scale prompt sets. To evaluate the generalization ability to unseen prompts, we evaluate the aesthetic score on ImageNet-animals and Activity-animals. The experiments are conducted on SDv14, using prompts from Simple-animals as the training set. As presented in Fig. 7, our method achieves promising aesthetic scores on the unseen datasets, which demonstrates the effectiveness of our method." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Generalization to more complex prompts. Additionally, we assess the performance on more complex prompts. To be specific, we further conduct experiments on SD-XL with a subset of the HPSv2 dataset [44], which contains complex prompts that are not limited to animals. We identify several categories in the prompt set and measure the prompt-image consistency using clip-score. We also provide the aesthetic score to assess the quality of generated images. As presented in Fig. 8, our method exhibits the best perfor" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 97 + ], + "type": "text", + "content": "mance in the aesthetic score and the maintenance of image-prompts alignment, which suggests our superiority." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 104, + 499, + 117 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 104, + 499, + 117 + ], + "spans": [ + { + "bbox": [ + 313, + 104, + 499, + 117 + ], + "type": "text", + "content": "5.6. Computation Efficiency of AdaEnt" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 289 + ], + "type": "text", + "content": "Since the adaptive mechanism introduced in our work allows for the early termination of the denoising process once denoising is essentially complete, it is expected to improve the training efficiency of diffusion models. Firstly, we provide the visualized distributions of denoising step number during training in Fig. 3 (b), which verifies the capability of adaptively reducing the step number. Quantitatively, we take the aesthetic score " + }, + { + "bbox": [ + 313, + 121, + 555, + 289 + ], + "type": "inline_equation", + "content": "= 7" + }, + { + "bbox": [ + 313, + 121, + 555, + 289 + ], + "type": "text", + "content": " as the training terminals of each method for evaluation. In Tab. 3, we provide GFLOPs and time for single batch (size=12) sampling, along with the total sampling step and time for the overall training process. The results show that the proposed AdaEnt can significantly accelerate the training process with almost no increase in computational cost (i.e., GFLOPS)." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 315, + 353, + 544, + 444 + ], + "blocks": [ + { + "bbox": [ + 313, + 300, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 344 + ], + "type": "text", + "content": "Table 3. Computation Efficiency of integrating our proposed dynamic early stop mechanism with existing RL frameworks. TDS refers to total denoising steps. Results of both overall training and single-batch inference are reported." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 353, + 544, + 444 + ], + "lines": [ + { + "bbox": [ + 315, + 353, + 544, + 444 + ], + "spans": [ + { + "bbox": [ + 315, + 353, + 544, + 444 + ], + "type": "table", + "html": "
MethodOverallSingle batch
TDSTimeGFLOPSTime
DDPO1.843+e611.98h1.034e+069.56s
DPOK2.015+e614.76h9.69s
DDPO+Ours0.242+e65.52h1.044e+066.01s
DPOK+Ours0.294+e67.16h7.12s
", + "image_path": "a610126ae975f8c24cac1332b69cd506a893a2a30ea9499ddf5eeda5bbee89cb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 466, + 388, + 478 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 713 + ], + "type": "text", + "content": "This work addresses the critical challenge of reward hacking in diffusion model optimization through reinforcement learning, where conventional RL frameworks tend to produce overfitted outputs that compromise distributional fidelity and generation diversity. Our theoretical analyses establish entropy as a proxy to enhance the exploratory capacity of denoising policies. The theoretical basis is instantiated as AdaEnt, a framework that achieves adaptive entropy control and step-wise optimization through three interconnected mechanisms: entropy-aware policy updates, distributional consistency regularization, and noise-level dependent exploration scheduling. Extensive experiments demonstrate that our method effectively mitigates reward hacking while preserving the diffusion model's inherent generation capabilities, establishing new state-of-the-art performance in quality-diversity trade-off across multiple diffusion-based RL benchmarks. The adaptive entropy proposed in this paper offers new perspectives for future RL fine-tuning of diffusion models." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 314, + 757 + ], + "type": "text", + "content": "1931" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 161, + 85 + ], + "type": "text", + "content": "7. Acknowledgment" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 218, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 218, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 218, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 296, + 262 + ], + "type": "text", + "content": "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 263, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 263, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 263, + 296, + 319 + ], + "type": "text", + "content": "[2] Bo Dai, Albert E. Shaw, Lihong Li, Lin Xiao, Niao He, Zhen Liu, Jianshu Chen, and Le Song. SBEED: convergent reinforcement learning with nonlinear function approximation. In Proceedings of International Conference on Machine Learning., 2018. 12" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 319, + 296, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 319, + 296, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 319, + 296, + 352 + ], + "type": "text", + "content": "[3] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 61, + 354, + 296, + 408 + ], + "type": "text", + "content": "[4] Luca Eyring, Shyamgopal Karthik, Karsten Roth, Alexey Dosovitskiy, and Zeynep Akata. Reno: Enhancing one-step text-to-image models through reward-based noise optimization. Advances in Neural Information Processing Systems, 37:125487-125519, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 410, + 296, + 475 + ], + "type": "text", + "content": "[5] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36, 2024. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 296, + 510 + ], + "type": "text", + "content": "[6] Yaozhong Gan, Renye Yan, Xiaoyang Tan, Zhe Wu, and Jun-liang Xing. Transductive off-policy proximal policy optimization. arXiv preprint arXiv:2406.03894, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 510, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 510, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 510, + 296, + 544 + ], + "type": "text", + "content": "[7] Yaozhong Gan, Renye Yan, Zhe Wu, and Junliang Xing. Reflective policy optimization. arXiv preprint arXiv:2406.03678, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 62, + 545, + 296, + 578 + ], + "type": "text", + "content": "[8] Leo Gao, John Schulman, and Jacob Hilton. Scaling laws for reward model overoptimization. In International Conference on Machine Learning, pages 10835-10866. PMLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 578, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 578, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 578, + 296, + 624 + ], + "type": "text", + "content": "[9] Tuomas Haarnoja, Haoran Tang, Pieter Abbeel, and Sergey Levine. Reinforcement learning with deep energy-based policies. In International conference on machine learning, pages 1352-1361. PMLR, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[10] Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In International conference on machine learning, pages 1861-1870. Pmlr, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 714 + ], + "type": "text", + "content": "[11] Seungyul Han and Youngchul Sung. A max-min entropy framework for reinforcement learning. Advances in Neural Information Processing Systems, 34:25732-25745, 2021. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 187 + ], + "type": "text", + "content": "This paper is supported by the National Science Foundation of China under Grant No.62025401, 62341407, U24B200170, 62406008; Beijing Municipal Science and Technology Program under Grant Z241100004224015; Beijing Natural Science Foundation under Grant No.L223004; '111' Project under Grant No.B18001. This work is also supported in part by the Natural Science Foundation of China under Grant No. 62222606." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 106 + ], + "type": "text", + "content": "[12] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 162 + ], + "type": "text", + "content": "[13] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 209 + ], + "type": "text", + "content": "[14] Rohit Jena, Ali Taghibakhshi, Sahil Jain, Gerald Shen, Nima Tajbakhsh, and Arash Vahdat. Elucidating optimal reward-diversity tradeoffs in text-to-image diffusion models. arXiv preprint arXiv:2409.06493, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 211, + 553, + 265 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 211, + 553, + 265 + ], + "spans": [ + { + "bbox": [ + 316, + 211, + 553, + 265 + ], + "type": "text", + "content": "[15] Bo-Kyeong Kim, Hyoung-Kyu Song, Thibault Castells, and Shinkook Choi. Bk-sdm: Architecturally compressed stable diffusion for efficient text-to-image generation. In Workshop on Efficient Systems for Foundation Models@ ICML2023, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 268, + 553, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 553, + 301 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 553, + 301 + ], + "type": "text", + "content": "[16] Diederik Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. Advances in neural information processing systems, 34:21696-21707, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 303, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 303, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 303, + 553, + 357 + ], + "type": "text", + "content": "[17] Yuval Kirstain, Adam Polyak, Uriel Singer, Shahbuland Matiana, Joe Penna, and Omer Levy. Pick-a-pic: An open dataset of user preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 404 + ], + "type": "text", + "content": "[18] Senmao Li, Taihang Hu, Fahad Shahbaz Khan, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster diffusion: Rethinking the role of unet encoder in diffusion models. CoRR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 316, + 406, + 555, + 450 + ], + "type": "text", + "content": "[19] Xinyin Ma, Gongfan Fang, and Xinchao Wang. Deepcache: Accelerating diffusion models for free. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15762-15772, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 452, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 496 + ], + "type": "text", + "content": "[20] Jincheng Mei, Chenjun Xiao, Ruitong Huang, Dale Schuurmans, and Martin Müller. On principled entropy exploration in policy optimization. In Proceedings of International Joint Conference on Artificial Intelligence, 2019. 12" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 553 + ], + "type": "text", + "content": "[21] Yuchun Miao, Sen Zhang, Liang Ding, Rong Bao, Lefei Zhang, and Dacheng Tao. Inform: Mitigating reward hacking in rlhf via information-theoretic reward modeling. Advances in Neural Information Processing Systems, 37: 134387-134429, 2025. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 609 + ], + "type": "text", + "content": "[22] Volodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International conference on machine learning, pages 1928-1937. PmLR, 2016. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 612, + 553, + 667 + ], + "type": "text", + "content": "[23] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[24] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International conference on machine learning, pages 8162-8171. PMLR, 2021. 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1932" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[25] Owen Oertell, Shikun Sun, Yiding Chen, Jin Peng Zhou, Zhiyong Wang, and Wen Sun. Efficient controllable diffusion via optimal classifier guidance. arXiv preprint arXiv:2505.21666, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[26] Mihir Prabhuesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "text", + "content": "[27] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning. PmLR. 5, 15" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 285 + ], + "type": "text", + "content": "[28] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[29] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 295, + 386 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 453 + ], + "type": "text", + "content": "[31] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. Advances in neural information processing systems, 35:36479-36494, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 498 + ], + "type": "text", + "content": "[32] Tim Salimans, Ian Goodfellow, Wojciech Zaremba, Vicki Cheung, Alec Radford, and Xi Chen. Improved techniques for training gans. Advances in neural information processing systems, 29, 2016. 5, 15" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 566 + ], + "type": "text", + "content": "[33] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in neural information processing systems, 2022. 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[34] John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Radford, and Oleg Klimov. Proximal policy optimization algorithms. arXiv preprint arXiv:1707.06347, 2017. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[35] Joar Skalse, Nikolaus Howe, Dmitrii Krasheninnikov, and David Krueger. Defining and characterizing reward gaming. Advances in Neural Information Processing Systems, 35:9460-9471, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 646, + 294, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 646, + 294, + 700 + ], + "spans": [ + { + "bbox": [ + 56, + 646, + 294, + 700 + ], + "type": "text", + "content": "[36] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International conference on machine learning, pages 2256-2265. PMLR, 2015. 1, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 294, + 713 + ], + "type": "text", + "content": "[37] Jiaming Song, Chenlin Meng, and Stefano Ermon." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020.3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[38] Qiao Sun, Zhicheng Jiang, Hanhong Zhao, and Kaiming He. Is noise conditioning necessary for denoising generative models? arXiv preprint arXiv:2502.13129, 2025. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 196 + ], + "type": "text", + "content": "[39] Shikun Sun, Chengrui Wang, Min Zhou, Zixuan Wang, Xiaoyu Qin, Tiezheng Ge, Bo Zheng, and Jia Jia. Depo: Enhancing e-commerce image background generation with short trajectory direct expected preference optimization. In Proceedings of the 33nd ACM International Conference on Multimedia, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 240 + ], + "type": "text", + "content": "[40] Marc Toussaint. Robot trajectory optimization using approximate inference. In Proceedings of the 26th annual international conference on machine learning, pages 1049-1056, 2009. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 308 + ], + "type": "text", + "content": "[41] Masatoshi Uehara, Yulai Zhao, Kevin Black, Ehsan Hajiramezanali, Gabriele Scalia, Nathaniel Lee Diamant, Alex M Tseng, Tommaso Biancalani, and Sergey Levine. Finetuning of continuous-time diffusion models as entropy-regularized control. arXiv preprint arXiv:2402.15194, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 376 + ], + "type": "text", + "content": "[42] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8228-8238, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 377, + 553, + 443 + ], + "type": "text", + "content": "[43] Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, et al. Cache me if you can: Accelerating diffusion models through block caching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6211-6220, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[44] Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341, 2023.8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[45] Xiaoshi Wu, Keqiang Sun, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score: Better aligning text-to-image models with human preference. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2096-2105, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 557, + 553, + 612 + ], + "type": "text", + "content": "[46] Jiazheng Xu, Xiao Liu, Yuchen Wu, Yuxuan Tong, Qinkai Li, Ming Ding, Jie Tang, and Yuxiao Dong. Imagereward: Learning and evaluating human preferences for text-to-image generation. Advances in Neural Information Processing Systems, 2023. 5, 15" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 668 + ], + "type": "text", + "content": "[47] Renye Yan, Zhe Wu, Yuan Zhan, Pin Tao, Zongwei Wang, Yimao Cai, and Junliang Xing. Mnemonic dictionary learning for intrinsic motivation in reinforcement learning. In 2023 International Joint Conference on Neural Networks (IJCNN), pages 1-7. IEEE, 2023. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[48] Renye Yan, Yaozhong Gan, You Wu, Ling Liang, Junliang Xing, Yimao Cai, and Ru Huang. The exploration-exploitation dilemma revisited: An entropy perspective. arXiv preprint arXiv:2408.09974, 2024. 2" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1933" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 228 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 137 + ], + "type": "text", + "content": "[49] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8941-8951, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 173 + ], + "type": "text", + "content": "[50] Hui Zhang, Zuxuan Wu, Zhen Xing, Jie Shao, and Yu-Gang Jiang. Adadiff: Adaptive step selection for fast diffusion. arXiv preprint arXiv:2311.14768, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 294, + 228 + ], + "type": "text", + "content": "[51] Ziyi Zhang, Sen Zhang, Yibing Zhan, Yong Luo, Yonggang Wen, and Dacheng Tao. Confronting reward overoptimization for diffusion models: A perspective of inductive and primacy biases. In International Conference on Machine Learning, pages 60396-60413. PMLR, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1934" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_content_list.json b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..1528db165adec6e9437371eb9df0e25b8c6dabb9 --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_content_list.json @@ -0,0 +1,1591 @@ +[ + { + "type": "text", + "text": "Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shaobo Zhang", + "bbox": [ + 183, + 202, + 308, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuhang Huang", + "bbox": [ + 354, + 202, + 482, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ziyu Guan", + "bbox": [ + 387, + 222, + 482, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wanqing Zhao $^{1*}$", + "bbox": [ + 526, + 203, + 658, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinye Peng $^{1*}$", + "bbox": [ + 529, + 222, + 630, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wei Zhao3", + "bbox": [ + 697, + 203, + 784, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Northwest University", + "bbox": [ + 125, + 238, + 303, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National University of Defense Technology", + "bbox": [ + 341, + 239, + 692, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Xidian University", + "bbox": [ + 732, + 238, + 883, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangshaobo, zhaowq, ziyuguan, pjy}@nwu.edu.cn, huangai@nudt.edu.cn, ywzhao@mail.xidian.edu.cn", + "bbox": [ + 101, + 258, + 913, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This paper introduces EA6D, a novel diffusion-based framework for 6D pose estimation that operates effectively in any environment. Traditional pose estimation methods struggle with the variability and complexity of real-world scenarios, often leading to overfitting on controlled datasets and poor generalization to new scenes. To address these challenges, we propose a generative pose estimation paradigm that generates environment-independent object representations for pose estimation, which are robust to environmental variations such as illumination, occlusion, and background clutter. Specifically, we propose the novel Environment Decoupling Diffusion Model (EDDM) which separates object representations from environmental factors while enabling efficient few-step sampling by leveraging input image priors instead of pure noise initialization. We validate our approach on four standard benchmarks and a self-made dataset DiverseScenes. The results demonstrate that EA6D, trained using only synthetic data, can outperform the state-of-the-art methods with both synthetic and realistic data. In particular, for fair comparisons with synthetic data, we can exceed the previous SOTA by $18.1\\%$ and $33.5\\%$ on Linemod and Linemod-Occluded datasets respectively. Project page: https://github.com/acf22/EA6D", + "bbox": [ + 89, + 340, + 483, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 717, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "6D Pose Estimation is a pivotal task in the field of computer vision and robotics, which aims to determine the exact position and orientation of an object in three-dimensional space relative to a camera or a coordinate system. The ability to accurately perceive an object's pose is fundamental for a wide range of applications, including robotic manipulation, augmented reality and autonomous navigation.", + "bbox": [ + 89, + 742, + 482, + 847 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, a series of deep-learning-based object 6D pose estimation methods [17, 22, 23, 28, 31, 35, 41, 44] have", + "bbox": [ + 89, + 848, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/58882ef3d1bc59fa258a1c3c65ea996bc0a68d9cec782bd97aeba59a485f1499.jpg", + "image_caption": [ + "Figure 1. Different from typical RGB-based pose estimation frameworks, which are easily influenced by environmental factors leading to inaccurate pose estimation, we propose a generative paradigm that leverages diffusion models to generate environment-independent object representations for pose estimation." + ], + "image_footnote": [], + "bbox": [ + 517, + 305, + 901, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "been proposed. These methods achieve state-of-the-art results, but require large annotated datasets. Generating accurate annotations for pose data is a labor-intensive and time-consuming process, leading most deep-learning-based 6D pose estimation methods to be trained and tested on closed datasets. These datasets [3, 10, 12, 16] typically contain only a few hundred images with limited variation and controlled factors, such as object arrangement, lighting conditions, and background, resulting in overfitting to the training set and the struggle to generalize to new scenes.", + "bbox": [ + 511, + 595, + 906, + 747 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, synthetic datasets have been employed as an additional complement to real datasets, or alternative to real datasets, as they can be easily and automatically annotated in simulated environments. The primary challenge with synthetic data is the domain discrepancy between real and synthetic datasets, which often leads to reduced accuracy when applied to real-world scenarios. To alleviate this problem, some self-supervised methods [20, 37, 39, 40, 46] are proposed, which leverage synthetic images with pose annotations and unannotated real images during training.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding authors", + "bbox": [ + 109, + 887, + 238, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8678", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These methods harness intrinsic structures to generate supervision signals and implement constraints on the discrepancies between real and synthetic images, thereby mitigating the domain gap.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Practical applications such as AR and autonomous driving, which need to operate in uncontrolled environments, make 6D pose estimation more challenging. Here, object appearances vary due to fluctuating environmental conditions, background clutter, and occlusions, which often exceed the variability present in training datasets. Previous methods usually use CNNs as image encoder to extract features for pose estimation (as shown on the top side of Figure 1). These features are highly sensitive to environmental factors. When the environment changes, CNNs struggle to accurately extract object features. This hinders the model's ability to generalize to new scenes, potentially diminishing the accuracy of pose estimation and even leading to object pose estimation failure. While self-supervised methods are proficient at leveraging unlabeled data, they may not fully account for the extensive real-world variability found in these uncontrolled environments, leading to models that are not sufficiently robust for novel conditions. Therefore, there is an urgent need for approaches that can adapt to a variety of dynamic and diverse environments, ensuring reliable pose estimation capabilities.", + "bbox": [ + 91, + 154, + 482, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel paradigm for pose estimation in uncertain environments, including unseen scenes, varying lighting, occlusion, and other possible situations. We consider a generative strategy that generates environment-independent object representations for pose estimation (bottom side of Figure 1). Diffusion models [7, 11, 34] have recently achieved promising results in various generative tasks, including image generating and editing. These models recover high-quality, deterministic samples from noisy and some conditions through a progressive denoising procedure. Motivated by the above, we design an Environment-Agnostic object 6D POSE estimation framework (named EA6D) that exploits the diffusion model to generate environment-independent representations in latent space for pose estimation, while input images serve as conditional cues, illustrated on the bottom side of Figure 1. Specifically, we propose an Environment Decoupling Diffusion Model (EDDM) which employs a novel environment decoupling diffusion process to intrinsically decouple the environment-independent object representations from input images. By performing validation on four popular benchmarks and a self-made dataset, the results show that our method significantly outperforms state-of-the-art self-supervised methods and provides a significant improvement in generalization to unseen scenarios.", + "bbox": [ + 91, + 472, + 482, + 848 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In sum, the main contributions in this paper include:", + "bbox": [ + 109, + 849, + 455, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "- We propose a new paradigm for environment-agnostic object 6D pose estimation, by generating environment", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "independent object representations rather than estimating poses directly from input images.", + "bbox": [ + 526, + 90, + 903, + 119 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel Environment Decoupling Diffusion Model which employs an explicit environmental decoupling mechanism via gradient modelling and conditional generation to separate object representations from environment, while enabling efficient several-step sampling.", + "- We conduct extensive experiments on four public datasets including Linemod, Linemod-Occluded, YCB-V, HomebrewedDB and a self-made dataset DiverseScenes, showing that our method surpasses current state-of-the-art self-supervised methods. Our method exhibits remarkable generalization capabilities, even in previously unseen scenes, which is crucial for practical applications." + ], + "bbox": [ + 513, + 121, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 316, + 650, + 330 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Fully-Supervised 6D Object Pose Estimation. Early approaches rely on local or global features to establish correspondences between keypoints on CAD models [1, 3, 4]. More recently, learning-based methods, particularly those utilizing Convolutional Neural Networks (CNNs), have become dominant. These methods extract deep features and establish mappings from RGB images to 6D object pose [14, 17, 41, 44]. Due to the inherent nonlinearity of the rotation space, the accuracy of directly regressed object pose is often limited [21]. To enhance accuracy, many methods employ CNNs to establish 2D-3D correspondences, followed by solving the object pose using the Perspective-n-Point (PnP) algorithm [19]. These methods can be categorized into sparse and dense correspondence methods. Sparse methods [28, 33, 38, 45] predict the 2D projections of predefined 3D keypoints, while dense methods [22, 23, 27, 35, 41, 48] predict 3D coordinates on the object model corresponding to image pixels thus providing better robustness to occlusions and clutter. However, fully-supervised 6D pose estimation methods are constrained by the need for accurate and diverse training data, which is often scarce due to the labor-intensive nature of precise annotation and the limited variation in available datasets, resulting in poor generalization to real-world and uncontrolled environments.", + "bbox": [ + 511, + 340, + 903, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Self-Supervised 6D Object Pose Estimation. Considering the tremendous effort to collect large amounts of annotations for 6D object pose, some methods [26, 29, 36] have turned to using synthetic images for training. However, because of the domain gap between real and synthetic images, the performances of these methods are significantly lower compared to the methods that use real images with object poses for training. To narrow the domain gap, a few recent self-supervised methods [5, 20, 37, 39, 40, 43, 46, 49] have begun to incorporate both synthetic images with known object poses and unannotated real images into their training datasets. However, in diverse real-world application sce", + "bbox": [ + 511, + 719, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8679", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "narios, models must contend with a variety of backgrounds, lighting conditions, and other environmental factors that are difficult to be adequately included in the training dataset. These environmental factors can significantly alter the appearance of objects within images. As a result, features learned by the model for the same object may vary across different scenes, leading to diminished generalization capabilities when encountering new scenes. Different from self-supervised learning methods, our proposed method directly generates environment-independent object representations to fundamentally eliminate the impact of environmental disturbances.", + "bbox": [ + 89, + 90, + 480, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion-based Models for Vision Tasks. Originally introduced for image generation, diffusion models [7, 11, 34] have demonstrated impressive generation capabilities and have been explored in various tasks, including edge detection [47], image inpainting [25], and 6D pose estimation [9, 42, 45]. DiffusionEdge [47] presents a novel diffusion-based edge detector capable of generating precise and sharp edge maps without post-processing. 6D-diff [45] formulates object keypoints detection as a reverse diffusion process and employs a MoC-based forward process to leverage distribution priors in intermediate representations. Method [42] introduces three novel aggregation networks designed to effectively aggregate diffusion features, showing superior generalizability for object pose estimation. Currently, no existing work has yet explored the potential of diffusion models to generate environment-independent object representations, which could significantly improve the robustness of pose estimation in diverse environments.", + "bbox": [ + 91, + 273, + 486, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 563, + 181, + 578 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 589, + 200, + 603 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Traditional pose estimation methods typically estimate pose directly from images. Environmental factors such as background clutter, illumination variations, and occlusion in images can limit the accuracy and robustness of pose estimation. Pursuing a different paradigm, we propose a diffusion-based method that generates environment-independent object representations for pose estimation. Traditional diffusion models are predominantly utilized for image generation tasks, where they excel at creating new synthetic images. These models operate on the principle of gradually adding noise to an input image and then learning to reverse this process, effectively denoising the image over multiple iterations to recover the original content. In this paper, we extend the concept of diffusion models to a completely new domain by recovering environment-independent object representations for pose estimation.", + "bbox": [ + 89, + 612, + 482, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall framework of the proposed method EA6D is shown in Figure 2. The first stage is environment-independent object representation encoding. A synthesized", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "pure object image with the same pose as the input image, but free from background, lighting, and occlusion, is encoded into a latent space using a Variational Autoencoder (VAE) [18]. This encoding results in an object representation $g_0 \\in R^{64 \\times 64 \\times 3}$ , which can be decoded back to the original pure object image by the decoder of VAE network, demonstrating that the representation contains only information about the object independent of the environment. Following this, a forward diffusion process is applied to $g_0$ to generate a noisy representation $g_t$ , which can be split into two sub-processes: the clean object representation $g_0$ to input image latent representation $r_0$ , meanwhile Gaussian noise is incrementally added. To denoise and recover the environment-independent object representations from $g_t$ , an Environment Decoupling Diffusion Model (EDDM) is introduced. A conditional encoder extracts multi-level features from the input images. These features are then fused with the features of the U-Net encoder and are input to U-Net decoders. The EDDM outputs a decoupled object representation, which potentially contains object pose information, independent of environmental factors. Finally, a pose decoder predicts the pose information, including rotation quaternion and translation vector. The inference process in EA6D is thus a combination of diffusion-based denoising and pose decoding, enabling robust and accurate pose estimation in diverse and complex environments.", + "bbox": [ + 511, + 90, + 906, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Environment Decoupling Diffusion for Generating Object Representations", + "text_level": 1, + "bbox": [ + 511, + 500, + 906, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Current 6D pose estimation methods often face robustness challenges in dynamic or unseen environments, as their object representations are influenced by complex environmental variations, resulting in multimodal distributions of features. CNNs-based methods typically use a single model to fit the entire data distribution. When facing multimodal distributions, the single model struggles to capture all the different patterns simultaneously, thereby resulting in a notable decrease in performance in new environments. To avoid the negative impact, we propose to extract environment-independent representations, enabling more accurate and generalizable pose estimation. Recent diffusion models show remarkable capabilities for modeling complex multimodal distribution, which is helpful for learning about environment-independent representations. Therefore, we propose to construct our framework based on the diffusion models to cover the complex distribution of object representations. More importantly, to acquire the environment-independent component from the input image distribution, we propose a novel environment decoupling diffusion model for generating environment-independent object representations inspired by the decoupled diffusion models [15, 47]. Different from the vanilla decoupled diffusion process that focuses on splitting the image and the", + "bbox": [ + 511, + 537, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8680", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bec0c1acf76e5faa60c7e4db7041f7da3ef6f44ef832801d32b4cb7bef0d4f39.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed method. In training phase, the proposed method involves training a Variational Autoencoder (VAE) to encode pure object images into environment-independent representations. These representations are then sampled to create noisy representations, which are processed by an Environment Decoupling Diffusion Model (EDDM). The EDDM uses a dual structure with one UNet encoder and two UNet decoders to decouple the environment-independent object representations from the input image with complex environments. The pose decoder is trained to predict the pose from the representations, enabling robust pose estimation across various environments. In testing phase, a random Gaussian noise is fed into the EDDM without VAE." + ], + "image_footnote": [], + "bbox": [ + 99, + 103, + 897, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "noise components, we aim to decouple the environment-independent object representations from the input image with complex environments. To this end, we introduce the improved forward and reversed diffusion processes.", + "bbox": [ + 88, + 417, + 480, + 479 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1.Forward Process", + "text_level": 1, + "bbox": [ + 89, + 484, + 254, + 498 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The vanilla decoupled diffusion process reformulates the conventional image-to-noise process into two stages: image-to-zero and zero-to-noise [15]. Corresponding to current object representation generation, there will be a clean object representation to zero process and a zero to noise process. However, such a paradigm only describes how to learn the environment-independent object representation from noise, without a clue on how to decouple it from the input images with complex environments. To address this problem, this paper proposes an environment decoupling diffusion process to intrinsically decouple the environment-independent object representation from the input image. We follow [30] to conduct the diffusion process in the latent space, i.e., our goal is to obtain the environment-independent object latent representations from the input image that mixes the object with the complex environments:", + "bbox": [ + 88, + 503, + 483, + 746 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(g _ {t} \\mid g _ {0}\\right) = \\mathcal {N} \\left(g _ {t}; g _ {0} + \\int_ {0} ^ {t} \\left(r _ {0} - g _ {0}\\right) d t, t I\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 766, + 482, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $g_{0}$ and $g_{t}$ denote the clean and noisy object latent representations respectively, $t\\in [0,1]$ is the time step, $r_0$ denotes the latent representation of input image. (A Variational Autoencoder features the latent representation.) According to the above equation, we can sample $g_{t}$ by $g_{t} = g_{0} + \\int_{0}^{t}(r_{0} - g_{0})dt + \\sqrt{t} n,$ where $n\\in \\mathcal{N}(0,I)$ . Ob", + "bbox": [ + 89, + 808, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "viously, the forward process describes two sub-processes: $g_0 + \\int_0^t (r_0 - g_0)dt$ is the clean object representation to input image latent representation and $\\sqrt{t} n$ is the zero to noise. It is worth noting that the first term contains the decay gradient from the input image to the environment-independent object representation, i.e., $r_0 - g_0$ . We call this term the environment decoupling gradient, which guides the generation from the input image representation to object representation. The environment decoupling gradient provides explicit evidence for our goal of decoupling the environment-independent object representations, allowing for accurate and efficient generation.", + "bbox": [ + 511, + 417, + 906, + 602 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.2. Reversed Process", + "text_level": 1, + "bbox": [ + 511, + 609, + 679, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The goal of the reversed process is to gradually recover the clean object representation $g_{0}$ from the noisy input $g_{1}$ . Following the derivation of [15], we can formally represent the reversed process as a conditional probability:", + "bbox": [ + 511, + 628, + 905, + 690 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} q \\left(g _ {t - \\triangle t} \\mid g _ {t}, g _ {0}\\right) = \\mathcal {N} \\left(g _ {t} + \\int_ {t} ^ {t - \\triangle t} \\left(r _ {0} - g _ {0}\\right) d t - \\frac {\\triangle t}{\\sqrt {t}} n, \\right. \\\\ \\frac {\\triangle t (t - \\triangle t)}{t} I), \\tag {2} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 700, + 903, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\triangle t$ is the smallest time step. Inheriting DDM, the reversed process benefits from the analytic gradient $r_0 - g_0$ , enabling approximate analytic solution for few-step generation (even with only one step). Different from general diffusion models that recover the goal content from a normal distribution, our reversed process aims to generate object representations from the combination of the latent image representations and the noise. We can easily obtain the ini", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "header", + "text": "- - - - → only training", + "bbox": [ + 99, + 89, + 220, + 103 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8681", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tial variable in the reversed process according to the forward process, i.e., $g_{1} = r_{0} + n$ . Intuitively, the generation process starts with the mixture of $r_0$ and $n$ instead of pure noise, which provides important prior information, thus speeding up the inference procedure and improving the generation quality. The reversed conditional probability $q(g_{t - \\triangle t}|g_t,g_0)$ includes $r_0 - g_0$ and $n$ , which are unknown in the reversed process. Therefore, we need to use a neural network $\\theta$ to parameterize $q(g_{t - \\triangle t}|g_t,g_0)$ . Practically, the neural network has two outputs: $d_{\\theta}$ and $n_{\\theta}$ , corresponding to the two unknown terms.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.3. Network Architecture", + "text_level": 1, + "bbox": [ + 89, + 263, + 290, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As shown in Figure 2, we construct our network based on a U-Net architecture including one encoder and two decoders. A pre-trained Swin Transformer [24] serves as a conditional encoder, extracting multi-level features from the input images. These features are then fused with the features of the U-Net encoder at corresponding levels through attention mechanisms, acting as inputs for the U-Net decoders. In practice, the input images will be obtained by an off-the-shelf object detector such as Yolo-v4 [2].", + "bbox": [ + 89, + 282, + 483, + 417 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Pose Decoder for Object Representations", + "text_level": 1, + "bbox": [ + 89, + 426, + 441, + 443 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The pose decoder is designed to predict the 6D pose, including both the rotation and translation, from the object representations generated from EDDM. The decoder consists of three modules: a ResNet block to further extract the features of $\\hat{g}_0^t$ , a rotation head for rotation estimation, and a translation head for translation prediction. The ResNet block enhances feature extraction through three convolutional layers. The first layer employs a $1 \\times 1$ kernel to expand the feature dimension, capturing richer information. The second layer uses a $3 \\times 3$ kernel for spatial feature reduction. The third layer also uses a $1 \\times 1$ kernel. The rotation head is designed with a series of $3 \\times 3$ convolutional layers, followed by a generalized average pooling layer and a fully connected layer, culminating in the prediction of the rotation quaternion $R$ .", + "bbox": [ + 89, + 449, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For 3D translation, since directly regressing the translation in 3D space does not work well in practice, we adopt Scale-invariant Translation Estimation (SITE) [22] to decouple the translation into the 2D location $(o_x,o_y)$ of the projected 3D centroid and the object's distance $T_{z}$ towards the camera. The translation head is constituted by a series of convolutional layers followed by fully connected layers, which predict the scale-invariant translation parameters $T_{site} = (\\Delta x,\\Delta y,\\Delta z)$ . These parameters are then utilized to compute the object's 2D coordinates and depth, as delineated by the equations:", + "bbox": [ + 89, + 675, + 483, + 842 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{l l} o _ {x} = & \\Delta x \\cdot w + c _ {x} \\\\ o _ {y} = & \\Delta y \\cdot h + c _ {y} \\\\ T _ {z} = & \\Delta z \\cdot r \\end{array} \\right. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 849, + 483, + 898 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w$ and $h$ represent the width and height of the bounding box, $c_{x}$ and $c_{y}$ denote the coordinates of the box's center, and $r$ is the scaling factor derived from the image's dimensions. Given the camera intrinsics $k$ , the translation $T$ can be calculated via back-projection $T = K^{-1}T_{z}[o_{x}, o_{y}, 1]^{T}$ . This method allows for a more accurate and robust estimation of the object's 3D translation, thereby enhancing the overall pose estimation accuracy.", + "bbox": [ + 511, + 90, + 905, + 212 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Training Procedure", + "text_level": 1, + "bbox": [ + 511, + 219, + 700, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our training object is meticulously designed to refine the performance of the VAE, EDDM, and pose decoder within our framework. Initially, the VAE is trained to encode synthetic images, which are environment-free and rendered with poses corresponding to the input images, into latent features. These latent features are then decoded to reconstruct the original synthesized images, ensuring the network's ability to capture the essential features of the object.", + "bbox": [ + 511, + 241, + 906, + 361 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the VAE's training, we proceed to train the EDDM while keeping the VAE's weights fixed. EDDM is trained following the training strategy and objective of DDM [15]:", + "bbox": [ + 511, + 362, + 905, + 407 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e d d m} = \\left\\| d _ {\\theta} - \\left(r _ {0} - g _ {0}\\right) \\right\\| _ {2} + \\left\\| n _ {\\theta} - n \\right\\| _ {2} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 421, + 903, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the first term and second term correspond to the supervision of predicting noise component and image component, respectively.", + "bbox": [ + 511, + 441, + 905, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "At last, we freeze VAE and EDDM to train the pose decoder. We employ a disentangled 6D pose loss via individually supervising the rotation $R$ , the scale-invariant 2D object center $(\\Delta x, \\Delta y)$ and the distance $\\Delta z$ :", + "bbox": [ + 511, + 486, + 905, + 547 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\text {p o s e}} = L _ {R} + L _ {\\text {c e n t e r}} + L _ {z} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 554, + 903, + 570 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Thereby,", + "bbox": [ + 513, + 577, + 576, + 590 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{ \\begin{array}{r c l} L _ {R} & = & \\left\\| R - \\bar {R} \\right\\| _ {2} \\\\ L _ {c e n t e r} & = & \\| \\Delta x - \\Delta \\bar {x} \\| _ {2} + \\| \\Delta y - \\Delta \\bar {y} \\| _ {2} \\\\ L _ {z} & = & \\| \\Delta z - \\Delta \\bar {z} \\| _ {2} \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 595, + 903, + 647 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(\\Delta x, \\Delta y, \\Delta z, R)$ and $(\\Delta \\bar{x}, \\Delta \\bar{y}, \\Delta \\bar{z}, \\bar{R})$ denote prediction and ground truth, respectively.", + "bbox": [ + 511, + 651, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the inference phase, object images from different environments are fed into the conditional encoder as cue information for the EDDM. The EDDM systematically eliminates environmental factors from the noisy latent space, yielding a representation that encapsulates solely the object's information. This purified representation is then fed into the pose decoder to extract the object pose, thereby achieving robust generalization across various scenes and conditions.", + "bbox": [ + 511, + 683, + 905, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 830, + 645, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we conduct experiments to verify the generalization capability of the proposed method in various environments. Therefore we follow existing works [5, 37] in", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8682", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "self-supervised object pose estimation which focus on narrowing the domain gap and enhancing the network's generalization ability across different domains. We compare our method to the state-of-the-art methods on four of the core datasets of the BOP challenge [13] and a self-made dataset. We also provide a detailed analysis of our proposed scheme in ablation study.", + "bbox": [ + 89, + 90, + 483, + 196 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 89, + 207, + 192, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our EA6D on four of the core BOP datasets [13], Linemod(LM) [10], Linemod-Occluded(LM-O) [3], HomebrewedDB(HB) [16] and YCB-V [44]. LM, LM-O, and HB include the same objects under different environments, and YCB-Video involves environmental changes, making them suitable for verifying the proposed method. HB is typically used to study the generalization ability of methods towards new scenes with changing illumination. However, the images in HB are from a single indoor environment. To expand the testing of unseen scenarios, the DiverseScenes Dataset is created by simulating both indoor and outdoor environments, to rigorously evaluate the model's ability to generalize across varied environments.", + "bbox": [ + 89, + 229, + 483, + 425 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "DiverseScenes Dataset. We select four distinct real-world environments, spanning both indoor and outdoor settings, and employ Physically Based Rendering (PBR) techniques to render the 3D models into these backgrounds. To capture the inherent diversity of real-world lighting, we apply a range of lighting conditions, thus amplifying the disparity with the training data and crafting previously unseen test scenarios. For each object category, we render 150 images per environment, amassing a total of 1950 images. Figure 3 illustrates some samples of the dataset. We give more samples of the dataset in the supplementary material and will release the dataset.", + "bbox": [ + 89, + 425, + 483, + 604 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7607c70aac8adbfc83d3a0029a09539cac16c89334eb0284f0754c1111ac35a2.jpg", + "image_caption": [ + "Figure 3. We simulated different indoor and outdoor backgrounds with varying lighting to construct four distinct scenarios, which have a significant difference from the training data." + ], + "image_footnote": [], + "bbox": [ + 91, + 609, + 483, + 790 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "It should be emphasized that EA6D does not require any real images for training. For the evaluation of EA6D, all images in the above dataset are used for testing only.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Metrics", + "text_level": 1, + "bbox": [ + 513, + 90, + 609, + 104 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We report our results referring to the common Average Distance of Distinguishable Model Points (ADD) metric [10]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. The estimated pose is considered to be correct if the average distance is smaller than $10\\%$ of the object's diameter. For symmetric objects, we use the ADD-S metric [44], where the mean distance is computed based on the closest point distance. For YCB-V, we report the AUC (area under curve) of ADD(-S) with a maximum threshold of $10\\mathrm{cm}$ [44].", + "bbox": [ + 511, + 112, + 906, + 280 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 287, + 730, + 304 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method is implemented by Pytorch, and all experiments are conducted on an NVIDIA 4090 GPU. We use Synthetic PBR Dataset [13] to train the 2D object detector Yolov4 [2] and the entire network. For each image, we utilize the object's CAD model to render it in the same pose onto a blank image to obtain the pure object image. We use pure object images to train VAE and use them as ground-truth for corresponding input PBR images to train EDDM from scratch. We train VAE using AdamW optimizer with an attenuated learning rate (from $1e^{-4}$ to $5e^{-7}$ ). We train the diffusion models using AdamW optimizer with an attenuated learning rate (from $2e^{-4}$ to $5e^{-6}$ ) for 20k iterations with a batch size of 12 and each training takes up about 30 GPU hours. The pose decoder is trained with the same optimizer setting for 10k iterations with a batch size of 24 and each training takes about 10 GPU hours. The conditional encoder is pre-trained on ImageNet [6].", + "bbox": [ + 511, + 310, + 908, + 568 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparative Evaluation", + "text_level": 1, + "bbox": [ + 511, + 575, + 738, + 592 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Here, we evaluate the proposed EA6D on the LM, LM-O and YCB-V in comparison to some state-of-the-art methods including six fully-supervised methods that are trained with annotated real images with object poses (DPOD [48], PVNet [28], SO-Pose [8]), ZebraPose [35], CheckerPose [23] and 6D-diff [45], two self-supervised methods that are trained with only synthetic data (AAE [36], MHP [26]), three self-supervised methods that are trained with both synthetic data and unannotated real images (DSC-PoseNet [46], SMOC-Net [37], TexPose [5]), and three self-supervised methods that are trained with synthetic data, unannotated real images and depth images (Self6D [39], Self6D++ [40], RKHSPose [43]). In addition, for a more comprehensive comparison, we also compare our method with Self6D++ [40] and RKHSpose [43] trained only with synthetic data. Table 1 reports the average metrics on each dataset and we provide detailed results in the supplementary material.", + "bbox": [ + 511, + 598, + 908, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparisons on Linemod. As seen from Table 1, there is a significant performance gap between fully-supervised", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8683", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/a8c22942ce88817d8be16c1bf795da4fc21f6b746236a7bc2415f1fc5aaa01eb.jpg", + "table_caption": [ + "Table 1. Comparison with state-of-the-art methods on LM, LM-O and YCB-V. The table reports results for the Average Recall (\\%) of ADD(-S) on LM and LM-O, and AUC of ADD-S and ADD(-S) on YCB-V. All results except ours are copied from SMOC-Net [37], TexPos [5] and RKHSPose [43]. $R$ : annotated real RGB data. $S$ : synthetic RGB data. $R^{-}$ : unannotated real RGB data. $D$ : depth data. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold. (-) denotes results missing from the original paper." + ], + "table_footnote": [], + "table_body": "
MethodsTraining dataLMLM-OYCB-V
ADD(-S)ADD-S AUCADD(-S) AUC
DPOD [48]R+S82.647.3--
PVNet [28]86.340.8-73.8
SO-Pose [8]96.062.390.983.9
ZebraPose [35]-76.990.185.3
CheckerPose [23]-77.591.386.4
6D-diff [45]-79.691.587.0
Self6D [39]S+R−+D58.932.1--
Self6D++ [40]88.564.791.480.0
RKHSPose [43]95.968.792.683.0
DSC-PoseNet [46]S+R−54.721.9--
Self6D++ [40]76.059.8--
SMOC-Net [37]91.363.3--
TexPose [5]91.766.7--
AAE [36]S31.4---
MHP [26]38.8---
Self6D++ [40]77.452.989.477.8
RKHSPose [43]78.254.390.276.5
EA6D96.387.893.188.3
", + "bbox": [ + 94, + 218, + 483, + 484 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "methods and approaches trained with synthetic data instead. Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities (unannotated real RGB data, depth data) to achieve comparable results. The proposed EA6D achieves a better performance than all the comparative methods (regardless of whether these methods are trained in either fully-supervised or self-supervised manner), and only uses synthetic images for training. This is attributed to EA6D generating environment-independent object representations for pose estimation rather than estimating poses directly from input images, avoiding the impact of domain gap on network performance.", + "bbox": [ + 88, + 492, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on Linemod-Occluded. In the inference stage, each object in an image will be inferred independently. The corresponding results are reported in Table 1 which shows that EA6D outperforms all these methods, demonstrating strong robustness against occlusion. Note that EA6D only uses synthetic images as training data and still outperforms fully supervised methods (6D-diff by $8.3\\%$ , CheckerPose by $10.3\\%$ , ZebraPose by $10.9\\%$ ).", + "bbox": [ + 89, + 688, + 482, + 809 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on YCB-V. As shown in Table 1, EA6D achieves the best performance on both the AUC of ADD-S and the AUC of ADD(-S) metrics, showing that EA6D is capable of handling a wide variety of object appearances and lighting conditions.", + "bbox": [ + 89, + 809, + 483, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparisons on Unseen Environments. In this experi", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8ebdd9dbdd030d822b529f54a793e1b7773114396fe5567fa9bdc939ca8f4b2c.jpg", + "image_caption": [ + "Figure 4. Qualitative results on Linemod (top-left), Linemod-Occluded (bottom) and HomebrewedDB (top-right). For each dataset, Top: input images and the visulizations of pose, green and blue bounding boxes represent GT poses and estimated poses, respectively. Middle: Visualization of generated environment-independent object representations. Bottom: The image after decoding the environment-independent object representations." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 300 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/adb1f1773f9583a386ad34504d1e340839dfcbd91481b5979e179d28ffb3fcb7.jpg", + "table_caption": [ + "Table 2. Evaluation results on HomebrewedDB dataset. All results except ours are copied from TexPose [5]. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold." + ], + "table_footnote": [], + "table_body": "
Training dataR− + SS
MethodSock[32]Self6D++[40]Texpose[5]GDR[41]EA6D
Benchvise57.375.792.988.897.6
Driller46.689.494.292.893.8
Phone41.576.881.278.792.6
Mean52.080.689.486.894.7
", + "bbox": [ + 517, + 479, + 906, + 569 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/bab49e680274139263050f773c7020a4a9a17bb3d9b2a44ebb99e158d100df75.jpg", + "table_caption": [ + "Table 3. Evaluation results on DiverseScenes dataset. All cited methods are re-implemented using their open source codes." + ], + "table_footnote": [], + "table_body": "
MethodsSelf6D++ [40]GDR [41]Texpose [5]EA6D
Scene 156.564.182.084.4
Scene 262.270.776.483.9
Scene 348.668.966.583.0
Scene 469.554.364.384.0
Mean59.264.572.383.8
", + "bbox": [ + 517, + 617, + 903, + 712 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment, we want to study the generalizability of our method towards new scenes. First, we evaluate our method on HomebrewedDB. From Table 2, we observe that EA6D outperforms all other methods and achieves the same results as testing in Linemod (95.7% vs.94.7%). Self6D++ and TexPose undergo a clear performance drop when testing on Linemod. Self6D++ drops from 94.1% to 80.6% and TexPose drops from 91.8% to 89.4%. It indicates that the generalization performance of existing self-supervised methods is limited by the unannotated real images used for training. Further, Table 3 shows the results of the test on the DiverseScenes dataset. In environments with significant changes", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8684", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in background and lighting, there is a certain decline in performance for all four methods, but our method is the least affected and outperforms the others by a huge leap.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/024e513529f972ef0c2f00d14fd22be4a5b401b83b9cf5b8f08bbd7248df3bca.jpg", + "image_caption": [ + "Scene 1" + ], + "image_footnote": [], + "bbox": [ + 96, + 147, + 285, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d97c730d4bb633fe2c5fb577054219dc46d1e484ce9a96b43ecad6ca72706461.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 287, + 147, + 478, + 185 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d8fcb9179be1e02562e51c14037d4b92e90c6f8ab3ba100c1a398f84f14bc7ba.jpg", + "image_caption": [ + "Scene 3" + ], + "image_footnote": [], + "bbox": [ + 96, + 196, + 287, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7660107b8d885ac4cde7ff7c3ffe51a78f248b6681f9b2a3c1c10d12e4381ac6.jpg", + "image_caption": [ + "Scene 4" + ], + "image_footnote": [], + "bbox": [ + 289, + 196, + 478, + 234 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/966e391bde2b3ad4051238f8728b4766afbaed64c69fcb86a495be0f2a82c9a8.jpg", + "image_caption": [ + "Same object in 7 different environments", + "Figure 5. The top two rows are qualitative results on DiverseScenes Dataset. Each scene has different backgrounds and lighting conditions. The bottom row is qualitative results of a same object in 7 different environments." + ], + "image_footnote": [], + "bbox": [ + 96, + 247, + 207, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/d3fe39d75992438083932905c4ba4743681b170ab3dfa80eb1c5f8fe6c853f79.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 247, + 341, + 290 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/874bed435b455bed5ce92be74c83876a00fb293ed0441a1d1d321c9f6c1296c8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 341, + 247, + 478, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We would like to emphasise that since our method is only trained on Synthetic PBR Dataset [13], the aforementioned datasets are all unseen environments for our method. Our method maintains accurate predictions (average $90.6\\%$ ) under different backgrounds, lighting conditions, and occlusions. It shows that the choice of adaptation domain (LM, HB, or other environments) has a marginal effect on our final performance. We attribute this to the generation of pure object features, eliminating the domain gap problem caused by environmental factors.", + "bbox": [ + 89, + 359, + 483, + 510 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Visulizations", + "text_level": 1, + "bbox": [ + 89, + 517, + 225, + 532 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this section, we will analyze our work by visualization of the test results. Figure 4 illustrates the visualizations of environment-independent object representations and pose estimation. It clearly shows that the images generated by the diffusion model filter out all environmental factors while accurately preserving the object information. The visualizations of results on LM-O explain why EA6D can effectively handle occlusions. With the strong generative capabilities of diffusion models, even when facing severe occlusion, the model is still able to accurately generate occluded parts. Figure 5 shows the visualizations of the results on the DiverseScenes Dataset and the visualizations of the same object in 7 different environments. These provide a clear demonstration of its capabilities in handling various environments. We give more visulizations results in supplementary material.", + "bbox": [ + 89, + 539, + 483, + 781 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 787, + 254, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of diffusion models and diffusion steps. We study the impact of different diffusion models on EA6D performance. We replace EDDM with traditional Denoising Diffusion Probabilistic Model (DDPM) [11] and Decoupled Diffusion Model (DDM) [15] and compare the results on Linemod. Also, the number of iterating steps is", + "bbox": [ + 89, + 809, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "another key parameter in diffusion models that determines the inference speed, which is essential for pose estimation. All the results are reported in Table 4. We can observe that the EA6D with EDDM achieves more accurate results with fewer iterating steps. This result verifies the effectiveness of mixing of input image latent representations and noise instead of pure noise in generation process, which provides prior information, thus speeding up the inference procedure and improving the generation quality.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/670720be74df8c225bbb4b30cd575c5cd34f866fd3290128ec97dae3610030e8.jpg", + "table_caption": [ + "Table 4. Mean results with ablation EA6D and the number of iterating steps on Linemod." + ], + "table_footnote": [], + "table_body": "
Method\\steps1510201001000
EA6D-DDPM [11]003.614.352.778.4
EA6D-DDM [15]93.196.195.696.396.296.3
EA6D w/o VAE68.471.373.175.176.277.4
EA6D*76.779.279.279.479.379.5
EA6D96.396.396.496.596.596.5
Time(s)EA6D0.230.510.871.606.9892.1
", + "bbox": [ + 532, + 265, + 883, + 400 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The effect of object representations. In our framework, we generate the object representations via performing the denoising process to estimate object pose. The object representations are the latent code of pure object images encoded by VAE. To evaluate the efficacy of such representations, we first remove the VAE and directly generate the pure object images as object representations through EDDM. Then we use the framework of EA6D, but decode the generated object representations to pure object images using the decoder of VAE and input pose decoder to estimate pose. We denote this setting as EA6D*. As shown in Table 4, the performance of these two types of object representations significantly drops, showing that the latent code of pure object images captures the rich semantic and structural information in the image and is more suitable for pose estimation than the generated images.", + "bbox": [ + 511, + 409, + 906, + 650 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 664, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a novel generative paradigm for 6D pose estimation. Different from typical RGB-based pose estimation frameworks which directly extract features from input images for pose estimation and are therefore easily influenced by environmental factors, we employ an environment decoupling model that separates the object from environmental factors during the denoising process, generating environment-independent object representations which are robust to environmental changes. A pose decoder is designed to estimate poses using these representations. The experimental results on five datasets demonstrate that EA6D significantly outperforms current state-of-the-art self-supervised methods, particularly in its ability to generalize to previously unseen scenarios.", + "bbox": [ + 511, + 689, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8685", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the National Key Research and Development Program of China (No.2024YFF0907600), the National Natural Science Foundation of China (Grant No.62273275), the China Postdoctoral Science Foundation (No.2022M722573), Shaanxi Postdoctoral Research Project (No.2023BSHEDZZ244), Shaanxi Natural Science Basic Research Program (No.2025JC-YBQN-928), Key Research and Development Plan of Shaanxi Province (2025GH-YBXM-018) and Xi'an Science and Technology Innovation and Qinchuangyuan Innovation Major Program (23ZDCYJSGG0009-2023).", + "bbox": [ + 89, + 114, + 483, + 268 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 297, + 187, + 313 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2", + "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5, 6", + "[3] E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 1, 2, 6", + "[4] E. Brachmann, F. Michel, A. Krull, M.Y. Yang, S. Gumhold, and C. Rother. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In CVPR, 2016. 2", + "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In CVPR, 2023. 2, 5, 6, 7", + "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 3", + "[8] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting self-occlusion for direct 6d pose estimation. In ICCV, 2021. 6, 7", + "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 3", + "[10] S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, and N. Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In ACCV, 2012. 1, 6", + "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 8", + "[12] Tomáš Hodan, P. Haluza, Štepan Obdržálek, J. Matas, M. Lourakis, and X. Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In WACV, 2017. 1" + ], + "bbox": [ + 93, + 323, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Tomás Hodař, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In ECCV Workshops, 2020. 6, 8", + "[14] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, 2020. 2", + "[15] Yuhang Huang, Zheng Qin, Xinwang Liu, and Kai Xu. Decoupled diffusion models with explicit transition probability. arXiv preprint arXiv:2306.13720, 2023. 3, 4, 5, 8", + "[16] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCV Workshops, 2019. 1, 6", + "[17] W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 1, 2", + "[18] DiederikP. KingmaandMaxWelling. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[19] V. Lepetit, F Moreno-Noguer, and P. Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 81(2):155-166, 2009. 2", + "[20] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, 2023. 1, 2", + "[21] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In ECCV, 2022. 2", + "[22] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, 2019. 1, 2, 5", + "[23] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In ICCV, 2023. 1, 2, 6, 7", + "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5", + "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 3", + "[26] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In ICCV, 2019. 2, 6, 7", + "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, 2019. 2", + "[28] S. Peng, Y. Liu, Q. Huang, X. Zhou, and H. Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2, 6, 7", + "[29] M. Rad and V. Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8686", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4", + "[31] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. TPAMI, 2021. 1", + "[32] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In 3DV, 2020. 7", + "[33] Chen Song, Jiaru Song, and Qixing Huang. Hybridpose: 6d object pose estimation under hybrid representations. In CVPR, 2020. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3", + "[35] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In CVPR, 2022. 1, 2, 6, 7", + "[36] M. Sundermeyer, Z. Marton, M. Durner, M. Brucker, and R. Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In ECCV, 2018. 2, 6, 7", + "[37] Tao Tan and Qiulei Dong. Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In CVPR, 2023. 1, 2, 5, 6, 7", + "[38] B. Tekin, S.N. Sinha, and P. Fua. Real-time seamless single shot 6d object pose prediction. In CVPR, 2018. 2", + "[39] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, 2020. 1, 2, 6, 7", + "[40] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6d object pose estimation. TPAMI, 2021. 1, 2, 6, 7", + "[41] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1, 2, 7", + "[42] Tianfu Wang, Guosheng Hu, and Hongguang Wang. Object pose estimation via the aggregation of diffusion features. In CVPR, 2024. 3", + "[43] Yangzheng Wu and Michael Greenspan. Pseudo-keypoint rkhs learning for self-supervised 6dof pose estimation. In ECCV, 2024. 2, 6, 7", + "[44] Y. Xiang, T. Schmidt, V. Narayanan, and D. Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1, 2, 6", + "[45] Li Xu, Haoxuan Qu, Yujun Cai, and Jun Liu. 6d-diff: A keypoint diffusion framework for 6d object pose estimation. In CVPR, 2024. 2, 3, 6, 7", + "[46] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In CVPR, 2021. 1, 2, 6, 7", + "[47] Yunfan Ye, Kai Xu, Yuhang Huang, Renjiao Yi, and Zhiping Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. In AAAI, 2024. 3" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] S. Zakharov, I. Shugurov, and S. Ilic. Dpod: 6d pose object detector and refiner. In ICCV, 2019. 2, 6, 7", + "[49] Shaobo Zhang, Wanqing Zhao, Ziyu Guan, Xianlin Peng, and Jinye Peng. Keypoint-graph-driven learning framework for object pose estimation. In CVPR, 2021. 2" + ], + "bbox": [ + 514, + 92, + 903, + 162 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8687", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_model.json b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_model.json new file mode 100644 index 0000000000000000000000000000000000000000..78900a7ca943ca8f2e82af6adf6b8496c4ef9f1a --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_model.json @@ -0,0 +1,2167 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation" + }, + { + "type": "text", + "bbox": [ + 0.184, + 0.203, + 0.31, + 0.223 + ], + "angle": 0, + "content": "Shaobo Zhang" + }, + { + "type": "text", + "bbox": [ + 0.355, + 0.203, + 0.483, + 0.222 + ], + "angle": 0, + "content": "Yuhang Huang" + }, + { + "type": "text", + "bbox": [ + 0.388, + 0.223, + 0.483, + 0.239 + ], + "angle": 0, + "content": "Ziyu Guan" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.204, + 0.66, + 0.222 + ], + "angle": 0, + "content": "Wanqing Zhao\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.53, + 0.223, + 0.632, + 0.24 + ], + "angle": 0, + "content": "Jinye Peng\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.699, + 0.204, + 0.785, + 0.222 + ], + "angle": 0, + "content": "Wei Zhao3" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.239, + 0.304, + 0.257 + ], + "angle": 0, + "content": "\\(^{1}\\)Northwest University" + }, + { + "type": "text", + "bbox": [ + 0.342, + 0.24, + 0.694, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)National University of Defense Technology" + }, + { + "type": "text", + "bbox": [ + 0.733, + 0.239, + 0.884, + 0.258 + ], + "angle": 0, + "content": "\\(^{3}\\)Xidian University" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.26, + 0.914, + 0.275 + ], + "angle": 0, + "content": "{zhangshaobo, zhaowq, ziyuguan, pjy}@nwu.edu.cn, huangai@nudt.edu.cn, ywzhao@mail.xidian.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.485, + 0.689 + ], + "angle": 0, + "content": "This paper introduces EA6D, a novel diffusion-based framework for 6D pose estimation that operates effectively in any environment. Traditional pose estimation methods struggle with the variability and complexity of real-world scenarios, often leading to overfitting on controlled datasets and poor generalization to new scenes. To address these challenges, we propose a generative pose estimation paradigm that generates environment-independent object representations for pose estimation, which are robust to environmental variations such as illumination, occlusion, and background clutter. Specifically, we propose the novel Environment Decoupling Diffusion Model (EDDM) which separates object representations from environmental factors while enabling efficient few-step sampling by leveraging input image priors instead of pure noise initialization. We validate our approach on four standard benchmarks and a self-made dataset DiverseScenes. The results demonstrate that EA6D, trained using only synthetic data, can outperform the state-of-the-art methods with both synthetic and realistic data. In particular, for fair comparisons with synthetic data, we can exceed the previous SOTA by \\(18.1\\%\\) and \\(33.5\\%\\) on Linemod and Linemod-Occluded datasets respectively. Project page: https://github.com/acf22/EA6D" + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.718, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.848 + ], + "angle": 0, + "content": "6D Pose Estimation is a pivotal task in the field of computer vision and robotics, which aims to determine the exact position and orientation of an object in three-dimensional space relative to a camera or a coordinate system. The ability to accurately perceive an object's pose is fundamental for a wide range of applications, including robotic manipulation, augmented reality and autonomous navigation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.849, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Recently, a series of deep-learning-based object 6D pose estimation methods [17, 22, 23, 28, 31, 35, 41, 44] have" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.306, + 0.903, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.582 + ], + "angle": 0, + "content": "Figure 1. Different from typical RGB-based pose estimation frameworks, which are easily influenced by environmental factors leading to inaccurate pose estimation, we propose a generative paradigm that leverages diffusion models to generate environment-independent object representations for pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.907, + 0.748 + ], + "angle": 0, + "content": "been proposed. These methods achieve state-of-the-art results, but require large annotated datasets. Generating accurate annotations for pose data is a labor-intensive and time-consuming process, leading most deep-learning-based 6D pose estimation methods to be trained and tested on closed datasets. These datasets [3, 10, 12, 16] typically contain only a few hundred images with limited variation and controlled factors, such as object arrangement, lighting conditions, and background, resulting in overfitting to the training set and the struggle to generalize to new scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Consequently, synthetic datasets have been employed as an additional complement to real datasets, or alternative to real datasets, as they can be easily and automatically annotated in simulated environments. The primary challenge with synthetic data is the domain discrepancy between real and synthetic datasets, which often leads to reduced accuracy when applied to real-world scenarios. To alleviate this problem, some self-supervised methods [20, 37, 39, 40, 46] are proposed, which leverage synthetic images with pose annotations and unannotated real images during training." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.24, + 0.901 + ], + "angle": 0, + "content": "*Corresponding authors" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8678" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "These methods harness intrinsic structures to generate supervision signals and implement constraints on the discrepancies between real and synthetic images, thereby mitigating the domain gap." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.155, + 0.483, + 0.47 + ], + "angle": 0, + "content": "Practical applications such as AR and autonomous driving, which need to operate in uncontrolled environments, make 6D pose estimation more challenging. Here, object appearances vary due to fluctuating environmental conditions, background clutter, and occlusions, which often exceed the variability present in training datasets. Previous methods usually use CNNs as image encoder to extract features for pose estimation (as shown on the top side of Figure 1). These features are highly sensitive to environmental factors. When the environment changes, CNNs struggle to accurately extract object features. This hinders the model's ability to generalize to new scenes, potentially diminishing the accuracy of pose estimation and even leading to object pose estimation failure. While self-supervised methods are proficient at leveraging unlabeled data, they may not fully account for the extensive real-world variability found in these uncontrolled environments, leading to models that are not sufficiently robust for novel conditions. Therefore, there is an urgent need for approaches that can adapt to a variety of dynamic and diverse environments, ensuring reliable pose estimation capabilities." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.473, + 0.483, + 0.849 + ], + "angle": 0, + "content": "In this paper, we introduce a novel paradigm for pose estimation in uncertain environments, including unseen scenes, varying lighting, occlusion, and other possible situations. We consider a generative strategy that generates environment-independent object representations for pose estimation (bottom side of Figure 1). Diffusion models [7, 11, 34] have recently achieved promising results in various generative tasks, including image generating and editing. These models recover high-quality, deterministic samples from noisy and some conditions through a progressive denoising procedure. Motivated by the above, we design an Environment-Agnostic object 6D POSE estimation framework (named EA6D) that exploits the diffusion model to generate environment-independent representations in latent space for pose estimation, while input images serve as conditional cues, illustrated on the bottom side of Figure 1. Specifically, we propose an Environment Decoupling Diffusion Model (EDDM) which employs a novel environment decoupling diffusion process to intrinsically decouple the environment-independent object representations from input images. By performing validation on four popular benchmarks and a self-made dataset, the results show that our method significantly outperforms state-of-the-art self-supervised methods and provides a significant improvement in generalization to unseen scenarios." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.851, + 0.456, + 0.866 + ], + "angle": 0, + "content": "In sum, the main contributions in this paper include:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We propose a new paradigm for environment-agnostic object 6D pose estimation, by generating environment" + }, + { + "type": "text", + "bbox": [ + 0.527, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "independent object representations rather than estimating poses directly from input images." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.197 + ], + "angle": 0, + "content": "- We design a novel Environment Decoupling Diffusion Model which employs an explicit environmental decoupling mechanism via gradient modelling and conditional generation to separate object representations from environment, while enabling efficient several-step sampling." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.198, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- We conduct extensive experiments on four public datasets including Linemod, Linemod-Occluded, YCB-V, HomebrewedDB and a self-made dataset DiverseScenes, showing that our method surpasses current state-of-the-art self-supervised methods. Our method exhibits remarkable generalization capabilities, even in previously unseen scenes, which is crucial for practical applications." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.122, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.317, + 0.651, + 0.332 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.905, + 0.718 + ], + "angle": 0, + "content": "Fully-Supervised 6D Object Pose Estimation. Early approaches rely on local or global features to establish correspondences between keypoints on CAD models [1, 3, 4]. More recently, learning-based methods, particularly those utilizing Convolutional Neural Networks (CNNs), have become dominant. These methods extract deep features and establish mappings from RGB images to 6D object pose [14, 17, 41, 44]. Due to the inherent nonlinearity of the rotation space, the accuracy of directly regressed object pose is often limited [21]. To enhance accuracy, many methods employ CNNs to establish 2D-3D correspondences, followed by solving the object pose using the Perspective-n-Point (PnP) algorithm [19]. These methods can be categorized into sparse and dense correspondence methods. Sparse methods [28, 33, 38, 45] predict the 2D projections of predefined 3D keypoints, while dense methods [22, 23, 27, 35, 41, 48] predict 3D coordinates on the object model corresponding to image pixels thus providing better robustness to occlusions and clutter. However, fully-supervised 6D pose estimation methods are constrained by the need for accurate and diverse training data, which is often scarce due to the labor-intensive nature of precise annotation and the limited variation in available datasets, resulting in poor generalization to real-world and uncontrolled environments." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.72, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Self-Supervised 6D Object Pose Estimation. Considering the tremendous effort to collect large amounts of annotations for 6D object pose, some methods [26, 29, 36] have turned to using synthetic images for training. However, because of the domain gap between real and synthetic images, the performances of these methods are significantly lower compared to the methods that use real images with object poses for training. To narrow the domain gap, a few recent self-supervised methods [5, 20, 37, 39, 40, 43, 46, 49] have begun to incorporate both synthetic images with known object poses and unannotated real images into their training datasets. However, in diverse real-world application sce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8679" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.272 + ], + "angle": 0, + "content": "narios, models must contend with a variety of backgrounds, lighting conditions, and other environmental factors that are difficult to be adequately included in the training dataset. These environmental factors can significantly alter the appearance of objects within images. As a result, features learned by the model for the same object may vary across different scenes, leading to diminished generalization capabilities when encountering new scenes. Different from self-supervised learning methods, our proposed method directly generates environment-independent object representations to fundamentally eliminate the impact of environmental disturbances." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.274, + 0.487, + 0.546 + ], + "angle": 0, + "content": "Diffusion-based Models for Vision Tasks. Originally introduced for image generation, diffusion models [7, 11, 34] have demonstrated impressive generation capabilities and have been explored in various tasks, including edge detection [47], image inpainting [25], and 6D pose estimation [9, 42, 45]. DiffusionEdge [47] presents a novel diffusion-based edge detector capable of generating precise and sharp edge maps without post-processing. 6D-diff [45] formulates object keypoints detection as a reverse diffusion process and employs a MoC-based forward process to leverage distribution priors in intermediate representations. Method [42] introduces three novel aggregation networks designed to effectively aggregate diffusion features, showing superior generalizability for object pose estimation. Currently, no existing work has yet explored the potential of diffusion models to generate environment-independent object representations, which could significantly improve the robustness of pose estimation in diverse environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.564, + 0.182, + 0.579 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.59, + 0.201, + 0.604 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.483, + 0.853 + ], + "angle": 0, + "content": "Traditional pose estimation methods typically estimate pose directly from images. Environmental factors such as background clutter, illumination variations, and occlusion in images can limit the accuracy and robustness of pose estimation. Pursuing a different paradigm, we propose a diffusion-based method that generates environment-independent object representations for pose estimation. Traditional diffusion models are predominantly utilized for image generation tasks, where they excel at creating new synthetic images. These models operate on the principle of gradually adding noise to an input image and then learning to reverse this process, effectively denoising the image over multiple iterations to recover the original content. In this paper, we extend the concept of diffusion models to a completely new domain by recovering environment-independent object representations for pose estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The overall framework of the proposed method EA6D is shown in Figure 2. The first stage is environment-independent object representation encoding. A synthesized" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.485 + ], + "angle": 0, + "content": "pure object image with the same pose as the input image, but free from background, lighting, and occlusion, is encoded into a latent space using a Variational Autoencoder (VAE) [18]. This encoding results in an object representation \\( g_0 \\in R^{64 \\times 64 \\times 3} \\), which can be decoded back to the original pure object image by the decoder of VAE network, demonstrating that the representation contains only information about the object independent of the environment. Following this, a forward diffusion process is applied to \\( g_0 \\) to generate a noisy representation \\( g_t \\), which can be split into two sub-processes: the clean object representation \\( g_0 \\) to input image latent representation \\( r_0 \\), meanwhile Gaussian noise is incrementally added. To denoise and recover the environment-independent object representations from \\( g_t \\), an Environment Decoupling Diffusion Model (EDDM) is introduced. A conditional encoder extracts multi-level features from the input images. These features are then fused with the features of the U-Net encoder and are input to U-Net decoders. The EDDM outputs a decoupled object representation, which potentially contains object pose information, independent of environmental factors. Finally, a pose decoder predicts the pose information, including rotation quaternion and translation vector. The inference process in EA6D is thus a combination of diffusion-based denoising and pose decoding, enabling robust and accurate pose estimation in diverse and complex environments." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.5, + 0.907, + 0.533 + ], + "angle": 0, + "content": "3.2. Environment Decoupling Diffusion for Generating Object Representations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Current 6D pose estimation methods often face robustness challenges in dynamic or unseen environments, as their object representations are influenced by complex environmental variations, resulting in multimodal distributions of features. CNNs-based methods typically use a single model to fit the entire data distribution. When facing multimodal distributions, the single model struggles to capture all the different patterns simultaneously, thereby resulting in a notable decrease in performance in new environments. To avoid the negative impact, we propose to extract environment-independent representations, enabling more accurate and generalizable pose estimation. Recent diffusion models show remarkable capabilities for modeling complex multimodal distribution, which is helpful for learning about environment-independent representations. Therefore, we propose to construct our framework based on the diffusion models to cover the complex distribution of object representations. More importantly, to acquire the environment-independent component from the input image distribution, we propose a novel environment decoupling diffusion model for generating environment-independent object representations inspired by the decoupled diffusion models [15, 47]. Different from the vanilla decoupled diffusion process that focuses on splitting the image and the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8680" + } + ], + [ + { + "type": "header", + "bbox": [ + 0.1, + 0.09, + 0.221, + 0.104 + ], + "angle": 0, + "content": "- - - - → only training" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.104, + 0.898, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.326, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed method. In training phase, the proposed method involves training a Variational Autoencoder (VAE) to encode pure object images into environment-independent representations. These representations are then sampled to create noisy representations, which are processed by an Environment Decoupling Diffusion Model (EDDM). The EDDM uses a dual structure with one UNet encoder and two UNet decoders to decouple the environment-independent object representations from the input image with complex environments. The pose decoder is trained to predict the pose from the representations, enabling robust pose estimation across various environments. In testing phase, a random Gaussian noise is fed into the EDDM without VAE." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.419, + 0.482, + 0.48 + ], + "angle": 0, + "content": "noise components, we aim to decouple the environment-independent object representations from the input image with complex environments. To this end, we introduce the improved forward and reversed diffusion processes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.485, + 0.255, + 0.499 + ], + "angle": 0, + "content": "3.2.1.Forward Process" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.504, + 0.484, + 0.747 + ], + "angle": 0, + "content": "The vanilla decoupled diffusion process reformulates the conventional image-to-noise process into two stages: image-to-zero and zero-to-noise [15]. Corresponding to current object representation generation, there will be a clean object representation to zero process and a zero to noise process. However, such a paradigm only describes how to learn the environment-independent object representation from noise, without a clue on how to decouple it from the input images with complex environments. To address this problem, this paper proposes an environment decoupling diffusion process to intrinsically decouple the environment-independent object representation from the input image. We follow [30] to conduct the diffusion process in the latent space, i.e., our goal is to obtain the environment-independent object latent representations from the input image that mixes the object with the complex environments:" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.767, + 0.483, + 0.802 + ], + "angle": 0, + "content": "\\[\nq \\left(g _ {t} \\mid g _ {0}\\right) = \\mathcal {N} \\left(g _ {t}; g _ {0} + \\int_ {0} ^ {t} \\left(r _ {0} - g _ {0}\\right) d t, t I\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.809, + 0.484, + 0.904 + ], + "angle": 0, + "content": "where \\(g_{0}\\) and \\(g_{t}\\) denote the clean and noisy object latent representations respectively, \\(t\\in [0,1]\\) is the time step, \\(r_0\\) denotes the latent representation of input image. (A Variational Autoencoder features the latent representation.) According to the above equation, we can sample \\(g_{t}\\) by \\(g_{t} = g_{0} + \\int_{0}^{t}(r_{0} - g_{0})dt + \\sqrt{t} n,\\) where \\(n\\in \\mathcal{N}(0,I)\\). Ob" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.907, + 0.603 + ], + "angle": 0, + "content": "viously, the forward process describes two sub-processes: \\( g_0 + \\int_0^t (r_0 - g_0)dt \\) is the clean object representation to input image latent representation and \\( \\sqrt{t} n \\) is the zero to noise. It is worth noting that the first term contains the decay gradient from the input image to the environment-independent object representation, i.e., \\( r_0 - g_0 \\). We call this term the environment decoupling gradient, which guides the generation from the input image representation to object representation. The environment decoupling gradient provides explicit evidence for our goal of decoupling the environment-independent object representations, allowing for accurate and efficient generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.681, + 0.625 + ], + "angle": 0, + "content": "3.2.2. Reversed Process" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.906, + 0.691 + ], + "angle": 0, + "content": "The goal of the reversed process is to gradually recover the clean object representation \\( g_{0} \\) from the noisy input \\( g_{1} \\). Following the derivation of [15], we can formally represent the reversed process as a conditional probability:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.701, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} q \\left(g _ {t - \\triangle t} \\mid g _ {t}, g _ {0}\\right) = \\mathcal {N} \\left(g _ {t} + \\int_ {t} ^ {t - \\triangle t} \\left(r _ {0} - g _ {0}\\right) d t - \\frac {\\triangle t}{\\sqrt {t}} n, \\right. \\\\ \\frac {\\triangle t (t - \\triangle t)}{t} I), \\tag {2} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\triangle t\\) is the smallest time step. Inheriting DDM, the reversed process benefits from the analytic gradient \\(r_0 - g_0\\), enabling approximate analytic solution for few-step generation (even with only one step). Different from general diffusion models that recover the goal content from a normal distribution, our reversed process aims to generate object representations from the combination of the latent image representations and the noise. We can easily obtain the ini" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8681" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.484, + 0.258 + ], + "angle": 0, + "content": "tial variable in the reversed process according to the forward process, i.e., \\( g_{1} = r_{0} + n \\). Intuitively, the generation process starts with the mixture of \\( r_0 \\) and \\( n \\) instead of pure noise, which provides important prior information, thus speeding up the inference procedure and improving the generation quality. The reversed conditional probability \\( q(g_{t - \\triangle t}|g_t,g_0) \\) includes \\( r_0 - g_0 \\) and \\( n \\), which are unknown in the reversed process. Therefore, we need to use a neural network \\( \\theta \\) to parameterize \\( q(g_{t - \\triangle t}|g_t,g_0) \\). Practically, the neural network has two outputs: \\( d_{\\theta} \\) and \\( n_{\\theta} \\), corresponding to the two unknown terms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.264, + 0.292, + 0.278 + ], + "angle": 0, + "content": "3.2.3. Network Architecture" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.283, + 0.484, + 0.419 + ], + "angle": 0, + "content": "As shown in Figure 2, we construct our network based on a U-Net architecture including one encoder and two decoders. A pre-trained Swin Transformer [24] serves as a conditional encoder, extracting multi-level features from the input images. These features are then fused with the features of the U-Net encoder at corresponding levels through attention mechanisms, acting as inputs for the U-Net decoders. In practice, the input images will be obtained by an off-the-shelf object detector such as Yolo-v4 [2]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.442, + 0.444 + ], + "angle": 0, + "content": "3.3. Pose Decoder for Object Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.675 + ], + "angle": 0, + "content": "The pose decoder is designed to predict the 6D pose, including both the rotation and translation, from the object representations generated from EDDM. The decoder consists of three modules: a ResNet block to further extract the features of \\(\\hat{g}_0^t\\), a rotation head for rotation estimation, and a translation head for translation prediction. The ResNet block enhances feature extraction through three convolutional layers. The first layer employs a \\(1 \\times 1\\) kernel to expand the feature dimension, capturing richer information. The second layer uses a \\(3 \\times 3\\) kernel for spatial feature reduction. The third layer also uses a \\(1 \\times 1\\) kernel. The rotation head is designed with a series of \\(3 \\times 3\\) convolutional layers, followed by a generalized average pooling layer and a fully connected layer, culminating in the prediction of the rotation quaternion \\(R\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.676, + 0.484, + 0.843 + ], + "angle": 0, + "content": "For 3D translation, since directly regressing the translation in 3D space does not work well in practice, we adopt Scale-invariant Translation Estimation (SITE) [22] to decouple the translation into the 2D location \\((o_x,o_y)\\) of the projected 3D centroid and the object's distance \\(T_{z}\\) towards the camera. The translation head is constituted by a series of convolutional layers followed by fully connected layers, which predict the scale-invariant translation parameters \\(T_{site} = (\\Delta x,\\Delta y,\\Delta z)\\). These parameters are then utilized to compute the object's 2D coordinates and depth, as delineated by the equations:" + }, + { + "type": "equation", + "bbox": [ + 0.211, + 0.85, + 0.484, + 0.899 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{l l} o _ {x} = & \\Delta x \\cdot w + c _ {x} \\\\ o _ {y} = & \\Delta y \\cdot h + c _ {y} \\\\ T _ {z} = & \\Delta z \\cdot r \\end{array} \\right. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.213 + ], + "angle": 0, + "content": "where \\( w \\) and \\( h \\) represent the width and height of the bounding box, \\( c_{x} \\) and \\( c_{y} \\) denote the coordinates of the box's center, and \\( r \\) is the scaling factor derived from the image's dimensions. Given the camera intrinsics \\( k \\), the translation \\( T \\) can be calculated via back-projection \\( T = K^{-1}T_{z}[o_{x}, o_{y}, 1]^{T} \\). This method allows for a more accurate and robust estimation of the object's 3D translation, thereby enhancing the overall pose estimation accuracy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.22, + 0.702, + 0.236 + ], + "angle": 0, + "content": "3.4. Training Procedure" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.242, + 0.907, + 0.362 + ], + "angle": 0, + "content": "Our training object is meticulously designed to refine the performance of the VAE, EDDM, and pose decoder within our framework. Initially, the VAE is trained to encode synthetic images, which are environment-free and rendered with poses corresponding to the input images, into latent features. These latent features are then decoded to reconstruct the original synthesized images, ensuring the network's ability to capture the essential features of the object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.363, + 0.906, + 0.409 + ], + "angle": 0, + "content": "After the VAE's training, we proceed to train the EDDM while keeping the VAE's weights fixed. EDDM is trained following the training strategy and objective of DDM [15]:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.422, + 0.905, + 0.44 + ], + "angle": 0, + "content": "\\[\nL _ {e d d m} = \\left\\| d _ {\\theta} - \\left(r _ {0} - g _ {0}\\right) \\right\\| _ {2} + \\left\\| n _ {\\theta} - n \\right\\| _ {2} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.443, + 0.906, + 0.487 + ], + "angle": 0, + "content": "where the first term and second term correspond to the supervision of predicting noise component and image component, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.906, + 0.548 + ], + "angle": 0, + "content": "At last, we freeze VAE and EDDM to train the pose decoder. We employ a disentangled 6D pose loss via individually supervising the rotation \\( R \\), the scale-invariant 2D object center \\( (\\Delta x, \\Delta y) \\) and the distance \\( \\Delta z \\):" + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.555, + 0.905, + 0.571 + ], + "angle": 0, + "content": "\\[\nL _ {\\text {p o s e}} = L _ {R} + L _ {\\text {c e n t e r}} + L _ {z} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.577, + 0.592 + ], + "angle": 0, + "content": "Thereby," + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.597, + 0.905, + 0.648 + ], + "angle": 0, + "content": "\\[\n\\left\\{ \\begin{array}{r c l} L _ {R} & = & \\left\\| R - \\bar {R} \\right\\| _ {2} \\\\ L _ {c e n t e r} & = & \\| \\Delta x - \\Delta \\bar {x} \\| _ {2} + \\| \\Delta y - \\Delta \\bar {y} \\| _ {2} \\\\ L _ {z} & = & \\| \\Delta z - \\Delta \\bar {z} \\| _ {2} \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.906, + 0.683 + ], + "angle": 0, + "content": "where \\((\\Delta x, \\Delta y, \\Delta z, R)\\) and \\((\\Delta \\bar{x}, \\Delta \\bar{y}, \\Delta \\bar{z}, \\bar{R})\\) denote prediction and ground truth, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.818 + ], + "angle": 0, + "content": "In the inference phase, object images from different environments are fed into the conditional encoder as cue information for the EDDM. The EDDM systematically eliminates environmental factors from the noisy latent space, yielding a representation that encapsulates solely the object's information. This purified representation is then fed into the pose decoder to extract the object pose, thereby achieving robust generalization across various scenes and conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.646, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this section, we conduct experiments to verify the generalization capability of the proposed method in various environments. Therefore we follow existing works [5, 37] in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8682" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.198 + ], + "angle": 0, + "content": "self-supervised object pose estimation which focus on narrowing the domain gap and enhancing the network's generalization ability across different domains. We compare our method to the state-of-the-art methods on four of the core datasets of the BOP challenge [13] and a self-made dataset. We also provide a detailed analysis of our proposed scheme in ablation study." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.208, + 0.194, + 0.222 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.23, + 0.485, + 0.426 + ], + "angle": 0, + "content": "We evaluate our EA6D on four of the core BOP datasets [13], Linemod(LM) [10], Linemod-Occluded(LM-O) [3], HomebrewedDB(HB) [16] and YCB-V [44]. LM, LM-O, and HB include the same objects under different environments, and YCB-Video involves environmental changes, making them suitable for verifying the proposed method. HB is typically used to study the generalization ability of methods towards new scenes with changing illumination. However, the images in HB are from a single indoor environment. To expand the testing of unseen scenarios, the DiverseScenes Dataset is created by simulating both indoor and outdoor environments, to rigorously evaluate the model's ability to generalize across varied environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.606 + ], + "angle": 0, + "content": "DiverseScenes Dataset. We select four distinct real-world environments, spanning both indoor and outdoor settings, and employ Physically Based Rendering (PBR) techniques to render the 3D models into these backgrounds. To capture the inherent diversity of real-world lighting, we apply a range of lighting conditions, thus amplifying the disparity with the training data and crafting previously unseen test scenarios. For each object category, we render 150 images per environment, amassing a total of 1950 images. Figure 3 illustrates some samples of the dataset. We give more samples of the dataset in the supplementary material and will release the dataset." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.611, + 0.484, + 0.791 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.805, + 0.483, + 0.847 + ], + "angle": 0, + "content": "Figure 3. We simulated different indoor and outdoor backgrounds with varying lighting to construct four distinct scenarios, which have a significant difference from the training data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "It should be emphasized that EA6D does not require any real images for training. For the evaluation of EA6D, all images in the above dataset are used for testing only." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.61, + 0.105 + ], + "angle": 0, + "content": "4.2. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.281 + ], + "angle": 0, + "content": "We report our results referring to the common Average Distance of Distinguishable Model Points (ADD) metric [10]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. The estimated pose is considered to be correct if the average distance is smaller than \\(10\\%\\) of the object's diameter. For symmetric objects, we use the ADD-S metric [44], where the mean distance is computed based on the closest point distance. For YCB-V, we report the AUC (area under curve) of ADD(-S) with a maximum threshold of \\(10\\mathrm{cm}\\) [44]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.731, + 0.305 + ], + "angle": 0, + "content": "4.3. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.311, + 0.909, + 0.569 + ], + "angle": 0, + "content": "Our method is implemented by Pytorch, and all experiments are conducted on an NVIDIA 4090 GPU. We use Synthetic PBR Dataset [13] to train the 2D object detector Yolov4 [2] and the entire network. For each image, we utilize the object's CAD model to render it in the same pose onto a blank image to obtain the pure object image. We use pure object images to train VAE and use them as ground-truth for corresponding input PBR images to train EDDM from scratch. We train VAE using AdamW optimizer with an attenuated learning rate (from \\(1e^{-4}\\) to \\(5e^{-7}\\)). We train the diffusion models using AdamW optimizer with an attenuated learning rate (from \\(2e^{-4}\\) to \\(5e^{-6}\\)) for 20k iterations with a batch size of 12 and each training takes up about 30 GPU hours. The pose decoder is trained with the same optimizer setting for 10k iterations with a batch size of 24 and each training takes about 10 GPU hours. The conditional encoder is pre-trained on ImageNet [6]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.577, + 0.74, + 0.593 + ], + "angle": 0, + "content": "4.4. Comparative Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.909, + 0.869 + ], + "angle": 0, + "content": "Here, we evaluate the proposed EA6D on the LM, LM-O and YCB-V in comparison to some state-of-the-art methods including six fully-supervised methods that are trained with annotated real images with object poses (DPOD [48], PVNet [28], SO-Pose [8]), ZebraPose [35], CheckerPose [23] and 6D-diff [45], two self-supervised methods that are trained with only synthetic data (AAE [36], MHP [26]), three self-supervised methods that are trained with both synthetic data and unannotated real images (DSC-PoseNet [46], SMOC-Net [37], TexPose [5]), and three self-supervised methods that are trained with synthetic data, unannotated real images and depth images (Self6D [39], Self6D++ [40], RKHSPose [43]). In addition, for a more comprehensive comparison, we also compare our method with Self6D++ [40] and RKHSpose [43] trained only with synthetic data. Table 1 reports the average metrics on each dataset and we provide detailed results in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "Comparisons on Linemod. As seen from Table 1, there is a significant performance gap between fully-supervised" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8683" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.485, + 0.215 + ], + "angle": 0, + "content": "Table 1. Comparison with state-of-the-art methods on LM, LM-O and YCB-V. The table reports results for the Average Recall (\\%) of ADD(-S) on LM and LM-O, and AUC of ADD-S and ADD(-S) on YCB-V. All results except ours are copied from SMOC-Net [37], TexPos [5] and RKHSPose [43]. \\( R \\): annotated real RGB data. \\( S \\): synthetic RGB data. \\( R^{-} \\): unannotated real RGB data. \\( D \\): depth data. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold. (-) denotes results missing from the original paper." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.219, + 0.485, + 0.485 + ], + "angle": 0, + "content": "
MethodsTraining dataLMLM-OYCB-V
ADD(-S)ADD-S AUCADD(-S) AUC
DPOD [48]R+S82.647.3--
PVNet [28]86.340.8-73.8
SO-Pose [8]96.062.390.983.9
ZebraPose [35]-76.990.185.3
CheckerPose [23]-77.591.386.4
6D-diff [45]-79.691.587.0
Self6D [39]S+R−+D58.932.1--
Self6D++ [40]88.564.791.480.0
RKHSPose [43]95.968.792.683.0
DSC-PoseNet [46]S+R−54.721.9--
Self6D++ [40]76.059.8--
SMOC-Net [37]91.363.3--
TexPose [5]91.766.7--
AAE [36]S31.4---
MHP [26]38.8---
Self6D++ [40]77.452.989.477.8
RKHSPose [43]78.254.390.276.5
EA6D96.387.893.188.3
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.483, + 0.688 + ], + "angle": 0, + "content": "methods and approaches trained with synthetic data instead. Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities (unannotated real RGB data, depth data) to achieve comparable results. The proposed EA6D achieves a better performance than all the comparative methods (regardless of whether these methods are trained in either fully-supervised or self-supervised manner), and only uses synthetic images for training. This is attributed to EA6D generating environment-independent object representations for pose estimation rather than estimating poses directly from input images, avoiding the impact of domain gap on network performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Comparisons on Linemod-Occluded. In the inference stage, each object in an image will be inferred independently. The corresponding results are reported in Table 1 which shows that EA6D outperforms all these methods, demonstrating strong robustness against occlusion. Note that EA6D only uses synthetic images as training data and still outperforms fully supervised methods (6D-diff by \\(8.3\\%\\), CheckerPose by \\(10.3\\%\\), ZebraPose by \\(10.9\\%\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.886 + ], + "angle": 0, + "content": "Comparisons on YCB-V. As shown in Table 1, EA6D achieves the best performance on both the AUC of ADD-S and the AUC of ADD(-S) metrics, showing that EA6D is capable of handling a wide variety of object appearances and lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Comparisons on Unseen Environments. In this experi" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.407 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on Linemod (top-left), Linemod-Occluded (bottom) and HomebrewedDB (top-right). For each dataset, Top: input images and the visulizations of pose, green and blue bounding boxes represent GT poses and estimated poses, respectively. Middle: Visualization of generated environment-independent object representations. Bottom: The image after decoding the environment-independent object representations." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.412, + 0.907, + 0.467 + ], + "angle": 0, + "content": "Table 2. Evaluation results on HomebrewedDB dataset. All results except ours are copied from TexPose [5]. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.48, + 0.907, + 0.57 + ], + "angle": 0, + "content": "
Training dataR− + SS
MethodSock[32]Self6D++[40]Texpose[5]GDR[41]EA6D
Benchvise57.375.792.988.897.6
Driller46.689.494.292.893.8
Phone41.576.881.278.792.6
Mean52.080.689.486.894.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.578, + 0.907, + 0.607 + ], + "angle": 0, + "content": "Table 3. Evaluation results on DiverseScenes dataset. All cited methods are re-implemented using their open source codes." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.618, + 0.905, + 0.713 + ], + "angle": 0, + "content": "
MethodsSelf6D++ [40]GDR [41]Texpose [5]EA6D
Scene 156.564.182.084.4
Scene 262.270.776.483.9
Scene 348.668.966.583.0
Scene 469.554.364.384.0
Mean59.264.572.383.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ment, we want to study the generalizability of our method towards new scenes. First, we evaluate our method on HomebrewedDB. From Table 2, we observe that EA6D outperforms all other methods and achieves the same results as testing in Linemod (95.7% vs.94.7%). Self6D++ and TexPose undergo a clear performance drop when testing on Linemod. Self6D++ drops from 94.1% to 80.6% and TexPose drops from 91.8% to 89.4%. It indicates that the generalization performance of existing self-supervised methods is limited by the unannotated real images used for training. Further, Table 3 shows the results of the test on the DiverseScenes dataset. In environments with significant changes" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8684" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "in background and lighting, there is a certain decline in performance for all four methods, but our method is the least affected and outperforms the others by a huge leap." + }, + { + "type": "image_caption", + "bbox": [ + 0.167, + 0.138, + 0.216, + 0.148 + ], + "angle": 0, + "content": "Scene 1" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.148, + 0.287, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.187, + 0.218, + 0.197 + ], + "angle": 0, + "content": "Scene 3" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.148, + 0.48, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.187, + 0.407, + 0.197 + ], + "angle": 0, + "content": "Scene 4" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.197, + 0.289, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.169, + 0.237, + 0.408, + 0.248 + ], + "angle": 0, + "content": "Same object in 7 different environments" + }, + { + "type": "image", + "bbox": [ + 0.29, + 0.197, + 0.48, + 0.236 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.248, + 0.208, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.248, + 0.342, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.343, + 0.248, + 0.479, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.301, + 0.483, + 0.356 + ], + "angle": 0, + "content": "Figure 5. The top two rows are qualitative results on DiverseScenes Dataset. Each scene has different backgrounds and lighting conditions. The bottom row is qualitative results of a same object in 7 different environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.484, + 0.511 + ], + "angle": 0, + "content": "We would like to emphasise that since our method is only trained on Synthetic PBR Dataset [13], the aforementioned datasets are all unseen environments for our method. Our method maintains accurate predictions (average \\(90.6\\%\\)) under different backgrounds, lighting conditions, and occlusions. It shows that the choice of adaptation domain (LM, HB, or other environments) has a marginal effect on our final performance. We attribute this to the generation of pure object features, eliminating the domain gap problem caused by environmental factors." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.518, + 0.226, + 0.533 + ], + "angle": 0, + "content": "4.5. Visulizations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.782 + ], + "angle": 0, + "content": "In this section, we will analyze our work by visualization of the test results. Figure 4 illustrates the visualizations of environment-independent object representations and pose estimation. It clearly shows that the images generated by the diffusion model filter out all environmental factors while accurately preserving the object information. The visualizations of results on LM-O explain why EA6D can effectively handle occlusions. With the strong generative capabilities of diffusion models, even when facing severe occlusion, the model is still able to accurately generate occluded parts. Figure 5 shows the visualizations of the results on the DiverseScenes Dataset and the visualizations of the same object in 7 different environments. These provide a clear demonstration of its capabilities in handling various environments. We give more visulizations results in supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.255, + 0.803 + ], + "angle": 0, + "content": "4.6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The effect of diffusion models and diffusion steps. We study the impact of different diffusion models on EA6D performance. We replace EDDM with traditional Denoising Diffusion Probabilistic Model (DDPM) [11] and Decoupled Diffusion Model (DDM) [15] and compare the results on Linemod. Also, the number of iterating steps is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "another key parameter in diffusion models that determines the inference speed, which is essential for pose estimation. All the results are reported in Table 4. We can observe that the EA6D with EDDM achieves more accurate results with fewer iterating steps. This result verifies the effectiveness of mixing of input image latent representations and noise instead of pure noise in generation process, which provides prior information, thus speeding up the inference procedure and improving the generation quality." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.234, + 0.905, + 0.262 + ], + "angle": 0, + "content": "Table 4. Mean results with ablation EA6D and the number of iterating steps on Linemod." + }, + { + "type": "table", + "bbox": [ + 0.534, + 0.266, + 0.884, + 0.401 + ], + "angle": 0, + "content": "
Method\\steps1510201001000
EA6D-DDPM [11]003.614.352.778.4
EA6D-DDM [15]93.196.195.696.396.296.3
EA6D w/o VAE68.471.373.175.176.277.4
EA6D*76.779.279.279.479.379.5
EA6D96.396.396.496.596.596.5
Time(s)EA6D0.230.510.871.606.9892.1
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.651 + ], + "angle": 0, + "content": "The effect of object representations. In our framework, we generate the object representations via performing the denoising process to estimate object pose. The object representations are the latent code of pure object images encoded by VAE. To evaluate the efficacy of such representations, we first remove the VAE and directly generate the pure object images as object representations through EDDM. Then we use the framework of EA6D, but decode the generated object representations to pure object images using the decoder of VAE and input pose decoder to estimate pose. We denote this setting as EA6D*. As shown in Table 4, the performance of these two types of object representations significantly drops, showing that the latent code of pure object images captures the rich semantic and structural information in the image and is more suitable for pose estimation than the generated images." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.665, + 0.634, + 0.68 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a novel generative paradigm for 6D pose estimation. Different from typical RGB-based pose estimation frameworks which directly extract features from input images for pose estimation and are therefore easily influenced by environmental factors, we employ an environment decoupling model that separates the object from environmental factors during the denoising process, generating environment-independent object representations which are robust to environmental changes. A pose decoder is designed to estimate poses using these representations. The experimental results on five datasets demonstrate that EA6D significantly outperforms current state-of-the-art self-supervised methods, particularly in its ability to generalize to previously unseen scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8685" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "6. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.27 + ], + "angle": 0, + "content": "This research was supported by the National Key Research and Development Program of China (No.2024YFF0907600), the National Natural Science Foundation of China (Grant No.62273275), the China Postdoctoral Science Foundation (No.2022M722573), Shaanxi Postdoctoral Research Project (No.2023BSHEDZZ244), Shaanxi Natural Science Basic Research Program (No.2025JC-YBQN-928), Key Research and Development Plan of Shaanxi Province (2025GH-YBXM-018) and Xi'an Science and Technology Innovation and Qinchuangyuan Innovation Major Program (23ZDCYJSGG0009-2023)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.298, + 0.188, + 0.314 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.324, + 0.481, + 0.351 + ], + "angle": 0, + "content": "[1] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.483, + 0.407 + ], + "angle": 0, + "content": "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.412, + 0.482, + 0.452 + ], + "angle": 0, + "content": "[3] E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.455, + 0.482, + 0.508 + ], + "angle": 0, + "content": "[4] E. Brachmann, F. Michel, A. Krull, M.Y. Yang, S. Gumhold, and C. Rother. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.513, + 0.482, + 0.567 + ], + "angle": 0, + "content": "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In CVPR, 2023. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.57, + 0.482, + 0.61 + ], + "angle": 0, + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.614, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.482, + 0.711 + ], + "angle": 0, + "content": "[8] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting self-occlusion for direct 6d pose estimation. In ICCV, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.715, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.482, + 0.813 + ], + "angle": 0, + "content": "[10] S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, and N. Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In ACCV, 2012. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[12] Tomáš Hodan, P. Haluza, Štepan Obdržálek, J. Matas, M. Lourakis, and X. Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In WACV, 2017. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.324, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[13] Tomás Hodař, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In ECCV Workshops, 2020. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.904, + 0.177 + ], + "angle": 0, + "content": "[14] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.18, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[15] Yuhang Huang, Zheng Qin, Xinwang Liu, and Kai Xu. Decoupled diffusion models with explicit transition probability. arXiv preprint arXiv:2306.13720, 2023. 3, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[16] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCV Workshops, 2019. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.264, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[17] W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.308, + 0.905, + 0.334 + ], + "angle": 0, + "content": "[18] DiederikP. KingmaandMaxWelling. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[19] V. Lepetit, F Moreno-Noguer, and P. Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 81(2):155-166, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.379, + 0.905, + 0.445 + ], + "angle": 0, + "content": "[20] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.904, + 0.489 + ], + "angle": 0, + "content": "[21] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.905, + 0.544 + ], + "angle": 0, + "content": "[22] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, 2019. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.588 + ], + "angle": 0, + "content": "[23] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In ICCV, 2023. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.591, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[26] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In ICCV, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] S. Peng, Y. Liu, Q. Huang, X. Zhou, and H. Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] M. Rad and V. Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8686" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[31] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. TPAMI, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[32] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In 3DV, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[33] Chen Song, Jiaru Song, and Qixing Huang. Hybridpose: 6d object pose estimation under hybrid representations. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[35] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In CVPR, 2022. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[36] M. Sundermeyer, Z. Marton, M. Durner, M. Brucker, and R. Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In ECCV, 2018. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.414, + 0.482, + 0.453 + ], + "angle": 0, + "content": "[37] Tao Tan and Qiulei Dong. Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In CVPR, 2023. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[38] B. Tekin, S.N. Sinha, and P. Fua. Real-time seamless single shot 6d object pose prediction. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[39] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, 2020. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[40] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6d object pose estimation. TPAMI, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[41] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.676 + ], + "angle": 0, + "content": "[42] Tianfu Wang, Guosheng Hu, and Hongguang Wang. Object pose estimation via the aggregation of diffusion features. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.482, + 0.718 + ], + "angle": 0, + "content": "[43] Yangzheng Wu and Michael Greenspan. Pseudo-keypoint rkhs learning for self-supervised 6dof pose estimation. In ECCV, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.774 + ], + "angle": 0, + "content": "[44] Y. Xiang, T. Schmidt, V. Narayanan, and D. Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[45] Li Xu, Haoxuan Qu, Yujun Cai, and Jun Liu. 6d-diff: A keypoint diffusion framework for 6d object pose estimation. In CVPR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[46] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[47] Yunfan Ye, Kai Xu, Yuhang Huang, Renjiao Yi, and Zhiping Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. In AAAI, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[48] S. Zakharov, I. Shugurov, and S. Ilic. Dpod: 6d pose object detector and refiner. In ICCV, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[49] Shaobo Zhang, Wanqing Zhao, Ziyu Guan, Xianlin Peng, and Jinye Peng. Keypoint-graph-driven learning framework for object pose estimation. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.093, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "8687" + } + ] +] \ No newline at end of file diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_origin.pdf b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..027ce2497298aeb82b56270b62f4762d72754ebb --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/5521ae38-4b00-475c-a94f-31fec5839328_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b395d58c68f4278811ab8cc3a46c99ca397276fa5e320d73f4595f6f3133f26a +size 1255001 diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/full.md b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8cba44663860601a27b169f885bdf5ee59a55027 --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/full.md @@ -0,0 +1,308 @@ +# Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation + +Shaobo Zhang + +Yuhang Huang + +Ziyu Guan + +Wanqing Zhao $^{1*}$ + +Jinye Peng $^{1*}$ + +Wei Zhao3 + +$^{1}$ Northwest University + +$^{2}$ National University of Defense Technology + +$^{3}$ Xidian University + +{zhangshaobo, zhaowq, ziyuguan, pjy}@nwu.edu.cn, huangai@nudt.edu.cn, ywzhao@mail.xidian.edu.cn + +# Abstract + +This paper introduces EA6D, a novel diffusion-based framework for 6D pose estimation that operates effectively in any environment. Traditional pose estimation methods struggle with the variability and complexity of real-world scenarios, often leading to overfitting on controlled datasets and poor generalization to new scenes. To address these challenges, we propose a generative pose estimation paradigm that generates environment-independent object representations for pose estimation, which are robust to environmental variations such as illumination, occlusion, and background clutter. Specifically, we propose the novel Environment Decoupling Diffusion Model (EDDM) which separates object representations from environmental factors while enabling efficient few-step sampling by leveraging input image priors instead of pure noise initialization. We validate our approach on four standard benchmarks and a self-made dataset DiverseScenes. The results demonstrate that EA6D, trained using only synthetic data, can outperform the state-of-the-art methods with both synthetic and realistic data. In particular, for fair comparisons with synthetic data, we can exceed the previous SOTA by $18.1\%$ and $33.5\%$ on Linemod and Linemod-Occluded datasets respectively. Project page: https://github.com/acf22/EA6D + +# 1. Introduction + +6D Pose Estimation is a pivotal task in the field of computer vision and robotics, which aims to determine the exact position and orientation of an object in three-dimensional space relative to a camera or a coordinate system. The ability to accurately perceive an object's pose is fundamental for a wide range of applications, including robotic manipulation, augmented reality and autonomous navigation. + +Recently, a series of deep-learning-based object 6D pose estimation methods [17, 22, 23, 28, 31, 35, 41, 44] have + +![](images/58882ef3d1bc59fa258a1c3c65ea996bc0a68d9cec782bd97aeba59a485f1499.jpg) +Figure 1. Different from typical RGB-based pose estimation frameworks, which are easily influenced by environmental factors leading to inaccurate pose estimation, we propose a generative paradigm that leverages diffusion models to generate environment-independent object representations for pose estimation. + +been proposed. These methods achieve state-of-the-art results, but require large annotated datasets. Generating accurate annotations for pose data is a labor-intensive and time-consuming process, leading most deep-learning-based 6D pose estimation methods to be trained and tested on closed datasets. These datasets [3, 10, 12, 16] typically contain only a few hundred images with limited variation and controlled factors, such as object arrangement, lighting conditions, and background, resulting in overfitting to the training set and the struggle to generalize to new scenes. + +Consequently, synthetic datasets have been employed as an additional complement to real datasets, or alternative to real datasets, as they can be easily and automatically annotated in simulated environments. The primary challenge with synthetic data is the domain discrepancy between real and synthetic datasets, which often leads to reduced accuracy when applied to real-world scenarios. To alleviate this problem, some self-supervised methods [20, 37, 39, 40, 46] are proposed, which leverage synthetic images with pose annotations and unannotated real images during training. + +These methods harness intrinsic structures to generate supervision signals and implement constraints on the discrepancies between real and synthetic images, thereby mitigating the domain gap. + +Practical applications such as AR and autonomous driving, which need to operate in uncontrolled environments, make 6D pose estimation more challenging. Here, object appearances vary due to fluctuating environmental conditions, background clutter, and occlusions, which often exceed the variability present in training datasets. Previous methods usually use CNNs as image encoder to extract features for pose estimation (as shown on the top side of Figure 1). These features are highly sensitive to environmental factors. When the environment changes, CNNs struggle to accurately extract object features. This hinders the model's ability to generalize to new scenes, potentially diminishing the accuracy of pose estimation and even leading to object pose estimation failure. While self-supervised methods are proficient at leveraging unlabeled data, they may not fully account for the extensive real-world variability found in these uncontrolled environments, leading to models that are not sufficiently robust for novel conditions. Therefore, there is an urgent need for approaches that can adapt to a variety of dynamic and diverse environments, ensuring reliable pose estimation capabilities. + +In this paper, we introduce a novel paradigm for pose estimation in uncertain environments, including unseen scenes, varying lighting, occlusion, and other possible situations. We consider a generative strategy that generates environment-independent object representations for pose estimation (bottom side of Figure 1). Diffusion models [7, 11, 34] have recently achieved promising results in various generative tasks, including image generating and editing. These models recover high-quality, deterministic samples from noisy and some conditions through a progressive denoising procedure. Motivated by the above, we design an Environment-Agnostic object 6D POSE estimation framework (named EA6D) that exploits the diffusion model to generate environment-independent representations in latent space for pose estimation, while input images serve as conditional cues, illustrated on the bottom side of Figure 1. Specifically, we propose an Environment Decoupling Diffusion Model (EDDM) which employs a novel environment decoupling diffusion process to intrinsically decouple the environment-independent object representations from input images. By performing validation on four popular benchmarks and a self-made dataset, the results show that our method significantly outperforms state-of-the-art self-supervised methods and provides a significant improvement in generalization to unseen scenarios. + +In sum, the main contributions in this paper include: + +- We propose a new paradigm for environment-agnostic object 6D pose estimation, by generating environment + +independent object representations rather than estimating poses directly from input images. + +- We design a novel Environment Decoupling Diffusion Model which employs an explicit environmental decoupling mechanism via gradient modelling and conditional generation to separate object representations from environment, while enabling efficient several-step sampling. +- We conduct extensive experiments on four public datasets including Linemod, Linemod-Occluded, YCB-V, HomebrewedDB and a self-made dataset DiverseScenes, showing that our method surpasses current state-of-the-art self-supervised methods. Our method exhibits remarkable generalization capabilities, even in previously unseen scenes, which is crucial for practical applications. + +# 2. Related work + +Fully-Supervised 6D Object Pose Estimation. Early approaches rely on local or global features to establish correspondences between keypoints on CAD models [1, 3, 4]. More recently, learning-based methods, particularly those utilizing Convolutional Neural Networks (CNNs), have become dominant. These methods extract deep features and establish mappings from RGB images to 6D object pose [14, 17, 41, 44]. Due to the inherent nonlinearity of the rotation space, the accuracy of directly regressed object pose is often limited [21]. To enhance accuracy, many methods employ CNNs to establish 2D-3D correspondences, followed by solving the object pose using the Perspective-n-Point (PnP) algorithm [19]. These methods can be categorized into sparse and dense correspondence methods. Sparse methods [28, 33, 38, 45] predict the 2D projections of predefined 3D keypoints, while dense methods [22, 23, 27, 35, 41, 48] predict 3D coordinates on the object model corresponding to image pixels thus providing better robustness to occlusions and clutter. However, fully-supervised 6D pose estimation methods are constrained by the need for accurate and diverse training data, which is often scarce due to the labor-intensive nature of precise annotation and the limited variation in available datasets, resulting in poor generalization to real-world and uncontrolled environments. + +Self-Supervised 6D Object Pose Estimation. Considering the tremendous effort to collect large amounts of annotations for 6D object pose, some methods [26, 29, 36] have turned to using synthetic images for training. However, because of the domain gap between real and synthetic images, the performances of these methods are significantly lower compared to the methods that use real images with object poses for training. To narrow the domain gap, a few recent self-supervised methods [5, 20, 37, 39, 40, 43, 46, 49] have begun to incorporate both synthetic images with known object poses and unannotated real images into their training datasets. However, in diverse real-world application sce + +narios, models must contend with a variety of backgrounds, lighting conditions, and other environmental factors that are difficult to be adequately included in the training dataset. These environmental factors can significantly alter the appearance of objects within images. As a result, features learned by the model for the same object may vary across different scenes, leading to diminished generalization capabilities when encountering new scenes. Different from self-supervised learning methods, our proposed method directly generates environment-independent object representations to fundamentally eliminate the impact of environmental disturbances. + +Diffusion-based Models for Vision Tasks. Originally introduced for image generation, diffusion models [7, 11, 34] have demonstrated impressive generation capabilities and have been explored in various tasks, including edge detection [47], image inpainting [25], and 6D pose estimation [9, 42, 45]. DiffusionEdge [47] presents a novel diffusion-based edge detector capable of generating precise and sharp edge maps without post-processing. 6D-diff [45] formulates object keypoints detection as a reverse diffusion process and employs a MoC-based forward process to leverage distribution priors in intermediate representations. Method [42] introduces three novel aggregation networks designed to effectively aggregate diffusion features, showing superior generalizability for object pose estimation. Currently, no existing work has yet explored the potential of diffusion models to generate environment-independent object representations, which could significantly improve the robustness of pose estimation in diverse environments. + +# 3. Method + +# 3.1. Overview + +Traditional pose estimation methods typically estimate pose directly from images. Environmental factors such as background clutter, illumination variations, and occlusion in images can limit the accuracy and robustness of pose estimation. Pursuing a different paradigm, we propose a diffusion-based method that generates environment-independent object representations for pose estimation. Traditional diffusion models are predominantly utilized for image generation tasks, where they excel at creating new synthetic images. These models operate on the principle of gradually adding noise to an input image and then learning to reverse this process, effectively denoising the image over multiple iterations to recover the original content. In this paper, we extend the concept of diffusion models to a completely new domain by recovering environment-independent object representations for pose estimation. + +The overall framework of the proposed method EA6D is shown in Figure 2. The first stage is environment-independent object representation encoding. A synthesized + +pure object image with the same pose as the input image, but free from background, lighting, and occlusion, is encoded into a latent space using a Variational Autoencoder (VAE) [18]. This encoding results in an object representation $g_0 \in R^{64 \times 64 \times 3}$ , which can be decoded back to the original pure object image by the decoder of VAE network, demonstrating that the representation contains only information about the object independent of the environment. Following this, a forward diffusion process is applied to $g_0$ to generate a noisy representation $g_t$ , which can be split into two sub-processes: the clean object representation $g_0$ to input image latent representation $r_0$ , meanwhile Gaussian noise is incrementally added. To denoise and recover the environment-independent object representations from $g_t$ , an Environment Decoupling Diffusion Model (EDDM) is introduced. A conditional encoder extracts multi-level features from the input images. These features are then fused with the features of the U-Net encoder and are input to U-Net decoders. The EDDM outputs a decoupled object representation, which potentially contains object pose information, independent of environmental factors. Finally, a pose decoder predicts the pose information, including rotation quaternion and translation vector. The inference process in EA6D is thus a combination of diffusion-based denoising and pose decoding, enabling robust and accurate pose estimation in diverse and complex environments. + +# 3.2. Environment Decoupling Diffusion for Generating Object Representations + +Current 6D pose estimation methods often face robustness challenges in dynamic or unseen environments, as their object representations are influenced by complex environmental variations, resulting in multimodal distributions of features. CNNs-based methods typically use a single model to fit the entire data distribution. When facing multimodal distributions, the single model struggles to capture all the different patterns simultaneously, thereby resulting in a notable decrease in performance in new environments. To avoid the negative impact, we propose to extract environment-independent representations, enabling more accurate and generalizable pose estimation. Recent diffusion models show remarkable capabilities for modeling complex multimodal distribution, which is helpful for learning about environment-independent representations. Therefore, we propose to construct our framework based on the diffusion models to cover the complex distribution of object representations. More importantly, to acquire the environment-independent component from the input image distribution, we propose a novel environment decoupling diffusion model for generating environment-independent object representations inspired by the decoupled diffusion models [15, 47]. Different from the vanilla decoupled diffusion process that focuses on splitting the image and the + +![](images/bec0c1acf76e5faa60c7e4db7041f7da3ef6f44ef832801d32b4cb7bef0d4f39.jpg) +Figure 2. Overview of the proposed method. In training phase, the proposed method involves training a Variational Autoencoder (VAE) to encode pure object images into environment-independent representations. These representations are then sampled to create noisy representations, which are processed by an Environment Decoupling Diffusion Model (EDDM). The EDDM uses a dual structure with one UNet encoder and two UNet decoders to decouple the environment-independent object representations from the input image with complex environments. The pose decoder is trained to predict the pose from the representations, enabling robust pose estimation across various environments. In testing phase, a random Gaussian noise is fed into the EDDM without VAE. + +noise components, we aim to decouple the environment-independent object representations from the input image with complex environments. To this end, we introduce the improved forward and reversed diffusion processes. + +# 3.2.1.Forward Process + +The vanilla decoupled diffusion process reformulates the conventional image-to-noise process into two stages: image-to-zero and zero-to-noise [15]. Corresponding to current object representation generation, there will be a clean object representation to zero process and a zero to noise process. However, such a paradigm only describes how to learn the environment-independent object representation from noise, without a clue on how to decouple it from the input images with complex environments. To address this problem, this paper proposes an environment decoupling diffusion process to intrinsically decouple the environment-independent object representation from the input image. We follow [30] to conduct the diffusion process in the latent space, i.e., our goal is to obtain the environment-independent object latent representations from the input image that mixes the object with the complex environments: + +$$ +q \left(g _ {t} \mid g _ {0}\right) = \mathcal {N} \left(g _ {t}; g _ {0} + \int_ {0} ^ {t} \left(r _ {0} - g _ {0}\right) d t, t I\right), \tag {1} +$$ + +where $g_{0}$ and $g_{t}$ denote the clean and noisy object latent representations respectively, $t\in [0,1]$ is the time step, $r_0$ denotes the latent representation of input image. (A Variational Autoencoder features the latent representation.) According to the above equation, we can sample $g_{t}$ by $g_{t} = g_{0} + \int_{0}^{t}(r_{0} - g_{0})dt + \sqrt{t} n,$ where $n\in \mathcal{N}(0,I)$ . Ob + +viously, the forward process describes two sub-processes: $g_0 + \int_0^t (r_0 - g_0)dt$ is the clean object representation to input image latent representation and $\sqrt{t} n$ is the zero to noise. It is worth noting that the first term contains the decay gradient from the input image to the environment-independent object representation, i.e., $r_0 - g_0$ . We call this term the environment decoupling gradient, which guides the generation from the input image representation to object representation. The environment decoupling gradient provides explicit evidence for our goal of decoupling the environment-independent object representations, allowing for accurate and efficient generation. + +# 3.2.2. Reversed Process + +The goal of the reversed process is to gradually recover the clean object representation $g_{0}$ from the noisy input $g_{1}$ . Following the derivation of [15], we can formally represent the reversed process as a conditional probability: + +$$ +\begin{array}{l} q \left(g _ {t - \triangle t} \mid g _ {t}, g _ {0}\right) = \mathcal {N} \left(g _ {t} + \int_ {t} ^ {t - \triangle t} \left(r _ {0} - g _ {0}\right) d t - \frac {\triangle t}{\sqrt {t}} n, \right. \\ \frac {\triangle t (t - \triangle t)}{t} I), \tag {2} \\ \end{array} +$$ + +where $\triangle t$ is the smallest time step. Inheriting DDM, the reversed process benefits from the analytic gradient $r_0 - g_0$ , enabling approximate analytic solution for few-step generation (even with only one step). Different from general diffusion models that recover the goal content from a normal distribution, our reversed process aims to generate object representations from the combination of the latent image representations and the noise. We can easily obtain the ini + +tial variable in the reversed process according to the forward process, i.e., $g_{1} = r_{0} + n$ . Intuitively, the generation process starts with the mixture of $r_0$ and $n$ instead of pure noise, which provides important prior information, thus speeding up the inference procedure and improving the generation quality. The reversed conditional probability $q(g_{t - \triangle t}|g_t,g_0)$ includes $r_0 - g_0$ and $n$ , which are unknown in the reversed process. Therefore, we need to use a neural network $\theta$ to parameterize $q(g_{t - \triangle t}|g_t,g_0)$ . Practically, the neural network has two outputs: $d_{\theta}$ and $n_{\theta}$ , corresponding to the two unknown terms. + +# 3.2.3. Network Architecture + +As shown in Figure 2, we construct our network based on a U-Net architecture including one encoder and two decoders. A pre-trained Swin Transformer [24] serves as a conditional encoder, extracting multi-level features from the input images. These features are then fused with the features of the U-Net encoder at corresponding levels through attention mechanisms, acting as inputs for the U-Net decoders. In practice, the input images will be obtained by an off-the-shelf object detector such as Yolo-v4 [2]. + +# 3.3. Pose Decoder for Object Representations + +The pose decoder is designed to predict the 6D pose, including both the rotation and translation, from the object representations generated from EDDM. The decoder consists of three modules: a ResNet block to further extract the features of $\hat{g}_0^t$ , a rotation head for rotation estimation, and a translation head for translation prediction. The ResNet block enhances feature extraction through three convolutional layers. The first layer employs a $1 \times 1$ kernel to expand the feature dimension, capturing richer information. The second layer uses a $3 \times 3$ kernel for spatial feature reduction. The third layer also uses a $1 \times 1$ kernel. The rotation head is designed with a series of $3 \times 3$ convolutional layers, followed by a generalized average pooling layer and a fully connected layer, culminating in the prediction of the rotation quaternion $R$ . + +For 3D translation, since directly regressing the translation in 3D space does not work well in practice, we adopt Scale-invariant Translation Estimation (SITE) [22] to decouple the translation into the 2D location $(o_x,o_y)$ of the projected 3D centroid and the object's distance $T_{z}$ towards the camera. The translation head is constituted by a series of convolutional layers followed by fully connected layers, which predict the scale-invariant translation parameters $T_{site} = (\Delta x,\Delta y,\Delta z)$ . These parameters are then utilized to compute the object's 2D coordinates and depth, as delineated by the equations: + +$$ +\left\{ \begin{array}{l l} o _ {x} = & \Delta x \cdot w + c _ {x} \\ o _ {y} = & \Delta y \cdot h + c _ {y} \\ T _ {z} = & \Delta z \cdot r \end{array} \right. \tag {3} +$$ + +where $w$ and $h$ represent the width and height of the bounding box, $c_{x}$ and $c_{y}$ denote the coordinates of the box's center, and $r$ is the scaling factor derived from the image's dimensions. Given the camera intrinsics $k$ , the translation $T$ can be calculated via back-projection $T = K^{-1}T_{z}[o_{x}, o_{y}, 1]^{T}$ . This method allows for a more accurate and robust estimation of the object's 3D translation, thereby enhancing the overall pose estimation accuracy. + +# 3.4. Training Procedure + +Our training object is meticulously designed to refine the performance of the VAE, EDDM, and pose decoder within our framework. Initially, the VAE is trained to encode synthetic images, which are environment-free and rendered with poses corresponding to the input images, into latent features. These latent features are then decoded to reconstruct the original synthesized images, ensuring the network's ability to capture the essential features of the object. + +After the VAE's training, we proceed to train the EDDM while keeping the VAE's weights fixed. EDDM is trained following the training strategy and objective of DDM [15]: + +$$ +L _ {e d d m} = \left\| d _ {\theta} - \left(r _ {0} - g _ {0}\right) \right\| _ {2} + \left\| n _ {\theta} - n \right\| _ {2} \tag {4} +$$ + +where the first term and second term correspond to the supervision of predicting noise component and image component, respectively. + +At last, we freeze VAE and EDDM to train the pose decoder. We employ a disentangled 6D pose loss via individually supervising the rotation $R$ , the scale-invariant 2D object center $(\Delta x, \Delta y)$ and the distance $\Delta z$ : + +$$ +L _ {\text {p o s e}} = L _ {R} + L _ {\text {c e n t e r}} + L _ {z} \tag {5} +$$ + +Thereby, + +$$ +\left\{ \begin{array}{r c l} L _ {R} & = & \left\| R - \bar {R} \right\| _ {2} \\ L _ {c e n t e r} & = & \| \Delta x - \Delta \bar {x} \| _ {2} + \| \Delta y - \Delta \bar {y} \| _ {2} \\ L _ {z} & = & \| \Delta z - \Delta \bar {z} \| _ {2} \end{array} \right. \tag {6} +$$ + +where $(\Delta x, \Delta y, \Delta z, R)$ and $(\Delta \bar{x}, \Delta \bar{y}, \Delta \bar{z}, \bar{R})$ denote prediction and ground truth, respectively. + +In the inference phase, object images from different environments are fed into the conditional encoder as cue information for the EDDM. The EDDM systematically eliminates environmental factors from the noisy latent space, yielding a representation that encapsulates solely the object's information. This purified representation is then fed into the pose decoder to extract the object pose, thereby achieving robust generalization across various scenes and conditions. + +# 4. Experiments + +In this section, we conduct experiments to verify the generalization capability of the proposed method in various environments. Therefore we follow existing works [5, 37] in + +self-supervised object pose estimation which focus on narrowing the domain gap and enhancing the network's generalization ability across different domains. We compare our method to the state-of-the-art methods on four of the core datasets of the BOP challenge [13] and a self-made dataset. We also provide a detailed analysis of our proposed scheme in ablation study. + +# 4.1. Datasets + +We evaluate our EA6D on four of the core BOP datasets [13], Linemod(LM) [10], Linemod-Occluded(LM-O) [3], HomebrewedDB(HB) [16] and YCB-V [44]. LM, LM-O, and HB include the same objects under different environments, and YCB-Video involves environmental changes, making them suitable for verifying the proposed method. HB is typically used to study the generalization ability of methods towards new scenes with changing illumination. However, the images in HB are from a single indoor environment. To expand the testing of unseen scenarios, the DiverseScenes Dataset is created by simulating both indoor and outdoor environments, to rigorously evaluate the model's ability to generalize across varied environments. + +DiverseScenes Dataset. We select four distinct real-world environments, spanning both indoor and outdoor settings, and employ Physically Based Rendering (PBR) techniques to render the 3D models into these backgrounds. To capture the inherent diversity of real-world lighting, we apply a range of lighting conditions, thus amplifying the disparity with the training data and crafting previously unseen test scenarios. For each object category, we render 150 images per environment, amassing a total of 1950 images. Figure 3 illustrates some samples of the dataset. We give more samples of the dataset in the supplementary material and will release the dataset. + +![](images/7607c70aac8adbfc83d3a0029a09539cac16c89334eb0284f0754c1111ac35a2.jpg) +Figure 3. We simulated different indoor and outdoor backgrounds with varying lighting to construct four distinct scenarios, which have a significant difference from the training data. + +It should be emphasized that EA6D does not require any real images for training. For the evaluation of EA6D, all images in the above dataset are used for testing only. + +# 4.2. Metrics + +We report our results referring to the common Average Distance of Distinguishable Model Points (ADD) metric [10]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. The estimated pose is considered to be correct if the average distance is smaller than $10\%$ of the object's diameter. For symmetric objects, we use the ADD-S metric [44], where the mean distance is computed based on the closest point distance. For YCB-V, we report the AUC (area under curve) of ADD(-S) with a maximum threshold of $10\mathrm{cm}$ [44]. + +# 4.3. Implementation Details + +Our method is implemented by Pytorch, and all experiments are conducted on an NVIDIA 4090 GPU. We use Synthetic PBR Dataset [13] to train the 2D object detector Yolov4 [2] and the entire network. For each image, we utilize the object's CAD model to render it in the same pose onto a blank image to obtain the pure object image. We use pure object images to train VAE and use them as ground-truth for corresponding input PBR images to train EDDM from scratch. We train VAE using AdamW optimizer with an attenuated learning rate (from $1e^{-4}$ to $5e^{-7}$ ). We train the diffusion models using AdamW optimizer with an attenuated learning rate (from $2e^{-4}$ to $5e^{-6}$ ) for 20k iterations with a batch size of 12 and each training takes up about 30 GPU hours. The pose decoder is trained with the same optimizer setting for 10k iterations with a batch size of 24 and each training takes about 10 GPU hours. The conditional encoder is pre-trained on ImageNet [6]. + +# 4.4. Comparative Evaluation + +Here, we evaluate the proposed EA6D on the LM, LM-O and YCB-V in comparison to some state-of-the-art methods including six fully-supervised methods that are trained with annotated real images with object poses (DPOD [48], PVNet [28], SO-Pose [8]), ZebraPose [35], CheckerPose [23] and 6D-diff [45], two self-supervised methods that are trained with only synthetic data (AAE [36], MHP [26]), three self-supervised methods that are trained with both synthetic data and unannotated real images (DSC-PoseNet [46], SMOC-Net [37], TexPose [5]), and three self-supervised methods that are trained with synthetic data, unannotated real images and depth images (Self6D [39], Self6D++ [40], RKHSPose [43]). In addition, for a more comprehensive comparison, we also compare our method with Self6D++ [40] and RKHSpose [43] trained only with synthetic data. Table 1 reports the average metrics on each dataset and we provide detailed results in the supplementary material. + +Comparisons on Linemod. As seen from Table 1, there is a significant performance gap between fully-supervised + +Table 1. Comparison with state-of-the-art methods on LM, LM-O and YCB-V. The table reports results for the Average Recall (\%) of ADD(-S) on LM and LM-O, and AUC of ADD-S and ADD(-S) on YCB-V. All results except ours are copied from SMOC-Net [37], TexPos [5] and RKHSPose [43]. $R$ : annotated real RGB data. $S$ : synthetic RGB data. $R^{-}$ : unannotated real RGB data. $D$ : depth data. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold. (-) denotes results missing from the original paper. + +
MethodsTraining dataLMLM-OYCB-V
ADD(-S)ADD-S AUCADD(-S) AUC
DPOD [48]R+S82.647.3--
PVNet [28]86.340.8-73.8
SO-Pose [8]96.062.390.983.9
ZebraPose [35]-76.990.185.3
CheckerPose [23]-77.591.386.4
6D-diff [45]-79.691.587.0
Self6D [39]S+R−+D58.932.1--
Self6D++ [40]88.564.791.480.0
RKHSPose [43]95.968.792.683.0
DSC-PoseNet [46]S+R−54.721.9--
Self6D++ [40]76.059.8--
SMOC-Net [37]91.363.3--
TexPose [5]91.766.7--
AAE [36]S31.4---
MHP [26]38.8---
Self6D++ [40]77.452.989.477.8
RKHSPose [43]78.254.390.276.5
EA6D96.387.893.188.3
+ +methods and approaches trained with synthetic data instead. Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities (unannotated real RGB data, depth data) to achieve comparable results. The proposed EA6D achieves a better performance than all the comparative methods (regardless of whether these methods are trained in either fully-supervised or self-supervised manner), and only uses synthetic images for training. This is attributed to EA6D generating environment-independent object representations for pose estimation rather than estimating poses directly from input images, avoiding the impact of domain gap on network performance. + +Comparisons on Linemod-Occluded. In the inference stage, each object in an image will be inferred independently. The corresponding results are reported in Table 1 which shows that EA6D outperforms all these methods, demonstrating strong robustness against occlusion. Note that EA6D only uses synthetic images as training data and still outperforms fully supervised methods (6D-diff by $8.3\%$ , CheckerPose by $10.3\%$ , ZebraPose by $10.9\%$ ). + +Comparisons on YCB-V. As shown in Table 1, EA6D achieves the best performance on both the AUC of ADD-S and the AUC of ADD(-S) metrics, showing that EA6D is capable of handling a wide variety of object appearances and lighting conditions. + +Comparisons on Unseen Environments. In this experi + +![](images/8ebdd9dbdd030d822b529f54a793e1b7773114396fe5567fa9bdc939ca8f4b2c.jpg) +Figure 4. Qualitative results on Linemod (top-left), Linemod-Occluded (bottom) and HomebrewedDB (top-right). For each dataset, Top: input images and the visulizations of pose, green and blue bounding boxes represent GT poses and estimated poses, respectively. Middle: Visualization of generated environment-independent object representations. Bottom: The image after decoding the environment-independent object representations. + +Table 2. Evaluation results on HomebrewedDB dataset. All results except ours are copied from TexPose [5]. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold. + +
Training dataR− + SS
MethodSock[32]Self6D++[40]Texpose[5]GDR[41]EA6D
Benchvise57.375.792.988.897.6
Driller46.689.494.292.893.8
Phone41.576.881.278.792.6
Mean52.080.689.486.894.7
+ +Table 3. Evaluation results on DiverseScenes dataset. All cited methods are re-implemented using their open source codes. + +
MethodsSelf6D++ [40]GDR [41]Texpose [5]EA6D
Scene 156.564.182.084.4
Scene 262.270.776.483.9
Scene 348.668.966.583.0
Scene 469.554.364.384.0
Mean59.264.572.383.8
+ +ment, we want to study the generalizability of our method towards new scenes. First, we evaluate our method on HomebrewedDB. From Table 2, we observe that EA6D outperforms all other methods and achieves the same results as testing in Linemod (95.7% vs.94.7%). Self6D++ and TexPose undergo a clear performance drop when testing on Linemod. Self6D++ drops from 94.1% to 80.6% and TexPose drops from 91.8% to 89.4%. It indicates that the generalization performance of existing self-supervised methods is limited by the unannotated real images used for training. Further, Table 3 shows the results of the test on the DiverseScenes dataset. In environments with significant changes + +in background and lighting, there is a certain decline in performance for all four methods, but our method is the least affected and outperforms the others by a huge leap. + +![](images/024e513529f972ef0c2f00d14fd22be4a5b401b83b9cf5b8f08bbd7248df3bca.jpg) +Scene 1 + +![](images/d97c730d4bb633fe2c5fb577054219dc46d1e484ce9a96b43ecad6ca72706461.jpg) + +![](images/d8fcb9179be1e02562e51c14037d4b92e90c6f8ab3ba100c1a398f84f14bc7ba.jpg) +Scene 3 + +![](images/7660107b8d885ac4cde7ff7c3ffe51a78f248b6681f9b2a3c1c10d12e4381ac6.jpg) +Scene 4 + +![](images/966e391bde2b3ad4051238f8728b4766afbaed64c69fcb86a495be0f2a82c9a8.jpg) +Same object in 7 different environments +Figure 5. The top two rows are qualitative results on DiverseScenes Dataset. Each scene has different backgrounds and lighting conditions. The bottom row is qualitative results of a same object in 7 different environments. + +![](images/d3fe39d75992438083932905c4ba4743681b170ab3dfa80eb1c5f8fe6c853f79.jpg) + +![](images/874bed435b455bed5ce92be74c83876a00fb293ed0441a1d1d321c9f6c1296c8.jpg) + +We would like to emphasise that since our method is only trained on Synthetic PBR Dataset [13], the aforementioned datasets are all unseen environments for our method. Our method maintains accurate predictions (average $90.6\%$ ) under different backgrounds, lighting conditions, and occlusions. It shows that the choice of adaptation domain (LM, HB, or other environments) has a marginal effect on our final performance. We attribute this to the generation of pure object features, eliminating the domain gap problem caused by environmental factors. + +# 4.5. Visulizations + +In this section, we will analyze our work by visualization of the test results. Figure 4 illustrates the visualizations of environment-independent object representations and pose estimation. It clearly shows that the images generated by the diffusion model filter out all environmental factors while accurately preserving the object information. The visualizations of results on LM-O explain why EA6D can effectively handle occlusions. With the strong generative capabilities of diffusion models, even when facing severe occlusion, the model is still able to accurately generate occluded parts. Figure 5 shows the visualizations of the results on the DiverseScenes Dataset and the visualizations of the same object in 7 different environments. These provide a clear demonstration of its capabilities in handling various environments. We give more visulizations results in supplementary material. + +# 4.6. Ablation Studies + +The effect of diffusion models and diffusion steps. We study the impact of different diffusion models on EA6D performance. We replace EDDM with traditional Denoising Diffusion Probabilistic Model (DDPM) [11] and Decoupled Diffusion Model (DDM) [15] and compare the results on Linemod. Also, the number of iterating steps is + +another key parameter in diffusion models that determines the inference speed, which is essential for pose estimation. All the results are reported in Table 4. We can observe that the EA6D with EDDM achieves more accurate results with fewer iterating steps. This result verifies the effectiveness of mixing of input image latent representations and noise instead of pure noise in generation process, which provides prior information, thus speeding up the inference procedure and improving the generation quality. + +Table 4. Mean results with ablation EA6D and the number of iterating steps on Linemod. + +
Method\steps1510201001000
EA6D-DDPM [11]003.614.352.778.4
EA6D-DDM [15]93.196.195.696.396.296.3
EA6D w/o VAE68.471.373.175.176.277.4
EA6D*76.779.279.279.479.379.5
EA6D96.396.396.496.596.596.5
Time(s)EA6D0.230.510.871.606.9892.1
+ +The effect of object representations. In our framework, we generate the object representations via performing the denoising process to estimate object pose. The object representations are the latent code of pure object images encoded by VAE. To evaluate the efficacy of such representations, we first remove the VAE and directly generate the pure object images as object representations through EDDM. Then we use the framework of EA6D, but decode the generated object representations to pure object images using the decoder of VAE and input pose decoder to estimate pose. We denote this setting as EA6D*. As shown in Table 4, the performance of these two types of object representations significantly drops, showing that the latent code of pure object images captures the rich semantic and structural information in the image and is more suitable for pose estimation than the generated images. + +# 5. Conclusion + +In this paper, we propose a novel generative paradigm for 6D pose estimation. Different from typical RGB-based pose estimation frameworks which directly extract features from input images for pose estimation and are therefore easily influenced by environmental factors, we employ an environment decoupling model that separates the object from environmental factors during the denoising process, generating environment-independent object representations which are robust to environmental changes. A pose decoder is designed to estimate poses using these representations. The experimental results on five datasets demonstrate that EA6D significantly outperforms current state-of-the-art self-supervised methods, particularly in its ability to generalize to previously unseen scenarios. + +# 6. Acknowledgments + +This research was supported by the National Key Research and Development Program of China (No.2024YFF0907600), the National Natural Science Foundation of China (Grant No.62273275), the China Postdoctoral Science Foundation (No.2022M722573), Shaanxi Postdoctoral Research Project (No.2023BSHEDZZ244), Shaanxi Natural Science Basic Research Program (No.2025JC-YBQN-928), Key Research and Development Plan of Shaanxi Province (2025GH-YBXM-018) and Xi'an Science and Technology Innovation and Qinchuangyuan Innovation Major Program (23ZDCYJSGG0009-2023). + +# References + +[1] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2 +[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5, 6 +[3] E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 1, 2, 6 +[4] E. Brachmann, F. Michel, A. Krull, M.Y. Yang, S. Gumhold, and C. Rother. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In CVPR, 2016. 2 +[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In CVPR, 2023. 2, 5, 6, 7 +[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 3 +[8] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting self-occlusion for direct 6d pose estimation. In ICCV, 2021. 6, 7 +[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 3 +[10] S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, and N. Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In ACCV, 2012. 1, 6 +[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 8 +[12] Tomáš Hodan, P. Haluza, Štepan Obdržálek, J. Matas, M. Lourakis, and X. Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In WACV, 2017. 1 + +[13] Tomás Hodař, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In ECCV Workshops, 2020. 6, 8 +[14] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, 2020. 2 +[15] Yuhang Huang, Zheng Qin, Xinwang Liu, and Kai Xu. Decoupled diffusion models with explicit transition probability. arXiv preprint arXiv:2306.13720, 2023. 3, 4, 5, 8 +[16] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCV Workshops, 2019. 1, 6 +[17] W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 1, 2 +[18] DiederikP. KingmaandMaxWelling. Auto-encoding variational bayes. In ICLR, 2014. 3 +[19] V. Lepetit, F Moreno-Noguer, and P. Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 81(2):155-166, 2009. 2 +[20] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, 2023. 1, 2 +[21] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In ECCV, 2022. 2 +[22] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, 2019. 1, 2, 5 +[23] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In ICCV, 2023. 1, 2, 6, 7 +[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5 +[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 3 +[26] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In ICCV, 2019. 2, 6, 7 +[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, 2019. 2 +[28] S. Peng, Y. Liu, Q. Huang, X. Zhou, and H. Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2, 6, 7 +[29] M. Rad and V. Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 2 + +[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4 +[31] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. TPAMI, 2021. 1 +[32] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In 3DV, 2020. 7 +[33] Chen Song, Jiaru Song, and Qixing Huang. Hybridpose: 6d object pose estimation under hybrid representations. In CVPR, 2020. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3 +[35] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In CVPR, 2022. 1, 2, 6, 7 +[36] M. Sundermeyer, Z. Marton, M. Durner, M. Brucker, and R. Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In ECCV, 2018. 2, 6, 7 +[37] Tao Tan and Qiulei Dong. Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In CVPR, 2023. 1, 2, 5, 6, 7 +[38] B. Tekin, S.N. Sinha, and P. Fua. Real-time seamless single shot 6d object pose prediction. In CVPR, 2018. 2 +[39] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, 2020. 1, 2, 6, 7 +[40] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6d object pose estimation. TPAMI, 2021. 1, 2, 6, 7 +[41] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1, 2, 7 +[42] Tianfu Wang, Guosheng Hu, and Hongguang Wang. Object pose estimation via the aggregation of diffusion features. In CVPR, 2024. 3 +[43] Yangzheng Wu and Michael Greenspan. Pseudo-keypoint rkhs learning for self-supervised 6dof pose estimation. In ECCV, 2024. 2, 6, 7 +[44] Y. Xiang, T. Schmidt, V. Narayanan, and D. Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1, 2, 6 +[45] Li Xu, Haoxuan Qu, Yujun Cai, and Jun Liu. 6d-diff: A keypoint diffusion framework for 6d object pose estimation. In CVPR, 2024. 2, 3, 6, 7 +[46] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In CVPR, 2021. 1, 2, 6, 7 +[47] Yunfan Ye, Kai Xu, Yuhang Huang, Renjiao Yi, and Zhiping Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. In AAAI, 2024. 3 + +[48] S. Zakharov, I. Shugurov, and S. Ilic. Dpod: 6d pose object detector and refiner. In ICCV, 2019. 2, 6, 7 +[49] Shaobo Zhang, Wanqing Zhao, Ziyu Guan, Xianlin Peng, and Jinye Peng. Keypoint-graph-driven learning framework for object pose estimation. In CVPR, 2021. 2 \ No newline at end of file diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/images.zip b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dffdaca54e1d3683a529c1067820c499ca24f04d --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fbadd99b8ffe49831fd5a31c840bc7cdc4e2b047c853c931b7f61c793307e25c +size 445479 diff --git a/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/layout.json b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4b48488361b83c8bad9aa83c38f01e6db0d3dcf9 --- /dev/null +++ b/2025/Environment-Agnostic Pose_ Generating Environment-independent Object Representations for 6D Pose Estimation/layout.json @@ -0,0 +1,8056 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "Environment-Agnostic Pose: Generating Environment-independent Object Representations for 6D Pose Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 112, + 160, + 189, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 160, + 189, + 176 + ], + "spans": [ + { + "bbox": [ + 112, + 160, + 189, + 176 + ], + "type": "text", + "content": "Shaobo Zhang" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 217, + 160, + 295, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 160, + 295, + 175 + ], + "spans": [ + { + "bbox": [ + 217, + 160, + 295, + 175 + ], + "type": "text", + "content": "Yuhang Huang" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 237, + 176, + 295, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 237, + 176, + 295, + 189 + ], + "spans": [ + { + "bbox": [ + 237, + 176, + 295, + 189 + ], + "type": "text", + "content": "Ziyu Guan" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 322, + 161, + 403, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 161, + 403, + 175 + ], + "spans": [ + { + "bbox": [ + 322, + 161, + 403, + 175 + ], + "type": "text", + "content": "Wanqing Zhao" + }, + { + "bbox": [ + 322, + 161, + 403, + 175 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 324, + 176, + 386, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 176, + 386, + 190 + ], + "spans": [ + { + "bbox": [ + 324, + 176, + 386, + 190 + ], + "type": "text", + "content": "Jinye Peng" + }, + { + "bbox": [ + 324, + 176, + 386, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 427, + 161, + 480, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 427, + 161, + 480, + 175 + ], + "spans": [ + { + "bbox": [ + 427, + 161, + 480, + 175 + ], + "type": "text", + "content": "Wei Zhao3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 77, + 189, + 186, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 189, + 186, + 203 + ], + "spans": [ + { + "bbox": [ + 77, + 189, + 186, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 77, + 189, + 186, + 203 + ], + "type": "text", + "content": "Northwest University" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 209, + 190, + 424, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 209, + 190, + 424, + 204 + ], + "spans": [ + { + "bbox": [ + 209, + 190, + 424, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 209, + 190, + 424, + 204 + ], + "type": "text", + "content": "National University of Defense Technology" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 448, + 189, + 541, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 448, + 189, + 541, + 204 + ], + "spans": [ + { + "bbox": [ + 448, + 189, + 541, + 204 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 448, + 189, + 541, + 204 + ], + "type": "text", + "content": "Xidian University" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 205, + 559, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 559, + 217 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 559, + 217 + ], + "type": "text", + "content": "{zhangshaobo, zhaowq, ziyuguan, pjy}@nwu.edu.cn, huangai@nudt.edu.cn, ywzhao@mail.xidian.edu.cn" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 152, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "text", + "content": "This paper introduces EA6D, a novel diffusion-based framework for 6D pose estimation that operates effectively in any environment. Traditional pose estimation methods struggle with the variability and complexity of real-world scenarios, often leading to overfitting on controlled datasets and poor generalization to new scenes. To address these challenges, we propose a generative pose estimation paradigm that generates environment-independent object representations for pose estimation, which are robust to environmental variations such as illumination, occlusion, and background clutter. Specifically, we propose the novel Environment Decoupling Diffusion Model (EDDM) which separates object representations from environmental factors while enabling efficient few-step sampling by leveraging input image priors instead of pure noise initialization. We validate our approach on four standard benchmarks and a self-made dataset DiverseScenes. The results demonstrate that EA6D, trained using only synthetic data, can outperform the state-of-the-art methods with both synthetic and realistic data. In particular, for fair comparisons with synthetic data, we can exceed the previous SOTA by " + }, + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "inline_equation", + "content": "18.1\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "inline_equation", + "content": "33.5\\%" + }, + { + "bbox": [ + 55, + 270, + 296, + 545 + ], + "type": "text", + "content": " on Linemod and Linemod-Occluded datasets respectively. Project page: https://github.com/acf22/EA6D" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 568, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 588, + 295, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 671 + ], + "type": "text", + "content": "6D Pose Estimation is a pivotal task in the field of computer vision and robotics, which aims to determine the exact position and orientation of an object in three-dimensional space relative to a camera or a coordinate system. The ability to accurately perceive an object's pose is fundamental for a wide range of applications, including robotic manipulation, augmented reality and autonomous navigation." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 295, + 696 + ], + "type": "text", + "content": "Recently, a series of deep-learning-based object 6D pose estimation methods [17, 22, 23, 28, 31, 35, 41, 44] have" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 242, + 552, + 397 + ], + "blocks": [ + { + "bbox": [ + 317, + 242, + 552, + 397 + ], + "lines": [ + { + "bbox": [ + 317, + 242, + 552, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 242, + 552, + 397 + ], + "type": "image", + "image_path": "58882ef3d1bc59fa258a1c3c65ea996bc0a68d9cec782bd97aeba59a485f1499.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 460 + ], + "type": "text", + "content": "Figure 1. Different from typical RGB-based pose estimation frameworks, which are easily influenced by environmental factors leading to inaccurate pose estimation, we propose a generative paradigm that leverages diffusion models to generate environment-independent object representations for pose estimation." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 472, + 555, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 592 + ], + "type": "text", + "content": "been proposed. These methods achieve state-of-the-art results, but require large annotated datasets. Generating accurate annotations for pose data is a labor-intensive and time-consuming process, leading most deep-learning-based 6D pose estimation methods to be trained and tested on closed datasets. These datasets [3, 10, 12, 16] typically contain only a few hundred images with limited variation and controlled factors, such as object arrangement, lighting conditions, and background, resulting in overfitting to the training set and the struggle to generalize to new scenes." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Consequently, synthetic datasets have been employed as an additional complement to real datasets, or alternative to real datasets, as they can be easily and automatically annotated in simulated environments. The primary challenge with synthetic data is the domain discrepancy between real and synthetic datasets, which often leads to reduced accuracy when applied to real-world scenarios. To alleviate this problem, some self-supervised methods [20, 37, 39, 40, 46] are proposed, which leverage synthetic images with pose annotations and unannotated real images during training." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 146, + 713 + ], + "type": "text", + "content": "*Corresponding authors" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8678" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "These methods harness intrinsic structures to generate supervision signals and implement constraints on the discrepancies between real and synthetic images, thereby mitigating the domain gap." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 56, + 122, + 295, + 372 + ], + "type": "text", + "content": "Practical applications such as AR and autonomous driving, which need to operate in uncontrolled environments, make 6D pose estimation more challenging. Here, object appearances vary due to fluctuating environmental conditions, background clutter, and occlusions, which often exceed the variability present in training datasets. Previous methods usually use CNNs as image encoder to extract features for pose estimation (as shown on the top side of Figure 1). These features are highly sensitive to environmental factors. When the environment changes, CNNs struggle to accurately extract object features. This hinders the model's ability to generalize to new scenes, potentially diminishing the accuracy of pose estimation and even leading to object pose estimation failure. While self-supervised methods are proficient at leveraging unlabeled data, they may not fully account for the extensive real-world variability found in these uncontrolled environments, leading to models that are not sufficiently robust for novel conditions. Therefore, there is an urgent need for approaches that can adapt to a variety of dynamic and diverse environments, ensuring reliable pose estimation capabilities." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 374, + 295, + 672 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 374, + 295, + 672 + ], + "spans": [ + { + "bbox": [ + 56, + 374, + 295, + 672 + ], + "type": "text", + "content": "In this paper, we introduce a novel paradigm for pose estimation in uncertain environments, including unseen scenes, varying lighting, occlusion, and other possible situations. We consider a generative strategy that generates environment-independent object representations for pose estimation (bottom side of Figure 1). Diffusion models [7, 11, 34] have recently achieved promising results in various generative tasks, including image generating and editing. These models recover high-quality, deterministic samples from noisy and some conditions through a progressive denoising procedure. Motivated by the above, we design an Environment-Agnostic object 6D POSE estimation framework (named EA6D) that exploits the diffusion model to generate environment-independent representations in latent space for pose estimation, while input images serve as conditional cues, illustrated on the bottom side of Figure 1. Specifically, we propose an Environment Decoupling Diffusion Model (EDDM) which employs a novel environment decoupling diffusion process to intrinsically decouple the environment-independent object representations from input images. By performing validation on four popular benchmarks and a self-made dataset, the results show that our method significantly outperforms state-of-the-art self-supervised methods and provides a significant improvement in generalization to unseen scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 673, + 279, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 673, + 279, + 685 + ], + "spans": [ + { + "bbox": [ + 67, + 673, + 279, + 685 + ], + "type": "text", + "content": "In sum, the main contributions in this paper include:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We propose a new paradigm for environment-agnostic object 6D pose estimation, by generating environment" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 322, + 72, + 553, + 95 + ], + "type": "text", + "content": "independent object representations rather than estimating poses directly from input images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 96, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 314, + 96, + 553, + 156 + ], + "type": "text", + "content": "- We design a novel Environment Decoupling Diffusion Model which employs an explicit environmental decoupling mechanism via gradient modelling and conditional generation to separate object representations from environment, while enabling efficient several-step sampling." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 156, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 156, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 314, + 156, + 553, + 239 + ], + "type": "text", + "content": "- We conduct extensive experiments on four public datasets including Linemod, Linemod-Occluded, YCB-V, HomebrewedDB and a self-made dataset DiverseScenes, showing that our method surpasses current state-of-the-art self-supervised methods. Our method exhibits remarkable generalization capabilities, even in previously unseen scenes, which is crucial for practical applications." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 251, + 398, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 398, + 262 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 398, + 262 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 553, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 568 + ], + "type": "text", + "content": "Fully-Supervised 6D Object Pose Estimation. Early approaches rely on local or global features to establish correspondences between keypoints on CAD models [1, 3, 4]. More recently, learning-based methods, particularly those utilizing Convolutional Neural Networks (CNNs), have become dominant. These methods extract deep features and establish mappings from RGB images to 6D object pose [14, 17, 41, 44]. Due to the inherent nonlinearity of the rotation space, the accuracy of directly regressed object pose is often limited [21]. To enhance accuracy, many methods employ CNNs to establish 2D-3D correspondences, followed by solving the object pose using the Perspective-n-Point (PnP) algorithm [19]. These methods can be categorized into sparse and dense correspondence methods. Sparse methods [28, 33, 38, 45] predict the 2D projections of predefined 3D keypoints, while dense methods [22, 23, 27, 35, 41, 48] predict 3D coordinates on the object model corresponding to image pixels thus providing better robustness to occlusions and clutter. However, fully-supervised 6D pose estimation methods are constrained by the need for accurate and diverse training data, which is often scarce due to the labor-intensive nature of precise annotation and the limited variation in available datasets, resulting in poor generalization to real-world and uncontrolled environments." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 553, + 713 + ], + "type": "text", + "content": "Self-Supervised 6D Object Pose Estimation. Considering the tremendous effort to collect large amounts of annotations for 6D object pose, some methods [26, 29, 36] have turned to using synthetic images for training. However, because of the domain gap between real and synthetic images, the performances of these methods are significantly lower compared to the methods that use real images with object poses for training. To narrow the domain gap, a few recent self-supervised methods [5, 20, 37, 39, 40, 43, 46, 49] have begun to incorporate both synthetic images with known object poses and unannotated real images into their training datasets. However, in diverse real-world application sce" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8679" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 215 + ], + "type": "text", + "content": "narios, models must contend with a variety of backgrounds, lighting conditions, and other environmental factors that are difficult to be adequately included in the training dataset. These environmental factors can significantly alter the appearance of objects within images. As a result, features learned by the model for the same object may vary across different scenes, leading to diminished generalization capabilities when encountering new scenes. Different from self-supervised learning methods, our proposed method directly generates environment-independent object representations to fundamentally eliminate the impact of environmental disturbances." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 217, + 298, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 298, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 298, + 432 + ], + "type": "text", + "content": "Diffusion-based Models for Vision Tasks. Originally introduced for image generation, diffusion models [7, 11, 34] have demonstrated impressive generation capabilities and have been explored in various tasks, including edge detection [47], image inpainting [25], and 6D pose estimation [9, 42, 45]. DiffusionEdge [47] presents a novel diffusion-based edge detector capable of generating precise and sharp edge maps without post-processing. 6D-diff [45] formulates object keypoints detection as a reverse diffusion process and employs a MoC-based forward process to leverage distribution priors in intermediate representations. Method [42] introduces three novel aggregation networks designed to effectively aggregate diffusion features, showing superior generalizability for object pose estimation. Currently, no existing work has yet explored the potential of diffusion models to generate environment-independent object representations, which could significantly improve the robustness of pose estimation in diverse environments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 111, + 458 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 467, + 123, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 467, + 123, + 478 + ], + "spans": [ + { + "bbox": [ + 55, + 467, + 123, + 478 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 485, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 295, + 675 + ], + "type": "text", + "content": "Traditional pose estimation methods typically estimate pose directly from images. Environmental factors such as background clutter, illumination variations, and occlusion in images can limit the accuracy and robustness of pose estimation. Pursuing a different paradigm, we propose a diffusion-based method that generates environment-independent object representations for pose estimation. Traditional diffusion models are predominantly utilized for image generation tasks, where they excel at creating new synthetic images. These models operate on the principle of gradually adding noise to an input image and then learning to reverse this process, effectively denoising the image over multiple iterations to recover the original content. In this paper, we extend the concept of diffusion models to a completely new domain by recovering environment-independent object representations for pose estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The overall framework of the proposed method EA6D is shown in Figure 2. The first stage is environment-independent object representation encoding. A synthesized" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": "pure object image with the same pose as the input image, but free from background, lighting, and occlusion, is encoded into a latent space using a Variational Autoencoder (VAE) [18]. This encoding results in an object representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "g_0 \\in R^{64 \\times 64 \\times 3}" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": ", which can be decoded back to the original pure object image by the decoder of VAE network, demonstrating that the representation contains only information about the object independent of the environment. Following this, a forward diffusion process is applied to " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "g_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": " to generate a noisy representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": ", which can be split into two sub-processes: the clean object representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "g_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": " to input image latent representation " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": ", meanwhile Gaussian noise is incrementally added. To denoise and recover the environment-independent object representations from " + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "inline_equation", + "content": "g_t" + }, + { + "bbox": [ + 313, + 72, + 555, + 384 + ], + "type": "text", + "content": ", an Environment Decoupling Diffusion Model (EDDM) is introduced. A conditional encoder extracts multi-level features from the input images. These features are then fused with the features of the U-Net encoder and are input to U-Net decoders. The EDDM outputs a decoupled object representation, which potentially contains object pose information, independent of environmental factors. Finally, a pose decoder predicts the pose information, including rotation quaternion and translation vector. The inference process in EA6D is thus a combination of diffusion-based denoising and pose decoding, enabling robust and accurate pose estimation in diverse and complex environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 396, + 555, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 396, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 396, + 555, + 422 + ], + "type": "text", + "content": "3.2. Environment Decoupling Diffusion for Generating Object Representations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 556, + 715 + ], + "type": "text", + "content": "Current 6D pose estimation methods often face robustness challenges in dynamic or unseen environments, as their object representations are influenced by complex environmental variations, resulting in multimodal distributions of features. CNNs-based methods typically use a single model to fit the entire data distribution. When facing multimodal distributions, the single model struggles to capture all the different patterns simultaneously, thereby resulting in a notable decrease in performance in new environments. To avoid the negative impact, we propose to extract environment-independent representations, enabling more accurate and generalizable pose estimation. Recent diffusion models show remarkable capabilities for modeling complex multimodal distribution, which is helpful for learning about environment-independent representations. Therefore, we propose to construct our framework based on the diffusion models to cover the complex distribution of object representations. More importantly, to acquire the environment-independent component from the input image distribution, we propose a novel environment decoupling diffusion model for generating environment-independent object representations inspired by the decoupled diffusion models [15, 47]. Different from the vanilla decoupled diffusion process that focuses on splitting the image and the" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8680" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 82, + 549, + 247 + ], + "blocks": [ + { + "bbox": [ + 61, + 82, + 549, + 247 + ], + "lines": [ + { + "bbox": [ + 61, + 82, + 549, + 247 + ], + "spans": [ + { + "bbox": [ + 61, + 82, + 549, + 247 + ], + "type": "image", + "image_path": "bec0c1acf76e5faa60c7e4db7041f7da3ef6f44ef832801d32b4cb7bef0d4f39.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 324 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed method. In training phase, the proposed method involves training a Variational Autoencoder (VAE) to encode pure object images into environment-independent representations. These representations are then sampled to create noisy representations, which are processed by an Environment Decoupling Diffusion Model (EDDM). The EDDM uses a dual structure with one UNet encoder and two UNet decoders to decouple the environment-independent object representations from the input image with complex environments. The pose decoder is trained to predict the pose from the representations, enabling robust pose estimation across various environments. In testing phase, a random Gaussian noise is fed into the EDDM without VAE." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 331, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 294, + 380 + ], + "type": "text", + "content": "noise components, we aim to decouple the environment-independent object representations from the input image with complex environments. To this end, we introduce the improved forward and reversed diffusion processes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 384, + 156, + 395 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 156, + 395 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 156, + 395 + ], + "type": "text", + "content": "3.2.1.Forward Process" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 399, + 296, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 399, + 296, + 591 + ], + "spans": [ + { + "bbox": [ + 54, + 399, + 296, + 591 + ], + "type": "text", + "content": "The vanilla decoupled diffusion process reformulates the conventional image-to-noise process into two stages: image-to-zero and zero-to-noise [15]. Corresponding to current object representation generation, there will be a clean object representation to zero process and a zero to noise process. However, such a paradigm only describes how to learn the environment-independent object representation from noise, without a clue on how to decouple it from the input images with complex environments. To address this problem, this paper proposes an environment decoupling diffusion process to intrinsically decouple the environment-independent object representation from the input image. We follow [30] to conduct the diffusion process in the latent space, i.e., our goal is to obtain the environment-independent object latent representations from the input image that mixes the object with the complex environments:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 86, + 607, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 607, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 86, + 607, + 295, + 635 + ], + "type": "interline_equation", + "content": "q \\left(g _ {t} \\mid g _ {0}\\right) = \\mathcal {N} \\left(g _ {t}; g _ {0} + \\int_ {0} ^ {t} \\left(r _ {0} - g _ {0}\\right) d t, t I\\right), \\tag {1}", + "image_path": "8746da7a0f443187b6bab63413f78b66e909da02dc0de8c4ccdfcf1586e7d280.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{0}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " denote the clean and noisy object latent representations respectively, " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "t\\in [0,1]" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " is the time step, " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " denotes the latent representation of input image. (A Variational Autoencoder features the latent representation.) According to the above equation, we can sample " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{t}" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{t} = g_{0} + \\int_{0}^{t}(r_{0} - g_{0})dt + \\sqrt{t} n," + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "inline_equation", + "content": "n\\in \\mathcal{N}(0,I)" + }, + { + "bbox": [ + 55, + 640, + 296, + 715 + ], + "type": "text", + "content": ". Ob" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": "viously, the forward process describes two sub-processes: " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "g_0 + \\int_0^t (r_0 - g_0)dt" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " is the clean object representation to input image latent representation and " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\sqrt{t} n" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": " is the zero to noise. It is worth noting that the first term contains the decay gradient from the input image to the environment-independent object representation, i.e., " + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "inline_equation", + "content": "r_0 - g_0" + }, + { + "bbox": [ + 313, + 331, + 555, + 477 + ], + "type": "text", + "content": ". We call this term the environment decoupling gradient, which guides the generation from the input image representation to object representation. The environment decoupling gradient provides explicit evidence for our goal of decoupling the environment-independent object representations, allowing for accurate and efficient generation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 416, + 495 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 416, + 495 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 416, + 495 + ], + "type": "text", + "content": "3.2.2. Reversed Process" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "text", + "content": "The goal of the reversed process is to gradually recover the clean object representation " + }, + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "inline_equation", + "content": "g_{0}" + }, + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "text", + "content": " from the noisy input " + }, + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "inline_equation", + "content": "g_{1}" + }, + { + "bbox": [ + 313, + 498, + 554, + 547 + ], + "type": "text", + "content": ". Following the derivation of [15], we can formally represent the reversed process as a conditional probability:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 322, + 555, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 555, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 322, + 555, + 553, + 617 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} q \\left(g _ {t - \\triangle t} \\mid g _ {t}, g _ {0}\\right) = \\mathcal {N} \\left(g _ {t} + \\int_ {t} ^ {t - \\triangle t} \\left(r _ {0} - g _ {0}\\right) d t - \\frac {\\triangle t}{\\sqrt {t}} n, \\right. \\\\ \\frac {\\triangle t (t - \\triangle t)}{t} I), \\tag {2} \\\\ \\end{array}", + "image_path": "8ab57f9e248b9918f9c8335b811737105edd1a145e8eadf5be91e07eab80556b.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\triangle t" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " is the smallest time step. Inheriting DDM, the reversed process benefits from the analytic gradient " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "r_0 - g_0" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ", enabling approximate analytic solution for few-step generation (even with only one step). Different from general diffusion models that recover the goal content from a normal distribution, our reversed process aims to generate object representations from the combination of the latent image representations and the noise. We can easily obtain the ini" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 61, + 71, + 135, + 82 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 71, + 135, + 82 + ], + "spans": [ + { + "bbox": [ + 61, + 71, + 135, + 82 + ], + "type": "text", + "content": "- - - - → only training" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8681" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "tial variable in the reversed process according to the forward process, i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "g_{1} = r_{0} + n" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". Intuitively, the generation process starts with the mixture of " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "r_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " instead of pure noise, which provides important prior information, thus speeding up the inference procedure and improving the generation quality. The reversed conditional probability " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "q(g_{t - \\triangle t}|g_t,g_0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " includes " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "r_0 - g_0" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", which are unknown in the reversed process. Therefore, we need to use a neural network " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " to parameterize " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "q(g_{t - \\triangle t}|g_t,g_0)" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ". Practically, the neural network has two outputs: " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "d_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "inline_equation", + "content": "n_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": ", corresponding to the two unknown terms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 209, + 178, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 209, + 178, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 209, + 178, + 220 + ], + "type": "text", + "content": "3.2.3. Network Architecture" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 296, + 331 + ], + "type": "text", + "content": "As shown in Figure 2, we construct our network based on a U-Net architecture including one encoder and two decoders. A pre-trained Swin Transformer [24] serves as a conditional encoder, extracting multi-level features from the input images. These features are then fused with the features of the U-Net encoder at corresponding levels through attention mechanisms, acting as inputs for the U-Net decoders. In practice, the input images will be obtained by an off-the-shelf object detector such as Yolo-v4 [2]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 270, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 270, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 270, + 351 + ], + "type": "text", + "content": "3.3. Pose Decoder for Object Representations" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": "The pose decoder is designed to predict the 6D pose, including both the rotation and translation, from the object representations generated from EDDM. The decoder consists of three modules: a ResNet block to further extract the features of " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "\\hat{g}_0^t" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": ", a rotation head for rotation estimation, and a translation head for translation prediction. The ResNet block enhances feature extraction through three convolutional layers. The first layer employs a " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": " kernel to expand the feature dimension, capturing richer information. The second layer uses a " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": " kernel for spatial feature reduction. The third layer also uses a " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "1 \\times 1" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": " kernel. The rotation head is designed with a series of " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": " convolutional layers, followed by a generalized average pooling layer and a fully connected layer, culminating in the prediction of the rotation quaternion " + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 356, + 296, + 534 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "text", + "content": "For 3D translation, since directly regressing the translation in 3D space does not work well in practice, we adopt Scale-invariant Translation Estimation (SITE) [22] to decouple the translation into the 2D location " + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "inline_equation", + "content": "(o_x,o_y)" + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "text", + "content": " of the projected 3D centroid and the object's distance " + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "inline_equation", + "content": "T_{z}" + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "text", + "content": " towards the camera. The translation head is constituted by a series of convolutional layers followed by fully connected layers, which predict the scale-invariant translation parameters " + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "inline_equation", + "content": "T_{site} = (\\Delta x,\\Delta y,\\Delta z)" + }, + { + "bbox": [ + 55, + 535, + 296, + 667 + ], + "type": "text", + "content": ". These parameters are then utilized to compute the object's 2D coordinates and depth, as delineated by the equations:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 129, + 673, + 296, + 712 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 673, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 129, + 673, + 296, + 712 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{l l} o _ {x} = & \\Delta x \\cdot w + c _ {x} \\\\ o _ {y} = & \\Delta y \\cdot h + c _ {y} \\\\ T _ {z} = & \\Delta z \\cdot r \\end{array} \\right. \\tag {3}", + "image_path": "94c23bae5a89ef3b9dd0db717e132e8df219642dbe2383bce7fbca3a9496101f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " represent the width and height of the bounding box, " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "c_{x}" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "c_{y}" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " denote the coordinates of the box's center, and " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " is the scaling factor derived from the image's dimensions. Given the camera intrinsics " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": ", the translation " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": " can be calculated via back-projection " + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "inline_equation", + "content": "T = K^{-1}T_{z}[o_{x}, o_{y}, 1]^{T}" + }, + { + "bbox": [ + 313, + 72, + 554, + 168 + ], + "type": "text", + "content": ". This method allows for a more accurate and robust estimation of the object's 3D translation, thereby enhancing the overall pose estimation accuracy." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 174, + 429, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 429, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 429, + 186 + ], + "type": "text", + "content": "3.4. Training Procedure" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 191, + 555, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 191, + 555, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 191, + 555, + 286 + ], + "type": "text", + "content": "Our training object is meticulously designed to refine the performance of the VAE, EDDM, and pose decoder within our framework. Initially, the VAE is trained to encode synthetic images, which are environment-free and rendered with poses corresponding to the input images, into latent features. These latent features are then decoded to reconstruct the original synthesized images, ensuring the network's ability to capture the essential features of the object." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 287, + 554, + 323 + ], + "type": "text", + "content": "After the VAE's training, we proceed to train the EDDM while keeping the VAE's weights fixed. EDDM is trained following the training strategy and objective of DDM [15]:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 347, + 334, + 553, + 348 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 334, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 347, + 334, + 553, + 348 + ], + "type": "interline_equation", + "content": "L _ {e d d m} = \\left\\| d _ {\\theta} - \\left(r _ {0} - g _ {0}\\right) \\right\\| _ {2} + \\left\\| n _ {\\theta} - n \\right\\| _ {2} \\tag {4}", + "image_path": "c5fc655a0e17095461f668283b96bdec82c7d76187626d3b312c5b5554fe2f7c.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 350, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 554, + 385 + ], + "type": "text", + "content": "where the first term and second term correspond to the supervision of predicting noise component and image component, respectively." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": "At last, we freeze VAE and EDDM to train the pose decoder. We employ a disentangled 6D pose loss via individually supervising the rotation " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": ", the scale-invariant 2D object center " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "(\\Delta x, \\Delta y)" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": " and the distance " + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "inline_equation", + "content": "\\Delta z" + }, + { + "bbox": [ + 313, + 385, + 554, + 434 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 375, + 439, + 553, + 452 + ], + "type": "interline_equation", + "content": "L _ {\\text {p o s e}} = L _ {R} + L _ {\\text {c e n t e r}} + L _ {z} \\tag {5}", + "image_path": "646af9c64bb52ef04e7e7d7db246a5bd4d1757e4d02ef1c126be367b750f183d.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 457, + 353, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 353, + 468 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 353, + 468 + ], + "type": "text", + "content": "Thereby," + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 472, + 553, + 513 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 472, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 339, + 472, + 553, + 513 + ], + "type": "interline_equation", + "content": "\\left\\{ \\begin{array}{r c l} L _ {R} & = & \\left\\| R - \\bar {R} \\right\\| _ {2} \\\\ L _ {c e n t e r} & = & \\| \\Delta x - \\Delta \\bar {x} \\| _ {2} + \\| \\Delta y - \\Delta \\bar {y} \\| _ {2} \\\\ L _ {z} & = & \\| \\Delta z - \\Delta \\bar {z} \\| _ {2} \\end{array} \\right. \\tag {6}", + "image_path": "339399f152ec4d3d1e86dda7f8ec5d2660ffdb589ce6e4155a8c4dabc9c8e9ff.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "inline_equation", + "content": "(\\Delta x, \\Delta y, \\Delta z, R)" + }, + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "inline_equation", + "content": "(\\Delta \\bar{x}, \\Delta \\bar{y}, \\Delta \\bar{z}, \\bar{R})" + }, + { + "bbox": [ + 313, + 516, + 554, + 540 + ], + "type": "text", + "content": " denote prediction and ground truth, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 541, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 647 + ], + "type": "text", + "content": "In the inference phase, object images from different environments are fed into the conditional encoder as cue information for the EDDM. The EDDM systematically eliminates environmental factors from the noisy latent space, yielding a representation that encapsulates solely the object's information. This purified representation is then fed into the pose decoder to extract the object pose, thereby achieving robust generalization across various scenes and conditions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 658, + 395, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 395, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 395, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "In this section, we conduct experiments to verify the generalization capability of the proposed method in various environments. Therefore we follow existing works [5, 37] in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8682" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 156 + ], + "type": "text", + "content": "self-supervised object pose estimation which focus on narrowing the domain gap and enhancing the network's generalization ability across different domains. We compare our method to the state-of-the-art methods on four of the core datasets of the BOP challenge [13] and a self-made dataset. We also provide a detailed analysis of our proposed scheme in ablation study." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 164, + 118, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 164, + 118, + 175 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 118, + 175 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 182, + 296, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 182, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 296, + 337 + ], + "type": "text", + "content": "We evaluate our EA6D on four of the core BOP datasets [13], Linemod(LM) [10], Linemod-Occluded(LM-O) [3], HomebrewedDB(HB) [16] and YCB-V [44]. LM, LM-O, and HB include the same objects under different environments, and YCB-Video involves environmental changes, making them suitable for verifying the proposed method. HB is typically used to study the generalization ability of methods towards new scenes with changing illumination. However, the images in HB are from a single indoor environment. To expand the testing of unseen scenarios, the DiverseScenes Dataset is created by simulating both indoor and outdoor environments, to rigorously evaluate the model's ability to generalize across varied environments." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 479 + ], + "type": "text", + "content": "DiverseScenes Dataset. We select four distinct real-world environments, spanning both indoor and outdoor settings, and employ Physically Based Rendering (PBR) techniques to render the 3D models into these backgrounds. To capture the inherent diversity of real-world lighting, we apply a range of lighting conditions, thus amplifying the disparity with the training data and crafting previously unseen test scenarios. For each object category, we render 150 images per environment, amassing a total of 1950 images. Figure 3 illustrates some samples of the dataset. We give more samples of the dataset in the supplementary material and will release the dataset." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 56, + 483, + 296, + 626 + ], + "blocks": [ + { + "bbox": [ + 56, + 483, + 296, + 626 + ], + "lines": [ + { + "bbox": [ + 56, + 483, + 296, + 626 + ], + "spans": [ + { + "bbox": [ + 56, + 483, + 296, + 626 + ], + "type": "image", + "image_path": "7607c70aac8adbfc83d3a0029a09539cac16c89334eb0284f0754c1111ac35a2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 637, + 295, + 670 + ], + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 670 + ], + "type": "text", + "content": "Figure 3. We simulated different indoor and outdoor backgrounds with varying lighting to construct four distinct scenarios, which have a significant difference from the training data." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "It should be emphasized that EA6D does not require any real images for training. For the evaluation of EA6D, all images in the above dataset are used for testing only." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 373, + 83 + ], + "type": "text", + "content": "4.2. Metrics" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "text", + "content": "We report our results referring to the common Average Distance of Distinguishable Model Points (ADD) metric [10]. This metric reports 6D pose error by transforming all object vertices with estimated pose and ground-truth pose and measuring the average distance of the two sets of points. The estimated pose is considered to be correct if the average distance is smaller than " + }, + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "text", + "content": " of the object's diameter. For symmetric objects, we use the ADD-S metric [44], where the mean distance is computed based on the closest point distance. For YCB-V, we report the AUC (area under curve) of ADD(-S) with a maximum threshold of " + }, + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 89, + 555, + 222 + ], + "type": "text", + "content": " [44]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 228, + 447, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 447, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 447, + 241 + ], + "type": "text", + "content": "4.3. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "spans": [ + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "content": "Our method is implemented by Pytorch, and all experiments are conducted on an NVIDIA 4090 GPU. We use Synthetic PBR Dataset [13] to train the 2D object detector Yolov4 [2] and the entire network. For each image, we utilize the object's CAD model to render it in the same pose onto a blank image to obtain the pure object image. We use pure object images to train VAE and use them as ground-truth for corresponding input PBR images to train EDDM from scratch. We train VAE using AdamW optimizer with an attenuated learning rate (from " + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "inline_equation", + "content": "1e^{-4}" + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "inline_equation", + "content": "5e^{-7}" + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "content": "). We train the diffusion models using AdamW optimizer with an attenuated learning rate (from " + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "inline_equation", + "content": "2e^{-4}" + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "inline_equation", + "content": "5e^{-6}" + }, + { + "bbox": [ + 313, + 246, + 556, + 450 + ], + "type": "text", + "content": ") for 20k iterations with a batch size of 12 and each training takes up about 30 GPU hours. The pose decoder is trained with the same optimizer setting for 10k iterations with a batch size of 24 and each training takes about 10 GPU hours. The conditional encoder is pre-trained on ImageNet [6]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 456, + 452, + 469 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 452, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 452, + 469 + ], + "type": "text", + "content": "4.4. Comparative Evaluation" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 474, + 556, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 556, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 556, + 688 + ], + "type": "text", + "content": "Here, we evaluate the proposed EA6D on the LM, LM-O and YCB-V in comparison to some state-of-the-art methods including six fully-supervised methods that are trained with annotated real images with object poses (DPOD [48], PVNet [28], SO-Pose [8]), ZebraPose [35], CheckerPose [23] and 6D-diff [45], two self-supervised methods that are trained with only synthetic data (AAE [36], MHP [26]), three self-supervised methods that are trained with both synthetic data and unannotated real images (DSC-PoseNet [46], SMOC-Net [37], TexPose [5]), and three self-supervised methods that are trained with synthetic data, unannotated real images and depth images (Self6D [39], Self6D++ [40], RKHSPose [43]). In addition, for a more comprehensive comparison, we also compare our method with Self6D++ [40] and RKHSpose [43] trained only with synthetic data. Table 1 reports the average metrics on each dataset and we provide detailed results in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Comparisons on Linemod. As seen from Table 1, there is a significant performance gap between fully-supervised" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8683" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 173, + 296, + 384 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": "Table 1. Comparison with state-of-the-art methods on LM, LM-O and YCB-V. The table reports results for the Average Recall (\\%) of ADD(-S) on LM and LM-O, and AUC of ADD-S and ADD(-S) on YCB-V. All results except ours are copied from SMOC-Net [37], TexPos [5] and RKHSPose [43]. " + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": ": annotated real RGB data. " + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": ": synthetic RGB data. " + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "inline_equation", + "content": "R^{-}" + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": ": unannotated real RGB data. " + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 70, + 296, + 170 + ], + "type": "text", + "content": ": depth data. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold. (-) denotes results missing from the original paper." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 173, + 296, + 384 + ], + "lines": [ + { + "bbox": [ + 58, + 173, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 58, + 173, + 296, + 384 + ], + "type": "table", + "html": "
MethodsTraining dataLMLM-OYCB-V
ADD(-S)ADD-S AUCADD(-S) AUC
DPOD [48]R+S82.647.3--
PVNet [28]86.340.8-73.8
SO-Pose [8]96.062.390.983.9
ZebraPose [35]-76.990.185.3
CheckerPose [23]-77.591.386.4
6D-diff [45]-79.691.587.0
Self6D [39]S+R−+D58.932.1--
Self6D++ [40]88.564.791.480.0
RKHSPose [43]95.968.792.683.0
DSC-PoseNet [46]S+R−54.721.9--
Self6D++ [40]76.059.8--
SMOC-Net [37]91.363.3--
TexPose [5]91.766.7--
AAE [36]S31.4---
MHP [26]38.8---
Self6D++ [40]77.452.989.477.8
RKHSPose [43]78.254.390.276.5
EA6D96.387.893.188.3
", + "image_path": "a8c22942ce88817d8be16c1bf795da4fc21f6b746236a7bc2415f1fc5aaa01eb.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 390, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 295, + 544 + ], + "type": "text", + "content": "methods and approaches trained with synthetic data instead. Self-supervised approaches have proven to be able to almost completely close the domain gap, yet typically require additional modalities (unannotated real RGB data, depth data) to achieve comparable results. The proposed EA6D achieves a better performance than all the comparative methods (regardless of whether these methods are trained in either fully-supervised or self-supervised manner), and only uses synthetic images for training. This is attributed to EA6D generating environment-independent object representations for pose estimation rather than estimating poses directly from input images, avoiding the impact of domain gap on network performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": "Comparisons on Linemod-Occluded. In the inference stage, each object in an image will be inferred independently. The corresponding results are reported in Table 1 which shows that EA6D outperforms all these methods, demonstrating strong robustness against occlusion. Note that EA6D only uses synthetic images as training data and still outperforms fully supervised methods (6D-diff by " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "8.3\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", CheckerPose by " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "10.3\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ", ZebraPose by " + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "inline_equation", + "content": "10.9\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 641 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 641, + 296, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 701 + ], + "type": "text", + "content": "Comparisons on YCB-V. As shown in Table 1, EA6D achieves the best performance on both the AUC of ADD-S and the AUC of ADD(-S) metrics, showing that EA6D is capable of handling a wide variety of object appearances and lighting conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Comparisons on Unseen Environments. In this experi" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 238 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 238 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 238 + ], + "type": "image", + "image_path": "8ebdd9dbdd030d822b529f54a793e1b7773114396fe5567fa9bdc939ca8f4b2c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 245, + 555, + 322 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 322 + ], + "type": "text", + "content": "Figure 4. Qualitative results on Linemod (top-left), Linemod-Occluded (bottom) and HomebrewedDB (top-right). For each dataset, Top: input images and the visulizations of pose, green and blue bounding boxes represent GT poses and estimated poses, respectively. Middle: Visualization of generated environment-independent object representations. Bottom: The image after decoding the environment-independent object representations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 380, + 555, + 451 + ], + "blocks": [ + { + "bbox": [ + 313, + 326, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 313, + 326, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 313, + 326, + 555, + 369 + ], + "type": "text", + "content": "Table 2. Evaluation results on HomebrewedDB dataset. All results except ours are copied from TexPose [5]. The best pose method using the same kind of training data is underlined, and the overall best method is marked in bold." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 380, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 317, + 380, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 317, + 380, + 555, + 451 + ], + "type": "table", + "html": "
Training dataR− + SS
MethodSock[32]Self6D++[40]Texpose[5]GDR[41]EA6D
Benchvise57.375.792.988.897.6
Driller46.689.494.292.893.8
Phone41.576.881.278.792.6
Mean52.080.689.486.894.7
", + "image_path": "adb1f1773f9583a386ad34504d1e340839dfcbd91481b5979e179d28ffb3fcb7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 489, + 553, + 564 + ], + "blocks": [ + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 480 + ], + "type": "text", + "content": "Table 3. Evaluation results on DiverseScenes dataset. All cited methods are re-implemented using their open source codes." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 489, + 553, + 564 + ], + "lines": [ + { + "bbox": [ + 317, + 489, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 553, + 564 + ], + "type": "table", + "html": "
MethodsSelf6D++ [40]GDR [41]Texpose [5]EA6D
Scene 156.564.182.084.4
Scene 262.270.776.483.9
Scene 348.668.966.583.0
Scene 469.554.364.384.0
Mean59.264.572.383.8
", + "image_path": "bab49e680274139263050f773c7020a4a9a17bb3d9b2a44ebb99e158d100df75.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "ment, we want to study the generalizability of our method towards new scenes. First, we evaluate our method on HomebrewedDB. From Table 2, we observe that EA6D outperforms all other methods and achieves the same results as testing in Linemod (95.7% vs.94.7%). Self6D++ and TexPose undergo a clear performance drop when testing on Linemod. Self6D++ drops from 94.1% to 80.6% and TexPose drops from 91.8% to 89.4%. It indicates that the generalization performance of existing self-supervised methods is limited by the unannotated real images used for training. Further, Table 3 shows the results of the test on the DiverseScenes dataset. In environments with significant changes" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8684" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "in background and lighting, there is a certain decline in performance for all four methods, but our method is the least affected and outperforms the others by a huge leap." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 117, + 175, + 147 + ], + "blocks": [ + { + "bbox": [ + 102, + 109, + 132, + 117 + ], + "lines": [ + { + "bbox": [ + 102, + 109, + 132, + 117 + ], + "spans": [ + { + "bbox": [ + 102, + 109, + 132, + 117 + ], + "type": "text", + "content": "Scene 1" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 117, + 175, + 147 + ], + "lines": [ + { + "bbox": [ + 59, + 117, + 175, + 147 + ], + "spans": [ + { + "bbox": [ + 59, + 117, + 175, + 147 + ], + "type": "image", + "image_path": "024e513529f972ef0c2f00d14fd22be4a5b401b83b9cf5b8f08bbd7248df3bca.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 176, + 117, + 293, + 147 + ], + "blocks": [ + { + "bbox": [ + 176, + 117, + 293, + 147 + ], + "lines": [ + { + "bbox": [ + 176, + 117, + 293, + 147 + ], + "spans": [ + { + "bbox": [ + 176, + 117, + 293, + 147 + ], + "type": "image", + "image_path": "d97c730d4bb633fe2c5fb577054219dc46d1e484ce9a96b43ecad6ca72706461.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 156, + 176, + 186 + ], + "blocks": [ + { + "bbox": [ + 103, + 148, + 133, + 156 + ], + "lines": [ + { + "bbox": [ + 103, + 148, + 133, + 156 + ], + "spans": [ + { + "bbox": [ + 103, + 148, + 133, + 156 + ], + "type": "text", + "content": "Scene 3" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 156, + 176, + 186 + ], + "lines": [ + { + "bbox": [ + 59, + 156, + 176, + 186 + ], + "spans": [ + { + "bbox": [ + 59, + 156, + 176, + 186 + ], + "type": "image", + "image_path": "d8fcb9179be1e02562e51c14037d4b92e90c6f8ab3ba100c1a398f84f14bc7ba.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 177, + 156, + 293, + 186 + ], + "blocks": [ + { + "bbox": [ + 219, + 148, + 249, + 156 + ], + "lines": [ + { + "bbox": [ + 219, + 148, + 249, + 156 + ], + "spans": [ + { + "bbox": [ + 219, + 148, + 249, + 156 + ], + "type": "text", + "content": "Scene 4" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 177, + 156, + 293, + 186 + ], + "lines": [ + { + "bbox": [ + 177, + 156, + 293, + 186 + ], + "spans": [ + { + "bbox": [ + 177, + 156, + 293, + 186 + ], + "type": "image", + "image_path": "7660107b8d885ac4cde7ff7c3ffe51a78f248b6681f9b2a3c1c10d12e4381ac6.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 59, + 196, + 127, + 230 + ], + "blocks": [ + { + "bbox": [ + 103, + 187, + 249, + 196 + ], + "lines": [ + { + "bbox": [ + 103, + 187, + 249, + 196 + ], + "spans": [ + { + "bbox": [ + 103, + 187, + 249, + 196 + ], + "type": "text", + "content": "Same object in 7 different environments" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 196, + 127, + 230 + ], + "lines": [ + { + "bbox": [ + 59, + 196, + 127, + 230 + ], + "spans": [ + { + "bbox": [ + 59, + 196, + 127, + 230 + ], + "type": "image", + "image_path": "966e391bde2b3ad4051238f8728b4766afbaed64c69fcb86a495be0f2a82c9a8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 238, + 295, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 238, + 295, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 295, + 281 + ], + "type": "text", + "content": "Figure 5. The top two rows are qualitative results on DiverseScenes Dataset. Each scene has different backgrounds and lighting conditions. The bottom row is qualitative results of a same object in 7 different environments." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 127, + 196, + 209, + 230 + ], + "blocks": [ + { + "bbox": [ + 127, + 196, + 209, + 230 + ], + "lines": [ + { + "bbox": [ + 127, + 196, + 209, + 230 + ], + "spans": [ + { + "bbox": [ + 127, + 196, + 209, + 230 + ], + "type": "image", + "image_path": "d3fe39d75992438083932905c4ba4743681b170ab3dfa80eb1c5f8fe6c853f79.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 209, + 196, + 293, + 230 + ], + "blocks": [ + { + "bbox": [ + 209, + 196, + 293, + 230 + ], + "lines": [ + { + "bbox": [ + 209, + 196, + 293, + 230 + ], + "spans": [ + { + "bbox": [ + 209, + 196, + 293, + 230 + ], + "type": "image", + "image_path": "874bed435b455bed5ce92be74c83876a00fb293ed0441a1d1d321c9f6c1296c8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 285, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 404 + ], + "type": "text", + "content": "We would like to emphasise that since our method is only trained on Synthetic PBR Dataset [13], the aforementioned datasets are all unseen environments for our method. Our method maintains accurate predictions (average " + }, + { + "bbox": [ + 55, + 285, + 296, + 404 + ], + "type": "inline_equation", + "content": "90.6\\%" + }, + { + "bbox": [ + 55, + 285, + 296, + 404 + ], + "type": "text", + "content": ") under different backgrounds, lighting conditions, and occlusions. It shows that the choice of adaptation domain (LM, HB, or other environments) has a marginal effect on our final performance. We attribute this to the generation of pure object features, eliminating the domain gap problem caused by environmental factors." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 410, + 138, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 138, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 138, + 422 + ], + "type": "text", + "content": "4.5. Visulizations" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 427, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 619 + ], + "type": "text", + "content": "In this section, we will analyze our work by visualization of the test results. Figure 4 illustrates the visualizations of environment-independent object representations and pose estimation. It clearly shows that the images generated by the diffusion model filter out all environmental factors while accurately preserving the object information. The visualizations of results on LM-O explain why EA6D can effectively handle occlusions. With the strong generative capabilities of diffusion models, even when facing severe occlusion, the model is still able to accurately generate occluded parts. Figure 5 shows the visualizations of the results on the DiverseScenes Dataset and the visualizations of the same object in 7 different environments. These provide a clear demonstration of its capabilities in handling various environments. We give more visulizations results in supplementary material." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 156, + 635 + ], + "type": "text", + "content": "4.6. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 713 + ], + "type": "text", + "content": "The effect of diffusion models and diffusion steps. We study the impact of different diffusion models on EA6D performance. We replace EDDM with traditional Denoising Diffusion Probabilistic Model (DDPM) [11] and Decoupled Diffusion Model (DDM) [15] and compare the results on Linemod. Also, the number of iterating steps is" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "another key parameter in diffusion models that determines the inference speed, which is essential for pose estimation. All the results are reported in Table 4. We can observe that the EA6D with EDDM achieves more accurate results with fewer iterating steps. This result verifies the effectiveness of mixing of input image latent representations and noise instead of pure noise in generation process, which provides prior information, thus speeding up the inference procedure and improving the generation quality." + } + ] + } + ], + "index": 18 + }, + { + "type": "table", + "bbox": [ + 326, + 210, + 541, + 317 + ], + "blocks": [ + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "lines": [ + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 185, + 553, + 207 + ], + "type": "text", + "content": "Table 4. Mean results with ablation EA6D and the number of iterating steps on Linemod." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 326, + 210, + 541, + 317 + ], + "lines": [ + { + "bbox": [ + 326, + 210, + 541, + 317 + ], + "spans": [ + { + "bbox": [ + 326, + 210, + 541, + 317 + ], + "type": "table", + "html": "
Method\\steps1510201001000
EA6D-DDPM [11]003.614.352.778.4
EA6D-DDM [15]93.196.195.696.396.296.3
EA6D w/o VAE68.471.373.175.176.277.4
EA6D*76.779.279.279.479.379.5
EA6D96.396.396.496.596.596.5
Time(s)EA6D0.230.510.871.606.9892.1
", + "image_path": "670720be74df8c225bbb4b30cd575c5cd34f866fd3290128ec97dae3610030e8.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_body" + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 324, + 555, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 515 + ], + "type": "text", + "content": "The effect of object representations. In our framework, we generate the object representations via performing the denoising process to estimate object pose. The object representations are the latent code of pure object images encoded by VAE. To evaluate the efficacy of such representations, we first remove the VAE and directly generate the pure object images as object representations through EDDM. Then we use the framework of EA6D, but decode the generated object representations to pure object images using the decoder of VAE and input pose decoder to estimate pose. We denote this setting as EA6D*. As shown in Table 4, the performance of these two types of object representations significantly drops, showing that the latent code of pure object images captures the rich semantic and structural information in the image and is more suitable for pose estimation than the generated images." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 388, + 538 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a novel generative paradigm for 6D pose estimation. Different from typical RGB-based pose estimation frameworks which directly extract features from input images for pose estimation and are therefore easily influenced by environmental factors, we employ an environment decoupling model that separates the object from environmental factors during the denoising process, generating environment-independent object representations which are robust to environmental changes. A pose decoder is designed to estimate poses using these representations. The experimental results on five datasets demonstrate that EA6D significantly outperforms current state-of-the-art self-supervised methods, particularly in its ability to generalize to previously unseen scenarios." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8685" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 165, + 85 + ], + "type": "text", + "content": "6. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 213 + ], + "type": "text", + "content": "This research was supported by the National Key Research and Development Program of China (No.2024YFF0907600), the National Natural Science Foundation of China (Grant No.62273275), the China Postdoctoral Science Foundation (No.2022M722573), Shaanxi Postdoctoral Research Project (No.2023BSHEDZZ244), Shaanxi Natural Science Basic Research Program (No.2025JC-YBQN-928), Key Research and Development Plan of Shaanxi Province (2025GH-YBXM-018) and Xi'an Science and Technology Innovation and Qinchuangyuan Innovation Major Program (23ZDCYJSGG0009-2023)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 236, + 115, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 236, + 115, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 236, + 115, + 248 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 256, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 256, + 294, + 277 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 256, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 61, + 256, + 294, + 277 + ], + "type": "text", + "content": "[1] H. Bay, T. Tuytelaars, and L. Van Gool. Surf: Speeded up robust features. In ECCV, 2006. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 280, + 295, + 322 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 295, + 322 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 295, + 322 + ], + "type": "text", + "content": "[2] Alexey Bochkovskiy, Chien-Yao Wang, and Hong-Yuan Mark Liao. Yolov4: Optimal speed and accuracy of object detection. arXiv preprint arXiv:2004.10934, 2020. 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "spans": [ + { + "bbox": [ + 62, + 326, + 294, + 357 + ], + "type": "text", + "content": "[3] E. Brachmann, A. Krull, F. Michel, S. Gumhold, J. Shotton, and C. Rother. Learning 6d object pose estimation using 3d object coordinates. In ECCV, 2014. 1, 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 360, + 294, + 402 + ], + "type": "text", + "content": "[4] E. Brachmann, F. Michel, A. Krull, M.Y. Yang, S. Gumhold, and C. Rother. Uncertainty-driven 6d pose estimation of objects and scenes from a single rgb image. In CVPR, 2016. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 406, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 406, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 62, + 406, + 294, + 449 + ], + "type": "text", + "content": "[5] Hanzhi Chen, Fabian Manhardt, Nassir Navab, and Benjamin Busam. Texpose: Neural texture learning for self-supervised 6d object pose estimation. In CVPR, 2023. 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 451, + 294, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 451, + 294, + 483 + ], + "spans": [ + { + "bbox": [ + 62, + 451, + 294, + 483 + ], + "type": "text", + "content": "[6] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 486, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 486, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 486, + 294, + 518 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 294, + 563 + ], + "type": "text", + "content": "[8] Yan Di, Fabian Manhardt, Gu Wang, Xiangyang Ji, Nassir Navab, and Federico Tombari. So-posed: Exploiting self-occlusion for direct 6d pose estimation. In ICCV, 2021. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 566, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 566, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 566, + 294, + 599 + ], + "type": "text", + "content": "[9] Jia Gong, Lin Geng Foo, Zhipeng Fan, Qiuhong Ke, Hossein Rahmani, and Jun Liu. Diffpose: Toward more reliable 3d pose estimation. In CVPR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 643 + ], + "type": "text", + "content": "[10] S. Hinterstoisser, V. Lepetit, S. Ilic, S. Holzer, G. Bradski, K. Konolige, and N. Navab. Model based training, detection and pose estimation of texture-less 3d objects in heavily cluttered scenes. In ACCV, 2012. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 678 + ], + "type": "text", + "content": "[11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2, 3, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[12] Tomáš Hodan, P. Haluza, Štepan Obdržálek, J. Matas, M. Lourakis, and X. Zabulis. T-less: An rgb-d dataset for 6d pose estimation of texture-less objects. In WACV, 2017. 1" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[13] Tomás Hodař, Martin Sundermeyer, Bertram Drost, Yann Labbe, Eric Brachmann, Frank Michel, Carsten Rother, and Jiri Matas. Bop challenge 2020 on 6d object localization. In ECCV Workshops, 2020. 6, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 140 + ], + "type": "text", + "content": "[14] Yinlin Hu, Pascal Fua, Wei Wang, and Mathieu Salzmann. Single-stage 6d object pose estimation. In CVPR, 2020. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 142, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 142, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 142, + 553, + 174 + ], + "type": "text", + "content": "[15] Yuhang Huang, Zheng Qin, Xinwang Liu, and Kai Xu. Decoupled diffusion models with explicit transition probability. arXiv preprint arXiv:2306.13720, 2023. 3, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 207 + ], + "type": "text", + "content": "[16] Roman Kaskman, Sergey Zakharov, Ivan Shugurov, and Slobodan Ilic. Homebreweddb: Rgb-d dataset for 6d pose estimation of 3d objects. In ICCV Workshops, 2019. 1, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 317, + 209, + 553, + 241 + ], + "type": "text", + "content": "[17] W. Kehl, F. Manhardt, F. Tombari, S. Ilic, and N. Navab. Ssd-6d: Making rgb-based 3d detection and 6d pose estimation great again. In ICCV, 2017. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 243, + 553, + 264 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 264 + ], + "type": "text", + "content": "[18] DiederikP. KingmaandMaxWelling. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 266, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 297 + ], + "type": "text", + "content": "[19] V. Lepetit, F Moreno-Noguer, and P. Fua. Epnp: An accurate o (n) solution to the pnp problem. IJCV, 81(2):155-166, 2009. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 300, + 553, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 300, + 553, + 352 + ], + "spans": [ + { + "bbox": [ + 317, + 300, + 553, + 352 + ], + "type": "text", + "content": "[20] Fu Li, Shishir Reddy Vutukur, Hao Yu, Ivan Shugurov, Benjamin Busam, Shaowu Yang, and Slobodan Ilic. Nerfpose: A first-reconstruct-then-regress approach for weakly-supervised 6d object pose estimation. In ICCV, 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 387 + ], + "type": "text", + "content": "[21] Hongyang Li, Jiehong Lin, and Kui Jia. Dcl-net: Deep correspondence learning network for 6d pose estimation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 389, + 553, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 553, + 430 + ], + "type": "text", + "content": "[22] Zhigang Li, Gu Wang, and Xiangyang Ji. Cdpn: Coordinates-based disentangled pose network for real-time rgb-based 6-dof object pose estimation. In ICCV, 2019. 1, 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 465 + ], + "type": "text", + "content": "[23] Ruyi Lian and Haibin Ling. Checkerpose: Progressive dense keypoint localization for object pose estimation with graph neural network. In ICCV, 2023. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 468, + 553, + 510 + ], + "type": "text", + "content": "[24] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 5" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[25] Andreas Lugmayr, Martin Danelljan, Andres Romero, Fisher Yu, Radu Timofte, and Luc Van Gool. Repaint: Inpainting using denoising diffusion probabilistic models. In CVPR, 2022. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 610 + ], + "type": "text", + "content": "[26] Fabian Manhardt, Diego Martin Arroyo, Christian Rupprecht, Benjamin Busam, Tolga Birdal, Nassir Navab, and Federico Tombari. Explaining the ambiguity of object detection and 6d pose from visual data. In ICCV, 2019. 2, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 644 + ], + "type": "text", + "content": "[27] Kiru Park, Timothy Patten, and Markus Vincze. Pix2pose: Pixel-wise coordinate regression of objects for 6d pose estimation. In ICCV, 2019. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 678 + ], + "type": "text", + "content": "[28] S. Peng, Y. Liu, Q. Huang, X. Zhou, and H. Bao. Pynet: Pixel-wise voting network for 6dof pose estimation. In CVPR, 2019. 1, 2, 6, 7" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] M. Rad and V. Lepetit. Bb8: A scalable, accurate, robust to partial occlusion method for predicting the 3d poses of challenging objects without using depth. In ICCV, 2017. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8686" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[30] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjorn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 138 + ], + "type": "text", + "content": "[31] Ivan Shugurov, Sergey Zakharov, and Slobodan Ilic. Dpodv2: Dense correspondence-based 6 dof pose estimation. TPAMI, 2021. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 183 + ], + "type": "text", + "content": "[32] Juil Sock, Guillermo Garcia-Hernando, Anil Armagan, and Tae-Kyun Kim. Introducing pose consistency and warp-alignment for self-supervised 6d object pose estimation in color images. In 3DV, 2020. 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 294, + 215 + ], + "type": "text", + "content": "[33] Chen Song, Jiaru Song, and Qixing Huang. Hybridpose: 6d object pose estimation under hybrid representations. In CVPR, 2020. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 248 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[35] Yongzhi Su, Mahdi Saleh, Torben Fetzer, Jason Rambach, Nassir Navab, Benjamin Busam, Didier Stricker, and Federico Tombari. Zebrapose: Coarse to fine surface encoding for 6 dof object pose estimation. In CVPR, 2022. 1, 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[36] M. Sundermeyer, Z. Marton, M. Durner, M. Brucker, and R. Triebel. Implicit 3d orientation learning for 6d object detection from rgb images. In ECCV, 2018. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 358 + ], + "type": "text", + "content": "[37] Tao Tan and Qiulei Dong. Smoc-net: Leveraging camera pose for self-supervised monocular object pose estimation. In CVPR, 2023. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 381 + ], + "type": "text", + "content": "[38] B. Tekin, S.N. Sinha, and P. Fua. Real-time seamless single shot 6d object pose prediction. In CVPR, 2018. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 425 + ], + "type": "text", + "content": "[39] Gu Wang, Fabian Manhardt, Jianzhun Shao, Xiangyang Ji, Nassir Navab, and Federico Tombari. Self6d: Self-supervised monocular 6d object pose estimation. In ECCV, 2020. 1, 2, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[40] Gu Wang, Fabian Manhardt, Xingyu Liu, Xiangyang Ji, and Federico Tombari. Occlusion-aware self-supervised monocular 6d object pose estimation. TPAMI, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[41] Gu Wang, Fabian Manhardt, Federico Tombari, and Xi-angyang Ji. Gdr-net: Geometry-guided direct regression network for monocular 6d object pose estimation. In CVPR, pages 16611-16621, 2021. 1, 2, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 535 + ], + "type": "text", + "content": "[42] Tianfu Wang, Guosheng Hu, and Hongguang Wang. Object pose estimation via the aggregation of diffusion features. In CVPR, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 294, + 568 + ], + "type": "text", + "content": "[43] Yangzheng Wu and Michael Greenspan. Pseudo-keypoint rkhs learning for self-supervised 6dof pose estimation. In ECCV, 2024. 2, 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[44] Y. Xiang, T. Schmidt, V. Narayanan, and D. Fox. PoseCNN: A convolutional neural network for 6d object pose estimation in cluttered scenes. arXiv preprint arXiv:1711.00199, 2017. 1, 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 646 + ], + "type": "text", + "content": "[45] Li Xu, Haoxuan Qu, Yujun Cai, and Jun Liu. 6d-diff: A keypoint diffusion framework for 6d object pose estimation. In CVPR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[46] Zongxin Yang, Xin Yu, and Yi Yang. Dsc-posenet: Learning 6dof object pose estimation via dual-scale consistency. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[47] Yunfan Ye, Kai Xu, Yuhang Huang, Renjiao Yi, and Zhiping Cai. Diffusionedge: Diffusion probabilistic model for crisp edge detection. In AAAI, 2024. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 129 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 95 + ], + "type": "text", + "content": "[48] S. Zakharov, I. Shugurov, and S. Ilic. Dpod: 6d pose object detector and refiner. In ICCV, 2019. 2, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 315, + 96, + 553, + 129 + ], + "type": "text", + "content": "[49] Shaobo Zhang, Wanqing Zhao, Ziyu Guan, Xianlin Peng, and Jinye Peng. Keypoint-graph-driven learning framework for object pose estimation. In CVPR, 2021. 2" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8687" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_content_list.json b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..67e7bdbb083465d97a81ba4e2b0ee84df0a70ebf --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_content_list.json @@ -0,0 +1,1232 @@ +[ + { + "type": "text", + "text": "Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation", + "text_level": 1, + "bbox": [ + 145, + 128, + 852, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chen Gao, Shuo Zhang* Youfang Lin \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China.", + "bbox": [ + 148, + 203, + 844, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{gaochen, zhangshuo, yflin}@bjtu.edu.cn", + "bbox": [ + 323, + 258, + 666, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Disparity estimation is an essential step in processing and analyzing Light Field (LF) images. Recent methods construct the cost volume to exploit the correspondence of the LFs over the preset maximum disparity, limiting them to process the large parallax scenes. Different from constructing cost volume, the self-attention mechanism calculates the parallax attention between epipolar lines to find the matching points. However, for LFs that have different views, the related disparity scales are different in parallax attention since the baselines with the central view are different. Moreover, if the matching information is occluded in one view, the disparity information can be explored through other views. Therefore, mapping these attentions to the same scale and selecting effective matching information are key points for disparity estimation from parallax attention. In this paper, we explore parallax attention for LF and design an unsupervised method, named Epipolar Consistent Attention Aggregation Network (ECAAN). We first introduce an epipolar consistent scale unification block by considering the consistency relationships to standardize disparity scales of the parallax attention maps. Based on the intra-properties and inter-reationships of parallax attention, we further propose a consistent occlusion-free aggregation block to integrate the information from the occlusion-free areas. In addition, we design an improved photometric loss to constrain the model. ECAAN achieves state-of-the-art performance in LF depth estimation. Notably, ECAAN attains a mean square error of 0.2 on large-disparity LF datasets, achieving a $68\\%$ error reduction compared to the second-best method.", + "bbox": [ + 86, + 340, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 821, + 220, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light Field (LF) disparity estimation aims to infer the 3D scene geometry from the captured LF images, which is ben-", + "bbox": [ + 89, + 847, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/32f3b947a660a00eb71532d343a7ce24feb3b21b187deeb19a8f48f8c43f7435.jpg", + "image_caption": [ + "Figure 1. Comparative analysis with state-of-the-art methods on the SLFD dataset. Left: Performance metrics of models on Bp across thresholds (0.3, 0.1, 0.05). Right: Quantitative results for MSE ( $\\times$ 100) and Inference Time (seconds). Our method achieves state-of-the-art performance across all metrics while maintaining competitive computational efficiency. Lower values of the Bp and MSE metrics indicate better performance." + ], + "image_footnote": [], + "bbox": [ + 526, + 306, + 895, + 452 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eficial for various advanced computer vision applications ranging from 3D reconstruction to robotics and autonomous driving [4, 5]. The key aspect of LF disparity estimation is to identify corresponding pixels across different views.", + "bbox": [ + 511, + 577, + 906, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "At present, both non-learning methods [15] and learning methods [8, 10, 16, 17, 24] explore geometry information between LF images based on the constructed cost volume. These methods usually construct a cost volume within a predefined disparity range, and compute the weighted sum of those pre-defined disparity candidates to obtain the final disparity prediction. However, such a design suffers from a high computational and memory burden, which limits them in constructing a full-range cost volume for handling large disparities scenes. Moreover, the predefined disparity range inherently constrains the performance of cost-volume-based approaches, making them less effective in handling scenes with extreme parallax variations.", + "bbox": [ + 509, + 640, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently, some methods [13] attempt to use self-attention to explore consistency in stereo matching and show great potential in learning stereo correspondence. PAM [25] explores the pixel movements between views by", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: zhangshuo@bjtu.edu.cn", + "bbox": [ + 107, + 886, + 367, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6488", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "finding the most similar features in parallax attention maps. Unlike stereo images, LFs have different views. The position offset scales in the attention maps are different since the baselines vary between the arranged views. In addition, if the matching information is occluded in one view, the disparity information of that point can be explored through other views. Therefore, how to map these attention maps to same scale and select information from the occlusion-free areas are key challenges that we should focus on.", + "bbox": [ + 89, + 90, + 480, + 224 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we adapt the parallax attention to LFs and propose a novel unsupervised framework to capture the 3d cues, termed the Epipolar Consistent Attention Aggregation Network (ECAAN). Specifically, we first explicitly calculate the correlation between the reference point and all possible matching points along the epipolar lines to construct the parallax attention maps. To standardize disparity scales of the parallax attention maps, we introduce an Epipolar Consistent Scale Unification (ECSU) block by considering the consistency relationships. Furthermore, we explore the intra-properties and inter-relationships of parallax attention maps and design a Consistent Occlusion-Free Aggregation (COFA) block to filter out occlusions. Finally, the fused attention map is used to regress the disparity map of the scene. In addition, we design an improved photometric loss to constrain the model during training. Our method demonstrates outstanding performance in LF disparity estimation, particularly in large-disparity scenes, achieving more accurate predictions with improved efficiency. For example, ECAAN achieves a $68\\%$ error reduction compared to the second-best method (see Fig. 1).", + "bbox": [ + 91, + 227, + 486, + 544 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 558, + 232, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Most existing learning-based LF disparity estimation methods primarily focus on addressing two challenges: occlusions and large disparities. To mitigate matching errors caused by occlusions, various methods integrate view selection modules or specialized loss functions. Tsai et al. [24] explored matching costs and introduced an SAI selection module to generate attention maps, extracting meaningful information from the available views. Chen et al. [1] proposed a four-branch architecture for matching cost learning, incorporating two attention mechanisms to further refine the information by prioritizing branches with minimal occlusion and richer textures. Guo et al. [3] introduced an occlusion detection network, trained using a designed occlusion map, to process occluded and non-occluded regions separately, ensuring accurate disparity estimation with sharp edges. In recent years, Jin et al. [10] partitioned the entire LF into sub-LFs, feeding them into a multi-scale network to predict disparity maps alongside confidence maps based on the constructed cost volume. In addition, they proposed a simple strategy to fuse multiple disparity maps to deal with occlusion. Zhang et al. [31] proposed a two-stage pipeline", + "bbox": [ + 89, + 583, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "to focus on the occlusion problem, in which they generate multiple occlusion priors and act on the loss functions. Yan et al. [28] used the generative adversarial network to estimate the disparity map in an unsupervised manner. Zhou et al. [33] introduced a geometry-based occlusion model along with two loss functions to learn occlusion- and noise-invariant representations of the LF.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Effectively managing large disparities in LFs is another open problem. Shi et al. [19] addressed this issue by decomposing LFs into multiple stereo-image pairs, employing a fine-tuned stereo estimation network to compute the disparity for each pair, which was subsequently fused into refined disparity maps. Some approaches [31] tackle large parallax scenes by reducing the angular resolution of the input. However, existing methods commonly generate cost volumes, leading to significant memory overhead, which constrains their ability to perform fine-grained disparity sampling and ultimately limits their performance. Given the advantages of global interaction, the self-attention mechanism in transformers facilitates pixel-wise similarity modeling, making them highly effective in stereo-matching [25]. However, their application in LF disparity estimation remains limited due to the relatively complex structure of LFs, which poses challenges in effectively integrating self-attention mechanisms within LF representations.", + "bbox": [ + 511, + 198, + 906, + 469 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Epipolar Parallax Attention", + "text_level": 1, + "bbox": [ + 511, + 481, + 772, + 500 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, the LFs are represented as $\\mathcal{L}(u,v,x,y)\\in$ $\\mathbb{R}^{U\\times V\\times H\\times W}$ , where $(u,v)$ represent angular coordinates, and $(x,y)$ represent spatial coordinates [12]. By fixing angular coordinates, the 2D view $I_{(u,v)}\\in \\mathbb{R}^{H\\times W}$ is obtained. The central view are represented as $I_{(0,0)}$", + "bbox": [ + 511, + 507, + 906, + 585 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. EPI Parallax Attention Calculation", + "text_level": 1, + "bbox": [ + 511, + 590, + 820, + 606 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Inspired by the PAM model [25], the self-attention mechanism explicitly computes the correlation between points in paired views. Unlike binocular vision, LFs exhibit correspondences along epipolar lines in multiple directions simultaneously. To better capture these relationships, we compute parallax attention maps on EPIs in various directions, rather than relying on pairwise view similarity as in [25]. Specifically, we design a horizontal EPI parallax attention calculation mechanism (H-EPA) and a vertical EPI parallax attention calculation mechanism (V-EPA) to capture the correlation between views. To introduce the H-EPA, we take a horizontal EPI feature $\\mathcal{F}_h\\in \\mathbb{R}^{V\\times W\\times C}$ that corresponding to the central view as an example. Firstly, 2D EPI convolutional layers are applied to the $\\mathcal{F}_h$ to realise the interaction of redundant information in EPIs. Then, we use the $1\\times 1$ convolutions to get the query features $Q_{h}$ and the key features $K_{h}$ from $\\mathcal{F}_h$ which are further flattened to a sequence of tokens with a length of $VW$ ( $Q_{h},V_{h}\\in \\mathbb{R}^{VW\\times C}$ ). $Q_{h}$ is multiplied with $K_{h}$ to obtain the hori", + "bbox": [ + 511, + 613, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6489", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a47acb8ff30d05ebb88177ece180b5269e22890ae905831d5e62332a133cfa12.jpg", + "image_caption": [ + "Figure 2. One typical example with the related parallax attention maps. The point 1 is visible in all views and the attention value achieves an apparent bright spot in one specific point in each pair view. The point 2, 3 are ambiguous in some views. These features are aggregated combining the occlusion masks to obtain accurate parallax attention." + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 898, + 354 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "zontal epipolar parallax attention map $C_h \\in \\mathbb{R}^{VW \\times VW}$ as $C_h = Q_h \\cdot K_h^T$ . Finally, $C_h$ is reshaped to $V \\times V$ parallax-attention maps $C_{(0,v_1) \\to (0,v_2)} \\in \\mathbb{R}^{W \\times W}$ , which records the similarity between any two positions along the horizontal epipolar lines. The process of V-EPA is the same as the H-EPA while dealing with the vertical EPI features $\\mathcal{F}_v \\in \\mathbb{R}^{U \\times H \\times C}$ .", + "bbox": [ + 88, + 415, + 480, + 521 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Parallax Attention Maps Analyse", + "text_level": 1, + "bbox": [ + 89, + 531, + 382, + 547 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Fig. 2, we visualized the $V$ horizontal parallax attention maps $C_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times W}$ in $C_h$ . These maps are calculated by evaluating the similarity between each point in the central view $I_{(0,0)}$ and every point along the corresponding epipolar line in the view $I_{(0,v)}$ , where brighter color means it is similar to the corresponding point in the central view. The disparity information is encoded through positional offsets of the most similar points relative to the diagonal line, which vary across different views. To effectively aggregate these parallax-attention maps, scale variations and occlusion remain key challenges.", + "bbox": [ + 88, + 551, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scale Issue: Since the disparity between views varies significantly for different baselines, the position offset scales of these parallax attention maps are different. In Fig. 2, for views $I_{(0, -3)}$ and $I_{(0,3)}$ that are far from the central view, the position offset is $3 \\times$ larger than the adjacent view $I_{(0, -1)}$ and $I_{(0,1)}$ due to the larger baseline. In the parallax attention maps, it is clear that the offsets relative to the diagonal line are larger in $M_{(0,0) \\to (0, -3)}$ and $M_{(0,0) \\to (0,3)}$ . The offsets of the left view and right view are also symmetric around the diagonal line since the view movement direction is opposite. Coarsely resizing the scale of these parallax attention maps may break the encoded consistency", + "bbox": [ + 88, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "information and introduce significant errors. To aggregate these attention maps, we propose an epipolar consistent re-sample and up-sample strategy (see Sec.4.2). We first crop and re-sample the attention map $M_{(0,0) \\to (0,v)}$ according to the disparity range and form a new attention map. Then, the re-sampled attention maps are up-sampled heuristically to the same scale by placing their original attention values in the corresponding positions. As in Fig. 2, all up-sampled attention maps are mapped to the same scale and are consistent with the initial parallax information.", + "bbox": [ + 511, + 416, + 906, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Occlusion Issue: Another fundamental challenge is occlusion, where view consistency does not hold. In LFs, the occlusions may arise from various directions, resulting in different occlusion relationships for each map. Here we take $M_{(0,0)\\to (0,v)}$ as a typical example in Fig. 2. If one point is visible in all views (as point 1), the attention value achieves an apparent bright spot at one specific point in each pair view, which is just its matching point. However, it is difficult to find the matching point for points that are occluded in part of views. For example, as the view moves from left to right, the point 2 is visible in the beginning but is then occluded by the foreground pencil. Therefore, the matching point with a bright spot is easy to find in view $I_{(0,-3/-2/-1/1)}$ but is ambiguous in view $I_{(0,2/3)}$ . Similarly, the background point 3 is occluded by the foreground pencil in the left views but is visible in the right views. The corresponding matching point is concentrated in view $I_{(0,-1/1/2/3)}$ but is dispersed in view $I_{(0,-3/-2)}$ . For these kinds of occluded points, labeled with blue and yellow boxes, directly fusing all the matching values or alignment information in all views together will bring errors. To alleviate the influence of the occluded views in disparity", + "bbox": [ + 511, + 568, + 908, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6490", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/c1aef4ad36e2fa5160877a654ea38e7d2b240e1bf5db45fe9b54156ef93a24a5.jpg", + "image_caption": [ + "Figure 3. An overview of our disparity estimation network. After feature extraction, the epipolar parallax attention maps are constructed by the H-EPA and V-EPA. The epipolar parallax attention maps are aggregated based on the epipolar consistent attention aggregation. Finally, the disparity map is generated through the disparity regression." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 859, + 435 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "estimation, we detect the occluded regions and aggregate the disparity information encoded in attention maps from non-occluded regions (see Sec. 4.3). Specifically, we calculate the occlusion masks for each view to label whether the point is occluded in the paired view. As shown, the two regions with blue boxes that are visible in the central view but occluded by the pencil in view $I_{(0, -3)}$ are accurately masked in the occlusion masks. Combining with the occlusion mask, we calculate the occlusion-aware aggregation weights for different views as the last row in Fig. 2. The fused attention map is also visualized, in which we can find a matching point with high similarity for each pixel by using the information of non-occluded viewpoints.", + "bbox": [ + 88, + 489, + 485, + 686 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Epipolar Consistent Attention Aggregation", + "text_level": 1, + "bbox": [ + 89, + 702, + 475, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The proposed Epipolar Consistent Attention Aggregation Network aims to estimate the disparity map of the central view $D_{c}\\in \\mathbb{R}^{H\\times W}$ from the LFs $\\mathcal{L}$ .", + "bbox": [ + 89, + 729, + 483, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 787, + 200, + 801 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall network structure is shown in Fig. 3. Given the input image $\\mathcal{L}$ , we first implement spatial convolution layers to encode the multi-scale features $\\mathcal{F}_e^i\\in \\mathbb{R}^{U\\times V\\times H_i\\times W_i\\times C}$ in which $H_{i} = \\frac{H}{2^{i}}$ , $W_{i} = \\frac{W}{2^{i}}$ . Then, these encoded features are processed through a hierarchical decoder structure that deploys with the H-EPA and V-EPA(Sec. 3.1) to ex", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "plore the epipolar consistency. At each decoding stage $i$ , the features produced in the previous stage $\\mathcal{F}_d^{i - 1}$ are up-sampled and fused with the $\\mathcal{F}_e^i$ through spatial convolutions, producing $\\mathcal{F}_{fuse}^{i}$ . The $\\mathcal{F}_{fuse}^{i}$ are then processed in parallel by H-EPA and V-EPA to obtain epipolar parallax attention maps $C_h^i /C_v^i$ along with refined features $\\mathcal{F}_h^i /\\mathcal{F}_v^i$ . These refined features $\\mathcal{F}_h^i /\\mathcal{F}_v^i$ are subsequently combined and fed into the next decoder stage. Once all hierarchical epipolar parallax attention maps $\\{C_h^i,C_c^i\\}_{i = 1}^N$ are obtained, they are up-sampled and summed to form the final attention maps $C_h$ and $C_v$ . To estimate depth maps of central views, the attention maps calculated from central views with other views are aggregated by the epipolar consistent attention aggregation (ECAA) mechanism, which consists of ECSU block (Sec. 4.2) and COFA block (Sec. 4.3). Finally, the disparity is estimated as the weighted sum of disparity candidates based on the aggregated epipolar consistent attention $P_{att}$ :", + "bbox": [ + 511, + 489, + 908, + 750 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {c} = \\sum_ {k = 0} ^ {D _ {u p}} k \\times P _ {a t t} (:,:, k) - \\frac {D _ {u p}}{2}. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 758, + 906, + 799 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{up}$ means the size of disparity candidates in $P_{att}$ .", + "bbox": [ + 511, + 806, + 893, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Epipolar Consistent Scale Unification", + "text_level": 1, + "bbox": [ + 511, + 832, + 836, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better aggregate the calculated attention maps from different views, we re-sample and up-sample these maps to a uniform scale under the constraint of epipolar consistency.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6491", + "bbox": [ + 482, + 944, + 513, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Epipolar Consistent Attention Re-Sample. For each $V - 1$ view that moves horizontally from the central view, there are $H$ attention maps $M_h$ available that calculate the pairwise similarity of these views. We add the disparity constraint to $C_h$ and get the constrained epipolar parallax attention maps $M_h$ . We re-sample each paired attention map $M_{(0,0)\\to (0,v)}$ according to the disparity range $[d_{min},d_{max}]$ and form a new map $P_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times (|\\mathrm{v}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}$ . Similarly, the attention maps related to the vertical moved views are also re-sampled as $P_{(0,0)\\to (u,0)}\\in \\mathbb{R}^{H\\times (|\\mathrm{u}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}$ .", + "bbox": [ + 89, + 90, + 480, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Epipolar Consistent Attention Up-Sample. For attention maps $P_{(0,0) \\to (0,v) / (u,0)}$ that have different kinds of scale, we propose to use deconvolution layers with different parameters to provide specific location information for similarity score interpolation. Taking $P_{(0,0) \\to (0,v)}$ as an example, the magnification factor $f$ in the deconvolution layer is:", + "bbox": [ + 89, + 258, + 480, + 349 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {(0, 0) \\rightarrow (0, v)} = N / | v |, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 364, + 480, + 378 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ indicates the least common multiple of $\\{1,2,\\dots ,\\frac{V}{2}\\}$ . In such a way, the original similarity scores in the small disparity scale space are put into the corresponding positions in the large disparity scale space as in Fig. 3, resulting in up-sampled $P_{(0,0)\\to (0,v)}^{up}\\in \\mathbb{R}^{W\\times D_{up}}$ $(D_{up} = N\\times (d_{max} - d_{min}) + 1)$ . Therefore, compared with other up-sampled methods such as pixel shuffle [20], the deconvolution layer with $f$ not only maps all attention to the same scale but also keeps the original matching information. Note that we manually flip half of the up-sampled parallax attentions since the pixel offsets are opposite for views that in different sides of the central view. We then combine all the processed consistent parallax attentions from the horizontal and vertical views $P_{(0,0)}^{up}\\in \\mathbb{R}^{(U + V - 2)\\times H\\times W\\times D_{up}}$ .", + "bbox": [ + 89, + 382, + 482, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Consistent Occlusion-Free Aggregation.", + "text_level": 1, + "bbox": [ + 89, + 606, + 431, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To effectively filter out occlusion-area information, we analyze both the intra-properties of individual parallax attention maps and the inter-relationships between them, leveraging this analysis to generate occlusion-aware fusion weights for consistent information aggregation.", + "bbox": [ + 89, + 628, + 480, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occlusion Masks. The intra-properties of individual parallax attention maps reveal that when a point is occluded in one view, its similarity remains relatively low as analyzed in Sec. 3.2. Based on this, we perform occlusion detection on parallax-attention map $M_{h}$ and $M_{v}$ and generate occlusion prior knowledge $O_{(0,v)}$ , $O_{(u,0)} \\in \\mathbb{R}^{H \\times W}$ of view $I_{(0,v)}$ , $I_{(u,0)}$ with respect to $I_{(0,0)}$ . Specifically, we add parallax attention maps $M_{(0,v) \\to (0,0)}$ and $M_{(u,0) \\to (0,0)}$ along the non-normalized dimension and set an appropriate threshold $\\tau$ (set to 0.9 in this paper) for binarization. Since occlusions usually appear at object edges, we utilize guided filter [7] to sharpen the edges before binarization. Finally, we obtain $(V - 1)$ binary mask $O_{(0,v)}$ and $(U - 1)$ binary mask $O_{(u,0)}$ .", + "bbox": [ + 89, + 704, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistency Weight. Given that disparity constitutes an intrinsic property of visual scenes, the inter-relationships between the up-sampled attention maps should inherently maintain consistency. Therefore, we focus the model on aggregating the attention maps that exhibit consistent disparity information. To implement this, we feed $P_{(0,0)}^{up}$ into the transformer model along with a linear layer to calculate the consistency weights of different attention maps. Specifically, the $P_{(0,0)}^{up}$ is reshaped to $H \\times W$ sequence of tokens $T \\in \\mathbb{R}^{(U + V - 2) \\times D_{up}}$ . Afterward, $T$ are multiplied separately by $W_{Q}, W_{K}, W_{V} \\in \\mathbb{R}^{D_{up} \\times c}$ to generate the corresponding query, key, and value. Then, the dot-product operation is used on the query and key, followed by a softmax function to obtain the attention scores on the tokens. The attention scores do matrix multiplication with value to get the fusion attention feature and the fusion attention feature is fed into a linear layer to get the initial fusion weight $W_{init}$ .", + "bbox": [ + 511, + 90, + 903, + 352 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Occlusion-Free Aggregation. Since occlusions often cause errors in the final regression results, we consider assigning small weights to the occluded parts when aggregating the attention maps. Specifically, we introduce the occlusion prior information $O_{(u,0)}$ and $O_{(0,v)}$ to suppress the information provided by the occluded views. First, we map occluded points in $O_{(u,0)}$ and $O_{(0,v)}$ to $-\\infty$ and non-occluded point to 0. The mapping results from $O_{(u,0)}$ and $O_{(0,v)}$ are concatenated and added to $W_{init}$ . We then fed the added weight into a softmax layer to produce the occlusion-aware fusion weight $W_{fuse} \\in \\mathbb{R}^{(U + V - 2)\\times H\\times W}$ .", + "bbox": [ + 511, + 352, + 903, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The up-sampled $P_{(0,0)}^{up}$ are then multiplied by the broadcasted $W_{fuse}$ and sum over the perspective dimension to obtain the fused epipolar consistent attention map $P_{att}$ :", + "bbox": [ + 511, + 517, + 903, + 563 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nP _ {a t t} = \\sum_ {k = 1} ^ {U + V - 2} W _ {f u s e} (k,:,:) \\times P _ {(0, 0)} ^ {u p} (k,:,:) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 569, + 903, + 606 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, we integrate attention maps from different views to increase the robustness of our results.", + "bbox": [ + 511, + 613, + 903, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.4. Loss Function", + "text_level": 1, + "bbox": [ + 511, + 651, + 660, + 666 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Based on the occlusion mask $O_{(u,0)}$ and $O_{(0,v)}$ , we improve the photometric loss [29] to make the model pay attention to non-occluded regions during training. Specifically, we use the final disparity map $D_{c}$ to warp all views toward the central view and calculate the L1 loss between them:", + "bbox": [ + 511, + 672, + 903, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {p} = \\sum_ {u} ^ {U} \\sum_ {v} ^ {V} L _ {1} \\left(w \\left(I _ {(u, v)}, D _ {c}\\right), I _ {(0, 0)}\\right) \\times O _ {(u, v)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 753, + 903, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $w(\\cdot)$ denotes a warping operator. Since we only have $O_{(u,0)}$ and $O_{(0,v)}$ that label the occluded regions in the cross-shape views in LFs, we propose a simple way to extend them to all views in LFs:", + "bbox": [ + 511, + 797, + 903, + 857 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {(u, v)} = \\left\\{ \\begin{array}{l l} 1, & O _ {(u, 0)} \\wedge O _ {(0, v)} = 1 \\\\ 0, & O _ {(u, 0)} \\vee O _ {(0, v)} = 0 \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 863, + 903, + 904 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6492", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/3485465f23cea27db9755e1c19ddad52cb1f82e4eab02ef370e95a855e8301d3.jpg", + "table_caption": [ + "Table 1. Quantitative comparison of MSE*100/Bp(0.3,0,1,0.05) on SLFD dataset. The best results are in bold." + ], + "table_footnote": [], + "table_body": "
MethodsLionRooster clockToy bricksElectro devices
MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)
SupEPINet [21]0.47631.45768.13882.9540.74036.86871.56285.1471.05734.65861.95777.4911.11239.80964.45581.793
LFAntNet [24]0.37212.99429.59449.3270.2786.83125.81350.9200.73811.60729.01852.5760.70313.81938.73160.993
UnSupUnCNN [16]1.44258.22184.86792.3739.21334.58675.02187.3832.25453.92282.17290.9502.18141.16372.97085.866
UnMonocular [32]0.76048.30587.37194.6730.66439.93173.60186.3701.22647.59471.09884.9822.73645.61275.23887.332
UnCon [14]1.19235.49375.39287.8910.49116.92853.88975.5971.74150.47179.47089.4743.60442.44375.65487.271
UnOcc [10]1.45453.46080.65090.2120.42111.14842.62768.1511.90661.45784.58791.5702.72040.83471.62284.935
UnMVFM [31]0.3608.76624.42047.9110.2615.79625.30348.6680.77210.04830.42755.7320.84213.63136.03857.880
ECAAN30.1731.5596.86416.4800.1772.58810.89631.1060.3323.74129.62454.4710.2909.09531.76755.710
ECAAN50.0871.1389.470332.9980.1010.8015.57814.3920.4083.20727.06847.2510.2115.63022.05744.819
", + "bbox": [ + 99, + 114, + 897, + 252 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0981def47ae9d544ddd90c8dcb29d58eaf26cab1d20dc5186b63f7800e9aedb4.jpg", + "image_caption": [ + "Figure 4. Visual comparison of the disparity maps estimated by different methods on SLFD." + ], + "image_footnote": [], + "bbox": [ + 112, + 265, + 880, + 573 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "If an occluder is labeled in $O_{(0,v)}$ , we consider that this occluder will also appear in the moving view along the vertical direction of $O_{(0,v)}$ . In addition, we also implement the photometric loss on disparity maps $D_{(0,v)}$ and $D_{(u,0)}$ using $L_p^h$ and $L_p^v$ to ensure the accuracy of upsampling, which is regressed using the up-sampled attention map $P_{(0,0)\\to (0,v)}^{up}$ and $P_{(0,0)\\to (u,0)}^{up}$ . Finally, we train our model in an unsupervised manner, the overall loss includes photometric loss $L_p$ , $L_p^h$ and $L_p^v$ , PAM loss $L_s$ and smoothness loss $L_{sm}$ :", + "bbox": [ + 89, + 611, + 483, + 753 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} L _ {u n s u p} = L _ {p} + \\lambda_ {h v} \\left(L _ {p} ^ {h} + L _ {p} ^ {v}\\right) + \\lambda_ {s} L _ {s m} \\tag {6} \\\\ + \\lambda_ {P} \\left(0. 2 L _ {s} ^ {1} + 0. 3 L _ {s} ^ {2} + 0. 5 L _ {s} ^ {3}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 776, + 482, + 815 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $L_{s}^{i}$ extended by the PAM loss [25] and is calculated on the all parallax attention maps obtained in the $i$ -th decoding stage. The $L_{sm}$ [23] is calculated on $D_{(0,v)}$ , $D_{(u,0)}$ and $D_{c}$ to encourage local smoothness of the disparity.", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiment", + "text_level": 1, + "bbox": [ + 513, + 609, + 638, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 633, + 732, + 650 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our model is implemented using the PyTorch framework and trained on an NVidia RTX A5000 GPU. During training, we crop views into $100 \\times 100$ patches as the model input. We train our model by minimizing the $L_{unsup}$ and $\\lambda_{hv}, \\lambda_s, \\lambda_P$ in Eq. 6 are set to 0.5, 0.01, 1. The models are optimized using the Adam optimizer [2] with $\\beta_1 = 0.9$ and $\\beta_2 = 0.999$ . The initial learning rate is set to $1e^{-3}$ and the batch size is set to 2. Besides, we randomly augment the datasets by flipping the images horizontally, vertically, or rotating 90 degrees. Typically, it roughly takes 2 days of training for our model to converge.", + "bbox": [ + 511, + 656, + 906, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Comparisons with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 832, + 888, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct a comparative analysis with SOTA methods, including two non-learning-based methods (i.e., SPO [30], OAVC [6]), four supervised learning-based meth", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6493", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/024eeed3f039cc25e662e8ea2d9c03a72b67136f709466f258c1724c91d3679f.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of the MSE*100/Bp(0.07) metrics on HCI dataset. The best results among all methods are in red. The best and second best results among unsupervised methods are in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodBackgammonDotsPyramidsStripsBoxesCottonDinoSideboardBedroomBicycleHerbsOrigamiAvg
MSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/Bp
NonSPO [30]4.59/3.785.24/16.270.04/0.866.96/14.999.11/15.891.31/2.590.31/2.181.02/9.300.21/4.865.57/10.9111.23/8.262.03/11.693.97/8.47
OAVC [6]3.84/3.1216.58/69.110.04/0.831.32/2.906.99/16.140.60/2.550.27/3.941.05/12.420.21/4.924.89/12.2210.36/8.731.48/12.563.97/12.45
SupEPINet [21]3.71/3.501.48/2.490.007/0.160.93/2.465.97/12.340.20/0.450.16/1.210.80/4.460.20/2.304.60/9.619.29/10.961.83/5.812.43/4.65
LFattNet [24]3.65/3.131.43/1.430.004/0.200.89/2.934.00/2.790.21/9.510.09/5.220.53/4.820.37/2.793.35/2.796.61/5.221.73/4.821.91/3.80
AttMLFNet [1]3.86/3.231.04/1.610.003/0.170.81/2.933.84/11.140.06/0.200.04/0.440.40/2.690.13/2.073.80/8.846.37/5.430.99/4.411.78/3.60
OACCNet [26]3.94/3.931.42/1.510.004/0.160.85/2.922.89/10.700.16/0.310.08/0.970.54/3.350.15/2.302.91/8.086.56/6.520.88/4.071.69/3.74
UnsupUnMonocular [32]11.83/12.312.54/3.650.027/0.262.68/11.149.27/22.882.91/3.811.03/5.402.77/10.950.42/7.419.23/20.1026.82/13.443.68/10.956.10/10.19
UnOcc [10]6.68/14.376.57/45.340.213/7.355.20/41.367.45/26.240.80/8.460.63/8.251.79/14.200.39/12.696.23/21.6513.94/16.961.92/19.824.32/19.72
UnLFDisp [33]6.86/5.201.44/1.530.016/0.271.80/5.128.32/20.182.04/3.040.78/4.171.85/7.680.35/7.367.49/16.5519.43/11.802.94/9.674.44/7.71
UnGAN [28]5.38/10.481.74/2.040.013/0.302.39/9.048.01/19.870.76/1.940.74/4.713.39/7.080.55/6.917.96/18.9215.58/12.763.61/10.334.18/8.70
ECAAN10.18/8.063.03/4.770.021/0.351.86/29.847.87/18.540.67/1.920.63/4.921.35/9.050.27/6.407.21/17.1214.81/10.811.99/11.044.16/10.24
", + "bbox": [ + 94, + 128, + 911, + 280 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3085ee2460d96a0f579e6afe2a712642525d767f5f884e2b6ec26564e9914e15.jpg", + "image_caption": [ + "Figure 5. Visual comparison of different methods on real-world scenes 'Lego Bulldozer' and 'Lego Knights'. The input $3 \\times 3$ views are selected (label in the above line) with different intervals to make the disparity ranges increase gradually." + ], + "image_footnote": [], + "bbox": [ + 112, + 292, + 888, + 481 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods (i.e., EPINet [21], LFAntNet [24], AttMLFNet [1], OACCNet [26]), and seven unsupervised learning-based methods (i.e., UnCNN [16], UnMonocular [32], UnCon [14], UnLFDisp [33], UnGAN [28], UnOcc [10], UnMVFM [31]). The Bad pixel (Bp) ratio [9], which measures the percentage of incorrectly estimated pixels $>t$ , and the Mean Squared Error (MSE) are used as metrics. To comprehensively evaluate model performance, we conduct experiments on both large and small disparity datasets.", + "bbox": [ + 88, + 537, + 485, + 674 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2.1. Large Disparity Scenes", + "text_level": 1, + "bbox": [ + 89, + 684, + 299, + 699 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Synthetic Datasets. Following most approaches [31], synthetic LFs in SLFD are selected for training and testing, with interview disparities in [-20, 20]. Table 1 illustrates the quantitative results of different methods on several scenes from SLFD. Due to the memory burden of constructing the cost-volume, these cost-volume based methods use the central $3 \\times 3$ views as input. To ensure fairness of the experiments, we respectively trained our models for $3 \\times 3$ and $5 \\times 5$ angular resolutions, labeled as $ECAAN_{3}$ and $ECAAN_{5}$ . Among the LF image methods, our method achieves the lowest MSE and Bp values across all scenes, indicating its ability to better capture the consistency relationship of large disparity scenes and produce superior disparity maps.", + "bbox": [ + 88, + 704, + 485, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Specifically, $ECAAN_{3}$ achieves a significant MSE reduction of up to 5.427 (Rooster clock) and 0.273 (Lion), while also lowering $\\mathrm{Bp}(0.3)$ by up to 7.628 (Lion) and 4.025 (Toy bricks) over the UnMVFM. The visual results are shown in Fig. 4, with a different color map from the LFs for distinction. Other approaches tend to generate noisy predictions and exhibit significant accuracy degradation. In contrast, our results have better visual qualities with smaller errors compared to others, especially in the edge regions.", + "bbox": [ + 511, + 537, + 906, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real Datasets. To assess the model's capability to process real-world scenes, we test the models on the LFs captured by the moving camera [22]. Since GT is not available, all models are trained in SLFD. For each scene, we select different $3 \\times 3$ views from the $17 \\times 17$ LFs with various disparity ranges to evaluate the robustness of the methods. As shown in Fig. 5, the disparity of LF is gradually larger when choosing views at different angular positions. Our method consistently achieves superior accuracy across inputs with varying disparities, preserving sharp boundaries and intricate structural details. Due to the predetermined fixed sampling intervals in cost volume construction, these models are prone to pronounced stratification artifacts. This issue is particularly evident in regions with smoothly varying disparities or occlusions, resulting in abrupt depth discontinu", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6494", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/52eb1ff0a0dc12fad056bdfcc0c68ca44dc6361e193e319f2b411faacf6e4ef0.jpg", + "table_caption": [ + "Table 3. Result of variant models." + ], + "table_footnote": [], + "table_body": "
ModuleLossMSE / Bp (0.3)
ECSUCOFALpLph/vLionRoosterToyElectro
(1)××all✓4.96/21.260.37/2.933.81/33.052.34/20.62
(2)Linear×all✓2.48/14.690.30/2.741.84/24.021.88/19.27
(3)Dconv×all✓2.06/6.790.25/2.461.14/13.421.51/18.66
(4)Dconv✓××0.26/2.170.17/1.020.83/3.990.52/9.91
(5)Dconv✓sub✓0.16/1.460.14/0.970.59/3.420.42/8.41
(6)Dconv✓all✓0.08/1.130.10/0.810.41/3.200.21/5.63
", + "bbox": [ + 93, + 107, + 478, + 229 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/94b469f546260e19b3c9dbe09474bb97284c2c3e9f72ae8209b3f65ba1988ba5.jpg", + "image_caption": [ + "Figure 6. Visual comparison on real-world small disparity LFs." + ], + "image_footnote": [], + "bbox": [ + 99, + 241, + 472, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ities and unnatural \"layering\" effects.", + "bbox": [ + 89, + 498, + 339, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2.2. Small Disparity Scenes", + "text_level": 1, + "bbox": [ + 89, + 537, + 295, + 553 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Synthetic Datasets. Regarding the synthetic LFs with small disparity, we use LFs from the HCI benchmark [9, 27] to verify the performance of the models on the synthetic datasets, in which the LFs have a disparity range of $[-4, 4]$ . The results obtained using LFs with an angular resolution of $7 \\times 7$ are illustrated in Table 2. Due to the lack of ground-truth supervision, the performance of our model and nonlearning models is worse than the supervised learning methods [1, 21, 24, 26] in some scenes. Compared with other unsupervised methods, our approach achieves the lowest MSE metric and outperforms others in the Bp metric across most scenes. More comparisons on HCI benchmark are provided in the supplementary material.", + "bbox": [ + 89, + 561, + 482, + 758 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Real Datasets. Following [10], the images from [11, 18, 22] are used for real-world dense LF comparison. All unsupervised methods are retrained on the 100 images from [11]. We show the results in real-world LFs [11, 18, 22] in Fig. 6. Due to the difference between the synthetic dataset and the real-world scene, the performance of the supervised models degrades in the real-world scene. Compared with other models, our model is able to better preserve fine structure in real scenes with complex occlusions.", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6ffd11fec07cd8a4fcac59f47c80bbf76953d1b3edfa3d7415473f0b3c045883.jpg", + "image_caption": [ + "Figure 7. Visual comparison of the variant models." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 903, + 198 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Investigation", + "text_level": 1, + "bbox": [ + 511, + 234, + 720, + 251 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this subsection, we conduct ablation experiments on SLFD using the $ECAAN_{5}$ . The results are shown in Table 3 and Fig. 7. We first design the following models to verify the effectiveness of our proposed epipolar consistent up-sample strategy: Model-1 directly regress all the parallax attention to disparity and average them as the final disparity map. Model-2 replaces the deconvolution layer using a simple linear layer to upsample the attention maps to the same scale. Model-2 performs better than Model-1, suggesting that integrating information in the attention level keeps more information than in the disparity level. Compared Model-2 with Model-3, our designed upsample strategy makes it easier to consistently map the attention to the same scale. We further verify the effectiveness of the proposed occlusion-aware aggregation. Model-3 removes the COFA while generating the mask for the loss. Compared Model-3 with the final model Model-6, the occlusion prior information reduces the weights of occluded regions and better uses the information extracted from the occlusion-free view. We further remove the occlusion mask on the loss, respectively. In Model-4 we remove all the occlusion masks in loss $L_{unsup}$ and in Model-5 we only apply subocclusion masks $O_{(u,0)}$ and $O_{(0,v)}$ . The results indicate that using the designed occlusion masks, the photometric supervision is more accurate and the performances are improved. More results are presented in the supplementary material.", + "bbox": [ + 511, + 257, + 906, + 651 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a LF disparity estimation network based on the proposed epipolar consistent parallax attention. First, we formulate the horizontal/vertical EPI parallax attention calculation mechanism and analyze the occlusion and scale issues in the LF process. Then, we apply the EPI parallax attention to the disparity estimation task. Specifically, we design an epipolar consistent attention aggregation mechanism to unify the offset scale and aggregate the consistency parallax information of the non-occluded areas. Experimental results show that our method performs favorably against state-of-the-art methods.", + "bbox": [ + 511, + 689, + 905, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032).", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6495", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jiaxin Chen, Shuo Zhang, and Youfang Lin. Attention-based multi-level fusion network for light field depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1009-1017, 2021. 2, 7, 8", + "[2] Yangling Chen, Shuo Zhang, Song Chang, and Youfang Lin. Light field reconstruction using efficient pseudo 4d epipolar-aware structure. IEEE Transactions on Computational Imaging, 8:397-410, 2022. 6", + "[3] Mantang Guo, Jing Jin, Hui Liu, and Junhui Hou. Learning dynamic interpolation for extremely sparse light fields with wide baselines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2450-2459, 2021. 2", + "[4] Yulan Guo, Mohammed Bennamoun, Ferdous Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11): 2270-2287, 2014. 1", + "[5] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1", + "[6] Kang Han, Wei Xiang, Eric Wang, and Tao Huang. A novel occlusion-aware vote cost for light field depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8022-8035, 2021. 6, 7", + "[7] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6):1397-1409, 2013. 5", + "[8] Stefan Heber, Wei Yu, and Thomas Pock. Neural epi-volume networks for shape from light field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2260, 2017. 1", + "[9] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision, pages 19-34. Springer, 2016. 7, 8", + "[10] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 1, 2, 6, 7, 8", + "[11] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM Transactions on Graphics, 35(6):1-10, 2016. 8", + "[12] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the annual conference on Computer Graphics and Interactive Techniques. ACM, 1996. 2", + "[13] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2023. 1", + "[14] Lili Lin, Qiujiang Li, Bin Gao, Yuxiang Yan, Wenhui Zhou, and Ercan Engin Kuruoglu. Unsupervised learning of light field depth estimation with spatial and angular consistencies. Neurocomput., 501(C):113-122, 2022. 6, 7" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] In Kyu Park, Kyoung Mu Lee, et al. Robust light field depth estimation using occlusion-noise aware data costs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2484-2497, 2017. 1", + "[16] Jiayong Peng, Zhiwei Xiong, Dong Liu, and Xuejin Chen. Unsupervised depth estimation from light field using a convolutional neural network. In Proceedings of the IEEE International Conference on 3D Vision, pages 295-303. IEEE, 2018. 1, 6, 7", + "[17] Jiayong Peng, Zhiwei Xiong, Yicheng Wang, Yueyi Zhang, and Dong Liu. Zero-shot depth estimation from light field using a convolutional neural network. IEEE Transactions on Computational Imaging, 6:682-696, 2020. 1", + "[18] Martin Rerabek and Touradj Ebrahimi. New light field image dataset. In Proceedings of the International Conference on Quality of Multimedia Experience, 2016. 8", + "[19] Jinglei Shi, Xiaoran Jiang, and Christine Guillemot. A framework for learning depth from a flexible subset of dense and sparse light field views. IEEE Transactions on Image Processing, 28(12):5867-5880, 2019. 2", + "[20] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 5", + "[21] Changha Shin, Hae-Gon Jeon, Youngjin Yoon, In So Kweon, and Seon Joo Kim. Epinet: A fully-convolutional neural network using epipolar geometry for depth from light field images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4748–4757, 2018. 6, 7, 8", + "[22] Abhilash Sunder Raj, Michael Lowney, Raj Shah, and Gordon Wetzstein. The stanford litro light field archive. http://lightfields.stanford.edu/LF2016.html, 2016. Accessed: Aug. 22, 2020. 7, 8", + "[23] C. Tomasi and R. Manduchi. Bilateral filtering for gray and color images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 839-846, 1998. 6", + "[24] Yu-Ju Tsai, Yu-Lun Liu, Ming Ouhyoung, and Yung-Yu Chuang. Attention-based view selection networks for lightfield disparity estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12095-12103, 2020. 1, 2, 6, 7, 8", + "[25] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2108-2125, 2022. 1, 2, 6", + "[26] Yingqian Wang, Longguang Wang, Zhengyu Liang, Jungang Yang, Wei An, and Yulan Guo. Occlusion-aware cost constructor for light field depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19809-19818, 2022. 7, 8", + "[27] Sven Wanner, Stephan Meister, and Bastian Goldluecke. Datasets and benchmarks for densely sampled 4d light fields. In Proceedings of the Conference on Vision, Modeling, and Visualization, pages 225-226. ACM, 2013. 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6496", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Wenbin Yan, Xiaogang Zhang, and Hua Chen. Occlusion-aware unsupervised light field depth estimation based on multi-scale gans. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6318-6333, 2024. 2, 7", + "[29] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1983-1992, 2018. 5", + "[30] Shuo Zhang, Hao Sheng, Chao Li, Jun Zhang, and Zhang Xiong. Robust depth estimation for light field via spinning parallelogram operator. Computer Vision and Image Understanding, 145:148-159, 2016. 6, 7", + "[31] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Unsupervised light field depth estimation via multi-view feature matching with occlusion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2261-2273, 2024. 2, 6, 7", + "[32] Wenhui Zhou, Enci Zhou, Gaomin Liu, Lili Lin, and Andrew Lumsdaine. Unsupervised monocular depth estimation from light field image. IEEE Transactions on Image Processing, 29:1606-1617, 2020. 6, 7", + "[33] Wenhui Zhou, Lili Lin, Yongjie Hong, Qiujuan Li, Xingfa Shen, and Ercan Engin Kuruoglu. Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2023. 2, 7" + ], + "bbox": [ + 89, + 92, + 482, + 458 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6497", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_model.json b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..51055606cb0d7c698bef5ad721b6778ad8752465 --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_model.json @@ -0,0 +1,1617 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.13, + 0.853, + 0.177 + ], + "angle": 0, + "content": "Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation" + }, + { + "type": "text", + "bbox": [ + 0.149, + 0.204, + 0.846, + 0.257 + ], + "angle": 0, + "content": "Chen Gao, Shuo Zhang* Youfang Lin \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China." + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.26, + 0.668, + 0.275 + ], + "angle": 0, + "content": "{gaochen, zhangshuo, yflin}@bjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.341, + 0.486, + 0.794 + ], + "angle": 0, + "content": "Disparity estimation is an essential step in processing and analyzing Light Field (LF) images. Recent methods construct the cost volume to exploit the correspondence of the LFs over the preset maximum disparity, limiting them to process the large parallax scenes. Different from constructing cost volume, the self-attention mechanism calculates the parallax attention between epipolar lines to find the matching points. However, for LFs that have different views, the related disparity scales are different in parallax attention since the baselines with the central view are different. Moreover, if the matching information is occluded in one view, the disparity information can be explored through other views. Therefore, mapping these attentions to the same scale and selecting effective matching information are key points for disparity estimation from parallax attention. In this paper, we explore parallax attention for LF and design an unsupervised method, named Epipolar Consistent Attention Aggregation Network (ECAAN). We first introduce an epipolar consistent scale unification block by considering the consistency relationships to standardize disparity scales of the parallax attention maps. Based on the intra-properties and inter-reationships of parallax attention, we further propose a consistent occlusion-free aggregation block to integrate the information from the occlusion-free areas. In addition, we design an improved photometric loss to constrain the model. ECAAN achieves state-of-the-art performance in LF depth estimation. Notably, ECAAN attains a mean square error of 0.2 on large-disparity LF datasets, achieving a \\(68\\%\\) error reduction compared to the second-best method." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.222, + 0.837 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.848, + 0.483, + 0.879 + ], + "angle": 0, + "content": "Light Field (LF) disparity estimation aims to infer the 3D scene geometry from the captured LF images, which is ben-" + }, + { + "type": "image", + "bbox": [ + 0.527, + 0.308, + 0.896, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.462, + 0.908, + 0.56 + ], + "angle": 0, + "content": "Figure 1. Comparative analysis with state-of-the-art methods on the SLFD dataset. Left: Performance metrics of models on Bp across thresholds (0.3, 0.1, 0.05). Right: Quantitative results for MSE (\\(\\times\\) 100) and Inference Time (seconds). Our method achieves state-of-the-art performance across all metrics while maintaining competitive computational efficiency. Lower values of the Bp and MSE metrics indicate better performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.639 + ], + "angle": 0, + "content": "eficial for various advanced computer vision applications ranging from 3D reconstruction to robotics and autonomous driving [4, 5]. The key aspect of LF disparity estimation is to identify corresponding pixels across different views." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.641, + 0.909, + 0.838 + ], + "angle": 0, + "content": "At present, both non-learning methods [15] and learning methods [8, 10, 16, 17, 24] explore geometry information between LF images based on the constructed cost volume. These methods usually construct a cost volume within a predefined disparity range, and compute the weighted sum of those pre-defined disparity candidates to obtain the final disparity prediction. However, such a design suffers from a high computational and memory burden, which limits them in constructing a full-range cost volume for handling large disparities scenes. Moreover, the predefined disparity range inherently constrains the performance of cost-volume-based approaches, making them less effective in handling scenes with extreme parallax variations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Recently, some methods [13] attempt to use self-attention to explore consistency in stereo matching and show great potential in learning stereo correspondence. PAM [25] explores the pixel movements between views by" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.368, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: zhangshuo@bjtu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6488" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.226 + ], + "angle": 0, + "content": "finding the most similar features in parallax attention maps. Unlike stereo images, LFs have different views. The position offset scales in the attention maps are different since the baselines vary between the arranged views. In addition, if the matching information is occluded in one view, the disparity information of that point can be explored through other views. Therefore, how to map these attention maps to same scale and select information from the occlusion-free areas are key challenges that we should focus on." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.228, + 0.487, + 0.545 + ], + "angle": 0, + "content": "In this paper, we adapt the parallax attention to LFs and propose a novel unsupervised framework to capture the 3d cues, termed the Epipolar Consistent Attention Aggregation Network (ECAAN). Specifically, we first explicitly calculate the correlation between the reference point and all possible matching points along the epipolar lines to construct the parallax attention maps. To standardize disparity scales of the parallax attention maps, we introduce an Epipolar Consistent Scale Unification (ECSU) block by considering the consistency relationships. Furthermore, we explore the intra-properties and inter-relationships of parallax attention maps and design a Consistent Occlusion-Free Aggregation (COFA) block to filter out occlusions. Finally, the fused attention map is used to regress the disparity map of the scene. In addition, we design an improved photometric loss to constrain the model during training. Our method demonstrates outstanding performance in LF disparity estimation, particularly in large-disparity scenes, achieving more accurate predictions with improved efficiency. For example, ECAAN achieves a \\(68\\%\\) error reduction compared to the second-best method (see Fig. 1)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.559, + 0.233, + 0.573 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Most existing learning-based LF disparity estimation methods primarily focus on addressing two challenges: occlusions and large disparities. To mitigate matching errors caused by occlusions, various methods integrate view selection modules or specialized loss functions. Tsai et al. [24] explored matching costs and introduced an SAI selection module to generate attention maps, extracting meaningful information from the available views. Chen et al. [1] proposed a four-branch architecture for matching cost learning, incorporating two attention mechanisms to further refine the information by prioritizing branches with minimal occlusion and richer textures. Guo et al. [3] introduced an occlusion detection network, trained using a designed occlusion map, to process occluded and non-occluded regions separately, ensuring accurate disparity estimation with sharp edges. In recent years, Jin et al. [10] partitioned the entire LF into sub-LFs, feeding them into a multi-scale network to predict disparity maps alongside confidence maps based on the constructed cost volume. In addition, they proposed a simple strategy to fuse multiple disparity maps to deal with occlusion. Zhang et al. [31] proposed a two-stage pipeline" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "to focus on the occlusion problem, in which they generate multiple occlusion priors and act on the loss functions. Yan et al. [28] used the generative adversarial network to estimate the disparity map in an unsupervised manner. Zhou et al. [33] introduced a geometry-based occlusion model along with two loss functions to learn occlusion- and noise-invariant representations of the LF." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.907, + 0.47 + ], + "angle": 0, + "content": "Effectively managing large disparities in LFs is another open problem. Shi et al. [19] addressed this issue by decomposing LFs into multiple stereo-image pairs, employing a fine-tuned stereo estimation network to compute the disparity for each pair, which was subsequently fused into refined disparity maps. Some approaches [31] tackle large parallax scenes by reducing the angular resolution of the input. However, existing methods commonly generate cost volumes, leading to significant memory overhead, which constrains their ability to perform fine-grained disparity sampling and ultimately limits their performance. Given the advantages of global interaction, the self-attention mechanism in transformers facilitates pixel-wise similarity modeling, making them highly effective in stereo-matching [25]. However, their application in LF disparity estimation remains limited due to the relatively complex structure of LFs, which poses challenges in effectively integrating self-attention mechanisms within LF representations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.482, + 0.773, + 0.5 + ], + "angle": 0, + "content": "3. Epipolar Parallax Attention" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.508, + 0.907, + 0.586 + ], + "angle": 0, + "content": "In this paper, the LFs are represented as \\(\\mathcal{L}(u,v,x,y)\\in\\) \\(\\mathbb{R}^{U\\times V\\times H\\times W}\\) , where \\((u,v)\\) represent angular coordinates, and \\((x,y)\\) represent spatial coordinates [12]. By fixing angular coordinates, the 2D view \\(I_{(u,v)}\\in \\mathbb{R}^{H\\times W}\\) is obtained. The central view are represented as \\(I_{(0,0)}\\)" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.821, + 0.607 + ], + "angle": 0, + "content": "3.1. EPI Parallax Attention Calculation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Inspired by the PAM model [25], the self-attention mechanism explicitly computes the correlation between points in paired views. Unlike binocular vision, LFs exhibit correspondences along epipolar lines in multiple directions simultaneously. To better capture these relationships, we compute parallax attention maps on EPIs in various directions, rather than relying on pairwise view similarity as in [25]. Specifically, we design a horizontal EPI parallax attention calculation mechanism (H-EPA) and a vertical EPI parallax attention calculation mechanism (V-EPA) to capture the correlation between views. To introduce the H-EPA, we take a horizontal EPI feature \\(\\mathcal{F}_h\\in \\mathbb{R}^{V\\times W\\times C}\\) that corresponding to the central view as an example. Firstly, 2D EPI convolutional layers are applied to the \\(\\mathcal{F}_h\\) to realise the interaction of redundant information in EPIs. Then, we use the \\(1\\times 1\\) convolutions to get the query features \\(Q_{h}\\) and the key features \\(K_{h}\\) from \\(\\mathcal{F}_h\\) which are further flattened to a sequence of tokens with a length of \\(VW\\) (\\(Q_{h},V_{h}\\in \\mathbb{R}^{VW\\times C}\\)). \\(Q_{h}\\) is multiplied with \\(K_{h}\\) to obtain the hori" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6489" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.9, + 0.355 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.363, + 0.908, + 0.406 + ], + "angle": 0, + "content": "Figure 2. One typical example with the related parallax attention maps. The point 1 is visible in all views and the attention value achieves an apparent bright spot in one specific point in each pair view. The point 2, 3 are ambiguous in some views. These features are aggregated combining the occlusion masks to obtain accurate parallax attention." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.416, + 0.482, + 0.522 + ], + "angle": 0, + "content": "zontal epipolar parallax attention map \\(C_h \\in \\mathbb{R}^{VW \\times VW}\\) as \\(C_h = Q_h \\cdot K_h^T\\). Finally, \\(C_h\\) is reshaped to \\(V \\times V\\) parallax-attention maps \\(C_{(0,v_1) \\to (0,v_2)} \\in \\mathbb{R}^{W \\times W}\\), which records the similarity between any two positions along the horizontal epipolar lines. The process of V-EPA is the same as the H-EPA while dealing with the vertical EPI features \\(\\mathcal{F}_v \\in \\mathbb{R}^{U \\times H \\times C}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.532, + 0.383, + 0.549 + ], + "angle": 0, + "content": "3.2. Parallax Attention Maps Analyse" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.483, + 0.719 + ], + "angle": 0, + "content": "As shown in Fig. 2, we visualized the \\(V\\) horizontal parallax attention maps \\(C_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times W}\\) in \\(C_h\\). These maps are calculated by evaluating the similarity between each point in the central view \\(I_{(0,0)}\\) and every point along the corresponding epipolar line in the view \\(I_{(0,v)}\\), where brighter color means it is similar to the corresponding point in the central view. The disparity information is encoded through positional offsets of the most similar points relative to the diagonal line, which vary across different views. To effectively aggregate these parallax-attention maps, scale variations and occlusion remain key challenges." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Scale Issue: Since the disparity between views varies significantly for different baselines, the position offset scales of these parallax attention maps are different. In Fig. 2, for views \\( I_{(0, -3)} \\) and \\( I_{(0,3)} \\) that are far from the central view, the position offset is \\( 3 \\times \\) larger than the adjacent view \\( I_{(0, -1)} \\) and \\( I_{(0,1)} \\) due to the larger baseline. In the parallax attention maps, it is clear that the offsets relative to the diagonal line are larger in \\( M_{(0,0) \\to (0, -3)} \\) and \\( M_{(0,0) \\to (0,3)} \\). The offsets of the left view and right view are also symmetric around the diagonal line since the view movement direction is opposite. Coarsely resizing the scale of these parallax attention maps may break the encoded consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.568 + ], + "angle": 0, + "content": "information and introduce significant errors. To aggregate these attention maps, we propose an epipolar consistent re-sample and up-sample strategy (see Sec.4.2). We first crop and re-sample the attention map \\( M_{(0,0) \\to (0,v)} \\) according to the disparity range and form a new attention map. Then, the re-sampled attention maps are up-sampled heuristically to the same scale by placing their original attention values in the corresponding positions. As in Fig. 2, all up-sampled attention maps are mapped to the same scale and are consistent with the initial parallax information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Occlusion Issue: Another fundamental challenge is occlusion, where view consistency does not hold. In LFs, the occlusions may arise from various directions, resulting in different occlusion relationships for each map. Here we take \\( M_{(0,0)\\to (0,v)} \\) as a typical example in Fig. 2. If one point is visible in all views (as point 1), the attention value achieves an apparent bright spot at one specific point in each pair view, which is just its matching point. However, it is difficult to find the matching point for points that are occluded in part of views. For example, as the view moves from left to right, the point 2 is visible in the beginning but is then occluded by the foreground pencil. Therefore, the matching point with a bright spot is easy to find in view \\( I_{(0,-3/-2/-1/1)} \\) but is ambiguous in view \\( I_{(0,2/3)} \\). Similarly, the background point 3 is occluded by the foreground pencil in the left views but is visible in the right views. The corresponding matching point is concentrated in view \\( I_{(0,-1/1/2/3)} \\) but is dispersed in view \\( I_{(0,-3/-2)} \\). For these kinds of occluded points, labeled with blue and yellow boxes, directly fusing all the matching values or alignment information in all views together will bring errors. To alleviate the influence of the occluded views in disparity" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6490" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.861, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.436, + 0.908, + 0.479 + ], + "angle": 0, + "content": "Figure 3. An overview of our disparity estimation network. After feature extraction, the epipolar parallax attention maps are constructed by the H-EPA and V-EPA. The epipolar parallax attention maps are aggregated based on the epipolar consistent attention aggregation. Finally, the disparity map is generated through the disparity regression." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.486, + 0.688 + ], + "angle": 0, + "content": "estimation, we detect the occluded regions and aggregate the disparity information encoded in attention maps from non-occluded regions (see Sec. 4.3). Specifically, we calculate the occlusion masks for each view to label whether the point is occluded in the paired view. As shown, the two regions with blue boxes that are visible in the central view but occluded by the pencil in view \\( I_{(0, -3)} \\) are accurately masked in the occlusion masks. Combining with the occlusion mask, we calculate the occlusion-aware aggregation weights for different views as the last row in Fig. 2. The fused attention map is also visualized, in which we can find a matching point with high similarity for each pixel by using the information of non-occluded viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.703, + 0.476, + 0.722 + ], + "angle": 0, + "content": "4. Epipolar Consistent Attention Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.73, + 0.485, + 0.775 + ], + "angle": 0, + "content": "The proposed Epipolar Consistent Attention Aggregation Network aims to estimate the disparity map of the central view \\(D_{c}\\in \\mathbb{R}^{H\\times W}\\) from the LFs \\(\\mathcal{L}\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.202, + 0.803 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.485, + 0.902 + ], + "angle": 0, + "content": "The overall network structure is shown in Fig. 3. Given the input image \\(\\mathcal{L}\\), we first implement spatial convolution layers to encode the multi-scale features \\(\\mathcal{F}_e^i\\in \\mathbb{R}^{U\\times V\\times H_i\\times W_i\\times C}\\) in which \\(H_{i} = \\frac{H}{2^{i}}\\), \\(W_{i} = \\frac{W}{2^{i}}\\). Then, these encoded features are processed through a hierarchical decoder structure that deploys with the H-EPA and V-EPA(Sec. 3.1) to ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.49, + 0.909, + 0.751 + ], + "angle": 0, + "content": "plore the epipolar consistency. At each decoding stage \\(i\\), the features produced in the previous stage \\(\\mathcal{F}_d^{i - 1}\\) are up-sampled and fused with the \\(\\mathcal{F}_e^i\\) through spatial convolutions, producing \\(\\mathcal{F}_{fuse}^{i}\\). The \\(\\mathcal{F}_{fuse}^{i}\\) are then processed in parallel by H-EPA and V-EPA to obtain epipolar parallax attention maps \\(C_h^i /C_v^i\\) along with refined features \\(\\mathcal{F}_h^i /\\mathcal{F}_v^i\\). These refined features \\(\\mathcal{F}_h^i /\\mathcal{F}_v^i\\) are subsequently combined and fed into the next decoder stage. Once all hierarchical epipolar parallax attention maps \\(\\{C_h^i,C_c^i\\}_{i = 1}^N\\) are obtained, they are up-sampled and summed to form the final attention maps \\(C_h\\) and \\(C_v\\). To estimate depth maps of central views, the attention maps calculated from central views with other views are aggregated by the epipolar consistent attention aggregation (ECAA) mechanism, which consists of ECSU block (Sec. 4.2) and COFA block (Sec. 4.3). Finally, the disparity is estimated as the weighted sum of disparity candidates based on the aggregated epipolar consistent attention \\(P_{att}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.759, + 0.907, + 0.8 + ], + "angle": 0, + "content": "\\[\nD _ {c} = \\sum_ {k = 0} ^ {D _ {u p}} k \\times P _ {a t t} (:,:, k) - \\frac {D _ {u p}}{2}. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.808, + 0.895, + 0.825 + ], + "angle": 0, + "content": "where \\(D_{up}\\) means the size of disparity candidates in \\(P_{att}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.838, + 0.85 + ], + "angle": 0, + "content": "4.2. Epipolar Consistent Scale Unification" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "To better aggregate the calculated attention maps from different views, we re-sample and up-sample these maps to a uniform scale under the constraint of epipolar consistency." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.958 + ], + "angle": 0, + "content": "6491" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.26 + ], + "angle": 0, + "content": "Epipolar Consistent Attention Re-Sample. For each \\(V - 1\\) view that moves horizontally from the central view, there are \\(H\\) attention maps \\(M_h\\) available that calculate the pairwise similarity of these views. We add the disparity constraint to \\(C_h\\) and get the constrained epipolar parallax attention maps \\(M_h\\). We re-sample each paired attention map \\(M_{(0,0)\\to (0,v)}\\) according to the disparity range \\([d_{min},d_{max}]\\) and form a new map \\(P_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times (|\\mathrm{v}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}\\). Similarly, the attention maps related to the vertical moved views are also re-sampled as \\(P_{(0,0)\\to (u,0)}\\in \\mathbb{R}^{H\\times (|\\mathrm{u}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.26, + 0.482, + 0.351 + ], + "angle": 0, + "content": "Epipolar Consistent Attention Up-Sample. For attention maps \\( P_{(0,0) \\to (0,v) / (u,0)} \\) that have different kinds of scale, we propose to use deconvolution layers with different parameters to provide specific location information for similarity score interpolation. Taking \\( P_{(0,0) \\to (0,v)} \\) as an example, the magnification factor \\( f \\) in the deconvolution layer is:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.365, + 0.482, + 0.38 + ], + "angle": 0, + "content": "\\[\nf _ {(0, 0) \\rightarrow (0, v)} = N / | v |, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.383, + 0.483, + 0.601 + ], + "angle": 0, + "content": "where \\(N\\) indicates the least common multiple of \\(\\{1,2,\\dots ,\\frac{V}{2}\\}\\). In such a way, the original similarity scores in the small disparity scale space are put into the corresponding positions in the large disparity scale space as in Fig. 3, resulting in up-sampled \\(P_{(0,0)\\to (0,v)}^{up}\\in \\mathbb{R}^{W\\times D_{up}}\\) \\((D_{up} = N\\times (d_{max} - d_{min}) + 1)\\). Therefore, compared with other up-sampled methods such as pixel shuffle [20], the deconvolution layer with \\(f\\) not only maps all attention to the same scale but also keeps the original matching information. Note that we manually flip half of the up-sampled parallax attentions since the pixel offsets are opposite for views that in different sides of the central view. We then combine all the processed consistent parallax attentions from the horizontal and vertical views \\(P_{(0,0)}^{up}\\in \\mathbb{R}^{(U + V - 2)\\times H\\times W\\times D_{up}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.607, + 0.432, + 0.623 + ], + "angle": 0, + "content": "4.3. Consistent Occlusion-Free Aggregation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.482, + 0.705 + ], + "angle": 0, + "content": "To effectively filter out occlusion-area information, we analyze both the intra-properties of individual parallax attention maps and the inter-relationships between them, leveraging this analysis to generate occlusion-aware fusion weights for consistent information aggregation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Occlusion Masks. The intra-properties of individual parallax attention maps reveal that when a point is occluded in one view, its similarity remains relatively low as analyzed in Sec. 3.2. Based on this, we perform occlusion detection on parallax-attention map \\( M_{h} \\) and \\( M_{v} \\) and generate occlusion prior knowledge \\( O_{(0,v)} \\), \\( O_{(u,0)} \\in \\mathbb{R}^{H \\times W} \\) of view \\( I_{(0,v)} \\), \\( I_{(u,0)} \\) with respect to \\( I_{(0,0)} \\). Specifically, we add parallax attention maps \\( M_{(0,v) \\to (0,0)} \\) and \\( M_{(u,0) \\to (0,0)} \\) along the non-normalized dimension and set an appropriate threshold \\( \\tau \\) (set to 0.9 in this paper) for binarization. Since occlusions usually appear at object edges, we utilize guided filter [7] to sharpen the edges before binarization. Finally, we obtain \\( (V - 1) \\) binary mask \\( O_{(0,v)} \\) and \\( (U - 1) \\) binary mask \\( O_{(u,0)} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.353 + ], + "angle": 0, + "content": "Consistency Weight. Given that disparity constitutes an intrinsic property of visual scenes, the inter-relationships between the up-sampled attention maps should inherently maintain consistency. Therefore, we focus the model on aggregating the attention maps that exhibit consistent disparity information. To implement this, we feed \\( P_{(0,0)}^{up} \\) into the transformer model along with a linear layer to calculate the consistency weights of different attention maps. Specifically, the \\( P_{(0,0)}^{up} \\) is reshaped to \\( H \\times W \\) sequence of tokens \\( T \\in \\mathbb{R}^{(U + V - 2) \\times D_{up}} \\). Afterward, \\( T \\) are multiplied separately by \\( W_{Q}, W_{K}, W_{V} \\in \\mathbb{R}^{D_{up} \\times c} \\) to generate the corresponding query, key, and value. Then, the dot-product operation is used on the query and key, followed by a softmax function to obtain the attention scores on the tokens. The attention scores do matrix multiplication with value to get the fusion attention feature and the fusion attention feature is fed into a linear layer to get the initial fusion weight \\( W_{init} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.518 + ], + "angle": 0, + "content": "Occlusion-Free Aggregation. Since occlusions often cause errors in the final regression results, we consider assigning small weights to the occluded parts when aggregating the attention maps. Specifically, we introduce the occlusion prior information \\( O_{(u,0)} \\) and \\( O_{(0,v)} \\) to suppress the information provided by the occluded views. First, we map occluded points in \\( O_{(u,0)} \\) and \\( O_{(0,v)} \\) to \\( -\\infty \\) and non-occluded point to 0. The mapping results from \\( O_{(u,0)} \\) and \\( O_{(0,v)} \\) are concatenated and added to \\( W_{init} \\). We then fed the added weight into a softmax layer to produce the occlusion-aware fusion weight \\( W_{fuse} \\in \\mathbb{R}^{(U + V - 2)\\times H\\times W} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.518, + 0.905, + 0.564 + ], + "angle": 0, + "content": "The up-sampled \\( P_{(0,0)}^{up} \\) are then multiplied by the broadcasted \\( W_{fuse} \\) and sum over the perspective dimension to obtain the fused epipolar consistent attention map \\( P_{att} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.57, + 0.905, + 0.607 + ], + "angle": 0, + "content": "\\[\nP _ {a t t} = \\sum_ {k = 1} ^ {U + V - 2} W _ {f u s e} (k,:,:) \\times P _ {(0, 0)} ^ {u p} (k,:,:) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.644 + ], + "angle": 0, + "content": "In this way, we integrate attention maps from different views to increase the robustness of our results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.652, + 0.661, + 0.667 + ], + "angle": 0, + "content": "4.4. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.905, + 0.749 + ], + "angle": 0, + "content": "Based on the occlusion mask \\(O_{(u,0)}\\) and \\(O_{(0,v)}\\), we improve the photometric loss [29] to make the model pay attention to non-occluded regions during training. Specifically, we use the final disparity map \\(D_{c}\\) to warp all views toward the central view and calculate the L1 loss between them:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.754, + 0.905, + 0.791 + ], + "angle": 0, + "content": "\\[\nL _ {p} = \\sum_ {u} ^ {U} \\sum_ {v} ^ {V} L _ {1} \\left(w \\left(I _ {(u, v)}, D _ {c}\\right), I _ {(0, 0)}\\right) \\times O _ {(u, v)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.905, + 0.858 + ], + "angle": 0, + "content": "where \\( w(\\cdot) \\) denotes a warping operator. Since we only have \\( O_{(u,0)} \\) and \\( O_{(0,v)} \\) that label the occluded regions in the cross-shape views in LFs, we propose a simple way to extend them to all views in LFs:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.864, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\nO _ {(u, v)} = \\left\\{ \\begin{array}{l l} 1, & O _ {(u, 0)} \\wedge O _ {(0, v)} = 1 \\\\ 0, & O _ {(u, 0)} \\vee O _ {(0, v)} = 0 \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6492" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.172, + 0.09, + 0.825, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison of MSE*100/Bp(0.3,0,1,0.05) on SLFD dataset. The best results are in bold." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.115, + 0.898, + 0.253 + ], + "angle": 0, + "content": "
MethodsLionRooster clockToy bricksElectro devices
MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)
SupEPINet [21]0.47631.45768.13882.9540.74036.86871.56285.1471.05734.65861.95777.4911.11239.80964.45581.793
LFAntNet [24]0.37212.99429.59449.3270.2786.83125.81350.9200.73811.60729.01852.5760.70313.81938.73160.993
UnSupUnCNN [16]1.44258.22184.86792.3739.21334.58675.02187.3832.25453.92282.17290.9502.18141.16372.97085.866
UnMonocular [32]0.76048.30587.37194.6730.66439.93173.60186.3701.22647.59471.09884.9822.73645.61275.23887.332
UnCon [14]1.19235.49375.39287.8910.49116.92853.88975.5971.74150.47179.47089.4743.60442.44375.65487.271
UnOcc [10]1.45453.46080.65090.2120.42111.14842.62768.1511.90661.45784.58791.5702.72040.83471.62284.935
UnMVFM [31]0.3608.76624.42047.9110.2615.79625.30348.6680.77210.04830.42755.7320.84213.63136.03857.880
ECAAN30.1731.5596.86416.4800.1772.58810.89631.1060.3323.74129.62454.4710.2909.09531.76755.710
ECAAN50.0871.1389.470332.9980.1010.8015.57814.3920.4083.20727.06847.2510.2115.63022.05744.819
" + }, + { + "type": "image", + "bbox": [ + 0.113, + 0.266, + 0.882, + 0.574 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.225, + 0.578, + 0.772, + 0.593 + ], + "angle": 0, + "content": "Figure 4. Visual comparison of the disparity maps estimated by different methods on SLFD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.612, + 0.484, + 0.754 + ], + "angle": 0, + "content": "If an occluder is labeled in \\( O_{(0,v)} \\), we consider that this occluder will also appear in the moving view along the vertical direction of \\( O_{(0,v)} \\). In addition, we also implement the photometric loss on disparity maps \\( D_{(0,v)} \\) and \\( D_{(u,0)} \\) using \\( L_p^h \\) and \\( L_p^v \\) to ensure the accuracy of upsampling, which is regressed using the up-sampled attention map \\( P_{(0,0)\\to (0,v)}^{up} \\) and \\( P_{(0,0)\\to (u,0)}^{up} \\). Finally, we train our model in an unsupervised manner, the overall loss includes photometric loss \\( L_p \\), \\( L_p^h \\) and \\( L_p^v \\), PAM loss \\( L_s \\) and smoothness loss \\( L_{sm} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.777, + 0.483, + 0.816 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} L _ {u n s u p} = L _ {p} + \\lambda_ {h v} \\left(L _ {p} ^ {h} + L _ {p} ^ {v}\\right) + \\lambda_ {s} L _ {s m} \\tag {6} \\\\ + \\lambda_ {P} \\left(0. 2 L _ {s} ^ {1} + 0. 3 L _ {s} ^ {2} + 0. 5 L _ {s} ^ {3}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( L_{s}^{i} \\) extended by the PAM loss [25] and is calculated on the all parallax attention maps obtained in the \\( i \\)-th decoding stage. The \\( L_{sm} \\) [23] is calculated on \\( D_{(0,v)} \\), \\( D_{(u,0)} \\) and \\( D_{c} \\) to encourage local smoothness of the disparity." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.61, + 0.64, + 0.628 + ], + "angle": 0, + "content": "5. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.733, + 0.651 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Our model is implemented using the PyTorch framework and trained on an NVidia RTX A5000 GPU. During training, we crop views into \\(100 \\times 100\\) patches as the model input. We train our model by minimizing the \\(L_{unsup}\\) and \\(\\lambda_{hv}, \\lambda_s, \\lambda_P\\) in Eq. 6 are set to 0.5, 0.01, 1. The models are optimized using the Adam optimizer [2] with \\(\\beta_1 = 0.9\\) and \\(\\beta_2 = 0.999\\). The initial learning rate is set to \\(1e^{-3}\\) and the batch size is set to 2. Besides, we randomly augment the datasets by flipping the images horizontally, vertically, or rotating 90 degrees. Typically, it roughly takes 2 days of training for our model to converge." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.89, + 0.85 + ], + "angle": 0, + "content": "5.2. Comparisons with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We conduct a comparative analysis with SOTA methods, including two non-learning-based methods (i.e., SPO [30], OAVC [6]), four supervised learning-based meth" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "6493" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of the MSE*100/Bp(0.07) metrics on HCI dataset. The best results among all methods are in red. The best and second best results among unsupervised methods are in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.129, + 0.912, + 0.281 + ], + "angle": 0, + "content": "
MethodBackgammonDotsPyramidsStripsBoxesCottonDinoSideboardBedroomBicycleHerbsOrigamiAvg
MSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/Bp
NonSPO [30]4.59/3.785.24/16.270.04/0.866.96/14.999.11/15.891.31/2.590.31/2.181.02/9.300.21/4.865.57/10.9111.23/8.262.03/11.693.97/8.47
OAVC [6]3.84/3.1216.58/69.110.04/0.831.32/2.906.99/16.140.60/2.550.27/3.941.05/12.420.21/4.924.89/12.2210.36/8.731.48/12.563.97/12.45
SupEPINet [21]3.71/3.501.48/2.490.007/0.160.93/2.465.97/12.340.20/0.450.16/1.210.80/4.460.20/2.304.60/9.619.29/10.961.83/5.812.43/4.65
LFattNet [24]3.65/3.131.43/1.430.004/0.200.89/2.934.00/2.790.21/9.510.09/5.220.53/4.820.37/2.793.35/2.796.61/5.221.73/4.821.91/3.80
AttMLFNet [1]3.86/3.231.04/1.610.003/0.170.81/2.933.84/11.140.06/0.200.04/0.440.40/2.690.13/2.073.80/8.846.37/5.430.99/4.411.78/3.60
OACCNet [26]3.94/3.931.42/1.510.004/0.160.85/2.922.89/10.700.16/0.310.08/0.970.54/3.350.15/2.302.91/8.086.56/6.520.88/4.071.69/3.74
UnsupUnMonocular [32]11.83/12.312.54/3.650.027/0.262.68/11.149.27/22.882.91/3.811.03/5.402.77/10.950.42/7.419.23/20.1026.82/13.443.68/10.956.10/10.19
UnOcc [10]6.68/14.376.57/45.340.213/7.355.20/41.367.45/26.240.80/8.460.63/8.251.79/14.200.39/12.696.23/21.6513.94/16.961.92/19.824.32/19.72
UnLFDisp [33]6.86/5.201.44/1.530.016/0.271.80/5.128.32/20.182.04/3.040.78/4.171.85/7.680.35/7.367.49/16.5519.43/11.802.94/9.674.44/7.71
UnGAN [28]5.38/10.481.74/2.040.013/0.302.39/9.048.01/19.870.76/1.940.74/4.713.39/7.080.55/6.917.96/18.9215.58/12.763.61/10.334.18/8.70
ECAAN10.18/8.063.03/4.770.021/0.351.86/29.847.87/18.540.67/1.920.63/4.921.35/9.050.27/6.407.21/17.1214.81/10.811.99/11.044.16/10.24
" + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.293, + 0.889, + 0.482 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.493, + 0.908, + 0.524 + ], + "angle": 0, + "content": "Figure 5. Visual comparison of different methods on real-world scenes 'Lego Bulldozer' and 'Lego Knights'. The input \\(3 \\times 3\\) views are selected (label in the above line) with different intervals to make the disparity ranges increase gradually." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.486, + 0.675 + ], + "angle": 0, + "content": "ods (i.e., EPINet [21], LFAntNet [24], AttMLFNet [1], OACCNet [26]), and seven unsupervised learning-based methods (i.e., UnCNN [16], UnMonocular [32], UnCon [14], UnLFDisp [33], UnGAN [28], UnOcc [10], UnMVFM [31]). The Bad pixel (Bp) ratio [9], which measures the percentage of incorrectly estimated pixels \\(>t\\), and the Mean Squared Error (MSE) are used as metrics. To comprehensively evaluate model performance, we conduct experiments on both large and small disparity datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.685, + 0.3, + 0.7 + ], + "angle": 0, + "content": "5.2.1. Large Disparity Scenes" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Synthetic Datasets. Following most approaches [31], synthetic LFs in SLFD are selected for training and testing, with interview disparities in [-20, 20]. Table 1 illustrates the quantitative results of different methods on several scenes from SLFD. Due to the memory burden of constructing the cost-volume, these cost-volume based methods use the central \\(3 \\times 3\\) views as input. To ensure fairness of the experiments, we respectively trained our models for \\(3 \\times 3\\) and \\(5 \\times 5\\) angular resolutions, labeled as \\(ECAAN_{3}\\) and \\(ECAAN_{5}\\). Among the LF image methods, our method achieves the lowest MSE and Bp values across all scenes, indicating its ability to better capture the consistency relationship of large disparity scenes and produce superior disparity maps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.674 + ], + "angle": 0, + "content": "Specifically, \\(ECAAN_{3}\\) achieves a significant MSE reduction of up to 5.427 (Rooster clock) and 0.273 (Lion), while also lowering \\(\\mathrm{Bp}(0.3)\\) by up to 7.628 (Lion) and 4.025 (Toy bricks) over the UnMVFM. The visual results are shown in Fig. 4, with a different color map from the LFs for distinction. Other approaches tend to generate noisy predictions and exhibit significant accuracy degradation. In contrast, our results have better visual qualities with smaller errors compared to others, especially in the edge regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Real Datasets. To assess the model's capability to process real-world scenes, we test the models on the LFs captured by the moving camera [22]. Since GT is not available, all models are trained in SLFD. For each scene, we select different \\(3 \\times 3\\) views from the \\(17 \\times 17\\) LFs with various disparity ranges to evaluate the robustness of the methods. As shown in Fig. 5, the disparity of LF is gradually larger when choosing views at different angular positions. Our method consistently achieves superior accuracy across inputs with varying disparities, preserving sharp boundaries and intricate structural details. Due to the predetermined fixed sampling intervals in cost volume construction, these models are prone to pronounced stratification artifacts. This issue is particularly evident in regions with smoothly varying disparities or occlusions, resulting in abrupt depth discontinu" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6494" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.186, + 0.09, + 0.388, + 0.102 + ], + "angle": 0, + "content": "Table 3. Result of variant models." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.108, + 0.48, + 0.23 + ], + "angle": 0, + "content": "
ModuleLossMSE / Bp (0.3)
ECSUCOFALpLph/vLionRoosterToyElectro
(1)××all✓4.96/21.260.37/2.933.81/33.052.34/20.62
(2)Linear×all✓2.48/14.690.30/2.741.84/24.021.88/19.27
(3)Dconv×all✓2.06/6.790.25/2.461.14/13.421.51/18.66
(4)Dconv✓××0.26/2.170.17/1.020.83/3.990.52/9.91
(5)Dconv✓sub✓0.16/1.460.14/0.970.59/3.420.42/8.41
(6)Dconv✓all✓0.08/1.130.10/0.810.41/3.200.21/5.63
" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.242, + 0.473, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.098, + 0.468, + 0.475, + 0.483 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on real-world small disparity LFs." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.499, + 0.341, + 0.515 + ], + "angle": 0, + "content": "ities and unnatural \"layering\" effects." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.539, + 0.296, + 0.554 + ], + "angle": 0, + "content": "5.2.2. Small Disparity Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.483, + 0.76 + ], + "angle": 0, + "content": "Synthetic Datasets. Regarding the synthetic LFs with small disparity, we use LFs from the HCI benchmark [9, 27] to verify the performance of the models on the synthetic datasets, in which the LFs have a disparity range of \\([-4, 4]\\). The results obtained using LFs with an angular resolution of \\(7 \\times 7\\) are illustrated in Table 2. Due to the lack of ground-truth supervision, the performance of our model and nonlearning models is worse than the supervised learning methods [1, 21, 24, 26] in some scenes. Compared with other unsupervised methods, our approach achieves the lowest MSE metric and outperforms others in the Bp metric across most scenes. More comparisons on HCI benchmark are provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Real Datasets. Following [10], the images from [11, 18, 22] are used for real-world dense LF comparison. All unsupervised methods are retrained on the 100 images from [11]. We show the results in real-world LFs [11, 18, 22] in Fig. 6. Due to the difference between the synthetic dataset and the real-world scene, the performance of the supervised models degrades in the real-world scene. Compared with other models, our model is able to better preserve fine structure in real scenes with complex occlusions." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.905, + 0.199 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.206, + 0.861, + 0.22 + ], + "angle": 0, + "content": "Figure 7. Visual comparison of the variant models." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.236, + 0.722, + 0.252 + ], + "angle": 0, + "content": "5.3. Ablation Investigation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.907, + 0.652 + ], + "angle": 0, + "content": "In this subsection, we conduct ablation experiments on SLFD using the \\(ECAAN_{5}\\). The results are shown in Table 3 and Fig. 7. We first design the following models to verify the effectiveness of our proposed epipolar consistent up-sample strategy: Model-1 directly regress all the parallax attention to disparity and average them as the final disparity map. Model-2 replaces the deconvolution layer using a simple linear layer to upsample the attention maps to the same scale. Model-2 performs better than Model-1, suggesting that integrating information in the attention level keeps more information than in the disparity level. Compared Model-2 with Model-3, our designed upsample strategy makes it easier to consistently map the attention to the same scale. We further verify the effectiveness of the proposed occlusion-aware aggregation. Model-3 removes the COFA while generating the mask for the loss. Compared Model-3 with the final model Model-6, the occlusion prior information reduces the weights of occluded regions and better uses the information extracted from the occlusion-free view. We further remove the occlusion mask on the loss, respectively. In Model-4 we remove all the occlusion masks in loss \\(L_{unsup}\\) and in Model-5 we only apply subocclusion masks \\(O_{(u,0)}\\) and \\(O_{(0,v)}\\). The results indicate that using the designed occlusion masks, the photometric supervision is more accurate and the performances are improved. More results are presented in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.856 + ], + "angle": 0, + "content": "In this paper, we propose a LF disparity estimation network based on the proposed epipolar consistent parallax attention. First, we formulate the horizontal/vertical EPI parallax attention calculation mechanism and analyze the occlusion and scale issues in the LF process. Then, we apply the EPI parallax attention to the disparity estimation task. Specifically, we design an epipolar consistent attention aggregation mechanism to unify the offset scale and aggregate the consistency parallax information of the non-occluded areas. Experimental results show that our method performs favorably against state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6495" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Jiaxin Chen, Shuo Zhang, and Youfang Lin. Attention-based multi-level fusion network for light field depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1009-1017, 2021. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.172, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Yangling Chen, Shuo Zhang, Song Chang, and Youfang Lin. Light field reconstruction using efficient pseudo 4d epipolar-aware structure. IEEE Transactions on Computational Imaging, 8:397-410, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.228, + 0.483, + 0.295 + ], + "angle": 0, + "content": "[3] Mantang Guo, Jing Jin, Hui Liu, and Junhui Hou. Learning dynamic interpolation for extremely sparse light fields with wide baselines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2450-2459, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.298, + 0.482, + 0.366 + ], + "angle": 0, + "content": "[4] Yulan Guo, Mohammed Bennamoun, Ferdous Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11): 2270-2287, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.482, + 0.422 + ], + "angle": 0, + "content": "[5] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.482, + 0.478 + ], + "angle": 0, + "content": "[6] Kang Han, Wei Xiang, Eric Wang, and Tao Huang. A novel occlusion-aware vote cost for light field depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8022-8035, 2021. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.482, + 0.521 + ], + "angle": 0, + "content": "[7] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6):1397-1409, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.523, + 0.482, + 0.577 + ], + "angle": 0, + "content": "[8] Stefan Heber, Wei Yu, and Thomas Pock. Neural epi-volume networks for shape from light field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2260, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.482, + 0.647 + ], + "angle": 0, + "content": "[9] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision, pages 19-34. Springer, 2016. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.649, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[10] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[11] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM Transactions on Graphics, 35(6):1-10, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.748, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[12] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the annual conference on Computer Graphics and Interactive Techniques. ACM, 1996. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[13] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Lili Lin, Qiujiang Li, Bin Gao, Yuxiang Yan, Wenhui Zhou, and Ercan Engin Kuruoglu. Unsupervised learning of light field depth estimation with spatial and angular consistencies. Neurocomput., 501(C):113-122, 2022. 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[15] In Kyu Park, Kyoung Mu Lee, et al. Robust light field depth estimation using occlusion-noise aware data costs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2484-2497, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[16] Jiayong Peng, Zhiwei Xiong, Dong Liu, and Xuejin Chen. Unsupervised depth estimation from light field using a convolutional neural network. In Proceedings of the IEEE International Conference on 3D Vision, pages 295-303. IEEE, 2018. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[17] Jiayong Peng, Zhiwei Xiong, Yicheng Wang, Yueyi Zhang, and Dong Liu. Zero-shot depth estimation from light field using a convolutional neural network. IEEE Transactions on Computational Imaging, 6:682-696, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.275, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[18] Martin Rerabek and Touradj Ebrahimi. New light field image dataset. In Proceedings of the International Conference on Quality of Multimedia Experience, 2016. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[19] Jinglei Shi, Xiaoran Jiang, and Christine Guillemot. A framework for learning depth from a flexible subset of dense and sparse light field views. IEEE Transactions on Image Processing, 28(12):5867-5880, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[20] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.457, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[21] Changha Shin, Hae-Gon Jeon, Youngjin Yoon, In So Kweon, and Seon Joo Kim. Epinet: A fully-convolutional neural network using epipolar geometry for depth from light field images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4748–4757, 2018. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.539, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[22] Abhilash Sunder Raj, Michael Lowney, Raj Shah, and Gordon Wetzstein. The stanford litro light field archive. http://lightfields.stanford.edu/LF2016.html, 2016. Accessed: Aug. 22, 2020. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.596, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[23] C. Tomasi and R. Manduchi. Bilateral filtering for gray and color images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 839-846, 1998. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.638, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[24] Yu-Ju Tsai, Yu-Lun Liu, Ming Ouhyoung, and Yung-Yu Chuang. Attention-based view selection networks for lightfield disparity estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12095-12103, 2020. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[25] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2108-2125, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.777, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[26] Yingqian Wang, Longguang Wang, Zhengyu Liang, Jungang Yang, Wei An, and Yulan Guo. Occlusion-aware cost constructor for light field depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19809-19818, 2022. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[27] Sven Wanner, Stephan Meister, and Bastian Goldluecke. Datasets and benchmarks for densely sampled 4d light fields. In Proceedings of the Conference on Vision, Modeling, and Visualization, pages 225-226. ACM, 2013. 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "6496" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Wenbin Yan, Xiaogang Zhang, and Hua Chen. Occlusion-aware unsupervised light field depth estimation based on multi-scale gans. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6318-6333, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.149, + 0.483, + 0.205 + ], + "angle": 0, + "content": "[29] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1983-1992, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.207, + 0.483, + 0.261 + ], + "angle": 0, + "content": "[30] Shuo Zhang, Hao Sheng, Chao Li, Jun Zhang, and Zhang Xiong. Robust depth estimation for light field via spinning parallelogram operator. Computer Vision and Image Understanding, 145:148-159, 2016. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[31] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Unsupervised light field depth estimation via multi-view feature matching with occlusion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2261-2273, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.333, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[32] Wenhui Zhou, Enci Zhou, Gaomin Liu, Lili Lin, and Andrew Lumsdaine. Unsupervised monocular depth estimation from light field image. IEEE Transactions on Image Processing, 29:1606-1617, 2020. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.39, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[33] Wenhui Zhou, Lili Lin, Yongjie Hong, Qiujuan Li, Xingfa Shen, and Ercan Engin Kuruoglu. Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2023. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.483, + 0.459 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6497" + } + ] +] \ No newline at end of file diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_origin.pdf b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9926cfd237d150498d50721e6b3b3893c3bbbe42 --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/74749a68-b688-4175-99d0-00c425f7c7e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b0617c7d10c68969e1783183cb3bf2e1f2f4678e061f21de1e6059d68f0bbead +size 13620780 diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/full.md b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..57c815631fc43b3a91ab5d25432e319aa529a67e --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/full.md @@ -0,0 +1,233 @@ +# Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation + +Chen Gao, Shuo Zhang* Youfang Lin +Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China. + +{gaochen, zhangshuo, yflin}@bjtu.edu.cn + +# Abstract + +Disparity estimation is an essential step in processing and analyzing Light Field (LF) images. Recent methods construct the cost volume to exploit the correspondence of the LFs over the preset maximum disparity, limiting them to process the large parallax scenes. Different from constructing cost volume, the self-attention mechanism calculates the parallax attention between epipolar lines to find the matching points. However, for LFs that have different views, the related disparity scales are different in parallax attention since the baselines with the central view are different. Moreover, if the matching information is occluded in one view, the disparity information can be explored through other views. Therefore, mapping these attentions to the same scale and selecting effective matching information are key points for disparity estimation from parallax attention. In this paper, we explore parallax attention for LF and design an unsupervised method, named Epipolar Consistent Attention Aggregation Network (ECAAN). We first introduce an epipolar consistent scale unification block by considering the consistency relationships to standardize disparity scales of the parallax attention maps. Based on the intra-properties and inter-reationships of parallax attention, we further propose a consistent occlusion-free aggregation block to integrate the information from the occlusion-free areas. In addition, we design an improved photometric loss to constrain the model. ECAAN achieves state-of-the-art performance in LF depth estimation. Notably, ECAAN attains a mean square error of 0.2 on large-disparity LF datasets, achieving a $68\%$ error reduction compared to the second-best method. + +# 1. Introduction + +Light Field (LF) disparity estimation aims to infer the 3D scene geometry from the captured LF images, which is ben- + +![](images/32f3b947a660a00eb71532d343a7ce24feb3b21b187deeb19a8f48f8c43f7435.jpg) +Figure 1. Comparative analysis with state-of-the-art methods on the SLFD dataset. Left: Performance metrics of models on Bp across thresholds (0.3, 0.1, 0.05). Right: Quantitative results for MSE ( $\times$ 100) and Inference Time (seconds). Our method achieves state-of-the-art performance across all metrics while maintaining competitive computational efficiency. Lower values of the Bp and MSE metrics indicate better performance. + +eficial for various advanced computer vision applications ranging from 3D reconstruction to robotics and autonomous driving [4, 5]. The key aspect of LF disparity estimation is to identify corresponding pixels across different views. + +At present, both non-learning methods [15] and learning methods [8, 10, 16, 17, 24] explore geometry information between LF images based on the constructed cost volume. These methods usually construct a cost volume within a predefined disparity range, and compute the weighted sum of those pre-defined disparity candidates to obtain the final disparity prediction. However, such a design suffers from a high computational and memory burden, which limits them in constructing a full-range cost volume for handling large disparities scenes. Moreover, the predefined disparity range inherently constrains the performance of cost-volume-based approaches, making them less effective in handling scenes with extreme parallax variations. + +Recently, some methods [13] attempt to use self-attention to explore consistency in stereo matching and show great potential in learning stereo correspondence. PAM [25] explores the pixel movements between views by + +finding the most similar features in parallax attention maps. Unlike stereo images, LFs have different views. The position offset scales in the attention maps are different since the baselines vary between the arranged views. In addition, if the matching information is occluded in one view, the disparity information of that point can be explored through other views. Therefore, how to map these attention maps to same scale and select information from the occlusion-free areas are key challenges that we should focus on. + +In this paper, we adapt the parallax attention to LFs and propose a novel unsupervised framework to capture the 3d cues, termed the Epipolar Consistent Attention Aggregation Network (ECAAN). Specifically, we first explicitly calculate the correlation between the reference point and all possible matching points along the epipolar lines to construct the parallax attention maps. To standardize disparity scales of the parallax attention maps, we introduce an Epipolar Consistent Scale Unification (ECSU) block by considering the consistency relationships. Furthermore, we explore the intra-properties and inter-relationships of parallax attention maps and design a Consistent Occlusion-Free Aggregation (COFA) block to filter out occlusions. Finally, the fused attention map is used to regress the disparity map of the scene. In addition, we design an improved photometric loss to constrain the model during training. Our method demonstrates outstanding performance in LF disparity estimation, particularly in large-disparity scenes, achieving more accurate predictions with improved efficiency. For example, ECAAN achieves a $68\%$ error reduction compared to the second-best method (see Fig. 1). + +# 2. Related Work + +Most existing learning-based LF disparity estimation methods primarily focus on addressing two challenges: occlusions and large disparities. To mitigate matching errors caused by occlusions, various methods integrate view selection modules or specialized loss functions. Tsai et al. [24] explored matching costs and introduced an SAI selection module to generate attention maps, extracting meaningful information from the available views. Chen et al. [1] proposed a four-branch architecture for matching cost learning, incorporating two attention mechanisms to further refine the information by prioritizing branches with minimal occlusion and richer textures. Guo et al. [3] introduced an occlusion detection network, trained using a designed occlusion map, to process occluded and non-occluded regions separately, ensuring accurate disparity estimation with sharp edges. In recent years, Jin et al. [10] partitioned the entire LF into sub-LFs, feeding them into a multi-scale network to predict disparity maps alongside confidence maps based on the constructed cost volume. In addition, they proposed a simple strategy to fuse multiple disparity maps to deal with occlusion. Zhang et al. [31] proposed a two-stage pipeline + +to focus on the occlusion problem, in which they generate multiple occlusion priors and act on the loss functions. Yan et al. [28] used the generative adversarial network to estimate the disparity map in an unsupervised manner. Zhou et al. [33] introduced a geometry-based occlusion model along with two loss functions to learn occlusion- and noise-invariant representations of the LF. + +Effectively managing large disparities in LFs is another open problem. Shi et al. [19] addressed this issue by decomposing LFs into multiple stereo-image pairs, employing a fine-tuned stereo estimation network to compute the disparity for each pair, which was subsequently fused into refined disparity maps. Some approaches [31] tackle large parallax scenes by reducing the angular resolution of the input. However, existing methods commonly generate cost volumes, leading to significant memory overhead, which constrains their ability to perform fine-grained disparity sampling and ultimately limits their performance. Given the advantages of global interaction, the self-attention mechanism in transformers facilitates pixel-wise similarity modeling, making them highly effective in stereo-matching [25]. However, their application in LF disparity estimation remains limited due to the relatively complex structure of LFs, which poses challenges in effectively integrating self-attention mechanisms within LF representations. + +# 3. Epipolar Parallax Attention + +In this paper, the LFs are represented as $\mathcal{L}(u,v,x,y)\in$ $\mathbb{R}^{U\times V\times H\times W}$ , where $(u,v)$ represent angular coordinates, and $(x,y)$ represent spatial coordinates [12]. By fixing angular coordinates, the 2D view $I_{(u,v)}\in \mathbb{R}^{H\times W}$ is obtained. The central view are represented as $I_{(0,0)}$ + +# 3.1. EPI Parallax Attention Calculation + +Inspired by the PAM model [25], the self-attention mechanism explicitly computes the correlation between points in paired views. Unlike binocular vision, LFs exhibit correspondences along epipolar lines in multiple directions simultaneously. To better capture these relationships, we compute parallax attention maps on EPIs in various directions, rather than relying on pairwise view similarity as in [25]. Specifically, we design a horizontal EPI parallax attention calculation mechanism (H-EPA) and a vertical EPI parallax attention calculation mechanism (V-EPA) to capture the correlation between views. To introduce the H-EPA, we take a horizontal EPI feature $\mathcal{F}_h\in \mathbb{R}^{V\times W\times C}$ that corresponding to the central view as an example. Firstly, 2D EPI convolutional layers are applied to the $\mathcal{F}_h$ to realise the interaction of redundant information in EPIs. Then, we use the $1\times 1$ convolutions to get the query features $Q_{h}$ and the key features $K_{h}$ from $\mathcal{F}_h$ which are further flattened to a sequence of tokens with a length of $VW$ ( $Q_{h},V_{h}\in \mathbb{R}^{VW\times C}$ ). $Q_{h}$ is multiplied with $K_{h}$ to obtain the hori + +![](images/a47acb8ff30d05ebb88177ece180b5269e22890ae905831d5e62332a133cfa12.jpg) +Figure 2. One typical example with the related parallax attention maps. The point 1 is visible in all views and the attention value achieves an apparent bright spot in one specific point in each pair view. The point 2, 3 are ambiguous in some views. These features are aggregated combining the occlusion masks to obtain accurate parallax attention. + +zontal epipolar parallax attention map $C_h \in \mathbb{R}^{VW \times VW}$ as $C_h = Q_h \cdot K_h^T$ . Finally, $C_h$ is reshaped to $V \times V$ parallax-attention maps $C_{(0,v_1) \to (0,v_2)} \in \mathbb{R}^{W \times W}$ , which records the similarity between any two positions along the horizontal epipolar lines. The process of V-EPA is the same as the H-EPA while dealing with the vertical EPI features $\mathcal{F}_v \in \mathbb{R}^{U \times H \times C}$ . + +# 3.2. Parallax Attention Maps Analyse + +As shown in Fig. 2, we visualized the $V$ horizontal parallax attention maps $C_{(0,0)\to (0,v)}\in \mathbb{R}^{W\times W}$ in $C_h$ . These maps are calculated by evaluating the similarity between each point in the central view $I_{(0,0)}$ and every point along the corresponding epipolar line in the view $I_{(0,v)}$ , where brighter color means it is similar to the corresponding point in the central view. The disparity information is encoded through positional offsets of the most similar points relative to the diagonal line, which vary across different views. To effectively aggregate these parallax-attention maps, scale variations and occlusion remain key challenges. + +Scale Issue: Since the disparity between views varies significantly for different baselines, the position offset scales of these parallax attention maps are different. In Fig. 2, for views $I_{(0, -3)}$ and $I_{(0,3)}$ that are far from the central view, the position offset is $3 \times$ larger than the adjacent view $I_{(0, -1)}$ and $I_{(0,1)}$ due to the larger baseline. In the parallax attention maps, it is clear that the offsets relative to the diagonal line are larger in $M_{(0,0) \to (0, -3)}$ and $M_{(0,0) \to (0,3)}$ . The offsets of the left view and right view are also symmetric around the diagonal line since the view movement direction is opposite. Coarsely resizing the scale of these parallax attention maps may break the encoded consistency + +information and introduce significant errors. To aggregate these attention maps, we propose an epipolar consistent re-sample and up-sample strategy (see Sec.4.2). We first crop and re-sample the attention map $M_{(0,0) \to (0,v)}$ according to the disparity range and form a new attention map. Then, the re-sampled attention maps are up-sampled heuristically to the same scale by placing their original attention values in the corresponding positions. As in Fig. 2, all up-sampled attention maps are mapped to the same scale and are consistent with the initial parallax information. + +Occlusion Issue: Another fundamental challenge is occlusion, where view consistency does not hold. In LFs, the occlusions may arise from various directions, resulting in different occlusion relationships for each map. Here we take $M_{(0,0)\to (0,v)}$ as a typical example in Fig. 2. If one point is visible in all views (as point 1), the attention value achieves an apparent bright spot at one specific point in each pair view, which is just its matching point. However, it is difficult to find the matching point for points that are occluded in part of views. For example, as the view moves from left to right, the point 2 is visible in the beginning but is then occluded by the foreground pencil. Therefore, the matching point with a bright spot is easy to find in view $I_{(0,-3/-2/-1/1)}$ but is ambiguous in view $I_{(0,2/3)}$ . Similarly, the background point 3 is occluded by the foreground pencil in the left views but is visible in the right views. The corresponding matching point is concentrated in view $I_{(0,-1/1/2/3)}$ but is dispersed in view $I_{(0,-3/-2)}$ . For these kinds of occluded points, labeled with blue and yellow boxes, directly fusing all the matching values or alignment information in all views together will bring errors. To alleviate the influence of the occluded views in disparity + +![](images/c1aef4ad36e2fa5160877a654ea38e7d2b240e1bf5db45fe9b54156ef93a24a5.jpg) +Figure 3. An overview of our disparity estimation network. After feature extraction, the epipolar parallax attention maps are constructed by the H-EPA and V-EPA. The epipolar parallax attention maps are aggregated based on the epipolar consistent attention aggregation. Finally, the disparity map is generated through the disparity regression. + +estimation, we detect the occluded regions and aggregate the disparity information encoded in attention maps from non-occluded regions (see Sec. 4.3). Specifically, we calculate the occlusion masks for each view to label whether the point is occluded in the paired view. As shown, the two regions with blue boxes that are visible in the central view but occluded by the pencil in view $I_{(0, -3)}$ are accurately masked in the occlusion masks. Combining with the occlusion mask, we calculate the occlusion-aware aggregation weights for different views as the last row in Fig. 2. The fused attention map is also visualized, in which we can find a matching point with high similarity for each pixel by using the information of non-occluded viewpoints. + +# 4. Epipolar Consistent Attention Aggregation + +The proposed Epipolar Consistent Attention Aggregation Network aims to estimate the disparity map of the central view $D_{c}\in \mathbb{R}^{H\times W}$ from the LFs $\mathcal{L}$ . + +# 4.1. Overview + +The overall network structure is shown in Fig. 3. Given the input image $\mathcal{L}$ , we first implement spatial convolution layers to encode the multi-scale features $\mathcal{F}_e^i\in \mathbb{R}^{U\times V\times H_i\times W_i\times C}$ in which $H_{i} = \frac{H}{2^{i}}$ , $W_{i} = \frac{W}{2^{i}}$ . Then, these encoded features are processed through a hierarchical decoder structure that deploys with the H-EPA and V-EPA(Sec. 3.1) to ex + +plore the epipolar consistency. At each decoding stage $i$ , the features produced in the previous stage $\mathcal{F}_d^{i - 1}$ are up-sampled and fused with the $\mathcal{F}_e^i$ through spatial convolutions, producing $\mathcal{F}_{fuse}^{i}$ . The $\mathcal{F}_{fuse}^{i}$ are then processed in parallel by H-EPA and V-EPA to obtain epipolar parallax attention maps $C_h^i /C_v^i$ along with refined features $\mathcal{F}_h^i /\mathcal{F}_v^i$ . These refined features $\mathcal{F}_h^i /\mathcal{F}_v^i$ are subsequently combined and fed into the next decoder stage. Once all hierarchical epipolar parallax attention maps $\{C_h^i,C_c^i\}_{i = 1}^N$ are obtained, they are up-sampled and summed to form the final attention maps $C_h$ and $C_v$ . To estimate depth maps of central views, the attention maps calculated from central views with other views are aggregated by the epipolar consistent attention aggregation (ECAA) mechanism, which consists of ECSU block (Sec. 4.2) and COFA block (Sec. 4.3). Finally, the disparity is estimated as the weighted sum of disparity candidates based on the aggregated epipolar consistent attention $P_{att}$ : + +$$ +D _ {c} = \sum_ {k = 0} ^ {D _ {u p}} k \times P _ {a t t} (:,:, k) - \frac {D _ {u p}}{2}. \tag {1} +$$ + +where $D_{up}$ means the size of disparity candidates in $P_{att}$ . + +# 4.2. Epipolar Consistent Scale Unification + +To better aggregate the calculated attention maps from different views, we re-sample and up-sample these maps to a uniform scale under the constraint of epipolar consistency. + +Epipolar Consistent Attention Re-Sample. For each $V - 1$ view that moves horizontally from the central view, there are $H$ attention maps $M_h$ available that calculate the pairwise similarity of these views. We add the disparity constraint to $C_h$ and get the constrained epipolar parallax attention maps $M_h$ . We re-sample each paired attention map $M_{(0,0)\to (0,v)}$ according to the disparity range $[d_{min},d_{max}]$ and form a new map $P_{(0,0)\to (0,v)}\in \mathbb{R}^{W\times (|\mathrm{v}|*(\mathrm{d}_{\mathrm{max}} - \mathrm{d}_{\mathrm{min}}) + 1)}$ . Similarly, the attention maps related to the vertical moved views are also re-sampled as $P_{(0,0)\to (u,0)}\in \mathbb{R}^{H\times (|\mathrm{u}|*(\mathrm{d}_{\mathrm{max}} - \mathrm{d}_{\mathrm{min}}) + 1)}$ . + +Epipolar Consistent Attention Up-Sample. For attention maps $P_{(0,0) \to (0,v) / (u,0)}$ that have different kinds of scale, we propose to use deconvolution layers with different parameters to provide specific location information for similarity score interpolation. Taking $P_{(0,0) \to (0,v)}$ as an example, the magnification factor $f$ in the deconvolution layer is: + +$$ +f _ {(0, 0) \rightarrow (0, v)} = N / | v |, \tag {2} +$$ + +where $N$ indicates the least common multiple of $\{1,2,\dots ,\frac{V}{2}\}$ . In such a way, the original similarity scores in the small disparity scale space are put into the corresponding positions in the large disparity scale space as in Fig. 3, resulting in up-sampled $P_{(0,0)\to (0,v)}^{up}\in \mathbb{R}^{W\times D_{up}}$ $(D_{up} = N\times (d_{max} - d_{min}) + 1)$ . Therefore, compared with other up-sampled methods such as pixel shuffle [20], the deconvolution layer with $f$ not only maps all attention to the same scale but also keeps the original matching information. Note that we manually flip half of the up-sampled parallax attentions since the pixel offsets are opposite for views that in different sides of the central view. We then combine all the processed consistent parallax attentions from the horizontal and vertical views $P_{(0,0)}^{up}\in \mathbb{R}^{(U + V - 2)\times H\times W\times D_{up}}$ . + +# 4.3. Consistent Occlusion-Free Aggregation. + +To effectively filter out occlusion-area information, we analyze both the intra-properties of individual parallax attention maps and the inter-relationships between them, leveraging this analysis to generate occlusion-aware fusion weights for consistent information aggregation. + +Occlusion Masks. The intra-properties of individual parallax attention maps reveal that when a point is occluded in one view, its similarity remains relatively low as analyzed in Sec. 3.2. Based on this, we perform occlusion detection on parallax-attention map $M_{h}$ and $M_{v}$ and generate occlusion prior knowledge $O_{(0,v)}$ , $O_{(u,0)} \in \mathbb{R}^{H \times W}$ of view $I_{(0,v)}$ , $I_{(u,0)}$ with respect to $I_{(0,0)}$ . Specifically, we add parallax attention maps $M_{(0,v) \to (0,0)}$ and $M_{(u,0) \to (0,0)}$ along the non-normalized dimension and set an appropriate threshold $\tau$ (set to 0.9 in this paper) for binarization. Since occlusions usually appear at object edges, we utilize guided filter [7] to sharpen the edges before binarization. Finally, we obtain $(V - 1)$ binary mask $O_{(0,v)}$ and $(U - 1)$ binary mask $O_{(u,0)}$ . + +Consistency Weight. Given that disparity constitutes an intrinsic property of visual scenes, the inter-relationships between the up-sampled attention maps should inherently maintain consistency. Therefore, we focus the model on aggregating the attention maps that exhibit consistent disparity information. To implement this, we feed $P_{(0,0)}^{up}$ into the transformer model along with a linear layer to calculate the consistency weights of different attention maps. Specifically, the $P_{(0,0)}^{up}$ is reshaped to $H \times W$ sequence of tokens $T \in \mathbb{R}^{(U + V - 2) \times D_{up}}$ . Afterward, $T$ are multiplied separately by $W_{Q}, W_{K}, W_{V} \in \mathbb{R}^{D_{up} \times c}$ to generate the corresponding query, key, and value. Then, the dot-product operation is used on the query and key, followed by a softmax function to obtain the attention scores on the tokens. The attention scores do matrix multiplication with value to get the fusion attention feature and the fusion attention feature is fed into a linear layer to get the initial fusion weight $W_{init}$ . + +Occlusion-Free Aggregation. Since occlusions often cause errors in the final regression results, we consider assigning small weights to the occluded parts when aggregating the attention maps. Specifically, we introduce the occlusion prior information $O_{(u,0)}$ and $O_{(0,v)}$ to suppress the information provided by the occluded views. First, we map occluded points in $O_{(u,0)}$ and $O_{(0,v)}$ to $-\infty$ and non-occluded point to 0. The mapping results from $O_{(u,0)}$ and $O_{(0,v)}$ are concatenated and added to $W_{init}$ . We then fed the added weight into a softmax layer to produce the occlusion-aware fusion weight $W_{fuse} \in \mathbb{R}^{(U + V - 2)\times H\times W}$ . + +The up-sampled $P_{(0,0)}^{up}$ are then multiplied by the broadcasted $W_{fuse}$ and sum over the perspective dimension to obtain the fused epipolar consistent attention map $P_{att}$ : + +$$ +P _ {a t t} = \sum_ {k = 1} ^ {U + V - 2} W _ {f u s e} (k,:,:) \times P _ {(0, 0)} ^ {u p} (k,:,:) \tag {3} +$$ + +In this way, we integrate attention maps from different views to increase the robustness of our results. + +# 4.4. Loss Function + +Based on the occlusion mask $O_{(u,0)}$ and $O_{(0,v)}$ , we improve the photometric loss [29] to make the model pay attention to non-occluded regions during training. Specifically, we use the final disparity map $D_{c}$ to warp all views toward the central view and calculate the L1 loss between them: + +$$ +L _ {p} = \sum_ {u} ^ {U} \sum_ {v} ^ {V} L _ {1} \left(w \left(I _ {(u, v)}, D _ {c}\right), I _ {(0, 0)}\right) \times O _ {(u, v)}, \tag {4} +$$ + +where $w(\cdot)$ denotes a warping operator. Since we only have $O_{(u,0)}$ and $O_{(0,v)}$ that label the occluded regions in the cross-shape views in LFs, we propose a simple way to extend them to all views in LFs: + +$$ +O _ {(u, v)} = \left\{ \begin{array}{l l} 1, & O _ {(u, 0)} \wedge O _ {(0, v)} = 1 \\ 0, & O _ {(u, 0)} \vee O _ {(0, v)} = 0 \end{array} \right. \tag {5} +$$ + +Table 1. Quantitative comparison of MSE*100/Bp(0.3,0,1,0.05) on SLFD dataset. The best results are in bold. + +
MethodsLionRooster clockToy bricksElectro devices
MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)
SupEPINet [21]0.47631.45768.13882.9540.74036.86871.56285.1471.05734.65861.95777.4911.11239.80964.45581.793
LFAntNet [24]0.37212.99429.59449.3270.2786.83125.81350.9200.73811.60729.01852.5760.70313.81938.73160.993
UnSupUnCNN [16]1.44258.22184.86792.3739.21334.58675.02187.3832.25453.92282.17290.9502.18141.16372.97085.866
UnMonocular [32]0.76048.30587.37194.6730.66439.93173.60186.3701.22647.59471.09884.9822.73645.61275.23887.332
UnCon [14]1.19235.49375.39287.8910.49116.92853.88975.5971.74150.47179.47089.4743.60442.44375.65487.271
UnOcc [10]1.45453.46080.65090.2120.42111.14842.62768.1511.90661.45784.58791.5702.72040.83471.62284.935
UnMVFM [31]0.3608.76624.42047.9110.2615.79625.30348.6680.77210.04830.42755.7320.84213.63136.03857.880
ECAAN30.1731.5596.86416.4800.1772.58810.89631.1060.3323.74129.62454.4710.2909.09531.76755.710
ECAAN50.0871.1389.470332.9980.1010.8015.57814.3920.4083.20727.06847.2510.2115.63022.05744.819
+ +![](images/0981def47ae9d544ddd90c8dcb29d58eaf26cab1d20dc5186b63f7800e9aedb4.jpg) +Figure 4. Visual comparison of the disparity maps estimated by different methods on SLFD. + +If an occluder is labeled in $O_{(0,v)}$ , we consider that this occluder will also appear in the moving view along the vertical direction of $O_{(0,v)}$ . In addition, we also implement the photometric loss on disparity maps $D_{(0,v)}$ and $D_{(u,0)}$ using $L_p^h$ and $L_p^v$ to ensure the accuracy of upsampling, which is regressed using the up-sampled attention map $P_{(0,0)\to (0,v)}^{up}$ and $P_{(0,0)\to (u,0)}^{up}$ . Finally, we train our model in an unsupervised manner, the overall loss includes photometric loss $L_p$ , $L_p^h$ and $L_p^v$ , PAM loss $L_s$ and smoothness loss $L_{sm}$ : + +$$ +\begin{array}{l} L _ {u n s u p} = L _ {p} + \lambda_ {h v} \left(L _ {p} ^ {h} + L _ {p} ^ {v}\right) + \lambda_ {s} L _ {s m} \tag {6} \\ + \lambda_ {P} \left(0. 2 L _ {s} ^ {1} + 0. 3 L _ {s} ^ {2} + 0. 5 L _ {s} ^ {3}\right), \\ \end{array} +$$ + +where $L_{s}^{i}$ extended by the PAM loss [25] and is calculated on the all parallax attention maps obtained in the $i$ -th decoding stage. The $L_{sm}$ [23] is calculated on $D_{(0,v)}$ , $D_{(u,0)}$ and $D_{c}$ to encourage local smoothness of the disparity. + +# 5. Experiment + +# 5.1. Implementation Details + +Our model is implemented using the PyTorch framework and trained on an NVidia RTX A5000 GPU. During training, we crop views into $100 \times 100$ patches as the model input. We train our model by minimizing the $L_{unsup}$ and $\lambda_{hv}, \lambda_s, \lambda_P$ in Eq. 6 are set to 0.5, 0.01, 1. The models are optimized using the Adam optimizer [2] with $\beta_1 = 0.9$ and $\beta_2 = 0.999$ . The initial learning rate is set to $1e^{-3}$ and the batch size is set to 2. Besides, we randomly augment the datasets by flipping the images horizontally, vertically, or rotating 90 degrees. Typically, it roughly takes 2 days of training for our model to converge. + +# 5.2. Comparisons with State-of-the-Art Methods + +We conduct a comparative analysis with SOTA methods, including two non-learning-based methods (i.e., SPO [30], OAVC [6]), four supervised learning-based meth + +Table 2. Quantitative comparison of the MSE*100/Bp(0.07) metrics on HCI dataset. The best results among all methods are in red. The best and second best results among unsupervised methods are in bold and underlined, respectively. + +
MethodBackgammonDotsPyramidsStripsBoxesCottonDinoSideboardBedroomBicycleHerbsOrigamiAvg
MSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/Bp
NonSPO [30]4.59/3.785.24/16.270.04/0.866.96/14.999.11/15.891.31/2.590.31/2.181.02/9.300.21/4.865.57/10.9111.23/8.262.03/11.693.97/8.47
OAVC [6]3.84/3.1216.58/69.110.04/0.831.32/2.906.99/16.140.60/2.550.27/3.941.05/12.420.21/4.924.89/12.2210.36/8.731.48/12.563.97/12.45
SupEPINet [21]3.71/3.501.48/2.490.007/0.160.93/2.465.97/12.340.20/0.450.16/1.210.80/4.460.20/2.304.60/9.619.29/10.961.83/5.812.43/4.65
LFattNet [24]3.65/3.131.43/1.430.004/0.200.89/2.934.00/2.790.21/9.510.09/5.220.53/4.820.37/2.793.35/2.796.61/5.221.73/4.821.91/3.80
AttMLFNet [1]3.86/3.231.04/1.610.003/0.170.81/2.933.84/11.140.06/0.200.04/0.440.40/2.690.13/2.073.80/8.846.37/5.430.99/4.411.78/3.60
OACCNet [26]3.94/3.931.42/1.510.004/0.160.85/2.922.89/10.700.16/0.310.08/0.970.54/3.350.15/2.302.91/8.086.56/6.520.88/4.071.69/3.74
UnsupUnMonocular [32]11.83/12.312.54/3.650.027/0.262.68/11.149.27/22.882.91/3.811.03/5.402.77/10.950.42/7.419.23/20.1026.82/13.443.68/10.956.10/10.19
UnOcc [10]6.68/14.376.57/45.340.213/7.355.20/41.367.45/26.240.80/8.460.63/8.251.79/14.200.39/12.696.23/21.6513.94/16.961.92/19.824.32/19.72
UnLFDisp [33]6.86/5.201.44/1.530.016/0.271.80/5.128.32/20.182.04/3.040.78/4.171.85/7.680.35/7.367.49/16.5519.43/11.802.94/9.674.44/7.71
UnGAN [28]5.38/10.481.74/2.040.013/0.302.39/9.048.01/19.870.76/1.940.74/4.713.39/7.080.55/6.917.96/18.9215.58/12.763.61/10.334.18/8.70
ECAAN10.18/8.063.03/4.770.021/0.351.86/29.847.87/18.540.67/1.920.63/4.921.35/9.050.27/6.407.21/17.1214.81/10.811.99/11.044.16/10.24
+ +![](images/3085ee2460d96a0f579e6afe2a712642525d767f5f884e2b6ec26564e9914e15.jpg) +Figure 5. Visual comparison of different methods on real-world scenes 'Lego Bulldozer' and 'Lego Knights'. The input $3 \times 3$ views are selected (label in the above line) with different intervals to make the disparity ranges increase gradually. + +ods (i.e., EPINet [21], LFAntNet [24], AttMLFNet [1], OACCNet [26]), and seven unsupervised learning-based methods (i.e., UnCNN [16], UnMonocular [32], UnCon [14], UnLFDisp [33], UnGAN [28], UnOcc [10], UnMVFM [31]). The Bad pixel (Bp) ratio [9], which measures the percentage of incorrectly estimated pixels $>t$ , and the Mean Squared Error (MSE) are used as metrics. To comprehensively evaluate model performance, we conduct experiments on both large and small disparity datasets. + +# 5.2.1. Large Disparity Scenes + +Synthetic Datasets. Following most approaches [31], synthetic LFs in SLFD are selected for training and testing, with interview disparities in [-20, 20]. Table 1 illustrates the quantitative results of different methods on several scenes from SLFD. Due to the memory burden of constructing the cost-volume, these cost-volume based methods use the central $3 \times 3$ views as input. To ensure fairness of the experiments, we respectively trained our models for $3 \times 3$ and $5 \times 5$ angular resolutions, labeled as $ECAAN_{3}$ and $ECAAN_{5}$ . Among the LF image methods, our method achieves the lowest MSE and Bp values across all scenes, indicating its ability to better capture the consistency relationship of large disparity scenes and produce superior disparity maps. + +Specifically, $ECAAN_{3}$ achieves a significant MSE reduction of up to 5.427 (Rooster clock) and 0.273 (Lion), while also lowering $\mathrm{Bp}(0.3)$ by up to 7.628 (Lion) and 4.025 (Toy bricks) over the UnMVFM. The visual results are shown in Fig. 4, with a different color map from the LFs for distinction. Other approaches tend to generate noisy predictions and exhibit significant accuracy degradation. In contrast, our results have better visual qualities with smaller errors compared to others, especially in the edge regions. + +Real Datasets. To assess the model's capability to process real-world scenes, we test the models on the LFs captured by the moving camera [22]. Since GT is not available, all models are trained in SLFD. For each scene, we select different $3 \times 3$ views from the $17 \times 17$ LFs with various disparity ranges to evaluate the robustness of the methods. As shown in Fig. 5, the disparity of LF is gradually larger when choosing views at different angular positions. Our method consistently achieves superior accuracy across inputs with varying disparities, preserving sharp boundaries and intricate structural details. Due to the predetermined fixed sampling intervals in cost volume construction, these models are prone to pronounced stratification artifacts. This issue is particularly evident in regions with smoothly varying disparities or occlusions, resulting in abrupt depth discontinu + +Table 3. Result of variant models. + +
ModuleLossMSE / Bp (0.3)
ECSUCOFALpLph/vLionRoosterToyElectro
(1)××all✓4.96/21.260.37/2.933.81/33.052.34/20.62
(2)Linear×all✓2.48/14.690.30/2.741.84/24.021.88/19.27
(3)Dconv×all✓2.06/6.790.25/2.461.14/13.421.51/18.66
(4)Dconv✓××0.26/2.170.17/1.020.83/3.990.52/9.91
(5)Dconv✓sub✓0.16/1.460.14/0.970.59/3.420.42/8.41
(6)Dconv✓all✓0.08/1.130.10/0.810.41/3.200.21/5.63
+ +![](images/94b469f546260e19b3c9dbe09474bb97284c2c3e9f72ae8209b3f65ba1988ba5.jpg) +Figure 6. Visual comparison on real-world small disparity LFs. + +ities and unnatural "layering" effects. + +# 5.2.2. Small Disparity Scenes + +Synthetic Datasets. Regarding the synthetic LFs with small disparity, we use LFs from the HCI benchmark [9, 27] to verify the performance of the models on the synthetic datasets, in which the LFs have a disparity range of $[-4, 4]$ . The results obtained using LFs with an angular resolution of $7 \times 7$ are illustrated in Table 2. Due to the lack of ground-truth supervision, the performance of our model and nonlearning models is worse than the supervised learning methods [1, 21, 24, 26] in some scenes. Compared with other unsupervised methods, our approach achieves the lowest MSE metric and outperforms others in the Bp metric across most scenes. More comparisons on HCI benchmark are provided in the supplementary material. + +Real Datasets. Following [10], the images from [11, 18, 22] are used for real-world dense LF comparison. All unsupervised methods are retrained on the 100 images from [11]. We show the results in real-world LFs [11, 18, 22] in Fig. 6. Due to the difference between the synthetic dataset and the real-world scene, the performance of the supervised models degrades in the real-world scene. Compared with other models, our model is able to better preserve fine structure in real scenes with complex occlusions. + +![](images/6ffd11fec07cd8a4fcac59f47c80bbf76953d1b3edfa3d7415473f0b3c045883.jpg) +Figure 7. Visual comparison of the variant models. + +# 5.3. Ablation Investigation + +In this subsection, we conduct ablation experiments on SLFD using the $ECAAN_{5}$ . The results are shown in Table 3 and Fig. 7. We first design the following models to verify the effectiveness of our proposed epipolar consistent up-sample strategy: Model-1 directly regress all the parallax attention to disparity and average them as the final disparity map. Model-2 replaces the deconvolution layer using a simple linear layer to upsample the attention maps to the same scale. Model-2 performs better than Model-1, suggesting that integrating information in the attention level keeps more information than in the disparity level. Compared Model-2 with Model-3, our designed upsample strategy makes it easier to consistently map the attention to the same scale. We further verify the effectiveness of the proposed occlusion-aware aggregation. Model-3 removes the COFA while generating the mask for the loss. Compared Model-3 with the final model Model-6, the occlusion prior information reduces the weights of occluded regions and better uses the information extracted from the occlusion-free view. We further remove the occlusion mask on the loss, respectively. In Model-4 we remove all the occlusion masks in loss $L_{unsup}$ and in Model-5 we only apply subocclusion masks $O_{(u,0)}$ and $O_{(0,v)}$ . The results indicate that using the designed occlusion masks, the photometric supervision is more accurate and the performances are improved. More results are presented in the supplementary material. + +# 6. Conclusion + +In this paper, we propose a LF disparity estimation network based on the proposed epipolar consistent parallax attention. First, we formulate the horizontal/vertical EPI parallax attention calculation mechanism and analyze the occlusion and scale issues in the LF process. Then, we apply the EPI parallax attention to the disparity estimation task. Specifically, we design an epipolar consistent attention aggregation mechanism to unify the offset scale and aggregate the consistency parallax information of the non-occluded areas. Experimental results show that our method performs favorably against state-of-the-art methods. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032). + +# References + +[1] Jiaxin Chen, Shuo Zhang, and Youfang Lin. Attention-based multi-level fusion network for light field depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1009-1017, 2021. 2, 7, 8 +[2] Yangling Chen, Shuo Zhang, Song Chang, and Youfang Lin. Light field reconstruction using efficient pseudo 4d epipolar-aware structure. IEEE Transactions on Computational Imaging, 8:397-410, 2022. 6 +[3] Mantang Guo, Jing Jin, Hui Liu, and Junhui Hou. Learning dynamic interpolation for extremely sparse light fields with wide baselines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2450-2459, 2021. 2 +[4] Yulan Guo, Mohammed Bennamoun, Ferdous Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11): 2270-2287, 2014. 1 +[5] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1 +[6] Kang Han, Wei Xiang, Eric Wang, and Tao Huang. A novel occlusion-aware vote cost for light field depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8022-8035, 2021. 6, 7 +[7] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6):1397-1409, 2013. 5 +[8] Stefan Heber, Wei Yu, and Thomas Pock. Neural epi-volume networks for shape from light field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2260, 2017. 1 +[9] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision, pages 19-34. Springer, 2016. 7, 8 +[10] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 1, 2, 6, 7, 8 +[11] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM Transactions on Graphics, 35(6):1-10, 2016. 8 +[12] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the annual conference on Computer Graphics and Interactive Techniques. ACM, 1996. 2 +[13] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2023. 1 +[14] Lili Lin, Qiujiang Li, Bin Gao, Yuxiang Yan, Wenhui Zhou, and Ercan Engin Kuruoglu. Unsupervised learning of light field depth estimation with spatial and angular consistencies. Neurocomput., 501(C):113-122, 2022. 6, 7 + +[15] In Kyu Park, Kyoung Mu Lee, et al. Robust light field depth estimation using occlusion-noise aware data costs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2484-2497, 2017. 1 +[16] Jiayong Peng, Zhiwei Xiong, Dong Liu, and Xuejin Chen. Unsupervised depth estimation from light field using a convolutional neural network. In Proceedings of the IEEE International Conference on 3D Vision, pages 295-303. IEEE, 2018. 1, 6, 7 +[17] Jiayong Peng, Zhiwei Xiong, Yicheng Wang, Yueyi Zhang, and Dong Liu. Zero-shot depth estimation from light field using a convolutional neural network. IEEE Transactions on Computational Imaging, 6:682-696, 2020. 1 +[18] Martin Rerabek and Touradj Ebrahimi. New light field image dataset. In Proceedings of the International Conference on Quality of Multimedia Experience, 2016. 8 +[19] Jinglei Shi, Xiaoran Jiang, and Christine Guillemot. A framework for learning depth from a flexible subset of dense and sparse light field views. IEEE Transactions on Image Processing, 28(12):5867-5880, 2019. 2 +[20] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 5 +[21] Changha Shin, Hae-Gon Jeon, Youngjin Yoon, In So Kweon, and Seon Joo Kim. Epinet: A fully-convolutional neural network using epipolar geometry for depth from light field images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4748–4757, 2018. 6, 7, 8 +[22] Abhilash Sunder Raj, Michael Lowney, Raj Shah, and Gordon Wetzstein. The stanford litro light field archive. http://lightfields.stanford.edu/LF2016.html, 2016. Accessed: Aug. 22, 2020. 7, 8 +[23] C. Tomasi and R. Manduchi. Bilateral filtering for gray and color images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 839-846, 1998. 6 +[24] Yu-Ju Tsai, Yu-Lun Liu, Ming Ouhyoung, and Yung-Yu Chuang. Attention-based view selection networks for lightfield disparity estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12095-12103, 2020. 1, 2, 6, 7, 8 +[25] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2108-2125, 2022. 1, 2, 6 +[26] Yingqian Wang, Longguang Wang, Zhengyu Liang, Jungang Yang, Wei An, and Yulan Guo. Occlusion-aware cost constructor for light field depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19809-19818, 2022. 7, 8 +[27] Sven Wanner, Stephan Meister, and Bastian Goldluecke. Datasets and benchmarks for densely sampled 4d light fields. In Proceedings of the Conference on Vision, Modeling, and Visualization, pages 225-226. ACM, 2013. 8 + +[28] Wenbin Yan, Xiaogang Zhang, and Hua Chen. Occlusion-aware unsupervised light field depth estimation based on multi-scale gans. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6318-6333, 2024. 2, 7 +[29] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1983-1992, 2018. 5 +[30] Shuo Zhang, Hao Sheng, Chao Li, Jun Zhang, and Zhang Xiong. Robust depth estimation for light field via spinning parallelogram operator. Computer Vision and Image Understanding, 145:148-159, 2016. 6, 7 +[31] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Unsupervised light field depth estimation via multi-view feature matching with occlusion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2261-2273, 2024. 2, 6, 7 +[32] Wenhui Zhou, Enci Zhou, Gaomin Liu, Lili Lin, and Andrew Lumsdaine. Unsupervised monocular depth estimation from light field image. IEEE Transactions on Image Processing, 29:1606-1617, 2020. 6, 7 +[33] Wenhui Zhou, Lili Lin, Yongjie Hong, Qiujuan Li, Xingfa Shen, and Ercan Engin Kuruoglu. Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2023. 2, 7 \ No newline at end of file diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/images.zip b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ef7060eb2d9d6f2f1ccb8df503974969ffe86e10 --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f18c8e137c2afcbb584b5db02dffc213237df0f9e1deb25a1de773421eb1d36b +size 909182 diff --git a/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/layout.json b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0b27a9dd77986d34b27b422542e2d16cb716febc --- /dev/null +++ b/2025/Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation/layout.json @@ -0,0 +1,8203 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 102, + 522, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 102, + 522, + 140 + ], + "spans": [ + { + "bbox": [ + 89, + 102, + 522, + 140 + ], + "type": "text", + "content": "Epipolar Consistent Attention Aggregation Network for Unsupervised Light Field Disparity Estimation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 91, + 161, + 517, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 161, + 517, + 203 + ], + "spans": [ + { + "bbox": [ + 91, + 161, + 517, + 203 + ], + "type": "text", + "content": "Chen Gao, Shuo Zhang* Youfang Lin \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science & Technology, Beijing Jiaotong University, Beijing, China." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 205, + 408, + 217 + ], + "type": "text", + "content": "{gaochen, zhangshuo, yflin}@bjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 53, + 270, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 270, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 53, + 270, + 297, + 628 + ], + "type": "text", + "content": "Disparity estimation is an essential step in processing and analyzing Light Field (LF) images. Recent methods construct the cost volume to exploit the correspondence of the LFs over the preset maximum disparity, limiting them to process the large parallax scenes. Different from constructing cost volume, the self-attention mechanism calculates the parallax attention between epipolar lines to find the matching points. However, for LFs that have different views, the related disparity scales are different in parallax attention since the baselines with the central view are different. Moreover, if the matching information is occluded in one view, the disparity information can be explored through other views. Therefore, mapping these attentions to the same scale and selecting effective matching information are key points for disparity estimation from parallax attention. In this paper, we explore parallax attention for LF and design an unsupervised method, named Epipolar Consistent Attention Aggregation Network (ECAAN). We first introduce an epipolar consistent scale unification block by considering the consistency relationships to standardize disparity scales of the parallax attention maps. Based on the intra-properties and inter-reationships of parallax attention, we further propose a consistent occlusion-free aggregation block to integrate the information from the occlusion-free areas. In addition, we design an improved photometric loss to constrain the model. ECAAN achieves state-of-the-art performance in LF depth estimation. Notably, ECAAN attains a mean square error of 0.2 on large-disparity LF datasets, achieving a " + }, + { + "bbox": [ + 53, + 270, + 297, + 628 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 53, + 270, + 297, + 628 + ], + "type": "text", + "content": " error reduction compared to the second-best method." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 651, + 135, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 135, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 135, + 662 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 295, + 696 + ], + "type": "text", + "content": "Light Field (LF) disparity estimation aims to infer the 3D scene geometry from the captured LF images, which is ben-" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 322, + 243, + 548, + 358 + ], + "blocks": [ + { + "bbox": [ + 322, + 243, + 548, + 358 + ], + "lines": [ + { + "bbox": [ + 322, + 243, + 548, + 358 + ], + "spans": [ + { + "bbox": [ + 322, + 243, + 548, + 358 + ], + "type": "image", + "image_path": "32f3b947a660a00eb71532d343a7ce24feb3b21b187deeb19a8f48f8c43f7435.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 365, + 555, + 443 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 443 + ], + "type": "text", + "content": "Figure 1. Comparative analysis with state-of-the-art methods on the SLFD dataset. Left: Performance metrics of models on Bp across thresholds (0.3, 0.1, 0.05). Right: Quantitative results for MSE (" + }, + { + "bbox": [ + 313, + 365, + 555, + 443 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 365, + 555, + 443 + ], + "type": "text", + "content": " 100) and Inference Time (seconds). Our method achieves state-of-the-art performance across all metrics while maintaining competitive computational efficiency. Lower values of the Bp and MSE metrics indicate better performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 457, + 555, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 506 + ], + "type": "text", + "content": "eficial for various advanced computer vision applications ranging from 3D reconstruction to robotics and autonomous driving [4, 5]. The key aspect of LF disparity estimation is to identify corresponding pixels across different views." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 507, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 507, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 312, + 507, + 556, + 663 + ], + "type": "text", + "content": "At present, both non-learning methods [15] and learning methods [8, 10, 16, 17, 24] explore geometry information between LF images based on the constructed cost volume. These methods usually construct a cost volume within a predefined disparity range, and compute the weighted sum of those pre-defined disparity candidates to obtain the final disparity prediction. However, such a design suffers from a high computational and memory burden, which limits them in constructing a full-range cost volume for handling large disparities scenes. Moreover, the predefined disparity range inherently constrains the performance of cost-volume-based approaches, making them less effective in handling scenes with extreme parallax variations." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Recently, some methods [13] attempt to use self-attention to explore consistency in stereo matching and show great potential in learning stereo correspondence. PAM [25] explores the pixel movements between views by" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 225, + 713 + ], + "type": "text", + "content": "*Corresponding author: zhangshuo@bjtu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6488" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 178 + ], + "type": "text", + "content": "finding the most similar features in parallax attention maps. Unlike stereo images, LFs have different views. The position offset scales in the attention maps are different since the baselines vary between the arranged views. In addition, if the matching information is occluded in one view, the disparity information of that point can be explored through other views. Therefore, how to map these attention maps to same scale and select information from the occlusion-free areas are key challenges that we should focus on." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 180, + 298, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 180, + 298, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 298, + 431 + ], + "type": "text", + "content": "In this paper, we adapt the parallax attention to LFs and propose a novel unsupervised framework to capture the 3d cues, termed the Epipolar Consistent Attention Aggregation Network (ECAAN). Specifically, we first explicitly calculate the correlation between the reference point and all possible matching points along the epipolar lines to construct the parallax attention maps. To standardize disparity scales of the parallax attention maps, we introduce an Epipolar Consistent Scale Unification (ECSU) block by considering the consistency relationships. Furthermore, we explore the intra-properties and inter-relationships of parallax attention maps and design a Consistent Occlusion-Free Aggregation (COFA) block to filter out occlusions. Finally, the fused attention map is used to regress the disparity map of the scene. In addition, we design an improved photometric loss to constrain the model during training. Our method demonstrates outstanding performance in LF disparity estimation, particularly in large-disparity scenes, achieving more accurate predictions with improved efficiency. For example, ECAAN achieves a " + }, + { + "bbox": [ + 56, + 180, + 298, + 431 + ], + "type": "inline_equation", + "content": "68\\%" + }, + { + "bbox": [ + 56, + 180, + 298, + 431 + ], + "type": "text", + "content": " error reduction compared to the second-best method (see Fig. 1)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 442, + 142, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 713 + ], + "type": "text", + "content": "Most existing learning-based LF disparity estimation methods primarily focus on addressing two challenges: occlusions and large disparities. To mitigate matching errors caused by occlusions, various methods integrate view selection modules or specialized loss functions. Tsai et al. [24] explored matching costs and introduced an SAI selection module to generate attention maps, extracting meaningful information from the available views. Chen et al. [1] proposed a four-branch architecture for matching cost learning, incorporating two attention mechanisms to further refine the information by prioritizing branches with minimal occlusion and richer textures. Guo et al. [3] introduced an occlusion detection network, trained using a designed occlusion map, to process occluded and non-occluded regions separately, ensuring accurate disparity estimation with sharp edges. In recent years, Jin et al. [10] partitioned the entire LF into sub-LFs, feeding them into a multi-scale network to predict disparity maps alongside confidence maps based on the constructed cost volume. In addition, they proposed a simple strategy to fuse multiple disparity maps to deal with occlusion. Zhang et al. [31] proposed a two-stage pipeline" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "to focus on the occlusion problem, in which they generate multiple occlusion priors and act on the loss functions. Yan et al. [28] used the generative adversarial network to estimate the disparity map in an unsupervised manner. Zhou et al. [33] introduced a geometry-based occlusion model along with two loss functions to learn occlusion- and noise-invariant representations of the LF." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 157, + 555, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 555, + 372 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 555, + 372 + ], + "type": "text", + "content": "Effectively managing large disparities in LFs is another open problem. Shi et al. [19] addressed this issue by decomposing LFs into multiple stereo-image pairs, employing a fine-tuned stereo estimation network to compute the disparity for each pair, which was subsequently fused into refined disparity maps. Some approaches [31] tackle large parallax scenes by reducing the angular resolution of the input. However, existing methods commonly generate cost volumes, leading to significant memory overhead, which constrains their ability to perform fine-grained disparity sampling and ultimately limits their performance. Given the advantages of global interaction, the self-attention mechanism in transformers facilitates pixel-wise similarity modeling, making them highly effective in stereo-matching [25]. However, their application in LF disparity estimation remains limited due to the relatively complex structure of LFs, which poses challenges in effectively integrating self-attention mechanisms within LF representations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 381, + 473, + 396 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 473, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 473, + 396 + ], + "type": "text", + "content": "3. Epipolar Parallax Attention" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "content": "In this paper, the LFs are represented as " + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(u,v,x,y)\\in" + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{U\\times V\\times H\\times W}" + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "content": " represent angular coordinates, and " + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "content": " represent spatial coordinates [12]. By fixing angular coordinates, the 2D view " + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "I_{(u,v)}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "text", + "content": " is obtained. The central view are represented as " + }, + { + "bbox": [ + 313, + 402, + 555, + 464 + ], + "type": "inline_equation", + "content": "I_{(0,0)}" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 502, + 480 + ], + "type": "text", + "content": "3.1. EPI Parallax Attention Calculation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "Inspired by the PAM model [25], the self-attention mechanism explicitly computes the correlation between points in paired views. Unlike binocular vision, LFs exhibit correspondences along epipolar lines in multiple directions simultaneously. To better capture these relationships, we compute parallax attention maps on EPIs in various directions, rather than relying on pairwise view similarity as in [25]. Specifically, we design a horizontal EPI parallax attention calculation mechanism (H-EPA) and a vertical EPI parallax attention calculation mechanism (V-EPA) to capture the correlation between views. To introduce the H-EPA, we take a horizontal EPI feature " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h\\in \\mathbb{R}^{V\\times W\\times C}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " that corresponding to the central view as an example. Firstly, 2D EPI convolutional layers are applied to the " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " to realise the interaction of redundant information in EPIs. Then, we use the " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " convolutions to get the query features " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q_{h}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " and the key features " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "K_{h}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " which are further flattened to a sequence of tokens with a length of " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "VW" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q_{h},V_{h}\\in \\mathbb{R}^{VW\\times C}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": "). " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "Q_{h}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " is multiplied with " + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "inline_equation", + "content": "K_{h}" + }, + { + "bbox": [ + 313, + 486, + 555, + 713 + ], + "type": "text", + "content": " to obtain the hori" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6489" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 550, + 281 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 550, + 281 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 550, + 281 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 550, + 281 + ], + "type": "image", + "image_path": "a47acb8ff30d05ebb88177ece180b5269e22890ae905831d5e62332a133cfa12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 287, + 555, + 321 + ], + "lines": [ + { + "bbox": [ + 54, + 287, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 54, + 287, + 555, + 321 + ], + "type": "text", + "content": "Figure 2. One typical example with the related parallax attention maps. The point 1 is visible in all views and the attention value achieves an apparent bright spot in one specific point in each pair view. The point 2, 3 are ambiguous in some views. These features are aggregated combining the occlusion masks to obtain accurate parallax attention." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": "zontal epipolar parallax attention map " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "C_h \\in \\mathbb{R}^{VW \\times VW}" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "C_h = Q_h \\cdot K_h^T" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": ". Finally, " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "V \\times V" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": " parallax-attention maps " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "C_{(0,v_1) \\to (0,v_2)} \\in \\mathbb{R}^{W \\times W}" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": ", which records the similarity between any two positions along the horizontal epipolar lines. The process of V-EPA is the same as the H-EPA while dealing with the vertical EPI features " + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_v \\in \\mathbb{R}^{U \\times H \\times C}" + }, + { + "bbox": [ + 54, + 329, + 294, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 421, + 234, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 234, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 234, + 434 + ], + "type": "text", + "content": "3.2. Parallax Attention Maps Analyse" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": "As shown in Fig. 2, we visualized the " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": " horizontal parallax attention maps " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "C_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times W}" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": ". These maps are calculated by evaluating the similarity between each point in the central view " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "I_{(0,0)}" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": " and every point along the corresponding epipolar line in the view " + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "inline_equation", + "content": "I_{(0,v)}" + }, + { + "bbox": [ + 54, + 437, + 295, + 569 + ], + "type": "text", + "content": ", where brighter color means it is similar to the corresponding point in the central view. The disparity information is encoded through positional offsets of the most similar points relative to the diagonal line, which vary across different views. To effectively aggregate these parallax-attention maps, scale variations and occlusion remain key challenges." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": "Scale Issue: Since the disparity between views varies significantly for different baselines, the position offset scales of these parallax attention maps are different. In Fig. 2, for views " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{(0, -3)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{(0,3)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " that are far from the central view, the position offset is " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "3 \\times" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " larger than the adjacent view " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{(0, -1)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "I_{(0,1)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " due to the larger baseline. In the parallax attention maps, it is clear that the offsets relative to the diagonal line are larger in " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{(0,0) \\to (0, -3)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "inline_equation", + "content": "M_{(0,0) \\to (0,3)}" + }, + { + "bbox": [ + 54, + 570, + 295, + 714 + ], + "type": "text", + "content": ". The offsets of the left view and right view are also symmetric around the diagonal line since the view movement direction is opposite. Coarsely resizing the scale of these parallax attention maps may break the encoded consistency" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 330, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 449 + ], + "type": "text", + "content": "information and introduce significant errors. To aggregate these attention maps, we propose an epipolar consistent re-sample and up-sample strategy (see Sec.4.2). We first crop and re-sample the attention map " + }, + { + "bbox": [ + 313, + 330, + 555, + 449 + ], + "type": "inline_equation", + "content": "M_{(0,0) \\to (0,v)}" + }, + { + "bbox": [ + 313, + 330, + 555, + 449 + ], + "type": "text", + "content": " according to the disparity range and form a new attention map. Then, the re-sampled attention maps are up-sampled heuristically to the same scale by placing their original attention values in the corresponding positions. As in Fig. 2, all up-sampled attention maps are mapped to the same scale and are consistent with the initial parallax information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": "Occlusion Issue: Another fundamental challenge is occlusion, where view consistency does not hold. In LFs, the occlusions may arise from various directions, resulting in different occlusion relationships for each map. Here we take " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "M_{(0,0)\\to (0,v)}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " as a typical example in Fig. 2. If one point is visible in all views (as point 1), the attention value achieves an apparent bright spot at one specific point in each pair view, which is just its matching point. However, it is difficult to find the matching point for points that are occluded in part of views. For example, as the view moves from left to right, the point 2 is visible in the beginning but is then occluded by the foreground pencil. Therefore, the matching point with a bright spot is easy to find in view " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,-3/-2/-1/1)}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " but is ambiguous in view " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,2/3)}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ". Similarly, the background point 3 is occluded by the foreground pencil in the left views but is visible in the right views. The corresponding matching point is concentrated in view " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,-1/1/2/3)}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": " but is dispersed in view " + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,-3/-2)}" + }, + { + "bbox": [ + 313, + 450, + 556, + 715 + ], + "type": "text", + "content": ". For these kinds of occluded points, labeled with blue and yellow boxes, directly fusing all the matching values or alignment information in all views together will bring errors. To alleviate the influence of the occluded views in disparity" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6490" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 526, + 345 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 526, + 345 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 526, + 345 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 526, + 345 + ], + "type": "image", + "image_path": "c1aef4ad36e2fa5160877a654ea38e7d2b240e1bf5db45fe9b54156ef93a24a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 345, + 555, + 379 + ], + "lines": [ + { + "bbox": [ + 54, + 345, + 555, + 379 + ], + "spans": [ + { + "bbox": [ + 54, + 345, + 555, + 379 + ], + "type": "text", + "content": "Figure 3. An overview of our disparity estimation network. After feature extraction, the epipolar parallax attention maps are constructed by the H-EPA and V-EPA. The epipolar parallax attention maps are aggregated based on the epipolar consistent attention aggregation. Finally, the disparity map is generated through the disparity regression." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 297, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 297, + 544 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 297, + 544 + ], + "type": "text", + "content": "estimation, we detect the occluded regions and aggregate the disparity information encoded in attention maps from non-occluded regions (see Sec. 4.3). Specifically, we calculate the occlusion masks for each view to label whether the point is occluded in the paired view. As shown, the two regions with blue boxes that are visible in the central view but occluded by the pencil in view " + }, + { + "bbox": [ + 54, + 388, + 297, + 544 + ], + "type": "inline_equation", + "content": "I_{(0, -3)}" + }, + { + "bbox": [ + 54, + 388, + 297, + 544 + ], + "type": "text", + "content": " are accurately masked in the occlusion masks. Combining with the occlusion mask, we calculate the occlusion-aware aggregation weights for different views as the last row in Fig. 2. The fused attention map is also visualized, in which we can find a matching point with high similarity for each pixel by using the information of non-occluded viewpoints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 556, + 291, + 571 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 291, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 291, + 571 + ], + "type": "text", + "content": "4. Epipolar Consistent Attention Aggregation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "text", + "content": "The proposed Epipolar Consistent Attention Aggregation Network aims to estimate the disparity map of the central view " + }, + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "inline_equation", + "content": "D_{c}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "text", + "content": " from the LFs " + }, + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 578, + 296, + 613 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 123, + 635 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "The overall network structure is shown in Fig. 3. Given the input image " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", we first implement spatial convolution layers to encode the multi-scale features " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e^i\\in \\mathbb{R}^{U\\times V\\times H_i\\times W_i\\times C}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": " in which " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "H_{i} = \\frac{H}{2^{i}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "inline_equation", + "content": "W_{i} = \\frac{W}{2^{i}}" + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": ". Then, these encoded features are processed through a hierarchical decoder structure that deploys with the H-EPA and V-EPA(Sec. 3.1) to ex" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": "plore the epipolar consistency. At each decoding stage " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": ", the features produced in the previous stage " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_d^{i - 1}" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " are up-sampled and fused with the " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_e^i" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " through spatial convolutions, producing " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{fuse}^{i}" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": ". The " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{fuse}^{i}" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " are then processed in parallel by H-EPA and V-EPA to obtain epipolar parallax attention maps " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "C_h^i /C_v^i" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " along with refined features " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h^i /\\mathcal{F}_v^i" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": ". These refined features " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_h^i /\\mathcal{F}_v^i" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " are subsequently combined and fed into the next decoder stage. Once all hierarchical epipolar parallax attention maps " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "\\{C_h^i,C_c^i\\}_{i = 1}^N" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " are obtained, they are up-sampled and summed to form the final attention maps " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "C_v" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": ". To estimate depth maps of central views, the attention maps calculated from central views with other views are aggregated by the epipolar consistent attention aggregation (ECAA) mechanism, which consists of ECSU block (Sec. 4.2) and COFA block (Sec. 4.3). Finally, the disparity is estimated as the weighted sum of disparity candidates based on the aggregated epipolar consistent attention " + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "inline_equation", + "content": "P_{att}" + }, + { + "bbox": [ + 313, + 388, + 556, + 594 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 368, + 601, + 555, + 633 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 601, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 368, + 601, + 555, + 633 + ], + "type": "interline_equation", + "content": "D _ {c} = \\sum_ {k = 0} ^ {D _ {u p}} k \\times P _ {a t t} (:,:, k) - \\frac {D _ {u p}}{2}. \\tag {1}", + "image_path": "3b8b83511f0aedd71ce7e1b0d9d5e69eb43370eae638ef7cff9ca73094589952.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "inline_equation", + "content": "D_{up}" + }, + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "text", + "content": " means the size of disparity candidates in " + }, + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "inline_equation", + "content": "P_{att}" + }, + { + "bbox": [ + 313, + 639, + 547, + 653 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 512, + 673 + ], + "type": "text", + "content": "4.2. Epipolar Consistent Scale Unification" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "To better aggregate the calculated attention maps from different views, we re-sample and up-sample these maps to a uniform scale under the constraint of epipolar consistency." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 758 + ], + "type": "text", + "content": "6491" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": "Epipolar Consistent Attention Re-Sample. For each " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "V - 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " view that moves horizontally from the central view, there are " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " attention maps " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "M_h" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " available that calculate the pairwise similarity of these views. We add the disparity constraint to " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "C_h" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " and get the constrained epipolar parallax attention maps " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "M_h" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": ". We re-sample each paired attention map " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "M_{(0,0)\\to (0,v)}" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " according to the disparity range " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "[d_{min},d_{max}]" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": " and form a new map " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "P_{(0,0)\\to (0,v)}\\in \\mathbb{R}^{W\\times (|\\mathrm{v}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": ". Similarly, the attention maps related to the vertical moved views are also re-sampled as " + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "inline_equation", + "content": "P_{(0,0)\\to (u,0)}\\in \\mathbb{R}^{H\\times (|\\mathrm{u}|*(\\mathrm{d}_{\\mathrm{max}} - \\mathrm{d}_{\\mathrm{min}}) + 1)}" + }, + { + "bbox": [ + 55, + 72, + 294, + 205 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "text", + "content": "Epipolar Consistent Attention Up-Sample. For attention maps " + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "inline_equation", + "content": "P_{(0,0) \\to (0,v) / (u,0)}" + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "text", + "content": " that have different kinds of scale, we propose to use deconvolution layers with different parameters to provide specific location information for similarity score interpolation. Taking " + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "inline_equation", + "content": "P_{(0,0) \\to (0,v)}" + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "text", + "content": " as an example, the magnification factor " + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 205, + 294, + 277 + ], + "type": "text", + "content": " in the deconvolution layer is:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 133, + 289, + 294, + 300 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 289, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 133, + 289, + 294, + 300 + ], + "type": "interline_equation", + "content": "f _ {(0, 0) \\rightarrow (0, v)} = N / | v |, \\tag {2}", + "image_path": "ae9471f605d0a68deb39e7392d31710133a5e6ba3c9b3b79e175123b3449e5b2.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": " indicates the least common multiple of " + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "\\{1,2,\\dots ,\\frac{V}{2}\\}" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": ". In such a way, the original similarity scores in the small disparity scale space are put into the corresponding positions in the large disparity scale space as in Fig. 3, resulting in up-sampled " + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "P_{(0,0)\\to (0,v)}^{up}\\in \\mathbb{R}^{W\\times D_{up}}" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "(D_{up} = N\\times (d_{max} - d_{min}) + 1)" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": ". Therefore, compared with other up-sampled methods such as pixel shuffle [20], the deconvolution layer with " + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": " not only maps all attention to the same scale but also keeps the original matching information. Note that we manually flip half of the up-sampled parallax attentions since the pixel offsets are opposite for views that in different sides of the central view. We then combine all the processed consistent parallax attentions from the horizontal and vertical views " + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "inline_equation", + "content": "P_{(0,0)}^{up}\\in \\mathbb{R}^{(U + V - 2)\\times H\\times W\\times D_{up}}" + }, + { + "bbox": [ + 55, + 303, + 295, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 480, + 264, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 264, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 264, + 493 + ], + "type": "text", + "content": "4.3. Consistent Occlusion-Free Aggregation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 294, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 294, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 294, + 558 + ], + "type": "text", + "content": "To effectively filter out occlusion-area information, we analyze both the intra-properties of individual parallax attention maps and the inter-relationships between them, leveraging this analysis to generate occlusion-aware fusion weights for consistent information aggregation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": "Occlusion Masks. The intra-properties of individual parallax attention maps reveal that when a point is occluded in one view, its similarity remains relatively low as analyzed in Sec. 3.2. Based on this, we perform occlusion detection on parallax-attention map " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{h}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{v}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " and generate occlusion prior knowledge " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "O_{(u,0)} \\in \\mathbb{R}^{H \\times W}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " of view " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,v)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{(u,0)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " with respect to " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "I_{(0,0)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": ". Specifically, we add parallax attention maps " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{(0,v) \\to (0,0)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "M_{(u,0) \\to (0,0)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " along the non-normalized dimension and set an appropriate threshold " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " (set to 0.9 in this paper) for binarization. Since occlusions usually appear at object edges, we utilize guided filter [7] to sharpen the edges before binarization. Finally, we obtain " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "(V - 1)" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " binary mask " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "(U - 1)" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": " binary mask " + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 55, + 558, + 295, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": "Consistency Weight. Given that disparity constitutes an intrinsic property of visual scenes, the inter-relationships between the up-sampled attention maps should inherently maintain consistency. Therefore, we focus the model on aggregating the attention maps that exhibit consistent disparity information. To implement this, we feed " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "P_{(0,0)}^{up}" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": " into the transformer model along with a linear layer to calculate the consistency weights of different attention maps. Specifically, the " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "P_{(0,0)}^{up}" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": " is reshaped to " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": " sequence of tokens " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "T \\in \\mathbb{R}^{(U + V - 2) \\times D_{up}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": ". Afterward, " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": " are multiplied separately by " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "W_{Q}, W_{K}, W_{V} \\in \\mathbb{R}^{D_{up} \\times c}" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": " to generate the corresponding query, key, and value. Then, the dot-product operation is used on the query and key, followed by a softmax function to obtain the attention scores on the tokens. The attention scores do matrix multiplication with value to get the fusion attention feature and the fusion attention feature is fed into a linear layer to get the initial fusion weight " + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "inline_equation", + "content": "W_{init}" + }, + { + "bbox": [ + 313, + 72, + 553, + 279 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": "Occlusion-Free Aggregation. Since occlusions often cause errors in the final regression results, we consider assigning small weights to the occluded parts when aggregating the attention maps. Specifically, we introduce the occlusion prior information " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " to suppress the information provided by the occluded views. First, we map occluded points in " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "-\\infty" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " and non-occluded point to 0. The mapping results from " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": " are concatenated and added to " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "W_{init}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": ". We then fed the added weight into a softmax layer to produce the occlusion-aware fusion weight " + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "inline_equation", + "content": "W_{fuse} \\in \\mathbb{R}^{(U + V - 2)\\times H\\times W}" + }, + { + "bbox": [ + 313, + 279, + 553, + 410 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "text", + "content": "The up-sampled " + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "inline_equation", + "content": "P_{(0,0)}^{up}" + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "text", + "content": " are then multiplied by the broadcasted " + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "inline_equation", + "content": "W_{fuse}" + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "text", + "content": " and sum over the perspective dimension to obtain the fused epipolar consistent attention map " + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "inline_equation", + "content": "P_{att}" + }, + { + "bbox": [ + 313, + 410, + 553, + 446 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 451, + 553, + 480 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 451, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 341, + 451, + 553, + 480 + ], + "type": "interline_equation", + "content": "P _ {a t t} = \\sum_ {k = 1} ^ {U + V - 2} W _ {f u s e} (k,:,:) \\times P _ {(0, 0)} ^ {u p} (k,:,:) \\tag {3}", + "image_path": "0b288488fb4925d07b79d2b0402b4c6787bbf09f497133947deb32b14a4612c9.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 486, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 510 + ], + "type": "text", + "content": "In this way, we integrate attention maps from different views to increase the robustness of our results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 516, + 404, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 404, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 404, + 528 + ], + "type": "text", + "content": "4.4. Loss Function" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "text", + "content": "Based on the occlusion mask " + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "text", + "content": ", we improve the photometric loss [29] to make the model pay attention to non-occluded regions during training. Specifically, we use the final disparity map " + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 313, + 533, + 553, + 593 + ], + "type": "text", + "content": " to warp all views toward the central view and calculate the L1 loss between them:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 597, + 553, + 626 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 597, + 553, + 626 + ], + "spans": [ + { + "bbox": [ + 337, + 597, + 553, + 626 + ], + "type": "interline_equation", + "content": "L _ {p} = \\sum_ {u} ^ {U} \\sum_ {v} ^ {V} L _ {1} \\left(w \\left(I _ {(u, v)}, D _ {c}\\right), I _ {(0, 0)}\\right) \\times O _ {(u, v)}, \\tag {4}", + "image_path": "51fa4fd300b6361a62a999ecc89f519d8ea2f6c48fb2ebf3df125dba2b3c5582.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "inline_equation", + "content": "w(\\cdot)" + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "text", + "content": " denotes a warping operator. Since we only have " + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 632, + 553, + 679 + ], + "type": "text", + "content": " that label the occluded regions in the cross-shape views in LFs, we propose a simple way to extend them to all views in LFs:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 354, + 684, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 684, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 354, + 684, + 553, + 716 + ], + "type": "interline_equation", + "content": "O _ {(u, v)} = \\left\\{ \\begin{array}{l l} 1, & O _ {(u, 0)} \\wedge O _ {(0, v)} = 1 \\\\ 0, & O _ {(u, 0)} \\vee O _ {(0, v)} = 0 \\end{array} \\right. \\tag {5}", + "image_path": "59cef37cb3b957302165f62da2c37b8a64f0e640e6a57d5011edabbcc85862c6.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6492" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 91, + 549, + 200 + ], + "blocks": [ + { + "bbox": [ + 105, + 71, + 504, + 82 + ], + "lines": [ + { + "bbox": [ + 105, + 71, + 504, + 82 + ], + "spans": [ + { + "bbox": [ + 105, + 71, + 504, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative comparison of MSE*100/Bp(0.3,0,1,0.05) on SLFD dataset. The best results are in bold." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 91, + 549, + 200 + ], + "lines": [ + { + "bbox": [ + 61, + 91, + 549, + 200 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 549, + 200 + ], + "type": "table", + "html": "
MethodsLionRooster clockToy bricksElectro devices
MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)MSEBp(0.3)Bp(0.1)Bp(0.05)
SupEPINet [21]0.47631.45768.13882.9540.74036.86871.56285.1471.05734.65861.95777.4911.11239.80964.45581.793
LFAntNet [24]0.37212.99429.59449.3270.2786.83125.81350.9200.73811.60729.01852.5760.70313.81938.73160.993
UnSupUnCNN [16]1.44258.22184.86792.3739.21334.58675.02187.3832.25453.92282.17290.9502.18141.16372.97085.866
UnMonocular [32]0.76048.30587.37194.6730.66439.93173.60186.3701.22647.59471.09884.9822.73645.61275.23887.332
UnCon [14]1.19235.49375.39287.8910.49116.92853.88975.5971.74150.47179.47089.4743.60442.44375.65487.271
UnOcc [10]1.45453.46080.65090.2120.42111.14842.62768.1511.90661.45784.58791.5702.72040.83471.62284.935
UnMVFM [31]0.3608.76624.42047.9110.2615.79625.30348.6680.77210.04830.42755.7320.84213.63136.03857.880
ECAAN30.1731.5596.86416.4800.1772.58810.89631.1060.3323.74129.62454.4710.2909.09531.76755.710
ECAAN50.0871.1389.470332.9980.1010.8015.57814.3920.4083.20727.06847.2510.2115.63022.05744.819
", + "image_path": "3485465f23cea27db9755e1c19ddad52cb1f82e4eab02ef370e95a855e8301d3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 210, + 539, + 454 + ], + "blocks": [ + { + "bbox": [ + 69, + 210, + 539, + 454 + ], + "lines": [ + { + "bbox": [ + 69, + 210, + 539, + 454 + ], + "spans": [ + { + "bbox": [ + 69, + 210, + 539, + 454 + ], + "type": "image", + "image_path": "0981def47ae9d544ddd90c8dcb29d58eaf26cab1d20dc5186b63f7800e9aedb4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 137, + 457, + 472, + 469 + ], + "lines": [ + { + "bbox": [ + 137, + 457, + 472, + 469 + ], + "spans": [ + { + "bbox": [ + 137, + 457, + 472, + 469 + ], + "type": "text", + "content": "Figure 4. Visual comparison of the disparity maps estimated by different methods on SLFD." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": "If an occluder is labeled in " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ", we consider that this occluder will also appear in the moving view along the vertical direction of " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ". In addition, we also implement the photometric loss on disparity maps " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "D_{(0,v)}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "D_{(u,0)}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_p^h" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_p^v" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " to ensure the accuracy of upsampling, which is regressed using the up-sampled attention map " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{(0,0)\\to (0,v)}^{up}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "P_{(0,0)\\to (u,0)}^{up}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ". Finally, we train our model in an unsupervised manner, the overall loss includes photometric loss " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_p" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_p^h" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_p^v" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ", PAM loss " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_s" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": " and smoothness loss " + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "inline_equation", + "content": "L_{sm}" + }, + { + "bbox": [ + 55, + 484, + 296, + 597 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 89, + 615, + 295, + 646 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 615, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 89, + 615, + 295, + 646 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} L _ {u n s u p} = L _ {p} + \\lambda_ {h v} \\left(L _ {p} ^ {h} + L _ {p} ^ {v}\\right) + \\lambda_ {s} L _ {s m} \\tag {6} \\\\ + \\lambda_ {P} \\left(0. 2 L _ {s} ^ {1} + 0. 3 L _ {s} ^ {2} + 0. 5 L _ {s} ^ {3}\\right), \\\\ \\end{array}", + "image_path": "3526f5e9531f3037cb5234dc4a04e44d52a5d0b03e45f26c86cee378b4635d58.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{s}^{i}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " extended by the PAM loss [25] and is calculated on the all parallax attention maps obtained in the " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "-th decoding stage. The " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "L_{sm}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " [23] is calculated on " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{(0,v)}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{(u,0)}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{c}" + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": " to encourage local smoothness of the disparity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 483, + 391, + 497 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 483, + 391, + 497 + ], + "spans": [ + { + "bbox": [ + 314, + 483, + 391, + 497 + ], + "type": "text", + "content": "5. Experiment" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 448, + 515 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 448, + 515 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 448, + 515 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": "Our model is implemented using the PyTorch framework and trained on an NVidia RTX A5000 GPU. During training, we crop views into " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "100 \\times 100" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": " patches as the model input. We train our model by minimizing the " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "L_{unsup}" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\lambda_{hv}, \\lambda_s, \\lambda_P" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": " in Eq. 6 are set to 0.5, 0.01, 1. The models are optimized using the Adam optimizer [2] with " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\beta_1 = 0.9" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "\\beta_2 = 0.999" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": ". The initial learning rate is set to " + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "inline_equation", + "content": "1e^{-3}" + }, + { + "bbox": [ + 313, + 520, + 555, + 653 + ], + "type": "text", + "content": " and the batch size is set to 2. Besides, we randomly augment the datasets by flipping the images horizontally, vertically, or rotating 90 degrees. Typically, it roughly takes 2 days of training for our model to converge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 544, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 544, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 544, + 673 + ], + "type": "text", + "content": "5.2. Comparisons with State-of-the-Art Methods" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "We conduct a comparative analysis with SOTA methods, including two non-learning-based methods (i.e., SPO [30], OAVC [6]), four supervised learning-based meth" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "6493" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 102, + 558, + 222 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of the MSE*100/Bp(0.07) metrics on HCI dataset. The best results among all methods are in red. The best and second best results among unsupervised methods are in bold and underlined, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 102, + 558, + 222 + ], + "lines": [ + { + "bbox": [ + 58, + 102, + 558, + 222 + ], + "spans": [ + { + "bbox": [ + 58, + 102, + 558, + 222 + ], + "type": "table", + "html": "
MethodBackgammonDotsPyramidsStripsBoxesCottonDinoSideboardBedroomBicycleHerbsOrigamiAvg
MSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/BpMSE/Bp
NonSPO [30]4.59/3.785.24/16.270.04/0.866.96/14.999.11/15.891.31/2.590.31/2.181.02/9.300.21/4.865.57/10.9111.23/8.262.03/11.693.97/8.47
OAVC [6]3.84/3.1216.58/69.110.04/0.831.32/2.906.99/16.140.60/2.550.27/3.941.05/12.420.21/4.924.89/12.2210.36/8.731.48/12.563.97/12.45
SupEPINet [21]3.71/3.501.48/2.490.007/0.160.93/2.465.97/12.340.20/0.450.16/1.210.80/4.460.20/2.304.60/9.619.29/10.961.83/5.812.43/4.65
LFattNet [24]3.65/3.131.43/1.430.004/0.200.89/2.934.00/2.790.21/9.510.09/5.220.53/4.820.37/2.793.35/2.796.61/5.221.73/4.821.91/3.80
AttMLFNet [1]3.86/3.231.04/1.610.003/0.170.81/2.933.84/11.140.06/0.200.04/0.440.40/2.690.13/2.073.80/8.846.37/5.430.99/4.411.78/3.60
OACCNet [26]3.94/3.931.42/1.510.004/0.160.85/2.922.89/10.700.16/0.310.08/0.970.54/3.350.15/2.302.91/8.086.56/6.520.88/4.071.69/3.74
UnsupUnMonocular [32]11.83/12.312.54/3.650.027/0.262.68/11.149.27/22.882.91/3.811.03/5.402.77/10.950.42/7.419.23/20.1026.82/13.443.68/10.956.10/10.19
UnOcc [10]6.68/14.376.57/45.340.213/7.355.20/41.367.45/26.240.80/8.460.63/8.251.79/14.200.39/12.696.23/21.6513.94/16.961.92/19.824.32/19.72
UnLFDisp [33]6.86/5.201.44/1.530.016/0.271.80/5.128.32/20.182.04/3.040.78/4.171.85/7.680.35/7.367.49/16.5519.43/11.802.94/9.674.44/7.71
UnGAN [28]5.38/10.481.74/2.040.013/0.302.39/9.048.01/19.870.76/1.940.74/4.713.39/7.080.55/6.917.96/18.9215.58/12.763.61/10.334.18/8.70
ECAAN10.18/8.063.03/4.770.021/0.351.86/29.847.87/18.540.67/1.920.63/4.921.35/9.050.27/6.407.21/17.1214.81/10.811.99/11.044.16/10.24
", + "image_path": "024eeed3f039cc25e662e8ea2d9c03a72b67136f709466f258c1724c91d3679f.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 69, + 232, + 544, + 381 + ], + "blocks": [ + { + "bbox": [ + 69, + 232, + 544, + 381 + ], + "lines": [ + { + "bbox": [ + 69, + 232, + 544, + 381 + ], + "spans": [ + { + "bbox": [ + 69, + 232, + 544, + 381 + ], + "type": "image", + "image_path": "3085ee2460d96a0f579e6afe2a712642525d767f5f884e2b6ec26564e9914e15.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 390, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 54, + 390, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 555, + 415 + ], + "type": "text", + "content": "Figure 5. Visual comparison of different methods on real-world scenes 'Lego Bulldozer' and 'Lego Knights'. The input " + }, + { + "bbox": [ + 54, + 390, + 555, + 415 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 390, + 555, + 415 + ], + "type": "text", + "content": " views are selected (label in the above line) with different intervals to make the disparity ranges increase gradually." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 426, + 297, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 297, + 534 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 297, + 534 + ], + "type": "text", + "content": "ods (i.e., EPINet [21], LFAntNet [24], AttMLFNet [1], OACCNet [26]), and seven unsupervised learning-based methods (i.e., UnCNN [16], UnMonocular [32], UnCon [14], UnLFDisp [33], UnGAN [28], UnOcc [10], UnMVFM [31]). The Bad pixel (Bp) ratio [9], which measures the percentage of incorrectly estimated pixels " + }, + { + "bbox": [ + 54, + 426, + 297, + 534 + ], + "type": "inline_equation", + "content": ">t" + }, + { + "bbox": [ + 54, + 426, + 297, + 534 + ], + "type": "text", + "content": ", and the Mean Squared Error (MSE) are used as metrics. To comprehensively evaluate model performance, we conduct experiments on both large and small disparity datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 542, + 183, + 554 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 183, + 554 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 183, + 554 + ], + "type": "text", + "content": "5.2.1. Large Disparity Scenes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": "Synthetic Datasets. Following most approaches [31], synthetic LFs in SLFD are selected for training and testing, with interview disparities in [-20, 20]. Table 1 illustrates the quantitative results of different methods on several scenes from SLFD. Due to the memory burden of constructing the cost-volume, these cost-volume based methods use the central " + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": " views as input. To ensure fairness of the experiments, we respectively trained our models for " + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": " angular resolutions, labeled as " + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "inline_equation", + "content": "ECAAN_{3}" + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "inline_equation", + "content": "ECAAN_{5}" + }, + { + "bbox": [ + 54, + 558, + 297, + 715 + ], + "type": "text", + "content": ". Among the LF image methods, our method achieves the lowest MSE and Bp values across all scenes, indicating its ability to better capture the consistency relationship of large disparity scenes and produce superior disparity maps." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "content": "Specifically, " + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "inline_equation", + "content": "ECAAN_{3}" + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "content": " achieves a significant MSE reduction of up to 5.427 (Rooster clock) and 0.273 (Lion), while also lowering " + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\mathrm{Bp}(0.3)" + }, + { + "bbox": [ + 313, + 426, + 555, + 533 + ], + "type": "text", + "content": " by up to 7.628 (Lion) and 4.025 (Toy bricks) over the UnMVFM. The visual results are shown in Fig. 4, with a different color map from the LFs for distinction. Other approaches tend to generate noisy predictions and exhibit significant accuracy degradation. In contrast, our results have better visual qualities with smaller errors compared to others, especially in the edge regions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Real Datasets. To assess the model's capability to process real-world scenes, we test the models on the LFs captured by the moving camera [22]. Since GT is not available, all models are trained in SLFD. For each scene, we select different " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " views from the " + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "inline_equation", + "content": "17 \\times 17" + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": " LFs with various disparity ranges to evaluate the robustness of the methods. As shown in Fig. 5, the disparity of LF is gradually larger when choosing views at different angular positions. Our method consistently achieves superior accuracy across inputs with varying disparities, preserving sharp boundaries and intricate structural details. Due to the predetermined fixed sampling intervals in cost volume construction, these models are prone to pronounced stratification artifacts. This issue is particularly evident in regions with smoothly varying disparities or occlusions, resulting in abrupt depth discontinu" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6494" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 85, + 293, + 182 + ], + "blocks": [ + { + "bbox": [ + 113, + 71, + 237, + 80 + ], + "lines": [ + { + "bbox": [ + 113, + 71, + 237, + 80 + ], + "spans": [ + { + "bbox": [ + 113, + 71, + 237, + 80 + ], + "type": "text", + "content": "Table 3. Result of variant models." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 85, + 293, + 182 + ], + "lines": [ + { + "bbox": [ + 57, + 85, + 293, + 182 + ], + "spans": [ + { + "bbox": [ + 57, + 85, + 293, + 182 + ], + "type": "table", + "html": "
ModuleLossMSE / Bp (0.3)
ECSUCOFALpLph/vLionRoosterToyElectro
(1)××all✓4.96/21.260.37/2.933.81/33.052.34/20.62
(2)Linear×all✓2.48/14.690.30/2.741.84/24.021.88/19.27
(3)Dconv×all✓2.06/6.790.25/2.461.14/13.421.51/18.66
(4)Dconv✓××0.26/2.170.17/1.020.83/3.990.52/9.91
(5)Dconv✓sub✓0.16/1.460.14/0.970.59/3.420.42/8.41
(6)Dconv✓all✓0.08/1.130.10/0.810.41/3.200.21/5.63
", + "image_path": "52eb1ff0a0dc12fad056bdfcc0c68ca44dc6361e193e319f2b411faacf6e4ef0.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 61, + 191, + 289, + 369 + ], + "blocks": [ + { + "bbox": [ + 61, + 191, + 289, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 191, + 289, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 289, + 369 + ], + "type": "image", + "image_path": "94b469f546260e19b3c9dbe09474bb97284c2c3e9f72ae8209b3f65ba1988ba5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 370, + 290, + 382 + ], + "lines": [ + { + "bbox": [ + 59, + 370, + 290, + 382 + ], + "spans": [ + { + "bbox": [ + 59, + 370, + 290, + 382 + ], + "type": "text", + "content": "Figure 6. Visual comparison on real-world small disparity LFs." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 395, + 208, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 395, + 208, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 395, + 208, + 407 + ], + "type": "text", + "content": "ities and unnatural \"layering\" effects." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 181, + 438 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 181, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 181, + 438 + ], + "type": "text", + "content": "5.2.2. Small Disparity Scenes" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "text", + "content": "Synthetic Datasets. Regarding the synthetic LFs with small disparity, we use LFs from the HCI benchmark [9, 27] to verify the performance of the models on the synthetic datasets, in which the LFs have a disparity range of " + }, + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "inline_equation", + "content": "[-4, 4]" + }, + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "text", + "content": ". The results obtained using LFs with an angular resolution of " + }, + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "inline_equation", + "content": "7 \\times 7" + }, + { + "bbox": [ + 55, + 445, + 295, + 601 + ], + "type": "text", + "content": " are illustrated in Table 2. Due to the lack of ground-truth supervision, the performance of our model and nonlearning models is worse than the supervised learning methods [1, 21, 24, 26] in some scenes. Compared with other unsupervised methods, our approach achieves the lowest MSE metric and outperforms others in the Bp metric across most scenes. More comparisons on HCI benchmark are provided in the supplementary material." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Real Datasets. Following [10], the images from [11, 18, 22] are used for real-world dense LF comparison. All unsupervised methods are retrained on the 100 images from [11]. We show the results in real-world LFs [11, 18, 22] in Fig. 6. Due to the difference between the synthetic dataset and the real-world scene, the performance of the supervised models degrades in the real-world scene. Compared with other models, our model is able to better preserve fine structure in real scenes with complex occlusions." + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 553, + 157 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 553, + 157 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 553, + 157 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 553, + 157 + ], + "type": "image", + "image_path": "6ffd11fec07cd8a4fcac59f47c80bbf76953d1b3edfa3d7415473f0b3c045883.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 163, + 526, + 174 + ], + "lines": [ + { + "bbox": [ + 340, + 163, + 526, + 174 + ], + "spans": [ + { + "bbox": [ + 340, + 163, + 526, + 174 + ], + "type": "text", + "content": "Figure 7. Visual comparison of the variant models." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 186, + 441, + 199 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 186, + 441, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 441, + 199 + ], + "type": "text", + "content": "5.3. Ablation Investigation" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "content": "In this subsection, we conduct ablation experiments on SLFD using the " + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "inline_equation", + "content": "ECAAN_{5}" + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "content": ". The results are shown in Table 3 and Fig. 7. We first design the following models to verify the effectiveness of our proposed epipolar consistent up-sample strategy: Model-1 directly regress all the parallax attention to disparity and average them as the final disparity map. Model-2 replaces the deconvolution layer using a simple linear layer to upsample the attention maps to the same scale. Model-2 performs better than Model-1, suggesting that integrating information in the attention level keeps more information than in the disparity level. Compared Model-2 with Model-3, our designed upsample strategy makes it easier to consistently map the attention to the same scale. We further verify the effectiveness of the proposed occlusion-aware aggregation. Model-3 removes the COFA while generating the mask for the loss. Compared Model-3 with the final model Model-6, the occlusion prior information reduces the weights of occluded regions and better uses the information extracted from the occlusion-free view. We further remove the occlusion mask on the loss, respectively. In Model-4 we remove all the occlusion masks in loss " + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "inline_equation", + "content": "L_{unsup}" + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "content": " and in Model-5 we only apply subocclusion masks " + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "inline_equation", + "content": "O_{(u,0)}" + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "inline_equation", + "content": "O_{(0,v)}" + }, + { + "bbox": [ + 313, + 204, + 555, + 516 + ], + "type": "text", + "content": ". The results indicate that using the designed occlusion masks, the photometric supervision is more accurate and the performances are improved. More results are presented in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 546, + 554, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 677 + ], + "type": "text", + "content": "In this paper, we propose a LF disparity estimation network based on the proposed epipolar consistent parallax attention. First, we formulate the horizontal/vertical EPI parallax attention calculation mechanism and analyze the occlusion and scale issues in the LF process. Then, we apply the EPI parallax attention to the disparity estimation task. Specifically, we design an epipolar consistent attention aggregation mechanism to unify the offset scale and aggregate the consistency parallax information of the non-occluded areas. Experimental results show that our method performs favorably against state-of-the-art methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032)." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6495" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Jiaxin Chen, Shuo Zhang, and Youfang Lin. Attention-based multi-level fusion network for light field depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1009-1017, 2021. 2, 7, 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 136, + 296, + 178 + ], + "type": "text", + "content": "[2] Yangling Chen, Shuo Zhang, Song Chang, and Youfang Lin. Light field reconstruction using efficient pseudo 4d epipolar-aware structure. IEEE Transactions on Computational Imaging, 8:397-410, 2022. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 180, + 295, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 180, + 295, + 233 + ], + "spans": [ + { + "bbox": [ + 62, + 180, + 295, + 233 + ], + "type": "text", + "content": "[3] Mantang Guo, Jing Jin, Hui Liu, and Junhui Hou. Learning dynamic interpolation for extremely sparse light fields with wide baselines. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2450-2459, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 62, + 236, + 294, + 289 + ], + "type": "text", + "content": "[4] Yulan Guo, Mohammed Bennamoun, Ferdous Sohel, Min Lu, and Jianwei Wan. 3d object recognition in cluttered scenes with local surface features: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 36(11): 2270-2287, 2014. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 294, + 334 + ], + "type": "text", + "content": "[5] Yulan Guo, Hanyun Wang, Qingyong Hu, Hao Liu, Li Liu, and Mohammed Bennamoun. Deep learning for 3d point clouds: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(12):4338-4364, 2021. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 294, + 378 + ], + "type": "text", + "content": "[6] Kang Han, Wei Xiang, Eric Wang, and Tao Huang. A novel occlusion-aware vote cost for light field depth estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(11):8022-8035, 2021. 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 294, + 412 + ], + "type": "text", + "content": "[7] Kaiming He, Jian Sun, and Xiaou Tang. Guided image filtering. IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6):1397-1409, 2013. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 414, + 294, + 456 + ], + "type": "text", + "content": "[8] Stefan Heber, Wei Yu, and Thomas Pock. Neural epi-volume networks for shape from light field. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2252-2260, 2017. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 294, + 512 + ], + "type": "text", + "content": "[9] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision, pages 19-34. Springer, 2016. 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 514, + 294, + 545 + ], + "type": "text", + "content": "[10] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 590 + ], + "type": "text", + "content": "[11] Nima Khademi Kalantari, Ting-Chun Wang, and Ravi Ramamoorthi. Learning-based view synthesis for light field cameras. ACM Transactions on Graphics, 35(6):1-10, 2016. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 592, + 294, + 624 + ], + "type": "text", + "content": "[12] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the annual conference on Computer Graphics and Interactive Techniques. ACM, 1996. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 625, + 294, + 667 + ], + "type": "text", + "content": "[13] Ming Li, Jun Liu, Ce Zheng, Xinming Huang, and Ziming Zhang. Exploiting multi-view part-wise correlation via an efficient transformer for vehicle re-identification. IEEE Transactions on Multimedia, 25:919–929, 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 712 + ], + "type": "text", + "content": "[14] Lili Lin, Qiujiang Li, Bin Gao, Yuxiang Yan, Wenhui Zhou, and Ercan Engin Kuruoglu. Unsupervised learning of light field depth estimation with spatial and angular consistencies. Neurocomput., 501(C):113-122, 2022. 6, 7" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[15] In Kyu Park, Kyoung Mu Lee, et al. Robust light field depth estimation using occlusion-noise aware data costs. IEEE Transactions on Pattern Analysis and Machine Intelligence, 40(10):2484-2497, 2017. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 171 + ], + "type": "text", + "content": "[16] Jiayong Peng, Zhiwei Xiong, Dong Liu, and Xuejin Chen. Unsupervised depth estimation from light field using a convolutional neural network. In Proceedings of the IEEE International Conference on 3D Vision, pages 295-303. IEEE, 2018. 1, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 216 + ], + "type": "text", + "content": "[17] Jiayong Peng, Zhiwei Xiong, Yicheng Wang, Yueyi Zhang, and Dong Liu. Zero-shot depth estimation from light field using a convolutional neural network. IEEE Transactions on Computational Imaging, 6:682-696, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 249 + ], + "type": "text", + "content": "[18] Martin Rerabek and Touradj Ebrahimi. New light field image dataset. In Proceedings of the International Conference on Quality of Multimedia Experience, 2016. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 293 + ], + "type": "text", + "content": "[19] Jinglei Shi, Xiaoran Jiang, and Christine Guillemot. A framework for learning depth from a flexible subset of dense and sparse light field views. IEEE Transactions on Image Processing, 28(12):5867-5880, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 295, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 360 + ], + "type": "text", + "content": "[20] Wenzhe Shi, Jose Caballero, Ferenc Huszar, Johannes Totz, Andrew P. Aitken, Rob Bishop, Daniel Rueckert, and Zehan Wang. Real-time single image and video super-resolution using an efficient sub-pixel convolutional neural network. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1874-1883, 2016. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 317, + 361, + 553, + 425 + ], + "type": "text", + "content": "[21] Changha Shin, Hae-Gon Jeon, Youngjin Yoon, In So Kweon, and Seon Joo Kim. Epinet: A fully-convolutional neural network using epipolar geometry for depth from light field images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4748–4757, 2018. 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 317, + 426, + 553, + 470 + ], + "type": "text", + "content": "[22] Abhilash Sunder Raj, Michael Lowney, Raj Shah, and Gordon Wetzstein. The stanford litro light field archive. http://lightfields.stanford.edu/LF2016.html, 2016. Accessed: Aug. 22, 2020. 7, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 472, + 553, + 503 + ], + "type": "text", + "content": "[23] C. Tomasi and R. Manduchi. Bilateral filtering for gray and color images. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 839-846, 1998. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 505, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 505, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 317, + 505, + 553, + 557 + ], + "type": "text", + "content": "[24] Yu-Ju Tsai, Yu-Lun Liu, Ming Ouhyoung, and Yung-Yu Chuang. Attention-based view selection networks for lightfield disparity estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 12095-12103, 2020. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 553, + 613 + ], + "type": "text", + "content": "[25] Longguang Wang, Yulan Guo, Yingqian Wang, Zhengfa Liang, Zaiping Lin, Jungang Yang, and Wei An. Parallax attention for unsupervised stereo correspondence learning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(4):2108-2125, 2022. 1, 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 615, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 615, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 615, + 555, + 669 + ], + "type": "text", + "content": "[26] Yingqian Wang, Longguang Wang, Zhengyu Liang, Jungang Yang, Wei An, and Yulan Guo. Occlusion-aware cost constructor for light field depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19809-19818, 2022. 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[27] Sven Wanner, Stephan Meister, and Bastian Goldluecke. Datasets and benchmarks for densely sampled 4d light fields. In Proceedings of the Conference on Vision, Modeling, and Visualization, pages 225-226. ACM, 2013. 8" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "6496" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 295, + 363 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Wenbin Yan, Xiaogang Zhang, and Hua Chen. Occlusion-aware unsupervised light field depth estimation based on multi-scale gans. IEEE Transactions on Circuits and Systems for Video Technology, 34(7):6318-6333, 2024. 2, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 162 + ], + "type": "text", + "content": "[29] Zhichao Yin and Jianping Shi. Geonet: Unsupervised learning of dense depth, optical flow and camera pose. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1983-1992, 2018. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "spans": [ + { + "bbox": [ + 57, + 163, + 295, + 206 + ], + "type": "text", + "content": "[30] Shuo Zhang, Hao Sheng, Chao Li, Jun Zhang, and Zhang Xiong. Robust depth estimation for light field via spinning parallelogram operator. Computer Vision and Image Understanding, 145:148-159, 2016. 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 294, + 262 + ], + "type": "text", + "content": "[31] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Unsupervised light field depth estimation via multi-view feature matching with occlusion prediction. IEEE Transactions on Circuits and Systems for Video Technology, 34(4):2261-2273, 2024. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 263, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 263, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 263, + 294, + 306 + ], + "type": "text", + "content": "[32] Wenhui Zhou, Enci Zhou, Gaomin Liu, Lili Lin, and Andrew Lumsdaine. Unsupervised monocular depth estimation from light field image. IEEE Transactions on Image Processing, 29:1606-1617, 2020. 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 308, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 308, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 57, + 308, + 294, + 363 + ], + "type": "text", + "content": "[33] Wenhui Zhou, Lili Lin, Yongjie Hong, Qiujuan Li, Xingfa Shen, and Ercan Engin Kuruoglu. Beyond photometric consistency: Geometry-based occlusion-aware unsupervised light field disparity estimation. IEEE Transactions on Neural Networks and Learning Systems, pages 1-15, 2023. 2, 7" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6497" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_content_list.json b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..364206022c9c2d14d261aeedbfb6e489ca994485 --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_content_list.json @@ -0,0 +1,1652 @@ +[ + { + "type": "text", + "text": "Epona: Autoregressive Diffusion World Model for Autonomous Driving", + "text_level": 1, + "bbox": [ + 137, + 94, + 859, + 118 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kaiwen Zhang $^{1,2*}$ Zhenyu Tang $^{1,3*}$ Xiaotao Hu $^{1,5}$ Xingang Pan $^{6}$ Xiaoyang Guo $^{1}$ Yuan Liu $^{5}$ Jingwei Huang $^{7}$ Li Yuan $^{3}$ Qian Zhang $^{1}$ Xiao-Xiao Long $^{4\\dagger}$ Xun Cao $^{4}$ Wei Yin $^{1}$ $\\S$", + "bbox": [ + 99, + 143, + 933, + 184 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Horizon Robotics $^{2}$ Tsinghua University $^{3}$ Shenzhen Gruadate School, Peking University $^{4}$ Nanjing University $^{5}$ Hong Kong University of Science and Technology $^{6}$ Nanyang Technological University $^{7}$ Tencent Hunyuan", + "bbox": [ + 120, + 189, + 875, + 246 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(A) Consistent High-resolution Long Video Generation", + "text_level": 1, + "bbox": [ + 125, + 251, + 501, + 266 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/481b9130b93e2a87f7b0b4c6e84e88da18c65d77ec21c4c48018bad82e58ee85.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 125, + 268, + 225, + 388 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/03bb646c4b8ad1b03b1bd3b5bdab749a995466dd7e3cd649b73c18191b662dd6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 233, + 268, + 848, + 388 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(B) Trajectory-controlled Video Generation", + "text_level": 1, + "bbox": [ + 127, + 393, + 419, + 407 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/b4f070a0c9b5dae181cea8b0c5d8fc80d65d71100299839a378d11ffb523da50.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 127, + 410, + 846, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(C)Traffic World Knowledge Understanding", + "text_level": 1, + "bbox": [ + 127, + 523, + 421, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4c8ede8bf3c6201cfe1a2019d3c646a6a018d88dd30ff90e1dfc2a3d0b523d6e.jpg", + "image_caption": [ + "Figure 1. Versatile capabilities of Epona. Given historical driving context, our Epona can generate consistent minutes-long future driving scenes at high resolution (A). It can be controlled by diverse trajectories (B), and understand real-world traffic knowledge (C). In addition, our world model can predict future trajectories and serve as an end-to-end real-time motion planner (D)." + ], + "image_footnote": [], + "bbox": [ + 125, + 539, + 635, + 616 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(D) End-to-End Trajectory Planning", + "text_level": 1, + "bbox": [ + 638, + 523, + 879, + 537 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6428349483a4095b34b420c3983bc4ec9c2c58dfd1f428356b7aea6b52bb1808.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 542, + 846, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 672, + 326, + 688 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Diffusion models have demonstrated exceptional visual quality in video generation, making them promising for autonomous driving world modeling. However, existing video diffusion-based world models struggle with flexible-length, long-horizon predictions and integrating trajectory planning. This is because conventional video diffusion models rely on global joint distribution modeling of fixed-length frame sequences rather than sequentially constructing localized distributions at each timestep. In this work, we propose Epona, an autoregressive diffusion world model that enables localized spatiotemporal distribution modeling", + "bbox": [ + 88, + 700, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "through two key innovations: 1) Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation, and 2) Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework. Our architecture enables high-resolution, long-duration generation while introducing a novel chain-of-forward training strategy to address error accumulation in autoregressive loops. Experimental results demonstrate state-of-the-art performance with $7.4\\%$ FVD improvement and minutes longer prediction duration compared to prior works. The learned world model further serves as a real-time motion planner, outperforming strong end-to-end planners on NAVSIM benchmarks.", + "bbox": [ + 511, + 672, + 908, + 885 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 47 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work. †Project advisor. §Project lead, Corresponding author.", + "bbox": [ + 89, + 875, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27220", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 90, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recently, with the rapid development of video generation models, world models have attracted significant attention and emerged as a powerful paradigm for physical world simulations and autonomous decision-making [1, 13, 21, 34]. These foundation models enable agents to understand inherent world knowledge and predict future dynamics, making them particularly promising for autonomous driving. Unlike traditional separate perception-planning pipelines, which require extensive annotations and explicit supervision, generative driving world models [11, 16, 18, 24, 26, 59, 60, 63, 70] integrate visual scene understanding with future prediction in a self-supervised manner, offering a new solution toward end-to-end autonomous driving.", + "bbox": [ + 89, + 114, + 480, + 311 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative world models primarily fall into two categories: diffusion-based approaches and GPT-style autoregressive methods. The diffusion-based paradigm (e.g., Vista [18]), while achieving impressive visual fidelity through joint distribution modeling of fixed-length videos [3, 23, 47, 50], fundamentally suffers from its inability to model per-timestep local distributions. This limitation manifests in critical deficiencies: failure to support variable-length long-range prediction crucial for dynamic world simulation, and infeasible trajectory planning due to the lack of multimodal prediction mechanism.", + "bbox": [ + 89, + 311, + 480, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Conversely, GPT-style approaches [5, 44, 45, 55] address temporal flexibility through autoregressive next-token prediction (as seen in GAIA-1 [24]). However, the quantization and tokenization process significantly degrades visual quality and planning precision. Moreover, the causal nature of autoregressive transformers constrains them to predicting only the next action rather than planning long-horizon trajectories [11], limiting their potential to serve as end-to-end driving planners. Both paradigms reveal complementary shortcomings - diffusion models lack temporal decomposition while autoregressive transformers sacrifice continuous visual precision - highlighting the need for a unified framework that reconciles these divergent advantages for competent driving world modeling.", + "bbox": [ + 89, + 478, + 482, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We introduce Epona, an autoregressive diffusion world model that achieves high-resolution long-horizon video generation and accurate trajectory planning. Our core innovation comes from three key designs: 1) Decoupled spatiotemporal factorization. While existing video diffusion methods model joint spatial-temporal distributions of past and future frames, we assume their temporal latent modeling lacks explicit causality constraints, leading to error accumulation in long sequences. Epona addresses this through spacetime-disentangled processing: A GPT-style transformer with causal attention handles temporal dynamics in compressed latent space, while twin diffusion transformers separately handle spatial rendering and trajectory generation. 2) Asynchronous multi-modal genera", + "bbox": [ + 89, + 688, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion. Building upon this foundation, we decouple trajectory planning from visual generation through parallel denoising processes. Two specialized DiTs [33, 41] asynchronously generate 3-second vehicle trajectories and the single next future frame. Both streams share flow-matching objectives [2, 36, 37] conditioned on the same temporal latent, ensuring alignment while preserving modality-specific optimizations. 3) A chain-of-forward training strategy addresses error accumulation and content drift in autoregressive loops.", + "bbox": [ + 511, + 90, + 903, + 242 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Epona offers several additional advantages. 1) Long-horizon generation. Our autoregressive diffusion model can achieve long-time generation over up to 2 minutes, significantly outperforming existing world models. 2) Real-time trajectory planning. The separate multi-modal generation architecture enables to perform trajectory planning solely while video prediction is deactivated, significantly reducing the inference FLOPS. This enables high-quality and even real-time trajectory planning, achieving rates of up to $20\\mathrm{Hz}$ . 3) Visual detail preservation. Our autoregressive formulation adopts continuous visual tokenizer instead of discrete ones, thus preserving rich scene details.", + "bbox": [ + 511, + 243, + 903, + 425 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments demonstrate the effectiveness and superiority of our world model. For video generation, our model achieves state-of-the-art FVD [52] on the NuScenes [6] benchmark, surpassing the best-performing Vista [18] by $7.4\\%$ , while extending generation length from 15 seconds to over 2 minutes (600 frames). Thanks to joint supervision with trajectory prediction, Epona allows action control to simulate diverse driving scenarios, as shown in Fig. 1 (B). For motion planning, our method outperforms strong end-to-end planners on NAVSIM [12] without perception inputs (e.g., 3D boxes/lanes). Notably, we observe that Epona learns essential traffic world knowledge (e.g., stop driving at red light) purely from self-supervised future prediction tasks, as shown in Fig. 1 (C). This suggests that our world model can implicitly learn real-world driving dynamics, making it a promising direction for next-generation autonomous driving systems.", + "bbox": [ + 511, + 426, + 903, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 700, + 653, + 715 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. World Models for Autonomous Driving", + "text_level": 1, + "bbox": [ + 511, + 726, + 849, + 742 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Constructing real-world driving world models have drawn considerable attention in recent years, among which vision-centric approaches gain prominence due to their superior sensor flexibility, data accessibility, and more human-like representation forms. Early efforts primarily focused on adapting pre-trained diffusion models (e.g., Stable Diffusion [3, 47]) to driving scenarios through fine-tuning. However, these methods either lacked critical planning modules [16, 18, 19] or were limited to low-resolution, short-term generation [17, 38, 59, 60, 64, 65], making them un", + "bbox": [ + 511, + 750, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27221", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/4d2234653f65cc239e1d00117dc7c8a02bea139b8ac5aaf2222cce526d8efdba.jpg", + "image_caption": [ + "Figure 2. Overview of Epona. Our world model utilizes a multimodal spatiotemporal transformer to process the historical context of the first $T$ frames and employs a next-frame prediction DiT to generate the frame at $T + 1$ and a trajectory planning DiT to forecast the future $N$ -frame pose trajectory. By adopting a chain-of-forward strategy, our approach enables high-quality and long-horizon video generation with an autoregressive manner." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 901, + 244 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "suitable for consistent long-range prediction and real-time planning. Recent work [11, 24, 26, 63, 70] explored harnessing GPT-like architecture to unify visual and action modeling and achieved long-range autoregressive generation. Yet, these methods require encoding images and trajectories into discrete tokens [14, 29, 54], which significantly degrades visual fidelity and trajectory precision. Similarly, while the newly released Cosmos [1] foundation model can serve as a driving world model, it does not introduce a new framework, facing the same limitations as the previous methods. In addition, its large parameter count and high computational demands limit its practicality. In contrast, we propose a novel autoregressive diffusion world model framework for autonomous driving, enabling long-range autoregressive generation in continuous visual and trajectory representations.", + "bbox": [ + 91, + 321, + 483, + 564 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Long Video Generation", + "text_level": 1, + "bbox": [ + 89, + 575, + 307, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Long-term prediction is not only a key challenge for current video generation models but also a crucial capability for robust world models, as it reflects the model's ability to learn consistent environment dynamics and accurately simulate real-world temporal progression [13]. Since original video diffusion models (e.g. SVD [3]) are limited to fixed-length short clips generation, prior methods have explored extrapolating video length by noise rescheduling [43], overlapped generation [57, 58] or hierarchical generation [66]. However, these techniques fail to resolve inherent model constraints, often resulting in inconsistencies and abrupt visual changes in long videos. Autoregressive approaches [11, 56, 61] naturally support variable-length generation but suffer from quality degradation due to domain shift between teacher-forcing training and error accumulation in sampling. GameNGen [53] and DrivingWorld [26] introduce noise augmentation and random token dropout during training to alleviate the problem by simulating the error in sampling, but still limited to specific model architectures. We propose a general chain-of-forward strat", + "bbox": [ + 91, + 599, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "egy allowing the model to directly learn inference errors during training, effectively reducing autoregressive drift. Meanwhile, recent works such as Diffusion Forcing [8, 49] and FIFO-Diffusion [31] explore integrating autoregressive generation in video diffusion by adjusting frame-wise noise levels and leveraging causal network designs. Our model adopts a similar causal temporal modeling strategy but redefines the architecture into a two-stage end-to-end framework, allowing joint generation of motion plans and next-frame images.", + "bbox": [ + 511, + 321, + 906, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 486, + 604, + 502 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we formally present the model framework and training techniques of Epona. We begin with preliminaries on diffusion models in Sec. 3.1 and discuss our world model formulation design insights in Sec. 3.2. Then we introduce our proposed autoregressive diffusion world model framework in Sec. 3.3, including three dedicated modules: a multimodal spatiotemporal transformer to capture historical context, a trajectory planning DiT to generate future 3-seconds trajectories, and a next-frame prediction DiT to generate the next-frame images. To mitigate autoregressive drift and enable long-horizon video generation, we propose a simple yet effective chain-of-forward training strategy, detailed in Sec.3.4. Additionally, to enhance video quality, we introduce a temporal-aware DCAE decoder in Sec.3.5. An overview of our method is illustrated in Fig. 2.", + "bbox": [ + 511, + 512, + 908, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 511, + 748, + 643, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models [23, 50] are a family of powerful generative models that transform noise samples $x_{(1)}$ drawn from a prior distribution $p_1 = \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ into data samples $x_{(0)}$ from the target distribution in terms of a differentiable equation:", + "bbox": [ + 511, + 768, + 906, + 830 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd x _ {(t)} = v _ {\\Theta} \\left(x _ {(t)}, t\\right) d t, t \\in [ 0, 1 ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 842, + 906, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where velocity $v$ is parametrized by a neural network $\\Theta$ . Rectified flow [2, 36, 37] proposes to define a straight prob-", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27222", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a6bdef404dd05a932485808fae60e82db343610c0f1412aa90acb47d3b9ac0d3.jpg", + "image_caption": [ + "Figure 3. Comparison of Different World Modeling Formulation. Up: Conventional autoregressive pipeline quantizes continuous images into discrete tokens and perform next-token prediction iteratively. Middle: The video-diffusion-based methods generate future $n$ frames simultaneously. Down: Our method autoregressively predicts fine-grained future frames in continuous space." + ], + "image_footnote": [], + "bbox": [ + 94, + 93, + 480, + 233 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ability path between $p_0$ and $p_1$ to improve training and sampling efficiency and optimize the network $\\Theta$ using a velocity prediction loss:", + "bbox": [ + 89, + 340, + 483, + 387 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nx _ {(t)} = (1 - t) x _ {(0)} + t \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 393, + 483, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {R F} = \\mathbb {E} _ {x _ {(0)}, \\epsilon , t} \\left[ \\| v _ {\\Theta} (x _ {(t)}, t) - (x _ {(0)} - \\epsilon) \\| ^ {2} \\right], \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 431, + 480, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which has proven to be an effective and scalable approach in state-of-the-art image and video generation models [15, 27, 33]. In Epona, we adopt the diffusion model and rectified flow objective for both next-frame image and trajectory generation. Particularly, to enhance efficiency, we encode images into compact latents using a pre-trained deep compression encoder [9] and adopt a latent diffusion model [47] for image synthesis.", + "bbox": [ + 89, + 454, + 483, + 575 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Reformulation of World Model Designs", + "text_level": 1, + "bbox": [ + 89, + 583, + 429, + 599 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we discuss different world model formulation design choices, as illustrated in Fig. 3. Given a sequence of previous front-view camera observations $\\{\\mathbf{O}_t\\}_{t=1}^T$ and the corresponding driving trajectory $\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T$ , the goal of driving world moels is to predict future driving dynamics based on historical context. Here, each driving action $\\mathbf{a}_{t_1 \\to t_2} := (\\Delta \\theta_{t_1 \\to t_2}, \\Delta x_{t_1 \\to t_2}, \\Delta y_{t_1 \\to t_2}) \\in \\mathbb{R}^3$ represents the vehicle's motion from $t_1$ to $t_2$ , where $\\Delta \\theta$ denotes the orientation change, and $(\\Delta x, \\Delta y)$ specify the relative displacement in the ego-coordinate frame. For consistency, we define $\\mathbf{a}_{0 \\to 1} = (0,0,0)$ . Existing methods tackle this problem by formulating world modeling in the following two ways:", + "bbox": [ + 89, + 604, + 483, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video diffusion-based world models. Current leading driving world models, like Vista [18], formulates world modeling in the form of video diffusion models [3], which jointly capture the global spatiotemporal distribution of both past and a fixed-length future,", + "bbox": [ + 89, + 801, + 483, + 876 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\{\\mathbf {O} _ {T + i} \\} _ {i = 1} ^ {n}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 183, + 883, + 388, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This formulation disrupts the causal temporal structure between historical observations and future predictions, limiting its ability to model real-world progressive dynamics and generate flexible-length long-term videos.", + "bbox": [ + 511, + 90, + 905, + 151 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GPT-based world models. Alternatively, autoregressive transformer-based world models [11, 24, 26] discretize image observations into token sequences $\\mathbf{O}_t = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots ,\\mathbf{t}_L]$ and model the conditional image distribution as a token-by-token prediction,", + "bbox": [ + 511, + 152, + 906, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\prod_ {i = 1} ^ {L} p (\\mathbf {t} _ {i} \\mid \\mathbf {t} _ {< i}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}).\n$$\n", + "text_format": "latex", + "bbox": [ + 612, + 238, + 807, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, this independent token modeling weakens spatial correlations and the quantization process distorts high-frequency details, leading to degraded generation quality.", + "bbox": [ + 511, + 290, + 905, + 335 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our approach. In contrast, we formulate world modeling as a sequential future prediction process in the temporal domain. Specifically, given past driving observations $\\{\\mathbf{O}_t\\}_{t=1}^T$ and the driving trajectory $\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T$ , our model predicts both a policy for future trajectory planning,", + "bbox": [ + 511, + 335, + 905, + 412 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi \\left(\\{\\mathbf {a} _ {T \\rightarrow T + i} \\} _ {i = 1} ^ {n} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 421, + 841, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and a conditional distribution over the next-frame camera observation as a whole,", + "bbox": [ + 511, + 450, + 905, + 479 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {O} _ {T + 1} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}, \\mathbf {a} _ {T \\rightarrow T + 1}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 492, + 846, + 512 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The next-frame prediction is conditioned on either a model-predicted action or an externally provided action $\\mathbf{a}_{T\\rightarrow T + 1}$ . By decoupling causal temporal modeling from fine-grained future prediction, our model can generate flexible-length long videos autoregressively in continuous representations. Moreover, by factorizing trajectory planning from visual generation, our model can seamlessly serve as a real-time motion planner, bridging a critical gap between current driving world models and end-to-end motion planners.", + "bbox": [ + 511, + 521, + 905, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Epona: Autoregressive Diffusion World Model", + "text_level": 1, + "bbox": [ + 511, + 667, + 905, + 683 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on the reformulated world modeling design, we propose Epona, an autoregressive diffusion world model for autonomous driving. Our framework consists of three key components. First, a Multimodal Spatiotemporal Transformer (MST) encodes historical context $\\{\\mathbf{O}_t,\\mathbf{a}_t\\}_{t = 1}^T$ into a compact latent representation, effectively capturing environmental context and driving dynamics. Then, based on the historical latents, we employ two specialized diffusion transformers to predict fine-grained future details, including a tiny Trajectory Planning Transformer (TrajDiT) that models the policy $\\pi$ for trajectory planning, and a Next-frame Prediction Transformer (VisDiT) that models the visual distribution $p$ for future image generation. This modular design enables a range of autonomous driving applications.", + "bbox": [ + 511, + 688, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27223", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For instance, MST and VisDiT can be used independently for controllable driving simulations, while MST and TrajDiT facilitate real-time motion planning.", + "bbox": [ + 89, + 90, + 482, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multimodal Spatiotemporal Transformer (MST). Given the encoded past driving scenes $\\{\\mathbf{Z}_t\\}_{t=1}^T$ and trajectory $\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T$ , we introduce a multimodal spatiotemporal transformer to effectively integrate temporal dynamics and multimodal information from historical context for future prediction. Inspired by prior work in video generation and world modeling [4, 26, 39], our approach employs interleaved multimodal spatial attention layers and causal temporal attention layers. This design progressively incorporates historical information into a compact latent representation while significantly reducing memory consumption compared to full-sequence attention. Additionally, this design naturally supports historical contexts of variable length.", + "bbox": [ + 89, + 136, + 482, + 347 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, we first project the flattened visual latent patches $\\mathbf{Z} \\in \\mathbb{R}^{B \\times T \\times L \\times C}$ and action sequences $\\mathbf{a} \\in \\mathbb{R}^{B \\times T \\times 3}$ into an embedding space. Then we concatenate them along the spatial dimension and add temporal positional embeddings to obtain the latent embedding sequence $\\mathbf{E} \\in \\mathbb{R}^{B \\times T \\times (L + 3) \\times D}$ . This sequence is processed through interleaved multimodal spatiotemporal layers as follows (using einops [46] notation):", + "bbox": [ + 89, + 347, + 483, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b t) l c \\rightarrow (b l) t c)) \\\\ \\mathbf {E} \\leftarrow \\text {C a u s a l T e m p o r a l L a y e r} (\\mathbf {E}, \\text {C a u s a l M a x}) \\\\ \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b l) t c \\rightarrow (b t) l c)) \\\\ \\mathbf {E} \\leftarrow \\operatorname {M u l t i m o d a l S p a t i a l L a y e r} (\\mathbf {E}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 93, + 474, + 477, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $B$ is the batch size, $T$ is the number of conditioning frames, $L = H \\times W$ is the number of flattened latents in an image, $C$ is the image latent channel dimension, $D$ is the embedding dimension, and CausalMask is the triangular causal attention mask. Finally, we use the latent embedding of the last frame, $\\mathbf{F} \\in \\mathbb{R}^{B \\times (L + 3) \\times D}$ , as the compact latent representation for the next-stage prediction. After training, this embedding encapsulates the historical context $\\{\\mathbf{O}_t, \\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T$ .", + "bbox": [ + 89, + 551, + 482, + 688 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Trajectory Planning Diffusion Transformer (TrajDiT). TrajDiT predicts future trajectories using a tiny diffusion transformer. Following the DiT frameworks in most advanced open-source text-to-image and video generation models [27, 33], we adopt a Dual-Single-Stream architecture. In the dual-stream phase, the historical latent representation $\\mathbf{F}$ and trajectory data are processed independently through transformer blocks, with only attention operations linking them. In the single-stream phase, they are concatenated to pass through subsequent transformer blocks for effective information fusion. Detailed architecture can be found in the supplementary material.", + "bbox": [ + 89, + 689, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During training, we add noise to the target trajectories $\\overline{\\mathbf{a}}\\in \\mathbb{R}^{B\\times N\\times 3}$ using Eq.2. The model then predicts velocity", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/64871f15011ea7d132dfb619eb87548b53b9522938672b693071e5e462e306ba.jpg", + "image_caption": [ + "Figure 4. Concept illustration of our training process. Here $x$ can be either image latents or trajectories." + ], + "image_footnote": [], + "bbox": [ + 532, + 80, + 883, + 218 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$v_{\\text{traj}}$ conditioned on $\\mathbf{F}$ , where $N$ is the planning horizon. We optimize using the rectified flow loss:", + "bbox": [ + 511, + 266, + 903, + 295 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t r a j} = \\mathbb {E} _ {\\overline {{\\mathbf {a}}}, \\epsilon , t} \\left[ \\| v _ {t r a j} (\\overline {{\\mathbf {a}}} _ {(t)}, t) - (\\overline {{\\mathbf {a}}} - \\epsilon) \\| ^ {2} \\right]. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 308, + 903, + 327 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For inference, random Gaussian noise is iteratively denoised conditioned on $\\mathbf{F}$ to generate future trajectory plans. Next-frame Prediction Diffusion Transformer (VisDiT). VisDiT has a similar architecture as TrajDiT, with an additional modulation [41] branch for action control $\\mathbf{a}_{T\\rightarrow T + 1}$ . We also use the flow loss for visual supervision:", + "bbox": [ + 511, + 339, + 903, + 431 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v i s} = \\mathbb {E} _ {\\mathbf {Z} _ {T + 1}, \\epsilon , t} \\left[ \\| v _ {v i s} \\left(\\mathbf {Z} _ {T + 1 (t)}, t\\right) - \\left(\\mathbf {Z} _ {T + 1} - \\epsilon\\right) \\| ^ {2} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 441, + 903, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Together, the total loss jointly optimizes the entire world model:", + "bbox": [ + 511, + 474, + 903, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {t r a j}} + \\mathcal {L} _ {\\text {v i s}}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 643, + 507, + 903, + 525 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During inference, VisDiT denoises $\\hat{\\mathbf{Z}}_{T + 1}$ , conditioned on $\\mathbf{F}$ and the action either predicted by TrajDiT or provided by user. The latents are then decoded using the DCAE decoder to generate the next-frame image $\\hat{\\mathbf{O}}_{T + 1}$ .", + "bbox": [ + 511, + 532, + 903, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Chain-of-Forward Training", + "text_level": 1, + "bbox": [ + 511, + 606, + 763, + 622 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With our proposed autoregressive diffusion world model, we can autoregressively generate future videos frame by frame. However, long-term generation suffers from a longstanding autoregressive drift problem [53]: during training, the model predicts the next frame using ground-truth historical context, whereas during inference, it relies on its own past predictions. This domain gap between teacher-forcing training and autoregressive sampling leads to error accumulation and rapid quality degradation.", + "bbox": [ + 511, + 628, + 903, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate this, we introduce a chain-of-forward training strategy. Periodically, we perform multiple forward passes using self-predicted frames to enhance the model's robustness to inference noise (see Fig. 4). Notably, to ensure training efficiency, instead of sampling next-frame latents from pure noise, we leverage the model-predicted velocity $v_{\\Theta}$ to estimate the denoised latents in one step:", + "bbox": [ + 511, + 765, + 903, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {(0)} = x _ {(t)} + t v _ {\\Theta} (x _ {(t)}, t) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 885, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27224", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/45788c2453ea7f8973e9ef24dec3ef354493d3389183695dc7f069595d2e691b.jpg", + "table_caption": [ + "Table 1. Comparisons of generated videos on the NuScenes [6] validation set. Our model achieves state-of-the-art FVD score compared to existing driving world models, while extending the video length to over two minutes. *The max duration number indicates the horizon that produces plausible results, following existing methods." + ], + "table_footnote": [], + "table_body": "
MetricDriveGAN [32]DriveDreamer [59]WoVoGen [38]Drive-WM [60]GenAD (OpenDV) [65]Vista [18]DrivingWorld [26]Ours
FID ↓73.452.627.615.815.46.97.47.5
FVD ↓502.3452.0417.7122.7184.089.490.982.8
Max Duration / Frames*N/A4s / 482.5s / 58s / 164s / 815s / 15040s / 400120s / 600
", + "bbox": [ + 91, + 130, + 906, + 186 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/03c723b6620506ffd57952f097423f4b6dc2b495c924bae75f3cab48c2140f76.jpg", + "image_caption": [ + "Figure 5. Qualitative Comparison between Vista [18] and Epona. Zoom in for better views." + ], + "image_footnote": [], + "bbox": [ + 96, + 204, + 890, + 328 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The estimated $\\hat{x}_{(0)}$ , along with previous conditioned frames, is then used in the next forward pass to autoregressively generate subsequent frames. This process simulates prediction noise, helping the model adapt to deviations and improving long-term video generation quality.", + "bbox": [ + 89, + 358, + 482, + 434 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.5. Temporal-aware DCAE Decoder", + "text_level": 1, + "bbox": [ + 89, + 441, + 377, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Unlike conventional autoencoders that downsample images by a factor of 8, DCAE [9] progressively increases this to 32, reducing latent tokens by $16 \\times$ . In our world model, we adopt DCAE for image encoding to improve training efficiency and reduce memory usage, enabling conditioning on longer historical contexts.", + "bbox": [ + 89, + 463, + 482, + 553 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "However, as an image autoencoder, DCAE lacks temporal interactions, causing flickering when decoding video frame by frame, which degrades visual quality. To address this, we propose a temporal-aware DCAE to enhance interframe consistency. Specifically, to maximize pretrained parameters while minimizing architectural changes, we introduce spatiotemporal self-attention layers before the DCAE decoder while keeping the encoder fixed during fine-tuning. This facilitates multi-frame interactions, greatly improving temporal consistency in generated videos.", + "bbox": [ + 89, + 555, + 482, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 718, + 214, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 742, + 305, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "World Model. Our Epona consists of $2.5\\,B$ parameters, including a 12-layer multimodal spatiotemporal transformer with $1.3\\,B$ parameters, a 12-layer next-frame prediction diffusion transformer with $1.2\\,B$ parameters, and a 2-layer trajectory planning diffusion transformer with $50\\,M$ parameters. It is trained on publicly available videos from the NuPlan dataset [7] and 700 scenes from the NuScenes dataset [6] from scratch, all images are resized to $512\\times 1024$ . We utilize the rectified flow [37] objective for both video", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/748afd99e6dfb5acd6dfc584d47665e877f88be5837bebf22edaca0dbf4d0bdf.jpg", + "table_caption": [ + "Table 2. Inference speed. We evaluate our inference speed for generating a 3-second trajectory and a $512 \\times 1024$ image per module on a single NVIDIA 4090 GPU." + ], + "table_footnote": [], + "table_body": "
DiT sampling stepsMSTTrajDiTVisDiT
10~0.02s~0.03s~0.3s
100~0.02s~0.3s~2s
", + "bbox": [ + 563, + 409, + 852, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "generation and trajectory planning tasks, training the entire model in an end-to-end manner. The training was conducted on 48 NVIDIA A100 GPUs for nearly two weeks, with a total of $600\\mathrm{k}$ iterations and a batch size of 96. During training, we apply Chain-of-Forward every 10 steps, each time performing three forward passes. We use the AdamW optimizer with a learning rate of $1\\times 10^{-4}$ and set the weight decay to $5\\times 10^{-2}$ . For inference, we report our speed for each module on a single NVIDIA 4090 GPU in Table 2. In all our experiments, we set our DiT sampling step to 100. Notice that with MST and TrajDiT, our world model can seamlessly serve as a real-time motion planner.", + "bbox": [ + 511, + 472, + 906, + 654 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "**Evaluations on video generation.** We employ 1628 video clips from the NuPlan test set [7] and 1646 video clips from the NuScenes validation dataset [6] for performance evaluation, respectively. During the test, our world model is conditioned on 10 consecutive past frames to generate the subsequent frame and repeat the process autoregressively to synthesize future video frames. We use the Frechet Video Distance (FVD) [52] and the Frechet Inception Distance (FID) [22] to evaluate the quality of the generated videos.", + "bbox": [ + 511, + 656, + 905, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluations on trajectory planning. We evaluate trajectory planning using the NuScenes benchmark [6] and the NAVSIM benchmark [12]. For the NuScenes, we use L2 error and collision rate as the evaluation metrics following the existing works [25, 48, 70] to evaluate the planning performance. L2 error measures the L2 distance between the predicted and ground truth trajectories, while the collision", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27225", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/a1ff9fc650f21fae731c82bf5a7210398edff5eb7a907a94e5e6162dbb8090d0.jpg", + "image_caption": [ + "Figure 6. Trajectory-controlled video generation. Our world model can generate controllable videos based on predefined trajectories." + ], + "image_footnote": [], + "bbox": [ + 169, + 87, + 823, + 303 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "rate measures the frequency of predicted trajectory intersections with objects. The NAVSIM benchmark assesses performance using the predictive driver model score (PDMS), derived from five factors, as shown in Table 4.", + "bbox": [ + 89, + 325, + 483, + 385 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation of Video Generation", + "text_level": 1, + "bbox": [ + 89, + 395, + 372, + 410 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison of Generated Videos. We present a quantitative comparison with existing methods on the NuScenes dataset [6] in Table 1. Since most methods are not publicly available, we compare with the reported results from their respective papers. Notably, the existing methods (e.g., Vista [18]) fine-tune video diffusion models pretrained on large-scale datasets, while our world model, including next-frame DiT, is trained from scratch. As shown in Table 1, our generated videos achieve state-of-the-art FVD scores, indicating smoother and more realistic video generation quality. Moreover, our world model can generate significantly longer video frames compared to existing approaches as shown in Table 1.", + "bbox": [ + 88, + 417, + 483, + 611 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Comparison of Generated Videos. We provide a qualitative comparison with the state-of-the-art open-source driving world model, Vista [18]. Since Vista is a 25-frame fixed-length video diffusion model, we perform rollout to generate longer videos as illustrated in their paper. As shown in Fig. 5, our Epona generates consistent long-horizon driving scenes with high-fidelity visuals and detailed structures and vehicles.", + "bbox": [ + 89, + 613, + 483, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Trajectory-controlled Video Generation. Fig. 6 illustrates the pose controllability of our model. Given the predefined pose trajectory, the different condition frames can generate future frames that conform to the corresponding motion path, which is crucial for obtaining autonomous driving videos under extreme scenarios.", + "bbox": [ + 89, + 734, + 483, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Extra Long-range Video Generation. Epona combines the strengths of autoregressive and diffusion models, facilitating the generation of high-quality, long-duration videos conditioned on input frames. As shown in Fig.5 and Fig.7, our model can autoregressively generate minute-long driv-", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing videos with high fidelity and consistency, without noticeable drift. More long-term generation videos are provided in the supplementary materials.", + "bbox": [ + 511, + 325, + 906, + 371 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation of Trajectory Planning", + "text_level": 1, + "bbox": [ + 511, + 378, + 812, + 395 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the NuScenes benchmark [6], we compare our world model with several existing methods, as shown in Table 3. Although our model does not achieve the best results, it attains competitive performance without any additional supervision. It is worth noting that incorporating more supervision typically leads to better performance but cost expensive annotations. Additionally, similar to Doe-1 [70], our model only utilizes the front camera, whereas other methods rely on multi-view inputs for planning. As shown in Table 3, our approach can generate reasonable trajectories while achieving the lowest collision rate for a 1-second horizon, which is crucial for long-term realistic video predictions. For the more challenging NAVSIM benchmark [12], as shown in Table 4, our method achieves state-of-the-art results in overall PDMS when conditioned on the past 2 seconds of observations to predict 4-second future trajectories, showing the strong motion planning capability.", + "bbox": [ + 511, + 401, + 906, + 659 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 667, + 666, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Shared Latent for Multi-modal Joint Prediction. To assess the benefit of jointly modeling scene and trajectory via a shared latent representation, we conduct an ablation by disabling video prediction and training the model solely for trajectory prediction. This variant is evaluated on the NAVISIM test set. As shown in Tab. 5, removing video prediction leads to a noticeable drop in planning performance. This result highlights the advantage of the shared latent $\\mathbf{F}$ , which encourages the world model to better capture complex driving dynamics by leveraging visual signals. This ablation confirms that coupling video and trajectory prediction within a unified latent space in world models significantly benefits downstream planning tasks.", + "bbox": [ + 511, + 688, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of Chain-of-Forward Training. To evaluate the", + "bbox": [ + 511, + 885, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27226", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/850e85a12901ccc0803d508487c57e1dd3b8f55520aeac6e5d06a58af680e119.jpg", + "image_caption": [ + "Figure 7. Qualitative Comparison between long videos generated by models w/ and w/o Chain-of-Forward training. Left: Visual quality deteriorates rapidly after 10–20 seconds. Right: The same driving scenes with Chain-of-Forward training maintain high visual quality, generating minute-long videos without significant degradation. Zoom in for better views." + ], + "image_footnote": [], + "bbox": [ + 101, + 80, + 500, + 195 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/f092074d6c47a2ac4f000ddcf052ca44a23197fc4eaa98dfbe27da2505a8e7a1.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 80, + 895, + 195 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2bafe04b0e29b1f4d7512057e238ec6aeadcd5206ac31102adbbe4d9e163fae7.jpg", + "table_caption": [ + "Table 3. End-to-end motion planning performance on the NuScenes [6] dataset. Note that our model achieves a low collision rate, demonstrating its understanding of basic traffic rules via simple next-frame prediction. * represents only using the front camera as input." + ], + "table_footnote": [], + "table_body": "
MethodInputAuxiliary SupervisionL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
ST-P3 [25]CameraMap & Box & Depth1.332.112.902.110.230.621.270.71
UniAD [48]CameraMap & Box & Motion & Tracklets & Occ0.480.961.651.030.050.170.710.31
OccNet [51]Camera3D-Occ & Map & Box1.292.132.992.140.210.591.370.72
OccWorld [68]Camera3D-Occ0.521.272.411.400.120.402.080.87
VAD-Tiny [30]CameraMap & Box & Motion0.601.232.061.300.310.531.330.72
VAD-Base [30]CameraMap & Box & Motion0.541.151.981.220.040.391.170.53
GenAD [69]CameraMap & Box & Motion0.360.831.550.910.060.231.000.43
Doe-1 [70]Camera*QA0.501.182.111.260.040.371.190.53
OursCamera*None0.611.171.981.250.010.220.850.36
", + "bbox": [ + 210, + 281, + 781, + 415 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d69014850276b8736e8551a70a2419668281efece069bd680158955cd37b11f2.jpg", + "table_caption": [ + "Table 4. End-to-end motion planning performance on the NAVSIM [12] test set. NC: no at-fault collision. DAC: drivable area compliance. TTC: time-to-collision. Conf.: comfort. EP: ego progress. PDMS: the predictive driver model score. LAW[35] is in the perception-free setting. Our world model outperforms strong end-to-end planners in the overall PDMS score." + ], + "table_footnote": [], + "table_body": "
MethodInputNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Human/10010010099.987.594.8
UniAD[48]Camera97.891.992.910078.883.4
PARA-Drive[62]Camera97.992.493.099.879.384.6
LAW[35]Camera96.495.488.799.981.784.6
TransFuser[42]Camera & Lidar97.792.892.810079.284.0
DRAMA[67]Camera & Lidar98.093.194.810080.185.5
VADv2[10]Camera & Lidar97.289.191.910076.080.9
OursCamera97.995.193.899.980.486.2
", + "bbox": [ + 93, + 517, + 480, + 616 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/3c3f61fc5472191db839e71cc042aec3e3039c95f0fefe5c69dd275366529988.jpg", + "table_caption": [ + "Table 5. Comparison of planning results on the NAVSIM test set. Jointly predicting the next scene using a shared latent significantly improves planning performance." + ], + "table_footnote": [], + "table_body": "
MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Ours w/o Joint Training94.589.788.199.974.778.1
Ours97.995.193.899.980.486.2
", + "bbox": [ + 93, + 670, + 480, + 712 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "impact of our chain-of-forward strategy on model performance, we conduct an ablation study comparing results with and without this strategy. Given that our model iteratively generates the next frame, the chain-of-forward approach simulates potential inference errors during training, thereby enhancing the model's robustness. As shown in Fig. 7 and Fig. 8, as the model autoregressively generates longer sequences, the gap in visual quality and FID score between models with and without the chain-of-forward strategy becomes increasingly significant, validating its effectiveness in long-term video generation. More ablations", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/e022709c0de5912cf52a031faa7a4ee59146b569c0f7162a0de4581b23bef58e.jpg", + "image_caption": [ + "Figure 8. Effect of Chain-of-Forward training. FID comparison in NuPlan test set between models w/ and w/o Chain-of-Forward training strategy." + ], + "image_footnote": [], + "bbox": [ + 540, + 436, + 870, + 601 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can be found in the supplementary material.", + "bbox": [ + 511, + 666, + 805, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 694, + 633, + 709 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented Epona, an autoregressive diffusion world model for autonomous driving that jointly predicts high-fidelity future trajectories and driving scenes based on historical driving context. Thanks to our proposed decoupled spatiotemporal modeling and asynchronous multimodal generation strategies, our model achieves high-quality and long-term prediction. In addition, our model could serves as a real-time motion planner via performing trajectory planning. We have demonstrated that our approach significantly advances the state of the art in driving world models, uncovering the large potential for building next-generation autonomous driving systems.", + "bbox": [ + 509, + 719, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27227", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3", + "[2] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. 2, 3", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3, 4", + "[4] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, pages 22563-22575, 2023. 5", + "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2", + "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 7, 8", + "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 6, 1", + "[8] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion, 2024. 3", + "[9] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 4, 6", + "[10] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024.8", + "[11] Yuntao Chen, Yuqi Wang, and Zhaoxiang Zhang. Driving-gpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers. arXiv preprint arXiv:2412.18607, 2024. 2, 3, 4, 1", + "[12] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7, 8" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models. arXiv preprint arXiv:2411.14499, 2024. 2, 3", + "[14] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 3", + "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 4", + "[16] Ruiyuan Gao, Kai Chen, Bo Xiao, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrivedit: High-resolution long video generation for autonomous driving with adaptive control. arXiv preprint arXiv:2411.13807, 2024. 2, 1", + "[17] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3d geometry control. In ICLR, 2024. 2", + "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 4, 6, 7", + "[19] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. Dome: Taming diffusion model into high-fidelity controllable occupancy world model. ArXiv, abs/2410.10429, 2024. 2", + "[20] Xi Guo, Chenjing Ding, Haoxuan Dou, Xin Zhang, Weixuan Tang, and Wei Wu. Infinitydrive: Breaking time limits in driving world models, 2024. 1", + "[21] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2", + "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6", + "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3", + "[24] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3, 4, 1", + "[25] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 6, 8", + "[26] Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. Drivingworld: Constructingworld model for autonomous driving via" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27228", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "video gpt. arXiv preprint arXiv:2412.19505, 2024. 2, 3, 4, 5, 6", + "[27] Tencent Hunyuan. Hunyuanvideo: A systematic framework for large video generative models, 2024. 4, 5, 1, 2", + "[28] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. Adriver-i: A general world model for autonomous driving, 2023. 1", + "[29] Mingkai Jia, Wei Yin, Xiaotao Hu, Jiaxin Guo, Xiaoyang Guo, Qian Zhang, Xiao-Xiao Long, and Ping Tan. Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization. arXiv preprint arXiv:2507.07997, 2025.3", + "[30] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 8", + "[31] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. In NeurIPS, 2024. 3", + "[32] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 6", + "[33] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 4, 5, 1", + "[34] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1):1-62, 2022. 2", + "[35] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. 2024. 8", + "[36] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 2, 3", + "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3, 6", + "[38] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. WoVoGen: World Volume-Aware Diffusion for Controllable Multi-Camera Driving Scene Generation. arXiv preprint arXiv:2312.02934, 2023. 2, 6", + "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation, 2024. 5", + "[40] Yiyang Ma, Xingchao Liu, Xi aokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Xingkai Yu, Liang Zhao, Yisong Wang, Jiaying Liu, and Chong Ruan. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. ArXiv, abs/2411.07975, 2024. 2", + "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "national conference on computer vision, pages 4195-4205, 2023. 2, 5", + "[42] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 8", + "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 3", + "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2", + "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2", + "[46] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2022. 5", + "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4", + "[48] Xiaogang Shi, Bin Cui, Gillian Dobbie, and Beng Chin Ooi. Uniad: A unified ad hoc data processing system. ACM Transactions on Database Systems (TODS), 42(1):1-42, 2016. 6, 8", + "[49] Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. History-guided video diffusion, 2025. 3", + "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3", + "[51] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 8", + "[52] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6", + "[53] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines, 2024. 3, 5", + "[54] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3", + "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2", + "[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27229", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 3", + "[58] Fu-Yun Wang, Zhaoyang Huang, Qiang Ma, Guanglu Song, Xudong Lu, Weikang Bian, Yijin Li, Yu Liu, and Hongsheng Li. Zola: Zero-shot creative long animation generation with short video model. In ECCV, 2024. 3", + "[59] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6", + "[60] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving. In CVPR, 2024. 2, 6", + "[61] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models, 2024. 3", + "[62] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 8", + "[63] Yanhao Wu, Haoyang Zhang, Tianwei Lin, Lichao Huang, Shujie Luo, Rui Wu, Congpei Qiu, Wei Ke, and Tong Zhang. Generating multimodal driving scenes via next-scene prediction. In CVPR, 2025. 2, 3", + "[64] Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goaldriven flow matching for multimodal trajectories generation in end-to-end autonomous driving. ArXiv, abs/2503.05689, 2025. 2", + "[65] Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu, Jia Zeng, Ping Luo, Jun Zhang, Andreas Geiger, Yu Qiao, and Hongyang Li. Generalized Predictive Model for Autonomous Driving. In CVPR, 2024. 2, 6", + "[66] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation, 2023. 3", + "[67] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, and Marcelo H. Ang Jr au2. Drama: An efficient end-to-end motion planner for autonomous driving with mamba, 2024. 8", + "[68] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In European conference on computer vision, pages 55-72. Springer, 2024. 8", + "[69] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87-104. Springer, 2024. 8" + ], + "bbox": [ + 91, + 90, + 480, + 901 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[70] Wenzhao Zheng, Zetian Xia, Yuanhui Huang, Sicheng Zuo, Jie Zhou, and Jiwen Lu. Doe-1: Closed-loop autonomous driving with large world model. arXiv preprint arXiv:2412.09627, 2024. 2, 3, 6, 7, 8", + "[71] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke S. Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. ArXiv, abs/2408.11039, 2024. 2" + ], + "bbox": [ + 514, + 90, + 905, + 217 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "27230", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_model.json b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_model.json new file mode 100644 index 0000000000000000000000000000000000000000..e7ba16e5c16e8d745e8851a06c1a380d96286315 --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_model.json @@ -0,0 +1,2433 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.048 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.138, + 0.095, + 0.861, + 0.119 + ], + "angle": 0, + "content": "Epona: Autoregressive Diffusion World Model for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.144, + 0.934, + 0.185 + ], + "angle": 0, + "content": "Kaiwen Zhang\\(^{1,2*}\\) Zhenyu Tang\\(^{1,3*}\\) Xiaotao Hu\\(^{1,5}\\) Xingang Pan\\(^{6}\\) Xiaoyang Guo\\(^{1}\\) Yuan Liu\\(^{5}\\) Jingwei Huang\\(^{7}\\) Li Yuan\\(^{3}\\) Qian Zhang\\(^{1}\\) Xiao-Xiao Long\\(^{4\\dagger}\\) Xun Cao\\(^{4}\\) Wei Yin\\(^{1}\\)\\(\\S\\)" + }, + { + "type": "text", + "bbox": [ + 0.121, + 0.19, + 0.877, + 0.247 + ], + "angle": 0, + "content": "\\(^{1}\\)Horizon Robotics \\(^{2}\\)Tsinghua University \\(^{3}\\)Shenzhen Gruadate School, Peking University \\(^{4}\\)Nanjing University \\(^{5}\\)Hong Kong University of Science and Technology \\(^{6}\\)Nanyang Technological University \\(^{7}\\)Tencent Hunyuan" + }, + { + "type": "title", + "bbox": [ + 0.127, + 0.252, + 0.502, + 0.267 + ], + "angle": 0, + "content": "(A) Consistent High-resolution Long Video Generation" + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.269, + 0.226, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.234, + 0.269, + 0.849, + 0.39 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.129, + 0.395, + 0.421, + 0.409 + ], + "angle": 0, + "content": "(B) Trajectory-controlled Video Generation" + }, + { + "type": "image", + "bbox": [ + 0.128, + 0.411, + 0.848, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.128, + 0.524, + 0.422, + 0.539 + ], + "angle": 0, + "content": "(C)Traffic World Knowledge Understanding" + }, + { + "type": "image", + "bbox": [ + 0.126, + 0.54, + 0.637, + 0.617 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.64, + 0.524, + 0.88, + 0.539 + ], + "angle": 0, + "content": "(D) End-to-End Trajectory Planning" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.544, + 0.848, + 0.61 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.624, + 0.908, + 0.667 + ], + "angle": 0, + "content": "Figure 1. Versatile capabilities of Epona. Given historical driving context, our Epona can generate consistent minutes-long future driving scenes at high resolution (A). It can be controlled by diverse trajectories (B), and understand real-world traffic knowledge (C). In addition, our world model can predict future trajectories and serve as an end-to-end real-time motion planner (D)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.673, + 0.327, + 0.689 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.701, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Diffusion models have demonstrated exceptional visual quality in video generation, making them promising for autonomous driving world modeling. However, existing video diffusion-based world models struggle with flexible-length, long-horizon predictions and integrating trajectory planning. This is because conventional video diffusion models rely on global joint distribution modeling of fixed-length frame sequences rather than sequentially constructing localized distributions at each timestep. In this work, we propose Epona, an autoregressive diffusion world model that enables localized spatiotemporal distribution modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.909, + 0.886 + ], + "angle": 0, + "content": "through two key innovations: 1) Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation, and 2) Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework. Our architecture enables high-resolution, long-duration generation while introducing a novel chain-of-forward training strategy to address error accumulation in autoregressive loops. Experimental results demonstrate state-of-the-art performance with \\(7.4\\%\\) FVD improvement and minutes longer prediction duration compared to prior works. The learned world model further serves as a real-time motion planner, outperforming strong end-to-end planners on NAVSIM benchmarks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.875, + 0.484, + 0.902 + ], + "angle": 0, + "content": "*These authors contributed equally to this work. †Project advisor. §Project lead, Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27220" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.312 + ], + "angle": 0, + "content": "Recently, with the rapid development of video generation models, world models have attracted significant attention and emerged as a powerful paradigm for physical world simulations and autonomous decision-making [1, 13, 21, 34]. These foundation models enable agents to understand inherent world knowledge and predict future dynamics, making them particularly promising for autonomous driving. Unlike traditional separate perception-planning pipelines, which require extensive annotations and explicit supervision, generative driving world models [11, 16, 18, 24, 26, 59, 60, 63, 70] integrate visual scene understanding with future prediction in a self-supervised manner, offering a new solution toward end-to-end autonomous driving." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.312, + 0.482, + 0.478 + ], + "angle": 0, + "content": "Generative world models primarily fall into two categories: diffusion-based approaches and GPT-style autoregressive methods. The diffusion-based paradigm (e.g., Vista [18]), while achieving impressive visual fidelity through joint distribution modeling of fixed-length videos [3, 23, 47, 50], fundamentally suffers from its inability to model per-timestep local distributions. This limitation manifests in critical deficiencies: failure to support variable-length long-range prediction crucial for dynamic world simulation, and infeasible trajectory planning due to the lack of multimodal prediction mechanism." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.689 + ], + "angle": 0, + "content": "Conversely, GPT-style approaches [5, 44, 45, 55] address temporal flexibility through autoregressive next-token prediction (as seen in GAIA-1 [24]). However, the quantization and tokenization process significantly degrades visual quality and planning precision. Moreover, the causal nature of autoregressive transformers constrains them to predicting only the next action rather than planning long-horizon trajectories [11], limiting their potential to serve as end-to-end driving planners. Both paradigms reveal complementary shortcomings - diffusion models lack temporal decomposition while autoregressive transformers sacrifice continuous visual precision - highlighting the need for a unified framework that reconciles these divergent advantages for competent driving world modeling." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We introduce Epona, an autoregressive diffusion world model that achieves high-resolution long-horizon video generation and accurate trajectory planning. Our core innovation comes from three key designs: 1) Decoupled spatiotemporal factorization. While existing video diffusion methods model joint spatial-temporal distributions of past and future frames, we assume their temporal latent modeling lacks explicit causality constraints, leading to error accumulation in long sequences. Epona addresses this through spacetime-disentangled processing: A GPT-style transformer with causal attention handles temporal dynamics in compressed latent space, while twin diffusion transformers separately handle spatial rendering and trajectory generation. 2) Asynchronous multi-modal genera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.243 + ], + "angle": 0, + "content": "tion. Building upon this foundation, we decouple trajectory planning from visual generation through parallel denoising processes. Two specialized DiTs [33, 41] asynchronously generate 3-second vehicle trajectories and the single next future frame. Both streams share flow-matching objectives [2, 36, 37] conditioned on the same temporal latent, ensuring alignment while preserving modality-specific optimizations. 3) A chain-of-forward training strategy addresses error accumulation and content drift in autoregressive loops." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.426 + ], + "angle": 0, + "content": "Epona offers several additional advantages. 1) Long-horizon generation. Our autoregressive diffusion model can achieve long-time generation over up to 2 minutes, significantly outperforming existing world models. 2) Real-time trajectory planning. The separate multi-modal generation architecture enables to perform trajectory planning solely while video prediction is deactivated, significantly reducing the inference FLOPS. This enables high-quality and even real-time trajectory planning, achieving rates of up to \\(20\\mathrm{Hz}\\). 3) Visual detail preservation. Our autoregressive formulation adopts continuous visual tokenizer instead of discrete ones, thus preserving rich scene details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.427, + 0.905, + 0.683 + ], + "angle": 0, + "content": "Extensive experiments demonstrate the effectiveness and superiority of our world model. For video generation, our model achieves state-of-the-art FVD [52] on the NuScenes [6] benchmark, surpassing the best-performing Vista [18] by \\(7.4\\%\\), while extending generation length from 15 seconds to over 2 minutes (600 frames). Thanks to joint supervision with trajectory prediction, Epona allows action control to simulate diverse driving scenarios, as shown in Fig. 1 (B). For motion planning, our method outperforms strong end-to-end planners on NAVSIM [12] without perception inputs (e.g., 3D boxes/lanes). Notably, we observe that Epona learns essential traffic world knowledge (e.g., stop driving at red light) purely from self-supervised future prediction tasks, as shown in Fig. 1 (C). This suggests that our world model can implicitly learn real-world driving dynamics, making it a promising direction for next-generation autonomous driving systems." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.701, + 0.655, + 0.716 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.727, + 0.851, + 0.743 + ], + "angle": 0, + "content": "2.1. World Models for Autonomous Driving" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Constructing real-world driving world models have drawn considerable attention in recent years, among which vision-centric approaches gain prominence due to their superior sensor flexibility, data accessibility, and more human-like representation forms. Early efforts primarily focused on adapting pre-trained diffusion models (e.g., Stable Diffusion [3, 47]) to driving scenarios through fine-tuning. However, these methods either lacked critical planning modules [16, 18, 19] or were limited to low-resolution, short-term generation [17, 38, 59, 60, 64, 65], making them un" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "27221" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.089, + 0.902, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.258, + 0.908, + 0.316 + ], + "angle": 0, + "content": "Figure 2. Overview of Epona. Our world model utilizes a multimodal spatiotemporal transformer to process the historical context of the first \\(T\\) frames and employs a next-frame prediction DiT to generate the frame at \\(T + 1\\) and a trajectory planning DiT to forecast the future \\(N\\)-frame pose trajectory. By adopting a chain-of-forward strategy, our approach enables high-quality and long-horizon video generation with an autoregressive manner." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.323, + 0.485, + 0.565 + ], + "angle": 0, + "content": "suitable for consistent long-range prediction and real-time planning. Recent work [11, 24, 26, 63, 70] explored harnessing GPT-like architecture to unify visual and action modeling and achieved long-range autoregressive generation. Yet, these methods require encoding images and trajectories into discrete tokens [14, 29, 54], which significantly degrades visual fidelity and trajectory precision. Similarly, while the newly released Cosmos [1] foundation model can serve as a driving world model, it does not introduce a new framework, facing the same limitations as the previous methods. In addition, its large parameter count and high computational demands limit its practicality. In contrast, we propose a novel autoregressive diffusion world model framework for autonomous driving, enabling long-range autoregressive generation in continuous visual and trajectory representations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.576, + 0.308, + 0.592 + ], + "angle": 0, + "content": "2.2. Long Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.6, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Long-term prediction is not only a key challenge for current video generation models but also a crucial capability for robust world models, as it reflects the model's ability to learn consistent environment dynamics and accurately simulate real-world temporal progression [13]. Since original video diffusion models (e.g. SVD [3]) are limited to fixed-length short clips generation, prior methods have explored extrapolating video length by noise rescheduling [43], overlapped generation [57, 58] or hierarchical generation [66]. However, these techniques fail to resolve inherent model constraints, often resulting in inconsistencies and abrupt visual changes in long videos. Autoregressive approaches [11, 56, 61] naturally support variable-length generation but suffer from quality degradation due to domain shift between teacher-forcing training and error accumulation in sampling. GameNGen [53] and DrivingWorld [26] introduce noise augmentation and random token dropout during training to alleviate the problem by simulating the error in sampling, but still limited to specific model architectures. We propose a general chain-of-forward strat" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.475 + ], + "angle": 0, + "content": "egy allowing the model to directly learn inference errors during training, effectively reducing autoregressive drift. Meanwhile, recent works such as Diffusion Forcing [8, 49] and FIFO-Diffusion [31] explore integrating autoregressive generation in video diffusion by adjusting frame-wise noise levels and leveraging causal network designs. Our model adopts a similar causal temporal modeling strategy but redefines the architecture into a two-stage end-to-end framework, allowing joint generation of motion plans and next-frame images." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.605, + 0.503 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.74 + ], + "angle": 0, + "content": "In this section, we formally present the model framework and training techniques of Epona. We begin with preliminaries on diffusion models in Sec. 3.1 and discuss our world model formulation design insights in Sec. 3.2. Then we introduce our proposed autoregressive diffusion world model framework in Sec. 3.3, including three dedicated modules: a multimodal spatiotemporal transformer to capture historical context, a trajectory planning DiT to generate future 3-seconds trajectories, and a next-frame prediction DiT to generate the next-frame images. To mitigate autoregressive drift and enable long-horizon video generation, we propose a simple yet effective chain-of-forward training strategy, detailed in Sec.3.4. Additionally, to enhance video quality, we introduce a temporal-aware DCAE decoder in Sec.3.5. An overview of our method is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.644, + 0.765 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.77, + 0.908, + 0.832 + ], + "angle": 0, + "content": "Diffusion models [23, 50] are a family of powerful generative models that transform noise samples \\( x_{(1)} \\) drawn from a prior distribution \\( p_1 = \\mathcal{N}(\\mathbf{0},\\mathbf{I}) \\) into data samples \\( x_{(0)} \\) from the target distribution in terms of a differentiable equation:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.843, + 0.907, + 0.861 + ], + "angle": 0, + "content": "\\[\nd x _ {(t)} = v _ {\\Theta} \\left(x _ {(t)}, t\\right) d t, t \\in [ 0, 1 ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.908, + 0.902 + ], + "angle": 0, + "content": "where velocity \\( v \\) is parametrized by a neural network \\( \\Theta \\). Rectified flow [2, 36, 37] proposes to define a straight prob-" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27222" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.094, + 0.482, + 0.234 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.246, + 0.485, + 0.332 + ], + "angle": 0, + "content": "Figure 3. Comparison of Different World Modeling Formulation. Up: Conventional autoregressive pipeline quantizes continuous images into discrete tokens and perform next-token prediction iteratively. Middle: The video-diffusion-based methods generate future \\( n \\) frames simultaneously. Down: Our method autoregressively predicts fine-grained future frames in continuous space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.342, + 0.484, + 0.388 + ], + "angle": 0, + "content": "ability path between \\( p_0 \\) and \\( p_1 \\) to improve training and sampling efficiency and optimize the network \\( \\Theta \\) using a velocity prediction loss:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.395, + 0.484, + 0.414 + ], + "angle": 0, + "content": "\\[\nx _ {(t)} = (1 - t) x _ {(0)} + t \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.433, + 0.482, + 0.453 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {R F} = \\mathbb {E} _ {x _ {(0)}, \\epsilon , t} \\left[ \\| v _ {\\Theta} (x _ {(t)}, t) - (x _ {(0)} - \\epsilon) \\| ^ {2} \\right], \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.455, + 0.484, + 0.576 + ], + "angle": 0, + "content": "which has proven to be an effective and scalable approach in state-of-the-art image and video generation models [15, 27, 33]. In Epona, we adopt the diffusion model and rectified flow objective for both next-frame image and trajectory generation. Particularly, to enhance efficiency, we encode images into compact latents using a pre-trained deep compression encoder [9] and adopt a latent diffusion model [47] for image synthesis." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.584, + 0.43, + 0.6 + ], + "angle": 0, + "content": "3.2. Reformulation of World Model Designs" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.605, + 0.484, + 0.801 + ], + "angle": 0, + "content": "In this section, we discuss different world model formulation design choices, as illustrated in Fig. 3. Given a sequence of previous front-view camera observations \\(\\{\\mathbf{O}_t\\}_{t=1}^T\\) and the corresponding driving trajectory \\(\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T\\), the goal of driving world moels is to predict future driving dynamics based on historical context. Here, each driving action \\(\\mathbf{a}_{t_1 \\to t_2} := (\\Delta \\theta_{t_1 \\to t_2}, \\Delta x_{t_1 \\to t_2}, \\Delta y_{t_1 \\to t_2}) \\in \\mathbb{R}^3\\) represents the vehicle's motion from \\(t_1\\) to \\(t_2\\), where \\(\\Delta \\theta\\) denotes the orientation change, and \\((\\Delta x, \\Delta y)\\) specify the relative displacement in the ego-coordinate frame. For consistency, we define \\(\\mathbf{a}_{0 \\to 1} = (0,0,0)\\). Existing methods tackle this problem by formulating world modeling in the following two ways:" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.802, + 0.484, + 0.877 + ], + "angle": 0, + "content": "Video diffusion-based world models. Current leading driving world models, like Vista [18], formulates world modeling in the form of video diffusion models [3], which jointly capture the global spatiotemporal distribution of both past and a fixed-length future," + }, + { + "type": "equation", + "bbox": [ + 0.184, + 0.884, + 0.389, + 0.904 + ], + "angle": 0, + "content": "\\[\np \\left(\\{\\mathbf {O} _ {T + i} \\} _ {i = 1} ^ {n}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.152 + ], + "angle": 0, + "content": "This formulation disrupts the causal temporal structure between historical observations and future predictions, limiting its ability to model real-world progressive dynamics and generate flexible-length long-term videos." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.153, + 0.907, + 0.228 + ], + "angle": 0, + "content": "GPT-based world models. Alternatively, autoregressive transformer-based world models [11, 24, 26] discretize image observations into token sequences \\(\\mathbf{O}_t = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots ,\\mathbf{t}_L]\\) and model the conditional image distribution as a token-by-token prediction," + }, + { + "type": "equation", + "bbox": [ + 0.613, + 0.239, + 0.808, + 0.281 + ], + "angle": 0, + "content": "\\[\n\\prod_ {i = 1} ^ {L} p (\\mathbf {t} _ {i} \\mid \\mathbf {t} _ {< i}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.906, + 0.336 + ], + "angle": 0, + "content": "However, this independent token modeling weakens spatial correlations and the quantization process distorts high-frequency details, leading to degraded generation quality." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.906, + 0.413 + ], + "angle": 0, + "content": "Our approach. In contrast, we formulate world modeling as a sequential future prediction process in the temporal domain. Specifically, given past driving observations \\(\\{\\mathbf{O}_t\\}_{t=1}^T\\) and the driving trajectory \\(\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T\\), our model predicts both a policy for future trajectory planning," + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.422, + 0.843, + 0.442 + ], + "angle": 0, + "content": "\\[\n\\pi \\left(\\{\\mathbf {a} _ {T \\rightarrow T + i} \\} _ {i = 1} ^ {n} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.452, + 0.906, + 0.481 + ], + "angle": 0, + "content": "and a conditional distribution over the next-frame camera observation as a whole," + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.493, + 0.847, + 0.513 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {O} _ {T + 1} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}, \\mathbf {a} _ {T \\rightarrow T + 1}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.906, + 0.659 + ], + "angle": 0, + "content": "The next-frame prediction is conditioned on either a model-predicted action or an externally provided action \\(\\mathbf{a}_{T\\rightarrow T + 1}\\). By decoupling causal temporal modeling from fine-grained future prediction, our model can generate flexible-length long videos autoregressively in continuous representations. Moreover, by factorizing trajectory planning from visual generation, our model can seamlessly serve as a real-time motion planner, bridging a critical gap between current driving world models and end-to-end motion planners." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.906, + 0.684 + ], + "angle": 0, + "content": "3.3. Epona: Autoregressive Diffusion World Model" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Based on the reformulated world modeling design, we propose Epona, an autoregressive diffusion world model for autonomous driving. Our framework consists of three key components. First, a Multimodal Spatiotemporal Transformer (MST) encodes historical context \\(\\{\\mathbf{O}_t,\\mathbf{a}_t\\}_{t = 1}^T\\) into a compact latent representation, effectively capturing environmental context and driving dynamics. Then, based on the historical latents, we employ two specialized diffusion transformers to predict fine-grained future details, including a tiny Trajectory Planning Transformer (TrajDiT) that models the policy \\(\\pi\\) for trajectory planning, and a Next-frame Prediction Transformer (VisDiT) that models the visual distribution \\(p\\) for future image generation. This modular design enables a range of autonomous driving applications." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27223" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.137 + ], + "angle": 0, + "content": "For instance, MST and VisDiT can be used independently for controllable driving simulations, while MST and TrajDiT facilitate real-time motion planning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.348 + ], + "angle": 0, + "content": "Multimodal Spatiotemporal Transformer (MST). Given the encoded past driving scenes \\(\\{\\mathbf{Z}_t\\}_{t=1}^T\\) and trajectory \\(\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T\\), we introduce a multimodal spatiotemporal transformer to effectively integrate temporal dynamics and multimodal information from historical context for future prediction. Inspired by prior work in video generation and world modeling [4, 26, 39], our approach employs interleaved multimodal spatial attention layers and causal temporal attention layers. This design progressively incorporates historical information into a compact latent representation while significantly reducing memory consumption compared to full-sequence attention. Additionally, this design naturally supports historical contexts of variable length." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.469 + ], + "angle": 0, + "content": "Specifically, we first project the flattened visual latent patches \\(\\mathbf{Z} \\in \\mathbb{R}^{B \\times T \\times L \\times C}\\) and action sequences \\(\\mathbf{a} \\in \\mathbb{R}^{B \\times T \\times 3}\\) into an embedding space. Then we concatenate them along the spatial dimension and add temporal positional embeddings to obtain the latent embedding sequence \\(\\mathbf{E} \\in \\mathbb{R}^{B \\times T \\times (L + 3) \\times D}\\). This sequence is processed through interleaved multimodal spatiotemporal layers as follows (using einops [46] notation):" + }, + { + "type": "equation", + "bbox": [ + 0.094, + 0.476, + 0.478, + 0.548 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b t) l c \\rightarrow (b l) t c)) \\\\ \\mathbf {E} \\leftarrow \\text {C a u s a l T e m p o r a l L a y e r} (\\mathbf {E}, \\text {C a u s a l M a x}) \\\\ \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b l) t c \\rightarrow (b t) l c)) \\\\ \\mathbf {E} \\leftarrow \\operatorname {M u l t i m o d a l S p a t i a l L a y e r} (\\mathbf {E}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\(B\\) is the batch size, \\(T\\) is the number of conditioning frames, \\(L = H \\times W\\) is the number of flattened latents in an image, \\(C\\) is the image latent channel dimension, \\(D\\) is the embedding dimension, and CausalMask is the triangular causal attention mask. Finally, we use the latent embedding of the last frame, \\(\\mathbf{F} \\in \\mathbb{R}^{B \\times (L + 3) \\times D}\\), as the compact latent representation for the next-stage prediction. After training, this embedding encapsulates the historical context \\(\\{\\mathbf{O}_t, \\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Trajectory Planning Diffusion Transformer (TrajDiT). TrajDiT predicts future trajectories using a tiny diffusion transformer. Following the DiT frameworks in most advanced open-source text-to-image and video generation models [27, 33], we adopt a Dual-Single-Stream architecture. In the dual-stream phase, the historical latent representation \\(\\mathbf{F}\\) and trajectory data are processed independently through transformer blocks, with only attention operations linking them. In the single-stream phase, they are concatenated to pass through subsequent transformer blocks for effective information fusion. Detailed architecture can be found in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "During training, we add noise to the target trajectories \\(\\overline{\\mathbf{a}}\\in \\mathbb{R}^{B\\times N\\times 3}\\) using Eq.2. The model then predicts velocity" + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.082, + 0.885, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.224, + 0.905, + 0.252 + ], + "angle": 0, + "content": "Figure 4. Concept illustration of our training process. Here \\( x \\) can be either image latents or trajectories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.267, + 0.905, + 0.296 + ], + "angle": 0, + "content": "\\(v_{\\text{traj}}\\) conditioned on \\(\\mathbf{F}\\), where \\(N\\) is the planning horizon. We optimize using the rectified flow loss:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.309, + 0.905, + 0.328 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t r a j} = \\mathbb {E} _ {\\overline {{\\mathbf {a}}}, \\epsilon , t} \\left[ \\| v _ {t r a j} (\\overline {{\\mathbf {a}}} _ {(t)}, t) - (\\overline {{\\mathbf {a}}} - \\epsilon) \\| ^ {2} \\right]. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.34, + 0.905, + 0.432 + ], + "angle": 0, + "content": "For inference, random Gaussian noise is iteratively denoised conditioned on \\(\\mathbf{F}\\) to generate future trajectory plans. Next-frame Prediction Diffusion Transformer (VisDiT). VisDiT has a similar architecture as TrajDiT, with an additional modulation [41] branch for action control \\(\\mathbf{a}_{T\\rightarrow T + 1}\\). We also use the flow loss for visual supervision:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.443, + 0.905, + 0.476 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v i s} = \\mathbb {E} _ {\\mathbf {Z} _ {T + 1}, \\epsilon , t} \\left[ \\| v _ {v i s} \\left(\\mathbf {Z} _ {T + 1 (t)}, t\\right) - \\left(\\mathbf {Z} _ {T + 1} - \\epsilon\\right) \\| ^ {2} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.476, + 0.905, + 0.504 + ], + "angle": 0, + "content": "Together, the total loss jointly optimizes the entire world model:" + }, + { + "type": "equation", + "bbox": [ + 0.645, + 0.508, + 0.905, + 0.526 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {t r a j}} + \\mathcal {L} _ {\\text {v i s}}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.534, + 0.905, + 0.596 + ], + "angle": 0, + "content": "During inference, VisDiT denoises \\(\\hat{\\mathbf{Z}}_{T + 1}\\), conditioned on \\(\\mathbf{F}\\) and the action either predicted by TrajDiT or provided by user. The latents are then decoded using the DCAE decoder to generate the next-frame image \\(\\hat{\\mathbf{O}}_{T + 1}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.764, + 0.623 + ], + "angle": 0, + "content": "3.4. Chain-of-Forward Training" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.765 + ], + "angle": 0, + "content": "With our proposed autoregressive diffusion world model, we can autoregressively generate future videos frame by frame. However, long-term generation suffers from a longstanding autoregressive drift problem [53]: during training, the model predicts the next frame using ground-truth historical context, whereas during inference, it relies on its own past predictions. This domain gap between teacher-forcing training and autoregressive sampling leads to error accumulation and rapid quality degradation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.766, + 0.905, + 0.872 + ], + "angle": 0, + "content": "To mitigate this, we introduce a chain-of-forward training strategy. Periodically, we perform multiple forward passes using self-predicted frames to enhance the model's robustness to inference noise (see Fig. 4). Notably, to ensure training efficiency, instead of sampling next-frame latents from pure noise, we leverage the model-predicted velocity \\( v_{\\Theta} \\) to estimate the denoised latents in one step:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.886, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {(0)} = x _ {(t)} + t v _ {\\Theta} (x _ {(t)}, t) \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27224" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.086, + 0.907, + 0.128 + ], + "angle": 0, + "content": "Table 1. Comparisons of generated videos on the NuScenes [6] validation set. Our model achieves state-of-the-art FVD score compared to existing driving world models, while extending the video length to over two minutes. *The max duration number indicates the horizon that produces plausible results, following existing methods." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.131, + 0.907, + 0.187 + ], + "angle": 0, + "content": "
MetricDriveGAN [32]DriveDreamer [59]WoVoGen [38]Drive-WM [60]GenAD (OpenDV) [65]Vista [18]DrivingWorld [26]Ours
FID ↓73.452.627.615.815.46.97.47.5
FVD ↓502.3452.0417.7122.7184.089.490.982.8
Max Duration / Frames*N/A4s / 482.5s / 58s / 164s / 815s / 15040s / 400120s / 600
" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.205, + 0.892, + 0.329 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.217, + 0.336, + 0.778, + 0.35 + ], + "angle": 0, + "content": "Figure 5. Qualitative Comparison between Vista [18] and Epona. Zoom in for better views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.359, + 0.483, + 0.435 + ], + "angle": 0, + "content": "The estimated \\(\\hat{x}_{(0)}\\), along with previous conditioned frames, is then used in the next forward pass to autoregressively generate subsequent frames. This process simulates prediction noise, helping the model adapt to deviations and improving long-term video generation quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.379, + 0.459 + ], + "angle": 0, + "content": "3.5. Temporal-aware DCAE Decoder" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.464, + 0.483, + 0.554 + ], + "angle": 0, + "content": "Unlike conventional autoencoders that downsample images by a factor of 8, DCAE [9] progressively increases this to 32, reducing latent tokens by \\(16 \\times\\). In our world model, we adopt DCAE for image encoding to improve training efficiency and reduce memory usage, enabling conditioning on longer historical contexts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.556, + 0.483, + 0.707 + ], + "angle": 0, + "content": "However, as an image autoencoder, DCAE lacks temporal interactions, causing flickering when decoding video frame by frame, which degrades visual quality. To address this, we propose a temporal-aware DCAE to enhance interframe consistency. Specifically, to maximize pretrained parameters while minimizing architectural changes, we introduce spatiotemporal self-attention layers before the DCAE decoder while keeping the encoder fixed during fine-tuning. This facilitates multi-frame interactions, greatly improving temporal consistency in generated videos." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.719, + 0.215, + 0.735 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.307, + 0.759 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "World Model. Our Epona consists of \\(2.5\\,B\\) parameters, including a 12-layer multimodal spatiotemporal transformer with \\(1.3\\,B\\) parameters, a 12-layer next-frame prediction diffusion transformer with \\(1.2\\,B\\) parameters, and a 2-layer trajectory planning diffusion transformer with \\(50\\,M\\) parameters. It is trained on publicly available videos from the NuPlan dataset [7] and 700 scenes from the NuScenes dataset [6] from scratch, all images are resized to \\(512\\times 1024\\). We utilize the rectified flow [37] objective for both video" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.357, + 0.907, + 0.398 + ], + "angle": 0, + "content": "Table 2. Inference speed. We evaluate our inference speed for generating a 3-second trajectory and a \\(512 \\times 1024\\) image per module on a single NVIDIA 4090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.565, + 0.41, + 0.853, + 0.461 + ], + "angle": 0, + "content": "
DiT sampling stepsMSTTrajDiTVisDiT
10~0.02s~0.03s~0.3s
100~0.02s~0.3s~2s
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.655 + ], + "angle": 0, + "content": "generation and trajectory planning tasks, training the entire model in an end-to-end manner. The training was conducted on 48 NVIDIA A100 GPUs for nearly two weeks, with a total of \\(600\\mathrm{k}\\) iterations and a batch size of 96. During training, we apply Chain-of-Forward every 10 steps, each time performing three forward passes. We use the AdamW optimizer with a learning rate of \\(1\\times 10^{-4}\\) and set the weight decay to \\(5\\times 10^{-2}\\). For inference, we report our speed for each module on a single NVIDIA 4090 GPU in Table 2. In all our experiments, we set our DiT sampling step to 100. Notice that with MST and TrajDiT, our world model can seamlessly serve as a real-time motion planner." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.657, + 0.906, + 0.793 + ], + "angle": 0, + "content": "**Evaluations on video generation.** We employ 1628 video clips from the NuPlan test set [7] and 1646 video clips from the NuScenes validation dataset [6] for performance evaluation, respectively. During the test, our world model is conditioned on 10 consecutive past frames to generate the subsequent frame and repeat the process autoregressively to synthesize future video frames. We use the Frechet Video Distance (FVD) [52] and the Frechet Inception Distance (FID) [22] to evaluate the quality of the generated videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Evaluations on trajectory planning. We evaluate trajectory planning using the NuScenes benchmark [6] and the NAVSIM benchmark [12]. For the NuScenes, we use L2 error and collision rate as the evaluation metrics following the existing works [25, 48, 70] to evaluate the planning performance. L2 error measures the L2 distance between the predicted and ground truth trajectories, while the collision" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27225" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.171, + 0.088, + 0.825, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.097, + 0.304, + 0.898, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Trajectory-controlled video generation. Our world model can generate controllable videos based on predefined trajectories." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.484, + 0.386 + ], + "angle": 0, + "content": "rate measures the frequency of predicted trajectory intersections with objects. The NAVSIM benchmark assesses performance using the predictive driver model score (PDMS), derived from five factors, as shown in Table 4." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.396, + 0.373, + 0.411 + ], + "angle": 0, + "content": "4.2. Evaluation of Video Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.418, + 0.484, + 0.612 + ], + "angle": 0, + "content": "Quantitative Comparison of Generated Videos. We present a quantitative comparison with existing methods on the NuScenes dataset [6] in Table 1. Since most methods are not publicly available, we compare with the reported results from their respective papers. Notably, the existing methods (e.g., Vista [18]) fine-tune video diffusion models pretrained on large-scale datasets, while our world model, including next-frame DiT, is trained from scratch. As shown in Table 1, our generated videos achieve state-of-the-art FVD scores, indicating smoother and more realistic video generation quality. Moreover, our world model can generate significantly longer video frames compared to existing approaches as shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Qualitative Comparison of Generated Videos. We provide a qualitative comparison with the state-of-the-art open-source driving world model, Vista [18]. Since Vista is a 25-frame fixed-length video diffusion model, we perform rollout to generate longer videos as illustrated in their paper. As shown in Fig. 5, our Epona generates consistent long-horizon driving scenes with high-fidelity visuals and detailed structures and vehicles." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.825 + ], + "angle": 0, + "content": "Trajectory-controlled Video Generation. Fig. 6 illustrates the pose controllability of our model. Given the predefined pose trajectory, the different condition frames can generate future frames that conform to the corresponding motion path, which is crucial for obtaining autonomous driving videos under extreme scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Extra Long-range Video Generation. Epona combines the strengths of autoregressive and diffusion models, facilitating the generation of high-quality, long-duration videos conditioned on input frames. As shown in Fig.5 and Fig.7, our model can autoregressively generate minute-long driv-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.907, + 0.372 + ], + "angle": 0, + "content": "ing videos with high fidelity and consistency, without noticeable drift. More long-term generation videos are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.38, + 0.813, + 0.396 + ], + "angle": 0, + "content": "4.3. Evaluation of Trajectory Planning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.66 + ], + "angle": 0, + "content": "For the NuScenes benchmark [6], we compare our world model with several existing methods, as shown in Table 3. Although our model does not achieve the best results, it attains competitive performance without any additional supervision. It is worth noting that incorporating more supervision typically leads to better performance but cost expensive annotations. Additionally, similar to Doe-1 [70], our model only utilizes the front camera, whereas other methods rely on multi-view inputs for planning. As shown in Table 3, our approach can generate reasonable trajectories while achieving the lowest collision rate for a 1-second horizon, which is crucial for long-term realistic video predictions. For the more challenging NAVSIM benchmark [12], as shown in Table 4, our method achieves state-of-the-art results in overall PDMS when conditioned on the past 2 seconds of observations to predict 4-second future trajectories, showing the strong motion planning capability." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.667, + 0.684 + ], + "angle": 0, + "content": "4.4. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.885 + ], + "angle": 0, + "content": "Effect of Shared Latent for Multi-modal Joint Prediction. To assess the benefit of jointly modeling scene and trajectory via a shared latent representation, we conduct an ablation by disabling video prediction and training the model solely for trajectory prediction. This variant is evaluated on the NAVISIM test set. As shown in Tab. 5, removing video prediction leads to a noticeable drop in planning performance. This result highlights the advantage of the shared latent \\(\\mathbf{F}\\), which encourages the world model to better capture complex driving dynamics by leveraging visual signals. This ablation confirms that coupling video and trajectory prediction within a unified latent space in world models significantly benefits downstream planning tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Effect of Chain-of-Forward Training. To evaluate the" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27226" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.082, + 0.5, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.082, + 0.897, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.206, + 0.908, + 0.248 + ], + "angle": 0, + "content": "Figure 7. Qualitative Comparison between long videos generated by models w/ and w/o Chain-of-Forward training. Left: Visual quality deteriorates rapidly after 10–20 seconds. Right: The same driving scenes with Chain-of-Forward training maintain high visual quality, generating minute-long videos without significant degradation. Zoom in for better views." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.252, + 0.908, + 0.281 + ], + "angle": 0, + "content": "Table 3. End-to-end motion planning performance on the NuScenes [6] dataset. Note that our model achieves a low collision rate, demonstrating its understanding of basic traffic rules via simple next-frame prediction. * represents only using the front camera as input." + }, + { + "type": "table", + "bbox": [ + 0.212, + 0.282, + 0.782, + 0.416 + ], + "angle": 0, + "content": "
MethodInputAuxiliary SupervisionL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
ST-P3 [25]CameraMap & Box & Depth1.332.112.902.110.230.621.270.71
UniAD [48]CameraMap & Box & Motion & Tracklets & Occ0.480.961.651.030.050.170.710.31
OccNet [51]Camera3D-Occ & Map & Box1.292.132.992.140.210.591.370.72
OccWorld [68]Camera3D-Occ0.521.272.411.400.120.402.080.87
VAD-Tiny [30]CameraMap & Box & Motion0.601.232.061.300.310.531.330.72
VAD-Base [30]CameraMap & Box & Motion0.541.151.981.220.040.391.170.53
GenAD [69]CameraMap & Box & Motion0.360.831.550.910.060.231.000.43
Doe-1 [70]Camera*QA0.501.182.111.260.040.371.190.53
OursCamera*None0.611.171.981.250.010.220.850.36
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.507 + ], + "angle": 0, + "content": "Table 4. End-to-end motion planning performance on the NAVSIM [12] test set. NC: no at-fault collision. DAC: drivable area compliance. TTC: time-to-collision. Conf.: comfort. EP: ego progress. PDMS: the predictive driver model score. LAW[35] is in the perception-free setting. Our world model outperforms strong end-to-end planners in the overall PDMS score." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.518, + 0.481, + 0.617 + ], + "angle": 0, + "content": "
MethodInputNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Human/10010010099.987.594.8
UniAD[48]Camera97.891.992.910078.883.4
PARA-Drive[62]Camera97.992.493.099.879.384.6
LAW[35]Camera96.495.488.799.981.784.6
TransFuser[42]Camera & Lidar97.792.892.810079.284.0
DRAMA[67]Camera & Lidar98.093.194.810080.185.5
VADv2[10]Camera & Lidar97.289.191.910076.080.9
OursCamera97.995.193.899.980.486.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.626, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Table 5. Comparison of planning results on the NAVSIM test set. Jointly predicting the next scene using a shared latent significantly improves planning performance." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.671, + 0.481, + 0.713 + ], + "angle": 0, + "content": "
MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Ours w/o Joint Training94.589.788.199.974.778.1
Ours97.995.193.899.980.486.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "impact of our chain-of-forward strategy on model performance, we conduct an ablation study comparing results with and without this strategy. Given that our model iteratively generates the next frame, the chain-of-forward approach simulates potential inference errors during training, thereby enhancing the model's robustness. As shown in Fig. 7 and Fig. 8, as the model autoregressively generates longer sequences, the gap in visual quality and FID score between models with and without the chain-of-forward strategy becomes increasingly significant, validating its effectiveness in long-term video generation. More ablations" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.437, + 0.871, + 0.602 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.617, + 0.907, + 0.659 + ], + "angle": 0, + "content": "Figure 8. Effect of Chain-of-Forward training. FID comparison in NuPlan test set between models w/ and w/o Chain-of-Forward training strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.667, + 0.807, + 0.683 + ], + "angle": 0, + "content": "can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.634, + 0.71 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.72, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We have presented Epona, an autoregressive diffusion world model for autonomous driving that jointly predicts high-fidelity future trajectories and driving scenes based on historical driving context. Thanks to our proposed decoupled spatiotemporal modeling and asynchronous multimodal generation strategies, our model achieves high-quality and long-term prediction. In addition, our model could serves as a real-time motion planner via performing trajectory planning. We have demonstrated that our approach significantly advances the state of the art in driving world models, uncovering the large potential for building next-generation autonomous driving systems." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27227" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.213 + ], + "angle": 0, + "content": "[2] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.285, + 0.483, + 0.34 + ], + "angle": 0, + "content": "[4] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, pages 22563-22575, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.342, + 0.484, + 0.422 + ], + "angle": 0, + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.485, + 0.508 + ], + "angle": 0, + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.509, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 6, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.579, + 0.484, + 0.633 + ], + "angle": 0, + "content": "[8] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.484, + 0.69 + ], + "angle": 0, + "content": "[9] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.692, + 0.483, + 0.759 + ], + "angle": 0, + "content": "[10] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.761, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[11] Yuntao Chen, Yuqi Wang, and Zhaoxiang Zhang. Driving-gpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers. arXiv preprint arXiv:2412.18607, 2024. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[12] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.485, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models. arXiv preprint arXiv:2411.14499, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.192 + ], + "angle": 0, + "content": "[14] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[16] Ruiyuan Gao, Kai Chen, Bo Xiao, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrivedit: High-resolution long video generation for autonomous driving with adaptive control. arXiv preprint arXiv:2411.13807, 2024. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[17] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3d geometry control. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[19] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. Dome: Taming diffusion model into high-fidelity controllable occupancy world model. ArXiv, abs/2410.10429, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.559 + ], + "angle": 0, + "content": "[20] Xi Guo, Chenjing Ding, Haoxuan Dou, Xin Zhang, Weixuan Tang, and Wei Wu. Infinitydrive: Breaking time limits in driving world models, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.905, + 0.589 + ], + "angle": 0, + "content": "[21] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.718, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[24] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3, 4, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[25] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. Drivingworld: Constructingworld model for autonomous driving via" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27228" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "video gpt. arXiv preprint arXiv:2412.19505, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.482, + 0.149 + ], + "angle": 0, + "content": "[27] Tencent Hunyuan. Hunyuanvideo: A systematic framework for large video generative models, 2024. 4, 5, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[28] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. Adriver-i: A general world model for autonomous driving, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[29] Mingkai Jia, Wei Yin, Xiaotao Hu, Jiaxin Guo, Xiaoyang Guo, Qian Zhang, Xiao-Xiao Long, and Ping Tan. Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization. arXiv preprint arXiv:2507.07997, 2025.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[30] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.362, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[31] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. In NeurIPS, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[32] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.476, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[33] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 4, 5, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.505, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[34] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1):1-62, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.547, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[35] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.591, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[36] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.745 + ], + "angle": 0, + "content": "[38] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. WoVoGen: World Volume-Aware Diffusion for Controllable Multi-Camera Driving Scene Generation. arXiv preprint arXiv:2312.02934, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[40] Yiyang Ma, Xingchao Liu, Xi aokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Xingkai Yu, Liang Zhao, Yisong Wang, Jiaying Liu, and Chong Ruan. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. ArXiv, abs/2411.07975, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "national conference on computer vision, pages 4195-4205, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[42] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.905, + 0.208 + ], + "angle": 0, + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.21, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.295 + ], + "angle": 0, + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.298, + 0.905, + 0.339 + ], + "angle": 0, + "content": "[46] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.341, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[48] Xiaogang Shi, Bin Cui, Gillian Dobbie, and Beng Chin Ooi. Uniad: A unified ad hoc data processing system. ACM Transactions on Database Systems (TODS), 42(1):1-42, 2016. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.443, + 0.905, + 0.483 + ], + "angle": 0, + "content": "[49] Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. History-guided video diffusion, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.486, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.612 + ], + "angle": 0, + "content": "[51] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.615, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[52] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.673, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[53] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.716, + 0.905, + 0.77 + ], + "angle": 0, + "content": "[54] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27229" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[57] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[58] Fu-Yun Wang, Zhaoyang Huang, Qiang Ma, Guanglu Song, Xudong Lu, Weikang Bian, Yijin Li, Yu Liu, and Hongsheng Li. Zola: Zero-shot creative long animation generation with short video model. In ECCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[59] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[60] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[61] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[62] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[63] Yanhao Wu, Haoyang Zhang, Tianwei Lin, Lichao Huang, Shujie Luo, Rui Wu, Congpei Qiu, Wei Ke, and Tong Zhang. Generating multimodal driving scenes via next-scene prediction. In CVPR, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[64] Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goaldriven flow matching for multimodal trajectories generation in end-to-end autonomous driving. ArXiv, abs/2503.05689, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.549, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[65] Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu, Jia Zeng, Ping Luo, Jun Zhang, Andreas Geiger, Yu Qiao, and Hongyang Li. Generalized Predictive Model for Autonomous Driving. In CVPR, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[66] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[67] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, and Marcelo H. Ang Jr au2. Drama: An efficient end-to-end motion planner for autonomous driving with mamba, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[68] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In European conference on computer vision, pages 55-72. Springer, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[69] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87-104. Springer, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[70] Wenzhao Zheng, Zetian Xia, Yuanhui Huang, Sicheng Zuo, Jie Zhou, and Jiwen Lu. Doe-1: Closed-loop autonomous driving with large world model. arXiv preprint arXiv:2412.09627, 2024. 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[71] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke S. Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. ArXiv, abs/2408.11039, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.218 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27230" + } + ] +] \ No newline at end of file diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_origin.pdf b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fd257934389e8846c81b263c3602f4c16c4a7fed --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/96492e1e-bfce-4329-bd89-949aa9fef588_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8ee07829be9c407bc49818469a2e047e2d8a28f2c475954eb0e85b3250d06392 +size 18540364 diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/full.md b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/full.md new file mode 100644 index 0000000000000000000000000000000000000000..0c8e085ae94b9cd5a2e2fcdeb40439133e6dec7c --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/full.md @@ -0,0 +1,346 @@ +# Epona: Autoregressive Diffusion World Model for Autonomous Driving + +Kaiwen Zhang $^{1,2*}$ Zhenyu Tang $^{1,3*}$ Xiaotao Hu $^{1,5}$ Xingang Pan $^{6}$ Xiaoyang Guo $^{1}$ Yuan Liu $^{5}$ Jingwei Huang $^{7}$ Li Yuan $^{3}$ Qian Zhang $^{1}$ Xiao-Xiao Long $^{4\dagger}$ Xun Cao $^{4}$ Wei Yin $^{1}$ $\S$ + +$^{1}$ Horizon Robotics $^{2}$ Tsinghua University $^{3}$ Shenzhen Gruadate School, Peking University $^{4}$ Nanjing University $^{5}$ Hong Kong University of Science and Technology $^{6}$ Nanyang Technological University $^{7}$ Tencent Hunyuan + +# (A) Consistent High-resolution Long Video Generation + +![](images/481b9130b93e2a87f7b0b4c6e84e88da18c65d77ec21c4c48018bad82e58ee85.jpg) + +![](images/03bb646c4b8ad1b03b1bd3b5bdab749a995466dd7e3cd649b73c18191b662dd6.jpg) + +# (B) Trajectory-controlled Video Generation + +![](images/b4f070a0c9b5dae181cea8b0c5d8fc80d65d71100299839a378d11ffb523da50.jpg) + +# (C)Traffic World Knowledge Understanding + +![](images/4c8ede8bf3c6201cfe1a2019d3c646a6a018d88dd30ff90e1dfc2a3d0b523d6e.jpg) +Figure 1. Versatile capabilities of Epona. Given historical driving context, our Epona can generate consistent minutes-long future driving scenes at high resolution (A). It can be controlled by diverse trajectories (B), and understand real-world traffic knowledge (C). In addition, our world model can predict future trajectories and serve as an end-to-end real-time motion planner (D). + +# (D) End-to-End Trajectory Planning + +![](images/6428349483a4095b34b420c3983bc4ec9c2c58dfd1f428356b7aea6b52bb1808.jpg) + +# Abstract + +Diffusion models have demonstrated exceptional visual quality in video generation, making them promising for autonomous driving world modeling. However, existing video diffusion-based world models struggle with flexible-length, long-horizon predictions and integrating trajectory planning. This is because conventional video diffusion models rely on global joint distribution modeling of fixed-length frame sequences rather than sequentially constructing localized distributions at each timestep. In this work, we propose Epona, an autoregressive diffusion world model that enables localized spatiotemporal distribution modeling + +through two key innovations: 1) Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation, and 2) Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework. Our architecture enables high-resolution, long-duration generation while introducing a novel chain-of-forward training strategy to address error accumulation in autoregressive loops. Experimental results demonstrate state-of-the-art performance with $7.4\%$ FVD improvement and minutes longer prediction duration compared to prior works. The learned world model further serves as a real-time motion planner, outperforming strong end-to-end planners on NAVSIM benchmarks. + +# 1. Introduction + +Recently, with the rapid development of video generation models, world models have attracted significant attention and emerged as a powerful paradigm for physical world simulations and autonomous decision-making [1, 13, 21, 34]. These foundation models enable agents to understand inherent world knowledge and predict future dynamics, making them particularly promising for autonomous driving. Unlike traditional separate perception-planning pipelines, which require extensive annotations and explicit supervision, generative driving world models [11, 16, 18, 24, 26, 59, 60, 63, 70] integrate visual scene understanding with future prediction in a self-supervised manner, offering a new solution toward end-to-end autonomous driving. + +Generative world models primarily fall into two categories: diffusion-based approaches and GPT-style autoregressive methods. The diffusion-based paradigm (e.g., Vista [18]), while achieving impressive visual fidelity through joint distribution modeling of fixed-length videos [3, 23, 47, 50], fundamentally suffers from its inability to model per-timestep local distributions. This limitation manifests in critical deficiencies: failure to support variable-length long-range prediction crucial for dynamic world simulation, and infeasible trajectory planning due to the lack of multimodal prediction mechanism. + +Conversely, GPT-style approaches [5, 44, 45, 55] address temporal flexibility through autoregressive next-token prediction (as seen in GAIA-1 [24]). However, the quantization and tokenization process significantly degrades visual quality and planning precision. Moreover, the causal nature of autoregressive transformers constrains them to predicting only the next action rather than planning long-horizon trajectories [11], limiting their potential to serve as end-to-end driving planners. Both paradigms reveal complementary shortcomings - diffusion models lack temporal decomposition while autoregressive transformers sacrifice continuous visual precision - highlighting the need for a unified framework that reconciles these divergent advantages for competent driving world modeling. + +We introduce Epona, an autoregressive diffusion world model that achieves high-resolution long-horizon video generation and accurate trajectory planning. Our core innovation comes from three key designs: 1) Decoupled spatiotemporal factorization. While existing video diffusion methods model joint spatial-temporal distributions of past and future frames, we assume their temporal latent modeling lacks explicit causality constraints, leading to error accumulation in long sequences. Epona addresses this through spacetime-disentangled processing: A GPT-style transformer with causal attention handles temporal dynamics in compressed latent space, while twin diffusion transformers separately handle spatial rendering and trajectory generation. 2) Asynchronous multi-modal genera + +tion. Building upon this foundation, we decouple trajectory planning from visual generation through parallel denoising processes. Two specialized DiTs [33, 41] asynchronously generate 3-second vehicle trajectories and the single next future frame. Both streams share flow-matching objectives [2, 36, 37] conditioned on the same temporal latent, ensuring alignment while preserving modality-specific optimizations. 3) A chain-of-forward training strategy addresses error accumulation and content drift in autoregressive loops. + +Epona offers several additional advantages. 1) Long-horizon generation. Our autoregressive diffusion model can achieve long-time generation over up to 2 minutes, significantly outperforming existing world models. 2) Real-time trajectory planning. The separate multi-modal generation architecture enables to perform trajectory planning solely while video prediction is deactivated, significantly reducing the inference FLOPS. This enables high-quality and even real-time trajectory planning, achieving rates of up to $20\mathrm{Hz}$ . 3) Visual detail preservation. Our autoregressive formulation adopts continuous visual tokenizer instead of discrete ones, thus preserving rich scene details. + +Extensive experiments demonstrate the effectiveness and superiority of our world model. For video generation, our model achieves state-of-the-art FVD [52] on the NuScenes [6] benchmark, surpassing the best-performing Vista [18] by $7.4\%$ , while extending generation length from 15 seconds to over 2 minutes (600 frames). Thanks to joint supervision with trajectory prediction, Epona allows action control to simulate diverse driving scenarios, as shown in Fig. 1 (B). For motion planning, our method outperforms strong end-to-end planners on NAVSIM [12] without perception inputs (e.g., 3D boxes/lanes). Notably, we observe that Epona learns essential traffic world knowledge (e.g., stop driving at red light) purely from self-supervised future prediction tasks, as shown in Fig. 1 (C). This suggests that our world model can implicitly learn real-world driving dynamics, making it a promising direction for next-generation autonomous driving systems. + +# 2. Related Work + +# 2.1. World Models for Autonomous Driving + +Constructing real-world driving world models have drawn considerable attention in recent years, among which vision-centric approaches gain prominence due to their superior sensor flexibility, data accessibility, and more human-like representation forms. Early efforts primarily focused on adapting pre-trained diffusion models (e.g., Stable Diffusion [3, 47]) to driving scenarios through fine-tuning. However, these methods either lacked critical planning modules [16, 18, 19] or were limited to low-resolution, short-term generation [17, 38, 59, 60, 64, 65], making them un + +![](images/4d2234653f65cc239e1d00117dc7c8a02bea139b8ac5aaf2222cce526d8efdba.jpg) +Figure 2. Overview of Epona. Our world model utilizes a multimodal spatiotemporal transformer to process the historical context of the first $T$ frames and employs a next-frame prediction DiT to generate the frame at $T + 1$ and a trajectory planning DiT to forecast the future $N$ -frame pose trajectory. By adopting a chain-of-forward strategy, our approach enables high-quality and long-horizon video generation with an autoregressive manner. + +suitable for consistent long-range prediction and real-time planning. Recent work [11, 24, 26, 63, 70] explored harnessing GPT-like architecture to unify visual and action modeling and achieved long-range autoregressive generation. Yet, these methods require encoding images and trajectories into discrete tokens [14, 29, 54], which significantly degrades visual fidelity and trajectory precision. Similarly, while the newly released Cosmos [1] foundation model can serve as a driving world model, it does not introduce a new framework, facing the same limitations as the previous methods. In addition, its large parameter count and high computational demands limit its practicality. In contrast, we propose a novel autoregressive diffusion world model framework for autonomous driving, enabling long-range autoregressive generation in continuous visual and trajectory representations. + +# 2.2. Long Video Generation + +Long-term prediction is not only a key challenge for current video generation models but also a crucial capability for robust world models, as it reflects the model's ability to learn consistent environment dynamics and accurately simulate real-world temporal progression [13]. Since original video diffusion models (e.g. SVD [3]) are limited to fixed-length short clips generation, prior methods have explored extrapolating video length by noise rescheduling [43], overlapped generation [57, 58] or hierarchical generation [66]. However, these techniques fail to resolve inherent model constraints, often resulting in inconsistencies and abrupt visual changes in long videos. Autoregressive approaches [11, 56, 61] naturally support variable-length generation but suffer from quality degradation due to domain shift between teacher-forcing training and error accumulation in sampling. GameNGen [53] and DrivingWorld [26] introduce noise augmentation and random token dropout during training to alleviate the problem by simulating the error in sampling, but still limited to specific model architectures. We propose a general chain-of-forward strat + +egy allowing the model to directly learn inference errors during training, effectively reducing autoregressive drift. Meanwhile, recent works such as Diffusion Forcing [8, 49] and FIFO-Diffusion [31] explore integrating autoregressive generation in video diffusion by adjusting frame-wise noise levels and leveraging causal network designs. Our model adopts a similar causal temporal modeling strategy but redefines the architecture into a two-stage end-to-end framework, allowing joint generation of motion plans and next-frame images. + +# 3. Method + +In this section, we formally present the model framework and training techniques of Epona. We begin with preliminaries on diffusion models in Sec. 3.1 and discuss our world model formulation design insights in Sec. 3.2. Then we introduce our proposed autoregressive diffusion world model framework in Sec. 3.3, including three dedicated modules: a multimodal spatiotemporal transformer to capture historical context, a trajectory planning DiT to generate future 3-seconds trajectories, and a next-frame prediction DiT to generate the next-frame images. To mitigate autoregressive drift and enable long-horizon video generation, we propose a simple yet effective chain-of-forward training strategy, detailed in Sec.3.4. Additionally, to enhance video quality, we introduce a temporal-aware DCAE decoder in Sec.3.5. An overview of our method is illustrated in Fig. 2. + +# 3.1. Preliminary + +Diffusion models [23, 50] are a family of powerful generative models that transform noise samples $x_{(1)}$ drawn from a prior distribution $p_1 = \mathcal{N}(\mathbf{0},\mathbf{I})$ into data samples $x_{(0)}$ from the target distribution in terms of a differentiable equation: + +$$ +d x _ {(t)} = v _ {\Theta} \left(x _ {(t)}, t\right) d t, t \in [ 0, 1 ], \tag {1} +$$ + +where velocity $v$ is parametrized by a neural network $\Theta$ . Rectified flow [2, 36, 37] proposes to define a straight prob- + +![](images/a6bdef404dd05a932485808fae60e82db343610c0f1412aa90acb47d3b9ac0d3.jpg) +Figure 3. Comparison of Different World Modeling Formulation. Up: Conventional autoregressive pipeline quantizes continuous images into discrete tokens and perform next-token prediction iteratively. Middle: The video-diffusion-based methods generate future $n$ frames simultaneously. Down: Our method autoregressively predicts fine-grained future frames in continuous space. + +ability path between $p_0$ and $p_1$ to improve training and sampling efficiency and optimize the network $\Theta$ using a velocity prediction loss: + +$$ +x _ {(t)} = (1 - t) x _ {(0)} + t \epsilon , \epsilon \sim \mathcal {N} (\mathbf {0}, \mathbf {I}), \tag {2} +$$ + +$$ +\mathcal {L} _ {R F} = \mathbb {E} _ {x _ {(0)}, \epsilon , t} \left[ \| v _ {\Theta} (x _ {(t)}, t) - (x _ {(0)} - \epsilon) \| ^ {2} \right], \tag {3} +$$ + +which has proven to be an effective and scalable approach in state-of-the-art image and video generation models [15, 27, 33]. In Epona, we adopt the diffusion model and rectified flow objective for both next-frame image and trajectory generation. Particularly, to enhance efficiency, we encode images into compact latents using a pre-trained deep compression encoder [9] and adopt a latent diffusion model [47] for image synthesis. + +# 3.2. Reformulation of World Model Designs + +In this section, we discuss different world model formulation design choices, as illustrated in Fig. 3. Given a sequence of previous front-view camera observations $\{\mathbf{O}_t\}_{t=1}^T$ and the corresponding driving trajectory $\{\mathbf{a}_{t-1 \to t}\}_{t=1}^T$ , the goal of driving world moels is to predict future driving dynamics based on historical context. Here, each driving action $\mathbf{a}_{t_1 \to t_2} := (\Delta \theta_{t_1 \to t_2}, \Delta x_{t_1 \to t_2}, \Delta y_{t_1 \to t_2}) \in \mathbb{R}^3$ represents the vehicle's motion from $t_1$ to $t_2$ , where $\Delta \theta$ denotes the orientation change, and $(\Delta x, \Delta y)$ specify the relative displacement in the ego-coordinate frame. For consistency, we define $\mathbf{a}_{0 \to 1} = (0,0,0)$ . Existing methods tackle this problem by formulating world modeling in the following two ways: + +Video diffusion-based world models. Current leading driving world models, like Vista [18], formulates world modeling in the form of video diffusion models [3], which jointly capture the global spatiotemporal distribution of both past and a fixed-length future, + +$$ +p \left(\{\mathbf {O} _ {T + i} \} _ {i = 1} ^ {n}, \{\mathbf {O} _ {t}, \mathbf {a} _ {t} \} _ {t = 1} ^ {T}\right). +$$ + +This formulation disrupts the causal temporal structure between historical observations and future predictions, limiting its ability to model real-world progressive dynamics and generate flexible-length long-term videos. + +GPT-based world models. Alternatively, autoregressive transformer-based world models [11, 24, 26] discretize image observations into token sequences $\mathbf{O}_t = [\mathbf{t}_1,\mathbf{t}_2,\dots ,\mathbf{t}_L]$ and model the conditional image distribution as a token-by-token prediction, + +$$ +\prod_ {i = 1} ^ {L} p (\mathbf {t} _ {i} \mid \mathbf {t} _ {< i}, \{\mathbf {O} _ {t}, \mathbf {a} _ {t} \} _ {t = 1} ^ {T}). +$$ + +However, this independent token modeling weakens spatial correlations and the quantization process distorts high-frequency details, leading to degraded generation quality. + +Our approach. In contrast, we formulate world modeling as a sequential future prediction process in the temporal domain. Specifically, given past driving observations $\{\mathbf{O}_t\}_{t=1}^T$ and the driving trajectory $\{\mathbf{a}_{t-1 \to t}\}_{t=1}^T$ , our model predicts both a policy for future trajectory planning, + +$$ +\pi \left(\{\mathbf {a} _ {T \rightarrow T + i} \} _ {i = 1} ^ {n} \mid \{\mathbf {O} _ {t}, \mathbf {a} _ {t - 1 \rightarrow t} \} _ {t = 1} ^ {T}\right), +$$ + +and a conditional distribution over the next-frame camera observation as a whole, + +$$ +p \left(\mathbf {O} _ {T + 1} \mid \{\mathbf {O} _ {t}, \mathbf {a} _ {t - 1 \rightarrow t} \} _ {t = 1} ^ {T}, \mathbf {a} _ {T \rightarrow T + 1}\right). +$$ + +The next-frame prediction is conditioned on either a model-predicted action or an externally provided action $\mathbf{a}_{T\rightarrow T + 1}$ . By decoupling causal temporal modeling from fine-grained future prediction, our model can generate flexible-length long videos autoregressively in continuous representations. Moreover, by factorizing trajectory planning from visual generation, our model can seamlessly serve as a real-time motion planner, bridging a critical gap between current driving world models and end-to-end motion planners. + +# 3.3. Epona: Autoregressive Diffusion World Model + +Based on the reformulated world modeling design, we propose Epona, an autoregressive diffusion world model for autonomous driving. Our framework consists of three key components. First, a Multimodal Spatiotemporal Transformer (MST) encodes historical context $\{\mathbf{O}_t,\mathbf{a}_t\}_{t = 1}^T$ into a compact latent representation, effectively capturing environmental context and driving dynamics. Then, based on the historical latents, we employ two specialized diffusion transformers to predict fine-grained future details, including a tiny Trajectory Planning Transformer (TrajDiT) that models the policy $\pi$ for trajectory planning, and a Next-frame Prediction Transformer (VisDiT) that models the visual distribution $p$ for future image generation. This modular design enables a range of autonomous driving applications. + +For instance, MST and VisDiT can be used independently for controllable driving simulations, while MST and TrajDiT facilitate real-time motion planning. + +Multimodal Spatiotemporal Transformer (MST). Given the encoded past driving scenes $\{\mathbf{Z}_t\}_{t=1}^T$ and trajectory $\{\mathbf{a}_{t-1 \to t}\}_{t=1}^T$ , we introduce a multimodal spatiotemporal transformer to effectively integrate temporal dynamics and multimodal information from historical context for future prediction. Inspired by prior work in video generation and world modeling [4, 26, 39], our approach employs interleaved multimodal spatial attention layers and causal temporal attention layers. This design progressively incorporates historical information into a compact latent representation while significantly reducing memory consumption compared to full-sequence attention. Additionally, this design naturally supports historical contexts of variable length. + +Specifically, we first project the flattened visual latent patches $\mathbf{Z} \in \mathbb{R}^{B \times T \times L \times C}$ and action sequences $\mathbf{a} \in \mathbb{R}^{B \times T \times 3}$ into an embedding space. Then we concatenate them along the spatial dimension and add temporal positional embeddings to obtain the latent embedding sequence $\mathbf{E} \in \mathbb{R}^{B \times T \times (L + 3) \times D}$ . This sequence is processed through interleaved multimodal spatiotemporal layers as follows (using einops [46] notation): + +$$ +\begin{array}{l} \mathbf {E} \leftarrow \operatorname {r e a r r a n g e} (\mathbf {E}, (b t) l c \rightarrow (b l) t c)) \\ \mathbf {E} \leftarrow \text {C a u s a l T e m p o r a l L a y e r} (\mathbf {E}, \text {C a u s a l M a x}) \\ \mathbf {E} \leftarrow \operatorname {r e a r r a n g e} (\mathbf {E}, (b l) t c \rightarrow (b t) l c)) \\ \mathbf {E} \leftarrow \operatorname {M u l t i m o d a l S p a t i a l L a y e r} (\mathbf {E}), \\ \end{array} +$$ + +where $B$ is the batch size, $T$ is the number of conditioning frames, $L = H \times W$ is the number of flattened latents in an image, $C$ is the image latent channel dimension, $D$ is the embedding dimension, and CausalMask is the triangular causal attention mask. Finally, we use the latent embedding of the last frame, $\mathbf{F} \in \mathbb{R}^{B \times (L + 3) \times D}$ , as the compact latent representation for the next-stage prediction. After training, this embedding encapsulates the historical context $\{\mathbf{O}_t, \mathbf{a}_{t-1 \to t}\}_{t=1}^T$ . + +Trajectory Planning Diffusion Transformer (TrajDiT). TrajDiT predicts future trajectories using a tiny diffusion transformer. Following the DiT frameworks in most advanced open-source text-to-image and video generation models [27, 33], we adopt a Dual-Single-Stream architecture. In the dual-stream phase, the historical latent representation $\mathbf{F}$ and trajectory data are processed independently through transformer blocks, with only attention operations linking them. In the single-stream phase, they are concatenated to pass through subsequent transformer blocks for effective information fusion. Detailed architecture can be found in the supplementary material. + +During training, we add noise to the target trajectories $\overline{\mathbf{a}}\in \mathbb{R}^{B\times N\times 3}$ using Eq.2. The model then predicts velocity + +![](images/64871f15011ea7d132dfb619eb87548b53b9522938672b693071e5e462e306ba.jpg) +Figure 4. Concept illustration of our training process. Here $x$ can be either image latents or trajectories. + +$v_{\text{traj}}$ conditioned on $\mathbf{F}$ , where $N$ is the planning horizon. We optimize using the rectified flow loss: + +$$ +\mathcal {L} _ {t r a j} = \mathbb {E} _ {\overline {{\mathbf {a}}}, \epsilon , t} \left[ \| v _ {t r a j} (\overline {{\mathbf {a}}} _ {(t)}, t) - (\overline {{\mathbf {a}}} - \epsilon) \| ^ {2} \right]. \tag {4} +$$ + +For inference, random Gaussian noise is iteratively denoised conditioned on $\mathbf{F}$ to generate future trajectory plans. Next-frame Prediction Diffusion Transformer (VisDiT). VisDiT has a similar architecture as TrajDiT, with an additional modulation [41] branch for action control $\mathbf{a}_{T\rightarrow T + 1}$ . We also use the flow loss for visual supervision: + +$$ +\mathcal {L} _ {v i s} = \mathbb {E} _ {\mathbf {Z} _ {T + 1}, \epsilon , t} \left[ \| v _ {v i s} \left(\mathbf {Z} _ {T + 1 (t)}, t\right) - \left(\mathbf {Z} _ {T + 1} - \epsilon\right) \| ^ {2} \right], \tag {5} +$$ + +Together, the total loss jointly optimizes the entire world model: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {t r a j}} + \mathcal {L} _ {\text {v i s}}. \tag {6} +$$ + +During inference, VisDiT denoises $\hat{\mathbf{Z}}_{T + 1}$ , conditioned on $\mathbf{F}$ and the action either predicted by TrajDiT or provided by user. The latents are then decoded using the DCAE decoder to generate the next-frame image $\hat{\mathbf{O}}_{T + 1}$ . + +# 3.4. Chain-of-Forward Training + +With our proposed autoregressive diffusion world model, we can autoregressively generate future videos frame by frame. However, long-term generation suffers from a longstanding autoregressive drift problem [53]: during training, the model predicts the next frame using ground-truth historical context, whereas during inference, it relies on its own past predictions. This domain gap between teacher-forcing training and autoregressive sampling leads to error accumulation and rapid quality degradation. + +To mitigate this, we introduce a chain-of-forward training strategy. Periodically, we perform multiple forward passes using self-predicted frames to enhance the model's robustness to inference noise (see Fig. 4). Notably, to ensure training efficiency, instead of sampling next-frame latents from pure noise, we leverage the model-predicted velocity $v_{\Theta}$ to estimate the denoised latents in one step: + +$$ +\hat {x} _ {(0)} = x _ {(t)} + t v _ {\Theta} (x _ {(t)}, t) \tag {7} +$$ + +Table 1. Comparisons of generated videos on the NuScenes [6] validation set. Our model achieves state-of-the-art FVD score compared to existing driving world models, while extending the video length to over two minutes. *The max duration number indicates the horizon that produces plausible results, following existing methods. + +
MetricDriveGAN [32]DriveDreamer [59]WoVoGen [38]Drive-WM [60]GenAD (OpenDV) [65]Vista [18]DrivingWorld [26]Ours
FID ↓73.452.627.615.815.46.97.47.5
FVD ↓502.3452.0417.7122.7184.089.490.982.8
Max Duration / Frames*N/A4s / 482.5s / 58s / 164s / 815s / 15040s / 400120s / 600
+ +![](images/03c723b6620506ffd57952f097423f4b6dc2b495c924bae75f3cab48c2140f76.jpg) +Figure 5. Qualitative Comparison between Vista [18] and Epona. Zoom in for better views. + +The estimated $\hat{x}_{(0)}$ , along with previous conditioned frames, is then used in the next forward pass to autoregressively generate subsequent frames. This process simulates prediction noise, helping the model adapt to deviations and improving long-term video generation quality. + +# 3.5. Temporal-aware DCAE Decoder + +Unlike conventional autoencoders that downsample images by a factor of 8, DCAE [9] progressively increases this to 32, reducing latent tokens by $16 \times$ . In our world model, we adopt DCAE for image encoding to improve training efficiency and reduce memory usage, enabling conditioning on longer historical contexts. + +However, as an image autoencoder, DCAE lacks temporal interactions, causing flickering when decoding video frame by frame, which degrades visual quality. To address this, we propose a temporal-aware DCAE to enhance interframe consistency. Specifically, to maximize pretrained parameters while minimizing architectural changes, we introduce spatiotemporal self-attention layers before the DCAE decoder while keeping the encoder fixed during fine-tuning. This facilitates multi-frame interactions, greatly improving temporal consistency in generated videos. + +# 4. Experiment + +# 4.1. Implementation Details + +World Model. Our Epona consists of $2.5\,B$ parameters, including a 12-layer multimodal spatiotemporal transformer with $1.3\,B$ parameters, a 12-layer next-frame prediction diffusion transformer with $1.2\,B$ parameters, and a 2-layer trajectory planning diffusion transformer with $50\,M$ parameters. It is trained on publicly available videos from the NuPlan dataset [7] and 700 scenes from the NuScenes dataset [6] from scratch, all images are resized to $512\times 1024$ . We utilize the rectified flow [37] objective for both video + +Table 2. Inference speed. We evaluate our inference speed for generating a 3-second trajectory and a $512 \times 1024$ image per module on a single NVIDIA 4090 GPU. + +
DiT sampling stepsMSTTrajDiTVisDiT
10~0.02s~0.03s~0.3s
100~0.02s~0.3s~2s
+ +generation and trajectory planning tasks, training the entire model in an end-to-end manner. The training was conducted on 48 NVIDIA A100 GPUs for nearly two weeks, with a total of $600\mathrm{k}$ iterations and a batch size of 96. During training, we apply Chain-of-Forward every 10 steps, each time performing three forward passes. We use the AdamW optimizer with a learning rate of $1\times 10^{-4}$ and set the weight decay to $5\times 10^{-2}$ . For inference, we report our speed for each module on a single NVIDIA 4090 GPU in Table 2. In all our experiments, we set our DiT sampling step to 100. Notice that with MST and TrajDiT, our world model can seamlessly serve as a real-time motion planner. + +**Evaluations on video generation.** We employ 1628 video clips from the NuPlan test set [7] and 1646 video clips from the NuScenes validation dataset [6] for performance evaluation, respectively. During the test, our world model is conditioned on 10 consecutive past frames to generate the subsequent frame and repeat the process autoregressively to synthesize future video frames. We use the Frechet Video Distance (FVD) [52] and the Frechet Inception Distance (FID) [22] to evaluate the quality of the generated videos. + +Evaluations on trajectory planning. We evaluate trajectory planning using the NuScenes benchmark [6] and the NAVSIM benchmark [12]. For the NuScenes, we use L2 error and collision rate as the evaluation metrics following the existing works [25, 48, 70] to evaluate the planning performance. L2 error measures the L2 distance between the predicted and ground truth trajectories, while the collision + +![](images/a1ff9fc650f21fae731c82bf5a7210398edff5eb7a907a94e5e6162dbb8090d0.jpg) +Figure 6. Trajectory-controlled video generation. Our world model can generate controllable videos based on predefined trajectories. + +rate measures the frequency of predicted trajectory intersections with objects. The NAVSIM benchmark assesses performance using the predictive driver model score (PDMS), derived from five factors, as shown in Table 4. + +# 4.2. Evaluation of Video Generation + +Quantitative Comparison of Generated Videos. We present a quantitative comparison with existing methods on the NuScenes dataset [6] in Table 1. Since most methods are not publicly available, we compare with the reported results from their respective papers. Notably, the existing methods (e.g., Vista [18]) fine-tune video diffusion models pretrained on large-scale datasets, while our world model, including next-frame DiT, is trained from scratch. As shown in Table 1, our generated videos achieve state-of-the-art FVD scores, indicating smoother and more realistic video generation quality. Moreover, our world model can generate significantly longer video frames compared to existing approaches as shown in Table 1. + +Qualitative Comparison of Generated Videos. We provide a qualitative comparison with the state-of-the-art open-source driving world model, Vista [18]. Since Vista is a 25-frame fixed-length video diffusion model, we perform rollout to generate longer videos as illustrated in their paper. As shown in Fig. 5, our Epona generates consistent long-horizon driving scenes with high-fidelity visuals and detailed structures and vehicles. + +Trajectory-controlled Video Generation. Fig. 6 illustrates the pose controllability of our model. Given the predefined pose trajectory, the different condition frames can generate future frames that conform to the corresponding motion path, which is crucial for obtaining autonomous driving videos under extreme scenarios. + +Extra Long-range Video Generation. Epona combines the strengths of autoregressive and diffusion models, facilitating the generation of high-quality, long-duration videos conditioned on input frames. As shown in Fig.5 and Fig.7, our model can autoregressively generate minute-long driv- + +ing videos with high fidelity and consistency, without noticeable drift. More long-term generation videos are provided in the supplementary materials. + +# 4.3. Evaluation of Trajectory Planning + +For the NuScenes benchmark [6], we compare our world model with several existing methods, as shown in Table 3. Although our model does not achieve the best results, it attains competitive performance without any additional supervision. It is worth noting that incorporating more supervision typically leads to better performance but cost expensive annotations. Additionally, similar to Doe-1 [70], our model only utilizes the front camera, whereas other methods rely on multi-view inputs for planning. As shown in Table 3, our approach can generate reasonable trajectories while achieving the lowest collision rate for a 1-second horizon, which is crucial for long-term realistic video predictions. For the more challenging NAVSIM benchmark [12], as shown in Table 4, our method achieves state-of-the-art results in overall PDMS when conditioned on the past 2 seconds of observations to predict 4-second future trajectories, showing the strong motion planning capability. + +# 4.4. Ablation Study + +Effect of Shared Latent for Multi-modal Joint Prediction. To assess the benefit of jointly modeling scene and trajectory via a shared latent representation, we conduct an ablation by disabling video prediction and training the model solely for trajectory prediction. This variant is evaluated on the NAVISIM test set. As shown in Tab. 5, removing video prediction leads to a noticeable drop in planning performance. This result highlights the advantage of the shared latent $\mathbf{F}$ , which encourages the world model to better capture complex driving dynamics by leveraging visual signals. This ablation confirms that coupling video and trajectory prediction within a unified latent space in world models significantly benefits downstream planning tasks. + +Effect of Chain-of-Forward Training. To evaluate the + +![](images/850e85a12901ccc0803d508487c57e1dd3b8f55520aeac6e5d06a58af680e119.jpg) +Figure 7. Qualitative Comparison between long videos generated by models w/ and w/o Chain-of-Forward training. Left: Visual quality deteriorates rapidly after 10–20 seconds. Right: The same driving scenes with Chain-of-Forward training maintain high visual quality, generating minute-long videos without significant degradation. Zoom in for better views. + +![](images/f092074d6c47a2ac4f000ddcf052ca44a23197fc4eaa98dfbe27da2505a8e7a1.jpg) + +Table 3. End-to-end motion planning performance on the NuScenes [6] dataset. Note that our model achieves a low collision rate, demonstrating its understanding of basic traffic rules via simple next-frame prediction. * represents only using the front camera as input. + +
MethodInputAuxiliary SupervisionL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
ST-P3 [25]CameraMap & Box & Depth1.332.112.902.110.230.621.270.71
UniAD [48]CameraMap & Box & Motion & Tracklets & Occ0.480.961.651.030.050.170.710.31
OccNet [51]Camera3D-Occ & Map & Box1.292.132.992.140.210.591.370.72
OccWorld [68]Camera3D-Occ0.521.272.411.400.120.402.080.87
VAD-Tiny [30]CameraMap & Box & Motion0.601.232.061.300.310.531.330.72
VAD-Base [30]CameraMap & Box & Motion0.541.151.981.220.040.391.170.53
GenAD [69]CameraMap & Box & Motion0.360.831.550.910.060.231.000.43
Doe-1 [70]Camera*QA0.501.182.111.260.040.371.190.53
OursCamera*None0.611.171.981.250.010.220.850.36
+ +Table 4. End-to-end motion planning performance on the NAVSIM [12] test set. NC: no at-fault collision. DAC: drivable area compliance. TTC: time-to-collision. Conf.: comfort. EP: ego progress. PDMS: the predictive driver model score. LAW[35] is in the perception-free setting. Our world model outperforms strong end-to-end planners in the overall PDMS score. + +
MethodInputNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Human/10010010099.987.594.8
UniAD[48]Camera97.891.992.910078.883.4
PARA-Drive[62]Camera97.992.493.099.879.384.6
LAW[35]Camera96.495.488.799.981.784.6
TransFuser[42]Camera & Lidar97.792.892.810079.284.0
DRAMA[67]Camera & Lidar98.093.194.810080.185.5
VADv2[10]Camera & Lidar97.289.191.910076.080.9
OursCamera97.995.193.899.980.486.2
+ +Table 5. Comparison of planning results on the NAVSIM test set. Jointly predicting the next scene using a shared latent significantly improves planning performance. + +
MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Ours w/o Joint Training94.589.788.199.974.778.1
Ours97.995.193.899.980.486.2
+ +impact of our chain-of-forward strategy on model performance, we conduct an ablation study comparing results with and without this strategy. Given that our model iteratively generates the next frame, the chain-of-forward approach simulates potential inference errors during training, thereby enhancing the model's robustness. As shown in Fig. 7 and Fig. 8, as the model autoregressively generates longer sequences, the gap in visual quality and FID score between models with and without the chain-of-forward strategy becomes increasingly significant, validating its effectiveness in long-term video generation. More ablations + +![](images/e022709c0de5912cf52a031faa7a4ee59146b569c0f7162a0de4581b23bef58e.jpg) +Figure 8. Effect of Chain-of-Forward training. FID comparison in NuPlan test set between models w/ and w/o Chain-of-Forward training strategy. + +can be found in the supplementary material. + +# 5. Conclusion + +We have presented Epona, an autoregressive diffusion world model for autonomous driving that jointly predicts high-fidelity future trajectories and driving scenes based on historical driving context. Thanks to our proposed decoupled spatiotemporal modeling and asynchronous multimodal generation strategies, our model achieves high-quality and long-term prediction. In addition, our model could serves as a real-time motion planner via performing trajectory planning. We have demonstrated that our approach significantly advances the state of the art in driving world models, uncovering the large potential for building next-generation autonomous driving systems. + +# References + +[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3 +[2] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. 2, 3 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3, 4 +[4] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, pages 22563-22575, 2023. 5 +[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2 +[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 7, 8 +[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 6, 1 +[8] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion, 2024. 3 +[9] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 4, 6 +[10] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024.8 +[11] Yuntao Chen, Yuqi Wang, and Zhaoxiang Zhang. Driving-gpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers. arXiv preprint arXiv:2412.18607, 2024. 2, 3, 4, 1 +[12] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7, 8 + +[13] Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models. arXiv preprint arXiv:2411.14499, 2024. 2, 3 +[14] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 3 +[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 4 +[16] Ruiyuan Gao, Kai Chen, Bo Xiao, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrivedit: High-resolution long video generation for autonomous driving with adaptive control. arXiv preprint arXiv:2411.13807, 2024. 2, 1 +[17] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3d geometry control. In ICLR, 2024. 2 +[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 4, 6, 7 +[19] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. Dome: Taming diffusion model into high-fidelity controllable occupancy world model. ArXiv, abs/2410.10429, 2024. 2 +[20] Xi Guo, Chenjing Ding, Haoxuan Dou, Xin Zhang, Weixuan Tang, and Wei Wu. Infinitydrive: Breaking time limits in driving world models, 2024. 1 +[21] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2 +[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6 +[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3 +[24] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3, 4, 1 +[25] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 6, 8 +[26] Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. Drivingworld: Constructingworld model for autonomous driving via + +video gpt. arXiv preprint arXiv:2412.19505, 2024. 2, 3, 4, 5, 6 +[27] Tencent Hunyuan. Hunyuanvideo: A systematic framework for large video generative models, 2024. 4, 5, 1, 2 +[28] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. Adriver-i: A general world model for autonomous driving, 2023. 1 +[29] Mingkai Jia, Wei Yin, Xiaotao Hu, Jiaxin Guo, Xiaoyang Guo, Qian Zhang, Xiao-Xiao Long, and Ping Tan. Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization. arXiv preprint arXiv:2507.07997, 2025.3 +[30] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 8 +[31] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. In NeurIPS, 2024. 3 +[32] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 6 +[33] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 4, 5, 1 +[34] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1):1-62, 2022. 2 +[35] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. 2024. 8 +[36] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 2, 3 +[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3, 6 +[38] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. WoVoGen: World Volume-Aware Diffusion for Controllable Multi-Camera Driving Scene Generation. arXiv preprint arXiv:2312.02934, 2023. 2, 6 +[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation, 2024. 5 +[40] Yiyang Ma, Xingchao Liu, Xi aokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Xingkai Yu, Liang Zhao, Yisong Wang, Jiaying Liu, and Chong Ruan. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. ArXiv, abs/2411.07975, 2024. 2 +[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter + +national conference on computer vision, pages 4195-4205, 2023. 2, 5 +[42] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 8 +[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 3 +[44] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2 +[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2 +[46] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2022. 5 +[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4 +[48] Xiaogang Shi, Bin Cui, Gillian Dobbie, and Beng Chin Ooi. Uniad: A unified ad hoc data processing system. ACM Transactions on Database Systems (TODS), 42(1):1-42, 2016. 6, 8 +[49] Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. History-guided video diffusion, 2025. 3 +[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3 +[51] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 8 +[52] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6 +[53] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines, 2024. 3, 5 +[54] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3 +[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2 +[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 3 + +[57] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 3 +[58] Fu-Yun Wang, Zhaoyang Huang, Qiang Ma, Guanglu Song, Xudong Lu, Weikang Bian, Yijin Li, Yu Liu, and Hongsheng Li. Zola: Zero-shot creative long animation generation with short video model. In ECCV, 2024. 3 +[59] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6 +[60] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving. In CVPR, 2024. 2, 6 +[61] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models, 2024. 3 +[62] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 8 +[63] Yanhao Wu, Haoyang Zhang, Tianwei Lin, Lichao Huang, Shujie Luo, Rui Wu, Congpei Qiu, Wei Ke, and Tong Zhang. Generating multimodal driving scenes via next-scene prediction. In CVPR, 2025. 2, 3 +[64] Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goaldriven flow matching for multimodal trajectories generation in end-to-end autonomous driving. ArXiv, abs/2503.05689, 2025. 2 +[65] Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu, Jia Zeng, Ping Luo, Jun Zhang, Andreas Geiger, Yu Qiao, and Hongyang Li. Generalized Predictive Model for Autonomous Driving. In CVPR, 2024. 2, 6 +[66] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation, 2023. 3 +[67] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, and Marcelo H. Ang Jr au2. Drama: An efficient end-to-end motion planner for autonomous driving with mamba, 2024. 8 +[68] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In European conference on computer vision, pages 55-72. Springer, 2024. 8 +[69] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87-104. Springer, 2024. 8 + +[70] Wenzhao Zheng, Zetian Xia, Yuanhui Huang, Sicheng Zuo, Jie Zhou, and Jiwen Lu. Doe-1: Closed-loop autonomous driving with large world model. arXiv preprint arXiv:2412.09627, 2024. 2, 3, 6, 7, 8 +[71] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke S. Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. ArXiv, abs/2408.11039, 2024. 2 \ No newline at end of file diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/images.zip b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..aa0b05da47bde1d3bfd8f5e0fa1c62cc3d975e14 --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:77007ac440c7dee31ec0069320e62b5c44ca09cbd7f8dad1064e13151eb22b3e +size 709705 diff --git a/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/layout.json b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ea472337a4fcaf82d0e999eb1d98ef923f56aa91 --- /dev/null +++ b/2025/Epona_ Autoregressive Diffusion World Model for Autonomous Driving/layout.json @@ -0,0 +1,9153 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 84, + 75, + 526, + 94 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 75, + 526, + 94 + ], + "spans": [ + { + "bbox": [ + 84, + 75, + 526, + 94 + ], + "type": "text", + "content": "Epona: Autoregressive Diffusion World Model for Autonomous Driving" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "spans": [ + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": "Kaiwen Zhang" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Zhenyu Tang" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1,3*}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Xiaotao Hu" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1,5}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Xingang Pan" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Xiaoyang Guo" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Yuan Liu" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Jingwei Huang" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Li Yuan" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Qian Zhang" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Xiao-Xiao Long" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{4\\dagger}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Xun Cao" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "text", + "content": " Wei Yin" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 114, + 571, + 146 + ], + "type": "inline_equation", + "content": "\\S" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "spans": [ + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Horizon Robotics " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Tsinghua University " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Shenzhen Gruadate School, Peking University " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Nanjing University " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Hong Kong University of Science and Technology " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Nanyang Technological University " + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 74, + 150, + 536, + 195 + ], + "type": "text", + "content": "Tencent Hunyuan" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 77, + 199, + 307, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 199, + 307, + 211 + ], + "spans": [ + { + "bbox": [ + 77, + 199, + 307, + 211 + ], + "type": "text", + "content": "(A) Consistent High-resolution Long Video Generation" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 77, + 213, + 138, + 308 + ], + "blocks": [ + { + "bbox": [ + 77, + 213, + 138, + 308 + ], + "lines": [ + { + "bbox": [ + 77, + 213, + 138, + 308 + ], + "spans": [ + { + "bbox": [ + 77, + 213, + 138, + 308 + ], + "type": "image", + "image_path": "481b9130b93e2a87f7b0b4c6e84e88da18c65d77ec21c4c48018bad82e58ee85.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 143, + 213, + 519, + 308 + ], + "blocks": [ + { + "bbox": [ + 143, + 213, + 519, + 308 + ], + "lines": [ + { + "bbox": [ + 143, + 213, + 519, + 308 + ], + "spans": [ + { + "bbox": [ + 143, + 213, + 519, + 308 + ], + "type": "image", + "image_path": "03bb646c4b8ad1b03b1bd3b5bdab749a995466dd7e3cd649b73c18191b662dd6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 78, + 312, + 257, + 323 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 312, + 257, + 323 + ], + "spans": [ + { + "bbox": [ + 78, + 312, + 257, + 323 + ], + "type": "text", + "content": "(B) Trajectory-controlled Video Generation" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 78, + 325, + 518, + 413 + ], + "blocks": [ + { + "bbox": [ + 78, + 325, + 518, + 413 + ], + "lines": [ + { + "bbox": [ + 78, + 325, + 518, + 413 + ], + "spans": [ + { + "bbox": [ + 78, + 325, + 518, + 413 + ], + "type": "image", + "image_path": "b4f070a0c9b5dae181cea8b0c5d8fc80d65d71100299839a378d11ffb523da50.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 78, + 415, + 258, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 78, + 415, + 258, + 426 + ], + "spans": [ + { + "bbox": [ + 78, + 415, + 258, + 426 + ], + "type": "text", + "content": "(C)Traffic World Knowledge Understanding" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 77, + 427, + 389, + 488 + ], + "blocks": [ + { + "bbox": [ + 77, + 427, + 389, + 488 + ], + "lines": [ + { + "bbox": [ + 77, + 427, + 389, + 488 + ], + "spans": [ + { + "bbox": [ + 77, + 427, + 389, + 488 + ], + "type": "image", + "image_path": "4c8ede8bf3c6201cfe1a2019d3c646a6a018d88dd30ff90e1dfc2a3d0b523d6e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 494, + 555, + 528 + ], + "lines": [ + { + "bbox": [ + 54, + 494, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 54, + 494, + 555, + 528 + ], + "type": "text", + "content": "Figure 1. Versatile capabilities of Epona. Given historical driving context, our Epona can generate consistent minutes-long future driving scenes at high resolution (A). It can be controlled by diverse trajectories (B), and understand real-world traffic knowledge (C). In addition, our world model can predict future trajectories and serve as an end-to-end real-time motion planner (D)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 391, + 415, + 538, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 391, + 415, + 538, + 426 + ], + "spans": [ + { + "bbox": [ + 391, + 415, + 538, + 426 + ], + "type": "text", + "content": "(D) End-to-End Trajectory Planning" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 402, + 430, + 518, + 483 + ], + "blocks": [ + { + "bbox": [ + 402, + 430, + 518, + 483 + ], + "lines": [ + { + "bbox": [ + 402, + 430, + 518, + 483 + ], + "spans": [ + { + "bbox": [ + 402, + 430, + 518, + 483 + ], + "type": "image", + "image_path": "6428349483a4095b34b420c3983bc4ec9c2c58dfd1f428356b7aea6b52bb1808.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 151, + 533, + 200, + 545 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 533, + 200, + 545 + ], + "spans": [ + { + "bbox": [ + 151, + 533, + 200, + 545 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 555, + 297, + 687 + ], + "type": "text", + "content": "Diffusion models have demonstrated exceptional visual quality in video generation, making them promising for autonomous driving world modeling. However, existing video diffusion-based world models struggle with flexible-length, long-horizon predictions and integrating trajectory planning. This is because conventional video diffusion models rely on global joint distribution modeling of fixed-length frame sequences rather than sequentially constructing localized distributions at each timestep. In this work, we propose Epona, an autoregressive diffusion world model that enables localized spatiotemporal distribution modeling" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 533, + 556, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 556, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 556, + 701 + ], + "type": "text", + "content": "through two key innovations: 1) Decoupled spatiotemporal factorization that separates temporal dynamics modeling from fine-grained future world generation, and 2) Modular trajectory and video prediction that seamlessly integrate motion planning with visual modeling in an end-to-end framework. Our architecture enables high-resolution, long-duration generation while introducing a novel chain-of-forward training strategy to address error accumulation in autoregressive loops. Experimental results demonstrate state-of-the-art performance with " + }, + { + "bbox": [ + 313, + 533, + 556, + 701 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 533, + 556, + 701 + ], + "type": "text", + "content": " FVD improvement and minutes longer prediction duration compared to prior works. The learned world model further serves as a real-time motion planner, outperforming strong end-to-end planners on NAVSIM benchmarks." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 38 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 693, + 296, + 714 + ], + "type": "text", + "content": "*These authors contributed equally to this work. †Project advisor. §Project lead, Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27220" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 247 + ], + "type": "text", + "content": "Recently, with the rapid development of video generation models, world models have attracted significant attention and emerged as a powerful paradigm for physical world simulations and autonomous decision-making [1, 13, 21, 34]. These foundation models enable agents to understand inherent world knowledge and predict future dynamics, making them particularly promising for autonomous driving. Unlike traditional separate perception-planning pipelines, which require extensive annotations and explicit supervision, generative driving world models [11, 16, 18, 24, 26, 59, 60, 63, 70] integrate visual scene understanding with future prediction in a self-supervised manner, offering a new solution toward end-to-end autonomous driving." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 247, + 294, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 247, + 294, + 378 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 294, + 378 + ], + "type": "text", + "content": "Generative world models primarily fall into two categories: diffusion-based approaches and GPT-style autoregressive methods. The diffusion-based paradigm (e.g., Vista [18]), while achieving impressive visual fidelity through joint distribution modeling of fixed-length videos [3, 23, 47, 50], fundamentally suffers from its inability to model per-timestep local distributions. This limitation manifests in critical deficiencies: failure to support variable-length long-range prediction crucial for dynamic world simulation, and infeasible trajectory planning due to the lack of multimodal prediction mechanism." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 379, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 545 + ], + "type": "text", + "content": "Conversely, GPT-style approaches [5, 44, 45, 55] address temporal flexibility through autoregressive next-token prediction (as seen in GAIA-1 [24]). However, the quantization and tokenization process significantly degrades visual quality and planning precision. Moreover, the causal nature of autoregressive transformers constrains them to predicting only the next action rather than planning long-horizon trajectories [11], limiting their potential to serve as end-to-end driving planners. Both paradigms reveal complementary shortcomings - diffusion models lack temporal decomposition while autoregressive transformers sacrifice continuous visual precision - highlighting the need for a unified framework that reconciles these divergent advantages for competent driving world modeling." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": "We introduce Epona, an autoregressive diffusion world model that achieves high-resolution long-horizon video generation and accurate trajectory planning. Our core innovation comes from three key designs: 1) Decoupled spatiotemporal factorization. While existing video diffusion methods model joint spatial-temporal distributions of past and future frames, we assume their temporal latent modeling lacks explicit causality constraints, leading to error accumulation in long sequences. Epona addresses this through spacetime-disentangled processing: A GPT-style transformer with causal attention handles temporal dynamics in compressed latent space, while twin diffusion transformers separately handle spatial rendering and trajectory generation. 2) Asynchronous multi-modal genera" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 192 + ], + "type": "text", + "content": "tion. Building upon this foundation, we decouple trajectory planning from visual generation through parallel denoising processes. Two specialized DiTs [33, 41] asynchronously generate 3-second vehicle trajectories and the single next future frame. Both streams share flow-matching objectives [2, 36, 37] conditioned on the same temporal latent, ensuring alignment while preserving modality-specific optimizations. 3) A chain-of-forward training strategy addresses error accumulation and content drift in autoregressive loops." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 553, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 337 + ], + "type": "text", + "content": "Epona offers several additional advantages. 1) Long-horizon generation. Our autoregressive diffusion model can achieve long-time generation over up to 2 minutes, significantly outperforming existing world models. 2) Real-time trajectory planning. The separate multi-modal generation architecture enables to perform trajectory planning solely while video prediction is deactivated, significantly reducing the inference FLOPS. This enables high-quality and even real-time trajectory planning, achieving rates of up to " + }, + { + "bbox": [ + 313, + 193, + 553, + 337 + ], + "type": "inline_equation", + "content": "20\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 193, + 553, + 337 + ], + "type": "text", + "content": ". 3) Visual detail preservation. Our autoregressive formulation adopts continuous visual tokenizer instead of discrete ones, thus preserving rich scene details." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 338, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 338, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 553, + 540 + ], + "type": "text", + "content": "Extensive experiments demonstrate the effectiveness and superiority of our world model. For video generation, our model achieves state-of-the-art FVD [52] on the NuScenes [6] benchmark, surpassing the best-performing Vista [18] by " + }, + { + "bbox": [ + 313, + 338, + 553, + 540 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 313, + 338, + 553, + 540 + ], + "type": "text", + "content": ", while extending generation length from 15 seconds to over 2 minutes (600 frames). Thanks to joint supervision with trajectory prediction, Epona allows action control to simulate diverse driving scenarios, as shown in Fig. 1 (B). For motion planning, our method outperforms strong end-to-end planners on NAVSIM [12] without perception inputs (e.g., 3D boxes/lanes). Notably, we observe that Epona learns essential traffic world knowledge (e.g., stop driving at red light) purely from self-supervised future prediction tasks, as shown in Fig. 1 (C). This suggests that our world model can implicitly learn real-world driving dynamics, making it a promising direction for next-generation autonomous driving systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 555, + 400, + 567 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 400, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 400, + 567 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 575, + 520, + 588 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 520, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 520, + 588 + ], + "type": "text", + "content": "2.1. World Models for Autonomous Driving" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 715 + ], + "type": "text", + "content": "Constructing real-world driving world models have drawn considerable attention in recent years, among which vision-centric approaches gain prominence due to their superior sensor flexibility, data accessibility, and more human-like representation forms. Early efforts primarily focused on adapting pre-trained diffusion models (e.g., Stable Diffusion [3, 47]) to driving scenarios through fine-tuning. However, these methods either lacked critical planning modules [16, 18, 19] or were limited to low-resolution, short-term generation [17, 38, 59, 60, 64, 65], making them un" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "27221" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 552, + 194 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 552, + 194 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 552, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 552, + 194 + ], + "type": "image", + "image_path": "4d2234653f65cc239e1d00117dc7c8a02bea139b8ac5aaf2222cce526d8efdba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": "Figure 2. Overview of Epona. Our world model utilizes a multimodal spatiotemporal transformer to process the historical context of the first " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": " frames and employs a next-frame prediction DiT to generate the frame at " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "T + 1" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": " and a trajectory planning DiT to forecast the future " + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 204, + 555, + 250 + ], + "type": "text", + "content": "-frame pose trajectory. By adopting a chain-of-forward strategy, our approach enables high-quality and long-horizon video generation with an autoregressive manner." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 255, + 296, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 296, + 447 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 296, + 447 + ], + "type": "text", + "content": "suitable for consistent long-range prediction and real-time planning. Recent work [11, 24, 26, 63, 70] explored harnessing GPT-like architecture to unify visual and action modeling and achieved long-range autoregressive generation. Yet, these methods require encoding images and trajectories into discrete tokens [14, 29, 54], which significantly degrades visual fidelity and trajectory precision. Similarly, while the newly released Cosmos [1] foundation model can serve as a driving world model, it does not introduce a new framework, facing the same limitations as the previous methods. In addition, its large parameter count and high computational demands limit its practicality. In contrast, we propose a novel autoregressive diffusion world model framework for autonomous driving, enabling long-range autoregressive generation in continuous visual and trajectory representations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 456, + 188, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 188, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 188, + 468 + ], + "type": "text", + "content": "2.2. Long Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 475, + 296, + 713 + ], + "type": "text", + "content": "Long-term prediction is not only a key challenge for current video generation models but also a crucial capability for robust world models, as it reflects the model's ability to learn consistent environment dynamics and accurately simulate real-world temporal progression [13]. Since original video diffusion models (e.g. SVD [3]) are limited to fixed-length short clips generation, prior methods have explored extrapolating video length by noise rescheduling [43], overlapped generation [57, 58] or hierarchical generation [66]. However, these techniques fail to resolve inherent model constraints, often resulting in inconsistencies and abrupt visual changes in long videos. Autoregressive approaches [11, 56, 61] naturally support variable-length generation but suffer from quality degradation due to domain shift between teacher-forcing training and error accumulation in sampling. GameNGen [53] and DrivingWorld [26] introduce noise augmentation and random token dropout during training to alleviate the problem by simulating the error in sampling, but still limited to specific model architectures. We propose a general chain-of-forward strat" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 255, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 376 + ], + "type": "text", + "content": "egy allowing the model to directly learn inference errors during training, effectively reducing autoregressive drift. Meanwhile, recent works such as Diffusion Forcing [8, 49] and FIFO-Diffusion [31] explore integrating autoregressive generation in video diffusion by adjusting frame-wise noise levels and leveraging causal network designs. Our model adopts a similar causal temporal modeling strategy but redefines the architecture into a two-stage end-to-end framework, allowing joint generation of motion plans and next-frame images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 370, + 398 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 586 + ], + "type": "text", + "content": "In this section, we formally present the model framework and training techniques of Epona. We begin with preliminaries on diffusion models in Sec. 3.1 and discuss our world model formulation design insights in Sec. 3.2. Then we introduce our proposed autoregressive diffusion world model framework in Sec. 3.3, including three dedicated modules: a multimodal spatiotemporal transformer to capture historical context, a trajectory planning DiT to generate future 3-seconds trajectories, and a next-frame prediction DiT to generate the next-frame images. To mitigate autoregressive drift and enable long-horizon video generation, we propose a simple yet effective chain-of-forward training strategy, detailed in Sec.3.4. Additionally, to enhance video quality, we introduce a temporal-aware DCAE decoder in Sec.3.5. An overview of our method is illustrated in Fig. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 593, + 394, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 394, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 394, + 605 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "text", + "content": "Diffusion models [23, 50] are a family of powerful generative models that transform noise samples " + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "inline_equation", + "content": "x_{(1)}" + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "text", + "content": " drawn from a prior distribution " + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "inline_equation", + "content": "p_1 = \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "text", + "content": " into data samples " + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "inline_equation", + "content": "x_{(0)}" + }, + { + "bbox": [ + 313, + 609, + 555, + 658 + ], + "type": "text", + "content": " from the target distribution in terms of a differentiable equation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 367, + 667, + 555, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 667, + 555, + 681 + ], + "spans": [ + { + "bbox": [ + 367, + 667, + 555, + 681 + ], + "type": "interline_equation", + "content": "d x _ {(t)} = v _ {\\Theta} \\left(x _ {(t)}, t\\right) d t, t \\in [ 0, 1 ], \\tag {1}", + "image_path": "37e0d7580468e38ce2f074c670de6b7a5c36238885f3367627d83d5ea1b8742d.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "where velocity " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": " is parametrized by a neural network " + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": ". Rectified flow [2, 36, 37] proposes to define a straight prob-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27222" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 74, + 294, + 185 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 294, + 185 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 294, + 185 + ], + "type": "image", + "image_path": "a6bdef404dd05a932485808fae60e82db343610c0f1412aa90acb47d3b9ac0d3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 194, + 296, + 262 + ], + "lines": [ + { + "bbox": [ + 55, + 194, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 194, + 296, + 262 + ], + "type": "text", + "content": "Figure 3. Comparison of Different World Modeling Formulation. Up: Conventional autoregressive pipeline quantizes continuous images into discrete tokens and perform next-token prediction iteratively. Middle: The video-diffusion-based methods generate future " + }, + { + "bbox": [ + 55, + 194, + 296, + 262 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 194, + 296, + 262 + ], + "type": "text", + "content": " frames simultaneously. Down: Our method autoregressively predicts fine-grained future frames in continuous space." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "text", + "content": "ability path between " + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "inline_equation", + "content": "p_0" + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "text", + "content": " to improve training and sampling efficiency and optimize the network " + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 55, + 270, + 296, + 307 + ], + "type": "text", + "content": " using a velocity prediction loss:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 97, + 312, + 296, + 327 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 312, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 97, + 312, + 296, + 327 + ], + "type": "interline_equation", + "content": "x _ {(t)} = (1 - t) x _ {(0)} + t \\epsilon , \\epsilon \\sim \\mathcal {N} (\\mathbf {0}, \\mathbf {I}), \\tag {2}", + "image_path": "dea5b8bd9a60d05675d54ba9a06e3b3490eb87f1ac8b009cdb13dbaab5f61dcd.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 79, + 342, + 294, + 358 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 342, + 294, + 358 + ], + "spans": [ + { + "bbox": [ + 79, + 342, + 294, + 358 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {R F} = \\mathbb {E} _ {x _ {(0)}, \\epsilon , t} \\left[ \\| v _ {\\Theta} (x _ {(t)}, t) - (x _ {(0)} - \\epsilon) \\| ^ {2} \\right], \\tag {3}", + "image_path": "729704d1bd2285a943af3fd74fbdd3dd223fdc4a14869c788cd448af3d7a2240.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 296, + 456 + ], + "type": "text", + "content": "which has proven to be an effective and scalable approach in state-of-the-art image and video generation models [15, 27, 33]. In Epona, we adopt the diffusion model and rectified flow objective for both next-frame image and trajectory generation. Particularly, to enhance efficiency, we encode images into compact latents using a pre-trained deep compression encoder [9] and adopt a latent diffusion model [47] for image synthesis." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 263, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 263, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 263, + 475 + ], + "type": "text", + "content": "3.2. Reformulation of World Model Designs" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": "In this section, we discuss different world model formulation design choices, as illustrated in Fig. 3. Given a sequence of previous front-view camera observations " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": " and the corresponding driving trajectory " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": ", the goal of driving world moels is to predict future driving dynamics based on historical context. Here, each driving action " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{t_1 \\to t_2} := (\\Delta \\theta_{t_1 \\to t_2}, \\Delta x_{t_1 \\to t_2}, \\Delta y_{t_1 \\to t_2}) \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": " represents the vehicle's motion from " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\Delta \\theta" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": " denotes the orientation change, and " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "(\\Delta x, \\Delta y)" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": " specify the relative displacement in the ego-coordinate frame. For consistency, we define " + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{0 \\to 1} = (0,0,0)" + }, + { + "bbox": [ + 55, + 479, + 296, + 634 + ], + "type": "text", + "content": ". Existing methods tackle this problem by formulating world modeling in the following two ways:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 296, + 694 + ], + "type": "text", + "content": "Video diffusion-based world models. Current leading driving world models, like Vista [18], formulates world modeling in the form of video diffusion models [3], which jointly capture the global spatiotemporal distribution of both past and a fixed-length future," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 112, + 700, + 238, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 112, + 700, + 238, + 715 + ], + "spans": [ + { + "bbox": [ + 112, + 700, + 238, + 715 + ], + "type": "interline_equation", + "content": "p \\left(\\{\\mathbf {O} _ {T + i} \\} _ {i = 1} ^ {n}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}\\right).", + "image_path": "fe05cb7caca822012c3970e8e345bad034632c649cf7abbe21a7d51be63cdd8c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 120 + ], + "type": "text", + "content": "This formulation disrupts the causal temporal structure between historical observations and future predictions, limiting its ability to model real-world progressive dynamics and generate flexible-length long-term videos." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 121, + 555, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 180 + ], + "type": "text", + "content": "GPT-based world models. Alternatively, autoregressive transformer-based world models [11, 24, 26] discretize image observations into token sequences " + }, + { + "bbox": [ + 313, + 121, + 555, + 180 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t = [\\mathbf{t}_1,\\mathbf{t}_2,\\dots ,\\mathbf{t}_L]" + }, + { + "bbox": [ + 313, + 121, + 555, + 180 + ], + "type": "text", + "content": " and model the conditional image distribution as a token-by-token prediction," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 375, + 189, + 494, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 375, + 189, + 494, + 222 + ], + "spans": [ + { + "bbox": [ + 375, + 189, + 494, + 222 + ], + "type": "interline_equation", + "content": "\\prod_ {i = 1} ^ {L} p (\\mathbf {t} _ {i} \\mid \\mathbf {t} _ {< i}, \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t} \\} _ {t = 1} ^ {T}).", + "image_path": "41084f0250fa7b32f46fbb2bd7af55709d804fb1be68246994da09cff8f7a947.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 230, + 554, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 554, + 266 + ], + "type": "text", + "content": "However, this independent token modeling weakens spatial correlations and the quantization process distorts high-frequency details, leading to degraded generation quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "text", + "content": "Our approach. In contrast, we formulate world modeling as a sequential future prediction process in the temporal domain. Specifically, given past driving observations " + }, + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "text", + "content": " and the driving trajectory " + }, + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T" + }, + { + "bbox": [ + 313, + 266, + 554, + 327 + ], + "type": "text", + "content": ", our model predicts both a policy for future trajectory planning," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 353, + 334, + 515, + 350 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 334, + 515, + 350 + ], + "spans": [ + { + "bbox": [ + 353, + 334, + 515, + 350 + ], + "type": "interline_equation", + "content": "\\pi \\left(\\{\\mathbf {a} _ {T \\rightarrow T + i} \\} _ {i = 1} ^ {n} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}\\right),", + "image_path": "0789ee5c5335b103d12344f978ede219f7130a5411e02b14a8f05e862625c4f9.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 357, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 380 + ], + "type": "text", + "content": "and a conditional distribution over the next-frame camera observation as a whole," + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 350, + 390, + 518, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 390, + 518, + 406 + ], + "spans": [ + { + "bbox": [ + 350, + 390, + 518, + 406 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {O} _ {T + 1} \\mid \\{\\mathbf {O} _ {t}, \\mathbf {a} _ {t - 1 \\rightarrow t} \\} _ {t = 1} ^ {T}, \\mathbf {a} _ {T \\rightarrow T + 1}\\right).", + "image_path": "f630189d330c470227f3dd3b74013c0514bd5fb9d895dfc8192a54995218d97a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "content": "The next-frame prediction is conditioned on either a model-predicted action or an externally provided action " + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{T\\rightarrow T + 1}" + }, + { + "bbox": [ + 313, + 413, + 554, + 521 + ], + "type": "text", + "content": ". By decoupling causal temporal modeling from fine-grained future prediction, our model can generate flexible-length long videos autoregressively in continuous representations. Moreover, by factorizing trajectory planning from visual generation, our model can seamlessly serve as a real-time motion planner, bridging a critical gap between current driving world models and end-to-end motion planners." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 529, + 554, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 541 + ], + "type": "text", + "content": "3.3. Epona: Autoregressive Diffusion World Model" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": "Based on the reformulated world modeling design, we propose Epona, an autoregressive diffusion world model for autonomous driving. Our framework consists of three key components. First, a Multimodal Spatiotemporal Transformer (MST) encodes historical context " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}_t,\\mathbf{a}_t\\}_{t = 1}^T" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " into a compact latent representation, effectively capturing environmental context and driving dynamics. Then, based on the historical latents, we employ two specialized diffusion transformers to predict fine-grained future details, including a tiny Trajectory Planning Transformer (TrajDiT) that models the policy " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " for trajectory planning, and a Next-frame Prediction Transformer (VisDiT) that models the visual distribution " + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 545, + 555, + 713 + ], + "type": "text", + "content": " for future image generation. This modular design enables a range of autonomous driving applications." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27223" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 108 + ], + "type": "text", + "content": "For instance, MST and VisDiT can be used independently for controllable driving simulations, while MST and TrajDiT facilitate real-time motion planning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "text", + "content": "Multimodal Spatiotemporal Transformer (MST). Given the encoded past driving scenes " + }, + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{Z}_t\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "text", + "content": " and trajectory " + }, + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 108, + 295, + 275 + ], + "type": "text", + "content": ", we introduce a multimodal spatiotemporal transformer to effectively integrate temporal dynamics and multimodal information from historical context for future prediction. Inspired by prior work in video generation and world modeling [4, 26, 39], our approach employs interleaved multimodal spatial attention layers and causal temporal attention layers. This design progressively incorporates historical information into a compact latent representation while significantly reducing memory consumption compared to full-sequence attention. Additionally, this design naturally supports historical contexts of variable length." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "text", + "content": "Specifically, we first project the flattened visual latent patches " + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{Z} \\in \\mathbb{R}^{B \\times T \\times L \\times C}" + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "text", + "content": " and action sequences " + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{a} \\in \\mathbb{R}^{B \\times T \\times 3}" + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "text", + "content": " into an embedding space. Then we concatenate them along the spatial dimension and add temporal positional embeddings to obtain the latent embedding sequence " + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{B \\times T \\times (L + 3) \\times D}" + }, + { + "bbox": [ + 55, + 275, + 296, + 371 + ], + "type": "text", + "content": ". This sequence is processed through interleaved multimodal spatiotemporal layers as follows (using einops [46] notation):" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 376, + 292, + 434 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 292, + 434 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 292, + 434 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b t) l c \\rightarrow (b l) t c)) \\\\ \\mathbf {E} \\leftarrow \\text {C a u s a l T e m p o r a l L a y e r} (\\mathbf {E}, \\text {C a u s a l M a x}) \\\\ \\mathbf {E} \\leftarrow \\operatorname {r e a r r a n g e} (\\mathbf {E}, (b l) t c \\rightarrow (b t) l c)) \\\\ \\mathbf {E} \\leftarrow \\operatorname {M u l t i m o d a l S p a t i a l L a y e r} (\\mathbf {E}), \\\\ \\end{array}", + "image_path": "a1d6175fad61af3add2defce928061b3cdf5675edfa1942ba64d7dbbdd8cf628.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": " is the batch size, " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": " is the number of conditioning frames, " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "L = H \\times W" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": " is the number of flattened latents in an image, " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": " is the image latent channel dimension, " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": " is the embedding dimension, and CausalMask is the triangular causal attention mask. Finally, we use the latent embedding of the last frame, " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{F} \\in \\mathbb{R}^{B \\times (L + 3) \\times D}" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": ", as the compact latent representation for the next-stage prediction. After training, this embedding encapsulates the historical context " + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{O}_t, \\mathbf{a}_{t-1 \\to t}\\}_{t=1}^T" + }, + { + "bbox": [ + 55, + 437, + 295, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "content": "Trajectory Planning Diffusion Transformer (TrajDiT). TrajDiT predicts future trajectories using a tiny diffusion transformer. Following the DiT frameworks in most advanced open-source text-to-image and video generation models [27, 33], we adopt a Dual-Single-Stream architecture. In the dual-stream phase, the historical latent representation " + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 55, + 546, + 295, + 689 + ], + "type": "text", + "content": " and trajectory data are processed independently through transformer blocks, with only attention operations linking them. In the single-stream phase, they are concatenated to pass through subsequent transformer blocks for effective information fusion. Detailed architecture can be found in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "During training, we add noise to the target trajectories " + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{a}}\\in \\mathbb{R}^{B\\times N\\times 3}" + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": " using Eq.2. The model then predicts velocity" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 64, + 541, + 173 + ], + "blocks": [ + { + "bbox": [ + 326, + 64, + 541, + 173 + ], + "lines": [ + { + "bbox": [ + 326, + 64, + 541, + 173 + ], + "spans": [ + { + "bbox": [ + 326, + 64, + 541, + 173 + ], + "type": "image", + "image_path": "64871f15011ea7d132dfb619eb87548b53b9522938672b693071e5e462e306ba.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 177, + 553, + 199 + ], + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 199 + ], + "type": "text", + "content": "Figure 4. Concept illustration of our training process. Here " + }, + { + "bbox": [ + 313, + 177, + 553, + 199 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 177, + 553, + 199 + ], + "type": "text", + "content": " can be either image latents or trajectories." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "inline_equation", + "content": "v_{\\text{traj}}" + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "text", + "content": " conditioned on " + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 211, + 553, + 234 + ], + "type": "text", + "content": " is the planning horizon. We optimize using the rectified flow loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 244, + 553, + 259 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 244, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 342, + 244, + 553, + 259 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t r a j} = \\mathbb {E} _ {\\overline {{\\mathbf {a}}}, \\epsilon , t} \\left[ \\| v _ {t r a j} (\\overline {{\\mathbf {a}}} _ {(t)}, t) - (\\overline {{\\mathbf {a}}} - \\epsilon) \\| ^ {2} \\right]. \\tag {4}", + "image_path": "dc1d199d4fb5beadfa76dcd1477d18e8322b653adf363dabb1dbca084dd5566b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "text", + "content": "For inference, random Gaussian noise is iteratively denoised conditioned on " + }, + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "text", + "content": " to generate future trajectory plans. Next-frame Prediction Diffusion Transformer (VisDiT). VisDiT has a similar architecture as TrajDiT, with an additional modulation [41] branch for action control " + }, + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "inline_equation", + "content": "\\mathbf{a}_{T\\rightarrow T + 1}" + }, + { + "bbox": [ + 313, + 269, + 553, + 342 + ], + "type": "text", + "content": ". We also use the flow loss for visual supervision:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 350, + 553, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 350, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 321, + 350, + 553, + 376 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v i s} = \\mathbb {E} _ {\\mathbf {Z} _ {T + 1}, \\epsilon , t} \\left[ \\| v _ {v i s} \\left(\\mathbf {Z} _ {T + 1 (t)}, t\\right) - \\left(\\mathbf {Z} _ {T + 1} - \\epsilon\\right) \\| ^ {2} \\right], \\tag {5}", + "image_path": "957c28ad55627937f82bac42790ca052e6b31c11478ce7ae54e8b6d300b19519.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 553, + 399 + ], + "type": "text", + "content": "Together, the total loss jointly optimizes the entire world model:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 394, + 402, + 553, + 416 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 394, + 402, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 394, + 402, + 553, + 416 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {t r a j}} + \\mathcal {L} _ {\\text {v i s}}. \\tag {6}", + "image_path": "21b993e877341d92e3f1a9476b4bac8223e1cbf6acb7df60cf6572a0fe3ada65.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "text", + "content": "During inference, VisDiT denoises " + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{Z}}_{T + 1}" + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "text", + "content": ", conditioned on " + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "text", + "content": " and the action either predicted by TrajDiT or provided by user. The latents are then decoded using the DCAE decoder to generate the next-frame image " + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{O}}_{T + 1}" + }, + { + "bbox": [ + 313, + 422, + 553, + 472 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 467, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 467, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 467, + 493 + ], + "type": "text", + "content": "3.4. Chain-of-Forward Training" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 605 + ], + "type": "text", + "content": "With our proposed autoregressive diffusion world model, we can autoregressively generate future videos frame by frame. However, long-term generation suffers from a longstanding autoregressive drift problem [53]: during training, the model predicts the next frame using ground-truth historical context, whereas during inference, it relies on its own past predictions. This domain gap between teacher-forcing training and autoregressive sampling leads to error accumulation and rapid quality degradation." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "content": "To mitigate this, we introduce a chain-of-forward training strategy. Periodically, we perform multiple forward passes using self-predicted frames to enhance the model's robustness to inference noise (see Fig. 4). Notably, to ensure training efficiency, instead of sampling next-frame latents from pure noise, we leverage the model-predicted velocity " + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "inline_equation", + "content": "v_{\\Theta}" + }, + { + "bbox": [ + 313, + 606, + 553, + 690 + ], + "type": "text", + "content": " to estimate the denoised latents in one step:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 380, + 701, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 701, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 380, + 701, + 553, + 715 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {(0)} = x _ {(t)} + t v _ {\\Theta} (x _ {(t)}, t) \\tag {7}", + "image_path": "c4e69d5c04c50a1aced733ecdda8a2d7d7f4127a254194149b5ae2b691fff79c.jpg" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27224" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 103, + 555, + 148 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 555, + 101 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 555, + 101 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 555, + 101 + ], + "type": "text", + "content": "Table 1. Comparisons of generated videos on the NuScenes [6] validation set. Our model achieves state-of-the-art FVD score compared to existing driving world models, while extending the video length to over two minutes. *The max duration number indicates the horizon that produces plausible results, following existing methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 103, + 555, + 148 + ], + "lines": [ + { + "bbox": [ + 56, + 103, + 555, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 103, + 555, + 148 + ], + "type": "table", + "html": "
MetricDriveGAN [32]DriveDreamer [59]WoVoGen [38]Drive-WM [60]GenAD (OpenDV) [65]Vista [18]DrivingWorld [26]Ours
FID ↓73.452.627.615.815.46.97.47.5
FVD ↓502.3452.0417.7122.7184.089.490.982.8
Max Duration / Frames*N/A4s / 482.5s / 58s / 164s / 815s / 15040s / 400120s / 600
", + "image_path": "45788c2453ea7f8973e9ef24dec3ef354493d3389183695dc7f069595d2e691b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 59, + 162, + 545, + 260 + ], + "blocks": [ + { + "bbox": [ + 59, + 162, + 545, + 260 + ], + "lines": [ + { + "bbox": [ + 59, + 162, + 545, + 260 + ], + "spans": [ + { + "bbox": [ + 59, + 162, + 545, + 260 + ], + "type": "image", + "image_path": "03c723b6620506ffd57952f097423f4b6dc2b495c924bae75f3cab48c2140f76.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 132, + 266, + 476, + 277 + ], + "lines": [ + { + "bbox": [ + 132, + 266, + 476, + 277 + ], + "spans": [ + { + "bbox": [ + 132, + 266, + 476, + 277 + ], + "type": "text", + "content": "Figure 5. Qualitative Comparison between Vista [18] and Epona. Zoom in for better views." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 284, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 284, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 284, + 295, + 344 + ], + "type": "text", + "content": "The estimated " + }, + { + "bbox": [ + 55, + 284, + 295, + 344 + ], + "type": "inline_equation", + "content": "\\hat{x}_{(0)}" + }, + { + "bbox": [ + 55, + 284, + 295, + 344 + ], + "type": "text", + "content": ", along with previous conditioned frames, is then used in the next forward pass to autoregressively generate subsequent frames. This process simulates prediction noise, helping the model adapt to deviations and improving long-term video generation quality." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 350, + 231, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 231, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 231, + 363 + ], + "type": "text", + "content": "3.5. Temporal-aware DCAE Decoder" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 367, + 295, + 438 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 295, + 438 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 295, + 438 + ], + "type": "text", + "content": "Unlike conventional autoencoders that downsample images by a factor of 8, DCAE [9] progressively increases this to 32, reducing latent tokens by " + }, + { + "bbox": [ + 55, + 367, + 295, + 438 + ], + "type": "inline_equation", + "content": "16 \\times" + }, + { + "bbox": [ + 55, + 367, + 295, + 438 + ], + "type": "text", + "content": ". In our world model, we adopt DCAE for image encoding to improve training efficiency and reduce memory usage, enabling conditioning on longer historical contexts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 440, + 295, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 440, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 440, + 295, + 559 + ], + "type": "text", + "content": "However, as an image autoencoder, DCAE lacks temporal interactions, causing flickering when decoding video frame by frame, which degrades visual quality. To address this, we propose a temporal-aware DCAE to enhance interframe consistency. Specifically, to maximize pretrained parameters while minimizing architectural changes, we introduce spatiotemporal self-attention layers before the DCAE decoder while keeping the encoder fixed during fine-tuning. This facilitates multi-frame interactions, greatly improving temporal consistency in generated videos." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 569, + 131, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 569, + 131, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 569, + 131, + 582 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 187, + 601 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "World Model. Our Epona consists of " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "2.5\\,B" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " parameters, including a 12-layer multimodal spatiotemporal transformer with " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.3\\,B" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " parameters, a 12-layer next-frame prediction diffusion transformer with " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "1.2\\,B" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " parameters, and a 2-layer trajectory planning diffusion transformer with " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "50\\,M" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": " parameters. It is trained on publicly available videos from the NuPlan dataset [7] and 700 scenes from the NuScenes dataset [6] from scratch, all images are resized to " + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "inline_equation", + "content": "512\\times 1024" + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": ". We utilize the rectified flow [37] objective for both video" + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 345, + 324, + 522, + 365 + ], + "blocks": [ + { + "bbox": [ + 313, + 282, + 555, + 315 + ], + "lines": [ + { + "bbox": [ + 313, + 282, + 555, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 555, + 315 + ], + "type": "text", + "content": "Table 2. Inference speed. We evaluate our inference speed for generating a 3-second trajectory and a " + }, + { + "bbox": [ + 313, + 282, + 555, + 315 + ], + "type": "inline_equation", + "content": "512 \\times 1024" + }, + { + "bbox": [ + 313, + 282, + 555, + 315 + ], + "type": "text", + "content": " image per module on a single NVIDIA 4090 GPU." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 345, + 324, + 522, + 365 + ], + "lines": [ + { + "bbox": [ + 345, + 324, + 522, + 365 + ], + "spans": [ + { + "bbox": [ + 345, + 324, + 522, + 365 + ], + "type": "table", + "html": "
DiT sampling stepsMSTTrajDiTVisDiT
10~0.02s~0.03s~0.3s
100~0.02s~0.3s~2s
", + "image_path": "748afd99e6dfb5acd6dfc584d47665e877f88be5837bebf22edaca0dbf4d0bdf.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "text", + "content": "generation and trajectory planning tasks, training the entire model in an end-to-end manner. The training was conducted on 48 NVIDIA A100 GPUs for nearly two weeks, with a total of " + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "inline_equation", + "content": "600\\mathrm{k}" + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "text", + "content": " iterations and a batch size of 96. During training, we apply Chain-of-Forward every 10 steps, each time performing three forward passes. We use the AdamW optimizer with a learning rate of " + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "inline_equation", + "content": "1\\times 10^{-4}" + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "text", + "content": " and set the weight decay to " + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "inline_equation", + "content": "5\\times 10^{-2}" + }, + { + "bbox": [ + 313, + 374, + 555, + 518 + ], + "type": "text", + "content": ". For inference, we report our speed for each module on a single NVIDIA 4090 GPU in Table 2. In all our experiments, we set our DiT sampling step to 100. Notice that with MST and TrajDiT, our world model can seamlessly serve as a real-time motion planner." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 520, + 554, + 628 + ], + "type": "text", + "content": "**Evaluations on video generation.** We employ 1628 video clips from the NuPlan test set [7] and 1646 video clips from the NuScenes validation dataset [6] for performance evaluation, respectively. During the test, our world model is conditioned on 10 consecutive past frames to generate the subsequent frame and repeat the process autoregressively to synthesize future video frames. We use the Frechet Video Distance (FVD) [52] and the Frechet Inception Distance (FID) [22] to evaluate the quality of the generated videos." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "Evaluations on trajectory planning. We evaluate trajectory planning using the NuScenes benchmark [6] and the NAVSIM benchmark [12]. For the NuScenes, we use L2 error and collision rate as the evaluation metrics following the existing works [25, 48, 70] to evaluate the planning performance. L2 error measures the L2 distance between the predicted and ground truth trajectories, while the collision" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27225" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 104, + 69, + 504, + 240 + ], + "blocks": [ + { + "bbox": [ + 104, + 69, + 504, + 240 + ], + "lines": [ + { + "bbox": [ + 104, + 69, + 504, + 240 + ], + "spans": [ + { + "bbox": [ + 104, + 69, + 504, + 240 + ], + "type": "image", + "image_path": "a1ff9fc650f21fae731c82bf5a7210398edff5eb7a907a94e5e6162dbb8090d0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 59, + 240, + 549, + 251 + ], + "lines": [ + { + "bbox": [ + 59, + 240, + 549, + 251 + ], + "spans": [ + { + "bbox": [ + 59, + 240, + 549, + 251 + ], + "type": "text", + "content": "Figure 6. Trajectory-controlled video generation. Our world model can generate controllable videos based on predefined trajectories." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 258, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 305 + ], + "type": "text", + "content": "rate measures the frequency of predicted trajectory intersections with objects. The NAVSIM benchmark assesses performance using the predictive driver model score (PDMS), derived from five factors, as shown in Table 4." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 313, + 228, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 228, + 325 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 228, + 325 + ], + "type": "text", + "content": "4.2. Evaluation of Video Generation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 331, + 296, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 296, + 484 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 296, + 484 + ], + "type": "text", + "content": "Quantitative Comparison of Generated Videos. We present a quantitative comparison with existing methods on the NuScenes dataset [6] in Table 1. Since most methods are not publicly available, we compare with the reported results from their respective papers. Notably, the existing methods (e.g., Vista [18]) fine-tune video diffusion models pretrained on large-scale datasets, while our world model, including next-frame DiT, is trained from scratch. As shown in Table 1, our generated videos achieve state-of-the-art FVD scores, indicating smoother and more realistic video generation quality. Moreover, our world model can generate significantly longer video frames compared to existing approaches as shown in Table 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 581 + ], + "type": "text", + "content": "Qualitative Comparison of Generated Videos. We provide a qualitative comparison with the state-of-the-art open-source driving world model, Vista [18]. Since Vista is a 25-frame fixed-length video diffusion model, we perform rollout to generate longer videos as illustrated in their paper. As shown in Fig. 5, our Epona generates consistent long-horizon driving scenes with high-fidelity visuals and detailed structures and vehicles." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 653 + ], + "type": "text", + "content": "Trajectory-controlled Video Generation. Fig. 6 illustrates the pose controllability of our model. Given the predefined pose trajectory, the different condition frames can generate future frames that conform to the corresponding motion path, which is crucial for obtaining autonomous driving videos under extreme scenarios." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "Extra Long-range Video Generation. Epona combines the strengths of autoregressive and diffusion models, facilitating the generation of high-quality, long-duration videos conditioned on input frames. As shown in Fig.5 and Fig.7, our model can autoregressively generate minute-long driv-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 294 + ], + "type": "text", + "content": "ing videos with high fidelity and consistency, without noticeable drift. More long-term generation videos are provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 497, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 497, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 497, + 313 + ], + "type": "text", + "content": "4.3. Evaluation of Trajectory Planning" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 318, + 555, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 522 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 522 + ], + "type": "text", + "content": "For the NuScenes benchmark [6], we compare our world model with several existing methods, as shown in Table 3. Although our model does not achieve the best results, it attains competitive performance without any additional supervision. It is worth noting that incorporating more supervision typically leads to better performance but cost expensive annotations. Additionally, similar to Doe-1 [70], our model only utilizes the front camera, whereas other methods rely on multi-view inputs for planning. As shown in Table 3, our approach can generate reasonable trajectories while achieving the lowest collision rate for a 1-second horizon, which is crucial for long-term realistic video predictions. For the more challenging NAVSIM benchmark [12], as shown in Table 4, our method achieves state-of-the-art results in overall PDMS when conditioned on the past 2 seconds of observations to predict 4-second future trajectories, showing the strong motion planning capability." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 408, + 541 + ], + "type": "text", + "content": "4.4. Ablation Study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 545, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 700 + ], + "type": "text", + "content": "Effect of Shared Latent for Multi-modal Joint Prediction. To assess the benefit of jointly modeling scene and trajectory via a shared latent representation, we conduct an ablation by disabling video prediction and training the model solely for trajectory prediction. This variant is evaluated on the NAVISIM test set. As shown in Tab. 5, removing video prediction leads to a noticeable drop in planning performance. This result highlights the advantage of the shared latent " + }, + { + "bbox": [ + 313, + 545, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 545, + 555, + 700 + ], + "type": "text", + "content": ", which encourages the world model to better capture complex driving dynamics by leveraging visual signals. This ablation confirms that coupling video and trajectory prediction within a unified latent space in world models significantly benefits downstream planning tasks." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 554, + 713 + ], + "type": "text", + "content": "Effect of Chain-of-Forward Training. To evaluate the" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27226" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 64, + 306, + 155 + ], + "blocks": [ + { + "bbox": [ + 62, + 64, + 306, + 155 + ], + "lines": [ + { + "bbox": [ + 62, + 64, + 306, + 155 + ], + "spans": [ + { + "bbox": [ + 62, + 64, + 306, + 155 + ], + "type": "image", + "image_path": "850e85a12901ccc0803d508487c57e1dd3b8f55520aeac6e5d06a58af680e119.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 163, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 54, + 163, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 54, + 163, + 555, + 196 + ], + "type": "text", + "content": "Figure 7. Qualitative Comparison between long videos generated by models w/ and w/o Chain-of-Forward training. Left: Visual quality deteriorates rapidly after 10–20 seconds. Right: The same driving scenes with Chain-of-Forward training maintain high visual quality, generating minute-long videos without significant degradation. Zoom in for better views." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 306, + 64, + 548, + 155 + ], + "blocks": [ + { + "bbox": [ + 306, + 64, + 548, + 155 + ], + "lines": [ + { + "bbox": [ + 306, + 64, + 548, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 64, + 548, + 155 + ], + "type": "image", + "image_path": "f092074d6c47a2ac4f000ddcf052ca44a23197fc4eaa98dfbe27da2505a8e7a1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 129, + 223, + 478, + 329 + ], + "blocks": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "lines": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 199, + 555, + 222 + ], + "type": "text", + "content": "Table 3. End-to-end motion planning performance on the NuScenes [6] dataset. Note that our model achieves a low collision rate, demonstrating its understanding of basic traffic rules via simple next-frame prediction. * represents only using the front camera as input." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 129, + 223, + 478, + 329 + ], + "lines": [ + { + "bbox": [ + 129, + 223, + 478, + 329 + ], + "spans": [ + { + "bbox": [ + 129, + 223, + 478, + 329 + ], + "type": "table", + "html": "
MethodInputAuxiliary SupervisionL2 (m) ↓Collision Rate (%) ↓
1s2s3sAvg.1s2s3sAvg.
ST-P3 [25]CameraMap & Box & Depth1.332.112.902.110.230.621.270.71
UniAD [48]CameraMap & Box & Motion & Tracklets & Occ0.480.961.651.030.050.170.710.31
OccNet [51]Camera3D-Occ & Map & Box1.292.132.992.140.210.591.370.72
OccWorld [68]Camera3D-Occ0.521.272.411.400.120.402.080.87
VAD-Tiny [30]CameraMap & Box & Motion0.601.232.061.300.310.531.330.72
VAD-Base [30]CameraMap & Box & Motion0.541.151.981.220.040.391.170.53
GenAD [69]CameraMap & Box & Motion0.360.831.550.910.060.231.000.43
Doe-1 [70]Camera*QA0.501.182.111.260.040.371.190.53
OursCamera*None0.611.171.981.250.010.220.850.36
", + "image_path": "2bafe04b0e29b1f4d7512057e238ec6aeadcd5206ac31102adbbe4d9e163fae7.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 57, + 410, + 294, + 488 + ], + "blocks": [ + { + "bbox": [ + 55, + 336, + 296, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 401 + ], + "type": "text", + "content": "Table 4. End-to-end motion planning performance on the NAVSIM [12] test set. NC: no at-fault collision. DAC: drivable area compliance. TTC: time-to-collision. Conf.: comfort. EP: ego progress. PDMS: the predictive driver model score. LAW[35] is in the perception-free setting. Our world model outperforms strong end-to-end planners in the overall PDMS score." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 410, + 294, + 488 + ], + "lines": [ + { + "bbox": [ + 57, + 410, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 57, + 410, + 294, + 488 + ], + "type": "table", + "html": "
MethodInputNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Human/10010010099.987.594.8
UniAD[48]Camera97.891.992.910078.883.4
PARA-Drive[62]Camera97.992.493.099.879.384.6
LAW[35]Camera96.495.488.799.981.784.6
TransFuser[42]Camera & Lidar97.792.892.810079.284.0
DRAMA[67]Camera & Lidar98.093.194.810080.185.5
VADv2[10]Camera & Lidar97.289.191.910076.080.9
OursCamera97.995.193.899.980.486.2
", + "image_path": "d69014850276b8736e8551a70a2419668281efece069bd680158955cd37b11f2.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 57, + 531, + 294, + 564 + ], + "blocks": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "lines": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 295, + 529 + ], + "type": "text", + "content": "Table 5. Comparison of planning results on the NAVSIM test set. Jointly predicting the next scene using a shared latent significantly improves planning performance." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 531, + 294, + 564 + ], + "lines": [ + { + "bbox": [ + 57, + 531, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 531, + 294, + 564 + ], + "type": "table", + "html": "
MethodNC ↑DAC ↑TTC ↑Comf. ↑EP ↑PDMS ↑
Ours w/o Joint Training94.589.788.199.974.778.1
Ours97.995.193.899.980.486.2
", + "image_path": "3c3f61fc5472191db839e71cc042aec3e3039c95f0fefe5c69dd275366529988.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "impact of our chain-of-forward strategy on model performance, we conduct an ablation study comparing results with and without this strategy. Given that our model iteratively generates the next frame, the chain-of-forward approach simulates potential inference errors during training, thereby enhancing the model's robustness. As shown in Fig. 7 and Fig. 8, as the model autoregressively generates longer sequences, the gap in visual quality and FID score between models with and without the chain-of-forward strategy becomes increasingly significant, validating its effectiveness in long-term video generation. More ablations" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 331, + 346, + 533, + 476 + ], + "blocks": [ + { + "bbox": [ + 331, + 346, + 533, + 476 + ], + "lines": [ + { + "bbox": [ + 331, + 346, + 533, + 476 + ], + "spans": [ + { + "bbox": [ + 331, + 346, + 533, + 476 + ], + "type": "image", + "image_path": "e022709c0de5912cf52a031faa7a4ee59146b569c0f7162a0de4581b23bef58e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 488, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 488, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 555, + 521 + ], + "type": "text", + "content": "Figure 8. Effect of Chain-of-Forward training. FID comparison in NuPlan test set between models w/ and w/o Chain-of-Forward training strategy." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 528, + 493, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 493, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 493, + 540 + ], + "type": "text", + "content": "can be found in the supplementary material." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 388, + 562 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 570, + 555, + 715 + ], + "type": "text", + "content": "We have presented Epona, an autoregressive diffusion world model for autonomous driving that jointly predicts high-fidelity future trajectories and driving scenes based on historical driving context. Thanks to our proposed decoupled spatiotemporal modeling and asynchronous multimodal generation strategies, our model achieves high-quality and long-term prediction. In addition, our model could serves as a real-time motion planner via performing trajectory planning. We have demonstrated that our approach significantly advances the state of the art in driving world models, uncovering the large potential for building next-generation autonomous driving systems." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27227" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chattopadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai. arXiv preprint arXiv:2501.03575, 2025. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 168 + ], + "type": "text", + "content": "[2] Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 224 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "spans": [ + { + "bbox": [ + 62, + 225, + 295, + 269 + ], + "type": "text", + "content": "[4] A. Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. CVPR, pages 22563-22575, 2023. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 270, + 296, + 334 + ], + "type": "text", + "content": "[5] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in Neural Information Processing Systems (NeurIPS), 33:1877-1901, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 335, + 296, + 402 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 296, + 402 + ], + "type": "text", + "content": "[6] Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621-11631, 2020. 2, 6, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 403, + 296, + 456 + ], + "type": "text", + "content": "[7] Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 6, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 458, + 296, + 501 + ], + "type": "text", + "content": "[8] Boyuan Chen, Diego Marti Monso, Yilun Du, Max Simchowitz, Russ Tedrake, and Vincent Sitzmann. Diffusion forcing: Next-token prediction meets full-sequence diffusion, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 296, + 546 + ], + "type": "text", + "content": "[9] Junyu Chen, Han Cai, Junsong Chen, Enze Xie, Shang Yang, Haotian Tang, Muyang Li, Yao Lu, and Song Han. Deep compression autoencoder for efficient high-resolution diffusion models. arXiv preprint arXiv:2410.10733, 2024. 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 548, + 295, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 548, + 295, + 601 + ], + "spans": [ + { + "bbox": [ + 57, + 548, + 295, + 601 + ], + "type": "text", + "content": "[10] Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243, 2024.8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 602, + 295, + 646 + ], + "type": "text", + "content": "[11] Yuntao Chen, Yuqi Wang, and Zhaoxiang Zhang. Driving-gpt: Unifying driving world modeling and planning with multi-modal autoregressive transformers. arXiv preprint arXiv:2412.18607, 2024. 2, 3, 4, 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[12] Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. In NeurIPS, 2024. 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[13] Jingtao Ding, Yunke Zhang, Yu Shang, Yuheng Zhang, Zefang Zong, Jie Feng, Yuan Yuan, Hongyuan Su, Nian Li, Nicholas Sukiennik, et al. Understanding world or predicting future? a comprehensive survey of world models. arXiv preprint arXiv:2411.14499, 2024. 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 152 + ], + "type": "text", + "content": "[14] Patrick Esser, Robin Rombach, and Björn Ommer. Taming transformers for high-resolution image synthesis, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 218 + ], + "type": "text", + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[16] Ruiyuan Gao, Kai Chen, Bo Xiao, Lanqing Hong, Zhenguo Li, and Qiang Xu. Magicdrivedit: High-resolution long video generation for autonomous driving with adaptive control. arXiv preprint arXiv:2411.13807, 2024. 2, 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "text", + "content": "[17] Ruiyuan Gao, Kai Chen, Enze Xie, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung, and Qiang Xu. MagicDrive: Street view generation with diverse 3d geometry control. In ICLR, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 363 + ], + "type": "text", + "content": "[18] Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024. 2, 4, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 408 + ], + "type": "text", + "content": "[19] Songen Gu, Wei Yin, Bu Jin, Xiaoyang Guo, Junming Wang, Haodong Li, Qian Zhang, and Xiaoxiao Long. Dome: Taming diffusion model into high-fidelity controllable occupancy world model. ArXiv, abs/2410.10429, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 442 + ], + "type": "text", + "content": "[20] Xi Guo, Chenjing Ding, Haoxuan Dou, Xin Zhang, Weixuan Tang, and Wei Wu. Infinitydrive: Breaking time limits in driving world models, 2024. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 553, + 466 + ], + "type": "text", + "content": "[21] David Ha and Jürgen Schmidhuber. World models. arXiv preprint arXiv:1803.10122, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 521 + ], + "type": "text", + "content": "[22] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a two time-scale update rule converge to a local nash equilibrium. Advances in neural information processing systems, 30, 2017. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 567 + ], + "type": "text", + "content": "[23] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, pages 6840-6851. Curran Associates, Inc., 2020. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 568, + 553, + 622 + ], + "type": "text", + "content": "[24] Anthony Hu, Lloyd Russell, Hudson Yeo, Zak Murez, George Fedoseev, Alex Kendall, Jamie Shotton, and Gianluca Corrado. Gaia-1: A generative world model for autonomous driving. arXiv preprint arXiv:2309.17080, 2023. 2, 3, 4, 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 678 + ], + "type": "text", + "content": "[25] Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based autonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533-549. Springer, 2022. 6, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] Xiaotao Hu, Wei Yin, Mingkai Jia, Junyuan Deng, Xiaoyang Guo, Qian Zhang, Xiaoxiao Long, and Ping Tan. Drivingworld: Constructingworld model for autonomous driving via" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27228" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "video gpt. arXiv preprint arXiv:2412.19505, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 294, + 118 + ], + "type": "text", + "content": "[27] Tencent Hunyuan. Hunyuanvideo: A systematic framework for large video generative models, 2024. 4, 5, 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 162 + ], + "type": "text", + "content": "[28] Fan Jia, Weixin Mao, Yingfei Liu, Yucheng Zhao, Yuqing Wen, Chi Zhang, Xiangyu Zhang, and Tiancai Wang. Adriver-i: A general world model for autonomous driving, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 294, + 217 + ], + "type": "text", + "content": "[29] Mingkai Jia, Wei Yin, Xiaotao Hu, Jiaxin Guo, Xiaoyang Guo, Qian Zhang, Xiao-Xiao Long, and Ping Tan. Mgvq: Could vq-vae beat vae? a generalizable tokenizer with multi-group quantization. arXiv preprint arXiv:2507.07997, 2025.3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[30] Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340-8350, 2023. 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 286, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 286, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 57, + 286, + 294, + 319 + ], + "type": "text", + "content": "[31] Jihwan Kim, Junoh Kang, Jinyoung Choi, and Bohyung Han. Fifo-diffusion: Generating infinite videos from text without training. In NeurIPS, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 374 + ], + "type": "text", + "content": "[32] Seung Wook Kim, Jonah Philion, Antonio Torralba, and Sanja Fidler. Drivegan: Towards a controllable high-quality neural simulation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5820-5829, 2021. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 376, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 376, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 57, + 376, + 294, + 398 + ], + "type": "text", + "content": "[33] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2024. 2, 4, 5, 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 399, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 399, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 399, + 294, + 430 + ], + "type": "text", + "content": "[34] Yann LeCun. A path towards autonomous machine intelligence version 0.9.2, 2022-06-27. Open Review, 62(1):1-62, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 433, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 433, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 57, + 433, + 294, + 465 + ], + "type": "text", + "content": "[35] Yingyan Li, Lue Fan, Jiawei He, Yuqi Wang, Yuntao Chen, Zhaoxiang Zhang, and Tieniu Tan. Enhancing end-to-end autonomous driving with latent world model. 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 468, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 468, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 57, + 468, + 294, + 499 + ], + "type": "text", + "content": "[36] Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling, 2023. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[37] Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. arXiv preprint arXiv:2209.03003, 2022. 2, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 590 + ], + "type": "text", + "content": "[38] Jiachen Lu, Ze Huang, Jiahui Zhang, Zeyu Yang, and Li Zhang. WoVoGen: World Volume-Aware Diffusion for Controllable Multi-Camera Driving Scene Generation. arXiv preprint arXiv:2312.02934, 2023. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 294, + 624 + ], + "type": "text", + "content": "[39] Xin Ma, Yaohui Wang, Gengyun Jia, Xinyuan Chen, Ziwei Liu, Yuan-Fang Li, Cunjian Chen, and Yu Qiao. Latte: Latent diffusion transformer for video generation, 2024. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 689 + ], + "type": "text", + "content": "[40] Yiyang Ma, Xingchao Liu, Xi aokang Chen, Wen Liu, Chengyue Wu, Zhiyu Wu, Zizheng Pan, Zhenda Xie, Haowei Zhang, Xingkai Yu, Liang Zhao, Yisong Wang, Jiaying Liu, and Chong Ruan. Janusflow: Harmonizing autoregression and rectified flow for unified multimodal understanding and generation. ArXiv, abs/2411.07975, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 713 + ], + "type": "text", + "content": "[41] William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF inter" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "national conference on computer vision, pages 4195-4205, 2023. 2, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[42] Aditya Prakash, Kashyap Chitta, and Andreas Geiger. Multimodal fusion transformer for end-to-end autonomous driving. In CVPR, 2021. 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 132, + 553, + 164 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 553, + 164 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 553, + 164 + ], + "type": "text", + "content": "[43] Haonan Qiu, Menghan Xia, Yong Zhang, Yingqing He, Xintao Wang, Ying Shan, and Ziwei Liu. Freenoise: Tuning-free longer video diffusion via noise rescheduling, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 166, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 166, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 316, + 166, + 553, + 199 + ], + "type": "text", + "content": "[44] Alec Radford, Karthik Narasimhan, Tim Salimans, and Ilya Sutskever. Improving language understanding by generative pre-training. OpenAI blog, 2018. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 233 + ], + "type": "text", + "content": "[45] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 236, + 553, + 268 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 553, + 268 + ], + "type": "text", + "content": "[46] Alex Rogozhnikov. Einops: Clear and reliable tensor manipulations with einstein-like notation. In International Conference on Learning Representations, 2022. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 553, + 302 + ], + "type": "text", + "content": "[47] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models, 2022. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 346 + ], + "type": "text", + "content": "[48] Xiaogang Shi, Bin Cui, Gillian Dobbie, and Beng Chin Ooi. Uniad: A unified ad hoc data processing system. ACM Transactions on Database Systems (TODS), 42(1):1-42, 2016. 6, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 382 + ], + "type": "text", + "content": "[49] Kiwhan Song, Boyuan Chen, Max Simchowitz, Yilun Du, Russ Tedrake, and Vincent Sitzmann. History-guided video diffusion, 2025. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 384, + 553, + 428 + ], + "type": "text", + "content": "[50] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2, 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 430, + 553, + 484 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 484 + ], + "type": "text", + "content": "[51] Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, et al. Scene as occupancy. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 8406-8415, 2023. 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 316, + 487, + 553, + 531 + ], + "type": "text", + "content": "[52] Thomas Unterthiner, Sjoerd Van Steenkiste, Karol Kurach, Raphael Marinier, Marcin Michalski, and Sylvain Gelly. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717, 2018. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 533, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 533, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 316, + 533, + 553, + 564 + ], + "type": "text", + "content": "[53] Dani Valevski, Yaniv Leviathan, Moab Arar, and Shlomi Fruchter. Diffusion models are real-time game engines, 2024. 3, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 553, + 609 + ], + "type": "text", + "content": "[54] Aaron van den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. In Advances in Neural Information Processing Systems. Curran Associates, Inc., 2017. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 656 + ], + "type": "text", + "content": "[55] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[56] Ruben Villegas, Mohammad Babaeizadeh, Pieter-Jan Kindermans, Hernan Moraldo, Han Zhang, Mohammad Taghi Saffar, Santiago Castro, Julius Kunze, and Dumitru Erhan. Phenaki: Variable length video generation from open domain textual description, 2022. 3" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27229" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[57] Fu-Yun Wang, Wenshuo Chen, Guanglu Song, Han-Jia Ye, Yu Liu, and Hongsheng Li. Gen-1-video: Multi-text to long video generation via temporal co-denoising. arXiv preprint arXiv:2305.18264, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[58] Fu-Yun Wang, Zhaoyang Huang, Qiang Ma, Guanglu Song, Xudong Lu, Weikang Bian, Yijin Li, Yu Liu, and Hongsheng Li. Zola: Zero-shot creative long animation generation with short video model. In ECCV, 2024. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[59] Xiaofeng Wang, Zheng Zhu, Guan Huang, Xinze Chen, Jiagang Zhu, and Jiwen Lu. Drivedreamer: Towards real-world-driven world models for autonomous driving. arXiv preprint arXiv:2309.09777, 2023. 2, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[60] Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the Future: Multiview Visual Forecasting and Planning with World Model for Autonomous Driving. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[61] Yuqing Wang, Tianwei Xiong, Daquan Zhou, Zhijie Lin, Yang Zhao, Bingyi Kang, Jiashi Feng, and Xihui Liu. Loong: Generating minute-level long videos with autoregressive language models, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[62] Xinshuo Weng, Boris Ivanovic, Yan Wang, Yue Wang, and Marco Pavone. Para-drive: Parallelized architecture for real-time autonomous driving. In CVPR, 2024. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "text", + "content": "[63] Yanhao Wu, Haoyang Zhang, Tianwei Lin, Lichao Huang, Shujie Luo, Rui Wu, Congpei Qiu, Wei Ke, and Tong Zhang. Generating multimodal driving scenes via next-scene prediction. In CVPR, 2025. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "text", + "content": "[64] Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goaldriven flow matching for multimodal trajectories generation in end-to-end autonomous driving. ArXiv, abs/2503.05689, 2025. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 294, + 488 + ], + "type": "text", + "content": "[65] Jiazhi Yang, Shenyuan Gao, Yihang Qiu, Li Chen, Tianyu Li, Bo Dai, Kashyap Chitta, Penghao Wu, Jia Zeng, Ping Luo, Jun Zhang, Andreas Geiger, Yu Qiao, and Hongyang Li. Generalized Predictive Model for Autonomous Driving. In CVPR, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 294, + 555 + ], + "type": "text", + "content": "[66] Shengming Yin, Chenfei Wu, Huan Yang, Jianfeng Wang, Xiaodong Wang, Minheng Ni, Zhengyuan Yang, Linjie Li, Shuguang Liu, Fan Yang, Jianlong Fu, Gong Ming, Lijuan Wang, Zicheng Liu, Houqiang Li, and Nan Duan. Nuwa-xl: Diffusion over diffusion for extremely long video generation, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[67] Chengran Yuan, Zhanqi Zhang, Jiawei Sun, Shuo Sun, Zefan Huang, Christina Dao Wen Lee, Dongen Li, Yuhang Han, Anthony Wong, Keng Peng Tee, and Marcelo H. Ang Jr au2. Drama: An efficient end-to-end motion planner for autonomous driving with mamba, 2024. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[68] Wenzhao Zheng, Weiliang Chen, Yuanhui Huang, Borui Zhang, Yueqi Duan, and Jiwen Lu. Occworld: Learning a 3d occupancy world model for autonomous driving. In European conference on computer vision, pages 55-72. Springer, 2024. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 714 + ], + "type": "text", + "content": "[69] Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end autonomous driving. In European Conference on Computer Vision, pages 87-104. Springer, 2024. 8" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 172 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 116 + ], + "type": "text", + "content": "[70] Wenzhao Zheng, Zetian Xia, Yuanhui Huang, Sicheng Zuo, Jie Zhou, and Jiwen Lu. Doe-1: Closed-loop autonomous driving with large world model. arXiv preprint arXiv:2412.09627, 2024. 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 172 + ], + "type": "text", + "content": "[71] Chunting Zhou, Lili Yu, Arun Babu, Kushal Tirumala, Michihiro Yasunaga, Leonid Shamis, Jacob Kahn, Xuezhe Ma, Luke S. Zettlemoyer, and Omer Levy. Transfusion: Predict the next token and diffuse images with one multi-modal model. ArXiv, abs/2408.11039, 2024. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "27230" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_content_list.json b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c37a36287a5a4acdf08d6ae2a379cf38c5a2a17d --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_content_list.json @@ -0,0 +1,1458 @@ +[ + { + "type": "text", + "text": "EquiCaps: Predictor-Free Pose-Aware Pre-Trained Capsule Networks", + "text_level": 1, + "bbox": [ + 145, + 130, + 851, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Athinoula Konstantinou1 Georgios Leontidis1 \n1University of Aberdeen, UK", + "bbox": [ + 125, + 179, + 529, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Mamatha Thota² Aiden Durrant¹ \n²University of Lincoln, UK", + "bbox": [ + 552, + 180, + 870, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{a.konstantinou.24, georgios.leontidis, aiden.durrant}@abdn.ac.uk, mthota@lincoln.ac.uk", + "bbox": [ + 111, + 218, + 879, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Learning self-supervised representations that are invariant and equivariant to transformations is crucial for advancing beyond traditional visual classification tasks. However, many methods rely on predictor architectures to encode equivariance, despite evidence that architectural choices, such as capsule networks, inherently excel at learning interpretable pose-aware representations. To explore this, we introduce EquiCaps (Equivariant Capsule Network), a capsule-based approach to pose-aware self-supervision that eliminates the need for a specialised predictor for enforcing equivariance. Instead, we leverage the intrinsic pose-awareness capabilities of capsules to improve performance in pose estimation tasks. To further challenge our assumptions, we increase task complexity via multi-geometric transformations to enable a more thorough evaluation of invariance and equivariance by introducing 3DIEBench-T, an extension of a 3D object-rendering benchmark dataset. Empirical results demonstrate that EquiCaps outperforms prior state-of-the-art equivariant methods on geometric tasks, including rotation and translation, achieving a supervised-level $R^2$ of 0.78 on the 3DIEBench rotation prediction benchmark and improving upon SIE and CapsIE by 0.05 and 0.04 $R^2$ , respectively. Moreover, in contrast to non-capsule-based equivariant approaches, EquiCaps maintains robust equivariant performance under combined geometric transformations, underscoring its generalisation capabilities and the promise of predictor-free capsule architectures. Code, dataset, and weights are released at http://github.com/AberdeenML/EquiCaps.", + "bbox": [ + 89, + 300, + 483, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 768, + 220, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of self-supervised learning (SSL) is largely attributed to the exploitation of invariant representations under data augmentations and transformations to learn rich semantic concepts [3, 4, 7, 8, 10, 12, 27, 31, 32, 62]. However, methods that strictly enforce invariance fail to preserve useful augmentation and transformation information, which could limit their applicability to downstream tasks where", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/caa2dbd19f946af1435a4c2980b43c2ef4537b2cbc3c342357c616c86a7cad4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 535, + 267, + 885, + 452 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64e325b9400dc2d2d0ac14b623fe1f3cb892e9d6cdfed833baeec4899340521c.jpg", + "image_caption": [ + "Figure 1. Equivariant evaluation of EquiCaps (ours) and competing methods on 3DIEBench benchmark. We indicate which subset (train or val) is used to generate embeddings, and, for PRE, which dataset (train, val, or all) is used for retrieval." + ], + "image_footnote": [], + "bbox": [ + 534, + 458, + 885, + 561 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "such information is vital [28, 40, 58]. In response, there has been increasing interest in developing more generalised self-supervised approaches, most notably through the investigation of pose-aware learning [24, 54, 56]. Achieving such generalisation, particularly to novel viewpoints, can consequently address real-world settings that inherently exhibit large data variations, e.g. autonomous driving [54].", + "bbox": [ + 509, + 657, + 906, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One promising strategy to preserve transformation information—and thus achieve pose-awareness—is to directly optimise for equivariant representations, enabling the model to capture how transformations steer the latent space. However, such methods often face limitations [19, 23, 24, 46]. First, they rely on a dedicated predictor network to encode equivariant properties, often further increasing architectural complexity through conditional projectors, partitioned representations, or hypernet", + "bbox": [ + 509, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7947", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "works. Second, equivariance in representations is typically enforced solely via objective functions with little regard taken to exploit architectural mechanisms—such as Capsule Networks (CapsNets)—that have been shown to effectively encode or structure pose information [15].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose to simplify and enhance the framework of equivariant self-supervision to address these concerns by using CapsNets [15]. CapsNets structure representations into capsules, each encoding both the existence of an entity and its instantiation parameters, where the latter is achieved through a routing algorithm based on agreement of part-whole relationships. This formulation introduces an inductive bias that has been shown to effectively capture equivariance w.r.t. viewpoints and invariance in the capsule activations, enabling them to exhibit viewpoint-invariant and viewpoint-equivariant properties by design.", + "bbox": [ + 88, + 167, + 482, + 332 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While prior attempts have been made to introduce CapsNets into equivariant SSL [23], they still rely on a separate transformation predictor, potentially underusing capsules' intrinsic pose-awareness capabilities. To overcome these limitations, we propose a streamlined pipeline that eliminates the predictor and directly leverages capsules' equivariant properties that prior equivariant SSL methods do not. CapsNets offer an intriguing prospect for capturing transformations, therefore we introduce an extension of the 3DIEBench dataset [24] that incorporates translational transformations in addition to rotations. By moving from rotation-only SO(3) transformations [24, 54] to the more general SE(3) group, our dataset increases the difficulty of both invariance and equivariance tasks, allowing for the evaluation of generalisation under multi-geometric settings.", + "bbox": [ + 91, + 333, + 483, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarise, our contributions are as follows:", + "bbox": [ + 109, + 559, + 424, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- 3DIEBench-T: We present 3DIEBench-T, an extension of the 3DIEBench that integrates translational transformations alongside rotations. This new benchmark enables the evaluation of representation methods under simultaneous multi-geometric 3D transformations.", + "- Predictor limitations: We empirically demonstrate that existing predictor-based equivariant architectures yield minimal improvement over purely invariant approaches when learning multi-geometric 3D transformations.", + "- Predictor-free CapsNet: We propose EquiCaps, removing the need for predictors by exploiting the intrinsic pose-awareness of CapsNets. Empirical results on rotation and translation tasks demonstrate that EquiCaps outperforms existing equivariant methods, and generalises successfully under multi-geometric settings, while also allowing intuitive manipulation of the embedding space." + ], + "bbox": [ + 89, + 575, + 482, + 818 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 829, + 228, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We first introduce key notions of equivariance and invariance in our setting [24, 46].", + "bbox": [ + 89, + 854, + 482, + 886 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Group actions. A group is a set $G$ with a binary operation $\\cdot$ that satisfies associativity, $G \\times G \\to G$ , the existence of an identity element, $g \\cdot 1 = 1 \\cdot g = g$ such that $\\forall g \\in G$ , and the existence of an inverse, $g \\cdot g^{-1} = g^{-1} \\cdot g = 1$ . A group action of $G$ on a set $S$ is defined as a map $\\alpha : G \\times S \\to S$ that is compatible with the composition of group elements, i.e., $\\alpha(g, \\alpha(h, s)) = \\alpha(gh, s)$ , $\\forall s \\in S$ and $\\forall g, h \\in G$ . The map $\\alpha(g, \\cdot)$ is called a group representation when $S$ is a vector space $V$ , such as $\\mathbb{R}^n$ , and $\\alpha$ is linear. A group representation is denoted as the map $\\rho : G \\to GL(V)$ , such that $\\rho(g) = \\alpha(g, \\cdot)$ . As in similar works [24], we use group representations to describe how transformations are applied both to the input data and latent space.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Equivalence learning. A function $f: X \\to Y$ is equivariant if for a group $G$ , with representations $\\rho_{X}$ and $\\rho_{Y}$ on $X$ and $Y$ , satisfies", + "bbox": [ + 511, + 287, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\rho_ {X} (g) \\cdot x\\right) = \\rho_ {Y} (g) \\cdot f (x), \\forall x \\in X, \\forall g \\in G. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 540, + 343, + 903, + 359 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This indicates the commutativity of group transformations and function application. The function $f$ retains the symmetry group $G$ yet changes how the group operates.", + "bbox": [ + 511, + 369, + 903, + 415 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Invariance learning. A function $f: X \\to Y$ is invariant if for a group $G$ , with representations $\\rho_{X}$ on $X$ , satisfies", + "bbox": [ + 511, + 416, + 903, + 446 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf \\left(\\rho_ {X} (g) \\cdot x\\right) = f (x), \\quad \\forall x \\in X, \\forall g \\in G. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 457, + 903, + 473 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Invariance is a trivial [14], special case of equivariance when $\\rho_Y(g)$ , in Eq. (1) is the identity map.", + "bbox": [ + 511, + 484, + 903, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Related works", + "text_level": 1, + "bbox": [ + 511, + 527, + 658, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Invariant SSL. Invariance to data augmentations has been the driving force of self-supervision, using this property to learn common semantic concepts under different transformations when there is no access to labels. Specifically, invariance in the latent space is enforced by aligning shared views that are transformations of the same source. There are various approaches to undertake this alignment and discrimination, notably contrastive methods [10, 12, 13, 31, 60], clustering-based learning [6-8, 45], and non-contrastive methods [3, 4, 11, 22, 27, 62] which instead employ architectural or objective modifications, gradient mechanisms, and information-theoretic regularisers.", + "bbox": [ + 511, + 553, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Implicit equivariant SSL. Invariant SSL methods are optimised to disregard data transformations, hence information is lost. To address this, predictive equivariant SSL encodes transformation-sensitive information by training to predict transformation parameters [25, 40, 50], while others integrate contrastive learning with transformation prediction [14]. Furthermore, modulating the contrastive objective itself, based on the strength of the applied augmentation has shown promise [59], while [58] introduces distinct embedding spaces that remain invariant to all but one augmentation. Focussing on the transformation, optimal augmented", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/e3362aef672282506dea597c1e511667fcfa0d5ae1dd875df349f6f48eaee286.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 117, + 93, + 519, + 224 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/ded19a7e06175cdd94f8cebfe06aa3e6a31eed696eb1836a5dfc351d3d8ee4b3.jpg", + "image_caption": [ + "(b)", + "Figure 2. (a) Framework Overview. Two views $x_{1}$ and $x_{1}^{\\prime} = \\rho_{X}(g_{rel}) \\cdot x_{1}$ of the same object are processed by a shared CNN encoder $f_{\\theta}$ and a capsule-based projector $h_{\\phi}$ . (b) Projector's Output. For each view, the projector produces an activation probability vector $(z_{act}, z_{act}')$ , indicating entity presence, and a pose matrix $(z_{pose}, z_{pose}')$ , encoding the pose for each entity. (c) Equivariant Objective. Equivariance is enforced by multiplying $z_{pose}$ by the relative transformation matrix $\\rho_{Y}(g_{rel})$ and aligning the result with $z_{pose}'$ ." + ], + "image_footnote": [], + "bbox": [ + 519, + 97, + 676, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/d6b3920457ede7f4cb24d3f97ac2254c2c23d40dc41ae4d581051579bf583c4d.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 679, + 98, + 890, + 219 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "views can also be exploited to retain task-relevant information [53]. However, these do not optimise for equivariance defined in Eq. (1). Hence, since they do not learn an explicit mapping from input transformations to corresponding changes in latent space, they do not guarantee equivariance, unlike explicit equivariant approaches [24, 56].", + "bbox": [ + 88, + 337, + 480, + 428 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explicit equivariant SSL. Avoiding transformation prediction tasks in favour of evaluating for equivariant representations, elicits a separate family of methods. These utilise representation reconstruction and maintain invariant and equivariant properties by splitting into or learning separate latent spaces [44, 57]. Others [51, 61] encourage equivariance through identically transformed image pairs, with [51] utilising the transformation group information for choosing optimal loss terms. Such approaches have also extended contrastive learning enforcing input transformations correspond to rotations in the embedding space [28]. Several methods adopt a predictor network to enforce an equivalent transformation mapping between latent representations through augmentation parameters [19, 46]. Building on these, SIE [24] incorporates a hypernetwork to learn such mapping, while dividing the representation into invariant and equivariant parts. CapsIE [23] adopts this module, instead leveraging a CapsNet projector, removing the need for multiple projectors and the split representation.", + "bbox": [ + 91, + 430, + 482, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CapsNets. CapsNets are designed to model part-whole hierarchical relationships by capturing the likelihood of an entity's presence and its attributes (pose, colour, texture, etc.) through neural activity. Lower-level capsules connect to higher-level ones via routing based on agreement in likelihood and positional attributes. This alignment enables capsules to identify objects by reinforcing pose consistency when multiple capsules converge on the same entity. Due to this positional agreement mechanism, CapsNets are explored as equivariant architectures, enabling tasks like image generation [33] and transformation learning [48]. Various routing methods have been proposed, including", + "bbox": [ + 89, + 719, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Gaussian-mixture clustering [34], Variational Bayes with uncertainty [47], self-routing via subordinate networks [29], and residual pose routing [42]. Like our approach, CapsNets have been adopted in Siamese SSL, notably by predicting canonical poses for 3D point clouds [52], or through optical flow analysis using successive video frames [49]. They have also been applied to video tasks, for example, by transforming 2D convolutional capsule routing to 3D [20] and by adopting attention-based routing [21]. Lastly, CapsNets share foundational concepts with slot-based object-centric methods [15], such as [43], which uses an iterative attention mechanism akin to capsule routing, whereas the slots in [41] specialise in distinct entities as capsules.", + "bbox": [ + 511, + 337, + 906, + 534 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. EquiCaps: Predictor-Free Pose-Aware SSL", + "text_level": 1, + "bbox": [ + 511, + 553, + 897, + 571 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Motivation. Prior equivariant self-supervised methods [19, 24, 46] commonly rely on predictor networks to enforce equivariant representations via augmentation-aware mappings. This leads to ad hoc vector representations that are both difficult to interpret or manipulate—an important consideration for evaluating equivariance. In effect, these methods attempt to enforce equivariance on architectures (i.e., CNNs) that are not naturally suited to it [15] and often introduce extra complexity through dual projectors, representation splits, or hypernetworks. Instead, we exploit the equivariant properties elicited by CapsNets [15], to more effectively encode positional information, particularly rotation and translation. Our design yields a simpler pose-aware SSL framework, outperforms previous equivariant methods on geometric tasks, and enables a more intuitive control and interpretability in the embedding space through the structured $4 \\times 4$ capsule pose matrices. We introduce and rationalise each component of our methodology as follows.", + "bbox": [ + 511, + 580, + 906, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Framework overview. Our method adopts a Siamese joint embedding setup with identical, weight-sharing branches, each formed by a CNN $f_{\\theta}$ and a CapsNet projec-", + "bbox": [ + 511, + 854, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "tor $h_{\\phi}$ . We omit global average pooling to preserve the 2D spatial structure necessary for capsule processing. Given an image $x \\in \\mathbb{R}^{c \\times h \\times w}$ from a dataset $D$ , we sample two transformations $g_{1}, g_{2}$ from a distribution over the group $G$ , as detailed in supplementary Sec. 14. These are applied to $x$ to produce two distinct views, $x_{g_1} = \\rho_X(g_1) \\cdot x$ and $x_{g_2} = \\rho_X(g_2) \\cdot x$ . Both views are passed through $f_{\\theta}$ , producing feature maps $F, F' \\in \\mathbb{R}^{c \\times h \\times w}$ . These are then processed by $h_{\\phi}$ , which contains $N$ capsules, yielding $z_{\\mathrm{act}}, z_{\\mathrm{pose}} = h_{\\phi}(F)$ and $z_{\\mathrm{act}}', z_{\\mathrm{pose}}' = h_{\\phi}(F')$ , where $z_{\\mathrm{act}}, z_{\\mathrm{act}}' \\in \\mathbb{R}^N$ are the activation probabilities, and $z_{\\mathrm{pose}}, z_{\\mathrm{pose}}' \\in \\mathbb{R}^{N \\times 4 \\times 4}$ are the corresponding pose matrices. The activation vectors encode transformation-invariant properties, representing semantic concepts that remain unchanged under $g_{1}, g_{2}$ . In contrast, the pose matrices capture transformation-equivariant properties, encoding an entity's pose, where each capsule represents a unique entity. A visual depiction of the general architecture is given in Fig. 2.", + "bbox": [ + 89, + 90, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Self-routing CapsNet projector. To reduce the computational overhead of iterative-routing, we adopt the noniterative self-routing algorithm [29] as our projector. The projector $h_{\\phi}$ consists of two layers: a primary capsule layer that transforms encoder feature maps into poses $(u_i)$ and activations $(a_i)$ , and a self-routing layer, where each capsule independently determines its routing coefficients via a subordinate single-layer perceptron. To do this, the routing coefficients $c_{ij}$ between each lower-layer capsule $i$ and each upper-level capsule $j$ are computed by multiplying the lower capsule's pose vector $u_i$ with a learnable weight matrix $W^{\\mathrm{route}}$ and applying a softmax function. The activation $a_j$ of each upper-level capsule is then computed by forming votes from lower-level capsules—each vote is the product $c_{ij}a_i$ , which are then summed and normalised by the total lower-level activation, to formulate weighted votes:", + "bbox": [ + 89, + 364, + 483, + 604 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc _ {i j} = \\operatorname {s o f t m a x} \\left(W _ {i} ^ {\\text {r o u t e}} u _ {i}\\right) _ {j}, \\quad a _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}{\\sum_ {i \\in \\Omega_ {l}} a _ {i}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 609, + 480, + 646 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The capsule layer's output pose is computed by multiplying a trainable weight matrix $W^{\\text{pose}}$ with each lower-level pose $u_i$ , yielding a predicted pose $\\hat{u}_{j|i}$ for each upper-level capsule. The final pose $u_j$ is the weighted average of $\\hat{u}_{j|i}$ :", + "bbox": [ + 89, + 654, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {u} _ {j \\mid i} = W _ {i j} ^ {\\text {p o s e}} u _ {i}, \\quad u _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i} \\hat {u} _ {j \\mid i}}{\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 719, + 480, + 756 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Enforcing equivariance. We define the relative transformation $g_{\\mathrm{rel}} = g_2g_1^{-1}$ , mapping between the views $x_{g_1}$ and $x_{g_2}$ , and let $\\rho_Y(g_{\\mathrm{rel}})$ denote its matrix representation. To enforce equivariance (defined in Eq. (1)) we multiply $z_{\\mathrm{pose}}$ by $\\rho_Y(g_{\\mathrm{rel}})$ , aligning the transformed result with $z_{\\mathrm{pose}}'$ . This indicates that the pose embeddings are equivariant w.r.t. the transformation group $G$ , ensuring that geometric changes in the input are consistently captured in the latent space. Each pose matrix $(z_{\\mathrm{pose}}\\cdot \\rho_Y(g_{\\mathrm{rel}}))$ and $z_{\\mathrm{pose}}'$", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "is normalised by its Frobenius norm and residing on a unit hypersphere. This direct manipulation in the latent space removes the need for a dedicated predictor module.", + "bbox": [ + 511, + 90, + 903, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Loss functions. To enforce invariance we minimise the cross-entropy $H(Z_{\\mathrm{act}}, Z_{\\mathrm{act}}^{\\prime})$ between the activation probability vectors $Z_{\\mathrm{act}}$ and $Z_{\\mathrm{act}}^{\\prime}$ , where $Z$ denotes the matrix embeddings computed over a batch. This aligns the capsule activations for different views of the same image. To ensure full capsule utilisation, we apply mean entropy maximisation (ME-MAX) regularisation [1, 2, 36] to the activation probabilities. Let $\\bar{Z}_{\\mathrm{act}} = \\frac{1}{B}\\sum_{i=1}^{B}Z_{\\mathrm{act}}$ , where $B$ is the batch size. Maximising $H(Z_{\\mathrm{act}})$ and $H(\\bar{Z}_{\\mathrm{act}}^{\\prime})$ encourages a diverse distribution of activation across all capsules, preventing the collapse to only a few highly active capsules. To enforce equivariance we minimise the mean-squared error between the normalised, transformed pose $(Z_{\\mathrm{pose}}\\cdot\\rho_{Y}(g_{\\mathrm{rel}}))$ and the corresponding pose $Z_{\\mathrm{pose}}^{\\prime}$ .", + "bbox": [ + 511, + 136, + 905, + 348 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To prevent collapse, we apply the variance-covariance regularisation criterion [3] to the concatenated capsule activation vectors and pose matrices, collectively denoted as $Z_{\\mathrm{cat}}$ . Regularising these combined embeddings encourages diversity and non-redundancy across both activation and pose dimensions. The regularisation loss is defined as:", + "bbox": [ + 511, + 348, + 905, + 438 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) = \\lambda_ {V} V \\left(Z _ {\\text {c a t}}\\right) + \\lambda_ {C} C \\left(Z _ {\\text {c a t}}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 452, + 906, + 470 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nV \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {j = 1} ^ {d} \\max \\left(0, 1 - \\sqrt {\\operatorname {V a r} \\left(Z _ {\\text {c a t} , j}\\right)}\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 474, + 905, + 517 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nC \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {i \\neq j} \\operatorname {C o v} \\left(Z _ {\\text {c a t}}\\right) _ {i, j} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 523, + 903, + 560 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $V(Z_{\\mathrm{cat}})$ ensures that each dimension is utilised and $C(Z_{\\mathrm{cat}})$ decorrelates different dimensions to reduce redundant information.", + "bbox": [ + 511, + 566, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While SIE [24] and CapsIE [23] apply this regularisation before incorporating transformation parameters, we compute it directly on the transformation-aligned embedding. This approach yields improved performance in our experiments, as it operates on the final, fully aligned representation. Our overall loss function is defined as:", + "bbox": [ + 511, + 611, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\lambda_ {\\mathrm {i n v}} H \\left(Z _ {\\mathrm {a c t}}, Z _ {\\mathrm {a c t}} ^ {\\prime}\\right) +\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 709, + 687, + 727 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH \\left(\\bar {Z} _ {\\mathrm {a c t}}\\right) + H \\left(\\bar {Z} _ {\\mathrm {a c t}} ^ {\\prime}\\right) +\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 728, + 691, + 744 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda_ {\\mathrm {e q u i}} \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\| \\frac {Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i})}{\\| Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i}) \\| _ {F}} - \\frac {Z _ {i , \\mathrm {p o s e}} ^ {\\prime}}{\\| Z _ {i , \\mathrm {p o s e}} ^ {\\prime} \\| _ {F}} \\right\\| _ {2} ^ {2} +\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 748, + 916, + 792 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) + \\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}} ^ {\\prime}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 795, + 903, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\| \\cdot \\|_F$ denotes the Frobenius norm and $\\lambda_{\\mathrm{inv}} = 0.1, \\lambda_V = 10, \\lambda_C = 1$ , and $\\lambda_{\\mathrm{equi}} = 5$ . Despite the loss calculation occurs on the embeddings, the encoder's globally pooled output is used for downstream tasks as it yields better results [10]. Ablations, including the convergence speed", + "bbox": [ + 511, + 825, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7950", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and the impact of the number of capsules and invariant loss components, are given in supplementary Sec. 10.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Removal of predictor module. Unlike previous equivariant methods such as SEN [46], EquiMod [19], SIE [24], and CapsIE [23] that use a dedicated predictor network to map one embedding to another, our approach eliminates this module. In those earlier works, the predictor $q$ is a network that enforces the equivariance condition (see Eq. (1)). Its input is $g_{rel}$ and the projector's output $z_{i}$ . It predicts the embedding of the transformed input, $z_{i}' = q(z_{i}, g_{rel})$ , where $q$ is a hypernetwork or linear layer. This removal simplifies the architecture and instead enforces the explicit mapping of transformations in the embedding space via affine transformation matrices applied to the pose embeddings rather than manipulating embeddings through an intermediate parameter-prediction stage.", + "bbox": [ + 89, + 121, + 483, + 333 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. 3DIEBench-T: a translation-enhanced benchmark for invariant-equivariant SSL", + "text_level": 1, + "bbox": [ + 89, + 344, + 483, + 380 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As highlighted by Garrido et al. [24], commonly used datasets for evaluating invariant or equivariant properties face limitations. Evaluations of invariant representations commonly benchmark on datasets like ImageNet [17] or CIFAR [39], where objects lack parameterisable transformations because augmentations occur only in the pixel domain. However, certain equivariant datasets [38, 46] lack visual complexity and transformational diversity to allow for fine-grained control. A more comprehensive alternative is 3DIEBench [24], which controls the scene and object parameters for 3D rotation prediction (equivariance) while retaining sufficient complexity for classification (invariance). Nonetheless, as with other similar datasets [54], 3DIEBench focuses only on rotation, evaluating methods under a single geometric transformation.", + "bbox": [ + 88, + 386, + 482, + 613 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Focussing only on rotation raises questions regarding the generalisability of methods [23, 24] to other geometric transformations, and their applicability in more realistic settings. To address this, we extend 3DIEBench by introducing additional translational transformations. Our new dataset, 3DIEBench-T (3D Invariant Equivariant Benchmark-Translated), goes beyond rotation-only benchmarks [24, 54], enabling translation prediction and providing a more challenging, realistic setting to assess how methods handle simultaneous multi-geometric transformations.", + "bbox": [ + 88, + 614, + 482, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To investigate the impact of the added transformations, we adhere to the original 3DIEBench data generation protocol, ensuring any observed performance differences are attributed to the inclusion of translations rather than broader dataset modifications. We use 52,472 3D object instances spanning 55 classes from ShapeNetCoreV2 [9], originally sourced from 3D Warehouse [35]. For each instance, we generate 50 uniformly sampled views within specified ranges via Blender-Proc [18], yielding 2,623,600 images.", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/43fc3ad6b664518ae8f3d9549addce56daa0e255e89acb03b06be8656b285fa6.jpg", + "image_caption": [ + "Figure 3. Sample object instances from 3DIEBench-T dataset under different rotation, translation, and colour hue transformations." + ], + "image_footnote": [], + "bbox": [ + 516, + 88, + 903, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Samples appear in Fig. 3, with further generation information and visualisations given in supplementary Sec. 14.", + "bbox": [ + 511, + 306, + 903, + 338 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "6. Experimental results", + "text_level": 1, + "bbox": [ + 511, + 353, + 712, + 369 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Setup. To ensure fair comparisons, we adopt the experimental setup from [24] across all methods, which aligns with standard SSL practices: a pre-training phase followed by various task-specific downstream evaluations. A detailed description of the experimental setup is given in supplementary Sec. 9.", + "bbox": [ + 511, + 378, + 905, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. As an upper bound, we include a supervised baseline using ResNet-18 [30]. For invariant self-supervision, we benchmark against VICReg [3], a noncontrastive method, and SimCLR [10], a contrastive approach. We also consider SimCLR + AugSelf [40], a parameter-prediction extension of SimCLR that learns the augmentation parameters but does not enforce explicit equivariance. For equivariant self-supervision, we compare against SEN [46], EquiMod [19], SIE [24], and CapsIE [23]. Following [24, 61], we replace SEN's triplet loss with the SimCLR loss to avoid hyperparameter tuning.", + "bbox": [ + 511, + 470, + 905, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation protocol. For the invariant task, we perform classification via the linear evaluation protocol, where a linear classification layer is trained on the frozen representations. To assess equivariant properties, we perform rotation and translation prediction tasks. For each task, a three-layer MLP is trained on the frozen representations. Its inputs are concatenated pairs of representations from two views of the same object, and is trained to regress the transformation parameters—rotation for the rotation task and translation for the translation task. To explore invariance to non-geometric transformations that are not optimised in pre-training, a linear layer is trained on the frozen representations to regress the floor and spot hue. The input representations are identical to those used in the geometric tasks. To assess the performance of predicting these parameters, the coefficient of determination $R^2 = 1 - \\frac{\\sum_i(y_i - \\hat{y}_i)^2}{\\sum_i(y_i - \\bar{y})^2}$ is used, where $y_{i}$ are the target values, $\\bar{y}$ is their mean, and $\\hat{y}_i$ are the predicted", + "bbox": [ + 511, + 638, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7951", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/243698cce5538542a3240e5495d1c1e72445349c229657ecf7b78ff6fe3439bc.jpg", + "table_caption": [ + "Table 1. Evaluation on 3DIEBench and 3DIEBench-T using a ResNet-18 backbone. Learned representations are evaluated on an invariant task (classification) and equivariant tasks (rotation, translation, and colour prediction). For each group of methods, the best value is shown in bold. Supervised baselines are trained for each task. Equivariant methods are trained to be rotation-equivariant, with no explicit constraints for translation or colour prediction." + ], + "table_footnote": [], + "table_body": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
3DIEBench3DIEBench-T3DIEBench3DIEBench-T3DIEBench-T3DIEBench3DIEBench-T
Supervised Methods
ResNet-1886.4580.130.770.730.670.990.99
Invariant and Parameter Prediction Methods
VICReg84.2874.710.450.390.220.100.50
SimCLR86.7380.080.520.440.250.290.50
SimCLR + AugSelf87.4480.860.750.690.500.280.51
Equivariant Methods
SEN86.9980.200.510.450.260.290.47
EquiMod87.3980.760.500.430.240.290.38
SIE82.9475.560.730.450.200.070.46
CapsIE79.1475.200.740.600.460.010.03
EquiCaps83.2476.910.780.730.600.090.05
", + "bbox": [ + 173, + 155, + 823, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "values. For geometric tasks, a larger $R^2$ denotes a better fit for the predictions, whereas for the colour task, an $R^2$ close to 0 is better, as no colour-specific pre-training constraint is enforced. Further details are given in supplementary Sec. 9.", + "bbox": [ + 89, + 359, + 482, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.1. Rotation-only pre-training", + "text_level": 1, + "bbox": [ + 89, + 433, + 331, + 450 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Tab. 1 summarises the downstream performance of the learned representations on both 3DIEBench and 3DIEBench-T. Specifically, 3DIEBench evaluates classification, rotation, and colour (hue) tasks, while 3DIEBench-T adds translations. Notably, all equivariant methods in Tab. 1 are optimised solely for rotation equivariance during pretraining. As observed in prior works, invariant methods fail to learn successful equivariant representations of the rotations and translations, while the addition of predictive tasks (AugSelf) improves their capability to quantify this change.", + "bbox": [ + 89, + 455, + 482, + 607 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We find that EquiCaps consistently outperforms SIE—its closest competitor—marginally on classification but significantly on geometric equivariant tasks, achieving $+0.28 R^2$ in rotation and $+0.40 R^2$ in translation on 3DIEBench-T. Notably, our capsule-based projector is able to intrinsically capture translational attributes, despite no direct supervision. CapsIE also generalises relatively better than other equivariant methods, but it underperforms our method, and validates our claim that its predictor-based design underuses the inherent pose-awareness capabilities of CapsNets. Nevertheless, we find that non-capsule architectures do not generalise well under more complex, combined 3D transformations. We attribute this limitation to their lack of an explicit, pose-aware mechanism. Results on the Objaverse [16] and MOVi-E [26] datasets are given in supplementary Sec. 11.", + "bbox": [ + 89, + 608, + 482, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "6.2. Rotation-translation pre-training", + "text_level": 1, + "bbox": [ + 89, + 847, + 382, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we present results when both rotation and translation parameters are optimised. As expected, all meth", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ods exhibit a small improvement in translation performance. However, the two capsule-based methods retain their substantial improvement on geometric tasks. Consistent with [55], we find that SIE's gains on 3DIEBench do not generalise well to multi-geometric learning. We further find that SIE exhibits minimal gains over purely invariant methods and maintains the weakest translation prediction, even when explicitly trained for translation equivariance.", + "bbox": [ + 511, + 359, + 906, + 482 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e018393a457d3c7d9ed37a319a7d8d35b4bc360045974324061479c7368c325d.jpg", + "table_caption": [ + "Table 2. Evaluation on 3DIEBench-T pre-trained for both rotation and translation. Learned representations (the encoder's outputs) are evaluated on an invariant and equivariant tasks: classification, rotation, translation, and colour prediction. We indicate the change relative to Tab. 1." + ], + "table_footnote": [], + "table_body": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
SimCLR + AugSelf81.04 ↑ 0.180.69 = 0.000.64 ↑ 0.140.51 = 0.00
SEN80.23 ↑ 0.030.46 ↑ 0.010.28 ↑ 0.020.50 ↑ 0.03
EquiMod80.89 ↑ 0.130.46 ↑ 0.030.37 ↑ 0.130.37 ↓ 0.01
SIE75.91 ↑ 0.350.48 ↑ 0.030.22 ↑ 0.020.36 ↓ 0.10
CapsIE76.31 ↑ 1.110.62 ↑ 0.020.53 ↑ 0.070.03 = 0.00
EquiCaps77.88 ↑ 0.970.71 ↓ 0.020.61 ↑ 0.010.02 ↓ 0.03
", + "bbox": [ + 514, + 589, + 906, + 667 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e1098ed328c1803ff6589f7f8c5d6cb9aeae6c7f3377caddee0658cdceda601b.jpg", + "table_caption": [ + "Table 3. Evaluation on 3DIEBench and 3DIEBench-T pre-trained for both rotation and translation. Learned embeddings (the projector's outputs) are evaluated on equivariant tasks: rotation and translation prediction." + ], + "table_footnote": [], + "table_body": "
MethodRotation (R2)Translation (R2)
3DIEBench3DIEBench-T3DIEBench-T
VICReg0.240.240.00
SimCLR0.240.240.00
SimCLR + AugSelf0.240.240.00
SEN0.240.240.00
EquiMod0.240.240.22
SIE0.600.340.00
CapsIE0.710.580.26
EquiCaps0.730.610.33
", + "bbox": [ + 514, + 752, + 906, + 897 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/22ba6832f5844be1c75bb1b18d4b4a23ea5e9d5e6da9230fb30853682c266161.jpg", + "table_caption": [ + "Table 4. Equivariance evaluation on 3DIEBench (top) and 3DIEBench-T (bottom), using MRR, H@k, and separate PRE measures for rotation and translation on 3DIEBench-T. We indicate which subset (train or val) is used to generate embeddings, and for PRE, which dataset (train, val, or all) is used for retrieval. For 3DIEBench-T, methods are pre-trained for both rotation and translation equivariance." + ], + "table_footnote": [], + "table_body": "
MethodMRR (↑)H@1 (↑)H@5 (↑)PRE (↓)
trainvaltrainvaltrainvaltrain-trainval-valval-all
AllAllAllAllAllAllRotationTranlationRotationTranlationRotationTranlation
3DIEBench
SEN0.090.090.020.020.100.100.61-0.61-0.61-
EquiMod0.090.090.020.020.100.100.61-0.61-0.61-
SIE0.510.410.410.300.600.510.26-0.29-0.27-
CapsIE0.600.470.500.360.710.580.17-0.21-0.20-
EquiCaps0.680.600.550.480.840.730.13-0.18-0.17-
3DIEBench-T
SEN0.090.090.020.020.100.100.612.710.622.720.622.72
EquiMod0.130.120.030.030.180.150.582.470.592.510.592.50
SIE0.150.140.060.050.200.180.542.730.542.720.542.73
CapsIE0.460.370.350.260.550.470.292.450.312.480.302.46
EquiCaps0.590.510.480.410.700.620.302.340.352.410.332.38
", + "bbox": [ + 173, + 142, + 823, + 348 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Colour prediction. EquiCaps maintains near-perfect invariance to hue transformations, consistent with prior work such as SIE and CapsIE, although SIE is less effective under 3DIEBench-T. By contrast, non-capsule methods exhibit substantially increased colour prediction when transitioning to 3DIEBench-T. We conjecture that the dataset's increased geometric complexity makes the colour task easier to solve for models lacking explicit pose-aware mechanisms.", + "bbox": [ + 89, + 375, + 483, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.3. Performance on embeddings", + "text_level": 1, + "bbox": [ + 89, + 502, + 346, + 518 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As in prior works [10, 24], we observe that representations (the outputs of the CNN encoder $f_{\\theta}$ ) learned by invariant methods achieve higher-than-expected performance on equivariant tasks. This arises because the invariance objective is enforced on the embeddings (the outputs of the projector $h_{\\phi}$ ) after the projector, while the representations can still retain information [10]. To explore this further, we evaluate rotation and translation predictions on the embeddings (see Tab. 3), where all methods show decreased performance compared to their representation-level results, highlighting that the projector absorbs more invariance. We also observe that invariant methods, along with SEN and EquiMod, converge to similar embedding-level results, except for EquiMod, which retains higher translation prediction performance, as in the representation-stage results. SIE aligns with invariant methods on translation prediction, consistent with its representation stage. EquiCaps maintains the best performance on both tasks, and capsule-based methods show the least decline from representations to embeddings, underscoring the advantages of capsules' pose-awareness.", + "bbox": [ + 88, + 523, + 482, + 825 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.4. Equivariance evaluation", + "text_level": 1, + "bbox": [ + 89, + 833, + 313, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We employ and adapt the Mean Reciprocal Rank (MRR), Hit Rate at k (H@k), and Prediction Retrieval Error (PRE) metrics from [24]. For compared methods, for each object's", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "source and target poses, we feed the source embeddings and the corresponding transformation parameters through the predictor, obtaining a predicted embedding. To adapt it to our method, we multiply the projector's pose matrix by the transformation matrix. We obtain the nearest neighbours of the predicted embedding among the views of the same object. MRR is computed as the mean of the reciprocal rank $\\frac{1}{r}$ , where $r$ is the rank of the target embedding among the neighbours. H@k yields 1 if the target is within the top- $k$ neighbours, and 0 otherwise. PRE measures the mean distance between the transformation parameters of the nearest neighbour and the target across the dataset. Rotation uses $d_r = 1 - \\langle q_1, q_2 \\rangle^2$ , where $q_1$ and $q_2$ are the nearest neighbour's and target's rotation represented as a quaternion, respectively. Translation uses the mean-squared error $d_t = ||t_1 - t_2||_2^2$ , where $t_1$ is the nearest neighbour's translation vector and $t_2$ is the target's translation vector.", + "bbox": [ + 511, + 375, + 906, + 631 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As shown in Tab. 4, EquiCaps outperforms all methods across each data split of 3DIEBench, achieving an MRR of 0.60 on the validation set, well above SIE (0.41) and CapsIE (0.47). Similar results are observed in H@1/5 metrics, where SEN and EquiMod perform randomly (2% and 10%) on both data splits, while CapsIE's PRE on all data splits approaches ours. Although EquiCaps generally outperforms all methods on 3DIEBench-T across most metrics, CapsIE achieves slightly lower rotation PRE. Despite this, our method remains significantly stronger in MRR and H@1/5 across all data splits, with EquiMod slightly improving. In contrast, SIE's overall performance drops with H@1 of 0.05 on the validation set, marginally above random.", + "bbox": [ + 511, + 633, + 908, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6.5. Qualitative results", + "text_level": 1, + "bbox": [ + 511, + 845, + 691, + 861 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Nearest-neighbour representations. We perform a nearest-neighbour representation search on the 3DIEBench-", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2ea6445c95794947c54b78747670a08ea7c63212438a01c07a0271e24b7bbe45.jpg", + "image_caption": [ + "Figure 4. Nearest-neighbour representation retrieval on 3DIEBench-T validation set directly after pre-training. The query image (left) is compared against each method's learned representations to find its top three nearest neighbours (in rows: 1-NN, 2-NN, 3-NN). Equivariant and parameter prediction methods are trained to be both rotationally and translationally equivariant." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 867, + 275 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "T validation set (Fig. 4) to illustrate the information captured by the representations. Invariant approaches should retrieve objects in the same class but varied poses, whereas equivariant methods should preserve the query pose. We find that only our method and CapsIE retrieve the nearest neighbour with a matching pose, though ours aligns also with the translation. For subsequent neighbours, only EquiCaps and AugSelf retrieve objects in similar poses, while CapsIE does so but less accurately. In contrast, the other methods, including equivariant methods, fail to yield pose-matching neighbours suggesting that invariance dominates and their representations leave out significant pose information. We replicate the retrieval on 3DIEBench Fig. A.6, supplementary Sec. 12, where ours, SIE, and CapsIE successfully return pose-matching neighbours, but remaining methods, only partially align poses, albeit less accurate. We conjecture this is attributed to the projector absorbing invariance, while partial rotation-related information remains at the representation stage. All retrieved objects belong to the car category, indicating that class semantics are preserved consistent with our quantitative results.", + "bbox": [ + 91, + 351, + 483, + 669 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Equivariance via pose manipulation. To further illustrate the equivariant properties of our approach, we examine the inversion property through the inverse rotation of an object in the embedding space. We rotate an object around the $z$ -axis within the range $[0, \\frac{\\pi}{2}]$ in $5^{\\circ}$ increments and feed these images through our trained network to obtain its pose embedding. Embedding is subsequently multiplied by its corresponding inverse transformation matrix. To examine the impact of the inverse transformation we retrieve the nearest neighbour among all embeddings. As shown in Fig. 5, the nearest neighbour closely (within $5^{\\circ}$ error) or exactly matches the source view, highlighting the invertibility of our representations, and the ability to intuitively and directly steer the latent space. Further qualitative results are provided in supplementary Sec. 12.", + "bbox": [ + 89, + 674, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/888dcae3d3a87acacc900a20db579cfbe82e8d99abddd932b8366290c0bd3f33.jpg", + "image_caption": [ + "Figure 5. Illustration of transformation inversion of equivariant embeddings through capsule-based pose manipulation." + ], + "image_footnote": [], + "bbox": [ + 514, + 348, + 906, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 553, + 633, + 569 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose EquiCaps, a capsule-based SSL method that jointly learns invariant and equivariant representations. By leveraging capsules' innate pose-awareness, EquiCaps shapes the latent space predictably, removing the need for a dedicated predictor. To enable richer benchmarking, we extend 3DIEBench to 3DIEBench-T, incorporating object translations alongside rotations. Empirically, EquiCaps achieves state-of-the-art rotation and translation prediction among equivariant methods. Tests on 3DIEBench-T highlight capsule architectures' superior generalisation under combined geometric transformations, demonstrating their potential in invariant-equivariant SSL.", + "bbox": [ + 509, + 580, + 906, + 762 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future work. Joint invariance-equivariance optimisation involves trade-offs; while EquiCaps balances these, task-specific priorities may vary. The method assumes known group elements—as prior methods do—restricting its use without such data, though partial data suffices. EquiCaps can theoretically handle any transformation which can be expressed as a matrix without architectural changes. Future work includes other 3D transformations (e.g., scaling, shear), and more object detection tasks.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgements", + "text_level": 1, + "bbox": [ + 89, + 90, + 277, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the UKRI AI Centre for Doctoral Training in Sustainable Understandable agri-food Systems Transformed by Artificial INtelligence (SUSTAIN) [grant reference: EP/Y03063X/1]. We also acknowledge the generous support of alumni and friends in establishing the University of Aberdeen's Interdisciplinary Institute, which partly enabled this research, including Dr Jane Hellman Caseley (MBChB 1956), Professor Patrick Meares (DSc 1959), Nancy Miller (MA 1942), Norman Robertson, Dr Ian Sessler (MBChB 1956) and Anne Young (MA 1957). Finally, we thank the University of Aberdeen's HPC Maxwell facility for providing the computational resources that facilitated our experimental work.", + "bbox": [ + 89, + 114, + 485, + 313 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 329, + 187, + 344 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, pages 8443-8452, 2021.", + "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022.", + "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022.", + "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VICRegl: Self-supervised learning of local visual features. In NeurIPS, pages 8799-8810, 2022.", + "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020.", + "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018.", + "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020.", + "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021.", + "[9] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, QiXing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015.", + "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge" + ], + "bbox": [ + 93, + 353, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020.", + "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021.", + "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020.", + "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021.", + "[14] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. In ICLR, 2022.", + "[15] Fabio De Sousa Ribeiro, Kevin Duarte, Miles Everett, Georgios Leontidis, and Mubarak Shah. Object-centric learning with capsule networks: A survey. ACM Comput. Surv., 56 (11), 2024.", + "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, pages 13142-13153, 2023.", + "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009.", + "[18] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. BlenderProc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023.", + "[19] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve self-supervised learning. In ICLR, 2023.", + "[20] Kevin Duarte, Yogesh Rawat, and Mubarak Shah. Videocapsulenet: A simplified network for action detection. In NeurIPS, 2018.", + "[21] Kevin Duarte, Yogesh S. Rawat, and Mubarak Shah. Capsulevos: Semi-supervised video object segmentation using capsule routing. In ICCV, 2019.", + "[22] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, pages 3015-3024, 2021.", + "[23] Miles Everett, Aiden Durrant, Mingjun Zhong, and Georgios Leontidis. Capsule network projectors are equivariant and invariant learners. arXiv preprint arXiv:2405.14386, 2024.", + "[24] Quentin Garrido, Laurent Najman, and Yann LeCun. Self-supervised learning of split invariant equivariant representations. In ICML, 2023.", + "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018.", + "[26] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek" + ], + "bbox": [ + 516, + 92, + 903, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. In CVPR, pages 3749-3761, 2022.", + "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020.", + "[28] Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and Stefanie Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. In ICLR, 2024.", + "[29] Taeyoung Hahn, Myeongjiang Pyeon, and Gunhee Kim. Selfrouting capsule networks. In NeurIPS, 2019.", + "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016.", + "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020.", + "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022.", + "[33] Geoffrey E. Hinton, Alex Krizhevsky, and Sida Wang. Transforming auto-encoders. In ICANN, pages 44–51, 2011.", + "[34] Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with EM routing. In ICLR, 2018.", + "[35] Trimble Inc. 3D warehouse. https://3dwarehouse.sketchup.com/. Accessed: 2025-03-01.", + "[36] Armand Joulin and Francis Bach. A convex relaxation for weakly supervised classifiers. In ICML, pages 1279-1286, 2012.", + "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015.", + "[38] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. In ICLR, 2020.", + "[39] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009.", + "[40] Hankook Lee, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. Improving transferability of representations via augmentation-aware self-supervision. In NeurIPS, 2021.", + "[41] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In ICCV, pages 1046-1055, 2021.", + "[42] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, 2024.", + "[43] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, pages 11525-11538, 2020.", + "[44] Giovanni Luca Marchetti, Gustaf Tegnér, Anastasiia Varava, and Danica Kragic. Equivariant representation learning via class-pose decomposition. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 4745-4756, 2023.", + "[45] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024.", + "[46] Jung Yeon Park, Ondrej Biza, Linfeng Zhao, Jan Willem van de Meent, and Robin Walters. Learning symmetric embeddings for equivariant world models. In ICML, 2022.", + "[47] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In AAAI, pages 3749-3756, 2020.", + "[48] Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. Dynamic routing between capsules. In NeurIPS, pages 3856-3866, 2017.", + "[49] Sara Sabour, Andrea Tagliasacchi, Soroosh Yazdani, Geoffrey E. Hinton, and David J. Fleet. Unsupervised part representation by flow capsules. In ICML, pages 9213-9223, 2021.", + "[50] Franz Scherr, Qinghai Guo, and Timoleon Moraitis. Self-supervised learning through efference copies. In NeurIPS, 2022.", + "[51] Mehran Shakerinava, Arnab Kumar Mondal, and Siamak Ravanbakhsh. Structuring representations using group invariants. In NeurIPS, pages 34162-34174, 2022.", + "[52] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey E. Hinton, and Kwang Moo Yi. Canonical capsules: Self-supervised capsules in canonical pose. In NeurIPS, pages 24993-25005, 2021.", + "[53] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020.", + "[54] Jiayun Wang, Yubei Chen, and Stella Yu. Pose-aware self-supervised learning with viewpoint trajectory regularization. In ECCV, 2024.", + "[55] Qin Wang, Kai Krajsek, and Hanno Scharr. Equivariant representation learning for augmentation-based self-supervised learning via image reconstruction. In NeurIPS 2024 Workshop: Self-Supervised Learning - Theory and Practice, 2024.", + "[56] Yifei Wang, Kaiwen Hu, Sharut Gupta, Ziyu Ye, Yisen Wang, and Stefanie Jegelka. Understanding the role of equivariance in self-supervised learning. In NeurIPS, pages 127483-127510, 2024." + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7956", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Robin Winter, Marco Bertolini, Tuan Le, Frank Noe, and Djork-Arné Clevert. Unsupervised learning of group invariant and equivariant representations. In NeurIPS, 2022.", + "[58] Tete Xiao, Xiaolong Wang, Alexei A. Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021.", + "[59] Yuyang Xie, Jianhong Wen, Kin Wai Lau, Yasar Abbas Ur Rehman, and Jiajun Shen. What should be equivariant in self-supervised learning. In CVPR Workshop, pages 4111-4120, 2022.", + "[60] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684, 2022.", + "[61] Jaemyung Yu, Jaehyun Choi, Dong-Jae Lee, HyeongGwon Hong, and Junmo Kim. Self-supervised transformation learning for equivariant representations. In NeurIPS, 2024.", + "[62] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320, 2021." + ], + "bbox": [ + 91, + 92, + 480, + 359 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_model.json b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..49619e23adb9c47c2f88bc3b283cfe2942e4f664 --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_model.json @@ -0,0 +1,2202 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.147, + 0.131, + 0.852, + 0.154 + ], + "angle": 0, + "content": "EquiCaps: Predictor-Free Pose-Aware Pre-Trained Capsule Networks" + }, + { + "type": "text", + "bbox": [ + 0.126, + 0.18, + 0.531, + 0.217 + ], + "angle": 0, + "content": "Athinoula Konstantinou1 Georgios Leontidis1 \n1University of Aberdeen, UK" + }, + { + "type": "text", + "bbox": [ + 0.553, + 0.181, + 0.871, + 0.217 + ], + "angle": 0, + "content": "Mamatha Thota² Aiden Durrant¹ \n²University of Lincoln, UK" + }, + { + "type": "text", + "bbox": [ + 0.112, + 0.219, + 0.88, + 0.234 + ], + "angle": 0, + "content": "{a.konstantinou.24, georgios.leontidis, aiden.durrant}@abdn.ac.uk, mthota@lincoln.ac.uk" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.485, + 0.741 + ], + "angle": 0, + "content": "Learning self-supervised representations that are invariant and equivariant to transformations is crucial for advancing beyond traditional visual classification tasks. However, many methods rely on predictor architectures to encode equivariance, despite evidence that architectural choices, such as capsule networks, inherently excel at learning interpretable pose-aware representations. To explore this, we introduce EquiCaps (Equivariant Capsule Network), a capsule-based approach to pose-aware self-supervision that eliminates the need for a specialised predictor for enforcing equivariance. Instead, we leverage the intrinsic pose-awareness capabilities of capsules to improve performance in pose estimation tasks. To further challenge our assumptions, we increase task complexity via multi-geometric transformations to enable a more thorough evaluation of invariance and equivariance by introducing 3DIEBench-T, an extension of a 3D object-rendering benchmark dataset. Empirical results demonstrate that EquiCaps outperforms prior state-of-the-art equivariant methods on geometric tasks, including rotation and translation, achieving a supervised-level \\( R^2 \\) of 0.78 on the 3DIEBench rotation prediction benchmark and improving upon SIE and CapsIE by 0.05 and 0.04 \\( R^2 \\), respectively. Moreover, in contrast to non-capsule-based equivariant approaches, EquiCaps maintains robust equivariant performance under combined geometric transformations, underscoring its generalisation capabilities and the promise of predictor-free capsule architectures. Code, dataset, and weights are released at http://github.com/AberdeenML/EquiCaps." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.769, + 0.222, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "The success of self-supervised learning (SSL) is largely attributed to the exploitation of invariant representations under data augmentations and transformations to learn rich semantic concepts [3, 4, 7, 8, 10, 12, 27, 31, 32, 62]. However, methods that strictly enforce invariance fail to preserve useful augmentation and transformation information, which could limit their applicability to downstream tasks where" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.268, + 0.886, + 0.453 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.459, + 0.886, + 0.563 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.573, + 0.908, + 0.63 + ], + "angle": 0, + "content": "Figure 1. Equivariant evaluation of EquiCaps (ours) and competing methods on 3DIEBench benchmark. We indicate which subset (train or val) is used to generate embeddings, and, for PRE, which dataset (train, val, or all) is used for retrieval." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.658, + 0.907, + 0.764 + ], + "angle": 0, + "content": "such information is vital [28, 40, 58]. In response, there has been increasing interest in developing more generalised self-supervised approaches, most notably through the investigation of pose-aware learning [24, 54, 56]. Achieving such generalisation, particularly to novel viewpoints, can consequently address real-world settings that inherently exhibit large data variations, e.g. autonomous driving [54]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "One promising strategy to preserve transformation information—and thus achieve pose-awareness—is to directly optimise for equivariant representations, enabling the model to capture how transformations steer the latent space. However, such methods often face limitations [19, 23, 24, 46]. First, they rely on a dedicated predictor network to encode equivariant properties, often further increasing architectural complexity through conditional projectors, partitioned representations, or hypernet" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7947" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "works. Second, equivariance in representations is typically enforced solely via objective functions with little regard taken to exploit architectural mechanisms—such as Capsule Networks (CapsNets)—that have been shown to effectively encode or structure pose information [15]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.168, + 0.483, + 0.333 + ], + "angle": 0, + "content": "We propose to simplify and enhance the framework of equivariant self-supervision to address these concerns by using CapsNets [15]. CapsNets structure representations into capsules, each encoding both the existence of an entity and its instantiation parameters, where the latter is achieved through a routing algorithm based on agreement of part-whole relationships. This formulation introduces an inductive bias that has been shown to effectively capture equivariance w.r.t. viewpoints and invariance in the capsule activations, enabling them to exhibit viewpoint-invariant and viewpoint-equivariant properties by design." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.334, + 0.485, + 0.56 + ], + "angle": 0, + "content": "While prior attempts have been made to introduce CapsNets into equivariant SSL [23], they still rely on a separate transformation predictor, potentially underusing capsules' intrinsic pose-awareness capabilities. To overcome these limitations, we propose a streamlined pipeline that eliminates the predictor and directly leverages capsules' equivariant properties that prior equivariant SSL methods do not. CapsNets offer an intriguing prospect for capturing transformations, therefore we introduce an extension of the 3DIEBench dataset [24] that incorporates translational transformations in addition to rotations. By moving from rotation-only SO(3) transformations [24, 54] to the more general SE(3) group, our dataset increases the difficulty of both invariance and equivariance tasks, allowing for the evaluation of generalisation under multi-geometric settings." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.56, + 0.425, + 0.574 + ], + "angle": 0, + "content": "To summarise, our contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.577, + 0.483, + 0.651 + ], + "angle": 0, + "content": "- 3DIEBench-T: We present 3DIEBench-T, an extension of the 3DIEBench that integrates translational transformations alongside rotations. This new benchmark enables the evaluation of representation methods under simultaneous multi-geometric 3D transformations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.653, + 0.483, + 0.711 + ], + "angle": 0, + "content": "- Predictor limitations: We empirically demonstrate that existing predictor-based equivariant architectures yield minimal improvement over purely invariant approaches when learning multi-geometric 3D transformations." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.713, + 0.483, + 0.819 + ], + "angle": 0, + "content": "- Predictor-free CapsNet: We propose EquiCaps, removing the need for predictors by exploiting the intrinsic pose-awareness of CapsNets. Empirical results on rotation and translation tasks demonstrate that EquiCaps outperforms existing equivariant methods, and generalises successfully under multi-geometric settings, while also allowing intuitive manipulation of the embedding space." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.577, + 0.483, + 0.819 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.83, + 0.229, + 0.846 + ], + "angle": 0, + "content": "2. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.887 + ], + "angle": 0, + "content": "We first introduce key notions of equivariance and invariance in our setting [24, 46]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Group actions. A group is a set \\( G \\) with a binary operation \\( \\cdot \\) that satisfies associativity, \\( G \\times G \\to G \\), the existence of an identity element, \\( g \\cdot 1 = 1 \\cdot g = g \\) such that \\( \\forall g \\in G \\), and the existence of an inverse, \\( g \\cdot g^{-1} = g^{-1} \\cdot g = 1 \\). A group action of \\( G \\) on a set \\( S \\) is defined as a map \\( \\alpha : G \\times S \\to S \\) that is compatible with the composition of group elements, i.e., \\( \\alpha(g, \\alpha(h, s)) = \\alpha(gh, s) \\), \\( \\forall s \\in S \\) and \\( \\forall g, h \\in G \\). The map \\( \\alpha(g, \\cdot) \\) is called a group representation when \\( S \\) is a vector space \\( V \\), such as \\( \\mathbb{R}^n \\), and \\( \\alpha \\) is linear. A group representation is denoted as the map \\( \\rho : G \\to GL(V) \\), such that \\( \\rho(g) = \\alpha(g, \\cdot) \\). As in similar works [24], we use group representations to describe how transformations are applied both to the input data and latent space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.334 + ], + "angle": 0, + "content": "Equivalence learning. A function \\( f: X \\to Y \\) is equivariant if for a group \\( G \\), with representations \\( \\rho_{X} \\) and \\( \\rho_{Y} \\) on \\( X \\) and \\( Y \\), satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.541, + 0.344, + 0.905, + 0.361 + ], + "angle": 0, + "content": "\\[\nf \\left(\\rho_ {X} (g) \\cdot x\\right) = \\rho_ {Y} (g) \\cdot f (x), \\forall x \\in X, \\forall g \\in G. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.371, + 0.905, + 0.416 + ], + "angle": 0, + "content": "This indicates the commutativity of group transformations and function application. The function \\( f \\) retains the symmetry group \\( G \\) yet changes how the group operates." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.905, + 0.448 + ], + "angle": 0, + "content": "Invariance learning. A function \\( f: X \\to Y \\) is invariant if for a group \\( G \\), with representations \\( \\rho_{X} \\) on \\( X \\), satisfies" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.458, + 0.905, + 0.474 + ], + "angle": 0, + "content": "\\[\nf \\left(\\rho_ {X} (g) \\cdot x\\right) = f (x), \\quad \\forall x \\in X, \\forall g \\in G. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.905, + 0.516 + ], + "angle": 0, + "content": "Invariance is a trivial [14], special case of equivariance when \\(\\rho_Y(g)\\), in Eq. (1) is the identity map." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.529, + 0.659, + 0.544 + ], + "angle": 0, + "content": "3. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Invariant SSL. Invariance to data augmentations has been the driving force of self-supervision, using this property to learn common semantic concepts under different transformations when there is no access to labels. Specifically, invariance in the latent space is enforced by aligning shared views that are transformations of the same source. There are various approaches to undertake this alignment and discrimination, notably contrastive methods [10, 12, 13, 31, 60], clustering-based learning [6-8, 45], and non-contrastive methods [3, 4, 11, 22, 27, 62] which instead employ architectural or objective modifications, gradient mechanisms, and information-theoretic regularisers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Implicit equivariant SSL. Invariant SSL methods are optimised to disregard data transformations, hence information is lost. To address this, predictive equivariant SSL encodes transformation-sensitive information by training to predict transformation parameters [25, 40, 50], while others integrate contrastive learning with transformation prediction [14]. Furthermore, modulating the contrastive objective itself, based on the strength of the applied augmentation has shown promise [59], while [58] introduces distinct embedding spaces that remain invariant to all but one augmentation. Focussing on the transformation, optimal augmented" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.118, + 0.094, + 0.52, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.298, + 0.228, + 0.317, + 0.241 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.098, + 0.678, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.602, + 0.227, + 0.623, + 0.241 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.099, + 0.891, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.771, + 0.228, + 0.79, + 0.241 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.256, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 2. (a) Framework Overview. Two views \\( x_{1} \\) and \\( x_{1}^{\\prime} = \\rho_{X}(g_{rel}) \\cdot x_{1} \\) of the same object are processed by a shared CNN encoder \\( f_{\\theta} \\) and a capsule-based projector \\( h_{\\phi} \\). (b) Projector's Output. For each view, the projector produces an activation probability vector \\( (z_{act}, z_{act}') \\), indicating entity presence, and a pose matrix \\( (z_{pose}, z_{pose}') \\), encoding the pose for each entity. (c) Equivariant Objective. Equivariance is enforced by multiplying \\( z_{pose} \\) by the relative transformation matrix \\( \\rho_{Y}(g_{rel}) \\) and aligning the result with \\( z_{pose}' \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.482, + 0.429 + ], + "angle": 0, + "content": "views can also be exploited to retain task-relevant information [53]. However, these do not optimise for equivariance defined in Eq. (1). Hence, since they do not learn an explicit mapping from input transformations to corresponding changes in latent space, they do not guarantee equivariance, unlike explicit equivariant approaches [24, 56]." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.431, + 0.483, + 0.718 + ], + "angle": 0, + "content": "Explicit equivariant SSL. Avoiding transformation prediction tasks in favour of evaluating for equivariant representations, elicits a separate family of methods. These utilise representation reconstruction and maintain invariant and equivariant properties by splitting into or learning separate latent spaces [44, 57]. Others [51, 61] encourage equivariance through identically transformed image pairs, with [51] utilising the transformation group information for choosing optimal loss terms. Such approaches have also extended contrastive learning enforcing input transformations correspond to rotations in the embedding space [28]. Several methods adopt a predictor network to enforce an equivalent transformation mapping between latent representations through augmentation parameters [19, 46]. Building on these, SIE [24] incorporates a hypernetwork to learn such mapping, while dividing the representation into invariant and equivariant parts. CapsIE [23] adopts this module, instead leveraging a CapsNet projector, removing the need for multiple projectors and the split representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.902 + ], + "angle": 0, + "content": "CapsNets. CapsNets are designed to model part-whole hierarchical relationships by capturing the likelihood of an entity's presence and its attributes (pose, colour, texture, etc.) through neural activity. Lower-level capsules connect to higher-level ones via routing based on agreement in likelihood and positional attributes. This alignment enables capsules to identify objects by reinforcing pose consistency when multiple capsules converge on the same entity. Due to this positional agreement mechanism, CapsNets are explored as equivariant architectures, enabling tasks like image generation [33] and transformation learning [48]. Various routing methods have been proposed, including" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.338, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Gaussian-mixture clustering [34], Variational Bayes with uncertainty [47], self-routing via subordinate networks [29], and residual pose routing [42]. Like our approach, CapsNets have been adopted in Siamese SSL, notably by predicting canonical poses for 3D point clouds [52], or through optical flow analysis using successive video frames [49]. They have also been applied to video tasks, for example, by transforming 2D convolutional capsule routing to 3D [20] and by adopting attention-based routing [21]. Lastly, CapsNets share foundational concepts with slot-based object-centric methods [15], such as [43], which uses an iterative attention mechanism akin to capsule routing, whereas the slots in [41] specialise in distinct entities as capsules." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.554, + 0.898, + 0.572 + ], + "angle": 0, + "content": "4. EquiCaps: Predictor-Free Pose-Aware SSL" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.907, + 0.853 + ], + "angle": 0, + "content": "Motivation. Prior equivariant self-supervised methods [19, 24, 46] commonly rely on predictor networks to enforce equivariant representations via augmentation-aware mappings. This leads to ad hoc vector representations that are both difficult to interpret or manipulate—an important consideration for evaluating equivariance. In effect, these methods attempt to enforce equivariance on architectures (i.e., CNNs) that are not naturally suited to it [15] and often introduce extra complexity through dual projectors, representation splits, or hypernetworks. Instead, we exploit the equivariant properties elicited by CapsNets [15], to more effectively encode positional information, particularly rotation and translation. Our design yields a simpler pose-aware SSL framework, outperforms previous equivariant methods on geometric tasks, and enables a more intuitive control and interpretability in the embedding space through the structured \\(4 \\times 4\\) capsule pose matrices. We introduce and rationalise each component of our methodology as follows." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Framework overview. Our method adopts a Siamese joint embedding setup with identical, weight-sharing branches, each formed by a CNN \\( f_{\\theta} \\) and a CapsNet projec-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7949" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.484, + 0.365 + ], + "angle": 0, + "content": "tor \\( h_{\\phi} \\). We omit global average pooling to preserve the 2D spatial structure necessary for capsule processing. Given an image \\( x \\in \\mathbb{R}^{c \\times h \\times w} \\) from a dataset \\( D \\), we sample two transformations \\( g_{1}, g_{2} \\) from a distribution over the group \\( G \\), as detailed in supplementary Sec. 14. These are applied to \\( x \\) to produce two distinct views, \\( x_{g_1} = \\rho_X(g_1) \\cdot x \\) and \\( x_{g_2} = \\rho_X(g_2) \\cdot x \\). Both views are passed through \\( f_{\\theta} \\), producing feature maps \\( F, F' \\in \\mathbb{R}^{c \\times h \\times w} \\). These are then processed by \\( h_{\\phi} \\), which contains \\( N \\) capsules, yielding \\( z_{\\mathrm{act}}, z_{\\mathrm{pose}} = h_{\\phi}(F) \\) and \\( z_{\\mathrm{act}}', z_{\\mathrm{pose}}' = h_{\\phi}(F') \\), where \\( z_{\\mathrm{act}}, z_{\\mathrm{act}}' \\in \\mathbb{R}^N \\) are the activation probabilities, and \\( z_{\\mathrm{pose}}, z_{\\mathrm{pose}}' \\in \\mathbb{R}^{N \\times 4 \\times 4} \\) are the corresponding pose matrices. The activation vectors encode transformation-invariant properties, representing semantic concepts that remain unchanged under \\( g_{1}, g_{2} \\). In contrast, the pose matrices capture transformation-equivariant properties, encoding an entity's pose, where each capsule represents a unique entity. A visual depiction of the general architecture is given in Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.365, + 0.484, + 0.606 + ], + "angle": 0, + "content": "Self-routing CapsNet projector. To reduce the computational overhead of iterative-routing, we adopt the noniterative self-routing algorithm [29] as our projector. The projector \\( h_{\\phi} \\) consists of two layers: a primary capsule layer that transforms encoder feature maps into poses \\( (u_i) \\) and activations \\( (a_i) \\), and a self-routing layer, where each capsule independently determines its routing coefficients via a subordinate single-layer perceptron. To do this, the routing coefficients \\( c_{ij} \\) between each lower-layer capsule \\( i \\) and each upper-level capsule \\( j \\) are computed by multiplying the lower capsule's pose vector \\( u_i \\) with a learnable weight matrix \\( W^{\\mathrm{route}} \\) and applying a softmax function. The activation \\( a_j \\) of each upper-level capsule is then computed by forming votes from lower-level capsules—each vote is the product \\( c_{ij}a_i \\), which are then summed and normalised by the total lower-level activation, to formulate weighted votes:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.611, + 0.482, + 0.647 + ], + "angle": 0, + "content": "\\[\nc _ {i j} = \\operatorname {s o f t m a x} \\left(W _ {i} ^ {\\text {r o u t e}} u _ {i}\\right) _ {j}, \\quad a _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}{\\sum_ {i \\in \\Omega_ {l}} a _ {i}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.716 + ], + "angle": 0, + "content": "The capsule layer's output pose is computed by multiplying a trainable weight matrix \\( W^{\\text{pose}} \\) with each lower-level pose \\( u_i \\), yielding a predicted pose \\( \\hat{u}_{j|i} \\) for each upper-level capsule. The final pose \\( u_j \\) is the weighted average of \\( \\hat{u}_{j|i} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.72, + 0.482, + 0.757 + ], + "angle": 0, + "content": "\\[\n\\hat {u} _ {j \\mid i} = W _ {i j} ^ {\\text {p o s e}} u _ {i}, \\quad u _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i} \\hat {u} _ {j \\mid i}}{\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Enforcing equivariance. We define the relative transformation \\( g_{\\mathrm{rel}} = g_2g_1^{-1} \\), mapping between the views \\( x_{g_1} \\) and \\( x_{g_2} \\), and let \\( \\rho_Y(g_{\\mathrm{rel}}) \\) denote its matrix representation. To enforce equivariance (defined in Eq. (1)) we multiply \\( z_{\\mathrm{pose}} \\) by \\( \\rho_Y(g_{\\mathrm{rel}}) \\), aligning the transformed result with \\( z_{\\mathrm{pose}}' \\). This indicates that the pose embeddings are equivariant w.r.t. the transformation group \\( G \\), ensuring that geometric changes in the input are consistently captured in the latent space. Each pose matrix \\( (z_{\\mathrm{pose}}\\cdot \\rho_Y(g_{\\mathrm{rel}})) \\) and \\( z_{\\mathrm{pose}}' \\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.136 + ], + "angle": 0, + "content": "is normalised by its Frobenius norm and residing on a unit hypersphere. This direct manipulation in the latent space removes the need for a dedicated predictor module." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.349 + ], + "angle": 0, + "content": "Loss functions. To enforce invariance we minimise the cross-entropy \\( H(Z_{\\mathrm{act}}, Z_{\\mathrm{act}}^{\\prime}) \\) between the activation probability vectors \\( Z_{\\mathrm{act}} \\) and \\( Z_{\\mathrm{act}}^{\\prime} \\), where \\( Z \\) denotes the matrix embeddings computed over a batch. This aligns the capsule activations for different views of the same image. To ensure full capsule utilisation, we apply mean entropy maximisation (ME-MAX) regularisation [1, 2, 36] to the activation probabilities. Let \\( \\bar{Z}_{\\mathrm{act}} = \\frac{1}{B}\\sum_{i=1}^{B}Z_{\\mathrm{act}} \\), where \\( B \\) is the batch size. Maximising \\( H(Z_{\\mathrm{act}}) \\) and \\( H(\\bar{Z}_{\\mathrm{act}}^{\\prime}) \\) encourages a diverse distribution of activation across all capsules, preventing the collapse to only a few highly active capsules. To enforce equivariance we minimise the mean-squared error between the normalised, transformed pose \\( (Z_{\\mathrm{pose}}\\cdot\\rho_{Y}(g_{\\mathrm{rel}})) \\) and the corresponding pose \\( Z_{\\mathrm{pose}}^{\\prime} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.906, + 0.439 + ], + "angle": 0, + "content": "To prevent collapse, we apply the variance-covariance regularisation criterion [3] to the concatenated capsule activation vectors and pose matrices, collectively denoted as \\( Z_{\\mathrm{cat}} \\). Regularising these combined embeddings encourages diversity and non-redundancy across both activation and pose dimensions. The regularisation loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.57, + 0.453, + 0.907, + 0.471 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) = \\lambda_ {V} V \\left(Z _ {\\text {c a t}}\\right) + \\lambda_ {C} C \\left(Z _ {\\text {c a t}}\\right), \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.476, + 0.906, + 0.518 + ], + "angle": 0, + "content": "\\[\nV \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {j = 1} ^ {d} \\max \\left(0, 1 - \\sqrt {\\operatorname {V a r} \\left(Z _ {\\text {c a t} , j}\\right)}\\right), \\tag {6}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.524, + 0.905, + 0.561 + ], + "angle": 0, + "content": "\\[\nC \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {i \\neq j} \\operatorname {C o v} \\left(Z _ {\\text {c a t}}\\right) _ {i, j} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.567, + 0.906, + 0.611 + ], + "angle": 0, + "content": "where \\( V(Z_{\\mathrm{cat}}) \\) ensures that each dimension is utilised and \\( C(Z_{\\mathrm{cat}}) \\) decorrelates different dimensions to reduce redundant information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.906, + 0.702 + ], + "angle": 0, + "content": "While SIE [24] and CapsIE [23] apply this regularisation before incorporating transformation parameters, we compute it directly on the transformation-aligned embedding. This approach yields improved performance in our experiments, as it operates on the final, fully aligned representation. Our overall loss function is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.71, + 0.689, + 0.728 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\lambda_ {\\mathrm {i n v}} H \\left(Z _ {\\mathrm {a c t}}, Z _ {\\mathrm {a c t}} ^ {\\prime}\\right) +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.729, + 0.692, + 0.746 + ], + "angle": 0, + "content": "\\[\nH \\left(\\bar {Z} _ {\\mathrm {a c t}}\\right) + H \\left(\\bar {Z} _ {\\mathrm {a c t}} ^ {\\prime}\\right) +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.749, + 0.918, + 0.793 + ], + "angle": 0, + "content": "\\[\n\\lambda_ {\\mathrm {e q u i}} \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\| \\frac {Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i})}{\\| Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i}) \\| _ {F}} - \\frac {Z _ {i , \\mathrm {p o s e}} ^ {\\prime}}{\\| Z _ {i , \\mathrm {p o s e}} ^ {\\prime} \\| _ {F}} \\right\\| _ {2} ^ {2} +\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.543, + 0.796, + 0.905, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) + \\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}} ^ {\\prime}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.901 + ], + "angle": 0, + "content": "where \\(\\| \\cdot \\|_F\\) denotes the Frobenius norm and \\(\\lambda_{\\mathrm{inv}} = 0.1, \\lambda_V = 10, \\lambda_C = 1\\), and \\(\\lambda_{\\mathrm{equi}} = 5\\). Despite the loss calculation occurs on the embeddings, the encoder's globally pooled output is used for downstream tasks as it yields better results [10]. Ablations, including the convergence speed" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "and the impact of the number of capsules and invariant loss components, are given in supplementary Sec. 10." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.484, + 0.334 + ], + "angle": 0, + "content": "Removal of predictor module. Unlike previous equivariant methods such as SEN [46], EquiMod [19], SIE [24], and CapsIE [23] that use a dedicated predictor network to map one embedding to another, our approach eliminates this module. In those earlier works, the predictor \\( q \\) is a network that enforces the equivariance condition (see Eq. (1)). Its input is \\( g_{rel} \\) and the projector's output \\( z_{i} \\). It predicts the embedding of the transformed input, \\( z_{i}' = q(z_{i}, g_{rel}) \\), where \\( q \\) is a hypernetwork or linear layer. This removal simplifies the architecture and instead enforces the explicit mapping of transformations in the embedding space via affine transformation matrices applied to the pose embeddings rather than manipulating embeddings through an intermediate parameter-prediction stage." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.345, + 0.484, + 0.381 + ], + "angle": 0, + "content": "5. 3DIEBench-T: a translation-enhanced benchmark for invariant-equivariant SSL" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.387, + 0.483, + 0.614 + ], + "angle": 0, + "content": "As highlighted by Garrido et al. [24], commonly used datasets for evaluating invariant or equivariant properties face limitations. Evaluations of invariant representations commonly benchmark on datasets like ImageNet [17] or CIFAR [39], where objects lack parameterisable transformations because augmentations occur only in the pixel domain. However, certain equivariant datasets [38, 46] lack visual complexity and transformational diversity to allow for fine-grained control. A more comprehensive alternative is 3DIEBench [24], which controls the scene and object parameters for 3D rotation prediction (equivariance) while retaining sufficient complexity for classification (invariance). Nonetheless, as with other similar datasets [54], 3DIEBench focuses only on rotation, evaluating methods under a single geometric transformation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.615, + 0.483, + 0.765 + ], + "angle": 0, + "content": "Focussing only on rotation raises questions regarding the generalisability of methods [23, 24] to other geometric transformations, and their applicability in more realistic settings. To address this, we extend 3DIEBench by introducing additional translational transformations. Our new dataset, 3DIEBench-T (3D Invariant Equivariant Benchmark-Translated), goes beyond rotation-only benchmarks [24, 54], enabling translation prediction and providing a more challenging, realistic setting to assess how methods handle simultaneous multi-geometric transformations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To investigate the impact of the added transformations, we adhere to the original 3DIEBench data generation protocol, ensuring any observed performance differences are attributed to the inclusion of translations rather than broader dataset modifications. We use 52,472 3D object instances spanning 55 classes from ShapeNetCoreV2 [9], originally sourced from 3D Warehouse [35]. For each instance, we generate 50 uniformly sampled views within specified ranges via Blender-Proc [18], yielding 2,623,600 images." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.089, + 0.905, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.25, + 0.905, + 0.278 + ], + "angle": 0, + "content": "Figure 3. Sample object instances from 3DIEBench-T dataset under different rotation, translation, and colour hue transformations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.307, + 0.905, + 0.339 + ], + "angle": 0, + "content": "Samples appear in Fig. 3, with further generation information and visualisations given in supplementary Sec. 14." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.354, + 0.714, + 0.371 + ], + "angle": 0, + "content": "6. Experimental results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.38, + 0.906, + 0.472 + ], + "angle": 0, + "content": "Setup. To ensure fair comparisons, we adopt the experimental setup from [24] across all methods, which aligns with standard SSL practices: a pre-training phase followed by various task-specific downstream evaluations. A detailed description of the experimental setup is given in supplementary Sec. 9." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.472, + 0.906, + 0.638 + ], + "angle": 0, + "content": "Baselines. As an upper bound, we include a supervised baseline using ResNet-18 [30]. For invariant self-supervision, we benchmark against VICReg [3], a noncontrastive method, and SimCLR [10], a contrastive approach. We also consider SimCLR + AugSelf [40], a parameter-prediction extension of SimCLR that learns the augmentation parameters but does not enforce explicit equivariance. For equivariant self-supervision, we compare against SEN [46], EquiMod [19], SIE [24], and CapsIE [23]. Following [24, 61], we replace SEN's triplet loss with the SimCLR loss to avoid hyperparameter tuning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Evaluation protocol. For the invariant task, we perform classification via the linear evaluation protocol, where a linear classification layer is trained on the frozen representations. To assess equivariant properties, we perform rotation and translation prediction tasks. For each task, a three-layer MLP is trained on the frozen representations. Its inputs are concatenated pairs of representations from two views of the same object, and is trained to regress the transformation parameters—rotation for the rotation task and translation for the translation task. To explore invariance to non-geometric transformations that are not optimised in pre-training, a linear layer is trained on the frozen representations to regress the floor and spot hue. The input representations are identical to those used in the geometric tasks. To assess the performance of predicting these parameters, the coefficient of determination \\( R^2 = 1 - \\frac{\\sum_i(y_i - \\hat{y}_i)^2}{\\sum_i(y_i - \\bar{y})^2} \\) is used, where \\( y_{i} \\) are the target values, \\( \\bar{y} \\) is their mean, and \\( \\hat{y}_i \\) are the predicted" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.956 + ], + "angle": 0, + "content": "7951" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Evaluation on 3DIEBench and 3DIEBench-T using a ResNet-18 backbone. Learned representations are evaluated on an invariant task (classification) and equivariant tasks (rotation, translation, and colour prediction). For each group of methods, the best value is shown in bold. Supervised baselines are trained for each task. Equivariant methods are trained to be rotation-equivariant, with no explicit constraints for translation or colour prediction." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.156, + 0.825, + 0.336 + ], + "angle": 0, + "content": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
3DIEBench3DIEBench-T3DIEBench3DIEBench-T3DIEBench-T3DIEBench3DIEBench-T
Supervised Methods
ResNet-1886.4580.130.770.730.670.990.99
Invariant and Parameter Prediction Methods
VICReg84.2874.710.450.390.220.100.50
SimCLR86.7380.080.520.440.250.290.50
SimCLR + AugSelf87.4480.860.750.690.500.280.51
Equivariant Methods
SEN86.9980.200.510.450.260.290.47
EquiMod87.3980.760.500.430.240.290.38
SIE82.9475.560.730.450.200.070.46
CapsIE79.1475.200.740.600.460.010.03
EquiCaps83.2476.910.780.730.600.090.05
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.483, + 0.422 + ], + "angle": 0, + "content": "values. For geometric tasks, a larger \\( R^2 \\) denotes a better fit for the predictions, whereas for the colour task, an \\( R^2 \\) close to 0 is better, as no colour-specific pre-training constraint is enforced. Further details are given in supplementary Sec. 9." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.434, + 0.332, + 0.451 + ], + "angle": 0, + "content": "6.1. Rotation-only pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.608 + ], + "angle": 0, + "content": "Tab. 1 summarises the downstream performance of the learned representations on both 3DIEBench and 3DIEBench-T. Specifically, 3DIEBench evaluates classification, rotation, and colour (hue) tasks, while 3DIEBench-T adds translations. Notably, all equivariant methods in Tab. 1 are optimised solely for rotation equivariance during pretraining. As observed in prior works, invariant methods fail to learn successful equivariant representations of the rotations and translations, while the addition of predictive tasks (AugSelf) improves their capability to quantify this change." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.835 + ], + "angle": 0, + "content": "We find that EquiCaps consistently outperforms SIE—its closest competitor—marginally on classification but significantly on geometric equivariant tasks, achieving \\(+0.28 R^2\\) in rotation and \\(+0.40 R^2\\) in translation on 3DIEBench-T. Notably, our capsule-based projector is able to intrinsically capture translational attributes, despite no direct supervision. CapsIE also generalises relatively better than other equivariant methods, but it underperforms our method, and validates our claim that its predictor-based design underuses the inherent pose-awareness capabilities of CapsNets. Nevertheless, we find that non-capsule architectures do not generalise well under more complex, combined 3D transformations. We attribute this limitation to their lack of an explicit, pose-aware mechanism. Results on the Objaverse [16] and MOVi-E [26] datasets are given in supplementary Sec. 11." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.383, + 0.865 + ], + "angle": 0, + "content": "6.2. Rotation-translation pre-training" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In Tab. 2, we present results when both rotation and translation parameters are optimised. As expected, all meth" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.361, + 0.907, + 0.483 + ], + "angle": 0, + "content": "ods exhibit a small improvement in translation performance. However, the two capsule-based methods retain their substantial improvement on geometric tasks. Consistent with [55], we find that SIE's gains on 3DIEBench do not generalise well to multi-geometric learning. We further find that SIE exhibits minimal gains over purely invariant methods and maintains the weakest translation prediction, even when explicitly trained for translation equivariance." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.509, + 0.907, + 0.577 + ], + "angle": 0, + "content": "Table 2. Evaluation on 3DIEBench-T pre-trained for both rotation and translation. Learned representations (the encoder's outputs) are evaluated on an invariant and equivariant tasks: classification, rotation, translation, and colour prediction. We indicate the change relative to Tab. 1." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.59, + 0.907, + 0.668 + ], + "angle": 0, + "content": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
SimCLR + AugSelf81.04 ↑ 0.180.69 = 0.000.64 ↑ 0.140.51 = 0.00
SEN80.23 ↑ 0.030.46 ↑ 0.010.28 ↑ 0.020.50 ↑ 0.03
EquiMod80.89 ↑ 0.130.46 ↑ 0.030.37 ↑ 0.130.37 ↓ 0.01
SIE75.91 ↑ 0.350.48 ↑ 0.030.22 ↑ 0.020.36 ↓ 0.10
CapsIE76.31 ↑ 1.110.62 ↑ 0.020.53 ↑ 0.070.03 = 0.00
EquiCaps77.88 ↑ 0.970.71 ↓ 0.020.61 ↑ 0.010.02 ↓ 0.03
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.686, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Table 3. Evaluation on 3DIEBench and 3DIEBench-T pre-trained for both rotation and translation. Learned embeddings (the projector's outputs) are evaluated on equivariant tasks: rotation and translation prediction." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.753, + 0.907, + 0.898 + ], + "angle": 0, + "content": "
MethodRotation (R2)Translation (R2)
3DIEBench3DIEBench-T3DIEBench-T
VICReg0.240.240.00
SimCLR0.240.240.00
SimCLR + AugSelf0.240.240.00
SEN0.240.240.00
EquiMod0.240.240.22
SIE0.600.340.00
CapsIE0.710.580.26
EquiCaps0.730.610.33
" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7952" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 4. Equivariance evaluation on 3DIEBench (top) and 3DIEBench-T (bottom), using MRR, H@k, and separate PRE measures for rotation and translation on 3DIEBench-T. We indicate which subset (train or val) is used to generate embeddings, and for PRE, which dataset (train, val, or all) is used for retrieval. For 3DIEBench-T, methods are pre-trained for both rotation and translation equivariance." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.143, + 0.825, + 0.349 + ], + "angle": 0, + "content": "
MethodMRR (↑)H@1 (↑)H@5 (↑)PRE (↓)
trainvaltrainvaltrainvaltrain-trainval-valval-all
AllAllAllAllAllAllRotationTranlationRotationTranlationRotationTranlation
3DIEBench
SEN0.090.090.020.020.100.100.61-0.61-0.61-
EquiMod0.090.090.020.020.100.100.61-0.61-0.61-
SIE0.510.410.410.300.600.510.26-0.29-0.27-
CapsIE0.600.470.500.360.710.580.17-0.21-0.20-
EquiCaps0.680.600.550.480.840.730.13-0.18-0.17-
3DIEBench-T
SEN0.090.090.020.020.100.100.612.710.622.720.622.72
EquiMod0.130.120.030.030.180.150.582.470.592.510.592.50
SIE0.150.140.060.050.200.180.542.730.542.720.542.73
CapsIE0.460.370.350.260.550.470.292.450.312.480.302.46
EquiCaps0.590.510.480.410.700.620.302.340.352.410.332.38
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.375, + 0.484, + 0.496 + ], + "angle": 0, + "content": "Colour prediction. EquiCaps maintains near-perfect invariance to hue transformations, consistent with prior work such as SIE and CapsIE, although SIE is less effective under 3DIEBench-T. By contrast, non-capsule methods exhibit substantially increased colour prediction when transitioning to 3DIEBench-T. We conjecture that the dataset's increased geometric complexity makes the colour task easier to solve for models lacking explicit pose-aware mechanisms." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.503, + 0.347, + 0.519 + ], + "angle": 0, + "content": "6.3. Performance on embeddings" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.827 + ], + "angle": 0, + "content": "As in prior works [10, 24], we observe that representations (the outputs of the CNN encoder \\( f_{\\theta} \\)) learned by invariant methods achieve higher-than-expected performance on equivariant tasks. This arises because the invariance objective is enforced on the embeddings (the outputs of the projector \\( h_{\\phi} \\)) after the projector, while the representations can still retain information [10]. To explore this further, we evaluate rotation and translation predictions on the embeddings (see Tab. 3), where all methods show decreased performance compared to their representation-level results, highlighting that the projector absorbs more invariance. We also observe that invariant methods, along with SEN and EquiMod, converge to similar embedding-level results, except for EquiMod, which retains higher translation prediction performance, as in the representation-stage results. SIE aligns with invariant methods on translation prediction, consistent with its representation stage. EquiCaps maintains the best performance on both tasks, and capsule-based methods show the least decline from representations to embeddings, underscoring the advantages of capsules' pose-awareness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.315, + 0.85 + ], + "angle": 0, + "content": "6.4. Equivariance evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We employ and adapt the Mean Reciprocal Rank (MRR), Hit Rate at k (H@k), and Prediction Retrieval Error (PRE) metrics from [24]. For compared methods, for each object's" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.632 + ], + "angle": 0, + "content": "source and target poses, we feed the source embeddings and the corresponding transformation parameters through the predictor, obtaining a predicted embedding. To adapt it to our method, we multiply the projector's pose matrix by the transformation matrix. We obtain the nearest neighbours of the predicted embedding among the views of the same object. MRR is computed as the mean of the reciprocal rank \\(\\frac{1}{r}\\), where \\(r\\) is the rank of the target embedding among the neighbours. H@k yields 1 if the target is within the top-\\(k\\) neighbours, and 0 otherwise. PRE measures the mean distance between the transformation parameters of the nearest neighbour and the target across the dataset. Rotation uses \\(d_r = 1 - \\langle q_1, q_2 \\rangle^2\\), where \\(q_1\\) and \\(q_2\\) are the nearest neighbour's and target's rotation represented as a quaternion, respectively. Translation uses the mean-squared error \\(d_t = ||t_1 - t_2||_2^2\\), where \\(t_1\\) is the nearest neighbour's translation vector and \\(t_2\\) is the target's translation vector." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.634, + 0.909, + 0.831 + ], + "angle": 0, + "content": "As shown in Tab. 4, EquiCaps outperforms all methods across each data split of 3DIEBench, achieving an MRR of 0.60 on the validation set, well above SIE (0.41) and CapsIE (0.47). Similar results are observed in H@1/5 metrics, where SEN and EquiMod perform randomly (2% and 10%) on both data splits, while CapsIE's PRE on all data splits approaches ours. Although EquiCaps generally outperforms all methods on 3DIEBench-T across most metrics, CapsIE achieves slightly lower rotation PRE. Despite this, our method remains significantly stronger in MRR and H@1/5 across all data splits, with EquiMod slightly improving. In contrast, SIE's overall performance drops with H@1 of 0.05 on the validation set, marginally above random." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.692, + 0.862 + ], + "angle": 0, + "content": "6.5. Qualitative results" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Nearest-neighbour representations. We perform a nearest-neighbour representation search on the 3DIEBench-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7953" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.087, + 0.868, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.284, + 0.908, + 0.327 + ], + "angle": 0, + "content": "Figure 4. Nearest-neighbour representation retrieval on 3DIEBench-T validation set directly after pre-training. The query image (left) is compared against each method's learned representations to find its top three nearest neighbours (in rows: 1-NN, 2-NN, 3-NN). Equivariant and parameter prediction methods are trained to be both rotationally and translationally equivariant." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.352, + 0.485, + 0.67 + ], + "angle": 0, + "content": "T validation set (Fig. 4) to illustrate the information captured by the representations. Invariant approaches should retrieve objects in the same class but varied poses, whereas equivariant methods should preserve the query pose. We find that only our method and CapsIE retrieve the nearest neighbour with a matching pose, though ours aligns also with the translation. For subsequent neighbours, only EquiCaps and AugSelf retrieve objects in similar poses, while CapsIE does so but less accurately. In contrast, the other methods, including equivariant methods, fail to yield pose-matching neighbours suggesting that invariance dominates and their representations leave out significant pose information. We replicate the retrieval on 3DIEBench Fig. A.6, supplementary Sec. 12, where ours, SIE, and CapsIE successfully return pose-matching neighbours, but remaining methods, only partially align poses, albeit less accurate. We conjecture this is attributed to the projector absorbing invariance, while partial rotation-related information remains at the representation stage. All retrieved objects belong to the car category, indicating that class semantics are preserved consistent with our quantitative results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Equivariance via pose manipulation. To further illustrate the equivariant properties of our approach, we examine the inversion property through the inverse rotation of an object in the embedding space. We rotate an object around the \\(z\\)-axis within the range \\([0, \\frac{\\pi}{2}]\\) in \\(5^{\\circ}\\) increments and feed these images through our trained network to obtain its pose embedding. Embedding is subsequently multiplied by its corresponding inverse transformation matrix. To examine the impact of the inverse transformation we retrieve the nearest neighbour among all embeddings. As shown in Fig. 5, the nearest neighbour closely (within \\(5^{\\circ}\\) error) or exactly matches the source view, highlighting the invertibility of our representations, and the ability to intuitively and directly steer the latent space. Further qualitative results are provided in supplementary Sec. 12." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.349, + 0.907, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.495, + 0.907, + 0.525 + ], + "angle": 0, + "content": "Figure 5. Illustration of transformation inversion of equivariant embeddings through capsule-based pose manipulation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.554, + 0.634, + 0.57 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.582, + 0.907, + 0.763 + ], + "angle": 0, + "content": "We propose EquiCaps, a capsule-based SSL method that jointly learns invariant and equivariant representations. By leveraging capsules' innate pose-awareness, EquiCaps shapes the latent space predictably, removing the need for a dedicated predictor. To enable richer benchmarking, we extend 3DIEBench to 3DIEBench-T, incorporating object translations alongside rotations. Empirically, EquiCaps achieves state-of-the-art rotation and translation prediction among equivariant methods. Tests on 3DIEBench-T highlight capsule architectures' superior generalisation under combined geometric transformations, demonstrating their potential in invariant-equivariant SSL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations and future work. Joint invariance-equivariance optimisation involves trade-offs; while EquiCaps balances these, task-specific priorities may vary. The method assumes known group elements—as prior methods do—restricting its use without such data, though partial data suffices. EquiCaps can theoretically handle any transformation which can be expressed as a matrix without architectural changes. Future work includes other 3D transformations (e.g., scaling, shear), and more object detection tasks." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7954" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.279, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.314 + ], + "angle": 0, + "content": "This work was supported by the UKRI AI Centre for Doctoral Training in Sustainable Understandable agri-food Systems Transformed by Artificial INtelligence (SUSTAIN) [grant reference: EP/Y03063X/1]. We also acknowledge the generous support of alumni and friends in establishing the University of Aberdeen's Interdisciplinary Institute, which partly enabled this research, including Dr Jane Hellman Caseley (MBChB 1956), Professor Patrick Meares (DSc 1959), Nancy Miller (MA 1942), Norman Robertson, Dr Ian Sessler (MBChB 1956) and Anne Young (MA 1957). Finally, we thank the University of Aberdeen's HPC Maxwell facility for providing the computational resources that facilitated our experimental work." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.33, + 0.188, + 0.345 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.354, + 0.483, + 0.425 + ], + "angle": 0, + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, pages 8443-8452, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.426, + 0.483, + 0.495 + ], + "angle": 0, + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.497, + 0.483, + 0.539 + ], + "angle": 0, + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.483, + 0.583 + ], + "angle": 0, + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VICRegl: Self-supervised learning of local visual features. In NeurIPS, pages 8799-8810, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.585, + 0.484, + 0.64 + ], + "angle": 0, + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.643, + 0.483, + 0.685 + ], + "angle": 0, + "content": "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.686, + 0.483, + 0.741 + ], + "angle": 0, + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.744, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.801, + 0.483, + 0.884 + ], + "angle": 0, + "content": "[9] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, QiXing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.354, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.15 + ], + "angle": 0, + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.152, + 0.904, + 0.193 + ], + "angle": 0, + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.196, + 0.905, + 0.237 + ], + "angle": 0, + "content": "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.239, + 0.905, + 0.281 + ], + "angle": 0, + "content": "[14] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. In ICLR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.283, + 0.905, + 0.338 + ], + "angle": 0, + "content": "[15] Fabio De Sousa Ribeiro, Kevin Duarte, Miles Everett, Georgios Leontidis, and Mubarak Shah. Object-centric learning with capsule networks: A survey. ACM Comput. Surv., 56 (11), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.408 + ], + "angle": 0, + "content": "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, pages 13142-13153, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[18] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. BlenderProc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.527, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[19] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve self-supervised learning. In ICLR, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.57, + 0.905, + 0.611 + ], + "angle": 0, + "content": "[20] Kevin Duarte, Yogesh Rawat, and Mubarak Shah. Videocapsulenet: A simplified network for action detection. In NeurIPS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.613, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[21] Kevin Duarte, Yogesh S. Rawat, and Mubarak Shah. Capsulevos: Semi-supervised video object segmentation using capsule routing. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[22] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, pages 3015-3024, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.701, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[23] Miles Everett, Aiden Durrant, Mingjun Zhong, and Georgios Leontidis. Capsule network projectors are equivariant and invariant learners. arXiv preprint arXiv:2405.14386, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[24] Quentin Garrido, Laurent Najman, and Yann LeCun. Self-supervised learning of split invariant equivariant representations. In ICML, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[26] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7955" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.483, + 0.176 + ], + "angle": 0, + "content": "Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. In CVPR, pages 3749-3761, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.483, + 0.272 + ], + "angle": 0, + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[28] Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and Stefanie Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. In ICLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[29] Taeyoung Hahn, Myeongjiang Pyeon, and Gunhee Kim. Selfrouting capsule networks. In NeurIPS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.36, + 0.482, + 0.402 + ], + "angle": 0, + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.402, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[33] Geoffrey E. Hinton, Alex Krizhevsky, and Sida Wang. Transforming auto-encoders. In ICANN, pages 44–51, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[34] Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with EM routing. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.48, + 0.573 + ], + "angle": 0, + "content": "[35] Trimble Inc. 3D warehouse. https://3dwarehouse.sketchup.com/. Accessed: 2025-03-01." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.615 + ], + "angle": 0, + "content": "[36] Armand Joulin and Francis Bach. A convex relaxation for weakly supervised classifiers. In ICML, pages 1279-1286, 2012." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.617, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[38] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. In ICLR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[39] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[40] Hankook Lee, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. Improving transferability of representations via augmentation-aware self-supervision. In NeurIPS, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[41] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In ICCV, pages 1046-1055, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[42] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[43] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, pages 11525-11538, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.135, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[44] Giovanni Luca Marchetti, Gustaf Tegnér, Anastasiia Varava, and Danica Kragic. Equivariant representation learning via class-pose decomposition. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 4745-4756, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[45] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.384 + ], + "angle": 0, + "content": "[46] Jung Yeon Park, Ondrej Biza, Linfeng Zhao, Jan Willem van de Meent, and Robin Walters. Learning symmetric embeddings for equivariant world models. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[47] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In AAAI, pages 3749-3756, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[48] Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. Dynamic routing between capsules. In NeurIPS, pages 3856-3866, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.523 + ], + "angle": 0, + "content": "[49] Sara Sabour, Andrea Tagliasacchi, Soroosh Yazdani, Geoffrey E. Hinton, and David J. Fleet. Unsupervised part representation by flow capsules. In ICML, pages 9213-9223, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.905, + 0.564 + ], + "angle": 0, + "content": "[50] Franz Scherr, Qinghai Guo, and Timoleon Moraitis. Self-supervised learning through efference copies. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.566, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[51] Mehran Shakerinava, Arnab Kumar Mondal, and Siamak Ravanbakhsh. Structuring representations using group invariants. In NeurIPS, pages 34162-34174, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[52] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey E. Hinton, and Kwang Moo Yi. Canonical capsules: Self-supervised capsules in canonical pose. In NeurIPS, pages 24993-25005, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[53] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[54] Jiayun Wang, Yubei Chen, and Stella Yu. Pose-aware self-supervised learning with viewpoint trajectory regularization. In ECCV, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[55] Qin Wang, Kai Krajsek, and Hanno Scharr. Equivariant representation learning for augmentation-based self-supervised learning via image reconstruction. In NeurIPS 2024 Workshop: Self-Supervised Learning - Theory and Practice, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[56] Yifei Wang, Kaiwen Hu, Sharut Gupta, Ziyu Ye, Yisen Wang, and Stefanie Jegelka. Understanding the role of equivariance in self-supervised learning. In NeurIPS, pages 127483-127510, 2024." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[57] Robin Winter, Marco Bertolini, Tuan Le, Frank Noe, and Djork-Arné Clevert. Unsupervised learning of group invariant and equivariant representations. In NeurIPS, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.175 + ], + "angle": 0, + "content": "[58] Tete Xiao, Xiaolong Wang, Alexei A. Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[59] Yuyang Xie, Jianhong Wen, Kin Wai Lau, Yasar Abbas Ur Rehman, and Jiajun Shen. What should be equivariant in self-supervised learning. In CVPR Workshop, pages 4111-4120, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[60] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[61] Jaemyung Yu, Jaehyun Choi, Dong-Jae Lee, HyeongGwon Hong, and Junmo Kim. Self-supervised transformation learning for equivariant representations. In NeurIPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.361 + ], + "angle": 0, + "content": "[62] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320, 2021." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7957" + } + ] +] \ No newline at end of file diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_origin.pdf b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..90c77b09478e81bdaf41cf93849a5807c78a0cfe --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/0be2dbf6-d8b3-41a1-9e9f-e1a2a2904ac5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:82cae1ca2a7d160b0b421d39803fe34c073b3d33872823326f439f3ec4c8d1d6 +size 2712949 diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/full.md b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ccc6f529a58225a7a387c57b20ccae9696205577 --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/full.md @@ -0,0 +1,311 @@ +# EquiCaps: Predictor-Free Pose-Aware Pre-Trained Capsule Networks + +Athinoula Konstantinou1 Georgios Leontidis1 +1University of Aberdeen, UK + +Mamatha Thota² Aiden Durrant¹ +²University of Lincoln, UK + +{a.konstantinou.24, georgios.leontidis, aiden.durrant}@abdn.ac.uk, mthota@lincoln.ac.uk + +# Abstract + +Learning self-supervised representations that are invariant and equivariant to transformations is crucial for advancing beyond traditional visual classification tasks. However, many methods rely on predictor architectures to encode equivariance, despite evidence that architectural choices, such as capsule networks, inherently excel at learning interpretable pose-aware representations. To explore this, we introduce EquiCaps (Equivariant Capsule Network), a capsule-based approach to pose-aware self-supervision that eliminates the need for a specialised predictor for enforcing equivariance. Instead, we leverage the intrinsic pose-awareness capabilities of capsules to improve performance in pose estimation tasks. To further challenge our assumptions, we increase task complexity via multi-geometric transformations to enable a more thorough evaluation of invariance and equivariance by introducing 3DIEBench-T, an extension of a 3D object-rendering benchmark dataset. Empirical results demonstrate that EquiCaps outperforms prior state-of-the-art equivariant methods on geometric tasks, including rotation and translation, achieving a supervised-level $R^2$ of 0.78 on the 3DIEBench rotation prediction benchmark and improving upon SIE and CapsIE by 0.05 and 0.04 $R^2$ , respectively. Moreover, in contrast to non-capsule-based equivariant approaches, EquiCaps maintains robust equivariant performance under combined geometric transformations, underscoring its generalisation capabilities and the promise of predictor-free capsule architectures. Code, dataset, and weights are released at http://github.com/AberdeenML/EquiCaps. + +# 1. Introduction + +The success of self-supervised learning (SSL) is largely attributed to the exploitation of invariant representations under data augmentations and transformations to learn rich semantic concepts [3, 4, 7, 8, 10, 12, 27, 31, 32, 62]. However, methods that strictly enforce invariance fail to preserve useful augmentation and transformation information, which could limit their applicability to downstream tasks where + +![](images/caa2dbd19f946af1435a4c2980b43c2ef4537b2cbc3c342357c616c86a7cad4b.jpg) + +![](images/64e325b9400dc2d2d0ac14b623fe1f3cb892e9d6cdfed833baeec4899340521c.jpg) +Figure 1. Equivariant evaluation of EquiCaps (ours) and competing methods on 3DIEBench benchmark. We indicate which subset (train or val) is used to generate embeddings, and, for PRE, which dataset (train, val, or all) is used for retrieval. + +such information is vital [28, 40, 58]. In response, there has been increasing interest in developing more generalised self-supervised approaches, most notably through the investigation of pose-aware learning [24, 54, 56]. Achieving such generalisation, particularly to novel viewpoints, can consequently address real-world settings that inherently exhibit large data variations, e.g. autonomous driving [54]. + +One promising strategy to preserve transformation information—and thus achieve pose-awareness—is to directly optimise for equivariant representations, enabling the model to capture how transformations steer the latent space. However, such methods often face limitations [19, 23, 24, 46]. First, they rely on a dedicated predictor network to encode equivariant properties, often further increasing architectural complexity through conditional projectors, partitioned representations, or hypernet + +works. Second, equivariance in representations is typically enforced solely via objective functions with little regard taken to exploit architectural mechanisms—such as Capsule Networks (CapsNets)—that have been shown to effectively encode or structure pose information [15]. + +We propose to simplify and enhance the framework of equivariant self-supervision to address these concerns by using CapsNets [15]. CapsNets structure representations into capsules, each encoding both the existence of an entity and its instantiation parameters, where the latter is achieved through a routing algorithm based on agreement of part-whole relationships. This formulation introduces an inductive bias that has been shown to effectively capture equivariance w.r.t. viewpoints and invariance in the capsule activations, enabling them to exhibit viewpoint-invariant and viewpoint-equivariant properties by design. + +While prior attempts have been made to introduce CapsNets into equivariant SSL [23], they still rely on a separate transformation predictor, potentially underusing capsules' intrinsic pose-awareness capabilities. To overcome these limitations, we propose a streamlined pipeline that eliminates the predictor and directly leverages capsules' equivariant properties that prior equivariant SSL methods do not. CapsNets offer an intriguing prospect for capturing transformations, therefore we introduce an extension of the 3DIEBench dataset [24] that incorporates translational transformations in addition to rotations. By moving from rotation-only SO(3) transformations [24, 54] to the more general SE(3) group, our dataset increases the difficulty of both invariance and equivariance tasks, allowing for the evaluation of generalisation under multi-geometric settings. + +To summarise, our contributions are as follows: + +- 3DIEBench-T: We present 3DIEBench-T, an extension of the 3DIEBench that integrates translational transformations alongside rotations. This new benchmark enables the evaluation of representation methods under simultaneous multi-geometric 3D transformations. +- Predictor limitations: We empirically demonstrate that existing predictor-based equivariant architectures yield minimal improvement over purely invariant approaches when learning multi-geometric 3D transformations. +- Predictor-free CapsNet: We propose EquiCaps, removing the need for predictors by exploiting the intrinsic pose-awareness of CapsNets. Empirical results on rotation and translation tasks demonstrate that EquiCaps outperforms existing equivariant methods, and generalises successfully under multi-geometric settings, while also allowing intuitive manipulation of the embedding space. + +# 2. Preliminaries + +We first introduce key notions of equivariance and invariance in our setting [24, 46]. + +Group actions. A group is a set $G$ with a binary operation $\cdot$ that satisfies associativity, $G \times G \to G$ , the existence of an identity element, $g \cdot 1 = 1 \cdot g = g$ such that $\forall g \in G$ , and the existence of an inverse, $g \cdot g^{-1} = g^{-1} \cdot g = 1$ . A group action of $G$ on a set $S$ is defined as a map $\alpha : G \times S \to S$ that is compatible with the composition of group elements, i.e., $\alpha(g, \alpha(h, s)) = \alpha(gh, s)$ , $\forall s \in S$ and $\forall g, h \in G$ . The map $\alpha(g, \cdot)$ is called a group representation when $S$ is a vector space $V$ , such as $\mathbb{R}^n$ , and $\alpha$ is linear. A group representation is denoted as the map $\rho : G \to GL(V)$ , such that $\rho(g) = \alpha(g, \cdot)$ . As in similar works [24], we use group representations to describe how transformations are applied both to the input data and latent space. + +Equivalence learning. A function $f: X \to Y$ is equivariant if for a group $G$ , with representations $\rho_{X}$ and $\rho_{Y}$ on $X$ and $Y$ , satisfies + +$$ +f \left(\rho_ {X} (g) \cdot x\right) = \rho_ {Y} (g) \cdot f (x), \forall x \in X, \forall g \in G. \tag {1} +$$ + +This indicates the commutativity of group transformations and function application. The function $f$ retains the symmetry group $G$ yet changes how the group operates. + +Invariance learning. A function $f: X \to Y$ is invariant if for a group $G$ , with representations $\rho_{X}$ on $X$ , satisfies + +$$ +f \left(\rho_ {X} (g) \cdot x\right) = f (x), \quad \forall x \in X, \forall g \in G. \tag {2} +$$ + +Invariance is a trivial [14], special case of equivariance when $\rho_Y(g)$ , in Eq. (1) is the identity map. + +# 3. Related works + +Invariant SSL. Invariance to data augmentations has been the driving force of self-supervision, using this property to learn common semantic concepts under different transformations when there is no access to labels. Specifically, invariance in the latent space is enforced by aligning shared views that are transformations of the same source. There are various approaches to undertake this alignment and discrimination, notably contrastive methods [10, 12, 13, 31, 60], clustering-based learning [6-8, 45], and non-contrastive methods [3, 4, 11, 22, 27, 62] which instead employ architectural or objective modifications, gradient mechanisms, and information-theoretic regularisers. + +Implicit equivariant SSL. Invariant SSL methods are optimised to disregard data transformations, hence information is lost. To address this, predictive equivariant SSL encodes transformation-sensitive information by training to predict transformation parameters [25, 40, 50], while others integrate contrastive learning with transformation prediction [14]. Furthermore, modulating the contrastive objective itself, based on the strength of the applied augmentation has shown promise [59], while [58] introduces distinct embedding spaces that remain invariant to all but one augmentation. Focussing on the transformation, optimal augmented + +![](images/e3362aef672282506dea597c1e511667fcfa0d5ae1dd875df349f6f48eaee286.jpg) +(a) + +![](images/ded19a7e06175cdd94f8cebfe06aa3e6a31eed696eb1836a5dfc351d3d8ee4b3.jpg) +(b) +Figure 2. (a) Framework Overview. Two views $x_{1}$ and $x_{1}^{\prime} = \rho_{X}(g_{rel}) \cdot x_{1}$ of the same object are processed by a shared CNN encoder $f_{\theta}$ and a capsule-based projector $h_{\phi}$ . (b) Projector's Output. For each view, the projector produces an activation probability vector $(z_{act}, z_{act}')$ , indicating entity presence, and a pose matrix $(z_{pose}, z_{pose}')$ , encoding the pose for each entity. (c) Equivariant Objective. Equivariance is enforced by multiplying $z_{pose}$ by the relative transformation matrix $\rho_{Y}(g_{rel})$ and aligning the result with $z_{pose}'$ . + +![](images/d6b3920457ede7f4cb24d3f97ac2254c2c23d40dc41ae4d581051579bf583c4d.jpg) +(c) + +views can also be exploited to retain task-relevant information [53]. However, these do not optimise for equivariance defined in Eq. (1). Hence, since they do not learn an explicit mapping from input transformations to corresponding changes in latent space, they do not guarantee equivariance, unlike explicit equivariant approaches [24, 56]. + +Explicit equivariant SSL. Avoiding transformation prediction tasks in favour of evaluating for equivariant representations, elicits a separate family of methods. These utilise representation reconstruction and maintain invariant and equivariant properties by splitting into or learning separate latent spaces [44, 57]. Others [51, 61] encourage equivariance through identically transformed image pairs, with [51] utilising the transformation group information for choosing optimal loss terms. Such approaches have also extended contrastive learning enforcing input transformations correspond to rotations in the embedding space [28]. Several methods adopt a predictor network to enforce an equivalent transformation mapping between latent representations through augmentation parameters [19, 46]. Building on these, SIE [24] incorporates a hypernetwork to learn such mapping, while dividing the representation into invariant and equivariant parts. CapsIE [23] adopts this module, instead leveraging a CapsNet projector, removing the need for multiple projectors and the split representation. + +CapsNets. CapsNets are designed to model part-whole hierarchical relationships by capturing the likelihood of an entity's presence and its attributes (pose, colour, texture, etc.) through neural activity. Lower-level capsules connect to higher-level ones via routing based on agreement in likelihood and positional attributes. This alignment enables capsules to identify objects by reinforcing pose consistency when multiple capsules converge on the same entity. Due to this positional agreement mechanism, CapsNets are explored as equivariant architectures, enabling tasks like image generation [33] and transformation learning [48]. Various routing methods have been proposed, including + +Gaussian-mixture clustering [34], Variational Bayes with uncertainty [47], self-routing via subordinate networks [29], and residual pose routing [42]. Like our approach, CapsNets have been adopted in Siamese SSL, notably by predicting canonical poses for 3D point clouds [52], or through optical flow analysis using successive video frames [49]. They have also been applied to video tasks, for example, by transforming 2D convolutional capsule routing to 3D [20] and by adopting attention-based routing [21]. Lastly, CapsNets share foundational concepts with slot-based object-centric methods [15], such as [43], which uses an iterative attention mechanism akin to capsule routing, whereas the slots in [41] specialise in distinct entities as capsules. + +# 4. EquiCaps: Predictor-Free Pose-Aware SSL + +Motivation. Prior equivariant self-supervised methods [19, 24, 46] commonly rely on predictor networks to enforce equivariant representations via augmentation-aware mappings. This leads to ad hoc vector representations that are both difficult to interpret or manipulate—an important consideration for evaluating equivariance. In effect, these methods attempt to enforce equivariance on architectures (i.e., CNNs) that are not naturally suited to it [15] and often introduce extra complexity through dual projectors, representation splits, or hypernetworks. Instead, we exploit the equivariant properties elicited by CapsNets [15], to more effectively encode positional information, particularly rotation and translation. Our design yields a simpler pose-aware SSL framework, outperforms previous equivariant methods on geometric tasks, and enables a more intuitive control and interpretability in the embedding space through the structured $4 \times 4$ capsule pose matrices. We introduce and rationalise each component of our methodology as follows. + +Framework overview. Our method adopts a Siamese joint embedding setup with identical, weight-sharing branches, each formed by a CNN $f_{\theta}$ and a CapsNet projec- + +tor $h_{\phi}$ . We omit global average pooling to preserve the 2D spatial structure necessary for capsule processing. Given an image $x \in \mathbb{R}^{c \times h \times w}$ from a dataset $D$ , we sample two transformations $g_{1}, g_{2}$ from a distribution over the group $G$ , as detailed in supplementary Sec. 14. These are applied to $x$ to produce two distinct views, $x_{g_1} = \rho_X(g_1) \cdot x$ and $x_{g_2} = \rho_X(g_2) \cdot x$ . Both views are passed through $f_{\theta}$ , producing feature maps $F, F' \in \mathbb{R}^{c \times h \times w}$ . These are then processed by $h_{\phi}$ , which contains $N$ capsules, yielding $z_{\mathrm{act}}, z_{\mathrm{pose}} = h_{\phi}(F)$ and $z_{\mathrm{act}}', z_{\mathrm{pose}}' = h_{\phi}(F')$ , where $z_{\mathrm{act}}, z_{\mathrm{act}}' \in \mathbb{R}^N$ are the activation probabilities, and $z_{\mathrm{pose}}, z_{\mathrm{pose}}' \in \mathbb{R}^{N \times 4 \times 4}$ are the corresponding pose matrices. The activation vectors encode transformation-invariant properties, representing semantic concepts that remain unchanged under $g_{1}, g_{2}$ . In contrast, the pose matrices capture transformation-equivariant properties, encoding an entity's pose, where each capsule represents a unique entity. A visual depiction of the general architecture is given in Fig. 2. + +Self-routing CapsNet projector. To reduce the computational overhead of iterative-routing, we adopt the noniterative self-routing algorithm [29] as our projector. The projector $h_{\phi}$ consists of two layers: a primary capsule layer that transforms encoder feature maps into poses $(u_i)$ and activations $(a_i)$ , and a self-routing layer, where each capsule independently determines its routing coefficients via a subordinate single-layer perceptron. To do this, the routing coefficients $c_{ij}$ between each lower-layer capsule $i$ and each upper-level capsule $j$ are computed by multiplying the lower capsule's pose vector $u_i$ with a learnable weight matrix $W^{\mathrm{route}}$ and applying a softmax function. The activation $a_j$ of each upper-level capsule is then computed by forming votes from lower-level capsules—each vote is the product $c_{ij}a_i$ , which are then summed and normalised by the total lower-level activation, to formulate weighted votes: + +$$ +c _ {i j} = \operatorname {s o f t m a x} \left(W _ {i} ^ {\text {r o u t e}} u _ {i}\right) _ {j}, \quad a _ {j} = \frac {\sum_ {i \in \Omega_ {l}} c _ {i j} a _ {i}}{\sum_ {i \in \Omega_ {l}} a _ {i}}. \tag {3} +$$ + +The capsule layer's output pose is computed by multiplying a trainable weight matrix $W^{\text{pose}}$ with each lower-level pose $u_i$ , yielding a predicted pose $\hat{u}_{j|i}$ for each upper-level capsule. The final pose $u_j$ is the weighted average of $\hat{u}_{j|i}$ : + +$$ +\hat {u} _ {j \mid i} = W _ {i j} ^ {\text {p o s e}} u _ {i}, \quad u _ {j} = \frac {\sum_ {i \in \Omega_ {l}} c _ {i j} a _ {i} \hat {u} _ {j \mid i}}{\sum_ {i \in \Omega_ {l}} c _ {i j} a _ {i}}. \tag {4} +$$ + +Enforcing equivariance. We define the relative transformation $g_{\mathrm{rel}} = g_2g_1^{-1}$ , mapping between the views $x_{g_1}$ and $x_{g_2}$ , and let $\rho_Y(g_{\mathrm{rel}})$ denote its matrix representation. To enforce equivariance (defined in Eq. (1)) we multiply $z_{\mathrm{pose}}$ by $\rho_Y(g_{\mathrm{rel}})$ , aligning the transformed result with $z_{\mathrm{pose}}'$ . This indicates that the pose embeddings are equivariant w.r.t. the transformation group $G$ , ensuring that geometric changes in the input are consistently captured in the latent space. Each pose matrix $(z_{\mathrm{pose}}\cdot \rho_Y(g_{\mathrm{rel}}))$ and $z_{\mathrm{pose}}'$ + +is normalised by its Frobenius norm and residing on a unit hypersphere. This direct manipulation in the latent space removes the need for a dedicated predictor module. + +Loss functions. To enforce invariance we minimise the cross-entropy $H(Z_{\mathrm{act}}, Z_{\mathrm{act}}^{\prime})$ between the activation probability vectors $Z_{\mathrm{act}}$ and $Z_{\mathrm{act}}^{\prime}$ , where $Z$ denotes the matrix embeddings computed over a batch. This aligns the capsule activations for different views of the same image. To ensure full capsule utilisation, we apply mean entropy maximisation (ME-MAX) regularisation [1, 2, 36] to the activation probabilities. Let $\bar{Z}_{\mathrm{act}} = \frac{1}{B}\sum_{i=1}^{B}Z_{\mathrm{act}}$ , where $B$ is the batch size. Maximising $H(Z_{\mathrm{act}})$ and $H(\bar{Z}_{\mathrm{act}}^{\prime})$ encourages a diverse distribution of activation across all capsules, preventing the collapse to only a few highly active capsules. To enforce equivariance we minimise the mean-squared error between the normalised, transformed pose $(Z_{\mathrm{pose}}\cdot\rho_{Y}(g_{\mathrm{rel}}))$ and the corresponding pose $Z_{\mathrm{pose}}^{\prime}$ . + +To prevent collapse, we apply the variance-covariance regularisation criterion [3] to the concatenated capsule activation vectors and pose matrices, collectively denoted as $Z_{\mathrm{cat}}$ . Regularising these combined embeddings encourages diversity and non-redundancy across both activation and pose dimensions. The regularisation loss is defined as: + +$$ +\mathcal {L} _ {\text {r e g}} \left(Z _ {\text {c a t}}\right) = \lambda_ {V} V \left(Z _ {\text {c a t}}\right) + \lambda_ {C} C \left(Z _ {\text {c a t}}\right), \tag {5} +$$ + +$$ +V \left(Z _ {\text {c a t}}\right) = \frac {1}{d} \sum_ {j = 1} ^ {d} \max \left(0, 1 - \sqrt {\operatorname {V a r} \left(Z _ {\text {c a t} , j}\right)}\right), \tag {6} +$$ + +$$ +C \left(Z _ {\text {c a t}}\right) = \frac {1}{d} \sum_ {i \neq j} \operatorname {C o v} \left(Z _ {\text {c a t}}\right) _ {i, j} ^ {2}, \tag {7} +$$ + +where $V(Z_{\mathrm{cat}})$ ensures that each dimension is utilised and $C(Z_{\mathrm{cat}})$ decorrelates different dimensions to reduce redundant information. + +While SIE [24] and CapsIE [23] apply this regularisation before incorporating transformation parameters, we compute it directly on the transformation-aligned embedding. This approach yields improved performance in our experiments, as it operates on the final, fully aligned representation. Our overall loss function is defined as: + +$$ +\mathcal {L} = \lambda_ {\mathrm {i n v}} H \left(Z _ {\mathrm {a c t}}, Z _ {\mathrm {a c t}} ^ {\prime}\right) + +$$ + +$$ +H \left(\bar {Z} _ {\mathrm {a c t}}\right) + H \left(\bar {Z} _ {\mathrm {a c t}} ^ {\prime}\right) + +$$ + +$$ +\lambda_ {\mathrm {e q u i}} \frac {1}{B} \sum_ {i = 1} ^ {B} \left\| \frac {Z _ {i , \mathrm {p o s e}} \cdot \rho_ {Y} (g _ {\mathrm {r e l} , i})}{\| Z _ {i , \mathrm {p o s e}} \cdot \rho_ {Y} (g _ {\mathrm {r e l} , i}) \| _ {F}} - \frac {Z _ {i , \mathrm {p o s e}} ^ {\prime}}{\| Z _ {i , \mathrm {p o s e}} ^ {\prime} \| _ {F}} \right\| _ {2} ^ {2} + +$$ + +$$ +\mathcal {L} _ {\text {r e g}} \left(Z _ {\text {c a t}}\right) + \mathcal {L} _ {\text {r e g}} \left(Z _ {\text {c a t}} ^ {\prime}\right), \tag {8} +$$ + +where $\| \cdot \|_F$ denotes the Frobenius norm and $\lambda_{\mathrm{inv}} = 0.1, \lambda_V = 10, \lambda_C = 1$ , and $\lambda_{\mathrm{equi}} = 5$ . Despite the loss calculation occurs on the embeddings, the encoder's globally pooled output is used for downstream tasks as it yields better results [10]. Ablations, including the convergence speed + +and the impact of the number of capsules and invariant loss components, are given in supplementary Sec. 10. + +Removal of predictor module. Unlike previous equivariant methods such as SEN [46], EquiMod [19], SIE [24], and CapsIE [23] that use a dedicated predictor network to map one embedding to another, our approach eliminates this module. In those earlier works, the predictor $q$ is a network that enforces the equivariance condition (see Eq. (1)). Its input is $g_{rel}$ and the projector's output $z_{i}$ . It predicts the embedding of the transformed input, $z_{i}' = q(z_{i}, g_{rel})$ , where $q$ is a hypernetwork or linear layer. This removal simplifies the architecture and instead enforces the explicit mapping of transformations in the embedding space via affine transformation matrices applied to the pose embeddings rather than manipulating embeddings through an intermediate parameter-prediction stage. + +# 5. 3DIEBench-T: a translation-enhanced benchmark for invariant-equivariant SSL + +As highlighted by Garrido et al. [24], commonly used datasets for evaluating invariant or equivariant properties face limitations. Evaluations of invariant representations commonly benchmark on datasets like ImageNet [17] or CIFAR [39], where objects lack parameterisable transformations because augmentations occur only in the pixel domain. However, certain equivariant datasets [38, 46] lack visual complexity and transformational diversity to allow for fine-grained control. A more comprehensive alternative is 3DIEBench [24], which controls the scene and object parameters for 3D rotation prediction (equivariance) while retaining sufficient complexity for classification (invariance). Nonetheless, as with other similar datasets [54], 3DIEBench focuses only on rotation, evaluating methods under a single geometric transformation. + +Focussing only on rotation raises questions regarding the generalisability of methods [23, 24] to other geometric transformations, and their applicability in more realistic settings. To address this, we extend 3DIEBench by introducing additional translational transformations. Our new dataset, 3DIEBench-T (3D Invariant Equivariant Benchmark-Translated), goes beyond rotation-only benchmarks [24, 54], enabling translation prediction and providing a more challenging, realistic setting to assess how methods handle simultaneous multi-geometric transformations. + +To investigate the impact of the added transformations, we adhere to the original 3DIEBench data generation protocol, ensuring any observed performance differences are attributed to the inclusion of translations rather than broader dataset modifications. We use 52,472 3D object instances spanning 55 classes from ShapeNetCoreV2 [9], originally sourced from 3D Warehouse [35]. For each instance, we generate 50 uniformly sampled views within specified ranges via Blender-Proc [18], yielding 2,623,600 images. + +![](images/43fc3ad6b664518ae8f3d9549addce56daa0e255e89acb03b06be8656b285fa6.jpg) +Figure 3. Sample object instances from 3DIEBench-T dataset under different rotation, translation, and colour hue transformations. + +Samples appear in Fig. 3, with further generation information and visualisations given in supplementary Sec. 14. + +# 6. Experimental results + +Setup. To ensure fair comparisons, we adopt the experimental setup from [24] across all methods, which aligns with standard SSL practices: a pre-training phase followed by various task-specific downstream evaluations. A detailed description of the experimental setup is given in supplementary Sec. 9. + +Baselines. As an upper bound, we include a supervised baseline using ResNet-18 [30]. For invariant self-supervision, we benchmark against VICReg [3], a noncontrastive method, and SimCLR [10], a contrastive approach. We also consider SimCLR + AugSelf [40], a parameter-prediction extension of SimCLR that learns the augmentation parameters but does not enforce explicit equivariance. For equivariant self-supervision, we compare against SEN [46], EquiMod [19], SIE [24], and CapsIE [23]. Following [24, 61], we replace SEN's triplet loss with the SimCLR loss to avoid hyperparameter tuning. + +Evaluation protocol. For the invariant task, we perform classification via the linear evaluation protocol, where a linear classification layer is trained on the frozen representations. To assess equivariant properties, we perform rotation and translation prediction tasks. For each task, a three-layer MLP is trained on the frozen representations. Its inputs are concatenated pairs of representations from two views of the same object, and is trained to regress the transformation parameters—rotation for the rotation task and translation for the translation task. To explore invariance to non-geometric transformations that are not optimised in pre-training, a linear layer is trained on the frozen representations to regress the floor and spot hue. The input representations are identical to those used in the geometric tasks. To assess the performance of predicting these parameters, the coefficient of determination $R^2 = 1 - \frac{\sum_i(y_i - \hat{y}_i)^2}{\sum_i(y_i - \bar{y})^2}$ is used, where $y_{i}$ are the target values, $\bar{y}$ is their mean, and $\hat{y}_i$ are the predicted + +Table 1. Evaluation on 3DIEBench and 3DIEBench-T using a ResNet-18 backbone. Learned representations are evaluated on an invariant task (classification) and equivariant tasks (rotation, translation, and colour prediction). For each group of methods, the best value is shown in bold. Supervised baselines are trained for each task. Equivariant methods are trained to be rotation-equivariant, with no explicit constraints for translation or colour prediction. + +
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
3DIEBench3DIEBench-T3DIEBench3DIEBench-T3DIEBench-T3DIEBench3DIEBench-T
Supervised Methods
ResNet-1886.4580.130.770.730.670.990.99
Invariant and Parameter Prediction Methods
VICReg84.2874.710.450.390.220.100.50
SimCLR86.7380.080.520.440.250.290.50
SimCLR + AugSelf87.4480.860.750.690.500.280.51
Equivariant Methods
SEN86.9980.200.510.450.260.290.47
EquiMod87.3980.760.500.430.240.290.38
SIE82.9475.560.730.450.200.070.46
CapsIE79.1475.200.740.600.460.010.03
EquiCaps83.2476.910.780.730.600.090.05
+ +values. For geometric tasks, a larger $R^2$ denotes a better fit for the predictions, whereas for the colour task, an $R^2$ close to 0 is better, as no colour-specific pre-training constraint is enforced. Further details are given in supplementary Sec. 9. + +# 6.1. Rotation-only pre-training + +Tab. 1 summarises the downstream performance of the learned representations on both 3DIEBench and 3DIEBench-T. Specifically, 3DIEBench evaluates classification, rotation, and colour (hue) tasks, while 3DIEBench-T adds translations. Notably, all equivariant methods in Tab. 1 are optimised solely for rotation equivariance during pretraining. As observed in prior works, invariant methods fail to learn successful equivariant representations of the rotations and translations, while the addition of predictive tasks (AugSelf) improves their capability to quantify this change. + +We find that EquiCaps consistently outperforms SIE—its closest competitor—marginally on classification but significantly on geometric equivariant tasks, achieving $+0.28 R^2$ in rotation and $+0.40 R^2$ in translation on 3DIEBench-T. Notably, our capsule-based projector is able to intrinsically capture translational attributes, despite no direct supervision. CapsIE also generalises relatively better than other equivariant methods, but it underperforms our method, and validates our claim that its predictor-based design underuses the inherent pose-awareness capabilities of CapsNets. Nevertheless, we find that non-capsule architectures do not generalise well under more complex, combined 3D transformations. We attribute this limitation to their lack of an explicit, pose-aware mechanism. Results on the Objaverse [16] and MOVi-E [26] datasets are given in supplementary Sec. 11. + +# 6.2. Rotation-translation pre-training + +In Tab. 2, we present results when both rotation and translation parameters are optimised. As expected, all meth + +ods exhibit a small improvement in translation performance. However, the two capsule-based methods retain their substantial improvement on geometric tasks. Consistent with [55], we find that SIE's gains on 3DIEBench do not generalise well to multi-geometric learning. We further find that SIE exhibits minimal gains over purely invariant methods and maintains the weakest translation prediction, even when explicitly trained for translation equivariance. + +Table 2. Evaluation on 3DIEBench-T pre-trained for both rotation and translation. Learned representations (the encoder's outputs) are evaluated on an invariant and equivariant tasks: classification, rotation, translation, and colour prediction. We indicate the change relative to Tab. 1. + +
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
SimCLR + AugSelf81.04 ↑ 0.180.69 = 0.000.64 ↑ 0.140.51 = 0.00
SEN80.23 ↑ 0.030.46 ↑ 0.010.28 ↑ 0.020.50 ↑ 0.03
EquiMod80.89 ↑ 0.130.46 ↑ 0.030.37 ↑ 0.130.37 ↓ 0.01
SIE75.91 ↑ 0.350.48 ↑ 0.030.22 ↑ 0.020.36 ↓ 0.10
CapsIE76.31 ↑ 1.110.62 ↑ 0.020.53 ↑ 0.070.03 = 0.00
EquiCaps77.88 ↑ 0.970.71 ↓ 0.020.61 ↑ 0.010.02 ↓ 0.03
+ +Table 3. Evaluation on 3DIEBench and 3DIEBench-T pre-trained for both rotation and translation. Learned embeddings (the projector's outputs) are evaluated on equivariant tasks: rotation and translation prediction. + +
MethodRotation (R2)Translation (R2)
3DIEBench3DIEBench-T3DIEBench-T
VICReg0.240.240.00
SimCLR0.240.240.00
SimCLR + AugSelf0.240.240.00
SEN0.240.240.00
EquiMod0.240.240.22
SIE0.600.340.00
CapsIE0.710.580.26
EquiCaps0.730.610.33
+ +Table 4. Equivariance evaluation on 3DIEBench (top) and 3DIEBench-T (bottom), using MRR, H@k, and separate PRE measures for rotation and translation on 3DIEBench-T. We indicate which subset (train or val) is used to generate embeddings, and for PRE, which dataset (train, val, or all) is used for retrieval. For 3DIEBench-T, methods are pre-trained for both rotation and translation equivariance. + +
MethodMRR (↑)H@1 (↑)H@5 (↑)PRE (↓)
trainvaltrainvaltrainvaltrain-trainval-valval-all
AllAllAllAllAllAllRotationTranlationRotationTranlationRotationTranlation
3DIEBench
SEN0.090.090.020.020.100.100.61-0.61-0.61-
EquiMod0.090.090.020.020.100.100.61-0.61-0.61-
SIE0.510.410.410.300.600.510.26-0.29-0.27-
CapsIE0.600.470.500.360.710.580.17-0.21-0.20-
EquiCaps0.680.600.550.480.840.730.13-0.18-0.17-
3DIEBench-T
SEN0.090.090.020.020.100.100.612.710.622.720.622.72
EquiMod0.130.120.030.030.180.150.582.470.592.510.592.50
SIE0.150.140.060.050.200.180.542.730.542.720.542.73
CapsIE0.460.370.350.260.550.470.292.450.312.480.302.46
EquiCaps0.590.510.480.410.700.620.302.340.352.410.332.38
+ +Colour prediction. EquiCaps maintains near-perfect invariance to hue transformations, consistent with prior work such as SIE and CapsIE, although SIE is less effective under 3DIEBench-T. By contrast, non-capsule methods exhibit substantially increased colour prediction when transitioning to 3DIEBench-T. We conjecture that the dataset's increased geometric complexity makes the colour task easier to solve for models lacking explicit pose-aware mechanisms. + +# 6.3. Performance on embeddings + +As in prior works [10, 24], we observe that representations (the outputs of the CNN encoder $f_{\theta}$ ) learned by invariant methods achieve higher-than-expected performance on equivariant tasks. This arises because the invariance objective is enforced on the embeddings (the outputs of the projector $h_{\phi}$ ) after the projector, while the representations can still retain information [10]. To explore this further, we evaluate rotation and translation predictions on the embeddings (see Tab. 3), where all methods show decreased performance compared to their representation-level results, highlighting that the projector absorbs more invariance. We also observe that invariant methods, along with SEN and EquiMod, converge to similar embedding-level results, except for EquiMod, which retains higher translation prediction performance, as in the representation-stage results. SIE aligns with invariant methods on translation prediction, consistent with its representation stage. EquiCaps maintains the best performance on both tasks, and capsule-based methods show the least decline from representations to embeddings, underscoring the advantages of capsules' pose-awareness. + +# 6.4. Equivariance evaluation + +We employ and adapt the Mean Reciprocal Rank (MRR), Hit Rate at k (H@k), and Prediction Retrieval Error (PRE) metrics from [24]. For compared methods, for each object's + +source and target poses, we feed the source embeddings and the corresponding transformation parameters through the predictor, obtaining a predicted embedding. To adapt it to our method, we multiply the projector's pose matrix by the transformation matrix. We obtain the nearest neighbours of the predicted embedding among the views of the same object. MRR is computed as the mean of the reciprocal rank $\frac{1}{r}$ , where $r$ is the rank of the target embedding among the neighbours. H@k yields 1 if the target is within the top- $k$ neighbours, and 0 otherwise. PRE measures the mean distance between the transformation parameters of the nearest neighbour and the target across the dataset. Rotation uses $d_r = 1 - \langle q_1, q_2 \rangle^2$ , where $q_1$ and $q_2$ are the nearest neighbour's and target's rotation represented as a quaternion, respectively. Translation uses the mean-squared error $d_t = ||t_1 - t_2||_2^2$ , where $t_1$ is the nearest neighbour's translation vector and $t_2$ is the target's translation vector. + +As shown in Tab. 4, EquiCaps outperforms all methods across each data split of 3DIEBench, achieving an MRR of 0.60 on the validation set, well above SIE (0.41) and CapsIE (0.47). Similar results are observed in H@1/5 metrics, where SEN and EquiMod perform randomly (2% and 10%) on both data splits, while CapsIE's PRE on all data splits approaches ours. Although EquiCaps generally outperforms all methods on 3DIEBench-T across most metrics, CapsIE achieves slightly lower rotation PRE. Despite this, our method remains significantly stronger in MRR and H@1/5 across all data splits, with EquiMod slightly improving. In contrast, SIE's overall performance drops with H@1 of 0.05 on the validation set, marginally above random. + +# 6.5. Qualitative results + +Nearest-neighbour representations. We perform a nearest-neighbour representation search on the 3DIEBench- + +![](images/2ea6445c95794947c54b78747670a08ea7c63212438a01c07a0271e24b7bbe45.jpg) +Figure 4. Nearest-neighbour representation retrieval on 3DIEBench-T validation set directly after pre-training. The query image (left) is compared against each method's learned representations to find its top three nearest neighbours (in rows: 1-NN, 2-NN, 3-NN). Equivariant and parameter prediction methods are trained to be both rotationally and translationally equivariant. + +T validation set (Fig. 4) to illustrate the information captured by the representations. Invariant approaches should retrieve objects in the same class but varied poses, whereas equivariant methods should preserve the query pose. We find that only our method and CapsIE retrieve the nearest neighbour with a matching pose, though ours aligns also with the translation. For subsequent neighbours, only EquiCaps and AugSelf retrieve objects in similar poses, while CapsIE does so but less accurately. In contrast, the other methods, including equivariant methods, fail to yield pose-matching neighbours suggesting that invariance dominates and their representations leave out significant pose information. We replicate the retrieval on 3DIEBench Fig. A.6, supplementary Sec. 12, where ours, SIE, and CapsIE successfully return pose-matching neighbours, but remaining methods, only partially align poses, albeit less accurate. We conjecture this is attributed to the projector absorbing invariance, while partial rotation-related information remains at the representation stage. All retrieved objects belong to the car category, indicating that class semantics are preserved consistent with our quantitative results. + +Equivariance via pose manipulation. To further illustrate the equivariant properties of our approach, we examine the inversion property through the inverse rotation of an object in the embedding space. We rotate an object around the $z$ -axis within the range $[0, \frac{\pi}{2}]$ in $5^{\circ}$ increments and feed these images through our trained network to obtain its pose embedding. Embedding is subsequently multiplied by its corresponding inverse transformation matrix. To examine the impact of the inverse transformation we retrieve the nearest neighbour among all embeddings. As shown in Fig. 5, the nearest neighbour closely (within $5^{\circ}$ error) or exactly matches the source view, highlighting the invertibility of our representations, and the ability to intuitively and directly steer the latent space. Further qualitative results are provided in supplementary Sec. 12. + +![](images/888dcae3d3a87acacc900a20db579cfbe82e8d99abddd932b8366290c0bd3f33.jpg) +Figure 5. Illustration of transformation inversion of equivariant embeddings through capsule-based pose manipulation. + +# 7. Conclusion + +We propose EquiCaps, a capsule-based SSL method that jointly learns invariant and equivariant representations. By leveraging capsules' innate pose-awareness, EquiCaps shapes the latent space predictably, removing the need for a dedicated predictor. To enable richer benchmarking, we extend 3DIEBench to 3DIEBench-T, incorporating object translations alongside rotations. Empirically, EquiCaps achieves state-of-the-art rotation and translation prediction among equivariant methods. Tests on 3DIEBench-T highlight capsule architectures' superior generalisation under combined geometric transformations, demonstrating their potential in invariant-equivariant SSL. + +Limitations and future work. Joint invariance-equivariance optimisation involves trade-offs; while EquiCaps balances these, task-specific priorities may vary. The method assumes known group elements—as prior methods do—restricting its use without such data, though partial data suffices. EquiCaps can theoretically handle any transformation which can be expressed as a matrix without architectural changes. Future work includes other 3D transformations (e.g., scaling, shear), and more object detection tasks. + +# 8. Acknowledgements + +This work was supported by the UKRI AI Centre for Doctoral Training in Sustainable Understandable agri-food Systems Transformed by Artificial INtelligence (SUSTAIN) [grant reference: EP/Y03063X/1]. We also acknowledge the generous support of alumni and friends in establishing the University of Aberdeen's Interdisciplinary Institute, which partly enabled this research, including Dr Jane Hellman Caseley (MBChB 1956), Professor Patrick Meares (DSc 1959), Nancy Miller (MA 1942), Norman Robertson, Dr Ian Sessler (MBChB 1956) and Anne Young (MA 1957). Finally, we thank the University of Aberdeen's HPC Maxwell facility for providing the computational resources that facilitated our experimental work. + +# References + +[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, pages 8443-8452, 2021. +[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022. +[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022. +[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VICRegl: Self-supervised learning of local visual features. In NeurIPS, pages 8799-8810, 2022. +[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020. +[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018. +[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020. +[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021. +[9] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, QiXing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015. +[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge + +offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020. +[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021. +[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020. +[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021. +[14] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. In ICLR, 2022. +[15] Fabio De Sousa Ribeiro, Kevin Duarte, Miles Everett, Georgios Leontidis, and Mubarak Shah. Object-centric learning with capsule networks: A survey. ACM Comput. Surv., 56 (11), 2024. +[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, pages 13142-13153, 2023. +[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009. +[18] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. BlenderProc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023. +[19] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve self-supervised learning. In ICLR, 2023. +[20] Kevin Duarte, Yogesh Rawat, and Mubarak Shah. Videocapsulenet: A simplified network for action detection. In NeurIPS, 2018. +[21] Kevin Duarte, Yogesh S. Rawat, and Mubarak Shah. Capsulevos: Semi-supervised video object segmentation using capsule routing. In ICCV, 2019. +[22] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, pages 3015-3024, 2021. +[23] Miles Everett, Aiden Durrant, Mingjun Zhong, and Georgios Leontidis. Capsule network projectors are equivariant and invariant learners. arXiv preprint arXiv:2405.14386, 2024. +[24] Quentin Garrido, Laurent Najman, and Yann LeCun. Self-supervised learning of split invariant equivariant representations. In ICML, 2023. +[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018. +[26] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek + +Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. In CVPR, pages 3749-3761, 2022. +[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020. +[28] Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and Stefanie Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. In ICLR, 2024. +[29] Taeyoung Hahn, Myeongjiang Pyeon, and Gunhee Kim. Selfrouting capsule networks. In NeurIPS, 2019. +[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. +[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020. +[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. +[33] Geoffrey E. Hinton, Alex Krizhevsky, and Sida Wang. Transforming auto-encoders. In ICANN, pages 44–51, 2011. +[34] Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with EM routing. In ICLR, 2018. +[35] Trimble Inc. 3D warehouse. https://3dwarehouse.sketchup.com/. Accessed: 2025-03-01. +[36] Armand Joulin and Francis Bach. A convex relaxation for weakly supervised classifiers. In ICML, pages 1279-1286, 2012. +[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015. +[38] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. In ICLR, 2020. +[39] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009. +[40] Hankook Lee, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. Improving transferability of representations via augmentation-aware self-supervision. In NeurIPS, 2021. +[41] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In ICCV, pages 1046-1055, 2021. +[42] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, 2024. +[43] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob + +Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, pages 11525-11538, 2020. +[44] Giovanni Luca Marchetti, Gustaf Tegnér, Anastasiia Varava, and Danica Kragic. Equivariant representation learning via class-pose decomposition. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 4745-4756, 2023. +[45] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024. +[46] Jung Yeon Park, Ondrej Biza, Linfeng Zhao, Jan Willem van de Meent, and Robin Walters. Learning symmetric embeddings for equivariant world models. In ICML, 2022. +[47] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In AAAI, pages 3749-3756, 2020. +[48] Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. Dynamic routing between capsules. In NeurIPS, pages 3856-3866, 2017. +[49] Sara Sabour, Andrea Tagliasacchi, Soroosh Yazdani, Geoffrey E. Hinton, and David J. Fleet. Unsupervised part representation by flow capsules. In ICML, pages 9213-9223, 2021. +[50] Franz Scherr, Qinghai Guo, and Timoleon Moraitis. Self-supervised learning through efference copies. In NeurIPS, 2022. +[51] Mehran Shakerinava, Arnab Kumar Mondal, and Siamak Ravanbakhsh. Structuring representations using group invariants. In NeurIPS, pages 34162-34174, 2022. +[52] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey E. Hinton, and Kwang Moo Yi. Canonical capsules: Self-supervised capsules in canonical pose. In NeurIPS, pages 24993-25005, 2021. +[53] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020. +[54] Jiayun Wang, Yubei Chen, and Stella Yu. Pose-aware self-supervised learning with viewpoint trajectory regularization. In ECCV, 2024. +[55] Qin Wang, Kai Krajsek, and Hanno Scharr. Equivariant representation learning for augmentation-based self-supervised learning via image reconstruction. In NeurIPS 2024 Workshop: Self-Supervised Learning - Theory and Practice, 2024. +[56] Yifei Wang, Kaiwen Hu, Sharut Gupta, Ziyu Ye, Yisen Wang, and Stefanie Jegelka. Understanding the role of equivariance in self-supervised learning. In NeurIPS, pages 127483-127510, 2024. + +[57] Robin Winter, Marco Bertolini, Tuan Le, Frank Noe, and Djork-Arné Clevert. Unsupervised learning of group invariant and equivariant representations. In NeurIPS, 2022. +[58] Tete Xiao, Xiaolong Wang, Alexei A. Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021. +[59] Yuyang Xie, Jianhong Wen, Kin Wai Lau, Yasar Abbas Ur Rehman, and Jiajun Shen. What should be equivariant in self-supervised learning. In CVPR Workshop, pages 4111-4120, 2022. +[60] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684, 2022. +[61] Jaemyung Yu, Jaehyun Choi, Dong-Jae Lee, HyeongGwon Hong, and Junmo Kim. Self-supervised transformation learning for equivariant representations. In NeurIPS, 2024. +[62] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320, 2021. \ No newline at end of file diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/images.zip b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..415430988c8698027016b865c14a8987141eb7d6 --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9259faab3c30da81e98b3ef6a6d6ea5e442ce142a563cbecd51ae63f3bd415b5 +size 409576 diff --git a/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/layout.json b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..14ae7a170d5bc1d929b2fe159e82a13eb05c9ee1 --- /dev/null +++ b/2025/EquiCaps_ Predictor-Free Pose-Aware Pre-Trained Capsule Networks/layout.json @@ -0,0 +1,9316 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "spans": [ + { + "bbox": [ + 89, + 103, + 521, + 121 + ], + "type": "text", + "content": "EquiCaps: Predictor-Free Pose-Aware Pre-Trained Capsule Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 142, + 324, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 142, + 324, + 171 + ], + "spans": [ + { + "bbox": [ + 77, + 142, + 324, + 171 + ], + "type": "text", + "content": "Athinoula Konstantinou1 Georgios Leontidis1 \n1University of Aberdeen, UK" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 338, + 143, + 533, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 143, + 533, + 171 + ], + "spans": [ + { + "bbox": [ + 338, + 143, + 533, + 171 + ], + "type": "text", + "content": "Mamatha Thota² Aiden Durrant¹ \n²University of Lincoln, UK" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 68, + 173, + 538, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 68, + 173, + 538, + 185 + ], + "spans": [ + { + "bbox": [ + 68, + 173, + 538, + 185 + ], + "type": "text", + "content": "{a.konstantinou.24, georgios.leontidis, aiden.durrant}@abdn.ac.uk, mthota@lincoln.ac.uk" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "text", + "content": "Learning self-supervised representations that are invariant and equivariant to transformations is crucial for advancing beyond traditional visual classification tasks. However, many methods rely on predictor architectures to encode equivariance, despite evidence that architectural choices, such as capsule networks, inherently excel at learning interpretable pose-aware representations. To explore this, we introduce EquiCaps (Equivariant Capsule Network), a capsule-based approach to pose-aware self-supervision that eliminates the need for a specialised predictor for enforcing equivariance. Instead, we leverage the intrinsic pose-awareness capabilities of capsules to improve performance in pose estimation tasks. To further challenge our assumptions, we increase task complexity via multi-geometric transformations to enable a more thorough evaluation of invariance and equivariance by introducing 3DIEBench-T, an extension of a 3D object-rendering benchmark dataset. Empirical results demonstrate that EquiCaps outperforms prior state-of-the-art equivariant methods on geometric tasks, including rotation and translation, achieving a supervised-level " + }, + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "text", + "content": " of 0.78 on the 3DIEBench rotation prediction benchmark and improving upon SIE and CapsIE by 0.05 and 0.04 " + }, + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 238, + 296, + 586 + ], + "type": "text", + "content": ", respectively. Moreover, in contrast to non-capsule-based equivariant approaches, EquiCaps maintains robust equivariant performance under combined geometric transformations, underscoring its generalisation capabilities and the promise of predictor-free capsule architectures. Code, dataset, and weights are released at http://github.com/AberdeenML/EquiCaps." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "The success of self-supervised learning (SSL) is largely attributed to the exploitation of invariant representations under data augmentations and transformations to learn rich semantic concepts [3, 4, 7, 8, 10, 12, 27, 31, 32, 62]. However, methods that strictly enforce invariance fail to preserve useful augmentation and transformation information, which could limit their applicability to downstream tasks where" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 212, + 542, + 358 + ], + "blocks": [ + { + "bbox": [ + 328, + 212, + 542, + 358 + ], + "lines": [ + { + "bbox": [ + 328, + 212, + 542, + 358 + ], + "spans": [ + { + "bbox": [ + 328, + 212, + 542, + 358 + ], + "type": "image", + "image_path": "caa2dbd19f946af1435a4c2980b43c2ef4537b2cbc3c342357c616c86a7cad4b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 327, + 363, + 542, + 445 + ], + "blocks": [ + { + "bbox": [ + 327, + 363, + 542, + 445 + ], + "lines": [ + { + "bbox": [ + 327, + 363, + 542, + 445 + ], + "spans": [ + { + "bbox": [ + 327, + 363, + 542, + 445 + ], + "type": "image", + "image_path": "64e325b9400dc2d2d0ac14b623fe1f3cb892e9d6cdfed833baeec4899340521c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 453, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 498 + ], + "type": "text", + "content": "Figure 1. Equivariant evaluation of EquiCaps (ours) and competing methods on 3DIEBench benchmark. We indicate which subset (train or val) is used to generate embeddings, and, for PRE, which dataset (train, val, or all) is used for retrieval." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 521, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 521, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 312, + 521, + 555, + 605 + ], + "type": "text", + "content": "such information is vital [28, 40, 58]. In response, there has been increasing interest in developing more generalised self-supervised approaches, most notably through the investigation of pose-aware learning [24, 54, 56]. Achieving such generalisation, particularly to novel viewpoints, can consequently address real-world settings that inherently exhibit large data variations, e.g. autonomous driving [54]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 605, + 556, + 715 + ], + "type": "text", + "content": "One promising strategy to preserve transformation information—and thus achieve pose-awareness—is to directly optimise for equivariant representations, enabling the model to capture how transformations steer the latent space. However, such methods often face limitations [19, 23, 24, 46]. First, they rely on a dedicated predictor network to encode equivariant properties, often further increasing architectural complexity through conditional projectors, partitioned representations, or hypernet" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7947" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "works. Second, equivariance in representations is typically enforced solely via objective functions with little regard taken to exploit architectural mechanisms—such as Capsule Networks (CapsNets)—that have been shown to effectively encode or structure pose information [15]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 133, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 133, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 54, + 133, + 295, + 263 + ], + "type": "text", + "content": "We propose to simplify and enhance the framework of equivariant self-supervision to address these concerns by using CapsNets [15]. CapsNets structure representations into capsules, each encoding both the existence of an entity and its instantiation parameters, where the latter is achieved through a routing algorithm based on agreement of part-whole relationships. This formulation introduces an inductive bias that has been shown to effectively capture equivariance w.r.t. viewpoints and invariance in the capsule activations, enabling them to exhibit viewpoint-invariant and viewpoint-equivariant properties by design." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 264, + 296, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 443 + ], + "type": "text", + "content": "While prior attempts have been made to introduce CapsNets into equivariant SSL [23], they still rely on a separate transformation predictor, potentially underusing capsules' intrinsic pose-awareness capabilities. To overcome these limitations, we propose a streamlined pipeline that eliminates the predictor and directly leverages capsules' equivariant properties that prior equivariant SSL methods do not. CapsNets offer an intriguing prospect for capturing transformations, therefore we introduce an extension of the 3DIEBench dataset [24] that incorporates translational transformations in addition to rotations. By moving from rotation-only SO(3) transformations [24, 54] to the more general SE(3) group, our dataset increases the difficulty of both invariance and equivariance tasks, allowing for the evaluation of generalisation under multi-geometric settings." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 443, + 260, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 443, + 260, + 454 + ], + "spans": [ + { + "bbox": [ + 67, + 443, + 260, + 454 + ], + "type": "text", + "content": "To summarise, our contributions are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 456, + 295, + 648 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 515 + ], + "type": "text", + "content": "- 3DIEBench-T: We present 3DIEBench-T, an extension of the 3DIEBench that integrates translational transformations alongside rotations. This new benchmark enables the evaluation of representation methods under simultaneous multi-geometric 3D transformations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 517, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 517, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 517, + 295, + 563 + ], + "type": "text", + "content": "- Predictor limitations: We empirically demonstrate that existing predictor-based equivariant architectures yield minimal improvement over purely invariant approaches when learning multi-geometric 3D transformations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 564, + 295, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 564, + 295, + 648 + ], + "spans": [ + { + "bbox": [ + 56, + 564, + 295, + 648 + ], + "type": "text", + "content": "- Predictor-free CapsNet: We propose EquiCaps, removing the need for predictors by exploiting the intrinsic pose-awareness of CapsNets. Empirical results on rotation and translation tasks demonstrate that EquiCaps outperforms existing equivariant methods, and generalises successfully under multi-geometric settings, while also allowing intuitive manipulation of the embedding space." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 657, + 140, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 657, + 140, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 657, + 140, + 670 + ], + "type": "text", + "content": "2. Preliminaries" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 295, + 702 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 702 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 702 + ], + "type": "text", + "content": "We first introduce key notions of equivariance and invariance in our setting [24, 46]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "Group actions. A group is a set " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " with a binary operation " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " that satisfies associativity, " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "G \\times G \\to G" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", the existence of an identity element, " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "g \\cdot 1 = 1 \\cdot g = g" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\forall g \\in G" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", and the existence of an inverse, " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "g \\cdot g^{-1} = g^{-1} \\cdot g = 1" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". A group action of " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " on a set " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " is defined as a map " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\alpha : G \\times S \\to S" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " that is compatible with the composition of group elements, i.e., " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\alpha(g, \\alpha(h, s)) = \\alpha(gh, s)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\forall s \\in S" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\forall g, h \\in G" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". The map " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\alpha(g, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " is called a group representation when " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " is a vector space " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", such as " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^n" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": " is linear. A group representation is denoted as the map " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\rho : G \\to GL(V)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ", such that " + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "inline_equation", + "content": "\\rho(g) = \\alpha(g, \\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": ". As in similar works [24], we use group representations to describe how transformations are applied both to the input data and latent space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": "Equivalence learning. A function " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "f: X \\to Y" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": " is equivariant if for a group " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": ", with representations " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\rho_{X}" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "\\rho_{Y}" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "inline_equation", + "content": "Y" + }, + { + "bbox": [ + 313, + 228, + 553, + 264 + ], + "type": "text", + "content": ", satisfies" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 331, + 272, + 553, + 285 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 272, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 331, + 272, + 553, + 285 + ], + "type": "interline_equation", + "content": "f \\left(\\rho_ {X} (g) \\cdot x\\right) = \\rho_ {Y} (g) \\cdot f (x), \\forall x \\in X, \\forall g \\in G. \\tag {1}", + "image_path": "193a57f94a58f29588d9b8f1331af2d94b384247a72d739be11dd7fdd36fee80.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "text", + "content": "This indicates the commutativity of group transformations and function application. The function " + }, + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "text", + "content": " retains the symmetry group " + }, + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 293, + 553, + 329 + ], + "type": "text", + "content": " yet changes how the group operates." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "content": "Invariance learning. A function " + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "inline_equation", + "content": "f: X \\to Y" + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "content": " is invariant if for a group " + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "content": ", with representations " + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "inline_equation", + "content": "\\rho_{X}" + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 313, + 330, + 553, + 354 + ], + "type": "text", + "content": ", satisfies" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 347, + 362, + 553, + 375 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 362, + 553, + 375 + ], + "spans": [ + { + "bbox": [ + 347, + 362, + 553, + 375 + ], + "type": "interline_equation", + "content": "f \\left(\\rho_ {X} (g) \\cdot x\\right) = f (x), \\quad \\forall x \\in X, \\forall g \\in G. \\tag {2}", + "image_path": "e55975f1f172b53cb216c1bf72ae7c9436ba2adb269cfea97cafed172f2b2827.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "text", + "content": "Invariance is a trivial [14], special case of equivariance when " + }, + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "inline_equation", + "content": "\\rho_Y(g)" + }, + { + "bbox": [ + 313, + 384, + 553, + 408 + ], + "type": "text", + "content": ", in Eq. (1) is the identity map." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 418, + 403, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 403, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 403, + 430 + ], + "type": "text", + "content": "3. Related works" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 438, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 553, + 581 + ], + "type": "text", + "content": "Invariant SSL. Invariance to data augmentations has been the driving force of self-supervision, using this property to learn common semantic concepts under different transformations when there is no access to labels. Specifically, invariance in the latent space is enforced by aligning shared views that are transformations of the same source. There are various approaches to undertake this alignment and discrimination, notably contrastive methods [10, 12, 13, 31, 60], clustering-based learning [6-8, 45], and non-contrastive methods [3, 4, 11, 22, 27, 62] which instead employ architectural or objective modifications, gradient mechanisms, and information-theoretic regularisers." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Implicit equivariant SSL. Invariant SSL methods are optimised to disregard data transformations, hence information is lost. To address this, predictive equivariant SSL encodes transformation-sensitive information by training to predict transformation parameters [25, 40, 50], while others integrate contrastive learning with transformation prediction [14]. Furthermore, modulating the contrastive objective itself, based on the strength of the applied augmentation has shown promise [59], while [58] introduces distinct embedding spaces that remain invariant to all but one augmentation. Focussing on the transformation, optimal augmented" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7948" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 72, + 74, + 318, + 178 + ], + "blocks": [ + { + "bbox": [ + 72, + 74, + 318, + 178 + ], + "lines": [ + { + "bbox": [ + 72, + 74, + 318, + 178 + ], + "spans": [ + { + "bbox": [ + 72, + 74, + 318, + 178 + ], + "type": "image", + "image_path": "e3362aef672282506dea597c1e511667fcfa0d5ae1dd875df349f6f48eaee286.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 182, + 180, + 194, + 190 + ], + "lines": [ + { + "bbox": [ + 182, + 180, + 194, + 190 + ], + "spans": [ + { + "bbox": [ + 182, + 180, + 194, + 190 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 318, + 77, + 414, + 175 + ], + "blocks": [ + { + "bbox": [ + 318, + 77, + 414, + 175 + ], + "lines": [ + { + "bbox": [ + 318, + 77, + 414, + 175 + ], + "spans": [ + { + "bbox": [ + 318, + 77, + 414, + 175 + ], + "type": "image", + "image_path": "ded19a7e06175cdd94f8cebfe06aa3e6a31eed696eb1836a5dfc351d3d8ee4b3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 368, + 179, + 381, + 190 + ], + "lines": [ + { + "bbox": [ + 368, + 179, + 381, + 190 + ], + "spans": [ + { + "bbox": [ + 368, + 179, + 381, + 190 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": "Figure 2. (a) Framework Overview. Two views " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "x_{1}" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "x_{1}^{\\prime} = \\rho_{X}(g_{rel}) \\cdot x_{1}" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": " of the same object are processed by a shared CNN encoder " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": " and a capsule-based projector " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": ". (b) Projector's Output. For each view, the projector produces an activation probability vector " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "(z_{act}, z_{act}')" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": ", indicating entity presence, and a pose matrix " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "(z_{pose}, z_{pose}')" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": ", encoding the pose for each entity. (c) Equivariant Objective. Equivariance is enforced by multiplying " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "z_{pose}" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": " by the relative transformation matrix " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "\\rho_{Y}(g_{rel})" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": " and aligning the result with " + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "inline_equation", + "content": "z_{pose}'" + }, + { + "bbox": [ + 55, + 202, + 555, + 248 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 416, + 78, + 545, + 174 + ], + "blocks": [ + { + "bbox": [ + 416, + 78, + 545, + 174 + ], + "lines": [ + { + "bbox": [ + 416, + 78, + 545, + 174 + ], + "spans": [ + { + "bbox": [ + 416, + 78, + 545, + 174 + ], + "type": "image", + "image_path": "d6b3920457ede7f4cb24d3f97ac2254c2c23d40dc41ae4d581051579bf583c4d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 180, + 483, + 190 + ], + "lines": [ + { + "bbox": [ + 471, + 180, + 483, + 190 + ], + "spans": [ + { + "bbox": [ + 471, + 180, + 483, + 190 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 267, + 294, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 294, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 294, + 339 + ], + "type": "text", + "content": "views can also be exploited to retain task-relevant information [53]. However, these do not optimise for equivariance defined in Eq. (1). Hence, since they do not learn an explicit mapping from input transformations to corresponding changes in latent space, they do not guarantee equivariance, unlike explicit equivariant approaches [24, 56]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 341, + 295, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 341, + 295, + 568 + ], + "spans": [ + { + "bbox": [ + 56, + 341, + 295, + 568 + ], + "type": "text", + "content": "Explicit equivariant SSL. Avoiding transformation prediction tasks in favour of evaluating for equivariant representations, elicits a separate family of methods. These utilise representation reconstruction and maintain invariant and equivariant properties by splitting into or learning separate latent spaces [44, 57]. Others [51, 61] encourage equivariance through identically transformed image pairs, with [51] utilising the transformation group information for choosing optimal loss terms. Such approaches have also extended contrastive learning enforcing input transformations correspond to rotations in the embedding space [28]. Several methods adopt a predictor network to enforce an equivalent transformation mapping between latent representations through augmentation parameters [19, 46]. Building on these, SIE [24] incorporates a hypernetwork to learn such mapping, while dividing the representation into invariant and equivariant parts. CapsIE [23] adopts this module, instead leveraging a CapsNet projector, removing the need for multiple projectors and the split representation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 714 + ], + "type": "text", + "content": "CapsNets. CapsNets are designed to model part-whole hierarchical relationships by capturing the likelihood of an entity's presence and its attributes (pose, colour, texture, etc.) through neural activity. Lower-level capsules connect to higher-level ones via routing based on agreement in likelihood and positional attributes. This alignment enables capsules to identify objects by reinforcing pose consistency when multiple capsules converge on the same entity. Due to this positional agreement mechanism, CapsNets are explored as equivariant architectures, enabling tasks like image generation [33] and transformation learning [48]. Various routing methods have been proposed, including" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 267, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 555, + 423 + ], + "type": "text", + "content": "Gaussian-mixture clustering [34], Variational Bayes with uncertainty [47], self-routing via subordinate networks [29], and residual pose routing [42]. Like our approach, CapsNets have been adopted in Siamese SSL, notably by predicting canonical poses for 3D point clouds [52], or through optical flow analysis using successive video frames [49]. They have also been applied to video tasks, for example, by transforming 2D convolutional capsule routing to 3D [20] and by adopting attention-based routing [21]. Lastly, CapsNets share foundational concepts with slot-based object-centric methods [15], such as [43], which uses an iterative attention mechanism akin to capsule routing, whereas the slots in [41] specialise in distinct entities as capsules." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 438, + 549, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 549, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 549, + 453 + ], + "type": "text", + "content": "4. EquiCaps: Predictor-Free Pose-Aware SSL" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 460, + 555, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 675 + ], + "type": "text", + "content": "Motivation. Prior equivariant self-supervised methods [19, 24, 46] commonly rely on predictor networks to enforce equivariant representations via augmentation-aware mappings. This leads to ad hoc vector representations that are both difficult to interpret or manipulate—an important consideration for evaluating equivariance. In effect, these methods attempt to enforce equivariance on architectures (i.e., CNNs) that are not naturally suited to it [15] and often introduce extra complexity through dual projectors, representation splits, or hypernetworks. Instead, we exploit the equivariant properties elicited by CapsNets [15], to more effectively encode positional information, particularly rotation and translation. Our design yields a simpler pose-aware SSL framework, outperforms previous equivariant methods on geometric tasks, and enables a more intuitive control and interpretability in the embedding space through the structured " + }, + { + "bbox": [ + 313, + 460, + 555, + 675 + ], + "type": "inline_equation", + "content": "4 \\times 4" + }, + { + "bbox": [ + 313, + 460, + 555, + 675 + ], + "type": "text", + "content": " capsule pose matrices. We introduce and rationalise each component of our methodology as follows." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": "Framework overview. Our method adopts a Siamese joint embedding setup with identical, weight-sharing branches, each formed by a CNN " + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 677, + 555, + 714 + ], + "type": "text", + "content": " and a CapsNet projec-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7949" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": "tor " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ". We omit global average pooling to preserve the 2D spatial structure necessary for capsule processing. Given an image " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " from a dataset " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ", we sample two transformations " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "g_{1}, g_{2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " from a distribution over the group " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ", as detailed in supplementary Sec. 14. These are applied to " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " to produce two distinct views, " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "x_{g_1} = \\rho_X(g_1) \\cdot x" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "x_{g_2} = \\rho_X(g_2) \\cdot x" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ". Both views are passed through " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ", producing feature maps " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "F, F' \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ". These are then processed by " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ", which contains " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " capsules, yielding " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{act}}, z_{\\mathrm{pose}} = h_{\\phi}(F)" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{act}}', z_{\\mathrm{pose}}' = h_{\\phi}(F')" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{act}}, z_{\\mathrm{act}}' \\in \\mathbb{R}^N" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " are the activation probabilities, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{pose}}, z_{\\mathrm{pose}}' \\in \\mathbb{R}^{N \\times 4 \\times 4}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": " are the corresponding pose matrices. The activation vectors encode transformation-invariant properties, representing semantic concepts that remain unchanged under " + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "inline_equation", + "content": "g_{1}, g_{2}" + }, + { + "bbox": [ + 55, + 72, + 296, + 289 + ], + "type": "text", + "content": ". In contrast, the pose matrices capture transformation-equivariant properties, encoding an entity's pose, where each capsule represents a unique entity. A visual depiction of the general architecture is given in Fig. 2." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": "Self-routing CapsNet projector. To reduce the computational overhead of iterative-routing, we adopt the noniterative self-routing algorithm [29] as our projector. The projector " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " consists of two layers: a primary capsule layer that transforms encoder feature maps into poses " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "(u_i)" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " and activations " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "(a_i)" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": ", and a self-routing layer, where each capsule independently determines its routing coefficients via a subordinate single-layer perceptron. To do this, the routing coefficients " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "c_{ij}" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " between each lower-layer capsule " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " and each upper-level capsule " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " are computed by multiplying the lower capsule's pose vector " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " with a learnable weight matrix " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "W^{\\mathrm{route}}" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " and applying a softmax function. The activation " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "a_j" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": " of each upper-level capsule is then computed by forming votes from lower-level capsules—each vote is the product " + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "inline_equation", + "content": "c_{ij}a_i" + }, + { + "bbox": [ + 55, + 289, + 296, + 479 + ], + "type": "text", + "content": ", which are then summed and normalised by the total lower-level activation, to formulate weighted votes:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 483, + 294, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 483, + 294, + 512 + ], + "spans": [ + { + "bbox": [ + 61, + 483, + 294, + 512 + ], + "type": "interline_equation", + "content": "c _ {i j} = \\operatorname {s o f t m a x} \\left(W _ {i} ^ {\\text {r o u t e}} u _ {i}\\right) _ {j}, \\quad a _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}{\\sum_ {i \\in \\Omega_ {l}} a _ {i}}. \\tag {3}", + "image_path": "2b8a89142980392987226803a79999e8f916649db77234408c52535ff2c3e0b3.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": "The capsule layer's output pose is computed by multiplying a trainable weight matrix " + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "inline_equation", + "content": "W^{\\text{pose}}" + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": " with each lower-level pose " + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": ", yielding a predicted pose " + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\hat{u}_{j|i}" + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": " for each upper-level capsule. The final pose " + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "inline_equation", + "content": "u_j" + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": " is the weighted average of " + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\hat{u}_{j|i}" + }, + { + "bbox": [ + 55, + 518, + 296, + 567 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 82, + 570, + 294, + 599 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 570, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 82, + 570, + 294, + 599 + ], + "type": "interline_equation", + "content": "\\hat {u} _ {j \\mid i} = W _ {i j} ^ {\\text {p o s e}} u _ {i}, \\quad u _ {j} = \\frac {\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i} \\hat {u} _ {j \\mid i}}{\\sum_ {i \\in \\Omega_ {l}} c _ {i j} a _ {i}}. \\tag {4}", + "image_path": "6fa53c53de8f5b8a57c2a31af81d0cfd0bf5d8657f78237770d7fb45b089fb33.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "Enforcing equivariance. We define the relative transformation " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "g_{\\mathrm{rel}} = g_2g_1^{-1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", mapping between the views " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{g_1}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{g_2}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", and let " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\rho_Y(g_{\\mathrm{rel}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " denote its matrix representation. To enforce equivariance (defined in Eq. (1)) we multiply " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{pose}}" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\rho_Y(g_{\\mathrm{rel}})" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", aligning the transformed result with " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{pose}}'" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ". This indicates that the pose embeddings are equivariant w.r.t. the transformation group " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", ensuring that geometric changes in the input are consistently captured in the latent space. Each pose matrix " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "(z_{\\mathrm{pose}}\\cdot \\rho_Y(g_{\\mathrm{rel}}))" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{pose}}'" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 107 + ], + "type": "text", + "content": "is normalised by its Frobenius norm and residing on a unit hypersphere. This direct manipulation in the latent space removes the need for a dedicated predictor module." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": "Loss functions. To enforce invariance we minimise the cross-entropy " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "H(Z_{\\mathrm{act}}, Z_{\\mathrm{act}}^{\\prime})" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " between the activation probability vectors " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{act}}" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{act}}^{\\prime}" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " denotes the matrix embeddings computed over a batch. This aligns the capsule activations for different views of the same image. To ensure full capsule utilisation, we apply mean entropy maximisation (ME-MAX) regularisation [1, 2, 36] to the activation probabilities. Let " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "\\bar{Z}_{\\mathrm{act}} = \\frac{1}{B}\\sum_{i=1}^{B}Z_{\\mathrm{act}}" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " is the batch size. Maximising " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "H(Z_{\\mathrm{act}})" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "H(\\bar{Z}_{\\mathrm{act}}^{\\prime})" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " encourages a diverse distribution of activation across all capsules, preventing the collapse to only a few highly active capsules. To enforce equivariance we minimise the mean-squared error between the normalised, transformed pose " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "(Z_{\\mathrm{pose}}\\cdot\\rho_{Y}(g_{\\mathrm{rel}}))" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": " and the corresponding pose " + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{pose}}^{\\prime}" + }, + { + "bbox": [ + 313, + 108, + 554, + 276 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 276, + 554, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 554, + 347 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 554, + 347 + ], + "type": "text", + "content": "To prevent collapse, we apply the variance-covariance regularisation criterion [3] to the concatenated capsule activation vectors and pose matrices, collectively denoted as " + }, + { + "bbox": [ + 313, + 276, + 554, + 347 + ], + "type": "inline_equation", + "content": "Z_{\\mathrm{cat}}" + }, + { + "bbox": [ + 313, + 276, + 554, + 347 + ], + "type": "text", + "content": ". Regularising these combined embeddings encourages diversity and non-redundancy across both activation and pose dimensions. The regularisation loss is defined as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 358, + 555, + 373 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 358, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 348, + 358, + 555, + 373 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) = \\lambda_ {V} V \\left(Z _ {\\text {c a t}}\\right) + \\lambda_ {C} C \\left(Z _ {\\text {c a t}}\\right), \\tag {5}", + "image_path": "4a2bdca1327e2bb1765d2ff4b389252fba40a3798514bfe2c793b3a456d02ea8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 330, + 376, + 554, + 410 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 376, + 554, + 410 + ], + "spans": [ + { + "bbox": [ + 330, + 376, + 554, + 410 + ], + "type": "interline_equation", + "content": "V \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {j = 1} ^ {d} \\max \\left(0, 1 - \\sqrt {\\operatorname {V a r} \\left(Z _ {\\text {c a t} , j}\\right)}\\right), \\tag {6}", + "image_path": "ad9c96dc5d039c406f90bf761c1054b0ad47dd79467317b000173eedaad1c607.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 369, + 415, + 553, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 415, + 553, + 444 + ], + "spans": [ + { + "bbox": [ + 369, + 415, + 553, + 444 + ], + "type": "interline_equation", + "content": "C \\left(Z _ {\\text {c a t}}\\right) = \\frac {1}{d} \\sum_ {i \\neq j} \\operatorname {C o v} \\left(Z _ {\\text {c a t}}\\right) _ {i, j} ^ {2}, \\tag {7}", + "image_path": "da5e55b4c7a86603a7fb83ab89480e1aa892f1a0bea75decba16dc6cf805deb1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "inline_equation", + "content": "V(Z_{\\mathrm{cat}})" + }, + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "text", + "content": " ensures that each dimension is utilised and " + }, + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "inline_equation", + "content": "C(Z_{\\mathrm{cat}})" + }, + { + "bbox": [ + 313, + 449, + 554, + 483 + ], + "type": "text", + "content": " decorrelates different dimensions to reduce redundant information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 554, + 555 + ], + "type": "text", + "content": "While SIE [24] and CapsIE [23] apply this regularisation before incorporating transformation parameters, we compute it directly on the transformation-aligned embedding. This approach yields improved performance in our experiments, as it operates on the final, fully aligned representation. Our overall loss function is defined as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 562, + 421, + 576 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 562, + 421, + 576 + ], + "spans": [ + { + "bbox": [ + 315, + 562, + 421, + 576 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\lambda_ {\\mathrm {i n v}} H \\left(Z _ {\\mathrm {a c t}}, Z _ {\\mathrm {a c t}} ^ {\\prime}\\right) +", + "image_path": "0b51beaf2f45dddd42368e7f38aa2ed935772dc5204adf6f669bf4635b060bf5.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 332, + 577, + 423, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 577, + 423, + 590 + ], + "spans": [ + { + "bbox": [ + 332, + 577, + 423, + 590 + ], + "type": "interline_equation", + "content": "H \\left(\\bar {Z} _ {\\mathrm {a c t}}\\right) + H \\left(\\bar {Z} _ {\\mathrm {a c t}} ^ {\\prime}\\right) +", + "image_path": "5d84abbb45ab1b84d8a160eab77873c9dc081ecde47be92579c23d20371cb53a.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 332, + 593, + 561, + 628 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 593, + 561, + 628 + ], + "spans": [ + { + "bbox": [ + 332, + 593, + 561, + 628 + ], + "type": "interline_equation", + "content": "\\lambda_ {\\mathrm {e q u i}} \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\left\\| \\frac {Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i})}{\\| Z _ {i , \\mathrm {p o s e}} \\cdot \\rho_ {Y} (g _ {\\mathrm {r e l} , i}) \\| _ {F}} - \\frac {Z _ {i , \\mathrm {p o s e}} ^ {\\prime}}{\\| Z _ {i , \\mathrm {p o s e}} ^ {\\prime} \\| _ {F}} \\right\\| _ {2} ^ {2} +", + "image_path": "7d732fc6164238b66ce1cb912198342f56ac31c0c67b352e3578e53c928eb260.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 332, + 630, + 553, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 630, + 553, + 653 + ], + "spans": [ + { + "bbox": [ + 332, + 630, + 553, + 653 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}}\\right) + \\mathcal {L} _ {\\text {r e g}} \\left(Z _ {\\text {c a t}} ^ {\\prime}\\right), \\tag {8}", + "image_path": "a4871a028b0501a940080663c0b45a1403eb216b3b9ef0c410ee090ec52e6727.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\|_F" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": " denotes the Frobenius norm and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{inv}} = 0.1, \\lambda_V = 10, \\lambda_C = 1" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{equi}} = 5" + }, + { + "bbox": [ + 313, + 654, + 554, + 713 + ], + "type": "text", + "content": ". Despite the loss calculation occurs on the embeddings, the encoder's globally pooled output is used for downstream tasks as it yields better results [10]. Ablations, including the convergence speed" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7950" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "and the impact of the number of capsules and invariant loss components, are given in supplementary Sec. 10." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": "Removal of predictor module. Unlike previous equivariant methods such as SEN [46], EquiMod [19], SIE [24], and CapsIE [23] that use a dedicated predictor network to map one embedding to another, our approach eliminates this module. In those earlier works, the predictor " + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": " is a network that enforces the equivariance condition (see Eq. (1)). Its input is " + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "inline_equation", + "content": "g_{rel}" + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": " and the projector's output " + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "inline_equation", + "content": "z_{i}" + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": ". It predicts the embedding of the transformed input, " + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "inline_equation", + "content": "z_{i}' = q(z_{i}, g_{rel})" + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 55, + 96, + 296, + 264 + ], + "type": "text", + "content": " is a hypernetwork or linear layer. This removal simplifies the architecture and instead enforces the explicit mapping of transformations in the embedding space via affine transformation matrices applied to the pose embeddings rather than manipulating embeddings through an intermediate parameter-prediction stage." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 273, + 296, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 296, + 301 + ], + "type": "text", + "content": "5. 3DIEBench-T: a translation-enhanced benchmark for invariant-equivariant SSL" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 306, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 306, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 54, + 306, + 295, + 486 + ], + "type": "text", + "content": "As highlighted by Garrido et al. [24], commonly used datasets for evaluating invariant or equivariant properties face limitations. Evaluations of invariant representations commonly benchmark on datasets like ImageNet [17] or CIFAR [39], where objects lack parameterisable transformations because augmentations occur only in the pixel domain. However, certain equivariant datasets [38, 46] lack visual complexity and transformational diversity to allow for fine-grained control. A more comprehensive alternative is 3DIEBench [24], which controls the scene and object parameters for 3D rotation prediction (equivariance) while retaining sufficient complexity for classification (invariance). Nonetheless, as with other similar datasets [54], 3DIEBench focuses only on rotation, evaluating methods under a single geometric transformation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 487, + 295, + 605 + ], + "type": "text", + "content": "Focussing only on rotation raises questions regarding the generalisability of methods [23, 24] to other geometric transformations, and their applicability in more realistic settings. To address this, we extend 3DIEBench by introducing additional translational transformations. Our new dataset, 3DIEBench-T (3D Invariant Equivariant Benchmark-Translated), goes beyond rotation-only benchmarks [24, 54], enabling translation prediction and providing a more challenging, realistic setting to assess how methods handle simultaneous multi-geometric transformations." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "To investigate the impact of the added transformations, we adhere to the original 3DIEBench data generation protocol, ensuring any observed performance differences are attributed to the inclusion of translations rather than broader dataset modifications. We use 52,472 3D object instances spanning 55 classes from ShapeNetCoreV2 [9], originally sourced from 3D Warehouse [35]. For each instance, we generate 50 uniformly sampled views within specified ranges via Blender-Proc [18], yielding 2,623,600 images." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 316, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 316, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 316, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 316, + 70, + 553, + 190 + ], + "type": "image", + "image_path": "43fc3ad6b664518ae8f3d9549addce56daa0e255e89acb03b06be8656b285fa6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 553, + 220 + ], + "type": "text", + "content": "Figure 3. Sample object instances from 3DIEBench-T dataset under different rotation, translation, and colour hue transformations." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 243, + 553, + 268 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 268 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 268 + ], + "type": "text", + "content": "Samples appear in Fig. 3, with further generation information and visualisations given in supplementary Sec. 14." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 280, + 436, + 293 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 280, + 436, + 293 + ], + "spans": [ + { + "bbox": [ + 313, + 280, + 436, + 293 + ], + "type": "text", + "content": "6. Experimental results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 554, + 373 + ], + "type": "text", + "content": "Setup. To ensure fair comparisons, we adopt the experimental setup from [24] across all methods, which aligns with standard SSL practices: a pre-training phase followed by various task-specific downstream evaluations. A detailed description of the experimental setup is given in supplementary Sec. 9." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 554, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 554, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 554, + 505 + ], + "type": "text", + "content": "Baselines. As an upper bound, we include a supervised baseline using ResNet-18 [30]. For invariant self-supervision, we benchmark against VICReg [3], a noncontrastive method, and SimCLR [10], a contrastive approach. We also consider SimCLR + AugSelf [40], a parameter-prediction extension of SimCLR that learns the augmentation parameters but does not enforce explicit equivariance. For equivariant self-supervision, we compare against SEN [46], EquiMod [19], SIE [24], and CapsIE [23]. Following [24, 61], we replace SEN's triplet loss with the SimCLR loss to avoid hyperparameter tuning." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "content": "Evaluation protocol. For the invariant task, we perform classification via the linear evaluation protocol, where a linear classification layer is trained on the frozen representations. To assess equivariant properties, we perform rotation and translation prediction tasks. For each task, a three-layer MLP is trained on the frozen representations. Its inputs are concatenated pairs of representations from two views of the same object, and is trained to regress the transformation parameters—rotation for the rotation task and translation for the translation task. To explore invariance to non-geometric transformations that are not optimised in pre-training, a linear layer is trained on the frozen representations to regress the floor and spot hue. The input representations are identical to those used in the geometric tasks. To assess the performance of predicting these parameters, the coefficient of determination " + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "inline_equation", + "content": "R^2 = 1 - \\frac{\\sum_i(y_i - \\hat{y}_i)^2}{\\sum_i(y_i - \\bar{y})^2}" + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "content": " is used, where " + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "content": " are the target values, " + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{y}" + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "content": " is their mean, and " + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{y}_i" + }, + { + "bbox": [ + 313, + 506, + 555, + 713 + ], + "type": "text", + "content": " are the predicted" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7951" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 123, + 504, + 266 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Evaluation on 3DIEBench and 3DIEBench-T using a ResNet-18 backbone. Learned representations are evaluated on an invariant task (classification) and equivariant tasks (rotation, translation, and colour prediction). For each group of methods, the best value is shown in bold. Supervised baselines are trained for each task. Equivariant methods are trained to be rotation-equivariant, with no explicit constraints for translation or colour prediction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 123, + 504, + 266 + ], + "lines": [ + { + "bbox": [ + 106, + 123, + 504, + 266 + ], + "spans": [ + { + "bbox": [ + 106, + 123, + 504, + 266 + ], + "type": "table", + "html": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
3DIEBench3DIEBench-T3DIEBench3DIEBench-T3DIEBench-T3DIEBench3DIEBench-T
Supervised Methods
ResNet-1886.4580.130.770.730.670.990.99
Invariant and Parameter Prediction Methods
VICReg84.2874.710.450.390.220.100.50
SimCLR86.7380.080.520.440.250.290.50
SimCLR + AugSelf87.4480.860.750.690.500.280.51
Equivariant Methods
SEN86.9980.200.510.450.260.290.47
EquiMod87.3980.760.500.430.240.290.38
SIE82.9475.560.730.450.200.070.46
CapsIE79.1475.200.740.600.460.010.03
EquiCaps83.2476.910.780.730.600.090.05
", + "image_path": "243698cce5538542a3240e5495d1c1e72445349c229657ecf7b78ff6fe3439bc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "text", + "content": "values. For geometric tasks, a larger " + }, + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "text", + "content": " denotes a better fit for the predictions, whereas for the colour task, an " + }, + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "inline_equation", + "content": "R^2" + }, + { + "bbox": [ + 55, + 285, + 295, + 334 + ], + "type": "text", + "content": " close to 0 is better, as no colour-specific pre-training constraint is enforced. Further details are given in supplementary Sec. 9." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 203, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 203, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 203, + 357 + ], + "type": "text", + "content": "6.1. Rotation-only pre-training" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 481 + ], + "type": "text", + "content": "Tab. 1 summarises the downstream performance of the learned representations on both 3DIEBench and 3DIEBench-T. Specifically, 3DIEBench evaluates classification, rotation, and colour (hue) tasks, while 3DIEBench-T adds translations. Notably, all equivariant methods in Tab. 1 are optimised solely for rotation equivariance during pretraining. As observed in prior works, invariant methods fail to learn successful equivariant representations of the rotations and translations, while the addition of predictive tasks (AugSelf) improves their capability to quantify this change." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "text", + "content": "We find that EquiCaps consistently outperforms SIE—its closest competitor—marginally on classification but significantly on geometric equivariant tasks, achieving " + }, + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "inline_equation", + "content": "+0.28 R^2" + }, + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "text", + "content": " in rotation and " + }, + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "inline_equation", + "content": "+0.40 R^2" + }, + { + "bbox": [ + 55, + 482, + 295, + 661 + ], + "type": "text", + "content": " in translation on 3DIEBench-T. Notably, our capsule-based projector is able to intrinsically capture translational attributes, despite no direct supervision. CapsIE also generalises relatively better than other equivariant methods, but it underperforms our method, and validates our claim that its predictor-based design underuses the inherent pose-awareness capabilities of CapsNets. Nevertheless, we find that non-capsule architectures do not generalise well under more complex, combined 3D transformations. We attribute this limitation to their lack of an explicit, pose-aware mechanism. Results on the Objaverse [16] and MOVi-E [26] datasets are given in supplementary Sec. 11." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 234, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 234, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 234, + 685 + ], + "type": "text", + "content": "6.2. Rotation-translation pre-training" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "In Tab. 2, we present results when both rotation and translation parameters are optimised. As expected, all meth" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 285, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 555, + 382 + ], + "type": "text", + "content": "ods exhibit a small improvement in translation performance. However, the two capsule-based methods retain their substantial improvement on geometric tasks. Consistent with [55], we find that SIE's gains on 3DIEBench do not generalise well to multi-geometric learning. We further find that SIE exhibits minimal gains over purely invariant methods and maintains the weakest translation prediction, even when explicitly trained for translation equivariance." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 315, + 467, + 555, + 529 + ], + "blocks": [ + { + "bbox": [ + 313, + 403, + 555, + 456 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 456 + ], + "type": "text", + "content": "Table 2. Evaluation on 3DIEBench-T pre-trained for both rotation and translation. Learned representations (the encoder's outputs) are evaluated on an invariant and equivariant tasks: classification, rotation, translation, and colour prediction. We indicate the change relative to Tab. 1." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 467, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 315, + 467, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 315, + 467, + 555, + 529 + ], + "type": "table", + "html": "
MethodClassification (Top-1)Rotation (R2)Translation (R2)Colour (R2)
SimCLR + AugSelf81.04 ↑ 0.180.69 = 0.000.64 ↑ 0.140.51 = 0.00
SEN80.23 ↑ 0.030.46 ↑ 0.010.28 ↑ 0.020.50 ↑ 0.03
EquiMod80.89 ↑ 0.130.46 ↑ 0.030.37 ↑ 0.130.37 ↓ 0.01
SIE75.91 ↑ 0.350.48 ↑ 0.030.22 ↑ 0.020.36 ↓ 0.10
CapsIE76.31 ↑ 1.110.62 ↑ 0.020.53 ↑ 0.070.03 = 0.00
EquiCaps77.88 ↑ 0.970.71 ↓ 0.020.61 ↑ 0.010.02 ↓ 0.03
", + "image_path": "e018393a457d3c7d9ed37a319a7d8d35b4bc360045974324061479c7368c325d.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 596, + 555, + 711 + ], + "blocks": [ + { + "bbox": [ + 313, + 543, + 554, + 588 + ], + "lines": [ + { + "bbox": [ + 313, + 543, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 543, + 554, + 588 + ], + "type": "text", + "content": "Table 3. Evaluation on 3DIEBench and 3DIEBench-T pre-trained for both rotation and translation. Learned embeddings (the projector's outputs) are evaluated on equivariant tasks: rotation and translation prediction." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 596, + 555, + 711 + ], + "lines": [ + { + "bbox": [ + 315, + 596, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 315, + 596, + 555, + 711 + ], + "type": "table", + "html": "
MethodRotation (R2)Translation (R2)
3DIEBench3DIEBench-T3DIEBench-T
VICReg0.240.240.00
SimCLR0.240.240.00
SimCLR + AugSelf0.240.240.00
SEN0.240.240.00
EquiMod0.240.240.22
SIE0.600.340.00
CapsIE0.710.580.26
EquiCaps0.730.610.33
", + "image_path": "e1098ed328c1803ff6589f7f8c5d6cb9aeae6c7f3377caddee0658cdceda601b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7952" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 106, + 113, + 504, + 276 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 4. Equivariance evaluation on 3DIEBench (top) and 3DIEBench-T (bottom), using MRR, H@k, and separate PRE measures for rotation and translation on 3DIEBench-T. We indicate which subset (train or val) is used to generate embeddings, and for PRE, which dataset (train, val, or all) is used for retrieval. For 3DIEBench-T, methods are pre-trained for both rotation and translation equivariance." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 113, + 504, + 276 + ], + "lines": [ + { + "bbox": [ + 106, + 113, + 504, + 276 + ], + "spans": [ + { + "bbox": [ + 106, + 113, + 504, + 276 + ], + "type": "table", + "html": "
MethodMRR (↑)H@1 (↑)H@5 (↑)PRE (↓)
trainvaltrainvaltrainvaltrain-trainval-valval-all
AllAllAllAllAllAllRotationTranlationRotationTranlationRotationTranlation
3DIEBench
SEN0.090.090.020.020.100.100.61-0.61-0.61-
EquiMod0.090.090.020.020.100.100.61-0.61-0.61-
SIE0.510.410.410.300.600.510.26-0.29-0.27-
CapsIE0.600.470.500.360.710.580.17-0.21-0.20-
EquiCaps0.680.600.550.480.840.730.13-0.18-0.17-
3DIEBench-T
SEN0.090.090.020.020.100.100.612.710.622.720.622.72
EquiMod0.130.120.030.030.180.150.582.470.592.510.592.50
SIE0.150.140.060.050.200.180.542.730.542.720.542.73
CapsIE0.460.370.350.260.550.470.292.450.312.480.302.46
EquiCaps0.590.510.480.410.700.620.302.340.352.410.332.38
", + "image_path": "22ba6832f5844be1c75bb1b18d4b4a23ea5e9d5e6da9230fb30853682c266161.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 297, + 296, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 297, + 296, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 297, + 296, + 392 + ], + "type": "text", + "content": "Colour prediction. EquiCaps maintains near-perfect invariance to hue transformations, consistent with prior work such as SIE and CapsIE, although SIE is less effective under 3DIEBench-T. By contrast, non-capsule methods exhibit substantially increased colour prediction when transitioning to 3DIEBench-T. We conjecture that the dataset's increased geometric complexity makes the colour task easier to solve for models lacking explicit pose-aware mechanisms." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 212, + 411 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 212, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 212, + 411 + ], + "type": "text", + "content": "6.3. Performance on embeddings" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "text", + "content": "As in prior works [10, 24], we observe that representations (the outputs of the CNN encoder " + }, + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "text", + "content": ") learned by invariant methods achieve higher-than-expected performance on equivariant tasks. This arises because the invariance objective is enforced on the embeddings (the outputs of the projector " + }, + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "inline_equation", + "content": "h_{\\phi}" + }, + { + "bbox": [ + 54, + 415, + 295, + 654 + ], + "type": "text", + "content": ") after the projector, while the representations can still retain information [10]. To explore this further, we evaluate rotation and translation predictions on the embeddings (see Tab. 3), where all methods show decreased performance compared to their representation-level results, highlighting that the projector absorbs more invariance. We also observe that invariant methods, along with SEN and EquiMod, converge to similar embedding-level results, except for EquiMod, which retains higher translation prediction performance, as in the representation-stage results. SIE aligns with invariant methods on translation prediction, consistent with its representation stage. EquiCaps maintains the best performance on both tasks, and capsule-based methods show the least decline from representations to embeddings, underscoring the advantages of capsules' pose-awareness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 192, + 673 + ], + "type": "text", + "content": "6.4. Equivariance evaluation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We employ and adapt the Mean Reciprocal Rank (MRR), Hit Rate at k (H@k), and Prediction Retrieval Error (PRE) metrics from [24]. For compared methods, for each object's" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": "source and target poses, we feed the source embeddings and the corresponding transformation parameters through the predictor, obtaining a predicted embedding. To adapt it to our method, we multiply the projector's pose matrix by the transformation matrix. We obtain the nearest neighbours of the predicted embedding among the views of the same object. MRR is computed as the mean of the reciprocal rank " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "\\frac{1}{r}" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " is the rank of the target embedding among the neighbours. H@k yields 1 if the target is within the top-" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " neighbours, and 0 otherwise. PRE measures the mean distance between the transformation parameters of the nearest neighbour and the target across the dataset. Rotation uses " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "d_r = 1 - \\langle q_1, q_2 \\rangle^2" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "q_1" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "q_2" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " are the nearest neighbour's and target's rotation represented as a quaternion, respectively. Translation uses the mean-squared error " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "d_t = ||t_1 - t_2||_2^2" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " is the nearest neighbour's translation vector and " + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "inline_equation", + "content": "t_2" + }, + { + "bbox": [ + 313, + 297, + 555, + 500 + ], + "type": "text", + "content": " is the target's translation vector." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 502, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 556, + 658 + ], + "type": "text", + "content": "As shown in Tab. 4, EquiCaps outperforms all methods across each data split of 3DIEBench, achieving an MRR of 0.60 on the validation set, well above SIE (0.41) and CapsIE (0.47). Similar results are observed in H@1/5 metrics, where SEN and EquiMod perform randomly (2% and 10%) on both data splits, while CapsIE's PRE on all data splits approaches ours. Although EquiCaps generally outperforms all methods on 3DIEBench-T across most metrics, CapsIE achieves slightly lower rotation PRE. Despite this, our method remains significantly stronger in MRR and H@1/5 across all data splits, with EquiMod slightly improving. In contrast, SIE's overall performance drops with H@1 of 0.05 on the validation set, marginally above random." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 670, + 423, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 423, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 423, + 682 + ], + "type": "text", + "content": "6.5. Qualitative results" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Nearest-neighbour representations. We perform a nearest-neighbour representation search on the 3DIEBench-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7953" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 531, + 218 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 531, + 218 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 531, + 218 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 531, + 218 + ], + "type": "image", + "image_path": "2ea6445c95794947c54b78747670a08ea7c63212438a01c07a0271e24b7bbe45.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "lines": [ + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 54, + 224, + 555, + 258 + ], + "type": "text", + "content": "Figure 4. Nearest-neighbour representation retrieval on 3DIEBench-T validation set directly after pre-training. The query image (left) is compared against each method's learned representations to find its top three nearest neighbours (in rows: 1-NN, 2-NN, 3-NN). Equivariant and parameter prediction methods are trained to be both rotationally and translationally equivariant." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 278, + 296, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 278, + 296, + 530 + ], + "spans": [ + { + "bbox": [ + 56, + 278, + 296, + 530 + ], + "type": "text", + "content": "T validation set (Fig. 4) to illustrate the information captured by the representations. Invariant approaches should retrieve objects in the same class but varied poses, whereas equivariant methods should preserve the query pose. We find that only our method and CapsIE retrieve the nearest neighbour with a matching pose, though ours aligns also with the translation. For subsequent neighbours, only EquiCaps and AugSelf retrieve objects in similar poses, while CapsIE does so but less accurately. In contrast, the other methods, including equivariant methods, fail to yield pose-matching neighbours suggesting that invariance dominates and their representations leave out significant pose information. We replicate the retrieval on 3DIEBench Fig. A.6, supplementary Sec. 12, where ours, SIE, and CapsIE successfully return pose-matching neighbours, but remaining methods, only partially align poses, albeit less accurate. We conjecture this is attributed to the projector absorbing invariance, while partial rotation-related information remains at the representation stage. All retrieved objects belong to the car category, indicating that class semantics are preserved consistent with our quantitative results." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "Equivariance via pose manipulation. To further illustrate the equivariant properties of our approach, we examine the inversion property through the inverse rotation of an object in the embedding space. We rotate an object around the " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": "-axis within the range " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "[0, \\frac{\\pi}{2}]" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " increments and feed these images through our trained network to obtain its pose embedding. Embedding is subsequently multiplied by its corresponding inverse transformation matrix. To examine the impact of the inverse transformation we retrieve the nearest neighbour among all embeddings. As shown in Fig. 5, the nearest neighbour closely (within " + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "inline_equation", + "content": "5^{\\circ}" + }, + { + "bbox": [ + 55, + 534, + 296, + 714 + ], + "type": "text", + "content": " error) or exactly matches the source view, highlighting the invertibility of our representations, and the ability to intuitively and directly steer the latent space. Further qualitative results are provided in supplementary Sec. 12." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 315, + 276, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 315, + 276, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 315, + 276, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 315, + 276, + 555, + 384 + ], + "type": "image", + "image_path": "888dcae3d3a87acacc900a20db579cfbe82e8d99abddd932b8366290c0bd3f33.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "lines": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 555, + 415 + ], + "type": "text", + "content": "Figure 5. Illustration of transformation inversion of equivariant embeddings through capsule-based pose manipulation." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 438, + 388, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 388, + 451 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 388, + 451 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 312, + 460, + 555, + 604 + ], + "type": "text", + "content": "We propose EquiCaps, a capsule-based SSL method that jointly learns invariant and equivariant representations. By leveraging capsules' innate pose-awareness, EquiCaps shapes the latent space predictably, removing the need for a dedicated predictor. To enable richer benchmarking, we extend 3DIEBench to 3DIEBench-T, incorporating object translations alongside rotations. Empirically, EquiCaps achieves state-of-the-art rotation and translation prediction among equivariant methods. Tests on 3DIEBench-T highlight capsule architectures' superior generalisation under combined geometric transformations, demonstrating their potential in invariant-equivariant SSL." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "Limitations and future work. Joint invariance-equivariance optimisation involves trade-offs; while EquiCaps balances these, task-specific priorities may vary. The method assumes known group elements—as prior methods do—restricting its use without such data, though partial data suffices. EquiCaps can theoretically handle any transformation which can be expressed as a matrix without architectural changes. Future work includes other 3D transformations (e.g., scaling, shear), and more object detection tasks." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7954" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 170, + 85 + ], + "type": "text", + "content": "8. Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 248 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 248 + ], + "type": "text", + "content": "This work was supported by the UKRI AI Centre for Doctoral Training in Sustainable Understandable agri-food Systems Transformed by Artificial INtelligence (SUSTAIN) [grant reference: EP/Y03063X/1]. We also acknowledge the generous support of alumni and friends in establishing the University of Aberdeen's Interdisciplinary Institute, which partly enabled this research, including Dr Jane Hellman Caseley (MBChB 1956), Professor Patrick Meares (DSc 1959), Nancy Miller (MA 1942), Norman Robertson, Dr Ian Sessler (MBChB 1956) and Anne Young (MA 1957). Finally, we thank the University of Aberdeen's HPC Maxwell facility for providing the computational resources that facilitated our experimental work." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 261, + 115, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 115, + 273 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 115, + 273 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 280, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 280, + 295, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 295, + 336 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 295, + 336 + ], + "type": "text", + "content": "[1] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Armand Joulin, Nicolas Ballas, and Michael Rabbat. Semi-supervised learning of visual features by nonparametrically predicting view assignments with support samples. In ICCV, pages 8443-8452, 2021." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 337, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 337, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 61, + 337, + 295, + 392 + ], + "type": "text", + "content": "[2] Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, and Nicolas Ballas. Masked siamese networks for label-efficient learning. In ECCV, pages 456-473, 2022." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 393, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 393, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 393, + 295, + 426 + ], + "type": "text", + "content": "[3] Adrien Bardes, Jean Ponce, and Yann LeCun. Vicreg: Variance-invariance-covariance regularization for self-supervised learning. In ICLR, 2022." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 428, + 295, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 295, + 461 + ], + "type": "text", + "content": "[4] Adrien Bardes, Jean Ponce, and Yann LeCun. VICRegl: Self-supervised learning of local visual features. In NeurIPS, pages 8799-8810, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 62, + 463, + 296, + 506 + ], + "type": "text", + "content": "[5] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In ECCV, pages 213-229, 2020." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 509, + 295, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 509, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 509, + 295, + 542 + ], + "type": "text", + "content": "[6] Mathilde Caron, Piotr Bojanowski, Armand Joulin, and Matthijs Douze. Deep clustering for unsupervised learning of visual features. In ECCV, pages 132-149, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 543, + 295, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 543, + 295, + 586 + ], + "spans": [ + { + "bbox": [ + 62, + 543, + 295, + 586 + ], + "type": "text", + "content": "[7] Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. In NeurIPS, pages 9912-9924, 2020." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 589, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 589, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 589, + 295, + 633 + ], + "type": "text", + "content": "[8] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, pages 9650-9660, 2021." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 634, + 295, + 700 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 634, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 62, + 634, + 295, + 700 + ], + "type": "text", + "content": "[9] Angel X. Chang, Thomas A. Funkhouser, Leonidas J. Guibas, Pat Hanrahan, QiXing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, Jianxiong Xiao, Li Yi, and Fisher Yu. ShapeNet: An information-rich 3D model repository. arXiv preprint arXiv:1512.03012, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 702, + 295, + 714 + ], + "type": "text", + "content": "[10] Ting Chen, Simon Kornblith, Mohammad Norouzi, and Ge" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "offrey Hinton. A simple framework for contrastive learning of visual representations. In ICML, pages 1597-1607, 2020." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 118 + ], + "type": "text", + "content": "[11] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In CVPR, pages 15750-15758, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "spans": [ + { + "bbox": [ + 317, + 120, + 553, + 152 + ], + "type": "text", + "content": "[12] Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 155, + 553, + 187 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 155, + 553, + 187 + ], + "spans": [ + { + "bbox": [ + 316, + 155, + 553, + 187 + ], + "type": "text", + "content": "[13] Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In ICCV, pages 9640-9649, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 189, + 553, + 222 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 189, + 553, + 222 + ], + "spans": [ + { + "bbox": [ + 316, + 189, + 553, + 222 + ], + "type": "text", + "content": "[14] Rumen Dangovski, Li Jing, Charlotte Loh, Seungwook Han, Akash Srivastava, Brian Cheung, Pulkit Agrawal, and Marin Soljacic. Equivariant contrastive learning. In ICLR, 2022." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 224, + 553, + 267 + ], + "type": "text", + "content": "[15] Fabio De Sousa Ribeiro, Kevin Duarte, Miles Everett, Georgios Leontidis, and Mubarak Shah. Object-centric learning with capsule networks: A survey. ACM Comput. Surv., 56 (11), 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "type": "text", + "content": "[16] Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objverse: A universe of annotated 3D objects. In CVPR, pages 13142-13153, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 325, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 358 + ], + "type": "text", + "content": "[17] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248-255, 2009." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 360, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 553, + 415 + ], + "type": "text", + "content": "[18] Maximilian Denninger, Dominik Winkelbauer, Martin Sundermeyer, Wout Boerdijk, Markus Knauer, Klaus H. Strobl, Matthias Hunt, and Rudolph Triebel. BlenderProc2: A procedural pipeline for photorealistic rendering. Journal of Open Source Software, 8(82):4901, 2023." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 417, + 553, + 449 + ], + "type": "text", + "content": "[19] Alexandre Devillers and Mathieu Lefort. Equimod: An equivariance module to improve self-supervised learning. In ICLR, 2023." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 316, + 451, + 553, + 483 + ], + "type": "text", + "content": "[20] Kevin Duarte, Yogesh Rawat, and Mubarak Shah. Videocapsulenet: A simplified network for action detection. In NeurIPS, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 485, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 485, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 485, + 553, + 518 + ], + "type": "text", + "content": "[21] Kevin Duarte, Yogesh S. Rawat, and Mubarak Shah. Capsulevos: Semi-supervised video object segmentation using capsule routing. In ICCV, 2019." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 553 + ], + "type": "text", + "content": "[22] Aleksandr Ermolov, Aliaksandr Siarohin, Enver Sangineto, and Nicu Sebe. Whitening for self-supervised representation learning. In ICML, pages 3015-3024, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 587 + ], + "type": "text", + "content": "[23] Miles Everett, Aiden Durrant, Mingjun Zhong, and Georgios Leontidis. Capsule network projectors are equivariant and invariant learners. arXiv preprint arXiv:2405.14386, 2024." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 553, + 621 + ], + "type": "text", + "content": "[24] Quentin Garrido, Laurent Najman, and Yann LeCun. Self-supervised learning of split invariant equivariant representations. In ICML, 2023." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[25] Spyros Gidaris, Praveer Singh, and Nikos Komodakis. Unsupervised representation learning by predicting image rotations. In ICLR, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[26] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J. Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, Thomas Kipf, Abhijit Kundu, Dmitry Lagun, Issam Laradji, Hsueh-Ti (Derek) Liu, Henning Meyer, Yishu Miao, Derek" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7955" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 76, + 72, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 295, + 139 + ], + "type": "text", + "content": "Nowrouzezahrai, Cengiz Oztireli, Etienne Pot, Noha Radwan, Daniel Rebain, Sara Sabour, Mehdi S. M. Sajjadi, Matan Sela, Vincent Sitzmann, Austin Stone, Deqing Sun, Suhani Vora, Ziyu Wang, Tianhao Wu, Kwang Moo Yi, Fangcheng Zhong, and Andrea Tagliasacchi. Kubric: A scalable dataset generator. In CVPR, pages 3749-3761, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 215 + ], + "type": "text", + "content": "[27] Jean-Bastien Grill, Florian Strub, Florent Altché, Corentin Tallec, Pierre H. Richemond, Elena Buchatskaya, Carl Doersch, Bernardo Ávila Pires, Zhaohan Daniel Guo, Mohammad Gheshlaghi Azar, Bilal Piot, Koray Kavukcuoglu, Rémi Munos, and Michal Valko. Bootstrap your own latent: A new approach to self-supervised learning. In NeurIPS, pages 21271-21284, 2020." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[28] Sharut Gupta, Joshua Robinson, Derek Lim, Soledad Villar, and Stefanie Jegelka. Structuring representation geometry with rotationally equivariant contrastive learning. In ICLR, 2024." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 261, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 284 + ], + "type": "text", + "content": "[29] Taeyoung Hahn, Myeongjiang Pyeon, and Gunhee Kim. Selfrouting capsule networks. In NeurIPS, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "spans": [ + { + "bbox": [ + 56, + 285, + 294, + 318 + ], + "type": "text", + "content": "[30] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 318, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 318, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 318, + 294, + 352 + ], + "type": "text", + "content": "[31] Kaiming He, Haoqi Fan, Yuxin Wu, Saining Xie, and Ross Girshick. Momentum contrast for unsupervised visual representation learning. In CVPR, pages 9729-9738, 2020." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[32] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 408 + ], + "type": "text", + "content": "[33] Geoffrey E. Hinton, Alex Krizhevsky, and Sida Wang. Transforming auto-encoders. In ICANN, pages 44–51, 2011." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 409, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 430 + ], + "type": "text", + "content": "[34] Geoffrey E. Hinton, Sara Sabour, and Nicholas Frosst. Matrix capsules with EM routing. In ICLR, 2018." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 432, + 293, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 293, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 293, + 453 + ], + "type": "text", + "content": "[35] Trimble Inc. 3D warehouse. https://3dwarehouse.sketchup.com/. Accessed: 2025-03-01." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 487 + ], + "type": "text", + "content": "[36] Armand Joulin and Francis Bach. A convex relaxation for weakly supervised classifiers. In ICML, pages 1279-1286, 2012." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 488, + 294, + 510 + ], + "type": "text", + "content": "[37] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 533 + ], + "type": "text", + "content": "[38] Thomas Kipf, Elise van der Pol, and Max Welling. Contrastive learning of structured world models. In ICLR, 2020." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 534, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 567 + ], + "type": "text", + "content": "[39] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, University of Toronto, 2009." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 601 + ], + "type": "text", + "content": "[40] Hankook Lee, Kibok Lee, Kimin Lee, Honglak Lee, and Jinwoo Shin. Improving transferability of representations via augmentation-aware self-supervision. In NeurIPS, 2021." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 646 + ], + "type": "text", + "content": "[41] Liangzhi Li, Bowen Wang, Manisha Verma, Yuta Nakashima, Ryo Kawasaki, and Hajime Nagahara. Scouter: Slot attention-based classifier for explainable image recognition. In ICCV, pages 1046-1055, 2021." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[42] Yi Liu, De Cheng, Dingwen Zhang, Shoukun Xu, and Jungong Han. Capsule networks with residual pose routing. IEEE Transactions on Neural Networks and Learning Systems, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[43] Francesco Locatello, Dirk Weissenborn, Thomas Unterthiner, Aravindh Mahendran, Georg Heigold, Jakob" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "Uszkoreit, Alexey Dosovitskiy, and Thomas Kipf. Object-centric learning with slot attention. In NeurIPS, pages 11525-11538, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 106, + 553, + 161 + ], + "type": "text", + "content": "[44] Giovanni Luca Marchetti, Gustaf Tegnér, Anastasiia Varava, and Danica Kragic. Equivariant representation learning via class-pose decomposition. In Proceedings of The 26th International Conference on Artificial Intelligence and Statistics, pages 4745-4756, 2023." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 162, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 270 + ], + "type": "text", + "content": "[45] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research, 2024." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 304 + ], + "type": "text", + "content": "[46] Jung Yeon Park, Ondrej Biza, Linfeng Zhao, Jan Willem van de Meent, and Robin Walters. Learning symmetric embeddings for equivariant world models. In ICML, 2022." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 553, + 337 + ], + "type": "text", + "content": "[47] Fabio De Sousa Ribeiro, Georgios Leontidis, and Stefanos Kollias. Capsule routing via variational bayes. In AAAI, pages 3749-3756, 2020." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[48] Sara Sabour, Nicholas Frosst, and Geoffrey E. Hinton. Dynamic routing between capsules. In NeurIPS, pages 3856-3866, 2017." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 414 + ], + "type": "text", + "content": "[49] Sara Sabour, Andrea Tagliasacchi, Soroosh Yazdani, Geoffrey E. Hinton, and David J. Fleet. Unsupervised part representation by flow capsules. In ICML, pages 9213-9223, 2021." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 553, + 446 + ], + "type": "text", + "content": "[50] Franz Scherr, Qinghai Guo, and Timoleon Moraitis. Self-supervised learning through efference copies. In NeurIPS, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 448, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 448, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 316, + 448, + 553, + 481 + ], + "type": "text", + "content": "[51] Mehran Shakerinava, Arnab Kumar Mondal, and Siamak Ravanbakhsh. Structuring representations using group invariants. In NeurIPS, pages 34162-34174, 2022." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[52] Weiwei Sun, Andrea Tagliasacchi, Boyang Deng, Sara Sabour, Soroosh Yazdani, Geoffrey E. Hinton, and Kwang Moo Yi. Canonical capsules: Self-supervised capsules in canonical pose. In NeurIPS, pages 24993-25005, 2021." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 536, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 579 + ], + "type": "text", + "content": "[53] Yonglong Tian, Chen Sun, Ben Poole, Dilip Krishnan, Cordelia Schmid, and Phillip Isola. What makes for good views for contrastive learning? In NeurIPS, pages 6827-6839, 2020." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 613 + ], + "type": "text", + "content": "[54] Jiayun Wang, Yubei Chen, and Stella Yu. Pose-aware self-supervised learning with viewpoint trajectory regularization. In ECCV, 2024." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "text", + "content": "[55] Qin Wang, Kai Krajsek, and Hanno Scharr. Equivariant representation learning for augmentation-based self-supervised learning via image reconstruction. In NeurIPS 2024 Workshop: Self-Supervised Learning - Theory and Practice, 2024." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[56] Yifei Wang, Kaiwen Hu, Sharut Gupta, Ziyu Ye, Yisen Wang, and Stefanie Jegelka. Understanding the role of equivariance in self-supervised learning. In NeurIPS, pages 127483-127510, 2024." + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7956" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 285 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[57] Robin Winter, Marco Bertolini, Tuan Le, Frank Noe, and Djork-Arné Clevert. Unsupervised learning of group invariant and equivariant representations. In NeurIPS, 2022." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 138 + ], + "type": "text", + "content": "[58] Tete Xiao, Xiaolong Wang, Alexei A. Efros, and Trevor Darryll. What should not be contrastive in contrastive learning. In ICLR, 2021." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 183 + ], + "type": "text", + "content": "[59] Yuyang Xie, Jianhong Wen, Kin Wai Lau, Yasar Abbas Ur Rehman, and Jiajun Shen. What should be equivariant in self-supervised learning. In CVPR Workshop, pages 4111-4120, 2022." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[60] Chun-Hsiao Yeh, Cheng-Yao Hong, Yen-Chi Hsu, Tyng-Luh Liu, Yubei Chen, and Yann LeCun. Decoupled contrastive learning. In ECCV, pages 668-684, 2022." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 251 + ], + "type": "text", + "content": "[61] Jaemyung Yu, Jaehyun Choi, Dong-Jae Lee, HyeongGwon Hong, and Junmo Kim. Self-supervised transformation learning for equivariant representations. In NeurIPS, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 285 + ], + "type": "text", + "content": "[62] Jure Zbontar, Li Jing, Ishan Misra, Yann LeCun, and Stéphane Deny. Barlow twins: Self-supervised learning via redundancy reduction. In ICML, pages 12310-12320, 2021." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7957" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_content_list.json b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5ba62d6d1f7d8498d92b38f4fbf5d87d106a34e5 --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_content_list.json @@ -0,0 +1,1602 @@ +[ + { + "type": "text", + "text": "Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection", + "text_level": 1, + "bbox": [ + 173, + 130, + 825, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shizhen Zhao Jiahui Liu Xin Wen Haoru Tan Xiaojuan Qi† The University of Hong Kong", + "bbox": [ + 230, + 203, + 764, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 273, + 326, + 289 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, we systematically investigate representative vision foundation models for OOD detection. Our findings reveal that a pre-trained DINOv2 model, even without fine-tuning on in-domain (ID) data, naturally provides a highly discriminative feature space for OOD detection, achieving performance comparable to existing state-of-the-art methods without requiring complex designs. Beyond this, we explore how fine-tuning foundation models on in-domain (ID) data can enhance OOD detection. However, we observe that the performance of vision foundation models remains unsatisfactory in scenarios with a large semantic space. This is due to the increased complexity of decision boundaries as the number of categories grows, which complicates the optimization process. To mitigate this, we propose the Mixture of Feature Experts (MoFE) module, which partitions features into subspaces, effectively capturing complex data distributions and refining decision boundaries. Further, we introduce a Dynamic- $\\beta$ Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming baseline methods. The project will be available at shizhen-zhao.github.io/OOD_MoFE/.", + "bbox": [ + 89, + 306, + 485, + 744 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 773, + 220, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The task of out-of-distribution (OOD) detection [16, 26, 31, 49] aims to equip models with the capability to discern whether input images originate from unknown OOD classes or belong to in-domain (ID) classes. Mainstream OOD detection methods [9, 10, 25, 63] focus on learning features", + "bbox": [ + 89, + 800, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/51a4b5db034172c05cb8f0c79ae340ecf6d91f9b0131e1efcf1f0ed6c2273953.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 524, + 273, + 890, + 431 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f220fba940cfbbb11dfef269d9272769b13ae241b780d1320ef0c5a2502abf37.jpg", + "image_caption": [ + "Figure 1. Holistic comparison to previous philosophy. (a) Traditional methods use a generalized model to project inputs onto a complex distribution; (b): Our approach leverages multiple experts to break the complex distribution into smaller ones, which leads to compact ID distribution and simplified decision boundary." + ], + "image_footnote": [], + "bbox": [ + 524, + 444, + 890, + 585 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and classifiers [31, 52, 53, 68] from ID data and then develop a score metric [14, 15, 35, 54] to determine whether a sample belongs to ID or OOD classes. Despite significant advancements, the fundamental challenge in OOD detection is establishing a feature space with high discriminative capacity that can effectively distinguish OOD samples from ID samples. Recently, vision foundation models [21, 38, 46, 51] trained on large-scale datasets have demonstrated the ability to learn robust and generalizable features, benefiting numerous tasks [29, 65, 76, 78]. This raises the question: with such powerful models and feature representations, does OOD detection remain a problem?", + "bbox": [ + 511, + 700, + 908, + 882 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Although several studies [11, 40, 41, 70] have explored", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 109, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1751", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "the use of foundation models for OOD detection, most focus on improving the performance of vision-language models like CLIP [46], while other foundation models, such as DINOv2, remain largely unexamined. In this study, we systematically investigate the feature spaces of different representative pre-trained foundation models, including vision-language models (e.g., CLIP) and self-supervised models (e.g., DINOv2), in the context of OOD detection. Our results reveal that DINOv2 provides the most discriminative feature space, enabling effective OOD detection without any finetuning. Notably, using a simple KNN metric [54], DINOv2 achieves performance comparable to more complex methods, establishing a strong baseline for further research.", + "bbox": [ + 89, + 90, + 485, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While vision foundation models [38, 45, 46, 70] have achieved impressive performance in OOD detection, there is still room for improvement, particularly on in-domain data with large semantic spaces [57-61, 71] (e.g., $29.27\\%$ FPR95 on the ImageNet-1K OOD benchmark [54]). This prompts us to investigate whether foundation models can be further optimized by leveraging available ID data. However, as the number of semantic classes increases, the complexity of the decision boundaries required to distinguish between ID and OOD data grows as well [17, 28, 32]. This heightened complexity creates challenges when fine-tuning foundation models on limited ID data. Previous methods (e.g. MOS [17]) decouple the complex space into simpler subspaces from the perspective of loss, which eases the optimization process and simplifies the decision boundaries. In this study, we tackle the problem orthogonally from the model perspective by designing a new Mixture of Experts (MoE) architecture to more thoroughly disentangle complex ID distribution.", + "bbox": [ + 91, + 289, + 483, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this issue, rather than directly optimizing the whole feature space [3, 33, 34, 37] (Fig. 1(a)), we propose a Mixture-of-Feature-Expert (MoFE) module, which utilizes multiple experts, and each expert specializes in a specific subspace and optimizes it accordingly (Fig. 1(b)). MoFE operates by partitioning the original feature space into $K$ subspaces based on semantics and feature similarities within the ID dataset. Each subspace is assigned to a dedicated expert, and a router assigns samples to the appropriate expert based on these partitions. Different from previous studies [12, 50], we use the [CLS] token as the input to the router network, since it encapsulates the semantic feature of the whole image. During training, the router is supervised by the partition assignments to ensure accurate sample-to-expert mapping. Each expert focuses solely on optimizing features within its designated partition, which helps prevent interference between features from different partitions. The results show that our approach significantly surpasses the previous approaches by a large margin (see Tab. 1), revealing the importance of learning expert models for OOD tasks.", + "bbox": [ + 89, + 565, + 485, + 867 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Additionally, given that data augmentation has been shown to enhance generalization for OOD, we introduce", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "a novel Mixup data augmentation strategy to further improve feature learning, which is better suited for advanced vision foundation models. Our design is based on the observation that different categories exhibit varying levels of discriminativeness with features from vision foundation models. In the original feature space, some categories show high discriminativeness, while others do not. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup [64] can blur the decision boundary between ID and OOD, leading to degraded performance (Appendix Fig. 6). Thus, unlike existing Mixup strategies that treat all categories equally [56, 62, 72], our approach makes Mixup weight sampling category-dependent by adjusting the sampling distribution (i.e. beta distribution) dynamically, taking into account their discriminativeness.", + "bbox": [ + 511, + 90, + 908, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our major contributions can be summarized as follows:", + "bbox": [ + 511, + 316, + 906, + 330 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We design a novel MoFE module to tailor pre-trained vision foundation models for OOD detection. This approach reduces the difficulty of fitting complex data distributions from limited data and eases the optimization process.", + "- We explore the effectiveness of the raw feature spaces from various vision foundation models for OOD detection. Through analysis, we leverage DINOv2 with simple scoring metric to establish a strong baseline. Additionally, we designed a Dynamic- $\\beta$ Mixup that is better suited for advanced vision foundation models.", + "- Our extensive experimental results demonstrate the effectiveness of the proposed model, achieving significant improvements over several competitive baseline methods on standard benchmarks." + ], + "bbox": [ + 511, + 333, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 556, + 656, + 571 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Out-of-Distribution Detection The goal of OOD detection is to detect OOD images from the test dataset (containing both ID and OOD images). Designing the score function is the most popular method in OOD detection tasks. The scores are mainly derived from three sources: the probability [14, 15], the logits [15, 35], and the feature [26, 44]. Some studies [20, 49, 69] focus on leveraging contrastive learning to enhance the feature representation. Other studies show that synthesizing pseudo samples [9, 49, 55, 55] as OOD instances is also a promising approach to make the feature space more compact. The methods [17, 28, 32] are the most relevant to ours, which also break the semantic space into smaller ones. Different from these approaches, in our design, we propose a novel MoE module, with each expert exclusively concentrating on optimizing features within its specific partition. Our results demonstrate that our approach outperforms them by a substantial margin.", + "bbox": [ + 511, + 580, + 908, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "OOD Detection with Foundation Models There are some existing OOD detection methods [11, 40-42, 45, 70] leveraging foundation models. Maximum Concept Matching", + "bbox": [ + 511, + 854, + 910, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1752", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(MCM) [41] proposes a simple yet effective zero-shot OOD detection method by aligning visual features with textual concepts. Some other studies [45, 70] explore negative prompts to learn the diversity of negative features, enabling more accurate detection of OOD samples. Although these studies have made great progress by leveraging CLIP to enhance the performance in existing benchmarks, they only explore and fine-tune CLIP. In our studies, we explore different foundation models and explore a better fine-tuning paradigm.", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Mixture of Experts Mixture of Experts has been studied independently in both computer vision [36, 43, 47, 73] and natural language processing [12, 22, 27, 50]. These works are studied in the context of conditional computation, which is to increase the number of model parameters without a proportional increase in computational cost. Currently, some studies [5, 24] explore improving expert specialization and leveraging MoE to mitigate data conflict problems, where some data might interfere with each other. In our study, we introduce MoFE to the out-of-distribution task in the context of foundation models and build specialized OOD detectors for different feature subspaces.", + "bbox": [ + 89, + 266, + 483, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Pilot Study", + "text_level": 1, + "bbox": [ + 89, + 472, + 209, + 489 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first introduce preliminaries for the OOD detection task in Sec. 3.1. Then, we explore the impact of foundation models on OOD detection performance and analyze their strengths and weaknesses in Sec. 3.2.", + "bbox": [ + 89, + 502, + 483, + 563 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 584, + 230, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We consider supervised multi-class classification, where $\\mathcal{X}$ represents the input image space and $\\mathcal{Y} = \\{1,2,\\dots,C\\}$ represents the label space. The training dataset $\\mathbb{D}_{in} = \\{(\\mathbf{x}_i,y_i)\\}_{i=1}^n$ is drawn independently and identically distributed (i.i.d.) from the joint data distribution $P_{\\mathcal{X}\\mathcal{Y}}$ . Let $\\mathcal{P}_{\\mathrm{in}}$ denote the marginal distribution on $\\mathcal{X}$ . Let $f:\\mathcal{X}\\mapsto \\mathbb{R}^{|\\mathcal{Y}|}$ be a neural network trained on samples drawn from $P_{\\mathcal{X}\\mathcal{Y}}$ to output a logit vector, which is used to predict the label of the input sample.", + "bbox": [ + 89, + 609, + 483, + 744 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Out-of-distribution Detection. When deploying a machine learning model in real-world scenarios, it is crucial for a reliable classifier not only to accurately classify known distribution (ID) samples, but also to recognize any out-of-distribution (OOD) inputs as \"unknown\". This can be accomplished by incorporating an OOD detector alongside the classification model $f$ . OOD can be formulated as a binary classification task. During testing, the objective is to determine whether a sample $\\mathbf{x} \\in \\mathcal{X}$ belongs to $\\mathcal{P}_{\\mathrm{in}}(\\mathrm{ID})$ or not (OOD). This decision can be made using a scoring", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "metric $S(\\mathbf{x})$", + "bbox": [ + 513, + 90, + 602, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG _ {\\lambda} (x) = \\left\\{ \\begin{array}{l l} \\mathrm {I D} & S (\\mathbf {x}) \\geq \\lambda \\\\ \\text {O O D} & S (\\mathbf {x}) < \\lambda \\end{array} , \\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 116, + 906, + 157 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where samples with higher scores $S(\\mathbf{x})$ are classified as ID and vice versa, and $\\lambda$ is the threshold. Some typically used metrics $S(\\mathbf{x})$ include MSP [14], MaxLogit [15], Energy [35] and KNN [54].", + "bbox": [ + 511, + 167, + 906, + 228 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a3be85820c442e52626f8d7e1a7769f131ca9bbdc9f0f2f42310cb5ad4971889.jpg", + "image_caption": [ + "Figure 2. Performance of vision foundation models across different OOD splits. The evaluation metric is FPR95, with lower values indicating better performance." + ], + "image_footnote": [], + "bbox": [ + 532, + 241, + 883, + 424 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evaluation of Vision Foundation Models", + "text_level": 1, + "bbox": [ + 513, + 486, + 859, + 501 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although several studies [11, 40, 41, 70] have explored the use of foundation models for OOD detection, they focus solely on vision-language foundation models such as CLIP [46]. Beyond CLIP, the community offers a variety of vision foundation models that provide robust raw feature space. This development has inspired us to re-examine which vision foundation model is best suited for OOD detection. In this section, we aim to investigate and analyze various pre-trained vision foundation models as effective OOD detectors without fine-tuning.", + "bbox": [ + 511, + 508, + 906, + 657 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Experimental Setup. We perform our evaluation on a challenging OOD detection benchmark that utilizes ImageNet-1K as ID data and selects samples from iNaturalist18, Sun, Places, and Textures as OOD samples. We also include two challenging OOD test sets: ImageNet-O [7] and NINCO [19]. We choose several representative vision foundation models, namely DINOv1 [2], DINOv2 [38], MAE [13], SigCLIP [75], OpenCLIP [18], and OpenAI CLIP [46]. For fair comparison, we use the ViT-B as the architecture of these models. The scoring functions for DINOv1, DINOv2, and MAE are set to KNN [54]. For the CLIP series, we report the best results among four scoring functions (MSP, MaxLogit, Energy and KNN). Without any model tuning, we directly use the features extracted from these models for OOD detection evaluation to assess whether they are already sufficiently capable of OOD detection. To emphasize the", + "bbox": [ + 511, + 659, + 908, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1753", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6854e2425378abcd7a908b1d1bfd772c286727043d21e0781bb31e9e4a923111.jpg", + "image_caption": [ + "(a) CLIP", + "Figure 3. Feature Visualization for Foundation Models. For fine-grained feature visualization, we randomly select fine-grained categories under 3 different super classes from ImageNet-1k." + ], + "image_footnote": [], + "bbox": [ + 106, + 106, + 277, + 239 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/621774629dafb82cfebe68739b3b61a219516034b5139d419a5ea6a24713a897.jpg", + "image_caption": [ + "(b) DINov2" + ], + "image_footnote": [], + "bbox": [ + 303, + 106, + 475, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "significance of our findings, we also compare them with the state-of-the-art method (NegPrompt [30]) that involve fine-tuning an ImageNet pre-trained model on the ID dataset. Additionally, we conduct further verification using datasets beyond ImageNet-1k as ID data in Appendix A.4 and the conclusion align with the experiments using ImageNet-1k.", + "bbox": [ + 89, + 349, + 483, + 441 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Result Analysis. (1) With traditional score metrics (i.e KNN), DINOv2 can outperform all other foundation models by a large margin. DINOv2+KNN shows the best results where the average FPR95 is $29.27\\%$ in the first four test sets, $8.9\\%$ in the latter challenging test sets, while SigCLIP only achieves $54.23\\%$ and $77.17\\%$ . This is potentially because DINOv2 leverages advanced self-supervised learning: iBot [80], which is a Mask Image Modeling (MIM) pretask for facilitating models to capture image details, and contrastive learning objective [2] that enhances the feature discriminativeness. (2) Without any fine-tuning, DINOv2 achieves performance comparable to the more complex method (i.e. NegPrompt) in the first four test sets. Notably, DINOv2+KNN still significantly outperforms NegPrompt on challenging test sets by $49.94\\%$ . The reason is that these two datasets contain images that are extremely similar to the ID categories. However, the paradigm of CLIP only provides image-level textual supervision without a supervision signal to retain detailed image information. Therefore, CLIP always fails in some fine-grained tasks, while DINOv2 consistently performs much better. As shown in Fig. 3a and Fig. 3b, where we randomly select 11 fine-grain categories under 3 different super classes, DINOv2 provide more discriminative boundaries, while CLIP can not.", + "bbox": [ + 91, + 443, + 483, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Further Challenges in OOD using Foundation models. In summary, DINov2, without requiring any fine-tuning, can already function as a high-performing OOD detector, surpassing previous approaches and underscoring the importance of discriminative and generalizable features for OOD detection. However, foundation models still have room to improve", + "bbox": [ + 89, + 810, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and cannot generalize well across the entire feature space. (1) Though there is a consensus that fine-tuning on the ID data can improve OOD performance [4, 15, 55, 67], we find that this doesn't hold in the context of foundation models, particularly on in-domain data with large semantic spaces. For instance, when we fine-tune DINOv2 on ImageNet-1K ID data and evaluate the fine-tuned model, the performance declines on three out of the four OOD datasets (see naive finetuning in Tab. 7 in Appendix). The implementation details of this finetuning can be referred to Appendix A.2. (2) Besides, as shown in Fig. 7c and Fig. 7f in Appendix, we also show some failure examples, where the models exhibit particularly poor feature discriminability, hindering effective OOD detection.", + "bbox": [ + 511, + 90, + 906, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 513, + 316, + 604, + 332 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section introduces our proposed methods for finetuning vision foundation models to enhance the OOD detection ability, which includes a Mixture of Feature Expert module in Sec. 4.1 and a Dynamic- $\\beta$ Mixup data augmentation strategy in Sec. 4.2.", + "bbox": [ + 511, + 342, + 906, + 417 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Mixture of Feature Experts", + "text_level": 1, + "bbox": [ + 511, + 428, + 761, + 445 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Fig. 4, we propose Mixture of Feature Experts (MoFE), which divides the complex semantic space into multiple subspaces and each expert specializes in a specific subspace. Each expert can tackle an easier problem instead of conducting OOD detection on a complicated distribution, which eases the optimization process while maintaining the generalizability of features. Below presents the detailed configuration of MoFE.", + "bbox": [ + 511, + 450, + 905, + 570 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given an RGB image $\\mathbf{v} \\in \\mathbb{R}^{H \\times W \\times 3}$ , where $H$ and $W$ are the origin resolution, we reshape the image $\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}$ into a sequence of flattened 2D patches $\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}$ , $C$ is the number of channels, $(P, P)$ is the resolution of each image patch. Next, we flatten the patches and map to $D$ dimensions with a trainable linear projection $\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}$ . A learnable embedding is pretended to the sequence of embedded patches $(\\mathbf{z}_0^0 = \\mathbf{x}_{\\mathrm{cls}}^0)$ and position embeddings are added to the patch embeddings $E_{pos} \\in \\mathbb{R}^{(N + 1) \\times D}$ . Then we input these embeddings to multiple transformer blocks. The output is processed by a MoFE layer to obtain the domain-specific features. This process is expressed as:", + "bbox": [ + 511, + 570, + 908, + 755 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {c l s}} ^ {0}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}; \\dots ; \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 766, + 906, + 785 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} ^ {\\prime} _ {\\ell} = \\operatorname {T r a n s f o r m e r} \\left(\\mathbf {z} _ {\\ell - 1}\\right) + \\mathbf {z} _ {\\ell - 1}, \\ell = 1 \\dots L, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 787, + 906, + 804 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} _ {\\ell} = \\operatorname {M o F E} \\left(\\mathrm {L N} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right)\\right) + \\mathbf {z} _ {\\ell} ^ {\\prime}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 806, + 906, + 824 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\mathrm {L N} (\\mathbf {z} _ {l}). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 825, + 906, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $LN$ denotes the layer norm.", + "bbox": [ + 511, + 854, + 743, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "MoFE Architecture. The MoFE layer consists of multiple expert networks, each of which is a transformer block. As an", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1754", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/6b9e2423c9dc87a8e500d123c6b9d8b942e60d81bb2bcc2daec7345460759837.jpg", + "image_caption": [ + "Figure 4. Illustration of our proposed Mixture of Feature Experts (MoFE). MoFE decomposes the large semantic space into multiple subspaces and each expert specializes in a specific subspace. Specifically, the image patches and the class token are input to obtain the preliminary patch embeddings and class embedding. A router is employed to determine the expert to further process the embeddings, and the input of the router is the class embedding. Finally, we apply associated experts to refine the class embeddings and the patch embeddings. We use the class embeddings output by MoFE and conduct the OOD detection in the corresponding subspace." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 467, + 309 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/7ea3225c529ab8b7e2a2e35a88ad2145d5e1fd49ba81c92347b53026d472b84d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 508, + 93, + 903, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "initialization step, we replicate the transformer blocks from the final layer of a foundation model to form an ensemble of experts $\\mathcal{E} = [e_1,e_2,\\dots ,e_E]$ . The router [50] is a linear layer that predicts the probability of each token being assigned to each expert. Routing accuracy is crucial for MoFE. The key question is what should be used to determine the results of feature routing? We explore various approaches, such as reinitializing a routing token, averaging patch embeddings, or utilizing class embeddings. We ultimately find that using the class embedding achieves the best results. Although it is not the embedding output from the last layer of the network, it is sufficiently discriminative. Therefore, we utilize the class embedding $\\mathbf{z}_l^0$ as the input of the router. The router is a linear layer that predicts the probability of each token being assigned to each expert. We formulate as:", + "bbox": [ + 88, + 412, + 485, + 638 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} = \\frac {e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i}}}{\\sum_ {j} ^ {E} e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {j}}}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 643, + 483, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where the router produces weight logits $f(\\mathbf{z}_l^{\\prime 0}) = \\mathbf{W} \\cdot \\mathbf{z}_l^{\\prime 0}$ , which are normalized by the softmax function. $\\mathbf{W} \\in \\mathbb{R}^{D \\times E}$ represents the lightweight training parameters and $E$ represents the number of experts. After determining the experts by using the class embedding, we input all embedding including the patch embeddings and class embedding to the activated experts. Each embedding is processed by the top- $k$ experts with the highest probabilities, and the weighted sum is calculated based on the softmax results of the probabilities:", + "bbox": [ + 88, + 693, + 486, + 832 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {M o F E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {k} \\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} \\cdot \\mathcal {E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) _ {i}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 838, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{E}$ represents the network of an expert [12]. In our", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoFE architecture, we route to only a single expert, thus $k = 1$ . We find that the router computation is reduced as we are only routing a token to a single expert and the performance does not increase when using more experts.", + "bbox": [ + 511, + 412, + 906, + 473 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature Space Separation. In MoFE, we aim to have different experts specialize in different subspaces. Therefore, we propose to first separate the whole feature space into multiple subspaces so that each expert specializes in learning features within its subspace. We use WordNet [39], which provides a good summary of the higher-level semantics of categories, to offer an initial partition of the subspace. Since the semantic and visual similarities are not completely equivalent, we further refine the clustering using K-Means to adjust for the discrepancies. Specifically, we extract feature representations $\\mathbf{z}_l^{\\prime 0}$ for each training image. Then, we calculate the class prototypes by averaging the features of the images from each category. Finally, we perform a K-Means clustering on categorical feature prototypes. The initial cluster centers are determined by the centroids of clusters, which are divided according to the original semantic space. After determining the clustering results, we assign different experts to different clusters, with samples from each category being routed to the corresponding expert model.", + "bbox": [ + 511, + 489, + 908, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "MoFE Training. We replace the final transformer block with a MoFE layer. Each transformer block within MoFE is initialized by the original final transformer block. We could set multiple layers as MoFE layers, but we find that using just the final layer achieves sufficiently good results. Then we randomly initialize a router layer and use the class token as the input. We use the labels generated by the above", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1755", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "clustering to supervise the routing:", + "bbox": [ + 89, + 90, + 323, + 107 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {r o u t e}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} y ^ {i} \\log \\left(\\mathcal {P} ^ {i} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 117, + 483, + 160 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For each expert, we leverage the categories within the corresponding cluster as the positive samples, and the categories beyond the cluster as the negative ones. Assuming that the category cluster of the $i$ th expert contains $Q_{i}$ classes, we set the categories beyond the cluster as the $Q_{i} + 1$ categories. The loss is designed as follows:", + "bbox": [ + 89, + 169, + 485, + 261 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e x p e r t}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} \\sum_ {q = 1} ^ {Q _ {i} + 1} y ^ {i} y ^ {q} \\log \\left(p _ {i} ^ {q} (\\mathbf {x})\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 137, + 272, + 483, + 316 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In order to achieve the sample balance for each cluster, we control the ratio of positive and negative samples as 1:1 during training. Therefore, the overall loss of MoFE is:", + "bbox": [ + 89, + 325, + 483, + 372 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M o F E}} = \\mathcal {L} _ {\\text {e x p e r t}} + \\mathcal {L} _ {\\text {r o u t e}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 202, + 385, + 483, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Discussion. MoFE is designed to address OOD issues under large-scale complex distributions. Similar to MOS [17], we decouple complex distributions into simpler subspace. However, while MOS approaches this solely from the perspective of the loss function, we approach it from the model perspective by assigning an expert model to each subspace. This allows each expert to focus on learning its assigned subspace, preventing interference between features from different partitions. To accurately assign expert models to different samples during inference, we have devised a new routing method that uses the [CLS] token as the input to the router network, since it encapsulates the semantic features of the entire image. With these designs, MoFE outperforms MOS by a substantial margin (see Tab. 1) and the feature visualization shows that MoFE can achieve more compact ID distribution and clearer decision boundaries between ID and OOD samples (Appendix Fig. 5).", + "bbox": [ + 89, + 419, + 485, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Dynamic- $\\beta$ Mixup", + "text_level": 1, + "bbox": [ + 89, + 686, + 269, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data augmentation (e.g., Mixup [64, 77]) has been proven to improve generalization during finetuning. Traditional Mixup [64, 77] augment samples and transform labels by:", + "bbox": [ + 89, + 708, + 483, + 755 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\quad \\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 116, + 766, + 483, + 782 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda \\sim \\mathrm{Beta}(\\sigma, \\sigma)$ . $\\lambda$ is the interpolation weight for generating new augmented samples. We observe that different categories exhibit varying levels of discriminativeness initialized by vision foundation models, as shown in Appendix Fig. 7c and Fig. 7f. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup can blur the decision boundary between ID", + "bbox": [ + 89, + 794, + 485, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and OOD, leading to degraded performance (Fig. 4 in Appendix). Therefore, we dynamically adjust the Beta distribution according to the feature discriminativeness per category. The reason is that when features of $x_{i}$ are discriminative enough, a small $\\lambda$ , which leads to a dissimilar sample, is not necessary for their representation learning. Instead, we should leverage similar samples from a large $\\lambda$ for building smooth decision boundaries. On the contrary, when features of a category show poor discriminativeness, we should set a relatively small $\\lambda$ to ease the feature learning. We use the accuracy of the validation set to measure the discriminativeness. Therefore, we set $\\lambda$ as:", + "bbox": [ + 511, + 90, + 908, + 272 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\lambda \\sim \\operatorname {B e t a} (\\sigma , \\sigma) \\text {f o r} \\sigma = 1 - w * s, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 593, + 282, + 906, + 297 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $w$ is a scaling factor and $s$ denotes the corresponding category's accuracy on the validation set. Because the probability density function of $\\mathrm{Beta}(\\sigma, \\sigma)$ is symmetric about 0.5 and ranges from 0 to 1, we need to ensure that with a larger $s$ , the probability of sampling larger values is greater. Therefore, we transform $\\lambda$ as:", + "bbox": [ + 511, + 306, + 908, + 397 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\lambda} = \\left\\{ \\begin{array}{l l} \\lambda & \\lambda \\geq 0. 5 \\\\ 1 - \\lambda & \\lambda < 0. 5 \\end{array} . \\right. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 405, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We determine the category difficulty at the beginning of the training and then update it during the training process. In our implementation, $x_{i}$ is the training sample, and $x_{j}$ is the instance used to corrupt $x_{i}$ . Therefore, we select the s from categories of $x_{i}$ , and we select samples from different classes. Additionally, we empirically find that using vanilla Mixup [64, 77] can cause feature norms to grow during finetuning vision foundation models (i.e., DINOv2), leading to performance degradation on the OOD task. In order to restrain the growth of feature norms, we propose to add a regularization term to suppress the increase in feature norm:", + "bbox": [ + 511, + 452, + 908, + 619 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {M i x u p}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\left(p ^ {c} (x)\\right) + R e g \\left(F ^ {0}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 627, + 906, + 669 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $C$ is the total number of categories, Reg denotes a regularization method, $F^0$ is the final class embeddings output by MoFE. By default, the regularization method has multiple choices, which can be $L_{2}$ norm or label smoothing. The final optimization objective is:", + "bbox": [ + 511, + 676, + 910, + 753 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M o F E}} + \\mathcal {L} _ {\\text {M i x u p}}. \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 762, + 906, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 800, + 645, + 816 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we set up a benchmark for evaluating OOD performance in Sec. 5.1. Then we compare our methods with the competitive baselines in Sec. 5.2. We conduct ablation studies and present more analysis on our designed method in Sec. 5.3.", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1756", + "bbox": [ + 483, + 945, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5c93cdad678b6240799c890f138f5ae0d35679407a6eddc76e9b2af39e96e9fc.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedEnergy [35]65.0087.1757.4087.3246.4391.1757.4087.3256.5588.2479.39
MSP [14]40.8988.6365.8181.2467.9080.1464.9678.1659.8982.0479.39
MaxLogit [15]60.8688.0355.587.4444.8191.1652.2586.0453.3588.1679.39
MCM [41]30.9194.6137.5992.5744.6989.7757.7786.1142.7490.7767.01
CLIPN [70]23.9495.2726.1793.9333.4592.2840.8390.9331.1093.1068.53
LSN [45]21.5695.8334.4891.2526.3294.3538.5490.4230.2292.9671.89
NegPrompt [30]6.3298.4327.6093.3422.8995.5535.2191.6023.0194.8166.84
Ours5.1997.2821.3294.6922.1095.1731.4792.1520.0294.8968.56
Dinov2-BasedMSP [14]25.0294.7657.0983.4553.6585.2248.7985.8148.1387.3186.01
MaxLogit [15]22.9694.5959.2178.4154.5281.8048.1784.1646.2184.7486.01
Energy [35]28.4893.1965.8874.4961.5478.7153.2981.9252.2982.0786.01
KNN [54]5.6797.6543.2588.2136.4290.2128.0492.6628.3492.1886.01
MOS [17]5.0197.8540.1590.3334.3291.8726.1492.9826.4093.2585.23
Ours2.7498.8224.3293.7317.3895.6518.5895.3817.0195.8986.40
", + "bbox": [ + 96, + 88, + 897, + 330 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative results of OOD detection performance for ImageNet-1k as ID. We employed our method on two pre-training paradigms (CLIP, and DINOv2). We use FPR95 and AUROC as evaluation metrics. We also report ID classification accuracy. The CLIP-based methods use ViT-B-16, and the DINOv2-based methods use ViT-B-14.", + "bbox": [ + 89, + 335, + 906, + 380 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Benchmark", + "text_level": 1, + "bbox": [ + 89, + 417, + 217, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In- and out-distribution Datasets. To validate the effectiveness of our proposed method, we conduct evaluation on standard benchmarks, which use ImageNet-1K [48] and ImageNet-100 [41] as the ID datasets. Following existing studies that leverage foundation models in OOD [30, 70], we use diverse OOD test datasets, including samples selected from iNaturalist18 [66], SUN [74], Places [79], and Textures [6].", + "bbox": [ + 88, + 440, + 482, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Method Comparison. We conduct method comparison on two pretaining paradigms(i.e. CLIP and DINOv2). For each group, we apply some traditional scoring metric (such as MSP [14], MaxLogit [15], Energy [35], KNN [54]). Moreover, we also involve the current CLIP-based state-of-the-art methods, such as MCM [41], CLIPN [70], NegPrompt [30], and LSN [45]. We use KNN [54] as the scoring function when using DINOv2, and follow the scoring metric of CLIPN [70] when applying our method to CLIP.", + "bbox": [ + 88, + 561, + 482, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Main Results", + "text_level": 1, + "bbox": [ + 89, + 710, + 228, + 724 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ImageNet-1K. We compare the proposed approach with the state-of-the-art methods for ImageNet-1K as ID on Tab. 1. These results show: 1) Based on DINOv2, our method reaches the best performance when setting ImageNet-1K as ID. Specifically, our approach reaches $17.01\\%$ FPR95 and $95.89\\%$ AUROC, averaging the results of all the OOD test sets. Our method surpasses MOS [17] by $9.39\\%$ in FPR95, and $2.64\\%$ in AUROC, which proves the importance of learning expert models for OOD detection. 2) When applying our method to CLIP, our method reaches $20.02\\%$ and $94.89\\%$ , which also outperforms NegPrompt [30] by", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "a large margin. These results indicate the effectiveness of the proposed MoFE and the dynamic regularized Mixup. 3) Our approach reaches $2.74\\%$ FPR95 on iNaturalist18 and increases the performance on all the test sets, which indicates that our MoFE design retains the discriminativeness of DINOv2 and facilitates feature learning on various feature subspaces.", + "bbox": [ + 511, + 417, + 906, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results on ImageNet-100. We compare the proposed approach with the state-of-the-art methods for ImageNet-100 as ID on Tab. 2. Based on DINOv2, our method reaches $8.10\\%$ FPR95 and $97.75\\%$ AUROC, surpassing the baseline by $4.40\\%$ FPR95, and $0.23\\%$ AUROC. This indicates that our proposed approach is also effective in a small-scale ID dataset. On the other hand, when applying our method to CLIP, we achieve $7.40\\%$ FPR95 and $98.10\\%$ AUROC, outperforming LSN [45] by $1.16\\%$ FPR95. The above experimental results validate the effectiveness of our approach, and we achieve the best performance on both small-scale and large-scale ID datasets.", + "bbox": [ + 511, + 523, + 908, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Analysis", + "text_level": 1, + "bbox": [ + 511, + 712, + 616, + 728 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this section, we conduct more analysis on the proposed methods. We use the DINOv2 as the pretaining paradigm and KNN [54] as the scoring function. We use ImageNet-1K as ID data and report the average performance on the four out-of-distribution datasets mentioned in Sec. 5.1. The baseline is set to DINOv2 with KNN scoring function without our designed method.", + "bbox": [ + 511, + 734, + 906, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Contributions of Individual Components. As shown in Tab. 3, we evaluate the contribution of each component of the full method. On ImageNet-1K, MoFE and Dynamic- $\\beta$ Mixup contribute $6.68\\%$ and $5.41\\%$ FPR95, respectively.", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1757", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/c5da24d9159e60b2806bbb8c946142b2826d5ea7cf4a1e27d8da369951939fa4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedMSP [14]23.5595.9240.4691.2337.0292.4524.4094.9031.4393.6391.93
MCM [41]18.1396.7734.5294.3636.4594.5441.2292.2532.5894.4887.88
CLIPN [70]4.8798.1613.6496.9313.5597.5615.7893.0211.9696.4191.64
LSN [45]4.9398.9212.8297.198.2397.988.2698.118.5698.0592.24
Ours3.2099.1710.0597.767.0698.399.3197.107.4098.1092.85
Dinov2-BasedMSP [14]5.0698.8526.5894.7827.6495.0226.4394.2721.4295.7294.50
MaxLogit [15]5.5598.7629.6994.1932.7394.2029.2793.7224.3195.2194.50
Energy [35]18.5796.6954.7288.9262.4287.1757.2888.4048.2190.2994.50
KNN [54]2.5899.0218.4595.1215.8996.1616.7996.3813.4296.6694.50
Ours2.2599.2312.8196.668.5197.868.8597.288.1097.7596.94
", + "bbox": [ + 96, + 88, + 898, + 277 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c98381c5b26e1755b6b94d678d73764e99ed708136e8cb4afaabb33c175d2425.jpg", + "table_caption": [ + "Table 2. Quantitative results of OOD detection performance for ImageNet-100 as ID." + ], + "table_footnote": [], + "table_body": "
SettingsBaseline+ MoFE+ D-β+ MoFE+D-β
FPR95↓29.2722.5923.8517.01
AUROC↑92.6794.0193.7295.89
", + "bbox": [ + 93, + 309, + 483, + 369 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0b24410b49ef9a3c9298aa57480fb1cc6362e7fe674cb5e3dfe5c24f00f048e4.jpg", + "table_caption": [ + "Table 3. Ablation study of individual components for ImageNet-1k as In-Distribution dataset." + ], + "table_footnote": [], + "table_body": "
Num.235789
Gain4.106.309.8012.2612.1012.09
", + "bbox": [ + 93, + 407, + 483, + 449 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/10971f6eb3cd1b400888644121afa48a7043434bf0e75f29554ad85c47a8499d.jpg", + "table_caption": [ + "Table 4. The effect of Cluster Number. We report the performance gain in FPR95 compared to the model without MoFE." + ], + "table_footnote": [], + "table_body": "
GroupingBaselineWordNetClusteringOurs
FPR95↓29.2725.6326.3422.59
AUROC↑92.6793.0192.9994.01
", + "bbox": [ + 94, + 481, + 483, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "When combined, the best performance (17.01% FPR95, and 95.89% AUROC) is achieved.", + "bbox": [ + 89, + 597, + 482, + 627 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Cluster Number. We conduct an experiment to validate the impact of cluster number on MoFE performance. We set different numbers of clusters. As shown in Tab. 4, we report the performance gain in FPR95. As the increasing of cluster number, the performance gradually increases.", + "bbox": [ + 89, + 628, + 482, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Feature Space Separation. As shown in Tab. 5, we validate the different strategies for determining the subspace. We compare our method with two methods: WordNet and clustering. The results show that using the ours (i.e. WordNet + Clustering) is the most promising approach. The reason might be that the features extracted by pretrained model are discriminative, especially at coarse-grained level.", + "bbox": [ + 88, + 704, + 482, + 809 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "More Analysis on Dynamic- $\\beta$ Mixup. As shown in Tab. 6, we conduct an ablation study on Dynamic- $\\beta$ Mixup. When we remove the regularization term, we find that the performance degrades (30.43% FPR95). Moreover, when we dynamic beta distribution is removed, the performance decreases to 24.96% FPR95.", + "bbox": [ + 89, + 810, + 485, + 898 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/90ad20f56a5338d37487e9b12a5d14fada5c7f4ae03811182471e71eba847855.jpg", + "table_caption": [ + "Table 5. Analysis on Grouping Strategy in MoFE." + ], + "table_footnote": [], + "table_body": "
MethodsBaselinew/o Regw/o D-βOurs
FPR95↓29.2730.4324.9623.85
AUROC↑92.6791.6593.3693.72
", + "bbox": [ + 517, + 327, + 905, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of Dynamic- $\\beta$ Mixup.", + "bbox": [ + 568, + 385, + 846, + 398 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 436, + 633, + 452 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper studies the OOD detection task within the context of foundation models. Our study shows that vision foundation models (e.g., DINOv2) are effective OOD detectors, suggesting high-quality and generalizable feature space is essential for OOD detection. Our study highlights that CLIP's pre-trained feature space is less effective for fine-grained tasks, where DINOv2 performs significantly better, which worths further exploration. In order to further optimize the performance of OOD detection when ID data is available for fine-tuning, we propose MoFE and a Dynamic- $\\beta$ Mixup data augmentation to enhance the feature learning. We conduct extensive experiments and ablation studies to validate the effectiveness of our approach. We hope our investigation could inspire more future studies.", + "bbox": [ + 511, + 462, + 906, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 686, + 671, + 704 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This research is supported by the National Key RD Program of China (Grant No. 2023YFB2806300). This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 172024221721292317215025), Theme-based Research (Grant No. T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust.", + "bbox": [ + 511, + 712, + 908, + 892 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1758", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Sparse mixture-of-experts are domain generalizable learners, author=Li, Bo and Shen, Yifei and Yang, Jingkang and Wang, Yezhen and Ren, Jiawei and Che, Tong and Zhang, Jun and Liu, Ziwei, booktitle=ICLR, year=2023. 1", + "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3, 4", + "[3] Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024. 2", + "[4] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV, 2020. 4", + "[5] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160, 2024.3", + "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 7", + "[7] Hendrycks Dan, Zhao Kevin, Basart Steven, Steinhardt Jacob, and Song Dawn. Natural adversarial examples. In CVPR, 2021. 3", + "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1", + "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2", + "[10] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In NeurIPS, 2023. 1", + "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, 2022. 1, 2, 3", + "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, 2021. 2, 3, 5", + "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022. 3", + "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 3, 7, 8", + "[15] Dan Hendrycks, Steven Basart, Mantas Mazeika, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 3, 4, 7, 8", + "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image with" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "out learning from out-of-distribution data. In CVPR, 2020.", + "[17] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In CVPR, 2021. 2, 6, 7, 3", + "[18] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3", + "[19] Bitterwolf Julian, Mueller Maximilian, and Matthias. Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, 2023. 3", + "[20] Umar Khalid, Ashkan Esmaeili, Nazmul Karim, and Nazanin Rahnavard. Rodd: A self-supervised approach for robust out-of-distribution detection. In CVPRW, 2022. 2", + "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1", + "[22] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. arXiv preprint arXiv:2212.05055, 2022. 3", + "[23] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 1, 3", + "[24] Yamuna Krishnamurthy, Chris Watkins, and Thomas Gaertner. Improving expert specialization in mixture of experts. arXiv preprint arXiv:2302.14703, 2023. 3", + "[25] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1", + "[26] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2", + "[27] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 3", + "[28] Jinglun Li, Xinyu Zhou, Pinxue Guo, Yixuan Sun, Yiwen Huang, Weifeng Ge, and Wenqiang Zhang. Hierarchical visual categories modeling: A joint representation learning and density estimation framework for out-of-distribution detection. In ICCV, 2023. 2", + "[29] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1" + ], + "bbox": [ + 516, + 93, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1759", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, 2024. 4, 7", + "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1", + "[32] Randolph Linderman, Jingyang Zhang, Nathan Inkawich, Hai Li, and Yiran Chen. Fine-grain inference on out-of-distribution data with hierarchical classification. arXiv preprint arXiv:2209.04493, 2022. 2", + "[33] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 2", + "[34] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 2", + "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 3, 7, 8", + "[36] Yuxuan Lou, Fuzhao Xue, Zangwei Zheng, and Yang You. Cross-token modeling with conditional computation. arXiv preprint arXiv:2109.02008, 2021. 3", + "[37] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 2", + "[38] Oquab Maxime, Darcet Timothee, Moutakanni Théo, Vo Huy, Szafraniec Marc, Khalidov Vasil, Fernandez Pierre, Haziza Daniel, Massa Francisco, El-Nouby Alaaeldin, Assran Mahmoud, Ballas Nicolas, Galuba Wojciech, Howes Russell, Huang Po-Yao, Li Shang-Wen, Misra Ishan, Rabbat Michael, Sharma Vasu, Synnaeve Gabriel, Xu Hu, Jegou Herve, Mairal Julien, Labatut Patrick, Joulin Armand, and Bojanowski Piotr. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 1, 2, 3", + "[39] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994\", 1994. 5", + "[40] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? IJCV, 2024. 1, 2, 3", + "[41] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 7, 8, 2", + "[42] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NIPS, 2023. 2", + "[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. arXiv preprint arXiv:2206.02770, 2022. 3" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[44] Ibrahim Ndiour, Nilesh Ahuja, and Omesh Tickoo. Out-of-distribution detection with subspace techniques and probabilistic modeling of features. arXiv preprint arXiv:2012.04250, 2020. 2", + "[45] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 7, 8", + "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3", + "[47] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. NeurIPS, 2021. 3, 1", + "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7", + "[49] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In ICLR, 2021. 1, 2", + "[50] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3, 5", + "[51] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, Rohit Girdhar, and Ishan Misra. The effectiveness of mae pre-pretraining for billion-scale pretraining. In ICCV, 2023. 1", + "[52] Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In ECCV, 2022. 1", + "[53] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 1", + "[54] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 1, 2, 3, 7, 8", + "[55] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 4", + "[56] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 2", + "[57] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023. 2", + "[58] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023." + ], + "bbox": [ + 516, + 90, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1760", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[59] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024.", + "[60] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024.", + "[61] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025. 2", + "[62] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 2", + "[63] Leitian Tao, Xuefeng Du, Xiaojin Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 1", + "[64] Sunil Thulasidasan, Gopinath Chennupati, Jeff Bilmes, Tan-moy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurIPS, 2019. 2, 6", + "[65] Changyao Tian, Wenhai Wang, Xizhou Zhu, Xiaogang Wang, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. arXiv preprint arXiv:2111.13579, 2021. 1", + "[66] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 7", + "[67] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: a good closed-set classifier is all you need? In ICLR, 2022. 4", + "[68] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, 2022. 1", + "[69] Haotao Wang, Aston Zhang, Yi Zhu, Shuai Zheng, Mu Li, Alex J Smola, and Zhangyang Wang. Partial and asymmetric contrastive learning for out-of-distribution detection in long-tailed recognition. In ICML, 2022. 2", + "[70] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In ICCV, 2023. 1, 2, 3, 7, 8", + "[71] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2", + "[72] Sitong Wu, Tianyi Wu, Haoru Tan, and Guodong Guo. Pale transformer: A general vision transformer backbone with pale-shaped attention. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2022. 2", + "[73] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 3", + "[74] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 7" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[75] Zhai Xiaohua, Mustafa Basil, Kolesnikov Alexander, and Beyer Lucas. Sigmoid loss for language image pre-training, 2023. 3", + "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1", + "[77] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2017. 6", + "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 1", + "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. PAMI, 2017. 7", + "[80] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 4" + ], + "bbox": [ + 514, + 90, + 905, + 375 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1761", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_model.json b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bcbf7078b20e81d3011887a4f51d9802203b53f8 --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_model.json @@ -0,0 +1,2477 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.174, + 0.131, + 0.826, + 0.175 + ], + "angle": 0, + "content": "Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.204, + 0.766, + 0.241 + ], + "angle": 0, + "content": "Shizhen Zhao Jiahui Liu Xin Wen Haoru Tan Xiaojuan Qi† The University of Hong Kong" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.274, + 0.327, + 0.29 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.307, + 0.486, + 0.745 + ], + "angle": 0, + "content": "Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, we systematically investigate representative vision foundation models for OOD detection. Our findings reveal that a pre-trained DINOv2 model, even without fine-tuning on in-domain (ID) data, naturally provides a highly discriminative feature space for OOD detection, achieving performance comparable to existing state-of-the-art methods without requiring complex designs. Beyond this, we explore how fine-tuning foundation models on in-domain (ID) data can enhance OOD detection. However, we observe that the performance of vision foundation models remains unsatisfactory in scenarios with a large semantic space. This is due to the increased complexity of decision boundaries as the number of categories grows, which complicates the optimization process. To mitigate this, we propose the Mixture of Feature Experts (MoFE) module, which partitions features into subspaces, effectively capturing complex data distributions and refining decision boundaries. Further, we introduce a Dynamic- \\(\\beta\\) Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming baseline methods. The project will be available at shizhen-zhao.github.io/OOD_MoFE/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.775, + 0.222, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.801, + 0.485, + 0.877 + ], + "angle": 0, + "content": "The task of out-of-distribution (OOD) detection [16, 26, 31, 49] aims to equip models with the capability to discern whether input images originate from unknown OOD classes or belong to in-domain (ID) classes. Mainstream OOD detection methods [9, 10, 25, 63] focus on learning features" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.274, + 0.891, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.445, + 0.891, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.598, + 0.908, + 0.669 + ], + "angle": 0, + "content": "Figure 1. Holistic comparison to previous philosophy. (a) Traditional methods use a generalized model to project inputs onto a complex distribution; (b): Our approach leverages multiple experts to break the complex distribution into smaller ones, which leads to compact ID distribution and simplified decision boundary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.909, + 0.883 + ], + "angle": 0, + "content": "and classifiers [31, 52, 53, 68] from ID data and then develop a score metric [14, 15, 35, 54] to determine whether a sample belongs to ID or OOD classes. Despite significant advancements, the fundamental challenge in OOD detection is establishing a feature space with high discriminative capacity that can effectively distinguish OOD samples from ID samples. Recently, vision foundation models [21, 38, 46, 51] trained on large-scale datasets have demonstrated the ability to learn robust and generalizable features, benefiting numerous tasks [29, 65, 76, 78]. This raises the question: with such powerful models and feature representations, does OOD detection remain a problem?" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Although several studies [11, 40, 41, 70] have explored" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.235, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1751" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.288 + ], + "angle": 0, + "content": "the use of foundation models for OOD detection, most focus on improving the performance of vision-language models like CLIP [46], while other foundation models, such as DINOv2, remain largely unexamined. In this study, we systematically investigate the feature spaces of different representative pre-trained foundation models, including vision-language models (e.g., CLIP) and self-supervised models (e.g., DINOv2), in the context of OOD detection. Our results reveal that DINOv2 provides the most discriminative feature space, enabling effective OOD detection without any finetuning. Notably, using a simple KNN metric [54], DINOv2 achieves performance comparable to more complex methods, establishing a strong baseline for further research." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.29, + 0.485, + 0.563 + ], + "angle": 0, + "content": "While vision foundation models [38, 45, 46, 70] have achieved impressive performance in OOD detection, there is still room for improvement, particularly on in-domain data with large semantic spaces [57-61, 71] (e.g., \\(29.27\\%\\) FPR95 on the ImageNet-1K OOD benchmark [54]). This prompts us to investigate whether foundation models can be further optimized by leveraging available ID data. However, as the number of semantic classes increases, the complexity of the decision boundaries required to distinguish between ID and OOD data grows as well [17, 28, 32]. This heightened complexity creates challenges when fine-tuning foundation models on limited ID data. Previous methods (e.g. MOS [17]) decouple the complex space into simpler subspaces from the perspective of loss, which eases the optimization process and simplifies the decision boundaries. In this study, we tackle the problem orthogonally from the model perspective by designing a new Mixture of Experts (MoE) architecture to more thoroughly disentangle complex ID distribution." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.486, + 0.868 + ], + "angle": 0, + "content": "To address this issue, rather than directly optimizing the whole feature space [3, 33, 34, 37] (Fig. 1(a)), we propose a Mixture-of-Feature-Expert (MoFE) module, which utilizes multiple experts, and each expert specializes in a specific subspace and optimizes it accordingly (Fig. 1(b)). MoFE operates by partitioning the original feature space into \\( K \\) subspaces based on semantics and feature similarities within the ID dataset. Each subspace is assigned to a dedicated expert, and a router assigns samples to the appropriate expert based on these partitions. Different from previous studies [12, 50], we use the [CLS] token as the input to the router network, since it encapsulates the semantic feature of the whole image. During training, the router is supervised by the partition assignments to ensure accurate sample-to-expert mapping. Each expert focuses solely on optimizing features within its designated partition, which helps prevent interference between features from different partitions. The results show that our approach significantly surpasses the previous approaches by a large margin (see Tab. 1), revealing the importance of learning expert models for OOD tasks." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Additionally, given that data augmentation has been shown to enhance generalization for OOD, we introduce" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.318 + ], + "angle": 0, + "content": "a novel Mixup data augmentation strategy to further improve feature learning, which is better suited for advanced vision foundation models. Our design is based on the observation that different categories exhibit varying levels of discriminativeness with features from vision foundation models. In the original feature space, some categories show high discriminativeness, while others do not. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup [64] can blur the decision boundary between ID and OOD, leading to degraded performance (Appendix Fig. 6). Thus, unlike existing Mixup strategies that treat all categories equally [56, 62, 72], our approach makes Mixup weight sampling category-dependent by adjusting the sampling distribution (i.e. beta distribution) dynamically, taking into account their discriminativeness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.318, + 0.908, + 0.332 + ], + "angle": 0, + "content": "Our major contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.334, + 0.907, + 0.393 + ], + "angle": 0, + "content": "- We design a novel MoFE module to tailor pre-trained vision foundation models for OOD detection. This approach reduces the difficulty of fitting complex data distributions from limited data and eases the optimization process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.394, + 0.907, + 0.483 + ], + "angle": 0, + "content": "- We explore the effectiveness of the raw feature spaces from various vision foundation models for OOD detection. Through analysis, we leverage DINOv2 with simple scoring metric to establish a strong baseline. Additionally, we designed a Dynamic-\\(\\beta\\) Mixup that is better suited for advanced vision foundation models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.543 + ], + "angle": 0, + "content": "- Our extensive experimental results demonstrate the effectiveness of the proposed model, achieving significant improvements over several competitive baseline methods on standard benchmarks." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.334, + 0.907, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.657, + 0.572 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.909, + 0.84 + ], + "angle": 0, + "content": "Out-of-Distribution Detection The goal of OOD detection is to detect OOD images from the test dataset (containing both ID and OOD images). Designing the score function is the most popular method in OOD detection tasks. The scores are mainly derived from three sources: the probability [14, 15], the logits [15, 35], and the feature [26, 44]. Some studies [20, 49, 69] focus on leveraging contrastive learning to enhance the feature representation. Other studies show that synthesizing pseudo samples [9, 49, 55, 55] as OOD instances is also a promising approach to make the feature space more compact. The methods [17, 28, 32] are the most relevant to ours, which also break the semantic space into smaller ones. Different from these approaches, in our design, we propose a novel MoE module, with each expert exclusively concentrating on optimizing features within its specific partition. Our results demonstrate that our approach outperforms them by a substantial margin." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.911, + 0.903 + ], + "angle": 0, + "content": "OOD Detection with Foundation Models There are some existing OOD detection methods [11, 40-42, 45, 70] leveraging foundation models. Maximum Concept Matching" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1752" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.228 + ], + "angle": 0, + "content": "(MCM) [41] proposes a simple yet effective zero-shot OOD detection method by aligning visual features with textual concepts. Some other studies [45, 70] explore negative prompts to learn the diversity of negative features, enabling more accurate detection of OOD samples. Although these studies have made great progress by leveraging CLIP to enhance the performance in existing benchmarks, they only explore and fine-tune CLIP. In our studies, we explore different foundation models and explore a better fine-tuning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.485, + 0.45 + ], + "angle": 0, + "content": "Mixture of Experts Mixture of Experts has been studied independently in both computer vision [36, 43, 47, 73] and natural language processing [12, 22, 27, 50]. These works are studied in the context of conditional computation, which is to increase the number of model parameters without a proportional increase in computational cost. Currently, some studies [5, 24] explore improving expert specialization and leveraging MoE to mitigate data conflict problems, where some data might interfere with each other. In our study, we introduce MoFE to the out-of-distribution task in the context of foundation models and build specialized OOD detectors for different feature subspaces." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.21, + 0.491 + ], + "angle": 0, + "content": "3. Pilot Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.564 + ], + "angle": 0, + "content": "In this section, we first introduce preliminaries for the OOD detection task in Sec. 3.1. Then, we explore the impact of foundation models on OOD detection performance and analyze their strengths and weaknesses in Sec. 3.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.585, + 0.232, + 0.599 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.61, + 0.484, + 0.746 + ], + "angle": 0, + "content": "We consider supervised multi-class classification, where \\(\\mathcal{X}\\) represents the input image space and \\(\\mathcal{Y} = \\{1,2,\\dots,C\\}\\) represents the label space. The training dataset \\(\\mathbb{D}_{in} = \\{(\\mathbf{x}_i,y_i)\\}_{i=1}^n\\) is drawn independently and identically distributed (i.i.d.) from the joint data distribution \\(P_{\\mathcal{X}\\mathcal{Y}}\\). Let \\(\\mathcal{P}_{\\mathrm{in}}\\) denote the marginal distribution on \\(\\mathcal{X}\\). Let \\(f:\\mathcal{X}\\mapsto \\mathbb{R}^{|\\mathcal{Y}|}\\) be a neural network trained on samples drawn from \\(P_{\\mathcal{X}\\mathcal{Y}}\\) to output a logit vector, which is used to predict the label of the input sample." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Out-of-distribution Detection. When deploying a machine learning model in real-world scenarios, it is crucial for a reliable classifier not only to accurately classify known distribution (ID) samples, but also to recognize any out-of-distribution (OOD) inputs as \"unknown\". This can be accomplished by incorporating an OOD detector alongside the classification model \\( f \\). OOD can be formulated as a binary classification task. During testing, the objective is to determine whether a sample \\( \\mathbf{x} \\in \\mathcal{X} \\) belongs to \\( \\mathcal{P}_{\\mathrm{in}}(\\mathrm{ID}) \\) or not (OOD). This decision can be made using a scoring" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.604, + 0.107 + ], + "angle": 0, + "content": "metric \\(S(\\mathbf{x})\\)" + }, + { + "type": "equation", + "bbox": [ + 0.605, + 0.117, + 0.907, + 0.159 + ], + "angle": 0, + "content": "\\[\nG _ {\\lambda} (x) = \\left\\{ \\begin{array}{l l} \\mathrm {I D} & S (\\mathbf {x}) \\geq \\lambda \\\\ \\text {O O D} & S (\\mathbf {x}) < \\lambda \\end{array} , \\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.907, + 0.229 + ], + "angle": 0, + "content": "where samples with higher scores \\( S(\\mathbf{x}) \\) are classified as ID and vice versa, and \\( \\lambda \\) is the threshold. Some typically used metrics \\( S(\\mathbf{x}) \\) include MSP [14], MaxLogit [15], Energy [35] and KNN [54]." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.242, + 0.885, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.433, + 0.907, + 0.475 + ], + "angle": 0, + "content": "Figure 2. Performance of vision foundation models across different OOD splits. The evaluation metric is FPR95, with lower values indicating better performance." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.487, + 0.86, + 0.502 + ], + "angle": 0, + "content": "3.2. Evaluation of Vision Foundation Models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.908, + 0.659 + ], + "angle": 0, + "content": "Although several studies [11, 40, 41, 70] have explored the use of foundation models for OOD detection, they focus solely on vision-language foundation models such as CLIP [46]. Beyond CLIP, the community offers a variety of vision foundation models that provide robust raw feature space. This development has inspired us to re-examine which vision foundation model is best suited for OOD detection. In this section, we aim to investigate and analyze various pre-trained vision foundation models as effective OOD detectors without fine-tuning." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Experimental Setup. We perform our evaluation on a challenging OOD detection benchmark that utilizes ImageNet-1K as ID data and selects samples from iNaturalist18, Sun, Places, and Textures as OOD samples. We also include two challenging OOD test sets: ImageNet-O [7] and NINCO [19]. We choose several representative vision foundation models, namely DINOv1 [2], DINOv2 [38], MAE [13], SigCLIP [75], OpenCLIP [18], and OpenAI CLIP [46]. For fair comparison, we use the ViT-B as the architecture of these models. The scoring functions for DINOv1, DINOv2, and MAE are set to KNN [54]. For the CLIP series, we report the best results among four scoring functions (MSP, MaxLogit, Energy and KNN). Without any model tuning, we directly use the features extracted from these models for OOD detection evaluation to assess whether they are already sufficiently capable of OOD detection. To emphasize the" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1753" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.107, + 0.278, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.25, + 0.22, + 0.262 + ], + "angle": 0, + "content": "(a) CLIP" + }, + { + "type": "image", + "bbox": [ + 0.304, + 0.107, + 0.476, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.358, + 0.25, + 0.426, + 0.262 + ], + "angle": 0, + "content": "(b) DINov2" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.274, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 3. Feature Visualization for Foundation Models. For fine-grained feature visualization, we randomly select fine-grained categories under 3 different super classes from ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.35, + 0.484, + 0.442 + ], + "angle": 0, + "content": "significance of our findings, we also compare them with the state-of-the-art method (NegPrompt [30]) that involve fine-tuning an ImageNet pre-trained model on the ID dataset. Additionally, we conduct further verification using datasets beyond ImageNet-1k as ID data in Appendix A.4 and the conclusion align with the experiments using ImageNet-1k." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.444, + 0.485, + 0.807 + ], + "angle": 0, + "content": "Result Analysis. (1) With traditional score metrics (i.e KNN), DINOv2 can outperform all other foundation models by a large margin. DINOv2+KNN shows the best results where the average FPR95 is \\(29.27\\%\\) in the first four test sets, \\(8.9\\%\\) in the latter challenging test sets, while SigCLIP only achieves \\(54.23\\%\\) and \\(77.17\\%\\). This is potentially because DINOv2 leverages advanced self-supervised learning: iBot [80], which is a Mask Image Modeling (MIM) pretask for facilitating models to capture image details, and contrastive learning objective [2] that enhances the feature discriminativeness. (2) Without any fine-tuning, DINOv2 achieves performance comparable to the more complex method (i.e. NegPrompt) in the first four test sets. Notably, DINOv2+KNN still significantly outperforms NegPrompt on challenging test sets by \\(49.94\\%\\). The reason is that these two datasets contain images that are extremely similar to the ID categories. However, the paradigm of CLIP only provides image-level textual supervision without a supervision signal to retain detailed image information. Therefore, CLIP always fails in some fine-grained tasks, while DINOv2 consistently performs much better. As shown in Fig. 3a and Fig. 3b, where we randomly select 11 fine-grain categories under 3 different super classes, DINOv2 provide more discriminative boundaries, while CLIP can not." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Further Challenges in OOD using Foundation models. In summary, DINov2, without requiring any fine-tuning, can already function as a high-performing OOD detector, surpassing previous approaches and underscoring the importance of discriminative and generalizable features for OOD detection. However, foundation models still have room to improve" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.304 + ], + "angle": 0, + "content": "and cannot generalize well across the entire feature space. (1) Though there is a consensus that fine-tuning on the ID data can improve OOD performance [4, 15, 55, 67], we find that this doesn't hold in the context of foundation models, particularly on in-domain data with large semantic spaces. For instance, when we fine-tune DINOv2 on ImageNet-1K ID data and evaluate the fine-tuned model, the performance declines on three out of the four OOD datasets (see naive finetuning in Tab. 7 in Appendix). The implementation details of this finetuning can be referred to Appendix A.2. (2) Besides, as shown in Fig. 7c and Fig. 7f in Appendix, we also show some failure examples, where the models exhibit particularly poor feature discriminability, hindering effective OOD detection." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.318, + 0.605, + 0.333 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.418 + ], + "angle": 0, + "content": "This section introduces our proposed methods for finetuning vision foundation models to enhance the OOD detection ability, which includes a Mixture of Feature Expert module in Sec. 4.1 and a Dynamic-\\(\\beta\\) Mixup data augmentation strategy in Sec. 4.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.429, + 0.762, + 0.446 + ], + "angle": 0, + "content": "4.1. Mixture of Feature Experts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.906, + 0.571 + ], + "angle": 0, + "content": "As shown in Fig. 4, we propose Mixture of Feature Experts (MoFE), which divides the complex semantic space into multiple subspaces and each expert specializes in a specific subspace. Each expert can tackle an easier problem instead of conducting OOD detection on a complicated distribution, which eases the optimization process while maintaining the generalizability of features. Below presents the detailed configuration of MoFE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.909, + 0.756 + ], + "angle": 0, + "content": "Given an RGB image \\(\\mathbf{v} \\in \\mathbb{R}^{H \\times W \\times 3}\\), where \\(H\\) and \\(W\\) are the origin resolution, we reshape the image \\(\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}\\) into a sequence of flattened 2D patches \\(\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}\\), \\(C\\) is the number of channels, \\((P, P)\\) is the resolution of each image patch. Next, we flatten the patches and map to \\(D\\) dimensions with a trainable linear projection \\(\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}\\). A learnable embedding is pretended to the sequence of embedded patches \\((\\mathbf{z}_0^0 = \\mathbf{x}_{\\mathrm{cls}}^0)\\) and position embeddings are added to the patch embeddings \\(E_{pos} \\in \\mathbb{R}^{(N + 1) \\times D}\\). Then we input these embeddings to multiple transformer blocks. The output is processed by a MoFE layer to obtain the domain-specific features. This process is expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.767, + 0.907, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {c l s}} ^ {0}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}; \\dots ; \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.789, + 0.907, + 0.805 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} ^ {\\prime} _ {\\ell} = \\operatorname {T r a n s f o r m e r} \\left(\\mathbf {z} _ {\\ell - 1}\\right) + \\mathbf {z} _ {\\ell - 1}, \\ell = 1 \\dots L, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.808, + 0.907, + 0.825 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} _ {\\ell} = \\operatorname {M o F E} \\left(\\mathrm {L N} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right)\\right) + \\mathbf {z} _ {\\ell} ^ {\\prime}, \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.827, + 0.907, + 0.843 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\mathrm {L N} (\\mathbf {z} _ {l}). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.745, + 0.87 + ], + "angle": 0, + "content": "where \\(LN\\) denotes the layer norm." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "MoFE Architecture. The MoFE layer consists of multiple expert networks, each of which is a transformer block. As an" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1754" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.468, + 0.31 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.509, + 0.094, + 0.904, + 0.316 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.33, + 0.907, + 0.402 + ], + "angle": 0, + "content": "Figure 4. Illustration of our proposed Mixture of Feature Experts (MoFE). MoFE decomposes the large semantic space into multiple subspaces and each expert specializes in a specific subspace. Specifically, the image patches and the class token are input to obtain the preliminary patch embeddings and class embedding. A router is employed to determine the expert to further process the embeddings, and the input of the router is the class embedding. Finally, we apply associated experts to refine the class embeddings and the patch embeddings. We use the class embeddings output by MoFE and conduct the OOD detection in the corresponding subspace." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.413, + 0.486, + 0.64 + ], + "angle": 0, + "content": "initialization step, we replicate the transformer blocks from the final layer of a foundation model to form an ensemble of experts \\(\\mathcal{E} = [e_1,e_2,\\dots ,e_E]\\). The router [50] is a linear layer that predicts the probability of each token being assigned to each expert. Routing accuracy is crucial for MoFE. The key question is what should be used to determine the results of feature routing? We explore various approaches, such as reinitializing a routing token, averaging patch embeddings, or utilizing class embeddings. We ultimately find that using the class embedding achieves the best results. Although it is not the embedding output from the last layer of the network, it is sufficiently discriminative. Therefore, we utilize the class embedding \\(\\mathbf{z}_l^0\\) as the input of the router. The router is a linear layer that predicts the probability of each token being assigned to each expert. We formulate as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.645, + 0.484, + 0.687 + ], + "angle": 0, + "content": "\\[\n\\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} = \\frac {e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i}}}{\\sum_ {j} ^ {E} e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {j}}}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.694, + 0.487, + 0.833 + ], + "angle": 0, + "content": "where the router produces weight logits \\( f(\\mathbf{z}_l^{\\prime 0}) = \\mathbf{W} \\cdot \\mathbf{z}_l^{\\prime 0} \\), which are normalized by the softmax function. \\( \\mathbf{W} \\in \\mathbb{R}^{D \\times E} \\) represents the lightweight training parameters and \\( E \\) represents the number of experts. After determining the experts by using the class embedding, we input all embedding including the patch embeddings and class embedding to the activated experts. Each embedding is processed by the top-\\( k \\) experts with the highest probabilities, and the weighted sum is calculated based on the softmax results of the probabilities:" + }, + { + "type": "equation", + "bbox": [ + 0.164, + 0.839, + 0.484, + 0.88 + ], + "angle": 0, + "content": "\\[\n\\operatorname {M o F E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {k} \\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} \\cdot \\mathcal {E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) _ {i}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathcal{E}\\) represents the network of an expert [12]. In our" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.413, + 0.907, + 0.474 + ], + "angle": 0, + "content": "MoFE architecture, we route to only a single expert, thus \\( k = 1 \\). We find that the router computation is reduced as we are only routing a token to a single expert and the performance does not increase when using more experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Feature Space Separation. In MoFE, we aim to have different experts specialize in different subspaces. Therefore, we propose to first separate the whole feature space into multiple subspaces so that each expert specializes in learning features within its subspace. We use WordNet [39], which provides a good summary of the higher-level semantics of categories, to offer an initial partition of the subspace. Since the semantic and visual similarities are not completely equivalent, we further refine the clustering using K-Means to adjust for the discrepancies. Specifically, we extract feature representations \\(\\mathbf{z}_l^{\\prime 0}\\) for each training image. Then, we calculate the class prototypes by averaging the features of the images from each category. Finally, we perform a K-Means clustering on categorical feature prototypes. The initial cluster centers are determined by the centroids of clusters, which are divided according to the original semantic space. After determining the clustering results, we assign different experts to different clusters, with samples from each category being routed to the corresponding expert model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "MoFE Training. We replace the final transformer block with a MoFE layer. Each transformer block within MoFE is initialized by the original final transformer block. We could set multiple layers as MoFE layers, but we find that using just the final layer achieves sufficiently good results. Then we randomly initialize a router layer and use the class token as the input. We use the labels generated by the above" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1755" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.325, + 0.108 + ], + "angle": 0, + "content": "clustering to supervise the routing:" + }, + { + "type": "equation", + "bbox": [ + 0.158, + 0.118, + 0.484, + 0.161 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {r o u t e}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} y ^ {i} \\log \\left(\\mathcal {P} ^ {i} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.486, + 0.262 + ], + "angle": 0, + "content": "For each expert, we leverage the categories within the corresponding cluster as the positive samples, and the categories beyond the cluster as the negative ones. Assuming that the category cluster of the \\(i\\)th expert contains \\(Q_{i}\\) classes, we set the categories beyond the cluster as the \\(Q_{i} + 1\\) categories. The loss is designed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.138, + 0.273, + 0.484, + 0.317 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e x p e r t}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} \\sum_ {q = 1} ^ {Q _ {i} + 1} y ^ {i} y ^ {q} \\log \\left(p _ {i} ^ {q} (\\mathbf {x})\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.485, + 0.373 + ], + "angle": 0, + "content": "In order to achieve the sample balance for each cluster, we control the ratio of positive and negative samples as 1:1 during training. Therefore, the overall loss of MoFE is:" + }, + { + "type": "equation", + "bbox": [ + 0.203, + 0.386, + 0.484, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M o F E}} = \\mathcal {L} _ {\\text {e x p e r t}} + \\mathcal {L} _ {\\text {r o u t e}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.486, + 0.678 + ], + "angle": 0, + "content": "Discussion. MoFE is designed to address OOD issues under large-scale complex distributions. Similar to MOS [17], we decouple complex distributions into simpler subspace. However, while MOS approaches this solely from the perspective of the loss function, we approach it from the model perspective by assigning an expert model to each subspace. This allows each expert to focus on learning its assigned subspace, preventing interference between features from different partitions. To accurately assign expert models to different samples during inference, we have devised a new routing method that uses the [CLS] token as the input to the router network, since it encapsulates the semantic features of the entire image. With these designs, MoFE outperforms MOS by a substantial margin (see Tab. 1) and the feature visualization shows that MoFE can achieve more compact ID distribution and clearer decision boundaries between ID and OOD samples (Appendix Fig. 5)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.687, + 0.27, + 0.704 + ], + "angle": 0, + "content": "4.2. Dynamic- \\(\\beta\\) Mixup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.485, + 0.756 + ], + "angle": 0, + "content": "Data augmentation (e.g., Mixup [64, 77]) has been proven to improve generalization during finetuning. Traditional Mixup [64, 77] augment samples and transform labels by:" + }, + { + "type": "equation", + "bbox": [ + 0.117, + 0.767, + 0.484, + 0.784 + ], + "angle": 0, + "content": "\\[\n\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\quad \\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.901 + ], + "angle": 0, + "content": "where \\(\\lambda \\sim \\mathrm{Beta}(\\sigma, \\sigma)\\). \\(\\lambda\\) is the interpolation weight for generating new augmented samples. We observe that different categories exhibit varying levels of discriminativeness initialized by vision foundation models, as shown in Appendix Fig. 7c and Fig. 7f. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup can blur the decision boundary between ID" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.273 + ], + "angle": 0, + "content": "and OOD, leading to degraded performance (Fig. 4 in Appendix). Therefore, we dynamically adjust the Beta distribution according to the feature discriminativeness per category. The reason is that when features of \\( x_{i} \\) are discriminative enough, a small \\( \\lambda \\), which leads to a dissimilar sample, is not necessary for their representation learning. Instead, we should leverage similar samples from a large \\( \\lambda \\) for building smooth decision boundaries. On the contrary, when features of a category show poor discriminativeness, we should set a relatively small \\( \\lambda \\) to ease the feature learning. We use the accuracy of the validation set to measure the discriminativeness. Therefore, we set \\( \\lambda \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.594, + 0.283, + 0.908, + 0.299 + ], + "angle": 0, + "content": "\\[\n\\lambda \\sim \\operatorname {B e t a} (\\sigma , \\sigma) \\text {f o r} \\sigma = 1 - w * s, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.909, + 0.398 + ], + "angle": 0, + "content": "where \\( w \\) is a scaling factor and \\( s \\) denotes the corresponding category's accuracy on the validation set. Because the probability density function of \\( \\mathrm{Beta}(\\sigma, \\sigma) \\) is symmetric about 0.5 and ranges from 0 to 1, we need to ensure that with a larger \\( s \\), the probability of sampling larger values is greater. Therefore, we transform \\( \\lambda \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.406, + 0.907, + 0.448 + ], + "angle": 0, + "content": "\\[\n\\hat {\\lambda} = \\left\\{ \\begin{array}{l l} \\lambda & \\lambda \\geq 0. 5 \\\\ 1 - \\lambda & \\lambda < 0. 5 \\end{array} . \\right. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.453, + 0.909, + 0.621 + ], + "angle": 0, + "content": "We determine the category difficulty at the beginning of the training and then update it during the training process. In our implementation, \\( x_{i} \\) is the training sample, and \\( x_{j} \\) is the instance used to corrupt \\( x_{i} \\). Therefore, we select the s from categories of \\( x_{i} \\), and we select samples from different classes. Additionally, we empirically find that using vanilla Mixup [64, 77] can cause feature norms to grow during finetuning vision foundation models (i.e., DINOv2), leading to performance degradation on the OOD task. In order to restrain the growth of feature norms, we propose to add a regularization term to suppress the increase in feature norm:" + }, + { + "type": "equation", + "bbox": [ + 0.531, + 0.628, + 0.907, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {M i x u p}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\left(p ^ {c} (x)\\right) + R e g \\left(F ^ {0}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.911, + 0.754 + ], + "angle": 0, + "content": "where \\( C \\) is the total number of categories, Reg denotes a regularization method, \\( F^0 \\) is the final class embeddings output by MoFE. By default, the regularization method has multiple choices, which can be \\( L_{2} \\) norm or label smoothing. The final optimization objective is:" + }, + { + "type": "equation", + "bbox": [ + 0.626, + 0.763, + 0.907, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M o F E}} + \\mathcal {L} _ {\\text {M i x u p}}. \\tag {15}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.646, + 0.817 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.908, + 0.9 + ], + "angle": 0, + "content": "In this section, we set up a benchmark for evaluating OOD performance in Sec. 5.1. Then we compare our methods with the competitive baselines in Sec. 5.2. We conduct ablation studies and present more analysis on our designed method in Sec. 5.3." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1756" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.898, + 0.332 + ], + "angle": 0, + "content": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedEnergy [35]65.0087.1757.4087.3246.4391.1757.4087.3256.5588.2479.39
MSP [14]40.8988.6365.8181.2467.9080.1464.9678.1659.8982.0479.39
MaxLogit [15]60.8688.0355.587.4444.8191.1652.2586.0453.3588.1679.39
MCM [41]30.9194.6137.5992.5744.6989.7757.7786.1142.7490.7767.01
CLIPN [70]23.9495.2726.1793.9333.4592.2840.8390.9331.1093.1068.53
LSN [45]21.5695.8334.4891.2526.3294.3538.5490.4230.2292.9671.89
NegPrompt [30]6.3298.4327.6093.3422.8995.5535.2191.6023.0194.8166.84
Ours5.1997.2821.3294.6922.1095.1731.4792.1520.0294.8968.56
Dinov2-BasedMSP [14]25.0294.7657.0983.4553.6585.2248.7985.8148.1387.3186.01
MaxLogit [15]22.9694.5959.2178.4154.5281.8048.1784.1646.2184.7486.01
Energy [35]28.4893.1965.8874.4961.5478.7153.2981.9252.2982.0786.01
KNN [54]5.6797.6543.2588.2136.4290.2128.0492.6628.3492.1886.01
MOS [17]5.0197.8540.1590.3334.3291.8726.1492.9826.4093.2585.23
Ours2.7498.8224.3293.7317.3895.6518.5895.3817.0195.8986.40
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.337, + 0.908, + 0.381 + ], + "angle": 0, + "content": "Table 1. Quantitative results of OOD detection performance for ImageNet-1k as ID. We employed our method on two pre-training paradigms (CLIP, and DINOv2). We use FPR95 and AUROC as evaluation metrics. We also report ID classification accuracy. The CLIP-based methods use ViT-B-16, and the DINOv2-based methods use ViT-B-14." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.418, + 0.218, + 0.433 + ], + "angle": 0, + "content": "5.1. Benchmark" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.562 + ], + "angle": 0, + "content": "In- and out-distribution Datasets. To validate the effectiveness of our proposed method, we conduct evaluation on standard benchmarks, which use ImageNet-1K [48] and ImageNet-100 [41] as the ID datasets. Following existing studies that leverage foundation models in OOD [30, 70], we use diverse OOD test datasets, including samples selected from iNaturalist18 [66], SUN [74], Places [79], and Textures [6]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.563, + 0.483, + 0.701 + ], + "angle": 0, + "content": "Method Comparison. We conduct method comparison on two pretaining paradigms(i.e. CLIP and DINOv2). For each group, we apply some traditional scoring metric (such as MSP [14], MaxLogit [15], Energy [35], KNN [54]). Moreover, we also involve the current CLIP-based state-of-the-art methods, such as MCM [41], CLIPN [70], NegPrompt [30], and LSN [45]. We use KNN [54] as the scoring function when using DINOv2, and follow the scoring metric of CLIPN [70] when applying our method to CLIP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.711, + 0.23, + 0.726 + ], + "angle": 0, + "content": "5.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Results on ImageNet-1K. We compare the proposed approach with the state-of-the-art methods for ImageNet-1K as ID on Tab. 1. These results show: 1) Based on DINOv2, our method reaches the best performance when setting ImageNet-1K as ID. Specifically, our approach reaches \\(17.01\\%\\) FPR95 and \\(95.89\\%\\) AUROC, averaging the results of all the OOD test sets. Our method surpasses MOS [17] by \\(9.39\\%\\) in FPR95, and \\(2.64\\%\\) in AUROC, which proves the importance of learning expert models for OOD detection. 2) When applying our method to CLIP, our method reaches \\(20.02\\%\\) and \\(94.89\\%\\), which also outperforms NegPrompt [30] by" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.908, + 0.525 + ], + "angle": 0, + "content": "a large margin. These results indicate the effectiveness of the proposed MoFE and the dynamic regularized Mixup. 3) Our approach reaches \\(2.74\\%\\) FPR95 on iNaturalist18 and increases the performance on all the test sets, which indicates that our MoFE design retains the discriminativeness of DINOv2 and facilitates feature learning on various feature subspaces." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.909, + 0.706 + ], + "angle": 0, + "content": "Results on ImageNet-100. We compare the proposed approach with the state-of-the-art methods for ImageNet-100 as ID on Tab. 2. Based on DINOv2, our method reaches \\(8.10\\%\\) FPR95 and \\(97.75\\%\\) AUROC, surpassing the baseline by \\(4.40\\%\\) FPR95, and \\(0.23\\%\\) AUROC. This indicates that our proposed approach is also effective in a small-scale ID dataset. On the other hand, when applying our method to CLIP, we achieve \\(7.40\\%\\) FPR95 and \\(98.10\\%\\) AUROC, outperforming LSN [45] by \\(1.16\\%\\) FPR95. The above experimental results validate the effectiveness of our approach, and we achieve the best performance on both small-scale and large-scale ID datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.617, + 0.729 + ], + "angle": 0, + "content": "5.3. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.84 + ], + "angle": 0, + "content": "In this section, we conduct more analysis on the proposed methods. We use the DINOv2 as the pretaining paradigm and KNN [54] as the scoring function. We use ImageNet-1K as ID data and report the average performance on the four out-of-distribution datasets mentioned in Sec. 5.1. The baseline is set to DINOv2 with KNN scoring function without our designed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Contributions of Individual Components. As shown in Tab. 3, we evaluate the contribution of each component of the full method. On ImageNet-1K, MoFE and Dynamic-\\(\\beta\\) Mixup contribute \\(6.68\\%\\) and \\(5.41\\%\\) FPR95, respectively." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1757" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.098, + 0.089, + 0.9, + 0.279 + ], + "angle": 0, + "content": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedMSP [14]23.5595.9240.4691.2337.0292.4524.4094.9031.4393.6391.93
MCM [41]18.1396.7734.5294.3636.4594.5441.2292.2532.5894.4887.88
CLIPN [70]4.8798.1613.6496.9313.5597.5615.7893.0211.9696.4191.64
LSN [45]4.9398.9212.8297.198.2397.988.2698.118.5698.0592.24
Ours3.2099.1710.0597.767.0698.399.3197.107.4098.1092.85
Dinov2-BasedMSP [14]5.0698.8526.5894.7827.6495.0226.4394.2721.4295.7294.50
MaxLogit [15]5.5598.7629.6994.1932.7394.2029.2793.7224.3195.2194.50
Energy [35]18.5796.6954.7288.9262.4287.1757.2888.4048.2190.2994.50
KNN [54]2.5899.0218.4595.1215.8996.1616.7996.3813.4296.6694.50
Ours2.2599.2312.8196.668.5197.868.8597.288.1097.7596.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.231, + 0.285, + 0.766, + 0.3 + ], + "angle": 0, + "content": "Table 2. Quantitative results of OOD detection performance for ImageNet-100 as ID." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.31, + 0.484, + 0.37 + ], + "angle": 0, + "content": "
SettingsBaseline+ MoFE+ D-β+ MoFE+D-β
FPR95↓29.2722.5923.8517.01
AUROC↑92.6794.0193.7295.89
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.372, + 0.484, + 0.399 + ], + "angle": 0, + "content": "Table 3. Ablation study of individual components for ImageNet-1k as In-Distribution dataset." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.409, + 0.484, + 0.45 + ], + "angle": 0, + "content": "
Num.235789
Gain4.106.309.8012.2612.1012.09
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.452, + 0.483, + 0.479 + ], + "angle": 0, + "content": "Table 4. The effect of Cluster Number. We report the performance gain in FPR95 compared to the model without MoFE." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.482, + 0.484, + 0.543 + ], + "angle": 0, + "content": "
GroupingBaselineWordNetClusteringOurs
FPR95↓29.2725.6326.3422.59
AUROC↑92.6793.0192.9994.01
" + }, + { + "type": "table_caption", + "bbox": [ + 0.137, + 0.544, + 0.436, + 0.558 + ], + "angle": 0, + "content": "Table 5. Analysis on Grouping Strategy in MoFE." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.483, + 0.628 + ], + "angle": 0, + "content": "When combined, the best performance (17.01% FPR95, and 95.89% AUROC) is achieved." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.703 + ], + "angle": 0, + "content": "Cluster Number. We conduct an experiment to validate the impact of cluster number on MoFE performance. We set different numbers of clusters. As shown in Tab. 4, we report the performance gain in FPR95. As the increasing of cluster number, the performance gradually increases." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.81 + ], + "angle": 0, + "content": "Feature Space Separation. As shown in Tab. 5, we validate the different strategies for determining the subspace. We compare our method with two methods: WordNet and clustering. The results show that using the ours (i.e. WordNet + Clustering) is the most promising approach. The reason might be that the features extracted by pretrained model are discriminative, especially at coarse-grained level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.9 + ], + "angle": 0, + "content": "More Analysis on Dynamic-\\(\\beta\\) Mixup. As shown in Tab. 6, we conduct an ablation study on Dynamic-\\(\\beta\\) Mixup. When we remove the regularization term, we find that the performance degrades (30.43% FPR95). Moreover, when we dynamic beta distribution is removed, the performance decreases to 24.96% FPR95." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.328, + 0.906, + 0.386 + ], + "angle": 0, + "content": "
MethodsBaselinew/o Regw/o D-βOurs
FPR95↓29.2730.4324.9623.85
AUROC↑92.6791.6593.3693.72
" + }, + { + "type": "table_caption", + "bbox": [ + 0.57, + 0.386, + 0.848, + 0.4 + ], + "angle": 0, + "content": "Table 6. Ablation study of Dynamic- \\( \\beta \\) Mixup." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.438, + 0.634, + 0.453 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.463, + 0.908, + 0.675 + ], + "angle": 0, + "content": "This paper studies the OOD detection task within the context of foundation models. Our study shows that vision foundation models (e.g., DINOv2) are effective OOD detectors, suggesting high-quality and generalizable feature space is essential for OOD detection. Our study highlights that CLIP's pre-trained feature space is less effective for fine-grained tasks, where DINOv2 performs significantly better, which worths further exploration. In order to further optimize the performance of OOD detection when ID data is available for fine-tuning, we propose MoFE and a Dynamic- \\(\\beta\\) Mixup data augmentation to enhance the feature learning. We conduct extensive experiments and ablation studies to validate the effectiveness of our approach. We hope our investigation could inspire more future studies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.688, + 0.673, + 0.705 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.713, + 0.909, + 0.893 + ], + "angle": 0, + "content": "This research is supported by the National Key RD Program of China (Grant No. 2023YFB2806300). This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 172024221721292317215025), Theme-based Research (Grant No. T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1758" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Sparse mixture-of-experts are domain generalizable learners, author=Li, Bo and Shen, Yifei and Yang, Jingkang and Wang, Yezhen and Ren, Jiawei and Che, Tong and Zhang, Jun and Liu, Ziwei, booktitle=ICLR, year=2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.226 + ], + "angle": 0, + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.229, + 0.483, + 0.271 + ], + "angle": 0, + "content": "[3] Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.329, + 0.484, + 0.383 + ], + "angle": 0, + "content": "[5] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.484, + 0.47 + ], + "angle": 0, + "content": "[7] Hendrycks Dan, Zhao Kevin, Basart Steven, Steinhardt Jacob, and Song Dawn. Natural adversarial examples. In CVPR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.484, + 0.555 + ], + "angle": 0, + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.558, + 0.482, + 0.598 + ], + "angle": 0, + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.601, + 0.482, + 0.641 + ], + "angle": 0, + "content": "[10] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.644, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, 2022. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, 2021. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.773, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[15] Dan Hendrycks, Steven Basart, Mantas Mazeika, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image with" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.907, + 0.119 + ], + "angle": 0, + "content": "out learning from out-of-distribution data. In CVPR, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.907, + 0.163 + ], + "angle": 0, + "content": "[17] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In CVPR, 2021. 2, 6, 7, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.907, + 0.233 + ], + "angle": 0, + "content": "[18] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.236, + 0.907, + 0.276 + ], + "angle": 0, + "content": "[19] Bitterwolf Julian, Mueller Maximilian, and Matthias. Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.906, + 0.32 + ], + "angle": 0, + "content": "[20] Umar Khalid, Ashkan Esmaeili, Nazmul Karim, and Nazanin Rahnavard. Rodd: A self-supervised approach for robust out-of-distribution detection. In CVPRW, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.907, + 0.377 + ], + "angle": 0, + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.379, + 0.907, + 0.448 + ], + "angle": 0, + "content": "[22] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. arXiv preprint arXiv:2212.05055, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.907, + 0.56 + ], + "angle": 0, + "content": "[23] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.562, + 0.907, + 0.603 + ], + "angle": 0, + "content": "[24] Yamuna Krishnamurthy, Chris Watkins, and Thomas Gaertner. Improving expert specialization in mixture of experts. arXiv preprint arXiv:2302.14703, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.605, + 0.907, + 0.646 + ], + "angle": 0, + "content": "[25] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.648, + 0.906, + 0.689 + ], + "angle": 0, + "content": "[26] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[27] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.907, + 0.83 + ], + "angle": 0, + "content": "[28] Jinglun Li, Xinyu Zhou, Pinxue Guo, Yixuan Sun, Yiwen Huang, Weifeng Ge, and Wenqiang Zhang. Hierarchical visual categories modeling: A joint representation learning and density estimation framework for out-of-distribution detection. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.907, + 0.899 + ], + "angle": 0, + "content": "[29] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.094, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1759" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.134 + ], + "angle": 0, + "content": "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[32] Randolph Linderman, Jingyang Zhang, Nathan Inkawich, Hai Li, and Yiran Chen. Fine-grain inference on out-of-distribution data with hierarchical classification. arXiv preprint arXiv:2209.04493, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.317 + ], + "angle": 0, + "content": "[33] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[34] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.483, + 0.418 + ], + "angle": 0, + "content": "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.483, + 0.462 + ], + "angle": 0, + "content": "[36] Yuxuan Lou, Fuzhao Xue, Zangwei Zheng, and Yang You. Cross-token modeling with conditional computation. arXiv preprint arXiv:2109.02008, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.463, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[37] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[38] Oquab Maxime, Darcet Timothee, Moutakanni Théo, Vo Huy, Szafraniec Marc, Khalidov Vasil, Fernandez Pierre, Haziza Daniel, Massa Francisco, El-Nouby Alaaeldin, Assran Mahmoud, Ballas Nicolas, Galuba Wojciech, Howes Russell, Huang Po-Yao, Li Shang-Wen, Misra Ishan, Rabbat Michael, Sharma Vasu, Synnaeve Gabriel, Xu Hu, Jegou Herve, Mairal Julien, Labatut Patrick, Joulin Armand, and Bojanowski Piotr. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[39] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994\", 1994. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.703, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[40] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? IJCV, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[41] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[42] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. arXiv preprint arXiv:2206.02770, 2022. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.147 + ], + "angle": 0, + "content": "[44] Ibrahim Ndiour, Nilesh Ahuja, and Omesh Tickoo. Out-of-distribution detection with subspace techniques and probabilistic modeling of features. arXiv preprint arXiv:2012.04250, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.906, + 0.192 + ], + "angle": 0, + "content": "[45] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.907, + 0.263 + ], + "angle": 0, + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.906, + 0.322 + ], + "angle": 0, + "content": "[47] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. NeurIPS, 2021. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.907, + 0.392 + ], + "angle": 0, + "content": "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.394, + 0.906, + 0.436 + ], + "angle": 0, + "content": "[49] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In ICLR, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.439, + 0.907, + 0.495 + ], + "angle": 0, + "content": "[50] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.907, + 0.579 + ], + "angle": 0, + "content": "[51] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, Rohit Girdhar, and Ishan Misra. The effectiveness of mae pre-pretraining for billion-scale pretraining. In ICCV, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.582, + 0.906, + 0.609 + ], + "angle": 0, + "content": "[52] Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.612, + 0.907, + 0.652 + ], + "angle": 0, + "content": "[53] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.656, + 0.907, + 0.697 + ], + "angle": 0, + "content": "[54] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.906, + 0.742 + ], + "angle": 0, + "content": "[55] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.744, + 0.907, + 0.8 + ], + "angle": 0, + "content": "[56] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.907, + 0.857 + ], + "angle": 0, + "content": "[57] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[58] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1760" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[59] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[60] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.485, + 0.258 + ], + "angle": 0, + "content": "[61] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.483, + 0.3 + ], + "angle": 0, + "content": "[62] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.483, + 0.329 + ], + "angle": 0, + "content": "[63] Leitian Tao, Xuefeng Du, Xiaojin Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.483, + 0.384 + ], + "angle": 0, + "content": "[64] Sunil Thulasidasan, Gopinath Chennupati, Jeff Bilmes, Tan-moy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurIPS, 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.386, + 0.483, + 0.441 + ], + "angle": 0, + "content": "[65] Changyao Tian, Wenhai Wang, Xizhou Zhu, Xiaogang Wang, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. arXiv preprint arXiv:2111.13579, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.483, + 0.496 + ], + "angle": 0, + "content": "[66] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.483, + 0.537 + ], + "angle": 0, + "content": "[67] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: a good closed-set classifier is all you need? In ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.58 + ], + "angle": 0, + "content": "[68] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.483, + 0.636 + ], + "angle": 0, + "content": "[69] Haotao Wang, Aston Zhang, Yi Zhu, Shuai Zheng, Mu Li, Alex J Smola, and Zhangyang Wang. Partial and asymmetric contrastive learning for out-of-distribution detection in long-tailed recognition. In ICML, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[70] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In ICCV, 2023. 1, 2, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[71] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[72] Sitong Wu, Tianyi Wu, Haoru Tan, and Guodong Guo. Pale transformer: A general vision transformer backbone with pale-shaped attention. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[73] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[74] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.133 + ], + "angle": 0, + "content": "[75] Zhai Xiaohua, Mustafa Basil, Kolesnikov Alexander, and Beyer Lucas. Sigmoid loss for language image pre-training, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.177 + ], + "angle": 0, + "content": "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.219 + ], + "angle": 0, + "content": "[77] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.906, + 0.333 + ], + "angle": 0, + "content": "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. PAMI, 2017. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.335, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[80] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 4" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.906, + 0.375 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1761" + } + ] +] \ No newline at end of file diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_origin.pdf b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db74944efb4d90fe511c84a3e2644486a4e6d473 --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/79e4b61c-cad3-4644-ae1d-912cf94be19d_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:857fc59363fb4c6a554f2a490579ee6bc902d24b3d46829bfd1c234667d4d4f5 +size 2236094 diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/full.md b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/full.md new file mode 100644 index 0000000000000000000000000000000000000000..bda90c85a37905761e80c19d2f522d3b145414fc --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/full.md @@ -0,0 +1,350 @@ +# Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection + +Shizhen Zhao Jiahui Liu Xin Wen Haoru Tan Xiaojuan Qi† The University of Hong Kong + +# Abstract + +Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, we systematically investigate representative vision foundation models for OOD detection. Our findings reveal that a pre-trained DINOv2 model, even without fine-tuning on in-domain (ID) data, naturally provides a highly discriminative feature space for OOD detection, achieving performance comparable to existing state-of-the-art methods without requiring complex designs. Beyond this, we explore how fine-tuning foundation models on in-domain (ID) data can enhance OOD detection. However, we observe that the performance of vision foundation models remains unsatisfactory in scenarios with a large semantic space. This is due to the increased complexity of decision boundaries as the number of categories grows, which complicates the optimization process. To mitigate this, we propose the Mixture of Feature Experts (MoFE) module, which partitions features into subspaces, effectively capturing complex data distributions and refining decision boundaries. Further, we introduce a Dynamic- $\beta$ Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming baseline methods. The project will be available at shizhen-zhao.github.io/OOD_MoFE/. + +# 1. Introduction + +The task of out-of-distribution (OOD) detection [16, 26, 31, 49] aims to equip models with the capability to discern whether input images originate from unknown OOD classes or belong to in-domain (ID) classes. Mainstream OOD detection methods [9, 10, 25, 63] focus on learning features + +![](images/51a4b5db034172c05cb8f0c79ae340ecf6d91f9b0131e1efcf1f0ed6c2273953.jpg) + +![](images/f220fba940cfbbb11dfef269d9272769b13ae241b780d1320ef0c5a2502abf37.jpg) +Figure 1. Holistic comparison to previous philosophy. (a) Traditional methods use a generalized model to project inputs onto a complex distribution; (b): Our approach leverages multiple experts to break the complex distribution into smaller ones, which leads to compact ID distribution and simplified decision boundary. + +and classifiers [31, 52, 53, 68] from ID data and then develop a score metric [14, 15, 35, 54] to determine whether a sample belongs to ID or OOD classes. Despite significant advancements, the fundamental challenge in OOD detection is establishing a feature space with high discriminative capacity that can effectively distinguish OOD samples from ID samples. Recently, vision foundation models [21, 38, 46, 51] trained on large-scale datasets have demonstrated the ability to learn robust and generalizable features, benefiting numerous tasks [29, 65, 76, 78]. This raises the question: with such powerful models and feature representations, does OOD detection remain a problem? + +Although several studies [11, 40, 41, 70] have explored + +the use of foundation models for OOD detection, most focus on improving the performance of vision-language models like CLIP [46], while other foundation models, such as DINOv2, remain largely unexamined. In this study, we systematically investigate the feature spaces of different representative pre-trained foundation models, including vision-language models (e.g., CLIP) and self-supervised models (e.g., DINOv2), in the context of OOD detection. Our results reveal that DINOv2 provides the most discriminative feature space, enabling effective OOD detection without any finetuning. Notably, using a simple KNN metric [54], DINOv2 achieves performance comparable to more complex methods, establishing a strong baseline for further research. + +While vision foundation models [38, 45, 46, 70] have achieved impressive performance in OOD detection, there is still room for improvement, particularly on in-domain data with large semantic spaces [57-61, 71] (e.g., $29.27\%$ FPR95 on the ImageNet-1K OOD benchmark [54]). This prompts us to investigate whether foundation models can be further optimized by leveraging available ID data. However, as the number of semantic classes increases, the complexity of the decision boundaries required to distinguish between ID and OOD data grows as well [17, 28, 32]. This heightened complexity creates challenges when fine-tuning foundation models on limited ID data. Previous methods (e.g. MOS [17]) decouple the complex space into simpler subspaces from the perspective of loss, which eases the optimization process and simplifies the decision boundaries. In this study, we tackle the problem orthogonally from the model perspective by designing a new Mixture of Experts (MoE) architecture to more thoroughly disentangle complex ID distribution. + +To address this issue, rather than directly optimizing the whole feature space [3, 33, 34, 37] (Fig. 1(a)), we propose a Mixture-of-Feature-Expert (MoFE) module, which utilizes multiple experts, and each expert specializes in a specific subspace and optimizes it accordingly (Fig. 1(b)). MoFE operates by partitioning the original feature space into $K$ subspaces based on semantics and feature similarities within the ID dataset. Each subspace is assigned to a dedicated expert, and a router assigns samples to the appropriate expert based on these partitions. Different from previous studies [12, 50], we use the [CLS] token as the input to the router network, since it encapsulates the semantic feature of the whole image. During training, the router is supervised by the partition assignments to ensure accurate sample-to-expert mapping. Each expert focuses solely on optimizing features within its designated partition, which helps prevent interference between features from different partitions. The results show that our approach significantly surpasses the previous approaches by a large margin (see Tab. 1), revealing the importance of learning expert models for OOD tasks. + +Additionally, given that data augmentation has been shown to enhance generalization for OOD, we introduce + +a novel Mixup data augmentation strategy to further improve feature learning, which is better suited for advanced vision foundation models. Our design is based on the observation that different categories exhibit varying levels of discriminativeness with features from vision foundation models. In the original feature space, some categories show high discriminativeness, while others do not. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup [64] can blur the decision boundary between ID and OOD, leading to degraded performance (Appendix Fig. 6). Thus, unlike existing Mixup strategies that treat all categories equally [56, 62, 72], our approach makes Mixup weight sampling category-dependent by adjusting the sampling distribution (i.e. beta distribution) dynamically, taking into account their discriminativeness. + +Our major contributions can be summarized as follows: + +- We design a novel MoFE module to tailor pre-trained vision foundation models for OOD detection. This approach reduces the difficulty of fitting complex data distributions from limited data and eases the optimization process. +- We explore the effectiveness of the raw feature spaces from various vision foundation models for OOD detection. Through analysis, we leverage DINOv2 with simple scoring metric to establish a strong baseline. Additionally, we designed a Dynamic- $\beta$ Mixup that is better suited for advanced vision foundation models. +- Our extensive experimental results demonstrate the effectiveness of the proposed model, achieving significant improvements over several competitive baseline methods on standard benchmarks. + +# 2. Related Work + +Out-of-Distribution Detection The goal of OOD detection is to detect OOD images from the test dataset (containing both ID and OOD images). Designing the score function is the most popular method in OOD detection tasks. The scores are mainly derived from three sources: the probability [14, 15], the logits [15, 35], and the feature [26, 44]. Some studies [20, 49, 69] focus on leveraging contrastive learning to enhance the feature representation. Other studies show that synthesizing pseudo samples [9, 49, 55, 55] as OOD instances is also a promising approach to make the feature space more compact. The methods [17, 28, 32] are the most relevant to ours, which also break the semantic space into smaller ones. Different from these approaches, in our design, we propose a novel MoE module, with each expert exclusively concentrating on optimizing features within its specific partition. Our results demonstrate that our approach outperforms them by a substantial margin. + +OOD Detection with Foundation Models There are some existing OOD detection methods [11, 40-42, 45, 70] leveraging foundation models. Maximum Concept Matching + +(MCM) [41] proposes a simple yet effective zero-shot OOD detection method by aligning visual features with textual concepts. Some other studies [45, 70] explore negative prompts to learn the diversity of negative features, enabling more accurate detection of OOD samples. Although these studies have made great progress by leveraging CLIP to enhance the performance in existing benchmarks, they only explore and fine-tune CLIP. In our studies, we explore different foundation models and explore a better fine-tuning paradigm. + +Mixture of Experts Mixture of Experts has been studied independently in both computer vision [36, 43, 47, 73] and natural language processing [12, 22, 27, 50]. These works are studied in the context of conditional computation, which is to increase the number of model parameters without a proportional increase in computational cost. Currently, some studies [5, 24] explore improving expert specialization and leveraging MoE to mitigate data conflict problems, where some data might interfere with each other. In our study, we introduce MoFE to the out-of-distribution task in the context of foundation models and build specialized OOD detectors for different feature subspaces. + +# 3. Pilot Study + +In this section, we first introduce preliminaries for the OOD detection task in Sec. 3.1. Then, we explore the impact of foundation models on OOD detection performance and analyze their strengths and weaknesses in Sec. 3.2. + +# 3.1. Preliminaries + +We consider supervised multi-class classification, where $\mathcal{X}$ represents the input image space and $\mathcal{Y} = \{1,2,\dots,C\}$ represents the label space. The training dataset $\mathbb{D}_{in} = \{(\mathbf{x}_i,y_i)\}_{i=1}^n$ is drawn independently and identically distributed (i.i.d.) from the joint data distribution $P_{\mathcal{X}\mathcal{Y}}$ . Let $\mathcal{P}_{\mathrm{in}}$ denote the marginal distribution on $\mathcal{X}$ . Let $f:\mathcal{X}\mapsto \mathbb{R}^{|\mathcal{Y}|}$ be a neural network trained on samples drawn from $P_{\mathcal{X}\mathcal{Y}}$ to output a logit vector, which is used to predict the label of the input sample. + +Out-of-distribution Detection. When deploying a machine learning model in real-world scenarios, it is crucial for a reliable classifier not only to accurately classify known distribution (ID) samples, but also to recognize any out-of-distribution (OOD) inputs as "unknown". This can be accomplished by incorporating an OOD detector alongside the classification model $f$ . OOD can be formulated as a binary classification task. During testing, the objective is to determine whether a sample $\mathbf{x} \in \mathcal{X}$ belongs to $\mathcal{P}_{\mathrm{in}}(\mathrm{ID})$ or not (OOD). This decision can be made using a scoring + +metric $S(\mathbf{x})$ + +$$ +G _ {\lambda} (x) = \left\{ \begin{array}{l l} \mathrm {I D} & S (\mathbf {x}) \geq \lambda \\ \text {O O D} & S (\mathbf {x}) < \lambda \end{array} , \right. \tag {1} +$$ + +where samples with higher scores $S(\mathbf{x})$ are classified as ID and vice versa, and $\lambda$ is the threshold. Some typically used metrics $S(\mathbf{x})$ include MSP [14], MaxLogit [15], Energy [35] and KNN [54]. + +![](images/a3be85820c442e52626f8d7e1a7769f131ca9bbdc9f0f2f42310cb5ad4971889.jpg) +Figure 2. Performance of vision foundation models across different OOD splits. The evaluation metric is FPR95, with lower values indicating better performance. + +# 3.2. Evaluation of Vision Foundation Models + +Although several studies [11, 40, 41, 70] have explored the use of foundation models for OOD detection, they focus solely on vision-language foundation models such as CLIP [46]. Beyond CLIP, the community offers a variety of vision foundation models that provide robust raw feature space. This development has inspired us to re-examine which vision foundation model is best suited for OOD detection. In this section, we aim to investigate and analyze various pre-trained vision foundation models as effective OOD detectors without fine-tuning. + +Experimental Setup. We perform our evaluation on a challenging OOD detection benchmark that utilizes ImageNet-1K as ID data and selects samples from iNaturalist18, Sun, Places, and Textures as OOD samples. We also include two challenging OOD test sets: ImageNet-O [7] and NINCO [19]. We choose several representative vision foundation models, namely DINOv1 [2], DINOv2 [38], MAE [13], SigCLIP [75], OpenCLIP [18], and OpenAI CLIP [46]. For fair comparison, we use the ViT-B as the architecture of these models. The scoring functions for DINOv1, DINOv2, and MAE are set to KNN [54]. For the CLIP series, we report the best results among four scoring functions (MSP, MaxLogit, Energy and KNN). Without any model tuning, we directly use the features extracted from these models for OOD detection evaluation to assess whether they are already sufficiently capable of OOD detection. To emphasize the + +![](images/6854e2425378abcd7a908b1d1bfd772c286727043d21e0781bb31e9e4a923111.jpg) +(a) CLIP +Figure 3. Feature Visualization for Foundation Models. For fine-grained feature visualization, we randomly select fine-grained categories under 3 different super classes from ImageNet-1k. + +![](images/621774629dafb82cfebe68739b3b61a219516034b5139d419a5ea6a24713a897.jpg) +(b) DINov2 + +significance of our findings, we also compare them with the state-of-the-art method (NegPrompt [30]) that involve fine-tuning an ImageNet pre-trained model on the ID dataset. Additionally, we conduct further verification using datasets beyond ImageNet-1k as ID data in Appendix A.4 and the conclusion align with the experiments using ImageNet-1k. + +Result Analysis. (1) With traditional score metrics (i.e KNN), DINOv2 can outperform all other foundation models by a large margin. DINOv2+KNN shows the best results where the average FPR95 is $29.27\%$ in the first four test sets, $8.9\%$ in the latter challenging test sets, while SigCLIP only achieves $54.23\%$ and $77.17\%$ . This is potentially because DINOv2 leverages advanced self-supervised learning: iBot [80], which is a Mask Image Modeling (MIM) pretask for facilitating models to capture image details, and contrastive learning objective [2] that enhances the feature discriminativeness. (2) Without any fine-tuning, DINOv2 achieves performance comparable to the more complex method (i.e. NegPrompt) in the first four test sets. Notably, DINOv2+KNN still significantly outperforms NegPrompt on challenging test sets by $49.94\%$ . The reason is that these two datasets contain images that are extremely similar to the ID categories. However, the paradigm of CLIP only provides image-level textual supervision without a supervision signal to retain detailed image information. Therefore, CLIP always fails in some fine-grained tasks, while DINOv2 consistently performs much better. As shown in Fig. 3a and Fig. 3b, where we randomly select 11 fine-grain categories under 3 different super classes, DINOv2 provide more discriminative boundaries, while CLIP can not. + +Further Challenges in OOD using Foundation models. In summary, DINov2, without requiring any fine-tuning, can already function as a high-performing OOD detector, surpassing previous approaches and underscoring the importance of discriminative and generalizable features for OOD detection. However, foundation models still have room to improve + +and cannot generalize well across the entire feature space. (1) Though there is a consensus that fine-tuning on the ID data can improve OOD performance [4, 15, 55, 67], we find that this doesn't hold in the context of foundation models, particularly on in-domain data with large semantic spaces. For instance, when we fine-tune DINOv2 on ImageNet-1K ID data and evaluate the fine-tuned model, the performance declines on three out of the four OOD datasets (see naive finetuning in Tab. 7 in Appendix). The implementation details of this finetuning can be referred to Appendix A.2. (2) Besides, as shown in Fig. 7c and Fig. 7f in Appendix, we also show some failure examples, where the models exhibit particularly poor feature discriminability, hindering effective OOD detection. + +# 4. Method + +This section introduces our proposed methods for finetuning vision foundation models to enhance the OOD detection ability, which includes a Mixture of Feature Expert module in Sec. 4.1 and a Dynamic- $\beta$ Mixup data augmentation strategy in Sec. 4.2. + +# 4.1. Mixture of Feature Experts + +As shown in Fig. 4, we propose Mixture of Feature Experts (MoFE), which divides the complex semantic space into multiple subspaces and each expert specializes in a specific subspace. Each expert can tackle an easier problem instead of conducting OOD detection on a complicated distribution, which eases the optimization process while maintaining the generalizability of features. Below presents the detailed configuration of MoFE. + +Given an RGB image $\mathbf{v} \in \mathbb{R}^{H \times W \times 3}$ , where $H$ and $W$ are the origin resolution, we reshape the image $\mathbf{x} \in \mathbb{R}^{H \times W \times C}$ into a sequence of flattened 2D patches $\mathbf{x}_p \in \mathbb{R}^{N \times (P^2 \cdot C)}$ , $C$ is the number of channels, $(P, P)$ is the resolution of each image patch. Next, we flatten the patches and map to $D$ dimensions with a trainable linear projection $\mathbf{E} \in \mathbb{R}^{(P^2 \cdot C) \times D}$ . A learnable embedding is pretended to the sequence of embedded patches $(\mathbf{z}_0^0 = \mathbf{x}_{\mathrm{cls}}^0)$ and position embeddings are added to the patch embeddings $E_{pos} \in \mathbb{R}^{(N + 1) \times D}$ . Then we input these embeddings to multiple transformer blocks. The output is processed by a MoFE layer to obtain the domain-specific features. This process is expressed as: + +$$ +\mathbf {z} _ {0} = \left[ \mathbf {x} _ {\mathrm {c l s}} ^ {0}; \mathbf {x} _ {p} ^ {1} \mathbf {E}; \mathbf {x} _ {p} ^ {2} \mathbf {E}; \dots ; \mathbf {x} _ {p} ^ {N} \mathbf {E} \right] + \mathbf {E} _ {p o s}, \tag {2} +$$ + +$$ +\mathbf {z} ^ {\prime} _ {\ell} = \operatorname {T r a n s f o r m e r} \left(\mathbf {z} _ {\ell - 1}\right) + \mathbf {z} _ {\ell - 1}, \ell = 1 \dots L, \tag {3} +$$ + +$$ +\mathbf {z} _ {\ell} = \operatorname {M o F E} \left(\mathrm {L N} \left(\mathbf {z} _ {\ell} ^ {\prime}\right)\right) + \mathbf {z} _ {\ell} ^ {\prime}, \tag {4} +$$ + +$$ +\mathbf {F} = \mathrm {L N} (\mathbf {z} _ {l}). \tag {5} +$$ + +where $LN$ denotes the layer norm. + +MoFE Architecture. The MoFE layer consists of multiple expert networks, each of which is a transformer block. As an + +![](images/6b9e2423c9dc87a8e500d123c6b9d8b942e60d81bb2bcc2daec7345460759837.jpg) +Figure 4. Illustration of our proposed Mixture of Feature Experts (MoFE). MoFE decomposes the large semantic space into multiple subspaces and each expert specializes in a specific subspace. Specifically, the image patches and the class token are input to obtain the preliminary patch embeddings and class embedding. A router is employed to determine the expert to further process the embeddings, and the input of the router is the class embedding. Finally, we apply associated experts to refine the class embeddings and the patch embeddings. We use the class embeddings output by MoFE and conduct the OOD detection in the corresponding subspace. + +![](images/7ea3225c529ab8b7e2a2e35a88ad2145d5e1fd49ba81c92347b53026d472b84d.jpg) + +initialization step, we replicate the transformer blocks from the final layer of a foundation model to form an ensemble of experts $\mathcal{E} = [e_1,e_2,\dots ,e_E]$ . The router [50] is a linear layer that predicts the probability of each token being assigned to each expert. Routing accuracy is crucial for MoFE. The key question is what should be used to determine the results of feature routing? We explore various approaches, such as reinitializing a routing token, averaging patch embeddings, or utilizing class embeddings. We ultimately find that using the class embedding achieves the best results. Although it is not the embedding output from the last layer of the network, it is sufficiently discriminative. Therefore, we utilize the class embedding $\mathbf{z}_l^0$ as the input of the router. The router is a linear layer that predicts the probability of each token being assigned to each expert. We formulate as: + +$$ +\mathcal {P} \left(\mathbf {z} _ {l} ^ {\prime 0}\right) _ {i} = \frac {e ^ {f \left(\mathbf {z} _ {l} ^ {\prime 0}\right) _ {i}}}{\sum_ {j} ^ {E} e ^ {f \left(\mathbf {z} _ {l} ^ {\prime 0}\right) _ {j}}}, \tag {6} +$$ + +where the router produces weight logits $f(\mathbf{z}_l^{\prime 0}) = \mathbf{W} \cdot \mathbf{z}_l^{\prime 0}$ , which are normalized by the softmax function. $\mathbf{W} \in \mathbb{R}^{D \times E}$ represents the lightweight training parameters and $E$ represents the number of experts. After determining the experts by using the class embedding, we input all embedding including the patch embeddings and class embedding to the activated experts. Each embedding is processed by the top- $k$ experts with the highest probabilities, and the weighted sum is calculated based on the softmax results of the probabilities: + +$$ +\operatorname {M o F E} \left(\mathbf {z} _ {\ell} ^ {\prime}\right) = \sum_ {i = 1} ^ {k} \mathcal {P} \left(\mathbf {z} _ {l} ^ {\prime 0}\right) _ {i} \cdot \mathcal {E} \left(\mathbf {z} _ {\ell} ^ {\prime}\right) _ {i}, \tag {7} +$$ + +where $\mathcal{E}$ represents the network of an expert [12]. In our + +MoFE architecture, we route to only a single expert, thus $k = 1$ . We find that the router computation is reduced as we are only routing a token to a single expert and the performance does not increase when using more experts. + +Feature Space Separation. In MoFE, we aim to have different experts specialize in different subspaces. Therefore, we propose to first separate the whole feature space into multiple subspaces so that each expert specializes in learning features within its subspace. We use WordNet [39], which provides a good summary of the higher-level semantics of categories, to offer an initial partition of the subspace. Since the semantic and visual similarities are not completely equivalent, we further refine the clustering using K-Means to adjust for the discrepancies. Specifically, we extract feature representations $\mathbf{z}_l^{\prime 0}$ for each training image. Then, we calculate the class prototypes by averaging the features of the images from each category. Finally, we perform a K-Means clustering on categorical feature prototypes. The initial cluster centers are determined by the centroids of clusters, which are divided according to the original semantic space. After determining the clustering results, we assign different experts to different clusters, with samples from each category being routed to the corresponding expert model. + +MoFE Training. We replace the final transformer block with a MoFE layer. Each transformer block within MoFE is initialized by the original final transformer block. We could set multiple layers as MoFE layers, but we find that using just the final layer achieves sufficiently good results. Then we randomly initialize a router layer and use the class token as the input. We use the labels generated by the above + +clustering to supervise the routing: + +$$ +\mathcal {L} _ {\text {r o u t e}} = - \frac {1}{N} \sum_ {n = 1} ^ {N} \sum_ {i = 1} ^ {E} y ^ {i} \log \left(\mathcal {P} ^ {i} \left(\mathbf {z} _ {l} ^ {\prime 0}\right)\right). \tag {8} +$$ + +For each expert, we leverage the categories within the corresponding cluster as the positive samples, and the categories beyond the cluster as the negative ones. Assuming that the category cluster of the $i$ th expert contains $Q_{i}$ classes, we set the categories beyond the cluster as the $Q_{i} + 1$ categories. The loss is designed as follows: + +$$ +\mathcal {L} _ {\text {e x p e r t}} = - \frac {1}{N} \sum_ {n = 1} ^ {N} \sum_ {i = 1} ^ {E} \sum_ {q = 1} ^ {Q _ {i} + 1} y ^ {i} y ^ {q} \log \left(p _ {i} ^ {q} (\mathbf {x})\right). \tag {9} +$$ + +In order to achieve the sample balance for each cluster, we control the ratio of positive and negative samples as 1:1 during training. Therefore, the overall loss of MoFE is: + +$$ +\mathcal {L} _ {\mathrm {M o F E}} = \mathcal {L} _ {\text {e x p e r t}} + \mathcal {L} _ {\text {r o u t e}}. \tag {10} +$$ + +Discussion. MoFE is designed to address OOD issues under large-scale complex distributions. Similar to MOS [17], we decouple complex distributions into simpler subspace. However, while MOS approaches this solely from the perspective of the loss function, we approach it from the model perspective by assigning an expert model to each subspace. This allows each expert to focus on learning its assigned subspace, preventing interference between features from different partitions. To accurately assign expert models to different samples during inference, we have devised a new routing method that uses the [CLS] token as the input to the router network, since it encapsulates the semantic features of the entire image. With these designs, MoFE outperforms MOS by a substantial margin (see Tab. 1) and the feature visualization shows that MoFE can achieve more compact ID distribution and clearer decision boundaries between ID and OOD samples (Appendix Fig. 5). + +# 4.2. Dynamic- $\beta$ Mixup + +Data augmentation (e.g., Mixup [64, 77]) has been proven to improve generalization during finetuning. Traditional Mixup [64, 77] augment samples and transform labels by: + +$$ +\tilde {x} = \lambda x _ {i} + (1 - \lambda) x _ {j}, \quad \tilde {y} = \lambda y _ {i} + (1 - \lambda) y _ {j}, \tag {11} +$$ + +where $\lambda \sim \mathrm{Beta}(\sigma, \sigma)$ . $\lambda$ is the interpolation weight for generating new augmented samples. We observe that different categories exhibit varying levels of discriminativeness initialized by vision foundation models, as shown in Appendix Fig. 7c and Fig. 7f. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup can blur the decision boundary between ID + +and OOD, leading to degraded performance (Fig. 4 in Appendix). Therefore, we dynamically adjust the Beta distribution according to the feature discriminativeness per category. The reason is that when features of $x_{i}$ are discriminative enough, a small $\lambda$ , which leads to a dissimilar sample, is not necessary for their representation learning. Instead, we should leverage similar samples from a large $\lambda$ for building smooth decision boundaries. On the contrary, when features of a category show poor discriminativeness, we should set a relatively small $\lambda$ to ease the feature learning. We use the accuracy of the validation set to measure the discriminativeness. Therefore, we set $\lambda$ as: + +$$ +\lambda \sim \operatorname {B e t a} (\sigma , \sigma) \text {f o r} \sigma = 1 - w * s, \tag {12} +$$ + +where $w$ is a scaling factor and $s$ denotes the corresponding category's accuracy on the validation set. Because the probability density function of $\mathrm{Beta}(\sigma, \sigma)$ is symmetric about 0.5 and ranges from 0 to 1, we need to ensure that with a larger $s$ , the probability of sampling larger values is greater. Therefore, we transform $\lambda$ as: + +$$ +\hat {\lambda} = \left\{ \begin{array}{l l} \lambda & \lambda \geq 0. 5 \\ 1 - \lambda & \lambda < 0. 5 \end{array} . \right. \tag {13} +$$ + +We determine the category difficulty at the beginning of the training and then update it during the training process. In our implementation, $x_{i}$ is the training sample, and $x_{j}$ is the instance used to corrupt $x_{i}$ . Therefore, we select the s from categories of $x_{i}$ , and we select samples from different classes. Additionally, we empirically find that using vanilla Mixup [64, 77] can cause feature norms to grow during finetuning vision foundation models (i.e., DINOv2), leading to performance degradation on the OOD task. In order to restrain the growth of feature norms, we propose to add a regularization term to suppress the increase in feature norm: + +$$ +\mathcal {L} _ {\text {M i x u p}} = - \frac {1}{N} \sum_ {n = 1} ^ {N} \sum_ {c = 1} ^ {C} y ^ {c} \log \left(p ^ {c} (x)\right) + R e g \left(F ^ {0}\right), \tag {14} +$$ + +where $C$ is the total number of categories, Reg denotes a regularization method, $F^0$ is the final class embeddings output by MoFE. By default, the regularization method has multiple choices, which can be $L_{2}$ norm or label smoothing. The final optimization objective is: + +$$ +\mathcal {L} _ {\text {f i n a l}} = \mathcal {L} _ {\text {M o F E}} + \mathcal {L} _ {\text {M i x u p}}. \tag {15} +$$ + +# 5. Experiments + +In this section, we set up a benchmark for evaluating OOD performance in Sec. 5.1. Then we compare our methods with the competitive baselines in Sec. 5.2. We conduct ablation studies and present more analysis on our designed method in Sec. 5.3. + +
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedEnergy [35]65.0087.1757.4087.3246.4391.1757.4087.3256.5588.2479.39
MSP [14]40.8988.6365.8181.2467.9080.1464.9678.1659.8982.0479.39
MaxLogit [15]60.8688.0355.587.4444.8191.1652.2586.0453.3588.1679.39
MCM [41]30.9194.6137.5992.5744.6989.7757.7786.1142.7490.7767.01
CLIPN [70]23.9495.2726.1793.9333.4592.2840.8390.9331.1093.1068.53
LSN [45]21.5695.8334.4891.2526.3294.3538.5490.4230.2292.9671.89
NegPrompt [30]6.3298.4327.6093.3422.8995.5535.2191.6023.0194.8166.84
Ours5.1997.2821.3294.6922.1095.1731.4792.1520.0294.8968.56
Dinov2-BasedMSP [14]25.0294.7657.0983.4553.6585.2248.7985.8148.1387.3186.01
MaxLogit [15]22.9694.5959.2178.4154.5281.8048.1784.1646.2184.7486.01
Energy [35]28.4893.1965.8874.4961.5478.7153.2981.9252.2982.0786.01
KNN [54]5.6797.6543.2588.2136.4290.2128.0492.6628.3492.1886.01
MOS [17]5.0197.8540.1590.3334.3291.8726.1492.9826.4093.2585.23
Ours2.7498.8224.3293.7317.3895.6518.5895.3817.0195.8986.40
+ +Table 1. Quantitative results of OOD detection performance for ImageNet-1k as ID. We employed our method on two pre-training paradigms (CLIP, and DINOv2). We use FPR95 and AUROC as evaluation metrics. We also report ID classification accuracy. The CLIP-based methods use ViT-B-16, and the DINOv2-based methods use ViT-B-14. + +# 5.1. Benchmark + +In- and out-distribution Datasets. To validate the effectiveness of our proposed method, we conduct evaluation on standard benchmarks, which use ImageNet-1K [48] and ImageNet-100 [41] as the ID datasets. Following existing studies that leverage foundation models in OOD [30, 70], we use diverse OOD test datasets, including samples selected from iNaturalist18 [66], SUN [74], Places [79], and Textures [6]. + +Method Comparison. We conduct method comparison on two pretaining paradigms(i.e. CLIP and DINOv2). For each group, we apply some traditional scoring metric (such as MSP [14], MaxLogit [15], Energy [35], KNN [54]). Moreover, we also involve the current CLIP-based state-of-the-art methods, such as MCM [41], CLIPN [70], NegPrompt [30], and LSN [45]. We use KNN [54] as the scoring function when using DINOv2, and follow the scoring metric of CLIPN [70] when applying our method to CLIP. + +# 5.2. Main Results + +Results on ImageNet-1K. We compare the proposed approach with the state-of-the-art methods for ImageNet-1K as ID on Tab. 1. These results show: 1) Based on DINOv2, our method reaches the best performance when setting ImageNet-1K as ID. Specifically, our approach reaches $17.01\%$ FPR95 and $95.89\%$ AUROC, averaging the results of all the OOD test sets. Our method surpasses MOS [17] by $9.39\%$ in FPR95, and $2.64\%$ in AUROC, which proves the importance of learning expert models for OOD detection. 2) When applying our method to CLIP, our method reaches $20.02\%$ and $94.89\%$ , which also outperforms NegPrompt [30] by + +a large margin. These results indicate the effectiveness of the proposed MoFE and the dynamic regularized Mixup. 3) Our approach reaches $2.74\%$ FPR95 on iNaturalist18 and increases the performance on all the test sets, which indicates that our MoFE design retains the discriminativeness of DINOv2 and facilitates feature learning on various feature subspaces. + +Results on ImageNet-100. We compare the proposed approach with the state-of-the-art methods for ImageNet-100 as ID on Tab. 2. Based on DINOv2, our method reaches $8.10\%$ FPR95 and $97.75\%$ AUROC, surpassing the baseline by $4.40\%$ FPR95, and $0.23\%$ AUROC. This indicates that our proposed approach is also effective in a small-scale ID dataset. On the other hand, when applying our method to CLIP, we achieve $7.40\%$ FPR95 and $98.10\%$ AUROC, outperforming LSN [45] by $1.16\%$ FPR95. The above experimental results validate the effectiveness of our approach, and we achieve the best performance on both small-scale and large-scale ID datasets. + +# 5.3. Analysis + +In this section, we conduct more analysis on the proposed methods. We use the DINOv2 as the pretaining paradigm and KNN [54] as the scoring function. We use ImageNet-1K as ID data and report the average performance on the four out-of-distribution datasets mentioned in Sec. 5.1. The baseline is set to DINOv2 with KNN scoring function without our designed method. + +Contributions of Individual Components. As shown in Tab. 3, we evaluate the contribution of each component of the full method. On ImageNet-1K, MoFE and Dynamic- $\beta$ Mixup contribute $6.68\%$ and $5.41\%$ FPR95, respectively. + +
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedMSP [14]23.5595.9240.4691.2337.0292.4524.4094.9031.4393.6391.93
MCM [41]18.1396.7734.5294.3636.4594.5441.2292.2532.5894.4887.88
CLIPN [70]4.8798.1613.6496.9313.5597.5615.7893.0211.9696.4191.64
LSN [45]4.9398.9212.8297.198.2397.988.2698.118.5698.0592.24
Ours3.2099.1710.0597.767.0698.399.3197.107.4098.1092.85
Dinov2-BasedMSP [14]5.0698.8526.5894.7827.6495.0226.4394.2721.4295.7294.50
MaxLogit [15]5.5598.7629.6994.1932.7394.2029.2793.7224.3195.2194.50
Energy [35]18.5796.6954.7288.9262.4287.1757.2888.4048.2190.2994.50
KNN [54]2.5899.0218.4595.1215.8996.1616.7996.3813.4296.6694.50
Ours2.2599.2312.8196.668.5197.868.8597.288.1097.7596.94
+ +Table 2. Quantitative results of OOD detection performance for ImageNet-100 as ID. + +
SettingsBaseline+ MoFE+ D-β+ MoFE+D-β
FPR95↓29.2722.5923.8517.01
AUROC↑92.6794.0193.7295.89
+ +Table 3. Ablation study of individual components for ImageNet-1k as In-Distribution dataset. + +
Num.235789
Gain4.106.309.8012.2612.1012.09
+ +Table 4. The effect of Cluster Number. We report the performance gain in FPR95 compared to the model without MoFE. + +
GroupingBaselineWordNetClusteringOurs
FPR95↓29.2725.6326.3422.59
AUROC↑92.6793.0192.9994.01
+ +When combined, the best performance (17.01% FPR95, and 95.89% AUROC) is achieved. + +Cluster Number. We conduct an experiment to validate the impact of cluster number on MoFE performance. We set different numbers of clusters. As shown in Tab. 4, we report the performance gain in FPR95. As the increasing of cluster number, the performance gradually increases. + +Feature Space Separation. As shown in Tab. 5, we validate the different strategies for determining the subspace. We compare our method with two methods: WordNet and clustering. The results show that using the ours (i.e. WordNet + Clustering) is the most promising approach. The reason might be that the features extracted by pretrained model are discriminative, especially at coarse-grained level. + +More Analysis on Dynamic- $\beta$ Mixup. As shown in Tab. 6, we conduct an ablation study on Dynamic- $\beta$ Mixup. When we remove the regularization term, we find that the performance degrades (30.43% FPR95). Moreover, when we dynamic beta distribution is removed, the performance decreases to 24.96% FPR95. + +Table 5. Analysis on Grouping Strategy in MoFE. + +
MethodsBaselinew/o Regw/o D-βOurs
FPR95↓29.2730.4324.9623.85
AUROC↑92.6791.6593.3693.72
+ +Table 6. Ablation study of Dynamic- $\beta$ Mixup. + +# 6. Conclusion + +This paper studies the OOD detection task within the context of foundation models. Our study shows that vision foundation models (e.g., DINOv2) are effective OOD detectors, suggesting high-quality and generalizable feature space is essential for OOD detection. Our study highlights that CLIP's pre-trained feature space is less effective for fine-grained tasks, where DINOv2 performs significantly better, which worths further exploration. In order to further optimize the performance of OOD detection when ID data is available for fine-tuning, we propose MoFE and a Dynamic- $\beta$ Mixup data augmentation to enhance the feature learning. We conduct extensive experiments and ablation studies to validate the effectiveness of our approach. We hope our investigation could inspire more future studies. + +# Acknowledgments + +This research is supported by the National Key RD Program of China (Grant No. 2023YFB2806300). This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 172024221721292317215025), Theme-based Research (Grant No. T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust. + +# References + +[1] Sparse mixture-of-experts are domain generalizable learners, author=Li, Bo and Shen, Yifei and Yang, Jingkang and Wang, Yezhen and Ren, Jiawei and Che, Tong and Zhang, Jun and Liu, Ziwei, booktitle=ICLR, year=2023. 1 +[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3, 4 +[3] Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024. 2 +[4] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV, 2020. 4 +[5] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160, 2024.3 +[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 7 +[7] Hendrycks Dan, Zhao Kevin, Basart Steven, Steinhardt Jacob, and Song Dawn. Natural adversarial examples. In CVPR, 2021. 3 +[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1 +[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2 +[10] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In NeurIPS, 2023. 1 +[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, 2022. 1, 2, 3 +[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, 2021. 2, 3, 5 +[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022. 3 +[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 3, 7, 8 +[15] Dan Hendrycks, Steven Basart, Mantas Mazeika, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 3, 4, 7, 8 +[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image with + +out learning from out-of-distribution data. In CVPR, 2020. +[17] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In CVPR, 2021. 2, 6, 7, 3 +[18] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3 +[19] Bitterwolf Julian, Mueller Maximilian, and Matthias. Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, 2023. 3 +[20] Umar Khalid, Ashkan Esmaeili, Nazmul Karim, and Nazanin Rahnavard. Rodd: A self-supervised approach for robust out-of-distribution detection. In CVPRW, 2022. 2 +[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1 +[22] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. arXiv preprint arXiv:2212.05055, 2022. 3 +[23] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 1, 3 +[24] Yamuna Krishnamurthy, Chris Watkins, and Thomas Gaertner. Improving expert specialization in mixture of experts. arXiv preprint arXiv:2302.14703, 2023. 3 +[25] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1 +[26] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2 +[27] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 3 +[28] Jinglun Li, Xinyu Zhou, Pinxue Guo, Yixuan Sun, Yiwen Huang, Weifeng Ge, and Wenqiang Zhang. Hierarchical visual categories modeling: A joint representation learning and density estimation framework for out-of-distribution detection. In ICCV, 2023. 2 +[29] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1 + +[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, 2024. 4, 7 +[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1 +[32] Randolph Linderman, Jingyang Zhang, Nathan Inkawich, Hai Li, and Yiran Chen. Fine-grain inference on out-of-distribution data with hierarchical classification. arXiv preprint arXiv:2209.04493, 2022. 2 +[33] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 2 +[34] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 2 +[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 3, 7, 8 +[36] Yuxuan Lou, Fuzhao Xue, Zangwei Zheng, and Yang You. Cross-token modeling with conditional computation. arXiv preprint arXiv:2109.02008, 2021. 3 +[37] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 2 +[38] Oquab Maxime, Darcet Timothee, Moutakanni Théo, Vo Huy, Szafraniec Marc, Khalidov Vasil, Fernandez Pierre, Haziza Daniel, Massa Francisco, El-Nouby Alaaeldin, Assran Mahmoud, Ballas Nicolas, Galuba Wojciech, Howes Russell, Huang Po-Yao, Li Shang-Wen, Misra Ishan, Rabbat Michael, Sharma Vasu, Synnaeve Gabriel, Xu Hu, Jegou Herve, Mairal Julien, Labatut Patrick, Joulin Armand, and Bojanowski Piotr. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 1, 2, 3 +[39] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994", 1994. 5 +[40] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? IJCV, 2024. 1, 2, 3 +[41] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 7, 8, 2 +[42] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NIPS, 2023. 2 +[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. arXiv preprint arXiv:2206.02770, 2022. 3 + +[44] Ibrahim Ndiour, Nilesh Ahuja, and Omesh Tickoo. Out-of-distribution detection with subspace techniques and probabilistic modeling of features. arXiv preprint arXiv:2012.04250, 2020. 2 +[45] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 7, 8 +[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3 +[47] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. NeurIPS, 2021. 3, 1 +[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7 +[49] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In ICLR, 2021. 1, 2 +[50] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3, 5 +[51] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, Rohit Girdhar, and Ishan Misra. The effectiveness of mae pre-pretraining for billion-scale pretraining. In ICCV, 2023. 1 +[52] Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In ECCV, 2022. 1 +[53] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 1 +[54] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 1, 2, 3, 7, 8 +[55] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 4 +[56] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 2 +[57] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023. 2 +[58] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023. + +[59] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024. +[60] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. +[61] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025. 2 +[62] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 2 +[63] Leitian Tao, Xuefeng Du, Xiaojin Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 1 +[64] Sunil Thulasidasan, Gopinath Chennupati, Jeff Bilmes, Tan-moy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurIPS, 2019. 2, 6 +[65] Changyao Tian, Wenhai Wang, Xizhou Zhu, Xiaogang Wang, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. arXiv preprint arXiv:2111.13579, 2021. 1 +[66] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 7 +[67] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: a good closed-set classifier is all you need? In ICLR, 2022. 4 +[68] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, 2022. 1 +[69] Haotao Wang, Aston Zhang, Yi Zhu, Shuai Zheng, Mu Li, Alex J Smola, and Zhangyang Wang. Partial and asymmetric contrastive learning for out-of-distribution detection in long-tailed recognition. In ICML, 2022. 2 +[70] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In ICCV, 2023. 1, 2, 3, 7, 8 +[71] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2 +[72] Sitong Wu, Tianyi Wu, Haoru Tan, and Guodong Guo. Pale transformer: A general vision transformer backbone with pale-shaped attention. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2022. 2 +[73] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 3 +[74] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 7 + +[75] Zhai Xiaohua, Mustafa Basil, Kolesnikov Alexander, and Beyer Lucas. Sigmoid loss for language image pre-training, 2023. 3 +[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1 +[77] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2017. 6 +[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 1 +[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. PAMI, 2017. 7 +[80] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 4 \ No newline at end of file diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/images.zip b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ea1767d50aa2a0a4a548086d54df6fa55b79e8a1 --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:62a06c36f230b2d7c491e3887541bbd015ffe5c98422587384cf8201126b54a8 +size 513637 diff --git a/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/layout.json b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b5600712c01d1b81e84246b4ac4de5de1d7c8093 --- /dev/null +++ b/2025/Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection/layout.json @@ -0,0 +1,9320 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 106, + 103, + 505, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 103, + 505, + 138 + ], + "spans": [ + { + "bbox": [ + 106, + 103, + 505, + 138 + ], + "type": "text", + "content": "Equipping Vision Foundation Model with Mixture of Experts for Out-of-Distribution Detection" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 161, + 468, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 161, + 468, + 190 + ], + "spans": [ + { + "bbox": [ + 141, + 161, + 468, + 190 + ], + "type": "text", + "content": "Shizhen Zhao Jiahui Liu Xin Wen Haoru Tan Xiaojuan Qi† The University of Hong Kong" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "spans": [ + { + "bbox": [ + 151, + 217, + 200, + 229 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 243, + 297, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 297, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 297, + 590 + ], + "type": "text", + "content": "Pre-trained vision foundation models have transformed many computer vision tasks. Despite their strong ability to learn discriminative and generalizable features crucial for out-of-distribution (OOD) detection, their impact on this task remains underexplored. Motivated by this gap, we systematically investigate representative vision foundation models for OOD detection. Our findings reveal that a pre-trained DINOv2 model, even without fine-tuning on in-domain (ID) data, naturally provides a highly discriminative feature space for OOD detection, achieving performance comparable to existing state-of-the-art methods without requiring complex designs. Beyond this, we explore how fine-tuning foundation models on in-domain (ID) data can enhance OOD detection. However, we observe that the performance of vision foundation models remains unsatisfactory in scenarios with a large semantic space. This is due to the increased complexity of decision boundaries as the number of categories grows, which complicates the optimization process. To mitigate this, we propose the Mixture of Feature Experts (MoFE) module, which partitions features into subspaces, effectively capturing complex data distributions and refining decision boundaries. Further, we introduce a Dynamic- " + }, + { + "bbox": [ + 55, + 243, + 297, + 590 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 243, + 297, + 590 + ], + "type": "text", + "content": " Mixup strategy, which samples interpolation weights from a dynamic beta distribution. This adapts to varying levels of learning difficulty across categories, improving feature learning for more challenging categories. Extensive experiments demonstrate the effectiveness of our approach, significantly outperforming baseline methods. The project will be available at shizhen-zhao.github.io/OOD_MoFE/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 135, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 634, + 296, + 694 + ], + "type": "text", + "content": "The task of out-of-distribution (OOD) detection [16, 26, 31, 49] aims to equip models with the capability to discern whether input images originate from unknown OOD classes or belong to in-domain (ID) classes. Mainstream OOD detection methods [9, 10, 25, 63] focus on learning features" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 217, + 545, + 342 + ], + "blocks": [ + { + "bbox": [ + 321, + 217, + 545, + 342 + ], + "lines": [ + { + "bbox": [ + 321, + 217, + 545, + 342 + ], + "spans": [ + { + "bbox": [ + 321, + 217, + 545, + 342 + ], + "type": "image", + "image_path": "51a4b5db034172c05cb8f0c79ae340ecf6d91f9b0131e1efcf1f0ed6c2273953.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 321, + 352, + 545, + 464 + ], + "blocks": [ + { + "bbox": [ + 321, + 352, + 545, + 464 + ], + "lines": [ + { + "bbox": [ + 321, + 352, + 545, + 464 + ], + "spans": [ + { + "bbox": [ + 321, + 352, + 545, + 464 + ], + "type": "image", + "image_path": "f220fba940cfbbb11dfef269d9272769b13ae241b780d1320ef0c5a2502abf37.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 473, + 555, + 529 + ], + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 529 + ], + "type": "text", + "content": "Figure 1. Holistic comparison to previous philosophy. (a) Traditional methods use a generalized model to project inputs onto a complex distribution; (b): Our approach leverages multiple experts to break the complex distribution into smaller ones, which leads to compact ID distribution and simplified decision boundary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 556, + 699 + ], + "type": "text", + "content": "and classifiers [31, 52, 53, 68] from ID data and then develop a score metric [14, 15, 35, 54] to determine whether a sample belongs to ID or OOD classes. Despite significant advancements, the fundamental challenge in OOD detection is establishing a feature space with high discriminative capacity that can effectively distinguish OOD samples from ID samples. Recently, vision foundation models [21, 38, 46, 51] trained on large-scale datasets have demonstrated the ability to learn robust and generalizable features, benefiting numerous tasks [29, 65, 76, 78]. This raises the question: with such powerful models and feature representations, does OOD detection remain a problem?" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "Although several studies [11, 40, 41, 70] have explored" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1751" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 228 + ], + "type": "text", + "content": "the use of foundation models for OOD detection, most focus on improving the performance of vision-language models like CLIP [46], while other foundation models, such as DINOv2, remain largely unexamined. In this study, we systematically investigate the feature spaces of different representative pre-trained foundation models, including vision-language models (e.g., CLIP) and self-supervised models (e.g., DINOv2), in the context of OOD detection. Our results reveal that DINOv2 provides the most discriminative feature space, enabling effective OOD detection without any finetuning. Notably, using a simple KNN metric [54], DINOv2 achieves performance comparable to more complex methods, establishing a strong baseline for further research." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 229, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 229, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 229, + 296, + 445 + ], + "type": "text", + "content": "While vision foundation models [38, 45, 46, 70] have achieved impressive performance in OOD detection, there is still room for improvement, particularly on in-domain data with large semantic spaces [57-61, 71] (e.g., " + }, + { + "bbox": [ + 56, + 229, + 296, + 445 + ], + "type": "inline_equation", + "content": "29.27\\%" + }, + { + "bbox": [ + 56, + 229, + 296, + 445 + ], + "type": "text", + "content": " FPR95 on the ImageNet-1K OOD benchmark [54]). This prompts us to investigate whether foundation models can be further optimized by leveraging available ID data. However, as the number of semantic classes increases, the complexity of the decision boundaries required to distinguish between ID and OOD data grows as well [17, 28, 32]. This heightened complexity creates challenges when fine-tuning foundation models on limited ID data. Previous methods (e.g. MOS [17]) decouple the complex space into simpler subspaces from the perspective of loss, which eases the optimization process and simplifies the decision boundaries. In this study, we tackle the problem orthogonally from the model perspective by designing a new Mixture of Experts (MoE) architecture to more thoroughly disentangle complex ID distribution." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 448, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 297, + 687 + ], + "type": "text", + "content": "To address this issue, rather than directly optimizing the whole feature space [3, 33, 34, 37] (Fig. 1(a)), we propose a Mixture-of-Feature-Expert (MoFE) module, which utilizes multiple experts, and each expert specializes in a specific subspace and optimizes it accordingly (Fig. 1(b)). MoFE operates by partitioning the original feature space into " + }, + { + "bbox": [ + 55, + 448, + 297, + 687 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 448, + 297, + 687 + ], + "type": "text", + "content": " subspaces based on semantics and feature similarities within the ID dataset. Each subspace is assigned to a dedicated expert, and a router assigns samples to the appropriate expert based on these partitions. Different from previous studies [12, 50], we use the [CLS] token as the input to the router network, since it encapsulates the semantic feature of the whole image. During training, the router is supervised by the partition assignments to ensure accurate sample-to-expert mapping. Each expert focuses solely on optimizing features within its designated partition, which helps prevent interference between features from different partitions. The results show that our approach significantly surpasses the previous approaches by a large margin (see Tab. 1), revealing the importance of learning expert models for OOD tasks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Additionally, given that data augmentation has been shown to enhance generalization for OOD, we introduce" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 556, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 251 + ], + "type": "text", + "content": "a novel Mixup data augmentation strategy to further improve feature learning, which is better suited for advanced vision foundation models. Our design is based on the observation that different categories exhibit varying levels of discriminativeness with features from vision foundation models. In the original feature space, some categories show high discriminativeness, while others do not. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup [64] can blur the decision boundary between ID and OOD, leading to degraded performance (Appendix Fig. 6). Thus, unlike existing Mixup strategies that treat all categories equally [56, 62, 72], our approach makes Mixup weight sampling category-dependent by adjusting the sampling distribution (i.e. beta distribution) dynamically, taking into account their discriminativeness." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 555, + 262 + ], + "type": "text", + "content": "Our major contributions can be summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 264, + 555, + 430 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 264, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 264, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 264, + 555, + 311 + ], + "type": "text", + "content": "- We design a novel MoFE module to tailor pre-trained vision foundation models for OOD detection. This approach reduces the difficulty of fitting complex data distributions from limited data and eases the optimization process." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 312, + 555, + 382 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 382 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 382 + ], + "type": "text", + "content": "- We explore the effectiveness of the raw feature spaces from various vision foundation models for OOD detection. Through analysis, we leverage DINOv2 with simple scoring metric to establish a strong baseline. Additionally, we designed a Dynamic-" + }, + { + "bbox": [ + 313, + 312, + 555, + 382 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 312, + 555, + 382 + ], + "type": "text", + "content": " Mixup that is better suited for advanced vision foundation models." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 383, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 430 + ], + "type": "text", + "content": "- Our extensive experimental results demonstrate the effectiveness of the proposed model, achieving significant improvements over several competitive baseline methods on standard benchmarks." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 441, + 402, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 402, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 402, + 453 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 665 + ], + "type": "text", + "content": "Out-of-Distribution Detection The goal of OOD detection is to detect OOD images from the test dataset (containing both ID and OOD images). Designing the score function is the most popular method in OOD detection tasks. The scores are mainly derived from three sources: the probability [14, 15], the logits [15, 35], and the feature [26, 44]. Some studies [20, 49, 69] focus on leveraging contrastive learning to enhance the feature representation. Other studies show that synthesizing pseudo samples [9, 49, 55, 55] as OOD instances is also a promising approach to make the feature space more compact. The methods [17, 28, 32] are the most relevant to ours, which also break the semantic space into smaller ones. Different from these approaches, in our design, we propose a novel MoE module, with each expert exclusively concentrating on optimizing features within its specific partition. Our results demonstrate that our approach outperforms them by a substantial margin." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 557, + 715 + ], + "type": "text", + "content": "OOD Detection with Foundation Models There are some existing OOD detection methods [11, 40-42, 45, 70] leveraging foundation models. Maximum Concept Matching" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1752" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "(MCM) [41] proposes a simple yet effective zero-shot OOD detection method by aligning visual features with textual concepts. Some other studies [45, 70] explore negative prompts to learn the diversity of negative features, enabling more accurate detection of OOD samples. Although these studies have made great progress by leveraging CLIP to enhance the performance in existing benchmarks, they only explore and fine-tune CLIP. In our studies, we explore different foundation models and explore a better fine-tuning paradigm." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 296, + 356 + ], + "type": "text", + "content": "Mixture of Experts Mixture of Experts has been studied independently in both computer vision [36, 43, 47, 73] and natural language processing [12, 22, 27, 50]. These works are studied in the context of conditional computation, which is to increase the number of model parameters without a proportional increase in computational cost. Currently, some studies [5, 24] explore improving expert specialization and leveraging MoE to mitigate data conflict problems, where some data might interfere with each other. In our study, we introduce MoFE to the out-of-distribution task in the context of foundation models and build specialized OOD detectors for different feature subspaces." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 374, + 128, + 388 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 128, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 128, + 388 + ], + "type": "text", + "content": "3. Pilot Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 446 + ], + "type": "text", + "content": "In this section, we first introduce preliminaries for the OOD detection task in Sec. 3.1. Then, we explore the impact of foundation models on OOD detection performance and analyze their strengths and weaknesses in Sec. 3.2." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 463, + 141, + 474 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 141, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 141, + 474 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": "We consider supervised multi-class classification, where " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " represents the input image space and " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{Y} = \\{1,2,\\dots,C\\}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " represents the label space. The training dataset " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathbb{D}_{in} = \\{(\\mathbf{x}_i,y_i)\\}_{i=1}^n" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " is drawn independently and identically distributed (i.i.d.) from the joint data distribution " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{X}\\mathcal{Y}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{in}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " denote the marginal distribution on " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "\\mathcal{X}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "f:\\mathcal{X}\\mapsto \\mathbb{R}^{|\\mathcal{Y}|}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " be a neural network trained on samples drawn from " + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "inline_equation", + "content": "P_{\\mathcal{X}\\mathcal{Y}}" + }, + { + "bbox": [ + 55, + 483, + 296, + 590 + ], + "type": "text", + "content": " to output a logit vector, which is used to predict the label of the input sample." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Out-of-distribution Detection. When deploying a machine learning model in real-world scenarios, it is crucial for a reliable classifier not only to accurately classify known distribution (ID) samples, but also to recognize any out-of-distribution (OOD) inputs as \"unknown\". This can be accomplished by incorporating an OOD detector alongside the classification model " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": ". OOD can be formulated as a binary classification task. During testing, the objective is to determine whether a sample " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathcal{X}" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " belongs to " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_{\\mathrm{in}}(\\mathrm{ID})" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " or not (OOD). This decision can be made using a scoring" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 72, + 369, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 369, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 369, + 84 + ], + "type": "text", + "content": "metric " + }, + { + "bbox": [ + 314, + 72, + 369, + 84 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x})" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 370, + 92, + 555, + 125 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 92, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 370, + 92, + 555, + 125 + ], + "type": "interline_equation", + "content": "G _ {\\lambda} (x) = \\left\\{ \\begin{array}{l l} \\mathrm {I D} & S (\\mathbf {x}) \\geq \\lambda \\\\ \\text {O O D} & S (\\mathbf {x}) < \\lambda \\end{array} , \\right. \\tag {1}", + "image_path": "bcec4afb6bac2a28a161278a73619dd58ee6637be4136d360a77e94b759c9af2.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": "where samples with higher scores " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " are classified as ID and vice versa, and " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " is the threshold. Some typically used metrics " + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "inline_equation", + "content": "S(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 133, + 555, + 181 + ], + "type": "text", + "content": " include MSP [14], MaxLogit [15], Energy [35] and KNN [54]." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 191, + 541, + 336 + ], + "blocks": [ + { + "bbox": [ + 326, + 191, + 541, + 336 + ], + "lines": [ + { + "bbox": [ + 326, + 191, + 541, + 336 + ], + "spans": [ + { + "bbox": [ + 326, + 191, + 541, + 336 + ], + "type": "image", + "image_path": "a3be85820c442e52626f8d7e1a7769f131ca9bbdc9f0f2f42310cb5ad4971889.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 555, + 376 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 555, + 376 + ], + "type": "text", + "content": "Figure 2. Performance of vision foundation models across different OOD splits. The evaluation metric is FPR95, with lower values indicating better performance." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 385, + 526, + 397 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 385, + 526, + 397 + ], + "spans": [ + { + "bbox": [ + 314, + 385, + 526, + 397 + ], + "type": "text", + "content": "3.2. Evaluation of Vision Foundation Models" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 521 + ], + "type": "text", + "content": "Although several studies [11, 40, 41, 70] have explored the use of foundation models for OOD detection, they focus solely on vision-language foundation models such as CLIP [46]. Beyond CLIP, the community offers a variety of vision foundation models that provide robust raw feature space. This development has inspired us to re-examine which vision foundation model is best suited for OOD detection. In this section, we aim to investigate and analyze various pre-trained vision foundation models as effective OOD detectors without fine-tuning." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 714 + ], + "type": "text", + "content": "Experimental Setup. We perform our evaluation on a challenging OOD detection benchmark that utilizes ImageNet-1K as ID data and selects samples from iNaturalist18, Sun, Places, and Textures as OOD samples. We also include two challenging OOD test sets: ImageNet-O [7] and NINCO [19]. We choose several representative vision foundation models, namely DINOv1 [2], DINOv2 [38], MAE [13], SigCLIP [75], OpenCLIP [18], and OpenAI CLIP [46]. For fair comparison, we use the ViT-B as the architecture of these models. The scoring functions for DINOv1, DINOv2, and MAE are set to KNN [54]. For the CLIP series, we report the best results among four scoring functions (MSP, MaxLogit, Energy and KNN). Without any model tuning, we directly use the features extracted from these models for OOD detection evaluation to assess whether they are already sufficiently capable of OOD detection. To emphasize the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1753" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 84, + 170, + 190 + ], + "blocks": [ + { + "bbox": [ + 65, + 84, + 170, + 190 + ], + "lines": [ + { + "bbox": [ + 65, + 84, + 170, + 190 + ], + "spans": [ + { + "bbox": [ + 65, + 84, + 170, + 190 + ], + "type": "image", + "image_path": "6854e2425378abcd7a908b1d1bfd772c286727043d21e0781bb31e9e4a923111.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 198, + 134, + 207 + ], + "lines": [ + { + "bbox": [ + 102, + 198, + 134, + 207 + ], + "spans": [ + { + "bbox": [ + 102, + 198, + 134, + 207 + ], + "type": "text", + "content": "(a) CLIP" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 251 + ], + "type": "text", + "content": "Figure 3. Feature Visualization for Foundation Models. For fine-grained feature visualization, we randomly select fine-grained categories under 3 different super classes from ImageNet-1k." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 186, + 84, + 291, + 190 + ], + "blocks": [ + { + "bbox": [ + 186, + 84, + 291, + 190 + ], + "lines": [ + { + "bbox": [ + 186, + 84, + 291, + 190 + ], + "spans": [ + { + "bbox": [ + 186, + 84, + 291, + 190 + ], + "type": "image", + "image_path": "621774629dafb82cfebe68739b3b61a219516034b5139d419a5ea6a24713a897.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 198, + 260, + 207 + ], + "lines": [ + { + "bbox": [ + 219, + 198, + 260, + 207 + ], + "spans": [ + { + "bbox": [ + 219, + 198, + 260, + 207 + ], + "type": "text", + "content": "(b) DINov2" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 277, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 296, + 350 + ], + "type": "text", + "content": "significance of our findings, we also compare them with the state-of-the-art method (NegPrompt [30]) that involve fine-tuning an ImageNet pre-trained model on the ID dataset. Additionally, we conduct further verification using datasets beyond ImageNet-1k as ID data in Appendix A.4 and the conclusion align with the experiments using ImageNet-1k." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": "Result Analysis. (1) With traditional score metrics (i.e KNN), DINOv2 can outperform all other foundation models by a large margin. DINOv2+KNN shows the best results where the average FPR95 is " + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "inline_equation", + "content": "29.27\\%" + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": " in the first four test sets, " + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "inline_equation", + "content": "8.9\\%" + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": " in the latter challenging test sets, while SigCLIP only achieves " + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "inline_equation", + "content": "54.23\\%" + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "inline_equation", + "content": "77.17\\%" + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": ". This is potentially because DINOv2 leverages advanced self-supervised learning: iBot [80], which is a Mask Image Modeling (MIM) pretask for facilitating models to capture image details, and contrastive learning objective [2] that enhances the feature discriminativeness. (2) Without any fine-tuning, DINOv2 achieves performance comparable to the more complex method (i.e. NegPrompt) in the first four test sets. Notably, DINOv2+KNN still significantly outperforms NegPrompt on challenging test sets by " + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "inline_equation", + "content": "49.94\\%" + }, + { + "bbox": [ + 56, + 351, + 296, + 639 + ], + "type": "text", + "content": ". The reason is that these two datasets contain images that are extremely similar to the ID categories. However, the paradigm of CLIP only provides image-level textual supervision without a supervision signal to retain detailed image information. Therefore, CLIP always fails in some fine-grained tasks, while DINOv2 consistently performs much better. As shown in Fig. 3a and Fig. 3b, where we randomly select 11 fine-grain categories under 3 different super classes, DINOv2 provide more discriminative boundaries, while CLIP can not." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 714 + ], + "type": "text", + "content": "Further Challenges in OOD using Foundation models. In summary, DINov2, without requiring any fine-tuning, can already function as a high-performing OOD detector, surpassing previous approaches and underscoring the importance of discriminative and generalizable features for OOD detection. However, foundation models still have room to improve" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 240 + ], + "type": "text", + "content": "and cannot generalize well across the entire feature space. (1) Though there is a consensus that fine-tuning on the ID data can improve OOD performance [4, 15, 55, 67], we find that this doesn't hold in the context of foundation models, particularly on in-domain data with large semantic spaces. For instance, when we fine-tune DINOv2 on ImageNet-1K ID data and evaluate the fine-tuned model, the performance declines on three out of the four OOD datasets (see naive finetuning in Tab. 7 in Appendix). The implementation details of this finetuning can be referred to Appendix A.2. (2) Besides, as shown in Fig. 7c and Fig. 7f in Appendix, we also show some failure examples, where the models exhibit particularly poor feature discriminability, hindering effective OOD detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 251, + 370, + 263 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 251, + 370, + 263 + ], + "spans": [ + { + "bbox": [ + 314, + 251, + 370, + 263 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "text", + "content": "This section introduces our proposed methods for finetuning vision foundation models to enhance the OOD detection ability, which includes a Mixture of Feature Expert module in Sec. 4.1 and a Dynamic-" + }, + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 271, + 555, + 331 + ], + "type": "text", + "content": " Mixup data augmentation strategy in Sec. 4.2." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 339, + 466, + 353 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 466, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 466, + 353 + ], + "type": "text", + "content": "4.1. Mixture of Feature Experts" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 452 + ], + "type": "text", + "content": "As shown in Fig. 4, we propose Mixture of Feature Experts (MoFE), which divides the complex semantic space into multiple subspaces and each expert specializes in a specific subspace. Each expert can tackle an easier problem instead of conducting OOD detection on a complicated distribution, which eases the optimization process while maintaining the generalizability of features. Below presents the detailed configuration of MoFE." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": "Given an RGB image " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{v} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " are the origin resolution, we reshape the image " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x} \\in \\mathbb{R}^{H \\times W \\times C}" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " into a sequence of flattened 2D patches " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_p \\in \\mathbb{R}^{N \\times (P^2 \\cdot C)}" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " is the number of channels, " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "(P, P)" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " is the resolution of each image patch. Next, we flatten the patches and map to " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " dimensions with a trainable linear projection " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "\\mathbf{E} \\in \\mathbb{R}^{(P^2 \\cdot C) \\times D}" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": ". A learnable embedding is pretended to the sequence of embedded patches " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "(\\mathbf{z}_0^0 = \\mathbf{x}_{\\mathrm{cls}}^0)" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": " and position embeddings are added to the patch embeddings " + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "inline_equation", + "content": "E_{pos} \\in \\mathbb{R}^{(N + 1) \\times D}" + }, + { + "bbox": [ + 313, + 452, + 556, + 598 + ], + "type": "text", + "content": ". Then we input these embeddings to multiple transformer blocks. The output is processed by a MoFE layer to obtain the domain-specific features. This process is expressed as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 341, + 607, + 555, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 607, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 341, + 607, + 555, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {0} = \\left[ \\mathbf {x} _ {\\mathrm {c l s}} ^ {0}; \\mathbf {x} _ {p} ^ {1} \\mathbf {E}; \\mathbf {x} _ {p} ^ {2} \\mathbf {E}; \\dots ; \\mathbf {x} _ {p} ^ {N} \\mathbf {E} \\right] + \\mathbf {E} _ {p o s}, \\tag {2}", + "image_path": "3a5d65a5b089023a18fbf944505e7c3bbbdb967276d4d1faeb84e06f88d30d42.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 341, + 624, + 555, + 637 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 624, + 555, + 637 + ], + "spans": [ + { + "bbox": [ + 341, + 624, + 555, + 637 + ], + "type": "interline_equation", + "content": "\\mathbf {z} ^ {\\prime} _ {\\ell} = \\operatorname {T r a n s f o r m e r} \\left(\\mathbf {z} _ {\\ell - 1}\\right) + \\mathbf {z} _ {\\ell - 1}, \\ell = 1 \\dots L, \\tag {3}", + "image_path": "9c6594c6d3b3e9cd8b71e4aa6b5ba762e4863572e1c9a5a931a1378476abe974.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 343, + 639, + 555, + 653 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 639, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 343, + 639, + 555, + 653 + ], + "type": "interline_equation", + "content": "\\mathbf {z} _ {\\ell} = \\operatorname {M o F E} \\left(\\mathrm {L N} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right)\\right) + \\mathbf {z} _ {\\ell} ^ {\\prime}, \\tag {4}", + "image_path": "3eb862a6fa929497f057c2abdd44f882776a0615bac2984ab906a5625d6ee5bd.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 345, + 654, + 555, + 667 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 654, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 345, + 654, + 555, + 667 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\mathrm {L N} (\\mathbf {z} _ {l}). \\tag {5}", + "image_path": "49576669378073bf31d5f3827a1571ebc17468b9988a45106360f2a815894b1e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 677, + 455, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 455, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 455, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 455, + 689 + ], + "type": "inline_equation", + "content": "LN" + }, + { + "bbox": [ + 313, + 677, + 455, + 689 + ], + "type": "text", + "content": " denotes the layer norm." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "MoFE Architecture. The MoFE layer consists of multiple expert networks, each of which is a transformer block. As an" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1754" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 286, + 245 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 286, + 245 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 286, + 245 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 286, + 245 + ], + "type": "image", + "image_path": "6b9e2423c9dc87a8e500d123c6b9d8b942e60d81bb2bcc2daec7345460759837.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "lines": [ + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "spans": [ + { + "bbox": [ + 54, + 261, + 555, + 318 + ], + "type": "text", + "content": "Figure 4. Illustration of our proposed Mixture of Feature Experts (MoFE). MoFE decomposes the large semantic space into multiple subspaces and each expert specializes in a specific subspace. Specifically, the image patches and the class token are input to obtain the preliminary patch embeddings and class embedding. A router is employed to determine the expert to further process the embeddings, and the input of the router is the class embedding. Finally, we apply associated experts to refine the class embeddings and the patch embeddings. We use the class embeddings output by MoFE and conduct the OOD detection in the corresponding subspace." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 311, + 74, + 553, + 250 + ], + "blocks": [ + { + "bbox": [ + 311, + 74, + 553, + 250 + ], + "lines": [ + { + "bbox": [ + 311, + 74, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 311, + 74, + 553, + 250 + ], + "type": "image", + "image_path": "7ea3225c529ab8b7e2a2e35a88ad2145d5e1fd49ba81c92347b53026d472b84d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "text", + "content": "initialization step, we replicate the transformer blocks from the final layer of a foundation model to form an ensemble of experts " + }, + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = [e_1,e_2,\\dots ,e_E]" + }, + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "text", + "content": ". The router [50] is a linear layer that predicts the probability of each token being assigned to each expert. Routing accuracy is crucial for MoFE. The key question is what should be used to determine the results of feature routing? We explore various approaches, such as reinitializing a routing token, averaging patch embeddings, or utilizing class embeddings. We ultimately find that using the class embedding achieves the best results. Although it is not the embedding output from the last layer of the network, it is sufficiently discriminative. Therefore, we utilize the class embedding " + }, + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_l^0" + }, + { + "bbox": [ + 54, + 327, + 297, + 506 + ], + "type": "text", + "content": " as the input of the router. The router is a linear layer that predicts the probability of each token being assigned to each expert. We formulate as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 510, + 296, + 544 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 510, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 123, + 510, + 296, + 544 + ], + "type": "interline_equation", + "content": "\\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} = \\frac {e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i}}}{\\sum_ {j} ^ {E} e ^ {f \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {j}}}, \\tag {6}", + "image_path": "b67273b9693aa8470fee8b0360ae148cfd970bb5c3afb2fb1e41d0da61faf551.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "spans": [ + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "content": "where the router produces weight logits " + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "inline_equation", + "content": "f(\\mathbf{z}_l^{\\prime 0}) = \\mathbf{W} \\cdot \\mathbf{z}_l^{\\prime 0}" + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "content": ", which are normalized by the softmax function. " + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "inline_equation", + "content": "\\mathbf{W} \\in \\mathbb{R}^{D \\times E}" + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "content": " represents the lightweight training parameters and " + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "content": " represents the number of experts. After determining the experts by using the class embedding, we input all embedding including the patch embeddings and class embedding to the activated experts. Each embedding is processed by the top-" + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 549, + 298, + 659 + ], + "type": "text", + "content": " experts with the highest probabilities, and the weighted sum is calculated based on the softmax results of the probabilities:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 100, + 664, + 296, + 696 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 664, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 100, + 664, + 296, + 696 + ], + "type": "interline_equation", + "content": "\\operatorname {M o F E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) = \\sum_ {i = 1} ^ {k} \\mathcal {P} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right) _ {i} \\cdot \\mathcal {E} \\left(\\mathbf {z} _ {\\ell} ^ {\\prime}\\right) _ {i}, \\tag {7}", + "image_path": "5db837817ae2aa83871063d92778efcbc36fe95f3bdc376cfffb49f9281cb331.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": " represents the network of an expert [12]. In our" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "text", + "content": "MoFE architecture, we route to only a single expert, thus " + }, + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "inline_equation", + "content": "k = 1" + }, + { + "bbox": [ + 313, + 327, + 555, + 375 + ], + "type": "text", + "content": ". We find that the router computation is reduced as we are only routing a token to a single expert and the performance does not increase when using more experts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "text", + "content": "Feature Space Separation. In MoFE, we aim to have different experts specialize in different subspaces. Therefore, we propose to first separate the whole feature space into multiple subspaces so that each expert specializes in learning features within its subspace. We use WordNet [39], which provides a good summary of the higher-level semantics of categories, to offer an initial partition of the subspace. Since the semantic and visual similarities are not completely equivalent, we further refine the clustering using K-Means to adjust for the discrepancies. Specifically, we extract feature representations " + }, + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_l^{\\prime 0}" + }, + { + "bbox": [ + 313, + 388, + 556, + 616 + ], + "type": "text", + "content": " for each training image. Then, we calculate the class prototypes by averaging the features of the images from each category. Finally, we perform a K-Means clustering on categorical feature prototypes. The initial cluster centers are determined by the centroids of clusters, which are divided according to the original semantic space. After determining the clustering results, we assign different experts to different clusters, with samples from each category being routed to the corresponding expert model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "MoFE Training. We replace the final transformer block with a MoFE layer. Each transformer block within MoFE is initialized by the original final transformer block. We could set multiple layers as MoFE layers, but we find that using just the final layer achieves sufficiently good results. Then we randomly initialize a router layer and use the class token as the input. We use the labels generated by the above" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1755" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 198, + 85 + ], + "type": "text", + "content": "clustering to supervise the routing:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 96, + 93, + 296, + 127 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 93, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 96, + 93, + 296, + 127 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {r o u t e}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} y ^ {i} \\log \\left(\\mathcal {P} ^ {i} \\left(\\mathbf {z} _ {l} ^ {\\prime 0}\\right)\\right). \\tag {8}", + "image_path": "989e23323868fc73317b17ae2d12176457bdec180736a7f64919e17d475d1d28.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "text", + "content": "For each expert, we leverage the categories within the corresponding cluster as the positive samples, and the categories beyond the cluster as the negative ones. Assuming that the category cluster of the " + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "text", + "content": "th expert contains " + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "inline_equation", + "content": "Q_{i}" + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "text", + "content": " classes, we set the categories beyond the cluster as the " + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "inline_equation", + "content": "Q_{i} + 1" + }, + { + "bbox": [ + 55, + 134, + 297, + 207 + ], + "type": "text", + "content": " categories. The loss is designed as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 84, + 216, + 296, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 84, + 216, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 84, + 216, + 296, + 251 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e x p e r t}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {i = 1} ^ {E} \\sum_ {q = 1} ^ {Q _ {i} + 1} y ^ {i} y ^ {q} \\log \\left(p _ {i} ^ {q} (\\mathbf {x})\\right). \\tag {9}", + "image_path": "a7b088fe2d5e2fd3beaf3d2f615c4e3c7197c7c2bcf047a860c43e090e3a0083.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 258, + 296, + 295 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 296, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 296, + 295 + ], + "type": "text", + "content": "In order to achieve the sample balance for each cluster, we control the ratio of positive and negative samples as 1:1 during training. Therefore, the overall loss of MoFE is:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 124, + 305, + 296, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 305, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 124, + 305, + 296, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M o F E}} = \\mathcal {L} _ {\\text {e x p e r t}} + \\mathcal {L} _ {\\text {r o u t e}}. \\tag {10}", + "image_path": "160e6b6213d9a3ccb7ac9f0a83f1666641b368d2ad1d411439ef5096c618e22c.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 332, + 297, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 297, + 536 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 297, + 536 + ], + "type": "text", + "content": "Discussion. MoFE is designed to address OOD issues under large-scale complex distributions. Similar to MOS [17], we decouple complex distributions into simpler subspace. However, while MOS approaches this solely from the perspective of the loss function, we approach it from the model perspective by assigning an expert model to each subspace. This allows each expert to focus on learning its assigned subspace, preventing interference between features from different partitions. To accurately assign expert models to different samples during inference, we have devised a new routing method that uses the [CLS] token as the input to the router network, since it encapsulates the semantic features of the entire image. With these designs, MoFE outperforms MOS by a substantial margin (see Tab. 1) and the feature visualization shows that MoFE can achieve more compact ID distribution and clearer decision boundaries between ID and OOD samples (Appendix Fig. 5)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 165, + 557 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 165, + 557 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 165, + 557 + ], + "type": "text", + "content": "4.2. Dynamic- " + }, + { + "bbox": [ + 55, + 544, + 165, + 557 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 544, + 165, + 557 + ], + "type": "text", + "content": " Mixup" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 561, + 296, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 598 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 598 + ], + "type": "text", + "content": "Data augmentation (e.g., Mixup [64, 77]) has been proven to improve generalization during finetuning. Traditional Mixup [64, 77] augment samples and transform labels by:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 71, + 607, + 296, + 620 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 607, + 296, + 620 + ], + "spans": [ + { + "bbox": [ + 71, + 607, + 296, + 620 + ], + "type": "interline_equation", + "content": "\\tilde {x} = \\lambda x _ {i} + (1 - \\lambda) x _ {j}, \\quad \\tilde {y} = \\lambda y _ {i} + (1 - \\lambda) y _ {j}, \\tag {11}", + "image_path": "1c8f98017a86ed3a0d4f265005099cad7520f1a45107f55f26caa5131b6b4c9d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda \\sim \\mathrm{Beta}(\\sigma, \\sigma)" + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 629, + 297, + 713 + ], + "type": "text", + "content": " is the interpolation weight for generating new augmented samples. We observe that different categories exhibit varying levels of discriminativeness initialized by vision foundation models, as shown in Appendix Fig. 7c and Fig. 7f. For categories that are already well-represented, synthesizing dissimilar samples via vanilla Mixup can blur the decision boundary between ID" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": "and OOD, leading to degraded performance (Fig. 4 in Appendix). Therefore, we dynamically adjust the Beta distribution according to the feature discriminativeness per category. The reason is that when features of " + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": " are discriminative enough, a small " + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": ", which leads to a dissimilar sample, is not necessary for their representation learning. Instead, we should leverage similar samples from a large " + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": " for building smooth decision boundaries. On the contrary, when features of a category show poor discriminativeness, we should set a relatively small " + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": " to ease the feature learning. We use the accuracy of the validation set to measure the discriminativeness. Therefore, we set " + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 72, + 556, + 216 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 363, + 224, + 555, + 236 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 363, + 224, + 555, + 236 + ], + "spans": [ + { + "bbox": [ + 363, + 224, + 555, + 236 + ], + "type": "interline_equation", + "content": "\\lambda \\sim \\operatorname {B e t a} (\\sigma , \\sigma) \\text {f o r} \\sigma = 1 - w * s, \\tag {12}", + "image_path": "feb91fa27231044da490ffda6b8cb5ffa0051829942c44e313194c4a4db05194.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": " is a scaling factor and " + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": " denotes the corresponding category's accuracy on the validation set. Because the probability density function of " + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "inline_equation", + "content": "\\mathrm{Beta}(\\sigma, \\sigma)" + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": " is symmetric about 0.5 and ranges from 0 to 1, we need to ensure that with a larger " + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": ", the probability of sampling larger values is greater. Therefore, we transform " + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 243, + 556, + 315 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 383, + 321, + 555, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 321, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 383, + 321, + 555, + 354 + ], + "type": "interline_equation", + "content": "\\hat {\\lambda} = \\left\\{ \\begin{array}{l l} \\lambda & \\lambda \\geq 0. 5 \\\\ 1 - \\lambda & \\lambda < 0. 5 \\end{array} . \\right. \\tag {13}", + "image_path": "9a44e8ad271b82e4f50da21f9f559414b080fe9b0d7c50432c58b6dc89ddaf05.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "content": "We determine the category difficulty at the beginning of the training and then update it during the training process. In our implementation, " + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "content": " is the training sample, and " + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "inline_equation", + "content": "x_{j}" + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "content": " is the instance used to corrupt " + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "content": ". Therefore, we select the s from categories of " + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "inline_equation", + "content": "x_{i}" + }, + { + "bbox": [ + 313, + 358, + 556, + 491 + ], + "type": "text", + "content": ", and we select samples from different classes. Additionally, we empirically find that using vanilla Mixup [64, 77] can cause feature norms to grow during finetuning vision foundation models (i.e., DINOv2), leading to performance degradation on the OOD task. In order to restrain the growth of feature norms, we propose to add a regularization term to suppress the increase in feature norm:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 324, + 497, + 555, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 497, + 555, + 530 + ], + "spans": [ + { + "bbox": [ + 324, + 497, + 555, + 530 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {M i x u p}} = - \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {c = 1} ^ {C} y ^ {c} \\log \\left(p ^ {c} (x)\\right) + R e g \\left(F ^ {0}\\right), \\tag {14}", + "image_path": "d5ec03d7676c5d41bf3f2c0109a815978690152eed5c0922f61adc50e8b9d0e2.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "text", + "content": " is the total number of categories, Reg denotes a regularization method, " + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "inline_equation", + "content": "F^0" + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "text", + "content": " is the final class embeddings output by MoFE. By default, the regularization method has multiple choices, which can be " + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 536, + 557, + 597 + ], + "type": "text", + "content": " norm or label smoothing. The final optimization objective is:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 383, + 604, + 555, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 604, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 383, + 604, + 555, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {f i n a l}} = \\mathcal {L} _ {\\text {M o F E}} + \\mathcal {L} _ {\\text {M i x u p}}. \\tag {15}", + "image_path": "610d43f7e839537f6f6a09dcf11fa892e0751a987909609b4902a05da14496bb.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 395, + 647 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "In this section, we set up a benchmark for evaluating OOD performance in Sec. 5.1. Then we compare our methods with the competitive baselines in Sec. 5.2. We conduct ablation studies and present more analysis on our designed method in Sec. 5.3." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 749, + 315, + 757 + ], + "type": "text", + "content": "1756" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 549, + 262 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 549, + 262 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 549, + 262 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 549, + 262 + ], + "type": "table", + "html": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedEnergy [35]65.0087.1757.4087.3246.4391.1757.4087.3256.5588.2479.39
MSP [14]40.8988.6365.8181.2467.9080.1464.9678.1659.8982.0479.39
MaxLogit [15]60.8688.0355.587.4444.8191.1652.2586.0453.3588.1679.39
MCM [41]30.9194.6137.5992.5744.6989.7757.7786.1142.7490.7767.01
CLIPN [70]23.9495.2726.1793.9333.4592.2840.8390.9331.1093.1068.53
LSN [45]21.5695.8334.4891.2526.3294.3538.5490.4230.2292.9671.89
NegPrompt [30]6.3298.4327.6093.3422.8995.5535.2191.6023.0194.8166.84
Ours5.1997.2821.3294.6922.1095.1731.4792.1520.0294.8968.56
Dinov2-BasedMSP [14]25.0294.7657.0983.4553.6585.2248.7985.8148.1387.3186.01
MaxLogit [15]22.9694.5959.2178.4154.5281.8048.1784.1646.2184.7486.01
Energy [35]28.4893.1965.8874.4961.5478.7153.2981.9252.2982.0786.01
KNN [54]5.6797.6543.2588.2136.4290.2128.0492.6628.3492.1886.01
MOS [17]5.0197.8540.1590.3334.3291.8726.1492.9826.4093.2585.23
Ours2.7498.8224.3293.7317.3895.6518.5895.3817.0195.8986.40
", + "image_path": "5c93cdad678b6240799c890f138f5ae0d35679407a6eddc76e9b2af39e96e9fc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 555, + 301 + ], + "type": "text", + "content": "Table 1. Quantitative results of OOD detection performance for ImageNet-1k as ID. We employed our method on two pre-training paradigms (CLIP, and DINOv2). We use FPR95 and AUROC as evaluation metrics. We also report ID classification accuracy. The CLIP-based methods use ViT-B-16, and the DINOv2-based methods use ViT-B-14." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 331, + 133, + 342 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 133, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 133, + 342 + ], + "type": "text", + "content": "5.1. Benchmark" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 445 + ], + "type": "text", + "content": "In- and out-distribution Datasets. To validate the effectiveness of our proposed method, we conduct evaluation on standard benchmarks, which use ImageNet-1K [48] and ImageNet-100 [41] as the ID datasets. Following existing studies that leverage foundation models in OOD [30, 70], we use diverse OOD test datasets, including samples selected from iNaturalist18 [66], SUN [74], Places [79], and Textures [6]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 445, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 295, + 555 + ], + "type": "text", + "content": "Method Comparison. We conduct method comparison on two pretaining paradigms(i.e. CLIP and DINOv2). For each group, we apply some traditional scoring metric (such as MSP [14], MaxLogit [15], Energy [35], KNN [54]). Moreover, we also involve the current CLIP-based state-of-the-art methods, such as MCM [41], CLIPN [70], NegPrompt [30], and LSN [45]. We use KNN [54] as the scoring function when using DINOv2, and follow the scoring metric of CLIPN [70] when applying our method to CLIP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 563, + 140, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 140, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 140, + 574 + ], + "type": "text", + "content": "5.2. Main Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "Results on ImageNet-1K. We compare the proposed approach with the state-of-the-art methods for ImageNet-1K as ID on Tab. 1. These results show: 1) Based on DINOv2, our method reaches the best performance when setting ImageNet-1K as ID. Specifically, our approach reaches " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "17.01\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " FPR95 and " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "95.89\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " AUROC, averaging the results of all the OOD test sets. Our method surpasses MOS [17] by " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "9.39\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " in FPR95, and " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "2.64\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " in AUROC, which proves the importance of learning expert models for OOD detection. 2) When applying our method to CLIP, our method reaches " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "20.02\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "94.89\\%" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": ", which also outperforms NegPrompt [30] by" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 555, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 555, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 555, + 415 + ], + "type": "text", + "content": "a large margin. These results indicate the effectiveness of the proposed MoFE and the dynamic regularized Mixup. 3) Our approach reaches " + }, + { + "bbox": [ + 313, + 331, + 555, + 415 + ], + "type": "inline_equation", + "content": "2.74\\%" + }, + { + "bbox": [ + 313, + 331, + 555, + 415 + ], + "type": "text", + "content": " FPR95 on iNaturalist18 and increases the performance on all the test sets, which indicates that our MoFE design retains the discriminativeness of DINOv2 and facilitates feature learning on various feature subspaces." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": "Results on ImageNet-100. We compare the proposed approach with the state-of-the-art methods for ImageNet-100 as ID on Tab. 2. Based on DINOv2, our method reaches " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "8.10\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " FPR95 and " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "97.75\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " AUROC, surpassing the baseline by " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "4.40\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " FPR95, and " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "0.23\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " AUROC. This indicates that our proposed approach is also effective in a small-scale ID dataset. On the other hand, when applying our method to CLIP, we achieve " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "7.40\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " FPR95 and " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "98.10\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " AUROC, outperforming LSN [45] by " + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "inline_equation", + "content": "1.16\\%" + }, + { + "bbox": [ + 313, + 415, + 556, + 559 + ], + "type": "text", + "content": " FPR95. The above experimental results validate the effectiveness of our approach, and we achieve the best performance on both small-scale and large-scale ID datasets." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 564, + 377, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 377, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 377, + 577 + ], + "type": "text", + "content": "5.3. Analysis" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 665 + ], + "type": "text", + "content": "In this section, we conduct more analysis on the proposed methods. We use the DINOv2 as the pretaining paradigm and KNN [54] as the scoring function. We use ImageNet-1K as ID data and report the average performance on the four out-of-distribution datasets mentioned in Sec. 5.1. The baseline is set to DINOv2 with KNN scoring function without our designed method." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Contributions of Individual Components. As shown in Tab. 3, we evaluate the contribution of each component of the full method. On ImageNet-1K, MoFE and Dynamic-" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " Mixup contribute " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "6.68\\%" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "inline_equation", + "content": "5.41\\%" + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": " FPR95, respectively." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1757" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 70, + 550, + 220 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 550, + 220 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 550, + 220 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 550, + 220 + ], + "type": "table", + "html": "
MethodiNaturalist18PlacesSunTexturesAverageID ACC↑
FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑
CLIP-BasedMSP [14]23.5595.9240.4691.2337.0292.4524.4094.9031.4393.6391.93
MCM [41]18.1396.7734.5294.3636.4594.5441.2292.2532.5894.4887.88
CLIPN [70]4.8798.1613.6496.9313.5597.5615.7893.0211.9696.4191.64
LSN [45]4.9398.9212.8297.198.2397.988.2698.118.5698.0592.24
Ours3.2099.1710.0597.767.0698.399.3197.107.4098.1092.85
Dinov2-BasedMSP [14]5.0698.8526.5894.7827.6495.0226.4394.2721.4295.7294.50
MaxLogit [15]5.5598.7629.6994.1932.7394.2029.2793.7224.3195.2194.50
Energy [35]18.5796.6954.7288.9262.4287.1757.2888.4048.2190.2994.50
KNN [54]2.5899.0218.4595.1215.8996.1616.7996.3813.4296.6694.50
Ours2.2599.2312.8196.668.5197.868.8597.288.1097.7596.94
", + "image_path": "c5da24d9159e60b2806bbb8c946142b2826d5ea7cf4a1e27d8da369951939fa4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 245, + 296, + 293 + ], + "blocks": [ + { + "bbox": [ + 141, + 225, + 468, + 237 + ], + "lines": [ + { + "bbox": [ + 141, + 225, + 468, + 237 + ], + "spans": [ + { + "bbox": [ + 141, + 225, + 468, + 237 + ], + "type": "text", + "content": "Table 2. Quantitative results of OOD detection performance for ImageNet-100 as ID." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 245, + 296, + 293 + ], + "lines": [ + { + "bbox": [ + 57, + 245, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 57, + 245, + 296, + 293 + ], + "type": "table", + "html": "
SettingsBaseline+ MoFE+ D-β+ MoFE+D-β
FPR95↓29.2722.5923.8517.01
AUROC↑92.6794.0193.7295.89
", + "image_path": "c98381c5b26e1755b6b94d678d73764e99ed708136e8cb4afaabb33c175d2425.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 57, + 323, + 296, + 356 + ], + "blocks": [ + { + "bbox": [ + 56, + 294, + 296, + 316 + ], + "lines": [ + { + "bbox": [ + 56, + 294, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 296, + 316 + ], + "type": "text", + "content": "Table 3. Ablation study of individual components for ImageNet-1k as In-Distribution dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 323, + 296, + 356 + ], + "lines": [ + { + "bbox": [ + 57, + 323, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 57, + 323, + 296, + 356 + ], + "type": "table", + "html": "
Num.235789
Gain4.106.309.8012.2612.1012.09
", + "image_path": "0b24410b49ef9a3c9298aa57480fb1cc6362e7fe674cb5e3dfe5c24f00f048e4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 58, + 381, + 296, + 430 + ], + "blocks": [ + { + "bbox": [ + 56, + 357, + 295, + 379 + ], + "lines": [ + { + "bbox": [ + 56, + 357, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 56, + 357, + 295, + 379 + ], + "type": "text", + "content": "Table 4. The effect of Cluster Number. We report the performance gain in FPR95 compared to the model without MoFE." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 381, + 296, + 430 + ], + "lines": [ + { + "bbox": [ + 58, + 381, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 58, + 381, + 296, + 430 + ], + "type": "table", + "html": "
GroupingBaselineWordNetClusteringOurs
FPR95↓29.2725.6326.3422.59
AUROC↑92.6793.0192.9994.01
", + "image_path": "10971f6eb3cd1b400888644121afa48a7043434bf0e75f29554ad85c47a8499d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 295, + 497 + ], + "type": "text", + "content": "When combined, the best performance (17.01% FPR95, and 95.89% AUROC) is achieved." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 498, + 295, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 556 + ], + "type": "text", + "content": "Cluster Number. We conduct an experiment to validate the impact of cluster number on MoFE performance. We set different numbers of clusters. As shown in Tab. 4, we report the performance gain in FPR95. As the increasing of cluster number, the performance gradually increases." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 558, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 641 + ], + "type": "text", + "content": "Feature Space Separation. As shown in Tab. 5, we validate the different strategies for determining the subspace. We compare our method with two methods: WordNet and clustering. The results show that using the ours (i.e. WordNet + Clustering) is the most promising approach. The reason might be that the features extracted by pretrained model are discriminative, especially at coarse-grained level." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "text", + "content": "More Analysis on Dynamic-" + }, + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "text", + "content": " Mixup. As shown in Tab. 6, we conduct an ablation study on Dynamic-" + }, + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 642, + 297, + 712 + ], + "type": "text", + "content": " Mixup. When we remove the regularization term, we find that the performance degrades (30.43% FPR95). Moreover, when we dynamic beta distribution is removed, the performance decreases to 24.96% FPR95." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 259, + 554, + 305 + ], + "blocks": [ + { + "bbox": [ + 83, + 430, + 266, + 441 + ], + "lines": [ + { + "bbox": [ + 83, + 430, + 266, + 441 + ], + "spans": [ + { + "bbox": [ + 83, + 430, + 266, + 441 + ], + "type": "text", + "content": "Table 5. Analysis on Grouping Strategy in MoFE." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 259, + 554, + 305 + ], + "lines": [ + { + "bbox": [ + 317, + 259, + 554, + 305 + ], + "spans": [ + { + "bbox": [ + 317, + 259, + 554, + 305 + ], + "type": "table", + "html": "
MethodsBaselinew/o Regw/o D-βOurs
FPR95↓29.2730.4324.9623.85
AUROC↑92.6791.6593.3693.72
", + "image_path": "90ad20f56a5338d37487e9b12a5d14fada5c7f4ae03811182471e71eba847855.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 348, + 305, + 518, + 316 + ], + "lines": [ + { + "bbox": [ + 348, + 305, + 518, + 316 + ], + "spans": [ + { + "bbox": [ + 348, + 305, + 518, + 316 + ], + "type": "text", + "content": "Table 6. Ablation study of Dynamic- " + }, + { + "bbox": [ + 348, + 305, + 518, + 316 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 348, + 305, + 518, + 316 + ], + "type": "text", + "content": " Mixup." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 388, + 358 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 366, + 555, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 366, + 555, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 366, + 555, + 534 + ], + "type": "text", + "content": "This paper studies the OOD detection task within the context of foundation models. Our study shows that vision foundation models (e.g., DINOv2) are effective OOD detectors, suggesting high-quality and generalizable feature space is essential for OOD detection. Our study highlights that CLIP's pre-trained feature space is less effective for fine-grained tasks, where DINOv2 performs significantly better, which worths further exploration. In order to further optimize the performance of OOD detection when ID data is available for fine-tuning, we propose MoFE and a Dynamic- " + }, + { + "bbox": [ + 313, + 366, + 555, + 534 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 366, + 555, + 534 + ], + "type": "text", + "content": " Mixup data augmentation to enhance the feature learning. We conduct extensive experiments and ablation studies to validate the effectiveness of our approach. We hope our investigation could inspire more future studies." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 544, + 411, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 411, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 411, + 558 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 564, + 556, + 707 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 556, + 707 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 556, + 707 + ], + "type": "text", + "content": "This research is supported by the National Key RD Program of China (Grant No. 2023YFB2806300). This work has been supported in part by Hong Kong Research Grant Council - Early Career Scheme (Grant No. 27209621), General Research Fund Scheme (Grant No. 172024221721292317215025), Theme-based Research (Grant No. T45-701/22-R) and Shenzhen Science and Technology Innovation Commission (SGDX20220530111405040). Part of the described research work is conducted in the JC STEM Lab of Robotics for Soft Materials funded by The Hong Kong Jockey Club Charities Trust." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1758" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Sparse mixture-of-experts are domain generalizable learners, author=Li, Bo and Shen, Yifei and Yang, Jingkang and Wang, Yezhen and Ren, Jiawei and Che, Tong and Zhang, Jun and Liu, Ziwei, booktitle=ICLR, year=2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 178 + ], + "type": "text", + "content": "[2] Mathilde Caron, Hugo Touvron, Ishan Misra, Herve Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. Emerging properties in self-supervised vision transformers. In ICCV, 2021. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 181, + 295, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 181, + 295, + 214 + ], + "spans": [ + { + "bbox": [ + 62, + 181, + 295, + 214 + ], + "type": "text", + "content": "[3] Chirui Chang, Zhengzhe Liu, Xiaoyang Lyu, and Xiaojuan Qi. What matters in detecting ai-generated videos like sora? arXiv preprint arXiv:2406.19568, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "text", + "content": "[4] Guangyao Chen, Limeng Qiao, Yemin Shi, Peixi Peng, Jia Li, Tiejun Huang, Shiliang Pu, and Yonghong Tian. Learning open set network with discriminative reciprocal points. In ECCV, 2020. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 260, + 296, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 260, + 296, + 303 + ], + "spans": [ + { + "bbox": [ + 62, + 260, + 296, + 303 + ], + "type": "text", + "content": "[5] Shaoxiang Chen, Zequn Jie, and Lin Ma. Llava-mole: Sparse mixture of lora experts for mitigating data conflicts in instruction finetuning mllms. arXiv preprint arXiv:2401.16160, 2024.3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 338 + ], + "type": "text", + "content": "[6] Mircea Cimpoi, Subhransu Maji, Iasonas Kokkinos, Sammy Mohamed, and Andrea Vedaldi. Describing textures in the wild. In CVPR, 2014. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 340, + 296, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 296, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 296, + 372 + ], + "type": "text", + "content": "[7] Hendrycks Dan, Zhao Kevin, Basart Steven, Steinhardt Jacob, and Song Dawn. Natural adversarial examples. In CVPR, 2021. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 374, + 296, + 439 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 296, + 439 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 296, + 439 + ], + "type": "text", + "content": "[8] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 441, + 294, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 441, + 294, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 441, + 294, + 473 + ], + "type": "text", + "content": "[9] Xuefeng Du, Zhaoning Wang, Mu Cai, and Yixuan Li. Vos: Learning what you don't know by virtual outlier synthesis. In ICLR, 2022. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 475, + 294, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 507 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 507 + ], + "type": "text", + "content": "[10] Xuefeng Du, Yiyou Sun, Xiaojin Zhu, and Yixuan Li. Dream the impossible: Outlier imagination with diffusion models. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 294, + 542 + ], + "type": "text", + "content": "[11] Sepideh Esmaeilpour, Bing Liu, Eric Robertson, and Lei Shu. Zero-shot out-of-distribution detection based on the pre-trained model clip. In AAAI, 2022. 1, 2, 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 576 + ], + "type": "text", + "content": "[12] William Fedus, Barret Zoph, and Noam Shazeer. Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity. CoRR, 2021. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 578, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 609 + ], + "type": "text", + "content": "[13] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. CVPR, 2022. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 612, + 294, + 644 + ], + "type": "text", + "content": "[14] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In ICLR, 2017. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 294, + 689 + ], + "type": "text", + "content": "[15] Dan Hendrycks, Steven Basart, Mantas Mazeika, Mohammadreza Mostajabi, Jacob Steinhardt, and Dawn Song. Scaling out-of-distribution detection for real-world settings. arXiv preprint arXiv:1911.11132, 2019. 1, 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[16] Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. Generalized odin: Detecting out-of-distribution image with" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 74, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 555, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 555, + 94 + ], + "type": "text", + "content": "out learning from out-of-distribution data. In CVPR, 2020." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 129 + ], + "type": "text", + "content": "[17] Rui Huang and Yixuan Li. Mos: Towards scaling out-of-distribution detection for large semantic space. In CVPR, 2021. 2, 6, 7, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 555, + 184 + ], + "type": "text", + "content": "[18] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 555, + 218 + ], + "type": "text", + "content": "[19] Bitterwolf Julian, Mueller Maximilian, and Matthias. Hein. In or out? fixing imagenet out-of-distribution detection evaluation. In ICML, 2023. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 554, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 554, + 253 + ], + "type": "text", + "content": "[20] Umar Khalid, Ashkan Esmaeili, Nazmul Karim, and Nazanin Rahnavard. Rodd: A self-supervised approach for robust out-of-distribution detection. In CVPRW, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 255, + 555, + 298 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 555, + 298 + ], + "type": "text", + "content": "[21] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dólar, and Ross Girshick. Segment anything. arXiv:2304.02643, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 300, + 555, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 300, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 300, + 555, + 354 + ], + "type": "text", + "content": "[22] Aran Komatsuzaki, Joan Puigcerver, James Lee-Thorp, Carlos Riquelme Ruiz, Basil Mustafa, Joshua Ainslie, Yi Tay, Mostafa Dehghani, and Neil Houlsby. Sparse upcycling: Training mixture-of-experts from dense checkpoints. arXiv preprint arXiv:2212.05055, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 555, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 443 + ], + "type": "text", + "content": "[23] Ivan Krasin, Tom Duerig, Neil Alldrin, Vittorio Ferrari, Sami Abu-El-Haija, Alina Kuznetsova, Hassan Rom, Jasper Uijlings, Stefan Popov, Andreas Veit, Serge Belongie, Victor Gomes, Abhinav Gupta, Chen Sun, Gal Chechik, David Cai, Zheyun Feng, Dhyanesh Narayanan, and Kevin Murphy. Openimages: A public dataset for large-scale multi-label and multi-class image classification. Dataset available from https://github.com/openimages, 2017. 1, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 445, + 555, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 445, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 445, + 555, + 477 + ], + "type": "text", + "content": "[24] Yamuna Krishnamurthy, Chris Watkins, and Thomas Gaertner. Improving expert specialization in mixture of experts. arXiv preprint arXiv:2302.14703, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 479, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 555, + 511 + ], + "type": "text", + "content": "[25] Kimin Lee, Honglak Lee, Kibok Lee, and Jinwoo Shin. Training confidence-calibrated classifiers for detecting out-of-distribution samples. In ICLR, 2018. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 513, + 554, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 513, + 554, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 513, + 554, + 545 + ], + "type": "text", + "content": "[26] Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. A simple unified framework for detecting out-of-distribution samples and adversarial attacks. In NeurIPS, 2018. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 554, + 601 + ], + "type": "text", + "content": "[27] Dmitry Lepikhin, HyoukJoong Lee, Yuanzhong Xu, Dehao Chen, Orhan Firat, Yanping Huang, Maxim Krikun, Noam Shazeer, and Zhifeng Chen. Gshard: Scaling giant models with conditional computation and automatic sharding. arXiv preprint arXiv:2006.16668, 2020. 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 555, + 657 + ], + "type": "text", + "content": "[28] Jinglun Li, Xinyu Zhou, Pinxue Guo, Yixuan Sun, Yiwen Huang, Weifeng Ge, and Wenqiang Zhang. Hierarchical visual categories modeling: A joint representation learning and density estimation framework for out-of-distribution detection. In ICCV, 2023. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 555, + 712 + ], + "type": "text", + "content": "[29] Liunian Harold Li, Pengchuan Zhang, Haotian Zhang, Jianwei Yang, Chunyuan Li, Yiwu Zhong, Lijuan Wang, Lu Yuan, Lei Zhang, Jenq-Neng Hwang, Kai-Wei Chang, and Jianfeng Gao. Grounded language-image pre-training. In CVPR, 2022. 1" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1759" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 106 + ], + "type": "text", + "content": "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In CVPR, 2024. 4, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[31] Shiyu Liang, Yixuan Li, and Rayadurgam Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In ICLR, 2018. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 295, + 185 + ], + "type": "text", + "content": "[32] Randolph Linderman, Jingyang Zhang, Nathan Inkawich, Hai Li, and Yiran Chen. Fine-grain inference on out-of-distribution data with hierarchical classification. arXiv preprint arXiv:2209.04493, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 251 + ], + "type": "text", + "content": "[33] Jiahui Liu, Chirui Chang, Jianhui Liu, Xiaoyang Wu, Lan Ma, and Xiaojuan Qi. Mars3d: A plug-and-play motion-aware model for semantic segmentation on multi-scan 3d point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9372-9381, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 295, + 297 + ], + "type": "text", + "content": "[34] Jiahui Liu, Xin Wen, Shizhen Zhao, Yingxian Chen, and Xiaojuan Qi. Can ood object detectors learn from foundation models? In European Conference on Computer Vision, pages 213-231. Springer, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 298, + 295, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 295, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 295, + 331 + ], + "type": "text", + "content": "[35] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. In NeurIPS, 2020. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 332, + 295, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 295, + 365 + ], + "type": "text", + "content": "[36] Yuxuan Lou, Fuzhao Xue, Zangwei Zheng, and Yang You. Cross-token modeling with conditional computation. arXiv preprint arXiv:2109.02008, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 366, + 295, + 421 + ], + "type": "text", + "content": "[37] Xiaoyang Lyu, Chirui Chang, Peng Dai, Yang-Tian Sun, and Xiaojuan Qi. Total-decom: decomposed 3d scene reconstruction with minimal interaction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20860-20869, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 295, + 520 + ], + "type": "text", + "content": "[38] Oquab Maxime, Darcet Timothee, Moutakanni Théo, Vo Huy, Szafraniec Marc, Khalidov Vasil, Fernandez Pierre, Haziza Daniel, Massa Francisco, El-Nouby Alaaeldin, Assran Mahmoud, Ballas Nicolas, Galuba Wojciech, Howes Russell, Huang Po-Yao, Li Shang-Wen, Misra Ishan, Rabbat Michael, Sharma Vasu, Synnaeve Gabriel, Xu Hu, Jegou Herve, Mairal Julien, Labatut Patrick, Joulin Armand, and Bojanowski Piotr. Dinov2: Learning robust visual features without supervision. arXiv:2304.07193, 2023. 1, 2, 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 295, + 555 + ], + "type": "text", + "content": "[39] George A. Miller. WordNet: A lexical database for English. In Human Language Technology: Proceedings of a Workshop held at Plainsboro, New Jersey, March 8-11, 1994\", 1994. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 556, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 556, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 556, + 295, + 588 + ], + "type": "text", + "content": "[40] Yifei Ming and Yixuan Li. How does fine-tuning impact out-of-distribution detection for vision-language models? IJCV, 2024. 1, 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 295, + 633 + ], + "type": "text", + "content": "[41] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In NeurIPS, 2022. 1, 3, 7, 8, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 669 + ], + "type": "text", + "content": "[42] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. In NIPS, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[43] Basil Mustafa, Carlos Riquelme, Joan Puigcerver, Rodolphe Jenatton, and Neil Houlsby. Multimodal contrastive learning with limoe: the language-image mixture of experts. arXiv preprint arXiv:2206.02770, 2022. 3" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 116 + ], + "type": "text", + "content": "[44] Ibrahim Ndiour, Nilesh Ahuja, and Omesh Tickoo. Out-of-distribution detection with subspace techniques and probabilistic modeling of features. arXiv preprint arXiv:2012.04250, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 554, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 554, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 554, + 152 + ], + "type": "text", + "content": "[45] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In ICLR, 2024. 2, 3, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 153, + 555, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 555, + 208 + ], + "type": "text", + "content": "[46] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, 2021. 1, 2, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 210, + 554, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 554, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 554, + 255 + ], + "type": "text", + "content": "[47] Carlos Riquelme, Joan Puigcerver, Basil Mustafa, Maxim Neumann, Rodolphe Jenatton, André Susano Pinto, Daniel Keysers, and Neil Houlsby. Scaling vision with sparse mixture of experts. NeurIPS, 2021. 3, 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 256, + 555, + 310 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 555, + 310 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 555, + 310 + ], + "type": "text", + "content": "[48] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. IJCV, 2015. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 312, + 554, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 554, + 345 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 554, + 345 + ], + "type": "text", + "content": "[49] Vikash Sehwag, Mung Chiang, and Prateek Mittal. Ssd: A unified framework for self-supervised outlier detection. In ICLR, 2021. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 347, + 555, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 347, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 347, + 555, + 392 + ], + "type": "text", + "content": "[50] Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc V. Le, Geoffrey E. Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer. In ICLR, 2017. 2, 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 393, + 555, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 555, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 555, + 458 + ], + "type": "text", + "content": "[51] Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollar, Christoph Feichtenhofer, Ross Girshick, Rohit Girdhar, and Ishan Misra. The effectiveness of mae pre-pretraining for billion-scale pretraining. In ICCV, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 554, + 482 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 554, + 482 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 554, + 482 + ], + "type": "text", + "content": "[52] Yiyou Sun and Yixuan Li. Dice: Leveraging sparsification for out-of-distribution detection. In ECCV, 2022. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 484, + 555, + 516 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 484, + 555, + 516 + ], + "spans": [ + { + "bbox": [ + 316, + 484, + 555, + 516 + ], + "type": "text", + "content": "[53] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. In NeurIPS, 2021. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 519, + 555, + 552 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 519, + 555, + 552 + ], + "spans": [ + { + "bbox": [ + 316, + 519, + 555, + 552 + ], + "type": "text", + "content": "[54] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In ICML, 2022. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 554, + 554, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 554, + 587 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 554, + 587 + ], + "type": "text", + "content": "[55] Jihoon Tack, Sangwoo Mo, Jongheon Jeong, and Jinwoo Shin. Csi: Novelty detection via contrastive learning on distributionally shifted instances. In NeurIPS, 2020. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 589, + 555, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 589, + 555, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 589, + 555, + 633 + ], + "type": "text", + "content": "[56] Haoru Tan, Chuang Wang, Sitong Wu, Tieqiang Wang, Xu-Yao Zhang, and Cheng-Lin Liu. Proxy graph matching with proximal matching networks. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 555, + 678 + ], + "type": "text", + "content": "[57] Haoru Tan, Sitong Wu, Fei Du, Yukang Chen, Zhibin Wang, Fan Wang, and Xiaojuan Qi. Data pruning via moving-one-sample-out. In Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 554, + 713 + ], + "type": "text", + "content": "[58] Haoru Tan, Sitong Wu, and Jimin Pi. Semantic diffusion network for semantic segmentation. In Neural Information Processing Systems (NeurIPS), 2023." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1760" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[59] Haoru Tan, Chuang Wang, Sitong Wu, Xu-Yao Zhang, Fei Yin, and Cheng-Lin Liu. Ensemble quadratic assignment network for graph matching. International Journal of Computer Vision (IJCV), 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[60] Haoru Tan, Sitong Wu, Zhuotao Tian, Yukang Chen, Xiaojuan Qi, and Jiaya Jia. Saco loss: Sample-wise affinity consistency for vision-language pre-training. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 204 + ], + "type": "text", + "content": "[61] Haoru Tan, Sitong Wu, Wei Huang, Shizhen Zhao, and Xiaojuan Qi. Data pruning by information maximization. In International Conference on Learning Representations (ICLR), 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 295, + 237 + ], + "type": "text", + "content": "[62] Haoru Tan, Sitong Wu, Bo Zhao, Zeke Xie, and XIAOJUAN QI. Diff-in: Data influence estimation with differential approximation, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 239, + 295, + 260 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 295, + 260 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 295, + 260 + ], + "type": "text", + "content": "[63] Leitian Tao, Xuefeng Du, Xiaojin Zhu, and Yixuan Li. Nonparametric outlier synthesis. In ICLR, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 295, + 304 + ], + "type": "text", + "content": "[64] Sunil Thulasidasan, Gopinath Chennupati, Jeff Bilmes, Tan-moy Bhattacharya, and Sarah Michalak. On mixup training: Improved calibration and predictive uncertainty for deep neural networks. In NeurIPS, 2019. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 305, + 295, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 305, + 295, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 305, + 295, + 349 + ], + "type": "text", + "content": "[65] Changyao Tian, Wenhai Wang, Xizhou Zhu, Xiaogang Wang, Jifeng Dai, and Yu Qiao. Vl-ltr: Learning class-wise visual-linguistic representation for long-tailed visual recognition. arXiv preprint arXiv:2111.13579, 2021. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 295, + 392 + ], + "type": "text", + "content": "[66] Grant Van Horn, Oisin Mac Aodha, Yang Song, Yin Cui, Chen Sun, Alex Shepard, Hartwig Adam, Pietro Perona, and Serge Belongie. The inaturalist species classification and detection dataset. In CVPR, 2018. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 295, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 295, + 425 + ], + "type": "text", + "content": "[67] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: a good closed-set classifier is all you need? In ICLR, 2022. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 459 + ], + "type": "text", + "content": "[68] Haoqi Wang, Zhizhong Li, Litong Feng, and Wayne Zhang. Vim: Out-of-distribution with virtual-logit matching. In CVPR, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 295, + 503 + ], + "type": "text", + "content": "[69] Haotao Wang, Aston Zhang, Yi Zhu, Shuai Zheng, Mu Li, Alex J Smola, and Zhangyang Wang. Partial and asymmetric contrastive learning for out-of-distribution detection in long-tailed recognition. In ICML, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 295, + 536 + ], + "type": "text", + "content": "[70] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In ICCV, 2023. 1, 2, 3, 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 581 + ], + "type": "text", + "content": "[71] Hai Wu, Ruifei He, Haoru Tan, Xiaojuan Qi, and Kaibin Huang. Vertical layering of quantized neural networks for heterogeneous inference. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 624 + ], + "type": "text", + "content": "[72] Sitong Wu, Tianyi Wu, Haoru Tan, and Guodong Guo. Pale transformer: A general vision transformer backbone with pale-shaped attention. In The Annual AAAI Conference on Artificial Intelligence (AAAI), 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 295, + 679 + ], + "type": "text", + "content": "[73] Sitong Wu, Haoru Tan, Yukang Chen, Shaofeng Zhang, Jingyao Li, Bei Yu, Xiaojuan Qi, and Jiaya Jia. Mixture-of-scores: Robust image-text data quality score via three lines of code. In International Conference on Computer Vision (ICCV), 2025. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 295, + 713 + ], + "type": "text", + "content": "[74] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, 2010. 7" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 554, + 297 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 315, + 72, + 554, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 554, + 105 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 554, + 105 + ], + "type": "text", + "content": "[75] Zhai Xiaohua, Mustafa Basil, Kolesnikov Alexander, and Beyer Lucas. Sigmoid loss for language image pre-training, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 140 + ], + "type": "text", + "content": "[76] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In CVPR, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 173 + ], + "type": "text", + "content": "[77] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In ICLR, 2017. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 554, + 228 + ], + "type": "text", + "content": "[78] Haotian Zhang, Pengchuan Zhang, Xiaowei Hu, Yen-Chun Chen, Liunian Harold Li, Xiyang Dai, Lijuan Wang, Lu Yuan, Jenq-Neng Hwang, and Jianfeng Gao. Glipv2: Unifying localization and vision-language understanding. arXiv preprint arXiv:2206.05836, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 554, + 263 + ], + "type": "text", + "content": "[79] Bolei Zhou, Agata Lapedriza, Aditya Khosla, Aude Oliva, and Antonio Torralba. Places: A 10 million image database for scene recognition. PAMI, 2017. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 265, + 554, + 297 + ], + "type": "text", + "content": "[80] Jinghao Zhou, Chen Wei, Huiyu Wang, Wei Shen, Cihang Xie, Alan Yuille, and Tao Kong. ibot: Image bert pre-training with online tokenizer. ICLR, 2022. 4" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1761" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_content_list.json b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8f0ca42e980a8b1e6623f56f42ba822cfa7e7239 --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_content_list.json @@ -0,0 +1,1246 @@ +[ + { + "type": "text", + "text": "Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts", + "text_level": 1, + "bbox": [ + 101, + 128, + 897, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ibtihel Amara $^{1,2\\dagger}$ , Ahmed Imtiaz Humayun $^{1,3\\dagger}$ , Ivana Kajic $^{4}$ , Zarana Parekh $^{4}$ , Natalie Harris $^{1}$ , Sarah Young $^{1}$ , Chirag Nagpal $^{1}$ , Najoung Kim $^{4}$ , Junfeng He $^{1}$ , Cristina Nader Vasconcelos $^{4}$ , Deepak Ramachandran $^{4}$ , Golnoosh Farnadi $^{1}$ , Katherine Heller $^{1}$ , Mohammad Havaei $^{1}$ , Negar Rostamzadeh $^{1}$", + "bbox": [ + 99, + 202, + 955, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Google Research, 2McGill University, 3Rice University, 4Google Deepmind", + "bbox": [ + 297, + 257, + 759, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 309, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality.", + "bbox": [ + 89, + 342, + 483, + 675 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 707, + 220, + 722 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As text-to-image (T2I) generative models [6, 20, 21, 38, 40, 45] have become increasingly popular and widely adopted, the demand for techniques that allow for the targeted removal of specific concepts has grown in parallel. Concept erasure methods [10, 11, 13, 18, 25, 26, 30, 32, 34, 44, 54] promise to address important concerns such as ensuring the safe deployment of models by removing undesirable or harmful concepts [10, 15, 52]. Whether applied to mitigate bias [4, 34, 46], protect privacy [10, 15, 22, 52], or filter out unwanted content, these techniques aim to create \"sanitized\" models capable of generating content without", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ee506efc645bb2c28150078b2ca6f7042a72b114f53625c688d251993ab02a9.jpg", + "image_caption": [ + "Figure 1. The effects of concept erasure on non-target concepts. Pre-erasure outputs (left) vs. post-erasure results (right) from Stable Diffusion (SD). Erasure negatively impacts the quality of unrelated concepts. EraseBench identifies such effects and offers a framework to evaluate the reliability of erasure methods." + ], + "image_footnote": [], + "bbox": [ + 521, + 309, + 906, + 569 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "triggering unwanted associations. However, the challenge of effectively erasing a concept without unintentionally distorting related or visually similar concepts remains unresolved. For example, erasing \"cat\" from a model should ideally remove all representations of a cat, yet leave semantically or visually related concepts like \"lion\" unaffected. Figure 1 shows various failure modes observed in state-of-the-art concept erasure techniques. These techniques exhibit instability after a concept is erased, often causing unintended consequences for closely related non-target concepts such as over-erasure and image artifacts (i.e. distortions). In this work, we argue that concept erasure techniques often struggle under real-world conditions. Even state-of-the-art methods fail when erasing concepts that are visually similar, binomially related, or semantically entangled. This lack of robustness raises concerns about their deployment in sen", + "bbox": [ + 511, + 659, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "16420", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "sitive applications, as unintended distortions of related non-target concepts affects the reliability of T2I models.", + "bbox": [ + 89, + 90, + 480, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "While recent works assess erasure techniques using arbitrary concepts, a standardized framework for stress-testing across diverse concept relationships is still lacking. To address this, we introduce EraseBench, a unified benchmark with curated concepts and evaluations designed to rigorously test T2I models post-erasure. Covering binomial pairs, visually similar objects, and semantic relations. EraseBench enables consistent evaluation of erasure effectiveness, unintended distortions, and concept leakage. Our findings highlight persistent challenges and the need for more robust and standardized testing. We envision EraseBench as a foundation for improving comparability and advancing safer and more reliable erasure methods. Additionally, our results demonstrate that these sanitized models are not yet ready for deployment, highlighting the need for more robust and reliable erasure techniques that can operate effectively in complex, real-world scenarios.", + "bbox": [ + 89, + 121, + 480, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 89, + 378, + 429, + 391 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We identify key evaluation dimensions where most concept erasure techniques exhibit vulnerabilities, particularly when handling visually similar, binomial, and semantically related concepts.", + "2. We present EraseBench, a comprehensive, multidimensional framework designed to rigorously evaluate the robustness and efficacy of concept erasure techniques across a diverse set of concepts and prompts.", + "3. We use a suite of evaluation metrics to measure the effectiveness of concept erasure methods that provides a more holistic assessment of their performance.", + "4. We evaluate five state-of-the-art concept erasure techniques on EraseBench and demonstrate that many of the current concept erasure techniques are not yet ready for real-world deployment, highlighting significant gaps in their reliability and robustness." + ], + "bbox": [ + 89, + 393, + 482, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 648, + 238, + 664 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Erasure. T2I models have witnessed remarkable advancements since the foundational works of [31, 57], driven by breakthroughs in model architectures [5, 19, 42, 49, 53, 56], state-of-the-art generative modeling techniques [1, 3, 12, 16, 23, 24, 29, 29, 33, 37-40, 50], and the availability of large high-quality datasets [27, 41]. These improvements have significantly enhanced their capability to generate realistic and diverse images from textual descriptions, pushing the boundaries of creativity and application. In light of recent advancements in T2I models, greater attention has been directed towards reliability considerations. There is a wide variety of concept erasure techniques, each utilizing distinct methods and technologies to achieve concept removal [7, 9-11, 14, 18, 22, 30, 30, 54].", + "bbox": [ + 89, + 674, + 482, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Challenges in Concept Erasure. Despite the rapid in", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "crease of these concept erasure methods, only a few works have critically examined the inherent limitations and vulnerabilities of concept erasure approaches. In particular, the work in [35] explores a fundamental vulnerability, which is the potential for erased concepts to be recovered through specific prompting. Meanwhile, other works, such as in [55], assess the effectiveness of concept erasure by examining a diverse range of objects and styles. Their approach evaluates whether the erased concepts have been completely removed from the model's generative capabilities, ensuring that no traces remain in the generated outputs. In our work, we delve into the broader impact of concept erasure, not only evaluating residual concept generation but also examining how erasure techniques may degrade the quality and fidelity of non-target concepts. In this work, we highlight the ripple effects of concept erasure on model performance and the stability of related concepts via the route of interrelated concepts.", + "bbox": [ + 511, + 90, + 903, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. EraseBench", + "text_level": 1, + "bbox": [ + 513, + 376, + 638, + 392 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this section, we provide a detailed breakdown of the process used to construct our benchmark.", + "bbox": [ + 511, + 402, + 903, + 430 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Gathering. We utilized the semantic knowledge embedded in large language models (LLMs) and structured sources, like the hierarchical taxonomy of ImageNet, to generate clusters of closely related concepts. By treating these language models akin to graph neural networks, we mapped similar concepts that share key attributes and gathered distinct sets for each dimension. This approach captures nuanced relationships crucial for testing erasure precision. For artistic styles, we selected artist names from the WikiArt dataset [47].", + "bbox": [ + 511, + 431, + 905, + 582 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Verification. To ensure quality and feasibility, we conducted human verification on the collected concepts. This included assessing whether existing T2I models could successfully generate accurate, high-quality representations of each concept. During this process, we identified and removed any concepts that models consistently struggled to render, preserving only those concepts that aligned with our benchmarks. This step ensures that the dataset maintains a high standard, with each concept suitable for rigorous testing of erasure techniques.", + "bbox": [ + 511, + 584, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Prompt Construction. We used Gemini Flash [43], to generate a diverse set of prompts for each concept. Prompts were crafted with variations in style, complexity, and length to allow for thorough evaluation. This includes prompts that differ in level of detail (simple vs. elaborate descriptions) and length (short vs. extended prompts). With this range, we enable a finer assessment of concept erasure techniques.", + "bbox": [ + 511, + 734, + 903, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Concept Dimensions. We introduce multiple evaluation dimensions, each capturing a distinct type of concept relationship or similarity, providing a structured way to evaluate the impact of erasure. Within each dimension, we pro", + "bbox": [ + 511, + 840, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "16421", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c12e62b9cb30586992e201acb12108a1b6e6f8415d0c61d0c201a34c113a0aa3.jpg", + "image_caption": [ + "Figure 2. Ripple effects of concept erasure methods across EraseBench entanglement dimensions. All the erasure baselines display failure cases across different EraseBench tasks. Receler and MACE frequently produce images that are unrelated to the text prompt, indicating they are the most sensitive of the five concept erasure techniques. In contrast, AdvUnlearn shows slightly better robustness across certain dimensions of the benchmark." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 893, + 588 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ee4c688bd9440dfee2a1a80929d28630c22ae348d9279e1c8f954bcc0368ef6.jpg", + "image_caption": [ + "Figure 3. Evaluation Dimensions of EraseBench." + ], + "image_footnote": [], + "bbox": [ + 119, + 667, + 455, + 781 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "vide a diverse set of concepts that challenge the model in different ways, ensuring a thorough evaluation. Figure 3 summarizes these dimensions in EraseBench. For each, we define a primary concept for erasure and related non-target concepts sharing visual, semantic, or contextual similari-", + "bbox": [ + 89, + 824, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ties to test unintended effects on the sanitized model. (1) Visual Similarity (Objects), A broad concept (e.g., \"cat\") is erased, along with synonyms (\"kitten,\" \"tabby\"), while visually similar non-targets (\"tiger,\" \"cheetah\") test unintended erasure. (2) Style Similarity (Artists), Erasing an artist (e.g., Van Gogh) tests model behavior on similar-style artists (e.g., Cézanne, Bernard). (3) Subset-Superset Relationship, Removing a concept (e.g., \"goldfish\") evaluates retention of related non-targets (\"guppy,\" \"koi\"). (4) Binomial Relations, Erasing one of a closely linked pair (e.g., \"sun\" from \"sun and moon\") tests impact on the other.", + "bbox": [ + 511, + 662, + 906, + 828 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Explicit Content. Under each dimension of EraseBench, we also integrate explicit concepts to evaluate the handling of sensitive and safety-critical content (details can be found in Table 8 in the supplemental material). These include", + "bbox": [ + 511, + 829, + 908, + 890 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "16422", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "categories such as NSFW, culturally related, and broader safety-related themes. Incorporating these explicit concepts allows for a more rigorous assessment of how well erasure methods balance effectiveness, generation quality, and sensitivity.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 181, + 279, + 199 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Baseline models. In our work, we narrowed down existing techniques and focused on leveraging more recent, advanced and diverse approaches to concept erasure, experimenting with recent techniques that incorporate variations such as fine-tuning model weights (e.g. ESD [10]), introducing targeted weight perturbations (e.g. UCE [11]), refining textual embeddings and adversarial training (e.g. Receler [18], AdvUnlearn [54]), and introducing parameter efficient fine-tuning (Receler, MACE [30]). This allows us to explore the nuanced dynamics of concept erasure and assess its impact under different dimensions. We provide detailed description about each of these techniques in the supplemental material.", + "bbox": [ + 89, + 205, + 483, + 400 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Experiments. We use Stable Diffusion (SD) as our T2I model and apply concept erasure techniques to each primary concept in EraseBench. For each concept, we perform erasure using both the default and best-reported settings from baseline methods. We maintain the default retain sets (e.g. COCO prompts for AdvUnlearn) and, for methods requiring anchor concepts, we map them to an empty concept (\"). We evaluate the sanitized models on related non-target concepts within the same EraseBench dimension, assessing the effects on visually and semantically similar concepts. Each concept is represented by over 10 unique prompts, ranging from simple to complex descriptions. To capture diverse outputs, we generate 10 images per prompt with different random seeds. This approach provides a comprehensive evaluation of erasure effectiveness and its unintended effects on related concepts.", + "bbox": [ + 89, + 402, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use different types of automated evaluation metrics to quantify and explain the detrimental effect erasure techniques can have on the quality of generated images of non-target concepts. We use three different automated evaluation metrics to quantify the post erasure performance: CLIP [36], Rich Automatic Human Feedback (RAHF [28]), and Gecko [48]. CLIP is a widely used joint image and text embedding model which quantifies the similarity between text and image representations. We leverage CLIP as a zero-shot classifier to determine whether a model generates the desired concept. The RAHF model is trained on rich human feedback for generated images with a range of scores to evaluate overall image quality, and also mark points about which regions are problematic. The RAHF quality scores include an aesthetic score, which assesses the visual appeal of an image, and an artifact score (i.e. plausibility score), which gauges the presence and extent", + "bbox": [ + 89, + 645, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of visual artifacts. We use both the aesthetic and artifact scores as measurements of quality for generated images, following [28], where higher scores indicate better quality (i.e. good aesthetics and minimal artifacts). In addition to scores, the RAHF model provides heatmaps for each image to highlight artifact-prone areas, which we use for visual inspection in our benchmark. Finally, we use a Visual Question Answering (VQA)-based metric called Gecko, a VQA evaluation framework. In our setup, we leverage the multimodal Gemini 1.5 model for both steps: first, to generate text-related questions, and second, to answer them based on the given image, similar to other VQA methods [2, 51]. One advantage of such metrics is that they do not require task-specific fine-tuning and provide an interpretable alignment score. This evaluation can be replicated using other available models. Typically, an LLM generates questions related to the text prompt in the first step, while a vision-language model (VLM) answers those questions based on the generated image. The final score is computed as the average of correctly answered questions, allowing for backtracing to identify which aspects of the text are misaligned with the generated image.", + "bbox": [ + 511, + 90, + 906, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 513, + 436, + 601, + 452 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "5.1. Concept Erasure Reduces T2I Alignment For Non-Target Concepts (Over-Erasure)", + "text_level": 1, + "bbox": [ + 511, + 460, + 906, + 492 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/340ad0189edabfbf75c37c294d78220a7ae1c5405fbf5ae0205ba7bf16f8bd55.jpg", + "image_caption": [], + "image_footnote": [ + "Q1: Is there a tiger in the image?" + ], + "bbox": [ + 540, + 507, + 702, + 633 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/86cfd8f20e7ca061ec337942e991d09e0805407ace5698d9e80565a9781b377b.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [ + "Q2: Is the tiger perched on something?" + ], + "bbox": [ + 714, + 507, + 875, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Alignment score", + "bbox": [ + 702, + 723, + 799, + 734 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4b6e1b5ff9c3ec449dcf77154277f2dbb165f9867eb40fb144bf205dbe4a11d6.jpg", + "image_caption": [ + "50", + "Figure 4. Erasure affects fine-grained alignment. Prompt: \"A tiger perched on a rocky outcrop surrounded by mountains and a serene blue sky.\" before (left) and after (right) erasing the concept cat using MACE [30])." + ], + "image_footnote": [ + "Q3: Is the tiger perched on a rocky outcrop?", + "Q4: Are there mountains in the image?", + "Q5: Is the sky a serene blue color?", + "Q6: Is there a sky visible in the image?" + ], + "bbox": [ + 805, + 632, + 831, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Table 1 presents the zero-shot predictions from CLIP as percentages (\\%). Following the definitions in [30], we measure two key attributes: Efficacy (Eff.) and Generality (Gen.). Efficacy represents the accuracy of concept erasure (accuracy on the erased class), where lower values indicate better erasure. Generality reflects accuracy on paraphrased or syn", + "bbox": [ + 511, + 810, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "16423", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/8b8b11f5a3b8372826621a519a7aaa8f453f98013b03b79fa3ddbe02be739942.jpg", + "table_caption": [ + "Table 1. Evaluation of Erasing concepts from EraseBench under the four different dimensions. We provide the average results across 10 concepts (for visual similarity (object)), 15 artists concepts (for artists similarity), 8 concepts for subset-superset, and 10 concepts (for binomial). We evaluate efficacy (Eff., capturing the efficacy of erasing the main concept), generality (Gen., capturing the impact on paraphrased or related concepts), and sensitivity (Sens., capturing the unintended effect on non-target yet related or similar concepts). All values are expressed as percentages (\\%). More detailed, per-concept breakdowns are provided in the Supplemental Material. For reference, we also include the classification accuracy of images generated by the original Stable Diffusion v1.4 model." + ], + "table_footnote": [], + "table_body": "
TechniquesVisual Similarity (Object)Artistic Style Similarity (Artists)Subset-SupersetBinomial
Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑
Original86.5 ± 7.190.2 ± 3.685.0 ± 4.015.9772.3 ± 6.380.7 ± 3.269.0 ± 4.329.2983.3 ± 8.992.8 ± 3.992.1 ± 3.514.3188.3 ± 5.289.8 ± 4.388.5 ± 4.515.40
ESD [10]24.5 ± 6.150.5 ± 4.165.9 ± 4.661.7015.1 ± 3.761.9 ± 4.840.3 ± 3.947.7418.1 ± 4.251.2 ± 4.265.1 ± 5.162.4220.8 ± 4.641.3 ± 4.370.6 ± 4.068.46
UCE [11]41.8 ± 5.568.3 ± 2.782.7 ± 3.149.3221.1 ± 4.352.2 ± 3.561.0 ± 4.360.0151.1 ± 8.862.7 ± 2.987.5 ± 2.951.1218.9 ± 2.231.4 ± 2.886.1 ± 4.3277.88
Receler [18]8.1 ± 3.220 ± 3.465.4 ± 3.877.588.7 ± 2.245.3 ± 3.822.8 ± 3.841.044.2 ± 0.820.4 ± 3.736.7 ± 7.559.7010.3 ± 2.412.6 ± 3.157.5 ± 9.275.04
MACE [30]15.6 ± 6.437.7 ± 3.966.4 ± 4.369.8320.2 ± 4.636.7 ± 4.749.2 ± 4.861.6613.9 ± 3.228.2 ± 3.866.9 ± 5.674.1011.3 ± 2.528.9 ± 5.170.7 ± 5.375.98
AdvUnlearn [54]8.7 ± 2.939.1 ± 6.564.3 ± 5.769.8814.5 ± 4.137.4 ± 3.927.4 ± 4.047.757.4 ± 2.630.1 ± 5.260.1 ± 5.971.879.3 ± 2.227.6 ± 2.164.6 ± 5.974.41
", + "bbox": [ + 91, + 183, + 903, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "onymous non-erased concepts, showing how well general information is erased. Additionally, we introduce Sensitivity (Sens.), defined as the accuracy for non-target but similar concepts. Most of the concept erasure techniques decreased accuracy on the target concept class (a desired outcome). Receler, MACE, and AdvUnlearn achieved the best efficacy in erasure overall per EraseBench dimension. Only a few techniques showed strong generality, indicating that erasure was not applied in a broad, global manner. In particular, Receler consistently displayed the best generality, preserving accuracy across target erased classes and paraphrased non-erased concepts. This may be attributed to the fact that Receler was trained using adversarial prompting and employs \"erasers\" designed to capture textual semantic concepts effectively. Overall, we observed a decrease in sensitivity, an undesirable outcome that highlights T2I misalignment for non-target non-erased concepts, essentially, over-erasure. After erasing a target concept, we noticed a clear mismatch between the text prompt and the generated image for related non-target concepts. Low sensitivity suggests that instead of selectively erasing only the target concept, the model inadvertently suppresses related concepts, causing generated images to fall out of distribution. This indicates that the model has not only erased the intended concept but also learned to overlook semantically or visually similar ones. As a result, the generated images fail to align with the intended prompts, revealing the model's inability to fully preserve non-target concepts after erasure. In Figure 2 (row 1), we observe that when models undergo erasure of the koala concept, prompting them with related concepts like Tree Kangaroo reveals unintended effect. These erased models appear to have also forgotten how to generate a Tree Kangaroo accurately. Similarly, when we erase the artist concept Bosch and prompt the sanitized model to generate artwork in the style of Altdorfer, the model struggles to capture the distinctive artistic style of the non-erased concept. This shows that the erasure process affects not only the target concept but also related artistic styles, leading to a loss of fidelity in the generated images. This demonstrates a significant vulnerability in these erasure tech", + "bbox": [ + 89, + 285, + 485, + 891 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0958fbcbf75ba8904226883cad7818947a26c2f4df29f24008fd816f385a24ca.jpg", + "table_caption": [ + "Table 2. Harmonic mean $(\\uparrow)$ of efficacy, generality, and sensitivity scores on explicit concepts across EraseBench dimensions. Each column represents a distinct relationship type: visually similar objects, artist style similarity, subset-superset relationships, and binomial pairs. Higher scores indicate better balance between erasure success and minimal unintended degradation. While all methods outperform the original model, performance varies by dimension, revealing trade-offs in handling sensitive and safety-critical content." + ], + "table_footnote": [], + "table_body": "
TechniqueVisual Sim.Subset-supersetBinomial
Original16.3723.4819.64
UCE [11]63.1662.6276.13
AdvUnlearn [54]67.6658.2871.33
Receler [18]68.4756.3174.88
", + "bbox": [ + 517, + 419, + 903, + 501 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7a8de9ee9fe99f90ebf742637d0e95067f22850d2dc800ec6e4bdff3ca82925b.jpg", + "table_caption": [ + "Table 3. Text-to-image alignment scores using the Gecko metric. The average scores and the standard errors of the mean for images generated with different techniques and prompts that either contain the erased concept or not. \"Original\" is the baseline model w/o any concept erasure. Drop in score compared to the original model shown in brackets." + ], + "table_footnote": [], + "table_body": "
TechniqueErased ConceptsNon-erased Concepts
Original84.1 ± 0.977.6 ± 0.7
UCE [11]57.6 ± 1.7 (−26.4)74.3 ± 0.8 (−3.4)
MACE [30]38.2 ± 1.3 (−45.9)67.9 ± 0.9 (−9.8)
AdvUnlearn [54]43.1 ± 1.4 (−41.0)68.6 ± 0.9 (−9.0)
", + "bbox": [ + 517, + 606, + 903, + 708 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "niques and highlights their risks associated with deploying them in real-world unrestricted settings. This trend of low sensitivity is consistent across dimensions of EraseBench. If we want to evaluate overall, UCE demonstrates superior preservation of visual concept components compared to other methods. Table 2 presents the harmonic mean of efficacy, generality, and sensitivity scores for explicit concepts (see Table 8 in the supplemental material for details) across three EraseBench dimensions: Visual Similarity, Subset-superset, and Binomial. UCE demonstrates consistently strong performance across most dimensions, while AdvUnlearn and Receler achieve particularly high scores in", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "16424", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/03f02e19b6359ad849ba68ca742b828317687d1fed841461a13a1ec1de780b1b.jpg", + "image_caption": [ + "Original Image" + ], + "image_footnote": [ + "Er: Goldfish" + ], + "bbox": [ + 96, + 102, + 220, + 200 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4f5915c06971e2de571135cc7e5a4eb175925ddb600a9c6d3a7dd1898c17887.jpg", + "image_caption": [ + "UCE" + ], + "image_footnote": [ + "Pr: A Guppy" + ], + "bbox": [ + 222, + 107, + 346, + 204 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/89734462cff06f4d4824c3ad95b18a477bdf021b50dddd787c4cd0009bb18ddf.jpg", + "image_caption": [ + "AdvUnlearn" + ], + "image_footnote": [], + "bbox": [ + 349, + 107, + 475, + 204 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/56bae0e5407c2dd75304e7c523b061ef449a6ca964c8a1e4498b5254b683dac0.jpg", + "image_caption": [ + "Figure 5. Erasure introduces artifacts during subset-superset concept generation. We erase concept \"goldfish\" and generate images for the prompt \"an image of a guppy\". We present the RAHF artifact heatmaps for images generated post-erasure via AdvUnlearn and UCE. We see that the artifact introduced by each method can vary spatially and by intensity, which prompts our inclusion of the artifact score in EraseBench. More heatmaps examples can be found in the supplemental material." + ], + "image_footnote": [ + "Artifact Scores:", + "Original:93.63", + "UCE: 84.21", + "AdvUnlearn:79.01" + ], + "bbox": [ + 223, + 207, + 346, + 303 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e9b0cd298957c065681039aac20e5d3cda7fb3f9624f132b61b9cebaccb6399b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 207, + 475, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Subset-superset and visual similarity settings. These results highlight the value of evaluating concept erasure across diverse relational categories to ensure method robustness and general applicability.", + "bbox": [ + 89, + 459, + 482, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Next, we use Gecko to evaluate image-text alignment with a VQA-based approach. Unlike CLIP, Gecko provides an interpretable alignment score as it is computed over question-answer pairs. Figure 4 shows an example of question-answer pairs for a pair of an original and a sanitized image. While some aspects of the prompt are still correctly depicted in the image, the score is lower for sanitized model as it is missing the animal. We compute alignment scores between images generated by original and sanitized models (UCE, MACE, and AdvUnlearn) and prompts corresponding to four different concepts (goat, goldfish, cat, and Vincent Van Gogh) from four different dimensions. In total, we compute alignment scores for 6246 text-image pairs. The results in Table 3 show an expected drop in scores for images generated with sanitized models compared to images generated with the original model (c.f. column \"Erased concepts\"), as we expect the erased concept to be missing in images generated by sanitized models. However, we also observe a drop in scores for non-erased concepts across different dimensions. While that drop is much smaller, it is statistically significant across all three comparisons ( $\\alpha = .01$ using the Wilcoxon signed-rank test). The UCE technique has the smallest gap with the original model for non-erased concepts, while MACE has the largest. The results for UCE are consistent with CLIP-score zero-shot", + "bbox": [ + 89, + 523, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "prediction in Table 1, where UCE was consistently the technique with highest sensitivity across all dimensions. We also manually inspected a subset of questions, answers and images, to ensure that the lower score is not caused by the VQA model latching on spurious correlations, but is indeed due to depicted concepts missing or the image being of lower quality. We observe that while some concepts are generally more difficult to depict correctly even in the original model (e.g. goldfish), this is reflected in the lower overall alignment score for the original model (e.g., for all three techniques the alignment score of the original model is in 60s for the concept goldfish, while it is in 80s and 90s for other concepts). Finally, question-answer breakdown in Gecko allows us to pinpoint some failures of image generation in sanitized models. When looking at answers to questions asking about object presence (eg., Is there an $\\langle$ object $\\rangle$ in the image?), we found that for all three techniques erased models had lower Gecko scores compared to original models on prompts with visually similar objects (UCE: $80.3\\%$ vs $89.0\\%$ , MACE: $67.1\\%$ vs $89.0\\%$ , and AdvUnlearn: $83.1\\%$ vs $89.0\\%$ ), which highlights unintentional consequences of concept erasure techniques.", + "bbox": [ + 511, + 90, + 906, + 424 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Concept Erasure Reduces Quality For Non-Target Concepts", + "text_level": 1, + "bbox": [ + 511, + 438, + 906, + 470 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Global Concepts. Beyond non-target misalignment, T2I models also exhibit other forms of degradation post-erasure, such as distortions (i.e. artifacts) and poor aesthetics. To explore this further, we specifically employed metrics that better align with human feedback, such as artifact measurements that quantify the degree of perceptible distortions and aesthetic quality (i.e. factors that reflect how humans perceive image fidelity). In Table 4, we present the RAHF alignment scores of individual concepts under different dimensions of EraseBench. Our results reveal a significant decline in overall quality compared to the original SD, suggesting that the erasure process leads to noticeable degradation in image quality. This indicates that similar and related concepts remain vulnerable to generation flaws, despite attempts at concept retention. In Figure 2, we observe that most models capable of retaining and generating the intended concept after erasure still exhibit various types of distortions. These include misaligned body parts for animal classes, cropped concepts, nonsensical text distortions, decreased size of the generated concept, and a general lack of sharpness. Additional examples can be found in the Supplemental Material. Artists Concepts. We observe a significant drop in T2I alignment, highlighting a major challenge in artistic style erasure as highlighted in Table 1. Similar styles or concepts across artists are at risk of being inadvertently erased if one artist's style is closely related to another. We observe this issue in the context of T2I alignment and found a marked degradation in quality, particularly in the", + "bbox": [ + 511, + 477, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "16425", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2074999bba1982d3e3cbe4f185ae63cb38a4854ac3705ed72fd53196ae8d722d.jpg", + "table_caption": [ + "Table 4. RAHF alignment scores [28]. Alignment scores for each concept under the overall dimension object similarity." + ], + "table_footnote": [], + "table_body": "
TechniquesVisual Similarity (Object)BinomialSubset- superset
Erase "Cat"Erase "Goat"Erase "Lock"Erase "Goldfish"
Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑
Original87.71 ± 0.0780.93 ± 0.0283.23 ± 0.678.44 ± 0.285.23 ± 1.576.72 ± 0.884.15 ± 0.178.21 ± 0.3
UCE [11]72.44 ± 0.973.77 ± 0.275.00 ± 1.074.19 ± 0.378.88 ± 3.478.50 ± 1.074.55 ± 0.672.20 ± 0.4
MACE [30]73.50 ± 0.871.80 ± 0.0772.00 ± 1.173.40 ± 1.272.79 ±4.077.22 ± 0.772.64 ± 0.375.33 ± 0.2
AdvUnlearn [54]74.60 ± 0.972.30 ± 1.269.78 ± 0.670.10 ± 0.966.23 ± 5.376.79 ± 1.171.87 ± 0.572.49 ± 0.2
", + "bbox": [ + 91, + 112, + 903, + 210 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "style of non-target concepts. In Figure 2, we show the erasure of $Degas$ (the artist) and its impact on the ability to generate images of other artists, such as Cassatt. Notably, we see a significant shift in style even for the non-erased concept. We specifically chose not to use the RAHF alignment scores for artistic style, as the models were not trained to evaluate artistic style based on their documentation.", + "bbox": [ + 89, + 223, + 480, + 328 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Empirical Validation with Human Preferences", + "text_level": 1, + "bbox": [ + 89, + 356, + 480, + 372 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We conduct a human preference study to additionally validate results obtained with automated evaluation techniques. We recruited 11 participants from our institution to judge images based on three different criteria: overall image quality, distortions and T2I alignment. All participants were instructed on how to complete the task, and have provided consent to participate in the study. Each participant was presented with 50 side-by-side image comparisons, corresponding to 50 pairs of images generated by the original model and by the model with a concept erased. The order of images was randomized for each pair during presentation. The participants were asked the following three questions: (i) Which image exhibits superior overall quality? (ii) Which of the following images displays LESS noticeable distortions? (iii) Which image most accurately reflects and is aligned with the text label?", + "bbox": [ + 88, + 380, + 482, + 622 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Participants selected one of the three possible answers: Image A, Image B, or Neutral. In total, we collected 1650 responses for UCE and 1485 for AdvUnlearn. We focus on images of non-erased concepts, as the primary hypothesis we wanted to test was whether non-target visually similar concepts were negatively affected with the erasure technique. The results in Figure 6 show that humans judge images generated with the original model as having better overall quality and having fewer distortions. The original model is also preferred for better alignment, although it is closer to the \"Neutral\" for UCE, meaning that either both the original image and sanitized image were of equally good or equally poor quality. We provide additional human preferences for AdvUnlearn in the supplemental material. Results in Figure 10 also show that most participants preferred the original images with respect to quality, alignment, and less artifacts. Overall, we find that human preferences corroborate findings we observe with automatic metrics.", + "bbox": [ + 89, + 628, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d6a3f2f71ca49af6b4282fc356124d1bf0a9eedf590d73381e70f616808f6406.jpg", + "image_caption": [ + "Figure 6. Human image preferences between images generated by the original and the erased model. The erased model used is UCE. Results show that humans prefer SD over UCE. We provide the results of AdvUnlearn in the supplemental material." + ], + "image_footnote": [], + "bbox": [ + 531, + 220, + 890, + 335 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Discussion", + "text_level": 1, + "bbox": [ + 513, + 422, + 625, + 438 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In our work, we have identified and emphasized critical issues surrounding T2I models post concept erasure, making it our primary focus. Addressing these challenges is the logical next step, but it raises a complex research question that demands further exploration. In this section, we discuss into possible mitigation strategies and assess their impacts. We stress that finding effective solutions is an ongoing, nontrivial process that requires continued investigation and innovation.", + "bbox": [ + 511, + 446, + 905, + 582 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The Role of Retain Sets in Mitigating Post-Erasure Distortions. A key question is whether the design of the retain set, specifically including concepts closely intertwined with the target erased concept can help alleviate post-generation distortions. We try to investigate whether adding non-target concepts that share semantic or visual similarities can reduce unintended side effects caused by erasure. To explore this, we performed erasure on targeted concepts using EraseBench while strategically incorporating related nontarget concepts into the retain set. For example, when erasing cat, we progressively introduce related concepts such as tiger and panther into the retain set and evaluate the impact on both generated outputs and artifact persistence. We show in Figure 7 the influence of the retain set on non-target concepts. We compare the original SD without erasure, erased model without a retain set (i.e. no explicit retain set), and erased model with a retain set in which we introduce the entangled non-target concepts one by one.", + "bbox": [ + 511, + 583, + 905, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "(i) Findings on non-target concepts. Our investigation indicates that incorporating entangled non-target concepts into the retain set partially alleviates over-erasure (misalign", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "16426", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ment on non-target concepts), preserves aspects of related non-target concepts. However, artifacts and distortions persist, particularly in non-targeted non-erased concepts suggesting that while carefully designing a retain set may help maintain concept structure, it does not fully resolve the unintended distortions.", + "bbox": [ + 89, + 90, + 480, + 180 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(ii) Findings on target-erased concepts. Examining the erased target concepts in Figure 8, we find that incorporating multiple entangled concepts into the retain set helps alleviate some post-erasure artifacts on the non-target concepts but comes at the cost of concept leakage. Specifically, this leakage manifests as spurious regeneration, where the erased concept partially or fully re-emerges in generated outputs. Our observations suggest that while including related non-target concepts in the retain set can mitigate over-erasure, it increases the risk of unintended regeneration. The more entangled concepts added, the greater the likelihood of leakage, highlighting a trade-off between preserving non-target concepts and ensuring effective erasure.", + "bbox": [ + 89, + 181, + 482, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure. We investigate whether anchor concepts can help in mitigating the unintended distortions in non-target concepts observed after concept erasure. An anchor may act as a reference point, defining boundaries for selective erasure while preserving related concepts. To test this, we conducted experiments in Figure 9 where we erased Van Gogh while introducing Post-Impressionism as an anchor concept. Similarly, we erased Grumpy Cat while anchoring the model with Internet meme, as outlined in [8]. The intuition behind this approach is that a well-defined anchor might help the model retain a coherent representation of related concepts, thereby reducing unintended consequences of erasure. We observed models struggle with generating good quality non-target concept images, suggesting anchors do not consistently improve stability. While anchors may help in alignment, they do not fully resolve artifacts or style preservation issues, warranting further investigation into their effectiveness.", + "bbox": [ + 89, + 378, + 482, + 662 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Multi-Concept Erasure: Intra-type and Inter-type Effects. We also investigate multi-concept erasure, differentiating between intra-type (main and paraphrased concepts) and inter-type (main and unrelated concepts) erasure. Average artifact scores in the visual similarity domain reach $78.3\\%$ with intra-type and $71.4\\%$ with inter-type erasure. The former outperforms the single concept erasure at $71.9\\%$ (see example in Figure 15 in the supplemental material). Our findings show that intra-type erasure (i.e. removing multiple related concepts) more effectively eliminates targets and reduces artifacts when faced with non-target but related prompts. In contrast, inter-type erasure offers minimal artifact reduction, suggesting that carefully grouping related concepts enhances erasure precision without degrading generation quality.", + "bbox": [ + 89, + 664, + 482, + 891 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c97d047a46c55dcef46adcc1774602d209664816cc1050cb5a1a83a62d4a1349.jpg", + "image_caption": [ + "Figure 7. Impact of adding visually similar concepts to the retain set with respect to non-target concepts. Without a retain set, image quality degrades significantly. Unrelated concepts (e.g. house, elephant, and boat) fail to prevent artifacts, while a curated, visually similar retain set improves quality but remains below pre-erasure levels." + ], + "image_footnote": [], + "bbox": [ + 517, + 71, + 903, + 224 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8249c162b0b8457e088a7a39bf84e9b1c191cb94a54034c4dd533e13ef87b39c.jpg", + "image_caption": [ + "Figure 8. Impact of adding visually similar concepts to the retain set with respect to the target concept. Without a retain set, erasure is effective, but as similar concepts are added, the erased concept gradually re-emerges." + ], + "image_footnote": [], + "bbox": [ + 517, + 325, + 903, + 421 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/be64878cb49855bbc2d2c9ce620db014d0ddd0a35ebcc6c56e7f1d640ceac727.jpg", + "image_caption": [ + "Figure 9. Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure." + ], + "image_footnote": [], + "bbox": [ + 514, + 505, + 689, + 583 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b005257329c1238324ca51a6380da723bb0b6a0d40c96d58384fbc1840f0b556.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 505, + 885, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion and Limitations", + "text_level": 1, + "bbox": [ + 513, + 637, + 769, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we present a thorough evaluation of concept erasure techniques and uncover shortcomings in their reliability. By introducing EraseBench, we reveal critical gaps in current methods. Our findings suggest that despite recent advancements, existing concept erasure techniques fall short in preserving model quality, which call for the need for more robust and nuanced approaches as well as more robust evaluation protocol and metrics. While our framework provides a diverse evaluation, automating concept selection could improve scalability. Given dimensions like binomiality, researchers can explore broader concept spaces beyond our curation. Additionally, a deeper exploration of optimal retain or erasure sets to minimize distortions remains an open question. We hope EraseBench inspires new research directions toward developing more reliable concept erasure techniques.", + "bbox": [ + 511, + 664, + 906, + 887 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "16427", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2", + "[2] Emanuele Bugliarello, Laurent Sartran, Aishwarya Agrawal, Lisa Anne Hendricks, and Aida Nematzadeh. Measuring progress in fine-grained vision-and-language understanding. arXiv preprint arXiv:2305.07558, 2023. 4", + "[3] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2", + "[4] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 1", + "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2", + "[6] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1", + "[7] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2", + "[8] Masane Fuchi and Tomohiro Takagi. Erasing concepts from text-to-image diffusion models with few-shot unlearning. arXiv preprint arXiv:2405.07288, 2, 2024. 8", + "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 6", + "[10] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 4, 5, 3, 6", + "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 4, 5, 7, 3, 6", + "[12] Roy Ganz and Michael Elad. Clipag: Towards generator-free text-to-image generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3843-3853, 2024. 2", + "[13] Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, and Weiming Zhang. Eraseanything: Enabling concept erasure in rectified" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "flow transformers. In Forty-second International Conference on Machine Learning, 2025. 1", + "[14] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36: 17170-17194, 2023. 2", + "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2", + "[17] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4", + "[18] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. ECCV, 2024. 1, 2, 4, 5, 6", + "[19] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. In European conference on computer vision, pages 91-109. Springer, 2022. 2", + "[20] Imagen-Team-Google, Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, Hongliang Fei, Nando de Freitas, Yilin Gao, Evgeny Gladchenko, Sergio Gomez Colmenarejo, Mandy Guo, Alex Haig, Will Hawkins, Hexiang Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Yelin Kim, Ksenia Konyushkova, Karol Langner, Eric Lau, Shixin Luo, Soña Mokrá, Henna Nandwani, Yasumasa Onoe, Aaron van den Oord, Zarana Parekh, Jordi Pont-Tuset, Hang Qi, Rui Qian, Deepak Ramachandran, Poorva Rane, Abdullah Rashwan, Ali Razavi, Robert Riachi, Hansa Srinivasan, Srivatsan Srinivasan, Robin Strudel, Benigno Uria, Oliver Wang, Su Wang, Austin Waters, Chris Wolff, Auriel Wright, Zhisheng Xiao, Hao Xiong, Keyang Xu, Marc van Zee, Junlin Zhang, Katie Zhang, Wenlei Zhou, Konrad Zolna, Ola Aboubakar, Canfer Akbulut, Oscar Akerlund, Isabela Albuquerque, Nina Anderson, Marco Andretto, Lora Aroyo, Ben Bariach, David Barker, Sherry Ben, Dana Berman, Courtney Biles, Irina Blok, Pankil Botadra, Jenny Brennan, Karla Brown, John Buckley, Rudy Bunel, Elie Bursztein, Christina Butterfield, Ben Caine, Viral Carpenter, Norman Casagrande, Ming-Wei Chang, Solomon Chang, Shamik Chaudhuri, Tony Chen, John Choi, Dmitry Churbanau, Nathan Clement, Matan Cohen, Forrester Cole, Mikhail Dektiarev, Vincent Du, Praneet Dutta, Tom Eccles, Ndidi Elue, Ashley Feden, Shlomi Fruchter, Frankie Garcia, Roopal Garg, Weina Ge, Ahmed Ghazy, Bryant Gipson, Andrew Goodman, Dawid Górný, Sven Gowal, Khyatti Gupta, Yoni Halpern, Yena Han, Susan Hao, Jamie Hayes, Amir Hertz, Ed Hirst Tingbo Hou Heidi Howard" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "16428", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mohamed Ibrahim, Dirichi Ike-Njoku, Joana Iljazi, Vlad Ionescu, William Isaac, Reena Jana, Gemma Jennings, Donovon Jenson, Xuhui Jia, Kerry Jones, Xiaoen Ju, Ivana Kajic, Christos Kaplanis, Burcu Karagol Ayan, Jacob Kelly, Suraj Kothawade, Christina Kouridi, Ira Ktena, Jolanda Kumakaw, Dana Kurniawan, Dmitry Lagun, Lily Lavitas, Jason Lee, Tao Li, Marco Liang, Maggie Li-Calis, Yuchi Liu, Javier Lopez Alberca, Peggy Lu, Kristian Lum, Yukun Ma, Chase Malik, John Mellor, Inbar Mosseri, Tom Murray, Aida Nematzadeh, Paul Nicholas, João Gabriel Oliveira, Guillermo Ortiz-Jimenez, Michela Paganini, Tom Le Paine, Roni Paiss, Alicia Parrish, Anne Peckham, Vikas Peswani, Igor Petrovski, Tobias Pfaff, Alex Pirozhenko, Ryan Poplin, Utsav Prabhu, Yuan Qi, Matthew Rahtz, Cyrus Rashtchian, Charvi Rastogi, Amit Raul, Ali Razavi, Sylvestre-Alvise Rebuffi, Susanna Ricco, Felix Riedel, Dirk Robinson, Pankaj Rohatgi, Bill Rosgen, Sarah Rumbley, Moonkyung Ryu, Anthony Salgado, Sahil Singla, Florian Schroff, Candice Schumann, Tanmay Shah, Brendan Shillingford, Kaushik Shivakumar, Dennis Shtatnov, Zach Singer, Evgeny Sluzhaev, Valerii Sokolov, Thibault Sottiaux, Florian Stimberg, Brad Stone, David Stutz, Yu-Chuan Su, Eric Tabellion, Shuai Tang, David Tao, Kurt Thomas, Gregory Thornton, Andeep Toor, Cristian Udrescu, Aayush Upadhyay, Cristina Vasconcelos, Alex Vasiloff, Andrey Voynov, Amanda Walker, Luyu Wang, Miaosen Wang, Simon Wang, Stanley Wang, Qifei Wang, Yuxiao Wang, Agoston Weisz, Olivia Wiles, Chenxia Wu, Xingyu Federico Xu, Andrew Xue, Jianbo Yang, Luo Yu, Mete Yurtoglu, Ali Zand, Han Zhang, Jia-geng Zhang, Catherine Zhao, Adilet Zhaxybay, Miao Zhou, Shengqi Zhu, Zhenkai Zhu, Dawn Bloxwich, Mahyar Bordbar, Luis C. Cobo, Eli Collins, Shengyang Dai, Tulsee Doshi, Anca Dragan, Douglas Eck, Demis Hassabis, Sissie Hsiao, Tom Hume, Koray Kavukcuoglu, Helen King, Jack Krawczyk, Yeqing Li, Kathy Meier-Hellstern, Andras Orban, Yury Pinsky, Amar Subramanya Oriol Vinyals Ting Yu and Yori Zwols. Imagen 3, 2024. 1", + "[21] Jun Young Koh, Sang Hyun Park, and Joy Song. Improving text generation on images with synthetic captions, 2024. 1", + "[22] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 7", + "[23] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[24] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R. Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9400-9409, 2024. 2", + "[25] Leyang Li, Shilin Lu, Yan Ren, and Adams Wai-Kin Kong. Set you straight: Auto-steering denoising trajectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782, 2025. 1" + ], + "bbox": [ + 93, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Ouxiang Li, Yuan Wang, Xinting Hu, Houcheng Jiang, Tao Liang, Yanbin Hao, Guojun Ma, and Fuli Feng. Speed: Scalable, precise, and efficient concept erasure for diffusion models. arXiv preprint arXiv:2503.07392, 2025. 1", + "[27] Zejian Li, Chenye Meng, Yize Li, Ling Yang, Shengyuan Zhang, Jiarui Ma, Jiayi Li, Guang Yang, Changyuan Yang, Zhiyuan Yang, et al. Laion-sg: An enhanced large-scale dataset for training complex image-text models with structural annotations. arXiv preprint arXiv:2412.08580, 2024. 2", + "[28] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19401-19411, 2024. 4, 7", + "[29] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 2", + "[30] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 7, 6", + "[31] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2", + "[32] Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3", + "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2", + "[34] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3", + "[35] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 2", + "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "16429", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1", + "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024.", + "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 1, 2", + "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2", + "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2", + "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2", + "[44] Zhihua Tian, Sirun Nan, Ming Xu, Shengfang Zhai, Wenjie Qu, Jian Liu, Kui Ren, Ruoxi Jia, and Jiaheng Zhang. Sparse autoencoder as a zero-shot classifier for concept erasing in text-to-image diffusion models. arXiv preprint arXiv:2503.09446, 2025.1", + "[45] Cristina Nader Vasconcelos, Abdullah Rashwan, Austin Waters, Trevor Walker, Keyang Xu, Jimmy Yan, Rui Qian, Yeqing Li, SHIXIN LUO, Yasumasa Onoe, Zarana Parekh, Ivana Kajic, Mandy Guo, Wenlei Zhou, Sarah Rosston, Roopal Garg, Hongliang Fei, Jordi Pont-Tuset, Su Wang, Henna Nandwani, Andrew Bunner, Kevin Swersky, David J. Fleet, Oliver Wang, and Jason Michael Baldridge. Greedy growing enables high-resolution pixel-based diffusion models. Transactions on Machine Learning Research, 2024. 1", + "[46] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for text-controlled vision models. arXiv preprint arXiv:2302.03693, 2, 2023. 1", + "[47] WikiArt.org. WikiArt - Visual Art Encyclopedia. https: //www.wikiart.org/, 2024. Accessed: June 30, 2024. 2", + "[48] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with gecko: On met" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "rics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 4", + "[49] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2", + "[50] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. Advances in Neural Information Processing Systems, 36, 2024. 2", + "[51] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 4", + "[52] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755-1764, 2024. 1", + "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2", + "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 5, 7, 6", + "[55] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2", + "[56] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2", + "[57] Xiaojin Zhu, Andrew B Goldberg, Mohamed Eldawy, Charles R Dyer, and Bradley Strock. A text-to-picture synthesis system for augmenting communication. In AAAI, pages 1590-1595, 2007. 2" + ], + "bbox": [ + 516, + 92, + 903, + 753 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "16430", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_model.json b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_model.json new file mode 100644 index 0000000000000000000000000000000000000000..72fa48a4ead866bd222e4413571337b542881107 --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_model.json @@ -0,0 +1,2136 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.898, + 0.178 + ], + "angle": 0, + "content": "Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.203, + 0.957, + 0.257 + ], + "angle": 0, + "content": "Ibtihel Amara\\(^{1,2\\dagger}\\), Ahmed Imtiaz Humayun\\(^{1,3\\dagger}\\), Ivana Kajic\\(^{4}\\), Zarana Parekh\\(^{4}\\), Natalie Harris\\(^{1}\\), Sarah Young\\(^{1}\\), Chirag Nagpal\\(^{1}\\), Najoung Kim\\(^{4}\\), Junfeng He\\(^{1}\\), Cristina Nader Vasconcelos\\(^{4}\\), Deepak Ramachandran\\(^{4}\\), Golnoosh Farnadi\\(^{1}\\), Katherine Heller\\(^{1}\\), Mohammad Havaei\\(^{1}\\), Negar Rostamzadeh\\(^{1}\\)" + }, + { + "type": "text", + "bbox": [ + 0.299, + 0.258, + 0.761, + 0.275 + ], + "angle": 0, + "content": "1Google Research, 2McGill University, 3Rice University, 4Google Deepmind" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.31, + 0.327, + 0.326 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.485, + 0.676 + ], + "angle": 0, + "content": "Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.708, + 0.222, + 0.723 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As text-to-image (T2I) generative models [6, 20, 21, 38, 40, 45] have become increasingly popular and widely adopted, the demand for techniques that allow for the targeted removal of specific concepts has grown in parallel. Concept erasure methods [10, 11, 13, 18, 25, 26, 30, 32, 34, 44, 54] promise to address important concerns such as ensuring the safe deployment of models by removing undesirable or harmful concepts [10, 15, 52]. Whether applied to mitigate bias [4, 34, 46], protect privacy [10, 15, 22, 52], or filter out unwanted content, these techniques aim to create \"sanitized\" models capable of generating content without" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.31, + 0.907, + 0.57 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.578, + 0.907, + 0.648 + ], + "angle": 0, + "content": "Figure 1. The effects of concept erasure on non-target concepts. Pre-erasure outputs (left) vs. post-erasure results (right) from Stable Diffusion (SD). Erasure negatively impacts the quality of unrelated concepts. EraseBench identifies such effects and offers a framework to evaluate the reliability of erasure methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.909, + 0.903 + ], + "angle": 0, + "content": "triggering unwanted associations. However, the challenge of effectively erasing a concept without unintentionally distorting related or visually similar concepts remains unresolved. For example, erasing \"cat\" from a model should ideally remove all representations of a cat, yet leave semantically or visually related concepts like \"lion\" unaffected. Figure 1 shows various failure modes observed in state-of-the-art concept erasure techniques. These techniques exhibit instability after a concept is erased, often causing unintended consequences for closely related non-target concepts such as over-erasure and image artifacts (i.e. distortions). In this work, we argue that concept erasure techniques often struggle under real-world conditions. Even state-of-the-art methods fail when erasing concepts that are visually similar, binomially related, or semantically entangled. This lack of robustness raises concerns about their deployment in sen" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16420" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "sitive applications, as unintended distortions of related non-target concepts affects the reliability of T2I models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.122, + 0.482, + 0.379 + ], + "angle": 0, + "content": "While recent works assess erasure techniques using arbitrary concepts, a standardized framework for stress-testing across diverse concept relationships is still lacking. To address this, we introduce EraseBench, a unified benchmark with curated concepts and evaluations designed to rigorously test T2I models post-erasure. Covering binomial pairs, visually similar objects, and semantic relations. EraseBench enables consistent evaluation of erasure effectiveness, unintended distortions, and concept leakage. Our findings highlight persistent challenges and the need for more robust and standardized testing. We envision EraseBench as a foundation for improving comparability and advancing safer and more reliable erasure methods. Additionally, our results demonstrate that these sanitized models are not yet ready for deployment, highlighting the need for more robust and reliable erasure techniques that can operate effectively in complex, real-world scenarios." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.379, + 0.431, + 0.392 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.394, + 0.483, + 0.455 + ], + "angle": 0, + "content": "1. We identify key evaluation dimensions where most concept erasure techniques exhibit vulnerabilities, particularly when handling visually similar, binomial, and semantically related concepts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.455, + 0.483, + 0.514 + ], + "angle": 0, + "content": "2. We present EraseBench, a comprehensive, multidimensional framework designed to rigorously evaluate the robustness and efficacy of concept erasure techniques across a diverse set of concepts and prompts." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.515, + 0.483, + 0.559 + ], + "angle": 0, + "content": "3. We use a suite of evaluation metrics to measure the effectiveness of concept erasure methods that provides a more holistic assessment of their performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.56, + 0.483, + 0.635 + ], + "angle": 0, + "content": "4. We evaluate five state-of-the-art concept erasure techniques on EraseBench and demonstrate that many of the current concept erasure techniques are not yet ready for real-world deployment, highlighting significant gaps in their reliability and robustness." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.394, + 0.483, + 0.635 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.24, + 0.665 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.885 + ], + "angle": 0, + "content": "Concept Erasure. T2I models have witnessed remarkable advancements since the foundational works of [31, 57], driven by breakthroughs in model architectures [5, 19, 42, 49, 53, 56], state-of-the-art generative modeling techniques [1, 3, 12, 16, 23, 24, 29, 29, 33, 37-40, 50], and the availability of large high-quality datasets [27, 41]. These improvements have significantly enhanced their capability to generate realistic and diverse images from textual descriptions, pushing the boundaries of creativity and application. In light of recent advancements in T2I models, greater attention has been directed towards reliability considerations. There is a wide variety of concept erasure techniques, each utilizing distinct methods and technologies to achieve concept removal [7, 9-11, 14, 18, 22, 30, 30, 54]." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Challenges in Concept Erasure. Despite the rapid in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.363 + ], + "angle": 0, + "content": "crease of these concept erasure methods, only a few works have critically examined the inherent limitations and vulnerabilities of concept erasure approaches. In particular, the work in [35] explores a fundamental vulnerability, which is the potential for erased concepts to be recovered through specific prompting. Meanwhile, other works, such as in [55], assess the effectiveness of concept erasure by examining a diverse range of objects and styles. Their approach evaluates whether the erased concepts have been completely removed from the model's generative capabilities, ensuring that no traces remain in the generated outputs. In our work, we delve into the broader impact of concept erasure, not only evaluating residual concept generation but also examining how erasure techniques may degrade the quality and fidelity of non-target concepts. In this work, we highlight the ripple effects of concept erasure on model performance and the stability of related concepts via the route of interrelated concepts." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.64, + 0.393 + ], + "angle": 0, + "content": "3. EraseBench" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.905, + 0.431 + ], + "angle": 0, + "content": "In this section, we provide a detailed breakdown of the process used to construct our benchmark." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.583 + ], + "angle": 0, + "content": "Concept Gathering. We utilized the semantic knowledge embedded in large language models (LLMs) and structured sources, like the hierarchical taxonomy of ImageNet, to generate clusters of closely related concepts. By treating these language models akin to graph neural networks, we mapped similar concepts that share key attributes and gathered distinct sets for each dimension. This approach captures nuanced relationships crucial for testing erasure precision. For artistic styles, we selected artist names from the WikiArt dataset [47]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Concept Verification. To ensure quality and feasibility, we conducted human verification on the collected concepts. This included assessing whether existing T2I models could successfully generate accurate, high-quality representations of each concept. During this process, we identified and removed any concepts that models consistently struggled to render, preserving only those concepts that aligned with our benchmarks. This step ensures that the dataset maintains a high standard, with each concept suitable for rigorous testing of erasure techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.84 + ], + "angle": 0, + "content": "Prompt Construction. We used Gemini Flash [43], to generate a diverse set of prompts for each concept. Prompts were crafted with variations in style, complexity, and length to allow for thorough evaluation. This includes prompts that differ in level of detail (simple vs. elaborate descriptions) and length (short vs. extended prompts). With this range, we enable a finer assessment of concept erasure techniques." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Concept Dimensions. We introduce multiple evaluation dimensions, each capturing a distinct type of concept relationship or similarity, providing a structured way to evaluate the impact of erasure. Within each dimension, we pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16421" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.086, + 0.895, + 0.589 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.595, + 0.908, + 0.651 + ], + "angle": 0, + "content": "Figure 2. Ripple effects of concept erasure methods across EraseBench entanglement dimensions. All the erasure baselines display failure cases across different EraseBench tasks. Receler and MACE frequently produce images that are unrelated to the text prompt, indicating they are the most sensitive of the five concept erasure techniques. In contrast, AdvUnlearn shows slightly better robustness across certain dimensions of the benchmark." + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.668, + 0.456, + 0.782 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.8, + 0.434, + 0.814 + ], + "angle": 0, + "content": "Figure 3. Evaluation Dimensions of EraseBench." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.901 + ], + "angle": 0, + "content": "vide a diverse set of concepts that challenge the model in different ways, ensuring a thorough evaluation. Figure 3 summarizes these dimensions in EraseBench. For each, we define a primary concept for erasure and related non-target concepts sharing visual, semantic, or contextual similari-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.663, + 0.907, + 0.829 + ], + "angle": 0, + "content": "ties to test unintended effects on the sanitized model. (1) Visual Similarity (Objects), A broad concept (e.g., \"cat\") is erased, along with synonyms (\"kitten,\" \"tabby\"), while visually similar non-targets (\"tiger,\" \"cheetah\") test unintended erasure. (2) Style Similarity (Artists), Erasing an artist (e.g., Van Gogh) tests model behavior on similar-style artists (e.g., Cézanne, Bernard). (3) Subset-Superset Relationship, Removing a concept (e.g., \"goldfish\") evaluates retention of related non-targets (\"guppy,\" \"koi\"). (4) Binomial Relations, Erasing one of a closely linked pair (e.g., \"sun\" from \"sun and moon\") tests impact on the other." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.83, + 0.909, + 0.891 + ], + "angle": 0, + "content": "Explicit Content. Under each dimension of EraseBench, we also integrate explicit concepts to evaluate the handling of sensitive and safety-critical content (details can be found in Table 8 in the supplemental material). These include" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "16422" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "categories such as NSFW, culturally related, and broader safety-related themes. Incorporating these explicit concepts allows for a more rigorous assessment of how well erasure methods balance effectiveness, generation quality, and sensitivity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.183, + 0.281, + 0.2 + ], + "angle": 0, + "content": "4. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.207, + 0.484, + 0.401 + ], + "angle": 0, + "content": "Baseline models. In our work, we narrowed down existing techniques and focused on leveraging more recent, advanced and diverse approaches to concept erasure, experimenting with recent techniques that incorporate variations such as fine-tuning model weights (e.g. ESD [10]), introducing targeted weight perturbations (e.g. UCE [11]), refining textual embeddings and adversarial training (e.g. Receler [18], AdvUnlearn [54]), and introducing parameter efficient fine-tuning (Receler, MACE [30]). This allows us to explore the nuanced dynamics of concept erasure and assess its impact under different dimensions. We provide detailed description about each of these techniques in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.403, + 0.484, + 0.644 + ], + "angle": 0, + "content": "Experiments. We use Stable Diffusion (SD) as our T2I model and apply concept erasure techniques to each primary concept in EraseBench. For each concept, we perform erasure using both the default and best-reported settings from baseline methods. We maintain the default retain sets (e.g. COCO prompts for AdvUnlearn) and, for methods requiring anchor concepts, we map them to an empty concept (\"). We evaluate the sanitized models on related non-target concepts within the same EraseBench dimension, assessing the effects on visually and semantically similar concepts. Each concept is represented by over 10 unique prompts, ranging from simple to complex descriptions. To capture diverse outputs, we generate 10 images per prompt with different random seeds. This approach provides a comprehensive evaluation of erasure effectiveness and its unintended effects on related concepts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.646, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We use different types of automated evaluation metrics to quantify and explain the detrimental effect erasure techniques can have on the quality of generated images of non-target concepts. We use three different automated evaluation metrics to quantify the post erasure performance: CLIP [36], Rich Automatic Human Feedback (RAHF [28]), and Gecko [48]. CLIP is a widely used joint image and text embedding model which quantifies the similarity between text and image representations. We leverage CLIP as a zero-shot classifier to determine whether a model generates the desired concept. The RAHF model is trained on rich human feedback for generated images with a range of scores to evaluate overall image quality, and also mark points about which regions are problematic. The RAHF quality scores include an aesthetic score, which assesses the visual appeal of an image, and an artifact score (i.e. plausibility score), which gauges the presence and extent" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.425 + ], + "angle": 0, + "content": "of visual artifacts. We use both the aesthetic and artifact scores as measurements of quality for generated images, following [28], where higher scores indicate better quality (i.e. good aesthetics and minimal artifacts). In addition to scores, the RAHF model provides heatmaps for each image to highlight artifact-prone areas, which we use for visual inspection in our benchmark. Finally, we use a Visual Question Answering (VQA)-based metric called Gecko, a VQA evaluation framework. In our setup, we leverage the multimodal Gemini 1.5 model for both steps: first, to generate text-related questions, and second, to answer them based on the given image, similar to other VQA methods [2, 51]. One advantage of such metrics is that they do not require task-specific fine-tuning and provide an interpretable alignment score. This evaluation can be replicated using other available models. Typically, an LLM generates questions related to the text prompt in the first step, while a vision-language model (VLM) answers those questions based on the generated image. The final score is computed as the average of correctly answered questions, allowing for backtracing to identify which aspects of the text are misaligned with the generated image." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.437, + 0.602, + 0.453 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.461, + 0.907, + 0.493 + ], + "angle": 0, + "content": "5.1. Concept Erasure Reduces T2I Alignment For Non-Target Concepts (Over-Erasure)" + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.508, + 0.703, + 0.634 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.715, + 0.508, + 0.877, + 0.633 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.647, + 0.724, + 0.659 + ], + "angle": 0, + "content": "Q1: Is there a tiger in the image?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.66, + 0.757, + 0.671 + ], + "angle": 0, + "content": "Q2: Is the tiger perched on something?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.671, + 0.784, + 0.683 + ], + "angle": 0, + "content": "Q3: Is the tiger perched on a rocky outcrop?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.684, + 0.754, + 0.695 + ], + "angle": 0, + "content": "Q4: Are there mountains in the image?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.696, + 0.735, + 0.707 + ], + "angle": 0, + "content": "Q5: Is the sky a serene blue color?" + }, + { + "type": "image_footnote", + "bbox": [ + 0.541, + 0.708, + 0.755, + 0.72 + ], + "angle": 0, + "content": "Q6: Is there a sky visible in the image?" + }, + { + "type": "list", + "bbox": [ + 0.541, + 0.647, + 0.784, + 0.72 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.703, + 0.724, + 0.8, + 0.736 + ], + "angle": 0, + "content": "Alignment score" + }, + { + "type": "image", + "bbox": [ + 0.807, + 0.633, + 0.833, + 0.721 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.634, + 0.864, + 0.721 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image_caption", + "bbox": [ + 0.845, + 0.724, + 0.864, + 0.735 + ], + "angle": 0, + "content": "50" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.747, + 0.908, + 0.802 + ], + "angle": 0, + "content": "Figure 4. Erasure affects fine-grained alignment. Prompt: \"A tiger perched on a rocky outcrop surrounded by mountains and a serene blue sky.\" before (left) and after (right) erasing the concept cat using MACE [30])." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Table 1 presents the zero-shot predictions from CLIP as percentages (\\%). Following the definitions in [30], we measure two key attributes: Efficacy (Eff.) and Generality (Gen.). Efficacy represents the accuracy of concept erasure (accuracy on the erased class), where lower values indicate better erasure. Generality reflects accuracy on paraphrased or syn" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16423" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.175 + ], + "angle": 0, + "content": "Table 1. Evaluation of Erasing concepts from EraseBench under the four different dimensions. We provide the average results across 10 concepts (for visual similarity (object)), 15 artists concepts (for artists similarity), 8 concepts for subset-superset, and 10 concepts (for binomial). We evaluate efficacy (Eff., capturing the efficacy of erasing the main concept), generality (Gen., capturing the impact on paraphrased or related concepts), and sensitivity (Sens., capturing the unintended effect on non-target yet related or similar concepts). All values are expressed as percentages (\\%). More detailed, per-concept breakdowns are provided in the Supplemental Material. For reference, we also include the classification accuracy of images generated by the original Stable Diffusion v1.4 model." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.184, + 0.905, + 0.271 + ], + "angle": 0, + "content": "
TechniquesVisual Similarity (Object)Artistic Style Similarity (Artists)Subset-SupersetBinomial
Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑
Original86.5 ± 7.190.2 ± 3.685.0 ± 4.015.9772.3 ± 6.380.7 ± 3.269.0 ± 4.329.2983.3 ± 8.992.8 ± 3.992.1 ± 3.514.3188.3 ± 5.289.8 ± 4.388.5 ± 4.515.40
ESD [10]24.5 ± 6.150.5 ± 4.165.9 ± 4.661.7015.1 ± 3.761.9 ± 4.840.3 ± 3.947.7418.1 ± 4.251.2 ± 4.265.1 ± 5.162.4220.8 ± 4.641.3 ± 4.370.6 ± 4.068.46
UCE [11]41.8 ± 5.568.3 ± 2.782.7 ± 3.149.3221.1 ± 4.352.2 ± 3.561.0 ± 4.360.0151.1 ± 8.862.7 ± 2.987.5 ± 2.951.1218.9 ± 2.231.4 ± 2.886.1 ± 4.3277.88
Receler [18]8.1 ± 3.220 ± 3.465.4 ± 3.877.588.7 ± 2.245.3 ± 3.822.8 ± 3.841.044.2 ± 0.820.4 ± 3.736.7 ± 7.559.7010.3 ± 2.412.6 ± 3.157.5 ± 9.275.04
MACE [30]15.6 ± 6.437.7 ± 3.966.4 ± 4.369.8320.2 ± 4.636.7 ± 4.749.2 ± 4.861.6613.9 ± 3.228.2 ± 3.866.9 ± 5.674.1011.3 ± 2.528.9 ± 5.170.7 ± 5.375.98
AdvUnlearn [54]8.7 ± 2.939.1 ± 6.564.3 ± 5.769.8814.5 ± 4.137.4 ± 3.927.4 ± 4.047.757.4 ± 2.630.1 ± 5.260.1 ± 5.971.879.3 ± 2.227.6 ± 2.164.6 ± 5.974.41
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.486, + 0.892 + ], + "angle": 0, + "content": "onymous non-erased concepts, showing how well general information is erased. Additionally, we introduce Sensitivity (Sens.), defined as the accuracy for non-target but similar concepts. Most of the concept erasure techniques decreased accuracy on the target concept class (a desired outcome). Receler, MACE, and AdvUnlearn achieved the best efficacy in erasure overall per EraseBench dimension. Only a few techniques showed strong generality, indicating that erasure was not applied in a broad, global manner. In particular, Receler consistently displayed the best generality, preserving accuracy across target erased classes and paraphrased non-erased concepts. This may be attributed to the fact that Receler was trained using adversarial prompting and employs \"erasers\" designed to capture textual semantic concepts effectively. Overall, we observed a decrease in sensitivity, an undesirable outcome that highlights T2I misalignment for non-target non-erased concepts, essentially, over-erasure. After erasing a target concept, we noticed a clear mismatch between the text prompt and the generated image for related non-target concepts. Low sensitivity suggests that instead of selectively erasing only the target concept, the model inadvertently suppresses related concepts, causing generated images to fall out of distribution. This indicates that the model has not only erased the intended concept but also learned to overlook semantically or visually similar ones. As a result, the generated images fail to align with the intended prompts, revealing the model's inability to fully preserve non-target concepts after erasure. In Figure 2 (row 1), we observe that when models undergo erasure of the koala concept, prompting them with related concepts like Tree Kangaroo reveals unintended effect. These erased models appear to have also forgotten how to generate a Tree Kangaroo accurately. Similarly, when we erase the artist concept Bosch and prompt the sanitized model to generate artwork in the style of Altdorfer, the model struggles to capture the distinctive artistic style of the non-erased concept. This shows that the erasure process affects not only the target concept but also related artistic styles, leading to a loss of fidelity in the generated images. This demonstrates a significant vulnerability in these erasure tech" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.284, + 0.908, + 0.409 + ], + "angle": 0, + "content": "Table 2. Harmonic mean \\((\\uparrow)\\) of efficacy, generality, and sensitivity scores on explicit concepts across EraseBench dimensions. Each column represents a distinct relationship type: visually similar objects, artist style similarity, subset-superset relationships, and binomial pairs. Higher scores indicate better balance between erasure success and minimal unintended degradation. While all methods outperform the original model, performance varies by dimension, revealing trade-offs in handling sensitive and safety-critical content." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.42, + 0.905, + 0.502 + ], + "angle": 0, + "content": "
TechniqueVisual Sim.Subset-supersetBinomial
Original16.3723.4819.64
UCE [11]63.1662.6276.13
AdvUnlearn [54]67.6658.2871.33
Receler [18]68.4756.3174.88
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.597 + ], + "angle": 0, + "content": "Table 3. Text-to-image alignment scores using the Gecko metric. The average scores and the standard errors of the mean for images generated with different techniques and prompts that either contain the erased concept or not. \"Original\" is the baseline model w/o any concept erasure. Drop in score compared to the original model shown in brackets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.607, + 0.905, + 0.709 + ], + "angle": 0, + "content": "
TechniqueErased ConceptsNon-erased Concepts
Original84.1 ± 0.977.6 ± 0.7
UCE [11]57.6 ± 1.7 (−26.4)74.3 ± 0.8 (−3.4)
MACE [30]38.2 ± 1.3 (−45.9)67.9 ± 0.9 (−9.8)
AdvUnlearn [54]43.1 ± 1.4 (−41.0)68.6 ± 0.9 (−9.0)
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "niques and highlights their risks associated with deploying them in real-world unrestricted settings. This trend of low sensitivity is consistent across dimensions of EraseBench. If we want to evaluate overall, UCE demonstrates superior preservation of visual concept components compared to other methods. Table 2 presents the harmonic mean of efficacy, generality, and sensitivity scores for explicit concepts (see Table 8 in the supplemental material for details) across three EraseBench dimensions: Visual Similarity, Subset-superset, and Binomial. UCE demonstrates consistently strong performance across most dimensions, while AdvUnlearn and Receler achieve particularly high scores in" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16424" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.112, + 0.089, + 0.205, + 0.102 + ], + "angle": 0, + "content": "Original Image" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.103, + 0.221, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.271, + 0.096, + 0.302, + 0.106 + ], + "angle": 0, + "content": "UCE" + }, + { + "type": "image", + "bbox": [ + 0.223, + 0.108, + 0.348, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.376, + 0.096, + 0.451, + 0.107 + ], + "angle": 0, + "content": "AdvUnlearn" + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.108, + 0.476, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.213, + 0.17, + 0.226 + ], + "angle": 0, + "content": "Er: Goldfish" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.228, + 0.172, + 0.242 + ], + "angle": 0, + "content": "Pr: A Guppy" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.244, + 0.194, + 0.256 + ], + "angle": 0, + "content": "Artifact Scores:" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.259, + 0.182, + 0.272 + ], + "angle": 0, + "content": "Original:93.63" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.274, + 0.165, + 0.286 + ], + "angle": 0, + "content": "UCE: 84.21" + }, + { + "type": "image_footnote", + "bbox": [ + 0.093, + 0.289, + 0.204, + 0.301 + ], + "angle": 0, + "content": "AdvUnlearn:79.01" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.213, + 0.204, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.224, + 0.208, + 0.348, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.351, + 0.208, + 0.476, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.324, + 0.483, + 0.436 + ], + "angle": 0, + "content": "Figure 5. Erasure introduces artifacts during subset-superset concept generation. We erase concept \"goldfish\" and generate images for the prompt \"an image of a guppy\". We present the RAHF artifact heatmaps for images generated post-erasure via AdvUnlearn and UCE. We see that the artifact introduced by each method can vary spatially and by intensity, which prompts our inclusion of the artifact score in EraseBench. More heatmaps examples can be found in the supplemental material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.483, + 0.522 + ], + "angle": 0, + "content": "Subset-superset and visual similarity settings. These results highlight the value of evaluating concept erasure across diverse relational categories to ensure method robustness and general applicability." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.524, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Next, we use Gecko to evaluate image-text alignment with a VQA-based approach. Unlike CLIP, Gecko provides an interpretable alignment score as it is computed over question-answer pairs. Figure 4 shows an example of question-answer pairs for a pair of an original and a sanitized image. While some aspects of the prompt are still correctly depicted in the image, the score is lower for sanitized model as it is missing the animal. We compute alignment scores between images generated by original and sanitized models (UCE, MACE, and AdvUnlearn) and prompts corresponding to four different concepts (goat, goldfish, cat, and Vincent Van Gogh) from four different dimensions. In total, we compute alignment scores for 6246 text-image pairs. The results in Table 3 show an expected drop in scores for images generated with sanitized models compared to images generated with the original model (c.f. column \"Erased concepts\"), as we expect the erased concept to be missing in images generated by sanitized models. However, we also observe a drop in scores for non-erased concepts across different dimensions. While that drop is much smaller, it is statistically significant across all three comparisons (\\(\\alpha = .01\\) using the Wilcoxon signed-rank test). The UCE technique has the smallest gap with the original model for non-erased concepts, while MACE has the largest. The results for UCE are consistent with CLIP-score zero-shot" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.425 + ], + "angle": 0, + "content": "prediction in Table 1, where UCE was consistently the technique with highest sensitivity across all dimensions. We also manually inspected a subset of questions, answers and images, to ensure that the lower score is not caused by the VQA model latching on spurious correlations, but is indeed due to depicted concepts missing or the image being of lower quality. We observe that while some concepts are generally more difficult to depict correctly even in the original model (e.g. goldfish), this is reflected in the lower overall alignment score for the original model (e.g., for all three techniques the alignment score of the original model is in 60s for the concept goldfish, while it is in 80s and 90s for other concepts). Finally, question-answer breakdown in Gecko allows us to pinpoint some failures of image generation in sanitized models. When looking at answers to questions asking about object presence (eg., Is there an \\(\\langle\\) object \\(\\rangle\\) in the image?), we found that for all three techniques erased models had lower Gecko scores compared to original models on prompts with visually similar objects (UCE: \\(80.3\\%\\) vs \\(89.0\\%\\), MACE: \\(67.1\\%\\) vs \\(89.0\\%\\), and AdvUnlearn: \\(83.1\\%\\) vs \\(89.0\\%\\)), which highlights unintentional consequences of concept erasure techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.907, + 0.471 + ], + "angle": 0, + "content": "5.2. Concept Erasure Reduces Quality For Non-Target Concepts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Global Concepts. Beyond non-target misalignment, T2I models also exhibit other forms of degradation post-erasure, such as distortions (i.e. artifacts) and poor aesthetics. To explore this further, we specifically employed metrics that better align with human feedback, such as artifact measurements that quantify the degree of perceptible distortions and aesthetic quality (i.e. factors that reflect how humans perceive image fidelity). In Table 4, we present the RAHF alignment scores of individual concepts under different dimensions of EraseBench. Our results reveal a significant decline in overall quality compared to the original SD, suggesting that the erasure process leads to noticeable degradation in image quality. This indicates that similar and related concepts remain vulnerable to generation flaws, despite attempts at concept retention. In Figure 2, we observe that most models capable of retaining and generating the intended concept after erasure still exhibit various types of distortions. These include misaligned body parts for animal classes, cropped concepts, nonsensical text distortions, decreased size of the generated concept, and a general lack of sharpness. Additional examples can be found in the Supplemental Material. Artists Concepts. We observe a significant drop in T2I alignment, highlighting a major challenge in artistic style erasure as highlighted in Table 1. Similar styles or concepts across artists are at risk of being inadvertently erased if one artist's style is closely related to another. We observe this issue in the context of T2I alignment and found a marked degradation in quality, particularly in the" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16425" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.141, + 0.09, + 0.856, + 0.104 + ], + "angle": 0, + "content": "Table 4. RAHF alignment scores [28]. Alignment scores for each concept under the overall dimension object similarity." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.113, + 0.905, + 0.212 + ], + "angle": 0, + "content": "
TechniquesVisual Similarity (Object)BinomialSubset- superset
Erase "Cat"Erase "Goat"Erase "Lock"Erase "Goldfish"
Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑
Original87.71 ± 0.0780.93 ± 0.0283.23 ± 0.678.44 ± 0.285.23 ± 1.576.72 ± 0.884.15 ± 0.178.21 ± 0.3
UCE [11]72.44 ± 0.973.77 ± 0.275.00 ± 1.074.19 ± 0.378.88 ± 3.478.50 ± 1.074.55 ± 0.672.20 ± 0.4
MACE [30]73.50 ± 0.871.80 ± 0.0772.00 ± 1.173.40 ± 1.272.79 ±4.077.22 ± 0.772.64 ± 0.375.33 ± 0.2
AdvUnlearn [54]74.60 ± 0.972.30 ± 1.269.78 ± 0.670.10 ± 0.966.23 ± 5.376.79 ± 1.171.87 ± 0.572.49 ± 0.2
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.224, + 0.482, + 0.329 + ], + "angle": 0, + "content": "style of non-target concepts. In Figure 2, we show the erasure of \\(Degas\\) (the artist) and its impact on the ability to generate images of other artists, such as Cassatt. Notably, we see a significant shift in style even for the non-erased concept. We specifically chose not to use the RAHF alignment scores for artistic style, as the models were not trained to evaluate artistic style based on their documentation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.357, + 0.482, + 0.373 + ], + "angle": 0, + "content": "5.3. Empirical Validation with Human Preferences" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.381, + 0.483, + 0.623 + ], + "angle": 0, + "content": "We conduct a human preference study to additionally validate results obtained with automated evaluation techniques. We recruited 11 participants from our institution to judge images based on three different criteria: overall image quality, distortions and T2I alignment. All participants were instructed on how to complete the task, and have provided consent to participate in the study. Each participant was presented with 50 side-by-side image comparisons, corresponding to 50 pairs of images generated by the original model and by the model with a concept erased. The order of images was randomized for each pair during presentation. The participants were asked the following three questions: (i) Which image exhibits superior overall quality? (ii) Which of the following images displays LESS noticeable distortions? (iii) Which image most accurately reflects and is aligned with the text label?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Participants selected one of the three possible answers: Image A, Image B, or Neutral. In total, we collected 1650 responses for UCE and 1485 for AdvUnlearn. We focus on images of non-erased concepts, as the primary hypothesis we wanted to test was whether non-target visually similar concepts were negatively affected with the erasure technique. The results in Figure 6 show that humans judge images generated with the original model as having better overall quality and having fewer distortions. The original model is also preferred for better alignment, although it is closer to the \"Neutral\" for UCE, meaning that either both the original image and sanitized image were of equally good or equally poor quality. We provide additional human preferences for AdvUnlearn in the supplemental material. Results in Figure 10 also show that most participants preferred the original images with respect to quality, alignment, and less artifacts. Overall, we find that human preferences corroborate findings we observe with automatic metrics." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.221, + 0.891, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.351, + 0.905, + 0.406 + ], + "angle": 0, + "content": "Figure 6. Human image preferences between images generated by the original and the erased model. The erased model used is UCE. Results show that humans prefer SD over UCE. We provide the results of AdvUnlearn in the supplemental material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.424, + 0.627, + 0.439 + ], + "angle": 0, + "content": "6. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.906, + 0.583 + ], + "angle": 0, + "content": "In our work, we have identified and emphasized critical issues surrounding T2I models post concept erasure, making it our primary focus. Addressing these challenges is the logical next step, but it raises a complex research question that demands further exploration. In this section, we discuss into possible mitigation strategies and assess their impacts. We stress that finding effective solutions is an ongoing, nontrivial process that requires continued investigation and innovation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.854 + ], + "angle": 0, + "content": "The Role of Retain Sets in Mitigating Post-Erasure Distortions. A key question is whether the design of the retain set, specifically including concepts closely intertwined with the target erased concept can help alleviate post-generation distortions. We try to investigate whether adding non-target concepts that share semantic or visual similarities can reduce unintended side effects caused by erasure. To explore this, we performed erasure on targeted concepts using EraseBench while strategically incorporating related nontarget concepts into the retain set. For example, when erasing cat, we progressively introduce related concepts such as tiger and panther into the retain set and evaluate the impact on both generated outputs and artifact persistence. We show in Figure 7 the influence of the retain set on non-target concepts. We compare the original SD without erasure, erased model without a retain set (i.e. no explicit retain set), and erased model with a retain set in which we introduce the entangled non-target concepts one by one." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "(i) Findings on non-target concepts. Our investigation indicates that incorporating entangled non-target concepts into the retain set partially alleviates over-erasure (misalign" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16426" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.181 + ], + "angle": 0, + "content": "ment on non-target concepts), preserves aspects of related non-target concepts. However, artifacts and distortions persist, particularly in non-targeted non-erased concepts suggesting that while carefully designing a retain set may help maintain concept structure, it does not fully resolve the unintended distortions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.182, + 0.483, + 0.379 + ], + "angle": 0, + "content": "(ii) Findings on target-erased concepts. Examining the erased target concepts in Figure 8, we find that incorporating multiple entangled concepts into the retain set helps alleviate some post-erasure artifacts on the non-target concepts but comes at the cost of concept leakage. Specifically, this leakage manifests as spurious regeneration, where the erased concept partially or fully re-emerges in generated outputs. Our observations suggest that while including related non-target concepts in the retain set can mitigate over-erasure, it increases the risk of unintended regeneration. The more entangled concepts added, the greater the likelihood of leakage, highlighting a trade-off between preserving non-target concepts and ensuring effective erasure." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.664 + ], + "angle": 0, + "content": "The Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure. We investigate whether anchor concepts can help in mitigating the unintended distortions in non-target concepts observed after concept erasure. An anchor may act as a reference point, defining boundaries for selective erasure while preserving related concepts. To test this, we conducted experiments in Figure 9 where we erased Van Gogh while introducing Post-Impressionism as an anchor concept. Similarly, we erased Grumpy Cat while anchoring the model with Internet meme, as outlined in [8]. The intuition behind this approach is that a well-defined anchor might help the model retain a coherent representation of related concepts, thereby reducing unintended consequences of erasure. We observed models struggle with generating good quality non-target concept images, suggesting anchors do not consistently improve stability. While anchors may help in alignment, they do not fully resolve artifacts or style preservation issues, warranting further investigation into their effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.665, + 0.483, + 0.892 + ], + "angle": 0, + "content": "Multi-Concept Erasure: Intra-type and Inter-type Effects. We also investigate multi-concept erasure, differentiating between intra-type (main and paraphrased concepts) and inter-type (main and unrelated concepts) erasure. Average artifact scores in the visual similarity domain reach \\(78.3\\%\\) with intra-type and \\(71.4\\%\\) with inter-type erasure. The former outperforms the single concept erasure at \\(71.9\\%\\) (see example in Figure 15 in the supplemental material). Our findings show that intra-type erasure (i.e. removing multiple related concepts) more effectively eliminates targets and reduces artifacts when faced with non-target but related prompts. In contrast, inter-type erasure offers minimal artifact reduction, suggesting that carefully grouping related concepts enhances erasure precision without degrading generation quality." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.073, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.235, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Figure 7. Impact of adding visually similar concepts to the retain set with respect to non-target concepts. Without a retain set, image quality degrades significantly. Unrelated concepts (e.g. house, elephant, and boat) fail to prevent artifacts, while a curated, visually similar retain set improves quality but remains below pre-erasure levels." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.327, + 0.905, + 0.422 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.432, + 0.905, + 0.488 + ], + "angle": 0, + "content": "Figure 8. Impact of adding visually similar concepts to the retain set with respect to the target concept. Without a retain set, erasure is effective, but as similar concepts are added, the erased concept gradually re-emerges." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.506, + 0.691, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.506, + 0.887, + 0.584 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.595, + 0.905, + 0.623 + ], + "angle": 0, + "content": "Figure 9. Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.638, + 0.771, + 0.654 + ], + "angle": 0, + "content": "7. Conclusion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.665, + 0.907, + 0.888 + ], + "angle": 0, + "content": "In this work, we present a thorough evaluation of concept erasure techniques and uncover shortcomings in their reliability. By introducing EraseBench, we reveal critical gaps in current methods. Our findings suggest that despite recent advancements, existing concept erasure techniques fall short in preserving model quality, which call for the need for more robust and nuanced approaches as well as more robust evaluation protocol and metrics. While our framework provides a diverse evaluation, automating concept selection could improve scalability. Given dimensions like binomiality, researchers can explore broader concept spaces beyond our curation. Additionally, a deeper exploration of optimal retain or erasure sets to minimize distortions remains an open question. We hope EraseBench inspires new research directions toward developing more reliable concept erasure techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "16427" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.185, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Emanuele Bugliarello, Laurent Sartran, Aishwarya Agrawal, Lisa Anne Hendricks, and Aida Nematzadeh. Measuring progress in fine-grained vision-and-language understanding. arXiv preprint arXiv:2305.07558, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.242, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[3] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[4] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.41 + ], + "angle": 0, + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.411, + 0.483, + 0.506 + ], + "angle": 0, + "content": "[6] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.576 + ], + "angle": 0, + "content": "[7] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[8] Masane Fuchi and Tomohiro Takagi. Erasing concepts from text-to-image diffusion models with few-shot unlearning. arXiv preprint arXiv:2405.07288, 2, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[10] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 4, 5, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 4, 5, 7, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.859 + ], + "angle": 0, + "content": "[12] Roy Ganz and Michael Elad. Clipag: Towards generator-free text-to-image generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3843-3853, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[13] Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, and Weiming Zhang. Eraseanything: Enabling concept erasure in rectified" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "flow transformers. In Forty-second International Conference on Machine Learning, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[14] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36: 17170-17194, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.276, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[17] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.346, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[18] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. ECCV, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[19] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. In European conference on computer vision, pages 91-109. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.472, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[20] Imagen-Team-Google, Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, Hongliang Fei, Nando de Freitas, Yilin Gao, Evgeny Gladchenko, Sergio Gomez Colmenarejo, Mandy Guo, Alex Haig, Will Hawkins, Hexiang Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Yelin Kim, Ksenia Konyushkova, Karol Langner, Eric Lau, Shixin Luo, Soña Mokrá, Henna Nandwani, Yasumasa Onoe, Aaron van den Oord, Zarana Parekh, Jordi Pont-Tuset, Hang Qi, Rui Qian, Deepak Ramachandran, Poorva Rane, Abdullah Rashwan, Ali Razavi, Robert Riachi, Hansa Srinivasan, Srivatsan Srinivasan, Robin Strudel, Benigno Uria, Oliver Wang, Su Wang, Austin Waters, Chris Wolff, Auriel Wright, Zhisheng Xiao, Hao Xiong, Keyang Xu, Marc van Zee, Junlin Zhang, Katie Zhang, Wenlei Zhou, Konrad Zolna, Ola Aboubakar, Canfer Akbulut, Oscar Akerlund, Isabela Albuquerque, Nina Anderson, Marco Andretto, Lora Aroyo, Ben Bariach, David Barker, Sherry Ben, Dana Berman, Courtney Biles, Irina Blok, Pankil Botadra, Jenny Brennan, Karla Brown, John Buckley, Rudy Bunel, Elie Bursztein, Christina Butterfield, Ben Caine, Viral Carpenter, Norman Casagrande, Ming-Wei Chang, Solomon Chang, Shamik Chaudhuri, Tony Chen, John Choi, Dmitry Churbanau, Nathan Clement, Matan Cohen, Forrester Cole, Mikhail Dektiarev, Vincent Du, Praneet Dutta, Tom Eccles, Ndidi Elue, Ashley Feden, Shlomi Fruchter, Frankie Garcia, Roopal Garg, Weina Ge, Ahmed Ghazy, Bryant Gipson, Andrew Goodman, Dawid Górný, Sven Gowal, Khyatti Gupta, Yoni Halpern, Yena Han, Susan Hao, Jamie Hayes, Amir Hertz, Ed Hirst Tingbo Hou Heidi Howard" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16428" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.605 + ], + "angle": 0, + "content": "Mohamed Ibrahim, Dirichi Ike-Njoku, Joana Iljazi, Vlad Ionescu, William Isaac, Reena Jana, Gemma Jennings, Donovon Jenson, Xuhui Jia, Kerry Jones, Xiaoen Ju, Ivana Kajic, Christos Kaplanis, Burcu Karagol Ayan, Jacob Kelly, Suraj Kothawade, Christina Kouridi, Ira Ktena, Jolanda Kumakaw, Dana Kurniawan, Dmitry Lagun, Lily Lavitas, Jason Lee, Tao Li, Marco Liang, Maggie Li-Calis, Yuchi Liu, Javier Lopez Alberca, Peggy Lu, Kristian Lum, Yukun Ma, Chase Malik, John Mellor, Inbar Mosseri, Tom Murray, Aida Nematzadeh, Paul Nicholas, João Gabriel Oliveira, Guillermo Ortiz-Jimenez, Michela Paganini, Tom Le Paine, Roni Paiss, Alicia Parrish, Anne Peckham, Vikas Peswani, Igor Petrovski, Tobias Pfaff, Alex Pirozhenko, Ryan Poplin, Utsav Prabhu, Yuan Qi, Matthew Rahtz, Cyrus Rashtchian, Charvi Rastogi, Amit Raul, Ali Razavi, Sylvestre-Alvise Rebuffi, Susanna Ricco, Felix Riedel, Dirk Robinson, Pankaj Rohatgi, Bill Rosgen, Sarah Rumbley, Moonkyung Ryu, Anthony Salgado, Sahil Singla, Florian Schroff, Candice Schumann, Tanmay Shah, Brendan Shillingford, Kaushik Shivakumar, Dennis Shtatnov, Zach Singer, Evgeny Sluzhaev, Valerii Sokolov, Thibault Sottiaux, Florian Stimberg, Brad Stone, David Stutz, Yu-Chuan Su, Eric Tabellion, Shuai Tang, David Tao, Kurt Thomas, Gregory Thornton, Andeep Toor, Cristian Udrescu, Aayush Upadhyay, Cristina Vasconcelos, Alex Vasiloff, Andrey Voynov, Amanda Walker, Luyu Wang, Miaosen Wang, Simon Wang, Stanley Wang, Qifei Wang, Yuxiao Wang, Agoston Weisz, Olivia Wiles, Chenxia Wu, Xingyu Federico Xu, Andrew Xue, Jianbo Yang, Luo Yu, Mete Yurtoglu, Ali Zand, Han Zhang, Jia-geng Zhang, Catherine Zhao, Adilet Zhaxybay, Miao Zhou, Shengqi Zhu, Zhenkai Zhu, Dawn Bloxwich, Mahyar Bordbar, Luis C. Cobo, Eli Collins, Shengyang Dai, Tulsee Doshi, Anca Dragan, Douglas Eck, Demis Hassabis, Sissie Hsiao, Tom Hume, Koray Kavukcuoglu, Helen King, Jack Krawczyk, Yeqing Li, Kathy Meier-Hellstern, Andras Orban, Yury Pinsky, Amar Subramanya Oriol Vinyals Ting Yu and Yori Zwols. Imagen 3, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.607, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[21] Jun Young Koh, Sang Hyun Park, and Joy Song. Improving text generation on images with synthetic captions, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.635, + 0.483, + 0.704 + ], + "angle": 0, + "content": "[22] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.705, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[23] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.762, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[24] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R. Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9400-9409, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[25] Leyang Li, Shilin Lu, Yan Ren, and Adams Wai-Kin Kong. Set you straight: Auto-steering denoising trajectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[26] Ouxiang Li, Yuan Wang, Xinting Hu, Houcheng Jiang, Tao Liang, Yanbin Hao, Guojun Ma, and Fuli Feng. Speed: Scalable, precise, and efficient concept erasure for diffusion models. arXiv preprint arXiv:2503.07392, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[27] Zejian Li, Chenye Meng, Yize Li, Ling Yang, Shengyuan Zhang, Jiarui Ma, Jiayi Li, Guang Yang, Changyuan Yang, Zhiyuan Yang, et al. Laion-sg: An enhanced large-scale dataset for training complex image-text models with structural annotations. arXiv preprint arXiv:2412.08580, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.234, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[28] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19401-19411, 2024. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.319, + 0.904, + 0.388 + ], + "angle": 0, + "content": "[29] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[30] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.904, + 0.503 + ], + "angle": 0, + "content": "[31] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[32] Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.562, + 0.905, + 0.63 + ], + "angle": 0, + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[34] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[35] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.76, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "16429" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.483, + 0.247 + ], + "angle": 0, + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.483, + 0.416 + ], + "angle": 0, + "content": "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.485, + 0.488 + ], + "angle": 0, + "content": "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.483, + 0.559 + ], + "angle": 0, + "content": "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[44] Zhihua Tian, Sirun Nan, Ming Xu, Shengfang Zhai, Wenjie Qu, Jian Liu, Kui Ren, Ruoxi Jia, and Jiaheng Zhang. Sparse autoencoder as a zero-shot classifier for concept erasing in text-to-image diffusion models. arXiv preprint arXiv:2503.09446, 2025.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[45] Cristina Nader Vasconcelos, Abdullah Rashwan, Austin Waters, Trevor Walker, Keyang Xu, Jimmy Yan, Rui Qian, Yeqing Li, SHIXIN LUO, Yasumasa Onoe, Zarana Parekh, Ivana Kajic, Mandy Guo, Wenlei Zhou, Sarah Rosston, Roopal Garg, Hongliang Fei, Jordi Pont-Tuset, Su Wang, Henna Nandwani, Andrew Bunner, Kevin Swersky, David J. Fleet, Oliver Wang, and Jason Michael Baldridge. Greedy growing enables high-resolution pixel-based diffusion models. Transactions on Machine Learning Research, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.759, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[46] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for text-controlled vision models. arXiv preprint arXiv:2302.03693, 2, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.483, + 0.841 + ], + "angle": 0, + "content": "[47] WikiArt.org. WikiArt - Visual Art Encyclopedia. https: //www.wikiart.org/, 2024. Accessed: June 30, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[48] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with gecko: On met" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "rics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[49] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[50] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. Advances in Neural Information Processing Systems, 36, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[51] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[52] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755-1764, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.471 + ], + "angle": 0, + "content": "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.474, + 0.905, + 0.557 + ], + "angle": 0, + "content": "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 5, 7, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.559, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[55] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.696 + ], + "angle": 0, + "content": "[56] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.7, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[57] Xiaojin Zhu, Andrew B Goldberg, Mohamed Eldawy, Charles R Dyer, and Bradley Strock. A text-to-picture synthesis system for augmenting communication. In AAAI, pages 1590-1595, 2007. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.755 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "16430" + } + ] +] \ No newline at end of file diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_origin.pdf b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..f48b90b993dbf094a19d70b74cf1b99da7d1146c --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/872a71e9-733b-4e03-be37-36d9dbc5bdd0_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a6aa920d42638fc897279b90005b04a0c1d49d5627f92bb9e43748a21c8f42da +size 24246738 diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/full.md b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/full.md new file mode 100644 index 0000000000000000000000000000000000000000..41de9b2c15ead40c33c7c6e40ce124f3125f44f7 --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/full.md @@ -0,0 +1,256 @@ +# Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts + +Ibtihel Amara $^{1,2\dagger}$ , Ahmed Imtiaz Humayun $^{1,3\dagger}$ , Ivana Kajic $^{4}$ , Zarana Parekh $^{4}$ , Natalie Harris $^{1}$ , Sarah Young $^{1}$ , Chirag Nagpal $^{1}$ , Najoung Kim $^{4}$ , Junfeng He $^{1}$ , Cristina Nader Vasconcelos $^{4}$ , Deepak Ramachandran $^{4}$ , Golnoosh Farnadi $^{1}$ , Katherine Heller $^{1}$ , Mohammad Havaei $^{1}$ , Negar Rostamzadeh $^{1}$ + +1Google Research, 2McGill University, 3Rice University, 4Google Deepmind + +# Abstract + +Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality. + +# 1. Introduction + +As text-to-image (T2I) generative models [6, 20, 21, 38, 40, 45] have become increasingly popular and widely adopted, the demand for techniques that allow for the targeted removal of specific concepts has grown in parallel. Concept erasure methods [10, 11, 13, 18, 25, 26, 30, 32, 34, 44, 54] promise to address important concerns such as ensuring the safe deployment of models by removing undesirable or harmful concepts [10, 15, 52]. Whether applied to mitigate bias [4, 34, 46], protect privacy [10, 15, 22, 52], or filter out unwanted content, these techniques aim to create "sanitized" models capable of generating content without + +![](images/2ee506efc645bb2c28150078b2ca6f7042a72b114f53625c688d251993ab02a9.jpg) +Figure 1. The effects of concept erasure on non-target concepts. Pre-erasure outputs (left) vs. post-erasure results (right) from Stable Diffusion (SD). Erasure negatively impacts the quality of unrelated concepts. EraseBench identifies such effects and offers a framework to evaluate the reliability of erasure methods. + +triggering unwanted associations. However, the challenge of effectively erasing a concept without unintentionally distorting related or visually similar concepts remains unresolved. For example, erasing "cat" from a model should ideally remove all representations of a cat, yet leave semantically or visually related concepts like "lion" unaffected. Figure 1 shows various failure modes observed in state-of-the-art concept erasure techniques. These techniques exhibit instability after a concept is erased, often causing unintended consequences for closely related non-target concepts such as over-erasure and image artifacts (i.e. distortions). In this work, we argue that concept erasure techniques often struggle under real-world conditions. Even state-of-the-art methods fail when erasing concepts that are visually similar, binomially related, or semantically entangled. This lack of robustness raises concerns about their deployment in sen + +sitive applications, as unintended distortions of related non-target concepts affects the reliability of T2I models. + +While recent works assess erasure techniques using arbitrary concepts, a standardized framework for stress-testing across diverse concept relationships is still lacking. To address this, we introduce EraseBench, a unified benchmark with curated concepts and evaluations designed to rigorously test T2I models post-erasure. Covering binomial pairs, visually similar objects, and semantic relations. EraseBench enables consistent evaluation of erasure effectiveness, unintended distortions, and concept leakage. Our findings highlight persistent challenges and the need for more robust and standardized testing. We envision EraseBench as a foundation for improving comparability and advancing safer and more reliable erasure methods. Additionally, our results demonstrate that these sanitized models are not yet ready for deployment, highlighting the need for more robust and reliable erasure techniques that can operate effectively in complex, real-world scenarios. + +Our main contributions are summarized as follows: + +1. We identify key evaluation dimensions where most concept erasure techniques exhibit vulnerabilities, particularly when handling visually similar, binomial, and semantically related concepts. +2. We present EraseBench, a comprehensive, multidimensional framework designed to rigorously evaluate the robustness and efficacy of concept erasure techniques across a diverse set of concepts and prompts. +3. We use a suite of evaluation metrics to measure the effectiveness of concept erasure methods that provides a more holistic assessment of their performance. +4. We evaluate five state-of-the-art concept erasure techniques on EraseBench and demonstrate that many of the current concept erasure techniques are not yet ready for real-world deployment, highlighting significant gaps in their reliability and robustness. + +# 2. Related Works + +Concept Erasure. T2I models have witnessed remarkable advancements since the foundational works of [31, 57], driven by breakthroughs in model architectures [5, 19, 42, 49, 53, 56], state-of-the-art generative modeling techniques [1, 3, 12, 16, 23, 24, 29, 29, 33, 37-40, 50], and the availability of large high-quality datasets [27, 41]. These improvements have significantly enhanced their capability to generate realistic and diverse images from textual descriptions, pushing the boundaries of creativity and application. In light of recent advancements in T2I models, greater attention has been directed towards reliability considerations. There is a wide variety of concept erasure techniques, each utilizing distinct methods and technologies to achieve concept removal [7, 9-11, 14, 18, 22, 30, 30, 54]. + +Challenges in Concept Erasure. Despite the rapid in + +crease of these concept erasure methods, only a few works have critically examined the inherent limitations and vulnerabilities of concept erasure approaches. In particular, the work in [35] explores a fundamental vulnerability, which is the potential for erased concepts to be recovered through specific prompting. Meanwhile, other works, such as in [55], assess the effectiveness of concept erasure by examining a diverse range of objects and styles. Their approach evaluates whether the erased concepts have been completely removed from the model's generative capabilities, ensuring that no traces remain in the generated outputs. In our work, we delve into the broader impact of concept erasure, not only evaluating residual concept generation but also examining how erasure techniques may degrade the quality and fidelity of non-target concepts. In this work, we highlight the ripple effects of concept erasure on model performance and the stability of related concepts via the route of interrelated concepts. + +# 3. EraseBench + +In this section, we provide a detailed breakdown of the process used to construct our benchmark. + +Concept Gathering. We utilized the semantic knowledge embedded in large language models (LLMs) and structured sources, like the hierarchical taxonomy of ImageNet, to generate clusters of closely related concepts. By treating these language models akin to graph neural networks, we mapped similar concepts that share key attributes and gathered distinct sets for each dimension. This approach captures nuanced relationships crucial for testing erasure precision. For artistic styles, we selected artist names from the WikiArt dataset [47]. + +Concept Verification. To ensure quality and feasibility, we conducted human verification on the collected concepts. This included assessing whether existing T2I models could successfully generate accurate, high-quality representations of each concept. During this process, we identified and removed any concepts that models consistently struggled to render, preserving only those concepts that aligned with our benchmarks. This step ensures that the dataset maintains a high standard, with each concept suitable for rigorous testing of erasure techniques. + +Prompt Construction. We used Gemini Flash [43], to generate a diverse set of prompts for each concept. Prompts were crafted with variations in style, complexity, and length to allow for thorough evaluation. This includes prompts that differ in level of detail (simple vs. elaborate descriptions) and length (short vs. extended prompts). With this range, we enable a finer assessment of concept erasure techniques. + +Concept Dimensions. We introduce multiple evaluation dimensions, each capturing a distinct type of concept relationship or similarity, providing a structured way to evaluate the impact of erasure. Within each dimension, we pro + +![](images/c12e62b9cb30586992e201acb12108a1b6e6f8415d0c61d0c201a34c113a0aa3.jpg) +Figure 2. Ripple effects of concept erasure methods across EraseBench entanglement dimensions. All the erasure baselines display failure cases across different EraseBench tasks. Receler and MACE frequently produce images that are unrelated to the text prompt, indicating they are the most sensitive of the five concept erasure techniques. In contrast, AdvUnlearn shows slightly better robustness across certain dimensions of the benchmark. + +![](images/4ee4c688bd9440dfee2a1a80929d28630c22ae348d9279e1c8f954bcc0368ef6.jpg) +Figure 3. Evaluation Dimensions of EraseBench. + +vide a diverse set of concepts that challenge the model in different ways, ensuring a thorough evaluation. Figure 3 summarizes these dimensions in EraseBench. For each, we define a primary concept for erasure and related non-target concepts sharing visual, semantic, or contextual similari- + +ties to test unintended effects on the sanitized model. (1) Visual Similarity (Objects), A broad concept (e.g., "cat") is erased, along with synonyms ("kitten," "tabby"), while visually similar non-targets ("tiger," "cheetah") test unintended erasure. (2) Style Similarity (Artists), Erasing an artist (e.g., Van Gogh) tests model behavior on similar-style artists (e.g., Cézanne, Bernard). (3) Subset-Superset Relationship, Removing a concept (e.g., "goldfish") evaluates retention of related non-targets ("guppy," "koi"). (4) Binomial Relations, Erasing one of a closely linked pair (e.g., "sun" from "sun and moon") tests impact on the other. + +Explicit Content. Under each dimension of EraseBench, we also integrate explicit concepts to evaluate the handling of sensitive and safety-critical content (details can be found in Table 8 in the supplemental material). These include + +categories such as NSFW, culturally related, and broader safety-related themes. Incorporating these explicit concepts allows for a more rigorous assessment of how well erasure methods balance effectiveness, generation quality, and sensitivity. + +# 4. Experimental setup + +Baseline models. In our work, we narrowed down existing techniques and focused on leveraging more recent, advanced and diverse approaches to concept erasure, experimenting with recent techniques that incorporate variations such as fine-tuning model weights (e.g. ESD [10]), introducing targeted weight perturbations (e.g. UCE [11]), refining textual embeddings and adversarial training (e.g. Receler [18], AdvUnlearn [54]), and introducing parameter efficient fine-tuning (Receler, MACE [30]). This allows us to explore the nuanced dynamics of concept erasure and assess its impact under different dimensions. We provide detailed description about each of these techniques in the supplemental material. + +Experiments. We use Stable Diffusion (SD) as our T2I model and apply concept erasure techniques to each primary concept in EraseBench. For each concept, we perform erasure using both the default and best-reported settings from baseline methods. We maintain the default retain sets (e.g. COCO prompts for AdvUnlearn) and, for methods requiring anchor concepts, we map them to an empty concept ("). We evaluate the sanitized models on related non-target concepts within the same EraseBench dimension, assessing the effects on visually and semantically similar concepts. Each concept is represented by over 10 unique prompts, ranging from simple to complex descriptions. To capture diverse outputs, we generate 10 images per prompt with different random seeds. This approach provides a comprehensive evaluation of erasure effectiveness and its unintended effects on related concepts. + +Evaluation Metrics. We use different types of automated evaluation metrics to quantify and explain the detrimental effect erasure techniques can have on the quality of generated images of non-target concepts. We use three different automated evaluation metrics to quantify the post erasure performance: CLIP [36], Rich Automatic Human Feedback (RAHF [28]), and Gecko [48]. CLIP is a widely used joint image and text embedding model which quantifies the similarity between text and image representations. We leverage CLIP as a zero-shot classifier to determine whether a model generates the desired concept. The RAHF model is trained on rich human feedback for generated images with a range of scores to evaluate overall image quality, and also mark points about which regions are problematic. The RAHF quality scores include an aesthetic score, which assesses the visual appeal of an image, and an artifact score (i.e. plausibility score), which gauges the presence and extent + +of visual artifacts. We use both the aesthetic and artifact scores as measurements of quality for generated images, following [28], where higher scores indicate better quality (i.e. good aesthetics and minimal artifacts). In addition to scores, the RAHF model provides heatmaps for each image to highlight artifact-prone areas, which we use for visual inspection in our benchmark. Finally, we use a Visual Question Answering (VQA)-based metric called Gecko, a VQA evaluation framework. In our setup, we leverage the multimodal Gemini 1.5 model for both steps: first, to generate text-related questions, and second, to answer them based on the given image, similar to other VQA methods [2, 51]. One advantage of such metrics is that they do not require task-specific fine-tuning and provide an interpretable alignment score. This evaluation can be replicated using other available models. Typically, an LLM generates questions related to the text prompt in the first step, while a vision-language model (VLM) answers those questions based on the generated image. The final score is computed as the average of correctly answered questions, allowing for backtracing to identify which aspects of the text are misaligned with the generated image. + +# 5. Results + +# 5.1. Concept Erasure Reduces T2I Alignment For Non-Target Concepts (Over-Erasure) + +![](images/340ad0189edabfbf75c37c294d78220a7ae1c5405fbf5ae0205ba7bf16f8bd55.jpg) +Q1: Is there a tiger in the image? + +B +![](images/86cfd8f20e7ca061ec337942e991d09e0805407ace5698d9e80565a9781b377b.jpg) +Q2: Is the tiger perched on something? + +Alignment score + +50 +Figure 4. Erasure affects fine-grained alignment. Prompt: "A tiger perched on a rocky outcrop surrounded by mountains and a serene blue sky." before (left) and after (right) erasing the concept cat using MACE [30]). +![](images/4b6e1b5ff9c3ec449dcf77154277f2dbb165f9867eb40fb144bf205dbe4a11d6.jpg) +Q3: Is the tiger perched on a rocky outcrop? +Q4: Are there mountains in the image? +Q5: Is the sky a serene blue color? +Q6: Is there a sky visible in the image? + +Table 1 presents the zero-shot predictions from CLIP as percentages (\%). Following the definitions in [30], we measure two key attributes: Efficacy (Eff.) and Generality (Gen.). Efficacy represents the accuracy of concept erasure (accuracy on the erased class), where lower values indicate better erasure. Generality reflects accuracy on paraphrased or syn + +Table 1. Evaluation of Erasing concepts from EraseBench under the four different dimensions. We provide the average results across 10 concepts (for visual similarity (object)), 15 artists concepts (for artists similarity), 8 concepts for subset-superset, and 10 concepts (for binomial). We evaluate efficacy (Eff., capturing the efficacy of erasing the main concept), generality (Gen., capturing the impact on paraphrased or related concepts), and sensitivity (Sens., capturing the unintended effect on non-target yet related or similar concepts). All values are expressed as percentages (\%). More detailed, per-concept breakdowns are provided in the Supplemental Material. For reference, we also include the classification accuracy of images generated by the original Stable Diffusion v1.4 model. + +
TechniquesVisual Similarity (Object)Artistic Style Similarity (Artists)Subset-SupersetBinomial
Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑
Original86.5 ± 7.190.2 ± 3.685.0 ± 4.015.9772.3 ± 6.380.7 ± 3.269.0 ± 4.329.2983.3 ± 8.992.8 ± 3.992.1 ± 3.514.3188.3 ± 5.289.8 ± 4.388.5 ± 4.515.40
ESD [10]24.5 ± 6.150.5 ± 4.165.9 ± 4.661.7015.1 ± 3.761.9 ± 4.840.3 ± 3.947.7418.1 ± 4.251.2 ± 4.265.1 ± 5.162.4220.8 ± 4.641.3 ± 4.370.6 ± 4.068.46
UCE [11]41.8 ± 5.568.3 ± 2.782.7 ± 3.149.3221.1 ± 4.352.2 ± 3.561.0 ± 4.360.0151.1 ± 8.862.7 ± 2.987.5 ± 2.951.1218.9 ± 2.231.4 ± 2.886.1 ± 4.3277.88
Receler [18]8.1 ± 3.220 ± 3.465.4 ± 3.877.588.7 ± 2.245.3 ± 3.822.8 ± 3.841.044.2 ± 0.820.4 ± 3.736.7 ± 7.559.7010.3 ± 2.412.6 ± 3.157.5 ± 9.275.04
MACE [30]15.6 ± 6.437.7 ± 3.966.4 ± 4.369.8320.2 ± 4.636.7 ± 4.749.2 ± 4.861.6613.9 ± 3.228.2 ± 3.866.9 ± 5.674.1011.3 ± 2.528.9 ± 5.170.7 ± 5.375.98
AdvUnlearn [54]8.7 ± 2.939.1 ± 6.564.3 ± 5.769.8814.5 ± 4.137.4 ± 3.927.4 ± 4.047.757.4 ± 2.630.1 ± 5.260.1 ± 5.971.879.3 ± 2.227.6 ± 2.164.6 ± 5.974.41
+ +onymous non-erased concepts, showing how well general information is erased. Additionally, we introduce Sensitivity (Sens.), defined as the accuracy for non-target but similar concepts. Most of the concept erasure techniques decreased accuracy on the target concept class (a desired outcome). Receler, MACE, and AdvUnlearn achieved the best efficacy in erasure overall per EraseBench dimension. Only a few techniques showed strong generality, indicating that erasure was not applied in a broad, global manner. In particular, Receler consistently displayed the best generality, preserving accuracy across target erased classes and paraphrased non-erased concepts. This may be attributed to the fact that Receler was trained using adversarial prompting and employs "erasers" designed to capture textual semantic concepts effectively. Overall, we observed a decrease in sensitivity, an undesirable outcome that highlights T2I misalignment for non-target non-erased concepts, essentially, over-erasure. After erasing a target concept, we noticed a clear mismatch between the text prompt and the generated image for related non-target concepts. Low sensitivity suggests that instead of selectively erasing only the target concept, the model inadvertently suppresses related concepts, causing generated images to fall out of distribution. This indicates that the model has not only erased the intended concept but also learned to overlook semantically or visually similar ones. As a result, the generated images fail to align with the intended prompts, revealing the model's inability to fully preserve non-target concepts after erasure. In Figure 2 (row 1), we observe that when models undergo erasure of the koala concept, prompting them with related concepts like Tree Kangaroo reveals unintended effect. These erased models appear to have also forgotten how to generate a Tree Kangaroo accurately. Similarly, when we erase the artist concept Bosch and prompt the sanitized model to generate artwork in the style of Altdorfer, the model struggles to capture the distinctive artistic style of the non-erased concept. This shows that the erasure process affects not only the target concept but also related artistic styles, leading to a loss of fidelity in the generated images. This demonstrates a significant vulnerability in these erasure tech + +Table 2. Harmonic mean $(\uparrow)$ of efficacy, generality, and sensitivity scores on explicit concepts across EraseBench dimensions. Each column represents a distinct relationship type: visually similar objects, artist style similarity, subset-superset relationships, and binomial pairs. Higher scores indicate better balance between erasure success and minimal unintended degradation. While all methods outperform the original model, performance varies by dimension, revealing trade-offs in handling sensitive and safety-critical content. + +
TechniqueVisual Sim.Subset-supersetBinomial
Original16.3723.4819.64
UCE [11]63.1662.6276.13
AdvUnlearn [54]67.6658.2871.33
Receler [18]68.4756.3174.88
+ +Table 3. Text-to-image alignment scores using the Gecko metric. The average scores and the standard errors of the mean for images generated with different techniques and prompts that either contain the erased concept or not. "Original" is the baseline model w/o any concept erasure. Drop in score compared to the original model shown in brackets. + +
TechniqueErased ConceptsNon-erased Concepts
Original84.1 ± 0.977.6 ± 0.7
UCE [11]57.6 ± 1.7 (−26.4)74.3 ± 0.8 (−3.4)
MACE [30]38.2 ± 1.3 (−45.9)67.9 ± 0.9 (−9.8)
AdvUnlearn [54]43.1 ± 1.4 (−41.0)68.6 ± 0.9 (−9.0)
+ +niques and highlights their risks associated with deploying them in real-world unrestricted settings. This trend of low sensitivity is consistent across dimensions of EraseBench. If we want to evaluate overall, UCE demonstrates superior preservation of visual concept components compared to other methods. Table 2 presents the harmonic mean of efficacy, generality, and sensitivity scores for explicit concepts (see Table 8 in the supplemental material for details) across three EraseBench dimensions: Visual Similarity, Subset-superset, and Binomial. UCE demonstrates consistently strong performance across most dimensions, while AdvUnlearn and Receler achieve particularly high scores in + +Original Image +![](images/03f02e19b6359ad849ba68ca742b828317687d1fed841461a13a1ec1de780b1b.jpg) +Er: Goldfish + +UCE +![](images/d4f5915c06971e2de571135cc7e5a4eb175925ddb600a9c6d3a7dd1898c17887.jpg) +Pr: A Guppy + +![](images/89734462cff06f4d4824c3ad95b18a477bdf021b50dddd787c4cd0009bb18ddf.jpg) +AdvUnlearn + +Figure 5. Erasure introduces artifacts during subset-superset concept generation. We erase concept "goldfish" and generate images for the prompt "an image of a guppy". We present the RAHF artifact heatmaps for images generated post-erasure via AdvUnlearn and UCE. We see that the artifact introduced by each method can vary spatially and by intensity, which prompts our inclusion of the artifact score in EraseBench. More heatmaps examples can be found in the supplemental material. +![](images/56bae0e5407c2dd75304e7c523b061ef449a6ca964c8a1e4498b5254b683dac0.jpg) +Artifact Scores: +Original:93.63 +UCE: 84.21 +AdvUnlearn:79.01 + +![](images/e9b0cd298957c065681039aac20e5d3cda7fb3f9624f132b61b9cebaccb6399b.jpg) + +Subset-superset and visual similarity settings. These results highlight the value of evaluating concept erasure across diverse relational categories to ensure method robustness and general applicability. + +Next, we use Gecko to evaluate image-text alignment with a VQA-based approach. Unlike CLIP, Gecko provides an interpretable alignment score as it is computed over question-answer pairs. Figure 4 shows an example of question-answer pairs for a pair of an original and a sanitized image. While some aspects of the prompt are still correctly depicted in the image, the score is lower for sanitized model as it is missing the animal. We compute alignment scores between images generated by original and sanitized models (UCE, MACE, and AdvUnlearn) and prompts corresponding to four different concepts (goat, goldfish, cat, and Vincent Van Gogh) from four different dimensions. In total, we compute alignment scores for 6246 text-image pairs. The results in Table 3 show an expected drop in scores for images generated with sanitized models compared to images generated with the original model (c.f. column "Erased concepts"), as we expect the erased concept to be missing in images generated by sanitized models. However, we also observe a drop in scores for non-erased concepts across different dimensions. While that drop is much smaller, it is statistically significant across all three comparisons ( $\alpha = .01$ using the Wilcoxon signed-rank test). The UCE technique has the smallest gap with the original model for non-erased concepts, while MACE has the largest. The results for UCE are consistent with CLIP-score zero-shot + +prediction in Table 1, where UCE was consistently the technique with highest sensitivity across all dimensions. We also manually inspected a subset of questions, answers and images, to ensure that the lower score is not caused by the VQA model latching on spurious correlations, but is indeed due to depicted concepts missing or the image being of lower quality. We observe that while some concepts are generally more difficult to depict correctly even in the original model (e.g. goldfish), this is reflected in the lower overall alignment score for the original model (e.g., for all three techniques the alignment score of the original model is in 60s for the concept goldfish, while it is in 80s and 90s for other concepts). Finally, question-answer breakdown in Gecko allows us to pinpoint some failures of image generation in sanitized models. When looking at answers to questions asking about object presence (eg., Is there an $\langle$ object $\rangle$ in the image?), we found that for all three techniques erased models had lower Gecko scores compared to original models on prompts with visually similar objects (UCE: $80.3\%$ vs $89.0\%$ , MACE: $67.1\%$ vs $89.0\%$ , and AdvUnlearn: $83.1\%$ vs $89.0\%$ ), which highlights unintentional consequences of concept erasure techniques. + +# 5.2. Concept Erasure Reduces Quality For Non-Target Concepts + +Global Concepts. Beyond non-target misalignment, T2I models also exhibit other forms of degradation post-erasure, such as distortions (i.e. artifacts) and poor aesthetics. To explore this further, we specifically employed metrics that better align with human feedback, such as artifact measurements that quantify the degree of perceptible distortions and aesthetic quality (i.e. factors that reflect how humans perceive image fidelity). In Table 4, we present the RAHF alignment scores of individual concepts under different dimensions of EraseBench. Our results reveal a significant decline in overall quality compared to the original SD, suggesting that the erasure process leads to noticeable degradation in image quality. This indicates that similar and related concepts remain vulnerable to generation flaws, despite attempts at concept retention. In Figure 2, we observe that most models capable of retaining and generating the intended concept after erasure still exhibit various types of distortions. These include misaligned body parts for animal classes, cropped concepts, nonsensical text distortions, decreased size of the generated concept, and a general lack of sharpness. Additional examples can be found in the Supplemental Material. Artists Concepts. We observe a significant drop in T2I alignment, highlighting a major challenge in artistic style erasure as highlighted in Table 1. Similar styles or concepts across artists are at risk of being inadvertently erased if one artist's style is closely related to another. We observe this issue in the context of T2I alignment and found a marked degradation in quality, particularly in the + +Table 4. RAHF alignment scores [28]. Alignment scores for each concept under the overall dimension object similarity. + +
TechniquesVisual Similarity (Object)BinomialSubset- superset
Erase "Cat"Erase "Goat"Erase "Lock"Erase "Goldfish"
Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑
Original87.71 ± 0.0780.93 ± 0.0283.23 ± 0.678.44 ± 0.285.23 ± 1.576.72 ± 0.884.15 ± 0.178.21 ± 0.3
UCE [11]72.44 ± 0.973.77 ± 0.275.00 ± 1.074.19 ± 0.378.88 ± 3.478.50 ± 1.074.55 ± 0.672.20 ± 0.4
MACE [30]73.50 ± 0.871.80 ± 0.0772.00 ± 1.173.40 ± 1.272.79 ±4.077.22 ± 0.772.64 ± 0.375.33 ± 0.2
AdvUnlearn [54]74.60 ± 0.972.30 ± 1.269.78 ± 0.670.10 ± 0.966.23 ± 5.376.79 ± 1.171.87 ± 0.572.49 ± 0.2
+ +style of non-target concepts. In Figure 2, we show the erasure of $Degas$ (the artist) and its impact on the ability to generate images of other artists, such as Cassatt. Notably, we see a significant shift in style even for the non-erased concept. We specifically chose not to use the RAHF alignment scores for artistic style, as the models were not trained to evaluate artistic style based on their documentation. + +# 5.3. Empirical Validation with Human Preferences + +We conduct a human preference study to additionally validate results obtained with automated evaluation techniques. We recruited 11 participants from our institution to judge images based on three different criteria: overall image quality, distortions and T2I alignment. All participants were instructed on how to complete the task, and have provided consent to participate in the study. Each participant was presented with 50 side-by-side image comparisons, corresponding to 50 pairs of images generated by the original model and by the model with a concept erased. The order of images was randomized for each pair during presentation. The participants were asked the following three questions: (i) Which image exhibits superior overall quality? (ii) Which of the following images displays LESS noticeable distortions? (iii) Which image most accurately reflects and is aligned with the text label? + +Participants selected one of the three possible answers: Image A, Image B, or Neutral. In total, we collected 1650 responses for UCE and 1485 for AdvUnlearn. We focus on images of non-erased concepts, as the primary hypothesis we wanted to test was whether non-target visually similar concepts were negatively affected with the erasure technique. The results in Figure 6 show that humans judge images generated with the original model as having better overall quality and having fewer distortions. The original model is also preferred for better alignment, although it is closer to the "Neutral" for UCE, meaning that either both the original image and sanitized image were of equally good or equally poor quality. We provide additional human preferences for AdvUnlearn in the supplemental material. Results in Figure 10 also show that most participants preferred the original images with respect to quality, alignment, and less artifacts. Overall, we find that human preferences corroborate findings we observe with automatic metrics. + +![](images/d6a3f2f71ca49af6b4282fc356124d1bf0a9eedf590d73381e70f616808f6406.jpg) +Figure 6. Human image preferences between images generated by the original and the erased model. The erased model used is UCE. Results show that humans prefer SD over UCE. We provide the results of AdvUnlearn in the supplemental material. + +# 6. Discussion + +In our work, we have identified and emphasized critical issues surrounding T2I models post concept erasure, making it our primary focus. Addressing these challenges is the logical next step, but it raises a complex research question that demands further exploration. In this section, we discuss into possible mitigation strategies and assess their impacts. We stress that finding effective solutions is an ongoing, nontrivial process that requires continued investigation and innovation. + +The Role of Retain Sets in Mitigating Post-Erasure Distortions. A key question is whether the design of the retain set, specifically including concepts closely intertwined with the target erased concept can help alleviate post-generation distortions. We try to investigate whether adding non-target concepts that share semantic or visual similarities can reduce unintended side effects caused by erasure. To explore this, we performed erasure on targeted concepts using EraseBench while strategically incorporating related nontarget concepts into the retain set. For example, when erasing cat, we progressively introduce related concepts such as tiger and panther into the retain set and evaluate the impact on both generated outputs and artifact persistence. We show in Figure 7 the influence of the retain set on non-target concepts. We compare the original SD without erasure, erased model without a retain set (i.e. no explicit retain set), and erased model with a retain set in which we introduce the entangled non-target concepts one by one. + +(i) Findings on non-target concepts. Our investigation indicates that incorporating entangled non-target concepts into the retain set partially alleviates over-erasure (misalign + +ment on non-target concepts), preserves aspects of related non-target concepts. However, artifacts and distortions persist, particularly in non-targeted non-erased concepts suggesting that while carefully designing a retain set may help maintain concept structure, it does not fully resolve the unintended distortions. + +(ii) Findings on target-erased concepts. Examining the erased target concepts in Figure 8, we find that incorporating multiple entangled concepts into the retain set helps alleviate some post-erasure artifacts on the non-target concepts but comes at the cost of concept leakage. Specifically, this leakage manifests as spurious regeneration, where the erased concept partially or fully re-emerges in generated outputs. Our observations suggest that while including related non-target concepts in the retain set can mitigate over-erasure, it increases the risk of unintended regeneration. The more entangled concepts added, the greater the likelihood of leakage, highlighting a trade-off between preserving non-target concepts and ensuring effective erasure. + +The Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure. We investigate whether anchor concepts can help in mitigating the unintended distortions in non-target concepts observed after concept erasure. An anchor may act as a reference point, defining boundaries for selective erasure while preserving related concepts. To test this, we conducted experiments in Figure 9 where we erased Van Gogh while introducing Post-Impressionism as an anchor concept. Similarly, we erased Grumpy Cat while anchoring the model with Internet meme, as outlined in [8]. The intuition behind this approach is that a well-defined anchor might help the model retain a coherent representation of related concepts, thereby reducing unintended consequences of erasure. We observed models struggle with generating good quality non-target concept images, suggesting anchors do not consistently improve stability. While anchors may help in alignment, they do not fully resolve artifacts or style preservation issues, warranting further investigation into their effectiveness. + +Multi-Concept Erasure: Intra-type and Inter-type Effects. We also investigate multi-concept erasure, differentiating between intra-type (main and paraphrased concepts) and inter-type (main and unrelated concepts) erasure. Average artifact scores in the visual similarity domain reach $78.3\%$ with intra-type and $71.4\%$ with inter-type erasure. The former outperforms the single concept erasure at $71.9\%$ (see example in Figure 15 in the supplemental material). Our findings show that intra-type erasure (i.e. removing multiple related concepts) more effectively eliminates targets and reduces artifacts when faced with non-target but related prompts. In contrast, inter-type erasure offers minimal artifact reduction, suggesting that carefully grouping related concepts enhances erasure precision without degrading generation quality. + +![](images/c97d047a46c55dcef46adcc1774602d209664816cc1050cb5a1a83a62d4a1349.jpg) +Figure 7. Impact of adding visually similar concepts to the retain set with respect to non-target concepts. Without a retain set, image quality degrades significantly. Unrelated concepts (e.g. house, elephant, and boat) fail to prevent artifacts, while a curated, visually similar retain set improves quality but remains below pre-erasure levels. + +![](images/8249c162b0b8457e088a7a39bf84e9b1c191cb94a54034c4dd533e13ef87b39c.jpg) +Figure 8. Impact of adding visually similar concepts to the retain set with respect to the target concept. Without a retain set, erasure is effective, but as similar concepts are added, the erased concept gradually re-emerges. + +![](images/be64878cb49855bbc2d2c9ce620db014d0ddd0a35ebcc6c56e7f1d640ceac727.jpg) +Figure 9. Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure. + +![](images/b005257329c1238324ca51a6380da723bb0b6a0d40c96d58384fbc1840f0b556.jpg) + +# 7. Conclusion and Limitations + +In this work, we present a thorough evaluation of concept erasure techniques and uncover shortcomings in their reliability. By introducing EraseBench, we reveal critical gaps in current methods. Our findings suggest that despite recent advancements, existing concept erasure techniques fall short in preserving model quality, which call for the need for more robust and nuanced approaches as well as more robust evaluation protocol and metrics. While our framework provides a diverse evaluation, automating concept selection could improve scalability. Given dimensions like binomiality, researchers can explore broader concept spaces beyond our curation. Additionally, a deeper exploration of optimal retain or erasure sets to minimize distortions remains an open question. We hope EraseBench inspires new research directions toward developing more reliable concept erasure techniques. + +# References + +[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2 +[2] Emanuele Bugliarello, Laurent Sartran, Aishwarya Agrawal, Lisa Anne Hendricks, and Aida Nematzadeh. Measuring progress in fine-grained vision-and-language understanding. arXiv preprint arXiv:2305.07558, 2023. 4 +[3] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2 +[4] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 1 +[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2 +[6] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1 +[7] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2 +[8] Masane Fuchi and Tomohiro Takagi. Erasing concepts from text-to-image diffusion models with few-shot unlearning. arXiv preprint arXiv:2405.07288, 2, 2024. 8 +[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 6 +[10] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 4, 5, 3, 6 +[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 4, 5, 7, 3, 6 +[12] Roy Ganz and Michael Elad. Clipag: Towards generator-free text-to-image generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3843-3853, 2024. 2 +[13] Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, and Weiming Zhang. Eraseanything: Enabling concept erasure in rectified + +flow transformers. In Forty-second International Conference on Machine Learning, 2025. 1 +[14] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36: 17170-17194, 2023. 2 +[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2 +[17] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4 +[18] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. ECCV, 2024. 1, 2, 4, 5, 6 +[19] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. In European conference on computer vision, pages 91-109. Springer, 2022. 2 +[20] Imagen-Team-Google, Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, Hongliang Fei, Nando de Freitas, Yilin Gao, Evgeny Gladchenko, Sergio Gomez Colmenarejo, Mandy Guo, Alex Haig, Will Hawkins, Hexiang Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Yelin Kim, Ksenia Konyushkova, Karol Langner, Eric Lau, Shixin Luo, Soña Mokrá, Henna Nandwani, Yasumasa Onoe, Aaron van den Oord, Zarana Parekh, Jordi Pont-Tuset, Hang Qi, Rui Qian, Deepak Ramachandran, Poorva Rane, Abdullah Rashwan, Ali Razavi, Robert Riachi, Hansa Srinivasan, Srivatsan Srinivasan, Robin Strudel, Benigno Uria, Oliver Wang, Su Wang, Austin Waters, Chris Wolff, Auriel Wright, Zhisheng Xiao, Hao Xiong, Keyang Xu, Marc van Zee, Junlin Zhang, Katie Zhang, Wenlei Zhou, Konrad Zolna, Ola Aboubakar, Canfer Akbulut, Oscar Akerlund, Isabela Albuquerque, Nina Anderson, Marco Andretto, Lora Aroyo, Ben Bariach, David Barker, Sherry Ben, Dana Berman, Courtney Biles, Irina Blok, Pankil Botadra, Jenny Brennan, Karla Brown, John Buckley, Rudy Bunel, Elie Bursztein, Christina Butterfield, Ben Caine, Viral Carpenter, Norman Casagrande, Ming-Wei Chang, Solomon Chang, Shamik Chaudhuri, Tony Chen, John Choi, Dmitry Churbanau, Nathan Clement, Matan Cohen, Forrester Cole, Mikhail Dektiarev, Vincent Du, Praneet Dutta, Tom Eccles, Ndidi Elue, Ashley Feden, Shlomi Fruchter, Frankie Garcia, Roopal Garg, Weina Ge, Ahmed Ghazy, Bryant Gipson, Andrew Goodman, Dawid Górný, Sven Gowal, Khyatti Gupta, Yoni Halpern, Yena Han, Susan Hao, Jamie Hayes, Amir Hertz, Ed Hirst Tingbo Hou Heidi Howard + +Mohamed Ibrahim, Dirichi Ike-Njoku, Joana Iljazi, Vlad Ionescu, William Isaac, Reena Jana, Gemma Jennings, Donovon Jenson, Xuhui Jia, Kerry Jones, Xiaoen Ju, Ivana Kajic, Christos Kaplanis, Burcu Karagol Ayan, Jacob Kelly, Suraj Kothawade, Christina Kouridi, Ira Ktena, Jolanda Kumakaw, Dana Kurniawan, Dmitry Lagun, Lily Lavitas, Jason Lee, Tao Li, Marco Liang, Maggie Li-Calis, Yuchi Liu, Javier Lopez Alberca, Peggy Lu, Kristian Lum, Yukun Ma, Chase Malik, John Mellor, Inbar Mosseri, Tom Murray, Aida Nematzadeh, Paul Nicholas, João Gabriel Oliveira, Guillermo Ortiz-Jimenez, Michela Paganini, Tom Le Paine, Roni Paiss, Alicia Parrish, Anne Peckham, Vikas Peswani, Igor Petrovski, Tobias Pfaff, Alex Pirozhenko, Ryan Poplin, Utsav Prabhu, Yuan Qi, Matthew Rahtz, Cyrus Rashtchian, Charvi Rastogi, Amit Raul, Ali Razavi, Sylvestre-Alvise Rebuffi, Susanna Ricco, Felix Riedel, Dirk Robinson, Pankaj Rohatgi, Bill Rosgen, Sarah Rumbley, Moonkyung Ryu, Anthony Salgado, Sahil Singla, Florian Schroff, Candice Schumann, Tanmay Shah, Brendan Shillingford, Kaushik Shivakumar, Dennis Shtatnov, Zach Singer, Evgeny Sluzhaev, Valerii Sokolov, Thibault Sottiaux, Florian Stimberg, Brad Stone, David Stutz, Yu-Chuan Su, Eric Tabellion, Shuai Tang, David Tao, Kurt Thomas, Gregory Thornton, Andeep Toor, Cristian Udrescu, Aayush Upadhyay, Cristina Vasconcelos, Alex Vasiloff, Andrey Voynov, Amanda Walker, Luyu Wang, Miaosen Wang, Simon Wang, Stanley Wang, Qifei Wang, Yuxiao Wang, Agoston Weisz, Olivia Wiles, Chenxia Wu, Xingyu Federico Xu, Andrew Xue, Jianbo Yang, Luo Yu, Mete Yurtoglu, Ali Zand, Han Zhang, Jia-geng Zhang, Catherine Zhao, Adilet Zhaxybay, Miao Zhou, Shengqi Zhu, Zhenkai Zhu, Dawn Bloxwich, Mahyar Bordbar, Luis C. Cobo, Eli Collins, Shengyang Dai, Tulsee Doshi, Anca Dragan, Douglas Eck, Demis Hassabis, Sissie Hsiao, Tom Hume, Koray Kavukcuoglu, Helen King, Jack Krawczyk, Yeqing Li, Kathy Meier-Hellstern, Andras Orban, Yury Pinsky, Amar Subramanya Oriol Vinyals Ting Yu and Yori Zwols. Imagen 3, 2024. 1 +[21] Jun Young Koh, Sang Hyun Park, and Joy Song. Improving text generation on images with synthetic captions, 2024. 1 +[22] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 7 +[23] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2 +[24] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R. Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9400-9409, 2024. 2 +[25] Leyang Li, Shilin Lu, Yan Ren, and Adams Wai-Kin Kong. Set you straight: Auto-steering denoising trajectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782, 2025. 1 + +[26] Ouxiang Li, Yuan Wang, Xinting Hu, Houcheng Jiang, Tao Liang, Yanbin Hao, Guojun Ma, and Fuli Feng. Speed: Scalable, precise, and efficient concept erasure for diffusion models. arXiv preprint arXiv:2503.07392, 2025. 1 +[27] Zejian Li, Chenye Meng, Yize Li, Ling Yang, Shengyuan Zhang, Jiarui Ma, Jiayi Li, Guang Yang, Changyuan Yang, Zhiyuan Yang, et al. Laion-sg: An enhanced large-scale dataset for training complex image-text models with structural annotations. arXiv preprint arXiv:2412.08580, 2024. 2 +[28] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19401-19411, 2024. 4, 7 +[29] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 2 +[30] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 7, 6 +[31] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2 +[32] Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3 +[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2 +[34] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3 +[35] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 2 +[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2 + +[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1 +[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024. +[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 1, 2 +[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2 +[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2 +[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2 +[44] Zhihua Tian, Sirun Nan, Ming Xu, Shengfang Zhai, Wenjie Qu, Jian Liu, Kui Ren, Ruoxi Jia, and Jiaheng Zhang. Sparse autoencoder as a zero-shot classifier for concept erasing in text-to-image diffusion models. arXiv preprint arXiv:2503.09446, 2025.1 +[45] Cristina Nader Vasconcelos, Abdullah Rashwan, Austin Waters, Trevor Walker, Keyang Xu, Jimmy Yan, Rui Qian, Yeqing Li, SHIXIN LUO, Yasumasa Onoe, Zarana Parekh, Ivana Kajic, Mandy Guo, Wenlei Zhou, Sarah Rosston, Roopal Garg, Hongliang Fei, Jordi Pont-Tuset, Su Wang, Henna Nandwani, Andrew Bunner, Kevin Swersky, David J. Fleet, Oliver Wang, and Jason Michael Baldridge. Greedy growing enables high-resolution pixel-based diffusion models. Transactions on Machine Learning Research, 2024. 1 +[46] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for text-controlled vision models. arXiv preprint arXiv:2302.03693, 2, 2023. 1 +[47] WikiArt.org. WikiArt - Visual Art Encyclopedia. https: //www.wikiart.org/, 2024. Accessed: June 30, 2024. 2 +[48] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with gecko: On met + +rics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 4 +[49] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2 +[50] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. Advances in Neural Information Processing Systems, 36, 2024. 2 +[51] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 4 +[52] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755-1764, 2024. 1 +[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2 +[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 5, 7, 6 +[55] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2 +[56] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2 +[57] Xiaojin Zhu, Andrew B Goldberg, Mohamed Eldawy, Charles R Dyer, and Bradley Strock. A text-to-picture synthesis system for augmenting communication. In AAAI, pages 1590-1595, 2007. 2 \ No newline at end of file diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/images.zip b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0de73924142c07bae94dff3d70d27c68013e3571 --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d9cf1a44d881e9087ed3e73be69708ccb30a26a71a3e43f80a641253aaf80020 +size 819446 diff --git a/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/layout.json b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a8a8f349af5457d478359af451974a7a9966e2bb --- /dev/null +++ b/2025/Erasing More Than Intended_ How Concept Erasure Degrades the Generation of Non-Target Concepts/layout.json @@ -0,0 +1,7078 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "text", + "content": "Erasing More Than Intended? How Concept Erasure Degrades the Generation of Non-Target Concepts" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": "Ibtihel Amara" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Ahmed Imtiaz Humayun" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1,3\\dagger}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Ivana Kajic" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Zarana Parekh" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Natalie Harris" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Sarah Young" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Chirag Nagpal" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Najoung Kim" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Junfeng He" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Cristina Nader Vasconcelos" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Deepak Ramachandran" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Golnoosh Farnadi" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Katherine Heller" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Mohammad Havaei" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "text", + "content": ", Negar Rostamzadeh" + }, + { + "bbox": [ + 61, + 160, + 585, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 182, + 204, + 465, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 182, + 204, + 465, + 217 + ], + "spans": [ + { + "bbox": [ + 182, + 204, + 465, + 217 + ], + "type": "text", + "content": "1Google Research, 2McGill University, 3Rice University, 4Google Deepmind" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 152, + 245, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 271, + 296, + 535 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 535 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 535 + ], + "type": "text", + "content": "Concept erasure techniques have recently gained significant attention for their potential to remove unwanted concepts from text-to-image models. While these methods often demonstrate promising results in controlled settings, their robustness in real-world applications and suitability for deployment remain uncertain. In this work, we (1) identify a critical gap in evaluating sanitized models, particularly in assessing their performance across diverse concept dimensions, and (2) systematically analyze the failure modes of text-to-image models post-erasure. We focus on the unintended consequences of concept removal on non-target concepts across different levels of interconnected relationships including visually similar, binomial, and semantically related concepts. To address this, we introduce EraseBench, a comprehensive benchmark for evaluating post-erasure performance. EraseBench includes over 100 curated concepts, targeted evaluation prompts, and a robust set of metrics to assess both effectiveness and side effects of erasure. Our findings reveal a phenomenon of concept entanglement, where erasure leads to unintended suppression of non-target concepts, causing spillover degradation that manifests as distortions and a decline in generation quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 135, + 572 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "As text-to-image (T2I) generative models [6, 20, 21, 38, 40, 45] have become increasingly popular and widely adopted, the demand for techniques that allow for the targeted removal of specific concepts has grown in parallel. Concept erasure methods [10, 11, 13, 18, 25, 26, 30, 32, 34, 44, 54] promise to address important concerns such as ensuring the safe deployment of models by removing undesirable or harmful concepts [10, 15, 52]. Whether applied to mitigate bias [4, 34, 46], protect privacy [10, 15, 22, 52], or filter out unwanted content, these techniques aim to create \"sanitized\" models capable of generating content without" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 319, + 245, + 555, + 451 + ], + "blocks": [ + { + "bbox": [ + 319, + 245, + 555, + 451 + ], + "lines": [ + { + "bbox": [ + 319, + 245, + 555, + 451 + ], + "spans": [ + { + "bbox": [ + 319, + 245, + 555, + 451 + ], + "type": "image", + "image_path": "2ee506efc645bb2c28150078b2ca6f7042a72b114f53625c688d251993ab02a9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 457, + 555, + 513 + ], + "lines": [ + { + "bbox": [ + 313, + 457, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 457, + 555, + 513 + ], + "type": "text", + "content": "Figure 1. The effects of concept erasure on non-target concepts. Pre-erasure outputs (left) vs. post-erasure results (right) from Stable Diffusion (SD). Erasure negatively impacts the quality of unrelated concepts. EraseBench identifies such effects and offers a framework to evaluate the reliability of erasure methods." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 556, + 715 + ], + "type": "text", + "content": "triggering unwanted associations. However, the challenge of effectively erasing a concept without unintentionally distorting related or visually similar concepts remains unresolved. For example, erasing \"cat\" from a model should ideally remove all representations of a cat, yet leave semantically or visually related concepts like \"lion\" unaffected. Figure 1 shows various failure modes observed in state-of-the-art concept erasure techniques. These techniques exhibit instability after a concept is erased, often causing unintended consequences for closely related non-target concepts such as over-erasure and image artifacts (i.e. distortions). In this work, we argue that concept erasure techniques often struggle under real-world conditions. Even state-of-the-art methods fail when erasing concepts that are visually similar, binomially related, or semantically entangled. This lack of robustness raises concerns about their deployment in sen" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16420" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 95 + ], + "type": "text", + "content": "sitive applications, as unintended distortions of related non-target concepts affects the reliability of T2I models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 96, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 96, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 96, + 294, + 300 + ], + "type": "text", + "content": "While recent works assess erasure techniques using arbitrary concepts, a standardized framework for stress-testing across diverse concept relationships is still lacking. To address this, we introduce EraseBench, a unified benchmark with curated concepts and evaluations designed to rigorously test T2I models post-erasure. Covering binomial pairs, visually similar objects, and semantic relations. EraseBench enables consistent evaluation of erasure effectiveness, unintended distortions, and concept leakage. Our findings highlight persistent challenges and the need for more robust and standardized testing. We envision EraseBench as a foundation for improving comparability and advancing safer and more reliable erasure methods. Additionally, our results demonstrate that these sanitized models are not yet ready for deployment, highlighting the need for more robust and reliable erasure techniques that can operate effectively in complex, real-world scenarios." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 263, + 310 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 263, + 310 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 263, + 310 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 312, + 295, + 502 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 312, + 295, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 312, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 312, + 295, + 360 + ], + "type": "text", + "content": "1. We identify key evaluation dimensions where most concept erasure techniques exhibit vulnerabilities, particularly when handling visually similar, binomial, and semantically related concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 360, + 295, + 407 + ], + "type": "text", + "content": "2. We present EraseBench, a comprehensive, multidimensional framework designed to rigorously evaluate the robustness and efficacy of concept erasure techniques across a diverse set of concepts and prompts." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 295, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 442 + ], + "type": "text", + "content": "3. We use a suite of evaluation metrics to measure the effectiveness of concept erasure methods that provides a more holistic assessment of their performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 443, + 295, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 443, + 295, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 295, + 502 + ], + "type": "text", + "content": "4. We evaluate five state-of-the-art concept erasure techniques on EraseBench and demonstrate that many of the current concept erasure techniques are not yet ready for real-world deployment, highlighting significant gaps in their reliability and robustness." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 514, + 146, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 146, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 146, + 526 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 534, + 295, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 700 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 700 + ], + "type": "text", + "content": "Concept Erasure. T2I models have witnessed remarkable advancements since the foundational works of [31, 57], driven by breakthroughs in model architectures [5, 19, 42, 49, 53, 56], state-of-the-art generative modeling techniques [1, 3, 12, 16, 23, 24, 29, 29, 33, 37-40, 50], and the availability of large high-quality datasets [27, 41]. These improvements have significantly enhanced their capability to generate realistic and diverse images from textual descriptions, pushing the boundaries of creativity and application. In light of recent advancements in T2I models, greater attention has been directed towards reliability considerations. There is a wide variety of concept erasure techniques, each utilizing distinct methods and technologies to achieve concept removal [7, 9-11, 14, 18, 22, 30, 30, 54]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Challenges in Concept Erasure. Despite the rapid in" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 287 + ], + "type": "text", + "content": "crease of these concept erasure methods, only a few works have critically examined the inherent limitations and vulnerabilities of concept erasure approaches. In particular, the work in [35] explores a fundamental vulnerability, which is the potential for erased concepts to be recovered through specific prompting. Meanwhile, other works, such as in [55], assess the effectiveness of concept erasure by examining a diverse range of objects and styles. Their approach evaluates whether the erased concepts have been completely removed from the model's generative capabilities, ensuring that no traces remain in the generated outputs. In our work, we delve into the broader impact of concept erasure, not only evaluating residual concept generation but also examining how erasure techniques may degrade the quality and fidelity of non-target concepts. In this work, we highlight the ripple effects of concept erasure on model performance and the stability of related concepts via the route of interrelated concepts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 298, + 391, + 311 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 391, + 311 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 391, + 311 + ], + "type": "text", + "content": "3. EraseBench" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 553, + 341 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 553, + 341 + ], + "type": "text", + "content": "In this section, we provide a detailed breakdown of the process used to construct our benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 461 + ], + "type": "text", + "content": "Concept Gathering. We utilized the semantic knowledge embedded in large language models (LLMs) and structured sources, like the hierarchical taxonomy of ImageNet, to generate clusters of closely related concepts. By treating these language models akin to graph neural networks, we mapped similar concepts that share key attributes and gathered distinct sets for each dimension. This approach captures nuanced relationships crucial for testing erasure precision. For artistic styles, we selected artist names from the WikiArt dataset [47]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 463, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 553, + 582 + ], + "type": "text", + "content": "Concept Verification. To ensure quality and feasibility, we conducted human verification on the collected concepts. This included assessing whether existing T2I models could successfully generate accurate, high-quality representations of each concept. During this process, we identified and removed any concepts that models consistently struggled to render, preserving only those concepts that aligned with our benchmarks. This step ensures that the dataset maintains a high standard, with each concept suitable for rigorous testing of erasure techniques." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 665 + ], + "type": "text", + "content": "Prompt Construction. We used Gemini Flash [43], to generate a diverse set of prompts for each concept. Prompts were crafted with variations in style, complexity, and length to allow for thorough evaluation. This includes prompts that differ in level of detail (simple vs. elaborate descriptions) and length (short vs. extended prompts). With this range, we enable a finer assessment of concept erasure techniques." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 713 + ], + "type": "text", + "content": "Concept Dimensions. We introduce multiple evaluation dimensions, each capturing a distinct type of concept relationship or similarity, providing a structured way to evaluate the impact of erasure. Within each dimension, we pro" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16421" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 547, + 466 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 547, + 466 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 547, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 547, + 466 + ], + "type": "image", + "image_path": "c12e62b9cb30586992e201acb12108a1b6e6f8415d0c61d0c201a34c113a0aa3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "lines": [ + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 555, + 515 + ], + "type": "text", + "content": "Figure 2. Ripple effects of concept erasure methods across EraseBench entanglement dimensions. All the erasure baselines display failure cases across different EraseBench tasks. Receler and MACE frequently produce images that are unrelated to the text prompt, indicating they are the most sensitive of the five concept erasure techniques. In contrast, AdvUnlearn shows slightly better robustness across certain dimensions of the benchmark." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 73, + 529, + 279, + 619 + ], + "blocks": [ + { + "bbox": [ + 73, + 529, + 279, + 619 + ], + "lines": [ + { + "bbox": [ + 73, + 529, + 279, + 619 + ], + "spans": [ + { + "bbox": [ + 73, + 529, + 279, + 619 + ], + "type": "image", + "image_path": "4ee4c688bd9440dfee2a1a80929d28630c22ae348d9279e1c8f954bcc0368ef6.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 633, + 265, + 644 + ], + "lines": [ + { + "bbox": [ + 85, + 633, + 265, + 644 + ], + "spans": [ + { + "bbox": [ + 85, + 633, + 265, + 644 + ], + "type": "text", + "content": "Figure 3. Evaluation Dimensions of EraseBench." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 713 + ], + "type": "text", + "content": "vide a diverse set of concepts that challenge the model in different ways, ensuring a thorough evaluation. Figure 3 summarizes these dimensions in EraseBench. For each, we define a primary concept for erasure and related non-target concepts sharing visual, semantic, or contextual similari-" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 525, + 555, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 555, + 656 + ], + "type": "text", + "content": "ties to test unintended effects on the sanitized model. (1) Visual Similarity (Objects), A broad concept (e.g., \"cat\") is erased, along with synonyms (\"kitten,\" \"tabby\"), while visually similar non-targets (\"tiger,\" \"cheetah\") test unintended erasure. (2) Style Similarity (Artists), Erasing an artist (e.g., Van Gogh) tests model behavior on similar-style artists (e.g., Cézanne, Bernard). (3) Subset-Superset Relationship, Removing a concept (e.g., \"goldfish\") evaluates retention of related non-targets (\"guppy,\" \"koi\"). (4) Binomial Relations, Erasing one of a closely linked pair (e.g., \"sun\" from \"sun and moon\") tests impact on the other." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 657, + 556, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 657, + 556, + 705 + ], + "spans": [ + { + "bbox": [ + 313, + 657, + 556, + 705 + ], + "type": "text", + "content": "Explicit Content. Under each dimension of EraseBench, we also integrate explicit concepts to evaluate the handling of sensitive and safety-critical content (details can be found in Table 8 in the supplemental material). These include" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "16422" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "categories such as NSFW, culturally related, and broader safety-related themes. Incorporating these explicit concepts allows for a more rigorous assessment of how well erasure methods balance effectiveness, generation quality, and sensitivity." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 171, + 158 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 171, + 158 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 171, + 158 + ], + "type": "text", + "content": "4. Experimental setup" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 163, + 296, + 317 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 163, + 296, + 317 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 296, + 317 + ], + "type": "text", + "content": "Baseline models. In our work, we narrowed down existing techniques and focused on leveraging more recent, advanced and diverse approaches to concept erasure, experimenting with recent techniques that incorporate variations such as fine-tuning model weights (e.g. ESD [10]), introducing targeted weight perturbations (e.g. UCE [11]), refining textual embeddings and adversarial training (e.g. Receler [18], AdvUnlearn [54]), and introducing parameter efficient fine-tuning (Receler, MACE [30]). This allows us to explore the nuanced dynamics of concept erasure and assess its impact under different dimensions. We provide detailed description about each of these techniques in the supplemental material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 510 + ], + "type": "text", + "content": "Experiments. We use Stable Diffusion (SD) as our T2I model and apply concept erasure techniques to each primary concept in EraseBench. For each concept, we perform erasure using both the default and best-reported settings from baseline methods. We maintain the default retain sets (e.g. COCO prompts for AdvUnlearn) and, for methods requiring anchor concepts, we map them to an empty concept (\"). We evaluate the sanitized models on related non-target concepts within the same EraseBench dimension, assessing the effects on visually and semantically similar concepts. Each concept is represented by over 10 unique prompts, ranging from simple to complex descriptions. To capture diverse outputs, we generate 10 images per prompt with different random seeds. This approach provides a comprehensive evaluation of erasure effectiveness and its unintended effects on related concepts." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 511, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 511, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 511, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We use different types of automated evaluation metrics to quantify and explain the detrimental effect erasure techniques can have on the quality of generated images of non-target concepts. We use three different automated evaluation metrics to quantify the post erasure performance: CLIP [36], Rich Automatic Human Feedback (RAHF [28]), and Gecko [48]. CLIP is a widely used joint image and text embedding model which quantifies the similarity between text and image representations. We leverage CLIP as a zero-shot classifier to determine whether a model generates the desired concept. The RAHF model is trained on rich human feedback for generated images with a range of scores to evaluate overall image quality, and also mark points about which regions are problematic. The RAHF quality scores include an aesthetic score, which assesses the visual appeal of an image, and an artifact score (i.e. plausibility score), which gauges the presence and extent" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": "of visual artifacts. We use both the aesthetic and artifact scores as measurements of quality for generated images, following [28], where higher scores indicate better quality (i.e. good aesthetics and minimal artifacts). In addition to scores, the RAHF model provides heatmaps for each image to highlight artifact-prone areas, which we use for visual inspection in our benchmark. Finally, we use a Visual Question Answering (VQA)-based metric called Gecko, a VQA evaluation framework. In our setup, we leverage the multimodal Gemini 1.5 model for both steps: first, to generate text-related questions, and second, to answer them based on the given image, similar to other VQA methods [2, 51]. One advantage of such metrics is that they do not require task-specific fine-tuning and provide an interpretable alignment score. This evaluation can be replicated using other available models. Typically, an LLM generates questions related to the text prompt in the first step, while a vision-language model (VLM) answers those questions based on the generated image. The final score is computed as the average of correctly answered questions, allowing for backtracing to identify which aspects of the text are misaligned with the generated image." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 346, + 368, + 358 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 346, + 368, + 358 + ], + "spans": [ + { + "bbox": [ + 314, + 346, + 368, + 358 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 365, + 555, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 390 + ], + "type": "text", + "content": "5.1. Concept Erasure Reduces T2I Alignment For Non-Target Concepts (Over-Erasure)" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 331, + 402, + 430, + 502 + ], + "blocks": [ + { + "bbox": [ + 331, + 402, + 430, + 502 + ], + "lines": [ + { + "bbox": [ + 331, + 402, + 430, + 502 + ], + "spans": [ + { + "bbox": [ + 331, + 402, + 430, + 502 + ], + "type": "image", + "image_path": "340ad0189edabfbf75c37c294d78220a7ae1c5405fbf5ae0205ba7bf16f8bd55.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 512, + 443, + 521 + ], + "lines": [ + { + "bbox": [ + 331, + 512, + 443, + 521 + ], + "spans": [ + { + "bbox": [ + 331, + 512, + 443, + 521 + ], + "type": "text", + "content": "Q1: Is there a tiger in the image?" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 437, + 402, + 536, + 501 + ], + "blocks": [ + { + "bbox": [ + 437, + 402, + 536, + 501 + ], + "lines": [ + { + "bbox": [ + 437, + 402, + 536, + 501 + ], + "spans": [ + { + "bbox": [ + 437, + 402, + 536, + 501 + ], + "type": "image", + "image_path": "86cfd8f20e7ca061ec337942e991d09e0805407ace5698d9e80565a9781b377b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 331, + 522, + 463, + 531 + ], + "lines": [ + { + "bbox": [ + 331, + 522, + 463, + 531 + ], + "spans": [ + { + "bbox": [ + 331, + 522, + 463, + 531 + ], + "type": "text", + "content": "Q2: Is the tiger perched on something?" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 517, + 502, + 528, + 571 + ], + "lines": [ + { + "bbox": [ + 517, + 502, + 528, + 571 + ], + "spans": [ + { + "bbox": [ + 517, + 502, + 528, + 571 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 430, + 573, + 489, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 430, + 573, + 489, + 582 + ], + "spans": [ + { + "bbox": [ + 430, + 573, + 489, + 582 + ], + "type": "text", + "content": "Alignment score" + } + ] + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 493, + 501, + 509, + 571 + ], + "blocks": [ + { + "bbox": [ + 331, + 531, + 479, + 540 + ], + "lines": [ + { + "bbox": [ + 331, + 531, + 479, + 540 + ], + "spans": [ + { + "bbox": [ + 331, + 531, + 479, + 540 + ], + "type": "text", + "content": "Q3: Is the tiger perched on a rocky outcrop?" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 331, + 541, + 461, + 550 + ], + "lines": [ + { + "bbox": [ + 331, + 541, + 461, + 550 + ], + "spans": [ + { + "bbox": [ + 331, + 541, + 461, + 550 + ], + "type": "text", + "content": "Q4: Are there mountains in the image?" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 331, + 551, + 449, + 559 + ], + "lines": [ + { + "bbox": [ + 331, + 551, + 449, + 559 + ], + "spans": [ + { + "bbox": [ + 331, + 551, + 449, + 559 + ], + "type": "text", + "content": "Q5: Is the sky a serene blue color?" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 331, + 560, + 462, + 570 + ], + "lines": [ + { + "bbox": [ + 331, + 560, + 462, + 570 + ], + "spans": [ + { + "bbox": [ + 331, + 560, + 462, + 570 + ], + "type": "text", + "content": "Q6: Is there a sky visible in the image?" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 493, + 501, + 509, + 571 + ], + "lines": [ + { + "bbox": [ + 493, + 501, + 509, + 571 + ], + "spans": [ + { + "bbox": [ + 493, + 501, + 509, + 571 + ], + "type": "image", + "image_path": "4b6e1b5ff9c3ec449dcf77154277f2dbb165f9867eb40fb144bf205dbe4a11d6.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 517, + 573, + 528, + 582 + ], + "lines": [ + { + "bbox": [ + 517, + 573, + 528, + 582 + ], + "spans": [ + { + "bbox": [ + 517, + 573, + 528, + 582 + ], + "type": "text", + "content": "50" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 591, + 555, + 635 + ], + "lines": [ + { + "bbox": [ + 313, + 591, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 555, + 635 + ], + "type": "text", + "content": "Figure 4. Erasure affects fine-grained alignment. Prompt: \"A tiger perched on a rocky outcrop surrounded by mountains and a serene blue sky.\" before (left) and after (right) erasing the concept cat using MACE [30])." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 715 + ], + "type": "text", + "content": "Table 1 presents the zero-shot predictions from CLIP as percentages (\\%). Following the definitions in [30], we measure two key attributes: Efficacy (Eff.) and Generality (Gen.). Efficacy represents the accuracy of concept erasure (accuracy on the erased class), where lower values indicate better erasure. Generality reflects accuracy on paraphrased or syn" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16423" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 145, + 553, + 214 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 138 + ], + "type": "text", + "content": "Table 1. Evaluation of Erasing concepts from EraseBench under the four different dimensions. We provide the average results across 10 concepts (for visual similarity (object)), 15 artists concepts (for artists similarity), 8 concepts for subset-superset, and 10 concepts (for binomial). We evaluate efficacy (Eff., capturing the efficacy of erasing the main concept), generality (Gen., capturing the impact on paraphrased or related concepts), and sensitivity (Sens., capturing the unintended effect on non-target yet related or similar concepts). All values are expressed as percentages (\\%). More detailed, per-concept breakdowns are provided in the Supplemental Material. For reference, we also include the classification accuracy of images generated by the original Stable Diffusion v1.4 model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 145, + 553, + 214 + ], + "lines": [ + { + "bbox": [ + 56, + 145, + 553, + 214 + ], + "spans": [ + { + "bbox": [ + 56, + 145, + 553, + 214 + ], + "type": "table", + "html": "
TechniquesVisual Similarity (Object)Artistic Style Similarity (Artists)Subset-SupersetBinomial
Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑Eff. ↓Gen. ↓Sens. ↑HM ↑
Original86.5 ± 7.190.2 ± 3.685.0 ± 4.015.9772.3 ± 6.380.7 ± 3.269.0 ± 4.329.2983.3 ± 8.992.8 ± 3.992.1 ± 3.514.3188.3 ± 5.289.8 ± 4.388.5 ± 4.515.40
ESD [10]24.5 ± 6.150.5 ± 4.165.9 ± 4.661.7015.1 ± 3.761.9 ± 4.840.3 ± 3.947.7418.1 ± 4.251.2 ± 4.265.1 ± 5.162.4220.8 ± 4.641.3 ± 4.370.6 ± 4.068.46
UCE [11]41.8 ± 5.568.3 ± 2.782.7 ± 3.149.3221.1 ± 4.352.2 ± 3.561.0 ± 4.360.0151.1 ± 8.862.7 ± 2.987.5 ± 2.951.1218.9 ± 2.231.4 ± 2.886.1 ± 4.3277.88
Receler [18]8.1 ± 3.220 ± 3.465.4 ± 3.877.588.7 ± 2.245.3 ± 3.822.8 ± 3.841.044.2 ± 0.820.4 ± 3.736.7 ± 7.559.7010.3 ± 2.412.6 ± 3.157.5 ± 9.275.04
MACE [30]15.6 ± 6.437.7 ± 3.966.4 ± 4.369.8320.2 ± 4.636.7 ± 4.749.2 ± 4.861.6613.9 ± 3.228.2 ± 3.866.9 ± 5.674.1011.3 ± 2.528.9 ± 5.170.7 ± 5.375.98
AdvUnlearn [54]8.7 ± 2.939.1 ± 6.564.3 ± 5.769.8814.5 ± 4.137.4 ± 3.927.4 ± 4.047.757.4 ± 2.630.1 ± 5.260.1 ± 5.971.879.3 ± 2.227.6 ± 2.164.6 ± 5.974.41
", + "image_path": "8b8b11f5a3b8372826621a519a7aaa8f453f98013b03b79fa3ddbe02be739942.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 226, + 297, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 297, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 297, + 706 + ], + "type": "text", + "content": "onymous non-erased concepts, showing how well general information is erased. Additionally, we introduce Sensitivity (Sens.), defined as the accuracy for non-target but similar concepts. Most of the concept erasure techniques decreased accuracy on the target concept class (a desired outcome). Receler, MACE, and AdvUnlearn achieved the best efficacy in erasure overall per EraseBench dimension. Only a few techniques showed strong generality, indicating that erasure was not applied in a broad, global manner. In particular, Receler consistently displayed the best generality, preserving accuracy across target erased classes and paraphrased non-erased concepts. This may be attributed to the fact that Receler was trained using adversarial prompting and employs \"erasers\" designed to capture textual semantic concepts effectively. Overall, we observed a decrease in sensitivity, an undesirable outcome that highlights T2I misalignment for non-target non-erased concepts, essentially, over-erasure. After erasing a target concept, we noticed a clear mismatch between the text prompt and the generated image for related non-target concepts. Low sensitivity suggests that instead of selectively erasing only the target concept, the model inadvertently suppresses related concepts, causing generated images to fall out of distribution. This indicates that the model has not only erased the intended concept but also learned to overlook semantically or visually similar ones. As a result, the generated images fail to align with the intended prompts, revealing the model's inability to fully preserve non-target concepts after erasure. In Figure 2 (row 1), we observe that when models undergo erasure of the koala concept, prompting them with related concepts like Tree Kangaroo reveals unintended effect. These erased models appear to have also forgotten how to generate a Tree Kangaroo accurately. Similarly, when we erase the artist concept Bosch and prompt the sanitized model to generate artwork in the style of Altdorfer, the model struggles to capture the distinctive artistic style of the non-erased concept. This shows that the erasure process affects not only the target concept but also related artistic styles, leading to a loss of fidelity in the generated images. This demonstrates a significant vulnerability in these erasure tech" + } + ] + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 317, + 332, + 553, + 397 + ], + "blocks": [ + { + "bbox": [ + 313, + 224, + 555, + 323 + ], + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 323 + ], + "type": "text", + "content": "Table 2. Harmonic mean " + }, + { + "bbox": [ + 313, + 224, + 555, + 323 + ], + "type": "inline_equation", + "content": "(\\uparrow)" + }, + { + "bbox": [ + 313, + 224, + 555, + 323 + ], + "type": "text", + "content": " of efficacy, generality, and sensitivity scores on explicit concepts across EraseBench dimensions. Each column represents a distinct relationship type: visually similar objects, artist style similarity, subset-superset relationships, and binomial pairs. Higher scores indicate better balance between erasure success and minimal unintended degradation. While all methods outperform the original model, performance varies by dimension, revealing trade-offs in handling sensitive and safety-critical content." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "lines": [ + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "spans": [ + { + "bbox": [ + 317, + 332, + 553, + 397 + ], + "type": "table", + "html": "
TechniqueVisual Sim.Subset-supersetBinomial
Original16.3723.4819.64
UCE [11]63.1662.6276.13
AdvUnlearn [54]67.6658.2871.33
Receler [18]68.4756.3174.88
", + "image_path": "0958fbcbf75ba8904226883cad7818947a26c2f4df29f24008fd816f385a24ca.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 480, + 553, + 561 + ], + "blocks": [ + { + "bbox": [ + 313, + 406, + 556, + 472 + ], + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 472 + ], + "type": "text", + "content": "Table 3. Text-to-image alignment scores using the Gecko metric. The average scores and the standard errors of the mean for images generated with different techniques and prompts that either contain the erased concept or not. \"Original\" is the baseline model w/o any concept erasure. Drop in score compared to the original model shown in brackets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 480, + 553, + 561 + ], + "lines": [ + { + "bbox": [ + 317, + 480, + 553, + 561 + ], + "spans": [ + { + "bbox": [ + 317, + 480, + 553, + 561 + ], + "type": "table", + "html": "
TechniqueErased ConceptsNon-erased Concepts
Original84.1 ± 0.977.6 ± 0.7
UCE [11]57.6 ± 1.7 (−26.4)74.3 ± 0.8 (−3.4)
MACE [30]38.2 ± 1.3 (−45.9)67.9 ± 0.9 (−9.8)
AdvUnlearn [54]43.1 ± 1.4 (−41.0)68.6 ± 0.9 (−9.0)
", + "image_path": "7a8de9ee9fe99f90ebf742637d0e95067f22850d2dc800ec6e4bdff3ca82925b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "niques and highlights their risks associated with deploying them in real-world unrestricted settings. This trend of low sensitivity is consistent across dimensions of EraseBench. If we want to evaluate overall, UCE demonstrates superior preservation of visual concept components compared to other methods. Table 2 presents the harmonic mean of efficacy, generality, and sensitivity scores for explicit concepts (see Table 8 in the supplemental material for details) across three EraseBench dimensions: Visual Similarity, Subset-superset, and Binomial. UCE demonstrates consistently strong performance across most dimensions, while AdvUnlearn and Receler achieve particularly high scores in" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "16424" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 81, + 135, + 159 + ], + "blocks": [ + { + "bbox": [ + 68, + 70, + 125, + 80 + ], + "lines": [ + { + "bbox": [ + 68, + 70, + 125, + 80 + ], + "spans": [ + { + "bbox": [ + 68, + 70, + 125, + 80 + ], + "type": "text", + "content": "Original Image" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 81, + 135, + 159 + ], + "lines": [ + { + "bbox": [ + 59, + 81, + 135, + 159 + ], + "spans": [ + { + "bbox": [ + 59, + 81, + 135, + 159 + ], + "type": "image", + "image_path": "03f02e19b6359ad849ba68ca742b828317687d1fed841461a13a1ec1de780b1b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 168, + 104, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 168, + 104, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 168, + 104, + 178 + ], + "type": "text", + "content": "Er: Goldfish" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 136, + 85, + 212, + 162 + ], + "blocks": [ + { + "bbox": [ + 165, + 76, + 184, + 83 + ], + "lines": [ + { + "bbox": [ + 165, + 76, + 184, + 83 + ], + "spans": [ + { + "bbox": [ + 165, + 76, + 184, + 83 + ], + "type": "text", + "content": "UCE" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 136, + 85, + 212, + 162 + ], + "lines": [ + { + "bbox": [ + 136, + 85, + 212, + 162 + ], + "spans": [ + { + "bbox": [ + 136, + 85, + 212, + 162 + ], + "type": "image", + "image_path": "d4f5915c06971e2de571135cc7e5a4eb175925ddb600a9c6d3a7dd1898c17887.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 180, + 105, + 191 + ], + "lines": [ + { + "bbox": [ + 56, + 180, + 105, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 180, + 105, + 191 + ], + "type": "text", + "content": "Pr: A Guppy" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_footnote" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 214, + 85, + 291, + 162 + ], + "blocks": [ + { + "bbox": [ + 230, + 76, + 276, + 84 + ], + "lines": [ + { + "bbox": [ + 230, + 76, + 276, + 84 + ], + "spans": [ + { + "bbox": [ + 230, + 76, + 276, + 84 + ], + "type": "text", + "content": "AdvUnlearn" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 214, + 85, + 291, + 162 + ], + "lines": [ + { + "bbox": [ + 214, + 85, + 291, + 162 + ], + "spans": [ + { + "bbox": [ + 214, + 85, + 291, + 162 + ], + "type": "image", + "image_path": "89734462cff06f4d4824c3ad95b18a477bdf021b50dddd787c4cd0009bb18ddf.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 137, + 164, + 212, + 240 + ], + "blocks": [ + { + "bbox": [ + 56, + 193, + 118, + 202 + ], + "lines": [ + { + "bbox": [ + 56, + 193, + 118, + 202 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 118, + 202 + ], + "type": "text", + "content": "Artifact Scores:" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 56, + 205, + 111, + 215 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 111, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 111, + 215 + ], + "type": "text", + "content": "Original:93.63" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 56, + 217, + 100, + 226 + ], + "lines": [ + { + "bbox": [ + 56, + 217, + 100, + 226 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 100, + 226 + ], + "type": "text", + "content": "UCE: 84.21" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 56, + 228, + 124, + 238 + ], + "lines": [ + { + "bbox": [ + 56, + 228, + 124, + 238 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 124, + 238 + ], + "type": "text", + "content": "AdvUnlearn:79.01" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_footnote" + }, + { + "bbox": [ + 137, + 164, + 212, + 240 + ], + "lines": [ + { + "bbox": [ + 137, + 164, + 212, + 240 + ], + "spans": [ + { + "bbox": [ + 137, + 164, + 212, + 240 + ], + "type": "image", + "image_path": "56bae0e5407c2dd75304e7c523b061ef449a6ca964c8a1e4498b5254b683dac0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 256, + 295, + 345 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 295, + 345 + ], + "type": "text", + "content": "Figure 5. Erasure introduces artifacts during subset-superset concept generation. We erase concept \"goldfish\" and generate images for the prompt \"an image of a guppy\". We present the RAHF artifact heatmaps for images generated post-erasure via AdvUnlearn and UCE. We see that the artifact introduced by each method can vary spatially and by intensity, which prompts our inclusion of the artifact score in EraseBench. More heatmaps examples can be found in the supplemental material." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 214, + 164, + 291, + 240 + ], + "blocks": [ + { + "bbox": [ + 214, + 164, + 291, + 240 + ], + "lines": [ + { + "bbox": [ + 214, + 164, + 291, + 240 + ], + "spans": [ + { + "bbox": [ + 214, + 164, + 291, + 240 + ], + "type": "image", + "image_path": "e9b0cd298957c065681039aac20e5d3cda7fb3f9624f132b61b9cebaccb6399b.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 364, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 295, + 413 + ], + "type": "text", + "content": "Subset-superset and visual similarity settings. These results highlight the value of evaluating concept erasure across diverse relational categories to ensure method robustness and general applicability." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "text", + "content": "Next, we use Gecko to evaluate image-text alignment with a VQA-based approach. Unlike CLIP, Gecko provides an interpretable alignment score as it is computed over question-answer pairs. Figure 4 shows an example of question-answer pairs for a pair of an original and a sanitized image. While some aspects of the prompt are still correctly depicted in the image, the score is lower for sanitized model as it is missing the animal. We compute alignment scores between images generated by original and sanitized models (UCE, MACE, and AdvUnlearn) and prompts corresponding to four different concepts (goat, goldfish, cat, and Vincent Van Gogh) from four different dimensions. In total, we compute alignment scores for 6246 text-image pairs. The results in Table 3 show an expected drop in scores for images generated with sanitized models compared to images generated with the original model (c.f. column \"Erased concepts\"), as we expect the erased concept to be missing in images generated by sanitized models. However, we also observe a drop in scores for non-erased concepts across different dimensions. While that drop is much smaller, it is statistically significant across all three comparisons (" + }, + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\alpha = .01" + }, + { + "bbox": [ + 55, + 415, + 295, + 713 + ], + "type": "text", + "content": " using the Wilcoxon signed-rank test). The UCE technique has the smallest gap with the original model for non-erased concepts, while MACE has the largest. The results for UCE are consistent with CLIP-score zero-shot" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": "prediction in Table 1, where UCE was consistently the technique with highest sensitivity across all dimensions. We also manually inspected a subset of questions, answers and images, to ensure that the lower score is not caused by the VQA model latching on spurious correlations, but is indeed due to depicted concepts missing or the image being of lower quality. We observe that while some concepts are generally more difficult to depict correctly even in the original model (e.g. goldfish), this is reflected in the lower overall alignment score for the original model (e.g., for all three techniques the alignment score of the original model is in 60s for the concept goldfish, while it is in 80s and 90s for other concepts). Finally, question-answer breakdown in Gecko allows us to pinpoint some failures of image generation in sanitized models. When looking at answers to questions asking about object presence (eg., Is there an " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\langle" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": " object " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "\\rangle" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": " in the image?), we found that for all three techniques erased models had lower Gecko scores compared to original models on prompts with visually similar objects (UCE: " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "80.3\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "89.0\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": ", MACE: " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "67.1\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "89.0\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": ", and AdvUnlearn: " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "83.1\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": " vs " + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "inline_equation", + "content": "89.0\\%" + }, + { + "bbox": [ + 313, + 72, + 555, + 336 + ], + "type": "text", + "content": "), which highlights unintentional consequences of concept erasure techniques." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 347, + 555, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 373 + ], + "type": "text", + "content": "5.2. Concept Erasure Reduces Quality For Non-Target Concepts" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 715 + ], + "type": "text", + "content": "Global Concepts. Beyond non-target misalignment, T2I models also exhibit other forms of degradation post-erasure, such as distortions (i.e. artifacts) and poor aesthetics. To explore this further, we specifically employed metrics that better align with human feedback, such as artifact measurements that quantify the degree of perceptible distortions and aesthetic quality (i.e. factors that reflect how humans perceive image fidelity). In Table 4, we present the RAHF alignment scores of individual concepts under different dimensions of EraseBench. Our results reveal a significant decline in overall quality compared to the original SD, suggesting that the erasure process leads to noticeable degradation in image quality. This indicates that similar and related concepts remain vulnerable to generation flaws, despite attempts at concept retention. In Figure 2, we observe that most models capable of retaining and generating the intended concept after erasure still exhibit various types of distortions. These include misaligned body parts for animal classes, cropped concepts, nonsensical text distortions, decreased size of the generated concept, and a general lack of sharpness. Additional examples can be found in the Supplemental Material. Artists Concepts. We observe a significant drop in T2I alignment, highlighting a major challenge in artistic style erasure as highlighted in Table 1. Similar styles or concepts across artists are at risk of being inadvertently erased if one artist's style is closely related to another. We observe this issue in the context of T2I alignment and found a marked degradation in quality, particularly in the" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16425" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 89, + 553, + 167 + ], + "blocks": [ + { + "bbox": [ + 86, + 71, + 523, + 82 + ], + "lines": [ + { + "bbox": [ + 86, + 71, + 523, + 82 + ], + "spans": [ + { + "bbox": [ + 86, + 71, + 523, + 82 + ], + "type": "text", + "content": "Table 4. RAHF alignment scores [28]. Alignment scores for each concept under the overall dimension object similarity." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 89, + 553, + 167 + ], + "lines": [ + { + "bbox": [ + 56, + 89, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 56, + 89, + 553, + 167 + ], + "type": "table", + "html": "
TechniquesVisual Similarity (Object)BinomialSubset- superset
Erase "Cat"Erase "Goat"Erase "Lock"Erase "Goldfish"
Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑Artifacts ↑Aesthetics ↑
Original87.71 ± 0.0780.93 ± 0.0283.23 ± 0.678.44 ± 0.285.23 ± 1.576.72 ± 0.884.15 ± 0.178.21 ± 0.3
UCE [11]72.44 ± 0.973.77 ± 0.275.00 ± 1.074.19 ± 0.378.88 ± 3.478.50 ± 1.074.55 ± 0.672.20 ± 0.4
MACE [30]73.50 ± 0.871.80 ± 0.0772.00 ± 1.173.40 ± 1.272.79 ±4.077.22 ± 0.772.64 ± 0.375.33 ± 0.2
AdvUnlearn [54]74.60 ± 0.972.30 ± 1.269.78 ± 0.670.10 ± 0.966.23 ± 5.376.79 ± 1.171.87 ± 0.572.49 ± 0.2
", + "image_path": "2074999bba1982d3e3cbe4f185ae63cb38a4854ac3705ed72fd53196ae8d722d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 177, + 294, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 177, + 294, + 260 + ], + "spans": [ + { + "bbox": [ + 55, + 177, + 294, + 260 + ], + "type": "text", + "content": "style of non-target concepts. In Figure 2, we show the erasure of " + }, + { + "bbox": [ + 55, + 177, + 294, + 260 + ], + "type": "inline_equation", + "content": "Degas" + }, + { + "bbox": [ + 55, + 177, + 294, + 260 + ], + "type": "text", + "content": " (the artist) and its impact on the ability to generate images of other artists, such as Cassatt. Notably, we see a significant shift in style even for the non-erased concept. We specifically chose not to use the RAHF alignment scores for artistic style, as the models were not trained to evaluate artistic style based on their documentation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 282, + 294, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 294, + 295 + ], + "type": "text", + "content": "5.3. Empirical Validation with Human Preferences" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 301, + 295, + 493 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 301, + 295, + 493 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 295, + 493 + ], + "type": "text", + "content": "We conduct a human preference study to additionally validate results obtained with automated evaluation techniques. We recruited 11 participants from our institution to judge images based on three different criteria: overall image quality, distortions and T2I alignment. All participants were instructed on how to complete the task, and have provided consent to participate in the study. Each participant was presented with 50 side-by-side image comparisons, corresponding to 50 pairs of images generated by the original model and by the model with a concept erased. The order of images was randomized for each pair during presentation. The participants were asked the following three questions: (i) Which image exhibits superior overall quality? (ii) Which of the following images displays LESS noticeable distortions? (iii) Which image most accurately reflects and is aligned with the text label?" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 713 + ], + "type": "text", + "content": "Participants selected one of the three possible answers: Image A, Image B, or Neutral. In total, we collected 1650 responses for UCE and 1485 for AdvUnlearn. We focus on images of non-erased concepts, as the primary hypothesis we wanted to test was whether non-target visually similar concepts were negatively affected with the erasure technique. The results in Figure 6 show that humans judge images generated with the original model as having better overall quality and having fewer distortions. The original model is also preferred for better alignment, although it is closer to the \"Neutral\" for UCE, meaning that either both the original image and sanitized image were of equally good or equally poor quality. We provide additional human preferences for AdvUnlearn in the supplemental material. Results in Figure 10 also show that most participants preferred the original images with respect to quality, alignment, and less artifacts. Overall, we find that human preferences corroborate findings we observe with automatic metrics." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 325, + 175, + 545, + 266 + ], + "blocks": [ + { + "bbox": [ + 325, + 175, + 545, + 266 + ], + "lines": [ + { + "bbox": [ + 325, + 175, + 545, + 266 + ], + "spans": [ + { + "bbox": [ + 325, + 175, + 545, + 266 + ], + "type": "image", + "image_path": "d6a3f2f71ca49af6b4282fc356124d1bf0a9eedf590d73381e70f616808f6406.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 277, + 553, + 321 + ], + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 321 + ], + "type": "text", + "content": "Figure 6. Human image preferences between images generated by the original and the erased model. The erased model used is UCE. Results show that humans prefer SD over UCE. We provide the results of AdvUnlearn in the supplemental material." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 335, + 383, + 347 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 335, + 383, + 347 + ], + "spans": [ + { + "bbox": [ + 314, + 335, + 383, + 347 + ], + "type": "text", + "content": "6. Discussion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 554, + 461 + ], + "type": "text", + "content": "In our work, we have identified and emphasized critical issues surrounding T2I models post concept erasure, making it our primary focus. Addressing these challenges is the logical next step, but it raises a complex research question that demands further exploration. In this section, we discuss into possible mitigation strategies and assess their impacts. We stress that finding effective solutions is an ongoing, nontrivial process that requires continued investigation and innovation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 676 + ], + "type": "text", + "content": "The Role of Retain Sets in Mitigating Post-Erasure Distortions. A key question is whether the design of the retain set, specifically including concepts closely intertwined with the target erased concept can help alleviate post-generation distortions. We try to investigate whether adding non-target concepts that share semantic or visual similarities can reduce unintended side effects caused by erasure. To explore this, we performed erasure on targeted concepts using EraseBench while strategically incorporating related nontarget concepts into the retain set. For example, when erasing cat, we progressively introduce related concepts such as tiger and panther into the retain set and evaluate the impact on both generated outputs and artifact persistence. We show in Figure 7 the influence of the retain set on non-target concepts. We compare the original SD without erasure, erased model without a retain set (i.e. no explicit retain set), and erased model with a retain set in which we introduce the entangled non-target concepts one by one." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "(i) Findings on non-target concepts. Our investigation indicates that incorporating entangled non-target concepts into the retain set partially alleviates over-erasure (misalign" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16426" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 143 + ], + "type": "text", + "content": "ment on non-target concepts), preserves aspects of related non-target concepts. However, artifacts and distortions persist, particularly in non-targeted non-erased concepts suggesting that while carefully designing a retain set may help maintain concept structure, it does not fully resolve the unintended distortions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 300 + ], + "type": "text", + "content": "(ii) Findings on target-erased concepts. Examining the erased target concepts in Figure 8, we find that incorporating multiple entangled concepts into the retain set helps alleviate some post-erasure artifacts on the non-target concepts but comes at the cost of concept leakage. Specifically, this leakage manifests as spurious regeneration, where the erased concept partially or fully re-emerges in generated outputs. Our observations suggest that while including related non-target concepts in the retain set can mitigate over-erasure, it increases the risk of unintended regeneration. The more entangled concepts added, the greater the likelihood of leakage, highlighting a trade-off between preserving non-target concepts and ensuring effective erasure." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 295, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 525 + ], + "type": "text", + "content": "The Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure. We investigate whether anchor concepts can help in mitigating the unintended distortions in non-target concepts observed after concept erasure. An anchor may act as a reference point, defining boundaries for selective erasure while preserving related concepts. To test this, we conducted experiments in Figure 9 where we erased Van Gogh while introducing Post-Impressionism as an anchor concept. Similarly, we erased Grumpy Cat while anchoring the model with Internet meme, as outlined in [8]. The intuition behind this approach is that a well-defined anchor might help the model retain a coherent representation of related concepts, thereby reducing unintended consequences of erasure. We observed models struggle with generating good quality non-target concept images, suggesting anchors do not consistently improve stability. While anchors may help in alignment, they do not fully resolve artifacts or style preservation issues, warranting further investigation into their effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "text", + "content": "Multi-Concept Erasure: Intra-type and Inter-type Effects. We also investigate multi-concept erasure, differentiating between intra-type (main and paraphrased concepts) and inter-type (main and unrelated concepts) erasure. Average artifact scores in the visual similarity domain reach " + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "inline_equation", + "content": "78.3\\%" + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "text", + "content": " with intra-type and " + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "inline_equation", + "content": "71.4\\%" + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "text", + "content": " with inter-type erasure. The former outperforms the single concept erasure at " + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "inline_equation", + "content": "71.9\\%" + }, + { + "bbox": [ + 55, + 526, + 295, + 706 + ], + "type": "text", + "content": " (see example in Figure 15 in the supplemental material). Our findings show that intra-type erasure (i.e. removing multiple related concepts) more effectively eliminates targets and reduces artifacts when faced with non-target but related prompts. In contrast, inter-type erasure offers minimal artifact reduction, suggesting that carefully grouping related concepts enhances erasure precision without degrading generation quality." + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 57, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 317, + 57, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 317, + 57, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 317, + 57, + 553, + 178 + ], + "type": "image", + "image_path": "c97d047a46c55dcef46adcc1774602d209664816cc1050cb5a1a83a62d4a1349.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 186, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 186, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 186, + 553, + 251 + ], + "type": "text", + "content": "Figure 7. Impact of adding visually similar concepts to the retain set with respect to non-target concepts. Without a retain set, image quality degrades significantly. Unrelated concepts (e.g. house, elephant, and boat) fail to prevent artifacts, while a curated, visually similar retain set improves quality but remains below pre-erasure levels." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 317, + 258, + 553, + 334 + ], + "blocks": [ + { + "bbox": [ + 317, + 258, + 553, + 334 + ], + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 334 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 334 + ], + "type": "image", + "image_path": "8249c162b0b8457e088a7a39bf84e9b1c191cb94a54034c4dd533e13ef87b39c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 342, + 553, + 386 + ], + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 386 + ], + "type": "text", + "content": "Figure 8. Impact of adding visually similar concepts to the retain set with respect to the target concept. Without a retain set, erasure is effective, but as similar concepts are added, the erased concept gradually re-emerges." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 400, + 422, + 462 + ], + "blocks": [ + { + "bbox": [ + 315, + 400, + 422, + 462 + ], + "lines": [ + { + "bbox": [ + 315, + 400, + 422, + 462 + ], + "spans": [ + { + "bbox": [ + 315, + 400, + 422, + 462 + ], + "type": "image", + "image_path": "be64878cb49855bbc2d2c9ce620db014d0ddd0a35ebcc6c56e7f1d640ceac727.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 471, + 553, + 493 + ], + "lines": [ + { + "bbox": [ + 313, + 471, + 553, + 493 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 553, + 493 + ], + "type": "text", + "content": "Figure 9. Impact of Anchor Sets on Mitigating Unintended Effects of Concept Erasure." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 434, + 400, + 542, + 462 + ], + "blocks": [ + { + "bbox": [ + 434, + 400, + 542, + 462 + ], + "lines": [ + { + "bbox": [ + 434, + 400, + 542, + 462 + ], + "spans": [ + { + "bbox": [ + 434, + 400, + 542, + 462 + ], + "type": "image", + "image_path": "b005257329c1238324ca51a6380da723bb0b6a0d40c96d58384fbc1840f0b556.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 505, + 471, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 505, + 471, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 505, + 471, + 517 + ], + "type": "text", + "content": "7. Conclusion and Limitations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 526, + 555, + 703 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 526, + 555, + 703 + ], + "spans": [ + { + "bbox": [ + 313, + 526, + 555, + 703 + ], + "type": "text", + "content": "In this work, we present a thorough evaluation of concept erasure techniques and uncover shortcomings in their reliability. By introducing EraseBench, we reveal critical gaps in current methods. Our findings suggest that despite recent advancements, existing concept erasure techniques fall short in preserving model quality, which call for the need for more robust and nuanced approaches as well as more robust evaluation protocol and metrics. While our framework provides a diverse evaluation, automating concept selection could improve scalability. Given dimensions like binomiality, researchers can explore broader concept spaces beyond our curation. Additionally, a deeper exploration of optimal retain or erasure sets to minimize distortions remains an open question. We hope EraseBench inspires new research directions toward developing more reliable concept erasure techniques." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16427" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 146, + 296, + 191 + ], + "type": "text", + "content": "[2] Emanuele Bugliarello, Laurent Sartran, Aishwarya Agrawal, Lisa Anne Hendricks, and Aida Nematzadeh. Measuring progress in fine-grained vision-and-language understanding. arXiv preprint arXiv:2305.07558, 2023. 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 191, + 295, + 246 + ], + "type": "text", + "content": "[3] Huiwen Chang, Han Zhang, Jarred Barber, AJ Maschinot, Jose Lezama, Lu Jiang, Ming-Hsuan Yang, Kevin Murphy, William T Freeman, Michael Rubinstein, et al. Muse: Text-to-image generation via masked generative transformers. arXiv preprint arXiv:2301.00704, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[4] Ching-Yao Chuang, Varun Jampani, Yuanzhen Li, Antonio Torralba, and Stefanie Jegelka. Debiasing vision-language models via biased prompts. arXiv preprint arXiv:2302.00070, 2023. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 324 + ], + "type": "text", + "content": "[5] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 325, + 295, + 400 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 325, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 62, + 325, + 295, + 400 + ], + "type": "text", + "content": "[6] Patrick Esser, Sumith Kulal, A. Blattmann, Rahim Entezari, Jonas Muller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, Dustin Podell, Tim Dockhorn, Zion English, Kyle Lacey, Alex Goodwin, Yannik Marek, and Robin Rombach. Scaling rectified flow transformers for high-resolution image synthesis. ArXiv, abs/2403.03206, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 456 + ], + "type": "text", + "content": "[7] Chongyu Fan, Jiancheng Liu, Yihua Zhang, Eric Wong, Dennis Wei, and Sijia Liu. Salun: Empowering machine unlearning via gradient-based weight saliency in both image classification and generation. arXiv preprint arXiv:2310.12508, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 457, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 295, + 491 + ], + "type": "text", + "content": "[8] Masane Fuchi and Tomohiro Takagi. Erasing concepts from text-to-image diffusion models with few-shot unlearning. arXiv preprint arXiv:2405.07288, 2, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[9] Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H Bermano, Gal Chechik, and Daniel Cohen-Or. An image is worth one word: Personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618, 2022. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 295, + 590 + ], + "type": "text", + "content": "[10] Rohit Gandikota, Joanna Materzyńska, Jaden Fiotto-Kaufman, and David Bau. Erasing concepts from diffusion models. In Proceedings of the 2023 IEEE International Conference on Computer Vision, 2023. 1, 4, 5, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 635 + ], + "type": "text", + "content": "[11] Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, and David Bau. Unified concept editing in diffusion models. IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 1, 2, 4, 5, 7, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 680 + ], + "type": "text", + "content": "[12] Roy Ganz and Michael Elad. Clipag: Towards generator-free text-to-image generation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 3843-3853, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[13] Daiheng Gao, Shilin Lu, Wenbo Zhou, Jiaming Chu, Jie Zhang, Mengxi Jia, Bang Zhang, Zhaoxin Fan, and Weiming Zhang. Eraseanything: Enabling concept erasure in rectified" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "flow transformers. In Forty-second International Conference on Machine Learning, 2025. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 139 + ], + "type": "text", + "content": "[14] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36: 17170-17194, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 183 + ], + "type": "text", + "content": "[15] Alvin Heng and Harold Soh. Selective amnesia: A continual learning approach to forgetting in deep generative models. Advances in Neural Information Processing Systems, 36, 2024. 1, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 217 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840-6851, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 218, + 553, + 271 + ], + "type": "text", + "content": "[17] Edward J Hu, yelong shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. In International Conference on Learning Representations, 2022. 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 327 + ], + "type": "text", + "content": "[18] Chi-Pin Huang, Kai-Po Chang, Chung-Ting Tsai, Yung-Hsuan Lai, Fu-En Yang, and Yu-Chiang Frank Wang. Receler: Reliable concept erasing of text-to-image diffusion models via lightweight erasers. ECCV, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 328, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 373 + ], + "type": "text", + "content": "[19] Xun Huang, Arun Mallya, Ting-Chun Wang, and Ming-Yu Liu. Multimodal conditional image synthesis with product-of-experts gans. In European conference on computer vision, pages 91-109. Springer, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 373, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 373, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 373, + 553, + 713 + ], + "type": "text", + "content": "[20] Imagen-Team-Google, Jason Baldridge, Jakob Bauer, Mukul Bhutani, Nicole Brichtova, Andrew Bunner, Kelvin Chan, Yichang Chen, Sander Dieleman, Yuqing Du, Zach Eaton-Rosen, Hongliang Fei, Nando de Freitas, Yilin Gao, Evgeny Gladchenko, Sergio Gomez Colmenarejo, Mandy Guo, Alex Haig, Will Hawkins, Hexiang Hu, Huilian Huang, Tobenna Peter Igwe, Christos Kaplanis, Siavash Khodadadeh, Yelin Kim, Ksenia Konyushkova, Karol Langner, Eric Lau, Shixin Luo, Soña Mokrá, Henna Nandwani, Yasumasa Onoe, Aaron van den Oord, Zarana Parekh, Jordi Pont-Tuset, Hang Qi, Rui Qian, Deepak Ramachandran, Poorva Rane, Abdullah Rashwan, Ali Razavi, Robert Riachi, Hansa Srinivasan, Srivatsan Srinivasan, Robin Strudel, Benigno Uria, Oliver Wang, Su Wang, Austin Waters, Chris Wolff, Auriel Wright, Zhisheng Xiao, Hao Xiong, Keyang Xu, Marc van Zee, Junlin Zhang, Katie Zhang, Wenlei Zhou, Konrad Zolna, Ola Aboubakar, Canfer Akbulut, Oscar Akerlund, Isabela Albuquerque, Nina Anderson, Marco Andretto, Lora Aroyo, Ben Bariach, David Barker, Sherry Ben, Dana Berman, Courtney Biles, Irina Blok, Pankil Botadra, Jenny Brennan, Karla Brown, John Buckley, Rudy Bunel, Elie Bursztein, Christina Butterfield, Ben Caine, Viral Carpenter, Norman Casagrande, Ming-Wei Chang, Solomon Chang, Shamik Chaudhuri, Tony Chen, John Choi, Dmitry Churbanau, Nathan Clement, Matan Cohen, Forrester Cole, Mikhail Dektiarev, Vincent Du, Praneet Dutta, Tom Eccles, Ndidi Elue, Ashley Feden, Shlomi Fruchter, Frankie Garcia, Roopal Garg, Weina Ge, Ahmed Ghazy, Bryant Gipson, Andrew Goodman, Dawid Górný, Sven Gowal, Khyatti Gupta, Yoni Halpern, Yena Han, Susan Hao, Jamie Hayes, Amir Hertz, Ed Hirst Tingbo Hou Heidi Howard" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "16428" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 479 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 479 + ], + "type": "text", + "content": "Mohamed Ibrahim, Dirichi Ike-Njoku, Joana Iljazi, Vlad Ionescu, William Isaac, Reena Jana, Gemma Jennings, Donovon Jenson, Xuhui Jia, Kerry Jones, Xiaoen Ju, Ivana Kajic, Christos Kaplanis, Burcu Karagol Ayan, Jacob Kelly, Suraj Kothawade, Christina Kouridi, Ira Ktena, Jolanda Kumakaw, Dana Kurniawan, Dmitry Lagun, Lily Lavitas, Jason Lee, Tao Li, Marco Liang, Maggie Li-Calis, Yuchi Liu, Javier Lopez Alberca, Peggy Lu, Kristian Lum, Yukun Ma, Chase Malik, John Mellor, Inbar Mosseri, Tom Murray, Aida Nematzadeh, Paul Nicholas, João Gabriel Oliveira, Guillermo Ortiz-Jimenez, Michela Paganini, Tom Le Paine, Roni Paiss, Alicia Parrish, Anne Peckham, Vikas Peswani, Igor Petrovski, Tobias Pfaff, Alex Pirozhenko, Ryan Poplin, Utsav Prabhu, Yuan Qi, Matthew Rahtz, Cyrus Rashtchian, Charvi Rastogi, Amit Raul, Ali Razavi, Sylvestre-Alvise Rebuffi, Susanna Ricco, Felix Riedel, Dirk Robinson, Pankaj Rohatgi, Bill Rosgen, Sarah Rumbley, Moonkyung Ryu, Anthony Salgado, Sahil Singla, Florian Schroff, Candice Schumann, Tanmay Shah, Brendan Shillingford, Kaushik Shivakumar, Dennis Shtatnov, Zach Singer, Evgeny Sluzhaev, Valerii Sokolov, Thibault Sottiaux, Florian Stimberg, Brad Stone, David Stutz, Yu-Chuan Su, Eric Tabellion, Shuai Tang, David Tao, Kurt Thomas, Gregory Thornton, Andeep Toor, Cristian Udrescu, Aayush Upadhyay, Cristina Vasconcelos, Alex Vasiloff, Andrey Voynov, Amanda Walker, Luyu Wang, Miaosen Wang, Simon Wang, Stanley Wang, Qifei Wang, Yuxiao Wang, Agoston Weisz, Olivia Wiles, Chenxia Wu, Xingyu Federico Xu, Andrew Xue, Jianbo Yang, Luo Yu, Mete Yurtoglu, Ali Zand, Han Zhang, Jia-geng Zhang, Catherine Zhao, Adilet Zhaxybay, Miao Zhou, Shengqi Zhu, Zhenkai Zhu, Dawn Bloxwich, Mahyar Bordbar, Luis C. Cobo, Eli Collins, Shengyang Dai, Tulsee Doshi, Anca Dragan, Douglas Eck, Demis Hassabis, Sissie Hsiao, Tom Hume, Koray Kavukcuoglu, Helen King, Jack Krawczyk, Yeqing Li, Kathy Meier-Hellstern, Andras Orban, Yury Pinsky, Amar Subramanya Oriol Vinyals Ting Yu and Yori Zwols. Imagen 3, 2024. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 480, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 480, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 57, + 480, + 294, + 502 + ], + "type": "text", + "content": "[21] Jun Young Koh, Sang Hyun Park, and Joy Song. Improving text generation on images with synthetic captions, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 502, + 295, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 502, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 57, + 502, + 295, + 557 + ], + "type": "text", + "content": "[22] Nupur Kumari, Bingliang Zhang, Sheng-Yu Wang, Eli Shechtman, Richard Zhang, and Jun-Yan Zhu. Ablating concepts in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 22691-22702, 2023. 1, 2, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 558, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 558, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 57, + 558, + 294, + 602 + ], + "type": "text", + "content": "[23] Dongxu Li, Junnan Li, and Steven Hoi. Blip-diffusion: Pretrained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 58, + 603, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 603, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 58, + 603, + 295, + 668 + ], + "type": "text", + "content": "[24] Hao Li, Yang Zou, Ying Wang, Orchid Majumder, Yusheng Xie, R. Manmatha, Ashwin Swaminathan, Zhuowen Tu, Stefano Ermon, and Stefano Soatto. On the scalability of diffusion-based text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9400-9409, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 294, + 712 + ], + "type": "text", + "content": "[25] Leyang Li, Shilin Lu, Yan Ren, and Adams Wai-Kin Kong. Set you straight: Auto-steering denoising trajectories to sidestep unwanted concepts. arXiv preprint arXiv:2504.12782, 2025. 1" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[26] Ouxiang Li, Yuan Wang, Xinting Hu, Houcheng Jiang, Tao Liang, Yanbin Hao, Guojun Ma, and Fuli Feng. Speed: Scalable, precise, and efficient concept erasure for diffusion models. arXiv preprint arXiv:2503.07392, 2025. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 183 + ], + "type": "text", + "content": "[27] Zejian Li, Chenye Meng, Yize Li, Ling Yang, Shengyuan Zhang, Jiarui Ma, Jiayi Li, Guang Yang, Changyuan Yang, Zhiyuan Yang, et al. Laion-sg: An enhanced large-scale dataset for training complex image-text models with structural annotations. arXiv preprint arXiv:2412.08580, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 185, + 553, + 251 + ], + "type": "text", + "content": "[28] Youwei Liang, Junfeng He, Gang Li, Peizhao Li, Arseniy Klimovskiy, Nicholas Carolan, Jiao Sun, Jordi Pont-Tuset, Sarah Young, Feng Yang, et al. Rich human feedback for text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19401-19411, 2024. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 317, + 252, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 252, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 252, + 553, + 307 + ], + "type": "text", + "content": "[29] Zhiqiu Lin, Deepak Pathak, Baiqi Li, Jiayao Li, Xide Xia, Graham Neubig, Pengchuan Zhang, and Deva Ramanan. Evaluating text-to-visual generation with image-to-text generation. In European Conference on Computer Vision, pages 366–384. Springer, 2025. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 363 + ], + "type": "text", + "content": "[30] Shilin Lu, Zilan Wang, Leyang Li, Yanzhu Liu, and Adams Wai-Kin Kong. Mace: Mass concept erasure in diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6430-6440, 2024. 1, 2, 4, 5, 7, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 398 + ], + "type": "text", + "content": "[31] Elman Mansimov, Emilio Parisotto, Jimmy Lei Ba, and Ruslan Salakhutdinov. Generating images from captions with attention. arXiv preprint arXiv:1511.02793, 2015. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 443 + ], + "type": "text", + "content": "[32] Kevin Meng, Arnab Sen Sharma, Alex J Andonian, Yonatan Belinkov, and David Bau. Mass-editing memory in a transformer. In The Eleventh International Conference on Learning Representations, 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 498 + ], + "type": "text", + "content": "[33] Alex Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob McGrew, Ilya Sutskever, and Mark Chen. Glide: Towards photorealistic image generation and editing with text-guided diffusion models. arXiv preprint arXiv:2112.10741, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 544 + ], + "type": "text", + "content": "[34] Hadas Orgad, Bahjat Kawar, and Yonatan Belinkov. Editing implicit assumptions in text-to-image diffusion models. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 599 + ], + "type": "text", + "content": "[35] Minh Pham, Kelly O Marshall, Niv Cohen, Govind Mittal, and Chinmay Hegde. Circumventing concept erasure methods for text-to-image generative models. In The Twelfth International Conference on Learning Representations, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 601, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 601, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 317, + 601, + 553, + 666 + ], + "type": "text", + "content": "[36] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[37] Scott Reed, Zeynep Akata, Xinchen Yan, Lajanugen Logeswaran, Bernt Schiele, and Honglak Lee. Generative adversarial text to image synthesis. In International conference on machine learning, pages 1060-1069. PMLR, 2016. 2" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "16429" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 127 + ], + "type": "text", + "content": "[38] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10684-10695, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 195 + ], + "type": "text", + "content": "[39] Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Wei Wei, Tingbo Hou, Yael Pritch, Neal Wadhwa, Michael Rubinstein, and Kfir Aberman. Hyperdream booth: Hypernetworks for fast personalization of text-to-image models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6527-6536, 2024." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 295, + 274 + ], + "type": "text", + "content": "[40] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with deep language understanding. In Advances in Neural Information Processing Systems, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 295, + 329 + ], + "type": "text", + "content": "[41] Christoph Schuhmann, Richard Vencu, Romain Beaumont, Robert Kaczmarczyk, Clayton Mullis, Aarush Katta, Theo Coombes, Jenia Jitsev, and Aran Komatsuzaki. Laion-400m: Open dataset of clip-filtered 400 million image-text pairs. arXiv preprint arXiv:2111.02114, 2021. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 296, + 386 + ], + "type": "text", + "content": "[42] Ming Tao, Hao Tang, Fei Wu, Xiao-Yuan Jing, Bing-Kun Bao, and Changsheng Xu. Df-gan: A simple and effective baseline for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16515-16525, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 442 + ], + "type": "text", + "content": "[43] Gemini Team, Petko Georgiev, Ving Ian Lei, Ryan Burnell, Libin Bai, Anmol Gulati, Garrett Tanzer, Damien Vincent, Zhufeng Pan, Shibo Wang, et al. Gemini 1.5: Unlocking multimodal understanding across millions of tokens of context. arXiv preprint arXiv:2403.05530, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "text", + "content": "[44] Zhihua Tian, Sirun Nan, Ming Xu, Shengfang Zhai, Wenjie Qu, Jian Liu, Kui Ren, Ruoxi Jia, and Jiaheng Zhang. Sparse autoencoder as a zero-shot classifier for concept erasing in text-to-image diffusion models. arXiv preprint arXiv:2503.09446, 2025.1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 501, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 599 + ], + "type": "text", + "content": "[45] Cristina Nader Vasconcelos, Abdullah Rashwan, Austin Waters, Trevor Walker, Keyang Xu, Jimmy Yan, Rui Qian, Yeqing Li, SHIXIN LUO, Yasumasa Onoe, Zarana Parekh, Ivana Kajic, Mandy Guo, Wenlei Zhou, Sarah Rosston, Roopal Garg, Hongliang Fei, Jordi Pont-Tuset, Su Wang, Henna Nandwani, Andrew Bunner, Kevin Swersky, David J. Fleet, Oliver Wang, and Jason Michael Baldridge. Greedy growing enables high-resolution pixel-based diffusion models. Transactions on Machine Learning Research, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 601, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 295, + 633 + ], + "type": "text", + "content": "[46] Zihao Wang, Lin Gui, Jeffrey Negrea, and Victor Veitch. Concept algebra for text-controlled vision models. arXiv preprint arXiv:2302.03693, 2, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 635, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 295, + 666 + ], + "type": "text", + "content": "[47] WikiArt.org. WikiArt - Visual Art Encyclopedia. https: //www.wikiart.org/, 2024. Accessed: June 30, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 295, + 713 + ], + "type": "text", + "content": "[48] Olivia Wiles, Chuhan Zhang, Isabela Albuquerque, Ivana Kajic, Su Wang, Emanuele Bugliarello, Yasumasa Onoe, Chris Knutsen, Cyrus Rashtchian, Jordi Pont-Tuset, et al. Revisiting text-to-image evaluation with gecko: On met" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 597 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "rics, prompts, and human ratings. arXiv preprint arXiv:2404.16820, 2024. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 161 + ], + "type": "text", + "content": "[49] Tao Xu, Pengchuan Zhang, Qiuyuan Huang, Han Zhang, Zhe Gan, Xiaolei Huang, and Xiaodong He. Attngan: Fine-grained text to image generation with attentional generative adversarial networks. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1316-1324, 2018. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 206 + ], + "type": "text", + "content": "[50] Zeyue Xue, Guanglu Song, Qiushan Guo, Boxiao Liu, Zhuofan Zong, Yu Liu, and Ping Luo. Raphael: Text-to-image generation via large mixture of diffusion paths. Advances in Neural Information Processing Systems, 36, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[51] Mert Yuksekgonul, Federico Bianchi, Pratyusha Kalluri, Dan Jurafsky, and James Zou. When and why vision-language models behave like bags-of-words, and what to do about it? In ICLR, 2022. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 307 + ], + "type": "text", + "content": "[52] Gong Zhang, Kai Wang, Xingqian Xu, Zhangyang Wang, and Humphrey Shi. Forget-me-not: Learning to forget in text-to-image diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1755-1764, 2024. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 373 + ], + "type": "text", + "content": "[53] Han Zhang, Tao Xu, Hongsheng Li, Shaoting Zhang, Xiaogang Wang, Xiaolei Huang, and Dimitris N Metaxas. Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 5907-5915, 2017. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 375, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 375, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 375, + 553, + 441 + ], + "type": "text", + "content": "[54] Yimeng Zhang, Xin Chen, Jinghan Jia, Yihua Zhang, Chongyu Fan, Jiancheng Liu, Mingyi Hong, Ke Ding, and Sijia Liu. Defensive unlearning with adversarial training for robust concept erasure in diffusion models. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 1, 2, 4, 5, 7, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 442, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 442, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 442, + 553, + 495 + ], + "type": "text", + "content": "[55] Yihua Zhang, Yimeng Zhang, Yuguang Yao, Jinghan Jia, Jiancheng Liu, Xiaoming Liu, and Sijia Liu. Unlearncanvas: A stylized image dataset to benchmark machine unlearning for diffusion models. arXiv preprint arXiv:2402.11846, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 498, + 553, + 551 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 551 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 551 + ], + "type": "text", + "content": "[56] Minfeng Zhu, Pingbo Pan, Wei Chen, and Yi Yang. Dm-gan: Dynamic memory generative adversarial networks for text-to-image synthesis. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5802-5810, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 554, + 553, + 597 + ], + "type": "text", + "content": "[57] Xiaojin Zhu, Andrew B Goldberg, Mohamed Eldawy, Charles R Dyer, and Bradley Strock. A text-to-picture synthesis system for augmenting communication. In AAAI, pages 1590-1595, 2007. 2" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "16430" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_content_list.json b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5613690b8243ad6db95e6d6184056ee012a52609 --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_content_list.json @@ -0,0 +1,1592 @@ +[ + { + "type": "text", + "text": "Error Recognition in Procedural Videos using Generalized Task Graph", + "text_level": 1, + "bbox": [ + 140, + 130, + 854, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shih-Po Lee Northeastern University", + "bbox": [ + 228, + 181, + 437, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "lee.shih@northeastern.edu", + "bbox": [ + 230, + 219, + 452, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Ehsan Elhamifar Northeastern University", + "bbox": [ + 540, + 181, + 733, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "e.elhamifar@northeastern.edu", + "bbox": [ + 511, + 219, + 759, + 232 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition.", + "bbox": [ + 88, + 301, + 483, + 650 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 676, + 220, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The integration of computer vision with augmented and virtual reality (AR & VR) is expected to significantly improve the way we learn and develop skills in our daily lives [83]. Imagine wearing a pair of AR glasses and trying to assemble a new piece of furniture. A visual task assistant should recognize your actions and provide guidance on what to do and how to perform them throughout the process. It should also recognize your possible errors and provide feedback to correct them, when needed. This has motivated exciting new research on action recognition [8, 10, 25, 26, 47, 50, 61, 89] and segmentation [23, 38, 45, 46, 54, 57, 58, 73, 77, 78, 84, 95, 96, 103], video grounding [5, 18, 20, 49, 59, 63, 86, 109], learning", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e3621cf74a1da1b7fca414e5900e646b458868cb1d4fe17cbe38b7e0f86d8862.jpg", + "image_caption": [ + "Figure 1. The illustration of our framework for joint temporal action segmentation and error recognition. It consists of 1) Generalized Task Graph (GTG)-Video Alignment for action segmentation and 2) an Error Recognition Module using features from LLM/VLM for error recognition." + ], + "image_footnote": [], + "bbox": [ + 524, + 268, + 903, + 500 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "from instructional videos [12, 33, 62, 64, 65, 85, 102] and very recently on error understanding [14, 27, 29, 41, 60, 90], as well as progress [17] and state [100] prediction.", + "bbox": [ + 511, + 564, + 905, + 609 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Successful operation of a visual task assistant, in particular, requires temporal action segmentation (assigning step labels to frames) and error recognition (recognizing the type of error, if any, in each frame) to seamlessly work together. Additionally, a task graph, which encodes plausible ways to perform a task, has been shown to be critical for temporal action segmentation to effectively mitigate oversegmentation [5, 19, 77, 106]. However, existing methods have three major limitations when it comes to understanding actions and errors.", + "bbox": [ + 511, + 609, + 905, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "SOTA Limitations. First, existing works on error understanding can only detect the presence of an error in a video frame without recognizing the type of the error. As recently studied in [41], different types of errors, such as omission, addition, modification, and slip errors, can occur during task execution. However, [41] cannot distinguish between the error types. A task assistant requires understanding the error type to provide appropriate feedback. For example, an actor knocking over a dripper while making coffee requires", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10009", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "immediate feedback for correction or spreading jelly onto tortilla using spoon instead of knife needs a feedback after task completion for the actor to better execute the step.", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Second, given the large number of possible errors, it is infeasible to gather real training videos and annotations for all possible errors. Also, recent datasets [41, 66] that contain videos with errors have imbalance annotations on different error types, making error recognition a challenging task, especially under the setting when we do not have error videos during training.", + "bbox": [ + 89, + 136, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Third, recent action segmentation methods [19, 77] that use task graphs assume that there are no errors in videos and assign step labels to the frames according to the task graph. In addition, two recent works that handle both action segmentation and error detection [29, 41] cannot use the task graph for action segmentation. This is because a task graph only contains the correct actions/steps and does not consider possible errors that may occur during the task execution. Thus, following the task graph leads to assigning every frame (including frames that have errors and ones that do not belong to any step) to correct actions/steps, which is undesired.", + "bbox": [ + 89, + 242, + 480, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Paper Contributions. We develop a unified error recognition and temporal action segmentation framework (see Figure 1) to address the aforementioned limitations.", + "bbox": [ + 89, + 426, + 480, + 469 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a Generalized Task Graph (GTG) whose nodes encode normal actions/steps and background (task-irrelevant actions).", + "- We develop a dynamic programming-based method, GTG-Video Alignment (GTG2Vid), with a dynamic frame and node dropping term to assign frames to GTG nodes for action segmentation and drop frames for error detection.", + "- Given the lack of error training videos, we leverage the rich knowledge of large language models (LLMs) in conjunction with the rich representation power of video language models (VLMs) to obtain error descriptions for different types of errors and their corresponding features for error recognition.", + "- We propose an Error Recognition Module (ERM) to recognize the error types, if any, in the frames predicted by GTG2Vid using joint similarities of video frame, normal type, and error type features.", + "- By extensive experiments on two egocentric datasets, EgoPER and CaptainCook4D, we show that without using error videos during training, our method 1) improves the performance of temporal action segmentation on error videos, 2) improves error detection, and 3) recognizes different errors." + ], + "bbox": [ + 89, + 473, + 480, + 829 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 842, + 238, + 858 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Temporal Action Segmentation. Learning to segment videos into several actions/steps has been studied under un", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "supervised [4, 21, 22, 28, 36, 37, 74, 79, 88, 91, 109], weakly-supervised [11, 16, 24, 42-44, 48, 55, 55, 56, 69-71, 76, 84], and fully-supervised [3, 7, 9, 15, 23, 31, 32, 35, 38, 45, 46, 51, 57, 73, 80-82, 84, 95, 96, 103, 104] setting. Since procedural step videos follow certain task graphs to complete the task, some prior works [5, 19, 77, 106] additionally adopt task graph information to guide segmentation process. Specifically, G2Vid [19] aims to assign steps/nodes in a task graph to video frames and drop the frames not corresponding to any node. However, G2Vid cannot drop unnecessary nodes in a task graph nor map background frames to nodes to distinguish between background and error frames. In contrast, our framework adds extra nodes to a task graph and allows droppable nodes to properly assign nodes to frames.", + "bbox": [ + 511, + 90, + 903, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Error Detection and Recognition for Procedural Tasks. Detecting errors for procedural tasks has recently drawn attention from the video understanding community. Recent works [14, 29, 68, 90] have released relevant datasets to facilitate error understanding. Meanwhile, several works [27, 41, 60, 75] have developed frameworks to detect if errors are present in videos and localize their temporal locations. Recent work [27] focuses on online error detection using action recognition and anticipation models based on videos and action sequences. Another work [41] uses only error-free (normal) videos during training to generate multiple prototypes for action segmentation and error detection. However, none of them has the ability to recognize error types. Our work is able to segment videos and recognize different errors simultaneously without error training videos.", + "bbox": [ + 511, + 323, + 903, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLMs/LLMs for Anomaly Detection. Anomaly detection in videos [2, 6, 13, 39, 40, 53, 67, 72, 87, 92, 97-99, 105, 107, 108] has been widely explored, and recent studies [34, 101] leverage VLMs and LLMs for further improvement due to their outstanding reasoning capabilities. Recent work in [101] proposes a training-free video anomaly detection framework using VLMs to generate captions over frames followed by LLMs to predict anomaly scores for those captions. Another work in [34] adopts LLMs to generate normal and abnormal descriptions under various scenes and classify frames into normal or abnormal according to the similarities between features of frames and descriptions. In our paper, we leverage LLMs to generate error descriptions for each error type of every step with crafted prompts, followed by VLMs to generate textual features aligned with visual features for error recognition.", + "bbox": [ + 511, + 570, + 903, + 811 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Graph to Video Alignment Review", + "text_level": 1, + "bbox": [ + 511, + 828, + 823, + 845 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our proposed GTG-Video Alignment method builds upon and generalizes the Graph-to-Video Alignment (G2Vid) [19], which we review here. A task graph (also called flow", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10010", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "graph) is a directed acyclic graph $G = (V, E)$ , where each node $v \\in V$ represents a procedure step and each edge $(v_k, v_i) \\in E$ from $v_k$ to $v_i$ implies that $v_k$ must be completed before proceeding to $v_i$ , see Figure 2(a). Thus, all predecessors of a node must be completed before starting that step. Thus, the task graph encodes all possible ways to perform a task (this can be obtained by topological sorting).", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To efficiently align a video with the task graph and therefore segment the video into task steps, G2Vid [19] first builds a tSort graph (see Figure 2(a)) that provides a compact representation of all topological sorts of the task graph (each path from the source to the sink node in the tSort graph represents a topological sort of $G$ ). G2Vid alignment recovers a traversal of the tSort graph that best aligns with the video. It uses a modified version of Drop-DTW [18] to align nodes and frames, allowing for i) one-to-many matching (one node/step can be assigned to many frames) and ii) unmatched sequence elements (to be able to drop frames that do not correspond to any node/step).", + "bbox": [ + 88, + 198, + 482, + 378 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $C_{i,j}$ denote the matching cost of node $i$ to frame $j$ , $d_j$ denote the dropping cost of frame $j$ , and $A(i)$ denote the set of predecessors of $i$ . The dynamic programming solution computes the cumulative matching $D_{i,j}^{+}$ , dropping $D_{i,j}^{-}$ and optimal $D_{i,j}$ costs up to the $i$ -th node and $j$ -th frame as", + "bbox": [ + 89, + 378, + 482, + 455 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i, j} ^ {+} = C _ {i, j} + \\min \\left\\{\\min _ {k \\in A (i)} \\left\\{D _ {k, j - 1} \\right\\}, D _ {i, j - 1} \\right\\}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 464, + 482, + 487 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i, j} ^ {-} = d _ {j} + D _ {i, j - 1}, \\quad D _ {i, j} = \\min \\{D _ {i, j} ^ {+}, D _ {i, j} ^ {-} \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 487, + 423, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "By iterating over all nodes and frames, the dynamic programming outputs the optimal cost and alignment by backtracking the path with minimum cost, starting from the top left and ending at the bottom right of the cumulative matrix.", + "bbox": [ + 89, + 508, + 482, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 583, + 264, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We develop a unified framework for Temporal Action Segmentation (TAS) and Error Recognition (ER) in procedural task videos. We assume having access to only normal (error-free) videos with their frame-wise step labels during training and the task graph during testing. Our proposed framework addresses three fundamental questions: (1) How can we build an efficient Generalized Task Graph (GTG) that enables TAS and Error Detection (ED) for videos? (2) How can we obtain features for different errors without collecting real error videos? (3) How can we leverage the error features to perform ER? In the following, we first formalize the problem, discuss the overarching goal of our framework, and discuss the details of the method to address these questions.", + "bbox": [ + 89, + 608, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Setting. For TAS and ER, given a sequence of video frame features $\\mathcal{X}^v = (x_1^v,x_2^v,\\ldots ,x_T^v)$ , our goal is to predict its frame-wise steps $\\hat{\\mathcal{Y}} = (\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_T)$ and error types $\\hat{\\mathcal{E}} = (\\hat{e}_1,\\hat{e}_2,\\dots ,\\hat{e}_T)$ , where $T$ is the number of frames, $x_{t}^{v}\\in \\mathbb{R}^{D}$ is a pre-extracted feature vector, $\\hat{y}_t\\in$", + "bbox": [ + 89, + 823, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a05146f0803a0830bf4e86937829de358e4db963586857d11a865f3a7259552c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 94, + 877, + 188 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/eed64a93e6f6d9f4cc02ef338eef0d57fb41cad71ff57662bcc2080c0a8012f6.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 544, + 189, + 877, + 268 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ee92c696037b5102bce497ccf0091ec0a7ba29d6e38758b6c259309a9745275.jpg", + "image_caption": [ + "Figure 2. The illustration of (a) an original task graph and its corresponding tSort graph, (b) a naive approach to capture background, (c) our proposed solution, which specifically illustrates the predecessors and successor of extra node $S_{3}$ ." + ], + "image_footnote": [], + "bbox": [ + 544, + 268, + 877, + 349 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$\\{-1,0,\\ldots ,S\\}$ , where $S$ is the number of steps and class $-1$ and $0$ indicate error and background (a task-irrelevant action, e.g., answering phone when making coffee), $\\hat{e}_t\\in \\{0,\\dots ,M\\}$ , where $M$ is the number of error types and class $0$ denotes no error happening. During training, we use only normal (error-free) videos and their ground-truth step labels $\\mathcal{V}$ . For testing, we assume that the task graph is given.", + "bbox": [ + 511, + 416, + 905, + 523 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "**Framework Overview.** We introduce Generalized Task Graph for handling both $TAS$ and $ED$ . Our key idea is to 1) insert extra nodes in the tSort graph to capture the background class that may occur before each step in the task graph, 2) perform $TAS$ and $ED$ by our GTG2Vid with a dynamic frame and node dropping term for different types of errors, 3) obtain error descriptions and features for all types of errors, respectively, by leveraging LLMs and VLMs and 4) use an Error Recognition Module (ERM) that computes joint similarities of video, normal type and error type features obtained by VLMs to classify the error types for the frames flagged by GTG2Vid.", + "bbox": [ + 511, + 527, + 906, + 709 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. Generalized Task Graph (GTG)", + "text_level": 1, + "bbox": [ + 511, + 720, + 795, + 736 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We build GTG to handle execution and omission errors and develop GTG2Vid for TAS and ED.", + "bbox": [ + 511, + 744, + 903, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Handling Execution Errors. We aim to use the task graph to segment videos into steps and to detect execution errors (e.g., slip, correction, modification, and addition errors defined in [41]). Some of the errors are step-relevant (e.g., spreading jelly using spoon instead of knife) and some are not in the task graph (e.g., adding honey to a recipe that does not include honey). However, G2Vid drops both frames with errors as well as background frames (e.g., answering", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10011", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 GTG-Video Alignment (GTG2Vid)" + ], + "code_body": "1: Inputs: $C\\in \\mathbb{R}^{2S\\times T}$ $A(\\cdot),d_i^z,d_j^x$ (matching cost matrix, predecessor mapping dictionary, node dropping costs, frame dropping costs.) $\\triangleright$ Initialize DP tables \n2: $D_{0,0}^{+} = 0;D_{i,0}^{+} = \\infty ;D_{0,j}^{+} = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}$ Matching table of $\\mathcal{X}_n^v$ (frames) and $\\mathcal{Z}$ (nodes) \n3: $D_{0,0}^{z - } = 0;D_{i,0}^{z - } = \\sum_{k = 1}^{i}d_{k}^{z};D_{0,j}^{z - } = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}$ Dropping table of $\\mathcal{Z}$ (nodes) \n4: $D_{0,0}^{-x} = 0;D_{i,0}^{-x} = \\infty ;D_{0,j}^{-x} = \\sum_{k = 1}^{j}d_{k}^{x};\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}$ Dropping table of $\\mathcal{X}_n^v$ (frames) \n5: $D_{0,0} = 0;D_{i,0} = D_{i,0}^{-x};D_{0,j} = D_{0,j}^{z - }; \\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}$ Optimal solution table \n6: for $i = 1,\\dots ,2S$ do Iterate nodes in Z \n7: for $j = 1,\\dots ,T$ do Iterate frames in $\\mathcal{X}_n^v$ \n8: $D_{i,j}^{+} = -C_{i,j} + \\min \\{\\min_{k\\in A(i)}\\{D_{k,j - 1}\\},D_{i,j - 1}\\}$ Consider matching node $\\mathcal{Z}_i$ to frame $x_{n,j}^v$ \n9: $D_{i,j}^{-x} = d_j^x +D_{i,j - 1}$ Consider dropping frame $x_{n,j}^v$ \n10: $D_{i,j}^{z - } = d_i^z +\\min_{k\\in A(i)}\\{D_{k,j}\\}$ Consider dropping node $\\mathcal{Z}_i$ \n11: $D_{i,j} = \\min \\{D_{i,j}^{+},D_{i,j}^{-x},D_{i,j}^{z - }\\}$ Select the optimal path \n12: end for \n13: end for \n14: $\\hat{y}^* =$ traceback(D) $\\triangleright$ Compute the optimal alignment by tracing back the path with minimum cost \n15: $\\hat{\\gamma} =$ frame-assign(y*) Flag dropping frames as error frames and frames belonging to extra nodes as background \n16: Outputs: $D_{2S,T},\\hat{\\gamma}$", + "bbox": [ + 93, + 108, + 906, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the phone while making a recipe) and thus cannot distinguish between them and understand when an error occurs.", + "bbox": [ + 89, + 401, + 480, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To handle this, we build GTG by inserting extra nodes into the task graph to capture background frames. Therefore, by aligning a video with GTG, we can understand which frames correspond to which steps and which frames correspond to background. Additionally, by allowing dropping frames, we can detect execution errors, and by allowing dropping nodes, we can detect omission errors which we will discuss next.", + "bbox": [ + 89, + 431, + 482, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A naive approach is to insert an extra node with a unique label between every two nodes (see Figure 2 (b)). Therefore, the task graph can potentially map frames to extra nodes. However, the complexity of the tSort graph grows exponentially. To solve this problem, our proposed approach inserts an extra node $S_{i}$ with a unique ID between every node $i$ and its predecessor in the tSort graph. For example, in Figure 2 (c), between node 2 and 3 as well as node 1 and 3, we insert the same extra node $S_{3}$ to capture the background frames with respect to step 3. Our design 1) prevents exponential growth of the tSort graph, only increasing the number of nodes from $S$ to $2S$ , and 2) inherits the correct execution orders in the original tSort graph.", + "bbox": [ + 89, + 551, + 482, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Handling Omission Errors. During execution, an actor may skip one or more steps in the task graph, resulting in omission error. In this case, forcing models to follow the task graph as done in prior works [5, 19, 77, 106] can lead to wrong TAS results where some frames are assigned to the skipped steps. To handle omission errors, we introduce a node dropping term to our GTG2Vid to make nodes of our GTG droppable,", + "bbox": [ + 89, + 752, + 482, + 873 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nD _ {i, j} ^ {z -} = d _ {i} ^ {z} + \\min _ {k \\in A (i)} \\left\\{D _ {k, j} \\right\\} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{i,j}^{z-}$ is the accumulative cost until node $i$ and frame $j$ in the dynamic programming (DP) table of dropping node and $d_i^z$ is the node dropping cost of node $i$ . A low dropping cost indicates that the corresponding GTG node is not in the video and should be omitted. This also allows dropping extra nodes, whenever there is no corresponding frame in the video.", + "bbox": [ + 511, + 401, + 906, + 506 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GTG-Video Alignment (GTG2Vid). Algorithm 1 shows the steps of GTG2Vid. $C_{i,j}$ is the matching cost of the $i$ -th node to the $j$ -th frame. We let $C_{i,j}$ to be the predicted probability of step $i$ for $j$ -th frame using an action segmentation model. We train the action segmentation model with smoothing and cross-entropy loss in [23] using framewise labels. In addition, to ensure that our proposed method can drop frames and nodes based on the confidence of predictions, we propose a dynamic dropping cost $d^{x}$ for frames and $d^{z}$ for nodes of GTG as", + "bbox": [ + 511, + 510, + 908, + 660 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {j} ^ {x} = - \\left(\\frac {\\sum_ {i = 1} ^ {2 S} C _ {i , j} \\log C _ {i , j}}{\\log \\frac {1}{2 S}} + \\eta\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 667, + 903, + 712 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {i} ^ {z} = - \\tau \\left(1 - \\max _ {j = 1, \\dots , T} C _ {i, j}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 713, + 792, + 744 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\eta$ and $\\tau$ are hyper-parameters (larger $\\eta$ and $\\tau$ lead to dropping more frames and nodes). Our frame dropping cost $d^{x}$ leads to drop frames with unconfident predictions, indicating that the frames do not correspond to any GTG node and may contain errors. Meanwhile, our node dropping cost $d^{z}$ allows dropping the nodes whose highest step probabilities over all frames are low. After obtaining the minimum cost path, we generate step predictions $\\hat{\\mathcal{V}}$ from $\\hat{\\mathcal{V}}^*$ by flagging the dropped frames as error frames and frames belonging to extra nodes as background. The complexity of our al", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10012", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/2de1b8e6b76815aac0da19092269e383b54895a79c7faf4dfd010c6750db33b6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTea
w-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAcc
Naive Predictor11.225.013.025.010.425.018.133.311.425.0
EgoPED (MV)7.575.05.575.07.475.010.9100.04.475.0
GTG2Vid (MV)5.150.02.750.04.450.00.633.34.050.0
GTG2Vid (w/o n.f.)16.375.016.475.029.9100.01.733.315.275.0
GTG2Vid31.7100.031.3100.017.8100.04.5100.022.1100.0
", + "bbox": [ + 109, + 88, + 883, + 200 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b6777a85b5711133c60a662e98b1170dd0bca1828787aa5aa2b6fb84fd6f113c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Hot Chocolate \nw-F1@0EAccCaptainCook4D
Sandwich \nw-F1@0EAccBurritos \nw-F1@0EAccRamen \nw-F1@0EAccRaita \nw-F1@0EAcc
Naive Predictor7.120.05.120.07.425.07.425.07.333.3
EgoPED (MV)7.380.03.480.07.175.03.375.04.366.6
GTG2Vid (MV)4.460.04.740.06.750.02.725.011.466.6
GTG2Vid (w/o n.f.)2.840.06.260.06.125.04.650.011.266.6
GTG2Vid10.5100.08.5100.09.475.04.475.023.0100.0
", + "bbox": [ + 109, + 202, + 883, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. ${ER}$ results on EgoPER and CaptainCook4D.", + "bbox": [ + 352, + 311, + 643, + 324 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "girthm is $\\mathcal{O}(SQT)$ , same as G2Vid [19], where $S, Q$ , and $T$ denote the number of GTG nodes, separate and linearly-ordered threads of GTG, and frames, respectively.", + "bbox": [ + 89, + 342, + 483, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Obtaining Error Features via LLM/VLM", + "text_level": 1, + "bbox": [ + 89, + 398, + 444, + 415 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our next step is to obtain error features for $ER$ . Since we are only given normal videos during training, we i) leverage LLMs to obtain error descriptions for each error type for every step, and ii) use VLMs to generate the textual features that are aligned with visual features. To begin, given a step $i$ and error type $m$ , we generate its error type description using LLMs. The prompt for LLMs combines a description for step $i$ , error definition for error type $m$ , and the number of different error type descriptions to produce. Notice that we directly set normal type description $(m = 0)$ as the description of step $i$ . Next, we generate the feature vector $x_{i,m}^{e}$ for the error description using the text encoder in VLMs, where $x^{e} \\in \\mathbb{R}^{S \\times M \\times D}$ is a tensor containing features of $M$ error type descriptions for $S$ steps. When LLMs are prompted to produce multiple error descriptions for an error type, we average all features to get a single vector. Meanwhile, we use the video encoder in VLMs to generate frame features for videos. In this way, we can ensure that the features of videos and error descriptions are aligned in the VLM feature space. See the supplementary materials for detailed prompts and generated error descriptions.", + "bbox": [ + 88, + 421, + 483, + 741 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Error Recognition Module (ERM)", + "text_level": 1, + "bbox": [ + 89, + 752, + 390, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our ERM classifies error types for error frames according to their joint similarities of frame, normal type, and error type features from Section 4.2 and obtain $ER$ predictions $\\hat{\\mathcal{E}}$ .", + "bbox": [ + 89, + 773, + 482, + 821 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Predicting Steps for Error Frames. First, we obtain TAS predictions $\\hat{\\mathcal{V}}$ by inferring GTG2Vid on a video. Next, we obtain another TAS predictions $\\hat{\\mathcal{V}}^{+} = (\\hat{y}_{1}^{+},\\hat{y}_{2}^{+},\\dots ,\\hat{y}_{T}^{+})$ where $\\hat{y}_t^+\\in \\{0,1,\\ldots ,S\\}$ , by inferring GTG2Vid without using the frame and node dropping term, forcing GTG2Vid", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to map all frames to nodes. It aims to obtain the step predictions for the error frames in $\\hat{\\mathcal{V}}$ to find the error type features of the corresponding steps.", + "bbox": [ + 511, + 342, + 903, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Computing Joint Similarities for Error Recognition. Next, we compute framewise joint similarities $p$ for $ER$ , where $p_{m,j} \\in [0,1]$ and $p \\in \\mathbb{R}^{M \\times T}$ . For each frame $j$ , given step prediction $\\hat{y}_j, \\hat{y}_j^+$ , and feature $x_j^v, x^e$ , we find the error type $\\hat{m}_j$ and compute its joint similarity $p_{\\hat{m}_j,j}$ as", + "bbox": [ + 511, + 390, + 906, + 468 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {m} _ {j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ \\underset {m \\in \\{1, \\dots , M \\}} {\\operatorname {a r g m a x}} s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, m\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 484, + 906, + 532 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\hat {m} _ {j}, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, \\hat {m} _ {j}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 540, + 905, + 579 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $s(\\cdot ,\\cdot ,\\cdot ,\\cdot)$ is our score function that outputs the joint similarity of the frame. Our score function is formulated as", + "bbox": [ + 511, + 583, + 905, + 613 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\ns \\left(x _ {j} ^ {v}, x ^ {e}, y, m\\right) = \\operatorname {s i g m o i d} \\left(\\left(x _ {y, 0} ^ {e} \\odot x _ {y, m} ^ {e}\\right) ^ {T} x _ {j} ^ {v}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 633, + 903, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ is element-wise multiplication and $x_{y,0}^{e}$ is the feature of step description (normal type) for step $y$ . The intuition of using normal type feature is that error frames often contain multiple sub-steps, including both normal and error ones (e.g., an actor opens a tortilla bag, grabs a tortilla from it, and accidentally drops it). A single error type feature may be unable to effectively match the corresponding error frames, especially when the number of frames is large and contain multiple sub-steps which the error description does not describe. Therefore, our score function computes the correlation between $x_{y,0}^{e}, x_{y,m}^{e}$ , and $x_{j}^{v}$ . Thus, it jointly considers the similarities of error frames with respect to both normal and error type features. We validate its effectiveness for $ER$ in Section 5. Finally, we apply a maximum filter on each error type of $p$ to smoothen the predictions, followed by argmax to obtain the error type predictions $\\hat{\\mathcal{E}}$ .", + "bbox": [ + 511, + 657, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10013", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/0c2cedf2080c3baa2350295406cee1c50cbcba29aac1f98ed2a5fcb14aeed9f5.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodQuesadillaOatmealEgoPER
F1@.5AccEditF1@.5AccEditPinwheelCoffeeTeaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid57.053.760.068.874.576.849.152.753.759.157.175.061.156.169.459.058.870.0
EgoPED56.961.447.362.154.870.349.854.741.053.959.655.469.464.558.356.459.054.5
GTG2Vid68.164.569.483.378.588.661.358.460.963.962.573.771.262.676.969.665.373.9
MethodHot ChocolateSandwichBurritosRamenRaitaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid17.119.223.821.321.739.833.427.542.928.033.950.518.523.126.123.725.136.6
EgoPED5.021.713.313.122.518.221.131.821.815.737.325.422.631.323.015.528.920.3
GTG2Vid20.843.527.827.238.538.836.634.041.331.333.642.922.130.826.127.636.135.4
", + "bbox": [ + 109, + 88, + 885, + 258 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. TAS results on EgoPER and CaptainCook4D.", + "bbox": [ + 351, + 262, + 643, + 275 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 292, + 223, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 319, + 279, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset. We evaluate our proposed method on EgoPER [41] and CaptainCook4D [66]. EgoPER consists of 5 tasks with 386 egocentric videos and contains 5 types of errors, including omission, addition, modification, slip, and correction. Captaincook4D has 24 tasks with 94.5 hours and contains 8 types of errors, including preparation, measurement, technique, timing, temperature, missing steps, ordering, and other. For EgoPER, we evaluate our method on all 5 tasks and error types with the same training and testing split in [41]. For CaptainCook4D, to have enough videos for training and testing, we select 5 of the tasks which have at least 5 normal and error videos: Spiced Hot Chocolate (Hot Chocolate), Microwave Egg Sandwich (Sandwich), Breakfast Burritos (Burritos), Ramen, and Cucumber Raita (Raita). We assign all normal videos to the training set and error videos to test set. To fit our setting for recognizing execution errors, we report the following error types in CaptainCook4D: preparation (Prep.), measurement (Mea.), technique (Tec.), timing (Time), and temperature (Temp). See the supplementary materials for breakdown analysis of EgoPER and CaptainCook4D and their detailed training and testing splits.", + "bbox": [ + 89, + 348, + 483, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We report results on $ER$ , $TAS$ , and $ED$ . In our experiments, we report the F1 score $(F1@\\beta)$ with an overlapping threshold $\\beta$ . We compute the F1 score by averaging (if not specified) over the F1 scores that are independently computed on the segments for each step or error type. For $ER$ , we report Error Accuracy (EAcc) and weighted $F1@0$ score $(w-F1@0)$ . As $ER$ is challenging under the setting of no error video during training, we set $\\beta$ to 0, meaning that a segment is true positive whenever any frame in the segment is correctly classified. EAcc is the percentage of error types that a model can recognize; that is, the error types with $F1@0$ score larger than 0. For $w-F1@0$ , we first compute $F1@0$ on the error type segments, excluding normal ones $(m=0)$ , to alleviate the effect of data imbalance (e.g.,", + "bbox": [ + 89, + 688, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "nearly $90\\%$ of the segments in EgoPER are normal). We then compute w-F1@0, where w-F1@0 = $\\frac{\\mathrm{F1}@\\mathrm{0}}{\\mathrm{M} - \\mathrm{EAcc} \\times \\mathrm{M} + 1}$ . w-F1@0 penalizes the model that fails to recognize all error types. For TAS, we report frame-wise Accuracy (Acc), segment-wise Edit score (Edit) on normal steps, and F1@.5 score. For ED, we report F1@.5 score on all segments. We also report F1@.5 score on normal (N.) and error segments (E.) separately for alleviating data imbalance. For omission error detection, we use Omission Accuracy (O-Acc), predicting the accuracy for omission steps, and Omission Intersection over Union (O-IoU), as in [41].", + "bbox": [ + 511, + 292, + 906, + 460 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Given the lack of baselines for joint TAS, $ER$ , and $ED$ , we report the performance of EgoPED (MV), a modified version of EgoPED [41] that can additionally perform $ER$ . EgoPED (MV) computes the cosine similarities between frames and error types using the error type features generated in Section 4.2. Then it assigns frames to the type with the highest similarity and uses the majority voting (MV) strategy in [41] over frames within a segment to decide the final error type for each segment. EgoPED (MV) detects error segments with step thresholds equal to mean similarities over videos in the validation set. For $ER$ and $ED$ , since both datasets have data imbalance, we report the performance of Naive Predictor as another baseline. Naive Predictor classifies frames into the same error type for $ER$ and into normal frames for $ED$ . Specifically, it classifies the frames into modification errors for EgoPER and preparation errors for CaptainCook4D. For $TAS$ , we compare with G2Vid [19] to show how G2Vid performs with standard task graphs. To run G2Vid with our cost matrix, we remove all the background entries.", + "bbox": [ + 511, + 468, + 908, + 772 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We employ GPT-4o mini [1] as our LLM for generating error descriptions and VideoCLIP [93] pre-trained on Ego4D [30] as our VLM for extracting both visual and textual features for videos and error type descriptions. For each task in both EgoPER and CaptainCook4D, we generate at least 2 descriptions for each error type of each action. We set the step description as the description of normal type. We use the backbone in [52] as", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10014", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/91a573e02fa094b6ecdc9ad205759c111d62f3592a2bb2957240a899753c8856.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTeaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor14.40.07.226.10.013.123.40.011.744.10.022.14.80.02.422.60.011.3
EgoPED30.125.227.729.312.921.114.624.519.69.58.99.236.334.035.124.021.122.5
GTG2Vid40.422.631.558.536.447.430.616.023.333.14.018.651.537.044.242.823.233.0
CaptainCook4D
Hot ChocolateSandwichBurritosRamenRaitaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor8.20.04.16.50.03.23.80.01.94.10.02.03.10.01.65.10.02.6
EgoPED8.12.45.216.36.811.516.38.212.212.410.711.614.19.511.813.47.510.5
GTG2Vid15.820.918.316.712.814.816.919.718.322.68.215.418.29.914.018.014.316.2
", + "bbox": [ + 117, + 88, + 877, + 262 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "our action segmentation model. We train the model using Adam optimizer with learning rate 0.0001 for 3000 iterations. For $ER$ , if $\\hat{y}_j = -1$ and $\\hat{y}_j^+$ is a background class, we directly assign an addition error and preparation error to $\\hat{m}_j$ for EgoPER and CaptainCook4D, respectively, and assign the corresponding similarity to 1. For $ED$ , we produce $ED$ predictions based on $\\hat{\\mathcal{E}}$ by classifying different error types into a single error type. In practice, we set $\\tau = S$ where $S$ is the number of nodes in GTG and $\\eta = 0$ .", + "bbox": [ + 89, + 295, + 483, + 431 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental Results", + "text_level": 1, + "bbox": [ + 89, + 438, + 292, + 453 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Error Recognition $(ER)$ . Table 1 compares $ER$ performance between different methods on EgoPER and CaptainCook4D. Our proposed framework, GTG2Vid, demonstrates the best recognition ability than Naive Predictor and EgoPED (MV). Specifically, our proposed method achieves $31.7\\%$ , $31.3\\%$ , $17.8\\%$ , and $22.1\\%$ on w-F1@0 compared to $7.5\\%$ , $5.5\\%$ , $7.4\\%$ , and $4.4\\%$ by EgoPED (MV) for quesadilla, oatmeal, pinwheel, and tea. On the other hand, our proposed method outperforms EgoPED (MV) on w-F1@0 for all tasks of CaptainCook4D. In addition, our proposed method achieves overall higher EAcc than EgoPED (MV), indicating that our method can recognize more error types on both datasets. Notice that our performance on w-F1@0 for coffee is worse than EgoPED (MV) because EgoPED predicts most of the segments as errors, shown in Table 3. Please see supplementary materials for the F1@0 score of each type of error.", + "bbox": [ + 89, + 462, + 483, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Next, we investigate the effectiveness of our ERM and our proposed score function (Eq. 6). By replacing ERM with MV, the performance drops significantly (more than $10\\%$ on w-F1@0 across all tasks of EgoPER). On the other hand, our method, GTG2Vid (w/o n.f.), without using normal type features suffers a drop on EAcc for most of the tasks, indicating that our proposed function for computing joint similarity has a better ability to recognize different error types. Furthermore, Table 4 shows the Top-1 and Top-2 accuracy of different scoring methods given the features of ground-truth error segments (average over frame features) and generated error type features. Our method gains a per", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/f5ba31dd5d4a7222eabc2307c6b076a98fc187f5dc7f13e326cbc2c4ad76b5c5.jpg", + "table_caption": [ + "Table 3. ED results of different methods on EgoPER and CaptainCook4D. N. and E. denote the F1@.5 score for normal and error segments." + ], + "table_footnote": [], + "table_body": "
MethodQuesadillaPinwheelTea
Top-1Top-2Top-1Top-2Top-1Top-2
Uniform25.050.025.050.025.050.0
Cos. Sim.42.665.827.748.427.845.3
ERM54.681.538.460.334.353.7
", + "bbox": [ + 519, + 291, + 897, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 4. Performance of feature matching methods on EgoPER. Top-1 and Top-2 denote the top-1 and top-2 accuracy over all error segments.", + "bbox": [ + 511, + 369, + 903, + 396 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "formance boost on Top-1 and Top-2 accuracy. Specifically, we obtain $10.7\\%$ improvement on Top-1 accuracy compared to cosine similarity (Cos. Sim.) for pinwheel, which verifies our claim in Section 4.3.", + "bbox": [ + 511, + 415, + 903, + 474 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Temporal Action Segmentation (TAS). For TAS performance shown in Table 2, our method obtains a $13.2\\%$ , $6.3\\%$ , and $19.4\\%$ higher F1@.5, Acc and Edit score than EgoPED on All of EgoPER. For CaptainCook4D, our method obtains a $12.1\\%$ , $7.2\\%$ , and $15.1\\%$ higher F1, Acc, and Edit score than EgoPED on All. Furthermore, our method has more robust TAS performance compared to EgoPED, which suffers a significant performance drop in Hot Chocolate and Sandwich. Notice that our Edit score on All is comparable to G2Vid since there are not many omission errors in test videos. Still, our method achieves a higher F1@.5 score than G2Vid, which cannot distinguish between background frames and errors.", + "bbox": [ + 511, + 481, + 906, + 676 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Error Detection (ED). Table 3 shows the $ED$ performance of different methods on EgoPER. Our method achieves $33.0\\%$ higher on F1@.5 score, compared to $22.5\\%$ by EgoPED for All in EgoPER. Although EgoPED can obtain a higher F1 score on error segments for quesadilla, pinwheel, and coffee, its detection ability for normal segments decreases significantly, specifically a $23.6\\%$ drop on F1@.5 score for coffee compared to our method. Therefore, compared to EgoPED, our method achieves a better balance between normal and error segments. For All in CaptainCook4D, GTG2Vid outperforms EgoPED, achieving $16.2\\%$ on F1@.5 compared to $10.5\\%$ by EgoPED.", + "bbox": [ + 511, + 683, + 905, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Ablation of Node and Frame Dropping Term. We show ablation studies and investigate the effectiveness of our pro", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10015", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5349db5d35961d469daadffdba256d69d20f1c92258ad904b8481d8cf19c743.jpg", + "image_caption": [ + "Figure 3. Qualitative results for quesadilla, tea, and oatmeal in EgoPER. The top part shows the results of TAS for ground-truths, EgoPED, and our method. The bottom part shows ER results for ground-truths, EgoPED, and our method. Seg. and E.T. denote action segmentation and error type, respectively." + ], + "image_footnote": [], + "bbox": [ + 138, + 94, + 864, + 256 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2e4dd901808e175f5aab560cb9062a1b11d4cccf2aab8bd3449d25e9b86292cc.jpg", + "image_caption": [ + "Figure 4. Ablation of our proposed node and frame dropping term." + ], + "image_footnote": [], + "bbox": [ + 98, + 296, + 477, + 435 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "posed node and frame dropping term in GTG2Vid. The two upper plots in Figure 4 show F1@.5 scores of our proposed frame dropping term (Ours) with different $\\eta$ and a naive dropping cost (G2Vid) [19] with different $k$ top percentile of the values in the cost matrix, for pinwheel and quesadilla. Our method achieves $26.0\\%$ and $33.8\\%$ on F1@.5 score compared to $23.2\\%$ and $30.9\\%$ by G2Vid. On the other hand, the two bottom plots in Figure 4 show the F1 scores for TAS with different $\\tau$ for our node dropping term. Smaller $\\tau$ indicates GTG2Vid drops a fewer number of nodes. When $\\tau$ is larger or equal to 0, GTG2Vid starts dropping nodes, improving TAS performance on videos with omission steps.", + "bbox": [ + 88, + 468, + 482, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In addition, Figure 5 shows the performance of omission detection for different methods on EgoPER. First, since G2Vid is unable to drop nodes of a graph, it cannot predict any omitted steps, resulting in $0\\%$ on O-Acc and O-IoU. Second, although our method suffers a performance drop on O-IoU ( $5.7\\%$ lower than EgoPED on All), we gain a performance boost on O-Acc ( $10\\%$ higher than EgoPED on All), indicating that our method successfully drops nodes of a task graph according to the input videos and has a stronger capability for detecting omitted steps. To conclude, our proposed method uses 1) dynamic frame dropping term to precisely drop error segments and 2) dynamic node dropping term to drop unnecessary nodes of task graphs for better TAS and omission detection performance.", + "bbox": [ + 89, + 665, + 482, + 876 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative Analysis. We visualize ER results in Fig. 3.", + "bbox": [ + 89, + 885, + 480, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d4438e99f1859d747038528fb7be64a76fe3a7a09c788b086040ab3a87a311a.jpg", + "image_caption": [ + "Figure 5. Omission detection results of different methods on EgoPER." + ], + "image_footnote": [], + "bbox": [ + 516, + 297, + 901, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our method can recognize different types of errors, e.g., addition errors (red) on tea, addition and correction (orange) errors on quesadilla, and modification errors (blue) on oatmeal. Our method obtains more accurate TAS, ED, and ER results for error videos on quesadilla, tea and oatmeal compared to EgoPED, thanks to dynamic frame and node dropping term in GTG2Vid and our proposed ERM. See supplementary materials for more qualitative visualization.", + "bbox": [ + 511, + 492, + 906, + 613 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 626, + 638, + 641 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We studied joint $ER$ and $TAS$ in procedural task videos. Our proposed framework consists of 1) generalized task graphs to capture correct steps and background, 2) a GTG-Video Alignment algorithm to find the optimal alignment between nodes in GTG and video frames with dynamic dropping terms, 3) a procedure for generating error descriptions for each error type of every step using LLMs and the corresponding features using VLMs, and 4) an Error Recognition Module to effectively perform $ER$ using joint similarities. Our experiments on two datasets showed that our proposed method can obtain promising results on $ER$ , $ED$ , and $TAS$ .", + "bbox": [ + 511, + 650, + 906, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 830, + 671, + 847 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was funded, in part, by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110), ONR (N000142512287) and ARO", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10016", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "(W911NF2110276). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the US Government.", + "bbox": [ + 89, + 90, + 485, + 151 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 162, + 187, + 179 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6", + "[2] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, , and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[3] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 2", + "[4] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2", + "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Neural Information Processing Systems, 2023. 1, 2, 4", + "[6] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. arXiv: 2110.09742, 2021. 2", + "[7] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2", + "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 1", + "[9] Daniele Di Mauro Mario Valerio Giuffrida Giovanni Maria Farinella Camillo Quattrocchi, Antonino Furnari. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 2", + "[10] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1", + "[11] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning" + ], + "bbox": [ + 101, + 188, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "in instructional videos. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI, 2020. 1", + "[13] Hanqiu Deng, Zhaoxiang Zhang, Shihao Zou, and Xingyu Li. Bi-directional frame interpolation for unsupervised video anomaly detection. IEEE Winter Conference on Applications of Computer Vision, 2023. 2", + "[14] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv: 2307.16453, 2023. 1, 2", + "[15] Guodong Ding, Hans Golong, and Angela Yao. Coherent temporal synthesis for incremental action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[16] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[17] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[18] Nikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan D Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 1, 3", + "[19] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 1, 2, 3, 4, 5, 6, 8", + "[20] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos Derpanis, Animesh Garg, Richard Wildes, and Allan Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1", + "[21] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2", + "[22] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2", + "[23] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4", + "[24] Mohsen Fayyaz and Jurgen Gall. Sect: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[25] C. Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1", + "[26] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10017", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Alessandro Flaborea, Guido Melendugno, Leonardo Pliniq, Luca Scofanoq, Edoardo Matteisq, Antonino Furnari, Giovanni Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2", + "[28] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2", + "[29] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2", + "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 6", + "[31] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2", + "[32] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2", + "[33] Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. European Conference on Computer Vision, 2024. 1" + ], + "bbox": [ + 99, + 92, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Jaehyun Kim, Seongwook Yoon, Taehyeon Choi, and Sanghoon Sull. Unsupervised video anomaly detection based on similarity with predefined text descriptions. Sensors, 2023. 2", + "[35] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition. IEEE Winter Conference on Applications of Computer Vision, 2016. 2", + "[36] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[37] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2", + "[38] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2", + "[39] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[40] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. Bman: Bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Transactions on Image Processing, 2019. 2", + "[41] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6", + "[42] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[43] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021.", + "[44] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2", + "[45] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[46] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2", + "[47] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers" + ], + "bbox": [ + 522, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10018", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "for classification and detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.1", + "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[49] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhan, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. European Conference on Computer Vision, 2024. 1", + "[50] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 1", + "[51] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 2", + "[52] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. 2023. 6", + "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[54] Yang Liu, Jiayu Huo, Jingjing Peng, Rachel Sparks, Prokar Dasgupta, Alejandro Granados, and Sebastien Ourselin. Skit: a fast key information video transformer for online surgical phase recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21074-21084, 2023. 1", + "[55] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2", + "[56] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[57] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2", + "[58] Z. Lu and E. Elhamifar. Multi-modal few-shot temporal action segmentation. International Conference on Computer Vision, 2025. 1", + "[59] Z. Lu, A. Iftekhar, G. Mittal, T. Meng, X. Wang, C. Zhao, R. Kukkala, E. Elhamifar, and M. Chen. Decafnet: Delegate and conquer for efficient temporal grounding in long videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1", + "[60] Antonino Furnari Luigi Seminara, Giovanni Maria Farinella. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. Neural Information Processing Systems, 2024. 1, 2" + ], + "bbox": [ + 99, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] K. Mangalam, H. Fan, Y. Li, C. Wu, B. Xiong, C. Feichtenhofer, and J. Malik. Reversible vision transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[62] A. Miech, J-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1", + "[63] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[64] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[65] Kumaranage Ravindu Yasas Nagasinghe, Honglu Zhou, Malitha Gunawardhana, Martin Renqiang Min, Daniel Harari, and Muhammad Haris Khan. Why not use your textbook? knowledge-enhanced procedure planning of instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[66] Rohith Peddi, Shivvrat Arya, Bharath Challa, Likhitha Pallapothula, Akshay Vyas, Bhavya Gouripeddi, Jikai Wang, Qifan Zhang, Vasundhara Komaragiri, Eric Ragan, Nicholas Ruozzi, Yu Xiang, and Vibhav Gogate. CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities, 2024. 2, 6", + "[67] Didik Purwanto, Yie-Tarng Chen, and Wen-Hsien Fang. Dance with self-attention: A new look of conditional random fields on anomaly detection in videos. IEEE International Conference on Computer Vision, 2021. 2", + "[68] Yicheng Qian, Weixin Luo, Dongze Lian, Xu Tang, Peilin Zhao, and Shenghua Gao. Svip: Sequence verification for procedures in videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[69] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2", + "[70] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018.", + "[71] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2", + "[72] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[73] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 1, 2" + ], + "bbox": [ + 522, + 92, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10019", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[75] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepop, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[76] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[77] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4", + "[78] Y. Shen and E. Elhamifar. Understanding multi-task activities from single-task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1", + "[79] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2", + "[80] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2", + "[81] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016.", + "[82] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. CoRR, abs/2105.10859, 2021. 2", + "[83] Dean Slawson. The case for vr-immersive and ai-adaptive soft skills training. Training Industry Magazine, 2018. 1", + "[84] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 1, 2", + "[85] Tomáš Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1", + "[86] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1", + "[87] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. IEEE Winter Conference on Applications of Computer Vision, 2023. 2" + ], + "bbox": [ + 99, + 92, + 482, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2", + "[89] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. CoRR, 2021. 1", + "[90] Xin Wang, Taein Kwon, Mahdi Radl Bowen Pan, Ishani Chakraborty, and Sean Andrist. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. IEEE International Conference on Computer Vision, 2023. 1, 2", + "[91] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2", + "[92] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. European Conference on Computer Vision, 2022. 2", + "[93] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. 2021. 6", + "[94] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[95] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 1, 2", + "[96] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2", + "[97] Guang Yu, Siqi Wang, Zhiping Cai, Xinwang Liu, Chuanfu Xu, and Chengkun Wu. Deep anomaly discovery from unlabeled videos via normality advantage and self-paced refinement. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[98] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. European Conference on Computer Vision, 2020.", + "[99] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[100] P. Zameni, Y. Shen, and E. Elhamifar. Moscato: Predicting multiple object state change through actions. International Conference on Computer Vision, 2025. 1" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10020", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[101] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[102] Ali Zare, Yulei Niu, Hammad Ayyubi, and Shih-Fu Chang. Rap: Retrieval-augmented planner for adaptive procedure planning in instructional videos. European Conference on Computer Vision, 2024. 1", + "[103] Junbin Zhang, Pei-Hsuan Tsai, and Meng-Hsun Tsai. Semantic2graph: Graph-based multi-modal feature fusion for action segmentation in videos, 2022. 1, 2", + "[104] Shrinivas Ramasubramanian Angela Yao Zhanzhong Pang, Fadime Sener. Long-tail temporal action segmentation with group-wise temporal logit adjustment. European Conference on Computer Vision, 2024. 2", + "[105] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. IEEE Conference on Computer Vision and Pattern Recognition, 2011. 2", + "[106] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 4", + "[107] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, , and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 2019. 2", + "[108] Yuansheng Zhu, Wentao Bao, , and Qi Yu. Towards open set video anomaly detection. European Conference on Computer Vision, 2022. 2", + "[109] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + ], + "bbox": [ + 91, + 90, + 480, + 628 + ], + "page_idx": 12 + }, + { + "type": "page_number", + "text": "10021", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 12 + } +] \ No newline at end of file diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_model.json b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bdb980ebb4db1c153074baadb81ecd5ea9857797 --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_model.json @@ -0,0 +1,2767 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.141, + 0.131, + 0.856, + 0.154 + ], + "angle": 0, + "content": "Error Recognition in Procedural Videos using Generalized Task Graph" + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.182, + 0.438, + 0.217 + ], + "angle": 0, + "content": "Shih-Po Lee Northeastern University" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.22, + 0.454, + 0.233 + ], + "angle": 0, + "content": "lee.shih@northeastern.edu" + }, + { + "type": "text", + "bbox": [ + 0.542, + 0.182, + 0.735, + 0.217 + ], + "angle": 0, + "content": "Ehsan Elhamifar Northeastern University" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.22, + 0.761, + 0.233 + ], + "angle": 0, + "content": "e.elhamifar@northeastern.edu" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.303, + 0.485, + 0.651 + ], + "angle": 0, + "content": "Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.678, + 0.222, + 0.693 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The integration of computer vision with augmented and virtual reality (AR & VR) is expected to significantly improve the way we learn and develop skills in our daily lives [83]. Imagine wearing a pair of AR glasses and trying to assemble a new piece of furniture. A visual task assistant should recognize your actions and provide guidance on what to do and how to perform them throughout the process. It should also recognize your possible errors and provide feedback to correct them, when needed. This has motivated exciting new research on action recognition [8, 10, 25, 26, 47, 50, 61, 89] and segmentation [23, 38, 45, 46, 54, 57, 58, 73, 77, 78, 84, 95, 96, 103], video grounding [5, 18, 20, 49, 59, 63, 86, 109], learning" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.27, + 0.904, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.503, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Figure 1. The illustration of our framework for joint temporal action segmentation and error recognition. It consists of 1) Generalized Task Graph (GTG)-Video Alignment for action segmentation and 2) an Error Recognition Module using features from LLM/VLM for error recognition." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.906, + 0.61 + ], + "angle": 0, + "content": "from instructional videos [12, 33, 62, 64, 65, 85, 102] and very recently on error understanding [14, 27, 29, 41, 60, 90], as well as progress [17] and state [100] prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.906, + 0.76 + ], + "angle": 0, + "content": "Successful operation of a visual task assistant, in particular, requires temporal action segmentation (assigning step labels to frames) and error recognition (recognizing the type of error, if any, in each frame) to seamlessly work together. Additionally, a task graph, which encodes plausible ways to perform a task, has been shown to be critical for temporal action segmentation to effectively mitigate oversegmentation [5, 19, 77, 106]. However, existing methods have three major limitations when it comes to understanding actions and errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "SOTA Limitations. First, existing works on error understanding can only detect the presence of an error in a video frame without recognizing the type of the error. As recently studied in [41], different types of errors, such as omission, addition, modification, and slip errors, can occur during task execution. However, [41] cannot distinguish between the error types. A task assistant requires understanding the error type to provide appropriate feedback. For example, an actor knocking over a dripper while making coffee requires" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "immediate feedback for correction or spreading jelly onto tortilla using spoon instead of knife needs a feedback after task completion for the actor to better execute the step." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Second, given the large number of possible errors, it is infeasible to gather real training videos and annotations for all possible errors. Also, recent datasets [41, 66] that contain videos with errors have imbalance annotations on different error types, making error recognition a challenging task, especially under the setting when we do not have error videos during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.243, + 0.482, + 0.422 + ], + "angle": 0, + "content": "Third, recent action segmentation methods [19, 77] that use task graphs assume that there are no errors in videos and assign step labels to the frames according to the task graph. In addition, two recent works that handle both action segmentation and error detection [29, 41] cannot use the task graph for action segmentation. This is because a task graph only contains the correct actions/steps and does not consider possible errors that may occur during the task execution. Thus, following the task graph leads to assigning every frame (including frames that have errors and ones that do not belong to any step) to correct actions/steps, which is undesired." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.482, + 0.47 + ], + "angle": 0, + "content": "Paper Contributions. We develop a unified error recognition and temporal action segmentation framework (see Figure 1) to address the aforementioned limitations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.482, + 0.519 + ], + "angle": 0, + "content": "- We propose a Generalized Task Graph (GTG) whose nodes encode normal actions/steps and background (task-irrelevant actions)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.482, + 0.581 + ], + "angle": 0, + "content": "- We develop a dynamic programming-based method, GTG-Video Alignment (GTG2Vid), with a dynamic frame and node dropping term to assign frames to GTG nodes for action segmentation and drop frames for error detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.482, + 0.675 + ], + "angle": 0, + "content": "- Given the lack of error training videos, we leverage the rich knowledge of large language models (LLMs) in conjunction with the rich representation power of video language models (VLMs) to obtain error descriptions for different types of errors and their corresponding features for error recognition." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.678, + 0.482, + 0.739 + ], + "angle": 0, + "content": "- We propose an Error Recognition Module (ERM) to recognize the error types, if any, in the frames predicted by GTG2Vid using joint similarities of video frame, normal type, and error type features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.741, + 0.482, + 0.83 + ], + "angle": 0, + "content": "- By extensive experiments on two egocentric datasets, EgoPER and CaptainCook4D, we show that without using error videos during training, our method 1) improves the performance of temporal action segmentation on error videos, 2) improves error detection, and 3) recognizes different errors." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.474, + 0.482, + 0.83 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.843, + 0.24, + 0.859 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Temporal Action Segmentation. Learning to segment videos into several actions/steps has been studied under un" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.318 + ], + "angle": 0, + "content": "supervised [4, 21, 22, 28, 36, 37, 74, 79, 88, 91, 109], weakly-supervised [11, 16, 24, 42-44, 48, 55, 55, 56, 69-71, 76, 84], and fully-supervised [3, 7, 9, 15, 23, 31, 32, 35, 38, 45, 46, 51, 57, 73, 80-82, 84, 95, 96, 103, 104] setting. Since procedural step videos follow certain task graphs to complete the task, some prior works [5, 19, 77, 106] additionally adopt task graph information to guide segmentation process. Specifically, G2Vid [19] aims to assign steps/nodes in a task graph to video frames and drop the frames not corresponding to any node. However, G2Vid cannot drop unnecessary nodes in a task graph nor map background frames to nodes to distinguish between background and error frames. In contrast, our framework adds extra nodes to a task graph and allows droppable nodes to properly assign nodes to frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.324, + 0.905, + 0.565 + ], + "angle": 0, + "content": "Error Detection and Recognition for Procedural Tasks. Detecting errors for procedural tasks has recently drawn attention from the video understanding community. Recent works [14, 29, 68, 90] have released relevant datasets to facilitate error understanding. Meanwhile, several works [27, 41, 60, 75] have developed frameworks to detect if errors are present in videos and localize their temporal locations. Recent work [27] focuses on online error detection using action recognition and anticipation models based on videos and action sequences. Another work [41] uses only error-free (normal) videos during training to generate multiple prototypes for action segmentation and error detection. However, none of them has the ability to recognize error types. Our work is able to segment videos and recognize different errors simultaneously without error training videos." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.571, + 0.905, + 0.813 + ], + "angle": 0, + "content": "VLMs/LLMs for Anomaly Detection. Anomaly detection in videos [2, 6, 13, 39, 40, 53, 67, 72, 87, 92, 97-99, 105, 107, 108] has been widely explored, and recent studies [34, 101] leverage VLMs and LLMs for further improvement due to their outstanding reasoning capabilities. Recent work in [101] proposes a training-free video anomaly detection framework using VLMs to generate captions over frames followed by LLMs to predict anomaly scores for those captions. Another work in [34] adopts LLMs to generate normal and abnormal descriptions under various scenes and classify frames into normal or abnormal according to the similarities between features of frames and descriptions. In our paper, we leverage LLMs to generate error descriptions for each error type of every step with crafted prompts, followed by VLMs to generate textual features aligned with visual features for error recognition." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.825, + 0.846 + ], + "angle": 0, + "content": "3. Graph to Video Alignment Review" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Our proposed GTG-Video Alignment method builds upon and generalizes the Graph-to-Video Alignment (G2Vid) [19], which we review here. A task graph (also called flow" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10010" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "graph) is a directed acyclic graph \\( G = (V, E) \\), where each node \\( v \\in V \\) represents a procedure step and each edge \\( (v_k, v_i) \\in E \\) from \\( v_k \\) to \\( v_i \\) implies that \\( v_k \\) must be completed before proceeding to \\( v_i \\), see Figure 2(a). Thus, all predecessors of a node must be completed before starting that step. Thus, the task graph encodes all possible ways to perform a task (this can be obtained by topological sorting)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.199, + 0.483, + 0.379 + ], + "angle": 0, + "content": "To efficiently align a video with the task graph and therefore segment the video into task steps, G2Vid [19] first builds a tSort graph (see Figure 2(a)) that provides a compact representation of all topological sorts of the task graph (each path from the source to the sink node in the tSort graph represents a topological sort of \\(G\\)). G2Vid alignment recovers a traversal of the tSort graph that best aligns with the video. It uses a modified version of Drop-DTW [18] to align nodes and frames, allowing for i) one-to-many matching (one node/step can be assigned to many frames) and ii) unmatched sequence elements (to be able to drop frames that do not correspond to any node/step)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.456 + ], + "angle": 0, + "content": "Let \\( C_{i,j} \\) denote the matching cost of node \\( i \\) to frame \\( j \\), \\( d_j \\) denote the dropping cost of frame \\( j \\), and \\( A(i) \\) denote the set of predecessors of \\( i \\). The dynamic programming solution computes the cumulative matching \\( D_{i,j}^{+} \\), dropping \\( D_{i,j}^{-} \\) and optimal \\( D_{i,j} \\) costs up to the \\( i \\)-th node and \\( j \\)-th frame as" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.465, + 0.483, + 0.488 + ], + "angle": 0, + "content": "\\[\nD _ {i, j} ^ {+} = C _ {i, j} + \\min \\left\\{\\min _ {k \\in A (i)} \\left\\{D _ {k, j - 1} \\right\\}, D _ {i, j - 1} \\right\\}, \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.488, + 0.424, + 0.501 + ], + "angle": 0, + "content": "\\[\nD _ {i, j} ^ {-} = d _ {j} + D _ {i, j - 1}, \\quad D _ {i, j} = \\min \\{D _ {i, j} ^ {+}, D _ {i, j} ^ {-} \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.483, + 0.571 + ], + "angle": 0, + "content": "By iterating over all nodes and frames, the dynamic programming outputs the optimal cost and alignment by backtracking the path with minimum cost, starting from the top left and ending at the bottom right of the cumulative matrix." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.584, + 0.265, + 0.601 + ], + "angle": 0, + "content": "4. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.609, + 0.483, + 0.82 + ], + "angle": 0, + "content": "We develop a unified framework for Temporal Action Segmentation (TAS) and Error Recognition (ER) in procedural task videos. We assume having access to only normal (error-free) videos with their frame-wise step labels during training and the task graph during testing. Our proposed framework addresses three fundamental questions: (1) How can we build an efficient Generalized Task Graph (GTG) that enables TAS and Error Detection (ED) for videos? (2) How can we obtain features for different errors without collecting real error videos? (3) How can we leverage the error features to perform ER? In the following, we first formalize the problem, discuss the overarching goal of our framework, and discuss the details of the method to address these questions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.824, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Problem Setting. For TAS and ER, given a sequence of video frame features \\(\\mathcal{X}^v = (x_1^v,x_2^v,\\ldots ,x_T^v)\\), our goal is to predict its frame-wise steps \\(\\hat{\\mathcal{Y}} = (\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_T)\\) and error types \\(\\hat{\\mathcal{E}} = (\\hat{e}_1,\\hat{e}_2,\\dots ,\\hat{e}_T)\\), where \\(T\\) is the number of frames, \\(x_{t}^{v}\\in \\mathbb{R}^{D}\\) is a pre-extracted feature vector, \\(\\hat{y}_t\\in\\)" + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.095, + 0.878, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.19, + 0.878, + 0.27 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.545, + 0.27, + 0.878, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.355, + 0.905, + 0.404 + ], + "angle": 0, + "content": "Figure 2. The illustration of (a) an original task graph and its corresponding tSort graph, (b) a naive approach to capture background, (c) our proposed solution, which specifically illustrates the predecessors and successor of extra node \\( S_{3} \\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.417, + 0.906, + 0.524 + ], + "angle": 0, + "content": "\\(\\{-1,0,\\ldots ,S\\}\\), where \\(S\\) is the number of steps and class \\(-1\\) and \\(0\\) indicate error and background (a task-irrelevant action, e.g., answering phone when making coffee), \\(\\hat{e}_t\\in \\{0,\\dots ,M\\}\\), where \\(M\\) is the number of error types and class \\(0\\) denotes no error happening. During training, we use only normal (error-free) videos and their ground-truth step labels \\(\\mathcal{V}\\). For testing, we assume that the task graph is given." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.907, + 0.71 + ], + "angle": 0, + "content": "**Framework Overview.** We introduce Generalized Task Graph for handling both \\( TAS \\) and \\( ED \\). Our key idea is to 1) insert extra nodes in the tSort graph to capture the background class that may occur before each step in the task graph, 2) perform \\( TAS \\) and \\( ED \\) by our GTG2Vid with a dynamic frame and node dropping term for different types of errors, 3) obtain error descriptions and features for all types of errors, respectively, by leveraging LLMs and VLMs and 4) use an Error Recognition Module (ERM) that computes joint similarities of video, normal type and error type features obtained by VLMs to classify the error types for the frames flagged by GTG2Vid." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.722, + 0.796, + 0.737 + ], + "angle": 0, + "content": "4.1. Generalized Task Graph (GTG)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.745, + 0.905, + 0.774 + ], + "angle": 0, + "content": "We build GTG to handle execution and omission errors and develop GTG2Vid for TAS and ED." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Handling Execution Errors. We aim to use the task graph to segment videos into steps and to detect execution errors (e.g., slip, correction, modification, and addition errors defined in [41]). Some of the errors are step-relevant (e.g., spreading jelly using spoon instead of knife) and some are not in the task graph (e.g., adding honey to a recipe that does not include honey). However, G2Vid drops both frames with errors as well as background frames (e.g., answering" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10011" + } + ], + [ + { + "type": "code_caption", + "bbox": [ + 0.093, + 0.091, + 0.416, + 0.107 + ], + "angle": 0, + "content": "Algorithm 1 GTG-Video Alignment (GTG2Vid)" + }, + { + "type": "algorithm", + "bbox": [ + 0.094, + 0.109, + 0.907, + 0.376 + ], + "angle": 0, + "content": "1: Inputs: \\(C\\in \\mathbb{R}^{2S\\times T}\\) \\(A(\\cdot),d_i^z,d_j^x\\) (matching cost matrix, predecessor mapping dictionary, node dropping costs, frame dropping costs.) \\(\\triangleright\\) Initialize DP tables \n2: \\(D_{0,0}^{+} = 0;D_{i,0}^{+} = \\infty ;D_{0,j}^{+} = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}\\) Matching table of \\(\\mathcal{X}_n^v\\) (frames) and \\(\\mathcal{Z}\\) (nodes) \n3: \\(D_{0,0}^{z - } = 0;D_{i,0}^{z - } = \\sum_{k = 1}^{i}d_{k}^{z};D_{0,j}^{z - } = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}\\) Dropping table of \\(\\mathcal{Z}\\) (nodes) \n4: \\(D_{0,0}^{-x} = 0;D_{i,0}^{-x} = \\infty ;D_{0,j}^{-x} = \\sum_{k = 1}^{j}d_{k}^{x};\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}\\) Dropping table of \\(\\mathcal{X}_n^v\\) (frames) \n5: \\(D_{0,0} = 0;D_{i,0} = D_{i,0}^{-x};D_{0,j} = D_{0,j}^{z - }; \\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}\\) Optimal solution table \n6: for \\(i = 1,\\dots ,2S\\) do Iterate nodes in Z \n7: for \\(j = 1,\\dots ,T\\) do Iterate frames in \\(\\mathcal{X}_n^v\\) \n8: \\(D_{i,j}^{+} = -C_{i,j} + \\min \\{\\min_{k\\in A(i)}\\{D_{k,j - 1}\\},D_{i,j - 1}\\}\\) Consider matching node \\(\\mathcal{Z}_i\\) to frame \\(x_{n,j}^v\\) \n9: \\(D_{i,j}^{-x} = d_j^x +D_{i,j - 1}\\) Consider dropping frame \\(x_{n,j}^v\\) \n10: \\(D_{i,j}^{z - } = d_i^z +\\min_{k\\in A(i)}\\{D_{k,j}\\}\\) Consider dropping node \\(\\mathcal{Z}_i\\) \n11: \\(D_{i,j} = \\min \\{D_{i,j}^{+},D_{i,j}^{-x},D_{i,j}^{z - }\\}\\) Select the optimal path \n12: end for \n13: end for \n14: \\(\\hat{y}^* =\\) traceback(D) \\(\\triangleright\\) Compute the optimal alignment by tracing back the path with minimum cost \n15: \\(\\hat{\\gamma} =\\) frame-assign(y*) Flag dropping frames as error frames and frames belonging to extra nodes as background \n16: Outputs: \\(D_{2S,T},\\hat{\\gamma}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.402, + 0.482, + 0.432 + ], + "angle": 0, + "content": "the phone while making a recipe) and thus cannot distinguish between them and understand when an error occurs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.552 + ], + "angle": 0, + "content": "To handle this, we build GTG by inserting extra nodes into the task graph to capture background frames. Therefore, by aligning a video with GTG, we can understand which frames correspond to which steps and which frames correspond to background. Additionally, by allowing dropping frames, we can detect execution errors, and by allowing dropping nodes, we can detect omission errors which we will discuss next." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.75 + ], + "angle": 0, + "content": "A naive approach is to insert an extra node with a unique label between every two nodes (see Figure 2 (b)). Therefore, the task graph can potentially map frames to extra nodes. However, the complexity of the tSort graph grows exponentially. To solve this problem, our proposed approach inserts an extra node \\( S_{i} \\) with a unique ID between every node \\( i \\) and its predecessor in the tSort graph. For example, in Figure 2 (c), between node 2 and 3 as well as node 1 and 3, we insert the same extra node \\( S_{3} \\) to capture the background frames with respect to step 3. Our design 1) prevents exponential growth of the tSort graph, only increasing the number of nodes from \\( S \\) to \\( 2S \\), and 2) inherits the correct execution orders in the original tSort graph." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.483, + 0.874 + ], + "angle": 0, + "content": "Handling Omission Errors. During execution, an actor may skip one or more steps in the task graph, resulting in omission error. In this case, forcing models to follow the task graph as done in prior works [5, 19, 77, 106] can lead to wrong TAS results where some frames are assigned to the skipped steps. To handle omission errors, we introduce a node dropping term to our GTG2Vid to make nodes of our GTG droppable," + }, + { + "type": "equation", + "bbox": [ + 0.185, + 0.885, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\nD _ {i, j} ^ {z -} = d _ {i} ^ {z} + \\min _ {k \\in A (i)} \\left\\{D _ {k, j} \\right\\} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.507 + ], + "angle": 0, + "content": "where \\( D_{i,j}^{z-} \\) is the accumulative cost until node \\( i \\) and frame \\( j \\) in the dynamic programming (DP) table of dropping node and \\( d_i^z \\) is the node dropping cost of node \\( i \\). A low dropping cost indicates that the corresponding GTG node is not in the video and should be omitted. This also allows dropping extra nodes, whenever there is no corresponding frame in the video." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.511, + 0.909, + 0.661 + ], + "angle": 0, + "content": "GTG-Video Alignment (GTG2Vid). Algorithm 1 shows the steps of GTG2Vid. \\( C_{i,j} \\) is the matching cost of the \\( i \\)-th node to the \\( j \\)-th frame. We let \\( C_{i,j} \\) to be the predicted probability of step \\( i \\) for \\( j \\)-th frame using an action segmentation model. We train the action segmentation model with smoothing and cross-entropy loss in [23] using framewise labels. In addition, to ensure that our proposed method can drop frames and nodes based on the confidence of predictions, we propose a dynamic dropping cost \\( d^{x} \\) for frames and \\( d^{z} \\) for nodes of GTG as" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.669, + 0.905, + 0.713 + ], + "angle": 0, + "content": "\\[\nd _ {j} ^ {x} = - \\left(\\frac {\\sum_ {i = 1} ^ {2 S} C _ {i , j} \\log C _ {i , j}}{\\log \\frac {1}{2 S}} + \\eta\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.714, + 0.794, + 0.746 + ], + "angle": 0, + "content": "\\[\nd _ {i} ^ {z} = - \\tau \\left(1 - \\max _ {j = 1, \\dots , T} C _ {i, j}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\eta\\) and \\(\\tau\\) are hyper-parameters (larger \\(\\eta\\) and \\(\\tau\\) lead to dropping more frames and nodes). Our frame dropping cost \\(d^{x}\\) leads to drop frames with unconfident predictions, indicating that the frames do not correspond to any GTG node and may contain errors. Meanwhile, our node dropping cost \\(d^{z}\\) allows dropping the nodes whose highest step probabilities over all frames are low. After obtaining the minimum cost path, we generate step predictions \\(\\hat{\\mathcal{V}}\\) from \\(\\hat{\\mathcal{V}}^*\\) by flagging the dropped frames as error frames and frames belonging to extra nodes as background. The complexity of our al" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10012" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.201 + ], + "angle": 0, + "content": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTea
w-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAcc
Naive Predictor11.225.013.025.010.425.018.133.311.425.0
EgoPED (MV)7.575.05.575.07.475.010.9100.04.475.0
GTG2Vid (MV)5.150.02.750.04.450.00.633.34.050.0
GTG2Vid (w/o n.f.)16.375.016.475.029.9100.01.733.315.275.0
GTG2Vid31.7100.031.3100.017.8100.04.5100.022.1100.0
" + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.203, + 0.885, + 0.31 + ], + "angle": 0, + "content": "
Hot Chocolate \nw-F1@0EAccCaptainCook4D
Sandwich \nw-F1@0EAccBurritos \nw-F1@0EAccRamen \nw-F1@0EAccRaita \nw-F1@0EAcc
Naive Predictor7.120.05.120.07.425.07.425.07.333.3
EgoPED (MV)7.380.03.480.07.175.03.375.04.366.6
GTG2Vid (MV)4.460.04.740.06.750.02.725.011.466.6
GTG2Vid (w/o n.f.)2.840.06.260.06.125.04.650.011.266.6
GTG2Vid10.5100.08.5100.09.475.04.475.023.0100.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.354, + 0.312, + 0.644, + 0.325 + ], + "angle": 0, + "content": "Table 1. \\( {ER} \\) results on EgoPER and CaptainCook4D." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.343, + 0.484, + 0.389 + ], + "angle": 0, + "content": "girthm is \\(\\mathcal{O}(SQT)\\), same as G2Vid [19], where \\(S, Q\\), and \\(T\\) denote the number of GTG nodes, separate and linearly-ordered threads of GTG, and frames, respectively." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.4, + 0.445, + 0.416 + ], + "angle": 0, + "content": "4.2. Obtaining Error Features via LLM/VLM" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.485, + 0.742 + ], + "angle": 0, + "content": "Our next step is to obtain error features for \\(ER\\). Since we are only given normal videos during training, we i) leverage LLMs to obtain error descriptions for each error type for every step, and ii) use VLMs to generate the textual features that are aligned with visual features. To begin, given a step \\(i\\) and error type \\(m\\), we generate its error type description using LLMs. The prompt for LLMs combines a description for step \\(i\\), error definition for error type \\(m\\), and the number of different error type descriptions to produce. Notice that we directly set normal type description \\((m = 0)\\) as the description of step \\(i\\). Next, we generate the feature vector \\(x_{i,m}^{e}\\) for the error description using the text encoder in VLMs, where \\(x^{e} \\in \\mathbb{R}^{S \\times M \\times D}\\) is a tensor containing features of \\(M\\) error type descriptions for \\(S\\) steps. When LLMs are prompted to produce multiple error descriptions for an error type, we average all features to get a single vector. Meanwhile, we use the video encoder in VLMs to generate frame features for videos. In this way, we can ensure that the features of videos and error descriptions are aligned in the VLM feature space. See the supplementary materials for detailed prompts and generated error descriptions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.753, + 0.391, + 0.769 + ], + "angle": 0, + "content": "4.3. Error Recognition Module (ERM)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.775, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Our ERM classifies error types for error frames according to their joint similarities of frame, normal type, and error type features from Section 4.2 and obtain \\(ER\\) predictions \\(\\hat{\\mathcal{E}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Predicting Steps for Error Frames. First, we obtain TAS predictions \\(\\hat{\\mathcal{V}}\\) by inferring GTG2Vid on a video. Next, we obtain another TAS predictions \\(\\hat{\\mathcal{V}}^{+} = (\\hat{y}_{1}^{+},\\hat{y}_{2}^{+},\\dots ,\\hat{y}_{T}^{+})\\) where \\(\\hat{y}_t^+\\in \\{0,1,\\ldots ,S\\}\\), by inferring GTG2Vid without using the frame and node dropping term, forcing GTG2Vid" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.905, + 0.389 + ], + "angle": 0, + "content": "to map all frames to nodes. It aims to obtain the step predictions for the error frames in \\(\\hat{\\mathcal{V}}\\) to find the error type features of the corresponding steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Computing Joint Similarities for Error Recognition. Next, we compute framewise joint similarities \\( p \\) for \\( ER \\), where \\( p_{m,j} \\in [0,1] \\) and \\( p \\in \\mathbb{R}^{M \\times T} \\). For each frame \\( j \\), given step prediction \\( \\hat{y}_j, \\hat{y}_j^+ \\), and feature \\( x_j^v, x^e \\), we find the error type \\( \\hat{m}_j \\) and compute its joint similarity \\( p_{\\hat{m}_j,j} \\) as" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.485, + 0.907, + 0.534 + ], + "angle": 0, + "content": "\\[\n\\hat {m} _ {j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ \\underset {m \\in \\{1, \\dots , M \\}} {\\operatorname {a r g m a x}} s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, m\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.541, + 0.906, + 0.58 + ], + "angle": 0, + "content": "\\[\np _ {\\hat {m} _ {j}, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, \\hat {m} _ {j}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.614 + ], + "angle": 0, + "content": "where \\(s(\\cdot ,\\cdot ,\\cdot ,\\cdot)\\) is our score function that outputs the joint similarity of the frame. Our score function is formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.635, + 0.905, + 0.656 + ], + "angle": 0, + "content": "\\[\ns \\left(x _ {j} ^ {v}, x ^ {e}, y, m\\right) = \\operatorname {s i g m o i d} \\left(\\left(x _ {y, 0} ^ {e} \\odot x _ {y, m} ^ {e}\\right) ^ {T} x _ {j} ^ {v}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(\\odot\\) is element-wise multiplication and \\(x_{y,0}^{e}\\) is the feature of step description (normal type) for step \\(y\\). The intuition of using normal type feature is that error frames often contain multiple sub-steps, including both normal and error ones (e.g., an actor opens a tortilla bag, grabs a tortilla from it, and accidentally drops it). A single error type feature may be unable to effectively match the corresponding error frames, especially when the number of frames is large and contain multiple sub-steps which the error description does not describe. Therefore, our score function computes the correlation between \\(x_{y,0}^{e}, x_{y,m}^{e}\\), and \\(x_{j}^{v}\\). Thus, it jointly considers the similarities of error frames with respect to both normal and error type features. We validate its effectiveness for \\(ER\\) in Section 5. Finally, we apply a maximum filter on each error type of \\(p\\) to smoothen the predictions, followed by argmax to obtain the error type predictions \\(\\hat{\\mathcal{E}}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10013" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.11, + 0.089, + 0.887, + 0.26 + ], + "angle": 0, + "content": "
MethodQuesadillaOatmealEgoPER
F1@.5AccEditF1@.5AccEditPinwheelCoffeeTeaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid57.053.760.068.874.576.849.152.753.759.157.175.061.156.169.459.058.870.0
EgoPED56.961.447.362.154.870.349.854.741.053.959.655.469.464.558.356.459.054.5
GTG2Vid68.164.569.483.378.588.661.358.460.963.962.573.771.262.676.969.665.373.9
MethodHot ChocolateSandwichBurritosRamenRaitaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid17.119.223.821.321.739.833.427.542.928.033.950.518.523.126.123.725.136.6
EgoPED5.021.713.313.122.518.221.131.821.815.737.325.422.631.323.015.528.920.3
GTG2Vid20.843.527.827.238.538.836.634.041.331.333.642.922.130.826.127.636.135.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.352, + 0.263, + 0.644, + 0.276 + ], + "angle": 0, + "content": "Table 2. TAS results on EgoPER and CaptainCook4D." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.293, + 0.224, + 0.31 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.32, + 0.281, + 0.336 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.349, + 0.485, + 0.683 + ], + "angle": 0, + "content": "Dataset. We evaluate our proposed method on EgoPER [41] and CaptainCook4D [66]. EgoPER consists of 5 tasks with 386 egocentric videos and contains 5 types of errors, including omission, addition, modification, slip, and correction. Captaincook4D has 24 tasks with 94.5 hours and contains 8 types of errors, including preparation, measurement, technique, timing, temperature, missing steps, ordering, and other. For EgoPER, we evaluate our method on all 5 tasks and error types with the same training and testing split in [41]. For CaptainCook4D, to have enough videos for training and testing, we select 5 of the tasks which have at least 5 normal and error videos: Spiced Hot Chocolate (Hot Chocolate), Microwave Egg Sandwich (Sandwich), Breakfast Burritos (Burritos), Ramen, and Cucumber Raita (Raita). We assign all normal videos to the training set and error videos to test set. To fit our setting for recognizing execution errors, we report the following error types in CaptainCook4D: preparation (Prep.), measurement (Mea.), technique (Tec.), timing (Time), and temperature (Temp). See the supplementary materials for breakdown analysis of EgoPER and CaptainCook4D and their detailed training and testing splits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We report results on \\(ER\\), \\(TAS\\), and \\(ED\\). In our experiments, we report the F1 score \\((F1@\\beta)\\) with an overlapping threshold \\(\\beta\\). We compute the F1 score by averaging (if not specified) over the F1 scores that are independently computed on the segments for each step or error type. For \\(ER\\), we report Error Accuracy (EAcc) and weighted \\(F1@0\\) score \\((w-F1@0)\\). As \\(ER\\) is challenging under the setting of no error video during training, we set \\(\\beta\\) to 0, meaning that a segment is true positive whenever any frame in the segment is correctly classified. EAcc is the percentage of error types that a model can recognize; that is, the error types with \\(F1@0\\) score larger than 0. For \\(w-F1@0\\), we first compute \\(F1@0\\) on the error type segments, excluding normal ones \\((m=0)\\), to alleviate the effect of data imbalance (e.g.," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.908, + 0.462 + ], + "angle": 0, + "content": "nearly \\(90\\%\\) of the segments in EgoPER are normal). We then compute w-F1@0, where w-F1@0 = \\(\\frac{\\mathrm{F1}@\\mathrm{0}}{\\mathrm{M} - \\mathrm{EAcc} \\times \\mathrm{M} + 1}\\). w-F1@0 penalizes the model that fails to recognize all error types. For TAS, we report frame-wise Accuracy (Acc), segment-wise Edit score (Edit) on normal steps, and F1@.5 score. For ED, we report F1@.5 score on all segments. We also report F1@.5 score on normal (N.) and error segments (E.) separately for alleviating data imbalance. For omission error detection, we use Omission Accuracy (O-Acc), predicting the accuracy for omission steps, and Omission Intersection over Union (O-IoU), as in [41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.469, + 0.909, + 0.773 + ], + "angle": 0, + "content": "Baselines. Given the lack of baselines for joint TAS, \\(ER\\), and \\(ED\\), we report the performance of EgoPED (MV), a modified version of EgoPED [41] that can additionally perform \\(ER\\). EgoPED (MV) computes the cosine similarities between frames and error types using the error type features generated in Section 4.2. Then it assigns frames to the type with the highest similarity and uses the majority voting (MV) strategy in [41] over frames within a segment to decide the final error type for each segment. EgoPED (MV) detects error segments with step thresholds equal to mean similarities over videos in the validation set. For \\(ER\\) and \\(ED\\), since both datasets have data imbalance, we report the performance of Naive Predictor as another baseline. Naive Predictor classifies frames into the same error type for \\(ER\\) and into normal frames for \\(ED\\). Specifically, it classifies the frames into modification errors for EgoPER and preparation errors for CaptainCook4D. For \\(TAS\\), we compare with G2Vid [19] to show how G2Vid performs with standard task graphs. To run G2Vid with our cost matrix, we remove all the background entries." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Implementation Details. We employ GPT-4o mini [1] as our LLM for generating error descriptions and VideoCLIP [93] pre-trained on Ego4D [30] as our VLM for extracting both visual and textual features for videos and error type descriptions. For each task in both EgoPER and CaptainCook4D, we generate at least 2 descriptions for each error type of each action. We set the step description as the description of normal type. We use the backbone in [52] as" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "10014" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.118, + 0.089, + 0.878, + 0.263 + ], + "angle": 0, + "content": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTeaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor14.40.07.226.10.013.123.40.011.744.10.022.14.80.02.422.60.011.3
EgoPED30.125.227.729.312.921.114.624.519.69.58.99.236.334.035.124.021.122.5
GTG2Vid40.422.631.558.536.447.430.616.023.333.14.018.651.537.044.242.823.233.0
CaptainCook4D
Hot ChocolateSandwichBurritosRamenRaitaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor8.20.04.16.50.03.23.80.01.94.10.02.03.10.01.65.10.02.6
EgoPED8.12.45.216.36.811.516.38.212.212.410.711.614.19.511.813.47.510.5
GTG2Vid15.820.918.316.712.814.816.919.718.322.68.215.418.29.914.018.014.316.2
" + }, + { + "type": "table_caption", + "bbox": [ + 0.128, + 0.265, + 0.868, + 0.278 + ], + "angle": 0, + "content": "Table 3. ED results of different methods on EgoPER and CaptainCook4D. N. and E. denote the F1@.5 score for normal and error segments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.484, + 0.432 + ], + "angle": 0, + "content": "our action segmentation model. We train the model using Adam optimizer with learning rate 0.0001 for 3000 iterations. For \\(ER\\), if \\(\\hat{y}_j = -1\\) and \\(\\hat{y}_j^+\\) is a background class, we directly assign an addition error and preparation error to \\(\\hat{m}_j\\) for EgoPER and CaptainCook4D, respectively, and assign the corresponding similarity to 1. For \\(ED\\), we produce \\(ED\\) predictions based on \\(\\hat{\\mathcal{E}}\\) by classifying different error types into a single error type. In practice, we set \\(\\tau = S\\) where \\(S\\) is the number of nodes in GTG and \\(\\eta = 0\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.293, + 0.454 + ], + "angle": 0, + "content": "5.2. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.719 + ], + "angle": 0, + "content": "Error Recognition \\((ER)\\). Table 1 compares \\(ER\\) performance between different methods on EgoPER and CaptainCook4D. Our proposed framework, GTG2Vid, demonstrates the best recognition ability than Naive Predictor and EgoPED (MV). Specifically, our proposed method achieves \\(31.7\\%\\), \\(31.3\\%\\), \\(17.8\\%\\), and \\(22.1\\%\\) on w-F1@0 compared to \\(7.5\\%\\), \\(5.5\\%\\), \\(7.4\\%\\), and \\(4.4\\%\\) by EgoPED (MV) for quesadilla, oatmeal, pinwheel, and tea. On the other hand, our proposed method outperforms EgoPED (MV) on w-F1@0 for all tasks of CaptainCook4D. In addition, our proposed method achieves overall higher EAcc than EgoPED (MV), indicating that our method can recognize more error types on both datasets. Notice that our performance on w-F1@0 for coffee is worse than EgoPED (MV) because EgoPED predicts most of the segments as errors, shown in Table 3. Please see supplementary materials for the F1@0 score of each type of error." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Next, we investigate the effectiveness of our ERM and our proposed score function (Eq. 6). By replacing ERM with MV, the performance drops significantly (more than \\(10\\%\\) on w-F1@0 across all tasks of EgoPER). On the other hand, our method, GTG2Vid (w/o n.f.), without using normal type features suffers a drop on EAcc for most of the tasks, indicating that our proposed function for computing joint similarity has a better ability to recognize different error types. Furthermore, Table 4 shows the Top-1 and Top-2 accuracy of different scoring methods given the features of ground-truth error segments (average over frame features) and generated error type features. Our method gains a per" + }, + { + "type": "table", + "bbox": [ + 0.52, + 0.292, + 0.898, + 0.369 + ], + "angle": 0, + "content": "
MethodQuesadillaPinwheelTea
Top-1Top-2Top-1Top-2Top-1Top-2
Uniform25.050.025.050.025.050.0
Cos. Sim.42.665.827.748.427.845.3
ERM54.681.538.460.334.353.7
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.371, + 0.905, + 0.397 + ], + "angle": 0, + "content": "Table 4. Performance of feature matching methods on EgoPER. Top-1 and Top-2 denote the top-1 and top-2 accuracy over all error segments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.416, + 0.905, + 0.476 + ], + "angle": 0, + "content": "formance boost on Top-1 and Top-2 accuracy. Specifically, we obtain \\(10.7\\%\\) improvement on Top-1 accuracy compared to cosine similarity (Cos. Sim.) for pinwheel, which verifies our claim in Section 4.3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.678 + ], + "angle": 0, + "content": "Temporal Action Segmentation (TAS). For TAS performance shown in Table 2, our method obtains a \\(13.2\\%\\), \\(6.3\\%\\), and \\(19.4\\%\\) higher F1@.5, Acc and Edit score than EgoPED on All of EgoPER. For CaptainCook4D, our method obtains a \\(12.1\\%\\), \\(7.2\\%\\), and \\(15.1\\%\\) higher F1, Acc, and Edit score than EgoPED on All. Furthermore, our method has more robust TAS performance compared to EgoPED, which suffers a significant performance drop in Hot Chocolate and Sandwich. Notice that our Edit score on All is comparable to G2Vid since there are not many omission errors in test videos. Still, our method achieves a higher F1@.5 score than G2Vid, which cannot distinguish between background frames and errors." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.684, + 0.906, + 0.866 + ], + "angle": 0, + "content": "Error Detection (ED). Table 3 shows the \\(ED\\) performance of different methods on EgoPER. Our method achieves \\(33.0\\%\\) higher on F1@.5 score, compared to \\(22.5\\%\\) by EgoPED for All in EgoPER. Although EgoPED can obtain a higher F1 score on error segments for quesadilla, pinwheel, and coffee, its detection ability for normal segments decreases significantly, specifically a \\(23.6\\%\\) drop on F1@.5 score for coffee compared to our method. Therefore, compared to EgoPED, our method achieves a better balance between normal and error segments. For All in CaptainCook4D, GTG2Vid outperforms EgoPED, achieving \\(16.2\\%\\) on F1@.5 compared to \\(10.5\\%\\) by EgoPED." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Ablation of Node and Frame Dropping Term. We show ablation studies and investigate the effectiveness of our pro" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10015" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.14, + 0.095, + 0.865, + 0.257 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.259, + 0.905, + 0.285 + ], + "angle": 0, + "content": "Figure 3. Qualitative results for quesadilla, tea, and oatmeal in EgoPER. The top part shows the results of TAS for ground-truths, EgoPED, and our method. The bottom part shows ER results for ground-truths, EgoPED, and our method. Seg. and E.T. denote action segmentation and error type, respectively." + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.297, + 0.478, + 0.436 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.439, + 0.466, + 0.452 + ], + "angle": 0, + "content": "Figure 4. Ablation of our proposed node and frame dropping term." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.469, + 0.483, + 0.665 + ], + "angle": 0, + "content": "posed node and frame dropping term in GTG2Vid. The two upper plots in Figure 4 show F1@.5 scores of our proposed frame dropping term (Ours) with different \\(\\eta\\) and a naive dropping cost (G2Vid) [19] with different \\(k\\) top percentile of the values in the cost matrix, for pinwheel and quesadilla. Our method achieves \\(26.0\\%\\) and \\(33.8\\%\\) on F1@.5 score compared to \\(23.2\\%\\) and \\(30.9\\%\\) by G2Vid. On the other hand, the two bottom plots in Figure 4 show the F1 scores for TAS with different \\(\\tau\\) for our node dropping term. Smaller \\(\\tau\\) indicates GTG2Vid drops a fewer number of nodes. When \\(\\tau\\) is larger or equal to 0, GTG2Vid starts dropping nodes, improving TAS performance on videos with omission steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.666, + 0.483, + 0.877 + ], + "angle": 0, + "content": "In addition, Figure 5 shows the performance of omission detection for different methods on EgoPER. First, since G2Vid is unable to drop nodes of a graph, it cannot predict any omitted steps, resulting in \\(0\\%\\) on O-Acc and O-IoU. Second, although our method suffers a performance drop on O-IoU (\\(5.7\\%\\) lower than EgoPED on All), we gain a performance boost on O-Acc (\\(10\\%\\) higher than EgoPED on All), indicating that our method successfully drops nodes of a task graph according to the input videos and has a stronger capability for detecting omitted steps. To conclude, our proposed method uses 1) dynamic frame dropping term to precisely drop error segments and 2) dynamic node dropping term to drop unnecessary nodes of task graphs for better TAS and omission detection performance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Qualitative Analysis. We visualize ER results in Fig. 3." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.299, + 0.903, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.472, + 0.899, + 0.486 + ], + "angle": 0, + "content": "Figure 5. Omission detection results of different methods on EgoPER." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.907, + 0.614 + ], + "angle": 0, + "content": "Our method can recognize different types of errors, e.g., addition errors (red) on tea, addition and correction (orange) errors on quesadilla, and modification errors (blue) on oatmeal. Our method obtains more accurate TAS, ED, and ER results for error videos on quesadilla, tea and oatmeal compared to EgoPED, thanks to dynamic frame and node dropping term in GTG2Vid and our proposed ERM. See supplementary materials for more qualitative visualization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.627, + 0.64, + 0.642 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.651, + 0.907, + 0.818 + ], + "angle": 0, + "content": "We studied joint \\(ER\\) and \\(TAS\\) in procedural task videos. Our proposed framework consists of 1) generalized task graphs to capture correct steps and background, 2) a GTG-Video Alignment algorithm to find the optimal alignment between nodes in GTG and video frames with dynamic dropping terms, 3) a procedure for generating error descriptions for each error type of every step using LLMs and the corresponding features using VLMs, and 4) an Error Recognition Module to effectively perform \\(ER\\) using joint similarities. Our experiments on two datasets showed that our proposed method can obtain promising results on \\(ER\\), \\(ED\\), and \\(TAS\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.831, + 0.673, + 0.848 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This work was funded, in part, by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110), ONR (N000142512287) and ARO" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10016" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.152 + ], + "angle": 0, + "content": "(W911NF2110276). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the US Government." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.164, + 0.188, + 0.18 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.108, + 0.189, + 0.484, + 0.258 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.26, + 0.484, + 0.342 + ], + "angle": 0, + "content": "[2] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, , and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.343, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[3] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.4, + 0.483, + 0.453 + ], + "angle": 0, + "content": "[4] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.455, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Neural Information Processing Systems, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.511, + 0.483, + 0.551 + ], + "angle": 0, + "content": "[6] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. arXiv: 2110.09742, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.553, + 0.483, + 0.607 + ], + "angle": 0, + "content": "[7] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.609, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.109, + 0.665, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[9] Daniele Di Mauro Mario Valerio Giuffrida Giovanni Maria Farinella Camillo Quattrocchi, Antonino Furnari. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.748, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.483, + 0.872 + ], + "angle": 0, + "content": "[11] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.874, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning" + }, + { + "type": "list", + "bbox": [ + 0.102, + 0.189, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.554, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "in instructional videos. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.136, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Hanqiu Deng, Zhaoxiang Zhang, Shihao Zou, and Xingyu Li. Bi-directional frame interpolation for unsupervised video anomaly detection. IEEE Winter Conference on Applications of Computer Vision, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.193, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv: 2307.16453, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.235, + 0.905, + 0.29 + ], + "angle": 0, + "content": "[15] Guodong Ding, Hans Golong, and Angela Yao. Coherent temporal synthesis for incremental action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.293, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[17] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[18] Nikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan D Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.462, + 0.905, + 0.532 + ], + "angle": 0, + "content": "[19] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 1, 2, 3, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.534, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[20] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos Derpanis, Animesh Garg, Richard Wildes, and Allan Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.604, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[21] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.647, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.689, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[23] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.76, + 0.905, + 0.814 + ], + "angle": 0, + "content": "[24] Mohsen Fayyaz and Jurgen Gall. Sect: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.817, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[25] C. Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10017" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[27] Alessandro Flaborea, Guido Melendugno, Leonardo Pliniq, Luca Scofanoq, Edoardo Matteisq, Antonino Furnari, Giovanni Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.482, + 0.245 + ], + "angle": 0, + "content": "[28] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.247, + 0.483, + 0.316 + ], + "angle": 0, + "content": "[29] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.318, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.706, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[31] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[32] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[33] Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. European Conference on Computer Vision, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[34] Jaehyun Kim, Seongwook Yoon, Taehyeon Choi, and Sanghoon Sull. Unsupervised video anomaly detection based on similarity with predefined text descriptions. Sensors, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[35] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition. IEEE Winter Conference on Applications of Computer Vision, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[36] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.264, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[37] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.349, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[38] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.406, + 0.905, + 0.472 + ], + "angle": 0, + "content": "[39] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.476, + 0.904, + 0.531 + ], + "angle": 0, + "content": "[40] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. Bman: Bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Transactions on Image Processing, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.533, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[41] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.59, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[42] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.633, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[43] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.676, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[44] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.718, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[45] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[46] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[47] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10018" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.133, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "for classification and detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[49] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhan, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. European Conference on Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.25, + 0.483, + 0.303 + ], + "angle": 0, + "content": "[50] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[51] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.35, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[52] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.393, + 0.482, + 0.461 + ], + "angle": 0, + "content": "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[54] Yang Liu, Jiayu Huo, Jingjing Peng, Rachel Sparks, Prokar Dasgupta, Alejandro Granados, and Sebastien Ourselin. Skit: a fast key information video transformer for online surgical phase recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21074-21084, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[55] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[56] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.662, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[57] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[58] Z. Lu and E. Elhamifar. Multi-modal few-shot temporal action segmentation. International Conference on Computer Vision, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[59] Z. Lu, A. Iftekhar, G. Mittal, T. Meng, X. Wang, C. Zhao, R. Kukkala, E. Elhamifar, and M. Chen. Decafnet: Delegate and conquer for efficient temporal grounding in long videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[60] Antonino Furnari Luigi Seminara, Giovanni Maria Farinella. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. Neural Information Processing Systems, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[61] K. Mangalam, H. Fan, Y. Li, C. Wu, B. Xiong, C. Feichtenhofer, and J. Malik. Reversible vision transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.151, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[62] A. Miech, J-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.208, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[63] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.251, + 0.904, + 0.291 + ], + "angle": 0, + "content": "[64] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.294, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[65] Kumaranage Ravindu Yasas Nagasinghe, Honglu Zhou, Malitha Gunawardhana, Martin Renqiang Min, Daniel Harari, and Muhammad Haris Khan. Why not use your textbook? knowledge-enhanced procedure planning of instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.378, + 0.905, + 0.459 + ], + "angle": 0, + "content": "[66] Rohith Peddi, Shivvrat Arya, Bharath Challa, Likhitha Pallapothula, Akshay Vyas, Bhavya Gouripeddi, Jikai Wang, Qifan Zhang, Vasundhara Komaragiri, Eric Ragan, Nicholas Ruozzi, Yu Xiang, and Vibhav Gogate. CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.463, + 0.904, + 0.517 + ], + "angle": 0, + "content": "[67] Didik Purwanto, Yie-Tarng Chen, and Wen-Hsien Fang. Dance with self-attention: A new look of conditional random fields on anomaly detection in videos. IEEE International Conference on Computer Vision, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.52, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[68] Yicheng Qian, Weixin Luo, Dongze Lian, Xu Tang, Peilin Zhao, and Shenghua Gao. Svip: Sequence verification for procedures in videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.577, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[69] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.648, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[70] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.704, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[71] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.762, + 0.905, + 0.841 + ], + "angle": 0, + "content": "[72] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[73] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.524, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10019" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[74] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[75] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepop, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.235, + 0.483, + 0.288 + ], + "angle": 0, + "content": "[76] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.291, + 0.483, + 0.346 + ], + "angle": 0, + "content": "[77] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.349, + 0.482, + 0.39 + ], + "angle": 0, + "content": "[78] Y. Shen and E. Elhamifar. Understanding multi-task activities from single-task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.392, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[79] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.448, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[80] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[81] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.562, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[82] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. CoRR, abs/2105.10859, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.605, + 0.482, + 0.633 + ], + "angle": 0, + "content": "[83] Dean Slawson. The case for vr-immersive and ai-adaptive soft skills training. Training Industry Magazine, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.635, + 0.482, + 0.688 + ], + "angle": 0, + "content": "[84] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[85] Tomáš Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.762, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[86] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[87] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. IEEE Winter Conference on Applications of Computer Vision, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[88] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.179, + 0.906, + 0.218 + ], + "angle": 0, + "content": "[89] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. CoRR, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.221, + 0.906, + 0.289 + ], + "angle": 0, + "content": "[90] Xin Wang, Taein Kwon, Mahdi Radl Bowen Pan, Ishani Chakraborty, and Sean Andrist. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. IEEE International Conference on Computer Vision, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.292, + 0.906, + 0.374 + ], + "angle": 0, + "content": "[91] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.377, + 0.906, + 0.431 + ], + "angle": 0, + "content": "[92] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.434, + 0.906, + 0.489 + ], + "angle": 0, + "content": "[93] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.491, + 0.906, + 0.545 + ], + "angle": 0, + "content": "[94] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.548, + 0.906, + 0.602 + ], + "angle": 0, + "content": "[95] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.605, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[96] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.648, + 0.906, + 0.716 + ], + "angle": 0, + "content": "[97] Guang Yu, Siqi Wang, Zhiping Cai, Xinwang Liu, Chuanfu Xu, and Chengkun Wu. Deep anomaly discovery from unlabeled videos via normality advantage and self-paced refinement. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.719, + 0.906, + 0.787 + ], + "angle": 0, + "content": "[98] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. European Conference on Computer Vision, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.525, + 0.789, + 0.906, + 0.856 + ], + "angle": 0, + "content": "[99] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[100] P. Zameni, Y. Shen, and E. Elhamifar. Moscato: Predicting multiple object state change through actions. International Conference on Computer Vision, 2025. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10020" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.16 + ], + "angle": 0, + "content": "[101] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[102] Ali Zare, Yulei Niu, Hammad Ayyubi, and Shih-Fu Chang. Rap: Retrieval-augmented planner for adaptive procedure planning in instructional videos. European Conference on Computer Vision, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[103] Junbin Zhang, Pei-Hsuan Tsai, and Meng-Hsun Tsai. Semantic2graph: Graph-based multi-modal feature fusion for action segmentation in videos, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[104] Shrinivas Ramasubramanian Angela Yao Zhanzhong Pang, Fadime Sener. Long-tail temporal action segmentation with group-wise temporal logit adjustment. European Conference on Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.374 + ], + "angle": 0, + "content": "[105] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. IEEE Conference on Computer Vision and Pattern Recognition, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[106] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.528 + ], + "angle": 0, + "content": "[107] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, , and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.531, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[108] Yuansheng Zhu, Wentao Bao, , and Qi Yu. Towards open set video anomaly detection. European Conference on Computer Vision, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[109] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.629 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10021" + } + ] +] \ No newline at end of file diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_origin.pdf b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7827cb6bd996d0867832e6a0639d800107b177b3 --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/5a14de37-2719-437f-aefb-79f29e2ef354_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b2a36fbab318cd135b899c7d5b6be1d6eaee3030be7e5218f90ac10a7cf2ecb0 +size 735565 diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/full.md b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8d67e95cefb9440d4464339e66a9436df68f1361 --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/full.md @@ -0,0 +1,371 @@ +# Error Recognition in Procedural Videos using Generalized Task Graph + +Shih-Po Lee Northeastern University + +lee.shih@northeastern.edu + +Ehsan Elhamifar Northeastern University + +e.elhamifar@northeastern.edu + +# Abstract + +Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition. + +# 1. Introduction + +The integration of computer vision with augmented and virtual reality (AR & VR) is expected to significantly improve the way we learn and develop skills in our daily lives [83]. Imagine wearing a pair of AR glasses and trying to assemble a new piece of furniture. A visual task assistant should recognize your actions and provide guidance on what to do and how to perform them throughout the process. It should also recognize your possible errors and provide feedback to correct them, when needed. This has motivated exciting new research on action recognition [8, 10, 25, 26, 47, 50, 61, 89] and segmentation [23, 38, 45, 46, 54, 57, 58, 73, 77, 78, 84, 95, 96, 103], video grounding [5, 18, 20, 49, 59, 63, 86, 109], learning + +![](images/e3621cf74a1da1b7fca414e5900e646b458868cb1d4fe17cbe38b7e0f86d8862.jpg) +Figure 1. The illustration of our framework for joint temporal action segmentation and error recognition. It consists of 1) Generalized Task Graph (GTG)-Video Alignment for action segmentation and 2) an Error Recognition Module using features from LLM/VLM for error recognition. + +from instructional videos [12, 33, 62, 64, 65, 85, 102] and very recently on error understanding [14, 27, 29, 41, 60, 90], as well as progress [17] and state [100] prediction. + +Successful operation of a visual task assistant, in particular, requires temporal action segmentation (assigning step labels to frames) and error recognition (recognizing the type of error, if any, in each frame) to seamlessly work together. Additionally, a task graph, which encodes plausible ways to perform a task, has been shown to be critical for temporal action segmentation to effectively mitigate oversegmentation [5, 19, 77, 106]. However, existing methods have three major limitations when it comes to understanding actions and errors. + +SOTA Limitations. First, existing works on error understanding can only detect the presence of an error in a video frame without recognizing the type of the error. As recently studied in [41], different types of errors, such as omission, addition, modification, and slip errors, can occur during task execution. However, [41] cannot distinguish between the error types. A task assistant requires understanding the error type to provide appropriate feedback. For example, an actor knocking over a dripper while making coffee requires + +immediate feedback for correction or spreading jelly onto tortilla using spoon instead of knife needs a feedback after task completion for the actor to better execute the step. + +Second, given the large number of possible errors, it is infeasible to gather real training videos and annotations for all possible errors. Also, recent datasets [41, 66] that contain videos with errors have imbalance annotations on different error types, making error recognition a challenging task, especially under the setting when we do not have error videos during training. + +Third, recent action segmentation methods [19, 77] that use task graphs assume that there are no errors in videos and assign step labels to the frames according to the task graph. In addition, two recent works that handle both action segmentation and error detection [29, 41] cannot use the task graph for action segmentation. This is because a task graph only contains the correct actions/steps and does not consider possible errors that may occur during the task execution. Thus, following the task graph leads to assigning every frame (including frames that have errors and ones that do not belong to any step) to correct actions/steps, which is undesired. + +Paper Contributions. We develop a unified error recognition and temporal action segmentation framework (see Figure 1) to address the aforementioned limitations. + +- We propose a Generalized Task Graph (GTG) whose nodes encode normal actions/steps and background (task-irrelevant actions). +- We develop a dynamic programming-based method, GTG-Video Alignment (GTG2Vid), with a dynamic frame and node dropping term to assign frames to GTG nodes for action segmentation and drop frames for error detection. +- Given the lack of error training videos, we leverage the rich knowledge of large language models (LLMs) in conjunction with the rich representation power of video language models (VLMs) to obtain error descriptions for different types of errors and their corresponding features for error recognition. +- We propose an Error Recognition Module (ERM) to recognize the error types, if any, in the frames predicted by GTG2Vid using joint similarities of video frame, normal type, and error type features. +- By extensive experiments on two egocentric datasets, EgoPER and CaptainCook4D, we show that without using error videos during training, our method 1) improves the performance of temporal action segmentation on error videos, 2) improves error detection, and 3) recognizes different errors. + +# 2. Related Works + +Temporal Action Segmentation. Learning to segment videos into several actions/steps has been studied under un + +supervised [4, 21, 22, 28, 36, 37, 74, 79, 88, 91, 109], weakly-supervised [11, 16, 24, 42-44, 48, 55, 55, 56, 69-71, 76, 84], and fully-supervised [3, 7, 9, 15, 23, 31, 32, 35, 38, 45, 46, 51, 57, 73, 80-82, 84, 95, 96, 103, 104] setting. Since procedural step videos follow certain task graphs to complete the task, some prior works [5, 19, 77, 106] additionally adopt task graph information to guide segmentation process. Specifically, G2Vid [19] aims to assign steps/nodes in a task graph to video frames and drop the frames not corresponding to any node. However, G2Vid cannot drop unnecessary nodes in a task graph nor map background frames to nodes to distinguish between background and error frames. In contrast, our framework adds extra nodes to a task graph and allows droppable nodes to properly assign nodes to frames. + +Error Detection and Recognition for Procedural Tasks. Detecting errors for procedural tasks has recently drawn attention from the video understanding community. Recent works [14, 29, 68, 90] have released relevant datasets to facilitate error understanding. Meanwhile, several works [27, 41, 60, 75] have developed frameworks to detect if errors are present in videos and localize their temporal locations. Recent work [27] focuses on online error detection using action recognition and anticipation models based on videos and action sequences. Another work [41] uses only error-free (normal) videos during training to generate multiple prototypes for action segmentation and error detection. However, none of them has the ability to recognize error types. Our work is able to segment videos and recognize different errors simultaneously without error training videos. + +VLMs/LLMs for Anomaly Detection. Anomaly detection in videos [2, 6, 13, 39, 40, 53, 67, 72, 87, 92, 97-99, 105, 107, 108] has been widely explored, and recent studies [34, 101] leverage VLMs and LLMs for further improvement due to their outstanding reasoning capabilities. Recent work in [101] proposes a training-free video anomaly detection framework using VLMs to generate captions over frames followed by LLMs to predict anomaly scores for those captions. Another work in [34] adopts LLMs to generate normal and abnormal descriptions under various scenes and classify frames into normal or abnormal according to the similarities between features of frames and descriptions. In our paper, we leverage LLMs to generate error descriptions for each error type of every step with crafted prompts, followed by VLMs to generate textual features aligned with visual features for error recognition. + +# 3. Graph to Video Alignment Review + +Our proposed GTG-Video Alignment method builds upon and generalizes the Graph-to-Video Alignment (G2Vid) [19], which we review here. A task graph (also called flow + +graph) is a directed acyclic graph $G = (V, E)$ , where each node $v \in V$ represents a procedure step and each edge $(v_k, v_i) \in E$ from $v_k$ to $v_i$ implies that $v_k$ must be completed before proceeding to $v_i$ , see Figure 2(a). Thus, all predecessors of a node must be completed before starting that step. Thus, the task graph encodes all possible ways to perform a task (this can be obtained by topological sorting). + +To efficiently align a video with the task graph and therefore segment the video into task steps, G2Vid [19] first builds a tSort graph (see Figure 2(a)) that provides a compact representation of all topological sorts of the task graph (each path from the source to the sink node in the tSort graph represents a topological sort of $G$ ). G2Vid alignment recovers a traversal of the tSort graph that best aligns with the video. It uses a modified version of Drop-DTW [18] to align nodes and frames, allowing for i) one-to-many matching (one node/step can be assigned to many frames) and ii) unmatched sequence elements (to be able to drop frames that do not correspond to any node/step). + +Let $C_{i,j}$ denote the matching cost of node $i$ to frame $j$ , $d_j$ denote the dropping cost of frame $j$ , and $A(i)$ denote the set of predecessors of $i$ . The dynamic programming solution computes the cumulative matching $D_{i,j}^{+}$ , dropping $D_{i,j}^{-}$ and optimal $D_{i,j}$ costs up to the $i$ -th node and $j$ -th frame as + +$$ +D _ {i, j} ^ {+} = C _ {i, j} + \min \left\{\min _ {k \in A (i)} \left\{D _ {k, j - 1} \right\}, D _ {i, j - 1} \right\}, \tag {1} +$$ + +$$ +D _ {i, j} ^ {-} = d _ {j} + D _ {i, j - 1}, \quad D _ {i, j} = \min \{D _ {i, j} ^ {+}, D _ {i, j} ^ {-} \}. +$$ + +By iterating over all nodes and frames, the dynamic programming outputs the optimal cost and alignment by backtracking the path with minimum cost, starting from the top left and ending at the bottom right of the cumulative matrix. + +# 4. Proposed Method + +We develop a unified framework for Temporal Action Segmentation (TAS) and Error Recognition (ER) in procedural task videos. We assume having access to only normal (error-free) videos with their frame-wise step labels during training and the task graph during testing. Our proposed framework addresses three fundamental questions: (1) How can we build an efficient Generalized Task Graph (GTG) that enables TAS and Error Detection (ED) for videos? (2) How can we obtain features for different errors without collecting real error videos? (3) How can we leverage the error features to perform ER? In the following, we first formalize the problem, discuss the overarching goal of our framework, and discuss the details of the method to address these questions. + +Problem Setting. For TAS and ER, given a sequence of video frame features $\mathcal{X}^v = (x_1^v,x_2^v,\ldots ,x_T^v)$ , our goal is to predict its frame-wise steps $\hat{\mathcal{Y}} = (\hat{y}_1,\hat{y}_2,\dots ,\hat{y}_T)$ and error types $\hat{\mathcal{E}} = (\hat{e}_1,\hat{e}_2,\dots ,\hat{e}_T)$ , where $T$ is the number of frames, $x_{t}^{v}\in \mathbb{R}^{D}$ is a pre-extracted feature vector, $\hat{y}_t\in$ + +![](images/a05146f0803a0830bf4e86937829de358e4db963586857d11a865f3a7259552c.jpg) + +![](images/eed64a93e6f6d9f4cc02ef338eef0d57fb41cad71ff57662bcc2080c0a8012f6.jpg) + +![](images/4ee92c696037b5102bce497ccf0091ec0a7ba29d6e38758b6c259309a9745275.jpg) +Figure 2. The illustration of (a) an original task graph and its corresponding tSort graph, (b) a naive approach to capture background, (c) our proposed solution, which specifically illustrates the predecessors and successor of extra node $S_{3}$ . + +$\{-1,0,\ldots ,S\}$ , where $S$ is the number of steps and class $-1$ and $0$ indicate error and background (a task-irrelevant action, e.g., answering phone when making coffee), $\hat{e}_t\in \{0,\dots ,M\}$ , where $M$ is the number of error types and class $0$ denotes no error happening. During training, we use only normal (error-free) videos and their ground-truth step labels $\mathcal{V}$ . For testing, we assume that the task graph is given. + +**Framework Overview.** We introduce Generalized Task Graph for handling both $TAS$ and $ED$ . Our key idea is to 1) insert extra nodes in the tSort graph to capture the background class that may occur before each step in the task graph, 2) perform $TAS$ and $ED$ by our GTG2Vid with a dynamic frame and node dropping term for different types of errors, 3) obtain error descriptions and features for all types of errors, respectively, by leveraging LLMs and VLMs and 4) use an Error Recognition Module (ERM) that computes joint similarities of video, normal type and error type features obtained by VLMs to classify the error types for the frames flagged by GTG2Vid. + +# 4.1. Generalized Task Graph (GTG) + +We build GTG to handle execution and omission errors and develop GTG2Vid for TAS and ED. + +Handling Execution Errors. We aim to use the task graph to segment videos into steps and to detect execution errors (e.g., slip, correction, modification, and addition errors defined in [41]). Some of the errors are step-relevant (e.g., spreading jelly using spoon instead of knife) and some are not in the task graph (e.g., adding honey to a recipe that does not include honey). However, G2Vid drops both frames with errors as well as background frames (e.g., answering + +Algorithm 1 GTG-Video Alignment (GTG2Vid) +1: Inputs: $C\in \mathbb{R}^{2S\times T}$ $A(\cdot),d_i^z,d_j^x$ (matching cost matrix, predecessor mapping dictionary, node dropping costs, frame dropping costs.) $\triangleright$ Initialize DP tables +2: $D_{0,0}^{+} = 0;D_{i,0}^{+} = \infty ;D_{0,j}^{+} = \infty ;\forall i\in \{1,\dots ,2S\} ,\forall j\in \{1,\dots ,T\}$ Matching table of $\mathcal{X}_n^v$ (frames) and $\mathcal{Z}$ (nodes) +3: $D_{0,0}^{z - } = 0;D_{i,0}^{z - } = \sum_{k = 1}^{i}d_{k}^{z};D_{0,j}^{z - } = \infty ;\forall i\in \{1,\dots ,2S\} ,\forall j\in \{1,\dots ,T\}$ Dropping table of $\mathcal{Z}$ (nodes) +4: $D_{0,0}^{-x} = 0;D_{i,0}^{-x} = \infty ;D_{0,j}^{-x} = \sum_{k = 1}^{j}d_{k}^{x};\forall i\in \{1,\dots ,2S\} ,\forall j\in \{1,\dots ,T\}$ Dropping table of $\mathcal{X}_n^v$ (frames) +5: $D_{0,0} = 0;D_{i,0} = D_{i,0}^{-x};D_{0,j} = D_{0,j}^{z - }; \forall i\in \{1,\dots ,2S\} ,\forall j\in \{1,\dots ,T\}$ Optimal solution table +6: for $i = 1,\dots ,2S$ do Iterate nodes in Z +7: for $j = 1,\dots ,T$ do Iterate frames in $\mathcal{X}_n^v$ +8: $D_{i,j}^{+} = -C_{i,j} + \min \{\min_{k\in A(i)}\{D_{k,j - 1}\},D_{i,j - 1}\}$ Consider matching node $\mathcal{Z}_i$ to frame $x_{n,j}^v$ +9: $D_{i,j}^{-x} = d_j^x +D_{i,j - 1}$ Consider dropping frame $x_{n,j}^v$ +10: $D_{i,j}^{z - } = d_i^z +\min_{k\in A(i)}\{D_{k,j}\}$ Consider dropping node $\mathcal{Z}_i$ +11: $D_{i,j} = \min \{D_{i,j}^{+},D_{i,j}^{-x},D_{i,j}^{z - }\}$ Select the optimal path +12: end for +13: end for +14: $\hat{y}^* =$ traceback(D) $\triangleright$ Compute the optimal alignment by tracing back the path with minimum cost +15: $\hat{\gamma} =$ frame-assign(y*) Flag dropping frames as error frames and frames belonging to extra nodes as background +16: Outputs: $D_{2S,T},\hat{\gamma}$ + +the phone while making a recipe) and thus cannot distinguish between them and understand when an error occurs. + +To handle this, we build GTG by inserting extra nodes into the task graph to capture background frames. Therefore, by aligning a video with GTG, we can understand which frames correspond to which steps and which frames correspond to background. Additionally, by allowing dropping frames, we can detect execution errors, and by allowing dropping nodes, we can detect omission errors which we will discuss next. + +A naive approach is to insert an extra node with a unique label between every two nodes (see Figure 2 (b)). Therefore, the task graph can potentially map frames to extra nodes. However, the complexity of the tSort graph grows exponentially. To solve this problem, our proposed approach inserts an extra node $S_{i}$ with a unique ID between every node $i$ and its predecessor in the tSort graph. For example, in Figure 2 (c), between node 2 and 3 as well as node 1 and 3, we insert the same extra node $S_{3}$ to capture the background frames with respect to step 3. Our design 1) prevents exponential growth of the tSort graph, only increasing the number of nodes from $S$ to $2S$ , and 2) inherits the correct execution orders in the original tSort graph. + +Handling Omission Errors. During execution, an actor may skip one or more steps in the task graph, resulting in omission error. In this case, forcing models to follow the task graph as done in prior works [5, 19, 77, 106] can lead to wrong TAS results where some frames are assigned to the skipped steps. To handle omission errors, we introduce a node dropping term to our GTG2Vid to make nodes of our GTG droppable, + +$$ +D _ {i, j} ^ {z -} = d _ {i} ^ {z} + \min _ {k \in A (i)} \left\{D _ {k, j} \right\} \tag {2} +$$ + +where $D_{i,j}^{z-}$ is the accumulative cost until node $i$ and frame $j$ in the dynamic programming (DP) table of dropping node and $d_i^z$ is the node dropping cost of node $i$ . A low dropping cost indicates that the corresponding GTG node is not in the video and should be omitted. This also allows dropping extra nodes, whenever there is no corresponding frame in the video. + +GTG-Video Alignment (GTG2Vid). Algorithm 1 shows the steps of GTG2Vid. $C_{i,j}$ is the matching cost of the $i$ -th node to the $j$ -th frame. We let $C_{i,j}$ to be the predicted probability of step $i$ for $j$ -th frame using an action segmentation model. We train the action segmentation model with smoothing and cross-entropy loss in [23] using framewise labels. In addition, to ensure that our proposed method can drop frames and nodes based on the confidence of predictions, we propose a dynamic dropping cost $d^{x}$ for frames and $d^{z}$ for nodes of GTG as + +$$ +d _ {j} ^ {x} = - \left(\frac {\sum_ {i = 1} ^ {2 S} C _ {i , j} \log C _ {i , j}}{\log \frac {1}{2 S}} + \eta\right), \tag {3} +$$ + +$$ +d _ {i} ^ {z} = - \tau \left(1 - \max _ {j = 1, \dots , T} C _ {i, j}\right), +$$ + +where $\eta$ and $\tau$ are hyper-parameters (larger $\eta$ and $\tau$ lead to dropping more frames and nodes). Our frame dropping cost $d^{x}$ leads to drop frames with unconfident predictions, indicating that the frames do not correspond to any GTG node and may contain errors. Meanwhile, our node dropping cost $d^{z}$ allows dropping the nodes whose highest step probabilities over all frames are low. After obtaining the minimum cost path, we generate step predictions $\hat{\mathcal{V}}$ from $\hat{\mathcal{V}}^*$ by flagging the dropped frames as error frames and frames belonging to extra nodes as background. The complexity of our al + +
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTea
w-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAcc
Naive Predictor11.225.013.025.010.425.018.133.311.425.0
EgoPED (MV)7.575.05.575.07.475.010.9100.04.475.0
GTG2Vid (MV)5.150.02.750.04.450.00.633.34.050.0
GTG2Vid (w/o n.f.)16.375.016.475.029.9100.01.733.315.275.0
GTG2Vid31.7100.031.3100.017.8100.04.5100.022.1100.0
+ +
Hot Chocolate +w-F1@0EAccCaptainCook4D
Sandwich +w-F1@0EAccBurritos +w-F1@0EAccRamen +w-F1@0EAccRaita +w-F1@0EAcc
Naive Predictor7.120.05.120.07.425.07.425.07.333.3
EgoPED (MV)7.380.03.480.07.175.03.375.04.366.6
GTG2Vid (MV)4.460.04.740.06.750.02.725.011.466.6
GTG2Vid (w/o n.f.)2.840.06.260.06.125.04.650.011.266.6
GTG2Vid10.5100.08.5100.09.475.04.475.023.0100.0
+ +Table 1. ${ER}$ results on EgoPER and CaptainCook4D. + +girthm is $\mathcal{O}(SQT)$ , same as G2Vid [19], where $S, Q$ , and $T$ denote the number of GTG nodes, separate and linearly-ordered threads of GTG, and frames, respectively. + +# 4.2. Obtaining Error Features via LLM/VLM + +Our next step is to obtain error features for $ER$ . Since we are only given normal videos during training, we i) leverage LLMs to obtain error descriptions for each error type for every step, and ii) use VLMs to generate the textual features that are aligned with visual features. To begin, given a step $i$ and error type $m$ , we generate its error type description using LLMs. The prompt for LLMs combines a description for step $i$ , error definition for error type $m$ , and the number of different error type descriptions to produce. Notice that we directly set normal type description $(m = 0)$ as the description of step $i$ . Next, we generate the feature vector $x_{i,m}^{e}$ for the error description using the text encoder in VLMs, where $x^{e} \in \mathbb{R}^{S \times M \times D}$ is a tensor containing features of $M$ error type descriptions for $S$ steps. When LLMs are prompted to produce multiple error descriptions for an error type, we average all features to get a single vector. Meanwhile, we use the video encoder in VLMs to generate frame features for videos. In this way, we can ensure that the features of videos and error descriptions are aligned in the VLM feature space. See the supplementary materials for detailed prompts and generated error descriptions. + +# 4.3. Error Recognition Module (ERM) + +Our ERM classifies error types for error frames according to their joint similarities of frame, normal type, and error type features from Section 4.2 and obtain $ER$ predictions $\hat{\mathcal{E}}$ . + +Predicting Steps for Error Frames. First, we obtain TAS predictions $\hat{\mathcal{V}}$ by inferring GTG2Vid on a video. Next, we obtain another TAS predictions $\hat{\mathcal{V}}^{+} = (\hat{y}_{1}^{+},\hat{y}_{2}^{+},\dots ,\hat{y}_{T}^{+})$ where $\hat{y}_t^+\in \{0,1,\ldots ,S\}$ , by inferring GTG2Vid without using the frame and node dropping term, forcing GTG2Vid + +to map all frames to nodes. It aims to obtain the step predictions for the error frames in $\hat{\mathcal{V}}$ to find the error type features of the corresponding steps. + +Computing Joint Similarities for Error Recognition. Next, we compute framewise joint similarities $p$ for $ER$ , where $p_{m,j} \in [0,1]$ and $p \in \mathbb{R}^{M \times T}$ . For each frame $j$ , given step prediction $\hat{y}_j, \hat{y}_j^+$ , and feature $x_j^v, x^e$ , we find the error type $\hat{m}_j$ and compute its joint similarity $p_{\hat{m}_j,j}$ as + +$$ +\hat {m} _ {j} = \left\{ \begin{array}{l l} 0, & \text {i f} \hat {y} _ {j} \neq - 1 \\ \underset {m \in \{1, \dots , M \}} {\operatorname {a r g m a x}} s \left(x _ {j} ^ {v}, x ^ {e}, \hat {y} _ {j} ^ {+}, m\right), & \text {o t h e r w i s e} \end{array} \right. \tag {4} +$$ + +$$ +p _ {\hat {m} _ {j}, j} = \left\{ \begin{array}{l l} 1, & \text {i f} \hat {y} _ {j} \neq - 1 \\ s \left(x _ {j} ^ {v}, x ^ {e}, \hat {y} _ {j} ^ {+}, \hat {m} _ {j}\right), & \text {o t h e r w i s e} \end{array} \right. \tag {5} +$$ + +where $s(\cdot ,\cdot ,\cdot ,\cdot)$ is our score function that outputs the joint similarity of the frame. Our score function is formulated as + +$$ +s \left(x _ {j} ^ {v}, x ^ {e}, y, m\right) = \operatorname {s i g m o i d} \left(\left(x _ {y, 0} ^ {e} \odot x _ {y, m} ^ {e}\right) ^ {T} x _ {j} ^ {v}\right) \tag {6} +$$ + +where $\odot$ is element-wise multiplication and $x_{y,0}^{e}$ is the feature of step description (normal type) for step $y$ . The intuition of using normal type feature is that error frames often contain multiple sub-steps, including both normal and error ones (e.g., an actor opens a tortilla bag, grabs a tortilla from it, and accidentally drops it). A single error type feature may be unable to effectively match the corresponding error frames, especially when the number of frames is large and contain multiple sub-steps which the error description does not describe. Therefore, our score function computes the correlation between $x_{y,0}^{e}, x_{y,m}^{e}$ , and $x_{j}^{v}$ . Thus, it jointly considers the similarities of error frames with respect to both normal and error type features. We validate its effectiveness for $ER$ in Section 5. Finally, we apply a maximum filter on each error type of $p$ to smoothen the predictions, followed by argmax to obtain the error type predictions $\hat{\mathcal{E}}$ . + +
MethodQuesadillaOatmealEgoPER
F1@.5AccEditF1@.5AccEditPinwheelCoffeeTeaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid57.053.760.068.874.576.849.152.753.759.157.175.061.156.169.459.058.870.0
EgoPED56.961.447.362.154.870.349.854.741.053.959.655.469.464.558.356.459.054.5
GTG2Vid68.164.569.483.378.588.661.358.460.963.962.573.771.262.676.969.665.373.9
MethodHot ChocolateSandwichBurritosRamenRaitaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid17.119.223.821.321.739.833.427.542.928.033.950.518.523.126.123.725.136.6
EgoPED5.021.713.313.122.518.221.131.821.815.737.325.422.631.323.015.528.920.3
GTG2Vid20.843.527.827.238.538.836.634.041.331.333.642.922.130.826.127.636.135.4
+ +Table 2. TAS results on EgoPER and CaptainCook4D. + +# 5. Experiments + +# 5.1. Experimental Setup + +Dataset. We evaluate our proposed method on EgoPER [41] and CaptainCook4D [66]. EgoPER consists of 5 tasks with 386 egocentric videos and contains 5 types of errors, including omission, addition, modification, slip, and correction. Captaincook4D has 24 tasks with 94.5 hours and contains 8 types of errors, including preparation, measurement, technique, timing, temperature, missing steps, ordering, and other. For EgoPER, we evaluate our method on all 5 tasks and error types with the same training and testing split in [41]. For CaptainCook4D, to have enough videos for training and testing, we select 5 of the tasks which have at least 5 normal and error videos: Spiced Hot Chocolate (Hot Chocolate), Microwave Egg Sandwich (Sandwich), Breakfast Burritos (Burritos), Ramen, and Cucumber Raita (Raita). We assign all normal videos to the training set and error videos to test set. To fit our setting for recognizing execution errors, we report the following error types in CaptainCook4D: preparation (Prep.), measurement (Mea.), technique (Tec.), timing (Time), and temperature (Temp). See the supplementary materials for breakdown analysis of EgoPER and CaptainCook4D and their detailed training and testing splits. + +Evaluation Metrics. We report results on $ER$ , $TAS$ , and $ED$ . In our experiments, we report the F1 score $(F1@\beta)$ with an overlapping threshold $\beta$ . We compute the F1 score by averaging (if not specified) over the F1 scores that are independently computed on the segments for each step or error type. For $ER$ , we report Error Accuracy (EAcc) and weighted $F1@0$ score $(w-F1@0)$ . As $ER$ is challenging under the setting of no error video during training, we set $\beta$ to 0, meaning that a segment is true positive whenever any frame in the segment is correctly classified. EAcc is the percentage of error types that a model can recognize; that is, the error types with $F1@0$ score larger than 0. For $w-F1@0$ , we first compute $F1@0$ on the error type segments, excluding normal ones $(m=0)$ , to alleviate the effect of data imbalance (e.g., + +nearly $90\%$ of the segments in EgoPER are normal). We then compute w-F1@0, where w-F1@0 = $\frac{\mathrm{F1}@\mathrm{0}}{\mathrm{M} - \mathrm{EAcc} \times \mathrm{M} + 1}$ . w-F1@0 penalizes the model that fails to recognize all error types. For TAS, we report frame-wise Accuracy (Acc), segment-wise Edit score (Edit) on normal steps, and F1@.5 score. For ED, we report F1@.5 score on all segments. We also report F1@.5 score on normal (N.) and error segments (E.) separately for alleviating data imbalance. For omission error detection, we use Omission Accuracy (O-Acc), predicting the accuracy for omission steps, and Omission Intersection over Union (O-IoU), as in [41]. + +Baselines. Given the lack of baselines for joint TAS, $ER$ , and $ED$ , we report the performance of EgoPED (MV), a modified version of EgoPED [41] that can additionally perform $ER$ . EgoPED (MV) computes the cosine similarities between frames and error types using the error type features generated in Section 4.2. Then it assigns frames to the type with the highest similarity and uses the majority voting (MV) strategy in [41] over frames within a segment to decide the final error type for each segment. EgoPED (MV) detects error segments with step thresholds equal to mean similarities over videos in the validation set. For $ER$ and $ED$ , since both datasets have data imbalance, we report the performance of Naive Predictor as another baseline. Naive Predictor classifies frames into the same error type for $ER$ and into normal frames for $ED$ . Specifically, it classifies the frames into modification errors for EgoPER and preparation errors for CaptainCook4D. For $TAS$ , we compare with G2Vid [19] to show how G2Vid performs with standard task graphs. To run G2Vid with our cost matrix, we remove all the background entries. + +Implementation Details. We employ GPT-4o mini [1] as our LLM for generating error descriptions and VideoCLIP [93] pre-trained on Ego4D [30] as our VLM for extracting both visual and textual features for videos and error type descriptions. For each task in both EgoPER and CaptainCook4D, we generate at least 2 descriptions for each error type of each action. We set the step description as the description of normal type. We use the backbone in [52] as + +
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTeaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor14.40.07.226.10.013.123.40.011.744.10.022.14.80.02.422.60.011.3
EgoPED30.125.227.729.312.921.114.624.519.69.58.99.236.334.035.124.021.122.5
GTG2Vid40.422.631.558.536.447.430.616.023.333.14.018.651.537.044.242.823.233.0
CaptainCook4D
Hot ChocolateSandwichBurritosRamenRaitaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor8.20.04.16.50.03.23.80.01.94.10.02.03.10.01.65.10.02.6
EgoPED8.12.45.216.36.811.516.38.212.212.410.711.614.19.511.813.47.510.5
GTG2Vid15.820.918.316.712.814.816.919.718.322.68.215.418.29.914.018.014.316.2
+ +our action segmentation model. We train the model using Adam optimizer with learning rate 0.0001 for 3000 iterations. For $ER$ , if $\hat{y}_j = -1$ and $\hat{y}_j^+$ is a background class, we directly assign an addition error and preparation error to $\hat{m}_j$ for EgoPER and CaptainCook4D, respectively, and assign the corresponding similarity to 1. For $ED$ , we produce $ED$ predictions based on $\hat{\mathcal{E}}$ by classifying different error types into a single error type. In practice, we set $\tau = S$ where $S$ is the number of nodes in GTG and $\eta = 0$ . + +# 5.2. Experimental Results + +Error Recognition $(ER)$ . Table 1 compares $ER$ performance between different methods on EgoPER and CaptainCook4D. Our proposed framework, GTG2Vid, demonstrates the best recognition ability than Naive Predictor and EgoPED (MV). Specifically, our proposed method achieves $31.7\%$ , $31.3\%$ , $17.8\%$ , and $22.1\%$ on w-F1@0 compared to $7.5\%$ , $5.5\%$ , $7.4\%$ , and $4.4\%$ by EgoPED (MV) for quesadilla, oatmeal, pinwheel, and tea. On the other hand, our proposed method outperforms EgoPED (MV) on w-F1@0 for all tasks of CaptainCook4D. In addition, our proposed method achieves overall higher EAcc than EgoPED (MV), indicating that our method can recognize more error types on both datasets. Notice that our performance on w-F1@0 for coffee is worse than EgoPED (MV) because EgoPED predicts most of the segments as errors, shown in Table 3. Please see supplementary materials for the F1@0 score of each type of error. + +Next, we investigate the effectiveness of our ERM and our proposed score function (Eq. 6). By replacing ERM with MV, the performance drops significantly (more than $10\%$ on w-F1@0 across all tasks of EgoPER). On the other hand, our method, GTG2Vid (w/o n.f.), without using normal type features suffers a drop on EAcc for most of the tasks, indicating that our proposed function for computing joint similarity has a better ability to recognize different error types. Furthermore, Table 4 shows the Top-1 and Top-2 accuracy of different scoring methods given the features of ground-truth error segments (average over frame features) and generated error type features. Our method gains a per + +Table 3. ED results of different methods on EgoPER and CaptainCook4D. N. and E. denote the F1@.5 score for normal and error segments. + +
MethodQuesadillaPinwheelTea
Top-1Top-2Top-1Top-2Top-1Top-2
Uniform25.050.025.050.025.050.0
Cos. Sim.42.665.827.748.427.845.3
ERM54.681.538.460.334.353.7
+ +Table 4. Performance of feature matching methods on EgoPER. Top-1 and Top-2 denote the top-1 and top-2 accuracy over all error segments. + +formance boost on Top-1 and Top-2 accuracy. Specifically, we obtain $10.7\%$ improvement on Top-1 accuracy compared to cosine similarity (Cos. Sim.) for pinwheel, which verifies our claim in Section 4.3. + +Temporal Action Segmentation (TAS). For TAS performance shown in Table 2, our method obtains a $13.2\%$ , $6.3\%$ , and $19.4\%$ higher F1@.5, Acc and Edit score than EgoPED on All of EgoPER. For CaptainCook4D, our method obtains a $12.1\%$ , $7.2\%$ , and $15.1\%$ higher F1, Acc, and Edit score than EgoPED on All. Furthermore, our method has more robust TAS performance compared to EgoPED, which suffers a significant performance drop in Hot Chocolate and Sandwich. Notice that our Edit score on All is comparable to G2Vid since there are not many omission errors in test videos. Still, our method achieves a higher F1@.5 score than G2Vid, which cannot distinguish between background frames and errors. + +Error Detection (ED). Table 3 shows the $ED$ performance of different methods on EgoPER. Our method achieves $33.0\%$ higher on F1@.5 score, compared to $22.5\%$ by EgoPED for All in EgoPER. Although EgoPED can obtain a higher F1 score on error segments for quesadilla, pinwheel, and coffee, its detection ability for normal segments decreases significantly, specifically a $23.6\%$ drop on F1@.5 score for coffee compared to our method. Therefore, compared to EgoPED, our method achieves a better balance between normal and error segments. For All in CaptainCook4D, GTG2Vid outperforms EgoPED, achieving $16.2\%$ on F1@.5 compared to $10.5\%$ by EgoPED. + +Ablation of Node and Frame Dropping Term. We show ablation studies and investigate the effectiveness of our pro + +![](images/b5349db5d35961d469daadffdba256d69d20f1c92258ad904b8481d8cf19c743.jpg) +Figure 3. Qualitative results for quesadilla, tea, and oatmeal in EgoPER. The top part shows the results of TAS for ground-truths, EgoPED, and our method. The bottom part shows ER results for ground-truths, EgoPED, and our method. Seg. and E.T. denote action segmentation and error type, respectively. + +![](images/2e4dd901808e175f5aab560cb9062a1b11d4cccf2aab8bd3449d25e9b86292cc.jpg) +Figure 4. Ablation of our proposed node and frame dropping term. + +posed node and frame dropping term in GTG2Vid. The two upper plots in Figure 4 show F1@.5 scores of our proposed frame dropping term (Ours) with different $\eta$ and a naive dropping cost (G2Vid) [19] with different $k$ top percentile of the values in the cost matrix, for pinwheel and quesadilla. Our method achieves $26.0\%$ and $33.8\%$ on F1@.5 score compared to $23.2\%$ and $30.9\%$ by G2Vid. On the other hand, the two bottom plots in Figure 4 show the F1 scores for TAS with different $\tau$ for our node dropping term. Smaller $\tau$ indicates GTG2Vid drops a fewer number of nodes. When $\tau$ is larger or equal to 0, GTG2Vid starts dropping nodes, improving TAS performance on videos with omission steps. + +In addition, Figure 5 shows the performance of omission detection for different methods on EgoPER. First, since G2Vid is unable to drop nodes of a graph, it cannot predict any omitted steps, resulting in $0\%$ on O-Acc and O-IoU. Second, although our method suffers a performance drop on O-IoU ( $5.7\%$ lower than EgoPED on All), we gain a performance boost on O-Acc ( $10\%$ higher than EgoPED on All), indicating that our method successfully drops nodes of a task graph according to the input videos and has a stronger capability for detecting omitted steps. To conclude, our proposed method uses 1) dynamic frame dropping term to precisely drop error segments and 2) dynamic node dropping term to drop unnecessary nodes of task graphs for better TAS and omission detection performance. + +Qualitative Analysis. We visualize ER results in Fig. 3. + +![](images/4d4438e99f1859d747038528fb7be64a76fe3a7a09c788b086040ab3a87a311a.jpg) +Figure 5. Omission detection results of different methods on EgoPER. + +Our method can recognize different types of errors, e.g., addition errors (red) on tea, addition and correction (orange) errors on quesadilla, and modification errors (blue) on oatmeal. Our method obtains more accurate TAS, ED, and ER results for error videos on quesadilla, tea and oatmeal compared to EgoPED, thanks to dynamic frame and node dropping term in GTG2Vid and our proposed ERM. See supplementary materials for more qualitative visualization. + +# 6. Conclusions + +We studied joint $ER$ and $TAS$ in procedural task videos. Our proposed framework consists of 1) generalized task graphs to capture correct steps and background, 2) a GTG-Video Alignment algorithm to find the optimal alignment between nodes in GTG and video frames with dynamic dropping terms, 3) a procedure for generating error descriptions for each error type of every step using LLMs and the corresponding features using VLMs, and 4) an Error Recognition Module to effectively perform $ER$ using joint similarities. Our experiments on two datasets showed that our proposed method can obtain promising results on $ER$ , $ED$ , and $TAS$ . + +# Acknowledgement + +This work was funded, in part, by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110), ONR (N000142512287) and ARO + +(W911NF2110276). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the US Government. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6 +[2] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, , and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[3] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 2 +[4] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2 +[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Neural Information Processing Systems, 2023. 1, 2, 4 +[6] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. arXiv: 2110.09742, 2021. 2 +[7] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2 +[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 1 +[9] Daniele Di Mauro Mario Valerio Giuffrida Giovanni Maria Farinella Camillo Quattrocchi, Antonino Furnari. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 2 +[10] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1 +[11] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning + +in instructional videos. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI, 2020. 1 +[13] Hanqiu Deng, Zhaoxiang Zhang, Shihao Zou, and Xingyu Li. Bi-directional frame interpolation for unsupervised video anomaly detection. IEEE Winter Conference on Applications of Computer Vision, 2023. 2 +[14] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv: 2307.16453, 2023. 1, 2 +[15] Guodong Ding, Hans Golong, and Angela Yao. Coherent temporal synthesis for incremental action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[16] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[17] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[18] Nikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan D Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 1, 3 +[19] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 1, 2, 3, 4, 5, 6, 8 +[20] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos Derpanis, Animesh Garg, Richard Wildes, and Allan Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1 +[21] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2 +[22] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2 +[23] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4 +[24] Mohsen Fayyaz and Jurgen Gall. Sect: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[25] C. Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1 +[26] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1 + +[27] Alessandro Flaborea, Guido Melendugno, Leonardo Pliniq, Luca Scofanoq, Edoardo Matteisq, Antonino Furnari, Giovanni Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[28] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2 +[29] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2 +[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 6 +[31] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2 +[32] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2 +[33] Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. European Conference on Computer Vision, 2024. 1 + +[34] Jaehyun Kim, Seongwook Yoon, Taehyeon Choi, and Sanghoon Sull. Unsupervised video anomaly detection based on similarity with predefined text descriptions. Sensors, 2023. 2 +[35] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition. IEEE Winter Conference on Applications of Computer Vision, 2016. 2 +[36] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[37] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2 +[38] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2 +[39] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[40] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. Bman: Bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Transactions on Image Processing, 2019. 2 +[41] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6 +[42] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2 +[43] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. +[44] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2 +[45] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[46] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2 +[47] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers + +for classification and detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.1 +[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[49] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhan, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. European Conference on Computer Vision, 2024. 1 +[50] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 1 +[51] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 2 +[52] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. 2023. 6 +[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[54] Yang Liu, Jiayu Huo, Jingjing Peng, Rachel Sparks, Prokar Dasgupta, Alejandro Granados, and Sebastien Ourselin. Skit: a fast key information video transformer for online surgical phase recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21074-21084, 2023. 1 +[55] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2 +[56] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[57] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[58] Z. Lu and E. Elhamifar. Multi-modal few-shot temporal action segmentation. International Conference on Computer Vision, 2025. 1 +[59] Z. Lu, A. Iftekhar, G. Mittal, T. Meng, X. Wang, C. Zhao, R. Kukkala, E. Elhamifar, and M. Chen. Decafnet: Delegate and conquer for efficient temporal grounding in long videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1 +[60] Antonino Furnari Luigi Seminara, Giovanni Maria Farinella. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. Neural Information Processing Systems, 2024. 1, 2 + +[61] K. Mangalam, H. Fan, Y. Li, C. Wu, B. Xiong, C. Feichtenhofer, and J. Malik. Reversible vision transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[62] A. Miech, J-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1 +[63] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[64] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[65] Kumaranage Ravindu Yasas Nagasinghe, Honglu Zhou, Malitha Gunawardhana, Martin Renqiang Min, Daniel Harari, and Muhammad Haris Khan. Why not use your textbook? knowledge-enhanced procedure planning of instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[66] Rohith Peddi, Shivvrat Arya, Bharath Challa, Likhitha Pallapothula, Akshay Vyas, Bhavya Gouripeddi, Jikai Wang, Qifan Zhang, Vasundhara Komaragiri, Eric Ragan, Nicholas Ruozzi, Yu Xiang, and Vibhav Gogate. CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities, 2024. 2, 6 +[67] Didik Purwanto, Yie-Tarng Chen, and Wen-Hsien Fang. Dance with self-attention: A new look of conditional random fields on anomaly detection in videos. IEEE International Conference on Computer Vision, 2021. 2 +[68] Yicheng Qian, Weixin Luo, Dongze Lian, Xu Tang, Peilin Zhao, and Shenghua Gao. Svip: Sequence verification for procedures in videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[69] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2 +[70] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018. +[71] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2 +[72] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[73] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 1, 2 + +[74] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2 +[75] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepop, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[76] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[77] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4 +[78] Y. Shen and E. Elhamifar. Understanding multi-task activities from single-task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1 +[79] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2 +[80] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2 +[81] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016. +[82] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. CoRR, abs/2105.10859, 2021. 2 +[83] Dean Slawson. The case for vr-immersive and ai-adaptive soft skills training. Training Industry Magazine, 2018. 1 +[84] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 1, 2 +[85] Tomáš Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1 +[86] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1 +[87] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. IEEE Winter Conference on Applications of Computer Vision, 2023. 2 + +[88] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2 +[89] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. CoRR, 2021. 1 +[90] Xin Wang, Taein Kwon, Mahdi Radl Bowen Pan, Ishani Chakraborty, and Sean Andrist. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. IEEE International Conference on Computer Vision, 2023. 1, 2 +[91] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2 +[92] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. European Conference on Computer Vision, 2022. 2 +[93] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. 2021. 6 +[94] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[95] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 1, 2 +[96] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2 +[97] Guang Yu, Siqi Wang, Zhiping Cai, Xinwang Liu, Chuanfu Xu, and Chengkun Wu. Deep anomaly discovery from unlabeled videos via normality advantage and self-paced refinement. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[98] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. European Conference on Computer Vision, 2020. +[99] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2 +[100] P. Zameni, Y. Shen, and E. Elhamifar. Moscato: Predicting multiple object state change through actions. International Conference on Computer Vision, 2025. 1 + +[101] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2 +[102] Ali Zare, Yulei Niu, Hammad Ayyubi, and Shih-Fu Chang. Rap: Retrieval-augmented planner for adaptive procedure planning in instructional videos. European Conference on Computer Vision, 2024. 1 +[103] Junbin Zhang, Pei-Hsuan Tsai, and Meng-Hsun Tsai. Semantic2graph: Graph-based multi-modal feature fusion for action segmentation in videos, 2022. 1, 2 +[104] Shrinivas Ramasubramanian Angela Yao Zhanzhong Pang, Fadime Sener. Long-tail temporal action segmentation with group-wise temporal logit adjustment. European Conference on Computer Vision, 2024. 2 +[105] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. IEEE Conference on Computer Vision and Pattern Recognition, 2011. 2 +[106] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 4 +[107] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, , and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 2019. 2 +[108] Yuansheng Zhu, Wentao Bao, , and Qi Yu. Towards open set video anomaly detection. European Conference on Computer Vision, 2022. 2 +[109] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2 \ No newline at end of file diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/images.zip b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..729e804159badef0c66b7e31b1e6cf41c32c9797 --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:21885ccfc4b551c868c9f866f62b85fecc53e5f59d0d3429472d9d1e8a73b0c9 +size 551246 diff --git a/2025/Error Recognition in Procedural Videos using Generalized Task Graph/layout.json b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..163e1d283c5ea4d6e665995ba7faa37b0a6b614d --- /dev/null +++ b/2025/Error Recognition in Procedural Videos using Generalized Task Graph/layout.json @@ -0,0 +1,12426 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "spans": [ + { + "bbox": [ + 86, + 103, + 523, + 121 + ], + "type": "text", + "content": "Error Recognition in Procedural Videos using Generalized Task Graph" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 140, + 144, + 268, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 144, + 268, + 171 + ], + "spans": [ + { + "bbox": [ + 140, + 144, + 268, + 171 + ], + "type": "text", + "content": "Shih-Po Lee Northeastern University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 141, + 174, + 277, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 174, + 277, + 184 + ], + "spans": [ + { + "bbox": [ + 141, + 174, + 277, + 184 + ], + "type": "text", + "content": "lee.shih@northeastern.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 331, + 144, + 449, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 331, + 144, + 449, + 171 + ], + "spans": [ + { + "bbox": [ + 331, + 144, + 449, + 171 + ], + "type": "text", + "content": "Ehsan Elhamifar Northeastern University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 174, + 465, + 184 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 174, + 465, + 184 + ], + "spans": [ + { + "bbox": [ + 313, + 174, + 465, + 184 + ], + "type": "text", + "content": "e.elhamifar@northeastern.edu" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 239, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 296, + 515 + ], + "type": "text", + "content": "Understanding user actions and their possible mistakes is essential for successful operation of task assistants. In this paper, we develop a unified framework for joint temporal action segmentation and error recognition (recognizing when and which type of error happens) in procedural task videos. We propose a Generalized Task Graph (GTG) whose nodes encode correct steps and background (task-irrelevant actions). We then develop a GTG-Video Alignment algorithm (GTG2Vid) to jointly segment videos into actions and detect frames containing errors. Given that it is infeasible to gather many videos and their annotations for different types of errors, we study a framework that only requires normal (error-free) videos during training. More specifically, we leverage large language models (LLMs) to obtain error descriptions and subsequently use video-language models (VLMs) to generate visually-aligned textual features, which we use for error recognition. We then propose an Error Recognition Module (ERM) to recognize the error frames predicted by GTG2Vid using the generated error features. By extensive experiments on two egocentric datasets of EgoPER and CaptainCook4D, we show that our framework outperforms other baselines on action segmentation, error detection and recognition." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 536, + 135, + 548 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "The integration of computer vision with augmented and virtual reality (AR & VR) is expected to significantly improve the way we learn and develop skills in our daily lives [83]. Imagine wearing a pair of AR glasses and trying to assemble a new piece of furniture. A visual task assistant should recognize your actions and provide guidance on what to do and how to perform them throughout the process. It should also recognize your possible errors and provide feedback to correct them, when needed. This has motivated exciting new research on action recognition [8, 10, 25, 26, 47, 50, 61, 89] and segmentation [23, 38, 45, 46, 54, 57, 58, 73, 77, 78, 84, 95, 96, 103], video grounding [5, 18, 20, 49, 59, 63, 86, 109], learning" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 321, + 213, + 553, + 396 + ], + "blocks": [ + { + "bbox": [ + 321, + 213, + 553, + 396 + ], + "lines": [ + { + "bbox": [ + 321, + 213, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 321, + 213, + 553, + 396 + ], + "type": "image", + "image_path": "e3621cf74a1da1b7fca414e5900e646b458868cb1d4fe17cbe38b7e0f86d8862.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 398, + 555, + 437 + ], + "lines": [ + { + "bbox": [ + 313, + 398, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 555, + 437 + ], + "type": "text", + "content": "Figure 1. The illustration of our framework for joint temporal action segmentation and error recognition. It consists of 1) Generalized Task Graph (GTG)-Video Alignment for action segmentation and 2) an Error Recognition Module using features from LLM/VLM for error recognition." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 554, + 483 + ], + "type": "text", + "content": "from instructional videos [12, 33, 62, 64, 65, 85, 102] and very recently on error understanding [14, 27, 29, 41, 60, 90], as well as progress [17] and state [100] prediction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 554, + 601 + ], + "type": "text", + "content": "Successful operation of a visual task assistant, in particular, requires temporal action segmentation (assigning step labels to frames) and error recognition (recognizing the type of error, if any, in each frame) to seamlessly work together. Additionally, a task graph, which encodes plausible ways to perform a task, has been shown to be critical for temporal action segmentation to effectively mitigate oversegmentation [5, 19, 77, 106]. However, existing methods have three major limitations when it comes to understanding actions and errors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "SOTA Limitations. First, existing works on error understanding can only detect the presence of an error in a video frame without recognizing the type of the error. As recently studied in [41], different types of errors, such as omission, addition, modification, and slip errors, can occur during task execution. However, [41] cannot distinguish between the error types. A task assistant requires understanding the error type to provide appropriate feedback. For example, an actor knocking over a dripper while making coffee requires" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10009" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "immediate feedback for correction or spreading jelly onto tortilla using spoon instead of knife needs a feedback after task completion for the actor to better execute the step." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 294, + 191 + ], + "type": "text", + "content": "Second, given the large number of possible errors, it is infeasible to gather real training videos and annotations for all possible errors. Also, recent datasets [41, 66] that contain videos with errors have imbalance annotations on different error types, making error recognition a challenging task, especially under the setting when we do not have error videos during training." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 192, + 294, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 192, + 294, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 192, + 294, + 334 + ], + "type": "text", + "content": "Third, recent action segmentation methods [19, 77] that use task graphs assume that there are no errors in videos and assign step labels to the frames according to the task graph. In addition, two recent works that handle both action segmentation and error detection [29, 41] cannot use the task graph for action segmentation. This is because a task graph only contains the correct actions/steps and does not consider possible errors that may occur during the task execution. Thus, following the task graph leads to assigning every frame (including frames that have errors and ones that do not belong to any step) to correct actions/steps, which is undesired." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 294, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 294, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 294, + 372 + ], + "type": "text", + "content": "Paper Contributions. We develop a unified error recognition and temporal action segmentation framework (see Figure 1) to address the aforementioned limitations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 375, + 294, + 657 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 375, + 294, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 294, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 294, + 411 + ], + "type": "text", + "content": "- We propose a Generalized Task Graph (GTG) whose nodes encode normal actions/steps and background (task-irrelevant actions)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 294, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 294, + 460 + ], + "type": "text", + "content": "- We develop a dynamic programming-based method, GTG-Video Alignment (GTG2Vid), with a dynamic frame and node dropping term to assign frames to GTG nodes for action segmentation and drop frames for error detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 463, + 294, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 294, + 534 + ], + "type": "text", + "content": "- Given the lack of error training videos, we leverage the rich knowledge of large language models (LLMs) in conjunction with the rich representation power of video language models (VLMs) to obtain error descriptions for different types of errors and their corresponding features for error recognition." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 536, + 294, + 585 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 294, + 585 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 294, + 585 + ], + "type": "text", + "content": "- We propose an Error Recognition Module (ERM) to recognize the error types, if any, in the frames predicted by GTG2Vid using joint similarities of video frame, normal type, and error type features." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 586, + 294, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 294, + 657 + ], + "type": "text", + "content": "- By extensive experiments on two egocentric datasets, EgoPER and CaptainCook4D, we show that without using error videos during training, our method 1) improves the performance of temporal action segmentation on error videos, 2) improves error detection, and 3) recognizes different errors." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 667, + 146, + 680 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 667, + 146, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 667, + 146, + 680 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "Temporal Action Segmentation. Learning to segment videos into several actions/steps has been studied under un" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 251 + ], + "type": "text", + "content": "supervised [4, 21, 22, 28, 36, 37, 74, 79, 88, 91, 109], weakly-supervised [11, 16, 24, 42-44, 48, 55, 55, 56, 69-71, 76, 84], and fully-supervised [3, 7, 9, 15, 23, 31, 32, 35, 38, 45, 46, 51, 57, 73, 80-82, 84, 95, 96, 103, 104] setting. Since procedural step videos follow certain task graphs to complete the task, some prior works [5, 19, 77, 106] additionally adopt task graph information to guide segmentation process. Specifically, G2Vid [19] aims to assign steps/nodes in a task graph to video frames and drop the frames not corresponding to any node. However, G2Vid cannot drop unnecessary nodes in a task graph nor map background frames to nodes to distinguish between background and error frames. In contrast, our framework adds extra nodes to a task graph and allows droppable nodes to properly assign nodes to frames." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 256, + 553, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 256, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 256, + 553, + 447 + ], + "type": "text", + "content": "Error Detection and Recognition for Procedural Tasks. Detecting errors for procedural tasks has recently drawn attention from the video understanding community. Recent works [14, 29, 68, 90] have released relevant datasets to facilitate error understanding. Meanwhile, several works [27, 41, 60, 75] have developed frameworks to detect if errors are present in videos and localize their temporal locations. Recent work [27] focuses on online error detection using action recognition and anticipation models based on videos and action sequences. Another work [41] uses only error-free (normal) videos during training to generate multiple prototypes for action segmentation and error detection. However, none of them has the ability to recognize error types. Our work is able to segment videos and recognize different errors simultaneously without error training videos." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 452, + 553, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 452, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 452, + 553, + 643 + ], + "type": "text", + "content": "VLMs/LLMs for Anomaly Detection. Anomaly detection in videos [2, 6, 13, 39, 40, 53, 67, 72, 87, 92, 97-99, 105, 107, 108] has been widely explored, and recent studies [34, 101] leverage VLMs and LLMs for further improvement due to their outstanding reasoning capabilities. Recent work in [101] proposes a training-free video anomaly detection framework using VLMs to generate captions over frames followed by LLMs to predict anomaly scores for those captions. Another work in [34] adopts LLMs to generate normal and abnormal descriptions under various scenes and classify frames into normal or abnormal according to the similarities between features of frames and descriptions. In our paper, we leverage LLMs to generate error descriptions for each error type of every step with crafted prompts, followed by VLMs to generate textual features aligned with visual features for error recognition." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 656, + 504, + 670 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 504, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 504, + 670 + ], + "type": "text", + "content": "3. Graph to Video Alignment Review" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Our proposed GTG-Video Alignment method builds upon and generalizes the Graph-to-Video Alignment (G2Vid) [19], which we review here. A task graph (also called flow" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10010" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "graph) is a directed acyclic graph " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "G = (V, E)" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", where each node " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "v \\in V" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " represents a procedure step and each edge " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "(v_k, v_i) \\in E" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "v_k" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " implies that " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "v_k" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": " must be completed before proceeding to " + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "inline_equation", + "content": "v_i" + }, + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": ", see Figure 2(a). Thus, all predecessors of a node must be completed before starting that step. Thus, the task graph encodes all possible ways to perform a task (this can be obtained by topological sorting)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 157, + 295, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 157, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 157, + 295, + 300 + ], + "type": "text", + "content": "To efficiently align a video with the task graph and therefore segment the video into task steps, G2Vid [19] first builds a tSort graph (see Figure 2(a)) that provides a compact representation of all topological sorts of the task graph (each path from the source to the sink node in the tSort graph represents a topological sort of " + }, + { + "bbox": [ + 54, + 157, + 295, + 300 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 54, + 157, + 295, + 300 + ], + "type": "text", + "content": "). G2Vid alignment recovers a traversal of the tSort graph that best aligns with the video. It uses a modified version of Drop-DTW [18] to align nodes and frames, allowing for i) one-to-many matching (one node/step can be assigned to many frames) and ii) unmatched sequence elements (to be able to drop frames that do not correspond to any node/step)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "C_{i,j}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " denote the matching cost of node " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "d_j" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " denote the dropping cost of frame " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "A(i)" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " denote the set of predecessors of " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ". The dynamic programming solution computes the cumulative matching " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "D_{i,j}^{+}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": ", dropping " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "D_{i,j}^{-}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " and optimal " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "D_{i,j}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " costs up to the " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "-th node and " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "-th frame as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 368, + 295, + 386 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 368, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 82, + 368, + 295, + 386 + ], + "type": "interline_equation", + "content": "D _ {i, j} ^ {+} = C _ {i, j} + \\min \\left\\{\\min _ {k \\in A (i)} \\left\\{D _ {k, j - 1} \\right\\}, D _ {i, j - 1} \\right\\}, \\tag {1}", + "image_path": "dc1f996ac778b59ad370e7b2c670d82c8c336c23911fc6c67b1ad5842ec7b2c3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 386, + 259, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 386, + 259, + 396 + ], + "spans": [ + { + "bbox": [ + 85, + 386, + 259, + 396 + ], + "type": "interline_equation", + "content": "D _ {i, j} ^ {-} = d _ {j} + D _ {i, j - 1}, \\quad D _ {i, j} = \\min \\{D _ {i, j} ^ {+}, D _ {i, j} ^ {-} \\}.", + "image_path": "25a3a4d82268d2a8bdfe30fc5c1d575dacf378a8e11192e75a6d0fa806638b60.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 403, + 295, + 452 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 295, + 452 + ], + "type": "text", + "content": "By iterating over all nodes and frames, the dynamic programming outputs the optimal cost and alignment by backtracking the path with minimum cost, starting from the top left and ending at the bottom right of the cumulative matrix." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 162, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 162, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 162, + 475 + ], + "type": "text", + "content": "4. Proposed Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 482, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 482, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 482, + 295, + 649 + ], + "type": "text", + "content": "We develop a unified framework for Temporal Action Segmentation (TAS) and Error Recognition (ER) in procedural task videos. We assume having access to only normal (error-free) videos with their frame-wise step labels during training and the task graph during testing. Our proposed framework addresses three fundamental questions: (1) How can we build an efficient Generalized Task Graph (GTG) that enables TAS and Error Detection (ED) for videos? (2) How can we obtain features for different errors without collecting real error videos? (3) How can we leverage the error features to perform ER? In the following, we first formalize the problem, discuss the overarching goal of our framework, and discuss the details of the method to address these questions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": "Problem Setting. For TAS and ER, given a sequence of video frame features " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{X}^v = (x_1^v,x_2^v,\\ldots ,x_T^v)" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", our goal is to predict its frame-wise steps " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{Y}} = (\\hat{y}_1,\\hat{y}_2,\\dots ,\\hat{y}_T)" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " and error types " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}} = (\\hat{e}_1,\\hat{e}_2,\\dots ,\\hat{e}_T)" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " is the number of frames, " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "x_{t}^{v}\\in \\mathbb{R}^{D}" + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "text", + "content": " is a pre-extracted feature vector, " + }, + { + "bbox": [ + 55, + 652, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_t\\in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 333, + 75, + 537, + 149 + ], + "blocks": [ + { + "bbox": [ + 333, + 75, + 537, + 149 + ], + "lines": [ + { + "bbox": [ + 333, + 75, + 537, + 149 + ], + "spans": [ + { + "bbox": [ + 333, + 75, + 537, + 149 + ], + "type": "image", + "image_path": "a05146f0803a0830bf4e86937829de358e4db963586857d11a865f3a7259552c.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 333, + 150, + 537, + 213 + ], + "blocks": [ + { + "bbox": [ + 333, + 150, + 537, + 213 + ], + "lines": [ + { + "bbox": [ + 333, + 150, + 537, + 213 + ], + "spans": [ + { + "bbox": [ + 333, + 150, + 537, + 213 + ], + "type": "image", + "image_path": "eed64a93e6f6d9f4cc02ef338eef0d57fb41cad71ff57662bcc2080c0a8012f6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 333, + 213, + 537, + 277 + ], + "blocks": [ + { + "bbox": [ + 333, + 213, + 537, + 277 + ], + "lines": [ + { + "bbox": [ + 333, + 213, + 537, + 277 + ], + "spans": [ + { + "bbox": [ + 333, + 213, + 537, + 277 + ], + "type": "image", + "image_path": "4ee92c696037b5102bce497ccf0091ec0a7ba29d6e38758b6c259309a9745275.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 281, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 553, + 319 + ], + "type": "text", + "content": "Figure 2. The illustration of (a) an original task graph and its corresponding tSort graph, (b) a naive approach to capture background, (c) our proposed solution, which specifically illustrates the predecessors and successor of extra node " + }, + { + "bbox": [ + 313, + 281, + 553, + 319 + ], + "type": "inline_equation", + "content": "S_{3}" + }, + { + "bbox": [ + 313, + 281, + 553, + 319 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\{-1,0,\\ldots ,S\\}" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": " is the number of steps and class " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": " indicate error and background (a task-irrelevant action, e.g., answering phone when making coffee), " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\hat{e}_t\\in \\{0,\\dots ,M\\}" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": " is the number of error types and class " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "0" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": " denotes no error happening. During training, we use only normal (error-free) videos and their ground-truth step labels " + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 330, + 554, + 415 + ], + "type": "text", + "content": ". For testing, we assume that the task graph is given." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "content": "**Framework Overview.** We introduce Generalized Task Graph for handling both " + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "content": ". Our key idea is to 1) insert extra nodes in the tSort graph to capture the background class that may occur before each step in the task graph, 2) perform " + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 418, + 555, + 562 + ], + "type": "text", + "content": " by our GTG2Vid with a dynamic frame and node dropping term for different types of errors, 3) obtain error descriptions and features for all types of errors, respectively, by leveraging LLMs and VLMs and 4) use an Error Recognition Module (ERM) that computes joint similarities of video, normal type and error type features obtained by VLMs to classify the error types for the frames flagged by GTG2Vid." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 571, + 487, + 583 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 571, + 487, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 571, + 487, + 583 + ], + "type": "text", + "content": "4.1. Generalized Task Graph (GTG)" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 590, + 553, + 613 + ], + "type": "text", + "content": "We build GTG to handle execution and omission errors and develop GTG2Vid for TAS and ED." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Handling Execution Errors. We aim to use the task graph to segment videos into steps and to detect execution errors (e.g., slip, correction, modification, and addition errors defined in [41]). Some of the errors are step-relevant (e.g., spreading jelly using spoon instead of knife) and some are not in the task graph (e.g., adding honey to a recipe that does not include honey). However, G2Vid drops both frames with errors as well as background frames (e.g., answering" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10011" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "code", + "bbox": [ + 57, + 86, + 555, + 297 + ], + "blocks": [ + { + "bbox": [ + 56, + 72, + 254, + 84 + ], + "lines": [ + { + "bbox": [ + 56, + 72, + 254, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 254, + 84 + ], + "type": "text", + "content": "Algorithm 1 GTG-Video Alignment (GTG2Vid)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "lines": [ + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": "1: Inputs: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "C\\in \\mathbb{R}^{2S\\times T}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "A(\\cdot),d_i^z,d_j^x" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " (matching cost matrix, predecessor mapping dictionary, node dropping costs, frame dropping costs.) " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Initialize DP tables \n2: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{0,0}^{+} = 0;D_{i,0}^{+} = \\infty ;D_{0,j}^{+} = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Matching table of " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_n^v" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " (frames) and " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " (nodes) \n3: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{0,0}^{z - } = 0;D_{i,0}^{z - } = \\sum_{k = 1}^{i}d_{k}^{z};D_{0,j}^{z - } = \\infty ;\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Dropping table of " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " (nodes) \n4: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{0,0}^{-x} = 0;D_{i,0}^{-x} = \\infty ;D_{0,j}^{-x} = \\sum_{k = 1}^{j}d_{k}^{x};\\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Dropping table of " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_n^v" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " (frames) \n5: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{0,0} = 0;D_{i,0} = D_{i,0}^{-x};D_{0,j} = D_{0,j}^{z - }; \\forall i\\in \\{1,\\dots ,2S\\} ,\\forall j\\in \\{1,\\dots ,T\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Optimal solution table \n6: for " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "i = 1,\\dots ,2S" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " do Iterate nodes in Z \n7: for " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "j = 1,\\dots ,T" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " do Iterate frames in " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{X}_n^v" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " \n8: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{i,j}^{+} = -C_{i,j} + \\min \\{\\min_{k\\in A(i)}\\{D_{k,j - 1}\\},D_{i,j - 1}\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Consider matching node " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_i" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " to frame " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{n,j}^v" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " \n9: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{i,j}^{-x} = d_j^x +D_{i,j - 1}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Consider dropping frame " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "x_{n,j}^v" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " \n10: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{i,j}^{z - } = d_i^z +\\min_{k\\in A(i)}\\{D_{k,j}\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Consider dropping node " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\mathcal{Z}_i" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " \n11: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{i,j} = \\min \\{D_{i,j}^{+},D_{i,j}^{-x},D_{i,j}^{z - }\\}" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Select the optimal path \n12: end for \n13: end for \n14: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\hat{y}^* =" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " traceback(D) " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\triangleright" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " Compute the optimal alignment by tracing back the path with minimum cost \n15: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma} =" + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "text", + "content": " frame-assign(y*) Flag dropping frames as error frames and frames belonging to extra nodes as background \n16: Outputs: " + }, + { + "bbox": [ + 57, + 86, + 555, + 297 + ], + "type": "inline_equation", + "content": "D_{2S,T},\\hat{\\gamma}" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_body" + } + ], + "index": 1, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 318, + 294, + 342 + ], + "type": "text", + "content": "the phone while making a recipe) and thus cannot distinguish between them and understand when an error occurs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 342, + 295, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 437 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 437 + ], + "type": "text", + "content": "To handle this, we build GTG by inserting extra nodes into the task graph to capture background frames. Therefore, by aligning a video with GTG, we can understand which frames correspond to which steps and which frames correspond to background. Additionally, by allowing dropping frames, we can detect execution errors, and by allowing dropping nodes, we can detect omission errors which we will discuss next." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": "A naive approach is to insert an extra node with a unique label between every two nodes (see Figure 2 (b)). Therefore, the task graph can potentially map frames to extra nodes. However, the complexity of the tSort graph grows exponentially. To solve this problem, our proposed approach inserts an extra node " + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "S_{i}" + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": " with a unique ID between every node " + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": " and its predecessor in the tSort graph. For example, in Figure 2 (c), between node 2 and 3 as well as node 1 and 3, we insert the same extra node " + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "S_{3}" + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": " to capture the background frames with respect to step 3. Our design 1) prevents exponential growth of the tSort graph, only increasing the number of nodes from " + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "inline_equation", + "content": "2S" + }, + { + "bbox": [ + 55, + 437, + 295, + 594 + ], + "type": "text", + "content": ", and 2) inherits the correct execution orders in the original tSort graph." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 596, + 295, + 692 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 295, + 692 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 295, + 692 + ], + "type": "text", + "content": "Handling Omission Errors. During execution, an actor may skip one or more steps in the task graph, resulting in omission error. In this case, forcing models to follow the task graph as done in prior works [5, 19, 77, 106] can lead to wrong TAS results where some frames are assigned to the skipped steps. To handle omission errors, we introduce a node dropping term to our GTG2Vid to make nodes of our GTG droppable," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 113, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 113, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "D _ {i, j} ^ {z -} = d _ {i} ^ {z} + \\min _ {k \\in A (i)} \\left\\{D _ {k, j} \\right\\} \\tag {2}", + "image_path": "204ae5c5a31bff45ac6d3ff80d1dbd93a9d7bf2d78c3ba9f5d2a3632b671d7af.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "inline_equation", + "content": "D_{i,j}^{z-}" + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": " is the accumulative cost until node " + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": " and frame " + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": " in the dynamic programming (DP) table of dropping node and " + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "inline_equation", + "content": "d_i^z" + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": " is the node dropping cost of node " + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 318, + 555, + 401 + ], + "type": "text", + "content": ". A low dropping cost indicates that the corresponding GTG node is not in the video and should be omitted. This also allows dropping extra nodes, whenever there is no corresponding frame in the video." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": "GTG-Video Alignment (GTG2Vid). Algorithm 1 shows the steps of GTG2Vid. " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "C_{i,j}" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": " is the matching cost of the " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": "-th node to the " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": "-th frame. We let " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "C_{i,j}" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": " to be the predicted probability of step " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": "-th frame using an action segmentation model. We train the action segmentation model with smoothing and cross-entropy loss in [23] using framewise labels. In addition, to ensure that our proposed method can drop frames and nodes based on the confidence of predictions, we propose a dynamic dropping cost " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "d^{x}" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": " for frames and " + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "inline_equation", + "content": "d^{z}" + }, + { + "bbox": [ + 313, + 404, + 556, + 523 + ], + "type": "text", + "content": " for nodes of GTG as" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 529, + 553, + 564 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 529, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 359, + 529, + 553, + 564 + ], + "type": "interline_equation", + "content": "d _ {j} ^ {x} = - \\left(\\frac {\\sum_ {i = 1} ^ {2 S} C _ {i , j} \\log C _ {i , j}}{\\log \\frac {1}{2 S}} + \\eta\\right), \\tag {3}", + "image_path": "19dc6a924faa87c74fe793d6f19ef62c4830667e7ca134f7786edaad497777a5.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 360, + 565, + 485, + 590 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 565, + 485, + 590 + ], + "spans": [ + { + "bbox": [ + 360, + 565, + 485, + 590 + ], + "type": "interline_equation", + "content": "d _ {i} ^ {z} = - \\tau \\left(1 - \\max _ {j = 1, \\dots , T} C _ {i, j}\\right),", + "image_path": "0681ec4d3fde7f6a3c343a45185315e726039c6e64d82c50d8a8b09e3a919cd4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " are hyper-parameters (larger " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " lead to dropping more frames and nodes). Our frame dropping cost " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "d^{x}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " leads to drop frames with unconfident predictions, indicating that the frames do not correspond to any GTG node and may contain errors. Meanwhile, our node dropping cost " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "d^{z}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " allows dropping the nodes whose highest step probabilities over all frames are low. After obtaining the minimum cost path, we generate step predictions " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}^*" + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": " by flagging the dropped frames as error frames and frames belonging to extra nodes as background. The complexity of our al" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10012" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 541, + 159 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 159 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 159 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 159 + ], + "type": "table", + "html": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTea
w-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAccw-F1@0EAcc
Naive Predictor11.225.013.025.010.425.018.133.311.425.0
EgoPED (MV)7.575.05.575.07.475.010.9100.04.475.0
GTG2Vid (MV)5.150.02.750.04.450.00.633.34.050.0
GTG2Vid (w/o n.f.)16.375.016.475.029.9100.01.733.315.275.0
GTG2Vid31.7100.031.3100.017.8100.04.5100.022.1100.0
", + "image_path": "2de1b8e6b76815aac0da19092269e383b54895a79c7faf4dfd010c6750db33b6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 160, + 541, + 245 + ], + "blocks": [ + { + "bbox": [ + 67, + 160, + 541, + 245 + ], + "lines": [ + { + "bbox": [ + 67, + 160, + 541, + 245 + ], + "spans": [ + { + "bbox": [ + 67, + 160, + 541, + 245 + ], + "type": "table", + "html": "
Hot Chocolate \nw-F1@0EAccCaptainCook4D
Sandwich \nw-F1@0EAccBurritos \nw-F1@0EAccRamen \nw-F1@0EAccRaita \nw-F1@0EAcc
Naive Predictor7.120.05.120.07.425.07.425.07.333.3
EgoPED (MV)7.380.03.480.07.175.03.375.04.366.6
GTG2Vid (MV)4.460.04.740.06.750.02.725.011.466.6
GTG2Vid (w/o n.f.)2.840.06.260.06.125.04.650.011.266.6
GTG2Vid10.5100.08.5100.09.475.04.475.023.0100.0
", + "image_path": "b6777a85b5711133c60a662e98b1170dd0bca1828787aa5aa2b6fb84fd6f113c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 216, + 247, + 394, + 257 + ], + "lines": [ + { + "bbox": [ + 216, + 247, + 394, + 257 + ], + "spans": [ + { + "bbox": [ + 216, + 247, + 394, + 257 + ], + "type": "text", + "content": "Table 1. " + }, + { + "bbox": [ + 216, + 247, + 394, + 257 + ], + "type": "inline_equation", + "content": "{ER}" + }, + { + "bbox": [ + 216, + 247, + 394, + 257 + ], + "type": "text", + "content": " results on EgoPER and CaptainCook4D." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "spans": [ + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "content": "girthm is " + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "inline_equation", + "content": "\\mathcal{O}(SQT)" + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "content": ", same as G2Vid [19], where " + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "inline_equation", + "content": "S, Q" + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 271, + 296, + 308 + ], + "type": "text", + "content": " denote the number of GTG nodes, separate and linearly-ordered threads of GTG, and frames, respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 272, + 329 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 272, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 272, + 329 + ], + "type": "text", + "content": "4.2. Obtaining Error Features via LLM/VLM" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": "Our next step is to obtain error features for " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": ". Since we are only given normal videos during training, we i) leverage LLMs to obtain error descriptions for each error type for every step, and ii) use VLMs to generate the textual features that are aligned with visual features. To begin, given a step " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " and error type " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": ", we generate its error type description using LLMs. The prompt for LLMs combines a description for step " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": ", error definition for error type " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": ", and the number of different error type descriptions to produce. Notice that we directly set normal type description " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "(m = 0)" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " as the description of step " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": ". Next, we generate the feature vector " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "x_{i,m}^{e}" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " for the error description using the text encoder in VLMs, where " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "x^{e} \\in \\mathbb{R}^{S \\times M \\times D}" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " is a tensor containing features of " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " error type descriptions for " + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 54, + 334, + 296, + 587 + ], + "type": "text", + "content": " steps. When LLMs are prompted to produce multiple error descriptions for an error type, we average all features to get a single vector. Meanwhile, we use the video encoder in VLMs to generate frame features for videos. In this way, we can ensure that the features of videos and error descriptions are aligned in the VLM feature space. See the supplementary materials for detailed prompts and generated error descriptions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 596, + 239, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 239, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 239, + 609 + ], + "type": "text", + "content": "4.3. Error Recognition Module (ERM)" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "text", + "content": "Our ERM classifies error types for error frames according to their joint similarities of frame, normal type, and error type features from Section 4.2 and obtain " + }, + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "text", + "content": " predictions " + }, + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}" + }, + { + "bbox": [ + 55, + 613, + 295, + 651 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Predicting Steps for Error Frames. First, we obtain TAS predictions " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " by inferring GTG2Vid on a video. Next, we obtain another TAS predictions " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}^{+} = (\\hat{y}_{1}^{+},\\hat{y}_{2}^{+},\\dots ,\\hat{y}_{T}^{+})" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\hat{y}_t^+\\in \\{0,1,\\ldots ,S\\}" + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": ", by inferring GTG2Vid without using the frame and node dropping term, forcing GTG2Vid" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 271, + 553, + 308 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 308 + ], + "type": "text", + "content": "to map all frames to nodes. It aims to obtain the step predictions for the error frames in " + }, + { + "bbox": [ + 313, + 271, + 553, + 308 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{V}}" + }, + { + "bbox": [ + 313, + 271, + 553, + 308 + ], + "type": "text", + "content": " to find the error type features of the corresponding steps." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": "Computing Joint Similarities for Error Recognition. Next, we compute framewise joint similarities " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "p_{m,j} \\in [0,1]" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "p \\in \\mathbb{R}^{M \\times T}" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": ". For each frame " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": ", given step prediction " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\hat{y}_j, \\hat{y}_j^+" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": ", and feature " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "x_j^v, x^e" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": ", we find the error type " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\hat{m}_j" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": " and compute its joint similarity " + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "inline_equation", + "content": "p_{\\hat{m}_j,j}" + }, + { + "bbox": [ + 313, + 309, + 555, + 371 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 325, + 384, + 555, + 422 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 384, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 325, + 384, + 555, + 422 + ], + "type": "interline_equation", + "content": "\\hat {m} _ {j} = \\left\\{ \\begin{array}{l l} 0, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ \\underset {m \\in \\{1, \\dots , M \\}} {\\operatorname {a r g m a x}} s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, m\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {4}", + "image_path": "e11bd4efd05abdab8b15efb66c65a62e2e0322a761a1e385b44146fd63fc37d3.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 345, + 428, + 554, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 428, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 345, + 428, + 554, + 459 + ], + "type": "interline_equation", + "content": "p _ {\\hat {m} _ {j}, j} = \\left\\{ \\begin{array}{l l} 1, & \\text {i f} \\hat {y} _ {j} \\neq - 1 \\\\ s \\left(x _ {j} ^ {v}, x ^ {e}, \\hat {y} _ {j} ^ {+}, \\hat {m} _ {j}\\right), & \\text {o t h e r w i s e} \\end{array} \\right. \\tag {5}", + "image_path": "cc8b11e434c83888da791228c3b641b3a8d2a1faebfb0cd004855203f65f95cb.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "inline_equation", + "content": "s(\\cdot ,\\cdot ,\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "text", + "content": " is our score function that outputs the joint similarity of the frame. Our score function is formulated as" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 337, + 502, + 553, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 502, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 337, + 502, + 553, + 519 + ], + "type": "interline_equation", + "content": "s \\left(x _ {j} ^ {v}, x ^ {e}, y, m\\right) = \\operatorname {s i g m o i d} \\left(\\left(x _ {y, 0} ^ {e} \\odot x _ {y, m} ^ {e}\\right) ^ {T} x _ {j} ^ {v}\\right) \\tag {6}", + "image_path": "aeb02cbfbdaf0cebb1576e5f5660a8f6fdd4f89271f2495f6779b1a9ec65630f.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": " is element-wise multiplication and " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{y,0}^{e}" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": " is the feature of step description (normal type) for step " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": ". The intuition of using normal type feature is that error frames often contain multiple sub-steps, including both normal and error ones (e.g., an actor opens a tortilla bag, grabs a tortilla from it, and accidentally drops it). A single error type feature may be unable to effectively match the corresponding error frames, especially when the number of frames is large and contain multiple sub-steps which the error description does not describe. Therefore, our score function computes the correlation between " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{y,0}^{e}, x_{y,m}^{e}" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "x_{j}^{v}" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": ". Thus, it jointly considers the similarities of error frames with respect to both normal and error type features. We validate its effectiveness for " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": " in Section 5. Finally, we apply a maximum filter on each error type of " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": " to smoothen the predictions, followed by argmax to obtain the error type predictions " + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10013" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 70, + 542, + 205 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 542, + 205 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 542, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 542, + 205 + ], + "type": "table", + "html": "
MethodQuesadillaOatmealEgoPER
F1@.5AccEditF1@.5AccEditPinwheelCoffeeTeaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid57.053.760.068.874.576.849.152.753.759.157.175.061.156.169.459.058.870.0
EgoPED56.961.447.362.154.870.349.854.741.053.959.655.469.464.558.356.459.054.5
GTG2Vid68.164.569.483.378.588.661.358.460.963.962.573.771.262.676.969.665.373.9
MethodHot ChocolateSandwichBurritosRamenRaitaAll
F1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEditF1@.5AccEdit
G2Vid17.119.223.821.321.739.833.427.542.928.033.950.518.523.126.123.725.136.6
EgoPED5.021.713.313.122.518.221.131.821.815.737.325.422.631.323.015.528.920.3
GTG2Vid20.843.527.827.238.538.836.634.041.331.333.642.922.130.826.127.636.135.4
", + "image_path": "0c2cedf2080c3baa2350295406cee1c50cbcba29aac1f98ed2a5fcb14aeed9f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 215, + 208, + 394, + 218 + ], + "lines": [ + { + "bbox": [ + 215, + 208, + 394, + 218 + ], + "spans": [ + { + "bbox": [ + 215, + 208, + 394, + 218 + ], + "type": "text", + "content": "Table 2. TAS results on EgoPER and CaptainCook4D." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 232, + 137, + 245 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 137, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 137, + 245 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 253, + 171, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 253, + 171, + 266 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 171, + 266 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 276, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 540 + ], + "type": "text", + "content": "Dataset. We evaluate our proposed method on EgoPER [41] and CaptainCook4D [66]. EgoPER consists of 5 tasks with 386 egocentric videos and contains 5 types of errors, including omission, addition, modification, slip, and correction. Captaincook4D has 24 tasks with 94.5 hours and contains 8 types of errors, including preparation, measurement, technique, timing, temperature, missing steps, ordering, and other. For EgoPER, we evaluate our method on all 5 tasks and error types with the same training and testing split in [41]. For CaptainCook4D, to have enough videos for training and testing, we select 5 of the tasks which have at least 5 normal and error videos: Spiced Hot Chocolate (Hot Chocolate), Microwave Egg Sandwich (Sandwich), Breakfast Burritos (Burritos), Ramen, and Cucumber Raita (Raita). We assign all normal videos to the training set and error videos to test set. To fit our setting for recognizing execution errors, we report the following error types in CaptainCook4D: preparation (Prep.), measurement (Mea.), technique (Tec.), timing (Time), and temperature (Temp). See the supplementary materials for breakdown analysis of EgoPER and CaptainCook4D and their detailed training and testing splits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We report results on " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". In our experiments, we report the F1 score " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "(F1@\\beta)" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " with an overlapping threshold " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". We compute the F1 score by averaging (if not specified) over the F1 scores that are independently computed on the segments for each step or error type. For " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", we report Error Accuracy (EAcc) and weighted " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "F1@0" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " score " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "(w-F1@0)" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " is challenging under the setting of no error video during training, we set " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " to 0, meaning that a segment is true positive whenever any frame in the segment is correctly classified. EAcc is the percentage of error types that a model can recognize; that is, the error types with " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "F1@0" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " score larger than 0. For " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "w-F1@0" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", we first compute " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "F1@0" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": " on the error type segments, excluding normal ones " + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "inline_equation", + "content": "(m=0)" + }, + { + "bbox": [ + 55, + 545, + 296, + 714 + ], + "type": "text", + "content": ", to alleviate the effect of data imbalance (e.g.," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "text", + "content": "nearly " + }, + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "inline_equation", + "content": "90\\%" + }, + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "text", + "content": " of the segments in EgoPER are normal). We then compute w-F1@0, where w-F1@0 = " + }, + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "inline_equation", + "content": "\\frac{\\mathrm{F1}@\\mathrm{0}}{\\mathrm{M} - \\mathrm{EAcc} \\times \\mathrm{M} + 1}" + }, + { + "bbox": [ + 313, + 232, + 555, + 365 + ], + "type": "text", + "content": ". w-F1@0 penalizes the model that fails to recognize all error types. For TAS, we report frame-wise Accuracy (Acc), segment-wise Edit score (Edit) on normal steps, and F1@.5 score. For ED, we report F1@.5 score on all segments. We also report F1@.5 score on normal (N.) and error segments (E.) separately for alleviating data imbalance. For omission error detection, we use Omission Accuracy (O-Acc), predicting the accuracy for omission steps, and Omission Intersection over Union (O-IoU), as in [41]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": "Baselines. Given the lack of baselines for joint TAS, " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ", we report the performance of EgoPED (MV), a modified version of EgoPED [41] that can additionally perform " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ". EgoPED (MV) computes the cosine similarities between frames and error types using the error type features generated in Section 4.2. Then it assigns frames to the type with the highest similarity and uses the majority voting (MV) strategy in [41] over frames within a segment to decide the final error type for each segment. EgoPED (MV) detects error segments with step thresholds equal to mean similarities over videos in the validation set. For " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ", since both datasets have data imbalance, we report the performance of Naive Predictor as another baseline. Naive Predictor classifies frames into the same error type for " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": " and into normal frames for " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ". Specifically, it classifies the frames into modification errors for EgoPER and preparation errors for CaptainCook4D. For " + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 313, + 371, + 556, + 612 + ], + "type": "text", + "content": ", we compare with G2Vid [19] to show how G2Vid performs with standard task graphs. To run G2Vid with our cost matrix, we remove all the background entries." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Implementation Details. We employ GPT-4o mini [1] as our LLM for generating error descriptions and VideoCLIP [93] pre-trained on Ego4D [30] as our VLM for extracting both visual and textual features for videos and error type descriptions. For each task in both EgoPER and CaptainCook4D, we generate at least 2 descriptions for each error type of each action. We set the step description as the description of normal type. We use the backbone in [52] as" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "10014" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 72, + 70, + 537, + 208 + ], + "blocks": [ + { + "bbox": [ + 72, + 70, + 537, + 208 + ], + "lines": [ + { + "bbox": [ + 72, + 70, + 537, + 208 + ], + "spans": [ + { + "bbox": [ + 72, + 70, + 537, + 208 + ], + "type": "table", + "html": "
MethodEgoPER
QuesadillaOatmealPinwheelCoffeeTeaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor14.40.07.226.10.013.123.40.011.744.10.022.14.80.02.422.60.011.3
EgoPED30.125.227.729.312.921.114.624.519.69.58.99.236.334.035.124.021.122.5
GTG2Vid40.422.631.558.536.447.430.616.023.333.14.018.651.537.044.242.823.233.0
CaptainCook4D
Hot ChocolateSandwichBurritosRamenRaitaAll
N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5N.E.F1@.5
Naive Predictor8.20.04.16.50.03.23.80.01.94.10.02.03.10.01.65.10.02.6
EgoPED8.12.45.216.36.811.516.38.212.212.410.711.614.19.511.813.47.510.5
GTG2Vid15.820.918.316.712.814.816.919.718.322.68.215.418.29.914.018.014.316.2
", + "image_path": "91a573e02fa094b6ecdc9ad205759c111d62f3592a2bb2957240a899753c8856.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": "our action segmentation model. We train the model using Adam optimizer with learning rate 0.0001 for 3000 iterations. For " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": ", if " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\hat{y}_j = -1" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\hat{y}_j^+" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " is a background class, we directly assign an addition error and preparation error to " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\hat{m}_j" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " for EgoPER and CaptainCook4D, respectively, and assign the corresponding similarity to 1. For " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": ", we produce " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " predictions based on " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{E}}" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " by classifying different error types into a single error type. In practice, we set " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\tau = S" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": " is the number of nodes in GTG and " + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "inline_equation", + "content": "\\eta = 0" + }, + { + "bbox": [ + 55, + 234, + 296, + 342 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 347, + 179, + 359 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 179, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 179, + 359 + ], + "type": "text", + "content": "5.2. Experimental Results" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": "Error Recognition " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "(ER)" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ". Table 1 compares " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": " performance between different methods on EgoPER and CaptainCook4D. Our proposed framework, GTG2Vid, demonstrates the best recognition ability than Naive Predictor and EgoPED (MV). Specifically, our proposed method achieves " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "31.7\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "31.3\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "17.8\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "22.1\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": " on w-F1@0 compared to " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "7.5\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "5.5\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "7.4\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "inline_equation", + "content": "4.4\\%" + }, + { + "bbox": [ + 55, + 366, + 296, + 569 + ], + "type": "text", + "content": " by EgoPED (MV) for quesadilla, oatmeal, pinwheel, and tea. On the other hand, our proposed method outperforms EgoPED (MV) on w-F1@0 for all tasks of CaptainCook4D. In addition, our proposed method achieves overall higher EAcc than EgoPED (MV), indicating that our method can recognize more error types on both datasets. Notice that our performance on w-F1@0 for coffee is worse than EgoPED (MV) because EgoPED predicts most of the segments as errors, shown in Table 3. Please see supplementary materials for the F1@0 score of each type of error." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Next, we investigate the effectiveness of our ERM and our proposed score function (Eq. 6). By replacing ERM with MV, the performance drops significantly (more than " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " on w-F1@0 across all tasks of EgoPER). On the other hand, our method, GTG2Vid (w/o n.f.), without using normal type features suffers a drop on EAcc for most of the tasks, indicating that our proposed function for computing joint similarity has a better ability to recognize different error types. Furthermore, Table 4 shows the Top-1 and Top-2 accuracy of different scoring methods given the features of ground-truth error segments (average over frame features) and generated error type features. Our method gains a per" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 318, + 231, + 549, + 292 + ], + "blocks": [ + { + "bbox": [ + 78, + 209, + 531, + 220 + ], + "lines": [ + { + "bbox": [ + 78, + 209, + 531, + 220 + ], + "spans": [ + { + "bbox": [ + 78, + 209, + 531, + 220 + ], + "type": "text", + "content": "Table 3. ED results of different methods on EgoPER and CaptainCook4D. N. and E. denote the F1@.5 score for normal and error segments." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 231, + 549, + 292 + ], + "lines": [ + { + "bbox": [ + 318, + 231, + 549, + 292 + ], + "spans": [ + { + "bbox": [ + 318, + 231, + 549, + 292 + ], + "type": "table", + "html": "
MethodQuesadillaPinwheelTea
Top-1Top-2Top-1Top-2Top-1Top-2
Uniform25.050.025.050.025.050.0
Cos. Sim.42.665.827.748.427.845.3
ERM54.681.538.460.334.353.7
", + "image_path": "f5ba31dd5d4a7222eabc2307c6b076a98fc187f5dc7f13e326cbc2c4ad76b5c5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 293, + 553, + 314 + ], + "lines": [ + { + "bbox": [ + 313, + 293, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 553, + 314 + ], + "type": "text", + "content": "Table 4. Performance of feature matching methods on EgoPER. Top-1 and Top-2 denote the top-1 and top-2 accuracy over all error segments." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "content": "formance boost on Top-1 and Top-2 accuracy. Specifically, we obtain " + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "inline_equation", + "content": "10.7\\%" + }, + { + "bbox": [ + 313, + 329, + 553, + 376 + ], + "type": "text", + "content": " improvement on Top-1 accuracy compared to cosine similarity (Cos. Sim.) for pinwheel, which verifies our claim in Section 4.3." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": "Temporal Action Segmentation (TAS). For TAS performance shown in Table 2, our method obtains a " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "13.2\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "6.3\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "19.4\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": " higher F1@.5, Acc and Edit score than EgoPED on All of EgoPER. For CaptainCook4D, our method obtains a " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "12.1\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "7.2\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "inline_equation", + "content": "15.1\\%" + }, + { + "bbox": [ + 313, + 381, + 555, + 536 + ], + "type": "text", + "content": " higher F1, Acc, and Edit score than EgoPED on All. Furthermore, our method has more robust TAS performance compared to EgoPED, which suffers a significant performance drop in Hot Chocolate and Sandwich. Notice that our Edit score on All is comparable to G2Vid since there are not many omission errors in test videos. Still, our method achieves a higher F1@.5 score than G2Vid, which cannot distinguish between background frames and errors." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": "Error Detection (ED). Table 3 shows the " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " performance of different methods on EgoPER. Our method achieves " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "33.0\\%" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " higher on F1@.5 score, compared to " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "22.5\\%" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " by EgoPED for All in EgoPER. Although EgoPED can obtain a higher F1 score on error segments for quesadilla, pinwheel, and coffee, its detection ability for normal segments decreases significantly, specifically a " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "23.6\\%" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " drop on F1@.5 score for coffee compared to our method. Therefore, compared to EgoPED, our method achieves a better balance between normal and error segments. For All in CaptainCook4D, GTG2Vid outperforms EgoPED, achieving " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "16.2\\%" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " on F1@.5 compared to " + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "inline_equation", + "content": "10.5\\%" + }, + { + "bbox": [ + 313, + 541, + 554, + 685 + ], + "type": "text", + "content": " by EgoPED." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Ablation of Node and Frame Dropping Term. We show ablation studies and investigate the effectiveness of our pro" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10015" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 75, + 529, + 203 + ], + "blocks": [ + { + "bbox": [ + 85, + 75, + 529, + 203 + ], + "lines": [ + { + "bbox": [ + 85, + 75, + 529, + 203 + ], + "spans": [ + { + "bbox": [ + 85, + 75, + 529, + 203 + ], + "type": "image", + "image_path": "b5349db5d35961d469daadffdba256d69d20f1c92258ad904b8481d8cf19c743.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "lines": [ + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 55, + 205, + 553, + 225 + ], + "type": "text", + "content": "Figure 3. Qualitative results for quesadilla, tea, and oatmeal in EgoPER. The top part shows the results of TAS for ground-truths, EgoPED, and our method. The bottom part shows ER results for ground-truths, EgoPED, and our method. Seg. and E.T. denote action segmentation and error type, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 235, + 292, + 345 + ], + "blocks": [ + { + "bbox": [ + 60, + 235, + 292, + 345 + ], + "lines": [ + { + "bbox": [ + 60, + 235, + 292, + 345 + ], + "spans": [ + { + "bbox": [ + 60, + 235, + 292, + 345 + ], + "type": "image", + "image_path": "2e4dd901808e175f5aab560cb9062a1b11d4cccf2aab8bd3449d25e9b86292cc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 347, + 285, + 357 + ], + "lines": [ + { + "bbox": [ + 65, + 347, + 285, + 357 + ], + "spans": [ + { + "bbox": [ + 65, + 347, + 285, + 357 + ], + "type": "text", + "content": "Figure 4. Ablation of our proposed node and frame dropping term." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": "posed node and frame dropping term in GTG2Vid. The two upper plots in Figure 4 show F1@.5 scores of our proposed frame dropping term (Ours) with different " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " and a naive dropping cost (G2Vid) [19] with different " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " top percentile of the values in the cost matrix, for pinwheel and quesadilla. Our method achieves " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "26.0\\%" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "33.8\\%" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " on F1@.5 score compared to " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "23.2\\%" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "30.9\\%" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " by G2Vid. On the other hand, the two bottom plots in Figure 4 show the F1 scores for TAS with different " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " for our node dropping term. Smaller " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " indicates GTG2Vid drops a fewer number of nodes. When " + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 54, + 371, + 295, + 526 + ], + "type": "text", + "content": " is larger or equal to 0, GTG2Vid starts dropping nodes, improving TAS performance on videos with omission steps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "text", + "content": "In addition, Figure 5 shows the performance of omission detection for different methods on EgoPER. First, since G2Vid is unable to drop nodes of a graph, it cannot predict any omitted steps, resulting in " + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "inline_equation", + "content": "0\\%" + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "text", + "content": " on O-Acc and O-IoU. Second, although our method suffers a performance drop on O-IoU (" + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "inline_equation", + "content": "5.7\\%" + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "text", + "content": " lower than EgoPED on All), we gain a performance boost on O-Acc (" + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 527, + 295, + 694 + ], + "type": "text", + "content": " higher than EgoPED on All), indicating that our method successfully drops nodes of a task graph according to the input videos and has a stronger capability for detecting omitted steps. To conclude, our proposed method uses 1) dynamic frame dropping term to precisely drop error segments and 2) dynamic node dropping term to drop unnecessary nodes of task graphs for better TAS and omission detection performance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 294, + 714 + ], + "type": "text", + "content": "Qualitative Analysis. We visualize ER results in Fig. 3." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 236, + 552, + 366 + ], + "blocks": [ + { + "bbox": [ + 316, + 236, + 552, + 366 + ], + "lines": [ + { + "bbox": [ + 316, + 236, + 552, + 366 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 552, + 366 + ], + "type": "image", + "image_path": "4d4438e99f1859d747038528fb7be64a76fe3a7a09c788b086040ab3a87a311a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 373, + 550, + 384 + ], + "lines": [ + { + "bbox": [ + 318, + 373, + 550, + 384 + ], + "spans": [ + { + "bbox": [ + 318, + 373, + 550, + 384 + ], + "type": "text", + "content": "Figure 5. Omission detection results of different methods on EgoPER." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 555, + 486 + ], + "type": "text", + "content": "Our method can recognize different types of errors, e.g., addition errors (red) on tea, addition and correction (orange) errors on quesadilla, and modification errors (blue) on oatmeal. Our method obtains more accurate TAS, ED, and ER results for error videos on quesadilla, tea and oatmeal compared to EgoPED, thanks to dynamic frame and node dropping term in GTG2Vid and our proposed ERM. See supplementary materials for more qualitative visualization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 496, + 391, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 496, + 391, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 496, + 391, + 508 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": "We studied joint " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": " in procedural task videos. Our proposed framework consists of 1) generalized task graphs to capture correct steps and background, 2) a GTG-Video Alignment algorithm to find the optimal alignment between nodes in GTG and video frames with dynamic dropping terms, 3) a procedure for generating error descriptions for each error type of every step using LLMs and the corresponding features using VLMs, and 4) an Error Recognition Module to effectively perform " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": " using joint similarities. Our experiments on two datasets showed that our proposed method can obtain promising results on " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "ER" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "ED" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "inline_equation", + "content": "TAS" + }, + { + "bbox": [ + 313, + 515, + 555, + 647 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "spans": [ + { + "bbox": [ + 313, + 658, + 411, + 671 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "This work was funded, in part, by ARPA-H (1AY2AX000062), DARPA PTG (HR00112220001), NSF (IIS-2115110), ONR (N000142512287) and ARO" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10016" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 120 + ], + "type": "text", + "content": "(W911NF2110276). The views and conclusions contained in this document are those of the authors and should not be interpreted as representing the official policies, either expressed or implied, of the US Government." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 115, + 142 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 149, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 66, + 149, + 296, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 149, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 66, + 149, + 296, + 204 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 205, + 296, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 205, + 296, + 270 + ], + "spans": [ + { + "bbox": [ + 66, + 205, + 296, + 270 + ], + "type": "text", + "content": "[2] Andra Acsintoae, Andrei Florescu, Mariana-Iuliana Georgescu, Tudor Mare, Paul Sumedrea, Radu Tudor Ionescu, Fahad Shahbaz Khan, , and Mubarak Shah. Ubnormal: New benchmark for supervised open-set video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 66, + 271, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 271, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 66, + 271, + 295, + 315 + ], + "type": "text", + "content": "[3] Hyemin Ahn and Dongheui Lee. Refining action segmentation with hierarchical video representations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 16302-16310, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 66, + 316, + 295, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 316, + 295, + 358 + ], + "spans": [ + { + "bbox": [ + 66, + 316, + 295, + 358 + ], + "type": "text", + "content": "[4] J. B. Alayrac, P. Bojanowski, N. Agrawal, J. Sivic, I. Laptev, and S. Lacoste-Julien. Unsupervised learning from narrated instruction videos. IEEE Conference on Computer Vision and Pattern Recognition, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 360, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 360, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 66, + 360, + 295, + 403 + ], + "type": "text", + "content": "[5] Kumar Ashutosh, Santhosh Kumar Ramakrishnan, Triantafyllos Afouras, and Kristen Grauman. Video-mined task graphs for keystep recognition in instructional videos. Neural Information Processing Systems, 2023. 1, 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 66, + 404, + 295, + 436 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 404, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 66, + 404, + 295, + 436 + ], + "type": "text", + "content": "[6] Marcella Astrid, Muhammad Zaigham Zaheer, Jae-Yeong Lee, and Seung-Ik Lee. Learning not to reconstruct anomalies. arXiv: 2110.09742, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 66, + 437, + 295, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 437, + 295, + 480 + ], + "spans": [ + { + "bbox": [ + 66, + 437, + 295, + 480 + ], + "type": "text", + "content": "[7] Nadine Behrmann, S. Alireza Golestaneh, Zico Kolter, Juergen Gall, and Mehdi Noroozi. Unified fully and timestamp supervised temporal action segmentation via sequence to sequence translation. In ECCV, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 66, + 482, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 482, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 66, + 482, + 295, + 525 + ], + "type": "text", + "content": "[8] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In Proceedings of the International Conference on Machine Learning (ICML), 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 66, + 526, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 526, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 66, + 526, + 295, + 590 + ], + "type": "text", + "content": "[9] Daniele Di Mauro Mario Valerio Giuffrida Giovanni Maria Farinella Camillo Quattrocchi, Antonino Furnari. Synchronization is all you need: Exocentric-to-egocentric transfer for temporal action segmentation with unlabeled synchronized video pairs. European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 592, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 592, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 592, + 295, + 634 + ], + "type": "text", + "content": "[10] J. Carreira and A. Zisserman. Quo vadis, action recognition? a new model and the kinetics dataset. In IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 636, + 295, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 295, + 690 + ], + "type": "text", + "content": "[11] Chien-Yi Chang, De-An Huang, Yanan Sui, Li Fei-Fei, and Juan Carlos Niebles. D3tw: Discriminative differentiable dynamic time warping for weakly supervised action alignment and segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 692, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 692, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 62, + 692, + 295, + 714 + ], + "type": "text", + "content": "[12] Chien-Yi Chang, De-An Huang, Danfei Xu, Ehsan Adeli, Li Fei-Fei, and Juan Carlos Niebles. Procedure planning" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 339, + 73, + 553, + 106 + ], + "type": "text", + "content": "in instructional videos. In Computer Vision - ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XI, 2020. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 321, + 107, + 553, + 150 + ], + "type": "text", + "content": "[13] Hanqiu Deng, Zhaoxiang Zhang, Shihao Zou, and Xingyu Li. Bi-directional frame interpolation for unsupervised video anomaly detection. IEEE Winter Conference on Applications of Computer Vision, 2023. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 152, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 152, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 320, + 152, + 553, + 184 + ], + "type": "text", + "content": "[14] Guodong Ding, Fadime Sener, Shugao Ma, and Angela Yao. Every mistake counts in assembly. arXiv: 2307.16453, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "spans": [ + { + "bbox": [ + 321, + 186, + 553, + 229 + ], + "type": "text", + "content": "[15] Guodong Ding, Hans Golong, and Angela Yao. Coherent temporal synthesis for incremental action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 321, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Li Ding and Chenliang Xu. Weakly-supervised action segmentation with iterative soft boundary assignment. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] G. Donahue and E. Elhamifar. Learning to predict activity progress by self-supervised video alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 553, + 364 + ], + "type": "text", + "content": "[18] Nikita Dvornik, Isma Hadji, Konstantinos G Derpanis, Animesh Garg, and Allan D Jepson. Drop-dtw: Aligning common signal between sequences while dropping outliers. In NeurIPS, 2021. 1, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 365, + 553, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 365, + 553, + 421 + ], + "spans": [ + { + "bbox": [ + 321, + 365, + 553, + 421 + ], + "type": "text", + "content": "[19] Nikita Dvornik, Isma Hadji, Hai Pham, Dhaivat Bhatt, Brais Martinez, Afsaneh Fazly, and Allan D Jepson. Flow graph to video grounding for weakly-supervised multi-step localization. In European Conference on Computer Vision, pages 319-335. Springer, 2022. 1, 2, 3, 4, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 422, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 476 + ], + "type": "text", + "content": "[20] Nikita Dvornik, Isma Hadji, Ran Zhang, Konstantinos Derpanis, Animesh Garg, Richard Wildes, and Allan Jepson. Stepformer: Self-supervised step discovery and localization in instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2023. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 478, + 553, + 510 + ], + "type": "text", + "content": "[21] E. Elhamifar and D. Huynh. Self-supervised multi-task procedure learning from instructional videos. European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 321, + 512, + 553, + 544 + ], + "type": "text", + "content": "[22] E. Elhamifar and Z. Naing. Unsupervised procedure learning via joint dynamic summarization. International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 545, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 545, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 321, + 545, + 553, + 599 + ], + "type": "text", + "content": "[23] Yazan Abu Farha and Jurgen Gall. Ms-tcn: Multi-stage temporal convolutional network for action segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3575-3584, 2019. 1, 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 321, + 601, + 553, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 601, + 553, + 644 + ], + "spans": [ + { + "bbox": [ + 321, + 601, + 553, + 644 + ], + "type": "text", + "content": "[24] Mohsen Fayyaz and Jurgen Gall. Sect: Set constrained temporal transformer for set supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 321, + 647, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 647, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 321, + 647, + 553, + 679 + ], + "type": "text", + "content": "[25] C. Feichtenhofer. X3d: Expanding architectures for efficient video recognition. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 713 + ], + "type": "text", + "content": "[26] C. Feichtenhofer, H. Fan, J. Malik, and K. He. Slowfast networks for video recognition. In 2019 IEEE/CVF International Conference on Computer Vision (ICCV), 2019. 1" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10017" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 137 + ], + "type": "text", + "content": "[27] Alessandro Flaborea, Guido Melendugno, Leonardo Pliniq, Luca Scofanoq, Edoardo Matteisq, Antonino Furnari, Giovanni Farinella, and Fabio Galasso. Prego: online mistake detection in procedural egocentric videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 294, + 194 + ], + "type": "text", + "content": "[28] Daniel Fried, Jean-Baptiste Alayrac, Phil Blunsom, Chris Dyer, Stephen Clark, and Aida Nematzadeh. Learning to segment actions from observation and narration. Annual Meeting of the Association for Computational Linguistics, 2020. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 195, + 295, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 195, + 295, + 250 + ], + "spans": [ + { + "bbox": [ + 62, + 195, + 295, + 250 + ], + "type": "text", + "content": "[29] Reza Ghoddoosian, Isht Dwivedi, Nakul Agarwal, and Behzad Dariush. Weakly-supervised action segmentation and unseen error detection in anomalous instructional videos. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10128-10138, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 251, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 251, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 251, + 294, + 557 + ], + "type": "text", + "content": "[30] Kristen Grauman, Andrew Westbury, Eugene Byrne, Zachary Chavis, Antonino Furnari, Rohit Girdhar, Jackson Hamburger, Hao Jiang, Miao Liu, Xingyu Liu, Miguel Martin, Tushar Nagarajan, Iija Radosavovic, Santhosh K. Ramakrishnan, Fiona Ryan, Jayant Sharma, Michael Wray, Mengmeng Xu, Eric Z. Xu, Chen Zhao, Siddhant Bansal, Dhruv Batra, Vincent Cartillier, Sean Crane, Tien Do, Morrie Doulaty, Akshay Erapalli, Christoph Feichtenhofer, Adriano Fragomeni, Qichen Fu, Christian Fuegen, Abraham Gebreselasie, Cristina Gonzalez, James M. Hillis, Xuhua Huang, Yifei Huang, Wenqi Jia, Weslie Khoo, Jachym Kolár, Satwik Kottur, Anurag Kumar, Federico Landini, Chao Li, Yanghao Li, Zhenqiang Li, Karttikeya Mangalam, Raghava Modhugu, Jonathan Munro, Tullie Murrell, Takumi Nishiyasu, Will Price, Paola Ruiz Puentes, Merey Ramazanova, Leda Sari, Kiran K. Somasundaram, Audrey Southerland, Yusuke Sugano, Ruijie Tao, Minh Vo, Yuchen Wang, Xindi Wu, Takuma Yagi, Yunyi Zhu, Pablo Arbeláez, David J. Crandall, Dima Damen, Giovanni Maria Farinella, Bernard Ghanem, Vamsi Krishna Ithapu, C. V. Jawahar, Hanbyul Joo, Kris Kitani, Haizhou Li, Richard A. Newcombe, Aude Oliva, Hyun Soo Park, James M. Rehg, Yoichi Sato, Jianbo Shi, Mike Zheng Shou, Antonio Torralba, Lorenzo Torresani, Mingfei Yan, and Jitendra Malik. Ego4d: Around the world in 3,000 hours of egocentric video. 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18973-18990, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 602 + ], + "type": "text", + "content": "[31] Yifei Huang, Yusuke Sugano, and Yoichi Sato. Improving action segmentation via graph-based temporal reasoning. In 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 658 + ], + "type": "text", + "content": "[32] Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka. Alleviating over-segmentation errors by detecting action boundaries. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 2322-2331, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 712 + ], + "type": "text", + "content": "[33] Mohaiminul Islam, Tushar Nagarajan, Huiyu Wang, Fu Jen Chu, Kris Kitani, Gedas Bertasius, and Xitong Yang. Propose, assess, search: Harnessing llms for goal-oriented planning in instructional videos. European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 553, + 116 + ], + "type": "text", + "content": "[34] Jaehyun Kim, Seongwook Yoon, Taehyeon Choi, and Sanghoon Sull. Unsupervised video anomaly detection based on similarity with predefined text descriptions. Sensors, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 320, + 118, + 553, + 161 + ], + "type": "text", + "content": "[35] H. Kuehne, J. Gall, and T. Serre. An end-to-end generative framework for video segmentation and recognition. IEEE Winter Conference on Applications of Computer Vision, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 320, + 163, + 553, + 206 + ], + "type": "text", + "content": "[36] Anna Kukleva, Hilde Kuehne, Fadime Sener, and Jurgen Gall. Unsupervised learning of action classes with continuous temporal embedding. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 320, + 209, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 209, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 320, + 209, + 553, + 274 + ], + "type": "text", + "content": "[37] Sateesh Kumar, Sanjay Haresh, Awais Ahmed, Andrey Konin, M. Zeeshan Zia, and Quoc-Huy Tran. Unsupervised action segmentation by joint representation learning and online clustering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 20174-20185, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 276, + 553, + 319 + ], + "type": "text", + "content": "[38] C. Lea, M. D. Flynn, R. Vidal, A. Reiter, and G. D. Hager. Temporal convolutional networks for action segmentation and detection. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 321, + 553, + 373 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 321, + 553, + 373 + ], + "spans": [ + { + "bbox": [ + 321, + 321, + 553, + 373 + ], + "type": "text", + "content": "[39] Jungbeom Lee, Jihun Yi, Chaehun Shin, and Sungroh Yoon. Bbam: Bounding box attribution map for weakly supervised semantic and instance segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 321, + 376, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 376, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 321, + 376, + 553, + 420 + ], + "type": "text", + "content": "[40] Sangmin Lee, Hak Gu Kim, and Yong Man Ro. Bman: Bidirectional multi-scale aggregation networks for abnormal event detection. IEEE Transactions on Image Processing, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 422, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 422, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 321, + 422, + 553, + 464 + ], + "type": "text", + "content": "[41] S. Lee, Z. Lu, Z. Zhang, M. Hoai, and E. Elhamifar. Error detection in egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 467, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 467, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 321, + 467, + 553, + 499 + ], + "type": "text", + "content": "[42] Jun Li and Sinisa Todorovic. Set-constrained viterbi for set-supervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 501, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 501, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 321, + 501, + 553, + 533 + ], + "type": "text", + "content": "[43] J. Li and S. Todorovic. Anchor-constrained viterbi for set-supervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 535, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 535, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 535, + 553, + 567 + ], + "type": "text", + "content": "[44] J. Li, P. Lei, and S. Todorovic. Weakly supervised energy-based learning for action segmentation. International Conference on Computer Vision, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 321, + 568, + 553, + 621 + ], + "type": "text", + "content": "[45] M. Li, L. Chen, Y. Duarr, Z. Hu, J. Feng, J. Zhou, and J. Lu. Bridge-prompt: Towards ordinal action understanding in instructional videos. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 553, + 678 + ], + "type": "text", + "content": "[46] Shi-Jie Li, Yazan AbuFarha, Yun Liu, Ming-Ming Cheng, and Juergen Gall. Ms-tcn++: Multi-stage temporal convolutional network for action segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, pages 1-1, 2020. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 321, + 681, + 553, + 712 + ], + "type": "text", + "content": "[47] Yanghao Li, Chao-Yuan Wu, Haoqi Fan, Karttikeya Mangalam, Bo Xiong, Jitendra Malik, and Christoph Feichtenhofer. Mvitv2: Improved multiscale vision transformers" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10018" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 81, + 72, + 294, + 105 + ], + "type": "text", + "content": "for classification and detection. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 61, + 107, + 295, + 150 + ], + "type": "text", + "content": "[48] Zhe Li, Yazan Abu Farha, and Jurgen Gall. Temporal action segmentation from timestamp supervision. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 152, + 294, + 195 + ], + "type": "text", + "content": "[49] Zeqian Li, Qirui Chen, Tengda Han, Ya Zhan, Yanfeng Wang, and Weidi Xie. Multi-sentence grounding for long-term instructional video. European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 198, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 198, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 62, + 198, + 295, + 239 + ], + "type": "text", + "content": "[50] Ji Lin, Chuang Gan, and Song Han. Tsm: Temporal shift module for efficient video understanding. In Proceedings of the IEEE International Conference on Computer Vision, 2019. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 294, + 275 + ], + "type": "text", + "content": "[51] Daochang Liu, Qiyue Li, AnhDung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. arXiv preprint arXiv:2303.17959, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 277, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 277, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 277, + 294, + 308 + ], + "type": "text", + "content": "[52] Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu. Diffusion action segmentation. 2023. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 311, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 311, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 62, + 311, + 294, + 365 + ], + "type": "text", + "content": "[53] Yunze Liu, Yun Liu, Che Jiang, Kangbo Lyu, Weikang Wan, Hao Shen, Boqiang Liang, Zhoujie Fu, He Wang, and Li Yi. Hoi4d: A 4d egocentric dataset for category-level human-object interaction. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 366, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 294, + 431 + ], + "type": "text", + "content": "[54] Yang Liu, Jiayu Huo, Jingjing Peng, Rachel Sparks, Prokar Dasgupta, Alejandro Granados, and Sebastien Ourselin. Skit: a fast key information video transformer for online surgical phase recognition. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 21074-21084, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 294, + 476 + ], + "type": "text", + "content": "[55] Z. Lu and E. Elhamifar. Weakly-supervised action segmentation and alignment via transcript-aware union-of-subspaces learning. International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 521 + ], + "type": "text", + "content": "[56] Z. Lu and E. Elhamifar. Set-supervised action learning in procedural task videos via pairwise order consistency. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 524, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 524, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 62, + 524, + 294, + 566 + ], + "type": "text", + "content": "[57] Z. Lu and E. Elhamifar. Fact: Frame-action cross-attention temporal modeling for efficient action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 294, + 601 + ], + "type": "text", + "content": "[58] Z. Lu and E. Elhamifar. Multi-modal few-shot temporal action segmentation. International Conference on Computer Vision, 2025. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "type": "text", + "content": "[59] Z. Lu, A. Iftekhar, G. Mittal, T. Meng, X. Wang, C. Zhao, R. Kukkala, E. Elhamifar, and M. Chen. Decafnet: Delegate and conquer for efficient temporal grounding in long videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 658, + 294, + 713 + ], + "type": "text", + "content": "[60] Antonino Furnari Luigi Seminara, Giovanni Maria Farinella. Differentiable task graph learning: Procedural activity representation and online mistake detection from egocentric videos. Neural Information Processing Systems, 2024. 1, 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 320, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 320, + 73, + 553, + 116 + ], + "type": "text", + "content": "[61] K. Mangalam, H. Fan, Y. Li, C. Wu, B. Xiong, C. Feichtenhofer, and J. Malik. Reversible vision transformers. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 119, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 119, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 320, + 119, + 553, + 162 + ], + "type": "text", + "content": "[62] A. Miech, J-B. Alayrac, L. Smaira, I. Laptev, J. Sivic, and A. Zisserman. End-to-end learning of visual representations from uncurated instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 320, + 164, + 553, + 196 + ], + "type": "text", + "content": "[63] Fangzhou Mu, Sicheng Mo, and Yin Li. Snag: Scalable and accurate video grounding. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 320, + 198, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 198, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 320, + 198, + 553, + 230 + ], + "type": "text", + "content": "[64] Tushar Nagarajan and Lorenzo Torresani. Step differences in instructional video. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 320, + 232, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 232, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 320, + 232, + 553, + 297 + ], + "type": "text", + "content": "[65] Kumaranage Ravindu Yasas Nagasinghe, Honglu Zhou, Malitha Gunawardhana, Martin Renqiang Min, Daniel Harari, and Muhammad Haris Khan. Why not use your textbook? knowledge-enhanced procedure planning of instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 320, + 299, + 553, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 299, + 553, + 363 + ], + "spans": [ + { + "bbox": [ + 320, + 299, + 553, + 363 + ], + "type": "text", + "content": "[66] Rohith Peddi, Shivvrat Arya, Bharath Challa, Likhitha Pallapothula, Akshay Vyas, Bhavya Gouripeddi, Jikai Wang, Qifan Zhang, Vasundhara Komaragiri, Eric Ragan, Nicholas Ruozzi, Yu Xiang, and Vibhav Gogate. CaptainCook4D: A Dataset for Understanding Errors in Procedural Activities, 2024. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 320, + 366, + 553, + 409 + ], + "type": "text", + "content": "[67] Didik Purwanto, Yie-Tarng Chen, and Wen-Hsien Fang. Dance with self-attention: A new look of conditional random fields on anomaly detection in videos. IEEE International Conference on Computer Vision, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 320, + 411, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 411, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 320, + 411, + 553, + 454 + ], + "type": "text", + "content": "[68] Yicheng Qian, Weixin Luo, Dongze Lian, Xu Tang, Peilin Zhao, and Shenghua Gao. Svip: Sequence verification for procedures in videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 320, + 456, + 553, + 510 + ], + "type": "text", + "content": "[69] Rahul Rahaman, Dipika Singhania, Alexandre Thiery, and Angela Yao. A generalized and robust framework for timestamp supervision in temporal action segmentation. In Computer Vision-ECCV 2022: 17th European Conference, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 320, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 320, + 513, + 553, + 555 + ], + "type": "text", + "content": "[70] A. Richard, H. Kuehne, and J. Gall. Action sets: Weakly supervised action segmentation without ordering constraints. IEEE Conference on Computer Vision and Pattern Recognition, 2018." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 320, + 557, + 553, + 601 + ], + "type": "text", + "content": "[71] A. Richard, H. Kuehne, A. Iqbal, and J. Gall. Neuralnetwork-viterbi: A framework for weakly supervised video learning. IEEE Conference on Computer Vision and Pattern Recognition, 2018. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 320, + 603, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 603, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 320, + 603, + 553, + 666 + ], + "type": "text", + "content": "[72] Nicolae-Catalin Ristea, Neelu Madan, Radu Tudor Ionescu, Kamal Nasrollahi, Fahad Shahbaz Khan, Thomas B. Moeslund, and Mubarak Shah. Self-supervised predictive convolutional attentive block for anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 320, + 670, + 553, + 713 + ], + "type": "text", + "content": "[73] M. Rohrbach, S. Amin, M. Andriluka, and B. Schiele. A database for fine grained activity detection of cooking activities. IEEE Conference on Computer Vision and Pattern Recognition, 2012. 1, 2" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10019" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 61, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 294, + 127 + ], + "type": "text", + "content": "[74] Saquib Sarfraz, Naila Murray, Vivek Sharma, Ali Diba, Luc Van Gool, and Rainer Stiefelhagen. Temporally-weighted hierarchical clustering for unsupervised action segmentation. IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 61, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 61, + 129, + 294, + 184 + ], + "type": "text", + "content": "[75] Fadime Sener, Dibyadip Chatterjee, Daniel Shelepop, Kun He, Dipika Singhania, Robert Wang, and Angela Yao. Assembly101: A large-scale multi-view video dataset for understanding procedural activities. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 295, + 228 + ], + "type": "text", + "content": "[76] Y. Shen and E. Elhamifar. Semi-weakly-supervised learning of complex actions from instructional task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "spans": [ + { + "bbox": [ + 62, + 230, + 295, + 274 + ], + "type": "text", + "content": "[77] Y. Shen and E. Elhamifar. Progress-aware online action segmentation for egocentric procedural task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1, 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 62, + 276, + 294, + 308 + ], + "type": "text", + "content": "[78] Y. Shen and E. Elhamifar. Understanding multi-task activities from single-task videos. IEEE Conference on Computer Vision and Pattern Recognition, 2025. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 62, + 310, + 294, + 353 + ], + "type": "text", + "content": "[79] Y. Shen, L. Wang, and E. Elhamifar. Learning to segment actions from visual and language instructions via differentiable weak sequence alignment. IEEE Conference on Computer Vision and Pattern Recognition, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 294, + 397 + ], + "type": "text", + "content": "[80] G. A. Sigurdsson, S. Divvala, A. Farhadi, and A. Gupta. Asynchronous temporal fields for action recognition. IEEE Conference on Computer Vision and Pattern Recognition, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 294, + 443 + ], + "type": "text", + "content": "[81] B. Singh, T. K. Marks, M. Jones, O. Tuzel, and M. Shao. A multi-stream bi-directional recurrent neural network for finegrained action detection. IEEE Conference on Computer Vision and Pattern Recognition, 2016." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 445, + 294, + 476 + ], + "type": "text", + "content": "[82] Dipika Singhania, Rahul Rahaman, and Angela Yao. Coarse to fine multi-resolution temporal convolutional network. CoRR, abs/2105.10859, 2021. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 294, + 501 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 294, + 501 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 294, + 501 + ], + "type": "text", + "content": "[83] Dean Slawson. The case for vr-immersive and ai-adaptive soft skills training. Training Industry Magazine, 2018. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 502, + 294, + 544 + ], + "type": "text", + "content": "[84] Yaser Souri, Mohsen Fayyaz, Luca Minciullo, Gianpiero Francesca, and Juergen Gall. Fast Weakly Supervised Action Segmentation Using Mutual Consistency. PAMI, 2021. 1, 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 294, + 601 + ], + "type": "text", + "content": "[85] Tomáš Souček, Dima Damen, Michael Wray, Ivan Laptev, and Josef Sivic. Genhowto: Learning to generate actions and state transformations from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 62, + 603, + 294, + 657 + ], + "type": "text", + "content": "[86] Yansong Tang, Dajun Ding, Yongming Rao, Yu Zheng, Danyang Zhang, Lili Zhao, Jiwen Lu, and Jie Zhou. Coin: A large-scale dataset for comprehensive instructional video analysis. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 295, + 713 + ], + "type": "text", + "content": "[87] Kamalakar Vijay Thakare, Yash Raghuwanshi, Debi Prosad Dogra, Heeseung Choi, and Ig-Jae Kim. Dyannet: A scene dynamicity guided self-trained video anomaly detection network. IEEE Winter Conference on Applications of Computer Vision, 2023. 2" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 321, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 321, + 73, + 555, + 139 + ], + "type": "text", + "content": "[88] Quoc-Huy Tran, Ahmed Mehmood, Muhammad Ahmed, Muhammad Naufil, Anas Zafar, Andrey Konin, and Zeeshan Zia. Permutation-aware activity segmentation via unsupervised frame-to-segment alignment. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 6426-6436, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 141, + 554, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 141, + 554, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 141, + 554, + 172 + ], + "type": "text", + "content": "[89] Mengmeng Wang, Jiazheng Xing, and Yong Liu. Actionclip: A new paradigm for video action recognition. CoRR, 2021. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 175, + 554, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 175, + 554, + 228 + ], + "spans": [ + { + "bbox": [ + 321, + 175, + 554, + 228 + ], + "type": "text", + "content": "[90] Xin Wang, Taein Kwon, Mahdi Radl Bowen Pan, Ishani Chakraborty, and Sean Andrist. Holoassist: an egocentric human interaction dataset for interactive ai assistants in the real world. IEEE International Conference on Computer Vision, 2023. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 321, + 231, + 554, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 231, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 321, + 231, + 554, + 296 + ], + "type": "text", + "content": "[91] Zhe Wang, Hao Chen, Xinyu Li, Chunhui Liu, Yuanjun Xiong, Joseph Tighe, and Charless Fowlkes. Sscap: Self-supervised co-occurrence action parsing for unsupervised temporal action segmentation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1819–1828, 2022. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 298, + 554, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 298, + 554, + 341 + ], + "spans": [ + { + "bbox": [ + 321, + 298, + 554, + 341 + ], + "type": "text", + "content": "[92] Jhih-Ciang Wu, He-Yen Hsieh, Ding-Jie Chen, Chiou-Shann Fuh, and Tyng-Luh Liu. Self-supervised sparse representation for video anomaly detection. European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 321, + 343, + 554, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 343, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 321, + 343, + 554, + 387 + ], + "type": "text", + "content": "[93] Hu Xu, Gargi Ghosh, Po-Yao Huang, Dmytro Okhonko, Armen Aghajanyan, Florian Metze, Luke Zettlemoyer, and Christoph Feichtenhofer. VideoCLIP: Contrastive pretraining for zero-shot video-text understanding. 2021. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 321, + 388, + 554, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 388, + 554, + 431 + ], + "spans": [ + { + "bbox": [ + 321, + 388, + 554, + 431 + ], + "type": "text", + "content": "[94] Ming Xu and Stephen Gould. Temporally consistent unbalanced optimal transport for unsupervised action segmentation. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 321, + 434, + 554, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 434, + 554, + 476 + ], + "spans": [ + { + "bbox": [ + 321, + 434, + 554, + 476 + ], + "type": "text", + "content": "[95] S. Yeung, O. Russakovsky, N. Jin, M. Andriluka, G. Mori, and L. Fei-Fei. Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 2018. 1, 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 321, + 479, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 479, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 321, + 479, + 554, + 510 + ], + "type": "text", + "content": "[96] Fangqiu Yi, Hongyu Wen, and Tingting Jiang. Asformer: Transformer for action segmentation. In The British Machine Vision Conference (BMVC), 2021. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 321, + 513, + 554, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 513, + 554, + 567 + ], + "spans": [ + { + "bbox": [ + 321, + 513, + 554, + 567 + ], + "type": "text", + "content": "[97] Guang Yu, Siqi Wang, Zhiping Cai, Xinwang Liu, Chuanfu Xu, and Chengkun Wu. Deep anomaly discovery from unlabeled videos via normality advantage and self-paced refinement. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 321, + 569, + 554, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 569, + 554, + 623 + ], + "spans": [ + { + "bbox": [ + 321, + 569, + 554, + 623 + ], + "type": "text", + "content": "[98] Muhammad Zaigham Zaheer, Arif Mahmood, Marcella Astrid, and Seung-Ik Lee. Claws: Clustering assisted weakly supervised learning with normalcy suppression for anomalous event detection. European Conference on Computer Vision, 2020." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 321, + 624, + 554, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 624, + 554, + 677 + ], + "spans": [ + { + "bbox": [ + 321, + 624, + 554, + 677 + ], + "type": "text", + "content": "[99] M Zaigham Zaheer, Arif Mahmood, M Haris Khan, Mattia Segu, Fisher Yu, and Seung-Ik Lee. Generative cooperative learning for unsupervised video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 554, + 713 + ], + "type": "text", + "content": "[100] P. Zameni, Y. Shen, and E. Elhamifar. Moscato: Predicting multiple object state change through actions. International Conference on Computer Vision, 2025. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "10020" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 498 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 126 + ], + "type": "text", + "content": "[101] Luca Zanella, Willi Menapace, Massimiliano Mancini, Yiming Wang, and Elisa Ricci. Harnessing large language models for training-free video anomaly detection. IEEE Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 173 + ], + "type": "text", + "content": "[102] Ali Zare, Yulei Niu, Hammad Ayyubi, and Shih-Fu Chang. Rap: Retrieval-augmented planner for adaptive procedure planning in instructional videos. European Conference on Computer Vision, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 206 + ], + "type": "text", + "content": "[103] Junbin Zhang, Pei-Hsuan Tsai, and Meng-Hsun Tsai. Semantic2graph: Graph-based multi-modal feature fusion for action segmentation in videos, 2022. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[104] Shrinivas Ramasubramanian Angela Yao Zhanzhong Pang, Fadime Sener. Long-tail temporal action segmentation with group-wise temporal logit adjustment. European Conference on Computer Vision, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 296 + ], + "type": "text", + "content": "[105] Bin Zhao, Li Fei-Fei, and Eric P. Xing. Online detection of unusual events in videos via dynamic sparse coding. IEEE Conference on Computer Vision and Pattern Recognition, 2011. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 361 + ], + "type": "text", + "content": "[106] Honglu Zhou, Roberto Martin-Martin, Mubbasir Kapadia, Silvio Savarese, and Juan Carlos Niebles. Procedure-aware pretraining for instructional video understanding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10727-10738, 2023. 1, 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 418 + ], + "type": "text", + "content": "[107] Joey Tianyi Zhou, Jiawei Du, Hongyuan Zhu, Xi Peng, Yong Liu, , and Rick Siow Mong Goh. Anomalynet: An anomaly detection network for video surveillance. IEEE Transactions on Information Forensics and Security, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 420, + 294, + 453 + ], + "type": "text", + "content": "[108] Yuansheng Zhu, Wentao Bao, , and Qi Yu. Towards open set video anomaly detection. European Conference on Computer Vision, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 498 + ], + "type": "text", + "content": "[109] D. Zhukov, J. B. Alayrac, R. G. Cinbis, D. Fouhey, I. Laptev, and J. Sivic. Cross-task weakly supervised learning from instructional videos. IEEE Conference on Computer Vision and Pattern Recognition, 2019. 1, 2" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10021" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 12 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_content_list.json b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7a962733159b307eb82785cdb8f78334f272feb2 --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_content_list.json @@ -0,0 +1,1614 @@ +[ + { + "type": "text", + "text": "Estimating 2D Camera Motion with Hybrid Motion Basis", + "text_level": 1, + "bbox": [ + 207, + 130, + 789, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Haipeng Li $^{1*}$ Tianhao Zhou $^{1*}$ Zhanglei Yang $^{1}$ Yi Wu $^{2}$ Yan Chen $^{2}$ Zijing Mao $^{2}$ Shen Cheng $^{3}$ Bing Zeng $^{1}$ Shuaicheng Liu $^{1\\dagger}$ $^{1}$ University of Electronic Science and Technology of China \n $^{2}$ Xiaomi Corporation ${}^{3}$ Dexmal", + "bbox": [ + 184, + 179, + 807, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating 2D camera motion is a fundamental computer vision task that models the projection of 3D camera movements onto the 2D image plane. Current methods rely on either homography-based approaches, limited to planar scenes, or meshflow techniques that use grid-based local homographies but struggle with complex non-linear transformations. We introduce CamFlow, a novel framework that represents camera motion using hybrid motion bases: physical bases derived from camera geometry and stochastic bases for complex scenarios. Our approach includes a hybrid probabilistic loss function based on the Laplace distribution that enhances training robustness. For evaluation, we create a new benchmark by masking dynamic objects in existing optical flow datasets to isolate pure camera motion. Experiments show CamFlow outperforms state-of-the-art methods across diverse scenarios, demonstrating superior robustness and generalization in zero-shot settings. Code and datasets are available at our project page: https://lhaippp.github.io/CamFlow/.", + "bbox": [ + 88, + 316, + 483, + 604 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 631, + 220, + 645 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating 2D camera motion, which involves recovering the projection of 3D rotation and translation onto 2D planes [14], is a cornerstone of computer vision. Given a 3D rotation matrix $\\mathbf{R}$ and translation vector $\\mathbf{t}$ , the camera motion $\\mathbf{M}$ can be expressed as:", + "bbox": [ + 89, + 656, + 483, + 731 + ], + "page_idx": 0 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} = \\mathbf {K} \\left(\\mathbf {R} + \\mathbf {t} \\frac {\\mathbf {n} ^ {T}}{d}\\right) \\mathbf {K} ^ {- 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 737, + 482, + 771 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "where $\\mathbf{n}^T$ is the transpose of the normal vector to planes in the scene, $d$ denotes the distance from the camera center to each plane, and $\\mathbf{K}$ represents the camera intrinsic matrix. The resulting 2D camera motion is inherently nonlinear due to its dependence on scene depth and plane geometry. In real-world images, scenes typically consist of", + "bbox": [ + 89, + 777, + 483, + 868 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/861dde72e0baeb08c17b1d10ee3e4f76087567a8043dcd50e64dab534f14b634.jpg", + "image_caption": [ + "Figure 1. Comparison of camera motion estimation approaches in multi-plane scenes. Visualizations show warped source images overlaid on targets, with brighter areas in heatmaps indicating higher error. BasesHomo [44] captures only background motion, while MeshHomoGAN [31] improves accuracy through local grid-based homographies. Our CamFlow, with hybrid motion basis, achieves superior representation of complex camera motion." + ], + "image_footnote": [], + "bbox": [ + 517, + 286, + 710, + 477 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17bdb60c4321a858e1967d29ecf7798a797140a370a10243d6c98baf20c30afa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 287, + 903, + 477 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "multiple depths and planes. As a result, different regions of an image undergo distinct transformations, leading to complex, non-linear motion patterns. This task is essential for various computational imaging applications, such as digital video stabilization [26], where accurate representation of camera motion directly enhances performance. Existing methods for modeling camera motion primarily fall into two categories: homography and meshflow-based approaches.", + "bbox": [ + 511, + 609, + 906, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Homography, a perspective transformation, aligns two views of a planar or nearly-planar 3D scene [14]. Traditional methods typically rely on artificially extracted and matched keypoints [33], excelling under standard conditions but often struggling in adverse scenarios (e.g., rain, snow, and low light) or when confronted with non-planar motion caused by depth parallax or dynamic objects. Recent deep learning approaches mitigate the reliance on keypoints, enabling direct estimation of homography from image pairs through robust, data-driven networks. A notable benchmark work, BasesHomo [44], reformulates the task", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 875, + 218, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author (liushuaicheng@uestc.edu.cn).", + "bbox": [ + 109, + 887, + 397, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7624", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "as learning a linear combination of an 8-dimensional motion basis, pioneering a new direction in motion estimation. Nevertheless, they are limited because: a homography can only align a single plane.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this limitation, MeshFlow [27] partitions the image into $N \\times N$ grids, estimating a local homography for each cell and smoothing them to model non-linear camera motion. This approach performs well in scenes with multiple small-baseline depth variations and has become a popular choice for digital video stabilization [26, 27]. Furthermore, deep meshflow variants [28] demonstrate enhanced robustness under challenging conditions. However, a key limitation persists in current camera motion representation: increasing the number of grids enhances the ability to model non-linearity but raises optimization challenges [43].", + "bbox": [ + 88, + 152, + 482, + 319 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we introduce CamFlow, a novel representation that models complex camera motion through hybrid motion bases. Our key insight is that the flow field resulting from the superposition of multiple homographies is inherently non-linear (as visualized in Fig. 3), enabling more sophisticated motion modeling beyond single-plane limitations. Building on this observation, we establish a comprehensive hybrid basis subspace comprising:", + "bbox": [ + 89, + 321, + 482, + 441 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Physical Motion Bases: Derived from Taylor expansion of homographic transformations up to second-order terms, our 12 physical bases model fundamental geometric transformations (rotation, translation, scaling, and perspective) that capture essential camera motion patterns;", + "- Noisy Motion Bases: To model complex residual motion, we construct $K$ orthogonal components through SVD decomposition of randomly sampled homographies from a Gaussian distribution, effectively complementing the physical bases by capturing higher-order motion patterns. To stabilize the training procedure and simplify complex loss designs, we propose a hybrid probabilistic loss function that assumes motion models follow a Laplace distribution [42], facilitating robust and efficient optimization." + ], + "bbox": [ + 89, + 444, + 482, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, CamFlow effectively represents complex 2D camera motion as illustrated in Fig. 1, capturing both background and foreground elements while accurately modeling depth-varying motion patterns that conventional approaches struggle to represent. To rigorously evaluate its performance, we introduce GHOF-Cam, a novel benchmark specifically designed for camera motion estimation by systematically masking dynamic objects and ill-posed occlusion regions in established optical flow datasets [24], thereby isolating pure camera-induced motion. Through comprehensive experiments across diverse datasets under both standard and challenging conditions, we demonstrate that CamFlow consistently outperforms state-of-the-art single-plane homography and multi-planar methods in both sparse and dense camera motion estimation tasks, exhibiting superior robustness and generalization capability in", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "real-world scenarios. Our main contributions are:", + "bbox": [ + 511, + 90, + 841, + 104 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A new hybrid motion representation that learns to model complex non-linear 2D camera motion through physically interpretable and stochastic motion bases.", + "- A novel probabilistic loss formulation based on Laplace distribution that simplifies training and stabilizes optimization without complex loss designs.", + "- A comprehensive benchmark for evaluating camera motion learning across diverse conditions. Experimental results confirm our approach's effectiveness, robustness, and generalization ability across real-world scenarios." + ], + "bbox": [ + 511, + 113, + 903, + 263 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 291, + 660, + 306 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Homography Methods", + "text_level": 1, + "bbox": [ + 513, + 319, + 723, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The traditional homography estimation follows three stages: feature detection (e.g., SIFT [33], ORB [35]), correspondence matching [5], and outlier rejection (e.g., RANSAC [9], MAGSAC++ [2]). Learning-based methods like LIFT [45], SuperPoint [7], and SOSNet [40] improve robustness. Optimization-based approaches [4, 8] iteratively refine homography estimates, while deep learning methods span supervised [3, 6, 17, 21, 23, 37] and unsupervised [15, 19, 34, 44, 47] frameworks. Unsupervised models have proven effective for real-world scenarios. Notable examples include CAHomo [47] and BaseSHom [44], which enhance feature extraction and motion constraints. HomoGAN [15] integrates GAN loss [12] and Transformers [32] for coarse-to-fine refinement. Despite this progress, homography remains a single-plane model, limiting its effectiveness in complex motion. In this paper, we introduce a hybrid motion-basis representation to model multi-plane, non-linear motion more effectively.", + "bbox": [ + 511, + 347, + 906, + 619 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Mesh-based Methods", + "text_level": 1, + "bbox": [ + 513, + 645, + 714, + 660 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mesh-based warping estimates local homographies per mesh cell, including dual-plane approaches [10], patchwise mixtures of homography [13], flexible warping techniques like APAP [46], Bundled Paths [26], grid-based methods [27], and cascade residual homography [38]. Deep learning variants include deepMeshFlow [43], MeshCA-Homo [29], which merges multi-resolution meshes, and BasesMesh [30], which applies motion bases per grid. MeshHomoGAN [31] incorporates a planarity-aware mechanism for local homography estimation. However, these methods still assume that local regions fit the homography relationship, limiting their ability to represent complex motion. To address this limitation, we propose a novel motion representation that combines multiple non-linear motion bases, enhancing performance.", + "bbox": [ + 511, + 674, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7625", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/a2fc9670de9cef2c7ab5a3e2d6be45eb09ffee994f94f597dcec02f7ed7bf34a.jpg", + "image_caption": [ + "Figure 2. Our proposed motion estimation framework. Given image pair $(I_a, I_b)$ , features are extracted through a multi-scale pyramid and processed by the motion estimation transformer (MET) to compute weights for physical (blue) and noisy (red) motion bases. These weights linearly combine predefined motion bases to generate flow maps for warping. A mask generator predicts uncertainty masks $\\mathbf{d}_{ab}$ and $\\mathbf{d}_{ba}$ to reject unreliable regions, enhancing estimation robustness." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 906, + 396 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 492, + 181, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion Basis", + "text_level": 1, + "bbox": [ + 89, + 523, + 228, + 537 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we propose a novel motion representation through motion basis learning with deep networks to better represent the camera motion via Physical and Noisy bases.", + "bbox": [ + 89, + 551, + 482, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Physical Basis. Consider a pixel with homogeneous coordinates $\\mathbf{P}(x,y) = [x,y,1]^T\\in \\mathbb{R}^3$ . The physical motion $m = [\\Delta x,\\Delta y,1]^T$ induced by a homography $\\mathbf{H}$ is defined:", + "bbox": [ + 89, + 603, + 483, + 650 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nm = \\mathbf {H} \\cdot \\mathbf {P} (x, y) - \\mathbf {P} (x, y), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 189, + 672, + 482, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After normalization, the two-dimensional coordinates (x-axis and y-axis) of the motion become:", + "bbox": [ + 89, + 714, + 482, + 744 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta x = \\frac {h _ {1} x + h _ {2} y + h _ {3}}{h _ {7} x + h _ {8} y + 1} - x, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 767, + 482, + 805 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta y = \\frac {h _ {4} x + h _ {5} y + h _ {6}}{h _ {7} x + h _ {8} y + 1} - y,\n$$\n", + "text_format": "latex", + "bbox": [ + 191, + 801, + 383, + 833 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $h_1, \\ldots, h_8$ denote the elements of matrix $\\mathbf{H}$ , and $h_9$ is constrained to 1. By applying a Taylor expansion, this motion can be mapped to another subspace. For instance,", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "expanding $\\Delta x$ around the point $(x,y) = (0,0)$ gives:", + "bbox": [ + 511, + 494, + 866, + 510 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta x = \\frac {\\left(h _ {1} - 1\\right) x + h _ {2} y + h _ {3} - h _ {7} x ^ {2} - h _ {8} x y}{h _ {7} x + h _ {8} y + 1} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 516, + 906, + 549 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\approx w _ {1} \\cdot 1 + w _ {2} \\cdot x + w _ {3} \\cdot y + w _ {4} \\cdot x y \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 553, + 903, + 568 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n+ w _ {5} \\cdot x ^ {2} + w _ {6} \\cdot y ^ {2} + \\Delta , \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 571, + 903, + 589 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $w_{i}, i \\in [1,6]$ are coefficients, the basis functions are $b = [1,x,y,xy,x^2,y^2]$ and $\\Delta$ denotes the higher-order infinitesimal. Similarly, $\\Delta y$ can be decomposed into this subspace. By combining the decompositions of both $\\Delta x, \\Delta y$ , the motion space can be represented using the 12 bases:", + "bbox": [ + 511, + 597, + 906, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {F} = \\left\\{\\left(b _ {i}, 0\\right) \\mid b _ {i} \\in b \\right\\} \\cup \\left\\{\\left(0, b _ {i}\\right) \\mid b _ {i} \\in b \\right\\}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 680, + 903, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $b_{i}$ denotes the $i$ -th element in $b$ . Each basis can be transformed into an optical flow according to the image coordinate, as shown in Fig. 2 (Physical motion bases).", + "bbox": [ + 511, + 704, + 903, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Non-Linearity of Adding Bases. A key finding is that combining flow fields derived from different homographies produces a non-linear motion field that cannot be represented by any single homography. This challenges the linear basis assumption in prior work [44]. As shown in Fig. 3, when two homography-derived flows (flow1 and flow2) are added together, the resulting flow (flow3) differs from the flow derived from multiplying the original homography matrices (homo3). Additionally, attempting to solve for a homography from points sampled on flow3 yields inconsistent", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7626", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a6325c9dde6565015401ca32f5c72a94dd1180fb1776522ef62d649ab436f46.jpg", + "image_caption": [ + "Figure 3. Non-linearity of flow addition. Two homography matrices generate flow1 and flow2. Adding these flows (flow3) differs from the flow derived by multiplying the original homography matrices (homo3). When sampling points from flow3 to solve for a homography, we get inconsistent solutions, proving that combined flow fields cannot be represented by a single homography." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 482, + 202 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "solutions, confirming that the combined flow cannot be represented by any single homography. This observation motivates our approach, which leverages more bases to model complex camera motion.", + "bbox": [ + 89, + 305, + 483, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Stochastic Basis. While the physical bases effectively capture fundamental motion patterns, the complete space of camera motions is infinitely dimensional when considering higher-order Taylor expansions. Exhaustively modeling all possible bases becomes computationally intractable. To address this challenge, we leverage the expressive power of random sampling to complement our physical bases. Specifically, we generate $K$ random $3 \\times 3$ matrices $\\left\\{\\mathbf{H}^{(k)}\\right\\}_{k=1}^{K} \\subseteq \\mathbb{R}^{3 \\times 3}$ , where each matrix is formulated as:", + "bbox": [ + 89, + 366, + 483, + 508 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {H} = \\left\\{h _ {i} \\right\\} _ {i = 1} ^ {9}, \\text {w h e r e} \\quad h _ {i} = \\left\\{ \\begin{array}{l l} \\epsilon_ {i} \\sim \\mathcal {N} (0, 1), & 1 \\leq i \\leq 8, \\\\ 1, & i = 9. \\end{array} \\right. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 518, + 482, + 555 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Following BasesHomo [44], we convert random matrices into flows and apply singular value decomposition (SVD) to extract principal components. This process yields $N - 12$ stochastic bases that capture diverse motion patterns beyond physical bases, as illustrated in Fig. 2 (Noisy motion bases). These stochastic bases, combined with the 12 physical basis vectors described earlier, form a comprehensive set of $N$ motion bases that enhances our ability to model complex non-linear camera motions.", + "bbox": [ + 89, + 565, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Network Structure", + "text_level": 1, + "bbox": [ + 89, + 710, + 272, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The network is illustrated in Fig. 2. Following practices from previous work [15, 47], we process video frames by: (1) random cropping to $320 \\times 576$ patches, (2) converting to grayscale, (3) projecting into a shallow feature space to handle luminance variations, and (4) generating a 3-layer feature pyramid for multi-scale processing. Then we propose a Motion Estimation Transformer (MET). The MET employs a specialized architecture to separately predict weights for both physical bases (12 weights) and stochastic bases ( $N - 12$ weights), yielding bidirectional weights $\\hat{\\mathbf{w}}_{ab}$ and $\\hat{\\mathbf{w}}_{ba}$ . The final camera motion is computed through a", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "linear combination of the predicted weights and their corresponding motion bases, producing bidirectional dense motion fields $\\hat{\\mathbf{m}}_{ab}$ and $\\hat{\\mathbf{m}}_{ba}$ .", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "It is noteworthy that the confidence masks $\\mathbf{d}_{ab}$ and $\\mathbf{d}_{ba}$ are also crucial in predicting camera motion, as they effectively filter out dynamic objects. First, we apply the predicted $\\hat{\\mathbf{m}}_{ab}$ and $\\hat{\\mathbf{m}}_{ba}$ to $f_{a}$ and $f_{b}$ , respectively, obtaining warped feature maps $f_{a}^{\\prime}$ and $f_{b}^{\\prime}$ . The mask network $\\mathcal{M}$ [22] inputs the concatenated features to generate weighted maps that highlight well-aligned regions:", + "bbox": [ + 511, + 136, + 905, + 243 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {d} _ {a b} = \\mathcal {M} \\left(\\left[ f _ {a}, f _ {b} ^ {\\prime} \\right]\\right), \\mathbf {d} _ {b a} = \\mathcal {M} \\left(\\left[ f _ {b}, f _ {a} ^ {\\prime} \\right]\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 250, + 906, + 270 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Loss Function", + "text_level": 1, + "bbox": [ + 513, + 277, + 660, + 292 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Probabilistic Motion Modeling. Camera motion estimation faces a fundamental challenge: distinguishing between camera-induced and object motion in the scene. We therefore formulate our approach as a probabilistic model that explicitly accounts for uncertainty in motion estimation.", + "bbox": [ + 511, + 299, + 905, + 375 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Building on previous findings that 2D motion follows a Laplace distribution [11, 16, 42], we model the conditional probability of the true camera motion given our prediction:", + "bbox": [ + 511, + 375, + 905, + 421 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 429, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\hat{\\mathbf{m}}_{ab}$ is our predicted motion field and $\\mathbf{d}_{ab}$ represents our confidence in each pixel's motion estimate. Higher confidence values indicate pixels likely following camera motion, while lower values suggest pixels belonging to independently moving objects.", + "bbox": [ + 511, + 455, + 905, + 530 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Laplace Distribution Model. We model the probability density using two conditionally independent Laplace distributions for the horizontal and vertical components:", + "bbox": [ + 511, + 530, + 903, + 575 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) = \\prod \\left(\\frac {1}{\\sqrt {2 \\sigma^ {2}}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | u - \\mu_ {u} |} \\right. \\tag {11} \\\\ \\left. \\cdot \\frac {1}{\\sqrt {2} \\sigma^ {2}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | v - \\mu_ {v} |}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 583, + 903, + 645 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $(u,v)$ are the ground-truth motion components, $(\\mu_u,\\mu_v)$ are the predicted motion components, and $\\sigma^2$ is derived from our confidence mask $\\mathbf{d}_{ab}$ . This formulation allows our model to express both its prediction and its uncertainty about that prediction.", + "bbox": [ + 511, + 651, + 905, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hybrid Loss Strategy. Another challenge in camera motion estimation is the scarcity of ground-truth labels. To overcome this, we introduce a hybrid loss strategy with two components: 1) Motion supervision loss $(\\ell_{NLL_m})$ : we generate pseudo-labels using existing methods and apply negative log-likelihood (NLL) loss in both forward and backward directions:", + "bbox": [ + 511, + 726, + 905, + 830 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {N L L _ {m}} = - \\log p \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(\\mathbf {m} _ {b a} \\mid \\hat {\\mathbf {m}} _ {b a}; \\mathbf {d} _ {b a}\\right). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 835, + 903, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2) Photometric loss $(\\ell_{NLL_p})$ : We apply the same probabilistic framework to enforce consistency between warped features. Given image features $f_{a}$ and $f_{b}$ , we use our predicted", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7627", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/5a40caf1a186e281aa23eb221e410ebb71b84dc83aeea338b19982c1bdce3251.jpg", + "image_caption": [ + "Figure 4. Using the GHOF benchmark, we leverage SAM to generate semantic maps and manually identify dynamic objects such as cars and people, producing corresponding masks. These masks are then dilated to encompass occlusion regions, reducing ill-posed artifacts. Finally, the masks are applied to both the images and optical flow, isolating camera-induced motion and constructing a camera-motion-only dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 156 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "motion to produce warped features $f_{a}^{\\prime}$ and $f_{b}^{\\prime}$ :", + "bbox": [ + 89, + 280, + 393, + 297 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {N L L _ {p}} = - \\log p \\left(f _ {a} \\mid f _ {b} ^ {\\prime}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(f _ {b} \\mid f _ {a} ^ {\\prime}; \\mathbf {d} _ {b a}\\right). \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 305, + 482, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Adaptive loss balancing: to ensure stable training despite the different scales of our loss components, we dynamically balance them using:", + "bbox": [ + 89, + 332, + 483, + 378 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {\\text {o v e r a l l}} = \\ell_ {N L L _ {p}} + \\mathbf {w} \\times \\frac {\\left| \\ell_ {N L L _ {p}} \\right|}{\\left| \\ell_ {N L L _ {m}} \\right|} \\cdot \\ell_ {N L L _ {m}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 387, + 482, + 420 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{w}$ is a predefined weight.", + "bbox": [ + 89, + 428, + 305, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 458, + 223, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 482, + 187, + 497 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate our method on: CAHomo [47] and GHOF [24]. We train on CAHomo (460K training pairs, 4.2K test pairs across regular, low-texture, low-light, and foreground scenes) with additional generated samples [23], and conduct zero-shot testing on GHOF (256 test pairs in Regular, Foggy, Low-light, Rainy, and Snowy conditions).", + "bbox": [ + 89, + 505, + 483, + 595 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "GHOF-Cam Benchmark. To isolate camera motion from dynamic scene elements, we propose a camera-motion-specific benchmark derived from GHOF. We employ the Segment Anything Model (SAM) [20] to generate semantic maps, from which we identify dynamic objects (e.g., cars, people) and create corresponding masks. These masks are dilated to encompass occlusion regions, mitigating ill-posed artifacts at object boundaries. For edge occlusions not detected by semantic segmentation, we mainly utilize the ground-truth homography to identify black edge regions as additional masks. The combined masks are then applied to both input images and ground-truth optical flow, resulting in a benchmark that exclusively captures camera-induced motion, as illustrated in Fig. 4.", + "bbox": [ + 89, + 595, + 483, + 808 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Comparison with Existing Methods", + "text_level": 1, + "bbox": [ + 89, + 816, + 398, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We assess CamFlow against a comprehensive set of methods across three primary categories. The first category encompasses feature-based methods: SIFT [33], ORB [35], SuperPoint[7] with SuperGlue (SPSG) [36], and LoFTR", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ef87aaccd922e224b10b850dd1f58749ba0598090a70080350e04cb84c07f9b7.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodsAVGRELTLLSFLF
1)\\( \\mathcal{I}_{3\\times 3} \\)6.707.757.657.217.533.39
2)SIFT[33] + RANSAC[9]1.410.301.344.030.810.57
3)SIFT[33] + MAGSAC[1]1.340.311.723.390.800.47
4)ORB[35] + RANSAC[9]1.480.852.591.671.101.24
5)ORB[35] + MAGSAC[1]1.690.973.341.581.151.40
6)SPSG[7, 36] + RANSAC[9]0.710.410.870.720.800.75
7)SPSG[7, 36] + MAGSAC[1]0.630.360.790.700.710.70
8)LoFTR[39] + RANSAC[9]1.440.562.701.361.051.52
9)LoFTR[39] + MAGSAC[1]1.390.552.571.331.051.41
10)DHN[6]2.871.514.482.762.623.00
11)LocalTrans[37]4.214.094.844.555.302.25
12)IHN[3]4.674.855.545.105.042.84
13)RealSH[17]0.340.220.350.440.420.29
14)DMHomo[23]0.310.190.330.400.380.28
15)CAHomo[47]0.880.731.011.030.920.70
16)BasesHomo[44]0.500.290.540.650.610.41
17)HomoGAN[15]0.390.220.410.570.440.31
18)Ours*0.330.200.330.410.390.30
19)Ours0.320.190.320.390.390.31
", + "bbox": [ + 517, + 88, + 903, + 324 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/68fdd02bcdd385a6b35b906cc4595f72ab6849d05aa9779d2722446caf135d42.jpg", + "table_caption": [ + "Table 1. The benchmark consists of 5 distinct scenarios, namely regular (RE), low-texture (LT), low-light (LL), small foreground (SF), and large foreground (LF). The point matching errors (PME) on the test set of CAHomo [47] are presented." + ], + "table_footnote": [], + "table_body": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×35.223.656.695.884.904.96
3)SIFT[33]2.820.602.437.090.613.37
4)SPSG[7, 36]3.073.991.576.880.792.16
5)CAHomo[47]2.812.022.034.562.842.61
6)BasesHomo[44]1.741.390.974.120.661.58
7)Meshflow[27]2.151.092.215.570.441.69
8)HM_Mix[13]4.351.024.038.751.536.42
9)RANSAC-F[38]3.262.813.145.122.213.04
10)Ours1.101.080.742.150.461.05
", + "bbox": [ + 517, + 387, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c1fdb222a4a6cc9da9061bd1e231eacfaf5e24e3a70a44b0bf85048f18c39bf4.jpg", + "table_caption": [ + "Table 2. To evaluate generalizability, we compute the end point errors (EPE) of pre-trained models from Table 1 on our proposed GHOF-Cam benchmark." + ], + "table_footnote": [], + "table_body": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT[33]4.800.594.4712.100.626.20
4)SPSG[7, 36]4.473.542.2110.660.835.10
5)DHN[6]6.616.046.027.686.996.32
6)LocalTrans[37]5.724.066.495.955.786.34
7)IHN[3]8.177.108.719.346.579.13
8)RealSH[17]1.721.600.884.420.431.28
9)DMHomo[23]1.750.640.854.160.392.74
10)CAHomo[47]3.874.103.846.991.273.17
11)BasesHomo[44]2.282.021.434.900.782.29
12)HomoGAN[15]1.951.730.603.950.473.02
13)Ours1.231.150.962.690.400.93
", + "bbox": [ + 517, + 580, + 905, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 3. To evaluate generalizability, we compute the point matching errors (PME) on the GHOF [24] test set using pre-trained models from Table 1.", + "bbox": [ + 511, + 768, + 903, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "[39], each evaluated with two outlier rejection techniques: RANSAC [9] and MAGSAC [1]. The second category includes supervised learning approaches: DHN [6], Local-Trans [37], IHN [3], RealSH [17], and DMHomo [23]. The third category comprises unsupervised methods: CAHomo", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7628", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/65eed8a36685bfcf2d52976052ab9c32a0d657918969a5c468e9b08e04c05b86.jpg", + "table_caption": [], + "table_footnote": [ + "Table 4. Quantitative comparison of different methods on various environmental conditions. We present three key perception metrics: PSNR (higher is better), SSIM (higher is better), and LPIPS (lower is better). Our method consistently outperforms existing approaches across all scenarios and metrics." + ], + "table_body": "
MethodAVGREFOGDARKRAINSNOW
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
\\( \\mathcal{I}_{3\\times 3} \\)24.050.74030.083621.060.69000.075026.570.77110.082125.700.85060.078521.530.53350.141125.370.85620.0412
GT-Homo32.780.91870.057028.390.86970.054935.230.95080.049231.880.94050.057530.110.85110.103338.310.98140.0199
SIFT28.440.90740.078129.230.91480.054529.420.90160.076827.370.90740.098230.000.86320.105526.160.94970.0558
SPSG28.010.86970.079621.830.75930.088630.880.90490.064527.600.90190.096628.860.82700.110330.880.95560.0379
CAHomo25.290.78370.084122.670.73410.080527.510.80480.075126.120.87430.084622.950.61300.142027.200.89240.0384
BasesHomo29.610.90260.067225.080.85220.066631.060.91700.062730.050.93030.070229.580.85120.107132.300.96220.0292
MeshFlow29.910.92390.068828.570.92160.057628.680.92800.074229.410.92540.077430.680.87470.104932.230.97000.0298
HM_Mix25.770.88960.088226.090.87210.059626.560.87530.088226.430.90370.100228.200.86720.110721.580.92960.0820
RANSAC-F26.040.83480.089026.090.88120.066529.220.89440.080127.290.90310.092321.680.55850.149525.900.93710.0566
Ours32.090.91420.057527.080.86150.055834.170.93710.051232.360.94210.056530.520.86080.102136.350.96920.0218
", + "bbox": [ + 91, + 88, + 903, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "[47], BasesHomo [44], and HomoGAN [15].", + "bbox": [ + 89, + 273, + 387, + 287 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For multi-plane camera motion modeling, we compare against both traditional approaches (MeshFlow [27], Homography Mixture [13], RANSAC-Flow [38]) and unsupervised deep methods (BasesMesh [30] and MeshHomogAN [31]). Regarding pre-training, DHN, LocalTrans and IHN use the MS-COCO dataset [25], while other deep learning methods are pre-trained on CAHomo. Additional qualitative results and visual comparisons are available on: https://lhaippp.github.io/CamFlow/.", + "bbox": [ + 88, + 291, + 480, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2.1. Quantitative Comparison.", + "text_level": 1, + "bbox": [ + 89, + 439, + 316, + 454 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Sparse camera motion. We evaluate using the points matching error (PME), which measures the average geometric distance between transformed source points and their corresponding ground-truth target points. Table 1 presents the quantitative performance of our method alongside various homography estimation approaches on the CAHomo test set, categorized as: feature-based methods (rows 2-9), supervised approaches (rows 10-14), and unsupervised ones (rows 15-17). The $\\mathcal{I}_{3\\times 3}$ baseline (row 1) represents the distance between point pairs without transformation.", + "bbox": [ + 88, + 459, + 482, + 609 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our CamFlow method achieves superior results across multiple categories, outperforming the leading unsupervised method, HomoGAN, by reducing PME by $17.95\\%$ (from 0.39 to 0.32). While feature-based methods like SIFT+RANSAC excel in regular (RE) scenes with abundant texture and keypoints, CamFlow surpasses them with a $36.67\\%$ improvement (reducing error by 0.11). To balance benchmark performance with generalization capability, we include Ours\\* (row 18), which represents an early-stopped training model that delivers better generalizability across datasets, even though it slightly underperforms our fully-trained model (row 19) on the CAHomo benchmark.", + "bbox": [ + 88, + 612, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In small foreground (SF) and large foreground (LF) scenarios, where dynamic objects disrupt camera motion estimation, our probabilistic loss and confidence masking yield lower PMEs compared to methods like CAHomo and HomoGAN, which employ explicit outlier rejection masks for robustness. In low-texture (LT) and low-light (LL) conditions, learning-based approaches generally exhibit superior", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "resilience due to their keypoint-free strategies, particularly among unsupervised techniques. However, these scenes often contain homogeneous regions [41], occupying large image areas and reducing photometric differences, which limit unsupervised methods' effectiveness. By contrast, CamFlow excels in LT and LL.", + "bbox": [ + 511, + 273, + 903, + 363 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Generalization experiment. Evaluating cross-dataset generalization presents a significant challenge in motion estimation, particularly for applications where real-world video often differs substantially from training data. We assess CamFlow's generalization capabilities in two ways: (1) for dense camera motion, we evaluate against traditional single-plane, multi-planar, and deep learning approaches using our proposed GHOF-Cam benchmark (Table 2 and Table 4); and (2) for sparse camera motion, we compare against traditional and homography-based methods on the original GHOF benchmark [24] (Table 3).", + "bbox": [ + 511, + 367, + 903, + 532 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2 presents End Point Error (EPE) results on the GHOF-Cam benchmark, which represents ground-truth camera motion. We evaluate against homography-based methods (rows 3-5) and non-single-plane approaches (rows 6-9). As illustrated in Fig. 3, we classify BasesHomo [44] as a multi-plane rather than single-homography method. The results demonstrate that CamFlow outperforms competing methods across nearly all categories, showcasing exceptional zero-shot capability in capturing non-linear camera motion patterns. Additionally, Table 4 presents perceptual quality metrics (PSNR, SSIM, and LPIPS) across various conditions. We also include ground-truth homography results for reference (GT-Homo). Our method, CamFlow, achieves state-of-the-art performance across multiple categories, consistently outperforming all baselines, particularly in challenging conditions such as snowy scenes. Notably, CamFlow approaches the performance of ground-truth homography, with only a $0.69\\mathrm{dB}$ difference in PSNR, 0.0045 in SSIM and 0.0005 in LPIPS, while significantly surpassing the second-best method, MeshFlow. These results support CamFlow's potential for applications requiring high-quality visual alignment.", + "bbox": [ + 511, + 535, + 903, + 868 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3 presents compelling evidence of CamFlow's generalization capability on the GHOF benchmark. Our", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7629", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3f2f9b429c92677a72b90585fddde43ca0b71d10471518a0aa11d8762c006578.jpg", + "image_caption": [ + "Figure 5. Qualitative results of CamFlow and methods from each category (i.e., supervised, unsupervised, and multi-homography) on the CAHomo testset [47]. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 906, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d4e7535b869100088323413231d5983607b9ec659392b10d32cb1745488ce234.jpg", + "image_caption": [ + "Figure 6. Qualitative results of our method and the best-performing generalizable methods from learning-based categories (i.e., supervised, unsupervised, and multi-homography) on the GHOF testset [24]. The examples are arranged from top to bottom and include RE, Fog, and Dark. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + ], + "image_footnote": [], + "bbox": [ + 91, + 266, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "method achieves the lowest average PME (1.23), representing a $28.5\\%$ improvement over the previous best supervised method (RealSH at 1.72) and a $36.9\\%$ improvement over the leading unsupervised approach (HomoGAN at 1.95). This performance advantage extends across all environmental conditions, with particularly significant gains in challenging scenarios. In low-light (LL) conditions, CamFlow reduces error by $39.1\\%$ compared to RealSH (from 4.42 to 2.69), while in snowy (SNOW) environments, it achieves a remarkable $69.2\\%$ error reduction compared to HomoGAN (from 3.02 to 0.93). These results confirm that CamFlow enables robust zero-shot transfer to unseen datasets, even under diverse and challenging environmental conditions.", + "bbox": [ + 89, + 635, + 485, + 833 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Qualitative Comparison", + "text_level": 1, + "bbox": [ + 89, + 845, + 312, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 5 and 6 present qualitative results of CamFlow alongside competing methods on the CAHomo and GHOF test", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "sets. Dynamic visualizations are available on our project page. For visualization, we employ red-blue ghosting and alignment heat maps following [18]. After transforming the source frame using estimated camera flow, misaligned regions appear as red and blue ghosting. We highlight specific areas with red boxes and generate alignment heat maps where brighter regions indicate poorer alignment.", + "bbox": [ + 511, + 635, + 906, + 742 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In both figures, the \"Identity\" column shows source and target images overlaid without warping. For CAHomo (Fig. 5), we focus on evaluating background camera motion modeling capabilities. We selected challenging test cases featuring extremely low-light conditions (first row) and sophisticated motion patterns (parallax and depth variation), comparing against leading methods from three categories: supervised (RealSH), unsupervised (HomoGAN), and deep meshflow (BasesMesh, MeshHomoGAN). CamFlow achieves superior alignment in background regions.", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7630", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d94fd4b5de40cdfec6bd1ad59315be78bb82041cde3c99570c013173ef00e74b.jpg", + "image_caption": [ + "Figure 7. Uncertainty masks and corresponding original images. The brighter the mask, the higher the uncertainty. The masks effectively highlight dynamic objects, indicating regions where camera motion estimation is less reliable." + ], + "image_footnote": [], + "bbox": [ + 91, + 88, + 285, + 146 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/28bfc84b23cf4b9b223c571ca1794ae20203f78a5ecfd0dc2d2d8e8cc5c91d1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 290, + 88, + 482, + 146 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The GHOF benchmark (Fig. 6) highlights CamFlow's advantages in handling complex camera motion and its zero-shot generalization capabilities. The first row shows a scene with parallax and foreground motion that challenges homography-based methods and reduces meshflow accuracy. The second row demonstrates CamFlow's generalization to foggy conditions unseen in training data, where depth variation complicates motion estimation. The last row presents an extreme case with motion blur, large parallax, and dynamic objects. In all scenarios, CamFlow delivers robust results where competing methods struggle, confirming its effectiveness in modeling complex camera motion patterns across diverse environmental conditions.", + "bbox": [ + 88, + 220, + 482, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Foreground Masks. Fig. 7 illustrates the uncertainty masks. We observe that the network effectively identifies dynamic objects and assigns higher uncertainty to regions containing these objects. This design helps focus the learning process on areas where motion is most relevant, while reducing noise from dynamic object regions, thereby enhancing the accuracy of camera motion estimation.", + "bbox": [ + 89, + 417, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 534, + 254, + 547 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/21ad81e3b087f4f6dd0f10b1929ef8deb1fb5aca6a30823dfdc31435dc81148f.jpg", + "table_caption": [ + "4.4.1. Motion Basis" + ], + "table_footnote": [], + "table_body": "
CAHomoGHOFGHOF-CamParamsInference Time
8 Bases0.371.681.452.658M76.42ms
12 Bases0.361.541.232.658M75.38ms
24 Bases0.331.231.102.660M79.63ms
200 Bases0.331.271.072.677M99.28ms
", + "bbox": [ + 93, + 579, + 483, + 654 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Performance comparison under different numbers of motion bases on three benchmarks. Results demonstrate the effectiveness of combining physical and stochastic motion bases, with 24 bases providing optimal balance between accuracy and computational efficiency.", + "bbox": [ + 89, + 657, + 482, + 728 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In Table 5, we evaluate with varying numbers of motion bases across three benchmarks: CAHomo (trained) and GHOF/GHOF-Cam (zero-shot). The average results demonstrate that: 1) Increasing physical bases from 8 to 12 improves performance across all benchmarks; 2) Introducing additional hybrid bases (24 total) yields further enhancements, particularly for generalization (GHOF and GHOF-Cam). Notably, while expanding to 200 bases provides marginal improvements on some benchmarks, it increases inference time by $24.7\\%$ . We therefore adopt 24 bases in our final model as the optimal balance between ac", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "curacy and computational efficiency.", + "bbox": [ + 511, + 90, + 756, + 106 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1a26a36ba212c272c78ac5f9ea468808016e8d70b3b9084f7667be73f5e1eabe.jpg", + "table_caption": [ + "4.4.2. Hybrid Probabilistic Loss" + ], + "table_footnote": [], + "table_body": "
\\( {\\ell }_{NLLm} \\)\\( {\\ell }_{NLLp} \\)CAHomoGHOFGHOF-Cam
✓0.412.212.13
✓0.361.581.42
✓✓0.331.231.10
", + "bbox": [ + 555, + 135, + 867, + 204 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6. Ablation study of different loss function combinations across three benchmarks. Results demonstrate that the hybrid approach combining motion loss and photometric loss achieves superior performance compared to individual loss components.", + "bbox": [ + 511, + 208, + 903, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 6 evaluates our probabilistic loss components across three benchmarks. Using only motion loss yields limited generalization performance, because pseudo motion labels provide approximate supervision. The photometric loss alone performs substantially better, particularly on zero-shot datasets, consistent with findings from prior unsupervised methods. However, our hybrid approach combining both losses achieves the best results across all benchmarks, with improvements on GHOF-Cam (22.5% error reduction compared to photometric-only). We believe motion labels provide coarse guidance while photometric loss enables fine-grained refinement, resulting in more accurate and generalizable camera motion estimation.", + "bbox": [ + 511, + 267, + 906, + 464 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 476, + 633, + 491 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "CamFlow presents a novel motion representation for modeling 2D camera motion using a hybrid motion basis approach. We identify a fundamental issue: superposing homographies by simply adding flow fields introduces nonlinear interactions, contradicting the assumption that homographies can be expressed as linear combinations of 8 basis flows. By expanding the previous 8-dimensional motion basis into a higher-dimensional space with both physical and stochastic motion bases, CamFlow effectively captures complex, non-linear motion patterns. The proposed probabilistic loss function enhances training stability. Our newly introduced GHOF-Cam benchmark demonstrates that CamFlow surpasses state-of-the-art homography and meshflow methods, showcasing superior robustness and generalization. We hope this work opens new avenues for camera motion modeling in video processing applications. Codebase, pre-trained models, and benchmark data are released at https://lhaippp.github.io/CamFlow/.", + "bbox": [ + 511, + 501, + 906, + 773 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (NSFC) under grants No. 62372091 and No. 62031009, and in part by the Hainan Province Key R&D Program under grant No. ZDYF2024(LALH)001. We would also like to thank Hao Xu, Mingbo Hong, Hai Jiang, Xinglong Luo, Nianjin Ye, and the anonymous reviewers for their valuable discussions, constructive suggestions, and helpful feedback.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7631", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 5", + "[2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proc. CVPR, pages 1304-1312, 2020. 2", + "[3] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5", + "[4] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2", + "[5] Padraig Cunningham and Sarah Jane Delany. K-nearest neighbour classifiers-a tutorial. ACM Computing Surveys, 54(6):1-25, 2021. 2", + "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 2, 5", + "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5", + "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2", + "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5", + "[10] Junhong Gao, Seon Joo Kim, and Michael S Brown. Constructing image panoramas using dual-homography warping. In Proc. CVPR, pages 49-56, 2011. 2", + "[11] Jochen Gast and Stefan Roth. Lightweight probabilistic deep networks. In Proc. CVPR, pages 3369-3378, 2018. 4", + "[12] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Proc. NIPS, 27, 2014. 2", + "[13] Matthias Grundmann, Vivek Kwatra, Daniel Castro, and Irfan Essa. Calibration-free rolling shutter removal. In Proc. ICCP, pages 1-8, 2012. 2, 5, 6", + "[14] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 1", + "[15] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 4, 5, 6", + "[16] Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. In Proc. ECCV, pages 652-667, 2018. 4", + "[17] Hai Jiang, Haipeng Li, Songchen Han, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Supervised homography learning" + ], + "bbox": [ + 93, + 114, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with realistic dataset generation. In Proc. ICCV, pages 9806-9815, 2023. 2, 5", + "[18] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 7", + "[19] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2", + "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proc. ICCV, pages 4015-4026, 2023. 5", + "[21] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2", + "[22] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 4", + "[23] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Dmhomo: Learning homography with diffusion models. ACM Transactions on Graphics (TOG), 43(3):1-16, 2024. 2, 5", + "[24] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. International Journal of Computer Vision (IJCV), 132(6):2331-2349, 2024. 2, 5, 6, 7", + "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 6", + "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1, 2", + "[27] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815, 2016. 2, 5, 6", + "[28] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), pages 1-14, 2022. 2", + "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(3):2849-2863, 2022. 2", + "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(6):7885-7899, 2023. 2, 6", + "[31] Shuaicheng Liu, Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, and Bing Zeng. Unsupervised global and local" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7632", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "homography estimation with coplanarity-aware gan. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2024. 1, 2, 6", + "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2", + "[33] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision (IJCV), 60(2):91-110, 2004. 1, 2, 5", + "[34] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters (RA-L), 3(3): 2346-2353, 2018. 2", + "[35] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 2, 5", + "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 5", + "[37] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5", + "[38] Xi Shen, François Darmon, Alexei A Efros, and Mathieu Aubry. Ransac-flow: generic two-stage image alignment. In Proc. ECCV, pages 618-637, 2020. 2, 5, 6", + "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching" + ], + "bbox": [ + 91, + 90, + 483, + 526 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "with transformers. In Proc. CVPR, pages 8922-8931, 2021. 5", + "[40] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2", + "[41] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 6", + "[42] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2, 4", + "[43] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 2", + "[44] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 1, 2, 3, 4, 5, 6", + "[45] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Proc. ECCV, pages 467-483, 2016. 2", + "[46] Julio Zaragoza, Tat-Jun Chin, Michael S Brown, and David Suter. As-projective-as-possible image stitching with moving dlt. In Proc. CVPR, pages 2339–2346, 2013. 2", + "[47] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 4, 5, 6, 7" + ], + "bbox": [ + 514, + 90, + 903, + 526 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7633", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_model.json b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9cb7d880805d2078b2e3368624f192ea90c323ed --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_model.json @@ -0,0 +1,2222 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.209, + 0.131, + 0.79, + 0.153 + ], + "angle": 0, + "content": "Estimating 2D Camera Motion with Hybrid Motion Basis" + }, + { + "type": "text", + "bbox": [ + 0.185, + 0.18, + 0.808, + 0.253 + ], + "angle": 0, + "content": "Haipeng Li\\(^{1*}\\) Tianhao Zhou\\(^{1*}\\) Zhanglei Yang\\(^{1}\\) Yi Wu\\(^{2}\\) Yan Chen\\(^{2}\\) Zijing Mao\\(^{2}\\) Shen Cheng\\(^{3}\\) Bing Zeng\\(^{1}\\) Shuaicheng Liu\\(^{1\\dagger}\\) \n\\(^{1}\\)University of Electronic Science and Technology of China \n\\(^{2}\\)Xiaomi Corporation \\({}^{3}\\)Dexmal" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.485, + 0.606 + ], + "angle": 0, + "content": "Estimating 2D camera motion is a fundamental computer vision task that models the projection of 3D camera movements onto the 2D image plane. Current methods rely on either homography-based approaches, limited to planar scenes, or meshflow techniques that use grid-based local homographies but struggle with complex non-linear transformations. We introduce CamFlow, a novel framework that represents camera motion using hybrid motion bases: physical bases derived from camera geometry and stochastic bases for complex scenarios. Our approach includes a hybrid probabilistic loss function based on the Laplace distribution that enhances training robustness. For evaluation, we create a new benchmark by masking dynamic objects in existing optical flow datasets to isolate pure camera motion. Experiments show CamFlow outperforms state-of-the-art methods across diverse scenarios, demonstrating superior robustness and generalization in zero-shot settings. Code and datasets are available at our project page: https://lhaippp.github.io/CamFlow/." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.632, + 0.222, + 0.646 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.657, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Estimating 2D camera motion, which involves recovering the projection of 3D rotation and translation onto 2D planes [14], is a cornerstone of computer vision. Given a 3D rotation matrix \\(\\mathbf{R}\\) and translation vector \\(\\mathbf{t}\\), the camera motion \\(\\mathbf{M}\\) can be expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.738, + 0.483, + 0.772 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} = \\mathbf {K} \\left(\\mathbf {R} + \\mathbf {t} \\frac {\\mathbf {n} ^ {T}}{d}\\right) \\mathbf {K} ^ {- 1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.869 + ], + "angle": 0, + "content": "where \\(\\mathbf{n}^T\\) is the transpose of the normal vector to planes in the scene, \\(d\\) denotes the distance from the camera center to each plane, and \\(\\mathbf{K}\\) represents the camera intrinsic matrix. The resulting 2D camera motion is inherently nonlinear due to its dependence on scene depth and plane geometry. In real-world images, scenes typically consist of" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.287, + 0.712, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.288, + 0.905, + 0.478 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.49, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Figure 1. Comparison of camera motion estimation approaches in multi-plane scenes. Visualizations show warped source images overlaid on targets, with brighter areas in heatmaps indicating higher error. BasesHomo [44] captures only background motion, while MeshHomoGAN [31] improves accuracy through local grid-based homographies. Our CamFlow, with hybrid motion basis, achieves superior representation of complex camera motion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.61, + 0.907, + 0.732 + ], + "angle": 0, + "content": "multiple depths and planes. As a result, different regions of an image undergo distinct transformations, leading to complex, non-linear motion patterns. This task is essential for various computational imaging applications, such as digital video stabilization [26], where accurate representation of camera motion directly enhances performance. Existing methods for modeling camera motion primarily fall into two categories: homography and meshflow-based approaches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Homography, a perspective transformation, aligns two views of a planar or nearly-planar 3D scene [14]. Traditional methods typically rely on artificially extracted and matched keypoints [33], excelling under standard conditions but often struggling in adverse scenarios (e.g., rain, snow, and low light) or when confronted with non-planar motion caused by depth parallax or dynamic objects. Recent deep learning approaches mitigate the reliance on keypoints, enabling direct estimation of homography from image pairs through robust, data-driven networks. A notable benchmark work, BasesHomo [44], reformulates the task" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.22, + 0.888 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.398, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author (liushuaicheng@uestc.edu.cn)." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.398, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7624" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "as learning a linear combination of an 8-dimensional motion basis, pioneering a new direction in motion estimation. Nevertheless, they are limited because: a homography can only align a single plane." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.154, + 0.483, + 0.32 + ], + "angle": 0, + "content": "To address this limitation, MeshFlow [27] partitions the image into \\( N \\times N \\) grids, estimating a local homography for each cell and smoothing them to model non-linear camera motion. This approach performs well in scenes with multiple small-baseline depth variations and has become a popular choice for digital video stabilization [26, 27]. Furthermore, deep meshflow variants [28] demonstrate enhanced robustness under challenging conditions. However, a key limitation persists in current camera motion representation: increasing the number of grids enhances the ability to model non-linearity but raises optimization challenges [43]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.483, + 0.442 + ], + "angle": 0, + "content": "In this work, we introduce CamFlow, a novel representation that models complex camera motion through hybrid motion bases. Our key insight is that the flow field resulting from the superposition of multiple homographies is inherently non-linear (as visualized in Fig. 3), enabling more sophisticated motion modeling beyond single-plane limitations. Building on this observation, we establish a comprehensive hybrid basis subspace comprising:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.445, + 0.483, + 0.52 + ], + "angle": 0, + "content": "- Physical Motion Bases: Derived from Taylor expansion of homographic transformations up to second-order terms, our 12 physical bases model fundamental geometric transformations (rotation, translation, scaling, and perspective) that capture essential camera motion patterns;" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.483, + 0.658 + ], + "angle": 0, + "content": "- Noisy Motion Bases: To model complex residual motion, we construct \\(K\\) orthogonal components through SVD decomposition of randomly sampled homographies from a Gaussian distribution, effectively complementing the physical bases by capturing higher-order motion patterns. To stabilize the training procedure and simplify complex loss designs, we propose a hybrid probabilistic loss function that assumes motion models follow a Laplace distribution [42], facilitating robust and efficient optimization." + }, + { + "type": "list", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.658 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In summary, CamFlow effectively represents complex 2D camera motion as illustrated in Fig. 1, capturing both background and foreground elements while accurately modeling depth-varying motion patterns that conventional approaches struggle to represent. To rigorously evaluate its performance, we introduce GHOF-Cam, a novel benchmark specifically designed for camera motion estimation by systematically masking dynamic objects and ill-posed occlusion regions in established optical flow datasets [24], thereby isolating pure camera-induced motion. Through comprehensive experiments across diverse datasets under both standard and challenging conditions, we demonstrate that CamFlow consistently outperforms state-of-the-art single-plane homography and multi-planar methods in both sparse and dense camera motion estimation tasks, exhibiting superior robustness and generalization capability in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.842, + 0.106 + ], + "angle": 0, + "content": "real-world scenarios. Our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.114, + 0.905, + 0.158 + ], + "angle": 0, + "content": "- A new hybrid motion representation that learns to model complex non-linear 2D camera motion through physically interpretable and stochastic motion bases." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.159, + 0.905, + 0.204 + ], + "angle": 0, + "content": "- A novel probabilistic loss formulation based on Laplace distribution that simplifies training and stabilizes optimization without complex loss designs." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.205, + 0.905, + 0.265 + ], + "angle": 0, + "content": "- A comprehensive benchmark for evaluating camera motion learning across diverse conditions. Experimental results confirm our approach's effectiveness, robustness, and generalization ability across real-world scenarios." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.114, + 0.905, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.292, + 0.661, + 0.308 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.724, + 0.336 + ], + "angle": 0, + "content": "2.1. Homography Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.348, + 0.908, + 0.621 + ], + "angle": 0, + "content": "The traditional homography estimation follows three stages: feature detection (e.g., SIFT [33], ORB [35]), correspondence matching [5], and outlier rejection (e.g., RANSAC [9], MAGSAC++ [2]). Learning-based methods like LIFT [45], SuperPoint [7], and SOSNet [40] improve robustness. Optimization-based approaches [4, 8] iteratively refine homography estimates, while deep learning methods span supervised [3, 6, 17, 21, 23, 37] and unsupervised [15, 19, 34, 44, 47] frameworks. Unsupervised models have proven effective for real-world scenarios. Notable examples include CAHomo [47] and BaseSHom [44], which enhance feature extraction and motion constraints. HomoGAN [15] integrates GAN loss [12] and Transformers [32] for coarse-to-fine refinement. Despite this progress, homography remains a single-plane model, limiting its effectiveness in complex motion. In this paper, we introduce a hybrid motion-basis representation to model multi-plane, non-linear motion more effectively." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.646, + 0.715, + 0.661 + ], + "angle": 0, + "content": "2.2. Mesh-based Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Mesh-based warping estimates local homographies per mesh cell, including dual-plane approaches [10], patchwise mixtures of homography [13], flexible warping techniques like APAP [46], Bundled Paths [26], grid-based methods [27], and cascade residual homography [38]. Deep learning variants include deepMeshFlow [43], MeshCA-Homo [29], which merges multi-resolution meshes, and BasesMesh [30], which applies motion bases per grid. MeshHomoGAN [31] incorporates a planarity-aware mechanism for local homography estimation. However, these methods still assume that local regions fit the homography relationship, limiting their ability to represent complex motion. To address this limitation, we propose a novel motion representation that combines multiple non-linear motion bases, enhancing performance." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7625" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.086, + 0.907, + 0.397 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.412, + 0.908, + 0.469 + ], + "angle": 0, + "content": "Figure 2. Our proposed motion estimation framework. Given image pair \\((I_a, I_b)\\), features are extracted through a multi-scale pyramid and processed by the motion estimation transformer (MET) to compute weights for physical (blue) and noisy (red) motion bases. These weights linearly combine predefined motion bases to generate flow maps for warping. A mask generator predicts uncertainty masks \\(\\mathbf{d}_{ab}\\) and \\(\\mathbf{d}_{ba}\\) to reject unreliable regions, enhancing estimation robustness." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.493, + 0.182, + 0.509 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.524, + 0.229, + 0.539 + ], + "angle": 0, + "content": "3.1. Motion Basis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.553, + 0.483, + 0.599 + ], + "angle": 0, + "content": "In this work, we propose a novel motion representation through motion basis learning with deep networks to better represent the camera motion via Physical and Noisy bases." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.651 + ], + "angle": 0, + "content": "Physical Basis. Consider a pixel with homogeneous coordinates \\(\\mathbf{P}(x,y) = [x,y,1]^T\\in \\mathbb{R}^3\\). The physical motion \\(m = [\\Delta x,\\Delta y,1]^T\\) induced by a homography \\(\\mathbf{H}\\) is defined:" + }, + { + "type": "equation", + "bbox": [ + 0.19, + 0.674, + 0.483, + 0.691 + ], + "angle": 0, + "content": "\\[\nm = \\mathbf {H} \\cdot \\mathbf {P} (x, y) - \\mathbf {P} (x, y), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.483, + 0.745 + ], + "angle": 0, + "content": "After normalization, the two-dimensional coordinates (x-axis and y-axis) of the motion become:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.768, + 0.483, + 0.806 + ], + "angle": 0, + "content": "\\[\n\\Delta x = \\frac {h _ {1} x + h _ {2} y + h _ {3}}{h _ {7} x + h _ {8} y + 1} - x, \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.192, + 0.802, + 0.384, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\Delta y = \\frac {h _ {4} x + h _ {5} y + h _ {6}}{h _ {7} x + h _ {8} y + 1} - y,\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(h_1, \\ldots, h_8\\) denote the elements of matrix \\(\\mathbf{H}\\), and \\(h_9\\) is constrained to 1. By applying a Taylor expansion, this motion can be mapped to another subspace. For instance," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.867, + 0.511 + ], + "angle": 0, + "content": "expanding \\(\\Delta x\\) around the point \\((x,y) = (0,0)\\) gives:" + }, + { + "type": "equation", + "bbox": [ + 0.555, + 0.517, + 0.907, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\Delta x = \\frac {\\left(h _ {1} - 1\\right) x + h _ {2} y + h _ {3} - h _ {7} x ^ {2} - h _ {8} x y}{h _ {7} x + h _ {8} y + 1} \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.554, + 0.905, + 0.569 + ], + "angle": 0, + "content": "\\[\n\\approx w _ {1} \\cdot 1 + w _ {2} \\cdot x + w _ {3} \\cdot y + w _ {4} \\cdot x y \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.573, + 0.905, + 0.59 + ], + "angle": 0, + "content": "\\[\n+ w _ {5} \\cdot x ^ {2} + w _ {6} \\cdot y ^ {2} + \\Delta , \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.673 + ], + "angle": 0, + "content": "where \\(w_{i}, i \\in [1,6]\\) are coefficients, the basis functions are \\(b = [1,x,y,xy,x^2,y^2]\\) and \\(\\Delta\\) denotes the higher-order infinitesimal. Similarly, \\(\\Delta y\\) can be decomposed into this subspace. By combining the decompositions of both \\(\\Delta x, \\Delta y\\), the motion space can be represented using the 12 bases:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.681, + 0.905, + 0.698 + ], + "angle": 0, + "content": "\\[\n\\mathbf {F} = \\left\\{\\left(b _ {i}, 0\\right) \\mid b _ {i} \\in b \\right\\} \\cup \\left\\{\\left(0, b _ {i}\\right) \\mid b _ {i} \\in b \\right\\}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.905, + 0.75 + ], + "angle": 0, + "content": "where \\( b_{i} \\) denotes the \\( i \\)-th element in \\( b \\). Each basis can be transformed into an optical flow according to the image coordinate, as shown in Fig. 2 (Physical motion bases)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Non-Linearity of Adding Bases. A key finding is that combining flow fields derived from different homographies produces a non-linear motion field that cannot be represented by any single homography. This challenges the linear basis assumption in prior work [44]. As shown in Fig. 3, when two homography-derived flows (flow1 and flow2) are added together, the resulting flow (flow3) differs from the flow derived from multiplying the original homography matrices (homo3). Additionally, attempting to solve for a homography from points sampled on flow3 yields inconsistent" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7626" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.483, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.485, + 0.293 + ], + "angle": 0, + "content": "Figure 3. Non-linearity of flow addition. Two homography matrices generate flow1 and flow2. Adding these flows (flow3) differs from the flow derived by multiplying the original homography matrices (homo3). When sampling points from flow3 to solve for a homography, we get inconsistent solutions, proving that combined flow fields cannot be represented by a single homography." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.306, + 0.484, + 0.366 + ], + "angle": 0, + "content": "solutions, confirming that the combined flow cannot be represented by any single homography. This observation motivates our approach, which leverages more bases to model complex camera motion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.484, + 0.51 + ], + "angle": 0, + "content": "Stochastic Basis. While the physical bases effectively capture fundamental motion patterns, the complete space of camera motions is infinitely dimensional when considering higher-order Taylor expansions. Exhaustively modeling all possible bases becomes computationally intractable. To address this challenge, we leverage the expressive power of random sampling to complement our physical bases. Specifically, we generate \\(K\\) random \\(3 \\times 3\\) matrices \\(\\left\\{\\mathbf{H}^{(k)}\\right\\}_{k=1}^{K} \\subseteq \\mathbb{R}^{3 \\times 3}\\), where each matrix is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.519, + 0.483, + 0.556 + ], + "angle": 0, + "content": "\\[\n\\mathbf {H} = \\left\\{h _ {i} \\right\\} _ {i = 1} ^ {9}, \\text {w h e r e} \\quad h _ {i} = \\left\\{ \\begin{array}{l l} \\epsilon_ {i} \\sim \\mathcal {N} (0, 1), & 1 \\leq i \\leq 8, \\\\ 1, & i = 9. \\end{array} \\right. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.566, + 0.484, + 0.703 + ], + "angle": 0, + "content": "Following BasesHomo [44], we convert random matrices into flows and apply singular value decomposition (SVD) to extract principal components. This process yields \\( N - 12 \\) stochastic bases that capture diverse motion patterns beyond physical bases, as illustrated in Fig. 2 (Noisy motion bases). These stochastic bases, combined with the 12 physical basis vectors described earlier, form a comprehensive set of \\( N \\) motion bases that enhances our ability to model complex non-linear camera motions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.712, + 0.273, + 0.727 + ], + "angle": 0, + "content": "3.2. Network Structure" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The network is illustrated in Fig. 2. Following practices from previous work [15, 47], we process video frames by: (1) random cropping to \\(320 \\times 576\\) patches, (2) converting to grayscale, (3) projecting into a shallow feature space to handle luminance variations, and (4) generating a 3-layer feature pyramid for multi-scale processing. Then we propose a Motion Estimation Transformer (MET). The MET employs a specialized architecture to separately predict weights for both physical bases (12 weights) and stochastic bases (\\(N - 12\\) weights), yielding bidirectional weights \\(\\hat{\\mathbf{w}}_{ab}\\) and \\(\\hat{\\mathbf{w}}_{ba}\\). The final camera motion is computed through a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "linear combination of the predicted weights and their corresponding motion bases, producing bidirectional dense motion fields \\(\\hat{\\mathbf{m}}_{ab}\\) and \\(\\hat{\\mathbf{m}}_{ba}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.137, + 0.906, + 0.244 + ], + "angle": 0, + "content": "It is noteworthy that the confidence masks \\(\\mathbf{d}_{ab}\\) and \\(\\mathbf{d}_{ba}\\) are also crucial in predicting camera motion, as they effectively filter out dynamic objects. First, we apply the predicted \\(\\hat{\\mathbf{m}}_{ab}\\) and \\(\\hat{\\mathbf{m}}_{ba}\\) to \\(f_{a}\\) and \\(f_{b}\\), respectively, obtaining warped feature maps \\(f_{a}^{\\prime}\\) and \\(f_{b}^{\\prime}\\). The mask network \\(\\mathcal{M}\\) [22] inputs the concatenated features to generate weighted maps that highlight well-aligned regions:" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.251, + 0.907, + 0.271 + ], + "angle": 0, + "content": "\\[\n\\mathbf {d} _ {a b} = \\mathcal {M} \\left(\\left[ f _ {a}, f _ {b} ^ {\\prime} \\right]\\right), \\mathbf {d} _ {b a} = \\mathcal {M} \\left(\\left[ f _ {b}, f _ {a} ^ {\\prime} \\right]\\right). \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.278, + 0.661, + 0.293 + ], + "angle": 0, + "content": "3.3. Loss Function" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.906, + 0.375 + ], + "angle": 0, + "content": "Probabilistic Motion Modeling. Camera motion estimation faces a fundamental challenge: distinguishing between camera-induced and object motion in the scene. We therefore formulate our approach as a probabilistic model that explicitly accounts for uncertainty in motion estimation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.376, + 0.906, + 0.422 + ], + "angle": 0, + "content": "Building on previous findings that 2D motion follows a Laplace distribution [11, 16, 42], we model the conditional probability of the true camera motion given our prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.641, + 0.43, + 0.905, + 0.448 + ], + "angle": 0, + "content": "\\[\np \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.906, + 0.531 + ], + "angle": 0, + "content": "where \\(\\hat{\\mathbf{m}}_{ab}\\) is our predicted motion field and \\(\\mathbf{d}_{ab}\\) represents our confidence in each pixel's motion estimate. Higher confidence values indicate pixels likely following camera motion, while lower values suggest pixels belonging to independently moving objects." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.531, + 0.905, + 0.577 + ], + "angle": 0, + "content": "Laplace Distribution Model. We model the probability density using two conditionally independent Laplace distributions for the horizontal and vertical components:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.584, + 0.905, + 0.646 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) = \\prod \\left(\\frac {1}{\\sqrt {2 \\sigma^ {2}}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | u - \\mu_ {u} |} \\right. \\tag {11} \\\\ \\left. \\cdot \\frac {1}{\\sqrt {2} \\sigma^ {2}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | v - \\mu_ {v} |}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.652, + 0.906, + 0.727 + ], + "angle": 0, + "content": "where \\((u,v)\\) are the ground-truth motion components, \\((\\mu_u,\\mu_v)\\) are the predicted motion components, and \\(\\sigma^2\\) is derived from our confidence mask \\(\\mathbf{d}_{ab}\\). This formulation allows our model to express both its prediction and its uncertainty about that prediction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.727, + 0.906, + 0.832 + ], + "angle": 0, + "content": "Hybrid Loss Strategy. Another challenge in camera motion estimation is the scarcity of ground-truth labels. To overcome this, we introduce a hybrid loss strategy with two components: 1) Motion supervision loss \\((\\ell_{NLL_m})\\): we generate pseudo-labels using existing methods and apply negative log-likelihood (NLL) loss in both forward and backward directions:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.837, + 0.905, + 0.852 + ], + "angle": 0, + "content": "\\[\n\\ell_ {N L L _ {m}} = - \\log p \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(\\mathbf {m} _ {b a} \\mid \\hat {\\mathbf {m}} _ {b a}; \\mathbf {d} _ {b a}\\right). \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "2) Photometric loss \\((\\ell_{NLL_p})\\): We apply the same probabilistic framework to enforce consistency between warped features. Given image features \\(f_{a}\\) and \\(f_{b}\\), we use our predicted" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7627" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.157 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.169, + 0.483, + 0.268 + ], + "angle": 0, + "content": "Figure 4. Using the GHOF benchmark, we leverage SAM to generate semantic maps and manually identify dynamic objects such as cars and people, producing corresponding masks. These masks are then dilated to encompass occlusion regions, reducing ill-posed artifacts. Finally, the masks are applied to both the images and optical flow, isolating camera-induced motion and constructing a camera-motion-only dataset." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.281, + 0.395, + 0.298 + ], + "angle": 0, + "content": "motion to produce warped features \\( f_{a}^{\\prime} \\) and \\( f_{b}^{\\prime} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.306, + 0.483, + 0.323 + ], + "angle": 0, + "content": "\\[\n\\ell_ {N L L _ {p}} = - \\log p \\left(f _ {a} \\mid f _ {b} ^ {\\prime}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(f _ {b} \\mid f _ {a} ^ {\\prime}; \\mathbf {d} _ {b a}\\right). \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.38 + ], + "angle": 0, + "content": "Adaptive loss balancing: to ensure stable training despite the different scales of our loss components, we dynamically balance them using:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.388, + 0.483, + 0.421 + ], + "angle": 0, + "content": "\\[\n\\ell_ {\\text {o v e r a l l}} = \\ell_ {N L L _ {p}} + \\mathbf {w} \\times \\frac {\\left| \\ell_ {N L L _ {p}} \\right|}{\\left| \\ell_ {N L L _ {m}} \\right|} \\cdot \\ell_ {N L L _ {m}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.306, + 0.445 + ], + "angle": 0, + "content": "where \\(\\mathbf{w}\\) is a predefined weight." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.459, + 0.224, + 0.475 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.483, + 0.188, + 0.498 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.596 + ], + "angle": 0, + "content": "We evaluate our method on: CAHomo [47] and GHOF [24]. We train on CAHomo (460K training pairs, 4.2K test pairs across regular, low-texture, low-light, and foreground scenes) with additional generated samples [23], and conduct zero-shot testing on GHOF (256 test pairs in Regular, Foggy, Low-light, Rainy, and Snowy conditions)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.484, + 0.809 + ], + "angle": 0, + "content": "GHOF-Cam Benchmark. To isolate camera motion from dynamic scene elements, we propose a camera-motion-specific benchmark derived from GHOF. We employ the Segment Anything Model (SAM) [20] to generate semantic maps, from which we identify dynamic objects (e.g., cars, people) and create corresponding masks. These masks are dilated to encompass occlusion regions, mitigating ill-posed artifacts at object boundaries. For edge occlusions not detected by semantic segmentation, we mainly utilize the ground-truth homography to identify black edge regions as additional masks. The combined masks are then applied to both input images and ground-truth optical flow, resulting in a benchmark that exclusively captures camera-induced motion, as illustrated in Fig. 4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.818, + 0.4, + 0.835 + ], + "angle": 0, + "content": "4.2. Comparison with Existing Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We assess CamFlow against a comprehensive set of methods across three primary categories. The first category encompasses feature-based methods: SIFT [33], ORB [35], SuperPoint[7] with SuperGlue (SPSG) [36], and LoFTR" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.325 + ], + "angle": 0, + "content": "
MethodsAVGRELTLLSFLF
1)\\( \\mathcal{I}_{3\\times 3} \\)6.707.757.657.217.533.39
2)SIFT[33] + RANSAC[9]1.410.301.344.030.810.57
3)SIFT[33] + MAGSAC[1]1.340.311.723.390.800.47
4)ORB[35] + RANSAC[9]1.480.852.591.671.101.24
5)ORB[35] + MAGSAC[1]1.690.973.341.581.151.40
6)SPSG[7, 36] + RANSAC[9]0.710.410.870.720.800.75
7)SPSG[7, 36] + MAGSAC[1]0.630.360.790.700.710.70
8)LoFTR[39] + RANSAC[9]1.440.562.701.361.051.52
9)LoFTR[39] + MAGSAC[1]1.390.552.571.331.051.41
10)DHN[6]2.871.514.482.762.623.00
11)LocalTrans[37]4.214.094.844.555.302.25
12)IHN[3]4.674.855.545.105.042.84
13)RealSH[17]0.340.220.350.440.420.29
14)DMHomo[23]0.310.190.330.400.380.28
15)CAHomo[47]0.880.731.011.030.920.70
16)BasesHomo[44]0.500.290.540.650.610.41
17)HomoGAN[15]0.390.220.410.570.440.31
18)Ours*0.330.200.330.410.390.30
19)Ours0.320.190.320.390.390.31
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.329, + 0.905, + 0.386 + ], + "angle": 0, + "content": "Table 1. The benchmark consists of 5 distinct scenarios, namely regular (RE), low-texture (LT), low-light (LL), small foreground (SF), and large foreground (LF). The point matching errors (PME) on the test set of CAHomo [47] are presented." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.388, + 0.906, + 0.533 + ], + "angle": 0, + "content": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×35.223.656.695.884.904.96
3)SIFT[33]2.820.602.437.090.613.37
4)SPSG[7, 36]3.073.991.576.880.792.16
5)CAHomo[47]2.812.022.034.562.842.61
6)BasesHomo[44]1.741.390.974.120.661.58
7)Meshflow[27]2.151.092.215.570.441.69
8)HM_Mix[13]4.351.024.038.751.536.42
9)RANSAC-F[38]3.262.813.145.122.213.04
10)Ours1.101.080.742.150.461.05
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.537, + 0.906, + 0.579 + ], + "angle": 0, + "content": "Table 2. To evaluate generalizability, we compute the end point errors (EPE) of pre-trained models from Table 1 on our proposed GHOF-Cam benchmark." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.582, + 0.906, + 0.766 + ], + "angle": 0, + "content": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT[33]4.800.594.4712.100.626.20
4)SPSG[7, 36]4.473.542.2110.660.835.10
5)DHN[6]6.616.046.027.686.996.32
6)LocalTrans[37]5.724.066.495.955.786.34
7)IHN[3]8.177.108.719.346.579.13
8)RealSH[17]1.721.600.884.420.431.28
9)DMHomo[23]1.750.640.854.160.392.74
10)CAHomo[47]3.874.103.846.991.273.17
11)BasesHomo[44]2.282.021.434.900.782.29
12)HomoGAN[15]1.951.730.603.950.473.02
13)Ours1.231.150.962.690.400.93
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.769, + 0.905, + 0.81 + ], + "angle": 0, + "content": "Table 3. To evaluate generalizability, we compute the point matching errors (PME) on the GHOF [24] test set using pre-trained models from Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "[39], each evaluated with two outlier rejection techniques: RANSAC [9] and MAGSAC [1]. The second category includes supervised learning approaches: DHN [6], Local-Trans [37], IHN [3], RealSH [17], and DMHomo [23]. The third category comprises unsupervised methods: CAHomo" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7628" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.093, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": "
MethodAVGREFOGDARKRAINSNOW
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
\\( \\mathcal{I}_{3\\times 3} \\)24.050.74030.083621.060.69000.075026.570.77110.082125.700.85060.078521.530.53350.141125.370.85620.0412
GT-Homo32.780.91870.057028.390.86970.054935.230.95080.049231.880.94050.057530.110.85110.103338.310.98140.0199
SIFT28.440.90740.078129.230.91480.054529.420.90160.076827.370.90740.098230.000.86320.105526.160.94970.0558
SPSG28.010.86970.079621.830.75930.088630.880.90490.064527.600.90190.096628.860.82700.110330.880.95560.0379
CAHomo25.290.78370.084122.670.73410.080527.510.80480.075126.120.87430.084622.950.61300.142027.200.89240.0384
BasesHomo29.610.90260.067225.080.85220.066631.060.91700.062730.050.93030.070229.580.85120.107132.300.96220.0292
MeshFlow29.910.92390.068828.570.92160.057628.680.92800.074229.410.92540.077430.680.87470.104932.230.97000.0298
HM_Mix25.770.88960.088226.090.87210.059626.560.87530.088226.430.90370.100228.200.86720.110721.580.92960.0820
RANSAC-F26.040.83480.089026.090.88120.066529.220.89440.080127.290.90310.092321.680.55850.149525.900.93710.0566
Ours32.090.91420.057527.080.86150.055834.170.93710.051232.360.94210.056530.520.86080.102136.350.96920.0218
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.09, + 0.226, + 0.905, + 0.266 + ], + "angle": 0, + "content": "Table 4. Quantitative comparison of different methods on various environmental conditions. We present three key perception metrics: PSNR (higher is better), SSIM (higher is better), and LPIPS (lower is better). Our method consistently outperforms existing approaches across all scenarios and metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.275, + 0.388, + 0.289 + ], + "angle": 0, + "content": "[47], BasesHomo [44], and HomoGAN [15]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.292, + 0.482, + 0.427 + ], + "angle": 0, + "content": "For multi-plane camera motion modeling, we compare against both traditional approaches (MeshFlow [27], Homography Mixture [13], RANSAC-Flow [38]) and unsupervised deep methods (BasesMesh [30] and MeshHomogAN [31]). Regarding pre-training, DHN, LocalTrans and IHN use the MS-COCO dataset [25], while other deep learning methods are pre-trained on CAHomo. Additional qualitative results and visual comparisons are available on: https://lhaippp.github.io/CamFlow/." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.318, + 0.455 + ], + "angle": 0, + "content": "4.2.1. Quantitative Comparison." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.46, + 0.483, + 0.61 + ], + "angle": 0, + "content": "Sparse camera motion. We evaluate using the points matching error (PME), which measures the average geometric distance between transformed source points and their corresponding ground-truth target points. Table 1 presents the quantitative performance of our method alongside various homography estimation approaches on the CAHomo test set, categorized as: feature-based methods (rows 2-9), supervised approaches (rows 10-14), and unsupervised ones (rows 15-17). The \\(\\mathcal{I}_{3\\times 3}\\) baseline (row 1) represents the distance between point pairs without transformation." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.483, + 0.793 + ], + "angle": 0, + "content": "Our CamFlow method achieves superior results across multiple categories, outperforming the leading unsupervised method, HomoGAN, by reducing PME by \\(17.95\\%\\) (from 0.39 to 0.32). While feature-based methods like SIFT+RANSAC excel in regular (RE) scenes with abundant texture and keypoints, CamFlow surpasses them with a \\(36.67\\%\\) improvement (reducing error by 0.11). To balance benchmark performance with generalization capability, we include Ours\\* (row 18), which represents an early-stopped training model that delivers better generalizability across datasets, even though it slightly underperforms our fully-trained model (row 19) on the CAHomo benchmark." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In small foreground (SF) and large foreground (LF) scenarios, where dynamic objects disrupt camera motion estimation, our probabilistic loss and confidence masking yield lower PMEs compared to methods like CAHomo and HomoGAN, which employ explicit outlier rejection masks for robustness. In low-texture (LT) and low-light (LL) conditions, learning-based approaches generally exhibit superior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.905, + 0.364 + ], + "angle": 0, + "content": "resilience due to their keypoint-free strategies, particularly among unsupervised techniques. However, these scenes often contain homogeneous regions [41], occupying large image areas and reducing photometric differences, which limit unsupervised methods' effectiveness. By contrast, CamFlow excels in LT and LL." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.368, + 0.905, + 0.533 + ], + "angle": 0, + "content": "Generalization experiment. Evaluating cross-dataset generalization presents a significant challenge in motion estimation, particularly for applications where real-world video often differs substantially from training data. We assess CamFlow's generalization capabilities in two ways: (1) for dense camera motion, we evaluate against traditional single-plane, multi-planar, and deep learning approaches using our proposed GHOF-Cam benchmark (Table 2 and Table 4); and (2) for sparse camera motion, we compare against traditional and homography-based methods on the original GHOF benchmark [24] (Table 3)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.536, + 0.905, + 0.869 + ], + "angle": 0, + "content": "Table 2 presents End Point Error (EPE) results on the GHOF-Cam benchmark, which represents ground-truth camera motion. We evaluate against homography-based methods (rows 3-5) and non-single-plane approaches (rows 6-9). As illustrated in Fig. 3, we classify BasesHomo [44] as a multi-plane rather than single-homography method. The results demonstrate that CamFlow outperforms competing methods across nearly all categories, showcasing exceptional zero-shot capability in capturing non-linear camera motion patterns. Additionally, Table 4 presents perceptual quality metrics (PSNR, SSIM, and LPIPS) across various conditions. We also include ground-truth homography results for reference (GT-Homo). Our method, CamFlow, achieves state-of-the-art performance across multiple categories, consistently outperforming all baselines, particularly in challenging conditions such as snowy scenes. Notably, CamFlow approaches the performance of ground-truth homography, with only a \\(0.69\\mathrm{dB}\\) difference in PSNR, 0.0045 in SSIM and 0.0005 in LPIPS, while significantly surpassing the second-best method, MeshFlow. These results support CamFlow's potential for applications requiring high-quality visual alignment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Table 3 presents compelling evidence of CamFlow's generalization capability on the GHOF benchmark. Our" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7629" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.907, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.223, + 0.907, + 0.265 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of CamFlow and methods from each category (i.e., supervised, unsupervised, and multi-homography) on the CAHomo testset [47]. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.267, + 0.907, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.566, + 0.907, + 0.622 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of our method and the best-performing generalizable methods from learning-based categories (i.e., supervised, unsupervised, and multi-homography) on the GHOF testset [24]. The examples are arranged from top to bottom and include RE, Fog, and Dark. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.636, + 0.486, + 0.834 + ], + "angle": 0, + "content": "method achieves the lowest average PME (1.23), representing a \\(28.5\\%\\) improvement over the previous best supervised method (RealSH at 1.72) and a \\(36.9\\%\\) improvement over the leading unsupervised approach (HomoGAN at 1.95). This performance advantage extends across all environmental conditions, with particularly significant gains in challenging scenarios. In low-light (LL) conditions, CamFlow reduces error by \\(39.1\\%\\) compared to RealSH (from 4.42 to 2.69), while in snowy (SNOW) environments, it achieves a remarkable \\(69.2\\%\\) error reduction compared to HomoGAN (from 3.02 to 0.93). These results confirm that CamFlow enables robust zero-shot transfer to unseen datasets, even under diverse and challenging environmental conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.847, + 0.313, + 0.864 + ], + "angle": 0, + "content": "4.3. Qualitative Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Fig. 5 and 6 present qualitative results of CamFlow alongside competing methods on the CAHomo and GHOF test" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.636, + 0.907, + 0.743 + ], + "angle": 0, + "content": "sets. Dynamic visualizations are available on our project page. For visualization, we employ red-blue ghosting and alignment heat maps following [18]. After transforming the source frame using estimated camera flow, misaligned regions appear as red and blue ghosting. We highlight specific areas with red boxes and generate alignment heat maps where brighter regions indicate poorer alignment." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In both figures, the \"Identity\" column shows source and target images overlaid without warping. For CAHomo (Fig. 5), we focus on evaluating background camera motion modeling capabilities. We selected challenging test cases featuring extremely low-light conditions (first row) and sophisticated motion patterns (parallax and depth variation), comparing against leading methods from three categories: supervised (RealSH), unsupervised (HomoGAN), and deep meshflow (BasesMesh, MeshHomoGAN). CamFlow achieves superior alignment in background regions." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7630" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.089, + 0.287, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.291, + 0.089, + 0.483, + 0.147 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.156, + 0.483, + 0.212 + ], + "angle": 0, + "content": "Figure 7. Uncertainty masks and corresponding original images. The brighter the mask, the higher the uncertainty. The masks effectively highlight dynamic objects, indicating regions where camera motion estimation is less reliable." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.221, + 0.483, + 0.417 + ], + "angle": 0, + "content": "The GHOF benchmark (Fig. 6) highlights CamFlow's advantages in handling complex camera motion and its zero-shot generalization capabilities. The first row shows a scene with parallax and foreground motion that challenges homography-based methods and reduces meshflow accuracy. The second row demonstrates CamFlow's generalization to foggy conditions unseen in training data, where depth variation complicates motion estimation. The last row presents an extreme case with motion blur, large parallax, and dynamic objects. In all scenarios, CamFlow delivers robust results where competing methods struggle, confirming its effectiveness in modeling complex camera motion patterns across diverse environmental conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.525 + ], + "angle": 0, + "content": "Foreground Masks. Fig. 7 illustrates the uncertainty masks. We observe that the network effectively identifies dynamic objects and assigns higher uncertainty to regions containing these objects. This design helps focus the learning process on areas where motion is most relevant, while reducing noise from dynamic object regions, thereby enhancing the accuracy of camera motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.535, + 0.255, + 0.549 + ], + "angle": 0, + "content": "4.4. Ablation Studies" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.557, + 0.229, + 0.57 + ], + "angle": 0, + "content": "4.4.1. Motion Basis" + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.58, + 0.485, + 0.655 + ], + "angle": 0, + "content": "
CAHomoGHOFGHOF-CamParamsInference Time
8 Bases0.371.681.452.658M76.42ms
12 Bases0.361.541.232.658M75.38ms
24 Bases0.331.231.102.660M79.63ms
200 Bases0.331.271.072.677M99.28ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.729 + ], + "angle": 0, + "content": "Table 5. Performance comparison under different numbers of motion bases on three benchmarks. Results demonstrate the effectiveness of combining physical and stochastic motion bases, with 24 bases providing optimal balance between accuracy and computational efficiency." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In Table 5, we evaluate with varying numbers of motion bases across three benchmarks: CAHomo (trained) and GHOF/GHOF-Cam (zero-shot). The average results demonstrate that: 1) Increasing physical bases from 8 to 12 improves performance across all benchmarks; 2) Introducing additional hybrid bases (24 total) yields further enhancements, particularly for generalization (GHOF and GHOF-Cam). Notably, while expanding to 200 bases provides marginal improvements on some benchmarks, it increases inference time by \\(24.7\\%\\). We therefore adopt 24 bases in our final model as the optimal balance between ac" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.757, + 0.107 + ], + "angle": 0, + "content": "curacy and computational efficiency." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.113, + 0.74, + 0.128 + ], + "angle": 0, + "content": "4.4.2. Hybrid Probabilistic Loss" + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.136, + 0.868, + 0.205 + ], + "angle": 0, + "content": "
\\( {\\ell }_{NLLm} \\)\\( {\\ell }_{NLLp} \\)CAHomoGHOFGHOF-Cam
✓0.412.212.13
✓0.361.581.42
✓✓0.331.231.10
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.209, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Table 6. Ablation study of different loss function combinations across three benchmarks. Results demonstrate that the hybrid approach combining motion loss and photometric loss achieves superior performance compared to individual loss components." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.268, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Table 6 evaluates our probabilistic loss components across three benchmarks. Using only motion loss yields limited generalization performance, because pseudo motion labels provide approximate supervision. The photometric loss alone performs substantially better, particularly on zero-shot datasets, consistent with findings from prior unsupervised methods. However, our hybrid approach combining both losses achieves the best results across all benchmarks, with improvements on GHOF-Cam (22.5% error reduction compared to photometric-only). We believe motion labels provide coarse guidance while photometric loss enables fine-grained refinement, resulting in more accurate and generalizable camera motion estimation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.477, + 0.634, + 0.492 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.907, + 0.775 + ], + "angle": 0, + "content": "CamFlow presents a novel motion representation for modeling 2D camera motion using a hybrid motion basis approach. We identify a fundamental issue: superposing homographies by simply adding flow fields introduces nonlinear interactions, contradicting the assumption that homographies can be expressed as linear combinations of 8 basis flows. By expanding the previous 8-dimensional motion basis into a higher-dimensional space with both physical and stochastic motion bases, CamFlow effectively captures complex, non-linear motion patterns. The proposed probabilistic loss function enhances training stability. Our newly introduced GHOF-Cam benchmark demonstrates that CamFlow surpasses state-of-the-art homography and meshflow methods, showcasing superior robustness and generalization. We hope this work opens new avenues for camera motion modeling in video processing applications. Codebase, pre-trained models, and benchmark data are released at https://lhaippp.github.io/CamFlow/." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (NSFC) under grants No. 62372091 and No. 62031009, and in part by the Hainan Province Key R&D Program under grant No. ZDYF2024(LALH)001. We would also like to thank Hao Xu, Mingbo Hong, Hai Jiang, Xinglong Luo, Nianjin Ye, and the anonymous reviewers for their valuable discussions, constructive suggestions, and helpful feedback." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7631" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.156 + ], + "angle": 0, + "content": "[1] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.2 + ], + "angle": 0, + "content": "[2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proc. CVPR, pages 1304-1312, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.482, + 0.242 + ], + "angle": 0, + "content": "[3] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.482, + 0.285 + ], + "angle": 0, + "content": "[4] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[5] Padraig Cunningham and Sarah Jane Delany. K-nearest neighbour classifiers-a tutorial. ACM Computing Surveys, 54(6):1-25, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.331, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.374, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.545, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[10] Junhong Gao, Seon Joo Kim, and Michael S Brown. Constructing image panoramas using dual-homography warping. In Proc. CVPR, pages 49-56, 2011. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.588, + 0.482, + 0.614 + ], + "angle": 0, + "content": "[11] Jochen Gast and Stefan Roth. Lightweight probabilistic deep networks. In Proc. CVPR, pages 3369-3378, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.617, + 0.482, + 0.67 + ], + "angle": 0, + "content": "[12] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Proc. NIPS, 27, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.674, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[13] Matthias Grundmann, Vivek Kwatra, Daniel Castro, and Irfan Essa. Calibration-free rolling shutter removal. In Proc. ICCP, pages 1-8, 2012. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.717, + 0.482, + 0.757 + ], + "angle": 0, + "content": "[14] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[15] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[16] Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. In Proc. ECCV, pages 652-667, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[17] Hai Jiang, Haipeng Li, Songchen Han, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Supervised homography learning" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "with realistic dataset generation. In Proc. ICCV, pages 9806-9815, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[18] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.246 + ], + "angle": 0, + "content": "[19] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.25, + 0.904, + 0.304 + ], + "angle": 0, + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proc. ICCV, pages 4015-4026, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.307, + 0.904, + 0.347 + ], + "angle": 0, + "content": "[21] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.35, + 0.904, + 0.39 + ], + "angle": 0, + "content": "[22] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[23] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Dmhomo: Learning homography with diffusion models. ACM Transactions on Graphics (TOG), 43(3):1-16, 2024. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.904, + 0.504 + ], + "angle": 0, + "content": "[24] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. International Journal of Computer Vision (IJCV), 132(6):2331-2349, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.506, + 0.905, + 0.56 + ], + "angle": 0, + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.563, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.606, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[27] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815, 2016. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.649, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[28] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), pages 1-14, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(3):2849-2863, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(6):7885-7899, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[31] Shuaicheng Liu, Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, and Bing Zeng. Unsupervised global and local" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7632" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "homography estimation with coplanarity-aware gan. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.135, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[33] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision (IJCV), 60(2):91-110, 2004. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[34] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters (RA-L), 3(3): 2346-2353, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[35] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.456 + ], + "angle": 0, + "content": "[37] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.482, + 0.498 + ], + "angle": 0, + "content": "[38] Xi Shen, François Darmon, Alexei A Efros, and Mathieu Aubry. Ransac-flow: generic two-stage image alignment. In Proc. ECCV, pages 618-637, 2020. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.498, + 0.482, + 0.527 + ], + "angle": 0, + "content": "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.092, + 0.905, + 0.119 + ], + "angle": 0, + "content": "with transformers. In Proc. CVPR, pages 8922-8931, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[40] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.18, + 0.905, + 0.223 + ], + "angle": 0, + "content": "[41] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.224, + 0.905, + 0.267 + ], + "angle": 0, + "content": "[42] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.268, + 0.905, + 0.324 + ], + "angle": 0, + "content": "[43] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.325, + 0.905, + 0.382 + ], + "angle": 0, + "content": "[44] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.383, + 0.905, + 0.425 + ], + "angle": 0, + "content": "[45] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Proc. ECCV, pages 467-483, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.427, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[46] Julio Zaragoza, Tat-Jun Chin, Michael S Brown, and David Suter. As-projective-as-possible image stitching with moving dlt. In Proc. CVPR, pages 2339–2346, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.47, + 0.905, + 0.527 + ], + "angle": 0, + "content": "[47] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 4, 5, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7633" + } + ] +] \ No newline at end of file diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_origin.pdf b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..979fbe4c03d6e338d7b289af8b25c2716aa1734c --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/46352bd4-609c-49a6-a102-ed3ec1467e54_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:81d9018c2a582fbaa3c568774a208fa410236b6cd3709fce38e8b9a6ad6557c1 +size 8238904 diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/full.md b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b6207ebb8308a567333fa1d9f04229d1c242ea7b --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/full.md @@ -0,0 +1,332 @@ +# Estimating 2D Camera Motion with Hybrid Motion Basis + +Haipeng Li $^{1*}$ Tianhao Zhou $^{1*}$ Zhanglei Yang $^{1}$ Yi Wu $^{2}$ Yan Chen $^{2}$ Zijing Mao $^{2}$ Shen Cheng $^{3}$ Bing Zeng $^{1}$ Shuaicheng Liu $^{1\dagger}$ $^{1}$ University of Electronic Science and Technology of China + $^{2}$ Xiaomi Corporation ${}^{3}$ Dexmal + +# Abstract + +Estimating 2D camera motion is a fundamental computer vision task that models the projection of 3D camera movements onto the 2D image plane. Current methods rely on either homography-based approaches, limited to planar scenes, or meshflow techniques that use grid-based local homographies but struggle with complex non-linear transformations. We introduce CamFlow, a novel framework that represents camera motion using hybrid motion bases: physical bases derived from camera geometry and stochastic bases for complex scenarios. Our approach includes a hybrid probabilistic loss function based on the Laplace distribution that enhances training robustness. For evaluation, we create a new benchmark by masking dynamic objects in existing optical flow datasets to isolate pure camera motion. Experiments show CamFlow outperforms state-of-the-art methods across diverse scenarios, demonstrating superior robustness and generalization in zero-shot settings. Code and datasets are available at our project page: https://lhaippp.github.io/CamFlow/. + +# 1. Introduction + +Estimating 2D camera motion, which involves recovering the projection of 3D rotation and translation onto 2D planes [14], is a cornerstone of computer vision. Given a 3D rotation matrix $\mathbf{R}$ and translation vector $\mathbf{t}$ , the camera motion $\mathbf{M}$ can be expressed as: + +$$ +\mathbf {M} = \mathbf {K} \left(\mathbf {R} + \mathbf {t} \frac {\mathbf {n} ^ {T}}{d}\right) \mathbf {K} ^ {- 1}, \tag {1} +$$ + +where $\mathbf{n}^T$ is the transpose of the normal vector to planes in the scene, $d$ denotes the distance from the camera center to each plane, and $\mathbf{K}$ represents the camera intrinsic matrix. The resulting 2D camera motion is inherently nonlinear due to its dependence on scene depth and plane geometry. In real-world images, scenes typically consist of + +![](images/861dde72e0baeb08c17b1d10ee3e4f76087567a8043dcd50e64dab534f14b634.jpg) +Figure 1. Comparison of camera motion estimation approaches in multi-plane scenes. Visualizations show warped source images overlaid on targets, with brighter areas in heatmaps indicating higher error. BasesHomo [44] captures only background motion, while MeshHomoGAN [31] improves accuracy through local grid-based homographies. Our CamFlow, with hybrid motion basis, achieves superior representation of complex camera motion. + +![](images/17bdb60c4321a858e1967d29ecf7798a797140a370a10243d6c98baf20c30afa.jpg) + +multiple depths and planes. As a result, different regions of an image undergo distinct transformations, leading to complex, non-linear motion patterns. This task is essential for various computational imaging applications, such as digital video stabilization [26], where accurate representation of camera motion directly enhances performance. Existing methods for modeling camera motion primarily fall into two categories: homography and meshflow-based approaches. + +Homography, a perspective transformation, aligns two views of a planar or nearly-planar 3D scene [14]. Traditional methods typically rely on artificially extracted and matched keypoints [33], excelling under standard conditions but often struggling in adverse scenarios (e.g., rain, snow, and low light) or when confronted with non-planar motion caused by depth parallax or dynamic objects. Recent deep learning approaches mitigate the reliance on keypoints, enabling direct estimation of homography from image pairs through robust, data-driven networks. A notable benchmark work, BasesHomo [44], reformulates the task + +as learning a linear combination of an 8-dimensional motion basis, pioneering a new direction in motion estimation. Nevertheless, they are limited because: a homography can only align a single plane. + +To address this limitation, MeshFlow [27] partitions the image into $N \times N$ grids, estimating a local homography for each cell and smoothing them to model non-linear camera motion. This approach performs well in scenes with multiple small-baseline depth variations and has become a popular choice for digital video stabilization [26, 27]. Furthermore, deep meshflow variants [28] demonstrate enhanced robustness under challenging conditions. However, a key limitation persists in current camera motion representation: increasing the number of grids enhances the ability to model non-linearity but raises optimization challenges [43]. + +In this work, we introduce CamFlow, a novel representation that models complex camera motion through hybrid motion bases. Our key insight is that the flow field resulting from the superposition of multiple homographies is inherently non-linear (as visualized in Fig. 3), enabling more sophisticated motion modeling beyond single-plane limitations. Building on this observation, we establish a comprehensive hybrid basis subspace comprising: + +- Physical Motion Bases: Derived from Taylor expansion of homographic transformations up to second-order terms, our 12 physical bases model fundamental geometric transformations (rotation, translation, scaling, and perspective) that capture essential camera motion patterns; +- Noisy Motion Bases: To model complex residual motion, we construct $K$ orthogonal components through SVD decomposition of randomly sampled homographies from a Gaussian distribution, effectively complementing the physical bases by capturing higher-order motion patterns. To stabilize the training procedure and simplify complex loss designs, we propose a hybrid probabilistic loss function that assumes motion models follow a Laplace distribution [42], facilitating robust and efficient optimization. + +In summary, CamFlow effectively represents complex 2D camera motion as illustrated in Fig. 1, capturing both background and foreground elements while accurately modeling depth-varying motion patterns that conventional approaches struggle to represent. To rigorously evaluate its performance, we introduce GHOF-Cam, a novel benchmark specifically designed for camera motion estimation by systematically masking dynamic objects and ill-posed occlusion regions in established optical flow datasets [24], thereby isolating pure camera-induced motion. Through comprehensive experiments across diverse datasets under both standard and challenging conditions, we demonstrate that CamFlow consistently outperforms state-of-the-art single-plane homography and multi-planar methods in both sparse and dense camera motion estimation tasks, exhibiting superior robustness and generalization capability in + +real-world scenarios. Our main contributions are: + +- A new hybrid motion representation that learns to model complex non-linear 2D camera motion through physically interpretable and stochastic motion bases. +- A novel probabilistic loss formulation based on Laplace distribution that simplifies training and stabilizes optimization without complex loss designs. +- A comprehensive benchmark for evaluating camera motion learning across diverse conditions. Experimental results confirm our approach's effectiveness, robustness, and generalization ability across real-world scenarios. + +# 2. Related works + +# 2.1. Homography Methods + +The traditional homography estimation follows three stages: feature detection (e.g., SIFT [33], ORB [35]), correspondence matching [5], and outlier rejection (e.g., RANSAC [9], MAGSAC++ [2]). Learning-based methods like LIFT [45], SuperPoint [7], and SOSNet [40] improve robustness. Optimization-based approaches [4, 8] iteratively refine homography estimates, while deep learning methods span supervised [3, 6, 17, 21, 23, 37] and unsupervised [15, 19, 34, 44, 47] frameworks. Unsupervised models have proven effective for real-world scenarios. Notable examples include CAHomo [47] and BaseSHom [44], which enhance feature extraction and motion constraints. HomoGAN [15] integrates GAN loss [12] and Transformers [32] for coarse-to-fine refinement. Despite this progress, homography remains a single-plane model, limiting its effectiveness in complex motion. In this paper, we introduce a hybrid motion-basis representation to model multi-plane, non-linear motion more effectively. + +# 2.2. Mesh-based Methods + +Mesh-based warping estimates local homographies per mesh cell, including dual-plane approaches [10], patchwise mixtures of homography [13], flexible warping techniques like APAP [46], Bundled Paths [26], grid-based methods [27], and cascade residual homography [38]. Deep learning variants include deepMeshFlow [43], MeshCA-Homo [29], which merges multi-resolution meshes, and BasesMesh [30], which applies motion bases per grid. MeshHomoGAN [31] incorporates a planarity-aware mechanism for local homography estimation. However, these methods still assume that local regions fit the homography relationship, limiting their ability to represent complex motion. To address this limitation, we propose a novel motion representation that combines multiple non-linear motion bases, enhancing performance. + +![](images/a2fc9670de9cef2c7ab5a3e2d6be45eb09ffee994f94f597dcec02f7ed7bf34a.jpg) +Figure 2. Our proposed motion estimation framework. Given image pair $(I_a, I_b)$ , features are extracted through a multi-scale pyramid and processed by the motion estimation transformer (MET) to compute weights for physical (blue) and noisy (red) motion bases. These weights linearly combine predefined motion bases to generate flow maps for warping. A mask generator predicts uncertainty masks $\mathbf{d}_{ab}$ and $\mathbf{d}_{ba}$ to reject unreliable regions, enhancing estimation robustness. + +# 3. Method + +# 3.1. Motion Basis + +In this work, we propose a novel motion representation through motion basis learning with deep networks to better represent the camera motion via Physical and Noisy bases. + +Physical Basis. Consider a pixel with homogeneous coordinates $\mathbf{P}(x,y) = [x,y,1]^T\in \mathbb{R}^3$ . The physical motion $m = [\Delta x,\Delta y,1]^T$ induced by a homography $\mathbf{H}$ is defined: + +$$ +m = \mathbf {H} \cdot \mathbf {P} (x, y) - \mathbf {P} (x, y), \tag {2} +$$ + +After normalization, the two-dimensional coordinates (x-axis and y-axis) of the motion become: + +$$ +\Delta x = \frac {h _ {1} x + h _ {2} y + h _ {3}}{h _ {7} x + h _ {8} y + 1} - x, \tag {3} +$$ + +$$ +\Delta y = \frac {h _ {4} x + h _ {5} y + h _ {6}}{h _ {7} x + h _ {8} y + 1} - y, +$$ + +where $h_1, \ldots, h_8$ denote the elements of matrix $\mathbf{H}$ , and $h_9$ is constrained to 1. By applying a Taylor expansion, this motion can be mapped to another subspace. For instance, + +expanding $\Delta x$ around the point $(x,y) = (0,0)$ gives: + +$$ +\Delta x = \frac {\left(h _ {1} - 1\right) x + h _ {2} y + h _ {3} - h _ {7} x ^ {2} - h _ {8} x y}{h _ {7} x + h _ {8} y + 1} \tag {4} +$$ + +$$ +\approx w _ {1} \cdot 1 + w _ {2} \cdot x + w _ {3} \cdot y + w _ {4} \cdot x y \tag {5} +$$ + +$$ ++ w _ {5} \cdot x ^ {2} + w _ {6} \cdot y ^ {2} + \Delta , \tag {6} +$$ + +where $w_{i}, i \in [1,6]$ are coefficients, the basis functions are $b = [1,x,y,xy,x^2,y^2]$ and $\Delta$ denotes the higher-order infinitesimal. Similarly, $\Delta y$ can be decomposed into this subspace. By combining the decompositions of both $\Delta x, \Delta y$ , the motion space can be represented using the 12 bases: + +$$ +\mathbf {F} = \left\{\left(b _ {i}, 0\right) \mid b _ {i} \in b \right\} \cup \left\{\left(0, b _ {i}\right) \mid b _ {i} \in b \right\}, \tag {7} +$$ + +where $b_{i}$ denotes the $i$ -th element in $b$ . Each basis can be transformed into an optical flow according to the image coordinate, as shown in Fig. 2 (Physical motion bases). + +Non-Linearity of Adding Bases. A key finding is that combining flow fields derived from different homographies produces a non-linear motion field that cannot be represented by any single homography. This challenges the linear basis assumption in prior work [44]. As shown in Fig. 3, when two homography-derived flows (flow1 and flow2) are added together, the resulting flow (flow3) differs from the flow derived from multiplying the original homography matrices (homo3). Additionally, attempting to solve for a homography from points sampled on flow3 yields inconsistent + +![](images/9a6325c9dde6565015401ca32f5c72a94dd1180fb1776522ef62d649ab436f46.jpg) +Figure 3. Non-linearity of flow addition. Two homography matrices generate flow1 and flow2. Adding these flows (flow3) differs from the flow derived by multiplying the original homography matrices (homo3). When sampling points from flow3 to solve for a homography, we get inconsistent solutions, proving that combined flow fields cannot be represented by a single homography. + +solutions, confirming that the combined flow cannot be represented by any single homography. This observation motivates our approach, which leverages more bases to model complex camera motion. + +Stochastic Basis. While the physical bases effectively capture fundamental motion patterns, the complete space of camera motions is infinitely dimensional when considering higher-order Taylor expansions. Exhaustively modeling all possible bases becomes computationally intractable. To address this challenge, we leverage the expressive power of random sampling to complement our physical bases. Specifically, we generate $K$ random $3 \times 3$ matrices $\left\{\mathbf{H}^{(k)}\right\}_{k=1}^{K} \subseteq \mathbb{R}^{3 \times 3}$ , where each matrix is formulated as: + +$$ +\mathbf {H} = \left\{h _ {i} \right\} _ {i = 1} ^ {9}, \text {w h e r e} \quad h _ {i} = \left\{ \begin{array}{l l} \epsilon_ {i} \sim \mathcal {N} (0, 1), & 1 \leq i \leq 8, \\ 1, & i = 9. \end{array} \right. \tag {8} +$$ + +Following BasesHomo [44], we convert random matrices into flows and apply singular value decomposition (SVD) to extract principal components. This process yields $N - 12$ stochastic bases that capture diverse motion patterns beyond physical bases, as illustrated in Fig. 2 (Noisy motion bases). These stochastic bases, combined with the 12 physical basis vectors described earlier, form a comprehensive set of $N$ motion bases that enhances our ability to model complex non-linear camera motions. + +# 3.2. Network Structure + +The network is illustrated in Fig. 2. Following practices from previous work [15, 47], we process video frames by: (1) random cropping to $320 \times 576$ patches, (2) converting to grayscale, (3) projecting into a shallow feature space to handle luminance variations, and (4) generating a 3-layer feature pyramid for multi-scale processing. Then we propose a Motion Estimation Transformer (MET). The MET employs a specialized architecture to separately predict weights for both physical bases (12 weights) and stochastic bases ( $N - 12$ weights), yielding bidirectional weights $\hat{\mathbf{w}}_{ab}$ and $\hat{\mathbf{w}}_{ba}$ . The final camera motion is computed through a + +linear combination of the predicted weights and their corresponding motion bases, producing bidirectional dense motion fields $\hat{\mathbf{m}}_{ab}$ and $\hat{\mathbf{m}}_{ba}$ . + +It is noteworthy that the confidence masks $\mathbf{d}_{ab}$ and $\mathbf{d}_{ba}$ are also crucial in predicting camera motion, as they effectively filter out dynamic objects. First, we apply the predicted $\hat{\mathbf{m}}_{ab}$ and $\hat{\mathbf{m}}_{ba}$ to $f_{a}$ and $f_{b}$ , respectively, obtaining warped feature maps $f_{a}^{\prime}$ and $f_{b}^{\prime}$ . The mask network $\mathcal{M}$ [22] inputs the concatenated features to generate weighted maps that highlight well-aligned regions: + +$$ +\mathbf {d} _ {a b} = \mathcal {M} \left(\left[ f _ {a}, f _ {b} ^ {\prime} \right]\right), \mathbf {d} _ {b a} = \mathcal {M} \left(\left[ f _ {b}, f _ {a} ^ {\prime} \right]\right). \tag {9} +$$ + +# 3.3. Loss Function + +Probabilistic Motion Modeling. Camera motion estimation faces a fundamental challenge: distinguishing between camera-induced and object motion in the scene. We therefore formulate our approach as a probabilistic model that explicitly accounts for uncertainty in motion estimation. + +Building on previous findings that 2D motion follows a Laplace distribution [11, 16, 42], we model the conditional probability of the true camera motion given our prediction: + +$$ +p \left(\mathbf {m} _ {a b} \mid \hat {\mathbf {m}} _ {a b}; \mathbf {d} _ {a b}\right), \tag {10} +$$ + +where $\hat{\mathbf{m}}_{ab}$ is our predicted motion field and $\mathbf{d}_{ab}$ represents our confidence in each pixel's motion estimate. Higher confidence values indicate pixels likely following camera motion, while lower values suggest pixels belonging to independently moving objects. + +Laplace Distribution Model. We model the probability density using two conditionally independent Laplace distributions for the horizontal and vertical components: + +$$ +\begin{array}{l} \mathcal {L} \left(\mathbf {m} _ {a b} \mid \hat {\mathbf {m}} _ {a b}; \mathbf {d} _ {a b}\right) = \prod \left(\frac {1}{\sqrt {2 \sigma^ {2}}} e ^ {- \sqrt {\frac {2}{\sigma^ {2}}} | u - \mu_ {u} |} \right. \tag {11} \\ \left. \cdot \frac {1}{\sqrt {2} \sigma^ {2}} e ^ {- \sqrt {\frac {2}{\sigma^ {2}}} | v - \mu_ {v} |}\right), \\ \end{array} +$$ + +where $(u,v)$ are the ground-truth motion components, $(\mu_u,\mu_v)$ are the predicted motion components, and $\sigma^2$ is derived from our confidence mask $\mathbf{d}_{ab}$ . This formulation allows our model to express both its prediction and its uncertainty about that prediction. + +Hybrid Loss Strategy. Another challenge in camera motion estimation is the scarcity of ground-truth labels. To overcome this, we introduce a hybrid loss strategy with two components: 1) Motion supervision loss $(\ell_{NLL_m})$ : we generate pseudo-labels using existing methods and apply negative log-likelihood (NLL) loss in both forward and backward directions: + +$$ +\ell_ {N L L _ {m}} = - \log p \left(\mathbf {m} _ {a b} \mid \hat {\mathbf {m}} _ {a b}; \mathbf {d} _ {a b}\right) - \log p \left(\mathbf {m} _ {b a} \mid \hat {\mathbf {m}} _ {b a}; \mathbf {d} _ {b a}\right). \tag {12} +$$ + +2) Photometric loss $(\ell_{NLL_p})$ : We apply the same probabilistic framework to enforce consistency between warped features. Given image features $f_{a}$ and $f_{b}$ , we use our predicted + +![](images/5a40caf1a186e281aa23eb221e410ebb71b84dc83aeea338b19982c1bdce3251.jpg) +Figure 4. Using the GHOF benchmark, we leverage SAM to generate semantic maps and manually identify dynamic objects such as cars and people, producing corresponding masks. These masks are then dilated to encompass occlusion regions, reducing ill-posed artifacts. Finally, the masks are applied to both the images and optical flow, isolating camera-induced motion and constructing a camera-motion-only dataset. + +motion to produce warped features $f_{a}^{\prime}$ and $f_{b}^{\prime}$ : + +$$ +\ell_ {N L L _ {p}} = - \log p \left(f _ {a} \mid f _ {b} ^ {\prime}; \mathbf {d} _ {a b}\right) - \log p \left(f _ {b} \mid f _ {a} ^ {\prime}; \mathbf {d} _ {b a}\right). \tag {13} +$$ + +Adaptive loss balancing: to ensure stable training despite the different scales of our loss components, we dynamically balance them using: + +$$ +\ell_ {\text {o v e r a l l}} = \ell_ {N L L _ {p}} + \mathbf {w} \times \frac {\left| \ell_ {N L L _ {p}} \right|}{\left| \ell_ {N L L _ {m}} \right|} \cdot \ell_ {N L L _ {m}}, \tag {14} +$$ + +where $\mathbf{w}$ is a predefined weight. + +# 4. Experiments + +# 4.1. Dataset + +We evaluate our method on: CAHomo [47] and GHOF [24]. We train on CAHomo (460K training pairs, 4.2K test pairs across regular, low-texture, low-light, and foreground scenes) with additional generated samples [23], and conduct zero-shot testing on GHOF (256 test pairs in Regular, Foggy, Low-light, Rainy, and Snowy conditions). + +GHOF-Cam Benchmark. To isolate camera motion from dynamic scene elements, we propose a camera-motion-specific benchmark derived from GHOF. We employ the Segment Anything Model (SAM) [20] to generate semantic maps, from which we identify dynamic objects (e.g., cars, people) and create corresponding masks. These masks are dilated to encompass occlusion regions, mitigating ill-posed artifacts at object boundaries. For edge occlusions not detected by semantic segmentation, we mainly utilize the ground-truth homography to identify black edge regions as additional masks. The combined masks are then applied to both input images and ground-truth optical flow, resulting in a benchmark that exclusively captures camera-induced motion, as illustrated in Fig. 4. + +# 4.2. Comparison with Existing Methods + +We assess CamFlow against a comprehensive set of methods across three primary categories. The first category encompasses feature-based methods: SIFT [33], ORB [35], SuperPoint[7] with SuperGlue (SPSG) [36], and LoFTR + +
MethodsAVGRELTLLSFLF
1)\( \mathcal{I}_{3\times 3} \)6.707.757.657.217.533.39
2)SIFT[33] + RANSAC[9]1.410.301.344.030.810.57
3)SIFT[33] + MAGSAC[1]1.340.311.723.390.800.47
4)ORB[35] + RANSAC[9]1.480.852.591.671.101.24
5)ORB[35] + MAGSAC[1]1.690.973.341.581.151.40
6)SPSG[7, 36] + RANSAC[9]0.710.410.870.720.800.75
7)SPSG[7, 36] + MAGSAC[1]0.630.360.790.700.710.70
8)LoFTR[39] + RANSAC[9]1.440.562.701.361.051.52
9)LoFTR[39] + MAGSAC[1]1.390.552.571.331.051.41
10)DHN[6]2.871.514.482.762.623.00
11)LocalTrans[37]4.214.094.844.555.302.25
12)IHN[3]4.674.855.545.105.042.84
13)RealSH[17]0.340.220.350.440.420.29
14)DMHomo[23]0.310.190.330.400.380.28
15)CAHomo[47]0.880.731.011.030.920.70
16)BasesHomo[44]0.500.290.540.650.610.41
17)HomoGAN[15]0.390.220.410.570.440.31
18)Ours*0.330.200.330.410.390.30
19)Ours0.320.190.320.390.390.31
+ +Table 1. The benchmark consists of 5 distinct scenarios, namely regular (RE), low-texture (LT), low-light (LL), small foreground (SF), and large foreground (LF). The point matching errors (PME) on the test set of CAHomo [47] are presented. + +
1)MethodsAVGREFOGLLRAINSNOW
2)I3×35.223.656.695.884.904.96
3)SIFT[33]2.820.602.437.090.613.37
4)SPSG[7, 36]3.073.991.576.880.792.16
5)CAHomo[47]2.812.022.034.562.842.61
6)BasesHomo[44]1.741.390.974.120.661.58
7)Meshflow[27]2.151.092.215.570.441.69
8)HM_Mix[13]4.351.024.038.751.536.42
9)RANSAC-F[38]3.262.813.145.122.213.04
10)Ours1.101.080.742.150.461.05
+ +Table 2. To evaluate generalizability, we compute the end point errors (EPE) of pre-trained models from Table 1 on our proposed GHOF-Cam benchmark. + +
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT[33]4.800.594.4712.100.626.20
4)SPSG[7, 36]4.473.542.2110.660.835.10
5)DHN[6]6.616.046.027.686.996.32
6)LocalTrans[37]5.724.066.495.955.786.34
7)IHN[3]8.177.108.719.346.579.13
8)RealSH[17]1.721.600.884.420.431.28
9)DMHomo[23]1.750.640.854.160.392.74
10)CAHomo[47]3.874.103.846.991.273.17
11)BasesHomo[44]2.282.021.434.900.782.29
12)HomoGAN[15]1.951.730.603.950.473.02
13)Ours1.231.150.962.690.400.93
+ +Table 3. To evaluate generalizability, we compute the point matching errors (PME) on the GHOF [24] test set using pre-trained models from Table 1. + +[39], each evaluated with two outlier rejection techniques: RANSAC [9] and MAGSAC [1]. The second category includes supervised learning approaches: DHN [6], Local-Trans [37], IHN [3], RealSH [17], and DMHomo [23]. The third category comprises unsupervised methods: CAHomo + +
MethodAVGREFOGDARKRAINSNOW
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
\( \mathcal{I}_{3\times 3} \)24.050.74030.083621.060.69000.075026.570.77110.082125.700.85060.078521.530.53350.141125.370.85620.0412
GT-Homo32.780.91870.057028.390.86970.054935.230.95080.049231.880.94050.057530.110.85110.103338.310.98140.0199
SIFT28.440.90740.078129.230.91480.054529.420.90160.076827.370.90740.098230.000.86320.105526.160.94970.0558
SPSG28.010.86970.079621.830.75930.088630.880.90490.064527.600.90190.096628.860.82700.110330.880.95560.0379
CAHomo25.290.78370.084122.670.73410.080527.510.80480.075126.120.87430.084622.950.61300.142027.200.89240.0384
BasesHomo29.610.90260.067225.080.85220.066631.060.91700.062730.050.93030.070229.580.85120.107132.300.96220.0292
MeshFlow29.910.92390.068828.570.92160.057628.680.92800.074229.410.92540.077430.680.87470.104932.230.97000.0298
HM_Mix25.770.88960.088226.090.87210.059626.560.87530.088226.430.90370.100228.200.86720.110721.580.92960.0820
RANSAC-F26.040.83480.089026.090.88120.066529.220.89440.080127.290.90310.092321.680.55850.149525.900.93710.0566
Ours32.090.91420.057527.080.86150.055834.170.93710.051232.360.94210.056530.520.86080.102136.350.96920.0218
+ +Table 4. Quantitative comparison of different methods on various environmental conditions. We present three key perception metrics: PSNR (higher is better), SSIM (higher is better), and LPIPS (lower is better). Our method consistently outperforms existing approaches across all scenarios and metrics. + +[47], BasesHomo [44], and HomoGAN [15]. + +For multi-plane camera motion modeling, we compare against both traditional approaches (MeshFlow [27], Homography Mixture [13], RANSAC-Flow [38]) and unsupervised deep methods (BasesMesh [30] and MeshHomogAN [31]). Regarding pre-training, DHN, LocalTrans and IHN use the MS-COCO dataset [25], while other deep learning methods are pre-trained on CAHomo. Additional qualitative results and visual comparisons are available on: https://lhaippp.github.io/CamFlow/. + +# 4.2.1. Quantitative Comparison. + +Sparse camera motion. We evaluate using the points matching error (PME), which measures the average geometric distance between transformed source points and their corresponding ground-truth target points. Table 1 presents the quantitative performance of our method alongside various homography estimation approaches on the CAHomo test set, categorized as: feature-based methods (rows 2-9), supervised approaches (rows 10-14), and unsupervised ones (rows 15-17). The $\mathcal{I}_{3\times 3}$ baseline (row 1) represents the distance between point pairs without transformation. + +Our CamFlow method achieves superior results across multiple categories, outperforming the leading unsupervised method, HomoGAN, by reducing PME by $17.95\%$ (from 0.39 to 0.32). While feature-based methods like SIFT+RANSAC excel in regular (RE) scenes with abundant texture and keypoints, CamFlow surpasses them with a $36.67\%$ improvement (reducing error by 0.11). To balance benchmark performance with generalization capability, we include Ours\* (row 18), which represents an early-stopped training model that delivers better generalizability across datasets, even though it slightly underperforms our fully-trained model (row 19) on the CAHomo benchmark. + +In small foreground (SF) and large foreground (LF) scenarios, where dynamic objects disrupt camera motion estimation, our probabilistic loss and confidence masking yield lower PMEs compared to methods like CAHomo and HomoGAN, which employ explicit outlier rejection masks for robustness. In low-texture (LT) and low-light (LL) conditions, learning-based approaches generally exhibit superior + +resilience due to their keypoint-free strategies, particularly among unsupervised techniques. However, these scenes often contain homogeneous regions [41], occupying large image areas and reducing photometric differences, which limit unsupervised methods' effectiveness. By contrast, CamFlow excels in LT and LL. + +Generalization experiment. Evaluating cross-dataset generalization presents a significant challenge in motion estimation, particularly for applications where real-world video often differs substantially from training data. We assess CamFlow's generalization capabilities in two ways: (1) for dense camera motion, we evaluate against traditional single-plane, multi-planar, and deep learning approaches using our proposed GHOF-Cam benchmark (Table 2 and Table 4); and (2) for sparse camera motion, we compare against traditional and homography-based methods on the original GHOF benchmark [24] (Table 3). + +Table 2 presents End Point Error (EPE) results on the GHOF-Cam benchmark, which represents ground-truth camera motion. We evaluate against homography-based methods (rows 3-5) and non-single-plane approaches (rows 6-9). As illustrated in Fig. 3, we classify BasesHomo [44] as a multi-plane rather than single-homography method. The results demonstrate that CamFlow outperforms competing methods across nearly all categories, showcasing exceptional zero-shot capability in capturing non-linear camera motion patterns. Additionally, Table 4 presents perceptual quality metrics (PSNR, SSIM, and LPIPS) across various conditions. We also include ground-truth homography results for reference (GT-Homo). Our method, CamFlow, achieves state-of-the-art performance across multiple categories, consistently outperforming all baselines, particularly in challenging conditions such as snowy scenes. Notably, CamFlow approaches the performance of ground-truth homography, with only a $0.69\mathrm{dB}$ difference in PSNR, 0.0045 in SSIM and 0.0005 in LPIPS, while significantly surpassing the second-best method, MeshFlow. These results support CamFlow's potential for applications requiring high-quality visual alignment. + +Table 3 presents compelling evidence of CamFlow's generalization capability on the GHOF benchmark. Our + +![](images/3f2f9b429c92677a72b90585fddde43ca0b71d10471518a0aa11d8762c006578.jpg) +Figure 5. Qualitative results of CamFlow and methods from each category (i.e., supervised, unsupervised, and multi-homography) on the CAHomo testset [47]. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in. + +![](images/d4e7535b869100088323413231d5983607b9ec659392b10d32cb1745488ce234.jpg) +Figure 6. Qualitative results of our method and the best-performing generalizable methods from learning-based categories (i.e., supervised, unsupervised, and multi-homography) on the GHOF testset [24]. The examples are arranged from top to bottom and include RE, Fog, and Dark. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in. + +method achieves the lowest average PME (1.23), representing a $28.5\%$ improvement over the previous best supervised method (RealSH at 1.72) and a $36.9\%$ improvement over the leading unsupervised approach (HomoGAN at 1.95). This performance advantage extends across all environmental conditions, with particularly significant gains in challenging scenarios. In low-light (LL) conditions, CamFlow reduces error by $39.1\%$ compared to RealSH (from 4.42 to 2.69), while in snowy (SNOW) environments, it achieves a remarkable $69.2\%$ error reduction compared to HomoGAN (from 3.02 to 0.93). These results confirm that CamFlow enables robust zero-shot transfer to unseen datasets, even under diverse and challenging environmental conditions. + +# 4.3. Qualitative Comparison + +Fig. 5 and 6 present qualitative results of CamFlow alongside competing methods on the CAHomo and GHOF test + +sets. Dynamic visualizations are available on our project page. For visualization, we employ red-blue ghosting and alignment heat maps following [18]. After transforming the source frame using estimated camera flow, misaligned regions appear as red and blue ghosting. We highlight specific areas with red boxes and generate alignment heat maps where brighter regions indicate poorer alignment. + +In both figures, the "Identity" column shows source and target images overlaid without warping. For CAHomo (Fig. 5), we focus on evaluating background camera motion modeling capabilities. We selected challenging test cases featuring extremely low-light conditions (first row) and sophisticated motion patterns (parallax and depth variation), comparing against leading methods from three categories: supervised (RealSH), unsupervised (HomoGAN), and deep meshflow (BasesMesh, MeshHomoGAN). CamFlow achieves superior alignment in background regions. + +![](images/d94fd4b5de40cdfec6bd1ad59315be78bb82041cde3c99570c013173ef00e74b.jpg) +Figure 7. Uncertainty masks and corresponding original images. The brighter the mask, the higher the uncertainty. The masks effectively highlight dynamic objects, indicating regions where camera motion estimation is less reliable. + +![](images/28bfc84b23cf4b9b223c571ca1794ae20203f78a5ecfd0dc2d2d8e8cc5c91d1e.jpg) + +The GHOF benchmark (Fig. 6) highlights CamFlow's advantages in handling complex camera motion and its zero-shot generalization capabilities. The first row shows a scene with parallax and foreground motion that challenges homography-based methods and reduces meshflow accuracy. The second row demonstrates CamFlow's generalization to foggy conditions unseen in training data, where depth variation complicates motion estimation. The last row presents an extreme case with motion blur, large parallax, and dynamic objects. In all scenarios, CamFlow delivers robust results where competing methods struggle, confirming its effectiveness in modeling complex camera motion patterns across diverse environmental conditions. + +Foreground Masks. Fig. 7 illustrates the uncertainty masks. We observe that the network effectively identifies dynamic objects and assigns higher uncertainty to regions containing these objects. This design helps focus the learning process on areas where motion is most relevant, while reducing noise from dynamic object regions, thereby enhancing the accuracy of camera motion estimation. + +# 4.4. Ablation Studies + +4.4.1. Motion Basis + +
CAHomoGHOFGHOF-CamParamsInference Time
8 Bases0.371.681.452.658M76.42ms
12 Bases0.361.541.232.658M75.38ms
24 Bases0.331.231.102.660M79.63ms
200 Bases0.331.271.072.677M99.28ms
+ +Table 5. Performance comparison under different numbers of motion bases on three benchmarks. Results demonstrate the effectiveness of combining physical and stochastic motion bases, with 24 bases providing optimal balance between accuracy and computational efficiency. + +In Table 5, we evaluate with varying numbers of motion bases across three benchmarks: CAHomo (trained) and GHOF/GHOF-Cam (zero-shot). The average results demonstrate that: 1) Increasing physical bases from 8 to 12 improves performance across all benchmarks; 2) Introducing additional hybrid bases (24 total) yields further enhancements, particularly for generalization (GHOF and GHOF-Cam). Notably, while expanding to 200 bases provides marginal improvements on some benchmarks, it increases inference time by $24.7\%$ . We therefore adopt 24 bases in our final model as the optimal balance between ac + +curacy and computational efficiency. + +4.4.2. Hybrid Probabilistic Loss + +
\( {\ell }_{NLLm} \)\( {\ell }_{NLLp} \)CAHomoGHOFGHOF-Cam
✓0.412.212.13
✓0.361.581.42
✓✓0.331.231.10
+ +Table 6. Ablation study of different loss function combinations across three benchmarks. Results demonstrate that the hybrid approach combining motion loss and photometric loss achieves superior performance compared to individual loss components. + +Table 6 evaluates our probabilistic loss components across three benchmarks. Using only motion loss yields limited generalization performance, because pseudo motion labels provide approximate supervision. The photometric loss alone performs substantially better, particularly on zero-shot datasets, consistent with findings from prior unsupervised methods. However, our hybrid approach combining both losses achieves the best results across all benchmarks, with improvements on GHOF-Cam (22.5% error reduction compared to photometric-only). We believe motion labels provide coarse guidance while photometric loss enables fine-grained refinement, resulting in more accurate and generalizable camera motion estimation. + +# 5. Conclusion + +CamFlow presents a novel motion representation for modeling 2D camera motion using a hybrid motion basis approach. We identify a fundamental issue: superposing homographies by simply adding flow fields introduces nonlinear interactions, contradicting the assumption that homographies can be expressed as linear combinations of 8 basis flows. By expanding the previous 8-dimensional motion basis into a higher-dimensional space with both physical and stochastic motion bases, CamFlow effectively captures complex, non-linear motion patterns. The proposed probabilistic loss function enhances training stability. Our newly introduced GHOF-Cam benchmark demonstrates that CamFlow surpasses state-of-the-art homography and meshflow methods, showcasing superior robustness and generalization. We hope this work opens new avenues for camera motion modeling in video processing applications. Codebase, pre-trained models, and benchmark data are released at https://lhaippp.github.io/CamFlow/. + +Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (NSFC) under grants No. 62372091 and No. 62031009, and in part by the Hainan Province Key R&D Program under grant No. ZDYF2024(LALH)001. We would also like to thank Hao Xu, Mingbo Hong, Hai Jiang, Xinglong Luo, Nianjin Ye, and the anonymous reviewers for their valuable discussions, constructive suggestions, and helpful feedback. + +# References + +[1] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 5 +[2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proc. CVPR, pages 1304-1312, 2020. 2 +[3] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5 +[4] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2 +[5] Padraig Cunningham and Sarah Jane Delany. K-nearest neighbour classifiers-a tutorial. ACM Computing Surveys, 54(6):1-25, 2021. 2 +[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 2, 5 +[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5 +[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2 +[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5 +[10] Junhong Gao, Seon Joo Kim, and Michael S Brown. Constructing image panoramas using dual-homography warping. In Proc. CVPR, pages 49-56, 2011. 2 +[11] Jochen Gast and Stefan Roth. Lightweight probabilistic deep networks. In Proc. CVPR, pages 3369-3378, 2018. 4 +[12] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Proc. NIPS, 27, 2014. 2 +[13] Matthias Grundmann, Vivek Kwatra, Daniel Castro, and Irfan Essa. Calibration-free rolling shutter removal. In Proc. ICCP, pages 1-8, 2012. 2, 5, 6 +[14] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 1 +[15] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 4, 5, 6 +[16] Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. In Proc. ECCV, pages 652-667, 2018. 4 +[17] Hai Jiang, Haipeng Li, Songchen Han, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Supervised homography learning + +with realistic dataset generation. In Proc. ICCV, pages 9806-9815, 2023. 2, 5 +[18] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 7 +[19] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2 +[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proc. ICCV, pages 4015-4026, 2023. 5 +[21] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2 +[22] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 4 +[23] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Dmhomo: Learning homography with diffusion models. ACM Transactions on Graphics (TOG), 43(3):1-16, 2024. 2, 5 +[24] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. International Journal of Computer Vision (IJCV), 132(6):2331-2349, 2024. 2, 5, 6, 7 +[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 6 +[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1, 2 +[27] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815, 2016. 2, 5, 6 +[28] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), pages 1-14, 2022. 2 +[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(3):2849-2863, 2022. 2 +[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(6):7885-7899, 2023. 2, 6 +[31] Shuaicheng Liu, Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, and Bing Zeng. Unsupervised global and local + +homography estimation with coplanarity-aware gan. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2024. 1, 2, 6 +[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2 +[33] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision (IJCV), 60(2):91-110, 2004. 1, 2, 5 +[34] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters (RA-L), 3(3): 2346-2353, 2018. 2 +[35] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 2, 5 +[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 5 +[37] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5 +[38] Xi Shen, François Darmon, Alexei A Efros, and Mathieu Aubry. Ransac-flow: generic two-stage image alignment. In Proc. ECCV, pages 618-637, 2020. 2, 5, 6 +[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching + +with transformers. In Proc. CVPR, pages 8922-8931, 2021. 5 +[40] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2 +[41] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 6 +[42] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2, 4 +[43] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 2 +[44] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 1, 2, 3, 4, 5, 6 +[45] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Proc. ECCV, pages 467-483, 2016. 2 +[46] Julio Zaragoza, Tat-Jun Chin, Michael S Brown, and David Suter. As-projective-as-possible image stitching with moving dlt. In Proc. CVPR, pages 2339–2346, 2013. 2 +[47] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 4, 5, 6, 7 \ No newline at end of file diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/images.zip b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7ebf1f2acd81d3117e573c9e6e92ff66351d14d3 --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:082082dbca7fcce55fb9d11769f4d3d7a2d5296d7d45c33332e0c5bb0368b248 +size 770010 diff --git a/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/layout.json b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..acc180e3722e1fd114123866c52bda6dd593ddda --- /dev/null +++ b/2025/Estimating 2D Camera Motion with Hybrid Motion Basis/layout.json @@ -0,0 +1,8380 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "spans": [ + { + "bbox": [ + 127, + 103, + 483, + 121 + ], + "type": "text", + "content": "Estimating 2D Camera Motion with Hybrid Motion Basis" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "spans": [ + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": "Haipeng Li" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Tianhao Zhou" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Zhanglei Yang" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Yi Wu" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Yan Chen" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Zijing Mao" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Shen Cheng" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Bing Zeng" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": " Shuaicheng Liu" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": "University of Electronic Science and Technology of China \n" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": "Xiaomi Corporation " + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "inline_equation", + "content": "{}^{3}" + }, + { + "bbox": [ + 113, + 142, + 494, + 200 + ], + "type": "text", + "content": "Dexmal" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 251, + 296, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 296, + 479 + ], + "type": "text", + "content": "Estimating 2D camera motion is a fundamental computer vision task that models the projection of 3D camera movements onto the 2D image plane. Current methods rely on either homography-based approaches, limited to planar scenes, or meshflow techniques that use grid-based local homographies but struggle with complex non-linear transformations. We introduce CamFlow, a novel framework that represents camera motion using hybrid motion bases: physical bases derived from camera geometry and stochastic bases for complex scenarios. Our approach includes a hybrid probabilistic loss function based on the Laplace distribution that enhances training robustness. For evaluation, we create a new benchmark by masking dynamic objects in existing optical flow datasets to isolate pure camera motion. Experiments show CamFlow outperforms state-of-the-art methods across diverse scenarios, demonstrating superior robustness and generalization in zero-shot settings. Code and datasets are available at our project page: https://lhaippp.github.io/CamFlow/." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 500, + 135, + 511 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 135, + 511 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 135, + 511 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": "Estimating 2D camera motion, which involves recovering the projection of 3D rotation and translation onto 2D planes [14], is a cornerstone of computer vision. Given a 3D rotation matrix " + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{R}" + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": " and translation vector " + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{t}" + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": ", the camera motion " + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 520, + 296, + 579 + ], + "type": "text", + "content": " can be expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 116, + 584, + 295, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 584, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 116, + 584, + 295, + 611 + ], + "type": "interline_equation", + "content": "\\mathbf {M} = \\mathbf {K} \\left(\\mathbf {R} + \\mathbf {t} \\frac {\\mathbf {n} ^ {T}}{d}\\right) \\mathbf {K} ^ {- 1}, \\tag {1}", + "image_path": "a401dba94f88d205e887d4e991cb0a0f94d826e689e1824249c089b42868222d.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{n}^T" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " is the transpose of the normal vector to planes in the scene, " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " denotes the distance from the camera center to each plane, and " + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 616, + 296, + 688 + ], + "type": "text", + "content": " represents the camera intrinsic matrix. The resulting 2D camera motion is inherently nonlinear due to its dependence on scene depth and plane geometry. In real-world images, scenes typically consist of" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 227, + 435, + 378 + ], + "blocks": [ + { + "bbox": [ + 317, + 227, + 435, + 378 + ], + "lines": [ + { + "bbox": [ + 317, + 227, + 435, + 378 + ], + "spans": [ + { + "bbox": [ + 317, + 227, + 435, + 378 + ], + "type": "image", + "image_path": "861dde72e0baeb08c17b1d10ee3e4f76087567a8043dcd50e64dab534f14b634.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 388, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 464 + ], + "type": "text", + "content": "Figure 1. Comparison of camera motion estimation approaches in multi-plane scenes. Visualizations show warped source images overlaid on targets, with brighter areas in heatmaps indicating higher error. BasesHomo [44] captures only background motion, while MeshHomoGAN [31] improves accuracy through local grid-based homographies. Our CamFlow, with hybrid motion basis, achieves superior representation of complex camera motion." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 436, + 228, + 553, + 378 + ], + "blocks": [ + { + "bbox": [ + 436, + 228, + 553, + 378 + ], + "lines": [ + { + "bbox": [ + 436, + 228, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 436, + 228, + 553, + 378 + ], + "type": "image", + "image_path": "17bdb60c4321a858e1967d29ecf7798a797140a370a10243d6c98baf20c30afa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 579 + ], + "type": "text", + "content": "multiple depths and planes. As a result, different regions of an image undergo distinct transformations, leading to complex, non-linear motion patterns. This task is essential for various computational imaging applications, such as digital video stabilization [26], where accurate representation of camera motion directly enhances performance. Existing methods for modeling camera motion primarily fall into two categories: homography and meshflow-based approaches." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Homography, a perspective transformation, aligns two views of a planar or nearly-planar 3D scene [14]. Traditional methods typically rely on artificially extracted and matched keypoints [33], excelling under standard conditions but often struggling in adverse scenarios (e.g., rain, snow, and low light) or when confronted with non-planar motion caused by depth parallax or dynamic objects. Recent deep learning approaches mitigate the reliance on keypoints, enabling direct estimation of homography from image pairs through robust, data-driven networks. A notable benchmark work, BasesHomo [44], reformulates the task" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 134, + 703 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 703, + 243, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 243, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 243, + 713 + ], + "type": "text", + "content": "† Corresponding author (liushuaicheng@uestc.edu.cn)." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7624" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "as learning a linear combination of an 8-dimensional motion basis, pioneering a new direction in motion estimation. Nevertheless, they are limited because: a homography can only align a single plane." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "content": "To address this limitation, MeshFlow [27] partitions the image into " + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "inline_equation", + "content": "N \\times N" + }, + { + "bbox": [ + 54, + 121, + 295, + 253 + ], + "type": "text", + "content": " grids, estimating a local homography for each cell and smoothing them to model non-linear camera motion. This approach performs well in scenes with multiple small-baseline depth variations and has become a popular choice for digital video stabilization [26, 27]. Furthermore, deep meshflow variants [28] demonstrate enhanced robustness under challenging conditions. However, a key limitation persists in current camera motion representation: increasing the number of grids enhances the ability to model non-linearity but raises optimization challenges [43]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 255, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 350 + ], + "type": "text", + "content": "In this work, we introduce CamFlow, a novel representation that models complex camera motion through hybrid motion bases. Our key insight is that the flow field resulting from the superposition of multiple homographies is inherently non-linear (as visualized in Fig. 3), enabling more sophisticated motion modeling beyond single-plane limitations. Building on this observation, we establish a comprehensive hybrid basis subspace comprising:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 352, + 295, + 521 + ], + "type": "list", + "angle": 0, + "index": 5, + "blocks": [ + { + "bbox": [ + 55, + 352, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 411 + ], + "type": "text", + "content": "- Physical Motion Bases: Derived from Taylor expansion of homographic transformations up to second-order terms, our 12 physical bases model fundamental geometric transformations (rotation, translation, scaling, and perspective) that capture essential camera motion patterns;" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 411, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 295, + 521 + ], + "type": "text", + "content": "- Noisy Motion Bases: To model complex residual motion, we construct " + }, + { + "bbox": [ + 55, + 411, + 295, + 521 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 411, + 295, + 521 + ], + "type": "text", + "content": " orthogonal components through SVD decomposition of randomly sampled homographies from a Gaussian distribution, effectively complementing the physical bases by capturing higher-order motion patterns. To stabilize the training procedure and simplify complex loss designs, we propose a hybrid probabilistic loss function that assumes motion models follow a Laplace distribution [42], facilitating robust and efficient optimization." + } + ] + } + ], + "index": 4 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "In summary, CamFlow effectively represents complex 2D camera motion as illustrated in Fig. 1, capturing both background and foreground elements while accurately modeling depth-varying motion patterns that conventional approaches struggle to represent. To rigorously evaluate its performance, we introduce GHOF-Cam, a novel benchmark specifically designed for camera motion estimation by systematically masking dynamic objects and ill-posed occlusion regions in established optical flow datasets [24], thereby isolating pure camera-induced motion. Through comprehensive experiments across diverse datasets under both standard and challenging conditions, we demonstrate that CamFlow consistently outperforms state-of-the-art single-plane homography and multi-planar methods in both sparse and dense camera motion estimation tasks, exhibiting superior robustness and generalization capability in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 515, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 515, + 83 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 515, + 83 + ], + "type": "text", + "content": "real-world scenarios. Our main contributions are:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 90, + 553, + 209 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 314, + 90, + 553, + 125 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 90, + 553, + 125 + ], + "spans": [ + { + "bbox": [ + 314, + 90, + 553, + 125 + ], + "type": "text", + "content": "- A new hybrid motion representation that learns to model complex non-linear 2D camera motion through physically interpretable and stochastic motion bases." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 125, + 553, + 161 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 125, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 125, + 553, + 161 + ], + "type": "text", + "content": "- A novel probabilistic loss formulation based on Laplace distribution that simplifies training and stabilizes optimization without complex loss designs." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 162, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 162, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 162, + 553, + 209 + ], + "type": "text", + "content": "- A comprehensive benchmark for evaluating camera motion learning across diverse conditions. Experimental results confirm our approach's effectiveness, robustness, and generalization ability across real-world scenarios." + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 231, + 404, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 404, + 243 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 404, + 243 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 253, + 443, + 266 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 443, + 266 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 443, + 266 + ], + "type": "text", + "content": "2.1. Homography Methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 275, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 275, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 275, + 555, + 491 + ], + "type": "text", + "content": "The traditional homography estimation follows three stages: feature detection (e.g., SIFT [33], ORB [35]), correspondence matching [5], and outlier rejection (e.g., RANSAC [9], MAGSAC++ [2]). Learning-based methods like LIFT [45], SuperPoint [7], and SOSNet [40] improve robustness. Optimization-based approaches [4, 8] iteratively refine homography estimates, while deep learning methods span supervised [3, 6, 17, 21, 23, 37] and unsupervised [15, 19, 34, 44, 47] frameworks. Unsupervised models have proven effective for real-world scenarios. Notable examples include CAHomo [47] and BaseSHom [44], which enhance feature extraction and motion constraints. HomoGAN [15] integrates GAN loss [12] and Transformers [32] for coarse-to-fine refinement. Despite this progress, homography remains a single-plane model, limiting its effectiveness in complex motion. In this paper, we introduce a hybrid motion-basis representation to model multi-plane, non-linear motion more effectively." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 511, + 437, + 523 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 511, + 437, + 523 + ], + "spans": [ + { + "bbox": [ + 314, + 511, + 437, + 523 + ], + "type": "text", + "content": "2.2. Mesh-based Methods" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 715 + ], + "type": "text", + "content": "Mesh-based warping estimates local homographies per mesh cell, including dual-plane approaches [10], patchwise mixtures of homography [13], flexible warping techniques like APAP [46], Bundled Paths [26], grid-based methods [27], and cascade residual homography [38]. Deep learning variants include deepMeshFlow [43], MeshCA-Homo [29], which merges multi-resolution meshes, and BasesMesh [30], which applies motion bases per grid. MeshHomoGAN [31] incorporates a planarity-aware mechanism for local homography estimation. However, these methods still assume that local regions fit the homography relationship, limiting their ability to represent complex motion. To address this limitation, we propose a novel motion representation that combines multiple non-linear motion bases, enhancing performance." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7625" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 555, + 314 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 555, + 314 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 555, + 314 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 555, + 314 + ], + "type": "image", + "image_path": "a2fc9670de9cef2c7ab5a3e2d6be45eb09ffee994f94f597dcec02f7ed7bf34a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "lines": [ + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": "Figure 2. Our proposed motion estimation framework. Given image pair " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "(I_a, I_b)" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": ", features are extracted through a multi-scale pyramid and processed by the motion estimation transformer (MET) to compute weights for physical (blue) and noisy (red) motion bases. These weights linearly combine predefined motion bases to generate flow maps for warping. A mask generator predicts uncertainty masks " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ab}" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ba}" + }, + { + "bbox": [ + 54, + 326, + 555, + 371 + ], + "type": "text", + "content": " to reject unreliable regions, enhancing estimation robustness." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 390, + 111, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 111, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 111, + 403 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 140, + 426 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 140, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 140, + 426 + ], + "type": "text", + "content": "3.1. Motion Basis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 437, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 474 + ], + "type": "text", + "content": "In this work, we propose a novel motion representation through motion basis learning with deep networks to better represent the camera motion via Physical and Noisy bases." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": "Physical Basis. Consider a pixel with homogeneous coordinates " + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{P}(x,y) = [x,y,1]^T\\in \\mathbb{R}^3" + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": ". The physical motion " + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "inline_equation", + "content": "m = [\\Delta x,\\Delta y,1]^T" + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": " induced by a homography " + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 478, + 296, + 515 + ], + "type": "text", + "content": " is defined:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 116, + 533, + 295, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 533, + 295, + 547 + ], + "spans": [ + { + "bbox": [ + 116, + 533, + 295, + 547 + ], + "type": "interline_equation", + "content": "m = \\mathbf {H} \\cdot \\mathbf {P} (x, y) - \\mathbf {P} (x, y), \\tag {2}", + "image_path": "d344db791b24e3a1e4cd3e615ed4645eb7da01cadc7729783078aa9e2c3ba3f3.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 566, + 295, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 295, + 590 + ], + "type": "text", + "content": "After normalization, the two-dimensional coordinates (x-axis and y-axis) of the motion become:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 608, + 295, + 638 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 608, + 295, + 638 + ], + "spans": [ + { + "bbox": [ + 115, + 608, + 295, + 638 + ], + "type": "interline_equation", + "content": "\\Delta x = \\frac {h _ {1} x + h _ {2} y + h _ {3}}{h _ {7} x + h _ {8} y + 1} - x, \\tag {3}", + "image_path": "bc0409109d23d09029e1e6123941c2b8bfb60aea1405473cf01a04474c90413c.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 117, + 635, + 235, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 635, + 235, + 660 + ], + "spans": [ + { + "bbox": [ + 117, + 635, + 235, + 660 + ], + "type": "interline_equation", + "content": "\\Delta y = \\frac {h _ {4} x + h _ {5} y + h _ {6}}{h _ {7} x + h _ {8} y + 1} - y,", + "image_path": "d9c98441915421e6b9bbb60f369678b0271bfc938e790b3202add2db07b6413d.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_1, \\ldots, h_8" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " denote the elements of matrix " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{H}" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "inline_equation", + "content": "h_9" + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": " is constrained to 1. By applying a Taylor expansion, this motion can be mapped to another subspace. For instance," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "text", + "content": "expanding " + }, + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "inline_equation", + "content": "\\Delta x" + }, + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "text", + "content": " around the point " + }, + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "inline_equation", + "content": "(x,y) = (0,0)" + }, + { + "bbox": [ + 313, + 392, + 530, + 404 + ], + "type": "text", + "content": " gives:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 339, + 409, + 555, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 409, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 339, + 409, + 555, + 435 + ], + "type": "interline_equation", + "content": "\\Delta x = \\frac {\\left(h _ {1} - 1\\right) x + h _ {2} y + h _ {3} - h _ {7} x ^ {2} - h _ {8} x y}{h _ {7} x + h _ {8} y + 1} \\tag {4}", + "image_path": "e5ffe56fcc3ada398ff5f4e6171262554d347880e7094bd9b8e7bb1abe059375.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 438, + 553, + 450 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 438, + 553, + 450 + ], + "spans": [ + { + "bbox": [ + 356, + 438, + 553, + 450 + ], + "type": "interline_equation", + "content": "\\approx w _ {1} \\cdot 1 + w _ {2} \\cdot x + w _ {3} \\cdot y + w _ {4} \\cdot x y \\tag {5}", + "image_path": "d4d68db5a717cacaabb75176425bdf57d2d94c0ab591268e624dfbe0f80680d6.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 366, + 453, + 553, + 467 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 453, + 553, + 467 + ], + "spans": [ + { + "bbox": [ + 366, + 453, + 553, + 467 + ], + "type": "interline_equation", + "content": "+ w _ {5} \\cdot x ^ {2} + w _ {6} \\cdot y ^ {2} + \\Delta , \\tag {6}", + "image_path": "793b1049166f8dc9edc6c471db435b86718fd24058767a7b13f3fc84e36c72b1.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "inline_equation", + "content": "w_{i}, i \\in [1,6]" + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": " are coefficients, the basis functions are " + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "inline_equation", + "content": "b = [1,x,y,xy,x^2,y^2]" + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": " denotes the higher-order infinitesimal. Similarly, " + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Delta y" + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": " can be decomposed into this subspace. By combining the decompositions of both " + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "inline_equation", + "content": "\\Delta x, \\Delta y" + }, + { + "bbox": [ + 313, + 473, + 555, + 533 + ], + "type": "text", + "content": ", the motion space can be represented using the 12 bases:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 539, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 539, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 347, + 539, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {F} = \\left\\{\\left(b _ {i}, 0\\right) \\mid b _ {i} \\in b \\right\\} \\cup \\left\\{\\left(0, b _ {i}\\right) \\mid b _ {i} \\in b \\right\\}, \\tag {7}", + "image_path": "bed5f666124d50938e282bcc2d57f3a3aa37cc20b0fcb87a2f811df862a2f2d7.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "content": "-th element in " + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 558, + 553, + 594 + ], + "type": "text", + "content": ". Each basis can be transformed into an optical flow according to the image coordinate, as shown in Fig. 2 (Physical motion bases)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Non-Linearity of Adding Bases. A key finding is that combining flow fields derived from different homographies produces a non-linear motion field that cannot be represented by any single homography. This challenges the linear basis assumption in prior work [44]. As shown in Fig. 3, when two homography-derived flows (flow1 and flow2) are added together, the resulting flow (flow3) differs from the flow derived from multiplying the original homography matrices (homo3). Additionally, attempting to solve for a homography from points sampled on flow3 yields inconsistent" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7626" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 295, + 160 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 295, + 160 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 295, + 160 + ], + "type": "image", + "image_path": "9a6325c9dde6565015401ca32f5c72a94dd1180fb1776522ef62d649ab436f46.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 164, + 296, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 296, + 232 + ], + "type": "text", + "content": "Figure 3. Non-linearity of flow addition. Two homography matrices generate flow1 and flow2. Adding these flows (flow3) differs from the flow derived by multiplying the original homography matrices (homo3). When sampling points from flow3 to solve for a homography, we get inconsistent solutions, proving that combined flow fields cannot be represented by a single homography." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 242, + 296, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 242, + 296, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 296, + 289 + ], + "type": "text", + "content": "solutions, confirming that the combined flow cannot be represented by any single homography. This observation motivates our approach, which leverages more bases to model complex camera motion." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "text", + "content": "Stochastic Basis. While the physical bases effectively capture fundamental motion patterns, the complete space of camera motions is infinitely dimensional when considering higher-order Taylor expansions. Exhaustively modeling all possible bases becomes computationally intractable. To address this challenge, we leverage the expressive power of random sampling to complement our physical bases. Specifically, we generate " + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "text", + "content": " random " + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "text", + "content": " matrices " + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "inline_equation", + "content": "\\left\\{\\mathbf{H}^{(k)}\\right\\}_{k=1}^{K} \\subseteq \\mathbb{R}^{3 \\times 3}" + }, + { + "bbox": [ + 55, + 290, + 296, + 403 + ], + "type": "text", + "content": ", where each matrix is formulated as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 411, + 295, + 440 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 411, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 61, + 411, + 295, + 440 + ], + "type": "interline_equation", + "content": "\\mathbf {H} = \\left\\{h _ {i} \\right\\} _ {i = 1} ^ {9}, \\text {w h e r e} \\quad h _ {i} = \\left\\{ \\begin{array}{l l} \\epsilon_ {i} \\sim \\mathcal {N} (0, 1), & 1 \\leq i \\leq 8, \\\\ 1, & i = 9. \\end{array} \\right. \\tag {8}", + "image_path": "271542eab046729922d5535b7c2ae8660479918ff6d390429c498c02ccbe4181.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "content": "Following BasesHomo [44], we convert random matrices into flows and apply singular value decomposition (SVD) to extract principal components. This process yields " + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "inline_equation", + "content": "N - 12" + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "content": " stochastic bases that capture diverse motion patterns beyond physical bases, as illustrated in Fig. 2 (Noisy motion bases). These stochastic bases, combined with the 12 physical basis vectors described earlier, form a comprehensive set of " + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 448, + 296, + 556 + ], + "type": "text", + "content": " motion bases that enhances our ability to model complex non-linear camera motions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 167, + 575 + ], + "type": "text", + "content": "3.2. Network Structure" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "The network is illustrated in Fig. 2. Following practices from previous work [15, 47], we process video frames by: (1) random cropping to " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "320 \\times 576" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " patches, (2) converting to grayscale, (3) projecting into a shallow feature space to handle luminance variations, and (4) generating a 3-layer feature pyramid for multi-scale processing. Then we propose a Motion Estimation Transformer (MET). The MET employs a specialized architecture to separately predict weights for both physical bases (12 weights) and stochastic bases (" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "N - 12" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " weights), yielding bidirectional weights " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{ab}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{w}}_{ba}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ". The final camera motion is computed through a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "linear combination of the predicted weights and their corresponding motion bases, producing bidirectional dense motion fields " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{ab}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{ba}" + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": "It is noteworthy that the confidence masks " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ab}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ba}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " are also crucial in predicting camera motion, as they effectively filter out dynamic objects. First, we apply the predicted " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{ab}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{ba}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": ", respectively, obtaining warped feature maps " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "f_{a}^{\\prime}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "f_{b}^{\\prime}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": ". The mask network " + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 108, + 554, + 193 + ], + "type": "text", + "content": " [22] inputs the concatenated features to generate weighted maps that highlight well-aligned regions:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 351, + 198, + 555, + 214 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 198, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 351, + 198, + 555, + 214 + ], + "type": "interline_equation", + "content": "\\mathbf {d} _ {a b} = \\mathcal {M} \\left(\\left[ f _ {a}, f _ {b} ^ {\\prime} \\right]\\right), \\mathbf {d} _ {b a} = \\mathcal {M} \\left(\\left[ f _ {b}, f _ {a} ^ {\\prime} \\right]\\right). \\tag {9}", + "image_path": "e25030b4274ef605cdbbce1916a2785340ec1528479078616c65c9d564a11dcb.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 220, + 404, + 232 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 220, + 404, + 232 + ], + "spans": [ + { + "bbox": [ + 314, + 220, + 404, + 232 + ], + "type": "text", + "content": "3.3. Loss Function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 237, + 554, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 554, + 297 + ], + "type": "text", + "content": "Probabilistic Motion Modeling. Camera motion estimation faces a fundamental challenge: distinguishing between camera-induced and object motion in the scene. We therefore formulate our approach as a probabilistic model that explicitly accounts for uncertainty in motion estimation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 297, + 554, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 554, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 554, + 334 + ], + "type": "text", + "content": "Building on previous findings that 2D motion follows a Laplace distribution [11, 16, 42], we model the conditional probability of the true camera motion given our prediction:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 392, + 340, + 553, + 354 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 340, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 392, + 340, + 553, + 354 + ], + "type": "interline_equation", + "content": "p \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right), \\tag {10}", + "image_path": "e275ba605c37414391e448ac1102523621b24a2b47e1e9b2d0877cde246eb342.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{m}}_{ab}" + }, + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "text", + "content": " is our predicted motion field and " + }, + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ab}" + }, + { + "bbox": [ + 313, + 361, + 554, + 420 + ], + "type": "text", + "content": " represents our confidence in each pixel's motion estimate. Higher confidence values indicate pixels likely following camera motion, while lower values suggest pixels belonging to independently moving objects." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 553, + 456 + ], + "type": "text", + "content": "Laplace Distribution Model. We model the probability density using two conditionally independent Laplace distributions for the horizontal and vertical components:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 332, + 462, + 553, + 511 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 462, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 332, + 462, + 553, + 511 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) = \\prod \\left(\\frac {1}{\\sqrt {2 \\sigma^ {2}}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | u - \\mu_ {u} |} \\right. \\tag {11} \\\\ \\left. \\cdot \\frac {1}{\\sqrt {2} \\sigma^ {2}} e ^ {- \\sqrt {\\frac {2}{\\sigma^ {2}}} | v - \\mu_ {v} |}\\right), \\\\ \\end{array}", + "image_path": "7fbf39f262ac2e39896821be1b75d3ee88ea5007eb6b9fa6df11d0e7aaf25495.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": " are the ground-truth motion components, " + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "inline_equation", + "content": "(\\mu_u,\\mu_v)" + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": " are the predicted motion components, and " + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\sigma^2" + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": " is derived from our confidence mask " + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ab}" + }, + { + "bbox": [ + 313, + 516, + 554, + 575 + ], + "type": "text", + "content": ". This formulation allows our model to express both its prediction and its uncertainty about that prediction." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "text", + "content": "Hybrid Loss Strategy. Another challenge in camera motion estimation is the scarcity of ground-truth labels. To overcome this, we introduce a hybrid loss strategy with two components: 1) Motion supervision loss " + }, + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "inline_equation", + "content": "(\\ell_{NLL_m})" + }, + { + "bbox": [ + 313, + 575, + 554, + 658 + ], + "type": "text", + "content": ": we generate pseudo-labels using existing methods and apply negative log-likelihood (NLL) loss in both forward and backward directions:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 318, + 662, + 553, + 674 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 662, + 553, + 674 + ], + "spans": [ + { + "bbox": [ + 318, + 662, + 553, + 674 + ], + "type": "interline_equation", + "content": "\\ell_ {N L L _ {m}} = - \\log p \\left(\\mathbf {m} _ {a b} \\mid \\hat {\\mathbf {m}} _ {a b}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(\\mathbf {m} _ {b a} \\mid \\hat {\\mathbf {m}} _ {b a}; \\mathbf {d} _ {b a}\\right). \\tag {12}", + "image_path": "0edd009bb161b3c53a2d624a82896a67361cecb28b8c392f62619cbefb3dc401.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "2) Photometric loss " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "(\\ell_{NLL_p})" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ": We apply the same probabilistic framework to enforce consistency between warped features. Given image features " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_{a}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "inline_equation", + "content": "f_{b}" + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": ", we use our predicted" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7627" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 124 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 124 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 124 + ], + "type": "image", + "image_path": "5a40caf1a186e281aa23eb221e410ebb71b84dc83aeea338b19982c1bdce3251.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 133, + 295, + 212 + ], + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 212 + ], + "type": "text", + "content": "Figure 4. Using the GHOF benchmark, we leverage SAM to generate semantic maps and manually identify dynamic objects such as cars and people, producing corresponding masks. These masks are then dilated to encompass occlusion regions, reducing ill-posed artifacts. Finally, the masks are applied to both the images and optical flow, isolating camera-induced motion and constructing a camera-motion-only dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "text", + "content": "motion to produce warped features " + }, + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "inline_equation", + "content": "f_{a}^{\\prime}" + }, + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "inline_equation", + "content": "f_{b}^{\\prime}" + }, + { + "bbox": [ + 55, + 222, + 241, + 236 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 242, + 295, + 255 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 242, + 295, + 255 + ], + "spans": [ + { + "bbox": [ + 69, + 242, + 295, + 255 + ], + "type": "interline_equation", + "content": "\\ell_ {N L L _ {p}} = - \\log p \\left(f _ {a} \\mid f _ {b} ^ {\\prime}; \\mathbf {d} _ {a b}\\right) - \\log p \\left(f _ {b} \\mid f _ {a} ^ {\\prime}; \\mathbf {d} _ {b a}\\right). \\tag {13}", + "image_path": "80a62228b95d7cd599390046fec2654e26b42d088d81159832f578c6ae9882f7.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 263, + 296, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 300 + ], + "type": "text", + "content": "Adaptive loss balancing: to ensure stable training despite the different scales of our loss components, we dynamically balance them using:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 307, + 295, + 333 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 307, + 295, + 333 + ], + "spans": [ + { + "bbox": [ + 94, + 307, + 295, + 333 + ], + "type": "interline_equation", + "content": "\\ell_ {\\text {o v e r a l l}} = \\ell_ {N L L _ {p}} + \\mathbf {w} \\times \\frac {\\left| \\ell_ {N L L _ {p}} \\right|}{\\left| \\ell_ {N L L _ {m}} \\right|} \\cdot \\ell_ {N L L _ {m}}, \\tag {14}", + "image_path": "90adf0823a34b446d712d598a8b484fc13eccc8ec2d7acd12057f094f8d3c496.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 339, + 187, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 187, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 187, + 352 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 339, + 187, + 352 + ], + "type": "inline_equation", + "content": "\\mathbf{w}" + }, + { + "bbox": [ + 55, + 339, + 187, + 352 + ], + "type": "text", + "content": " is a predefined weight." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 363, + 137, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 137, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 137, + 376 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 382, + 115, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 382, + 115, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 382, + 115, + 394 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 400, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 472 + ], + "type": "text", + "content": "We evaluate our method on: CAHomo [47] and GHOF [24]. We train on CAHomo (460K training pairs, 4.2K test pairs across regular, low-texture, low-light, and foreground scenes) with additional generated samples [23], and conduct zero-shot testing on GHOF (256 test pairs in Regular, Foggy, Low-light, Rainy, and Snowy conditions)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 640 + ], + "type": "text", + "content": "GHOF-Cam Benchmark. To isolate camera motion from dynamic scene elements, we propose a camera-motion-specific benchmark derived from GHOF. We employ the Segment Anything Model (SAM) [20] to generate semantic maps, from which we identify dynamic objects (e.g., cars, people) and create corresponding masks. These masks are dilated to encompass occlusion regions, mitigating ill-posed artifacts at object boundaries. For edge occlusions not detected by semantic segmentation, we mainly utilize the ground-truth homography to identify black edge regions as additional masks. The combined masks are then applied to both input images and ground-truth optical flow, resulting in a benchmark that exclusively captures camera-induced motion, as illustrated in Fig. 4." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 647, + 244, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 244, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 244, + 661 + ], + "type": "text", + "content": "4.2. Comparison with Existing Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We assess CamFlow against a comprehensive set of methods across three primary categories. The first category encompasses feature-based methods: SIFT [33], ORB [35], SuperPoint[7] with SuperGlue (SPSG) [36], and LoFTR" + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 257 + ], + "blocks": [ + { + "bbox": [ + 317, + 70, + 553, + 257 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 257 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 257 + ], + "type": "table", + "html": "
MethodsAVGRELTLLSFLF
1)\\( \\mathcal{I}_{3\\times 3} \\)6.707.757.657.217.533.39
2)SIFT[33] + RANSAC[9]1.410.301.344.030.810.57
3)SIFT[33] + MAGSAC[1]1.340.311.723.390.800.47
4)ORB[35] + RANSAC[9]1.480.852.591.671.101.24
5)ORB[35] + MAGSAC[1]1.690.973.341.581.151.40
6)SPSG[7, 36] + RANSAC[9]0.710.410.870.720.800.75
7)SPSG[7, 36] + MAGSAC[1]0.630.360.790.700.710.70
8)LoFTR[39] + RANSAC[9]1.440.562.701.361.051.52
9)LoFTR[39] + MAGSAC[1]1.390.552.571.331.051.41
10)DHN[6]2.871.514.482.762.623.00
11)LocalTrans[37]4.214.094.844.555.302.25
12)IHN[3]4.674.855.545.105.042.84
13)RealSH[17]0.340.220.350.440.420.29
14)DMHomo[23]0.310.190.330.400.380.28
15)CAHomo[47]0.880.731.011.030.920.70
16)BasesHomo[44]0.500.290.540.650.610.41
17)HomoGAN[15]0.390.220.410.570.440.31
18)Ours*0.330.200.330.410.390.30
19)Ours0.320.190.320.390.390.31
", + "image_path": "ef87aaccd922e224b10b850dd1f58749ba0598090a70080350e04cb84c07f9b7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 317, + 307, + 554, + 422 + ], + "blocks": [ + { + "bbox": [ + 313, + 260, + 553, + 305 + ], + "lines": [ + { + "bbox": [ + 313, + 260, + 553, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 553, + 305 + ], + "type": "text", + "content": "Table 1. The benchmark consists of 5 distinct scenarios, namely regular (RE), low-texture (LT), low-light (LL), small foreground (SF), and large foreground (LF). The point matching errors (PME) on the test set of CAHomo [47] are presented." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 307, + 554, + 422 + ], + "lines": [ + { + "bbox": [ + 317, + 307, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 317, + 307, + 554, + 422 + ], + "type": "table", + "html": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×35.223.656.695.884.904.96
3)SIFT[33]2.820.602.437.090.613.37
4)SPSG[7, 36]3.073.991.576.880.792.16
5)CAHomo[47]2.812.022.034.562.842.61
6)BasesHomo[44]1.741.390.974.120.661.58
7)Meshflow[27]2.151.092.215.570.441.69
8)HM_Mix[13]4.351.024.038.751.536.42
9)RANSAC-F[38]3.262.813.145.122.213.04
10)Ours1.101.080.742.150.461.05
", + "image_path": "68fdd02bcdd385a6b35b906cc4595f72ab6849d05aa9779d2722446caf135d42.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 460, + 554, + 606 + ], + "blocks": [ + { + "bbox": [ + 313, + 425, + 554, + 458 + ], + "lines": [ + { + "bbox": [ + 313, + 425, + 554, + 458 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 554, + 458 + ], + "type": "text", + "content": "Table 2. To evaluate generalizability, we compute the end point errors (EPE) of pre-trained models from Table 1 on our proposed GHOF-Cam benchmark." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 460, + 554, + 606 + ], + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 606 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 606 + ], + "type": "table", + "html": "
1)MethodsAVGREFOGLLRAINSNOW
2)I3×36.334.947.248.095.485.89
3)SIFT[33]4.800.594.4712.100.626.20
4)SPSG[7, 36]4.473.542.2110.660.835.10
5)DHN[6]6.616.046.027.686.996.32
6)LocalTrans[37]5.724.066.495.955.786.34
7)IHN[3]8.177.108.719.346.579.13
8)RealSH[17]1.721.600.884.420.431.28
9)DMHomo[23]1.750.640.854.160.392.74
10)CAHomo[47]3.874.103.846.991.273.17
11)BasesHomo[44]2.282.021.434.900.782.29
12)HomoGAN[15]1.951.730.603.950.473.02
13)Ours1.231.150.962.690.400.93
", + "image_path": "c1fdb222a4a6cc9da9061bd1e231eacfaf5e24e3a70a44b0bf85048f18c39bf4.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 609, + 553, + 641 + ], + "lines": [ + { + "bbox": [ + 313, + 609, + 553, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 609, + 553, + 641 + ], + "type": "text", + "content": "Table 3. To evaluate generalizability, we compute the point matching errors (PME) on the GHOF [24] test set using pre-trained models from Table 1." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "[39], each evaluated with two outlier rejection techniques: RANSAC [9] and MAGSAC [1]. The second category includes supervised learning approaches: DHN [6], Local-Trans [37], IHN [3], RealSH [17], and DMHomo [23]. The third category comprises unsupervised methods: CAHomo" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7628" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 178 + ], + "type": "table", + "html": "
MethodAVGREFOGDARKRAINSNOW
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
\\( \\mathcal{I}_{3\\times 3} \\)24.050.74030.083621.060.69000.075026.570.77110.082125.700.85060.078521.530.53350.141125.370.85620.0412
GT-Homo32.780.91870.057028.390.86970.054935.230.95080.049231.880.94050.057530.110.85110.103338.310.98140.0199
SIFT28.440.90740.078129.230.91480.054529.420.90160.076827.370.90740.098230.000.86320.105526.160.94970.0558
SPSG28.010.86970.079621.830.75930.088630.880.90490.064527.600.90190.096628.860.82700.110330.880.95560.0379
CAHomo25.290.78370.084122.670.73410.080527.510.80480.075126.120.87430.084622.950.61300.142027.200.89240.0384
BasesHomo29.610.90260.067225.080.85220.066631.060.91700.062730.050.93030.070229.580.85120.107132.300.96220.0292
MeshFlow29.910.92390.068828.570.92160.057628.680.92800.074229.410.92540.077430.680.87470.104932.230.97000.0298
HM_Mix25.770.88960.088226.090.87210.059626.560.87530.088226.430.90370.100228.200.86720.110721.580.92960.0820
RANSAC-F26.040.83480.089026.090.88120.066529.220.89440.080127.290.90310.092321.680.55850.149525.900.93710.0566
Ours32.090.91420.057527.080.86150.055834.170.93710.051232.360.94210.056530.520.86080.102136.350.96920.0218
", + "image_path": "65eed8a36685bfcf2d52976052ab9c32a0d657918969a5c468e9b08e04c05b86.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 55, + 178, + 553, + 210 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 553, + 210 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 553, + 210 + ], + "type": "text", + "content": "Table 4. Quantitative comparison of different methods on various environmental conditions. We present three key perception metrics: PSNR (higher is better), SSIM (higher is better), and LPIPS (lower is better). Our method consistently outperforms existing approaches across all scenarios and metrics." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 237, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 237, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 237, + 228 + ], + "type": "text", + "content": "[47], BasesHomo [44], and HomoGAN [15]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 231, + 294, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 231, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 54, + 231, + 294, + 338 + ], + "type": "text", + "content": "For multi-plane camera motion modeling, we compare against both traditional approaches (MeshFlow [27], Homography Mixture [13], RANSAC-Flow [38]) and unsupervised deep methods (BasesMesh [30] and MeshHomogAN [31]). Regarding pre-training, DHN, LocalTrans and IHN use the MS-COCO dataset [25], while other deep learning methods are pre-trained on CAHomo. Additional qualitative results and visual comparisons are available on: https://lhaippp.github.io/CamFlow/." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 348, + 194, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 194, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 194, + 360 + ], + "type": "text", + "content": "4.2.1. Quantitative Comparison." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 364, + 295, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 364, + 295, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 364, + 295, + 483 + ], + "type": "text", + "content": "Sparse camera motion. We evaluate using the points matching error (PME), which measures the average geometric distance between transformed source points and their corresponding ground-truth target points. Table 1 presents the quantitative performance of our method alongside various homography estimation approaches on the CAHomo test set, categorized as: feature-based methods (rows 2-9), supervised approaches (rows 10-14), and unsupervised ones (rows 15-17). The " + }, + { + "bbox": [ + 54, + 364, + 295, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{3\\times 3}" + }, + { + "bbox": [ + 54, + 364, + 295, + 483 + ], + "type": "text", + "content": " baseline (row 1) represents the distance between point pairs without transformation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "content": "Our CamFlow method achieves superior results across multiple categories, outperforming the leading unsupervised method, HomoGAN, by reducing PME by " + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "inline_equation", + "content": "17.95\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "content": " (from 0.39 to 0.32). While feature-based methods like SIFT+RANSAC excel in regular (RE) scenes with abundant texture and keypoints, CamFlow surpasses them with a " + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "inline_equation", + "content": "36.67\\%" + }, + { + "bbox": [ + 54, + 485, + 295, + 628 + ], + "type": "text", + "content": " improvement (reducing error by 0.11). To balance benchmark performance with generalization capability, we include Ours\\* (row 18), which represents an early-stopped training model that delivers better generalizability across datasets, even though it slightly underperforms our fully-trained model (row 19) on the CAHomo benchmark." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "In small foreground (SF) and large foreground (LF) scenarios, where dynamic objects disrupt camera motion estimation, our probabilistic loss and confidence masking yield lower PMEs compared to methods like CAHomo and HomoGAN, which employ explicit outlier rejection masks for robustness. In low-texture (LT) and low-light (LL) conditions, learning-based approaches generally exhibit superior" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 288 + ], + "type": "text", + "content": "resilience due to their keypoint-free strategies, particularly among unsupervised techniques. However, these scenes often contain homogeneous regions [41], occupying large image areas and reducing photometric differences, which limit unsupervised methods' effectiveness. By contrast, CamFlow excels in LT and LL." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 291, + 553, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 422 + ], + "type": "text", + "content": "Generalization experiment. Evaluating cross-dataset generalization presents a significant challenge in motion estimation, particularly for applications where real-world video often differs substantially from training data. We assess CamFlow's generalization capabilities in two ways: (1) for dense camera motion, we evaluate against traditional single-plane, multi-planar, and deep learning approaches using our proposed GHOF-Cam benchmark (Table 2 and Table 4); and (2) for sparse camera motion, we compare against traditional and homography-based methods on the original GHOF benchmark [24] (Table 3)." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 424, + 553, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 424, + 553, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 424, + 553, + 688 + ], + "type": "text", + "content": "Table 2 presents End Point Error (EPE) results on the GHOF-Cam benchmark, which represents ground-truth camera motion. We evaluate against homography-based methods (rows 3-5) and non-single-plane approaches (rows 6-9). As illustrated in Fig. 3, we classify BasesHomo [44] as a multi-plane rather than single-homography method. The results demonstrate that CamFlow outperforms competing methods across nearly all categories, showcasing exceptional zero-shot capability in capturing non-linear camera motion patterns. Additionally, Table 4 presents perceptual quality metrics (PSNR, SSIM, and LPIPS) across various conditions. We also include ground-truth homography results for reference (GT-Homo). Our method, CamFlow, achieves state-of-the-art performance across multiple categories, consistently outperforming all baselines, particularly in challenging conditions such as snowy scenes. Notably, CamFlow approaches the performance of ground-truth homography, with only a " + }, + { + "bbox": [ + 313, + 424, + 553, + 688 + ], + "type": "inline_equation", + "content": "0.69\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 424, + 553, + 688 + ], + "type": "text", + "content": " difference in PSNR, 0.0045 in SSIM and 0.0005 in LPIPS, while significantly surpassing the second-best method, MeshFlow. These results support CamFlow's potential for applications requiring high-quality visual alignment." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Table 3 presents compelling evidence of CamFlow's generalization capability on the GHOF benchmark. Our" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7629" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 555, + 173 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 555, + 173 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 555, + 173 + ], + "type": "image", + "image_path": "3f2f9b429c92677a72b90585fddde43ca0b71d10471518a0aa11d8762c006578.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 176, + 555, + 209 + ], + "lines": [ + { + "bbox": [ + 55, + 176, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 55, + 176, + 555, + 209 + ], + "type": "text", + "content": "Figure 5. Qualitative results of CamFlow and methods from each category (i.e., supervised, unsupervised, and multi-homography) on the CAHomo testset [47]. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 56, + 211, + 555, + 445 + ], + "blocks": [ + { + "bbox": [ + 56, + 211, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 56, + 211, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 555, + 445 + ], + "type": "image", + "image_path": "d4e7535b869100088323413231d5983607b9ec659392b10d32cb1745488ce234.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 448, + 555, + 492 + ], + "lines": [ + { + "bbox": [ + 55, + 448, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 448, + 555, + 492 + ], + "type": "text", + "content": "Figure 6. Qualitative results of our method and the best-performing generalizable methods from learning-based categories (i.e., supervised, unsupervised, and multi-homography) on the GHOF testset [24]. The examples are arranged from top to bottom and include RE, Fog, and Dark. The images are generated by superimposing the warped source images on the target image. Error-prone regions are highlighted with red boxes, which are further converted into alignment heatmaps for better distinction when zoomed in." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "content": "method achieves the lowest average PME (1.23), representing a " + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "inline_equation", + "content": "28.5\\%" + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "content": " improvement over the previous best supervised method (RealSH at 1.72) and a " + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "inline_equation", + "content": "36.9\\%" + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "content": " improvement over the leading unsupervised approach (HomoGAN at 1.95). This performance advantage extends across all environmental conditions, with particularly significant gains in challenging scenarios. In low-light (LL) conditions, CamFlow reduces error by " + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "inline_equation", + "content": "39.1\\%" + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "content": " compared to RealSH (from 4.42 to 2.69), while in snowy (SNOW) environments, it achieves a remarkable " + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "inline_equation", + "content": "69.2\\%" + }, + { + "bbox": [ + 55, + 503, + 297, + 660 + ], + "type": "text", + "content": " error reduction compared to HomoGAN (from 3.02 to 0.93). These results confirm that CamFlow enables robust zero-shot transfer to unseen datasets, even under diverse and challenging environmental conditions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 670, + 191, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 191, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 191, + 684 + ], + "type": "text", + "content": "4.3. Qualitative Comparison" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Fig. 5 and 6 present qualitative results of CamFlow alongside competing methods on the CAHomo and GHOF test" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 503, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 503, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 503, + 555, + 588 + ], + "type": "text", + "content": "sets. Dynamic visualizations are available on our project page. For visualization, we employ red-blue ghosting and alignment heat maps following [18]. After transforming the source frame using estimated camera flow, misaligned regions appear as red and blue ghosting. We highlight specific areas with red boxes and generate alignment heat maps where brighter regions indicate poorer alignment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "In both figures, the \"Identity\" column shows source and target images overlaid without warping. For CAHomo (Fig. 5), we focus on evaluating background camera motion modeling capabilities. We selected challenging test cases featuring extremely low-light conditions (first row) and sophisticated motion patterns (parallax and depth variation), comparing against leading methods from three categories: supervised (RealSH), unsupervised (HomoGAN), and deep meshflow (BasesMesh, MeshHomoGAN). CamFlow achieves superior alignment in background regions." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7630" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 175, + 116 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 175, + 116 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 175, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 175, + 116 + ], + "type": "image", + "image_path": "d94fd4b5de40cdfec6bd1ad59315be78bb82041cde3c99570c013173ef00e74b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 123, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 123, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 123, + 295, + 167 + ], + "type": "text", + "content": "Figure 7. Uncertainty masks and corresponding original images. The brighter the mask, the higher the uncertainty. The masks effectively highlight dynamic objects, indicating regions where camera motion estimation is less reliable." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 178, + 70, + 295, + 116 + ], + "blocks": [ + { + "bbox": [ + 178, + 70, + 295, + 116 + ], + "lines": [ + { + "bbox": [ + 178, + 70, + 295, + 116 + ], + "spans": [ + { + "bbox": [ + 178, + 70, + 295, + 116 + ], + "type": "image", + "image_path": "28bfc84b23cf4b9b223c571ca1794ae20203f78a5ecfd0dc2d2d8e8cc5c91d1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 175, + 295, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 175, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 54, + 175, + 295, + 330 + ], + "type": "text", + "content": "The GHOF benchmark (Fig. 6) highlights CamFlow's advantages in handling complex camera motion and its zero-shot generalization capabilities. The first row shows a scene with parallax and foreground motion that challenges homography-based methods and reduces meshflow accuracy. The second row demonstrates CamFlow's generalization to foggy conditions unseen in training data, where depth variation complicates motion estimation. The last row presents an extreme case with motion blur, large parallax, and dynamic objects. In all scenarios, CamFlow delivers robust results where competing methods struggle, confirming its effectiveness in modeling complex camera motion patterns across diverse environmental conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 415 + ], + "type": "text", + "content": "Foreground Masks. Fig. 7 illustrates the uncertainty masks. We observe that the network effectively identifies dynamic objects and assigns higher uncertainty to regions containing these objects. This design helps focus the learning process on areas where motion is most relevant, while reducing noise from dynamic object regions, thereby enhancing the accuracy of camera motion estimation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 423, + 156, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 156, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 156, + 434 + ], + "type": "text", + "content": "4.4. Ablation Studies" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 57, + 459, + 296, + 518 + ], + "blocks": [ + { + "bbox": [ + 55, + 441, + 140, + 451 + ], + "lines": [ + { + "bbox": [ + 55, + 441, + 140, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 140, + 451 + ], + "type": "text", + "content": "4.4.1. Motion Basis" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 459, + 296, + 518 + ], + "lines": [ + { + "bbox": [ + 57, + 459, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 57, + 459, + 296, + 518 + ], + "type": "table", + "html": "
CAHomoGHOFGHOF-CamParamsInference Time
8 Bases0.371.681.452.658M76.42ms
12 Bases0.361.541.232.658M75.38ms
24 Bases0.331.231.102.660M79.63ms
200 Bases0.331.271.072.677M99.28ms
", + "image_path": "21ad81e3b087f4f6dd0f10b1929ef8deb1fb5aca6a30823dfdc31435dc81148f.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 521, + 295, + 577 + ], + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 577 + ], + "type": "text", + "content": "Table 5. Performance comparison under different numbers of motion bases on three benchmarks. Results demonstrate the effectiveness of combining physical and stochastic motion bases, with 24 bases providing optimal balance between accuracy and computational efficiency." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "In Table 5, we evaluate with varying numbers of motion bases across three benchmarks: CAHomo (trained) and GHOF/GHOF-Cam (zero-shot). The average results demonstrate that: 1) Increasing physical bases from 8 to 12 improves performance across all benchmarks; 2) Introducing additional hybrid bases (24 total) yields further enhancements, particularly for generalization (GHOF and GHOF-Cam). Notably, while expanding to 200 bases provides marginal improvements on some benchmarks, it increases inference time by " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "24.7\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". We therefore adopt 24 bases in our final model as the optimal balance between ac" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 463, + 84 + ], + "type": "text", + "content": "curacy and computational efficiency." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 340, + 107, + 531, + 162 + ], + "blocks": [ + { + "bbox": [ + 313, + 89, + 452, + 101 + ], + "lines": [ + { + "bbox": [ + 313, + 89, + 452, + 101 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 452, + 101 + ], + "type": "text", + "content": "4.4.2. Hybrid Probabilistic Loss" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 107, + 531, + 162 + ], + "lines": [ + { + "bbox": [ + 340, + 107, + 531, + 162 + ], + "spans": [ + { + "bbox": [ + 340, + 107, + 531, + 162 + ], + "type": "table", + "html": "
\\( {\\ell }_{NLLm} \\)\\( {\\ell }_{NLLp} \\)CAHomoGHOFGHOF-Cam
✓0.412.212.13
✓0.361.581.42
✓✓0.331.231.10
", + "image_path": "1a26a36ba212c272c78ac5f9ea468808016e8d70b3b9084f7667be73f5e1eabe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 165, + 553, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 165, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 165, + 553, + 209 + ], + "type": "text", + "content": "Table 6. Ablation study of different loss function combinations across three benchmarks. Results demonstrate that the hybrid approach combining motion loss and photometric loss achieves superior performance compared to individual loss components." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 212, + 555, + 368 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 555, + 368 + ], + "type": "text", + "content": "Table 6 evaluates our probabilistic loss components across three benchmarks. Using only motion loss yields limited generalization performance, because pseudo motion labels provide approximate supervision. The photometric loss alone performs substantially better, particularly on zero-shot datasets, consistent with findings from prior unsupervised methods. However, our hybrid approach combining both losses achieves the best results across all benchmarks, with improvements on GHOF-Cam (22.5% error reduction compared to photometric-only). We believe motion labels provide coarse guidance while photometric loss enables fine-grained refinement, resulting in more accurate and generalizable camera motion estimation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 377, + 388, + 389 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 388, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 388, + 389 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 397, + 555, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 555, + 613 + ], + "type": "text", + "content": "CamFlow presents a novel motion representation for modeling 2D camera motion using a hybrid motion basis approach. We identify a fundamental issue: superposing homographies by simply adding flow fields introduces nonlinear interactions, contradicting the assumption that homographies can be expressed as linear combinations of 8 basis flows. By expanding the previous 8-dimensional motion basis into a higher-dimensional space with both physical and stochastic motion bases, CamFlow effectively captures complex, non-linear motion patterns. The proposed probabilistic loss function enhances training stability. Our newly introduced GHOF-Cam benchmark demonstrates that CamFlow surpasses state-of-the-art homography and meshflow methods, showcasing superior robustness and generalization. We hope this work opens new avenues for camera motion modeling in video processing applications. Codebase, pre-trained models, and benchmark data are released at https://lhaippp.github.io/CamFlow/." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "Acknowledgements. This work was supported in part by the National Natural Science Foundation of China (NSFC) under grants No. 62372091 and No. 62031009, and in part by the Hainan Province Key R&D Program under grant No. ZDYF2024(LALH)001. We would also like to thank Hao Xu, Mingbo Hong, Hai Jiang, Xinglong Luo, Nianjin Ye, and the anonymous reviewers for their valuable discussions, constructive suggestions, and helpful feedback." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7631" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 123 + ], + "type": "text", + "content": "[1] Daniel Barath, Jiri Matas, and Jana Noskova. Magsac: marginalizing sample consensus. In Proc. CVPR, pages 10197-10205, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 158 + ], + "type": "text", + "content": "[2] Daniel Barath, Jana Noskova, Maksym Ivashechkin, and Jiri Matas. Magsac++, a fast, reliable and accurate robust estimator. In Proc. CVPR, pages 1304-1312, 2020. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 294, + 191 + ], + "type": "text", + "content": "[3] Si-Yuan Cao, Jianxin Hu, Zehua Sheng, and Hui-Liang Shen. Iterative deep homography estimation. In Proc. CVPR, pages 1879-1888, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 294, + 225 + ], + "type": "text", + "content": "[4] Che-Han Chang, Chun-Nan Chou, and Edward Y Chang. Clkn: Cascaded lucas-kanade networks for image alignment. In Proc. CVPR, pages 2213–2221, 2017. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 294, + 259 + ], + "type": "text", + "content": "[5] Padraig Cunningham and Sarah Jane Delany. K-nearest neighbour classifiers-a tutorial. ACM Computing Surveys, 54(6):1-25, 2021. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 294, + 294 + ], + "type": "text", + "content": "[6] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Deep image homography estimation. arXiv preprint arXiv:1606.03798, 2016. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 296, + 294, + 338 + ], + "type": "text", + "content": "[7] Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superpoint: Self-supervised interest point detection and description. In Proc. CVPRW, pages 224-236, 2018. 2, 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 294, + 384 + ], + "type": "text", + "content": "[8] Tianjiao Ding, Yunchen Yang, Zhihui Zhu, Daniel P Robinson, René Vidal, Laurent Kneip, and Manolis C Tsakiris. Robust homography estimation via dual principal component pursuit. In Proc. CVPR, pages 6080-6089, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 430 + ], + "type": "text", + "content": "[9] Martin A Fischler and Robert C Bolles. Random sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography. Communications of the ACM, 24(6):381-395, 1981. 2, 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 431, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 431, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 57, + 431, + 294, + 463 + ], + "type": "text", + "content": "[10] Junhong Gao, Seon Joo Kim, and Michael S Brown. Constructing image panoramas using dual-homography warping. In Proc. CVPR, pages 49-56, 2011. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 465, + 294, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 465, + 294, + 486 + ], + "spans": [ + { + "bbox": [ + 57, + 465, + 294, + 486 + ], + "type": "text", + "content": "[11] Jochen Gast and Stefan Roth. Lightweight probabilistic deep networks. In Proc. CVPR, pages 3369-3378, 2018. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 488, + 294, + 530 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 488, + 294, + 530 + ], + "spans": [ + { + "bbox": [ + 57, + 488, + 294, + 530 + ], + "type": "text", + "content": "[12] Ian J Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Proc. NIPS, 27, 2014. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 294, + 566 + ], + "type": "text", + "content": "[13] Matthias Grundmann, Vivek Kwatra, Daniel Castro, and Irfan Essa. Calibration-free rolling shutter removal. In Proc. ICCP, pages 1-8, 2012. 2, 5, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 599 + ], + "type": "text", + "content": "[14] Richard Hartley and Andrew Zisserman. Multiple View Geometry in Computer Vision. Cambridge University Press, 2003. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 294, + 644 + ], + "type": "text", + "content": "[15] Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, Qijun Zhao, and Shuaicheng Liu. Unsupervised homography estimation with coplanarity-aware gan. In Proc. CVPR, pages 17663-17672, 2022. 2, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 689 + ], + "type": "text", + "content": "[16] Eddy Ilg, Ozgun Cicek, Silvio Galesso, Aaron Klein, Osama Makansi, Frank Hutter, and Thomas Brox. Uncertainty estimates and multi-hypotheses networks for optical flow. In Proc. ECCV, pages 652-667, 2018. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[17] Hai Jiang, Haipeng Li, Songchen Han, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Supervised homography learning" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "with realistic dataset generation. In Proc. ICCV, pages 9806-9815, 2023. 2, 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[18] Hai Jiang, Haipeng Li, Yuhang Lu, Songchen Han, and Shuaicheng Liu. Semi-supervised deep large-baseline homography estimation with progressive equivalence constraint. In Proc. AAAI, pages 1024-1032, 2023. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 141, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 194 + ], + "type": "text", + "content": "[19] Dewi Endah Kharismawati, Hadi Ali Akbarpour, Rumana Aktar, Filiz Bunyak, Kannappan Palaniappan, and Toni Kazic. Cornet: Unsupervised deep homography estimation for agricultural aerial imagery. In Proc. ECCV, pages 400-417, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 553, + 240 + ], + "type": "text", + "content": "[20] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proc. ICCV, pages 4015-4026, 2023. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 243, + 553, + 274 + ], + "type": "text", + "content": "[21] Hoang Le, Feng Liu, Shu Zhang, and Aseem Agarwala. Deep homography estimation for dynamic scenes. In Proc. CVPR, pages 7652-7661, 2020. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 277, + 553, + 308 + ], + "type": "text", + "content": "[22] Haipeng Li, Kunming Luo, and Shuaicheng Liu. Gyroflow: Gyroscope-guided unsupervised optical flow learning. In Proc. ICCV, pages 12869-12878, 2021. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 553, + 354 + ], + "type": "text", + "content": "[23] Haipeng Li, Hai Jiang, Ao Luo, Ping Tan, Haoqiang Fan, Bing Zeng, and Shuaicheng Liu. Dmhomo: Learning homography with diffusion models. ACM Transactions on Graphics (TOG), 43(3):1-16, 2024. 2, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 399 + ], + "type": "text", + "content": "[24] Haipeng Li, Kunming Luo, Bing Zeng, and Shuaicheng Liu. Gyroflow+: Gyroscope-guided unsupervised deep homography and optical flow learning. International Journal of Computer Vision (IJCV), 132(6):2331-2349, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 400, + 553, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 400, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 317, + 400, + 553, + 443 + ], + "type": "text", + "content": "[25] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. ECCV, pages 740-755, 2014. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 445, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 445, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 317, + 445, + 553, + 477 + ], + "type": "text", + "content": "[26] Shuaicheng Liu, Lu Yuan, Ping Tan, and Jian Sun. Bundled camera paths for video stabilization. ACM Transactions on Graphics (TOG), 32(4):1-10, 2013. 1, 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 479, + 553, + 511 + ], + "type": "text", + "content": "[27] Shuaicheng Liu, Ping Tan, Lu Yuan, Jian Sun, and Bing Zeng. Meshflow: Minimum latency online video stabilization. In Proc. ECCV, pages 800-815, 2016. 2, 5, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 514, + 553, + 567 + ], + "type": "text", + "content": "[28] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), pages 1-14, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 569, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 632 + ], + "type": "text", + "content": "[29] Shuaicheng Liu, Nianjin Ye, Chuan Wang, Jirong Zhang, Lanpeng Jia, Kunming Luo, Jue Wang, and Jian Sun. Content-aware unsupervised deep homography estimation and its extensions. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(3):2849-2863, 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[30] Shuaicheng Liu, Yuhang Lu, Hai Jiang, Nianjin Ye, Chuan Wang, and Bing Zeng. Unsupervised global and local homography estimation with motion basis learning. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 45(6):7885-7899, 2023. 2, 6" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[31] Shuaicheng Liu, Mingbo Hong, Yuhang Lu, Nianjin Ye, Chunyu Lin, and Bing Zeng. Unsupervised global and local" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7632" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 417 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "homography estimation with coplanarity-aware gan. IEEE Transactions on Pattern Analysis and Machine Intelligence (PAMI), 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 295, + 150 + ], + "type": "text", + "content": "[32] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In Proc. ICCV, pages 10012-10022, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[33] David G Lowe. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision (IJCV), 60(2):91-110, 2004. 1, 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 237 + ], + "type": "text", + "content": "[34] Ty Nguyen, Steven W Chen, Shreyas S Shivakumar, Camillo Jose Taylor, and Vijay Kumar. Unsupervised deep homography: A fast and robust homography estimation model. IEEE Robotics and Automation Letters (RA-L), 3(3): 2346-2353, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 238, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 294, + 272 + ], + "type": "text", + "content": "[35] Ethan Rublee, Vincent Rabaud, Kurt Konolige, and Gary Bradski. Orb: An efficient alternative to sift or surf. In Proc. ICCV, pages 2564-2571, 2011. 2, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[36] Paul-Edouard Sarlin, Daniel DeTone, Tomasz Malisiewicz, and Andrew Rabinovich. Superglue: Learning feature matching with graph neural networks. In Proc. CVPR, pages 4938–4947, 2020. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 361 + ], + "type": "text", + "content": "[37] Ruizhi Shao, Gaochang Wu, Yuemei Zhou, Ying Fu, Lu Fang, and Yebin Liu. Localtrans: A multiscale local transformer network for cross-resolution homography estimation. In Proc. ICCV, pages 14890-14899, 2021. 2, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 361, + 294, + 394 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 294, + 394 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 294, + 394 + ], + "type": "text", + "content": "[38] Xi Shen, François Darmon, Alexei A Efros, and Mathieu Aubry. Ransac-flow: generic two-stage image alignment. In Proc. ECCV, pages 618-637, 2020. 2, 5, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 394, + 294, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 394, + 294, + 417 + ], + "spans": [ + { + "bbox": [ + 56, + 394, + 294, + 417 + ], + "type": "text", + "content": "[39] Jiaming Sun, Zehong Shen, Yuang Wang, Hujun Bao, and Xiaowei Zhou. Loftr: Detector-free local feature matching" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 417 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 72, + 553, + 94 + ], + "type": "text", + "content": "with transformers. In Proc. CVPR, pages 8922-8931, 2021. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[40] Yurun Tian, Xin Yu, Bin Fan, Fuchao Wu, Huub Heijnen, and Vassileios Balntas. Sosnet: Second order similarity regularization for local descriptor learning. In Proc. CVPR, pages 11016-11025, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 142, + 553, + 176 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 142, + 553, + 176 + ], + "spans": [ + { + "bbox": [ + 315, + 142, + 553, + 176 + ], + "type": "text", + "content": "[41] Prune Truong, Martin Danelljan, and Radu Timofte. Glunet: Global-local universal network for dense flow and correspondences. In Proc. CVPR, pages 6258-6268, 2020. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 315, + 177, + 553, + 211 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 177, + 553, + 211 + ], + "spans": [ + { + "bbox": [ + 315, + 177, + 553, + 211 + ], + "type": "text", + "content": "[42] Prune Truong, Martin Danelljan, Luc Van Gool, and Radu Timofte. Learning accurate dense correspondences and when to trust them. In Proc. CVPR, pages 5714-5724, 2021. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 315, + 212, + 553, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 212, + 553, + 256 + ], + "spans": [ + { + "bbox": [ + 315, + 212, + 553, + 256 + ], + "type": "text", + "content": "[43] Nianjin Ye, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Jue Wang, and Yongqing Cui. Deepmeshflow: Content adaptive mesh deformation for robust image registration. arXiv preprint arXiv:1912.05131, 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 315, + 257, + 553, + 302 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 257, + 553, + 302 + ], + "spans": [ + { + "bbox": [ + 315, + 257, + 553, + 302 + ], + "type": "text", + "content": "[44] Nianjin Ye, Chuan Wang, Haoqiang Fan, and Shuaicheng Liu. Motion basis learning for unsupervised deep homography estimation with subspace projection. In Proc. ICCV, pages 13117-13125, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 303, + 553, + 336 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 303, + 553, + 336 + ], + "spans": [ + { + "bbox": [ + 315, + 303, + 553, + 336 + ], + "type": "text", + "content": "[45] Kwang Moo Yi, Eduard Trulls, Vincent Lepetit, and Pascal Fua. Lift: Learned invariant feature transform. In Proc. ECCV, pages 467-483, 2016. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 338, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 338, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 315, + 338, + 553, + 371 + ], + "type": "text", + "content": "[46] Julio Zaragoza, Tat-Jun Chin, Michael S Brown, and David Suter. As-projective-as-possible image stitching with moving dlt. In Proc. CVPR, pages 2339–2346, 2013. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 372, + 553, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 372, + 553, + 417 + ], + "spans": [ + { + "bbox": [ + 315, + 372, + 553, + 417 + ], + "type": "text", + "content": "[47] Jirong Zhang, Chuan Wang, Shuaicheng Liu, Lanpeng Jia, Nianjin Ye, Jue Wang, Ji Zhou, and Jian Sun. Content-aware unsupervised deep homography estimation. In Proc. ECCV, pages 653-669, 2020. 2, 4, 5, 6, 7" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7633" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_content_list.json b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..938c1d9744d556c8b2edf26fc616d7dddac9e4c9 --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_content_list.json @@ -0,0 +1,1627 @@ +[ + { + "type": "text", + "text": "EvRT-DETR: Latent Space Adaptation of Image Detectors for Event-based Vision", + "text_level": 1, + "bbox": [ + 202, + 130, + 795, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dmitrii Torbunov1, Yihui Ren1, Animesh Ghose1, Odera Dim1, Yonggang Cui1 \n1Brookhaven National Laboratory, Upton, NY, USA", + "bbox": [ + 184, + 202, + 810, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "dtorbunov@bnl.gov, yren@bnl.gov, aghose@bnl.gov, dodera@bnl.gov, ycui@bnl.gov", + "bbox": [ + 153, + 255, + 836, + 270 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 305, + 326, + 320 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-based cameras (EBCs) have emerged as a bioinspired alternative to traditional cameras, offering advantages in power efficiency, temporal resolution, and high dynamic range. However, development of image analysis methods for EBCs is challenging due to the sparse and asynchronous nature of the data. This work addresses the problem of object detection for EBC cameras. The current approaches to EBC object detection focus on constructing complex data representations and rely on specialized architectures. We introduce I2EvDet (Image-to-Event Detection), a novel adaptation framework that bridges mainstream object detection with temporal event data processing. First, we demonstrate that a Real-Time DEetection Transformer, or RT-DETR, a state-of-the-art natural image detector, trained on a simple image-like representation of the EBC data achieves performance comparable to specialized EBC methods. Next, as part of our framework, we develop an efficient adaptation technique that transforms image-based detectors into event-based detection models by modifying their frozen latent representation space via minimal architectural additions. The resulting EvRT-DETR model reaches state-of-the-art performance on the standard benchmark datasets Gen1 (mAP +2.3) and 1Mpx/Gen4 (mAP +1.4). These results demonstrate a fundamentally new approach to EBC object detection through principled adaptation of mainstream architectures, offering an efficient alternative with potential applications to other temporal visual domains. The code is available at: https://github.com/realtime-intelligence/evrt-detr.", + "bbox": [ + 89, + 339, + 485, + 792 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 828, + 220, + 843 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Event-based cameras (EBCs) present a biologically inspired alternative to traditional frame-based cameras. Unlike conventional cameras that capture frames at predetermined in", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1f4beb4a8a1e624ea438d69e0ef38ecb723a8f5f86ba498b6069617facc1e5df.jpg", + "image_caption": [ + "Figure 1. Object Detection Performance vs. Inference Time. A summary of object detection performance (COCO mAP) versus inference time (ms) of various models on the 1Mpx automotive dataset. Circle size is proportional to the number of model parameters. Inference times are reported for NVIDIA T4 GPU. Our EvRT-DETR models achieve state-of-the-art accuracy while maintaining competitive inference speeds." + ], + "image_footnote": [], + "bbox": [ + 519, + 306, + 901, + 526 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tervals, EBC pixels operate asynchronously, generating data only when brightness changes exceed a threshold. This results in sparse, asynchronous data streams where each event is a tuple containing pixel location, timestamp, and polarity (whether brightness increased or decreased), as illustrated in Figure 2. This approach offers remarkable advantages: low power consumption (as low as $10\\mathrm{mW}$ ), reduced data transfer rates, exceptional temporal resolution (on the order of $\\mu s$ ), and high dynamic range ( $>100$ dB). These properties have led to widespread adoption in autonomous driving, robotics, and wearable electronics [6].", + "bbox": [ + 511, + 670, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "These unique characteristics of EBC data - sparsity, asynchronicity, and temporal nature - create fundamental challenges when applying traditional computer vision techniques. Early attempts to directly apply conventional", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "9812", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cdf3dd14f6e384cce330ebd2cff576c77f5ead735509b710f19ae1d2ab2b7679.jpg", + "image_caption": [ + "Figure 2. Event-based Camera Output Visualization. Sample frames from the 1Mpx dataset showing event polarity. Red indicates positive events, and blue denotes negative events." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 482, + 180 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "image-based object detectors to simple event representations yielded poor performance [22], highlighting the gap between these domains. This led to two predominant research directions: (1) constructing sophisticated event representations that transform EBC data into formats compatible with existing detectors (e.g., [33]) or (2) designing specialized architectures tailored to handle the temporal nature of event data (e.g., [20]). A common assumption in these approaches is that EBC data processing requires either specialized representations or architectures due to the fundamental differences between event- and frame-based vision, leading to development paths that often diverge from mainstream computer vision research.", + "bbox": [ + 88, + 262, + 482, + 458 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Meanwhile, mainstream object detection has seen remarkable progress, evolving from convolutional neural network (CNN)-based architectures, e.g.,YOLO [24] and Faster R-CNN [25], to transformer-based models like DETR (DEtection Transformer) [2]. Recent advances have culminated in highly efficient and accurate models, such as the Real-Time DETection TRansformer (RT-DETR) [30] that achieves exceptional performance while maintaining real-time inference speeds. Though primarily optimized for processing individual images, these state-of-the-art detectors offer powerful feature extraction capabilities that potentially could be leveraged for temporal data.", + "bbox": [ + 89, + 460, + 482, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This evolution in mainstream detection creates an opportunity to bridge the gap between conventional and event-based vision. Rather than assuming fundamental incompatibility between these domains, we explore a complementary direction: adapting state-of-the-art frame-based detectors to process event data effectively with minimal modifications. In this work, we investigate how RT-DETR can be strategically adapted to the event-based domain, demonstrating that modern object detection architectures can successfully work on event data and opening a new research direction.", + "bbox": [ + 89, + 642, + 482, + 792 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach proceeds in two stages to systematically adapt RT-DETR to event data. First, we train RT-DETR directly on simple image-like representations of EBC data using standard training procedures without special modifications. This straightforward approach yields surprising results - even when processing only fixed $50\\mathrm{ms}$ time frames, the detector achieves performance comparable to special", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ized EBC-specific methods explicitly designed for video object detection on EBC data. This is particularly unexpected as these specialized methods deliberately exploit the temporal nature of EBC data and process information from the distant past, while our approach performs detection using only a single fixed time window (frame) of $50~\\mathrm{ms}$ , similar to traditional frame-based cameras.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this foundation, we introduce I2EvDet (Image-to-Event Detection), a technique that enables straightforward transformation of image-based object detectors into video-capable models. This approach leverages latent space adaptation principles [14, 15] where we freeze the pre-trained RT-DETR model and insert a lightweight recurrent neural network (RNN) module within the encoder's latent representation space. This design allows the model to capture temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead. The resulting adaptation effectively leverages information across multiple time windows without requiring a specialized video detection architecture.", + "bbox": [ + 511, + 209, + 906, + 421 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The combined approach, called EvRT-DETR, achieves state-of-the-art performance on standard EBC benchmarks, outperforming specialized methods on both Gen1 [4] (mAP +2.3) and 1Mpx [22] (mAP +1.4) datasets. Importantly, the results are achieved using only standard modules from natural image and video analysis without requiring domain-specific modifications. This demonstrates not only the effectiveness of our approach for EBC data, but also suggests a broader paradigm for adapting image detectors to temporal domains. As shown in Figure 1, our model achieves superior accuracy while maintaining competitive inference speeds, making it practical for real-world applications.", + "bbox": [ + 511, + 433, + 908, + 617 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contributions of this work:", + "bbox": [ + 531, + 628, + 715, + 642 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We show that when trained on simple image-like representations of EBC data using standard procedures, mainstream object detectors like RT-DETR can achieve performance comparable to specialized EBC methods, offering a complementary approach to domain-specific architectures.", + "- We propose I2EvDet, an efficient adaptation technique that transforms image-based detectors into video-capable models through minimal architectural additions to the frozen pre-trained model and with potential applications to various temporal visual domains.", + "- We validate our approach in the challenging domain of event-based cameras, where EvRT-DETR achieves state-of-the-art performance on the standard Gen1 (mAP +2.3) and 1Mpx (mAP +1.4) EBC benchmarks while maintaining competitive inference speeds." + ], + "bbox": [ + 513, + 656, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "9813", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 91, + 89, + 232, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Research at the intersection of event-based cameras and object detection spans multiple domains in computer vision. Current approaches primarily focus on two directions: designing efficient representations for the unique properties of event data and developing specialized neural architectures to process these representations. However, despite its success in other domains, a third direction – adapting mainstream computer vision models to event data – remains underexplored. This section reviews relevant work across these areas, highlighting how our approach bridges the gap between specialized event-based methods and mainstream object detection advances.", + "bbox": [ + 89, + 114, + 482, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. EBC Data Representations", + "text_level": 1, + "bbox": [ + 89, + 304, + 333, + 321 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "An event camera produces a stream of event data in the form $(t, p, x, y)$ , where $t$ is the timestamp of an event, $p$ is the polarity (positive or negative) of the brightness change, and $(x, y)$ is the location of the event pixel. To simplify the analysis, the event data often are transformed into alternative representations more suitable for conventional image analysis algorithms.", + "bbox": [ + 89, + 325, + 482, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The simplest image-like representations of event camera data are Event Frames or 2D Histograms [6], where events within fixed time windows are accumulated into two-dimensional frames of shape $(2, H, W)$ with polarity as the first dimension and spatial coordinates as the remaining dimensions. A natural evolution is the Stacked 2D Histogram, which further partitions each frame into $T$ time intervals, creating a representation of shape $(2, T, H, W)$ that typically is reshaped to $(2T, H, W)$ and treated as a natural image. While these representations are straightforward, directly applying existing computer vision algorithms to them has yielded poor performance in the past [22].", + "bbox": [ + 89, + 431, + 482, + 613 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To improve performance, other representations are being explored. For example, a Time Surface (TS) representation [6, 33] is an image-like representation where pixel values encode the time elapsed since the last event in a given pixel. Thus, unlike the simplest fixed time window representations, TS potentially can encode arbitrarily distant past. Empirically, it has been shown that using the TS representations offers significantly better performance on the object detection tasks compared to the 2D Histogram representations [22].", + "bbox": [ + 89, + 614, + 482, + 763 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many other EBC data representations also have been proposed. The recent ERGO-12 (Event Representation through Gromov-Wasserstein Optimization) [33] work has developed efficient image-like representations, achieving state-of-the-art object detection performance on the Gen1 dataset. Meanwhile, other approaches reformulate object detection as a three-dimensional (3D) problem by treating events as points in spatial-temporal space or discretizing them into voxel grids [6]. While computationally heavier,", + "bbox": [ + 89, + 763, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "these representations can preserve exact event timing, potentially enabling better handling of overlapping objects and complex motions.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. EBC Object Detection Architectures", + "text_level": 1, + "bbox": [ + 511, + 147, + 830, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Once the data representation is selected, one can start to develop object detection algorithms. There are several families of object detection methods designed for particular data representations.", + "bbox": [ + 511, + 169, + 903, + 229 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most straightforward approach applies conventional image detection methods directly to 2D Histogram frames, but this yields relatively poor performance [22]. More advanced methods, such as Recurrent Vision Transformer (RVT) [9] and State Space Vision Transformer (S5-ViT-B) [34], extend this approach by incorporating temporal memory to capture sequential information, significantly improving detection quality compared to frame-by-frame processing.", + "bbox": [ + 511, + 229, + 903, + 366 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The most recent family of methods attempts to use more efficient event representations to further improve object detection performance. For example, 12-channel ERGO-12 [33] constructs an optimized image-like event representation and shows that the standard object detection methods perform exceptionally well on it. Alternatively, Asynchronous Spatio-Temporal Memory Network (ASTMNet) [16] and Adaptive Event Conversion (AEC) [19] attempt to combine better event representation while using novel neural architectures to expand state-of-the-art results.", + "bbox": [ + 511, + 367, + 903, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Other methods have been developed that treat EBC object detection as a 3D detection problem [8, 22, 26], exploit benefits of neuromorphic architectures [3], or explore various hybrid approaches [20, 21].", + "bbox": [ + 511, + 518, + 903, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Transformer-based Object Detection", + "text_level": 1, + "bbox": [ + 511, + 590, + 830, + 606 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Object detection has been revolutionized by transformer architectures in recent years. DETR [2] was the first widely successful transformer-based [28] model for object detection, offering a simple and elegant alternative to traditional CNN-based architectures [10, 24].", + "bbox": [ + 511, + 612, + 903, + 686 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DETR follows a hybrid CNN-Transformer design with a backbone-encoder-decoder structure. It uses a CNN backbone (typically ResNet [13]) for feature extraction, followed by a Transformer encoder that processes these features into tokens and captures their correlations. The Transformer decoder then identifies objects through cross-attention to the encoder outputs.", + "bbox": [ + 511, + 688, + 903, + 794 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite its innovative approach, the original DETR suffered from limitations, including training instability [29], poor performance on small objects [2], and high computational costs during inference [32]. Subsequent research has addressed these issues through architectural improvements such as Feature Pyramid Networks (FPN), specialized decoder tokens, and deformable attention mechanisms [2, 32].", + "bbox": [ + 511, + 795, + 903, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "9814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bf19cc5e0a7f5a6e87ddbbf7189f387b4035aee94ea1d2429bad0b5ee0590aab.jpg", + "image_caption": [ + "Figure 3. Overview of the I2EvDet Framework Applied to RT-DETR. This approach adapts a pre-trained RT-DETR model to process temporal event data using minimal architectural modifications. The backbone and encoder components remain frozen during temporal adaptation, while trainable temporal modules are strategically inserted at multiple feature scales. Left: Event camera frames are processed through the frozen RT-DETR backbone and encoder. Center: Our temporal adaptation modules operate on multiscale features through residual connections (red pathways), enabling information flow across time steps while preserving spatial representation integrity. Right: Feature fusion and decoder stages produce final detections. The detailed view of the Temporal Adaptation Module (bottom right) shows the projection operations $(W_{i}^{down}, W_{i}^{up})$ and recurrent memory module $(\\mathbf{R}_{i})$ that enable efficient temporal modeling with minimal additional parameters." + ], + "image_footnote": [], + "bbox": [ + 96, + 85, + 911, + 368 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These incremental advances have culminated in the RT-DETR [30], which achieves exceptional performance across all object scales with stable training and significantly faster inference times. RT-DETR represents the current state-of-the-art in efficient transformer-based detection and serves as an ideal foundation for adaptation to specialized domains, such as event-based vision.", + "bbox": [ + 89, + 517, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "2.4. Model Adaptation Techniques", + "text_level": 1, + "bbox": [ + 89, + 635, + 359, + 652 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While specialized architectures dominate event-based vision, adapting mainstream architectures offers compelling advantages. Parameter-efficient methods [12, 14, 15] modify pre-trained models with minimal changes by adding compact modules to frozen backbones. These approaches have proven effective across language [5], vision [31], and multimodal systems [11], but they remain underexplored for event-based vision despite their efficiency and transfer learning benefits.", + "bbox": [ + 89, + 657, + 482, + 792 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the success of adaptation techniques in other domains and the power of transformer-based detection models, there is a clear opportunity to bridge these approaches for event-based vision. Our work explores this intersection, demonstrating that mainstream object detectors can be effectively adapted to EBC data with minimal modifications rather than requiring specialized architectures.", + "bbox": [ + 89, + 795, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 513, + 515, + 612, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This section presents I2EvDet, our adaptation framework for event-based object detection. Unlike approaches that rely on specialized architectures or complex representations, I2EvDet bridges mainstream object detection with temporal event data via strategic, minimal modifications. Our framework demonstrates how image-based detectors can be efficiently transformed into video-capable models while preserving their core strengths. We first describe our image-like representation of event data then introduce the I2EvDet framework. Finally, we detail the technical implementation and discuss how our approach addresses the unique temporal dependencies in event camera data.", + "bbox": [ + 511, + 542, + 906, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Event Representation", + "text_level": 1, + "bbox": [ + 511, + 738, + 718, + 755 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our adaptation framework begins with transforming EBC data into an image-like representation that is compatible with mainstream object detection models. We employ the Stacked 2D Histogram representation [6] as it provides a straightforward bridge between event data and conventional image processing architectures.", + "bbox": [ + 511, + 762, + 905, + 853 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To ensure direct comparability with existing methods, we adopt the frame construction parameters from RVT [9], which allows us to isolate the impact of our architectural", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "9815", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "choices from data pre-processing. Specifically, we take the stream of events and partition it into a series of consecutive fixed time windows of $T_{\\mathrm{frame}} = 50$ ms. Each such window corresponds to a single frame. Next, we further subdivide each frame into 10 intervals of $T_{\\mathrm{bin}} = 5$ ms.", + "bbox": [ + 89, + 90, + 480, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To construct a frame corresponding to an interval $[t_0, t_0 + T_{\\mathrm{frame}}]$ , we create an intermediate stacked histogram $S$ from a set of events $\\mathcal{E}$ in that interval:", + "bbox": [ + 89, + 167, + 482, + 210 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS(p,t_{i},y,x) = \\sum_{\\{e|t_{e}\\in [t_{0},t_{0} + T_{\\text{frame}})\\}}\\delta_{x}^{x_{e}}\\delta_{y}^{y_{e}}\\delta_{p}^{p_{e}}\\mathbf{1}_{[t_{i},t_{i + 1})}(t_{e})\n$$\n", + "text_format": "latex", + "bbox": [ + 96, + 220, + 475, + 255 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $e = (t_e, p_e, x_e, y_e)$ represents an individual event with time stamp $t_e$ , polarity $p_e$ , and spatial coordinates $(x_e, y_e)$ ; $t_i = t_0 + i \\cdot T_{\\mathrm{bin}}$ , $i \\in [0, 10)$ is the histogram bin index; $(x, y)$ are the spatial indices in the output histogram; $\\delta_a^b$ is the Kronecker delta function; and $\\mathbf{1}$ is an indicator function. Once a stacked histogram $S$ of shape $(2, 10, H, W)$ is constructed, we merge the polarity and time dimensions to obtain an image-like 2D frame $F$ of shape $(20, H, W)$ .", + "bbox": [ + 89, + 265, + 482, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. I2EvDet: Image-to-Event Detection Framework", + "text_level": 1, + "bbox": [ + 89, + 395, + 482, + 425 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "I2EvDet is a principled framework for adapting image-based object detectors to process temporal data. While specialized architectures typically are developed for event-based object detection, this approach demonstrates that strategic adaptation of mainstream detectors can achieve superior performance with minimal architectural modifications.", + "bbox": [ + 89, + 431, + 482, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The I2EvDet framework applies latent space adaptation principles [12, 14, 15], which enable domain adaptation through minimal changes to pre-trained models. The I2EvDet framework follows a two-stage process, and it is applicable to any object detection model with a clean separation between feature extraction and object detection modules (e.g.,YOLO and DETR families).", + "bbox": [ + 89, + 537, + 482, + 643 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "First, we establish a solid foundation by training an object detector on individual EBC frames (subsection 3.1), treating them as conventional images. This creates a detector with powerful spatial representation capabilities optimized for event data.", + "bbox": [ + 89, + 643, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In the second stage, we transform this image-based detector into a video-capable model. We freeze the pre-trained detector parameters and insert lightweight RNN modules between the feature extractor and detection components. We train the inserted temporal modules on the EBC video object detection task while keeping the original detector frozen. The resulting model captures temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead.", + "bbox": [ + 89, + 719, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This two-stage approach retains learned spatial representations of the original image detector while efficiently", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "adding temporal capabilities. In this work, we apply I2EvDet to the RT-DETR model as the primary detector, resulting in the EvRT-DETR model. Additionally, we demonstrate the generalizability of the I2EvDet framework using experiments with YOLOX detectors.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Temporal Modeling Considerations", + "text_level": 1, + "bbox": [ + 511, + 175, + 821, + 191 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key theoretical consideration for this adaptation framework is the choice of temporal memory mechanism. Event cameras present a unique challenge: static objects become \"invisible\" as they generate no events when motionless. Their detection requires having a persistent memory of past events. While transformers have dominated recent vision architectures, their fixed context windows limit temporal memory. We employ RNNs for their theoretically unbounded memory capacity through recurrent state, making them uniquely suited for maintaining object persistence during periods of invisibility in the event stream. This theoretical foundation has informed the EvRT-DETR implementation.", + "bbox": [ + 511, + 196, + 906, + 392 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Technical Implementation of EvRT-DETR", + "text_level": 1, + "bbox": [ + 511, + 402, + 872, + 419 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The EvRT-DETR architecture implements the I2EvDet framework via strategic, minimal modifications to the RT-DETR model. As shown in Figure 3, we identify key integration points within the pre-trained model's latent space, where temporal information can be effectively incorporated without disrupting spatial feature extraction.", + "bbox": [ + 511, + 424, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The RT-DETR encoder produces multiscale feature representations $\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}$ , where each element corresponds to a transformer token encoding rich spatial information. These features, originating from the last three stages of the backbone $\\{S_3,S_4,S_5\\}$ [30], provide an ideal latent manifold for temporal adaptation.", + "bbox": [ + 511, + 515, + 905, + 604 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following latent space adaptation principles, we introduce specialized temporal processing modules at each scale, implementing three parallel RNN-based memory modules $\\{\\mathbf{R}_3,\\mathbf{R}_4,\\mathbf{R}_5\\}$ . This multiscale temporal adaptation design ensures that temporal dynamics are captured at different levels of feature abstraction. Critically, each RNN module interfaces with its corresponding encoder feature map through residual connections, mathematically expressed as:", + "bbox": [ + 511, + 606, + 905, + 727 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {i} ^ {t, p r o j} = W _ {i} ^ {d o w n} \\cdot \\mathcal {E} _ {i} ^ {t} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 750, + 903, + 770 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\mathcal {O} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t}\\right) = \\mathbf {R} _ {i} \\left(\\mathcal {E} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t - 1}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 772, + 903, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {O} _ {i} ^ {t} = W _ {i} ^ {u p} \\cdot \\mathcal {O} _ {i} ^ {t, p r o j} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 660, + 800, + 903, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {E}} _ {i} ^ {t} = \\mathcal {E} _ {i} ^ {t} + \\alpha_ {i} \\cdot \\mathcal {O} _ {i} ^ {t}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 661, + 821, + 903, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{i}^{down}\\in \\mathbb{R}^{C_{i}^{RNN}\\times C_{i}}$ and $W_{i}^{up}\\in \\mathbb{R}^{C_{i}\\times C_{i}^{RNN}}$ are the projection matrices that transform features into and out of the temporal processing space, $\\mathcal{M}_i^t\\in \\mathbb{R}^{C_i^{RNN}\\times H_i\\times W_i}$", + "bbox": [ + 511, + 849, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "9816", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a25c8b8d2bafef5fdaf3a86cf3a6f7f6bacec7cb889372b5dc816088a99d152b.jpg", + "image_caption": [ + "B" + ], + "image_footnote": [], + "bbox": [ + 163, + 88, + 302, + 150 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c8deaa347e462b8fe6bc41dfc358f059ca0987cae5cacba0e9138690e3969c0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 88, + 441, + 150 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2854ed932a2c056a16761c76c64e74bbab82b448b302eb8e0b90e8ba3f6b6c8d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 88, + 583, + 150 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4c246c15051eb71ba1eed325546ed123c0ca0a95cb047d22f8086cadca6dccc2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 89, + 722, + 151 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bd45114d2465646c2d6672a47b8eba8b576245a1fb8a5c6b303bb00da5fec233.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 89, + 864, + 151 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d4f8d120c7db1903aa90174a3b32261c075bf5246981b7b8292cd7168ffe380d.jpg", + "image_caption": [ + "Figure 4. Temporal Adaptation Benefits of EvRT-DETR. Visualization of object detection across sequential 1Mpx frames captured by a vehicle-mounted camera that gradually comes to a stop. While the camera is in motion (left frames), both detectors maintain accurate bounding boxes of stationary vehicles. As camera motion decreases (middle frames), event data becomes increasingly sparse. When the camera stops completely (right frames), RT-DETR (top row) loses detection capability, while EvRT-DETR (bottom row) maintains consistent detection by engaging its temporal memory module." + ], + "image_footnote": [], + "bbox": [ + 163, + 160, + 302, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/615b91d1f381ebb892b4ae92e06f6450bfb463353537c761ccca95655bb13157.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 303, + 160, + 441, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d428df93af6a95718f09b10ede61dc34ee458a16916893b5893782e53d25b028.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 444, + 160, + 583, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/05610d3f100e0d20f8b1dc2271f7d9c7f62f6dde6cc744b2918332a766334dd0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 586, + 160, + 722, + 222 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d11a37e3cf9f1dade38af3dc95a6d085faf0e55a41376b33f5c7a47aa2f3c300.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 723, + 160, + 864, + 222 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "represents the temporal memory state for the $i$ -th feature scale at time $t$ , $\\mathcal{O}_i^t$ denotes the corresponding temporal adaptation output, and $\\alpha_{i}$ is a learnable scaling factor that modulates the influence of temporal information on the spatial representation inspired by the ReZero technique [1]. In our implementation, the input projection matrix $W_{i}^{down}$ is integrated within the RNN architecture, which already handles dimensionality reduction through its internal gating mechanisms. Meanwhile, the output projection $W_{i}^{up}$ is explicitly implemented to ensure proper integration back into the frozen RT-DETR architecture. Thus, the transformed feature representation $\\tilde{\\mathcal{E}}_i^t$ incorporates temporal context while maintaining the integrity of the original spatial representation $\\mathcal{E}_i^t$ . In addition, our ablation studies show that $C_i^{RNN}$ can be reduced 4x relative to $C_i$ with minimal performance impact.", + "bbox": [ + 91, + 334, + 482, + 574 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For specific implementation of the RNN modules, we employ the ConvLSTM architecture [27], which effectively balances representational capacity with computational efficiency. This choice was informed by ConvLSTM's demonstrated effectiveness in handling spatio-temporal data in the RVT model [9], allowing our adaptation to employ established temporal modeling techniques.", + "bbox": [ + 89, + 575, + 482, + 681 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation of the EvRT-DETR Model", + "text_level": 1, + "bbox": [ + 89, + 693, + 429, + 709 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The EvRT-DETR model is evaluated on two standard EBC object detection benchmarks: Gen1 [4] and 1Mpx [22] (also known as Gen4) datasets. These datasets contain diverse driving scenarios captured by Prophesee's neuromorphic vision sensors, providing an ideal testbed for validating our adaptation approach under real-world conditions.", + "bbox": [ + 89, + 719, + 482, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setting", + "text_level": 1, + "bbox": [ + 89, + 818, + 290, + 834 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We evaluate on two standard Propheesee automotive datasets: Gen1 [4] and 1Mpx [22]. Gen1 contains $39+$ hours of recordings at $304 \\times 240$ resolution with manual annotations (1-4 Hz) for cars and pedestrians, presenting", + "bbox": [ + 89, + 840, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "challenges with sparse labels and stationary objects. 1Mpx offers $14+$ hours at higher resolution $(1280 \\times 720)$ with denser automatic annotations $(60\\mathrm{Hz})$ for cars, pedestrians, and two-wheelers. Both datasets are organized as 60-second clips with standard train/test splits. 1Mpx's higher resolution and frame rate enable evaluation on more complex scenes, while Gen1's annotation pattern tests robustness to sparse temporal supervision.", + "bbox": [ + 511, + 334, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Event Representations. We adopt the well-established stacked 2D Histogram approach from RVT protocols [9], constructing event frames from each $50\\mathrm{ms}$ time window. Each frame is subdivided into 10 equally time-spaced $5\\mathrm{ms}$ intervals, creating a representation with sufficient temporal resolution for our adaptation framework while remaining compatible with standard image processing architectures.", + "bbox": [ + 511, + 473, + 905, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For the Gen1 dataset (original size (240, 304)), we pad frames to (256, 320) to ensure divisibility by 32 as required by the feature pyramid architecture of RT-DETR. For 1Mpx, we downsize from (720, 1280) to (360, 640) using bilinear interpolation and pad to (384, 640). This maintains consistency with prior work while allowing for fair computational comparison.", + "bbox": [ + 511, + 579, + 905, + 685 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Two-stage Training Process. We train EvRT-DETR in two stages: (1) training the RT-DETR detector on individual event frames using standard practices (Adam optimizer; EMA of weights) then (2) freezing these parameters and training only the lightweight RNN modules on both random and sequential clips to develop temporal memory. The first stage establishes robust spatial feature extraction, while the second adds temporal processing with minimal parameter overhead. For temporal training, we maintain RNN memory continuity across sequential clips while resetting for random clips using different clip lengths for Gen1 (21 frames) and 1Mpx (10 frames). Full implementation details are provided in the supplementary material.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "9817", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/69aba0c7e85d5dac0f7bb1de890ca83dd17138ac9c06703e2c6654fd597706e0.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGen11Mpx/Gen4Params (M)
mAP (%)Runtime (ms)mAP (%)Runtime (ms)
AsyNet [18]14.5---11.4
AEGNN [26]16.3---20.0
Spiking DenseNet [3]18.9---8.2
Events-RetinaNet [22]34.0-18.0-32.8
RED [22]40.016.743.039.324.1
AEC [19]47.010.6 (3.9)48.437.6 (13.8)46.5
RVT-B [9]47.210.247.411.918.5
GET-T [20]47.916.848.418.221.9
S5-ViT-B [34]47.48.247.29.617.5
ASTMNet [16]46.735.648.372.3>100
SAST-CB [21]48.2-48.757.5 (19.7)18.9
ERGO-12 [33]50.469.940.6100.059.6
RT-DETR-T (ours)46.06.444.18.620.1
RT-DETR-B (ours)47.610.545.214.942.8
EvRT-DETR-T (ours)52.38.449.912.534.4
EvRT-DETR-B (ours)52.712.750.118.857.1
", + "bbox": [ + 168, + 87, + 828, + 368 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Performance Comparison on Event-based Camera Datasets. Our approach achieves state-of-the-art results on both Gen1 and 1Mpx benchmarks with significant improvements over previous methods. The first stage of the framework (RT-DETR-T/B) already achieves competitive performance using only image-based detection, while the proposed I2EvDet-adapted models (EvRT-DETR-T/B) surpass all existing approaches. Runtime measurements are reported in milliseconds on an NVIDIA T4 GPU. Values in parentheses represent original reported times on different hardware with T4-equivalent times calculated using FLOPS/throughput conversion.", + "bbox": [ + 89, + 378, + 906, + 450 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Evaluation Protocol. We evaluate EvRT-DETR using the standard COCO (Common Objects in Context) Mean Average Precision (mAP) measure [17] that allows direct comparison with existing methods. For consistency with previous works, we use the EBC-specific implementation of the COCO metrics provided by Propheesee's Automotive Dataset Toolbox [22, 23]. All inference times reported in Table 2 are standardized to NVIDIA T4 GPU performance for fair comparison. Inference times include model forward pass and post-processing but exclude event preprocessing, following standard reporting practices. When prior work reported only preprocessing-inclusive times, we use their reported values directly.", + "bbox": [ + 88, + 474, + 482, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Performance Analysis", + "text_level": 1, + "bbox": [ + 89, + 681, + 294, + 698 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1 presents a comprehensive comparison of our EvRT-DETR models against state-of-the-art approaches on the Gen1 and 1Mpx benchmarks.", + "bbox": [ + 89, + 704, + 482, + 748 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "First, our baseline adaptation using two RT-DETR variants, RT-DETR-T (Tiny, with ResNet-18 backbone) and RT-DETR-B (Base, with ResNet-50 backbone), and trained on simple event frames achieves competitive performance (46.0% and 47.6% mAP on Gen1; 44.1% and 45.2% mAP on 1Mpx) without requiring specialized architectures or complex event representations. This demonstrates that modern mainstream object detectors can be effectively adapted to event data through appropriate training.", + "bbox": [ + 89, + 750, + 482, + 885 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "More importantly, our complete EvRT-DETR models", + "bbox": [ + 109, + 885, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "with temporal adaptation significantly outperform all existing methods, achieving new state-of-the-art results on both benchmarks. EvRT-DETR-B reaches $52.7\\%$ mAP on Gen1 $(+2.3\\%$ over the previous best) and $50.1\\%$ mAP on 1Mpx $(+1.4\\%$ over the previous best). Even our lighter EvRT-DETR-T model surpasses all existing approaches while maintaining faster inference times.", + "bbox": [ + 511, + 474, + 906, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 shows how the temporal adaptation module provides crucial capabilities for handling stationary objects - a fundamental challenge for event cameras. When objects are in motion, both RT-DETR and EvRT-DETR perform well, yet, when motion stops and event generation becomes minimal, EvRT-DETRs temporal memory enables consistent detection.", + "bbox": [ + 511, + 580, + 906, + 685 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "These performance improvements come with minimal computational overhead. While we focus primarily on detection accuracy in this work, the runtime measurements in Table 1 demonstrate that our approach maintains competitive inference speeds despite adding temporal processing capabilities.", + "bbox": [ + 511, + 686, + 908, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Framework Generalizability", + "text_level": 1, + "bbox": [ + 511, + 787, + 769, + 804 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate that our results are not specific to the RT-DETR architecture, we apply the I2EvDet framework to YOLOX [7]. YOLOX is a CNN-based detector that contrasts with RT-DETR's transformer architecture and has been used in other EBC applications [9, 34]. Table 2 shows that our temporal adaptation approach consistently", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "9818", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a4eb61df46bf09a353800c6020e6a0439b906190c0a1495bfe55dda1f012a60c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
DetectormAP (%)
OriginalI2EvDetImprovement
RT-DETR-T46.052.3+6.3
RT-DETR-B47.652.7+5.1
YOLOX-T36.042.4+6.4
YOLOX-X43.447.8+4.4
", + "bbox": [ + 130, + 88, + 439, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "improves performance across different baseline architectures on the Gen1 dataset. The substantial improvements (4.4-6.4 mAP) across both RT-DETR and YOLOX variants validate that I2EvDet represents a general framework for adapting image-based detectors to temporal domains rather than an architecture-specific optimization. Additional details are provided in the supplementary material.", + "bbox": [ + 89, + 261, + 483, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Analysis of Adaptation Design Choices", + "text_level": 1, + "bbox": [ + 89, + 376, + 423, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We systematically analyze key design choices in our temporal adaptation module to better understand the factors contributing to EvRT-DETR's performance. This analysis provides insights into efficient adaptation strategies for event-based object detection. Additional ablation studies are provided in the supplementary material.", + "bbox": [ + 89, + 398, + 483, + 489 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/593c5f52f2ba016805122c8eae52ecb02e17d14de48b351fae34f5300937cabe.jpg", + "table_caption": [ + "Table 2. Generalizability of I2EvDet Framework Across Different Object Detection Architectures on the Gen1 Dataset. Our temporal adaptation consistently improves performance regardless of the base detector architecture." + ], + "table_footnote": [], + "table_body": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)
RT-DETR-B (ours)47.676.349.5
EvRT-DETR-B (0,1,1)51.080.754.0
EvRT-DETR-B (1,0,1)52.281.355.1
EvRT-DETR-B (1,1,0)52.481.755.3
EvRT-DETR-B (1,1,1)52.782.056.0
", + "bbox": [ + 91, + 500, + 478, + 583 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.1. Multiscale Temporal Adaptation", + "text_level": 1, + "bbox": [ + 89, + 685, + 364, + 700 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "A crucial aspect of our approach is the insertion of temporal modules at multiple scales in the feature hierarchy. Table 3 shows performance when selectively removing ConvLSTM cells from specific feature levels with the array notation $(x,y,z)$ indicating presence (1) or absence (0) at each scale $\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}$ .", + "bbox": [ + 89, + 704, + 482, + 795 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The results reveal a clear pattern: the lowest-level features $(\\mathcal{E}_3)$ benefit most significantly from temporal adaptation as performance drops substantially when this module is removed. The impact diminishes at higher feature levels, suggesting that temporal information is most valuable for fine-grained, lower-level features that capture detailed motion patterns. This insight can guide more efficient adapta", + "bbox": [ + 89, + 795, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion designs, wherein computational resources are directed to where they provide maximum benefit.", + "bbox": [ + 511, + 90, + 906, + 122 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/101f0915dd8319a68ace1df8b6b9f675e734f17884f15ffb31cca604262e10d6.jpg", + "table_caption": [ + "Table 3. Effect of Temporal Module Placement. Performance on the Gen1 dataset with ConvLSTM modules at different feature scales. The array notation $(x,y,z)$ indicates presence (1) or absence (0) of temporal modules at scales $\\mathcal{E}_3$ , $\\mathcal{E}_4$ , and $\\mathcal{E}_5$ , respectively. Low-level features benefit most from temporal adaptation." + ], + "table_footnote": [], + "table_body": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)\\( N_T \\) (M)
EvRT-DETR-B (M=64)52.181.354.92.3
EvRT-DETR-B (M=128)52.581.755.65.4
EvRT-DETR-B (M=256)52.782.056.014.4
EvRT-DETR-B (M=512)52.981.956.042.9
", + "bbox": [ + 514, + 136, + 901, + 195 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Impact of Temporal Module Capacity. Performance on Gen1 dataset with varying ConvLSTM hidden dimensions (M). The rightmost column shows the number of trainable parameters (in millions) in the temporal adaptation module.", + "bbox": [ + 511, + 205, + 906, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.2. Parameter Efficiency in Adaptation", + "text_level": 1, + "bbox": [ + 511, + 291, + 807, + 306 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Inspired by parameter-efficient techniques such as LoRA (Low-Rank Adaptation) [15], we investigate the minimal temporal module capacity needed for effective adaptation. Table 4 shows EvRT-DETR-B performance with varying hidden dimensions in the ConvLSTM modules.", + "bbox": [ + 511, + 311, + 905, + 386 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Remarkably, reducing hidden dimensions to 64 features (a $4\\mathrm{x}$ reduction from our baseline) results in only modest performance degradation $(-0.6\\% \\mathrm{mAP})$ while requiring just $2.3\\mathrm{M}$ additional parameters - merely $5.4\\%$ of the base RT-DETR-B model's $42.8\\mathrm{M}$ parameters. This configuration closely resembles LoRA-style adaptation, achieving substantial performance gains with minimal parameter overhead. Even our most parameter-efficient adaptation $(\\mathbf{M} = 64)$ outperforms all previous state-of-the-art approaches, demonstrating that effective temporal adaptation does not require extensive architectural modifications. Increasing the hidden dimension to 512 features $(42.9\\mathrm{M}$ parameters, nearly doubling the model size) yields only a marginal $0.2\\% \\mathrm{mAP}$ improvement, indicating diminishing returns.", + "bbox": [ + 511, + 388, + 906, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 632, + 638, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work introduces I2EvDet, a framework that efficiently transforms image-based detectors into video-capable models using minimal modifications. Applied to event-based cameras, our approach demonstrates that mainstream object detection architectures can be effectively adapted to temporal domains without specialized redesign. RT-DETR trained on simple event representations achieves performance comparable to specialized methods, while our I2EvDet-adapted model EvRT-DETR achieves state-of-the-art results on Gen1 $(+2.3\\mathrm{mAP})$ and 1Mpx $(+1.4\\mathrm{mAP})$ benchmarks. Ablation studies provide insights about optimal temporal module design and parameter efficiency. These results suggest that bridging mainstream computer vision and specialized domains can be achieved through targeted adaptation rather than complete architectural redesign, potentially accelerating progress across diverse temporal visual domains.", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "9819", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The authors extend their sincere gratitude to Yuewei Lin (Brookhaven Lab) for his valuable feedback and suggestions. This work presented in this paper was funded by the U.S. Department of Energy (DOE), National Nuclear Security Administration, Office of Defense Nuclear Nonproliferation Research and Development. The manuscript has been authored by Brookhaven National Laboratory managed by Brookhaven Science Associates, LLC under the Contract No. DE-SC0012704 with the U.S. Department of Energy.", + "bbox": [ + 89, + 114, + 485, + 252 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 263, + 187, + 277 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Gary Cottrell, and Julian J. McAuley. ReZero is all you need: fast convergence at large depth. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, UAI 2021, Virtual Event, 27-30 July 2021, pages 1352-1361. AUAI Press, 2021. 6", + "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213-229. Springer, 2020. 2, 3", + "[3] Loïc Cordone, Benoit Miramond, and Philippe Thiérison. Object detection with spiking neural networks on automotive event data. In International Joint Conference on Neural Networks, IJCNN 2022, Padua, Italy, July 18-23, 2022, pages 1-8. IEEE, 2022. 3, 7", + "[4] Pierre de Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. CoRR, abs/2001.08499, 2020. 2, 6", + "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4", + "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jorg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3, 4", + "[7] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 7", + "[8] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. CoRR, abs/2211.12324, 2022. 3", + "[9] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13884-13893. IEEE, 2023. 3, 4, 6, 7" + ], + "bbox": [ + 99, + 287, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 580-587. IEEE Computer Society, 2014. 3", + "[11] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. Trans. Mach. Learn. Res., 2024, 2024. 4", + "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4, 5", + "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 3", + "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 2790-2799. PMLR, 2019. 2, 4, 5", + "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 4, 5, 8", + "[16] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Trans. Image Process., 31:2975-2987, 2022. 3, 7", + "[17] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7", + "[18] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII, pages 415-431. Springer, 2020. 7", + "[19] Yansong Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington," + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "9820", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "DC, USA, February 7-14, 2023, pages 2056-2064. AAAI Press, 2023. 3, 7", + "[20] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. GET: group event transformer for event-based vision. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 6015-6025. IEEE, 2023. 2, 3, 7", + "[21] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16794-16804. IEEE, 2024. 3, 7", + "[22] Etienne Perot, Pierre de Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6, 7", + "[23] Propheesee. Propheesee's automotive dataset toolbox. https://github.com/propheesee-ai/propheesee-automotive-dataset-toolbox, 2024.7", + "[24] Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 779-788. IEEE Computer Society, 2016. 2, 3", + "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 91-99, 2015. 2", + "[26] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. AEGNN: asynchronous event-based graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12361-12371. IEEE, 2022. 3, 7", + "[27] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 802-810, 2015. 6", + "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 3", + "[29] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3", + "[30] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on Real-time object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16965-16974. IEEE, 2024. 2, 4, 5", + "[31] Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. Convolution meets lora: Parameter efficient finetuning for segment anything model. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 4", + "[32] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3", + "[33] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 12800-12810. IEEE, 2023. 2, 3, 7", + "[34] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5819-5828. IEEE, 2024. 3, 7" + ], + "bbox": [ + 516, + 92, + 903, + 512 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "9821", + "bbox": [ + 482, + 945, + 513, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_model.json b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_model.json new file mode 100644 index 0000000000000000000000000000000000000000..de435d3299d77a4bee48fe3d535253ff8fb8e04e --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.203, + 0.131, + 0.797, + 0.175 + ], + "angle": 0, + "content": "EvRT-DETR: Latent Space Adaptation of Image Detectors for Event-based Vision" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.203, + 0.812, + 0.248 + ], + "angle": 0, + "content": "Dmitrii Torbunov1, Yihui Ren1, Animesh Ghose1, Odera Dim1, Yonggang Cui1 \n1Brookhaven National Laboratory, Upton, NY, USA" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.256, + 0.838, + 0.271 + ], + "angle": 0, + "content": "dtorbunov@bnl.gov, yren@bnl.gov, aghose@bnl.gov, dodera@bnl.gov, ycui@bnl.gov" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.306, + 0.327, + 0.321 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.34, + 0.486, + 0.794 + ], + "angle": 0, + "content": "Event-based cameras (EBCs) have emerged as a bioinspired alternative to traditional cameras, offering advantages in power efficiency, temporal resolution, and high dynamic range. However, development of image analysis methods for EBCs is challenging due to the sparse and asynchronous nature of the data. This work addresses the problem of object detection for EBC cameras. The current approaches to EBC object detection focus on constructing complex data representations and rely on specialized architectures. We introduce I2EvDet (Image-to-Event Detection), a novel adaptation framework that bridges mainstream object detection with temporal event data processing. First, we demonstrate that a Real-Time DEetection Transformer, or RT-DETR, a state-of-the-art natural image detector, trained on a simple image-like representation of the EBC data achieves performance comparable to specialized EBC methods. Next, as part of our framework, we develop an efficient adaptation technique that transforms image-based detectors into event-based detection models by modifying their frozen latent representation space via minimal architectural additions. The resulting EvRT-DETR model reaches state-of-the-art performance on the standard benchmark datasets Gen1 (mAP +2.3) and 1Mpx/Gen4 (mAP +1.4). These results demonstrate a fundamentally new approach to EBC object detection through principled adaptation of mainstream architectures, offering an efficient alternative with potential applications to other temporal visual domains. The code is available at: https://github.com/realtime-intelligence/evrt-detr." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.829, + 0.222, + 0.844 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Event-based cameras (EBCs) present a biologically inspired alternative to traditional frame-based cameras. Unlike conventional cameras that capture frames at predetermined in" + }, + { + "type": "image", + "bbox": [ + 0.521, + 0.308, + 0.903, + 0.527 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.541, + 0.908, + 0.64 + ], + "angle": 0, + "content": "Figure 1. Object Detection Performance vs. Inference Time. A summary of object detection performance (COCO mAP) versus inference time (ms) of various models on the 1Mpx automotive dataset. Circle size is proportional to the number of model parameters. Inference times are reported for NVIDIA T4 GPU. Our EvRT-DETR models achieve state-of-the-art accuracy while maintaining competitive inference speeds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.671, + 0.909, + 0.838 + ], + "angle": 0, + "content": "tervals, EBC pixels operate asynchronously, generating data only when brightness changes exceed a threshold. This results in sparse, asynchronous data streams where each event is a tuple containing pixel location, timestamp, and polarity (whether brightness increased or decreased), as illustrated in Figure 2. This approach offers remarkable advantages: low power consumption (as low as \\(10\\mathrm{mW}\\)), reduced data transfer rates, exceptional temporal resolution (on the order of \\(\\mu s\\)), and high dynamic range (\\(>100\\) dB). These properties have led to widespread adoption in autonomous driving, robotics, and wearable electronics [6]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "These unique characteristics of EBC data - sparsity, asynchronicity, and temporal nature - create fundamental challenges when applying traditional computer vision techniques. Early attempts to directly apply conventional" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "9812" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.483, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.191, + 0.483, + 0.234 + ], + "angle": 0, + "content": "Figure 2. Event-based Camera Output Visualization. Sample frames from the 1Mpx dataset showing event polarity. Red indicates positive events, and blue denotes negative events." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.263, + 0.483, + 0.459 + ], + "angle": 0, + "content": "image-based object detectors to simple event representations yielded poor performance [22], highlighting the gap between these domains. This led to two predominant research directions: (1) constructing sophisticated event representations that transform EBC data into formats compatible with existing detectors (e.g., [33]) or (2) designing specialized architectures tailored to handle the temporal nature of event data (e.g., [20]). A common assumption in these approaches is that EBC data processing requires either specialized representations or architectures due to the fundamental differences between event- and frame-based vision, leading to development paths that often diverge from mainstream computer vision research." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.642 + ], + "angle": 0, + "content": "Meanwhile, mainstream object detection has seen remarkable progress, evolving from convolutional neural network (CNN)-based architectures, e.g.,YOLO [24] and Faster R-CNN [25], to transformer-based models like DETR (DEtection Transformer) [2]. Recent advances have culminated in highly efficient and accurate models, such as the Real-Time DETection TRansformer (RT-DETR) [30] that achieves exceptional performance while maintaining real-time inference speeds. Though primarily optimized for processing individual images, these state-of-the-art detectors offer powerful feature extraction capabilities that potentially could be leveraged for temporal data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.794 + ], + "angle": 0, + "content": "This evolution in mainstream detection creates an opportunity to bridge the gap between conventional and event-based vision. Rather than assuming fundamental incompatibility between these domains, we explore a complementary direction: adapting state-of-the-art frame-based detectors to process event data effectively with minimal modifications. In this work, we investigate how RT-DETR can be strategically adapted to the event-based domain, demonstrating that modern object detection architectures can successfully work on event data and opening a new research direction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our approach proceeds in two stages to systematically adapt RT-DETR to event data. First, we train RT-DETR directly on simple image-like representations of EBC data using standard training procedures without special modifications. This straightforward approach yields surprising results - even when processing only fixed \\(50\\mathrm{ms}\\) time frames, the detector achieves performance comparable to special" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ized EBC-specific methods explicitly designed for video object detection on EBC data. This is particularly unexpected as these specialized methods deliberately exploit the temporal nature of EBC data and process information from the distant past, while our approach performs detection using only a single fixed time window (frame) of \\(50~\\mathrm{ms}\\), similar to traditional frame-based cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.21, + 0.907, + 0.422 + ], + "angle": 0, + "content": "Building on this foundation, we introduce I2EvDet (Image-to-Event Detection), a technique that enables straightforward transformation of image-based object detectors into video-capable models. This approach leverages latent space adaptation principles [14, 15] where we freeze the pre-trained RT-DETR model and insert a lightweight recurrent neural network (RNN) module within the encoder's latent representation space. This design allows the model to capture temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead. The resulting adaptation effectively leverages information across multiple time windows without requiring a specialized video detection architecture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.909, + 0.618 + ], + "angle": 0, + "content": "The combined approach, called EvRT-DETR, achieves state-of-the-art performance on standard EBC benchmarks, outperforming specialized methods on both Gen1 [4] (mAP +2.3) and 1Mpx [22] (mAP +1.4) datasets. Importantly, the results are achieved using only standard modules from natural image and video analysis without requiring domain-specific modifications. This demonstrates not only the effectiveness of our approach for EBC data, but also suggests a broader paradigm for adapting image detectors to temporal domains. As shown in Figure 1, our model achieves superior accuracy while maintaining competitive inference speeds, making it practical for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.629, + 0.716, + 0.643 + ], + "angle": 0, + "content": "Contributions of this work:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.657, + 0.905, + 0.746 + ], + "angle": 0, + "content": "- We show that when trained on simple image-like representations of EBC data using standard procedures, mainstream object detectors like RT-DETR can achieve performance comparable to specialized EBC methods, offering a complementary approach to domain-specific architectures." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.748, + 0.905, + 0.822 + ], + "angle": 0, + "content": "- We propose I2EvDet, an efficient adaptation technique that transforms image-based detectors into video-capable models through minimal architectural additions to the frozen pre-trained model and with potential applications to various temporal visual domains." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.823, + 0.905, + 0.899 + ], + "angle": 0, + "content": "- We validate our approach in the challenging domain of event-based cameras, where EvRT-DETR achieves state-of-the-art performance on the standard Gen1 (mAP +2.3) and 1Mpx (mAP +1.4) EBC benchmarks while maintaining competitive inference speeds." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.657, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9813" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.233, + 0.107 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.483, + 0.297 + ], + "angle": 0, + "content": "Research at the intersection of event-based cameras and object detection spans multiple domains in computer vision. Current approaches primarily focus on two directions: designing efficient representations for the unique properties of event data and developing specialized neural architectures to process these representations. However, despite its success in other domains, a third direction – adapting mainstream computer vision models to event data – remains underexplored. This section reviews relevant work across these areas, highlighting how our approach bridges the gap between specialized event-based methods and mainstream object detection advances." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.305, + 0.334, + 0.322 + ], + "angle": 0, + "content": "2.1. EBC Data Representations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.327, + 0.483, + 0.432 + ], + "angle": 0, + "content": "An event camera produces a stream of event data in the form \\((t, p, x, y)\\), where \\(t\\) is the timestamp of an event, \\(p\\) is the polarity (positive or negative) of the brightness change, and \\((x, y)\\) is the location of the event pixel. To simplify the analysis, the event data often are transformed into alternative representations more suitable for conventional image analysis algorithms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.614 + ], + "angle": 0, + "content": "The simplest image-like representations of event camera data are Event Frames or 2D Histograms [6], where events within fixed time windows are accumulated into two-dimensional frames of shape \\((2, H, W)\\) with polarity as the first dimension and spatial coordinates as the remaining dimensions. A natural evolution is the Stacked 2D Histogram, which further partitions each frame into \\(T\\) time intervals, creating a representation of shape \\((2, T, H, W)\\) that typically is reshaped to \\((2T, H, W)\\) and treated as a natural image. While these representations are straightforward, directly applying existing computer vision algorithms to them has yielded poor performance in the past [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.615, + 0.483, + 0.765 + ], + "angle": 0, + "content": "To improve performance, other representations are being explored. For example, a Time Surface (TS) representation [6, 33] is an image-like representation where pixel values encode the time elapsed since the last event in a given pixel. Thus, unlike the simplest fixed time window representations, TS potentially can encode arbitrarily distant past. Empirically, it has been shown that using the TS representations offers significantly better performance on the object detection tasks compared to the 2D Histogram representations [22]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Many other EBC data representations also have been proposed. The recent ERGO-12 (Event Representation through Gromov-Wasserstein Optimization) [33] work has developed efficient image-like representations, achieving state-of-the-art object detection performance on the Gen1 dataset. Meanwhile, other approaches reformulate object detection as a three-dimensional (3D) problem by treating events as points in spatial-temporal space or discretizing them into voxel grids [6]. While computationally heavier," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "these representations can preserve exact event timing, potentially enabling better handling of overlapping objects and complex motions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.148, + 0.831, + 0.164 + ], + "angle": 0, + "content": "2.2. EBC Object Detection Architectures" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.17, + 0.905, + 0.23 + ], + "angle": 0, + "content": "Once the data representation is selected, one can start to develop object detection algorithms. There are several families of object detection methods designed for particular data representations." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.231, + 0.905, + 0.367 + ], + "angle": 0, + "content": "The most straightforward approach applies conventional image detection methods directly to 2D Histogram frames, but this yields relatively poor performance [22]. More advanced methods, such as Recurrent Vision Transformer (RVT) [9] and State Space Vision Transformer (S5-ViT-B) [34], extend this approach by incorporating temporal memory to capture sequential information, significantly improving detection quality compared to frame-by-frame processing." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.368, + 0.905, + 0.519 + ], + "angle": 0, + "content": "The most recent family of methods attempts to use more efficient event representations to further improve object detection performance. For example, 12-channel ERGO-12 [33] constructs an optimized image-like event representation and shows that the standard object detection methods perform exceptionally well on it. Alternatively, Asynchronous Spatio-Temporal Memory Network (ASTMNet) [16] and Adaptive Event Conversion (AEC) [19] attempt to combine better event representation while using novel neural architectures to expand state-of-the-art results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.519, + 0.905, + 0.58 + ], + "angle": 0, + "content": "Other methods have been developed that treat EBC object detection as a 3D detection problem [8, 22, 26], exploit benefits of neuromorphic architectures [3], or explore various hybrid approaches [20, 21]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.591, + 0.831, + 0.607 + ], + "angle": 0, + "content": "2.3. Transformer-based Object Detection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.688 + ], + "angle": 0, + "content": "Object detection has been revolutionized by transformer architectures in recent years. DETR [2] was the first widely successful transformer-based [28] model for object detection, offering a simple and elegant alternative to traditional CNN-based architectures [10, 24]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.689, + 0.905, + 0.795 + ], + "angle": 0, + "content": "DETR follows a hybrid CNN-Transformer design with a backbone-encoder-decoder structure. It uses a CNN backbone (typically ResNet [13]) for feature extraction, followed by a Transformer encoder that processes these features into tokens and captures their correlations. The Transformer decoder then identifies objects through cross-attention to the encoder outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.905, + 0.903 + ], + "angle": 0, + "content": "Despite its innovative approach, the original DETR suffered from limitations, including training instability [29], poor performance on small objects [2], and high computational costs during inference [32]. Subsequent research has addressed these issues through architectural improvements such as Feature Pyramid Networks (FPN), specialized decoder tokens, and deformable attention mechanisms [2, 32]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9814" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.087, + 0.913, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.492 + ], + "angle": 0, + "content": "Figure 3. Overview of the I2EvDet Framework Applied to RT-DETR. This approach adapts a pre-trained RT-DETR model to process temporal event data using minimal architectural modifications. The backbone and encoder components remain frozen during temporal adaptation, while trainable temporal modules are strategically inserted at multiple feature scales. Left: Event camera frames are processed through the frozen RT-DETR backbone and encoder. Center: Our temporal adaptation modules operate on multiscale features through residual connections (red pathways), enabling information flow across time steps while preserving spatial representation integrity. Right: Feature fusion and decoder stages produce final detections. The detailed view of the Temporal Adaptation Module (bottom right) shows the projection operations \\((W_{i}^{down}, W_{i}^{up})\\) and recurrent memory module \\((\\mathbf{R}_{i})\\) that enable efficient temporal modeling with minimal additional parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.625 + ], + "angle": 0, + "content": "These incremental advances have culminated in the RT-DETR [30], which achieves exceptional performance across all object scales with stable training and significantly faster inference times. RT-DETR represents the current state-of-the-art in efficient transformer-based detection and serves as an ideal foundation for adaptation to specialized domains, such as event-based vision." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.636, + 0.36, + 0.654 + ], + "angle": 0, + "content": "2.4. Model Adaptation Techniques" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.794 + ], + "angle": 0, + "content": "While specialized architectures dominate event-based vision, adapting mainstream architectures offers compelling advantages. Parameter-efficient methods [12, 14, 15] modify pre-trained models with minimal changes by adding compact modules to frozen backbones. These approaches have proven effective across language [5], vision [31], and multimodal systems [11], but they remain underexplored for event-based vision despite their efficiency and transfer learning benefits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Given the success of adaptation techniques in other domains and the power of transformer-based detection models, there is a clear opportunity to bridge these approaches for event-based vision. Our work explores this intersection, demonstrating that mainstream object detectors can be effectively adapted to EBC data with minimal modifications rather than requiring specialized architectures." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.516, + 0.613, + 0.532 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.543, + 0.907, + 0.726 + ], + "angle": 0, + "content": "This section presents I2EvDet, our adaptation framework for event-based object detection. Unlike approaches that rely on specialized architectures or complex representations, I2EvDet bridges mainstream object detection with temporal event data via strategic, minimal modifications. Our framework demonstrates how image-based detectors can be efficiently transformed into video-capable models while preserving their core strengths. We first describe our image-like representation of event data then introduce the I2EvDet framework. Finally, we detail the technical implementation and discuss how our approach addresses the unique temporal dependencies in event camera data." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.719, + 0.756 + ], + "angle": 0, + "content": "3.1. Event Representation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.763, + 0.906, + 0.854 + ], + "angle": 0, + "content": "Our adaptation framework begins with transforming EBC data into an image-like representation that is compatible with mainstream object detection models. We employ the Stacked 2D Histogram representation [6] as it provides a straightforward bridge between event data and conventional image processing architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To ensure direct comparability with existing methods, we adopt the frame construction parameters from RVT [9], which allows us to isolate the impact of our architectural" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.167 + ], + "angle": 0, + "content": "choices from data pre-processing. Specifically, we take the stream of events and partition it into a series of consecutive fixed time windows of \\( T_{\\mathrm{frame}} = 50 \\) ms. Each such window corresponds to a single frame. Next, we further subdivide each frame into 10 intervals of \\( T_{\\mathrm{bin}} = 5 \\) ms." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.168, + 0.483, + 0.211 + ], + "angle": 0, + "content": "To construct a frame corresponding to an interval \\([t_0, t_0 + T_{\\mathrm{frame}}]\\), we create an intermediate stacked histogram \\(S\\) from a set of events \\(\\mathcal{E}\\) in that interval:" + }, + { + "type": "equation", + "bbox": [ + 0.098, + 0.221, + 0.476, + 0.256 + ], + "angle": 0, + "content": "\\[\nS(p,t_{i},y,x) = \\sum_{\\{e|t_{e}\\in [t_{0},t_{0} + T_{\\text{frame}})\\}}\\delta_{x}^{x_{e}}\\delta_{y}^{y_{e}}\\delta_{p}^{p_{e}}\\mathbf{1}_{[t_{i},t_{i + 1})}(t_{e})\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.266, + 0.483, + 0.388 + ], + "angle": 0, + "content": "where \\( e = (t_e, p_e, x_e, y_e) \\) represents an individual event with time stamp \\( t_e \\), polarity \\( p_e \\), and spatial coordinates \\( (x_e, y_e) \\); \\( t_i = t_0 + i \\cdot T_{\\mathrm{bin}} \\), \\( i \\in [0, 10) \\) is the histogram bin index; \\( (x, y) \\) are the spatial indices in the output histogram; \\( \\delta_a^b \\) is the Kronecker delta function; and \\( \\mathbf{1} \\) is an indicator function. Once a stacked histogram \\( S \\) of shape \\( (2, 10, H, W) \\) is constructed, we merge the polarity and time dimensions to obtain an image-like 2D frame \\( F \\) of shape \\( (20, H, W) \\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.396, + 0.483, + 0.426 + ], + "angle": 0, + "content": "3.2. I2EvDet: Image-to-Event Detection Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.538 + ], + "angle": 0, + "content": "I2EvDet is a principled framework for adapting image-based object detectors to process temporal data. While specialized architectures typically are developed for event-based object detection, this approach demonstrates that strategic adaptation of mainstream detectors can achieve superior performance with minimal architectural modifications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.644 + ], + "angle": 0, + "content": "The I2EvDet framework applies latent space adaptation principles [12, 14, 15], which enable domain adaptation through minimal changes to pre-trained models. The I2EvDet framework follows a two-stage process, and it is applicable to any object detection model with a clean separation between feature extraction and object detection modules (e.g.,YOLO and DETR families)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.483, + 0.719 + ], + "angle": 0, + "content": "First, we establish a solid foundation by training an object detector on individual EBC frames (subsection 3.1), treating them as conventional images. This creates a detector with powerful spatial representation capabilities optimized for event data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.87 + ], + "angle": 0, + "content": "In the second stage, we transform this image-based detector into a video-capable model. We freeze the pre-trained detector parameters and insert lightweight RNN modules between the feature extractor and detection components. We train the inserted temporal modules on the EBC video object detection task while keeping the original detector frozen. The resulting model captures temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "This two-stage approach retains learned spatial representations of the original image detector while efficiently" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "adding temporal capabilities. In this work, we apply I2EvDet to the RT-DETR model as the primary detector, resulting in the EvRT-DETR model. Additionally, we demonstrate the generalizability of the I2EvDet framework using experiments with YOLOX detectors." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.176, + 0.822, + 0.192 + ], + "angle": 0, + "content": "3.3. Temporal Modeling Considerations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.198, + 0.907, + 0.393 + ], + "angle": 0, + "content": "A key theoretical consideration for this adaptation framework is the choice of temporal memory mechanism. Event cameras present a unique challenge: static objects become \"invisible\" as they generate no events when motionless. Their detection requires having a persistent memory of past events. While transformers have dominated recent vision architectures, their fixed context windows limit temporal memory. We employ RNNs for their theoretically unbounded memory capacity through recurrent state, making them uniquely suited for maintaining object persistence during periods of invisibility in the event stream. This theoretical foundation has informed the EvRT-DETR implementation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.403, + 0.874, + 0.42 + ], + "angle": 0, + "content": "3.4. Technical Implementation of EvRT-DETR" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.425, + 0.906, + 0.516 + ], + "angle": 0, + "content": "The EvRT-DETR architecture implements the I2EvDet framework via strategic, minimal modifications to the RT-DETR model. As shown in Figure 3, we identify key integration points within the pre-trained model's latent space, where temporal information can be effectively incorporated without disrupting spatial feature extraction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.606 + ], + "angle": 0, + "content": "The RT-DETR encoder produces multiscale feature representations \\(\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}\\), where each element corresponds to a transformer token encoding rich spatial information. These features, originating from the last three stages of the backbone \\(\\{S_3,S_4,S_5\\}\\) [30], provide an ideal latent manifold for temporal adaptation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.607, + 0.906, + 0.728 + ], + "angle": 0, + "content": "Following latent space adaptation principles, we introduce specialized temporal processing modules at each scale, implementing three parallel RNN-based memory modules \\(\\{\\mathbf{R}_3,\\mathbf{R}_4,\\mathbf{R}_5\\}\\). This multiscale temporal adaptation design ensures that temporal dynamics are captured at different levels of feature abstraction. Critically, each RNN module interfaces with its corresponding encoder feature map through residual connections, mathematically expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.751, + 0.905, + 0.771 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {i} ^ {t, p r o j} = W _ {i} ^ {d o w n} \\cdot \\mathcal {E} _ {i} ^ {t} \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.773, + 0.905, + 0.799 + ], + "angle": 0, + "content": "\\[\n\\left(\\mathcal {O} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t}\\right) = \\mathbf {R} _ {i} \\left(\\mathcal {E} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t - 1}\\right) \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.661, + 0.801, + 0.905, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {O} _ {i} ^ {t} = W _ {i} ^ {u p} \\cdot \\mathcal {O} _ {i} ^ {t, p r o j} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.662, + 0.822, + 0.905, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {E}} _ {i} ^ {t} = \\mathcal {E} _ {i} ^ {t} + \\alpha_ {i} \\cdot \\mathcal {O} _ {i} ^ {t}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.851, + 0.905, + 0.902 + ], + "angle": 0, + "content": "where \\(W_{i}^{down}\\in \\mathbb{R}^{C_{i}^{RNN}\\times C_{i}}\\) and \\(W_{i}^{up}\\in \\mathbb{R}^{C_{i}\\times C_{i}^{RNN}}\\) are the projection matrices that transform features into and out of the temporal processing space, \\(\\mathcal{M}_i^t\\in \\mathbb{R}^{C_i^{RNN}\\times H_i\\times W_i}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9816" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.131, + 0.092, + 0.148, + 0.227 + ], + "angle": 0, + "content": "B" + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.089, + 0.303, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.089, + 0.442, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.089, + 0.584, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.09, + 0.723, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.09, + 0.865, + 0.152 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.164, + 0.161, + 0.303, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.305, + 0.161, + 0.442, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.446, + 0.161, + 0.584, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.587, + 0.161, + 0.723, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.161, + 0.865, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.238, + 0.907, + 0.308 + ], + "angle": 0, + "content": "Figure 4. Temporal Adaptation Benefits of EvRT-DETR. Visualization of object detection across sequential 1Mpx frames captured by a vehicle-mounted camera that gradually comes to a stop. While the camera is in motion (left frames), both detectors maintain accurate bounding boxes of stationary vehicles. As camera motion decreases (middle frames), event data becomes increasingly sparse. When the camera stops completely (right frames), RT-DETR (top row) loses detection capability, while EvRT-DETR (bottom row) maintains consistent detection by engaging its temporal memory module." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.575 + ], + "angle": 0, + "content": "represents the temporal memory state for the \\(i\\)-th feature scale at time \\(t\\), \\(\\mathcal{O}_i^t\\) denotes the corresponding temporal adaptation output, and \\(\\alpha_{i}\\) is a learnable scaling factor that modulates the influence of temporal information on the spatial representation inspired by the ReZero technique [1]. In our implementation, the input projection matrix \\(W_{i}^{down}\\) is integrated within the RNN architecture, which already handles dimensionality reduction through its internal gating mechanisms. Meanwhile, the output projection \\(W_{i}^{up}\\) is explicitly implemented to ensure proper integration back into the frozen RT-DETR architecture. Thus, the transformed feature representation \\(\\tilde{\\mathcal{E}}_i^t\\) incorporates temporal context while maintaining the integrity of the original spatial representation \\(\\mathcal{E}_i^t\\). In addition, our ablation studies show that \\(C_i^{RNN}\\) can be reduced 4x relative to \\(C_i\\) with minimal performance impact." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.576, + 0.483, + 0.682 + ], + "angle": 0, + "content": "For specific implementation of the RNN modules, we employ the ConvLSTM architecture [27], which effectively balances representational capacity with computational efficiency. This choice was informed by ConvLSTM's demonstrated effectiveness in handling spatio-temporal data in the RVT model [9], allowing our adaptation to employ established temporal modeling techniques." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.694, + 0.431, + 0.71 + ], + "angle": 0, + "content": "4. Evaluation of the EvRT-DETR Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.81 + ], + "angle": 0, + "content": "The EvRT-DETR model is evaluated on two standard EBC object detection benchmarks: Gen1 [4] and 1Mpx [22] (also known as Gen4) datasets. These datasets contain diverse driving scenarios captured by Prophesee's neuromorphic vision sensors, providing an ideal testbed for validating our adaptation approach under real-world conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.819, + 0.292, + 0.835 + ], + "angle": 0, + "content": "4.1. Experimental Setting" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.841, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Datasets. We evaluate on two standard Propheesee automotive datasets: Gen1 [4] and 1Mpx [22]. Gen1 contains \\(39+\\) hours of recordings at \\(304 \\times 240\\) resolution with manual annotations (1-4 Hz) for cars and pedestrians, presenting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.907, + 0.456 + ], + "angle": 0, + "content": "challenges with sparse labels and stationary objects. 1Mpx offers \\(14+\\) hours at higher resolution \\((1280 \\times 720)\\) with denser automatic annotations \\((60\\mathrm{Hz})\\) for cars, pedestrians, and two-wheelers. Both datasets are organized as 60-second clips with standard train/test splits. 1Mpx's higher resolution and frame rate enable evaluation on more complex scenes, while Gen1's annotation pattern tests robustness to sparse temporal supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.474, + 0.906, + 0.58 + ], + "angle": 0, + "content": "Event Representations. We adopt the well-established stacked 2D Histogram approach from RVT protocols [9], constructing event frames from each \\(50\\mathrm{ms}\\) time window. Each frame is subdivided into 10 equally time-spaced \\(5\\mathrm{ms}\\) intervals, creating a representation with sufficient temporal resolution for our adaptation framework while remaining compatible with standard image processing architectures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.58, + 0.906, + 0.686 + ], + "angle": 0, + "content": "For the Gen1 dataset (original size (240, 304)), we pad frames to (256, 320) to ensure divisibility by 32 as required by the feature pyramid architecture of RT-DETR. For 1Mpx, we downsize from (720, 1280) to (360, 640) using bilinear interpolation and pad to (384, 640). This maintains consistency with prior work while allowing for fair computational comparison." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Two-stage Training Process. We train EvRT-DETR in two stages: (1) training the RT-DETR detector on individual event frames using standard practices (Adam optimizer; EMA of weights) then (2) freezing these parameters and training only the lightweight RNN modules on both random and sequential clips to develop temporal memory. The first stage establishes robust spatial feature extraction, while the second adds temporal processing with minimal parameter overhead. For temporal training, we maintain RNN memory continuity across sequential clips while resetting for random clips using different clip lengths for Gen1 (21 frames) and 1Mpx (10 frames). Full implementation details are provided in the supplementary material." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9817" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.169, + 0.088, + 0.83, + 0.369 + ], + "angle": 0, + "content": "
ModelGen11Mpx/Gen4Params (M)
mAP (%)Runtime (ms)mAP (%)Runtime (ms)
AsyNet [18]14.5---11.4
AEGNN [26]16.3---20.0
Spiking DenseNet [3]18.9---8.2
Events-RetinaNet [22]34.0-18.0-32.8
RED [22]40.016.743.039.324.1
AEC [19]47.010.6 (3.9)48.437.6 (13.8)46.5
RVT-B [9]47.210.247.411.918.5
GET-T [20]47.916.848.418.221.9
S5-ViT-B [34]47.48.247.29.617.5
ASTMNet [16]46.735.648.372.3>100
SAST-CB [21]48.2-48.757.5 (19.7)18.9
ERGO-12 [33]50.469.940.6100.059.6
RT-DETR-T (ours)46.06.444.18.620.1
RT-DETR-B (ours)47.610.545.214.942.8
EvRT-DETR-T (ours)52.38.449.912.534.4
EvRT-DETR-B (ours)52.712.750.118.857.1
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.379, + 0.907, + 0.451 + ], + "angle": 0, + "content": "Table 1. Performance Comparison on Event-based Camera Datasets. Our approach achieves state-of-the-art results on both Gen1 and 1Mpx benchmarks with significant improvements over previous methods. The first stage of the framework (RT-DETR-T/B) already achieves competitive performance using only image-based detection, while the proposed I2EvDet-adapted models (EvRT-DETR-T/B) surpass all existing approaches. Runtime measurements are reported in milliseconds on an NVIDIA T4 GPU. Values in parentheses represent original reported times on different hardware with T4-equivalent times calculated using FLOPS/throughput conversion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.476, + 0.483, + 0.673 + ], + "angle": 0, + "content": "Evaluation Protocol. We evaluate EvRT-DETR using the standard COCO (Common Objects in Context) Mean Average Precision (mAP) measure [17] that allows direct comparison with existing methods. For consistency with previous works, we use the EBC-specific implementation of the COCO metrics provided by Propheesee's Automotive Dataset Toolbox [22, 23]. All inference times reported in Table 2 are standardized to NVIDIA T4 GPU performance for fair comparison. Inference times include model forward pass and post-processing but exclude event preprocessing, following standard reporting practices. When prior work reported only preprocessing-inclusive times, we use their reported values directly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.295, + 0.699 + ], + "angle": 0, + "content": "4.2. Performance Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Table 1 presents a comprehensive comparison of our EvRT-DETR models against state-of-the-art approaches on the Gen1 and 1Mpx benchmarks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.886 + ], + "angle": 0, + "content": "First, our baseline adaptation using two RT-DETR variants, RT-DETR-T (Tiny, with ResNet-18 backbone) and RT-DETR-B (Base, with ResNet-50 backbone), and trained on simple event frames achieves competitive performance (46.0% and 47.6% mAP on Gen1; 44.1% and 45.2% mAP on 1Mpx) without requiring specialized architectures or complex event representations. This demonstrates that modern mainstream object detectors can be effectively adapted to event data through appropriate training." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.901 + ], + "angle": 0, + "content": "More importantly, our complete EvRT-DETR models" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.476, + 0.907, + 0.581 + ], + "angle": 0, + "content": "with temporal adaptation significantly outperform all existing methods, achieving new state-of-the-art results on both benchmarks. EvRT-DETR-B reaches \\(52.7\\%\\) mAP on Gen1 \\((+2.3\\%\\) over the previous best) and \\(50.1\\%\\) mAP on 1Mpx \\((+1.4\\%\\) over the previous best). Even our lighter EvRT-DETR-T model surpasses all existing approaches while maintaining faster inference times." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.686 + ], + "angle": 0, + "content": "Figure 4 shows how the temporal adaptation module provides crucial capabilities for handling stationary objects - a fundamental challenge for event cameras. When objects are in motion, both RT-DETR and EvRT-DETR perform well, yet, when motion stops and event generation becomes minimal, EvRT-DETRs temporal memory enables consistent detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.688, + 0.909, + 0.78 + ], + "angle": 0, + "content": "These performance improvements come with minimal computational overhead. While we focus primarily on detection accuracy in this work, the runtime measurements in Table 1 demonstrate that our approach maintains competitive inference speeds despite adding temporal processing capabilities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.771, + 0.805 + ], + "angle": 0, + "content": "4.3. Framework Generalizability" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To demonstrate that our results are not specific to the RT-DETR architecture, we apply the I2EvDet framework to YOLOX [7]. YOLOX is a CNN-based detector that contrasts with RT-DETR's transformer architecture and has been used in other EBC applications [9, 34]. Table 2 shows that our temporal adaptation approach consistently" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "9818" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.132, + 0.089, + 0.441, + 0.171 + ], + "angle": 0, + "content": "
DetectormAP (%)
OriginalI2EvDetImprovement
RT-DETR-T46.052.3+6.3
RT-DETR-B47.652.7+5.1
YOLOX-T36.042.4+6.4
YOLOX-X43.447.8+4.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.237 + ], + "angle": 0, + "content": "Table 2. Generalizability of I2EvDet Framework Across Different Object Detection Architectures on the Gen1 Dataset. Our temporal adaptation consistently improves performance regardless of the base detector architecture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.262, + 0.484, + 0.369 + ], + "angle": 0, + "content": "improves performance across different baseline architectures on the Gen1 dataset. The substantial improvements (4.4-6.4 mAP) across both RT-DETR and YOLOX variants validate that I2EvDet represents a general framework for adapting image-based detectors to temporal domains rather than an architecture-specific optimization. Additional details are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.377, + 0.424, + 0.394 + ], + "angle": 0, + "content": "4.4. Analysis of Adaptation Design Choices" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.49 + ], + "angle": 0, + "content": "We systematically analyze key design choices in our temporal adaptation module to better understand the factors contributing to EvRT-DETR's performance. This analysis provides insights into efficient adaptation strategies for event-based object detection. Additional ablation studies are provided in the supplementary material." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.501, + 0.48, + 0.584 + ], + "angle": 0, + "content": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)
RT-DETR-B (ours)47.676.349.5
EvRT-DETR-B (0,1,1)51.080.754.0
EvRT-DETR-B (1,0,1)52.281.355.1
EvRT-DETR-B (1,1,0)52.481.755.3
EvRT-DETR-B (1,1,1)52.782.056.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.664 + ], + "angle": 0, + "content": "Table 3. Effect of Temporal Module Placement. Performance on the Gen1 dataset with ConvLSTM modules at different feature scales. The array notation \\((x,y,z)\\) indicates presence (1) or absence (0) of temporal modules at scales \\(\\mathcal{E}_3\\), \\(\\mathcal{E}_4\\), and \\(\\mathcal{E}_5\\), respectively. Low-level features benefit most from temporal adaptation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.686, + 0.365, + 0.701 + ], + "angle": 0, + "content": "4.4.1. Multiscale Temporal Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.796 + ], + "angle": 0, + "content": "A crucial aspect of our approach is the insertion of temporal modules at multiple scales in the feature hierarchy. Table 3 shows performance when selectively removing ConvLSTM cells from specific feature levels with the array notation \\((x,y,z)\\) indicating presence (1) or absence (0) at each scale \\(\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The results reveal a clear pattern: the lowest-level features \\((\\mathcal{E}_3)\\) benefit most significantly from temporal adaptation as performance drops substantially when this module is removed. The impact diminishes at higher feature levels, suggesting that temporal information is most valuable for fine-grained, lower-level features that capture detailed motion patterns. This insight can guide more efficient adapta" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.123 + ], + "angle": 0, + "content": "tion designs, wherein computational resources are directed to where they provide maximum benefit." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.137, + 0.903, + 0.196 + ], + "angle": 0, + "content": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)\\( N_T \\) (M)
EvRT-DETR-B (M=64)52.181.354.92.3
EvRT-DETR-B (M=128)52.581.755.65.4
EvRT-DETR-B (M=256)52.782.056.014.4
EvRT-DETR-B (M=512)52.981.956.042.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.207, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Table 4. Impact of Temporal Module Capacity. Performance on Gen1 dataset with varying ConvLSTM hidden dimensions (M). The rightmost column shows the number of trainable parameters (in millions) in the temporal adaptation module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.292, + 0.808, + 0.307 + ], + "angle": 0, + "content": "4.4.2. Parameter Efficiency in Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.312, + 0.906, + 0.387 + ], + "angle": 0, + "content": "Inspired by parameter-efficient techniques such as LoRA (Low-Rank Adaptation) [15], we investigate the minimal temporal module capacity needed for effective adaptation. Table 4 shows EvRT-DETR-B performance with varying hidden dimensions in the ConvLSTM modules." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Remarkably, reducing hidden dimensions to 64 features (a \\(4\\mathrm{x}\\) reduction from our baseline) results in only modest performance degradation \\((-0.6\\% \\mathrm{mAP})\\) while requiring just \\(2.3\\mathrm{M}\\) additional parameters - merely \\(5.4\\%\\) of the base RT-DETR-B model's \\(42.8\\mathrm{M}\\) parameters. This configuration closely resembles LoRA-style adaptation, achieving substantial performance gains with minimal parameter overhead. Even our most parameter-efficient adaptation \\((\\mathbf{M} = 64)\\) outperforms all previous state-of-the-art approaches, demonstrating that effective temporal adaptation does not require extensive architectural modifications. Increasing the hidden dimension to 512 features \\((42.9\\mathrm{M}\\) parameters, nearly doubling the model size) yields only a marginal \\(0.2\\% \\mathrm{mAP}\\) improvement, indicating diminishing returns." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.64, + 0.648 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work introduces I2EvDet, a framework that efficiently transforms image-based detectors into video-capable models using minimal modifications. Applied to event-based cameras, our approach demonstrates that mainstream object detection architectures can be effectively adapted to temporal domains without specialized redesign. RT-DETR trained on simple event representations achieves performance comparable to specialized methods, while our I2EvDet-adapted model EvRT-DETR achieves state-of-the-art results on Gen1 \\((+2.3\\mathrm{mAP})\\) and 1Mpx \\((+1.4\\mathrm{mAP})\\) benchmarks. Ablation studies provide insights about optimal temporal module design and parameter efficiency. These results suggest that bridging mainstream computer vision and specialized domains can be achieved through targeted adaptation rather than complete architectural redesign, potentially accelerating progress across diverse temporal visual domains." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "9819" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.253 + ], + "angle": 0, + "content": "The authors extend their sincere gratitude to Yuewei Lin (Brookhaven Lab) for his valuable feedback and suggestions. This work presented in this paper was funded by the U.S. Department of Energy (DOE), National Nuclear Security Administration, Office of Defense Nuclear Nonproliferation Research and Development. The manuscript has been authored by Brookhaven National Laboratory managed by Brookhaven Science Associates, LLC under the Contract No. DE-SC0012704 with the U.S. Department of Energy." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.264, + 0.188, + 0.279 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.288, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[1] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Gary Cottrell, and Julian J. McAuley. ReZero is all you need: fast convergence at large depth. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, UAI 2021, Virtual Event, 27-30 July 2021, pages 1352-1361. AUAI Press, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.372, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213-229. Springer, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.457, + 0.484, + 0.523 + ], + "angle": 0, + "content": "[3] Loïc Cordone, Benoit Miramond, and Philippe Thiérison. Object detection with spiking neural networks on automotive event data. In International Joint Conference on Neural Networks, IJCNN 2022, Padua, Italy, July 18-23, 2022, pages 1-8. IEEE, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.526, + 0.484, + 0.579 + ], + "angle": 0, + "content": "[4] Pierre de Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. CoRR, abs/2001.08499, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.582, + 0.484, + 0.663 + ], + "angle": 0, + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.665, + 0.484, + 0.746 + ], + "angle": 0, + "content": "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jorg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.749, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[7] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.79, + 0.484, + 0.831 + ], + "angle": 0, + "content": "[8] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. CoRR, abs/2211.12324, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.833, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13884-13893. IEEE, 2023. 3, 4, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.288, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[10] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 580-587. IEEE Computer Society, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.906, + 0.232 + ], + "angle": 0, + "content": "[11] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. Trans. Mach. Learn. Res., 2024, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.907, + 0.317 + ], + "angle": 0, + "content": "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.322, + 0.906, + 0.39 + ], + "angle": 0, + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.392, + 0.906, + 0.488 + ], + "angle": 0, + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 2790-2799. PMLR, 2019. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.492, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 4, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[16] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Trans. Image Process., 31:2975-2987, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.634, + 0.906, + 0.73 + ], + "angle": 0, + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.733, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[18] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII, pages 415-431. Springer, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[19] Yansong Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "9820" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "DC, USA, February 7-14, 2023, pages 2056-2064. AAAI Press, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[20] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. GET: group event transformer for event-based vision. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 6015-6025. IEEE, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[21] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16794-16804. IEEE, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[22] Etienne Perot, Pierre de Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.363, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[23] Propheesee. Propheesee's automotive dataset toolbox. https://github.com/propheesee-ai/propheesee-automotive-dataset-toolbox, 2024.7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[24] Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 779-788. IEEE Computer Society, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.589 + ], + "angle": 0, + "content": "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 91-99, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[26] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. AEGNN: asynchronous event-based graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12361-12371. IEEE, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.662, + 0.482, + 0.759 + ], + "angle": 0, + "content": "[27] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 802-810, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[29] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[30] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on Real-time object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16965-16974. IEEE, 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[31] Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. Convolution meets lora: Parameter efficient finetuning for segment anything model. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.29, + 0.905, + 0.359 + ], + "angle": 0, + "content": "[32] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.442 + ], + "angle": 0, + "content": "[33] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 12800-12810. IEEE, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.446, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[34] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5819-5828. IEEE, 2024. 3, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.513 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.514, + 0.956 + ], + "angle": 0, + "content": "9821" + } + ] +] \ No newline at end of file diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_origin.pdf b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d1ce32c013b4cfd1ba3951f3230194b1de61fa61 --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/b434c8f2-0109-4d2a-8b19-39a0325126c5_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:152299f0cc84c36ed0598b3ca427207a210078bcb12e5bb86d46534abe3796ed +size 3810993 diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/full.md b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/full.md new file mode 100644 index 0000000000000000000000000000000000000000..61b46652cf1f95ffa4c12e92e49fa1d1f14531ac --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/full.md @@ -0,0 +1,301 @@ +# EvRT-DETR: Latent Space Adaptation of Image Detectors for Event-based Vision + +Dmitrii Torbunov1, Yihui Ren1, Animesh Ghose1, Odera Dim1, Yonggang Cui1 +1Brookhaven National Laboratory, Upton, NY, USA + +dtorbunov@bnl.gov, yren@bnl.gov, aghose@bnl.gov, dodera@bnl.gov, ycui@bnl.gov + +# Abstract + +Event-based cameras (EBCs) have emerged as a bioinspired alternative to traditional cameras, offering advantages in power efficiency, temporal resolution, and high dynamic range. However, development of image analysis methods for EBCs is challenging due to the sparse and asynchronous nature of the data. This work addresses the problem of object detection for EBC cameras. The current approaches to EBC object detection focus on constructing complex data representations and rely on specialized architectures. We introduce I2EvDet (Image-to-Event Detection), a novel adaptation framework that bridges mainstream object detection with temporal event data processing. First, we demonstrate that a Real-Time DEetection Transformer, or RT-DETR, a state-of-the-art natural image detector, trained on a simple image-like representation of the EBC data achieves performance comparable to specialized EBC methods. Next, as part of our framework, we develop an efficient adaptation technique that transforms image-based detectors into event-based detection models by modifying their frozen latent representation space via minimal architectural additions. The resulting EvRT-DETR model reaches state-of-the-art performance on the standard benchmark datasets Gen1 (mAP +2.3) and 1Mpx/Gen4 (mAP +1.4). These results demonstrate a fundamentally new approach to EBC object detection through principled adaptation of mainstream architectures, offering an efficient alternative with potential applications to other temporal visual domains. The code is available at: https://github.com/realtime-intelligence/evrt-detr. + +# 1. Introduction + +Event-based cameras (EBCs) present a biologically inspired alternative to traditional frame-based cameras. Unlike conventional cameras that capture frames at predetermined in + +![](images/1f4beb4a8a1e624ea438d69e0ef38ecb723a8f5f86ba498b6069617facc1e5df.jpg) +Figure 1. Object Detection Performance vs. Inference Time. A summary of object detection performance (COCO mAP) versus inference time (ms) of various models on the 1Mpx automotive dataset. Circle size is proportional to the number of model parameters. Inference times are reported for NVIDIA T4 GPU. Our EvRT-DETR models achieve state-of-the-art accuracy while maintaining competitive inference speeds. + +tervals, EBC pixels operate asynchronously, generating data only when brightness changes exceed a threshold. This results in sparse, asynchronous data streams where each event is a tuple containing pixel location, timestamp, and polarity (whether brightness increased or decreased), as illustrated in Figure 2. This approach offers remarkable advantages: low power consumption (as low as $10\mathrm{mW}$ ), reduced data transfer rates, exceptional temporal resolution (on the order of $\mu s$ ), and high dynamic range ( $>100$ dB). These properties have led to widespread adoption in autonomous driving, robotics, and wearable electronics [6]. + +These unique characteristics of EBC data - sparsity, asynchronicity, and temporal nature - create fundamental challenges when applying traditional computer vision techniques. Early attempts to directly apply conventional + +![](images/cdf3dd14f6e384cce330ebd2cff576c77f5ead735509b710f19ae1d2ab2b7679.jpg) +Figure 2. Event-based Camera Output Visualization. Sample frames from the 1Mpx dataset showing event polarity. Red indicates positive events, and blue denotes negative events. + +image-based object detectors to simple event representations yielded poor performance [22], highlighting the gap between these domains. This led to two predominant research directions: (1) constructing sophisticated event representations that transform EBC data into formats compatible with existing detectors (e.g., [33]) or (2) designing specialized architectures tailored to handle the temporal nature of event data (e.g., [20]). A common assumption in these approaches is that EBC data processing requires either specialized representations or architectures due to the fundamental differences between event- and frame-based vision, leading to development paths that often diverge from mainstream computer vision research. + +Meanwhile, mainstream object detection has seen remarkable progress, evolving from convolutional neural network (CNN)-based architectures, e.g.,YOLO [24] and Faster R-CNN [25], to transformer-based models like DETR (DEtection Transformer) [2]. Recent advances have culminated in highly efficient and accurate models, such as the Real-Time DETection TRansformer (RT-DETR) [30] that achieves exceptional performance while maintaining real-time inference speeds. Though primarily optimized for processing individual images, these state-of-the-art detectors offer powerful feature extraction capabilities that potentially could be leveraged for temporal data. + +This evolution in mainstream detection creates an opportunity to bridge the gap between conventional and event-based vision. Rather than assuming fundamental incompatibility between these domains, we explore a complementary direction: adapting state-of-the-art frame-based detectors to process event data effectively with minimal modifications. In this work, we investigate how RT-DETR can be strategically adapted to the event-based domain, demonstrating that modern object detection architectures can successfully work on event data and opening a new research direction. + +Our approach proceeds in two stages to systematically adapt RT-DETR to event data. First, we train RT-DETR directly on simple image-like representations of EBC data using standard training procedures without special modifications. This straightforward approach yields surprising results - even when processing only fixed $50\mathrm{ms}$ time frames, the detector achieves performance comparable to special + +ized EBC-specific methods explicitly designed for video object detection on EBC data. This is particularly unexpected as these specialized methods deliberately exploit the temporal nature of EBC data and process information from the distant past, while our approach performs detection using only a single fixed time window (frame) of $50~\mathrm{ms}$ , similar to traditional frame-based cameras. + +Building on this foundation, we introduce I2EvDet (Image-to-Event Detection), a technique that enables straightforward transformation of image-based object detectors into video-capable models. This approach leverages latent space adaptation principles [14, 15] where we freeze the pre-trained RT-DETR model and insert a lightweight recurrent neural network (RNN) module within the encoder's latent representation space. This design allows the model to capture temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead. The resulting adaptation effectively leverages information across multiple time windows without requiring a specialized video detection architecture. + +The combined approach, called EvRT-DETR, achieves state-of-the-art performance on standard EBC benchmarks, outperforming specialized methods on both Gen1 [4] (mAP +2.3) and 1Mpx [22] (mAP +1.4) datasets. Importantly, the results are achieved using only standard modules from natural image and video analysis without requiring domain-specific modifications. This demonstrates not only the effectiveness of our approach for EBC data, but also suggests a broader paradigm for adapting image detectors to temporal domains. As shown in Figure 1, our model achieves superior accuracy while maintaining competitive inference speeds, making it practical for real-world applications. + +Contributions of this work: + +- We show that when trained on simple image-like representations of EBC data using standard procedures, mainstream object detectors like RT-DETR can achieve performance comparable to specialized EBC methods, offering a complementary approach to domain-specific architectures. +- We propose I2EvDet, an efficient adaptation technique that transforms image-based detectors into video-capable models through minimal architectural additions to the frozen pre-trained model and with potential applications to various temporal visual domains. +- We validate our approach in the challenging domain of event-based cameras, where EvRT-DETR achieves state-of-the-art performance on the standard Gen1 (mAP +2.3) and 1Mpx (mAP +1.4) EBC benchmarks while maintaining competitive inference speeds. + +# 2. Related Work + +Research at the intersection of event-based cameras and object detection spans multiple domains in computer vision. Current approaches primarily focus on two directions: designing efficient representations for the unique properties of event data and developing specialized neural architectures to process these representations. However, despite its success in other domains, a third direction – adapting mainstream computer vision models to event data – remains underexplored. This section reviews relevant work across these areas, highlighting how our approach bridges the gap between specialized event-based methods and mainstream object detection advances. + +# 2.1. EBC Data Representations + +An event camera produces a stream of event data in the form $(t, p, x, y)$ , where $t$ is the timestamp of an event, $p$ is the polarity (positive or negative) of the brightness change, and $(x, y)$ is the location of the event pixel. To simplify the analysis, the event data often are transformed into alternative representations more suitable for conventional image analysis algorithms. + +The simplest image-like representations of event camera data are Event Frames or 2D Histograms [6], where events within fixed time windows are accumulated into two-dimensional frames of shape $(2, H, W)$ with polarity as the first dimension and spatial coordinates as the remaining dimensions. A natural evolution is the Stacked 2D Histogram, which further partitions each frame into $T$ time intervals, creating a representation of shape $(2, T, H, W)$ that typically is reshaped to $(2T, H, W)$ and treated as a natural image. While these representations are straightforward, directly applying existing computer vision algorithms to them has yielded poor performance in the past [22]. + +To improve performance, other representations are being explored. For example, a Time Surface (TS) representation [6, 33] is an image-like representation where pixel values encode the time elapsed since the last event in a given pixel. Thus, unlike the simplest fixed time window representations, TS potentially can encode arbitrarily distant past. Empirically, it has been shown that using the TS representations offers significantly better performance on the object detection tasks compared to the 2D Histogram representations [22]. + +Many other EBC data representations also have been proposed. The recent ERGO-12 (Event Representation through Gromov-Wasserstein Optimization) [33] work has developed efficient image-like representations, achieving state-of-the-art object detection performance on the Gen1 dataset. Meanwhile, other approaches reformulate object detection as a three-dimensional (3D) problem by treating events as points in spatial-temporal space or discretizing them into voxel grids [6]. While computationally heavier, + +these representations can preserve exact event timing, potentially enabling better handling of overlapping objects and complex motions. + +# 2.2. EBC Object Detection Architectures + +Once the data representation is selected, one can start to develop object detection algorithms. There are several families of object detection methods designed for particular data representations. + +The most straightforward approach applies conventional image detection methods directly to 2D Histogram frames, but this yields relatively poor performance [22]. More advanced methods, such as Recurrent Vision Transformer (RVT) [9] and State Space Vision Transformer (S5-ViT-B) [34], extend this approach by incorporating temporal memory to capture sequential information, significantly improving detection quality compared to frame-by-frame processing. + +The most recent family of methods attempts to use more efficient event representations to further improve object detection performance. For example, 12-channel ERGO-12 [33] constructs an optimized image-like event representation and shows that the standard object detection methods perform exceptionally well on it. Alternatively, Asynchronous Spatio-Temporal Memory Network (ASTMNet) [16] and Adaptive Event Conversion (AEC) [19] attempt to combine better event representation while using novel neural architectures to expand state-of-the-art results. + +Other methods have been developed that treat EBC object detection as a 3D detection problem [8, 22, 26], exploit benefits of neuromorphic architectures [3], or explore various hybrid approaches [20, 21]. + +# 2.3. Transformer-based Object Detection + +Object detection has been revolutionized by transformer architectures in recent years. DETR [2] was the first widely successful transformer-based [28] model for object detection, offering a simple and elegant alternative to traditional CNN-based architectures [10, 24]. + +DETR follows a hybrid CNN-Transformer design with a backbone-encoder-decoder structure. It uses a CNN backbone (typically ResNet [13]) for feature extraction, followed by a Transformer encoder that processes these features into tokens and captures their correlations. The Transformer decoder then identifies objects through cross-attention to the encoder outputs. + +Despite its innovative approach, the original DETR suffered from limitations, including training instability [29], poor performance on small objects [2], and high computational costs during inference [32]. Subsequent research has addressed these issues through architectural improvements such as Feature Pyramid Networks (FPN), specialized decoder tokens, and deformable attention mechanisms [2, 32]. + +![](images/bf19cc5e0a7f5a6e87ddbbf7189f387b4035aee94ea1d2429bad0b5ee0590aab.jpg) +Figure 3. Overview of the I2EvDet Framework Applied to RT-DETR. This approach adapts a pre-trained RT-DETR model to process temporal event data using minimal architectural modifications. The backbone and encoder components remain frozen during temporal adaptation, while trainable temporal modules are strategically inserted at multiple feature scales. Left: Event camera frames are processed through the frozen RT-DETR backbone and encoder. Center: Our temporal adaptation modules operate on multiscale features through residual connections (red pathways), enabling information flow across time steps while preserving spatial representation integrity. Right: Feature fusion and decoder stages produce final detections. The detailed view of the Temporal Adaptation Module (bottom right) shows the projection operations $(W_{i}^{down}, W_{i}^{up})$ and recurrent memory module $(\mathbf{R}_{i})$ that enable efficient temporal modeling with minimal additional parameters. + +These incremental advances have culminated in the RT-DETR [30], which achieves exceptional performance across all object scales with stable training and significantly faster inference times. RT-DETR represents the current state-of-the-art in efficient transformer-based detection and serves as an ideal foundation for adaptation to specialized domains, such as event-based vision. + +# 2.4. Model Adaptation Techniques + +While specialized architectures dominate event-based vision, adapting mainstream architectures offers compelling advantages. Parameter-efficient methods [12, 14, 15] modify pre-trained models with minimal changes by adding compact modules to frozen backbones. These approaches have proven effective across language [5], vision [31], and multimodal systems [11], but they remain underexplored for event-based vision despite their efficiency and transfer learning benefits. + +Given the success of adaptation techniques in other domains and the power of transformer-based detection models, there is a clear opportunity to bridge these approaches for event-based vision. Our work explores this intersection, demonstrating that mainstream object detectors can be effectively adapted to EBC data with minimal modifications rather than requiring specialized architectures. + +# 3. Methods + +This section presents I2EvDet, our adaptation framework for event-based object detection. Unlike approaches that rely on specialized architectures or complex representations, I2EvDet bridges mainstream object detection with temporal event data via strategic, minimal modifications. Our framework demonstrates how image-based detectors can be efficiently transformed into video-capable models while preserving their core strengths. We first describe our image-like representation of event data then introduce the I2EvDet framework. Finally, we detail the technical implementation and discuss how our approach addresses the unique temporal dependencies in event camera data. + +# 3.1. Event Representation + +Our adaptation framework begins with transforming EBC data into an image-like representation that is compatible with mainstream object detection models. We employ the Stacked 2D Histogram representation [6] as it provides a straightforward bridge between event data and conventional image processing architectures. + +To ensure direct comparability with existing methods, we adopt the frame construction parameters from RVT [9], which allows us to isolate the impact of our architectural + +choices from data pre-processing. Specifically, we take the stream of events and partition it into a series of consecutive fixed time windows of $T_{\mathrm{frame}} = 50$ ms. Each such window corresponds to a single frame. Next, we further subdivide each frame into 10 intervals of $T_{\mathrm{bin}} = 5$ ms. + +To construct a frame corresponding to an interval $[t_0, t_0 + T_{\mathrm{frame}}]$ , we create an intermediate stacked histogram $S$ from a set of events $\mathcal{E}$ in that interval: + +$$ +S(p,t_{i},y,x) = \sum_{\{e|t_{e}\in [t_{0},t_{0} + T_{\text{frame}})\}}\delta_{x}^{x_{e}}\delta_{y}^{y_{e}}\delta_{p}^{p_{e}}\mathbf{1}_{[t_{i},t_{i + 1})}(t_{e}) +$$ + +where $e = (t_e, p_e, x_e, y_e)$ represents an individual event with time stamp $t_e$ , polarity $p_e$ , and spatial coordinates $(x_e, y_e)$ ; $t_i = t_0 + i \cdot T_{\mathrm{bin}}$ , $i \in [0, 10)$ is the histogram bin index; $(x, y)$ are the spatial indices in the output histogram; $\delta_a^b$ is the Kronecker delta function; and $\mathbf{1}$ is an indicator function. Once a stacked histogram $S$ of shape $(2, 10, H, W)$ is constructed, we merge the polarity and time dimensions to obtain an image-like 2D frame $F$ of shape $(20, H, W)$ . + +# 3.2. I2EvDet: Image-to-Event Detection Framework + +I2EvDet is a principled framework for adapting image-based object detectors to process temporal data. While specialized architectures typically are developed for event-based object detection, this approach demonstrates that strategic adaptation of mainstream detectors can achieve superior performance with minimal architectural modifications. + +The I2EvDet framework applies latent space adaptation principles [12, 14, 15], which enable domain adaptation through minimal changes to pre-trained models. The I2EvDet framework follows a two-stage process, and it is applicable to any object detection model with a clean separation between feature extraction and object detection modules (e.g.,YOLO and DETR families). + +First, we establish a solid foundation by training an object detector on individual EBC frames (subsection 3.1), treating them as conventional images. This creates a detector with powerful spatial representation capabilities optimized for event data. + +In the second stage, we transform this image-based detector into a video-capable model. We freeze the pre-trained detector parameters and insert lightweight RNN modules between the feature extractor and detection components. We train the inserted temporal modules on the EBC video object detection task while keeping the original detector frozen. The resulting model captures temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead. + +This two-stage approach retains learned spatial representations of the original image detector while efficiently + +adding temporal capabilities. In this work, we apply I2EvDet to the RT-DETR model as the primary detector, resulting in the EvRT-DETR model. Additionally, we demonstrate the generalizability of the I2EvDet framework using experiments with YOLOX detectors. + +# 3.3. Temporal Modeling Considerations + +A key theoretical consideration for this adaptation framework is the choice of temporal memory mechanism. Event cameras present a unique challenge: static objects become "invisible" as they generate no events when motionless. Their detection requires having a persistent memory of past events. While transformers have dominated recent vision architectures, their fixed context windows limit temporal memory. We employ RNNs for their theoretically unbounded memory capacity through recurrent state, making them uniquely suited for maintaining object persistence during periods of invisibility in the event stream. This theoretical foundation has informed the EvRT-DETR implementation. + +# 3.4. Technical Implementation of EvRT-DETR + +The EvRT-DETR architecture implements the I2EvDet framework via strategic, minimal modifications to the RT-DETR model. As shown in Figure 3, we identify key integration points within the pre-trained model's latent space, where temporal information can be effectively incorporated without disrupting spatial feature extraction. + +The RT-DETR encoder produces multiscale feature representations $\{\mathcal{E}_3,\mathcal{E}_4,\mathcal{E}_5\}$ , where each element corresponds to a transformer token encoding rich spatial information. These features, originating from the last three stages of the backbone $\{S_3,S_4,S_5\}$ [30], provide an ideal latent manifold for temporal adaptation. + +Following latent space adaptation principles, we introduce specialized temporal processing modules at each scale, implementing three parallel RNN-based memory modules $\{\mathbf{R}_3,\mathbf{R}_4,\mathbf{R}_5\}$ . This multiscale temporal adaptation design ensures that temporal dynamics are captured at different levels of feature abstraction. Critically, each RNN module interfaces with its corresponding encoder feature map through residual connections, mathematically expressed as: + +$$ +\mathcal {E} _ {i} ^ {t, p r o j} = W _ {i} ^ {d o w n} \cdot \mathcal {E} _ {i} ^ {t} \tag {1} +$$ + +$$ +\left(\mathcal {O} _ {i} ^ {t, p r o j}, \mathcal {M} _ {i} ^ {t}\right) = \mathbf {R} _ {i} \left(\mathcal {E} _ {i} ^ {t, p r o j}, \mathcal {M} _ {i} ^ {t - 1}\right) \tag {2} +$$ + +$$ +\mathcal {O} _ {i} ^ {t} = W _ {i} ^ {u p} \cdot \mathcal {O} _ {i} ^ {t, p r o j} \tag {3} +$$ + +$$ +\tilde {\mathcal {E}} _ {i} ^ {t} = \mathcal {E} _ {i} ^ {t} + \alpha_ {i} \cdot \mathcal {O} _ {i} ^ {t}, \tag {4} +$$ + +where $W_{i}^{down}\in \mathbb{R}^{C_{i}^{RNN}\times C_{i}}$ and $W_{i}^{up}\in \mathbb{R}^{C_{i}\times C_{i}^{RNN}}$ are the projection matrices that transform features into and out of the temporal processing space, $\mathcal{M}_i^t\in \mathbb{R}^{C_i^{RNN}\times H_i\times W_i}$ + +![](images/a25c8b8d2bafef5fdaf3a86cf3a6f7f6bacec7cb889372b5dc816088a99d152b.jpg) +B + +![](images/c8deaa347e462b8fe6bc41dfc358f059ca0987cae5cacba0e9138690e3969c0f.jpg) + +![](images/2854ed932a2c056a16761c76c64e74bbab82b448b302eb8e0b90e8ba3f6b6c8d.jpg) + +![](images/4c246c15051eb71ba1eed325546ed123c0ca0a95cb047d22f8086cadca6dccc2.jpg) + +![](images/bd45114d2465646c2d6672a47b8eba8b576245a1fb8a5c6b303bb00da5fec233.jpg) + +![](images/d4f8d120c7db1903aa90174a3b32261c075bf5246981b7b8292cd7168ffe380d.jpg) +Figure 4. Temporal Adaptation Benefits of EvRT-DETR. Visualization of object detection across sequential 1Mpx frames captured by a vehicle-mounted camera that gradually comes to a stop. While the camera is in motion (left frames), both detectors maintain accurate bounding boxes of stationary vehicles. As camera motion decreases (middle frames), event data becomes increasingly sparse. When the camera stops completely (right frames), RT-DETR (top row) loses detection capability, while EvRT-DETR (bottom row) maintains consistent detection by engaging its temporal memory module. + +![](images/615b91d1f381ebb892b4ae92e06f6450bfb463353537c761ccca95655bb13157.jpg) + +![](images/d428df93af6a95718f09b10ede61dc34ee458a16916893b5893782e53d25b028.jpg) + +![](images/05610d3f100e0d20f8b1dc2271f7d9c7f62f6dde6cc744b2918332a766334dd0.jpg) + +![](images/d11a37e3cf9f1dade38af3dc95a6d085faf0e55a41376b33f5c7a47aa2f3c300.jpg) + +represents the temporal memory state for the $i$ -th feature scale at time $t$ , $\mathcal{O}_i^t$ denotes the corresponding temporal adaptation output, and $\alpha_{i}$ is a learnable scaling factor that modulates the influence of temporal information on the spatial representation inspired by the ReZero technique [1]. In our implementation, the input projection matrix $W_{i}^{down}$ is integrated within the RNN architecture, which already handles dimensionality reduction through its internal gating mechanisms. Meanwhile, the output projection $W_{i}^{up}$ is explicitly implemented to ensure proper integration back into the frozen RT-DETR architecture. Thus, the transformed feature representation $\tilde{\mathcal{E}}_i^t$ incorporates temporal context while maintaining the integrity of the original spatial representation $\mathcal{E}_i^t$ . In addition, our ablation studies show that $C_i^{RNN}$ can be reduced 4x relative to $C_i$ with minimal performance impact. + +For specific implementation of the RNN modules, we employ the ConvLSTM architecture [27], which effectively balances representational capacity with computational efficiency. This choice was informed by ConvLSTM's demonstrated effectiveness in handling spatio-temporal data in the RVT model [9], allowing our adaptation to employ established temporal modeling techniques. + +# 4. Evaluation of the EvRT-DETR Model + +The EvRT-DETR model is evaluated on two standard EBC object detection benchmarks: Gen1 [4] and 1Mpx [22] (also known as Gen4) datasets. These datasets contain diverse driving scenarios captured by Prophesee's neuromorphic vision sensors, providing an ideal testbed for validating our adaptation approach under real-world conditions. + +# 4.1. Experimental Setting + +Datasets. We evaluate on two standard Propheesee automotive datasets: Gen1 [4] and 1Mpx [22]. Gen1 contains $39+$ hours of recordings at $304 \times 240$ resolution with manual annotations (1-4 Hz) for cars and pedestrians, presenting + +challenges with sparse labels and stationary objects. 1Mpx offers $14+$ hours at higher resolution $(1280 \times 720)$ with denser automatic annotations $(60\mathrm{Hz})$ for cars, pedestrians, and two-wheelers. Both datasets are organized as 60-second clips with standard train/test splits. 1Mpx's higher resolution and frame rate enable evaluation on more complex scenes, while Gen1's annotation pattern tests robustness to sparse temporal supervision. + +Event Representations. We adopt the well-established stacked 2D Histogram approach from RVT protocols [9], constructing event frames from each $50\mathrm{ms}$ time window. Each frame is subdivided into 10 equally time-spaced $5\mathrm{ms}$ intervals, creating a representation with sufficient temporal resolution for our adaptation framework while remaining compatible with standard image processing architectures. + +For the Gen1 dataset (original size (240, 304)), we pad frames to (256, 320) to ensure divisibility by 32 as required by the feature pyramid architecture of RT-DETR. For 1Mpx, we downsize from (720, 1280) to (360, 640) using bilinear interpolation and pad to (384, 640). This maintains consistency with prior work while allowing for fair computational comparison. + +Two-stage Training Process. We train EvRT-DETR in two stages: (1) training the RT-DETR detector on individual event frames using standard practices (Adam optimizer; EMA of weights) then (2) freezing these parameters and training only the lightweight RNN modules on both random and sequential clips to develop temporal memory. The first stage establishes robust spatial feature extraction, while the second adds temporal processing with minimal parameter overhead. For temporal training, we maintain RNN memory continuity across sequential clips while resetting for random clips using different clip lengths for Gen1 (21 frames) and 1Mpx (10 frames). Full implementation details are provided in the supplementary material. + +
ModelGen11Mpx/Gen4Params (M)
mAP (%)Runtime (ms)mAP (%)Runtime (ms)
AsyNet [18]14.5---11.4
AEGNN [26]16.3---20.0
Spiking DenseNet [3]18.9---8.2
Events-RetinaNet [22]34.0-18.0-32.8
RED [22]40.016.743.039.324.1
AEC [19]47.010.6 (3.9)48.437.6 (13.8)46.5
RVT-B [9]47.210.247.411.918.5
GET-T [20]47.916.848.418.221.9
S5-ViT-B [34]47.48.247.29.617.5
ASTMNet [16]46.735.648.372.3>100
SAST-CB [21]48.2-48.757.5 (19.7)18.9
ERGO-12 [33]50.469.940.6100.059.6
RT-DETR-T (ours)46.06.444.18.620.1
RT-DETR-B (ours)47.610.545.214.942.8
EvRT-DETR-T (ours)52.38.449.912.534.4
EvRT-DETR-B (ours)52.712.750.118.857.1
+ +Table 1. Performance Comparison on Event-based Camera Datasets. Our approach achieves state-of-the-art results on both Gen1 and 1Mpx benchmarks with significant improvements over previous methods. The first stage of the framework (RT-DETR-T/B) already achieves competitive performance using only image-based detection, while the proposed I2EvDet-adapted models (EvRT-DETR-T/B) surpass all existing approaches. Runtime measurements are reported in milliseconds on an NVIDIA T4 GPU. Values in parentheses represent original reported times on different hardware with T4-equivalent times calculated using FLOPS/throughput conversion. + +Evaluation Protocol. We evaluate EvRT-DETR using the standard COCO (Common Objects in Context) Mean Average Precision (mAP) measure [17] that allows direct comparison with existing methods. For consistency with previous works, we use the EBC-specific implementation of the COCO metrics provided by Propheesee's Automotive Dataset Toolbox [22, 23]. All inference times reported in Table 2 are standardized to NVIDIA T4 GPU performance for fair comparison. Inference times include model forward pass and post-processing but exclude event preprocessing, following standard reporting practices. When prior work reported only preprocessing-inclusive times, we use their reported values directly. + +# 4.2. Performance Analysis + +Table 1 presents a comprehensive comparison of our EvRT-DETR models against state-of-the-art approaches on the Gen1 and 1Mpx benchmarks. + +First, our baseline adaptation using two RT-DETR variants, RT-DETR-T (Tiny, with ResNet-18 backbone) and RT-DETR-B (Base, with ResNet-50 backbone), and trained on simple event frames achieves competitive performance (46.0% and 47.6% mAP on Gen1; 44.1% and 45.2% mAP on 1Mpx) without requiring specialized architectures or complex event representations. This demonstrates that modern mainstream object detectors can be effectively adapted to event data through appropriate training. + +More importantly, our complete EvRT-DETR models + +with temporal adaptation significantly outperform all existing methods, achieving new state-of-the-art results on both benchmarks. EvRT-DETR-B reaches $52.7\%$ mAP on Gen1 $(+2.3\%$ over the previous best) and $50.1\%$ mAP on 1Mpx $(+1.4\%$ over the previous best). Even our lighter EvRT-DETR-T model surpasses all existing approaches while maintaining faster inference times. + +Figure 4 shows how the temporal adaptation module provides crucial capabilities for handling stationary objects - a fundamental challenge for event cameras. When objects are in motion, both RT-DETR and EvRT-DETR perform well, yet, when motion stops and event generation becomes minimal, EvRT-DETRs temporal memory enables consistent detection. + +These performance improvements come with minimal computational overhead. While we focus primarily on detection accuracy in this work, the runtime measurements in Table 1 demonstrate that our approach maintains competitive inference speeds despite adding temporal processing capabilities. + +# 4.3. Framework Generalizability + +To demonstrate that our results are not specific to the RT-DETR architecture, we apply the I2EvDet framework to YOLOX [7]. YOLOX is a CNN-based detector that contrasts with RT-DETR's transformer architecture and has been used in other EBC applications [9, 34]. Table 2 shows that our temporal adaptation approach consistently + +
DetectormAP (%)
OriginalI2EvDetImprovement
RT-DETR-T46.052.3+6.3
RT-DETR-B47.652.7+5.1
YOLOX-T36.042.4+6.4
YOLOX-X43.447.8+4.4
+ +improves performance across different baseline architectures on the Gen1 dataset. The substantial improvements (4.4-6.4 mAP) across both RT-DETR and YOLOX variants validate that I2EvDet represents a general framework for adapting image-based detectors to temporal domains rather than an architecture-specific optimization. Additional details are provided in the supplementary material. + +# 4.4. Analysis of Adaptation Design Choices + +We systematically analyze key design choices in our temporal adaptation module to better understand the factors contributing to EvRT-DETR's performance. This analysis provides insights into efficient adaptation strategies for event-based object detection. Additional ablation studies are provided in the supplementary material. + +Table 2. Generalizability of I2EvDet Framework Across Different Object Detection Architectures on the Gen1 Dataset. Our temporal adaptation consistently improves performance regardless of the base detector architecture. + +
ModelmAP (%)\( \mathrm{mAP}_{50} \) (%)\( \mathrm{mAP}_{75} \) (%)
RT-DETR-B (ours)47.676.349.5
EvRT-DETR-B (0,1,1)51.080.754.0
EvRT-DETR-B (1,0,1)52.281.355.1
EvRT-DETR-B (1,1,0)52.481.755.3
EvRT-DETR-B (1,1,1)52.782.056.0
+ +# 4.4.1. Multiscale Temporal Adaptation + +A crucial aspect of our approach is the insertion of temporal modules at multiple scales in the feature hierarchy. Table 3 shows performance when selectively removing ConvLSTM cells from specific feature levels with the array notation $(x,y,z)$ indicating presence (1) or absence (0) at each scale $\{\mathcal{E}_3,\mathcal{E}_4,\mathcal{E}_5\}$ . + +The results reveal a clear pattern: the lowest-level features $(\mathcal{E}_3)$ benefit most significantly from temporal adaptation as performance drops substantially when this module is removed. The impact diminishes at higher feature levels, suggesting that temporal information is most valuable for fine-grained, lower-level features that capture detailed motion patterns. This insight can guide more efficient adapta + +tion designs, wherein computational resources are directed to where they provide maximum benefit. + +Table 3. Effect of Temporal Module Placement. Performance on the Gen1 dataset with ConvLSTM modules at different feature scales. The array notation $(x,y,z)$ indicates presence (1) or absence (0) of temporal modules at scales $\mathcal{E}_3$ , $\mathcal{E}_4$ , and $\mathcal{E}_5$ , respectively. Low-level features benefit most from temporal adaptation. + +
ModelmAP (%)\( \mathrm{mAP}_{50} \) (%)\( \mathrm{mAP}_{75} \) (%)\( N_T \) (M)
EvRT-DETR-B (M=64)52.181.354.92.3
EvRT-DETR-B (M=128)52.581.755.65.4
EvRT-DETR-B (M=256)52.782.056.014.4
EvRT-DETR-B (M=512)52.981.956.042.9
+ +Table 4. Impact of Temporal Module Capacity. Performance on Gen1 dataset with varying ConvLSTM hidden dimensions (M). The rightmost column shows the number of trainable parameters (in millions) in the temporal adaptation module. + +# 4.4.2. Parameter Efficiency in Adaptation + +Inspired by parameter-efficient techniques such as LoRA (Low-Rank Adaptation) [15], we investigate the minimal temporal module capacity needed for effective adaptation. Table 4 shows EvRT-DETR-B performance with varying hidden dimensions in the ConvLSTM modules. + +Remarkably, reducing hidden dimensions to 64 features (a $4\mathrm{x}$ reduction from our baseline) results in only modest performance degradation $(-0.6\% \mathrm{mAP})$ while requiring just $2.3\mathrm{M}$ additional parameters - merely $5.4\%$ of the base RT-DETR-B model's $42.8\mathrm{M}$ parameters. This configuration closely resembles LoRA-style adaptation, achieving substantial performance gains with minimal parameter overhead. Even our most parameter-efficient adaptation $(\mathbf{M} = 64)$ outperforms all previous state-of-the-art approaches, demonstrating that effective temporal adaptation does not require extensive architectural modifications. Increasing the hidden dimension to 512 features $(42.9\mathrm{M}$ parameters, nearly doubling the model size) yields only a marginal $0.2\% \mathrm{mAP}$ improvement, indicating diminishing returns. + +# 5. Conclusions + +This work introduces I2EvDet, a framework that efficiently transforms image-based detectors into video-capable models using minimal modifications. Applied to event-based cameras, our approach demonstrates that mainstream object detection architectures can be effectively adapted to temporal domains without specialized redesign. RT-DETR trained on simple event representations achieves performance comparable to specialized methods, while our I2EvDet-adapted model EvRT-DETR achieves state-of-the-art results on Gen1 $(+2.3\mathrm{mAP})$ and 1Mpx $(+1.4\mathrm{mAP})$ benchmarks. Ablation studies provide insights about optimal temporal module design and parameter efficiency. These results suggest that bridging mainstream computer vision and specialized domains can be achieved through targeted adaptation rather than complete architectural redesign, potentially accelerating progress across diverse temporal visual domains. + +# Acknowledgments + +The authors extend their sincere gratitude to Yuewei Lin (Brookhaven Lab) for his valuable feedback and suggestions. This work presented in this paper was funded by the U.S. Department of Energy (DOE), National Nuclear Security Administration, Office of Defense Nuclear Nonproliferation Research and Development. The manuscript has been authored by Brookhaven National Laboratory managed by Brookhaven Science Associates, LLC under the Contract No. DE-SC0012704 with the U.S. Department of Energy. + +# References + +[1] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Gary Cottrell, and Julian J. McAuley. ReZero is all you need: fast convergence at large depth. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, UAI 2021, Virtual Event, 27-30 July 2021, pages 1352-1361. AUAI Press, 2021. 6 +[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213-229. Springer, 2020. 2, 3 +[3] Loïc Cordone, Benoit Miramond, and Philippe Thiérison. Object detection with spiking neural networks on automotive event data. In International Joint Conference on Neural Networks, IJCNN 2022, Padua, Italy, July 18-23, 2022, pages 1-8. IEEE, 2022. 3, 7 +[4] Pierre de Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. CoRR, abs/2001.08499, 2020. 2, 6 +[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4 +[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jorg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3, 4 +[7] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 7 +[8] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. CoRR, abs/2211.12324, 2022. 3 +[9] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13884-13893. IEEE, 2023. 3, 4, 6, 7 + +[10] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 580-587. IEEE Computer Society, 2014. 3 +[11] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. Trans. Mach. Learn. Res., 2024, 2024. 4 +[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4, 5 +[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 3 +[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 2790-2799. PMLR, 2019. 2, 4, 5 +[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 4, 5, 8 +[16] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Trans. Image Process., 31:2975-2987, 2022. 3, 7 +[17] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7 +[18] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII, pages 415-431. Springer, 2020. 7 +[19] Yansong Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington, + +DC, USA, February 7-14, 2023, pages 2056-2064. AAAI Press, 2023. 3, 7 +[20] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. GET: group event transformer for event-based vision. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 6015-6025. IEEE, 2023. 2, 3, 7 +[21] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16794-16804. IEEE, 2024. 3, 7 +[22] Etienne Perot, Pierre de Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6, 7 +[23] Propheesee. Propheesee's automotive dataset toolbox. https://github.com/propheesee-ai/propheesee-automotive-dataset-toolbox, 2024.7 +[24] Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 779-788. IEEE Computer Society, 2016. 2, 3 +[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 91-99, 2015. 2 +[26] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. AEGNN: asynchronous event-based graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12361-12371. IEEE, 2022. 3, 7 +[27] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 802-810, 2015. 6 +[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 3 +[29] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on + +Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3 +[30] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on Real-time object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16965-16974. IEEE, 2024. 2, 4, 5 +[31] Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. Convolution meets lora: Parameter efficient finetuning for segment anything model. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 4 +[32] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3 +[33] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 12800-12810. IEEE, 2023. 2, 3, 7 +[34] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5819-5828. IEEE, 2024. 3, 7 \ No newline at end of file diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/images.zip b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a01ff93d789d56ca50d696a0a8bbc0e288e074b5 --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:0a41db3860cf79b9babd8a46132ed38fe1cc628f997bce362515c5d55914026b +size 458576 diff --git a/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/layout.json b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..05f2b6d87866812434b41e478cf82d9379e5fce6 --- /dev/null +++ b/2025/EvRT-DETR_ Latent Space Adaptation of Image Detectors for Event-based Vision/layout.json @@ -0,0 +1,7770 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 487, + 138 + ], + "type": "text", + "content": "EvRT-DETR: Latent Space Adaptation of Image Detectors for Event-based Vision" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 113, + 160, + 496, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 160, + 496, + 196 + ], + "spans": [ + { + "bbox": [ + 113, + 160, + 496, + 196 + ], + "type": "text", + "content": "Dmitrii Torbunov1, Yihui Ren1, Animesh Ghose1, Odera Dim1, Yonggang Cui1 \n1Brookhaven National Laboratory, Upton, NY, USA" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 94, + 202, + 512, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 202, + 512, + 214 + ], + "spans": [ + { + "bbox": [ + 94, + 202, + 512, + 214 + ], + "type": "text", + "content": "dtorbunov@bnl.gov, yren@bnl.gov, aghose@bnl.gov, dodera@bnl.gov, ycui@bnl.gov" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "spans": [ + { + "bbox": [ + 151, + 242, + 200, + 254 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 269, + 297, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 269, + 297, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 269, + 297, + 628 + ], + "type": "text", + "content": "Event-based cameras (EBCs) have emerged as a bioinspired alternative to traditional cameras, offering advantages in power efficiency, temporal resolution, and high dynamic range. However, development of image analysis methods for EBCs is challenging due to the sparse and asynchronous nature of the data. This work addresses the problem of object detection for EBC cameras. The current approaches to EBC object detection focus on constructing complex data representations and rely on specialized architectures. We introduce I2EvDet (Image-to-Event Detection), a novel adaptation framework that bridges mainstream object detection with temporal event data processing. First, we demonstrate that a Real-Time DEetection Transformer, or RT-DETR, a state-of-the-art natural image detector, trained on a simple image-like representation of the EBC data achieves performance comparable to specialized EBC methods. Next, as part of our framework, we develop an efficient adaptation technique that transforms image-based detectors into event-based detection models by modifying their frozen latent representation space via minimal architectural additions. The resulting EvRT-DETR model reaches state-of-the-art performance on the standard benchmark datasets Gen1 (mAP +2.3) and 1Mpx/Gen4 (mAP +1.4). These results demonstrate a fundamentally new approach to EBC object detection through principled adaptation of mainstream architectures, offering an efficient alternative with potential applications to other temporal visual domains. The code is available at: https://github.com/realtime-intelligence/evrt-detr." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 656, + 135, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 656, + 135, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 656, + 135, + 668 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Event-based cameras (EBCs) present a biologically inspired alternative to traditional frame-based cameras. Unlike conventional cameras that capture frames at predetermined in" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 318, + 243, + 552, + 417 + ], + "blocks": [ + { + "bbox": [ + 318, + 243, + 552, + 417 + ], + "lines": [ + { + "bbox": [ + 318, + 243, + 552, + 417 + ], + "spans": [ + { + "bbox": [ + 318, + 243, + 552, + 417 + ], + "type": "image", + "image_path": "1f4beb4a8a1e624ea438d69e0ef38ecb723a8f5f86ba498b6069617facc1e5df.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 428, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 506 + ], + "type": "text", + "content": "Figure 1. Object Detection Performance vs. Inference Time. A summary of object detection performance (COCO mAP) versus inference time (ms) of various models on the 1Mpx automotive dataset. Circle size is proportional to the number of model parameters. Inference times are reported for NVIDIA T4 GPU. Our EvRT-DETR models achieve state-of-the-art accuracy while maintaining competitive inference speeds." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": "tervals, EBC pixels operate asynchronously, generating data only when brightness changes exceed a threshold. This results in sparse, asynchronous data streams where each event is a tuple containing pixel location, timestamp, and polarity (whether brightness increased or decreased), as illustrated in Figure 2. This approach offers remarkable advantages: low power consumption (as low as " + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "inline_equation", + "content": "10\\mathrm{mW}" + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": "), reduced data transfer rates, exceptional temporal resolution (on the order of " + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "inline_equation", + "content": "\\mu s" + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": "), and high dynamic range (" + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "inline_equation", + "content": ">100" + }, + { + "bbox": [ + 313, + 531, + 556, + 663 + ], + "type": "text", + "content": " dB). These properties have led to widespread adoption in autonomous driving, robotics, and wearable electronics [6]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "These unique characteristics of EBC data - sparsity, asynchronicity, and temporal nature - create fundamental challenges when applying traditional computer vision techniques. Early attempts to directly apply conventional" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9812" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 295, + 143 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 295, + 143 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 295, + 143 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 295, + 143 + ], + "type": "image", + "image_path": "cdf3dd14f6e384cce330ebd2cff576c77f5ead735509b710f19ae1d2ab2b7679.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "lines": [ + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 295, + 185 + ], + "type": "text", + "content": "Figure 2. Event-based Camera Output Visualization. Sample frames from the 1Mpx dataset showing event polarity. Red indicates positive events, and blue denotes negative events." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 208, + 295, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 295, + 363 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 295, + 363 + ], + "type": "text", + "content": "image-based object detectors to simple event representations yielded poor performance [22], highlighting the gap between these domains. This led to two predominant research directions: (1) constructing sophisticated event representations that transform EBC data into formats compatible with existing detectors (e.g., [33]) or (2) designing specialized architectures tailored to handle the temporal nature of event data (e.g., [20]). A common assumption in these approaches is that EBC data processing requires either specialized representations or architectures due to the fundamental differences between event- and frame-based vision, leading to development paths that often diverge from mainstream computer vision research." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "text", + "content": "Meanwhile, mainstream object detection has seen remarkable progress, evolving from convolutional neural network (CNN)-based architectures, e.g.,YOLO [24] and Faster R-CNN [25], to transformer-based models like DETR (DEtection Transformer) [2]. Recent advances have culminated in highly efficient and accurate models, such as the Real-Time DETection TRansformer (RT-DETR) [30] that achieves exceptional performance while maintaining real-time inference speeds. Though primarily optimized for processing individual images, these state-of-the-art detectors offer powerful feature extraction capabilities that potentially could be leveraged for temporal data." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 628 + ], + "type": "text", + "content": "This evolution in mainstream detection creates an opportunity to bridge the gap between conventional and event-based vision. Rather than assuming fundamental incompatibility between these domains, we explore a complementary direction: adapting state-of-the-art frame-based detectors to process event data effectively with minimal modifications. In this work, we investigate how RT-DETR can be strategically adapted to the event-based domain, demonstrating that modern object detection architectures can successfully work on event data and opening a new research direction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Our approach proceeds in two stages to systematically adapt RT-DETR to event data. First, we train RT-DETR directly on simple image-like representations of EBC data using standard training procedures without special modifications. This straightforward approach yields surprising results - even when processing only fixed " + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "inline_equation", + "content": "50\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": " time frames, the detector achieves performance comparable to special" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ized EBC-specific methods explicitly designed for video object detection on EBC data. This is particularly unexpected as these specialized methods deliberately exploit the temporal nature of EBC data and process information from the distant past, while our approach performs detection using only a single fixed time window (frame) of " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "50~\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", similar to traditional frame-based cameras." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 166, + 555, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 166, + 555, + 334 + ], + "spans": [ + { + "bbox": [ + 313, + 166, + 555, + 334 + ], + "type": "text", + "content": "Building on this foundation, we introduce I2EvDet (Image-to-Event Detection), a technique that enables straightforward transformation of image-based object detectors into video-capable models. This approach leverages latent space adaptation principles [14, 15] where we freeze the pre-trained RT-DETR model and insert a lightweight recurrent neural network (RNN) module within the encoder's latent representation space. This design allows the model to capture temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead. The resulting adaptation effectively leverages information across multiple time windows without requiring a specialized video detection architecture." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 343, + 556, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 556, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 556, + 489 + ], + "type": "text", + "content": "The combined approach, called EvRT-DETR, achieves state-of-the-art performance on standard EBC benchmarks, outperforming specialized methods on both Gen1 [4] (mAP +2.3) and 1Mpx [22] (mAP +1.4) datasets. Importantly, the results are achieved using only standard modules from natural image and video analysis without requiring domain-specific modifications. This demonstrates not only the effectiveness of our approach for EBC data, but also suggests a broader paradigm for adapting image detectors to temporal domains. As shown in Figure 1, our model achieves superior accuracy while maintaining competitive inference speeds, making it practical for real-world applications." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 325, + 498, + 438, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 498, + 438, + 509 + ], + "spans": [ + { + "bbox": [ + 325, + 498, + 438, + 509 + ], + "type": "text", + "content": "Contributions of this work:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 520, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 520, + 553, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 520, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 314, + 520, + 553, + 590 + ], + "type": "text", + "content": "- We show that when trained on simple image-like representations of EBC data using standard procedures, mainstream object detectors like RT-DETR can achieve performance comparable to specialized EBC methods, offering a complementary approach to domain-specific architectures." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 592, + 553, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 592, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 592, + 553, + 651 + ], + "type": "text", + "content": "- We propose I2EvDet, an efficient adaptation technique that transforms image-based detectors into video-capable models through minimal architectural additions to the frozen pre-trained model and with potential applications to various temporal visual domains." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 651, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 651, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 651, + 553, + 712 + ], + "type": "text", + "content": "- We validate our approach in the challenging domain of event-based cameras, where EvRT-DETR achieves state-of-the-art performance on the standard Gen1 (mAP +2.3) and 1Mpx (mAP +1.4) EBC benchmarks while maintaining competitive inference speeds." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9813" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 142, + 84 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 295, + 235 + ], + "type": "text", + "content": "Research at the intersection of event-based cameras and object detection spans multiple domains in computer vision. Current approaches primarily focus on two directions: designing efficient representations for the unique properties of event data and developing specialized neural architectures to process these representations. However, despite its success in other domains, a third direction – adapting mainstream computer vision models to event data – remains underexplored. This section reviews relevant work across these areas, highlighting how our approach bridges the gap between specialized event-based methods and mainstream object detection advances." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 241, + 204, + 255 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 241, + 204, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 241, + 204, + 255 + ], + "type": "text", + "content": "2.1. EBC Data Representations" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "spans": [ + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "content": "An event camera produces a stream of event data in the form " + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "inline_equation", + "content": "(t, p, x, y)" + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "content": " is the timestamp of an event, " + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "content": " is the polarity (positive or negative) of the brightness change, and " + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 258, + 295, + 342 + ], + "type": "text", + "content": " is the location of the event pixel. To simplify the analysis, the event data often are transformed into alternative representations more suitable for conventional image analysis algorithms." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": "The simplest image-like representations of event camera data are Event Frames or 2D Histograms [6], where events within fixed time windows are accumulated into two-dimensional frames of shape " + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "inline_equation", + "content": "(2, H, W)" + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": " with polarity as the first dimension and spatial coordinates as the remaining dimensions. A natural evolution is the Stacked 2D Histogram, which further partitions each frame into " + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": " time intervals, creating a representation of shape " + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "inline_equation", + "content": "(2, T, H, W)" + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": " that typically is reshaped to " + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "inline_equation", + "content": "(2T, H, W)" + }, + { + "bbox": [ + 55, + 342, + 295, + 486 + ], + "type": "text", + "content": " and treated as a natural image. While these representations are straightforward, directly applying existing computer vision algorithms to them has yielded poor performance in the past [22]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 487, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 487, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 487, + 295, + 605 + ], + "type": "text", + "content": "To improve performance, other representations are being explored. For example, a Time Surface (TS) representation [6, 33] is an image-like representation where pixel values encode the time elapsed since the last event in a given pixel. Thus, unlike the simplest fixed time window representations, TS potentially can encode arbitrarily distant past. Empirically, it has been shown that using the TS representations offers significantly better performance on the object detection tasks compared to the 2D Histogram representations [22]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 715 + ], + "type": "text", + "content": "Many other EBC data representations also have been proposed. The recent ERGO-12 (Event Representation through Gromov-Wasserstein Optimization) [33] work has developed efficient image-like representations, achieving state-of-the-art object detection performance on the Gen1 dataset. Meanwhile, other approaches reformulate object detection as a three-dimensional (3D) problem by treating events as points in spatial-temporal space or discretizing them into voxel grids [6]. While computationally heavier," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "these representations can preserve exact event timing, potentially enabling better handling of overlapping objects and complex motions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 117, + 508, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 508, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 508, + 129 + ], + "type": "text", + "content": "2.2. EBC Object Detection Architectures" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 134, + 553, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 134, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 313, + 134, + 553, + 182 + ], + "type": "text", + "content": "Once the data representation is selected, one can start to develop object detection algorithms. There are several families of object detection methods designed for particular data representations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 182, + 553, + 290 + ], + "type": "text", + "content": "The most straightforward approach applies conventional image detection methods directly to 2D Histogram frames, but this yields relatively poor performance [22]. More advanced methods, such as Recurrent Vision Transformer (RVT) [9] and State Space Vision Transformer (S5-ViT-B) [34], extend this approach by incorporating temporal memory to capture sequential information, significantly improving detection quality compared to frame-by-frame processing." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 291, + 553, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 291, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 553, + 411 + ], + "type": "text", + "content": "The most recent family of methods attempts to use more efficient event representations to further improve object detection performance. For example, 12-channel ERGO-12 [33] constructs an optimized image-like event representation and shows that the standard object detection methods perform exceptionally well on it. Alternatively, Asynchronous Spatio-Temporal Memory Network (ASTMNet) [16] and Adaptive Event Conversion (AEC) [19] attempt to combine better event representation while using novel neural architectures to expand state-of-the-art results." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 411, + 553, + 459 + ], + "type": "text", + "content": "Other methods have been developed that treat EBC object detection as a 3D detection problem [8, 22, 26], exploit benefits of neuromorphic architectures [3], or explore various hybrid approaches [20, 21]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 468, + 508, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 508, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 508, + 480 + ], + "type": "text", + "content": "2.3. Transformer-based Object Detection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 544 + ], + "type": "text", + "content": "Object detection has been revolutionized by transformer architectures in recent years. DETR [2] was the first widely successful transformer-based [28] model for object detection, offering a simple and elegant alternative to traditional CNN-based architectures [10, 24]." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 553, + 629 + ], + "type": "text", + "content": "DETR follows a hybrid CNN-Transformer design with a backbone-encoder-decoder structure. It uses a CNN backbone (typically ResNet [13]) for feature extraction, followed by a Transformer encoder that processes these features into tokens and captures their correlations. The Transformer decoder then identifies objects through cross-attention to the encoder outputs." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 553, + 715 + ], + "type": "text", + "content": "Despite its innovative approach, the original DETR suffered from limitations, including training instability [29], poor performance on small objects [2], and high computational costs during inference [32]. Subsequent research has addressed these issues through architectural improvements such as Feature Pyramid Networks (FPN), specialized decoder tokens, and deformable attention mechanisms [2, 32]." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9814" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 68, + 558, + 292 + ], + "blocks": [ + { + "bbox": [ + 59, + 68, + 558, + 292 + ], + "lines": [ + { + "bbox": [ + 59, + 68, + 558, + 292 + ], + "spans": [ + { + "bbox": [ + 59, + 68, + 558, + 292 + ], + "type": "image", + "image_path": "bf19cc5e0a7f5a6e87ddbbf7189f387b4035aee94ea1d2429bad0b5ee0590aab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "type": "text", + "content": "Figure 3. Overview of the I2EvDet Framework Applied to RT-DETR. This approach adapts a pre-trained RT-DETR model to process temporal event data using minimal architectural modifications. The backbone and encoder components remain frozen during temporal adaptation, while trainable temporal modules are strategically inserted at multiple feature scales. Left: Event camera frames are processed through the frozen RT-DETR backbone and encoder. Center: Our temporal adaptation modules operate on multiscale features through residual connections (red pathways), enabling information flow across time steps while preserving spatial representation integrity. Right: Feature fusion and decoder stages produce final detections. The detailed view of the Temporal Adaptation Module (bottom right) shows the projection operations " + }, + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "type": "inline_equation", + "content": "(W_{i}^{down}, W_{i}^{up})" + }, + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "type": "text", + "content": " and recurrent memory module " + }, + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "type": "inline_equation", + "content": "(\\mathbf{R}_{i})" + }, + { + "bbox": [ + 55, + 300, + 555, + 389 + ], + "type": "text", + "content": " that enable efficient temporal modeling with minimal additional parameters." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 410, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 495 + ], + "type": "text", + "content": "These incremental advances have culminated in the RT-DETR [30], which achieves exceptional performance across all object scales with stable training and significantly faster inference times. RT-DETR represents the current state-of-the-art in efficient transformer-based detection and serves as an ideal foundation for adaptation to specialized domains, such as event-based vision." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 503, + 220, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 503, + 220, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 503, + 220, + 517 + ], + "type": "text", + "content": "2.4. Model Adaptation Techniques" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 628 + ], + "type": "text", + "content": "While specialized architectures dominate event-based vision, adapting mainstream architectures offers compelling advantages. Parameter-efficient methods [12, 14, 15] modify pre-trained models with minimal changes by adding compact modules to frozen backbones. These approaches have proven effective across language [5], vision [31], and multimodal systems [11], but they remain underexplored for event-based vision despite their efficiency and transfer learning benefits." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 295, + 715 + ], + "type": "text", + "content": "Given the success of adaptation techniques in other domains and the power of transformer-based detection models, there is a clear opportunity to bridge these approaches for event-based vision. Our work explores this intersection, demonstrating that mainstream object detectors can be effectively adapted to EBC data with minimal modifications rather than requiring specialized architectures." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 408, + 375, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 408, + 375, + 421 + ], + "spans": [ + { + "bbox": [ + 314, + 408, + 375, + 421 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 430, + 555, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 555, + 574 + ], + "type": "text", + "content": "This section presents I2EvDet, our adaptation framework for event-based object detection. Unlike approaches that rely on specialized architectures or complex representations, I2EvDet bridges mainstream object detection with temporal event data via strategic, minimal modifications. Our framework demonstrates how image-based detectors can be efficiently transformed into video-capable models while preserving their core strengths. We first describe our image-like representation of event data then introduce the I2EvDet framework. Finally, we detail the technical implementation and discuss how our approach addresses the unique temporal dependencies in event camera data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 585, + 440, + 598 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 440, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 440, + 598 + ], + "type": "text", + "content": "3.1. Event Representation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 554, + 676 + ], + "type": "text", + "content": "Our adaptation framework begins with transforming EBC data into an image-like representation that is compatible with mainstream object detection models. We employ the Stacked 2D Histogram representation [6] as it provides a straightforward bridge between event data and conventional image processing architectures." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "To ensure direct comparability with existing methods, we adopt the frame construction parameters from RVT [9], which allows us to isolate the impact of our architectural" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9815" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": "choices from data pre-processing. Specifically, we take the stream of events and partition it into a series of consecutive fixed time windows of " + }, + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{frame}} = 50" + }, + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": " ms. Each such window corresponds to a single frame. Next, we further subdivide each frame into 10 intervals of " + }, + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "inline_equation", + "content": "T_{\\mathrm{bin}} = 5" + }, + { + "bbox": [ + 55, + 72, + 294, + 132 + ], + "type": "text", + "content": " ms." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "text", + "content": "To construct a frame corresponding to an interval " + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "inline_equation", + "content": "[t_0, t_0 + T_{\\mathrm{frame}}]" + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "text", + "content": ", we create an intermediate stacked histogram " + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "text", + "content": " from a set of events " + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 133, + 295, + 167 + ], + "type": "text", + "content": " in that interval:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 59, + 175, + 291, + 202 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 59, + 175, + 291, + 202 + ], + "spans": [ + { + "bbox": [ + 59, + 175, + 291, + 202 + ], + "type": "interline_equation", + "content": "S(p,t_{i},y,x) = \\sum_{\\{e|t_{e}\\in [t_{0},t_{0} + T_{\\text{frame}})\\}}\\delta_{x}^{x_{e}}\\delta_{y}^{y_{e}}\\delta_{p}^{p_{e}}\\mathbf{1}_{[t_{i},t_{i + 1})}(t_{e})", + "image_path": "4eae8e71941df2233ab3ca2a558be0e20c8f2609d9af9c38b3d3ba8feff91abd.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "e = (t_e, p_e, x_e, y_e)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " represents an individual event with time stamp " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "t_e" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": ", polarity " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "p_e" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": ", and spatial coordinates " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "(x_e, y_e)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "t_i = t_0 + i \\cdot T_{\\mathrm{bin}}" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "i \\in [0, 10)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " is the histogram bin index; " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "(x, y)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " are the spatial indices in the output histogram; " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "\\delta_a^b" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " is the Kronecker delta function; and " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "\\mathbf{1}" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " is an indicator function. Once a stacked histogram " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "(2, 10, H, W)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " is constructed, we merge the polarity and time dimensions to obtain an image-like 2D frame " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": " of shape " + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "inline_equation", + "content": "(20, H, W)" + }, + { + "bbox": [ + 55, + 210, + 295, + 307 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 313, + 295, + 337 + ], + "type": "text", + "content": "3.2. I2EvDet: Image-to-Event Detection Framework" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 426 + ], + "type": "text", + "content": "I2EvDet is a principled framework for adapting image-based object detectors to process temporal data. While specialized architectures typically are developed for event-based object detection, this approach demonstrates that strategic adaptation of mainstream detectors can achieve superior performance with minimal architectural modifications." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 426, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 510 + ], + "type": "text", + "content": "The I2EvDet framework applies latent space adaptation principles [12, 14, 15], which enable domain adaptation through minimal changes to pre-trained models. The I2EvDet framework follows a two-stage process, and it is applicable to any object detection model with a clean separation between feature extraction and object detection modules (e.g.,YOLO and DETR families)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 569 + ], + "type": "text", + "content": "First, we establish a solid foundation by training an object detector on individual EBC frames (subsection 3.1), treating them as conventional images. This creates a detector with powerful spatial representation capabilities optimized for event data." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 689 + ], + "type": "text", + "content": "In the second stage, we transform this image-based detector into a video-capable model. We freeze the pre-trained detector parameters and insert lightweight RNN modules between the feature extractor and detection components. We train the inserted temporal modules on the EBC video object detection task while keeping the original detector frozen. The resulting model captures temporal dynamics across frames while preserving the powerful representation capabilities of the original detector and adding minimal computational overhead." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "This two-stage approach retains learned spatial representations of the original image detector while efficiently" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "adding temporal capabilities. In this work, we apply I2EvDet to the RT-DETR model as the primary detector, resulting in the EvRT-DETR model. Additionally, we demonstrate the generalizability of the I2EvDet framework using experiments with YOLOX detectors." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 139, + 503, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 139, + 503, + 152 + ], + "spans": [ + { + "bbox": [ + 313, + 139, + 503, + 152 + ], + "type": "text", + "content": "3.3. Temporal Modeling Considerations" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 156, + 555, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 311 + ], + "type": "text", + "content": "A key theoretical consideration for this adaptation framework is the choice of temporal memory mechanism. Event cameras present a unique challenge: static objects become \"invisible\" as they generate no events when motionless. Their detection requires having a persistent memory of past events. While transformers have dominated recent vision architectures, their fixed context windows limit temporal memory. We employ RNNs for their theoretically unbounded memory capacity through recurrent state, making them uniquely suited for maintaining object persistence during periods of invisibility in the event stream. This theoretical foundation has informed the EvRT-DETR implementation." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 319, + 534, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 534, + 332 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 534, + 332 + ], + "type": "text", + "content": "3.4. Technical Implementation of EvRT-DETR" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 336, + 554, + 408 + ], + "type": "text", + "content": "The EvRT-DETR architecture implements the I2EvDet framework via strategic, minimal modifications to the RT-DETR model. As shown in Figure 3, we identify key integration points within the pre-trained model's latent space, where temporal information can be effectively incorporated without disrupting spatial feature extraction." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "text", + "content": "The RT-DETR encoder produces multiscale feature representations " + }, + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}" + }, + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "text", + "content": ", where each element corresponds to a transformer token encoding rich spatial information. These features, originating from the last three stages of the backbone " + }, + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "inline_equation", + "content": "\\{S_3,S_4,S_5\\}" + }, + { + "bbox": [ + 313, + 408, + 554, + 479 + ], + "type": "text", + "content": " [30], provide an ideal latent manifold for temporal adaptation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 480, + 554, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 554, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 554, + 576 + ], + "type": "text", + "content": "Following latent space adaptation principles, we introduce specialized temporal processing modules at each scale, implementing three parallel RNN-based memory modules " + }, + { + "bbox": [ + 313, + 480, + 554, + 576 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{R}_3,\\mathbf{R}_4,\\mathbf{R}_5\\}" + }, + { + "bbox": [ + 313, + 480, + 554, + 576 + ], + "type": "text", + "content": ". This multiscale temporal adaptation design ensures that temporal dynamics are captured at different levels of feature abstraction. Critically, each RNN module interfaces with its corresponding encoder feature map through residual connections, mathematically expressed as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 388, + 594, + 553, + 610 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 594, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 388, + 594, + 553, + 610 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {i} ^ {t, p r o j} = W _ {i} ^ {d o w n} \\cdot \\mathcal {E} _ {i} ^ {t} \\tag {1}", + "image_path": "8b1f9fc346c9ac0dca013f5e79a1a5ccc10874e6a2be49e01ddd11d280808d0e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 354, + 612, + 553, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 612, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 354, + 612, + 553, + 632 + ], + "type": "interline_equation", + "content": "\\left(\\mathcal {O} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t}\\right) = \\mathbf {R} _ {i} \\left(\\mathcal {E} _ {i} ^ {t, p r o j}, \\mathcal {M} _ {i} ^ {t - 1}\\right) \\tag {2}", + "image_path": "caf774de8c9cd9bfb21943fc59903b094272706d1eb34d7a47ce7e7d749c2a65.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 404, + 634, + 553, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 404, + 634, + 553, + 649 + ], + "spans": [ + { + "bbox": [ + 404, + 634, + 553, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {O} _ {i} ^ {t} = W _ {i} ^ {u p} \\cdot \\mathcal {O} _ {i} ^ {t, p r o j} \\tag {3}", + "image_path": "ad962ff594119c724371593fe3d04b0a4fb099ca116c4128b2c349b6d3334637.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 405, + 651, + 553, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 405, + 651, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 405, + 651, + 553, + 666 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {E}} _ {i} ^ {t} = \\mathcal {E} _ {i} ^ {t} + \\alpha_ {i} \\cdot \\mathcal {O} _ {i} ^ {t}, \\tag {4}", + "image_path": "5d1cbb95fa07fd25ff0752e9d5628b76865c13866ced407124494b16a4659537.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "inline_equation", + "content": "W_{i}^{down}\\in \\mathbb{R}^{C_{i}^{RNN}\\times C_{i}}" + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "inline_equation", + "content": "W_{i}^{up}\\in \\mathbb{R}^{C_{i}\\times C_{i}^{RNN}}" + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "text", + "content": " are the projection matrices that transform features into and out of the temporal processing space, " + }, + { + "bbox": [ + 313, + 673, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_i^t\\in \\mathbb{R}^{C_i^{RNN}\\times H_i\\times W_i}" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "9816" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 100, + 70, + 185, + 119 + ], + "blocks": [ + { + "bbox": [ + 80, + 72, + 90, + 179 + ], + "lines": [ + { + "bbox": [ + 80, + 72, + 90, + 179 + ], + "spans": [ + { + "bbox": [ + 80, + 72, + 90, + 179 + ], + "type": "text", + "content": "B" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 100, + 70, + 185, + 119 + ], + "lines": [ + { + "bbox": [ + 100, + 70, + 185, + 119 + ], + "spans": [ + { + "bbox": [ + 100, + 70, + 185, + 119 + ], + "type": "image", + "image_path": "a25c8b8d2bafef5fdaf3a86cf3a6f7f6bacec7cb889372b5dc816088a99d152b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 186, + 70, + 270, + 119 + ], + "blocks": [ + { + "bbox": [ + 186, + 70, + 270, + 119 + ], + "lines": [ + { + "bbox": [ + 186, + 70, + 270, + 119 + ], + "spans": [ + { + "bbox": [ + 186, + 70, + 270, + 119 + ], + "type": "image", + "image_path": "c8deaa347e462b8fe6bc41dfc358f059ca0987cae5cacba0e9138690e3969c0f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 272, + 70, + 357, + 119 + ], + "blocks": [ + { + "bbox": [ + 272, + 70, + 357, + 119 + ], + "lines": [ + { + "bbox": [ + 272, + 70, + 357, + 119 + ], + "spans": [ + { + "bbox": [ + 272, + 70, + 357, + 119 + ], + "type": "image", + "image_path": "2854ed932a2c056a16761c76c64e74bbab82b448b302eb8e0b90e8ba3f6b6c8d.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 359, + 71, + 442, + 120 + ], + "blocks": [ + { + "bbox": [ + 359, + 71, + 442, + 120 + ], + "lines": [ + { + "bbox": [ + 359, + 71, + 442, + 120 + ], + "spans": [ + { + "bbox": [ + 359, + 71, + 442, + 120 + ], + "type": "image", + "image_path": "4c246c15051eb71ba1eed325546ed123c0ca0a95cb047d22f8086cadca6dccc2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 443, + 71, + 529, + 120 + ], + "blocks": [ + { + "bbox": [ + 443, + 71, + 529, + 120 + ], + "lines": [ + { + "bbox": [ + 443, + 71, + 529, + 120 + ], + "spans": [ + { + "bbox": [ + 443, + 71, + 529, + 120 + ], + "type": "image", + "image_path": "bd45114d2465646c2d6672a47b8eba8b576245a1fb8a5c6b303bb00da5fec233.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 100, + 127, + 185, + 176 + ], + "blocks": [ + { + "bbox": [ + 100, + 127, + 185, + 176 + ], + "lines": [ + { + "bbox": [ + 100, + 127, + 185, + 176 + ], + "spans": [ + { + "bbox": [ + 100, + 127, + 185, + 176 + ], + "type": "image", + "image_path": "d4f8d120c7db1903aa90174a3b32261c075bf5246981b7b8292cd7168ffe380d.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 188, + 555, + 243 + ], + "lines": [ + { + "bbox": [ + 55, + 188, + 555, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 188, + 555, + 243 + ], + "type": "text", + "content": "Figure 4. Temporal Adaptation Benefits of EvRT-DETR. Visualization of object detection across sequential 1Mpx frames captured by a vehicle-mounted camera that gradually comes to a stop. While the camera is in motion (left frames), both detectors maintain accurate bounding boxes of stationary vehicles. As camera motion decreases (middle frames), event data becomes increasingly sparse. When the camera stops completely (right frames), RT-DETR (top row) loses detection capability, while EvRT-DETR (bottom row) maintains consistent detection by engaging its temporal memory module." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 186, + 127, + 270, + 176 + ], + "blocks": [ + { + "bbox": [ + 186, + 127, + 270, + 176 + ], + "lines": [ + { + "bbox": [ + 186, + 127, + 270, + 176 + ], + "spans": [ + { + "bbox": [ + 186, + 127, + 270, + 176 + ], + "type": "image", + "image_path": "615b91d1f381ebb892b4ae92e06f6450bfb463353537c761ccca95655bb13157.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 272, + 127, + 357, + 176 + ], + "blocks": [ + { + "bbox": [ + 272, + 127, + 357, + 176 + ], + "lines": [ + { + "bbox": [ + 272, + 127, + 357, + 176 + ], + "spans": [ + { + "bbox": [ + 272, + 127, + 357, + 176 + ], + "type": "image", + "image_path": "d428df93af6a95718f09b10ede61dc34ee458a16916893b5893782e53d25b028.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 359, + 127, + 442, + 176 + ], + "blocks": [ + { + "bbox": [ + 359, + 127, + 442, + 176 + ], + "lines": [ + { + "bbox": [ + 359, + 127, + 442, + 176 + ], + "spans": [ + { + "bbox": [ + 359, + 127, + 442, + 176 + ], + "type": "image", + "image_path": "05610d3f100e0d20f8b1dc2271f7d9c7f62f6dde6cc744b2918332a766334dd0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 443, + 127, + 529, + 176 + ], + "blocks": [ + { + "bbox": [ + 443, + 127, + 529, + 176 + ], + "lines": [ + { + "bbox": [ + 443, + 127, + 529, + 176 + ], + "spans": [ + { + "bbox": [ + 443, + 127, + 529, + 176 + ], + "type": "image", + "image_path": "d11a37e3cf9f1dade38af3dc95a6d085faf0e55a41376b33f5c7a47aa2f3c300.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": "represents the temporal memory state for the " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": "-th feature scale at time " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{O}_i^t" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " denotes the corresponding temporal adaptation output, and " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " is a learnable scaling factor that modulates the influence of temporal information on the spatial representation inspired by the ReZero technique [1]. In our implementation, the input projection matrix " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "W_{i}^{down}" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " is integrated within the RNN architecture, which already handles dimensionality reduction through its internal gating mechanisms. Meanwhile, the output projection " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "W_{i}^{up}" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " is explicitly implemented to ensure proper integration back into the frozen RT-DETR architecture. Thus, the transformed feature representation " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{E}}_i^t" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " incorporates temporal context while maintaining the integrity of the original spatial representation " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i^t" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": ". In addition, our ablation studies show that " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "C_i^{RNN}" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " can be reduced 4x relative to " + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "inline_equation", + "content": "C_i" + }, + { + "bbox": [ + 56, + 265, + 295, + 455 + ], + "type": "text", + "content": " with minimal performance impact." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 456, + 295, + 540 + ], + "type": "text", + "content": "For specific implementation of the RNN modules, we employ the ConvLSTM architecture [27], which effectively balances representational capacity with computational efficiency. This choice was informed by ConvLSTM's demonstrated effectiveness in handling spatio-temporal data in the RVT model [9], allowing our adaptation to employ established temporal modeling techniques." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 549, + 263, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 549, + 263, + 562 + ], + "spans": [ + { + "bbox": [ + 55, + 549, + 263, + 562 + ], + "type": "text", + "content": "4. Evaluation of the EvRT-DETR Model" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 641 + ], + "type": "text", + "content": "The EvRT-DETR model is evaluated on two standard EBC object detection benchmarks: Gen1 [4] and 1Mpx [22] (also known as Gen4) datasets. These datasets contain diverse driving scenarios captured by Prophesee's neuromorphic vision sensors, providing an ideal testbed for validating our adaptation approach under real-world conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 648, + 178, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 178, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 178, + 661 + ], + "type": "text", + "content": "4.1. Experimental Setting" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": "Datasets. We evaluate on two standard Propheesee automotive datasets: Gen1 [4] and 1Mpx [22]. Gen1 contains " + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "inline_equation", + "content": "39+" + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": " hours of recordings at " + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "inline_equation", + "content": "304 \\times 240" + }, + { + "bbox": [ + 55, + 666, + 295, + 714 + ], + "type": "text", + "content": " resolution with manual annotations (1-4 Hz) for cars and pedestrians, presenting" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "content": "challenges with sparse labels and stationary objects. 1Mpx offers " + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "inline_equation", + "content": "14+" + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "content": " hours at higher resolution " + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "inline_equation", + "content": "(1280 \\times 720)" + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "content": " with denser automatic annotations " + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "inline_equation", + "content": "(60\\mathrm{Hz})" + }, + { + "bbox": [ + 313, + 265, + 555, + 361 + ], + "type": "text", + "content": " for cars, pedestrians, and two-wheelers. Both datasets are organized as 60-second clips with standard train/test splits. 1Mpx's higher resolution and frame rate enable evaluation on more complex scenes, while Gen1's annotation pattern tests robustness to sparse temporal supervision." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "text", + "content": "Event Representations. We adopt the well-established stacked 2D Histogram approach from RVT protocols [9], constructing event frames from each " + }, + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "inline_equation", + "content": "50\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "text", + "content": " time window. Each frame is subdivided into 10 equally time-spaced " + }, + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "inline_equation", + "content": "5\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 375, + 554, + 459 + ], + "type": "text", + "content": " intervals, creating a representation with sufficient temporal resolution for our adaptation framework while remaining compatible with standard image processing architectures." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 459, + 554, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 459, + 554, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 459, + 554, + 543 + ], + "type": "text", + "content": "For the Gen1 dataset (original size (240, 304)), we pad frames to (256, 320) to ensure divisibility by 32 as required by the feature pyramid architecture of RT-DETR. For 1Mpx, we downsize from (720, 1280) to (360, 640) using bilinear interpolation and pad to (384, 640). This maintains consistency with prior work while allowing for fair computational comparison." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "Two-stage Training Process. We train EvRT-DETR in two stages: (1) training the RT-DETR detector on individual event frames using standard practices (Adam optimizer; EMA of weights) then (2) freezing these parameters and training only the lightweight RNN modules on both random and sequential clips to develop temporal memory. The first stage establishes robust spatial feature extraction, while the second adds temporal processing with minimal parameter overhead. For temporal training, we maintain RNN memory continuity across sequential clips while resetting for random clips using different clip lengths for Gen1 (21 frames) and 1Mpx (10 frames). Full implementation details are provided in the supplementary material." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9817" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 103, + 69, + 507, + 292 + ], + "blocks": [ + { + "bbox": [ + 103, + 69, + 507, + 292 + ], + "lines": [ + { + "bbox": [ + 103, + 69, + 507, + 292 + ], + "spans": [ + { + "bbox": [ + 103, + 69, + 507, + 292 + ], + "type": "table", + "html": "
ModelGen11Mpx/Gen4Params (M)
mAP (%)Runtime (ms)mAP (%)Runtime (ms)
AsyNet [18]14.5---11.4
AEGNN [26]16.3---20.0
Spiking DenseNet [3]18.9---8.2
Events-RetinaNet [22]34.0-18.0-32.8
RED [22]40.016.743.039.324.1
AEC [19]47.010.6 (3.9)48.437.6 (13.8)46.5
RVT-B [9]47.210.247.411.918.5
GET-T [20]47.916.848.418.221.9
S5-ViT-B [34]47.48.247.29.617.5
ASTMNet [16]46.735.648.372.3>100
SAST-CB [21]48.2-48.757.5 (19.7)18.9
ERGO-12 [33]50.469.940.6100.059.6
RT-DETR-T (ours)46.06.444.18.620.1
RT-DETR-B (ours)47.610.545.214.942.8
EvRT-DETR-T (ours)52.38.449.912.534.4
EvRT-DETR-B (ours)52.712.750.118.857.1
", + "image_path": "69aba0c7e85d5dac0f7bb1de890ca83dd17138ac9c06703e2c6654fd597706e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 555, + 357 + ], + "type": "text", + "content": "Table 1. Performance Comparison on Event-based Camera Datasets. Our approach achieves state-of-the-art results on both Gen1 and 1Mpx benchmarks with significant improvements over previous methods. The first stage of the framework (RT-DETR-T/B) already achieves competitive performance using only image-based detection, while the proposed I2EvDet-adapted models (EvRT-DETR-T/B) surpass all existing approaches. Runtime measurements are reported in milliseconds on an NVIDIA T4 GPU. Values in parentheses represent original reported times on different hardware with T4-equivalent times calculated using FLOPS/throughput conversion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 376, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 376, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 376, + 295, + 533 + ], + "type": "text", + "content": "Evaluation Protocol. We evaluate EvRT-DETR using the standard COCO (Common Objects in Context) Mean Average Precision (mAP) measure [17] that allows direct comparison with existing methods. For consistency with previous works, we use the EBC-specific implementation of the COCO metrics provided by Propheesee's Automotive Dataset Toolbox [22, 23]. All inference times reported in Table 2 are standardized to NVIDIA T4 GPU performance for fair comparison. Inference times include model forward pass and post-processing but exclude event preprocessing, following standard reporting practices. When prior work reported only preprocessing-inclusive times, we use their reported values directly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 540, + 180, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 180, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 180, + 553 + ], + "type": "text", + "content": "4.2. Performance Analysis" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 593 + ], + "type": "text", + "content": "Table 1 presents a comprehensive comparison of our EvRT-DETR models against state-of-the-art approaches on the Gen1 and 1Mpx benchmarks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 701 + ], + "type": "text", + "content": "First, our baseline adaptation using two RT-DETR variants, RT-DETR-T (Tiny, with ResNet-18 backbone) and RT-DETR-B (Base, with ResNet-50 backbone), and trained on simple event frames achieves competitive performance (46.0% and 47.6% mAP on Gen1; 44.1% and 45.2% mAP on 1Mpx) without requiring specialized architectures or complex event representations. This demonstrates that modern mainstream object detectors can be effectively adapted to event data through appropriate training." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 713 + ], + "type": "text", + "content": "More importantly, our complete EvRT-DETR models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "content": "with temporal adaptation significantly outperform all existing methods, achieving new state-of-the-art results on both benchmarks. EvRT-DETR-B reaches " + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "inline_equation", + "content": "52.7\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "content": " mAP on Gen1 " + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "inline_equation", + "content": "(+2.3\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "content": " over the previous best) and " + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "inline_equation", + "content": "50.1\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "content": " mAP on 1Mpx " + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "inline_equation", + "content": "(+1.4\\%" + }, + { + "bbox": [ + 313, + 376, + 555, + 460 + ], + "type": "text", + "content": " over the previous best). Even our lighter EvRT-DETR-T model surpasses all existing approaches while maintaining faster inference times." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 460, + 555, + 543 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 543 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 543 + ], + "type": "text", + "content": "Figure 4 shows how the temporal adaptation module provides crucial capabilities for handling stationary objects - a fundamental challenge for event cameras. When objects are in motion, both RT-DETR and EvRT-DETR perform well, yet, when motion stops and event generation becomes minimal, EvRT-DETRs temporal memory enables consistent detection." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 544, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 556, + 617 + ], + "type": "text", + "content": "These performance improvements come with minimal computational overhead. While we focus primarily on detection accuracy in this work, the runtime measurements in Table 1 demonstrate that our approach maintains competitive inference speeds despite adding temporal processing capabilities." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 624, + 471, + 637 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 471, + 637 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 471, + 637 + ], + "type": "text", + "content": "4.3. Framework Generalizability" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "To demonstrate that our results are not specific to the RT-DETR architecture, we apply the I2EvDet framework to YOLOX [7]. YOLOX is a CNN-based detector that contrasts with RT-DETR's transformer architecture and has been used in other EBC applications [9, 34]. Table 2 shows that our temporal adaptation approach consistently" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "9818" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 80, + 70, + 269, + 135 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 269, + 135 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 269, + 135 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 269, + 135 + ], + "type": "table", + "html": "
DetectormAP (%)
OriginalI2EvDetImprovement
RT-DETR-T46.052.3+6.3
RT-DETR-B47.652.7+5.1
YOLOX-T36.042.4+6.4
YOLOX-X43.447.8+4.4
", + "image_path": "a4eb61df46bf09a353800c6020e6a0439b906190c0a1495bfe55dda1f012a60c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 207, + 296, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 207, + 296, + 292 + ], + "spans": [ + { + "bbox": [ + 55, + 207, + 296, + 292 + ], + "type": "text", + "content": "improves performance across different baseline architectures on the Gen1 dataset. The substantial improvements (4.4-6.4 mAP) across both RT-DETR and YOLOX variants validate that I2EvDet represents a general framework for adapting image-based detectors to temporal domains rather than an architecture-specific optimization. Additional details are provided in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 298, + 259, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 259, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 259, + 312 + ], + "type": "text", + "content": "4.4. Analysis of Adaptation Design Choices" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 388 + ], + "type": "text", + "content": "We systematically analyze key design choices in our temporal adaptation module to better understand the factors contributing to EvRT-DETR's performance. This analysis provides insights into efficient adaptation strategies for event-based object detection. Additional ablation studies are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 396, + 293, + 462 + ], + "blocks": [ + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 187 + ], + "type": "text", + "content": "Table 2. Generalizability of I2EvDet Framework Across Different Object Detection Architectures on the Gen1 Dataset. Our temporal adaptation consistently improves performance regardless of the base detector architecture." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 396, + 293, + 462 + ], + "lines": [ + { + "bbox": [ + 56, + 396, + 293, + 462 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 293, + 462 + ], + "type": "table", + "html": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)
RT-DETR-B (ours)47.676.349.5
EvRT-DETR-B (0,1,1)51.080.754.0
EvRT-DETR-B (1,0,1)52.281.355.1
EvRT-DETR-B (1,1,0)52.481.755.3
EvRT-DETR-B (1,1,1)52.782.056.0
", + "image_path": "593c5f52f2ba016805122c8eae52ecb02e17d14de48b351fae34f5300937cabe.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 543, + 223, + 555 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 543, + 223, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 543, + 223, + 555 + ], + "type": "text", + "content": "4.4.1. Multiscale Temporal Adaptation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "text", + "content": "A crucial aspect of our approach is the insertion of temporal modules at multiple scales in the feature hierarchy. Table 3 shows performance when selectively removing ConvLSTM cells from specific feature levels with the array notation " + }, + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "text", + "content": " indicating presence (1) or absence (0) at each scale " + }, + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{E}_3,\\mathcal{E}_4,\\mathcal{E}_5\\}" + }, + { + "bbox": [ + 55, + 558, + 295, + 630 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": "The results reveal a clear pattern: the lowest-level features " + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{E}_3)" + }, + { + "bbox": [ + 55, + 630, + 296, + 715 + ], + "type": "text", + "content": " benefit most significantly from temporal adaptation as performance drops substantially when this module is removed. The impact diminishes at higher feature levels, suggesting that temporal information is most valuable for fine-grained, lower-level features that capture detailed motion patterns. This insight can guide more efficient adapta" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 97 + ], + "type": "text", + "content": "tion designs, wherein computational resources are directed to where they provide maximum benefit." + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 108, + 552, + 155 + ], + "blocks": [ + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "text", + "content": "Table 3. Effect of Temporal Module Placement. Performance on the Gen1 dataset with ConvLSTM modules at different feature scales. The array notation " + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "inline_equation", + "content": "(x,y,z)" + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "text", + "content": " indicates presence (1) or absence (0) of temporal modules at scales " + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_3" + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_4" + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_5" + }, + { + "bbox": [ + 55, + 470, + 295, + 525 + ], + "type": "text", + "content": ", respectively. Low-level features benefit most from temporal adaptation." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 108, + 552, + 155 + ], + "lines": [ + { + "bbox": [ + 315, + 108, + 552, + 155 + ], + "spans": [ + { + "bbox": [ + 315, + 108, + 552, + 155 + ], + "type": "table", + "html": "
ModelmAP (%)\\( \\mathrm{mAP}_{50} \\) (%)\\( \\mathrm{mAP}_{75} \\) (%)\\( N_T \\) (M)
EvRT-DETR-B (M=64)52.181.354.92.3
EvRT-DETR-B (M=128)52.581.755.65.4
EvRT-DETR-B (M=256)52.782.056.014.4
EvRT-DETR-B (M=512)52.981.956.042.9
", + "image_path": "101f0915dd8319a68ace1df8b6b9f675e734f17884f15ffb31cca604262e10d6.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 163, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 555, + 208 + ], + "type": "text", + "content": "Table 4. Impact of Temporal Module Capacity. Performance on Gen1 dataset with varying ConvLSTM hidden dimensions (M). The rightmost column shows the number of trainable parameters (in millions) in the temporal adaptation module." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 231, + 494, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 494, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 494, + 243 + ], + "type": "text", + "content": "4.4.2. Parameter Efficiency in Adaptation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 247, + 554, + 306 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 554, + 306 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 554, + 306 + ], + "type": "text", + "content": "Inspired by parameter-efficient techniques such as LoRA (Low-Rank Adaptation) [15], we investigate the minimal temporal module capacity needed for effective adaptation. Table 4 shows EvRT-DETR-B performance with varying hidden dimensions in the ConvLSTM modules." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": "Remarkably, reducing hidden dimensions to 64 features (a " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "4\\mathrm{x}" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " reduction from our baseline) results in only modest performance degradation " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "(-0.6\\% \\mathrm{mAP})" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " while requiring just " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "2.3\\mathrm{M}" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " additional parameters - merely " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "5.4\\%" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " of the base RT-DETR-B model's " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "42.8\\mathrm{M}" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " parameters. This configuration closely resembles LoRA-style adaptation, achieving substantial performance gains with minimal parameter overhead. Even our most parameter-efficient adaptation " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "(\\mathbf{M} = 64)" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " outperforms all previous state-of-the-art approaches, demonstrating that effective temporal adaptation does not require extensive architectural modifications. Increasing the hidden dimension to 512 features " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "(42.9\\mathrm{M}" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " parameters, nearly doubling the model size) yields only a marginal " + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "inline_equation", + "content": "0.2\\% \\mathrm{mAP}" + }, + { + "bbox": [ + 313, + 308, + 555, + 488 + ], + "type": "text", + "content": " improvement, indicating diminishing returns." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 501, + 391, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 391, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 391, + 513 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "This work introduces I2EvDet, a framework that efficiently transforms image-based detectors into video-capable models using minimal modifications. Applied to event-based cameras, our approach demonstrates that mainstream object detection architectures can be effectively adapted to temporal domains without specialized redesign. RT-DETR trained on simple event representations achieves performance comparable to specialized methods, while our I2EvDet-adapted model EvRT-DETR achieves state-of-the-art results on Gen1 " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "(+2.3\\mathrm{mAP})" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " and 1Mpx " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "(+1.4\\mathrm{mAP})" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " benchmarks. Ablation studies provide insights about optimal temporal module design and parameter efficiency. These results suggest that bridging mainstream computer vision and specialized domains can be achieved through targeted adaptation rather than complete architectural redesign, potentially accelerating progress across diverse temporal visual domains." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "9819" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 200 + ], + "type": "text", + "content": "The authors extend their sincere gratitude to Yuewei Lin (Brookhaven Lab) for his valuable feedback and suggestions. This work presented in this paper was funded by the U.S. Department of Energy (DOE), National Nuclear Security Administration, Office of Defense Nuclear Nonproliferation Research and Development. The manuscript has been authored by Brookhaven National Laboratory managed by Brookhaven Science Associates, LLC under the Contract No. DE-SC0012704 with the U.S. Department of Energy." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 115, + 220 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 228, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 228, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 294, + 293 + ], + "type": "text", + "content": "[1] Thomas Bachlechner, Bodhisattwa Prasad Majumder, Huanru Henry Mao, Gary Cottrell, and Julian J. McAuley. ReZero is all you need: fast convergence at large depth. In Proceedings of the Thirty-Seventh Conference on Uncertainty in Artificial Intelligence, UAI 2021, Virtual Event, 27-30 July 2021, pages 1352-1361. AUAI Press, 2021. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 294, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 294, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 294, + 296, + 360 + ], + "type": "text", + "content": "[2] Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, and Sergey Zagoruyko. End-to-end object detection with transformers. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part I, pages 213-229. Springer, 2020. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 361, + 296, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 361, + 296, + 414 + ], + "spans": [ + { + "bbox": [ + 61, + 361, + 296, + 414 + ], + "type": "text", + "content": "[3] Loïc Cordone, Benoit Miramond, and Philippe Thiérison. Object detection with spiking neural networks on automotive event data. In International Joint Conference on Neural Networks, IJCNN 2022, Padua, Italy, July 18-23, 2022, pages 1-8. IEEE, 2022. 3, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 416, + 296, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 416, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 62, + 416, + 296, + 458 + ], + "type": "text", + "content": "[4] Pierre de Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. CoRR, abs/2001.08499, 2020. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 460, + 296, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 460, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 62, + 460, + 296, + 525 + ], + "type": "text", + "content": "[5] Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, and Luke Zettlemoyer. Qlora: Efficient finetuning of quantized llms. In Advances in Neural Information Processing Systems 36: Annual Conference on Neural Information Processing Systems 2023, NeurIPS 2023, New Orleans, LA, USA, December 10 - 16, 2023, 2023. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 526, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 526, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 526, + 296, + 590 + ], + "type": "text", + "content": "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jorg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3, 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 593, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 593, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 593, + 296, + 624 + ], + "type": "text", + "content": "[7] Zheng Ge, Songtao Liu, Feng Wang, Zeming Li, and Jian Sun. YOLOX: exceeding YOLO series in 2021. CoRR, abs/2107.08430, 2021. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "spans": [ + { + "bbox": [ + 62, + 625, + 296, + 658 + ], + "type": "text", + "content": "[8] Daniel Gehrig and Davide Scaramuzza. Pushing the limits of asynchronous graph-based object detection with event cameras. CoRR, abs/2211.12324, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 659, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 659, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 659, + 296, + 713 + ], + "type": "text", + "content": "[9] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2023, Vancouver, BC, Canada, June 17-24, 2023, pages 13884-13893. IEEE, 2023. 3, 4, 6, 7" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 139 + ], + "type": "text", + "content": "[10] Ross B. Girshick, Jeff Donahue, Trevor Darrell, and Jitendra Malik. Rich feature hierarchies for accurate object detection and semantic segmentation. In 2014 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2014, Columbus, OH, USA, June 23-28, 2014, pages 580-587. IEEE Computer Society, 2014. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 554, + 183 + ], + "type": "text", + "content": "[11] Zeyu Han, Chao Gao, Jinyang Liu, Jeff Zhang, and Sai Qian Zhang. Parameter-efficient fine-tuning for large models: A comprehensive survey. Trans. Mach. Learn. Res., 2024, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 555, + 251 + ], + "type": "text", + "content": "[12] Junxian He, Chunting Zhou, Xuezhe Ma, Taylor Berg-Kirkpatrick, and Graham Neubig. Towards a unified view of parameter-efficient transfer learning. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 4, 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 317, + 255, + 554, + 308 + ], + "type": "text", + "content": "[13] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 770-778. IEEE Computer Society, 2016. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 310, + 554, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 310, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 310, + 554, + 386 + ], + "type": "text", + "content": "[14] Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin de Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for NLP. In Proceedings of the 36th International Conference on Machine Learning, ICML 2019, 9-15 June 2019, Long Beach, California, USA, pages 2790-2799. PMLR, 2019. 2, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 389, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 389, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 317, + 389, + 554, + 453 + ], + "type": "text", + "content": "[15] Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: low-rank adaptation of large language models. In The Tenth International Conference on Learning Representations, ICLR 2022, Virtual Event, April 25-29, 2022. OpenReview.net, 2022. 2, 4, 5, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 554, + 499 + ], + "type": "text", + "content": "[16] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE Trans. Image Process., 31:2975-2987, 2022. 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 502, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 502, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 317, + 502, + 554, + 578 + ], + "type": "text", + "content": "[17] Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dollár, and C. Lawrence Zitnick. Microsoft COCO: common objects in context. In Computer Vision - ECCV 2014 - 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V, pages 740-755. Springer, 2014. 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 580, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 580, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 580, + 554, + 635 + ], + "type": "text", + "content": "[18] Nico Messikommer, Daniel Gehrig, Antonio Loquercio, and Davide Scaramuzza. Event-based asynchronous sparse convolutional networks. In Computer Vision - ECCV 2020 - 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VIII, pages 415-431. Springer, 2020. 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 636, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 554, + 713 + ], + "type": "text", + "content": "[19] Yansong Peng, Yueyi Zhang, Peilin Xiao, Xiaoyan Sun, and Feng Wu. Better and faster: Adaptive event conversion for event-based object detection. In Thirty-Seventh AAAI Conference on Artificial Intelligence, AAAI 2023, Thirty-Fifth Conference on Innovative Applications of Artificial Intelligence, IAAI 2023, Thirteenth Symposium on Educational Advances in Artificial Intelligence, EAAI 2023, Washington," + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "9820" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 11, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "DC, USA, February 7-14, 2023, pages 2056-2064. AAAI Press, 2023. 3, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 151 + ], + "type": "text", + "content": "[20] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. GET: group event transformer for event-based vision. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 6015-6025. IEEE, 2023. 2, 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 217 + ], + "type": "text", + "content": "[21] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16794-16804. IEEE, 2024. 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 285 + ], + "type": "text", + "content": "[22] Etienne Perot, Pierre de Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual, 2020. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 287, + 294, + 330 + ], + "type": "text", + "content": "[23] Propheesee. Propheesee's automotive dataset toolbox. https://github.com/propheesee-ai/propheesee-automotive-dataset-toolbox, 2024.7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 333, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 397 + ], + "type": "text", + "content": "[24] Joseph Redmon, Santosh Kumar Divvala, Ross B. Girshick, and Ali Farhadi. You only look once: Unified, real-time object detection. In 2016 IEEE Conference on Computer Vision and Pattern Recognition, CVPR 2016, Las Vegas, NV, USA, June 27-30, 2016, pages 779-788. IEEE Computer Society, 2016. 2, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 466 + ], + "type": "text", + "content": "[25] Shaoqing Ren, Kaiming He, Ross B. Girshick, and Jian Sun. Faster R-CNN: towards real-time object detection with region proposal networks. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 91-99, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 468, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 522 + ], + "type": "text", + "content": "[26] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. AEGNN: asynchronous event-based graph neural networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2022, New Orleans, LA, USA, June 18-24, 2022, pages 12361-12371. IEEE, 2022. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 524, + 294, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 524, + 294, + 601 + ], + "spans": [ + { + "bbox": [ + 56, + 524, + 294, + 601 + ], + "type": "text", + "content": "[27] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems 28: Annual Conference on Neural Information Processing Systems 2015, December 7-12, 2015, Montreal, Quebec, Canada, pages 802-810, 2015. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 668 + ], + "type": "text", + "content": "[28] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. In Advances in Neural Information Processing Systems 30: Annual Conference on Neural Information Processing Systems 2017, December 4-9, 2017, Long Beach, CA, USA, pages 5998-6008, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[29] Hao Zhang, Feng Li, Shilong Liu, Lei Zhang, Hang Su, Jun Zhu, Lionel M. Ni, and Heung-Yeung Shum. DINO: DETR with improved denoising anchor boxes for end-to-end object detection. In The Eleventh International Conference on" + } + ] + } + ], + "index": 10 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 406 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Learning Representations, ICLR 2023, Kigali, Rwanda, May 1-5, 2023. OpenReview.net, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 162 + ], + "type": "text", + "content": "[30] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. DETRs beat YOLOs on Real-time object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 16965-16974. IEEE, 2024. 2, 4, 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 228 + ], + "type": "text", + "content": "[31] Zihan Zhong, Zhiqiang Tang, Tong He, Haoyang Fang, and Chun Yuan. Convolution meets lora: Parameter efficient finetuning for segment anything model. In The Twelfth International Conference on Learning Representations, ICLR 2024, Vienna, Austria, May 7-11, 2024. OpenReview.net, 2024. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "spans": [ + { + "bbox": [ + 316, + 229, + 553, + 284 + ], + "type": "text", + "content": "[32] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable DETR: deformable transformers for end-to-end object detection. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021. OpenReview.net, 2021. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 350 + ], + "type": "text", + "content": "[33] Nikola Zubic, Daniel Gehrig, Mathias Gehrig, and Davide Scaramuzza. From chaos comes order: Ordering event representations for object recognition and detection. In IEEE/CVF International Conference on Computer Vision, ICCV 2023, Paris, France, October 1-6, 2023, pages 12800-12810. IEEE, 2023. 2, 3, 7" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 353, + 553, + 406 + ], + "type": "text", + "content": "[34] Nikola Zubic, Mathias Gehrig, and Davide Scaramuzza. State space models for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, CVPR 2024, Seattle, WA, USA, June 16-22, 2024, pages 5819-5828. IEEE, 2024. 3, 7" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 314, + 757 + ], + "type": "text", + "content": "9821" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_content_list.json b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..7e4502a466c8b8e209807378c329fa7d91f9198a --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_content_list.json @@ -0,0 +1,1722 @@ +[ + { + "type": "text", + "text": "EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images", + "text_level": 1, + "bbox": [ + 106, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wangbo Yu $^{1,2*}$ , Chaoran Feng $^{1*}$ , Jianing Li $^{3}$ , Jiye Tang $^{4}$ , Jiashu Yang $^{5}$ , Zhenyu Tang $^{1}$ , Meng Cao $^{6}$ , Xu Jia $^{5}$ , Yuchao Yang $^{1}$ , Li Yuan $^{1,2\\dagger}$ and Yonghong Tian $^{1,2\\dagger}$", + "bbox": [ + 106, + 179, + 888, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1Shenzhen Graduate School, Peking University, 2Peng Cheng Laboratory, 3School of Computer and Science, Peking University \n4University of Science and Technology of China, 5Dalian University of Technology, 6MBZUAI {wbyu, chaoran.feng}@stu.pku.edu.cn, {yuanli-ece, yhtian}@pu.edu.cn", + "bbox": [ + 117, + 234, + 879, + 287 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2152ac3dce439fe973cb9b3ed60d1974db601256af66caec9056d6bd2e562052.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 311, + 272, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/24e6ee784102928c42531161649139aedb027a483b80747d64d122673ca06bd3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 272, + 311, + 450, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/18d76e1fc6782a46660d4a8bcaa50afb52f8d981e1aa95d910fd71f7c55782b4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 311, + 630, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/55c6a84091b04e4bdf398e896c20b4edcf348420a70a152111f4cb5fd9a2a731.jpg", + "image_caption": [ + "Figure 1. EvaGaussians integrates frame-based cameras and event cameras to reconstruct sharp 3D-GS from motion-blurred images. It achieves higher-quality novel view synthesis compared to previous methods." + ], + "image_footnote": [], + "bbox": [ + 94, + 407, + 272, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa21a3ac57b6ab17d068127ee646ece87416a594ba078a2c2658559e1a02001a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 274, + 407, + 450, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a0c2082c7dba596bd96e1068e01cb9d18deccdf200263997b911f35dee4cd5bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 452, + 407, + 630, + 501 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/474585339f000993377c94f4c229b1fc1ca6be93d9212381ef00bfb036f6ebcd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 311, + 900, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/01cc40e305b7342c95bf0557567eca0f29e05280db0fa380ad75d0aad95ed5f0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 635, + 443, + 900, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 556, + 326, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D Gaussian Splatting (3D-GS) has demonstrated exceptional capabilities in synthesizing novel views of 3D scenes. However, its training is heavily reliant on high-quality images and precise camera poses. Meeting these criteria can be challenging in non-ideal real-world conditions, where motion-blurred images frequently occur due to high-speed camera movements. To address these challenges, we introduce Event Stream Assisted Gaussian Splatting (EvaGaus-sians), a novel approach that harnesses event streams captured by event cameras to facilitate the learning of high-quality 3D-GS from motion-blurred images. Capitalizing on the high temporal resolution and dynamic range offered by event streams, we seamlessly integrate them into the initialization and optimization of 3D-GS, thereby enhancing the acquisition of high-fidelity novel views with intricate texture details. We also contribute two novel datasets comprising RGB frames, event streams, and corresponding camera parameters, featuring a wide variety of scenes", + "bbox": [ + 88, + 588, + 485, + 861 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "and various camera motions. The comparison results reveal that our approach not only excels in generating high-fidelity novel views, but also offers improved training and inference efficiency. Video and code are available at the project page.", + "bbox": [ + 511, + 556, + 906, + 619 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 648, + 643, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Novel view synthesis from 2D image collections has presented a persistent challenge within the field of computer vision and computer graphics. This task stands as a fundamental component in various vision applications, such as virtual reality [17, 60], robotics navigation [42, 62, 74], scene understanding [15, 24, 25], and many others, thereby prompting significant research efforts over the last decades. Amid pioneering works, 3D Gaussian Splatting (3D-GS) [14] achieves notable success in generating high-fidelity novel views. However, the optimization of 3D-GS heavily relies on accurate camera poses and point cloud initialization produced by COLMAP [44], which necessitates high-quality images without blurring. Fulfilling such conditions can be challenging in real-world situations. For example, in UAVs and robotics, rapid camera movement is", + "bbox": [ + 509, + 672, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*These authors contributed equally to this work.", + "bbox": [ + 107, + 875, + 364, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24780", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "common when capturing images or recording videos, which often result in significant motion blur.", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies have demonstrated the significant potential of event-based cameras in alleviating motion blur in images captured by conventional frame-based cameras. [13, 23, 34]. Serving as an innovative bio-inspired visual sensor, event cameras asynchronously report the logarithmic intensity changes of each pixel captured, and can record higher temporal resolution and dynamic range data in contrast to conventional cameras. The continuous event streams naturally align with the frame-rate agnostic rendering property of 3D-GS, offering sharper scene geometry information than frame-based cameras, making it highly promising in facilitating 3D-GS training from fast camera motion.", + "bbox": [ + 89, + 128, + 483, + 309 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by this, we propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first method that seamlessly incorporates continuous event streams captured by event cameras to enhance the learning of high-quality 3D-GS from motion-blurred images. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Furthermore, to tackle the geometric ambiguity caused by blurry images, we introduce two event-assisted geometry regularization terms to stabilize the geometry of 3D-GS. Through optimizing the 3D-GS in a progressive manner, our method can recover a high-quality 3D-GS that facilitates the real-time generation of high-fidelity novel views. Finally, to comprehensively evaluate novel view synthesis with hybrid frame- and event-based cameras, we contribute a synthetic dataset and a real-world dataset, both containing diverse scenes with varying scales and camera motions, setting a benchmark for related researches. To summarize, our contributions can be delineated as follows:", + "bbox": [ + 91, + 316, + 483, + 662 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first framework designed to reconstruct high-quality 3D-GS from motion-blurred images with the assistance of event cameras. It not only recovers intricate details from the motion-blurred images but also enables high-fidelity novel view synthesis.", + "- We contribute a synthetic dataset and a real-world dataset for evaluating novel view synthesis with hybrid frame-and event-based cameras. Both datasets contain diverse scenes with varying scales and camera motions.", + "- We conduct a comprehensive evaluation of the proposed method and compare it with several strong baselines. As shown in Figure. 1 and Figure. 2, our method not only excels in generating higher-fidelity novel views but also enables improved training and inference efficiency." + ], + "bbox": [ + 89, + 671, + 483, + 898 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2763364a299093d5f3e4b8de97327ad8b5e3773cd95150f44f71090338857d6a.jpg", + "image_caption": [ + "Figure 2. Performance, rendering speed, and training time (indicated by the area of each circle) comparison. EvaGaussians achieves better results while maintaining real-time rendering and training efficiency." + ], + "image_footnote": [], + "bbox": [ + 531, + 88, + 883, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 420, + 663, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Reconstructing 3D Scene from Blurry Images", + "text_level": 1, + "bbox": [ + 511, + 446, + 898, + 463 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Reconstructing a high-quality 3D Scene typically requires high-fidelity, sharp images as supervision. However, motion-blurred images often occur in real world scenarios, thus hindering accurate reconstruction of 3D scenes. Several studies have been proposed to address this issue. For example, Deblur-NeRF [28], DP-NeRF [18] and PDRF [37] attempted to learn a blur formation kernel to model the image blurring process. BAD-NeRF [56] further physically modeled the blurry images formation process, and adopted a bundle-adjustment strategy to jointly optimize NeRF parameters and the camera poses during the exposure time. These NeRF-based methods lacked real-time rendering capabilities and suffered from extended training times. With the rapid advancement of 3D-GS, concurrent works like BAD-Gaussians [70] and BAGS [38] proposed to utilize 3D-GS as representation to achieve deblurring reconstruction. Although they achieved real-time rendering and faster convergence compared with prior works, they still struggled to handle severely blurred images in which COLMAP [44] will fail to produce the initial point clouds and camera calibration.", + "bbox": [ + 511, + 470, + 906, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Reconstructing 3D Scene from Event Streams", + "text_level": 1, + "bbox": [ + 511, + 801, + 898, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the exceptional properties offered by event cameras [1, 2, 6, 7, 20-22, 29, 33, 35, 36, 45, 47-49, 51, 54, 58, 61, 68, 72, 75], several studies attempted to reconstruct 3D scenes from event streams captured by event cameras, particularly in low-light conditions with fast camera mo", + "bbox": [ + 511, + 825, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24781", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tion [8, 11, 12, 27, 43, 55, 59, 61, 63-66, 69, 71]. These methods were mostly designed to recover scene geometry from event streams, and are not typically designed for reconstructing 3D Scene from blurry images. Recently, E-NeRF [16], $\\mathrm{E}^2\\mathrm{NeRF}$ [39], EBAD-NeRF [40], BENeRF [19] and EvDNeRF [3] proposed to jointly utilize event streams captured by event cameras and motion-blurred images captured by standard frame-based cameras to reconstruct a NeRF representation. In contrast to frame-based deblurring methods, they are better at handling motion blur due to the introduction of event information. However, these NeRF-based methods suffer from long training and inference times, which limit their further application.", + "bbox": [ + 89, + 90, + 483, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 300, + 181, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 324, + 220, + 340 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event camera is a type of bio-inspired sensor that can asynchronously record intensity changes [7]. In contrast to conventional cameras that are restricted to sequentially produce frames at a fixed frame rate, event cameras asynchronously trigger events in each pixel when their intensity change exceeds a constant threshold, featuring properties such as low latency and high dynamic range. Formally, let $\\mathbf{I}_{xy}(t)$ denote the instantaneous intensity at pixel coordinate $(x,y)$ at time $t$ , and $\\mathbf{L}_{xy}(t)$ denotes its logarithm. An event $p = \\pm 1$ will be triggered whenever the change of $\\mathbf{L}_{xy}(t)$ surpasses the threshold $c$ , where the polarity represents the direction (increase or decrease) of changes. Let $\\delta_{t_0}(t)$ be the impulse function at time $t_0$ with a unit integral, the event can therefore be expressed as a continuous-time signal $\\mathbf{e}_{xy}(t) = p\\delta_{t_0}(t)$ , where $t_0$ signifies the time at which the event occurs. Then, the proportional intensity change during a time interval $[s,t]$ can be computed as the integral of events that occurred between times $s$ and $t$ , expressed as $\\mathbf{E}_{xy}(t) = \\int_s^t\\mathbf{e}_{xy}(h)dh$ . Given that each pixel can be treated separately in the event camera, the subscripts can be omitted:", + "bbox": [ + 89, + 345, + 483, + 664 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {E} (t) = \\int_ {s} ^ {t} \\mathbf {e} (h) d h. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 217, + 671, + 482, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We can then represent the logarithmic intensity change as: $\\mathbf{L}(t) - \\mathbf{L}(s) = c\\mathbf{E}(t)$ , rewrite as $\\mathbf{L}(t) = \\mathbf{L}(s) + c\\mathbf{E}(t)$ , and subsequently obtain the actual intensity change:", + "bbox": [ + 89, + 713, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {I} (t) = \\mathbf {I} (s) \\cdot \\exp (c \\mathbf {E} (t)). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 768, + 482, + 785 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, when an image $\\mathbf{I}(s)$ is captured at time $s$ , and the event stream is recorded during the time interval $[s,t]$ , the image $\\mathbf{I}(t)$ can be obtained by warping $\\mathbf{I}(s)$ using Eq. 2.", + "bbox": [ + 89, + 794, + 483, + 842 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Event-assisted Initialization", + "text_level": 1, + "bbox": [ + 89, + 848, + 339, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The optimization of 3D-GS requires camera calibration and point cloud initialization using COLMAP [44]. However,", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "as shown in Table.4 in the supplementary, this process can fail when dealing with images that have significant motion blur. Motion-blurred images are resulted from camera movements during the exposure time, which can be mathematically represented as:", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\mathbf {I} (t) d t, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 176, + 906, + 214 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{B}$ denotes a captured blurry image, which is equivalent to averaging the instantaneous latent images $\\mathbf{I}(t)$ during the exposure time $[s - \\tau /2,s + \\tau /2]$ .", + "bbox": [ + 511, + 223, + 903, + 268 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To obtain initial camera poses and point clouds for 3D-GS optimization, we first preprocess the motion-blurred images using the Event-based Double Integral (EDI) [34] model, which can be derived through substituting Eq. 2 into Eq. 3:", + "bbox": [ + 511, + 268, + 906, + 344 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {B} = \\mathbf {I} (s) \\cdot \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\exp (c \\mathbf {E} (t)) d t. \\qquad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 584, + 354, + 906, + 392 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the predefined threshold $c$ , a blurry image $\\mathbf{B}$ , and the recorded event stream $\\mathbf{E}(t)$ , the EDI model (Eq. 4) allows the derivation of $\\mathbf{I}(s)$ , following which the latent image $\\mathbf{I}(t)$ at any moment within the exposure time can be estimated through Eq. 2. As shown in Figure. 3, given a total of $K$ blurry images $\\{\\mathbf{B}^j\\}_{j=1}^K$ , for each of them, we uniformly sample $n$ time stamps during their exposure time to obtain a series of EDI-processed latent images rich in texture features, denoted as $\\{\\mathbf{I}_i\\}_{i=1}^n$ , then obtain their poses $\\{\\mathbf{P}_i\\}_{i=1}^n$ and the initial point cloud of the scene using COLMAP [44].", + "bbox": [ + 511, + 401, + 905, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After initialization, a straightforward approach to optimizing the 3D-GS is to use the EDI-processed latent images and their camera poses as supervision. However, although these latent images provide more texture features than the original blurry image, they still do not fully recover the ideal latent image and exhibit relatively low visual quality, which also introduces inaccuracies into the camera poses, thereby leading to unsatisfactory optimization results. To more robustly recover a sharp 3D-GS from motion-blurred images, we propose to harness the advantages of event streams and seamlessly integrate them into the optimization process of 3D-GS.", + "bbox": [ + 511, + 551, + 906, + 734 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Event-assisted Bundle-Adjustment", + "text_level": 1, + "bbox": [ + 511, + 742, + 816, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As introduced in Eq. 3, during the exposure time, a motion-blurred image can be decomposed into a series of latent images along a specific camera trajectory, which can be roughly approximated by the EDI-produced camera poses $\\{\\mathbf{P}_i\\}_{i=1}^n$ according to Eq. 4. Motivated by this, we jointly optimize these camera poses and the 3D-GS attributes in a bundle-adjustment manner [56] to simultaneously recover the blur-formation camera trajectories and a sharp 3D-GS. As shown in Figure. 3, we add each of the EDI-produced", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24782", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/416f929965d0e231266637a7b3841749bd7acc39b560a0de43f9ee4dcccb8017.jpg", + "image_caption": [ + "Figure 3. Overview of EvaGaussians. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Leveraging the continuously recorded event streams, we introduce two event-assisted geometry regularization terms that can function beyond the exposure time to stabilize the geometry of 3D-GS. Once trained, it can not only recover intricate details from motion-blurred images but also enable high-fidelity novel view synthesis." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "camera poses a learnable offset $\\{\\mathbf{d}_i\\}_{i = 1}^n$ as correction parameters, resulting a learnable camera trajectory $\\{\\widetilde{\\mathbf{P}}_i\\}_{i = 1}^n$ where $\\widetilde{\\mathbf{P}}_i = \\mathbf{P}_i + \\mathbf{d}_i$ . In each training iteration, we simultaneously render $n$ images $\\{\\widetilde{\\mathbf{I}}_i\\}_{i = 1}^n$ from the 3D-GS along the camera trajectory of the blurry view, and then simulate the formation of the motion-blurred images using a discrete approximation of Eq 3, denoted as $\\widetilde{\\mathbf{B}} = \\frac{1}{n}\\sum_{i = 1}^{n}\\widetilde{\\mathbf{I}}_i$ . Consequently, for a total of $K$ real-captured blurry images $\\{\\mathbf{B}^j\\}_{j = 1}^K$ , we can obtain their simulated versions $\\{\\widetilde{\\mathbf{B}}^j\\}_{j = 1}^K$ through each corresponding learnable camera trajectory.", + "bbox": [ + 89, + 438, + 480, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Blur Reconstruction Loss. With the simulated blurry images, we use the captured real blurry images $\\{\\mathbf{B}^j\\}_{j=1}^K$ to serve as image level supervision. Specifically, for each blurry image $\\mathbf{B}^j$ and its simulated version $\\tilde{\\mathbf{B}}^j$ , we employ a blur reconstruction loss to minimize their photometric error, expressed as", + "bbox": [ + 89, + 593, + 482, + 684 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {b l u r} = \\left(1 - \\lambda_ {1}\\right) \\cdot \\| \\mathbf {B} ^ {j} - \\widetilde {\\mathbf {B}} ^ {j} \\| _ {1} + \\lambda_ {1} \\cdot \\mathrm {D - S S I M} \\left(\\mathbf {B} ^ {j}, \\widetilde {\\mathbf {B}} ^ {j}\\right). (5)\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 691, + 482, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The formulation of blur reconstruction loss is the same as in the original 3D-GS [14], it differs in utilizing blurry images as supervision and jointly optimizing the 3D-GS attributes and the camera trajectories, thus facilitating an initial deblurring reconstruction of 3D-GS.", + "bbox": [ + 89, + 719, + 482, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Event Reconstruction Loss. Leveraging the abundant high-frequency information offered by the event streams, we further adopt an event reconstruction loss to aid in 3D-GS optimization. Specifically, we uniformly divide the exposure time into $m = n - 1$ intervals, each with a duration of $\\frac{\\tau}{m}$ . Subsequently, we integrate the recorded event stream along these time intervals using Eq. 1, resulting in", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$m$ event maps $\\{\\mathbf{E}_i\\}_{i=1}^m$ to serve as event level supervision. During training, for the $j$ -th blurry view, we convert the rendered image sequence $\\{\\widetilde{\\mathbf{I}}_i\\}_{i=1}^n$ on the camera trajectory into event maps $\\{\\widetilde{\\mathbf{E}}_i\\}_{i=1}^m$ , using a differentiable event simulator [10, 41], and constrain the discrepancies between the simulated event maps and the ground truth event maps, expressed as:", + "bbox": [ + 511, + 438, + 906, + 546 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e v e n t}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\| \\mathbf {E} _ {i} - \\widetilde {\\mathbf {E}} _ {i} \\| _ {1}. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 558, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The event reconstruction loss further aids in recovering a sharp 3D-GS with improved texture details.", + "bbox": [ + 511, + 609, + 906, + 640 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Event-assisted Geometry Regularization", + "text_level": 1, + "bbox": [ + 511, + 650, + 859, + 667 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The blurry color images are captured only during the exposure time and are much sparser than the event stream. Relying on such low-quality image-level supervision may cause the 3D-GS to overfit on the training images, resulting in significant floaters and inferior geometry. Leveraging Eq. 1 and Eq. 4, given the continuously recorded event streams $\\mathbf{E}(t)$ , we can derive continuous grayscale intensity images $\\mathbf{G}(t)$ that are rich in geometric information and can function beyond the exposure time. Motivated by this, we further propose two event-assisted geometry regularization terms to aid in 3D-GS training.", + "bbox": [ + 511, + 672, + 906, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intensity Reconstruction Loss. During training, we randomly sample continuous time $t$ between the interval of two adjacent blurry image, and derive the grayscale intensity image $\\mathbf{G}(t)$ using Eq. 2. We then minimize the difference", + "bbox": [ + 511, + 840, + 908, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24783", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "between it and the rendered intensity image from 3D-GS, expressed as:", + "bbox": [ + 89, + 90, + 482, + 121 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i n t} = (1 - \\lambda_ {2}) \\cdot \\| \\mathbf {G} (t) - \\widetilde {\\mathbf {G}} (t) \\| _ {1} + \\lambda_ {2} \\cdot D - S S I M (\\mathbf {G} (t), \\widetilde {\\mathbf {G}} (t)), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 130, + 483, + 162 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{G}}(t)$ is converted from the colored render result.", + "bbox": [ + 89, + 162, + 457, + 178 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Intensity-aware Depth Regularization Loss. As shown in Figure. 3, to further improve the geometry of 3D-GS, inspired by [4, 9], we adopt an intensity-aware depth regularization loss during training, defined as:", + "bbox": [ + 89, + 179, + 483, + 239 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {x, y} \\left(\\left| \\partial_ {x} \\widetilde {\\mathbf {D}} _ {x y} (t) \\right| e ^ {- \\beta \\left| \\partial_ {x} \\mathbf {G} _ {x y} (t) \\right|} \\right. \\tag {8} \\\\ + | \\partial_ {y} \\widetilde {\\mathbf {D}} _ {x y} (t) | e ^ {- \\beta | \\partial_ {y} \\mathbf {G} _ {x y} (t) |}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 250, + 482, + 309 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\widetilde{\\mathbf{D}}(t)$ is the rendered depth map, $(x,y)$ denotes the pixel location, $N$ is the total number of pixels, and $\\beta$ is set to 2 in our experiments. The horizontal and vertical gradients are calculated by applying convolution operations with $5 \\times 5$ Sobel kernels [52]. This regularization is founded on the observation that depth transitions in an image often correspond to changes in intensity. Therefore, it ensures that the spatial variation of depth closely matches that of the intensity image, thereby reducing geometric artifacts at object boundaries.", + "bbox": [ + 89, + 320, + 483, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The total loss function is the combination of the above losses, defined as:", + "bbox": [ + 89, + 473, + 482, + 502 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {b l u r}} \\mathcal {L} _ {\\text {b l u r}} + \\lambda_ {\\text {e v e n t}} \\mathcal {L} _ {\\text {e v e n t}} \\tag {9} \\\\ + \\lambda_ {i n t} \\mathcal {L} _ {i n t} + \\lambda_ {d e p t h} \\mathcal {L} _ {d e p t h}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 160, + 513, + 482, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 571, + 223, + 589 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this section, we first provide the implementation details of our method in 4.1, followed by an illustration of the introduced dataset in 4.2, and the experimental settings in 4.3. The comparison results and ablation study are presented in 4.4 and 4.5. For more detailed experiments.", + "bbox": [ + 89, + 597, + 483, + 674 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 681, + 307, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Progressive Training. We implemented EvaGaussians based on the official code of 3D-GS [14]. The training process spans 50,000 iterations, with an event reconstruction loss introduced after a 3,000-iteration warmup and we omit the densification process to streamline and simplify the subsequent optimization. Additionally, we adopt a coarse-to-fine training strategy, starting with rendering at a low resolution (0.3× downsampling in the early 30% iterations) and progressively increasing the size of the rendered views to full resolution. All experiments were conducted using a single NVIDIA RTX 4090 GPU.", + "bbox": [ + 89, + 704, + 482, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Hyperparameter Setting. During the training process, we set $\\lambda_{blur} = 1.0$ , $\\lambda_{depth} = 1.0e^{-2}$ , $\\lambda_{event} = 5.0e^{-3}$ and", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$\\lambda_{int} = 1.0e^{-3}$ for the loss function, and used $n = 9$ for the number of poses to be optimized during the exposure time. In implementing the loss $\\mathcal{L}_{event}$ , we configured the positive threshold as $c_{pos} = 0.25$ and the negative threshold as $c_{neg} = 0.25$ for synthetic scenes, and set $c_{pos} = 0.197$ and $c_{neg} = 0.241$ for real scenes.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Datasets", + "text_level": 1, + "bbox": [ + 513, + 193, + 614, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To enable a comprehensive evaluation, we introduce two novel datasets, with an overview provided below. Additionally, we present a detailed dataset introduction in the supplementary and will release the dataset construction pipeline for future research.", + "bbox": [ + 511, + 215, + 905, + 290 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EvaGaussians-Blender Dataset. We construct a synthetic dataset covering a variety of scene scales, coupling with diverse camera trajectories and event data. It consists 6 large-scale scenes, 7 medium-scale scenes, and 6 object-level scenes. We simulate motion blur by manually placing multi-view cameras, randomly adjusting camera poses, and performing linear interpolation between the original and perturbed positions for each view. The corresponding event streams are simulated using ESIM [41] and V2E [10].", + "bbox": [ + 511, + 291, + 905, + 428 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "EvaGaussians-DAVIS Dataset. We manually recorded five real-world scenes using the Color DAVIS346 event camera [50]. The dataset includes three object-level scenes and two indoor scenes. After processing, the final dataset consists of 30 images per scene, along with the recorded event streams, each showcasing various blur and lighting conditions.", + "bbox": [ + 511, + 428, + 905, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Comparison with Existing Datasets. The goal of contributing this dataset is to create a more comprehensive evaluation benchmark for event-assisted deblurring reconstruction tasks. While E2NeRF [39] and EvDNeRF [3] have introduced similar datasets, we observed that these datasets are limited in terms of scene variety, camera motion diversity, and data size. As shown in Table.1 and Table.2 of the supplementary, our proposed dataset includes a broader range of scenes with various scales, making it more suitable for comprehensive evaluation.", + "bbox": [ + 511, + 534, + 905, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Experiment Settings", + "text_level": 1, + "bbox": [ + 513, + 696, + 705, + 713 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Baselines. We compare our method with three types of baselines: 1) NeRF [30] and 3D-GS [14] directly trained on the blurry images, referring to as B-NeRF and B-3DGS. 2) Deblur rendering methods, including BAD-NeRF [56], BAD-GS [70], BAGS [38], $\\mathrm{E}^2\\mathrm{NeRF}$ [39], and EvDNeRF [3]. Among these, BAD-NeRF and BAD-GS simulate motion blur and optimize camera trajectories without using event stream, while BAGS addresses motion blur through kernel modeling, also without event stream. In contrast, $\\mathrm{E}^2\\mathrm{NeRF}$ and EvDNeRF are event-assisted methods without camera trajectory optimization. 3) Image deblur methods, including UFP [5] (single-image deblurring),", + "bbox": [ + 511, + 719, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24784", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8b2d07830431f64054add0d7b8d3d90ba653c374e34190d80d65ea35b64c1085.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on the synthetic dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Full results are provided in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 132, + 88, + 864, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "EDI [34] (event-based deblurring), and EFNet [46] (learnable event-based deblurring). We process input blurry images with them and train the vanilla 3D-GS with predeblurred images. The resulting baselines are referred to as UFP-GS, EDI-GS, and EFN-GS.", + "bbox": [ + 89, + 584, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. For synthetic datasets, we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [57], and VGG-based Learned Perceptual Image Patch Similarity (LPIPS) [67] to evaluate the similarity between rendered novel views and ground-truth novel views. For real-world datasets, since the sharp ground-truth images are unavailable, we utilize several No-Reference Image Quality Assessment (NR-IQA) metrics for evaluation, including BRISQUE [31], NIQE [32], PIQE [53], RankIQA [26], and MetaIQA [73], which allow for image evaluation when lacking ground truth images.", + "bbox": [ + 89, + 659, + 482, + 825 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Comparisons", + "text_level": 1, + "bbox": [ + 89, + 833, + 230, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate our approach across a variety of scenes, including large-scale scenes, medium-scale scenes, object-level scenes and real scenes. Quantitative assessments of", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "novel view synthesis are shown in Table. 1. It can be found that our method achieves substantial improvements in most of the metrics, especially in challenging large-scale scenes. Specifically, both B-NeRF and B-3DGS produce blurry novel views since they are directly trained on motion-blurred images. The image deblurring-based baselines, UFP-GS, EDI-GS and EFN-GS, also produced inferior results, because the image deblurring process potentially corrupts the 3D consistency of the training images. Notably, our approach outperforms BAD-NeRF [56], BAD-GS [70] and BAGS [38], due to their limited capability in modeling complex textures. In addition, our method also surpasses the event-assisted methods $\\mathrm{E}^2\\mathrm{NeRF}$ [39] and EvDNeRF [3] in producing high-quality novel views with intricate details, with better training and rendering efficiency. Notably, since directly using COLMAP on the blurry images fails in most cases (as shown in Table.4 of the supplementary), we use the EDI-produced poses for all the comparison baselines and our method to ensure a fair comparison.", + "bbox": [ + 511, + 584, + 906, + 871 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The qualitative results of five top-performing methods", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24785", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2ded9169f56cf91f41121cf49e8fdcde63ca8429cdbbc2b4311a28fe36975e48.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of novel view synthesis across large-scale, medium-scale, object-level, and real-world scenes. The table reports the average performance for each scale, demonstrating that our method consistently surpasses previous state-of-the-art approaches across all metrics. Best-performing results are highlighted in bold and second-best results in underline." + ], + "table_footnote": [], + "table_body": "
Scene TypeMetricsB-NeRFB-3DGSUFP-GS [5]EDI-GS [34]EFN-GS [46]E2NeRF [39]BAD-NeRF [56]BAD-GS [70]BAGS [38]EvDNeRF [3]Ours
LargePSNR↑21.3321.4821.3622.3122.6922.9623.8523.8624.9324.6326.02
SSIM↑0.6780.6880.6600.6860.6830.7070.7320.7330.7450.7530.806
LPIPS↓0.4250.3970.3740.3820.3630.3750.3480.3470.3270.3280.268
MediumPSNR↑24.0824.8026.3826.4426.1327.7828.4428.4628.7428.9130.47
SSIM↑0.7170.7510.8020.8010.7980.8660.8790.8730.8740.8850.916
LPIPS↓0.3620.3190.2640.2580.2730.1990.1820.1820.1740.1690.152
ObjectPSNR↑22.2822.3425.1624.9425.4529.6127.3327.8628.4129.8330.24
SSIM↑0.9040.9050.9280.9250.9290.9640.9480.9500.9560.9660.970
LPIPS↓0.1480.1470.1170.1210.1100.0740.0930.0910.0840.0720.070
RealBRISQUE↓92.2573.8062.9462.7562.9361.5261.5060.8958.5858.6353.96
NIQE↓15.0012.0110.1710.2010.219.44010.009.9029.2869.0118.371
PIQE↓65.9252.7445.0344.8344.8446.7643.9543.5145.2744.6341.53
RankIQA↓9.4287.5426.4396.4116.4115.5736.2856.2235.6415.3204.895
MetaIQA↑0.1240.1420.1730.1740.1730.1810.1770.1790.1980.1910.213
", + "bbox": [ + 94, + 133, + 903, + 359 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7692f256e148173b3bd5b7166d6e45370dad6c37f0859613432644ff851831e4.jpg", + "image_caption": [ + "Blurry Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 373, + 230, + 531 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/61362ef85eca0d2cc84d792f893ae4a872608c25a4441972fe6d66b8d991f144.jpg", + "image_caption": [ + "$\\mathrm{E}^2\\mathrm{NeRF}$" + ], + "image_footnote": [], + "bbox": [ + 230, + 375, + 362, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2b9482e2f3521bb08dd1c83bc3b86d250021b668fd6d869ecf996743f93e701.jpg", + "image_caption": [ + "BAD-GS", + "Figure 5. Qualitative comparison on the real-world dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Compared with baselines, our method is capable of reconstructing shaper geometry from blurry inputs." + ], + "image_footnote": [], + "bbox": [ + 364, + 375, + 496, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10c45d0baf38537598bef3cb3109049782f3b912bea66e564f44d08d685dd7b1.jpg", + "image_caption": [ + "BAGS" + ], + "image_footnote": [], + "bbox": [ + 500, + 375, + 632, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d7f51291244177e455030f2a6884167696e734244f327adf99b78733de49a27e.jpg", + "image_caption": [ + "EvDNeRF" + ], + "image_footnote": [], + "bbox": [ + 632, + 375, + 766, + 530 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/86971b530a316f0511cc918201512d63acc49306abe955665146864bd54360e8.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 767, + 375, + 903, + 530 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "are presented in Figure. 4 and Figure. 5, and the complete qualitative results are provided in the supplementary. In Figure. 4, it can be found that $\\mathrm{E}^2\\mathrm{NeRF}$ [39] struggles in medium and large-scale scene modeling, producing significant blurring results. Additionally, BAD-GS [70] and BAGS [38] falls short in regions with significant color and depth variations, and produces overly smooth background textures. Although EvDNeRF [3] exhibits overall satisfactory performance, its complex network architecture prolongs the training time and precludes real-time rendering. In comparison, our method outperforms the baselines in producing high-fidelity novel views, and significantly reducing training time as well as demonstrating advantages in real-time applications. Figure. 5 further demonstrates that ours is capable of reconstructing detailed geometry from blurry inputs, achieving better performance than baselines.", + "bbox": [ + 88, + 593, + 485, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 847, + 243, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of Poses Optimization. We conduct experiments on the large-scale subset of synthetic data to evaluate", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "the impact of pose optimization. As shown in Table 3, we calculate the ATE (Average Trajectory Error) between the initial COLMAP poses and GT poses, as well as the ATE between the optimized poses and GT poses. The results demonstrate the effectiveness of our method in correcting the initial poses and recovering more accurate camera trajectories. Additionally, the improved image quality metrics further demonstrate the effectiveness of pose optimization.", + "bbox": [ + 511, + 593, + 906, + 715 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Analysis of Number of Latent Camera Poses. We conduct ablations to investigate the effect of camera poses number in the exposure time, using the large-scale subset of the synthetic data. Specifically, we vary the number of latent camera poses $n$ from 5, 9, 13, and 17, and visualize the quantitative novel view synthesis results in Figure. 6. It can be observed that the performance reaches a bottleneck at 9 latent poses. Beyond this point, further improvements are limited and may potentially lead to local convergence issues due to over-parameterization. Based on these experiments, we choose $n = 9$ camera poses to achieve a balance between rendering performance and training efficiency. Here,", + "bbox": [ + 511, + 719, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24786", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8520e65798fe81ba266bfac77956743714ac18ad0463c960ba8d622479411ba5.jpg", + "image_caption": [ + "Figure 6. Ablation on number of latent camera poses. We vary the number of latent camera poses $n$ from 5, 9, 13, and 17, and compare with BAD-NeRF [56] and BAD-GS [70] in each setting." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 225, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/02ae19ec502b17aa1e5decd7a117e525e836373aff07661ff03dc115010f0e05.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 90, + 356, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feef9527cf143b5b60abee2f8a7c77dac8abea66dca88c197bb097a610994dc0.jpg", + "image_caption": [ + "Figure 7. Ablation on the impact of event threshold. We vary the threshold from $c = 0.1$ to $c = 0.4$ . Our method shows robustness to varying event thresholds since it does not rely solely on event data." + ], + "image_footnote": [], + "bbox": [ + 357, + 90, + 486, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac827d0b3c523bcb7cc253c2008c57e4ccc5f1d5d2ab6ee4f9620d4ce5958751.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 90, + 640, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0776ba992b5af5d7dc960cde74a694849fcdd9546f9a55743a4414e0ccc811f9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 90, + 767, + 219 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/68aa3348edbb3046bf37f4fea86a3ad5915805725438bae75a567827bc1adc3f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 769, + 90, + 898, + 219 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/7b6806ee00b81483a54903e2191d24af4d1eebd8e9421ea43b87e25f32b3dd99.jpg", + "table_caption": [ + "Table 2. Ablation on the loss functions. Best-performing results are highlighted in bold and second results in underline." + ], + "table_footnote": [], + "table_body": "
Large-scaleMedium-scaleReal Scene
\\( \\mathcal{L}_{\\text{blur}} \\)\\( \\mathcal{L}_{\\text{event}} \\)\\( \\mathcal{L}_{\\text{depth}} \\)\\( \\mathcal{L}_{\\text{int}} \\)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓BRISQUE↓NIQE↓PIQE↓MetaIQA↑RankIQA↓
✓✘✘✘24.980.7870.29929.100.8950.16757.328.79943.260.1986.125
✓✓✘✘25.710.7950.27529.940.9100.15556.168.58242.510.2015.066
✓✓✓✘25.540.8000.27130.050.9140.15456.178.49942.230.2004.923
✓✓✓✓26.020.8060.26830.470.9160.15253.968.37141.530.2134.895
", + "bbox": [ + 93, + 301, + 903, + 386 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82f13b489900f17fe6d50f699240baa53dc781cc264c4bc0f8cac7fb1573dc70.jpg", + "table_caption": [ + "Table 3. Ablation on the effectiveness of pose optimization. We calculate ATE between the initial COLMAP poses and GT poses, as well as ATE between the optimized poses and GT poses." + ], + "table_footnote": [], + "table_body": "
ATE↓PSNR↑SSIM↑LPIPS↓
w/o pose. opt0.383 ± 0.12524.790.7880.293
w/ pose. opt0.049 ± 0.01126.020.8060.268
", + "bbox": [ + 94, + 448, + 478, + 498 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1abf781521a9f1e66d4d250fd475b417e4c55d13769828c558581fb2ab9a3ca7.jpg", + "table_caption": [ + "Table 4. Ablation of different motion blur levels. Mild blur corresponds to $0.5 \\times$ camera motion speed, medium blur to the default speed, and strong blur to the $2.0 \\times$ camera motion speed." + ], + "table_footnote": [], + "table_body": "
Blur levelMild (0.5×)Medium (1.0×)Strong (2.0×)
PSNR↑26.3825.7125.04
SSIM↑0.8160.7950.789
LPIPS↓0.2690.2750.280
", + "bbox": [ + 94, + 559, + 475, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "we also provide a comparison with BAD-NeRF [56] and BAD-GS [70]. These two methods typically use linear interpolation to obtain camera trajectory, while our camera trajectories are estimated from the decomposed latent images, which provides more accurate initialization and helps our method achieves better performance.", + "bbox": [ + 89, + 642, + 482, + 732 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of The Loss Functions. We conduct novel view synthesis experiments on the challenging medium-scale and large-scale subsets of the synthetic dataset and the real-world dataset to validate the effectiveness of the training losses. The quantitative results are shown in Table. 2. It indicates that using only the blur reconstruction loss $\\mathcal{L}_{\\mathrm{blur}}$ leads to suboptimal outputs. In contrast, incorporating $\\mathcal{L}_{\\mathrm{event}}, \\mathcal{L}_{\\mathrm{depth}}$ , and $\\mathcal{L}_{\\mathrm{int}}$ enables our method to achieve better performance.", + "bbox": [ + 89, + 733, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Event Threshold. We conduct experiments on the medium-scale subset of the synthetic data to evaluate", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the impact of different event thresholds. We vary the threshold from $c = 0.1$ to $c = 0.4$ , and the results are shown in Figure 7. Since our method does not rely solely on event data and incorporates RGB inputs as supervision, it demonstrates robustness to varying event thresholds.", + "bbox": [ + 511, + 405, + 903, + 479 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to Motion Blur Levels. To validate the robustness of our method in handling different levels of motion blur, we set up three different camera speeds in the City Blocks scene of the synthetic dataset to obtain images with varying levels of blur. Images with mild blur are captured at $0.5 \\times$ camera motion speed, images with medium blur are captured at the default motion speed, and images with strong blur are captured at $2.0 \\times$ motion speed. The quantitative results of novel view synthesis are listed in Table. 4. It can be observed that the results show no significant fluctuations across different levels of blur, demonstrating the robustness of our method to varying motion blur levels.", + "bbox": [ + 511, + 481, + 905, + 661 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 676, + 638, + 690 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces Event Stream Assisted Gaussian Splitting (EvaGaussians), a novel framework that seamlessly integrates the event streams captured by an event camera into the training of 3D-GS, effectively addressing the challenges of reconstructing high-quality 3D-GS from motion-blurred images. We contribute two novel datasets and conduct comprehensive experiments. The results demonstrate that our method outperforms previous state-of-the-art deblurring rendering techniques in terms of novel view synthesis quality, without sacrificing inference efficiency. Despite its promising performance, our method may still face challenges when reconstructing scenes with extremely intricate textures from severely blurred images.", + "bbox": [ + 511, + 700, + 906, + 897 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24787", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by Natural Science Foundation of China (No.62332002, No.62202014), and Shenzhen KQTD (No.20240729102051063).", + "bbox": [ + 89, + 114, + 483, + 160 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 172, + 187, + 188 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In CVPR, 2016. 2", + "[2] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame- and event-based davis sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686–689, 2014. 2", + "[3] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6, 7", + "[4] Mauro Comi, Alessio Tonioni, Max Yang, Jonathan Tremblay, Valts Blukis, Yijiong Lin, Nathan F Lepora, and Laurence Aitchison. Snap-it, tap-it, splat-it: Tactile-informed 3d gaussian splatting for reconstructing challenging surfaces. arXiv preprint arXiv:2403.20275, 2024. 5", + "[5] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 5, 7", + "[6] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2924–2932, 2025. 2", + "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 2020. 2, 3", + "[8] Haiqian Han, Jianing Li, Henglu Wei, and Xiangyang Ji. Event-3dgs: Event-based 3d reconstruction using 3d gaussian splatting. In Advances in Neural Information Processing Systems, 2025. 3", + "[9] Philipp Heise, Sebastian Klose, Brian Jensen, and Alois Knoll. Pm-huber: Patchmatch with huber regularization for stereo matching. In ICCV, 2013. 5", + "[10] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 4, 5", + "[11] Jian Huang, Chengrui Dong, and Peidong Liu. Inceventgs: Pose-free gaussian splatting from a single event camera. In CVPR, 2025. 3", + "[12] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 3" + ], + "bbox": [ + 93, + 198, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[13] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In CVPR, 2020. 2", + "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 4, 5", + "[15] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In CVPR, 2023. 1", + "[16] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 2023. 3", + "[17] Joseph J LaViola Jr. Bringing vr and spatial 3d interaction to the masses through video games. IEEE Computer Graphics and Applications, 2008. 1", + "[18] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2", + "[19] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision, pages 416-434. Springer, 2025. 3", + "[20] Yuchen Li*, Chaoran Feng*, Zhenyu Tang, Kaiyuan Deng, Wangbo Yu, Yonghong Tian, and Li Yuan. Gs2e: Gaussian splatting is an effective data generator for event stream generation. arXiv preprint arXiv:2505.15287, 2025. 2", + "[21] Guixu Lin, Jin Han, Mingdeng Cao, Zhihang Zhong, and Yinqiang Zheng. Event-guided frame interpolation and dynamic range expansion of single rolling shutter image. In Proceedings of the 31st ACM International Conference on Multimedia, 2023.", + "[22] Mingyuan Lin, Yangguang Wang, Xiang Zhang, Boxin Shi, Wen Yang, Chu He, Gui-song Xia, and Lei Yu. Self-supervised shutter unrolling with events. IJCV, 2025. 2", + "[23] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In ECCV, 2020. 2", + "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 1", + "[25] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2024. 1", + "[26] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 6", + "[27] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[28] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural Radiance Fields from Blurry Images. In CVPR, 2022. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24788", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[29] Nico Messikommer, Carter Fang, Mathias Gehrig, Giovanni Cioffi, and Davide Scaramuzza. Data-driven feature tracking for event cameras with and without frames. TPAMI, 2025. 2", + "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5", + "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 21(12):4695-4708, 2012. 6", + "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6", + "[33] Genady Paikin, Yotam Ater, Roy Shaul, and Evgeny Soloveichik. Efi-net: Video frame interpolation from fusion of events and frames. In CVPR, 2021. 2", + "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 2, 3, 6, 7", + "[35] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In CVPR, 2020. 2", + "[36] Roberto Pellerito, Marco Cannici, Daniel Gehrig, Joris Bel-hadj, Olivier Dubois-Matra, Massimo Casasco, and Davide Scaramuzza. Deep visual odometry with events and frames. In IROS, 2024. 2", + "[37] Cheng Peng and Rama Chellappa. Pdrf: Progressively deblurring radiance field for fast and robust scene reconstruction from blurry images. arXiv preprint arXiv:2208.08049, 2022. 2", + "[38] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In ECCV, 2024. 2, 5, 6, 7", + "[39] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 3, 5, 6, 7", + "[40] Yunshan Qi, Lin Zhu, Yifan Zhao, Nan Bao, and Jia Li. Deblurring neural radiance fields with event-driven bundle adjustment. In MM, 2024. 3", + "[41] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 4, 5", + "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. 2023. 1", + "[43] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, 2016. 1, 2, 3", + "[45] Zihang Shao*, Chaoran Feng*, Xuanye Fang, Yaxin Li, Jiangrong Shen, and Qi Xu. Eicil: joint excitatory inhibitory cycle iteration learning for deep spiking neural networks. Advances in Neural Information Processing Systems, 36: 32117-32128, 2023. 2" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 6, 7", + "[47] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022. 2", + "[48] Zhenyu Tang*, Chaoran Feng*, Xinhua Cheng, Wangbo Yu, Junwu Zhang, Yuan Liu, Xiaoxiao Long, Wenping Wang, and Li Yuan. Neuralgs: Bridging neural fields and 3d gaussian splatting for compact 3d representations. arXiv preprint arXiv:2503.23162, 2025.", + "[49] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7320–7328, 2025. 2", + "[50] Gemma Taverni. Applications of Silicon Retinas: From Neuroscience to Computer Vision. PhD thesis, Universität Zürich, 2020. 5", + "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, pages 16155-16164, 2021. 2", + "[52] Manoj K Vairalkar and SU Nimbhorkar. Edge detection of images using sobel operator. International Journal of Emerging Technology and Advanced Engineering, 2(1):291-293, 2012. 5", + "[53] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 6", + "[54] Zhexiong Wan, Bin Fan, Le Hui, Yuchao Dai, and Gim Hee Lee. Instance-level moving object segmentation from a single image with events. IJCV, 2025. 2", + "[55] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. In ICML, 2024. 3", + "[56] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In CVPR, 2023. 2, 3, 5, 6, 7, 8", + "[57] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[58] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2", + "[59] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24789", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[60] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, et al. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia, 2023. 1", + "[61] Wenhao Xu, Wenming Weng, Yueyi Zhang, Ruikang Xu, and Zhiwei Xiong. Event-boosted deformable 3d gaussians for fast dynamic scene reconstruction. arXiv preprint arXiv:2411.16180, 2024. 2, 3", + "[62] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. 2021. 1", + "[63] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 3", + "[64] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024.", + "[65] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, 2024.", + "[66] Sohaib Zahid, Viktor Rudnev, Eddy Ilg, and Vladislav Golyanik. E-3dgs: Event-based novel view rendering of large-scale scenes using 3d gaussian splatting. In 3DV, 2025. 3", + "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6", + "[68] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, 2022. 2", + "[69] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 3", + "[70] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In ECCV, 2024. 2, 5, 6, 7, 8", + "[71] Haiyang Zhou, Wangbo Yu, Jiawen Guan, Xinhua Cheng, Yonghong Tian, and Li Yuan. Holotime: Taming video diffusion models for panoramic 4d scene generation. arXiv preprint arXiv:2504.21650, 2025. 3", + "[72] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2", + "[73] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143-14152, 2020. 6", + "[74] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "feys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 1", + "[75] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. TPAMI. 2" + ], + "bbox": [ + 516, + 92, + 903, + 162 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24790", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_model.json b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c06495d716b0351ec3f2d5c83204e0434f0efb77 --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_model.json @@ -0,0 +1,2598 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.131, + 0.892, + 0.154 + ], + "angle": 0, + "content": "EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images" + }, + { + "type": "text", + "bbox": [ + 0.107, + 0.18, + 0.89, + 0.218 + ], + "angle": 0, + "content": "Wangbo Yu\\(^{1,2*}\\), Chaoran Feng\\(^{1*}\\), Jianing Li\\(^{3}\\), Jiye Tang\\(^{4}\\), Jiashu Yang\\(^{5}\\), Zhenyu Tang\\(^{1}\\), Meng Cao\\(^{6}\\), Xu Jia\\(^{5}\\), Yuchao Yang\\(^{1}\\), Li Yuan\\(^{1,2\\dagger}\\) and Yonghong Tian\\(^{1,2\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.118, + 0.236, + 0.88, + 0.288 + ], + "angle": 0, + "content": "1Shenzhen Graduate School, Peking University, 2Peng Cheng Laboratory, 3School of Computer and Science, Peking University \n4University of Science and Technology of China, 5Dalian University of Technology, 6MBZUAI {wbyu, chaoran.feng}@stu.pku.edu.cn, {yuanli-ece, yhtian}@pu.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.312, + 0.273, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.274, + 0.312, + 0.452, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.312, + 0.631, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.409, + 0.274, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.275, + 0.409, + 0.452, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.453, + 0.409, + 0.631, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.312, + 0.901, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.444, + 0.901, + 0.502 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.511, + 0.908, + 0.54 + ], + "angle": 0, + "content": "Figure 1. EvaGaussians integrates frame-based cameras and event cameras to reconstruct sharp 3D-GS from motion-blurred images. It achieves higher-quality novel view synthesis compared to previous methods." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.557, + 0.327, + 0.572 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.589, + 0.486, + 0.862 + ], + "angle": 0, + "content": "3D Gaussian Splatting (3D-GS) has demonstrated exceptional capabilities in synthesizing novel views of 3D scenes. However, its training is heavily reliant on high-quality images and precise camera poses. Meeting these criteria can be challenging in non-ideal real-world conditions, where motion-blurred images frequently occur due to high-speed camera movements. To address these challenges, we introduce Event Stream Assisted Gaussian Splatting (EvaGaus-sians), a novel approach that harnesses event streams captured by event cameras to facilitate the learning of high-quality 3D-GS from motion-blurred images. Capitalizing on the high temporal resolution and dynamic range offered by event streams, we seamlessly integrate them into the initialization and optimization of 3D-GS, thereby enhancing the acquisition of high-fidelity novel views with intricate texture details. We also contribute two novel datasets comprising RGB frames, event streams, and corresponding camera parameters, featuring a wide variety of scenes" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.62 + ], + "angle": 0, + "content": "and various camera motions. The comparison results reveal that our approach not only excels in generating high-fidelity novel views, but also offers improved training and inference efficiency. Video and code are available at the project page." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.649, + 0.645, + 0.665 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.674, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Novel view synthesis from 2D image collections has presented a persistent challenge within the field of computer vision and computer graphics. This task stands as a fundamental component in various vision applications, such as virtual reality [17, 60], robotics navigation [42, 62, 74], scene understanding [15, 24, 25], and many others, thereby prompting significant research efforts over the last decades. Amid pioneering works, 3D Gaussian Splatting (3D-GS) [14] achieves notable success in generating high-fidelity novel views. However, the optimization of 3D-GS heavily relies on accurate camera poses and point cloud initialization produced by COLMAP [44], which necessitates high-quality images without blurring. Fulfilling such conditions can be challenging in real-world situations. For example, in UAVs and robotics, rapid camera movement is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.888 + ], + "angle": 0, + "content": "*These authors contributed equally to this work." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.875, + 0.365, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24780" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "common when capturing images or recording videos, which often result in significant motion blur." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.129, + 0.484, + 0.31 + ], + "angle": 0, + "content": "Recent studies have demonstrated the significant potential of event-based cameras in alleviating motion blur in images captured by conventional frame-based cameras. [13, 23, 34]. Serving as an innovative bio-inspired visual sensor, event cameras asynchronously report the logarithmic intensity changes of each pixel captured, and can record higher temporal resolution and dynamic range data in contrast to conventional cameras. The continuous event streams naturally align with the frame-rate agnostic rendering property of 3D-GS, offering sharper scene geometry information than frame-based cameras, making it highly promising in facilitating 3D-GS training from fast camera motion." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.318, + 0.484, + 0.664 + ], + "angle": 0, + "content": "Motivated by this, we propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first method that seamlessly incorporates continuous event streams captured by event cameras to enhance the learning of high-quality 3D-GS from motion-blurred images. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Furthermore, to tackle the geometric ambiguity caused by blurry images, we introduce two event-assisted geometry regularization terms to stabilize the geometry of 3D-GS. Through optimizing the 3D-GS in a progressive manner, our method can recover a high-quality 3D-GS that facilitates the real-time generation of high-fidelity novel views. Finally, to comprehensively evaluate novel view synthesis with hybrid frame- and event-based cameras, we contribute a synthetic dataset and a real-world dataset, both containing diverse scenes with varying scales and camera motions, setting a benchmark for related researches. To summarize, our contributions can be delineated as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.763 + ], + "angle": 0, + "content": "- We propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first framework designed to reconstruct high-quality 3D-GS from motion-blurred images with the assistance of event cameras. It not only recovers intricate details from the motion-blurred images but also enables high-fidelity novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.763, + 0.483, + 0.823 + ], + "angle": 0, + "content": "- We contribute a synthetic dataset and a real-world dataset for evaluating novel view synthesis with hybrid frame-and event-based cameras. Both datasets contain diverse scenes with varying scales and camera motions." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.823, + 0.484, + 0.899 + ], + "angle": 0, + "content": "- We conduct a comprehensive evaluation of the proposed method and compare it with several strong baselines. As shown in Figure. 1 and Figure. 2, our method not only excels in generating higher-fidelity novel views but also enables improved training and inference efficiency." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.672, + 0.484, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.089, + 0.885, + 0.356 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.358, + 0.907, + 0.413 + ], + "angle": 0, + "content": "Figure 2. Performance, rendering speed, and training time (indicated by the area of each circle) comparison. EvaGaussians achieves better results while maintaining real-time rendering and training efficiency." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.421, + 0.665, + 0.437 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.447, + 0.9, + 0.464 + ], + "angle": 0, + "content": "2.1. Reconstructing 3D Scene from Blurry Images" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.907, + 0.786 + ], + "angle": 0, + "content": "Reconstructing a high-quality 3D Scene typically requires high-fidelity, sharp images as supervision. However, motion-blurred images often occur in real world scenarios, thus hindering accurate reconstruction of 3D scenes. Several studies have been proposed to address this issue. For example, Deblur-NeRF [28], DP-NeRF [18] and PDRF [37] attempted to learn a blur formation kernel to model the image blurring process. BAD-NeRF [56] further physically modeled the blurry images formation process, and adopted a bundle-adjustment strategy to jointly optimize NeRF parameters and the camera poses during the exposure time. These NeRF-based methods lacked real-time rendering capabilities and suffered from extended training times. With the rapid advancement of 3D-GS, concurrent works like BAD-Gaussians [70] and BAGS [38] proposed to utilize 3D-GS as representation to achieve deblurring reconstruction. Although they achieved real-time rendering and faster convergence compared with prior works, they still struggled to handle severely blurred images in which COLMAP [44] will fail to produce the initial point clouds and camera calibration." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.802, + 0.9, + 0.818 + ], + "angle": 0, + "content": "2.2. Reconstructing 3D Scene from Event Streams" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Motivated by the exceptional properties offered by event cameras [1, 2, 6, 7, 20-22, 29, 33, 35, 36, 45, 47-49, 51, 54, 58, 61, 68, 72, 75], several studies attempted to reconstruct 3D scenes from event streams captured by event cameras, particularly in low-light conditions with fast camera mo" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24781" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.289 + ], + "angle": 0, + "content": "tion [8, 11, 12, 27, 43, 55, 59, 61, 63-66, 69, 71]. These methods were mostly designed to recover scene geometry from event streams, and are not typically designed for reconstructing 3D Scene from blurry images. Recently, E-NeRF [16], \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [39], EBAD-NeRF [40], BENeRF [19] and EvDNeRF [3] proposed to jointly utilize event streams captured by event cameras and motion-blurred images captured by standard frame-based cameras to reconstruct a NeRF representation. In contrast to frame-based deblurring methods, they are better at handling motion blur due to the introduction of event information. However, these NeRF-based methods suffer from long training and inference times, which limit their further application." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.301, + 0.182, + 0.317 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.325, + 0.221, + 0.342 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.347, + 0.485, + 0.665 + ], + "angle": 0, + "content": "Event camera is a type of bio-inspired sensor that can asynchronously record intensity changes [7]. In contrast to conventional cameras that are restricted to sequentially produce frames at a fixed frame rate, event cameras asynchronously trigger events in each pixel when their intensity change exceeds a constant threshold, featuring properties such as low latency and high dynamic range. Formally, let \\(\\mathbf{I}_{xy}(t)\\) denote the instantaneous intensity at pixel coordinate \\((x,y)\\) at time \\(t\\), and \\(\\mathbf{L}_{xy}(t)\\) denotes its logarithm. An event \\(p = \\pm 1\\) will be triggered whenever the change of \\(\\mathbf{L}_{xy}(t)\\) surpasses the threshold \\(c\\), where the polarity represents the direction (increase or decrease) of changes. Let \\(\\delta_{t_0}(t)\\) be the impulse function at time \\(t_0\\) with a unit integral, the event can therefore be expressed as a continuous-time signal \\(\\mathbf{e}_{xy}(t) = p\\delta_{t_0}(t)\\), where \\(t_0\\) signifies the time at which the event occurs. Then, the proportional intensity change during a time interval \\([s,t]\\) can be computed as the integral of events that occurred between times \\(s\\) and \\(t\\), expressed as \\(\\mathbf{E}_{xy}(t) = \\int_s^t\\mathbf{e}_{xy}(h)dh\\). Given that each pixel can be treated separately in the event camera, the subscripts can be omitted:" + }, + { + "type": "equation", + "bbox": [ + 0.218, + 0.672, + 0.483, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathbf {E} (t) = \\int_ {s} ^ {t} \\mathbf {e} (h) d h. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.714, + 0.484, + 0.76 + ], + "angle": 0, + "content": "We can then represent the logarithmic intensity change as: \\(\\mathbf{L}(t) - \\mathbf{L}(s) = c\\mathbf{E}(t)\\), rewrite as \\(\\mathbf{L}(t) = \\mathbf{L}(s) + c\\mathbf{E}(t)\\), and subsequently obtain the actual intensity change:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.769, + 0.483, + 0.786 + ], + "angle": 0, + "content": "\\[\n\\mathbf {I} (t) = \\mathbf {I} (s) \\cdot \\exp (c \\mathbf {E} (t)). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.843 + ], + "angle": 0, + "content": "Therefore, when an image \\(\\mathbf{I}(s)\\) is captured at time \\(s\\), and the event stream is recorded during the time interval \\([s,t]\\), the image \\(\\mathbf{I}(t)\\) can be obtained by warping \\(\\mathbf{I}(s)\\) using Eq. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.341, + 0.865 + ], + "angle": 0, + "content": "3.2. Event-assisted Initialization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The optimization of 3D-GS requires camera calibration and point cloud initialization using COLMAP [44]. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "as shown in Table.4 in the supplementary, this process can fail when dealing with images that have significant motion blur. Motion-blurred images are resulted from camera movements during the exposure time, which can be mathematically represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.178, + 0.907, + 0.215 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\mathbf {I} (t) d t, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.905, + 0.269 + ], + "angle": 0, + "content": "where \\(\\mathbf{B}\\) denotes a captured blurry image, which is equivalent to averaging the instantaneous latent images \\(\\mathbf{I}(t)\\) during the exposure time \\([s - \\tau /2,s + \\tau /2]\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.345 + ], + "angle": 0, + "content": "To obtain initial camera poses and point clouds for 3D-GS optimization, we first preprocess the motion-blurred images using the Event-based Double Integral (EDI) [34] model, which can be derived through substituting Eq. 2 into Eq. 3:" + }, + { + "type": "equation", + "bbox": [ + 0.585, + 0.355, + 0.907, + 0.393 + ], + "angle": 0, + "content": "\\[\n\\mathbf {B} = \\mathbf {I} (s) \\cdot \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\exp (c \\mathbf {E} (t)) d t. \\qquad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.906, + 0.553 + ], + "angle": 0, + "content": "Given the predefined threshold \\( c \\), a blurry image \\( \\mathbf{B} \\), and the recorded event stream \\( \\mathbf{E}(t) \\), the EDI model (Eq. 4) allows the derivation of \\( \\mathbf{I}(s) \\), following which the latent image \\( \\mathbf{I}(t) \\) at any moment within the exposure time can be estimated through Eq. 2. As shown in Figure. 3, given a total of \\( K \\) blurry images \\( \\{\\mathbf{B}^j\\}_{j=1}^K \\), for each of them, we uniformly sample \\( n \\) time stamps during their exposure time to obtain a series of EDI-processed latent images rich in texture features, denoted as \\( \\{\\mathbf{I}_i\\}_{i=1}^n \\), then obtain their poses \\( \\{\\mathbf{P}_i\\}_{i=1}^n \\) and the initial point cloud of the scene using COLMAP [44]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.735 + ], + "angle": 0, + "content": "After initialization, a straightforward approach to optimizing the 3D-GS is to use the EDI-processed latent images and their camera poses as supervision. However, although these latent images provide more texture features than the original blurry image, they still do not fully recover the ideal latent image and exhibit relatively low visual quality, which also introduces inaccuracies into the camera poses, thereby leading to unsatisfactory optimization results. To more robustly recover a sharp 3D-GS from motion-blurred images, we propose to harness the advantages of event streams and seamlessly integrate them into the optimization process of 3D-GS." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.818, + 0.76 + ], + "angle": 0, + "content": "3.3. Event-assisted Bundle-Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "As introduced in Eq. 3, during the exposure time, a motion-blurred image can be decomposed into a series of latent images along a specific camera trajectory, which can be roughly approximated by the EDI-produced camera poses \\(\\{\\mathbf{P}_i\\}_{i=1}^n\\) according to Eq. 4. Motivated by this, we jointly optimize these camera poses and the 3D-GS attributes in a bundle-adjustment manner [56] to simultaneously recover the blur-formation camera trajectories and a sharp 3D-GS. As shown in Figure. 3, we add each of the EDI-produced" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24782" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.317 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.329, + 0.908, + 0.413 + ], + "angle": 0, + "content": "Figure 3. Overview of EvaGaussians. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Leveraging the continuously recorded event streams, we introduce two event-assisted geometry regularization terms that can function beyond the exposure time to stabilize the geometry of 3D-GS. Once trained, it can not only recover intricate details from motion-blurred images but also enable high-fidelity novel view synthesis." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.482, + 0.594 + ], + "angle": 0, + "content": "camera poses a learnable offset \\(\\{\\mathbf{d}_i\\}_{i = 1}^n\\) as correction parameters, resulting a learnable camera trajectory \\(\\{\\widetilde{\\mathbf{P}}_i\\}_{i = 1}^n\\) where \\(\\widetilde{\\mathbf{P}}_i = \\mathbf{P}_i + \\mathbf{d}_i\\). In each training iteration, we simultaneously render \\(n\\) images \\(\\{\\widetilde{\\mathbf{I}}_i\\}_{i = 1}^n\\) from the 3D-GS along the camera trajectory of the blurry view, and then simulate the formation of the motion-blurred images using a discrete approximation of Eq 3, denoted as \\(\\widetilde{\\mathbf{B}} = \\frac{1}{n}\\sum_{i = 1}^{n}\\widetilde{\\mathbf{I}}_i\\). Consequently, for a total of \\(K\\) real-captured blurry images \\(\\{\\mathbf{B}^j\\}_{j = 1}^K\\), we can obtain their simulated versions \\(\\{\\widetilde{\\mathbf{B}}^j\\}_{j = 1}^K\\) through each corresponding learnable camera trajectory." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.483, + 0.685 + ], + "angle": 0, + "content": "Blur Reconstruction Loss. With the simulated blurry images, we use the captured real blurry images \\(\\{\\mathbf{B}^j\\}_{j=1}^K\\) to serve as image level supervision. Specifically, for each blurry image \\(\\mathbf{B}^j\\) and its simulated version \\(\\tilde{\\mathbf{B}}^j\\), we employ a blur reconstruction loss to minimize their photometric error, expressed as" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.692, + 0.483, + 0.712 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {b l u r} = \\left(1 - \\lambda_ {1}\\right) \\cdot \\| \\mathbf {B} ^ {j} - \\widetilde {\\mathbf {B}} ^ {j} \\| _ {1} + \\lambda_ {1} \\cdot \\mathrm {D - S S I M} \\left(\\mathbf {B} ^ {j}, \\widetilde {\\mathbf {B}} ^ {j}\\right). (5)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.795 + ], + "angle": 0, + "content": "The formulation of blur reconstruction loss is the same as in the original 3D-GS [14], it differs in utilizing blurry images as supervision and jointly optimizing the 3D-GS attributes and the camera trajectories, thus facilitating an initial deblurring reconstruction of 3D-GS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Event Reconstruction Loss. Leveraging the abundant high-frequency information offered by the event streams, we further adopt an event reconstruction loss to aid in 3D-GS optimization. Specifically, we uniformly divide the exposure time into \\( m = n - 1 \\) intervals, each with a duration of \\( \\frac{\\tau}{m} \\). Subsequently, we integrate the recorded event stream along these time intervals using Eq. 1, resulting in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.547 + ], + "angle": 0, + "content": "\\(m\\) event maps \\(\\{\\mathbf{E}_i\\}_{i=1}^m\\) to serve as event level supervision. During training, for the \\(j\\)-th blurry view, we convert the rendered image sequence \\(\\{\\widetilde{\\mathbf{I}}_i\\}_{i=1}^n\\) on the camera trajectory into event maps \\(\\{\\widetilde{\\mathbf{E}}_i\\}_{i=1}^m\\), using a differentiable event simulator [10, 41], and constrain the discrepancies between the simulated event maps and the ground truth event maps, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.559, + 0.907, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e v e n t}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\| \\mathbf {E} _ {i} - \\widetilde {\\mathbf {E}} _ {i} \\| _ {1}. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.61, + 0.907, + 0.641 + ], + "angle": 0, + "content": "The event reconstruction loss further aids in recovering a sharp 3D-GS with improved texture details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.86, + 0.668 + ], + "angle": 0, + "content": "3.4. Event-assisted Geometry Regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.84 + ], + "angle": 0, + "content": "The blurry color images are captured only during the exposure time and are much sparser than the event stream. Relying on such low-quality image-level supervision may cause the 3D-GS to overfit on the training images, resulting in significant floaters and inferior geometry. Leveraging Eq. 1 and Eq. 4, given the continuously recorded event streams \\(\\mathbf{E}(t)\\), we can derive continuous grayscale intensity images \\(\\mathbf{G}(t)\\) that are rich in geometric information and can function beyond the exposure time. Motivated by this, we further propose two event-assisted geometry regularization terms to aid in 3D-GS training." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Intensity Reconstruction Loss. During training, we randomly sample continuous time \\(t\\) between the interval of two adjacent blurry image, and derive the grayscale intensity image \\(\\mathbf{G}(t)\\) using Eq. 2. We then minimize the difference" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24783" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.122 + ], + "angle": 0, + "content": "between it and the rendered intensity image from 3D-GS, expressed as:" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.131, + 0.484, + 0.164 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i n t} = (1 - \\lambda_ {2}) \\cdot \\| \\mathbf {G} (t) - \\widetilde {\\mathbf {G}} (t) \\| _ {1} + \\lambda_ {2} \\cdot D - S S I M (\\mathbf {G} (t), \\widetilde {\\mathbf {G}} (t)), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.164, + 0.458, + 0.179 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{G}}(t)\\) is converted from the colored render result." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.18, + 0.484, + 0.24 + ], + "angle": 0, + "content": "Intensity-aware Depth Regularization Loss. As shown in Figure. 3, to further improve the geometry of 3D-GS, inspired by [4, 9], we adopt an intensity-aware depth regularization loss during training, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.25, + 0.483, + 0.31 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {x, y} \\left(\\left| \\partial_ {x} \\widetilde {\\mathbf {D}} _ {x y} (t) \\right| e ^ {- \\beta \\left| \\partial_ {x} \\mathbf {G} _ {x y} (t) \\right|} \\right. \\tag {8} \\\\ + | \\partial_ {y} \\widetilde {\\mathbf {D}} _ {x y} (t) | e ^ {- \\beta | \\partial_ {y} \\mathbf {G} _ {x y} (t) |}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.321, + 0.484, + 0.472 + ], + "angle": 0, + "content": "where \\(\\widetilde{\\mathbf{D}}(t)\\) is the rendered depth map, \\((x,y)\\) denotes the pixel location, \\(N\\) is the total number of pixels, and \\(\\beta\\) is set to 2 in our experiments. The horizontal and vertical gradients are calculated by applying convolution operations with \\(5 \\times 5\\) Sobel kernels [52]. This regularization is founded on the observation that depth transitions in an image often correspond to changes in intensity. Therefore, it ensures that the spatial variation of depth closely matches that of the intensity image, thereby reducing geometric artifacts at object boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.503 + ], + "angle": 0, + "content": "The total loss function is the combination of the above losses, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.161, + 0.515, + 0.483, + 0.55 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {b l u r}} \\mathcal {L} _ {\\text {b l u r}} + \\lambda_ {\\text {e v e n t}} \\mathcal {L} _ {\\text {e v e n t}} \\tag {9} \\\\ + \\lambda_ {i n t} \\mathcal {L} _ {i n t} + \\lambda_ {d e p t h} \\mathcal {L} _ {d e p t h}. \\\\ \\end{array}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.224, + 0.59 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.675 + ], + "angle": 0, + "content": "In this section, we first provide the implementation details of our method in 4.1, followed by an illustration of the introduced dataset in 4.2, and the experimental settings in 4.3. The comparison results and ablation study are presented in 4.4 and 4.5. For more detailed experiments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.683, + 0.308, + 0.699 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Progressive Training. We implemented EvaGaussians based on the official code of 3D-GS [14]. The training process spans 50,000 iterations, with an event reconstruction loss introduced after a 3,000-iteration warmup and we omit the densification process to streamline and simplify the subsequent optimization. Additionally, we adopt a coarse-to-fine training strategy, starting with rendering at a low resolution (0.3× downsampling in the early 30% iterations) and progressively increasing the size of the rendered views to full resolution. All experiments were conducted using a single NVIDIA RTX 4090 GPU." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hyperparameter Setting. During the training process, we set \\(\\lambda_{blur} = 1.0\\), \\(\\lambda_{depth} = 1.0e^{-2}\\), \\(\\lambda_{event} = 5.0e^{-3}\\) and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "\\(\\lambda_{int} = 1.0e^{-3}\\) for the loss function, and used \\(n = 9\\) for the number of poses to be optimized during the exposure time. In implementing the loss \\(\\mathcal{L}_{event}\\), we configured the positive threshold as \\(c_{pos} = 0.25\\) and the negative threshold as \\(c_{neg} = 0.25\\) for synthetic scenes, and set \\(c_{pos} = 0.197\\) and \\(c_{neg} = 0.241\\) for real scenes." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.194, + 0.616, + 0.208 + ], + "angle": 0, + "content": "4.2. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.217, + 0.906, + 0.291 + ], + "angle": 0, + "content": "To enable a comprehensive evaluation, we introduce two novel datasets, with an overview provided below. Additionally, we present a detailed dataset introduction in the supplementary and will release the dataset construction pipeline for future research." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.292, + 0.906, + 0.429 + ], + "angle": 0, + "content": "EvaGaussians-Blender Dataset. We construct a synthetic dataset covering a variety of scene scales, coupling with diverse camera trajectories and event data. It consists 6 large-scale scenes, 7 medium-scale scenes, and 6 object-level scenes. We simulate motion blur by manually placing multi-view cameras, randomly adjusting camera poses, and performing linear interpolation between the original and perturbed positions for each view. The corresponding event streams are simulated using ESIM [41] and V2E [10]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.429, + 0.906, + 0.533 + ], + "angle": 0, + "content": "EvaGaussians-DAVIS Dataset. We manually recorded five real-world scenes using the Color DAVIS346 event camera [50]. The dataset includes three object-level scenes and two indoor scenes. After processing, the final dataset consists of 30 images per scene, along with the recorded event streams, each showcasing various blur and lighting conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.535, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Comparison with Existing Datasets. The goal of contributing this dataset is to create a more comprehensive evaluation benchmark for event-assisted deblurring reconstruction tasks. While E2NeRF [39] and EvDNeRF [3] have introduced similar datasets, we observed that these datasets are limited in terms of scene variety, camera motion diversity, and data size. As shown in Table.1 and Table.2 of the supplementary, our proposed dataset includes a broader range of scenes with various scales, making it more suitable for comprehensive evaluation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.697, + 0.707, + 0.714 + ], + "angle": 0, + "content": "4.3. Experiment Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Baselines. We compare our method with three types of baselines: 1) NeRF [30] and 3D-GS [14] directly trained on the blurry images, referring to as B-NeRF and B-3DGS. 2) Deblur rendering methods, including BAD-NeRF [56], BAD-GS [70], BAGS [38], \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [39], and EvDNeRF [3]. Among these, BAD-NeRF and BAD-GS simulate motion blur and optimize camera trajectories without using event stream, while BAGS addresses motion blur through kernel modeling, also without event stream. In contrast, \\(\\mathrm{E}^2\\mathrm{NeRF}\\) and EvDNeRF are event-assisted methods without camera trajectory optimization. 3) Image deblur methods, including UFP [5] (single-image deblurring)," + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24784" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.865, + 0.535 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.54, + 0.908, + 0.569 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on the synthetic dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Full results are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.483, + 0.659 + ], + "angle": 0, + "content": "EDI [34] (event-based deblurring), and EFNet [46] (learnable event-based deblurring). We process input blurry images with them and train the vanilla 3D-GS with predeblurred images. The resulting baselines are referred to as UFP-GS, EDI-GS, and EFN-GS." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.827 + ], + "angle": 0, + "content": "Evaluation Metrics. For synthetic datasets, we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [57], and VGG-based Learned Perceptual Image Patch Similarity (LPIPS) [67] to evaluate the similarity between rendered novel views and ground-truth novel views. For real-world datasets, since the sharp ground-truth images are unavailable, we utilize several No-Reference Image Quality Assessment (NR-IQA) metrics for evaluation, including BRISQUE [31], NIQE [32], PIQE [53], RankIQA [26], and MetaIQA [73], which allow for image evaluation when lacking ground truth images." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.231, + 0.85 + ], + "angle": 0, + "content": "4.4. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.485, + 0.901 + ], + "angle": 0, + "content": "We evaluate our approach across a variety of scenes, including large-scale scenes, medium-scale scenes, object-level scenes and real scenes. Quantitative assessments of" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.585, + 0.907, + 0.872 + ], + "angle": 0, + "content": "novel view synthesis are shown in Table. 1. It can be found that our method achieves substantial improvements in most of the metrics, especially in challenging large-scale scenes. Specifically, both B-NeRF and B-3DGS produce blurry novel views since they are directly trained on motion-blurred images. The image deblurring-based baselines, UFP-GS, EDI-GS and EFN-GS, also produced inferior results, because the image deblurring process potentially corrupts the 3D consistency of the training images. Notably, our approach outperforms BAD-NeRF [56], BAD-GS [70] and BAGS [38], due to their limited capability in modeling complex textures. In addition, our method also surpasses the event-assisted methods \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [39] and EvDNeRF [3] in producing high-quality novel views with intricate details, with better training and rendering efficiency. Notably, since directly using COLMAP on the blurry images fails in most cases (as shown in Table.4 of the supplementary), we use the EDI-produced poses for all the comparison baselines and our method to ensure a fair comparison." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The qualitative results of five top-performing methods" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24785" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.132 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of novel view synthesis across large-scale, medium-scale, object-level, and real-world scenes. The table reports the average performance for each scale, demonstrating that our method consistently surpasses previous state-of-the-art approaches across all metrics. Best-performing results are highlighted in bold and second-best results in underline." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.134, + 0.904, + 0.361 + ], + "angle": 0, + "content": "
Scene TypeMetricsB-NeRFB-3DGSUFP-GS [5]EDI-GS [34]EFN-GS [46]E2NeRF [39]BAD-NeRF [56]BAD-GS [70]BAGS [38]EvDNeRF [3]Ours
LargePSNR↑21.3321.4821.3622.3122.6922.9623.8523.8624.9324.6326.02
SSIM↑0.6780.6880.6600.6860.6830.7070.7320.7330.7450.7530.806
LPIPS↓0.4250.3970.3740.3820.3630.3750.3480.3470.3270.3280.268
MediumPSNR↑24.0824.8026.3826.4426.1327.7828.4428.4628.7428.9130.47
SSIM↑0.7170.7510.8020.8010.7980.8660.8790.8730.8740.8850.916
LPIPS↓0.3620.3190.2640.2580.2730.1990.1820.1820.1740.1690.152
ObjectPSNR↑22.2822.3425.1624.9425.4529.6127.3327.8628.4129.8330.24
SSIM↑0.9040.9050.9280.9250.9290.9640.9480.9500.9560.9660.970
LPIPS↓0.1480.1470.1170.1210.1100.0740.0930.0910.0840.0720.070
RealBRISQUE↓92.2573.8062.9462.7562.9361.5261.5060.8958.5858.6353.96
NIQE↓15.0012.0110.1710.2010.219.44010.009.9029.2869.0118.371
PIQE↓65.9252.7445.0344.8344.8446.7643.9543.5145.2744.6341.53
RankIQA↓9.4287.5426.4396.4116.4115.5736.2856.2235.6415.3204.895
MetaIQA↑0.1240.1420.1730.1740.1730.1810.1770.1790.1980.1910.213
" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.374, + 0.231, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.533, + 0.204, + 0.545 + ], + "angle": 0, + "content": "Blurry Input" + }, + { + "type": "image", + "bbox": [ + 0.231, + 0.375, + 0.364, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.279, + 0.532, + 0.328, + 0.544 + ], + "angle": 0, + "content": "\\(\\mathrm{E}^2\\mathrm{NeRF}\\)" + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.375, + 0.498, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.407, + 0.533, + 0.461, + 0.543 + ], + "angle": 0, + "content": "BAD-GS" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.375, + 0.633, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.533, + 0.588, + 0.543 + ], + "angle": 0, + "content": "BAGS" + }, + { + "type": "image", + "bbox": [ + 0.633, + 0.375, + 0.767, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.675, + 0.532, + 0.735, + 0.543 + ], + "angle": 0, + "content": "EvDNeRF" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.375, + 0.904, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.826, + 0.533, + 0.856, + 0.543 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.55, + 0.907, + 0.581 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on the real-world dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Compared with baselines, our method is capable of reconstructing shaper geometry from blurry inputs." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.594, + 0.486, + 0.837 + ], + "angle": 0, + "content": "are presented in Figure. 4 and Figure. 5, and the complete qualitative results are provided in the supplementary. In Figure. 4, it can be found that \\(\\mathrm{E}^2\\mathrm{NeRF}\\) [39] struggles in medium and large-scale scene modeling, producing significant blurring results. Additionally, BAD-GS [70] and BAGS [38] falls short in regions with significant color and depth variations, and produces overly smooth background textures. Although EvDNeRF [3] exhibits overall satisfactory performance, its complex network architecture prolongs the training time and precludes real-time rendering. In comparison, our method outperforms the baselines in producing high-fidelity novel views, and significantly reducing training time as well as demonstrating advantages in real-time applications. Figure. 5 further demonstrates that ours is capable of reconstructing detailed geometry from blurry inputs, achieving better performance than baselines." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.245, + 0.865 + ], + "angle": 0, + "content": "4.5. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Poses Optimization. We conduct experiments on the large-scale subset of synthetic data to evaluate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.594, + 0.907, + 0.716 + ], + "angle": 0, + "content": "the impact of pose optimization. As shown in Table 3, we calculate the ATE (Average Trajectory Error) between the initial COLMAP poses and GT poses, as well as the ATE between the optimized poses and GT poses. The results demonstrate the effectiveness of our method in correcting the initial poses and recovering more accurate camera trajectories. Additionally, the improved image quality metrics further demonstrate the effectiveness of pose optimization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Analysis of Number of Latent Camera Poses. We conduct ablations to investigate the effect of camera poses number in the exposure time, using the large-scale subset of the synthetic data. Specifically, we vary the number of latent camera poses \\( n \\) from 5, 9, 13, and 17, and visualize the quantitative novel view synthesis results in Figure. 6. It can be observed that the performance reaches a bottleneck at 9 latent poses. Beyond this point, further improvements are limited and may potentially lead to local convergence issues due to over-parameterization. Based on these experiments, we choose \\( n = 9 \\) camera poses to achieve a balance between rendering performance and training efficiency. Here," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24786" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.226, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.092, + 0.357, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.358, + 0.092, + 0.488, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.23, + 0.493, + 0.273 + ], + "angle": 0, + "content": "Figure 6. Ablation on number of latent camera poses. We vary the number of latent camera poses \\( n \\) from 5, 9, 13, and 17, and compare with BAD-NeRF [56] and BAD-GS [70] in each setting." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.092, + 0.642, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.092, + 0.769, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.77, + 0.092, + 0.9, + 0.22 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.502, + 0.23, + 0.905, + 0.272 + ], + "angle": 0, + "content": "Figure 7. Ablation on the impact of event threshold. We vary the threshold from \\( c = 0.1 \\) to \\( c = 0.4 \\). Our method shows robustness to varying event thresholds since it does not rely solely on event data." + }, + { + "type": "table_caption", + "bbox": [ + 0.14, + 0.287, + 0.854, + 0.3 + ], + "angle": 0, + "content": "Table 2. Ablation on the loss functions. Best-performing results are highlighted in bold and second results in underline." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.302, + 0.905, + 0.387 + ], + "angle": 0, + "content": "
Large-scaleMedium-scaleReal Scene
\\( \\mathcal{L}_{\\text{blur}} \\)\\( \\mathcal{L}_{\\text{event}} \\)\\( \\mathcal{L}_{\\text{depth}} \\)\\( \\mathcal{L}_{\\text{int}} \\)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓BRISQUE↓NIQE↓PIQE↓MetaIQA↑RankIQA↓
✓✘✘✘24.980.7870.29929.100.8950.16757.328.79943.260.1986.125
✓✓✘✘25.710.7950.27529.940.9100.15556.168.58242.510.2015.066
✓✓✓✘25.540.8000.27130.050.9140.15456.178.49942.230.2004.923
✓✓✓✓26.020.8060.26830.470.9160.15253.968.37141.530.2134.895
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.404, + 0.482, + 0.446 + ], + "angle": 0, + "content": "Table 3. Ablation on the effectiveness of pose optimization. We calculate ATE between the initial COLMAP poses and GT poses, as well as ATE between the optimized poses and GT poses." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.449, + 0.48, + 0.499 + ], + "angle": 0, + "content": "
ATE↓PSNR↑SSIM↑LPIPS↓
w/o pose. opt0.383 ± 0.12524.790.7880.293
w/ pose. opt0.049 ± 0.01126.020.8060.268
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.515, + 0.483, + 0.556 + ], + "angle": 0, + "content": "Table 4. Ablation of different motion blur levels. Mild blur corresponds to \\(0.5 \\times\\) camera motion speed, medium blur to the default speed, and strong blur to the \\(2.0 \\times\\) camera motion speed." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.56, + 0.476, + 0.62 + ], + "angle": 0, + "content": "
Blur levelMild (0.5×)Medium (1.0×)Strong (2.0×)
PSNR↑26.3825.7125.04
SSIM↑0.8160.7950.789
LPIPS↓0.2690.2750.280
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.733 + ], + "angle": 0, + "content": "we also provide a comparison with BAD-NeRF [56] and BAD-GS [70]. These two methods typically use linear interpolation to obtain camera trajectory, while our camera trajectories are estimated from the decomposed latent images, which provides more accurate initialization and helps our method achieves better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.734, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Effectiveness of The Loss Functions. We conduct novel view synthesis experiments on the challenging medium-scale and large-scale subsets of the synthetic dataset and the real-world dataset to validate the effectiveness of the training losses. The quantitative results are shown in Table. 2. It indicates that using only the blur reconstruction loss \\(\\mathcal{L}_{\\mathrm{blur}}\\) leads to suboptimal outputs. In contrast, incorporating \\(\\mathcal{L}_{\\mathrm{event}}, \\mathcal{L}_{\\mathrm{depth}}\\), and \\(\\mathcal{L}_{\\mathrm{int}}\\) enables our method to achieve better performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of Event Threshold. We conduct experiments on the medium-scale subset of the synthetic data to evaluate" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.406, + 0.905, + 0.481 + ], + "angle": 0, + "content": "the impact of different event thresholds. We vary the threshold from \\( c = 0.1 \\) to \\( c = 0.4 \\), and the results are shown in Figure 7. Since our method does not rely solely on event data and incorporates RGB inputs as supervision, it demonstrates robustness to varying event thresholds." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.906, + 0.662 + ], + "angle": 0, + "content": "Robustness to Motion Blur Levels. To validate the robustness of our method in handling different levels of motion blur, we set up three different camera speeds in the City Blocks scene of the synthetic dataset to obtain images with varying levels of blur. Images with mild blur are captured at \\(0.5 \\times\\) camera motion speed, images with medium blur are captured at the default motion speed, and images with strong blur are captured at \\(2.0 \\times\\) motion speed. The quantitative results of novel view synthesis are listed in Table. 4. It can be observed that the results show no significant fluctuations across different levels of blur, demonstrating the robustness of our method to varying motion blur levels." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.677, + 0.64, + 0.691 + ], + "angle": 0, + "content": "5. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.701, + 0.907, + 0.898 + ], + "angle": 0, + "content": "This paper introduces Event Stream Assisted Gaussian Splitting (EvaGaussians), a novel framework that seamlessly integrates the event streams captured by an event camera into the training of 3D-GS, effectively addressing the challenges of reconstructing high-quality 3D-GS from motion-blurred images. We contribute two novel datasets and conduct comprehensive experiments. The results demonstrate that our method outperforms previous state-of-the-art deblurring rendering techniques in terms of novel view synthesis quality, without sacrificing inference efficiency. Despite its promising performance, our method may still face challenges when reconstructing scenes with extremely intricate textures from severely blurred images." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24787" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.115, + 0.484, + 0.161 + ], + "angle": 0, + "content": "This work was supported in part by Natural Science Foundation of China (No.62332002, No.62202014), and Shenzhen KQTD (No.20240729102051063)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.173, + 0.188, + 0.189 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.199, + 0.484, + 0.24 + ], + "angle": 0, + "content": "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.242, + 0.484, + 0.31 + ], + "angle": 0, + "content": "[2] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame- and event-based davis sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686–689, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[3] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.436 + ], + "angle": 0, + "content": "[4] Mauro Comi, Alessio Tonioni, Max Yang, Jonathan Tremblay, Valts Blukis, Yijiong Lin, Nathan F Lepora, and Laurence Aitchison. Snap-it, tap-it, splat-it: Tactile-informed 3d gaussian splatting for reconstructing challenging surfaces. arXiv preprint arXiv:2403.20275, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.438, + 0.484, + 0.52 + ], + "angle": 0, + "content": "[5] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.522, + 0.483, + 0.59 + ], + "angle": 0, + "content": "[6] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2924–2932, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.592, + 0.483, + 0.647 + ], + "angle": 0, + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.483, + 0.702 + ], + "angle": 0, + "content": "[8] Haiqian Han, Jianing Li, Henglu Wei, and Xiangyang Ji. Event-3dgs: Event-based 3d reconstruction using 3d gaussian splatting. In Advances in Neural Information Processing Systems, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.483, + 0.746 + ], + "angle": 0, + "content": "[9] Philipp Heise, Sebastian Klose, Brian Jensen, and Alois Knoll. Pm-huber: Patchmatch with huber regularization for stereo matching. In ICCV, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[10] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[11] Jian Huang, Chengrui Dong, and Peidong Liu. Inceventgs: Pose-free gaussian splatting from a single event camera. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.199, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[13] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.175 + ], + "angle": 0, + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.178, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[15] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.219, + 0.905, + 0.272 + ], + "angle": 0, + "content": "[16] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.316 + ], + "angle": 0, + "content": "[17] Joseph J LaViola Jr. Bringing vr and spatial 3d interaction to the masses through video games. IEEE Computer Graphics and Applications, 2008. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.317, + 0.905, + 0.357 + ], + "angle": 0, + "content": "[18] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[19] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision, pages 416-434. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.415, + 0.905, + 0.469 + ], + "angle": 0, + "content": "[20] Yuchen Li*, Chaoran Feng*, Zhenyu Tang, Kaiyuan Deng, Wangbo Yu, Yonghong Tian, and Li Yuan. Gs2e: Gaussian splatting is an effective data generator for event stream generation. arXiv preprint arXiv:2505.15287, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[21] Guixu Lin, Jin Han, Mingdeng Cao, Zhihang Zhong, and Yinqiang Zheng. Event-guided frame interpolation and dynamic range expansion of single rolling shutter image. In Proceedings of the 31st ACM International Conference on Multimedia, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[22] Mingyuan Lin, Yangguang Wang, Xiang Zhang, Boxin Shi, Wen Yang, Chu He, Gui-song Xia, and Lei Yu. Self-supervised shutter unrolling with events. IJCV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.635 + ], + "angle": 0, + "content": "[23] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.637, + 0.905, + 0.692 + ], + "angle": 0, + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.693, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[25] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[26] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[27] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[28] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural Radiance Fields from Blurry Images. In CVPR, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24788" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[29] Nico Messikommer, Carter Fang, Mathias Gehrig, Giovanni Cioffi, and Davide Scaramuzza. Data-driven feature tracking for event cameras with and without frames. TPAMI, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 21(12):4695-4708, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.315 + ], + "angle": 0, + "content": "[33] Genady Paikin, Yotam Ater, Roy Shaul, and Evgeny Soloveichik. Efi-net: Video frame interpolation from fusion of events and frames. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[35] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In CVPR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.467 + ], + "angle": 0, + "content": "[36] Roberto Pellerito, Marco Cannici, Daniel Gehrig, Joris Bel-hadj, Olivier Dubois-Matra, Massimo Casasco, and Davide Scaramuzza. Deep visual odometry with events and frames. In IROS, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.523 + ], + "angle": 0, + "content": "[37] Cheng Peng and Rama Chellappa. Pdrf: Progressively deblurring radiance field for fast and robust scene reconstruction from blurry images. arXiv preprint arXiv:2208.08049, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.579 + ], + "angle": 0, + "content": "[38] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In ECCV, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.622 + ], + "angle": 0, + "content": "[39] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.623, + 0.482, + 0.664 + ], + "angle": 0, + "content": "[40] Yunshan Qi, Lin Zhu, Yifan Zhao, Nan Bao, and Jia Li. Deblurring neural radiance fields with event-driven bundle adjustment. In MM, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.482, + 0.706 + ], + "angle": 0, + "content": "[41] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.747 + ], + "angle": 0, + "content": "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[43] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, 2016. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[45] Zihang Shao*, Chaoran Feng*, Xuanye Fang, Yaxin Li, Jiangrong Shen, and Qi Xu. Eicil: joint excitatory inhibitory cycle iteration learning for deep spiking neural networks. Advances in Neural Information Processing Systems, 36: 32117-32128, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.218 + ], + "angle": 0, + "content": "[47] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.289 + ], + "angle": 0, + "content": "[48] Zhenyu Tang*, Chaoran Feng*, Xinhua Cheng, Wangbo Yu, Junwu Zhang, Yuan Liu, Xiaoxiao Long, Wenping Wang, and Li Yuan. Neuralgs: Bridging neural fields and 3d gaussian splatting for compact 3d representations. arXiv preprint arXiv:2503.23162, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[49] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7320–7328, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.375, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[50] Gemma Taverni. Applications of Silicon Retinas: From Neuroscience to Computer Vision. PhD thesis, Universität Zürich, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, pages 16155-16164, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[52] Manoj K Vairalkar and SU Nimbhorkar. Edge detection of images using sobel operator. International Journal of Emerging Technology and Advanced Engineering, 2(1):291-293, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.532, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[53] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.602, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[54] Zhexiong Wan, Bin Fan, Le Hui, Yuchao Dai, and Gim Hee Lee. Instance-level moving object segmentation from a single image with events. IJCV, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[55] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. In ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[56] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[57] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[58] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[59] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24789" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[60] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, et al. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[61] Wenhao Xu, Wenming Weng, Yueyi Zhang, Ruikang Xu, and Zhiwei Xiong. Event-boosted deformable 3d gaussians for fast dynamic scene reconstruction. arXiv preprint arXiv:2411.16180, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[62] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[63] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[64] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[65] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[66] Sohaib Zahid, Viktor Rudnev, Eddy Ilg, and Vladislav Golyanik. E-3dgs: Event-based novel view rendering of large-scale scenes using 3d gaussian splatting. In 3DV, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.504, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.575, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[68] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[69] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[70] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In ECCV, 2024. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[71] Haiyang Zhou, Wangbo Yu, Jiawen Guan, Xinhua Cheng, Yonghong Tian, and Li Yuan. Holotime: Taming video diffusion models for panoramic 4d scene generation. arXiv preprint arXiv:2504.21650, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.801 + ], + "angle": 0, + "content": "[72] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[73] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143-14152, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[74] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "feys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[75] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. TPAMI. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.163 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24790" + } + ] +] \ No newline at end of file diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_origin.pdf b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..db8f1c833c1403307cd8e0797ab3538262c139c7 --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/4ff03f6c-2177-47f2-8440-fcfd6f8d42f1_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ceb225c2c5176828f06cc1c539f70357a9c17936961ae49132ec9c2c92e7f7ed +size 5433155 diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/full.md b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/full.md new file mode 100644 index 0000000000000000000000000000000000000000..2de8bb3d4b01da4e33812f5ded84a579743b7f27 --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/full.md @@ -0,0 +1,355 @@ +# EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images + +Wangbo Yu $^{1,2*}$ , Chaoran Feng $^{1*}$ , Jianing Li $^{3}$ , Jiye Tang $^{4}$ , Jiashu Yang $^{5}$ , Zhenyu Tang $^{1}$ , Meng Cao $^{6}$ , Xu Jia $^{5}$ , Yuchao Yang $^{1}$ , Li Yuan $^{1,2\dagger}$ and Yonghong Tian $^{1,2\dagger}$ + +1Shenzhen Graduate School, Peking University, 2Peng Cheng Laboratory, 3School of Computer and Science, Peking University +4University of Science and Technology of China, 5Dalian University of Technology, 6MBZUAI {wbyu, chaoran.feng}@stu.pku.edu.cn, {yuanli-ece, yhtian}@pu.edu.cn + +![](images/2152ac3dce439fe973cb9b3ed60d1974db601256af66caec9056d6bd2e562052.jpg) + +![](images/24e6ee784102928c42531161649139aedb027a483b80747d64d122673ca06bd3.jpg) + +![](images/18d76e1fc6782a46660d4a8bcaa50afb52f8d981e1aa95d910fd71f7c55782b4.jpg) + +![](images/55c6a84091b04e4bdf398e896c20b4edcf348420a70a152111f4cb5fd9a2a731.jpg) +Figure 1. EvaGaussians integrates frame-based cameras and event cameras to reconstruct sharp 3D-GS from motion-blurred images. It achieves higher-quality novel view synthesis compared to previous methods. + +![](images/fa21a3ac57b6ab17d068127ee646ece87416a594ba078a2c2658559e1a02001a.jpg) + +![](images/a0c2082c7dba596bd96e1068e01cb9d18deccdf200263997b911f35dee4cd5bb.jpg) + +![](images/474585339f000993377c94f4c229b1fc1ca6be93d9212381ef00bfb036f6ebcd.jpg) + +![](images/01cc40e305b7342c95bf0557567eca0f29e05280db0fa380ad75d0aad95ed5f0.jpg) + +# Abstract + +3D Gaussian Splatting (3D-GS) has demonstrated exceptional capabilities in synthesizing novel views of 3D scenes. However, its training is heavily reliant on high-quality images and precise camera poses. Meeting these criteria can be challenging in non-ideal real-world conditions, where motion-blurred images frequently occur due to high-speed camera movements. To address these challenges, we introduce Event Stream Assisted Gaussian Splatting (EvaGaus-sians), a novel approach that harnesses event streams captured by event cameras to facilitate the learning of high-quality 3D-GS from motion-blurred images. Capitalizing on the high temporal resolution and dynamic range offered by event streams, we seamlessly integrate them into the initialization and optimization of 3D-GS, thereby enhancing the acquisition of high-fidelity novel views with intricate texture details. We also contribute two novel datasets comprising RGB frames, event streams, and corresponding camera parameters, featuring a wide variety of scenes + +and various camera motions. The comparison results reveal that our approach not only excels in generating high-fidelity novel views, but also offers improved training and inference efficiency. Video and code are available at the project page. + +# 1. Introduction + +Novel view synthesis from 2D image collections has presented a persistent challenge within the field of computer vision and computer graphics. This task stands as a fundamental component in various vision applications, such as virtual reality [17, 60], robotics navigation [42, 62, 74], scene understanding [15, 24, 25], and many others, thereby prompting significant research efforts over the last decades. Amid pioneering works, 3D Gaussian Splatting (3D-GS) [14] achieves notable success in generating high-fidelity novel views. However, the optimization of 3D-GS heavily relies on accurate camera poses and point cloud initialization produced by COLMAP [44], which necessitates high-quality images without blurring. Fulfilling such conditions can be challenging in real-world situations. For example, in UAVs and robotics, rapid camera movement is + +common when capturing images or recording videos, which often result in significant motion blur. + +Recent studies have demonstrated the significant potential of event-based cameras in alleviating motion blur in images captured by conventional frame-based cameras. [13, 23, 34]. Serving as an innovative bio-inspired visual sensor, event cameras asynchronously report the logarithmic intensity changes of each pixel captured, and can record higher temporal resolution and dynamic range data in contrast to conventional cameras. The continuous event streams naturally align with the frame-rate agnostic rendering property of 3D-GS, offering sharper scene geometry information than frame-based cameras, making it highly promising in facilitating 3D-GS training from fast camera motion. + +Motivated by this, we propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first method that seamlessly incorporates continuous event streams captured by event cameras to enhance the learning of high-quality 3D-GS from motion-blurred images. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Furthermore, to tackle the geometric ambiguity caused by blurry images, we introduce two event-assisted geometry regularization terms to stabilize the geometry of 3D-GS. Through optimizing the 3D-GS in a progressive manner, our method can recover a high-quality 3D-GS that facilitates the real-time generation of high-fidelity novel views. Finally, to comprehensively evaluate novel view synthesis with hybrid frame- and event-based cameras, we contribute a synthetic dataset and a real-world dataset, both containing diverse scenes with varying scales and camera motions, setting a benchmark for related researches. To summarize, our contributions can be delineated as follows: + +- We propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first framework designed to reconstruct high-quality 3D-GS from motion-blurred images with the assistance of event cameras. It not only recovers intricate details from the motion-blurred images but also enables high-fidelity novel view synthesis. +- We contribute a synthetic dataset and a real-world dataset for evaluating novel view synthesis with hybrid frame-and event-based cameras. Both datasets contain diverse scenes with varying scales and camera motions. +- We conduct a comprehensive evaluation of the proposed method and compare it with several strong baselines. As shown in Figure. 1 and Figure. 2, our method not only excels in generating higher-fidelity novel views but also enables improved training and inference efficiency. + +![](images/2763364a299093d5f3e4b8de97327ad8b5e3773cd95150f44f71090338857d6a.jpg) +Figure 2. Performance, rendering speed, and training time (indicated by the area of each circle) comparison. EvaGaussians achieves better results while maintaining real-time rendering and training efficiency. + +# 2. Related Works + +# 2.1. Reconstructing 3D Scene from Blurry Images + +Reconstructing a high-quality 3D Scene typically requires high-fidelity, sharp images as supervision. However, motion-blurred images often occur in real world scenarios, thus hindering accurate reconstruction of 3D scenes. Several studies have been proposed to address this issue. For example, Deblur-NeRF [28], DP-NeRF [18] and PDRF [37] attempted to learn a blur formation kernel to model the image blurring process. BAD-NeRF [56] further physically modeled the blurry images formation process, and adopted a bundle-adjustment strategy to jointly optimize NeRF parameters and the camera poses during the exposure time. These NeRF-based methods lacked real-time rendering capabilities and suffered from extended training times. With the rapid advancement of 3D-GS, concurrent works like BAD-Gaussians [70] and BAGS [38] proposed to utilize 3D-GS as representation to achieve deblurring reconstruction. Although they achieved real-time rendering and faster convergence compared with prior works, they still struggled to handle severely blurred images in which COLMAP [44] will fail to produce the initial point clouds and camera calibration. + +# 2.2. Reconstructing 3D Scene from Event Streams + +Motivated by the exceptional properties offered by event cameras [1, 2, 6, 7, 20-22, 29, 33, 35, 36, 45, 47-49, 51, 54, 58, 61, 68, 72, 75], several studies attempted to reconstruct 3D scenes from event streams captured by event cameras, particularly in low-light conditions with fast camera mo + +tion [8, 11, 12, 27, 43, 55, 59, 61, 63-66, 69, 71]. These methods were mostly designed to recover scene geometry from event streams, and are not typically designed for reconstructing 3D Scene from blurry images. Recently, E-NeRF [16], $\mathrm{E}^2\mathrm{NeRF}$ [39], EBAD-NeRF [40], BENeRF [19] and EvDNeRF [3] proposed to jointly utilize event streams captured by event cameras and motion-blurred images captured by standard frame-based cameras to reconstruct a NeRF representation. In contrast to frame-based deblurring methods, they are better at handling motion blur due to the introduction of event information. However, these NeRF-based methods suffer from long training and inference times, which limit their further application. + +# 3. Method + +# 3.1. Preliminary + +Event camera is a type of bio-inspired sensor that can asynchronously record intensity changes [7]. In contrast to conventional cameras that are restricted to sequentially produce frames at a fixed frame rate, event cameras asynchronously trigger events in each pixel when their intensity change exceeds a constant threshold, featuring properties such as low latency and high dynamic range. Formally, let $\mathbf{I}_{xy}(t)$ denote the instantaneous intensity at pixel coordinate $(x,y)$ at time $t$ , and $\mathbf{L}_{xy}(t)$ denotes its logarithm. An event $p = \pm 1$ will be triggered whenever the change of $\mathbf{L}_{xy}(t)$ surpasses the threshold $c$ , where the polarity represents the direction (increase or decrease) of changes. Let $\delta_{t_0}(t)$ be the impulse function at time $t_0$ with a unit integral, the event can therefore be expressed as a continuous-time signal $\mathbf{e}_{xy}(t) = p\delta_{t_0}(t)$ , where $t_0$ signifies the time at which the event occurs. Then, the proportional intensity change during a time interval $[s,t]$ can be computed as the integral of events that occurred between times $s$ and $t$ , expressed as $\mathbf{E}_{xy}(t) = \int_s^t\mathbf{e}_{xy}(h)dh$ . Given that each pixel can be treated separately in the event camera, the subscripts can be omitted: + +$$ +\mathbf {E} (t) = \int_ {s} ^ {t} \mathbf {e} (h) d h. \tag {1} +$$ + +We can then represent the logarithmic intensity change as: $\mathbf{L}(t) - \mathbf{L}(s) = c\mathbf{E}(t)$ , rewrite as $\mathbf{L}(t) = \mathbf{L}(s) + c\mathbf{E}(t)$ , and subsequently obtain the actual intensity change: + +$$ +\mathbf {I} (t) = \mathbf {I} (s) \cdot \exp (c \mathbf {E} (t)). \tag {2} +$$ + +Therefore, when an image $\mathbf{I}(s)$ is captured at time $s$ , and the event stream is recorded during the time interval $[s,t]$ , the image $\mathbf{I}(t)$ can be obtained by warping $\mathbf{I}(s)$ using Eq. 2. + +# 3.2. Event-assisted Initialization + +The optimization of 3D-GS requires camera calibration and point cloud initialization using COLMAP [44]. However, + +as shown in Table.4 in the supplementary, this process can fail when dealing with images that have significant motion blur. Motion-blurred images are resulted from camera movements during the exposure time, which can be mathematically represented as: + +$$ +\mathbf {B} = \frac {1}{\tau} \int_ {s - \tau / 2} ^ {s + \tau / 2} \mathbf {I} (t) d t, \tag {3} +$$ + +where $\mathbf{B}$ denotes a captured blurry image, which is equivalent to averaging the instantaneous latent images $\mathbf{I}(t)$ during the exposure time $[s - \tau /2,s + \tau /2]$ . + +To obtain initial camera poses and point clouds for 3D-GS optimization, we first preprocess the motion-blurred images using the Event-based Double Integral (EDI) [34] model, which can be derived through substituting Eq. 2 into Eq. 3: + +$$ +\mathbf {B} = \mathbf {I} (s) \cdot \frac {1}{\tau} \int_ {s - \tau / 2} ^ {s + \tau / 2} \exp (c \mathbf {E} (t)) d t. \qquad (4) +$$ + +Given the predefined threshold $c$ , a blurry image $\mathbf{B}$ , and the recorded event stream $\mathbf{E}(t)$ , the EDI model (Eq. 4) allows the derivation of $\mathbf{I}(s)$ , following which the latent image $\mathbf{I}(t)$ at any moment within the exposure time can be estimated through Eq. 2. As shown in Figure. 3, given a total of $K$ blurry images $\{\mathbf{B}^j\}_{j=1}^K$ , for each of them, we uniformly sample $n$ time stamps during their exposure time to obtain a series of EDI-processed latent images rich in texture features, denoted as $\{\mathbf{I}_i\}_{i=1}^n$ , then obtain their poses $\{\mathbf{P}_i\}_{i=1}^n$ and the initial point cloud of the scene using COLMAP [44]. + +After initialization, a straightforward approach to optimizing the 3D-GS is to use the EDI-processed latent images and their camera poses as supervision. However, although these latent images provide more texture features than the original blurry image, they still do not fully recover the ideal latent image and exhibit relatively low visual quality, which also introduces inaccuracies into the camera poses, thereby leading to unsatisfactory optimization results. To more robustly recover a sharp 3D-GS from motion-blurred images, we propose to harness the advantages of event streams and seamlessly integrate them into the optimization process of 3D-GS. + +# 3.3. Event-assisted Bundle-Adjustment + +As introduced in Eq. 3, during the exposure time, a motion-blurred image can be decomposed into a series of latent images along a specific camera trajectory, which can be roughly approximated by the EDI-produced camera poses $\{\mathbf{P}_i\}_{i=1}^n$ according to Eq. 4. Motivated by this, we jointly optimize these camera poses and the 3D-GS attributes in a bundle-adjustment manner [56] to simultaneously recover the blur-formation camera trajectories and a sharp 3D-GS. As shown in Figure. 3, we add each of the EDI-produced + +![](images/416f929965d0e231266637a7b3841749bd7acc39b560a0de43f9ee4dcccb8017.jpg) +Figure 3. Overview of EvaGaussians. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Leveraging the continuously recorded event streams, we introduce two event-assisted geometry regularization terms that can function beyond the exposure time to stabilize the geometry of 3D-GS. Once trained, it can not only recover intricate details from motion-blurred images but also enable high-fidelity novel view synthesis. + +camera poses a learnable offset $\{\mathbf{d}_i\}_{i = 1}^n$ as correction parameters, resulting a learnable camera trajectory $\{\widetilde{\mathbf{P}}_i\}_{i = 1}^n$ where $\widetilde{\mathbf{P}}_i = \mathbf{P}_i + \mathbf{d}_i$ . In each training iteration, we simultaneously render $n$ images $\{\widetilde{\mathbf{I}}_i\}_{i = 1}^n$ from the 3D-GS along the camera trajectory of the blurry view, and then simulate the formation of the motion-blurred images using a discrete approximation of Eq 3, denoted as $\widetilde{\mathbf{B}} = \frac{1}{n}\sum_{i = 1}^{n}\widetilde{\mathbf{I}}_i$ . Consequently, for a total of $K$ real-captured blurry images $\{\mathbf{B}^j\}_{j = 1}^K$ , we can obtain their simulated versions $\{\widetilde{\mathbf{B}}^j\}_{j = 1}^K$ through each corresponding learnable camera trajectory. + +Blur Reconstruction Loss. With the simulated blurry images, we use the captured real blurry images $\{\mathbf{B}^j\}_{j=1}^K$ to serve as image level supervision. Specifically, for each blurry image $\mathbf{B}^j$ and its simulated version $\tilde{\mathbf{B}}^j$ , we employ a blur reconstruction loss to minimize their photometric error, expressed as + +$$ +\mathcal {L} _ {b l u r} = \left(1 - \lambda_ {1}\right) \cdot \| \mathbf {B} ^ {j} - \widetilde {\mathbf {B}} ^ {j} \| _ {1} + \lambda_ {1} \cdot \mathrm {D - S S I M} \left(\mathbf {B} ^ {j}, \widetilde {\mathbf {B}} ^ {j}\right). (5) +$$ + +The formulation of blur reconstruction loss is the same as in the original 3D-GS [14], it differs in utilizing blurry images as supervision and jointly optimizing the 3D-GS attributes and the camera trajectories, thus facilitating an initial deblurring reconstruction of 3D-GS. + +Event Reconstruction Loss. Leveraging the abundant high-frequency information offered by the event streams, we further adopt an event reconstruction loss to aid in 3D-GS optimization. Specifically, we uniformly divide the exposure time into $m = n - 1$ intervals, each with a duration of $\frac{\tau}{m}$ . Subsequently, we integrate the recorded event stream along these time intervals using Eq. 1, resulting in + +$m$ event maps $\{\mathbf{E}_i\}_{i=1}^m$ to serve as event level supervision. During training, for the $j$ -th blurry view, we convert the rendered image sequence $\{\widetilde{\mathbf{I}}_i\}_{i=1}^n$ on the camera trajectory into event maps $\{\widetilde{\mathbf{E}}_i\}_{i=1}^m$ , using a differentiable event simulator [10, 41], and constrain the discrepancies between the simulated event maps and the ground truth event maps, expressed as: + +$$ +\mathcal {L} _ {\text {e v e n t}} = \frac {1}{m} \sum_ {i = 1} ^ {m} \| \mathbf {E} _ {i} - \widetilde {\mathbf {E}} _ {i} \| _ {1}. \tag {6} +$$ + +The event reconstruction loss further aids in recovering a sharp 3D-GS with improved texture details. + +# 3.4. Event-assisted Geometry Regularization + +The blurry color images are captured only during the exposure time and are much sparser than the event stream. Relying on such low-quality image-level supervision may cause the 3D-GS to overfit on the training images, resulting in significant floaters and inferior geometry. Leveraging Eq. 1 and Eq. 4, given the continuously recorded event streams $\mathbf{E}(t)$ , we can derive continuous grayscale intensity images $\mathbf{G}(t)$ that are rich in geometric information and can function beyond the exposure time. Motivated by this, we further propose two event-assisted geometry regularization terms to aid in 3D-GS training. + +Intensity Reconstruction Loss. During training, we randomly sample continuous time $t$ between the interval of two adjacent blurry image, and derive the grayscale intensity image $\mathbf{G}(t)$ using Eq. 2. We then minimize the difference + +between it and the rendered intensity image from 3D-GS, expressed as: + +$$ +\mathcal {L} _ {i n t} = (1 - \lambda_ {2}) \cdot \| \mathbf {G} (t) - \widetilde {\mathbf {G}} (t) \| _ {1} + \lambda_ {2} \cdot D - S S I M (\mathbf {G} (t), \widetilde {\mathbf {G}} (t)), \tag {7} +$$ + +where $\widetilde{\mathbf{G}}(t)$ is converted from the colored render result. + +Intensity-aware Depth Regularization Loss. As shown in Figure. 3, to further improve the geometry of 3D-GS, inspired by [4, 9], we adopt an intensity-aware depth regularization loss during training, defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {d e p t h}} = \frac {1}{N} \sum_ {x, y} \left(\left| \partial_ {x} \widetilde {\mathbf {D}} _ {x y} (t) \right| e ^ {- \beta \left| \partial_ {x} \mathbf {G} _ {x y} (t) \right|} \right. \tag {8} \\ + | \partial_ {y} \widetilde {\mathbf {D}} _ {x y} (t) | e ^ {- \beta | \partial_ {y} \mathbf {G} _ {x y} (t) |}), \\ \end{array} +$$ + +where $\widetilde{\mathbf{D}}(t)$ is the rendered depth map, $(x,y)$ denotes the pixel location, $N$ is the total number of pixels, and $\beta$ is set to 2 in our experiments. The horizontal and vertical gradients are calculated by applying convolution operations with $5 \times 5$ Sobel kernels [52]. This regularization is founded on the observation that depth transitions in an image often correspond to changes in intensity. Therefore, it ensures that the spatial variation of depth closely matches that of the intensity image, thereby reducing geometric artifacts at object boundaries. + +The total loss function is the combination of the above losses, defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {b l u r}} \mathcal {L} _ {\text {b l u r}} + \lambda_ {\text {e v e n t}} \mathcal {L} _ {\text {e v e n t}} \tag {9} \\ + \lambda_ {i n t} \mathcal {L} _ {i n t} + \lambda_ {d e p t h} \mathcal {L} _ {d e p t h}. \\ \end{array} +$$ + +# 4. Experiments + +In this section, we first provide the implementation details of our method in 4.1, followed by an illustration of the introduced dataset in 4.2, and the experimental settings in 4.3. The comparison results and ablation study are presented in 4.4 and 4.5. For more detailed experiments. + +# 4.1. Implementation Details + +Progressive Training. We implemented EvaGaussians based on the official code of 3D-GS [14]. The training process spans 50,000 iterations, with an event reconstruction loss introduced after a 3,000-iteration warmup and we omit the densification process to streamline and simplify the subsequent optimization. Additionally, we adopt a coarse-to-fine training strategy, starting with rendering at a low resolution (0.3× downsampling in the early 30% iterations) and progressively increasing the size of the rendered views to full resolution. All experiments were conducted using a single NVIDIA RTX 4090 GPU. + +Hyperparameter Setting. During the training process, we set $\lambda_{blur} = 1.0$ , $\lambda_{depth} = 1.0e^{-2}$ , $\lambda_{event} = 5.0e^{-3}$ and + +$\lambda_{int} = 1.0e^{-3}$ for the loss function, and used $n = 9$ for the number of poses to be optimized during the exposure time. In implementing the loss $\mathcal{L}_{event}$ , we configured the positive threshold as $c_{pos} = 0.25$ and the negative threshold as $c_{neg} = 0.25$ for synthetic scenes, and set $c_{pos} = 0.197$ and $c_{neg} = 0.241$ for real scenes. + +# 4.2. Datasets + +To enable a comprehensive evaluation, we introduce two novel datasets, with an overview provided below. Additionally, we present a detailed dataset introduction in the supplementary and will release the dataset construction pipeline for future research. + +EvaGaussians-Blender Dataset. We construct a synthetic dataset covering a variety of scene scales, coupling with diverse camera trajectories and event data. It consists 6 large-scale scenes, 7 medium-scale scenes, and 6 object-level scenes. We simulate motion blur by manually placing multi-view cameras, randomly adjusting camera poses, and performing linear interpolation between the original and perturbed positions for each view. The corresponding event streams are simulated using ESIM [41] and V2E [10]. + +EvaGaussians-DAVIS Dataset. We manually recorded five real-world scenes using the Color DAVIS346 event camera [50]. The dataset includes three object-level scenes and two indoor scenes. After processing, the final dataset consists of 30 images per scene, along with the recorded event streams, each showcasing various blur and lighting conditions. + +Comparison with Existing Datasets. The goal of contributing this dataset is to create a more comprehensive evaluation benchmark for event-assisted deblurring reconstruction tasks. While E2NeRF [39] and EvDNeRF [3] have introduced similar datasets, we observed that these datasets are limited in terms of scene variety, camera motion diversity, and data size. As shown in Table.1 and Table.2 of the supplementary, our proposed dataset includes a broader range of scenes with various scales, making it more suitable for comprehensive evaluation. + +# 4.3. Experiment Settings + +Baselines. We compare our method with three types of baselines: 1) NeRF [30] and 3D-GS [14] directly trained on the blurry images, referring to as B-NeRF and B-3DGS. 2) Deblur rendering methods, including BAD-NeRF [56], BAD-GS [70], BAGS [38], $\mathrm{E}^2\mathrm{NeRF}$ [39], and EvDNeRF [3]. Among these, BAD-NeRF and BAD-GS simulate motion blur and optimize camera trajectories without using event stream, while BAGS addresses motion blur through kernel modeling, also without event stream. In contrast, $\mathrm{E}^2\mathrm{NeRF}$ and EvDNeRF are event-assisted methods without camera trajectory optimization. 3) Image deblur methods, including UFP [5] (single-image deblurring), + +![](images/8b2d07830431f64054add0d7b8d3d90ba653c374e34190d80d65ea35b64c1085.jpg) +Figure 4. Qualitative comparison on the synthetic dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Full results are provided in the supplementary. + +EDI [34] (event-based deblurring), and EFNet [46] (learnable event-based deblurring). We process input blurry images with them and train the vanilla 3D-GS with predeblurred images. The resulting baselines are referred to as UFP-GS, EDI-GS, and EFN-GS. + +Evaluation Metrics. For synthetic datasets, we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [57], and VGG-based Learned Perceptual Image Patch Similarity (LPIPS) [67] to evaluate the similarity between rendered novel views and ground-truth novel views. For real-world datasets, since the sharp ground-truth images are unavailable, we utilize several No-Reference Image Quality Assessment (NR-IQA) metrics for evaluation, including BRISQUE [31], NIQE [32], PIQE [53], RankIQA [26], and MetaIQA [73], which allow for image evaluation when lacking ground truth images. + +# 4.4. Comparisons + +We evaluate our approach across a variety of scenes, including large-scale scenes, medium-scale scenes, object-level scenes and real scenes. Quantitative assessments of + +novel view synthesis are shown in Table. 1. It can be found that our method achieves substantial improvements in most of the metrics, especially in challenging large-scale scenes. Specifically, both B-NeRF and B-3DGS produce blurry novel views since they are directly trained on motion-blurred images. The image deblurring-based baselines, UFP-GS, EDI-GS and EFN-GS, also produced inferior results, because the image deblurring process potentially corrupts the 3D consistency of the training images. Notably, our approach outperforms BAD-NeRF [56], BAD-GS [70] and BAGS [38], due to their limited capability in modeling complex textures. In addition, our method also surpasses the event-assisted methods $\mathrm{E}^2\mathrm{NeRF}$ [39] and EvDNeRF [3] in producing high-quality novel views with intricate details, with better training and rendering efficiency. Notably, since directly using COLMAP on the blurry images fails in most cases (as shown in Table.4 of the supplementary), we use the EDI-produced poses for all the comparison baselines and our method to ensure a fair comparison. + +The qualitative results of five top-performing methods + +Table 1. Quantitative comparisons of novel view synthesis across large-scale, medium-scale, object-level, and real-world scenes. The table reports the average performance for each scale, demonstrating that our method consistently surpasses previous state-of-the-art approaches across all metrics. Best-performing results are highlighted in bold and second-best results in underline. + +
Scene TypeMetricsB-NeRFB-3DGSUFP-GS [5]EDI-GS [34]EFN-GS [46]E2NeRF [39]BAD-NeRF [56]BAD-GS [70]BAGS [38]EvDNeRF [3]Ours
LargePSNR↑21.3321.4821.3622.3122.6922.9623.8523.8624.9324.6326.02
SSIM↑0.6780.6880.6600.6860.6830.7070.7320.7330.7450.7530.806
LPIPS↓0.4250.3970.3740.3820.3630.3750.3480.3470.3270.3280.268
MediumPSNR↑24.0824.8026.3826.4426.1327.7828.4428.4628.7428.9130.47
SSIM↑0.7170.7510.8020.8010.7980.8660.8790.8730.8740.8850.916
LPIPS↓0.3620.3190.2640.2580.2730.1990.1820.1820.1740.1690.152
ObjectPSNR↑22.2822.3425.1624.9425.4529.6127.3327.8628.4129.8330.24
SSIM↑0.9040.9050.9280.9250.9290.9640.9480.9500.9560.9660.970
LPIPS↓0.1480.1470.1170.1210.1100.0740.0930.0910.0840.0720.070
RealBRISQUE↓92.2573.8062.9462.7562.9361.5261.5060.8958.5858.6353.96
NIQE↓15.0012.0110.1710.2010.219.44010.009.9029.2869.0118.371
PIQE↓65.9252.7445.0344.8344.8446.7643.9543.5145.2744.6341.53
RankIQA↓9.4287.5426.4396.4116.4115.5736.2856.2235.6415.3204.895
MetaIQA↑0.1240.1420.1730.1740.1730.1810.1770.1790.1980.1910.213
+ +![](images/7692f256e148173b3bd5b7166d6e45370dad6c37f0859613432644ff851831e4.jpg) +Blurry Input + +![](images/61362ef85eca0d2cc84d792f893ae4a872608c25a4441972fe6d66b8d991f144.jpg) +$\mathrm{E}^2\mathrm{NeRF}$ + +![](images/e2b9482e2f3521bb08dd1c83bc3b86d250021b668fd6d869ecf996743f93e701.jpg) +BAD-GS +Figure 5. Qualitative comparison on the real-world dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Compared with baselines, our method is capable of reconstructing shaper geometry from blurry inputs. + +![](images/10c45d0baf38537598bef3cb3109049782f3b912bea66e564f44d08d685dd7b1.jpg) +BAGS + +![](images/d7f51291244177e455030f2a6884167696e734244f327adf99b78733de49a27e.jpg) +EvDNeRF + +![](images/86971b530a316f0511cc918201512d63acc49306abe955665146864bd54360e8.jpg) +Ours + +are presented in Figure. 4 and Figure. 5, and the complete qualitative results are provided in the supplementary. In Figure. 4, it can be found that $\mathrm{E}^2\mathrm{NeRF}$ [39] struggles in medium and large-scale scene modeling, producing significant blurring results. Additionally, BAD-GS [70] and BAGS [38] falls short in regions with significant color and depth variations, and produces overly smooth background textures. Although EvDNeRF [3] exhibits overall satisfactory performance, its complex network architecture prolongs the training time and precludes real-time rendering. In comparison, our method outperforms the baselines in producing high-fidelity novel views, and significantly reducing training time as well as demonstrating advantages in real-time applications. Figure. 5 further demonstrates that ours is capable of reconstructing detailed geometry from blurry inputs, achieving better performance than baselines. + +# 4.5. Ablation Study + +Effectiveness of Poses Optimization. We conduct experiments on the large-scale subset of synthetic data to evaluate + +the impact of pose optimization. As shown in Table 3, we calculate the ATE (Average Trajectory Error) between the initial COLMAP poses and GT poses, as well as the ATE between the optimized poses and GT poses. The results demonstrate the effectiveness of our method in correcting the initial poses and recovering more accurate camera trajectories. Additionally, the improved image quality metrics further demonstrate the effectiveness of pose optimization. + +Analysis of Number of Latent Camera Poses. We conduct ablations to investigate the effect of camera poses number in the exposure time, using the large-scale subset of the synthetic data. Specifically, we vary the number of latent camera poses $n$ from 5, 9, 13, and 17, and visualize the quantitative novel view synthesis results in Figure. 6. It can be observed that the performance reaches a bottleneck at 9 latent poses. Beyond this point, further improvements are limited and may potentially lead to local convergence issues due to over-parameterization. Based on these experiments, we choose $n = 9$ camera poses to achieve a balance between rendering performance and training efficiency. Here, + +![](images/8520e65798fe81ba266bfac77956743714ac18ad0463c960ba8d622479411ba5.jpg) +Figure 6. Ablation on number of latent camera poses. We vary the number of latent camera poses $n$ from 5, 9, 13, and 17, and compare with BAD-NeRF [56] and BAD-GS [70] in each setting. + +![](images/02ae19ec502b17aa1e5decd7a117e525e836373aff07661ff03dc115010f0e05.jpg) + +![](images/feef9527cf143b5b60abee2f8a7c77dac8abea66dca88c197bb097a610994dc0.jpg) +Figure 7. Ablation on the impact of event threshold. We vary the threshold from $c = 0.1$ to $c = 0.4$ . Our method shows robustness to varying event thresholds since it does not rely solely on event data. + +![](images/ac827d0b3c523bcb7cc253c2008c57e4ccc5f1d5d2ab6ee4f9620d4ce5958751.jpg) + +![](images/0776ba992b5af5d7dc960cde74a694849fcdd9546f9a55743a4414e0ccc811f9.jpg) + +![](images/68aa3348edbb3046bf37f4fea86a3ad5915805725438bae75a567827bc1adc3f.jpg) + +Table 2. Ablation on the loss functions. Best-performing results are highlighted in bold and second results in underline. + +
Large-scaleMedium-scaleReal Scene
\( \mathcal{L}_{\text{blur}} \)\( \mathcal{L}_{\text{event}} \)\( \mathcal{L}_{\text{depth}} \)\( \mathcal{L}_{\text{int}} \)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓BRISQUE↓NIQE↓PIQE↓MetaIQA↑RankIQA↓
✓✘✘✘24.980.7870.29929.100.8950.16757.328.79943.260.1986.125
✓✓✘✘25.710.7950.27529.940.9100.15556.168.58242.510.2015.066
✓✓✓✘25.540.8000.27130.050.9140.15456.178.49942.230.2004.923
✓✓✓✓26.020.8060.26830.470.9160.15253.968.37141.530.2134.895
+ +Table 3. Ablation on the effectiveness of pose optimization. We calculate ATE between the initial COLMAP poses and GT poses, as well as ATE between the optimized poses and GT poses. + +
ATE↓PSNR↑SSIM↑LPIPS↓
w/o pose. opt0.383 ± 0.12524.790.7880.293
w/ pose. opt0.049 ± 0.01126.020.8060.268
+ +Table 4. Ablation of different motion blur levels. Mild blur corresponds to $0.5 \times$ camera motion speed, medium blur to the default speed, and strong blur to the $2.0 \times$ camera motion speed. + +
Blur levelMild (0.5×)Medium (1.0×)Strong (2.0×)
PSNR↑26.3825.7125.04
SSIM↑0.8160.7950.789
LPIPS↓0.2690.2750.280
+ +we also provide a comparison with BAD-NeRF [56] and BAD-GS [70]. These two methods typically use linear interpolation to obtain camera trajectory, while our camera trajectories are estimated from the decomposed latent images, which provides more accurate initialization and helps our method achieves better performance. + +Effectiveness of The Loss Functions. We conduct novel view synthesis experiments on the challenging medium-scale and large-scale subsets of the synthetic dataset and the real-world dataset to validate the effectiveness of the training losses. The quantitative results are shown in Table. 2. It indicates that using only the blur reconstruction loss $\mathcal{L}_{\mathrm{blur}}$ leads to suboptimal outputs. In contrast, incorporating $\mathcal{L}_{\mathrm{event}}, \mathcal{L}_{\mathrm{depth}}$ , and $\mathcal{L}_{\mathrm{int}}$ enables our method to achieve better performance. + +Analysis of Event Threshold. We conduct experiments on the medium-scale subset of the synthetic data to evaluate + +the impact of different event thresholds. We vary the threshold from $c = 0.1$ to $c = 0.4$ , and the results are shown in Figure 7. Since our method does not rely solely on event data and incorporates RGB inputs as supervision, it demonstrates robustness to varying event thresholds. + +Robustness to Motion Blur Levels. To validate the robustness of our method in handling different levels of motion blur, we set up three different camera speeds in the City Blocks scene of the synthetic dataset to obtain images with varying levels of blur. Images with mild blur are captured at $0.5 \times$ camera motion speed, images with medium blur are captured at the default motion speed, and images with strong blur are captured at $2.0 \times$ motion speed. The quantitative results of novel view synthesis are listed in Table. 4. It can be observed that the results show no significant fluctuations across different levels of blur, demonstrating the robustness of our method to varying motion blur levels. + +# 5. Conclusions + +This paper introduces Event Stream Assisted Gaussian Splitting (EvaGaussians), a novel framework that seamlessly integrates the event streams captured by an event camera into the training of 3D-GS, effectively addressing the challenges of reconstructing high-quality 3D-GS from motion-blurred images. We contribute two novel datasets and conduct comprehensive experiments. The results demonstrate that our method outperforms previous state-of-the-art deblurring rendering techniques in terms of novel view synthesis quality, without sacrificing inference efficiency. Despite its promising performance, our method may still face challenges when reconstructing scenes with extremely intricate textures from severely blurred images. + +# Acknowledgement + +This work was supported in part by Natural Science Foundation of China (No.62332002, No.62202014), and Shenzhen KQTD (No.20240729102051063). + +# References + +[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In CVPR, 2016. 2 +[2] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame- and event-based davis sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686–689, 2014. 2 +[3] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6, 7 +[4] Mauro Comi, Alessio Tonioni, Max Yang, Jonathan Tremblay, Valts Blukis, Yijiong Lin, Nathan F Lepora, and Laurence Aitchison. Snap-it, tap-it, splat-it: Tactile-informed 3d gaussian splatting for reconstructing challenging surfaces. arXiv preprint arXiv:2403.20275, 2024. 5 +[5] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 5, 7 +[6] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2924–2932, 2025. 2 +[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 2020. 2, 3 +[8] Haiqian Han, Jianing Li, Henglu Wei, and Xiangyang Ji. Event-3dgs: Event-based 3d reconstruction using 3d gaussian splatting. In Advances in Neural Information Processing Systems, 2025. 3 +[9] Philipp Heise, Sebastian Klose, Brian Jensen, and Alois Knoll. Pm-huber: Patchmatch with huber regularization for stereo matching. In ICCV, 2013. 5 +[10] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 4, 5 +[11] Jian Huang, Chengrui Dong, and Peidong Liu. Inceventgs: Pose-free gaussian splatting from a single event camera. In CVPR, 2025. 3 +[12] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 3 + +[13] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In CVPR, 2020. 2 +[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 4, 5 +[15] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In CVPR, 2023. 1 +[16] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 2023. 3 +[17] Joseph J LaViola Jr. Bringing vr and spatial 3d interaction to the masses through video games. IEEE Computer Graphics and Applications, 2008. 1 +[18] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2 +[19] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision, pages 416-434. Springer, 2025. 3 +[20] Yuchen Li*, Chaoran Feng*, Zhenyu Tang, Kaiyuan Deng, Wangbo Yu, Yonghong Tian, and Li Yuan. Gs2e: Gaussian splatting is an effective data generator for event stream generation. arXiv preprint arXiv:2505.15287, 2025. 2 +[21] Guixu Lin, Jin Han, Mingdeng Cao, Zhihang Zhong, and Yinqiang Zheng. Event-guided frame interpolation and dynamic range expansion of single rolling shutter image. In Proceedings of the 31st ACM International Conference on Multimedia, 2023. +[22] Mingyuan Lin, Yangguang Wang, Xiang Zhang, Boxin Shi, Wen Yang, Chu He, Gui-song Xia, and Lei Yu. Self-supervised shutter unrolling with events. IJCV, 2025. 2 +[23] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In ECCV, 2020. 2 +[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 1 +[25] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2024. 1 +[26] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 6 +[27] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[28] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural Radiance Fields from Blurry Images. In CVPR, 2022. 2 + +[29] Nico Messikommer, Carter Fang, Mathias Gehrig, Giovanni Cioffi, and Davide Scaramuzza. Data-driven feature tracking for event cameras with and without frames. TPAMI, 2025. 2 +[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5 +[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 21(12):4695-4708, 2012. 6 +[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6 +[33] Genady Paikin, Yotam Ater, Roy Shaul, and Evgeny Soloveichik. Efi-net: Video frame interpolation from fusion of events and frames. In CVPR, 2021. 2 +[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 2, 3, 6, 7 +[35] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In CVPR, 2020. 2 +[36] Roberto Pellerito, Marco Cannici, Daniel Gehrig, Joris Bel-hadj, Olivier Dubois-Matra, Massimo Casasco, and Davide Scaramuzza. Deep visual odometry with events and frames. In IROS, 2024. 2 +[37] Cheng Peng and Rama Chellappa. Pdrf: Progressively deblurring radiance field for fast and robust scene reconstruction from blurry images. arXiv preprint arXiv:2208.08049, 2022. 2 +[38] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In ECCV, 2024. 2, 5, 6, 7 +[39] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 3, 5, 6, 7 +[40] Yunshan Qi, Lin Zhu, Yifan Zhao, Nan Bao, and Jia Li. Deblurring neural radiance fields with event-driven bundle adjustment. In MM, 2024. 3 +[41] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 4, 5 +[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. 2023. 1 +[43] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, 2016. 1, 2, 3 +[45] Zihang Shao*, Chaoran Feng*, Xuanye Fang, Yaxin Li, Jiangrong Shen, and Qi Xu. Eicil: joint excitatory inhibitory cycle iteration learning for deep spiking neural networks. Advances in Neural Information Processing Systems, 36: 32117-32128, 2023. 2 + +[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 6, 7 +[47] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022. 2 +[48] Zhenyu Tang*, Chaoran Feng*, Xinhua Cheng, Wangbo Yu, Junwu Zhang, Yuan Liu, Xiaoxiao Long, Wenping Wang, and Li Yuan. Neuralgs: Bridging neural fields and 3d gaussian splatting for compact 3d representations. arXiv preprint arXiv:2503.23162, 2025. +[49] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7320–7328, 2025. 2 +[50] Gemma Taverni. Applications of Silicon Retinas: From Neuroscience to Computer Vision. PhD thesis, Universität Zürich, 2020. 5 +[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, pages 16155-16164, 2021. 2 +[52] Manoj K Vairalkar and SU Nimbhorkar. Edge detection of images using sobel operator. International Journal of Emerging Technology and Advanced Engineering, 2(1):291-293, 2012. 5 +[53] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 6 +[54] Zhexiong Wan, Bin Fan, Le Hui, Yuchao Dai, and Gim Hee Lee. Instance-level moving object segmentation from a single image with events. IJCV, 2025. 2 +[55] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. In ICML, 2024. 3 +[56] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In CVPR, 2023. 2, 3, 5, 6, 7, 8 +[57] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[58] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2 +[59] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 3 + +[60] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, et al. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia, 2023. 1 +[61] Wenhao Xu, Wenming Weng, Yueyi Zhang, Ruikang Xu, and Zhiwei Xiong. Event-boosted deformable 3d gaussians for fast dynamic scene reconstruction. arXiv preprint arXiv:2411.16180, 2024. 2, 3 +[62] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. 2021. 1 +[63] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 3 +[64] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024. +[65] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, 2024. +[66] Sohaib Zahid, Viktor Rudnev, Eddy Ilg, and Vladislav Golyanik. E-3dgs: Event-based novel view rendering of large-scale scenes using 3d gaussian splatting. In 3DV, 2025. 3 +[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6 +[68] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, 2022. 2 +[69] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 3 +[70] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In ECCV, 2024. 2, 5, 6, 7, 8 +[71] Haiyang Zhou, Wangbo Yu, Jiawen Guan, Xinhua Cheng, Yonghong Tian, and Li Yuan. Holotime: Taming video diffusion models for panoramic 4d scene generation. arXiv preprint arXiv:2504.21650, 2025. 3 +[72] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2 +[73] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143-14152, 2020. 6 +[74] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle + +feys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 1 +[75] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. TPAMI. 2 \ No newline at end of file diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/images.zip b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dad9ca78141d6bc07f93f2d5286c5c398ba4af58 --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a5fe29f171c04da9a9d5c576f3ff4bf02c2ced92680ed380d3e94c9cef1cb523 +size 822765 diff --git a/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/layout.json b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..b75479c1dd8f81b38ec64a1b86f133a38c18447f --- /dev/null +++ b/2025/EvaGaussians_ Event Stream Assisted Gaussian Splatting from Blurry Images/layout.json @@ -0,0 +1,10090 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 65, + 103, + 545, + 121 + ], + "type": "text", + "content": "EvaGaussians: Event Stream Assisted Gaussian Splatting from Blurry Images" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "spans": [ + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": "Wangbo Yu" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1,2*}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Chaoran Feng" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Jianing Li" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Jiye Tang" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Jiashu Yang" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Zhenyu Tang" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Meng Cao" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Xu Jia" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Yuchao Yang" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": ", Li Yuan" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "text", + "content": " and Yonghong Tian" + }, + { + "bbox": [ + 65, + 142, + 544, + 172 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 72, + 186, + 538, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 72, + 186, + 538, + 228 + ], + "spans": [ + { + "bbox": [ + 72, + 186, + 538, + 228 + ], + "type": "text", + "content": "1Shenzhen Graduate School, Peking University, 2Peng Cheng Laboratory, 3School of Computer and Science, Peking University \n4University of Science and Technology of China, 5Dalian University of Technology, 6MBZUAI {wbyu, chaoran.feng}@stu.pku.edu.cn, {yuanli-ece, yhtian}@pu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 167, + 321 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 167, + 321 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 167, + 321 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 167, + 321 + ], + "type": "image", + "image_path": "2152ac3dce439fe973cb9b3ed60d1974db601256af66caec9056d6bd2e562052.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 167, + 247, + 276, + 321 + ], + "blocks": [ + { + "bbox": [ + 167, + 247, + 276, + 321 + ], + "lines": [ + { + "bbox": [ + 167, + 247, + 276, + 321 + ], + "spans": [ + { + "bbox": [ + 167, + 247, + 276, + 321 + ], + "type": "image", + "image_path": "24e6ee784102928c42531161649139aedb027a483b80747d64d122673ca06bd3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 277, + 247, + 386, + 321 + ], + "blocks": [ + { + "bbox": [ + 277, + 247, + 386, + 321 + ], + "lines": [ + { + "bbox": [ + 277, + 247, + 386, + 321 + ], + "spans": [ + { + "bbox": [ + 277, + 247, + 386, + 321 + ], + "type": "image", + "image_path": "18d76e1fc6782a46660d4a8bcaa50afb52f8d981e1aa95d910fd71f7c55782b4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 58, + 323, + 167, + 397 + ], + "blocks": [ + { + "bbox": [ + 58, + 323, + 167, + 397 + ], + "lines": [ + { + "bbox": [ + 58, + 323, + 167, + 397 + ], + "spans": [ + { + "bbox": [ + 58, + 323, + 167, + 397 + ], + "type": "image", + "image_path": "55c6a84091b04e4bdf398e896c20b4edcf348420a70a152111f4cb5fd9a2a731.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 555, + 427 + ], + "type": "text", + "content": "Figure 1. EvaGaussians integrates frame-based cameras and event cameras to reconstruct sharp 3D-GS from motion-blurred images. It achieves higher-quality novel view synthesis compared to previous methods." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 168, + 323, + 276, + 397 + ], + "blocks": [ + { + "bbox": [ + 168, + 323, + 276, + 397 + ], + "lines": [ + { + "bbox": [ + 168, + 323, + 276, + 397 + ], + "spans": [ + { + "bbox": [ + 168, + 323, + 276, + 397 + ], + "type": "image", + "image_path": "fa21a3ac57b6ab17d068127ee646ece87416a594ba078a2c2658559e1a02001a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 277, + 323, + 386, + 397 + ], + "blocks": [ + { + "bbox": [ + 277, + 323, + 386, + 397 + ], + "lines": [ + { + "bbox": [ + 277, + 323, + 386, + 397 + ], + "spans": [ + { + "bbox": [ + 277, + 323, + 386, + 397 + ], + "type": "image", + "image_path": "a0c2082c7dba596bd96e1068e01cb9d18deccdf200263997b911f35dee4cd5bb.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 389, + 247, + 551, + 350 + ], + "blocks": [ + { + "bbox": [ + 389, + 247, + 551, + 350 + ], + "lines": [ + { + "bbox": [ + 389, + 247, + 551, + 350 + ], + "spans": [ + { + "bbox": [ + 389, + 247, + 551, + 350 + ], + "type": "image", + "image_path": "474585339f000993377c94f4c229b1fc1ca6be93d9212381ef00bfb036f6ebcd.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 389, + 351, + 551, + 397 + ], + "blocks": [ + { + "bbox": [ + 389, + 351, + 551, + 397 + ], + "lines": [ + { + "bbox": [ + 389, + 351, + 551, + 397 + ], + "spans": [ + { + "bbox": [ + 389, + 351, + 551, + 397 + ], + "type": "image", + "image_path": "01cc40e305b7342c95bf0557567eca0f29e05280db0fa380ad75d0aad95ed5f0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 151, + 441, + 200, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 441, + 200, + 453 + ], + "spans": [ + { + "bbox": [ + 151, + 441, + 200, + 453 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 466, + 297, + 682 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 466, + 297, + 682 + ], + "spans": [ + { + "bbox": [ + 54, + 466, + 297, + 682 + ], + "type": "text", + "content": "3D Gaussian Splatting (3D-GS) has demonstrated exceptional capabilities in synthesizing novel views of 3D scenes. However, its training is heavily reliant on high-quality images and precise camera poses. Meeting these criteria can be challenging in non-ideal real-world conditions, where motion-blurred images frequently occur due to high-speed camera movements. To address these challenges, we introduce Event Stream Assisted Gaussian Splatting (EvaGaus-sians), a novel approach that harnesses event streams captured by event cameras to facilitate the learning of high-quality 3D-GS from motion-blurred images. Capitalizing on the high temporal resolution and dynamic range offered by event streams, we seamlessly integrate them into the initialization and optimization of 3D-GS, thereby enhancing the acquisition of high-fidelity novel views with intricate texture details. We also contribute two novel datasets comprising RGB frames, event streams, and corresponding camera parameters, featuring a wide variety of scenes" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 441, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 491 + ], + "type": "text", + "content": "and various camera motions. The comparison results reveal that our approach not only excels in generating high-fidelity novel views, but also offers improved training and inference efficiency. Video and code are available at the project page." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "spans": [ + { + "bbox": [ + 314, + 514, + 394, + 526 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 713 + ], + "type": "text", + "content": "Novel view synthesis from 2D image collections has presented a persistent challenge within the field of computer vision and computer graphics. This task stands as a fundamental component in various vision applications, such as virtual reality [17, 60], robotics navigation [42, 62, 74], scene understanding [15, 24, 25], and many others, thereby prompting significant research efforts over the last decades. Amid pioneering works, 3D Gaussian Splatting (3D-GS) [14] achieves notable success in generating high-fidelity novel views. However, the optimization of 3D-GS heavily relies on accurate camera poses and point cloud initialization produced by COLMAP [44], which necessitates high-quality images without blurring. Fulfilling such conditions can be challenging in real-world situations. For example, in UAVs and robotics, rapid camera movement is" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 223, + 703 + ], + "type": "text", + "content": "*These authors contributed equally to this work." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24780" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "common when capturing images or recording videos, which often result in significant motion blur." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 102, + 296, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 102, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 55, + 102, + 296, + 245 + ], + "type": "text", + "content": "Recent studies have demonstrated the significant potential of event-based cameras in alleviating motion blur in images captured by conventional frame-based cameras. [13, 23, 34]. Serving as an innovative bio-inspired visual sensor, event cameras asynchronously report the logarithmic intensity changes of each pixel captured, and can record higher temporal resolution and dynamic range data in contrast to conventional cameras. The continuous event streams naturally align with the frame-rate agnostic rendering property of 3D-GS, offering sharper scene geometry information than frame-based cameras, making it highly promising in facilitating 3D-GS training from fast camera motion." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 251, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 251, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 251, + 296, + 525 + ], + "type": "text", + "content": "Motivated by this, we propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first method that seamlessly incorporates continuous event streams captured by event cameras to enhance the learning of high-quality 3D-GS from motion-blurred images. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Furthermore, to tackle the geometric ambiguity caused by blurry images, we introduce two event-assisted geometry regularization terms to stabilize the geometry of 3D-GS. Through optimizing the 3D-GS in a progressive manner, our method can recover a high-quality 3D-GS that facilitates the real-time generation of high-fidelity novel views. Finally, to comprehensively evaluate novel view synthesis with hybrid frame- and event-based cameras, we contribute a synthetic dataset and a real-world dataset, both containing diverse scenes with varying scales and camera motions, setting a benchmark for related researches. To summarize, our contributions can be delineated as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 604 + ], + "type": "text", + "content": "- We propose EvaGaussians (Event Stream Assisted Gaussian Splatting), the first framework designed to reconstruct high-quality 3D-GS from motion-blurred images with the assistance of event cameras. It not only recovers intricate details from the motion-blurred images but also enables high-fidelity novel view synthesis." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 295, + 651 + ], + "type": "text", + "content": "- We contribute a synthetic dataset and a real-world dataset for evaluating novel view synthesis with hybrid frame-and event-based cameras. Both datasets contain diverse scenes with varying scales and camera motions." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 651, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 651, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 651, + 296, + 712 + ], + "type": "text", + "content": "- We conduct a comprehensive evaluation of the proposed method and compare it with several strong baselines. As shown in Figure. 1 and Figure. 2, our method not only excels in generating higher-fidelity novel views but also enables improved training and inference efficiency." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "type": "image", + "bbox": [ + 325, + 70, + 541, + 281 + ], + "blocks": [ + { + "bbox": [ + 325, + 70, + 541, + 281 + ], + "lines": [ + { + "bbox": [ + 325, + 70, + 541, + 281 + ], + "spans": [ + { + "bbox": [ + 325, + 70, + 541, + 281 + ], + "type": "image", + "image_path": "2763364a299093d5f3e4b8de97327ad8b5e3773cd95150f44f71090338857d6a.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 283, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 283, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 555, + 327 + ], + "type": "text", + "content": "Figure 2. Performance, rendering speed, and training time (indicated by the area of each circle) comparison. EvaGaussians achieves better results while maintaining real-time rendering and training efficiency." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 333, + 406, + 346 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 333, + 406, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 406, + 346 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 354, + 550, + 367 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 550, + 367 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 550, + 367 + ], + "type": "text", + "content": "2.1. Reconstructing 3D Scene from Blurry Images" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 373, + 555, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 555, + 622 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 555, + 622 + ], + "type": "text", + "content": "Reconstructing a high-quality 3D Scene typically requires high-fidelity, sharp images as supervision. However, motion-blurred images often occur in real world scenarios, thus hindering accurate reconstruction of 3D scenes. Several studies have been proposed to address this issue. For example, Deblur-NeRF [28], DP-NeRF [18] and PDRF [37] attempted to learn a blur formation kernel to model the image blurring process. BAD-NeRF [56] further physically modeled the blurry images formation process, and adopted a bundle-adjustment strategy to jointly optimize NeRF parameters and the camera poses during the exposure time. These NeRF-based methods lacked real-time rendering capabilities and suffered from extended training times. With the rapid advancement of 3D-GS, concurrent works like BAD-Gaussians [70] and BAGS [38] proposed to utilize 3D-GS as representation to achieve deblurring reconstruction. Although they achieved real-time rendering and faster convergence compared with prior works, they still struggled to handle severely blurred images in which COLMAP [44] will fail to produce the initial point clouds and camera calibration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 550, + 647 + ], + "type": "text", + "content": "2.2. Reconstructing 3D Scene from Event Streams" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 715 + ], + "type": "text", + "content": "Motivated by the exceptional properties offered by event cameras [1, 2, 6, 7, 20-22, 29, 33, 35, 36, 45, 47-49, 51, 54, 58, 61, 68, 72, 75], several studies attempted to reconstruct 3D scenes from event streams captured by event cameras, particularly in low-light conditions with fast camera mo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24781" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": "tion [8, 11, 12, 27, 43, 55, 59, 61, 63-66, 69, 71]. These methods were mostly designed to recover scene geometry from event streams, and are not typically designed for reconstructing 3D Scene from blurry images. Recently, E-NeRF [16], " + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 55, + 72, + 296, + 228 + ], + "type": "text", + "content": " [39], EBAD-NeRF [40], BENeRF [19] and EvDNeRF [3] proposed to jointly utilize event streams captured by event cameras and motion-blurred images captured by standard frame-based cameras to reconstruct a NeRF representation. In contrast to frame-based deblurring methods, they are better at handling motion blur due to the introduction of event information. However, these NeRF-based methods suffer from long training and inference times, which limit their further application." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 238, + 111, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 111, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 111, + 251 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 257, + 135, + 270 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 135, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 135, + 270 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": "Event camera is a type of bio-inspired sensor that can asynchronously record intensity changes [7]. In contrast to conventional cameras that are restricted to sequentially produce frames at a fixed frame rate, event cameras asynchronously trigger events in each pixel when their intensity change exceeds a constant threshold, featuring properties such as low latency and high dynamic range. Formally, let " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{xy}(t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " denote the instantaneous intensity at pixel coordinate " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{xy}(t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " denotes its logarithm. An event " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "p = \\pm 1" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " will be triggered whenever the change of " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{L}_{xy}(t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " surpasses the threshold " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": ", where the polarity represents the direction (increase or decrease) of changes. Let " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\delta_{t_0}(t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " be the impulse function at time " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " with a unit integral, the event can therefore be expressed as a continuous-time signal " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{e}_{xy}(t) = p\\delta_{t_0}(t)" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " signifies the time at which the event occurs. Then, the proportional intensity change during a time interval " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "[s,t]" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " can be computed as the integral of events that occurred between times " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": ", expressed as " + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_{xy}(t) = \\int_s^t\\mathbf{e}_{xy}(h)dh" + }, + { + "bbox": [ + 55, + 274, + 296, + 526 + ], + "type": "text", + "content": ". Given that each pixel can be treated separately in the event camera, the subscripts can be omitted:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 532, + 295, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 532, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 133, + 532, + 295, + 559 + ], + "type": "interline_equation", + "content": "\\mathbf {E} (t) = \\int_ {s} ^ {t} \\mathbf {e} (h) d h. \\tag {1}", + "image_path": "832e16e2da233f35d02c7f2f4c05e46b80066b52399743d56f81294606f82331.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": "We can then represent the logarithmic intensity change as: " + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(t) - \\mathbf{L}(s) = c\\mathbf{E}(t)" + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": ", rewrite as " + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(t) = \\mathbf{L}(s) + c\\mathbf{E}(t)" + }, + { + "bbox": [ + 55, + 565, + 296, + 601 + ], + "type": "text", + "content": ", and subsequently obtain the actual intensity change:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 609, + 295, + 622 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 609, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 121, + 609, + 295, + 622 + ], + "type": "interline_equation", + "content": "\\mathbf {I} (t) = \\mathbf {I} (s) \\cdot \\exp (c \\mathbf {E} (t)). \\tag {2}", + "image_path": "bbd6c1a538ff49366cd3c653eb91697687fef8a063df2e9988bb9170b910460e.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": "Therefore, when an image " + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(s)" + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": " is captured at time " + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": ", and the event stream is recorded during the time interval " + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "inline_equation", + "content": "[s,t]" + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": ", the image " + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t)" + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": " can be obtained by warping " + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(s)" + }, + { + "bbox": [ + 55, + 629, + 296, + 667 + ], + "type": "text", + "content": " using Eq. 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 208, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 208, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 208, + 685 + ], + "type": "text", + "content": "3.2. Event-assisted Initialization" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "The optimization of 3D-GS requires camera calibration and point cloud initialization using COLMAP [44]. However," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "as shown in Table.4 in the supplementary, this process can fail when dealing with images that have significant motion blur. Motion-blurred images are resulted from camera movements during the exposure time, which can be mathematically represented as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 386, + 140, + 555, + 170 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 140, + 555, + 170 + ], + "spans": [ + { + "bbox": [ + 386, + 140, + 555, + 170 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\mathbf {I} (t) d t, \\tag {3}", + "image_path": "a5a5cda003d804adc25c87005acb3b44964de744c3c1d0aac197fb85af13e7b4.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": " denotes a captured blurry image, which is equivalent to averaging the instantaneous latent images " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t)" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": " during the exposure time " + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "inline_equation", + "content": "[s - \\tau /2,s + \\tau /2]" + }, + { + "bbox": [ + 313, + 177, + 553, + 213 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 213, + 555, + 273 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 273 + ], + "type": "text", + "content": "To obtain initial camera poses and point clouds for 3D-GS optimization, we first preprocess the motion-blurred images using the Event-based Double Integral (EDI) [34] model, which can be derived through substituting Eq. 2 into Eq. 3:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 358, + 281, + 555, + 311 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 358, + 281, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 358, + 281, + 555, + 311 + ], + "type": "interline_equation", + "content": "\\mathbf {B} = \\mathbf {I} (s) \\cdot \\frac {1}{\\tau} \\int_ {s - \\tau / 2} ^ {s + \\tau / 2} \\exp (c \\mathbf {E} (t)) d t. \\qquad (4)", + "image_path": "43c6066dc7b8d56e406292f5a2f10e8743783b2cbd7a4af3484f648666e087d6.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": "Given the predefined threshold " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", a blurry image " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", and the recorded event stream " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(t)" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", the EDI model (Eq. 4) allows the derivation of " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(s)" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", following which the latent image " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\mathbf{I}(t)" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": " at any moment within the exposure time can be estimated through Eq. 2. As shown in Figure. 3, given a total of " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": " blurry images " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}^j\\}_{j=1}^K" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", for each of them, we uniformly sample " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": " time stamps during their exposure time to obtain a series of EDI-processed latent images rich in texture features, denoted as " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{I}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": ", then obtain their poses " + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 318, + 554, + 437 + ], + "type": "text", + "content": " and the initial point cloud of the scene using COLMAP [44]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 582 + ], + "type": "text", + "content": "After initialization, a straightforward approach to optimizing the 3D-GS is to use the EDI-processed latent images and their camera poses as supervision. However, although these latent images provide more texture features than the original blurry image, they still do not fully recover the ideal latent image and exhibit relatively low visual quality, which also introduces inaccuracies into the camera poses, thereby leading to unsatisfactory optimization results. To more robustly recover a sharp 3D-GS from motion-blurred images, we propose to harness the advantages of event streams and seamlessly integrate them into the optimization process of 3D-GS." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 588, + 500, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 500, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 500, + 601 + ], + "type": "text", + "content": "3.3. Event-assisted Bundle-Adjustment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "As introduced in Eq. 3, during the exposure time, a motion-blurred image can be decomposed into a series of latent images along a specific camera trajectory, which can be roughly approximated by the EDI-produced camera poses " + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": " according to Eq. 4. Motivated by this, we jointly optimize these camera poses and the 3D-GS attributes in a bundle-adjustment manner [56] to simultaneously recover the blur-formation camera trajectories and a sharp 3D-GS. As shown in Figure. 3, we add each of the EDI-produced" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24782" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 251 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 251 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 251 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 251 + ], + "type": "image", + "image_path": "416f929965d0e231266637a7b3841749bd7acc39b560a0de43f9ee4dcccb8017.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 260, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 54, + 260, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 260, + 555, + 327 + ], + "type": "text", + "content": "Figure 3. Overview of EvaGaussians. Harnessing the exceptional temporal resolution and dynamic range offered by event streams, we begin by using them to assist in the initialization of 3D-GS and camera poses, then utilize a blur formation model and an event reconstruction loss to jointly optimize the 3D-GS parameters and recover blur-formation camera trajectories in a bundle-adjustment manner. Leveraging the continuously recorded event streams, we introduce two event-assisted geometry regularization terms that can function beyond the exposure time to stabilize the geometry of 3D-GS. Once trained, it can not only recover intricate details from motion-blurred images but also enable high-fidelity novel view synthesis." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": "camera poses a learnable offset " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{d}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " as correction parameters, resulting a learnable camera trajectory " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{P}}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{P}}_i = \\mathbf{P}_i + \\mathbf{d}_i" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": ". In each training iteration, we simultaneously render " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " images " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{I}}_i\\}_{i = 1}^n" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " from the 3D-GS along the camera trajectory of the blurry view, and then simulate the formation of the motion-blurred images using a discrete approximation of Eq 3, denoted as " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{B}} = \\frac{1}{n}\\sum_{i = 1}^{n}\\widetilde{\\mathbf{I}}_i" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": ". Consequently, for a total of " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " real-captured blurry images " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}^j\\}_{j = 1}^K" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": ", we can obtain their simulated versions " + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{B}}^j\\}_{j = 1}^K" + }, + { + "bbox": [ + 55, + 347, + 294, + 470 + ], + "type": "text", + "content": " through each corresponding learnable camera trajectory." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "text", + "content": "Blur Reconstruction Loss. With the simulated blurry images, we use the captured real blurry images " + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{B}^j\\}_{j=1}^K" + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "text", + "content": " to serve as image level supervision. Specifically, for each blurry image " + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\mathbf{B}^j" + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "text", + "content": " and its simulated version " + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{B}}^j" + }, + { + "bbox": [ + 55, + 470, + 295, + 542 + ], + "type": "text", + "content": ", we employ a blur reconstruction loss to minimize their photometric error, expressed as" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 548, + 295, + 563 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 548, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 61, + 548, + 295, + 563 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {b l u r} = \\left(1 - \\lambda_ {1}\\right) \\cdot \\| \\mathbf {B} ^ {j} - \\widetilde {\\mathbf {B}} ^ {j} \\| _ {1} + \\lambda_ {1} \\cdot \\mathrm {D - S S I M} \\left(\\mathbf {B} ^ {j}, \\widetilde {\\mathbf {B}} ^ {j}\\right). (5)", + "image_path": "0cc37ba41c4f743a119c32829284eb88f848f12c5bc9ec61c28a55643f3a81d5.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 629 + ], + "type": "text", + "content": "The formulation of blur reconstruction loss is the same as in the original 3D-GS [14], it differs in utilizing blurry images as supervision and jointly optimizing the 3D-GS attributes and the camera trajectories, thus facilitating an initial deblurring reconstruction of 3D-GS." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "Event Reconstruction Loss. Leveraging the abundant high-frequency information offered by the event streams, we further adopt an event reconstruction loss to aid in 3D-GS optimization. Specifically, we uniformly divide the exposure time into " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "m = n - 1" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": " intervals, each with a duration of " + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\frac{\\tau}{m}" + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": ". Subsequently, we integrate the recorded event stream along these time intervals using Eq. 1, resulting in" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "content": " event maps " + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{E}_i\\}_{i=1}^m" + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "content": " to serve as event level supervision. During training, for the " + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "content": "-th blurry view, we convert the rendered image sequence " + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{I}}_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "content": " on the camera trajectory into event maps " + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "inline_equation", + "content": "\\{\\widetilde{\\mathbf{E}}_i\\}_{i=1}^m" + }, + { + "bbox": [ + 313, + 347, + 555, + 433 + ], + "type": "text", + "content": ", using a differentiable event simulator [10, 41], and constrain the discrepancies between the simulated event maps and the ground truth event maps, expressed as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 442, + 555, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 442, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 372, + 442, + 555, + 474 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e v e n t}} = \\frac {1}{m} \\sum_ {i = 1} ^ {m} \\| \\mathbf {E} _ {i} - \\widetilde {\\mathbf {E}} _ {i} \\| _ {1}. \\tag {6}", + "image_path": "3094f9e62499af8783129039ddad980dd3e2eb0a303d81704668adcb0d7cb733.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 507 + ], + "type": "text", + "content": "The event reconstruction loss further aids in recovering a sharp 3D-GS with improved texture details." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 515, + 526, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 526, + 529 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 526, + 529 + ], + "type": "text", + "content": "3.4. Event-assisted Geometry Regularization" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": "The blurry color images are captured only during the exposure time and are much sparser than the event stream. Relying on such low-quality image-level supervision may cause the 3D-GS to overfit on the training images, resulting in significant floaters and inferior geometry. Leveraging Eq. 1 and Eq. 4, given the continuously recorded event streams " + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{E}(t)" + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": ", we can derive continuous grayscale intensity images " + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(t)" + }, + { + "bbox": [ + 313, + 533, + 555, + 665 + ], + "type": "text", + "content": " that are rich in geometric information and can function beyond the exposure time. Motivated by this, we further propose two event-assisted geometry regularization terms to aid in 3D-GS training." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": "Intensity Reconstruction Loss. During training, we randomly sample continuous time " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " between the interval of two adjacent blurry image, and derive the grayscale intensity image " + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{G}(t)" + }, + { + "bbox": [ + 313, + 666, + 556, + 714 + ], + "type": "text", + "content": " using Eq. 2. We then minimize the difference" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24783" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 96 + ], + "type": "text", + "content": "between it and the rendered intensity image from 3D-GS, expressed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 103, + 296, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 103, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 55, + 103, + 296, + 129 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i n t} = (1 - \\lambda_ {2}) \\cdot \\| \\mathbf {G} (t) - \\widetilde {\\mathbf {G}} (t) \\| _ {1} + \\lambda_ {2} \\cdot D - S S I M (\\mathbf {G} (t), \\widetilde {\\mathbf {G}} (t)), \\tag {7}", + "image_path": "383b66ed4c7e5d1e7223a311176b4dcf9c9ee75aac3b35c8aa2cbd1ab1d9026b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 129, + 280, + 141 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 129, + 280, + 141 + ], + "spans": [ + { + "bbox": [ + 55, + 129, + 280, + 141 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 129, + 280, + 141 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{G}}(t)" + }, + { + "bbox": [ + 55, + 129, + 280, + 141 + ], + "type": "text", + "content": " is converted from the colored render result." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 142, + 296, + 190 + ], + "type": "text", + "content": "Intensity-aware Depth Regularization Loss. As shown in Figure. 3, to further improve the geometry of 3D-GS, inspired by [4, 9], we adopt an intensity-aware depth regularization loss during training, defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 87, + 198, + 295, + 245 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 198, + 295, + 245 + ], + "spans": [ + { + "bbox": [ + 87, + 198, + 295, + 245 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {d e p t h}} = \\frac {1}{N} \\sum_ {x, y} \\left(\\left| \\partial_ {x} \\widetilde {\\mathbf {D}} _ {x y} (t) \\right| e ^ {- \\beta \\left| \\partial_ {x} \\mathbf {G} _ {x y} (t) \\right|} \\right. \\tag {8} \\\\ + | \\partial_ {y} \\widetilde {\\mathbf {D}} _ {x y} (t) | e ^ {- \\beta | \\partial_ {y} \\mathbf {G} _ {x y} (t) |}), \\\\ \\end{array}", + "image_path": "0250fc4ca58d28685f288f0c64399fe94d8fa6d782ad9bc5a5a7a72d24e29c45.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\widetilde{\\mathbf{D}}(t)" + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": " is the rendered depth map, " + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": " denotes the pixel location, " + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": " is the total number of pixels, and " + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": " is set to 2 in our experiments. The horizontal and vertical gradients are calculated by applying convolution operations with " + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 55, + 254, + 296, + 373 + ], + "type": "text", + "content": " Sobel kernels [52]. This regularization is founded on the observation that depth transitions in an image often correspond to changes in intensity. Therefore, it ensures that the spatial variation of depth closely matches that of the intensity image, thereby reducing geometric artifacts at object boundaries." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 375, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 398 + ], + "type": "text", + "content": "The total loss function is the combination of the above losses, defined as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 98, + 407, + 295, + 435 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 98, + 407, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 98, + 407, + 295, + 435 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {b l u r}} \\mathcal {L} _ {\\text {b l u r}} + \\lambda_ {\\text {e v e n t}} \\mathcal {L} _ {\\text {e v e n t}} \\tag {9} \\\\ + \\lambda_ {i n t} \\mathcal {L} _ {i n t} + \\lambda_ {d e p t h} \\mathcal {L} _ {d e p t h}. \\\\ \\end{array}", + "image_path": "cd8795088bfb791c3e411a895274763f09c3e5db6beb0b03bfa780cbfe6d9aa2.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 137, + 467 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 534 + ], + "type": "text", + "content": "In this section, we first provide the implementation details of our method in 4.1, followed by an illustration of the introduced dataset in 4.2, and the experimental settings in 4.3. The comparison results and ablation study are presented in 4.4 and 4.5. For more detailed experiments." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 188, + 553 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 689 + ], + "type": "text", + "content": "Progressive Training. We implemented EvaGaussians based on the official code of 3D-GS [14]. The training process spans 50,000 iterations, with an event reconstruction loss introduced after a 3,000-iteration warmup and we omit the densification process to streamline and simplify the subsequent optimization. Additionally, we adopt a coarse-to-fine training strategy, starting with rendering at a low resolution (0.3× downsampling in the early 30% iterations) and progressively increasing the size of the rendered views to full resolution. All experiments were conducted using a single NVIDIA RTX 4090 GPU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Hyperparameter Setting. During the training process, we set " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{blur} = 1.0" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{depth} = 1.0e^{-2}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{event} = 5.0e^{-3}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " and" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\lambda_{int} = 1.0e^{-3}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " for the loss function, and used " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "n = 9" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " for the number of poses to be optimized during the exposure time. In implementing the loss " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{event}" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": ", we configured the positive threshold as " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c_{pos} = 0.25" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and the negative threshold as " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c_{neg} = 0.25" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " for synthetic scenes, and set " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c_{pos} = 0.197" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "c_{neg} = 0.241" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " for real scenes." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 153, + 376, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 153, + 376, + 164 + ], + "spans": [ + { + "bbox": [ + 314, + 153, + 376, + 164 + ], + "type": "text", + "content": "4.2. Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 171, + 554, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 171, + 554, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 554, + 230 + ], + "type": "text", + "content": "To enable a comprehensive evaluation, we introduce two novel datasets, with an overview provided below. Additionally, we present a detailed dataset introduction in the supplementary and will release the dataset construction pipeline for future research." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 231, + 554, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 231, + 554, + 339 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 554, + 339 + ], + "type": "text", + "content": "EvaGaussians-Blender Dataset. We construct a synthetic dataset covering a variety of scene scales, coupling with diverse camera trajectories and event data. It consists 6 large-scale scenes, 7 medium-scale scenes, and 6 object-level scenes. We simulate motion blur by manually placing multi-view cameras, randomly adjusting camera poses, and performing linear interpolation between the original and perturbed positions for each view. The corresponding event streams are simulated using ESIM [41] and V2E [10]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 339, + 554, + 422 + ], + "type": "text", + "content": "EvaGaussians-DAVIS Dataset. We manually recorded five real-world scenes using the Color DAVIS346 event camera [50]. The dataset includes three object-level scenes and two indoor scenes. After processing, the final dataset consists of 30 images per scene, along with the recorded event streams, each showcasing various blur and lighting conditions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 423, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 554, + 544 + ], + "type": "text", + "content": "Comparison with Existing Datasets. The goal of contributing this dataset is to create a more comprehensive evaluation benchmark for event-assisted deblurring reconstruction tasks. While E2NeRF [39] and EvDNeRF [3] have introduced similar datasets, we observed that these datasets are limited in terms of scene variety, camera motion diversity, and data size. As shown in Table.1 and Table.2 of the supplementary, our proposed dataset includes a broader range of scenes with various scales, making it more suitable for comprehensive evaluation." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 552, + 432, + 565 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 552, + 432, + 565 + ], + "spans": [ + { + "bbox": [ + 314, + 552, + 432, + 565 + ], + "type": "text", + "content": "4.3. Experiment Settings" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": "Baselines. We compare our method with three types of baselines: 1) NeRF [30] and 3D-GS [14] directly trained on the blurry images, referring to as B-NeRF and B-3DGS. 2) Deblur rendering methods, including BAD-NeRF [56], BAD-GS [70], BAGS [38], " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " [39], and EvDNeRF [3]. Among these, BAD-NeRF and BAD-GS simulate motion blur and optimize camera trajectories without using event stream, while BAGS addresses motion blur through kernel modeling, also without event stream. In contrast, " + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 313, + 570, + 555, + 714 + ], + "type": "text", + "content": " and EvDNeRF are event-assisted methods without camera trajectory optimization. 3) Image deblur methods, including UFP [5] (single-image deblurring)," + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24784" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 529, + 423 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 529, + 423 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 529, + 423 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 529, + 423 + ], + "type": "image", + "image_path": "8b2d07830431f64054add0d7b8d3d90ba653c374e34190d80d65ea35b64c1085.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 427, + 555, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 427, + 555, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 555, + 450 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on the synthetic dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Full results are provided in the supplementary." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 463, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 295, + 521 + ], + "type": "text", + "content": "EDI [34] (event-based deblurring), and EFNet [46] (learnable event-based deblurring). We process input blurry images with them and train the vanilla 3D-GS with predeblurred images. The resulting baselines are referred to as UFP-GS, EDI-GS, and EFN-GS." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 654 + ], + "type": "text", + "content": "Evaluation Metrics. For synthetic datasets, we employ the Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM) [57], and VGG-based Learned Perceptual Image Patch Similarity (LPIPS) [67] to evaluate the similarity between rendered novel views and ground-truth novel views. For real-world datasets, since the sharp ground-truth images are unavailable, we utilize several No-Reference Image Quality Assessment (NR-IQA) metrics for evaluation, including BRISQUE [31], NIQE [32], PIQE [53], RankIQA [26], and MetaIQA [73], which allow for image evaluation when lacking ground truth images." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 141, + 673 + ], + "type": "text", + "content": "4.4. Comparisons" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "We evaluate our approach across a variety of scenes, including large-scale scenes, medium-scale scenes, object-level scenes and real scenes. Quantitative assessments of" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 463, + 555, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 463, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 313, + 463, + 555, + 690 + ], + "type": "text", + "content": "novel view synthesis are shown in Table. 1. It can be found that our method achieves substantial improvements in most of the metrics, especially in challenging large-scale scenes. Specifically, both B-NeRF and B-3DGS produce blurry novel views since they are directly trained on motion-blurred images. The image deblurring-based baselines, UFP-GS, EDI-GS and EFN-GS, also produced inferior results, because the image deblurring process potentially corrupts the 3D consistency of the training images. Notably, our approach outperforms BAD-NeRF [56], BAD-GS [70] and BAGS [38], due to their limited capability in modeling complex textures. In addition, our method also surpasses the event-assisted methods " + }, + { + "bbox": [ + 313, + 463, + 555, + 690 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 313, + 463, + 555, + 690 + ], + "type": "text", + "content": " [39] and EvDNeRF [3] in producing high-quality novel views with intricate details, with better training and rendering efficiency. Notably, since directly using COLMAP on the blurry images fails in most cases (as shown in Table.4 of the supplementary), we use the EDI-produced poses for all the comparison baselines and our method to ensure a fair comparison." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "The qualitative results of five top-performing methods" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24785" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 106, + 553, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 104 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of novel view synthesis across large-scale, medium-scale, object-level, and real-world scenes. The table reports the average performance for each scale, demonstrating that our method consistently surpasses previous state-of-the-art approaches across all metrics. Best-performing results are highlighted in bold and second-best results in underline." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 106, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 58, + 106, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 58, + 106, + 553, + 285 + ], + "type": "table", + "html": "
Scene TypeMetricsB-NeRFB-3DGSUFP-GS [5]EDI-GS [34]EFN-GS [46]E2NeRF [39]BAD-NeRF [56]BAD-GS [70]BAGS [38]EvDNeRF [3]Ours
LargePSNR↑21.3321.4821.3622.3122.6922.9623.8523.8624.9324.6326.02
SSIM↑0.6780.6880.6600.6860.6830.7070.7320.7330.7450.7530.806
LPIPS↓0.4250.3970.3740.3820.3630.3750.3480.3470.3270.3280.268
MediumPSNR↑24.0824.8026.3826.4426.1327.7828.4428.4628.7428.9130.47
SSIM↑0.7170.7510.8020.8010.7980.8660.8790.8730.8740.8850.916
LPIPS↓0.3620.3190.2640.2580.2730.1990.1820.1820.1740.1690.152
ObjectPSNR↑22.2822.3425.1624.9425.4529.6127.3327.8628.4129.8330.24
SSIM↑0.9040.9050.9280.9250.9290.9640.9480.9500.9560.9660.970
LPIPS↓0.1480.1470.1170.1210.1100.0740.0930.0910.0840.0720.070
RealBRISQUE↓92.2573.8062.9462.7562.9361.5261.5060.8958.5858.6353.96
NIQE↓15.0012.0110.1710.2010.219.44010.009.9029.2869.0118.371
PIQE↓65.9252.7445.0344.8344.8446.7643.9543.5145.2744.6341.53
RankIQA↓9.4287.5426.4396.4116.4115.5736.2856.2235.6415.3204.895
MetaIQA↑0.1240.1420.1730.1740.1730.1810.1770.1790.1980.1910.213
", + "image_path": "2ded9169f56cf91f41121cf49e8fdcde63ca8429cdbbc2b4311a28fe36975e48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 296, + 141, + 421 + ], + "blocks": [ + { + "bbox": [ + 58, + 296, + 141, + 421 + ], + "lines": [ + { + "bbox": [ + 58, + 296, + 141, + 421 + ], + "spans": [ + { + "bbox": [ + 58, + 296, + 141, + 421 + ], + "type": "image", + "image_path": "7692f256e148173b3bd5b7166d6e45370dad6c37f0859613432644ff851831e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 422, + 124, + 431 + ], + "lines": [ + { + "bbox": [ + 80, + 422, + 124, + 431 + ], + "spans": [ + { + "bbox": [ + 80, + 422, + 124, + 431 + ], + "type": "text", + "content": "Blurry Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 141, + 297, + 222, + 420 + ], + "blocks": [ + { + "bbox": [ + 141, + 297, + 222, + 420 + ], + "lines": [ + { + "bbox": [ + 141, + 297, + 222, + 420 + ], + "spans": [ + { + "bbox": [ + 141, + 297, + 222, + 420 + ], + "type": "image", + "image_path": "61362ef85eca0d2cc84d792f893ae4a872608c25a4441972fe6d66b8d991f144.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 170, + 421, + 200, + 430 + ], + "lines": [ + { + "bbox": [ + 170, + 421, + 200, + 430 + ], + "spans": [ + { + "bbox": [ + 170, + 421, + 200, + 430 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 223, + 297, + 304, + 420 + ], + "blocks": [ + { + "bbox": [ + 223, + 297, + 304, + 420 + ], + "lines": [ + { + "bbox": [ + 223, + 297, + 304, + 420 + ], + "spans": [ + { + "bbox": [ + 223, + 297, + 304, + 420 + ], + "type": "image", + "image_path": "e2b9482e2f3521bb08dd1c83bc3b86d250021b668fd6d869ecf996743f93e701.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 249, + 422, + 282, + 430 + ], + "lines": [ + { + "bbox": [ + 249, + 422, + 282, + 430 + ], + "spans": [ + { + "bbox": [ + 249, + 422, + 282, + 430 + ], + "type": "text", + "content": "BAD-GS" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 435, + 555, + 460 + ], + "lines": [ + { + "bbox": [ + 55, + 435, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 555, + 460 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on the real-world dataset. We present rendered novel views and zoomed-in patches with highlighted colors for detailed comparison. Compared with baselines, our method is capable of reconstructing shaper geometry from blurry inputs." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 306, + 297, + 387, + 420 + ], + "blocks": [ + { + "bbox": [ + 306, + 297, + 387, + 420 + ], + "lines": [ + { + "bbox": [ + 306, + 297, + 387, + 420 + ], + "spans": [ + { + "bbox": [ + 306, + 297, + 387, + 420 + ], + "type": "image", + "image_path": "10c45d0baf38537598bef3cb3109049782f3b912bea66e564f44d08d685dd7b1.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 422, + 359, + 430 + ], + "lines": [ + { + "bbox": [ + 335, + 422, + 359, + 430 + ], + "spans": [ + { + "bbox": [ + 335, + 422, + 359, + 430 + ], + "type": "text", + "content": "BAGS" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 387, + 297, + 469, + 420 + ], + "blocks": [ + { + "bbox": [ + 387, + 297, + 469, + 420 + ], + "lines": [ + { + "bbox": [ + 387, + 297, + 469, + 420 + ], + "spans": [ + { + "bbox": [ + 387, + 297, + 469, + 420 + ], + "type": "image", + "image_path": "d7f51291244177e455030f2a6884167696e734244f327adf99b78733de49a27e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 413, + 421, + 449, + 430 + ], + "lines": [ + { + "bbox": [ + 413, + 421, + 449, + 430 + ], + "spans": [ + { + "bbox": [ + 413, + 421, + 449, + 430 + ], + "type": "text", + "content": "EvDNeRF" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 470, + 297, + 553, + 420 + ], + "blocks": [ + { + "bbox": [ + 470, + 297, + 553, + 420 + ], + "lines": [ + { + "bbox": [ + 470, + 297, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 470, + 297, + 553, + 420 + ], + "type": "image", + "image_path": "86971b530a316f0511cc918201512d63acc49306abe955665146864bd54360e8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 505, + 422, + 523, + 430 + ], + "lines": [ + { + "bbox": [ + 505, + 422, + 523, + 430 + ], + "spans": [ + { + "bbox": [ + 505, + 422, + 523, + 430 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 470, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 470, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 297, + 662 + ], + "type": "text", + "content": "are presented in Figure. 4 and Figure. 5, and the complete qualitative results are provided in the supplementary. In Figure. 4, it can be found that " + }, + { + "bbox": [ + 54, + 470, + 297, + 662 + ], + "type": "inline_equation", + "content": "\\mathrm{E}^2\\mathrm{NeRF}" + }, + { + "bbox": [ + 54, + 470, + 297, + 662 + ], + "type": "text", + "content": " [39] struggles in medium and large-scale scene modeling, producing significant blurring results. Additionally, BAD-GS [70] and BAGS [38] falls short in regions with significant color and depth variations, and produces overly smooth background textures. Although EvDNeRF [3] exhibits overall satisfactory performance, its complex network architecture prolongs the training time and precludes real-time rendering. In comparison, our method outperforms the baselines in producing high-fidelity novel views, and significantly reducing training time as well as demonstrating advantages in real-time applications. Figure. 5 further demonstrates that ours is capable of reconstructing detailed geometry from blurry inputs, achieving better performance than baselines." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 671, + 149, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 149, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 149, + 685 + ], + "type": "text", + "content": "4.5. Ablation Study" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Effectiveness of Poses Optimization. We conduct experiments on the large-scale subset of synthetic data to evaluate" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 470, + 555, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 470, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 313, + 470, + 555, + 567 + ], + "type": "text", + "content": "the impact of pose optimization. As shown in Table 3, we calculate the ATE (Average Trajectory Error) between the initial COLMAP poses and GT poses, as well as the ATE between the optimized poses and GT poses. The results demonstrate the effectiveness of our method in correcting the initial poses and recovering more accurate camera trajectories. Additionally, the improved image quality metrics further demonstrate the effectiveness of pose optimization." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": "Analysis of Number of Latent Camera Poses. We conduct ablations to investigate the effect of camera poses number in the exposure time, using the large-scale subset of the synthetic data. Specifically, we vary the number of latent camera poses " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " from 5, 9, 13, and 17, and visualize the quantitative novel view synthesis results in Figure. 6. It can be observed that the performance reaches a bottleneck at 9 latent poses. Beyond this point, further improvements are limited and may potentially lead to local convergence issues due to over-parameterization. Based on these experiments, we choose " + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "inline_equation", + "content": "n = 9" + }, + { + "bbox": [ + 313, + 570, + 556, + 714 + ], + "type": "text", + "content": " camera poses to achieve a balance between rendering performance and training efficiency. Here," + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24786" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 138, + 174 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 138, + 174 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 138, + 174 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 138, + 174 + ], + "type": "image", + "image_path": "8520e65798fe81ba266bfac77956743714ac18ad0463c960ba8d622479411ba5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 301, + 216 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 301, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 301, + 216 + ], + "type": "text", + "content": "Figure 6. Ablation on number of latent camera poses. We vary the number of latent camera poses " + }, + { + "bbox": [ + 55, + 182, + 301, + 216 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 182, + 301, + 216 + ], + "type": "text", + "content": " from 5, 9, 13, and 17, and compare with BAD-NeRF [56] and BAD-GS [70] in each setting." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 72, + 218, + 174 + ], + "blocks": [ + { + "bbox": [ + 139, + 72, + 218, + 174 + ], + "lines": [ + { + "bbox": [ + 139, + 72, + 218, + 174 + ], + "spans": [ + { + "bbox": [ + 139, + 72, + 218, + 174 + ], + "type": "image", + "image_path": "02ae19ec502b17aa1e5decd7a117e525e836373aff07661ff03dc115010f0e05.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 219, + 72, + 298, + 174 + ], + "blocks": [ + { + "bbox": [ + 219, + 72, + 298, + 174 + ], + "lines": [ + { + "bbox": [ + 219, + 72, + 298, + 174 + ], + "spans": [ + { + "bbox": [ + 219, + 72, + 298, + 174 + ], + "type": "image", + "image_path": "feef9527cf143b5b60abee2f8a7c77dac8abea66dca88c197bb097a610994dc0.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "lines": [ + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "type": "text", + "content": "Figure 7. Ablation on the impact of event threshold. We vary the threshold from " + }, + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "type": "inline_equation", + "content": "c = 0.1" + }, + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "type": "inline_equation", + "content": "c = 0.4" + }, + { + "bbox": [ + 307, + 182, + 553, + 215 + ], + "type": "text", + "content": ". Our method shows robustness to varying event thresholds since it does not rely solely on event data." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 316, + 72, + 392, + 174 + ], + "blocks": [ + { + "bbox": [ + 316, + 72, + 392, + 174 + ], + "lines": [ + { + "bbox": [ + 316, + 72, + 392, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 392, + 174 + ], + "type": "image", + "image_path": "ac827d0b3c523bcb7cc253c2008c57e4ccc5f1d5d2ab6ee4f9620d4ce5958751.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 392, + 72, + 470, + 174 + ], + "blocks": [ + { + "bbox": [ + 392, + 72, + 470, + 174 + ], + "lines": [ + { + "bbox": [ + 392, + 72, + 470, + 174 + ], + "spans": [ + { + "bbox": [ + 392, + 72, + 470, + 174 + ], + "type": "image", + "image_path": "0776ba992b5af5d7dc960cde74a694849fcdd9546f9a55743a4414e0ccc811f9.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 471, + 72, + 550, + 174 + ], + "blocks": [ + { + "bbox": [ + 471, + 72, + 550, + 174 + ], + "lines": [ + { + "bbox": [ + 471, + 72, + 550, + 174 + ], + "spans": [ + { + "bbox": [ + 471, + 72, + 550, + 174 + ], + "type": "image", + "image_path": "68aa3348edbb3046bf37f4fea86a3ad5915805725438bae75a567827bc1adc3f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 57, + 239, + 553, + 306 + ], + "blocks": [ + { + "bbox": [ + 85, + 227, + 522, + 237 + ], + "lines": [ + { + "bbox": [ + 85, + 227, + 522, + 237 + ], + "spans": [ + { + "bbox": [ + 85, + 227, + 522, + 237 + ], + "type": "text", + "content": "Table 2. Ablation on the loss functions. Best-performing results are highlighted in bold and second results in underline." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 239, + 553, + 306 + ], + "lines": [ + { + "bbox": [ + 57, + 239, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 57, + 239, + 553, + 306 + ], + "type": "table", + "html": "
Large-scaleMedium-scaleReal Scene
\\( \\mathcal{L}_{\\text{blur}} \\)\\( \\mathcal{L}_{\\text{event}} \\)\\( \\mathcal{L}_{\\text{depth}} \\)\\( \\mathcal{L}_{\\text{int}} \\)PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓BRISQUE↓NIQE↓PIQE↓MetaIQA↑RankIQA↓
✓✘✘✘24.980.7870.29929.100.8950.16757.328.79943.260.1986.125
✓✓✘✘25.710.7950.27529.940.9100.15556.168.58242.510.2015.066
✓✓✓✘25.540.8000.27130.050.9140.15456.178.49942.230.2004.923
✓✓✓✓26.020.8060.26830.470.9160.15253.968.37141.530.2134.895
", + "image_path": "7b6806ee00b81483a54903e2191d24af4d1eebd8e9421ea43b87e25f32b3dd99.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 58, + 355, + 293, + 395 + ], + "blocks": [ + { + "bbox": [ + 55, + 319, + 294, + 353 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 294, + 353 + ], + "type": "text", + "content": "Table 3. Ablation on the effectiveness of pose optimization. We calculate ATE between the initial COLMAP poses and GT poses, as well as ATE between the optimized poses and GT poses." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "lines": [ + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "spans": [ + { + "bbox": [ + 58, + 355, + 293, + 395 + ], + "type": "table", + "html": "
ATE↓PSNR↑SSIM↑LPIPS↓
w/o pose. opt0.383 ± 0.12524.790.7880.293
w/ pose. opt0.049 ± 0.01126.020.8060.268
", + "image_path": "82f13b489900f17fe6d50f699240baa53dc781cc264c4bc0f8cac7fb1573dc70.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 58, + 443, + 291, + 491 + ], + "blocks": [ + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "lines": [ + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": "Table 4. Ablation of different motion blur levels. Mild blur corresponds to " + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "inline_equation", + "content": "0.5 \\times" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": " camera motion speed, medium blur to the default speed, and strong blur to the " + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 55, + 407, + 295, + 440 + ], + "type": "text", + "content": " camera motion speed." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 443, + 291, + 491 + ], + "lines": [ + { + "bbox": [ + 58, + 443, + 291, + 491 + ], + "spans": [ + { + "bbox": [ + 58, + 443, + 291, + 491 + ], + "type": "table", + "html": "
Blur levelMild (0.5×)Medium (1.0×)Strong (2.0×)
PSNR↑26.3825.7125.04
SSIM↑0.8160.7950.789
LPIPS↓0.2690.2750.280
", + "image_path": "1abf781521a9f1e66d4d250fd475b417e4c55d13769828c558581fb2ab9a3ca7.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 580 + ], + "type": "text", + "content": "we also provide a comparison with BAD-NeRF [56] and BAD-GS [70]. These two methods typically use linear interpolation to obtain camera trajectory, while our camera trajectories are estimated from the decomposed latent images, which provides more accurate initialization and helps our method achieves better performance." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": "Effectiveness of The Loss Functions. We conduct novel view synthesis experiments on the challenging medium-scale and large-scale subsets of the synthetic dataset and the real-world dataset to validate the effectiveness of the training losses. The quantitative results are shown in Table. 2. It indicates that using only the blur reconstruction loss " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{blur}}" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": " leads to suboptimal outputs. In contrast, incorporating " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{event}}, \\mathcal{L}_{\\mathrm{depth}}" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 581, + 295, + 689 + ], + "type": "text", + "content": " enables our method to achieve better performance." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Analysis of Event Threshold. We conduct experiments on the medium-scale subset of the synthetic data to evaluate" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "text", + "content": "the impact of different event thresholds. We vary the threshold from " + }, + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "inline_equation", + "content": "c = 0.1" + }, + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "inline_equation", + "content": "c = 0.4" + }, + { + "bbox": [ + 313, + 321, + 553, + 380 + ], + "type": "text", + "content": ", and the results are shown in Figure 7. Since our method does not rely solely on event data and incorporates RGB inputs as supervision, it demonstrates robustness to varying event thresholds." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "text", + "content": "Robustness to Motion Blur Levels. To validate the robustness of our method in handling different levels of motion blur, we set up three different camera speeds in the City Blocks scene of the synthetic dataset to obtain images with varying levels of blur. Images with mild blur are captured at " + }, + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "inline_equation", + "content": "0.5 \\times" + }, + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "text", + "content": " camera motion speed, images with medium blur are captured at the default motion speed, and images with strong blur are captured at " + }, + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "inline_equation", + "content": "2.0 \\times" + }, + { + "bbox": [ + 313, + 381, + 554, + 524 + ], + "type": "text", + "content": " motion speed. The quantitative results of novel view synthesis are listed in Table. 4. It can be observed that the results show no significant fluctuations across different levels of blur, demonstrating the robustness of our method to varying motion blur levels." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 536, + 391, + 547 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 391, + 547 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 391, + 547 + ], + "type": "text", + "content": "5. Conclusions" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 555, + 555, + 711 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 711 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 711 + ], + "type": "text", + "content": "This paper introduces Event Stream Assisted Gaussian Splitting (EvaGaussians), a novel framework that seamlessly integrates the event streams captured by an event camera into the training of 3D-GS, effectively addressing the challenges of reconstructing high-quality 3D-GS from motion-blurred images. We contribute two novel datasets and conduct comprehensive experiments. The results demonstrate that our method outperforms previous state-of-the-art deblurring rendering techniques in terms of novel view synthesis quality, without sacrificing inference efficiency. Despite its promising performance, our method may still face challenges when reconstructing scenes with extremely intricate textures from severely blurred images." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24787" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 127 + ], + "type": "text", + "content": "This work was supported in part by Natural Science Foundation of China (No.62332002, No.62202014), and Shenzhen KQTD (No.20240729102051063)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 137, + 115, + 149 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 157, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 157, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 157, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 157, + 296, + 190 + ], + "type": "text", + "content": "[1] Patrick Bardow, Andrew J Davison, and Stefan Leutenegger. Simultaneous optical flow and intensity estimation from an event camera. In CVPR, 2016. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 191, + 296, + 245 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 296, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 296, + 245 + ], + "type": "text", + "content": "[2] Christian Brandli, Lorenz Muller, and Tobi Delbruck. Realtime, high-speed video decompression using a frame- and event-based davis sensor. In 2014 IEEE International Symposium on Circuits and Systems (ISCAS), pages 686–689, 2014. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[3] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3, 5, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 345 + ], + "type": "text", + "content": "[4] Mauro Comi, Alessio Tonioni, Max Yang, Jonathan Tremblay, Valts Blukis, Yijiong Lin, Nathan F Lepora, and Laurence Aitchison. Snap-it, tap-it, splat-it: Tactile-informed 3d gaussian splatting for reconstructing challenging surfaces. arXiv preprint arXiv:2403.20275, 2024. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 346, + 296, + 411 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 346, + 296, + 411 + ], + "spans": [ + { + "bbox": [ + 62, + 346, + 296, + 411 + ], + "type": "text", + "content": "[5] Zhenxuan Fang, Fangfang Wu, Weisheng Dong, Xin Li, Jinjian Wu, and Guangming Shi. Self-supervised non-uniform kernel estimation with flow-based motion prior for blind image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18105-18114, 2023. 5, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 62, + 413, + 295, + 467 + ], + "type": "text", + "content": "[6] Chaoran Feng, Wangbo Yu, Xinhua Cheng, Zhenyu Tang, Junwu Zhang, Li Yuan, and Yonghong Tian. Ae-nerf: Aug-mentation event-based neural radiance fields for non-ideal conditions and larger scene. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2924–2932, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 295, + 512 + ], + "type": "text", + "content": "[7] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE TPAMI, 2020. 2, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 295, + 555 + ], + "type": "text", + "content": "[8] Haiqian Han, Jianing Li, Henglu Wei, and Xiangyang Ji. Event-3dgs: Event-based 3d reconstruction using 3d gaussian splatting. In Advances in Neural Information Processing Systems, 2025. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 295, + 590 + ], + "type": "text", + "content": "[9] Philipp Heise, Sebastian Klose, Brian Jensen, and Alois Knoll. Pm-huber: Patchmatch with huber regularization for stereo matching. In ICCV, 2013. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 296, + 635 + ], + "type": "text", + "content": "[10] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1312-1321, 2021. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 636, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 667 + ], + "type": "text", + "content": "[11] Jian Huang, Chengrui Dong, and Peidong Liu. Inceventgs: Pose-free gaussian splatting from a single event camera. In CVPR, 2025. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 837–847, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[13] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In CVPR, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 138 + ], + "type": "text", + "content": "[14] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splattering for real-time radiance field rendering. ACM TOG, 2023. 1, 4, 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 553, + 171 + ], + "type": "text", + "content": "[15] Justin Kerr, Chung Min Kim, Ken Goldberg, Angjoo Kanazawa, and Matthew Tancik. Leref: Language embedded radiance fields. In CVPR, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 317, + 173, + 553, + 215 + ], + "type": "text", + "content": "[16] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 250 + ], + "type": "text", + "content": "[17] Joseph J LaViola Jr. Bringing vr and spatial 3d interaction to the masses through video games. IEEE Computer Graphics and Applications, 2008. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 251, + 553, + 282 + ], + "type": "text", + "content": "[18] Dogyoon Lee, Minhyeok Lee, Chajin Shin, and Sangyoun Lee. Dp-nerf: Deblurred neural radiance field with physical scene priors. In CVPR, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 553, + 327 + ], + "type": "text", + "content": "[19] Wenpu Li, Pian Wan, Peng Wang, Jinghang Li, Yi Zhou, and Peidong Liu. Benerf: neural radiance fields from a single blurry image and event stream. In European Conference on Computer Vision, pages 416-434. Springer, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 316, + 328, + 553, + 371 + ], + "type": "text", + "content": "[20] Yuchen Li*, Chaoran Feng*, Zhenyu Tang, Kaiyuan Deng, Wangbo Yu, Yonghong Tian, and Li Yuan. Gs2e: Gaussian splatting is an effective data generator for event stream generation. arXiv preprint arXiv:2505.15287, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 425 + ], + "type": "text", + "content": "[21] Guixu Lin, Jin Han, Mingdeng Cao, Zhihang Zhong, and Yinqiang Zheng. Event-guided frame interpolation and dynamic range expansion of single rolling shutter image. In Proceedings of the 31st ACM International Conference on Multimedia, 2023." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 459 + ], + "type": "text", + "content": "[22] Mingyuan Lin, Yangguang Wang, Xiang Zhang, Boxin Shi, Wen Yang, Chu He, Gui-song Xia, and Lei Yu. Self-supervised shutter unrolling with events. IJCV, 2025. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 502 + ], + "type": "text", + "content": "[23] Songnan Lin, Jiawei Zhang, Jinshan Pan, Zhe Jiang, Dongqing Zou, Yongtian Wang, Jing Chen, and Jimmy Ren. Learning event-driven video deblurring and interpolation. In ECCV, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 317, + 504, + 553, + 548 + ], + "type": "text", + "content": "[24] Kunhao Liu, Fangneng Zhan, Jiahui Zhang, Muyu Xu, Yingchen Yu, Abdulmotaleb El Saddik, Christian Theobalt, Eric Xing, and Shijian Lu. Weakly supervised 3d open-vocabulary segmentation. In NeurIPS, 2023. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[25] Minghua Liu, Ruoxi Shi, Kaiming Kuang, Yinhao Zhu, Xuanlin Li, Shizhong Han, Hong Cai, Fatih Porikli, and Hao Su. Openshape: Scaling up 3d shape representation towards open-world understanding. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 635 + ], + "type": "text", + "content": "[26] Xialei Liu, Joost Van De Weijer, and Andrew D Bagdanov. Rankiaq: Learning from rankings for no-reference image quality assessment. In Proceedings of the IEEE international conference on computer vision, pages 1040-1049, 2017. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 680 + ], + "type": "text", + "content": "[27] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[28] Li Ma, Xiaoyu Li, Jing Liao, Qi Zhang, Xuan Wang, Jue Wang, and Pedro V Sander. Deblur-NeRF: Neural Radiance Fields from Blurry Images. In CVPR, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24788" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[29] Nico Messikommer, Carter Fang, Mathias Gehrig, Giovanni Cioffi, and Davide Scaramuzza. Data-driven feature tracking for event cameras with and without frames. TPAMI, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 149 + ], + "type": "text", + "content": "[30] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. NeRF: Representing Scenes as Neural Radiance Fields for View Synthesis. In ECCV, 2020. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 183 + ], + "type": "text", + "content": "[31] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE TIP, 21(12):4695-4708, 2012. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 216 + ], + "type": "text", + "content": "[32] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal processing letters, 20(3):209-212, 2012. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 249 + ], + "type": "text", + "content": "[33] Genady Paikin, Yotam Ater, Roy Shaul, and Evgeny Soloveichik. Efi-net: Video frame interpolation from fusion of events and frames. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 293 + ], + "type": "text", + "content": "[34] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In CVPR, 2019. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 326 + ], + "type": "text", + "content": "[35] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In CVPR, 2020. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 369 + ], + "type": "text", + "content": "[36] Roberto Pellerito, Marco Cannici, Daniel Gehrig, Joris Bel-hadj, Olivier Dubois-Matra, Massimo Casasco, and Davide Scaramuzza. Deep visual odometry with events and frames. In IROS, 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 414 + ], + "type": "text", + "content": "[37] Cheng Peng and Rama Chellappa. Pdrf: Progressively deblurring radiance field for fast and robust scene reconstruction from blurry images. arXiv preprint arXiv:2208.08049, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 458 + ], + "type": "text", + "content": "[38] Cheng Peng, Yutao Tang, Yifan Zhou, Nengyu Wang, Xijun Liu, Deming Li, and Rama Chellappa. Bags: Blur agnostic gaussian splatting through multi-scale kernel modeling. In ECCV, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 492 + ], + "type": "text", + "content": "[39] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 493, + 294, + 525 + ], + "type": "text", + "content": "[40] Yunshan Qi, Lin Zhu, Yifan Zhao, Nan Bao, and Jia Li. Deblurring neural radiance fields with event-driven bundle adjustment. In MM, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 526, + 294, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 294, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 294, + 559 + ], + "type": "text", + "content": "[41] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In Conference on robot learning, pages 969-982. PMLR, 2018. 4, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 591 + ], + "type": "text", + "content": "[42] Antoni Rosinol, John J Leonard, and Luca Carlone. Nerf-slam: Real-time dense monocular slam with neural radiance fields. 2023. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[43] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 658 + ], + "type": "text", + "content": "[44] Johannes L Schonberger and Jan-Michael Frahm. Structure-from-motion Revisited. In CVPR, 2016. 1, 2, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[45] Zihang Shao*, Chaoran Feng*, Xuanye Fang, Yaxin Li, Jiangrong Shen, and Qi Xu. Eicil: joint excitatory inhibitory cycle iteration learning for deep spiking neural networks. Advances in Neural Information Processing Systems, 36: 32117-32128, 2023. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 128 + ], + "type": "text", + "content": "[46] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European Conference on Computer Vision, pages 412-428. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 172 + ], + "type": "text", + "content": "[47] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In ECCV, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 228 + ], + "type": "text", + "content": "[48] Zhenyu Tang*, Chaoran Feng*, Xinhua Cheng, Wangbo Yu, Junwu Zhang, Yuan Liu, Xiaoxiao Long, Wenping Wang, and Li Yuan. Neuralgs: Bridging neural fields and 3d gaussian splatting for compact 3d representations. arXiv preprint arXiv:2503.23162, 2025." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 295 + ], + "type": "text", + "content": "[49] Zhenyu Tang, Junwu Zhang, Xinhua Cheng, Wangbo Yu, Chaoran Feng, Yatian Pang, Bin Lin, and Li Yuan. Cycle3d: High-quality and consistent image-to-3d generation via generation-reconstruction cycle. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7320–7328, 2025. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[50] Gemma Taverni. Applications of Silicon Retinas: From Neuroscience to Computer Vision. PhD thesis, Universität Zürich, 2020. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[51] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In CVPR, pages 16155-16164, 2021. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 419 + ], + "type": "text", + "content": "[52] Manoj K Vairalkar and SU Nimbhorkar. Edge detection of images using sobel operator. International Journal of Emerging Technology and Advanced Engineering, 2(1):291-293, 2012. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 316, + 421, + 553, + 475 + ], + "type": "text", + "content": "[53] N Venkatanath, D Praneeth, Maruthi Chandrasekhar Bh, Sumohana S Channappayya, and Swarup S Medasani. Blind image quality evaluation using perception based features. In 2015 twenty first national conference on communications (NCC), pages 1-6. IEEE, 2015. 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 476, + 553, + 510 + ], + "type": "text", + "content": "[54] Zhexiong Wan, Bin Fan, Le Hui, Yuchao Dai, and Gim Hee Lee. Instance-level moving object segmentation from a single image with events. IJCV, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 555 + ], + "type": "text", + "content": "[55] Jiaxu Wang, Junhao He, Ziyi Zhang, Mingyuan Sun, Jingkai Sun, and Renjing Xu. Evggs: A collaborative learning framework for event-based generalizable gaussian splatting. In ICML, 2024. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 553, + 589 + ], + "type": "text", + "content": "[56] Peng Wang, Lingzhe Zhao, Ruijie Ma, and Peidong Liu. BAD-NeRF: Bundle Adjusted Deblur Neural Radiance Fields. In CVPR, 2023. 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 553, + 633 + ], + "type": "text", + "content": "[57] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 666 + ], + "type": "text", + "content": "[58] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[59] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher A Metzler. Event3dgs: Event-based 3d gaussian splattering for fast egomotion. arXiv preprint arXiv:2406.02972, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24789" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[60] Linning Xu, Vasu Agrawal, William Laney, Tony Garcia, Aayush Bansal, Changil Kim, Samuel Rota Bulò, Lorenzo Porzi, Peter Kontschieder, Aljaž Božić, et al. Vr-nerf: High-fidelity virtualized walkable spaces. In SIGGRAPH Asia, 2023. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[61] Wenhao Xu, Wenming Weng, Yueyi Zhang, Ruikang Xu, and Zhiwei Xiong. Event-boosted deformable 3d gaussians for fast dynamic scene reconstruction. arXiv preprint arXiv:2411.16180, 2024. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 207 + ], + "type": "text", + "content": "[62] Lin Yen-Chen, Pete Florence, Jonathan T Barron, Alberto Rodriguez, Phillip Isola, and Tsung-Yi Lin. inerf: Inverting neural radiance fields for pose estimation. 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 251 + ], + "type": "text", + "content": "[63] Mark YU, Wenbo Hu, Jinbo Xing, and Ying Shan. Trajectorycrafter: Redirecting camera trajectory for monocular videos via diffusion models. arXiv preprint arXiv:2503.05638, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[64] Wangbo Yu, Jinbo Xing, Li Yuan, Wenbo Hu, Xiaoyu Li, Zhipeng Huang, Xiangjun Gao, Tien-Tsin Wong, Ying Shan, and Yonghong Tian. Viewcrafter: Taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048, 2024." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 352 + ], + "type": "text", + "content": "[65] Wangbo Yu, Li Yuan, Yan-Pei Cao, Xiangjun Gao, Xiaoyu Li, Wenbo Hu, Long Quan, Ying Shan, and Yonghong Tian. Hifi-123: Towards high-fidelity one image to 3d content generation. In European Conference on Computer Vision, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 396 + ], + "type": "text", + "content": "[66] Sohaib Zahid, Viktor Rudnev, Eddy Ilg, and Vladislav Golyanik. E-3dgs: Event-based novel view rendering of large-scale scenes using 3d gaussian splatting. In 3DV, 2025. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[67] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 455, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 455, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 455, + 294, + 475 + ], + "type": "text", + "content": "[68] Xiang Zhang and Lei Yu. Unifying motion deblurring and frame interpolation with events. In CVPR, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 520 + ], + "type": "text", + "content": "[69] Zixin Zhang, Kanghao Chen, and Lin Wang. Elite-evgs: Learning event-based 3d gaussian splatting by distilling event-to-video priors. arXiv preprint arXiv:2409.13392, 2024. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 523, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 554 + ], + "type": "text", + "content": "[70] Lingzhe Zhao, Peng Wang, and Peidong Liu. Bad-gaussians: Bundle adjusted deblur gaussian splatting. In ECCV, 2024. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 600 + ], + "type": "text", + "content": "[71] Haiyang Zhou, Wangbo Yu, Jiawen Guan, Xinhua Cheng, Yonghong Tian, and Li Yuan. Holotime: Taming video diffusion models for panoramic 4d scene generation. arXiv preprint arXiv:2504.21650, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 634 + ], + "type": "text", + "content": "[72] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[73] Hancheng Zhu, Leida Li, Jinjian Wu, Weisheng Dong, and Guangming Shi. Metaiaq: Deep meta-learning for no-reference image quality assessment. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14143-14152, 2020. 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[74] Zihan Zhu, Songyou Peng, Viktor Larsson, Weiwei Xu, Hujun Bao, Zhaopeng Cui, Martin R Oswald, and Marc Polle" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 129 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "feys. Nice-slam: Neural implicit scalable encoding for slam. In CVPR, 2022. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[75] Yunhao Zou, Ying Fu, Tsuyoshi Takatani, and Yinqiang Zheng. Eventhrd: From event to high-speed hdr videos and beyond. TPAMI. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24790" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_content_list.json b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ee55cbeea1ab862cf953464bad489f69fe353d69 --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_content_list.json @@ -0,0 +1,1629 @@ +[ + { + "type": "text", + "text": "Evading Data Provenance in Deep Neural Networks", + "text_level": 1, + "bbox": [ + 236, + 130, + 759, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Hongyu Zhu $^{1*}$ Sichu Liang $^{2*}$ Wenwen Wang $^{3}$ Zhuomeng Zhang $^{1}$ Fangqi Li $^{1}$ Shi-Lin Wang $^{1\\dagger}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Southeast University $^{3}$ Carnegie Mellon University Code: https://github.com/dbsxfz/EscapingDOV", + "bbox": [ + 99, + 178, + 916, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Modern over-parameterized deep models are highly data-dependent, with large scale general-purpose and domain-specific datasets serving as the bedrock for rapid advancements. However, many datasets are proprietary or contain sensitive information, making unrestricted model training problematic. In the open world where data thefts cannot be fully prevented, Dataset Ownership Verification (DOV) has emerged as a promising method to protect copyright by detecting unauthorized model training and tracing illicit activities. Due to its diversity and superior stealth, evading DOV is considered extremely challenging. However, this paper identifies that previous studies have relied on oversimplistic evasion attacks for evaluation, leading to a false sense of security. We introduce a unified evasion framework, in which a teacher model first learns from the copyright dataset and then transfers task-relevant yet identifier-independent domain knowledge to a surrogate student using an out-of-distribution (OOD) dataset as the intermediary. Leveraging Vision-Language Models and Large Language Models, we curate the most informative and reliable subsets from the OOD gallery set as the final transfer set, and propose selectively transferring task-oriented knowledge to achieve a better trade-off between generalization and evasion effectiveness. Experiments across diverse datasets covering eleven DOV methods demonstrate our approach simultaneously eliminates all copyright identifiers and significantly outperforms nine state-of-the-art evasion attacks in both generalization and effectiveness, with moderate computational overhead. As a proof of concept, we reveal key vulnerabilities in current DOV methods, highlighting the need for long-term development to enhance practicality.", + "bbox": [ + 89, + 300, + 485, + 770 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 801, + 220, + 816 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The success of modern deep learning hinges heavily on abundant datasets, spanning from large-scale vision and multi-modal resources such as ImageNet [9] and LAION-", + "bbox": [ + 89, + 828, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5B [52] to datasets tailored for vertical industries like face recognition [37] and computational pathology [71]. Curated meticulously through extensive human effort for collection, cleaning, and labeling, these datasets are protected as intellectual property and are available solely through licensed access. Some are released for academic purposes with restrictions on commercial exploitation, while others are proprietary assets, strictly prohibiting third-party model training. Moreover, personal data—such as photos shared on social platforms—is increasingly at risk of being scraped without consent [44]. Consequently, unauthorized model training on restricted data poses an escalating threat to both intellectual property and privacy rights.", + "bbox": [ + 511, + 268, + 906, + 467 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite growing regulatory enforcement, like the General Data Protection Regulation (GDPR) [10], preventing unauthorized training remains a formidable challenge. Every stage of the data supply chain is vulnerable to attacks that funnel data into illicit markets [66], including insider threats [59], side-channel exploits [21], illegal mobile app access [80], and leakage from pretrained models [27, 70]. High-profile cases of data misuse in model training, such as the Cambridge Analytica scandal [69] and the Flickr Leakage [44], have intensified public scrutiny.", + "bbox": [ + 511, + 470, + 908, + 621 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Consequently, preventing unauthorized model training at its source—the dataset itself—is crucial. Unlearnable Examples [23] introduce perturbations to prevent models from learning meaningful features. However, beyond their susceptibility to adaptive attacks [51], they entirely obstruct legitimate dataset use. As an alternative, post hoc verification for data provenance, known as Dataset Ownership Verification (DOV), has garnered substantial attention. Inspired by watermarks used to track duplication, remixing, or exploitation of multi-modal content [6], the backdoor watermark embeds carefully crafted \"poison\" samples selected by the dataset owner. Models trained on watermarked data exhibit predefined behaviors in response to backdoor triggers, enabling model-agnostic owner attribution [33, 63].", + "bbox": [ + 511, + 625, + 908, + 837 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To mitigate the impact of backdoor on authorized training, recent techniques introduce non-poisoning watermarks that adjust prediction confidence without inducing misclassification, employing hypothesis testing for ownership trac", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution. † Corresponding author.", + "bbox": [ + 107, + 886, + 346, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "1249", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [50, 68, 79]. Non-intrusive dataset fingerprint have also emerged, allowing verification via elevated confidence on memorized training samples [35, 42]. DOV is widely regarded as an effective, if not the only, solution for tracking unauthorized training without disrupting legitimate usage [17, 32], with rapid advancements underway.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "However, when deployed in the real world, the arms race between attackers and defenders forms an ongoing strategic game. Adversaries attempt to suppress verification behaviors to evade detection, while defenders continually strengthen verification mechanisms to reliably identify unauthorized training. Validating the robustness of DOV against evasion attempts is thus essential for practical deployment. Despite substantial progress in DOV, advances in evasion attacks remain underdeveloped, relying on simple regularization techniques and generic adaptations from poisoning defenses like fine-tuning [17, 32, 35, 79], which lack the versatility to counter diverse DOV techniques. Meanwhile, advanced DOV strategies, such as clean-label, invisible, or untargeted backdoor watermarks provide exceptional flexibility; non-poisoning watermarks and dataset fingerprints, which avoid misclassification or require no dataset modifications, remain nearly undetectable. It is widely agreed that designing a universal attack capable of evading all forms of verification is nearly impossible [42, 68]. This gap between the advancement of DOV and evasion techniques hinders reliable assessments.", + "bbox": [ + 91, + 185, + 486, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper we ask: is current DOV robust enough to preclude any evasion attempt? Our findings reveal that prior evasion attacks are too weak, leading to a false sense of security. Through analysis of verification behaviors, we identify shared characteristics: they are both exclusive and subtle, serving as unique markers that distinguish the protected dataset while remain incompatible with underlying semantic distribution of the main task.", + "bbox": [ + 89, + 505, + 483, + 625 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Building on this insight, we propose the first universal evasion framework, Escaping DOV, which requires no additional clean in-distribution data or assumptions about the verification process. In Escaping DOV, a teacher model is initially trained directly on the copyright dataset to fully absorb domain knowledge. Then, an OOD transfer set, unrelated to the protected data, serves as an intermediary to extract task-oriented yet identifier-free knowledge to a surrogate student for deployment. Intuitively, exclusive and subtle identifiers, acting as side-channel signals orthogonal to the task distribution, make it improbable for OOD samples to trigger verification behaviors. Thus, only essential, identifier-invariant task knowledge is likely to transfer to the student model. To enhance Escaping DOV, we propose Transfer Set Curation, leveraging the unbiased knowledge embedded in large language models (LLMs) and vision-language models (VLMs) to retrieve the most informative and reliable samples from a general OOD gallery set. Ad", + "bbox": [ + 89, + 628, + 486, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ditionally, we introduce Selective Knowledge Transfer to block suspicious verification behaviors, achieving better balance between generalization and evasion. The overall pipeline for Escaping DOV is shown in Figure 1.", + "bbox": [ + 511, + 90, + 906, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are summarized as follows:", + "bbox": [ + 532, + 152, + 834, + 166 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. We reveal that current DOV robustness evaluations are insufficiently rigorous. We introduce Escaping DOV, the first universal evasion framework that achieves simplicity, efficacy, and generalizability.", + "2. We propose Transfer Set Curation and Selective Knowledge Transfer to achieve balance between generalization and evasion with moderate overhead.", + "3. Experiments on diverse datasets and 11 DOV methods, alongside comparisons with 9 SOTA evasion methods, validate the efficacy of Escaping DOV, establishing it as a reliable benchmark for future DOV assessments." + ], + "bbox": [ + 513, + 167, + 903, + 333 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 351, + 655, + 367 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dataset Ownership Verification", + "text_level": 1, + "bbox": [ + 513, + 376, + 790, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Backdoor attacks inject poisoned samples to associate predictions with a specific trigger [16], which induces misclassification whenever the trigger appears [15]. When controlled by copyright owners, backdoors can also act as identifiers for model ownership [1]. Inspired by this, the straightforward backdoor watermark embeds backdoor samples in the dataset. Models trained on this marked data produce a predefined label with high probability when evaluated on trigger samples [33]. The defender's knowledge of this secret prediction behavior serves as proof of ownership.", + "bbox": [ + 511, + 400, + 906, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early backdoors link triggers to target classes by manipulating the labels of poisoned samples. However, due to the obvious nature of poisoned labels, these samples are easily filtered through automatic data sanitization [73]. Clean-label backdoor watermark [63] utilizes advanced poisoning techniques [65] to enforce backdoor behavior without modifying labels. Later advancements further constrain modifications within an $L_{\\infty}$ norm ball [74], making such clean-label, invisible backdoor watermarks highly stealthy [73]. Recently, untargeted backdoor watermark [32] induces nondeterministic misclassification without a fixed target class, enhancing resilience against post-training defenses such as trigger synthesis and output distribution analysis.", + "bbox": [ + 511, + 551, + 905, + 748 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Advanced backdoor watermark provides a stealthy and resilient verification mechanism. However, it inevitably embeds harmful shortcuts [76], problematic when authorized training is essential [17]. Non-poisoning watermark subtly embeds watermark features into the dataset without inducing misclassification. Verification is performed by measuring loss differences between watermarked and clean samples via hypothesis testing. Various mechanisms have been developed to generate stealthy watermarks: Radioactive data [50] and ML Auditor [24] optimize", + "bbox": [ + 511, + 750, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "1250", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3ce822d7d903579f10901ce1d1ad52061dc5d06557d686f6c473c2304b50b515.jpg", + "image_caption": [ + "Figure 1. Overall Pipeline of Escaping DOV. (a) Transfer Set Curation ( $\\S 3.2$ ): Samples from the OOD gallery set are clustered via a VLM, ranked by distribution proximity, and filtered through consensus ensemble to curate an optimal transfer set. (b) Selective Knowledge Transfer ( $\\S 3.3$ ): Promoting invariance to teacher's worst perturbations, facilitating task-oriented yet identifier-free knowledge transfer." + ], + "image_footnote": [], + "bbox": [ + 99, + 92, + 903, + 352 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "$L_{\\infty}$ -perturbations, Anti-Neuron Watermarking [79] applies color transformations in hue space, Domain Watermark [17] converts samples into a hardly-generalizable domain, while Isotope [68] blends external features into the dataset.", + "bbox": [ + 89, + 407, + 482, + 469 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Additionally, dataset fingerprint utilizes intrinsic features for verification without modifying the dataset. Dataset inference [42] analyzes the margins between training and external samples relative to the decision boundary, while MeFA [35] aggregates membership inference across multiple samples. Although dataset fingerprints are non-intrusive and support post hoc protection for published datasets, they are susceptible to high false positives [61, 62].", + "bbox": [ + 89, + 470, + 483, + 592 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Evasion Attacks against Data Provenance", + "text_level": 1, + "bbox": [ + 89, + 604, + 444, + 621 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Despite rapid progress in DOV, current evasion strategies rely on basic regularization and backdoor defenses like finetuning [17, 32, 36]. As they barely evade any DOV methods, we present their results in Appendix C.1 and use SOTA backdoor defenses as baselines, which are categorized by stage of application: (1) Data sanitization removes suspicious samples pre-training [73]; (2) Robust learning mitigates poisoning effects during training [4, 31]; (3) Backdoor unlearning synthesizes triggers post-training to neutralize the model [72]; and (4) Input preprocessing detects or purifies triggers during testing [56]. Together, these defenses span the model lifecycle; yet advanced invisible, clean-label, or untargeted backdoor watermarks can bypass most of them [17, 32]. Further, non-poisoning watermarks and non-intrusive fingerprints render backdoor defenses ineffective, making universal evasion extremely challenging. This work is the first to establish a unified evasion perspective, opening new directions for robust evaluation of DOV.", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We discuss the contemporaneous works [77] and [55] that emerged at the time of acceptance in Appendix E.", + "bbox": [ + 511, + 409, + 906, + 439 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Escaping DOV", + "text_level": 1, + "bbox": [ + 513, + 455, + 661, + 473 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 513, + 482, + 624, + 497 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In $K$ -class image classification, an adversary gains unauthorized access to the copyright dataset $\\mathcal{D}$ to train a model $f_{\\theta}$ , which maps images $x \\in \\{0,1,\\ldots,255\\}^{C \\times W \\times H}$ to predictions $y \\in \\mathbb{R}^K$ . For data provenance, the defender employs a DOV method $\\operatorname{Verif}(f_{\\theta},\\mathcal{D},\\mathcal{M})$ , inspecting the output of the suspicious model $f_{\\theta}$ on specific queries. This process yields a verification metric indicating the likelihood that $f_{\\theta}$ was trained on $\\mathcal{D}$ , such as the Attack Success Rate for backdoor watermarks or the confidence for rejecting the null hypothesis in non-poisoning DOV. Here, $\\mathcal{M}$ represents external materials (e.g., watermark trigger patterns) controlled by the defender to uniquely identify the dataset.", + "bbox": [ + 511, + 507, + 906, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "DOV methods are flexible and diverse. Backdoor watermark achieves superior stealth through invisible, clean-label, or untargeted poison techniques. Non-poisoning watermarks avoid misclassification, while dataset fingerprints require no dataset modifications. Hence, direct sanitization or suppression of all potential verification behaviors becomes a formidable challenge; most evasion attempts that simply adapt poison defenses have consistently failed [17, 32, 42, 68, 79]. However, we identify a common trait among all DOV strategies: verification behaviors are both exclusive and subtle. On one hand, these behaviors must exclusively belong to the copyright dataset to prevent false attribution to models trained on external datasets. On the other hand, they must be subtle enough to remain concealed", + "bbox": [ + 511, + 689, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "1251", + "bbox": [ + 483, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "during data sanitization, ensuring they are rarely triggered by regular samples to preserve the generalization of models trained for authorized use.", + "bbox": [ + 89, + 90, + 480, + 135 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Accordingly, we propose a universal evasion strategy: first, a teacher model $f_{\\theta_t}$ is trained on the copyright dataset $\\mathcal{D}$ to absorb task-domain knowledge, inevitably marked by DOV. Next, an unrelated OOD transfer set $\\mathcal{T}$ serves as the medium to extract task-oriented yet identifier-invariant knowledge into a surrogate student $f_{\\theta_s}$ . Intuitively, the exclusive and subtle watermark triggers are absent in the natural samples of $\\mathcal{T}$ , preventing the activation and transmission of verification behaviors to the surrogate student. For fingerprints based on exaggerated confidence, the student only indirectly inherits knowledge from the marked teacher via the unrelated OOD data, significantly reducing excessive memorization associated with $\\mathcal{D}$ .", + "bbox": [ + 89, + 136, + 482, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The overall objective of Escaping DOV can be formalized as simultaneously minimizing the generalization error on the task distribution and the verification metric:", + "bbox": [ + 89, + 332, + 482, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\theta_ {s}} \\mathbb {E} _ {(x, y) \\in P (x, y)} [ \\mathcal {L} \\left(f _ {\\theta_ {s}} (x), y\\right) ] + \\alpha \\cdot \\operatorname {V e r i f} \\left(f _ {\\theta_ {s}}, \\mathcal {D}, \\mathcal {M}\\right) \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 386, + 482, + 409 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $f_{\\theta_s}$ represents the surrogate student ultimately deployed, $P(x,y)$ the underlying distribution of samples $x,y \\in \\mathcal{D}$ , and $\\mathcal{L}$ the loss function measuring prediction discrepancies. The tuning factor $\\alpha$ balances the pursuit of generalization and evasion for $f_{\\theta_s}$ . However, $P(x,y)$ and the process of $\\mathrm{Verif}(f_{\\theta_s},\\mathcal{D},\\mathcal{M})$ are agnostic to the adversary and intractable during optimization. Therefore, we propose two modules to better achieve and balance these objectives: (1) For generalization, we curate the most informative and reliable transfer set from a large-scale OOD gallery set to support high-performance and identifier-invariant knowledge extraction; (2) For evasion, to prevent the transmission of suspicious verification knowledge via dark knowledge in soft labels, we propose selective knowledge transfer to filter out such knowledge and better trade off between the two objectives. Next, we elaborate on these two components.", + "bbox": [ + 89, + 417, + 482, + 659 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Transfer Set Curation", + "text_level": 1, + "bbox": [ + 89, + 667, + 295, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Our objective is to curate an optimal subset $\\mathcal{T}$ from an OOD gallery set $\\mathcal{G}$ (e.g., ImageNet [9] or DataComp [12]), selecting samples that are both informative and reliable to facilitate effective knowledge transfer. While this task may resemble classical core-set selection [5, 53], it presents significantly greater challenge: (1) Our algorithm operates in a fully OOD context, where $\\mathcal{G}$ and $\\mathcal{D}$ do not necessarily overlap in classes, and their distributions could vary substantially; (2) $\\mathcal{D}$ could include identifiers manipulated by the defender (e.g., trigger samples), and the teacher $f_{\\theta_t}$ trained directly on $\\mathcal{D}$ is already marked. Consequently, selection based on the unreliable $\\mathcal{D}$ and $f_{\\theta_t}$ inadvertently introduce samples from $\\mathcal{G}$ that activate verification behaviors, risking the induction of these behaviors in the surrogate student $f_{\\theta_s}$ .", + "bbox": [ + 89, + 688, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this paper, we leverage a vision-language model (VLM) with unbiased pretrained knowledge and robust zero-shot capabilities (e.g., CLIP [49]) to facilitate reliable selection alongside the marked teacher $f_{\\theta_t}$ . CLIP aligns visual and textual features in a shared embedding space, where the image encoder $\\phi_I(\\cdot)$ maps an image $x$ into $\\phi_I(x)$ , and text templates with category names (e.g., \"a photo of {class name}\") $c_1, \\ldots, c_K \\in \\mathcal{D}$ are mapped by the text encoder $\\phi_t(\\cdot)$ to $\\phi_t(c_1), \\ldots, \\phi_t(c_K)$ . By calculating the cosine similarity between $\\phi_I(x)$ and $\\phi_t(c_i)$ , each sample $x \\in \\mathcal{G}$ can be uniquely assigned to a category $t$ in $\\mathcal{D}$ .", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, class-name-only text templates lack foreground semantics [48], limiting the ability to distinguish target categories. While few-shot tuning allows the VLM to rapidly adapt to tasks within $\\mathcal{D}$ [26], it risks being compromised by verification behaviors embedded in $\\mathcal{D}$ [3]. Inspired by advances in zero-shot prompt learning [48], we adjust the VLM with unbiased category semantics from a large language model (LLM). The LLM generates a description set $\\mathrm{Desc}_{c_i}$ for each category $c_{i}$ in $\\mathcal{D}$ , averaging descriptions in the feature space to create representation prototypes for each category, enhancing the zero-shot capabilities of the VLM. Consequently, each $x\\in \\mathcal{G}$ is assigned to class $t$ in a zero-shot fashion as follows in Equation 2:", + "bbox": [ + 511, + 257, + 903, + 454 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nt = \\arg \\max _ {c _ {i}} \\frac {1}{\\left| \\operatorname {D e s c} _ {c _ {i}} \\right|} \\sum_ {d \\in \\operatorname {D e s c} _ {c _ {i}}} (\\sin (\\phi_ {I} (x), \\phi_ {I} (d))) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 459, + 903, + 500 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where sim denotes cosine similarity shown in Equation 3:", + "bbox": [ + 511, + 506, + 893, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {s i m} \\left(\\phi_ {I} (x), \\phi_ {I} (d)\\right) = \\frac {\\phi_ {I} (x) \\cdot \\phi_ {I} (d)}{\\left| \\phi_ {I} (x) \\right| \\cdot \\left| \\phi_ {I} (d) \\right|} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 527, + 903, + 561 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although the VLM assigns all gallery set samples to categories in $\\mathcal{D}$ , it lacks a criterion for the most informative samples. To leverage the distribution from $\\mathcal{D}$ while avoiding biases from potential verification behaviors, we project all samples in $\\mathcal{D}$ into the feature space with $\\phi_I(\\cdot)$ and calculate the density centroids $\\mathrm{Cent}_t$ for each class $t\\in \\{1,\\dots ,K\\}$ . These centroids, acting as distribution digests, encapsulate the conditional distribution of $\\mathcal{D}$ . Samples closer to $\\mathrm{Cent}_t$ in feature space are prioritized for selection, a robust and straightforward criterion that is theoretically resistant to manipulation by a small proportion of outlier samples [25].", + "bbox": [ + 511, + 568, + 903, + 733 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then establish a consensus-based ensemble for final transfer set curation: (1) The gallery set $\\mathcal{G}$ is divided by the LLM-enhanced VLM into $K$ bins corresponding to classes in $\\mathcal{D}$ . Within each bin $\\mathcal{G}_t$ , (2) samples are ordered by proximity to distribution digest $\\mathrm{Cent}_t$ ; (3) Only samples classified by the teacher $f_{\\theta_t}$ as $t$ that are closest to $\\mathrm{Cent}_t$ are selected, until the selected samples match the class count $|\\mathcal{D}_t|$ . Illustrated in Figure 1, this process ensures all predictions on the transfer set $\\mathcal{T}$ by $f_{\\theta_t}$ are consistent with VLM, minimizing the risk of triggering verification behaviors while closely approximating the distribution in $\\mathcal{D}$ .", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "1252", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The Feature Bank. The curation process involves projecting all gallery set samples into the feature space, which is the most time-consuming step. However, when the gallery set is potentially reused, this projection only needs to be performed once. The resulting features and indexes are saved in a feature bank for efficient reuse.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Selective Knowledge Transfer", + "text_level": 1, + "bbox": [ + 89, + 190, + 354, + 207 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The transfer set $\\mathcal{T}$ serves as an intermediary for extraction of task-oriented and identifier-invariant knowledge. This process employs the standard knowledge distillation framework [20] with the following optimization objective:", + "bbox": [ + 89, + 212, + 483, + 273 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 281, + 483, + 314 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}$ measures prediction discrepancies, such as the Kullback-Leibler divergence, modulated by the temperature $\\tau$ . This process effectively transfers knowledge despite potential misalignments between the transfer set $\\mathcal{T}$ and the training distribution [46]. However, the soft labels $f_{\\theta_t}(x)$ contain dark knowledge [11], which inadvertently transmit verification behaviors. Thus, we introduce Selective Knowledge Transfer (SKT) to filter out suspicious verification knowledge during distillation.", + "bbox": [ + 89, + 321, + 483, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The exclusive and subtle verification behaviors reflect unique biases inherent to the marked dataset $\\mathcal{D}$ , distinguishing it within the task distribution. Preventing the student $f_{\\theta_s}$ from overfitting to biases limits DOV's ability to infer connections between $f_{\\theta_s}$ and $\\mathcal{D}$ through trigger-specific outputs or excessive confidence. Drawing on insights from shortcut learning [14], we mitigate predictive shortcuts during the student's learning process, reducing reliance on spurious features. Since dataset biases typically induce universal shortcut behaviors, a straightforward solution is to apply universal adversarial training (UAT) on the student:", + "bbox": [ + 89, + 458, + 483, + 623 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {T}} \\mathcal {L} \\left(f _ {\\theta_ {s}} (x + \\delta), y\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 631, + 482, + 667 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "UAT effectively suppresses atypical predictive behaviors in $f_{\\theta_s}$ induced by perturbations $\\delta$ . However, the min-max problem involves a notoriously hard bi-level optimization, requiring iterative solutions within each iteration. Furthermore, the absence of ground-truth labels in the OOD transfer set $\\mathcal{T}$ necessitates predictions from the marked teacher as labels, introducing significant calibration errors.", + "bbox": [ + 89, + 676, + 483, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given the objective to mitigate biases specific to the copyrighted dataset $\\mathcal{D}$ , and with all verification behaviors encapsulated in the teacher model, we propose using only the teacher model $f_{\\theta_t}$ to generate perturbations on $\\mathcal{D}$ as a practical approximation for the inner max-problem:", + "bbox": [ + 89, + 782, + 483, + 858 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\delta^ {*} = \\arg \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {D}} \\mathcal {L} \\left(f _ {\\theta_ {t}} (x + \\delta), y\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 866, + 482, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This strategy decouples the bi-level optimization, rendering it asynchronously solvable. By pre-generating a perturbation pool $\\{\\delta_i\\}_{i=1}^n$ in an offline manner with the teacher $f_{\\theta_t}$ and reusing it in the outer min-problem during distillation, computational overhead is significantly reduced.", + "bbox": [ + 511, + 90, + 903, + 166 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To solve Equation 6, the perturbation $\\delta$ is typically constrained by an $L_{p}$ norm to preserve core image semantics [40]. However, without prior knowledge of the specific form of $\\delta$ associated with verification behaviors (e.g., watermark triggers), choosing an inappropriate norm undermine the efficacy of UAT in suppressing spurious features [64]. Thus, we seek to diversify the teacher-generated perturbations beyond a single norm constraint.", + "bbox": [ + 511, + 167, + 905, + 286 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We solve the max-problem using mini-batch stochastic gradient descent [54], applying $L_{2}$ constraints at each batch update to preserve gradient directions. After each iteration, we project the perturbation onto the norm constraint $(L_0, L_2, \\text{or} L_\\infty)$ that maximizes the loss. This steepest descent projection [41] adaptively seeks the optimal norm constraint, facilitating the generation of perturbations closely resembling those employed by DOV. Moreover, extreme projections onto $L_{0}$ or $L_{\\infty}$ norms approximate a convex hull spanning multiple norm constraints, broadening the range of potential perturbations [7].", + "bbox": [ + 511, + 287, + 905, + 452 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Ideally, perturbation should preserve core image semantics without being limited to specific norms [57]. Generative models can achieve this by producing unrestricted adversarial perturbations [76], though training such models on task-specific distributions is resource-intensive. Therefore, we propose a lightweight approach using image corruptions instead of generative models to produce unrestricted perturbations. Starting with the 15 corruptions defined in ImageNet-C [19], we employ a genetic algorithm [8, 43] to select the combination and sequence of corruptions that solves Equation 6, forming a corruption chain:", + "bbox": [ + 511, + 454, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\max \\mathcal {L} \\left(f _ {\\theta_ {t}} \\left(\\operatorname {C o r r u p t} _ {1: N} (x)\\right), y\\right), \\quad x, y \\in \\mathcal {D} \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 631, + 903, + 648 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The perturbations and corruption chain generated with the teacher $f_{\\theta_t}$ on $\\mathcal{D}$ are collectively referred to as $A(\\cdot)$ , leading to the final objective for Selective Knowledge Transfer:", + "bbox": [ + 511, + 659, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right) \\tag {8} \\\\ + \\beta \\cdot \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (A (x))}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 715, + 903, + 784 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The second term promotes the student's invariance to the worst-case perturbations from the teacher while encouraging predictive divergence to reduce bias toward the teacher. The tuning factor $\\beta$ balances generalization with evasion efficacy, applied as a sampling probability to selectively introduce perturbations or corruption chain to a subset of samples, adding minimal overhead.", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "1253", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/88faa8a65205d756da8fbb972754f8c7bd50b6b01ad8eede06acbb420ee48a02.jpg", + "table_caption": [ + "Table 1. Escaping DOV on CIFAR-10 and Tiny ImageNet. Red indicates detection by DOV, while green indicates successful evasion." + ], + "table_footnote": [], + "table_body": "
DOVCIFAR-10Tiny-Imagenet
VanillaEvasionVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)ACCVSRACC(↑)VSR(↓)
Badnets94.44 ± 0.74100.00 ± 0.0093.46 ± 0.201.36 ± 0.0769.05 ± 0.2598.98 ± 0.1166.48 ± 0.141.48 ± 0.11
UBW93.98 ± 0.2595.54 ± 0.8493.41 ± 0.191.74 ± 0.0666.37 ± 0.1275.56 ± 0.5065.34 ± 0.084.60 ± 0.11
Label-Consistent94.74 ± 0.5496.18 ± 2.7093.19 ± 0.143.74 ± 0.4069.07 ± 0.2434.99 ± 1.5366.46 ± 0.103.51 ± 0.22
Narcissus94.76 ± 0.5487.34 ± 1.8294.37 ± 0.184.59 ± 0.6468.99 ± 0.1457.49 ± 0.7866.39 ± 0.1812.33 ± 6.45
Non-Poisoning DOVACCp-valueACC(↑)p-value(↑)ACCp-valueACC(↑)p-value(↑)
Radioactive Data94.96 ± 0.473.03e-0394.07 ± 0.189.45e-0168.84 ± 0.305.33e-0366.64 ± 0.244.94e-01
ANW94.66 ± 0.311.37e-0993.91 ± 0.281.00e+0068.75 ± 0.305.57e-2866.44 ± 0.131.00e+00
Domain Watermark94.38 ± 0.531.67e-2293.90 ± 0.131.00e+0068.55 ± 0.393.91e-1666.38 ± 0.161.00e+00
Isotope94.75 ± 0.282.87e-0393.99 ± 0.492.84e-0168.97 ± 0.371.63e-0366.76 ± 0.221.66e-01
ML Auditor94.22 ± 0.177.47e-0593.45 ± 0.217.83e-0168.41 ± 0.281.55e-0665.35 ± 0.161.00e+00
Dataset Inference94.83 ± 0.471.87e-0393.97 ± 0.494.76e-0169.15 ± 0.349.27e-1466.06 ± 0.184.47e-01
MeFA94.83 ± 0.472.62e-1493.93 ± 0.551.00e+0069.15 ± 0.346.17e-2766.22 ± 0.209.99e-01
", + "bbox": [ + 96, + 108, + 901, + 325 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 330, + 222, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 357, + 279, + 373 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following established DOV methods [17, 33, 42, 79], we adopt ResNet-18 [18] for both teacher and students. For transfer set curation, we utilize MobileCLIP [67] and GPT-40 mini [45] as lightweight VLM and LLM, respectively.", + "bbox": [ + 89, + 380, + 482, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We assess backdoor watermarks with Verification Success Rate (VSR)-the probability of a trigger-induced misclassification to the target class. A VSR above $30\\%$ confirms successful verification [76], indicating unauthorized training. Non-poisoning watermarks and fingerprints, collectively termed non-poisoning DOV, are evaluated via the p-value from a one-tailed T-test. The p-value represents the confidence level that the model was not trained on the copyright dataset; lower p-values indicate successful verification. The common threshold for p-values is set at 0.01 [42], equating to a $99\\%$ confidence level of unauthorized training. Experiments were repeated 3 times, reporting mean $\\pm$ std for ACC and VSR, and harmonic mean for p-values.", + "bbox": [ + 88, + 441, + 482, + 638 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Escaping Advanced DOV Methods", + "text_level": 1, + "bbox": [ + 89, + 651, + 393, + 667 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate Escaping DOV on natural image datasets commonly used as DOV benchmarks, including CIFAR-10 [28] and the 200-class Tiny ImageNet [29]. We further consider six datasets with significant distribution shifts from the gallery set (e.g., facial recognition and medical diagnosis), with a summary of results in Section 4.4.5 and detailed findings in Appendix C.3. For the gallery set, we use LSVRC-2012 [9]. Given overlapping samples between Tiny ImageNet and LSVRC-2012, we create a challenging setting by removing all overlapping classes from LSVRC-2012 to produce a purely OOD gallery set. The targeted DOV approaches cover advanced algorithms across all categories: (1) backdoor watermarks, including BadNets [16], UBW [32], Label-Consistent [65], and Narcissus [74], encompassing clean-label, invisible, and untargeted poisoning", + "bbox": [ + 89, + 674, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "techniques; (2) non-poisoning watermarks, including Radioactive Data [50], ANW [79], Domain Watermark [17], Isotope [68], and ML Auditor [24]; and (3) fingerprints, including Dataset Inference [42] and MeFA [35].", + "bbox": [ + 511, + 332, + 903, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1 presents the generalization and verification metrics for the models before and after evasion attacks (the teacher and surrogate student in Escaping DOV). All DOV methods successfully identify the vanilla teacher trained directly on the copyright dataset. However, none meet their verification thresholds (VSR > 30% or p-value < 0.01) on the student produced by Escaping DOV, where VSR is near random guesses and p-values exceed 0.1 in most cases. Moreover, Escaping DOV incurs minimal impact on generalization: test accuracy on CIFAR-10 decreases by less than 1%, while on the challenging Tiny ImageNet set, with no overlapping gallery classes, accuracy drops by only 2-3%. Thus, Escaping DOV achieves a robust balance between universal evasion and preserved generalization.", + "bbox": [ + 511, + 393, + 905, + 606 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f858d9435b29ef401f6b61bff1b7372c374673963b9ea698120d10e8d6e344ca.jpg", + "image_caption": [ + "Figure 2. Loss Barrier between Teacher and Student Parameters." + ], + "image_footnote": [], + "bbox": [ + 519, + 609, + 898, + 724 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We further observe that Escaping DOV encourages the student to diverge from the teacher's predictive behavior, forming its own mechanism. In Figure 2, we linearly interpolate between teacher and student parameters, computing training and test losses at each point. The results reveal two key findings: (1) A loss barrier appears during interpolation from teacher $\\theta_{t}$ to student $\\theta_{s}$ . This lack of linear connectivity suggests that the surrogate student develops a distinct prediction mechanism, bypassing the spurious features central to the teacher's verification behavior [39]. (2)", + "bbox": [ + 511, + 750, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "1254", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/4ce117b5151f7e7a8159e353331ae2cafffe9284d2fce273cfd373555ed8133e.jpg", + "table_caption": [ + "Table 2. Comparison of Escaping DOV with Evasion Attacks. Red indicates detection by DOV, green indicates successful evasion." + ], + "table_footnote": [], + "table_body": "
MethodBadnetsNarcissusIsotopeDataset Inference
ACC(↑)VSR(↓)ACC(↑)VSR(↓)ACC(↑)p-value(↑)ACC(↑)p-value(↑)
Fine-pruning86.21 ± 0.1699.37 ± 0.2786.58 ± 0.1543.75 ± 14.4186.78 ± 0.120.053286.88 ± 0.480.0095
Meta-Sift88.68 ± 0.4599.70 ± 0.3288.64 ± 0.6249.40 ± 8.2688.56 ± 0.470.026488.07 ± 0.210.0963
Differential Privacy90.49 ± 1.2076.70 ± 4.6287.70 ± 0.6872.77 ± 3.9188.69 ± 2.610.067889.64 ± 4.860.0440
I-BAU89.09 ± 0.2514.80 ± 0.4688.76 ± 1.9472.74 ± 2.1787.07 ± 2.590.155289.93 ± 0.680.0355
ZIP82.44 ± 0.1079.88 ± 0.3983.51 ± 0.3753.21 ± 4.4782.97 ± 0.420.009883.21 ± 0.900.1050
NAD89.86 ± 0.373.11 ± 0.5191.04 ± 0.3675.41 ± 5.1291.77 ± 0.450.096192.12 ± 0.340.0014
BCU92.75 ± 0.051.56 ± 0.2592.33 ± 0.1161.72 ± 6.4392.77 ± 0.360.162293.09 ± 0.180.0773
ABD84.94 ± 3.047.07 ± 3.1585.42 ± 2.1022.10 ± 1.1586.29 ± 3.460.006684.87 ± 3.020.3696
IPRemoval83.96 ± 0.238.09 ± 0.5085.06 ± 0.1155.97 ± 2.1484.64 ± 0.060.191685.28 ± 0.050.1322
Escaping DOV (Ours)93.46 ± 0.201.36 ± 0.0794.37 ± 0.184.59 ± 0.6493.99 ± 0.490.284593.97 ± 0.490.4759
", + "bbox": [ + 96, + 108, + 906, + 275 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the teacher, test loss is over $10^{3}$ times the training loss, allowing fingerprints to easily infer dataset association. In contrast, the student's test loss is typically under twice the training loss, falling within the variance across subsets and making it far less distinguishable. Thus, Escaping DOV significantly suppresses both watermarks and fingerprints.", + "bbox": [ + 89, + 282, + 482, + 376 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Comparison with SOTA Evasion Attacks", + "text_level": 1, + "bbox": [ + 89, + 386, + 439, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We compare with SOTA evasion attacks on CIFAR-10 across the seminal Badnets watermark and three most resistant DOVs from each category—Narcissus, Isotope, and Dataset Inference (see Table 1). Results for all DOV methods are provided in Appendix C.2. Evasion attacks include Fine-Pruning [36], stronger combination of fine-tuning and pruning widely used in DOV evaluations, along with SOTA backdoor defenses across four stages: data sanitization with Meta-Sift [73], robust training with Differential Privacy [4], backdoor unlearning with I-BAU [72], and input purification with ZIP [56]. We also compare with methods that use knowledge distillation to mitigate backdoors, such as NAD [31], BCU [47] and ABD [22], as well as IPRemoval [78], which aims to eliminate model watermarks. Results in Table 2 and Appendix C.2 show that Escaping DOV consistently surpasses all comparison methods in both generalization and evasion, and is the only approach that fully evades all DOV methods. For instance, only ABD and Escaping DOV reduce VSR below $40\\%$ on the clean-label and invisible Narcissus watermark; however, ABD fails to evade Isotope, while Escaping DOV achieves nearly $10\\%$ higher accuracy with superior evasion performance. We provide a rigorous analysis in Appendix D, explaining why SOTA evasion attacks, particularly those relying on distillation, are less effective than our Escaping DOV.", + "bbox": [ + 91, + 409, + 483, + 786 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablations and Discussions", + "text_level": 1, + "bbox": [ + 89, + 797, + 326, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Transfer Set Curation", + "text_level": 1, + "bbox": [ + 89, + 820, + 290, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We begin with a qualitative analysis of the transfer set. Figure 3a visualizes the feature space for CIFAR-10 samples as well as selected and unselected ones from the transfer set, on a clean ResNet-18. The selected samples are notably", + "bbox": [ + 89, + 839, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "closer to original CIFAR-10 samples. Figure 3b illustrates the feature space of a Badnets-marked model, where trigger samples form distinct clusters with almost no transfer set samples nearby. The exclusive and subtle verification behaviors are rarely activated by OOD samples, and the curation process effectively excludes suspicious samples.", + "bbox": [ + 511, + 282, + 906, + 375 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/30aa5cca0a72ec5148e512446ea08bc55c86c7d8996762fa937e449053f6e80a.jpg", + "image_caption": [ + "Figure 3. Sample Distribution in the Feature Space." + ], + "image_footnote": [], + "bbox": [ + 517, + 376, + 710, + 494 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/91d664e9df5cabf3b5cbb3ccee4961cf9a966c9c3c63fab22cd7e739105b89ef.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 712, + 376, + 903, + 494 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative analysis follows, using Optimal Transport Dataset Distance (OTDD) [2] to assess the distance from the target dataset to each proxy distribution, as shown in Table 3. The curated transfer set is significantly closer to the target distribution than random sampling. Further, we compare the curated transfer set with random and unrelative samples on CIFAR-10 with Narcissus. Table 4 shows that transfer set curation not only improves generalization but also suppresses DOV by selecting more reliable samples.", + "bbox": [ + 511, + 521, + 906, + 657 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d5e41254984145be5f7018af7fcb72db1497dc89d546c6f78c504b58e399afd4.jpg", + "table_caption": [ + "Table 3. Optimal Transport Distances Between Datasets." + ], + "table_footnote": [], + "table_body": "
DistanceselfLSVRC(curation)LSVRC(random)GTSRB [58]
CIFAR-100.7710.9371.1161.129
Tiny-Imagenet0.8550.9880.9931.066
", + "bbox": [ + 517, + 675, + 905, + 722 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d30284928ea4cdd9ae71e8bf3a508c89c346b577b43f72a2a664ba83b12603e9.jpg", + "table_caption": [ + "Table 4. Escaping DOV across different Transfer Sets." + ], + "table_footnote": [], + "table_body": "
Transfer SetGTSRB [58]LSVRC(random)LSVRC(curation)
ACC(↑)66.80 ± 0.4191.63 ± 0.4994.37 ± 0.18
VSR(↓)8.96 ± 1.637.87 ± 0.524.59 ± 0.64
", + "bbox": [ + 517, + 743, + 903, + 795 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Selective Knowledge Transfer", + "text_level": 1, + "bbox": [ + 511, + 806, + 767, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 4 illustrates the impact of Selective Knowledge Transfer (SKT) on the most resistant Narcissus watermark across distillation temperatures $\\tau$ . As $\\tau$ increases, omitting SKT causes a substantial increase in VSR, while SKT consistently suppresses verification behavior across all temper", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "1255", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "atures. Additionally, VSR escalates sharply at higher temperatures, while test accuracy remains nearly unchanged. We hypothesis that task and verification knowledge possess distinct \"boiling points\", naturally separating at lower temperatures. Thus, we adopt $\\tau = 1$ as a reliable default.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/faf7efc6e4773da8f48a83aacd5b94d43ca8e3191607e62a79c70b3a612db076.jpg", + "image_caption": [ + "Figure 4. Impact of SKT and Distillation Temperature." + ], + "image_footnote": [], + "bbox": [ + 93, + 169, + 480, + 271 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further compare the offline perturbation pool in SKT with direct UAT on the student. While UAT lowers VSR to $2.07\\%$ , it incurs a $3.54\\%$ accuracy reduction and more than triples computational cost. As SKT balances evasion and generalization, it serves as a lightweight approximation. Detailed time complexity is analysed in Appendix C.8.", + "bbox": [ + 89, + 292, + 483, + 385 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Escaping Adaptive Defenses", + "text_level": 1, + "bbox": [ + 89, + 397, + 333, + 412 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We assume a knowledgeable defender employing adaptive verification, specifically a semantic backdoor [60] and an anti-distillation backdoor [13]. For the semantic backdoor, we assume the defender knows the gallery set and selects 500 \"tench\" samples from LSVRC-2012, relabeling them as \"plane\" and inserting them into CIFAR-10 to test their classification probability as \"plane\" (VSR). Since the teacher memorizes this verification behavior, using it for sample selection allows triggers enter the transfer set. For instance, DFND [13], an entropy-based selection method, produces a student with VSR of $70.2\\%$ by including 254 \"tench\" samples in the transfer set. In contrast, our curated transfer set contains fewer than 10 \"tench\" samples, reducing VSR to only $17.4\\%$ . Even if the defender knows the gallery set, Transfer Set Curation still selects reliable samples. We also evaluate the entropy of teacher annotations: DFND's entropy is 2.90, while ours reaches 3.32, near the theoretical optimum $\\log_2(10)$ , indicating better class balance.", + "bbox": [ + 89, + 417, + 483, + 690 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6eae561c53ea74d1a26b27280048a4169ef13a3b0fbec2f50a0ad04fdeebf195.jpg", + "table_caption": [ + "Table 5. Escaping DOV against the Adaptive ADB Watermark." + ], + "table_footnote": [], + "table_body": "
μVanillaEvasionIndependent
ACCVSRACCVSRACCVSR
0.191.59%100.00%90.14%94.42%95.13%92.89%
190.90%100.00%89.23%78.48%95.13%71.74%
1091.26%12.88%88.66%14.21%95.13%10.01%
", + "bbox": [ + 94, + 709, + 483, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The anti-distillation backdoor (ADB) [13] is designed to transfer to student during distillation. As we employ distillation for knowledge transfer, we test ADB under a hypothetical scenario where the defender controls the teacher's training process to optimize the trigger. Table 5 reports the VSR across teacher, student, and unmarked models. A smaller $\\mu$ allows larger perturbations, turning", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the trigger into an adversarial perturbation that induces false positives on unmarked models. Despite the defender's unrealistic control over the teacher, students' VSR remains closer to that of an unmarked model. As $\\mu$ increases, ADB becomes ineffective. Thus, even under defender-controlled teacher training, Escaping DOV remains a potent threat.", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.4. Enhanced Evasion Tactics", + "text_level": 1, + "bbox": [ + 511, + 193, + 741, + 207 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Escaping DOV involves only post-training adjustments and operates orthogonally to pre-training, in-training and test-time strategies, all of which can collectively enhance evasion. As shown in Table 6, replacing data augmentation with mixup [75] during teacher training—an approach that mitigates backdoors and overfitting—further boosts evasion efficacy in conjunction with Escaping DOV on CIFAR-10.", + "bbox": [ + 511, + 212, + 905, + 319 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/6ae04f2c307e49353b6feff8ac20bb1e2fbc1348f46ec23f71c9717f2136ec89.jpg", + "table_caption": [ + "Table 6. Mixup Teacher Training for Escaping DOV." + ], + "table_footnote": [], + "table_body": "
DOVVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)
Narcissus94.04%72.47%93.03%3.66%
Non-poisoning DOVACCp-valueACC(↑)p-value(↑)
Isotope94.03%0.296492.71%0.7013
Dataset Inference94.60%0.093792.78%0.5196
", + "bbox": [ + 517, + 338, + 905, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.5. Case Study on Domain-Specific Datasets", + "text_level": 1, + "bbox": [ + 511, + 458, + 841, + 472 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Datasets in vertical industries often require stronger protection than natural images due to collection challenges and privacy concerns. We conduct a case study on RAFDB (facial emotion) [30] and OrganCMNIST (medical imaging) [71] in Table 7, with detailed results in Appendix C.3. Escaping DOV maintains generalization and efficacy despite unique distribution, while a larger domain-related gallery and VLM could further improve performance [34, 38].", + "bbox": [ + 511, + 478, + 905, + 599 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/82b0e4ffd1eb072bc77d53599eed70c6e74a8d0e260a262a5a1ebc67c6a242a8.jpg", + "table_caption": [ + "Table 7. Escaping DOV on Domain-specific Datasets." + ], + "table_footnote": [], + "table_body": "
SettingsVanillaEvasion
DOVDataSetACCVSRACC(↑)VSR(↓)
BadnetsRAFDB85.21 ± 0.3399.95 ± 0.0280.53 ± 0.422.19 ± 0.06
OrganC95.32 ± 0.0699.99 ± 0.0190.57 ± 0.700.83 ± 0.15
DOVDataSetACCp-valueACC(↑)p-value(↑)
IsotopeRAFDB82.46 ± 0.591.50e-0680.37 ± 0.131.05e-01
OrganC94.63 ± 0.192.65e-0388.96 ± 0.565.04e-01
", + "bbox": [ + 517, + 617, + 905, + 719 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 738, + 633, + 753 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we reveal that Data Ownership Verification, widely used to prevent unauthorized model training, is vulnerable to strong evasion attacks. We introduce Escaping DOV, which leverages transfer set curation and selective knowledge transfer to achieve superior generalization and universal evasion capabilities. We propose that Escaping DOV serve as a evaluation framework to help establish truly reliable DOV methods, with plans for extensive cross-modal experiments to explore broader societal benefits.", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "1256", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "The work was supported by the National Natural Science Foundation of China under Grant 62271307 and 61771310.", + "bbox": [ + 91, + 113, + 483, + 142 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 170, + 187, + 186 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdooring. In 27th USENIX security symposium (USENIX Security 18), pages 1615-1631, 2018. 2", + "[2] David Alvarez-Melis and Nicolo Fusi. Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems, 33:21428-21439, 2020. 7", + "[3] Jiawang Bai, Kuofeng Gao, Shaobo Min, Shu-Tao Xia, Zhifeng Li, and Wei Liu. Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24239-24250, 2024. 4", + "[4] Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. Automatic clipping: Differentially private deep learning made easier and stronger. Advances in Neural Information Processing Systems, 36, 2023. 3, 7", + "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6428-6437, 2021. 4", + "[6] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE transactions on image processing, 6(12):1673-1687, 1997. 1", + "[7] Francesco Croce and Matthias Hein. Adversarial robustness against multiple and single $l$ - $p$ -threat models via quick finetuning of robust classifiers. In International Conference on Machine Learning, pages 4436-4454. PMLR, 2022. 5", + "[8] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation, 6(2):182-197, 2002. 5", + "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 4, 6", + "[10] European Parliament and Council of the European Union. General data protection regulation (gdpr): Regulation (eu) 2016/679 on data protection and free movement, 2016. Available at https://gdpr-info.eu/.1", + "[11] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In International conference on machine learning, pages 1607-1616. PMLR, 2018. 5", + "[12] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal" + ], + "bbox": [ + 93, + 195, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "datasets. Advances in Neural Information Processing Systems, 36, 2024. 4", + "[13] Yunjie Ge, Qian Wang, Baolin Zheng, Xinlu Zhuang, Qi Li, Chao Shen, and Cong Wang. Anti-distillation backdoor attacks: Backdoors can really survive in knowledge distillation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 826-834, 2021. 8", + "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 5", + "[15] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1563-1580, 2022. 2", + "[16] T Gu, B Dolan-Gavitt, and S BadNets. Identifying vulnerabilities in the machine learning model supply chain. In Proceedings of the Neural Information Processing Symposium Workshop Mach. Learning Security (MLSec), pages 1-5, 2017. 2, 6", + "[17] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6", + "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6", + "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 5", + "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5", + "[21] Mathew Hogan, Yan Michalevsky, and Saba Eskandarian. Dbreach: Stealing from databases using compression side channels. In 2023 IEEE Symposium on Security and Privacy (SP), pages 182-198. IEEE, 2023. 1", + "[22] Junyuan Hong, Yi Zeng, Shuyang Yu, Lingjuan Lyu, Ruoxi Jia, and Jiayu Zhou. Revisiting data-free knowledge distillation with poisoned teachers. In International Conference on Machine Learning, pages 13199-13212. PMLR, 2023. 7", + "[23] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 1", + "[24] Zonghao Huang, Neil Zhenqiang Gong, and Michael K Reiter. A general framework for data-use auditing of ml models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 1300-1314, 2024. 2, 6", + "[25] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "1257", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 4", + "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4", + "[27] Xiao Jin, Pin-Yu Chen, Chia-Yi Hsu, Chia-Mu Yu, and Tianyi Chen. Cafe: Catastrophic data leakage in vertical federated learning. Advances in Neural Information Processing Systems, 34:994-1006, 2021. 1", + "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6", + "[29] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6", + "[30] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 8", + "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021. 3, 7", + "[32] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 6", + "[33] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 1, 2, 6", + "[34] Yudong Li, Xianxu Hou, Zheng Dezhi, Linlin Shen, and Zhe Zhao. Flip-80m: 80 million visual-linguistic pairs for facial language-image pre-training. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 58–67, 2024. 8", + "[35] Gaoyang Liu, Tianlong Xu, Xiaoqiang Ma, and Chen Wang. Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security, 17:1024-1037, 2022. 2, 3, 6", + "[36] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 3, 7", + "[37] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 1", + "[38] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 8" + ], + "bbox": [ + 91, + 90, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka. Mechanistic mode connectivity. In International Conference on Machine Learning, pages 22965-23004. PMLR, 2023. 6", + "[40] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5", + "[41] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 5", + "[42] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 2, 3, 6", + "[43] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 8884-8892, 2021. 5", + "[44] NBC News. Facial recognition's 'dirty little secret': Millions of online photos scraped without consent, 2019. Available at https://www.nbcnews.com/news/all/facial-recognition-s-dirty-little-secret-millionphotos-online-photos-scraped-n981921.1", + "[45] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Available at https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/.6", + "[46] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4954-4963, 2019. 5", + "[47] Lu Pang, Tianlong Sun, Huan Ling, and Changyou Chen. Backdoor cleansing with unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12218-12227, 2023. 7", + "[48] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 4", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4", + "[50] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 6", + "[51] Pedro Sandoval-Segura, Vasu Singla, Jonas Geiping, Micah Goldblum, and Tom Goldstein. What can we learn from unlearnable datasets? Advances in Neural Information Processing Systems, 36, 2023. 1" + ], + "bbox": [ + 516, + 90, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "1258", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1", + "[53] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 4", + "[54] Ali Shafahi, Mahyar Najibi, Zheng Xu, John Dickerson, Larry S Davis, and Tom Goldstein. Universal adversarial training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5636-5643, 2020. 5", + "[55] Shuo Shao, Yiming Li, Mengren Zheng, Zhiyang Hu, Yukun Chen, Boheng Li, Yu He, Junfeng Guo, Tianwei Zhang, Dacheng Tao, et al. Databench: Evaluating dataset auditing in deep learning from an adversarial perspective. arXiv preprint arXiv:2507.05622, 2025. 3", + "[56] Yucheng Shi, Mengnan Du, Xuansheng Wu, Zihan Guan, Jin Sun, and Ninghao Liu. Black-box backdoor defense via zero-shot image purification. Advances in Neural Information Processing Systems, 36:57336-57366, 2023. 3, 7", + "[57] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in neural information processing systems, 31, 2018. 5", + "[58] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011 international joint conference on neural networks*, pages 1453–1460. IEEE, 2011. 7", + "[59] Salvatore J Stolfo, Malek Ben Salem, and Angelos D Keromytis. Fog computing: Mitigating insider data theft attacks in the cloud. In 2012 IEEE symposium on security and privacy workshops, pages 125-128. IEEE, 2012. 1", + "[60] Bing Sun, Jun Sun, Wayne Koh, and Jie Shi. Neural network semantic backdoor detection and mitigation: A causality-based approach. In Proceedings of the 33rd USENIX Security Symposium. USENIX Association, San Francisco, CA, USA, 2024. 8", + "[61] Sebastian Szyller and N Asokan. Conflicting interactions among protection mechanisms for machine learning models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15179-15187, 2023. 3", + "[62] Sebastian Szyller, Rui Zhang, Jian Liu, and N. Asokan. On the robustness of dataset inference. Transactions on Machine Learning Research, 2023. 3", + "[63] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 1, 2", + "[64] Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. Advances in neural information processing systems, 32, 2019. 5", + "[65] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[66] Rolf Van Wegberg, Samaneh Tajalizadehkhoob, Kyle Soska, Ugur Akyazi, Carlos Hernandez Ganan, Bram Klievink, Nicolas Christin, and Michel Van Eeten. Plug and prey? measuring the commoditization of cybercrime via online anonymous markets. In 27th USENIX security symposium (USENIX security 18), pages 1009-1026, 2018. 1", + "[67] Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, and Oncel Tuzel. Mobile-clip: Fast image-text models through multi-modal reinforced training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15963–15974, 2024. 6", + "[68] Emily Wenger, Xiuyu Li, Ben Y Zhao, and Vitaly Shmatikov. Data isotopes for data provenance in dnns. Proceedings on Privacy Enhancing Technologies, 2024(1), 2024. 2, 3, 6", + "[69] Wikipedia contributors. Cambridge analytica — wikipedia, the free encyclopedia. https://en.wikipedia.org/wiki/Cambridge_Analytica, 2024.1", + "[70] Nuo Xu, Qi Liu, Tao Liu, Zihao Liu, Xiaochen Guo, and Wujie Wen. Stealing your data from compressed machine learning models. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. 1", + "[71] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 1, 8", + "[72] Yi Zeng, Si Chen, Won Park, Zhuoqing Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. In International Conference on Learning Representations, 2022. 3, 7", + "[73] Yi Zeng, Minzhou Pan, Himanshu Jahagirdar, Ming Jin, Lingjuan Lyu, and Ruoxi Jia. Meta-Sift: How to sift out a clean subset in the presence of data poisoning? In 32nd USENIX Security Symposium (USENIX Security 23), pages 1667-1684, Anaheim, CA, 2023. USENIX Association. 2, 3, 7", + "[74] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pages 771-785, 2023. 2, 6", + "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 8", + "[76] Hongyu Zhu, Sichu Liang, Wentao Hu, Li Fangqi, Ju Jia, and Shi-Lin Wang. Reliable model watermarking: Defending against theft without compromising on evasion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10124-10133, 2024. 2, 5, 6", + "[77] Hongyu Zhu, Sichu Liang, Wentao Hu, et al. Stealing knowledge from auditable datasets. In Proceedings of the European Conference on Artificial Intelligence (ECAI), 2025. 3", + "[78] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "1259", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "model inversion attack against deep neural network fingerprinting and watermarking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7837-7845, 2024. 7", + "[79] Zihang Zou, Boqing Gong, and Liqiang Wang. Anti-neuron watermarking: protecting personal data against unauthorized neural networks. In European Conference on Computer Vision, pages 449-465. Springer, 2022. 2, 3, 6", + "[80] Chaoshun Zuo, Zhiqiang Lin, and Yinqian Zhang. Why does your data leak? uncovering the data leakage in cloud from mobile apps. In 2019 IEEE Symposium on Security and Privacy (SP), pages 1296-1310. IEEE, 2019. 1" + ], + "bbox": [ + 91, + 90, + 480, + 247 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "1260", + "bbox": [ + 483, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_model.json b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0bc0e107912e6f67f979ce3e4cbd178539add874 --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_model.json @@ -0,0 +1,2534 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.237, + 0.131, + 0.761, + 0.154 + ], + "angle": 0, + "content": "Evading Data Provenance in Deep Neural Networks" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.179, + 0.918, + 0.234 + ], + "angle": 0, + "content": "Hongyu Zhu\\(^{1*}\\) Sichu Liang\\(^{2*}\\) Wenwen Wang\\(^{3}\\) Zhuomeng Zhang\\(^{1}\\) Fangqi Li\\(^{1}\\) Shi-Lin Wang\\(^{1\\dagger}\\) \n\\(^{1}\\)Shanghai Jiao Tong University \\(^{2}\\)Southeast University \\(^{3}\\)Carnegie Mellon University Code: https://github.com/dbsxfz/EscapingDOV" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.301, + 0.486, + 0.771 + ], + "angle": 0, + "content": "Modern over-parameterized deep models are highly data-dependent, with large scale general-purpose and domain-specific datasets serving as the bedrock for rapid advancements. However, many datasets are proprietary or contain sensitive information, making unrestricted model training problematic. In the open world where data thefts cannot be fully prevented, Dataset Ownership Verification (DOV) has emerged as a promising method to protect copyright by detecting unauthorized model training and tracing illicit activities. Due to its diversity and superior stealth, evading DOV is considered extremely challenging. However, this paper identifies that previous studies have relied on oversimplistic evasion attacks for evaluation, leading to a false sense of security. We introduce a unified evasion framework, in which a teacher model first learns from the copyright dataset and then transfers task-relevant yet identifier-independent domain knowledge to a surrogate student using an out-of-distribution (OOD) dataset as the intermediary. Leveraging Vision-Language Models and Large Language Models, we curate the most informative and reliable subsets from the OOD gallery set as the final transfer set, and propose selectively transferring task-oriented knowledge to achieve a better trade-off between generalization and evasion effectiveness. Experiments across diverse datasets covering eleven DOV methods demonstrate our approach simultaneously eliminates all copyright identifiers and significantly outperforms nine state-of-the-art evasion attacks in both generalization and effectiveness, with moderate computational overhead. As a proof of concept, we reveal key vulnerabilities in current DOV methods, highlighting the need for long-term development to enhance practicality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.802, + 0.222, + 0.818 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.829, + 0.485, + 0.875 + ], + "angle": 0, + "content": "The success of modern deep learning hinges heavily on abundant datasets, spanning from large-scale vision and multi-modal resources such as ImageNet [9] and LAION-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.908, + 0.468 + ], + "angle": 0, + "content": "5B [52] to datasets tailored for vertical industries like face recognition [37] and computational pathology [71]. Curated meticulously through extensive human effort for collection, cleaning, and labeling, these datasets are protected as intellectual property and are available solely through licensed access. Some are released for academic purposes with restrictions on commercial exploitation, while others are proprietary assets, strictly prohibiting third-party model training. Moreover, personal data—such as photos shared on social platforms—is increasingly at risk of being scraped without consent [44]. Consequently, unauthorized model training on restricted data poses an escalating threat to both intellectual property and privacy rights." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.471, + 0.909, + 0.622 + ], + "angle": 0, + "content": "Despite growing regulatory enforcement, like the General Data Protection Regulation (GDPR) [10], preventing unauthorized training remains a formidable challenge. Every stage of the data supply chain is vulnerable to attacks that funnel data into illicit markets [66], including insider threats [59], side-channel exploits [21], illegal mobile app access [80], and leakage from pretrained models [27, 70]. High-profile cases of data misuse in model training, such as the Cambridge Analytica scandal [69] and the Flickr Leakage [44], have intensified public scrutiny." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.626, + 0.909, + 0.838 + ], + "angle": 0, + "content": "Consequently, preventing unauthorized model training at its source—the dataset itself—is crucial. Unlearnable Examples [23] introduce perturbations to prevent models from learning meaningful features. However, beyond their susceptibility to adaptive attacks [51], they entirely obstruct legitimate dataset use. As an alternative, post hoc verification for data provenance, known as Dataset Ownership Verification (DOV), has garnered substantial attention. Inspired by watermarks used to track duplication, remixing, or exploitation of multi-modal content [6], the backdoor watermark embeds carefully crafted \"poison\" samples selected by the dataset owner. Models trained on watermarked data exhibit predefined behaviors in response to backdoor triggers, enabling model-agnostic owner attribution [33, 63]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To mitigate the impact of backdoor on authorized training, recent techniques introduce non-poisoning watermarks that adjust prediction confidence without inducing misclassification, employing hypothesis testing for ownership trac" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.348, + 0.901 + ], + "angle": 0, + "content": "*Equal contribution. † Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1249" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "ing [50, 68, 79]. Non-intrusive dataset fingerprint have also emerged, allowing verification via elevated confidence on memorized training samples [35, 42]. DOV is widely regarded as an effective, if not the only, solution for tracking unauthorized training without disrupting legitimate usage [17, 32], with rapid advancements underway." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.186, + 0.487, + 0.502 + ], + "angle": 0, + "content": "However, when deployed in the real world, the arms race between attackers and defenders forms an ongoing strategic game. Adversaries attempt to suppress verification behaviors to evade detection, while defenders continually strengthen verification mechanisms to reliably identify unauthorized training. Validating the robustness of DOV against evasion attempts is thus essential for practical deployment. Despite substantial progress in DOV, advances in evasion attacks remain underdeveloped, relying on simple regularization techniques and generic adaptations from poisoning defenses like fine-tuning [17, 32, 35, 79], which lack the versatility to counter diverse DOV techniques. Meanwhile, advanced DOV strategies, such as clean-label, invisible, or untargeted backdoor watermarks provide exceptional flexibility; non-poisoning watermarks and dataset fingerprints, which avoid misclassification or require no dataset modifications, remain nearly undetectable. It is widely agreed that designing a universal attack capable of evading all forms of verification is nearly impossible [42, 68]. This gap between the advancement of DOV and evasion techniques hinders reliable assessments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.485, + 0.626 + ], + "angle": 0, + "content": "In this paper we ask: is current DOV robust enough to preclude any evasion attempt? Our findings reveal that prior evasion attacks are too weak, leading to a false sense of security. Through analysis of verification behaviors, we identify shared characteristics: they are both exclusive and subtle, serving as unique markers that distinguish the protected dataset while remain incompatible with underlying semantic distribution of the main task." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Building on this insight, we propose the first universal evasion framework, Escaping DOV, which requires no additional clean in-distribution data or assumptions about the verification process. In Escaping DOV, a teacher model is initially trained directly on the copyright dataset to fully absorb domain knowledge. Then, an OOD transfer set, unrelated to the protected data, serves as an intermediary to extract task-oriented yet identifier-free knowledge to a surrogate student for deployment. Intuitively, exclusive and subtle identifiers, acting as side-channel signals orthogonal to the task distribution, make it improbable for OOD samples to trigger verification behaviors. Thus, only essential, identifier-invariant task knowledge is likely to transfer to the student model. To enhance Escaping DOV, we propose Transfer Set Curation, leveraging the unbiased knowledge embedded in large language models (LLMs) and vision-language models (VLMs) to retrieve the most informative and reliable samples from a general OOD gallery set. Ad" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.152 + ], + "angle": 0, + "content": "ditionally, we introduce Selective Knowledge Transfer to block suspicious verification behaviors, achieving better balance between generalization and evasion. The overall pipeline for Escaping DOV is shown in Figure 1." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.154, + 0.836, + 0.167 + ], + "angle": 0, + "content": "Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.515, + 0.169, + 0.905, + 0.229 + ], + "angle": 0, + "content": "1. We reveal that current DOV robustness evaluations are insufficiently rigorous. We introduce Escaping DOV, the first universal evasion framework that achieves simplicity, efficacy, and generalizability." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.23, + 0.905, + 0.273 + ], + "angle": 0, + "content": "2. We propose Transfer Set Curation and Selective Knowledge Transfer to achieve balance between generalization and evasion with moderate overhead." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.275, + 0.905, + 0.334 + ], + "angle": 0, + "content": "3. Experiments on diverse datasets and 11 DOV methods, alongside comparisons with 9 SOTA evasion methods, validate the efficacy of Escaping DOV, establishing it as a reliable benchmark for future DOV assessments." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.169, + 0.905, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.352, + 0.656, + 0.368 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.377, + 0.792, + 0.394 + ], + "angle": 0, + "content": "2.1. Dataset Ownership Verification" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.401, + 0.907, + 0.552 + ], + "angle": 0, + "content": "Backdoor attacks inject poisoned samples to associate predictions with a specific trigger [16], which induces misclassification whenever the trigger appears [15]. When controlled by copyright owners, backdoors can also act as identifiers for model ownership [1]. Inspired by this, the straightforward backdoor watermark embeds backdoor samples in the dataset. Models trained on this marked data produce a predefined label with high probability when evaluated on trigger samples [33]. The defender's knowledge of this secret prediction behavior serves as proof of ownership." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.553, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Early backdoors link triggers to target classes by manipulating the labels of poisoned samples. However, due to the obvious nature of poisoned labels, these samples are easily filtered through automatic data sanitization [73]. Clean-label backdoor watermark [63] utilizes advanced poisoning techniques [65] to enforce backdoor behavior without modifying labels. Later advancements further constrain modifications within an \\( L_{\\infty} \\) norm ball [74], making such clean-label, invisible backdoor watermarks highly stealthy [73]. Recently, untargeted backdoor watermark [32] induces nondeterministic misclassification without a fixed target class, enhancing resilience against post-training defenses such as trigger synthesis and output distribution analysis." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Advanced backdoor watermark provides a stealthy and resilient verification mechanism. However, it inevitably embeds harmful shortcuts [76], problematic when authorized training is essential [17]. Non-poisoning watermark subtly embeds watermark features into the dataset without inducing misclassification. Verification is performed by measuring loss differences between watermarked and clean samples via hypothesis testing. Various mechanisms have been developed to generate stealthy watermarks: Radioactive data [50] and ML Auditor [24] optimize" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1250" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.101, + 0.093, + 0.904, + 0.353 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.357, + 0.907, + 0.4 + ], + "angle": 0, + "content": "Figure 1. Overall Pipeline of Escaping DOV. (a) Transfer Set Curation (\\(\\S 3.2\\)): Samples from the OOD gallery set are clustered via a VLM, ranked by distribution proximity, and filtered through consensus ensemble to curate an optimal transfer set. (b) Selective Knowledge Transfer (\\(\\S 3.3\\)): Promoting invariance to teacher's worst perturbations, facilitating task-oriented yet identifier-free knowledge transfer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.409, + 0.483, + 0.47 + ], + "angle": 0, + "content": "\\(L_{\\infty}\\) -perturbations, Anti-Neuron Watermarking [79] applies color transformations in hue space, Domain Watermark [17] converts samples into a hardly-generalizable domain, while Isotope [68] blends external features into the dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.471, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Additionally, dataset fingerprint utilizes intrinsic features for verification without modifying the dataset. Dataset inference [42] analyzes the margins between training and external samples relative to the decision boundary, while MeFA [35] aggregates membership inference across multiple samples. Although dataset fingerprints are non-intrusive and support post hoc protection for published datasets, they are susceptible to high false positives [61, 62]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.446, + 0.622 + ], + "angle": 0, + "content": "2.2. Evasion Attacks against Data Provenance" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Despite rapid progress in DOV, current evasion strategies rely on basic regularization and backdoor defenses like finetuning [17, 32, 36]. As they barely evade any DOV methods, we present their results in Appendix C.1 and use SOTA backdoor defenses as baselines, which are categorized by stage of application: (1) Data sanitization removes suspicious samples pre-training [73]; (2) Robust learning mitigates poisoning effects during training [4, 31]; (3) Backdoor unlearning synthesizes triggers post-training to neutralize the model [72]; and (4) Input preprocessing detects or purifies triggers during testing [56]. Together, these defenses span the model lifecycle; yet advanced invisible, clean-label, or untargeted backdoor watermarks can bypass most of them [17, 32]. Further, non-poisoning watermarks and non-intrusive fingerprints render backdoor defenses ineffective, making universal evasion extremely challenging. This work is the first to establish a unified evasion perspective, opening new directions for robust evaluation of DOV." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.41, + 0.907, + 0.44 + ], + "angle": 0, + "content": "We discuss the contemporaneous works [77] and [55] that emerged at the time of acceptance in Appendix E." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.457, + 0.663, + 0.474 + ], + "angle": 0, + "content": "3. Escaping DOV" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.483, + 0.625, + 0.498 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.508, + 0.907, + 0.689 + ], + "angle": 0, + "content": "In \\(K\\)-class image classification, an adversary gains unauthorized access to the copyright dataset \\(\\mathcal{D}\\) to train a model \\(f_{\\theta}\\), which maps images \\(x \\in \\{0,1,\\ldots,255\\}^{C \\times W \\times H}\\) to predictions \\(y \\in \\mathbb{R}^K\\). For data provenance, the defender employs a DOV method \\(\\operatorname{Verif}(f_{\\theta},\\mathcal{D},\\mathcal{M})\\), inspecting the output of the suspicious model \\(f_{\\theta}\\) on specific queries. This process yields a verification metric indicating the likelihood that \\(f_{\\theta}\\) was trained on \\(\\mathcal{D}\\), such as the Attack Success Rate for backdoor watermarks or the confidence for rejecting the null hypothesis in non-poisoning DOV. Here, \\(\\mathcal{M}\\) represents external materials (e.g., watermark trigger patterns) controlled by the defender to uniquely identify the dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.909, + 0.901 + ], + "angle": 0, + "content": "DOV methods are flexible and diverse. Backdoor watermark achieves superior stealth through invisible, clean-label, or untargeted poison techniques. Non-poisoning watermarks avoid misclassification, while dataset fingerprints require no dataset modifications. Hence, direct sanitization or suppression of all potential verification behaviors becomes a formidable challenge; most evasion attempts that simply adapt poison defenses have consistently failed [17, 32, 42, 68, 79]. However, we identify a common trait among all DOV strategies: verification behaviors are both exclusive and subtle. On one hand, these behaviors must exclusively belong to the copyright dataset to prevent false attribution to models trained on external datasets. On the other hand, they must be subtle enough to remain concealed" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "1251" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.136 + ], + "angle": 0, + "content": "during data sanitization, ensuring they are rarely triggered by regular samples to preserve the generalization of models trained for authorized use." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.137, + 0.483, + 0.332 + ], + "angle": 0, + "content": "Accordingly, we propose a universal evasion strategy: first, a teacher model \\( f_{\\theta_t} \\) is trained on the copyright dataset \\( \\mathcal{D} \\) to absorb task-domain knowledge, inevitably marked by DOV. Next, an unrelated OOD transfer set \\( \\mathcal{T} \\) serves as the medium to extract task-oriented yet identifier-invariant knowledge into a surrogate student \\( f_{\\theta_s} \\). Intuitively, the exclusive and subtle watermark triggers are absent in the natural samples of \\( \\mathcal{T} \\), preventing the activation and transmission of verification behaviors to the surrogate student. For fingerprints based on exaggerated confidence, the student only indirectly inherits knowledge from the marked teacher via the unrelated OOD data, significantly reducing excessive memorization associated with \\( \\mathcal{D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.378 + ], + "angle": 0, + "content": "The overall objective of Escaping DOV can be formalized as simultaneously minimizing the generalization error on the task distribution and the verification metric:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.387, + 0.483, + 0.41 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\theta_ {s}} \\mathbb {E} _ {(x, y) \\in P (x, y)} [ \\mathcal {L} \\left(f _ {\\theta_ {s}} (x), y\\right) ] + \\alpha \\cdot \\operatorname {V e r i f} \\left(f _ {\\theta_ {s}}, \\mathcal {D}, \\mathcal {M}\\right) \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.483, + 0.66 + ], + "angle": 0, + "content": "Here, \\( f_{\\theta_s} \\) represents the surrogate student ultimately deployed, \\( P(x,y) \\) the underlying distribution of samples \\( x,y \\in \\mathcal{D} \\), and \\( \\mathcal{L} \\) the loss function measuring prediction discrepancies. The tuning factor \\( \\alpha \\) balances the pursuit of generalization and evasion for \\( f_{\\theta_s} \\). However, \\( P(x,y) \\) and the process of \\( \\mathrm{Verif}(f_{\\theta_s},\\mathcal{D},\\mathcal{M}) \\) are agnostic to the adversary and intractable during optimization. Therefore, we propose two modules to better achieve and balance these objectives: (1) For generalization, we curate the most informative and reliable transfer set from a large-scale OOD gallery set to support high-performance and identifier-invariant knowledge extraction; (2) For evasion, to prevent the transmission of suspicious verification knowledge via dark knowledge in soft labels, we propose selective knowledge transfer to filter out such knowledge and better trade off between the two objectives. Next, we elaborate on these two components." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.296, + 0.682 + ], + "angle": 0, + "content": "3.2. Transfer Set Curation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Our objective is to curate an optimal subset \\(\\mathcal{T}\\) from an OOD gallery set \\(\\mathcal{G}\\) (e.g., ImageNet [9] or DataComp [12]), selecting samples that are both informative and reliable to facilitate effective knowledge transfer. While this task may resemble classical core-set selection [5, 53], it presents significantly greater challenge: (1) Our algorithm operates in a fully OOD context, where \\(\\mathcal{G}\\) and \\(\\mathcal{D}\\) do not necessarily overlap in classes, and their distributions could vary substantially; (2) \\(\\mathcal{D}\\) could include identifiers manipulated by the defender (e.g., trigger samples), and the teacher \\(f_{\\theta_t}\\) trained directly on \\(\\mathcal{D}\\) is already marked. Consequently, selection based on the unreliable \\(\\mathcal{D}\\) and \\(f_{\\theta_t}\\) inadvertently introduce samples from \\(\\mathcal{G}\\) that activate verification behaviors, risking the induction of these behaviors in the surrogate student \\(f_{\\theta_s}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "In this paper, we leverage a vision-language model (VLM) with unbiased pretrained knowledge and robust zero-shot capabilities (e.g., CLIP [49]) to facilitate reliable selection alongside the marked teacher \\( f_{\\theta_t} \\). CLIP aligns visual and textual features in a shared embedding space, where the image encoder \\( \\phi_I(\\cdot) \\) maps an image \\( x \\) into \\( \\phi_I(x) \\), and text templates with category names (e.g., \"a photo of {class name}\") \\( c_1, \\ldots, c_K \\in \\mathcal{D} \\) are mapped by the text encoder \\( \\phi_t(\\cdot) \\) to \\( \\phi_t(c_1), \\ldots, \\phi_t(c_K) \\). By calculating the cosine similarity between \\( \\phi_I(x) \\) and \\( \\phi_t(c_i) \\), each sample \\( x \\in \\mathcal{G} \\) can be uniquely assigned to a category \\( t \\) in \\( \\mathcal{D} \\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.258, + 0.905, + 0.455 + ], + "angle": 0, + "content": "However, class-name-only text templates lack foreground semantics [48], limiting the ability to distinguish target categories. While few-shot tuning allows the VLM to rapidly adapt to tasks within \\(\\mathcal{D}\\) [26], it risks being compromised by verification behaviors embedded in \\(\\mathcal{D}\\) [3]. Inspired by advances in zero-shot prompt learning [48], we adjust the VLM with unbiased category semantics from a large language model (LLM). The LLM generates a description set \\(\\mathrm{Desc}_{c_i}\\) for each category \\(c_{i}\\) in \\(\\mathcal{D}\\), averaging descriptions in the feature space to create representation prototypes for each category, enhancing the zero-shot capabilities of the VLM. Consequently, each \\(x\\in \\mathcal{G}\\) is assigned to class \\(t\\) in a zero-shot fashion as follows in Equation 2:" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.46, + 0.905, + 0.5 + ], + "angle": 0, + "content": "\\[\nt = \\arg \\max _ {c _ {i}} \\frac {1}{\\left| \\operatorname {D e s c} _ {c _ {i}} \\right|} \\sum_ {d \\in \\operatorname {D e s c} _ {c _ {i}}} (\\sin (\\phi_ {I} (x), \\phi_ {I} (d))) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.507, + 0.895, + 0.523 + ], + "angle": 0, + "content": "where sim denotes cosine similarity shown in Equation 3:" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.529, + 0.905, + 0.563 + ], + "angle": 0, + "content": "\\[\n\\operatorname {s i m} \\left(\\phi_ {I} (x), \\phi_ {I} (d)\\right) = \\frac {\\phi_ {I} (x) \\cdot \\phi_ {I} (d)}{\\left| \\phi_ {I} (x) \\right| \\cdot \\left| \\phi_ {I} (d) \\right|} \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Although the VLM assigns all gallery set samples to categories in \\(\\mathcal{D}\\), it lacks a criterion for the most informative samples. To leverage the distribution from \\(\\mathcal{D}\\) while avoiding biases from potential verification behaviors, we project all samples in \\(\\mathcal{D}\\) into the feature space with \\(\\phi_I(\\cdot)\\) and calculate the density centroids \\(\\mathrm{Cent}_t\\) for each class \\(t\\in \\{1,\\dots ,K\\}\\). These centroids, acting as distribution digests, encapsulate the conditional distribution of \\(\\mathcal{D}\\). Samples closer to \\(\\mathrm{Cent}_t\\) in feature space are prioritized for selection, a robust and straightforward criterion that is theoretically resistant to manipulation by a small proportion of outlier samples [25]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We then establish a consensus-based ensemble for final transfer set curation: (1) The gallery set \\(\\mathcal{G}\\) is divided by the LLM-enhanced VLM into \\(K\\) bins corresponding to classes in \\(\\mathcal{D}\\). Within each bin \\(\\mathcal{G}_t\\), (2) samples are ordered by proximity to distribution digest \\(\\mathrm{Cent}_t\\); (3) Only samples classified by the teacher \\(f_{\\theta_t}\\) as \\(t\\) that are closest to \\(\\mathrm{Cent}_t\\) are selected, until the selected samples match the class count \\(|\\mathcal{D}_t|\\). Illustrated in Figure 1, this process ensures all predictions on the transfer set \\(\\mathcal{T}\\) by \\(f_{\\theta_t}\\) are consistent with VLM, minimizing the risk of triggering verification behaviors while closely approximating the distribution in \\(\\mathcal{D}\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1252" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "The Feature Bank. The curation process involves projecting all gallery set samples into the feature space, which is the most time-consuming step. However, when the gallery set is potentially reused, this projection only needs to be performed once. The resulting features and indexes are saved in a feature bank for efficient reuse." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.191, + 0.356, + 0.208 + ], + "angle": 0, + "content": "3.3. Selective Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.484, + 0.274 + ], + "angle": 0, + "content": "The transfer set \\(\\mathcal{T}\\) serves as an intermediary for extraction of task-oriented and identifier-invariant knowledge. This process employs the standard knowledge distillation framework [20] with the following optimization objective:" + }, + { + "type": "equation", + "bbox": [ + 0.134, + 0.282, + 0.484, + 0.315 + ], + "angle": 0, + "content": "\\[\n\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.322, + 0.484, + 0.457 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}\\) measures prediction discrepancies, such as the Kullback-Leibler divergence, modulated by the temperature \\(\\tau\\). This process effectively transfers knowledge despite potential misalignments between the transfer set \\(\\mathcal{T}\\) and the training distribution [46]. However, the soft labels \\(f_{\\theta_t}(x)\\) contain dark knowledge [11], which inadvertently transmit verification behaviors. Thus, we introduce Selective Knowledge Transfer (SKT) to filter out suspicious verification knowledge during distillation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.624 + ], + "angle": 0, + "content": "The exclusive and subtle verification behaviors reflect unique biases inherent to the marked dataset \\(\\mathcal{D}\\), distinguishing it within the task distribution. Preventing the student \\(f_{\\theta_s}\\) from overfitting to biases limits DOV's ability to infer connections between \\(f_{\\theta_s}\\) and \\(\\mathcal{D}\\) through trigger-specific outputs or excessive confidence. Drawing on insights from shortcut learning [14], we mitigate predictive shortcuts during the student's learning process, reducing reliance on spurious features. Since dataset biases typically induce universal shortcut behaviors, a straightforward solution is to apply universal adversarial training (UAT) on the student:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.632, + 0.483, + 0.669 + ], + "angle": 0, + "content": "\\[\n\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {T}} \\mathcal {L} \\left(f _ {\\theta_ {s}} (x + \\delta), y\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.677, + 0.484, + 0.782 + ], + "angle": 0, + "content": "UAT effectively suppresses atypical predictive behaviors in \\( f_{\\theta_s} \\) induced by perturbations \\( \\delta \\). However, the min-max problem involves a notoriously hard bi-level optimization, requiring iterative solutions within each iteration. Furthermore, the absence of ground-truth labels in the OOD transfer set \\( \\mathcal{T} \\) necessitates predictions from the marked teacher as labels, introducing significant calibration errors." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.784, + 0.484, + 0.859 + ], + "angle": 0, + "content": "Given the objective to mitigate biases specific to the copyrighted dataset \\(\\mathcal{D}\\), and with all verification behaviors encapsulated in the teacher model, we propose using only the teacher model \\(f_{\\theta_t}\\) to generate perturbations on \\(\\mathcal{D}\\) as a practical approximation for the inner max-problem:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.867, + 0.483, + 0.904 + ], + "angle": 0, + "content": "\\[\n\\delta^ {*} = \\arg \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {D}} \\mathcal {L} \\left(f _ {\\theta_ {t}} (x + \\delta), y\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.167 + ], + "angle": 0, + "content": "This strategy decouples the bi-level optimization, rendering it asynchronously solvable. By pre-generating a perturbation pool \\(\\{\\delta_i\\}_{i=1}^n\\) in an offline manner with the teacher \\(f_{\\theta_t}\\) and reusing it in the outer min-problem during distillation, computational overhead is significantly reduced." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.906, + 0.287 + ], + "angle": 0, + "content": "To solve Equation 6, the perturbation \\(\\delta\\) is typically constrained by an \\(L_{p}\\) norm to preserve core image semantics [40]. However, without prior knowledge of the specific form of \\(\\delta\\) associated with verification behaviors (e.g., watermark triggers), choosing an inappropriate norm undermine the efficacy of UAT in suppressing spurious features [64]. Thus, we seek to diversify the teacher-generated perturbations beyond a single norm constraint." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.906, + 0.453 + ], + "angle": 0, + "content": "We solve the max-problem using mini-batch stochastic gradient descent [54], applying \\( L_{2} \\) constraints at each batch update to preserve gradient directions. After each iteration, we project the perturbation onto the norm constraint \\( (L_0, L_2, \\text{or} L_\\infty) \\) that maximizes the loss. This steepest descent projection [41] adaptively seeks the optimal norm constraint, facilitating the generation of perturbations closely resembling those employed by DOV. Moreover, extreme projections onto \\( L_{0} \\) or \\( L_{\\infty} \\) norms approximate a convex hull spanning multiple norm constraints, broadening the range of potential perturbations [7]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.906, + 0.621 + ], + "angle": 0, + "content": "Ideally, perturbation should preserve core image semantics without being limited to specific norms [57]. Generative models can achieve this by producing unrestricted adversarial perturbations [76], though training such models on task-specific distributions is resource-intensive. Therefore, we propose a lightweight approach using image corruptions instead of generative models to produce unrestricted perturbations. Starting with the 15 corruptions defined in ImageNet-C [19], we employ a genetic algorithm [8, 43] to select the combination and sequence of corruptions that solves Equation 6, forming a corruption chain:" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.632, + 0.905, + 0.65 + ], + "angle": 0, + "content": "\\[\n\\max \\mathcal {L} \\left(f _ {\\theta_ {t}} \\left(\\operatorname {C o r r u p t} _ {1: N} (x)\\right), y\\right), \\quad x, y \\in \\mathcal {D} \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.66, + 0.905, + 0.707 + ], + "angle": 0, + "content": "The perturbations and corruption chain generated with the teacher \\( f_{\\theta_t} \\) on \\( \\mathcal{D} \\) are collectively referred to as \\( A(\\cdot) \\), leading to the final objective for Selective Knowledge Transfer:" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.717, + 0.905, + 0.785 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right) \\tag {8} \\\\ + \\beta \\cdot \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (A (x))}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "The second term promotes the student's invariance to the worst-case perturbations from the teacher while encouraging predictive divergence to reduce bias toward the teacher. The tuning factor \\(\\beta\\) balances generalization with evasion efficacy, applied as a sampling probability to selectively introduce perturbations or corruption chain to a subset of samples, adding minimal overhead." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1253" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.103, + 0.092, + 0.892, + 0.107 + ], + "angle": 0, + "content": "Table 1. Escaping DOV on CIFAR-10 and Tiny ImageNet. Red indicates detection by DOV, while green indicates successful evasion." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.109, + 0.903, + 0.326 + ], + "angle": 0, + "content": "
DOVCIFAR-10Tiny-Imagenet
VanillaEvasionVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)ACCVSRACC(↑)VSR(↓)
Badnets94.44 ± 0.74100.00 ± 0.0093.46 ± 0.201.36 ± 0.0769.05 ± 0.2598.98 ± 0.1166.48 ± 0.141.48 ± 0.11
UBW93.98 ± 0.2595.54 ± 0.8493.41 ± 0.191.74 ± 0.0666.37 ± 0.1275.56 ± 0.5065.34 ± 0.084.60 ± 0.11
Label-Consistent94.74 ± 0.5496.18 ± 2.7093.19 ± 0.143.74 ± 0.4069.07 ± 0.2434.99 ± 1.5366.46 ± 0.103.51 ± 0.22
Narcissus94.76 ± 0.5487.34 ± 1.8294.37 ± 0.184.59 ± 0.6468.99 ± 0.1457.49 ± 0.7866.39 ± 0.1812.33 ± 6.45
Non-Poisoning DOVACCp-valueACC(↑)p-value(↑)ACCp-valueACC(↑)p-value(↑)
Radioactive Data94.96 ± 0.473.03e-0394.07 ± 0.189.45e-0168.84 ± 0.305.33e-0366.64 ± 0.244.94e-01
ANW94.66 ± 0.311.37e-0993.91 ± 0.281.00e+0068.75 ± 0.305.57e-2866.44 ± 0.131.00e+00
Domain Watermark94.38 ± 0.531.67e-2293.90 ± 0.131.00e+0068.55 ± 0.393.91e-1666.38 ± 0.161.00e+00
Isotope94.75 ± 0.282.87e-0393.99 ± 0.492.84e-0168.97 ± 0.371.63e-0366.76 ± 0.221.66e-01
ML Auditor94.22 ± 0.177.47e-0593.45 ± 0.217.83e-0168.41 ± 0.281.55e-0665.35 ± 0.161.00e+00
Dataset Inference94.83 ± 0.471.87e-0393.97 ± 0.494.76e-0169.15 ± 0.349.27e-1466.06 ± 0.184.47e-01
MeFA94.83 ± 0.472.62e-1493.93 ± 0.551.00e+0069.15 ± 0.346.17e-2766.22 ± 0.209.99e-01
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.332, + 0.223, + 0.349 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.281, + 0.374 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.381, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Following established DOV methods [17, 33, 42, 79], we adopt ResNet-18 [18] for both teacher and students. For transfer set curation, we utilize MobileCLIP [67] and GPT-40 mini [45] as lightweight VLM and LLM, respectively." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.442, + 0.483, + 0.639 + ], + "angle": 0, + "content": "We assess backdoor watermarks with Verification Success Rate (VSR)-the probability of a trigger-induced misclassification to the target class. A VSR above \\(30\\%\\) confirms successful verification [76], indicating unauthorized training. Non-poisoning watermarks and fingerprints, collectively termed non-poisoning DOV, are evaluated via the p-value from a one-tailed T-test. The p-value represents the confidence level that the model was not trained on the copyright dataset; lower p-values indicate successful verification. The common threshold for p-values is set at 0.01 [42], equating to a \\(99\\%\\) confidence level of unauthorized training. Experiments were repeated 3 times, reporting mean \\(\\pm\\) std for ACC and VSR, and harmonic mean for p-values." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.394, + 0.668 + ], + "angle": 0, + "content": "4.2. Escaping Advanced DOV Methods" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We evaluate Escaping DOV on natural image datasets commonly used as DOV benchmarks, including CIFAR-10 [28] and the 200-class Tiny ImageNet [29]. We further consider six datasets with significant distribution shifts from the gallery set (e.g., facial recognition and medical diagnosis), with a summary of results in Section 4.4.5 and detailed findings in Appendix C.3. For the gallery set, we use LSVRC-2012 [9]. Given overlapping samples between Tiny ImageNet and LSVRC-2012, we create a challenging setting by removing all overlapping classes from LSVRC-2012 to produce a purely OOD gallery set. The targeted DOV approaches cover advanced algorithms across all categories: (1) backdoor watermarks, including BadNets [16], UBW [32], Label-Consistent [65], and Narcissus [74], encompassing clean-label, invisible, and untargeted poisoning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.905, + 0.394 + ], + "angle": 0, + "content": "techniques; (2) non-poisoning watermarks, including Radioactive Data [50], ANW [79], Domain Watermark [17], Isotope [68], and ML Auditor [24]; and (3) fingerprints, including Dataset Inference [42] and MeFA [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.607 + ], + "angle": 0, + "content": "Table 1 presents the generalization and verification metrics for the models before and after evasion attacks (the teacher and surrogate student in Escaping DOV). All DOV methods successfully identify the vanilla teacher trained directly on the copyright dataset. However, none meet their verification thresholds (VSR > 30% or p-value < 0.01) on the student produced by Escaping DOV, where VSR is near random guesses and p-values exceed 0.1 in most cases. Moreover, Escaping DOV incurs minimal impact on generalization: test accuracy on CIFAR-10 decreases by less than 1%, while on the challenging Tiny ImageNet set, with no overlapping gallery classes, accuracy drops by only 2-3%. Thus, Escaping DOV achieves a robust balance between universal evasion and preserved generalization." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.611, + 0.9, + 0.725 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.517, + 0.73, + 0.9, + 0.744 + ], + "angle": 0, + "content": "Figure 2. Loss Barrier between Teacher and Student Parameters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.906, + 0.901 + ], + "angle": 0, + "content": "We further observe that Escaping DOV encourages the student to diverge from the teacher's predictive behavior, forming its own mechanism. In Figure 2, we linearly interpolate between teacher and student parameters, computing training and test losses at each point. The results reveal two key findings: (1) A loss barrier appears during interpolation from teacher \\(\\theta_{t}\\) to student \\(\\theta_{s}\\). This lack of linear connectivity suggests that the surrogate student develops a distinct prediction mechanism, bypassing the spurious features central to the teacher's verification behavior [39]. (2)" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1254" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.112, + 0.092, + 0.884, + 0.107 + ], + "angle": 0, + "content": "Table 2. Comparison of Escaping DOV with Evasion Attacks. Red indicates detection by DOV, green indicates successful evasion." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.109, + 0.907, + 0.276 + ], + "angle": 0, + "content": "
MethodBadnetsNarcissusIsotopeDataset Inference
ACC(↑)VSR(↓)ACC(↑)VSR(↓)ACC(↑)p-value(↑)ACC(↑)p-value(↑)
Fine-pruning86.21 ± 0.1699.37 ± 0.2786.58 ± 0.1543.75 ± 14.4186.78 ± 0.120.053286.88 ± 0.480.0095
Meta-Sift88.68 ± 0.4599.70 ± 0.3288.64 ± 0.6249.40 ± 8.2688.56 ± 0.470.026488.07 ± 0.210.0963
Differential Privacy90.49 ± 1.2076.70 ± 4.6287.70 ± 0.6872.77 ± 3.9188.69 ± 2.610.067889.64 ± 4.860.0440
I-BAU89.09 ± 0.2514.80 ± 0.4688.76 ± 1.9472.74 ± 2.1787.07 ± 2.590.155289.93 ± 0.680.0355
ZIP82.44 ± 0.1079.88 ± 0.3983.51 ± 0.3753.21 ± 4.4782.97 ± 0.420.009883.21 ± 0.900.1050
NAD89.86 ± 0.373.11 ± 0.5191.04 ± 0.3675.41 ± 5.1291.77 ± 0.450.096192.12 ± 0.340.0014
BCU92.75 ± 0.051.56 ± 0.2592.33 ± 0.1161.72 ± 6.4392.77 ± 0.360.162293.09 ± 0.180.0773
ABD84.94 ± 3.047.07 ± 3.1585.42 ± 2.1022.10 ± 1.1586.29 ± 3.460.006684.87 ± 3.020.3696
IPRemoval83.96 ± 0.238.09 ± 0.5085.06 ± 0.1155.97 ± 2.1484.64 ± 0.060.191685.28 ± 0.050.1322
Escaping DOV (Ours)93.46 ± 0.201.36 ± 0.0794.37 ± 0.184.59 ± 0.6493.99 ± 0.490.284593.97 ± 0.490.4759
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.284, + 0.483, + 0.377 + ], + "angle": 0, + "content": "For the teacher, test loss is over \\(10^{3}\\) times the training loss, allowing fingerprints to easily infer dataset association. In contrast, the student's test loss is typically under twice the training loss, falling within the variance across subsets and making it far less distinguishable. Thus, Escaping DOV significantly suppresses both watermarks and fingerprints." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.387, + 0.441, + 0.403 + ], + "angle": 0, + "content": "4.3. Comparison with SOTA Evasion Attacks" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.41, + 0.485, + 0.787 + ], + "angle": 0, + "content": "We compare with SOTA evasion attacks on CIFAR-10 across the seminal Badnets watermark and three most resistant DOVs from each category—Narcissus, Isotope, and Dataset Inference (see Table 1). Results for all DOV methods are provided in Appendix C.2. Evasion attacks include Fine-Pruning [36], stronger combination of fine-tuning and pruning widely used in DOV evaluations, along with SOTA backdoor defenses across four stages: data sanitization with Meta-Sift [73], robust training with Differential Privacy [4], backdoor unlearning with I-BAU [72], and input purification with ZIP [56]. We also compare with methods that use knowledge distillation to mitigate backdoors, such as NAD [31], BCU [47] and ABD [22], as well as IPRemoval [78], which aims to eliminate model watermarks. Results in Table 2 and Appendix C.2 show that Escaping DOV consistently surpasses all comparison methods in both generalization and evasion, and is the only approach that fully evades all DOV methods. For instance, only ABD and Escaping DOV reduce VSR below \\(40\\%\\) on the clean-label and invisible Narcissus watermark; however, ABD fails to evade Isotope, while Escaping DOV achieves nearly \\(10\\%\\) higher accuracy with superior evasion performance. We provide a rigorous analysis in Appendix D, explaining why SOTA evasion attacks, particularly those relying on distillation, are less effective than our Escaping DOV." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.799, + 0.328, + 0.813 + ], + "angle": 0, + "content": "4.4. Ablations and Discussions" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.821, + 0.292, + 0.835 + ], + "angle": 0, + "content": "4.4.1. Transfer Set Curation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We begin with a qualitative analysis of the transfer set. Figure 3a visualizes the feature space for CIFAR-10 samples as well as selected and unselected ones from the transfer set, on a clean ResNet-18. The selected samples are notably" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.375 + ], + "angle": 0, + "content": "closer to original CIFAR-10 samples. Figure 3b illustrates the feature space of a Badnets-marked model, where trigger samples form distinct clusters with almost no transfer set samples nearby. The exclusive and subtle verification behaviors are rarely activated by OOD samples, and the curation process effectively excludes suspicious samples." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.377, + 0.712, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.713, + 0.377, + 0.904, + 0.496 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.501, + 0.863, + 0.516 + ], + "angle": 0, + "content": "Figure 3. Sample Distribution in the Feature Space." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Quantitative analysis follows, using Optimal Transport Dataset Distance (OTDD) [2] to assess the distance from the target dataset to each proxy distribution, as shown in Table 3. The curated transfer set is significantly closer to the target distribution than random sampling. Further, we compare the curated transfer set with random and unrelative samples on CIFAR-10 with Narcissus. Table 4 shows that transfer set curation not only improves generalization but also suppresses DOV by selecting more reliable samples." + }, + { + "type": "table_caption", + "bbox": [ + 0.541, + 0.661, + 0.878, + 0.675 + ], + "angle": 0, + "content": "Table 3. Optimal Transport Distances Between Datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.676, + 0.906, + 0.723 + ], + "angle": 0, + "content": "
DistanceselfLSVRC(curation)LSVRC(random)GTSRB [58]
CIFAR-100.7710.9371.1161.129
Tiny-Imagenet0.8550.9880.9931.066
" + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.728, + 0.871, + 0.742 + ], + "angle": 0, + "content": "Table 4. Escaping DOV across different Transfer Sets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.744, + 0.905, + 0.796 + ], + "angle": 0, + "content": "
Transfer SetGTSRB [58]LSVRC(random)LSVRC(curation)
ACC(↑)66.80 ± 0.4191.63 ± 0.4994.37 ± 0.18
VSR(↓)8.96 ± 1.637.87 ± 0.524.59 ± 0.64
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.807, + 0.768, + 0.822 + ], + "angle": 0, + "content": "4.4.2. Selective Knowledge Transfer" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Figure 4 illustrates the impact of Selective Knowledge Transfer (SKT) on the most resistant Narcissus watermark across distillation temperatures \\(\\tau\\). As \\(\\tau\\) increases, omitting SKT causes a substantial increase in VSR, while SKT consistently suppresses verification behavior across all temper" + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1255" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "atures. Additionally, VSR escalates sharply at higher temperatures, while test accuracy remains nearly unchanged. We hypothesis that task and verification knowledge possess distinct \"boiling points\", naturally separating at lower temperatures. Thus, we adopt \\(\\tau = 1\\) as a reliable default." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.17, + 0.482, + 0.272 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.124, + 0.274, + 0.45, + 0.287 + ], + "angle": 0, + "content": "Figure 4. Impact of SKT and Distillation Temperature." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.386 + ], + "angle": 0, + "content": "We further compare the offline perturbation pool in SKT with direct UAT on the student. While UAT lowers VSR to \\(2.07\\%\\), it incurs a \\(3.54\\%\\) accuracy reduction and more than triples computational cost. As SKT balances evasion and generalization, it serves as a lightweight approximation. Detailed time complexity is analysed in Appendix C.8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.398, + 0.334, + 0.413 + ], + "angle": 0, + "content": "4.4.3. Escaping Adaptive Defenses" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.418, + 0.484, + 0.691 + ], + "angle": 0, + "content": "We assume a knowledgeable defender employing adaptive verification, specifically a semantic backdoor [60] and an anti-distillation backdoor [13]. For the semantic backdoor, we assume the defender knows the gallery set and selects 500 \"tench\" samples from LSVRC-2012, relabeling them as \"plane\" and inserting them into CIFAR-10 to test their classification probability as \"plane\" (VSR). Since the teacher memorizes this verification behavior, using it for sample selection allows triggers enter the transfer set. For instance, DFND [13], an entropy-based selection method, produces a student with VSR of \\(70.2\\%\\) by including 254 \"tench\" samples in the transfer set. In contrast, our curated transfer set contains fewer than 10 \"tench\" samples, reducing VSR to only \\(17.4\\%\\). Even if the defender knows the gallery set, Transfer Set Curation still selects reliable samples. We also evaluate the entropy of teacher annotations: DFND's entropy is 2.90, while ours reaches 3.32, near the theoretical optimum \\(\\log_2(10)\\), indicating better class balance." + }, + { + "type": "table_caption", + "bbox": [ + 0.1, + 0.694, + 0.473, + 0.708 + ], + "angle": 0, + "content": "Table 5. Escaping DOV against the Adaptive ADB Watermark." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.71, + 0.484, + 0.793 + ], + "angle": 0, + "content": "
μVanillaEvasionIndependent
ACCVSRACCVSRACCVSR
0.191.59%100.00%90.14%94.42%95.13%92.89%
190.90%100.00%89.23%78.48%95.13%71.74%
1091.26%12.88%88.66%14.21%95.13%10.01%
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "The anti-distillation backdoor (ADB) [13] is designed to transfer to student during distillation. As we employ distillation for knowledge transfer, we test ADB under a hypothetical scenario where the defender controls the teacher's training process to optimize the trigger. Table 5 reports the VSR across teacher, student, and unmarked models. A smaller \\(\\mu\\) allows larger perturbations, turning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "the trigger into an adversarial perturbation that induces false positives on unmarked models. Despite the defender's unrealistic control over the teacher, students' VSR remains closer to that of an unmarked model. As \\(\\mu\\) increases, ADB becomes ineffective. Thus, even under defender-controlled teacher training, Escaping DOV remains a potent threat." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.194, + 0.742, + 0.208 + ], + "angle": 0, + "content": "4.4.4. Enhanced Evasion Tactics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.213, + 0.906, + 0.32 + ], + "angle": 0, + "content": "Escaping DOV involves only post-training adjustments and operates orthogonally to pre-training, in-training and test-time strategies, all of which can collectively enhance evasion. As shown in Table 6, replacing data augmentation with mixup [75] during teacher training—an approach that mitigates backdoors and overfitting—further boosts evasion efficacy in conjunction with Escaping DOV on CIFAR-10." + }, + { + "type": "table_caption", + "bbox": [ + 0.551, + 0.322, + 0.867, + 0.337 + ], + "angle": 0, + "content": "Table 6. Mixup Teacher Training for Escaping DOV." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.339, + 0.906, + 0.446 + ], + "angle": 0, + "content": "
DOVVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)
Narcissus94.04%72.47%93.03%3.66%
Non-poisoning DOVACCp-valueACC(↑)p-value(↑)
Isotope94.03%0.296492.71%0.7013
Dataset Inference94.60%0.093792.78%0.5196
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.459, + 0.842, + 0.473 + ], + "angle": 0, + "content": "4.4.5. Case Study on Domain-Specific Datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.906, + 0.6 + ], + "angle": 0, + "content": "Datasets in vertical industries often require stronger protection than natural images due to collection challenges and privacy concerns. We conduct a case study on RAFDB (facial emotion) [30] and OrganCMNIST (medical imaging) [71] in Table 7, with detailed results in Appendix C.3. Escaping DOV maintains generalization and efficacy despite unique distribution, while a larger domain-related gallery and VLM could further improve performance [34, 38]." + }, + { + "type": "table_caption", + "bbox": [ + 0.549, + 0.603, + 0.871, + 0.617 + ], + "angle": 0, + "content": "Table 7. Escaping DOV on Domain-specific Datasets." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.618, + 0.906, + 0.72 + ], + "angle": 0, + "content": "
SettingsVanillaEvasion
DOVDataSetACCVSRACC(↑)VSR(↓)
BadnetsRAFDB85.21 ± 0.3399.95 ± 0.0280.53 ± 0.422.19 ± 0.06
OrganC95.32 ± 0.0699.99 ± 0.0190.57 ± 0.700.83 ± 0.15
DOVDataSetACCp-valueACC(↑)p-value(↑)
IsotopeRAFDB82.46 ± 0.591.50e-0680.37 ± 0.131.05e-01
OrganC94.63 ± 0.192.65e-0388.96 ± 0.565.04e-01
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.739, + 0.634, + 0.754 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "In this paper, we reveal that Data Ownership Verification, widely used to prevent unauthorized model training, is vulnerable to strong evasion attacks. We introduce Escaping DOV, which leverages transfer set curation and selective knowledge transfer to achieve superior generalization and universal evasion capabilities. We propose that Escaping DOV serve as a evaluation framework to help establish truly reliable DOV methods, with plans for extensive cross-modal experiments to explore broader societal benefits." + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1256" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.114, + 0.484, + 0.143 + ], + "angle": 0, + "content": "The work was supported by the National Natural Science Foundation of China under Grant 62271307 and 61771310." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.171, + 0.188, + 0.187 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.196, + 0.484, + 0.265 + ], + "angle": 0, + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdooring. In 27th USENIX security symposium (USENIX Security 18), pages 1615-1631, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.266, + 0.484, + 0.307 + ], + "angle": 0, + "content": "[2] David Alvarez-Melis and Nicolo Fusi. Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems, 33:21428-21439, 2020. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.309, + 0.484, + 0.378 + ], + "angle": 0, + "content": "[3] Jiawang Bai, Kuofeng Gao, Shaobo Min, Shu-Tao Xia, Zhifeng Li, and Wei Liu. Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24239-24250, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.483, + 0.434 + ], + "angle": 0, + "content": "[4] Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. Automatic clipping: Differentially private deep learning made easier and stronger. Advances in Neural Information Processing Systems, 36, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6428-6437, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.506, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[6] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE transactions on image processing, 6(12):1673-1687, 1997. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.483, + 0.618 + ], + "angle": 0, + "content": "[7] Francesco Croce and Matthias Hein. Adversarial robustness against multiple and single \\( l \\)-\\( p \\)-threat models via quick finetuning of robust classifiers. In International Conference on Machine Learning, pages 4436-4454. PMLR, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.62, + 0.483, + 0.674 + ], + "angle": 0, + "content": "[8] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation, 6(2):182-197, 2002. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.676, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[10] European Parliament and Council of the European Union. General data protection regulation (gdpr): Regulation (eu) 2016/679 on data protection and free movement, 2016. Available at https://gdpr-info.eu/.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.844 + ], + "angle": 0, + "content": "[11] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In International conference on machine learning, pages 1607-1616. PMLR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.196, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "datasets. Advances in Neural Information Processing Systems, 36, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] Yunjie Ge, Qian Wang, Baolin Zheng, Xinlu Zhuang, Qi Li, Chao Shen, and Cong Wang. Anti-distillation backdoor attacks: Backdoors can really survive in knowledge distillation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 826-834, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[15] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1563-1580, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[16] T Gu, B Dolan-Gavitt, and S BadNets. Identifying vulnerabilities in the machine learning model supply chain. In Proceedings of the Neural Information Processing Symposium Workshop Mach. Learning Security (MLSec), pages 1-5, 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[17] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.525 + ], + "angle": 0, + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[21] Mathew Hogan, Yan Michalevsky, and Saba Eskandarian. Dbreach: Stealing from databases using compression side channels. In 2023 IEEE Symposium on Security and Privacy (SP), pages 182-198. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[22] Junyuan Hong, Yi Zeng, Shuyang Yu, Lingjuan Lyu, Ruoxi Jia, and Jiayu Zhou. Revisiting data-free knowledge distillation with poisoned teachers. In International Conference on Machine Learning, pages 13199-13212. PMLR, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.802 + ], + "angle": 0, + "content": "[23] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[24] Zonghao Huang, Neil Zhenqiang Gong, and Michael K Reiter. A general framework for data-use auditing of ml models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 1300-1314, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "1257" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.134 + ], + "angle": 0, + "content": "data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[27] Xiao Jin, Pin-Yu Chen, Chia-Yi Hsu, Chia-Mu Yu, and Tianyi Chen. Cafe: Catastrophic data leakage in vertical federated learning. Advances in Neural Information Processing Systems, 34:994-1006, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.278 + ], + "angle": 0, + "content": "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.279, + 0.482, + 0.307 + ], + "angle": 0, + "content": "[29] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.308, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[30] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.434 + ], + "angle": 0, + "content": "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.435, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[32] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.561 + ], + "angle": 0, + "content": "[33] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[34] Yudong Li, Xianxu Hou, Zheng Dezhi, Linlin Shen, and Zhe Zhao. Flip-80m: 80 million visual-linguistic pairs for facial language-image pre-training. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 58–67, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.702 + ], + "angle": 0, + "content": "[35] Gaoyang Liu, Tianlong Xu, Xiaoqiang Ma, and Chen Wang. Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security, 17:1024-1037, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[36] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.83 + ], + "angle": 0, + "content": "[37] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 8" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[39] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka. Mechanistic mode connectivity. In International Conference on Machine Learning, pages 22965-23004. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.217 + ], + "angle": 0, + "content": "[40] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.906, + 0.272 + ], + "angle": 0, + "content": "[41] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[42] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.384 + ], + "angle": 0, + "content": "[43] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 8884-8892, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.906, + 0.468 + ], + "angle": 0, + "content": "[44] NBC News. Facial recognition's 'dirty little secret': Millions of online photos scraped without consent, 2019. Available at https://www.nbcnews.com/news/all/facial-recognition-s-dirty-little-secret-millionphotos-online-photos-scraped-n981921.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.906, + 0.524 + ], + "angle": 0, + "content": "[45] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Available at https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.525, + 0.906, + 0.58 + ], + "angle": 0, + "content": "[46] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4954-4963, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.907, + 0.636 + ], + "angle": 0, + "content": "[47] Lu Pang, Tianlong Sun, Huan Ling, and Changyou Chen. Backdoor cleansing with unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12218-12227, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[48] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.844 + ], + "angle": 0, + "content": "[50] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.906, + 0.9 + ], + "angle": 0, + "content": "[51] Pedro Sandoval-Segura, Vasu Singla, Jonas Geiping, Micah Goldblum, and Tom Goldstein. What can we learn from unlearnable datasets? Advances in Neural Information Processing Systems, 36, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1258" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[53] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.276 + ], + "angle": 0, + "content": "[54] Ali Shafahi, Mahyar Najibi, Zheng Xu, John Dickerson, Larry S Davis, and Tom Goldstein. Universal adversarial training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5636-5643, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.278, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[55] Shuo Shao, Yiming Li, Mengren Zheng, Zhiyang Hu, Yukun Chen, Boheng Li, Yu He, Junfeng Guo, Tianwei Zhang, Dacheng Tao, et al. Databench: Evaluating dataset auditing in deep learning from an adversarial perspective. arXiv preprint arXiv:2507.05622, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[56] Yucheng Shi, Mengnan Du, Xuansheng Wu, Zihan Guan, Jin Sun, and Ninghao Liu. Black-box backdoor defense via zero-shot image purification. Advances in Neural Information Processing Systems, 36:57336-57366, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[57] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in neural information processing systems, 31, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.461, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[58] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011 international joint conference on neural networks*, pages 1453–1460. IEEE, 2011. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.588 + ], + "angle": 0, + "content": "[59] Salvatore J Stolfo, Malek Ben Salem, and Angelos D Keromytis. Fog computing: Mitigating insider data theft attacks in the cloud. In 2012 IEEE symposium on security and privacy workshops, pages 125-128. IEEE, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[60] Bing Sun, Jun Sun, Wayne Koh, and Jie Shi. Neural network semantic backdoor detection and mitigation: A causality-based approach. In Proceedings of the 33rd USENIX Security Symposium. USENIX Association, San Francisco, CA, USA, 2024. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.66, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[61] Sebastian Szyller and N Asokan. Conflicting interactions among protection mechanisms for machine learning models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15179-15187, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[62] Sebastian Szyller, Rui Zhang, Jian Liu, and N. Asokan. On the robustness of dataset inference. Transactions on Machine Learning Research, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.815 + ], + "angle": 0, + "content": "[63] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[64] Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. Advances in neural information processing systems, 32, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[65] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[66] Rolf Van Wegberg, Samaneh Tajalizadehkhoob, Kyle Soska, Ugur Akyazi, Carlos Hernandez Ganan, Bram Klievink, Nicolas Christin, and Michel Van Eeten. Plug and prey? measuring the commoditization of cybercrime via online anonymous markets. In 27th USENIX security symposium (USENIX security 18), pages 1009-1026, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[67] Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, and Oncel Tuzel. Mobile-clip: Fast image-text models through multi-modal reinforced training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15963–15974, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[68] Emily Wenger, Xiuyu Li, Ben Y Zhao, and Vitaly Shmatikov. Data isotopes for data provenance in dnns. Proceedings on Privacy Enhancing Technologies, 2024(1), 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[69] Wikipedia contributors. Cambridge analytica — wikipedia, the free encyclopedia. https://en.wikipedia.org/wiki/Cambridge_Analytica, 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[70] Nuo Xu, Qi Liu, Tao Liu, Zihao Liu, Xiaochen Guo, and Wujie Wen. Stealing your data from compressed machine learning models. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[71] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 1, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[72] Yi Zeng, Si Chen, Won Park, Zhuoqing Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. In International Conference on Learning Representations, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[73] Yi Zeng, Minzhou Pan, Himanshu Jahagirdar, Ming Jin, Lingjuan Lyu, and Ruoxi Jia. Meta-Sift: How to sift out a clean subset in the presence of data poisoning? In 32nd USENIX Security Symposium (USENIX Security 23), pages 1667-1684, Anaheim, CA, 2023. USENIX Association. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.701 + ], + "angle": 0, + "content": "[74] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pages 771-785, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.703, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.759, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[76] Hongyu Zhu, Sichu Liang, Wentao Hu, Li Fangqi, Ju Jia, and Shi-Lin Wang. Reliable model watermarking: Defending against theft without compromising on evasion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10124-10133, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.831, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[77] Hongyu Zhu, Sichu Liang, Wentao Hu, et al. Stealing knowledge from auditable datasets. In Proceedings of the European Conference on Artificial Intelligence (ECAI), 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[78] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1259" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "model inversion attack against deep neural network fingerprinting and watermarking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7837-7845, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[79] Zihang Zou, Boqing Gong, and Liqiang Wang. Anti-neuron watermarking: protecting personal data against unauthorized neural networks. In European Conference on Computer Vision, pages 449-465. Springer, 2022. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[80] Chaoshun Zuo, Zhiqiang Lin, and Yinqian Zhang. Why does your data leak? uncovering the data leakage in cloud from mobile apps. In 2019 IEEE Symposium on Security and Privacy (SP), pages 1296-1310. IEEE, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.484, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "1260" + } + ] +] \ No newline at end of file diff --git a/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_origin.pdf b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e9e1ae031d22715b61a0792b6db94a9654fa1f3c --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/22b6c5cf-c3b2-4e69-9394-11210dd9a3ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:46dabacad69dadae2fc49276eb68a8b794f7d67eef70caea743b7328f1f58445 +size 3704165 diff --git a/2025/Evading Data Provenance in Deep Neural Networks/full.md b/2025/Evading Data Provenance in Deep Neural Networks/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7b76ac41d59d0ca16945e81926aacbfe54b5a600 --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/full.md @@ -0,0 +1,343 @@ +# Evading Data Provenance in Deep Neural Networks + +Hongyu Zhu $^{1*}$ Sichu Liang $^{2*}$ Wenwen Wang $^{3}$ Zhuomeng Zhang $^{1}$ Fangqi Li $^{1}$ Shi-Lin Wang $^{1\dagger}$ $^{1}$ Shanghai Jiao Tong University $^{2}$ Southeast University $^{3}$ Carnegie Mellon University Code: https://github.com/dbsxfz/EscapingDOV + +# Abstract + +Modern over-parameterized deep models are highly data-dependent, with large scale general-purpose and domain-specific datasets serving as the bedrock for rapid advancements. However, many datasets are proprietary or contain sensitive information, making unrestricted model training problematic. In the open world where data thefts cannot be fully prevented, Dataset Ownership Verification (DOV) has emerged as a promising method to protect copyright by detecting unauthorized model training and tracing illicit activities. Due to its diversity and superior stealth, evading DOV is considered extremely challenging. However, this paper identifies that previous studies have relied on oversimplistic evasion attacks for evaluation, leading to a false sense of security. We introduce a unified evasion framework, in which a teacher model first learns from the copyright dataset and then transfers task-relevant yet identifier-independent domain knowledge to a surrogate student using an out-of-distribution (OOD) dataset as the intermediary. Leveraging Vision-Language Models and Large Language Models, we curate the most informative and reliable subsets from the OOD gallery set as the final transfer set, and propose selectively transferring task-oriented knowledge to achieve a better trade-off between generalization and evasion effectiveness. Experiments across diverse datasets covering eleven DOV methods demonstrate our approach simultaneously eliminates all copyright identifiers and significantly outperforms nine state-of-the-art evasion attacks in both generalization and effectiveness, with moderate computational overhead. As a proof of concept, we reveal key vulnerabilities in current DOV methods, highlighting the need for long-term development to enhance practicality. + +# 1. Introduction + +The success of modern deep learning hinges heavily on abundant datasets, spanning from large-scale vision and multi-modal resources such as ImageNet [9] and LAION- + +5B [52] to datasets tailored for vertical industries like face recognition [37] and computational pathology [71]. Curated meticulously through extensive human effort for collection, cleaning, and labeling, these datasets are protected as intellectual property and are available solely through licensed access. Some are released for academic purposes with restrictions on commercial exploitation, while others are proprietary assets, strictly prohibiting third-party model training. Moreover, personal data—such as photos shared on social platforms—is increasingly at risk of being scraped without consent [44]. Consequently, unauthorized model training on restricted data poses an escalating threat to both intellectual property and privacy rights. + +Despite growing regulatory enforcement, like the General Data Protection Regulation (GDPR) [10], preventing unauthorized training remains a formidable challenge. Every stage of the data supply chain is vulnerable to attacks that funnel data into illicit markets [66], including insider threats [59], side-channel exploits [21], illegal mobile app access [80], and leakage from pretrained models [27, 70]. High-profile cases of data misuse in model training, such as the Cambridge Analytica scandal [69] and the Flickr Leakage [44], have intensified public scrutiny. + +Consequently, preventing unauthorized model training at its source—the dataset itself—is crucial. Unlearnable Examples [23] introduce perturbations to prevent models from learning meaningful features. However, beyond their susceptibility to adaptive attacks [51], they entirely obstruct legitimate dataset use. As an alternative, post hoc verification for data provenance, known as Dataset Ownership Verification (DOV), has garnered substantial attention. Inspired by watermarks used to track duplication, remixing, or exploitation of multi-modal content [6], the backdoor watermark embeds carefully crafted "poison" samples selected by the dataset owner. Models trained on watermarked data exhibit predefined behaviors in response to backdoor triggers, enabling model-agnostic owner attribution [33, 63]. + +To mitigate the impact of backdoor on authorized training, recent techniques introduce non-poisoning watermarks that adjust prediction confidence without inducing misclassification, employing hypothesis testing for ownership trac + +ing [50, 68, 79]. Non-intrusive dataset fingerprint have also emerged, allowing verification via elevated confidence on memorized training samples [35, 42]. DOV is widely regarded as an effective, if not the only, solution for tracking unauthorized training without disrupting legitimate usage [17, 32], with rapid advancements underway. + +However, when deployed in the real world, the arms race between attackers and defenders forms an ongoing strategic game. Adversaries attempt to suppress verification behaviors to evade detection, while defenders continually strengthen verification mechanisms to reliably identify unauthorized training. Validating the robustness of DOV against evasion attempts is thus essential for practical deployment. Despite substantial progress in DOV, advances in evasion attacks remain underdeveloped, relying on simple regularization techniques and generic adaptations from poisoning defenses like fine-tuning [17, 32, 35, 79], which lack the versatility to counter diverse DOV techniques. Meanwhile, advanced DOV strategies, such as clean-label, invisible, or untargeted backdoor watermarks provide exceptional flexibility; non-poisoning watermarks and dataset fingerprints, which avoid misclassification or require no dataset modifications, remain nearly undetectable. It is widely agreed that designing a universal attack capable of evading all forms of verification is nearly impossible [42, 68]. This gap between the advancement of DOV and evasion techniques hinders reliable assessments. + +In this paper we ask: is current DOV robust enough to preclude any evasion attempt? Our findings reveal that prior evasion attacks are too weak, leading to a false sense of security. Through analysis of verification behaviors, we identify shared characteristics: they are both exclusive and subtle, serving as unique markers that distinguish the protected dataset while remain incompatible with underlying semantic distribution of the main task. + +Building on this insight, we propose the first universal evasion framework, Escaping DOV, which requires no additional clean in-distribution data or assumptions about the verification process. In Escaping DOV, a teacher model is initially trained directly on the copyright dataset to fully absorb domain knowledge. Then, an OOD transfer set, unrelated to the protected data, serves as an intermediary to extract task-oriented yet identifier-free knowledge to a surrogate student for deployment. Intuitively, exclusive and subtle identifiers, acting as side-channel signals orthogonal to the task distribution, make it improbable for OOD samples to trigger verification behaviors. Thus, only essential, identifier-invariant task knowledge is likely to transfer to the student model. To enhance Escaping DOV, we propose Transfer Set Curation, leveraging the unbiased knowledge embedded in large language models (LLMs) and vision-language models (VLMs) to retrieve the most informative and reliable samples from a general OOD gallery set. Ad + +ditionally, we introduce Selective Knowledge Transfer to block suspicious verification behaviors, achieving better balance between generalization and evasion. The overall pipeline for Escaping DOV is shown in Figure 1. + +Our contributions are summarized as follows: + +1. We reveal that current DOV robustness evaluations are insufficiently rigorous. We introduce Escaping DOV, the first universal evasion framework that achieves simplicity, efficacy, and generalizability. +2. We propose Transfer Set Curation and Selective Knowledge Transfer to achieve balance between generalization and evasion with moderate overhead. +3. Experiments on diverse datasets and 11 DOV methods, alongside comparisons with 9 SOTA evasion methods, validate the efficacy of Escaping DOV, establishing it as a reliable benchmark for future DOV assessments. + +# 2. Related Work + +# 2.1. Dataset Ownership Verification + +Backdoor attacks inject poisoned samples to associate predictions with a specific trigger [16], which induces misclassification whenever the trigger appears [15]. When controlled by copyright owners, backdoors can also act as identifiers for model ownership [1]. Inspired by this, the straightforward backdoor watermark embeds backdoor samples in the dataset. Models trained on this marked data produce a predefined label with high probability when evaluated on trigger samples [33]. The defender's knowledge of this secret prediction behavior serves as proof of ownership. + +Early backdoors link triggers to target classes by manipulating the labels of poisoned samples. However, due to the obvious nature of poisoned labels, these samples are easily filtered through automatic data sanitization [73]. Clean-label backdoor watermark [63] utilizes advanced poisoning techniques [65] to enforce backdoor behavior without modifying labels. Later advancements further constrain modifications within an $L_{\infty}$ norm ball [74], making such clean-label, invisible backdoor watermarks highly stealthy [73]. Recently, untargeted backdoor watermark [32] induces nondeterministic misclassification without a fixed target class, enhancing resilience against post-training defenses such as trigger synthesis and output distribution analysis. + +Advanced backdoor watermark provides a stealthy and resilient verification mechanism. However, it inevitably embeds harmful shortcuts [76], problematic when authorized training is essential [17]. Non-poisoning watermark subtly embeds watermark features into the dataset without inducing misclassification. Verification is performed by measuring loss differences between watermarked and clean samples via hypothesis testing. Various mechanisms have been developed to generate stealthy watermarks: Radioactive data [50] and ML Auditor [24] optimize + +![](images/3ce822d7d903579f10901ce1d1ad52061dc5d06557d686f6c473c2304b50b515.jpg) +Figure 1. Overall Pipeline of Escaping DOV. (a) Transfer Set Curation ( $\S 3.2$ ): Samples from the OOD gallery set are clustered via a VLM, ranked by distribution proximity, and filtered through consensus ensemble to curate an optimal transfer set. (b) Selective Knowledge Transfer ( $\S 3.3$ ): Promoting invariance to teacher's worst perturbations, facilitating task-oriented yet identifier-free knowledge transfer. + +$L_{\infty}$ -perturbations, Anti-Neuron Watermarking [79] applies color transformations in hue space, Domain Watermark [17] converts samples into a hardly-generalizable domain, while Isotope [68] blends external features into the dataset. + +Additionally, dataset fingerprint utilizes intrinsic features for verification without modifying the dataset. Dataset inference [42] analyzes the margins between training and external samples relative to the decision boundary, while MeFA [35] aggregates membership inference across multiple samples. Although dataset fingerprints are non-intrusive and support post hoc protection for published datasets, they are susceptible to high false positives [61, 62]. + +# 2.2. Evasion Attacks against Data Provenance + +Despite rapid progress in DOV, current evasion strategies rely on basic regularization and backdoor defenses like finetuning [17, 32, 36]. As they barely evade any DOV methods, we present their results in Appendix C.1 and use SOTA backdoor defenses as baselines, which are categorized by stage of application: (1) Data sanitization removes suspicious samples pre-training [73]; (2) Robust learning mitigates poisoning effects during training [4, 31]; (3) Backdoor unlearning synthesizes triggers post-training to neutralize the model [72]; and (4) Input preprocessing detects or purifies triggers during testing [56]. Together, these defenses span the model lifecycle; yet advanced invisible, clean-label, or untargeted backdoor watermarks can bypass most of them [17, 32]. Further, non-poisoning watermarks and non-intrusive fingerprints render backdoor defenses ineffective, making universal evasion extremely challenging. This work is the first to establish a unified evasion perspective, opening new directions for robust evaluation of DOV. + +We discuss the contemporaneous works [77] and [55] that emerged at the time of acceptance in Appendix E. + +# 3. Escaping DOV + +# 3.1. Overview + +In $K$ -class image classification, an adversary gains unauthorized access to the copyright dataset $\mathcal{D}$ to train a model $f_{\theta}$ , which maps images $x \in \{0,1,\ldots,255\}^{C \times W \times H}$ to predictions $y \in \mathbb{R}^K$ . For data provenance, the defender employs a DOV method $\operatorname{Verif}(f_{\theta},\mathcal{D},\mathcal{M})$ , inspecting the output of the suspicious model $f_{\theta}$ on specific queries. This process yields a verification metric indicating the likelihood that $f_{\theta}$ was trained on $\mathcal{D}$ , such as the Attack Success Rate for backdoor watermarks or the confidence for rejecting the null hypothesis in non-poisoning DOV. Here, $\mathcal{M}$ represents external materials (e.g., watermark trigger patterns) controlled by the defender to uniquely identify the dataset. + +DOV methods are flexible and diverse. Backdoor watermark achieves superior stealth through invisible, clean-label, or untargeted poison techniques. Non-poisoning watermarks avoid misclassification, while dataset fingerprints require no dataset modifications. Hence, direct sanitization or suppression of all potential verification behaviors becomes a formidable challenge; most evasion attempts that simply adapt poison defenses have consistently failed [17, 32, 42, 68, 79]. However, we identify a common trait among all DOV strategies: verification behaviors are both exclusive and subtle. On one hand, these behaviors must exclusively belong to the copyright dataset to prevent false attribution to models trained on external datasets. On the other hand, they must be subtle enough to remain concealed + +during data sanitization, ensuring they are rarely triggered by regular samples to preserve the generalization of models trained for authorized use. + +Accordingly, we propose a universal evasion strategy: first, a teacher model $f_{\theta_t}$ is trained on the copyright dataset $\mathcal{D}$ to absorb task-domain knowledge, inevitably marked by DOV. Next, an unrelated OOD transfer set $\mathcal{T}$ serves as the medium to extract task-oriented yet identifier-invariant knowledge into a surrogate student $f_{\theta_s}$ . Intuitively, the exclusive and subtle watermark triggers are absent in the natural samples of $\mathcal{T}$ , preventing the activation and transmission of verification behaviors to the surrogate student. For fingerprints based on exaggerated confidence, the student only indirectly inherits knowledge from the marked teacher via the unrelated OOD data, significantly reducing excessive memorization associated with $\mathcal{D}$ . + +The overall objective of Escaping DOV can be formalized as simultaneously minimizing the generalization error on the task distribution and the verification metric: + +$$ +\min _ {\theta_ {s}} \mathbb {E} _ {(x, y) \in P (x, y)} [ \mathcal {L} \left(f _ {\theta_ {s}} (x), y\right) ] + \alpha \cdot \operatorname {V e r i f} \left(f _ {\theta_ {s}}, \mathcal {D}, \mathcal {M}\right) \tag {1} +$$ + +Here, $f_{\theta_s}$ represents the surrogate student ultimately deployed, $P(x,y)$ the underlying distribution of samples $x,y \in \mathcal{D}$ , and $\mathcal{L}$ the loss function measuring prediction discrepancies. The tuning factor $\alpha$ balances the pursuit of generalization and evasion for $f_{\theta_s}$ . However, $P(x,y)$ and the process of $\mathrm{Verif}(f_{\theta_s},\mathcal{D},\mathcal{M})$ are agnostic to the adversary and intractable during optimization. Therefore, we propose two modules to better achieve and balance these objectives: (1) For generalization, we curate the most informative and reliable transfer set from a large-scale OOD gallery set to support high-performance and identifier-invariant knowledge extraction; (2) For evasion, to prevent the transmission of suspicious verification knowledge via dark knowledge in soft labels, we propose selective knowledge transfer to filter out such knowledge and better trade off between the two objectives. Next, we elaborate on these two components. + +# 3.2. Transfer Set Curation + +Our objective is to curate an optimal subset $\mathcal{T}$ from an OOD gallery set $\mathcal{G}$ (e.g., ImageNet [9] or DataComp [12]), selecting samples that are both informative and reliable to facilitate effective knowledge transfer. While this task may resemble classical core-set selection [5, 53], it presents significantly greater challenge: (1) Our algorithm operates in a fully OOD context, where $\mathcal{G}$ and $\mathcal{D}$ do not necessarily overlap in classes, and their distributions could vary substantially; (2) $\mathcal{D}$ could include identifiers manipulated by the defender (e.g., trigger samples), and the teacher $f_{\theta_t}$ trained directly on $\mathcal{D}$ is already marked. Consequently, selection based on the unreliable $\mathcal{D}$ and $f_{\theta_t}$ inadvertently introduce samples from $\mathcal{G}$ that activate verification behaviors, risking the induction of these behaviors in the surrogate student $f_{\theta_s}$ . + +In this paper, we leverage a vision-language model (VLM) with unbiased pretrained knowledge and robust zero-shot capabilities (e.g., CLIP [49]) to facilitate reliable selection alongside the marked teacher $f_{\theta_t}$ . CLIP aligns visual and textual features in a shared embedding space, where the image encoder $\phi_I(\cdot)$ maps an image $x$ into $\phi_I(x)$ , and text templates with category names (e.g., "a photo of {class name}") $c_1, \ldots, c_K \in \mathcal{D}$ are mapped by the text encoder $\phi_t(\cdot)$ to $\phi_t(c_1), \ldots, \phi_t(c_K)$ . By calculating the cosine similarity between $\phi_I(x)$ and $\phi_t(c_i)$ , each sample $x \in \mathcal{G}$ can be uniquely assigned to a category $t$ in $\mathcal{D}$ . + +However, class-name-only text templates lack foreground semantics [48], limiting the ability to distinguish target categories. While few-shot tuning allows the VLM to rapidly adapt to tasks within $\mathcal{D}$ [26], it risks being compromised by verification behaviors embedded in $\mathcal{D}$ [3]. Inspired by advances in zero-shot prompt learning [48], we adjust the VLM with unbiased category semantics from a large language model (LLM). The LLM generates a description set $\mathrm{Desc}_{c_i}$ for each category $c_{i}$ in $\mathcal{D}$ , averaging descriptions in the feature space to create representation prototypes for each category, enhancing the zero-shot capabilities of the VLM. Consequently, each $x\in \mathcal{G}$ is assigned to class $t$ in a zero-shot fashion as follows in Equation 2: + +$$ +t = \arg \max _ {c _ {i}} \frac {1}{\left| \operatorname {D e s c} _ {c _ {i}} \right|} \sum_ {d \in \operatorname {D e s c} _ {c _ {i}}} (\sin (\phi_ {I} (x), \phi_ {I} (d))) \tag {2} +$$ + +where sim denotes cosine similarity shown in Equation 3: + +$$ +\operatorname {s i m} \left(\phi_ {I} (x), \phi_ {I} (d)\right) = \frac {\phi_ {I} (x) \cdot \phi_ {I} (d)}{\left| \phi_ {I} (x) \right| \cdot \left| \phi_ {I} (d) \right|} \tag {3} +$$ + +Although the VLM assigns all gallery set samples to categories in $\mathcal{D}$ , it lacks a criterion for the most informative samples. To leverage the distribution from $\mathcal{D}$ while avoiding biases from potential verification behaviors, we project all samples in $\mathcal{D}$ into the feature space with $\phi_I(\cdot)$ and calculate the density centroids $\mathrm{Cent}_t$ for each class $t\in \{1,\dots ,K\}$ . These centroids, acting as distribution digests, encapsulate the conditional distribution of $\mathcal{D}$ . Samples closer to $\mathrm{Cent}_t$ in feature space are prioritized for selection, a robust and straightforward criterion that is theoretically resistant to manipulation by a small proportion of outlier samples [25]. + +We then establish a consensus-based ensemble for final transfer set curation: (1) The gallery set $\mathcal{G}$ is divided by the LLM-enhanced VLM into $K$ bins corresponding to classes in $\mathcal{D}$ . Within each bin $\mathcal{G}_t$ , (2) samples are ordered by proximity to distribution digest $\mathrm{Cent}_t$ ; (3) Only samples classified by the teacher $f_{\theta_t}$ as $t$ that are closest to $\mathrm{Cent}_t$ are selected, until the selected samples match the class count $|\mathcal{D}_t|$ . Illustrated in Figure 1, this process ensures all predictions on the transfer set $\mathcal{T}$ by $f_{\theta_t}$ are consistent with VLM, minimizing the risk of triggering verification behaviors while closely approximating the distribution in $\mathcal{D}$ . + +The Feature Bank. The curation process involves projecting all gallery set samples into the feature space, which is the most time-consuming step. However, when the gallery set is potentially reused, this projection only needs to be performed once. The resulting features and indexes are saved in a feature bank for efficient reuse. + +# 3.3. Selective Knowledge Transfer + +The transfer set $\mathcal{T}$ serves as an intermediary for extraction of task-oriented and identifier-invariant knowledge. This process employs the standard knowledge distillation framework [20] with the following optimization objective: + +$$ +\theta_ {s} ^ {*} = \arg \min _ {\theta_ {s}} \mathcal {L} \left(\frac {f _ {\theta_ {s}} (x)}{\tau}, \frac {f _ {\theta_ {t}} (x)}{\tau}\right), \quad x \in \mathcal {T} \tag {4} +$$ + +Here, $\mathcal{L}$ measures prediction discrepancies, such as the Kullback-Leibler divergence, modulated by the temperature $\tau$ . This process effectively transfers knowledge despite potential misalignments between the transfer set $\mathcal{T}$ and the training distribution [46]. However, the soft labels $f_{\theta_t}(x)$ contain dark knowledge [11], which inadvertently transmit verification behaviors. Thus, we introduce Selective Knowledge Transfer (SKT) to filter out suspicious verification knowledge during distillation. + +The exclusive and subtle verification behaviors reflect unique biases inherent to the marked dataset $\mathcal{D}$ , distinguishing it within the task distribution. Preventing the student $f_{\theta_s}$ from overfitting to biases limits DOV's ability to infer connections between $f_{\theta_s}$ and $\mathcal{D}$ through trigger-specific outputs or excessive confidence. Drawing on insights from shortcut learning [14], we mitigate predictive shortcuts during the student's learning process, reducing reliance on spurious features. Since dataset biases typically induce universal shortcut behaviors, a straightforward solution is to apply universal adversarial training (UAT) on the student: + +$$ +\theta_ {s} ^ {*} = \arg \min _ {\theta_ {s}} \max _ {\delta} \frac {1}{n} \sum_ {x, y \in \mathcal {T}} \mathcal {L} \left(f _ {\theta_ {s}} (x + \delta), y\right) \tag {5} +$$ + +UAT effectively suppresses atypical predictive behaviors in $f_{\theta_s}$ induced by perturbations $\delta$ . However, the min-max problem involves a notoriously hard bi-level optimization, requiring iterative solutions within each iteration. Furthermore, the absence of ground-truth labels in the OOD transfer set $\mathcal{T}$ necessitates predictions from the marked teacher as labels, introducing significant calibration errors. + +Given the objective to mitigate biases specific to the copyrighted dataset $\mathcal{D}$ , and with all verification behaviors encapsulated in the teacher model, we propose using only the teacher model $f_{\theta_t}$ to generate perturbations on $\mathcal{D}$ as a practical approximation for the inner max-problem: + +$$ +\delta^ {*} = \arg \max _ {\delta} \frac {1}{n} \sum_ {x, y \in \mathcal {D}} \mathcal {L} \left(f _ {\theta_ {t}} (x + \delta), y\right) \tag {6} +$$ + +This strategy decouples the bi-level optimization, rendering it asynchronously solvable. By pre-generating a perturbation pool $\{\delta_i\}_{i=1}^n$ in an offline manner with the teacher $f_{\theta_t}$ and reusing it in the outer min-problem during distillation, computational overhead is significantly reduced. + +To solve Equation 6, the perturbation $\delta$ is typically constrained by an $L_{p}$ norm to preserve core image semantics [40]. However, without prior knowledge of the specific form of $\delta$ associated with verification behaviors (e.g., watermark triggers), choosing an inappropriate norm undermine the efficacy of UAT in suppressing spurious features [64]. Thus, we seek to diversify the teacher-generated perturbations beyond a single norm constraint. + +We solve the max-problem using mini-batch stochastic gradient descent [54], applying $L_{2}$ constraints at each batch update to preserve gradient directions. After each iteration, we project the perturbation onto the norm constraint $(L_0, L_2, \text{or} L_\infty)$ that maximizes the loss. This steepest descent projection [41] adaptively seeks the optimal norm constraint, facilitating the generation of perturbations closely resembling those employed by DOV. Moreover, extreme projections onto $L_{0}$ or $L_{\infty}$ norms approximate a convex hull spanning multiple norm constraints, broadening the range of potential perturbations [7]. + +Ideally, perturbation should preserve core image semantics without being limited to specific norms [57]. Generative models can achieve this by producing unrestricted adversarial perturbations [76], though training such models on task-specific distributions is resource-intensive. Therefore, we propose a lightweight approach using image corruptions instead of generative models to produce unrestricted perturbations. Starting with the 15 corruptions defined in ImageNet-C [19], we employ a genetic algorithm [8, 43] to select the combination and sequence of corruptions that solves Equation 6, forming a corruption chain: + +$$ +\max \mathcal {L} \left(f _ {\theta_ {t}} \left(\operatorname {C o r r u p t} _ {1: N} (x)\right), y\right), \quad x, y \in \mathcal {D} \tag {7} +$$ + +The perturbations and corruption chain generated with the teacher $f_{\theta_t}$ on $\mathcal{D}$ are collectively referred to as $A(\cdot)$ , leading to the final objective for Selective Knowledge Transfer: + +$$ +\begin{array}{l} \arg \min _ {\theta_ {s}} \mathcal {L} \left(\frac {f _ {\theta_ {s}} (x)}{\tau}, \frac {f _ {\theta_ {t}} (x)}{\tau}\right) \tag {8} \\ + \beta \cdot \mathcal {L} \left(\frac {f _ {\theta_ {s}} (A (x))}{\tau}, \frac {f _ {\theta_ {t}} (x)}{\tau}\right), \quad x \in \mathcal {T} \\ \end{array} +$$ + +The second term promotes the student's invariance to the worst-case perturbations from the teacher while encouraging predictive divergence to reduce bias toward the teacher. The tuning factor $\beta$ balances generalization with evasion efficacy, applied as a sampling probability to selectively introduce perturbations or corruption chain to a subset of samples, adding minimal overhead. + +Table 1. Escaping DOV on CIFAR-10 and Tiny ImageNet. Red indicates detection by DOV, while green indicates successful evasion. + +
DOVCIFAR-10Tiny-Imagenet
VanillaEvasionVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)ACCVSRACC(↑)VSR(↓)
Badnets94.44 ± 0.74100.00 ± 0.0093.46 ± 0.201.36 ± 0.0769.05 ± 0.2598.98 ± 0.1166.48 ± 0.141.48 ± 0.11
UBW93.98 ± 0.2595.54 ± 0.8493.41 ± 0.191.74 ± 0.0666.37 ± 0.1275.56 ± 0.5065.34 ± 0.084.60 ± 0.11
Label-Consistent94.74 ± 0.5496.18 ± 2.7093.19 ± 0.143.74 ± 0.4069.07 ± 0.2434.99 ± 1.5366.46 ± 0.103.51 ± 0.22
Narcissus94.76 ± 0.5487.34 ± 1.8294.37 ± 0.184.59 ± 0.6468.99 ± 0.1457.49 ± 0.7866.39 ± 0.1812.33 ± 6.45
Non-Poisoning DOVACCp-valueACC(↑)p-value(↑)ACCp-valueACC(↑)p-value(↑)
Radioactive Data94.96 ± 0.473.03e-0394.07 ± 0.189.45e-0168.84 ± 0.305.33e-0366.64 ± 0.244.94e-01
ANW94.66 ± 0.311.37e-0993.91 ± 0.281.00e+0068.75 ± 0.305.57e-2866.44 ± 0.131.00e+00
Domain Watermark94.38 ± 0.531.67e-2293.90 ± 0.131.00e+0068.55 ± 0.393.91e-1666.38 ± 0.161.00e+00
Isotope94.75 ± 0.282.87e-0393.99 ± 0.492.84e-0168.97 ± 0.371.63e-0366.76 ± 0.221.66e-01
ML Auditor94.22 ± 0.177.47e-0593.45 ± 0.217.83e-0168.41 ± 0.281.55e-0665.35 ± 0.161.00e+00
Dataset Inference94.83 ± 0.471.87e-0393.97 ± 0.494.76e-0169.15 ± 0.349.27e-1466.06 ± 0.184.47e-01
MeFA94.83 ± 0.472.62e-1493.93 ± 0.551.00e+0069.15 ± 0.346.17e-2766.22 ± 0.209.99e-01
+ +# 4. Experiments + +# 4.1. Experimental Setup + +Following established DOV methods [17, 33, 42, 79], we adopt ResNet-18 [18] for both teacher and students. For transfer set curation, we utilize MobileCLIP [67] and GPT-40 mini [45] as lightweight VLM and LLM, respectively. + +We assess backdoor watermarks with Verification Success Rate (VSR)-the probability of a trigger-induced misclassification to the target class. A VSR above $30\%$ confirms successful verification [76], indicating unauthorized training. Non-poisoning watermarks and fingerprints, collectively termed non-poisoning DOV, are evaluated via the p-value from a one-tailed T-test. The p-value represents the confidence level that the model was not trained on the copyright dataset; lower p-values indicate successful verification. The common threshold for p-values is set at 0.01 [42], equating to a $99\%$ confidence level of unauthorized training. Experiments were repeated 3 times, reporting mean $\pm$ std for ACC and VSR, and harmonic mean for p-values. + +# 4.2. Escaping Advanced DOV Methods + +We evaluate Escaping DOV on natural image datasets commonly used as DOV benchmarks, including CIFAR-10 [28] and the 200-class Tiny ImageNet [29]. We further consider six datasets with significant distribution shifts from the gallery set (e.g., facial recognition and medical diagnosis), with a summary of results in Section 4.4.5 and detailed findings in Appendix C.3. For the gallery set, we use LSVRC-2012 [9]. Given overlapping samples between Tiny ImageNet and LSVRC-2012, we create a challenging setting by removing all overlapping classes from LSVRC-2012 to produce a purely OOD gallery set. The targeted DOV approaches cover advanced algorithms across all categories: (1) backdoor watermarks, including BadNets [16], UBW [32], Label-Consistent [65], and Narcissus [74], encompassing clean-label, invisible, and untargeted poisoning + +techniques; (2) non-poisoning watermarks, including Radioactive Data [50], ANW [79], Domain Watermark [17], Isotope [68], and ML Auditor [24]; and (3) fingerprints, including Dataset Inference [42] and MeFA [35]. + +Table 1 presents the generalization and verification metrics for the models before and after evasion attacks (the teacher and surrogate student in Escaping DOV). All DOV methods successfully identify the vanilla teacher trained directly on the copyright dataset. However, none meet their verification thresholds (VSR > 30% or p-value < 0.01) on the student produced by Escaping DOV, where VSR is near random guesses and p-values exceed 0.1 in most cases. Moreover, Escaping DOV incurs minimal impact on generalization: test accuracy on CIFAR-10 decreases by less than 1%, while on the challenging Tiny ImageNet set, with no overlapping gallery classes, accuracy drops by only 2-3%. Thus, Escaping DOV achieves a robust balance between universal evasion and preserved generalization. + +![](images/f858d9435b29ef401f6b61bff1b7372c374673963b9ea698120d10e8d6e344ca.jpg) +Figure 2. Loss Barrier between Teacher and Student Parameters. + +We further observe that Escaping DOV encourages the student to diverge from the teacher's predictive behavior, forming its own mechanism. In Figure 2, we linearly interpolate between teacher and student parameters, computing training and test losses at each point. The results reveal two key findings: (1) A loss barrier appears during interpolation from teacher $\theta_{t}$ to student $\theta_{s}$ . This lack of linear connectivity suggests that the surrogate student develops a distinct prediction mechanism, bypassing the spurious features central to the teacher's verification behavior [39]. (2) + +Table 2. Comparison of Escaping DOV with Evasion Attacks. Red indicates detection by DOV, green indicates successful evasion. + +
MethodBadnetsNarcissusIsotopeDataset Inference
ACC(↑)VSR(↓)ACC(↑)VSR(↓)ACC(↑)p-value(↑)ACC(↑)p-value(↑)
Fine-pruning86.21 ± 0.1699.37 ± 0.2786.58 ± 0.1543.75 ± 14.4186.78 ± 0.120.053286.88 ± 0.480.0095
Meta-Sift88.68 ± 0.4599.70 ± 0.3288.64 ± 0.6249.40 ± 8.2688.56 ± 0.470.026488.07 ± 0.210.0963
Differential Privacy90.49 ± 1.2076.70 ± 4.6287.70 ± 0.6872.77 ± 3.9188.69 ± 2.610.067889.64 ± 4.860.0440
I-BAU89.09 ± 0.2514.80 ± 0.4688.76 ± 1.9472.74 ± 2.1787.07 ± 2.590.155289.93 ± 0.680.0355
ZIP82.44 ± 0.1079.88 ± 0.3983.51 ± 0.3753.21 ± 4.4782.97 ± 0.420.009883.21 ± 0.900.1050
NAD89.86 ± 0.373.11 ± 0.5191.04 ± 0.3675.41 ± 5.1291.77 ± 0.450.096192.12 ± 0.340.0014
BCU92.75 ± 0.051.56 ± 0.2592.33 ± 0.1161.72 ± 6.4392.77 ± 0.360.162293.09 ± 0.180.0773
ABD84.94 ± 3.047.07 ± 3.1585.42 ± 2.1022.10 ± 1.1586.29 ± 3.460.006684.87 ± 3.020.3696
IPRemoval83.96 ± 0.238.09 ± 0.5085.06 ± 0.1155.97 ± 2.1484.64 ± 0.060.191685.28 ± 0.050.1322
Escaping DOV (Ours)93.46 ± 0.201.36 ± 0.0794.37 ± 0.184.59 ± 0.6493.99 ± 0.490.284593.97 ± 0.490.4759
+ +For the teacher, test loss is over $10^{3}$ times the training loss, allowing fingerprints to easily infer dataset association. In contrast, the student's test loss is typically under twice the training loss, falling within the variance across subsets and making it far less distinguishable. Thus, Escaping DOV significantly suppresses both watermarks and fingerprints. + +# 4.3. Comparison with SOTA Evasion Attacks + +We compare with SOTA evasion attacks on CIFAR-10 across the seminal Badnets watermark and three most resistant DOVs from each category—Narcissus, Isotope, and Dataset Inference (see Table 1). Results for all DOV methods are provided in Appendix C.2. Evasion attacks include Fine-Pruning [36], stronger combination of fine-tuning and pruning widely used in DOV evaluations, along with SOTA backdoor defenses across four stages: data sanitization with Meta-Sift [73], robust training with Differential Privacy [4], backdoor unlearning with I-BAU [72], and input purification with ZIP [56]. We also compare with methods that use knowledge distillation to mitigate backdoors, such as NAD [31], BCU [47] and ABD [22], as well as IPRemoval [78], which aims to eliminate model watermarks. Results in Table 2 and Appendix C.2 show that Escaping DOV consistently surpasses all comparison methods in both generalization and evasion, and is the only approach that fully evades all DOV methods. For instance, only ABD and Escaping DOV reduce VSR below $40\%$ on the clean-label and invisible Narcissus watermark; however, ABD fails to evade Isotope, while Escaping DOV achieves nearly $10\%$ higher accuracy with superior evasion performance. We provide a rigorous analysis in Appendix D, explaining why SOTA evasion attacks, particularly those relying on distillation, are less effective than our Escaping DOV. + +# 4.4. Ablations and Discussions + +# 4.4.1. Transfer Set Curation + +We begin with a qualitative analysis of the transfer set. Figure 3a visualizes the feature space for CIFAR-10 samples as well as selected and unselected ones from the transfer set, on a clean ResNet-18. The selected samples are notably + +closer to original CIFAR-10 samples. Figure 3b illustrates the feature space of a Badnets-marked model, where trigger samples form distinct clusters with almost no transfer set samples nearby. The exclusive and subtle verification behaviors are rarely activated by OOD samples, and the curation process effectively excludes suspicious samples. + +![](images/30aa5cca0a72ec5148e512446ea08bc55c86c7d8996762fa937e449053f6e80a.jpg) +Figure 3. Sample Distribution in the Feature Space. + +![](images/91d664e9df5cabf3b5cbb3ccee4961cf9a966c9c3c63fab22cd7e739105b89ef.jpg) + +Quantitative analysis follows, using Optimal Transport Dataset Distance (OTDD) [2] to assess the distance from the target dataset to each proxy distribution, as shown in Table 3. The curated transfer set is significantly closer to the target distribution than random sampling. Further, we compare the curated transfer set with random and unrelative samples on CIFAR-10 with Narcissus. Table 4 shows that transfer set curation not only improves generalization but also suppresses DOV by selecting more reliable samples. + +Table 3. Optimal Transport Distances Between Datasets. + +
DistanceselfLSVRC(curation)LSVRC(random)GTSRB [58]
CIFAR-100.7710.9371.1161.129
Tiny-Imagenet0.8550.9880.9931.066
+ +Table 4. Escaping DOV across different Transfer Sets. + +
Transfer SetGTSRB [58]LSVRC(random)LSVRC(curation)
ACC(↑)66.80 ± 0.4191.63 ± 0.4994.37 ± 0.18
VSR(↓)8.96 ± 1.637.87 ± 0.524.59 ± 0.64
+ +# 4.4.2. Selective Knowledge Transfer + +Figure 4 illustrates the impact of Selective Knowledge Transfer (SKT) on the most resistant Narcissus watermark across distillation temperatures $\tau$ . As $\tau$ increases, omitting SKT causes a substantial increase in VSR, while SKT consistently suppresses verification behavior across all temper + +atures. Additionally, VSR escalates sharply at higher temperatures, while test accuracy remains nearly unchanged. We hypothesis that task and verification knowledge possess distinct "boiling points", naturally separating at lower temperatures. Thus, we adopt $\tau = 1$ as a reliable default. + +![](images/faf7efc6e4773da8f48a83aacd5b94d43ca8e3191607e62a79c70b3a612db076.jpg) +Figure 4. Impact of SKT and Distillation Temperature. + +We further compare the offline perturbation pool in SKT with direct UAT on the student. While UAT lowers VSR to $2.07\%$ , it incurs a $3.54\%$ accuracy reduction and more than triples computational cost. As SKT balances evasion and generalization, it serves as a lightweight approximation. Detailed time complexity is analysed in Appendix C.8. + +# 4.4.3. Escaping Adaptive Defenses + +We assume a knowledgeable defender employing adaptive verification, specifically a semantic backdoor [60] and an anti-distillation backdoor [13]. For the semantic backdoor, we assume the defender knows the gallery set and selects 500 "tench" samples from LSVRC-2012, relabeling them as "plane" and inserting them into CIFAR-10 to test their classification probability as "plane" (VSR). Since the teacher memorizes this verification behavior, using it for sample selection allows triggers enter the transfer set. For instance, DFND [13], an entropy-based selection method, produces a student with VSR of $70.2\%$ by including 254 "tench" samples in the transfer set. In contrast, our curated transfer set contains fewer than 10 "tench" samples, reducing VSR to only $17.4\%$ . Even if the defender knows the gallery set, Transfer Set Curation still selects reliable samples. We also evaluate the entropy of teacher annotations: DFND's entropy is 2.90, while ours reaches 3.32, near the theoretical optimum $\log_2(10)$ , indicating better class balance. + +Table 5. Escaping DOV against the Adaptive ADB Watermark. + +
μVanillaEvasionIndependent
ACCVSRACCVSRACCVSR
0.191.59%100.00%90.14%94.42%95.13%92.89%
190.90%100.00%89.23%78.48%95.13%71.74%
1091.26%12.88%88.66%14.21%95.13%10.01%
+ +The anti-distillation backdoor (ADB) [13] is designed to transfer to student during distillation. As we employ distillation for knowledge transfer, we test ADB under a hypothetical scenario where the defender controls the teacher's training process to optimize the trigger. Table 5 reports the VSR across teacher, student, and unmarked models. A smaller $\mu$ allows larger perturbations, turning + +the trigger into an adversarial perturbation that induces false positives on unmarked models. Despite the defender's unrealistic control over the teacher, students' VSR remains closer to that of an unmarked model. As $\mu$ increases, ADB becomes ineffective. Thus, even under defender-controlled teacher training, Escaping DOV remains a potent threat. + +# 4.4.4. Enhanced Evasion Tactics + +Escaping DOV involves only post-training adjustments and operates orthogonally to pre-training, in-training and test-time strategies, all of which can collectively enhance evasion. As shown in Table 6, replacing data augmentation with mixup [75] during teacher training—an approach that mitigates backdoors and overfitting—further boosts evasion efficacy in conjunction with Escaping DOV on CIFAR-10. + +Table 6. Mixup Teacher Training for Escaping DOV. + +
DOVVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)
Narcissus94.04%72.47%93.03%3.66%
Non-poisoning DOVACCp-valueACC(↑)p-value(↑)
Isotope94.03%0.296492.71%0.7013
Dataset Inference94.60%0.093792.78%0.5196
+ +# 4.4.5. Case Study on Domain-Specific Datasets + +Datasets in vertical industries often require stronger protection than natural images due to collection challenges and privacy concerns. We conduct a case study on RAFDB (facial emotion) [30] and OrganCMNIST (medical imaging) [71] in Table 7, with detailed results in Appendix C.3. Escaping DOV maintains generalization and efficacy despite unique distribution, while a larger domain-related gallery and VLM could further improve performance [34, 38]. + +Table 7. Escaping DOV on Domain-specific Datasets. + +
SettingsVanillaEvasion
DOVDataSetACCVSRACC(↑)VSR(↓)
BadnetsRAFDB85.21 ± 0.3399.95 ± 0.0280.53 ± 0.422.19 ± 0.06
OrganC95.32 ± 0.0699.99 ± 0.0190.57 ± 0.700.83 ± 0.15
DOVDataSetACCp-valueACC(↑)p-value(↑)
IsotopeRAFDB82.46 ± 0.591.50e-0680.37 ± 0.131.05e-01
OrganC94.63 ± 0.192.65e-0388.96 ± 0.565.04e-01
+ +# 5. Conclusion + +In this paper, we reveal that Data Ownership Verification, widely used to prevent unauthorized model training, is vulnerable to strong evasion attacks. We introduce Escaping DOV, which leverages transfer set curation and selective knowledge transfer to achieve superior generalization and universal evasion capabilities. We propose that Escaping DOV serve as a evaluation framework to help establish truly reliable DOV methods, with plans for extensive cross-modal experiments to explore broader societal benefits. + +# Acknowledgments + +The work was supported by the National Natural Science Foundation of China under Grant 62271307 and 61771310. + +# References + +[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdooring. In 27th USENIX security symposium (USENIX Security 18), pages 1615-1631, 2018. 2 +[2] David Alvarez-Melis and Nicolo Fusi. Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems, 33:21428-21439, 2020. 7 +[3] Jiawang Bai, Kuofeng Gao, Shaobo Min, Shu-Tao Xia, Zhifeng Li, and Wei Liu. Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24239-24250, 2024. 4 +[4] Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. Automatic clipping: Differentially private deep learning made easier and stronger. Advances in Neural Information Processing Systems, 36, 2023. 3, 7 +[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6428-6437, 2021. 4 +[6] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE transactions on image processing, 6(12):1673-1687, 1997. 1 +[7] Francesco Croce and Matthias Hein. Adversarial robustness against multiple and single $l$ - $p$ -threat models via quick finetuning of robust classifiers. In International Conference on Machine Learning, pages 4436-4454. PMLR, 2022. 5 +[8] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation, 6(2):182-197, 2002. 5 +[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 4, 6 +[10] European Parliament and Council of the European Union. General data protection regulation (gdpr): Regulation (eu) 2016/679 on data protection and free movement, 2016. Available at https://gdpr-info.eu/.1 +[11] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In International conference on machine learning, pages 1607-1616. PMLR, 2018. 5 +[12] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal + +datasets. Advances in Neural Information Processing Systems, 36, 2024. 4 +[13] Yunjie Ge, Qian Wang, Baolin Zheng, Xinlu Zhuang, Qi Li, Chao Shen, and Cong Wang. Anti-distillation backdoor attacks: Backdoors can really survive in knowledge distillation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 826-834, 2021. 8 +[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 5 +[15] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1563-1580, 2022. 2 +[16] T Gu, B Dolan-Gavitt, and S BadNets. Identifying vulnerabilities in the machine learning model supply chain. In Proceedings of the Neural Information Processing Symposium Workshop Mach. Learning Security (MLSec), pages 1-5, 2017. 2, 6 +[17] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6 +[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6 +[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 5 +[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5 +[21] Mathew Hogan, Yan Michalevsky, and Saba Eskandarian. Dbreach: Stealing from databases using compression side channels. In 2023 IEEE Symposium on Security and Privacy (SP), pages 182-198. IEEE, 2023. 1 +[22] Junyuan Hong, Yi Zeng, Shuyang Yu, Lingjuan Lyu, Ruoxi Jia, and Jiayu Zhou. Revisiting data-free knowledge distillation with poisoned teachers. In International Conference on Machine Learning, pages 13199-13212. PMLR, 2023. 7 +[23] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 1 +[24] Zonghao Huang, Neil Zhenqiang Gong, and Michael K Reiter. A general framework for data-use auditing of ml models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 1300-1314, 2024. 2, 6 +[25] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against + +data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 4 +[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4 +[27] Xiao Jin, Pin-Yu Chen, Chia-Yi Hsu, Chia-Mu Yu, and Tianyi Chen. Cafe: Catastrophic data leakage in vertical federated learning. Advances in Neural Information Processing Systems, 34:994-1006, 2021. 1 +[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6 +[29] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6 +[30] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 8 +[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021. 3, 7 +[32] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 6 +[33] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 1, 2, 6 +[34] Yudong Li, Xianxu Hou, Zheng Dezhi, Linlin Shen, and Zhe Zhao. Flip-80m: 80 million visual-linguistic pairs for facial language-image pre-training. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 58–67, 2024. 8 +[35] Gaoyang Liu, Tianlong Xu, Xiaoqiang Ma, and Chen Wang. Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security, 17:1024-1037, 2022. 2, 3, 6 +[36] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 3, 7 +[37] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 1 +[38] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 8 + +[39] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka. Mechanistic mode connectivity. In International Conference on Machine Learning, pages 22965-23004. PMLR, 2023. 6 +[40] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5 +[41] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 5 +[42] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 2, 3, 6 +[43] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 8884-8892, 2021. 5 +[44] NBC News. Facial recognition's 'dirty little secret': Millions of online photos scraped without consent, 2019. Available at https://www.nbcnews.com/news/all/facial-recognition-s-dirty-little-secret-millionphotos-online-photos-scraped-n981921.1 +[45] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Available at https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/.6 +[46] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4954-4963, 2019. 5 +[47] Lu Pang, Tianlong Sun, Huan Ling, and Changyou Chen. Backdoor cleansing with unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12218-12227, 2023. 7 +[48] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 4 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4 +[50] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 6 +[51] Pedro Sandoval-Segura, Vasu Singla, Jonas Geiping, Micah Goldblum, and Tom Goldstein. What can we learn from unlearnable datasets? Advances in Neural Information Processing Systems, 36, 2023. 1 + +[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1 +[53] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 4 +[54] Ali Shafahi, Mahyar Najibi, Zheng Xu, John Dickerson, Larry S Davis, and Tom Goldstein. Universal adversarial training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5636-5643, 2020. 5 +[55] Shuo Shao, Yiming Li, Mengren Zheng, Zhiyang Hu, Yukun Chen, Boheng Li, Yu He, Junfeng Guo, Tianwei Zhang, Dacheng Tao, et al. Databench: Evaluating dataset auditing in deep learning from an adversarial perspective. arXiv preprint arXiv:2507.05622, 2025. 3 +[56] Yucheng Shi, Mengnan Du, Xuansheng Wu, Zihan Guan, Jin Sun, and Ninghao Liu. Black-box backdoor defense via zero-shot image purification. Advances in Neural Information Processing Systems, 36:57336-57366, 2023. 3, 7 +[57] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in neural information processing systems, 31, 2018. 5 +[58] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011 international joint conference on neural networks*, pages 1453–1460. IEEE, 2011. 7 +[59] Salvatore J Stolfo, Malek Ben Salem, and Angelos D Keromytis. Fog computing: Mitigating insider data theft attacks in the cloud. In 2012 IEEE symposium on security and privacy workshops, pages 125-128. IEEE, 2012. 1 +[60] Bing Sun, Jun Sun, Wayne Koh, and Jie Shi. Neural network semantic backdoor detection and mitigation: A causality-based approach. In Proceedings of the 33rd USENIX Security Symposium. USENIX Association, San Francisco, CA, USA, 2024. 8 +[61] Sebastian Szyller and N Asokan. Conflicting interactions among protection mechanisms for machine learning models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15179-15187, 2023. 3 +[62] Sebastian Szyller, Rui Zhang, Jian Liu, and N. Asokan. On the robustness of dataset inference. Transactions on Machine Learning Research, 2023. 3 +[63] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 1, 2 +[64] Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. Advances in neural information processing systems, 32, 2019. 5 +[65] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 6 + +[66] Rolf Van Wegberg, Samaneh Tajalizadehkhoob, Kyle Soska, Ugur Akyazi, Carlos Hernandez Ganan, Bram Klievink, Nicolas Christin, and Michel Van Eeten. Plug and prey? measuring the commoditization of cybercrime via online anonymous markets. In 27th USENIX security symposium (USENIX security 18), pages 1009-1026, 2018. 1 +[67] Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, and Oncel Tuzel. Mobile-clip: Fast image-text models through multi-modal reinforced training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15963–15974, 2024. 6 +[68] Emily Wenger, Xiuyu Li, Ben Y Zhao, and Vitaly Shmatikov. Data isotopes for data provenance in dnns. Proceedings on Privacy Enhancing Technologies, 2024(1), 2024. 2, 3, 6 +[69] Wikipedia contributors. Cambridge analytica — wikipedia, the free encyclopedia. https://en.wikipedia.org/wiki/Cambridge_Analytica, 2024.1 +[70] Nuo Xu, Qi Liu, Tao Liu, Zihao Liu, Xiaochen Guo, and Wujie Wen. Stealing your data from compressed machine learning models. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. 1 +[71] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 1, 8 +[72] Yi Zeng, Si Chen, Won Park, Zhuoqing Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. In International Conference on Learning Representations, 2022. 3, 7 +[73] Yi Zeng, Minzhou Pan, Himanshu Jahagirdar, Ming Jin, Lingjuan Lyu, and Ruoxi Jia. Meta-Sift: How to sift out a clean subset in the presence of data poisoning? In 32nd USENIX Security Symposium (USENIX Security 23), pages 1667-1684, Anaheim, CA, 2023. USENIX Association. 2, 3, 7 +[74] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pages 771-785, 2023. 2, 6 +[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 8 +[76] Hongyu Zhu, Sichu Liang, Wentao Hu, Li Fangqi, Ju Jia, and Shi-Lin Wang. Reliable model watermarking: Defending against theft without compromising on evasion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10124-10133, 2024. 2, 5, 6 +[77] Hongyu Zhu, Sichu Liang, Wentao Hu, et al. Stealing knowledge from auditable datasets. In Proceedings of the European Conference on Artificial Intelligence (ECAI), 2025. 3 +[78] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative + +model inversion attack against deep neural network fingerprinting and watermarking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7837-7845, 2024. 7 +[79] Zihang Zou, Boqing Gong, and Liqiang Wang. Anti-neuron watermarking: protecting personal data against unauthorized neural networks. In European Conference on Computer Vision, pages 449-465. Springer, 2022. 2, 3, 6 +[80] Chaoshun Zuo, Zhiqiang Lin, and Yinqian Zhang. Why does your data leak? uncovering the data leakage in cloud from mobile apps. In 2019 IEEE Symposium on Security and Privacy (SP), pages 1296-1310. IEEE, 2019. 1 \ No newline at end of file diff --git a/2025/Evading Data Provenance in Deep Neural Networks/images.zip b/2025/Evading Data Provenance in Deep Neural Networks/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..5122f5f9f804820878c698ce99661d5cc439a6d8 --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4c398a5b50a0ae90080f1bfcbad6d046d663a1aca3861ab174b2061f06872b41 +size 665323 diff --git a/2025/Evading Data Provenance in Deep Neural Networks/layout.json b/2025/Evading Data Provenance in Deep Neural Networks/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..2c1f0f6cb097534fa19e82f17eb8fb3093f86d4b --- /dev/null +++ b/2025/Evading Data Provenance in Deep Neural Networks/layout.json @@ -0,0 +1,10349 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "spans": [ + { + "bbox": [ + 145, + 103, + 465, + 121 + ], + "type": "text", + "content": "Evading Data Provenance in Deep Neural Networks" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "spans": [ + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": "Hongyu Zhu" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": " Sichu Liang" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": " Wenwen Wang" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": " Zhuomeng Zhang" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": " Fangqi Li" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": " Shi-Lin Wang" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": "Shanghai Jiao Tong University " + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": "Southeast University " + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 141, + 561, + 185 + ], + "type": "text", + "content": "Carnegie Mellon University Code: https://github.com/dbsxfz/EscapingDOV" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 151, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 238, + 297, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 238, + 297, + 610 + ], + "spans": [ + { + "bbox": [ + 55, + 238, + 297, + 610 + ], + "type": "text", + "content": "Modern over-parameterized deep models are highly data-dependent, with large scale general-purpose and domain-specific datasets serving as the bedrock for rapid advancements. However, many datasets are proprietary or contain sensitive information, making unrestricted model training problematic. In the open world where data thefts cannot be fully prevented, Dataset Ownership Verification (DOV) has emerged as a promising method to protect copyright by detecting unauthorized model training and tracing illicit activities. Due to its diversity and superior stealth, evading DOV is considered extremely challenging. However, this paper identifies that previous studies have relied on oversimplistic evasion attacks for evaluation, leading to a false sense of security. We introduce a unified evasion framework, in which a teacher model first learns from the copyright dataset and then transfers task-relevant yet identifier-independent domain knowledge to a surrogate student using an out-of-distribution (OOD) dataset as the intermediary. Leveraging Vision-Language Models and Large Language Models, we curate the most informative and reliable subsets from the OOD gallery set as the final transfer set, and propose selectively transferring task-oriented knowledge to achieve a better trade-off between generalization and evasion effectiveness. Experiments across diverse datasets covering eleven DOV methods demonstrate our approach simultaneously eliminates all copyright identifiers and significantly outperforms nine state-of-the-art evasion attacks in both generalization and effectiveness, with moderate computational overhead. As a proof of concept, we reveal key vulnerabilities in current DOV methods, highlighting the need for long-term development to enhance practicality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 135, + 647 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 656, + 296, + 693 + ], + "type": "text", + "content": "The success of modern deep learning hinges heavily on abundant datasets, spanning from large-scale vision and multi-modal resources such as ImageNet [9] and LAION-" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 370 + ], + "type": "text", + "content": "5B [52] to datasets tailored for vertical industries like face recognition [37] and computational pathology [71]. Curated meticulously through extensive human effort for collection, cleaning, and labeling, these datasets are protected as intellectual property and are available solely through licensed access. Some are released for academic purposes with restrictions on commercial exploitation, while others are proprietary assets, strictly prohibiting third-party model training. Moreover, personal data—such as photos shared on social platforms—is increasingly at risk of being scraped without consent [44]. Consequently, unauthorized model training on restricted data poses an escalating threat to both intellectual property and privacy rights." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 373, + 556, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 556, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 556, + 492 + ], + "type": "text", + "content": "Despite growing regulatory enforcement, like the General Data Protection Regulation (GDPR) [10], preventing unauthorized training remains a formidable challenge. Every stage of the data supply chain is vulnerable to attacks that funnel data into illicit markets [66], including insider threats [59], side-channel exploits [21], illegal mobile app access [80], and leakage from pretrained models [27, 70]. High-profile cases of data misuse in model training, such as the Cambridge Analytica scandal [69] and the Flickr Leakage [44], have intensified public scrutiny." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 556, + 663 + ], + "type": "text", + "content": "Consequently, preventing unauthorized model training at its source—the dataset itself—is crucial. Unlearnable Examples [23] introduce perturbations to prevent models from learning meaningful features. However, beyond their susceptibility to adaptive attacks [51], they entirely obstruct legitimate dataset use. As an alternative, post hoc verification for data provenance, known as Dataset Ownership Verification (DOV), has garnered substantial attention. Inspired by watermarks used to track duplication, remixing, or exploitation of multi-modal content [6], the backdoor watermark embeds carefully crafted \"poison\" samples selected by the dataset owner. Models trained on watermarked data exhibit predefined behaviors in response to backdoor triggers, enabling model-agnostic owner attribution [33, 63]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "To mitigate the impact of backdoor on authorized training, recent techniques introduce non-poisoning watermarks that adjust prediction confidence without inducing misclassification, employing hypothesis testing for ownership trac" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 212, + 713 + ], + "type": "text", + "content": "*Equal contribution. † Corresponding author." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1249" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "ing [50, 68, 79]. Non-intrusive dataset fingerprint have also emerged, allowing verification via elevated confidence on memorized training samples [35, 42]. DOV is widely regarded as an effective, if not the only, solution for tracking unauthorized training without disrupting legitimate usage [17, 32], with rapid advancements underway." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 147, + 298, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 147, + 298, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 147, + 298, + 397 + ], + "type": "text", + "content": "However, when deployed in the real world, the arms race between attackers and defenders forms an ongoing strategic game. Adversaries attempt to suppress verification behaviors to evade detection, while defenders continually strengthen verification mechanisms to reliably identify unauthorized training. Validating the robustness of DOV against evasion attempts is thus essential for practical deployment. Despite substantial progress in DOV, advances in evasion attacks remain underdeveloped, relying on simple regularization techniques and generic adaptations from poisoning defenses like fine-tuning [17, 32, 35, 79], which lack the versatility to counter diverse DOV techniques. Meanwhile, advanced DOV strategies, such as clean-label, invisible, or untargeted backdoor watermarks provide exceptional flexibility; non-poisoning watermarks and dataset fingerprints, which avoid misclassification or require no dataset modifications, remain nearly undetectable. It is widely agreed that designing a universal attack capable of evading all forms of verification is nearly impossible [42, 68]. This gap between the advancement of DOV and evasion techniques hinders reliable assessments." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 495 + ], + "type": "text", + "content": "In this paper we ask: is current DOV robust enough to preclude any evasion attempt? Our findings reveal that prior evasion attacks are too weak, leading to a false sense of security. Through analysis of verification behaviors, we identify shared characteristics: they are both exclusive and subtle, serving as unique markers that distinguish the protected dataset while remain incompatible with underlying semantic distribution of the main task." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 298, + 714 + ], + "type": "text", + "content": "Building on this insight, we propose the first universal evasion framework, Escaping DOV, which requires no additional clean in-distribution data or assumptions about the verification process. In Escaping DOV, a teacher model is initially trained directly on the copyright dataset to fully absorb domain knowledge. Then, an OOD transfer set, unrelated to the protected data, serves as an intermediary to extract task-oriented yet identifier-free knowledge to a surrogate student for deployment. Intuitively, exclusive and subtle identifiers, acting as side-channel signals orthogonal to the task distribution, make it improbable for OOD samples to trigger verification behaviors. Thus, only essential, identifier-invariant task knowledge is likely to transfer to the student model. To enhance Escaping DOV, we propose Transfer Set Curation, leveraging the unbiased knowledge embedded in large language models (LLMs) and vision-language models (VLMs) to retrieve the most informative and reliable samples from a general OOD gallery set. Ad" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 120 + ], + "type": "text", + "content": "ditionally, we introduce Selective Knowledge Transfer to block suspicious verification behaviors, achieving better balance between generalization and evasion. The overall pipeline for Escaping DOV is shown in Figure 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 121, + 511, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 121, + 511, + 132 + ], + "spans": [ + { + "bbox": [ + 326, + 121, + 511, + 132 + ], + "type": "text", + "content": "Our contributions are summarized as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 133, + 553, + 264 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 315, + 133, + 553, + 181 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 133, + 553, + 181 + ], + "spans": [ + { + "bbox": [ + 315, + 133, + 553, + 181 + ], + "type": "text", + "content": "1. We reveal that current DOV robustness evaluations are insufficiently rigorous. We introduce Escaping DOV, the first universal evasion framework that achieves simplicity, efficacy, and generalizability." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 182, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 182, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 314, + 182, + 553, + 216 + ], + "type": "text", + "content": "2. We propose Transfer Set Curation and Selective Knowledge Transfer to achieve balance between generalization and evasion with moderate overhead." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 217, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 217, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 314, + 217, + 553, + 264 + ], + "type": "text", + "content": "3. Experiments on diverse datasets and 11 DOV methods, alongside comparisons with 9 SOTA evasion methods, validate the efficacy of Escaping DOV, establishing it as a reliable benchmark for future DOV assessments." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 278, + 401, + 291 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 278, + 401, + 291 + ], + "spans": [ + { + "bbox": [ + 314, + 278, + 401, + 291 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 298, + 484, + 312 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 298, + 484, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 298, + 484, + 312 + ], + "type": "text", + "content": "2.1. Dataset Ownership Verification" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 317, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 317, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 317, + 555, + 437 + ], + "type": "text", + "content": "Backdoor attacks inject poisoned samples to associate predictions with a specific trigger [16], which induces misclassification whenever the trigger appears [15]. When controlled by copyright owners, backdoors can also act as identifiers for model ownership [1]. Inspired by this, the straightforward backdoor watermark embeds backdoor samples in the dataset. Models trained on this marked data produce a predefined label with high probability when evaluated on trigger samples [33]. The defender's knowledge of this secret prediction behavior serves as proof of ownership." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "text", + "content": "Early backdoors link triggers to target classes by manipulating the labels of poisoned samples. However, due to the obvious nature of poisoned labels, these samples are easily filtered through automatic data sanitization [73]. Clean-label backdoor watermark [63] utilizes advanced poisoning techniques [65] to enforce backdoor behavior without modifying labels. Later advancements further constrain modifications within an " + }, + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 313, + 437, + 554, + 593 + ], + "type": "text", + "content": " norm ball [74], making such clean-label, invisible backdoor watermarks highly stealthy [73]. Recently, untargeted backdoor watermark [32] induces nondeterministic misclassification without a fixed target class, enhancing resilience against post-training defenses such as trigger synthesis and output distribution analysis." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 714 + ], + "type": "text", + "content": "Advanced backdoor watermark provides a stealthy and resilient verification mechanism. However, it inevitably embeds harmful shortcuts [76], problematic when authorized training is essential [17]. Non-poisoning watermark subtly embeds watermark features into the dataset without inducing misclassification. Verification is performed by measuring loss differences between watermarked and clean samples via hypothesis testing. Various mechanisms have been developed to generate stealthy watermarks: Radioactive data [50] and ML Auditor [24] optimize" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1250" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 73, + 553, + 279 + ], + "blocks": [ + { + "bbox": [ + 61, + 73, + 553, + 279 + ], + "lines": [ + { + "bbox": [ + 61, + 73, + 553, + 279 + ], + "spans": [ + { + "bbox": [ + 61, + 73, + 553, + 279 + ], + "type": "image", + "image_path": "3ce822d7d903579f10901ce1d1ad52061dc5d06557d686f6c473c2304b50b515.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "text", + "content": "Figure 1. Overall Pipeline of Escaping DOV. (a) Transfer Set Curation (" + }, + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\S 3.2" + }, + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "text", + "content": "): Samples from the OOD gallery set are clustered via a VLM, ranked by distribution proximity, and filtered through consensus ensemble to curate an optimal transfer set. (b) Selective Knowledge Transfer (" + }, + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\S 3.3" + }, + { + "bbox": [ + 55, + 282, + 555, + 316 + ], + "type": "text", + "content": "): Promoting invariance to teacher's worst perturbations, facilitating task-oriented yet identifier-free knowledge transfer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 323, + 295, + 372 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 295, + 372 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 55, + 323, + 295, + 372 + ], + "type": "text", + "content": " -perturbations, Anti-Neuron Watermarking [79] applies color transformations in hue space, Domain Watermark [17] converts samples into a hardly-generalizable domain, while Isotope [68] blends external features into the dataset." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 373, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 373, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 373, + 296, + 469 + ], + "type": "text", + "content": "Additionally, dataset fingerprint utilizes intrinsic features for verification without modifying the dataset. Dataset inference [42] analyzes the margins between training and external samples relative to the decision boundary, while MeFA [35] aggregates membership inference across multiple samples. Although dataset fingerprints are non-intrusive and support post hoc protection for published datasets, they are susceptible to high false positives [61, 62]." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 272, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 272, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 272, + 492 + ], + "type": "text", + "content": "2.2. Evasion Attacks against Data Provenance" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Despite rapid progress in DOV, current evasion strategies rely on basic regularization and backdoor defenses like finetuning [17, 32, 36]. As they barely evade any DOV methods, we present their results in Appendix C.1 and use SOTA backdoor defenses as baselines, which are categorized by stage of application: (1) Data sanitization removes suspicious samples pre-training [73]; (2) Robust learning mitigates poisoning effects during training [4, 31]; (3) Backdoor unlearning synthesizes triggers post-training to neutralize the model [72]; and (4) Input preprocessing detects or purifies triggers during testing [56]. Together, these defenses span the model lifecycle; yet advanced invisible, clean-label, or untargeted backdoor watermarks can bypass most of them [17, 32]. Further, non-poisoning watermarks and non-intrusive fingerprints render backdoor defenses ineffective, making universal evasion extremely challenging. This work is the first to establish a unified evasion perspective, opening new directions for robust evaluation of DOV." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 348 + ], + "type": "text", + "content": "We discuss the contemporaneous works [77] and [55] that emerged at the time of acceptance in Appendix E." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 361, + 405, + 375 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 361, + 405, + 375 + ], + "spans": [ + { + "bbox": [ + 314, + 361, + 405, + 375 + ], + "type": "text", + "content": "3. Escaping DOV" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 382, + 382, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 382, + 382, + 394 + ], + "spans": [ + { + "bbox": [ + 314, + 382, + 382, + 394 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": "In " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": "-class image classification, an adversary gains unauthorized access to the copyright dataset " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " to train a model " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": ", which maps images " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "x \\in \\{0,1,\\ldots,255\\}^{C \\times W \\times H}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " to predictions " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "y \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": ". For data provenance, the defender employs a DOV method " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\operatorname{Verif}(f_{\\theta},\\mathcal{D},\\mathcal{M})" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": ", inspecting the output of the suspicious model " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " on specific queries. This process yields a verification metric indicating the likelihood that " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " was trained on " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": ", such as the Attack Success Rate for backdoor watermarks or the confidence for rejecting the null hypothesis in non-poisoning DOV. Here, " + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 402, + 555, + 545 + ], + "type": "text", + "content": " represents external materials (e.g., watermark trigger patterns) controlled by the defender to uniquely identify the dataset." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 556, + 713 + ], + "type": "text", + "content": "DOV methods are flexible and diverse. Backdoor watermark achieves superior stealth through invisible, clean-label, or untargeted poison techniques. Non-poisoning watermarks avoid misclassification, while dataset fingerprints require no dataset modifications. Hence, direct sanitization or suppression of all potential verification behaviors becomes a formidable challenge; most evasion attempts that simply adapt poison defenses have consistently failed [17, 32, 42, 68, 79]. However, we identify a common trait among all DOV strategies: verification behaviors are both exclusive and subtle. On one hand, these behaviors must exclusively belong to the copyright dataset to prevent false attribution to models trained on external datasets. On the other hand, they must be subtle enough to remain concealed" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 314, + 757 + ], + "type": "text", + "content": "1251" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 107 + ], + "type": "text", + "content": "during data sanitization, ensuring they are rarely triggered by regular samples to preserve the generalization of models trained for authorized use." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": "Accordingly, we propose a universal evasion strategy: first, a teacher model " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": " is trained on the copyright dataset " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": " to absorb task-domain knowledge, inevitably marked by DOV. Next, an unrelated OOD transfer set " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": " serves as the medium to extract task-oriented yet identifier-invariant knowledge into a surrogate student " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": ". Intuitively, the exclusive and subtle watermark triggers are absent in the natural samples of " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": ", preventing the activation and transmission of verification behaviors to the surrogate student. For fingerprints based on exaggerated confidence, the student only indirectly inherits knowledge from the marked teacher via the unrelated OOD data, significantly reducing excessive memorization associated with " + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 108, + 295, + 262 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 263, + 295, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 299 + ], + "type": "text", + "content": "The overall objective of Escaping DOV can be formalized as simultaneously minimizing the generalization error on the task distribution and the verification metric:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 306, + 295, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 306, + 295, + 324 + ], + "spans": [ + { + "bbox": [ + 61, + 306, + 295, + 324 + ], + "type": "interline_equation", + "content": "\\min _ {\\theta_ {s}} \\mathbb {E} _ {(x, y) \\in P (x, y)} [ \\mathcal {L} \\left(f _ {\\theta_ {s}} (x), y\\right) ] + \\alpha \\cdot \\operatorname {V e r i f} \\left(f _ {\\theta_ {s}}, \\mathcal {D}, \\mathcal {M}\\right) \\tag {1}", + "image_path": "fb3ae4400264a6ef481d3721af29c0df0f226f6edd0e9a60a406ca8eb7c80639.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " represents the surrogate student ultimately deployed, " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "P(x,y)" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " the underlying distribution of samples " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "x,y \\in \\mathcal{D}" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " the loss function measuring prediction discrepancies. The tuning factor " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " balances the pursuit of generalization and evasion for " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": ". However, " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "P(x,y)" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " and the process of " + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "inline_equation", + "content": "\\mathrm{Verif}(f_{\\theta_s},\\mathcal{D},\\mathcal{M})" + }, + { + "bbox": [ + 55, + 331, + 295, + 522 + ], + "type": "text", + "content": " are agnostic to the adversary and intractable during optimization. Therefore, we propose two modules to better achieve and balance these objectives: (1) For generalization, we curate the most informative and reliable transfer set from a large-scale OOD gallery set to support high-performance and identifier-invariant knowledge extraction; (2) For evasion, to prevent the transmission of suspicious verification knowledge via dark knowledge in soft labels, we propose selective knowledge transfer to filter out such knowledge and better trade off between the two objectives. Next, we elaborate on these two components." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 181, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 181, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 181, + 540 + ], + "type": "text", + "content": "3.2. Transfer Set Curation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": "Our objective is to curate an optimal subset " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " from an OOD gallery set " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " (e.g., ImageNet [9] or DataComp [12]), selecting samples that are both informative and reliable to facilitate effective knowledge transfer. While this task may resemble classical core-set selection [5, 53], it presents significantly greater challenge: (1) Our algorithm operates in a fully OOD context, where " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " do not necessarily overlap in classes, and their distributions could vary substantially; (2) " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " could include identifiers manipulated by the defender (e.g., trigger samples), and the teacher " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " trained directly on " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " is already marked. Consequently, selection based on the unreliable " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " inadvertently introduce samples from " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": " that activate verification behaviors, risking the induction of these behaviors in the surrogate student " + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 545, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "In this paper, we leverage a vision-language model (VLM) with unbiased pretrained knowledge and robust zero-shot capabilities (e.g., CLIP [49]) to facilitate reliable selection alongside the marked teacher " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ". CLIP aligns visual and textual features in a shared embedding space, where the image encoder " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_I(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " maps an image " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_I(x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", and text templates with category names (e.g., \"a photo of {class name}\") " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "c_1, \\ldots, c_K \\in \\mathcal{D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " are mapped by the text encoder " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_t(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_t(c_1), \\ldots, \\phi_t(c_K)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ". By calculating the cosine similarity between " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_I(x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\phi_t(c_i)" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": ", each sample " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "x \\in \\mathcal{G}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " can be uniquely assigned to a category " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": "However, class-name-only text templates lack foreground semantics [48], limiting the ability to distinguish target categories. While few-shot tuning allows the VLM to rapidly adapt to tasks within " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " [26], it risks being compromised by verification behaviors embedded in " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " [3]. Inspired by advances in zero-shot prompt learning [48], we adjust the VLM with unbiased category semantics from a large language model (LLM). The LLM generates a description set " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathrm{Desc}_{c_i}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " for each category " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " in " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": ", averaging descriptions in the feature space to create representation prototypes for each category, enhancing the zero-shot capabilities of the VLM. Consequently, each " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "x\\in \\mathcal{G}" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " is assigned to class " + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 204, + 553, + 360 + ], + "type": "text", + "content": " in a zero-shot fashion as follows in Equation 2:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 324, + 364, + 553, + 396 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 364, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 324, + 364, + 553, + 396 + ], + "type": "interline_equation", + "content": "t = \\arg \\max _ {c _ {i}} \\frac {1}{\\left| \\operatorname {D e s c} _ {c _ {i}} \\right|} \\sum_ {d \\in \\operatorname {D e s c} _ {c _ {i}}} (\\sin (\\phi_ {I} (x), \\phi_ {I} (d))) \\tag {2}", + "image_path": "2bf0d31ab1794c0647a83e0f4b1a7e9db778ee0224c90990e3c08f9428e9c69c.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 401, + 547, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 547, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 547, + 414 + ], + "type": "text", + "content": "where sim denotes cosine similarity shown in Equation 3:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 418, + 553, + 445 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 418, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 354, + 418, + 553, + 445 + ], + "type": "interline_equation", + "content": "\\operatorname {s i m} \\left(\\phi_ {I} (x), \\phi_ {I} (d)\\right) = \\frac {\\phi_ {I} (x) \\cdot \\phi_ {I} (d)}{\\left| \\phi_ {I} (x) \\right| \\cdot \\left| \\phi_ {I} (d) \\right|} \\tag {3}", + "image_path": "c293117ecd868c2c46d48d1d7772d9ea5ae8ebe0d44003c5e3c3173b5783c5c7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": "Although the VLM assigns all gallery set samples to categories in " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": ", it lacks a criterion for the most informative samples. To leverage the distribution from " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": " while avoiding biases from potential verification behaviors, we project all samples in " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": " into the feature space with " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\phi_I(\\cdot)" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": " and calculate the density centroids " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{Cent}_t" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": " for each class " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "t\\in \\{1,\\dots ,K\\}" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": ". These centroids, acting as distribution digests, encapsulate the conditional distribution of " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": ". Samples closer to " + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "inline_equation", + "content": "\\mathrm{Cent}_t" + }, + { + "bbox": [ + 313, + 450, + 553, + 581 + ], + "type": "text", + "content": " in feature space are prioritized for selection, a robust and straightforward criterion that is theoretically resistant to manipulation by a small proportion of outlier samples [25]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "We then establish a consensus-based ensemble for final transfer set curation: (1) The gallery set " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " is divided by the LLM-enhanced VLM into " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " bins corresponding to classes in " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Within each bin " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ", (2) samples are ordered by proximity to distribution digest " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Cent}_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "; (3) Only samples classified by the teacher " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " that are closest to " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathrm{Cent}_t" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " are selected, until the selected samples match the class count " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "|\\mathcal{D}_t|" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": ". Illustrated in Figure 1, this process ensures all predictions on the transfer set " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": " are consistent with VLM, minimizing the risk of triggering verification behaviors while closely approximating the distribution in " + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1252" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "The Feature Bank. The curation process involves projecting all gallery set samples into the feature space, which is the most time-consuming step. However, when the gallery set is potentially reused, this projection only needs to be performed once. The resulting features and indexes are saved in a feature bank for efficient reuse." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 151, + 217, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 151, + 217, + 164 + ], + "spans": [ + { + "bbox": [ + 55, + 151, + 217, + 164 + ], + "type": "text", + "content": "3.3. Selective Knowledge Transfer" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 168, + 296, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 296, + 217 + ], + "type": "text", + "content": "The transfer set " + }, + { + "bbox": [ + 55, + 168, + 296, + 217 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 168, + 296, + 217 + ], + "type": "text", + "content": " serves as an intermediary for extraction of task-oriented and identifier-invariant knowledge. This process employs the standard knowledge distillation framework [20] with the following optimization objective:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 82, + 223, + 296, + 249 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 223, + 296, + 249 + ], + "spans": [ + { + "bbox": [ + 82, + 223, + 296, + 249 + ], + "type": "interline_equation", + "content": "\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\tag {4}", + "image_path": "492fc0b0e71b59ebcc9aec8453d5b517d55c68950fedd4e841e11ec969823833.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{L}" + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": " measures prediction discrepancies, such as the Kullback-Leibler divergence, modulated by the temperature " + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": ". This process effectively transfers knowledge despite potential misalignments between the transfer set " + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": " and the training distribution [46]. However, the soft labels " + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}(x)" + }, + { + "bbox": [ + 55, + 255, + 296, + 361 + ], + "type": "text", + "content": " contain dark knowledge [11], which inadvertently transmit verification behaviors. Thus, we introduce Selective Knowledge Transfer (SKT) to filter out suspicious verification knowledge during distillation." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "content": "The exclusive and subtle verification behaviors reflect unique biases inherent to the marked dataset " + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "content": ", distinguishing it within the task distribution. Preventing the student " + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "content": " from overfitting to biases limits DOV's ability to infer connections between " + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 363, + 296, + 494 + ], + "type": "text", + "content": " through trigger-specific outputs or excessive confidence. Drawing on insights from shortcut learning [14], we mitigate predictive shortcuts during the student's learning process, reducing reliance on spurious features. Since dataset biases typically induce universal shortcut behaviors, a straightforward solution is to apply universal adversarial training (UAT) on the student:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 500, + 295, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 500, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 79, + 500, + 295, + 529 + ], + "type": "interline_equation", + "content": "\\theta_ {s} ^ {*} = \\arg \\min _ {\\theta_ {s}} \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {T}} \\mathcal {L} \\left(f _ {\\theta_ {s}} (x + \\delta), y\\right) \\tag {5}", + "image_path": "1b85c026874f28aa1dc7570b984db70d0f1a25639f622138067b9fc86c4653d1.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "text", + "content": "UAT effectively suppresses atypical predictive behaviors in " + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "inline_equation", + "content": "f_{\\theta_s}" + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "text", + "content": " induced by perturbations " + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "text", + "content": ". However, the min-max problem involves a notoriously hard bi-level optimization, requiring iterative solutions within each iteration. Furthermore, the absence of ground-truth labels in the OOD transfer set " + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 536, + 296, + 619 + ], + "type": "text", + "content": " necessitates predictions from the marked teacher as labels, introducing significant calibration errors." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "spans": [ + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "text", + "content": "Given the objective to mitigate biases specific to the copyrighted dataset " + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "text", + "content": ", and with all verification behaviors encapsulated in the teacher model, we propose using only the teacher model " + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "text", + "content": " to generate perturbations on " + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 620, + 296, + 680 + ], + "type": "text", + "content": " as a practical approximation for the inner max-problem:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 88, + 686, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 686, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 88, + 686, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\delta^ {*} = \\arg \\max _ {\\delta} \\frac {1}{n} \\sum_ {x, y \\in \\mathcal {D}} \\mathcal {L} \\left(f _ {\\theta_ {t}} (x + \\delta), y\\right) \\tag {6}", + "image_path": "8795626559892327c979b3edd236aea48ea54d95e5148b2943a793585c3f50c7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": "This strategy decouples the bi-level optimization, rendering it asynchronously solvable. By pre-generating a perturbation pool " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "\\{\\delta_i\\}_{i=1}^n" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " in an offline manner with the teacher " + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 313, + 72, + 553, + 132 + ], + "type": "text", + "content": " and reusing it in the outer min-problem during distillation, computational overhead is significantly reduced." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": "To solve Equation 6, the perturbation " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " is typically constrained by an " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "L_{p}" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " norm to preserve core image semantics [40]. However, without prior knowledge of the specific form of " + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 133, + 554, + 227 + ], + "type": "text", + "content": " associated with verification behaviors (e.g., watermark triggers), choosing an inappropriate norm undermine the efficacy of UAT in suppressing spurious features [64]. Thus, we seek to diversify the teacher-generated perturbations beyond a single norm constraint." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "content": "We solve the max-problem using mini-batch stochastic gradient descent [54], applying " + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "content": " constraints at each batch update to preserve gradient directions. After each iteration, we project the perturbation onto the norm constraint " + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "inline_equation", + "content": "(L_0, L_2, \\text{or} L_\\infty)" + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "content": " that maximizes the loss. This steepest descent projection [41] adaptively seeks the optimal norm constraint, facilitating the generation of perturbations closely resembling those employed by DOV. Moreover, extreme projections onto " + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "inline_equation", + "content": "L_{0}" + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "inline_equation", + "content": "L_{\\infty}" + }, + { + "bbox": [ + 313, + 228, + 554, + 358 + ], + "type": "text", + "content": " norms approximate a convex hull spanning multiple norm constraints, broadening the range of potential perturbations [7]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 360, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 491 + ], + "type": "text", + "content": "Ideally, perturbation should preserve core image semantics without being limited to specific norms [57]. Generative models can achieve this by producing unrestricted adversarial perturbations [76], though training such models on task-specific distributions is resource-intensive. Therefore, we propose a lightweight approach using image corruptions instead of generative models to produce unrestricted perturbations. Starting with the 15 corruptions defined in ImageNet-C [19], we employ a genetic algorithm [8, 43] to select the combination and sequence of corruptions that solves Equation 6, forming a corruption chain:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 345, + 500, + 553, + 514 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 500, + 553, + 514 + ], + "spans": [ + { + "bbox": [ + 345, + 500, + 553, + 514 + ], + "type": "interline_equation", + "content": "\\max \\mathcal {L} \\left(f _ {\\theta_ {t}} \\left(\\operatorname {C o r r u p t} _ {1: N} (x)\\right), y\\right), \\quad x, y \\in \\mathcal {D} \\tag {7}", + "image_path": "6d8bc0de203f7d26480f7bc3e6f8f1a08b15dde9597c02f4e66cf6b48171decc.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "text", + "content": "The perturbations and corruption chain generated with the teacher " + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "inline_equation", + "content": "f_{\\theta_t}" + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "text", + "content": " are collectively referred to as " + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "inline_equation", + "content": "A(\\cdot)" + }, + { + "bbox": [ + 313, + 522, + 553, + 559 + ], + "type": "text", + "content": ", leading to the final objective for Selective Knowledge Transfer:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 328, + 567, + 553, + 621 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 567, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 328, + 567, + 553, + 621 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\arg \\min _ {\\theta_ {s}} \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (x)}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right) \\tag {8} \\\\ + \\beta \\cdot \\mathcal {L} \\left(\\frac {f _ {\\theta_ {s}} (A (x))}{\\tau}, \\frac {f _ {\\theta_ {t}} (x)}{\\tau}\\right), \\quad x \\in \\mathcal {T} \\\\ \\end{array}", + "image_path": "fad06c2f7da2a29bff2f2718cfd8f70f7c95eac65d238d080085eadc02b7ab42.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "The second term promotes the student's invariance to the worst-case perturbations from the teacher while encouraging predictive divergence to reduce bias toward the teacher. The tuning factor " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " balances generalization with evasion efficacy, applied as a sampling probability to selectively introduce perturbations or corruption chain to a subset of samples, adding minimal overhead." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1253" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 86, + 552, + 258 + ], + "blocks": [ + { + "bbox": [ + 63, + 72, + 545, + 84 + ], + "lines": [ + { + "bbox": [ + 63, + 72, + 545, + 84 + ], + "spans": [ + { + "bbox": [ + 63, + 72, + 545, + 84 + ], + "type": "text", + "content": "Table 1. Escaping DOV on CIFAR-10 and Tiny ImageNet. Red indicates detection by DOV, while green indicates successful evasion." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 86, + 552, + 258 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 552, + 258 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 552, + 258 + ], + "type": "table", + "html": "
DOVCIFAR-10Tiny-Imagenet
VanillaEvasionVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)ACCVSRACC(↑)VSR(↓)
Badnets94.44 ± 0.74100.00 ± 0.0093.46 ± 0.201.36 ± 0.0769.05 ± 0.2598.98 ± 0.1166.48 ± 0.141.48 ± 0.11
UBW93.98 ± 0.2595.54 ± 0.8493.41 ± 0.191.74 ± 0.0666.37 ± 0.1275.56 ± 0.5065.34 ± 0.084.60 ± 0.11
Label-Consistent94.74 ± 0.5496.18 ± 2.7093.19 ± 0.143.74 ± 0.4069.07 ± 0.2434.99 ± 1.5366.46 ± 0.103.51 ± 0.22
Narcissus94.76 ± 0.5487.34 ± 1.8294.37 ± 0.184.59 ± 0.6468.99 ± 0.1457.49 ± 0.7866.39 ± 0.1812.33 ± 6.45
Non-Poisoning DOVACCp-valueACC(↑)p-value(↑)ACCp-valueACC(↑)p-value(↑)
Radioactive Data94.96 ± 0.473.03e-0394.07 ± 0.189.45e-0168.84 ± 0.305.33e-0366.64 ± 0.244.94e-01
ANW94.66 ± 0.311.37e-0993.91 ± 0.281.00e+0068.75 ± 0.305.57e-2866.44 ± 0.131.00e+00
Domain Watermark94.38 ± 0.531.67e-2293.90 ± 0.131.00e+0068.55 ± 0.393.91e-1666.38 ± 0.161.00e+00
Isotope94.75 ± 0.282.87e-0393.99 ± 0.492.84e-0168.97 ± 0.371.63e-0366.76 ± 0.221.66e-01
ML Auditor94.22 ± 0.177.47e-0593.45 ± 0.217.83e-0168.41 ± 0.281.55e-0665.35 ± 0.161.00e+00
Dataset Inference94.83 ± 0.471.87e-0393.97 ± 0.494.76e-0169.15 ± 0.349.27e-1466.06 ± 0.184.47e-01
MeFA94.83 ± 0.472.62e-1493.93 ± 0.551.00e+0069.15 ± 0.346.17e-2766.22 ± 0.209.99e-01
", + "image_path": "88faa8a65205d756da8fbb972754f8c7bd50b6b01ad8eede06acbb420ee48a02.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 136, + 276 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 171, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 171, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 171, + 296 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 301, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 301, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 301, + 295, + 350 + ], + "type": "text", + "content": "Following established DOV methods [17, 33, 42, 79], we adopt ResNet-18 [18] for both teacher and students. For transfer set curation, we utilize MobileCLIP [67] and GPT-40 mini [45] as lightweight VLM and LLM, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "text", + "content": "We assess backdoor watermarks with Verification Success Rate (VSR)-the probability of a trigger-induced misclassification to the target class. A VSR above " + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "inline_equation", + "content": "30\\%" + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "text", + "content": " confirms successful verification [76], indicating unauthorized training. Non-poisoning watermarks and fingerprints, collectively termed non-poisoning DOV, are evaluated via the p-value from a one-tailed T-test. The p-value represents the confidence level that the model was not trained on the copyright dataset; lower p-values indicate successful verification. The common threshold for p-values is set at 0.01 [42], equating to a " + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "inline_equation", + "content": "99\\%" + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "text", + "content": " confidence level of unauthorized training. Experiments were repeated 3 times, reporting mean " + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "inline_equation", + "content": "\\pm" + }, + { + "bbox": [ + 54, + 350, + 295, + 506 + ], + "type": "text", + "content": " std for ACC and VSR, and harmonic mean for p-values." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 516, + 241, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 241, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 241, + 529 + ], + "type": "text", + "content": "4.2. Escaping Advanced DOV Methods" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 714 + ], + "type": "text", + "content": "We evaluate Escaping DOV on natural image datasets commonly used as DOV benchmarks, including CIFAR-10 [28] and the 200-class Tiny ImageNet [29]. We further consider six datasets with significant distribution shifts from the gallery set (e.g., facial recognition and medical diagnosis), with a summary of results in Section 4.4.5 and detailed findings in Appendix C.3. For the gallery set, we use LSVRC-2012 [9]. Given overlapping samples between Tiny ImageNet and LSVRC-2012, we create a challenging setting by removing all overlapping classes from LSVRC-2012 to produce a purely OOD gallery set. The targeted DOV approaches cover advanced algorithms across all categories: (1) backdoor watermarks, including BadNets [16], UBW [32], Label-Consistent [65], and Narcissus [74], encompassing clean-label, invisible, and untargeted poisoning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 553, + 312 + ], + "type": "text", + "content": "techniques; (2) non-poisoning watermarks, including Radioactive Data [50], ANW [79], Domain Watermark [17], Isotope [68], and ML Auditor [24]; and (3) fingerprints, including Dataset Inference [42] and MeFA [35]." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 312, + 554, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 480 + ], + "type": "text", + "content": "Table 1 presents the generalization and verification metrics for the models before and after evasion attacks (the teacher and surrogate student in Escaping DOV). All DOV methods successfully identify the vanilla teacher trained directly on the copyright dataset. However, none meet their verification thresholds (VSR > 30% or p-value < 0.01) on the student produced by Escaping DOV, where VSR is near random guesses and p-values exceed 0.1 in most cases. Moreover, Escaping DOV incurs minimal impact on generalization: test accuracy on CIFAR-10 decreases by less than 1%, while on the challenging Tiny ImageNet set, with no overlapping gallery classes, accuracy drops by only 2-3%. Thus, Escaping DOV achieves a robust balance between universal evasion and preserved generalization." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 483, + 550, + 574 + ], + "blocks": [ + { + "bbox": [ + 318, + 483, + 550, + 574 + ], + "lines": [ + { + "bbox": [ + 318, + 483, + 550, + 574 + ], + "spans": [ + { + "bbox": [ + 318, + 483, + 550, + 574 + ], + "type": "image", + "image_path": "f858d9435b29ef401f6b61bff1b7372c374673963b9ea698120d10e8d6e344ca.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 316, + 578, + 550, + 589 + ], + "lines": [ + { + "bbox": [ + 316, + 578, + 550, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 550, + 589 + ], + "type": "text", + "content": "Figure 2. Loss Barrier between Teacher and Student Parameters." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": "We further observe that Escaping DOV encourages the student to diverge from the teacher's predictive behavior, forming its own mechanism. In Figure 2, we linearly interpolate between teacher and student parameters, computing training and test losses at each point. The results reveal two key findings: (1) A loss barrier appears during interpolation from teacher " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{t}" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": " to student " + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\theta_{s}" + }, + { + "bbox": [ + 313, + 594, + 554, + 713 + ], + "type": "text", + "content": ". This lack of linear connectivity suggests that the surrogate student develops a distinct prediction mechanism, bypassing the spurious features central to the teacher's verification behavior [39]. (2)" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1254" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 86, + 555, + 218 + ], + "blocks": [ + { + "bbox": [ + 68, + 72, + 541, + 84 + ], + "lines": [ + { + "bbox": [ + 68, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 68, + 72, + 541, + 84 + ], + "type": "text", + "content": "Table 2. Comparison of Escaping DOV with Evasion Attacks. Red indicates detection by DOV, green indicates successful evasion." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 86, + 555, + 218 + ], + "lines": [ + { + "bbox": [ + 59, + 86, + 555, + 218 + ], + "spans": [ + { + "bbox": [ + 59, + 86, + 555, + 218 + ], + "type": "table", + "html": "
MethodBadnetsNarcissusIsotopeDataset Inference
ACC(↑)VSR(↓)ACC(↑)VSR(↓)ACC(↑)p-value(↑)ACC(↑)p-value(↑)
Fine-pruning86.21 ± 0.1699.37 ± 0.2786.58 ± 0.1543.75 ± 14.4186.78 ± 0.120.053286.88 ± 0.480.0095
Meta-Sift88.68 ± 0.4599.70 ± 0.3288.64 ± 0.6249.40 ± 8.2688.56 ± 0.470.026488.07 ± 0.210.0963
Differential Privacy90.49 ± 1.2076.70 ± 4.6287.70 ± 0.6872.77 ± 3.9188.69 ± 2.610.067889.64 ± 4.860.0440
I-BAU89.09 ± 0.2514.80 ± 0.4688.76 ± 1.9472.74 ± 2.1787.07 ± 2.590.155289.93 ± 0.680.0355
ZIP82.44 ± 0.1079.88 ± 0.3983.51 ± 0.3753.21 ± 4.4782.97 ± 0.420.009883.21 ± 0.900.1050
NAD89.86 ± 0.373.11 ± 0.5191.04 ± 0.3675.41 ± 5.1291.77 ± 0.450.096192.12 ± 0.340.0014
BCU92.75 ± 0.051.56 ± 0.2592.33 ± 0.1161.72 ± 6.4392.77 ± 0.360.162293.09 ± 0.180.0773
ABD84.94 ± 3.047.07 ± 3.1585.42 ± 2.1022.10 ± 1.1586.29 ± 3.460.006684.87 ± 3.020.3696
IPRemoval83.96 ± 0.238.09 ± 0.5085.06 ± 0.1155.97 ± 2.1484.64 ± 0.060.191685.28 ± 0.050.1322
Escaping DOV (Ours)93.46 ± 0.201.36 ± 0.0794.37 ± 0.184.59 ± 0.6493.99 ± 0.490.284593.97 ± 0.490.4759
", + "image_path": "4ce117b5151f7e7a8159e353331ae2cafffe9284d2fce273cfd373555ed8133e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "spans": [ + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "text", + "content": "For the teacher, test loss is over " + }, + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "inline_equation", + "content": "10^{3}" + }, + { + "bbox": [ + 55, + 224, + 295, + 298 + ], + "type": "text", + "content": " times the training loss, allowing fingerprints to easily infer dataset association. In contrast, the student's test loss is typically under twice the training loss, falling within the variance across subsets and making it far less distinguishable. Thus, Escaping DOV significantly suppresses both watermarks and fingerprints." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 306, + 269, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 306, + 269, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 306, + 269, + 319 + ], + "type": "text", + "content": "4.3. Comparison with SOTA Evasion Attacks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "text", + "content": "We compare with SOTA evasion attacks on CIFAR-10 across the seminal Badnets watermark and three most resistant DOVs from each category—Narcissus, Isotope, and Dataset Inference (see Table 1). Results for all DOV methods are provided in Appendix C.2. Evasion attacks include Fine-Pruning [36], stronger combination of fine-tuning and pruning widely used in DOV evaluations, along with SOTA backdoor defenses across four stages: data sanitization with Meta-Sift [73], robust training with Differential Privacy [4], backdoor unlearning with I-BAU [72], and input purification with ZIP [56]. We also compare with methods that use knowledge distillation to mitigate backdoors, such as NAD [31], BCU [47] and ABD [22], as well as IPRemoval [78], which aims to eliminate model watermarks. Results in Table 2 and Appendix C.2 show that Escaping DOV consistently surpasses all comparison methods in both generalization and evasion, and is the only approach that fully evades all DOV methods. For instance, only ABD and Escaping DOV reduce VSR below " + }, + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "inline_equation", + "content": "40\\%" + }, + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "text", + "content": " on the clean-label and invisible Narcissus watermark; however, ABD fails to evade Isotope, while Escaping DOV achieves nearly " + }, + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 56, + 324, + 296, + 623 + ], + "type": "text", + "content": " higher accuracy with superior evasion performance. We provide a rigorous analysis in Appendix D, explaining why SOTA evasion attacks, particularly those relying on distillation, are less effective than our Escaping DOV." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 632, + 200, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 632, + 200, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 632, + 200, + 643 + ], + "type": "text", + "content": "4.4. Ablations and Discussions" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 650, + 178, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 650, + 178, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 650, + 178, + 661 + ], + "type": "text", + "content": "4.4.1. Transfer Set Curation" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 714 + ], + "type": "text", + "content": "We begin with a qualitative analysis of the transfer set. Figure 3a visualizes the feature space for CIFAR-10 samples as well as selected and unselected ones from the transfer set, on a clean ResNet-18. The selected samples are notably" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 297 + ], + "type": "text", + "content": "closer to original CIFAR-10 samples. Figure 3b illustrates the feature space of a Badnets-marked model, where trigger samples form distinct clusters with almost no transfer set samples nearby. The exclusive and subtle verification behaviors are rarely activated by OOD samples, and the curation process effectively excludes suspicious samples." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 298, + 435, + 392 + ], + "blocks": [ + { + "bbox": [ + 317, + 298, + 435, + 392 + ], + "lines": [ + { + "bbox": [ + 317, + 298, + 435, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 435, + 392 + ], + "type": "image", + "image_path": "30aa5cca0a72ec5148e512446ea08bc55c86c7d8996762fa937e449053f6e80a.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 396, + 528, + 408 + ], + "lines": [ + { + "bbox": [ + 339, + 396, + 528, + 408 + ], + "spans": [ + { + "bbox": [ + 339, + 396, + 528, + 408 + ], + "type": "text", + "content": "Figure 3. Sample Distribution in the Feature Space." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 436, + 298, + 553, + 392 + ], + "blocks": [ + { + "bbox": [ + 436, + 298, + 553, + 392 + ], + "lines": [ + { + "bbox": [ + 436, + 298, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 436, + 298, + 553, + 392 + ], + "type": "image", + "image_path": "91d664e9df5cabf3b5cbb3ccee4961cf9a966c9c3c63fab22cd7e739105b89ef.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 413, + 555, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 521 + ], + "type": "text", + "content": "Quantitative analysis follows, using Optimal Transport Dataset Distance (OTDD) [2] to assess the distance from the target dataset to each proxy distribution, as shown in Table 3. The curated transfer set is significantly closer to the target distribution than random sampling. Further, we compare the curated transfer set with random and unrelative samples on CIFAR-10 with Narcissus. Table 4 shows that transfer set curation not only improves generalization but also suppresses DOV by selecting more reliable samples." + } + ] + } + ], + "index": 12 + }, + { + "type": "table", + "bbox": [ + 317, + 535, + 554, + 572 + ], + "blocks": [ + { + "bbox": [ + 331, + 523, + 537, + 534 + ], + "lines": [ + { + "bbox": [ + 331, + 523, + 537, + 534 + ], + "spans": [ + { + "bbox": [ + 331, + 523, + 537, + 534 + ], + "type": "text", + "content": "Table 3. Optimal Transport Distances Between Datasets." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 535, + 554, + 572 + ], + "lines": [ + { + "bbox": [ + 317, + 535, + 554, + 572 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 554, + 572 + ], + "type": "table", + "html": "
DistanceselfLSVRC(curation)LSVRC(random)GTSRB [58]
CIFAR-100.7710.9371.1161.129
Tiny-Imagenet0.8550.9880.9931.066
", + "image_path": "d5e41254984145be5f7018af7fcb72db1497dc89d546c6f78c504b58e399afd4.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 317, + 589, + 553, + 630 + ], + "blocks": [ + { + "bbox": [ + 335, + 576, + 533, + 587 + ], + "lines": [ + { + "bbox": [ + 335, + 576, + 533, + 587 + ], + "spans": [ + { + "bbox": [ + 335, + 576, + 533, + 587 + ], + "type": "text", + "content": "Table 4. Escaping DOV across different Transfer Sets." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 589, + 553, + 630 + ], + "lines": [ + { + "bbox": [ + 317, + 589, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 317, + 589, + 553, + 630 + ], + "type": "table", + "html": "
Transfer SetGTSRB [58]LSVRC(random)LSVRC(curation)
ACC(↑)66.80 ± 0.4191.63 ± 0.4994.37 ± 0.18
VSR(↓)8.96 ± 1.637.87 ± 0.524.59 ± 0.64
", + "image_path": "d30284928ea4cdd9ae71e8bf3a508c89c346b577b43f72a2a664ba83b12603e9.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_body" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 639, + 470, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 470, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 470, + 651 + ], + "type": "text", + "content": "4.4.2. Selective Knowledge Transfer" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Figure 4 illustrates the impact of Selective Knowledge Transfer (SKT) on the most resistant Narcissus watermark across distillation temperatures " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ". As " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " increases, omitting SKT causes a substantial increase in VSR, while SKT consistently suppresses verification behavior across all temper" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1255" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "atures. Additionally, VSR escalates sharply at higher temperatures, while test accuracy remains nearly unchanged. We hypothesis that task and verification knowledge possess distinct \"boiling points\", naturally separating at lower temperatures. Thus, we adopt " + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\tau = 1" + }, + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": " as a reliable default." + } + ] + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 134, + 294, + 215 + ], + "blocks": [ + { + "bbox": [ + 57, + 134, + 294, + 215 + ], + "lines": [ + { + "bbox": [ + 57, + 134, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 57, + 134, + 294, + 215 + ], + "type": "image", + "image_path": "faf7efc6e4773da8f48a83aacd5b94d43ca8e3191607e62a79c70b3a612db076.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 75, + 217, + 275, + 227 + ], + "lines": [ + { + "bbox": [ + 75, + 217, + 275, + 227 + ], + "spans": [ + { + "bbox": [ + 75, + 217, + 275, + 227 + ], + "type": "text", + "content": "Figure 4. Impact of SKT and Distillation Temperature." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "content": "We further compare the offline perturbation pool in SKT with direct UAT on the student. While UAT lowers VSR to " + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "inline_equation", + "content": "2.07\\%" + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "content": ", it incurs a " + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "inline_equation", + "content": "3.54\\%" + }, + { + "bbox": [ + 55, + 232, + 296, + 305 + ], + "type": "text", + "content": " accuracy reduction and more than triples computational cost. As SKT balances evasion and generalization, it serves as a lightweight approximation. Detailed time complexity is analysed in Appendix C.8." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 315, + 204, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 315, + 204, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 204, + 327 + ], + "type": "text", + "content": "4.4.3. Escaping Adaptive Defenses" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "spans": [ + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "text", + "content": "We assume a knowledgeable defender employing adaptive verification, specifically a semantic backdoor [60] and an anti-distillation backdoor [13]. For the semantic backdoor, we assume the defender knows the gallery set and selects 500 \"tench\" samples from LSVRC-2012, relabeling them as \"plane\" and inserting them into CIFAR-10 to test their classification probability as \"plane\" (VSR). Since the teacher memorizes this verification behavior, using it for sample selection allows triggers enter the transfer set. For instance, DFND [13], an entropy-based selection method, produces a student with VSR of " + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "inline_equation", + "content": "70.2\\%" + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "text", + "content": " by including 254 \"tench\" samples in the transfer set. In contrast, our curated transfer set contains fewer than 10 \"tench\" samples, reducing VSR to only " + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "inline_equation", + "content": "17.4\\%" + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "text", + "content": ". Even if the defender knows the gallery set, Transfer Set Curation still selects reliable samples. We also evaluate the entropy of teacher annotations: DFND's entropy is 2.90, while ours reaches 3.32, near the theoretical optimum " + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "inline_equation", + "content": "\\log_2(10)" + }, + { + "bbox": [ + 55, + 331, + 296, + 547 + ], + "type": "text", + "content": ", indicating better class balance." + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 58, + 562, + 296, + 628 + ], + "blocks": [ + { + "bbox": [ + 61, + 549, + 289, + 560 + ], + "lines": [ + { + "bbox": [ + 61, + 549, + 289, + 560 + ], + "spans": [ + { + "bbox": [ + 61, + 549, + 289, + 560 + ], + "type": "text", + "content": "Table 5. Escaping DOV against the Adaptive ADB Watermark." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 562, + 296, + 628 + ], + "lines": [ + { + "bbox": [ + 58, + 562, + 296, + 628 + ], + "spans": [ + { + "bbox": [ + 58, + 562, + 296, + 628 + ], + "type": "table", + "html": "
μVanillaEvasionIndependent
ACCVSRACCVSRACCVSR
0.191.59%100.00%90.14%94.42%95.13%92.89%
190.90%100.00%89.23%78.48%95.13%71.74%
1091.26%12.88%88.66%14.21%95.13%10.01%
", + "image_path": "6eae561c53ea74d1a26b27280048a4169ef13a3b0fbec2f50a0ad04fdeebf195.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "The anti-distillation backdoor (ADB) [13] is designed to transfer to student during distillation. As we employ distillation for knowledge transfer, we test ADB under a hypothetical scenario where the defender controls the teacher's training process to optimize the trigger. Table 5 reports the VSR across teacher, student, and unmarked models. A smaller " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " allows larger perturbations, turning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "the trigger into an adversarial perturbation that induces false positives on unmarked models. Despite the defender's unrealistic control over the teacher, students' VSR remains closer to that of an unmarked model. As " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " increases, ADB becomes ineffective. Thus, even under defender-controlled teacher training, Escaping DOV remains a potent threat." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 153, + 454, + 164 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 153, + 454, + 164 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 454, + 164 + ], + "type": "text", + "content": "4.4.4. Enhanced Evasion Tactics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 168, + 554, + 253 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 168, + 554, + 253 + ], + "spans": [ + { + "bbox": [ + 313, + 168, + 554, + 253 + ], + "type": "text", + "content": "Escaping DOV involves only post-training adjustments and operates orthogonally to pre-training, in-training and test-time strategies, all of which can collectively enhance evasion. As shown in Table 6, replacing data augmentation with mixup [75] during teacher training—an approach that mitigates backdoors and overfitting—further boosts evasion efficacy in conjunction with Escaping DOV on CIFAR-10." + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 268, + 554, + 353 + ], + "blocks": [ + { + "bbox": [ + 337, + 255, + 530, + 266 + ], + "lines": [ + { + "bbox": [ + 337, + 255, + 530, + 266 + ], + "spans": [ + { + "bbox": [ + 337, + 255, + 530, + 266 + ], + "type": "text", + "content": "Table 6. Mixup Teacher Training for Escaping DOV." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 268, + 554, + 353 + ], + "lines": [ + { + "bbox": [ + 317, + 268, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 554, + 353 + ], + "type": "table", + "html": "
DOVVanillaEvasion
Poisoning DOVACCVSRACC(↑)VSR(↓)
Narcissus94.04%72.47%93.03%3.66%
Non-poisoning DOVACCp-valueACC(↑)p-value(↑)
Isotope94.03%0.296492.71%0.7013
Dataset Inference94.60%0.093792.78%0.5196
", + "image_path": "6ae04f2c307e49353b6feff8ac20bb1e2fbc1348f46ec23f71c9717f2136ec89.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 363, + 515, + 374 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 515, + 374 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 515, + 374 + ], + "type": "text", + "content": "4.4.5. Case Study on Domain-Specific Datasets" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 379, + 554, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 554, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 554, + 475 + ], + "type": "text", + "content": "Datasets in vertical industries often require stronger protection than natural images due to collection challenges and privacy concerns. We conduct a case study on RAFDB (facial emotion) [30] and OrganCMNIST (medical imaging) [71] in Table 7, with detailed results in Appendix C.3. Escaping DOV maintains generalization and efficacy despite unique distribution, while a larger domain-related gallery and VLM could further improve performance [34, 38]." + } + ] + } + ], + "index": 15 + }, + { + "type": "table", + "bbox": [ + 317, + 489, + 554, + 570 + ], + "blocks": [ + { + "bbox": [ + 335, + 477, + 533, + 488 + ], + "lines": [ + { + "bbox": [ + 335, + 477, + 533, + 488 + ], + "spans": [ + { + "bbox": [ + 335, + 477, + 533, + 488 + ], + "type": "text", + "content": "Table 7. Escaping DOV on Domain-specific Datasets." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 489, + 554, + 570 + ], + "lines": [ + { + "bbox": [ + 317, + 489, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 489, + 554, + 570 + ], + "type": "table", + "html": "
SettingsVanillaEvasion
DOVDataSetACCVSRACC(↑)VSR(↓)
BadnetsRAFDB85.21 ± 0.3399.95 ± 0.0280.53 ± 0.422.19 ± 0.06
OrganC95.32 ± 0.0699.99 ± 0.0190.57 ± 0.700.83 ± 0.15
DOVDataSetACCp-valueACC(↑)p-value(↑)
IsotopeRAFDB82.46 ± 0.591.50e-0680.37 ± 0.131.05e-01
OrganC94.63 ± 0.192.65e-0388.96 ± 0.565.04e-01
", + "image_path": "82b0e4ffd1eb072bc77d53599eed70c6e74a8d0e260a262a5a1ebc67c6a242a8.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "spans": [ + { + "bbox": [ + 313, + 585, + 388, + 597 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "In this paper, we reveal that Data Ownership Verification, widely used to prevent unauthorized model training, is vulnerable to strong evasion attacks. We introduce Escaping DOV, which leverages transfer set curation and selective knowledge transfer to achieve superior generalization and universal evasion capabilities. We propose that Escaping DOV serve as a evaluation framework to help establish truly reliable DOV methods, with plans for extensive cross-modal experiments to explore broader societal benefits." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1256" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 90, + 296, + 113 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 90, + 296, + 113 + ], + "spans": [ + { + "bbox": [ + 56, + 90, + 296, + 113 + ], + "type": "text", + "content": "The work was supported by the National Natural Science Foundation of China under Grant 62271307 and 61771310." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 135, + 115, + 148 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 155, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 155, + 296, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 155, + 296, + 209 + ], + "spans": [ + { + "bbox": [ + 61, + 155, + 296, + 209 + ], + "type": "text", + "content": "[1] Yossi Adi, Carsten Baum, Moustapha Cisse, Benny Pinkas, and Joseph Keshet. Turning your weakness into a strength: Watermarking deep neural networks by backdooring. In 27th USENIX security symposium (USENIX Security 18), pages 1615-1631, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "spans": [ + { + "bbox": [ + 61, + 210, + 296, + 243 + ], + "type": "text", + "content": "[2] David Alvarez-Melis and Nicolo Fusi. Geometric dataset distances via optimal transport. Advances in Neural Information Processing Systems, 33:21428-21439, 2020. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 244, + 296, + 299 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 244, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 62, + 244, + 296, + 299 + ], + "type": "text", + "content": "[3] Jiawang Bai, Kuofeng Gao, Shaobo Min, Shu-Tao Xia, Zhifeng Li, and Wei Liu. Badclip: Trigger-aware prompt learning for backdoor attacks on clip. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24239-24250, 2024. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 300, + 295, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 295, + 343 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 295, + 343 + ], + "type": "text", + "content": "[4] Zhiqi Bu, Yu-Xiang Wang, Sheng Zha, and George Karypis. Automatic clipping: Differentially private deep learning made easier and stronger. Advances in Neural Information Processing Systems, 36, 2023. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 295, + 399 + ], + "type": "text", + "content": "[5] Hanting Chen, Tianyu Guo, Chang Xu, Wenshuo Li, Chunjing Xu, Chao Xu, and Yunhe Wang. Learning student networks in the wild. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6428-6437, 2021. 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 400, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 400, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 400, + 295, + 443 + ], + "type": "text", + "content": "[6] Ingemar J Cox, Joe Kilian, F Thomson Leighton, and Talal Shamoon. Secure spread spectrum watermarking for multimedia. IEEE transactions on image processing, 6(12):1673-1687, 1997. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "text", + "content": "[7] Francesco Croce and Matthias Hein. Adversarial robustness against multiple and single " + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "text", + "content": "-" + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 62, + 446, + 295, + 489 + ], + "type": "text", + "content": "-threat models via quick finetuning of robust classifiers. In International Conference on Machine Learning, pages 4436-4454. PMLR, 2022. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 533 + ], + "type": "text", + "content": "[8] Kalyanmoy Deb, Amrit Pratap, Sameer Agarwal, and TAMT Meyarivan. A fast and elitist multiobjective genetic algorithm: Nsga-ii. IEEE transactions on evolutionary computation, 6(2):182-197, 2002. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 535, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 535, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 535, + 295, + 579 + ], + "type": "text", + "content": "[9] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition, pages 248-255. IEEE, 2009. 1, 4, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 294, + 624 + ], + "type": "text", + "content": "[10] European Parliament and Council of the European Union. General data protection regulation (gdpr): Regulation (eu) 2016/679 on data protection and free movement, 2016. Available at https://gdpr-info.eu/.1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 668 + ], + "type": "text", + "content": "[11] Tommaso Furlanello, Zachary Lipton, Michael Tschannen, Laurent Itti, and Anima Anandkumar. Born again neural networks. In International conference on machine learning, pages 1607-1616. PMLR, 2018. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[12] Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Datacomp: In search of the next generation of multimodal" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 94 + ], + "type": "text", + "content": "datasets. Advances in Neural Information Processing Systems, 36, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 150 + ], + "type": "text", + "content": "[13] Yunjie Ge, Qian Wang, Baolin Zheng, Xinlu Zhuang, Qi Li, Chao Shen, and Cong Wang. Anti-distillation backdoor attacks: Backdoors can really survive in knowledge distillation. In Proceedings of the 29th ACM International Conference on Multimedia, pages 826-834, 2021. 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[14] Robert Geirhos, Jorn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nature Machine Intelligence, 2(11):665-673, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[15] Micah Goldblum, Dimitris Tsipras, Chulin Xie, Xinyun Chen, Avi Schwarzschild, Dawn Song, Aleksander Madry, Bo Li, and Tom Goldstein. Dataset security for machine learning: Data poisoning, backdoor attacks, and defenses. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2):1563-1580, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 315 + ], + "type": "text", + "content": "[16] T Gu, B Dolan-Gavitt, and S BadNets. Identifying vulnerabilities in the machine learning model supply chain. In Proceedings of the Neural Information Processing Symposium Workshop Mach. Learning Security (MLSec), pages 1-5, 2017. 2, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 370 + ], + "type": "text", + "content": "[17] Junfeng Guo, Yiming Li, Lixu Wang, Shu-Tao Xia, Heng Huang, Cong Liu, and Bo Li. Domain watermark: Effective and harmless dataset copyright protection is closed at hand. Advances in Neural Information Processing Systems, 36, 2024. 2, 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 415 + ], + "type": "text", + "content": "[18] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 458 + ], + "type": "text", + "content": "[19] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In International Conference on Learning Representations, 2018. 5" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 491 + ], + "type": "text", + "content": "[20] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 536 + ], + "type": "text", + "content": "[21] Mathew Hogan, Yan Michalevsky, and Saba Eskandarian. Dbreach: Stealing from databases using compression side channels. In 2023 IEEE Symposium on Security and Privacy (SP), pages 182-198. IEEE, 2023. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[22] Junyuan Hong, Yi Zeng, Shuyang Yu, Lingjuan Lyu, Ruoxi Jia, and Jiayu Zhou. Revisiting data-free knowledge distillation with poisoned teachers. In International Conference on Machine Learning, pages 13199-13212. PMLR, 2023. 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 635 + ], + "type": "text", + "content": "[23] Hanxun Huang, Xingjun Ma, Sarah Monazam Erfani, James Bailey, and Yisen Wang. Unlearnable examples: Making personal data unexploitable. In Proceedings of the International Conference on Learning Representations (ICLR), 2021. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 689 + ], + "type": "text", + "content": "[24] Zonghao Huang, Neil Zhenqiang Gong, and Michael K Reiter. A general framework for data-use auditing of ml models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, pages 1300-1314, 2024. 2, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[25] Jinyuan Jia, Yupei Liu, Xiaoyu Cao, and Neil Zhenqiang Gong. Certified robustness of nearest neighbors against" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1257" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 106 + ], + "type": "text", + "content": "data poisoning and backdoor attacks. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 9575-9583, 2022. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 152 + ], + "type": "text", + "content": "[26] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[27] Xiao Jin, Pin-Yu Chen, Chia-Yi Hsu, Chia-Mu Yu, and Tianyi Chen. Cafe: Catastrophic data leakage in vertical federated learning. Advances in Neural Information Processing Systems, 34:994-1006, 2021. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 220 + ], + "type": "text", + "content": "[28] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 220, + 294, + 243 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 243 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 243 + ], + "type": "text", + "content": "[29] Ya Le and Xuan Yang. Tiny imagenet visual recognition challenge. CS 231N, 7(7):3, 2015. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[30] Shan Li, Weihong Deng, and JunPing Du. Reliable crowdsourcing and deep locality-preserving learning for expression recognition in the wild. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2852-2861, 2017. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 343 + ], + "type": "text", + "content": "[31] Yige Li, Xixiang Lyu, Nodens Koren, Lingjuan Lyu, Bo Li, and Xingjun Ma. Neural attention distillation: Erasing back-door triggers from deep neural networks. In International Conference on Learning Representations, 2021. 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 344, + 294, + 399 + ], + "type": "text", + "content": "[32] Yiming Li, Yang Bai, Yong Jiang, Yong Yang, Shu-Tao Xia, and Bo Li. Untargeted backdoor watermark: Towards harmless and stealthy dataset copyright protection. Advances in Neural Information Processing Systems, 35:13238-13250, 2022. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 444 + ], + "type": "text", + "content": "[33] Yiming Li, Mingyan Zhu, Xue Yang, Yong Jiang, Tao Wei, and Shu-Tao Xia. Black-box dataset ownership verification via backdoor watermarking. IEEE Transactions on Information Forensics and Security, 18:2318-2332, 2023. 1, 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[34] Yudong Li, Xianxu Hou, Zheng Dezhi, Linlin Shen, and Zhe Zhao. Flip-80m: 80 million visual-linguistic pairs for facial language-image pre-training. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 58–67, 2024. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[35] Gaoyang Liu, Tianlong Xu, Xiaoqiang Ma, and Chen Wang. Your model trains on my data? protecting intellectual property of training data via membership fingerprint authentication. IEEE Transactions on Information Forensics and Security, 17:1024-1037, 2022. 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[36] Kang Liu, Brendan Dolan-Gavitt, and Siddharth Garg. Finepruning: Defending against backdooring attacks on deep neural networks. In International symposium on research in attacks, intrusions, and defenses, pages 273-294. Springer, 2018. 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 657 + ], + "type": "text", + "content": "[37] Ziwei Liu, Ping Luo, Xiaogang Wang, and Xiaou Tang. Deep learning face attributes in the wild. In Proceedings of the IEEE international conference on computer vision, pages 3730-3738, 2015. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[38] Ming Y Lu, Bowen Chen, Drew FK Williamson, Richard J Chen, Ivy Liang, Tong Ding, Guillaume Jaume, Igor Odintsov, Long Phi Le, Georg Gerber, et al. A visual-language foundation model for computational pathology. Nature Medicine, 30(3):863-874, 2024. 8" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 554, + 116 + ], + "type": "text", + "content": "[39] Ekdeep Singh Lubana, Eric J Bigelow, Robert P Dick, David Krueger, and Hidenori Tanaka. Mechanistic mode connectivity. In International Conference on Machine Learning, pages 22965-23004. PMLR, 2023. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "text", + "content": "[40] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In Proceedings of the International Conference on Learning Representations (ICLR), 2018. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 554, + 215 + ], + "type": "text", + "content": "[41] Pratyush Maini, Eric Wong, and Zico Kolter. Adversarial robustness against the union of multiple perturbation models. In International Conference on Machine Learning, pages 6640-6650. PMLR, 2020. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 554, + 259 + ], + "type": "text", + "content": "[42] Pratyush Maini, Mohammad Yaghini, and Nicolas Papernot. Dataset inference: Ownership resolution in machine learning. In International Conference on Learning Representations, 2021. 2, 3, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "text", + "content": "[43] Xiaofeng Mao, Yuefeng Chen, Shuhui Wang, Hang Su, Yuan He, and Hui Xue. Composite adversarial attacks. In Proceedings of the AAAI conference on artificial intelligence, pages 8884-8892, 2021. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 304, + 554, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 554, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 554, + 370 + ], + "type": "text", + "content": "[44] NBC News. Facial recognition's 'dirty little secret': Millions of online photos scraped without consent, 2019. Available at https://www.nbcnews.com/news/all/facial-recognition-s-dirty-little-secret-millionphotos-online-photos-scraped-n981921.1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 371, + 554, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 554, + 415 + ], + "type": "text", + "content": "[45] OpenAI. Gpt-4o mini: Advancing cost-efficient intelligence, 2024. Available at https://openai.com/index/gpt-4o-mini-advancing-cost-efficient-intelligence/.6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 415, + 554, + 459 + ], + "type": "text", + "content": "[46] Tribhuvanesh Orekondy, Bernt Schiele, and Mario Fritz. Knockoff nets: Stealing functionality of black-box models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4954-4963, 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 460, + 555, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 555, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 555, + 503 + ], + "type": "text", + "content": "[47] Lu Pang, Tianlong Sun, Huan Ling, and Changyou Chen. Backdoor cleansing with unlabeled data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12218-12227, 2023. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 554, + 558 + ], + "type": "text", + "content": "[48] Sarah Pratt, Ian Covert, Rosanne Liu, and Ali Farhadi. What does a platypus look like? generating customized prompts for zero-shot image classification. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15691-15701, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 559, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 554, + 624 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 668 + ], + "type": "text", + "content": "[50] Alexandre Sablayrolles, Matthijs Douze, Cordelia Schmid, and Hervé Jégou. Radioactive data: tracing through training. In International Conference on Machine Learning, pages 8326-8335. PMLR, 2020. 2, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 554, + 712 + ], + "type": "text", + "content": "[51] Pedro Sandoval-Segura, Vasu Singla, Jonas Geiping, Micah Goldblum, and Tom Goldstein. What can we learn from unlearnable datasets? Advances in Neural Information Processing Systems, 36, 2023. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1258" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 139 + ], + "type": "text", + "content": "[52] Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, et al. Laion-5b: An open large-scale dataset for training next generation image-text models. Advances in Neural Information Processing Systems, 35:25278-25294, 2022. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 295, + 173 + ], + "type": "text", + "content": "[53] Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. In International Conference on Learning Representations, 2018. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 218 + ], + "type": "text", + "content": "[54] Ali Shafahi, Mahyar Najibi, Zheng Xu, John Dickerson, Larry S Davis, and Tom Goldstein. Universal adversarial training. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5636-5643, 2020. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 220, + 294, + 274 + ], + "type": "text", + "content": "[55] Shuo Shao, Yiming Li, Mengren Zheng, Zhiyang Hu, Yukun Chen, Boheng Li, Yu He, Junfeng Guo, Tianwei Zhang, Dacheng Tao, et al. Databench: Evaluating dataset auditing in deep learning from an adversarial perspective. arXiv preprint arXiv:2507.05622, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[56] Yucheng Shi, Mengnan Du, Xuansheng Wu, Zihan Guan, Jin Sun, and Ninghao Liu. Black-box backdoor defense via zero-shot image purification. Advances in Neural Information Processing Systems, 36:57336-57366, 2023. 3, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "text", + "content": "[57] Yang Song, Rui Shu, Nate Kushman, and Stefano Ermon. Constructing unrestricted adversarial examples with generative models. Advances in neural information processing systems, 31, 2018. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 419 + ], + "type": "text", + "content": "[58] Johannes Stallkamp, Marc Schlipsing, Jan Salmen, and Christian Igel. The german traffic sign recognition benchmark: a multi-class classification competition. In *The 2011 international joint conference on neural networks*, pages 1453–1460. IEEE, 2011. 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 465 + ], + "type": "text", + "content": "[59] Salvatore J Stolfo, Malek Ben Salem, and Angelos D Keromytis. Fog computing: Mitigating insider data theft attacks in the cloud. In 2012 IEEE symposium on security and privacy workshops, pages 125-128. IEEE, 2012. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 520 + ], + "type": "text", + "content": "[60] Bing Sun, Jun Sun, Wayne Koh, and Jie Shi. Neural network semantic backdoor detection and mitigation: A causality-based approach. In Proceedings of the 33rd USENIX Security Symposium. USENIX Association, San Francisco, CA, USA, 2024. 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 522, + 294, + 566 + ], + "type": "text", + "content": "[61] Sebastian Szyller and N Asokan. Conflicting interactions among protection mechanisms for machine learning models. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 15179-15187, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 600 + ], + "type": "text", + "content": "[62] Sebastian Szyller, Rui Zhang, Jian Liu, and N. Asokan. On the robustness of dataset inference. Transactions on Machine Learning Research, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 645 + ], + "type": "text", + "content": "[63] Ruixiang Tang, Qizhang Feng, Ninghao Liu, Fan Yang, and Xia Hu. Did you train on my dataset? towards public dataset protection with cleanlabel backdoor watermarking. ACM SIGKDD Explorations Newsletter, 25(1):43-53, 2023. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 679 + ], + "type": "text", + "content": "[64] Florian Tramer and Dan Boneh. Adversarial training and robustness for multiple perturbations. Advances in neural information processing systems, 32, 2019. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[65] Alexander Turner, Dimitris Tsipras, and Aleksander Madry. Label-consistent backdoor attacks. arXiv preprint arXiv:1912.02771, 2019. 2, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 139 + ], + "type": "text", + "content": "[66] Rolf Van Wegberg, Samaneh Tajalizadehkhoob, Kyle Soska, Ugur Akyazi, Carlos Hernandez Ganan, Bram Klievink, Nicolas Christin, and Michel Van Eeten. Plug and prey? measuring the commoditization of cybercrime via online anonymous markets. In 27th USENIX security symposium (USENIX security 18), pages 1009-1026, 2018. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 205 + ], + "type": "text", + "content": "[67] Pavan Kumar Anasosalu Vasu, Hadi Pouransari, Fartash Faghri, Raviteja Vemulapalli, and Oncel Tuzel. Mobile-clip: Fast image-text models through multi-modal reinforced training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15963–15974, 2024. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 553, + 251 + ], + "type": "text", + "content": "[68] Emily Wenger, Xiuyu Li, Ben Y Zhao, and Vitaly Shmatikov. Data isotopes for data provenance in dnns. Proceedings on Privacy Enhancing Technologies, 2024(1), 2024. 2, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 553, + 285 + ], + "type": "text", + "content": "[69] Wikipedia contributors. Cambridge analytica — wikipedia, the free encyclopedia. https://en.wikipedia.org/wiki/Cambridge_Analytica, 2024.1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 330 + ], + "type": "text", + "content": "[70] Nuo Xu, Qi Liu, Tao Liu, Zihao Liu, Xiaochen Guo, and Wujie Wen. Stealing your data from compressed machine learning models. In 2020 57th ACM/IEEE Design Automation Conference (DAC), pages 1-6. IEEE, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "text", + "content": "[71] Jiancheng Yang, Rui Shi, Donglai Wei, Zequan Liu, Lin Zhao, Bilian Ke, Hanspeter Pfister, and Bingbing Ni. Medmnist v2-a large-scale lightweight benchmark for 2d and 3d biomedical image classification. Scientific Data, 10(1):41, 2023. 1, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 431 + ], + "type": "text", + "content": "[72] Yi Zeng, Si Chen, Won Park, Zhuoqing Mao, Ming Jin, and Ruoxi Jia. Adversarial unlearning of backdoors via implicit hypergradient. In International Conference on Learning Representations, 2022. 3, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 553, + 498 + ], + "type": "text", + "content": "[73] Yi Zeng, Minzhou Pan, Himanshu Jahagirdar, Ming Jin, Lingjuan Lyu, and Ruoxi Jia. Meta-Sift: How to sift out a clean subset in the presence of data poisoning? In 32nd USENIX Security Symposium (USENIX Security 23), pages 1667-1684, Anaheim, CA, 2023. USENIX Association. 2, 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 555 + ], + "type": "text", + "content": "[74] Yi Zeng, Minzhou Pan, Hoang Anh Just, Lingjuan Lyu, Meikang Qiu, and Ruoxi Jia. Narcissus: A practical clean-label backdoor attack with limited information. In Proceedings of the 2023 ACM SIGSAC Conference on Computer and Communications Security, pages 771-785, 2023. 2, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 556, + 553, + 599 + ], + "type": "text", + "content": "[75] Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, and David Lopez-Paz. mixup: Beyond empirical risk minimization. In International Conference on Learning Representations, 2018. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 553, + 656 + ], + "type": "text", + "content": "[76] Hongyu Zhu, Sichu Liang, Wentao Hu, Li Fangqi, Ju Jia, and Shi-Lin Wang. Reliable model watermarking: Defending against theft without compromising on evasion. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 10124-10133, 2024. 2, 5, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 690 + ], + "type": "text", + "content": "[77] Hongyu Zhu, Sichu Liang, Wentao Hu, et al. Stealing knowledge from auditable datasets. In Proceedings of the European Conference on Artificial Intelligence (ECAI), 2025. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 712 + ], + "type": "text", + "content": "[78] Wei Zong, Yang-Wai Chow, Willy Susilo, Joonsang Baek, Jongkil Kim, and Seyit Camtepe. Ipremover: A generative" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1259" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 196 + ], + "type": "list", + "angle": 0, + "index": 3, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "model inversion attack against deep neural network fingerprinting and watermarking. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 7837-7845, 2024. 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 150 + ], + "type": "text", + "content": "[79] Zihang Zou, Boqing Gong, and Liqiang Wang. Anti-neuron watermarking: protecting personal data against unauthorized neural networks. In European Conference on Computer Vision, pages 449-465. Springer, 2022. 2, 3, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[80] Chaoshun Zuo, Zhiqiang Lin, and Yinqian Zhang. Why does your data leak? uncovering the data leakage in cloud from mobile apps. In 2019 IEEE Symposium on Security and Privacy (SP), pages 1296-1310. IEEE, 2019. 1" + } + ] + } + ], + "index": 2 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 296, + 748, + 315, + 757 + ], + "type": "text", + "content": "1260" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_content_list.json b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..18d7a574e64f6afa07d122dab92b05088f1c9b10 --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_content_list.json @@ -0,0 +1,1378 @@ +[ + { + "type": "text", + "text": "Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene", + "text_level": 1, + "bbox": [ + 138, + 130, + 857, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Donggeun Lim $^{1}$ Jinseok Bae $^{1}$ Inwoo Hwang $^{1}$ Seungmin Lee $^{1}$ Hwanhee Lee $^{2}$ Young Min Kim $^{1\\dagger}$", + "bbox": [ + 225, + 179, + 769, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Seoul National University $^{2}$ Chung-Ang University", + "bbox": [ + 320, + 220, + 676, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}\\{\\mathrm{rms}2836$ , capoo95, inusu0818, rsual, youngmin.kim $\\} \\in$ snu.ac.kr $^{2}$ hwanheelee@cau.ac.kr", + "bbox": [ + 122, + 239, + 867, + 253 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ba5a4f3b68252eb66586ef36117577f3dce395ad52a54384a8229c3f9223f601.jpg", + "image_caption": [ + "Figure 1. Our framework populates 3D scenes with multiple characters. The generated characters interact with their surroundings and other people, bringing the space to life." + ], + "image_footnote": [], + "bbox": [ + 96, + 266, + 359, + 448 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0a9c2520a27f79d3803801e501ac6ffa2909308936575c3cc8353ed33494e951.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 364, + 266, + 622, + 446 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/37d24f6da7963b2eab81ec7b6bbdb7b2bdc684b39e18dc750edd078c677799ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 625, + 267, + 900, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 493, + 326, + 508 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this work, we propose a framework that creates a lively virtual dynamic scene with contextual motions of multiple humans. Generating multi-human contextual motion requires holistic reasoning over dynamic relationships among human-human and human-scene interactions. We adapt the power of a large language model (LLM) to digest the contextual complexity within textual input and convert the task into tangible subproblems such that we can generate multi-agent behavior beyond the scale that was not considered before. Specifically, our event generator formulates the temporal progression of a dynamic scene into a sequence of small events. Each event calls for a well-defined motion involving relevant characters and objects. Next, we synthesize the motions of characters at positions sampled based on spatial guidance. We employ a high-level module to deliver scalable yet comprehensive context, translating events into relative descriptions that enable the retrieval of precise coordinates. As the first to address this problem at scale and with diversity, we offer a benchmark to assess diverse aspects of contextual reasoning. Benchmark results and user studies show that our framework effectively captures scene context with high scalability. The code and benchmark, along with result videos, are available", + "bbox": [ + 88, + 527, + 485, + 875 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "at our project page: https://rms0329.github.io/Event-Driven-Storytelling/.", + "bbox": [ + 511, + 494, + 908, + 525 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 556, + 643, + 571 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "A scene with digital humans creates lively atmosphere and has a wide range of applications in VR, games, and movies. The realism incurs from natural motions of multiple characters that seamlessly integrate into the surrounding environment. However, existing motion synthesis works focus on individual characters independently or extend to interactions with either the surrounding environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35], remaining limited to a single type of interaction. Hence, they severely lack generalizability and scalability to concurrently generate plausible interactions of multiple characters in a crowded scene. It is highly challenging to correctly assign the behavior of individual characters at the correct time stamp as it requires extensive dynamic contextual and spatial reasoning.", + "bbox": [ + 509, + 582, + 906, + 808 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To fulfill such holistic requirements, we take inspiration from recent advances in large language models (LLMs). LLMs demonstrated phenomenal performance in various areas, especially for high-level planning [11, 15, 24, 30, 34, 36] or behavior modeling [3, 29]. We deduce that LLMs have the potential to interpret nuanced semantics from", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Corresponding author", + "bbox": [ + 109, + 886, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11654", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "open-ended textual inputs and dependency between natural motion and scene context. However, simply generating commands from the input request may suffer from hallucinations or struggle with localization errors, which are well-known limitations of foundation models [18, 27, 32, 42].", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We propose an LLM-powered framework that populates multiple lifelike virtual humans within a 3D scene. As shown in Figure 1, the characters produced by our framework can understand and interact with their surroundings, including both 3D scenes and other characters. The proposed LLM module can flexibly understand the holistic context in open-ended scenarios and make high-level decisions. The characters may exhibit emergent yet plausible behavior that was not explicitly defined within the original input, providing a lively atmosphere with rich interactions. The framework can further adapt to user-interactive scenarios, where users may interact with the characters through high-level text instructions.", + "bbox": [ + 89, + 169, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Towards more performant and scalable planning with LLMs, our modular framework is deliberately designed to reduce the reasoning complexity. We introduce an event-based planning approach with two LLM modules: narrator and event parser. The narrator progressively weaves the flow of the scene by generating a textual description of one event at a time. Each event considers only a well-defined subset of characters and objects, detached from the collective behavior of the holistic flow. The event parser then transforms high-level event representation into detailed information for existing motion synthesis frameworks.", + "bbox": [ + 89, + 367, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We provide devised scene information tailored for individual modules to guide necessary spatial reasoning. Our scene descriptor converts the 3D scene graph into a textual summary and provides it as input to the narrator and the event parser. Our prompt encourages contextual information on different utilization of spaces, such that the description delivers the overall structure. The event parser provides accurate 3D scene grounding to the low-level motion synthesis module, such that it can generate fine-grained and realistic interaction. Inspired by previous studies [24, 34, 36, 44], the event parser employs spatial reasoning tools for LLM and specifies a relevant location within a programming framework. Furthermore, we propose an area-conditioned position sampling technique to compensate for the weakness of LLMs in precise localization at the coordinate level.", + "bbox": [ + 89, + 536, + 482, + 777 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our framework successfully generates long-term motions with more than 4-5 characters in various multi-room scale scenes, and is robust to the choice of LLM engines. We also propose evaluation criteria for scene-aware multi-agent behavior planning tasks and create a benchmark to assess the compatibility of our framework comprehensively. The extensive results demonstrate that our proposed pipeline outperforms in generating a natural and plausible", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "sequence of actions with adequate trajectories at appropriate relative timing in a shared space, especially for larger scenes with more number of characters.", + "bbox": [ + 511, + 90, + 906, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 152, + 663, + 169 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Contextual Human Motion Synthesis Synthesizing human motions is a long-standing research topic in the fields of computer vision and graphics. In particular, an increasing number of studies have incorporated contextual information in motion synthesis, such as a 3D environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35]. Although these studies have shown notable progress in reproducing physically accurate and high-fidelity motions for a single type of interaction, most of them do not consider more than one contextual constraint or expand to long-term sequences of actions. Some recent works [3, 5] have explored motion generation considering both humans and scenes of contextual information. Digital Life Project [3] builds autonomous characters with social intelligence in a 3D scene, while Sitcom-Crafter [5] introduces a unified motion generation framework that integrates different types of interaction motions with plot-driven guidance. The research expanded on previous boundaries, yet it did not take into account the coordination of interactions between more than two characters. Furthermore, in terms of scene understanding, these models lack the complexity needed for scenarios requiring detailed spatial reasoning because of their simplistic scene descriptions. Compared to these studies, our framework demonstrates successful multi-agent contextual behavior planning at a larger scale, including an increased number of characters and a scene with greater size and complexity.", + "bbox": [ + 511, + 180, + 906, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "LLM-based Planning In recent years, there has been growing attention to using pre-trained LLMs in various domains for their impressive zero-shot reasoning capabilities [8, 16, 43]. Particularly in the field of robotics and embodied AI, many recent works have successfully demonstrated the abilities of LLMs for task planning [11, 15, 20, 22, 24, 28, 30, 34, 36]. However, robotic applications primarily focus on goal-oriented tasks, which do not require complex contextual reasoning. While the majority of these works are limited to single-agent scenarios, a few studies [20, 46] have explored LLM-based multiagent task planning recently. CoELA [46] introduces a novel cognitive-inspired modular framework that enables cooperation and communication between multiple agents. However, since CoELA requires continuous message exchanges between agents for behavior coordination, this can result in inefficiency when scaling to an increased number of agents. On the other hand, SMART-LLM [20] demonstrated multi-agent task planning in larger-scale scenarios.", + "bbox": [ + 511, + 613, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11655", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/44d0ecc5690a7c89cf5be99672997194baeae64dc3d4fa22c53a2786ecbd6889.jpg", + "image_caption": [ + "Figure 2. Overview of our framework. The framework consists of two main modules: a high-level action planning module and a low-level motion synthesis module. The action planning module coordinates and plans multi-character behavior as a sequence of events. The motion synthesis module receives a detailed description of the event and synthesizes 3D motions." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 501, + 224 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "SMART-LLM accomplishes multi-robot cooperation by decomposing tasks into several sub-stages. However, SMART-LLM represents a scene as a mere list of objects, leading to a lack of comprehensive scene understanding. Distinct from earlier approaches, ours tackles the problem of multi-agent planning while maintaining scalability and in-depth scene understanding.", + "bbox": [ + 89, + 329, + 483, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 450, + 181, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a 3D scene $S$ , characters $\\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\}$ , and optional user instructions $\\mathcal{T}$ , our framework generates 3D multi-human motions $\\mathcal{M} = \\{M_1, M_2, \\dots, M_N\\}$ within the scene. Here, $N$ represents the number of characters and is determined by the user at the system initialization. The 3D scene $S$ is provided with instance segmentation, and can be in various representations including VR scenes with CAD models or real-world scenes composed of 3D point cloud scans or mesh reconstructions. The user instructions $\\mathcal{T}$ are provided in free text forms, leveraging the generalizability of LLM engines.", + "bbox": [ + 89, + 476, + 483, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In an overview, our framework operates around an intermediate representation called event $e$ . Figure 2 contains examples of textual descriptions of events along with detailed action labels and positions for the motion synthesis module. The events are generated in sequence to realize the given instruction while coordinating with the surrounding environments. Each event has its own lifecycle with a different time span, and multiple events can coexist within the 3D scene. Through the event-based formulation, our system can generate scene dynamics with infinitely long-horizon, and adaptively process user instructions at an arbitrary time during operation.", + "bbox": [ + 89, + 643, + 483, + 824 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, our framework consists of two main components: a high-level action planning module and a low-level motion synthesis module. The action planning module performs planning on an event basis, driving the scene's timeline forward (Section 3.1). The motion syn", + "bbox": [ + 89, + 825, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "thesis module takes the generated events as input, assigns them to the corresponding characters, and creates motions that reflect the assigned event (Section 3.2).", + "bbox": [ + 511, + 90, + 905, + 136 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. High-level Action Planning Module", + "text_level": 1, + "bbox": [ + 511, + 146, + 818, + 162 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The primary function of the action planning module is to generate new events by comprehensively considering the given 3D scene $S$ , existing event history $\\mathcal{H}$ , and the behaviors of multiple characters within the scene. To effectively satisfy the holistic requirements, we rely on the powerful reasoning capabilities of LLMs. As shown in Figure 3, there are three LLM submodules: the scene descriptor, the narrator, and the event parser. The scene descriptor generates a textual scene description $\\mathcal{D}$ from the given 3D scene $S$ such that our planning modules can understand the necessary context from the spatial arrangement. The narrator generates a sequence of events, which are converted into detailed information by the event parser. Each module is assigned to smaller, well-defined tasks such that the system stays performant and scalable for multi-human motion generation.", + "bbox": [ + 511, + 167, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Scene Descriptor", + "text_level": 1, + "bbox": [ + 511, + 417, + 669, + 430 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the preprocessing stage before the motion planning, the scene descriptor generates a textual description $\\mathcal{D}$ of the given 3D scene $S$ . We use a 3D scene graph, which may lead an LLM module to absorb rich spatial relationships [14]. The scene descriptor first extracts a 3D scene graph in an automated way as proposed in [17] and converts it into textual data in JSON format. Then, we use it to prompt the LLM module and generate a scene description. In addition to the explicit pairwise relationships in the scene graph, further contextual layouts with regional information enhance the naturalness of scene-aware motions. To enhance contextual reasoning, we encourage our scene descriptor to discover a spatial arrangement of objects in proximity, which semantically creates a functional space, for example, a dining area or study zone. Specifically, we extract a regional cluster of objects through the DBSCAN algorithm [9] as an additional input and promote detecting areas of interest. We also provide concrete examples of the same input and output format, allowing the scene descriptor to perform the given task more effectively through in-context learning [2]. We provide additional details and full prompts in the supplementary material.", + "bbox": [ + 511, + 436, + 908, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. Narrator", + "text_level": 1, + "bbox": [ + 511, + 776, + 624, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As the core of our behavior planning, the narrator generates a sequence of events $\\mathcal{E} = \\{e_1,e_2,\\ldots \\}$ . Given the scene description $\\mathcal{D}$ from the scene descriptor and the optional user instruction $\\mathcal{T}$ , the narrator keeps the previous event history $\\mathcal{H}$ and generates one event at a time based on the high-level context. Focusing on one event at a time, the subsequent module can alleviate the burden of holistic spatio-temporal", + "bbox": [ + 511, + 795, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11656", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/9a71ebeff84852c8a41fd1a76d6edf2c2a11c3255ae0f8c394a9643bfeeadd5f.jpg", + "image_caption": [ + "Figure 3. Overview of the high-level action planning module." + ], + "image_footnote": [], + "bbox": [ + 109, + 85, + 890, + 214 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "reasoning, making the storytelling scalable and interactive. The event-based design also allows the incorporation of additional user instructions between events to refine the storyline.", + "bbox": [ + 88, + 250, + 480, + 309 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As shown in Figure 2, events are expressed in seminarrative natural language, focusing on contextual description. The textual description is further parsed in the subsequent event parser, relieving the burden of detailed spatiotemporal localization. The narrator also receives the state of a generated event as one of two states, 'ongoing' or 'completed'. From this, the narrator can compactly conceive the necessary information to progress the timeline. Our event representation is particularly efficient for group events with multiple characters, which makes the broad scene context easily understood. The narrator coordinates the high-level behaviors of multiple characters and generates the storyline from a macroscopic perspective. The detailed spatiotemporal progression of individual characters is assigned to the event parser.", + "bbox": [ + 91, + 311, + 483, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to the scene descriptor, we also incorporate several widely used prompting techniques in the narrator's prompt, the in-context learning and chain-of-thought reasoning [41]. For chain-of-thought reasoning, we guide the narrator to first reason about the current planning state and other characters' statuses before generating an event. We provide concrete examples of event generation with such a reasoning process to reinforce structured decision-making and improve the coherence of generated events. Further details and the full prompts are provided in the supplementary material.", + "bbox": [ + 89, + 539, + 483, + 705 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1.3.Event Parser", + "text_level": 1, + "bbox": [ + 89, + 714, + 227, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The event parser takes the event $e$ generated by the narrator as input and concretizes it into low-level details that the motion synthesis module can digest. The output event of the narrator is designed to be a short description of the region and the character for the motion, leaving ambiguities on the exact location within the scene. The event parser observes the current state of the scene and refines the event description into detailed 3D grounding suitable for motion synthesis. Specifically, it converts the event description $e$ into $e = (\\mathcal{C}_e, \\{p_i\\}, \\{d_i\\}, \\{a_i\\})$ , where $\\mathcal{C}_e = \\{c_i\\}$ is the set of involved characters, $p_i \\in \\mathbb{R}^2$ is the target position,", + "bbox": [ + 89, + 734, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/2919c4aaa9cd9d79817cb05a7c6eeadf16157fa02b7a6e01f9a6db6a740a196d.jpg", + "image_caption": [ + "Figure 4. An example of a programming-structured prompt used in our event parser. The event parser leverages provided spatial reasoning tools, such as the get_distance_between() function in this example, to procedurally deduce solutions for given spatial reasoning tasks." + ], + "image_footnote": [], + "bbox": [ + 511, + 252, + 910, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$d_{i}\\in \\mathbb{R}^{1}$ is the target orientation, and $a_{i}$ is the target action label for each character $c_{i}$ .", + "bbox": [ + 511, + 551, + 905, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Inspired by [24, 34, 44], we adopt a Python programming structure in the event parser's prompt and utilize spatial reasoning tools that are provided as a form of Python functions. The programming framework can infer solutions in a well-defined sequence of inquiries, leading to more powerful spatial reasoning capabilities. Figure 4 shows a specific example. The event parser uses the get_distance_between() function to deduce spatial details missing from the scene description, such as identifying the farthest seat from a given reference point. The event parser compiles the output format by calling the get_character() function, and for each character $c_i$ , it specifies $p_i$ , $d_i$ , and $a_i$ using the set_position(), set_ORIENTATION(), and set_target_action() functions, respectively. If a specific target orientation is not necessary, it can be left unspecified. When retrieving the positions, the event parser first outputs semantic descriptions using scene objects, and precise coordinates are found afterward to compensate for possible errors of LLMs in coordinate-level reasoning [7, 27]. Specifically, we define a position by its spatial relationship with an anchor object", + "bbox": [ + 511, + 583, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11657", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9713893b4dc8d03263da5179c1dfdc0ed2aad564f83ae10feb76a45832b8e1ed.jpg", + "image_caption": [ + "Figure 5. Qualitative results of our framework across diverse 3D scenes." + ], + "image_footnote": [], + "bbox": [ + 135, + 88, + 864, + 223 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and sample the coordinate representation $p_i$ within the corresponding area. Then, if a target orientation is specified, we calculate $d_i$ as the object or character the subject faces after sampling $p_i$ . Otherwise, $d_i$ is set to the character's expected forward direction upon reaching $p_i$ .", + "bbox": [ + 88, + 262, + 480, + 337 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As with the narrator, the event parser incorporates in-context learning, chain-of-thought reasoning, and self-feedback techniques in its prompt. Further details about sampling and the full prompts are provided in the supplementary material.", + "bbox": [ + 88, + 338, + 482, + 412 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Low-level Motion Synthesis Module", + "text_level": 1, + "bbox": [ + 89, + 424, + 401, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After the high-level action planning module provides the narrowed location and character action to generate, the subsequent motion synthesis module concentrates solely on converting the set of events into character motions in a 3D scene in an online manner, detached from the holistic flow of the scene. Given the positions and actions of individual characters, the motion synthesis may employ any existing framework to reach the location and perform the designated action. An event is completed and removed from the scene when all characters in $\\{c_i\\}$ complete their own target actions $a_{i}$ . We briefly describe our motion synthesis module below, and more details are available in the supplementary material.", + "bbox": [ + 88, + 446, + 482, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We first plan a collision-free trajectory and generate locomotion along it such that the character $c_{i}$ moves to the target position $p_{i}$ , aligning its facing direction with the target orientation $d_{i}$ . While reaching the target position $p_{i}$ , the character should avoid collisions with other characters and the surrounding environment. We first create a 2D grid map that represents navigable areas in the scene, following the approach used in [38]. On this grid map, our motion synthesis module finds collision-free multi-agent paths using the windowed cooperative $A^{*}$ algorithm proposed in [33].", + "bbox": [ + 88, + 643, + 482, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After reaching the target position, the motion synthesis module creates motions assigned with the action label $a_{i}$ . We leverage the motion matching [6] to quickly generate plausible motions. If $p_i$ is deliberately set to overlap with an object (e.g., on a chair), the motion synthesis module first makes a transition to a desired posture (e.g., sitting) before creating motions for $a_{i}$ . There are also group events,", + "bbox": [ + 88, + 795, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where multiple characters are involved, such as chatting or handshake. Because temporal synchronization is critical, the characters who reach $p_i$ before others maintain an idle motion and wait for the remaining characters. Motion synthesis for $a_i$ occurs only after all characters have reached their respective target position $p_i$ .", + "bbox": [ + 511, + 262, + 905, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 368, + 645, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The key functionality of our framework is to perform LLM-based scene-aware multi-agent planning to enable natural interactions and activities of characters within a given 3D scene. We can successfully populate natural motions of multiple characters within diverse scenes ranging from virtual scenes composed of CAD models (Figure 1) to 3D scans (Figure 5). More qualitative results are available in the supplementary materials.", + "bbox": [ + 511, + 393, + 905, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To the best of our knowledge, our framework is the first to address this task at this scale; there is no existing baseline for direct comparison with our framework. Therefore, we propose evaluation criteria to comprehensively assess our framework (Section 4.1) and create a benchmark designed for this assessment (Section 4.2). Using the developed benchmark, we evaluate the effectiveness of our proposed key methodologies through an ablation study (Section 4.3). Furthermore, we conduct a user study to visually review and compare the generated results across the different ablation settings (Section 4.4).", + "bbox": [ + 511, + 515, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Evaluation Criteria", + "text_level": 1, + "bbox": [ + 511, + 693, + 700, + 707 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We propose a set of criteria to assess several capabilities required for the task as shown below. We define each evaluation criterion as a tag and use these tags to categorize test cases in the benchmark.", + "bbox": [ + 511, + 715, + 903, + 773 + ], + "page_idx": 4 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Object arrangement reasoning (OA): Object arrangement reasoning is essential for identifying objects based on spatial relationships (e.g., the chair farthest from the window) or ensuring that characters interact correctly with their surroundings (e.g., to study at a desk, a character needs to sit on the associated chair).", + "- Regional context reasoning $(RC)$ : Regional context reasoning is required for generating character behaviors that" + ], + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11658", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/841df82702aa02d5c36d30dd6a4ff7be29d9782ac412cbf2fa81eb0bd52400f5.jpg", + "image_caption": [ + "Figure 6. An example of a test case. In this test case, a system is required to identify the chairs in the kitchen and select one that is not currently occupied." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 472, + 223 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "are well-aligned with the contextual meaning of a space (e.g., cooking in the kitchen rather than in the living room).", + "bbox": [ + 102, + 290, + 482, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "- Scene state reasoning (SS): Scene state reasoning requires considering the overall scene flow and the current state of other characters (e.g., if a character is using the coffee machine, another character should wait to use it).", + "bbox": [ + 89, + 335, + 482, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In addition to the holistic spatial context defined above, we separately evaluate the low-level performance on Position inference $(PI)$ . The cases with the tag analyze if the system can find the precise positions of characters at the coordinate level.", + "bbox": [ + 89, + 398, + 483, + 472 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Benchmark Creation", + "text_level": 1, + "bbox": [ + 89, + 483, + 289, + 497 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Scenes We create three scenes—House, Office, and Restaurant—by placing the objects from the HSSD dataset [21]. Each scene is designed to include two to three separate areas with distinct contextual meaning. To ensure the LLM infers contextual meaning of areas without relying on object labels, we replace several object labels with general ones (e.g., kitchen_cabinet $\\rightarrow$ cabinet). We also randomize indices for identically labeled objects to eliminate index-based arrangement cues. We define a distinct set of available action labels $A_{S}$ for each scene, where $S \\in \\{\\text{house}, \\text{office}, \\text{restaurant}\\}$ . This $A_{S}$ is utilized when determining the pass/fail of test cases that use the scene $S$ .", + "bbox": [ + 89, + 505, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Test Cases Our framework can create natural behaviors for multiple characters without any pre-scripted scenarios. However, for rigorous evaluation, we introduce specific test scenarios in our benchmark. Each test case consists of a scene $S$ , characters $\\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\}$ , test scenario $\\mathcal{T}$ , mock planning history $\\mathcal{H}$ , and expected plan $\\tau_e = (\\mathcal{A}_e, \\tilde{p}_e)$ . Here, $N$ is the number of characters, $\\mathcal{A}_e$ is the set of expected action labels, and $\\tilde{p}_e$ is the expected area where the target character's position is expected to be included. $\\tilde{p}_e$ is expressed through the union or intersection of multiple areas, each defined as a semantic description based on scene objects. For simplicity, we ignore the character's orientation in the benchmark, and the target character for the expected", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "plan $\\tau_{e}$ is always fixed to the first character $c_{1}$ in all test cases. An example of a test case is depicted in Figure 6.", + "bbox": [ + 511, + 90, + 905, + 121 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "When running a test case, we input the test scenario $\\mathcal{T}$ and mock planning history $\\mathcal{H}$ into the system. The mock planning history $\\mathcal{H}$ simulates the prior planning history of the scene, building the necessary context required for the test case. Then, the system generates the most appropriate next plan $\\tau_{g} = (a_{g},\\tilde{p}_{g}$ or $p_g)$ by evaluating the given mock planning history and the test scenario. Here, $\\tau_{g}$ is the generated next plan for the target character, $a_{g}$ is the planned action label, and $\\tilde{p}_g$ or $p_g$ is the planned position. If the test case evaluates coordinate-level localization, the system outputs a coordinate-level position $p_{g}$ . Otherwise, it outputs an area-level position $\\tilde{p}_g$ . As the final step, the test case's pass/fail is determined by comparing the expected plan $\\tau_{e}$ and the generated plan $\\tau_{g}$ . For a test case to pass, the following conditions must all be satisfied:", + "bbox": [ + 511, + 125, + 906, + 349 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "1. A plan for the target character is generated $(\\tau_{g}\\neq \\emptyset)$", + "2. The action label is included in both the available action labels and the expected action labels $(a_{g}\\in \\mathcal{A}_{\\mathcal{S}}\\cap \\mathcal{A}_{e})$", + "3. The character's position is within the expected area $(\\tilde{p}_g \\subseteq \\tilde{p}_e$ or $p_g \\in \\tilde{p}_e)$ ." + ], + "bbox": [ + 511, + 352, + 903, + 429 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Benchmark Summary We construct the benchmark based on the previously defined evaluation criteria, namely, tags. Our benchmark consists of 40 test cases in total. 10 test cases are designed for the $PI$ tag, while the remaining 30 test cases involve $OA$ , $RC$ , and $SS$ tags. The test cases for the $PI$ tag deliberately use straightforward scenarios to focus solely on the localization performance. Among the latter 30 test cases, half incorporate two tags for increased complexity. Additionally, for the latter 30 test cases, we vary the number of characters in the scene and the size of the mock planning history across three levels within the same scenario. This expands the number of effective test cases and allows us to evaluate the system's scalability by assessing its robustness against increasing agent counts and planning history complexity. We design our benchmark to ensure that each tag is evenly represented, preventing bias toward any specific evaluation factor.", + "bbox": [ + 511, + 460, + 906, + 717 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics Our benchmark employs two evaluation metrics: success rate and execution rate. The success rate measures the proportion of passed runs among all test case runs. On the other hand, the execution rate measures the proportion of runs that are executed without any runtime errors, which are included in the parenthesis next to the success rate in tables. Runtime errors typically occur when the LLM module references a nonexistent object or generates a response with incorrect syntax. Any test case run with a runtime error is automatically marked as failed.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11659", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/750e8d6b21c95af4e8452769d6f23b401386abf8892d2f905b368589f8246ca9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.9 (0.98)0.93 (0.99)0.9 (0.98)0.92 (0.98)0.74 (0.96)0.72 (0.95)0.72 (0.97)0.78 (0.93)0.72 (0.87)0.78 (0.91)0.76 (0.87)0.61 (0.81)
w/o Event0.82 (0.92)0.88 (0.92)0.86 (0.93)0.77 (0.92)0.6 (0.95)0.56 (0.99)0.6 (0.92)0.54 (0.92)0.6 (0.81)0.6 (0.78)0.62 (0.86)0.54 (0.76)
Object List0.51 (0.85)0.61 (0.78)0.28 (0.88)0.65 (0.97)0.34 (0.88)0.37 (0.88)0.12 (0.86)0.51 (0.9)0.35 (0.68)0.31 (0.7)0.29 (0.61)0.41 (0.69)
Scene Graph0.82 (0.96)0.8 (0.96)0.82 (0.94)0.87 (0.95)0.49 (0.9)0.52 (0.93)0.32 (0.9)0.51 (0.83)0.72 (0.92)0.69 (0.91)0.77 (0.94)0.64 (0.85)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.35 (0.74)0.37 (0.8)0.37 (0.77)0.35 (0.67)0.71 (0.9)0.71 (0.91)0.72 (0.95)0.65 (0.85)0.39 (0.87)0.38 (0.88)0.36 (0.87)0.35 (0.85)
w/o Event0.31 (0.58)0.3 (0.57)0.32 (0.6)0.28 (0.63)0.62 (0.85)0.74 (0.85)0.65 (0.9)0.59 (0.85)0.35 (0.79)0.41 (0.86)0.32 (0.79)0.32 (0.74)
Object List0.22 (0.58)0.26 (0.65)0.13 (0.62)0.23 (0.44)0.44 (0.77)0.45 (0.77)0.36 (0.83)0.52 (0.72)0.14 (0.77)0.09 (0.74)0.07 (0.77)0.24 (0.67)
Scene Graph0.29 (0.75)0.31 (0.76)0.35 (0.86)0.16 (0.61)0.69 (0.84)0.69 (0.81)0.77 (0.93)0.55 (0.72)0.31 (0.84)0.27 (0.88)0.23 (0.84)0.35 (0.84)
", + "bbox": [ + 93, + 88, + 903, + 239 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Benchmark result for test cases with object arrangement reasoning (OA), regional context reasoning (RC), and scene state reasoning (SS) tags. In total, our method achieves the highest success rate across various LLM models and sizes.", + "bbox": [ + 89, + 250, + 906, + 277 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45e1ee3cf38af64f0a0e53a406289310acbb6f71238197d5d152b45869456e3f.jpg", + "image_caption": [ + "Figure 7. Benchmark results by number of characters. Our approach exhibits stable success rate for stories with different number of characters while the performance gap increases as more characters are involved, compared to the version without event representation." + ], + "image_footnote": [], + "bbox": [ + 109, + 290, + 459, + 424 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5c907f3c4ddf01e44942aaf55a63c1f808eb977e6c75de650e5684f1c11684f9.jpg", + "image_caption": [ + "Figure 8. Comparison of average token usage. Object List uses the least number of tokens, and Scene Graph uses significantly more number of tokens, especially when the scene complexity increases (Office)." + ], + "image_footnote": [], + "bbox": [ + 109, + 496, + 454, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Benchmark Results", + "text_level": 1, + "bbox": [ + 89, + 657, + 277, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We use GPT-4o [16], Llama-3.1 [8], and Qwen2.5 [43] in various sizes as LLM backbones to obtain benchmark results. The backbones are selected to represent both commercial and open-source modules that are popular and demonstrate competitive performance. We repeat each test case five times and average the results.", + "bbox": [ + 89, + 679, + 482, + 772 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Event-Driven Planning Table 1 presents the benchmark results for test cases with $OA$ , $RC$ , and $SS$ tags. Each column represents the average score for test cases that include the corresponding tag and each cell displays the success rate, with the execution rate shown in parentheses. The ablation without event-driven planning (w/o Event) handles the entire process with a single unified LLM module without using high-level event representations. Therefore, all in", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "put and output information, except for the scene description and test scenario, is represented in a low-level structured format. Compared to $w/o$ Event, our method benefits from a cascaded planning pipeline that effectively decomposes the entire planning complexity. This leads to an overall improvement in planning performance. In particular, the performance gain in SS is relatively larger compared to other tags in GPT-40 models. The SS tag requires a comprehensive understanding of the overall planning state of multiple characters in the scene. We attribute this to the advantages of our event-based representation, which provides a more efficient and intuitive way to capture the overall scene state. Additionally, as shown in Figure 7, our planning pipeline proves to be more effective in terms of scalability. $w/o$ Event struggles with planning as the number of characters in the scene increases, whereas our approach maintains stable performance.", + "bbox": [ + 511, + 292, + 906, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Scene Description We also analyze the effect of using scene description generated from an LLM module. We create ablation settings that retain our pipeline but modify the scene description method. Object List represents the scene as a simple list of objects present in the scene, while Scene Graph uses the JSON-formatted scene graph representation as the scene description. Our approach contributes to more effective planning compared to other description methods as shown in Table 1. Additionally, we present the average token usage for each description method in Figure 8 to provide a deeper understanding of their differences. As presented in Figure 8, the object list is the most efficient way to represent scene information and has often been used in previous studies such as [5, 20, 34]. However, since the object list does not have any spatial information, the planning module with the description is ignorant of object arrangement or spatial layout. As a result, the object list records the lowest scores in our benchmark, which requires sophisticated scene understanding for planning.", + "bbox": [ + 511, + 566, + 908, + 854 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Unlike the object list, the scene graph representation enables necessary spatial reasoning by incorporating detailed attributes of objects and spatial relationships between ob", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11660", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1afa5562a2826a805df55a656824600dd25a926da20e4e72652fd9871f1afc91.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
Ours0.94 (1.0)0.9 (1.0)0.94 (1.0)
Direct Inference0.46 (1.0)0.38 (1.0)0.28 (1.0)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
Ours0.86 (0.88)1.0 (1.0)0.94 (1.0)
Direct Inference0.14 (0.9)0.14 (0.48)0.1 (0.8)
", + "bbox": [ + 122, + 88, + 452, + 172 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Benchmark result for test cases with position inference tag (PI). Our proposed method can stably infer accurate positions, whereas Direct Inference with LLMs often struggles.", + "bbox": [ + 89, + 181, + 482, + 224 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "jects. However, the representation contains exhaustive detail, including even minor and insignificant ones, without any abstraction or compression. This inefficiency leads to excessive token consumption (Figure 8) and can even interfere with the LLM's reasoning due to unnecessary details (Table 1). The token inefficiency becomes more pronounced as the scene size and complexity increase. As shown in Figure 8, when comparing the relatively simple 'House' scene to the more complex 'Office' scene, the scene graph exhibits a significantly larger increase in token consumption compared to alternative approaches. Compared to the scene graph, our approach allows for a more efficient description of the given scene while preserving key spatial information.", + "bbox": [ + 88, + 234, + 485, + 445 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Positional Inference (PI) In Table 2, we demonstrate the effectiveness of our area-conditioned position sampling method based on the results for the PI-tagged test cases. Direct Inference is an ablation setting where the same planning pipeline is used, but the LLM directly outputs coordinates when determining character positions. Although all test cases with the PI tag are intentionally designed to be very simple, all LLM models with Direct Inference struggle to accurately determine the coordinates of character positions. In contrast, our system allows the LLM to process information semantically, and obtains precise locations through the proposed area-conditioned position sampling technique.", + "bbox": [ + 89, + 457, + 483, + 640 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. User Study Results", + "text_level": 1, + "bbox": [ + 89, + 651, + 272, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also validate through a user study that our approach represents the given scenario more effectively than other ablation settings. For the user study, we use a total of four test scenarios, each set in a different scene. The scenes include House, Office, and Restaurant from the benchmark, as well as MPH11 from the PROX dataset [12]. Compared to the other scenes, MPH11 is relatively simple in both scale and complexity. Given its limited space, we use a simple scenario containing only two events for the MPH11. In contrast, for other test scenes, we use more complex scenarios that include 4-5 events. For each test scenario, users are asked to visually compare results generated from different ablation settings and select the one that best represents the given scenario. If multiple results are perceived as equally effective, users are allowed to choose more than one. We", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/95d295ffc686b4380dde9ff5b00d25bd99e59bf7f9feba4c3539b75741b2e782.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 532, + 87, + 883, + 164 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/040e2bb95e94b12f51e808d1a098e2522ee7da6fb983f8f442327e5299fcec06.jpg", + "image_caption": [ + "Figure 9. User study results. Our system correctly generates the dynamic story line with multiple characters in large scenes, significantly outperforming other baselines, except for simple scenarios with MPH11." + ], + "image_footnote": [], + "bbox": [ + 532, + 167, + 883, + 241 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generate the results for the user study using GPT-4o mini with a temperature setting of 0.0.", + "bbox": [ + 511, + 314, + 903, + 343 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We collected responses from 50 general participants. The responses to the user study are summarized in Figure 9. The chart labels SF, OL, and No E correspond to the same ablation settings used in the benchmark: Scene Graph, Object List, and w/o Event, respectively. In three of our scenarios, our approach received the most selections by a large margin. Furthermore, the noticeable difference in selection distribution between the MPH11 scene, which involves a relatively smaller-scale problem, and the other scenes suggests that our approach contributes to building a more scalable system. We provide the test scenarios and visualization of generated results in the supplementary material.", + "bbox": [ + 511, + 345, + 906, + 526 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 542, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this study, we propose an LLM-based framework for generating multiple lifelike virtual humans within a given 3D scene, dynamically aligning their behaviors with an emergent storyline. Characters generated by our framework interact contextually with their surroundings and other individuals, enhancing scene realism. Our framework performs multi-character behavior planning based on semi-narrative events, allowing the LLM to efficiently capture scene progression and group behaviors. The modular design of our planning pipeline decomposes the overall problem, reducing reasoning complexity for the LLM and enabling more scalable planning. Additionally, we introduce a context-centric scene description to efficiently convey key spatial information and propose area-conditioned position sampling to mitigate the LLM's weaknesses in numerical reasoning. To evaluate our system, we construct a benchmark for a comprehensive assessment of LLM-based scene-aware multi-agent behavior planning. Through benchmark results and a user study, we demonstrate that our proposed methods contribute to building a scalable and effective system for coordinating multi-agent behaviors, while carefully considering the scene.", + "bbox": [ + 511, + 568, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11661", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work was supported by the IITP grant [RS-2021-II211341, Artificial Intelligence Graduate School Program (Chung-Ang University)], the NRF grant (No. RS-2023-00208197) funded by the Korea government (MSIT), and Creative-Pioneering Researchers Program through Seoul National University.", + "bbox": [ + 89, + 90, + 485, + 183 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 193, + 187, + 209 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 1, 2", + "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3", + "[3] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 1, 2", + "[4] Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Generating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1855-1866, 2024. 1, 2", + "[5] Jianqi Chen, Panwen Hu, Xiaojun Chang, Zhenwei Shi, Michael Christian Kampffmeyer, and Xiaodan Liang. Sitcom-crafter: A plot-driven human motion generation system in 3d scenes. arXiv preprint arXiv:2410.10790, 2024. 2, 7", + "[6] Simon Clavet et al. Motion matching and the road to next-gen animation. In Proc. of GDC, page 4, 2016. 5", + "[7] Arash Gholami Davoodi, Seyed Pouyan Mousavi Davoudi, and Pouya Pezeshkpour. Llms are not intelligent thinkers: Introducing mathematical topic tree benchmark for comprehensive evaluation of llms. arXiv preprint arXiv:2406.05194, 2024. 4", + "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 7", + "[9] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 3", + "[10] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision, pages 418-437. Springer, 2024. 1, 2" + ], + "bbox": [ + 93, + 218, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[11] Maitrey Gramopadhye and Daniel Szafir. Generating executable action plans with environmentally-aware language models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3568-3575. IEEE, 2023. 1, 2", + "[12] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 8", + "[13] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 1, 2", + "[14] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In Proceedings of the 41st International Conference on Machine Learning, pages 20413-20451, 2024. 3", + "[15] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2", + "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 7", + "[17] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In European Conference on Computer Vision, pages 289-310. Springer, 2024. 3", + "[18] Bowen Jiang, Yangxinyu Xie, Zhuoqun Hao, Xiaomeng Wang, Tanwi Mallick, Weijie Su, Camillo Taylor, and Dan Roth. A peek into token bias: Large language models are not yet genuine reasoners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4722-4756, 2024. 2", + "[19] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 1, 2", + "[20] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12140–12147. IEEE, 2024. 2, 7", + "[21] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 6" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11662", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[22] Boyi Li, Philipp Wu, Pieter Abbeel, and Jitendra Malik. Interactive task planning with language models. arXiv preprint arXiv:2310.10645, 2023. 2", + "[23] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, 132(9):3463-3483, 2024. 1, 2", + "[24] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 1, 2, 4", + "[25] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 1, 2", + "[26] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 903-913. IEEE, 2024. 1, 2", + "[27] Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. 2, 4", + "[28] Zhe Ni, Xiaoxin Deng, Cong Tai, Xinyue Zhu, Qinghongbing Xie, Weihang Huang, Xiang Wu, and Long Zeng. Grid: Scene-graph-based instruction-driven robotic task planning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13765-13772. IEEE, 2024. 2", + "[29] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1-22, 2023. 1", + "[30] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Sunderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, 2023. 1, 2", + "[31] Mengyi Shan, Lu Dong, Yutao Han, Yuan Yao, Tao Liu, Ifeoma Nwogu, Guo-Jun Qi, and Mitch Hill. Towards open domain text-driven synthesis of multi-person motions. In European Conference on Computer Vision, pages 67-86. Springer, 2024. 1, 2", + "[32] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Scharli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR, 2023. 2", + "[33] David Silver. Cooperative pathfinding. In Proceedings of the aaai conference on artificial intelligence and interactive digital entertainment, pages 117-122, 2005. 5", + "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot" + ], + "bbox": [ + 89, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 1, 2, 4, 7", + "[35] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 1, 2", + "[36] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024. 1, 2", + "[37] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 1, 2", + "[38] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022. 5", + "[39] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems, 35:14959-14971, 2022.", + "[40] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 1, 2", + "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4", + "[42] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2", + "[43] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 7", + "[44] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 2, 4", + "[45] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision, pages 246-263. Springer, 2024. 1, 2", + "[46] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11663", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Gan. Building cooperative embodied agents modularly with large language models. In The Twelfth International Conference on Learning Representations, 2024. 2", + "[47] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14738-14749, 2023. 1, 2" + ], + "bbox": [ + 91, + 90, + 480, + 204 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "11664", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_model.json b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..54db01c6fea202d99bf48733ed71ac8c014acb7f --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_model.json @@ -0,0 +1,1960 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.14, + 0.131, + 0.859, + 0.153 + ], + "angle": 0, + "content": "Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene" + }, + { + "type": "text", + "bbox": [ + 0.226, + 0.18, + 0.771, + 0.217 + ], + "angle": 0, + "content": "Donggeun Lim\\(^{1}\\) Jinseok Bae\\(^{1}\\) Inwoo Hwang\\(^{1}\\) Seungmin Lee\\(^{1}\\) Hwanhee Lee\\(^{2}\\) Young Min Kim\\(^{1\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.321, + 0.221, + 0.678, + 0.238 + ], + "angle": 0, + "content": "\\(^{1}\\)Seoul National University \\(^{2}\\)Chung-Ang University" + }, + { + "type": "text", + "bbox": [ + 0.124, + 0.24, + 0.868, + 0.255 + ], + "angle": 0, + "content": "\\(^{1}\\{\\mathrm{rms}2836\\) , capoo95, inusu0818, rsual, youngmin.kim \\(\\} \\in\\) snu.ac.kr \\(^{2}\\) hwanheelee@cau.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.267, + 0.361, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.365, + 0.267, + 0.624, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.626, + 0.268, + 0.901, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.452, + 0.908, + 0.481 + ], + "angle": 0, + "content": "Figure 1. Our framework populates 3D scenes with multiple characters. The generated characters interact with their surroundings and other people, bringing the space to life." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.494, + 0.327, + 0.509 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.528, + 0.486, + 0.875 + ], + "angle": 0, + "content": "In this work, we propose a framework that creates a lively virtual dynamic scene with contextual motions of multiple humans. Generating multi-human contextual motion requires holistic reasoning over dynamic relationships among human-human and human-scene interactions. We adapt the power of a large language model (LLM) to digest the contextual complexity within textual input and convert the task into tangible subproblems such that we can generate multi-agent behavior beyond the scale that was not considered before. Specifically, our event generator formulates the temporal progression of a dynamic scene into a sequence of small events. Each event calls for a well-defined motion involving relevant characters and objects. Next, we synthesize the motions of characters at positions sampled based on spatial guidance. We employ a high-level module to deliver scalable yet comprehensive context, translating events into relative descriptions that enable the retrieval of precise coordinates. As the first to address this problem at scale and with diversity, we offer a benchmark to assess diverse aspects of contextual reasoning. Benchmark results and user studies show that our framework effectively captures scene context with high scalability. The code and benchmark, along with result videos, are available" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.495, + 0.909, + 0.526 + ], + "angle": 0, + "content": "at our project page: https://rms0329.github.io/Event-Driven-Storytelling/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.557, + 0.645, + 0.572 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.583, + 0.907, + 0.809 + ], + "angle": 0, + "content": "A scene with digital humans creates lively atmosphere and has a wide range of applications in VR, games, and movies. The realism incurs from natural motions of multiple characters that seamlessly integrate into the surrounding environment. However, existing motion synthesis works focus on individual characters independently or extend to interactions with either the surrounding environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35], remaining limited to a single type of interaction. Hence, they severely lack generalizability and scalability to concurrently generate plausible interactions of multiple characters in a crowded scene. It is highly challenging to correctly assign the behavior of individual characters at the correct time stamp as it requires extensive dynamic contextual and spatial reasoning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To fulfill such holistic requirements, we take inspiration from recent advances in large language models (LLMs). LLMs demonstrated phenomenal performance in various areas, especially for high-level planning [11, 15, 24, 30, 34, 36] or behavior modeling [3, 29]. We deduce that LLMs have the potential to interpret nuanced semantics from" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.235, + 0.9 + ], + "angle": 0, + "content": "†Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11654" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "open-ended textual inputs and dependency between natural motion and scene context. However, simply generating commands from the input request may suffer from hallucinations or struggle with localization errors, which are well-known limitations of foundation models [18, 27, 32, 42]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.365 + ], + "angle": 0, + "content": "We propose an LLM-powered framework that populates multiple lifelike virtual humans within a 3D scene. As shown in Figure 1, the characters produced by our framework can understand and interact with their surroundings, including both 3D scenes and other characters. The proposed LLM module can flexibly understand the holistic context in open-ended scenarios and make high-level decisions. The characters may exhibit emergent yet plausible behavior that was not explicitly defined within the original input, providing a lively atmosphere with rich interactions. The framework can further adapt to user-interactive scenarios, where users may interact with the characters through high-level text instructions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Towards more performant and scalable planning with LLMs, our modular framework is deliberately designed to reduce the reasoning complexity. We introduce an event-based planning approach with two LLM modules: narrator and event parser. The narrator progressively weaves the flow of the scene by generating a textual description of one event at a time. Each event considers only a well-defined subset of characters and objects, detached from the collective behavior of the holistic flow. The event parser then transforms high-level event representation into detailed information for existing motion synthesis frameworks." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.537, + 0.483, + 0.779 + ], + "angle": 0, + "content": "We provide devised scene information tailored for individual modules to guide necessary spatial reasoning. Our scene descriptor converts the 3D scene graph into a textual summary and provides it as input to the narrator and the event parser. Our prompt encourages contextual information on different utilization of spaces, such that the description delivers the overall structure. The event parser provides accurate 3D scene grounding to the low-level motion synthesis module, such that it can generate fine-grained and realistic interaction. Inspired by previous studies [24, 34, 36, 44], the event parser employs spatial reasoning tools for LLM and specifies a relevant location within a programming framework. Furthermore, we propose an area-conditioned position sampling technique to compensate for the weakness of LLMs in precise localization at the coordinate level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Our framework successfully generates long-term motions with more than 4-5 characters in various multi-room scale scenes, and is robust to the choice of LLM engines. We also propose evaluation criteria for scene-aware multi-agent behavior planning tasks and create a benchmark to assess the compatibility of our framework comprehensively. The extensive results demonstrate that our proposed pipeline outperforms in generating a natural and plausible" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.907, + 0.138 + ], + "angle": 0, + "content": "sequence of actions with adequate trajectories at appropriate relative timing in a shared space, especially for larger scenes with more number of characters." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.154, + 0.665, + 0.17 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.181, + 0.908, + 0.591 + ], + "angle": 0, + "content": "Contextual Human Motion Synthesis Synthesizing human motions is a long-standing research topic in the fields of computer vision and graphics. In particular, an increasing number of studies have incorporated contextual information in motion synthesis, such as a 3D environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35]. Although these studies have shown notable progress in reproducing physically accurate and high-fidelity motions for a single type of interaction, most of them do not consider more than one contextual constraint or expand to long-term sequences of actions. Some recent works [3, 5] have explored motion generation considering both humans and scenes of contextual information. Digital Life Project [3] builds autonomous characters with social intelligence in a 3D scene, while Sitcom-Crafter [5] introduces a unified motion generation framework that integrates different types of interaction motions with plot-driven guidance. The research expanded on previous boundaries, yet it did not take into account the coordination of interactions between more than two characters. Furthermore, in terms of scene understanding, these models lack the complexity needed for scenarios requiring detailed spatial reasoning because of their simplistic scene descriptions. Compared to these studies, our framework demonstrates successful multi-agent contextual behavior planning at a larger scale, including an increased number of characters and a scene with greater size and complexity." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.614, + 0.909, + 0.903 + ], + "angle": 0, + "content": "LLM-based Planning In recent years, there has been growing attention to using pre-trained LLMs in various domains for their impressive zero-shot reasoning capabilities [8, 16, 43]. Particularly in the field of robotics and embodied AI, many recent works have successfully demonstrated the abilities of LLMs for task planning [11, 15, 20, 22, 24, 28, 30, 34, 36]. However, robotic applications primarily focus on goal-oriented tasks, which do not require complex contextual reasoning. While the majority of these works are limited to single-agent scenarios, a few studies [20, 46] have explored LLM-based multiagent task planning recently. CoELA [46] introduces a novel cognitive-inspired modular framework that enables cooperation and communication between multiple agents. However, since CoELA requires continuous message exchanges between agents for behavior coordination, this can result in inefficiency when scaling to an increased number of agents. On the other hand, SMART-LLM [20] demonstrated multi-agent task planning in larger-scale scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11655" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.502, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.237, + 0.485, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Overview of our framework. The framework consists of two main modules: a high-level action planning module and a low-level motion synthesis module. The action planning module coordinates and plans multi-character behavior as a sequence of events. The motion synthesis module receives a detailed description of the event and synthesizes 3D motions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.33, + 0.484, + 0.435 + ], + "angle": 0, + "content": "SMART-LLM accomplishes multi-robot cooperation by decomposing tasks into several sub-stages. However, SMART-LLM represents a scene as a mere list of objects, leading to a lack of comprehensive scene understanding. Distinct from earlier approaches, ours tackles the problem of multi-agent planning while maintaining scalability and in-depth scene understanding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.451, + 0.182, + 0.466 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.477, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Given a 3D scene \\(S\\), characters \\(\\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\}\\), and optional user instructions \\(\\mathcal{T}\\), our framework generates 3D multi-human motions \\(\\mathcal{M} = \\{M_1, M_2, \\dots, M_N\\}\\) within the scene. Here, \\(N\\) represents the number of characters and is determined by the user at the system initialization. The 3D scene \\(S\\) is provided with instance segmentation, and can be in various representations including VR scenes with CAD models or real-world scenes composed of 3D point cloud scans or mesh reconstructions. The user instructions \\(\\mathcal{T}\\) are provided in free text forms, leveraging the generalizability of LLM engines." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.825 + ], + "angle": 0, + "content": "In an overview, our framework operates around an intermediate representation called event \\( e \\). Figure 2 contains examples of textual descriptions of events along with detailed action labels and positions for the motion synthesis module. The events are generated in sequence to realize the given instruction while coordinating with the surrounding environments. Each event has its own lifecycle with a different time span, and multiple events can coexist within the 3D scene. Through the event-based formulation, our system can generate scene dynamics with infinitely long-horizon, and adaptively process user instructions at an arbitrary time during operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.901 + ], + "angle": 0, + "content": "As shown in Figure 2, our framework consists of two main components: a high-level action planning module and a low-level motion synthesis module. The action planning module performs planning on an event basis, driving the scene's timeline forward (Section 3.1). The motion syn" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.906, + 0.137 + ], + "angle": 0, + "content": "thesis module takes the generated events as input, assigns them to the corresponding characters, and creates motions that reflect the assigned event (Section 3.2)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.147, + 0.82, + 0.163 + ], + "angle": 0, + "content": "3.1. High-level Action Planning Module" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.169, + 0.907, + 0.41 + ], + "angle": 0, + "content": "The primary function of the action planning module is to generate new events by comprehensively considering the given 3D scene \\( S \\), existing event history \\( \\mathcal{H} \\), and the behaviors of multiple characters within the scene. To effectively satisfy the holistic requirements, we rely on the powerful reasoning capabilities of LLMs. As shown in Figure 3, there are three LLM submodules: the scene descriptor, the narrator, and the event parser. The scene descriptor generates a textual scene description \\( \\mathcal{D} \\) from the given 3D scene \\( S \\) such that our planning modules can understand the necessary context from the spatial arrangement. The narrator generates a sequence of events, which are converted into detailed information by the event parser. Each module is assigned to smaller, well-defined tasks such that the system stays performant and scalable for multi-human motion generation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.418, + 0.671, + 0.431 + ], + "angle": 0, + "content": "3.1.1. Scene Descriptor" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.437, + 0.909, + 0.77 + ], + "angle": 0, + "content": "In the preprocessing stage before the motion planning, the scene descriptor generates a textual description \\(\\mathcal{D}\\) of the given 3D scene \\(S\\). We use a 3D scene graph, which may lead an LLM module to absorb rich spatial relationships [14]. The scene descriptor first extracts a 3D scene graph in an automated way as proposed in [17] and converts it into textual data in JSON format. Then, we use it to prompt the LLM module and generate a scene description. In addition to the explicit pairwise relationships in the scene graph, further contextual layouts with regional information enhance the naturalness of scene-aware motions. To enhance contextual reasoning, we encourage our scene descriptor to discover a spatial arrangement of objects in proximity, which semantically creates a functional space, for example, a dining area or study zone. Specifically, we extract a regional cluster of objects through the DBSCAN algorithm [9] as an additional input and promote detecting areas of interest. We also provide concrete examples of the same input and output format, allowing the scene descriptor to perform the given task more effectively through in-context learning [2]. We provide additional details and full prompts in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.625, + 0.79 + ], + "angle": 0, + "content": "3.1.2. Narrator" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As the core of our behavior planning, the narrator generates a sequence of events \\(\\mathcal{E} = \\{e_1,e_2,\\ldots \\}\\). Given the scene description \\(\\mathcal{D}\\) from the scene descriptor and the optional user instruction \\(\\mathcal{T}\\), the narrator keeps the previous event history \\(\\mathcal{H}\\) and generates one event at a time based on the high-level context. Focusing on one event at a time, the subsequent module can alleviate the burden of holistic spatio-temporal" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11656" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.087, + 0.891, + 0.215 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.314, + 0.223, + 0.68, + 0.236 + ], + "angle": 0, + "content": "Figure 3. Overview of the high-level action planning module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.251, + 0.482, + 0.31 + ], + "angle": 0, + "content": "reasoning, making the storytelling scalable and interactive. The event-based design also allows the incorporation of additional user instructions between events to refine the storyline." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.312, + 0.484, + 0.538 + ], + "angle": 0, + "content": "As shown in Figure 2, events are expressed in seminarrative natural language, focusing on contextual description. The textual description is further parsed in the subsequent event parser, relieving the burden of detailed spatiotemporal localization. The narrator also receives the state of a generated event as one of two states, 'ongoing' or 'completed'. From this, the narrator can compactly conceive the necessary information to progress the timeline. Our event representation is particularly efficient for group events with multiple characters, which makes the broad scene context easily understood. The narrator coordinates the high-level behaviors of multiple characters and generates the storyline from a macroscopic perspective. The detailed spatiotemporal progression of individual characters is assigned to the event parser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.54, + 0.484, + 0.706 + ], + "angle": 0, + "content": "Similar to the scene descriptor, we also incorporate several widely used prompting techniques in the narrator's prompt, the in-context learning and chain-of-thought reasoning [41]. For chain-of-thought reasoning, we guide the narrator to first reason about the current planning state and other characters' statuses before generating an event. We provide concrete examples of event generation with such a reasoning process to reinforce structured decision-making and improve the coherence of generated events. Further details and the full prompts are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.715, + 0.228, + 0.729 + ], + "angle": 0, + "content": "3.1.3.Event Parser" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.902 + ], + "angle": 0, + "content": "The event parser takes the event \\( e \\) generated by the narrator as input and concretizes it into low-level details that the motion synthesis module can digest. The output event of the narrator is designed to be a short description of the region and the character for the motion, leaving ambiguities on the exact location within the scene. The event parser observes the current state of the scene and refines the event description into detailed 3D grounding suitable for motion synthesis. Specifically, it converts the event description \\( e \\) into \\( e = (\\mathcal{C}_e, \\{p_i\\}, \\{d_i\\}, \\{a_i\\}) \\), where \\( \\mathcal{C}_e = \\{c_i\\} \\) is the set of involved characters, \\( p_i \\in \\mathbb{R}^2 \\) is the target position," + }, + { + "type": "image", + "bbox": [ + 0.513, + 0.253, + 0.911, + 0.463 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.469, + 0.907, + 0.54 + ], + "angle": 0, + "content": "Figure 4. An example of a programming-structured prompt used in our event parser. The event parser leverages provided spatial reasoning tools, such as the get_distance_between() function in this example, to procedurally deduce solutions for given spatial reasoning tasks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.584 + ], + "angle": 0, + "content": "\\(d_{i}\\in \\mathbb{R}^{1}\\) is the target orientation, and \\(a_{i}\\) is the target action label for each character \\(c_{i}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Inspired by [24, 34, 44], we adopt a Python programming structure in the event parser's prompt and utilize spatial reasoning tools that are provided as a form of Python functions. The programming framework can infer solutions in a well-defined sequence of inquiries, leading to more powerful spatial reasoning capabilities. Figure 4 shows a specific example. The event parser uses the get_distance_between() function to deduce spatial details missing from the scene description, such as identifying the farthest seat from a given reference point. The event parser compiles the output format by calling the get_character() function, and for each character \\( c_i \\), it specifies \\( p_i \\), \\( d_i \\), and \\( a_i \\) using the set_position(), set_ORIENTATION(), and set_target_action() functions, respectively. If a specific target orientation is not necessary, it can be left unspecified. When retrieving the positions, the event parser first outputs semantic descriptions using scene objects, and precise coordinates are found afterward to compensate for possible errors of LLMs in coordinate-level reasoning [7, 27]. Specifically, we define a position by its spatial relationship with an anchor object" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11657" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.089, + 0.865, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.283, + 0.234, + 0.713, + 0.248 + ], + "angle": 0, + "content": "Figure 5. Qualitative results of our framework across diverse 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.263, + 0.482, + 0.338 + ], + "angle": 0, + "content": "and sample the coordinate representation \\( p_i \\) within the corresponding area. Then, if a target orientation is specified, we calculate \\( d_i \\) as the object or character the subject faces after sampling \\( p_i \\). Otherwise, \\( d_i \\) is set to the character's expected forward direction upon reaching \\( p_i \\)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.339, + 0.483, + 0.414 + ], + "angle": 0, + "content": "As with the narrator, the event parser incorporates in-context learning, chain-of-thought reasoning, and self-feedback techniques in its prompt. Further details about sampling and the full prompts are provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.425, + 0.403, + 0.44 + ], + "angle": 0, + "content": "3.2. Low-level Motion Synthesis Module" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.447, + 0.483, + 0.642 + ], + "angle": 0, + "content": "After the high-level action planning module provides the narrowed location and character action to generate, the subsequent motion synthesis module concentrates solely on converting the set of events into character motions in a 3D scene in an online manner, detached from the holistic flow of the scene. Given the positions and actions of individual characters, the motion synthesis may employ any existing framework to reach the location and perform the designated action. An event is completed and removed from the scene when all characters in \\(\\{c_i\\}\\) complete their own target actions \\(a_{i}\\). We briefly describe our motion synthesis module below, and more details are available in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.795 + ], + "angle": 0, + "content": "We first plan a collision-free trajectory and generate locomotion along it such that the character \\( c_{i} \\) moves to the target position \\( p_{i} \\), aligning its facing direction with the target orientation \\( d_{i} \\). While reaching the target position \\( p_{i} \\), the character should avoid collisions with other characters and the surrounding environment. We first create a 2D grid map that represents navigable areas in the scene, following the approach used in [38]. On this grid map, our motion synthesis module finds collision-free multi-agent paths using the windowed cooperative \\( A^{*} \\) algorithm proposed in [33]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.901 + ], + "angle": 0, + "content": "After reaching the target position, the motion synthesis module creates motions assigned with the action label \\( a_{i} \\). We leverage the motion matching [6] to quickly generate plausible motions. If \\( p_i \\) is deliberately set to overlap with an object (e.g., on a chair), the motion synthesis module first makes a transition to a desired posture (e.g., sitting) before creating motions for \\( a_{i} \\). There are also group events," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.906, + 0.354 + ], + "angle": 0, + "content": "where multiple characters are involved, such as chatting or handshake. Because temporal synchronization is critical, the characters who reach \\( p_i \\) before others maintain an idle motion and wait for the remaining characters. Motion synthesis for \\( a_i \\) occurs only after all characters have reached their respective target position \\( p_i \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.646, + 0.386 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.395, + 0.906, + 0.516 + ], + "angle": 0, + "content": "The key functionality of our framework is to perform LLM-based scene-aware multi-agent planning to enable natural interactions and activities of characters within a given 3D scene. We can successfully populate natural motions of multiple characters within diverse scenes ranging from virtual scenes composed of CAD models (Figure 1) to 3D scans (Figure 5). More qualitative results are available in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.906, + 0.682 + ], + "angle": 0, + "content": "To the best of our knowledge, our framework is the first to address this task at this scale; there is no existing baseline for direct comparison with our framework. Therefore, we propose evaluation criteria to comprehensively assess our framework (Section 4.1) and create a benchmark designed for this assessment (Section 4.2). Using the developed benchmark, we evaluate the effectiveness of our proposed key methodologies through an ablation study (Section 4.3). Furthermore, we conduct a user study to visually review and compare the generated results across the different ablation settings (Section 4.4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.694, + 0.702, + 0.708 + ], + "angle": 0, + "content": "4.1. Evaluation Criteria" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.716, + 0.905, + 0.775 + ], + "angle": 0, + "content": "We propose a set of criteria to assess several capabilities required for the task as shown below. We define each evaluation criterion as a tag and use these tags to categorize test cases in the benchmark." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.87 + ], + "angle": 0, + "content": "- Object arrangement reasoning (OA): Object arrangement reasoning is essential for identifying objects based on spatial relationships (e.g., the chair farthest from the window) or ensuring that characters interact correctly with their surroundings (e.g., to study at a desk, a character needs to sit on the associated chair)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "- Regional context reasoning \\((RC)\\): Regional context reasoning is required for generating character behaviors that" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11658" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.473, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.483, + 0.278 + ], + "angle": 0, + "content": "Figure 6. An example of a test case. In this test case, a system is required to identify the chairs in the kitchen and select one that is not currently occupied." + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.291, + 0.483, + 0.336 + ], + "angle": 0, + "content": "are well-aligned with the contextual meaning of a space (e.g., cooking in the kitchen rather than in the living room)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.337, + 0.483, + 0.396 + ], + "angle": 0, + "content": "- Scene state reasoning (SS): Scene state reasoning requires considering the overall scene flow and the current state of other characters (e.g., if a character is using the coffee machine, another character should wait to use it)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.473 + ], + "angle": 0, + "content": "In addition to the holistic spatial context defined above, we separately evaluate the low-level performance on Position inference \\((PI)\\). The cases with the tag analyze if the system can find the precise positions of characters at the coordinate level." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.484, + 0.29, + 0.498 + ], + "angle": 0, + "content": "4.2. Benchmark Creation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.687 + ], + "angle": 0, + "content": "Test Scenes We create three scenes—House, Office, and Restaurant—by placing the objects from the HSSD dataset [21]. Each scene is designed to include two to three separate areas with distinct contextual meaning. To ensure the LLM infers contextual meaning of areas without relying on object labels, we replace several object labels with general ones (e.g., kitchen_cabinet \\(\\rightarrow\\) cabinet). We also randomize indices for identically labeled objects to eliminate index-based arrangement cues. We define a distinct set of available action labels \\(A_{S}\\) for each scene, where \\(S \\in \\{\\text{house}, \\text{office}, \\text{restaurant}\\}\\). This \\(A_{S}\\) is utilized when determining the pass/fail of test cases that use the scene \\(S\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Test Cases Our framework can create natural behaviors for multiple characters without any pre-scripted scenarios. However, for rigorous evaluation, we introduce specific test scenarios in our benchmark. Each test case consists of a scene \\( S \\), characters \\( \\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\} \\), test scenario \\( \\mathcal{T} \\), mock planning history \\( \\mathcal{H} \\), and expected plan \\( \\tau_e = (\\mathcal{A}_e, \\tilde{p}_e) \\). Here, \\( N \\) is the number of characters, \\( \\mathcal{A}_e \\) is the set of expected action labels, and \\( \\tilde{p}_e \\) is the expected area where the target character's position is expected to be included. \\( \\tilde{p}_e \\) is expressed through the union or intersection of multiple areas, each defined as a semantic description based on scene objects. For simplicity, we ignore the character's orientation in the benchmark, and the target character for the expected" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.122 + ], + "angle": 0, + "content": "plan \\(\\tau_{e}\\) is always fixed to the first character \\(c_{1}\\) in all test cases. An example of a test case is depicted in Figure 6." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.125, + 0.907, + 0.35 + ], + "angle": 0, + "content": "When running a test case, we input the test scenario \\(\\mathcal{T}\\) and mock planning history \\(\\mathcal{H}\\) into the system. The mock planning history \\(\\mathcal{H}\\) simulates the prior planning history of the scene, building the necessary context required for the test case. Then, the system generates the most appropriate next plan \\(\\tau_{g} = (a_{g},\\tilde{p}_{g}\\) or \\(p_g)\\) by evaluating the given mock planning history and the test scenario. Here, \\(\\tau_{g}\\) is the generated next plan for the target character, \\(a_{g}\\) is the planned action label, and \\(\\tilde{p}_g\\) or \\(p_g\\) is the planned position. If the test case evaluates coordinate-level localization, the system outputs a coordinate-level position \\(p_{g}\\). Otherwise, it outputs an area-level position \\(\\tilde{p}_g\\). As the final step, the test case's pass/fail is determined by comparing the expected plan \\(\\tau_{e}\\) and the generated plan \\(\\tau_{g}\\). For a test case to pass, the following conditions must all be satisfied:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.353, + 0.882, + 0.369 + ], + "angle": 0, + "content": "1. A plan for the target character is generated \\((\\tau_{g}\\neq \\emptyset)\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.369, + 0.905, + 0.399 + ], + "angle": 0, + "content": "2. The action label is included in both the available action labels and the expected action labels \\((a_{g}\\in \\mathcal{A}_{\\mathcal{S}}\\cap \\mathcal{A}_{e})\\)" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.399, + 0.905, + 0.43 + ], + "angle": 0, + "content": "3. The character's position is within the expected area \\((\\tilde{p}_g \\subseteq \\tilde{p}_e\\) or \\(p_g \\in \\tilde{p}_e)\\)." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.353, + 0.905, + 0.43 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.907, + 0.718 + ], + "angle": 0, + "content": "Benchmark Summary We construct the benchmark based on the previously defined evaluation criteria, namely, tags. Our benchmark consists of 40 test cases in total. 10 test cases are designed for the \\( PI \\) tag, while the remaining 30 test cases involve \\( OA \\), \\( RC \\), and \\( SS \\) tags. The test cases for the \\( PI \\) tag deliberately use straightforward scenarios to focus solely on the localization performance. Among the latter 30 test cases, half incorporate two tags for increased complexity. Additionally, for the latter 30 test cases, we vary the number of characters in the scene and the size of the mock planning history across three levels within the same scenario. This expands the number of effective test cases and allows us to evaluate the system's scalability by assessing its robustness against increasing agent counts and planning history complexity. We design our benchmark to ensure that each tag is evenly represented, preventing bias toward any specific evaluation factor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Evaluation Metrics Our benchmark employs two evaluation metrics: success rate and execution rate. The success rate measures the proportion of passed runs among all test case runs. On the other hand, the execution rate measures the proportion of runs that are executed without any runtime errors, which are included in the parenthesis next to the success rate in tables. Runtime errors typically occur when the LLM module references a nonexistent object or generates a response with incorrect syntax. Any test case run with a runtime error is automatically marked as failed." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11659" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.905, + 0.24 + ], + "angle": 0, + "content": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.9 (0.98)0.93 (0.99)0.9 (0.98)0.92 (0.98)0.74 (0.96)0.72 (0.95)0.72 (0.97)0.78 (0.93)0.72 (0.87)0.78 (0.91)0.76 (0.87)0.61 (0.81)
w/o Event0.82 (0.92)0.88 (0.92)0.86 (0.93)0.77 (0.92)0.6 (0.95)0.56 (0.99)0.6 (0.92)0.54 (0.92)0.6 (0.81)0.6 (0.78)0.62 (0.86)0.54 (0.76)
Object List0.51 (0.85)0.61 (0.78)0.28 (0.88)0.65 (0.97)0.34 (0.88)0.37 (0.88)0.12 (0.86)0.51 (0.9)0.35 (0.68)0.31 (0.7)0.29 (0.61)0.41 (0.69)
Scene Graph0.82 (0.96)0.8 (0.96)0.82 (0.94)0.87 (0.95)0.49 (0.9)0.52 (0.93)0.32 (0.9)0.51 (0.83)0.72 (0.92)0.69 (0.91)0.77 (0.94)0.64 (0.85)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.35 (0.74)0.37 (0.8)0.37 (0.77)0.35 (0.67)0.71 (0.9)0.71 (0.91)0.72 (0.95)0.65 (0.85)0.39 (0.87)0.38 (0.88)0.36 (0.87)0.35 (0.85)
w/o Event0.31 (0.58)0.3 (0.57)0.32 (0.6)0.28 (0.63)0.62 (0.85)0.74 (0.85)0.65 (0.9)0.59 (0.85)0.35 (0.79)0.41 (0.86)0.32 (0.79)0.32 (0.74)
Object List0.22 (0.58)0.26 (0.65)0.13 (0.62)0.23 (0.44)0.44 (0.77)0.45 (0.77)0.36 (0.83)0.52 (0.72)0.14 (0.77)0.09 (0.74)0.07 (0.77)0.24 (0.67)
Scene Graph0.29 (0.75)0.31 (0.76)0.35 (0.86)0.16 (0.61)0.69 (0.84)0.69 (0.81)0.77 (0.93)0.55 (0.72)0.31 (0.84)0.27 (0.88)0.23 (0.84)0.35 (0.84)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.907, + 0.278 + ], + "angle": 0, + "content": "Table 1. Benchmark result for test cases with object arrangement reasoning (OA), regional context reasoning (RC), and scene state reasoning (SS) tags. In total, our method achieves the highest success rate across various LLM models and sizes." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.291, + 0.46, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.425, + 0.483, + 0.495 + ], + "angle": 0, + "content": "Figure 7. Benchmark results by number of characters. Our approach exhibits stable success rate for stories with different number of characters while the performance gap increases as more characters are involved, compared to the version without event representation." + }, + { + "type": "image", + "bbox": [ + 0.111, + 0.497, + 0.455, + 0.587 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.588, + 0.483, + 0.643 + ], + "angle": 0, + "content": "Figure 8. Comparison of average token usage. Object List uses the least number of tokens, and Scene Graph uses significantly more number of tokens, especially when the scene complexity increases (Office)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.658, + 0.279, + 0.673 + ], + "angle": 0, + "content": "4.3. Benchmark Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.68, + 0.483, + 0.773 + ], + "angle": 0, + "content": "We use GPT-4o [16], Llama-3.1 [8], and Qwen2.5 [43] in various sizes as LLM backbones to obtain benchmark results. The backbones are selected to represent both commercial and open-source modules that are popular and demonstrate competitive performance. We repeat each test case five times and average the results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Event-Driven Planning Table 1 presents the benchmark results for test cases with \\(OA\\), \\(RC\\), and \\(SS\\) tags. Each column represents the average score for test cases that include the corresponding tag and each cell displays the success rate, with the execution rate shown in parentheses. The ablation without event-driven planning (w/o Event) handles the entire process with a single unified LLM module without using high-level event representations. Therefore, all in" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.908, + 0.55 + ], + "angle": 0, + "content": "put and output information, except for the scene description and test scenario, is represented in a low-level structured format. Compared to \\( w/o \\) Event, our method benefits from a cascaded planning pipeline that effectively decomposes the entire planning complexity. This leads to an overall improvement in planning performance. In particular, the performance gain in SS is relatively larger compared to other tags in GPT-40 models. The SS tag requires a comprehensive understanding of the overall planning state of multiple characters in the scene. We attribute this to the advantages of our event-based representation, which provides a more efficient and intuitive way to capture the overall scene state. Additionally, as shown in Figure 7, our planning pipeline proves to be more effective in terms of scalability. \\( w/o \\) Event struggles with planning as the number of characters in the scene increases, whereas our approach maintains stable performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.909, + 0.856 + ], + "angle": 0, + "content": "Scene Description We also analyze the effect of using scene description generated from an LLM module. We create ablation settings that retain our pipeline but modify the scene description method. Object List represents the scene as a simple list of objects present in the scene, while Scene Graph uses the JSON-formatted scene graph representation as the scene description. Our approach contributes to more effective planning compared to other description methods as shown in Table 1. Additionally, we present the average token usage for each description method in Figure 8 to provide a deeper understanding of their differences. As presented in Figure 8, the object list is the most efficient way to represent scene information and has often been used in previous studies such as [5, 20, 34]. However, since the object list does not have any spatial information, the planning module with the description is ignorant of object arrangement or spatial layout. As a result, the object list records the lowest scores in our benchmark, which requires sophisticated scene understanding for planning." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "Unlike the object list, the scene graph representation enables necessary spatial reasoning by incorporating detailed attributes of objects and spatial relationships between ob" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11660" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.124, + 0.089, + 0.454, + 0.173 + ], + "angle": 0, + "content": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
Ours0.94 (1.0)0.9 (1.0)0.94 (1.0)
Direct Inference0.46 (1.0)0.38 (1.0)0.28 (1.0)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
Ours0.86 (0.88)1.0 (1.0)0.94 (1.0)
Direct Inference0.14 (0.9)0.14 (0.48)0.1 (0.8)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.226 + ], + "angle": 0, + "content": "Table 2. Benchmark result for test cases with position inference tag (PI). Our proposed method can stably infer accurate positions, whereas Direct Inference with LLMs often struggles." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.235, + 0.486, + 0.446 + ], + "angle": 0, + "content": "jects. However, the representation contains exhaustive detail, including even minor and insignificant ones, without any abstraction or compression. This inefficiency leads to excessive token consumption (Figure 8) and can even interfere with the LLM's reasoning due to unnecessary details (Table 1). The token inefficiency becomes more pronounced as the scene size and complexity increase. As shown in Figure 8, when comparing the relatively simple 'House' scene to the more complex 'Office' scene, the scene graph exhibits a significantly larger increase in token consumption compared to alternative approaches. Compared to the scene graph, our approach allows for a more efficient description of the given scene while preserving key spatial information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.458, + 0.484, + 0.641 + ], + "angle": 0, + "content": "Positional Inference (PI) In Table 2, we demonstrate the effectiveness of our area-conditioned position sampling method based on the results for the PI-tagged test cases. Direct Inference is an ablation setting where the same planning pipeline is used, but the LLM directly outputs coordinates when determining character positions. Although all test cases with the PI tag are intentionally designed to be very simple, all LLM models with Direct Inference struggle to accurately determine the coordinates of character positions. In contrast, our system allows the LLM to process information semantically, and obtains precise locations through the proposed area-conditioned position sampling technique." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.652, + 0.274, + 0.668 + ], + "angle": 0, + "content": "4.4. User Study Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We also validate through a user study that our approach represents the given scenario more effectively than other ablation settings. For the user study, we use a total of four test scenarios, each set in a different scene. The scenes include House, Office, and Restaurant from the benchmark, as well as MPH11 from the PROX dataset [12]. Compared to the other scenes, MPH11 is relatively simple in both scale and complexity. Given its limited space, we use a simple scenario containing only two events for the MPH11. In contrast, for other test scenes, we use more complex scenarios that include 4-5 events. For each test scenario, users are asked to visually compare results generated from different ablation settings and select the one that best represents the given scenario. If multiple results are perceived as equally effective, users are allowed to choose more than one. We" + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.088, + 0.885, + 0.165 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.533, + 0.169, + 0.885, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.298 + ], + "angle": 0, + "content": "Figure 9. User study results. Our system correctly generates the dynamic story line with multiple characters in large scenes, significantly outperforming other baselines, except for simple scenarios with MPH11." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.315, + 0.905, + 0.344 + ], + "angle": 0, + "content": "generate the results for the user study using GPT-4o mini with a temperature setting of 0.0." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.346, + 0.907, + 0.527 + ], + "angle": 0, + "content": "We collected responses from 50 general participants. The responses to the user study are summarized in Figure 9. The chart labels SF, OL, and No E correspond to the same ablation settings used in the benchmark: Scene Graph, Object List, and w/o Event, respectively. In three of our scenarios, our approach received the most selections by a large margin. Furthermore, the noticeable difference in selection distribution between the MPH11 scene, which involves a relatively smaller-scale problem, and the other scenes suggests that our approach contributes to building a more scalable system. We provide the test scenarios and visualization of generated results in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.543, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.907, + 0.903 + ], + "angle": 0, + "content": "In this study, we propose an LLM-based framework for generating multiple lifelike virtual humans within a given 3D scene, dynamically aligning their behaviors with an emergent storyline. Characters generated by our framework interact contextually with their surroundings and other individuals, enhancing scene realism. Our framework performs multi-character behavior planning based on semi-narrative events, allowing the LLM to efficiently capture scene progression and group behaviors. The modular design of our planning pipeline decomposes the overall problem, reducing reasoning complexity for the LLM and enabling more scalable planning. Additionally, we introduce a context-centric scene description to efficiently convey key spatial information and propose area-conditioned position sampling to mitigate the LLM's weaknesses in numerical reasoning. To evaluate our system, we construct a benchmark for a comprehensive assessment of LLM-based scene-aware multi-agent behavior planning. Through benchmark results and a user study, we demonstrate that our proposed methods contribute to building a scalable and effective system for coordinating multi-agent behaviors, while carefully considering the scene." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11661" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.184 + ], + "angle": 0, + "content": "Acknowledgments This work was supported by the IITP grant [RS-2021-II211341, Artificial Intelligence Graduate School Program (Chung-Ang University)], the NRF grant (No. RS-2023-00208197) funded by the Korea government (MSIT), and Creative-Pioneering Researchers Program through Seoul National University." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.194, + 0.188, + 0.21 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.219, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[1] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.303, + 0.484, + 0.371 + ], + "angle": 0, + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.373, + 0.484, + 0.455 + ], + "angle": 0, + "content": "[3] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.457, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[4] Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Generating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1855-1866, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.54, + 0.482, + 0.607 + ], + "angle": 0, + "content": "[5] Jianqi Chen, Panwen Hu, Xiaojun Chang, Zhenwei Shi, Michael Christian Kampffmeyer, and Xiaodan Liang. Sitcom-crafter: A plot-driven human motion generation system in 3d scenes. arXiv preprint arXiv:2410.10790, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.609, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[6] Simon Clavet et al. Motion matching and the road to next-gen animation. In Proc. of GDC, page 4, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.638, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[7] Arash Gholami Davoodi, Seyed Pouyan Mousavi Davoudi, and Pouya Pezeshkpour. Llms are not intelligent thinkers: Introducing mathematical topic tree benchmark for comprehensive evaluation of llms. arXiv preprint arXiv:2406.05194, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.707, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[9] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[10] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision, pages 418-437. Springer, 2024. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.219, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[11] Maitrey Gramopadhye and Daniel Szafir. Generating executable action plans with environmentally-aware language models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3568-3575. IEEE, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.163, + 0.905, + 0.231 + ], + "angle": 0, + "content": "[12] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.233, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[13] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.383 + ], + "angle": 0, + "content": "[14] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In Proceedings of the 41st International Conference on Machine Learning, pages 20413-20451, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.904, + 0.453 + ], + "angle": 0, + "content": "[15] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.904, + 0.511 + ], + "angle": 0, + "content": "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[17] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In European Conference on Computer Vision, pages 289-310. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[18] Bowen Jiang, Yangxinyu Xie, Zhuoqun Hao, Xiaomeng Wang, Tanwi Mallick, Weijie Su, Camillo Taylor, and Dan Roth. A peek into token bias: Large language models are not yet genuine reasoners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4722-4756, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.665, + 0.905, + 0.72 + ], + "angle": 0, + "content": "[19] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[20] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12140–12147. IEEE, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[21] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11662" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.092, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[22] Boyi Li, Philipp Wu, Pieter Abbeel, and Jitendra Malik. Interactive task planning with language models. arXiv preprint arXiv:2310.10645, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.136, + 0.484, + 0.192 + ], + "angle": 0, + "content": "[23] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, 132(9):3463-3483, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.483, + 0.262 + ], + "angle": 0, + "content": "[24] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.264, + 0.483, + 0.332 + ], + "angle": 0, + "content": "[25] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.334, + 0.483, + 0.39 + ], + "angle": 0, + "content": "[26] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 903-913. IEEE, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.483, + 0.46 + ], + "angle": 0, + "content": "[27] Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.462, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[28] Zhe Ni, Xiaoxin Deng, Cong Tai, Xinyue Zhu, Qinghongbing Xie, Weihang Huang, Xiang Wu, and Long Zeng. Grid: Scene-graph-based instruction-driven robotic task planning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13765-13772. IEEE, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.547, + 0.483, + 0.616 + ], + "angle": 0, + "content": "[29] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1-22, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.618, + 0.483, + 0.673 + ], + "angle": 0, + "content": "[30] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Sunderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.675, + 0.483, + 0.743 + ], + "angle": 0, + "content": "[31] Mengyi Shan, Lu Dong, Yutao Han, Yuan Yao, Tao Liu, Ifeoma Nwogu, Guo-Jun Qi, and Mitch Hill. Towards open domain text-driven synthesis of multi-person motions. In European Conference on Computer Vision, pages 67-86. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[32] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Scharli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[33] David Silver. Cooperative pathfinding. In Proceedings of the aaai conference on artificial intelligence and interactive digital entertainment, pages 117-122, 2005. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.135 + ], + "angle": 0, + "content": "task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 1, 2, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.907, + 0.192 + ], + "angle": 0, + "content": "[35] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[36] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[37] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[38] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[39] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems, 35:14959-14971, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.905, + 0.558 + ], + "angle": 0, + "content": "[40] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.561, + 0.905, + 0.629 + ], + "angle": 0, + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[42] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[43] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.813 + ], + "angle": 0, + "content": "[44] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[45] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision, pages 246-263. Springer, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[46] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11663" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "Gan. Building cooperative embodied agents modularly with large language models. In The Twelfth International Conference on Learning Representations, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[47] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14738-14749, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.205 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11664" + } + ] +] \ No newline at end of file diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_origin.pdf b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..3473dca574e8be9fc223a10735384be3ffdd47ea --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/d626a47c-0fd6-4b74-b622-4b22ef5d348b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2795f3e5f5f5770f09275cd6a362ff511163cee4da8842f9f7092fc3f53ce0d7 +size 12456849 diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/full.md b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/full.md new file mode 100644 index 0000000000000000000000000000000000000000..982e66b4cc29e6051c9b46bfdca63817c3b0ddac --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/full.md @@ -0,0 +1,253 @@ +# Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene + +Donggeun Lim $^{1}$ Jinseok Bae $^{1}$ Inwoo Hwang $^{1}$ Seungmin Lee $^{1}$ Hwanhee Lee $^{2}$ Young Min Kim $^{1\dagger}$ + +$^{1}$ Seoul National University $^{2}$ Chung-Ang University + +$^{1}\{\mathrm{rms}2836$ , capoo95, inusu0818, rsual, youngmin.kim $\} \in$ snu.ac.kr $^{2}$ hwanheelee@cau.ac.kr + +![](images/ba5a4f3b68252eb66586ef36117577f3dce395ad52a54384a8229c3f9223f601.jpg) +Figure 1. Our framework populates 3D scenes with multiple characters. The generated characters interact with their surroundings and other people, bringing the space to life. + +![](images/0a9c2520a27f79d3803801e501ac6ffa2909308936575c3cc8353ed33494e951.jpg) + +![](images/37d24f6da7963b2eab81ec7b6bbdb7b2bdc684b39e18dc750edd078c677799ec.jpg) + +# Abstract + +In this work, we propose a framework that creates a lively virtual dynamic scene with contextual motions of multiple humans. Generating multi-human contextual motion requires holistic reasoning over dynamic relationships among human-human and human-scene interactions. We adapt the power of a large language model (LLM) to digest the contextual complexity within textual input and convert the task into tangible subproblems such that we can generate multi-agent behavior beyond the scale that was not considered before. Specifically, our event generator formulates the temporal progression of a dynamic scene into a sequence of small events. Each event calls for a well-defined motion involving relevant characters and objects. Next, we synthesize the motions of characters at positions sampled based on spatial guidance. We employ a high-level module to deliver scalable yet comprehensive context, translating events into relative descriptions that enable the retrieval of precise coordinates. As the first to address this problem at scale and with diversity, we offer a benchmark to assess diverse aspects of contextual reasoning. Benchmark results and user studies show that our framework effectively captures scene context with high scalability. The code and benchmark, along with result videos, are available + +at our project page: https://rms0329.github.io/Event-Driven-Storytelling/. + +# 1. Introduction + +A scene with digital humans creates lively atmosphere and has a wide range of applications in VR, games, and movies. The realism incurs from natural motions of multiple characters that seamlessly integrate into the surrounding environment. However, existing motion synthesis works focus on individual characters independently or extend to interactions with either the surrounding environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35], remaining limited to a single type of interaction. Hence, they severely lack generalizability and scalability to concurrently generate plausible interactions of multiple characters in a crowded scene. It is highly challenging to correctly assign the behavior of individual characters at the correct time stamp as it requires extensive dynamic contextual and spatial reasoning. + +To fulfill such holistic requirements, we take inspiration from recent advances in large language models (LLMs). LLMs demonstrated phenomenal performance in various areas, especially for high-level planning [11, 15, 24, 30, 34, 36] or behavior modeling [3, 29]. We deduce that LLMs have the potential to interpret nuanced semantics from + +open-ended textual inputs and dependency between natural motion and scene context. However, simply generating commands from the input request may suffer from hallucinations or struggle with localization errors, which are well-known limitations of foundation models [18, 27, 32, 42]. + +We propose an LLM-powered framework that populates multiple lifelike virtual humans within a 3D scene. As shown in Figure 1, the characters produced by our framework can understand and interact with their surroundings, including both 3D scenes and other characters. The proposed LLM module can flexibly understand the holistic context in open-ended scenarios and make high-level decisions. The characters may exhibit emergent yet plausible behavior that was not explicitly defined within the original input, providing a lively atmosphere with rich interactions. The framework can further adapt to user-interactive scenarios, where users may interact with the characters through high-level text instructions. + +Towards more performant and scalable planning with LLMs, our modular framework is deliberately designed to reduce the reasoning complexity. We introduce an event-based planning approach with two LLM modules: narrator and event parser. The narrator progressively weaves the flow of the scene by generating a textual description of one event at a time. Each event considers only a well-defined subset of characters and objects, detached from the collective behavior of the holistic flow. The event parser then transforms high-level event representation into detailed information for existing motion synthesis frameworks. + +We provide devised scene information tailored for individual modules to guide necessary spatial reasoning. Our scene descriptor converts the 3D scene graph into a textual summary and provides it as input to the narrator and the event parser. Our prompt encourages contextual information on different utilization of spaces, such that the description delivers the overall structure. The event parser provides accurate 3D scene grounding to the low-level motion synthesis module, such that it can generate fine-grained and realistic interaction. Inspired by previous studies [24, 34, 36, 44], the event parser employs spatial reasoning tools for LLM and specifies a relevant location within a programming framework. Furthermore, we propose an area-conditioned position sampling technique to compensate for the weakness of LLMs in precise localization at the coordinate level. + +Our framework successfully generates long-term motions with more than 4-5 characters in various multi-room scale scenes, and is robust to the choice of LLM engines. We also propose evaluation criteria for scene-aware multi-agent behavior planning tasks and create a benchmark to assess the compatibility of our framework comprehensively. The extensive results demonstrate that our proposed pipeline outperforms in generating a natural and plausible + +sequence of actions with adequate trajectories at appropriate relative timing in a shared space, especially for larger scenes with more number of characters. + +# 2. Related Works + +Contextual Human Motion Synthesis Synthesizing human motions is a long-standing research topic in the fields of computer vision and graphics. In particular, an increasing number of studies have incorporated contextual information in motion synthesis, such as a 3D environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35]. Although these studies have shown notable progress in reproducing physically accurate and high-fidelity motions for a single type of interaction, most of them do not consider more than one contextual constraint or expand to long-term sequences of actions. Some recent works [3, 5] have explored motion generation considering both humans and scenes of contextual information. Digital Life Project [3] builds autonomous characters with social intelligence in a 3D scene, while Sitcom-Crafter [5] introduces a unified motion generation framework that integrates different types of interaction motions with plot-driven guidance. The research expanded on previous boundaries, yet it did not take into account the coordination of interactions between more than two characters. Furthermore, in terms of scene understanding, these models lack the complexity needed for scenarios requiring detailed spatial reasoning because of their simplistic scene descriptions. Compared to these studies, our framework demonstrates successful multi-agent contextual behavior planning at a larger scale, including an increased number of characters and a scene with greater size and complexity. + +LLM-based Planning In recent years, there has been growing attention to using pre-trained LLMs in various domains for their impressive zero-shot reasoning capabilities [8, 16, 43]. Particularly in the field of robotics and embodied AI, many recent works have successfully demonstrated the abilities of LLMs for task planning [11, 15, 20, 22, 24, 28, 30, 34, 36]. However, robotic applications primarily focus on goal-oriented tasks, which do not require complex contextual reasoning. While the majority of these works are limited to single-agent scenarios, a few studies [20, 46] have explored LLM-based multiagent task planning recently. CoELA [46] introduces a novel cognitive-inspired modular framework that enables cooperation and communication between multiple agents. However, since CoELA requires continuous message exchanges between agents for behavior coordination, this can result in inefficiency when scaling to an increased number of agents. On the other hand, SMART-LLM [20] demonstrated multi-agent task planning in larger-scale scenarios. + +![](images/44d0ecc5690a7c89cf5be99672997194baeae64dc3d4fa22c53a2786ecbd6889.jpg) +Figure 2. Overview of our framework. The framework consists of two main modules: a high-level action planning module and a low-level motion synthesis module. The action planning module coordinates and plans multi-character behavior as a sequence of events. The motion synthesis module receives a detailed description of the event and synthesizes 3D motions. + +SMART-LLM accomplishes multi-robot cooperation by decomposing tasks into several sub-stages. However, SMART-LLM represents a scene as a mere list of objects, leading to a lack of comprehensive scene understanding. Distinct from earlier approaches, ours tackles the problem of multi-agent planning while maintaining scalability and in-depth scene understanding. + +# 3. Method + +Given a 3D scene $S$ , characters $\mathcal{C} = \{c_1, c_2, \dots, c_N\}$ , and optional user instructions $\mathcal{T}$ , our framework generates 3D multi-human motions $\mathcal{M} = \{M_1, M_2, \dots, M_N\}$ within the scene. Here, $N$ represents the number of characters and is determined by the user at the system initialization. The 3D scene $S$ is provided with instance segmentation, and can be in various representations including VR scenes with CAD models or real-world scenes composed of 3D point cloud scans or mesh reconstructions. The user instructions $\mathcal{T}$ are provided in free text forms, leveraging the generalizability of LLM engines. + +In an overview, our framework operates around an intermediate representation called event $e$ . Figure 2 contains examples of textual descriptions of events along with detailed action labels and positions for the motion synthesis module. The events are generated in sequence to realize the given instruction while coordinating with the surrounding environments. Each event has its own lifecycle with a different time span, and multiple events can coexist within the 3D scene. Through the event-based formulation, our system can generate scene dynamics with infinitely long-horizon, and adaptively process user instructions at an arbitrary time during operation. + +As shown in Figure 2, our framework consists of two main components: a high-level action planning module and a low-level motion synthesis module. The action planning module performs planning on an event basis, driving the scene's timeline forward (Section 3.1). The motion syn + +thesis module takes the generated events as input, assigns them to the corresponding characters, and creates motions that reflect the assigned event (Section 3.2). + +# 3.1. High-level Action Planning Module + +The primary function of the action planning module is to generate new events by comprehensively considering the given 3D scene $S$ , existing event history $\mathcal{H}$ , and the behaviors of multiple characters within the scene. To effectively satisfy the holistic requirements, we rely on the powerful reasoning capabilities of LLMs. As shown in Figure 3, there are three LLM submodules: the scene descriptor, the narrator, and the event parser. The scene descriptor generates a textual scene description $\mathcal{D}$ from the given 3D scene $S$ such that our planning modules can understand the necessary context from the spatial arrangement. The narrator generates a sequence of events, which are converted into detailed information by the event parser. Each module is assigned to smaller, well-defined tasks such that the system stays performant and scalable for multi-human motion generation. + +# 3.1.1. Scene Descriptor + +In the preprocessing stage before the motion planning, the scene descriptor generates a textual description $\mathcal{D}$ of the given 3D scene $S$ . We use a 3D scene graph, which may lead an LLM module to absorb rich spatial relationships [14]. The scene descriptor first extracts a 3D scene graph in an automated way as proposed in [17] and converts it into textual data in JSON format. Then, we use it to prompt the LLM module and generate a scene description. In addition to the explicit pairwise relationships in the scene graph, further contextual layouts with regional information enhance the naturalness of scene-aware motions. To enhance contextual reasoning, we encourage our scene descriptor to discover a spatial arrangement of objects in proximity, which semantically creates a functional space, for example, a dining area or study zone. Specifically, we extract a regional cluster of objects through the DBSCAN algorithm [9] as an additional input and promote detecting areas of interest. We also provide concrete examples of the same input and output format, allowing the scene descriptor to perform the given task more effectively through in-context learning [2]. We provide additional details and full prompts in the supplementary material. + +# 3.1.2. Narrator + +As the core of our behavior planning, the narrator generates a sequence of events $\mathcal{E} = \{e_1,e_2,\ldots \}$ . Given the scene description $\mathcal{D}$ from the scene descriptor and the optional user instruction $\mathcal{T}$ , the narrator keeps the previous event history $\mathcal{H}$ and generates one event at a time based on the high-level context. Focusing on one event at a time, the subsequent module can alleviate the burden of holistic spatio-temporal + +![](images/9a71ebeff84852c8a41fd1a76d6edf2c2a11c3255ae0f8c394a9643bfeeadd5f.jpg) +Figure 3. Overview of the high-level action planning module. + +reasoning, making the storytelling scalable and interactive. The event-based design also allows the incorporation of additional user instructions between events to refine the storyline. + +As shown in Figure 2, events are expressed in seminarrative natural language, focusing on contextual description. The textual description is further parsed in the subsequent event parser, relieving the burden of detailed spatiotemporal localization. The narrator also receives the state of a generated event as one of two states, 'ongoing' or 'completed'. From this, the narrator can compactly conceive the necessary information to progress the timeline. Our event representation is particularly efficient for group events with multiple characters, which makes the broad scene context easily understood. The narrator coordinates the high-level behaviors of multiple characters and generates the storyline from a macroscopic perspective. The detailed spatiotemporal progression of individual characters is assigned to the event parser. + +Similar to the scene descriptor, we also incorporate several widely used prompting techniques in the narrator's prompt, the in-context learning and chain-of-thought reasoning [41]. For chain-of-thought reasoning, we guide the narrator to first reason about the current planning state and other characters' statuses before generating an event. We provide concrete examples of event generation with such a reasoning process to reinforce structured decision-making and improve the coherence of generated events. Further details and the full prompts are provided in the supplementary material. + +# 3.1.3.Event Parser + +The event parser takes the event $e$ generated by the narrator as input and concretizes it into low-level details that the motion synthesis module can digest. The output event of the narrator is designed to be a short description of the region and the character for the motion, leaving ambiguities on the exact location within the scene. The event parser observes the current state of the scene and refines the event description into detailed 3D grounding suitable for motion synthesis. Specifically, it converts the event description $e$ into $e = (\mathcal{C}_e, \{p_i\}, \{d_i\}, \{a_i\})$ , where $\mathcal{C}_e = \{c_i\}$ is the set of involved characters, $p_i \in \mathbb{R}^2$ is the target position, + +![](images/2919c4aaa9cd9d79817cb05a7c6eeadf16157fa02b7a6e01f9a6db6a740a196d.jpg) +Figure 4. An example of a programming-structured prompt used in our event parser. The event parser leverages provided spatial reasoning tools, such as the get_distance_between() function in this example, to procedurally deduce solutions for given spatial reasoning tasks. + +$d_{i}\in \mathbb{R}^{1}$ is the target orientation, and $a_{i}$ is the target action label for each character $c_{i}$ . + +Inspired by [24, 34, 44], we adopt a Python programming structure in the event parser's prompt and utilize spatial reasoning tools that are provided as a form of Python functions. The programming framework can infer solutions in a well-defined sequence of inquiries, leading to more powerful spatial reasoning capabilities. Figure 4 shows a specific example. The event parser uses the get_distance_between() function to deduce spatial details missing from the scene description, such as identifying the farthest seat from a given reference point. The event parser compiles the output format by calling the get_character() function, and for each character $c_i$ , it specifies $p_i$ , $d_i$ , and $a_i$ using the set_position(), set_ORIENTATION(), and set_target_action() functions, respectively. If a specific target orientation is not necessary, it can be left unspecified. When retrieving the positions, the event parser first outputs semantic descriptions using scene objects, and precise coordinates are found afterward to compensate for possible errors of LLMs in coordinate-level reasoning [7, 27]. Specifically, we define a position by its spatial relationship with an anchor object + +![](images/9713893b4dc8d03263da5179c1dfdc0ed2aad564f83ae10feb76a45832b8e1ed.jpg) +Figure 5. Qualitative results of our framework across diverse 3D scenes. + +and sample the coordinate representation $p_i$ within the corresponding area. Then, if a target orientation is specified, we calculate $d_i$ as the object or character the subject faces after sampling $p_i$ . Otherwise, $d_i$ is set to the character's expected forward direction upon reaching $p_i$ . + +As with the narrator, the event parser incorporates in-context learning, chain-of-thought reasoning, and self-feedback techniques in its prompt. Further details about sampling and the full prompts are provided in the supplementary material. + +# 3.2. Low-level Motion Synthesis Module + +After the high-level action planning module provides the narrowed location and character action to generate, the subsequent motion synthesis module concentrates solely on converting the set of events into character motions in a 3D scene in an online manner, detached from the holistic flow of the scene. Given the positions and actions of individual characters, the motion synthesis may employ any existing framework to reach the location and perform the designated action. An event is completed and removed from the scene when all characters in $\{c_i\}$ complete their own target actions $a_{i}$ . We briefly describe our motion synthesis module below, and more details are available in the supplementary material. + +We first plan a collision-free trajectory and generate locomotion along it such that the character $c_{i}$ moves to the target position $p_{i}$ , aligning its facing direction with the target orientation $d_{i}$ . While reaching the target position $p_{i}$ , the character should avoid collisions with other characters and the surrounding environment. We first create a 2D grid map that represents navigable areas in the scene, following the approach used in [38]. On this grid map, our motion synthesis module finds collision-free multi-agent paths using the windowed cooperative $A^{*}$ algorithm proposed in [33]. + +After reaching the target position, the motion synthesis module creates motions assigned with the action label $a_{i}$ . We leverage the motion matching [6] to quickly generate plausible motions. If $p_i$ is deliberately set to overlap with an object (e.g., on a chair), the motion synthesis module first makes a transition to a desired posture (e.g., sitting) before creating motions for $a_{i}$ . There are also group events, + +where multiple characters are involved, such as chatting or handshake. Because temporal synchronization is critical, the characters who reach $p_i$ before others maintain an idle motion and wait for the remaining characters. Motion synthesis for $a_i$ occurs only after all characters have reached their respective target position $p_i$ . + +# 4. Experiments + +The key functionality of our framework is to perform LLM-based scene-aware multi-agent planning to enable natural interactions and activities of characters within a given 3D scene. We can successfully populate natural motions of multiple characters within diverse scenes ranging from virtual scenes composed of CAD models (Figure 1) to 3D scans (Figure 5). More qualitative results are available in the supplementary materials. + +To the best of our knowledge, our framework is the first to address this task at this scale; there is no existing baseline for direct comparison with our framework. Therefore, we propose evaluation criteria to comprehensively assess our framework (Section 4.1) and create a benchmark designed for this assessment (Section 4.2). Using the developed benchmark, we evaluate the effectiveness of our proposed key methodologies through an ablation study (Section 4.3). Furthermore, we conduct a user study to visually review and compare the generated results across the different ablation settings (Section 4.4). + +# 4.1. Evaluation Criteria + +We propose a set of criteria to assess several capabilities required for the task as shown below. We define each evaluation criterion as a tag and use these tags to categorize test cases in the benchmark. + +- Object arrangement reasoning (OA): Object arrangement reasoning is essential for identifying objects based on spatial relationships (e.g., the chair farthest from the window) or ensuring that characters interact correctly with their surroundings (e.g., to study at a desk, a character needs to sit on the associated chair). +- Regional context reasoning $(RC)$ : Regional context reasoning is required for generating character behaviors that + +![](images/841df82702aa02d5c36d30dd6a4ff7be29d9782ac412cbf2fa81eb0bd52400f5.jpg) +Figure 6. An example of a test case. In this test case, a system is required to identify the chairs in the kitchen and select one that is not currently occupied. + +are well-aligned with the contextual meaning of a space (e.g., cooking in the kitchen rather than in the living room). + +- Scene state reasoning (SS): Scene state reasoning requires considering the overall scene flow and the current state of other characters (e.g., if a character is using the coffee machine, another character should wait to use it). + +In addition to the holistic spatial context defined above, we separately evaluate the low-level performance on Position inference $(PI)$ . The cases with the tag analyze if the system can find the precise positions of characters at the coordinate level. + +# 4.2. Benchmark Creation + +Test Scenes We create three scenes—House, Office, and Restaurant—by placing the objects from the HSSD dataset [21]. Each scene is designed to include two to three separate areas with distinct contextual meaning. To ensure the LLM infers contextual meaning of areas without relying on object labels, we replace several object labels with general ones (e.g., kitchen_cabinet $\rightarrow$ cabinet). We also randomize indices for identically labeled objects to eliminate index-based arrangement cues. We define a distinct set of available action labels $A_{S}$ for each scene, where $S \in \{\text{house}, \text{office}, \text{restaurant}\}$ . This $A_{S}$ is utilized when determining the pass/fail of test cases that use the scene $S$ . + +Test Cases Our framework can create natural behaviors for multiple characters without any pre-scripted scenarios. However, for rigorous evaluation, we introduce specific test scenarios in our benchmark. Each test case consists of a scene $S$ , characters $\mathcal{C} = \{c_1, c_2, \dots, c_N\}$ , test scenario $\mathcal{T}$ , mock planning history $\mathcal{H}$ , and expected plan $\tau_e = (\mathcal{A}_e, \tilde{p}_e)$ . Here, $N$ is the number of characters, $\mathcal{A}_e$ is the set of expected action labels, and $\tilde{p}_e$ is the expected area where the target character's position is expected to be included. $\tilde{p}_e$ is expressed through the union or intersection of multiple areas, each defined as a semantic description based on scene objects. For simplicity, we ignore the character's orientation in the benchmark, and the target character for the expected + +plan $\tau_{e}$ is always fixed to the first character $c_{1}$ in all test cases. An example of a test case is depicted in Figure 6. + +When running a test case, we input the test scenario $\mathcal{T}$ and mock planning history $\mathcal{H}$ into the system. The mock planning history $\mathcal{H}$ simulates the prior planning history of the scene, building the necessary context required for the test case. Then, the system generates the most appropriate next plan $\tau_{g} = (a_{g},\tilde{p}_{g}$ or $p_g)$ by evaluating the given mock planning history and the test scenario. Here, $\tau_{g}$ is the generated next plan for the target character, $a_{g}$ is the planned action label, and $\tilde{p}_g$ or $p_g$ is the planned position. If the test case evaluates coordinate-level localization, the system outputs a coordinate-level position $p_{g}$ . Otherwise, it outputs an area-level position $\tilde{p}_g$ . As the final step, the test case's pass/fail is determined by comparing the expected plan $\tau_{e}$ and the generated plan $\tau_{g}$ . For a test case to pass, the following conditions must all be satisfied: + +1. A plan for the target character is generated $(\tau_{g}\neq \emptyset)$ +2. The action label is included in both the available action labels and the expected action labels $(a_{g}\in \mathcal{A}_{\mathcal{S}}\cap \mathcal{A}_{e})$ +3. The character's position is within the expected area $(\tilde{p}_g \subseteq \tilde{p}_e$ or $p_g \in \tilde{p}_e)$ . + +Benchmark Summary We construct the benchmark based on the previously defined evaluation criteria, namely, tags. Our benchmark consists of 40 test cases in total. 10 test cases are designed for the $PI$ tag, while the remaining 30 test cases involve $OA$ , $RC$ , and $SS$ tags. The test cases for the $PI$ tag deliberately use straightforward scenarios to focus solely on the localization performance. Among the latter 30 test cases, half incorporate two tags for increased complexity. Additionally, for the latter 30 test cases, we vary the number of characters in the scene and the size of the mock planning history across three levels within the same scenario. This expands the number of effective test cases and allows us to evaluate the system's scalability by assessing its robustness against increasing agent counts and planning history complexity. We design our benchmark to ensure that each tag is evenly represented, preventing bias toward any specific evaluation factor. + +Evaluation Metrics Our benchmark employs two evaluation metrics: success rate and execution rate. The success rate measures the proportion of passed runs among all test case runs. On the other hand, the execution rate measures the proportion of runs that are executed without any runtime errors, which are included in the parenthesis next to the success rate in tables. Runtime errors typically occur when the LLM module references a nonexistent object or generates a response with incorrect syntax. Any test case run with a runtime error is automatically marked as failed. + +
ModelGPT-4oGPT-4o miniLlama-3.1-70B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.9 (0.98)0.93 (0.99)0.9 (0.98)0.92 (0.98)0.74 (0.96)0.72 (0.95)0.72 (0.97)0.78 (0.93)0.72 (0.87)0.78 (0.91)0.76 (0.87)0.61 (0.81)
w/o Event0.82 (0.92)0.88 (0.92)0.86 (0.93)0.77 (0.92)0.6 (0.95)0.56 (0.99)0.6 (0.92)0.54 (0.92)0.6 (0.81)0.6 (0.78)0.62 (0.86)0.54 (0.76)
Object List0.51 (0.85)0.61 (0.78)0.28 (0.88)0.65 (0.97)0.34 (0.88)0.37 (0.88)0.12 (0.86)0.51 (0.9)0.35 (0.68)0.31 (0.7)0.29 (0.61)0.41 (0.69)
Scene Graph0.82 (0.96)0.8 (0.96)0.82 (0.94)0.87 (0.95)0.49 (0.9)0.52 (0.93)0.32 (0.9)0.51 (0.83)0.72 (0.92)0.69 (0.91)0.77 (0.94)0.64 (0.85)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.35 (0.74)0.37 (0.8)0.37 (0.77)0.35 (0.67)0.71 (0.9)0.71 (0.91)0.72 (0.95)0.65 (0.85)0.39 (0.87)0.38 (0.88)0.36 (0.87)0.35 (0.85)
w/o Event0.31 (0.58)0.3 (0.57)0.32 (0.6)0.28 (0.63)0.62 (0.85)0.74 (0.85)0.65 (0.9)0.59 (0.85)0.35 (0.79)0.41 (0.86)0.32 (0.79)0.32 (0.74)
Object List0.22 (0.58)0.26 (0.65)0.13 (0.62)0.23 (0.44)0.44 (0.77)0.45 (0.77)0.36 (0.83)0.52 (0.72)0.14 (0.77)0.09 (0.74)0.07 (0.77)0.24 (0.67)
Scene Graph0.29 (0.75)0.31 (0.76)0.35 (0.86)0.16 (0.61)0.69 (0.84)0.69 (0.81)0.77 (0.93)0.55 (0.72)0.31 (0.84)0.27 (0.88)0.23 (0.84)0.35 (0.84)
+ +Table 1. Benchmark result for test cases with object arrangement reasoning (OA), regional context reasoning (RC), and scene state reasoning (SS) tags. In total, our method achieves the highest success rate across various LLM models and sizes. + +![](images/45e1ee3cf38af64f0a0e53a406289310acbb6f71238197d5d152b45869456e3f.jpg) +Figure 7. Benchmark results by number of characters. Our approach exhibits stable success rate for stories with different number of characters while the performance gap increases as more characters are involved, compared to the version without event representation. + +![](images/5c907f3c4ddf01e44942aaf55a63c1f808eb977e6c75de650e5684f1c11684f9.jpg) +Figure 8. Comparison of average token usage. Object List uses the least number of tokens, and Scene Graph uses significantly more number of tokens, especially when the scene complexity increases (Office). + +# 4.3. Benchmark Results + +We use GPT-4o [16], Llama-3.1 [8], and Qwen2.5 [43] in various sizes as LLM backbones to obtain benchmark results. The backbones are selected to represent both commercial and open-source modules that are popular and demonstrate competitive performance. We repeat each test case five times and average the results. + +Event-Driven Planning Table 1 presents the benchmark results for test cases with $OA$ , $RC$ , and $SS$ tags. Each column represents the average score for test cases that include the corresponding tag and each cell displays the success rate, with the execution rate shown in parentheses. The ablation without event-driven planning (w/o Event) handles the entire process with a single unified LLM module without using high-level event representations. Therefore, all in + +put and output information, except for the scene description and test scenario, is represented in a low-level structured format. Compared to $w/o$ Event, our method benefits from a cascaded planning pipeline that effectively decomposes the entire planning complexity. This leads to an overall improvement in planning performance. In particular, the performance gain in SS is relatively larger compared to other tags in GPT-40 models. The SS tag requires a comprehensive understanding of the overall planning state of multiple characters in the scene. We attribute this to the advantages of our event-based representation, which provides a more efficient and intuitive way to capture the overall scene state. Additionally, as shown in Figure 7, our planning pipeline proves to be more effective in terms of scalability. $w/o$ Event struggles with planning as the number of characters in the scene increases, whereas our approach maintains stable performance. + +Scene Description We also analyze the effect of using scene description generated from an LLM module. We create ablation settings that retain our pipeline but modify the scene description method. Object List represents the scene as a simple list of objects present in the scene, while Scene Graph uses the JSON-formatted scene graph representation as the scene description. Our approach contributes to more effective planning compared to other description methods as shown in Table 1. Additionally, we present the average token usage for each description method in Figure 8 to provide a deeper understanding of their differences. As presented in Figure 8, the object list is the most efficient way to represent scene information and has often been used in previous studies such as [5, 20, 34]. However, since the object list does not have any spatial information, the planning module with the description is ignorant of object arrangement or spatial layout. As a result, the object list records the lowest scores in our benchmark, which requires sophisticated scene understanding for planning. + +Unlike the object list, the scene graph representation enables necessary spatial reasoning by incorporating detailed attributes of objects and spatial relationships between ob + +
ModelGPT-4oGPT-4o miniLlama-3.1-70B
Ours0.94 (1.0)0.9 (1.0)0.94 (1.0)
Direct Inference0.46 (1.0)0.38 (1.0)0.28 (1.0)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
Ours0.86 (0.88)1.0 (1.0)0.94 (1.0)
Direct Inference0.14 (0.9)0.14 (0.48)0.1 (0.8)
+ +Table 2. Benchmark result for test cases with position inference tag (PI). Our proposed method can stably infer accurate positions, whereas Direct Inference with LLMs often struggles. + +jects. However, the representation contains exhaustive detail, including even minor and insignificant ones, without any abstraction or compression. This inefficiency leads to excessive token consumption (Figure 8) and can even interfere with the LLM's reasoning due to unnecessary details (Table 1). The token inefficiency becomes more pronounced as the scene size and complexity increase. As shown in Figure 8, when comparing the relatively simple 'House' scene to the more complex 'Office' scene, the scene graph exhibits a significantly larger increase in token consumption compared to alternative approaches. Compared to the scene graph, our approach allows for a more efficient description of the given scene while preserving key spatial information. + +Positional Inference (PI) In Table 2, we demonstrate the effectiveness of our area-conditioned position sampling method based on the results for the PI-tagged test cases. Direct Inference is an ablation setting where the same planning pipeline is used, but the LLM directly outputs coordinates when determining character positions. Although all test cases with the PI tag are intentionally designed to be very simple, all LLM models with Direct Inference struggle to accurately determine the coordinates of character positions. In contrast, our system allows the LLM to process information semantically, and obtains precise locations through the proposed area-conditioned position sampling technique. + +# 4.4. User Study Results + +We also validate through a user study that our approach represents the given scenario more effectively than other ablation settings. For the user study, we use a total of four test scenarios, each set in a different scene. The scenes include House, Office, and Restaurant from the benchmark, as well as MPH11 from the PROX dataset [12]. Compared to the other scenes, MPH11 is relatively simple in both scale and complexity. Given its limited space, we use a simple scenario containing only two events for the MPH11. In contrast, for other test scenes, we use more complex scenarios that include 4-5 events. For each test scenario, users are asked to visually compare results generated from different ablation settings and select the one that best represents the given scenario. If multiple results are perceived as equally effective, users are allowed to choose more than one. We + +![](images/95d295ffc686b4380dde9ff5b00d25bd99e59bf7f9feba4c3539b75741b2e782.jpg) + +![](images/040e2bb95e94b12f51e808d1a098e2522ee7da6fb983f8f442327e5299fcec06.jpg) +Figure 9. User study results. Our system correctly generates the dynamic story line with multiple characters in large scenes, significantly outperforming other baselines, except for simple scenarios with MPH11. + +generate the results for the user study using GPT-4o mini with a temperature setting of 0.0. + +We collected responses from 50 general participants. The responses to the user study are summarized in Figure 9. The chart labels SF, OL, and No E correspond to the same ablation settings used in the benchmark: Scene Graph, Object List, and w/o Event, respectively. In three of our scenarios, our approach received the most selections by a large margin. Furthermore, the noticeable difference in selection distribution between the MPH11 scene, which involves a relatively smaller-scale problem, and the other scenes suggests that our approach contributes to building a more scalable system. We provide the test scenarios and visualization of generated results in the supplementary material. + +# 5. Conclusion + +In this study, we propose an LLM-based framework for generating multiple lifelike virtual humans within a given 3D scene, dynamically aligning their behaviors with an emergent storyline. Characters generated by our framework interact contextually with their surroundings and other individuals, enhancing scene realism. Our framework performs multi-character behavior planning based on semi-narrative events, allowing the LLM to efficiently capture scene progression and group behaviors. The modular design of our planning pipeline decomposes the overall problem, reducing reasoning complexity for the LLM and enabling more scalable planning. Additionally, we introduce a context-centric scene description to efficiently convey key spatial information and propose area-conditioned position sampling to mitigate the LLM's weaknesses in numerical reasoning. To evaluate our system, we construct a benchmark for a comprehensive assessment of LLM-based scene-aware multi-agent behavior planning. Through benchmark results and a user study, we demonstrate that our proposed methods contribute to building a scalable and effective system for coordinating multi-agent behaviors, while carefully considering the scene. + +Acknowledgments This work was supported by the IITP grant [RS-2021-II211341, Artificial Intelligence Graduate School Program (Chung-Ang University)], the NRF grant (No. RS-2023-00208197) funded by the Korea government (MSIT), and Creative-Pioneering Researchers Program through Seoul National University. + +# References + +[1] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 1, 2 +[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3 +[3] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 1, 2 +[4] Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Generating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1855-1866, 2024. 1, 2 +[5] Jianqi Chen, Panwen Hu, Xiaojun Chang, Zhenwei Shi, Michael Christian Kampffmeyer, and Xiaodan Liang. Sitcom-crafter: A plot-driven human motion generation system in 3d scenes. arXiv preprint arXiv:2410.10790, 2024. 2, 7 +[6] Simon Clavet et al. Motion matching and the road to next-gen animation. In Proc. of GDC, page 4, 2016. 5 +[7] Arash Gholami Davoodi, Seyed Pouyan Mousavi Davoudi, and Pouya Pezeshkpour. Llms are not intelligent thinkers: Introducing mathematical topic tree benchmark for comprehensive evaluation of llms. arXiv preprint arXiv:2406.05194, 2024. 4 +[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 7 +[9] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 3 +[10] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision, pages 418-437. Springer, 2024. 1, 2 + +[11] Maitrey Gramopadhye and Daniel Szafir. Generating executable action plans with environmentally-aware language models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3568-3575. IEEE, 2023. 1, 2 +[12] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 8 +[13] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 1, 2 +[14] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In Proceedings of the 41st International Conference on Machine Learning, pages 20413-20451, 2024. 3 +[15] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2 +[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 7 +[17] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In European Conference on Computer Vision, pages 289-310. Springer, 2024. 3 +[18] Bowen Jiang, Yangxinyu Xie, Zhuoqun Hao, Xiaomeng Wang, Tanwi Mallick, Weijie Su, Camillo Taylor, and Dan Roth. A peek into token bias: Large language models are not yet genuine reasoners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4722-4756, 2024. 2 +[19] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 1, 2 +[20] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12140–12147. IEEE, 2024. 2, 7 +[21] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 6 + +[22] Boyi Li, Philipp Wu, Pieter Abbeel, and Jitendra Malik. Interactive task planning with language models. arXiv preprint arXiv:2310.10645, 2023. 2 +[23] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, 132(9):3463-3483, 2024. 1, 2 +[24] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 1, 2, 4 +[25] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 1, 2 +[26] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 903-913. IEEE, 2024. 1, 2 +[27] Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. 2, 4 +[28] Zhe Ni, Xiaoxin Deng, Cong Tai, Xinyue Zhu, Qinghongbing Xie, Weihang Huang, Xiang Wu, and Long Zeng. Grid: Scene-graph-based instruction-driven robotic task planning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13765-13772. IEEE, 2024. 2 +[29] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1-22, 2023. 1 +[30] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Sunderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, 2023. 1, 2 +[31] Mengyi Shan, Lu Dong, Yutao Han, Yuan Yao, Tao Liu, Ifeoma Nwogu, Guo-Jun Qi, and Mitch Hill. Towards open domain text-driven synthesis of multi-person motions. In European Conference on Computer Vision, pages 67-86. Springer, 2024. 1, 2 +[32] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Scharli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR, 2023. 2 +[33] David Silver. Cooperative pathfinding. In Proceedings of the aaai conference on artificial intelligence and interactive digital entertainment, pages 117-122, 2005. 5 +[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot + +task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 1, 2, 4, 7 +[35] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 1, 2 +[36] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024. 1, 2 +[37] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 1, 2 +[38] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022. 5 +[39] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems, 35:14959-14971, 2022. +[40] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 1, 2 +[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4 +[42] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2 +[43] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 7 +[44] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 2, 4 +[45] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision, pages 246-263. Springer, 2024. 1, 2 +[46] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang + +Gan. Building cooperative embodied agents modularly with large language models. In The Twelfth International Conference on Learning Representations, 2024. 2 +[47] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14738-14749, 2023. 1, 2 \ No newline at end of file diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/images.zip b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d9143f92ea60e46574dbd3c82a825ebe5d2bce62 --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1049272cc7314e3b0eda233bd9f41116da8f0d965feeed77d13794dcd9958208 +size 432857 diff --git a/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/layout.json b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..892e73943d5347c6e2ba4d4c4006c2cf8f3f143c --- /dev/null +++ b/2025/Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene/layout.json @@ -0,0 +1,7992 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "spans": [ + { + "bbox": [ + 85, + 103, + 525, + 121 + ], + "type": "text", + "content": "Event-Driven Storytelling with Multiple Lifelike Humans in a 3D Scene" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "spans": [ + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": "Donggeun Lim" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": " Jinseok Bae" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": " Inwoo Hwang" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": " Seungmin Lee" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": " Hwanhee Lee" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "text", + "content": " Young Min Kim" + }, + { + "bbox": [ + 138, + 142, + 471, + 171 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "spans": [ + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "type": "text", + "content": "Seoul National University " + }, + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 196, + 175, + 414, + 188 + ], + "type": "text", + "content": "Chung-Ang University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "spans": [ + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "inline_equation", + "content": "^{1}\\{\\mathrm{rms}2836" + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "text", + "content": " , capoo95, inusu0818, rsual, youngmin.kim " + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "inline_equation", + "content": "\\} \\in" + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "text", + "content": " snu.ac.kr " + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 75, + 190, + 531, + 201 + ], + "type": "text", + "content": " hwanheelee@cau.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 59, + 211, + 220, + 355 + ], + "blocks": [ + { + "bbox": [ + 59, + 211, + 220, + 355 + ], + "lines": [ + { + "bbox": [ + 59, + 211, + 220, + 355 + ], + "spans": [ + { + "bbox": [ + 59, + 211, + 220, + 355 + ], + "type": "image", + "image_path": "ba5a4f3b68252eb66586ef36117577f3dce395ad52a54384a8229c3f9223f601.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 555, + 380 + ], + "type": "text", + "content": "Figure 1. Our framework populates 3D scenes with multiple characters. The generated characters interact with their surroundings and other people, bringing the space to life." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 223, + 211, + 381, + 354 + ], + "blocks": [ + { + "bbox": [ + 223, + 211, + 381, + 354 + ], + "lines": [ + { + "bbox": [ + 223, + 211, + 381, + 354 + ], + "spans": [ + { + "bbox": [ + 223, + 211, + 381, + 354 + ], + "type": "image", + "image_path": "0a9c2520a27f79d3803801e501ac6ffa2909308936575c3cc8353ed33494e951.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 383, + 212, + 551, + 355 + ], + "blocks": [ + { + "bbox": [ + 383, + 212, + 551, + 355 + ], + "lines": [ + { + "bbox": [ + 383, + 212, + 551, + 355 + ], + "spans": [ + { + "bbox": [ + 383, + 212, + 551, + 355 + ], + "type": "image", + "image_path": "37d24f6da7963b2eab81ec7b6bbdb7b2bdc684b39e18dc750edd078c677799ec.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 391, + 200, + 403 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 391, + 200, + 403 + ], + "spans": [ + { + "bbox": [ + 151, + 391, + 200, + 403 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 418, + 297, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 418, + 297, + 693 + ], + "spans": [ + { + "bbox": [ + 54, + 418, + 297, + 693 + ], + "type": "text", + "content": "In this work, we propose a framework that creates a lively virtual dynamic scene with contextual motions of multiple humans. Generating multi-human contextual motion requires holistic reasoning over dynamic relationships among human-human and human-scene interactions. We adapt the power of a large language model (LLM) to digest the contextual complexity within textual input and convert the task into tangible subproblems such that we can generate multi-agent behavior beyond the scale that was not considered before. Specifically, our event generator formulates the temporal progression of a dynamic scene into a sequence of small events. Each event calls for a well-defined motion involving relevant characters and objects. Next, we synthesize the motions of characters at positions sampled based on spatial guidance. We employ a high-level module to deliver scalable yet comprehensive context, translating events into relative descriptions that enable the retrieval of precise coordinates. As the first to address this problem at scale and with diversity, we offer a benchmark to assess diverse aspects of contextual reasoning. Benchmark results and user studies show that our framework effectively captures scene context with high scalability. The code and benchmark, along with result videos, are available" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 392, + 556, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 556, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 556, + 416 + ], + "type": "text", + "content": "at our project page: https://rms0329.github.io/Event-Driven-Storytelling/." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "spans": [ + { + "bbox": [ + 314, + 441, + 394, + 453 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 461, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 461, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 312, + 461, + 555, + 640 + ], + "type": "text", + "content": "A scene with digital humans creates lively atmosphere and has a wide range of applications in VR, games, and movies. The realism incurs from natural motions of multiple characters that seamlessly integrate into the surrounding environment. However, existing motion synthesis works focus on individual characters independently or extend to interactions with either the surrounding environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35], remaining limited to a single type of interaction. Hence, they severely lack generalizability and scalability to concurrently generate plausible interactions of multiple characters in a crowded scene. It is highly challenging to correctly assign the behavior of individual characters at the correct time stamp as it requires extensive dynamic contextual and spatial reasoning." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "To fulfill such holistic requirements, we take inspiration from recent advances in large language models (LLMs). LLMs demonstrated phenomenal performance in various areas, especially for high-level planning [11, 15, 24, 30, 34, 36] or behavior modeling [3, 29]. We deduce that LLMs have the potential to interpret nuanced semantics from" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 143, + 712 + ], + "type": "text", + "content": "†Corresponding author" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11654" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "open-ended textual inputs and dependency between natural motion and scene context. However, simply generating commands from the input request may suffer from hallucinations or struggle with localization errors, which are well-known limitations of foundation models [18, 27, 32, 42]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 289 + ], + "type": "text", + "content": "We propose an LLM-powered framework that populates multiple lifelike virtual humans within a 3D scene. As shown in Figure 1, the characters produced by our framework can understand and interact with their surroundings, including both 3D scenes and other characters. The proposed LLM module can flexibly understand the holistic context in open-ended scenarios and make high-level decisions. The characters may exhibit emergent yet plausible behavior that was not explicitly defined within the original input, providing a lively atmosphere with rich interactions. The framework can further adapt to user-interactive scenarios, where users may interact with the characters through high-level text instructions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 291, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 422 + ], + "type": "text", + "content": "Towards more performant and scalable planning with LLMs, our modular framework is deliberately designed to reduce the reasoning complexity. We introduce an event-based planning approach with two LLM modules: narrator and event parser. The narrator progressively weaves the flow of the scene by generating a textual description of one event at a time. Each event considers only a well-defined subset of characters and objects, detached from the collective behavior of the holistic flow. The event parser then transforms high-level event representation into detailed information for existing motion synthesis frameworks." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 425, + 295, + 616 + ], + "type": "text", + "content": "We provide devised scene information tailored for individual modules to guide necessary spatial reasoning. Our scene descriptor converts the 3D scene graph into a textual summary and provides it as input to the narrator and the event parser. Our prompt encourages contextual information on different utilization of spaces, such that the description delivers the overall structure. The event parser provides accurate 3D scene grounding to the low-level motion synthesis module, such that it can generate fine-grained and realistic interaction. Inspired by previous studies [24, 34, 36, 44], the event parser employs spatial reasoning tools for LLM and specifies a relevant location within a programming framework. Furthermore, we propose an area-conditioned position sampling technique to compensate for the weakness of LLMs in precise localization at the coordinate level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "Our framework successfully generates long-term motions with more than 4-5 characters in various multi-room scale scenes, and is robust to the choice of LLM engines. We also propose evaluation criteria for scene-aware multi-agent behavior planning tasks and create a benchmark to assess the compatibility of our framework comprehensively. The extensive results demonstrate that our proposed pipeline outperforms in generating a natural and plausible" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 109 + ], + "type": "text", + "content": "sequence of actions with adequate trajectories at appropriate relative timing in a shared space, especially for larger scenes with more number of characters." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 121, + 406, + 134 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 406, + 134 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 406, + 134 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 143, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 143, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 143, + 555, + 468 + ], + "type": "text", + "content": "Contextual Human Motion Synthesis Synthesizing human motions is a long-standing research topic in the fields of computer vision and graphics. In particular, an increasing number of studies have incorporated contextual information in motion synthesis, such as a 3D environment [1, 4, 13, 19, 26, 37-40, 45, 47] or another human [10, 23, 25, 31, 35]. Although these studies have shown notable progress in reproducing physically accurate and high-fidelity motions for a single type of interaction, most of them do not consider more than one contextual constraint or expand to long-term sequences of actions. Some recent works [3, 5] have explored motion generation considering both humans and scenes of contextual information. Digital Life Project [3] builds autonomous characters with social intelligence in a 3D scene, while Sitcom-Crafter [5] introduces a unified motion generation framework that integrates different types of interaction motions with plot-driven guidance. The research expanded on previous boundaries, yet it did not take into account the coordination of interactions between more than two characters. Furthermore, in terms of scene understanding, these models lack the complexity needed for scenarios requiring detailed spatial reasoning because of their simplistic scene descriptions. Compared to these studies, our framework demonstrates successful multi-agent contextual behavior planning at a larger scale, including an increased number of characters and a scene with greater size and complexity." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 715 + ], + "type": "text", + "content": "LLM-based Planning In recent years, there has been growing attention to using pre-trained LLMs in various domains for their impressive zero-shot reasoning capabilities [8, 16, 43]. Particularly in the field of robotics and embodied AI, many recent works have successfully demonstrated the abilities of LLMs for task planning [11, 15, 20, 22, 24, 28, 30, 34, 36]. However, robotic applications primarily focus on goal-oriented tasks, which do not require complex contextual reasoning. While the majority of these works are limited to single-agent scenarios, a few studies [20, 46] have explored LLM-based multiagent task planning recently. CoELA [46] introduces a novel cognitive-inspired modular framework that enables cooperation and communication between multiple agents. However, since CoELA requires continuous message exchanges between agents for behavior coordination, this can result in inefficiency when scaling to an increased number of agents. On the other hand, SMART-LLM [20] demonstrated multi-agent task planning in larger-scale scenarios." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11655" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 307, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 307, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 307, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 307, + 178 + ], + "type": "image", + "image_path": "44d0ecc5690a7c89cf5be99672997194baeae64dc3d4fa22c53a2786ecbd6889.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 187, + 296, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 187, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 187, + 296, + 254 + ], + "type": "text", + "content": "Figure 2. Overview of our framework. The framework consists of two main modules: a high-level action planning module and a low-level motion synthesis module. The action planning module coordinates and plans multi-character behavior as a sequence of events. The motion synthesis module receives a detailed description of the event and synthesizes 3D motions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 261, + 296, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 261, + 296, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 261, + 296, + 344 + ], + "type": "text", + "content": "SMART-LLM accomplishes multi-robot cooperation by decomposing tasks into several sub-stages. However, SMART-LLM represents a scene as a mere list of objects, leading to a lack of comprehensive scene understanding. Distinct from earlier approaches, ours tackles the problem of multi-agent planning while maintaining scalability and in-depth scene understanding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 357, + 111, + 369 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 111, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 111, + 369 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": "Given a 3D scene " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": ", characters " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\}" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": ", and optional user instructions " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": ", our framework generates 3D multi-human motions " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{M} = \\{M_1, M_2, \\dots, M_N\\}" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": " within the scene. Here, " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": " represents the number of characters and is determined by the user at the system initialization. The 3D scene " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": " is provided with instance segmentation, and can be in various representations including VR scenes with CAD models or real-world scenes composed of 3D point cloud scans or mesh reconstructions. The user instructions " + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 377, + 296, + 509 + ], + "type": "text", + "content": " are provided in free text forms, leveraging the generalizability of LLM engines." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 653 + ], + "type": "text", + "content": "In an overview, our framework operates around an intermediate representation called event " + }, + { + "bbox": [ + 55, + 510, + 296, + 653 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 510, + 296, + 653 + ], + "type": "text", + "content": ". Figure 2 contains examples of textual descriptions of events along with detailed action labels and positions for the motion synthesis module. The events are generated in sequence to realize the given instruction while coordinating with the surrounding environments. Each event has its own lifecycle with a different time span, and multiple events can coexist within the 3D scene. Through the event-based formulation, our system can generate scene dynamics with infinitely long-horizon, and adaptively process user instructions at an arbitrary time during operation." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 713 + ], + "type": "text", + "content": "As shown in Figure 2, our framework consists of two main components: a high-level action planning module and a low-level motion synthesis module. The action planning module performs planning on an event basis, driving the scene's timeline forward (Section 3.1). The motion syn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 108 + ], + "type": "text", + "content": "thesis module takes the generated events as input, assigns them to the corresponding characters, and creates motions that reflect the assigned event (Section 3.2)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 116, + 501, + 129 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 116, + 501, + 129 + ], + "spans": [ + { + "bbox": [ + 313, + 116, + 501, + 129 + ], + "type": "text", + "content": "3.1. High-level Action Planning Module" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "content": "The primary function of the action planning module is to generate new events by comprehensively considering the given 3D scene " + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "content": ", existing event history " + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "content": ", and the behaviors of multiple characters within the scene. To effectively satisfy the holistic requirements, we rely on the powerful reasoning capabilities of LLMs. As shown in Figure 3, there are three LLM submodules: the scene descriptor, the narrator, and the event parser. The scene descriptor generates a textual scene description " + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "content": " from the given 3D scene " + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 133, + 555, + 324 + ], + "type": "text", + "content": " such that our planning modules can understand the necessary context from the spatial arrangement. The narrator generates a sequence of events, which are converted into detailed information by the event parser. Each module is assigned to smaller, well-defined tasks such that the system stays performant and scalable for multi-human motion generation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 331, + 410, + 341 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 410, + 341 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 410, + 341 + ], + "type": "text", + "content": "3.1.1. Scene Descriptor" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "text", + "content": "In the preprocessing stage before the motion planning, the scene descriptor generates a textual description " + }, + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "text", + "content": " of the given 3D scene " + }, + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 346, + 556, + 609 + ], + "type": "text", + "content": ". We use a 3D scene graph, which may lead an LLM module to absorb rich spatial relationships [14]. The scene descriptor first extracts a 3D scene graph in an automated way as proposed in [17] and converts it into textual data in JSON format. Then, we use it to prompt the LLM module and generate a scene description. In addition to the explicit pairwise relationships in the scene graph, further contextual layouts with regional information enhance the naturalness of scene-aware motions. To enhance contextual reasoning, we encourage our scene descriptor to discover a spatial arrangement of objects in proximity, which semantically creates a functional space, for example, a dining area or study zone. Specifically, we extract a regional cluster of objects through the DBSCAN algorithm [9] as an additional input and promote detecting areas of interest. We also provide concrete examples of the same input and output format, allowing the scene descriptor to perform the given task more effectively through in-context learning [2]. We provide additional details and full prompts in the supplementary material." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 615, + 382, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 382, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 382, + 625 + ], + "type": "text", + "content": "3.1.2. Narrator" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": "As the core of our behavior planning, the narrator generates a sequence of events " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_1,e_2,\\ldots \\}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ". Given the scene description " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " from the scene descriptor and the optional user instruction " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": ", the narrator keeps the previous event history " + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 630, + 554, + 713 + ], + "type": "text", + "content": " and generates one event at a time based on the high-level context. Focusing on one event at a time, the subsequent module can alleviate the burden of holistic spatio-temporal" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11656" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 68, + 545, + 170 + ], + "blocks": [ + { + "bbox": [ + 67, + 68, + 545, + 170 + ], + "lines": [ + { + "bbox": [ + 67, + 68, + 545, + 170 + ], + "spans": [ + { + "bbox": [ + 67, + 68, + 545, + 170 + ], + "type": "image", + "image_path": "9a71ebeff84852c8a41fd1a76d6edf2c2a11c3255ae0f8c394a9643bfeeadd5f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 176, + 416, + 186 + ], + "lines": [ + { + "bbox": [ + 192, + 176, + 416, + 186 + ], + "spans": [ + { + "bbox": [ + 192, + 176, + 416, + 186 + ], + "type": "text", + "content": "Figure 3. Overview of the high-level action planning module." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 198, + 294, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 198, + 294, + 245 + ], + "spans": [ + { + "bbox": [ + 54, + 198, + 294, + 245 + ], + "type": "text", + "content": "reasoning, making the storytelling scalable and interactive. The event-based design also allows the incorporation of additional user instructions between events to refine the storyline." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 247, + 296, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 247, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 247, + 296, + 426 + ], + "type": "text", + "content": "As shown in Figure 2, events are expressed in seminarrative natural language, focusing on contextual description. The textual description is further parsed in the subsequent event parser, relieving the burden of detailed spatiotemporal localization. The narrator also receives the state of a generated event as one of two states, 'ongoing' or 'completed'. From this, the narrator can compactly conceive the necessary information to progress the timeline. Our event representation is particularly efficient for group events with multiple characters, which makes the broad scene context easily understood. The narrator coordinates the high-level behaviors of multiple characters and generates the storyline from a macroscopic perspective. The detailed spatiotemporal progression of individual characters is assigned to the event parser." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 296, + 559 + ], + "type": "text", + "content": "Similar to the scene descriptor, we also incorporate several widely used prompting techniques in the narrator's prompt, the in-context learning and chain-of-thought reasoning [41]. For chain-of-thought reasoning, we guide the narrator to first reason about the current planning state and other characters' statuses before generating an event. We provide concrete examples of event generation with such a reasoning process to reinforce structured decision-making and improve the coherence of generated events. Further details and the full prompts are provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 566, + 139, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 139, + 577 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 139, + 577 + ], + "type": "text", + "content": "3.1.3.Event Parser" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": "The event parser takes the event " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " generated by the narrator as input and concretizes it into low-level details that the motion synthesis module can digest. The output event of the narrator is designed to be a short description of the region and the character for the motion, leaving ambiguities on the exact location within the scene. The event parser observes the current state of the scene and refines the event description into detailed 3D grounding suitable for motion synthesis. Specifically, it converts the event description " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "e = (\\mathcal{C}_e, \\{p_i\\}, \\{d_i\\}, \\{a_i\\})" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{C}_e = \\{c_i\\}" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is the set of involved characters, " + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "inline_equation", + "content": "p_i \\in \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 582, + 296, + 714 + ], + "type": "text", + "content": " is the target position," + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 313, + 200, + 557, + 366 + ], + "blocks": [ + { + "bbox": [ + 313, + 200, + 557, + 366 + ], + "lines": [ + { + "bbox": [ + 313, + 200, + 557, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 557, + 366 + ], + "type": "image", + "image_path": "2919c4aaa9cd9d79817cb05a7c6eeadf16157fa02b7a6e01f9a6db6a740a196d.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 371, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 371, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 371, + 555, + 427 + ], + "type": "text", + "content": "Figure 4. An example of a programming-structured prompt used in our event parser. The event parser leverages provided spatial reasoning tools, such as the get_distance_between() function in this example, to procedurally deduce solutions for given spatial reasoning tasks." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "inline_equation", + "content": "d_{i}\\in \\mathbb{R}^{1}" + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "text", + "content": " is the target orientation, and " + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "text", + "content": " is the target action label for each character " + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 313, + 437, + 554, + 462 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": "Inspired by [24, 34, 44], we adopt a Python programming structure in the event parser's prompt and utilize spatial reasoning tools that are provided as a form of Python functions. The programming framework can infer solutions in a well-defined sequence of inquiries, leading to more powerful spatial reasoning capabilities. Figure 4 shows a specific example. The event parser uses the get_distance_between() function to deduce spatial details missing from the scene description, such as identifying the farthest seat from a given reference point. The event parser compiles the output format by calling the get_character() function, and for each character " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "c_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", it specifies " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 313, + 462, + 555, + 715 + ], + "type": "text", + "content": " using the set_position(), set_ORIENTATION(), and set_target_action() functions, respectively. If a specific target orientation is not necessary, it can be left unspecified. When retrieving the positions, the event parser first outputs semantic descriptions using scene objects, and precise coordinates are found afterward to compensate for possible errors of LLMs in coordinate-level reasoning [7, 27]. Specifically, we define a position by its spatial relationship with an anchor object" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11657" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 70, + 529, + 177 + ], + "blocks": [ + { + "bbox": [ + 83, + 70, + 529, + 177 + ], + "lines": [ + { + "bbox": [ + 83, + 70, + 529, + 177 + ], + "spans": [ + { + "bbox": [ + 83, + 70, + 529, + 177 + ], + "type": "image", + "image_path": "9713893b4dc8d03263da5179c1dfdc0ed2aad564f83ae10feb76a45832b8e1ed.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 185, + 436, + 196 + ], + "lines": [ + { + "bbox": [ + 173, + 185, + 436, + 196 + ], + "spans": [ + { + "bbox": [ + 173, + 185, + 436, + 196 + ], + "type": "text", + "content": "Figure 5. Qualitative results of our framework across diverse 3D scenes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "spans": [ + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": "and sample the coordinate representation " + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": " within the corresponding area. Then, if a target orientation is specified, we calculate " + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": " as the object or character the subject faces after sampling " + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": ". Otherwise, " + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "inline_equation", + "content": "d_i" + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": " is set to the character's expected forward direction upon reaching " + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 54, + 208, + 294, + 267 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 268, + 295, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 268, + 295, + 327 + ], + "spans": [ + { + "bbox": [ + 54, + 268, + 295, + 327 + ], + "type": "text", + "content": "As with the narrator, the event parser incorporates in-context learning, chain-of-thought reasoning, and self-feedback techniques in its prompt. Further details about sampling and the full prompts are provided in the supplementary material." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 336, + 246, + 348 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 246, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 246, + 348 + ], + "type": "text", + "content": "3.2. Low-level Motion Synthesis Module" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": "After the high-level action planning module provides the narrowed location and character action to generate, the subsequent motion synthesis module concentrates solely on converting the set of events into character motions in a 3D scene in an online manner, detached from the holistic flow of the scene. Given the positions and actions of individual characters, the motion synthesis may employ any existing framework to reach the location and perform the designated action. An event is completed and removed from the scene when all characters in " + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "inline_equation", + "content": "\\{c_i\\}" + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": " complete their own target actions " + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 54, + 354, + 295, + 508 + ], + "type": "text", + "content": ". We briefly describe our motion synthesis module below, and more details are available in the supplementary material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": "We first plan a collision-free trajectory and generate locomotion along it such that the character " + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": " moves to the target position " + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": ", aligning its facing direction with the target orientation " + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "d_{i}" + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": ". While reaching the target position " + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "p_{i}" + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": ", the character should avoid collisions with other characters and the surrounding environment. We first create a 2D grid map that represents navigable areas in the scene, following the approach used in [38]. On this grid map, our motion synthesis module finds collision-free multi-agent paths using the windowed cooperative " + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "inline_equation", + "content": "A^{*}" + }, + { + "bbox": [ + 54, + 510, + 295, + 629 + ], + "type": "text", + "content": " algorithm proposed in [33]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": "After reaching the target position, the motion synthesis module creates motions assigned with the action label " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": ". We leverage the motion matching [6] to quickly generate plausible motions. If " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": " is deliberately set to overlap with an object (e.g., on a chair), the motion synthesis module first makes a transition to a desired posture (e.g., sitting) before creating motions for " + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "inline_equation", + "content": "a_{i}" + }, + { + "bbox": [ + 54, + 630, + 295, + 713 + ], + "type": "text", + "content": ". There are also group events," + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": "where multiple characters are involved, such as chatting or handshake. Because temporal synchronization is critical, the characters who reach " + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": " before others maintain an idle motion and wait for the remaining characters. Motion synthesis for " + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "inline_equation", + "content": "a_i" + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": " occurs only after all characters have reached their respective target position " + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 313, + 208, + 554, + 280 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 292, + 395, + 305 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 395, + 305 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 395, + 305 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 312, + 554, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 312, + 554, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 554, + 408 + ], + "type": "text", + "content": "The key functionality of our framework is to perform LLM-based scene-aware multi-agent planning to enable natural interactions and activities of characters within a given 3D scene. We can successfully populate natural motions of multiple characters within diverse scenes ranging from virtual scenes composed of CAD models (Figure 1) to 3D scans (Figure 5). More qualitative results are available in the supplementary materials." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 554, + 540 + ], + "type": "text", + "content": "To the best of our knowledge, our framework is the first to address this task at this scale; there is no existing baseline for direct comparison with our framework. Therefore, we propose evaluation criteria to comprehensively assess our framework (Section 4.1) and create a benchmark designed for this assessment (Section 4.2). Using the developed benchmark, we evaluate the effectiveness of our proposed key methodologies through an ablation study (Section 4.3). Furthermore, we conduct a user study to visually review and compare the generated results across the different ablation settings (Section 4.4)." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 549, + 429, + 560 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 549, + 429, + 560 + ], + "spans": [ + { + "bbox": [ + 313, + 549, + 429, + 560 + ], + "type": "text", + "content": "4.1. Evaluation Criteria" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 567, + 553, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 613 + ], + "type": "text", + "content": "We propose a set of criteria to assess several capabilities required for the task as shown below. We define each evaluation criterion as a tag and use these tags to categorize test cases in the benchmark." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 689 + ], + "type": "text", + "content": "- Object arrangement reasoning (OA): Object arrangement reasoning is essential for identifying objects based on spatial relationships (e.g., the chair farthest from the window) or ensuring that characters interact correctly with their surroundings (e.g., to study at a desk, a character needs to sit on the associated chair)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "- Regional context reasoning " + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "inline_equation", + "content": "(RC)" + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": ": Regional context reasoning is required for generating character behaviors that" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11658" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 289, + 177 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 289, + 177 + ], + "type": "image", + "image_path": "841df82702aa02d5c36d30dd6a4ff7be29d9782ac412cbf2fa81eb0bd52400f5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 295, + 220 + ], + "type": "text", + "content": "Figure 6. An example of a test case. In this test case, a system is required to identify the chairs in the kitchen and select one that is not currently occupied." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 63, + 230, + 295, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 230, + 295, + 266 + ], + "spans": [ + { + "bbox": [ + 63, + 230, + 295, + 266 + ], + "type": "text", + "content": "are well-aligned with the contextual meaning of a space (e.g., cooking in the kitchen rather than in the living room)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 266, + 295, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 295, + 313 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 295, + 313 + ], + "type": "text", + "content": "- Scene state reasoning (SS): Scene state reasoning requires considering the overall scene flow and the current state of other characters (e.g., if a character is using the coffee machine, another character should wait to use it)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 316, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 374 + ], + "type": "text", + "content": "In addition to the holistic spatial context defined above, we separately evaluate the low-level performance on Position inference " + }, + { + "bbox": [ + 55, + 316, + 296, + 374 + ], + "type": "inline_equation", + "content": "(PI)" + }, + { + "bbox": [ + 55, + 316, + 296, + 374 + ], + "type": "text", + "content": ". The cases with the tag analyze if the system can find the precise positions of characters at the coordinate level." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 383, + 177, + 394 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 177, + 394 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 177, + 394 + ], + "type": "text", + "content": "4.2. Benchmark Creation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": "Test Scenes We create three scenes—House, Office, and Restaurant—by placing the objects from the HSSD dataset [21]. Each scene is designed to include two to three separate areas with distinct contextual meaning. To ensure the LLM infers contextual meaning of areas without relying on object labels, we replace several object labels with general ones (e.g., kitchen_cabinet " + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": " cabinet). We also randomize indices for identically labeled objects to eliminate index-based arrangement cues. We define a distinct set of available action labels " + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "inline_equation", + "content": "A_{S}" + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": " for each scene, where " + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "inline_equation", + "content": "S \\in \\{\\text{house}, \\text{office}, \\text{restaurant}\\}" + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": ". This " + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "inline_equation", + "content": "A_{S}" + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": " is utilized when determining the pass/fail of test cases that use the scene " + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 400, + 296, + 544 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Test Cases Our framework can create natural behaviors for multiple characters without any pre-scripted scenarios. However, for rigorous evaluation, we introduce specific test scenarios in our benchmark. Each test case consists of a scene " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", characters " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{C} = \\{c_1, c_2, \\dots, c_N\\}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", test scenario " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", mock planning history " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ", and expected plan " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tau_e = (\\mathcal{A}_e, \\tilde{p}_e)" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": ". Here, " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " is the number of characters, " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_e" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " is the set of expected action labels, and " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{p}_e" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " is the expected area where the target character's position is expected to be included. " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\tilde{p}_e" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " is expressed through the union or intersection of multiple areas, each defined as a semantic description based on scene objects. For simplicity, we ignore the character's orientation in the benchmark, and the target character for the expected" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "text", + "content": "plan " + }, + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "inline_equation", + "content": "\\tau_{e}" + }, + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "text", + "content": " is always fixed to the first character " + }, + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "inline_equation", + "content": "c_{1}" + }, + { + "bbox": [ + 313, + 72, + 554, + 96 + ], + "type": "text", + "content": " in all test cases. An example of a test case is depicted in Figure 6." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": "When running a test case, we input the test scenario " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " and mock planning history " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " into the system. The mock planning history " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " simulates the prior planning history of the scene, building the necessary context required for the test case. Then, the system generates the most appropriate next plan " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tau_{g} = (a_{g},\\tilde{p}_{g}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "p_g)" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " by evaluating the given mock planning history and the test scenario. Here, " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " is the generated next plan for the target character, " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "a_{g}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " is the planned action label, and " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{p}_g" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "p_g" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " is the planned position. If the test case evaluates coordinate-level localization, the system outputs a coordinate-level position " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "p_{g}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": ". Otherwise, it outputs an area-level position " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tilde{p}_g" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": ". As the final step, the test case's pass/fail is determined by comparing the expected plan " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tau_{e}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": " and the generated plan " + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "inline_equation", + "content": "\\tau_{g}" + }, + { + "bbox": [ + 313, + 99, + 555, + 277 + ], + "type": "text", + "content": ". For a test case to pass, the following conditions must all be satisfied:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 279, + 553, + 340 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 314, + 279, + 539, + 292 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 279, + 539, + 292 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 539, + 292 + ], + "type": "text", + "content": "1. A plan for the target character is generated " + }, + { + "bbox": [ + 314, + 279, + 539, + 292 + ], + "type": "inline_equation", + "content": "(\\tau_{g}\\neq \\emptyset)" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 292, + 553, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 553, + 316 + ], + "type": "text", + "content": "2. The action label is included in both the available action labels and the expected action labels " + }, + { + "bbox": [ + 313, + 292, + 553, + 316 + ], + "type": "inline_equation", + "content": "(a_{g}\\in \\mathcal{A}_{\\mathcal{S}}\\cap \\mathcal{A}_{e})" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "text", + "content": "3. The character's position is within the expected area " + }, + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "inline_equation", + "content": "(\\tilde{p}_g \\subseteq \\tilde{p}_e" + }, + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "inline_equation", + "content": "p_g \\in \\tilde{p}_e)" + }, + { + "bbox": [ + 313, + 316, + 553, + 340 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": "Benchmark Summary We construct the benchmark based on the previously defined evaluation criteria, namely, tags. Our benchmark consists of 40 test cases in total. 10 test cases are designed for the " + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "inline_equation", + "content": "PI" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": " tag, while the remaining 30 test cases involve " + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "inline_equation", + "content": "OA" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "inline_equation", + "content": "RC" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "inline_equation", + "content": "SS" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": " tags. The test cases for the " + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "inline_equation", + "content": "PI" + }, + { + "bbox": [ + 313, + 365, + 555, + 568 + ], + "type": "text", + "content": " tag deliberately use straightforward scenarios to focus solely on the localization performance. Among the latter 30 test cases, half incorporate two tags for increased complexity. Additionally, for the latter 30 test cases, we vary the number of characters in the scene and the size of the mock planning history across three levels within the same scenario. This expands the number of effective test cases and allows us to evaluate the system's scalability by assessing its robustness against increasing agent counts and planning history complexity. We design our benchmark to ensure that each tag is evenly represented, preventing bias toward any specific evaluation factor." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Evaluation Metrics Our benchmark employs two evaluation metrics: success rate and execution rate. The success rate measures the proportion of passed runs among all test case runs. On the other hand, the execution rate measures the proportion of runs that are executed without any runtime errors, which are included in the parenthesis next to the success rate in tables. Runtime errors typically occur when the LLM module references a nonexistent object or generates a response with incorrect syntax. Any test case run with a runtime error is automatically marked as failed." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11659" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 553, + 190 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 190 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 190 + ], + "type": "table", + "html": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.9 (0.98)0.93 (0.99)0.9 (0.98)0.92 (0.98)0.74 (0.96)0.72 (0.95)0.72 (0.97)0.78 (0.93)0.72 (0.87)0.78 (0.91)0.76 (0.87)0.61 (0.81)
w/o Event0.82 (0.92)0.88 (0.92)0.86 (0.93)0.77 (0.92)0.6 (0.95)0.56 (0.99)0.6 (0.92)0.54 (0.92)0.6 (0.81)0.6 (0.78)0.62 (0.86)0.54 (0.76)
Object List0.51 (0.85)0.61 (0.78)0.28 (0.88)0.65 (0.97)0.34 (0.88)0.37 (0.88)0.12 (0.86)0.51 (0.9)0.35 (0.68)0.31 (0.7)0.29 (0.61)0.41 (0.69)
Scene Graph0.82 (0.96)0.8 (0.96)0.82 (0.94)0.87 (0.95)0.49 (0.9)0.52 (0.93)0.32 (0.9)0.51 (0.83)0.72 (0.92)0.69 (0.91)0.77 (0.94)0.64 (0.85)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
MetricsTotalOARCSSTotalOARCSSTotalOARCSS
Ours0.35 (0.74)0.37 (0.8)0.37 (0.77)0.35 (0.67)0.71 (0.9)0.71 (0.91)0.72 (0.95)0.65 (0.85)0.39 (0.87)0.38 (0.88)0.36 (0.87)0.35 (0.85)
w/o Event0.31 (0.58)0.3 (0.57)0.32 (0.6)0.28 (0.63)0.62 (0.85)0.74 (0.85)0.65 (0.9)0.59 (0.85)0.35 (0.79)0.41 (0.86)0.32 (0.79)0.32 (0.74)
Object List0.22 (0.58)0.26 (0.65)0.13 (0.62)0.23 (0.44)0.44 (0.77)0.45 (0.77)0.36 (0.83)0.52 (0.72)0.14 (0.77)0.09 (0.74)0.07 (0.77)0.24 (0.67)
Scene Graph0.29 (0.75)0.31 (0.76)0.35 (0.86)0.16 (0.61)0.69 (0.84)0.69 (0.81)0.77 (0.93)0.55 (0.72)0.31 (0.84)0.27 (0.88)0.23 (0.84)0.35 (0.84)
", + "image_path": "750e8d6b21c95af4e8452769d6f23b401386abf8892d2f905b368589f8246ca9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 555, + 220 + ], + "type": "text", + "content": "Table 1. Benchmark result for test cases with object arrangement reasoning (OA), regional context reasoning (RC), and scene state reasoning (SS) tags. In total, our method achieves the highest success rate across various LLM models and sizes." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 67, + 230, + 281, + 336 + ], + "blocks": [ + { + "bbox": [ + 67, + 230, + 281, + 336 + ], + "lines": [ + { + "bbox": [ + 67, + 230, + 281, + 336 + ], + "spans": [ + { + "bbox": [ + 67, + 230, + 281, + 336 + ], + "type": "image", + "image_path": "45e1ee3cf38af64f0a0e53a406289310acbb6f71238197d5d152b45869456e3f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 336, + 295, + 392 + ], + "lines": [ + { + "bbox": [ + 55, + 336, + 295, + 392 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 295, + 392 + ], + "type": "text", + "content": "Figure 7. Benchmark results by number of characters. Our approach exhibits stable success rate for stories with different number of characters while the performance gap increases as more characters are involved, compared to the version without event representation." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 67, + 393, + 278, + 464 + ], + "blocks": [ + { + "bbox": [ + 67, + 393, + 278, + 464 + ], + "lines": [ + { + "bbox": [ + 67, + 393, + 278, + 464 + ], + "spans": [ + { + "bbox": [ + 67, + 393, + 278, + 464 + ], + "type": "image", + "image_path": "5c907f3c4ddf01e44942aaf55a63c1f808eb977e6c75de650e5684f1c11684f9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 465, + 295, + 509 + ], + "lines": [ + { + "bbox": [ + 55, + 465, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 465, + 295, + 509 + ], + "type": "text", + "content": "Figure 8. Comparison of average token usage. Object List uses the least number of tokens, and Scene Graph uses significantly more number of tokens, especially when the scene complexity increases (Office)." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 521, + 170, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 170, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 170, + 533 + ], + "type": "text", + "content": "4.3. Benchmark Results" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 538, + 295, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 538, + 295, + 612 + ], + "spans": [ + { + "bbox": [ + 55, + 538, + 295, + 612 + ], + "type": "text", + "content": "We use GPT-4o [16], Llama-3.1 [8], and Qwen2.5 [43] in various sizes as LLM backbones to obtain benchmark results. The backbones are selected to represent both commercial and open-source modules that are popular and demonstrate competitive performance. We repeat each test case five times and average the results." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "Event-Driven Planning Table 1 presents the benchmark results for test cases with " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "OA" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "RC" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "inline_equation", + "content": "SS" + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": " tags. Each column represents the average score for test cases that include the corresponding tag and each cell displays the success rate, with the execution rate shown in parentheses. The ablation without event-driven planning (w/o Event) handles the entire process with a single unified LLM module without using high-level event representations. Therefore, all in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "text", + "content": "put and output information, except for the scene description and test scenario, is represented in a low-level structured format. Compared to " + }, + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "text", + "content": " Event, our method benefits from a cascaded planning pipeline that effectively decomposes the entire planning complexity. This leads to an overall improvement in planning performance. In particular, the performance gain in SS is relatively larger compared to other tags in GPT-40 models. The SS tag requires a comprehensive understanding of the overall planning state of multiple characters in the scene. We attribute this to the advantages of our event-based representation, which provides a more efficient and intuitive way to capture the overall scene state. Additionally, as shown in Figure 7, our planning pipeline proves to be more effective in terms of scalability. " + }, + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "inline_equation", + "content": "w/o" + }, + { + "bbox": [ + 313, + 232, + 555, + 435 + ], + "type": "text", + "content": " Event struggles with planning as the number of characters in the scene increases, whereas our approach maintains stable performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 449, + 556, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 677 + ], + "type": "text", + "content": "Scene Description We also analyze the effect of using scene description generated from an LLM module. We create ablation settings that retain our pipeline but modify the scene description method. Object List represents the scene as a simple list of objects present in the scene, while Scene Graph uses the JSON-formatted scene graph representation as the scene description. Our approach contributes to more effective planning compared to other description methods as shown in Table 1. Additionally, we present the average token usage for each description method in Figure 8 to provide a deeper understanding of their differences. As presented in Figure 8, the object list is the most efficient way to represent scene information and has often been used in previous studies such as [5, 20, 34]. However, since the object list does not have any spatial information, the planning module with the description is ignorant of object arrangement or spatial layout. As a result, the object list records the lowest scores in our benchmark, which requires sophisticated scene understanding for planning." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Unlike the object list, the scene graph representation enables necessary spatial reasoning by incorporating detailed attributes of objects and spatial relationships between ob" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11660" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 75, + 70, + 277, + 137 + ], + "blocks": [ + { + "bbox": [ + 75, + 70, + 277, + 137 + ], + "lines": [ + { + "bbox": [ + 75, + 70, + 277, + 137 + ], + "spans": [ + { + "bbox": [ + 75, + 70, + 277, + 137 + ], + "type": "table", + "html": "
ModelGPT-4oGPT-4o miniLlama-3.1-70B
Ours0.94 (1.0)0.9 (1.0)0.94 (1.0)
Direct Inference0.46 (1.0)0.38 (1.0)0.28 (1.0)
ModelLlama-3.1-8BQwen2.5-72BQwen2.5-7B
Ours0.86 (0.88)1.0 (1.0)0.94 (1.0)
Direct Inference0.14 (0.9)0.14 (0.48)0.1 (0.8)
", + "image_path": "1afa5562a2826a805df55a656824600dd25a926da20e4e72652fd9871f1afc91.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 178 + ], + "type": "text", + "content": "Table 2. Benchmark result for test cases with position inference tag (PI). Our proposed method can stably infer accurate positions, whereas Direct Inference with LLMs often struggles." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 186, + 297, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 186, + 297, + 353 + ], + "spans": [ + { + "bbox": [ + 54, + 186, + 297, + 353 + ], + "type": "text", + "content": "jects. However, the representation contains exhaustive detail, including even minor and insignificant ones, without any abstraction or compression. This inefficiency leads to excessive token consumption (Figure 8) and can even interfere with the LLM's reasoning due to unnecessary details (Table 1). The token inefficiency becomes more pronounced as the scene size and complexity increase. As shown in Figure 8, when comparing the relatively simple 'House' scene to the more complex 'Office' scene, the scene graph exhibits a significantly larger increase in token consumption compared to alternative approaches. Compared to the scene graph, our approach allows for a more efficient description of the given scene while preserving key spatial information." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 362, + 296, + 507 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 362, + 296, + 507 + ], + "spans": [ + { + "bbox": [ + 55, + 362, + 296, + 507 + ], + "type": "text", + "content": "Positional Inference (PI) In Table 2, we demonstrate the effectiveness of our area-conditioned position sampling method based on the results for the PI-tagged test cases. Direct Inference is an ablation setting where the same planning pipeline is used, but the LLM directly outputs coordinates when determining character positions. Although all test cases with the PI tag are intentionally designed to be very simple, all LLM models with Direct Inference struggle to accurately determine the coordinates of character positions. In contrast, our system allows the LLM to process information semantically, and obtains precise locations through the proposed area-conditioned position sampling technique." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 167, + 529 + ], + "type": "text", + "content": "4.4. User Study Results" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "We also validate through a user study that our approach represents the given scenario more effectively than other ablation settings. For the user study, we use a total of four test scenarios, each set in a different scene. The scenes include House, Office, and Restaurant from the benchmark, as well as MPH11 from the PROX dataset [12]. Compared to the other scenes, MPH11 is relatively simple in both scale and complexity. Given its limited space, we use a simple scenario containing only two events for the MPH11. In contrast, for other test scenes, we use more complex scenarios that include 4-5 events. For each test scenario, users are asked to visually compare results generated from different ablation settings and select the one that best represents the given scenario. If multiple results are perceived as equally effective, users are allowed to choose more than one. We" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 326, + 69, + 541, + 130 + ], + "blocks": [ + { + "bbox": [ + 326, + 69, + 541, + 130 + ], + "lines": [ + { + "bbox": [ + 326, + 69, + 541, + 130 + ], + "spans": [ + { + "bbox": [ + 326, + 69, + 541, + 130 + ], + "type": "image", + "image_path": "95d295ffc686b4380dde9ff5b00d25bd99e59bf7f9feba4c3539b75741b2e782.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 326, + 133, + 541, + 191 + ], + "blocks": [ + { + "bbox": [ + 326, + 133, + 541, + 191 + ], + "lines": [ + { + "bbox": [ + 326, + 133, + 541, + 191 + ], + "spans": [ + { + "bbox": [ + 326, + 133, + 541, + 191 + ], + "type": "image", + "image_path": "040e2bb95e94b12f51e808d1a098e2522ee7da6fb983f8f442327e5299fcec06.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 193, + 553, + 236 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 236 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 236 + ], + "type": "text", + "content": "Figure 9. User study results. Our system correctly generates the dynamic story line with multiple characters in large scenes, significantly outperforming other baselines, except for simple scenarios with MPH11." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 249, + 553, + 272 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 249, + 553, + 272 + ], + "spans": [ + { + "bbox": [ + 313, + 249, + 553, + 272 + ], + "type": "text", + "content": "generate the results for the user study using GPT-4o mini with a temperature setting of 0.0." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 417 + ], + "type": "text", + "content": "We collected responses from 50 general participants. The responses to the user study are summarized in Figure 9. The chart labels SF, OL, and No E correspond to the same ablation settings used in the benchmark: Scene Graph, Object List, and w/o Event, respectively. In three of our scenarios, our approach received the most selections by a large margin. Furthermore, the noticeable difference in selection distribution between the MPH11 scene, which involves a relatively smaller-scale problem, and the other scenes suggests that our approach contributes to building a more scalable system. We provide the test scenarios and visualization of generated results in the supplementary material." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 430, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 430, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 430, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 555, + 715 + ], + "type": "text", + "content": "In this study, we propose an LLM-based framework for generating multiple lifelike virtual humans within a given 3D scene, dynamically aligning their behaviors with an emergent storyline. Characters generated by our framework interact contextually with their surroundings and other individuals, enhancing scene realism. Our framework performs multi-character behavior planning based on semi-narrative events, allowing the LLM to efficiently capture scene progression and group behaviors. The modular design of our planning pipeline decomposes the overall problem, reducing reasoning complexity for the LLM and enabling more scalable planning. Additionally, we introduce a context-centric scene description to efficiently convey key spatial information and propose area-conditioned position sampling to mitigate the LLM's weaknesses in numerical reasoning. To evaluate our system, we construct a benchmark for a comprehensive assessment of LLM-based scene-aware multi-agent behavior planning. Through benchmark results and a user study, we demonstrate that our proposed methods contribute to building a scalable and effective system for coordinating multi-agent behaviors, while carefully considering the scene." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11661" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 145 + ], + "type": "text", + "content": "Acknowledgments This work was supported by the IITP grant [RS-2021-II211341, Artificial Intelligence Graduate School Program (Chung-Ang University)], the NRF grant (No. RS-2023-00208197) funded by the Korea government (MSIT), and Creative-Pioneering Researchers Program through Seoul National University." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 115, + 166 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 173, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 173, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 173, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 61, + 173, + 294, + 237 + ], + "type": "text", + "content": "[1] Joao Pedro Araujo, Jiaman Li, Karthik Vetrivel, Rishi Agarwal, Jiajun Wu, Deepak Gopinath, Alexander William Clegg, and Karen Liu. Circle: Capture in rich contextual environments. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21211-21221, 2023. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 293 + ], + "type": "text", + "content": "[2] Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances in neural information processing systems, 33:1877-1901, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 295, + 296, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 295, + 296, + 360 + ], + "spans": [ + { + "bbox": [ + 61, + 295, + 296, + 360 + ], + "type": "text", + "content": "[3] Zhongang Cai, Jianping Jiang, Zhongfei Qing, Xinying Guo, Mingyuan Zhang, Zhengyu Lin, Haiyi Mei, Chen Wei, Ruisi Wang, Wanqi Yin, et al. Digital life project: Autonomous 3d characters with social intelligence. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 582-592, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 62, + 361, + 294, + 425 + ], + "type": "text", + "content": "[4] Zhi Cen, Huaijin Pi, Sida Peng, Zehong Shen, Minghui Yang, Shuai Zhu, Hujun Bao, and Xiaowei Zhou. Generating human motion in 3d scenes from text descriptions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1855-1866, 2024. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "spans": [ + { + "bbox": [ + 62, + 427, + 294, + 480 + ], + "type": "text", + "content": "[5] Jianqi Chen, Panwen Hu, Xiaojun Chang, Zhenwei Shi, Michael Christian Kampffmeyer, and Xiaodan Liang. Sitcom-crafter: A plot-driven human motion generation system in 3d scenes. arXiv preprint arXiv:2410.10790, 2024. 2, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 482, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 482, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 62, + 482, + 294, + 503 + ], + "type": "text", + "content": "[6] Simon Clavet et al. Motion matching and the road to next-gen animation. In Proc. of GDC, page 4, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 505, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 505, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 62, + 505, + 294, + 557 + ], + "type": "text", + "content": "[7] Arash Gholami Davoodi, Seyed Pouyan Mousavi Davoudi, and Pouya Pezeshkpour. Llms are not intelligent thinkers: Introducing mathematical topic tree benchmark for comprehensive evaluation of llms. arXiv preprint arXiv:2406.05194, 2024. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 559, + 294, + 612 + ], + "type": "text", + "content": "[8] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024. 2, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 294, + 656 + ], + "type": "text", + "content": "[9] Martin Ester, Hans-Peter Kriegel, Jörg Sander, Xiaowei Xu, et al. A density-based algorithm for discovering clusters in large spatial databases with noise. In kdd, pages 226-231, 1996. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 712 + ], + "type": "text", + "content": "[10] Anindita Ghosh, Rishabh Dabral, Vladislav Golyanik, Christian Theobalt, and Philipp Slusallek. Remos: 3d motion-conditioned reaction synthesis for two-person interactions. In European Conference on Computer Vision, pages 418-437. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[11] Maitrey Gramopadhye and Daniel Szafir. Generating executable action plans with environmentally-aware language models. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3568-3575. IEEE, 2023. 1, 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 182 + ], + "type": "text", + "content": "[12] Mohamed Hassan, Vasileios Choutas, Dimitrios Tzionas, and Michael J Black. Resolving 3d human pose ambiguities with 3d scene constraints. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2282-2292, 2019. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 317, + 184, + 553, + 237 + ], + "type": "text", + "content": "[13] Mohamed Hassan, Duygu Ceylan, Ruben Villegas, Jun Saito, Jimei Yang, Yi Zhou, and Michael J Black. Stochastic scene-aware motion prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 11374-11384, 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 303 + ], + "type": "text", + "content": "[14] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. In Proceedings of the 41st International Conference on Machine Learning, pages 20413-20451, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 358 + ], + "type": "text", + "content": "[15] Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. Language models as zero-shot planners: Extracting actionable knowledge for embodied agents. In International conference on machine learning, pages 9118-9147. PMLR, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[16] Aaron Hurst, Adam Lerer, Adam P Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, et al. Gpt-4o system card. arXiv preprint arXiv:2410.21276, 2024. 2, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 459 + ], + "type": "text", + "content": "[17] Baoxiong Jia, Yixin Chen, Huangyue Yu, Yan Wang, Xuesong Niu, Tengyu Liu, Qing Li, and Siyuan Huang. Sceneverse: Scaling 3d vision-language learning for grounded scene understanding. In European Conference on Computer Vision, pages 289-310. Springer, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 553, + 525 + ], + "type": "text", + "content": "[18] Bowen Jiang, Yangxinyu Xie, Zhuoqun Hao, Xiaomeng Wang, Tanwi Mallick, Weijie Su, Camillo Taylor, and Dan Roth. A peek into token bias: Large language models are not yet genuine reasoners. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 4722-4756, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "spans": [ + { + "bbox": [ + 317, + 526, + 553, + 570 + ], + "type": "text", + "content": "[19] Nan Jiang, Zimo He, Zi Wang, Hongjie Li, Yixin Chen, Siyuan Huang, and Yixin Zhu. Autonomous character-scene interaction synthesis from text instruction. In SIGGRAPH Asia 2024 Conference Papers, pages 1-11, 2024. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 624 + ], + "type": "text", + "content": "[20] Shyam Sundar Kannan, Vishnunandan LN Venkatesh, and Byung-Cheol Min. Smart-llm: Smart multi-agent robot task planning using large language models. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 12140–12147. IEEE, 2024. 2, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 625, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 712 + ], + "type": "text", + "content": "[21] Mukul Khanna, Yongsen Mao, Hanxiao Jiang, Sanjay Haresh, Brennan Shacklett, Dhruv Batra, Alexander Clegg, Eric Undersander, Angel X Chang, and Manolis Savva. Habitat synthetic scenes dataset (hssd-200): An analysis of 3d scene scale and realism tradeoffs for objectgoal navigation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16384–16393, 2024. 6" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11662" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[22] Boyi Li, Philipp Wu, Pieter Abbeel, and Jitendra Malik. Interactive task planning with language models. arXiv preprint arXiv:2310.10645, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 107, + 296, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 107, + 296, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 107, + 296, + 152 + ], + "type": "text", + "content": "[23] Han Liang, Wenqian Zhang, Wenxuan Li, Jingyi Yu, and Lan Xu. Intergen: Diffusion-based multi-human motion generation under complex interactions. International Journal of Computer Vision, 132(9):3463-3483, 2024. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 153, + 295, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 295, + 207 + ], + "type": "text", + "content": "[24] Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493-9500. IEEE, 2023. 1, 2, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 57, + 209, + 295, + 262 + ], + "type": "text", + "content": "[25] Yunze Liu, Changxi Chen, Chenjing Ding, and Li Yi. Phys-reaction: Physically plausible real-time humanoid reaction synthesis via forward dynamics guided 4d imitation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 3771-3780, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 57, + 264, + 295, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 264, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 57, + 264, + 295, + 308 + ], + "type": "text", + "content": "[26] Aymen Mir, Xavier Puig, Angjoo Kanazawa, and Gerard Pons-Moll. Generating continual human motion in diverse 3d scenes. In 2024 International Conference on 3D Vision (3DV), pages 903-913. IEEE, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 310, + 295, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 295, + 364 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 295, + 364 + ], + "type": "text", + "content": "[27] Iman Mirzadeh, Keivan Alizadeh, Hooman Shahrokhi, Oncel Tuzel, Samy Bengio, and Mehrdad Farajtabar. Gsm-symbolic: Understanding the limitations of mathematical reasoning in large language models. arXiv preprint arXiv:2410.05229, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 57, + 365, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 365, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 57, + 365, + 295, + 430 + ], + "type": "text", + "content": "[28] Zhe Ni, Xiaoxin Deng, Cong Tai, Xinyue Zhu, Qinghongbing Xie, Weihang Huang, Xiang Wu, and Long Zeng. Grid: Scene-graph-based instruction-driven robotic task planning. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13765-13772. IEEE, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "spans": [ + { + "bbox": [ + 57, + 433, + 295, + 487 + ], + "type": "text", + "content": "[29] Joon Sung Park, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1-22, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 489, + 295, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 489, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 57, + 489, + 295, + 533 + ], + "type": "text", + "content": "[30] Krishan Rana, Jesse Haviland, Sourav Garg, Jad Abou-Chakra, Ian D Reid, and Niko Sunderhauf. Sayplan: Grounding large language models using 3d scene graphs for scalable robot task planning. In CoRL, 2023. 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "spans": [ + { + "bbox": [ + 57, + 534, + 295, + 588 + ], + "type": "text", + "content": "[31] Mengyi Shan, Lu Dong, Yutao Han, Yuan Yao, Tao Liu, Ifeoma Nwogu, Guo-Jun Qi, and Mitch Hill. Towards open domain text-driven synthesis of multi-person motions. In European Conference on Computer Vision, pages 67-86. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 644 + ], + "type": "text", + "content": "[32] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael Scharli, and Denny Zhou. Large language models can be easily distracted by irrelevant context. In International Conference on Machine Learning, pages 31210-31227. PMLR, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[33] David Silver. Cooperative pathfinding. In Proceedings of the aaai conference on artificial intelligence and interactive digital entertainment, pages 117-122, 2005. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 713 + ], + "type": "text", + "content": "[34] Ishika Singh, Valts Blukis, Arsalan Mousavian, Ankit Goyal, Danfei Xu, Jonathan Tremblay, Dieter Fox, Jesse Thomason, and Animesh Garg. Progprompt: Generating situated robot" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "task plans using large language models. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 11523-11530. IEEE, 2023. 1, 2, 4, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 555, + 152 + ], + "type": "text", + "content": "[35] Mikihiro Tanaka and Kent Fujiwara. Role-aware interaction generation from textual description. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15999-16009, 2023. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 153, + 553, + 206 + ], + "type": "text", + "content": "[36] Guanzhi Wang, Yuqi Xie, Yunfan Jiang, Ajay Mandlekar, Chaowei Xiao, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Voyager: An open-ended embodied agent with large language models. Transactions on Machine Learning Research, 2024. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 262 + ], + "type": "text", + "content": "[37] Jiashun Wang, Huazhe Xu, Jingwei Xu, Sifei Liu, and Xiaolong Wang. Synthesizing long-term 3d human motion and interaction in 3d scenes. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9401-9411, 2021. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 319 + ], + "type": "text", + "content": "[38] Jingbo Wang, Yu Rong, Jingyuan Liu, Sijie Yan, Dahua Lin, and Bo Dai. Towards diverse and natural scene-aware 3d human motion synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20460-20469, 2022. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[39] Zan Wang, Yixin Chen, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Humanise: Language-conditioned human motion generation in 3d scenes. Advances in Neural Information Processing Systems, 35:14959-14971, 2022." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 365, + 553, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 553, + 441 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 553, + 441 + ], + "type": "text", + "content": "[40] Zan Wang, Yixin Chen, Baoxiong Jia, Puhao Li, Jinlu Zhang, Jingze Zhang, Tengyu Liu, Yixin Zhu, Wei Liang, and Siyuan Huang. Move as you say interact as you can: Language-guided human motion generation with scene affordance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 433-444, 2024. 1, 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 316, + 444, + 553, + 498 + ], + "type": "text", + "content": "[41] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824-24837, 2022. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 553, + 533 + ], + "type": "text", + "content": "[42] Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is inevitable: An innate limitation of large language models. arXiv preprint arXiv:2401.11817, 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[43] An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, et al. Qwen2.5 technical report. arXiv preprint arXiv:2412.15115, 2024. 2, 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 643 + ], + "type": "text", + "content": "[44] Jianing Yang, Xuweiyi Chen, Shengyi Qian, Nikhil Madaan, Madhavan Iyengar, David F Fouhey, and Joyce Chai. Llm-grounder: Open-vocabulary 3d visual grounding with large language model as an agent. In 2024 IEEE International Conference on Robotics and Automation (ICRA), pages 7694-7701. IEEE, 2024. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[45] Hongwei Yi, Justus Thies, Michael J Black, Xue Bin Peng, and Davis Rempe. Generating human interaction motions in scenes with text control. In European Conference on Computer Vision, pages 246-263. Springer, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[46] Hongxin Zhang, Weihua Du, Jiaming Shan, Qinhong Zhou, Yilun Du, Joshua B Tenenbaum, Tianmin Shu, and Chuang" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11663" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 162 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "Gan. Building cooperative embodied agents modularly with large language models. In The Twelfth International Conference on Learning Representations, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 162 + ], + "type": "text", + "content": "[47] Kaifeng Zhao, Yan Zhang, Shaofei Wang, Thabo Beeler, and Siyu Tang. Synthesizing diverse human motions in 3d indoor scenes. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14738-14749, 2023. 1, 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11664" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_content_list.json b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..cf2b9a9e9672414a3e2f419566c99f84a2481257 --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Event-aided Dense and Continuous Point Tracking: Everywhere and Anytime", + "text_level": 1, + "bbox": [ + 107, + 130, + 890, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Zhexiong Wan $^{1,2}$ Jianqin Luo $^{1}$ Yuchao Dai $^{1}$ Gim Hee Lee $^{2}$", + "bbox": [ + 235, + 179, + 759, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ School of Electronics and Information, Northwestern Polytechnical University & Shaanxi Key Laboratory of Information Acquisition and Processing", + "bbox": [ + 174, + 198, + 820, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Department of Computer Science, National University of Singapore", + "bbox": [ + 227, + 233, + 769, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{wanzhexiong,luojianqin}@mail.nwpu.edu.cn，daiyuchao@nwpu.edu.cn，gimhee.lee@nus.edu.sg", + "bbox": [ + 107, + 253, + 883, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 304, + 326, + 319 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent point tracking methods have made great strides in recovering the trajectories of any point (especially key points) in long video sequences associated with large motions. However, the spatial and temporal granularities of point trajectories remain constrained by limited motion estimation accuracy and video frame rate. Leveraging the high temporal resolution and motion sensitivity of event cameras, we introduce event data for the first time to recover spatially dense and temporally continuous trajectories of every point at any time. Specifically, we define the dense and continuous point trajectory representation as estimating multiple control points of curves for each pixel and model the movement of sparse events triggered along continuous point trajectories. Building on this, we propose a novel multi-frame iterative streaming framework that first estimates local inter-frame motion representations from two consecutive frames with inter-frame events, then aggregates them into a global long-term motion representation to utilize input full video and event data with an arbitrary number of frames. Extensive experiments on simulated and real data demonstrate the significant improvement of our framework over state-of-the-art methods and the crucial role of introducing events to model continuous point trajectories.", + "bbox": [ + 89, + 335, + 485, + 684 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 713, + 220, + 729 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Estimating fine-grained motion from input videos is a crucial task in computer vision with widespread applications such as video compression [1], frame interpolation [23, 55], motion segmentation [2, 34], and dynamic scene reconstruction [16]. Early studies [21, 44] based on two-frame optical flow can model the spatially dense motion, but they suffer from the challenge of modeling long-term motion from more frames. With the proposal of tracking any point (TAP) task [5, 19], using sparse points as query indexes to", + "bbox": [ + 89, + 739, + 483, + 876 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "estimate pointwise long-term motions also draws attention. Despite significant progress, the sparse and independent representation of pointwise trajectories remains inherently incompatible with the spatially dense representation of input video. Thus, recently there are methods [4, 35, 46] that continue to use dense optical flow to represent long-term motion. However, the temporal frame rate of the input video is constrained by traditional shutter cameras, which makes accurately modeling complex motion trajectories difficult for video-only methods from the data acquisition perspective. Consequently, accurate estimation of fine-grained spatially dense and temporally continuous motion remains a challenging and worthwhile research problem.", + "bbox": [ + 511, + 305, + 906, + 501 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unlike traditional shutter cameras that expose the entire image at fixed frame rates, the new bio-inspired event cameras [9] can independently and asynchronously detect per-pixel brightness changes with microsecond precision. This unique design makes it inherently sensitive to visual motion changes, leading to impressive motion-related applications such as optical flow estimation [17, 62], motion segmentation [20, 61], feature tracking [33], object tracking [63], and frame interpolation [45]. However, events are typically triggered only with motion contours or rich textures, making it challenging to comprehensively perceive spatially dense motion. Recently, integrating the advantages of event camera and traditional shutter camera has become a new direction [38, 57]. Leveraging on these successes, we propose for the first time taking event data as an auxiliary input for the point tracking task, where we comprehensively model fine-grained spatially dense and temporally continuous motion from the input image and event modalities.", + "bbox": [ + 511, + 505, + 908, + 776 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The introduction of event cameras offers the potential to model continuous motion from the data perspective. However, to effectively model continuous long-term motion, we need a new representation that replaces optical flow and parametrically associates the dense temporal properties of event data. BFlow [14] proposes to learn trajectories with Bezier curves from events, but is limited to input fixed frames and cannot adapt to longer sequences. CPFlow [31]", + "bbox": [ + 511, + 779, + 910, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://npuctr.github.io/EDCPT", + "bbox": [ + 112, + 887, + 426, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7936", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "proposes to learn control points represented as B-spline curves from a fixed number of image slices. Although continuous motion can be successfully modeled using these curve representations in normalized timescales, their fixed number of control points made it hard to handle complex dynamics and varying lengths of video sequences. Based on the curve representation, we propose a new streaming pipeline for accumulating multiple local curves to address these limitations. In addition, how to adequately fuse multi-frame temporal information to achieve long-term tracking is also a concern. Compared to not explicitly associating multi-frames in DOT [35] and aggregating over estimated motion vectors in AccFlow [52] and FlowTrack [4], we perform recurrent fusion in the feature space, dedicated to preserving multi-frame sequence information.", + "bbox": [ + 89, + 90, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present the first event-aided point tracking framework for recovering spatially dense and temporally continuous point trajectories from input videos and event sequences. Specifically, we first propose a new point trajectory representation with parametric curves that accumulate multiple local curves to adapt to multi-frame input videos at any length. We then design a new framework for combining two frames with events to estimate dense point curve trajectories and extend to multi-frame iterative accumulation. In addition, we establish the association between continuous curve trajectory and event triggering as a part of the learning objective for continuous motion modeling. Extensive experiments on both simulated and real-world data demonstrate that the proposed framework significantly outperforms state-of-the-art methods. Particularly, our ablation studies illustrate the effectiveness of the proposed global aggregation and highlight the crucial role of incorporating event data in continuous trajectory modeling.", + "bbox": [ + 91, + 319, + 483, + 590 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are summarized as follows:", + "bbox": [ + 109, + 592, + 449, + 606 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a new setup that, for the first time, enables long-term spatially dense and temporally continuous point tracking by integrating the strengths of both images and events.", + "- We present a novel global curve representation of continuous point trajectories through multi-frame aggregation, establishing a connection between event triggering and continuous motion.", + "- We propose a novel event-aided streaming framework that iteratively accumulates the local tracks from two frames with inter-frame events, resulting in global, long-term dense and continuous trajectories through iterative temporal global motion aggregation." + ], + "bbox": [ + 109, + 608, + 483, + 804 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 89, + 821, + 240, + 835 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Image-based query point tracking", + "text_level": 1, + "bbox": [ + 89, + 847, + 387, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The goal of point tracking is to recover the corresponding positions of query points in each frame, which has attracted", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "wide attention with the proposal of the TAP benchmark and the baseline model TAPNet [5]. PIPs [19] proposes to extract independent point representation for 8-frame tracking, then $\\mathrm{PIPs}++$ [58] extends to long-term trajectories. TAPIR [6] proposes a two-stage matching framework that fuses TAPNet and PIPs, and BootsTAPIR [7] performs self-supervised training on more data. Unlike these methods that track only one query at a time, CoTracker [24] and Context-PIPs [51] use additional tracks and pixel featuresto improve global tracking. SpatialTracker [54] introduces the triplane representation for group pixels in 3D.DINOTracker [46] performs test-time finetuning on the pre-trained DINOViT [37] model and achieves fine-grained tracking per video. When applying these query points-based methods to achieve dense tracking, full points need to be processed individually or in batches, which brings computational hurdles and limits their downstream applications [35].", + "bbox": [ + 511, + 90, + 903, + 349 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Image-based dense point tracking", + "text_level": 1, + "bbox": [ + 511, + 359, + 808, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies turn to tracking every point within a frame in a single run, aiming to enhance neighborhood relationships while reducing computational requirements. OmniMotion [49] performs pixel-wise tracking between local and canonical space to maintain the global consistency of the motion, and then FastOmniTrack [42] and DecoMotion [26] improve from the perspectives of computation cost and object motion decomposition. CPFlow [31] proposes to estimate spatio-temporally dense motion curves, but it can only input 4 images and needs pre-sampling for longer video. AccFlow [52] proposes the forward and backward aggregation pipeline, extending inter-frame dense optical flow to multi-frame long ranges. MFT [36] select chaining multiframe candidates and FlowTrack [4] automatically apply error compensation in instances of tracking inaccuracies. DOT [35] unifies point tracking and optical flow, upgrading sparse tracks to dense flow fields between every frame. However, limited by the video frame rate, these methods are struggling to model challenging dynamics. In this work, we propose to introduce continuous event data into the input video to enable temporally continuous point tracking.", + "bbox": [ + 511, + 382, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. Event-based motion estimation", + "text_level": 1, + "bbox": [ + 511, + 710, + 785, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Thanks to the motion-sensitive nature of event cameras, extensive motion estimation studies in recent years have highlighted their potential applications to challenging dynamics. The feature tracking methods [11, 27, 33, 50] show the benefits of event cameras for low-latency tracking, but can only track sparse, specific textured locations. Recently estimating optical flow from events has become mainstream. Using only sparse event data [13, 32, 62] allows to estimate satisfactory dense optical flow, while introducing data from other sensors such as images [47, 59] and point clouds [48, 60] achieves significant performance", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7937", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "gains. BFlow [14] and MotionPriorCMax [18] exploit the continuous property of events to estimate parametric Bezier trajectories, but can only estimate motion within a fixed consecutive frame interval and cannot be directly adapted to long-term sequences. Recently, FE-TAP [29] proposes to recover point trajectories from a fixed number of images and events based on TAPVid [5], but does not take full advantage of the continuous nature of events. We propose to combine the advantages of images and events to enable temporally continuous point tracking by modeling long-term global motion with an arbitrary number of frames.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 280, + 181, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview. To address the limitations of sparse events and low video frame rates in modeling fine-grained motion, we present the first framework that recovers dense and continuous point trajectories from a video with corresponding event sequences. Our framework consists of four parts: 1) A parametric multi-frame continuous point trajectories representation; 2) A model of event triggering along the point trajectories; 3) A two-frame basis motion estimation model; 4) A multi-frame motion aggregation and streaming framework.", + "bbox": [ + 89, + 308, + 483, + 445 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation. Conventional shutter camera captures a video with $N_{v}$ frames $\\{I_i\\}_{i=1}^{N_v}$ at a fixed frame rate. Event camera triggers an unbounded event sequence $\\{e_i\\}_{i=1}^{N_e}$ , where $N_e$ is the number of events. Each event $e_i = \\{\\mathbf{x}_i, t_i, p_i\\}$ consists of the pixel position $\\mathbf{x}_i = (x,y)$ , timestamp $t_i$ with microsecond precision, and the brightness change polarity $p_i$ in log domain. Our goal is to combine these two modalities to recover the spatially dense and temporally continuous trajectories $\\mathbf{T}_{1 \\to t}$ of all points starting from the first frame at time 1 to last frame at time $t$ .", + "bbox": [ + 89, + 448, + 483, + 599 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Motion model", + "text_level": 1, + "bbox": [ + 89, + 617, + 235, + 632 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Trajectory representation. Previous point tracking methods typically estimate two-channel motion vectors in the XY directions, which is the optical flow when representing dense trajectories [4, 35]. To learn the curve trajectory from the deep network, we instead learn the multiple control points as curved trajectories [31]. Specifically, we choose the B-spline curve as our curve representation, which is defined by $N_{c}$ control points $\\{\\mathbf{P}_i\\}_{i=1}^{N_c}$ and basis functions $\\{B_{i,p}(t)\\}_{i=1}^{N_c}$ with degree $p$ . The continuous trajectory $\\mathbf{T}(t)$ represented by b-spline curve in time variable $t$ is a collection of piecewise polynomial functions $\\mathbf{T}(t) = \\sum_{i=1}^{N_c} B_{i,p}(t) \\mathbf{P}_i$ . Similar to learning dense flow, each pixel has an independent curve with control points denoted as $\\mathbb{P} \\in \\mathbb{R}^{N_c \\times 2 \\times H \\times W}$ , where $H \\times W$ is the image size. This enables learnable motion modeling of dense and continuous point trajectories $\\mathbf{T}$ with parametric curve representation. More details are provided in the Supp.", + "bbox": [ + 89, + 642, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Multi-frame global trajectories accumulation. Existing parametric motion modeling methods are fixed in the number of frames they can handle, e.g., BFlow [14] is limited to between two frames, and CPFlow [31] struggles to benefit from more than 4 frame inputs, resulting in suboptimal long-term trajectory modeling. Inspired by the practice of multi-frame optical flow accumulation [36, 52], we propose a new multi-frame curve trajectories accumulation strategy to handle long-term videos with arbitrary frames.", + "bbox": [ + 511, + 90, + 903, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our accumulation framework works on a streaming pipeline, where the previous global trajectory $\\mathbf{T}_{1\\rightarrow t}$ with $(t - 1)\\times N_{c}$ control points has been accumulated from the previous $t - 1$ local trajectories $\\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1}$ when processing the $t$ -th step. When we get updated $t$ -th local trajectory $\\mathbf{T}_{t\\rightarrow t + 1}$ with $N_{c}$ control points from time $t$ to $t + 1$ , a simple approach is to directly accumulate the new global trajectory with $t\\times N_{c}$ control points from time 1 to $t + 1$ by $\\mathbf{T}_{1\\rightarrow t + 1}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathrm{Warp}(\\mathbf{T}_{t\\rightarrow t + 1},\\mathbf{T}_{1\\rightarrow t})(\\mathbf{x})\\right]$ , where Warp represents backward warping, the operator [,] combines the control points of two sub-curves and creates a more complex curve. However, there are two problems with just warping operation: 1) It suffers from numerical error as integer sampling with floating-point coordinates, i.e., for warping vectors from $\\mathbf{b}$ to $\\mathbf{a}$ , $\\mathrm{Warp}(\\mathbf{a},\\mathbf{b})(\\mathbf{x}) = \\mathbf{a}(\\mathbf{x} + \\mathbf{b}(\\mathbf{x}))$ , $\\mathbf{x}$ is integer coordinates but not $\\mathbf{x} + \\mathbf{b}(\\mathbf{x})$ . 2) Some points may be occluded at time $t$ , resulting in failing to find the corresponding points. More details are provided in the Supp.", + "bbox": [ + 511, + 232, + 906, + 503 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our framework iteratively maintains and learns to integrate from a global motion representation $\\mathbf{M}_{1\\rightarrow t}^{global}$ in the streaming process (cf. Sec. 3.2). For the first numerical problem, we estimate a start point offsets $\\mathbf{O}_t\\in \\mathbb{R}^{2\\times H\\times W}$ learned from $\\mathbf{M}_{1\\rightarrow t}^{global}$ and normalized to the range $[-1,1]$ . Sampling compensation is achieved by adding this offset directly during warping. For the second occlusion problem, we introduce an occlusion solving strategy for occluded pixels. We additionally estimate the visibility map $\\mathbf{V}_{1\\rightarrow t}$ of each point from the initial frame to the $t$ -th frame as well as the trajectory update $\\Delta \\mathbf{T}_t$ . Aggregation is based on a warp with offset when the point $x$ is visible. When point $x$ is occluded, a learnable module Fusion is introduced to regress the point's coarse motion trajectory in $t\\to t + 1$ from $\\mathbf{M}_{1\\rightarrow t}^{global}$ . Finally, the trajectory update $\\Delta \\mathbf{T}_t$ as a uniformly refinement serves the global trajectory accumulation goal. The refinement process can be modeled as follows:", + "bbox": [ + 511, + 508, + 908, + 765 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {T} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {W a r p} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {O} _ {t}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 1,\\\\\\operatorname {F u s i o n} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {M} _ {1 \\rightarrow t} ^ {\\text {g l o b a l}}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 0,\\end{array}\\right. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 509, + 786, + 908, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{T}'$ is the local curve in $t\\to t + 1$ warped to the starting coordinates at time 1. Then we direct combine it with the cached previous global curve to obtain the accumulated current global curve $\\mathbf{T}_{1\\rightarrow t + 1}^{accum}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathbf{T}'(\\mathbf{x})\\right]$ .", + "bbox": [ + 511, + 840, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7938", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd78a4ad66fb8187085af31e2b7a244c880e9f25cb1dc399fc61216c6b96771e.jpg", + "image_caption": [ + "Figure 1. Our proposed event-aided dense and continuous point tracking framework consists of two main steps. 1) Local motion estimation: estimating short-term curve trajectories with $N_{c}$ control points from two consecutive images and inter-frame events, while concurrently updating the local motion representation. 2) Global motion accumulation: iteratively fusing the latest local motion representation with the previous global motion representation in a streaming manner for aggregating the latest global motion representation. Subsequently, the global long-term curve trajectories with $t \\times N_{c}$ control points are optimized on trajectory combinations." + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 861, + 349 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Events along the trajectory. Following the contrast maximization framework [8], we assume that events are triggered along with the pixel motion trajectories at the moving boundary. For a motion trajectory $\\mathbf{T}$ starting from pixel $\\mathbf{x}_1$ at time $t_1$ , the generated event's coordinates satisfy the trajectory, i.e., $\\mathbf{x}_1 = \\mathbf{T}(t_1), \\mathbf{x}_i - \\mathbf{x}_1 = \\mathbf{T}(t_i) - \\mathbf{T}(t_1)$ . We can thus use the motion trajectory to transform the events $e_i \\dot{=} \\{\\mathbf{x}_i, t_i, p_i\\}$ back to time $t_1$ :", + "bbox": [ + 88, + 438, + 480, + 559 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ne _ {i} ^ {\\prime} \\doteq \\left\\{\\mathbf {x} _ {i} ^ {\\prime} = \\operatorname {W a r p} \\left(\\mathbf {x} _ {i}; \\mathbf {T} (t _ {i}) - \\mathbf {T} (t _ {1})\\right), t _ {1}, p _ {i} \\right\\}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 569, + 482, + 587 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Assuming the trajectory $\\mathbf{T}$ is accurate, this process transforms the event $e_i$ to the starting point position $\\mathbf{x}_1$ of the trajectory, i.e., $\\mathbf{x}_1 = \\mathrm{Warp}(\\mathbf{x}_1; \\mathbf{T}(t_i - t_1))$ . Based on the correlated motion modeling of events and point trajectories, we build additional self-supervised training objectives in Sec. 3.3 to alleviate the lack of continuous trajectory annotations in the training datasets.", + "bbox": [ + 89, + 597, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Framework", + "text_level": 1, + "bbox": [ + 89, + 712, + 218, + 726 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Two-frame basis model. The two-frame basis model is designed to recover inter-frame short-term trajectories $\\mathbf{T}_{t\\rightarrow t + 1}$ from the encoded features of input two consecutive frames $F_{t},F_{t + 1}$ and inter-frame events $E_{t\\rightarrow t + 1}$ . In the feature extraction phase, we first convert the raw event data into a dense grid representation [41], followed by two feature encoders for two images and an event grid. Subsequently, we construct the local correlation between two frame features by matrix multiplication [44]. By leveraging the local correlation and events, we learn the local motion representation $M_{t\\rightarrow t + 1}^{local}$ by a motion extractor which allows", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "recovery of local dense trajectories $\\mathbf{T}_{t\\rightarrow t + 1}$ by a trajectory decoder. Specifically, the former utilizes a transformer fusion module in GMA [22], while the latter is based on the typical FPN decoder in RAFT [43]. The trajectory decoder estimates the coordinates of $N_{c}$ control points $\\mathbb{P}_{t\\to t + 1}$ and a single-channel visibility map $\\{\\mathbf{V}\\}_{t\\to t + 1}$ , which is essential for establishing multi-frame global accumulation in Eq. (1).", + "bbox": [ + 511, + 438, + 906, + 544 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Global motion aggregation module. For processing a video comprising $N_{v}$ frames, the above two-frame basis model needs to be streamed sequentially $N_{v} - 1$ times yielding multiple local motion representations $\\{\\mathbf{M}_{i\\rightarrow i + 1}^{local}\\}_{i = 1}^{t - 1}$ and curve trajectories $\\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1}$ . To facilitate the accumulation of global multi-frame trajectories according to Sec. 3.1, the cached global motion representation $\\mathbf{h}_{t - 1}\\doteq \\mathbf{M}_{1\\to t}^{global}$ from the previous $t$ -frames is utilized as the query, while the current local motion representation $\\mathbf{h}_t^l\\doteq \\mathbf{M}_{t\\to t + 1}^{local}$ serves as the key and value. We first perform the linear projections and compute the cross-attention:", + "bbox": [ + 511, + 566, + 908, + 734 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\operatorname {C A} \\left(\\mathbf {h} _ {t - 1}, \\mathbf {h} _ {t} ^ {l}, W _ {Q, K, V}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {Q} \\cdot \\mathbf {h} _ {t - 1}\\right) \\left(W _ {K} \\cdot \\mathbf {h} _ {t} ^ {l}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\left(W _ {V} \\cdot \\mathbf {h} _ {t} ^ {l}\\right), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 758, + 903, + 828 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $d_{k}$ is the channel size, $\\cdot$ is the linear projection and $W_{Q,K,V}$ are the corresponding weights. We then conduct iterative fusion based on the gated activation unit (GRU) [3], where the update gate is $\\mathbf{z}_t =$", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7939", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "sigmoid(CA( $\\mathbf{h}_{t-1}$ , $\\mathbf{h}_t^l$ , $W_{Q,K,V}$ ), the reset gate is $\\mathbf{r}_t =$ sigmoid(CA( $\\mathbf{h}_{t-1}$ , $\\mathbf{h}_t^l$ , $W_{Q,K,V}^\\prime$ ), and the hidden state is $\\mathbf{s}_t = \\tanh(\\mathrm{CA}(\\mathbf{r}_t \\odot \\mathbf{h}_{t-1}, \\mathbf{h}_t^l, W_{Q,K,V}^{\\prime\\prime}))$ , $\\odot$ is the elementwise multiplication. The superscript of $W_{Q,K,V}$ denotes the different projection weights taken independently in each attention calculation. Finally, we iteratively update the current global motion representation at the feature level by:", + "bbox": [ + 89, + 90, + 483, + 200 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} _ {1 \\rightarrow t + 1} ^ {\\text {g l o b a l}} \\doteq \\mathbf {h} _ {t} = \\left(1 - \\mathbf {z} _ {t}\\right) \\odot \\mathbf {h} _ {t - 1} + \\mathbf {z} _ {t} \\odot \\mathbf {s} _ {t}. \\quad (4)\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 210, + 482, + 232 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The simple and effective recurrent temporal aggregation we take is naturally compatible with the streaming pipeline, and we verify its effectiveness compared to previous fusion solutions in ablation experiments (cf. Sec. 4.4).", + "bbox": [ + 89, + 246, + 483, + 308 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Multi-frame iterative streaming framework. As depicted in Fig. 1, our framework iteratively processes the input video and event data through local motion estimation and global motion accumulation. We aggregate the local motion representations from each frame interval to the global motion representation at the feature level through the above global aggregation module in Sec. 3.2. Subsequently, the multi-frame accumulation step described in Sec. 3.1 sequentially combines each inter-frame short-term curve into a global long-term motion trajectory at the trajectory level, providing the dense and continuous point tracking representation as the model output. On the right side of Fig. 1, local motion is visualized with dense optical flow, and global motion is represented with deformations of dense point grid and curve trajectories of sparse query points.", + "bbox": [ + 89, + 334, + 483, + 561 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objective", + "text_level": 1, + "bbox": [ + 89, + 574, + 202, + 590 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Temporal discrete trajectory supervision. The available point tracking datasets provide only temporally discrete point tracks with no ground truth for continuous interframe trajectories. Following DOT [35], we first adopt supervised losses based on the temporal discrete ground-truth point tracks provided by the dataset, which consists of the L1 loss $L_{\\text{traj}}$ for sampled discrete trajectory prediction and the binary cross-entropy loss $L_{\\text{vis}}$ for visibility map. We then randomly select different frame intervals for augmented training. Local correlation is not constructed when the frames are skipped, therefore the corresponding event features are taken into streaming for iteratively updating the global motion representation. There are cases where some images are not used as input when the frame interval is greater than 1, but the corresponding input events and ground-truth tracks can be regarded as inter-frame motion contributing to curve trajectory learning. Such sampling-based augmentation ensures the model learning through diverse long- and short-term motions, capitalizing on the continuity of events to estimate continuous trajectories.", + "bbox": [ + 89, + 598, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Event consistency with continuous trajectory. Since events are usually generated along motion trajectories, we propose to leverage the continuous property of events for self-supervised continuous trajectory learning in conjunction with discrete supervision of point trajectories. However, events are computationally intensive to process one by one and are generally accompanied by noise. We thus first introduce event temporal chunking to process events in batches within a fixed duration to reduce the noise impact and computation. For the $b$ -th interval of $B$ chunks, we isolate the events within that $b$ -th chunk and aggregate them after warping them to $t_b$ as Eq. (2). For each chunk, the events then are summed into an image of warped events (IWE) [8], i.e., $\\mathbf{EB}(\\mathbf{x}_i,b)\\doteq \\sum_{i = 1}^{Ne}\\mathcal{N}(\\mathbf{x}_i;x_i',\\sigma^2)$ , where $t_b\\leq t_i < t_{b + 1}$ and $\\sigma$ is the neighboring range which is usually chosen as 1 pixel. This IWE essentially counts the number of warped events $e_i^\\prime$ per pixel and chunk. The chunking intervals are chosen randomly to exploit the continuous nature of events. Thus we can establish consistent connections between event chunks and continuous trajectories:", + "bbox": [ + 511, + 90, + 906, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {e c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {b _ {1} \\neq b _ {2}} ^ {B} \\rho \\binom {\\mathbf {E B} (\\mathbf {x}, b _ {1}),} {\\text {W a r p} \\left(\\mathbf {E B} (\\mathbf {x}, b _ {2}); \\mathbf {T} \\left(t _ {b _ {2}}\\right) - \\mathbf {T} \\left(t _ {b _ {1}}\\right)\\right)}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 519, + 409, + 906, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\rho$ is the consistency measure with L1 norm. Since events are spatially sparse, we only establish connections with the dense trajectories at locations $\\Omega$ with valid events.", + "bbox": [ + 511, + 481, + 906, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image consistency with discrete trajectory. Similar to the unsupervised optical flow task [30], we can also establish the discrete consistency of the motion trajectory with the images at discrete times, to compensate for the spatial sparsity issue of ground-truth point tracks in the training data. In addition, in our sampling-based augmented training, skipped images can be used as additional continuity training objectives.", + "bbox": [ + 511, + 559, + 905, + 679 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the accumulated continuous global trajectory $\\mathbf{T}_{1\\rightarrow t}$ we sample the the discrete optical flow $\\mathbf{F}_{i\\to j}$ from $I_{i}$ to $I_{j}$ via timestamps. Similar to Eq. (5), the consistency of images can be modeled as:", + "bbox": [ + 511, + 681, + 906, + 742 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {i c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {i \\neq j} ^ {N _ {v}} \\rho \\left(I _ {i} (\\mathbf {x}), \\operatorname {W a r p} \\left(I _ {k} (\\mathbf {x}); \\mathbf {F} _ {i \\rightarrow j} (\\mathbf {x})\\right)\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 758, + 906, + 801 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Total objective. Our total training objective is a weighted combination of the above objectives, i.e., $L = L_{\\text{traj}} + \\lambda_1 L_{vis} + \\lambda_2 L_{ec} + \\lambda_3 L_{ic}$ , $\\lambda$ are manual hyperparameters. Our ablations verify that combined self-supervised training can compensate for the temporal continuity of trajectories.", + "bbox": [ + 511, + 825, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7940", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20926011ecdde905bc7cfdce5b6eee5f09588e736a811fc2ef07a45180e95c12.jpg", + "image_caption": [ + "Input images with GT query points: frame 1, 7, 19, 27" + ], + "image_footnote": [], + "bbox": [ + 130, + 88, + 223, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2d45f50ff938260353577c4235090cb96a364dd5ef4e944a02738893e6efa3c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 236, + 88, + 331, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/27f2044ab6b3377ffd94a49d78365fcab712c97a7432fd01f61ac863c2748b2f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 88, + 437, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3315f716dc28c0314ebcaf0092f1fd62dd651191e4e9ad230ca31e7a28c2f3e9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 88, + 542, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9f5c2870b91081d75692e8bbcf34d2b15e60f3e5a64fb7a4fc4ab746c74358f2.jpg", + "image_caption": [ + "Init Coords", + "Figure 2. Visual comparisons of long-term dense point tracking on the TAP-Vid-DAVIS [5] dataset, with the ground-truth sparse query points with input images. Due to space limitations, we provide more sequences and event data visualization in the Supp. and demo video." + ], + "image_footnote": [], + "bbox": [ + 130, + 175, + 223, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e95ad5b036df431cddbe71bd95aa089fa07429b2f0e090853e734070c58597d3.jpg", + "image_caption": [ + "DOT: frame 7, 19, 27" + ], + "image_footnote": [], + "bbox": [ + 236, + 175, + 331, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3532de9b9c32b6da9664495a0b985024862babb7e711754dfa432ce73dd9cd4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 343, + 176, + 437, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f6fa006b79904cf61a6bd22771e1d03cddfb44e1418c7c280f8720f6c48d85bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 450, + 176, + 542, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c262e1b5c82bd005854405afb531dde6a1471dc41fb333acc787bb409b4f6326.jpg", + "image_caption": [ + "CoTracker: frame 7, 19, 27" + ], + "image_footnote": [], + "bbox": [ + 555, + 88, + 647, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/585f9913958a1e636dcba20093ed6804f66fbe5c65d1f129f4a034c030f6fc4b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 89, + 753, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3a733254be8ee7ec3a82c13ab25d185818bb623f85a2753ac7e70728ec36775e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 89, + 859, + 159 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/22ac09a90fd9fad4530145fdb9a3c6e9f9b7a0253fa59b1deaa87614d7db97c5.jpg", + "image_caption": [ + "Ours: frame 7, 19, 27" + ], + "image_footnote": [], + "bbox": [ + 557, + 175, + 648, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6eb6b97bc24766c478ce95dd317c02396badee7d7cfa26369d7dd45953033f3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 661, + 176, + 754, + 247 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/7c0e298d0dcdb6f1dcf88a53ec7709bbb9e24e792e5f6964308a32c5b0f159e0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 176, + 861, + 247 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/399048cff99b027d35387223de467e7e9364fa8f5645dd11399429ccd09228c7.jpg", + "table_caption": [ + "Table 1. Quantitative results of dense evaluation on the CVO test [52] and extended set [35]." + ], + "table_footnote": [], + "table_body": "
MethodCVO (Clean)CVO (Final)CVO (Extended)
EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑
QueryPIPs++ [58]9.05 / 6.62 / 21.533.39.49 / 7.06 / 22.032.718.4 / 10.0 / 32.158.7
TAPIR [6]3.80 / 1.49 / 14.773.54.19 / 1.86 / 15.372.419.8 / 4.74 / 42.568.4
CoTracker [24]1.51 / 0.88 / 4.5775.51.52 / 0.93 / 4.3875.35.20 / 3.84 / 7.7070.4
DenseGMA [22]2.42 / 1.38 / 7.1460.52.57 / 1.52 / 7.2259.721.8 / 15.7 / 32.865.6
MFT [36]2.91 / 1.39 / 9.9319.43.16 / 1.56 / 10.319.521.4 / 9.20 / 41.837.6
AccFlow [52]1.69 / 1.08 / 4.7048.11.73 / 1.15 / 4.6347.536.7 / 28.1 / 52.936.5
DOT [35]1.32 / 0.74 / 4.1280.41.38 / 0.82 / 4.1080.25.07 / 3.67 / 7.3471.0
EDCPT (Ours)1.23 / 0.71 / 3.8382.11.31 / 0.76 / 3.8681.94.88 / 3.44 / 7.4671.9
", + "bbox": [ + 173, + 311, + 818, + 467 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/562352820b0b2d107319d84ade9f7a16e050a1b4ec343d05e564c27792a0fe08.jpg", + "table_caption": [ + "Table 2. Quantitative results of sparse evaluation on the TAP-VidDAVIS [5]point tracking benchmark." + ], + "table_footnote": [], + "table_body": "
MethodDAVIS (First)DAVIS (Strided)
AJ ↑<δxavg↑OA ↑AJ ↑<δxavg↑OA ↑
QueryTAP-Net [5]33.048.678.838.453.182.3
Context-PIPs [51]42.760.379.548.964.083.4
TAPIR [6]56.270.086.561.373.688.8
CoTracker [24]61.174.689.163.579.887.8
SpatialTracker [54]61.176.389.5---
DenseCPFlow [31]9.614.6----
MFT [36]47.366.877.856.170.886.9
DecoMotion [26]53.069.984.260.274.487.2
DinoTracker [46]---62.378.287.5
FlowTrack [4]---63.276.389.2
DOT [35]61.675.589.566.780.690.4
EDCPT (Ours)63.876.390.667.580.591.1
", + "bbox": [ + 106, + 507, + 465, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 717, + 223, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental details", + "text_level": 1, + "bbox": [ + 89, + 742, + 285, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our model on MOVI-F [15] dataset and directly evaluate it on CVO [52] and TAP-Vid-DAVIS [5] benchmarks, using the vid2e [10] simulator to generate events.Following DOT [35], we train the model for 500k steps on $4\\times$ NVIDIA L40 48G GPUs, using the Adam optimizer and OneCycle learning rate decay with a maximum of $10^{-4}$ . We choose 3 frames as training samples, along with the random selection of up to 10 frames in different frame intervals. The loss hyperparameters are set to 1.0,", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "0.1, 0.1. For dense evaluation, we report the dense absolute error $\\mathrm{EPE}_{\\mathrm{all / vis / occ}}$ for all/visible/occluded points, as well as occlusion IoU accuracy OA for estimated visible mask. For sparse evaluation, we follow TAPNet [5] and report average Jaccard AJ, position accuracy $< \\delta_{\\mathrm{avg}}^x$ , and occlusion accuracy OA. Additionally, we adopt the real-captured event-based optical flow benchmark, DSEC [12, 13], to verify our adaptation capacity. More details are provided in the Supp.", + "bbox": [ + 511, + 479, + 906, + 602 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Standard spatially dense point tracking", + "text_level": 1, + "bbox": [ + 511, + 617, + 852, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance on CVO and TAP-Vid-DAVIS benchmarks. We conduct comprehensive evaluations on two common standard point tracking benchmarks. Consistent with DOT [35], we report the quantitative results of the spatially dense optical flow from the last to the first frame of CVO [52] dataset in Tab. 1. Our proposed new EDCPT framework achieves significant performance improvements, whether comparing methods that only predict the partial Query points or directly estimating spatially Dense trajectories within a single inference. Particularly, we achieve 0.19 $\\mathrm{EPE}_{\\mathrm{all}}$ and 0.9 OA improvements on the extended set of 476 videos with 48 frames when compared to DOT [35].", + "bbox": [ + 511, + 641, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, the real TAP-Vid-DAVIS dataset [5] only provides ground-truth trajectories for selected query points. As a result, we only sparsely evaluate these points for a fair comparison despite the output trajectories of our model and some compared methods are spatially dense. The", + "bbox": [ + 511, + 824, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7941", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c08b675dfba08ef858393f7b0a353e69d582a04423327d6ff1aa24e19c8627e2.jpg", + "image_caption": [ + "Figure 3. Visual comparisons of dense and continuous point trajectories on TAP-Vid-DAVIS [5]. Zoom in for detailed curve trajectories." + ], + "image_footnote": [], + "bbox": [ + 132, + 88, + 369, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/765a1a10e4c12ae8e59dde07123a44f8fb8422a4be3c7a7e2844aa5761b84e3a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 370, + 88, + 612, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/db3f7af69fd85254f5392812c74c36d4fe34112186ee69fc27f0443e98446d45.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 89, + 856, + 223 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8bd3752f65657b0a13367ea3ba74981164256cf90732d38c01eb726dcb4e87cb.jpg", + "table_caption": [ + "Table 3. Quantitative results on the real DSEC optical flow benchmark [13]. The first two methods are categorized as self-supervised learning and the remaining as supervised learning." + ], + "table_footnote": [], + "table_body": "
MethodInputEPE ↓AE ↓%Out ↓
Taming [39]Events2.3310.5617.77
MPCMax [18]Events3.208.5315.21
E-RAFT [13]Events0.792.852.68
TMA [28]Events0.742.682.30
IDNet [53]Events0.722.722.04
ECDDP [56]Events0.702.581.96
BFlow [14]Events0.752.682.44
BFlow [14]Images+Events0.692.421.88
EDCPT (Ours)Images+Events0.632.171.52
", + "bbox": [ + 114, + 291, + 455, + 435 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "quantitative results in Tab. 2 demonstrate the superiority of our framework, outperforming the existing state-of-the-art methods DOT [35] and SpatialTracker [54] with up to 2.7 AJ and 1.1 OA. We argue that ours $< \\delta_{\\mathrm{avg}}^{x}$ is close to DOT's because the sparse to dense upgrade strategy from sparse point matches is capable of modeling the rigid object motion and thus benefiting the proportion evaluation. We also perform qualitative visual comparisons in Fig. 2. Combining the above quantitative and qualitative comparisons with previous image-based methods, the new attempts of incorporating events by our framework significantly improve the accuracy with the standard dense point tracking setting.", + "bbox": [ + 88, + 449, + 483, + 630 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance on DSEC benchmark. We further conduct experiments on the DSEC benchmark [13] with real captured event data. Because the DSEC online leaderboard only measures the optical flow between two consecutive frames, we therefore finetune the local motion estimation on the DSEC training set from our pre-trained long-term model. The submission results are shown in Tab. 3. Notably, while BFlow [14] can estimate curve trajectories, they only submitted another optical flow version. Our framework fuses image and event data, yielding state-of-the-art performance with improvements of 0.05 EPE and 0.25 AE.", + "bbox": [ + 89, + 654, + 483, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Temporally continuous point tracking", + "text_level": 1, + "bbox": [ + 89, + 832, + 418, + 848 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We adapt the above standard evaluation to input only a portion of the full video frames into the model to evaluate the temporal continuity with the ground truths of skipped", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "frames. For the CVO final set with 7 frames per video, we skip 1 frame (half) and 2 frames (one-third) as model inputs, because the longer extended set lacks multi-frame ground-truth tracks. For the TAP-Vid-DAVIS dataset with $\\sim 100$ frames, we report skip 1 frame (half) and skip 3 frames (quarter). The compared image-based methods cannot model inter-frame motion, thus we take linear motion interpolation to generate trajectory when frames are skipped. We retrain AccFlow [52] and mark it with * as its public version does not support forward point tracking.", + "bbox": [ + 511, + 251, + 903, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As reported in Tab. 4, our proposed new framework with global continuous trajectory accumulation significantly outperforms existing methods. Especially in nonlinear motion scenarios of TAP-Vid-DAVIS datasets, the larger frame intervals lead to greater performance gaps. In addition to the ablation of different motion assumptions in Tab. 6, the B-spline representation we adopt achieves better performance. We also provide visual comparisons in Fig. 3 and a demo video of continuous trajectory visualization in the Supp., that includes four sequences with diverse motion complexity and event sparsity from TAP-Vid-DAVIS and real-captured ERF-X170FPS [25] dataset. With extensive experiments on both simulated events (CVO and TAP-Vid-DAVIS) and real events (DSEC and ERF-X170FPS), we fully validate the robustness of the proposed global motion accumulation in modeling complex continuous trajectories.", + "bbox": [ + 511, + 402, + 905, + 645 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Ablation experiments and discussions", + "text_level": 1, + "bbox": [ + 511, + 657, + 836, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To perform progressive ablations in Tab. 5, Tab. 6, and Tab. 7, the underlined settings are those utilized in the previous table, and the bolded ones represent the choices for our final framework. To validate the capability for continuous point tracking, the ablation results are reported with CVO third and DAVIS quarter settings as in Sec. 4.3 and Tab. 4.", + "bbox": [ + 511, + 680, + 905, + 771 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Global motion aggregation. One of our key contributions is the global aggregation of local motion representations in the streaming pipeline. In Tab. 5, N/A indicates do not explicitly model sequential motion as in DOT [35]. post is the forward aggregation in AccFlow [52], which performs post-processing only on the output optical flow and ignores the internal long-term motion information. We aggregate", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7942", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4f57aa91c8775197a33e55c85e5b0abf26aaa0aeaf1a11533f8596fe8b06fb54.jpg", + "table_caption": [ + "Table 4. Continuous point tracking evaluation results on the CVO extended set [35] and TAP-Vid-DAVIS dataset [5]." + ], + "table_footnote": [], + "table_body": "
MethodCVO (Final) - EPEall/vis/occ ↓DAVIS (First) - AJ / <δxavg↑
fullhalfthirdfullhalfquarter
RAFT2.09 / 0.81 / 8.022.44 / 0.95 / 9.073.02 / 1.16 / 11.4233.9 / 46.628.6 / 40.122.4 / 34.2
GMA1.99 / 0.77 / 7.572.35 / 0.89 / 8.452.92 / 1.09 / 10.9739.3 / 52.531.7 / 44.326.5 / 38.3
AccFlow*2.28 / 0.60 / 11.182.39 / 0.80 / 10.742.79 / 1.02 / 11.3747.2 / 62.337.5 / 49.230.9 / 42.4
CoTracker1.89 / 0.63 / 7.052.11 / 0.82 / 8.023.17 / 1.65 / 11.1361.1 / 74.654.3 / 68.848.9 / 63.9
DOT1.83 / 0.59 / 6.952.10 / 0.73 / 7.882.69 / 0.97 / 10.8461.6 / 75.555.6 / 70.150.4 / 65.3
EDCPT (Ours)1.76 / 0.55 / 6.731.97 / 0.66 / 7.612.16 / 0.73 / 8.7663.8 / 76.359.7 / 73.156.2 / 70.9
", + "bbox": [ + 153, + 104, + 841, + 215 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/0c74a03fcba7e50527e63c410fa7a3424b13206a4f1c6f4cfec7d6b559ba4a24.jpg", + "table_caption": [ + "Table 5. Ablations on global motion aggre- Table 6. Ablations on parametric curve rep- Table 7. Ablations on input data and supervi-gation. resentation. sion." + ], + "table_footnote": [], + "table_body": "
EPEall/vis/occ ↓AJ / <δxavg↑
N/A2.54 / 0.89 / 10.3351.9 / 66.4
post2.49 / 0.85 / 9.7352.5 / 66.9
solo2.45 / 0.82 / 9.8952.7 / 67.0
–offsets2.43 / 0.82 / 9.7353.0 / 67.2
stream2.42 / 0.80 / 9.6853.3 / 67.4
", + "bbox": [ + 98, + 251, + 359, + 347 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/f960c4a1739216187b9161ddb88165d8f9625561acb4800e259156de1099fb2b.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EPEall/vis/occ ↓AJ / <δxavg↑
linear2.42 / 0.80 / 9.6853.3 / 67.4
quad2.49 / 0.84 / 9.4654.2 / 68.1
Nc=32.32 / 0.79 / 9.3354.4 / 68.6
Nc=42.23 / 0.76 / 8.9755.4 / 69.7
Nc=52.26 / 0.75 / 9.2055.0 / 69.3
", + "bbox": [ + 364, + 251, + 625, + 349 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/20d0e93eaf1cbb56f7d6f113f9f5e2e7a3568bd4059280131c909d7a97f597f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
EPEall/vis/occ ↓AJ / <δxavg ↑
Images2.50 / 0.86 / 9.9252.5 / 66.7
+Events2.23 / 0.76 / 8.9755.4 / 69.7
+Lic2.20 / 0.75 / 8.9055.8 / 70.2
+Lec2.16 / 0.73 / 8.7656.2 / 70.9
", + "bbox": [ + 630, + 251, + 895, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "motion representations at the feature level instead of dealing directly with motion vectors. solo is the short and long term fusion module in SOLOFusion [40]. Unlike its temporal fusion with a fixed number of frames and low resolution, we adopt sequential modeling with an unspecified number of frames in a streaming pipeline. Our proposed motion aggregation framework fuses image correspondence and event features from local to global with recurrent streaming structure, which better preserves multi-frame motion information and achieves optimal performance. In addition, we also verified that removing the additional offset estimation to address the numerical problem in Warping leads to a slight performance degradation, as this would require the subsequent refinement to handle it simultaneously.", + "bbox": [ + 88, + 364, + 482, + 575 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Curve representation. Based on our streaming framework, we compare the performance improvement of taking the B-spline curve representation compared to interpolating with linear and quadratic motion assumptions in Tab. 6. For the control points number of B-spline curves, we chose $N_{c} = 4$ , since further increasing the number of control points does not significantly improve the performance under CVO third and DAVIS quarter settings.", + "bbox": [ + 89, + 594, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Input data and supervision. Since previous methods usually use only image data, we verify the advantages of incorporating event data for high-precision continuous point tracking by removing events, as depicted in Tab. 7. Moreover, the comparison results between this image-only variant and DOT [35] in Tab. 4 demonstrates that our proposed streaming aggregation and curve representation are beneficial even in the absence of event data. Furthermore, we validate that training using the proposed image and event-to-point trajectory consistencies as additional supervision complements the lack of continuous inter-frame tracks in", + "bbox": [ + 89, + 734, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the training data and can further improve performance.", + "bbox": [ + 511, + 364, + 874, + 378 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. On the same RTX 3090 PC, our framework processes a 48-frame video at $512 \\times 512$ resolution in 12.6 seconds, significantly faster than CoTracker's 11 minutes, while only processing partial query points in a single run. However, it is slower than the two-frame optical flow method GMA of only 2.1 seconds and slightly slower than the multi-frame point tracking method DOT of 9.5 seconds. In addition, there is currently a lack of point-tracking datasets with real events, and it is difficult to obtain long-term tracking annotations. We evaluate on standard benchmarks with simulated events, as well as visualize point tracking on real events, and also evaluate optical flow estimation on a real event benchmark. Our future work plans to improve on model efficiency and the evaluation setup.", + "bbox": [ + 511, + 402, + 906, + 616 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 632, + 633, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a new long-term dense point tracking framework for integrating image and event data to estimate continuous motion trajectories. Specifically, we streamingly process two adjacent frames and inter-frame events to generate local motion representations, and accumulate previous caches to global motion representation at the feature level to produce new global trajectories at the trajectory level. We utilize multi-frame parametric curve accumulation to represent continuous trajectories with any number of frames, complemented by image and event-to-trajectory consistencies for model training. We validate the effectiveness of our framework by conducting extensive benchmark experiments as well as qualitative verification on real challenge data. We believe this work provides new insights into the point tracking task from the perspective of event modality and continuous global curve representations.", + "bbox": [ + 511, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7943", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research/project is supported by the National Natural Science Foundation of China (62271410, 12150007), the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008), and the Tier 1 grant T1-251RES2305 from the Singapore Ministry of Education. Zhexiong Wan was also supported by the Program of China Scholarship Council (202306290193), and the Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University (CX2023013).", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 280, + 187, + 296 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scalespace flow for end-to-end optimized video compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1", + "[2] Adam Bielski and Paolo Favaro. Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems (NeurIPS), 35:33371-33386, 2022. 1", + "[3] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 4", + "[4] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268–19277, 2024. 1, 2, 3, 6", + "[5] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re-casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, pages 13610–13626, 2022. 1, 2, 3, 6, 7, 8", + "[6] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10061-10072, 2023. 2, 6", + "[7] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. In Asian Conference on Computer Vision (ACCV), 2024. 2", + "[8] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 4, 5", + "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and" + ], + "bbox": [ + 99, + 305, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Davide Scaramuzza. Event-based vision: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):154-180, 2022. 1", + "[10] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3586-3595, 2020. 6", + "[11] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision (IJCV), 128(3):601-618, 2020. 2", + "[12] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters (RA-L), 2021. 6", + "[13] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 2, 6, 7", + "[14] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 46(7):4736-4746, 2024. 1, 3, 7", + "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3749-3761, 2022. 6", + "[16] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16022-16033, 2023. 1", + "[17] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 7167-7179, 2021. 1", + "[18] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In European Conference on Computer Vision (ECCV), 2024. 3, 7", + "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022. 1, 2", + "[20] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. Progressive spatio-temporal alignment for efficient event-based motion estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1537-1546, 2023. 1", + "[21] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. FlowNet 2.0: Evolution of optical flow estimation with deep networks." + ], + "bbox": [ + 516, + 92, + 908, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7944", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 1", + "[22] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9772-9781, 2021. 4, 6", + "[23] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2023. 1", + "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision (ECCV), 2024. 2, 6", + "[25] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 7", + "[26] Rui Li and Dong Liu. Decomposition betters tracking everything everywhere. In European Conference on Computer Vision (ECCV), 2024. 2, 6", + "[27] Siqi Li, Zhikuan Zhou, Zhou Xue, Yipeng Li, Shaoyi Du, and Yue Gao. 3d feature tracking via event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18974-18983, 2024. 2", + "[28] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9685-9694, 2023. 7", + "[29] Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate. arXiv preprint arXiv:2409.11953, 2024. 3", + "[30] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6489-6498, 2020. 5", + "[31] Jianqin Luo, Zhexiong Wan, Yuxin Mao, Bo Li, and Yuchao Dai. Continuous parametric optical flow. In Advances in Neural Information Processing Systems (NeurIPS), pages 23520-23532, 2023. 1, 2, 3, 6", + "[32] Xinglong Luo, Ao Luo, Zhengning Wang, Chunyu Lin, Bing Zeng, and Shuaicheng Liu. Efficient meshflow and optical flow estimation from event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19198-19207, 2024. 2", + "[33] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5642-5651, 2023. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Etienne Meunier, Anais Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4):4462-4473, 2023. 1", + "[35] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 8", + "[36] Michal Neoral, Jonás Šerzych, and Jérí Matas. Mft: Long-term tracking of every pixel. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 6837-6847, 2024. 2, 3, 6", + "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (TMLR), 2024. 2", + "[38] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1669-1678, 2020. 1", + "[39] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9695-9705, 2023. 7", + "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris M. Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multiview 3d object detection. In International Conference on Learning Representations (ICLR), 2023. 8", + "[41] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 4", + "[42] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision (ECCV), 2024. 2", + "[43] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. CRAFT: Cross-attentional flow transformer for robust optical flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17602-17611, 2022. 4", + "[44] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419, 2020. 1, 4", + "[45] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time Lens: Event-based video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, 2021. 1" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7945", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 6", + "[47] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing (TIP), 31:7237-7251, 2022. 2", + "[48] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. RPEFlow: Multimodal fusion of RGB-pointcloud-event for joint optical flow and scene flow estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10030-10040, 2023. 2", + "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19795-19806, 2023. 2", + "[50] Xiangyuan Wang, Huai Yu, Lei Yu, Wen Yang, and GuiSong Xia. Towards robust keypoint detection and tracking: A fusion approach with event-aligned image features. IEEE Robotics and Automation Letters (RA-L), 9(9):8059-8066, 2024. 2", + "[51] BIAN Weikang, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands spatial context features. In Advances in Neural Information Processing Systems (NeurIPS), pages 55285-55298, 2023. 2, 6", + "[52] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: backward accumulation for long-range optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12119-12128, 2023. 2, 3, 6, 7", + "[53] Yilun Wu, Federico Paredes-Vallés, and Guido CHE De Croon. Lightweight event-based optical flow estimation via iterative deblurring. In IEEE International Conference on Robotics and Automation (ICRA), pages 14708-14715. IEEE, 2024. 7", + "[54] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7", + "[55] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In Advances in Neural Information Processing Systems (NeurIPS), pages 1647-1656, 2019. 1", + "[56] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data dense pre-training. In European Conference on Computer Vision (ECCV), pages 292-310. Springer, 2024. 7", + "[57] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9781-9790, 2023. 1", + "[58] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthetic dataset for long-term point tracking. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19855-19865, 2023. 2, 6", + "[59] Hanyu Zhou, Yi Chang, Haoyue Liu, Wending Yan, Yuxing Duan, Zhiwei Shi, and Luxin Yan. Exploring the common appearance-boundary adaptation for nighttime optical flow. In International Conference on Learning Representations (ICLR), 2024. 2", + "[60] Hanyu Zhou, Yi Chang, and Zhiwei Shi. Bring event into rgb and lidar: Hierarchical visual-motion fusion for scene flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26477-26486, 2024. 2", + "[61] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 34(8):4868-4880, 2021. 1", + "[62] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 1, 2", + "[63] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for RGB-event transformer-trackers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 22045-22055, 2023. 1" + ], + "bbox": [ + 516, + 92, + 903, + 455 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "7946", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_model.json b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_model.json new file mode 100644 index 0000000000000000000000000000000000000000..c85ac064839693c1c3ad32614104d0dad5593a97 --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_model.json @@ -0,0 +1,2334 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.108, + 0.131, + 0.892, + 0.153 + ], + "angle": 0, + "content": "Event-aided Dense and Continuous Point Tracking: Everywhere and Anytime" + }, + { + "type": "text", + "bbox": [ + 0.236, + 0.18, + 0.76, + 0.199 + ], + "angle": 0, + "content": "Zhexiong Wan\\(^{1,2}\\) Jianqin Luo\\(^{1}\\) Yuchao Dai\\(^{1}\\) Gim Hee Lee\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.199, + 0.821, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)School of Electronics and Information, Northwestern Polytechnical University & Shaanxi Key Laboratory of Information Acquisition and Processing" + }, + { + "type": "text", + "bbox": [ + 0.228, + 0.234, + 0.77, + 0.253 + ], + "angle": 0, + "content": "\\(^{2}\\)Department of Computer Science, National University of Singapore" + }, + { + "type": "text", + "bbox": [ + 0.108, + 0.255, + 0.885, + 0.27 + ], + "angle": 0, + "content": "{wanzhexiong,luojianqin}@mail.nwpu.edu.cn，daiyuchao@nwpu.edu.cn，gimhee.lee@nus.edu.sg" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.305, + 0.327, + 0.32 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.337, + 0.486, + 0.685 + ], + "angle": 0, + "content": "Recent point tracking methods have made great strides in recovering the trajectories of any point (especially key points) in long video sequences associated with large motions. However, the spatial and temporal granularities of point trajectories remain constrained by limited motion estimation accuracy and video frame rate. Leveraging the high temporal resolution and motion sensitivity of event cameras, we introduce event data for the first time to recover spatially dense and temporally continuous trajectories of every point at any time. Specifically, we define the dense and continuous point trajectory representation as estimating multiple control points of curves for each pixel and model the movement of sparse events triggered along continuous point trajectories. Building on this, we propose a novel multi-frame iterative streaming framework that first estimates local inter-frame motion representations from two consecutive frames with inter-frame events, then aggregates them into a global long-term motion representation to utilize input full video and event data with an arbitrary number of frames. Extensive experiments on simulated and real data demonstrate the significant improvement of our framework over state-of-the-art methods and the crucial role of introducing events to model continuous point trajectories." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.714, + 0.222, + 0.73 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.74, + 0.485, + 0.877 + ], + "angle": 0, + "content": "Estimating fine-grained motion from input videos is a crucial task in computer vision with widespread applications such as video compression [1], frame interpolation [23, 55], motion segmentation [2, 34], and dynamic scene reconstruction [16]. Early studies [21, 44] based on two-frame optical flow can model the spatially dense motion, but they suffer from the challenge of modeling long-term motion from more frames. With the proposal of tracking any point (TAP) task [5, 19], using sparse points as query indexes to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.908, + 0.502 + ], + "angle": 0, + "content": "estimate pointwise long-term motions also draws attention. Despite significant progress, the sparse and independent representation of pointwise trajectories remains inherently incompatible with the spatially dense representation of input video. Thus, recently there are methods [4, 35, 46] that continue to use dense optical flow to represent long-term motion. However, the temporal frame rate of the input video is constrained by traditional shutter cameras, which makes accurately modeling complex motion trajectories difficult for video-only methods from the data acquisition perspective. Consequently, accurate estimation of fine-grained spatially dense and temporally continuous motion remains a challenging and worthwhile research problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.909, + 0.777 + ], + "angle": 0, + "content": "Unlike traditional shutter cameras that expose the entire image at fixed frame rates, the new bio-inspired event cameras [9] can independently and asynchronously detect per-pixel brightness changes with microsecond precision. This unique design makes it inherently sensitive to visual motion changes, leading to impressive motion-related applications such as optical flow estimation [17, 62], motion segmentation [20, 61], feature tracking [33], object tracking [63], and frame interpolation [45]. However, events are typically triggered only with motion contours or rich textures, making it challenging to comprehensively perceive spatially dense motion. Recently, integrating the advantages of event camera and traditional shutter camera has become a new direction [38, 57]. Leveraging on these successes, we propose for the first time taking event data as an auxiliary input for the point tracking task, where we comprehensively model fine-grained spatially dense and temporally continuous motion from the input image and event modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.903 + ], + "angle": 0, + "content": "The introduction of event cameras offers the potential to model continuous motion from the data perspective. However, to effectively model continuous long-term motion, we need a new representation that replaces optical flow and parametrically associates the dense temporal properties of event data. BFlow [14] proposes to learn trajectories with Bezier curves from events, but is limited to input fixed frames and cannot adapt to longer sequences. CPFlow [31]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.888, + 0.428, + 0.901 + ], + "angle": 0, + "content": "Project page: https://npuctr.github.io/EDCPT" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7936" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.318 + ], + "angle": 0, + "content": "proposes to learn control points represented as B-spline curves from a fixed number of image slices. Although continuous motion can be successfully modeled using these curve representations in normalized timescales, their fixed number of control points made it hard to handle complex dynamics and varying lengths of video sequences. Based on the curve representation, we propose a new streaming pipeline for accumulating multiple local curves to address these limitations. In addition, how to adequately fuse multi-frame temporal information to achieve long-term tracking is also a concern. Compared to not explicitly associating multi-frames in DOT [35] and aggregating over estimated motion vectors in AccFlow [52] and FlowTrack [4], we perform recurrent fusion in the feature space, dedicated to preserving multi-frame sequence information." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.32, + 0.485, + 0.591 + ], + "angle": 0, + "content": "In this paper, we present the first event-aided point tracking framework for recovering spatially dense and temporally continuous point trajectories from input videos and event sequences. Specifically, we first propose a new point trajectory representation with parametric curves that accumulate multiple local curves to adapt to multi-frame input videos at any length. We then design a new framework for combining two frames with events to estimate dense point curve trajectories and extend to multi-frame iterative accumulation. In addition, we establish the association between continuous curve trajectory and event triggering as a part of the learning objective for continuous motion modeling. Extensive experiments on both simulated and real-world data demonstrate that the proposed framework significantly outperforms state-of-the-art methods. Particularly, our ablation studies illustrate the effectiveness of the proposed global aggregation and highlight the crucial role of incorporating event data in continuous trajectory modeling." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.593, + 0.45, + 0.607 + ], + "angle": 0, + "content": "Our main contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.609, + 0.483, + 0.668 + ], + "angle": 0, + "content": "- We introduce a new setup that, for the first time, enables long-term spatially dense and temporally continuous point tracking by integrating the strengths of both images and events." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.669, + 0.483, + 0.729 + ], + "angle": 0, + "content": "- We present a novel global curve representation of continuous point trajectories through multi-frame aggregation, establishing a connection between event triggering and continuous motion." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.73, + 0.484, + 0.805 + ], + "angle": 0, + "content": "- We propose a novel event-aided streaming framework that iteratively accumulates the local tracks from two frames with inter-frame events, resulting in global, long-term dense and continuous trajectories through iterative temporal global motion aggregation." + }, + { + "type": "list", + "bbox": [ + 0.11, + 0.609, + 0.484, + 0.805 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.822, + 0.241, + 0.837 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.388, + 0.865 + ], + "angle": 0, + "content": "2.1. Image-based query point tracking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "The goal of point tracking is to recover the corresponding positions of query points in each frame, which has attracted" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.35 + ], + "angle": 0, + "content": "wide attention with the proposal of the TAP benchmark and the baseline model TAPNet [5]. PIPs [19] proposes to extract independent point representation for 8-frame tracking, then \\(\\mathrm{PIPs}++\\) [58] extends to long-term trajectories. TAPIR [6] proposes a two-stage matching framework that fuses TAPNet and PIPs, and BootsTAPIR [7] performs self-supervised training on more data. Unlike these methods that track only one query at a time, CoTracker [24] and Context-PIPs [51] use additional tracks and pixel featuresto improve global tracking. SpatialTracker [54] introduces the triplane representation for group pixels in 3D.DINOTracker [46] performs test-time finetuning on the pre-trained DINOViT [37] model and achieves fine-grained tracking per video. When applying these query points-based methods to achieve dense tracking, full points need to be processed individually or in batches, which brings computational hurdles and limits their downstream applications [35]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.361, + 0.81, + 0.377 + ], + "angle": 0, + "content": "2.2. Image-based dense point tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Recent studies turn to tracking every point within a frame in a single run, aiming to enhance neighborhood relationships while reducing computational requirements. OmniMotion [49] performs pixel-wise tracking between local and canonical space to maintain the global consistency of the motion, and then FastOmniTrack [42] and DecoMotion [26] improve from the perspectives of computation cost and object motion decomposition. CPFlow [31] proposes to estimate spatio-temporally dense motion curves, but it can only input 4 images and needs pre-sampling for longer video. AccFlow [52] proposes the forward and backward aggregation pipeline, extending inter-frame dense optical flow to multi-frame long ranges. MFT [36] select chaining multiframe candidates and FlowTrack [4] automatically apply error compensation in instances of tracking inaccuracies. DOT [35] unifies point tracking and optical flow, upgrading sparse tracks to dense flow fields between every frame. However, limited by the video frame rate, these methods are struggling to model challenging dynamics. In this work, we propose to introduce continuous event data into the input video to enable temporally continuous point tracking." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.787, + 0.727 + ], + "angle": 0, + "content": "2.3. Event-based motion estimation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Thanks to the motion-sensitive nature of event cameras, extensive motion estimation studies in recent years have highlighted their potential applications to challenging dynamics. The feature tracking methods [11, 27, 33, 50] show the benefits of event cameras for low-latency tracking, but can only track sparse, specific textured locations. Recently estimating optical flow from events has become mainstream. Using only sparse event data [13, 32, 62] allows to estimate satisfactory dense optical flow, while introducing data from other sensors such as images [47, 59] and point clouds [48, 60] achieves significant performance" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7937" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "gains. BFlow [14] and MotionPriorCMax [18] exploit the continuous property of events to estimate parametric Bezier trajectories, but can only estimate motion within a fixed consecutive frame interval and cannot be directly adapted to long-term sequences. Recently, FE-TAP [29] proposes to recover point trajectories from a fixed number of images and events based on TAPVid [5], but does not take full advantage of the continuous nature of events. We propose to combine the advantages of images and events to enable temporally continuous point tracking by modeling long-term global motion with an arbitrary number of frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.281, + 0.182, + 0.297 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.484, + 0.446 + ], + "angle": 0, + "content": "Overview. To address the limitations of sparse events and low video frame rates in modeling fine-grained motion, we present the first framework that recovers dense and continuous point trajectories from a video with corresponding event sequences. Our framework consists of four parts: 1) A parametric multi-frame continuous point trajectories representation; 2) A model of event triggering along the point trajectories; 3) A two-frame basis motion estimation model; 4) A multi-frame motion aggregation and streaming framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.6 + ], + "angle": 0, + "content": "Problem Formulation. Conventional shutter camera captures a video with \\( N_{v} \\) frames \\( \\{I_i\\}_{i=1}^{N_v} \\) at a fixed frame rate. Event camera triggers an unbounded event sequence \\( \\{e_i\\}_{i=1}^{N_e} \\), where \\( N_e \\) is the number of events. Each event \\( e_i = \\{\\mathbf{x}_i, t_i, p_i\\} \\) consists of the pixel position \\( \\mathbf{x}_i = (x,y) \\), timestamp \\( t_i \\) with microsecond precision, and the brightness change polarity \\( p_i \\) in log domain. Our goal is to combine these two modalities to recover the spatially dense and temporally continuous trajectories \\( \\mathbf{T}_{1 \\to t} \\) of all points starting from the first frame at time 1 to last frame at time \\( t \\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.618, + 0.236, + 0.633 + ], + "angle": 0, + "content": "3.1. Motion model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.643, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Trajectory representation. Previous point tracking methods typically estimate two-channel motion vectors in the XY directions, which is the optical flow when representing dense trajectories [4, 35]. To learn the curve trajectory from the deep network, we instead learn the multiple control points as curved trajectories [31]. Specifically, we choose the B-spline curve as our curve representation, which is defined by \\( N_{c} \\) control points \\( \\{\\mathbf{P}_i\\}_{i=1}^{N_c} \\) and basis functions \\( \\{B_{i,p}(t)\\}_{i=1}^{N_c} \\) with degree \\( p \\). The continuous trajectory \\( \\mathbf{T}(t) \\) represented by b-spline curve in time variable \\( t \\) is a collection of piecewise polynomial functions \\( \\mathbf{T}(t) = \\sum_{i=1}^{N_c} B_{i,p}(t) \\mathbf{P}_i \\). Similar to learning dense flow, each pixel has an independent curve with control points denoted as \\( \\mathbb{P} \\in \\mathbb{R}^{N_c \\times 2 \\times H \\times W} \\), where \\( H \\times W \\) is the image size. This enables learnable motion modeling of dense and continuous point trajectories \\( \\mathbf{T} \\) with parametric curve representation. More details are provided in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Multi-frame global trajectories accumulation. Existing parametric motion modeling methods are fixed in the number of frames they can handle, e.g., BFlow [14] is limited to between two frames, and CPFlow [31] struggles to benefit from more than 4 frame inputs, resulting in suboptimal long-term trajectory modeling. Inspired by the practice of multi-frame optical flow accumulation [36, 52], we propose a new multi-frame curve trajectories accumulation strategy to handle long-term videos with arbitrary frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.233, + 0.907, + 0.504 + ], + "angle": 0, + "content": "Our accumulation framework works on a streaming pipeline, where the previous global trajectory \\(\\mathbf{T}_{1\\rightarrow t}\\) with \\((t - 1)\\times N_{c}\\) control points has been accumulated from the previous \\(t - 1\\) local trajectories \\(\\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1}\\) when processing the \\(t\\) -th step. When we get updated \\(t\\) -th local trajectory \\(\\mathbf{T}_{t\\rightarrow t + 1}\\) with \\(N_{c}\\) control points from time \\(t\\) to \\(t + 1\\), a simple approach is to directly accumulate the new global trajectory with \\(t\\times N_{c}\\) control points from time 1 to \\(t + 1\\) by \\(\\mathbf{T}_{1\\rightarrow t + 1}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathrm{Warp}(\\mathbf{T}_{t\\rightarrow t + 1},\\mathbf{T}_{1\\rightarrow t})(\\mathbf{x})\\right]\\), where Warp represents backward warping, the operator [,] combines the control points of two sub-curves and creates a more complex curve. However, there are two problems with just warping operation: 1) It suffers from numerical error as integer sampling with floating-point coordinates, i.e., for warping vectors from \\(\\mathbf{b}\\) to \\(\\mathbf{a}\\), \\(\\mathrm{Warp}(\\mathbf{a},\\mathbf{b})(\\mathbf{x}) = \\mathbf{a}(\\mathbf{x} + \\mathbf{b}(\\mathbf{x}))\\), \\(\\mathbf{x}\\) is integer coordinates but not \\(\\mathbf{x} + \\mathbf{b}(\\mathbf{x})\\). 2) Some points may be occluded at time \\(t\\), resulting in failing to find the corresponding points. More details are provided in the Supp." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.51, + 0.909, + 0.766 + ], + "angle": 0, + "content": "Our framework iteratively maintains and learns to integrate from a global motion representation \\(\\mathbf{M}_{1\\rightarrow t}^{global}\\) in the streaming process (cf. Sec. 3.2). For the first numerical problem, we estimate a start point offsets \\(\\mathbf{O}_t\\in \\mathbb{R}^{2\\times H\\times W}\\) learned from \\(\\mathbf{M}_{1\\rightarrow t}^{global}\\) and normalized to the range \\([-1,1]\\). Sampling compensation is achieved by adding this offset directly during warping. For the second occlusion problem, we introduce an occlusion solving strategy for occluded pixels. We additionally estimate the visibility map \\(\\mathbf{V}_{1\\rightarrow t}\\) of each point from the initial frame to the \\(t\\)-th frame as well as the trajectory update \\(\\Delta \\mathbf{T}_t\\). Aggregation is based on a warp with offset when the point \\(x\\) is visible. When point \\(x\\) is occluded, a learnable module Fusion is introduced to regress the point's coarse motion trajectory in \\(t\\to t + 1\\) from \\(\\mathbf{M}_{1\\rightarrow t}^{global}\\). Finally, the trajectory update \\(\\Delta \\mathbf{T}_t\\) as a uniformly refinement serves the global trajectory accumulation goal. The refinement process can be modeled as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.51, + 0.787, + 0.909, + 0.841 + ], + "angle": 0, + "content": "\\[\n\\mathbf {T} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {W a r p} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {O} _ {t}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 1,\\\\\\operatorname {F u s i o n} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {M} _ {1 \\rightarrow t} ^ {\\text {g l o b a l}}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 0,\\end{array}\\right. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.908, + 0.903 + ], + "angle": 0, + "content": "where \\(\\mathbf{T}'\\) is the local curve in \\(t\\to t + 1\\) warped to the starting coordinates at time 1. Then we direct combine it with the cached previous global curve to obtain the accumulated current global curve \\(\\mathbf{T}_{1\\rightarrow t + 1}^{accum}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathbf{T}'(\\mathbf{x})\\right]\\)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7938" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.862, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.353, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 1. Our proposed event-aided dense and continuous point tracking framework consists of two main steps. 1) Local motion estimation: estimating short-term curve trajectories with \\( N_{c} \\) control points from two consecutive images and inter-frame events, while concurrently updating the local motion representation. 2) Global motion accumulation: iteratively fusing the latest local motion representation with the previous global motion representation in a streaming manner for aggregating the latest global motion representation. Subsequently, the global long-term curve trajectories with \\( t \\times N_{c} \\) control points are optimized on trajectory combinations." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.439, + 0.482, + 0.56 + ], + "angle": 0, + "content": "Events along the trajectory. Following the contrast maximization framework [8], we assume that events are triggered along with the pixel motion trajectories at the moving boundary. For a motion trajectory \\(\\mathbf{T}\\) starting from pixel \\(\\mathbf{x}_1\\) at time \\(t_1\\), the generated event's coordinates satisfy the trajectory, i.e., \\(\\mathbf{x}_1 = \\mathbf{T}(t_1), \\mathbf{x}_i - \\mathbf{x}_1 = \\mathbf{T}(t_i) - \\mathbf{T}(t_1)\\). We can thus use the motion trajectory to transform the events \\(e_i \\dot{=} \\{\\mathbf{x}_i, t_i, p_i\\}\\) back to time \\(t_1\\):" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.57, + 0.483, + 0.588 + ], + "angle": 0, + "content": "\\[\ne _ {i} ^ {\\prime} \\doteq \\left\\{\\mathbf {x} _ {i} ^ {\\prime} = \\operatorname {W a r p} \\left(\\mathbf {x} _ {i}; \\mathbf {T} (t _ {i}) - \\mathbf {T} (t _ {1})\\right), t _ {1}, p _ {i} \\right\\}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.705 + ], + "angle": 0, + "content": "Assuming the trajectory \\(\\mathbf{T}\\) is accurate, this process transforms the event \\(e_i\\) to the starting point position \\(\\mathbf{x}_1\\) of the trajectory, i.e., \\(\\mathbf{x}_1 = \\mathrm{Warp}(\\mathbf{x}_1; \\mathbf{T}(t_i - t_1))\\). Based on the correlated motion modeling of events and point trajectories, we build additional self-supervised training objectives in Sec. 3.3 to alleviate the lack of continuous trajectory annotations in the training datasets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.22, + 0.727 + ], + "angle": 0, + "content": "3.2. Framework" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Two-frame basis model. The two-frame basis model is designed to recover inter-frame short-term trajectories \\(\\mathbf{T}_{t\\rightarrow t + 1}\\) from the encoded features of input two consecutive frames \\(F_{t},F_{t + 1}\\) and inter-frame events \\(E_{t\\rightarrow t + 1}\\). In the feature extraction phase, we first convert the raw event data into a dense grid representation [41], followed by two feature encoders for two images and an event grid. Subsequently, we construct the local correlation between two frame features by matrix multiplication [44]. By leveraging the local correlation and events, we learn the local motion representation \\(M_{t\\rightarrow t + 1}^{local}\\) by a motion extractor which allows" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.439, + 0.907, + 0.545 + ], + "angle": 0, + "content": "recovery of local dense trajectories \\(\\mathbf{T}_{t\\rightarrow t + 1}\\) by a trajectory decoder. Specifically, the former utilizes a transformer fusion module in GMA [22], while the latter is based on the typical FPN decoder in RAFT [43]. The trajectory decoder estimates the coordinates of \\(N_{c}\\) control points \\(\\mathbb{P}_{t\\to t + 1}\\) and a single-channel visibility map \\(\\{\\mathbf{V}\\}_{t\\to t + 1}\\), which is essential for establishing multi-frame global accumulation in Eq. (1)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.567, + 0.909, + 0.735 + ], + "angle": 0, + "content": "Global motion aggregation module. For processing a video comprising \\( N_{v} \\) frames, the above two-frame basis model needs to be streamed sequentially \\( N_{v} - 1 \\) times yielding multiple local motion representations \\( \\{\\mathbf{M}_{i\\rightarrow i + 1}^{local}\\}_{i = 1}^{t - 1} \\) and curve trajectories \\( \\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1} \\). To facilitate the accumulation of global multi-frame trajectories according to Sec. 3.1, the cached global motion representation \\( \\mathbf{h}_{t - 1}\\doteq \\mathbf{M}_{1\\to t}^{global} \\) from the previous \\( t \\)-frames is utilized as the query, while the current local motion representation \\( \\mathbf{h}_t^l\\doteq \\mathbf{M}_{t\\to t + 1}^{local} \\) serves as the key and value. We first perform the linear projections and compute the cross-attention:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.759, + 0.905, + 0.829 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\operatorname {C A} \\left(\\mathbf {h} _ {t - 1}, \\mathbf {h} _ {t} ^ {l}, W _ {Q, K, V}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {Q} \\cdot \\mathbf {h} _ {t - 1}\\right) \\left(W _ {K} \\cdot \\mathbf {h} _ {t} ^ {l}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\left(W _ {V} \\cdot \\mathbf {h} _ {t} ^ {l}\\right), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(d_{k}\\) is the channel size, \\(\\cdot\\) is the linear projection and \\(W_{Q,K,V}\\) are the corresponding weights. We then conduct iterative fusion based on the gated activation unit (GRU) [3], where the update gate is \\(\\mathbf{z}_t =\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7939" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.201 + ], + "angle": 0, + "content": "sigmoid(CA(\\(\\mathbf{h}_{t-1}\\), \\(\\mathbf{h}_t^l\\), \\(W_{Q,K,V}\\)), the reset gate is \\(\\mathbf{r}_t =\\) sigmoid(CA(\\(\\mathbf{h}_{t-1}\\), \\(\\mathbf{h}_t^l\\), \\(W_{Q,K,V}^\\prime\\)), and the hidden state is \\(\\mathbf{s}_t = \\tanh(\\mathrm{CA}(\\mathbf{r}_t \\odot \\mathbf{h}_{t-1}, \\mathbf{h}_t^l, W_{Q,K,V}^{\\prime\\prime}))\\), \\(\\odot\\) is the elementwise multiplication. The superscript of \\(W_{Q,K,V}\\) denotes the different projection weights taken independently in each attention calculation. Finally, we iteratively update the current global motion representation at the feature level by:" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.212, + 0.483, + 0.233 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} _ {1 \\rightarrow t + 1} ^ {\\text {g l o b a l}} \\doteq \\mathbf {h} _ {t} = \\left(1 - \\mathbf {z} _ {t}\\right) \\odot \\mathbf {h} _ {t - 1} + \\mathbf {z} _ {t} \\odot \\mathbf {s} _ {t}. \\quad (4)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.247, + 0.485, + 0.309 + ], + "angle": 0, + "content": "The simple and effective recurrent temporal aggregation we take is naturally compatible with the streaming pipeline, and we verify its effectiveness compared to previous fusion solutions in ablation experiments (cf. Sec. 4.4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.335, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Multi-frame iterative streaming framework. As depicted in Fig. 1, our framework iteratively processes the input video and event data through local motion estimation and global motion accumulation. We aggregate the local motion representations from each frame interval to the global motion representation at the feature level through the above global aggregation module in Sec. 3.2. Subsequently, the multi-frame accumulation step described in Sec. 3.1 sequentially combines each inter-frame short-term curve into a global long-term motion trajectory at the trajectory level, providing the dense and continuous point tracking representation as the model output. On the right side of Fig. 1, local motion is visualized with dense optical flow, and global motion is represented with deformations of dense point grid and curve trajectories of sparse query points." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.204, + 0.592 + ], + "angle": 0, + "content": "3.3. Objective" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Temporal discrete trajectory supervision. The available point tracking datasets provide only temporally discrete point tracks with no ground truth for continuous interframe trajectories. Following DOT [35], we first adopt supervised losses based on the temporal discrete ground-truth point tracks provided by the dataset, which consists of the L1 loss \\(L_{\\text{traj}}\\) for sampled discrete trajectory prediction and the binary cross-entropy loss \\(L_{\\text{vis}}\\) for visibility map. We then randomly select different frame intervals for augmented training. Local correlation is not constructed when the frames are skipped, therefore the corresponding event features are taken into streaming for iteratively updating the global motion representation. There are cases where some images are not used as input when the frame interval is greater than 1, but the corresponding input events and ground-truth tracks can be regarded as inter-frame motion contributing to curve trajectory learning. Such sampling-based augmentation ensures the model learning through diverse long- and short-term motions, capitalizing on the continuity of events to estimate continuous trajectories." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Event consistency with continuous trajectory. Since events are usually generated along motion trajectories, we propose to leverage the continuous property of events for self-supervised continuous trajectory learning in conjunction with discrete supervision of point trajectories. However, events are computationally intensive to process one by one and are generally accompanied by noise. We thus first introduce event temporal chunking to process events in batches within a fixed duration to reduce the noise impact and computation. For the \\(b\\)-th interval of \\(B\\) chunks, we isolate the events within that \\(b\\)-th chunk and aggregate them after warping them to \\(t_b\\) as Eq. (2). For each chunk, the events then are summed into an image of warped events (IWE) [8], i.e., \\(\\mathbf{EB}(\\mathbf{x}_i,b)\\doteq \\sum_{i = 1}^{Ne}\\mathcal{N}(\\mathbf{x}_i;x_i',\\sigma^2)\\), where \\(t_b\\leq t_i < t_{b + 1}\\) and \\(\\sigma\\) is the neighboring range which is usually chosen as 1 pixel. This IWE essentially counts the number of warped events \\(e_i^\\prime\\) per pixel and chunk. The chunking intervals are chosen randomly to exploit the continuous nature of events. Thus we can establish consistent connections between event chunks and continuous trajectories:" + }, + { + "type": "equation", + "bbox": [ + 0.52, + 0.41, + 0.907, + 0.467 + ], + "angle": 0, + "content": "\\[\nL _ {e c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {b _ {1} \\neq b _ {2}} ^ {B} \\rho \\binom {\\mathbf {E B} (\\mathbf {x}, b _ {1}),} {\\text {W a r p} \\left(\\mathbf {E B} (\\mathbf {x}, b _ {2}); \\mathbf {T} \\left(t _ {b _ {2}}\\right) - \\mathbf {T} \\left(t _ {b _ {1}}\\right)\\right)}, \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.528 + ], + "angle": 0, + "content": "where \\(\\rho\\) is the consistency measure with L1 norm. Since events are spatially sparse, we only establish connections with the dense trajectories at locations \\(\\Omega\\) with valid events." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.56, + 0.906, + 0.68 + ], + "angle": 0, + "content": "Image consistency with discrete trajectory. Similar to the unsupervised optical flow task [30], we can also establish the discrete consistency of the motion trajectory with the images at discrete times, to compensate for the spatial sparsity issue of ground-truth point tracks in the training data. In addition, in our sampling-based augmented training, skipped images can be used as additional continuity training objectives." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.683, + 0.907, + 0.743 + ], + "angle": 0, + "content": "For the accumulated continuous global trajectory \\(\\mathbf{T}_{1\\rightarrow t}\\) we sample the the discrete optical flow \\(\\mathbf{F}_{i\\to j}\\) from \\(I_{i}\\) to \\(I_{j}\\) via timestamps. Similar to Eq. (5), the consistency of images can be modeled as:" + }, + { + "type": "equation", + "bbox": [ + 0.533, + 0.759, + 0.907, + 0.803 + ], + "angle": 0, + "content": "\\[\nL _ {i c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {i \\neq j} ^ {N _ {v}} \\rho \\left(I _ {i} (\\mathbf {x}), \\operatorname {W a r p} \\left(I _ {k} (\\mathbf {x}); \\mathbf {F} _ {i \\rightarrow j} (\\mathbf {x})\\right)\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Total objective. Our total training objective is a weighted combination of the above objectives, i.e., \\( L = L_{\\text{traj}} + \\lambda_1 L_{vis} + \\lambda_2 L_{ec} + \\lambda_3 L_{ic} \\), \\( \\lambda \\) are manual hyperparameters. Our ablations verify that combined self-supervised training can compensate for the temporal continuity of trajectories." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7940" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.132, + 0.089, + 0.225, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.161, + 0.507, + 0.173 + ], + "angle": 0, + "content": "Input images with GT query points: frame 1, 7, 19, 27" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.089, + 0.332, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.089, + 0.438, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.089, + 0.543, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.132, + 0.176, + 0.225, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.142, + 0.248, + 0.214, + 0.259 + ], + "angle": 0, + "content": "Init Coords" + }, + { + "type": "image", + "bbox": [ + 0.238, + 0.176, + 0.332, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.344, + 0.177, + 0.438, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.451, + 0.177, + 0.543, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.324, + 0.248, + 0.459, + 0.26 + ], + "angle": 0, + "content": "DOT: frame 7, 19, 27" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.089, + 0.648, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.09, + 0.754, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.09, + 0.861, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.623, + 0.16, + 0.796, + 0.173 + ], + "angle": 0, + "content": "CoTracker: frame 7, 19, 27" + }, + { + "type": "image", + "bbox": [ + 0.558, + 0.176, + 0.649, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.663, + 0.177, + 0.755, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.177, + 0.862, + 0.248 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.248, + 0.778, + 0.26 + ], + "angle": 0, + "content": "Ours: frame 7, 19, 27" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.266, + 0.905, + 0.294 + ], + "angle": 0, + "content": "Figure 2. Visual comparisons of long-term dense point tracking on the TAP-Vid-DAVIS [5] dataset, with the ground-truth sparse query points with input images. Due to space limitations, we provide more sequences and event data visualization in the Supp. and demo video." + }, + { + "type": "table_caption", + "bbox": [ + 0.225, + 0.297, + 0.771, + 0.31 + ], + "angle": 0, + "content": "Table 1. Quantitative results of dense evaluation on the CVO test [52] and extended set [35]." + }, + { + "type": "table", + "bbox": [ + 0.174, + 0.313, + 0.819, + 0.468 + ], + "angle": 0, + "content": "
MethodCVO (Clean)CVO (Final)CVO (Extended)
EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑
QueryPIPs++ [58]9.05 / 6.62 / 21.533.39.49 / 7.06 / 22.032.718.4 / 10.0 / 32.158.7
TAPIR [6]3.80 / 1.49 / 14.773.54.19 / 1.86 / 15.372.419.8 / 4.74 / 42.568.4
CoTracker [24]1.51 / 0.88 / 4.5775.51.52 / 0.93 / 4.3875.35.20 / 3.84 / 7.7070.4
DenseGMA [22]2.42 / 1.38 / 7.1460.52.57 / 1.52 / 7.2259.721.8 / 15.7 / 32.865.6
MFT [36]2.91 / 1.39 / 9.9319.43.16 / 1.56 / 10.319.521.4 / 9.20 / 41.837.6
AccFlow [52]1.69 / 1.08 / 4.7048.11.73 / 1.15 / 4.6347.536.7 / 28.1 / 52.936.5
DOT [35]1.32 / 0.74 / 4.1280.41.38 / 0.82 / 4.1080.25.07 / 3.67 / 7.3471.0
EDCPT (Ours)1.23 / 0.71 / 3.8382.11.31 / 0.76 / 3.8681.94.88 / 3.44 / 7.4671.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.479, + 0.483, + 0.507 + ], + "angle": 0, + "content": "Table 2. Quantitative results of sparse evaluation on the TAP-VidDAVIS [5]point tracking benchmark." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.508, + 0.466, + 0.707 + ], + "angle": 0, + "content": "
MethodDAVIS (First)DAVIS (Strided)
AJ ↑<δxavg↑OA ↑AJ ↑<δxavg↑OA ↑
QueryTAP-Net [5]33.048.678.838.453.182.3
Context-PIPs [51]42.760.379.548.964.083.4
TAPIR [6]56.270.086.561.373.688.8
CoTracker [24]61.174.689.163.579.887.8
SpatialTracker [54]61.176.389.5---
DenseCPFlow [31]9.614.6----
MFT [36]47.366.877.856.170.886.9
DecoMotion [26]53.069.984.260.274.487.2
DinoTracker [46]---62.378.287.5
FlowTrack [4]---63.276.389.2
DOT [35]61.675.589.566.780.690.4
EDCPT (Ours)63.876.390.667.580.591.1
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.224, + 0.735 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.287, + 0.759 + ], + "angle": 0, + "content": "4.1. Experimental details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We train our model on MOVI-F [15] dataset and directly evaluate it on CVO [52] and TAP-Vid-DAVIS [5] benchmarks, using the vid2e [10] simulator to generate events.Following DOT [35], we train the model for 500k steps on \\(4\\times\\) NVIDIA L40 48G GPUs, using the Adam optimizer and OneCycle learning rate decay with a maximum of \\(10^{-4}\\) . We choose 3 frames as training samples, along with the random selection of up to 10 frames in different frame intervals. The loss hyperparameters are set to 1.0," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.907, + 0.603 + ], + "angle": 0, + "content": "0.1, 0.1. For dense evaluation, we report the dense absolute error \\(\\mathrm{EPE}_{\\mathrm{all / vis / occ}}\\) for all/visible/occluded points, as well as occlusion IoU accuracy OA for estimated visible mask. For sparse evaluation, we follow TAPNet [5] and report average Jaccard AJ, position accuracy \\(< \\delta_{\\mathrm{avg}}^x\\), and occlusion accuracy OA. Additionally, we adopt the real-captured event-based optical flow benchmark, DSEC [12, 13], to verify our adaptation capacity. More details are provided in the Supp." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.618, + 0.854, + 0.635 + ], + "angle": 0, + "content": "4.2. Standard spatially dense point tracking" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.907, + 0.823 + ], + "angle": 0, + "content": "Performance on CVO and TAP-Vid-DAVIS benchmarks. We conduct comprehensive evaluations on two common standard point tracking benchmarks. Consistent with DOT [35], we report the quantitative results of the spatially dense optical flow from the last to the first frame of CVO [52] dataset in Tab. 1. Our proposed new EDCPT framework achieves significant performance improvements, whether comparing methods that only predict the partial Query points or directly estimating spatially Dense trajectories within a single inference. Particularly, we achieve 0.19 \\(\\mathrm{EPE}_{\\mathrm{all}}\\) and 0.9 OA improvements on the extended set of 476 videos with 48 frames when compared to DOT [35]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In contrast, the real TAP-Vid-DAVIS dataset [5] only provides ground-truth trajectories for selected query points. As a result, we only sparsely evaluate these points for a fair comparison despite the output trajectories of our model and some compared methods are spatially dense. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7941" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.089, + 0.37, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.372, + 0.089, + 0.613, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.09, + 0.857, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.095, + 0.224, + 0.9, + 0.238 + ], + "angle": 0, + "content": "Figure 3. Visual comparisons of dense and continuous point trajectories on TAP-Vid-DAVIS [5]. Zoom in for detailed curve trajectories." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.292 + ], + "angle": 0, + "content": "Table 3. Quantitative results on the real DSEC optical flow benchmark [13]. The first two methods are categorized as self-supervised learning and the remaining as supervised learning." + }, + { + "type": "table", + "bbox": [ + 0.115, + 0.292, + 0.457, + 0.436 + ], + "angle": 0, + "content": "
MethodInputEPE ↓AE ↓%Out ↓
Taming [39]Events2.3310.5617.77
MPCMax [18]Events3.208.5315.21
E-RAFT [13]Events0.792.852.68
TMA [28]Events0.742.682.30
IDNet [53]Events0.722.722.04
ECDDP [56]Events0.702.581.96
BFlow [14]Events0.752.682.44
BFlow [14]Images+Events0.692.421.88
EDCPT (Ours)Images+Events0.632.171.52
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.484, + 0.631 + ], + "angle": 0, + "content": "quantitative results in Tab. 2 demonstrate the superiority of our framework, outperforming the existing state-of-the-art methods DOT [35] and SpatialTracker [54] with up to 2.7 AJ and 1.1 OA. We argue that ours \\( < \\delta_{\\mathrm{avg}}^{x} \\) is close to DOT's because the sparse to dense upgrade strategy from sparse point matches is capable of modeling the rigid object motion and thus benefiting the proportion evaluation. We also perform qualitative visual comparisons in Fig. 2. Combining the above quantitative and qualitative comparisons with previous image-based methods, the new attempts of incorporating events by our framework significantly improve the accuracy with the standard dense point tracking setting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.484, + 0.82 + ], + "angle": 0, + "content": "Performance on DSEC benchmark. We further conduct experiments on the DSEC benchmark [13] with real captured event data. Because the DSEC online leaderboard only measures the optical flow between two consecutive frames, we therefore finetune the local motion estimation on the DSEC training set from our pre-trained long-term model. The submission results are shown in Tab. 3. Notably, while BFlow [14] can estimate curve trajectories, they only submitted another optical flow version. Our framework fuses image and event data, yielding state-of-the-art performance with improvements of 0.05 EPE and 0.25 AE." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.833, + 0.419, + 0.849 + ], + "angle": 0, + "content": "4.3. Temporally continuous point tracking" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We adapt the above standard evaluation to input only a portion of the full video frames into the model to evaluate the temporal continuity with the ground truths of skipped" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.905, + 0.403 + ], + "angle": 0, + "content": "frames. For the CVO final set with 7 frames per video, we skip 1 frame (half) and 2 frames (one-third) as model inputs, because the longer extended set lacks multi-frame ground-truth tracks. For the TAP-Vid-DAVIS dataset with \\(\\sim 100\\) frames, we report skip 1 frame (half) and skip 3 frames (quarter). The compared image-based methods cannot model inter-frame motion, thus we take linear motion interpolation to generate trajectory when frames are skipped. We retrain AccFlow [52] and mark it with * as its public version does not support forward point tracking." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.906, + 0.646 + ], + "angle": 0, + "content": "As reported in Tab. 4, our proposed new framework with global continuous trajectory accumulation significantly outperforms existing methods. Especially in nonlinear motion scenarios of TAP-Vid-DAVIS datasets, the larger frame intervals lead to greater performance gaps. In addition to the ablation of different motion assumptions in Tab. 6, the B-spline representation we adopt achieves better performance. We also provide visual comparisons in Fig. 3 and a demo video of continuous trajectory visualization in the Supp., that includes four sequences with diverse motion complexity and event sparsity from TAP-Vid-DAVIS and real-captured ERF-X170FPS [25] dataset. With extensive experiments on both simulated events (CVO and TAP-Vid-DAVIS) and real events (DSEC and ERF-X170FPS), we fully validate the robustness of the proposed global motion accumulation in modeling complex continuous trajectories." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.658, + 0.838, + 0.674 + ], + "angle": 0, + "content": "4.4. Ablation experiments and discussions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.681, + 0.906, + 0.772 + ], + "angle": 0, + "content": "To perform progressive ablations in Tab. 5, Tab. 6, and Tab. 7, the underlined settings are those utilized in the previous table, and the bolded ones represent the choices for our final framework. To validate the capability for continuous point tracking, the ablation results are reported with CVO third and DAVIS quarter settings as in Sec. 4.3 and Tab. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Global motion aggregation. One of our key contributions is the global aggregation of local motion representations in the streaming pipeline. In Tab. 5, N/A indicates do not explicitly model sequential motion as in DOT [35]. post is the forward aggregation in AccFlow [52], which performs post-processing only on the output optical flow and ignores the internal long-term motion information. We aggregate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7942" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.154, + 0.089, + 0.843, + 0.104 + ], + "angle": 0, + "content": "Table 4. Continuous point tracking evaluation results on the CVO extended set [35] and TAP-Vid-DAVIS dataset [5]." + }, + { + "type": "table", + "bbox": [ + 0.155, + 0.106, + 0.842, + 0.216 + ], + "angle": 0, + "content": "
MethodCVO (Final) - EPEall/vis/occ ↓DAVIS (First) - AJ / <δxavg↑
fullhalfthirdfullhalfquarter
RAFT2.09 / 0.81 / 8.022.44 / 0.95 / 9.073.02 / 1.16 / 11.4233.9 / 46.628.6 / 40.122.4 / 34.2
GMA1.99 / 0.77 / 7.572.35 / 0.89 / 8.452.92 / 1.09 / 10.9739.3 / 52.531.7 / 44.326.5 / 38.3
AccFlow*2.28 / 0.60 / 11.182.39 / 0.80 / 10.742.79 / 1.02 / 11.3747.2 / 62.337.5 / 49.230.9 / 42.4
CoTracker1.89 / 0.63 / 7.052.11 / 0.82 / 8.023.17 / 1.65 / 11.1361.1 / 74.654.3 / 68.848.9 / 63.9
DOT1.83 / 0.59 / 6.952.10 / 0.73 / 7.882.69 / 0.97 / 10.8461.6 / 75.555.6 / 70.150.4 / 65.3
EDCPT (Ours)1.76 / 0.55 / 6.731.97 / 0.66 / 7.612.16 / 0.73 / 8.7663.8 / 76.359.7 / 73.156.2 / 70.9
" + }, + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.223, + 0.899, + 0.25 + ], + "angle": 0, + "content": "Table 5. Ablations on global motion aggre- Table 6. Ablations on parametric curve rep- Table 7. Ablations on input data and supervi-gation. resentation. sion." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.252, + 0.36, + 0.348 + ], + "angle": 0, + "content": "
EPEall/vis/occ ↓AJ / <δxavg↑
N/A2.54 / 0.89 / 10.3351.9 / 66.4
post2.49 / 0.85 / 9.7352.5 / 66.9
solo2.45 / 0.82 / 9.8952.7 / 67.0
–offsets2.43 / 0.82 / 9.7353.0 / 67.2
stream2.42 / 0.80 / 9.6853.3 / 67.4
" + }, + { + "type": "table", + "bbox": [ + 0.366, + 0.252, + 0.627, + 0.35 + ], + "angle": 0, + "content": "
EPEall/vis/occ ↓AJ / <δxavg↑
linear2.42 / 0.80 / 9.6853.3 / 67.4
quad2.49 / 0.84 / 9.4654.2 / 68.1
Nc=32.32 / 0.79 / 9.3354.4 / 68.6
Nc=42.23 / 0.76 / 8.9755.4 / 69.7
Nc=52.26 / 0.75 / 9.2055.0 / 69.3
" + }, + { + "type": "table", + "bbox": [ + 0.632, + 0.252, + 0.896, + 0.331 + ], + "angle": 0, + "content": "
EPEall/vis/occ ↓AJ / <δxavg ↑
Images2.50 / 0.86 / 9.9252.5 / 66.7
+Events2.23 / 0.76 / 8.9755.4 / 69.7
+Lic2.20 / 0.75 / 8.9055.8 / 70.2
+Lec2.16 / 0.73 / 8.7656.2 / 70.9
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.365, + 0.483, + 0.576 + ], + "angle": 0, + "content": "motion representations at the feature level instead of dealing directly with motion vectors. solo is the short and long term fusion module in SOLOFusion [40]. Unlike its temporal fusion with a fixed number of frames and low resolution, we adopt sequential modeling with an unspecified number of frames in a streaming pipeline. Our proposed motion aggregation framework fuses image correspondence and event features from local to global with recurrent streaming structure, which better preserves multi-frame motion information and achieves optimal performance. In addition, we also verified that removing the additional offset estimation to address the numerical problem in Warping leads to a slight performance degradation, as this would require the subsequent refinement to handle it simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.483, + 0.716 + ], + "angle": 0, + "content": "Curve representation. Based on our streaming framework, we compare the performance improvement of taking the B-spline curve representation compared to interpolating with linear and quadratic motion assumptions in Tab. 6. For the control points number of B-spline curves, we chose \\( N_{c} = 4 \\), since further increasing the number of control points does not significantly improve the performance under CVO third and DAVIS quarter settings." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Input data and supervision. Since previous methods usually use only image data, we verify the advantages of incorporating event data for high-precision continuous point tracking by removing events, as depicted in Tab. 7. Moreover, the comparison results between this image-only variant and DOT [35] in Tab. 4 demonstrates that our proposed streaming aggregation and curve representation are beneficial even in the absence of event data. Furthermore, we validate that training using the proposed image and event-to-point trajectory consistencies as additional supervision complements the lack of continuous inter-frame tracks in" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.365, + 0.875, + 0.38 + ], + "angle": 0, + "content": "the training data and can further improve performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.404, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Limitations. On the same RTX 3090 PC, our framework processes a 48-frame video at \\(512 \\times 512\\) resolution in 12.6 seconds, significantly faster than CoTracker's 11 minutes, while only processing partial query points in a single run. However, it is slower than the two-frame optical flow method GMA of only 2.1 seconds and slightly slower than the multi-frame point tracking method DOT of 9.5 seconds. In addition, there is currently a lack of point-tracking datasets with real events, and it is difficult to obtain long-term tracking annotations. We evaluate on standard benchmarks with simulated events, as well as visualize point tracking on real events, and also evaluate optical flow estimation on a real event benchmark. Our future work plans to improve on model efficiency and the evaluation setup." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.633, + 0.634, + 0.649 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this paper, we propose a new long-term dense point tracking framework for integrating image and event data to estimate continuous motion trajectories. Specifically, we streamingly process two adjacent frames and inter-frame events to generate local motion representations, and accumulate previous caches to global motion representation at the feature level to produce new global trajectories at the trajectory level. We utilize multi-frame parametric curve accumulation to represent continuous trajectories with any number of frames, complemented by image and event-to-trajectory consistencies for model training. We validate the effectiveness of our framework by conducting extensive benchmark experiments as well as qualitative verification on real challenge data. We believe this work provides new insights into the point tracking task from the perspective of event modality and continuous global curve representations." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7943" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This research/project is supported by the National Natural Science Foundation of China (62271410, 12150007), the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008), and the Tier 1 grant T1-251RES2305 from the Singapore Ministry of Education. Zhexiong Wan was also supported by the Program of China Scholarship Council (202306290193), and the Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University (CX2023013)." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.281, + 0.188, + 0.297 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.306, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scalespace flow for end-to-end optimized video compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.377, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[2] Adam Bielski and Paolo Favaro. Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems (NeurIPS), 35:33371-33386, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.435, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[3] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.491, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[4] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268–19277, 2024. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.561, + 0.484, + 0.646 + ], + "angle": 0, + "content": "[5] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re-casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, pages 13610–13626, 2022. 1, 2, 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.647, + 0.484, + 0.717 + ], + "angle": 0, + "content": "[6] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10061-10072, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.718, + 0.484, + 0.787 + ], + "angle": 0, + "content": "[7] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. In Asian Conference on Computer Vision (ACCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.788, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[8] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and" + }, + { + "type": "list", + "bbox": [ + 0.1, + 0.306, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.134 + ], + "angle": 0, + "content": "Davide Scaramuzza. Event-based vision: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):154-180, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[10] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3586-3595, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.909, + 0.262 + ], + "angle": 0, + "content": "[11] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision (IJCV), 128(3):601-618, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[12] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters (RA-L), 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.376 + ], + "angle": 0, + "content": "[13] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.906, + 0.434 + ], + "angle": 0, + "content": "[14] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 46(7):4736-4746, 2024. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.435, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3749-3761, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.906, + 0.589 + ], + "angle": 0, + "content": "[16] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16022-16033, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.658 + ], + "angle": 0, + "content": "[17] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 7167-7179, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.66, + 0.906, + 0.731 + ], + "angle": 0, + "content": "[18] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In European Conference on Computer Vision (ECCV), 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.906, + 0.788 + ], + "angle": 0, + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.906, + 0.857 + ], + "angle": 0, + "content": "[20] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. Progressive spatio-temporal alignment for efficient event-based motion estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1537-1546, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.858, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[21] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. FlowNet 2.0: Evolution of optical flow estimation with deep networks." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.909, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7944" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.121 + ], + "angle": 0, + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.123, + 0.483, + 0.191 + ], + "angle": 0, + "content": "[22] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9772-9781, 2021. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.194, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[23] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.267, + 0.482, + 0.322 + ], + "angle": 0, + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision (ECCV), 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.325, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[25] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.438 + ], + "angle": 0, + "content": "[26] Rui Li and Dong Liu. Decomposition betters tracking everything everywhere. In European Conference on Computer Vision (ECCV), 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[27] Siqi Li, Zhikuan Zhou, Zhou Xue, Yipeng Li, Shaoyi Du, and Yue Gao. 3d feature tracking via event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18974-18983, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.5, + 0.482, + 0.568 + ], + "angle": 0, + "content": "[28] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9685-9694, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.572, + 0.482, + 0.626 + ], + "angle": 0, + "content": "[29] Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate. arXiv preprint arXiv:2409.11953, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.63, + 0.482, + 0.713 + ], + "angle": 0, + "content": "[30] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6489-6498, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.716, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[31] Jianqin Luo, Zhexiong Wan, Yuxin Mao, Bo Li, and Yuchao Dai. Continuous parametric optical flow. In Advances in Neural Information Processing Systems (NeurIPS), pages 23520-23532, 2023. 1, 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.774, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[32] Xinglong Luo, Ao Luo, Zhengning Wang, Chunyu Lin, Bing Zeng, and Shuaicheng Liu. Efficient meshflow and optical flow estimation from event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19198-19207, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[33] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5642-5651, 2023. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[34] Etienne Meunier, Anais Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4):4462-4473, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[35] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.259 + ], + "angle": 0, + "content": "[36] Michal Neoral, Jonás Šerzych, and Jérí Matas. Mft: Long-term tracking of every pixel. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 6837-6847, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.261, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (TMLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[38] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1669-1678, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.524 + ], + "angle": 0, + "content": "[39] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9695-9705, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.526, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris M. Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multiview 3d object detection. In International Conference on Learning Representations (ICLR), 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[41] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[42] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.72, + 0.905, + 0.789 + ], + "angle": 0, + "content": "[43] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. CRAFT: Cross-attentional flow transformer for robust optical flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17602-17611, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[44] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419, 2020. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[45] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time Lens: Event-based video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, 2021. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7945" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.203 + ], + "angle": 0, + "content": "[47] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing (TIP), 31:7237-7251, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.272 + ], + "angle": 0, + "content": "[48] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. RPEFlow: Multimodal fusion of RGB-pointcloud-event for joint optical flow and scene flow estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10030-10040, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19795-19806, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[50] Xiangyuan Wang, Huai Yu, Lei Yu, Wen Yang, and GuiSong Xia. Towards robust keypoint detection and tracking: A fusion approach with event-aligned image features. IEEE Robotics and Automation Letters (RA-L), 9(9):8059-8066, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[51] BIAN Weikang, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands spatial context features. In Advances in Neural Information Processing Systems (NeurIPS), pages 55285-55298, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.564 + ], + "angle": 0, + "content": "[52] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: backward accumulation for long-range optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12119-12128, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.482, + 0.634 + ], + "angle": 0, + "content": "[53] Yilun Wu, Federico Paredes-Vallés, and Guido CHE De Croon. Lightweight event-based optical flow estimation via iterative deblurring. In IEEE International Conference on Robotics and Automation (ICRA), pages 14708-14715. IEEE, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[54] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.482, + 0.761 + ], + "angle": 0, + "content": "[55] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In Advances in Neural Information Processing Systems (NeurIPS), pages 1647-1656, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.763, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[56] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data dense pre-training. In European Conference on Computer Vision (ECCV), pages 292-310. Springer, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.805, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[57] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9781-9790, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[58] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "synthetic dataset for long-term point tracking. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19855-19865, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[59] Hanyu Zhou, Yi Chang, Haoyue Liu, Wending Yan, Yuxing Duan, Zhiwei Shi, and Luxin Yan. Exploring the common appearance-boundary adaptation for nighttime optical flow. In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[60] Hanyu Zhou, Yi Chang, and Zhiwei Shi. Bring event into rgb and lidar: Hierarchical visual-motion fusion for scene flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26477-26486, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[61] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 34(8):4868-4880, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[62] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.457 + ], + "angle": 0, + "content": "[63] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for RGB-event transformer-trackers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 22045-22055, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.956 + ], + "angle": 0, + "content": "7946" + } + ] +] \ No newline at end of file diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_origin.pdf b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cd60b29170d6ef9a96497fd8be5d0a189affb4a3 --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/9ca8a16e-beb5-412b-96ff-28204d9a5d86_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:dc216b4f77747c28556dc2917150be473902d123e687253e56b1f6856eac6f33 +size 2116998 diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/full.md b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/full.md new file mode 100644 index 0000000000000000000000000000000000000000..f362024bdc2c3a30a5c0c937ca914bbd2cbed533 --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/full.md @@ -0,0 +1,319 @@ +# Event-aided Dense and Continuous Point Tracking: Everywhere and Anytime + +Zhexiong Wan $^{1,2}$ Jianqin Luo $^{1}$ Yuchao Dai $^{1}$ Gim Hee Lee $^{2}$ + +$^{1}$ School of Electronics and Information, Northwestern Polytechnical University & Shaanxi Key Laboratory of Information Acquisition and Processing + +$^{2}$ Department of Computer Science, National University of Singapore + +{wanzhexiong,luojianqin}@mail.nwpu.edu.cn，daiyuchao@nwpu.edu.cn，gimhee.lee@nus.edu.sg + +# Abstract + +Recent point tracking methods have made great strides in recovering the trajectories of any point (especially key points) in long video sequences associated with large motions. However, the spatial and temporal granularities of point trajectories remain constrained by limited motion estimation accuracy and video frame rate. Leveraging the high temporal resolution and motion sensitivity of event cameras, we introduce event data for the first time to recover spatially dense and temporally continuous trajectories of every point at any time. Specifically, we define the dense and continuous point trajectory representation as estimating multiple control points of curves for each pixel and model the movement of sparse events triggered along continuous point trajectories. Building on this, we propose a novel multi-frame iterative streaming framework that first estimates local inter-frame motion representations from two consecutive frames with inter-frame events, then aggregates them into a global long-term motion representation to utilize input full video and event data with an arbitrary number of frames. Extensive experiments on simulated and real data demonstrate the significant improvement of our framework over state-of-the-art methods and the crucial role of introducing events to model continuous point trajectories. + +# 1. Introduction + +Estimating fine-grained motion from input videos is a crucial task in computer vision with widespread applications such as video compression [1], frame interpolation [23, 55], motion segmentation [2, 34], and dynamic scene reconstruction [16]. Early studies [21, 44] based on two-frame optical flow can model the spatially dense motion, but they suffer from the challenge of modeling long-term motion from more frames. With the proposal of tracking any point (TAP) task [5, 19], using sparse points as query indexes to + +estimate pointwise long-term motions also draws attention. Despite significant progress, the sparse and independent representation of pointwise trajectories remains inherently incompatible with the spatially dense representation of input video. Thus, recently there are methods [4, 35, 46] that continue to use dense optical flow to represent long-term motion. However, the temporal frame rate of the input video is constrained by traditional shutter cameras, which makes accurately modeling complex motion trajectories difficult for video-only methods from the data acquisition perspective. Consequently, accurate estimation of fine-grained spatially dense and temporally continuous motion remains a challenging and worthwhile research problem. + +Unlike traditional shutter cameras that expose the entire image at fixed frame rates, the new bio-inspired event cameras [9] can independently and asynchronously detect per-pixel brightness changes with microsecond precision. This unique design makes it inherently sensitive to visual motion changes, leading to impressive motion-related applications such as optical flow estimation [17, 62], motion segmentation [20, 61], feature tracking [33], object tracking [63], and frame interpolation [45]. However, events are typically triggered only with motion contours or rich textures, making it challenging to comprehensively perceive spatially dense motion. Recently, integrating the advantages of event camera and traditional shutter camera has become a new direction [38, 57]. Leveraging on these successes, we propose for the first time taking event data as an auxiliary input for the point tracking task, where we comprehensively model fine-grained spatially dense and temporally continuous motion from the input image and event modalities. + +The introduction of event cameras offers the potential to model continuous motion from the data perspective. However, to effectively model continuous long-term motion, we need a new representation that replaces optical flow and parametrically associates the dense temporal properties of event data. BFlow [14] proposes to learn trajectories with Bezier curves from events, but is limited to input fixed frames and cannot adapt to longer sequences. CPFlow [31] + +proposes to learn control points represented as B-spline curves from a fixed number of image slices. Although continuous motion can be successfully modeled using these curve representations in normalized timescales, their fixed number of control points made it hard to handle complex dynamics and varying lengths of video sequences. Based on the curve representation, we propose a new streaming pipeline for accumulating multiple local curves to address these limitations. In addition, how to adequately fuse multi-frame temporal information to achieve long-term tracking is also a concern. Compared to not explicitly associating multi-frames in DOT [35] and aggregating over estimated motion vectors in AccFlow [52] and FlowTrack [4], we perform recurrent fusion in the feature space, dedicated to preserving multi-frame sequence information. + +In this paper, we present the first event-aided point tracking framework for recovering spatially dense and temporally continuous point trajectories from input videos and event sequences. Specifically, we first propose a new point trajectory representation with parametric curves that accumulate multiple local curves to adapt to multi-frame input videos at any length. We then design a new framework for combining two frames with events to estimate dense point curve trajectories and extend to multi-frame iterative accumulation. In addition, we establish the association between continuous curve trajectory and event triggering as a part of the learning objective for continuous motion modeling. Extensive experiments on both simulated and real-world data demonstrate that the proposed framework significantly outperforms state-of-the-art methods. Particularly, our ablation studies illustrate the effectiveness of the proposed global aggregation and highlight the crucial role of incorporating event data in continuous trajectory modeling. + +Our main contributions are summarized as follows: + +- We introduce a new setup that, for the first time, enables long-term spatially dense and temporally continuous point tracking by integrating the strengths of both images and events. +- We present a novel global curve representation of continuous point trajectories through multi-frame aggregation, establishing a connection between event triggering and continuous motion. +- We propose a novel event-aided streaming framework that iteratively accumulates the local tracks from two frames with inter-frame events, resulting in global, long-term dense and continuous trajectories through iterative temporal global motion aggregation. + +# 2. Related Works + +# 2.1. Image-based query point tracking + +The goal of point tracking is to recover the corresponding positions of query points in each frame, which has attracted + +wide attention with the proposal of the TAP benchmark and the baseline model TAPNet [5]. PIPs [19] proposes to extract independent point representation for 8-frame tracking, then $\mathrm{PIPs}++$ [58] extends to long-term trajectories. TAPIR [6] proposes a two-stage matching framework that fuses TAPNet and PIPs, and BootsTAPIR [7] performs self-supervised training on more data. Unlike these methods that track only one query at a time, CoTracker [24] and Context-PIPs [51] use additional tracks and pixel featuresto improve global tracking. SpatialTracker [54] introduces the triplane representation for group pixels in 3D.DINOTracker [46] performs test-time finetuning on the pre-trained DINOViT [37] model and achieves fine-grained tracking per video. When applying these query points-based methods to achieve dense tracking, full points need to be processed individually or in batches, which brings computational hurdles and limits their downstream applications [35]. + +# 2.2. Image-based dense point tracking + +Recent studies turn to tracking every point within a frame in a single run, aiming to enhance neighborhood relationships while reducing computational requirements. OmniMotion [49] performs pixel-wise tracking between local and canonical space to maintain the global consistency of the motion, and then FastOmniTrack [42] and DecoMotion [26] improve from the perspectives of computation cost and object motion decomposition. CPFlow [31] proposes to estimate spatio-temporally dense motion curves, but it can only input 4 images and needs pre-sampling for longer video. AccFlow [52] proposes the forward and backward aggregation pipeline, extending inter-frame dense optical flow to multi-frame long ranges. MFT [36] select chaining multiframe candidates and FlowTrack [4] automatically apply error compensation in instances of tracking inaccuracies. DOT [35] unifies point tracking and optical flow, upgrading sparse tracks to dense flow fields between every frame. However, limited by the video frame rate, these methods are struggling to model challenging dynamics. In this work, we propose to introduce continuous event data into the input video to enable temporally continuous point tracking. + +# 2.3. Event-based motion estimation + +Thanks to the motion-sensitive nature of event cameras, extensive motion estimation studies in recent years have highlighted their potential applications to challenging dynamics. The feature tracking methods [11, 27, 33, 50] show the benefits of event cameras for low-latency tracking, but can only track sparse, specific textured locations. Recently estimating optical flow from events has become mainstream. Using only sparse event data [13, 32, 62] allows to estimate satisfactory dense optical flow, while introducing data from other sensors such as images [47, 59] and point clouds [48, 60] achieves significant performance + +gains. BFlow [14] and MotionPriorCMax [18] exploit the continuous property of events to estimate parametric Bezier trajectories, but can only estimate motion within a fixed consecutive frame interval and cannot be directly adapted to long-term sequences. Recently, FE-TAP [29] proposes to recover point trajectories from a fixed number of images and events based on TAPVid [5], but does not take full advantage of the continuous nature of events. We propose to combine the advantages of images and events to enable temporally continuous point tracking by modeling long-term global motion with an arbitrary number of frames. + +# 3. Method + +Overview. To address the limitations of sparse events and low video frame rates in modeling fine-grained motion, we present the first framework that recovers dense and continuous point trajectories from a video with corresponding event sequences. Our framework consists of four parts: 1) A parametric multi-frame continuous point trajectories representation; 2) A model of event triggering along the point trajectories; 3) A two-frame basis motion estimation model; 4) A multi-frame motion aggregation and streaming framework. + +Problem Formulation. Conventional shutter camera captures a video with $N_{v}$ frames $\{I_i\}_{i=1}^{N_v}$ at a fixed frame rate. Event camera triggers an unbounded event sequence $\{e_i\}_{i=1}^{N_e}$ , where $N_e$ is the number of events. Each event $e_i = \{\mathbf{x}_i, t_i, p_i\}$ consists of the pixel position $\mathbf{x}_i = (x,y)$ , timestamp $t_i$ with microsecond precision, and the brightness change polarity $p_i$ in log domain. Our goal is to combine these two modalities to recover the spatially dense and temporally continuous trajectories $\mathbf{T}_{1 \to t}$ of all points starting from the first frame at time 1 to last frame at time $t$ . + +# 3.1. Motion model + +Trajectory representation. Previous point tracking methods typically estimate two-channel motion vectors in the XY directions, which is the optical flow when representing dense trajectories [4, 35]. To learn the curve trajectory from the deep network, we instead learn the multiple control points as curved trajectories [31]. Specifically, we choose the B-spline curve as our curve representation, which is defined by $N_{c}$ control points $\{\mathbf{P}_i\}_{i=1}^{N_c}$ and basis functions $\{B_{i,p}(t)\}_{i=1}^{N_c}$ with degree $p$ . The continuous trajectory $\mathbf{T}(t)$ represented by b-spline curve in time variable $t$ is a collection of piecewise polynomial functions $\mathbf{T}(t) = \sum_{i=1}^{N_c} B_{i,p}(t) \mathbf{P}_i$ . Similar to learning dense flow, each pixel has an independent curve with control points denoted as $\mathbb{P} \in \mathbb{R}^{N_c \times 2 \times H \times W}$ , where $H \times W$ is the image size. This enables learnable motion modeling of dense and continuous point trajectories $\mathbf{T}$ with parametric curve representation. More details are provided in the Supp. + +Multi-frame global trajectories accumulation. Existing parametric motion modeling methods are fixed in the number of frames they can handle, e.g., BFlow [14] is limited to between two frames, and CPFlow [31] struggles to benefit from more than 4 frame inputs, resulting in suboptimal long-term trajectory modeling. Inspired by the practice of multi-frame optical flow accumulation [36, 52], we propose a new multi-frame curve trajectories accumulation strategy to handle long-term videos with arbitrary frames. + +Our accumulation framework works on a streaming pipeline, where the previous global trajectory $\mathbf{T}_{1\rightarrow t}$ with $(t - 1)\times N_{c}$ control points has been accumulated from the previous $t - 1$ local trajectories $\{\mathbf{T}_{i\rightarrow i + 1}\}_{i = 1}^{t - 1}$ when processing the $t$ -th step. When we get updated $t$ -th local trajectory $\mathbf{T}_{t\rightarrow t + 1}$ with $N_{c}$ control points from time $t$ to $t + 1$ , a simple approach is to directly accumulate the new global trajectory with $t\times N_{c}$ control points from time 1 to $t + 1$ by $\mathbf{T}_{1\rightarrow t + 1}(\mathbf{x}) = \left[\mathbf{T}_{1\rightarrow t}(\mathbf{x}),\mathrm{Warp}(\mathbf{T}_{t\rightarrow t + 1},\mathbf{T}_{1\rightarrow t})(\mathbf{x})\right]$ , where Warp represents backward warping, the operator [,] combines the control points of two sub-curves and creates a more complex curve. However, there are two problems with just warping operation: 1) It suffers from numerical error as integer sampling with floating-point coordinates, i.e., for warping vectors from $\mathbf{b}$ to $\mathbf{a}$ , $\mathrm{Warp}(\mathbf{a},\mathbf{b})(\mathbf{x}) = \mathbf{a}(\mathbf{x} + \mathbf{b}(\mathbf{x}))$ , $\mathbf{x}$ is integer coordinates but not $\mathbf{x} + \mathbf{b}(\mathbf{x})$ . 2) Some points may be occluded at time $t$ , resulting in failing to find the corresponding points. More details are provided in the Supp. + +Our framework iteratively maintains and learns to integrate from a global motion representation $\mathbf{M}_{1\rightarrow t}^{global}$ in the streaming process (cf. Sec. 3.2). For the first numerical problem, we estimate a start point offsets $\mathbf{O}_t\in \mathbb{R}^{2\times H\times W}$ learned from $\mathbf{M}_{1\rightarrow t}^{global}$ and normalized to the range $[-1,1]$ . Sampling compensation is achieved by adding this offset directly during warping. For the second occlusion problem, we introduce an occlusion solving strategy for occluded pixels. We additionally estimate the visibility map $\mathbf{V}_{1\rightarrow t}$ of each point from the initial frame to the $t$ -th frame as well as the trajectory update $\Delta \mathbf{T}_t$ . Aggregation is based on a warp with offset when the point $x$ is visible. When point $x$ is occluded, a learnable module Fusion is introduced to regress the point's coarse motion trajectory in $t\to t + 1$ from $\mathbf{M}_{1\rightarrow t}^{global}$ . Finally, the trajectory update $\Delta \mathbf{T}_t$ as a uniformly refinement serves the global trajectory accumulation goal. The refinement process can be modeled as follows: + +$$ +\mathbf {T} ^ {\prime} = \left\{\begin{array}{l l}\operatorname {W a r p} \left(\mathbf {T} _ {t \rightarrow t + 1}, \mathbf {T} _ {1 \rightarrow t}, \mathbf {O} _ {t}\right) + \Delta \mathbf {T} _ {t}&\text {i f} \mathbf {V} _ {1 \rightarrow t} (\mathbf {x}) = 1,\\\operatorname {F u s i o n} \left(\mathbf {T} _ {t \rightarrow t + 1}, \mathbf {T} _ {1 \rightarrow t}, \mathbf {M} _ {1 \rightarrow t} ^ {\text {g l o b a l}}\right) + \Delta \mathbf {T} _ {t}&\text {i f} \mathbf {V} _ {1 \rightarrow t} (\mathbf {x}) = 0,\end{array}\right. \tag {1} +$$ + +where $\mathbf{T}'$ is the local curve in $t\to t + 1$ warped to the starting coordinates at time 1. Then we direct combine it with the cached previous global curve to obtain the accumulated current global curve $\mathbf{T}_{1\rightarrow t + 1}^{accum}(\mathbf{x}) = \left[\mathbf{T}_{1\rightarrow t}(\mathbf{x}),\mathbf{T}'(\mathbf{x})\right]$ . + +![](images/dd78a4ad66fb8187085af31e2b7a244c880e9f25cb1dc399fc61216c6b96771e.jpg) +Figure 1. Our proposed event-aided dense and continuous point tracking framework consists of two main steps. 1) Local motion estimation: estimating short-term curve trajectories with $N_{c}$ control points from two consecutive images and inter-frame events, while concurrently updating the local motion representation. 2) Global motion accumulation: iteratively fusing the latest local motion representation with the previous global motion representation in a streaming manner for aggregating the latest global motion representation. Subsequently, the global long-term curve trajectories with $t \times N_{c}$ control points are optimized on trajectory combinations. + +Events along the trajectory. Following the contrast maximization framework [8], we assume that events are triggered along with the pixel motion trajectories at the moving boundary. For a motion trajectory $\mathbf{T}$ starting from pixel $\mathbf{x}_1$ at time $t_1$ , the generated event's coordinates satisfy the trajectory, i.e., $\mathbf{x}_1 = \mathbf{T}(t_1), \mathbf{x}_i - \mathbf{x}_1 = \mathbf{T}(t_i) - \mathbf{T}(t_1)$ . We can thus use the motion trajectory to transform the events $e_i \dot{=} \{\mathbf{x}_i, t_i, p_i\}$ back to time $t_1$ : + +$$ +e _ {i} ^ {\prime} \doteq \left\{\mathbf {x} _ {i} ^ {\prime} = \operatorname {W a r p} \left(\mathbf {x} _ {i}; \mathbf {T} (t _ {i}) - \mathbf {T} (t _ {1})\right), t _ {1}, p _ {i} \right\}. \tag {2} +$$ + +Assuming the trajectory $\mathbf{T}$ is accurate, this process transforms the event $e_i$ to the starting point position $\mathbf{x}_1$ of the trajectory, i.e., $\mathbf{x}_1 = \mathrm{Warp}(\mathbf{x}_1; \mathbf{T}(t_i - t_1))$ . Based on the correlated motion modeling of events and point trajectories, we build additional self-supervised training objectives in Sec. 3.3 to alleviate the lack of continuous trajectory annotations in the training datasets. + +# 3.2. Framework + +Two-frame basis model. The two-frame basis model is designed to recover inter-frame short-term trajectories $\mathbf{T}_{t\rightarrow t + 1}$ from the encoded features of input two consecutive frames $F_{t},F_{t + 1}$ and inter-frame events $E_{t\rightarrow t + 1}$ . In the feature extraction phase, we first convert the raw event data into a dense grid representation [41], followed by two feature encoders for two images and an event grid. Subsequently, we construct the local correlation between two frame features by matrix multiplication [44]. By leveraging the local correlation and events, we learn the local motion representation $M_{t\rightarrow t + 1}^{local}$ by a motion extractor which allows + +recovery of local dense trajectories $\mathbf{T}_{t\rightarrow t + 1}$ by a trajectory decoder. Specifically, the former utilizes a transformer fusion module in GMA [22], while the latter is based on the typical FPN decoder in RAFT [43]. The trajectory decoder estimates the coordinates of $N_{c}$ control points $\mathbb{P}_{t\to t + 1}$ and a single-channel visibility map $\{\mathbf{V}\}_{t\to t + 1}$ , which is essential for establishing multi-frame global accumulation in Eq. (1). + +Global motion aggregation module. For processing a video comprising $N_{v}$ frames, the above two-frame basis model needs to be streamed sequentially $N_{v} - 1$ times yielding multiple local motion representations $\{\mathbf{M}_{i\rightarrow i + 1}^{local}\}_{i = 1}^{t - 1}$ and curve trajectories $\{\mathbf{T}_{i\rightarrow i + 1}\}_{i = 1}^{t - 1}$ . To facilitate the accumulation of global multi-frame trajectories according to Sec. 3.1, the cached global motion representation $\mathbf{h}_{t - 1}\doteq \mathbf{M}_{1\to t}^{global}$ from the previous $t$ -frames is utilized as the query, while the current local motion representation $\mathbf{h}_t^l\doteq \mathbf{M}_{t\to t + 1}^{local}$ serves as the key and value. We first perform the linear projections and compute the cross-attention: + +$$ +\begin{array}{l} \operatorname {C A} \left(\mathbf {h} _ {t - 1}, \mathbf {h} _ {t} ^ {l}, W _ {Q, K, V}\right) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {T}}{\sqrt {d _ {k}}}\right) V \tag {3} \\ = \operatorname {s o f t m a x} \left(\frac {\left(W _ {Q} \cdot \mathbf {h} _ {t - 1}\right) \left(W _ {K} \cdot \mathbf {h} _ {t} ^ {l}\right) ^ {T}}{\sqrt {d _ {k}}}\right) \left(W _ {V} \cdot \mathbf {h} _ {t} ^ {l}\right), \\ \end{array} +$$ + +where $d_{k}$ is the channel size, $\cdot$ is the linear projection and $W_{Q,K,V}$ are the corresponding weights. We then conduct iterative fusion based on the gated activation unit (GRU) [3], where the update gate is $\mathbf{z}_t =$ + +sigmoid(CA( $\mathbf{h}_{t-1}$ , $\mathbf{h}_t^l$ , $W_{Q,K,V}$ ), the reset gate is $\mathbf{r}_t =$ sigmoid(CA( $\mathbf{h}_{t-1}$ , $\mathbf{h}_t^l$ , $W_{Q,K,V}^\prime$ ), and the hidden state is $\mathbf{s}_t = \tanh(\mathrm{CA}(\mathbf{r}_t \odot \mathbf{h}_{t-1}, \mathbf{h}_t^l, W_{Q,K,V}^{\prime\prime}))$ , $\odot$ is the elementwise multiplication. The superscript of $W_{Q,K,V}$ denotes the different projection weights taken independently in each attention calculation. Finally, we iteratively update the current global motion representation at the feature level by: + +$$ +\mathbf {M} _ {1 \rightarrow t + 1} ^ {\text {g l o b a l}} \doteq \mathbf {h} _ {t} = \left(1 - \mathbf {z} _ {t}\right) \odot \mathbf {h} _ {t - 1} + \mathbf {z} _ {t} \odot \mathbf {s} _ {t}. \quad (4) +$$ + +The simple and effective recurrent temporal aggregation we take is naturally compatible with the streaming pipeline, and we verify its effectiveness compared to previous fusion solutions in ablation experiments (cf. Sec. 4.4). + +Multi-frame iterative streaming framework. As depicted in Fig. 1, our framework iteratively processes the input video and event data through local motion estimation and global motion accumulation. We aggregate the local motion representations from each frame interval to the global motion representation at the feature level through the above global aggregation module in Sec. 3.2. Subsequently, the multi-frame accumulation step described in Sec. 3.1 sequentially combines each inter-frame short-term curve into a global long-term motion trajectory at the trajectory level, providing the dense and continuous point tracking representation as the model output. On the right side of Fig. 1, local motion is visualized with dense optical flow, and global motion is represented with deformations of dense point grid and curve trajectories of sparse query points. + +# 3.3. Objective + +Temporal discrete trajectory supervision. The available point tracking datasets provide only temporally discrete point tracks with no ground truth for continuous interframe trajectories. Following DOT [35], we first adopt supervised losses based on the temporal discrete ground-truth point tracks provided by the dataset, which consists of the L1 loss $L_{\text{traj}}$ for sampled discrete trajectory prediction and the binary cross-entropy loss $L_{\text{vis}}$ for visibility map. We then randomly select different frame intervals for augmented training. Local correlation is not constructed when the frames are skipped, therefore the corresponding event features are taken into streaming for iteratively updating the global motion representation. There are cases where some images are not used as input when the frame interval is greater than 1, but the corresponding input events and ground-truth tracks can be regarded as inter-frame motion contributing to curve trajectory learning. Such sampling-based augmentation ensures the model learning through diverse long- and short-term motions, capitalizing on the continuity of events to estimate continuous trajectories. + +Event consistency with continuous trajectory. Since events are usually generated along motion trajectories, we propose to leverage the continuous property of events for self-supervised continuous trajectory learning in conjunction with discrete supervision of point trajectories. However, events are computationally intensive to process one by one and are generally accompanied by noise. We thus first introduce event temporal chunking to process events in batches within a fixed duration to reduce the noise impact and computation. For the $b$ -th interval of $B$ chunks, we isolate the events within that $b$ -th chunk and aggregate them after warping them to $t_b$ as Eq. (2). For each chunk, the events then are summed into an image of warped events (IWE) [8], i.e., $\mathbf{EB}(\mathbf{x}_i,b)\doteq \sum_{i = 1}^{Ne}\mathcal{N}(\mathbf{x}_i;x_i',\sigma^2)$ , where $t_b\leq t_i < t_{b + 1}$ and $\sigma$ is the neighboring range which is usually chosen as 1 pixel. This IWE essentially counts the number of warped events $e_i^\prime$ per pixel and chunk. The chunking intervals are chosen randomly to exploit the continuous nature of events. Thus we can establish consistent connections between event chunks and continuous trajectories: + +$$ +L _ {e c} = \sum_ {\mathbf {x}} ^ {\Omega} \sum_ {b _ {1} \neq b _ {2}} ^ {B} \rho \binom {\mathbf {E B} (\mathbf {x}, b _ {1}),} {\text {W a r p} \left(\mathbf {E B} (\mathbf {x}, b _ {2}); \mathbf {T} \left(t _ {b _ {2}}\right) - \mathbf {T} \left(t _ {b _ {1}}\right)\right)}, \tag {5} +$$ + +where $\rho$ is the consistency measure with L1 norm. Since events are spatially sparse, we only establish connections with the dense trajectories at locations $\Omega$ with valid events. + +Image consistency with discrete trajectory. Similar to the unsupervised optical flow task [30], we can also establish the discrete consistency of the motion trajectory with the images at discrete times, to compensate for the spatial sparsity issue of ground-truth point tracks in the training data. In addition, in our sampling-based augmented training, skipped images can be used as additional continuity training objectives. + +For the accumulated continuous global trajectory $\mathbf{T}_{1\rightarrow t}$ we sample the the discrete optical flow $\mathbf{F}_{i\to j}$ from $I_{i}$ to $I_{j}$ via timestamps. Similar to Eq. (5), the consistency of images can be modeled as: + +$$ +L _ {i c} = \sum_ {\mathbf {x}} ^ {\Omega} \sum_ {i \neq j} ^ {N _ {v}} \rho \left(I _ {i} (\mathbf {x}), \operatorname {W a r p} \left(I _ {k} (\mathbf {x}); \mathbf {F} _ {i \rightarrow j} (\mathbf {x})\right)\right). \tag {6} +$$ + +Total objective. Our total training objective is a weighted combination of the above objectives, i.e., $L = L_{\text{traj}} + \lambda_1 L_{vis} + \lambda_2 L_{ec} + \lambda_3 L_{ic}$ , $\lambda$ are manual hyperparameters. Our ablations verify that combined self-supervised training can compensate for the temporal continuity of trajectories. + +![](images/20926011ecdde905bc7cfdce5b6eee5f09588e736a811fc2ef07a45180e95c12.jpg) +Input images with GT query points: frame 1, 7, 19, 27 + +![](images/2d45f50ff938260353577c4235090cb96a364dd5ef4e944a02738893e6efa3c9.jpg) + +![](images/27f2044ab6b3377ffd94a49d78365fcab712c97a7432fd01f61ac863c2748b2f.jpg) + +![](images/3315f716dc28c0314ebcaf0092f1fd62dd651191e4e9ad230ca31e7a28c2f3e9.jpg) + +![](images/9f5c2870b91081d75692e8bbcf34d2b15e60f3e5a64fb7a4fc4ab746c74358f2.jpg) +Init Coords +Figure 2. Visual comparisons of long-term dense point tracking on the TAP-Vid-DAVIS [5] dataset, with the ground-truth sparse query points with input images. Due to space limitations, we provide more sequences and event data visualization in the Supp. and demo video. + +![](images/e95ad5b036df431cddbe71bd95aa089fa07429b2f0e090853e734070c58597d3.jpg) +DOT: frame 7, 19, 27 + +![](images/f3532de9b9c32b6da9664495a0b985024862babb7e711754dfa432ce73dd9cd4.jpg) + +![](images/f6fa006b79904cf61a6bd22771e1d03cddfb44e1418c7c280f8720f6c48d85bb.jpg) + +![](images/c262e1b5c82bd005854405afb531dde6a1471dc41fb333acc787bb409b4f6326.jpg) +CoTracker: frame 7, 19, 27 + +![](images/585f9913958a1e636dcba20093ed6804f66fbe5c65d1f129f4a034c030f6fc4b.jpg) + +![](images/3a733254be8ee7ec3a82c13ab25d185818bb623f85a2753ac7e70728ec36775e.jpg) + +![](images/22ac09a90fd9fad4530145fdb9a3c6e9f9b7a0253fa59b1deaa87614d7db97c5.jpg) +Ours: frame 7, 19, 27 + +![](images/6eb6b97bc24766c478ce95dd317c02396badee7d7cfa26369d7dd45953033f3a.jpg) + +![](images/7c0e298d0dcdb6f1dcf88a53ec7709bbb9e24e792e5f6964308a32c5b0f159e0.jpg) + +Table 1. Quantitative results of dense evaluation on the CVO test [52] and extended set [35]. + +
MethodCVO (Clean)CVO (Final)CVO (Extended)
EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑
QueryPIPs++ [58]9.05 / 6.62 / 21.533.39.49 / 7.06 / 22.032.718.4 / 10.0 / 32.158.7
TAPIR [6]3.80 / 1.49 / 14.773.54.19 / 1.86 / 15.372.419.8 / 4.74 / 42.568.4
CoTracker [24]1.51 / 0.88 / 4.5775.51.52 / 0.93 / 4.3875.35.20 / 3.84 / 7.7070.4
DenseGMA [22]2.42 / 1.38 / 7.1460.52.57 / 1.52 / 7.2259.721.8 / 15.7 / 32.865.6
MFT [36]2.91 / 1.39 / 9.9319.43.16 / 1.56 / 10.319.521.4 / 9.20 / 41.837.6
AccFlow [52]1.69 / 1.08 / 4.7048.11.73 / 1.15 / 4.6347.536.7 / 28.1 / 52.936.5
DOT [35]1.32 / 0.74 / 4.1280.41.38 / 0.82 / 4.1080.25.07 / 3.67 / 7.3471.0
EDCPT (Ours)1.23 / 0.71 / 3.8382.11.31 / 0.76 / 3.8681.94.88 / 3.44 / 7.4671.9
+ +Table 2. Quantitative results of sparse evaluation on the TAP-VidDAVIS [5]point tracking benchmark. + +
MethodDAVIS (First)DAVIS (Strided)
AJ ↑<δxavg↑OA ↑AJ ↑<δxavg↑OA ↑
QueryTAP-Net [5]33.048.678.838.453.182.3
Context-PIPs [51]42.760.379.548.964.083.4
TAPIR [6]56.270.086.561.373.688.8
CoTracker [24]61.174.689.163.579.887.8
SpatialTracker [54]61.176.389.5---
DenseCPFlow [31]9.614.6----
MFT [36]47.366.877.856.170.886.9
DecoMotion [26]53.069.984.260.274.487.2
DinoTracker [46]---62.378.287.5
FlowTrack [4]---63.276.389.2
DOT [35]61.675.589.566.780.690.4
EDCPT (Ours)63.876.390.667.580.591.1
+ +# 4. Experiments + +# 4.1. Experimental details + +We train our model on MOVI-F [15] dataset and directly evaluate it on CVO [52] and TAP-Vid-DAVIS [5] benchmarks, using the vid2e [10] simulator to generate events.Following DOT [35], we train the model for 500k steps on $4\times$ NVIDIA L40 48G GPUs, using the Adam optimizer and OneCycle learning rate decay with a maximum of $10^{-4}$ . We choose 3 frames as training samples, along with the random selection of up to 10 frames in different frame intervals. The loss hyperparameters are set to 1.0, + +0.1, 0.1. For dense evaluation, we report the dense absolute error $\mathrm{EPE}_{\mathrm{all / vis / occ}}$ for all/visible/occluded points, as well as occlusion IoU accuracy OA for estimated visible mask. For sparse evaluation, we follow TAPNet [5] and report average Jaccard AJ, position accuracy $< \delta_{\mathrm{avg}}^x$ , and occlusion accuracy OA. Additionally, we adopt the real-captured event-based optical flow benchmark, DSEC [12, 13], to verify our adaptation capacity. More details are provided in the Supp. + +# 4.2. Standard spatially dense point tracking + +Performance on CVO and TAP-Vid-DAVIS benchmarks. We conduct comprehensive evaluations on two common standard point tracking benchmarks. Consistent with DOT [35], we report the quantitative results of the spatially dense optical flow from the last to the first frame of CVO [52] dataset in Tab. 1. Our proposed new EDCPT framework achieves significant performance improvements, whether comparing methods that only predict the partial Query points or directly estimating spatially Dense trajectories within a single inference. Particularly, we achieve 0.19 $\mathrm{EPE}_{\mathrm{all}}$ and 0.9 OA improvements on the extended set of 476 videos with 48 frames when compared to DOT [35]. + +In contrast, the real TAP-Vid-DAVIS dataset [5] only provides ground-truth trajectories for selected query points. As a result, we only sparsely evaluate these points for a fair comparison despite the output trajectories of our model and some compared methods are spatially dense. The + +![](images/c08b675dfba08ef858393f7b0a353e69d582a04423327d6ff1aa24e19c8627e2.jpg) +Figure 3. Visual comparisons of dense and continuous point trajectories on TAP-Vid-DAVIS [5]. Zoom in for detailed curve trajectories. + +![](images/765a1a10e4c12ae8e59dde07123a44f8fb8422a4be3c7a7e2844aa5761b84e3a.jpg) + +![](images/db3f7af69fd85254f5392812c74c36d4fe34112186ee69fc27f0443e98446d45.jpg) + +Table 3. Quantitative results on the real DSEC optical flow benchmark [13]. The first two methods are categorized as self-supervised learning and the remaining as supervised learning. + +
MethodInputEPE ↓AE ↓%Out ↓
Taming [39]Events2.3310.5617.77
MPCMax [18]Events3.208.5315.21
E-RAFT [13]Events0.792.852.68
TMA [28]Events0.742.682.30
IDNet [53]Events0.722.722.04
ECDDP [56]Events0.702.581.96
BFlow [14]Events0.752.682.44
BFlow [14]Images+Events0.692.421.88
EDCPT (Ours)Images+Events0.632.171.52
+ +quantitative results in Tab. 2 demonstrate the superiority of our framework, outperforming the existing state-of-the-art methods DOT [35] and SpatialTracker [54] with up to 2.7 AJ and 1.1 OA. We argue that ours $< \delta_{\mathrm{avg}}^{x}$ is close to DOT's because the sparse to dense upgrade strategy from sparse point matches is capable of modeling the rigid object motion and thus benefiting the proportion evaluation. We also perform qualitative visual comparisons in Fig. 2. Combining the above quantitative and qualitative comparisons with previous image-based methods, the new attempts of incorporating events by our framework significantly improve the accuracy with the standard dense point tracking setting. + +Performance on DSEC benchmark. We further conduct experiments on the DSEC benchmark [13] with real captured event data. Because the DSEC online leaderboard only measures the optical flow between two consecutive frames, we therefore finetune the local motion estimation on the DSEC training set from our pre-trained long-term model. The submission results are shown in Tab. 3. Notably, while BFlow [14] can estimate curve trajectories, they only submitted another optical flow version. Our framework fuses image and event data, yielding state-of-the-art performance with improvements of 0.05 EPE and 0.25 AE. + +# 4.3. Temporally continuous point tracking + +We adapt the above standard evaluation to input only a portion of the full video frames into the model to evaluate the temporal continuity with the ground truths of skipped + +frames. For the CVO final set with 7 frames per video, we skip 1 frame (half) and 2 frames (one-third) as model inputs, because the longer extended set lacks multi-frame ground-truth tracks. For the TAP-Vid-DAVIS dataset with $\sim 100$ frames, we report skip 1 frame (half) and skip 3 frames (quarter). The compared image-based methods cannot model inter-frame motion, thus we take linear motion interpolation to generate trajectory when frames are skipped. We retrain AccFlow [52] and mark it with * as its public version does not support forward point tracking. + +As reported in Tab. 4, our proposed new framework with global continuous trajectory accumulation significantly outperforms existing methods. Especially in nonlinear motion scenarios of TAP-Vid-DAVIS datasets, the larger frame intervals lead to greater performance gaps. In addition to the ablation of different motion assumptions in Tab. 6, the B-spline representation we adopt achieves better performance. We also provide visual comparisons in Fig. 3 and a demo video of continuous trajectory visualization in the Supp., that includes four sequences with diverse motion complexity and event sparsity from TAP-Vid-DAVIS and real-captured ERF-X170FPS [25] dataset. With extensive experiments on both simulated events (CVO and TAP-Vid-DAVIS) and real events (DSEC and ERF-X170FPS), we fully validate the robustness of the proposed global motion accumulation in modeling complex continuous trajectories. + +# 4.4. Ablation experiments and discussions + +To perform progressive ablations in Tab. 5, Tab. 6, and Tab. 7, the underlined settings are those utilized in the previous table, and the bolded ones represent the choices for our final framework. To validate the capability for continuous point tracking, the ablation results are reported with CVO third and DAVIS quarter settings as in Sec. 4.3 and Tab. 4. + +Global motion aggregation. One of our key contributions is the global aggregation of local motion representations in the streaming pipeline. In Tab. 5, N/A indicates do not explicitly model sequential motion as in DOT [35]. post is the forward aggregation in AccFlow [52], which performs post-processing only on the output optical flow and ignores the internal long-term motion information. We aggregate + +Table 4. Continuous point tracking evaluation results on the CVO extended set [35] and TAP-Vid-DAVIS dataset [5]. + +
MethodCVO (Final) - EPEall/vis/occ ↓DAVIS (First) - AJ / <δxavg↑
fullhalfthirdfullhalfquarter
RAFT2.09 / 0.81 / 8.022.44 / 0.95 / 9.073.02 / 1.16 / 11.4233.9 / 46.628.6 / 40.122.4 / 34.2
GMA1.99 / 0.77 / 7.572.35 / 0.89 / 8.452.92 / 1.09 / 10.9739.3 / 52.531.7 / 44.326.5 / 38.3
AccFlow*2.28 / 0.60 / 11.182.39 / 0.80 / 10.742.79 / 1.02 / 11.3747.2 / 62.337.5 / 49.230.9 / 42.4
CoTracker1.89 / 0.63 / 7.052.11 / 0.82 / 8.023.17 / 1.65 / 11.1361.1 / 74.654.3 / 68.848.9 / 63.9
DOT1.83 / 0.59 / 6.952.10 / 0.73 / 7.882.69 / 0.97 / 10.8461.6 / 75.555.6 / 70.150.4 / 65.3
EDCPT (Ours)1.76 / 0.55 / 6.731.97 / 0.66 / 7.612.16 / 0.73 / 8.7663.8 / 76.359.7 / 73.156.2 / 70.9
+ +Table 5. Ablations on global motion aggre- Table 6. Ablations on parametric curve rep- Table 7. Ablations on input data and supervi-gation. resentation. sion. + +
EPEall/vis/occ ↓AJ / <δxavg↑
N/A2.54 / 0.89 / 10.3351.9 / 66.4
post2.49 / 0.85 / 9.7352.5 / 66.9
solo2.45 / 0.82 / 9.8952.7 / 67.0
–offsets2.43 / 0.82 / 9.7353.0 / 67.2
stream2.42 / 0.80 / 9.6853.3 / 67.4
+ +
EPEall/vis/occ ↓AJ / <δxavg↑
linear2.42 / 0.80 / 9.6853.3 / 67.4
quad2.49 / 0.84 / 9.4654.2 / 68.1
Nc=32.32 / 0.79 / 9.3354.4 / 68.6
Nc=42.23 / 0.76 / 8.9755.4 / 69.7
Nc=52.26 / 0.75 / 9.2055.0 / 69.3
+ +
EPEall/vis/occ ↓AJ / <δxavg ↑
Images2.50 / 0.86 / 9.9252.5 / 66.7
+Events2.23 / 0.76 / 8.9755.4 / 69.7
+Lic2.20 / 0.75 / 8.9055.8 / 70.2
+Lec2.16 / 0.73 / 8.7656.2 / 70.9
+ +motion representations at the feature level instead of dealing directly with motion vectors. solo is the short and long term fusion module in SOLOFusion [40]. Unlike its temporal fusion with a fixed number of frames and low resolution, we adopt sequential modeling with an unspecified number of frames in a streaming pipeline. Our proposed motion aggregation framework fuses image correspondence and event features from local to global with recurrent streaming structure, which better preserves multi-frame motion information and achieves optimal performance. In addition, we also verified that removing the additional offset estimation to address the numerical problem in Warping leads to a slight performance degradation, as this would require the subsequent refinement to handle it simultaneously. + +Curve representation. Based on our streaming framework, we compare the performance improvement of taking the B-spline curve representation compared to interpolating with linear and quadratic motion assumptions in Tab. 6. For the control points number of B-spline curves, we chose $N_{c} = 4$ , since further increasing the number of control points does not significantly improve the performance under CVO third and DAVIS quarter settings. + +Input data and supervision. Since previous methods usually use only image data, we verify the advantages of incorporating event data for high-precision continuous point tracking by removing events, as depicted in Tab. 7. Moreover, the comparison results between this image-only variant and DOT [35] in Tab. 4 demonstrates that our proposed streaming aggregation and curve representation are beneficial even in the absence of event data. Furthermore, we validate that training using the proposed image and event-to-point trajectory consistencies as additional supervision complements the lack of continuous inter-frame tracks in + +the training data and can further improve performance. + +Limitations. On the same RTX 3090 PC, our framework processes a 48-frame video at $512 \times 512$ resolution in 12.6 seconds, significantly faster than CoTracker's 11 minutes, while only processing partial query points in a single run. However, it is slower than the two-frame optical flow method GMA of only 2.1 seconds and slightly slower than the multi-frame point tracking method DOT of 9.5 seconds. In addition, there is currently a lack of point-tracking datasets with real events, and it is difficult to obtain long-term tracking annotations. We evaluate on standard benchmarks with simulated events, as well as visualize point tracking on real events, and also evaluate optical flow estimation on a real event benchmark. Our future work plans to improve on model efficiency and the evaluation setup. + +# 5. Conclusion + +In this paper, we propose a new long-term dense point tracking framework for integrating image and event data to estimate continuous motion trajectories. Specifically, we streamingly process two adjacent frames and inter-frame events to generate local motion representations, and accumulate previous caches to global motion representation at the feature level to produce new global trajectories at the trajectory level. We utilize multi-frame parametric curve accumulation to represent continuous trajectories with any number of frames, complemented by image and event-to-trajectory consistencies for model training. We validate the effectiveness of our framework by conducting extensive benchmark experiments as well as qualitative verification on real challenge data. We believe this work provides new insights into the point tracking task from the perspective of event modality and continuous global curve representations. + +# Acknowledgments + +This research/project is supported by the National Natural Science Foundation of China (62271410, 12150007), the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008), and the Tier 1 grant T1-251RES2305 from the Singapore Ministry of Education. Zhexiong Wan was also supported by the Program of China Scholarship Council (202306290193), and the Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University (CX2023013). + +# References + +[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scalespace flow for end-to-end optimized video compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1 +[2] Adam Bielski and Paolo Favaro. Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems (NeurIPS), 35:33371-33386, 2022. 1 +[3] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 4 +[4] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268–19277, 2024. 1, 2, 3, 6 +[5] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re-casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, pages 13610–13626, 2022. 1, 2, 3, 6, 7, 8 +[6] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10061-10072, 2023. 2, 6 +[7] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. In Asian Conference on Computer Vision (ACCV), 2024. 2 +[8] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 4, 5 +[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and + +Davide Scaramuzza. Event-based vision: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):154-180, 2022. 1 +[10] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3586-3595, 2020. 6 +[11] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision (IJCV), 128(3):601-618, 2020. 2 +[12] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters (RA-L), 2021. 6 +[13] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 2, 6, 7 +[14] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 46(7):4736-4746, 2024. 1, 3, 7 +[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3749-3761, 2022. 6 +[16] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16022-16033, 2023. 1 +[17] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 7167-7179, 2021. 1 +[18] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In European Conference on Computer Vision (ECCV), 2024. 3, 7 +[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022. 1, 2 +[20] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. Progressive spatio-temporal alignment for efficient event-based motion estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1537-1546, 2023. 1 +[21] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. FlowNet 2.0: Evolution of optical flow estimation with deep networks. + +In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 1 +[22] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9772-9781, 2021. 4, 6 +[23] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2023. 1 +[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision (ECCV), 2024. 2, 6 +[25] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 7 +[26] Rui Li and Dong Liu. Decomposition betters tracking everything everywhere. In European Conference on Computer Vision (ECCV), 2024. 2, 6 +[27] Siqi Li, Zhikuan Zhou, Zhou Xue, Yipeng Li, Shaoyi Du, and Yue Gao. 3d feature tracking via event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18974-18983, 2024. 2 +[28] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9685-9694, 2023. 7 +[29] Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate. arXiv preprint arXiv:2409.11953, 2024. 3 +[30] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6489-6498, 2020. 5 +[31] Jianqin Luo, Zhexiong Wan, Yuxin Mao, Bo Li, and Yuchao Dai. Continuous parametric optical flow. In Advances in Neural Information Processing Systems (NeurIPS), pages 23520-23532, 2023. 1, 2, 3, 6 +[32] Xinglong Luo, Ao Luo, Zhengning Wang, Chunyu Lin, Bing Zeng, and Shuaicheng Liu. Efficient meshflow and optical flow estimation from event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19198-19207, 2024. 2 +[33] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5642-5651, 2023. 1, 2 + +[34] Etienne Meunier, Anais Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4):4462-4473, 2023. 1 +[35] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 8 +[36] Michal Neoral, Jonás Šerzych, and Jérí Matas. Mft: Long-term tracking of every pixel. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 6837-6847, 2024. 2, 3, 6 +[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (TMLR), 2024. 2 +[38] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1669-1678, 2020. 1 +[39] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9695-9705, 2023. 7 +[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris M. Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multiview 3d object detection. In International Conference on Learning Representations (ICLR), 2023. 8 +[41] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 4 +[42] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision (ECCV), 2024. 2 +[43] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. CRAFT: Cross-attentional flow transformer for robust optical flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17602-17611, 2022. 4 +[44] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419, 2020. 1, 4 +[45] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time Lens: Event-based video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, 2021. 1 + +[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 6 +[47] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing (TIP), 31:7237-7251, 2022. 2 +[48] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. RPEFlow: Multimodal fusion of RGB-pointcloud-event for joint optical flow and scene flow estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10030-10040, 2023. 2 +[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19795-19806, 2023. 2 +[50] Xiangyuan Wang, Huai Yu, Lei Yu, Wen Yang, and GuiSong Xia. Towards robust keypoint detection and tracking: A fusion approach with event-aligned image features. IEEE Robotics and Automation Letters (RA-L), 9(9):8059-8066, 2024. 2 +[51] BIAN Weikang, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands spatial context features. In Advances in Neural Information Processing Systems (NeurIPS), pages 55285-55298, 2023. 2, 6 +[52] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: backward accumulation for long-range optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12119-12128, 2023. 2, 3, 6, 7 +[53] Yilun Wu, Federico Paredes-Vallés, and Guido CHE De Croon. Lightweight event-based optical flow estimation via iterative deblurring. In IEEE International Conference on Robotics and Automation (ICRA), pages 14708-14715. IEEE, 2024. 7 +[54] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7 +[55] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In Advances in Neural Information Processing Systems (NeurIPS), pages 1647-1656, 2019. 1 +[56] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data dense pre-training. In European Conference on Computer Vision (ECCV), pages 292-310. Springer, 2024. 7 +[57] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9781-9790, 2023. 1 +[58] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale + +synthetic dataset for long-term point tracking. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19855-19865, 2023. 2, 6 +[59] Hanyu Zhou, Yi Chang, Haoyue Liu, Wending Yan, Yuxing Duan, Zhiwei Shi, and Luxin Yan. Exploring the common appearance-boundary adaptation for nighttime optical flow. In International Conference on Learning Representations (ICLR), 2024. 2 +[60] Hanyu Zhou, Yi Chang, and Zhiwei Shi. Bring event into rgb and lidar: Hierarchical visual-motion fusion for scene flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26477-26486, 2024. 2 +[61] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 34(8):4868-4880, 2021. 1 +[62] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 1, 2 +[63] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for RGB-event transformer-trackers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 22045-22055, 2023. 1 \ No newline at end of file diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/images.zip b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..1d527a1bcc7c58dfdce19170342e53caa20bba9c --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:1358c2e986bda2da35cc4733e9475461abdeddef2e75d7c8015e10a3d80db174 +size 625640 diff --git a/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/layout.json b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..8dcc540842c86ad0e188f614191ba1a1e34a774d --- /dev/null +++ b/2025/Event-aided Dense and Continuous Point Tracking_ Everywhere and Anytime/layout.json @@ -0,0 +1,9777 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 545, + 121 + ], + "type": "text", + "content": "Event-aided Dense and Continuous Point Tracking: Everywhere and Anytime" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "spans": [ + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "text", + "content": "Zhexiong Wan" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "text", + "content": " Jianqin Luo" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "text", + "content": " Yuchao Dai" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "text", + "content": " Gim Hee Lee" + }, + { + "bbox": [ + 144, + 142, + 465, + 157 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 107, + 157, + 502, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 157, + 502, + 185 + ], + "spans": [ + { + "bbox": [ + 107, + 157, + 502, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 157, + 502, + 185 + ], + "type": "text", + "content": "School of Electronics and Information, Northwestern Polytechnical University & Shaanxi Key Laboratory of Information Acquisition and Processing" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 139, + 185, + 471, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 139, + 185, + 471, + 200 + ], + "spans": [ + { + "bbox": [ + 139, + 185, + 471, + 200 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 139, + 185, + 471, + 200 + ], + "type": "text", + "content": "Department of Computer Science, National University of Singapore" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 66, + 201, + 541, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 201, + 541, + 213 + ], + "spans": [ + { + "bbox": [ + 66, + 201, + 541, + 213 + ], + "type": "text", + "content": "{wanzhexiong,luojianqin}@mail.nwpu.edu.cn，daiyuchao@nwpu.edu.cn，gimhee.lee@nus.edu.sg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "spans": [ + { + "bbox": [ + 151, + 241, + 200, + 253 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 266, + 297, + 542 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 266, + 297, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 266, + 297, + 542 + ], + "type": "text", + "content": "Recent point tracking methods have made great strides in recovering the trajectories of any point (especially key points) in long video sequences associated with large motions. However, the spatial and temporal granularities of point trajectories remain constrained by limited motion estimation accuracy and video frame rate. Leveraging the high temporal resolution and motion sensitivity of event cameras, we introduce event data for the first time to recover spatially dense and temporally continuous trajectories of every point at any time. Specifically, we define the dense and continuous point trajectory representation as estimating multiple control points of curves for each pixel and model the movement of sparse events triggered along continuous point trajectories. Building on this, we propose a novel multi-frame iterative streaming framework that first estimates local inter-frame motion representations from two consecutive frames with inter-frame events, then aggregates them into a global long-term motion representation to utilize input full video and event data with an arbitrary number of frames. Extensive experiments on simulated and real data demonstrate the significant improvement of our framework over state-of-the-art methods and the crucial role of introducing events to model continuous point trajectories." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 565, + 135, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 565, + 135, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 565, + 135, + 578 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 296, + 694 + ], + "type": "text", + "content": "Estimating fine-grained motion from input videos is a crucial task in computer vision with widespread applications such as video compression [1], frame interpolation [23, 55], motion segmentation [2, 34], and dynamic scene reconstruction [16]. Early studies [21, 44] based on two-frame optical flow can model the spatially dense motion, but they suffer from the challenge of modeling long-term motion from more frames. With the proposal of tracking any point (TAP) task [5, 19], using sparse points as query indexes to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 397 + ], + "type": "text", + "content": "estimate pointwise long-term motions also draws attention. Despite significant progress, the sparse and independent representation of pointwise trajectories remains inherently incompatible with the spatially dense representation of input video. Thus, recently there are methods [4, 35, 46] that continue to use dense optical flow to represent long-term motion. However, the temporal frame rate of the input video is constrained by traditional shutter cameras, which makes accurately modeling complex motion trajectories difficult for video-only methods from the data acquisition perspective. Consequently, accurate estimation of fine-grained spatially dense and temporally continuous motion remains a challenging and worthwhile research problem." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 400, + 556, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 556, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 556, + 615 + ], + "type": "text", + "content": "Unlike traditional shutter cameras that expose the entire image at fixed frame rates, the new bio-inspired event cameras [9] can independently and asynchronously detect per-pixel brightness changes with microsecond precision. This unique design makes it inherently sensitive to visual motion changes, leading to impressive motion-related applications such as optical flow estimation [17, 62], motion segmentation [20, 61], feature tracking [33], object tracking [63], and frame interpolation [45]. However, events are typically triggered only with motion contours or rich textures, making it challenging to comprehensively perceive spatially dense motion. Recently, integrating the advantages of event camera and traditional shutter camera has become a new direction [38, 57]. Leveraging on these successes, we propose for the first time taking event data as an auxiliary input for the point tracking task, where we comprehensively model fine-grained spatially dense and temporally continuous motion from the input image and event modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 715 + ], + "type": "text", + "content": "The introduction of event cameras offers the potential to model continuous motion from the data perspective. However, to effectively model continuous long-term motion, we need a new representation that replaces optical flow and parametrically associates the dense temporal properties of event data. BFlow [14] proposes to learn trajectories with Bezier curves from events, but is limited to input fixed frames and cannot adapt to longer sequences. CPFlow [31]" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 703, + 261, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 703, + 261, + 713 + ], + "spans": [ + { + "bbox": [ + 69, + 703, + 261, + 713 + ], + "type": "text", + "content": "Project page: https://npuctr.github.io/EDCPT" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7936" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 251 + ], + "type": "text", + "content": "proposes to learn control points represented as B-spline curves from a fixed number of image slices. Although continuous motion can be successfully modeled using these curve representations in normalized timescales, their fixed number of control points made it hard to handle complex dynamics and varying lengths of video sequences. Based on the curve representation, we propose a new streaming pipeline for accumulating multiple local curves to address these limitations. In addition, how to adequately fuse multi-frame temporal information to achieve long-term tracking is also a concern. Compared to not explicitly associating multi-frames in DOT [35] and aggregating over estimated motion vectors in AccFlow [52] and FlowTrack [4], we perform recurrent fusion in the feature space, dedicated to preserving multi-frame sequence information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 253, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 296, + 468 + ], + "type": "text", + "content": "In this paper, we present the first event-aided point tracking framework for recovering spatially dense and temporally continuous point trajectories from input videos and event sequences. Specifically, we first propose a new point trajectory representation with parametric curves that accumulate multiple local curves to adapt to multi-frame input videos at any length. We then design a new framework for combining two frames with events to estimate dense point curve trajectories and extend to multi-frame iterative accumulation. In addition, we establish the association between continuous curve trajectory and event triggering as a part of the learning objective for continuous motion modeling. Extensive experiments on both simulated and real-world data demonstrate that the proposed framework significantly outperforms state-of-the-art methods. Particularly, our ablation studies illustrate the effectiveness of the proposed global aggregation and highlight the crucial role of incorporating event data in continuous trajectory modeling." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 469, + 275, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 469, + 275, + 480 + ], + "spans": [ + { + "bbox": [ + 67, + 469, + 275, + 480 + ], + "type": "text", + "content": "Our main contributions are summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 482, + 296, + 637 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 67, + 482, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 482, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 67, + 482, + 295, + 529 + ], + "type": "text", + "content": "- We introduce a new setup that, for the first time, enables long-term spatially dense and temporally continuous point tracking by integrating the strengths of both images and events." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 529, + 295, + 577 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 529, + 295, + 577 + ], + "spans": [ + { + "bbox": [ + 67, + 529, + 295, + 577 + ], + "type": "text", + "content": "- We present a novel global curve representation of continuous point trajectories through multi-frame aggregation, establishing a connection between event triggering and continuous motion." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 67, + 578, + 296, + 637 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 296, + 637 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 296, + 637 + ], + "type": "text", + "content": "- We propose a novel event-aided streaming framework that iteratively accumulates the local tracks from two frames with inter-frame events, resulting in global, long-term dense and continuous trajectories through iterative temporal global motion aggregation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 651, + 147, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 651, + 147, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 651, + 147, + 662 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 671, + 237, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 237, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 237, + 685 + ], + "type": "text", + "content": "2.1. Image-based query point tracking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "The goal of point tracking is to recover the corresponding positions of query points in each frame, which has attracted" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 277 + ], + "type": "text", + "content": "wide attention with the proposal of the TAP benchmark and the baseline model TAPNet [5]. PIPs [19] proposes to extract independent point representation for 8-frame tracking, then " + }, + { + "bbox": [ + 313, + 72, + 553, + 277 + ], + "type": "inline_equation", + "content": "\\mathrm{PIPs}++" + }, + { + "bbox": [ + 313, + 72, + 553, + 277 + ], + "type": "text", + "content": " [58] extends to long-term trajectories. TAPIR [6] proposes a two-stage matching framework that fuses TAPNet and PIPs, and BootsTAPIR [7] performs self-supervised training on more data. Unlike these methods that track only one query at a time, CoTracker [24] and Context-PIPs [51] use additional tracks and pixel featuresto improve global tracking. SpatialTracker [54] introduces the triplane representation for group pixels in 3D.DINOTracker [46] performs test-time finetuning on the pre-trained DINOViT [37] model and achieves fine-grained tracking per video. When applying these query points-based methods to achieve dense tracking, full points need to be processed individually or in batches, which brings computational hurdles and limits their downstream applications [35]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 285, + 495, + 298 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 495, + 298 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 495, + 298 + ], + "type": "text", + "content": "2.2. Image-based dense point tracking" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 303, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 555 + ], + "type": "text", + "content": "Recent studies turn to tracking every point within a frame in a single run, aiming to enhance neighborhood relationships while reducing computational requirements. OmniMotion [49] performs pixel-wise tracking between local and canonical space to maintain the global consistency of the motion, and then FastOmniTrack [42] and DecoMotion [26] improve from the perspectives of computation cost and object motion decomposition. CPFlow [31] proposes to estimate spatio-temporally dense motion curves, but it can only input 4 images and needs pre-sampling for longer video. AccFlow [52] proposes the forward and backward aggregation pipeline, extending inter-frame dense optical flow to multi-frame long ranges. MFT [36] select chaining multiframe candidates and FlowTrack [4] automatically apply error compensation in instances of tracking inaccuracies. DOT [35] unifies point tracking and optical flow, upgrading sparse tracks to dense flow fields between every frame. However, limited by the video frame rate, these methods are struggling to model challenging dynamics. In this work, we propose to introduce continuous event data into the input video to enable temporally continuous point tracking." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 563, + 481, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 481, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 481, + 575 + ], + "type": "text", + "content": "2.3. Event-based motion estimation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "Thanks to the motion-sensitive nature of event cameras, extensive motion estimation studies in recent years have highlighted their potential applications to challenging dynamics. The feature tracking methods [11, 27, 33, 50] show the benefits of event cameras for low-latency tracking, but can only track sparse, specific textured locations. Recently estimating optical flow from events has become mainstream. Using only sparse event data [13, 32, 62] allows to estimate satisfactory dense optical flow, while introducing data from other sensors such as images [47, 59] and point clouds [48, 60] achieves significant performance" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7937" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "gains. BFlow [14] and MotionPriorCMax [18] exploit the continuous property of events to estimate parametric Bezier trajectories, but can only estimate motion within a fixed consecutive frame interval and cannot be directly adapted to long-term sequences. Recently, FE-TAP [29] proposes to recover point trajectories from a fixed number of images and events based on TAPVid [5], but does not take full advantage of the continuous nature of events. We propose to combine the advantages of images and events to enable temporally continuous point tracking by modeling long-term global motion with an arbitrary number of frames." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 222, + 111, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 111, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 111, + 235 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 244, + 296, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 296, + 353 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 296, + 353 + ], + "type": "text", + "content": "Overview. To address the limitations of sparse events and low video frame rates in modeling fine-grained motion, we present the first framework that recovers dense and continuous point trajectories from a video with corresponding event sequences. Our framework consists of four parts: 1) A parametric multi-frame continuous point trajectories representation; 2) A model of event triggering along the point trajectories; 3) A two-frame basis motion estimation model; 4) A multi-frame motion aggregation and streaming framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": "Problem Formulation. Conventional shutter camera captures a video with " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " frames " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\{I_i\\}_{i=1}^{N_v}" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " at a fixed frame rate. Event camera triggers an unbounded event sequence " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_{i=1}^{N_e}" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "N_e" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " is the number of events. Each event " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "e_i = \\{\\mathbf{x}_i, t_i, p_i\\}" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " consists of the pixel position " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i = (x,y)" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": ", timestamp " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " with microsecond precision, and the brightness change polarity " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " in log domain. Our goal is to combine these two modalities to recover the spatially dense and temporally continuous trajectories " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{1 \\to t}" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": " of all points starting from the first frame at time 1 to last frame at time " + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 355, + 296, + 475 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 489, + 144, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 489, + 144, + 501 + ], + "spans": [ + { + "bbox": [ + 55, + 489, + 144, + 501 + ], + "type": "text", + "content": "3.1. Motion model" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": "Trajectory representation. Previous point tracking methods typically estimate two-channel motion vectors in the XY directions, which is the optical flow when representing dense trajectories [4, 35]. To learn the curve trajectory from the deep network, we instead learn the multiple control points as curved trajectories [31]. Specifically, we choose the B-spline curve as our curve representation, which is defined by " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " control points " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{P}_i\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " and basis functions " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\{B_{i,p}(t)\\}_{i=1}^{N_c}" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " with degree " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": ". The continuous trajectory " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t)" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " represented by b-spline curve in time variable " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " is a collection of piecewise polynomial functions " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}(t) = \\sum_{i=1}^{N_c} B_{i,p}(t) \\mathbf{P}_i" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": ". Similar to learning dense flow, each pixel has an independent curve with control points denoted as " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbb{P} \\in \\mathbb{R}^{N_c \\times 2 \\times H \\times W}" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "H \\times W" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " is the image size. This enables learnable motion modeling of dense and continuous point trajectories " + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 509, + 295, + 714 + ], + "type": "text", + "content": " with parametric curve representation. More details are provided in the Supp." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 180 + ], + "type": "text", + "content": "Multi-frame global trajectories accumulation. Existing parametric motion modeling methods are fixed in the number of frames they can handle, e.g., BFlow [14] is limited to between two frames, and CPFlow [31] struggles to benefit from more than 4 frame inputs, resulting in suboptimal long-term trajectory modeling. Inspired by the practice of multi-frame optical flow accumulation [36, 52], we propose a new multi-frame curve trajectories accumulation strategy to handle long-term videos with arbitrary frames." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": "Our accumulation framework works on a streaming pipeline, where the previous global trajectory " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{1\\rightarrow t}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "(t - 1)\\times N_{c}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " control points has been accumulated from the previous " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " local trajectories " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " when processing the " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " -th step. When we get updated " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " -th local trajectory " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t\\rightarrow t + 1}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " control points from time " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ", a simple approach is to directly accumulate the new global trajectory with " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t\\times N_{c}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " control points from time 1 to " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{1\\rightarrow t + 1}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathrm{Warp}(\\mathbf{T}_{t\\rightarrow t + 1},\\mathbf{T}_{1\\rightarrow t})(\\mathbf{x})\\right]" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ", where Warp represents backward warping, the operator [,] combines the control points of two sub-curves and creates a more complex curve. However, there are two problems with just warping operation: 1) It suffers from numerical error as integer sampling with floating-point coordinates, i.e., for warping vectors from " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{a}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathrm{Warp}(\\mathbf{a},\\mathbf{b})(\\mathbf{x}) = \\mathbf{a}(\\mathbf{x} + \\mathbf{b}(\\mathbf{x}))" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": " is integer coordinates but not " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{x} + \\mathbf{b}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ". 2) Some points may be occluded at time " + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 184, + 555, + 399 + ], + "type": "text", + "content": ", resulting in failing to find the corresponding points. More details are provided in the Supp." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": "Our framework iteratively maintains and learns to integrate from a global motion representation " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1\\rightarrow t}^{global}" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " in the streaming process (cf. Sec. 3.2). For the first numerical problem, we estimate a start point offsets " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{O}_t\\in \\mathbb{R}^{2\\times H\\times W}" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " learned from " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1\\rightarrow t}^{global}" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " and normalized to the range " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "[-1,1]" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": ". Sampling compensation is achieved by adding this offset directly during warping. For the second occlusion problem, we introduce an occlusion solving strategy for occluded pixels. We additionally estimate the visibility map " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_{1\\rightarrow t}" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " of each point from the initial frame to the " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": "-th frame as well as the trajectory update " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_t" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": ". Aggregation is based on a warp with offset when the point " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " is visible. When point " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " is occluded, a learnable module Fusion is introduced to regress the point's coarse motion trajectory in " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "t\\to t + 1" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{1\\rightarrow t}^{global}" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": ". Finally, the trajectory update " + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{T}_t" + }, + { + "bbox": [ + 313, + 403, + 556, + 606 + ], + "type": "text", + "content": " as a uniformly refinement serves the global trajectory accumulation goal. The refinement process can be modeled as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 623, + 556, + 666 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 623, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 312, + 623, + 556, + 666 + ], + "type": "interline_equation", + "content": "\\mathbf {T} ^ {\\prime} = \\left\\{\\begin{array}{l l}\\operatorname {W a r p} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {O} _ {t}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 1,\\\\\\operatorname {F u s i o n} \\left(\\mathbf {T} _ {t \\rightarrow t + 1}, \\mathbf {T} _ {1 \\rightarrow t}, \\mathbf {M} _ {1 \\rightarrow t} ^ {\\text {g l o b a l}}\\right) + \\Delta \\mathbf {T} _ {t}&\\text {i f} \\mathbf {V} _ {1 \\rightarrow t} (\\mathbf {x}) = 0,\\end{array}\\right. \\tag {1}", + "image_path": "a13ed3ec45d2c19c53c15e613b4dc5cb9dbabd1148e5690668a5bad8f9670644.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}'" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " is the local curve in " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "t\\to t + 1" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": " warped to the starting coordinates at time 1. Then we direct combine it with the cached previous global curve to obtain the accumulated current global curve " + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{1\\rightarrow t + 1}^{accum}(\\mathbf{x}) = \\left[\\mathbf{T}_{1\\rightarrow t}(\\mathbf{x}),\\mathbf{T}'(\\mathbf{x})\\right]" + }, + { + "bbox": [ + 313, + 666, + 555, + 715 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7938" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 527, + 277 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 527, + 277 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 527, + 277 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 527, + 277 + ], + "type": "image", + "image_path": "dd78a4ad66fb8187085af31e2b7a244c880e9f25cb1dc399fc61216c6b96771e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "type": "text", + "content": "Figure 1. Our proposed event-aided dense and continuous point tracking framework consists of two main steps. 1) Local motion estimation: estimating short-term curve trajectories with " + }, + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "type": "text", + "content": " control points from two consecutive images and inter-frame events, while concurrently updating the local motion representation. 2) Global motion accumulation: iteratively fusing the latest local motion representation with the previous global motion representation in a streaming manner for aggregating the latest global motion representation. Subsequently, the global long-term curve trajectories with " + }, + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "type": "inline_equation", + "content": "t \\times N_{c}" + }, + { + "bbox": [ + 54, + 279, + 555, + 336 + ], + "type": "text", + "content": " control points are optimized on trajectory combinations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": "Events along the trajectory. Following the contrast maximization framework [8], we assume that events are triggered along with the pixel motion trajectories at the moving boundary. For a motion trajectory " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": " starting from pixel " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": ", the generated event's coordinates satisfy the trajectory, i.e., " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1 = \\mathbf{T}(t_1), \\mathbf{x}_i - \\mathbf{x}_1 = \\mathbf{T}(t_i) - \\mathbf{T}(t_1)" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": ". We can thus use the motion trajectory to transform the events " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "e_i \\dot{=} \\{\\mathbf{x}_i, t_i, p_i\\}" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": " back to time " + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "inline_equation", + "content": "t_1" + }, + { + "bbox": [ + 54, + 347, + 294, + 443 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 69, + 451, + 295, + 465 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 451, + 295, + 465 + ], + "spans": [ + { + "bbox": [ + 69, + 451, + 295, + 465 + ], + "type": "interline_equation", + "content": "e _ {i} ^ {\\prime} \\doteq \\left\\{\\mathbf {x} _ {i} ^ {\\prime} = \\operatorname {W a r p} \\left(\\mathbf {x} _ {i}; \\mathbf {T} (t _ {i}) - \\mathbf {T} (t _ {1})\\right), t _ {1}, p _ {i} \\right\\}. \\tag {2}", + "image_path": "c1d4b5711a953a1905dcee7de2f25d046084e623748e2a748fe8fce0c6c311fe.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": "Assuming the trajectory " + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{T}" + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": " is accurate, this process transforms the event " + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "inline_equation", + "content": "e_i" + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": " to the starting point position " + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1" + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": " of the trajectory, i.e., " + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_1 = \\mathrm{Warp}(\\mathbf{x}_1; \\mathbf{T}(t_i - t_1))" + }, + { + "bbox": [ + 55, + 473, + 296, + 558 + ], + "type": "text", + "content": ". Based on the correlated motion modeling of events and point trajectories, we build additional self-supervised training objectives in Sec. 3.3 to alleviate the lack of continuous trajectory annotations in the training datasets." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 564, + 134, + 575 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 134, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 134, + 575 + ], + "type": "text", + "content": "3.2. Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Two-frame basis model. The two-frame basis model is designed to recover inter-frame short-term trajectories " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t\\rightarrow t + 1}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " from the encoded features of input two consecutive frames " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "F_{t},F_{t + 1}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " and inter-frame events " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "E_{t\\rightarrow t + 1}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ". In the feature extraction phase, we first convert the raw event data into a dense grid representation [41], followed by two feature encoders for two images and an event grid. Subsequently, we construct the local correlation between two frame features by matrix multiplication [44]. By leveraging the local correlation and events, we learn the local motion representation " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "M_{t\\rightarrow t + 1}^{local}" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " by a motion extractor which allows" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "content": "recovery of local dense trajectories " + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{t\\rightarrow t + 1}" + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "content": " by a trajectory decoder. Specifically, the former utilizes a transformer fusion module in GMA [22], while the latter is based on the typical FPN decoder in RAFT [43]. The trajectory decoder estimates the coordinates of " + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "content": " control points " + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\mathbb{P}_{t\\to t + 1}" + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "content": " and a single-channel visibility map " + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{V}\\}_{t\\to t + 1}" + }, + { + "bbox": [ + 313, + 347, + 555, + 431 + ], + "type": "text", + "content": ", which is essential for establishing multi-frame global accumulation in Eq. (1)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": "Global motion aggregation module. For processing a video comprising " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "N_{v}" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": " frames, the above two-frame basis model needs to be streamed sequentially " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "N_{v} - 1" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": " times yielding multiple local motion representations " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{M}_{i\\rightarrow i + 1}^{local}\\}_{i = 1}^{t - 1}" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": " and curve trajectories " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\{\\mathbf{T}_{i\\rightarrow i + 1}\\}_{i = 1}^{t - 1}" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": ". To facilitate the accumulation of global multi-frame trajectories according to Sec. 3.1, the cached global motion representation " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{t - 1}\\doteq \\mathbf{M}_{1\\to t}^{global}" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": " from the previous " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": "-frames is utilized as the query, while the current local motion representation " + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^l\\doteq \\mathbf{M}_{t\\to t + 1}^{local}" + }, + { + "bbox": [ + 313, + 449, + 556, + 582 + ], + "type": "text", + "content": " serves as the key and value. We first perform the linear projections and compute the cross-attention:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 332, + 601, + 553, + 656 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 601, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 332, + 601, + 553, + 656 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\operatorname {C A} \\left(\\mathbf {h} _ {t - 1}, \\mathbf {h} _ {t} ^ {l}, W _ {Q, K, V}\\right) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {T}}{\\sqrt {d _ {k}}}\\right) V \\tag {3} \\\\ = \\operatorname {s o f t m a x} \\left(\\frac {\\left(W _ {Q} \\cdot \\mathbf {h} _ {t - 1}\\right) \\left(W _ {K} \\cdot \\mathbf {h} _ {t} ^ {l}\\right) ^ {T}}{\\sqrt {d _ {k}}}\\right) \\left(W _ {V} \\cdot \\mathbf {h} _ {t} ^ {l}\\right), \\\\ \\end{array}", + "image_path": "6b41db190392adf59451c861310ce03a8881e31360ce967a76406b66644c6fde.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "d_{k}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the channel size, " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\cdot" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " is the linear projection and " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "W_{Q,K,V}" + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": " are the corresponding weights. We then conduct iterative fusion based on the gated activation unit (GRU) [3], where the update gate is " + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{z}_t =" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7939" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": "sigmoid(CA(" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^l" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "W_{Q,K,V}" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": "), the reset gate is " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{r}_t =" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": " sigmoid(CA(" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{h}_t^l" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "W_{Q,K,V}^\\prime" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": "), and the hidden state is " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\mathbf{s}_t = \\tanh(\\mathrm{CA}(\\mathbf{r}_t \\odot \\mathbf{h}_{t-1}, \\mathbf{h}_t^l, W_{Q,K,V}^{\\prime\\prime}))" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": " is the elementwise multiplication. The superscript of " + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "inline_equation", + "content": "W_{Q,K,V}" + }, + { + "bbox": [ + 55, + 72, + 296, + 159 + ], + "type": "text", + "content": " denotes the different projection weights taken independently in each attention calculation. Finally, we iteratively update the current global motion representation at the feature level by:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 82, + 167, + 295, + 184 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 167, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 82, + 167, + 295, + 184 + ], + "type": "interline_equation", + "content": "\\mathbf {M} _ {1 \\rightarrow t + 1} ^ {\\text {g l o b a l}} \\doteq \\mathbf {h} _ {t} = \\left(1 - \\mathbf {z} _ {t}\\right) \\odot \\mathbf {h} _ {t - 1} + \\mathbf {z} _ {t} \\odot \\mathbf {s} _ {t}. \\quad (4)", + "image_path": "dda0b5c8781a4c4a081caff1e8ba76727a04d690124845b857a16fdbcdce8014.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 195, + 296, + 244 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 195, + 296, + 244 + ], + "spans": [ + { + "bbox": [ + 55, + 195, + 296, + 244 + ], + "type": "text", + "content": "The simple and effective recurrent temporal aggregation we take is naturally compatible with the streaming pipeline, and we verify its effectiveness compared to previous fusion solutions in ablation experiments (cf. Sec. 4.4)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 265, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 445 + ], + "type": "text", + "content": "Multi-frame iterative streaming framework. As depicted in Fig. 1, our framework iteratively processes the input video and event data through local motion estimation and global motion accumulation. We aggregate the local motion representations from each frame interval to the global motion representation at the feature level through the above global aggregation module in Sec. 3.2. Subsequently, the multi-frame accumulation step described in Sec. 3.1 sequentially combines each inter-frame short-term curve into a global long-term motion trajectory at the trajectory level, providing the dense and continuous point tracking representation as the model output. On the right side of Fig. 1, local motion is visualized with dense optical flow, and global motion is represented with deformations of dense point grid and curve trajectories of sparse query points." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 124, + 468 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 124, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 124, + 468 + ], + "type": "text", + "content": "3.3. Objective" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": "Temporal discrete trajectory supervision. The available point tracking datasets provide only temporally discrete point tracks with no ground truth for continuous interframe trajectories. Following DOT [35], we first adopt supervised losses based on the temporal discrete ground-truth point tracks provided by the dataset, which consists of the L1 loss " + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\text{traj}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": " for sampled discrete trajectory prediction and the binary cross-entropy loss " + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "inline_equation", + "content": "L_{\\text{vis}}" + }, + { + "bbox": [ + 55, + 474, + 296, + 714 + ], + "type": "text", + "content": " for visibility map. We then randomly select different frame intervals for augmented training. Local correlation is not constructed when the frames are skipped, therefore the corresponding event features are taken into streaming for iteratively updating the global motion representation. There are cases where some images are not used as input when the frame interval is greater than 1, but the corresponding input events and ground-truth tracks can be regarded as inter-frame motion contributing to curve trajectory learning. Such sampling-based augmentation ensures the model learning through diverse long- and short-term motions, capitalizing on the continuity of events to estimate continuous trajectories." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "Event consistency with continuous trajectory. Since events are usually generated along motion trajectories, we propose to leverage the continuous property of events for self-supervised continuous trajectory learning in conjunction with discrete supervision of point trajectories. However, events are computationally intensive to process one by one and are generally accompanied by noise. We thus first introduce event temporal chunking to process events in batches within a fixed duration to reduce the noise impact and computation. For the " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "-th interval of " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " chunks, we isolate the events within that " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": "-th chunk and aggregate them after warping them to " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "t_b" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " as Eq. (2). For each chunk, the events then are summed into an image of warped events (IWE) [8], i.e., " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\mathbf{EB}(\\mathbf{x}_i,b)\\doteq \\sum_{i = 1}^{Ne}\\mathcal{N}(\\mathbf{x}_i;x_i',\\sigma^2)" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "t_b\\leq t_i < t_{b + 1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " is the neighboring range which is usually chosen as 1 pixel. This IWE essentially counts the number of warped events " + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "inline_equation", + "content": "e_i^\\prime" + }, + { + "bbox": [ + 313, + 72, + 555, + 312 + ], + "type": "text", + "content": " per pixel and chunk. The chunking intervals are chosen randomly to exploit the continuous nature of events. Thus we can establish consistent connections between event chunks and continuous trajectories:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 324, + 555, + 369 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 324, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 318, + 324, + 555, + 369 + ], + "type": "interline_equation", + "content": "L _ {e c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {b _ {1} \\neq b _ {2}} ^ {B} \\rho \\binom {\\mathbf {E B} (\\mathbf {x}, b _ {1}),} {\\text {W a r p} \\left(\\mathbf {E B} (\\mathbf {x}, b _ {2}); \\mathbf {T} \\left(t _ {b _ {2}}\\right) - \\mathbf {T} \\left(t _ {b _ {1}}\\right)\\right)}, \\tag {5}", + "image_path": "69bb240a819c4d2ecea956088ba29eb430bb8a1a34b6882fdaeee09f1613d090.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": " is the consistency measure with L1 norm. Since events are spatially sparse, we only establish connections with the dense trajectories at locations " + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "inline_equation", + "content": "\\Omega" + }, + { + "bbox": [ + 313, + 381, + 555, + 418 + ], + "type": "text", + "content": " with valid events." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 443, + 554, + 538 + ], + "type": "text", + "content": "Image consistency with discrete trajectory. Similar to the unsupervised optical flow task [30], we can also establish the discrete consistency of the motion trajectory with the images at discrete times, to compensate for the spatial sparsity issue of ground-truth point tracks in the training data. In addition, in our sampling-based augmented training, skipped images can be used as additional continuity training objectives." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "content": "For the accumulated continuous global trajectory " + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{T}_{1\\rightarrow t}" + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "content": " we sample the the discrete optical flow " + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_{i\\to j}" + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "inline_equation", + "content": "I_{i}" + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "inline_equation", + "content": "I_{j}" + }, + { + "bbox": [ + 313, + 540, + 555, + 588 + ], + "type": "text", + "content": " via timestamps. Similar to Eq. (5), the consistency of images can be modeled as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 326, + 601, + 555, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 601, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 326, + 601, + 555, + 635 + ], + "type": "interline_equation", + "content": "L _ {i c} = \\sum_ {\\mathbf {x}} ^ {\\Omega} \\sum_ {i \\neq j} ^ {N _ {v}} \\rho \\left(I _ {i} (\\mathbf {x}), \\operatorname {W a r p} \\left(I _ {k} (\\mathbf {x}); \\mathbf {F} _ {i \\rightarrow j} (\\mathbf {x})\\right)\\right). \\tag {6}", + "image_path": "f1fd13f5dd22b0b3d824b18f1ec68dfa6f0c5668f007ee785d44ac253b49af93.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": "Total objective. Our total training objective is a weighted combination of the above objectives, i.e., " + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "L = L_{\\text{traj}} + \\lambda_1 L_{vis} + \\lambda_2 L_{ec} + \\lambda_3 L_{ic}" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 654, + 555, + 714 + ], + "type": "text", + "content": " are manual hyperparameters. Our ablations verify that combined self-supervised training can compensate for the temporal continuity of trajectories." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7940" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 80, + 70, + 137, + 126 + ], + "blocks": [ + { + "bbox": [ + 80, + 70, + 137, + 126 + ], + "lines": [ + { + "bbox": [ + 80, + 70, + 137, + 126 + ], + "spans": [ + { + "bbox": [ + 80, + 70, + 137, + 126 + ], + "type": "image", + "image_path": "20926011ecdde905bc7cfdce5b6eee5f09588e736a811fc2ef07a45180e95c12.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 102, + 127, + 310, + 137 + ], + "lines": [ + { + "bbox": [ + 102, + 127, + 310, + 137 + ], + "spans": [ + { + "bbox": [ + 102, + 127, + 310, + 137 + ], + "type": "text", + "content": "Input images with GT query points: frame 1, 7, 19, 27" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 145, + 70, + 203, + 126 + ], + "blocks": [ + { + "bbox": [ + 145, + 70, + 203, + 126 + ], + "lines": [ + { + "bbox": [ + 145, + 70, + 203, + 126 + ], + "spans": [ + { + "bbox": [ + 145, + 70, + 203, + 126 + ], + "type": "image", + "image_path": "2d45f50ff938260353577c4235090cb96a364dd5ef4e944a02738893e6efa3c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 210, + 70, + 268, + 126 + ], + "blocks": [ + { + "bbox": [ + 210, + 70, + 268, + 126 + ], + "lines": [ + { + "bbox": [ + 210, + 70, + 268, + 126 + ], + "spans": [ + { + "bbox": [ + 210, + 70, + 268, + 126 + ], + "type": "image", + "image_path": "27f2044ab6b3377ffd94a49d78365fcab712c97a7432fd01f61ac863c2748b2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 276, + 70, + 332, + 126 + ], + "blocks": [ + { + "bbox": [ + 276, + 70, + 332, + 126 + ], + "lines": [ + { + "bbox": [ + 276, + 70, + 332, + 126 + ], + "spans": [ + { + "bbox": [ + 276, + 70, + 332, + 126 + ], + "type": "image", + "image_path": "3315f716dc28c0314ebcaf0092f1fd62dd651191e4e9ad230ca31e7a28c2f3e9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 80, + 139, + 137, + 196 + ], + "blocks": [ + { + "bbox": [ + 80, + 139, + 137, + 196 + ], + "lines": [ + { + "bbox": [ + 80, + 139, + 137, + 196 + ], + "spans": [ + { + "bbox": [ + 80, + 139, + 137, + 196 + ], + "type": "image", + "image_path": "9f5c2870b91081d75692e8bbcf34d2b15e60f3e5a64fb7a4fc4ab746c74358f2.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 86, + 196, + 130, + 205 + ], + "lines": [ + { + "bbox": [ + 86, + 196, + 130, + 205 + ], + "spans": [ + { + "bbox": [ + 86, + 196, + 130, + 205 + ], + "type": "text", + "content": "Init Coords" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 553, + 232 + ], + "type": "text", + "content": "Figure 2. Visual comparisons of long-term dense point tracking on the TAP-Vid-DAVIS [5] dataset, with the ground-truth sparse query points with input images. Due to space limitations, we provide more sequences and event data visualization in the Supp. and demo video." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 145, + 139, + 203, + 196 + ], + "blocks": [ + { + "bbox": [ + 145, + 139, + 203, + 196 + ], + "lines": [ + { + "bbox": [ + 145, + 139, + 203, + 196 + ], + "spans": [ + { + "bbox": [ + 145, + 139, + 203, + 196 + ], + "type": "image", + "image_path": "e95ad5b036df431cddbe71bd95aa089fa07429b2f0e090853e734070c58597d3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 196, + 280, + 205 + ], + "lines": [ + { + "bbox": [ + 198, + 196, + 280, + 205 + ], + "spans": [ + { + "bbox": [ + 198, + 196, + 280, + 205 + ], + "type": "text", + "content": "DOT: frame 7, 19, 27" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 210, + 140, + 268, + 196 + ], + "blocks": [ + { + "bbox": [ + 210, + 140, + 268, + 196 + ], + "lines": [ + { + "bbox": [ + 210, + 140, + 268, + 196 + ], + "spans": [ + { + "bbox": [ + 210, + 140, + 268, + 196 + ], + "type": "image", + "image_path": "f3532de9b9c32b6da9664495a0b985024862babb7e711754dfa432ce73dd9cd4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 276, + 140, + 332, + 196 + ], + "blocks": [ + { + "bbox": [ + 276, + 140, + 332, + 196 + ], + "lines": [ + { + "bbox": [ + 276, + 140, + 332, + 196 + ], + "spans": [ + { + "bbox": [ + 276, + 140, + 332, + 196 + ], + "type": "image", + "image_path": "f6fa006b79904cf61a6bd22771e1d03cddfb44e1418c7c280f8720f6c48d85bb.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 70, + 396, + 126 + ], + "blocks": [ + { + "bbox": [ + 340, + 70, + 396, + 126 + ], + "lines": [ + { + "bbox": [ + 340, + 70, + 396, + 126 + ], + "spans": [ + { + "bbox": [ + 340, + 70, + 396, + 126 + ], + "type": "image", + "image_path": "c262e1b5c82bd005854405afb531dde6a1471dc41fb333acc787bb409b4f6326.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 381, + 126, + 487, + 137 + ], + "lines": [ + { + "bbox": [ + 381, + 126, + 487, + 137 + ], + "spans": [ + { + "bbox": [ + 381, + 126, + 487, + 137 + ], + "type": "text", + "content": "CoTracker: frame 7, 19, 27" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 405, + 71, + 461, + 126 + ], + "blocks": [ + { + "bbox": [ + 405, + 71, + 461, + 126 + ], + "lines": [ + { + "bbox": [ + 405, + 71, + 461, + 126 + ], + "spans": [ + { + "bbox": [ + 405, + 71, + 461, + 126 + ], + "type": "image", + "image_path": "585f9913958a1e636dcba20093ed6804f66fbe5c65d1f129f4a034c030f6fc4b.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 470, + 71, + 526, + 126 + ], + "blocks": [ + { + "bbox": [ + 470, + 71, + 526, + 126 + ], + "lines": [ + { + "bbox": [ + 470, + 71, + 526, + 126 + ], + "spans": [ + { + "bbox": [ + 470, + 71, + 526, + 126 + ], + "type": "image", + "image_path": "3a733254be8ee7ec3a82c13ab25d185818bb623f85a2753ac7e70728ec36775e.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 341, + 139, + 397, + 196 + ], + "blocks": [ + { + "bbox": [ + 341, + 139, + 397, + 196 + ], + "lines": [ + { + "bbox": [ + 341, + 139, + 397, + 196 + ], + "spans": [ + { + "bbox": [ + 341, + 139, + 397, + 196 + ], + "type": "image", + "image_path": "22ac09a90fd9fad4530145fdb9a3c6e9f9b7a0253fa59b1deaa87614d7db97c5.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 196, + 476, + 205 + ], + "lines": [ + { + "bbox": [ + 392, + 196, + 476, + 205 + ], + "spans": [ + { + "bbox": [ + 392, + 196, + 476, + 205 + ], + "type": "text", + "content": "Ours: frame 7, 19, 27" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 405, + 140, + 462, + 196 + ], + "blocks": [ + { + "bbox": [ + 405, + 140, + 462, + 196 + ], + "lines": [ + { + "bbox": [ + 405, + 140, + 462, + 196 + ], + "spans": [ + { + "bbox": [ + 405, + 140, + 462, + 196 + ], + "type": "image", + "image_path": "6eb6b97bc24766c478ce95dd317c02396badee7d7cfa26369d7dd45953033f3a.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 470, + 140, + 527, + 196 + ], + "blocks": [ + { + "bbox": [ + 470, + 140, + 527, + 196 + ], + "lines": [ + { + "bbox": [ + 470, + 140, + 527, + 196 + ], + "spans": [ + { + "bbox": [ + 470, + 140, + 527, + 196 + ], + "type": "image", + "image_path": "7c0e298d0dcdb6f1dcf88a53ec7709bbb9e24e792e5f6964308a32c5b0f159e0.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "table", + "bbox": [ + 106, + 247, + 501, + 370 + ], + "blocks": [ + { + "bbox": [ + 137, + 235, + 471, + 245 + ], + "lines": [ + { + "bbox": [ + 137, + 235, + 471, + 245 + ], + "spans": [ + { + "bbox": [ + 137, + 235, + 471, + 245 + ], + "type": "text", + "content": "Table 1. Quantitative results of dense evaluation on the CVO test [52] and extended set [35]." + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 106, + 247, + 501, + 370 + ], + "lines": [ + { + "bbox": [ + 106, + 247, + 501, + 370 + ], + "spans": [ + { + "bbox": [ + 106, + 247, + 501, + 370 + ], + "type": "table", + "html": "
MethodCVO (Clean)CVO (Final)CVO (Extended)
EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑EPEall/vis/occ ↓OA ↑
QueryPIPs++ [58]9.05 / 6.62 / 21.533.39.49 / 7.06 / 22.032.718.4 / 10.0 / 32.158.7
TAPIR [6]3.80 / 1.49 / 14.773.54.19 / 1.86 / 15.372.419.8 / 4.74 / 42.568.4
CoTracker [24]1.51 / 0.88 / 4.5775.51.52 / 0.93 / 4.3875.35.20 / 3.84 / 7.7070.4
DenseGMA [22]2.42 / 1.38 / 7.1460.52.57 / 1.52 / 7.2259.721.8 / 15.7 / 32.865.6
MFT [36]2.91 / 1.39 / 9.9319.43.16 / 1.56 / 10.319.521.4 / 9.20 / 41.837.6
AccFlow [52]1.69 / 1.08 / 4.7048.11.73 / 1.15 / 4.6347.536.7 / 28.1 / 52.936.5
DOT [35]1.32 / 0.74 / 4.1280.41.38 / 0.82 / 4.1080.25.07 / 3.67 / 7.3471.0
EDCPT (Ours)1.23 / 0.71 / 3.8382.11.31 / 0.76 / 3.8681.94.88 / 3.44 / 7.4671.9
", + "image_path": "399048cff99b027d35387223de467e7e9364fa8f5645dd11399429ccd09228c7.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "table_body" + } + ], + "index": 21 + }, + { + "type": "table", + "bbox": [ + 65, + 402, + 285, + 559 + ], + "blocks": [ + { + "bbox": [ + 55, + 379, + 295, + 401 + ], + "lines": [ + { + "bbox": [ + 55, + 379, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 379, + 295, + 401 + ], + "type": "text", + "content": "Table 2. Quantitative results of sparse evaluation on the TAP-VidDAVIS [5]point tracking benchmark." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 402, + 285, + 559 + ], + "lines": [ + { + "bbox": [ + 65, + 402, + 285, + 559 + ], + "spans": [ + { + "bbox": [ + 65, + 402, + 285, + 559 + ], + "type": "table", + "html": "
MethodDAVIS (First)DAVIS (Strided)
AJ ↑<δxavg↑OA ↑AJ ↑<δxavg↑OA ↑
QueryTAP-Net [5]33.048.678.838.453.182.3
Context-PIPs [51]42.760.379.548.964.083.4
TAPIR [6]56.270.086.561.373.688.8
CoTracker [24]61.174.689.163.579.887.8
SpatialTracker [54]61.176.389.5---
DenseCPFlow [31]9.614.6----
MFT [36]47.366.877.856.170.886.9
DecoMotion [26]53.069.984.260.274.487.2
DinoTracker [46]---62.378.287.5
FlowTrack [4]---63.276.389.2
DOT [35]61.675.589.566.780.690.4
EDCPT (Ours)63.876.390.667.580.591.1
", + "image_path": "562352820b0b2d107319d84ade9f7a16e050a1b4ec343d05e564c27792a0fe08.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 568, + 137, + 582 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 137, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 137, + 582 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 55, + 588, + 175, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 175, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 175, + 601 + ], + "type": "text", + "content": "4.1. Experimental details" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "We train our model on MOVI-F [15] dataset and directly evaluate it on CVO [52] and TAP-Vid-DAVIS [5] benchmarks, using the vid2e [10] simulator to generate events.Following DOT [35], we train the model for 500k steps on " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "4\\times" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " NVIDIA L40 48G GPUs, using the Adam optimizer and OneCycle learning rate decay with a maximum of " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "10^{-4}" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " . We choose 3 frames as training samples, along with the random selection of up to 10 frames in different frame intervals. The loss hyperparameters are set to 1.0," + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "text", + "content": "0.1, 0.1. For dense evaluation, we report the dense absolute error " + }, + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "inline_equation", + "content": "\\mathrm{EPE}_{\\mathrm{all / vis / occ}}" + }, + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "text", + "content": " for all/visible/occluded points, as well as occlusion IoU accuracy OA for estimated visible mask. For sparse evaluation, we follow TAPNet [5] and report average Jaccard AJ, position accuracy " + }, + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "inline_equation", + "content": "< \\delta_{\\mathrm{avg}}^x" + }, + { + "bbox": [ + 313, + 380, + 555, + 477 + ], + "type": "text", + "content": ", and occlusion accuracy OA. Additionally, we adopt the real-captured event-based optical flow benchmark, DSEC [12, 13], to verify our adaptation capacity. More details are provided in the Supp." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 489, + 522, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 489, + 522, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 489, + 522, + 502 + ], + "type": "text", + "content": "4.2. Standard spatially dense point tracking" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 313, + 508, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 555, + 651 + ], + "type": "text", + "content": "Performance on CVO and TAP-Vid-DAVIS benchmarks. We conduct comprehensive evaluations on two common standard point tracking benchmarks. Consistent with DOT [35], we report the quantitative results of the spatially dense optical flow from the last to the first frame of CVO [52] dataset in Tab. 1. Our proposed new EDCPT framework achieves significant performance improvements, whether comparing methods that only predict the partial Query points or directly estimating spatially Dense trajectories within a single inference. Particularly, we achieve 0.19 " + }, + { + "bbox": [ + 313, + 508, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{EPE}_{\\mathrm{all}}" + }, + { + "bbox": [ + 313, + 508, + 555, + 651 + ], + "type": "text", + "content": " and 0.9 OA improvements on the extended set of 476 videos with 48 frames when compared to DOT [35]." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 713 + ], + "type": "text", + "content": "In contrast, the real TAP-Vid-DAVIS dataset [5] only provides ground-truth trajectories for selected query points. As a result, we only sparsely evaluate these points for a fair comparison despite the output trajectories of our model and some compared methods are spatially dense. The" + } + ] + } + ], + "index": 30 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7941" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 70, + 226, + 177 + ], + "blocks": [ + { + "bbox": [ + 81, + 70, + 226, + 177 + ], + "lines": [ + { + "bbox": [ + 81, + 70, + 226, + 177 + ], + "spans": [ + { + "bbox": [ + 81, + 70, + 226, + 177 + ], + "type": "image", + "image_path": "c08b675dfba08ef858393f7b0a353e69d582a04423327d6ff1aa24e19c8627e2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 58, + 177, + 550, + 188 + ], + "lines": [ + { + "bbox": [ + 58, + 177, + 550, + 188 + ], + "spans": [ + { + "bbox": [ + 58, + 177, + 550, + 188 + ], + "type": "text", + "content": "Figure 3. Visual comparisons of dense and continuous point trajectories on TAP-Vid-DAVIS [5]. Zoom in for detailed curve trajectories." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 227, + 70, + 375, + 177 + ], + "blocks": [ + { + "bbox": [ + 227, + 70, + 375, + 177 + ], + "lines": [ + { + "bbox": [ + 227, + 70, + 375, + 177 + ], + "spans": [ + { + "bbox": [ + 227, + 70, + 375, + 177 + ], + "type": "image", + "image_path": "765a1a10e4c12ae8e59dde07123a44f8fb8422a4be3c7a7e2844aa5761b84e3a.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 376, + 71, + 524, + 177 + ], + "blocks": [ + { + "bbox": [ + 376, + 71, + 524, + 177 + ], + "lines": [ + { + "bbox": [ + 376, + 71, + 524, + 177 + ], + "spans": [ + { + "bbox": [ + 376, + 71, + 524, + 177 + ], + "type": "image", + "image_path": "db3f7af69fd85254f5392812c74c36d4fe34112186ee69fc27f0443e98446d45.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 70, + 231, + 279, + 345 + ], + "blocks": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 231 + ], + "type": "text", + "content": "Table 3. Quantitative results on the real DSEC optical flow benchmark [13]. The first two methods are categorized as self-supervised learning and the remaining as supervised learning." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 70, + 231, + 279, + 345 + ], + "lines": [ + { + "bbox": [ + 70, + 231, + 279, + 345 + ], + "spans": [ + { + "bbox": [ + 70, + 231, + 279, + 345 + ], + "type": "table", + "html": "
MethodInputEPE ↓AE ↓%Out ↓
Taming [39]Events2.3310.5617.77
MPCMax [18]Events3.208.5315.21
E-RAFT [13]Events0.792.852.68
TMA [28]Events0.742.682.30
IDNet [53]Events0.722.722.04
ECDDP [56]Events0.702.581.96
BFlow [14]Events0.752.682.44
BFlow [14]Images+Events0.692.421.88
EDCPT (Ours)Images+Events0.632.171.52
", + "image_path": "8bd3752f65657b0a13367ea3ba74981164256cf90732d38c01eb726dcb4e87cb.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 356, + 296, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 296, + 499 + ], + "type": "text", + "content": "quantitative results in Tab. 2 demonstrate the superiority of our framework, outperforming the existing state-of-the-art methods DOT [35] and SpatialTracker [54] with up to 2.7 AJ and 1.1 OA. We argue that ours " + }, + { + "bbox": [ + 54, + 356, + 296, + 499 + ], + "type": "inline_equation", + "content": "< \\delta_{\\mathrm{avg}}^{x}" + }, + { + "bbox": [ + 54, + 356, + 296, + 499 + ], + "type": "text", + "content": " is close to DOT's because the sparse to dense upgrade strategy from sparse point matches is capable of modeling the rigid object motion and thus benefiting the proportion evaluation. We also perform qualitative visual comparisons in Fig. 2. Combining the above quantitative and qualitative comparisons with previous image-based methods, the new attempts of incorporating events by our framework significantly improve the accuracy with the standard dense point tracking setting." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 518, + 296, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 649 + ], + "type": "text", + "content": "Performance on DSEC benchmark. We further conduct experiments on the DSEC benchmark [13] with real captured event data. Because the DSEC online leaderboard only measures the optical flow between two consecutive frames, we therefore finetune the local motion estimation on the DSEC training set from our pre-trained long-term model. The submission results are shown in Tab. 3. Notably, while BFlow [14] can estimate curve trajectories, they only submitted another optical flow version. Our framework fuses image and event data, yielding state-of-the-art performance with improvements of 0.05 EPE and 0.25 AE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 256, + 672 + ], + "type": "text", + "content": "4.3. Temporally continuous point tracking" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "We adapt the above standard evaluation to input only a portion of the full video frames into the model to evaluate the temporal continuity with the ground truths of skipped" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 199, + 553, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 553, + 319 + ], + "type": "text", + "content": "frames. For the CVO final set with 7 frames per video, we skip 1 frame (half) and 2 frames (one-third) as model inputs, because the longer extended set lacks multi-frame ground-truth tracks. For the TAP-Vid-DAVIS dataset with " + }, + { + "bbox": [ + 313, + 199, + 553, + 319 + ], + "type": "inline_equation", + "content": "\\sim 100" + }, + { + "bbox": [ + 313, + 199, + 553, + 319 + ], + "type": "text", + "content": " frames, we report skip 1 frame (half) and skip 3 frames (quarter). The compared image-based methods cannot model inter-frame motion, thus we take linear motion interpolation to generate trajectory when frames are skipped. We retrain AccFlow [52] and mark it with * as its public version does not support forward point tracking." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 319, + 554, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 511 + ], + "type": "text", + "content": "As reported in Tab. 4, our proposed new framework with global continuous trajectory accumulation significantly outperforms existing methods. Especially in nonlinear motion scenarios of TAP-Vid-DAVIS datasets, the larger frame intervals lead to greater performance gaps. In addition to the ablation of different motion assumptions in Tab. 6, the B-spline representation we adopt achieves better performance. We also provide visual comparisons in Fig. 3 and a demo video of continuous trajectory visualization in the Supp., that includes four sequences with diverse motion complexity and event sparsity from TAP-Vid-DAVIS and real-captured ERF-X170FPS [25] dataset. With extensive experiments on both simulated events (CVO and TAP-Vid-DAVIS) and real events (DSEC and ERF-X170FPS), we fully validate the robustness of the proposed global motion accumulation in modeling complex continuous trajectories." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 521, + 512, + 533 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 512, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 512, + 533 + ], + "type": "text", + "content": "4.4. Ablation experiments and discussions" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 539, + 554, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 539, + 554, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 539, + 554, + 611 + ], + "type": "text", + "content": "To perform progressive ablations in Tab. 5, Tab. 6, and Tab. 7, the underlined settings are those utilized in the previous table, and the bolded ones represent the choices for our final framework. To validate the capability for continuous point tracking, the ablation results are reported with CVO third and DAVIS quarter settings as in Sec. 4.3 and Tab. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Global motion aggregation. One of our key contributions is the global aggregation of local motion representations in the streaming pipeline. In Tab. 5, N/A indicates do not explicitly model sequential motion as in DOT [35]. post is the forward aggregation in AccFlow [52], which performs post-processing only on the output optical flow and ignores the internal long-term motion information. We aggregate" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7942" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 94, + 83, + 515, + 171 + ], + "blocks": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "lines": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "spans": [ + { + "bbox": [ + 94, + 70, + 515, + 82 + ], + "type": "text", + "content": "Table 4. Continuous point tracking evaluation results on the CVO extended set [35] and TAP-Vid-DAVIS dataset [5]." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 94, + 83, + 515, + 171 + ], + "lines": [ + { + "bbox": [ + 94, + 83, + 515, + 171 + ], + "spans": [ + { + "bbox": [ + 94, + 83, + 515, + 171 + ], + "type": "table", + "html": "
MethodCVO (Final) - EPEall/vis/occ ↓DAVIS (First) - AJ / <δxavg↑
fullhalfthirdfullhalfquarter
RAFT2.09 / 0.81 / 8.022.44 / 0.95 / 9.073.02 / 1.16 / 11.4233.9 / 46.628.6 / 40.122.4 / 34.2
GMA1.99 / 0.77 / 7.572.35 / 0.89 / 8.452.92 / 1.09 / 10.9739.3 / 52.531.7 / 44.326.5 / 38.3
AccFlow*2.28 / 0.60 / 11.182.39 / 0.80 / 10.742.79 / 1.02 / 11.3747.2 / 62.337.5 / 49.230.9 / 42.4
CoTracker1.89 / 0.63 / 7.052.11 / 0.82 / 8.023.17 / 1.65 / 11.1361.1 / 74.654.3 / 68.848.9 / 63.9
DOT1.83 / 0.59 / 6.952.10 / 0.73 / 7.882.69 / 0.97 / 10.8461.6 / 75.555.6 / 70.150.4 / 65.3
EDCPT (Ours)1.76 / 0.55 / 6.731.97 / 0.66 / 7.612.16 / 0.73 / 8.7663.8 / 76.359.7 / 73.156.2 / 70.9
", + "image_path": "4f57aa91c8775197a33e55c85e5b0abf26aaa0aeaf1a11533f8596fe8b06fb54.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 199, + 220, + 275 + ], + "blocks": [ + { + "bbox": [ + 58, + 176, + 550, + 198 + ], + "lines": [ + { + "bbox": [ + 58, + 176, + 550, + 198 + ], + "spans": [ + { + "bbox": [ + 58, + 176, + 550, + 198 + ], + "type": "text", + "content": "Table 5. Ablations on global motion aggre- Table 6. Ablations on parametric curve rep- Table 7. Ablations on input data and supervi-gation. resentation. sion." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 199, + 220, + 275 + ], + "lines": [ + { + "bbox": [ + 60, + 199, + 220, + 275 + ], + "spans": [ + { + "bbox": [ + 60, + 199, + 220, + 275 + ], + "type": "table", + "html": "
EPEall/vis/occ ↓AJ / <δxavg↑
N/A2.54 / 0.89 / 10.3351.9 / 66.4
post2.49 / 0.85 / 9.7352.5 / 66.9
solo2.45 / 0.82 / 9.8952.7 / 67.0
–offsets2.43 / 0.82 / 9.7353.0 / 67.2
stream2.42 / 0.80 / 9.6853.3 / 67.4
", + "image_path": "0c74a03fcba7e50527e63c410fa7a3424b13206a4f1c6f4cfec7d6b559ba4a24.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 223, + 199, + 383, + 277 + ], + "blocks": [ + { + "bbox": [ + 223, + 199, + 383, + 277 + ], + "lines": [ + { + "bbox": [ + 223, + 199, + 383, + 277 + ], + "spans": [ + { + "bbox": [ + 223, + 199, + 383, + 277 + ], + "type": "table", + "html": "
EPEall/vis/occ ↓AJ / <δxavg↑
linear2.42 / 0.80 / 9.6853.3 / 67.4
quad2.49 / 0.84 / 9.4654.2 / 68.1
Nc=32.32 / 0.79 / 9.3354.4 / 68.6
Nc=42.23 / 0.76 / 8.9755.4 / 69.7
Nc=52.26 / 0.75 / 9.2055.0 / 69.3
", + "image_path": "f960c4a1739216187b9161ddb88165d8f9625561acb4800e259156de1099fb2b.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 386, + 199, + 548, + 262 + ], + "blocks": [ + { + "bbox": [ + 386, + 199, + 548, + 262 + ], + "lines": [ + { + "bbox": [ + 386, + 199, + 548, + 262 + ], + "spans": [ + { + "bbox": [ + 386, + 199, + 548, + 262 + ], + "type": "table", + "html": "
EPEall/vis/occ ↓AJ / <δxavg ↑
Images2.50 / 0.86 / 9.9252.5 / 66.7
+Events2.23 / 0.76 / 8.9755.4 / 69.7
+Lic2.20 / 0.75 / 8.9055.8 / 70.2
+Lec2.16 / 0.73 / 8.7656.2 / 70.9
", + "image_path": "20d0e93eaf1cbb56f7d6f113f9f5e2e7a3568bd4059280131c909d7a97f597f3.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "spans": [ + { + "bbox": [ + 54, + 289, + 295, + 456 + ], + "type": "text", + "content": "motion representations at the feature level instead of dealing directly with motion vectors. solo is the short and long term fusion module in SOLOFusion [40]. Unlike its temporal fusion with a fixed number of frames and low resolution, we adopt sequential modeling with an unspecified number of frames in a streaming pipeline. Our proposed motion aggregation framework fuses image correspondence and event features from local to global with recurrent streaming structure, which better preserves multi-frame motion information and achieves optimal performance. In addition, we also verified that removing the additional offset estimation to address the numerical problem in Warping leads to a slight performance degradation, as this would require the subsequent refinement to handle it simultaneously." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 471, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 295, + 567 + ], + "type": "text", + "content": "Curve representation. Based on our streaming framework, we compare the performance improvement of taking the B-spline curve representation compared to interpolating with linear and quadratic motion assumptions in Tab. 6. For the control points number of B-spline curves, we chose " + }, + { + "bbox": [ + 55, + 471, + 295, + 567 + ], + "type": "inline_equation", + "content": "N_{c} = 4" + }, + { + "bbox": [ + 55, + 471, + 295, + 567 + ], + "type": "text", + "content": ", since further increasing the number of control points does not significantly improve the performance under CVO third and DAVIS quarter settings." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 713 + ], + "type": "text", + "content": "Input data and supervision. Since previous methods usually use only image data, we verify the advantages of incorporating event data for high-precision continuous point tracking by removing events, as depicted in Tab. 7. Moreover, the comparison results between this image-only variant and DOT [35] in Tab. 4 demonstrates that our proposed streaming aggregation and curve representation are beneficial even in the absence of event data. Furthermore, we validate that training using the proposed image and event-to-point trajectory consistencies as additional supervision complements the lack of continuous inter-frame tracks in" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 289, + 535, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 535, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 535, + 300 + ], + "type": "text", + "content": "the training data and can further improve performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 319, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 555, + 488 + ], + "type": "text", + "content": "Limitations. On the same RTX 3090 PC, our framework processes a 48-frame video at " + }, + { + "bbox": [ + 313, + 319, + 555, + 488 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 313, + 319, + 555, + 488 + ], + "type": "text", + "content": " resolution in 12.6 seconds, significantly faster than CoTracker's 11 minutes, while only processing partial query points in a single run. However, it is slower than the two-frame optical flow method GMA of only 2.1 seconds and slightly slower than the multi-frame point tracking method DOT of 9.5 seconds. In addition, there is currently a lack of point-tracking datasets with real events, and it is difficult to obtain long-term tracking annotations. We evaluate on standard benchmarks with simulated events, as well as visualize point tracking on real events, and also evaluate optical flow estimation on a real event benchmark. Our future work plans to improve on model efficiency and the evaluation setup." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 388, + 514 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 555, + 713 + ], + "type": "text", + "content": "In this paper, we propose a new long-term dense point tracking framework for integrating image and event data to estimate continuous motion trajectories. Specifically, we streamingly process two adjacent frames and inter-frame events to generate local motion representations, and accumulate previous caches to global motion representation at the feature level to produce new global trajectories at the trajectory level. We utilize multi-frame parametric curve accumulation to represent continuous trajectories with any number of frames, complemented by image and event-to-trajectory consistencies for model training. We validate the effectiveness of our framework by conducting extensive benchmark experiments as well as qualitative verification on real challenge data. We believe this work provides new insights into the point tracking task from the perspective of event modality and continuous global curve representations." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7943" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This research/project is supported by the National Natural Science Foundation of China (62271410, 12150007), the National Research Foundation, Singapore, under its NRF-Investigatorship Programme (Award ID. NRF-NRFI09-0008), and the Tier 1 grant T1-251RES2305 from the Singapore Ministry of Education. Zhexiong Wan was also supported by the Program of China Scholarship Council (202306290193), and the Innovation Foundation for Doctor Dissertation of Northwestern Polytechnical University (CX2023013)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 115, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 242, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 61, + 242, + 296, + 297 + ], + "type": "text", + "content": "[1] Eirikur Agustsson, David Minnen, Nick Johnston, Johannes Balle, Sung Jin Hwang, and George Toderici. Scalespace flow for end-to-end optimized video compression. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 8503-8512, 2020. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 298, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 298, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 298, + 296, + 343 + ], + "type": "text", + "content": "[2] Adam Bielski and Paolo Favaro. Move: Unsupervised movable object segmentation and detection. Advances in Neural Information Processing Systems (NeurIPS), 35:33371-33386, 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 61, + 344, + 296, + 388 + ], + "type": "text", + "content": "[3] Kyunghyun Cho, Bart van Merrienboer, Dzmitry Bahdanau, and Yoshua Bengio. On the properties of neural machine translation: Encoder-decoder approaches. arXiv preprint arXiv:1409.1259, 2014. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 61, + 388, + 296, + 443 + ], + "type": "text", + "content": "[4] Seokju Cho, Jiahui Huang, Seungryong Kim, and Joon-Young Lee. Flowtrack: Revisiting optical flow for long-range dense tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19268–19277, 2024. 1, 2, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 61, + 444, + 296, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 444, + 296, + 511 + ], + "spans": [ + { + "bbox": [ + 61, + 444, + 296, + 511 + ], + "type": "text", + "content": "[5] Carl Doersch, Ankush Gupta, Larisa Markeeva, Adria Re-casens, Lucas Smaira, Yusuf Aytar, Joao Carreira, Andrew Zisserman, and Yi Yang. Tap-vid: A benchmark for tracking any point in a video. In Advances in Neural Information Processing Systems (NeurIPS) Datasets and Benchmarks Track, pages 13610–13626, 2022. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 61, + 512, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 512, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 61, + 512, + 296, + 567 + ], + "type": "text", + "content": "[6] Carl Doersch, Yi Yang, Mel Vecerik, Dilara Gokay, Ankush Gupta, Yusuf Aytar, Joao Carreira, and Andrew Zisserman. Tapir: Tracking any point with per-frame initialization and temporal refinement. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10061-10072, 2023. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 61, + 568, + 296, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 568, + 296, + 623 + ], + "spans": [ + { + "bbox": [ + 61, + 568, + 296, + 623 + ], + "type": "text", + "content": "[7] Carl Doersch, Yi Yang, Dilara Gokay, Pauline Luc, Skanda Koppula, Ankush Gupta, Joseph Heyward, Ross Goroshin, João Carreira, and Andrew Zisserman. Bootstrap: Bootstrapped training for tracking-any-point. In Asian Conference on Computer Vision (ACCV), 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 61, + 624, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 624, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 61, + 624, + 296, + 679 + ], + "type": "text", + "content": "[8] Guillermo Gallego, Henri Rebecq, and Davide Scaramuzza. A unifying contrast maximization framework for event cameras, with applications to motion, depth, and optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3867-3876, 2018. 4, 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 61, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 61, + 680, + 296, + 713 + ], + "type": "text", + "content": "[9] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J. Davison, Jörg Conradt, Kostas Daniilidis, and" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 556, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 106 + ], + "type": "text", + "content": "Davide Scaramuzza. Event-based vision: a survey. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 44(1):154-180, 2022. 1" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 162 + ], + "type": "text", + "content": "[10] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3586-3595, 2020. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 163, + 556, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 556, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 556, + 207 + ], + "type": "text", + "content": "[11] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: Asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision (IJCV), 128(3):601-618, 2020. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 554, + 252 + ], + "type": "text", + "content": "[12] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. DSEC: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters (RA-L), 2021. 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 297 + ], + "type": "text", + "content": "[13] Mathias Gehrig, Mario Millhäsler, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), pages 197–206, 2021. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 298, + 554, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 554, + 343 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 554, + 343 + ], + "type": "text", + "content": "[14] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 46(7):4736-4746, 2024. 1, 3, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 344, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 344, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 344, + 554, + 409 + ], + "type": "text", + "content": "[15] Klaus Greff, Francois Belletti, Lucas Beyer, Carl Doersch, Yilun Du, Daniel Duckworth, David J Fleet, Dan Gnanapragasam, Florian Golemo, Charles Herrmann, et al. Kubric: A scalable dataset generator. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3749-3761, 2022. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 554, + 466 + ], + "type": "text", + "content": "[16] Xiang Guo, Jiadai Sun, Yuchao Dai, Guanying Chen, Xiaoqing Ye, Xiao Tan, Errui Ding, Yumeng Zhang, and Jingdong Wang. Forward flow for novel view synthesis of dynamic scenes. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 16022-16033, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 467, + 554, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 521 + ], + "type": "text", + "content": "[17] Jesse Hagenaars, Federico Paredes-Vallés, and Guido De Croon. Self-supervised learning of event-based optical flow with spiking neural networks. In Advances in Neural Information Processing Systems (NeurIPS), pages 7167-7179, 2021. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 522, + 554, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 522, + 554, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 522, + 554, + 578 + ], + "type": "text", + "content": "[18] Friedhelm Hamann, Ziyun Wang, Ioannis Asmanis, Kenneth Chaney, Guillermo Gallego, and Kostas Daniilidis. Motion-prior contrast maximization for dense continuous-time motion estimation. In European Conference on Computer Vision (ECCV), 2024. 3, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 554, + 624 + ], + "type": "text", + "content": "[19] Adam W Harley, Zhaoyuan Fang, and Katerina Fragkiadaki. Particle video revisited: Tracking through occlusions using point trajectories. In European Conference on Computer Vision (ECCV), pages 59-75. Springer, 2022. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 554, + 678 + ], + "type": "text", + "content": "[20] Xueyan Huang, Yueyi Zhang, and Zhiwei Xiong. Progressive spatio-temporal alignment for efficient event-based motion estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1537-1546, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 679, + 554, + 713 + ], + "type": "text", + "content": "[21] Eddy Ilg, Nikolaus Mayer, Tonmoy Saikia, Margret Keuper, Alexey Dosovitskiy, and Thomas Brox. FlowNet 2.0: Evolution of optical flow estimation with deep networks." + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7944" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 2462-2470, 2017. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 97, + 295, + 151 + ], + "type": "text", + "content": "[22] Shihao Jiang, Dylan Campbell, Yao Lu, Hongdong Li, and Richard Hartley. Learning to estimate hidden motions with global motion aggregation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9772-9781, 2021. 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 153, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 153, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 153, + 295, + 208 + ], + "type": "text", + "content": "[23] Xin Jin, Longhai Wu, Jie Chen, Youxin Chen, Jayoon Koo, and Cheul-hee Hahm. A unified pyramid recurrent network for video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1578-1587, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 211, + 294, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 211, + 294, + 255 + ], + "spans": [ + { + "bbox": [ + 56, + 211, + 294, + 255 + ], + "type": "text", + "content": "[24] Nikita Karaev, Ignacio Rocco, Benjamin Graham, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht. Co-tracker: It is better to track together. In European Conference on Computer Vision (ECCV), 2024. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 257, + 294, + 312 + ], + "type": "text", + "content": "[25] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 315, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 346 + ], + "type": "text", + "content": "[26] Rui Li and Dong Liu. Decomposition betters tracking everything everywhere. In European Conference on Computer Vision (ECCV), 2024. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 392 + ], + "type": "text", + "content": "[27] Siqi Li, Zhikuan Zhou, Zhou Xue, Yipeng Li, Shaoyi Du, and Yue Gao. 3d feature tracking via event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18974-18983, 2024. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 396, + 294, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 396, + 294, + 449 + ], + "spans": [ + { + "bbox": [ + 56, + 396, + 294, + 449 + ], + "type": "text", + "content": "[28] Haotian Liu, Guang Chen, Sanqing Qu, Yanping Zhang, Zhijun Li, Alois Knoll, and Changjun Jiang. Tma: Temporal motion aggregation for event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9685-9694, 2023. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "spans": [ + { + "bbox": [ + 56, + 453, + 294, + 495 + ], + "type": "text", + "content": "[29] Jiaxiong Liu, Bo Wang, Zhen Tan, Jinpu Zhang, Hui Shen, and Dewen Hu. Tracking any point with frame-event fusion network at high frame rate. arXiv preprint arXiv:2409.11953, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 294, + 564 + ], + "type": "text", + "content": "[30] Liang Liu, Jiangning Zhang, Ruifei He, Yong Liu, Yabiao Wang, Ying Tai, Donghao Luo, Chengjie Wang, Jilin Li, and Feiyue Huang. Learning by analogy: Reliable supervision from transformations for unsupervised optical flow estimation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 6489-6498, 2020. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 609 + ], + "type": "text", + "content": "[31] Jianqin Luo, Zhexiong Wan, Yuxin Mao, Bo Li, and Yuchao Dai. Continuous parametric optical flow. In Advances in Neural Information Processing Systems (NeurIPS), pages 23520-23532, 2023. 1, 2, 3, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 667 + ], + "type": "text", + "content": "[32] Xinglong Luo, Ao Luo, Zhengning Wang, Chunyu Lin, Bing Zeng, and Shuaicheng Liu. Efficient meshflow and optical flow estimation from event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 19198-19207, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[33] Nico Messikommer, Carter Fang, Mathias Gehrig, and Davide Scaramuzza. Data-driven feature tracking for event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5642-5651, 2023. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[34] Etienne Meunier, Anais Badoual, and Patrick Bouthemy. Em-driven unsupervised learning for efficient motion segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 45(4):4462-4473, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[35] Guillaume Le Moing, Jean Ponce, and Cordelia Schmid. Dense optical tracking: Connecting the dots. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 205 + ], + "type": "text", + "content": "[36] Michal Neoral, Jonás Šerzych, and Jérí Matas. Mft: Long-term tracking of every pixel. In IEEE Winter Conference on Applications of Computer Vision (WACV), pages 6837-6847, 2024. 2, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 206, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 206, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 206, + 553, + 315 + ], + "type": "text", + "content": "[37] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel HAZIZA, Francisco Massa, Alaaeldin El-Nouby, Mido Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. DINOv2: Learning robust visual features without supervision. Transactions on Machine Learning Research (TMLR), 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[38] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1669-1678, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 415 + ], + "type": "text", + "content": "[39] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 9695-9705, 2023. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 416, + 553, + 470 + ], + "type": "text", + "content": "[40] Jinhyung Park, Chenfeng Xu, Shijia Yang, Kurt Keutzer, Kris M. Kitani, Masayoshi Tomizuka, and Wei Zhan. Time will tell: New outlooks and a baseline for temporal multiview 3d object detection. In International Conference on Learning Representations (ICLR), 2023. 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 525 + ], + "type": "text", + "content": "[41] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 3857-3866, 2019. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 568 + ], + "type": "text", + "content": "[42] Yunzhou Song, Jiahui Lei, Ziyun Wang, Lingjie Liu, and Kostas Daniilidis. Track everything everywhere fast and robustly. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 570, + 553, + 624 + ], + "type": "text", + "content": "[43] Xiuchao Sui, Shaohua Li, Xue Geng, Yan Wu, Xinxing Xu, Yong Liu, Rick Goh, and Hongyuan Zhu. CRAFT: Cross-attentional flow transformer for robust optical flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17602-17611, 2022. 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 658 + ], + "type": "text", + "content": "[44] Zachary Teed and Jia Deng. RAFT: Recurrent all-pairs field transforms for optical flow. In European Conference on Computer Vision (ECCV), pages 402-419, 2020. 1, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[45] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time Lens: Event-based video frame interpolation. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16155-16164, 2021. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7945" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[46] Narek Tumanyan, Assaf Singer, Shai Bagon, and Tali Dekel. Dino-tracker: Taming dino for self-supervised point tracking in a single video. In European Conference on Computer Vision (ECCV), 2024. 1, 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 160 + ], + "type": "text", + "content": "[47] Zhexiong Wan, Yuchao Dai, and Yuxin Mao. Learning dense and continuous optical flow from an event camera. IEEE Transactions on Image Processing (TIP), 31:7237-7251, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 215 + ], + "type": "text", + "content": "[48] Zhexiong Wan, Yuxin Mao, Jing Zhang, and Yuchao Dai. RPEFlow: Multimodal fusion of RGB-pointcloud-event for joint optical flow and scene flow estimation. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 10030-10040, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 270 + ], + "type": "text", + "content": "[49] Qianqian Wang, Yen-Yu Chang, Ruojin Cai, Zhengqi Li, Bharath Hariharan, Aleksander Holynski, and Noah Snavely. Tracking everything everywhere all at once. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19795-19806, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "text", + "content": "[50] Xiangyuan Wang, Huai Yu, Lei Yu, Wen Yang, and GuiSong Xia. Towards robust keypoint detection and tracking: A fusion approach with event-aligned image features. IEEE Robotics and Automation Letters (RA-L), 9(9):8059-8066, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 381 + ], + "type": "text", + "content": "[51] BIAN Weikang, Zhaoyang Huang, Xiaoyu Shi, Yitong Dong, Yijin Li, and Hongsheng Li. Context-pips: Persistent independent particles demands spatial context features. In Advances in Neural Information Processing Systems (NeurIPS), pages 55285-55298, 2023. 2, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 446 + ], + "type": "text", + "content": "[52] Guangyang Wu, Xiaohong Liu, Kunming Luo, Xi Liu, Qingqing Zheng, Shuaicheng Liu, Xinyang Jiang, Guangtao Zhai, and Wenyi Wang. Accflow: backward accumulation for long-range optical flow. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 12119-12128, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 294, + 502 + ], + "type": "text", + "content": "[53] Yilun Wu, Federico Paredes-Vallés, and Guido CHE De Croon. Lightweight event-based optical flow estimation via iterative deblurring. In IEEE International Conference on Robotics and Automation (ICRA), pages 14708-14715. IEEE, 2024. 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 557 + ], + "type": "text", + "content": "[54] Yuxi Xiao, Qianqian Wang, Shangzhan Zhang, Nan Xue, Sida Peng, Yujun Shen, and Xiaowei Zhou. Spatialtracker: Tracking any 2d pixels in 3d space. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 602 + ], + "type": "text", + "content": "[55] Xiangyu Xu, Li Siyao, Wenxiu Sun, Qian Yin, and Ming-Hsuan Yang. Quadratic video interpolation. In Advances in Neural Information Processing Systems (NeurIPS), pages 1647-1656, 2019. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 604, + 294, + 635 + ], + "type": "text", + "content": "[56] Yan Yang, Liyuan Pan, and Liu Liu. Event camera data dense pre-training. In European Conference on Computer Vision (ECCV), pages 292-310. Springer, 2024. 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 56, + 637, + 294, + 690 + ], + "type": "text", + "content": "[57] Jiqing Zhang, Yuanchen Wang, Wenxi Liu, Meng Li, Jinpeng Bai, Baocai Yin, and Xin Yang. Frame-event alignment and fusion network for high frame rate tracking. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9781-9790, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 294, + 713 + ], + "type": "text", + "content": "[58] Yang Zheng, Adam W Harley, Bokui Shen, Gordon Wetzstein, and Leonidas J Guibas. Pointodyssey: A large-scale" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 361 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 105 + ], + "type": "text", + "content": "synthetic dataset for long-term point tracking. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 19855-19865, 2023. 2, 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 161 + ], + "type": "text", + "content": "[59] Hanyu Zhou, Yi Chang, Haoyue Liu, Wending Yan, Yuxing Duan, Zhiwei Shi, and Luxin Yan. Exploring the common appearance-boundary adaptation for nighttime optical flow. In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[60] Hanyu Zhou, Yi Chang, and Zhiwei Shi. Bring event into rgb and lidar: Hierarchical visual-motion fusion for scene flow. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 26477-26486, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[61] Yi Zhou, Guillermo Gallego, Xiuyuan Lu, Siqi Liu, and Shaojie Shen. Event-based motion segmentation with spatiotemporal graph cuts. IEEE Transactions on Neural Networks and Learning Systems (TNNLS), 34(8):4868-4880, 2021. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 306 + ], + "type": "text", + "content": "[62] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 989-997, 2019. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 361 + ], + "type": "text", + "content": "[63] Zhiyu Zhu, Junhui Hou, and Dapeng Oliver Wu. Cross-modal orthogonal high-rank augmentation for RGB-event transformer-trackers. In IEEE/CVF International Conference on Computer Vision (ICCV), pages 22045-22055, 2023. 1" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7946" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_content_list.json b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..b04a508800133e4bfdfc90485670c0adf95c40b5 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_content_list.json @@ -0,0 +1,1619 @@ +[ + { + "type": "text", + "text": "Event-based Tiny Object Detection: A Benchmark Dataset and Baseline", + "text_level": 1, + "bbox": [ + 135, + 130, + 861, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Nuo Chen $^{1}$ , Chao Xiao $^{1}$ , Yimian Dai $^{2}$ , Shiman He $^{1}$ , Miao Li $^{1}$ , Wei An $^{1\\boxtimes}$ $^{1}$ National University of Defense Technology, $^{2}$ Nankai University", + "bbox": [ + 209, + 186, + 787, + 229 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{chennuo97,anwei}@nudt.edu.cn", + "bbox": [ + 369, + 234, + 627, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 285, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Small object detection (SOD) in anti-UAV task is a challenging problem due to the small size of UAVs and complex backgrounds. Traditional frame-based cameras struggle to detect small objects in complex environments due to their low frame rates, limited dynamic range, and data redundancy. Event cameras, with microsecond temporal resolution and high dynamic range, provide a more effective solution for SOD. However, existing event-based object detection datasets are limited in scale, feature large targets size, and lack diverse backgrounds, making them unsuitable for SOD benchmarks. In this paper, we introduce a Event-based Small object detection (EVSOD) dataset (namely EVUAV), the first large-scale, highly diverse benchmark for anti-UAV tasks. It includes 147 sequences with over 2.3 million event-level annotations, featuring extremely small targets (averaging $6.8 \\times 5.4$ pixels) and diverse scenarios such as urban clutter and extreme lighting conditions. Furthermore, based on the observation that small moving targets form continuous curves in spatiotemporal event point clouds, we propose Event based Sparse Segmentation Network (EV-SpSegNet), a novel baseline for event segmentation in point cloud space, along with a Spatiotemporal Correlation (STC) loss that leverages motion continuity to guide the network in retaining target events. Extensive experiments on the EV-UAV dataset demonstrate the superiority of our method and provide a benchmark for future research in EVSOD. The dataset and code are at https://github.com/ChenYichen9527/Ev-UAV.", + "bbox": [ + 88, + 316, + 485, + 739 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 768, + 220, + 784 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unmanned Aerial Vehicles (UAVs) have made significant contributions to fields such as geological exploration [22] and precision agriculture [28] due to their high mobility and low cost. However, the rapid increase in UAV numbers has also introduced potential threats to public safety, making anti-UAV technologies increasingly critical. Small Object Detection (SOD), which aims to accurately and real-time", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a3b5155344e5a1beedae1b05318592a94a67ff0f6eb755844ede6553091cf76b.jpg", + "image_caption": [ + "Figure 1. Comparison between event camera and RGB camera. The RGB camera can only capture the objects under normal light, while the event camera can capture objects under various extreme lighting conditions. And the event camera can capture the continuous motion trajectory of the small object (shown as the red curve)." + ], + "image_footnote": [], + "bbox": [ + 522, + 286, + 908, + 446 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "locate tiny targets, is fundamental to anti-UAV tasks. However, it remains a challenging problem due to the small size of UAVs and the complexity of their surrounding environments.", + "bbox": [ + 511, + 547, + 906, + 607 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Conventional frame-based cameras suffer from limited frame rates (typically $30 - 60\\mathrm{Hz}$ ), narrow dynamic range, and redundant data capture, hindering the development of SOD [10, 11]. Event cameras [6, 27], bio-inspired sensors that asynchronously capture per-pixel intensity changes with microsecond temporal resolution ( $\\geq 10^{6}\\mathrm{Hz}$ ) and high dynamic range (120dB), offer a promising alternative to SOD under real-world scenarios. Their ability to encode motion as sparse spatio-temporal event streams inherently avoids data redundancy and preserves fine-grained details of high-speed objects, even in challenging illumination.", + "bbox": [ + 511, + 611, + 908, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite this potential, the following challenges hinder their adoption in SOD: First, the small size and limited appearance cues of small objects hinder feature learning, while background clutter degrades detection performance and increases false alarms. Second, current event-based methods [7, 26] often convert events into synchronized image representations for frame-based detection. This disrupts the continuity and sparsity of event data, causing loss of", + "bbox": [ + 511, + 779, + 910, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "temporal information and unnecessary processing of redundant background data. Furthermore, the lack of large-scale, high-quality datasets significantly impedes the development of event-based small moving object detection.", + "bbox": [ + 89, + 90, + 480, + 151 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, in this paper, we aim to address the aforementioned challenges to advance the development of the event-based small moving object (especially for UAVs) detection (EVSOD) by firstly building a large-scale dataset for EVSOD, then developing a novel detection method tailored for EVSOD based on the spatio-temporal event point cloud representation, and finally extensively evaluating various deep feature learning methods for EVSOD with the developed dataset and the proposed method as baseline.", + "bbox": [ + 88, + 152, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Specifically, we construct a new dataset for EVSOD, for the first time, to address the following issues: (1) Existing SOD datasets primarily focus on visible modality [13, 45, 46], infrared modality [38, 49], or the fusion of these two modalities [47], and few research is conducted to explore the EVSOD. (2) Although several event-based object detection datasets [7, 26, 44] have been proposed, they focus on detecting large objects in simple backgrounds, failing to capture real-world scenarios where small objects appear in cluttered environments. (3) Existing event-based object detection datasets [18, 19] lack event-level annotations, restricting the development of algorithms using other event representations (e.g., sparse point clouds and SNNs). The aforementioned issues urge us to introduce the first large-scale dataset with event-level annotations for EVSOD, advancing the development of anti-UAV.", + "bbox": [ + 91, + 291, + 482, + 532 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Then, we propose Event based Sparse Segmentation Network (EV-SpSegNet), which specially designed for efficient and robust EVSOD. We observe that small moving objects in spatiotemporal event point clouds form continuous, elongated curves, which are distinct from the surfaces formed by the background and the discrete points formed by noise. Leveraging this characteristic, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module to extract multi-scale local and long-term contextual features of curve structures. Based on the proposed GDSCA, we construct our network to detect small moving objects. Furthermore, we design a new Spatiotemporal Correlation (STC) loss, which evaluates the local spatiotemporal correlation of events to guide the network in retaining more events with continuous curves.", + "bbox": [ + 89, + 534, + 482, + 760 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, we conduct extensive performance evaluations on the proposed dataset, comparing our method with 13 state-of-the-art algorithms, including 4 frame-based methods [16, 30, 37, 51], 6 event-based methods [7, 17, 24-26, 34], and 3 point cloud segmentation methods [2, 9, 36], enabling fair and reproducible comparisons for EVSOD.", + "bbox": [ + 89, + 762, + 480, + 853 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows: (1) We build the first large-scale dataset EV-UAV, with high diversity for EVSOD. It contains 147 event sequences with", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "event-level annotations, covering challenging scenarios like high-brightness and low-light conditions, with targets averaging 1/50 the size in existing datasets. (2) We propose EVSpSegNet, a novel detection framework that distinguishes small targets from background and noise based on the continuous curves formed by targets. Additionally, we introduce STC loss, which retains target events with continuous motion through local spatiotemporal correlation. (3) Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 511, + 287, + 650, + 304 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Anti-UAV detection dataset", + "text_level": 1, + "bbox": [ + 511, + 313, + 759, + 329 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Currently, event-based object detection datasets primarily focus on autonomous driving [5, 26] and general object detection [39, 41, 43]. There is limited attention given to datasets that are exclusively designed for UAV detection. In Tab. 1, we provide a comprehensive summary of existing datasets, highlighting the scarcity of event-based datasets for UAV object detection.", + "bbox": [ + 511, + 335, + 903, + 441 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early datasets often included UAVs only as a minor subcategory within larger datasets [39, 41, 43]. Due to their limited scale and diversity, these datasets are insufficient for training modern deep learning-based methods. The recently proposed F-UAV-D dataset [19] is specifically designed for event-based UAV object detection. However, it only covers indoor environments, which limits its applicability. NeRDD [18], the first dataset dedicated to event-based UAV detection, considers only large objects against clean backgrounds. This fails to capture real-world scenarios where UAVs often appear as small objects in dynamic and cluttered backgrounds. To address these limitations, we introduce EV-UAV, a dataset designed to better represent challenging real-world scenarios. EV-UAV features a large number of detection scenes with small targets and dynamic backgrounds, providing a robust benchmark for developing and evaluating anti-UAV methods.", + "bbox": [ + 511, + 443, + 905, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Object detection for event cameras", + "text_level": 1, + "bbox": [ + 511, + 710, + 815, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing object detection methods for event cameras can be broadly divided into two categories. The first category focuses on directly processing sparse event data. For example, graph-based methods model events as dynamic spatiotemporal graphs [15, 21, 32], restricting information propagation to a finite set of sparse connections to reduce computational complexity. Similarly, Spiking Neural Network (SNN)-based methods process asynchronous and sparse event streams directly within the network architecture [4, 23, 33, 48]. By transmitting only discrete spike signals between neurons, these methods significantly re", + "bbox": [ + 511, + 734, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7210", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/809bbc2909354d3512eba1e3b743fef7ea2ad73d3de5380bca82193f1d894934.jpg", + "table_caption": [ + "Table 1. Summary of existing event UAV detection datasets. BL: bright light, NL: normal light, LL: low light, MS: multi-scene, MT: multi-target." + ], + "table_footnote": [], + "table_body": "
Dataset#AGV.UAV scaleLabel TypeUAV Sequence Ratio(%)UAV centricLighting conditionsObjectYear
BLNLLLMSMT
VisEvent [40]84×66 pixelsBBox15.97××✓✓✓×2023
EventVOT [44]129×100 pixelsBBox8.41××✓✓✓×2024
EvDET200K [42]68×45 pixelsBBox3.57××✓✓✓✓2024
F-UAV-D [19]-BBox100✓×✓×××2024
NeRDD [18]55×31 pixelsBBox100✓×✓××✓2024
EV-UAV6.8×5.4 pixelsSeg100✓✓✓✓✓✓2025
", + "bbox": [ + 138, + 128, + 859, + 250 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "duce power consumption during training. However, sparse representation-based approaches are limited by their reliance on specialized hardware or suboptimal performance in complex scenarios.", + "bbox": [ + 88, + 273, + 480, + 333 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The second category transforms sparse event streams into image-like representations, enabling compatibility with dense convolutional operations. Early dense-based methods directly used event representations generated from short time windows for detection [3, 20, 29, 50]. These methods, however, only considered event information within a limited time, resulting in limited performance. Later work addressed this issue by incorporating recurrent neural networks [7, 14, 26], which improved detection performance by capturing temporal dependencies. Despite these advancements, such methods consume significant computational resources for event transformation and processing artificially introduced redundant background data, while struggling to detect small targets due to their lack of distinct texture and contour features. In this work, we observe that small targets form continuous curves in spatiotemporal event point clouds, which are distinct from background and noise. Leveraging this characteristic, we design EVSpSegNet, which directly segments continuous target trajectories from sparse point clouds. This approach avoids the computational overhead caused by converting events into image-like representations. Moreover, the distinct curve features help the network better detect small targets.", + "bbox": [ + 91, + 335, + 483, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. EV-UAV benchmark", + "text_level": 1, + "bbox": [ + 89, + 699, + 287, + 714 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 724, + 232, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event cameras work totally differently from frame-based cameras with fixed frame rates. Specifically, each pixel of an event camera independently responds to changes in light intensity. When the logarithmic brightness change reaches a predefined threshold, an event $E = (x,y,t,p)$ is triggered. In this tuple, $x$ and $y$ represent the pixel coordinates, $p$ indicates the polarity of the brightness change (positive or negative), and $t$ is the timestamp of the event.", + "bbox": [ + 89, + 748, + 482, + 868 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike previous annotations for object detection [7, 26] or semantic segmentation [1, 8, 35], we do not convert", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "events into synchronized frames. Instead, we segment events directly in the point cloud space, event by event. The sources of events can be categorized into three types: events generated by the target $(E_T)$ , events generated by the background $(E_B)$ , and events caused by noise $(E_N)$ . This can be represented as: $E = E_T + E_B + E_N$ . Our goal is to segment all events generated by target motion from the raw event stream. The final annotation is a set of discrete events $E_{label} \\in E_T$ .", + "bbox": [ + 511, + 273, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Data Collection and Annotations", + "text_level": 1, + "bbox": [ + 511, + 421, + 802, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.1. Data Collection", + "text_level": 1, + "bbox": [ + 511, + 443, + 668, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We collect data using a DAVIS346 camera mounted on a tripod to monitor a fixed range of scenes. This camera has a spatial resolution of $346 \\times 260$ and a temporal resolution of up to $10^{6}$ HZ. Our data collection spans from May 1, 2023, to June 1, 2024, covering various times of the day, seasons, and weather conditions. To accurately reflect the movement patterns of UAVs and the complexity of real-world scenarios, we capture UAVs performing diverse actions in various environments. These actions include forward and backward motion, lateral movements, ascending and descending maneuvers, as well as complex combinations of these behaviors. The dataset encompasses three different lighting conditions (normal, bright and low light) and various backgrounds, such as buildings, mountains, forests, urban areas, clouds, water surfaces, etc. We carefully annotated 147 sequences of data, totaling over 20.3 million target events.", + "bbox": [ + 511, + 463, + 908, + 705 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.2.Event by Event Annotation", + "text_level": 1, + "bbox": [ + 511, + 715, + 750, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Due to the asynchronous and sparse output of event cameras, manually annotating per-event data is highly time-consuming and inefficient. Existing event camera datasets are typically recorded using a combination of event cameras and standard RGB cameras. Labels are first extracted from the frame-based cameras and then transferred to the pixel coordinates of the event cameras through geometric transformations. However, this labeling strategy is limited to producing low frame rate bounding box annotations, which cannot match the high-frequency output characteristics of event cameras or meet the fine-grained, event-", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7211", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/bb4e18c1f386c3ffd75789b83a3b4420cb5fcdd628f839db99905201e67a639c.jpg", + "image_caption": [ + "Train" + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 230, + 152 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a12fba785f58aa730c6b3c85c4952ab984acd82577f79bc30de6ac2291fbf524.jpg", + "image_caption": [ + "Val" + ], + "image_footnote": [], + "bbox": [ + 98, + 164, + 230, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d3a5c188208442400bf692ee4ae2dc69452ad31868294460d1dcd34ee182ddb6.jpg", + "image_caption": [ + "Test", + "(a) Position Distribution", + "Extremely tiny" + ], + "image_footnote": [], + "bbox": [ + 98, + 234, + 230, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/145dfde06b8a68d89d2e4cca763629db5c5f1567a000fc30c46659e9a4e4c782.jpg", + "image_caption": [ + "45.1%", + "34.1%", + "Tiny", + "20.8%" + ], + "image_footnote": [], + "bbox": [ + 235, + 90, + 454, + 262 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe73cf2aa51bbd4e501e8a8893dc97cb83a15c2e97e4c2a0db94318a72c4fecd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 459, + 90, + 575, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/889655e607a4b646daae5327cfcb31bd76ffd3f9e910d904065a7527383b8891.jpg", + "image_caption": [ + "illumination:highlight $\\rightarrow$ lowlight" + ], + "image_footnote": [], + "bbox": [ + 580, + 90, + 687, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/49346c06e4c8037fef4e9aa5a23f6d2dc57bcf49b0d6a40dd431d2e0b9985e9e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 697, + 90, + 808, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/870e145cc8dab9f7fb2a70b4018e916a280cd6b824e278ace52e35af75cea424.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 812, + 90, + 928, + 150 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/34f6d63a72d1a01f7892a1a092dc6e29d0de00ae2dc8a5ba8d98d73494779458.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 460, + 162, + 575, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/1f38712daafa71fc4d3801a430ff03b8f9778f21e103f6de445f2f915593bd7e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 576, + 162, + 692, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/852905a2a8ff0cba2ffc1c80a28d79d3dae825bbbe5a8c066a3d9c02c0807e6b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 162, + 807, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/164bcf02ee65a9a17b2a8865cc3fb66c48bc739fc6cd5793b760a69bfdddb298.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 162, + 928, + 222 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/773817104496077a6772820be3efa6df8132a679a2bddf6972a2d00f4b28fdb2.jpg", + "image_caption": [ + "Small", + "(b) Target size Distribution" + ], + "image_footnote": [], + "bbox": [ + 459, + 234, + 573, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9ab140ff0971fb1ac46dffe06a16cb49f4fa9ed0b03c82b9a002eb84494b7181.jpg", + "image_caption": [ + "Target Scale: small→large", + "(c) Diversity of EV-UAV" + ], + "image_footnote": [], + "bbox": [ + 576, + 234, + 692, + 304 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a5cb2b9a2e6743d4664478abe835cbba71d435eb9d4f8aa284f00db6187b400f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 694, + 234, + 807, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/3875e344e40fc6e401f948ab37afaf21856314b9f8907d62fc36493e1204a0fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 810, + 234, + 928, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/f89e97b9cd390ae3be73f865acddf7d4de2f875b41acee2ebc9022b8cfe5e14f.jpg", + "image_caption": [ + "Figure 2. Illustration of the proposed dataset EV-UAV. (a) Position distribution of EV-UAV. (b) Distribution of noise level and illumination of the proposed dataset. (c) EV-UAV collects diverse event streams across various conditions of illumination and target number and scale. N represents the noise level, reflecting the proportion of events in spatiotemporal space. T denotes the number of targets. S indicates the size of targets in the image plane.", + "Figure 3. Event-by-event annotation method. Firstly, the events are accumulated into event frames, and targets are annotated with 2D bounding boxes on the event frames. Then, according to the time interval $\\Delta t$ corresponding to each frame, the 2D annotations on the image plane are extended into 3D bounding box annotations in XYT 3D space, and all events within the 3D boxes are filtered as targets." + ], + "image_footnote": [], + "bbox": [ + 94, + 410, + 488, + 511 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "by-event labeling requirements. Although manual event-by-event labeling could theoretically improve annotation accuracy, the process is highly labor-intensive and time-consuming, making it impractical for real-world applications. To address this issue, we propose a novel method that generates approximate event-by-event annotations using bounding box labels, as illustrated in the Fig. 3. This approach not only achieves fine-grained event-level annotations, enabling microsecond-level object detection, but also allows for degradation into bounding box annotations at any given time, thus supporting the learning requirements of some asynchronous event-based object detection algorithms.", + "bbox": [ + 88, + 626, + 482, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Benchmark Features and Statistics", + "text_level": 1, + "bbox": [ + 89, + 832, + 393, + 848 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dataset Splitting. The EV-UAV dataset is divided into training (99 sequences), validation (24 sequences), and test sets (24 sequences), totaling 147 event sequences.", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Position Distribution. As shown in Fig. 2(a), the targets are mostly located in the central area of the image. The larger horizontal variance suggests that the drone's movement is predominantly horizontal. Additionally, the range of target movements in the validation set is more varied. The fluctuations in the validation set are greater than those in the test set in corresponding directions.", + "bbox": [ + 511, + 409, + 906, + 515 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Small-Scale Targets. In real-world scenarios, due to the long imaging distance and the small size of UAVs, drone targets appear at very small scales in the images. The general scale ranks $^1$ of [31] are no longer applicable. Following previous work [47], we further divide small targets into three levels: extremely tiny $\\in [1, 8^2)$ , tiny $\\in [8^2, 16^2)$ , and small $\\in [16^2, \\infty)$ . Fig. 2(b) shows the number of annotations for each target category. It can be observed that the scale of all targets is smaller than $32^2$ , with extremely tiny targets accounting for the largest proportion, approximately $45\\%$ . Compared to other UAV detection datasets, our EV-UAV dataset has the smallest average UAV scale, as indicated in Tab. 1.", + "bbox": [ + 511, + 515, + 908, + 710 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Rich Diversity. To represent real-world anti-UAV target detection scenarios, we endow EV-UAV with diversity. This helps the dataset cover various complex situations in anti-UAV detection both in time and space. As shown in Fig. 2(c), the diversity of EV-UAV is reflected not just in illumination levels but also in background complexity, target quantity, and scale.", + "bbox": [ + 511, + 710, + 908, + 818 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 825, + 697, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this work, we evaluate the event-by-event segmentation performance of the algorithm using Intersection over Union", + "bbox": [ + 511, + 848, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "${}^{1}$ small $\\in \\left\\lbrack {1,{32}^{2}}\\right)$ ,medium $\\in \\left\\lbrack {{32}^{2},{96}^{2}}\\right)$ ,and large $\\in \\left\\lbrack {{96}^{2},\\infty }\\right)$ .", + "bbox": [ + 529, + 885, + 854, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7212", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/31fd6e9fb455be82f46e0d7537741caf16547b592e8df14c069dbb4fe4cccd40.jpg", + "image_caption": [ + "Figure 4. (a) The overview of the proposed Ev-SpSegNet. The network is built on a U-shaped architecture consisting of symmetric encoder and decoder. (b) The GDSCA module. The grouped dilated sparse convolution block is applied to features from different channels to extract local temporal features at multiple scales. The Sp-SE module is used to integrate these features. And the Patch Attention block is used to downsample voxel features and enable global context feature interaction." + ], + "image_footnote": [], + "bbox": [ + 102, + 88, + 897, + 388 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(IoU) and accuracy (ACC), and assess the localization performance using probability of detection $(P_d)$ and false alarm rate $(F_a)$ . More details regarding evaluation metrics can be found in the supplementary material.", + "bbox": [ + 88, + 481, + 486, + 542 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Method", + "text_level": 1, + "bbox": [ + 89, + 556, + 183, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Frame-based detectors require converting events into synchronized image-like representations, which compresses the temporal information of events and makes it even more difficult to detect small targets that already lack distinct appearance cues. In raw event data, events generated by moving targets exhibit strong spatiotemporal correlations, forming elongated continuous curves in event point cloud space. Leveraging this characteristic, we propose EV-SpSegNet, which directly segments events generated by moving targets in sparse event point clouds. Additionally, we introduce a novel spatiotemporal correlation loss that evaluates the local spatiotemporal correlation of events to guide the network in retaining target events with continuous curves.", + "bbox": [ + 88, + 580, + 482, + 777 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. EV-SpSegNet", + "text_level": 1, + "bbox": [ + 89, + 787, + 232, + 803 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Small moving targets form distinct continuous curves in spatiotemporal event point clouds. However, networks struggle to extract their features because these curves span long temporal ranges, and traditional convolution operations with limited receptive fields cannot effectively capture global features. While attention mechanisms can ex", + "bbox": [ + 88, + 810, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tract global context, their quadratic computational complexity restricts their ability to process large-scale event data. To effectively extract features from continuous curves in spatiotemporal event point clouds, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module, as shown in Fig. 4(b). This module first captures local multiscale features through the Grouped Dilated Sparse Convolution (GDSC) Block and then enables global feature interaction across patches using patch attention. Specifically, the GDSC Block groups input features along the channel dimension, with each group applying convolutions of different dilation rates to extract multi-scale local features. These features are then fused in the Sparse Squeeze and Excitation (Sp-SE) block. Finally, the patch attention block divides the point cloud into larger sub-regions and implements global feature interaction between sub-regions using self-attention. This allows the network to model the differences between targets and background noise in greater detail, enhancing its ability to distinguish moving targets.", + "bbox": [ + 511, + 481, + 906, + 770 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Spatiotemporal Correlation Loss", + "text_level": 1, + "bbox": [ + 511, + 781, + 803, + 797 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In general image segmentation tasks, the loss function used for training the network is typically the standard pixel-wise binary cross-entropy (BCE) loss:", + "bbox": [ + 511, + 804, + 906, + 849 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nL _ {b c e} (p, y) = \\left\\{ \\begin{array}{c} - \\log (p), y = 1 \\\\ - \\log (1 - p), y = 0 \\end{array} \\right.. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 863, + 906, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7213", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37d6bcf7677342d7d683588c5fd4ea7dc5e26760ed8aa3de541b0c357eaaaf2a.jpg", + "image_caption": [ + "Figure 5. (a) The principle of spatiotemporal correlation loss. (b) Small object segmentation results of BCE loss and STC loss." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 488, + 195 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Although EV-SpSegNet computes structured outputs and focuses on large neighborhood information to better utilize long-term features, losses like BCE loss process each pixel independently. They do not consider the structure of neighborhoods, such as the spatiotemporal continuity of objects, which is crucial for describing the continuous curves formed by objects. As shown in Fig. 5(a), compared to the isolated noise $p_2$ , the event $p_1$ with more high-confidence supporting events is more likely to be a true target. However, misclassifying both of them only results in a low coat of BCE loss.", + "bbox": [ + 88, + 255, + 482, + 419 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address this issue, we introduce a spatiotemporal correlation loss that encourages the network to retain more events with high spatiotemporal correlation while discarding more isolated noise. Specifically, this loss calculates the sum of the confidence of all supporting events within a $k \\times k \\times \\tau$ neighborhood around the events, and defines it as the spatiotemporal correlation weight. We incorporate this weight into the cross-entropy loss and the spatiotemporal correlation loss is defined as follows:", + "bbox": [ + 89, + 421, + 483, + 556 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nL _ {s t c} (p, y) = \\left\\{ \\begin{array}{c} - w _ {s t c} ^ {\\gamma} \\log (p), y = 1 \\\\ - \\left(1 - w _ {s t c}\\right) ^ {\\gamma} \\log (1 - p), y = 0 \\end{array} , \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 561, + 482, + 603 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the $y \\in \\{0,1\\}$ and $p \\in (0,1)$ represent the segmentation label and prediction result of the event, respectively. $\\gamma$ is a hyperparameters to control the influence of spatiotemporal correlations. $w_{stc}$ is the spatiotemporal correlation weight of each event, which is defined as:", + "bbox": [ + 89, + 609, + 483, + 686 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nw _ {s t c} (p) = \\operatorname {s i g m o i d} \\left(\\sum_ {e ^ {j} \\in V ^ {k \\tau}} p ^ {j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 693, + 482, + 726 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $V^{k\\tau}$ represents the neighborhood of events with a size of $k \\times k \\times \\tau$ . As illustrated in Fig. 5 (b1-b2), the STC loss preserves the continuous event trajectories generated by moving objects and eliminates false alarms caused by misclassification.", + "bbox": [ + 89, + 734, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 89, + 848, + 297, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We used 8-second events as input and set the size of each voxel to $1\\mathrm{pixel} \\times 1\\mathrm{pixel} \\times 1\\mathrm{ms}$ . The batch size was set to", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "1, and all layers of the network were randomly initialized. We trained our network using the Adam optimizer [12] for 50 epochs with the learning rate linearly decaying from 1e-2 to 1e-3. The model was trained on the predefined training set of EV-UAV dataset, and we chose the models based on their best performance on the validation set and evaluate them on the test set. The hyperparameters $k$ , $\\tau$ , and $\\gamma$ of the STC loss are set to 3, 5, and 2, respectively. All models were implemented with PyTorch on a single NVIDIA RTX 3090 GPU.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Benchmark Comparisons", + "text_level": 1, + "bbox": [ + 511, + 252, + 746, + 268 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare the proposed method with several state-of-the-art methods, including 4 frame-based generic object detection methods (i.e., SSD [16], Faster RCNN [30], Deformable DETR [51], YOLOV10 [37]), 6 event based methods (i.e., RVT [7], SAST [25], GET [24], RED [26], EMS-YOLO [34], Spike-YOLO [17]), and 3 point cloud segmentation-based methods (i.e., KPCov, RandLA-Net, COSeg). We convert events within a $50~ms$ time window into event frames [20] as input for frame-based methods.", + "bbox": [ + 511, + 273, + 906, + 410 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. As shown in Tab. 2, our EV-SpSegNet achieve the best detection performance with all evaluation metrics. We attribute this to two reasons. On the one hand, the GDSCA module combines local multiscale information with global context, enabling the network to better model the differences between targets and background noise. This enhances the ability of the network to distinguish moving targets. On the other hand, the spatiotemporal correlation loss encourages the network to identify events with high spatiotemporal correlation, improving detection performance. Additionally, by using sparse convolution to build the network, our method can inference 8 seconds of event data in $35.9~ms$ , significantly outperforming other methods.", + "bbox": [ + 511, + 410, + 905, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Results. Qualitative results of different methods are shown in Fig. 6. It can be observed that our method can detect the most moving targets in all scenes, while the compared methods exhibit more or less missed detections and false positives. Particularly in scenes with complex backgrounds, such as the second and fourth rows with strong noise, the improvement of our method is more significant. Much to our surprise, our method even detects some targets that were missed by manual annotation due to strong background clutter, as shown in the third and fourth rows. We attribute this to the spatiotemporal correlation loss, which encourages the network to retain events with high spatiotemporal correlation.", + "bbox": [ + 511, + 622, + 906, + 819 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 828, + 678, + 843 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3.1. Contribution of Components", + "text_level": 1, + "bbox": [ + 511, + 851, + 761, + 866 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Tab. 3, adding the GDSC block to the baseline improves the model performance in terms of IoU, ACC,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7214", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e16c7b6640ff3e05a7d0fbdc7eeb581e10029a4778135be3f72acd8e81bef8c9.jpg", + "table_caption": [ + "Table 2. Quantitative comparison of the proposed method to state-of-the-art methods." + ], + "table_footnote": [ + "The bold and the underline represent the best and second-best performance, respectively." + ], + "table_body": "
MethodsPublicationsEvent Rep.TemporalIoU(%)↑ACC(%)↑\\(P_d\\)(%)↑\\(F_a(10^{\\text{~-4~}})\\downarrow\\)#Params.Runtime(ms)
SSD [16]ECCV 2016Event CountNo25.3128.5626.31486.6325.2M2113
Faster RCNN [30]TPAMI 2016Event CountNo26.9329.6827.39689.6841.2M3962
DETR [51]ICLR 2020Event CountNo30.3533.6331.64631.3739.8M3136
YOLOV10-S [37]NIPS 2025Event CountNo32.5533.3932.18589.677.3M1627
EMS-YOLO [34]ICCV 2023SNNYes36.7742.9250.68112.363.3M1229
Spike-YOLO [17]ECCV 2024SNNYes43.9448.2659.6255.3869.0M1883
GET [24]ICCV 2023Groupv TokenYes40.3148.9160.7346.3518.4M2168
RED [26]NeurIPS 2020Voxel GridYes35.9945.5453.76102.2724.1M3427
RVT [7]CVPR 2023Voxel GridYes43.2151.3860.3555.689.9M1737
SAST [25]CVPR 2024Voxel GridYes34.3140.2251.21150.3218.5M3075
KPCov [36]ICCV 2019PointsYes48.1957.2868.5916.3250.1M562
RandLA-Net [9]CVPR 2020PointsYes50.3259.2970.566.951.2M353
COSeg [2]CVPR 2024PointsYes51.8960.9371.329.2123.4M364
Ours-PointsYes55.1865.0277.531.634.0M35.9
", + "bbox": [ + 91, + 113, + 916, + 333 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed49ab74805cc3b98d74b4d3788f92edb573fcbc12a5600795508baf07ae0b4e.jpg", + "image_caption": [ + "Figure 6. Qualitative results of different methods. Our method accurately detects small moving targets, while other methods produce numerous missed detections and false alarms." + ], + "image_footnote": [], + "bbox": [ + 122, + 359, + 875, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and $P_{d}$ by 1.37, 2.43 and 1.18, which demonstrates that extracting multi-scale local features using GDSC is beneficial. However, adding only the PA block causes a performance drop. This is because downsampling is applied before computing global attention, and without sufficient local receptive fields, the model cannot effectively capture global features, leading to feature confusion and performance degradation. It is worth noting that when the GDSC block and PA block are used together, the performance improves significantly. This indicates that capturing sufficient local features within patches before enabling global feature interaction is", + "bbox": [ + 89, + 732, + 485, + 898 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "essential. Additionally, using the STC loss consistently improves model performance, which shows that guiding the network to retain events with high spatiotemporal correlation through STCLoss is effective.", + "bbox": [ + 511, + 732, + 906, + 792 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3.2. The Influence of the Number of Branches", + "text_level": 1, + "bbox": [ + 511, + 805, + 846, + 819 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As depicted in Tab. 4, we conduct an ablation experiment to study the influence of the number of branches in GDSC block on the final detection performance. It can be observed that the network achieves best performance when the number of branches is set to 4. This is because too", + "bbox": [ + 511, + 825, + 906, + 898 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7215", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9bfdd2b2523290491d4058ad51fb33842a6cdab61505279bc0ee3e08929317ca.jpg", + "table_caption": [ + "Table 3. Ablation study on framework scheme and loss function." + ], + "table_footnote": [], + "table_body": "
GDSCPALstcIoU↑ACC↑Pd↑Fa↓#Params.
51.3660.2171.944.815.6M
✓52.7362.6473.124.283.8M
✓51.2659.5371.674.315.8M
✓53.0761.0174.133.665.6M
✓✓53.6264.0276.761.934.0M
✓✓✓55.1865.0277.531.634.0M
", + "bbox": [ + 91, + 113, + 480, + 220 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5ba3846b7103981576a36b70cc00833b02e1c1eae762297b1b9cd23bad475b3a.jpg", + "table_caption": [ + "Table 4. The influence of the number of branches in GDSC block." + ], + "table_footnote": [], + "table_body": "
#BranchesIoU↑ACC↑Pd↑Fa↓#Params.
150.2162.3673.652.975.6M
251.4164.1675.212.534.5M
353.5264.2776.662.014.3M
455.1865.0277.531.634.0M
553.1264.3176.532.113.8M
", + "bbox": [ + 91, + 258, + 483, + 351 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0d7b90d21ce889253175f5b40e3da3723828e23c4de910fcc6b16014b018c90.jpg", + "table_caption": [ + "Table 5. Combination of convolutions with different dilation rates." + ], + "table_footnote": [], + "table_body": "
Dilated ratesIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
(1,2,3,4)55.1865.0277.531.63
(1,2,3,5)53.2164.1176.932.41
(1,3,5,7)52.7564.0276.894.61
(1,3,5,9)51.3563.9275.924.76
", + "bbox": [ + 91, + 391, + 486, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "many branches result in insufficient features assigned to each channel, making it difficult to capture features at the corresponding scale. On the other hand, too few branches lead to insufficient dilation rates, preventing the network from capturing features at multiple scales.", + "bbox": [ + 89, + 496, + 482, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.3. The Influence of Dilation Rate Combinations", + "text_level": 1, + "bbox": [ + 89, + 580, + 447, + 594 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We next conduct an ablation study to analyze the relationship between detection performance and different dilation rate combinations. Tab. 5 illustrates that setting the dilation rate to (1,2,3,4) achieves the best performance. This indicates that the dilation rate is not always better when larger. An excessively large dilation rate causes the corresponding events to be too far apart, making it difficult to effectively determine whether these events have temporal correlation.", + "bbox": [ + 89, + 599, + 482, + 720 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3.4. The Effectiveness of STC Loss", + "text_level": 1, + "bbox": [ + 89, + 729, + 346, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further validate the generalizability of our proposed spatiotemporal correlation Loss, we integrate it with several state-of-the-art methods, including KPConv [36], RandLA-Net [9], and COSeg [2]. As shown in Tab. 6, the addition of $L_{stc}$ consistently improves the performance of all methods. For example, KPConv with $L_{stc}$ achieves improvements of $2.13\\%$ , $1.75\\%$ and $1.4\\%$ in terms of IoU, ACC and $P_d$ , respectively, and the false alarm rate decreases by $5.72\\%$ . These results show that $L_{stc}$ effectively enhances the ability of existing methods to capture spatiotemporal correlations,", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/046b71c7356eaa8ea785580fb2d1b841d31c5014c40da400650767a03a97e737.jpg", + "table_caption": [ + "Table 6. The performance of spatiotemporal correlation loss function on different methods." + ], + "table_footnote": [ + "* represent the methods with STC loss." + ], + "table_body": "
MethodIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
KPConv48.1957.2868.5916.32
KPConv*50.32(+2.13)59.03(+1.75)69.99(+1.40)10.60(-5.72)
RandLA50.3259.2970.569.21
RandLA*51.15(+0.83)60.63(+1.34)72.36(+2.37)8.91(-1.69)
COSeg51.8960.9371.326.95
COSeg*53.12(+1.23)62.31(0.99)73.96(+2.64)5.86(-1.09)
", + "bbox": [ + 514, + 128, + 911, + 233 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "leading to better segmentation accuracy and reduced false alarms. The consistent performance gains across diverse methods prove the robustness and general applicability of our proposed loss function. In addition, we explore the selection of optimal hyperparameters for the STC loss. For more details, please refer to the supplementary material.", + "bbox": [ + 511, + 275, + 906, + 367 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Discussion and Limitations", + "text_level": 1, + "bbox": [ + 511, + 380, + 764, + 396 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Although our EV-UAV accurately replicates real-world scenarios for detecting tiny UAVs in the wild, event cameras only capture motion information. When targets are stationary or move slowly, event cameras produce no output, leading to detection failures. In contrast, frame-based cameras can directly provide detailed texture information (i.e., absolute brightness) in static scenes, making them complementary to event cameras. We believe that developing a dual-modal dataset and extending our method to incorporate multi-modal inputs to maximize object detection accuracy represents a promising direction for future work.", + "bbox": [ + 511, + 406, + 906, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 587, + 633, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce EV-UAV, the first large-scale benchmark for anti-UAV tasks. EV-UAV is a highly challenging benchmark, featuring diverse scenarios and extremely small targets. It provides event-level fine-grained annotations, supporting algorithms with different event representations. Furthermore, based on the observation that small moving targets form continuous curves in event point cloud space, we propose EV-SpSegNet, a novel baseline for segmenting target motion trajectories from sparse event point clouds, along with a Spatiotemporal Correlation (STC) loss that leverages spatiotemporal local correlation to guide the network in retaining more target events. Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD.", + "bbox": [ + 511, + 612, + 906, + 853 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement: This work was supported by the Young Elite Scientist Sponsorship Program by China Association for Science and Technology (NO.YESS20240427).", + "bbox": [ + 511, + 854, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7216", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Inigo Alonso and Ana C Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In CVPRW, pages 0–0, 2019. 3", + "[2] Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, and Serge Belongie. Rethinking few-shot 3d point cloud semantic segmentation. In CVPR, pages 3996-4006, 2024. 2, 7, 8", + "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Networks and Learning Systems, 25(2):407-417, 2013. 3", + "[4] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IEEE International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2022. 2", + "[5] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 2", + "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. 44(1):154-180, 2020. 1", + "[7] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1, 2, 3, 6, 7", + "[8] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 3", + "[9] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 2, 7, 8", + "[10] Bo Huang, Jianan Li, Junjie Chen, Gang Wang, Jian Zhao, and Tingfa Xu. Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild. 46 (5):2852-2865, 2023. 1", + "[11] Nan Jiang, Kuiran Wang, Xiaoke Peng, Xuehui Yu, Qiang Wang, Junliang Xing, Guorong Li, Jian Zhao, Guodong Guo, and Zhenjun Han. Anti-uav: A large multi-modal benchmark for uav tracking. arXiv preprint arXiv:2101.08466, 2021. 1", + "[12] Diederik P Kingma and Jimmy Lei Ba. Adam: Amethod for stochastic optimization. In ICLR, pages 1-15, 2014. 6", + "[13] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 2", + "[14] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE TIP, 31:2975-2987, 2022. 3", + "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[16] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 2, 6, 7", + "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In ECCV, 2024. 2, 6, 7", + "[18] Gabriele Magrini, Federico Becattini, Pietro Pala, Alberto Del Bimbo, and Antonio Porta. Neuromorphic drone detection: an event-rgb multimodal approach. arXiv preprint arXiv:2409.16099, 2024. 2, 3", + "[19] Jakub Mandula, Jonas Kühne, Luca Pascarella, and Michele Magno. Towards real-time fast unmanned aerial vehicle detection using dynamic vision sensors. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1-6. IEEE, 2024. 2, 3", + "[20] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso Garcia, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In CVPR, pages 5419-5427, 2018. 3, 6", + "[21] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In CVPR, pages 14414-14423, 2020. 2", + "[22] Francesco Nex and Fabio Remondino. Uav for 3d mapping applications: a review. Applied geomatics, 6:1-15, 2014. 1", + "[23] Federico Paredes-Valles, Kirk YW Scheper, and Guido CHE De Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE TPAMI, 42(8):2051-2064, 2019. 2", + "[24] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: group event transformer for event-based vision. In ICCV, pages 6038-6048, 2023. 2, 6, 7", + "[25] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In CVPR, pages 16794-16804, 2024. 6, 7", + "[26] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 33:16639-16652, 2020. 1, 2, 3, 6, 7", + "[27] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1", + "[28] Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, Thomas Lagkas, and Ioannis Moscholios. A compilation of uav applications for precision agriculture. Computer Networks, 172:107148, 2020. 1", + "[29] Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In BMVC, 2017. 3", + "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7217", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 4", + "[32] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, pages 12371-12381, 2022. 2", + "[33] Sumit Bam Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. arXiv preprint arXiv.1810.08646, 2018. 2", + "[34] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In ICCV, pages 6555-6565, 2023. 2, 6, 7", + "[35] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, pages 341-357. Springer, 2022. 3", + "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 2, 7, 8", + "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. NeurIPS, 37:107984-108011, 2025. 2, 6, 7", + "[38] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In ICCV, pages 8509-8518, 2019. 2", + "[39] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 2", + "[40] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 3", + "[41] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 2", + "[42] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 3", + "[43] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, pages 19248-19257, 2024. 2", + "[44] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and a novel baseline. In CVPR, pages 19248-19257, 2024. 2,3", + "[45] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543. Springer, 2022. 2", + "[46] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 2", + "[47] Xinyi Ying, Chao Xiao, Ruojing Li, Xu He, Boyang Li, Xu Cao, Zhaoxu Li, Yingqian Wang, Mingyuan Hu, Qingyu Xu, et al. Visible-thermal tiny object detection: A benchmark dataset and baselines. arXiv preprint arXiv:2406.14482, 2024. 2, 4", + "[48] Emre O. Neftci; Hesham Mostafa; Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, pages 51-63, 2019. 2", + "[49] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In CVPR, pages 877-886, 2022. 2", + "[50] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, pages 989-997, 2019. 3", + "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 516 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7218", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_model.json b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0093cf57b56691bb8c845941c485add6c7350b20 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_model.json @@ -0,0 +1,2321 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.137, + 0.131, + 0.862, + 0.152 + ], + "angle": 0, + "content": "Event-based Tiny Object Detection: A Benchmark Dataset and Baseline" + }, + { + "type": "text", + "bbox": [ + 0.21, + 0.188, + 0.788, + 0.23 + ], + "angle": 0, + "content": "Nuo Chen\\(^{1}\\), Chao Xiao\\(^{1}\\), Yimian Dai\\(^{2}\\), Shiman He\\(^{1}\\), Miao Li\\(^{1}\\), Wei An\\(^{1\\boxtimes}\\) \n\\(^{1}\\)National University of Defense Technology, \\(^{2}\\)Nankai University" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.235, + 0.629, + 0.248 + ], + "angle": 0, + "content": "{chennuo97,anwei}@nudt.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.286, + 0.327, + 0.302 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.318, + 0.486, + 0.741 + ], + "angle": 0, + "content": "Small object detection (SOD) in anti-UAV task is a challenging problem due to the small size of UAVs and complex backgrounds. Traditional frame-based cameras struggle to detect small objects in complex environments due to their low frame rates, limited dynamic range, and data redundancy. Event cameras, with microsecond temporal resolution and high dynamic range, provide a more effective solution for SOD. However, existing event-based object detection datasets are limited in scale, feature large targets size, and lack diverse backgrounds, making them unsuitable for SOD benchmarks. In this paper, we introduce a Event-based Small object detection (EVSOD) dataset (namely EVUAV), the first large-scale, highly diverse benchmark for anti-UAV tasks. It includes 147 sequences with over 2.3 million event-level annotations, featuring extremely small targets (averaging \\(6.8 \\times 5.4\\) pixels) and diverse scenarios such as urban clutter and extreme lighting conditions. Furthermore, based on the observation that small moving targets form continuous curves in spatiotemporal event point clouds, we propose Event based Sparse Segmentation Network (EV-SpSegNet), a novel baseline for event segmentation in point cloud space, along with a Spatiotemporal Correlation (STC) loss that leverages motion continuity to guide the network in retaining target events. Extensive experiments on the EV-UAV dataset demonstrate the superiority of our method and provide a benchmark for future research in EVSOD. The dataset and code are at https://github.com/ChenYichen9527/Ev-UAV." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.77, + 0.222, + 0.785 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Unmanned Aerial Vehicles (UAVs) have made significant contributions to fields such as geological exploration [22] and precision agriculture [28] due to their high mobility and low cost. However, the rapid increase in UAV numbers has also introduced potential threats to public safety, making anti-UAV technologies increasingly critical. Small Object Detection (SOD), which aims to accurately and real-time" + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.287, + 0.909, + 0.448 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.459, + 0.907, + 0.529 + ], + "angle": 0, + "content": "Figure 1. Comparison between event camera and RGB camera. The RGB camera can only capture the objects under normal light, while the event camera can capture objects under various extreme lighting conditions. And the event camera can capture the continuous motion trajectory of the small object (shown as the red curve)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.907, + 0.608 + ], + "angle": 0, + "content": "locate tiny targets, is fundamental to anti-UAV tasks. However, it remains a challenging problem due to the small size of UAVs and the complexity of their surrounding environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.779 + ], + "angle": 0, + "content": "Conventional frame-based cameras suffer from limited frame rates (typically \\(30 - 60\\mathrm{Hz}\\)), narrow dynamic range, and redundant data capture, hindering the development of SOD [10, 11]. Event cameras [6, 27], bio-inspired sensors that asynchronously capture per-pixel intensity changes with microsecond temporal resolution (\\(\\geq 10^{6}\\mathrm{Hz}\\)) and high dynamic range (120dB), offer a promising alternative to SOD under real-world scenarios. Their ability to encode motion as sparse spatio-temporal event streams inherently avoids data redundancy and preserves fine-grained details of high-speed objects, even in challenging illumination." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.911, + 0.902 + ], + "angle": 0, + "content": "Despite this potential, the following challenges hinder their adoption in SOD: First, the small size and limited appearance cues of small objects hinder feature learning, while background clutter degrades detection performance and increases false alarms. Second, current event-based methods [7, 26] often convert events into synchronized image representations for frame-based detection. This disrupts the continuity and sparsity of event data, causing loss of" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7209" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.152 + ], + "angle": 0, + "content": "temporal information and unnecessary processing of redundant background data. Furthermore, the lack of large-scale, high-quality datasets significantly impedes the development of event-based small moving object detection." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.154, + 0.482, + 0.289 + ], + "angle": 0, + "content": "Therefore, in this paper, we aim to address the aforementioned challenges to advance the development of the event-based small moving object (especially for UAVs) detection (EVSOD) by firstly building a large-scale dataset for EVSOD, then developing a novel detection method tailored for EVSOD based on the spatio-temporal event point cloud representation, and finally extensively evaluating various deep feature learning methods for EVSOD with the developed dataset and the proposed method as baseline." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.534 + ], + "angle": 0, + "content": "Specifically, we construct a new dataset for EVSOD, for the first time, to address the following issues: (1) Existing SOD datasets primarily focus on visible modality [13, 45, 46], infrared modality [38, 49], or the fusion of these two modalities [47], and few research is conducted to explore the EVSOD. (2) Although several event-based object detection datasets [7, 26, 44] have been proposed, they focus on detecting large objects in simple backgrounds, failing to capture real-world scenarios where small objects appear in cluttered environments. (3) Existing event-based object detection datasets [18, 19] lack event-level annotations, restricting the development of algorithms using other event representations (e.g., sparse point clouds and SNNs). The aforementioned issues urge us to introduce the first large-scale dataset with event-level annotations for EVSOD, advancing the development of anti-UAV." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.761 + ], + "angle": 0, + "content": "Then, we propose Event based Sparse Segmentation Network (EV-SpSegNet), which specially designed for efficient and robust EVSOD. We observe that small moving objects in spatiotemporal event point clouds form continuous, elongated curves, which are distinct from the surfaces formed by the background and the discrete points formed by noise. Leveraging this characteristic, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module to extract multi-scale local and long-term contextual features of curve structures. Based on the proposed GDSCA, we construct our network to detect small moving objects. Furthermore, we design a new Spatiotemporal Correlation (STC) loss, which evaluates the local spatiotemporal correlation of events to guide the network in retaining more events with continuous curves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.482, + 0.854 + ], + "angle": 0, + "content": "Finally, we conduct extensive performance evaluations on the proposed dataset, comparing our method with 13 state-of-the-art algorithms, including 4 frame-based methods [16, 30, 37, 51], 6 event-based methods [7, 17, 24-26, 34], and 3 point cloud segmentation methods [2, 9, 36], enabling fair and reproducible comparisons for EVSOD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The main contributions are summarized as follows: (1) We build the first large-scale dataset EV-UAV, with high diversity for EVSOD. It contains 147 event sequences with" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "event-level annotations, covering challenging scenarios like high-brightness and low-light conditions, with targets averaging 1/50 the size in existing datasets. (2) We propose EVSpSegNet, a novel detection framework that distinguishes small targets from background and noise based on the continuous curves formed by targets. Additionally, we introduce STC loss, which retains target events with continuous motion through local spatiotemporal correlation. (3) Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.651, + 0.305 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.76, + 0.33 + ], + "angle": 0, + "content": "2.1. Anti-UAV detection dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.905, + 0.442 + ], + "angle": 0, + "content": "Currently, event-based object detection datasets primarily focus on autonomous driving [5, 26] and general object detection [39, 41, 43]. There is limited attention given to datasets that are exclusively designed for UAV detection. In Tab. 1, we provide a comprehensive summary of existing datasets, highlighting the scarcity of event-based datasets for UAV object detection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.906, + 0.701 + ], + "angle": 0, + "content": "Early datasets often included UAVs only as a minor subcategory within larger datasets [39, 41, 43]. Due to their limited scale and diversity, these datasets are insufficient for training modern deep learning-based methods. The recently proposed F-UAV-D dataset [19] is specifically designed for event-based UAV object detection. However, it only covers indoor environments, which limits its applicability. NeRDD [18], the first dataset dedicated to event-based UAV detection, considers only large objects against clean backgrounds. This fails to capture real-world scenarios where UAVs often appear as small objects in dynamic and cluttered backgrounds. To address these limitations, we introduce EV-UAV, a dataset designed to better represent challenging real-world scenarios. EV-UAV features a large number of detection scenes with small targets and dynamic backgrounds, providing a robust benchmark for developing and evaluating anti-UAV methods." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.816, + 0.728 + ], + "angle": 0, + "content": "2.2. Object detection for event cameras" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Existing object detection methods for event cameras can be broadly divided into two categories. The first category focuses on directly processing sparse event data. For example, graph-based methods model events as dynamic spatiotemporal graphs [15, 21, 32], restricting information propagation to a finite set of sparse connections to reduce computational complexity. Similarly, Spiking Neural Network (SNN)-based methods process asynchronous and sparse event streams directly within the network architecture [4, 23, 33, 48]. By transmitting only discrete spike signals between neurons, these methods significantly re" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7210" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Summary of existing event UAV detection datasets. BL: bright light, NL: normal light, LL: low light, MS: multi-scene, MT: multi-target." + }, + { + "type": "table", + "bbox": [ + 0.139, + 0.129, + 0.86, + 0.25 + ], + "angle": 0, + "content": "
Dataset#AGV.UAV scaleLabel TypeUAV Sequence Ratio(%)UAV centricLighting conditionsObjectYear
BLNLLLMSMT
VisEvent [40]84×66 pixelsBBox15.97××✓✓✓×2023
EventVOT [44]129×100 pixelsBBox8.41××✓✓✓×2024
EvDET200K [42]68×45 pixelsBBox3.57××✓✓✓✓2024
F-UAV-D [19]-BBox100✓×✓×××2024
NeRDD [18]55×31 pixelsBBox100✓×✓××✓2024
EV-UAV6.8×5.4 pixelsSeg100✓✓✓✓✓✓2025
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.274, + 0.482, + 0.334 + ], + "angle": 0, + "content": "duce power consumption during training. However, sparse representation-based approaches are limited by their reliance on specialized hardware or suboptimal performance in complex scenarios." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.336, + 0.485, + 0.683 + ], + "angle": 0, + "content": "The second category transforms sparse event streams into image-like representations, enabling compatibility with dense convolutional operations. Early dense-based methods directly used event representations generated from short time windows for detection [3, 20, 29, 50]. These methods, however, only considered event information within a limited time, resulting in limited performance. Later work addressed this issue by incorporating recurrent neural networks [7, 14, 26], which improved detection performance by capturing temporal dependencies. Despite these advancements, such methods consume significant computational resources for event transformation and processing artificially introduced redundant background data, while struggling to detect small targets due to their lack of distinct texture and contour features. In this work, we observe that small targets form continuous curves in spatiotemporal event point clouds, which are distinct from background and noise. Leveraging this characteristic, we design EVSpSegNet, which directly segments continuous target trajectories from sparse point clouds. This approach avoids the computational overhead caused by converting events into image-like representations. Moreover, the distinct curve features help the network better detect small targets." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.7, + 0.288, + 0.715 + ], + "angle": 0, + "content": "3. EV-UAV benchmark" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.233, + 0.74 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.483, + 0.869 + ], + "angle": 0, + "content": "Event cameras work totally differently from frame-based cameras with fixed frame rates. Specifically, each pixel of an event camera independently responds to changes in light intensity. When the logarithmic brightness change reaches a predefined threshold, an event \\( E = (x,y,t,p) \\) is triggered. In this tuple, \\( x \\) and \\( y \\) represent the pixel coordinates, \\( p \\) indicates the polarity of the brightness change (positive or negative), and \\( t \\) is the timestamp of the event." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Unlike previous annotations for object detection [7, 26] or semantic segmentation [1, 8, 35], we do not convert" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.907, + 0.41 + ], + "angle": 0, + "content": "events into synchronized frames. Instead, we segment events directly in the point cloud space, event by event. The sources of events can be categorized into three types: events generated by the target \\((E_T)\\), events generated by the background \\((E_B)\\), and events caused by noise \\((E_N)\\). This can be represented as: \\( E = E_T + E_B + E_N \\). Our goal is to segment all events generated by target motion from the raw event stream. The final annotation is a set of discrete events \\( E_{label} \\in E_T \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.422, + 0.803, + 0.436 + ], + "angle": 0, + "content": "3.2. Data Collection and Annotations" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.444, + 0.669, + 0.457 + ], + "angle": 0, + "content": "3.2.1. Data Collection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.706 + ], + "angle": 0, + "content": "We collect data using a DAVIS346 camera mounted on a tripod to monitor a fixed range of scenes. This camera has a spatial resolution of \\(346 \\times 260\\) and a temporal resolution of up to \\(10^{6}\\) HZ. Our data collection spans from May 1, 2023, to June 1, 2024, covering various times of the day, seasons, and weather conditions. To accurately reflect the movement patterns of UAVs and the complexity of real-world scenarios, we capture UAVs performing diverse actions in various environments. These actions include forward and backward motion, lateral movements, ascending and descending maneuvers, as well as complex combinations of these behaviors. The dataset encompasses three different lighting conditions (normal, bright and low light) and various backgrounds, such as buildings, mountains, forests, urban areas, clouds, water surfaces, etc. We carefully annotated 147 sequences of data, totaling over 20.3 million target events." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.716, + 0.75, + 0.731 + ], + "angle": 0, + "content": "3.2.2.Event by Event Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Due to the asynchronous and sparse output of event cameras, manually annotating per-event data is highly time-consuming and inefficient. Existing event camera datasets are typically recorded using a combination of event cameras and standard RGB cameras. Labels are first extracted from the frame-based cameras and then transferred to the pixel coordinates of the event cameras through geometric transformations. However, this labeling strategy is limited to producing low frame rate bounding box annotations, which cannot match the high-frequency output characteristics of event cameras or meet the fine-grained, event-" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7211" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.232, + 0.153 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.154, + 0.155, + 0.178, + 0.163 + ], + "angle": 0, + "content": "Train" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.165, + 0.231, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.158, + 0.225, + 0.175, + 0.233 + ], + "angle": 0, + "content": "Val" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.235, + 0.231, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.297, + 0.178, + 0.305 + ], + "angle": 0, + "content": "Test" + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.309, + 0.226, + 0.319 + ], + "angle": 0, + "content": "(a) Position Distribution" + }, + { + "type": "image", + "bbox": [ + 0.236, + 0.092, + 0.455, + 0.263 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.259, + 0.28, + 0.292, + 0.289 + ], + "angle": 0, + "content": "45.1%" + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.296, + 0.308, + 0.307 + ], + "angle": 0, + "content": "Extremely tiny" + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.28, + 0.362, + 0.289 + ], + "angle": 0, + "content": "34.1%" + }, + { + "type": "image_caption", + "bbox": [ + 0.336, + 0.296, + 0.358, + 0.306 + ], + "angle": 0, + "content": "Tiny" + }, + { + "type": "image_caption", + "bbox": [ + 0.399, + 0.28, + 0.432, + 0.289 + ], + "angle": 0, + "content": "20.8%" + }, + { + "type": "image_caption", + "bbox": [ + 0.404, + 0.296, + 0.431, + 0.305 + ], + "angle": 0, + "content": "Small" + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.31, + 0.424, + 0.32 + ], + "angle": 0, + "content": "(b) Target size Distribution" + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.091, + 0.576, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.581, + 0.092, + 0.689, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.699, + 0.092, + 0.809, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.813, + 0.092, + 0.929, + 0.151 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.625, + 0.152, + 0.762, + 0.162 + ], + "angle": 0, + "content": "illumination:highlight \\(\\rightarrow\\) lowlight" + }, + { + "type": "image", + "bbox": [ + 0.461, + 0.164, + 0.576, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.578, + 0.164, + 0.694, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.164, + 0.808, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.164, + 0.929, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.46, + 0.235, + 0.574, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.577, + 0.235, + 0.694, + 0.305 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.696, + 0.235, + 0.808, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.812, + 0.235, + 0.929, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.639, + 0.297, + 0.756, + 0.306 + ], + "angle": 0, + "content": "Target Scale: small→large" + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.309, + 0.75, + 0.32 + ], + "angle": 0, + "content": "(c) Diversity of EV-UAV" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.331, + 0.907, + 0.388 + ], + "angle": 0, + "content": "Figure 2. Illustration of the proposed dataset EV-UAV. (a) Position distribution of EV-UAV. (b) Distribution of noise level and illumination of the proposed dataset. (c) EV-UAV collects diverse event streams across various conditions of illumination and target number and scale. N represents the noise level, reflecting the proportion of events in spatiotemporal space. T denotes the number of targets. S indicates the size of targets in the image plane." + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.411, + 0.49, + 0.512 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.616 + ], + "angle": 0, + "content": "Figure 3. Event-by-event annotation method. Firstly, the events are accumulated into event frames, and targets are annotated with 2D bounding boxes on the event frames. Then, according to the time interval \\(\\Delta t\\) corresponding to each frame, the 2D annotations on the image plane are extended into 3D bounding box annotations in XYT 3D space, and all events within the 3D boxes are filtered as targets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.627, + 0.483, + 0.822 + ], + "angle": 0, + "content": "by-event labeling requirements. Although manual event-by-event labeling could theoretically improve annotation accuracy, the process is highly labor-intensive and time-consuming, making it impractical for real-world applications. To address this issue, we propose a novel method that generates approximate event-by-event annotations using bounding box labels, as illustrated in the Fig. 3. This approach not only achieves fine-grained event-level annotations, enabling microsecond-level object detection, but also allows for degradation into bounding box annotations at any given time, thus supporting the learning requirements of some asynchronous event-based object detection algorithms." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.395, + 0.849 + ], + "angle": 0, + "content": "3.3. Benchmark Features and Statistics" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Dataset Splitting. The EV-UAV dataset is divided into training (99 sequences), validation (24 sequences), and test sets (24 sequences), totaling 147 event sequences." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.516 + ], + "angle": 0, + "content": "Position Distribution. As shown in Fig. 2(a), the targets are mostly located in the central area of the image. The larger horizontal variance suggests that the drone's movement is predominantly horizontal. Additionally, the range of target movements in the validation set is more varied. The fluctuations in the validation set are greater than those in the test set in corresponding directions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.516, + 0.909, + 0.711 + ], + "angle": 0, + "content": "Small-Scale Targets. In real-world scenarios, due to the long imaging distance and the small size of UAVs, drone targets appear at very small scales in the images. The general scale ranks\\(^1\\) of [31] are no longer applicable. Following previous work [47], we further divide small targets into three levels: extremely tiny \\(\\in [1, 8^2)\\), tiny \\(\\in [8^2, 16^2)\\), and small \\(\\in [16^2, \\infty)\\). Fig. 2(b) shows the number of annotations for each target category. It can be observed that the scale of all targets is smaller than \\(32^2\\), with extremely tiny targets accounting for the largest proportion, approximately \\(45\\%\\). Compared to other UAV detection datasets, our EV-UAV dataset has the smallest average UAV scale, as indicated in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.909, + 0.819 + ], + "angle": 0, + "content": "Rich Diversity. To represent real-world anti-UAV target detection scenarios, we endow EV-UAV with diversity. This helps the dataset cover various complex situations in anti-UAV detection both in time and space. As shown in Fig. 2(c), the diversity of EV-UAV is reflected not just in illumination levels but also in background complexity, target quantity, and scale." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.827, + 0.698, + 0.842 + ], + "angle": 0, + "content": "3.4. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.849, + 0.908, + 0.88 + ], + "angle": 0, + "content": "In this work, we evaluate the event-by-event segmentation performance of the algorithm using Intersection over Union" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.886, + 0.855, + 0.901 + ], + "angle": 0, + "content": "\\( {}^{1} \\) small \\( \\in \\left\\lbrack {1,{32}^{2}}\\right) \\) ,medium \\( \\in \\left\\lbrack {{32}^{2},{96}^{2}}\\right) \\) ,and large \\( \\in \\left\\lbrack {{96}^{2},\\infty }\\right) \\) ." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7212" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.104, + 0.089, + 0.898, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.399, + 0.908, + 0.457 + ], + "angle": 0, + "content": "Figure 4. (a) The overview of the proposed Ev-SpSegNet. The network is built on a U-shaped architecture consisting of symmetric encoder and decoder. (b) The GDSCA module. The grouped dilated sparse convolution block is applied to features from different channels to extract local temporal features at multiple scales. The Sp-SE module is used to integrate these features. And the Patch Attention block is used to downsample voxel features and enable global context feature interaction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.482, + 0.487, + 0.543 + ], + "angle": 0, + "content": "(IoU) and accuracy (ACC), and assess the localization performance using probability of detection \\((P_d)\\) and false alarm rate \\((F_a)\\). More details regarding evaluation metrics can be found in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.557, + 0.184, + 0.572 + ], + "angle": 0, + "content": "4. Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.582, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Frame-based detectors require converting events into synchronized image-like representations, which compresses the temporal information of events and makes it even more difficult to detect small targets that already lack distinct appearance cues. In raw event data, events generated by moving targets exhibit strong spatiotemporal correlations, forming elongated continuous curves in event point cloud space. Leveraging this characteristic, we propose EV-SpSegNet, which directly segments events generated by moving targets in sparse event point clouds. Additionally, we introduce a novel spatiotemporal correlation loss that evaluates the local spatiotemporal correlation of events to guide the network in retaining target events with continuous curves." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.788, + 0.233, + 0.804 + ], + "angle": 0, + "content": "4.1. EV-SpSegNet" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.811, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Small moving targets form distinct continuous curves in spatiotemporal event point clouds. However, networks struggle to extract their features because these curves span long temporal ranges, and traditional convolution operations with limited receptive fields cannot effectively capture global features. While attention mechanisms can ex" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.482, + 0.907, + 0.771 + ], + "angle": 0, + "content": "tract global context, their quadratic computational complexity restricts their ability to process large-scale event data. To effectively extract features from continuous curves in spatiotemporal event point clouds, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module, as shown in Fig. 4(b). This module first captures local multiscale features through the Grouped Dilated Sparse Convolution (GDSC) Block and then enables global feature interaction across patches using patch attention. Specifically, the GDSC Block groups input features along the channel dimension, with each group applying convolutions of different dilation rates to extract multi-scale local features. These features are then fused in the Sparse Squeeze and Excitation (Sp-SE) block. Finally, the patch attention block divides the point cloud into larger sub-regions and implements global feature interaction between sub-regions using self-attention. This allows the network to model the differences between targets and background noise in greater detail, enhancing its ability to distinguish moving targets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.782, + 0.804, + 0.799 + ], + "angle": 0, + "content": "4.2. Spatiotemporal Correlation Loss" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.805, + 0.908, + 0.851 + ], + "angle": 0, + "content": "In general image segmentation tasks, the loss function used for training the network is typically the standard pixel-wise binary cross-entropy (BCE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.579, + 0.864, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\nL _ {b c e} (p, y) = \\left\\{ \\begin{array}{c} - \\log (p), y = 1 \\\\ - \\log (1 - p), y = 0 \\end{array} \\right.. \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7213" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.093, + 0.088, + 0.49, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.203, + 0.483, + 0.233 + ], + "angle": 0, + "content": "Figure 5. (a) The principle of spatiotemporal correlation loss. (b) Small object segmentation results of BCE loss and STC loss." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.256, + 0.483, + 0.42 + ], + "angle": 0, + "content": "Although EV-SpSegNet computes structured outputs and focuses on large neighborhood information to better utilize long-term features, losses like BCE loss process each pixel independently. They do not consider the structure of neighborhoods, such as the spatiotemporal continuity of objects, which is crucial for describing the continuous curves formed by objects. As shown in Fig. 5(a), compared to the isolated noise \\( p_2 \\), the event \\( p_1 \\) with more high-confidence supporting events is more likely to be a true target. However, misclassifying both of them only results in a low coat of BCE loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.557 + ], + "angle": 0, + "content": "To address this issue, we introduce a spatiotemporal correlation loss that encourages the network to retain more events with high spatiotemporal correlation while discarding more isolated noise. Specifically, this loss calculates the sum of the confidence of all supporting events within a \\( k \\times k \\times \\tau \\) neighborhood around the events, and defines it as the spatiotemporal correlation weight. We incorporate this weight into the cross-entropy loss and the spatiotemporal correlation loss is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.563, + 0.483, + 0.604 + ], + "angle": 0, + "content": "\\[\nL _ {s t c} (p, y) = \\left\\{ \\begin{array}{c} - w _ {s t c} ^ {\\gamma} \\log (p), y = 1 \\\\ - \\left(1 - w _ {s t c}\\right) ^ {\\gamma} \\log (1 - p), y = 0 \\end{array} , \\right. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.484, + 0.687 + ], + "angle": 0, + "content": "where the \\(y \\in \\{0,1\\}\\) and \\(p \\in (0,1)\\) represent the segmentation label and prediction result of the event, respectively. \\(\\gamma\\) is a hyperparameters to control the influence of spatiotemporal correlations. \\(w_{stc}\\) is the spatiotemporal correlation weight of each event, which is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.694, + 0.483, + 0.727 + ], + "angle": 0, + "content": "\\[\nw _ {s t c} (p) = \\operatorname {s i g m o i d} \\left(\\sum_ {e ^ {j} \\in V ^ {k \\tau}} p ^ {j}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.812 + ], + "angle": 0, + "content": "where \\( V^{k\\tau} \\) represents the neighborhood of events with a size of \\( k \\times k \\times \\tau \\). As illustrated in Fig. 5 (b1-b2), the STC loss preserves the continuous event trajectories generated by moving objects and eliminates false alarms caused by misclassification." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.299, + 0.866 + ], + "angle": 0, + "content": "5.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We used 8-second events as input and set the size of each voxel to \\(1\\mathrm{pixel} \\times 1\\mathrm{pixel} \\times 1\\mathrm{ms}\\). The batch size was set to" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.243 + ], + "angle": 0, + "content": "1, and all layers of the network were randomly initialized. We trained our network using the Adam optimizer [12] for 50 epochs with the learning rate linearly decaying from 1e-2 to 1e-3. The model was trained on the predefined training set of EV-UAV dataset, and we chose the models based on their best performance on the validation set and evaluate them on the test set. The hyperparameters \\( k \\), \\( \\tau \\), and \\( \\gamma \\) of the STC loss are set to 3, 5, and 2, respectively. All models were implemented with PyTorch on a single NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.253, + 0.747, + 0.269 + ], + "angle": 0, + "content": "5.2. Benchmark Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.275, + 0.907, + 0.411 + ], + "angle": 0, + "content": "We compare the proposed method with several state-of-the-art methods, including 4 frame-based generic object detection methods (i.e., SSD [16], Faster RCNN [30], Deformable DETR [51], YOLOV10 [37]), 6 event based methods (i.e., RVT [7], SAST [25], GET [24], RED [26], EMS-YOLO [34], Spike-YOLO [17]), and 3 point cloud segmentation-based methods (i.e., KPCov, RandLA-Net, COSeg). We convert events within a \\(50~ms\\) time window into event frames [20] as input for frame-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.411, + 0.906, + 0.622 + ], + "angle": 0, + "content": "Quantitative Results. As shown in Tab. 2, our EV-SpSegNet achieve the best detection performance with all evaluation metrics. We attribute this to two reasons. On the one hand, the GDSCA module combines local multiscale information with global context, enabling the network to better model the differences between targets and background noise. This enhances the ability of the network to distinguish moving targets. On the other hand, the spatiotemporal correlation loss encourages the network to identify events with high spatiotemporal correlation, improving detection performance. Additionally, by using sparse convolution to build the network, our method can inference 8 seconds of event data in \\(35.9~ms\\), significantly outperforming other methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.82 + ], + "angle": 0, + "content": "Qualitative Results. Qualitative results of different methods are shown in Fig. 6. It can be observed that our method can detect the most moving targets in all scenes, while the compared methods exhibit more or less missed detections and false positives. Particularly in scenes with complex backgrounds, such as the second and fourth rows with strong noise, the improvement of our method is more significant. Much to our surprise, our method even detects some targets that were missed by manual annotation due to strong background clutter, as shown in the third and fourth rows. We attribute this to the spatiotemporal correlation loss, which encourages the network to retain events with high spatiotemporal correlation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.829, + 0.679, + 0.844 + ], + "angle": 0, + "content": "5.3. Ablation Studies" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.762, + 0.867 + ], + "angle": 0, + "content": "5.3.1. Contribution of Components" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "As shown in Tab. 3, adding the GDSC block to the baseline improves the model performance in terms of IoU, ACC," + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7214" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.251, + 0.09, + 0.757, + 0.104 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison of the proposed method to state-of-the-art methods." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.114, + 0.918, + 0.334 + ], + "angle": 0, + "content": "
MethodsPublicationsEvent Rep.TemporalIoU(%)↑ACC(%)↑\\(P_d\\)(%)↑\\(F_a(10^{\\text{~-4~}})\\downarrow\\)#Params.Runtime(ms)
SSD [16]ECCV 2016Event CountNo25.3128.5626.31486.6325.2M2113
Faster RCNN [30]TPAMI 2016Event CountNo26.9329.6827.39689.6841.2M3962
DETR [51]ICLR 2020Event CountNo30.3533.6331.64631.3739.8M3136
YOLOV10-S [37]NIPS 2025Event CountNo32.5533.3932.18589.677.3M1627
EMS-YOLO [34]ICCV 2023SNNYes36.7742.9250.68112.363.3M1229
Spike-YOLO [17]ECCV 2024SNNYes43.9448.2659.6255.3869.0M1883
GET [24]ICCV 2023Groupv TokenYes40.3148.9160.7346.3518.4M2168
RED [26]NeurIPS 2020Voxel GridYes35.9945.5453.76102.2724.1M3427
RVT [7]CVPR 2023Voxel GridYes43.2151.3860.3555.689.9M1737
SAST [25]CVPR 2024Voxel GridYes34.3140.2251.21150.3218.5M3075
KPCov [36]ICCV 2019PointsYes48.1957.2868.5916.3250.1M562
RandLA-Net [9]CVPR 2020PointsYes50.3259.2970.566.951.2M353
COSeg [2]CVPR 2024PointsYes51.8960.9371.329.2123.4M364
Ours-PointsYes55.1865.0277.531.634.0M35.9
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.111, + 0.337, + 0.581, + 0.35 + ], + "angle": 0, + "content": "The bold and the underline represent the best and second-best performance, respectively." + }, + { + "type": "image", + "bbox": [ + 0.123, + 0.361, + 0.877, + 0.668 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.678, + 0.908, + 0.707 + ], + "angle": 0, + "content": "Figure 6. Qualitative results of different methods. Our method accurately detects small moving targets, while other methods produce numerous missed detections and false alarms." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.733, + 0.486, + 0.9 + ], + "angle": 0, + "content": "and \\( P_{d} \\) by 1.37, 2.43 and 1.18, which demonstrates that extracting multi-scale local features using GDSC is beneficial. However, adding only the PA block causes a performance drop. This is because downsampling is applied before computing global attention, and without sufficient local receptive fields, the model cannot effectively capture global features, leading to feature confusion and performance degradation. It is worth noting that when the GDSC block and PA block are used together, the performance improves significantly. This indicates that capturing sufficient local features within patches before enabling global feature interaction is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.733, + 0.907, + 0.793 + ], + "angle": 0, + "content": "essential. Additionally, using the STC loss consistently improves model performance, which shows that guiding the network to retain events with high spatiotemporal correlation through STCLoss is effective." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.806, + 0.848, + 0.82 + ], + "angle": 0, + "content": "5.3.2. The Influence of the Number of Branches" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.907, + 0.9 + ], + "angle": 0, + "content": "As depicted in Tab. 4, we conduct an ablation experiment to study the influence of the number of branches in GDSC block on the final detection performance. It can be observed that the network achieves best performance when the number of branches is set to 4. This is because too" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7215" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.095, + 0.09, + 0.479, + 0.104 + ], + "angle": 0, + "content": "Table 3. Ablation study on framework scheme and loss function." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.114, + 0.481, + 0.222 + ], + "angle": 0, + "content": "
GDSCPALstcIoU↑ACC↑Pd↑Fa↓#Params.
51.3660.2171.944.815.6M
✓52.7362.6473.124.283.8M
✓51.2659.5371.674.315.8M
✓53.0761.0174.133.665.6M
✓✓53.6264.0276.761.934.0M
✓✓✓55.1865.0277.531.634.0M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.236, + 0.482, + 0.249 + ], + "angle": 0, + "content": "Table 4. The influence of the number of branches in GDSC block." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.26, + 0.485, + 0.352 + ], + "angle": 0, + "content": "
#BranchesIoU↑ACC↑Pd↑Fa↓#Params.
150.2162.3673.652.975.6M
251.4164.1675.212.534.5M
353.5264.2776.662.014.3M
455.1865.0277.531.634.0M
553.1264.3176.532.113.8M
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.367, + 0.482, + 0.38 + ], + "angle": 0, + "content": "Table 5. Combination of convolutions with different dilation rates." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.392, + 0.487, + 0.469 + ], + "angle": 0, + "content": "
Dilated ratesIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
(1,2,3,4)55.1865.0277.531.63
(1,2,3,5)53.2164.1176.932.41
(1,3,5,7)52.7564.0276.894.61
(1,3,5,9)51.3563.9275.924.76
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.497, + 0.483, + 0.572 + ], + "angle": 0, + "content": "many branches result in insufficient features assigned to each channel, making it difficult to capture features at the corresponding scale. On the other hand, too few branches lead to insufficient dilation rates, preventing the network from capturing features at multiple scales." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.581, + 0.449, + 0.595 + ], + "angle": 0, + "content": "5.3.3. The Influence of Dilation Rate Combinations" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.6, + 0.483, + 0.722 + ], + "angle": 0, + "content": "We next conduct an ablation study to analyze the relationship between detection performance and different dilation rate combinations. Tab. 5 illustrates that setting the dilation rate to (1,2,3,4) achieves the best performance. This indicates that the dilation rate is not always better when larger. An excessively large dilation rate causes the corresponding events to be too far apart, making it difficult to effectively determine whether these events have temporal correlation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.731, + 0.348, + 0.744 + ], + "angle": 0, + "content": "5.3.4. The Effectiveness of STC Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To further validate the generalizability of our proposed spatiotemporal correlation Loss, we integrate it with several state-of-the-art methods, including KPConv [36], RandLA-Net [9], and COSeg [2]. As shown in Tab. 6, the addition of \\( L_{stc} \\) consistently improves the performance of all methods. For example, KPConv with \\( L_{stc} \\) achieves improvements of \\( 2.13\\% \\), \\( 1.75\\% \\) and \\( 1.4\\% \\) in terms of IoU, ACC and \\( P_d \\), respectively, and the false alarm rate decreases by \\( 5.72\\% \\). These results show that \\( L_{stc} \\) effectively enhances the ability of existing methods to capture spatiotemporal correlations," + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.117 + ], + "angle": 0, + "content": "Table 6. The performance of spatiotemporal correlation loss function on different methods." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.129, + 0.912, + 0.234 + ], + "angle": 0, + "content": "
MethodIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
KPConv48.1957.2868.5916.32
KPConv*50.32(+2.13)59.03(+1.75)69.99(+1.40)10.60(-5.72)
RandLA50.3259.2970.569.21
RandLA*51.15(+0.83)60.63(+1.34)72.36(+2.37)8.91(-1.69)
COSeg51.8960.9371.326.95
COSeg*53.12(+1.23)62.31(0.99)73.96(+2.64)5.86(-1.09)
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.534, + 0.235, + 0.742, + 0.248 + ], + "angle": 0, + "content": "* represent the methods with STC loss." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.276, + 0.907, + 0.368 + ], + "angle": 0, + "content": "leading to better segmentation accuracy and reduced false alarms. The consistent performance gains across diverse methods prove the robustness and general applicability of our proposed loss function. In addition, we explore the selection of optimal hyperparameters for the STC loss. For more details, please refer to the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.766, + 0.397 + ], + "angle": 0, + "content": "6. Discussion and Limitations" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.407, + 0.907, + 0.573 + ], + "angle": 0, + "content": "Although our EV-UAV accurately replicates real-world scenarios for detecting tiny UAVs in the wild, event cameras only capture motion information. When targets are stationary or move slowly, event cameras produce no output, leading to detection failures. In contrast, frame-based cameras can directly provide detailed texture information (i.e., absolute brightness) in static scenes, making them complementary to event cameras. We believe that developing a dual-modal dataset and extending our method to incorporate multi-modal inputs to maximize object detection accuracy represents a promising direction for future work." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.588, + 0.634, + 0.604 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.854 + ], + "angle": 0, + "content": "In this paper, we introduce EV-UAV, the first large-scale benchmark for anti-UAV tasks. EV-UAV is a highly challenging benchmark, featuring diverse scenarios and extremely small targets. It provides event-level fine-grained annotations, supporting algorithms with different event representations. Furthermore, based on the observation that small moving targets form continuous curves in event point cloud space, we propose EV-SpSegNet, a novel baseline for segmenting target motion trajectories from sparse event point clouds, along with a Spatiotemporal Correlation (STC) loss that leverages spatiotemporal local correlation to guide the network in retaining more target events. Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement: This work was supported by the Young Elite Scientist Sponsorship Program by China Association for Science and Technology (NO.YESS20240427)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7216" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.155 + ], + "angle": 0, + "content": "[1] Inigo Alonso and Ana C Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In CVPRW, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.213 + ], + "angle": 0, + "content": "[2] Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, and Serge Belongie. Rethinking few-shot 3d point cloud semantic segmentation. In CVPR, pages 3996-4006, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.216, + 0.483, + 0.27 + ], + "angle": 0, + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Networks and Learning Systems, 25(2):407-417, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.327 + ], + "angle": 0, + "content": "[4] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IEEE International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.482, + 0.384 + ], + "angle": 0, + "content": "[5] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.482, + 0.442 + ], + "angle": 0, + "content": "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[7] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.482, + 0.542 + ], + "angle": 0, + "content": "[8] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.482, + 0.6 + ], + "angle": 0, + "content": "[9] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.482, + 0.656 + ], + "angle": 0, + "content": "[10] Bo Huang, Jianan Li, Junjie Chen, Gang Wang, Jian Zhao, and Tingfa Xu. Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild. 46 (5):2852-2865, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.659, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[11] Nan Jiang, Kuiran Wang, Xiaoke Peng, Xuehui Yu, Qiang Wang, Junliang Xing, Guorong Li, Jian Zhao, Guodong Guo, and Zhenjun Han. Anti-uav: A large multi-modal benchmark for uav tracking. arXiv preprint arXiv:2101.08466, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[12] Diederik P Kingma and Jimmy Lei Ba. Adam: Amethod for stochastic optimization. In ICLR, pages 1-15, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[13] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[14] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE TIP, 31:2975-2987, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[16] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In ECCV, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[18] Gabriele Magrini, Federico Becattini, Pietro Pala, Alberto Del Bimbo, and Antonio Porta. Neuromorphic drone detection: an event-rgb multimodal approach. arXiv preprint arXiv:2409.16099, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[19] Jakub Mandula, Jonas Kühne, Luca Pascarella, and Michele Magno. Towards real-time fast unmanned aerial vehicle detection using dynamic vision sensors. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1-6. IEEE, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.334, + 0.905, + 0.388 + ], + "angle": 0, + "content": "[20] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso Garcia, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In CVPR, pages 5419-5427, 2018. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[21] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In CVPR, pages 14414-14423, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.434, + 0.905, + 0.461 + ], + "angle": 0, + "content": "[22] Francesco Nex and Fabio Remondino. Uav for 3d mapping applications: a review. Applied geomatics, 6:1-15, 2014. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.53 + ], + "angle": 0, + "content": "[23] Federico Paredes-Valles, Kirk YW Scheper, and Guido CHE De Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE TPAMI, 42(8):2051-2064, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.533, + 0.905, + 0.574 + ], + "angle": 0, + "content": "[24] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: group event transformer for event-based vision. In ICCV, pages 6038-6048, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[25] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In CVPR, pages 16794-16804, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[26] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 33:16639-16652, 2020. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[27] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[28] Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, Thomas Lagkas, and Ioannis Moscholios. A compilation of uav applications for precision agriculture. Computer Networks, 172:107148, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.844 + ], + "angle": 0, + "content": "[29] Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In BMVC, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7217" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.191 + ], + "angle": 0, + "content": "[32] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, pages 12371-12381, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[33] Sumit Bam Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. arXiv preprint arXiv.1810.08646, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.289 + ], + "angle": 0, + "content": "[34] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In ICCV, pages 6555-6565, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.347 + ], + "angle": 0, + "content": "[35] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, pages 341-357. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.447 + ], + "angle": 0, + "content": "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. NeurIPS, 37:107984-108011, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[38] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In ICCV, pages 8509-8518, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[39] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.577, + 0.482, + 0.645 + ], + "angle": 0, + "content": "[40] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[41] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[42] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[43] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, pages 19248-19257, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[44] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "and a novel baseline. In CVPR, pages 19248-19257, 2024. 2,3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[45] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[46] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[47] Xinyi Ying, Chao Xiao, Ruojing Li, Xu He, Boyang Li, Xu Cao, Zhaoxu Li, Yingqian Wang, Mingyuan Hu, Qingyu Xu, et al. Visible-thermal tiny object detection: A benchmark dataset and baselines. arXiv preprint arXiv:2406.14482, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.373 + ], + "angle": 0, + "content": "[48] Emre O. Neftci; Hesham Mostafa; Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, pages 51-63, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.417 + ], + "angle": 0, + "content": "[49] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In CVPR, pages 877-886, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.905, + 0.473 + ], + "angle": 0, + "content": "[50] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, pages 989-997, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.475, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7218" + } + ] +] \ No newline at end of file diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_origin.pdf b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..fa1395ad01b2ad7760be474cc7c5dfafd2f46e53 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/4ab2a4db-cd75-44f4-866e-39574b448472_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9f68e7398b73aa97a019f14b254e90a6690378a69e4bc8c46e49641f09eeb5e0 +size 919960 diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/full.md b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/full.md new file mode 100644 index 0000000000000000000000000000000000000000..486c415bbe6c116bf43570d12bf2fcd2a76ca632 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/full.md @@ -0,0 +1,321 @@ +# Event-based Tiny Object Detection: A Benchmark Dataset and Baseline + +Nuo Chen $^{1}$ , Chao Xiao $^{1}$ , Yimian Dai $^{2}$ , Shiman He $^{1}$ , Miao Li $^{1}$ , Wei An $^{1\boxtimes}$ $^{1}$ National University of Defense Technology, $^{2}$ Nankai University + +{chennuo97,anwei}@nudt.edu.cn + +# Abstract + +Small object detection (SOD) in anti-UAV task is a challenging problem due to the small size of UAVs and complex backgrounds. Traditional frame-based cameras struggle to detect small objects in complex environments due to their low frame rates, limited dynamic range, and data redundancy. Event cameras, with microsecond temporal resolution and high dynamic range, provide a more effective solution for SOD. However, existing event-based object detection datasets are limited in scale, feature large targets size, and lack diverse backgrounds, making them unsuitable for SOD benchmarks. In this paper, we introduce a Event-based Small object detection (EVSOD) dataset (namely EVUAV), the first large-scale, highly diverse benchmark for anti-UAV tasks. It includes 147 sequences with over 2.3 million event-level annotations, featuring extremely small targets (averaging $6.8 \times 5.4$ pixels) and diverse scenarios such as urban clutter and extreme lighting conditions. Furthermore, based on the observation that small moving targets form continuous curves in spatiotemporal event point clouds, we propose Event based Sparse Segmentation Network (EV-SpSegNet), a novel baseline for event segmentation in point cloud space, along with a Spatiotemporal Correlation (STC) loss that leverages motion continuity to guide the network in retaining target events. Extensive experiments on the EV-UAV dataset demonstrate the superiority of our method and provide a benchmark for future research in EVSOD. The dataset and code are at https://github.com/ChenYichen9527/Ev-UAV. + +# 1. Introduction + +Unmanned Aerial Vehicles (UAVs) have made significant contributions to fields such as geological exploration [22] and precision agriculture [28] due to their high mobility and low cost. However, the rapid increase in UAV numbers has also introduced potential threats to public safety, making anti-UAV technologies increasingly critical. Small Object Detection (SOD), which aims to accurately and real-time + +![](images/a3b5155344e5a1beedae1b05318592a94a67ff0f6eb755844ede6553091cf76b.jpg) +Figure 1. Comparison between event camera and RGB camera. The RGB camera can only capture the objects under normal light, while the event camera can capture objects under various extreme lighting conditions. And the event camera can capture the continuous motion trajectory of the small object (shown as the red curve). + +locate tiny targets, is fundamental to anti-UAV tasks. However, it remains a challenging problem due to the small size of UAVs and the complexity of their surrounding environments. + +Conventional frame-based cameras suffer from limited frame rates (typically $30 - 60\mathrm{Hz}$ ), narrow dynamic range, and redundant data capture, hindering the development of SOD [10, 11]. Event cameras [6, 27], bio-inspired sensors that asynchronously capture per-pixel intensity changes with microsecond temporal resolution ( $\geq 10^{6}\mathrm{Hz}$ ) and high dynamic range (120dB), offer a promising alternative to SOD under real-world scenarios. Their ability to encode motion as sparse spatio-temporal event streams inherently avoids data redundancy and preserves fine-grained details of high-speed objects, even in challenging illumination. + +Despite this potential, the following challenges hinder their adoption in SOD: First, the small size and limited appearance cues of small objects hinder feature learning, while background clutter degrades detection performance and increases false alarms. Second, current event-based methods [7, 26] often convert events into synchronized image representations for frame-based detection. This disrupts the continuity and sparsity of event data, causing loss of + +temporal information and unnecessary processing of redundant background data. Furthermore, the lack of large-scale, high-quality datasets significantly impedes the development of event-based small moving object detection. + +Therefore, in this paper, we aim to address the aforementioned challenges to advance the development of the event-based small moving object (especially for UAVs) detection (EVSOD) by firstly building a large-scale dataset for EVSOD, then developing a novel detection method tailored for EVSOD based on the spatio-temporal event point cloud representation, and finally extensively evaluating various deep feature learning methods for EVSOD with the developed dataset and the proposed method as baseline. + +Specifically, we construct a new dataset for EVSOD, for the first time, to address the following issues: (1) Existing SOD datasets primarily focus on visible modality [13, 45, 46], infrared modality [38, 49], or the fusion of these two modalities [47], and few research is conducted to explore the EVSOD. (2) Although several event-based object detection datasets [7, 26, 44] have been proposed, they focus on detecting large objects in simple backgrounds, failing to capture real-world scenarios where small objects appear in cluttered environments. (3) Existing event-based object detection datasets [18, 19] lack event-level annotations, restricting the development of algorithms using other event representations (e.g., sparse point clouds and SNNs). The aforementioned issues urge us to introduce the first large-scale dataset with event-level annotations for EVSOD, advancing the development of anti-UAV. + +Then, we propose Event based Sparse Segmentation Network (EV-SpSegNet), which specially designed for efficient and robust EVSOD. We observe that small moving objects in spatiotemporal event point clouds form continuous, elongated curves, which are distinct from the surfaces formed by the background and the discrete points formed by noise. Leveraging this characteristic, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module to extract multi-scale local and long-term contextual features of curve structures. Based on the proposed GDSCA, we construct our network to detect small moving objects. Furthermore, we design a new Spatiotemporal Correlation (STC) loss, which evaluates the local spatiotemporal correlation of events to guide the network in retaining more events with continuous curves. + +Finally, we conduct extensive performance evaluations on the proposed dataset, comparing our method with 13 state-of-the-art algorithms, including 4 frame-based methods [16, 30, 37, 51], 6 event-based methods [7, 17, 24-26, 34], and 3 point cloud segmentation methods [2, 9, 36], enabling fair and reproducible comparisons for EVSOD. + +The main contributions are summarized as follows: (1) We build the first large-scale dataset EV-UAV, with high diversity for EVSOD. It contains 147 event sequences with + +event-level annotations, covering challenging scenarios like high-brightness and low-light conditions, with targets averaging 1/50 the size in existing datasets. (2) We propose EVSpSegNet, a novel detection framework that distinguishes small targets from background and noise based on the continuous curves formed by targets. Additionally, we introduce STC loss, which retains target events with continuous motion through local spatiotemporal correlation. (3) Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD. + +# 2. Related work + +# 2.1. Anti-UAV detection dataset + +Currently, event-based object detection datasets primarily focus on autonomous driving [5, 26] and general object detection [39, 41, 43]. There is limited attention given to datasets that are exclusively designed for UAV detection. In Tab. 1, we provide a comprehensive summary of existing datasets, highlighting the scarcity of event-based datasets for UAV object detection. + +Early datasets often included UAVs only as a minor subcategory within larger datasets [39, 41, 43]. Due to their limited scale and diversity, these datasets are insufficient for training modern deep learning-based methods. The recently proposed F-UAV-D dataset [19] is specifically designed for event-based UAV object detection. However, it only covers indoor environments, which limits its applicability. NeRDD [18], the first dataset dedicated to event-based UAV detection, considers only large objects against clean backgrounds. This fails to capture real-world scenarios where UAVs often appear as small objects in dynamic and cluttered backgrounds. To address these limitations, we introduce EV-UAV, a dataset designed to better represent challenging real-world scenarios. EV-UAV features a large number of detection scenes with small targets and dynamic backgrounds, providing a robust benchmark for developing and evaluating anti-UAV methods. + +# 2.2. Object detection for event cameras + +Existing object detection methods for event cameras can be broadly divided into two categories. The first category focuses on directly processing sparse event data. For example, graph-based methods model events as dynamic spatiotemporal graphs [15, 21, 32], restricting information propagation to a finite set of sparse connections to reduce computational complexity. Similarly, Spiking Neural Network (SNN)-based methods process asynchronous and sparse event streams directly within the network architecture [4, 23, 33, 48]. By transmitting only discrete spike signals between neurons, these methods significantly re + +Table 1. Summary of existing event UAV detection datasets. BL: bright light, NL: normal light, LL: low light, MS: multi-scene, MT: multi-target. + +
Dataset#AGV.UAV scaleLabel TypeUAV Sequence Ratio(%)UAV centricLighting conditionsObjectYear
BLNLLLMSMT
VisEvent [40]84×66 pixelsBBox15.97××✓✓✓×2023
EventVOT [44]129×100 pixelsBBox8.41××✓✓✓×2024
EvDET200K [42]68×45 pixelsBBox3.57××✓✓✓✓2024
F-UAV-D [19]-BBox100✓×✓×××2024
NeRDD [18]55×31 pixelsBBox100✓×✓××✓2024
EV-UAV6.8×5.4 pixelsSeg100✓✓✓✓✓✓2025
+ +duce power consumption during training. However, sparse representation-based approaches are limited by their reliance on specialized hardware or suboptimal performance in complex scenarios. + +The second category transforms sparse event streams into image-like representations, enabling compatibility with dense convolutional operations. Early dense-based methods directly used event representations generated from short time windows for detection [3, 20, 29, 50]. These methods, however, only considered event information within a limited time, resulting in limited performance. Later work addressed this issue by incorporating recurrent neural networks [7, 14, 26], which improved detection performance by capturing temporal dependencies. Despite these advancements, such methods consume significant computational resources for event transformation and processing artificially introduced redundant background data, while struggling to detect small targets due to their lack of distinct texture and contour features. In this work, we observe that small targets form continuous curves in spatiotemporal event point clouds, which are distinct from background and noise. Leveraging this characteristic, we design EVSpSegNet, which directly segments continuous target trajectories from sparse point clouds. This approach avoids the computational overhead caused by converting events into image-like representations. Moreover, the distinct curve features help the network better detect small targets. + +# 3. EV-UAV benchmark + +# 3.1. Preliminaries + +Event cameras work totally differently from frame-based cameras with fixed frame rates. Specifically, each pixel of an event camera independently responds to changes in light intensity. When the logarithmic brightness change reaches a predefined threshold, an event $E = (x,y,t,p)$ is triggered. In this tuple, $x$ and $y$ represent the pixel coordinates, $p$ indicates the polarity of the brightness change (positive or negative), and $t$ is the timestamp of the event. + +Unlike previous annotations for object detection [7, 26] or semantic segmentation [1, 8, 35], we do not convert + +events into synchronized frames. Instead, we segment events directly in the point cloud space, event by event. The sources of events can be categorized into three types: events generated by the target $(E_T)$ , events generated by the background $(E_B)$ , and events caused by noise $(E_N)$ . This can be represented as: $E = E_T + E_B + E_N$ . Our goal is to segment all events generated by target motion from the raw event stream. The final annotation is a set of discrete events $E_{label} \in E_T$ . + +# 3.2. Data Collection and Annotations + +# 3.2.1. Data Collection + +We collect data using a DAVIS346 camera mounted on a tripod to monitor a fixed range of scenes. This camera has a spatial resolution of $346 \times 260$ and a temporal resolution of up to $10^{6}$ HZ. Our data collection spans from May 1, 2023, to June 1, 2024, covering various times of the day, seasons, and weather conditions. To accurately reflect the movement patterns of UAVs and the complexity of real-world scenarios, we capture UAVs performing diverse actions in various environments. These actions include forward and backward motion, lateral movements, ascending and descending maneuvers, as well as complex combinations of these behaviors. The dataset encompasses three different lighting conditions (normal, bright and low light) and various backgrounds, such as buildings, mountains, forests, urban areas, clouds, water surfaces, etc. We carefully annotated 147 sequences of data, totaling over 20.3 million target events. + +# 3.2.2.Event by Event Annotation + +Due to the asynchronous and sparse output of event cameras, manually annotating per-event data is highly time-consuming and inefficient. Existing event camera datasets are typically recorded using a combination of event cameras and standard RGB cameras. Labels are first extracted from the frame-based cameras and then transferred to the pixel coordinates of the event cameras through geometric transformations. However, this labeling strategy is limited to producing low frame rate bounding box annotations, which cannot match the high-frequency output characteristics of event cameras or meet the fine-grained, event- + +![](images/bb4e18c1f386c3ffd75789b83a3b4420cb5fcdd628f839db99905201e67a639c.jpg) +Train + +![](images/a12fba785f58aa730c6b3c85c4952ab984acd82577f79bc30de6ac2291fbf524.jpg) +Val + +![](images/d3a5c188208442400bf692ee4ae2dc69452ad31868294460d1dcd34ee182ddb6.jpg) +Test +(a) Position Distribution +Extremely tiny + +![](images/145dfde06b8a68d89d2e4cca763629db5c5f1567a000fc30c46659e9a4e4c782.jpg) +45.1% +34.1% +Tiny +20.8% + +![](images/fe73cf2aa51bbd4e501e8a8893dc97cb83a15c2e97e4c2a0db94318a72c4fecd.jpg) + +![](images/889655e607a4b646daae5327cfcb31bd76ffd3f9e910d904065a7527383b8891.jpg) +illumination:highlight $\rightarrow$ lowlight + +![](images/49346c06e4c8037fef4e9aa5a23f6d2dc57bcf49b0d6a40dd431d2e0b9985e9e.jpg) + +![](images/870e145cc8dab9f7fb2a70b4018e916a280cd6b824e278ace52e35af75cea424.jpg) + +![](images/34f6d63a72d1a01f7892a1a092dc6e29d0de00ae2dc8a5ba8d98d73494779458.jpg) + +![](images/1f38712daafa71fc4d3801a430ff03b8f9778f21e103f6de445f2f915593bd7e.jpg) + +![](images/852905a2a8ff0cba2ffc1c80a28d79d3dae825bbbe5a8c066a3d9c02c0807e6b.jpg) + +![](images/164bcf02ee65a9a17b2a8865cc3fb66c48bc739fc6cd5793b760a69bfdddb298.jpg) + +![](images/773817104496077a6772820be3efa6df8132a679a2bddf6972a2d00f4b28fdb2.jpg) +Small +(b) Target size Distribution + +![](images/9ab140ff0971fb1ac46dffe06a16cb49f4fa9ed0b03c82b9a002eb84494b7181.jpg) +Target Scale: small→large +(c) Diversity of EV-UAV + +![](images/a5cb2b9a2e6743d4664478abe835cbba71d435eb9d4f8aa284f00db6187b400f.jpg) + +![](images/3875e344e40fc6e401f948ab37afaf21856314b9f8907d62fc36493e1204a0fd.jpg) + +![](images/f89e97b9cd390ae3be73f865acddf7d4de2f875b41acee2ebc9022b8cfe5e14f.jpg) +Figure 2. Illustration of the proposed dataset EV-UAV. (a) Position distribution of EV-UAV. (b) Distribution of noise level and illumination of the proposed dataset. (c) EV-UAV collects diverse event streams across various conditions of illumination and target number and scale. N represents the noise level, reflecting the proportion of events in spatiotemporal space. T denotes the number of targets. S indicates the size of targets in the image plane. +Figure 3. Event-by-event annotation method. Firstly, the events are accumulated into event frames, and targets are annotated with 2D bounding boxes on the event frames. Then, according to the time interval $\Delta t$ corresponding to each frame, the 2D annotations on the image plane are extended into 3D bounding box annotations in XYT 3D space, and all events within the 3D boxes are filtered as targets. + +by-event labeling requirements. Although manual event-by-event labeling could theoretically improve annotation accuracy, the process is highly labor-intensive and time-consuming, making it impractical for real-world applications. To address this issue, we propose a novel method that generates approximate event-by-event annotations using bounding box labels, as illustrated in the Fig. 3. This approach not only achieves fine-grained event-level annotations, enabling microsecond-level object detection, but also allows for degradation into bounding box annotations at any given time, thus supporting the learning requirements of some asynchronous event-based object detection algorithms. + +# 3.3. Benchmark Features and Statistics + +Dataset Splitting. The EV-UAV dataset is divided into training (99 sequences), validation (24 sequences), and test sets (24 sequences), totaling 147 event sequences. + +Position Distribution. As shown in Fig. 2(a), the targets are mostly located in the central area of the image. The larger horizontal variance suggests that the drone's movement is predominantly horizontal. Additionally, the range of target movements in the validation set is more varied. The fluctuations in the validation set are greater than those in the test set in corresponding directions. + +Small-Scale Targets. In real-world scenarios, due to the long imaging distance and the small size of UAVs, drone targets appear at very small scales in the images. The general scale ranks $^1$ of [31] are no longer applicable. Following previous work [47], we further divide small targets into three levels: extremely tiny $\in [1, 8^2)$ , tiny $\in [8^2, 16^2)$ , and small $\in [16^2, \infty)$ . Fig. 2(b) shows the number of annotations for each target category. It can be observed that the scale of all targets is smaller than $32^2$ , with extremely tiny targets accounting for the largest proportion, approximately $45\%$ . Compared to other UAV detection datasets, our EV-UAV dataset has the smallest average UAV scale, as indicated in Tab. 1. + +Rich Diversity. To represent real-world anti-UAV target detection scenarios, we endow EV-UAV with diversity. This helps the dataset cover various complex situations in anti-UAV detection both in time and space. As shown in Fig. 2(c), the diversity of EV-UAV is reflected not just in illumination levels but also in background complexity, target quantity, and scale. + +# 3.4. Evaluation Metrics + +In this work, we evaluate the event-by-event segmentation performance of the algorithm using Intersection over Union + +![](images/31fd6e9fb455be82f46e0d7537741caf16547b592e8df14c069dbb4fe4cccd40.jpg) +Figure 4. (a) The overview of the proposed Ev-SpSegNet. The network is built on a U-shaped architecture consisting of symmetric encoder and decoder. (b) The GDSCA module. The grouped dilated sparse convolution block is applied to features from different channels to extract local temporal features at multiple scales. The Sp-SE module is used to integrate these features. And the Patch Attention block is used to downsample voxel features and enable global context feature interaction. + +(IoU) and accuracy (ACC), and assess the localization performance using probability of detection $(P_d)$ and false alarm rate $(F_a)$ . More details regarding evaluation metrics can be found in the supplementary material. + +# 4. Method + +Frame-based detectors require converting events into synchronized image-like representations, which compresses the temporal information of events and makes it even more difficult to detect small targets that already lack distinct appearance cues. In raw event data, events generated by moving targets exhibit strong spatiotemporal correlations, forming elongated continuous curves in event point cloud space. Leveraging this characteristic, we propose EV-SpSegNet, which directly segments events generated by moving targets in sparse event point clouds. Additionally, we introduce a novel spatiotemporal correlation loss that evaluates the local spatiotemporal correlation of events to guide the network in retaining target events with continuous curves. + +# 4.1. EV-SpSegNet + +Small moving targets form distinct continuous curves in spatiotemporal event point clouds. However, networks struggle to extract their features because these curves span long temporal ranges, and traditional convolution operations with limited receptive fields cannot effectively capture global features. While attention mechanisms can ex + +tract global context, their quadratic computational complexity restricts their ability to process large-scale event data. To effectively extract features from continuous curves in spatiotemporal event point clouds, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module, as shown in Fig. 4(b). This module first captures local multiscale features through the Grouped Dilated Sparse Convolution (GDSC) Block and then enables global feature interaction across patches using patch attention. Specifically, the GDSC Block groups input features along the channel dimension, with each group applying convolutions of different dilation rates to extract multi-scale local features. These features are then fused in the Sparse Squeeze and Excitation (Sp-SE) block. Finally, the patch attention block divides the point cloud into larger sub-regions and implements global feature interaction between sub-regions using self-attention. This allows the network to model the differences between targets and background noise in greater detail, enhancing its ability to distinguish moving targets. + +# 4.2. Spatiotemporal Correlation Loss + +In general image segmentation tasks, the loss function used for training the network is typically the standard pixel-wise binary cross-entropy (BCE) loss: + +$$ +L _ {b c e} (p, y) = \left\{ \begin{array}{c} - \log (p), y = 1 \\ - \log (1 - p), y = 0 \end{array} \right.. \tag {1} +$$ + +![](images/37d6bcf7677342d7d683588c5fd4ea7dc5e26760ed8aa3de541b0c357eaaaf2a.jpg) +Figure 5. (a) The principle of spatiotemporal correlation loss. (b) Small object segmentation results of BCE loss and STC loss. + +Although EV-SpSegNet computes structured outputs and focuses on large neighborhood information to better utilize long-term features, losses like BCE loss process each pixel independently. They do not consider the structure of neighborhoods, such as the spatiotemporal continuity of objects, which is crucial for describing the continuous curves formed by objects. As shown in Fig. 5(a), compared to the isolated noise $p_2$ , the event $p_1$ with more high-confidence supporting events is more likely to be a true target. However, misclassifying both of them only results in a low coat of BCE loss. + +To address this issue, we introduce a spatiotemporal correlation loss that encourages the network to retain more events with high spatiotemporal correlation while discarding more isolated noise. Specifically, this loss calculates the sum of the confidence of all supporting events within a $k \times k \times \tau$ neighborhood around the events, and defines it as the spatiotemporal correlation weight. We incorporate this weight into the cross-entropy loss and the spatiotemporal correlation loss is defined as follows: + +$$ +L _ {s t c} (p, y) = \left\{ \begin{array}{c} - w _ {s t c} ^ {\gamma} \log (p), y = 1 \\ - \left(1 - w _ {s t c}\right) ^ {\gamma} \log (1 - p), y = 0 \end{array} , \right. \tag {2} +$$ + +where the $y \in \{0,1\}$ and $p \in (0,1)$ represent the segmentation label and prediction result of the event, respectively. $\gamma$ is a hyperparameters to control the influence of spatiotemporal correlations. $w_{stc}$ is the spatiotemporal correlation weight of each event, which is defined as: + +$$ +w _ {s t c} (p) = \operatorname {s i g m o i d} \left(\sum_ {e ^ {j} \in V ^ {k \tau}} p ^ {j}\right), \tag {3} +$$ + +where $V^{k\tau}$ represents the neighborhood of events with a size of $k \times k \times \tau$ . As illustrated in Fig. 5 (b1-b2), the STC loss preserves the continuous event trajectories generated by moving objects and eliminates false alarms caused by misclassification. + +# 5. Experiments + +# 5.1. Experimental Settings + +We used 8-second events as input and set the size of each voxel to $1\mathrm{pixel} \times 1\mathrm{pixel} \times 1\mathrm{ms}$ . The batch size was set to + +1, and all layers of the network were randomly initialized. We trained our network using the Adam optimizer [12] for 50 epochs with the learning rate linearly decaying from 1e-2 to 1e-3. The model was trained on the predefined training set of EV-UAV dataset, and we chose the models based on their best performance on the validation set and evaluate them on the test set. The hyperparameters $k$ , $\tau$ , and $\gamma$ of the STC loss are set to 3, 5, and 2, respectively. All models were implemented with PyTorch on a single NVIDIA RTX 3090 GPU. + +# 5.2. Benchmark Comparisons + +We compare the proposed method with several state-of-the-art methods, including 4 frame-based generic object detection methods (i.e., SSD [16], Faster RCNN [30], Deformable DETR [51], YOLOV10 [37]), 6 event based methods (i.e., RVT [7], SAST [25], GET [24], RED [26], EMS-YOLO [34], Spike-YOLO [17]), and 3 point cloud segmentation-based methods (i.e., KPCov, RandLA-Net, COSeg). We convert events within a $50~ms$ time window into event frames [20] as input for frame-based methods. + +Quantitative Results. As shown in Tab. 2, our EV-SpSegNet achieve the best detection performance with all evaluation metrics. We attribute this to two reasons. On the one hand, the GDSCA module combines local multiscale information with global context, enabling the network to better model the differences between targets and background noise. This enhances the ability of the network to distinguish moving targets. On the other hand, the spatiotemporal correlation loss encourages the network to identify events with high spatiotemporal correlation, improving detection performance. Additionally, by using sparse convolution to build the network, our method can inference 8 seconds of event data in $35.9~ms$ , significantly outperforming other methods. + +Qualitative Results. Qualitative results of different methods are shown in Fig. 6. It can be observed that our method can detect the most moving targets in all scenes, while the compared methods exhibit more or less missed detections and false positives. Particularly in scenes with complex backgrounds, such as the second and fourth rows with strong noise, the improvement of our method is more significant. Much to our surprise, our method even detects some targets that were missed by manual annotation due to strong background clutter, as shown in the third and fourth rows. We attribute this to the spatiotemporal correlation loss, which encourages the network to retain events with high spatiotemporal correlation. + +# 5.3. Ablation Studies + +# 5.3.1. Contribution of Components + +As shown in Tab. 3, adding the GDSC block to the baseline improves the model performance in terms of IoU, ACC, + +Table 2. Quantitative comparison of the proposed method to state-of-the-art methods. + +
MethodsPublicationsEvent Rep.TemporalIoU(%)↑ACC(%)↑\(P_d\)(%)↑\(F_a(10^{\text{~-4~}})\downarrow\)#Params.Runtime(ms)
SSD [16]ECCV 2016Event CountNo25.3128.5626.31486.6325.2M2113
Faster RCNN [30]TPAMI 2016Event CountNo26.9329.6827.39689.6841.2M3962
DETR [51]ICLR 2020Event CountNo30.3533.6331.64631.3739.8M3136
YOLOV10-S [37]NIPS 2025Event CountNo32.5533.3932.18589.677.3M1627
EMS-YOLO [34]ICCV 2023SNNYes36.7742.9250.68112.363.3M1229
Spike-YOLO [17]ECCV 2024SNNYes43.9448.2659.6255.3869.0M1883
GET [24]ICCV 2023Groupv TokenYes40.3148.9160.7346.3518.4M2168
RED [26]NeurIPS 2020Voxel GridYes35.9945.5453.76102.2724.1M3427
RVT [7]CVPR 2023Voxel GridYes43.2151.3860.3555.689.9M1737
SAST [25]CVPR 2024Voxel GridYes34.3140.2251.21150.3218.5M3075
KPCov [36]ICCV 2019PointsYes48.1957.2868.5916.3250.1M562
RandLA-Net [9]CVPR 2020PointsYes50.3259.2970.566.951.2M353
COSeg [2]CVPR 2024PointsYes51.8960.9371.329.2123.4M364
Ours-PointsYes55.1865.0277.531.634.0M35.9
+ +The bold and the underline represent the best and second-best performance, respectively. + +![](images/ed49ab74805cc3b98d74b4d3788f92edb573fcbc12a5600795508baf07ae0b4e.jpg) +Figure 6. Qualitative results of different methods. Our method accurately detects small moving targets, while other methods produce numerous missed detections and false alarms. + +and $P_{d}$ by 1.37, 2.43 and 1.18, which demonstrates that extracting multi-scale local features using GDSC is beneficial. However, adding only the PA block causes a performance drop. This is because downsampling is applied before computing global attention, and without sufficient local receptive fields, the model cannot effectively capture global features, leading to feature confusion and performance degradation. It is worth noting that when the GDSC block and PA block are used together, the performance improves significantly. This indicates that capturing sufficient local features within patches before enabling global feature interaction is + +essential. Additionally, using the STC loss consistently improves model performance, which shows that guiding the network to retain events with high spatiotemporal correlation through STCLoss is effective. + +# 5.3.2. The Influence of the Number of Branches + +As depicted in Tab. 4, we conduct an ablation experiment to study the influence of the number of branches in GDSC block on the final detection performance. It can be observed that the network achieves best performance when the number of branches is set to 4. This is because too + +Table 3. Ablation study on framework scheme and loss function. + +
GDSCPALstcIoU↑ACC↑Pd↑Fa↓#Params.
51.3660.2171.944.815.6M
✓52.7362.6473.124.283.8M
✓51.2659.5371.674.315.8M
✓53.0761.0174.133.665.6M
✓✓53.6264.0276.761.934.0M
✓✓✓55.1865.0277.531.634.0M
+ +Table 4. The influence of the number of branches in GDSC block. + +
#BranchesIoU↑ACC↑Pd↑Fa↓#Params.
150.2162.3673.652.975.6M
251.4164.1675.212.534.5M
353.5264.2776.662.014.3M
455.1865.0277.531.634.0M
553.1264.3176.532.113.8M
+ +Table 5. Combination of convolutions with different dilation rates. + +
Dilated ratesIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
(1,2,3,4)55.1865.0277.531.63
(1,2,3,5)53.2164.1176.932.41
(1,3,5,7)52.7564.0276.894.61
(1,3,5,9)51.3563.9275.924.76
+ +many branches result in insufficient features assigned to each channel, making it difficult to capture features at the corresponding scale. On the other hand, too few branches lead to insufficient dilation rates, preventing the network from capturing features at multiple scales. + +# 5.3.3. The Influence of Dilation Rate Combinations + +We next conduct an ablation study to analyze the relationship between detection performance and different dilation rate combinations. Tab. 5 illustrates that setting the dilation rate to (1,2,3,4) achieves the best performance. This indicates that the dilation rate is not always better when larger. An excessively large dilation rate causes the corresponding events to be too far apart, making it difficult to effectively determine whether these events have temporal correlation. + +# 5.3.4. The Effectiveness of STC Loss + +To further validate the generalizability of our proposed spatiotemporal correlation Loss, we integrate it with several state-of-the-art methods, including KPConv [36], RandLA-Net [9], and COSeg [2]. As shown in Tab. 6, the addition of $L_{stc}$ consistently improves the performance of all methods. For example, KPConv with $L_{stc}$ achieves improvements of $2.13\%$ , $1.75\%$ and $1.4\%$ in terms of IoU, ACC and $P_d$ , respectively, and the false alarm rate decreases by $5.72\%$ . These results show that $L_{stc}$ effectively enhances the ability of existing methods to capture spatiotemporal correlations, + +Table 6. The performance of spatiotemporal correlation loss function on different methods. + +
MethodIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
KPConv48.1957.2868.5916.32
KPConv*50.32(+2.13)59.03(+1.75)69.99(+1.40)10.60(-5.72)
RandLA50.3259.2970.569.21
RandLA*51.15(+0.83)60.63(+1.34)72.36(+2.37)8.91(-1.69)
COSeg51.8960.9371.326.95
COSeg*53.12(+1.23)62.31(0.99)73.96(+2.64)5.86(-1.09)
+ +* represent the methods with STC loss. + +leading to better segmentation accuracy and reduced false alarms. The consistent performance gains across diverse methods prove the robustness and general applicability of our proposed loss function. In addition, we explore the selection of optimal hyperparameters for the STC loss. For more details, please refer to the supplementary material. + +# 6. Discussion and Limitations + +Although our EV-UAV accurately replicates real-world scenarios for detecting tiny UAVs in the wild, event cameras only capture motion information. When targets are stationary or move slowly, event cameras produce no output, leading to detection failures. In contrast, frame-based cameras can directly provide detailed texture information (i.e., absolute brightness) in static scenes, making them complementary to event cameras. We believe that developing a dual-modal dataset and extending our method to incorporate multi-modal inputs to maximize object detection accuracy represents a promising direction for future work. + +# 7. Conclusion + +In this paper, we introduce EV-UAV, the first large-scale benchmark for anti-UAV tasks. EV-UAV is a highly challenging benchmark, featuring diverse scenarios and extremely small targets. It provides event-level fine-grained annotations, supporting algorithms with different event representations. Furthermore, based on the observation that small moving targets form continuous curves in event point cloud space, we propose EV-SpSegNet, a novel baseline for segmenting target motion trajectories from sparse event point clouds, along with a Spatiotemporal Correlation (STC) loss that leverages spatiotemporal local correlation to guide the network in retaining more target events. Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD. + +Acknowledgement: This work was supported by the Young Elite Scientist Sponsorship Program by China Association for Science and Technology (NO.YESS20240427). + +# References + +[1] Inigo Alonso and Ana C Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In CVPRW, pages 0–0, 2019. 3 +[2] Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, and Serge Belongie. Rethinking few-shot 3d point cloud semantic segmentation. In CVPR, pages 3996-4006, 2024. 2, 7, 8 +[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Networks and Learning Systems, 25(2):407-417, 2013. 3 +[4] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IEEE International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2022. 2 +[5] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 2 +[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. 44(1):154-180, 2020. 1 +[7] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1, 2, 3, 6, 7 +[8] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 3 +[9] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 2, 7, 8 +[10] Bo Huang, Jianan Li, Junjie Chen, Gang Wang, Jian Zhao, and Tingfa Xu. Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild. 46 (5):2852-2865, 2023. 1 +[11] Nan Jiang, Kuiran Wang, Xiaoke Peng, Xuehui Yu, Qiang Wang, Junliang Xing, Guorong Li, Jian Zhao, Guodong Guo, and Zhenjun Han. Anti-uav: A large multi-modal benchmark for uav tracking. arXiv preprint arXiv:2101.08466, 2021. 1 +[12] Diederik P Kingma and Jimmy Lei Ba. Adam: Amethod for stochastic optimization. In ICLR, pages 1-15, 2014. 6 +[13] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 2 +[14] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE TIP, 31:2975-2987, 2022. 3 +[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2 + +[16] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 2, 6, 7 +[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In ECCV, 2024. 2, 6, 7 +[18] Gabriele Magrini, Federico Becattini, Pietro Pala, Alberto Del Bimbo, and Antonio Porta. Neuromorphic drone detection: an event-rgb multimodal approach. arXiv preprint arXiv:2409.16099, 2024. 2, 3 +[19] Jakub Mandula, Jonas Kühne, Luca Pascarella, and Michele Magno. Towards real-time fast unmanned aerial vehicle detection using dynamic vision sensors. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1-6. IEEE, 2024. 2, 3 +[20] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso Garcia, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In CVPR, pages 5419-5427, 2018. 3, 6 +[21] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In CVPR, pages 14414-14423, 2020. 2 +[22] Francesco Nex and Fabio Remondino. Uav for 3d mapping applications: a review. Applied geomatics, 6:1-15, 2014. 1 +[23] Federico Paredes-Valles, Kirk YW Scheper, and Guido CHE De Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE TPAMI, 42(8):2051-2064, 2019. 2 +[24] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: group event transformer for event-based vision. In ICCV, pages 6038-6048, 2023. 2, 6, 7 +[25] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In CVPR, pages 16794-16804, 2024. 6, 7 +[26] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 33:16639-16652, 2020. 1, 2, 3, 6, 7 +[27] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1 +[28] Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, Thomas Lagkas, and Ioannis Moscholios. A compilation of uav applications for precision agriculture. Computer Networks, 172:107148, 2020. 1 +[29] Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In BMVC, 2017. 3 +[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2, 6, 7 + +[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 4 +[32] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, pages 12371-12381, 2022. 2 +[33] Sumit Bam Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. arXiv preprint arXiv.1810.08646, 2018. 2 +[34] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In ICCV, pages 6555-6565, 2023. 2, 6, 7 +[35] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, pages 341-357. Springer, 2022. 3 +[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 2, 7, 8 +[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. NeurIPS, 37:107984-108011, 2025. 2, 6, 7 +[38] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In ICCV, pages 8509-8518, 2019. 2 +[39] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 2 +[40] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 3 +[41] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 2 +[42] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 3 +[43] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, pages 19248-19257, 2024. 2 +[44] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset + +and a novel baseline. In CVPR, pages 19248-19257, 2024. 2,3 +[45] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543. Springer, 2022. 2 +[46] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 2 +[47] Xinyi Ying, Chao Xiao, Ruojing Li, Xu He, Boyang Li, Xu Cao, Zhaoxu Li, Yingqian Wang, Mingyuan Hu, Qingyu Xu, et al. Visible-thermal tiny object detection: A benchmark dataset and baselines. arXiv preprint arXiv:2406.14482, 2024. 2, 4 +[48] Emre O. Neftci; Hesham Mostafa; Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, pages 51-63, 2019. 2 +[49] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In CVPR, pages 877-886, 2022. 2 +[50] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, pages 989-997, 2019. 3 +[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 6, 7 \ No newline at end of file diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/images.zip b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7520c7e97bd6d2ee3eafc1f8f55c12b2392368a0 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:5d391a5d9b04b048c1ff501017541cd17ebe5d90862fe61a7a3f4366af801aab +size 645385 diff --git a/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/layout.json b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..d4d736d4756d6b0bce3675fa134934b5cf032d11 --- /dev/null +++ b/2025/Event-based Tiny Object Detection_ A Benchmark Dataset and Baseline/layout.json @@ -0,0 +1,8470 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 83, + 103, + 527, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 103, + 527, + 120 + ], + "spans": [ + { + "bbox": [ + 83, + 103, + 527, + 120 + ], + "type": "text", + "content": "Event-based Tiny Object Detection: A Benchmark Dataset and Baseline" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "spans": [ + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": "Nuo Chen" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": ", Chao Xiao" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": ", Yimian Dai" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": ", Shiman He" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": ", Miao Li" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": ", Wei An" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1\\boxtimes}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": "National University of Defense Technology, " + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 128, + 148, + 482, + 182 + ], + "type": "text", + "content": "Nankai University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 226, + 186, + 384, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 186, + 384, + 196 + ], + "spans": [ + { + "bbox": [ + 226, + 186, + 384, + 196 + ], + "type": "text", + "content": "{chennuo97,anwei}@nudt.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 226, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 251, + 297, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 251, + 297, + 586 + ], + "spans": [ + { + "bbox": [ + 54, + 251, + 297, + 586 + ], + "type": "text", + "content": "Small object detection (SOD) in anti-UAV task is a challenging problem due to the small size of UAVs and complex backgrounds. Traditional frame-based cameras struggle to detect small objects in complex environments due to their low frame rates, limited dynamic range, and data redundancy. Event cameras, with microsecond temporal resolution and high dynamic range, provide a more effective solution for SOD. However, existing event-based object detection datasets are limited in scale, feature large targets size, and lack diverse backgrounds, making them unsuitable for SOD benchmarks. In this paper, we introduce a Event-based Small object detection (EVSOD) dataset (namely EVUAV), the first large-scale, highly diverse benchmark for anti-UAV tasks. It includes 147 sequences with over 2.3 million event-level annotations, featuring extremely small targets (averaging " + }, + { + "bbox": [ + 54, + 251, + 297, + 586 + ], + "type": "inline_equation", + "content": "6.8 \\times 5.4" + }, + { + "bbox": [ + 54, + 251, + 297, + 586 + ], + "type": "text", + "content": " pixels) and diverse scenarios such as urban clutter and extreme lighting conditions. Furthermore, based on the observation that small moving targets form continuous curves in spatiotemporal event point clouds, we propose Event based Sparse Segmentation Network (EV-SpSegNet), a novel baseline for event segmentation in point cloud space, along with a Spatiotemporal Correlation (STC) loss that leverages motion continuity to guide the network in retaining target events. Extensive experiments on the EV-UAV dataset demonstrate the superiority of our method and provide a benchmark for future research in EVSOD. The dataset and code are at https://github.com/ChenYichen9527/Ev-UAV." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "spans": [ + { + "bbox": [ + 56, + 609, + 135, + 621 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "Unmanned Aerial Vehicles (UAVs) have made significant contributions to fields such as geological exploration [22] and precision agriculture [28] due to their high mobility and low cost. However, the rapid increase in UAV numbers has also introduced potential threats to public safety, making anti-UAV technologies increasingly critical. Small Object Detection (SOD), which aims to accurately and real-time" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 320, + 227, + 556, + 354 + ], + "blocks": [ + { + "bbox": [ + 320, + 227, + 556, + 354 + ], + "lines": [ + { + "bbox": [ + 320, + 227, + 556, + 354 + ], + "spans": [ + { + "bbox": [ + 320, + 227, + 556, + 354 + ], + "type": "image", + "image_path": "a3b5155344e5a1beedae1b05318592a94a67ff0f6eb755844ede6553091cf76b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 363, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. Comparison between event camera and RGB camera. The RGB camera can only capture the objects under normal light, while the event camera can capture objects under various extreme lighting conditions. And the event camera can capture the continuous motion trajectory of the small object (shown as the red curve)." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 555, + 481 + ], + "type": "text", + "content": "locate tiny targets, is fundamental to anti-UAV tasks. However, it remains a challenging problem due to the small size of UAVs and the complexity of their surrounding environments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "text", + "content": "Conventional frame-based cameras suffer from limited frame rates (typically " + }, + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "inline_equation", + "content": "30 - 60\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "text", + "content": "), narrow dynamic range, and redundant data capture, hindering the development of SOD [10, 11]. Event cameras [6, 27], bio-inspired sensors that asynchronously capture per-pixel intensity changes with microsecond temporal resolution (" + }, + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "inline_equation", + "content": "\\geq 10^{6}\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 484, + 556, + 616 + ], + "type": "text", + "content": ") and high dynamic range (120dB), offer a promising alternative to SOD under real-world scenarios. Their ability to encode motion as sparse spatio-temporal event streams inherently avoids data redundancy and preserves fine-grained details of high-speed objects, even in challenging illumination." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 557, + 714 + ], + "type": "text", + "content": "Despite this potential, the following challenges hinder their adoption in SOD: First, the small size and limited appearance cues of small objects hinder feature learning, while background clutter degrades detection performance and increases false alarms. Second, current event-based methods [7, 26] often convert events into synchronized image representations for frame-based detection. This disrupts the continuity and sparsity of event data, causing loss of" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7209" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 120 + ], + "type": "text", + "content": "temporal information and unnecessary processing of redundant background data. Furthermore, the lack of large-scale, high-quality datasets significantly impedes the development of event-based small moving object detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 121, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 121, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 54, + 121, + 294, + 228 + ], + "type": "text", + "content": "Therefore, in this paper, we aim to address the aforementioned challenges to advance the development of the event-based small moving object (especially for UAVs) detection (EVSOD) by firstly building a large-scale dataset for EVSOD, then developing a novel detection method tailored for EVSOD based on the spatio-temporal event point cloud representation, and finally extensively evaluating various deep feature learning methods for EVSOD with the developed dataset and the proposed method as baseline." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 231, + 295, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 422 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 422 + ], + "type": "text", + "content": "Specifically, we construct a new dataset for EVSOD, for the first time, to address the following issues: (1) Existing SOD datasets primarily focus on visible modality [13, 45, 46], infrared modality [38, 49], or the fusion of these two modalities [47], and few research is conducted to explore the EVSOD. (2) Although several event-based object detection datasets [7, 26, 44] have been proposed, they focus on detecting large objects in simple backgrounds, failing to capture real-world scenarios where small objects appear in cluttered environments. (3) Existing event-based object detection datasets [18, 19] lack event-level annotations, restricting the development of algorithms using other event representations (e.g., sparse point clouds and SNNs). The aforementioned issues urge us to introduce the first large-scale dataset with event-level annotations for EVSOD, advancing the development of anti-UAV." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 423, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 602 + ], + "type": "text", + "content": "Then, we propose Event based Sparse Segmentation Network (EV-SpSegNet), which specially designed for efficient and robust EVSOD. We observe that small moving objects in spatiotemporal event point clouds form continuous, elongated curves, which are distinct from the surfaces formed by the background and the discrete points formed by noise. Leveraging this characteristic, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module to extract multi-scale local and long-term contextual features of curve structures. Based on the proposed GDSCA, we construct our network to detect small moving objects. Furthermore, we design a new Spatiotemporal Correlation (STC) loss, which evaluates the local spatiotemporal correlation of events to guide the network in retaining more events with continuous curves." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 604, + 294, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 294, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 294, + 676 + ], + "type": "text", + "content": "Finally, we conduct extensive performance evaluations on the proposed dataset, comparing our method with 13 state-of-the-art algorithms, including 4 frame-based methods [16, 30, 37, 51], 6 event-based methods [7, 17, 24-26, 34], and 3 point cloud segmentation methods [2, 9, 36], enabling fair and reproducible comparisons for EVSOD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "The main contributions are summarized as follows: (1) We build the first large-scale dataset EV-UAV, with high diversity for EVSOD. It contains 147 event sequences with" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "event-level annotations, covering challenging scenarios like high-brightness and low-light conditions, with targets averaging 1/50 the size in existing datasets. (2) We propose EVSpSegNet, a novel detection framework that distinguishes small targets from background and noise based on the continuous curves formed by targets. Additionally, we introduce STC loss, which retains target events with continuous motion through local spatiotemporal correlation. (3) Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 228, + 398, + 241 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 398, + 241 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 398, + 241 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 248, + 465, + 261 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 465, + 261 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 465, + 261 + ], + "type": "text", + "content": "2.1. Anti-UAV detection dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 266, + 553, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 553, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 553, + 350 + ], + "type": "text", + "content": "Currently, event-based object detection datasets primarily focus on autonomous driving [5, 26] and general object detection [39, 41, 43]. There is limited attention given to datasets that are exclusively designed for UAV detection. In Tab. 1, we provide a comprehensive summary of existing datasets, highlighting the scarcity of event-based datasets for UAV object detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 554, + 555 + ], + "type": "text", + "content": "Early datasets often included UAVs only as a minor subcategory within larger datasets [39, 41, 43]. Due to their limited scale and diversity, these datasets are insufficient for training modern deep learning-based methods. The recently proposed F-UAV-D dataset [19] is specifically designed for event-based UAV object detection. However, it only covers indoor environments, which limits its applicability. NeRDD [18], the first dataset dedicated to event-based UAV detection, considers only large objects against clean backgrounds. This fails to capture real-world scenarios where UAVs often appear as small objects in dynamic and cluttered backgrounds. To address these limitations, we introduce EV-UAV, a dataset designed to better represent challenging real-world scenarios. EV-UAV features a large number of detection scenes with small targets and dynamic backgrounds, providing a robust benchmark for developing and evaluating anti-UAV methods." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 563, + 499, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 499, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 499, + 576 + ], + "type": "text", + "content": "2.2. Object detection for event cameras" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 713 + ], + "type": "text", + "content": "Existing object detection methods for event cameras can be broadly divided into two categories. The first category focuses on directly processing sparse event data. For example, graph-based methods model events as dynamic spatiotemporal graphs [15, 21, 32], restricting information propagation to a finite set of sparse connections to reduce computational complexity. Similarly, Spiking Neural Network (SNN)-based methods process asynchronous and sparse event streams directly within the network architecture [4, 23, 33, 48]. By transmitting only discrete spike signals between neurons, these methods significantly re" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7210" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 85, + 102, + 526, + 198 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Summary of existing event UAV detection datasets. BL: bright light, NL: normal light, LL: low light, MS: multi-scene, MT: multi-target." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 85, + 102, + 526, + 198 + ], + "lines": [ + { + "bbox": [ + 85, + 102, + 526, + 198 + ], + "spans": [ + { + "bbox": [ + 85, + 102, + 526, + 198 + ], + "type": "table", + "html": "
Dataset#AGV.UAV scaleLabel TypeUAV Sequence Ratio(%)UAV centricLighting conditionsObjectYear
BLNLLLMSMT
VisEvent [40]84×66 pixelsBBox15.97××✓✓✓×2023
EventVOT [44]129×100 pixelsBBox8.41××✓✓✓×2024
EvDET200K [42]68×45 pixelsBBox3.57××✓✓✓✓2024
F-UAV-D [19]-BBox100✓×✓×××2024
NeRDD [18]55×31 pixelsBBox100✓×✓××✓2024
EV-UAV6.8×5.4 pixelsSeg100✓✓✓✓✓✓2025
", + "image_path": "809bbc2909354d3512eba1e3b743fef7ea2ad73d3de5380bca82193f1d894934.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 217, + 294, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 217, + 294, + 264 + ], + "spans": [ + { + "bbox": [ + 54, + 217, + 294, + 264 + ], + "type": "text", + "content": "duce power consumption during training. However, sparse representation-based approaches are limited by their reliance on specialized hardware or suboptimal performance in complex scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 266, + 296, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 266, + 296, + 540 + ], + "spans": [ + { + "bbox": [ + 56, + 266, + 296, + 540 + ], + "type": "text", + "content": "The second category transforms sparse event streams into image-like representations, enabling compatibility with dense convolutional operations. Early dense-based methods directly used event representations generated from short time windows for detection [3, 20, 29, 50]. These methods, however, only considered event information within a limited time, resulting in limited performance. Later work addressed this issue by incorporating recurrent neural networks [7, 14, 26], which improved detection performance by capturing temporal dependencies. Despite these advancements, such methods consume significant computational resources for event transformation and processing artificially introduced redundant background data, while struggling to detect small targets due to their lack of distinct texture and contour features. In this work, we observe that small targets form continuous curves in spatiotemporal event point clouds, which are distinct from background and noise. Leveraging this characteristic, we design EVSpSegNet, which directly segments continuous target trajectories from sparse point clouds. This approach avoids the computational overhead caused by converting events into image-like representations. Moreover, the distinct curve features help the network better detect small targets." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 554, + 176, + 566 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 176, + 566 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 176, + 566 + ], + "type": "text", + "content": "3. EV-UAV benchmark" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 142, + 586 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": "Event cameras work totally differently from frame-based cameras with fixed frame rates. Specifically, each pixel of an event camera independently responds to changes in light intensity. When the logarithmic brightness change reaches a predefined threshold, an event " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "E = (x,y,t,p)" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " is triggered. In this tuple, " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " represent the pixel coordinates, " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " indicates the polarity of the brightness change (positive or negative), and " + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 593, + 295, + 688 + ], + "type": "text", + "content": " is the timestamp of the event." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Unlike previous annotations for object detection [7, 26] or semantic segmentation [1, 8, 35], we do not convert" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": "events into synchronized frames. Instead, we segment events directly in the point cloud space, event by event. The sources of events can be categorized into three types: events generated by the target " + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "inline_equation", + "content": "(E_T)" + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": ", events generated by the background " + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "inline_equation", + "content": "(E_B)" + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": ", and events caused by noise " + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "inline_equation", + "content": "(E_N)" + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": ". This can be represented as: " + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "inline_equation", + "content": "E = E_T + E_B + E_N" + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": ". Our goal is to segment all events generated by target motion from the raw event stream. The final annotation is a set of discrete events " + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "inline_equation", + "content": "E_{label} \\in E_T" + }, + { + "bbox": [ + 313, + 217, + 555, + 324 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 334, + 491, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 491, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 491, + 345 + ], + "type": "text", + "content": "3.2. Data Collection and Annotations" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 351, + 409, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 409, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 409, + 361 + ], + "type": "text", + "content": "3.2.1. Data Collection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "text", + "content": "We collect data using a DAVIS346 camera mounted on a tripod to monitor a fixed range of scenes. This camera has a spatial resolution of " + }, + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "inline_equation", + "content": "346 \\times 260" + }, + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "text", + "content": " and a temporal resolution of up to " + }, + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "inline_equation", + "content": "10^{6}" + }, + { + "bbox": [ + 313, + 367, + 556, + 559 + ], + "type": "text", + "content": " HZ. Our data collection spans from May 1, 2023, to June 1, 2024, covering various times of the day, seasons, and weather conditions. To accurately reflect the movement patterns of UAVs and the complexity of real-world scenarios, we capture UAVs performing diverse actions in various environments. These actions include forward and backward motion, lateral movements, ascending and descending maneuvers, as well as complex combinations of these behaviors. The dataset encompasses three different lighting conditions (normal, bright and low light) and various backgrounds, such as buildings, mountains, forests, urban areas, clouds, water surfaces, etc. We carefully annotated 147 sequences of data, totaling over 20.3 million target events." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 567, + 459, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 459, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 459, + 578 + ], + "type": "text", + "content": "3.2.2.Event by Event Annotation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Due to the asynchronous and sparse output of event cameras, manually annotating per-event data is highly time-consuming and inefficient. Existing event camera datasets are typically recorded using a combination of event cameras and standard RGB cameras. Labels are first extracted from the frame-based cameras and then transferred to the pixel coordinates of the event cameras through geometric transformations. However, this labeling strategy is limited to producing low frame rate bounding box annotations, which cannot match the high-frequency output characteristics of event cameras or meet the fine-grained, event-" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7211" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 141, + 121 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 141, + 121 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 141, + 121 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 141, + 121 + ], + "type": "image", + "image_path": "bb4e18c1f386c3ffd75789b83a3b4420cb5fcdd628f839db99905201e67a639c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 94, + 122, + 108, + 129 + ], + "lines": [ + { + "bbox": [ + 94, + 122, + 108, + 129 + ], + "spans": [ + { + "bbox": [ + 94, + 122, + 108, + 129 + ], + "type": "text", + "content": "Train" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 60, + 130, + 141, + 177 + ], + "blocks": [ + { + "bbox": [ + 60, + 130, + 141, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 130, + 141, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 130, + 141, + 177 + ], + "type": "image", + "image_path": "a12fba785f58aa730c6b3c85c4952ab984acd82577f79bc30de6ac2291fbf524.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 96, + 178, + 107, + 184 + ], + "lines": [ + { + "bbox": [ + 96, + 178, + 107, + 184 + ], + "spans": [ + { + "bbox": [ + 96, + 178, + 107, + 184 + ], + "type": "text", + "content": "Val" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 60, + 186, + 141, + 233 + ], + "blocks": [ + { + "bbox": [ + 60, + 186, + 141, + 233 + ], + "lines": [ + { + "bbox": [ + 60, + 186, + 141, + 233 + ], + "spans": [ + { + "bbox": [ + 60, + 186, + 141, + 233 + ], + "type": "image", + "image_path": "d3a5c188208442400bf692ee4ae2dc69452ad31868294460d1dcd34ee182ddb6.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 95, + 235, + 108, + 241 + ], + "lines": [ + { + "bbox": [ + 95, + 235, + 108, + 241 + ], + "spans": [ + { + "bbox": [ + 95, + 235, + 108, + 241 + ], + "type": "text", + "content": "Test" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 66, + 244, + 138, + 252 + ], + "lines": [ + { + "bbox": [ + 66, + 244, + 138, + 252 + ], + "spans": [ + { + "bbox": [ + 66, + 244, + 138, + 252 + ], + "type": "text", + "content": "(a) Position Distribution" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 148, + 234, + 188, + 243 + ], + "lines": [ + { + "bbox": [ + 148, + 234, + 188, + 243 + ], + "spans": [ + { + "bbox": [ + 148, + 234, + 188, + 243 + ], + "type": "text", + "content": "Extremely tiny" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 144, + 72, + 278, + 208 + ], + "blocks": [ + { + "bbox": [ + 144, + 72, + 278, + 208 + ], + "lines": [ + { + "bbox": [ + 144, + 72, + 278, + 208 + ], + "spans": [ + { + "bbox": [ + 144, + 72, + 278, + 208 + ], + "type": "image", + "image_path": "145dfde06b8a68d89d2e4cca763629db5c5f1567a000fc30c46659e9a4e4c782.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 158, + 221, + 178, + 228 + ], + "lines": [ + { + "bbox": [ + 158, + 221, + 178, + 228 + ], + "spans": [ + { + "bbox": [ + 158, + 221, + 178, + 228 + ], + "type": "text", + "content": "45.1%" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 203, + 221, + 221, + 228 + ], + "lines": [ + { + "bbox": [ + 203, + 221, + 221, + 228 + ], + "spans": [ + { + "bbox": [ + 203, + 221, + 221, + 228 + ], + "type": "text", + "content": "34.1%" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 205, + 234, + 219, + 242 + ], + "lines": [ + { + "bbox": [ + 205, + 234, + 219, + 242 + ], + "spans": [ + { + "bbox": [ + 205, + 234, + 219, + 242 + ], + "type": "text", + "content": "Tiny" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 244, + 221, + 264, + 228 + ], + "lines": [ + { + "bbox": [ + 244, + 221, + 264, + 228 + ], + "spans": [ + { + "bbox": [ + 244, + 221, + 264, + 228 + ], + "type": "text", + "content": "20.8%" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 281, + 72, + 352, + 119 + ], + "blocks": [ + { + "bbox": [ + 281, + 72, + 352, + 119 + ], + "lines": [ + { + "bbox": [ + 281, + 72, + 352, + 119 + ], + "spans": [ + { + "bbox": [ + 281, + 72, + 352, + 119 + ], + "type": "image", + "image_path": "fe73cf2aa51bbd4e501e8a8893dc97cb83a15c2e97e4c2a0db94318a72c4fecd.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 355, + 72, + 421, + 119 + ], + "blocks": [ + { + "bbox": [ + 355, + 72, + 421, + 119 + ], + "lines": [ + { + "bbox": [ + 355, + 72, + 421, + 119 + ], + "spans": [ + { + "bbox": [ + 355, + 72, + 421, + 119 + ], + "type": "image", + "image_path": "889655e607a4b646daae5327cfcb31bd76ffd3f9e910d904065a7527383b8891.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 382, + 120, + 466, + 128 + ], + "lines": [ + { + "bbox": [ + 382, + 120, + 466, + 128 + ], + "spans": [ + { + "bbox": [ + 382, + 120, + 466, + 128 + ], + "type": "text", + "content": "illumination:highlight " + }, + { + "bbox": [ + 382, + 120, + 466, + 128 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 382, + 120, + 466, + 128 + ], + "type": "text", + "content": " lowlight" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 427, + 72, + 495, + 119 + ], + "blocks": [ + { + "bbox": [ + 427, + 72, + 495, + 119 + ], + "lines": [ + { + "bbox": [ + 427, + 72, + 495, + 119 + ], + "spans": [ + { + "bbox": [ + 427, + 72, + 495, + 119 + ], + "type": "image", + "image_path": "49346c06e4c8037fef4e9aa5a23f6d2dc57bcf49b0d6a40dd431d2e0b9985e9e.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 497, + 72, + 568, + 119 + ], + "blocks": [ + { + "bbox": [ + 497, + 72, + 568, + 119 + ], + "lines": [ + { + "bbox": [ + 497, + 72, + 568, + 119 + ], + "spans": [ + { + "bbox": [ + 497, + 72, + 568, + 119 + ], + "type": "image", + "image_path": "870e145cc8dab9f7fb2a70b4018e916a280cd6b824e278ace52e35af75cea424.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 282, + 129, + 352, + 176 + ], + "blocks": [ + { + "bbox": [ + 282, + 129, + 352, + 176 + ], + "lines": [ + { + "bbox": [ + 282, + 129, + 352, + 176 + ], + "spans": [ + { + "bbox": [ + 282, + 129, + 352, + 176 + ], + "type": "image", + "image_path": "34f6d63a72d1a01f7892a1a092dc6e29d0de00ae2dc8a5ba8d98d73494779458.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 353, + 129, + 424, + 176 + ], + "blocks": [ + { + "bbox": [ + 353, + 129, + 424, + 176 + ], + "lines": [ + { + "bbox": [ + 353, + 129, + 424, + 176 + ], + "spans": [ + { + "bbox": [ + 353, + 129, + 424, + 176 + ], + "type": "image", + "image_path": "1f38712daafa71fc4d3801a430ff03b8f9778f21e103f6de445f2f915593bd7e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 425, + 129, + 494, + 176 + ], + "blocks": [ + { + "bbox": [ + 425, + 129, + 494, + 176 + ], + "lines": [ + { + "bbox": [ + 425, + 129, + 494, + 176 + ], + "spans": [ + { + "bbox": [ + 425, + 129, + 494, + 176 + ], + "type": "image", + "image_path": "852905a2a8ff0cba2ffc1c80a28d79d3dae825bbbe5a8c066a3d9c02c0807e6b.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 496, + 129, + 568, + 176 + ], + "blocks": [ + { + "bbox": [ + 496, + 129, + 568, + 176 + ], + "lines": [ + { + "bbox": [ + 496, + 129, + 568, + 176 + ], + "spans": [ + { + "bbox": [ + 496, + 129, + 568, + 176 + ], + "type": "image", + "image_path": "164bcf02ee65a9a17b2a8865cc3fb66c48bc739fc6cd5793b760a69bfdddb298.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 281, + 186, + 351, + 241 + ], + "blocks": [ + { + "bbox": [ + 247, + 234, + 263, + 241 + ], + "lines": [ + { + "bbox": [ + 247, + 234, + 263, + 241 + ], + "spans": [ + { + "bbox": [ + 247, + 234, + 263, + 241 + ], + "type": "text", + "content": "Small" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 179, + 245, + 259, + 253 + ], + "lines": [ + { + "bbox": [ + 179, + 245, + 259, + 253 + ], + "spans": [ + { + "bbox": [ + 179, + 245, + 259, + 253 + ], + "type": "text", + "content": "(b) Target size Distribution" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 281, + 186, + 351, + 241 + ], + "lines": [ + { + "bbox": [ + 281, + 186, + 351, + 241 + ], + "spans": [ + { + "bbox": [ + 281, + 186, + 351, + 241 + ], + "type": "image", + "image_path": "773817104496077a6772820be3efa6df8132a679a2bddf6972a2d00f4b28fdb2.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 353, + 186, + 424, + 241 + ], + "blocks": [ + { + "bbox": [ + 353, + 186, + 424, + 241 + ], + "lines": [ + { + "bbox": [ + 353, + 186, + 424, + 241 + ], + "spans": [ + { + "bbox": [ + 353, + 186, + 424, + 241 + ], + "type": "image", + "image_path": "9ab140ff0971fb1ac46dffe06a16cb49f4fa9ed0b03c82b9a002eb84494b7181.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 391, + 235, + 462, + 242 + ], + "lines": [ + { + "bbox": [ + 391, + 235, + 462, + 242 + ], + "spans": [ + { + "bbox": [ + 391, + 235, + 462, + 242 + ], + "type": "text", + "content": "Target Scale: small→large" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 244, + 459, + 253 + ], + "lines": [ + { + "bbox": [ + 389, + 244, + 459, + 253 + ], + "spans": [ + { + "bbox": [ + 389, + 244, + 459, + 253 + ], + "type": "text", + "content": "(c) Diversity of EV-UAV" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 425, + 186, + 494, + 233 + ], + "blocks": [ + { + "bbox": [ + 425, + 186, + 494, + 233 + ], + "lines": [ + { + "bbox": [ + 425, + 186, + 494, + 233 + ], + "spans": [ + { + "bbox": [ + 425, + 186, + 494, + 233 + ], + "type": "image", + "image_path": "a5cb2b9a2e6743d4664478abe835cbba71d435eb9d4f8aa284f00db6187b400f.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 496, + 186, + 568, + 233 + ], + "blocks": [ + { + "bbox": [ + 496, + 186, + 568, + 233 + ], + "lines": [ + { + "bbox": [ + 496, + 186, + 568, + 233 + ], + "spans": [ + { + "bbox": [ + 496, + 186, + 568, + 233 + ], + "type": "image", + "image_path": "3875e344e40fc6e401f948ab37afaf21856314b9f8907d62fc36493e1204a0fd.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 58, + 325, + 299, + 405 + ], + "blocks": [ + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "lines": [ + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 555, + 307 + ], + "type": "text", + "content": "Figure 2. Illustration of the proposed dataset EV-UAV. (a) Position distribution of EV-UAV. (b) Distribution of noise level and illumination of the proposed dataset. (c) EV-UAV collects diverse event streams across various conditions of illumination and target number and scale. N represents the noise level, reflecting the proportion of events in spatiotemporal space. T denotes the number of targets. S indicates the size of targets in the image plane." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 325, + 299, + 405 + ], + "lines": [ + { + "bbox": [ + 58, + 325, + 299, + 405 + ], + "spans": [ + { + "bbox": [ + 58, + 325, + 299, + 405 + ], + "type": "image", + "image_path": "f89e97b9cd390ae3be73f865acddf7d4de2f875b41acee2ebc9022b8cfe5e14f.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 409, + 296, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 487 + ], + "type": "text", + "content": "Figure 3. Event-by-event annotation method. Firstly, the events are accumulated into event frames, and targets are annotated with 2D bounding boxes on the event frames. Then, according to the time interval " + }, + { + "bbox": [ + 55, + 409, + 296, + 487 + ], + "type": "inline_equation", + "content": "\\Delta t" + }, + { + "bbox": [ + 55, + 409, + 296, + 487 + ], + "type": "text", + "content": " corresponding to each frame, the 2D annotations on the image plane are extended into 3D bounding box annotations in XYT 3D space, and all events within the 3D boxes are filtered as targets." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 31 + }, + { + "bbox": [ + 54, + 496, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 496, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 496, + 295, + 651 + ], + "type": "text", + "content": "by-event labeling requirements. Although manual event-by-event labeling could theoretically improve annotation accuracy, the process is highly labor-intensive and time-consuming, making it impractical for real-world applications. To address this issue, we propose a novel method that generates approximate event-by-event annotations using bounding box labels, as illustrated in the Fig. 3. This approach not only achieves fine-grained event-level annotations, enabling microsecond-level object detection, but also allows for degradation into bounding box annotations at any given time, thus supporting the learning requirements of some asynchronous event-based object detection algorithms." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 659, + 241, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 241, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 241, + 672 + ], + "type": "text", + "content": "3.3. Benchmark Features and Statistics" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Dataset Splitting. The EV-UAV dataset is divided into training (99 sequences), validation (24 sequences), and test sets (24 sequences), totaling 147 event sequences." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 324, + 555, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 408 + ], + "type": "text", + "content": "Position Distribution. As shown in Fig. 2(a), the targets are mostly located in the central area of the image. The larger horizontal variance suggests that the drone's movement is predominantly horizontal. Additionally, the range of target movements in the validation set is more varied. The fluctuations in the validation set are greater than those in the test set in corresponding directions." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": "Small-Scale Targets. In real-world scenarios, due to the long imaging distance and the small size of UAVs, drone targets appear at very small scales in the images. The general scale ranks" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "^1" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": " of [31] are no longer applicable. Following previous work [47], we further divide small targets into three levels: extremely tiny " + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "\\in [1, 8^2)" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": ", tiny " + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "\\in [8^2, 16^2)" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": ", and small " + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "\\in [16^2, \\infty)" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": ". Fig. 2(b) shows the number of annotations for each target category. It can be observed that the scale of all targets is smaller than " + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "32^2" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": ", with extremely tiny targets accounting for the largest proportion, approximately " + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "inline_equation", + "content": "45\\%" + }, + { + "bbox": [ + 313, + 408, + 556, + 563 + ], + "type": "text", + "content": ". Compared to other UAV detection datasets, our EV-UAV dataset has the smallest average UAV scale, as indicated in Tab. 1." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 563, + 556, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 648 + ], + "type": "text", + "content": "Rich Diversity. To represent real-world anti-UAV target detection scenarios, we endow EV-UAV with diversity. This helps the dataset cover various complex situations in anti-UAV detection both in time and space. As shown in Fig. 2(c), the diversity of EV-UAV is reflected not just in illumination levels but also in background complexity, target quantity, and scale." + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 654, + 427, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 427, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 427, + 666 + ], + "type": "text", + "content": "3.4. Evaluation Metrics" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 672, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 555, + 696 + ], + "type": "text", + "content": "In this work, we evaluate the event-by-event segmentation performance of the algorithm using Intersection over Union" + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "inline_equation", + "content": "{}^{1}" + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "text", + "content": " small " + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "inline_equation", + "content": "\\in \\left\\lbrack {1,{32}^{2}}\\right)" + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "text", + "content": " ,medium " + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "inline_equation", + "content": "\\in \\left\\lbrack {{32}^{2},{96}^{2}}\\right)" + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "text", + "content": " ,and large " + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "inline_equation", + "content": "\\in \\left\\lbrack {{96}^{2},\\infty }\\right)" + }, + { + "bbox": [ + 324, + 701, + 523, + 713 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7212" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 63, + 70, + 549, + 308 + ], + "blocks": [ + { + "bbox": [ + 63, + 70, + 549, + 308 + ], + "lines": [ + { + "bbox": [ + 63, + 70, + 549, + 308 + ], + "spans": [ + { + "bbox": [ + 63, + 70, + 549, + 308 + ], + "type": "image", + "image_path": "31fd6e9fb455be82f46e0d7537741caf16547b592e8df14c069dbb4fe4cccd40.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "lines": [ + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 54, + 316, + 555, + 361 + ], + "type": "text", + "content": "Figure 4. (a) The overview of the proposed Ev-SpSegNet. The network is built on a U-shaped architecture consisting of symmetric encoder and decoder. (b) The GDSCA module. The grouped dilated sparse convolution block is applied to features from different channels to extract local temporal features at multiple scales. The Sp-SE module is used to integrate these features. And the Patch Attention block is used to downsample voxel features and enable global context feature interaction." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "spans": [ + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "text", + "content": "(IoU) and accuracy (ACC), and assess the localization performance using probability of detection " + }, + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "inline_equation", + "content": "(P_d)" + }, + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "text", + "content": " and false alarm rate " + }, + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "inline_equation", + "content": "(F_a)" + }, + { + "bbox": [ + 54, + 381, + 298, + 430 + ], + "type": "text", + "content": ". More details regarding evaluation metrics can be found in the supplementary material." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 441, + 112, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 441, + 112, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 441, + 112, + 453 + ], + "type": "text", + "content": "4. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 295, + 616 + ], + "type": "text", + "content": "Frame-based detectors require converting events into synchronized image-like representations, which compresses the temporal information of events and makes it even more difficult to detect small targets that already lack distinct appearance cues. In raw event data, events generated by moving targets exhibit strong spatiotemporal correlations, forming elongated continuous curves in event point cloud space. Leveraging this characteristic, we propose EV-SpSegNet, which directly segments events generated by moving targets in sparse event point clouds. Additionally, we introduce a novel spatiotemporal correlation loss that evaluates the local spatiotemporal correlation of events to guide the network in retaining target events with continuous curves." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 624, + 142, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 142, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 142, + 636 + ], + "type": "text", + "content": "4.1. EV-SpSegNet" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 642, + 296, + 714 + ], + "type": "text", + "content": "Small moving targets form distinct continuous curves in spatiotemporal event point clouds. However, networks struggle to extract their features because these curves span long temporal ranges, and traditional convolution operations with limited receptive fields cannot effectively capture global features. While attention mechanisms can ex" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 381, + 555, + 610 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 610 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 610 + ], + "type": "text", + "content": "tract global context, their quadratic computational complexity restricts their ability to process large-scale event data. To effectively extract features from continuous curves in spatiotemporal event point clouds, we design a Grouped Dilated Sparse Convolution Attention (GDSCA) module, as shown in Fig. 4(b). This module first captures local multiscale features through the Grouped Dilated Sparse Convolution (GDSC) Block and then enables global feature interaction across patches using patch attention. Specifically, the GDSC Block groups input features along the channel dimension, with each group applying convolutions of different dilation rates to extract multi-scale local features. These features are then fused in the Sparse Squeeze and Excitation (Sp-SE) block. Finally, the patch attention block divides the point cloud into larger sub-regions and implements global feature interaction between sub-regions using self-attention. This allows the network to model the differences between targets and background noise in greater detail, enhancing its ability to distinguish moving targets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 619, + 492, + 632 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 619, + 492, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 492, + 632 + ], + "type": "text", + "content": "4.2. Spatiotemporal Correlation Loss" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 637, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 637, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 637, + 555, + 673 + ], + "type": "text", + "content": "In general image segmentation tasks, the loss function used for training the network is typically the standard pixel-wise binary cross-entropy (BCE) loss:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 354, + 684, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 684, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 354, + 684, + 555, + 717 + ], + "type": "interline_equation", + "content": "L _ {b c e} (p, y) = \\left\\{ \\begin{array}{c} - \\log (p), y = 1 \\\\ - \\log (1 - p), y = 0 \\end{array} \\right.. \\tag {1}", + "image_path": "3704637a50e3035764458a42f2040ae306a7acc6faa618f1b9f82ce647765b60.jpg" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7213" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 299, + 155 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 299, + 155 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 299, + 155 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 299, + 155 + ], + "type": "image", + "image_path": "37d6bcf7677342d7d683588c5fd4ea7dc5e26760ed8aa3de541b0c357eaaaf2a.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 160, + 295, + 184 + ], + "lines": [ + { + "bbox": [ + 55, + 160, + 295, + 184 + ], + "spans": [ + { + "bbox": [ + 55, + 160, + 295, + 184 + ], + "type": "text", + "content": "Figure 5. (a) The principle of spatiotemporal correlation loss. (b) Small object segmentation results of BCE loss and STC loss." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "spans": [ + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "text", + "content": "Although EV-SpSegNet computes structured outputs and focuses on large neighborhood information to better utilize long-term features, losses like BCE loss process each pixel independently. They do not consider the structure of neighborhoods, such as the spatiotemporal continuity of objects, which is crucial for describing the continuous curves formed by objects. As shown in Fig. 5(a), compared to the isolated noise " + }, + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "inline_equation", + "content": "p_2" + }, + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "text", + "content": ", the event " + }, + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "inline_equation", + "content": "p_1" + }, + { + "bbox": [ + 54, + 202, + 295, + 332 + ], + "type": "text", + "content": " with more high-confidence supporting events is more likely to be a true target. However, misclassifying both of them only results in a low coat of BCE loss." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": "To address this issue, we introduce a spatiotemporal correlation loss that encourages the network to retain more events with high spatiotemporal correlation while discarding more isolated noise. Specifically, this loss calculates the sum of the confidence of all supporting events within a " + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "inline_equation", + "content": "k \\times k \\times \\tau" + }, + { + "bbox": [ + 55, + 334, + 296, + 441 + ], + "type": "text", + "content": " neighborhood around the events, and defines it as the spatiotemporal correlation weight. We incorporate this weight into the cross-entropy loss and the spatiotemporal correlation loss is defined as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 65, + 445, + 295, + 478 + ], + "type": "interline_equation", + "content": "L _ {s t c} (p, y) = \\left\\{ \\begin{array}{c} - w _ {s t c} ^ {\\gamma} \\log (p), y = 1 \\\\ - \\left(1 - w _ {s t c}\\right) ^ {\\gamma} \\log (1 - p), y = 0 \\end{array} , \\right. \\tag {2}", + "image_path": "711640c8ad43808ae7be253c846b328239539a66fb09efae131140179a68bf4c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "content": "where the " + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "inline_equation", + "content": "y \\in \\{0,1\\}" + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "inline_equation", + "content": "p \\in (0,1)" + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "content": " represent the segmentation label and prediction result of the event, respectively. " + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "content": " is a hyperparameters to control the influence of spatiotemporal correlations. " + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "inline_equation", + "content": "w_{stc}" + }, + { + "bbox": [ + 55, + 483, + 296, + 544 + ], + "type": "text", + "content": " is the spatiotemporal correlation weight of each event, which is defined as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 109, + 549, + 295, + 575 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 549, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 109, + 549, + 295, + 575 + ], + "type": "interline_equation", + "content": "w _ {s t c} (p) = \\operatorname {s i g m o i d} \\left(\\sum_ {e ^ {j} \\in V ^ {k \\tau}} p ^ {j}\\right), \\tag {3}", + "image_path": "da8951ef2bde341b4a0333743f5c631c5442372599ece8bcf0aa5e8d9b1095bb.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "inline_equation", + "content": "V^{k\\tau}" + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "content": " represents the neighborhood of events with a size of " + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "inline_equation", + "content": "k \\times k \\times \\tau" + }, + { + "bbox": [ + 55, + 582, + 296, + 643 + ], + "type": "text", + "content": ". As illustrated in Fig. 5 (b1-b2), the STC loss preserves the continuous event trajectories generated by moving objects and eliminates false alarms caused by misclassification." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 182, + 685 + ], + "type": "text", + "content": "5.1. Experimental Settings" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We used 8-second events as input and set the size of each voxel to " + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "inline_equation", + "content": "1\\mathrm{pixel} \\times 1\\mathrm{pixel} \\times 1\\mathrm{ms}" + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": ". The batch size was set to" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "1, and all layers of the network were randomly initialized. We trained our network using the Adam optimizer [12] for 50 epochs with the learning rate linearly decaying from 1e-2 to 1e-3. The model was trained on the predefined training set of EV-UAV dataset, and we chose the models based on their best performance on the validation set and evaluate them on the test set. The hyperparameters " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": " of the STC loss are set to 3, 5, and 2, respectively. All models were implemented with PyTorch on a single NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 200, + 457, + 213 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 200, + 457, + 213 + ], + "spans": [ + { + "bbox": [ + 313, + 200, + 457, + 213 + ], + "type": "text", + "content": "5.2. Benchmark Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": "We compare the proposed method with several state-of-the-art methods, including 4 frame-based generic object detection methods (i.e., SSD [16], Faster RCNN [30], Deformable DETR [51], YOLOV10 [37]), 6 event based methods (i.e., RVT [7], SAST [25], GET [24], RED [26], EMS-YOLO [34], Spike-YOLO [17]), and 3 point cloud segmentation-based methods (i.e., KPCov, RandLA-Net, COSeg). We convert events within a " + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "inline_equation", + "content": "50~ms" + }, + { + "bbox": [ + 313, + 217, + 555, + 325 + ], + "type": "text", + "content": " time window into event frames [20] as input for frame-based methods." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 325, + 554, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 492 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 492 + ], + "type": "text", + "content": "Quantitative Results. As shown in Tab. 2, our EV-SpSegNet achieve the best detection performance with all evaluation metrics. We attribute this to two reasons. On the one hand, the GDSCA module combines local multiscale information with global context, enabling the network to better model the differences between targets and background noise. This enhances the ability of the network to distinguish moving targets. On the other hand, the spatiotemporal correlation loss encourages the network to identify events with high spatiotemporal correlation, improving detection performance. Additionally, by using sparse convolution to build the network, our method can inference 8 seconds of event data in " + }, + { + "bbox": [ + 313, + 325, + 554, + 492 + ], + "type": "inline_equation", + "content": "35.9~ms" + }, + { + "bbox": [ + 313, + 325, + 554, + 492 + ], + "type": "text", + "content": ", significantly outperforming other methods." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 493, + 555, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 649 + ], + "type": "text", + "content": "Qualitative Results. Qualitative results of different methods are shown in Fig. 6. It can be observed that our method can detect the most moving targets in all scenes, while the compared methods exhibit more or less missed detections and false positives. Particularly in scenes with complex backgrounds, such as the second and fourth rows with strong noise, the improvement of our method is more significant. Much to our surprise, our method even detects some targets that were missed by manual annotation due to strong background clutter, as shown in the third and fourth rows. We attribute this to the spatiotemporal correlation loss, which encourages the network to retain events with high spatiotemporal correlation." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 656, + 415, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 656, + 415, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 656, + 415, + 668 + ], + "type": "text", + "content": "5.3. Ablation Studies" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 674, + 466, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 466, + 686 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 466, + 686 + ], + "type": "text", + "content": "5.3.1. Contribution of Components" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "As shown in Tab. 3, adding the GDSC block to the baseline improves the model performance in terms of IoU, ACC," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7214" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 90, + 561, + 264 + ], + "blocks": [ + { + "bbox": [ + 153, + 71, + 463, + 82 + ], + "lines": [ + { + "bbox": [ + 153, + 71, + 463, + 82 + ], + "spans": [ + { + "bbox": [ + 153, + 71, + 463, + 82 + ], + "type": "text", + "content": "Table 2. Quantitative comparison of the proposed method to state-of-the-art methods." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 90, + 561, + 264 + ], + "lines": [ + { + "bbox": [ + 56, + 90, + 561, + 264 + ], + "spans": [ + { + "bbox": [ + 56, + 90, + 561, + 264 + ], + "type": "table", + "html": "
MethodsPublicationsEvent Rep.TemporalIoU(%)↑ACC(%)↑\\(P_d\\)(%)↑\\(F_a(10^{\\text{~-4~}})\\downarrow\\)#Params.Runtime(ms)
SSD [16]ECCV 2016Event CountNo25.3128.5626.31486.6325.2M2113
Faster RCNN [30]TPAMI 2016Event CountNo26.9329.6827.39689.6841.2M3962
DETR [51]ICLR 2020Event CountNo30.3533.6331.64631.3739.8M3136
YOLOV10-S [37]NIPS 2025Event CountNo32.5533.3932.18589.677.3M1627
EMS-YOLO [34]ICCV 2023SNNYes36.7742.9250.68112.363.3M1229
Spike-YOLO [17]ECCV 2024SNNYes43.9448.2659.6255.3869.0M1883
GET [24]ICCV 2023Groupv TokenYes40.3148.9160.7346.3518.4M2168
RED [26]NeurIPS 2020Voxel GridYes35.9945.5453.76102.2724.1M3427
RVT [7]CVPR 2023Voxel GridYes43.2151.3860.3555.689.9M1737
SAST [25]CVPR 2024Voxel GridYes34.3140.2251.21150.3218.5M3075
KPCov [36]ICCV 2019PointsYes48.1957.2868.5916.3250.1M562
RandLA-Net [9]CVPR 2020PointsYes50.3259.2970.566.951.2M353
COSeg [2]CVPR 2024PointsYes51.8960.9371.329.2123.4M364
Ours-PointsYes55.1865.0277.531.634.0M35.9
", + "image_path": "e16c7b6640ff3e05a7d0fbdc7eeb581e10029a4778135be3f72acd8e81bef8c9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 67, + 266, + 355, + 277 + ], + "lines": [ + { + "bbox": [ + 67, + 266, + 355, + 277 + ], + "spans": [ + { + "bbox": [ + 67, + 266, + 355, + 277 + ], + "type": "text", + "content": "The bold and the underline represent the best and second-best performance, respectively." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 75, + 285, + 536, + 529 + ], + "blocks": [ + { + "bbox": [ + 75, + 285, + 536, + 529 + ], + "lines": [ + { + "bbox": [ + 75, + 285, + 536, + 529 + ], + "spans": [ + { + "bbox": [ + 75, + 285, + 536, + 529 + ], + "type": "image", + "image_path": "ed49ab74805cc3b98d74b4d3788f92edb573fcbc12a5600795508baf07ae0b4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "lines": [ + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 536, + 555, + 559 + ], + "type": "text", + "content": "Figure 6. Qualitative results of different methods. Our method accurately detects small moving targets, while other methods produce numerous missed detections and false alarms." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 580, + 297, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 580, + 297, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 580, + 297, + 712 + ], + "type": "text", + "content": "and " + }, + { + "bbox": [ + 55, + 580, + 297, + 712 + ], + "type": "inline_equation", + "content": "P_{d}" + }, + { + "bbox": [ + 55, + 580, + 297, + 712 + ], + "type": "text", + "content": " by 1.37, 2.43 and 1.18, which demonstrates that extracting multi-scale local features using GDSC is beneficial. However, adding only the PA block causes a performance drop. This is because downsampling is applied before computing global attention, and without sufficient local receptive fields, the model cannot effectively capture global features, leading to feature confusion and performance degradation. It is worth noting that when the GDSC block and PA block are used together, the performance improves significantly. This indicates that capturing sufficient local features within patches before enabling global feature interaction is" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 580, + 555, + 628 + ], + "type": "text", + "content": "essential. Additionally, using the STC loss consistently improves model performance, which shows that guiding the network to retain events with high spatiotemporal correlation through STCLoss is effective." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 638, + 518, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 638, + 518, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 638, + 518, + 649 + ], + "type": "text", + "content": "5.3.2. The Influence of the Number of Branches" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 555, + 712 + ], + "type": "text", + "content": "As depicted in Tab. 4, we conduct an ablation experiment to study the influence of the number of branches in GDSC block on the final detection performance. It can be observed that the network achieves best performance when the number of branches is set to 4. This is because too" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7215" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 90, + 294, + 175 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 293, + 82 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 293, + 82 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 293, + 82 + ], + "type": "text", + "content": "Table 3. Ablation study on framework scheme and loss function." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 90, + 294, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 90, + 294, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 90, + 294, + 175 + ], + "type": "table", + "html": "
GDSCPALstcIoU↑ACC↑Pd↑Fa↓#Params.
51.3660.2171.944.815.6M
✓52.7362.6473.124.283.8M
✓51.2659.5371.674.315.8M
✓53.0761.0174.133.665.6M
✓✓53.6264.0276.761.934.0M
✓✓✓55.1865.0277.531.634.0M
", + "image_path": "9bfdd2b2523290491d4058ad51fb33842a6cdab61505279bc0ee3e08929317ca.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 205, + 296, + 278 + ], + "blocks": [ + { + "bbox": [ + 56, + 186, + 294, + 197 + ], + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 197 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 197 + ], + "type": "text", + "content": "Table 4. The influence of the number of branches in GDSC block." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 205, + 296, + 278 + ], + "lines": [ + { + "bbox": [ + 56, + 205, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 296, + 278 + ], + "type": "table", + "html": "
#BranchesIoU↑ACC↑Pd↑Fa↓#Params.
150.2162.3673.652.975.6M
251.4164.1675.212.534.5M
353.5264.2776.662.014.3M
455.1865.0277.531.634.0M
553.1264.3176.532.113.8M
", + "image_path": "5ba3846b7103981576a36b70cc00833b02e1c1eae762297b1b9cd23bad475b3a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 310, + 298, + 371 + ], + "blocks": [ + { + "bbox": [ + 56, + 290, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 290, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 290, + 294, + 300 + ], + "type": "text", + "content": "Table 5. Combination of convolutions with different dilation rates." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 310, + 298, + 371 + ], + "lines": [ + { + "bbox": [ + 56, + 310, + 298, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 298, + 371 + ], + "type": "table", + "html": "
Dilated ratesIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
(1,2,3,4)55.1865.0277.531.63
(1,2,3,5)53.2164.1176.932.41
(1,3,5,7)52.7564.0276.894.61
(1,3,5,9)51.3563.9275.924.76
", + "image_path": "e0d7b90d21ce889253175f5b40e3da3723828e23c4de910fcc6b16014b018c90.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 393, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 393, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 393, + 295, + 453 + ], + "type": "text", + "content": "many branches result in insufficient features assigned to each channel, making it difficult to capture features at the corresponding scale. On the other hand, too few branches lead to insufficient dilation rates, preventing the network from capturing features at multiple scales." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 460, + 274, + 471 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 274, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 274, + 471 + ], + "type": "text", + "content": "5.3.3. The Influence of Dilation Rate Combinations" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 295, + 571 + ], + "type": "text", + "content": "We next conduct an ablation study to analyze the relationship between detection performance and different dilation rate combinations. Tab. 5 illustrates that setting the dilation rate to (1,2,3,4) achieves the best performance. This indicates that the dilation rate is not always better when larger. An excessively large dilation rate causes the corresponding events to be too far apart, making it difficult to effectively determine whether these events have temporal correlation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 578, + 212, + 589 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 578, + 212, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 578, + 212, + 589 + ], + "type": "text", + "content": "5.3.4. The Effectiveness of STC Loss" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To further validate the generalizability of our proposed spatiotemporal correlation Loss, we integrate it with several state-of-the-art methods, including KPConv [36], RandLA-Net [9], and COSeg [2]. As shown in Tab. 6, the addition of " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{stc}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " consistently improves the performance of all methods. For example, KPConv with " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{stc}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " achieves improvements of " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "2.13\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.75\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.4\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " in terms of IoU, ACC and " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "P_d" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", respectively, and the false alarm rate decreases by " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "5.72\\%" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". These results show that " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "L_{stc}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " effectively enhances the ability of existing methods to capture spatiotemporal correlations," + } + ] + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 315, + 102, + 558, + 185 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 92 + ], + "type": "text", + "content": "Table 6. The performance of spatiotemporal correlation loss function on different methods." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 102, + 558, + 185 + ], + "lines": [ + { + "bbox": [ + 315, + 102, + 558, + 185 + ], + "spans": [ + { + "bbox": [ + 315, + 102, + 558, + 185 + ], + "type": "table", + "html": "
MethodIoU(%)↑ACC(%)↑Pd(%)↑Fa(10-4)↓
KPConv48.1957.2868.5916.32
KPConv*50.32(+2.13)59.03(+1.75)69.99(+1.40)10.60(-5.72)
RandLA50.3259.2970.569.21
RandLA*51.15(+0.83)60.63(+1.34)72.36(+2.37)8.91(-1.69)
COSeg51.8960.9371.326.95
COSeg*53.12(+1.23)62.31(0.99)73.96(+2.64)5.86(-1.09)
", + "image_path": "046b71c7356eaa8ea785580fb2d1b841d31c5014c40da400650767a03a97e737.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 326, + 186, + 454, + 196 + ], + "lines": [ + { + "bbox": [ + 326, + 186, + 454, + 196 + ], + "spans": [ + { + "bbox": [ + 326, + 186, + 454, + 196 + ], + "type": "text", + "content": "* represent the methods with STC loss." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 555, + 291 + ], + "type": "text", + "content": "leading to better segmentation accuracy and reduced false alarms. The consistent performance gains across diverse methods prove the robustness and general applicability of our proposed loss function. In addition, we explore the selection of optimal hyperparameters for the STC loss. For more details, please refer to the supplementary material." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 301, + 468, + 314 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 468, + 314 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 468, + 314 + ], + "type": "text", + "content": "6. Discussion and Limitations" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 322, + 555, + 453 + ], + "type": "text", + "content": "Although our EV-UAV accurately replicates real-world scenarios for detecting tiny UAVs in the wild, event cameras only capture motion information. When targets are stationary or move slowly, event cameras produce no output, leading to detection failures. In contrast, frame-based cameras can directly provide detailed texture information (i.e., absolute brightness) in static scenes, making them complementary to event cameras. We believe that developing a dual-modal dataset and extending our method to incorporate multi-modal inputs to maximize object detection accuracy represents a promising direction for future work." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 388, + 478 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 676 + ], + "type": "text", + "content": "In this paper, we introduce EV-UAV, the first large-scale benchmark for anti-UAV tasks. EV-UAV is a highly challenging benchmark, featuring diverse scenarios and extremely small targets. It provides event-level fine-grained annotations, supporting algorithms with different event representations. Furthermore, based on the observation that small moving targets form continuous curves in event point cloud space, we propose EV-SpSegNet, a novel baseline for segmenting target motion trajectories from sparse event point clouds, along with a Spatiotemporal Correlation (STC) loss that leverages spatiotemporal local correlation to guide the network in retaining more target events. Based on the proposed dataset, we conduct a comprehensive evaluation of our method and 13 state-of-the-art approaches, demonstrating the superiority of our method and providing a benchmark for future research in EVSOD." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 713 + ], + "type": "text", + "content": "Acknowledgement: This work was supported by the Young Elite Scientist Sponsorship Program by China Association for Science and Technology (NO.YESS20240427)." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7216" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 122 + ], + "type": "text", + "content": "[1] Inigo Alonso and Ana C Murillo. Ev-segnet: Semantic segmentation for event-based cameras. In CVPRW, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 168 + ], + "type": "text", + "content": "[2] Zhaochong An, Guolei Sun, Yun Liu, Fayao Liu, Zongwei Wu, Dan Wang, Luc Van Gool, and Serge Belongie. Rethinking few-shot 3d point cloud semantic segmentation. In CVPR, pages 3996-4006, 2024. 2, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 213 + ], + "type": "text", + "content": "[3] Ryad Benosman, Charles Clercq, Xavier Lagorce, Sio-Hoi Ieng, and Chiara Bartolozzi. Event-based visual flow. IEEE Trans. Neural Networks and Learning Systems, 25(2):407-417, 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 258 + ], + "type": "text", + "content": "[4] Loic Cordone, Benoit Miramond, and Philippe Thierion. Object detection with spiking neural networks on automotive event data. In IEEE International Joint Conference on Neural Networks (IJCNN), pages 1-8, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 294, + 304 + ], + "type": "text", + "content": "[5] Pierre De Tournemire, Davide Nitti, Etienne Perot, Davide Migliore, and Amos Sironi. A large scale event-based detection dataset for automotive. arXiv preprint arXiv:2001.08499, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 294, + 350 + ], + "type": "text", + "content": "[6] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 294, + 384 + ], + "type": "text", + "content": "[7] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In CVPR, pages 13884-13893, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 294, + 429 + ], + "type": "text", + "content": "[8] Mathias Gehrig, Willem Aarents, Daniel Gehrig, and Davide Scaramuzza. Dsec: A stereo event camera dataset for driving scenarios. IEEE Robotics and Automation Letters, 6(3): 4947-4954, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 294, + 475 + ], + "type": "text", + "content": "[9] Qingyong Hu, Bo Yang, Linhai Xie, Stefano Rosa, Yulan Guo, Zhihua Wang, Niki Trigoni, and Andrew Markham. Randla-net: Efficient semantic segmentation of large-scale point clouds. In CVPR, pages 11108-11117, 2020. 2, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 294, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 294, + 519 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 294, + 519 + ], + "type": "text", + "content": "[10] Bo Huang, Jianan Li, Junjie Chen, Gang Wang, Jian Zhao, and Tingfa Xu. Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild. 46 (5):2852-2865, 2023. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 521, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 521, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 57, + 521, + 294, + 565 + ], + "type": "text", + "content": "[11] Nan Jiang, Kuiran Wang, Xiaoke Peng, Xuehui Yu, Qiang Wang, Junliang Xing, Guorong Li, Jian Zhao, Guodong Guo, and Zhenjun Han. Anti-uav: A large multi-modal benchmark for uav tracking. arXiv preprint arXiv:2101.08466, 2021. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 567, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 294, + 589 + ], + "type": "text", + "content": "[12] Diederik P Kingma and Jimmy Lei Ba. Adam: Amethod for stochastic optimization. In ICLR, pages 1-15, 2014. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[13] Chunggi Lee, Seonwook Park, Heon Song, Jeongun Ryu, Sanghoon Kim, Haejoon Kim, Sérgio Pereira, and Donggeun Yoo. Interactive multi-class tiny-object detection. In CVPR, pages 14136-14145, 2022. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 678 + ], + "type": "text", + "content": "[14] Jianing Li, Jia Li, Lin Zhu, Xijie Xiang, Tiejun Huang, and Yonghong Tian. Asynchronous spatio-temporal memory network for continuous event-based object detection. IEEE TIP, 31:2975-2987, 2022. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 294, + 713 + ], + "type": "text", + "content": "[15] Zhengqi Li, Simon Niklaus, Noah Snavely, and Oliver Wang. Neural scene flow fields for space-time view synthesis of dynamic scenes. In CVPR, pages 6498-6508, 2021. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[16] Wei Liu, Dragomir Anguelov, Dumitru Erhan, Christian Szegedy, Scott Reed, Cheng-Yang Fu, and Alexander C Berg. Ssd: Single shot multibox detector. In ECCV, pages 21-37. Springer, 2016. 2, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[17] Xinhao Luo, Man Yao, Yuhong Chou, Bo Xu, and Guoqi Li. Integer-valued training and spike-driven inference spiking neural network for high-performance and energy-efficient object detection. In ECCV, 2024. 2, 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 206 + ], + "type": "text", + "content": "[18] Gabriele Magrini, Federico Becattini, Pietro Pala, Alberto Del Bimbo, and Antonio Porta. Neuromorphic drone detection: an event-rgb multimodal approach. arXiv preprint arXiv:2409.16099, 2024. 2, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 553, + 262 + ], + "type": "text", + "content": "[19] Jakub Mandula, Jonas Kühne, Luca Pascarella, and Michele Magno. Towards real-time fast unmanned aerial vehicle detection using dynamic vision sensors. In 2024 IEEE International Instrumentation and Measurement Technology Conference (I2MTC), pages 1-6. IEEE, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 317, + 264, + 553, + 307 + ], + "type": "text", + "content": "[20] Ana I. Maqueda, Antonio Loquercio, Guillermo Gallego, Narciso Garcia, and Davide Scaramuzza. Event-based vision meets deep learning on steering prediction for self-driving cars. In CVPR, pages 5419-5427, 2018. 3, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[21] Anton Mitrokhin, Zhiyuan Hua, Cornelia Fermuller, and Yiannis Aloimonos. Learning visual motion segmentation using event surfaces. In CVPR, pages 14414-14423, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 343, + 553, + 365 + ], + "type": "text", + "content": "[22] Francesco Nex and Fabio Remondino. Uav for 3d mapping applications: a review. Applied geomatics, 6:1-15, 2014. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 419 + ], + "type": "text", + "content": "[23] Federico Paredes-Valles, Kirk YW Scheper, and Guido CHE De Croon. Unsupervised learning of a hierarchical spiking neural network for optical flow estimation: From events to global motion perception. IEEE TPAMI, 42(8):2051-2064, 2019. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "spans": [ + { + "bbox": [ + 316, + 422, + 553, + 454 + ], + "type": "text", + "content": "[24] Yansong Peng, Yueyi Zhang, Zhiwei Xiong, Xiaoyan Sun, and Feng Wu. Get: group event transformer for event-based vision. In ICCV, pages 6038-6048, 2023. 2, 6, 7" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[25] Yansong Peng, Hebei Li, Yueyi Zhang, Xiaoyan Sun, and Feng Wu. Scene adaptive sparse transformer for event-based object detection. In CVPR, pages 16794-16804, 2024. 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 533 + ], + "type": "text", + "content": "[26] Etienne Perot, Pierre De Tournemire, Davide Nitti, Jonathan Masci, and Amos Sironi. Learning to detect objects with a 1 megapixel event camera. NeurIPS, 33:16639-16652, 2020. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 553, + 588 + ], + "type": "text", + "content": "[27] Christoph Posch, Daniel Matolin, and Rainer Wohlgenannt. A qvga 143 db dynamic range frame-free pwm image sensor with lossless pixel-level video compression and time-domain cds. IEEE Journal of Solid-State Circuits, 46(1):259-275, 2010. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[28] Panagiotis Radoglou-Grammatikis, Panagiotis Sarigiannidis, Thomas Lagkas, and Ioannis Moscholios. A compilation of uav applications for precision agriculture. Computer Networks, 172:107148, 2020. 1" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 668 + ], + "type": "text", + "content": "[29] Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Real-time visual-inertial odometry for event cameras using keyframe-based nonlinear optimization. In BMVC, 2017. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[30] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. IEEE transactions on pattern analysis and machine intelligence, 39(6):1137-1149, 2016. 2, 6, 7" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7217" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[31] Hamid Rezatofighi, Nathan Tsoi, JunYoung Gwak, Amir Sadeghian, Ian Reid, and Silvio Savarese. Generalized intersection over union: A metric and a loss for bounding box regression. In CVPR, pages 658-666, 2019. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 151 + ], + "type": "text", + "content": "[32] Simon Schaefer, Daniel Gehrig, and Davide Scaramuzza. Aegnn: Asynchronous event-based graph neural networks. In CVPR, pages 12371-12381, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[33] Sumit Bam Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. arXiv preprint arXiv.1810.08646, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 228 + ], + "type": "text", + "content": "[34] Qiaoyi Su, Yuhong Chou, Yifan Hu, Jianing Li, Shijie Mei, Ziyang Zhang, and Guoqi Li. Deep directly-trained spiking neural networks for object detection. In ICCV, pages 6555-6565, 2023. 2, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[35] Zhaoning Sun, Nico Messikommer, Daniel Gehrig, and Davide Scaramuzza. Ess: Learning event-based semantic segmentation from still images. In ECCV, pages 341-357. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[36] Hugues Thomas, Charles R Qi, Jean-Emmanuel Deschaud, Beatriz Marcotegui, François Goulette, and Leonidas J Guibas. Kpconv: Flexible and deformable convolution for point clouds. In ICCV, pages 6411–6420, 2019. 2, 7, 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 354 + ], + "type": "text", + "content": "[37] Ao Wang, Hui Chen, Lihao Liu, Kai Chen, Zijia Lin, Jungong Han, et al. Yolov10: Real-time end-to-end object detection. NeurIPS, 37:107984-108011, 2025. 2, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "text", + "content": "[38] Huan Wang, Luping Zhou, and Lei Wang. Miss detection vs. false alarm: Adversarial learning for small object segmentation in infrared images. In ICCV, pages 8509-8518, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 400, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 454 + ], + "type": "text", + "content": "[39] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 510 + ], + "type": "text", + "content": "[40] Xiao Wang, Jianing Li, Lin Zhu, Zhipeng Zhang, Zhe Chen, Xin Li, Yaowei Wang, Yonghong Tian, and Feng Wu. Visevent: Reliable object tracking via collaboration of frame and event flows. IEEE Transactions on Cybernetics, 54(3): 1997-2010, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 294, + 566 + ], + "type": "text", + "content": "[41] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 622 + ], + "type": "text", + "content": "[42] Xiao Wang, Yu Jin, Wentao Wu, Wei Zhang, Lin Zhu, Bo Jiang, and Yonghong Tian. Object detection using event camera: A moe heat conduction based detector and a new benchmark dataset. arXiv preprint arXiv:2412.06647, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 677 + ], + "type": "text", + "content": "[43] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset and a novel baseline. In CVPR, pages 19248-19257, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[44] Xiao Wang, Shiao Wang, Chuanming Tang, Lin Zhu, Bo Jiang, Yonghong Tian, and Jin Tang. Event stream-based visual object tracking: A high-resolution benchmark dataset" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 409 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "and a novel baseline. In CVPR, pages 19248-19257, 2024. 2,3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[45] Chang Xu, Jinwang Wang, Wen Yang, Huai Yu, Lei Yu, and Gui-Song Xia. Rfla: Gaussian receptive field based label assignment for tiny object detection. In ECCV, pages 526-543. Springer, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 184 + ], + "type": "text", + "content": "[46] Chenhongyi Yang, Zehao Huang, and Naiyan Wang. Querydet: Cascaded sparse query for accelerating high-resolution small object detection. In CVPR, pages 13668-13677, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 186, + 553, + 239 + ], + "type": "text", + "content": "[47] Xinyi Ying, Chao Xiao, Ruojing Li, Xu He, Boyang Li, Xu Cao, Zhaoxu Li, Yingqian Wang, Mingyuan Hu, Qingyu Xu, et al. Visible-thermal tiny object detection: A benchmark dataset and baselines. arXiv preprint arXiv:2406.14482, 2024. 2, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 295 + ], + "type": "text", + "content": "[48] Emre O. Neftci; Hesham Mostafa; Friedemann Zenke. Surrogate gradient learning in spiking neural networks: Bringing the power of gradient-based optimization to spiking neural networks. IEEE Signal Processing Magazine, pages 51-63, 2019. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 330 + ], + "type": "text", + "content": "[49] Mingjin Zhang, Rui Zhang, Yuxiang Yang, Haichen Bai, Jing Zhang, and Jie Guo. Isnet: Shape matters for infrared small target detection. In CVPR, pages 877-886, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 374 + ], + "type": "text", + "content": "[50] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, pages 989-997, 2019. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 553, + 409 + ], + "type": "text", + "content": "[51] Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transformers for end-to-end object detection. In ICLR, 2020. 2, 6, 7" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7218" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_content_list.json b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a777c55b716a59cb8d7b77cd82b780d881fb101f --- /dev/null +++ b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_content_list.json @@ -0,0 +1,1752 @@ +[ + { + "type": "text", + "text": "Event-based Visual Vibrometry", + "text_level": 1, + "bbox": [ + 334, + 131, + 656, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Xinyu Zhou $^{1}$ Peiqi Duan $^{2,3}$ Yeliduosi Xiaokaiti $^{2,3}$ Chao Xu $^{1}$ Boxin Shi $^{2,3*}$", + "bbox": [ + 174, + 172, + 833, + 191 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University", + "bbox": [ + 140, + 193, + 870, + 209 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{2}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University", + "$^{3}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 174, + 210, + 833, + 243 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhouxiny, duanqi0001, shiboxin}@pku.edu.cn, yongqiye@stu.pku.edu.cn, xuchao@cis.pku.edu.cn", + "bbox": [ + 101, + 247, + 903, + 262 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e87abb35354628ef3e818bbf9bfaf522a6f8bb51ecd41d96ff80fcb78a13f4d0.jpg", + "image_caption": [ + "(a) Hybrid camera system" + ], + "image_footnote": [], + "bbox": [ + 99, + 273, + 338, + 420 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a563896e06d22b14780ef5168626cd4c2a85ed587f15ebe298f91fadd1ef7e0d.jpg", + "image_caption": [ + "(b) Frame + Event stream" + ], + "image_footnote": [], + "bbox": [ + 348, + 297, + 552, + 398 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb2ab6f4cd3402565b4c0d7bd9267e5f3b00fe1e3137c76260d95342e97c0402.jpg", + "image_caption": [ + "Pixel Displacement", + "Figure 1. Overview of event-based visual vibrometry. A hybrid camera system, comprising a frame-based camera and an event camera, is utilized to capture subtle object vibrations (a). We propose an event-based subtle motion estimation method that leverages the event stream and a reference frame as input (b). The subtle motion field is extracted at each pixel location across time (c). This motion field can be further analyzed to extract the global motion spectrum and independent vibration modes (d). Each mode is characterized by a resonant frequency and a corresponding mode shape, which manifest as distinct peaks in the motion spectrum." + ], + "image_footnote": [], + "bbox": [ + 584, + 276, + 725, + 441 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92c97e841170be6ade1bd30cc895c9c0933f399b26f34b6821aff66712ef12e8.jpg", + "image_caption": [ + "Motion Spectrum" + ], + "image_footnote": [], + "bbox": [ + 741, + 276, + 893, + 441 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 550, + 326, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Visual vibrometry has emerged as a powerful technique for remote acquisition of audio and the physical properties of materials. To capture high-frequency vibrations, frame-based approaches often require a high-speed video camera and bright lighting to compensate for the short exposure time. In this paper, we introduce event-based visual vibrometry, a new high-speed visual vibration sensing method using an event camera. By leveraging the high temporal resolution and low bandwidth characteristics of event cameras, event-based visual vibrometry enables high-speed vibration sensing under ambient lighting conditions with improved data efficiency. Specifically, we leverage a hybrid camera system and propose an event-based subtle motion estimation framework that integrates an optimization-based approach based on the event generation model and a motion refinement network. We demonstrate our method by capturing vibration caused by audio sources and estimating material properties for various objects.", + "bbox": [ + 88, + 584, + 485, + 858 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 550, + 643, + 565 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vibrations are pervasive phenomena that arise from a wide range of sources, including engines, percussive impacts, and musical instruments. These vibrations typically exhibit small amplitudes and a broad spectrum of frequencies, spanning from a few Hz to MHz, and carry an immense amount of information. Vibration analysis is an established tool in various engineering and scientific fields, including the remote acquisition of audio signals [8], the monitoring of human heart rate [38, 56], and the estimation of material properties [2, 9, 13].", + "bbox": [ + 509, + 579, + 906, + 731 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Various devices have been used to measure and analyze vibrations. Among these, contact sensors [48] and laser vibrometers [11] are two widely used traditional vibration sensors, both of which are specialized instruments and only measure the vibration of a single surface point. Recently, much progress has been made on visual vibrometry [2, 4, 9, 13], an emerging technique that utilizes general-purpose video cameras for vibration measurement. In contrast to traditional vibration sensors, video-based approaches enable spatially dense measurements of surface motion. However, the acquisition of high-frequency vi", + "bbox": [ + 509, + 734, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author", + "bbox": [ + 114, + 886, + 243, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "24666", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "briations requires high-speed cameras coupled with bright lighting to compensate for the short exposure time, which is impractical in many real-world applications. Additionally, high-speed cameras demand substantial bandwidth requirements. To eliminate the need for high-speed cameras, Sheinin et al. [43] propose a dual-shutter system combined with active speckle-based techniques. Nevertheless, the nature of speckle-based vibration imaging restricts its applicability and confines vibration sampling to only several surface points. In short, frame-based visual vibrometry encounters challenges in efficiently capturing high-frequency vibrations while maintaining bandwidth efficiency and measurement convenience.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event cameras [14, 25], characterized by their high temporal resolution, high dynamic range, and low bandwidth requirements, have emerged as a promising solution for high-speed sensing [16]. Unlike conventional cameras, event cameras asynchronously record per-pixel logarithmic scene radiance changes. This signal-triggering mechanism is particularly advantageous for vibration sensing, where radiance changes occur only in a small fraction of pixels, leading to a significantly reduced data rate. The high temporal resolution and dynamic range inherent to event cameras enable the measurement of high-frequency vibrations under ambient lighting. These characteristics position event cameras as potential alternatives to address the constraints of frame-based visual vibrometry.", + "bbox": [ + 89, + 290, + 482, + 501 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite the potential of event cameras, estimating subtle motion from event data remains an open problem. Firstly, although event cameras have been applied to small motion scenarios for motion magnification [7], this approach focuses on visualizing small motions and adopts an implicit motion representation. However, the explicit estimation of subtle motion is essential in visual vibrometry for vibration analysis. Secondly, adapting existing event-based motion estimation techniques to vibration scenarios presents significant difficulties. Existing event-based optical flow estimation methods [19, 37, 47, 58] generally assume continuous apparent motion of visual patterns, which is violated by the oscillatory nature of vibrations. Events caused by vibrations are primarily localized along object edges, and this spatial sparsity hinders the effectiveness of event alignment techniques like Contrast Maximization [37, 47]. Moreover, in comparison to general scenarios, vibrations tend to produce significantly fewer events. Estimating subtle motion from such sparse event data is highly vulnerable to noise inherent in event cameras [23].", + "bbox": [ + 89, + 505, + 482, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we propose event-based visual vibrometry as an efficient and practical solution for vibration measurement and analysis. To address the challenges mentioned above, we employ an Event-RGB hybrid camera system [7, 20, 55] (Fig. 1(a)), which combines event streams with spatial texture information from conventional frames. We pro", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "pose an event-based subtle motion estimation framework that takes event streams and a single reference frame as inputs (Fig. 1(b)). Initially, we employ an optimization-based approach grounded in the event generation model [21, 36] to derive coarse motion estimates within short temporal windows. The sparsity of event data restricts the validity of these coarse estimates to only event-triggered regions, while simultaneously introducing undesirable smoothing effects. Additionally, quantization errors in event cameras compromise the accuracy of motion amplitude estimation. To address these issues, we adopt a motion refinement network that aggregates historical motion information to enhance the coarse motion estimation and obtain long-range vibration displacements (Fig. 1(c)). The refined motion fields across time can subsequently be analyzed (Fig. 1(d)) for various visual vibrometry applications. The key contributions of our work are summarized as follows:", + "bbox": [ + 511, + 90, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- a bandwidth-efficient visual vibrometry approach that leverages a hybrid camera system to measure high-frequency vibration under ambient lighting conditions;", + "- an event-based subtle motion estimation framework that integrates a model-based optimization approach and a motion refinement network;", + "- a thorough evaluation on both synthetic data and real-world vibrations caused by various sources, demonstrating the effectiveness of event-based visual vibrometry." + ], + "bbox": [ + 513, + 348, + 903, + 484 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 500, + 661, + 515 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Traditional vibration sensors. This section reviews two prevalent traditional vibration measurement devices: contact sensors and active sensors. Contact sensors operate through direct physical contact with the target object [48]. Accelerometers and piezoelectric pickups are representative examples of contact sensors that respond to acceleration. Among active sensors, laser Doppler vibrometers (LDVs) are the most common type used for vibration measurement [11, 41]. LDVs operate by projecting a laser beam onto the target object and calculating the surface velocity from the phase shift in the reflected light. LDVs have been used to find defects in composite materials [3, 5], examine the integrity of building structures [40], and modal analysis [30, 31]. Both contact sensors and laser vibrometers are typically limited to measuring vibrations at a single point on the surface.", + "bbox": [ + 511, + 531, + 903, + 772 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Frame-based visual vibrometry. Recent studies on video motion magnification [53, 54, 56] have demonstrated the capability to detect subtle motions captured in videos. These methods typically exploit spatial phase variations of the complex steerable pyramid [39, 49, 50] to estimate small local motions. Chen et al. [6] applied this phase-based subtle motion estimation method to quantify the vibration modes of pipes and cantilever beams. Davis", + "bbox": [ + 511, + 780, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "24667", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "et al. [8] analyzed object vibrations induced by acoustic excitation to reconstruct the original sound. Davis et al. [9] employed the motion spectra extracted from video to estimate the physical properties of objects. Feng et al. [13] extended this methodology to heterogeneous properties and proposed a physics-constrained optimization approach. Capturing high-frequency vibrations requires expensive 2D high-speed cameras. To address this cost, some works [1, 57] utilize fast 1D sensors for high-frequency acquisition, but they can only measure vibrations along one dimension. Sheinin et al. [43] proposed a dual-shutter system with high-frequency sensing capabilities, integrated with active speckle-based techniques that optically magnify small-amplitude vibrations through laser illumination.", + "bbox": [ + 89, + 90, + 483, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-based vibration measurement. Recent studies have investigated the application of event cameras for vibration measurement [10, 29, 33]. Dorn et al. [10] adapted the phase-based subtle motion extraction method to event data, assuming that each event contributes an equal phase change. Subsequently, they developed an event-based structural monitoring method. Shi et al. [44] incorporated laser active illumination to enhance event-based vibration measurements and introduced a frequency estimation technique based on Gaussian mixture distributions. Niwa et al. [34] presented a non-contact audio recovery setup using an event camera, employing the phase-based approach proposed in [10]. Howard et al. [22] implemented a similar setup, but recovered audio signals by analyzing the zero-crossings of pixels. Chen et al. [7] developed an event-based motion magnification method for amplifying and visualizing subtle motions. However, this approach does not directly estimate subtle motions; instead, motion information is implicitly derived from event signals via neural network processing. In this work, we focus on passive capture and estimation of subtle vibrations through the synergistic use of event signals and conventional frames.", + "bbox": [ + 91, + 308, + 486, + 641 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 655, + 181, + 671 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 680, + 294, + 695 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vibration modal analysis. The vibrations of objects are often well-approximated by linear systems. In modal analysis, a solid object is modeled as a system of point masses interconnected by springs [42]. In this discretized model, the mass matrix $\\mathbf{M}$ and the stiffness matrix $\\mathbf{K}$ describe mass concentrated and stiffness between each pair of degrees of freedom (DOF), respectively. The equation of motion for this system is given by:", + "bbox": [ + 89, + 707, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {M} \\ddot {x} + \\mathbf {K} x = 0, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 230, + 842, + 482, + 857 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $x$ and $\\ddot{x}$ denote the displacement and acceleration of each DOF. The object's vibrations exhibit resonant frequen", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "cies, which correspond to the eigenvalues of the system:", + "bbox": [ + 511, + 90, + 883, + 106 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {K} \\mathbf {u} _ {i} = \\omega_ {i} ^ {2} \\mathbf {M} \\mathbf {u} _ {i}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 117, + 906, + 135 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\omega_{i}$ represents a resonant frequency and the corresponding eigenvector $\\mathbf{u}_i$ describes the mode shape. To make the theory of vibration accessible to the computer vision audience, we highlight the following key concepts:", + "bbox": [ + 511, + 147, + 903, + 208 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Mode Shapes $(\\mathbf{u}_1\\dots \\mathbf{u}_n)$ : Each mode shape represents a distinct pattern of vibration for the object. The set of mode shapes forms an orthogonal basis for the vibration.", + "- Resonant Frequencies $(\\omega_{1}\\dots \\omega_{n})$ : Each mode is associated with a resonant frequency. Resonant frequencies are spatially invariant across the object's surface." + ], + "bbox": [ + 511, + 210, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The global motion power spectrum for an object can be computed by averaging the power spectra of local motions extracted at each pixel location. The modes of vibrations manifest as peaks in the motion power spectrum (Fig. 1(d)). Estimating mechanical properties from vibrations is based on the insight that variations in mechanical properties induce changes in the resonant frequencies and mode shapes for a fixed geometry.", + "bbox": [ + 511, + 304, + 905, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event formation model. When the logarithmic change of brightness at pixel $\\mathbf{x}_k$ and time $t$ exceeds a specific threshold $C$ , an event signal $e_k = (\\mathbf{x}_k, t_k, p_k)$ will be triggered:", + "bbox": [ + 511, + 430, + 905, + 491 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left| \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) \\right| \\geq C, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 503, + 903, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $p_k \\in \\{1, -1\\}$ is the polarity of the brightness change, $\\mathbf{L}(\\mathbf{x}_k, t_k)$ denotes the logarithmic brightness at pixel $\\mathbf{x}_k$ and time $t_k$ , and $t_k - \\Delta t_k$ is the time at which the preceding event was triggered at pixel $\\mathbf{x}_k$ . Let $\\mathcal{E} = \\{e_k\\}_{k=1}^{N_e}$ denote a set of $N_e$ events triggered within a short time interval, the brightness change can be computed by summing their polarities:", + "bbox": [ + 511, + 531, + 905, + 638 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\mathbf {L} (\\mathbf {x}) = \\sum_ {k = 1} ^ {N _ {e}} p _ {k} C \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 648, + 903, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the Dirac delta function $\\delta$ selects the pixel $\\mathbf{x}_k$ . Assuming brightness constancy over the interval, the brightness change can be approximated as the product of the brightness gradient $\\nabla \\mathbf{L}$ and the displacement $\\Delta \\mathbf{x}$ , where the displacement is a result of motion with velocity $\\mathbf{v}$ on the image plane [18]:", + "bbox": [ + 511, + 700, + 905, + 792 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\Delta \\widehat {\\mathbf {L}} (\\mathbf {x}) = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\Delta \\mathbf {x} = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\mathbf {v} (\\mathbf {x}) \\Delta t. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 539, + 801, + 903, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Coarse Motion by Optimization", + "text_level": 1, + "bbox": [ + 511, + 832, + 794, + 849 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the sparsity of event signals in vibration scenarios, event alignment and correlation techniques [19, 37, 47], commonly used in previous methods, are not applicable.", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "24668", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/2621cadf4486f7003d5eb57b8e0f68b69baf85cf9464a2a83942c25bb944caf8.jpg", + "image_caption": [ + "(a) Events $(\\Delta \\mathbf{L})$" + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 222, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/88d3d31806afc2055e0fc1c53ec5e903ec718880e718561981bb9663e14ff6f3.jpg", + "image_caption": [ + "(b) Coarse Motion" + ], + "image_footnote": [], + "bbox": [ + 223, + 88, + 349, + 186 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/461fa197b596e8905fdf7ded099d0160339c92f3ce1db7c57828a0f7f3c5b000.jpg", + "image_caption": [ + "(c) Predicted $(\\Delta \\hat{\\mathbf{L}})$", + "Figure 2. An example (corresponding to the jelly cube in Fig. 1(b)) of coarse motion through optimization. The predicted brightness change (c), denoted as $\\Delta \\widehat{\\mathbf{L}}$ in Eq. (5) aligns well with the event frame (a), represented by $\\Delta \\mathbf{L}$ in Eq. (4)." + ], + "image_footnote": [], + "bbox": [ + 352, + 88, + 478, + 186 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we employ a hybrid Event-RGB camera system, integrating high-speed event signals with spatially dense texture information from images, to facilitate subtle motion estimation. Our method takes event streams and a single reference frame as inputs.", + "bbox": [ + 89, + 295, + 482, + 369 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Leveraging the event formation model, we formulate the problem of event-based subtle motion estimation as an optimization task, where we minimize the mismatch between the brightness change $\\Delta \\mathbf{L}$ , derived from event accumulation (as shown in Eq. (4)), and its prediction $\\Delta \\widehat{\\mathbf{L}}$ , obtained from moving brightness edges (as detailed in Eq. (5)).", + "bbox": [ + 89, + 371, + 483, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization objective. The composite loss function is a combination of data-fidelity term $E_{\\mathrm{data}}$ and regularization term $E_{\\mathrm{reg}}$ :", + "bbox": [ + 89, + 467, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE (\\mathbf {v}) = E _ {\\text {d a t a}} (\\mathbf {v}; \\mathcal {E}) + \\lambda E _ {\\text {r e g}} (\\nabla \\mathbf {v}), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 165, + 523, + 482, + 542 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda$ serves as a balancing parameter, empirically set to 0.5 in our experiments.", + "bbox": [ + 89, + 553, + 482, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data-fidelity term $E_{\\mathrm{data}}$ measures the difference between the observed brightness change $\\Delta \\mathbf{L}$ derived from event accumulation and its prediction $\\Delta \\widehat{\\mathbf{L}}$ based on our coarse motion estimation $\\mathbf{v}$ :", + "bbox": [ + 89, + 584, + 483, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\mathrm {d a t a}} (\\mathbf {v}; \\mathcal {E}) = \\left\\| \\frac {\\Delta \\widehat {\\mathbf {L}}}{\\| \\Delta \\widehat {\\mathbf {L}} \\| _ {2}} \\cdot \\mathbf {M} - \\frac {\\Delta \\mathbf {L}}{\\| \\Delta \\mathbf {L} \\| _ {2}} \\right\\| _ {1}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 652, + 482, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{M}$ is a mask selecting pixels with events triggered. Due to spatial-temporal variations in the contrast threshold $C$ [23], we compute the difference between normalized brightness changes across the image plane to enhance the robustness of the optimization, following established methodologies [21, 46]. To analyze high-frequency vibrations, the subtle motion is estimated within a very short time interval (typically less than $0.5~ms$ ), during which very few event signals are triggered. It is challenging to accurately estimate flow in regions devoid of events, which often correspond to areas with low contrast or zero flow. To mitigate these inaccuracies, we only calculate the data-fidelity loss term at pixel locations with events triggered.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We assume that flow vectors vary smoothly with sparse discontinuities occurring at object boundaries. The regularization term $E_{\\mathrm{reg}}$ encourages smoothness of the estimated motion $\\mathbf{v}$ by minimizing the differences between neighboring pixels. This term is defined as:", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nE _ {\\text {r e g}} (\\nabla \\mathbf {v}) = \\| \\omega (\\mathbf {x}) \\nabla \\mathbf {v} (\\mathbf {x}) \\| _ {1}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 176, + 903, + 194 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\omega$ represents the smoothness weight for each pixel. Following prior research [36], the weight $\\omega$ is computed based on the gradients of the reference image:", + "bbox": [ + 511, + 203, + 903, + 248 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\omega (\\mathbf {x}) = \\exp (- \\mu | \\nabla \\mathbf {L} (\\mathbf {x}) |), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 616, + 258, + 903, + 276 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\nabla \\mathbf{L}(\\mathbf{x})$ denotes the spatial gradient of the logarithmic brightness at pixel $\\mathbf{x}$ .", + "bbox": [ + 511, + 284, + 903, + 315 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Optimization details. To enhance the convergence speed and robustness of the optimization process, we employ a patch-based optimization strategy implemented in a coarse-to-fine manner [47]. This strategy involves three resolution scales in a pyramid structure, with the coarsest patch size at $32 \\times 32$ pixels and the finest at $8 \\times 8$ pixels. Bilinear interpolation is utilized for upsampling between successive scales. $\\mathbf{v}$ is initialized uniformly within the range $[-0.1, 0.1]$ at the coarsest scale. For the subsequent scale, the initialization is derived from the optimization results of the preceding scale. We utilize the Adam optimizer [24] with a learning rate of 0.05, and the overall number of iterations is set to 40.", + "bbox": [ + 511, + 321, + 906, + 516 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Analysis. An example of coarse motion estimation is presented in Fig. 2. From this example, it is observed that the predicted brightness change $\\Delta \\mathbf{L}$ , derived from the coarse motion (Fig. 2(b)), aligns well with the event frame (Fig. 2(a)), which corresponds to $\\Delta \\mathbf{L}$ in Eq. (4). This demonstrates the efficacy of the coarse motion estimation. However, the coarse motion results manifest several limitations. First, due to the sparsity of events, the estimated coarse motion yields valid results only in regions where events are triggered, and it is susceptible to noise. Second, quantization errors in event cameras hinder the accurate recovery of motion amplitude in the coarse motion results.", + "bbox": [ + 511, + 522, + 905, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Motion Refinement by Learning", + "text_level": 1, + "bbox": [ + 511, + 712, + 797, + 729 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "With the optimization-based approach described in Sec. 3.2, we obtain the coarse motion results, denoted as $f^c$ for ease of presentation. While $f^c$ enables the brightness change prediction $\\Delta \\hat{\\mathbf{L}}$ to align with the event signals, it exhibits issues including excessive smoothness, susceptibility to noise, and inaccurate motion amplitude estimation. To address these problems, we propose EVibNet, a motion refinement network specifically designed to enhance the coarse motion results.", + "bbox": [ + 511, + 734, + 905, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The pipeline of EVibNet is shown in Fig. 3. The network takes event signals $\\mathcal{E}$ , a single reference frame $\\mathbf{I}_{\\mathrm{ref}}$", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "24669", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/9e114cc8f5467df240c7e27df4a23d3ac749f6788727b5ced6e05ee1fe492273.jpg", + "image_caption": [ + "Figure 3. Pipeline of our EVibNet. (Input: $\\mathbf{I}_{\\mathrm{ref}}$ and $\\varepsilon_{i}$ ) Initially, features are extracted from the reference frame $\\mathbf{I}_{\\mathrm{ref}}$ and events triggered within the time interval $[t_i,t_{i + 1}]$ using $\\mathcal{F}_{\\mathrm{L}}$ and $\\mathcal{F}_{\\mathrm{E}}$ , respectively. Then the extracted features are fused with a fusion module $\\mathcal{F}_{\\mathrm{fuse}}$ . The coarse motion obtained in Sec. 3.2 is fed into a motion encoder $\\mathcal{F}_{\\mathrm{M}}$ , generating the motion feature $\\mathbf{f}_{t_i,t_{i + 1}}^c$ , which is further refined by the motion refine module $\\mathcal{F}_{\\mathrm{refine}}$ . Finally, the refined motion feature is accumulated to obtain the vibration displacement (Outout: $\\mathbf{f}_{t_0,t_i + 1}$ )." + ], + "image_footnote": [], + "bbox": [ + 98, + 93, + 898, + 281 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and coarse motion results $f^c$ as inputs. Specifically, EVibNet comprises two principal components: 1) two modality-specific encoders $\\mathcal{F}_{\\mathrm{L}}$ and $\\mathcal{F}_{\\mathrm{E}}$ and a fusion module $\\mathcal{F}_{\\mathrm{fuse}}$ ; 2) a coarse motion refinement module $\\mathcal{F}_{\\mathrm{refine}}$ and a motion accumulation module $\\mathcal{F}_{\\mathrm{acc}}$ .", + "bbox": [ + 89, + 378, + 482, + 455 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature extraction. For events $\\mathcal{E}_i$ triggered in the time interval $[t_i, t_{i+1}]$ , we transform $\\mathcal{E}_i$ into two distinct voxel grids $\\mathbf{E}_i^+$ and $\\mathbf{E}_i^-$ for positive and negative events respectively, following the procedure in [17]. Due to the modality gap between frames and events, we use two modality-specific encoders $\\mathcal{F}_{\\mathrm{L}}$ and $\\mathcal{F}_{\\mathrm{E}}$ to extract features from the reference frame $\\mathbf{I}_{\\mathrm{ref}}$ and events $\\mathcal{E}_i$ , respectively. To solve the issue of inaccurate motion amplitude estimation caused by quantization error, we leverage the temporal structure inherent in the event data using a recurrent convolutional encoder. Specifically, a ConvLSTM [45] layer is integrated into each encoder stage of $\\mathcal{F}_{\\mathrm{E}}$ . The feature extraction process for the reference frame and event stack is performed using the feature encoders as follows:", + "bbox": [ + 89, + 459, + 483, + 671 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} ^ {\\mathrm {L}} = \\mathcal {F} _ {\\mathrm {L}} \\left(\\mathbf {I} _ {\\text {r e f}}\\right), \\quad \\mathbf {f} _ {i} ^ {\\mathrm {E}} = \\mathcal {F} _ {\\mathrm {E}} \\left(\\mathbf {E} _ {i} ^ {+}, \\mathbf {E} _ {i} ^ {-}, h _ {i - 1}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 125, + 678, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}^{\\mathrm{L}},\\mathbf{f}_i^{\\mathrm{E}}\\in \\mathbb{R}^{h\\times w\\times D}$ , the feature resolution $(h,w)$ is $1 / 4$ of the original image resolution $(H,W)$ , and $h_{i - 1}$ refers to the hidden states in the recurrent event encoder.", + "bbox": [ + 89, + 702, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given that event signals are triggered in only a few regions when capturing vibrating objects, it is crucial to leverage texture features extracted from the reference frame to regulate the motion features extracted from the events. Following [51], we employ a cross-modal attention fusion block, denoted as $\\mathcal{F}_{\\mathrm{fuse}}$ , to adaptively fuse the information from these two modalities, yielding the fused feature $\\mathbf{f}_i^{\\mathrm{H}}$ . The cross-modal attention in $\\mathcal{F}_{\\mathrm{fuse}}$ takes as input the queries $\\mathbf{Q}^{\\mathrm{E}}$ derived from the event features $\\mathbf{f}_i^{\\mathrm{E}}$ , and the keys $\\mathbf{K}^{\\mathrm{L}}$ and values $\\mathbf{V}^{\\mathrm{L}}$ derived from the image features $\\mathbf{f}^{\\mathrm{L}}$ .", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Motion refinement and accumulation. With the frame-event fused feature $\\mathbf{f}_i^{\\mathrm{H}}$ , we refine the coarse motion $f_{t_i,t_{i + 1}}^c$ through a motion refinement module, denoted as $\\mathcal{F}_{\\mathrm{refine}}$ . The coarse motion $f_{t_i,t_{i + 1}}^c$ is initially encoded into coarse motion features $\\mathbf{f}_{t_i,t_{i + 1}}^c$ using a motion encoder $\\mathcal{F}_{\\mathrm{M}}$ . Subsequently, the motion information embedded in $\\mathbf{f}_i^{\\mathrm{H}}$ is leveraged to refine $\\mathbf{f}_{t_i,t_{i + 1}}^c$ , producing the refined motion feature $\\mathbf{f}_{t_i,t_{i + 1}}^r$ :", + "bbox": [ + 511, + 378, + 906, + 503 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r} = \\mathcal {F} _ {\\text {r e f i n e}} \\left(\\mathbf {f} _ {i} ^ {\\mathrm {H}}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {c}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 503, + 903, + 522 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Following the refinement of coarse motion, we further accumulate short-term motion into long-range displacement. This approach is motivated by two factors: First, motion within a very short time interval is typically minimal and thus more susceptible to noise. Second, estimating the displacement of points on vibrating objects relative to a reference state facilitates subsequent analysis. However, direct accumulation of motion at each time step results in significant accumulation errors. Therefore, we integrate short-term motion using a motion accumulation module $\\mathcal{F}_{\\mathrm{acc}}$ . The accumulated motion feature $\\mathbf{f}_{t_0,t_i}$ is computed as follows:", + "bbox": [ + 511, + 525, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {t _ {0}, t _ {i + 1}} ^ {a c c} = \\mathcal {F} _ {\\text {a c c}} \\left(\\mathbf {f} _ {t _ {0}, t _ {i}} ^ {a c c}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r}, \\mathbf {f} ^ {\\mathrm {L}}\\right), \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 705, + 903, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{f}_{t_0,t_i}^{acc}$ is derived from the preceding time interval, and $\\mathbf{f}^{\\mathrm{L}}$ serves as the context feature to guide the motion accumulation. For simplicity, $\\mathcal{F}_{\\mathrm{acc}}$ is implemented as a simple RNN comprising two ConvNext blocks [27]. Finally, the accumulated motion feature is decoded into the target displacement $f_{t_0,t_{i + 1}}$ using a motion decoder $\\mathcal{F}_{\\mathrm{D}}$ .", + "bbox": [ + 511, + 727, + 906, + 820 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 827, + 732, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss functions. Our loss functions are derived from prior work on optical flow estimation [52]. The network is supervised using the $l_{1}$ distance between the predicted and", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "24670", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ground truth flow across the entire sequence of predictions, $\\{f_{t_0,t_1},\\dots ,f_{t_0,t_N}\\}$ . Given the ground truth flow sequence $\\{f_{t_0,t_1}^{gt},\\dots ,f_{t_0,t_N}^{gt}\\}$ , the loss is defined as:", + "bbox": [ + 89, + 90, + 480, + 138 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {t _ {0}, t _ {i}} ^ {g t} - f _ {t _ {0}, t _ {i}} \\| _ {1}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 147, + 480, + 188 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic dataset. Obtaining ground truth data for real-world subtle motion is challenging. Therefore, following previous studies [7, 35], we simulate subtle motion by composing foreground elements onto a background. We sample images from the MS COCO dataset [26] for the background and use segmented objects from the PASCAL VOC dataset [12] for the foreground. The magnitude and direction of motion for both the background and each foreground object are generated randomly. To simulate sub-pixel motions, we initially generate high-resolution videos, where the motion appears larger, and subsequently downsample each frame to the desired resolution. After obtaining the videos, we use the event simulator V2E [23] to simulate event data from downsampled videos.", + "bbox": [ + 89, + 200, + 483, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Training details. We implement our approach using the PyTorch framework and conduct all experiments on a single NVIDIA GeForce RTX 3090 GPU. During training, we use the AdamW optimizer [28] and configure the batch size to 4. The EVibNet is trained together for 20 epochs. The initial learning rate is set to $2 \\times 10^{-4}$ , and a cosine annealing learning rate scheduling strategy is employed.", + "bbox": [ + 89, + 419, + 483, + 525 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Hybrid camera system. Our hybrid camera system consists of a machine vision camera (HIKVISION MV-CA050-12UC) and an event camera (PROPHESEE GEN4.1), which are co-aligned using a beam splitter (Thorlabs CCM1-BS013). We utilize the calibration method proposed in [32] to achieve pixel alignment.", + "bbox": [ + 89, + 530, + 483, + 621 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 633, + 223, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we first evaluate the sub-pixel motion estimation accuracy of the proposed method on the synthetic dataset in Sec. 4.1. Then we validate our method's ability to detect high-frequency vibrations by capturing and replaying the vibration caused by audio sources in Sec. 4.2. Finally, we demonstrate our method by estimating both homogeneous and heterogeneous material properties of objects with known geometry in Sec. 4.3.", + "bbox": [ + 89, + 657, + 482, + 779 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation on Synthetic Dataset", + "text_level": 1, + "bbox": [ + 89, + 787, + 372, + 803 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method to the phase-based subtle motion estimation approach (denoted as RGBPhase) that is widely used in frame-based visual vibrometry. This approach computes local motion signals from phase shifts within a complex steerable pyramid (CSP) [39, 49, 50]. For the comparison with high-speed videos, we apply RGBPhase to the", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/d711477ebe5b7af9989866481aba75177bec1ce996a850c085f4f78bc9e38164.jpg", + "table_caption": [ + "Table 1. Motion estimation comparison on the synthetic dataset. $\\mathbf{E}$ and $\\mathbf{F}$ denote events and frames, respectively. $\\downarrow$ indicates the lower, the better throughout this paper. The best performances are highlighted in bold." + ], + "table_footnote": [], + "table_body": "
MethodsInputsEPE ↓% Out ↓
EV-FlowNet [37]E0.180547.95
MCM [47]E0.451166.87
RGBPhase [15]F (960Hz)0.283354.78
EMM [7]+RAFT [52]E + F (30Hz)0.219047.63
OursE + F (single)0.083421.29
", + "bbox": [ + 517, + 155, + 903, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "original video sequences used to simulate event streams in our synthetic dataset. Furthermore, we compare our method with two representative event-based optical flow estimation methods: Multi-reference Contrast Maximization (MCM) [47] and the recurrent version of EV-FlowNet [37, 58]. To validate the superior suitability of our method for subtle motion estimation compared to the event-based motion magnification method (EMM) [7], we retrain EMM using our synthetic dataset and leverage an off-the-shelf optical flow estimation method RAFT [52] to estimate optical flow from the motion-magnified videos generated by EMM.", + "bbox": [ + 511, + 280, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the average endpoint error (EPE) and the percentage of pixels with EPE $>0.1$ px (denoted as $\\%$ Out\") as evaluation metrics. The quantitative comparison results are presented in Tab. 1. The results demonstrate that our method substantially outperforms previous event-based optical flow estimation techniques, primarily due to its specific design for subtle motion estimation. Notably, RGBPhase exhibits unsatisfactory performance on our synthetic dataset, likely attributable to the noise in local phase information encoded in the CSP. The inferior performance of EMM + RAFT compared to our method suggests that subtle motion estimation cannot be effectively achieved by simply cascading an optical flow method with EMM. This performance gap may be attributed to the blurring artifacts often introduced by motion magnification techniques.", + "bbox": [ + 511, + 446, + 908, + 674 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Capturing audio signals", + "text_level": 1, + "bbox": [ + 511, + 681, + 733, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the efficacy of our method in real-world high-frequency vibration sensing, we apply it to remote acquisition of audio signals. In the experimental setup depicted in Fig. 4, we capture the membrane vibrations of a speaker playing a linear up-chirp signal ranging from 0 to $1000\\mathrm{Hz}$ . Concurrently, we obtain reference audio recordings using a microphone. With the subtle motion extracted by our method, the 1D audio signal is reconstructed by averaging the local motions across all pixel locations and orientations [8]. We compare our method with an event-based sound recovery method (EBVM) proposed by Niwa et al. [34]. By integrating events and frames, our method estimates subtle motion more accurately than the phase-based ap", + "bbox": [ + 511, + 703, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "24671", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e845338d9e6893f378e543826ed912fa86d609c191fbf170ff91bac7e5e308c9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 101, + 90, + 277, + 179 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b8628e31ac83d5ef426f719964cf20594135941fef79d2e245e54ae6f00fc430.jpg", + "image_caption": [ + "(b) Microphone" + ], + "image_footnote": [], + "bbox": [ + 279, + 90, + 475, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a2eec5cb4e11dc7073874c3907988bc455360449ae602c159123c11e34db945.jpg", + "image_caption": [ + "(a) Experimental setup", + "(c) EBVM", + "Figure 4. Speaker membrane experiments. (a) A speaker is playing an up-chirp signal ranging from 0 to $1000\\mathrm{Hz}$ . (b) The spectrogram of the source sound recorded by a microphone. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + ], + "image_footnote": [], + "bbox": [ + 94, + 193, + 284, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c3459ef65fb27c831e4aca8b388935a37e265b498fd74fc4bb332174e94c132e.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 287, + 193, + 477, + 285 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/11f6541417c2378b7fad369a65ed9fcc06ea148be5ae96aeae357be132974403.jpg", + "table_caption": [ + "Table 2. Percent error in estimating the Young's Modulus for each rod. The best performances are highlighted in bold. The content in each cell refers to the results for rods clamped to a length of 35 cm and 48 cm respectively." + ], + "table_footnote": [], + "table_body": "
% Error ↓LightingAluminumSteelCopper
Video [9]low10.01/10.42-11.64/-9.28-9.40/4.64
medium4.13/-3.70-10.72/-7.62-4.41/-2.70
high-1.45/-1.99-7.94/-5.96-3.14/-2.10
Ourslow2.90/3.21-7.61/-6.58-5.65/2.28
medium-1.26/1.64-6.77/-5.08-3.39/-2.07
high0.39/-1.47-5.93/-5.08-2.25/-2.07
", + "bbox": [ + 94, + 463, + 483, + 575 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proach adopted in EBVM. The experimental results demonstrate significantly reduced noise levels and enhanced high-frequency reconstruction fidelity compared to the baseline method.", + "bbox": [ + 89, + 601, + 482, + 660 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Fig. 5, we conduct an experiment similar to Davis et al. [8]. A speaker playing the \"MaryMIDI\" audio excites the chip bag. We measure the vibration of a chip bag and recover the audio. The spectrograms of recovered audio reveal superior waveform fidelity achieved by our method. These results demonstrate our method's robustness and practical applicability in complex real-world vibration sensing scenarios.", + "bbox": [ + 89, + 662, + 483, + 784 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Estimating material properties", + "text_level": 1, + "bbox": [ + 89, + 795, + 362, + 811 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We apply our method to estimate homogeneous and heterogeneous material properties from vibrations.", + "bbox": [ + 89, + 818, + 482, + 849 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Homogeneous material properties. Considering objects with homogeneous material properties, it can be obtained from Eq. (2) that different objects with the same ge", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/bcb04966c96a449a84c6d96aa65c5705a971bddbd41285994eedcf2c503205cc.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 90, + 702, + 178 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7cfa53b66d52518adf7fcebcae6fa5384118495061dae17efd92ba87b3daff2e.jpg", + "image_caption": [ + "(a) Experimental setup", + "(c) EBVM", + "Figure 5. Audio recovery from a bag of chips. (a) Sound from an audio source (speaker) excites the chip bag. (b) The spectrogram of the input signal sent to the speaker. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + ], + "image_footnote": [], + "bbox": [ + 517, + 193, + 707, + 285 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e2507dae63c5e301867152e273afcae7c8a02f506a4bc6892ffd30b94d2502de.jpg", + "image_caption": [ + "(b) Input signal" + ], + "image_footnote": [], + "bbox": [ + 702, + 90, + 901, + 180 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5e7e702c9b1a8a7654f8bf35644ecd7cbda2e4df2799a7056b3b99398185c1aa.jpg", + "image_caption": [ + "(d) Ours" + ], + "image_footnote": [], + "bbox": [ + 709, + 193, + 901, + 282 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ometry have an identical set of mode shapes, but their resonant frequencies scale proportionally to material properties [9]. Following Davis et al. [9], we apply our method to estimate the homogeneous material properties of various metal rods. The simple geometry of clamped rods makes their vibration well-studied [42]. The fundamental frequency $\\omega_{1}$ (first resonant frequency) of a rod is given by:", + "bbox": [ + 511, + 404, + 905, + 511 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\omega_ {1} = 0. 1 3 9 9 \\frac {d}{L ^ {2}} \\sqrt {\\frac {E}{\\rho}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 518, + 905, + 560 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $d$ is the diameter of the rod, $L$ is the length, $E$ is the Young's modulus, and $\\rho$ is the density. While length, diameter, and density can be easily measured, Young's modulus is often measured with a tensile test, which will damage the object. In this experiment, we try to estimate the Young's modulus of the rods by finding their fundamental frequency, which is the highest peak in the motion spectrum.", + "bbox": [ + 511, + 568, + 905, + 672 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We test rods with three different metals: aluminum, steel, and copper. Each rod is tested at two lengths (35 cm and 48 cm) and under three brightness conditions (100, 500, and 1500 lux). For comparison with frame-based visual vibrometry, we capture 1000 fps videos under identical settings. Resonant frequencies are global properties and invariant to the viewpoint, ensuring that estimation accuracy is not affected by the viewpoint. In Tab. 2, we compare the estimated Young's modulus values with those provided by the manufacturer. The results demonstrate that our method achieves more robust performance across varying lighting conditions. We further compare the data sizes of 1000 fps videos with signals captured by our imaging system in the rod experiments. Our method requires only a single image; consequently, the data size of the reference frame is negligible.", + "bbox": [ + 511, + 674, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "24672", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fae9ad03b961d72790633bf248f0322b93ee8d6acbea9d61fd33f58489bf1300.jpg", + "image_caption": [ + "Figure 6. Heterogenous material properties estimation result of the defect cube (a) and a homogeneous cube (b), the results are plotted with the same colormaps. Based on the results, we can easily differentiate between a cube with a defect and a homogeneous one." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 478, + 239 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "gible. At the same 720p resolution, event cameras produce event streams at $1 - 5\\mathrm{MB / s}$ , about $20\\%$ of the compressed 1000 fps videos (H. 264 encoded). The discrepancy is more pronounced when considering the actual transmission bandwidth required for high-speed cameras. We also test existing event-based optical flow estimation methods. However, both EV-FlowNet [37] and MCM [47] struggle to accurately estimate the vibrations of the rods, and we can not identify reasonable resonant frequencies in the motion spectra.", + "bbox": [ + 89, + 351, + 483, + 488 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Heterogeneous material properties. When considering objects with heterogeneous material properties, spatial inhomogeneities in these properties influence both the resonant frequencies and mode shapes. Based on the mathematical relationship between modes and material properties, Feng et al. [13] propose a physical-constrained optimization approach to estimate heterogeneous material properties from several image-space vibration modes.", + "bbox": [ + 89, + 500, + 483, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To demonstrate the applicability of our method to visual vibration tomography, we experiment on two jelly cubes: one homogeneous and the other with an interior defect (a rubber doll embedded within the jelly cube). We capture data under an initial deformation condition and identify unique modes in the estimated motion fields, as shown in Fig. 1. For each cube, we infer spatially-varying Young's modulus and density values on a $10 \\times 10 \\times 10$ hexahedral mesh with the optimization algorithm proposed by Feng et al. [13]. The reconstruction results are shown in Fig. 6. The reconstruction is obtained using only five unique image-space modes. The defective cube has a rubber doll on top, which is stiffer than the jelly cube. The results Fig. 6(a) show the existence of the defect, and clearly differ from that of the homogeneous cube. Due to the high damping of the jelly cubes, the reconstruction results are not perfectly aligned with the data, but we can still easily differentiate between the two cubes.", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/5f05e698c4d39da279f29e2e59146dc332cc479b2cc61388678a5016a50a9da6.jpg", + "table_caption": [ + "Table 3. Ablation study of the method design on the synthetic data." + ], + "table_footnote": [], + "table_body": "
CoarseW/o cmW/o accW/o frameOurs
EPE ↓0.36320.09690.12990.17340.0834
% Out ↓51.4525.9136.9942.2521.29
", + "bbox": [ + 517, + 114, + 906, + 170 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Ablation studies", + "text_level": 1, + "bbox": [ + 513, + 198, + 674, + 212 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness and necessity of each part of our method, we conduct several ablation studies and show the results in Tab. 3. To assess the necessity of motion refinement using EVibNet, we test the performance of coarse motion (denoted as \"coarse\"). To verify the contribution of the coarse motion obtained from optimization, we remove it from the inputs of EVibNet (denoted as \"w/o cm\"). To demonstrate the effectiveness of the motion accumulation module, we directly accumulate motions at each time step (denoted as \"w/o acc\"). Finally, we show the necessity of the reference frame by removing it from inputs (denoted as \"w/o frame\"). As indicated in Tab. 3, our complete method achieves the lowest EPE and % Out, which demonstrates the contribution of each component of our method.", + "bbox": [ + 511, + 222, + 906, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 454, + 633, + 469 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We present a bandwidth-efficient approach for high-speed vibration sensing with an Event-RGB hybrid camera system. To extract subtle motion frame event data, we integrate an optimization-based approach with a learning-based motion refinement network. We demonstrate our method through the capture of vibrations induced by audio sources and the estimation of material properties.", + "bbox": [ + 511, + 482, + 905, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. In the synthetic dataset, subtle motion is simulated by compositing foreground elements onto a background, which exhibits a gap from real-world vibration scenarios. In the future, simulating more realistic synthetic datasets using the finite element method (FEM) will further improve the performance of event-based visual vibrometry.", + "bbox": [ + 511, + 595, + 905, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 513, + 707, + 671, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62402014, 62276007), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). Peiqi Duan was also supported by China National Postdoctoral Program for Innovative Talents (Grant No. BX20230010) and China Postdoctoral Science Foundation (Grant No. 2023M740076). The authors thank openbayes.com for providing computing resource.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "24673", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Silvio Bianchi and Emanuele Giacomozzi. Long-range detection of acoustic vibrations by speckle tracking. Applied optics, 58(28):7805-7809, 2019. 3", + "[2] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proc. of IEEE International Conference on Computer Vision, 2013. 1", + "[3] Oral Büyüközürk, Mehmet Ali Taşdemir, O Büyüközürk, R Haupt, C Tuakta, and J Chen. Remote detection of debonding in frp-strengthened concrete structures using acoustic-laser technique. In Proc. of Nondestructive Testing of Materials and Structures, 2013. 2", + "[4] Oral Buyukozturk, Justin G Chen, Neal Wadhwa, Abe Davis, Frédo Durand, and William T Freeman. Smaller than the eye can see: Vibration analysis with video cameras. In Proc. of World Conference on Non-Destructive Testing, 2016. 1", + "[5] Justin G Chen, Robert W Haupt, and Oral Buyukozturk. The acoustic-laser vibrometry technique for the noncontact detection of discontinuities in fiber reinforced polymer-retrofitted concrete. Materials evaluation, 72(10), 2014. 2", + "[6] Justin G Chen, Neal Wadhwa, Young-Jin Cha, Fredo Durand, William T Freeman, and Oral Buyukozturk. Modal identification of simple structures with high-speed video using motion magnification. Journal of Sound and Vibration, 345:58-71, 2015. 2", + "[7] Yutian Chen, Shi Guo, Fangzheng Yu, Feng Zhang, Jinwei Gu, and Tianfan Xue. Event-based motion magnification. In Proc. of European Conference on Computer Vision, 2024. 2, 3, 6", + "[8] Abe Davis, Michael Rubinstein, Neal Wadhwa, Gautham J. Mysore, Frédo Durand, and William T. Freeman. The visual microphone: passive recovery of sound from video. ACM Transactions on Graphics, 33(4):79:1-79:10, 2014. 1, 3, 6, 7", + "[9] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Oral Buyukozturk, Frédo Durand, and William T Freeman. Visual Vibrometry: Estimating material properties from small motions in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):732-745, 2017. 1, 3, 7", + "[10] Charles Dorn, Sudeep Dasari, Yongchao Yang, Garrett Kenyon, Paul Welch, and David Mascareñas. Efficient fullfield operational modal analysis using neuromorphic event-based imaging. In Proc. of Conference and Exposition on Structural Dynamics, 2017. 3", + "[11] Franz Durst, Adrian Melling, and James H Whitelaw. Principles and practice of laser-doppler anemometry. NASA STI/Recon Technical Report A, 76:47019, 1976. 1, 2", + "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6", + "[13] Berthy T Feng, Alexander C Ogren, Chiara Daraio, and Katherine L Bouman. Visual vibration tomography: Estimating interior material properties from monocular video. In" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 8", + "[14] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. A $1280 \\times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\\mu \\mathrm{m}$ pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In Proc. of IEEE International Solid-State Circuits Conference, 2020. 2", + "[15] David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information. International Journal of Computer Vision, 5:77-104, 1990. 6", + "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2", + "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-End learning of representations for asynchronous event-based data. In Proc. of IEEE International Conference on Computer Vision, 2019. 5", + "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3):601-618, 2020. 3", + "[19] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4736-4746, 2024. 2, 3", + "[20] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8553-8565, 2023. 2", + "[21] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4", + "[22] Matthew Howard and Keigo Hirakawa. Event-based visual microphone. In Proc. of International Conference on Acoustics, Speech and Signal Processing, 2023. 3", + "[23] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbrück. V2E: From video frames to realistic DVS events. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 2, 4, 6", + "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of International Conference on Learning Representations, 2015. 4", + "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \\times 128$ 120 db $15~\\mu \\mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 2", + "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. of European Conference on Computer Vision, 2014. 6" + ], + "bbox": [ + 517, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "24674", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5", + "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of International Conference on Learning Representations, 2017. 6", + "[29] Yuanyuan Lv, Liang Zhou, Zhaohui Liu, and Haiyang Zhang. Structural vibration frequency monitoring based on event camera. Measurement Science and Technology, 35(8): 085007, 2024. 3", + "[30] William N MacPherson, Mark Reeves, David P Towers, Andrew J Moore, Julian DC Jones, Martin Dale, and Craig Edwards. Multipoint laser vibrometer for modal analysis. Applied optics, 46(16):3126-3132, 2007. 2", + "[31] M Martarelli, Gian Marco Revel, and C Santolini. Automated modal analysis by scanning laser vibrometry: problems and uncertainties associated with the scanning system calibration. Mechanical systems and signal processing, 15 (3):581-601, 2001. 2", + "[32] Manasi Muglikar, Mathias Gehrig, Daniel Gehrig, and Davide Scaramuzza. How to calibrate your event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 6", + "[33] Woong-jae Na, Kyung Ho Sun, Byeong Chan Jeon, Jaeyun Lee, and Yun-ho Shin. Event-based micro vibration measurement using phase correlation template matching with event filter optimization. Measurement, 215:112867, 2023. 3", + "[34] Ryogo Niwa, Tatsuki Fushimi, Kenta Yamamoto, and Yoichi Ochiai. Live demonstration: Event-based visual microphone. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023. 3, 6, 7", + "[35] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed A. Elgharib, Frédo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification. In Proc. of European Conference on Computer Vision, 2018. 6", + "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 4", + "[37] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Proc. of IEEE International Conference on Computer Vision, 2023. 2, 3, 6, 8", + "[38] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics Express, 18(10):10762-10774, 2010. 1", + "[39] Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. International Journal of Computer Vision, 40:49-70, 2000. 2, 6", + "[40] Nicolaas Bernardus Roozen, Ludovic Labelle, Monika Rychtarikova, and Christ Glorieux. Determining radiated" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "sound power of building structures by means of laser doppler vibrometry. Journal of Sound and Vibration, 346:81-99, 2015. 2", + "[41] Steve Rothberg, JR Baker, and Neil A Halliwell. Laser vibrometry: pseudo-vibrations. Journal of Sound and Vibration, 135(3):516-522, 1989. 2", + "[42] Ahmed A Shabana. Theory of vibration. Springer, 1991. 3, 7", + "[43] Mark Sheinin, Dorian Chan, Matthew O'Toole, and Srinivasa G Narasimhan. Dual-shutter optical vibration sensing. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 3", + "[44] Chenyang Shi, Ningfang Song, Boyi Wei, Yuzhen Li, Yibo Zhang, Wenzhuo Li, and Jing Jin. Event-based vibration frequency measurement with laser-assisted illumination based on mixture gaussian distribution. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 3", + "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems, 2015. 5", + "[46] Shintaro Shiba, Friedhelm Hamann, Yoshimitsu Aoki, and Guillermo Gallego. Event-based background-oriented schlieren. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2011-2026, 2024. 4", + "[47] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):7742-7759, 2024. 2, 3, 4, 6, 8", + "[48] Peter J Shull. Nondestructive evaluation: theory, techniques, and applications. CRC press, 2002. 1, 2", + "[49] Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. In Proc. of International Conference on Image Processing, 1995. 2, 6", + "[50] Eero P Simoncelli, William T Freeman, Edward H Adelson, and David J Heeger. Shiftable multiscale transforms. IEEE transactions on Information Theory, 38(2):587-607, 1992. 2, 6", + "[51] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In Proc. of European Conference on Computer Vision, 2022. 5", + "[52] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Proc. of European Conference on Computer Vision, 2020. 5, 6", + "[53] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics, 32(4):1-10, 2013. 2", + "[54] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Riesz pyramids for fast phase-based video magnification. In Proc. of International Conference on Computational Photography, 2014. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "24675", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[55] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2", + "[56] Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, and William Freeman. Eulerian video magnification for revealing subtle changes in the world. ACM Transactions on Graphics, 31(4):1-8, 2012. 1, 2", + "[57] Nan Wu and Shinichiro Haruyama. The 20k samples-persecond real time detection of acoustic vibration based on displacement estimation of one-dimensional laser speckle images. Sensors, 21(9):2938, 2021. 3", + "[58] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Proc. of Robotics: Science and Systems, 2018. 2, 6" + ], + "bbox": [ + 91, + 92, + 480, + 330 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "24676", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_model.json b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1bf7d227c6720208c878d2162922aedf03156717 --- /dev/null +++ b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_model.json @@ -0,0 +1,2576 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.336, + 0.132, + 0.658, + 0.154 + ], + "angle": 0, + "content": "Event-based Visual Vibrometry" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.173, + 0.834, + 0.192 + ], + "angle": 0, + "content": "Xinyu Zhou\\(^{1}\\) Peiqi Duan\\(^{2,3}\\) Yeliduosi Xiaokaiti\\(^{2,3}\\) Chao Xu\\(^{1}\\) Boxin Shi\\(^{2,3*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.142, + 0.194, + 0.871, + 0.21 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.211, + 0.834, + 0.227 + ], + "angle": 0, + "content": "\\(^{2}\\) State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.176, + 0.228, + 0.834, + 0.244 + ], + "angle": 0, + "content": "\\(^{3}\\)National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.176, + 0.211, + 0.834, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.248, + 0.905, + 0.263 + ], + "angle": 0, + "content": "{zhouxiny, duanqi0001, shiboxin}@pku.edu.cn, yongqiye@stu.pku.edu.cn, xuchao@cis.pku.edu.cn" + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.274, + 0.339, + 0.421 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.17, + 0.421, + 0.338, + 0.435 + ], + "angle": 0, + "content": "(a) Hybrid camera system" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.299, + 0.553, + 0.4 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.374, + 0.403, + 0.541, + 0.416 + ], + "angle": 0, + "content": "(b) Frame + Event stream" + }, + { + "type": "image", + "bbox": [ + 0.586, + 0.277, + 0.727, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.616, + 0.443, + 0.718, + 0.455 + ], + "angle": 0, + "content": "Pixel Displacement" + }, + { + "type": "image", + "bbox": [ + 0.743, + 0.277, + 0.895, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.779, + 0.443, + 0.872, + 0.455 + ], + "angle": 0, + "content": "Motion Spectrum" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.468, + 0.907, + 0.539 + ], + "angle": 0, + "content": "Figure 1. Overview of event-based visual vibrometry. A hybrid camera system, comprising a frame-based camera and an event camera, is utilized to capture subtle object vibrations (a). We propose an event-based subtle motion estimation method that leverages the event stream and a reference frame as input (b). The subtle motion field is extracted at each pixel location across time (c). This motion field can be further analyzed to extract the global motion spectrum and independent vibration modes (d). Each mode is characterized by a resonant frequency and a corresponding mode shape, which manifest as distinct peaks in the motion spectrum." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.551, + 0.327, + 0.566 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.585, + 0.486, + 0.859 + ], + "angle": 0, + "content": "Visual vibrometry has emerged as a powerful technique for remote acquisition of audio and the physical properties of materials. To capture high-frequency vibrations, frame-based approaches often require a high-speed video camera and bright lighting to compensate for the short exposure time. In this paper, we introduce event-based visual vibrometry, a new high-speed visual vibration sensing method using an event camera. By leveraging the high temporal resolution and low bandwidth characteristics of event cameras, event-based visual vibrometry enables high-speed vibration sensing under ambient lighting conditions with improved data efficiency. Specifically, we leverage a hybrid camera system and propose an event-based subtle motion estimation framework that integrates an optimization-based approach based on the event generation model and a motion refinement network. We demonstrate our method by capturing vibration caused by audio sources and estimating material properties for various objects." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.551, + 0.645, + 0.566 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.58, + 0.907, + 0.732 + ], + "angle": 0, + "content": "Vibrations are pervasive phenomena that arise from a wide range of sources, including engines, percussive impacts, and musical instruments. These vibrations typically exhibit small amplitudes and a broad spectrum of frequencies, spanning from a few Hz to MHz, and carry an immense amount of information. Vibration analysis is an established tool in various engineering and scientific fields, including the remote acquisition of audio signals [8], the monitoring of human heart rate [38, 56], and the estimation of material properties [2, 9, 13]." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.735, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Various devices have been used to measure and analyze vibrations. Among these, contact sensors [48] and laser vibrometers [11] are two widely used traditional vibration sensors, both of which are specialized instruments and only measure the vibration of a single surface point. Recently, much progress has been made on visual vibrometry [2, 4, 9, 13], an emerging technique that utilizes general-purpose video cameras for vibration measurement. In contrast to traditional vibration sensors, video-based approaches enable spatially dense measurements of surface motion. However, the acquisition of high-frequency vi" + }, + { + "type": "page_footnote", + "bbox": [ + 0.116, + 0.887, + 0.245, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24666" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "briations requires high-speed cameras coupled with bright lighting to compensate for the short exposure time, which is impractical in many real-world applications. Additionally, high-speed cameras demand substantial bandwidth requirements. To eliminate the need for high-speed cameras, Sheinin et al. [43] propose a dual-shutter system combined with active speckle-based techniques. Nevertheless, the nature of speckle-based vibration imaging restricts its applicability and confines vibration sampling to only several surface points. In short, frame-based visual vibrometry encounters challenges in efficiently capturing high-frequency vibrations while maintaining bandwidth efficiency and measurement convenience." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.291, + 0.483, + 0.502 + ], + "angle": 0, + "content": "Event cameras [14, 25], characterized by their high temporal resolution, high dynamic range, and low bandwidth requirements, have emerged as a promising solution for high-speed sensing [16]. Unlike conventional cameras, event cameras asynchronously record per-pixel logarithmic scene radiance changes. This signal-triggering mechanism is particularly advantageous for vibration sensing, where radiance changes occur only in a small fraction of pixels, leading to a significantly reduced data rate. The high temporal resolution and dynamic range inherent to event cameras enable the measurement of high-frequency vibrations under ambient lighting. These characteristics position event cameras as potential alternatives to address the constraints of frame-based visual vibrometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.483, + 0.807 + ], + "angle": 0, + "content": "Despite the potential of event cameras, estimating subtle motion from event data remains an open problem. Firstly, although event cameras have been applied to small motion scenarios for motion magnification [7], this approach focuses on visualizing small motions and adopts an implicit motion representation. However, the explicit estimation of subtle motion is essential in visual vibrometry for vibration analysis. Secondly, adapting existing event-based motion estimation techniques to vibration scenarios presents significant difficulties. Existing event-based optical flow estimation methods [19, 37, 47, 58] generally assume continuous apparent motion of visual patterns, which is violated by the oscillatory nature of vibrations. Events caused by vibrations are primarily localized along object edges, and this spatial sparsity hinders the effectiveness of event alignment techniques like Contrast Maximization [37, 47]. Moreover, in comparison to general scenarios, vibrations tend to produce significantly fewer events. Estimating subtle motion from such sparse event data is highly vulnerable to noise inherent in event cameras [23]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "In this paper, we propose event-based visual vibrometry as an efficient and practical solution for vibration measurement and analysis. To address the challenges mentioned above, we employ an Event-RGB hybrid camera system [7, 20, 55] (Fig. 1(a)), which combines event streams with spatial texture information from conventional frames. We pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.348 + ], + "angle": 0, + "content": "pose an event-based subtle motion estimation framework that takes event streams and a single reference frame as inputs (Fig. 1(b)). Initially, we employ an optimization-based approach grounded in the event generation model [21, 36] to derive coarse motion estimates within short temporal windows. The sparsity of event data restricts the validity of these coarse estimates to only event-triggered regions, while simultaneously introducing undesirable smoothing effects. Additionally, quantization errors in event cameras compromise the accuracy of motion amplitude estimation. To address these issues, we adopt a motion refinement network that aggregates historical motion information to enhance the coarse motion estimation and obtain long-range vibration displacements (Fig. 1(c)). The refined motion fields across time can subsequently be analyzed (Fig. 1(d)) for various visual vibrometry applications. The key contributions of our work are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.349, + 0.905, + 0.394 + ], + "angle": 0, + "content": "- a bandwidth-efficient visual vibrometry approach that leverages a hybrid camera system to measure high-frequency vibration under ambient lighting conditions;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.395, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- an event-based subtle motion estimation framework that integrates a model-based optimization approach and a motion refinement network;" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.44, + 0.904, + 0.485 + ], + "angle": 0, + "content": "- a thorough evaluation on both synthetic data and real-world vibrations caused by various sources, demonstrating the effectiveness of event-based visual vibrometry." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.349, + 0.905, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.501, + 0.663, + 0.516 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.905, + 0.773 + ], + "angle": 0, + "content": "Traditional vibration sensors. This section reviews two prevalent traditional vibration measurement devices: contact sensors and active sensors. Contact sensors operate through direct physical contact with the target object [48]. Accelerometers and piezoelectric pickups are representative examples of contact sensors that respond to acceleration. Among active sensors, laser Doppler vibrometers (LDVs) are the most common type used for vibration measurement [11, 41]. LDVs operate by projecting a laser beam onto the target object and calculating the surface velocity from the phase shift in the reflected light. LDVs have been used to find defects in composite materials [3, 5], examine the integrity of building structures [40], and modal analysis [30, 31]. Both contact sensors and laser vibrometers are typically limited to measuring vibrations at a single point on the surface." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Frame-based visual vibrometry. Recent studies on video motion magnification [53, 54, 56] have demonstrated the capability to detect subtle motions captured in videos. These methods typically exploit spatial phase variations of the complex steerable pyramid [39, 49, 50] to estimate small local motions. Chen et al. [6] applied this phase-based subtle motion estimation method to quantify the vibration modes of pipes and cantilever beams. Davis" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.304 + ], + "angle": 0, + "content": "et al. [8] analyzed object vibrations induced by acoustic excitation to reconstruct the original sound. Davis et al. [9] employed the motion spectra extracted from video to estimate the physical properties of objects. Feng et al. [13] extended this methodology to heterogeneous properties and proposed a physics-constrained optimization approach. Capturing high-frequency vibrations requires expensive 2D high-speed cameras. To address this cost, some works [1, 57] utilize fast 1D sensors for high-frequency acquisition, but they can only measure vibrations along one dimension. Sheinin et al. [43] proposed a dual-shutter system with high-frequency sensing capabilities, integrated with active speckle-based techniques that optically magnify small-amplitude vibrations through laser illumination." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.309, + 0.487, + 0.642 + ], + "angle": 0, + "content": "Event-based vibration measurement. Recent studies have investigated the application of event cameras for vibration measurement [10, 29, 33]. Dorn et al. [10] adapted the phase-based subtle motion extraction method to event data, assuming that each event contributes an equal phase change. Subsequently, they developed an event-based structural monitoring method. Shi et al. [44] incorporated laser active illumination to enhance event-based vibration measurements and introduced a frequency estimation technique based on Gaussian mixture distributions. Niwa et al. [34] presented a non-contact audio recovery setup using an event camera, employing the phase-based approach proposed in [10]. Howard et al. [22] implemented a similar setup, but recovered audio signals by analyzing the zero-crossings of pixels. Chen et al. [7] developed an event-based motion magnification method for amplifying and visualizing subtle motions. However, this approach does not directly estimate subtle motions; instead, motion information is implicitly derived from event signals via neural network processing. In this work, we focus on passive capture and estimation of subtle vibrations through the synergistic use of event signals and conventional frames." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.656, + 0.183, + 0.672 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.681, + 0.295, + 0.696 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.708, + 0.485, + 0.83 + ], + "angle": 0, + "content": "Vibration modal analysis. The vibrations of objects are often well-approximated by linear systems. In modal analysis, a solid object is modeled as a system of point masses interconnected by springs [42]. In this discretized model, the mass matrix \\(\\mathbf{M}\\) and the stiffness matrix \\(\\mathbf{K}\\) describe mass concentrated and stiffness between each pair of degrees of freedom (DOF), respectively. The equation of motion for this system is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.232, + 0.843, + 0.483, + 0.858 + ], + "angle": 0, + "content": "\\[\n\\mathbf {M} \\ddot {x} + \\mathbf {K} x = 0, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "where \\( x \\) and \\( \\ddot{x} \\) denote the displacement and acceleration of each DOF. The object's vibrations exhibit resonant frequen" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.885, + 0.107 + ], + "angle": 0, + "content": "cies, which correspond to the eigenvalues of the system:" + }, + { + "type": "equation", + "bbox": [ + 0.653, + 0.118, + 0.907, + 0.136 + ], + "angle": 0, + "content": "\\[\n\\mathbf {K} \\mathbf {u} _ {i} = \\omega_ {i} ^ {2} \\mathbf {M} \\mathbf {u} _ {i}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.148, + 0.905, + 0.209 + ], + "angle": 0, + "content": "where \\(\\omega_{i}\\) represents a resonant frequency and the corresponding eigenvector \\(\\mathbf{u}_i\\) describes the mode shape. To make the theory of vibration accessible to the computer vision audience, we highlight the following key concepts:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.211, + 0.907, + 0.257 + ], + "angle": 0, + "content": "- Mode Shapes \\((\\mathbf{u}_1\\dots \\mathbf{u}_n)\\): Each mode shape represents a distinct pattern of vibration for the object. The set of mode shapes forms an orthogonal basis for the vibration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.257, + 0.905, + 0.302 + ], + "angle": 0, + "content": "- Resonant Frequencies \\((\\omega_{1}\\dots \\omega_{n})\\) : Each mode is associated with a resonant frequency. Resonant frequencies are spatially invariant across the object's surface." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.211, + 0.907, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.426 + ], + "angle": 0, + "content": "The global motion power spectrum for an object can be computed by averaging the power spectra of local motions extracted at each pixel location. The modes of vibrations manifest as peaks in the motion power spectrum (Fig. 1(d)). Estimating mechanical properties from vibrations is based on the insight that variations in mechanical properties induce changes in the resonant frequencies and mode shapes for a fixed geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.431, + 0.906, + 0.492 + ], + "angle": 0, + "content": "Event formation model. When the logarithmic change of brightness at pixel \\(\\mathbf{x}_k\\) and time \\(t\\) exceeds a specific threshold \\(C\\), an event signal \\(e_k = (\\mathbf{x}_k, t_k, p_k)\\) will be triggered:" + }, + { + "type": "equation", + "bbox": [ + 0.587, + 0.504, + 0.905, + 0.521 + ], + "angle": 0, + "content": "\\[\n\\left| \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) \\right| \\geq C, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.906, + 0.639 + ], + "angle": 0, + "content": "where \\(p_k \\in \\{1, -1\\}\\) is the polarity of the brightness change, \\(\\mathbf{L}(\\mathbf{x}_k, t_k)\\) denotes the logarithmic brightness at pixel \\(\\mathbf{x}_k\\) and time \\(t_k\\), and \\(t_k - \\Delta t_k\\) is the time at which the preceding event was triggered at pixel \\(\\mathbf{x}_k\\). Let \\(\\mathcal{E} = \\{e_k\\}_{k=1}^{N_e}\\) denote a set of \\(N_e\\) events triggered within a short time interval, the brightness change can be computed by summing their polarities:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.65, + 0.905, + 0.692 + ], + "angle": 0, + "content": "\\[\n\\Delta \\mathbf {L} (\\mathbf {x}) = \\sum_ {k = 1} ^ {N _ {e}} p _ {k} C \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.794 + ], + "angle": 0, + "content": "where the Dirac delta function \\(\\delta\\) selects the pixel \\(\\mathbf{x}_k\\). Assuming brightness constancy over the interval, the brightness change can be approximated as the product of the brightness gradient \\(\\nabla \\mathbf{L}\\) and the displacement \\(\\Delta \\mathbf{x}\\), where the displacement is a result of motion with velocity \\(\\mathbf{v}\\) on the image plane [18]:" + }, + { + "type": "equation", + "bbox": [ + 0.54, + 0.803, + 0.905, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\Delta \\widehat {\\mathbf {L}} (\\mathbf {x}) = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\Delta \\mathbf {x} = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\mathbf {v} (\\mathbf {x}) \\Delta t. \\tag {5}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.795, + 0.85 + ], + "angle": 0, + "content": "3.2. Coarse Motion by Optimization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Given the sparsity of event signals in vibration scenarios, event alignment and correlation techniques [19, 37, 47], commonly used in previous methods, are not applicable." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "24668" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.223, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.12, + 0.19, + 0.201, + 0.202 + ], + "angle": 0, + "content": "(a) Events \\((\\Delta \\mathbf{L})\\)" + }, + { + "type": "image", + "bbox": [ + 0.225, + 0.089, + 0.351, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.241, + 0.19, + 0.335, + 0.201 + ], + "angle": 0, + "content": "(b) Coarse Motion" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.089, + 0.48, + 0.188 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.37, + 0.19, + 0.463, + 0.202 + ], + "angle": 0, + "content": "(c) Predicted \\((\\Delta \\hat{\\mathbf{L}})\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.269 + ], + "angle": 0, + "content": "Figure 2. An example (corresponding to the jelly cube in Fig. 1(b)) of coarse motion through optimization. The predicted brightness change (c), denoted as \\(\\Delta \\widehat{\\mathbf{L}}\\) in Eq. (5) aligns well with the event frame (a), represented by \\(\\Delta \\mathbf{L}\\) in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.296, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Therefore, we employ a hybrid Event-RGB camera system, integrating high-speed event signals with spatially dense texture information from images, to facilitate subtle motion estimation. Our method takes event streams and a single reference frame as inputs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.372, + 0.484, + 0.463 + ], + "angle": 0, + "content": "Leveraging the event formation model, we formulate the problem of event-based subtle motion estimation as an optimization task, where we minimize the mismatch between the brightness change \\(\\Delta \\mathbf{L}\\), derived from event accumulation (as shown in Eq. (4)), and its prediction \\(\\Delta \\widehat{\\mathbf{L}}\\), obtained from moving brightness edges (as detailed in Eq. (5))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.468, + 0.484, + 0.515 + ], + "angle": 0, + "content": "Optimization objective. The composite loss function is a combination of data-fidelity term \\( E_{\\mathrm{data}} \\) and regularization term \\( E_{\\mathrm{reg}} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.166, + 0.525, + 0.483, + 0.543 + ], + "angle": 0, + "content": "\\[\nE (\\mathbf {v}) = E _ {\\text {d a t a}} (\\mathbf {v}; \\mathcal {E}) + \\lambda E _ {\\text {r e g}} (\\nabla \\mathbf {v}), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.483, + 0.584 + ], + "angle": 0, + "content": "where \\(\\lambda\\) serves as a balancing parameter, empirically set to 0.5 in our experiments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.585, + 0.484, + 0.645 + ], + "angle": 0, + "content": "The data-fidelity term \\( E_{\\mathrm{data}} \\) measures the difference between the observed brightness change \\( \\Delta \\mathbf{L} \\) derived from event accumulation and its prediction \\( \\Delta \\widehat{\\mathbf{L}} \\) based on our coarse motion estimation \\( \\mathbf{v} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.653, + 0.483, + 0.695 + ], + "angle": 0, + "content": "\\[\nE _ {\\mathrm {d a t a}} (\\mathbf {v}; \\mathcal {E}) = \\left\\| \\frac {\\Delta \\widehat {\\mathbf {L}}}{\\| \\Delta \\widehat {\\mathbf {L}} \\| _ {2}} \\cdot \\mathbf {M} - \\frac {\\Delta \\mathbf {L}}{\\| \\Delta \\mathbf {L} \\| _ {2}} \\right\\| _ {1}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\mathbf{M}\\) is a mask selecting pixels with events triggered. Due to spatial-temporal variations in the contrast threshold \\(C\\) [23], we compute the difference between normalized brightness changes across the image plane to enhance the robustness of the optimization, following established methodologies [21, 46]. To analyze high-frequency vibrations, the subtle motion is estimated within a very short time interval (typically less than \\(0.5~ms\\)), during which very few event signals are triggered. It is challenging to accurately estimate flow in regions devoid of events, which often correspond to areas with low contrast or zero flow. To mitigate these inaccuracies, we only calculate the data-fidelity loss term at pixel locations with events triggered." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "We assume that flow vectors vary smoothly with sparse discontinuities occurring at object boundaries. The regularization term \\( E_{\\mathrm{reg}} \\) encourages smoothness of the estimated motion \\( \\mathbf{v} \\) by minimizing the differences between neighboring pixels. This term is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.177, + 0.905, + 0.195 + ], + "angle": 0, + "content": "\\[\nE _ {\\text {r e g}} (\\nabla \\mathbf {v}) = \\| \\omega (\\mathbf {x}) \\nabla \\mathbf {v} (\\mathbf {x}) \\| _ {1}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.204, + 0.905, + 0.249 + ], + "angle": 0, + "content": "where \\(\\omega\\) represents the smoothness weight for each pixel. Following prior research [36], the weight \\(\\omega\\) is computed based on the gradients of the reference image:" + }, + { + "type": "equation", + "bbox": [ + 0.617, + 0.259, + 0.905, + 0.277 + ], + "angle": 0, + "content": "\\[\n\\omega (\\mathbf {x}) = \\exp (- \\mu | \\nabla \\mathbf {L} (\\mathbf {x}) |), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.285, + 0.905, + 0.316 + ], + "angle": 0, + "content": "where \\(\\nabla \\mathbf{L}(\\mathbf{x})\\) denotes the spatial gradient of the logarithmic brightness at pixel \\(\\mathbf{x}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.322, + 0.907, + 0.517 + ], + "angle": 0, + "content": "Optimization details. To enhance the convergence speed and robustness of the optimization process, we employ a patch-based optimization strategy implemented in a coarse-to-fine manner [47]. This strategy involves three resolution scales in a pyramid structure, with the coarsest patch size at \\(32 \\times 32\\) pixels and the finest at \\(8 \\times 8\\) pixels. Bilinear interpolation is utilized for upsampling between successive scales. \\(\\mathbf{v}\\) is initialized uniformly within the range \\([-0.1, 0.1]\\) at the coarsest scale. For the subsequent scale, the initialization is derived from the optimization results of the preceding scale. We utilize the Adam optimizer [24] with a learning rate of 0.05, and the overall number of iterations is set to 40." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.523, + 0.906, + 0.705 + ], + "angle": 0, + "content": "Analysis. An example of coarse motion estimation is presented in Fig. 2. From this example, it is observed that the predicted brightness change \\(\\Delta \\mathbf{L}\\), derived from the coarse motion (Fig. 2(b)), aligns well with the event frame (Fig. 2(a)), which corresponds to \\(\\Delta \\mathbf{L}\\) in Eq. (4). This demonstrates the efficacy of the coarse motion estimation. However, the coarse motion results manifest several limitations. First, due to the sparsity of events, the estimated coarse motion yields valid results only in regions where events are triggered, and it is susceptible to noise. Second, quantization errors in event cameras hinder the accurate recovery of motion amplitude in the coarse motion results." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.799, + 0.73 + ], + "angle": 0, + "content": "3.3. Motion Refinement by Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.906, + 0.87 + ], + "angle": 0, + "content": "With the optimization-based approach described in Sec. 3.2, we obtain the coarse motion results, denoted as \\( f^c \\) for ease of presentation. While \\( f^c \\) enables the brightness change prediction \\( \\Delta \\hat{\\mathbf{L}} \\) to align with the event signals, it exhibits issues including excessive smoothness, susceptibility to noise, and inaccurate motion amplitude estimation. To address these problems, we propose EVibNet, a motion refinement network specifically designed to enhance the coarse motion results." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The pipeline of EVibNet is shown in Fig. 3. The network takes event signals \\(\\mathcal{E}\\), a single reference frame \\(\\mathbf{I}_{\\mathrm{ref}}\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.094, + 0.9, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.296, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Figure 3. Pipeline of our EVibNet. (Input: \\(\\mathbf{I}_{\\mathrm{ref}}\\) and \\(\\varepsilon_{i}\\)) Initially, features are extracted from the reference frame \\(\\mathbf{I}_{\\mathrm{ref}}\\) and events triggered within the time interval \\([t_i,t_{i + 1}]\\) using \\(\\mathcal{F}_{\\mathrm{L}}\\) and \\(\\mathcal{F}_{\\mathrm{E}}\\), respectively. Then the extracted features are fused with a fusion module \\(\\mathcal{F}_{\\mathrm{fuse}}\\). The coarse motion obtained in Sec. 3.2 is fed into a motion encoder \\(\\mathcal{F}_{\\mathrm{M}}\\), generating the motion feature \\(\\mathbf{f}_{t_i,t_{i + 1}}^c\\), which is further refined by the motion refine module \\(\\mathcal{F}_{\\mathrm{refine}}\\). Finally, the refined motion feature is accumulated to obtain the vibration displacement (Outout: \\(\\mathbf{f}_{t_0,t_i + 1}\\))." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.379, + 0.483, + 0.456 + ], + "angle": 0, + "content": "and coarse motion results \\( f^c \\) as inputs. Specifically, EVibNet comprises two principal components: 1) two modality-specific encoders \\( \\mathcal{F}_{\\mathrm{L}} \\) and \\( \\mathcal{F}_{\\mathrm{E}} \\) and a fusion module \\( \\mathcal{F}_{\\mathrm{fuse}} \\); 2) a coarse motion refinement module \\( \\mathcal{F}_{\\mathrm{refine}} \\) and a motion accumulation module \\( \\mathcal{F}_{\\mathrm{acc}} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.46, + 0.484, + 0.672 + ], + "angle": 0, + "content": "Feature extraction. For events \\(\\mathcal{E}_i\\) triggered in the time interval \\([t_i, t_{i+1}]\\), we transform \\(\\mathcal{E}_i\\) into two distinct voxel grids \\(\\mathbf{E}_i^+\\) and \\(\\mathbf{E}_i^-\\) for positive and negative events respectively, following the procedure in [17]. Due to the modality gap between frames and events, we use two modality-specific encoders \\(\\mathcal{F}_{\\mathrm{L}}\\) and \\(\\mathcal{F}_{\\mathrm{E}}\\) to extract features from the reference frame \\(\\mathbf{I}_{\\mathrm{ref}}\\) and events \\(\\mathcal{E}_i\\), respectively. To solve the issue of inaccurate motion amplitude estimation caused by quantization error, we leverage the temporal structure inherent in the event data using a recurrent convolutional encoder. Specifically, a ConvLSTM [45] layer is integrated into each encoder stage of \\(\\mathcal{F}_{\\mathrm{E}}\\). The feature extraction process for the reference frame and event stack is performed using the feature encoders as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.126, + 0.679, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} ^ {\\mathrm {L}} = \\mathcal {F} _ {\\mathrm {L}} \\left(\\mathbf {I} _ {\\text {r e f}}\\right), \\quad \\mathbf {f} _ {i} ^ {\\mathrm {E}} = \\mathcal {F} _ {\\mathrm {E}} \\left(\\mathbf {E} _ {i} ^ {+}, \\mathbf {E} _ {i} ^ {-}, h _ {i - 1}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}^{\\mathrm{L}},\\mathbf{f}_i^{\\mathrm{E}}\\in \\mathbb{R}^{h\\times w\\times D}\\), the feature resolution \\((h,w)\\) is \\(1 / 4\\) of the original image resolution \\((H,W)\\), and \\(h_{i - 1}\\) refers to the hidden states in the recurrent event encoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Given that event signals are triggered in only a few regions when capturing vibrating objects, it is crucial to leverage texture features extracted from the reference frame to regulate the motion features extracted from the events. Following [51], we employ a cross-modal attention fusion block, denoted as \\(\\mathcal{F}_{\\mathrm{fuse}}\\), to adaptively fuse the information from these two modalities, yielding the fused feature \\(\\mathbf{f}_i^{\\mathrm{H}}\\). The cross-modal attention in \\(\\mathcal{F}_{\\mathrm{fuse}}\\) takes as input the queries \\(\\mathbf{Q}^{\\mathrm{E}}\\) derived from the event features \\(\\mathbf{f}_i^{\\mathrm{E}}\\), and the keys \\(\\mathbf{K}^{\\mathrm{L}}\\) and values \\(\\mathbf{V}^{\\mathrm{L}}\\) derived from the image features \\(\\mathbf{f}^{\\mathrm{L}}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.907, + 0.505 + ], + "angle": 0, + "content": "Motion refinement and accumulation. With the frame-event fused feature \\(\\mathbf{f}_i^{\\mathrm{H}}\\), we refine the coarse motion \\(f_{t_i,t_{i + 1}}^c\\) through a motion refinement module, denoted as \\(\\mathcal{F}_{\\mathrm{refine}}\\). The coarse motion \\(f_{t_i,t_{i + 1}}^c\\) is initially encoded into coarse motion features \\(\\mathbf{f}_{t_i,t_{i + 1}}^c\\) using a motion encoder \\(\\mathcal{F}_{\\mathrm{M}}\\). Subsequently, the motion information embedded in \\(\\mathbf{f}_i^{\\mathrm{H}}\\) is leveraged to refine \\(\\mathbf{f}_{t_i,t_{i + 1}}^c\\), producing the refined motion feature \\(\\mathbf{f}_{t_i,t_{i + 1}}^r\\):" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.504, + 0.905, + 0.523 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r} = \\mathcal {F} _ {\\text {r e f i n e}} \\left(\\mathbf {f} _ {i} ^ {\\mathrm {H}}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {c}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.526, + 0.907, + 0.706 + ], + "angle": 0, + "content": "Following the refinement of coarse motion, we further accumulate short-term motion into long-range displacement. This approach is motivated by two factors: First, motion within a very short time interval is typically minimal and thus more susceptible to noise. Second, estimating the displacement of points on vibrating objects relative to a reference state facilitates subsequent analysis. However, direct accumulation of motion at each time step results in significant accumulation errors. Therefore, we integrate short-term motion using a motion accumulation module \\(\\mathcal{F}_{\\mathrm{acc}}\\). The accumulated motion feature \\(\\mathbf{f}_{t_0,t_i}\\) is computed as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.596, + 0.706, + 0.905, + 0.725 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {t _ {0}, t _ {i + 1}} ^ {a c c} = \\mathcal {F} _ {\\text {a c c}} \\left(\\mathbf {f} _ {t _ {0}, t _ {i}} ^ {a c c}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r}, \\mathbf {f} ^ {\\mathrm {L}}\\right), \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.728, + 0.907, + 0.821 + ], + "angle": 0, + "content": "where \\(\\mathbf{f}_{t_0,t_i}^{acc}\\) is derived from the preceding time interval, and \\(\\mathbf{f}^{\\mathrm{L}}\\) serves as the context feature to guide the motion accumulation. For simplicity, \\(\\mathcal{F}_{\\mathrm{acc}}\\) is implemented as a simple RNN comprising two ConvNext blocks [27]. Finally, the accumulated motion feature is decoded into the target displacement \\(f_{t_0,t_{i + 1}}\\) using a motion decoder \\(\\mathcal{F}_{\\mathrm{D}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.828, + 0.733, + 0.844 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Loss functions. Our loss functions are derived from prior work on optical flow estimation [52]. The network is supervised using the \\(l_{1}\\) distance between the predicted and" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24670" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.14 + ], + "angle": 0, + "content": "ground truth flow across the entire sequence of predictions, \\(\\{f_{t_0,t_1},\\dots ,f_{t_0,t_N}\\}\\) . Given the ground truth flow sequence \\(\\{f_{t_0,t_1}^{gt},\\dots ,f_{t_0,t_N}^{gt}\\}\\) , the loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.148, + 0.482, + 0.189 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {t _ {0}, t _ {i}} ^ {g t} - f _ {t _ {0}, t _ {i}} \\| _ {1}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Synthetic dataset. Obtaining ground truth data for real-world subtle motion is challenging. Therefore, following previous studies [7, 35], we simulate subtle motion by composing foreground elements onto a background. We sample images from the MS COCO dataset [26] for the background and use segmented objects from the PASCAL VOC dataset [12] for the foreground. The magnitude and direction of motion for both the background and each foreground object are generated randomly. To simulate sub-pixel motions, we initially generate high-resolution videos, where the motion appears larger, and subsequently downsample each frame to the desired resolution. After obtaining the videos, we use the event simulator V2E [23] to simulate event data from downsampled videos." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.42, + 0.484, + 0.526 + ], + "angle": 0, + "content": "Training details. We implement our approach using the PyTorch framework and conduct all experiments on a single NVIDIA GeForce RTX 3090 GPU. During training, we use the AdamW optimizer [28] and configure the batch size to 4. The EVibNet is trained together for 20 epochs. The initial learning rate is set to \\(2 \\times 10^{-4}\\), and a cosine annealing learning rate scheduling strategy is employed." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.531, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Hybrid camera system. Our hybrid camera system consists of a machine vision camera (HIKVISION MV-CA050-12UC) and an event camera (PROPHESEE GEN4.1), which are co-aligned using a beam splitter (Thorlabs CCM1-BS013). We utilize the calibration method proposed in [32] to achieve pixel alignment." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.224, + 0.65 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.78 + ], + "angle": 0, + "content": "In this section, we first evaluate the sub-pixel motion estimation accuracy of the proposed method on the synthetic dataset in Sec. 4.1. Then we validate our method's ability to detect high-frequency vibrations by capturing and replaying the vibration caused by audio sources in Sec. 4.2. Finally, we demonstrate our method by estimating both homogeneous and heterogeneous material properties of objects with known geometry in Sec. 4.3." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.789, + 0.373, + 0.804 + ], + "angle": 0, + "content": "4.1. Evaluation on Synthetic Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "We compare our method to the phase-based subtle motion estimation approach (denoted as RGBPhase) that is widely used in frame-based visual vibrometry. This approach computes local motion signals from phase shifts within a complex steerable pyramid (CSP) [39, 49, 50]. For the comparison with high-speed videos, we apply RGBPhase to the" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.145 + ], + "angle": 0, + "content": "Table 1. Motion estimation comparison on the synthetic dataset. \\(\\mathbf{E}\\) and \\(\\mathbf{F}\\) denote events and frames, respectively. \\(\\downarrow\\) indicates the lower, the better throughout this paper. The best performances are highlighted in bold." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.156, + 0.905, + 0.257 + ], + "angle": 0, + "content": "
MethodsInputsEPE ↓% Out ↓
EV-FlowNet [37]E0.180547.95
MCM [47]E0.451166.87
RGBPhase [15]F (960Hz)0.283354.78
EMM [7]+RAFT [52]E + F (30Hz)0.219047.63
OursE + F (single)0.083421.29
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.281, + 0.907, + 0.447 + ], + "angle": 0, + "content": "original video sequences used to simulate event streams in our synthetic dataset. Furthermore, we compare our method with two representative event-based optical flow estimation methods: Multi-reference Contrast Maximization (MCM) [47] and the recurrent version of EV-FlowNet [37, 58]. To validate the superior suitability of our method for subtle motion estimation compared to the event-based motion magnification method (EMM) [7], we retrain EMM using our synthetic dataset and leverage an off-the-shelf optical flow estimation method RAFT [52] to estimate optical flow from the motion-magnified videos generated by EMM." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.448, + 0.909, + 0.675 + ], + "angle": 0, + "content": "We adopt the average endpoint error (EPE) and the percentage of pixels with EPE \\(>0.1\\) px (denoted as \\(\\%\\) Out\") as evaluation metrics. The quantitative comparison results are presented in Tab. 1. The results demonstrate that our method substantially outperforms previous event-based optical flow estimation techniques, primarily due to its specific design for subtle motion estimation. Notably, RGBPhase exhibits unsatisfactory performance on our synthetic dataset, likely attributable to the noise in local phase information encoded in the CSP. The inferior performance of EMM + RAFT compared to our method suggests that subtle motion estimation cannot be effectively achieved by simply cascading an optical flow method with EMM. This performance gap may be attributed to the blurring artifacts often introduced by motion magnification techniques." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.734, + 0.699 + ], + "angle": 0, + "content": "4.2. Capturing audio signals" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.909, + 0.902 + ], + "angle": 0, + "content": "To validate the efficacy of our method in real-world high-frequency vibration sensing, we apply it to remote acquisition of audio signals. In the experimental setup depicted in Fig. 4, we capture the membrane vibrations of a speaker playing a linear up-chirp signal ranging from 0 to \\(1000\\mathrm{Hz}\\). Concurrently, we obtain reference audio recordings using a microphone. With the subtle motion extracted by our method, the 1D audio signal is reconstructed by averaging the local motions across all pixel locations and orientations [8]. We compare our method with an event-based sound recovery method (EBVM) proposed by Niwa et al. [34]. By integrating events and frames, our method estimates subtle motion more accurately than the phase-based ap" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "24671" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.092, + 0.278, + 0.18 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.128, + 0.181, + 0.262, + 0.194 + ], + "angle": 0, + "content": "(a) Experimental setup" + }, + { + "type": "image", + "bbox": [ + 0.28, + 0.091, + 0.477, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.181, + 0.437, + 0.194 + ], + "angle": 0, + "content": "(b) Microphone" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.194, + 0.285, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.286, + 0.221, + 0.298 + ], + "angle": 0, + "content": "(c) EBVM" + }, + { + "type": "image", + "bbox": [ + 0.288, + 0.194, + 0.478, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.365, + 0.286, + 0.415, + 0.299 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.312, + 0.484, + 0.381 + ], + "angle": 0, + "content": "Figure 4. Speaker membrane experiments. (a) A speaker is playing an up-chirp signal ranging from 0 to \\(1000\\mathrm{Hz}\\). (b) The spectrogram of the source sound recorded by a microphone. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.398, + 0.484, + 0.454 + ], + "angle": 0, + "content": "Table 2. Percent error in estimating the Young's Modulus for each rod. The best performances are highlighted in bold. The content in each cell refers to the results for rods clamped to a length of 35 cm and 48 cm respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.464, + 0.484, + 0.576 + ], + "angle": 0, + "content": "
% Error ↓LightingAluminumSteelCopper
Video [9]low10.01/10.42-11.64/-9.28-9.40/4.64
medium4.13/-3.70-10.72/-7.62-4.41/-2.70
high-1.45/-1.99-7.94/-5.96-3.14/-2.10
Ourslow2.90/3.21-7.61/-6.58-5.65/2.28
medium-1.26/1.64-6.77/-5.08-3.39/-2.07
high0.39/-1.47-5.93/-5.08-2.25/-2.07
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.602, + 0.483, + 0.661 + ], + "angle": 0, + "content": "proach adopted in EBVM. The experimental results demonstrate significantly reduced noise levels and enhanced high-frequency reconstruction fidelity compared to the baseline method." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.663, + 0.484, + 0.785 + ], + "angle": 0, + "content": "In Fig. 5, we conduct an experiment similar to Davis et al. [8]. A speaker playing the \"MaryMIDI\" audio excites the chip bag. We measure the vibration of a chip bag and recover the audio. The spectrograms of recovered audio reveal superior waveform fidelity achieved by our method. These results demonstrate our method's robustness and practical applicability in complex real-world vibration sensing scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.796, + 0.364, + 0.813 + ], + "angle": 0, + "content": "4.3. Estimating material properties" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.819, + 0.483, + 0.85 + ], + "angle": 0, + "content": "We apply our method to estimate homogeneous and heterogeneous material properties from vibrations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Homogeneous material properties. Considering objects with homogeneous material properties, it can be obtained from Eq. (2) that different objects with the same ge" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.704, + 0.179 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.181, + 0.682, + 0.194 + ], + "angle": 0, + "content": "(a) Experimental setup" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.194, + 0.709, + 0.286 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.591, + 0.286, + 0.65, + 0.299 + ], + "angle": 0, + "content": "(c) EBVM" + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.091, + 0.903, + 0.181 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.77, + 0.181, + 0.862, + 0.195 + ], + "angle": 0, + "content": "(b) Input signal" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.194, + 0.903, + 0.284 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.786, + 0.286, + 0.838, + 0.299 + ], + "angle": 0, + "content": "(d) Ours" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.31, + 0.907, + 0.38 + ], + "angle": 0, + "content": "Figure 5. Audio recovery from a bag of chips. (a) Sound from an audio source (speaker) excites the chip bag. (b) The spectrogram of the input signal sent to the speaker. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.405, + 0.906, + 0.512 + ], + "angle": 0, + "content": "ometry have an identical set of mode shapes, but their resonant frequencies scale proportionally to material properties [9]. Following Davis et al. [9], we apply our method to estimate the homogeneous material properties of various metal rods. The simple geometry of clamped rods makes their vibration well-studied [42]. The fundamental frequency \\(\\omega_{1}\\) (first resonant frequency) of a rod is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.52, + 0.906, + 0.561 + ], + "angle": 0, + "content": "\\[\n\\omega_ {1} = 0. 1 3 9 9 \\frac {d}{L ^ {2}} \\sqrt {\\frac {E}{\\rho}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.906, + 0.674 + ], + "angle": 0, + "content": "where \\(d\\) is the diameter of the rod, \\(L\\) is the length, \\(E\\) is the Young's modulus, and \\(\\rho\\) is the density. While length, diameter, and density can be easily measured, Young's modulus is often measured with a tensile test, which will damage the object. In this experiment, we try to estimate the Young's modulus of the rods by finding their fundamental frequency, which is the highest peak in the motion spectrum." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We test rods with three different metals: aluminum, steel, and copper. Each rod is tested at two lengths (35 cm and 48 cm) and under three brightness conditions (100, 500, and 1500 lux). For comparison with frame-based visual vibrometry, we capture 1000 fps videos under identical settings. Resonant frequencies are global properties and invariant to the viewpoint, ensuring that estimation accuracy is not affected by the viewpoint. In Tab. 2, we compare the estimated Young's modulus values with those provided by the manufacturer. The results demonstrate that our method achieves more robust performance across varying lighting conditions. We further compare the data sizes of 1000 fps videos with signals captured by our imaging system in the rod experiments. Our method requires only a single image; consequently, the data size of the reference frame is negligible." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "24672" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.089, + 0.48, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.483, + 0.311 + ], + "angle": 0, + "content": "Figure 6. Heterogenous material properties estimation result of the defect cube (a) and a homogeneous cube (b), the results are plotted with the same colormaps. Based on the results, we can easily differentiate between a cube with a defect and a homogeneous one." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.484, + 0.489 + ], + "angle": 0, + "content": "gible. At the same 720p resolution, event cameras produce event streams at \\(1 - 5\\mathrm{MB / s}\\), about \\(20\\%\\) of the compressed 1000 fps videos (H. 264 encoded). The discrepancy is more pronounced when considering the actual transmission bandwidth required for high-speed cameras. We also test existing event-based optical flow estimation methods. However, both EV-FlowNet [37] and MCM [47] struggle to accurately estimate the vibrations of the rods, and we can not identify reasonable resonant frequencies in the motion spectra." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.501, + 0.484, + 0.622 + ], + "angle": 0, + "content": "Heterogeneous material properties. When considering objects with heterogeneous material properties, spatial inhomogeneities in these properties influence both the resonant frequencies and mode shapes. Based on the mathematical relationship between modes and material properties, Feng et al. [13] propose a physical-constrained optimization approach to estimate heterogeneous material properties from several image-space vibration modes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To demonstrate the applicability of our method to visual vibration tomography, we experiment on two jelly cubes: one homogeneous and the other with an interior defect (a rubber doll embedded within the jelly cube). We capture data under an initial deformation condition and identify unique modes in the estimated motion fields, as shown in Fig. 1. For each cube, we infer spatially-varying Young's modulus and density values on a \\(10 \\times 10 \\times 10\\) hexahedral mesh with the optimization algorithm proposed by Feng et al. [13]. The reconstruction results are shown in Fig. 6. The reconstruction is obtained using only five unique image-space modes. The defective cube has a rubber doll on top, which is stiffer than the jelly cube. The results Fig. 6(a) show the existence of the defect, and clearly differ from that of the homogeneous cube. Due to the high damping of the jelly cubes, the reconstruction results are not perfectly aligned with the data, but we can still easily differentiate between the two cubes." + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.09, + 0.907, + 0.105 + ], + "angle": 0, + "content": "Table 3. Ablation study of the method design on the synthetic data." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.115, + 0.907, + 0.171 + ], + "angle": 0, + "content": "
CoarseW/o cmW/o accW/o frameOurs
EPE ↓0.36320.09690.12990.17340.0834
% Out ↓51.4525.9136.9942.2521.29
" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.199, + 0.676, + 0.213 + ], + "angle": 0, + "content": "4.4. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.907, + 0.435 + ], + "angle": 0, + "content": "To validate the effectiveness and necessity of each part of our method, we conduct several ablation studies and show the results in Tab. 3. To assess the necessity of motion refinement using EVibNet, we test the performance of coarse motion (denoted as \"coarse\"). To verify the contribution of the coarse motion obtained from optimization, we remove it from the inputs of EVibNet (denoted as \"w/o cm\"). To demonstrate the effectiveness of the motion accumulation module, we directly accumulate motions at each time step (denoted as \"w/o acc\"). Finally, we show the necessity of the reference frame by removing it from inputs (denoted as \"w/o frame\"). As indicated in Tab. 3, our complete method achieves the lowest EPE and % Out, which demonstrates the contribution of each component of our method." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.455, + 0.634, + 0.47 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.483, + 0.906, + 0.589 + ], + "angle": 0, + "content": "We present a bandwidth-efficient approach for high-speed vibration sensing with an Event-RGB hybrid camera system. To extract subtle motion frame event data, we integrate an optimization-based approach with a learning-based motion refinement network. We demonstrate our method through the capture of vibrations induced by audio sources and the estimation of material properties." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.906, + 0.687 + ], + "angle": 0, + "content": "Limitations. In the synthetic dataset, subtle motion is simulated by compositing foreground elements onto a background, which exhibits a gap from real-world vibration scenarios. In the future, simulating more realistic synthetic datasets using the finite element method (FEM) will further improve the performance of event-based visual vibrometry." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.708, + 0.673, + 0.724 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62402014, 62276007), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). Peiqi Duan was also supported by China National Postdoctoral Program for Innovative Talents (Grant No. BX20230010) and China Postdoctoral Science Foundation (Grant No. 2023M740076). The authors thank openbayes.com for providing computing resource." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24673" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.157 + ], + "angle": 0, + "content": "[1] Silvio Bianchi and Emanuele Giacomozzi. Long-range detection of acoustic vibrations by speckle tracking. Applied optics, 58(28):7805-7809, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.484, + 0.214 + ], + "angle": 0, + "content": "[2] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proc. of IEEE International Conference on Computer Vision, 2013. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.217, + 0.483, + 0.285 + ], + "angle": 0, + "content": "[3] Oral Büyüközürk, Mehmet Ali Taşdemir, O Büyüközürk, R Haupt, C Tuakta, and J Chen. Remote detection of debonding in frp-strengthened concrete structures using acoustic-laser technique. In Proc. of Nondestructive Testing of Materials and Structures, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.288, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[4] Oral Buyukozturk, Justin G Chen, Neal Wadhwa, Abe Davis, Frédo Durand, and William T Freeman. Smaller than the eye can see: Vibration analysis with video cameras. In Proc. of World Conference on Non-Destructive Testing, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.345, + 0.483, + 0.401 + ], + "angle": 0, + "content": "[5] Justin G Chen, Robert W Haupt, and Oral Buyukozturk. The acoustic-laser vibrometry technique for the noncontact detection of discontinuities in fiber reinforced polymer-retrofitted concrete. Materials evaluation, 72(10), 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[6] Justin G Chen, Neal Wadhwa, Young-Jin Cha, Fredo Durand, William T Freeman, and Oral Buyukozturk. Modal identification of simple structures with high-speed video using motion magnification. Journal of Sound and Vibration, 345:58-71, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.528 + ], + "angle": 0, + "content": "[7] Yutian Chen, Shi Guo, Fangzheng Yu, Feng Zhang, Jinwei Gu, and Tianfan Xue. Event-based motion magnification. In Proc. of European Conference on Computer Vision, 2024. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.531, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[8] Abe Davis, Michael Rubinstein, Neal Wadhwa, Gautham J. Mysore, Frédo Durand, and William T. Freeman. The visual microphone: passive recovery of sound from video. ACM Transactions on Graphics, 33(4):79:1-79:10, 2014. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.684 + ], + "angle": 0, + "content": "[9] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Oral Buyukozturk, Frédo Durand, and William T Freeman. Visual Vibrometry: Estimating material properties from small motions in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):732-745, 2017. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.687, + 0.483, + 0.756 + ], + "angle": 0, + "content": "[10] Charles Dorn, Sudeep Dasari, Yongchao Yang, Garrett Kenyon, Paul Welch, and David Mascareñas. Efficient fullfield operational modal analysis using neuromorphic event-based imaging. In Proc. of Conference and Exposition on Structural Dynamics, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[11] Franz Durst, Adrian Melling, and James H Whitelaw. Principles and practice of laser-doppler anemometry. NASA STI/Recon Technical Report A, 76:47019, 1976. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.802, + 0.483, + 0.857 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[13] Berthy T Feng, Alexander C Ogren, Chiara Daraio, and Katherine L Bouman. Visual vibration tomography: Estimating interior material properties from monocular video. In" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[14] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. A \\(1280 \\times 720\\) back-illuminated stacked temporal contrast event-based vision sensor with \\(4.86\\mu \\mathrm{m}\\) pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In Proc. of IEEE International Solid-State Circuits Conference, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[15] David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information. International Journal of Computer Vision, 5:77-104, 1990. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.278, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.907, + 0.403 + ], + "angle": 0, + "content": "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-End learning of representations for asynchronous event-based data. In Proc. of IEEE International Conference on Computer Vision, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.405, + 0.907, + 0.46 + ], + "angle": 0, + "content": "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3):601-618, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.462, + 0.905, + 0.516 + ], + "angle": 0, + "content": "[19] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4736-4746, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.519, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[20] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8553-8565, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.589, + 0.905, + 0.644 + ], + "angle": 0, + "content": "[21] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.647, + 0.905, + 0.687 + ], + "angle": 0, + "content": "[22] Matthew Howard and Keigo Hirakawa. Event-based visual microphone. In Proc. of International Conference on Acoustics, Speech and Signal Processing, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[23] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbrück. V2E: From video frames to realistic DVS events. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.787 + ], + "angle": 0, + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of International Conference on Learning Representations, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.789, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A \\(128 \\times 128\\) 120 db \\(15~\\mu \\mathrm{s}\\) latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. of European Conference on Computer Vision, 2014. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "24674" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of International Conference on Learning Representations, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Yuanyuan Lv, Liang Zhou, Zhaohui Liu, and Haiyang Zhang. Structural vibration frequency monitoring based on event camera. Measurement Science and Technology, 35(8): 085007, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.305 + ], + "angle": 0, + "content": "[30] William N MacPherson, Mark Reeves, David P Towers, Andrew J Moore, Julian DC Jones, Martin Dale, and Craig Edwards. Multipoint laser vibrometer for modal analysis. Applied optics, 46(16):3126-3132, 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[31] M Martarelli, Gian Marco Revel, and C Santolini. Automated modal analysis by scanning laser vibrometry: problems and uncertainties associated with the scanning system calibration. Mechanical systems and signal processing, 15 (3):581-601, 2001. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[32] Manasi Muglikar, Mathias Gehrig, Daniel Gehrig, and Davide Scaramuzza. How to calibrate your event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.502 + ], + "angle": 0, + "content": "[33] Woong-jae Na, Kyung Ho Sun, Byeong Chan Jeon, Jaeyun Lee, and Yun-ho Shin. Event-based micro vibration measurement using phase correlation template matching with event filter optimization. Measurement, 215:112867, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[34] Ryogo Niwa, Tatsuki Fushimi, Kenta Yamamoto, and Yoichi Ochiai. Live demonstration: Event-based visual microphone. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.629 + ], + "angle": 0, + "content": "[35] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed A. Elgharib, Frédo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification. In Proc. of European Conference on Computer Vision, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[37] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Proc. of IEEE International Conference on Computer Vision, 2023. 2, 3, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[38] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics Express, 18(10):10762-10774, 2010. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[39] Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. International Journal of Computer Vision, 40:49-70, 2000. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[40] Nicolaas Bernardus Roozen, Ludovic Labelle, Monika Rychtarikova, and Christ Glorieux. Determining radiated" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "sound power of building structures by means of laser doppler vibrometry. Journal of Sound and Vibration, 346:81-99, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.137, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[41] Steve Rothberg, JR Baker, and Neil A Halliwell. Laser vibrometry: pseudo-vibrations. Journal of Sound and Vibration, 135(3):516-522, 1989. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[42] Ahmed A Shabana. Theory of vibration. Springer, 1991. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.211, + 0.905, + 0.265 + ], + "angle": 0, + "content": "[43] Mark Sheinin, Dorian Chan, Matthew O'Toole, and Srinivasa G Narasimhan. Dual-shutter optical vibration sensing. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.269, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[44] Chenyang Shi, Ningfang Song, Boyi Wei, Yuzhen Li, Yibo Zhang, Wenzhuo Li, and Jing Jin. Event-based vibration frequency measurement with laser-assisted illumination based on mixture gaussian distribution. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.34, + 0.905, + 0.408 + ], + "angle": 0, + "content": "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.466 + ], + "angle": 0, + "content": "[46] Shintaro Shiba, Friedhelm Hamann, Yoshimitsu Aoki, and Guillermo Gallego. Event-based background-oriented schlieren. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2011-2026, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[47] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):7742-7759, 2024. 2, 3, 4, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.541, + 0.905, + 0.568 + ], + "angle": 0, + "content": "[48] Peter J Shull. Nondestructive evaluation: theory, techniques, and applications. CRC press, 2002. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.571, + 0.905, + 0.626 + ], + "angle": 0, + "content": "[49] Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. In Proc. of International Conference on Image Processing, 1995. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.683 + ], + "angle": 0, + "content": "[50] Eero P Simoncelli, William T Freeman, Edward H Adelson, and David J Heeger. Shiftable multiscale transforms. IEEE transactions on Information Theory, 38(2):587-607, 1992. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[51] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In Proc. of European Conference on Computer Vision, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.799 + ], + "angle": 0, + "content": "[52] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Proc. of European Conference on Computer Vision, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[53] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics, 32(4):1-10, 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[54] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Riesz pyramids for fast phase-based video magnification. In Proc. of International Conference on Computational Photography, 2014. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "24675" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[55] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[56] Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, and William Freeman. Eulerian video magnification for revealing subtle changes in the world. ACM Transactions on Graphics, 31(4):1-8, 2012. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.275 + ], + "angle": 0, + "content": "[57] Nan Wu and Shinichiro Haruyama. The 20k samples-persecond real time detection of acoustic vibration based on displacement estimation of one-dimensional laser speckle images. Sensors, 21(9):2938, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[58] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Proc. of Robotics: Science and Systems, 2018. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.332 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.956 + ], + "angle": 0, + "content": "24676" + } + ] +] \ No newline at end of file diff --git a/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_origin.pdf b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..61361d7eeb8aba977e89c0597662d92e46d59134 --- /dev/null +++ b/2025/Event-based Visual Vibrometry/b38fc5f7-5d86-436d-8460-9c3eb27d9fce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d2cfc8235088ec01dc8926e24816c30713a1a03e82e5011bfc4494c737853600 +size 2007553 diff --git a/2025/Event-based Visual Vibrometry/full.md b/2025/Event-based Visual Vibrometry/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d3292c51dc6d5a478316cd9dc800344ee4212a97 --- /dev/null +++ b/2025/Event-based Visual Vibrometry/full.md @@ -0,0 +1,367 @@ +# Event-based Visual Vibrometry + +Xinyu Zhou $^{1}$ Peiqi Duan $^{2,3}$ Yeliduosi Xiaokaiti $^{2,3}$ Chao Xu $^{1}$ Boxin Shi $^{2,3*}$ + +$^{1}$ State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University + +$^{2}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University +$^{3}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +{zhouxiny, duanqi0001, shiboxin}@pku.edu.cn, yongqiye@stu.pku.edu.cn, xuchao@cis.pku.edu.cn + +![](images/e87abb35354628ef3e818bbf9bfaf522a6f8bb51ecd41d96ff80fcb78a13f4d0.jpg) +(a) Hybrid camera system + +![](images/a563896e06d22b14780ef5168626cd4c2a85ed587f15ebe298f91fadd1ef7e0d.jpg) +(b) Frame + Event stream + +![](images/bb2ab6f4cd3402565b4c0d7bd9267e5f3b00fe1e3137c76260d95342e97c0402.jpg) +Pixel Displacement +Figure 1. Overview of event-based visual vibrometry. A hybrid camera system, comprising a frame-based camera and an event camera, is utilized to capture subtle object vibrations (a). We propose an event-based subtle motion estimation method that leverages the event stream and a reference frame as input (b). The subtle motion field is extracted at each pixel location across time (c). This motion field can be further analyzed to extract the global motion spectrum and independent vibration modes (d). Each mode is characterized by a resonant frequency and a corresponding mode shape, which manifest as distinct peaks in the motion spectrum. + +![](images/92c97e841170be6ade1bd30cc895c9c0933f399b26f34b6821aff66712ef12e8.jpg) +Motion Spectrum + +# Abstract + +Visual vibrometry has emerged as a powerful technique for remote acquisition of audio and the physical properties of materials. To capture high-frequency vibrations, frame-based approaches often require a high-speed video camera and bright lighting to compensate for the short exposure time. In this paper, we introduce event-based visual vibrometry, a new high-speed visual vibration sensing method using an event camera. By leveraging the high temporal resolution and low bandwidth characteristics of event cameras, event-based visual vibrometry enables high-speed vibration sensing under ambient lighting conditions with improved data efficiency. Specifically, we leverage a hybrid camera system and propose an event-based subtle motion estimation framework that integrates an optimization-based approach based on the event generation model and a motion refinement network. We demonstrate our method by capturing vibration caused by audio sources and estimating material properties for various objects. + +# 1. Introduction + +Vibrations are pervasive phenomena that arise from a wide range of sources, including engines, percussive impacts, and musical instruments. These vibrations typically exhibit small amplitudes and a broad spectrum of frequencies, spanning from a few Hz to MHz, and carry an immense amount of information. Vibration analysis is an established tool in various engineering and scientific fields, including the remote acquisition of audio signals [8], the monitoring of human heart rate [38, 56], and the estimation of material properties [2, 9, 13]. + +Various devices have been used to measure and analyze vibrations. Among these, contact sensors [48] and laser vibrometers [11] are two widely used traditional vibration sensors, both of which are specialized instruments and only measure the vibration of a single surface point. Recently, much progress has been made on visual vibrometry [2, 4, 9, 13], an emerging technique that utilizes general-purpose video cameras for vibration measurement. In contrast to traditional vibration sensors, video-based approaches enable spatially dense measurements of surface motion. However, the acquisition of high-frequency vi + +briations requires high-speed cameras coupled with bright lighting to compensate for the short exposure time, which is impractical in many real-world applications. Additionally, high-speed cameras demand substantial bandwidth requirements. To eliminate the need for high-speed cameras, Sheinin et al. [43] propose a dual-shutter system combined with active speckle-based techniques. Nevertheless, the nature of speckle-based vibration imaging restricts its applicability and confines vibration sampling to only several surface points. In short, frame-based visual vibrometry encounters challenges in efficiently capturing high-frequency vibrations while maintaining bandwidth efficiency and measurement convenience. + +Event cameras [14, 25], characterized by their high temporal resolution, high dynamic range, and low bandwidth requirements, have emerged as a promising solution for high-speed sensing [16]. Unlike conventional cameras, event cameras asynchronously record per-pixel logarithmic scene radiance changes. This signal-triggering mechanism is particularly advantageous for vibration sensing, where radiance changes occur only in a small fraction of pixels, leading to a significantly reduced data rate. The high temporal resolution and dynamic range inherent to event cameras enable the measurement of high-frequency vibrations under ambient lighting. These characteristics position event cameras as potential alternatives to address the constraints of frame-based visual vibrometry. + +Despite the potential of event cameras, estimating subtle motion from event data remains an open problem. Firstly, although event cameras have been applied to small motion scenarios for motion magnification [7], this approach focuses on visualizing small motions and adopts an implicit motion representation. However, the explicit estimation of subtle motion is essential in visual vibrometry for vibration analysis. Secondly, adapting existing event-based motion estimation techniques to vibration scenarios presents significant difficulties. Existing event-based optical flow estimation methods [19, 37, 47, 58] generally assume continuous apparent motion of visual patterns, which is violated by the oscillatory nature of vibrations. Events caused by vibrations are primarily localized along object edges, and this spatial sparsity hinders the effectiveness of event alignment techniques like Contrast Maximization [37, 47]. Moreover, in comparison to general scenarios, vibrations tend to produce significantly fewer events. Estimating subtle motion from such sparse event data is highly vulnerable to noise inherent in event cameras [23]. + +In this paper, we propose event-based visual vibrometry as an efficient and practical solution for vibration measurement and analysis. To address the challenges mentioned above, we employ an Event-RGB hybrid camera system [7, 20, 55] (Fig. 1(a)), which combines event streams with spatial texture information from conventional frames. We pro + +pose an event-based subtle motion estimation framework that takes event streams and a single reference frame as inputs (Fig. 1(b)). Initially, we employ an optimization-based approach grounded in the event generation model [21, 36] to derive coarse motion estimates within short temporal windows. The sparsity of event data restricts the validity of these coarse estimates to only event-triggered regions, while simultaneously introducing undesirable smoothing effects. Additionally, quantization errors in event cameras compromise the accuracy of motion amplitude estimation. To address these issues, we adopt a motion refinement network that aggregates historical motion information to enhance the coarse motion estimation and obtain long-range vibration displacements (Fig. 1(c)). The refined motion fields across time can subsequently be analyzed (Fig. 1(d)) for various visual vibrometry applications. The key contributions of our work are summarized as follows: + +- a bandwidth-efficient visual vibrometry approach that leverages a hybrid camera system to measure high-frequency vibration under ambient lighting conditions; +- an event-based subtle motion estimation framework that integrates a model-based optimization approach and a motion refinement network; +- a thorough evaluation on both synthetic data and real-world vibrations caused by various sources, demonstrating the effectiveness of event-based visual vibrometry. + +# 2. Related Works + +Traditional vibration sensors. This section reviews two prevalent traditional vibration measurement devices: contact sensors and active sensors. Contact sensors operate through direct physical contact with the target object [48]. Accelerometers and piezoelectric pickups are representative examples of contact sensors that respond to acceleration. Among active sensors, laser Doppler vibrometers (LDVs) are the most common type used for vibration measurement [11, 41]. LDVs operate by projecting a laser beam onto the target object and calculating the surface velocity from the phase shift in the reflected light. LDVs have been used to find defects in composite materials [3, 5], examine the integrity of building structures [40], and modal analysis [30, 31]. Both contact sensors and laser vibrometers are typically limited to measuring vibrations at a single point on the surface. + +Frame-based visual vibrometry. Recent studies on video motion magnification [53, 54, 56] have demonstrated the capability to detect subtle motions captured in videos. These methods typically exploit spatial phase variations of the complex steerable pyramid [39, 49, 50] to estimate small local motions. Chen et al. [6] applied this phase-based subtle motion estimation method to quantify the vibration modes of pipes and cantilever beams. Davis + +et al. [8] analyzed object vibrations induced by acoustic excitation to reconstruct the original sound. Davis et al. [9] employed the motion spectra extracted from video to estimate the physical properties of objects. Feng et al. [13] extended this methodology to heterogeneous properties and proposed a physics-constrained optimization approach. Capturing high-frequency vibrations requires expensive 2D high-speed cameras. To address this cost, some works [1, 57] utilize fast 1D sensors for high-frequency acquisition, but they can only measure vibrations along one dimension. Sheinin et al. [43] proposed a dual-shutter system with high-frequency sensing capabilities, integrated with active speckle-based techniques that optically magnify small-amplitude vibrations through laser illumination. + +Event-based vibration measurement. Recent studies have investigated the application of event cameras for vibration measurement [10, 29, 33]. Dorn et al. [10] adapted the phase-based subtle motion extraction method to event data, assuming that each event contributes an equal phase change. Subsequently, they developed an event-based structural monitoring method. Shi et al. [44] incorporated laser active illumination to enhance event-based vibration measurements and introduced a frequency estimation technique based on Gaussian mixture distributions. Niwa et al. [34] presented a non-contact audio recovery setup using an event camera, employing the phase-based approach proposed in [10]. Howard et al. [22] implemented a similar setup, but recovered audio signals by analyzing the zero-crossings of pixels. Chen et al. [7] developed an event-based motion magnification method for amplifying and visualizing subtle motions. However, this approach does not directly estimate subtle motions; instead, motion information is implicitly derived from event signals via neural network processing. In this work, we focus on passive capture and estimation of subtle vibrations through the synergistic use of event signals and conventional frames. + +# 3. Method + +# 3.1. Problem Formulation + +Vibration modal analysis. The vibrations of objects are often well-approximated by linear systems. In modal analysis, a solid object is modeled as a system of point masses interconnected by springs [42]. In this discretized model, the mass matrix $\mathbf{M}$ and the stiffness matrix $\mathbf{K}$ describe mass concentrated and stiffness between each pair of degrees of freedom (DOF), respectively. The equation of motion for this system is given by: + +$$ +\mathbf {M} \ddot {x} + \mathbf {K} x = 0, \tag {1} +$$ + +where $x$ and $\ddot{x}$ denote the displacement and acceleration of each DOF. The object's vibrations exhibit resonant frequen + +cies, which correspond to the eigenvalues of the system: + +$$ +\mathbf {K} \mathbf {u} _ {i} = \omega_ {i} ^ {2} \mathbf {M} \mathbf {u} _ {i}, \tag {2} +$$ + +where $\omega_{i}$ represents a resonant frequency and the corresponding eigenvector $\mathbf{u}_i$ describes the mode shape. To make the theory of vibration accessible to the computer vision audience, we highlight the following key concepts: + +- Mode Shapes $(\mathbf{u}_1\dots \mathbf{u}_n)$ : Each mode shape represents a distinct pattern of vibration for the object. The set of mode shapes forms an orthogonal basis for the vibration. +- Resonant Frequencies $(\omega_{1}\dots \omega_{n})$ : Each mode is associated with a resonant frequency. Resonant frequencies are spatially invariant across the object's surface. + +The global motion power spectrum for an object can be computed by averaging the power spectra of local motions extracted at each pixel location. The modes of vibrations manifest as peaks in the motion power spectrum (Fig. 1(d)). Estimating mechanical properties from vibrations is based on the insight that variations in mechanical properties induce changes in the resonant frequencies and mode shapes for a fixed geometry. + +Event formation model. When the logarithmic change of brightness at pixel $\mathbf{x}_k$ and time $t$ exceeds a specific threshold $C$ , an event signal $e_k = (\mathbf{x}_k, t_k, p_k)$ will be triggered: + +$$ +\left| \mathbf {L} \left(\mathbf {x} _ {k}, t _ {k}\right) - \mathbf {L} \left(\mathbf {x} _ {k}, t _ {k} - \Delta t _ {k}\right) \right| \geq C, \tag {3} +$$ + +where $p_k \in \{1, -1\}$ is the polarity of the brightness change, $\mathbf{L}(\mathbf{x}_k, t_k)$ denotes the logarithmic brightness at pixel $\mathbf{x}_k$ and time $t_k$ , and $t_k - \Delta t_k$ is the time at which the preceding event was triggered at pixel $\mathbf{x}_k$ . Let $\mathcal{E} = \{e_k\}_{k=1}^{N_e}$ denote a set of $N_e$ events triggered within a short time interval, the brightness change can be computed by summing their polarities: + +$$ +\Delta \mathbf {L} (\mathbf {x}) = \sum_ {k = 1} ^ {N _ {e}} p _ {k} C \delta \left(\mathbf {x} - \mathbf {x} _ {k}\right), \tag {4} +$$ + +where the Dirac delta function $\delta$ selects the pixel $\mathbf{x}_k$ . Assuming brightness constancy over the interval, the brightness change can be approximated as the product of the brightness gradient $\nabla \mathbf{L}$ and the displacement $\Delta \mathbf{x}$ , where the displacement is a result of motion with velocity $\mathbf{v}$ on the image plane [18]: + +$$ +\Delta \widehat {\mathbf {L}} (\mathbf {x}) = - \nabla \mathbf {L} (\mathbf {x}) \cdot \Delta \mathbf {x} = - \nabla \mathbf {L} (\mathbf {x}) \cdot \mathbf {v} (\mathbf {x}) \Delta t. \tag {5} +$$ + +# 3.2. Coarse Motion by Optimization + +Given the sparsity of event signals in vibration scenarios, event alignment and correlation techniques [19, 37, 47], commonly used in previous methods, are not applicable. + +![](images/2621cadf4486f7003d5eb57b8e0f68b69baf85cf9464a2a83942c25bb944caf8.jpg) +(a) Events $(\Delta \mathbf{L})$ + +![](images/88d3d31806afc2055e0fc1c53ec5e903ec718880e718561981bb9663e14ff6f3.jpg) +(b) Coarse Motion + +![](images/461fa197b596e8905fdf7ded099d0160339c92f3ce1db7c57828a0f7f3c5b000.jpg) +(c) Predicted $(\Delta \hat{\mathbf{L}})$ +Figure 2. An example (corresponding to the jelly cube in Fig. 1(b)) of coarse motion through optimization. The predicted brightness change (c), denoted as $\Delta \widehat{\mathbf{L}}$ in Eq. (5) aligns well with the event frame (a), represented by $\Delta \mathbf{L}$ in Eq. (4). + +Therefore, we employ a hybrid Event-RGB camera system, integrating high-speed event signals with spatially dense texture information from images, to facilitate subtle motion estimation. Our method takes event streams and a single reference frame as inputs. + +Leveraging the event formation model, we formulate the problem of event-based subtle motion estimation as an optimization task, where we minimize the mismatch between the brightness change $\Delta \mathbf{L}$ , derived from event accumulation (as shown in Eq. (4)), and its prediction $\Delta \widehat{\mathbf{L}}$ , obtained from moving brightness edges (as detailed in Eq. (5)). + +Optimization objective. The composite loss function is a combination of data-fidelity term $E_{\mathrm{data}}$ and regularization term $E_{\mathrm{reg}}$ : + +$$ +E (\mathbf {v}) = E _ {\text {d a t a}} (\mathbf {v}; \mathcal {E}) + \lambda E _ {\text {r e g}} (\nabla \mathbf {v}), \tag {6} +$$ + +where $\lambda$ serves as a balancing parameter, empirically set to 0.5 in our experiments. + +The data-fidelity term $E_{\mathrm{data}}$ measures the difference between the observed brightness change $\Delta \mathbf{L}$ derived from event accumulation and its prediction $\Delta \widehat{\mathbf{L}}$ based on our coarse motion estimation $\mathbf{v}$ : + +$$ +E _ {\mathrm {d a t a}} (\mathbf {v}; \mathcal {E}) = \left\| \frac {\Delta \widehat {\mathbf {L}}}{\| \Delta \widehat {\mathbf {L}} \| _ {2}} \cdot \mathbf {M} - \frac {\Delta \mathbf {L}}{\| \Delta \mathbf {L} \| _ {2}} \right\| _ {1}, \tag {7} +$$ + +where $\mathbf{M}$ is a mask selecting pixels with events triggered. Due to spatial-temporal variations in the contrast threshold $C$ [23], we compute the difference between normalized brightness changes across the image plane to enhance the robustness of the optimization, following established methodologies [21, 46]. To analyze high-frequency vibrations, the subtle motion is estimated within a very short time interval (typically less than $0.5~ms$ ), during which very few event signals are triggered. It is challenging to accurately estimate flow in regions devoid of events, which often correspond to areas with low contrast or zero flow. To mitigate these inaccuracies, we only calculate the data-fidelity loss term at pixel locations with events triggered. + +We assume that flow vectors vary smoothly with sparse discontinuities occurring at object boundaries. The regularization term $E_{\mathrm{reg}}$ encourages smoothness of the estimated motion $\mathbf{v}$ by minimizing the differences between neighboring pixels. This term is defined as: + +$$ +E _ {\text {r e g}} (\nabla \mathbf {v}) = \| \omega (\mathbf {x}) \nabla \mathbf {v} (\mathbf {x}) \| _ {1}, \tag {8} +$$ + +where $\omega$ represents the smoothness weight for each pixel. Following prior research [36], the weight $\omega$ is computed based on the gradients of the reference image: + +$$ +\omega (\mathbf {x}) = \exp (- \mu | \nabla \mathbf {L} (\mathbf {x}) |), \tag {9} +$$ + +where $\nabla \mathbf{L}(\mathbf{x})$ denotes the spatial gradient of the logarithmic brightness at pixel $\mathbf{x}$ . + +Optimization details. To enhance the convergence speed and robustness of the optimization process, we employ a patch-based optimization strategy implemented in a coarse-to-fine manner [47]. This strategy involves three resolution scales in a pyramid structure, with the coarsest patch size at $32 \times 32$ pixels and the finest at $8 \times 8$ pixels. Bilinear interpolation is utilized for upsampling between successive scales. $\mathbf{v}$ is initialized uniformly within the range $[-0.1, 0.1]$ at the coarsest scale. For the subsequent scale, the initialization is derived from the optimization results of the preceding scale. We utilize the Adam optimizer [24] with a learning rate of 0.05, and the overall number of iterations is set to 40. + +Analysis. An example of coarse motion estimation is presented in Fig. 2. From this example, it is observed that the predicted brightness change $\Delta \mathbf{L}$ , derived from the coarse motion (Fig. 2(b)), aligns well with the event frame (Fig. 2(a)), which corresponds to $\Delta \mathbf{L}$ in Eq. (4). This demonstrates the efficacy of the coarse motion estimation. However, the coarse motion results manifest several limitations. First, due to the sparsity of events, the estimated coarse motion yields valid results only in regions where events are triggered, and it is susceptible to noise. Second, quantization errors in event cameras hinder the accurate recovery of motion amplitude in the coarse motion results. + +# 3.3. Motion Refinement by Learning + +With the optimization-based approach described in Sec. 3.2, we obtain the coarse motion results, denoted as $f^c$ for ease of presentation. While $f^c$ enables the brightness change prediction $\Delta \hat{\mathbf{L}}$ to align with the event signals, it exhibits issues including excessive smoothness, susceptibility to noise, and inaccurate motion amplitude estimation. To address these problems, we propose EVibNet, a motion refinement network specifically designed to enhance the coarse motion results. + +The pipeline of EVibNet is shown in Fig. 3. The network takes event signals $\mathcal{E}$ , a single reference frame $\mathbf{I}_{\mathrm{ref}}$ + +![](images/9e114cc8f5467df240c7e27df4a23d3ac749f6788727b5ced6e05ee1fe492273.jpg) +Figure 3. Pipeline of our EVibNet. (Input: $\mathbf{I}_{\mathrm{ref}}$ and $\varepsilon_{i}$ ) Initially, features are extracted from the reference frame $\mathbf{I}_{\mathrm{ref}}$ and events triggered within the time interval $[t_i,t_{i + 1}]$ using $\mathcal{F}_{\mathrm{L}}$ and $\mathcal{F}_{\mathrm{E}}$ , respectively. Then the extracted features are fused with a fusion module $\mathcal{F}_{\mathrm{fuse}}$ . The coarse motion obtained in Sec. 3.2 is fed into a motion encoder $\mathcal{F}_{\mathrm{M}}$ , generating the motion feature $\mathbf{f}_{t_i,t_{i + 1}}^c$ , which is further refined by the motion refine module $\mathcal{F}_{\mathrm{refine}}$ . Finally, the refined motion feature is accumulated to obtain the vibration displacement (Outout: $\mathbf{f}_{t_0,t_i + 1}$ ). + +and coarse motion results $f^c$ as inputs. Specifically, EVibNet comprises two principal components: 1) two modality-specific encoders $\mathcal{F}_{\mathrm{L}}$ and $\mathcal{F}_{\mathrm{E}}$ and a fusion module $\mathcal{F}_{\mathrm{fuse}}$ ; 2) a coarse motion refinement module $\mathcal{F}_{\mathrm{refine}}$ and a motion accumulation module $\mathcal{F}_{\mathrm{acc}}$ . + +Feature extraction. For events $\mathcal{E}_i$ triggered in the time interval $[t_i, t_{i+1}]$ , we transform $\mathcal{E}_i$ into two distinct voxel grids $\mathbf{E}_i^+$ and $\mathbf{E}_i^-$ for positive and negative events respectively, following the procedure in [17]. Due to the modality gap between frames and events, we use two modality-specific encoders $\mathcal{F}_{\mathrm{L}}$ and $\mathcal{F}_{\mathrm{E}}$ to extract features from the reference frame $\mathbf{I}_{\mathrm{ref}}$ and events $\mathcal{E}_i$ , respectively. To solve the issue of inaccurate motion amplitude estimation caused by quantization error, we leverage the temporal structure inherent in the event data using a recurrent convolutional encoder. Specifically, a ConvLSTM [45] layer is integrated into each encoder stage of $\mathcal{F}_{\mathrm{E}}$ . The feature extraction process for the reference frame and event stack is performed using the feature encoders as follows: + +$$ +\mathbf {f} ^ {\mathrm {L}} = \mathcal {F} _ {\mathrm {L}} \left(\mathbf {I} _ {\text {r e f}}\right), \quad \mathbf {f} _ {i} ^ {\mathrm {E}} = \mathcal {F} _ {\mathrm {E}} \left(\mathbf {E} _ {i} ^ {+}, \mathbf {E} _ {i} ^ {-}, h _ {i - 1}\right), \tag {10} +$$ + +where $\mathbf{f}^{\mathrm{L}},\mathbf{f}_i^{\mathrm{E}}\in \mathbb{R}^{h\times w\times D}$ , the feature resolution $(h,w)$ is $1 / 4$ of the original image resolution $(H,W)$ , and $h_{i - 1}$ refers to the hidden states in the recurrent event encoder. + +Given that event signals are triggered in only a few regions when capturing vibrating objects, it is crucial to leverage texture features extracted from the reference frame to regulate the motion features extracted from the events. Following [51], we employ a cross-modal attention fusion block, denoted as $\mathcal{F}_{\mathrm{fuse}}$ , to adaptively fuse the information from these two modalities, yielding the fused feature $\mathbf{f}_i^{\mathrm{H}}$ . The cross-modal attention in $\mathcal{F}_{\mathrm{fuse}}$ takes as input the queries $\mathbf{Q}^{\mathrm{E}}$ derived from the event features $\mathbf{f}_i^{\mathrm{E}}$ , and the keys $\mathbf{K}^{\mathrm{L}}$ and values $\mathbf{V}^{\mathrm{L}}$ derived from the image features $\mathbf{f}^{\mathrm{L}}$ . + +Motion refinement and accumulation. With the frame-event fused feature $\mathbf{f}_i^{\mathrm{H}}$ , we refine the coarse motion $f_{t_i,t_{i + 1}}^c$ through a motion refinement module, denoted as $\mathcal{F}_{\mathrm{refine}}$ . The coarse motion $f_{t_i,t_{i + 1}}^c$ is initially encoded into coarse motion features $\mathbf{f}_{t_i,t_{i + 1}}^c$ using a motion encoder $\mathcal{F}_{\mathrm{M}}$ . Subsequently, the motion information embedded in $\mathbf{f}_i^{\mathrm{H}}$ is leveraged to refine $\mathbf{f}_{t_i,t_{i + 1}}^c$ , producing the refined motion feature $\mathbf{f}_{t_i,t_{i + 1}}^r$ : + +$$ +\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r} = \mathcal {F} _ {\text {r e f i n e}} \left(\mathbf {f} _ {i} ^ {\mathrm {H}}, \mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {c}\right). \tag {11} +$$ + +Following the refinement of coarse motion, we further accumulate short-term motion into long-range displacement. This approach is motivated by two factors: First, motion within a very short time interval is typically minimal and thus more susceptible to noise. Second, estimating the displacement of points on vibrating objects relative to a reference state facilitates subsequent analysis. However, direct accumulation of motion at each time step results in significant accumulation errors. Therefore, we integrate short-term motion using a motion accumulation module $\mathcal{F}_{\mathrm{acc}}$ . The accumulated motion feature $\mathbf{f}_{t_0,t_i}$ is computed as follows: + +$$ +\mathbf {f} _ {t _ {0}, t _ {i + 1}} ^ {a c c} = \mathcal {F} _ {\text {a c c}} \left(\mathbf {f} _ {t _ {0}, t _ {i}} ^ {a c c}, \mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r}, \mathbf {f} ^ {\mathrm {L}}\right), \tag {12} +$$ + +where $\mathbf{f}_{t_0,t_i}^{acc}$ is derived from the preceding time interval, and $\mathbf{f}^{\mathrm{L}}$ serves as the context feature to guide the motion accumulation. For simplicity, $\mathcal{F}_{\mathrm{acc}}$ is implemented as a simple RNN comprising two ConvNext blocks [27]. Finally, the accumulated motion feature is decoded into the target displacement $f_{t_0,t_{i + 1}}$ using a motion decoder $\mathcal{F}_{\mathrm{D}}$ . + +# 3.4. Implementation Details + +Loss functions. Our loss functions are derived from prior work on optical flow estimation [52]. The network is supervised using the $l_{1}$ distance between the predicted and + +ground truth flow across the entire sequence of predictions, $\{f_{t_0,t_1},\dots ,f_{t_0,t_N}\}$ . Given the ground truth flow sequence $\{f_{t_0,t_1}^{gt},\dots ,f_{t_0,t_N}^{gt}\}$ , the loss is defined as: + +$$ +\mathcal {L} = \sum_ {i = 1} ^ {N} \| f _ {t _ {0}, t _ {i}} ^ {g t} - f _ {t _ {0}, t _ {i}} \| _ {1}. \tag {13} +$$ + +Synthetic dataset. Obtaining ground truth data for real-world subtle motion is challenging. Therefore, following previous studies [7, 35], we simulate subtle motion by composing foreground elements onto a background. We sample images from the MS COCO dataset [26] for the background and use segmented objects from the PASCAL VOC dataset [12] for the foreground. The magnitude and direction of motion for both the background and each foreground object are generated randomly. To simulate sub-pixel motions, we initially generate high-resolution videos, where the motion appears larger, and subsequently downsample each frame to the desired resolution. After obtaining the videos, we use the event simulator V2E [23] to simulate event data from downsampled videos. + +Training details. We implement our approach using the PyTorch framework and conduct all experiments on a single NVIDIA GeForce RTX 3090 GPU. During training, we use the AdamW optimizer [28] and configure the batch size to 4. The EVibNet is trained together for 20 epochs. The initial learning rate is set to $2 \times 10^{-4}$ , and a cosine annealing learning rate scheduling strategy is employed. + +Hybrid camera system. Our hybrid camera system consists of a machine vision camera (HIKVISION MV-CA050-12UC) and an event camera (PROPHESEE GEN4.1), which are co-aligned using a beam splitter (Thorlabs CCM1-BS013). We utilize the calibration method proposed in [32] to achieve pixel alignment. + +# 4. Experiments + +In this section, we first evaluate the sub-pixel motion estimation accuracy of the proposed method on the synthetic dataset in Sec. 4.1. Then we validate our method's ability to detect high-frequency vibrations by capturing and replaying the vibration caused by audio sources in Sec. 4.2. Finally, we demonstrate our method by estimating both homogeneous and heterogeneous material properties of objects with known geometry in Sec. 4.3. + +# 4.1. Evaluation on Synthetic Dataset + +We compare our method to the phase-based subtle motion estimation approach (denoted as RGBPhase) that is widely used in frame-based visual vibrometry. This approach computes local motion signals from phase shifts within a complex steerable pyramid (CSP) [39, 49, 50]. For the comparison with high-speed videos, we apply RGBPhase to the + +Table 1. Motion estimation comparison on the synthetic dataset. $\mathbf{E}$ and $\mathbf{F}$ denote events and frames, respectively. $\downarrow$ indicates the lower, the better throughout this paper. The best performances are highlighted in bold. + +
MethodsInputsEPE ↓% Out ↓
EV-FlowNet [37]E0.180547.95
MCM [47]E0.451166.87
RGBPhase [15]F (960Hz)0.283354.78
EMM [7]+RAFT [52]E + F (30Hz)0.219047.63
OursE + F (single)0.083421.29
+ +original video sequences used to simulate event streams in our synthetic dataset. Furthermore, we compare our method with two representative event-based optical flow estimation methods: Multi-reference Contrast Maximization (MCM) [47] and the recurrent version of EV-FlowNet [37, 58]. To validate the superior suitability of our method for subtle motion estimation compared to the event-based motion magnification method (EMM) [7], we retrain EMM using our synthetic dataset and leverage an off-the-shelf optical flow estimation method RAFT [52] to estimate optical flow from the motion-magnified videos generated by EMM. + +We adopt the average endpoint error (EPE) and the percentage of pixels with EPE $>0.1$ px (denoted as $\%$ Out") as evaluation metrics. The quantitative comparison results are presented in Tab. 1. The results demonstrate that our method substantially outperforms previous event-based optical flow estimation techniques, primarily due to its specific design for subtle motion estimation. Notably, RGBPhase exhibits unsatisfactory performance on our synthetic dataset, likely attributable to the noise in local phase information encoded in the CSP. The inferior performance of EMM + RAFT compared to our method suggests that subtle motion estimation cannot be effectively achieved by simply cascading an optical flow method with EMM. This performance gap may be attributed to the blurring artifacts often introduced by motion magnification techniques. + +# 4.2. Capturing audio signals + +To validate the efficacy of our method in real-world high-frequency vibration sensing, we apply it to remote acquisition of audio signals. In the experimental setup depicted in Fig. 4, we capture the membrane vibrations of a speaker playing a linear up-chirp signal ranging from 0 to $1000\mathrm{Hz}$ . Concurrently, we obtain reference audio recordings using a microphone. With the subtle motion extracted by our method, the 1D audio signal is reconstructed by averaging the local motions across all pixel locations and orientations [8]. We compare our method with an event-based sound recovery method (EBVM) proposed by Niwa et al. [34]. By integrating events and frames, our method estimates subtle motion more accurately than the phase-based ap + +![](images/e845338d9e6893f378e543826ed912fa86d609c191fbf170ff91bac7e5e308c9.jpg) + +![](images/b8628e31ac83d5ef426f719964cf20594135941fef79d2e245e54ae6f00fc430.jpg) +(b) Microphone + +![](images/0a2eec5cb4e11dc7073874c3907988bc455360449ae602c159123c11e34db945.jpg) +(a) Experimental setup +(c) EBVM +Figure 4. Speaker membrane experiments. (a) A speaker is playing an up-chirp signal ranging from 0 to $1000\mathrm{Hz}$ . (b) The spectrogram of the source sound recorded by a microphone. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound. + +![](images/c3459ef65fb27c831e4aca8b388935a37e265b498fd74fc4bb332174e94c132e.jpg) +(d) Ours + +Table 2. Percent error in estimating the Young's Modulus for each rod. The best performances are highlighted in bold. The content in each cell refers to the results for rods clamped to a length of 35 cm and 48 cm respectively. + +
% Error ↓LightingAluminumSteelCopper
Video [9]low10.01/10.42-11.64/-9.28-9.40/4.64
medium4.13/-3.70-10.72/-7.62-4.41/-2.70
high-1.45/-1.99-7.94/-5.96-3.14/-2.10
Ourslow2.90/3.21-7.61/-6.58-5.65/2.28
medium-1.26/1.64-6.77/-5.08-3.39/-2.07
high0.39/-1.47-5.93/-5.08-2.25/-2.07
+ +proach adopted in EBVM. The experimental results demonstrate significantly reduced noise levels and enhanced high-frequency reconstruction fidelity compared to the baseline method. + +In Fig. 5, we conduct an experiment similar to Davis et al. [8]. A speaker playing the "MaryMIDI" audio excites the chip bag. We measure the vibration of a chip bag and recover the audio. The spectrograms of recovered audio reveal superior waveform fidelity achieved by our method. These results demonstrate our method's robustness and practical applicability in complex real-world vibration sensing scenarios. + +# 4.3. Estimating material properties + +We apply our method to estimate homogeneous and heterogeneous material properties from vibrations. + +Homogeneous material properties. Considering objects with homogeneous material properties, it can be obtained from Eq. (2) that different objects with the same ge + +![](images/bcb04966c96a449a84c6d96aa65c5705a971bddbd41285994eedcf2c503205cc.jpg) + +![](images/7cfa53b66d52518adf7fcebcae6fa5384118495061dae17efd92ba87b3daff2e.jpg) +(a) Experimental setup +(c) EBVM +Figure 5. Audio recovery from a bag of chips. (a) Sound from an audio source (speaker) excites the chip bag. (b) The spectrogram of the input signal sent to the speaker. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound. + +![](images/e2507dae63c5e301867152e273afcae7c8a02f506a4bc6892ffd30b94d2502de.jpg) +(b) Input signal + +![](images/5e7e702c9b1a8a7654f8bf35644ecd7cbda2e4df2799a7056b3b99398185c1aa.jpg) +(d) Ours + +ometry have an identical set of mode shapes, but their resonant frequencies scale proportionally to material properties [9]. Following Davis et al. [9], we apply our method to estimate the homogeneous material properties of various metal rods. The simple geometry of clamped rods makes their vibration well-studied [42]. The fundamental frequency $\omega_{1}$ (first resonant frequency) of a rod is given by: + +$$ +\omega_ {1} = 0. 1 3 9 9 \frac {d}{L ^ {2}} \sqrt {\frac {E}{\rho}}, \tag {14} +$$ + +where $d$ is the diameter of the rod, $L$ is the length, $E$ is the Young's modulus, and $\rho$ is the density. While length, diameter, and density can be easily measured, Young's modulus is often measured with a tensile test, which will damage the object. In this experiment, we try to estimate the Young's modulus of the rods by finding their fundamental frequency, which is the highest peak in the motion spectrum. + +We test rods with three different metals: aluminum, steel, and copper. Each rod is tested at two lengths (35 cm and 48 cm) and under three brightness conditions (100, 500, and 1500 lux). For comparison with frame-based visual vibrometry, we capture 1000 fps videos under identical settings. Resonant frequencies are global properties and invariant to the viewpoint, ensuring that estimation accuracy is not affected by the viewpoint. In Tab. 2, we compare the estimated Young's modulus values with those provided by the manufacturer. The results demonstrate that our method achieves more robust performance across varying lighting conditions. We further compare the data sizes of 1000 fps videos with signals captured by our imaging system in the rod experiments. Our method requires only a single image; consequently, the data size of the reference frame is negligible. + +![](images/fae9ad03b961d72790633bf248f0322b93ee8d6acbea9d61fd33f58489bf1300.jpg) +Figure 6. Heterogenous material properties estimation result of the defect cube (a) and a homogeneous cube (b), the results are plotted with the same colormaps. Based on the results, we can easily differentiate between a cube with a defect and a homogeneous one. + +gible. At the same 720p resolution, event cameras produce event streams at $1 - 5\mathrm{MB / s}$ , about $20\%$ of the compressed 1000 fps videos (H. 264 encoded). The discrepancy is more pronounced when considering the actual transmission bandwidth required for high-speed cameras. We also test existing event-based optical flow estimation methods. However, both EV-FlowNet [37] and MCM [47] struggle to accurately estimate the vibrations of the rods, and we can not identify reasonable resonant frequencies in the motion spectra. + +Heterogeneous material properties. When considering objects with heterogeneous material properties, spatial inhomogeneities in these properties influence both the resonant frequencies and mode shapes. Based on the mathematical relationship between modes and material properties, Feng et al. [13] propose a physical-constrained optimization approach to estimate heterogeneous material properties from several image-space vibration modes. + +To demonstrate the applicability of our method to visual vibration tomography, we experiment on two jelly cubes: one homogeneous and the other with an interior defect (a rubber doll embedded within the jelly cube). We capture data under an initial deformation condition and identify unique modes in the estimated motion fields, as shown in Fig. 1. For each cube, we infer spatially-varying Young's modulus and density values on a $10 \times 10 \times 10$ hexahedral mesh with the optimization algorithm proposed by Feng et al. [13]. The reconstruction results are shown in Fig. 6. The reconstruction is obtained using only five unique image-space modes. The defective cube has a rubber doll on top, which is stiffer than the jelly cube. The results Fig. 6(a) show the existence of the defect, and clearly differ from that of the homogeneous cube. Due to the high damping of the jelly cubes, the reconstruction results are not perfectly aligned with the data, but we can still easily differentiate between the two cubes. + +Table 3. Ablation study of the method design on the synthetic data. + +
CoarseW/o cmW/o accW/o frameOurs
EPE ↓0.36320.09690.12990.17340.0834
% Out ↓51.4525.9136.9942.2521.29
+ +# 4.4. Ablation studies + +To validate the effectiveness and necessity of each part of our method, we conduct several ablation studies and show the results in Tab. 3. To assess the necessity of motion refinement using EVibNet, we test the performance of coarse motion (denoted as "coarse"). To verify the contribution of the coarse motion obtained from optimization, we remove it from the inputs of EVibNet (denoted as "w/o cm"). To demonstrate the effectiveness of the motion accumulation module, we directly accumulate motions at each time step (denoted as "w/o acc"). Finally, we show the necessity of the reference frame by removing it from inputs (denoted as "w/o frame"). As indicated in Tab. 3, our complete method achieves the lowest EPE and % Out, which demonstrates the contribution of each component of our method. + +# 5. Conclusion + +We present a bandwidth-efficient approach for high-speed vibration sensing with an Event-RGB hybrid camera system. To extract subtle motion frame event data, we integrate an optimization-based approach with a learning-based motion refinement network. We demonstrate our method through the capture of vibrations induced by audio sources and the estimation of material properties. + +Limitations. In the synthetic dataset, subtle motion is simulated by compositing foreground elements onto a background, which exhibits a gap from real-world vibration scenarios. In the future, simulating more realistic synthetic datasets using the finite element method (FEM) will further improve the performance of event-based visual vibrometry. + +# Acknowledgement + +This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62402014, 62276007), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). Peiqi Duan was also supported by China National Postdoctoral Program for Innovative Talents (Grant No. BX20230010) and China Postdoctoral Science Foundation (Grant No. 2023M740076). The authors thank openbayes.com for providing computing resource. + +# References + +[1] Silvio Bianchi and Emanuele Giacomozzi. Long-range detection of acoustic vibrations by speckle tracking. Applied optics, 58(28):7805-7809, 2019. 3 +[2] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proc. of IEEE International Conference on Computer Vision, 2013. 1 +[3] Oral Büyüközürk, Mehmet Ali Taşdemir, O Büyüközürk, R Haupt, C Tuakta, and J Chen. Remote detection of debonding in frp-strengthened concrete structures using acoustic-laser technique. In Proc. of Nondestructive Testing of Materials and Structures, 2013. 2 +[4] Oral Buyukozturk, Justin G Chen, Neal Wadhwa, Abe Davis, Frédo Durand, and William T Freeman. Smaller than the eye can see: Vibration analysis with video cameras. In Proc. of World Conference on Non-Destructive Testing, 2016. 1 +[5] Justin G Chen, Robert W Haupt, and Oral Buyukozturk. The acoustic-laser vibrometry technique for the noncontact detection of discontinuities in fiber reinforced polymer-retrofitted concrete. Materials evaluation, 72(10), 2014. 2 +[6] Justin G Chen, Neal Wadhwa, Young-Jin Cha, Fredo Durand, William T Freeman, and Oral Buyukozturk. Modal identification of simple structures with high-speed video using motion magnification. Journal of Sound and Vibration, 345:58-71, 2015. 2 +[7] Yutian Chen, Shi Guo, Fangzheng Yu, Feng Zhang, Jinwei Gu, and Tianfan Xue. Event-based motion magnification. In Proc. of European Conference on Computer Vision, 2024. 2, 3, 6 +[8] Abe Davis, Michael Rubinstein, Neal Wadhwa, Gautham J. Mysore, Frédo Durand, and William T. Freeman. The visual microphone: passive recovery of sound from video. ACM Transactions on Graphics, 33(4):79:1-79:10, 2014. 1, 3, 6, 7 +[9] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Oral Buyukozturk, Frédo Durand, and William T Freeman. Visual Vibrometry: Estimating material properties from small motions in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):732-745, 2017. 1, 3, 7 +[10] Charles Dorn, Sudeep Dasari, Yongchao Yang, Garrett Kenyon, Paul Welch, and David Mascareñas. Efficient fullfield operational modal analysis using neuromorphic event-based imaging. In Proc. of Conference and Exposition on Structural Dynamics, 2017. 3 +[11] Franz Durst, Adrian Melling, and James H Whitelaw. Principles and practice of laser-doppler anemometry. NASA STI/Recon Technical Report A, 76:47019, 1976. 1, 2 +[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6 +[13] Berthy T Feng, Alexander C Ogren, Chiara Daraio, and Katherine L Bouman. Visual vibration tomography: Estimating interior material properties from monocular video. In + +Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 8 +[14] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. A $1280 \times 720$ back-illuminated stacked temporal contrast event-based vision sensor with $4.86\mu \mathrm{m}$ pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In Proc. of IEEE International Solid-State Circuits Conference, 2020. 2 +[15] David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information. International Journal of Computer Vision, 5:77-104, 1990. 6 +[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2 +[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-End learning of representations for asynchronous event-based data. In Proc. of IEEE International Conference on Computer Vision, 2019. 5 +[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3):601-618, 2020. 3 +[19] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4736-4746, 2024. 2, 3 +[20] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8553-8565, 2023. 2 +[21] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4 +[22] Matthew Howard and Keigo Hirakawa. Event-based visual microphone. In Proc. of International Conference on Acoustics, Speech and Signal Processing, 2023. 3 +[23] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbrück. V2E: From video frames to realistic DVS events. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 2, 4, 6 +[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of International Conference on Learning Representations, 2015. 4 +[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A $128 \times 128$ 120 db $15~\mu \mathrm{s}$ latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 2 +[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. of European Conference on Computer Vision, 2014. 6 + +[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5 +[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of International Conference on Learning Representations, 2017. 6 +[29] Yuanyuan Lv, Liang Zhou, Zhaohui Liu, and Haiyang Zhang. Structural vibration frequency monitoring based on event camera. Measurement Science and Technology, 35(8): 085007, 2024. 3 +[30] William N MacPherson, Mark Reeves, David P Towers, Andrew J Moore, Julian DC Jones, Martin Dale, and Craig Edwards. Multipoint laser vibrometer for modal analysis. Applied optics, 46(16):3126-3132, 2007. 2 +[31] M Martarelli, Gian Marco Revel, and C Santolini. Automated modal analysis by scanning laser vibrometry: problems and uncertainties associated with the scanning system calibration. Mechanical systems and signal processing, 15 (3):581-601, 2001. 2 +[32] Manasi Muglikar, Mathias Gehrig, Daniel Gehrig, and Davide Scaramuzza. How to calibrate your event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 6 +[33] Woong-jae Na, Kyung Ho Sun, Byeong Chan Jeon, Jaeyun Lee, and Yun-ho Shin. Event-based micro vibration measurement using phase correlation template matching with event filter optimization. Measurement, 215:112867, 2023. 3 +[34] Ryogo Niwa, Tatsuki Fushimi, Kenta Yamamoto, and Yoichi Ochiai. Live demonstration: Event-based visual microphone. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023. 3, 6, 7 +[35] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed A. Elgharib, Frédo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification. In Proc. of European Conference on Computer Vision, 2018. 6 +[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 4 +[37] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Proc. of IEEE International Conference on Computer Vision, 2023. 2, 3, 6, 8 +[38] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics Express, 18(10):10762-10774, 2010. 1 +[39] Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. International Journal of Computer Vision, 40:49-70, 2000. 2, 6 +[40] Nicolaas Bernardus Roozen, Ludovic Labelle, Monika Rychtarikova, and Christ Glorieux. Determining radiated + +sound power of building structures by means of laser doppler vibrometry. Journal of Sound and Vibration, 346:81-99, 2015. 2 +[41] Steve Rothberg, JR Baker, and Neil A Halliwell. Laser vibrometry: pseudo-vibrations. Journal of Sound and Vibration, 135(3):516-522, 1989. 2 +[42] Ahmed A Shabana. Theory of vibration. Springer, 1991. 3, 7 +[43] Mark Sheinin, Dorian Chan, Matthew O'Toole, and Srinivasa G Narasimhan. Dual-shutter optical vibration sensing. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 3 +[44] Chenyang Shi, Ningfang Song, Boyi Wei, Yuzhen Li, Yibo Zhang, Wenzhuo Li, and Jing Jin. Event-based vibration frequency measurement with laser-assisted illumination based on mixture gaussian distribution. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 3 +[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems, 2015. 5 +[46] Shintaro Shiba, Friedhelm Hamann, Yoshimitsu Aoki, and Guillermo Gallego. Event-based background-oriented schlieren. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2011-2026, 2024. 4 +[47] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):7742-7759, 2024. 2, 3, 4, 6, 8 +[48] Peter J Shull. Nondestructive evaluation: theory, techniques, and applications. CRC press, 2002. 1, 2 +[49] Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. In Proc. of International Conference on Image Processing, 1995. 2, 6 +[50] Eero P Simoncelli, William T Freeman, Edward H Adelson, and David J Heeger. Shiftable multiscale transforms. IEEE transactions on Information Theory, 38(2):587-607, 1992. 2, 6 +[51] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In Proc. of European Conference on Computer Vision, 2022. 5 +[52] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Proc. of European Conference on Computer Vision, 2020. 5, 6 +[53] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics, 32(4):1-10, 2013. 2 +[54] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Riesz pyramids for fast phase-based video magnification. In Proc. of International Conference on Computational Photography, 2014. 2 + +[55] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2 +[56] Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, and William Freeman. Eulerian video magnification for revealing subtle changes in the world. ACM Transactions on Graphics, 31(4):1-8, 2012. 1, 2 +[57] Nan Wu and Shinichiro Haruyama. The 20k samples-persecond real time detection of acoustic vibration based on displacement estimation of one-dimensional laser speckle images. Sensors, 21(9):2938, 2021. 3 +[58] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Proc. of Robotics: Science and Systems, 2018. 2, 6 \ No newline at end of file diff --git a/2025/Event-based Visual Vibrometry/images.zip b/2025/Event-based Visual Vibrometry/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2f66e3a9908f9bf93084458902e3d8c11bc29798 --- /dev/null +++ b/2025/Event-based Visual Vibrometry/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2bdf9868b8d9a93636cd363d323ea2578a658c791c70546c86fbe783fcd68ef3 +size 438819 diff --git a/2025/Event-based Visual Vibrometry/layout.json b/2025/Event-based Visual Vibrometry/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3e34fdee86d323a488325bcbffbbff7eb6fe8fa1 --- /dev/null +++ b/2025/Event-based Visual Vibrometry/layout.json @@ -0,0 +1,10718 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 205, + 104, + 402, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 205, + 104, + 402, + 121 + ], + "spans": [ + { + "bbox": [ + 205, + 104, + 402, + 121 + ], + "type": "text", + "content": "Event-based Visual Vibrometry" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "spans": [ + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "content": "Xinyu Zhou" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "content": " Peiqi Duan" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "content": " Yeliduosi Xiaokaiti" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "inline_equation", + "content": "^{2,3}" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "content": " Chao Xu" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 107, + 137, + 510, + 152 + ], + "type": "inline_equation", + "content": "^{2,3*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 86, + 153, + 533, + 166 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 153, + 533, + 166 + ], + "spans": [ + { + "bbox": [ + 86, + 153, + 533, + 166 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 86, + 153, + 533, + 166 + ], + "type": "text", + "content": " State Key Laboratory of General Artificial Intelligence, School of Intelligence Science and Technology, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 107, + 167, + 510, + 193 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 107, + 167, + 510, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 167, + 510, + 179 + ], + "spans": [ + { + "bbox": [ + 107, + 167, + 510, + 179 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 167, + 510, + 179 + ], + "type": "text", + "content": " State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 180, + 510, + 193 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 180, + 510, + 193 + ], + "spans": [ + { + "bbox": [ + 107, + 180, + 510, + 193 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 180, + 510, + 193 + ], + "type": "text", + "content": "National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 62, + 196, + 553, + 208 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 196, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 62, + 196, + 553, + 208 + ], + "type": "text", + "content": "{zhouxiny, duanqi0001, shiboxin}@pku.edu.cn, yongqiye@stu.pku.edu.cn, xuchao@cis.pku.edu.cn" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 61, + 217, + 207, + 333 + ], + "blocks": [ + { + "bbox": [ + 61, + 217, + 207, + 333 + ], + "lines": [ + { + "bbox": [ + 61, + 217, + 207, + 333 + ], + "spans": [ + { + "bbox": [ + 61, + 217, + 207, + 333 + ], + "type": "image", + "image_path": "e87abb35354628ef3e818bbf9bfaf522a6f8bb51ecd41d96ff80fcb78a13f4d0.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 333, + 206, + 344 + ], + "lines": [ + { + "bbox": [ + 104, + 333, + 206, + 344 + ], + "spans": [ + { + "bbox": [ + 104, + 333, + 206, + 344 + ], + "type": "text", + "content": "(a) Hybrid camera system" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 213, + 236, + 338, + 316 + ], + "blocks": [ + { + "bbox": [ + 213, + 236, + 338, + 316 + ], + "lines": [ + { + "bbox": [ + 213, + 236, + 338, + 316 + ], + "spans": [ + { + "bbox": [ + 213, + 236, + 338, + 316 + ], + "type": "image", + "image_path": "a563896e06d22b14780ef5168626cd4c2a85ed587f15ebe298f91fadd1ef7e0d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 319, + 331, + 329 + ], + "lines": [ + { + "bbox": [ + 228, + 319, + 331, + 329 + ], + "spans": [ + { + "bbox": [ + 228, + 319, + 331, + 329 + ], + "type": "text", + "content": "(b) Frame + Event stream" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 358, + 219, + 444, + 350 + ], + "blocks": [ + { + "bbox": [ + 358, + 219, + 444, + 350 + ], + "lines": [ + { + "bbox": [ + 358, + 219, + 444, + 350 + ], + "spans": [ + { + "bbox": [ + 358, + 219, + 444, + 350 + ], + "type": "image", + "image_path": "bb2ab6f4cd3402565b4c0d7bd9267e5f3b00fe1e3137c76260d95342e97c0402.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 376, + 350, + 439, + 360 + ], + "lines": [ + { + "bbox": [ + 376, + 350, + 439, + 360 + ], + "spans": [ + { + "bbox": [ + 376, + 350, + 439, + 360 + ], + "type": "text", + "content": "Pixel Displacement" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 370, + 555, + 426 + ], + "lines": [ + { + "bbox": [ + 54, + 370, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 555, + 426 + ], + "type": "text", + "content": "Figure 1. Overview of event-based visual vibrometry. A hybrid camera system, comprising a frame-based camera and an event camera, is utilized to capture subtle object vibrations (a). We propose an event-based subtle motion estimation method that leverages the event stream and a reference frame as input (b). The subtle motion field is extracted at each pixel location across time (c). This motion field can be further analyzed to extract the global motion spectrum and independent vibration modes (d). Each mode is characterized by a resonant frequency and a corresponding mode shape, which manifest as distinct peaks in the motion spectrum." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 454, + 219, + 547, + 350 + ], + "blocks": [ + { + "bbox": [ + 454, + 219, + 547, + 350 + ], + "lines": [ + { + "bbox": [ + 454, + 219, + 547, + 350 + ], + "spans": [ + { + "bbox": [ + 454, + 219, + 547, + 350 + ], + "type": "image", + "image_path": "92c97e841170be6ade1bd30cc895c9c0933f399b26f34b6821aff66712ef12e8.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 476, + 350, + 533, + 360 + ], + "lines": [ + { + "bbox": [ + 476, + 350, + 533, + 360 + ], + "spans": [ + { + "bbox": [ + 476, + 350, + 533, + 360 + ], + "type": "text", + "content": "Motion Spectrum" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 151, + 436, + 200, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 436, + 200, + 448 + ], + "spans": [ + { + "bbox": [ + 151, + 436, + 200, + 448 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 54, + 463, + 297, + 680 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 463, + 297, + 680 + ], + "spans": [ + { + "bbox": [ + 54, + 463, + 297, + 680 + ], + "type": "text", + "content": "Visual vibrometry has emerged as a powerful technique for remote acquisition of audio and the physical properties of materials. To capture high-frequency vibrations, frame-based approaches often require a high-speed video camera and bright lighting to compensate for the short exposure time. In this paper, we introduce event-based visual vibrometry, a new high-speed visual vibration sensing method using an event camera. By leveraging the high temporal resolution and low bandwidth characteristics of event cameras, event-based visual vibrometry enables high-speed vibration sensing under ambient lighting conditions with improved data efficiency. Specifically, we leverage a hybrid camera system and propose an event-based subtle motion estimation framework that integrates an optimization-based approach based on the event generation model and a motion refinement network. We demonstrate our method by capturing vibration caused by audio sources and estimating material properties for various objects." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 436, + 394, + 448 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 436, + 394, + 448 + ], + "spans": [ + { + "bbox": [ + 315, + 436, + 394, + 448 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 312, + 459, + 555, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 459, + 555, + 579 + ], + "spans": [ + { + "bbox": [ + 312, + 459, + 555, + 579 + ], + "type": "text", + "content": "Vibrations are pervasive phenomena that arise from a wide range of sources, including engines, percussive impacts, and musical instruments. These vibrations typically exhibit small amplitudes and a broad spectrum of frequencies, spanning from a few Hz to MHz, and carry an immense amount of information. Vibration analysis is an established tool in various engineering and scientific fields, including the remote acquisition of audio signals [8], the monitoring of human heart rate [38, 56], and the estimation of material properties [2, 9, 13]." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 582, + 556, + 715 + ], + "type": "text", + "content": "Various devices have been used to measure and analyze vibrations. Among these, contact sensors [48] and laser vibrometers [11] are two widely used traditional vibration sensors, both of which are specialized instruments and only measure the vibration of a single surface point. Recently, much progress has been made on visual vibrometry [2, 4, 9, 13], an emerging technique that utilizes general-purpose video cameras for vibration measurement. In contrast to traditional vibration sensors, video-based approaches enable spatially dense measurements of surface motion. However, the acquisition of high-frequency vi" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "spans": [ + { + "bbox": [ + 70, + 702, + 149, + 712 + ], + "type": "text", + "content": "* Corresponding author" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24666" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "briations requires high-speed cameras coupled with bright lighting to compensate for the short exposure time, which is impractical in many real-world applications. Additionally, high-speed cameras demand substantial bandwidth requirements. To eliminate the need for high-speed cameras, Sheinin et al. [43] propose a dual-shutter system combined with active speckle-based techniques. Nevertheless, the nature of speckle-based vibration imaging restricts its applicability and confines vibration sampling to only several surface points. In short, frame-based visual vibrometry encounters challenges in efficiently capturing high-frequency vibrations while maintaining bandwidth efficiency and measurement convenience." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 230, + 295, + 397 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 295, + 397 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 295, + 397 + ], + "type": "text", + "content": "Event cameras [14, 25], characterized by their high temporal resolution, high dynamic range, and low bandwidth requirements, have emerged as a promising solution for high-speed sensing [16]. Unlike conventional cameras, event cameras asynchronously record per-pixel logarithmic scene radiance changes. This signal-triggering mechanism is particularly advantageous for vibration sensing, where radiance changes occur only in a small fraction of pixels, leading to a significantly reduced data rate. The high temporal resolution and dynamic range inherent to event cameras enable the measurement of high-frequency vibrations under ambient lighting. These characteristics position event cameras as potential alternatives to address the constraints of frame-based visual vibrometry." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 400, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 295, + 639 + ], + "type": "text", + "content": "Despite the potential of event cameras, estimating subtle motion from event data remains an open problem. Firstly, although event cameras have been applied to small motion scenarios for motion magnification [7], this approach focuses on visualizing small motions and adopts an implicit motion representation. However, the explicit estimation of subtle motion is essential in visual vibrometry for vibration analysis. Secondly, adapting existing event-based motion estimation techniques to vibration scenarios presents significant difficulties. Existing event-based optical flow estimation methods [19, 37, 47, 58] generally assume continuous apparent motion of visual patterns, which is violated by the oscillatory nature of vibrations. Events caused by vibrations are primarily localized along object edges, and this spatial sparsity hinders the effectiveness of event alignment techniques like Contrast Maximization [37, 47]. Moreover, in comparison to general scenarios, vibrations tend to produce significantly fewer events. Estimating subtle motion from such sparse event data is highly vulnerable to noise inherent in event cameras [23]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "In this paper, we propose event-based visual vibrometry as an efficient and practical solution for vibration measurement and analysis. To address the challenges mentioned above, we employ an Event-RGB hybrid camera system [7, 20, 55] (Fig. 1(a)), which combines event streams with spatial texture information from conventional frames. We pro" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 275 + ], + "type": "text", + "content": "pose an event-based subtle motion estimation framework that takes event streams and a single reference frame as inputs (Fig. 1(b)). Initially, we employ an optimization-based approach grounded in the event generation model [21, 36] to derive coarse motion estimates within short temporal windows. The sparsity of event data restricts the validity of these coarse estimates to only event-triggered regions, while simultaneously introducing undesirable smoothing effects. Additionally, quantization errors in event cameras compromise the accuracy of motion amplitude estimation. To address these issues, we adopt a motion refinement network that aggregates historical motion information to enhance the coarse motion estimation and obtain long-range vibration displacements (Fig. 1(c)). The refined motion fields across time can subsequently be analyzed (Fig. 1(d)) for various visual vibrometry applications. The key contributions of our work are summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 276, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 276, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 276, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 314, + 276, + 553, + 312 + ], + "type": "text", + "content": "- a bandwidth-efficient visual vibrometry approach that leverages a hybrid camera system to measure high-frequency vibration under ambient lighting conditions;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 312, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 312, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 314, + 312, + 553, + 346 + ], + "type": "text", + "content": "- an event-based subtle motion estimation framework that integrates a model-based optimization approach and a motion refinement network;" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 314, + 348, + 553, + 384 + ], + "type": "text", + "content": "- a thorough evaluation on both synthetic data and real-world vibrations caused by various sources, demonstrating the effectiveness of event-based visual vibrometry." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 396, + 405, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 396, + 405, + 408 + ], + "spans": [ + { + "bbox": [ + 314, + 396, + 405, + 408 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 421, + 553, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 553, + 612 + ], + "type": "text", + "content": "Traditional vibration sensors. This section reviews two prevalent traditional vibration measurement devices: contact sensors and active sensors. Contact sensors operate through direct physical contact with the target object [48]. Accelerometers and piezoelectric pickups are representative examples of contact sensors that respond to acceleration. Among active sensors, laser Doppler vibrometers (LDVs) are the most common type used for vibration measurement [11, 41]. LDVs operate by projecting a laser beam onto the target object and calculating the surface velocity from the phase shift in the reflected light. LDVs have been used to find defects in composite materials [3, 5], examine the integrity of building structures [40], and modal analysis [30, 31]. Both contact sensors and laser vibrometers are typically limited to measuring vibrations at a single point on the surface." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 553, + 713 + ], + "type": "text", + "content": "Frame-based visual vibrometry. Recent studies on video motion magnification [53, 54, 56] have demonstrated the capability to detect subtle motions captured in videos. These methods typically exploit spatial phase variations of the complex steerable pyramid [39, 49, 50] to estimate small local motions. Chen et al. [6] applied this phase-based subtle motion estimation method to quantify the vibration modes of pipes and cantilever beams. Davis" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24667" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 240 + ], + "type": "text", + "content": "et al. [8] analyzed object vibrations induced by acoustic excitation to reconstruct the original sound. Davis et al. [9] employed the motion spectra extracted from video to estimate the physical properties of objects. Feng et al. [13] extended this methodology to heterogeneous properties and proposed a physics-constrained optimization approach. Capturing high-frequency vibrations requires expensive 2D high-speed cameras. To address this cost, some works [1, 57] utilize fast 1D sensors for high-frequency acquisition, but they can only measure vibrations along one dimension. Sheinin et al. [43] proposed a dual-shutter system with high-frequency sensing capabilities, integrated with active speckle-based techniques that optically magnify small-amplitude vibrations through laser illumination." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 244, + 298, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 244, + 298, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 244, + 298, + 508 + ], + "type": "text", + "content": "Event-based vibration measurement. Recent studies have investigated the application of event cameras for vibration measurement [10, 29, 33]. Dorn et al. [10] adapted the phase-based subtle motion extraction method to event data, assuming that each event contributes an equal phase change. Subsequently, they developed an event-based structural monitoring method. Shi et al. [44] incorporated laser active illumination to enhance event-based vibration measurements and introduced a frequency estimation technique based on Gaussian mixture distributions. Niwa et al. [34] presented a non-contact audio recovery setup using an event camera, employing the phase-based approach proposed in [10]. Howard et al. [22] implemented a similar setup, but recovered audio signals by analyzing the zero-crossings of pixels. Chen et al. [7] developed an event-based motion magnification method for amplifying and visualizing subtle motions. However, this approach does not directly estimate subtle motions; instead, motion information is implicitly derived from event signals via neural network processing. In this work, we focus on passive capture and estimation of subtle vibrations through the synergistic use of event signals and conventional frames." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 519, + 111, + 532 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 111, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 111, + 532 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 539, + 180, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 539, + 180, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 539, + 180, + 551 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "text", + "content": "Vibration modal analysis. The vibrations of objects are often well-approximated by linear systems. In modal analysis, a solid object is modeled as a system of point masses interconnected by springs [42]. In this discretized model, the mass matrix " + }, + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "text", + "content": " and the stiffness matrix " + }, + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\mathbf{K}" + }, + { + "bbox": [ + 55, + 560, + 296, + 657 + ], + "type": "text", + "content": " describe mass concentrated and stiffness between each pair of degrees of freedom (DOF), respectively. The equation of motion for this system is given by:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 141, + 667, + 295, + 679 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 667, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 141, + 667, + 295, + 679 + ], + "type": "interline_equation", + "content": "\\mathbf {M} \\ddot {x} + \\mathbf {K} x = 0, \\tag {1}", + "image_path": "036ad03b740abf1478a07dbb38912ac795114711557af740e64309d7f24fc4e5.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\ddot{x}" + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": " denote the displacement and acceleration of each DOF. The object's vibrations exhibit resonant frequen" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 541, + 84 + ], + "type": "text", + "content": "cies, which correspond to the eigenvalues of the system:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 399, + 93, + 555, + 107 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 93, + 555, + 107 + ], + "spans": [ + { + "bbox": [ + 399, + 93, + 555, + 107 + ], + "type": "interline_equation", + "content": "\\mathbf {K} \\mathbf {u} _ {i} = \\omega_ {i} ^ {2} \\mathbf {M} \\mathbf {u} _ {i}, \\tag {2}", + "image_path": "f4e7d3b8099a965dd4478c446590e64222c3af4176ff7da1743a05100a37cc1b.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "spans": [ + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "inline_equation", + "content": "\\omega_{i}" + }, + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "text", + "content": " represents a resonant frequency and the corresponding eigenvector " + }, + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_i" + }, + { + "bbox": [ + 313, + 117, + 553, + 165 + ], + "type": "text", + "content": " describes the mode shape. To make the theory of vibration accessible to the computer vision audience, we highlight the following key concepts:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 167, + 555, + 239 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 313, + 167, + 555, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 167, + 555, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 167, + 555, + 203 + ], + "type": "text", + "content": "- Mode Shapes " + }, + { + "bbox": [ + 313, + 167, + 555, + 203 + ], + "type": "inline_equation", + "content": "(\\mathbf{u}_1\\dots \\mathbf{u}_n)" + }, + { + "bbox": [ + 313, + 167, + 555, + 203 + ], + "type": "text", + "content": ": Each mode shape represents a distinct pattern of vibration for the object. The set of mode shapes forms an orthogonal basis for the vibration." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 203, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 203, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 203, + 553, + 239 + ], + "type": "text", + "content": "- Resonant Frequencies " + }, + { + "bbox": [ + 313, + 203, + 553, + 239 + ], + "type": "inline_equation", + "content": "(\\omega_{1}\\dots \\omega_{n})" + }, + { + "bbox": [ + 313, + 203, + 553, + 239 + ], + "type": "text", + "content": " : Each mode is associated with a resonant frequency. Resonant frequencies are spatially invariant across the object's surface." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 241, + 554, + 337 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 337 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 337 + ], + "type": "text", + "content": "The global motion power spectrum for an object can be computed by averaging the power spectra of local motions extracted at each pixel location. The modes of vibrations manifest as peaks in the motion power spectrum (Fig. 1(d)). Estimating mechanical properties from vibrations is based on the insight that variations in mechanical properties induce changes in the resonant frequencies and mode shapes for a fixed geometry." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "content": "Event formation model. When the logarithmic change of brightness at pixel " + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "content": " exceeds a specific threshold " + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "content": ", an event signal " + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "inline_equation", + "content": "e_k = (\\mathbf{x}_k, t_k, p_k)" + }, + { + "bbox": [ + 313, + 341, + 554, + 389 + ], + "type": "text", + "content": " will be triggered:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 399, + 553, + 412 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 399, + 553, + 412 + ], + "spans": [ + { + "bbox": [ + 359, + 399, + 553, + 412 + ], + "type": "interline_equation", + "content": "\\left| \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k}\\right) - \\mathbf {L} \\left(\\mathbf {x} _ {k}, t _ {k} - \\Delta t _ {k}\\right) \\right| \\geq C, \\tag {3}", + "image_path": "2e8df9d395ee1b5734bc0e4272e743678c2474e14d656b9b79b3aa0e8abf5905.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "p_k \\in \\{1, -1\\}" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " is the polarity of the brightness change, " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\mathbf{x}_k, t_k)" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " denotes the logarithmic brightness at pixel " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " and time " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "t_k - \\Delta t_k" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " is the time at which the preceding event was triggered at pixel " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": ". Let " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "\\mathcal{E} = \\{e_k\\}_{k=1}^{N_e}" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " denote a set of " + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "inline_equation", + "content": "N_e" + }, + { + "bbox": [ + 313, + 421, + 554, + 506 + ], + "type": "text", + "content": " events triggered within a short time interval, the brightness change can be computed by summing their polarities:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 372, + 514, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 514, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 372, + 514, + 553, + 548 + ], + "type": "interline_equation", + "content": "\\Delta \\mathbf {L} (\\mathbf {x}) = \\sum_ {k = 1} ^ {N _ {e}} p _ {k} C \\delta \\left(\\mathbf {x} - \\mathbf {x} _ {k}\\right), \\tag {4}", + "image_path": "268b91f801bf65b51739e13defa62f46c754646bceba26ce094e7c8d917a7d27.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": "where the Dirac delta function " + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": " selects the pixel " + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_k" + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": ". Assuming brightness constancy over the interval, the brightness change can be approximated as the product of the brightness gradient " + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{L}" + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": " and the displacement " + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{x}" + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": ", where the displacement is a result of motion with velocity " + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 555, + 554, + 628 + ], + "type": "text", + "content": " on the image plane [18]:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 330, + 635, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 330, + 635, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 330, + 635, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\Delta \\widehat {\\mathbf {L}} (\\mathbf {x}) = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\Delta \\mathbf {x} = - \\nabla \\mathbf {L} (\\mathbf {x}) \\cdot \\mathbf {v} (\\mathbf {x}) \\Delta t. \\tag {5}", + "image_path": "4ef36ef887076167c08dd758dd6934b559174062ec0bd04fd3be738bee14d0e3.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 659, + 486, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 486, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 486, + 673 + ], + "type": "text", + "content": "3.2. Coarse Motion by Optimization" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "Given the sparsity of event signals in vibration scenarios, event alignment and correlation techniques [19, 37, 47], commonly used in previous methods, are not applicable." + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24668" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 136, + 148 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 136, + 148 + ], + "type": "image", + "image_path": "2621cadf4486f7003d5eb57b8e0f68b69baf85cf9464a2a83942c25bb944caf8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 73, + 150, + 123, + 159 + ], + "lines": [ + { + "bbox": [ + 73, + 150, + 123, + 159 + ], + "spans": [ + { + "bbox": [ + 73, + 150, + 123, + 159 + ], + "type": "text", + "content": "(a) Events " + }, + { + "bbox": [ + 73, + 150, + 123, + 159 + ], + "type": "inline_equation", + "content": "(\\Delta \\mathbf{L})" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 137, + 70, + 214, + 148 + ], + "blocks": [ + { + "bbox": [ + 137, + 70, + 214, + 148 + ], + "lines": [ + { + "bbox": [ + 137, + 70, + 214, + 148 + ], + "spans": [ + { + "bbox": [ + 137, + 70, + 214, + 148 + ], + "type": "image", + "image_path": "88d3d31806afc2055e0fc1c53ec5e903ec718880e718561981bb9663e14ff6f3.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 147, + 150, + 205, + 159 + ], + "lines": [ + { + "bbox": [ + 147, + 150, + 205, + 159 + ], + "spans": [ + { + "bbox": [ + 147, + 150, + 205, + 159 + ], + "type": "text", + "content": "(b) Coarse Motion" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 70, + 293, + 148 + ], + "blocks": [ + { + "bbox": [ + 216, + 70, + 293, + 148 + ], + "lines": [ + { + "bbox": [ + 216, + 70, + 293, + 148 + ], + "spans": [ + { + "bbox": [ + 216, + 70, + 293, + 148 + ], + "type": "image", + "image_path": "461fa197b596e8905fdf7ded099d0160339c92f3ce1db7c57828a0f7f3c5b000.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 226, + 150, + 283, + 159 + ], + "lines": [ + { + "bbox": [ + 226, + 150, + 283, + 159 + ], + "spans": [ + { + "bbox": [ + 226, + 150, + 283, + 159 + ], + "type": "text", + "content": "(c) Predicted " + }, + { + "bbox": [ + 226, + 150, + 283, + 159 + ], + "type": "inline_equation", + "content": "(\\Delta \\hat{\\mathbf{L}})" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "type": "text", + "content": "Figure 2. An example (corresponding to the jelly cube in Fig. 1(b)) of coarse motion through optimization. The predicted brightness change (c), denoted as " + }, + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "type": "inline_equation", + "content": "\\Delta \\widehat{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "type": "text", + "content": " in Eq. (5) aligns well with the event frame (a), represented by " + }, + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{L}" + }, + { + "bbox": [ + 55, + 168, + 295, + 213 + ], + "type": "text", + "content": " in Eq. (4)." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 234, + 295, + 293 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 234, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 295, + 293 + ], + "type": "text", + "content": "Therefore, we employ a hybrid Event-RGB camera system, integrating high-speed event signals with spatially dense texture information from images, to facilitate subtle motion estimation. Our method takes event streams and a single reference frame as inputs." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "spans": [ + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "text", + "content": "Leveraging the event formation model, we formulate the problem of event-based subtle motion estimation as an optimization task, where we minimize the mismatch between the brightness change " + }, + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{L}" + }, + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "text", + "content": ", derived from event accumulation (as shown in Eq. (4)), and its prediction " + }, + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "inline_equation", + "content": "\\Delta \\widehat{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 294, + 296, + 366 + ], + "type": "text", + "content": ", obtained from moving brightness edges (as detailed in Eq. (5))." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "text", + "content": "Optimization objective. The composite loss function is a combination of data-fidelity term " + }, + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{data}}" + }, + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "text", + "content": " and regularization term " + }, + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{reg}}" + }, + { + "bbox": [ + 55, + 370, + 296, + 407 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 101, + 415, + 295, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 101, + 415, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 101, + 415, + 295, + 430 + ], + "type": "interline_equation", + "content": "E (\\mathbf {v}) = E _ {\\text {d a t a}} (\\mathbf {v}; \\mathcal {E}) + \\lambda E _ {\\text {r e g}} (\\nabla \\mathbf {v}), \\tag {6}", + "image_path": "4dadcd2c915a3c3802fcf1f23e7c7be42dffb5dde43f5ffd7fe0052e850536b1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 438, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 295, + 462 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 438, + 295, + 462 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 55, + 438, + 295, + 462 + ], + "type": "text", + "content": " serves as a balancing parameter, empirically set to 0.5 in our experiments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "content": "The data-fidelity term " + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{data}}" + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "content": " measures the difference between the observed brightness change " + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{L}" + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "content": " derived from event accumulation and its prediction " + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\Delta \\widehat{\\mathbf{L}}" + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "content": " based on our coarse motion estimation " + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 55, + 463, + 296, + 510 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 86, + 517, + 295, + 550 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 517, + 295, + 550 + ], + "spans": [ + { + "bbox": [ + 86, + 517, + 295, + 550 + ], + "type": "interline_equation", + "content": "E _ {\\mathrm {d a t a}} (\\mathbf {v}; \\mathcal {E}) = \\left\\| \\frac {\\Delta \\widehat {\\mathbf {L}}}{\\| \\Delta \\widehat {\\mathbf {L}} \\| _ {2}} \\cdot \\mathbf {M} - \\frac {\\Delta \\mathbf {L}}{\\| \\Delta \\mathbf {L} \\| _ {2}} \\right\\| _ {1}, \\tag {7}", + "image_path": "7a31220791c7b0714025edec522c800a56b2466b5106455a4bb4cf998a92d3d2.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{M}" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " is a mask selecting pixels with events triggered. Due to spatial-temporal variations in the contrast threshold " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " [23], we compute the difference between normalized brightness changes across the image plane to enhance the robustness of the optimization, following established methodologies [21, 46]. To analyze high-frequency vibrations, the subtle motion is estimated within a very short time interval (typically less than " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "0.5~ms" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "), during which very few event signals are triggered. It is challenging to accurately estimate flow in regions devoid of events, which often correspond to areas with low contrast or zero flow. To mitigate these inaccuracies, we only calculate the data-fidelity loss term at pixel locations with events triggered." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "We assume that flow vectors vary smoothly with sparse discontinuities occurring at object boundaries. The regularization term " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "E_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " encourages smoothness of the estimated motion " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " by minimizing the differences between neighboring pixels. This term is defined as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 372, + 140, + 553, + 154 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 140, + 553, + 154 + ], + "spans": [ + { + "bbox": [ + 372, + 140, + 553, + 154 + ], + "type": "interline_equation", + "content": "E _ {\\text {r e g}} (\\nabla \\mathbf {v}) = \\| \\omega (\\mathbf {x}) \\nabla \\mathbf {v} (\\mathbf {x}) \\| _ {1}, \\tag {8}", + "image_path": "e978ba436e000b197c99581a19b357ea1349c109c8d7cc9c91643510c2632361.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "text", + "content": " represents the smoothness weight for each pixel. Following prior research [36], the weight " + }, + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "inline_equation", + "content": "\\omega" + }, + { + "bbox": [ + 313, + 161, + 553, + 197 + ], + "type": "text", + "content": " is computed based on the gradients of the reference image:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 377, + 205, + 553, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 377, + 205, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 377, + 205, + 553, + 219 + ], + "type": "interline_equation", + "content": "\\omega (\\mathbf {x}) = \\exp (- \\mu | \\nabla \\mathbf {L} (\\mathbf {x}) |), \\tag {9}", + "image_path": "f6ae18c83a9e2fa12bd742f665dc04cc429d5e7f0c7b84f4ab67060fec5111d5.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "inline_equation", + "content": "\\nabla \\mathbf{L}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "text", + "content": " denotes the spatial gradient of the logarithmic brightness at pixel " + }, + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 313, + 225, + 553, + 250 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "content": "Optimization details. To enhance the convergence speed and robustness of the optimization process, we employ a patch-based optimization strategy implemented in a coarse-to-fine manner [47]. This strategy involves three resolution scales in a pyramid structure, with the coarsest patch size at " + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "content": " pixels and the finest at " + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "content": " pixels. Bilinear interpolation is utilized for upsampling between successive scales. " + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "content": " is initialized uniformly within the range " + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "inline_equation", + "content": "[-0.1, 0.1]" + }, + { + "bbox": [ + 313, + 255, + 555, + 409 + ], + "type": "text", + "content": " at the coarsest scale. For the subsequent scale, the initialization is derived from the optimization results of the preceding scale. We utilize the Adam optimizer [24] with a learning rate of 0.05, and the overall number of iterations is set to 40." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": "Analysis. An example of coarse motion estimation is presented in Fig. 2. From this example, it is observed that the predicted brightness change " + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{L}" + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": ", derived from the coarse motion (Fig. 2(b)), aligns well with the event frame (Fig. 2(a)), which corresponds to " + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "inline_equation", + "content": "\\Delta \\mathbf{L}" + }, + { + "bbox": [ + 313, + 414, + 554, + 558 + ], + "type": "text", + "content": " in Eq. (4). This demonstrates the efficacy of the coarse motion estimation. However, the coarse motion results manifest several limitations. First, due to the sparsity of events, the estimated coarse motion yields valid results only in regions where events are triggered, and it is susceptible to noise. Second, quantization errors in event cameras hinder the accurate recovery of motion amplitude in the coarse motion results." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 564, + 488, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 488, + 578 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 488, + 578 + ], + "type": "text", + "content": "3.3. Motion Refinement by Learning" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": "With the optimization-based approach described in Sec. 3.2, we obtain the coarse motion results, denoted as " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "f^c" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " for ease of presentation. While " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "f^c" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " enables the brightness change prediction " + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "inline_equation", + "content": "\\Delta \\hat{\\mathbf{L}}" + }, + { + "bbox": [ + 313, + 582, + 554, + 689 + ], + "type": "text", + "content": " to align with the event signals, it exhibits issues including excessive smoothness, susceptibility to noise, and inaccurate motion amplitude estimation. To address these problems, we propose EVibNet, a motion refinement network specifically designed to enhance the coarse motion results." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "The pipeline of EVibNet is shown in Fig. 3. The network takes event signals " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": ", a single reference frame " + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24669" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 74, + 550, + 223 + ], + "blocks": [ + { + "bbox": [ + 60, + 74, + 550, + 223 + ], + "lines": [ + { + "bbox": [ + 60, + 74, + 550, + 223 + ], + "spans": [ + { + "bbox": [ + 60, + 74, + 550, + 223 + ], + "type": "image", + "image_path": "9e114cc8f5467df240c7e27df4a23d3ac749f6788727b5ced6e05ee1fe492273.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": "Figure 3. Pipeline of our EVibNet. (Input: " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\varepsilon_{i}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ") Initially, features are extracted from the reference frame " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": " and events triggered within the time interval " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "[t_i,t_{i + 1}]" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ", respectively. Then the extracted features are fused with a fusion module " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ". The coarse motion obtained in Sec. 3.2 is fed into a motion encoder " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{M}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ", generating the motion feature " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_i,t_{i + 1}}^c" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ", which is further refined by the motion refine module " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{refine}}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ". Finally, the refined motion feature is accumulated to obtain the vibration displacement (Outout: " + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_0,t_i + 1}" + }, + { + "bbox": [ + 55, + 234, + 555, + 281 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "and coarse motion results " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "f^c" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " as inputs. Specifically, EVibNet comprises two principal components: 1) two modality-specific encoders " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " and a fusion module " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "; 2) a coarse motion refinement module " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{refine}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": " and a motion accumulation module " + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 55, + 300, + 295, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "spans": [ + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": "Feature extraction. For events " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " triggered in the time interval " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "[t_i, t_{i+1}]" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": ", we transform " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " into two distinct voxel grids " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i^+" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{E}_i^-" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " for positive and negative events respectively, following the procedure in [17]. Due to the modality gap between frames and events, we use two modality-specific encoders " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " to extract features from the reference frame " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathbf{I}_{\\mathrm{ref}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": " and events " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{E}_i" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": ", respectively. To solve the issue of inaccurate motion amplitude estimation caused by quantization error, we leverage the temporal structure inherent in the event data using a recurrent convolutional encoder. Specifically, a ConvLSTM [45] layer is integrated into each encoder stage of " + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 364, + 296, + 532 + ], + "type": "text", + "content": ". The feature extraction process for the reference frame and event stack is performed using the feature encoders as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 77, + 537, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 537, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 77, + 537, + 295, + 552 + ], + "type": "interline_equation", + "content": "\\mathbf {f} ^ {\\mathrm {L}} = \\mathcal {F} _ {\\mathrm {L}} \\left(\\mathbf {I} _ {\\text {r e f}}\\right), \\quad \\mathbf {f} _ {i} ^ {\\mathrm {E}} = \\mathcal {F} _ {\\mathrm {E}} \\left(\\mathbf {E} _ {i} ^ {+}, \\mathbf {E} _ {i} ^ {-}, h _ {i - 1}\\right), \\tag {10}", + "image_path": "1a14ec7e0bbd5146bd7e9578cb776888993a10bfc818915595a1312275f60beb.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{L}},\\mathbf{f}_i^{\\mathrm{E}}\\in \\mathbb{R}^{h\\times w\\times D}" + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": ", the feature resolution " + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "inline_equation", + "content": "(h,w)" + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "inline_equation", + "content": "1 / 4" + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": " of the original image resolution " + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "inline_equation", + "content": "(H,W)" + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "inline_equation", + "content": "h_{i - 1}" + }, + { + "bbox": [ + 55, + 556, + 295, + 593 + ], + "type": "text", + "content": " refers to the hidden states in the recurrent event encoder." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "Given that event signals are triggered in only a few regions when capturing vibrating objects, it is crucial to leverage texture features extracted from the reference frame to regulate the motion features extracted from the events. Following [51], we employ a cross-modal attention fusion block, denoted as " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", to adaptively fuse the information from these two modalities, yielding the fused feature " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{H}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ". The cross-modal attention in " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{fuse}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " takes as input the queries " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{Q}^{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " derived from the event features " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{E}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": ", and the keys " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{K}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{V}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " derived from the image features " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{L}}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": "Motion refinement and accumulation. With the frame-event fused feature " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{H}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": ", we refine the coarse motion " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "f_{t_i,t_{i + 1}}^c" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": " through a motion refinement module, denoted as " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{refine}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": ". The coarse motion " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "f_{t_i,t_{i + 1}}^c" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": " is initially encoded into coarse motion features " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_i,t_{i + 1}}^c" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": " using a motion encoder " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{M}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": ". Subsequently, the motion information embedded in " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_i^{\\mathrm{H}}" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": " is leveraged to refine " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_i,t_{i + 1}}^c" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": ", producing the refined motion feature " + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_i,t_{i + 1}}^r" + }, + { + "bbox": [ + 313, + 300, + 555, + 399 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 372, + 399, + 553, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 399, + 553, + 414 + ], + "spans": [ + { + "bbox": [ + 372, + 399, + 553, + 414 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r} = \\mathcal {F} _ {\\text {r e f i n e}} \\left(\\mathbf {f} _ {i} ^ {\\mathrm {H}}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {c}\\right). \\tag {11}", + "image_path": "acedf657866b70efb13ca7931b84a703dcd3d7e4484d957a86f9fdd5cfa941c7.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "text", + "content": "Following the refinement of coarse motion, we further accumulate short-term motion into long-range displacement. This approach is motivated by two factors: First, motion within a very short time interval is typically minimal and thus more susceptible to noise. Second, estimating the displacement of points on vibrating objects relative to a reference state facilitates subsequent analysis. However, direct accumulation of motion at each time step results in significant accumulation errors. Therefore, we integrate short-term motion using a motion accumulation module " + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "text", + "content": ". The accumulated motion feature " + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_0,t_i}" + }, + { + "bbox": [ + 313, + 416, + 555, + 559 + ], + "type": "text", + "content": " is computed as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 364, + 559, + 553, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 559, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 364, + 559, + 553, + 574 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {t _ {0}, t _ {i + 1}} ^ {a c c} = \\mathcal {F} _ {\\text {a c c}} \\left(\\mathbf {f} _ {t _ {0}, t _ {i}} ^ {a c c}, \\mathbf {f} _ {t _ {i}, t _ {i + 1}} ^ {r}, \\mathbf {f} ^ {\\mathrm {L}}\\right), \\tag {12}", + "image_path": "d195962411c5c42beeacc352c8aa94333797217c75bc55532040b107874a3c80.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "spans": [ + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{f}_{t_0,t_i}^{acc}" + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": " is derived from the preceding time interval, and " + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathbf{f}^{\\mathrm{L}}" + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": " serves as the context feature to guide the motion accumulation. For simplicity, " + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{acc}}" + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": " is implemented as a simple RNN comprising two ConvNext blocks [27]. Finally, the accumulated motion feature is decoded into the target displacement " + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "inline_equation", + "content": "f_{t_0,t_{i + 1}}" + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": " using a motion decoder " + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\mathrm{D}}" + }, + { + "bbox": [ + 313, + 576, + 555, + 650 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 655, + 448, + 668 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 655, + 448, + 668 + ], + "spans": [ + { + "bbox": [ + 314, + 655, + 448, + 668 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "Loss functions. Our loss functions are derived from prior work on optical flow estimation [52]. The network is supervised using the " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "l_{1}" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " distance between the predicted and" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24670" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "text", + "content": "ground truth flow across the entire sequence of predictions, " + }, + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "inline_equation", + "content": "\\{f_{t_0,t_1},\\dots ,f_{t_0,t_N}\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "text", + "content": " . Given the ground truth flow sequence " + }, + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "inline_equation", + "content": "\\{f_{t_0,t_1}^{gt},\\dots ,f_{t_0,t_N}^{gt}\\}" + }, + { + "bbox": [ + 55, + 72, + 294, + 110 + ], + "type": "text", + "content": " , the loss is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 121, + 117, + 294, + 149 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 117, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 121, + 117, + 294, + 149 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\sum_ {i = 1} ^ {N} \\| f _ {t _ {0}, t _ {i}} ^ {g t} - f _ {t _ {0}, t _ {i}} \\| _ {1}. \\tag {13}", + "image_path": "5e7b630a734a1ebb112c70bd8520c52fc64f7cf9924ac09bf052c4bf8b1d2299.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 327 + ], + "type": "text", + "content": "Synthetic dataset. Obtaining ground truth data for real-world subtle motion is challenging. Therefore, following previous studies [7, 35], we simulate subtle motion by composing foreground elements onto a background. We sample images from the MS COCO dataset [26] for the background and use segmented objects from the PASCAL VOC dataset [12] for the foreground. The magnitude and direction of motion for both the background and each foreground object are generated randomly. To simulate sub-pixel motions, we initially generate high-resolution videos, where the motion appears larger, and subsequently downsample each frame to the desired resolution. After obtaining the videos, we use the event simulator V2E [23] to simulate event data from downsampled videos." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 332, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 332, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 332, + 296, + 416 + ], + "type": "text", + "content": "Training details. We implement our approach using the PyTorch framework and conduct all experiments on a single NVIDIA GeForce RTX 3090 GPU. During training, we use the AdamW optimizer [28] and configure the batch size to 4. The EVibNet is trained together for 20 epochs. The initial learning rate is set to " + }, + { + "bbox": [ + 55, + 332, + 296, + 416 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 55, + 332, + 296, + 416 + ], + "type": "text", + "content": ", and a cosine annealing learning rate scheduling strategy is employed." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 420, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 420, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 420, + 296, + 492 + ], + "type": "text", + "content": "Hybrid camera system. Our hybrid camera system consists of a machine vision camera (HIKVISION MV-CA050-12UC) and an event camera (PROPHESEE GEN4.1), which are co-aligned using a beam splitter (Thorlabs CCM1-BS013). We utilize the calibration method proposed in [32] to achieve pixel alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 617 + ], + "type": "text", + "content": "In this section, we first evaluate the sub-pixel motion estimation accuracy of the proposed method on the synthetic dataset in Sec. 4.1. Then we validate our method's ability to detect high-frequency vibrations by capturing and replaying the vibration caused by audio sources in Sec. 4.2. Finally, we demonstrate our method by estimating both homogeneous and heterogeneous material properties of objects with known geometry in Sec. 4.3." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 624, + 228, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 228, + 636 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 228, + 636 + ], + "type": "text", + "content": "4.1. Evaluation on Synthetic Dataset" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "We compare our method to the phase-based subtle motion estimation approach (denoted as RGBPhase) that is widely used in frame-based visual vibrometry. This approach computes local motion signals from phase shifts within a complex steerable pyramid (CSP) [39, 49, 50]. For the comparison with high-speed videos, we apply RGBPhase to the" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 123, + 553, + 203 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": "Table 1. Motion estimation comparison on the synthetic dataset. " + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "inline_equation", + "content": "\\mathbf{E}" + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "inline_equation", + "content": "\\mathbf{F}" + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": " denote events and frames, respectively. " + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "inline_equation", + "content": "\\downarrow" + }, + { + "bbox": [ + 313, + 71, + 553, + 114 + ], + "type": "text", + "content": " indicates the lower, the better throughout this paper. The best performances are highlighted in bold." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 123, + 553, + 203 + ], + "lines": [ + { + "bbox": [ + 317, + 123, + 553, + 203 + ], + "spans": [ + { + "bbox": [ + 317, + 123, + 553, + 203 + ], + "type": "table", + "html": "
MethodsInputsEPE ↓% Out ↓
EV-FlowNet [37]E0.180547.95
MCM [47]E0.451166.87
RGBPhase [15]F (960Hz)0.283354.78
EMM [7]+RAFT [52]E + F (30Hz)0.219047.63
OursE + F (single)0.083421.29
", + "image_path": "d711477ebe5b7af9989866481aba75177bec1ce996a850c085f4f78bc9e38164.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 222, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 555, + 354 + ], + "type": "text", + "content": "original video sequences used to simulate event streams in our synthetic dataset. Furthermore, we compare our method with two representative event-based optical flow estimation methods: Multi-reference Contrast Maximization (MCM) [47] and the recurrent version of EV-FlowNet [37, 58]. To validate the superior suitability of our method for subtle motion estimation compared to the event-based motion magnification method (EMM) [7], we retrain EMM using our synthetic dataset and leverage an off-the-shelf optical flow estimation method RAFT [52] to estimate optical flow from the motion-magnified videos generated by EMM." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "text", + "content": "We adopt the average endpoint error (EPE) and the percentage of pixels with EPE " + }, + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "inline_equation", + "content": ">0.1" + }, + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "text", + "content": " px (denoted as " + }, + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "inline_equation", + "content": "\\%" + }, + { + "bbox": [ + 313, + 354, + 556, + 534 + ], + "type": "text", + "content": " Out\") as evaluation metrics. The quantitative comparison results are presented in Tab. 1. The results demonstrate that our method substantially outperforms previous event-based optical flow estimation techniques, primarily due to its specific design for subtle motion estimation. Notably, RGBPhase exhibits unsatisfactory performance on our synthetic dataset, likely attributable to the noise in local phase information encoded in the CSP. The inferior performance of EMM + RAFT compared to our method suggests that subtle motion estimation cannot be effectively achieved by simply cascading an optical flow method with EMM. This performance gap may be attributed to the blurring artifacts often introduced by motion magnification techniques." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 540, + 449, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 449, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 449, + 553 + ], + "type": "text", + "content": "4.2. Capturing audio signals" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "text", + "content": "To validate the efficacy of our method in real-world high-frequency vibration sensing, we apply it to remote acquisition of audio signals. In the experimental setup depicted in Fig. 4, we capture the membrane vibrations of a speaker playing a linear up-chirp signal ranging from 0 to " + }, + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "inline_equation", + "content": "1000\\mathrm{Hz}" + }, + { + "bbox": [ + 313, + 557, + 556, + 714 + ], + "type": "text", + "content": ". Concurrently, we obtain reference audio recordings using a microphone. With the subtle motion extracted by our method, the 1D audio signal is reconstructed by averaging the local motions across all pixel locations and orientations [8]. We compare our method with an event-based sound recovery method (EBVM) proposed by Niwa et al. [34]. By integrating events and frames, our method estimates subtle motion more accurately than the phase-based ap" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24671" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 72, + 170, + 142 + ], + "blocks": [ + { + "bbox": [ + 62, + 72, + 170, + 142 + ], + "lines": [ + { + "bbox": [ + 62, + 72, + 170, + 142 + ], + "spans": [ + { + "bbox": [ + 62, + 72, + 170, + 142 + ], + "type": "image", + "image_path": "e845338d9e6893f378e543826ed912fa86d609c191fbf170ff91bac7e5e308c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 171, + 72, + 291, + 143 + ], + "blocks": [ + { + "bbox": [ + 171, + 72, + 291, + 143 + ], + "lines": [ + { + "bbox": [ + 171, + 72, + 291, + 143 + ], + "spans": [ + { + "bbox": [ + 171, + 72, + 291, + 143 + ], + "type": "image", + "image_path": "b8628e31ac83d5ef426f719964cf20594135941fef79d2e245e54ae6f00fc430.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 143, + 267, + 153 + ], + "lines": [ + { + "bbox": [ + 211, + 143, + 267, + 153 + ], + "spans": [ + { + "bbox": [ + 211, + 143, + 267, + 153 + ], + "type": "text", + "content": "(b) Microphone" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 58, + 153, + 174, + 226 + ], + "blocks": [ + { + "bbox": [ + 78, + 143, + 160, + 153 + ], + "lines": [ + { + "bbox": [ + 78, + 143, + 160, + 153 + ], + "spans": [ + { + "bbox": [ + 78, + 143, + 160, + 153 + ], + "type": "text", + "content": "(a) Experimental setup" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 153, + 174, + 226 + ], + "lines": [ + { + "bbox": [ + 58, + 153, + 174, + 226 + ], + "spans": [ + { + "bbox": [ + 58, + 153, + 174, + 226 + ], + "type": "image", + "image_path": "0a2eec5cb4e11dc7073874c3907988bc455360449ae602c159123c11e34db945.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 226, + 135, + 236 + ], + "lines": [ + { + "bbox": [ + 99, + 226, + 135, + 236 + ], + "spans": [ + { + "bbox": [ + 99, + 226, + 135, + 236 + ], + "type": "text", + "content": "(c) EBVM" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 247, + 296, + 301 + ], + "lines": [ + { + "bbox": [ + 55, + 247, + 296, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 247, + 296, + 301 + ], + "type": "text", + "content": "Figure 4. Speaker membrane experiments. (a) A speaker is playing an up-chirp signal ranging from 0 to " + }, + { + "bbox": [ + 55, + 247, + 296, + 301 + ], + "type": "inline_equation", + "content": "1000\\mathrm{Hz}" + }, + { + "bbox": [ + 55, + 247, + 296, + 301 + ], + "type": "text", + "content": ". (b) The spectrogram of the source sound recorded by a microphone. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 176, + 153, + 292, + 226 + ], + "blocks": [ + { + "bbox": [ + 176, + 153, + 292, + 226 + ], + "lines": [ + { + "bbox": [ + 176, + 153, + 292, + 226 + ], + "spans": [ + { + "bbox": [ + 176, + 153, + 292, + 226 + ], + "type": "image", + "image_path": "c3459ef65fb27c831e4aca8b388935a37e265b498fd74fc4bb332174e94c132e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 223, + 226, + 253, + 236 + ], + "lines": [ + { + "bbox": [ + 223, + 226, + 253, + 236 + ], + "spans": [ + { + "bbox": [ + 223, + 226, + 253, + 236 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 58, + 367, + 296, + 456 + ], + "blocks": [ + { + "bbox": [ + 55, + 315, + 296, + 359 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 296, + 359 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 296, + 359 + ], + "type": "text", + "content": "Table 2. Percent error in estimating the Young's Modulus for each rod. The best performances are highlighted in bold. The content in each cell refers to the results for rods clamped to a length of 35 cm and 48 cm respectively." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 367, + 296, + 456 + ], + "lines": [ + { + "bbox": [ + 58, + 367, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 58, + 367, + 296, + 456 + ], + "type": "table", + "html": "
% Error ↓LightingAluminumSteelCopper
Video [9]low10.01/10.42-11.64/-9.28-9.40/4.64
medium4.13/-3.70-10.72/-7.62-4.41/-2.70
high-1.45/-1.99-7.94/-5.96-3.14/-2.10
Ourslow2.90/3.21-7.61/-6.58-5.65/2.28
medium-1.26/1.64-6.77/-5.08-3.39/-2.07
high0.39/-1.47-5.93/-5.08-2.25/-2.07
", + "image_path": "11f6541417c2378b7fad369a65ed9fcc06ea148be5ae96aeae357be132974403.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 476, + 295, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 476, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 476, + 295, + 523 + ], + "type": "text", + "content": "proach adopted in EBVM. The experimental results demonstrate significantly reduced noise levels and enhanced high-frequency reconstruction fidelity compared to the baseline method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 621 + ], + "type": "text", + "content": "In Fig. 5, we conduct an experiment similar to Davis et al. [8]. A speaker playing the \"MaryMIDI\" audio excites the chip bag. We measure the vibration of a chip bag and recover the audio. The spectrograms of recovered audio reveal superior waveform fidelity achieved by our method. These results demonstrate our method's robustness and practical applicability in complex real-world vibration sensing scenarios." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 630, + 222, + 643 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 222, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 222, + 643 + ], + "type": "text", + "content": "4.3. Estimating material properties" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 648, + 295, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 648, + 295, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 648, + 295, + 673 + ], + "type": "text", + "content": "We apply our method to estimate homogeneous and heterogeneous material properties from vibrations." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Homogeneous material properties. Considering objects with homogeneous material properties, it can be obtained from Eq. (2) that different objects with the same ge" + } + ] + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 430, + 141 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 430, + 141 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 430, + 141 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 430, + 141 + ], + "type": "image", + "image_path": "bcb04966c96a449a84c6d96aa65c5705a971bddbd41285994eedcf2c503205cc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 317, + 153, + 433, + 226 + ], + "blocks": [ + { + "bbox": [ + 333, + 143, + 417, + 153 + ], + "lines": [ + { + "bbox": [ + 333, + 143, + 417, + 153 + ], + "spans": [ + { + "bbox": [ + 333, + 143, + 417, + 153 + ], + "type": "text", + "content": "(a) Experimental setup" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 153, + 433, + 226 + ], + "lines": [ + { + "bbox": [ + 317, + 153, + 433, + 226 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 433, + 226 + ], + "type": "image", + "image_path": "7cfa53b66d52518adf7fcebcae6fa5384118495061dae17efd92ba87b3daff2e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 361, + 226, + 397, + 236 + ], + "lines": [ + { + "bbox": [ + 361, + 226, + 397, + 236 + ], + "spans": [ + { + "bbox": [ + 361, + 226, + 397, + 236 + ], + "type": "text", + "content": "(c) EBVM" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 245, + 555, + 300 + ], + "lines": [ + { + "bbox": [ + 313, + 245, + 555, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 245, + 555, + 300 + ], + "type": "text", + "content": "Figure 5. Audio recovery from a bag of chips. (a) Sound from an audio source (speaker) excites the chip bag. (b) The spectrogram of the input signal sent to the speaker. (c) The spectrogram of the recovered audio from EBVM [34]. (d) The spectrogram of our recovered sound." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 430, + 72, + 552, + 143 + ], + "blocks": [ + { + "bbox": [ + 430, + 72, + 552, + 143 + ], + "lines": [ + { + "bbox": [ + 430, + 72, + 552, + 143 + ], + "spans": [ + { + "bbox": [ + 430, + 72, + 552, + 143 + ], + "type": "image", + "image_path": "e2507dae63c5e301867152e273afcae7c8a02f506a4bc6892ffd30b94d2502de.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 471, + 143, + 527, + 154 + ], + "lines": [ + { + "bbox": [ + 471, + 143, + 527, + 154 + ], + "spans": [ + { + "bbox": [ + 471, + 143, + 527, + 154 + ], + "type": "text", + "content": "(b) Input signal" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 434, + 153, + 552, + 224 + ], + "blocks": [ + { + "bbox": [ + 434, + 153, + 552, + 224 + ], + "lines": [ + { + "bbox": [ + 434, + 153, + 552, + 224 + ], + "spans": [ + { + "bbox": [ + 434, + 153, + 552, + 224 + ], + "type": "image", + "image_path": "5e7e702c9b1a8a7654f8bf35644ecd7cbda2e4df2799a7056b3b99398185c1aa.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 481, + 226, + 512, + 236 + ], + "lines": [ + { + "bbox": [ + 481, + 226, + 512, + 236 + ], + "spans": [ + { + "bbox": [ + 481, + 226, + 512, + 236 + ], + "type": "text", + "content": "(d) Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 320, + 554, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 320, + 554, + 405 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 554, + 405 + ], + "type": "text", + "content": "ometry have an identical set of mode shapes, but their resonant frequencies scale proportionally to material properties [9]. Following Davis et al. [9], we apply our method to estimate the homogeneous material properties of various metal rods. The simple geometry of clamped rods makes their vibration well-studied [42]. The fundamental frequency " + }, + { + "bbox": [ + 313, + 320, + 554, + 405 + ], + "type": "inline_equation", + "content": "\\omega_{1}" + }, + { + "bbox": [ + 313, + 320, + 554, + 405 + ], + "type": "text", + "content": " (first resonant frequency) of a rod is given by:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 388, + 411, + 554, + 444 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 411, + 554, + 444 + ], + "spans": [ + { + "bbox": [ + 388, + 411, + 554, + 444 + ], + "type": "interline_equation", + "content": "\\omega_ {1} = 0. 1 3 9 9 \\frac {d}{L ^ {2}} \\sqrt {\\frac {E}{\\rho}}, \\tag {14}", + "image_path": "1695a87f2d44b67d36414ee511a3393eeb7433a213ce8f88c5bda7849c276613.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": " is the diameter of the rod, " + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": " is the length, " + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": " is the Young's modulus, and " + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "inline_equation", + "content": "\\rho" + }, + { + "bbox": [ + 313, + 450, + 554, + 533 + ], + "type": "text", + "content": " is the density. While length, diameter, and density can be easily measured, Young's modulus is often measured with a tensile test, which will damage the object. In this experiment, we try to estimate the Young's modulus of the rods by finding their fundamental frequency, which is the highest peak in the motion spectrum." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 555, + 713 + ], + "type": "text", + "content": "We test rods with three different metals: aluminum, steel, and copper. Each rod is tested at two lengths (35 cm and 48 cm) and under three brightness conditions (100, 500, and 1500 lux). For comparison with frame-based visual vibrometry, we capture 1000 fps videos under identical settings. Resonant frequencies are global properties and invariant to the viewpoint, ensuring that estimation accuracy is not affected by the viewpoint. In Tab. 2, we compare the estimated Young's modulus values with those provided by the manufacturer. The results demonstrate that our method achieves more robust performance across varying lighting conditions. We further compare the data sizes of 1000 fps videos with signals captured by our imaging system in the rod experiments. Our method requires only a single image; consequently, the data size of the reference frame is negligible." + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "24672" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 293, + 190 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 293, + 190 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 293, + 190 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 293, + 190 + ], + "type": "image", + "image_path": "fae9ad03b961d72790633bf248f0322b93ee8d6acbea9d61fd33f58489bf1300.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 295, + 246 + ], + "type": "text", + "content": "Figure 6. Heterogenous material properties estimation result of the defect cube (a) and a homogeneous cube (b), the results are plotted with the same colormaps. Based on the results, we can easily differentiate between a cube with a defect and a homogeneous one." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "text", + "content": "gible. At the same 720p resolution, event cameras produce event streams at " + }, + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "inline_equation", + "content": "1 - 5\\mathrm{MB / s}" + }, + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "text", + "content": ", about " + }, + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 278, + 296, + 387 + ], + "type": "text", + "content": " of the compressed 1000 fps videos (H. 264 encoded). The discrepancy is more pronounced when considering the actual transmission bandwidth required for high-speed cameras. We also test existing event-based optical flow estimation methods. However, both EV-FlowNet [37] and MCM [47] struggle to accurately estimate the vibrations of the rods, and we can not identify reasonable resonant frequencies in the motion spectra." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 296, + 492 + ], + "type": "text", + "content": "Heterogeneous material properties. When considering objects with heterogeneous material properties, spatial inhomogeneities in these properties influence both the resonant frequencies and mode shapes. Based on the mathematical relationship between modes and material properties, Feng et al. [13] propose a physical-constrained optimization approach to estimate heterogeneous material properties from several image-space vibration modes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "To demonstrate the applicability of our method to visual vibration tomography, we experiment on two jelly cubes: one homogeneous and the other with an interior defect (a rubber doll embedded within the jelly cube). We capture data under an initial deformation condition and identify unique modes in the estimated motion fields, as shown in Fig. 1. For each cube, we infer spatially-varying Young's modulus and density values on a " + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "inline_equation", + "content": "10 \\times 10 \\times 10" + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": " hexahedral mesh with the optimization algorithm proposed by Feng et al. [13]. The reconstruction results are shown in Fig. 6. The reconstruction is obtained using only five unique image-space modes. The defective cube has a rubber doll on top, which is stiffer than the jelly cube. The results Fig. 6(a) show the existence of the defect, and clearly differ from that of the homogeneous cube. Due to the high damping of the jelly cubes, the reconstruction results are not perfectly aligned with the data, but we can still easily differentiate between the two cubes." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 91, + 555, + 135 + ], + "blocks": [ + { + "bbox": [ + 314, + 71, + 555, + 83 + ], + "lines": [ + { + "bbox": [ + 314, + 71, + 555, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 555, + 83 + ], + "type": "text", + "content": "Table 3. Ablation study of the method design on the synthetic data." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 91, + 555, + 135 + ], + "lines": [ + { + "bbox": [ + 317, + 91, + 555, + 135 + ], + "spans": [ + { + "bbox": [ + 317, + 91, + 555, + 135 + ], + "type": "table", + "html": "
CoarseW/o cmW/o accW/o frameOurs
EPE ↓0.36320.09690.12990.17340.0834
% Out ↓51.4525.9136.9942.2521.29
", + "image_path": "5f05e698c4d39da279f29e2e59146dc332cc479b2cc61388678a5016a50a9da6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 157, + 413, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 157, + 413, + 168 + ], + "spans": [ + { + "bbox": [ + 314, + 157, + 413, + 168 + ], + "type": "text", + "content": "4.4. Ablation studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 176, + 555, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 555, + 344 + ], + "type": "text", + "content": "To validate the effectiveness and necessity of each part of our method, we conduct several ablation studies and show the results in Tab. 3. To assess the necessity of motion refinement using EVibNet, we test the performance of coarse motion (denoted as \"coarse\"). To verify the contribution of the coarse motion obtained from optimization, we remove it from the inputs of EVibNet (denoted as \"w/o cm\"). To demonstrate the effectiveness of the motion accumulation module, we directly accumulate motions at each time step (denoted as \"w/o acc\"). Finally, we show the necessity of the reference frame by removing it from inputs (denoted as \"w/o frame\"). As indicated in Tab. 3, our complete method achieves the lowest EPE and % Out, which demonstrates the contribution of each component of our method." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "spans": [ + { + "bbox": [ + 314, + 360, + 388, + 372 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 382, + 554, + 466 + ], + "type": "text", + "content": "We present a bandwidth-efficient approach for high-speed vibration sensing with an Event-RGB hybrid camera system. To extract subtle motion frame event data, we integrate an optimization-based approach with a learning-based motion refinement network. We demonstrate our method through the capture of vibrations induced by audio sources and the estimation of material properties." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 544 + ], + "type": "text", + "content": "Limitations. In the synthetic dataset, subtle motion is simulated by compositing foreground elements onto a background, which exhibits a gap from real-world vibration scenarios. In the future, simulating more realistic synthetic datasets using the finite element method (FEM) will further improve the performance of event-based visual vibrometry." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 560, + 411, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 560, + 411, + 573 + ], + "spans": [ + { + "bbox": [ + 314, + 560, + 411, + 573 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001, 62402014, 62276007), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). Peiqi Duan was also supported by China National Postdoctoral Program for Innovative Talents (Grant No. BX20230010) and China Postdoctoral Science Foundation (Grant No. 2023M740076). The authors thank openbayes.com for providing computing resource." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "24673" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 124 + ], + "type": "text", + "content": "[1] Silvio Bianchi and Emanuele Giacomozzi. Long-range detection of acoustic vibrations by speckle tracking. Applied optics, 58(28):7805-7809, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 296, + 169 + ], + "type": "text", + "content": "[2] Katherine L Bouman, Bei Xiao, Peter Battaglia, and William T Freeman. Estimating the material properties of fabric from video. In Proc. of IEEE International Conference on Computer Vision, 2013. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "spans": [ + { + "bbox": [ + 62, + 171, + 295, + 225 + ], + "type": "text", + "content": "[3] Oral Büyüközürk, Mehmet Ali Taşdemir, O Büyüközürk, R Haupt, C Tuakta, and J Chen. Remote detection of debonding in frp-strengthened concrete structures using acoustic-laser technique. In Proc. of Nondestructive Testing of Materials and Structures, 2013. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 62, + 228, + 296, + 271 + ], + "type": "text", + "content": "[4] Oral Buyukozturk, Justin G Chen, Neal Wadhwa, Abe Davis, Frédo Durand, and William T Freeman. Smaller than the eye can see: Vibration analysis with video cameras. In Proc. of World Conference on Non-Destructive Testing, 2016. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 273, + 295, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 273, + 295, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 273, + 295, + 317 + ], + "type": "text", + "content": "[5] Justin G Chen, Robert W Haupt, and Oral Buyukozturk. The acoustic-laser vibrometry technique for the noncontact detection of discontinuities in fiber reinforced polymer-retrofitted concrete. Materials evaluation, 72(10), 2014. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 295, + 372 + ], + "type": "text", + "content": "[6] Justin G Chen, Neal Wadhwa, Young-Jin Cha, Fredo Durand, William T Freeman, and Oral Buyukozturk. Modal identification of simple structures with high-speed video using motion magnification. Journal of Sound and Vibration, 345:58-71, 2015. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 418 + ], + "type": "text", + "content": "[7] Yutian Chen, Shi Guo, Fangzheng Yu, Feng Zhang, Jinwei Gu, and Tianfan Xue. Event-based motion magnification. In Proc. of European Conference on Computer Vision, 2024. 2, 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 420, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 420, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 420, + 295, + 473 + ], + "type": "text", + "content": "[8] Abe Davis, Michael Rubinstein, Neal Wadhwa, Gautham J. Mysore, Frédo Durand, and William T. Freeman. The visual microphone: passive recovery of sound from video. ACM Transactions on Graphics, 33(4):79:1-79:10, 2014. 1, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 476, + 295, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 541 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 541 + ], + "type": "text", + "content": "[9] Abe Davis, Katherine L Bouman, Justin G Chen, Michael Rubinstein, Oral Buyukozturk, Frédo Durand, and William T Freeman. Visual Vibrometry: Estimating material properties from small motions in video. IEEE Transactions on Pattern Analysis and Machine Intelligence, 39(4):732-745, 2017. 1, 3, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 295, + 598 + ], + "type": "text", + "content": "[10] Charles Dorn, Sudeep Dasari, Yongchao Yang, Garrett Kenyon, Paul Welch, and David Mascareñas. Efficient fullfield operational modal analysis using neuromorphic event-based imaging. In Proc. of Conference and Exposition on Structural Dynamics, 2017. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 633 + ], + "type": "text", + "content": "[11] Franz Durst, Adrian Melling, and James H Whitelaw. Principles and practice of laser-doppler anemometry. NASA STI/Recon Technical Report A, 76:47019, 1976. 1, 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 678 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes (VOC) challenge. International Journal of Computer Vision, 88(2):303-338, 2010. 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[13] Berthy T Feng, Alexander C Ogren, Chiara Daraio, and Katherine L Bouman. Visual vibration tomography: Estimating interior material properties from monocular video. In" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 1, 3, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "text", + "content": "[14] Thomas Finateu, Atsumi Niwa, Daniel Matolin, Koya Tsuchimoto, Andrea Mascheroni, Etienne Reynaud, Pooria Mostafalu, Frederick Brady, Ludovic Chotard, Florian LeGoff, et al. A " + }, + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "text", + "content": " back-illuminated stacked temporal contrast event-based vision sensor with " + }, + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "inline_equation", + "content": "4.86\\mu \\mathrm{m}" + }, + { + "bbox": [ + 317, + 96, + 553, + 184 + ], + "type": "text", + "content": " pixels, 1.066 geps readout, programmable event-rate controller and compressive data-formating pipeline. In Proc. of IEEE International Solid-State Circuits Conference, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "text", + "content": "[15] David J Fleet and Allan D Jepson. Computation of component image velocity from local phase information. International Journal of Computer Vision, 5:77-104, 1990. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 274 + ], + "type": "text", + "content": "[16] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 555, + 319 + ], + "type": "text", + "content": "[17] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-End learning of representations for asynchronous event-based data. In Proc. of IEEE International Conference on Computer Vision, 2019. 5" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 320, + 555, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 320, + 555, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 320, + 555, + 364 + ], + "type": "text", + "content": "[18] Daniel Gehrig, Henri Rebecq, Guillermo Gallego, and Davide Scaramuzza. EKLT: asynchronous photometric feature tracking using events and frames. International Journal of Computer Vision, 128(3):601-618, 2020. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "spans": [ + { + "bbox": [ + 317, + 365, + 553, + 408 + ], + "type": "text", + "content": "[19] Mathias Gehrig, Manasi Muglikar, and Davide Scaramuzza. Dense continuous-time optical flow from event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(7):4736-4746, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 464 + ], + "type": "text", + "content": "[20] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):8553-8565, 2023. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 466, + 553, + 510 + ], + "type": "text", + "content": "[21] Javier Hidalgo-Carrió, Guillermo Gallego, and Davide Scaramuzza. Event-aided direct sparse odometry. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 512, + 553, + 544 + ], + "type": "text", + "content": "[22] Matthew Howard and Keigo Hirakawa. Event-based visual microphone. In Proc. of International Conference on Acoustics, Speech and Signal Processing, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[23] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbrück. V2E: From video frames to realistic DVS events. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 2, 4, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 623 + ], + "type": "text", + "content": "[24] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In Proc. of International Conference on Learning Representations, 2015. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": "[25] Patrick Lichtsteiner, Christoph Posch, and Tobi Delbruck. A " + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": " 120 db " + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "inline_equation", + "content": "15~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 317, + 624, + 553, + 667 + ], + "type": "text", + "content": " latency asynchronous temporal contrast vision sensor. IEEE Journal of Solid-State Circuits, 43 (2):566-576, 2008. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[26] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólár, and C Lawrence Zitnick. Microsoft coco: Common objects in context. In Proc. of European Conference on Computer Vision, 2014. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "24674" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[27] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[28] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In Proc. of International Conference on Learning Representations, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[29] Yuanyuan Lv, Liang Zhou, Zhaohui Liu, and Haiyang Zhang. Structural vibration frequency monitoring based on event camera. Measurement Science and Technology, 35(8): 085007, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 241 + ], + "type": "text", + "content": "[30] William N MacPherson, Mark Reeves, David P Towers, Andrew J Moore, Julian DC Jones, Martin Dale, and Craig Edwards. Multipoint laser vibrometer for modal analysis. Applied optics, 46(16):3126-3132, 2007. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 297 + ], + "type": "text", + "content": "[31] M Martarelli, Gian Marco Revel, and C Santolini. Automated modal analysis by scanning laser vibrometry: problems and uncertainties associated with the scanning system calibration. Mechanical systems and signal processing, 15 (3):581-601, 2001. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 342 + ], + "type": "text", + "content": "[32] Manasi Muglikar, Mathias Gehrig, Daniel Gehrig, and Davide Scaramuzza. How to calibrate your event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 397 + ], + "type": "text", + "content": "[33] Woong-jae Na, Kyung Ho Sun, Byeong Chan Jeon, Jaeyun Lee, and Yun-ho Shin. Event-based micro vibration measurement using phase correlation template matching with event filter optimization. Measurement, 215:112867, 2023. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 443 + ], + "type": "text", + "content": "[34] Ryogo Niwa, Tatsuki Fushimi, Kenta Yamamoto, and Yoichi Ochiai. Live demonstration: Event-based visual microphone. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, 2023. 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[35] Tae-Hyun Oh, Ronnachai Jaroensri, Changil Kim, Mohamed A. Elgharib, Frédo Durand, William T. Freeman, and Wojciech Matusik. Learning-based video motion magnification. In Proc. of European Conference on Computer Vision, 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 544 + ], + "type": "text", + "content": "[36] Liyuan Pan, Miaomiao Liu, and Richard Hartley. Single image optical flow estimation with an event camera. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "text", + "content": "[37] Federico Paredes-Vallés, Kirk YW Scheper, Christophe De Wagter, and Guido CHE De Croon. Taming contrast maximization for learning sequential, low-latency, event-based optical flow. In Proc. of IEEE International Conference on Computer Vision, 2023. 2, 3, 6, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[38] Ming-Zher Poh, Daniel J McDuff, and Rosalind W Picard. Non-contact, automated cardiac pulse measurements using video imaging and blind source separation. Optics Express, 18(10):10762-10774, 2010. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[39] Javier Portilla and Eero P Simoncelli. A parametric texture model based on joint statistics of complex wavelet coefficients. International Journal of Computer Vision, 40:49-70, 2000. 2, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[40] Nicolaas Bernardus Roozen, Ludovic Labelle, Monika Rychtarikova, and Christ Glorieux. Determining radiated" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "sound power of building structures by means of laser doppler vibrometry. Journal of Sound and Vibration, 346:81-99, 2015. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 108, + 553, + 140 + ], + "type": "text", + "content": "[41] Steve Rothberg, JR Baker, and Neil A Halliwell. Laser vibrometry: pseudo-vibrations. Journal of Sound and Vibration, 135(3):516-522, 1989. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 143, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 163 + ], + "type": "text", + "content": "[42] Ahmed A Shabana. Theory of vibration. Springer, 1991. 3, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 167, + 553, + 209 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 167, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 316, + 167, + 553, + 209 + ], + "type": "text", + "content": "[43] Mark Sheinin, Dorian Chan, Matthew O'Toole, and Srinivasa G Narasimhan. Dual-shutter optical vibration sensing. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 213, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 213, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 213, + 553, + 266 + ], + "type": "text", + "content": "[44] Chenyang Shi, Ningfang Song, Boyi Wei, Yuzhen Li, Yibo Zhang, Wenzhuo Li, and Jing Jin. Event-based vibration frequency measurement with laser-assisted illumination based on mixture gaussian distribution. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 269, + 553, + 323 + ], + "type": "text", + "content": "[45] Xingjian Shi, Zhourong Chen, Hao Wang, Dit-Yan Yeung, Wai-Kin Wong, and Wang-chun Woo. Convolutional LSTM network: A machine learning approach for precipitation nowcasting. In Advances in Neural Information Processing Systems, 2015. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 369 + ], + "type": "text", + "content": "[46] Shintaro Shiba, Friedhelm Hamann, Yoshimitsu Aoki, and Guillermo Gallego. Event-based background-oriented schlieren. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(4):2011-2026, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 426 + ], + "type": "text", + "content": "[47] Shintaro Shiba, Yannick Klose, Yoshimitsu Aoki, and Guillermo Gallego. Secrets of event-based optical flow, depth and ego-motion estimation by contrast maximization. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(12):7742-7759, 2024. 2, 3, 4, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 428, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 428, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 428, + 553, + 449 + ], + "type": "text", + "content": "[48] Peter J Shull. Nondestructive evaluation: theory, techniques, and applications. CRC press, 2002. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "spans": [ + { + "bbox": [ + 316, + 452, + 553, + 495 + ], + "type": "text", + "content": "[49] Eero P Simoncelli and William T Freeman. The steerable pyramid: A flexible architecture for multi-scale derivative computation. In Proc. of International Conference on Image Processing, 1995. 2, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 540 + ], + "type": "text", + "content": "[50] Eero P Simoncelli, William T Freeman, Edward H Adelson, and David J Heeger. Shiftable multiscale transforms. IEEE transactions on Information Theory, 38(2):587-607, 1992. 2, 6" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "text", + "content": "[51] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In Proc. of European Conference on Computer Vision, 2022. 5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "type": "text", + "content": "[52] Zachary Teed and Jia Deng. RAFT: recurrent all-pairs field transforms for optical flow. In Proc. of European Conference on Computer Vision, 2020. 5, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 667 + ], + "type": "text", + "content": "[53] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Phase-based video motion processing. ACM Transactions on Graphics, 32(4):1-10, 2013. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[54] Neal Wadhwa, Michael Rubinstein, Frédo Durand, and William T Freeman. Riesz pyramids for fast phase-based video magnification. In Proc. of International Conference on Computational Photography, 2014. 2" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24675" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 262 + ], + "type": "list", + "angle": 0, + "index": 4, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[55] Zihao W Wang, Peiqi Duan, Oliver Cossairt, Aggelos Katsaggelos, Tiejun Huang, and Boxin Shi. Joint filtering of intensity images and neuromorphic events for high-resolution noise-robust imaging. In Proc. of IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 172 + ], + "type": "text", + "content": "[56] Hao-Yu Wu, Michael Rubinstein, Eugene Shih, John Guttag, Frédo Durand, and William Freeman. Eulerian video magnification for revealing subtle changes in the world. ACM Transactions on Graphics, 31(4):1-8, 2012. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 217 + ], + "type": "text", + "content": "[57] Nan Wu and Shinichiro Haruyama. The 20k samples-persecond real time detection of acoustic vibration based on displacement estimation of one-dimensional laser speckle images. Sensors, 21(9):2938, 2021. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 262 + ], + "type": "text", + "content": "[58] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. EV-FlowNet: Self-supervised optical flow estimation for event-based cameras. In Proc. of Robotics: Science and Systems, 2018. 2, 6" + } + ] + } + ], + "index": 3 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "24676" + } + ] + } + ], + "index": 5 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_content_list.json b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef64263a88172da82784689dbdf8b507db5b4f01 --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_content_list.json @@ -0,0 +1,1713 @@ +[ + { + "type": "text", + "text": "Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction", + "text_level": 1, + "bbox": [ + 114, + 130, + 880, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Wenhao Xu1 Wenming Weng1 Yueyi Zhang2, * Ruikang Xu1 Zhiwei Xiong1, * \n1University of Science and Technology of China 2MiroMind", + "bbox": [ + 153, + 178, + 843, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "wh-xu@mail.ustc.edu.cn, yueyi.zhang@miromind.ai, zwxiong@ustc.edu.cn", + "bbox": [ + 196, + 219, + 799, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://github.com/wh-xul/E-D3DGS", + "bbox": [ + 346, + 237, + 643, + 250 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/701fea225884661de94c68de553537d4cafda88441bab97a8cd5120039349391.jpg", + "image_caption": [ + "Figure 1. Left: Quality comparison with the baselines 4D-GS [42], Event-4DGS (the event-extended version of [46]), and our variants, highlighting the superior rendering quality of our method. Our GS-threshold joint modeling (GTJM) effectively eliminates event-induced artifacts (particularly the purple haze), while our dynamic-static decomposition (DSD) improves the quality of dynamic regions. Middle: Separate rendering of dynamic and static Gaussians from our DSD. Right: The scatter plot illustrates our method's ability to achieve both high fidelity and fast rendering, where dot radii correspond to different resolutions $(400\\times 400,600\\times 600,$ and $800\\times 800)$ ." + ], + "image_footnote": [], + "bbox": [ + 94, + 282, + 906, + 448 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 532, + 326, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Deformable 3D Gaussian Splating (3D-GS) is limited by missing intermediate motion information due to the low temporal resolution of RGB cameras. To address this, we introduce the first approach combining event cameras, which capture high-temporal-resolution, continuous motion data, with deformable 3D-GS for dynamic scene reconstruction. We observe that threshold modeling for events plays a crucial role in achieving high-quality reconstruction. Therefore, we propose a GS-Threshold Joint Modeling strategy, creating a mutually reinforcing process that greatly improves both 3D reconstruction and threshold modeling. Moreover, we introduce a Dynamic-Static Decomposition strategy that first identifies dynamic areas by exploiting the inability of static Gaussians to represent motions, then applies a buffer-based soft decomposition to separate dynamic and static areas. This strategy accelerates rendering by avoiding unnecessary deformation in static areas, and focuses on dynamic areas to enhance fidelity. Additionally, we contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance.", + "bbox": [ + 88, + 565, + 485, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 532, + 643, + 547 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dynamic scene reconstruction and novel view synthesis are essential for immersive applications in virtual/augmented reality and entertainment [20, 24, 34, 42, 46, 47, 49]. While Neural Radiance Fields (NeRF) [2, 26, 36] offer unprecedented photorealism, they are constrained by slow training and rendering speeds. Despite recent advances in optimization techniques [5, 7-9], real-time rendering remains elusive. Recently, 3D Gaussian Splatting (3D-GS) [17] addresses this limitation through efficient differentiable rasterization, yet existing dynamic extensions [24, 42, 46, 47] are constrained by inherent limitations of RGB cameras, including low frame rates and motion blur.", + "bbox": [ + 511, + 561, + 906, + 743 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we present event-boosted deformable 3D Gaussians for dynamic scene reconstruction. Event cameras [10, 19, 38, 39], with their microsecond-level temporal resolution, can provide continuous motion information and near-infinite viewpoints that traditional RGB cameras often fail to capture. These advantages make event cameras particularly valuable for dynamic scene reconstruction.", + "bbox": [ + 511, + 746, + 908, + 852 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "However, integrating events into 3D scene reconstruction faces new challenges. Specifically, event supervision for 3D-GS relies on an accurate event generation model", + "bbox": [ + 511, + 854, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28334", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "[4, 18], where the threshold undergoes complex variations across polarity, space, and time [13, 18]. Previous methods [4, 6, 18, 25, 30, 32, 43, 44, 48] adopt a constant threshold, yet this simplification significantly degrades the quality of event supervision. While recent works [15, 22] attempt to model threshold variations using event data alone, they achieve limited success due to the inherent binary nature of events, which only indicate brightness change directions. To address this challenge, we propose a novel GS-threshold joint modeling strategy. First, we leverage the brightness change values from RGB frames to supervise threshold optimization. Second, since the sparsity of RGB frames weakens supervision, we use 3D-GS rendered results as pseudo-intermediate frames to enhance the supervision. This finally creates a mutually reinforcing process where RGB-optimized threshold enables better event supervision for 3D-GS, while improved 3D-GS in turn provides accurate geometric constraints for threshold refinement.", + "bbox": [ + 89, + 90, + 480, + 363 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Furthermore, we observe that existing dynamic 3D-GS methods inefficiently use dynamic Gaussians solely to model both static and dynamic regions [1, 12, 14, 23, 24, 42, 46]. This unified treatment leads to reduced rendering speed, wasted deformation field capacity, and degraded reconstruction quality. While some methods have explored dynamic-static decomposition, they are limited by either inaccurate dynamic Gaussians initialization [20] or constraints in multi-view scenarios [34]. To address these limitations, we propose a novel dynamic-static decomposition strategy that first identifies dynamic regions based on the inherent inability of static Gaussians to represent motions, and then employs a buffer-based soft decomposition to adaptively search for the optimal decomposition boundary. This decomposition not only accelerates rendering by eliminating unnecessary deformation computations in static regions, but also enhances reconstruction quality by focusing the deformation field exclusively on dynamic regions.", + "bbox": [ + 89, + 369, + 482, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions can be summarized as follows:", + "bbox": [ + 109, + 648, + 470, + 664 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We present the first method integrating event cameras with deformable 3D-GS for dynamic scene reconstruction, enabling high-fidelity and fast rendering.", + "- We propose a novel GS-threshold joint modeling strategy that combines RGB-assisted initial estimation with GS-boosted refinement, creating a mutually reinforcing process that significantly improves both threshold modeling and 3D reconstruction.", + "- We introduce an effective dynamic-static decomposition strategy that not only accelerates rendering through selective deformation computation but also enhances reconstruction quality by focusing on dynamic regions.", + "- We contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance." + ], + "bbox": [ + 89, + 671, + 482, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Neural Rendering for Dynamic Scenes. Neural rendering techniques have revolutionized dynamic scene reconstruction in recent years. Pioneering works like D-NeRF [29] and Nerfies [28] extend Neural Radiance Fields [26] through deformation field, mapping observations into a canonical space for modeling non-rigid motion. Despite their impressive reconstruction quality, these methods are constrained by extensive computational demands due to dense MLP evaluations during training and rendering. Various acceleration strategies have been proposed to address these limitations. K-Planes [9] introduces an efficient explicit representation using six feature planes, while Tensor4D [33] and DTensorRF [16] employ tensor decomposition techniques to achieve compact spatiotemporal encoding. A significant breakthrough came with 3D Gaussian Splatting (3D-GS) [17], which leverages efficient differentiable rasterization for real-time rendering. This advancement has spawned several dynamic scene extensions, including 4D-GS [42], Deformable-3DGS [46], and related works [24, 34], which successfully achieve real-time rendering for high-quality dynamic scene reconstruction.", + "bbox": [ + 511, + 114, + 903, + 431 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based Neural Rendering. The integration of neural representations with event-based 3D reconstruction has emerged as a promising research direction. Pioneering works like EventNeRF [32], E-NeRF [18], and Ev-NeRF [15] first demonstrated the potential of pure event-based static scene reconstruction, albeit with different assumptions about event camera characteristics. To further improve reconstruction quality, E2NeRF [30] and Ev-DeblurNeRF [4] incorporated blurry RGB images alongside events. A significant milestone was achieved by DE-NeRF [25], which pioneered the combination of events and RGB frames for dynamic scene reconstruction. The recent advent of 3D Gaussian Splatting [17] has catalyzed new developments in event-based methods. While Ev-GS [43] adapted the pure event-based paradigm to 3D-GS, subsequent works including E2GS [6], EaDeblur-GS [41], and Event3DGS [44] primarily addressed deblurring challenges. To the best of our knowledge, our work is the first to integrate events with deformable 3D-GS for dynamic scene reconstruction. Our novel method and benchmark, specifically tailored for events and dynamic scenes, highlight significant and independent contributions that set our work apart.", + "bbox": [ + 511, + 431, + 903, + 765 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 776, + 604, + 791 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Event Cameras for 3D-GS", + "text_level": 1, + "bbox": [ + 513, + 801, + 751, + 816 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Gaussian Splatting Preliminary. 3D Gaussian Splatting (3D-GS) [17] represents a scene as anisotropic 3D Gaussians, each characterized by a covariance matrix $\\Sigma$ and center position $\\mu$ : $GS(\\mathbf{x}) = e^{-\\frac{1}{2} (\\mathbf{x} - \\mu)^T\\boldsymbol{\\Sigma}^{-1}(\\mathbf{x} - \\mu)}$ . The covariance matrix $\\Sigma$ is parameterized using scaling matrix $S$", + "bbox": [ + 511, + 823, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28335", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/aad825284d12801d28d74671a61736e0a28da0b29ef58cf084d5c3ffc6f4ad44.jpg", + "image_caption": [ + "Figure 2. Overview of GS-threshold joint modeling strategy. $\\mathcal{L}_{rgb}$ optimizes 3D-GS, $\\mathcal{L}_{thres}$ optimizes the threshold, and $\\mathcal{L}_{event}$ jointly optimizes both 3D-GS and threshold." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 486, + 241 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and rotation matrix $R$ to ensure positive semi-definiteness: $\\Sigma = RSS^T R^T$ . Each Gaussian is further defined by spherical harmonic coefficients $\\mathcal{C}$ and opacity $\\sigma$ . Final pixel colors $c$ are computed using differentiable tile-based rasterization", + "bbox": [ + 89, + 315, + 483, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nc = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 196, + 383, + 483, + 425 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $c_{i}$ denotes spherical harmonic color, and $\\alpha_{i}$ combines opacity $\\sigma$ with projected $GS(\\mathbf{x})$ .", + "bbox": [ + 89, + 431, + 482, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event/RGB Rendering Loss. Event cameras [10, 19, 21, 40, 45] are novel sensors, that asynchronously capture pixel-wise brightness changes with microsecond-level temporal resolution. Their high temporal precision enables capturing crucial motions between RGB frames and provides near-infinite viewpoint supervision, making them ideal for monocular dynamic scene reconstruction.", + "bbox": [ + 89, + 462, + 483, + 566 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Each event is represented as $e_{x,y}(\\tau) = p\\delta(\\tau)$ , where $(x,y)$ is pixel position, $\\tau$ is timestamp, $p \\in \\{+1, -1\\}$ indicates brightness change direction relative to threshold $C$ , and $\\delta(t)$ is a unit integral impulse function. Omitting pixel subscripts, the brightness change over interval $\\triangle t$ can be formulated as", + "bbox": [ + 89, + 566, + 483, + 657 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE (t, t + \\Delta t) = \\int_ {t} ^ {t + \\Delta t} C \\cdot e (\\tau) d \\tau . \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 661, + 482, + 696 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This change can also be estimated from rendered brightness", + "bbox": [ + 89, + 702, + 482, + 718 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} (t, t + \\triangle t) \\stackrel {\\mathrm {d e f}} {=} \\log (\\hat {I} (t + \\triangle t)) - \\log (I (t)), \\quad (3)\n$$\n", + "text_format": "latex", + "bbox": [ + 135, + 726, + 482, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{I}$ and $I$ respectively denote 3D-GS rendered and ground truth brightnesses. The event rendering loss is", + "bbox": [ + 89, + 752, + 483, + 784 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| E (t, t + \\triangle t) - \\hat {E} (t, t + \\triangle t) \\right\\| _ {2} ^ {2}. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 790, + 482, + 819 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Similarly, we utilize the RGB rendering loss [17] combining L1 and D-SSIM losses as", + "bbox": [ + 89, + 824, + 483, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {r g b} = \\left(1 - \\lambda_ {s}\\right) \\left\\| \\hat {I} (t) - I (t) \\right\\| _ {1} + \\lambda_ {s} \\mathcal {L} _ {D - S S I M} (\\hat {I} (t), I (t)), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 862, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a190dd5d3358b89e676b2579f01d3827ab2e4bd5966c4c0772fccca9ecb5bc61.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 720, + 202 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f81b37a6d672990ea6842acdda2747088896f6a9e38752b03822396a9307ab57.jpg", + "image_caption": [ + "(b)", + "Figure 3. (a) The effect of different ranges of threshold variation for 3D reconstruction. (b) The effect of different number of RGB frames for threshold estimation." + ], + "image_footnote": [], + "bbox": [ + 723, + 88, + 903, + 203 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\lambda_{s}$ is a weighting factor that controls the balance. 3D-GS is optimized to minimize $\\mathcal{L}_{event}$ and $\\mathcal{L}_{rgb}$ jointly", + "bbox": [ + 511, + 280, + 903, + 311 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nG S ^ {*} = \\arg \\min _ {G S} \\left(\\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 319, + 903, + 342 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. GS-threshold Joint Modeling", + "text_level": 1, + "bbox": [ + 511, + 358, + 774, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Eq. (2), the threshold $C$ critically affects event integration and supervision quality. While existing methods [4, 18, 25, 30, 32] typically assume a constant threshold, real event cameras exhibit threshold variations across polarity, space, and time [13, 18]. Fig. 3 (a) demonstrates how increasing threshold variation significantly degrades reconstruction quality. To model threshold variations, we propose a GS-threshold joint modeling (GTJM) strategy (see Fig. 2), consisting of RGB-assisted threshold estimation and GS-boosted threshold refinement.", + "bbox": [ + 511, + 380, + 906, + 530 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "RGB-assisted Threshold Estimation. Recent works [15, 22] attempt to model threshold variations solely from event data but face significant limitations due to the difficulty of inferring brightness changes from events, which only indicate the direction of change. We propose to leverage brightness change values from RGB frames for robust threshold estimation. Given two RGB frames $I(t)$ and $I(f)$ at times $t$ and $f$ , we define the threshold modeling loss as", + "bbox": [ + 511, + 531, + 908, + 652 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t h r e s}} = \\left\\| E _ {\\text {t h r e s}} (t, f) - \\hat {E} _ {\\text {t h r e s}} (t, f) \\right\\| _ {2} ^ {2}, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 660, + 903, + 690 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\hat{E}_{thres}(t,f) := \\int_t^f \\hat{C} \\cdot e(\\tau)d\\tau$ , $E_{thres}(t,f) = \\log(I(f)) - \\log(I(t))$ . In practice, we adopt a simple yet fast way to compute $\\hat{E}_{thres}(t,f)$ . We first accumulate events to obtain event count maps [11] $ECM_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}$ , where $B$ denotes time bins and $P = 2$ corresponds to the event polarity. Using learnable threshold parameters $\\hat{C}_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}$ , we compute", + "bbox": [ + 511, + 698, + 906, + 808 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {E} _ {\\text {t h r e s}} (t, f) = \\sum_ {b = 1} ^ {B} \\sum_ {p = 1} ^ {P} \\left(E C M _ {t, f} \\odot \\hat {C} _ {t, f}\\right) _ {b, p,:,:}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 535, + 816, + 903, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where threshold $\\hat{C}_{t,f}$ is optimized by minimizing Eq. (7) in an end-to-end manner.", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28336", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/9e57a07cfd095382cce5f58ad10f905992322aa40026eb9c50b0ac5edd896c1e.jpg", + "table_caption": [ + "Table 1. Step-by-step validation of mutual boosting between threshold modeling (TM) and 3D reconstruction (3D Rec.). Abbreviations: \"Fro.\" : \"Frozen\"; \"Ft.\" : \"Fine-tuning\". Note that TM is evaluated by MSE between the estimated and GT thresholds from the simulator." + ], + "table_footnote": [], + "table_body": "
TM for 3D Rec.3D Rec. for TM
Stage1Stage23D Rec.Stage1Stage2TM
TM3D Rec. (Fro. C)PSNR↑3D Rec.TM (Fro. GS)MSE↓ (×10-4)
×✓24.46×✓8.317
✓✓26.63✓✓7.077
", + "bbox": [ + 94, + 162, + 480, + 258 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/a8d1ca8e0392ea26015c2cc88b7d70f5856ce94aec925118069266963127c241.jpg", + "table_caption": [ + "Joint TM and 3D Rec. Optimization" + ], + "table_footnote": [], + "table_body": "
Stage1Stage23D Rec.TM
TM3D Rec. & TM (Ft. C̅ & Ft. GS)PSNR↑MSE↓ (×10-4)
✓✓28.016.322
", + "bbox": [ + 98, + 273, + 478, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that accurate threshold modeling improves 3D reconstruction quality. As shown in \"TM for 3D Rec.\" in Tab. 1, our RGB-assisted threshold optimization approach benefits threshold estimation, thus significantly enhancing 3D-GS reconstruction quality and achieving a 2.17 dB PSNR improvement.", + "bbox": [ + 89, + 349, + 482, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "GS-boosted Threshold Refinement. While RGB frames facilitate threshold estimation, their effectiveness is constrained by low frame rate. As illustrated in Fig. 3 (b), sparse RGB frames lead to longer integration intervals, reducing supervision quality and threshold estimation accuracy. To overcome this limitation, we found that once a 3D-GS is trained first by Eq. (4) and Eq. (5), it can be used to render intermediate frames as additional pseudosupervision. Specifically, we freeze the trained 3D-GS and reuse Eq. (4) to enhance Eq. (7) for optimizing threshold $\\hat{C}$", + "bbox": [ + 88, + 440, + 482, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} ^ {*} = \\arg \\min _ {\\hat {C}} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}}\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 599, + 482, + 625 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that the incorporation of 3D-GS significantly enhances threshold estimation accuracy. The underlying reason is that events may provide unreliable supervision in some regions with inaccurate thresholds or noise, whereas 3D-GS can correct these errors via geometric consistency. As demonstrated in \"3D Rec. for TM\" part in Tab. 1, using trained and frozen 3D-GS for threshold modeling substantially reduces MSE, leading to more precise threshold estimation.", + "bbox": [ + 89, + 632, + 482, + 766 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Joint Threshold and GS Optimization. Having demonstrated the mutual benefits between threshold modeling and 3D reconstruction, we propose jointly optimizing both threshold $\\hat{C}$ and 3D Gaussians $GS$ through", + "bbox": [ + 89, + 768, + 482, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {C} ^ {*}, G S ^ {*} = \\arg \\min _ {\\hat {C}, G S} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 835, + 482, + 861 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We observe that this joint optimization enables a beneficial cycle where optimized thresholds enhance event su", + "bbox": [ + 88, + 869, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b741992a7bf752ce457db7a807f1a1ffcc7ea919f4abecbcda42f88090fb6fe9.jpg", + "image_caption": [ + "Figure 4. The effect of dynamic-static decomposition strategy, which improves the rendering quality of dynamic regions." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 883, + 252 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "pervision for 3D-GS, while improved 3D-GS refines threshold estimates through geometric consistency. As shown in \"Joint TM and 3D Rec. Optimization\" part in Tab. 1, this approach achieves superior threshold modeling and reconstruction quality.", + "bbox": [ + 511, + 306, + 903, + 382 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In summary, our optimization proceeds in two stages: first optimizing the threshold using $\\mathcal{L}_{thres}$ , then jointly optimizing both threshold and 3D-GS using all three losses $\\mathcal{L}_{thres},\\mathcal{L}_{event}$ , and $\\mathcal{L}_{rgb}$ .", + "bbox": [ + 511, + 385, + 903, + 446 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Dynamic-static Decomposition", + "text_level": 1, + "bbox": [ + 513, + 460, + 782, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic scenes typically contain substantial static regions (e.g., tables, walls) that require no deformation. Unlike existing methods [1, 12, 14, 23, 24, 42, 46] that use dynamic Gaussians throughout, we separately model dynamic and static regions with corresponding Gaussian types. This decomposition offers dual benefits: accelerated rendering by bypassing deformation field computation for static Gaussians, and enhanced deformation fidelity through focused MLP capacity optimization for dynamic regions, as demonstrated in Fig. 4.", + "bbox": [ + 511, + 484, + 905, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The key challenge lies in accurately initializing dynamic Gaussians in dynamic regions and static Gaussians in static regions. We address this through a proposed dynamic-static decomposition (DSD) strategy, as illustrated in Fig. 5.", + "bbox": [ + 511, + 638, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Dynamic-static Decomposition on 2D. We leverage the inherent inability of static Gaussians in representing motion to decompose dynamic and static regions in 2D images. During the first $3k$ iterations, we perform scene reconstruction using only static Gaussians for initialization. This naturally results in poor reconstruction in dynamic regions while achieving high fidelity in static areas (illustrated in Fig. 5 (a)). This distinct performance difference enables decomposition of training images into dynamic and static regions.", + "bbox": [ + 511, + 700, + 905, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, using a pretrained VGG19 [35] network $\\mathcal{F}_{\\phi}$ we extract multi-scale features from both rendered image $\\hat{I}(t)$ and ground truth $I(t)$ . The cosine similarities computed at each scale are upsampled to a uniform resolution", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28337", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e58d64e958d6e0898b474036ac15ced927e96aa93508d6c69100542265c1aafb.jpg", + "image_caption": [ + "Dynamic-static Decomposition on 2D", + "Decomposition Correspondence from 2D to 3D", + "Joint Rendering", + "Figure 5. Overview of dynamic-static decomposition strategy. First, we decompose dynamic and static regions in 2D images based on the inherent inability of static Gaussians to represent motions. Next, we establish a correspondence to extend 2D decomposition to 3D Gaussians. Finally, the decomposed dynamic and static Gaussians are jointly rendered to reconstruct the complete dynamic scene." + ], + "image_footnote": [], + "bbox": [ + 98, + 103, + 903, + 292 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and averaged to generate a fused similarity map", + "bbox": [ + 89, + 359, + 406, + 375 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {S i m} = \\sum_ {l} U p \\left(\\frac {\\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\cdot \\mathcal {F} _ {\\phi} ^ {l} (I (t))}{\\left\\| \\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\right\\| \\left\\| \\mathcal {F} _ {\\phi} ^ {l} (I (t)) \\right\\|}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 386, + 480, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{F}_{\\phi}^{l}(\\cdot)$ represents the $l$ -th layer output of VGG19, and $Up(\\cdot)$ indicates bilinear upsampling. The histogram of the resulting similarity map exhibits a bimodal distribution as shown in Fig. 5 (b), enabling dynamic region mask generation through Otsu's [27] method", + "bbox": [ + 89, + 448, + 483, + 525 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nM a s k = \\mathbf {1} _ {\\text {S i m} < \\text {O t s u} (\\text {S i m})}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 537, + 480, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathbf{1}_{\\{\\cdot\\}}$ denotes the indicator function, which returns 1 if the condition is true. The mask is then multiplied by the ground truth image to extract the dynamic region.", + "bbox": [ + 89, + 566, + 482, + 612 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Decomposition Correspondence from 2D to 3D. To extend 2D dynamic-static decomposition to 3D Gaussians, we establish view-independent correspondences by leveraging depth information from 3D-GS rendering. By unprojecting pixels from masked dynamic regions across multiple views and merging the resulting 3D points, we obtain a comprehensive representation of dynamic regions in 3D space.", + "bbox": [ + 89, + 613, + 482, + 718 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Next, we map merged points to dynamic Gaussians based on spatial proximity. Each point expands spherically with radius $r$ to form a 3D volume (Fig. 5 (c)), initially classifying enclosed Gaussians as dynamic and others as static. To overcome potential decomposition inaccuracies and radius sensitivity, we implement a buffer-based soft decomposition strategy using two radii, $r_1$ and $r_2$ . Gaussians within $r_1$ are marked as dynamic, beyond $r_2$ as static, while those between are pruned to create a buffer zone. This strategy enables 3D-GS to optimize decomposition boundaries through adaptive density control (ADC) [17], enhancing both rendering quality and speed. As demonstrated in Fig. 12, the", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "strategy also exhibits improved robustness to radius parameter selection.", + "bbox": [ + 511, + 359, + 903, + 387 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It should be noted that, our DSD method is performed only once during the entire training process and requires only about one minute, introducing minimal overhead to the training pipeline.", + "bbox": [ + 511, + 388, + 905, + 449 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint Rendering of Dynamic and Static Gaussians. With the decomposed dynamic and static Gaussians, we jointly render the entire dynamic scene. Particularly, a deformation field [46] learns to map dynamic Gaussians from canonical space to arbitrary time. Taking time $t$ and the center position $\\mu$ of dynamic Gaussians as inputs, the deformation field outputs the displacement of their position $\\delta_{\\mu}$ , rotation $\\delta_r$ , and scaling $\\delta_s$", + "bbox": [ + 511, + 450, + 906, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\delta_ {\\mu}, \\delta_ {r}, \\delta_ {s}\\right) = \\mathcal {F} _ {\\theta} (\\gamma (s g (\\mu)), \\gamma (t)), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 580, + 903, + 598 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $sg(\\cdot)$ indicates a stop-gradient operation and $\\gamma (\\cdot)$ denotes the positional encoding [46]. Then, the deformed dynamic gaussians can be addressed as", + "bbox": [ + 511, + 607, + 905, + 654 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {r} ^ {\\prime}, \\boldsymbol {s} ^ {\\prime}\\right) = \\left(\\boldsymbol {\\mu} + \\delta_ {\\boldsymbol {\\mu}}, \\boldsymbol {r} + \\delta_ {\\boldsymbol {r}}, \\boldsymbol {s} + \\delta_ {\\boldsymbol {s}}\\right). \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 662, + 903, + 680 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finally, static Gaussians bypass the deformation field and merge with the deformed dynamic Gaussians as inputs to the rasterizer, enabling high-frame-rate dynamic rendering.", + "bbox": [ + 511, + 689, + 905, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 748, + 638, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Settings", + "text_level": 1, + "bbox": [ + 511, + 772, + 720, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Current datasets for event-based dynamic scene reconstruction are highly limited, with only three synthetic and three real-world scenes [25], all unpublished. Notably, the only three publicly available real-world scenes [25] were captured with a static camera, making novel view evaluation infeasible. To facilitate future research, we build the first event-inclusive 4D benchmark featuring 8 synthetic", + "bbox": [ + 511, + 794, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28338", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/25f6230e20b752f137c2060e18548b97dc2532698e7a45f8349c7120611a82dc.jpg", + "table_caption": [ + "Table 2. Quantitative results on our synthetic dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events. The best and second-best result are marked in bold and underlined, respectively." + ], + "table_footnote": [], + "table_body": "
MethodLegoHotdogMaterialsMusic box
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]23.600.9180.08822330.010.9510.06426028.070.9670.06126219.200.9050.122239
TiNeuVox [7]22.390.8910.0710.5330.810.9530.0350.4926.630.9380.0540.5220.450.8310.1520.62
K-Planes [9]24.550.9310.0352.3431.360.9580.0162.3530.620.9760.0092.2920.770.8580.0712.39
4D-GS [42]26.300.9370.07210433.480.9650.05213230.400.9790.05411124.060.9370.07164
Deformable-3DGS [46]23.790.9230.0537332.910.9620.01713234.000.9860.0049122.080.9240.05251
Event-4DGS28.000.9430.0405434.610.9690.0199635.600.9890.0067428.580.9500.04342
Ours31.850.9670.01818936.150.9740.01324138.020.9930.00324030.780.9630.02992
MethodCelestial globeFanWater wheelMan
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]19.050.9150.11018221.260.8910.11827019.430.8870.10921520.790.8700.114210
TiNeuVox [7]13.620.7360.2900.6219.900.8890.1070.5317.030.8500.1470.5622.810.8870.0710.50
K-Planes [9]15.490.8570.0882.4622.100.9090.0622.3820.960.9200.0462.4121.020.8570.0732.33
4D-GS [42]20.970.9420.0725225.050.9360.08010820.960.9170.0777723.800.9140.07664
Deformable-3DGS [46]23.070.9620.0364124.660.9290.0519020.790.9120.0514323.060.9060.05137
Event-4DGS24.300.9480.0453627.660.9490.0417126.340.9320.0523025.550.9210.06333
Ours28.830.9760.0207330.180.9640.02516828.470.9500.03311228.210.9430.037129
", + "bbox": [ + 96, + 119, + 903, + 347 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/77edba42499e35ecfe1c62fb98e27603efedd1a9037531464040b83af849bf2f.jpg", + "table_caption": [ + "Table 3. Quantitative results on our real-world dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events." + ], + "table_footnote": [], + "table_body": "
MethodExcavatorJeepFlowersEagle
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
4D-GS [42]28.350.9110.11011528.340.8780.0936126.820.8730.1236327.590.9000.128105
Deformable-3DGS [46]26.120.9030.1208126.300.8700.1045226.400.9030.0796427.440.9030.12570
Event-4DGS29.670.9140.0925729.640.9010.0794727.530.9050.0844029.080.8960.10463
Ours31.280.9250.07017930.410.9050.0688928.570.9130.06914931.290.9180.074192
", + "bbox": [ + 94, + 366, + 901, + 448 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and 4 real-world dynamic scenes, encompassing diverse complexities, intricate structures, and rapid motions, thus enabling effective evaluation of dynamic reconstruction.", + "bbox": [ + 89, + 460, + 482, + 505 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For synthetic scenes, we use Blender [3] to generate one-second, $360^{\\circ}$ monocular camera rotations, producing thousands of continuous frames per scene. These high-temporal-resolution sequences are processed through ESIM [31] to generate events. For each sequence, we uniformly sample 30 frames (equivalent to 30 FPS) for training, and select intermediate frames as far apart as possible from the training frames for testing. Particularly, both \"Fan\" representing a typical high-speed 4D scene, and \"Man\" featuring large object displacements, present significant challenges.", + "bbox": [ + 88, + 506, + 482, + 657 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For real-world scenes, as shown in Fig. 6, we construct a hybrid camera system consisting of a beam splitter, an event camera (Prophesee Gen4), a frame camera (Basler ace), and a microcontroller (STM32) for outputting synchronization signals. Following [32], we keep the camera system static and place the objects on a motorized optical rotating turntable, which is equivalent to camera motion. Following prior work [25], we downsample the original high-FPS video for training and use intermediate frames for testing.", + "bbox": [ + 88, + 657, + 482, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. For RGB-only settings, we benchmarked our method against the representative NeRF baselines K-Planes [9] and TiNeuVox [7], along with Gaussians baselines 3D-GS [17], 4D-GS [42], and Deformable-3DGS [46]. For event-assisted settings, DE-NeRF [25] is the only baseline; however, it could not be directly compared, as its code was still closed-source. DE-NeRF relies on NeRF's volume ren", + "bbox": [ + 88, + 795, + 482, + 898 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4837165c12f3e9cca44395f814b8bb07f094bfdf45f1cadec074b25a7b391136.jpg", + "image_caption": [ + "Figure 6. Real-world data acquisition setup (left) and our hybrid camera system (right)." + ], + "image_footnote": [], + "bbox": [ + 545, + 458, + 870, + 584 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "dering techniques [26], leading to predictably slow rendering speeds. Moreover, its reconstruction quality is also predictably limited due to the absence of threshold modeling for events. To provide a comparable baseline, we introduce Event-4DGS, an extension of Deformable-3DGS [46] that incorporates the event rendering loss in Eq. (4).", + "bbox": [ + 511, + 627, + 906, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics. We evaluate rendering quality using PSNR, SSIM [37], and LPIPS [50] (based on AlexNet) and measure rendering speed in FPS on an NVIDIA RTX 3090 GPU.", + "bbox": [ + 511, + 718, + 906, + 763 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparisons", + "text_level": 1, + "bbox": [ + 511, + 772, + 653, + 789 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. We report the quantitative results of the comparison on the synthetic and real-world datasets in Tab. 2 and Tab. 3, respectively. Although 4D-GS [42] and Deformable-3DGS [46] achieve relatively higher FPS compared to NeRF baselines [7, 9], their reconstruction quality is limited by the sparsity of the RGB training frames. In contrast, Event-4DGS leverages the rich intermediate mo", + "bbox": [ + 511, + 794, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28339", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/55a10940b76dbe35079c00e8a2e212076c7d4f52971277923b0b2b3f8c2a20e6.jpg", + "image_caption": [ + "Figure 7. Qualitative comparisons on our synthetic dataset. Please see the supplementary video for details." + ], + "image_footnote": [], + "bbox": [ + 132, + 87, + 859, + 465 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5b8ea59bf6d16982b93e27873198469c02f1862e1a213f7a8be0d3050c31edbb.jpg", + "image_caption": [ + "Figure 8. Qualitative comparisons on our real-world dataset. Please see the supplementary video for details." + ], + "image_footnote": [], + "bbox": [ + 132, + 483, + 859, + 731 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "tion and viewpoint information provided by events, significantly outperforming other baselines in reconstruction quality, with an average PSNR improvement of 3.28 dB over Deformable-3DGS across all synthetic scenes. This notable improvement underscores the effectiveness of high-temporal-resolution event cameras for dynamic scene reconstruction. However, Event-4DGS still suffers from threshold variation, whereas our method with GTJM enables accurate threshold modeling and better event super", + "bbox": [ + 89, + 758, + 482, + 893 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vision, achieving an average PSNR improvement of 2.73 dB over Event-4DGS on synthetic datasets. Meanwhile, our method maintains exceptionally fast rendering speeds, averaging $1.71 \\times$ faster than 4D-GS on synthetic datasets. In summary, our method enjoys both the highest rendering quality and exceptional rendering speed on both synthetic and real-world datasets.", + "bbox": [ + 511, + 758, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. For a more visual assessment, we present qualitative results on the synthetic and real-world", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28340", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/08705781c372cdec7a26040c40ac742de40f88930948f974e4a91ad27608aa48.jpg", + "table_caption": [ + "Table 4. Ablation studies on synthetic dataset. For real-world ablation studies, please refer to the supplementary document." + ], + "table_footnote": [], + "table_body": "
MethodPSNR↑SSIM↑LPIPS↓FPS↑
w/o GTJM29.390.9560.034153
w/o Joint Optimization in GTJM30.870.9630.026152
w/o DSD30.780.9610.02657
w/o Buffer-based Soft Decomposition31.020.9630.025138
Full31.560.9660.022156
", + "bbox": [ + 94, + 117, + 480, + 210 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "datasets in Fig. 7 and Fig. 8, respectively. These comparisons highlight the capability of our method to deliver high-fidelity dynamic scene modeling. Notably, our method effectively captures intricate motion details, while other baselines exhibit structural deficiencies and distortions.", + "bbox": [ + 89, + 215, + 482, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic Blurry Scene Comparisons. Motion blur is another common challenge in dynamic scenes. To address this, we extend both baselines and our method with blur loss and EDI from [48], and build blurry scenes for evaluations. Fig. 9 shows that, by leveraging events' deblurring advantage, our method outperforms Deformable-3DGS by 4.79 dB in PSNR, achieving the best results. For detailed methods and quantitative results, see the supplementary document.", + "bbox": [ + 89, + 292, + 483, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 438, + 243, + 454 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "GS-threshold Joint Modeling. Using a constant threshold fails to properly neutralize opposing polarity events during accumulation, resulting in motion trajectory artifacts as shown in Fig. 10 (b). These artifacts, when used for Gaussian supervision, produce undesirable purple haze in rendered outputs, such as the Event-4DGS results in Fig. 7. Our RGB-assisted threshold estimation significantly reduces these artifacts (Fig. 10 (c)), while subsequent joint threshold and GS optimization effectively eliminates remaining distortions (Fig. 10 (d)). As demonstrated in Tab. 4, this improved event supervision yields a $2.17\\mathrm{dB}$ average PSNR improvement across all scenes, validating our GTJM strategy's effectiveness in handling threshold variations.", + "bbox": [ + 89, + 460, + 483, + 657 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Dynamic-static Decomposition. Our DSD method successfully identifies dynamic regions of varying sizes and geometries, as demonstrated in Fig. 11. Modeling the entire scene with dynamic Gaussians without DSD misallocates deformation field capacity to static regions, compromising dynamic region reconstruction quality as shown in Fig. 4. Quantitative results in Tab. 4 demonstrate that using DSD improves the average PSNR by $0.78\\mathrm{dB}$ and accelerates the rendering speed to 2.74 times the original FPS. This underscores DSD's crucial role in achieving both high-fidelity dynamic scene reconstruction and efficient rendering.", + "bbox": [ + 89, + 657, + 482, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Buffer-based Soft Decomposition. Our buffer-based soft decomposition enables adaptive optimization of decomposition boundaries, yielding a 0.54 dB improvement in average PSNR (Tab. 4). Sensitivity analysis reveals that reconstruction quality stabilizes when buffer size $(r_2 - r_1)$", + "bbox": [ + 89, + 824, + 483, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/43c8ae281459886f7adf86aa88137b11d9d07dfa13ba5de358b78ccebfe4c4bf.jpg", + "image_caption": [ + "Figure 9. Extended comparisons on the dynamic blurry scene." + ], + "image_footnote": [], + "bbox": [ + 524, + 89, + 602, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/57e620d0531d10c7d0e089b7d89b72ac5e1b00708fea5dae23063395b7c52e77.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 90, + 647, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd504b106d097fe1f5c4c75d83033e87b52371c70da886e1486c6563cbb16c84.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 651, + 90, + 723, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb78c0a3864da654c8d61b6868639ad13d787540e2c945d46a3cab0a4c53f038.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 725, + 90, + 771, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/7dcfd7deda0d68aa09981543df298986b22ea8900ab46724b372af067d760bfe.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 90, + 848, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d372c9d3ebd06f9071764a067c54c17a57427531103b1bd86a4ae41edd81428.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 851, + 90, + 893, + 152 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3e641b04f76e0d0925fd3be5f4923fa51518e238ebc8d8c4eb8a64fd65560154.jpg", + "image_caption": [ + "(a) GT" + ], + "image_footnote": [], + "bbox": [ + 522, + 171, + 614, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6400539520cccf496f75b6598a337f0807b97faa780bed6a046e9607e0715d91.jpg", + "image_caption": [ + "(b) w/o GTJM", + "c) w/o Joint Optimization" + ], + "image_footnote": [], + "bbox": [ + 616, + 171, + 709, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4ec7e36bf2924f7ffd0658f4dfa86d60e695575564f5df5217d1c37166c42d59.jpg", + "image_caption": [ + "in GTJM" + ], + "image_footnote": [], + "bbox": [ + 709, + 171, + 802, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/9b2b290f2151706660f24c79818c96da05bc2de5953cc34195e24643c2f18e3b.jpg", + "image_caption": [ + "(d) Full" + ], + "image_footnote": [], + "bbox": [ + 802, + 171, + 893, + 311 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b286634b0b8acac45b7eae191b2554ffeab58c56741a496b7ba91b7b4027d41c.jpg", + "image_caption": [ + "Figure 10. The effect of GS-threshold joint modeling strategy, which eliminates event artifacts caused by threshold variations.", + "Lego Hotdog", + "Hotdog" + ], + "image_footnote": [], + "bbox": [ + 531, + 361, + 614, + 460 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/063d96bced7e5dca0cdf0b5d0e453c296433544b033b055abcfea113105dc542.jpg", + "image_caption": [ + "Materials", + "Cel", + "estial globe", + "Man", + "Figure 11. Rendering results of dynamic and static Gaussians separated by our dynamic-static decomposition strategy.", + "Figure 12. Sensitivity analysis on buffer size $(r_2 - r_1)$ ." + ], + "image_footnote": [], + "bbox": [ + 542, + 506, + 874, + 636 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "exceeds approximately 12 basic units (normalized by average inter-Gaussian distance to account for scene variations), as shown in Fig. 12. This stability demonstrates the robustness of our DSD method through adaptive boundary search, highlighting the effectiveness of the buffer-based strategy.", + "bbox": [ + 511, + 661, + 905, + 737 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 753, + 633, + 768 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present an event-boosted deformable 3D Gaussian framework for high-quality dynamic scene reconstruction. Our GS-threshold joint modeling effectively addresses threshold variation challenges, enabling reliable event supervision. The proposed dynamic-static decomposition method enhances both rendering efficiency and reconstruction quality through optimized resource allocation between static and dynamic regions.", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28341", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 259, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001, and the Fundamental Research Funds for the Central Universities under Grant WK210000059.", + "bbox": [ + 89, + 114, + 485, + 191 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 205, + 187, + 220 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In ECCV, 2024. 2, 4", + "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1", + "[3] Blender. A 3d modelling and rendering package. https://www.blender.org/, 2018.6", + "[4] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In CVPR, 2024. 2, 3", + "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 1", + "[6] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In ICIP, 2024. 2", + "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia, 2022. 1, 6", + "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022.", + "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 1, 2, 6", + "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3", + "[11] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019. 3", + "[12] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 35(4):3119-3133, 2024. 2, 4", + "[13] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR workshop, 2021. 2, 3" + ], + "bbox": [ + 93, + 231, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR, 2024. 2, 4", + "[15] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In WACV, 2023. 2, 3", + "[16] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 2", + "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):139-1, 2023. 1, 2, 3, 5, 6", + "[18] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 2, 3", + "[19] Chenghan Li, Christian Brandli, Raphael Berner, Hongjie Liu, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. Design of an rgb color vga rolling and global shutter dynamic and active-pixel vision sensor. In ISCAS, 2015. 1, 3", + "[20] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. In WACV, 2025. 1, 2", + "[21] Yutong Liu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. S2d-lfe: Sparse-to-dense light field event generation. In CVPR, 2025. 3", + "[22] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In ICCV, 2023. 2, 3", + "[23] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, 2024. 2, 4", + "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 1, 2, 4", + "[25] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, 2023. 2, 3, 5, 6", + "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 6", + "[27] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62-66, 1979. 5", + "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2", + "[29] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28342", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 2, 3", + "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 6", + "[32] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, 2023. 2, 3, 6", + "[33] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2", + "[34] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In ECCV, 2024. 1, 2", + "[35] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 4", + "[36] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1", + "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6", + "[38] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 1", + "[39] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Boosting event stream super-resolution with a recurrent neural network. In ECCV, 2022. 1", + "[40] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In CVPR, 2023. 3", + "[41] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2", + "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 1, 2, 4, 6", + "[43] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. In MLSP, 2024. 2", + "[44] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In CoRL, 2024. 2", + "[45] Wenhao Xu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Ceia: Clip-based event-image alignment for open-world event-based understanding. In ECCV workshop, 2024. 3", + "[46] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 4, 5, 6" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[47] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 1", + "[48] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2, 8", + "[49] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: a structured and explicit radiance representation for 3d generative modeling. In NeurIPS, 2024. 1", + "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + ], + "bbox": [ + 514, + 90, + 903, + 290 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28343", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_model.json b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5dde3185fb85d4b42a520b64d1e592bd5ac4fc56 --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_model.json @@ -0,0 +1,2486 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.131, + 0.882, + 0.152 + ], + "angle": 0, + "content": "Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.154, + 0.179, + 0.844, + 0.218 + ], + "angle": 0, + "content": "Wenhao Xu1 Wenming Weng1 Yueyi Zhang2, * Ruikang Xu1 Zhiwei Xiong1, * \n1University of Science and Technology of China 2MiroMind" + }, + { + "type": "text", + "bbox": [ + 0.197, + 0.22, + 0.8, + 0.234 + ], + "angle": 0, + "content": "wh-xu@mail.ustc.edu.cn, yueyi.zhang@miromind.ai, zwxiong@ustc.edu.cn" + }, + { + "type": "text", + "bbox": [ + 0.348, + 0.238, + 0.645, + 0.251 + ], + "angle": 0, + "content": "https://github.com/wh-xul/E-D3DGS" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.284, + 0.907, + 0.449 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.451, + 0.907, + 0.522 + ], + "angle": 0, + "content": "Figure 1. Left: Quality comparison with the baselines 4D-GS [42], Event-4DGS (the event-extended version of [46]), and our variants, highlighting the superior rendering quality of our method. Our GS-threshold joint modeling (GTJM) effectively eliminates event-induced artifacts (particularly the purple haze), while our dynamic-static decomposition (DSD) improves the quality of dynamic regions. Middle: Separate rendering of dynamic and static Gaussians from our DSD. Right: The scatter plot illustrates our method's ability to achieve both high fidelity and fast rendering, where dot radii correspond to different resolutions \\((400\\times 400,600\\times 600,\\) and \\(800\\times 800)\\)." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.534, + 0.327, + 0.549 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.566, + 0.486, + 0.884 + ], + "angle": 0, + "content": "Deformable 3D Gaussian Splating (3D-GS) is limited by missing intermediate motion information due to the low temporal resolution of RGB cameras. To address this, we introduce the first approach combining event cameras, which capture high-temporal-resolution, continuous motion data, with deformable 3D-GS for dynamic scene reconstruction. We observe that threshold modeling for events plays a crucial role in achieving high-quality reconstruction. Therefore, we propose a GS-Threshold Joint Modeling strategy, creating a mutually reinforcing process that greatly improves both 3D reconstruction and threshold modeling. Moreover, we introduce a Dynamic-Static Decomposition strategy that first identifies dynamic areas by exploiting the inability of static Gaussians to represent motions, then applies a buffer-based soft decomposition to separate dynamic and static areas. This strategy accelerates rendering by avoiding unnecessary deformation in static areas, and focuses on dynamic areas to enhance fidelity. Additionally, we contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.534, + 0.645, + 0.549 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.563, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Dynamic scene reconstruction and novel view synthesis are essential for immersive applications in virtual/augmented reality and entertainment [20, 24, 34, 42, 46, 47, 49]. While Neural Radiance Fields (NeRF) [2, 26, 36] offer unprecedented photorealism, they are constrained by slow training and rendering speeds. Despite recent advances in optimization techniques [5, 7-9], real-time rendering remains elusive. Recently, 3D Gaussian Splatting (3D-GS) [17] addresses this limitation through efficient differentiable rasterization, yet existing dynamic extensions [24, 42, 46, 47] are constrained by inherent limitations of RGB cameras, including low frame rates and motion blur." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.747, + 0.909, + 0.853 + ], + "angle": 0, + "content": "In this paper, we present event-boosted deformable 3D Gaussians for dynamic scene reconstruction. Event cameras [10, 19, 38, 39], with their microsecond-level temporal resolution, can provide continuous motion information and near-infinite viewpoints that traditional RGB cameras often fail to capture. These advantages make event cameras particularly valuable for dynamic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.901 + ], + "angle": 0, + "content": "However, integrating events into 3D scene reconstruction faces new challenges. Specifically, event supervision for 3D-GS relies on an accurate event generation model" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28334" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.364 + ], + "angle": 0, + "content": "[4, 18], where the threshold undergoes complex variations across polarity, space, and time [13, 18]. Previous methods [4, 6, 18, 25, 30, 32, 43, 44, 48] adopt a constant threshold, yet this simplification significantly degrades the quality of event supervision. While recent works [15, 22] attempt to model threshold variations using event data alone, they achieve limited success due to the inherent binary nature of events, which only indicate brightness change directions. To address this challenge, we propose a novel GS-threshold joint modeling strategy. First, we leverage the brightness change values from RGB frames to supervise threshold optimization. Second, since the sparsity of RGB frames weakens supervision, we use 3D-GS rendered results as pseudo-intermediate frames to enhance the supervision. This finally creates a mutually reinforcing process where RGB-optimized threshold enables better event supervision for 3D-GS, while improved 3D-GS in turn provides accurate geometric constraints for threshold refinement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.371, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Furthermore, we observe that existing dynamic 3D-GS methods inefficiently use dynamic Gaussians solely to model both static and dynamic regions [1, 12, 14, 23, 24, 42, 46]. This unified treatment leads to reduced rendering speed, wasted deformation field capacity, and degraded reconstruction quality. While some methods have explored dynamic-static decomposition, they are limited by either inaccurate dynamic Gaussians initialization [20] or constraints in multi-view scenarios [34]. To address these limitations, we propose a novel dynamic-static decomposition strategy that first identifies dynamic regions based on the inherent inability of static Gaussians to represent motions, and then employs a buffer-based soft decomposition to adaptively search for the optimal decomposition boundary. This decomposition not only accelerates rendering by eliminating unnecessary deformation computations in static regions, but also enhances reconstruction quality by focusing the deformation field exclusively on dynamic regions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.65, + 0.472, + 0.665 + ], + "angle": 0, + "content": "Our main contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.717 + ], + "angle": 0, + "content": "- We present the first method integrating event cameras with deformable 3D-GS for dynamic scene reconstruction, enabling high-fidelity and fast rendering." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.718, + 0.483, + 0.792 + ], + "angle": 0, + "content": "- We propose a novel GS-threshold joint modeling strategy that combines RGB-assisted initial estimation with GS-boosted refinement, creating a mutually reinforcing process that significantly improves both threshold modeling and 3D reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.793, + 0.483, + 0.853 + ], + "angle": 0, + "content": "- We introduce an effective dynamic-static decomposition strategy that not only accelerates rendering through selective deformation computation but also enhances reconstruction quality by focusing on dynamic regions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.853, + 0.483, + 0.899 + ], + "angle": 0, + "content": "- We contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.672, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.905, + 0.432 + ], + "angle": 0, + "content": "Neural Rendering for Dynamic Scenes. Neural rendering techniques have revolutionized dynamic scene reconstruction in recent years. Pioneering works like D-NeRF [29] and Nerfies [28] extend Neural Radiance Fields [26] through deformation field, mapping observations into a canonical space for modeling non-rigid motion. Despite their impressive reconstruction quality, these methods are constrained by extensive computational demands due to dense MLP evaluations during training and rendering. Various acceleration strategies have been proposed to address these limitations. K-Planes [9] introduces an efficient explicit representation using six feature planes, while Tensor4D [33] and DTensorRF [16] employ tensor decomposition techniques to achieve compact spatiotemporal encoding. A significant breakthrough came with 3D Gaussian Splatting (3D-GS) [17], which leverages efficient differentiable rasterization for real-time rendering. This advancement has spawned several dynamic scene extensions, including 4D-GS [42], Deformable-3DGS [46], and related works [24, 34], which successfully achieve real-time rendering for high-quality dynamic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.433, + 0.905, + 0.766 + ], + "angle": 0, + "content": "Event-based Neural Rendering. The integration of neural representations with event-based 3D reconstruction has emerged as a promising research direction. Pioneering works like EventNeRF [32], E-NeRF [18], and Ev-NeRF [15] first demonstrated the potential of pure event-based static scene reconstruction, albeit with different assumptions about event camera characteristics. To further improve reconstruction quality, E2NeRF [30] and Ev-DeblurNeRF [4] incorporated blurry RGB images alongside events. A significant milestone was achieved by DE-NeRF [25], which pioneered the combination of events and RGB frames for dynamic scene reconstruction. The recent advent of 3D Gaussian Splatting [17] has catalyzed new developments in event-based methods. While Ev-GS [43] adapted the pure event-based paradigm to 3D-GS, subsequent works including E2GS [6], EaDeblur-GS [41], and Event3DGS [44] primarily addressed deblurring challenges. To the best of our knowledge, our work is the first to integrate events with deformable 3D-GS for dynamic scene reconstruction. Our novel method and benchmark, specifically tailored for events and dynamic scenes, highlight significant and independent contributions that set our work apart." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.777, + 0.605, + 0.792 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.802, + 0.753, + 0.817 + ], + "angle": 0, + "content": "3.1. Event Cameras for 3D-GS" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.824, + 0.905, + 0.902 + ], + "angle": 0, + "content": "3D Gaussian Splatting Preliminary. 3D Gaussian Splatting (3D-GS) [17] represents a scene as anisotropic 3D Gaussians, each characterized by a covariance matrix \\(\\Sigma\\) and center position \\(\\mu\\): \\(GS(\\mathbf{x}) = e^{-\\frac{1}{2} (\\mathbf{x} - \\mu)^T\\boldsymbol{\\Sigma}^{-1}(\\mathbf{x} - \\mu)}\\). The covariance matrix \\(\\Sigma\\) is parameterized using scaling matrix \\(S\\)" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28335" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.488, + 0.242 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.251, + 0.485, + 0.294 + ], + "angle": 0, + "content": "Figure 2. Overview of GS-threshold joint modeling strategy. \\(\\mathcal{L}_{rgb}\\) optimizes 3D-GS, \\(\\mathcal{L}_{thres}\\) optimizes the threshold, and \\(\\mathcal{L}_{event}\\) jointly optimizes both 3D-GS and threshold." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.316, + 0.484, + 0.378 + ], + "angle": 0, + "content": "and rotation matrix \\( R \\) to ensure positive semi-definiteness: \\( \\Sigma = RSS^T R^T \\). Each Gaussian is further defined by spherical harmonic coefficients \\( \\mathcal{C} \\) and opacity \\( \\sigma \\). Final pixel colors \\( c \\) are computed using differentiable tile-based rasterization" + }, + { + "type": "equation", + "bbox": [ + 0.197, + 0.384, + 0.484, + 0.426 + ], + "angle": 0, + "content": "\\[\nc = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.433, + 0.483, + 0.462 + ], + "angle": 0, + "content": "where \\(c_{i}\\) denotes spherical harmonic color, and \\(\\alpha_{i}\\) combines opacity \\(\\sigma\\) with projected \\(GS(\\mathbf{x})\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.568 + ], + "angle": 0, + "content": "Event/RGB Rendering Loss. Event cameras [10, 19, 21, 40, 45] are novel sensors, that asynchronously capture pixel-wise brightness changes with microsecond-level temporal resolution. Their high temporal precision enables capturing crucial motions between RGB frames and provides near-infinite viewpoint supervision, making them ideal for monocular dynamic scene reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.658 + ], + "angle": 0, + "content": "Each event is represented as \\( e_{x,y}(\\tau) = p\\delta(\\tau) \\), where \\( (x,y) \\) is pixel position, \\( \\tau \\) is timestamp, \\( p \\in \\{+1, -1\\} \\) indicates brightness change direction relative to threshold \\( C \\), and \\( \\delta(t) \\) is a unit integral impulse function. Omitting pixel subscripts, the brightness change over interval \\( \\triangle t \\) can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.662, + 0.483, + 0.698 + ], + "angle": 0, + "content": "\\[\nE (t, t + \\Delta t) = \\int_ {t} ^ {t + \\Delta t} C \\cdot e (\\tau) d \\tau . \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.703, + 0.483, + 0.719 + ], + "angle": 0, + "content": "This change can also be estimated from rendered brightness" + }, + { + "type": "equation", + "bbox": [ + 0.136, + 0.727, + 0.483, + 0.748 + ], + "angle": 0, + "content": "\\[\n\\hat {E} (t, t + \\triangle t) \\stackrel {\\mathrm {d e f}} {=} \\log (\\hat {I} (t + \\triangle t)) - \\log (I (t)), \\quad (3)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.785 + ], + "angle": 0, + "content": "where \\(\\hat{I}\\) and \\(I\\) respectively denote 3D-GS rendered and ground truth brightnesses. The event rendering loss is" + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.791, + 0.483, + 0.82 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| E (t, t + \\triangle t) - \\hat {E} (t, t + \\triangle t) \\right\\| _ {2} ^ {2}. \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Similarly, we utilize the RGB rendering loss [17] combining L1 and D-SSIM losses as" + }, + { + "type": "equation", + "bbox": [ + 0.091, + 0.863, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {r g b} = \\left(1 - \\lambda_ {s}\\right) \\left\\| \\hat {I} (t) - I (t) \\right\\| _ {1} + \\lambda_ {s} \\mathcal {L} _ {D - S S I M} (\\hat {I} (t), I (t)), \\tag {5}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.722, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.608, + 0.205, + 0.625, + 0.216 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.725, + 0.089, + 0.905, + 0.204 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.817, + 0.204, + 0.834, + 0.215 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.222, + 0.905, + 0.264 + ], + "angle": 0, + "content": "Figure 3. (a) The effect of different ranges of threshold variation for 3D reconstruction. (b) The effect of different number of RGB frames for threshold estimation." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.281, + 0.905, + 0.313 + ], + "angle": 0, + "content": "where \\(\\lambda_{s}\\) is a weighting factor that controls the balance. 3D-GS is optimized to minimize \\(\\mathcal{L}_{event}\\) and \\(\\mathcal{L}_{rgb}\\) jointly" + }, + { + "type": "equation", + "bbox": [ + 0.598, + 0.32, + 0.905, + 0.343 + ], + "angle": 0, + "content": "\\[\nG S ^ {*} = \\arg \\min _ {G S} \\left(\\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {6}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.359, + 0.776, + 0.376 + ], + "angle": 0, + "content": "3.2. GS-threshold Joint Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.531 + ], + "angle": 0, + "content": "As shown in Eq. (2), the threshold \\( C \\) critically affects event integration and supervision quality. While existing methods [4, 18, 25, 30, 32] typically assume a constant threshold, real event cameras exhibit threshold variations across polarity, space, and time [13, 18]. Fig. 3 (a) demonstrates how increasing threshold variation significantly degrades reconstruction quality. To model threshold variations, we propose a GS-threshold joint modeling (GTJM) strategy (see Fig. 2), consisting of RGB-assisted threshold estimation and GS-boosted threshold refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.532, + 0.909, + 0.654 + ], + "angle": 0, + "content": "RGB-assisted Threshold Estimation. Recent works [15, 22] attempt to model threshold variations solely from event data but face significant limitations due to the difficulty of inferring brightness changes from events, which only indicate the direction of change. We propose to leverage brightness change values from RGB frames for robust threshold estimation. Given two RGB frames \\( I(t) \\) and \\( I(f) \\) at times \\( t \\) and \\( f \\), we define the threshold modeling loss as" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.661, + 0.905, + 0.691 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t h r e s}} = \\left\\| E _ {\\text {t h r e s}} (t, f) - \\hat {E} _ {\\text {t h r e s}} (t, f) \\right\\| _ {2} ^ {2}, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.699, + 0.907, + 0.809 + ], + "angle": 0, + "content": "where \\(\\hat{E}_{thres}(t,f) := \\int_t^f \\hat{C} \\cdot e(\\tau)d\\tau\\), \\(E_{thres}(t,f) = \\log(I(f)) - \\log(I(t))\\). In practice, we adopt a simple yet fast way to compute \\(\\hat{E}_{thres}(t,f)\\). We first accumulate events to obtain event count maps [11] \\(ECM_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}\\), where \\(B\\) denotes time bins and \\(P = 2\\) corresponds to the event polarity. Using learnable threshold parameters \\(\\hat{C}_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}\\), we compute" + }, + { + "type": "equation", + "bbox": [ + 0.537, + 0.817, + 0.905, + 0.86 + ], + "angle": 0, + "content": "\\[\n\\hat {E} _ {\\text {t h r e s}} (t, f) = \\sum_ {b = 1} ^ {B} \\sum_ {p = 1} ^ {P} \\left(E C M _ {t, f} \\odot \\hat {C} _ {t, f}\\right) _ {b, p,:,:}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.87, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where threshold \\(\\hat{C}_{t,f}\\) is optimized by minimizing Eq. (7) in an end-to-end manner." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28336" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.483, + 0.158 + ], + "angle": 0, + "content": "Table 1. Step-by-step validation of mutual boosting between threshold modeling (TM) and 3D reconstruction (3D Rec.). Abbreviations: \"Fro.\" : \"Frozen\"; \"Ft.\" : \"Fine-tuning\". Note that TM is evaluated by MSE between the estimated and GT thresholds from the simulator." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.163, + 0.482, + 0.26 + ], + "angle": 0, + "content": "
TM for 3D Rec.3D Rec. for TM
Stage1Stage23D Rec.Stage1Stage2TM
TM3D Rec. (Fro. C)PSNR↑3D Rec.TM (Fro. GS)MSE↓ (×10-4)
×✓24.46×✓8.317
✓✓26.63✓✓7.077
" + }, + { + "type": "table_caption", + "bbox": [ + 0.194, + 0.263, + 0.383, + 0.274 + ], + "angle": 0, + "content": "Joint TM and 3D Rec. Optimization" + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.275, + 0.48, + 0.342 + ], + "angle": 0, + "content": "
Stage1Stage23D Rec.TM
TM3D Rec. & TM (Ft. C̅ & Ft. GS)PSNR↑MSE↓ (×10-4)
✓✓28.016.322
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.351, + 0.483, + 0.44 + ], + "angle": 0, + "content": "We observe that accurate threshold modeling improves 3D reconstruction quality. As shown in \"TM for 3D Rec.\" in Tab. 1, our RGB-assisted threshold optimization approach benefits threshold estimation, thus significantly enhancing 3D-GS reconstruction quality and achieving a 2.17 dB PSNR improvement." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.441, + 0.483, + 0.593 + ], + "angle": 0, + "content": "GS-boosted Threshold Refinement. While RGB frames facilitate threshold estimation, their effectiveness is constrained by low frame rate. As illustrated in Fig. 3 (b), sparse RGB frames lead to longer integration intervals, reducing supervision quality and threshold estimation accuracy. To overcome this limitation, we found that once a 3D-GS is trained first by Eq. (4) and Eq. (5), it can be used to render intermediate frames as additional pseudosupervision. Specifically, we freeze the trained 3D-GS and reuse Eq. (4) to enhance Eq. (7) for optimizing threshold \\(\\hat{C}\\)" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.6, + 0.483, + 0.625 + ], + "angle": 0, + "content": "\\[\n\\hat {C} ^ {*} = \\arg \\min _ {\\hat {C}} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}}\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.483, + 0.767 + ], + "angle": 0, + "content": "We observe that the incorporation of 3D-GS significantly enhances threshold estimation accuracy. The underlying reason is that events may provide unreliable supervision in some regions with inaccurate thresholds or noise, whereas 3D-GS can correct these errors via geometric consistency. As demonstrated in \"3D Rec. for TM\" part in Tab. 1, using trained and frozen 3D-GS for threshold modeling substantially reduces MSE, leading to more precise threshold estimation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.769, + 0.483, + 0.829 + ], + "angle": 0, + "content": "Joint Threshold and GS Optimization. Having demonstrated the mutual benefits between threshold modeling and 3D reconstruction, we propose jointly optimizing both threshold \\(\\hat{C}\\) and 3D Gaussians \\(GS\\) through" + }, + { + "type": "equation", + "bbox": [ + 0.113, + 0.837, + 0.483, + 0.862 + ], + "angle": 0, + "content": "\\[\n\\hat {C} ^ {*}, G S ^ {*} = \\arg \\min _ {\\hat {C}, G S} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We observe that this joint optimization enables a beneficial cycle where optimized thresholds enhance event su" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.089, + 0.885, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.262, + 0.905, + 0.291 + ], + "angle": 0, + "content": "Figure 4. The effect of dynamic-static decomposition strategy, which improves the rendering quality of dynamic regions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.308, + 0.905, + 0.383 + ], + "angle": 0, + "content": "pervision for 3D-GS, while improved 3D-GS refines threshold estimates through geometric consistency. As shown in \"Joint TM and 3D Rec. Optimization\" part in Tab. 1, this approach achieves superior threshold modeling and reconstruction quality." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.386, + 0.905, + 0.447 + ], + "angle": 0, + "content": "In summary, our optimization proceeds in two stages: first optimizing the threshold using \\(\\mathcal{L}_{thres}\\), then jointly optimizing both threshold and 3D-GS using all three losses \\(\\mathcal{L}_{thres},\\mathcal{L}_{event}\\), and \\(\\mathcal{L}_{rgb}\\)." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.462, + 0.783, + 0.479 + ], + "angle": 0, + "content": "3.3. Dynamic-static Decomposition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.906, + 0.637 + ], + "angle": 0, + "content": "Dynamic scenes typically contain substantial static regions (e.g., tables, walls) that require no deformation. Unlike existing methods [1, 12, 14, 23, 24, 42, 46] that use dynamic Gaussians throughout, we separately model dynamic and static regions with corresponding Gaussian types. This decomposition offers dual benefits: accelerated rendering by bypassing deformation field computation for static Gaussians, and enhanced deformation fidelity through focused MLP capacity optimization for dynamic regions, as demonstrated in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.906, + 0.701 + ], + "angle": 0, + "content": "The key challenge lies in accurately initializing dynamic Gaussians in dynamic regions and static Gaussians in static regions. We address this through a proposed dynamic-static decomposition (DSD) strategy, as illustrated in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.906, + 0.839 + ], + "angle": 0, + "content": "Dynamic-static Decomposition on 2D. We leverage the inherent inability of static Gaussians in representing motion to decompose dynamic and static regions in 2D images. During the first \\(3k\\) iterations, we perform scene reconstruction using only static Gaussians for initialization. This naturally results in poor reconstruction in dynamic regions while achieving high fidelity in static areas (illustrated in Fig. 5 (a)). This distinct performance difference enables decomposition of training images into dynamic and static regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Specifically, using a pretrained VGG19 [35] network \\(\\mathcal{F}_{\\phi}\\) we extract multi-scale features from both rendered image \\(\\hat{I}(t)\\) and ground truth \\(I(t)\\). The cosine similarities computed at each scale are upsampled to a uniform resolution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28337" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.089, + 0.31, + 0.102 + ], + "angle": 0, + "content": "Dynamic-static Decomposition on 2D" + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.09, + 0.669, + 0.102 + ], + "angle": 0, + "content": "Decomposition Correspondence from 2D to 3D" + }, + { + "type": "image_caption", + "bbox": [ + 0.788, + 0.09, + 0.874, + 0.102 + ], + "angle": 0, + "content": "Joint Rendering" + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.104, + 0.904, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.306, + 0.908, + 0.349 + ], + "angle": 0, + "content": "Figure 5. Overview of dynamic-static decomposition strategy. First, we decompose dynamic and static regions in 2D images based on the inherent inability of static Gaussians to represent motions. Next, we establish a correspondence to extend 2D decomposition to 3D Gaussians. Finally, the decomposed dynamic and static Gaussians are jointly rendered to reconstruct the complete dynamic scene." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.36, + 0.408, + 0.376 + ], + "angle": 0, + "content": "and averaged to generate a fused similarity map" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.387, + 0.482, + 0.439 + ], + "angle": 0, + "content": "\\[\n\\operatorname {S i m} = \\sum_ {l} U p \\left(\\frac {\\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\cdot \\mathcal {F} _ {\\phi} ^ {l} (I (t))}{\\left\\| \\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\right\\| \\left\\| \\mathcal {F} _ {\\phi} ^ {l} (I (t)) \\right\\|}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.526 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}_{\\phi}^{l}(\\cdot)\\) represents the \\(l\\)-th layer output of VGG19, and \\(Up(\\cdot)\\) indicates bilinear upsampling. The histogram of the resulting similarity map exhibits a bimodal distribution as shown in Fig. 5 (b), enabling dynamic region mask generation through Otsu's [27] method" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.539, + 0.482, + 0.557 + ], + "angle": 0, + "content": "\\[\nM a s k = \\mathbf {1} _ {\\text {S i m} < \\text {O t s u} (\\text {S i m})}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.483, + 0.613 + ], + "angle": 0, + "content": "where \\(\\mathbf{1}_{\\{\\cdot\\}}\\) denotes the indicator function, which returns 1 if the condition is true. The mask is then multiplied by the ground truth image to extract the dynamic region." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.719 + ], + "angle": 0, + "content": "Decomposition Correspondence from 2D to 3D. To extend 2D dynamic-static decomposition to 3D Gaussians, we establish view-independent correspondences by leveraging depth information from 3D-GS rendering. By unprojecting pixels from masked dynamic regions across multiple views and merging the resulting 3D points, we obtain a comprehensive representation of dynamic regions in 3D space." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Next, we map merged points to dynamic Gaussians based on spatial proximity. Each point expands spherically with radius \\( r \\) to form a 3D volume (Fig. 5 (c)), initially classifying enclosed Gaussians as dynamic and others as static. To overcome potential decomposition inaccuracies and radius sensitivity, we implement a buffer-based soft decomposition strategy using two radii, \\( r_1 \\) and \\( r_2 \\). Gaussians within \\( r_1 \\) are marked as dynamic, beyond \\( r_2 \\) as static, while those between are pruned to create a buffer zone. This strategy enables 3D-GS to optimize decomposition boundaries through adaptive density control (ADC) [17], enhancing both rendering quality and speed. As demonstrated in Fig. 12, the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.36, + 0.905, + 0.388 + ], + "angle": 0, + "content": "strategy also exhibits improved robustness to radius parameter selection." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.45 + ], + "angle": 0, + "content": "It should be noted that, our DSD method is performed only once during the entire training process and requires only about one minute, introducing minimal overhead to the training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.907, + 0.572 + ], + "angle": 0, + "content": "Joint Rendering of Dynamic and Static Gaussians. With the decomposed dynamic and static Gaussians, we jointly render the entire dynamic scene. Particularly, a deformation field [46] learns to map dynamic Gaussians from canonical space to arbitrary time. Taking time \\( t \\) and the center position \\( \\mu \\) of dynamic Gaussians as inputs, the deformation field outputs the displacement of their position \\( \\delta_{\\mu} \\), rotation \\( \\delta_r \\), and scaling \\( \\delta_s \\)" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.581, + 0.905, + 0.599 + ], + "angle": 0, + "content": "\\[\n\\left(\\delta_ {\\mu}, \\delta_ {r}, \\delta_ {s}\\right) = \\mathcal {F} _ {\\theta} (\\gamma (s g (\\mu)), \\gamma (t)), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.608, + 0.906, + 0.655 + ], + "angle": 0, + "content": "where \\( sg(\\cdot) \\) indicates a stop-gradient operation and \\( \\gamma (\\cdot) \\) denotes the positional encoding [46]. Then, the deformed dynamic gaussians can be addressed as" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.663, + 0.905, + 0.681 + ], + "angle": 0, + "content": "\\[\n\\left(\\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {r} ^ {\\prime}, \\boldsymbol {s} ^ {\\prime}\\right) = \\left(\\boldsymbol {\\mu} + \\delta_ {\\boldsymbol {\\mu}}, \\boldsymbol {r} + \\delta_ {\\boldsymbol {r}}, \\boldsymbol {s} + \\delta_ {\\boldsymbol {s}}\\right). \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.69, + 0.906, + 0.737 + ], + "angle": 0, + "content": "Finally, static Gaussians bypass the deformation field and merge with the deformed dynamic Gaussians as inputs to the rasterizer, enabling high-frame-rate dynamic rendering." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.749, + 0.64, + 0.765 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.722, + 0.79 + ], + "angle": 0, + "content": "4.1. Experimental Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Datasets. Current datasets for event-based dynamic scene reconstruction are highly limited, with only three synthetic and three real-world scenes [25], all unpublished. Notably, the only three publicly available real-world scenes [25] were captured with a static camera, making novel view evaluation infeasible. To facilitate future research, we build the first event-inclusive 4D benchmark featuring 8 synthetic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28338" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.118 + ], + "angle": 0, + "content": "Table 2. Quantitative results on our synthetic dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events. The best and second-best result are marked in bold and underlined, respectively." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.12, + 0.905, + 0.348 + ], + "angle": 0, + "content": "
MethodLegoHotdogMaterialsMusic box
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]23.600.9180.08822330.010.9510.06426028.070.9670.06126219.200.9050.122239
TiNeuVox [7]22.390.8910.0710.5330.810.9530.0350.4926.630.9380.0540.5220.450.8310.1520.62
K-Planes [9]24.550.9310.0352.3431.360.9580.0162.3530.620.9760.0092.2920.770.8580.0712.39
4D-GS [42]26.300.9370.07210433.480.9650.05213230.400.9790.05411124.060.9370.07164
Deformable-3DGS [46]23.790.9230.0537332.910.9620.01713234.000.9860.0049122.080.9240.05251
Event-4DGS28.000.9430.0405434.610.9690.0199635.600.9890.0067428.580.9500.04342
Ours31.850.9670.01818936.150.9740.01324138.020.9930.00324030.780.9630.02992
MethodCelestial globeFanWater wheelMan
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]19.050.9150.11018221.260.8910.11827019.430.8870.10921520.790.8700.114210
TiNeuVox [7]13.620.7360.2900.6219.900.8890.1070.5317.030.8500.1470.5622.810.8870.0710.50
K-Planes [9]15.490.8570.0882.4622.100.9090.0622.3820.960.9200.0462.4121.020.8570.0732.33
4D-GS [42]20.970.9420.0725225.050.9360.08010820.960.9170.0777723.800.9140.07664
Deformable-3DGS [46]23.070.9620.0364124.660.9290.0519020.790.9120.0514323.060.9060.05137
Event-4DGS24.300.9480.0453627.660.9490.0417126.340.9320.0523025.550.9210.06333
Ours28.830.9760.0207330.180.9640.02516828.470.9500.03311228.210.9430.037129
" + }, + { + "type": "table_caption", + "bbox": [ + 0.099, + 0.351, + 0.899, + 0.365 + ], + "angle": 0, + "content": "Table 3. Quantitative results on our real-world dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.367, + 0.903, + 0.449 + ], + "angle": 0, + "content": "
MethodExcavatorJeepFlowersEagle
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
4D-GS [42]28.350.9110.11011528.340.8780.0936126.820.8730.1236327.590.9000.128105
Deformable-3DGS [46]26.120.9030.1208126.300.8700.1045226.400.9030.0796427.440.9030.12570
Event-4DGS29.670.9140.0925729.640.9010.0794727.530.9050.0844029.080.8960.10463
Ours31.280.9250.07017930.410.9050.0688928.570.9130.06914931.290.9180.074192
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.461, + 0.483, + 0.506 + ], + "angle": 0, + "content": "and 4 real-world dynamic scenes, encompassing diverse complexities, intricate structures, and rapid motions, thus enabling effective evaluation of dynamic reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.507, + 0.483, + 0.658 + ], + "angle": 0, + "content": "For synthetic scenes, we use Blender [3] to generate one-second, \\(360^{\\circ}\\) monocular camera rotations, producing thousands of continuous frames per scene. These high-temporal-resolution sequences are processed through ESIM [31] to generate events. For each sequence, we uniformly sample 30 frames (equivalent to 30 FPS) for training, and select intermediate frames as far apart as possible from the training frames for testing. Particularly, both \"Fan\" representing a typical high-speed 4D scene, and \"Man\" featuring large object displacements, present significant challenges." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.794 + ], + "angle": 0, + "content": "For real-world scenes, as shown in Fig. 6, we construct a hybrid camera system consisting of a beam splitter, an event camera (Prophesee Gen4), a frame camera (Basler ace), and a microcontroller (STM32) for outputting synchronization signals. Following [32], we keep the camera system static and place the objects on a motorized optical rotating turntable, which is equivalent to camera motion. Following prior work [25], we downsample the original high-FPS video for training and use intermediate frames for testing." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Baselines. For RGB-only settings, we benchmarked our method against the representative NeRF baselines K-Planes [9] and TiNeuVox [7], along with Gaussians baselines 3D-GS [17], 4D-GS [42], and Deformable-3DGS [46]. For event-assisted settings, DE-NeRF [25] is the only baseline; however, it could not be directly compared, as its code was still closed-source. DE-NeRF relies on NeRF's volume ren" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.459, + 0.872, + 0.585 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.588, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Figure 6. Real-world data acquisition setup (left) and our hybrid camera system (right)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.628, + 0.907, + 0.719 + ], + "angle": 0, + "content": "dering techniques [26], leading to predictably slow rendering speeds. Moreover, its reconstruction quality is also predictably limited due to the absence of threshold modeling for events. To provide a comparable baseline, we introduce Event-4DGS, an extension of Deformable-3DGS [46] that incorporates the event rendering loss in Eq. (4)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.719, + 0.907, + 0.765 + ], + "angle": 0, + "content": "Metrics. We evaluate rendering quality using PSNR, SSIM [37], and LPIPS [50] (based on AlexNet) and measure rendering speed in FPS on an NVIDIA RTX 3090 GPU." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.655, + 0.79 + ], + "angle": 0, + "content": "4.2. Comparisons" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Quantitative Results. We report the quantitative results of the comparison on the synthetic and real-world datasets in Tab. 2 and Tab. 3, respectively. Although 4D-GS [42] and Deformable-3DGS [46] achieve relatively higher FPS compared to NeRF baselines [7, 9], their reconstruction quality is limited by the sparsity of the RGB training frames. In contrast, Event-4DGS leverages the rich intermediate mo" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28339" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.133, + 0.088, + 0.861, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.176, + 0.468, + 0.82, + 0.482 + ], + "angle": 0, + "content": "Figure 7. Qualitative comparisons on our synthetic dataset. Please see the supplementary video for details." + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.484, + 0.861, + 0.732 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.733, + 0.825, + 0.747 + ], + "angle": 0, + "content": "Figure 8. Qualitative comparisons on our real-world dataset. Please see the supplementary video for details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.759, + 0.483, + 0.895 + ], + "angle": 0, + "content": "tion and viewpoint information provided by events, significantly outperforming other baselines in reconstruction quality, with an average PSNR improvement of 3.28 dB over Deformable-3DGS across all synthetic scenes. This notable improvement underscores the effectiveness of high-temporal-resolution event cameras for dynamic scene reconstruction. However, Event-4DGS still suffers from threshold variation, whereas our method with GTJM enables accurate threshold modeling and better event super" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.759, + 0.907, + 0.865 + ], + "angle": 0, + "content": "vision, achieving an average PSNR improvement of 2.73 dB over Event-4DGS on synthetic datasets. Meanwhile, our method maintains exceptionally fast rendering speeds, averaging \\(1.71 \\times\\) faster than 4D-GS on synthetic datasets. In summary, our method enjoys both the highest rendering quality and exceptional rendering speed on both synthetic and real-world datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Qualitative Results. For a more visual assessment, we present qualitative results on the synthetic and real-world" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "28340" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 4. Ablation studies on synthetic dataset. For real-world ablation studies, please refer to the supplementary document." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.118, + 0.481, + 0.212 + ], + "angle": 0, + "content": "
MethodPSNR↑SSIM↑LPIPS↓FPS↑
w/o GTJM29.390.9560.034153
w/o Joint Optimization in GTJM30.870.9630.026152
w/o DSD30.780.9610.02657
w/o Buffer-based Soft Decomposition31.020.9630.025138
Full31.560.9660.022156
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.291 + ], + "angle": 0, + "content": "datasets in Fig. 7 and Fig. 8, respectively. These comparisons highlight the capability of our method to deliver high-fidelity dynamic scene modeling. Notably, our method effectively captures intricate motion details, while other baselines exhibit structural deficiencies and distortions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.293, + 0.484, + 0.429 + ], + "angle": 0, + "content": "Dynamic Blurry Scene Comparisons. Motion blur is another common challenge in dynamic scenes. To address this, we extend both baselines and our method with blur loss and EDI from [48], and build blurry scenes for evaluations. Fig. 9 shows that, by leveraging events' deblurring advantage, our method outperforms Deformable-3DGS by 4.79 dB in PSNR, achieving the best results. For detailed methods and quantitative results, see the supplementary document." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.439, + 0.245, + 0.455 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.462, + 0.484, + 0.658 + ], + "angle": 0, + "content": "GS-threshold Joint Modeling. Using a constant threshold fails to properly neutralize opposing polarity events during accumulation, resulting in motion trajectory artifacts as shown in Fig. 10 (b). These artifacts, when used for Gaussian supervision, produce undesirable purple haze in rendered outputs, such as the Event-4DGS results in Fig. 7. Our RGB-assisted threshold estimation significantly reduces these artifacts (Fig. 10 (c)), while subsequent joint threshold and GS optimization effectively eliminates remaining distortions (Fig. 10 (d)). As demonstrated in Tab. 4, this improved event supervision yields a \\(2.17\\mathrm{dB}\\) average PSNR improvement across all scenes, validating our GTJM strategy's effectiveness in handling threshold variations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Dynamic-static Decomposition. Our DSD method successfully identifies dynamic regions of varying sizes and geometries, as demonstrated in Fig. 11. Modeling the entire scene with dynamic Gaussians without DSD misallocates deformation field capacity to static regions, compromising dynamic region reconstruction quality as shown in Fig. 4. Quantitative results in Tab. 4 demonstrate that using DSD improves the average PSNR by \\(0.78\\mathrm{dB}\\) and accelerates the rendering speed to 2.74 times the original FPS. This underscores DSD's crucial role in achieving both high-fidelity dynamic scene reconstruction and efficient rendering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Buffer-based Soft Decomposition. Our buffer-based soft decomposition enables adaptive optimization of decomposition boundaries, yielding a 0.54 dB improvement in average PSNR (Tab. 4). Sensitivity analysis reveals that reconstruction quality stabilizes when buffer size \\((r_2 - r_1)\\)" + }, + { + "type": "image", + "bbox": [ + 0.525, + 0.09, + 0.603, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.091, + 0.648, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.652, + 0.091, + 0.725, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.727, + 0.091, + 0.772, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.091, + 0.849, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.852, + 0.091, + 0.894, + 0.154 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.526, + 0.156, + 0.894, + 0.17 + ], + "angle": 0, + "content": "Figure 9. Extended comparisons on the dynamic blurry scene." + }, + { + "type": "image", + "bbox": [ + 0.523, + 0.172, + 0.616, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.313, + 0.585, + 0.321 + ], + "angle": 0, + "content": "(a) GT" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.172, + 0.71, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.628, + 0.313, + 0.698, + 0.321 + ], + "angle": 0, + "content": "(b) w/o GTJM" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.172, + 0.803, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.704, + 0.313, + 0.811, + 0.321 + ], + "angle": 0, + "content": "c) w/o Joint Optimization" + }, + { + "type": "image", + "bbox": [ + 0.803, + 0.172, + 0.895, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.83, + 0.313, + 0.864, + 0.321 + ], + "angle": 0, + "content": "(d) Full" + }, + { + "type": "image_caption", + "bbox": [ + 0.734, + 0.322, + 0.775, + 0.331 + ], + "angle": 0, + "content": "in GTJM" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.333, + 0.905, + 0.361 + ], + "angle": 0, + "content": "Figure 10. The effect of GS-threshold joint modeling strategy, which eliminates event artifacts caused by threshold variations." + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.362, + 0.616, + 0.462 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.568, + 0.464, + 0.665, + 0.473 + ], + "angle": 0, + "content": "Lego Hotdog" + }, + { + "type": "image_caption", + "bbox": [ + 0.617, + 0.463, + 0.665, + 0.473 + ], + "angle": 0, + "content": "Hotdog" + }, + { + "type": "image_caption", + "bbox": [ + 0.668, + 0.463, + 0.738, + 0.473 + ], + "angle": 0, + "content": "Materials" + }, + { + "type": "image_caption", + "bbox": [ + 0.74, + 0.463, + 0.763, + 0.472 + ], + "angle": 0, + "content": "Cel" + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.463, + 0.812, + 0.472 + ], + "angle": 0, + "content": "estial globe" + }, + { + "type": "image_caption", + "bbox": [ + 0.814, + 0.463, + 0.859, + 0.472 + ], + "angle": 0, + "content": "Man" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.477, + 0.905, + 0.505 + ], + "angle": 0, + "content": "Figure 11. Rendering results of dynamic and static Gaussians separated by our dynamic-static decomposition strategy." + }, + { + "type": "image", + "bbox": [ + 0.544, + 0.507, + 0.875, + 0.637 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.638, + 0.874, + 0.651 + ], + "angle": 0, + "content": "Figure 12. Sensitivity analysis on buffer size \\((r_2 - r_1)\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.662, + 0.906, + 0.738 + ], + "angle": 0, + "content": "exceeds approximately 12 basic units (normalized by average inter-Gaussian distance to account for scene variations), as shown in Fig. 12. This stability demonstrates the robustness of our DSD method through adaptive boundary search, highlighting the effectiveness of the buffer-based strategy." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.754, + 0.634, + 0.77 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In this paper, we present an event-boosted deformable 3D Gaussian framework for high-quality dynamic scene reconstruction. Our GS-threshold joint modeling effectively addresses threshold variation challenges, enabling reliable event supervision. The proposed dynamic-static decomposition method enhances both rendering efficiency and reconstruction quality through optimized resource allocation between static and dynamic regions." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "28341" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.26, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.192 + ], + "angle": 0, + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001, and the Fundamental Research Funds for the Central Universities under Grant WK210000059." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.207, + 0.188, + 0.222 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.232, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[1] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In ECCV, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.289, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[3] Blender. A 3d modelling and rendering package. https://www.blender.org/, 2018.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.375, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[4] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.483, + 0.445 + ], + "angle": 0, + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[6] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In ICIP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.545 + ], + "angle": 0, + "content": "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia, 2022. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.548, + 0.484, + 0.588 + ], + "angle": 0, + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.591, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.647, + 0.484, + 0.729 + ], + "angle": 0, + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[11] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.484, + 0.857 + ], + "angle": 0, + "content": "[12] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 35(4):3119-3133, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[13] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR workshop, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.232, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.151, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[15] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In WACV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.181, + 0.906, + 0.221 + ], + "angle": 0, + "content": "[16] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.906, + 0.279 + ], + "angle": 0, + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):139-1, 2023. 1, 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.906, + 0.336 + ], + "angle": 0, + "content": "[18] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.339, + 0.906, + 0.395 + ], + "angle": 0, + "content": "[19] Chenghan Li, Christian Brandli, Raphael Berner, Hongjie Liu, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. Design of an rgb color vga rolling and global shutter dynamic and active-pixel vision sensor. In ISCAS, 2015. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.398, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[20] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. In WACV, 2025. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.454, + 0.905, + 0.495 + ], + "angle": 0, + "content": "[21] Yutong Liu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. S2d-lfe: Sparse-to-dense light field event generation. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[22] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[23] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.6, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 1, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.643, + 0.905, + 0.684 + ], + "angle": 0, + "content": "[25] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, 2023. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.687, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.906, + 0.8 + ], + "angle": 0, + "content": "[27] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62-66, 1979. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "28342" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[30] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[32] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.204, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[33] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[34] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In ECCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.344 + ], + "angle": 0, + "content": "[35] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.399 + ], + "angle": 0, + "content": "[36] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.482, + 0.455 + ], + "angle": 0, + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.455, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[38] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.496, + 0.482, + 0.538 + ], + "angle": 0, + "content": "[39] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Boosting event stream super-resolution with a recurrent neural network. In ECCV, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[40] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.58, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[41] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.692 + ], + "angle": 0, + "content": "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[43] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. In MLSP, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.789 + ], + "angle": 0, + "content": "[44] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In CoRL, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[45] Wenhao Xu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Ceia: Clip-based event-image alignment for open-world event-based understanding. In ECCV workshop, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[46] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.135 + ], + "angle": 0, + "content": "[47] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.134, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[48] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.19, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[49] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: a structured and explicit radiance representation for 3d generative modeling. In NeurIPS, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.248, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "28343" + } + ] +] \ No newline at end of file diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_origin.pdf b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..9dd17142571c0490a6de25484105d6b9f2b5ab63 --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/0f6daa46-496f-4815-9cc1-8b104f1b964b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a3b7fd24620283e2b29f79a8ef1fc584709372e61b7bba5eb08a07f6e7c0b1c5 +size 7050828 diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/full.md b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d72f2e1d4d753b47b15c9ffeea2ae195d52aea7f --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/full.md @@ -0,0 +1,365 @@ +# Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction + +Wenhao Xu1 Wenming Weng1 Yueyi Zhang2, * Ruikang Xu1 Zhiwei Xiong1, * +1University of Science and Technology of China 2MiroMind + +wh-xu@mail.ustc.edu.cn, yueyi.zhang@miromind.ai, zwxiong@ustc.edu.cn + +https://github.com/wh-xul/E-D3DGS + +![](images/701fea225884661de94c68de553537d4cafda88441bab97a8cd5120039349391.jpg) +Figure 1. Left: Quality comparison with the baselines 4D-GS [42], Event-4DGS (the event-extended version of [46]), and our variants, highlighting the superior rendering quality of our method. Our GS-threshold joint modeling (GTJM) effectively eliminates event-induced artifacts (particularly the purple haze), while our dynamic-static decomposition (DSD) improves the quality of dynamic regions. Middle: Separate rendering of dynamic and static Gaussians from our DSD. Right: The scatter plot illustrates our method's ability to achieve both high fidelity and fast rendering, where dot radii correspond to different resolutions $(400\times 400,600\times 600,$ and $800\times 800)$ . + +# Abstract + +Deformable 3D Gaussian Splating (3D-GS) is limited by missing intermediate motion information due to the low temporal resolution of RGB cameras. To address this, we introduce the first approach combining event cameras, which capture high-temporal-resolution, continuous motion data, with deformable 3D-GS for dynamic scene reconstruction. We observe that threshold modeling for events plays a crucial role in achieving high-quality reconstruction. Therefore, we propose a GS-Threshold Joint Modeling strategy, creating a mutually reinforcing process that greatly improves both 3D reconstruction and threshold modeling. Moreover, we introduce a Dynamic-Static Decomposition strategy that first identifies dynamic areas by exploiting the inability of static Gaussians to represent motions, then applies a buffer-based soft decomposition to separate dynamic and static areas. This strategy accelerates rendering by avoiding unnecessary deformation in static areas, and focuses on dynamic areas to enhance fidelity. Additionally, we contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance. + +# 1. Introduction + +Dynamic scene reconstruction and novel view synthesis are essential for immersive applications in virtual/augmented reality and entertainment [20, 24, 34, 42, 46, 47, 49]. While Neural Radiance Fields (NeRF) [2, 26, 36] offer unprecedented photorealism, they are constrained by slow training and rendering speeds. Despite recent advances in optimization techniques [5, 7-9], real-time rendering remains elusive. Recently, 3D Gaussian Splatting (3D-GS) [17] addresses this limitation through efficient differentiable rasterization, yet existing dynamic extensions [24, 42, 46, 47] are constrained by inherent limitations of RGB cameras, including low frame rates and motion blur. + +In this paper, we present event-boosted deformable 3D Gaussians for dynamic scene reconstruction. Event cameras [10, 19, 38, 39], with their microsecond-level temporal resolution, can provide continuous motion information and near-infinite viewpoints that traditional RGB cameras often fail to capture. These advantages make event cameras particularly valuable for dynamic scene reconstruction. + +However, integrating events into 3D scene reconstruction faces new challenges. Specifically, event supervision for 3D-GS relies on an accurate event generation model + +[4, 18], where the threshold undergoes complex variations across polarity, space, and time [13, 18]. Previous methods [4, 6, 18, 25, 30, 32, 43, 44, 48] adopt a constant threshold, yet this simplification significantly degrades the quality of event supervision. While recent works [15, 22] attempt to model threshold variations using event data alone, they achieve limited success due to the inherent binary nature of events, which only indicate brightness change directions. To address this challenge, we propose a novel GS-threshold joint modeling strategy. First, we leverage the brightness change values from RGB frames to supervise threshold optimization. Second, since the sparsity of RGB frames weakens supervision, we use 3D-GS rendered results as pseudo-intermediate frames to enhance the supervision. This finally creates a mutually reinforcing process where RGB-optimized threshold enables better event supervision for 3D-GS, while improved 3D-GS in turn provides accurate geometric constraints for threshold refinement. + +Furthermore, we observe that existing dynamic 3D-GS methods inefficiently use dynamic Gaussians solely to model both static and dynamic regions [1, 12, 14, 23, 24, 42, 46]. This unified treatment leads to reduced rendering speed, wasted deformation field capacity, and degraded reconstruction quality. While some methods have explored dynamic-static decomposition, they are limited by either inaccurate dynamic Gaussians initialization [20] or constraints in multi-view scenarios [34]. To address these limitations, we propose a novel dynamic-static decomposition strategy that first identifies dynamic regions based on the inherent inability of static Gaussians to represent motions, and then employs a buffer-based soft decomposition to adaptively search for the optimal decomposition boundary. This decomposition not only accelerates rendering by eliminating unnecessary deformation computations in static regions, but also enhances reconstruction quality by focusing the deformation field exclusively on dynamic regions. + +Our main contributions can be summarized as follows: + +- We present the first method integrating event cameras with deformable 3D-GS for dynamic scene reconstruction, enabling high-fidelity and fast rendering. +- We propose a novel GS-threshold joint modeling strategy that combines RGB-assisted initial estimation with GS-boosted refinement, creating a mutually reinforcing process that significantly improves both threshold modeling and 3D reconstruction. +- We introduce an effective dynamic-static decomposition strategy that not only accelerates rendering through selective deformation computation but also enhances reconstruction quality by focusing on dynamic regions. +- We contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance. + +# 2. Related Work + +Neural Rendering for Dynamic Scenes. Neural rendering techniques have revolutionized dynamic scene reconstruction in recent years. Pioneering works like D-NeRF [29] and Nerfies [28] extend Neural Radiance Fields [26] through deformation field, mapping observations into a canonical space for modeling non-rigid motion. Despite their impressive reconstruction quality, these methods are constrained by extensive computational demands due to dense MLP evaluations during training and rendering. Various acceleration strategies have been proposed to address these limitations. K-Planes [9] introduces an efficient explicit representation using six feature planes, while Tensor4D [33] and DTensorRF [16] employ tensor decomposition techniques to achieve compact spatiotemporal encoding. A significant breakthrough came with 3D Gaussian Splatting (3D-GS) [17], which leverages efficient differentiable rasterization for real-time rendering. This advancement has spawned several dynamic scene extensions, including 4D-GS [42], Deformable-3DGS [46], and related works [24, 34], which successfully achieve real-time rendering for high-quality dynamic scene reconstruction. + +Event-based Neural Rendering. The integration of neural representations with event-based 3D reconstruction has emerged as a promising research direction. Pioneering works like EventNeRF [32], E-NeRF [18], and Ev-NeRF [15] first demonstrated the potential of pure event-based static scene reconstruction, albeit with different assumptions about event camera characteristics. To further improve reconstruction quality, E2NeRF [30] and Ev-DeblurNeRF [4] incorporated blurry RGB images alongside events. A significant milestone was achieved by DE-NeRF [25], which pioneered the combination of events and RGB frames for dynamic scene reconstruction. The recent advent of 3D Gaussian Splatting [17] has catalyzed new developments in event-based methods. While Ev-GS [43] adapted the pure event-based paradigm to 3D-GS, subsequent works including E2GS [6], EaDeblur-GS [41], and Event3DGS [44] primarily addressed deblurring challenges. To the best of our knowledge, our work is the first to integrate events with deformable 3D-GS for dynamic scene reconstruction. Our novel method and benchmark, specifically tailored for events and dynamic scenes, highlight significant and independent contributions that set our work apart. + +# 3. Method + +# 3.1. Event Cameras for 3D-GS + +3D Gaussian Splatting Preliminary. 3D Gaussian Splatting (3D-GS) [17] represents a scene as anisotropic 3D Gaussians, each characterized by a covariance matrix $\Sigma$ and center position $\mu$ : $GS(\mathbf{x}) = e^{-\frac{1}{2} (\mathbf{x} - \mu)^T\boldsymbol{\Sigma}^{-1}(\mathbf{x} - \mu)}$ . The covariance matrix $\Sigma$ is parameterized using scaling matrix $S$ + +![](images/aad825284d12801d28d74671a61736e0a28da0b29ef58cf084d5c3ffc6f4ad44.jpg) +Figure 2. Overview of GS-threshold joint modeling strategy. $\mathcal{L}_{rgb}$ optimizes 3D-GS, $\mathcal{L}_{thres}$ optimizes the threshold, and $\mathcal{L}_{event}$ jointly optimizes both 3D-GS and threshold. + +and rotation matrix $R$ to ensure positive semi-definiteness: $\Sigma = RSS^T R^T$ . Each Gaussian is further defined by spherical harmonic coefficients $\mathcal{C}$ and opacity $\sigma$ . Final pixel colors $c$ are computed using differentiable tile-based rasterization + +$$ +c = \sum_ {i \in N} c _ {i} \alpha_ {i} \prod_ {j = 1} ^ {i - 1} (1 - \alpha_ {j}), \tag {1} +$$ + +where $c_{i}$ denotes spherical harmonic color, and $\alpha_{i}$ combines opacity $\sigma$ with projected $GS(\mathbf{x})$ . + +Event/RGB Rendering Loss. Event cameras [10, 19, 21, 40, 45] are novel sensors, that asynchronously capture pixel-wise brightness changes with microsecond-level temporal resolution. Their high temporal precision enables capturing crucial motions between RGB frames and provides near-infinite viewpoint supervision, making them ideal for monocular dynamic scene reconstruction. + +Each event is represented as $e_{x,y}(\tau) = p\delta(\tau)$ , where $(x,y)$ is pixel position, $\tau$ is timestamp, $p \in \{+1, -1\}$ indicates brightness change direction relative to threshold $C$ , and $\delta(t)$ is a unit integral impulse function. Omitting pixel subscripts, the brightness change over interval $\triangle t$ can be formulated as + +$$ +E (t, t + \Delta t) = \int_ {t} ^ {t + \Delta t} C \cdot e (\tau) d \tau . \tag {2} +$$ + +This change can also be estimated from rendered brightness + +$$ +\hat {E} (t, t + \triangle t) \stackrel {\mathrm {d e f}} {=} \log (\hat {I} (t + \triangle t)) - \log (I (t)), \quad (3) +$$ + +where $\hat{I}$ and $I$ respectively denote 3D-GS rendered and ground truth brightnesses. The event rendering loss is + +$$ +\mathcal {L} _ {\text {e v e n t}} = \left\| E (t, t + \triangle t) - \hat {E} (t, t + \triangle t) \right\| _ {2} ^ {2}. \tag {4} +$$ + +Similarly, we utilize the RGB rendering loss [17] combining L1 and D-SSIM losses as + +$$ +\mathcal {L} _ {r g b} = \left(1 - \lambda_ {s}\right) \left\| \hat {I} (t) - I (t) \right\| _ {1} + \lambda_ {s} \mathcal {L} _ {D - S S I M} (\hat {I} (t), I (t)), \tag {5} +$$ + +![](images/a190dd5d3358b89e676b2579f01d3827ab2e4bd5966c4c0772fccca9ecb5bc61.jpg) +(a) + +![](images/f81b37a6d672990ea6842acdda2747088896f6a9e38752b03822396a9307ab57.jpg) +(b) +Figure 3. (a) The effect of different ranges of threshold variation for 3D reconstruction. (b) The effect of different number of RGB frames for threshold estimation. + +where $\lambda_{s}$ is a weighting factor that controls the balance. 3D-GS is optimized to minimize $\mathcal{L}_{event}$ and $\mathcal{L}_{rgb}$ jointly + +$$ +G S ^ {*} = \arg \min _ {G S} \left(\mathcal {L} _ {\text {e v e n t}} + \mathcal {L} _ {\text {r g b}}\right). \tag {6} +$$ + +# 3.2. GS-threshold Joint Modeling + +As shown in Eq. (2), the threshold $C$ critically affects event integration and supervision quality. While existing methods [4, 18, 25, 30, 32] typically assume a constant threshold, real event cameras exhibit threshold variations across polarity, space, and time [13, 18]. Fig. 3 (a) demonstrates how increasing threshold variation significantly degrades reconstruction quality. To model threshold variations, we propose a GS-threshold joint modeling (GTJM) strategy (see Fig. 2), consisting of RGB-assisted threshold estimation and GS-boosted threshold refinement. + +RGB-assisted Threshold Estimation. Recent works [15, 22] attempt to model threshold variations solely from event data but face significant limitations due to the difficulty of inferring brightness changes from events, which only indicate the direction of change. We propose to leverage brightness change values from RGB frames for robust threshold estimation. Given two RGB frames $I(t)$ and $I(f)$ at times $t$ and $f$ , we define the threshold modeling loss as + +$$ +\mathcal {L} _ {\text {t h r e s}} = \left\| E _ {\text {t h r e s}} (t, f) - \hat {E} _ {\text {t h r e s}} (t, f) \right\| _ {2} ^ {2}, \tag {7} +$$ + +where $\hat{E}_{thres}(t,f) := \int_t^f \hat{C} \cdot e(\tau)d\tau$ , $E_{thres}(t,f) = \log(I(f)) - \log(I(t))$ . In practice, we adopt a simple yet fast way to compute $\hat{E}_{thres}(t,f)$ . We first accumulate events to obtain event count maps [11] $ECM_{t,f} \in \mathbb{R}^{B \times P \times H \times W}$ , where $B$ denotes time bins and $P = 2$ corresponds to the event polarity. Using learnable threshold parameters $\hat{C}_{t,f} \in \mathbb{R}^{B \times P \times H \times W}$ , we compute + +$$ +\hat {E} _ {\text {t h r e s}} (t, f) = \sum_ {b = 1} ^ {B} \sum_ {p = 1} ^ {P} \left(E C M _ {t, f} \odot \hat {C} _ {t, f}\right) _ {b, p,:,:}, \tag {8} +$$ + +where threshold $\hat{C}_{t,f}$ is optimized by minimizing Eq. (7) in an end-to-end manner. + +Table 1. Step-by-step validation of mutual boosting between threshold modeling (TM) and 3D reconstruction (3D Rec.). Abbreviations: "Fro." : "Frozen"; "Ft." : "Fine-tuning". Note that TM is evaluated by MSE between the estimated and GT thresholds from the simulator. + +
TM for 3D Rec.3D Rec. for TM
Stage1Stage23D Rec.Stage1Stage2TM
TM3D Rec. (Fro. C)PSNR↑3D Rec.TM (Fro. GS)MSE↓ (×10-4)
×✓24.46×✓8.317
✓✓26.63✓✓7.077
+ +Joint TM and 3D Rec. Optimization + +
Stage1Stage23D Rec.TM
TM3D Rec. & TM (Ft. C̅ & Ft. GS)PSNR↑MSE↓ (×10-4)
✓✓28.016.322
+ +We observe that accurate threshold modeling improves 3D reconstruction quality. As shown in "TM for 3D Rec." in Tab. 1, our RGB-assisted threshold optimization approach benefits threshold estimation, thus significantly enhancing 3D-GS reconstruction quality and achieving a 2.17 dB PSNR improvement. + +GS-boosted Threshold Refinement. While RGB frames facilitate threshold estimation, their effectiveness is constrained by low frame rate. As illustrated in Fig. 3 (b), sparse RGB frames lead to longer integration intervals, reducing supervision quality and threshold estimation accuracy. To overcome this limitation, we found that once a 3D-GS is trained first by Eq. (4) and Eq. (5), it can be used to render intermediate frames as additional pseudosupervision. Specifically, we freeze the trained 3D-GS and reuse Eq. (4) to enhance Eq. (7) for optimizing threshold $\hat{C}$ + +$$ +\hat {C} ^ {*} = \arg \min _ {\hat {C}} \left(\mathcal {L} _ {\text {t h r e s}} + \mathcal {L} _ {\text {e v e n t}}\right). \tag {9} +$$ + +We observe that the incorporation of 3D-GS significantly enhances threshold estimation accuracy. The underlying reason is that events may provide unreliable supervision in some regions with inaccurate thresholds or noise, whereas 3D-GS can correct these errors via geometric consistency. As demonstrated in "3D Rec. for TM" part in Tab. 1, using trained and frozen 3D-GS for threshold modeling substantially reduces MSE, leading to more precise threshold estimation. + +Joint Threshold and GS Optimization. Having demonstrated the mutual benefits between threshold modeling and 3D reconstruction, we propose jointly optimizing both threshold $\hat{C}$ and 3D Gaussians $GS$ through + +$$ +\hat {C} ^ {*}, G S ^ {*} = \arg \min _ {\hat {C}, G S} \left(\mathcal {L} _ {\text {t h r e s}} + \mathcal {L} _ {\text {e v e n t}} + \mathcal {L} _ {\text {r g b}}\right). \tag {10} +$$ + +We observe that this joint optimization enables a beneficial cycle where optimized thresholds enhance event su + +![](images/b741992a7bf752ce457db7a807f1a1ffcc7ea919f4abecbcda42f88090fb6fe9.jpg) +Figure 4. The effect of dynamic-static decomposition strategy, which improves the rendering quality of dynamic regions. + +pervision for 3D-GS, while improved 3D-GS refines threshold estimates through geometric consistency. As shown in "Joint TM and 3D Rec. Optimization" part in Tab. 1, this approach achieves superior threshold modeling and reconstruction quality. + +In summary, our optimization proceeds in two stages: first optimizing the threshold using $\mathcal{L}_{thres}$ , then jointly optimizing both threshold and 3D-GS using all three losses $\mathcal{L}_{thres},\mathcal{L}_{event}$ , and $\mathcal{L}_{rgb}$ . + +# 3.3. Dynamic-static Decomposition + +Dynamic scenes typically contain substantial static regions (e.g., tables, walls) that require no deformation. Unlike existing methods [1, 12, 14, 23, 24, 42, 46] that use dynamic Gaussians throughout, we separately model dynamic and static regions with corresponding Gaussian types. This decomposition offers dual benefits: accelerated rendering by bypassing deformation field computation for static Gaussians, and enhanced deformation fidelity through focused MLP capacity optimization for dynamic regions, as demonstrated in Fig. 4. + +The key challenge lies in accurately initializing dynamic Gaussians in dynamic regions and static Gaussians in static regions. We address this through a proposed dynamic-static decomposition (DSD) strategy, as illustrated in Fig. 5. + +Dynamic-static Decomposition on 2D. We leverage the inherent inability of static Gaussians in representing motion to decompose dynamic and static regions in 2D images. During the first $3k$ iterations, we perform scene reconstruction using only static Gaussians for initialization. This naturally results in poor reconstruction in dynamic regions while achieving high fidelity in static areas (illustrated in Fig. 5 (a)). This distinct performance difference enables decomposition of training images into dynamic and static regions. + +Specifically, using a pretrained VGG19 [35] network $\mathcal{F}_{\phi}$ we extract multi-scale features from both rendered image $\hat{I}(t)$ and ground truth $I(t)$ . The cosine similarities computed at each scale are upsampled to a uniform resolution + +![](images/e58d64e958d6e0898b474036ac15ced927e96aa93508d6c69100542265c1aafb.jpg) +Dynamic-static Decomposition on 2D +Decomposition Correspondence from 2D to 3D +Joint Rendering +Figure 5. Overview of dynamic-static decomposition strategy. First, we decompose dynamic and static regions in 2D images based on the inherent inability of static Gaussians to represent motions. Next, we establish a correspondence to extend 2D decomposition to 3D Gaussians. Finally, the decomposed dynamic and static Gaussians are jointly rendered to reconstruct the complete dynamic scene. + +and averaged to generate a fused similarity map + +$$ +\operatorname {S i m} = \sum_ {l} U p \left(\frac {\mathcal {F} _ {\phi} ^ {l} (\hat {I} (t)) \cdot \mathcal {F} _ {\phi} ^ {l} (I (t))}{\left\| \mathcal {F} _ {\phi} ^ {l} (\hat {I} (t)) \right\| \left\| \mathcal {F} _ {\phi} ^ {l} (I (t)) \right\|}\right), \tag {11} +$$ + +where $\mathcal{F}_{\phi}^{l}(\cdot)$ represents the $l$ -th layer output of VGG19, and $Up(\cdot)$ indicates bilinear upsampling. The histogram of the resulting similarity map exhibits a bimodal distribution as shown in Fig. 5 (b), enabling dynamic region mask generation through Otsu's [27] method + +$$ +M a s k = \mathbf {1} _ {\text {S i m} < \text {O t s u} (\text {S i m})}, \tag {12} +$$ + +where $\mathbf{1}_{\{\cdot\}}$ denotes the indicator function, which returns 1 if the condition is true. The mask is then multiplied by the ground truth image to extract the dynamic region. + +Decomposition Correspondence from 2D to 3D. To extend 2D dynamic-static decomposition to 3D Gaussians, we establish view-independent correspondences by leveraging depth information from 3D-GS rendering. By unprojecting pixels from masked dynamic regions across multiple views and merging the resulting 3D points, we obtain a comprehensive representation of dynamic regions in 3D space. + +Next, we map merged points to dynamic Gaussians based on spatial proximity. Each point expands spherically with radius $r$ to form a 3D volume (Fig. 5 (c)), initially classifying enclosed Gaussians as dynamic and others as static. To overcome potential decomposition inaccuracies and radius sensitivity, we implement a buffer-based soft decomposition strategy using two radii, $r_1$ and $r_2$ . Gaussians within $r_1$ are marked as dynamic, beyond $r_2$ as static, while those between are pruned to create a buffer zone. This strategy enables 3D-GS to optimize decomposition boundaries through adaptive density control (ADC) [17], enhancing both rendering quality and speed. As demonstrated in Fig. 12, the + +strategy also exhibits improved robustness to radius parameter selection. + +It should be noted that, our DSD method is performed only once during the entire training process and requires only about one minute, introducing minimal overhead to the training pipeline. + +Joint Rendering of Dynamic and Static Gaussians. With the decomposed dynamic and static Gaussians, we jointly render the entire dynamic scene. Particularly, a deformation field [46] learns to map dynamic Gaussians from canonical space to arbitrary time. Taking time $t$ and the center position $\mu$ of dynamic Gaussians as inputs, the deformation field outputs the displacement of their position $\delta_{\mu}$ , rotation $\delta_r$ , and scaling $\delta_s$ + +$$ +\left(\delta_ {\mu}, \delta_ {r}, \delta_ {s}\right) = \mathcal {F} _ {\theta} (\gamma (s g (\mu)), \gamma (t)), \tag {13} +$$ + +where $sg(\cdot)$ indicates a stop-gradient operation and $\gamma (\cdot)$ denotes the positional encoding [46]. Then, the deformed dynamic gaussians can be addressed as + +$$ +\left(\boldsymbol {\mu} ^ {\prime}, \boldsymbol {r} ^ {\prime}, \boldsymbol {s} ^ {\prime}\right) = \left(\boldsymbol {\mu} + \delta_ {\boldsymbol {\mu}}, \boldsymbol {r} + \delta_ {\boldsymbol {r}}, \boldsymbol {s} + \delta_ {\boldsymbol {s}}\right). \tag {14} +$$ + +Finally, static Gaussians bypass the deformation field and merge with the deformed dynamic Gaussians as inputs to the rasterizer, enabling high-frame-rate dynamic rendering. + +# 4. Experiment + +# 4.1. Experimental Settings + +Datasets. Current datasets for event-based dynamic scene reconstruction are highly limited, with only three synthetic and three real-world scenes [25], all unpublished. Notably, the only three publicly available real-world scenes [25] were captured with a static camera, making novel view evaluation infeasible. To facilitate future research, we build the first event-inclusive 4D benchmark featuring 8 synthetic + +Table 2. Quantitative results on our synthetic dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events. The best and second-best result are marked in bold and underlined, respectively. + +
MethodLegoHotdogMaterialsMusic box
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]23.600.9180.08822330.010.9510.06426028.070.9670.06126219.200.9050.122239
TiNeuVox [7]22.390.8910.0710.5330.810.9530.0350.4926.630.9380.0540.5220.450.8310.1520.62
K-Planes [9]24.550.9310.0352.3431.360.9580.0162.3530.620.9760.0092.2920.770.8580.0712.39
4D-GS [42]26.300.9370.07210433.480.9650.05213230.400.9790.05411124.060.9370.07164
Deformable-3DGS [46]23.790.9230.0537332.910.9620.01713234.000.9860.0049122.080.9240.05251
Event-4DGS28.000.9430.0405434.610.9690.0199635.600.9890.0067428.580.9500.04342
Ours31.850.9670.01818936.150.9740.01324138.020.9930.00324030.780.9630.02992
MethodCelestial globeFanWater wheelMan
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]19.050.9150.11018221.260.8910.11827019.430.8870.10921520.790.8700.114210
TiNeuVox [7]13.620.7360.2900.6219.900.8890.1070.5317.030.8500.1470.5622.810.8870.0710.50
K-Planes [9]15.490.8570.0882.4622.100.9090.0622.3820.960.9200.0462.4121.020.8570.0732.33
4D-GS [42]20.970.9420.0725225.050.9360.08010820.960.9170.0777723.800.9140.07664
Deformable-3DGS [46]23.070.9620.0364124.660.9290.0519020.790.9120.0514323.060.9060.05137
Event-4DGS24.300.9480.0453627.660.9490.0417126.340.9320.0523025.550.9210.06333
Ours28.830.9760.0207330.180.9640.02516828.470.9500.03311228.210.9430.037129
+ +Table 3. Quantitative results on our real-world dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events. + +
MethodExcavatorJeepFlowersEagle
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
4D-GS [42]28.350.9110.11011528.340.8780.0936126.820.8730.1236327.590.9000.128105
Deformable-3DGS [46]26.120.9030.1208126.300.8700.1045226.400.9030.0796427.440.9030.12570
Event-4DGS29.670.9140.0925729.640.9010.0794727.530.9050.0844029.080.8960.10463
Ours31.280.9250.07017930.410.9050.0688928.570.9130.06914931.290.9180.074192
+ +and 4 real-world dynamic scenes, encompassing diverse complexities, intricate structures, and rapid motions, thus enabling effective evaluation of dynamic reconstruction. + +For synthetic scenes, we use Blender [3] to generate one-second, $360^{\circ}$ monocular camera rotations, producing thousands of continuous frames per scene. These high-temporal-resolution sequences are processed through ESIM [31] to generate events. For each sequence, we uniformly sample 30 frames (equivalent to 30 FPS) for training, and select intermediate frames as far apart as possible from the training frames for testing. Particularly, both "Fan" representing a typical high-speed 4D scene, and "Man" featuring large object displacements, present significant challenges. + +For real-world scenes, as shown in Fig. 6, we construct a hybrid camera system consisting of a beam splitter, an event camera (Prophesee Gen4), a frame camera (Basler ace), and a microcontroller (STM32) for outputting synchronization signals. Following [32], we keep the camera system static and place the objects on a motorized optical rotating turntable, which is equivalent to camera motion. Following prior work [25], we downsample the original high-FPS video for training and use intermediate frames for testing. + +Baselines. For RGB-only settings, we benchmarked our method against the representative NeRF baselines K-Planes [9] and TiNeuVox [7], along with Gaussians baselines 3D-GS [17], 4D-GS [42], and Deformable-3DGS [46]. For event-assisted settings, DE-NeRF [25] is the only baseline; however, it could not be directly compared, as its code was still closed-source. DE-NeRF relies on NeRF's volume ren + +![](images/4837165c12f3e9cca44395f814b8bb07f094bfdf45f1cadec074b25a7b391136.jpg) +Figure 6. Real-world data acquisition setup (left) and our hybrid camera system (right). + +dering techniques [26], leading to predictably slow rendering speeds. Moreover, its reconstruction quality is also predictably limited due to the absence of threshold modeling for events. To provide a comparable baseline, we introduce Event-4DGS, an extension of Deformable-3DGS [46] that incorporates the event rendering loss in Eq. (4). + +Metrics. We evaluate rendering quality using PSNR, SSIM [37], and LPIPS [50] (based on AlexNet) and measure rendering speed in FPS on an NVIDIA RTX 3090 GPU. + +# 4.2. Comparisons + +Quantitative Results. We report the quantitative results of the comparison on the synthetic and real-world datasets in Tab. 2 and Tab. 3, respectively. Although 4D-GS [42] and Deformable-3DGS [46] achieve relatively higher FPS compared to NeRF baselines [7, 9], their reconstruction quality is limited by the sparsity of the RGB training frames. In contrast, Event-4DGS leverages the rich intermediate mo + +![](images/55a10940b76dbe35079c00e8a2e212076c7d4f52971277923b0b2b3f8c2a20e6.jpg) +Figure 7. Qualitative comparisons on our synthetic dataset. Please see the supplementary video for details. + +![](images/5b8ea59bf6d16982b93e27873198469c02f1862e1a213f7a8be0d3050c31edbb.jpg) +Figure 8. Qualitative comparisons on our real-world dataset. Please see the supplementary video for details. + +tion and viewpoint information provided by events, significantly outperforming other baselines in reconstruction quality, with an average PSNR improvement of 3.28 dB over Deformable-3DGS across all synthetic scenes. This notable improvement underscores the effectiveness of high-temporal-resolution event cameras for dynamic scene reconstruction. However, Event-4DGS still suffers from threshold variation, whereas our method with GTJM enables accurate threshold modeling and better event super + +vision, achieving an average PSNR improvement of 2.73 dB over Event-4DGS on synthetic datasets. Meanwhile, our method maintains exceptionally fast rendering speeds, averaging $1.71 \times$ faster than 4D-GS on synthetic datasets. In summary, our method enjoys both the highest rendering quality and exceptional rendering speed on both synthetic and real-world datasets. + +Qualitative Results. For a more visual assessment, we present qualitative results on the synthetic and real-world + +Table 4. Ablation studies on synthetic dataset. For real-world ablation studies, please refer to the supplementary document. + +
MethodPSNR↑SSIM↑LPIPS↓FPS↑
w/o GTJM29.390.9560.034153
w/o Joint Optimization in GTJM30.870.9630.026152
w/o DSD30.780.9610.02657
w/o Buffer-based Soft Decomposition31.020.9630.025138
Full31.560.9660.022156
+ +datasets in Fig. 7 and Fig. 8, respectively. These comparisons highlight the capability of our method to deliver high-fidelity dynamic scene modeling. Notably, our method effectively captures intricate motion details, while other baselines exhibit structural deficiencies and distortions. + +Dynamic Blurry Scene Comparisons. Motion blur is another common challenge in dynamic scenes. To address this, we extend both baselines and our method with blur loss and EDI from [48], and build blurry scenes for evaluations. Fig. 9 shows that, by leveraging events' deblurring advantage, our method outperforms Deformable-3DGS by 4.79 dB in PSNR, achieving the best results. For detailed methods and quantitative results, see the supplementary document. + +# 4.3. Ablation Study + +GS-threshold Joint Modeling. Using a constant threshold fails to properly neutralize opposing polarity events during accumulation, resulting in motion trajectory artifacts as shown in Fig. 10 (b). These artifacts, when used for Gaussian supervision, produce undesirable purple haze in rendered outputs, such as the Event-4DGS results in Fig. 7. Our RGB-assisted threshold estimation significantly reduces these artifacts (Fig. 10 (c)), while subsequent joint threshold and GS optimization effectively eliminates remaining distortions (Fig. 10 (d)). As demonstrated in Tab. 4, this improved event supervision yields a $2.17\mathrm{dB}$ average PSNR improvement across all scenes, validating our GTJM strategy's effectiveness in handling threshold variations. + +Dynamic-static Decomposition. Our DSD method successfully identifies dynamic regions of varying sizes and geometries, as demonstrated in Fig. 11. Modeling the entire scene with dynamic Gaussians without DSD misallocates deformation field capacity to static regions, compromising dynamic region reconstruction quality as shown in Fig. 4. Quantitative results in Tab. 4 demonstrate that using DSD improves the average PSNR by $0.78\mathrm{dB}$ and accelerates the rendering speed to 2.74 times the original FPS. This underscores DSD's crucial role in achieving both high-fidelity dynamic scene reconstruction and efficient rendering. + +Buffer-based Soft Decomposition. Our buffer-based soft decomposition enables adaptive optimization of decomposition boundaries, yielding a 0.54 dB improvement in average PSNR (Tab. 4). Sensitivity analysis reveals that reconstruction quality stabilizes when buffer size $(r_2 - r_1)$ + +![](images/43c8ae281459886f7adf86aa88137b11d9d07dfa13ba5de358b78ccebfe4c4bf.jpg) +Figure 9. Extended comparisons on the dynamic blurry scene. + +![](images/57e620d0531d10c7d0e089b7d89b72ac5e1b00708fea5dae23063395b7c52e77.jpg) + +![](images/cd504b106d097fe1f5c4c75d83033e87b52371c70da886e1486c6563cbb16c84.jpg) + +![](images/eb78c0a3864da654c8d61b6868639ad13d787540e2c945d46a3cab0a4c53f038.jpg) + +![](images/7dcfd7deda0d68aa09981543df298986b22ea8900ab46724b372af067d760bfe.jpg) + +![](images/5d372c9d3ebd06f9071764a067c54c17a57427531103b1bd86a4ae41edd81428.jpg) + +![](images/3e641b04f76e0d0925fd3be5f4923fa51518e238ebc8d8c4eb8a64fd65560154.jpg) +(a) GT + +![](images/6400539520cccf496f75b6598a337f0807b97faa780bed6a046e9607e0715d91.jpg) +(b) w/o GTJM +c) w/o Joint Optimization + +![](images/4ec7e36bf2924f7ffd0658f4dfa86d60e695575564f5df5217d1c37166c42d59.jpg) +in GTJM + +![](images/9b2b290f2151706660f24c79818c96da05bc2de5953cc34195e24643c2f18e3b.jpg) +(d) Full + +![](images/b286634b0b8acac45b7eae191b2554ffeab58c56741a496b7ba91b7b4027d41c.jpg) +Figure 10. The effect of GS-threshold joint modeling strategy, which eliminates event artifacts caused by threshold variations. +Lego Hotdog +Hotdog + +![](images/063d96bced7e5dca0cdf0b5d0e453c296433544b033b055abcfea113105dc542.jpg) +Materials +Cel +estial globe +Man +Figure 11. Rendering results of dynamic and static Gaussians separated by our dynamic-static decomposition strategy. +Figure 12. Sensitivity analysis on buffer size $(r_2 - r_1)$ . + +exceeds approximately 12 basic units (normalized by average inter-Gaussian distance to account for scene variations), as shown in Fig. 12. This stability demonstrates the robustness of our DSD method through adaptive boundary search, highlighting the effectiveness of the buffer-based strategy. + +# 5. Conclusion + +In this paper, we present an event-boosted deformable 3D Gaussian framework for high-quality dynamic scene reconstruction. Our GS-threshold joint modeling effectively addresses threshold variation challenges, enabling reliable event supervision. The proposed dynamic-static decomposition method enhances both rendering efficiency and reconstruction quality through optimized resource allocation between static and dynamic regions. + +# Acknowledgements + +This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001, and the Fundamental Research Funds for the Central Universities under Grant WK210000059. + +# References + +[1] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In ECCV, 2024. 2, 4 +[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1 +[3] Blender. A 3d modelling and rendering package. https://www.blender.org/, 2018.6 +[4] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In CVPR, 2024. 2, 3 +[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 1 +[6] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In ICIP, 2024. 2 +[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia, 2022. 1, 6 +[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022. +[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 1, 2, 6 +[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3 +[11] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019. 3 +[12] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 35(4):3119-3133, 2024. 2, 4 +[13] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR workshop, 2021. 2, 3 + +[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR, 2024. 2, 4 +[15] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In WACV, 2023. 2, 3 +[16] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 2 +[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):139-1, 2023. 1, 2, 3, 5, 6 +[18] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 2, 3 +[19] Chenghan Li, Christian Brandli, Raphael Berner, Hongjie Liu, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. Design of an rgb color vga rolling and global shutter dynamic and active-pixel vision sensor. In ISCAS, 2015. 1, 3 +[20] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. In WACV, 2025. 1, 2 +[21] Yutong Liu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. S2d-lfe: Sparse-to-dense light field event generation. In CVPR, 2025. 3 +[22] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In ICCV, 2023. 2, 3 +[23] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, 2024. 2, 4 +[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 1, 2, 4 +[25] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, 2023. 2, 3, 5, 6 +[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 6 +[27] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62-66, 1979. 5 +[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2 +[29] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2 + +[30] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 2, 3 +[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 6 +[32] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, 2023. 2, 3, 6 +[33] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2 +[34] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In ECCV, 2024. 1, 2 +[35] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 4 +[36] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1 +[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6 +[38] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 1 +[39] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Boosting event stream super-resolution with a recurrent neural network. In ECCV, 2022. 1 +[40] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In CVPR, 2023. 3 +[41] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2 +[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 1, 2, 4, 6 +[43] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. In MLSP, 2024. 2 +[44] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In CoRL, 2024. 2 +[45] Wenhao Xu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Ceia: Clip-based event-image alignment for open-world event-based understanding. In ECCV workshop, 2024. 3 +[46] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 4, 5, 6 + +[47] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 1 +[48] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2, 8 +[49] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: a structured and explicit radiance representation for 3d generative modeling. In NeurIPS, 2024. 1 +[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 \ No newline at end of file diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/images.zip b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..d58577f9f2a6dd689a07f34e1677e9e649f94cbb --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cb5b8a7efc0c0c669fad70478941c9f6f44c0313938498de7e36941686d4be0 +size 996815 diff --git a/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/layout.json b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ab383c654b1715bf7fda71fa0cc73a024e7d4d1c --- /dev/null +++ b/2025/Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction/layout.json @@ -0,0 +1,9224 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "spans": [ + { + "bbox": [ + 70, + 103, + 539, + 120 + ], + "type": "text", + "content": "Event-boosted Deformable 3D Gaussians for Dynamic Scene Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 141, + 516, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 141, + 516, + 172 + ], + "spans": [ + { + "bbox": [ + 94, + 141, + 516, + 172 + ], + "type": "text", + "content": "Wenhao Xu1 Wenming Weng1 Yueyi Zhang2, * Ruikang Xu1 Zhiwei Xiong1, * \n1University of Science and Technology of China 2MiroMind" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 120, + 174, + 489, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 174, + 489, + 185 + ], + "spans": [ + { + "bbox": [ + 120, + 174, + 489, + 185 + ], + "type": "text", + "content": "wh-xu@mail.ustc.edu.cn, yueyi.zhang@miromind.ai, zwxiong@ustc.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 212, + 188, + 394, + 198 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 212, + 188, + 394, + 198 + ], + "spans": [ + { + "bbox": [ + 212, + 188, + 394, + 198 + ], + "type": "text", + "content": "https://github.com/wh-xul/E-D3DGS" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 58, + 224, + 555, + 355 + ], + "blocks": [ + { + "bbox": [ + 58, + 224, + 555, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 224, + 555, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 224, + 555, + 355 + ], + "type": "image", + "image_path": "701fea225884661de94c68de553537d4cafda88441bab97a8cd5120039349391.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "lines": [ + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "type": "text", + "content": "Figure 1. Left: Quality comparison with the baselines 4D-GS [42], Event-4DGS (the event-extended version of [46]), and our variants, highlighting the superior rendering quality of our method. Our GS-threshold joint modeling (GTJM) effectively eliminates event-induced artifacts (particularly the purple haze), while our dynamic-static decomposition (DSD) improves the quality of dynamic regions. Middle: Separate rendering of dynamic and static Gaussians from our DSD. Right: The scatter plot illustrates our method's ability to achieve both high fidelity and fast rendering, where dot radii correspond to different resolutions " + }, + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "type": "inline_equation", + "content": "(400\\times 400,600\\times 600," + }, + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "type": "inline_equation", + "content": "800\\times 800)" + }, + { + "bbox": [ + 54, + 357, + 555, + 413 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "spans": [ + { + "bbox": [ + 151, + 422, + 200, + 434 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "spans": [ + { + "bbox": [ + 54, + 448, + 297, + 700 + ], + "type": "text", + "content": "Deformable 3D Gaussian Splating (3D-GS) is limited by missing intermediate motion information due to the low temporal resolution of RGB cameras. To address this, we introduce the first approach combining event cameras, which capture high-temporal-resolution, continuous motion data, with deformable 3D-GS for dynamic scene reconstruction. We observe that threshold modeling for events plays a crucial role in achieving high-quality reconstruction. Therefore, we propose a GS-Threshold Joint Modeling strategy, creating a mutually reinforcing process that greatly improves both 3D reconstruction and threshold modeling. Moreover, we introduce a Dynamic-Static Decomposition strategy that first identifies dynamic areas by exploiting the inability of static Gaussians to represent motions, then applies a buffer-based soft decomposition to separate dynamic and static areas. This strategy accelerates rendering by avoiding unnecessary deformation in static areas, and focuses on dynamic areas to enhance fidelity. Additionally, we contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "spans": [ + { + "bbox": [ + 315, + 422, + 394, + 434 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 589 + ], + "type": "text", + "content": "Dynamic scene reconstruction and novel view synthesis are essential for immersive applications in virtual/augmented reality and entertainment [20, 24, 34, 42, 46, 47, 49]. While Neural Radiance Fields (NeRF) [2, 26, 36] offer unprecedented photorealism, they are constrained by slow training and rendering speeds. Despite recent advances in optimization techniques [5, 7-9], real-time rendering remains elusive. Recently, 3D Gaussian Splatting (3D-GS) [17] addresses this limitation through efficient differentiable rasterization, yet existing dynamic extensions [24, 42, 46, 47] are constrained by inherent limitations of RGB cameras, including low frame rates and motion blur." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 591, + 556, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 591, + 556, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 591, + 556, + 675 + ], + "type": "text", + "content": "In this paper, we present event-boosted deformable 3D Gaussians for dynamic scene reconstruction. Event cameras [10, 19, 38, 39], with their microsecond-level temporal resolution, can provide continuous motion information and near-infinite viewpoints that traditional RGB cameras often fail to capture. These advantages make event cameras particularly valuable for dynamic scene reconstruction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 713 + ], + "type": "text", + "content": "However, integrating events into 3D scene reconstruction faces new challenges. Specifically, event supervision for 3D-GS relies on an accurate event generation model" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28334" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 288 + ], + "type": "text", + "content": "[4, 18], where the threshold undergoes complex variations across polarity, space, and time [13, 18]. Previous methods [4, 6, 18, 25, 30, 32, 43, 44, 48] adopt a constant threshold, yet this simplification significantly degrades the quality of event supervision. While recent works [15, 22] attempt to model threshold variations using event data alone, they achieve limited success due to the inherent binary nature of events, which only indicate brightness change directions. To address this challenge, we propose a novel GS-threshold joint modeling strategy. First, we leverage the brightness change values from RGB frames to supervise threshold optimization. Second, since the sparsity of RGB frames weakens supervision, we use 3D-GS rendered results as pseudo-intermediate frames to enhance the supervision. This finally creates a mutually reinforcing process where RGB-optimized threshold enables better event supervision for 3D-GS, while improved 3D-GS in turn provides accurate geometric constraints for threshold refinement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 293, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 510 + ], + "type": "text", + "content": "Furthermore, we observe that existing dynamic 3D-GS methods inefficiently use dynamic Gaussians solely to model both static and dynamic regions [1, 12, 14, 23, 24, 42, 46]. This unified treatment leads to reduced rendering speed, wasted deformation field capacity, and degraded reconstruction quality. While some methods have explored dynamic-static decomposition, they are limited by either inaccurate dynamic Gaussians initialization [20] or constraints in multi-view scenarios [34]. To address these limitations, we propose a novel dynamic-static decomposition strategy that first identifies dynamic regions based on the inherent inability of static Gaussians to represent motions, and then employs a buffer-based soft decomposition to adaptively search for the optimal decomposition boundary. This decomposition not only accelerates rendering by eliminating unnecessary deformation computations in static regions, but also enhances reconstruction quality by focusing the deformation field exclusively on dynamic regions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 514, + 288, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 514, + 288, + 526 + ], + "spans": [ + { + "bbox": [ + 67, + 514, + 288, + 526 + ], + "type": "text", + "content": "Our main contributions can be summarized as follows:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 532, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 295, + 567 + ], + "type": "text", + "content": "- We present the first method integrating event cameras with deformable 3D-GS for dynamic scene reconstruction, enabling high-fidelity and fast rendering." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 568, + 295, + 627 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 295, + 627 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 295, + 627 + ], + "type": "text", + "content": "- We propose a novel GS-threshold joint modeling strategy that combines RGB-assisted initial estimation with GS-boosted refinement, creating a mutually reinforcing process that significantly improves both threshold modeling and 3D reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "spans": [ + { + "bbox": [ + 55, + 628, + 295, + 675 + ], + "type": "text", + "content": "- We introduce an effective dynamic-static decomposition strategy that not only accelerates rendering through selective deformation computation but also enhances reconstruction quality by focusing on dynamic regions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 675, + 295, + 712 + ], + "type": "text", + "content": "- We contribute the first event-inclusive 4D benchmark with synthetic and real-world dynamic scenes, on which our method achieves state-of-the-art performance." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 553, + 342 + ], + "type": "text", + "content": "Neural Rendering for Dynamic Scenes. Neural rendering techniques have revolutionized dynamic scene reconstruction in recent years. Pioneering works like D-NeRF [29] and Nerfies [28] extend Neural Radiance Fields [26] through deformation field, mapping observations into a canonical space for modeling non-rigid motion. Despite their impressive reconstruction quality, these methods are constrained by extensive computational demands due to dense MLP evaluations during training and rendering. Various acceleration strategies have been proposed to address these limitations. K-Planes [9] introduces an efficient explicit representation using six feature planes, while Tensor4D [33] and DTensorRF [16] employ tensor decomposition techniques to achieve compact spatiotemporal encoding. A significant breakthrough came with 3D Gaussian Splatting (3D-GS) [17], which leverages efficient differentiable rasterization for real-time rendering. This advancement has spawned several dynamic scene extensions, including 4D-GS [42], Deformable-3DGS [46], and related works [24, 34], which successfully achieve real-time rendering for high-quality dynamic scene reconstruction." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 553, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 606 + ], + "type": "text", + "content": "Event-based Neural Rendering. The integration of neural representations with event-based 3D reconstruction has emerged as a promising research direction. Pioneering works like EventNeRF [32], E-NeRF [18], and Ev-NeRF [15] first demonstrated the potential of pure event-based static scene reconstruction, albeit with different assumptions about event camera characteristics. To further improve reconstruction quality, E2NeRF [30] and Ev-DeblurNeRF [4] incorporated blurry RGB images alongside events. A significant milestone was achieved by DE-NeRF [25], which pioneered the combination of events and RGB frames for dynamic scene reconstruction. The recent advent of 3D Gaussian Splatting [17] has catalyzed new developments in event-based methods. While Ev-GS [43] adapted the pure event-based paradigm to 3D-GS, subsequent works including E2GS [6], EaDeblur-GS [41], and Event3DGS [44] primarily addressed deblurring challenges. To the best of our knowledge, our work is the first to integrate events with deformable 3D-GS for dynamic scene reconstruction. Our novel method and benchmark, specifically tailored for events and dynamic scenes, highlight significant and independent contributions that set our work apart." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 615, + 370, + 627 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 615, + 370, + 627 + ], + "spans": [ + { + "bbox": [ + 314, + 615, + 370, + 627 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 635, + 460, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 635, + 460, + 647 + ], + "spans": [ + { + "bbox": [ + 314, + 635, + 460, + 647 + ], + "type": "text", + "content": "3.1. Event Cameras for 3D-GS" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "content": "3D Gaussian Splatting Preliminary. 3D Gaussian Splatting (3D-GS) [17] represents a scene as anisotropic 3D Gaussians, each characterized by a covariance matrix " + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "content": " and center position " + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "content": ": " + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "inline_equation", + "content": "GS(\\mathbf{x}) = e^{-\\frac{1}{2} (\\mathbf{x} - \\mu)^T\\boldsymbol{\\Sigma}^{-1}(\\mathbf{x} - \\mu)}" + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "content": ". The covariance matrix " + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\Sigma" + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "text", + "content": " is parameterized using scaling matrix " + }, + { + "bbox": [ + 313, + 652, + 553, + 714 + ], + "type": "inline_equation", + "content": "S" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "28335" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 298, + 191 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 298, + 191 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 298, + 191 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 298, + 191 + ], + "type": "image", + "image_path": "aad825284d12801d28d74671a61736e0a28da0b29ef58cf084d5c3ffc6f4ad44.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "text", + "content": "Figure 2. Overview of GS-threshold joint modeling strategy. " + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "text", + "content": " optimizes 3D-GS, " + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{thres}" + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "text", + "content": " optimizes the threshold, and " + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{event}" + }, + { + "bbox": [ + 55, + 198, + 296, + 232 + ], + "type": "text", + "content": " jointly optimizes both 3D-GS and threshold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": "and rotation matrix " + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": " to ensure positive semi-definiteness: " + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\Sigma = RSS^T R^T" + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": ". Each Gaussian is further defined by spherical harmonic coefficients " + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": " and opacity " + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": ". Final pixel colors " + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 250, + 296, + 299 + ], + "type": "text", + "content": " are computed using differentiable tile-based rasterization" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 120, + 304, + 296, + 337 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 120, + 304, + 296, + 337 + ], + "spans": [ + { + "bbox": [ + 120, + 304, + 296, + 337 + ], + "type": "interline_equation", + "content": "c = \\sum_ {i \\in N} c _ {i} \\alpha_ {i} \\prod_ {j = 1} ^ {i - 1} (1 - \\alpha_ {j}), \\tag {1}", + "image_path": "f76e262b141c05e71fbe7960384c3debd1fd915a87c346b13115d34bf0047245.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "c_{i}" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " denotes spherical harmonic color, and " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "\\alpha_{i}" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " combines opacity " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": " with projected " + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "inline_equation", + "content": "GS(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 342, + 295, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 366, + 296, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 449 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 449 + ], + "type": "text", + "content": "Event/RGB Rendering Loss. Event cameras [10, 19, 21, 40, 45] are novel sensors, that asynchronously capture pixel-wise brightness changes with microsecond-level temporal resolution. Their high temporal precision enables capturing crucial motions between RGB frames and provides near-infinite viewpoint supervision, making them ideal for monocular dynamic scene reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": "Each event is represented as " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "e_{x,y}(\\tau) = p\\delta(\\tau)" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " is pixel position, " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " is timestamp, " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "p \\in \\{+1, -1\\}" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " indicates brightness change direction relative to threshold " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\delta(t)" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " is a unit integral impulse function. Omitting pixel subscripts, the brightness change over interval " + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "inline_equation", + "content": "\\triangle t" + }, + { + "bbox": [ + 55, + 449, + 296, + 521 + ], + "type": "text", + "content": " can be formulated as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 524, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 524, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 102, + 524, + 295, + 552 + ], + "type": "interline_equation", + "content": "E (t, t + \\Delta t) = \\int_ {t} ^ {t + \\Delta t} C \\cdot e (\\tau) d \\tau . \\tag {2}", + "image_path": "a3d7ef75631335fb23108e5e7ef47e110c49c6d51fca6aa1890a48ec650e7abb.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 556, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 569 + ], + "type": "text", + "content": "This change can also be estimated from rendered brightness" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 83, + 575, + 295, + 592 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 575, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 83, + 575, + 295, + 592 + ], + "type": "interline_equation", + "content": "\\hat {E} (t, t + \\triangle t) \\stackrel {\\mathrm {d e f}} {=} \\log (\\hat {I} (t + \\triangle t)) - \\log (I (t)), \\quad (3)", + "image_path": "5538c51e9c4bda83b9b20f53b573c5485a40a12756d3f8a55e3e892f3720612a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "inline_equation", + "content": "\\hat{I}" + }, + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "inline_equation", + "content": "I" + }, + { + "bbox": [ + 55, + 596, + 296, + 621 + ], + "type": "text", + "content": " respectively denote 3D-GS rendered and ground truth brightnesses. The event rendering loss is" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 85, + 626, + 295, + 649 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 626, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 85, + 626, + 295, + 649 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {e v e n t}} = \\left\\| E (t, t + \\triangle t) - \\hat {E} (t, t + \\triangle t) \\right\\| _ {2} ^ {2}. \\tag {4}", + "image_path": "14e7bd8d2a8be5305e410f274dc41669cc8080672b4e7a45378133b72e73f6d7.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 677 + ], + "type": "text", + "content": "Similarly, we utilize the RGB rendering loss [17] combining L1 and D-SSIM losses as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {r g b} = \\left(1 - \\lambda_ {s}\\right) \\left\\| \\hat {I} (t) - I (t) \\right\\| _ {1} + \\lambda_ {s} \\mathcal {L} _ {D - S S I M} (\\hat {I} (t), I (t)), \\tag {5}", + "image_path": "908c5fca1e1a59edb629f2e5885211198c549c17377fbc411d74ade7d303304e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 441, + 160 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 441, + 160 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 441, + 160 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 441, + 160 + ], + "type": "image", + "image_path": "a190dd5d3358b89e676b2579f01d3827ab2e4bd5966c4c0772fccca9ecb5bc61.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 372, + 162, + 382, + 171 + ], + "lines": [ + { + "bbox": [ + 372, + 162, + 382, + 171 + ], + "spans": [ + { + "bbox": [ + 372, + 162, + 382, + 171 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 443, + 70, + 553, + 161 + ], + "blocks": [ + { + "bbox": [ + 443, + 70, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 443, + 70, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 443, + 70, + 553, + 161 + ], + "type": "image", + "image_path": "f81b37a6d672990ea6842acdda2747088896f6a9e38752b03822396a9307ab57.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 161, + 510, + 170 + ], + "lines": [ + { + "bbox": [ + 500, + 161, + 510, + 170 + ], + "spans": [ + { + "bbox": [ + 500, + 161, + 510, + 170 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 175, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 553, + 209 + ], + "type": "text", + "content": "Figure 3. (a) The effect of different ranges of threshold variation for 3D reconstruction. (b) The effect of different number of RGB frames for threshold estimation." + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\lambda_{s}" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": " is a weighting factor that controls the balance. 3D-GS is optimized to minimize " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{event}" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 313, + 222, + 553, + 247 + ], + "type": "text", + "content": " jointly" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 365, + 253, + 553, + 271 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 253, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 365, + 253, + 553, + 271 + ], + "type": "interline_equation", + "content": "G S ^ {*} = \\arg \\min _ {G S} \\left(\\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {6}", + "image_path": "502d5febde88759dcc54161f6026aa96d04ff523159c32aa49b89ea5569b5c76.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 284, + 474, + 297 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 474, + 297 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 474, + 297 + ], + "type": "text", + "content": "3.2. GS-threshold Joint Modeling" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 301, + 555, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 420 + ], + "type": "text", + "content": "As shown in Eq. (2), the threshold " + }, + { + "bbox": [ + 313, + 301, + 555, + 420 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 301, + 555, + 420 + ], + "type": "text", + "content": " critically affects event integration and supervision quality. While existing methods [4, 18, 25, 30, 32] typically assume a constant threshold, real event cameras exhibit threshold variations across polarity, space, and time [13, 18]. Fig. 3 (a) demonstrates how increasing threshold variation significantly degrades reconstruction quality. To model threshold variations, we propose a GS-threshold joint modeling (GTJM) strategy (see Fig. 2), consisting of RGB-assisted threshold estimation and GS-boosted threshold refinement." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "content": "RGB-assisted Threshold Estimation. Recent works [15, 22] attempt to model threshold variations solely from event data but face significant limitations due to the difficulty of inferring brightness changes from events, which only indicate the direction of change. We propose to leverage brightness change values from RGB frames for robust threshold estimation. Given two RGB frames " + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "inline_equation", + "content": "I(f)" + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "content": " at times " + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 421, + 556, + 517 + ], + "type": "text", + "content": ", we define the threshold modeling loss as" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 347, + 523, + 553, + 547 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 523, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 347, + 523, + 553, + 547 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t h r e s}} = \\left\\| E _ {\\text {t h r e s}} (t, f) - \\hat {E} _ {\\text {t h r e s}} (t, f) \\right\\| _ {2} ^ {2}, \\tag {7}", + "image_path": "19e531aa5f68c3614dcc7412ed61d173d038241f73fd28adf19dcbddea1f0032.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\hat{E}_{thres}(t,f) := \\int_t^f \\hat{C} \\cdot e(\\tau)d\\tau" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "E_{thres}(t,f) = \\log(I(f)) - \\log(I(t))" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": ". In practice, we adopt a simple yet fast way to compute " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\hat{E}_{thres}(t,f)" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": ". We first accumulate events to obtain event count maps [11] " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "ECM_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": " denotes time bins and " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "P = 2" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": " corresponds to the event polarity. Using learnable threshold parameters " + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "inline_equation", + "content": "\\hat{C}_{t,f} \\in \\mathbb{R}^{B \\times P \\times H \\times W}" + }, + { + "bbox": [ + 313, + 553, + 555, + 640 + ], + "type": "text", + "content": ", we compute" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 328, + 647, + 553, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 647, + 553, + 681 + ], + "spans": [ + { + "bbox": [ + 328, + 647, + 553, + 681 + ], + "type": "interline_equation", + "content": "\\hat {E} _ {\\text {t h r e s}} (t, f) = \\sum_ {b = 1} ^ {B} \\sum_ {p = 1} ^ {P} \\left(E C M _ {t, f} \\odot \\hat {C} _ {t, f}\\right) _ {b, p,:,:}, \\tag {8}", + "image_path": "75a00d9a3cd7cf83c6abb6ded13f570b669ea573cd2d727e096ca158ca14f774.jpg" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "where threshold " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\hat{C}_{t,f}" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " is optimized by minimizing Eq. (7) in an end-to-end manner." + } + ] + } + ], + "index": 27 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28336" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 129, + 294, + 205 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 295, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 295, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 295, + 125 + ], + "type": "text", + "content": "Table 1. Step-by-step validation of mutual boosting between threshold modeling (TM) and 3D reconstruction (3D Rec.). Abbreviations: \"Fro.\" : \"Frozen\"; \"Ft.\" : \"Fine-tuning\". Note that TM is evaluated by MSE between the estimated and GT thresholds from the simulator." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 129, + 294, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 129, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 129, + 294, + 205 + ], + "type": "table", + "html": "
TM for 3D Rec.3D Rec. for TM
Stage1Stage23D Rec.Stage1Stage2TM
TM3D Rec. (Fro. C)PSNR↑3D Rec.TM (Fro. GS)MSE↓ (×10-4)
×✓24.46×✓8.317
✓✓26.63✓✓7.077
", + "image_path": "9e57a07cfd095382cce5f58ad10f905992322aa40026eb9c50b0ac5edd896c1e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 60, + 217, + 293, + 270 + ], + "blocks": [ + { + "bbox": [ + 118, + 208, + 234, + 217 + ], + "lines": [ + { + "bbox": [ + 118, + 208, + 234, + 217 + ], + "spans": [ + { + "bbox": [ + 118, + 208, + 234, + 217 + ], + "type": "text", + "content": "Joint TM and 3D Rec. Optimization" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 217, + 293, + 270 + ], + "lines": [ + { + "bbox": [ + 60, + 217, + 293, + 270 + ], + "spans": [ + { + "bbox": [ + 60, + 217, + 293, + 270 + ], + "type": "table", + "html": "
Stage1Stage23D Rec.TM
TM3D Rec. & TM (Ft. C̅ & Ft. GS)PSNR↑MSE↓ (×10-4)
✓✓28.016.322
", + "image_path": "a8d1ca8e0392ea26015c2cc88b7d70f5856ce94aec925118069266963127c241.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 277, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 277, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 277, + 295, + 348 + ], + "type": "text", + "content": "We observe that accurate threshold modeling improves 3D reconstruction quality. As shown in \"TM for 3D Rec.\" in Tab. 1, our RGB-assisted threshold optimization approach benefits threshold estimation, thus significantly enhancing 3D-GS reconstruction quality and achieving a 2.17 dB PSNR improvement." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 349, + 295, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 349, + 295, + 469 + ], + "spans": [ + { + "bbox": [ + 54, + 349, + 295, + 469 + ], + "type": "text", + "content": "GS-boosted Threshold Refinement. While RGB frames facilitate threshold estimation, their effectiveness is constrained by low frame rate. As illustrated in Fig. 3 (b), sparse RGB frames lead to longer integration intervals, reducing supervision quality and threshold estimation accuracy. To overcome this limitation, we found that once a 3D-GS is trained first by Eq. (4) and Eq. (5), it can be used to render intermediate frames as additional pseudosupervision. Specifically, we freeze the trained 3D-GS and reuse Eq. (4) to enhance Eq. (7) for optimizing threshold " + }, + { + "bbox": [ + 54, + 349, + 295, + 469 + ], + "type": "inline_equation", + "content": "\\hat{C}" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 106, + 475, + 295, + 495 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 475, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 106, + 475, + 295, + 495 + ], + "type": "interline_equation", + "content": "\\hat {C} ^ {*} = \\arg \\min _ {\\hat {C}} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}}\\right). \\tag {9}", + "image_path": "f7d99072d0e1229c7d7ec912c13f634a93c0e90931fcb1e5e610d51bd51d8556.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 295, + 607 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 295, + 607 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 295, + 607 + ], + "type": "text", + "content": "We observe that the incorporation of 3D-GS significantly enhances threshold estimation accuracy. The underlying reason is that events may provide unreliable supervision in some regions with inaccurate thresholds or noise, whereas 3D-GS can correct these errors via geometric consistency. As demonstrated in \"3D Rec. for TM\" part in Tab. 1, using trained and frozen 3D-GS for threshold modeling substantially reduces MSE, leading to more precise threshold estimation." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "text", + "content": "Joint Threshold and GS Optimization. Having demonstrated the mutual benefits between threshold modeling and 3D reconstruction, we propose jointly optimizing both threshold " + }, + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "inline_equation", + "content": "\\hat{C}" + }, + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "text", + "content": " and 3D Gaussians " + }, + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "inline_equation", + "content": "GS" + }, + { + "bbox": [ + 55, + 609, + 295, + 656 + ], + "type": "text", + "content": " through" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 69, + 662, + 295, + 682 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 662, + 295, + 682 + ], + "spans": [ + { + "bbox": [ + 69, + 662, + 295, + 682 + ], + "type": "interline_equation", + "content": "\\hat {C} ^ {*}, G S ^ {*} = \\arg \\min _ {\\hat {C}, G S} \\left(\\mathcal {L} _ {\\text {t h r e s}} + \\mathcal {L} _ {\\text {e v e n t}} + \\mathcal {L} _ {\\text {r g b}}\\right). \\tag {10}", + "image_path": "70f9f0ba13ac40753dd828459e6a7b756a082164270c1cab2a1c92f274f94d1a.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 689, + 295, + 713 + ], + "type": "text", + "content": "We observe that this joint optimization enables a beneficial cycle where optimized thresholds enhance event su" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 541, + 200 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 541, + 200 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 541, + 200 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 541, + 200 + ], + "type": "image", + "image_path": "b741992a7bf752ce457db7a807f1a1ffcc7ea919f4abecbcda42f88090fb6fe9.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 207, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 313, + 207, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 207, + 553, + 230 + ], + "type": "text", + "content": "Figure 4. The effect of dynamic-static decomposition strategy, which improves the rendering quality of dynamic regions." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 243, + 553, + 303 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 243, + 553, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 243, + 553, + 303 + ], + "type": "text", + "content": "pervision for 3D-GS, while improved 3D-GS refines threshold estimates through geometric consistency. As shown in \"Joint TM and 3D Rec. Optimization\" part in Tab. 1, this approach achieves superior threshold modeling and reconstruction quality." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "text", + "content": "In summary, our optimization proceeds in two stages: first optimizing the threshold using " + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{thres}" + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "text", + "content": ", then jointly optimizing both threshold and 3D-GS using all three losses " + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{thres},\\mathcal{L}_{event}" + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{rgb}" + }, + { + "bbox": [ + 313, + 305, + 553, + 354 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 365, + 479, + 379 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 365, + 479, + 379 + ], + "spans": [ + { + "bbox": [ + 314, + 365, + 479, + 379 + ], + "type": "text", + "content": "3.3. Dynamic-static Decomposition" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 504 + ], + "type": "text", + "content": "Dynamic scenes typically contain substantial static regions (e.g., tables, walls) that require no deformation. Unlike existing methods [1, 12, 14, 23, 24, 42, 46] that use dynamic Gaussians throughout, we separately model dynamic and static regions with corresponding Gaussian types. This decomposition offers dual benefits: accelerated rendering by bypassing deformation field computation for static Gaussians, and enhanced deformation fidelity through focused MLP capacity optimization for dynamic regions, as demonstrated in Fig. 4." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "content": "The key challenge lies in accurately initializing dynamic Gaussians in dynamic regions and static Gaussians in static regions. We address this through a proposed dynamic-static decomposition (DSD) strategy, as illustrated in Fig. 5." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 555, + 554, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 554, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 554, + 664 + ], + "type": "text", + "content": "Dynamic-static Decomposition on 2D. We leverage the inherent inability of static Gaussians in representing motion to decompose dynamic and static regions in 2D images. During the first " + }, + { + "bbox": [ + 313, + 555, + 554, + 664 + ], + "type": "inline_equation", + "content": "3k" + }, + { + "bbox": [ + 313, + 555, + 554, + 664 + ], + "type": "text", + "content": " iterations, we perform scene reconstruction using only static Gaussians for initialization. This naturally results in poor reconstruction in dynamic regions while achieving high fidelity in static areas (illustrated in Fig. 5 (a)). This distinct performance difference enables decomposition of training images into dynamic and static regions." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Specifically, using a pretrained VGG19 [35] network " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\phi}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " we extract multi-scale features from both rendered image " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\hat{I}(t)" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " and ground truth " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "I(t)" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ". The cosine similarities computed at each scale are upsampled to a uniform resolution" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28337" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 82, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 189, + 80 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 189, + 80 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 189, + 80 + ], + "type": "text", + "content": "Dynamic-static Decomposition on 2D" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 255, + 71, + 409, + 80 + ], + "lines": [ + { + "bbox": [ + 255, + 71, + 409, + 80 + ], + "spans": [ + { + "bbox": [ + 255, + 71, + 409, + 80 + ], + "type": "text", + "content": "Decomposition Correspondence from 2D to 3D" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 482, + 71, + 534, + 80 + ], + "lines": [ + { + "bbox": [ + 482, + 71, + 534, + 80 + ], + "spans": [ + { + "bbox": [ + 482, + 71, + 534, + 80 + ], + "type": "text", + "content": "Joint Rendering" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 60, + 82, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 60, + 82, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 60, + 82, + 553, + 232 + ], + "type": "image", + "image_path": "e58d64e958d6e0898b474036ac15ced927e96aa93508d6c69100542265c1aafb.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "lines": [ + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "spans": [ + { + "bbox": [ + 55, + 242, + 555, + 276 + ], + "type": "text", + "content": "Figure 5. Overview of dynamic-static decomposition strategy. First, we decompose dynamic and static regions in 2D images based on the inherent inability of static Gaussians to represent motions. Next, we establish a correspondence to extend 2D decomposition to 3D Gaussians. Finally, the decomposed dynamic and static Gaussians are jointly rendered to reconstruct the complete dynamic scene." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 249, + 297 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 249, + 297 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 249, + 297 + ], + "type": "text", + "content": "and averaged to generate a fused similarity map" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 73, + 306, + 294, + 347 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 306, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 73, + 306, + 294, + 347 + ], + "type": "interline_equation", + "content": "\\operatorname {S i m} = \\sum_ {l} U p \\left(\\frac {\\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\cdot \\mathcal {F} _ {\\phi} ^ {l} (I (t))}{\\left\\| \\mathcal {F} _ {\\phi} ^ {l} (\\hat {I} (t)) \\right\\| \\left\\| \\mathcal {F} _ {\\phi} ^ {l} (I (t)) \\right\\|}\\right), \\tag {11}", + "image_path": "c6464f9930c0690587265f59f48d5ebba3f28960067b70894879e320a1bea979.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "inline_equation", + "content": "\\mathcal{F}_{\\phi}^{l}(\\cdot)" + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "text", + "content": "-th layer output of VGG19, and " + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "inline_equation", + "content": "Up(\\cdot)" + }, + { + "bbox": [ + 55, + 355, + 296, + 416 + ], + "type": "text", + "content": " indicates bilinear upsampling. The histogram of the resulting similarity map exhibits a bimodal distribution as shown in Fig. 5 (b), enabling dynamic region mask generation through Otsu's [27] method" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 119, + 426, + 294, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 426, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 119, + 426, + 294, + 441 + ], + "type": "interline_equation", + "content": "M a s k = \\mathbf {1} _ {\\text {S i m} < \\text {O t s u} (\\text {S i m})}, \\tag {12}", + "image_path": "5b4fef1a9e7a7337713636f110521c907d6a6d3fb2c0aae86ee658b92c9b0425.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "inline_equation", + "content": "\\mathbf{1}_{\\{\\cdot\\}}" + }, + { + "bbox": [ + 55, + 449, + 295, + 485 + ], + "type": "text", + "content": " denotes the indicator function, which returns 1 if the condition is true. The mask is then multiplied by the ground truth image to extract the dynamic region." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 486, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 569 + ], + "type": "text", + "content": "Decomposition Correspondence from 2D to 3D. To extend 2D dynamic-static decomposition to 3D Gaussians, we establish view-independent correspondences by leveraging depth information from 3D-GS rendering. By unprojecting pixels from masked dynamic regions across multiple views and merging the resulting 3D points, we obtain a comprehensive representation of dynamic regions in 3D space." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Next, we map merged points to dynamic Gaussians based on spatial proximity. Each point expands spherically with radius " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "r" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " to form a 3D volume (Fig. 5 (c)), initially classifying enclosed Gaussians as dynamic and others as static. To overcome potential decomposition inaccuracies and radius sensitivity, we implement a buffer-based soft decomposition strategy using two radii, " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ". Gaussians within " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "r_1" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are marked as dynamic, beyond " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "r_2" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " as static, while those between are pruned to create a buffer zone. This strategy enables 3D-GS to optimize decomposition boundaries through adaptive density control (ADC) [17], enhancing both rendering quality and speed. As demonstrated in Fig. 12, the" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "spans": [ + { + "bbox": [ + 313, + 285, + 553, + 307 + ], + "type": "text", + "content": "strategy also exhibits improved robustness to radius parameter selection." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 308, + 554, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 356 + ], + "type": "text", + "content": "It should be noted that, our DSD method is performed only once during the entire training process and requires only about one minute, introducing minimal overhead to the training pipeline." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": "Joint Rendering of Dynamic and Static Gaussians. With the decomposed dynamic and static Gaussians, we jointly render the entire dynamic scene. Particularly, a deformation field [46] learns to map dynamic Gaussians from canonical space to arbitrary time. Taking time " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": " and the center position " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": " of dynamic Gaussians as inputs, the deformation field outputs the displacement of their position " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\delta_{\\mu}" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": ", rotation " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\delta_r" + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "text", + "content": ", and scaling " + }, + { + "bbox": [ + 313, + 357, + 555, + 453 + ], + "type": "inline_equation", + "content": "\\delta_s" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 361, + 460, + 553, + 474 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 460, + 553, + 474 + ], + "spans": [ + { + "bbox": [ + 361, + 460, + 553, + 474 + ], + "type": "interline_equation", + "content": "\\left(\\delta_ {\\mu}, \\delta_ {r}, \\delta_ {s}\\right) = \\mathcal {F} _ {\\theta} (\\gamma (s g (\\mu)), \\gamma (t)), \\tag {13}", + "image_path": "e50a00859d499b359743e10066f436095562b344b0b300f225e96984d64846eb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "inline_equation", + "content": "sg(\\cdot)" + }, + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "text", + "content": " indicates a stop-gradient operation and " + }, + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "inline_equation", + "content": "\\gamma (\\cdot)" + }, + { + "bbox": [ + 313, + 481, + 554, + 518 + ], + "type": "text", + "content": " denotes the positional encoding [46]. Then, the deformed dynamic gaussians can be addressed as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 525, + 553, + 539 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 525, + 553, + 539 + ], + "spans": [ + { + "bbox": [ + 351, + 525, + 553, + 539 + ], + "type": "interline_equation", + "content": "\\left(\\boldsymbol {\\mu} ^ {\\prime}, \\boldsymbol {r} ^ {\\prime}, \\boldsymbol {s} ^ {\\prime}\\right) = \\left(\\boldsymbol {\\mu} + \\delta_ {\\boldsymbol {\\mu}}, \\boldsymbol {r} + \\delta_ {\\boldsymbol {r}}, \\boldsymbol {s} + \\delta_ {\\boldsymbol {s}}\\right). \\tag {14}", + "image_path": "fb6898521174b626689c9840d78507f2dfb10f4235f45526c19166e32d47b217.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 546, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 554, + 583 + ], + "type": "text", + "content": "Finally, static Gaussians bypass the deformation field and merge with the deformed dynamic Gaussians as inputs to the rasterizer, enabling high-frame-rate dynamic rendering." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 593, + 391, + 605 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 391, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 391, + 605 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 612, + 441, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 441, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 441, + 625 + ], + "type": "text", + "content": "4.1. Experimental Settings" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 713 + ], + "type": "text", + "content": "Datasets. Current datasets for event-based dynamic scene reconstruction are highly limited, with only three synthetic and three real-world scenes [25], all unpublished. Notably, the only three publicly available real-world scenes [25] were captured with a static camera, making novel view evaluation infeasible. To facilitate future research, we build the first event-inclusive 4D benchmark featuring 8 synthetic" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28338" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 95, + 553, + 275 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 93 + ], + "type": "text", + "content": "Table 2. Quantitative results on our synthetic dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events. The best and second-best result are marked in bold and underlined, respectively." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 95, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 59, + 95, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 59, + 95, + 553, + 275 + ], + "type": "table", + "html": "
MethodLegoHotdogMaterialsMusic box
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]23.600.9180.08822330.010.9510.06426028.070.9670.06126219.200.9050.122239
TiNeuVox [7]22.390.8910.0710.5330.810.9530.0350.4926.630.9380.0540.5220.450.8310.1520.62
K-Planes [9]24.550.9310.0352.3431.360.9580.0162.3530.620.9760.0092.2920.770.8580.0712.39
4D-GS [42]26.300.9370.07210433.480.9650.05213230.400.9790.05411124.060.9370.07164
Deformable-3DGS [46]23.790.9230.0537332.910.9620.01713234.000.9860.0049122.080.9240.05251
Event-4DGS28.000.9430.0405434.610.9690.0199635.600.9890.0067428.580.9500.04342
Ours31.850.9670.01818936.150.9740.01324138.020.9930.00324030.780.9630.02992
MethodCelestial globeFanWater wheelMan
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
3D-GS [17]19.050.9150.11018221.260.8910.11827019.430.8870.10921520.790.8700.114210
TiNeuVox [7]13.620.7360.2900.6219.900.8890.1070.5317.030.8500.1470.5622.810.8870.0710.50
K-Planes [9]15.490.8570.0882.4622.100.9090.0622.3820.960.9200.0462.4121.020.8570.0732.33
4D-GS [42]20.970.9420.0725225.050.9360.08010820.960.9170.0777723.800.9140.07664
Deformable-3DGS [46]23.070.9620.0364124.660.9290.0519020.790.9120.0514323.060.9060.05137
Event-4DGS24.300.9480.0453627.660.9490.0417126.340.9320.0523025.550.9210.06333
Ours28.830.9760.0207330.180.9640.02516828.470.9500.03311228.210.9430.037129
", + "image_path": "25f6230e20b752f137c2060e18548b97dc2532698e7a45f8349c7120611a82dc.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 290, + 552, + 355 + ], + "blocks": [ + { + "bbox": [ + 60, + 277, + 550, + 289 + ], + "lines": [ + { + "bbox": [ + 60, + 277, + 550, + 289 + ], + "spans": [ + { + "bbox": [ + 60, + 277, + 550, + 289 + ], + "type": "text", + "content": "Table 3. Quantitative results on our real-world dataset. Event-4DGS is an extension of Deformable-3DGS [46] by incorporating events." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 290, + 552, + 355 + ], + "lines": [ + { + "bbox": [ + 58, + 290, + 552, + 355 + ], + "spans": [ + { + "bbox": [ + 58, + 290, + 552, + 355 + ], + "type": "table", + "html": "
MethodExcavatorJeepFlowersEagle
PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑PSNR↑SSIM↑LPIPS↓FPS↑
4D-GS [42]28.350.9110.11011528.340.8780.0936126.820.8730.1236327.590.9000.128105
Deformable-3DGS [46]26.120.9030.1208126.300.8700.1045226.400.9030.0796427.440.9030.12570
Event-4DGS29.670.9140.0925729.640.9010.0794727.530.9050.0844029.080.8960.10463
Ours31.280.9250.07017930.410.9050.0688928.570.9130.06914931.290.9180.074192
", + "image_path": "77edba42499e35ecfe1c62fb98e27603efedd1a9037531464040b83af849bf2f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 400 + ], + "type": "text", + "content": "and 4 real-world dynamic scenes, encompassing diverse complexities, intricate structures, and rapid motions, thus enabling effective evaluation of dynamic reconstruction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 401, + 295, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 401, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 54, + 401, + 295, + 521 + ], + "type": "text", + "content": "For synthetic scenes, we use Blender [3] to generate one-second, " + }, + { + "bbox": [ + 54, + 401, + 295, + 521 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 54, + 401, + 295, + 521 + ], + "type": "text", + "content": " monocular camera rotations, producing thousands of continuous frames per scene. These high-temporal-resolution sequences are processed through ESIM [31] to generate events. For each sequence, we uniformly sample 30 frames (equivalent to 30 FPS) for training, and select intermediate frames as far apart as possible from the training frames for testing. Particularly, both \"Fan\" representing a typical high-speed 4D scene, and \"Man\" featuring large object displacements, present significant challenges." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 628 + ], + "type": "text", + "content": "For real-world scenes, as shown in Fig. 6, we construct a hybrid camera system consisting of a beam splitter, an event camera (Prophesee Gen4), a frame camera (Basler ace), and a microcontroller (STM32) for outputting synchronization signals. Following [32], we keep the camera system static and place the objects on a motorized optical rotating turntable, which is equivalent to camera motion. Following prior work [25], we downsample the original high-FPS video for training and use intermediate frames for testing." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 630, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 712 + ], + "type": "text", + "content": "Baselines. For RGB-only settings, we benchmarked our method against the representative NeRF baselines K-Planes [9] and TiNeuVox [7], along with Gaussians baselines 3D-GS [17], 4D-GS [42], and Deformable-3DGS [46]. For event-assisted settings, DE-NeRF [25] is the only baseline; however, it could not be directly compared, as its code was still closed-source. DE-NeRF relies on NeRF's volume ren" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 334, + 363, + 533, + 463 + ], + "blocks": [ + { + "bbox": [ + 334, + 363, + 533, + 463 + ], + "lines": [ + { + "bbox": [ + 334, + 363, + 533, + 463 + ], + "spans": [ + { + "bbox": [ + 334, + 363, + 533, + 463 + ], + "type": "image", + "image_path": "4837165c12f3e9cca44395f814b8bb07f094bfdf45f1cadec074b25a7b391136.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 465, + 555, + 488 + ], + "lines": [ + { + "bbox": [ + 313, + 465, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 555, + 488 + ], + "type": "text", + "content": "Figure 6. Real-world data acquisition setup (left) and our hybrid camera system (right)." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 497, + 555, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 497, + 555, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 497, + 555, + 569 + ], + "type": "text", + "content": "dering techniques [26], leading to predictably slow rendering speeds. Moreover, its reconstruction quality is also predictably limited due to the absence of threshold modeling for events. To provide a comparable baseline, we introduce Event-4DGS, an extension of Deformable-3DGS [46] that incorporates the event rendering loss in Eq. (4)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 569, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 569, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 569, + 555, + 605 + ], + "type": "text", + "content": "Metrics. We evaluate rendering quality using PSNR, SSIM [37], and LPIPS [50] (based on AlexNet) and measure rendering speed in FPS on an NVIDIA RTX 3090 GPU." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 612, + 400, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 400, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 400, + 625 + ], + "type": "text", + "content": "4.2. Comparisons" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 555, + 713 + ], + "type": "text", + "content": "Quantitative Results. We report the quantitative results of the comparison on the synthetic and real-world datasets in Tab. 2 and Tab. 3, respectively. Although 4D-GS [42] and Deformable-3DGS [46] achieve relatively higher FPS compared to NeRF baselines [7, 9], their reconstruction quality is limited by the sparsity of the RGB training frames. In contrast, Event-4DGS leverages the rich intermediate mo" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28339" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 81, + 69, + 526, + 369 + ], + "blocks": [ + { + "bbox": [ + 81, + 69, + 526, + 369 + ], + "lines": [ + { + "bbox": [ + 81, + 69, + 526, + 369 + ], + "spans": [ + { + "bbox": [ + 81, + 69, + 526, + 369 + ], + "type": "image", + "image_path": "55a10940b76dbe35079c00e8a2e212076c7d4f52971277923b0b2b3f8c2a20e6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 107, + 370, + 501, + 381 + ], + "lines": [ + { + "bbox": [ + 107, + 370, + 501, + 381 + ], + "spans": [ + { + "bbox": [ + 107, + 370, + 501, + 381 + ], + "type": "text", + "content": "Figure 7. Qualitative comparisons on our synthetic dataset. Please see the supplementary video for details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 81, + 383, + 526, + 579 + ], + "blocks": [ + { + "bbox": [ + 81, + 383, + 526, + 579 + ], + "lines": [ + { + "bbox": [ + 81, + 383, + 526, + 579 + ], + "spans": [ + { + "bbox": [ + 81, + 383, + 526, + 579 + ], + "type": "image", + "image_path": "5b8ea59bf6d16982b93e27873198469c02f1862e1a213f7a8be0d3050c31edbb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 580, + 504, + 591 + ], + "lines": [ + { + "bbox": [ + 104, + 580, + 504, + 591 + ], + "spans": [ + { + "bbox": [ + 104, + 580, + 504, + 591 + ], + "type": "text", + "content": "Figure 8. Qualitative comparisons on our real-world dataset. Please see the supplementary video for details." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 601, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 601, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 601, + 295, + 708 + ], + "type": "text", + "content": "tion and viewpoint information provided by events, significantly outperforming other baselines in reconstruction quality, with an average PSNR improvement of 3.28 dB over Deformable-3DGS across all synthetic scenes. This notable improvement underscores the effectiveness of high-temporal-resolution event cameras for dynamic scene reconstruction. However, Event-4DGS still suffers from threshold variation, whereas our method with GTJM enables accurate threshold modeling and better event super" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "content": "vision, achieving an average PSNR improvement of 2.73 dB over Event-4DGS on synthetic datasets. Meanwhile, our method maintains exceptionally fast rendering speeds, averaging " + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "inline_equation", + "content": "1.71 \\times" + }, + { + "bbox": [ + 313, + 601, + 555, + 685 + ], + "type": "text", + "content": " faster than 4D-GS on synthetic datasets. In summary, our method enjoys both the highest rendering quality and exceptional rendering speed on both synthetic and real-world datasets." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Qualitative Results. For a more visual assessment, we present qualitative results on the synthetic and real-world" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28340" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 93, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 4. Ablation studies on synthetic dataset. For real-world ablation studies, please refer to the supplementary document." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 93, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 58, + 93, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 58, + 93, + 294, + 167 + ], + "type": "table", + "html": "
MethodPSNR↑SSIM↑LPIPS↓FPS↑
w/o GTJM29.390.9560.034153
w/o Joint Optimization in GTJM30.870.9630.026152
w/o DSD30.780.9610.02657
w/o Buffer-based Soft Decomposition31.020.9630.025138
Full31.560.9660.022156
", + "image_path": "08705781c372cdec7a26040c40ac742de40f88930948f974e4a91ad27608aa48.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 230 + ], + "type": "text", + "content": "datasets in Fig. 7 and Fig. 8, respectively. These comparisons highlight the capability of our method to deliver high-fidelity dynamic scene modeling. Notably, our method effectively captures intricate motion details, while other baselines exhibit structural deficiencies and distortions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 232, + 296, + 339 + ], + "type": "text", + "content": "Dynamic Blurry Scene Comparisons. Motion blur is another common challenge in dynamic scenes. To address this, we extend both baselines and our method with blur loss and EDI from [48], and build blurry scenes for evaluations. Fig. 9 shows that, by leveraging events' deblurring advantage, our method outperforms Deformable-3DGS by 4.79 dB in PSNR, achieving the best results. For detailed methods and quantitative results, see the supplementary document." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 149, + 360 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 365, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 296, + 521 + ], + "type": "text", + "content": "GS-threshold Joint Modeling. Using a constant threshold fails to properly neutralize opposing polarity events during accumulation, resulting in motion trajectory artifacts as shown in Fig. 10 (b). These artifacts, when used for Gaussian supervision, produce undesirable purple haze in rendered outputs, such as the Event-4DGS results in Fig. 7. Our RGB-assisted threshold estimation significantly reduces these artifacts (Fig. 10 (c)), while subsequent joint threshold and GS optimization effectively eliminates remaining distortions (Fig. 10 (d)). As demonstrated in Tab. 4, this improved event supervision yields a " + }, + { + "bbox": [ + 55, + 365, + 296, + 521 + ], + "type": "inline_equation", + "content": "2.17\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 365, + 296, + 521 + ], + "type": "text", + "content": " average PSNR improvement across all scenes, validating our GTJM strategy's effectiveness in handling threshold variations." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": "Dynamic-static Decomposition. Our DSD method successfully identifies dynamic regions of varying sizes and geometries, as demonstrated in Fig. 11. Modeling the entire scene with dynamic Gaussians without DSD misallocates deformation field capacity to static regions, compromising dynamic region reconstruction quality as shown in Fig. 4. Quantitative results in Tab. 4 demonstrate that using DSD improves the average PSNR by " + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "inline_equation", + "content": "0.78\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": " and accelerates the rendering speed to 2.74 times the original FPS. This underscores DSD's crucial role in achieving both high-fidelity dynamic scene reconstruction and efficient rendering." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "text", + "content": "Buffer-based Soft Decomposition. Our buffer-based soft decomposition enables adaptive optimization of decomposition boundaries, yielding a 0.54 dB improvement in average PSNR (Tab. 4). Sensitivity analysis reveals that reconstruction quality stabilizes when buffer size " + }, + { + "bbox": [ + 55, + 653, + 296, + 715 + ], + "type": "inline_equation", + "content": "(r_2 - r_1)" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 321, + 71, + 369, + 121 + ], + "blocks": [ + { + "bbox": [ + 321, + 71, + 369, + 121 + ], + "lines": [ + { + "bbox": [ + 321, + 71, + 369, + 121 + ], + "spans": [ + { + "bbox": [ + 321, + 71, + 369, + 121 + ], + "type": "image", + "image_path": "43c8ae281459886f7adf86aa88137b11d9d07dfa13ba5de358b78ccebfe4c4bf.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 321, + 123, + 547, + 134 + ], + "lines": [ + { + "bbox": [ + 321, + 123, + 547, + 134 + ], + "spans": [ + { + "bbox": [ + 321, + 123, + 547, + 134 + ], + "type": "text", + "content": "Figure 9. Extended comparisons on the dynamic blurry scene." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 396, + 121 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 396, + 121 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 396, + 121 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 396, + 121 + ], + "type": "image", + "image_path": "57e620d0531d10c7d0e089b7d89b72ac5e1b00708fea5dae23063395b7c52e77.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 399, + 72, + 443, + 121 + ], + "blocks": [ + { + "bbox": [ + 399, + 72, + 443, + 121 + ], + "lines": [ + { + "bbox": [ + 399, + 72, + 443, + 121 + ], + "spans": [ + { + "bbox": [ + 399, + 72, + 443, + 121 + ], + "type": "image", + "image_path": "cd504b106d097fe1f5c4c75d83033e87b52371c70da886e1486c6563cbb16c84.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 444, + 72, + 472, + 121 + ], + "blocks": [ + { + "bbox": [ + 444, + 72, + 472, + 121 + ], + "lines": [ + { + "bbox": [ + 444, + 72, + 472, + 121 + ], + "spans": [ + { + "bbox": [ + 444, + 72, + 472, + 121 + ], + "type": "image", + "image_path": "eb78c0a3864da654c8d61b6868639ad13d787540e2c945d46a3cab0a4c53f038.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 72, + 519, + 121 + ], + "blocks": [ + { + "bbox": [ + 474, + 72, + 519, + 121 + ], + "lines": [ + { + "bbox": [ + 474, + 72, + 519, + 121 + ], + "spans": [ + { + "bbox": [ + 474, + 72, + 519, + 121 + ], + "type": "image", + "image_path": "7dcfd7deda0d68aa09981543df298986b22ea8900ab46724b372af067d760bfe.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 521, + 72, + 547, + 121 + ], + "blocks": [ + { + "bbox": [ + 521, + 72, + 547, + 121 + ], + "lines": [ + { + "bbox": [ + 521, + 72, + 547, + 121 + ], + "spans": [ + { + "bbox": [ + 521, + 72, + 547, + 121 + ], + "type": "image", + "image_path": "5d372c9d3ebd06f9071764a067c54c17a57427531103b1bd86a4ae41edd81428.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 320, + 136, + 376, + 247 + ], + "blocks": [ + { + "bbox": [ + 320, + 136, + 376, + 247 + ], + "lines": [ + { + "bbox": [ + 320, + 136, + 376, + 247 + ], + "spans": [ + { + "bbox": [ + 320, + 136, + 376, + 247 + ], + "type": "image", + "image_path": "3e641b04f76e0d0925fd3be5f4923fa51518e238ebc8d8c4eb8a64fd65560154.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 247, + 358, + 254 + ], + "lines": [ + { + "bbox": [ + 339, + 247, + 358, + 254 + ], + "spans": [ + { + "bbox": [ + 339, + 247, + 358, + 254 + ], + "type": "text", + "content": "(a) GT" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 377, + 136, + 434, + 247 + ], + "blocks": [ + { + "bbox": [ + 377, + 136, + 434, + 247 + ], + "lines": [ + { + "bbox": [ + 377, + 136, + 434, + 247 + ], + "spans": [ + { + "bbox": [ + 377, + 136, + 434, + 247 + ], + "type": "image", + "image_path": "6400539520cccf496f75b6598a337f0807b97faa780bed6a046e9607e0715d91.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 384, + 247, + 427, + 254 + ], + "lines": [ + { + "bbox": [ + 384, + 247, + 427, + 254 + ], + "spans": [ + { + "bbox": [ + 384, + 247, + 427, + 254 + ], + "type": "text", + "content": "(b) w/o GTJM" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 430, + 247, + 496, + 254 + ], + "lines": [ + { + "bbox": [ + 430, + 247, + 496, + 254 + ], + "spans": [ + { + "bbox": [ + 430, + 247, + 496, + 254 + ], + "type": "text", + "content": "c) w/o Joint Optimization" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 434, + 136, + 491, + 247 + ], + "blocks": [ + { + "bbox": [ + 434, + 136, + 491, + 247 + ], + "lines": [ + { + "bbox": [ + 434, + 136, + 491, + 247 + ], + "spans": [ + { + "bbox": [ + 434, + 136, + 491, + 247 + ], + "type": "image", + "image_path": "4ec7e36bf2924f7ffd0658f4dfa86d60e695575564f5df5217d1c37166c42d59.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 449, + 255, + 474, + 262 + ], + "lines": [ + { + "bbox": [ + 449, + 255, + 474, + 262 + ], + "spans": [ + { + "bbox": [ + 449, + 255, + 474, + 262 + ], + "type": "text", + "content": "in GTJM" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 491, + 136, + 547, + 247 + ], + "blocks": [ + { + "bbox": [ + 491, + 136, + 547, + 247 + ], + "lines": [ + { + "bbox": [ + 491, + 136, + 547, + 247 + ], + "spans": [ + { + "bbox": [ + 491, + 136, + 547, + 247 + ], + "type": "image", + "image_path": "9b2b290f2151706660f24c79818c96da05bc2de5953cc34195e24643c2f18e3b.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 507, + 247, + 528, + 254 + ], + "lines": [ + { + "bbox": [ + 507, + 247, + 528, + 254 + ], + "spans": [ + { + "bbox": [ + 507, + 247, + 528, + 254 + ], + "type": "text", + "content": "(d) Full" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 325, + 286, + 376, + 365 + ], + "blocks": [ + { + "bbox": [ + 313, + 263, + 553, + 285 + ], + "lines": [ + { + "bbox": [ + 313, + 263, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 553, + 285 + ], + "type": "text", + "content": "Figure 10. The effect of GS-threshold joint modeling strategy, which eliminates event artifacts caused by threshold variations." + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 325, + 286, + 376, + 365 + ], + "lines": [ + { + "bbox": [ + 325, + 286, + 376, + 365 + ], + "spans": [ + { + "bbox": [ + 325, + 286, + 376, + 365 + ], + "type": "image", + "image_path": "b286634b0b8acac45b7eae191b2554ffeab58c56741a496b7ba91b7b4027d41c.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 367, + 406, + 374 + ], + "lines": [ + { + "bbox": [ + 347, + 367, + 406, + 374 + ], + "spans": [ + { + "bbox": [ + 347, + 367, + 406, + 374 + ], + "type": "text", + "content": "Lego Hotdog" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 377, + 366, + 406, + 374 + ], + "lines": [ + { + "bbox": [ + 377, + 366, + 406, + 374 + ], + "spans": [ + { + "bbox": [ + 377, + 366, + 406, + 374 + ], + "type": "text", + "content": "Hotdog" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 332, + 401, + 535, + 504 + ], + "blocks": [ + { + "bbox": [ + 408, + 366, + 451, + 374 + ], + "lines": [ + { + "bbox": [ + 408, + 366, + 451, + 374 + ], + "spans": [ + { + "bbox": [ + 408, + 366, + 451, + 374 + ], + "type": "text", + "content": "Materials" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 452, + 366, + 466, + 373 + ], + "lines": [ + { + "bbox": [ + 452, + 366, + 466, + 373 + ], + "spans": [ + { + "bbox": [ + 452, + 366, + 466, + 373 + ], + "type": "text", + "content": "Cel" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 468, + 366, + 496, + 373 + ], + "lines": [ + { + "bbox": [ + 468, + 366, + 496, + 373 + ], + "spans": [ + { + "bbox": [ + 468, + 366, + 496, + 373 + ], + "type": "text", + "content": "estial globe" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 498, + 366, + 525, + 373 + ], + "lines": [ + { + "bbox": [ + 498, + 366, + 525, + 373 + ], + "spans": [ + { + "bbox": [ + 498, + 366, + 525, + 373 + ], + "type": "text", + "content": "Man" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 377, + 553, + 399 + ], + "lines": [ + { + "bbox": [ + 313, + 377, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 553, + 399 + ], + "type": "text", + "content": "Figure 11. Rendering results of dynamic and static Gaussians separated by our dynamic-static decomposition strategy." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 332, + 401, + 535, + 504 + ], + "lines": [ + { + "bbox": [ + 332, + 401, + 535, + 504 + ], + "spans": [ + { + "bbox": [ + 332, + 401, + 535, + 504 + ], + "type": "image", + "image_path": "063d96bced7e5dca0cdf0b5d0e453c296433544b033b055abcfea113105dc542.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 505, + 534, + 515 + ], + "lines": [ + { + "bbox": [ + 333, + 505, + 534, + 515 + ], + "spans": [ + { + "bbox": [ + 333, + 505, + 534, + 515 + ], + "type": "text", + "content": "Figure 12. Sensitivity analysis on buffer size " + }, + { + "bbox": [ + 333, + 505, + 534, + 515 + ], + "type": "inline_equation", + "content": "(r_2 - r_1)" + }, + { + "bbox": [ + 333, + 505, + 534, + 515 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "bbox": [ + 313, + 524, + 554, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 524, + 554, + 584 + ], + "spans": [ + { + "bbox": [ + 313, + 524, + 554, + 584 + ], + "type": "text", + "content": "exceeds approximately 12 basic units (normalized by average inter-Gaussian distance to account for scene variations), as shown in Fig. 12. This stability demonstrates the robustness of our DSD method through adaptive boundary search, highlighting the effectiveness of the buffer-based strategy." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "spans": [ + { + "bbox": [ + 313, + 597, + 388, + 609 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "In this paper, we present an event-boosted deformable 3D Gaussian framework for high-quality dynamic scene reconstruction. Our GS-threshold joint modeling effectively addresses threshold variation challenges, enabling reliable event supervision. The proposed dynamic-static decomposition method enhances both rendering efficiency and reconstruction quality through optimized resource allocation between static and dynamic regions." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28341" + } + ] + } + ], + "index": 38 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 159, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 152 + ], + "type": "text", + "content": "This work was supported in part by the National Natural Science Foundation of China under Grants 62472399, 62131003 and 62021001, and the Fundamental Research Funds for the Central Universities under Grant WK210000059." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 115, + 175 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 183, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 183, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 183, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 183, + 296, + 226 + ], + "type": "text", + "content": "[1] Jeongmin Bae, Seoha Kim, Youngsik Yun, Hahyun Lee, Gun Bang, and Youngjung Uh. Per-gaussian embedding-based deformation for deformable 3d gaussian splatting. In ECCV, 2024. 2, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 271 + ], + "type": "text", + "content": "[2] Jonathan T Barron, Ben Mildenhall, Matthew Tancik, Peter Hedman, Ricardo Martin-Brualla, and Pratul P Srinivasan. Mip-nerf: A multiscale representation for anti-aliasing neural radiance fields. In ICCV, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 294, + 295 + ], + "type": "text", + "content": "[3] Blender. A 3d modelling and rendering package. https://www.blender.org/, 2018.6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 297, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 297, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 61, + 297, + 295, + 328 + ], + "type": "text", + "content": "[4] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 295, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 295, + 352 + ], + "type": "text", + "content": "[5] Ang Cao and Justin Johnson. Hexplane: A fast representation for dynamic scenes. In CVPR, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "type": "text", + "content": "[6] Hiroyuki Deguchi, Mana Masuda, Takuya Nakabayashi, and Hideo Saito. E2gs: Event enhanced gaussian splatting. In ICIP, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 431 + ], + "type": "text", + "content": "[7] Jiemin Fang, Taoran Yi, Xinggang Wang, Lingxi Xie, Xiaopeng Zhang, Wenyu Liu, Matthias Nießner, and Qi Tian. Fast dynamic radiance fields with time-aware neural voxels. In SIGGRAPH Asia, 2022. 1, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "spans": [ + { + "bbox": [ + 62, + 434, + 296, + 465 + ], + "type": "text", + "content": "[8] Sara Fridovich-Keil, Alex Yu, Matthew Tancik, Qinhong Chen, Benjamin Recht, and Angjoo Kanazawa. Plenoxels: Radiance fields without neural networks. In CVPR, 2022." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 468, + 296, + 510 + ], + "type": "text", + "content": "[9] Sara Fridovich-Keil, Giacomo Meanti, Frederik Rahbæk Warburg, Benjamin Recht, and Angjoo Kanazawa. K-planes: Explicit radiance fields in space, time, and appearance. In CVPR, 2023. 1, 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 512, + 296, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 512, + 296, + 577 + ], + "spans": [ + { + "bbox": [ + 57, + 512, + 296, + 577 + ], + "type": "text", + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 1, 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 57, + 579, + 295, + 622 + ], + "type": "text", + "content": "[11] Daniel Gehrig, Antonio Loquercio, Konstantinos G Derpanis, and Davide Scaramuzza. End-to-end learning of representations for asynchronous event-based data. In ICCV, 2019. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 296, + 678 + ], + "type": "text", + "content": "[12] Zhiyang Guo, Wengang Zhou, Li Li, Min Wang, and Houqiang Li. Motion-aware 3d gaussian splatting for efficient dynamic scene reconstruction. IEEE Transactions on Circuits and Systems for Video Technology, 35(4):3119-3133, 2024. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 296, + 712 + ], + "type": "text", + "content": "[13] Yuhuang Hu, Shih-Chii Liu, and Tobi Delbruck. v2e: From video frames to realistic dvs events. In CVPR workshop, 2021. 2, 3" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 33, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[14] Yi-Hua Huang, Yang-Tian Sun, Ziyi Yang, Xiaoyang Lyu, Yan-Pei Cao, and Xiaojuan Qi. Sc-gs: Sparse-controlled gaussian splatting for editable dynamic scenes. In CVPR, 2024. 2, 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 119, + 553, + 140 + ], + "type": "text", + "content": "[15] Inwoo Hwang, Junho Kim, and Young Min Kim. Ev-nerf: Event based neural radiance field. In WACV, 2023. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 143, + 554, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 143, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 143, + 554, + 175 + ], + "type": "text", + "content": "[16] Hankyu Jang and Daeyoung Kim. D-tensorf: Tensorial radiance fields for dynamic scenes. arXiv preprint arXiv:2212.02375, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 177, + 554, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 554, + 220 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 554, + 220 + ], + "type": "text", + "content": "[17] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Transactions on Graphics, 42 (4):139-1, 2023. 1, 2, 3, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 554, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 554, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 554, + 266 + ], + "type": "text", + "content": "[18] Simon Klenk, Lukas Koestler, Davide Scaramuzza, and Daniel Cremers. E-nerf: Neural radiance fields from a moving event camera. IEEE Robotics and Automation Letters, 8 (3):1587–1594, 2023. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "spans": [ + { + "bbox": [ + 316, + 268, + 554, + 312 + ], + "type": "text", + "content": "[19] Chenghan Li, Christian Brandli, Raphael Berner, Hongjie Liu, Minhao Yang, Shih-Chii Liu, and Tobi Delbruck. Design of an rgb color vga rolling and global shutter dynamic and active-pixel vision sensor. In ISCAS, 2015. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 315, + 553, + 357 + ], + "type": "text", + "content": "[20] Yiqing Liang, Numair Khan, Zhengqin Li, Thu Nguyen-Phuoc, Douglas Lanman, James Tompkin, and Lei Xiao. Gaufre: Gaussian deformation fields for real-time dynamic novel view synthesis. In WACV, 2025. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 359, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 359, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 359, + 553, + 392 + ], + "type": "text", + "content": "[21] Yutong Liu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. S2d-lfe: Sparse-to-dense light field event generation. In CVPR, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 426 + ], + "type": "text", + "content": "[22] Weng Fei Low and Gim Hee Lee. Robust e-nerf: Nerf from sparse & noisy events under non-uniform motion. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 472 + ], + "type": "text", + "content": "[23] Zhicheng Lu, Xiang Guo, Le Hui, Tianrui Chen, Min Yang, Xiao Tang, Feng Zhu, and Yuchao Dai. 3d geometry-aware deformable gaussian splattering for dynamic view synthesis. In CVPR, 2024. 2, 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 507 + ], + "type": "text", + "content": "[24] Jonathon Luiten, Georgios Kopanas, Bastian Leibe, and Deva Ramanan. Dynamic 3d gaussians: Tracking by persistent dynamic view synthesis. In 3DV, 2024. 1, 2, 4" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 509, + 553, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 509, + 553, + 541 + ], + "spans": [ + { + "bbox": [ + 316, + 509, + 553, + 541 + ], + "type": "text", + "content": "[25] Qi Ma, Danda Pani Paudel, Ajad Chhatkuli, and Luc Van Gool. Deformable neural radiance fields using rgb and event cameras. In ICCV, 2023. 2, 3, 5, 6" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 544, + 553, + 597 + ], + "type": "text", + "content": "[26] Ben Mildenhall, Pratul P Srinivasan, Matthew Tancik, Jonathan T Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. Communications of the ACM, 65(1):99-106, 2021. 1, 2, 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 600, + 554, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 554, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 554, + 633 + ], + "type": "text", + "content": "[27] Nobuyuki Otsu. A threshold selection method from gray-level histograms. IEEE Transactions on Systems, Man, and Cybernetics, 9(1):62-66, 1979. 5" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[28] Keunhong Park, Utkarsh Sinha, Jonathan T Barron, Sofien Bouaziz, Dan B Goldman, Steven M Seitz, and Ricardo Martin-Brualla. Nerfies: Deformable neural radiance fields. In ICCV, 2021. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[29] Albert Pumarola, Enric Corona, Gerard Pons-Moll, and Francesc Moreno-Noguer. D-nerf: Neural radiance fields for dynamic scenes. In CVPR, 2021. 2" + } + ] + } + ], + "index": 32 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "28342" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[30] Yunshan Qi, Lin Zhu, Yu Zhang, and Jia Li. E2nerf: Event enhanced neural radiance fields from blurry images. In ICCV, 2023. 2, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "text", + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. Esim: an open event camera simulator. In CoRL, 2018. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "text", + "content": "[32] Viktor Rudnev, Mohamed Elgharib, Christian Theobalt, and Vladislav Golyanik. Eventnerf: Neural radiance fields from a single colour event camera. In CVPR, 2023. 2, 3, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 161, + 294, + 205 + ], + "type": "text", + "content": "[33] Ruizhi Shao, Zerong Zheng, Hanzhang Tu, Boning Liu, Hongwen Zhang, and Yebin Liu. Tensor4d: Efficient neural 4d decomposition for high-fidelity dynamic reconstruction and rendering. In CVPR, 2023. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "text", + "content": "[34] Richard Shaw, Michal Nazarczuk, Jifei Song, Arthur Moreau, Sibi Catley-Chandar, Helisa Dhamo, and Eduardo Pérez-Pellitero. Swings: sliding windows for dynamic 3d gaussian splatting. In ECCV, 2024. 1, 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 272 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 272 + ], + "type": "text", + "content": "[35] K Simonyan and A Zisserman. Very deep convolutional networks for large-scale image recognition. In ICLR, 2015. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 316 + ], + "type": "text", + "content": "[36] Dor Verbin, Peter Hedman, Ben Mildenhall, Todd Zickler, Jonathan T Barron, and Pratul P Srinivasan. Ref-nerf: Structured view-dependent appearance for neural radiance fields. In CVPR, 2022. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 294, + 360 + ], + "type": "text", + "content": "[37] Zhou Wang, Alan C Bovik, Hamid R Sheikh, and Eero P Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing, 13(4):600-612, 2004. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 360, + 294, + 392 + ], + "type": "text", + "content": "[38] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In ICCV, 2021. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 392, + 294, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 392, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 392, + 294, + 426 + ], + "type": "text", + "content": "[39] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Boosting event stream super-resolution with a recurrent neural network. In ECCV, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[40] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In CVPR, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 459, + 294, + 503 + ], + "type": "text", + "content": "[41] Yuchen Weng, Zhengwen Shen, Ruofan Chen, Qi Wang, and Jun Wang. Eadeblur-gs: Event assisted 3d deblur reconstruction with gaussian splatting. arXiv preprint arXiv:2407.13520, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 503, + 294, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 548 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 548 + ], + "type": "text", + "content": "[42] Guanjun Wu, Taoran Yi, Jiemin Fang, Lingxi Xie, Xiaopeng Zhang, Wei Wei, Wenyu Liu, Qi Tian, and Xinggang Wang. 4d gaussian splatting for real-time dynamic scene rendering. In CVPR, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[43] Jingqian Wu, Shuo Zhu, Chutian Wang, and Edmund Y Lam. Ev-gs: Event-based gaussian splatting for efficient and accurate radiance field rendering. In MLSP, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 624 + ], + "type": "text", + "content": "[44] Tianyi Xiong, Jiayi Wu, Botao He, Cornelia Fermuller, Yiannis Aloimonos, Heng Huang, and Christopher Metzler. Event3dgs: Event-based 3d gaussian splatting for high-speed robot egomotion. In CoRL, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[45] Wenhao Xu, Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Ceia: Clip-based event-image alignment for open-world event-based understanding. In ECCV workshop, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[46] Ziyi Yang, Xinyu Gao, Wen Zhou, Shaohui Jiao, Yuqing Zhang, and Xiaogang Jin. Deformable 3d gaussians for high-fidelity monocular dynamic scene reconstruction. In CVPR, 2024. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 230 + ], + "type": "list", + "angle": 0, + "index": 22, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[47] Zeyu Yang, Hongye Yang, Zijie Pan, and Li Zhang. Real-time photorealistic dynamic scene representation and rendering with 4d gaussian splatting. In ICLR, 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 106, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 106, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 553, + 150 + ], + "type": "text", + "content": "[48] Wangbo Yu, Chaoran Feng, Jiye Tang, Xu Jia, Li Yuan, and Yonghong Tian. Evagaussians: Event stream assisted gaussian splatting from blurry images. arXiv preprint arXiv:2405.20224, 2024. 2, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 150, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 150, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 315, + 150, + 553, + 196 + ], + "type": "text", + "content": "[49] Bowen Zhang, Yiji Cheng, Jiaolong Yang, Chunyu Wang, Feng Zhao, Yansong Tang, Dong Chen, and Baining Guo. Gaussiancube: a structured and explicit radiance representation for 3d generative modeling. In NeurIPS, 2024. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 196, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 196, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 315, + 196, + 553, + 230 + ], + "type": "text", + "content": "[50] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 21 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "28343" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_content_list.json b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..a66d090832633f7d57dd0f3b62d8b621f675a951 --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_content_list.json @@ -0,0 +1,1796 @@ +[ + { + "type": "text", + "text": "Event-guided HDR Reconstruction with Diffusion Priors", + "text_level": 1, + "bbox": [ + 212, + 130, + 785, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yixin Yang $^{1,2,\\dagger}$ Jiawei Zhang $^{3}$ Yang Zhang $^{1,2}$", + "bbox": [ + 305, + 179, + 679, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yunxuan Wei $^{3}$ Dongqing Zou $^{4}$ Jimmy S. Ren $^{3,5}$ Boxin Shi $^{1,2,*}$", + "bbox": [ + 238, + 198, + 754, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University", + "bbox": [ + 116, + 215, + 880, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University", + "bbox": [ + 114, + 234, + 880, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ SenseTime Research $^{4}$ PBVR $^{5}$ Hong Kong Metropolitan University", + "bbox": [ + 220, + 251, + 774, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{yangyixin93, shiboxin}@pku.edu.cn, Github Page: github.com/YixinYang-00/HDRev-Diff", + "bbox": [ + 187, + 271, + 807, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Events provide High Dynamic Range (HDR) intensity change which can guide Low Dynamic Range (LDR) image for HDR reconstruction. However, events only provide temporal intensity differences and it is still ill-posed in over-/under-exposed areas due to missing initial reference brightness and color information. With strong generation ability, diffusion models have shown their potential for tackling ill-posed problems. Therefore, we introduce conditional diffusion models to hallucinate missing information. Whereas, directly adopting events and LDR image as conditions is complicated for diffusion models to sufficiently utilize their information. Thus we introduce a pretrained events-image encoder tailored for HDR reconstruction and a pyramid fusion module to provide HDR conditions, which can be efficiently and effectively utilized by the diffusion model. Moreover, the generation results of diffusion models usually exhibit distortion, particularly for fine-grained details. To better preserve fidelity and suppress distortion, we propose a fine-grained detail recovery approach using a histogram-based structural loss. Experiments on real and synthetic data show the effectiveness of the proposed method in terms of both detail preservation and information hallucination.", + "bbox": [ + 88, + 340, + 483, + 689 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 717, + 220, + 733 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extending the dynamic range of an image (usually with low dynamic range, LDR) to record the brightness of real scenes plausibly is particularly useful for daily photography and computer vision tasks. High dynamic range (HDR) imaging can be, in the most direct way, achieved by utilizing neural networks to hallucinate missing information in over-/under-exposed areas [4, 5, 22, 24], as shown in Figure 1 (a), which is challenging to faithfully recover", + "bbox": [ + 89, + 743, + 483, + 864 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/62146720e129dfa968c35c47d423c59e5f4698ccedfafb5fdf21f9a824d6bd02.jpg", + "image_caption": [ + "(a) Dille et. al." + ], + "image_footnote": [], + "bbox": [ + 517, + 308, + 638, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c51f95e84dc932d2802ce39488b0639c37b04c35a729cfbf67c117216778dd0.jpg", + "image_caption": [ + "(b) Events" + ], + "image_footnote": [], + "bbox": [ + 640, + 308, + 764, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6f2c90e6df82210e9bd3da228d4f921e98a1810a649abd06611c19768d0947ea.jpg", + "image_caption": [ + "(c) EventHDR" + ], + "image_footnote": [], + "bbox": [ + 782, + 308, + 903, + 402 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ae2c101566f5800d0e01250561b958a7c4bd48811ea929a68dd66440aa727a0.jpg", + "image_caption": [ + "(d) LDR image" + ], + "image_footnote": [], + "bbox": [ + 517, + 420, + 638, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9e79e0012e1eaa40b5ac7c9a54f1fec3426ac3f4b01ec6d8e95ee81e3e1f684c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 640, + 420, + 777, + 541 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a8d7b7f26b188735f773ee416cfae1762946492602decfa1771b4d41dba7fe87.jpg", + "image_caption": [ + "(e) HDRev" + ], + "image_footnote": [], + "bbox": [ + 782, + 420, + 903, + 518 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/dfd498d74bd0f46b0354755a13961313ae0a3d4b1bc598eb7c55aa2a43b8e4e1.jpg", + "image_caption": [ + "(f) Sagiri", + "Figure 1. HDR image can be recovered by only using a single LDR image (d) as input (e.g., Dille et al. [4]), which is a difficult problem as shown in (a). With only events (b) as input, event-to-image reconstruction (e.g., EventHDR [53]) (c) recovers an HDR scene without color. Event-guided HDR methods (e.g., HDRev [48]) (e) take both events and the LDR image as input, which recovers colorful HDR scenes with moderate details. Diffusion-based methods (e.g., Sagiri [22]) (f) take the LDR image (and optionally the text prompt) as input showing visually pleasing results but may have challenges in maintaining consistent contents with the original scene. The proposed method (g) takes events and the LDR image as input and leverages the diffusion priors to recover a plausible and realistic HDR image with rich details and less distortion. scene details without any auxiliary information. Neuromorphic cameras, such as event cameras, are proven to be helpful in providing additional details for HDR purpose [28, 32, 53]. Event cameras capture intensity changes" + ], + "image_footnote": [], + "bbox": [ + 517, + 541, + 638, + 635 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce99c953bdb05d0ee04ced2ae08a814ced006d2ba1e7a2af30c8bb62749a0928.jpg", + "image_caption": [ + "(g) Ours" + ], + "image_footnote": [], + "bbox": [ + 640, + 541, + 764, + 635 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$\\Delta$ : Events", + "$\\Delta$ :Colorful results", + "$\\Delta$ : More realistic", + "$\\Delta$ : Less distortion", + ": CNN models", + "Diffusion models" + ], + "bbox": [ + 776, + 551, + 900, + 627 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "This work is done during Yixin's internship at SenseTime.", + "bbox": [ + 107, + 875, + 423, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11787", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "by triggering events asynchronously, which inherently contains HDR information (e.g., 120 dB for DAVIS340), and events can be directly converted to HDR images, as demonstrated by EventHDR [53] in Figure 1 (c). The event-to-image reconstruction [28, 32, 53] can recover details in over-/under-exposed regions, but fail to provide colorful results due to lacking color information.", + "bbox": [ + 89, + 90, + 480, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate colorful HDR images, event-guided HDR reconstructions [13, 14, 48] take both LDR image and events as inputs, complementarily learning to predict color and intensity in over-/under-exposed areas, as demonstrated in Figure 1 (e). They can reconstruct HDR information, e.g., the cloud in the green box, and provide colorful results. Although regression models are adopted to build the relationship between the predicted pixel values and its neighboring pixels, those methods cannot recover proper intensity, especially when the LDR image has large over-/under-exposed areas. Because absolute HDR intensity (events only contain thresholded radiance changes) and color information are still missing in these regions, it is difficult for regression models to appropriately hallucinate their values.", + "bbox": [ + 89, + 199, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Diffusion models [16, 37-39] are emerging generative models that show great potential for generating plausible and high-quality images. Even if for ill-posed problems, they can provide visually pleasing results thanks to the expressive power of diffusion priors. Diffusion models have also been adopted for image restoration [23, 43, 49]. Recent attempts, e.g., Sagiri [22], have shown the potential of diffusion priors for HDR reconstruction by providing more visually realistic details in over-exposed areas, e.g., the cloud in the sky of Figure 1 (f). However, the generated HDR images still cannot appropriately keep the fidelity of the original scene, and additional distortion are brought into well-exposed regions, as shown by the red box in Figure 1 (f).", + "bbox": [ + 89, + 414, + 482, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To reconstruct visually pleasing and faithful HDR images with less distortion, we introduce events to provide differential HDR intensity and adopt conditional diffusion models [50] to compensate for the ill-posed nature of hallucinating missing information and constrain the generative model to be consistent with the original scene. Specially, we focus on solving the following two key problems via employing diffusion models in event-guided HDR reconstruction: 1) Condition extraction: Since events and LDR image are two different modalities of visual representation, directly adopting them as conditions is non-compatible for the conditional diffusion models to conduct effective fusion. Inspired by the modality alignment of HDRev [48], we introduce an event-image encoder and a pyramid fusion module to fuse events and LDR image in different resolutions and provide conditions to guide the generation process. 2) Details preservation: Applying diffusion models to HDR reconstruction [22, 23] suffers from fine-grained detail distortion, as shown in Figure 1 (f). Such artifacts become", + "bbox": [ + 89, + 613, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "more obvious when LDR images are high quality and have large well-exposed areas. To refine the fine-grained details and make them consistent with LDR images and events, we design a refinement module with histogram-based structure loss. The proposed method is validated on both synthetic and real data and outperforms existing methods by providing fidelity-preserving HDR reconstruction, as illustrated in Figure 1 (g), with the following technical contributions:", + "bbox": [ + 511, + 90, + 903, + 210 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- integrating events and the conditional diffusion models to recover missing information faithfully;", + "- adopting the pretrained event-image encoder and pyramid fusion to effectively apply conditions; and", + "- designing the refinement module and histogram-based structure loss to further strengthen fine-grained details." + ], + "bbox": [ + 511, + 213, + 903, + 301 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related works", + "text_level": 1, + "bbox": [ + 513, + 319, + 660, + 334 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Image-based HDR reconstruction Many approaches leverage the capabilities of neural networks to directly hallucinate HDR information from training data, such as CNNs [5, 6, 9], Generative Adversarial Networks (GANs) [40], and diffusion models [7, 22]. Liu et al. [24] incorporates the LDR image formulation pipeline into its hallucination process to perform step-by-step prediction. For a more comprehensive view, we refer to GTA-HDR [2]. Since information is unavoidably missing in over-/underexposed areas, the above methods only rely on hallucination, which may not correspond to real scenes.", + "bbox": [ + 511, + 338, + 903, + 503 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Event-based HDR reconstruction Benefiting from the HDR property of event cameras, previous works reconstruct HDR images from events with recurrent neural networks [32, 33, 53] or GANs [28, 41]. Since events only record logarithmic irradiance changes, it's difficult to reconstruct color images. To reconstruct color HDR images, Han et al. [13, 14] adopted a hybrid camera system to capture events and frame, enabling event-guided HDR reconstruction. Yang et al. [48] treated events and frame as different modalities and proposed multi-modalities alignment and fusion to provide HDR videos. Other methods [26, 35] incorporate events and bracketed exposure for HDR reconstruction, while long-exposed images bring blurry artifacts. The proposed method takes events and LDR image as input, incorporating diffusion models to compensate for realistic and faithful information in over-/under-exposed areas.", + "bbox": [ + 511, + 508, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generative priors Generative Priors capture the underlying data distribution, enabling the generated images to follow real-world image distribution. GAN [10] has shown its potential to leverage generative priors for image restoration [20, 21, 30]. Recently, diffusion models [16, 37-39] have widely adopted as effective generative priors for producing realistic images [29, 34, 42, 50], which also demonstrate strong potential for image restorations [23, 43, 49], specifically, for HDR image reconstruction from bracketed exposure [12, 17, 46]. There are some attempts [11, 22] to", + "bbox": [ + 511, + 750, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11788", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/48d481cd855603f94ca7c593645d5fe45d625f9a99def7ac4b0b1f6b7bb25e1c.jpg", + "image_caption": [ + "Figure 2. Pipeline of the proposed method consists of three parts: diffusion model and denoising process (green part), event-image conditioning (blue part), and detail refinement (orange part). The blue part fuses LDR image and events to generate injected features for Denoiser as described in Section 3.2. The green part adopts the injected features as conditions to estimate noise from latent $z_{t}$ step by step with fixed pretrained parameters. To refine the distortion of denoised images, the refinement module in the orange part, whose architecture is shown on the right, is proposed to leverage the HDR features and denoised latent $z_{0}$ . Considering the uncertainty of diffusion results in color and brightness, we introduce structure loss, which adopts adaptive Ground Truth (GT) as the supervision target, as described in Section 3.3. After refinement, we obtain a plausible and realistic refined image without distortion." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 277 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "take advantage of diffusion models to generate HDR images from single LDR image, while it cannot provide faithful reconstruction due to the inevitable dynamic clipping.", + "bbox": [ + 88, + 383, + 480, + 429 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 89, + 441, + 264, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Section 3.1 introduces basic concepts about diffusion models and event cameras. The pipeline of the proposed method is illustrated in Figure 2. The green part is the pretrained latent diffusion model, which performs denoising in latent space step by step with the injected feature from the blue part. The blue part is the proposed event-guided conditioning and generation process, which adopts an event-image encoder and a pyramid fusion to utilize events and the LDR image efficiently and effectively, as described in Section 3.2. The orange part is the fine-grained detail enhancement based on the Variant Auto Encoder (VAE) decoder, which learns to refine fine-grained details by the proposed histogram-based structure loss as described in Section 3.3. Training details are described in Section 3.4.", + "bbox": [ + 88, + 467, + 482, + 680 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 688, + 220, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion models Diffusion models [16, 34, 38] can generate realistic high-quality images by diffusing and denoising processes. Latent diffusion models [34] are a type of diffusion models that perform those processes on latent space. During training, the latent diffusion models project the clean image into the latent space with a VAE encoder as $z_0$ , then diffuse $z_0$ by adding Gaussian noise for $n$ steps to get the noisy latent $z_n$ . Adding noise in one step is as:", + "bbox": [ + 89, + 709, + 483, + 832 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n} = \\sqrt {a _ {n}} z _ {n - 1} + \\sqrt {1 - a _ {n}} \\epsilon_ {n - 1}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 842, + 482, + 859 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $n \\in [1, N]$ means the $n$ -th diffusion step, $N$ is the total diffusion steps, $\\epsilon_n \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})$ is the added noise, $a_n$ is", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the pre-defined parameter of the noise scheduler. Defining $\\bar{a}_n = \\prod_{i=1}^n a_i$ , the closed form equation of Eq. (1) [16] is:", + "bbox": [ + 511, + 383, + 906, + 415 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {n} = \\sqrt {\\bar {a} _ {n}} z _ {0} + \\sqrt {1 - \\bar {a} _ {n}} \\epsilon , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 420, + 903, + 436 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The denoising process trains a denoiser $\\epsilon_{\\theta}$ to estimate the noise $\\epsilon$ usually by minimizing Mean Square Error (MSE) $\\mathcal{L}_{n,z_0,\\epsilon} = \\| \\epsilon -\\epsilon_{\\theta}(z_n,n)\\|$ . As the variable $z_{n}$ will be an approximately standard Gaussian distribution when $n$ is large enough, during inference, the denoising process often starts from an i.i.d. noise $z_{N}$ and generates a clean image $z_{0}$ using $\\epsilon_{\\theta}$ over $N$ steps.", + "bbox": [ + 511, + 443, + 906, + 550 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event triggering and stacking Event cameras trigger an event $(t,x,y,p)$ when illuminance changes of pixel $(x,y)$ in the logarithm domain over a predefined threshold $\\eta$ in time $t$ , where the polarity $p\\in \\{-1,1\\}$ indicates the decrease and increase of the illuminance. Mathematically, an event is generated when the following inequation occurs:", + "bbox": [ + 511, + 551, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\| \\log I _ {x, y} (t) - \\log I _ {x, y} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\eta , \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 596, + 650, + 903, + 667 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $I_{x,y}(t)$ is the illuminance at time $t$ , $I_{x,y}(t_{\\mathrm{ref}})$ is the reference illuminance level, $t_\\mathrm{ref}$ is the last event triggered time of pixel $(x,y)$ , $\\eta$ is a pre-defined event threshold. We convert the stream-like events to tensors using voxel grid [52], which encodes temporal information in a C-channel 3D volume by discretizing event timestamps into $C$ temporal bins. Following Rebecq et al. [32, 33], $k$ -th event $(t_k,x_k,y_k,p_k)$ distributes its polarity $p_k$ to the two closest temporal bins related to its normalized timestamp by:", + "bbox": [ + 511, + 672, + 906, + 809 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nE _ {j} ^ {x, y} = \\sum_ {x _ {k} = x, y _ {k} = y} p _ {k} \\max (0, 1 - | \\tilde {t} _ {k} - j |), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 815, + 903, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\tilde{t}_k = \\frac{C - 1}{\\Delta t} (t_k - t_0)$ is the normalized timestamp, $\\Delta_t = \\max_k(t_k) - t_0$ is the time span of events, $t_0$ is the start timestamp, and $j \\in [0, C - 1]$ is the index of the bin.", + "bbox": [ + 511, + 853, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11789", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/26d7d1d70d68d9dedeb95de0f48359d962b474bc969b0c535093b34a0c8c15cc.jpg", + "image_caption": [ + "Figure 3. Pipeline of the event-guided conditioning process, with comparison to conditioning using HDRev [48] restored image. Directly concatenating events and LDR image cannot fully utilize events, thus we introduce an event-image encoder to fuse them as shown in the orange part. The HDRev [48] restored images by decoding HDR features can serve as input to the embedding module, as shown in the yellow part. The proposed pyramid fusion module, as shown in the blue part, fuses different levels HDR features to provide conditions." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 339, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b044f0e8017dfb15c9ff581463b50135748688fc6798c14fa2b0279ffb810a06.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 339, + 88, + 575, + 224 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/8bf7ba3d618c03b8f712d7598c448f854c0f5a1140c77866026f4ca6a04cff98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 575, + 88, + 903, + 224 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Event-guided Conditioning and Generation", + "text_level": 1, + "bbox": [ + 89, + 290, + 460, + 306 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To hallucinate missing information for HDR reconstruction, previous methods mainly come in two ways: adopting better hallucination models (e.g., diffusion models [22]) or introducing HDR sensors (e.g., event cameras [13, 14, 48]). Diffusion models generate high-quality and realistic images by modeling real-world image distributions with diffusion priors. Event cameras provide differential HDR intensity, making events well-suited as conditions for diffusion models to generate more accurate details in over-/under-exposed areas, thereby compensating the dynamic range of LDR images. Thus we propose to leverage both event cameras and diffusion models to hallucinate realistic HDR images to be more consistent with real scenes.", + "bbox": [ + 88, + 311, + 483, + 507 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate images corresponding to conditions, conditional models [18, 50] are proposed to generate features injected into specific layers of the well-trained denoiser. ControlNet [50] is one of the widelyconditional models for image restoration [23, 43, 44], which generates and injects features for each down-sample and middle layers of the denoiser. As illustrated in the interaction between the blue and green parts of Figure 2, we perform feature injection similar to ControlNet. In the green part, we utilize Stable Diffusion [34] V1.5 as the pretrained denoiser $\\epsilon_{\\theta}$ to denoise from the initial latent $z_{N}$ step by step. The generated image is obtained by decoding the estimated initial latent $\\hat{z}_0$ with VAE decoder. With an image-like input, ControlNet employs an embedding module to generate conditions, then adopts a control module to obtain injected features. However, it is challenging for the embedding module to effectively fuse LDR image with events and extract useful information.", + "bbox": [ + 89, + 508, + 483, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Events record temporal intensity changes, making them distinct from LDR images. Directly concatenating stacked events $E$ and LDR images $I_{\\mathrm{LDR}}$ together as the input for the embedding module $\\mathcal{T}_{\\mathrm{con}}$ introduces HDR information, where the conditioning process is:", + "bbox": [ + 89, + 765, + 483, + 842 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {C o n C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {c o n}} \\left(\\left[ I _ {\\text {L D R}}, E \\right]\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 847, + 482, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "[.,] denotes concatenation. But it cannot fully utilize HDR information encoded in events, e.g., decoration and cable", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "in over-exposed areas cannot be accurately reconstructed in Figure 4 (d). Therefore, a specifically designed HDR information extraction model is needed. HDRev [48] performs event-guided HDR reconstruction through representation alignment and feature fusion, which addresses the modality gap between events and LDR images. It performs representation alignment through pre-training to obtain modality-specific encoders as shown in the orange part of Figure 3. The LDR image and events are encoded into modality-specific features, represented by green features (LDR image) and yellow features (events), respectively. Those two features are fused through the fusion module to obtain HDR features. The restoration process can be written into:", + "bbox": [ + 511, + 290, + 906, + 487 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\text {r e s t o r e d}} = \\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 498, + 906, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{B}$ and $\\mathcal{H}$ are the decoder and encoder in HDRev [48]. The restoration process utilizes events to recover more details, as shown by Figure 4 (e), which offers a better reconstruction of the decoration and cable than (d). The intuitive way to better utilize events for conditioning is adopting $I_{\\mathrm{restored}}$ as input for the embedding module $\\mathcal{T}_{\\mathrm{rest}}$ as shown in the yellow part of Figure 3. The conditioning process1 is:", + "bbox": [ + 511, + 523, + 905, + 630 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} _ {\\text {R e s t C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {r e s t}} \\left(\\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\text {L D R}}, E\\right)\\right)\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 640, + 906, + 656 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Diffusion models improve image quality as shown by the improved shape and color of the table in Figure 4 (f) compared to (e). Adopting $I_{\\mathrm{restored}}$ as conditions (\"RestCond\") better utilizes events and provides more details than \"Con-Cond\", e.g., the decoration in Figure 4 (f) is better than (d).", + "bbox": [ + 511, + 667, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Although HDRev is an encoder-decoder architecture, the fusion between LDR image and events is performed only in the encoder. Applying $I_{\\mathrm{restored}}$ as conditions not only suffers from a decoder-encoder redundant computation $\\mathcal{T}_{\\mathrm{rest}}(\\mathcal{B}(\\cdot))$ as shown in the yellow part of Figure 3, but also fails to reconstruct details from events, e.g., the shape of table in the red box of Figure 4, and is difficult to adjust the proper brightness of over-exposed areas, e.g., the cloud in Figure 1 (e). The embedding module $\\mathcal{T}_{\\mathrm{rest}}$ may misunderstand", + "bbox": [ + 511, + 743, + 906, + 878 + ], + "page_idx": 3 + }, + { + "type": "page_footnote", + "text": "1 More details are provided in the supplementary materials.", + "bbox": [ + 529, + 886, + 841, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11790", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fbe39092ac18e63fb7979b084c7aa0bd69decec4bbd5a69fc9070d9d7e63707e.jpg", + "image_caption": [ + "(a) LDR" + ], + "image_footnote": [], + "bbox": [ + 94, + 66, + 204, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/de5214b870f828a3a944b400bfdade9305f5cfb4659b4df235384bf0022ea3dc.jpg", + "image_caption": [ + "(b) Events" + ], + "image_footnote": [], + "bbox": [ + 210, + 66, + 321, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/693b024506a71d8b6f956f9215074ea3e6ba8cec76ec9038e176069b3f5ae773.jpg", + "image_caption": [ + "(c) GroundTruth", + "Figure 4. Comparison with different conditioning ways: directly concatenating events and LDR image (\"ConCond\") and HDRev [48] restored image (\"RestCond\"). \"ConCond\" (d) cannot fully utilize events in over-exposed areas to reconstruct cable. Illustrated by \"Rest-Cond\" (f), the unsatisfactory results of decoration and table in \"Restored\" (e) failed to provide HDR information for conditioning in over-exposed areas. The proposed conditioning module can restore missing information as depicted in (g)." + ], + "image_footnote": [], + "bbox": [ + 328, + 66, + 437, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/1be280b534db4d0a78dc333062b31f958dd19452ec93a405299693334ef80f20.jpg", + "image_caption": [ + "(d) ConCond" + ], + "image_footnote": [], + "bbox": [ + 444, + 66, + 553, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2f741ba38a4d6d50b3f018e32a3fd7564e8bc0beaeb1e54220ff8b30aa1155b4.jpg", + "image_caption": [ + "(e) Restored" + ], + "image_footnote": [], + "bbox": [ + 560, + 66, + 669, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/02580f47034e2d670432cce1c7136d2190f2b1c5e96f5b4a1b47eab3ef544d82.jpg", + "image_caption": [ + "(f) RestCond" + ], + "image_footnote": [], + "bbox": [ + 676, + 66, + 787, + 190 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9e70a1d9dba1fa16735cc8f97e08d05f4953dabd41cb91b72545e91219bf06a5.jpg", + "image_caption": [ + "(g) Proposed" + ], + "image_footnote": [], + "bbox": [ + 794, + 66, + 903, + 190 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "these unsatisfactory areas as shown by the shape of the decoration in the red box of Figure 4 (f).", + "bbox": [ + 88, + 268, + 482, + 299 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate issues of poor brightness adjustment in HDRev, and leverage its advantage in event-image fusion, we adopt HDR features from each layer of its event-image encoder to provide conditions. Directly adopting HDR features as conditions is not only more efficient, because of removing the redundant decoder-encoder process, but also more effective, because of avoiding introducing unsatisfactory reconstruction from the decoder $\\mathcal{B}$ . However, the HDR features are pyramidal, in which different levels of visual information are encoded separately. To fuse different level HDR features, we propose a pyramid fusion module $\\mathcal{P}$ as shown in the blue part of Figure 3, which fuses different level features with down-sample layers and feature concatenation to provide conditions. More accurate details and colors can be recovered with it as shown in Figure 4 (g). The proposed conditioning process $\\mathcal{E}$ consists of an event-image encoder and pyramid fusion module, denoted by:", + "bbox": [ + 91, + 299, + 483, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {E} \\left(I _ {\\mathrm {L D R}}, E\\right) = \\mathcal {P} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 563, + 482, + 580 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Then the predicted noise is $\\epsilon_{\\theta}(z_n,n,\\mathcal{C}(\\mathcal{E}(I_{\\mathrm{LDR}},E))$ , where $\\mathcal{C}$ is the control module derived from ControlNet [50]. With the pretrained denoiser $\\epsilon_{\\theta}$ , we are able to train the control module $\\mathcal{C}$ and conditioning process $\\mathcal{E}$ to hallucinate HDR images by minimizing the noise loss:", + "bbox": [ + 89, + 585, + 483, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {n, z _ {0}, \\epsilon_ {n}} = \\| \\epsilon_ {n} - \\epsilon_ {\\theta} (z _ {n}, n, \\mathcal {C} (\\mathcal {E} (I _ {\\mathrm {L D R}}, E))) \\| ^ {2}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 132, + 667, + 482, + 685 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Fine-grained Detail Refinement", + "text_level": 1, + "bbox": [ + 89, + 690, + 370, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although diffusion models can reconstruct realistic HDR images, it suffers from distortion, especially for fine-grained details as shown in Figure 1 (f) and Figure 5 \"W/o Refinement\". Therefore, we aim to introduce information from events and LDR image to refine those distortion. Typically, we adopt HDR features from the event-image encoder for refinement. As shown in the orange part of Figure 2, our refinement module is a decoder $\\mathcal{D}$ in which the up-sample layers are initialized from the VAE decoder. It fuses HDR features $\\mathcal{H}(I_{\\mathrm{LDR}},E)$ with latent $z_0$ to provide the refined HDR image $H$ :", + "bbox": [ + 89, + 710, + 483, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nH = \\mathcal {D} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right), z _ {0}\\right). \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 883, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By training the refinement module with the Ground Truth (GT), we can refine some distortion as shown in Figure 5 \"W/o Structure\". Due to the inconsistent color between GT and diffusion results as illustrated in Figure 5, applying GT as a supervision target brings color artifacts similar to regression-based methods [24, 48] as depicted in the red box of Figure 5 \"W/o Structure\". Those artifacts come from the discrepancy between the deterministic GT and the uncertainty in diffusion results during the training process of refinement. Additionally, it also depicts that the details in diffusion results have been wiped (the road curb in the red box) after this refinement. Because it is easier for the refinement to minimize the total loss by reducing the brightness and color difference than the detail difference.", + "bbox": [ + 511, + 268, + 906, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Meanwhile, although the diffusion results suffer from distortion, they already have visually pleasant brightness and color consistent with LDR images. Therefore, we tend to maintain the brightness and color properties of diffusion results and only focus on detail refinement by introducing local histogram matching [45] to adjust the GT to serve as a new supervision target. Local histogram matching [45] is able to adjust the pixel intensity distribution to match the histogram of a target image for each local area, which is suitable for adjusting only color and brightness, and maintaining details of GT. Adopting local histogram matching to adjust ground truth image $O$ to diffusion result $H_{\\mathrm{diff}}$ is as:", + "bbox": [ + 511, + 479, + 908, + 662 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nO _ {\\mathrm {a d j}} = \\operatorname {H i s t} (O, H _ {\\text {d i f f}}). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 674, + 903, + 691 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our histogram-based structure loss adopts the adjusted image $O_{\\mathrm{adj}}$ as ground truth and is composed of two losses. The first one is MSE loss:", + "bbox": [ + 511, + 703, + 906, + 748 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M S E}} (H, O _ {\\mathrm {a d j}}) = \\| H - O _ {\\mathrm {a d j}} \\| ^ {2}, \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 601, + 760, + 903, + 779 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the perceptual loss $\\mathcal{L}_{perc}$ based on the feature maps extracted by the pretrained VGG-16 [36] network:", + "bbox": [ + 511, + 790, + 906, + 821 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l} \\| \\phi_ {l} (H) - \\phi_ {l} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2} + \\| \\mathcal {G} _ {l} ^ {\\phi} (H) - \\mathcal {G} _ {l} ^ {\\phi} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 832, + 903, + 876 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "2Diverse diffusion results are shown in the supplementary materials.", + "bbox": [ + 529, + 886, + 890, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/091bd60993080a16a587d3cd8e95327b29cb7070d8899032bf3e24a0d1b0cab6.jpg", + "image_caption": [ + "Figure 5. Validation for the effectiveness of the structure loss. With only conditioning and diffusion process, the results of \"W/o Refinement\" suffer from distortion as shown in the green box. Training the refinement module with ground truth, some distortions are refined, but introducing unnatural saturation adjustment and unsatisfactory detail prediction as shown in the red box of \"W/o Structure\". We introduce local histogram matching [45] to adjust the brightness and color of ground truth to diffusion results, making the refinement module focus on refining the details. The adapted ground truth is shown in \"Adapted GT\". As shown in \"Proposed\", by adopting adapted ground truth as a supervision target, the refinement module with structure loss refines distortion with more natural colors and details." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 204, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cbb780887490dc6c0d0ba6ad31d1b4bd61311ee38169c07ad948e49afecf553e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 207, + 88, + 320, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2c8e0d3deaa7378e30eb5b74d25ae0885c4f85e96648f82b41d861bdda21a1d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 325, + 88, + 437, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/765b5887378ad2204b9be39ccbea1de9ac1c1c9e370b3451465e5eb39d1aa5c2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 442, + 88, + 553, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f957d898bf9dc599bcdd8e9216c0ab098ecce1d032a55024b188210e832dd08a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 88, + 671, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8918109efd6f8d94ac290a21a5d650ab71a3b36a73bf67bdbd5656d45009e362.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 676, + 88, + 787, + 224 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/15dfe72c48d5cc6e9f93f560bb9449fd27012baa0ca7c4dc18b2a4709e8fd4a3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 792, + 88, + 903, + 224 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\phi_l$ is the extracted feature from $l$ -th layer of VGG-16, $\\mathcal{G}_l^\\phi$ is the Gram matrix of $\\phi_l$ . Our structure loss is:", + "bbox": [ + 89, + 332, + 482, + 364 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {s t r u c t}} = \\alpha \\mathcal {L} _ {\\mathrm {M S E}} + \\beta \\mathcal {L} _ {\\text {p e r c}}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 378, + 480, + 395 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\alpha = 0.01, \\beta = 0.001$ are the balance weights of different terms. Using $O_{\\mathrm{adj}}$ as the supervision target enables the refinement module to focus on fine-grained detail refinement as illustrated in Figure 5 \"Complete\", in which the building is undistorted and the road curb is natural.", + "bbox": [ + 89, + 409, + 483, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Training details", + "text_level": 1, + "bbox": [ + 89, + 500, + 250, + 515 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a large number of paired events, LDR image, and GT is needed for training, we adopt 733 HDR images collected by Yang et al. [48] and follow the same data generation process to generate training and testing data with resolution (512, 512). We first simulate 733 HDR videos with random global motion by generating random camera motion trajectories [3]. We utilize the event simulator [31] and virtual camera [6] to simulate the events and LDR images, respectively. The dataset generated from 733 images is separated into training and testing following the setting of Yang et al. [48], while 663 for training and 70 for testing. During training, we randomly over- or under-exposed $20\\%$ to $50\\%$ pixels of the HDR images to generate LDR images.", + "bbox": [ + 88, + 523, + 483, + 720 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Other implementation details. The proposed method is implemented with the Pytorch framework and runs on a single NVIDIA GeForce RTX 3090 GPU. For optimization, we use the ADAM optimizer [19] with a linear decay learning rate scheduler starting from step 0 and an initial learning rate of $\\gamma = 10^{-5}$ . DDIM noise scheduler is adopted both in training and sampling, with training steps $N = 1000$ and linear-scaled beta schedule. We adopt $C = 5$ same to HDRev [48] to apply their pretrain model. During sampling, we adopt classifier-free guidance with guidance scale $u = 1.5$ and inference steps $n = 9$ to obtain our results.", + "bbox": [ + 89, + 734, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/352c6f914f42a1c5db480dbae694a975523d22d1c2006a5186e910a8b7c7c0e0.jpg", + "table_caption": [ + "Table 1. Quantitative evaluation of synthetic data on dataset collected by Yang et al. [48]. The red metric shows the best performance. $\\uparrow (\\downarrow)$ means higher (lower) is better." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
Liu et al. [24]18.350.7710.27615.3378.414.02
EventHDR [53]11.040.3340.44723.44182.114.58
Sagiri [22]12.500.4530.41417.6783.465.35
HDRev [48]14.050.6190.23817.6946.233.88
NeurImg [14]18.530.6210.33817.42105.074.04
Dille et al. [4]19.730.8200.2439.7676.834.22
Ours25.670.9260.0996.0127.093.86
", + "bbox": [ + 514, + 376, + 906, + 500 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 513, + 645, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The comparison with existing methods including four categories: event-based EventHDR [53], single-image CNN-based Liu et al. [24] and Dille et al. [4], single-image diffusion-based Sagiri [22], and event-guided NeurImg-HDR [14] and HDRev [48]. Experiments are conducted both on synthetic data and real data, as described in Section 4.1 and Section 4.2 respectively.", + "bbox": [ + 511, + 539, + 906, + 645 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Metrics For fidelity preserving, we adopt reference metrics including peak signal-to-noise ratio (PSNR), structural similarity (SSIM), the perceptual error with learned perceptual image patch similarity (LPIPS) [51], the Frechet Inception Distance (FID) [15], and the Color Difference Evaluation 2000 (CIEDE) [25], respectively. A non-reference metric named the Natural Image Quality Evaluator (NIQE) [27] is adopted to evaluate the image quality.", + "bbox": [ + 511, + 648, + 908, + 771 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Evaluation on Synthetic Data", + "text_level": 1, + "bbox": [ + 511, + 776, + 774, + 792 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The quantitative evaluation is reported in Table 1. The proposed method outperforms the previous methods in terms of fidelity and structure preservation as shown by reference metrics. Also, ours image quality outperforms other methods, as indicated by NIQE [27]. The qualitative evaluation is illustrated in Figure 6. EventHDR [53] reconstructs faithful grayscale HDR images without color. Liu et al. [24]", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11792", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cce0b5cd0446b2cbf19bee4561fb0576067408b861fd8430c1f37cd9d9bc97bd.jpg", + "image_caption": [ + "Figure 6. Qualitative evaluation of synthetic data on dataset collected by Yang et al. [48]." + ], + "image_footnote": [], + "bbox": [ + 96, + 90, + 903, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a80edf8c56f24607ea4705587b770068af85c464680f34370ce212d7326d3ee5.jpg", + "image_caption": [ + "Figure 7. Qualitative evaluation of real data on DSEC [8]." + ], + "image_footnote": [], + "bbox": [ + 93, + 318, + 903, + 545 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/046e149a28427118415d145e4390d9bec1c5e18e71c9dd5f906953cf0bf80783.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation (NIQE [27]↓/MANIQA [47]↑/QualiClip [1]↑) of real data. Blue metrics indicate the second-best." + ], + "table_footnote": [], + "table_body": "
DatasetLiu et al. [24]EventHDR [53]Sagiri [22]HDRev [48]NeurImg [14]Dille et al.Ours
DSEC [8](3.65/0.466)(4.00/0.270)(5.87/0.617)(3.41/0.402)(3.80/0.570)(3.63/0.467)(3.12/0.693)
HES-HDR [14](5.67/0.335)(5.52/0.203)(6.40/0.576)(5.01/0.322)(4.61/0.442)(5.72/0.367)(4.81/0.530)
", + "bbox": [ + 119, + 587, + 874, + 648 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "maintains details from LDR images but is hard to compensate for missing information. Sagiri [22] adopts diffusion models to provide high-quality images but fails to preserve consistency with inputs. NeurImg [14] has difficulty combining events with LDR image to predict over-exposed areas and has noise-like artifacts. HDRev [48] recovers HDR information, while it is difficult to deal with brightness and color adjustment in over-/under-exposed areas. Our method provides natural results and outperforms other methods both on missing information generation and fidelity preservation.", + "bbox": [ + 88, + 655, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Evaluation on Real Data", + "text_level": 1, + "bbox": [ + 89, + 814, + 316, + 829 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on DSEC dataset DSEC [8] is a stereo event camera dataset for driving scenarios. We choose parts3 of data recording HDR scenes as our testing data.", + "bbox": [ + 89, + 833, + 483, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The quantitative evaluation using non-reference metrics is shown in Table 2, indicating that the proposed method achieves higher image quality than others. The qualitative evaluation of the DSEC dataset is depicted in Figure 7. EventHDR [53] fails to restore grayscale results. It's challenging for Liu et al. [24] and Sagiri [22] to hallucinate missing information. NeurImg [14] and HDRev [48] have problems utilizing the HDR information provided by events. The proposed method not only provides realistic and faithful results compared with other methods but also exhibits generation ability even with fewer events.", + "bbox": [ + 511, + 655, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison on HES-HDR dataset HES-HDR [14] is a hybrid event and spike HDR dataset. Its hybrid event dataset is suitable for our task. Although the non-reference metrics are not the best, we show better recovery ability as shown by qualitative evaluation3. The qualitative evalua", + "bbox": [ + 511, + 824, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_footnote", + "text": "3More details are provided in the supplementary materials.", + "bbox": [ + 107, + 886, + 419, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11793", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a8bb1686632dcfa21009aa04b4ed909e9cb103dd165faea8e19686e631e4ffce.jpg", + "image_caption": [ + "Figure 8. Qualitative evaluation of real data on HES-HDR [14]." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 903, + 314 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/d8b6cac3fe629e9bf18b060588a7a0af793f9e574a78fd0ff9a8bd3a08cef6aa.jpg", + "table_caption": [ + "Table 3. Ablation studies of synthetic data on the dataset collected by Yang et al. [48]. \"Proposed*\" is calculated with adjusted GT $O_{\\mathrm{adj}}$ . The red metric shows the best performance among different ablation parts. $\\uparrow (\\downarrow)$ means higher (lower) is better." + ], + "table_footnote": [], + "table_body": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
ConCond22.600.7490.2007.4848.144.23
RestCond20.790.7250.2479.3361.744.20
W/o Pyramid21.620.7280.2238.6746.993.89
W/o Refinement23.290.770.1707.2938.133.63
W/o structure26.650.9330.0935.6725.903.90
Proposed25.670.9260.0996.0127.093.86
Proposed*29.770.9390.0803.7922.623.86
", + "bbox": [ + 91, + 395, + 483, + 523 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tion of the HES-HDR dataset is depicted in Figure 8, which shows similar conclusions as DSEC [8]. Besides, even with unnatural color-shifted LDR image, our results have a more natural appearance and better fidelity.", + "bbox": [ + 89, + 529, + 483, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 604, + 243, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Our ablation study includes two parts, different conditioning processes, and different refinement targets. The quantitative evaluations of different conditioning processes are shown in the upper part of Table 3. \"ConCond\" and \"Rest-Cond\" are as described in Section 3.2. \"W/o Pyramid\" represents removing the pyramid fusion module and adopting the lowest HDR features as conditions. The proposed one (\"W/o Refinement\") outperforms others. The qualitative evaluations in Figure 4 show that the proposed one can better utilize events and LDR information to provide faithful and realistic images. The quantitative evaluations of refinement and structure loss are depicted in the lower parts of Table 3. The ablation of removing the refinement module is shown by \"W/o Refinement\". Compared with diffusion-only method \"W/o Refinement\", the refinement module improves the fidelity in terms of supervised metrics. Employing GT as a supervision target (\"W/o structure\") makes the reference metrics a little better than \"Proposed\". However,", + "bbox": [ + 89, + 628, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "due to the different supervision targets, directly comparing the proposed method with GT cannot show our structure-preserving capacity. Therefore, we include \"Proposed*\", which is calculated with the adjusted GT $O_{\\mathrm{adj}}$ . The image quality reflected by NIQE of \"Proposed\" is better than \"W/o structure\". As shown by Figure 5, more natural and higher-quality images without distortion can be generated with the proposed refinement module.", + "bbox": [ + 511, + 335, + 906, + 458 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 469, + 633, + 484 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we present a fidelity-preserving HDR reconstruction method based on diffusion models. Our method consists of two parts, event-guided conditioning generation and fine-grained detail refinement. We utilize an event-image encoder and a pyramid fusion module to efficiently and effectively fuse events and images, providing HDR conditions to guide the diffusion process and generate missing information. A refinement module with histogram-based structure loss is proposed to tackle distortion without changing the brightness and contrast. Experiments on both synthetic and real data demonstrate the generation and fidelity preservation ability of our method.", + "bbox": [ + 511, + 488, + 906, + 672 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Since diffusion models have uncertainty in the generation process, it is difficult to apply our method for video generation, for the challenge of maintaining consistency over consecutive frames.", + "bbox": [ + 511, + 688, + 905, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement", + "text_level": 1, + "bbox": [ + 511, + 761, + 671, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resource.", + "bbox": [ + 511, + 779, + 906, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11794", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-aware image-text alignment for opinion-unaware image quality assessment. arXiv preprint arXiv:2403.11176, 2024. 7", + "[2] Hrishav Bakul Barua, Kalin Stefanov, KokSheik Wong, Abhinav Dhall, and Ganesh Krishnasamy. GTA-HDR: A large-scale synthetic dataset for HDR image reconstruction. arXiv preprint arXiv:2403.17837, 2024. 2", + "[3] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE TIP, 2012. 6", + "[4] Sebastian Dille, Chris Careaga, and Yaqiz Aksoy. Intrinsic single-image hdr reconstruction. In ECCV, 2024. 1, 6", + "[5] Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafal K Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 1, 2", + "[6] Yuki Endo, Yoshihiro Kanamori, and Jun Mitani. Deep reverse tone mapping. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 2, 6", + "[7] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In CVPR, 2023. 2", + "[8] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), 2021. 7, 8", + "[9] Michael Gharbi, Jiawen Chen, Jonathan Barron, Samuel Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2017. 2", + "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2", + "[11] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024. 2", + "[12] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2", + "[13] Jin Han, Chu Zhou, Peiqi Duan, Yehui Tang, Chang Xu, Chao Xu, Tiejun Huang, and Boxin Shi. Neuromorphic camera guided high dynamic range imaging. In CVPR, 2020. 2, 4", + "[14] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 4, 6, 7, 8", + "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3", + "[17] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for hdr deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In Proceedings of the European Conference on Computer Vision (ECCV). ECCV, 2024. 4", + "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[20] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In CVPR, 2018. 2", + "[21] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2", + "[22] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arxiv, 2024. 1, 2, 4, 6, 7", + "[23] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-BIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2024. 2, 4", + "[24] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 1, 2, 5, 6, 7", + "[25] M Ronnier Luo, Guihua Cui, and Bryan Rigg. The development of the cie 2000 colour-difference formula: Ciede2000. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Français de la Couleur, 26(5):340-350, 2001. 6", + "[26] Nico Messikommer, Stamatios Georgoulis, Daniel Gehrig, Stepan Tulyakov, Julius Erbach, Alfredo Bochicchio, Yuanyou Li, and Davide Scaramuzza. Multi-bracket high dynamic range imaging with event cameras. In CVPR, 2022. 2", + "[27] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6, 7", + "[28] Mohammad Mostafavi, Lin Wang, and Kuk-Jin Yoon. Learning to reconstruct hdr images from events, with applications to depth and flow prediction. IJCV, 2021. 1, 2", + "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11795", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2", + "[30] Yohan Poirier-Ginter and Jean-François Lalonde. Robust unsupervised stylegan image restoration. In CVPR, 2023. 2", + "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: An open event camera simulator. In Proc. of Conference on Robotics Learning, 2018. 6", + "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 1, 2, 3", + "[33] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 2, 3", + "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4", + "[35] Richard Shaw, Sibi Catley-Chandar, Ales Leonardis, and Eduardo Pérez-Pellitero. HDR reconstruction from bracketed exposures and events. In BMVC, 2022. 2", + "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5", + "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 2", + "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2", + "[40] Chao Wang, Ana Serrano, Xingang Pan, Bin Chen, Karol Myszkowski, Hans-Peter Seidel, Christian Theobalt, and Thomas Leimkuhler. GlowGAN: Unsupervised learning of HDR images from LDR images in the wild. In ICCV, 2023. 2", + "[41] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In CVPR, 2019. 2", + "[42] Zejia Weng, Xitong Yang, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Genrec: Unifying video generation and recognition with diffusion models. arXiv preprint arXiv:2408.15241, 2024. 2", + "[43] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2, 4", + "[44] Wenhan Wu, Xian Liu, and Jie Chen. Image restoration with controlnet. IEEE Transactions on Image Processing, 2023. 4", + "[45] Garima Yadav, Saurabh Maheshwari, and Anjali Agarwal. Contrast limited adaptive histogram equalization based enhancement for real time video system. In 2014 International Conference on Advances in Computing, Communications and Informatics (ICACCI), 2014. 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[46] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2", + "[47] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Mania: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 7", + "[48] Yixin Yang, Jin Han, Jinxiu Liang, Imari Sato, and Boxin Shi. Learning event guided high dynamic range video reconstruction. In CVPR, 2023. 1, 2, 4, 5, 6, 7, 8", + "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2", + "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5", + "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6", + "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 3", + "[53] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, 2021. 1, 2, 6, 7" + ], + "bbox": [ + 516, + 92, + 903, + 488 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11796", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_model.json b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ea2a4ff6a6b34fb7029a0c38e7f88d0ed4199577 --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_model.json @@ -0,0 +1,2585 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.213, + 0.131, + 0.787, + 0.154 + ], + "angle": 0, + "content": "Event-guided HDR Reconstruction with Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.18, + 0.68, + 0.199 + ], + "angle": 0, + "content": "Yixin Yang\\(^{1,2,\\dagger}\\) Jiawei Zhang\\(^{3}\\) Yang Zhang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.24, + 0.199, + 0.756, + 0.217 + ], + "angle": 0, + "content": "Yunxuan Wei\\(^{3}\\) Dongqing Zou\\(^{4}\\) Jimmy S. Ren\\(^{3,5}\\) Boxin Shi\\(^{1,2,*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.117, + 0.217, + 0.882, + 0.235 + ], + "angle": 0, + "content": "1 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.116, + 0.235, + 0.882, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\) Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.221, + 0.252, + 0.776, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\) SenseTime Research \\(^{4}\\)PBVR \\(^{5}\\)Hong Kong Metropolitan University" + }, + { + "type": "text", + "bbox": [ + 0.188, + 0.272, + 0.808, + 0.288 + ], + "angle": 0, + "content": "{yangyixin93, shiboxin}@pku.edu.cn, Github Page: github.com/YixinYang-00/HDRev-Diff" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.485, + 0.69 + ], + "angle": 0, + "content": "Events provide High Dynamic Range (HDR) intensity change which can guide Low Dynamic Range (LDR) image for HDR reconstruction. However, events only provide temporal intensity differences and it is still ill-posed in over-/under-exposed areas due to missing initial reference brightness and color information. With strong generation ability, diffusion models have shown their potential for tackling ill-posed problems. Therefore, we introduce conditional diffusion models to hallucinate missing information. Whereas, directly adopting events and LDR image as conditions is complicated for diffusion models to sufficiently utilize their information. Thus we introduce a pretrained events-image encoder tailored for HDR reconstruction and a pyramid fusion module to provide HDR conditions, which can be efficiently and effectively utilized by the diffusion model. Moreover, the generation results of diffusion models usually exhibit distortion, particularly for fine-grained details. To better preserve fidelity and suppress distortion, we propose a fine-grained detail recovery approach using a histogram-based structural loss. Experiments on real and synthetic data show the effectiveness of the proposed method in terms of both detail preservation and information hallucination." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.222, + 0.734 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.744, + 0.484, + 0.866 + ], + "angle": 0, + "content": "Extending the dynamic range of an image (usually with low dynamic range, LDR) to record the brightness of real scenes plausibly is particularly useful for daily photography and computer vision tasks. High dynamic range (HDR) imaging can be, in the most direct way, achieved by utilizing neural networks to hallucinate missing information in over-/under-exposed areas [4, 5, 22, 24], as shown in Figure 1 (a), which is challenging to faithfully recover" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.309, + 0.64, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.404, + 0.622, + 0.417 + ], + "angle": 0, + "content": "(a) Dille et. al." + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.309, + 0.765, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.671, + 0.404, + 0.736, + 0.417 + ], + "angle": 0, + "content": "(b) Events" + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.309, + 0.905, + 0.403 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.798, + 0.404, + 0.885, + 0.417 + ], + "angle": 0, + "content": "(c) EventHDR" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.421, + 0.64, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.533, + 0.521, + 0.625, + 0.535 + ], + "angle": 0, + "content": "(d) LDR image" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.421, + 0.779, + 0.542 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.421, + 0.905, + 0.52 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.809, + 0.521, + 0.877, + 0.535 + ], + "angle": 0, + "content": "(e) HDRev" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.542, + 0.639, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.552, + 0.638, + 0.607, + 0.652 + ], + "angle": 0, + "content": "(f) Sagiri" + }, + { + "type": "image", + "bbox": [ + 0.642, + 0.542, + 0.765, + 0.636 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.677, + 0.638, + 0.729, + 0.652 + ], + "angle": 0, + "content": "(g) Ours" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.553, + 0.842, + 0.565 + ], + "angle": 0, + "content": "\\(\\Delta\\) : Events" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.566, + 0.89, + 0.577 + ], + "angle": 0, + "content": "\\(\\Delta\\) :Colorful results" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.578, + 0.882, + 0.59 + ], + "angle": 0, + "content": "\\(\\Delta\\) : More realistic" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.591, + 0.888, + 0.603 + ], + "angle": 0, + "content": "\\(\\Delta\\) : Less distortion" + }, + { + "type": "text", + "bbox": [ + 0.778, + 0.604, + 0.877, + 0.615 + ], + "angle": 0, + "content": ": CNN models" + }, + { + "type": "text", + "bbox": [ + 0.777, + 0.616, + 0.901, + 0.628 + ], + "angle": 0, + "content": "Diffusion models" + }, + { + "type": "list", + "bbox": [ + 0.777, + 0.553, + 0.901, + 0.628 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.658, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Figure 1. HDR image can be recovered by only using a single LDR image (d) as input (e.g., Dille et al. [4]), which is a difficult problem as shown in (a). With only events (b) as input, event-to-image reconstruction (e.g., EventHDR [53]) (c) recovers an HDR scene without color. Event-guided HDR methods (e.g., HDRev [48]) (e) take both events and the LDR image as input, which recovers colorful HDR scenes with moderate details. Diffusion-based methods (e.g., Sagiri [22]) (f) take the LDR image (and optionally the text prompt) as input showing visually pleasing results but may have challenges in maintaining consistent contents with the original scene. The proposed method (g) takes events and the LDR image as input and leverages the diffusion priors to recover a plausible and realistic HDR image with rich details and less distortion. scene details without any auxiliary information. Neuromorphic cameras, such as event cameras, are proven to be helpful in providing additional details for HDR purpose [28, 32, 53]. Event cameras capture intensity changes" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.875, + 0.424, + 0.888 + ], + "angle": 0, + "content": "This work is done during Yixin's internship at SenseTime." + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "* Corresponding author." + }, + { + "type": "list", + "bbox": [ + 0.108, + 0.875, + 0.424, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11787" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.198 + ], + "angle": 0, + "content": "by triggering events asynchronously, which inherently contains HDR information (e.g., 120 dB for DAVIS340), and events can be directly converted to HDR images, as demonstrated by EventHDR [53] in Figure 1 (c). The event-to-image reconstruction [28, 32, 53] can recover details in over-/under-exposed regions, but fail to provide colorful results due to lacking color information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.412 + ], + "angle": 0, + "content": "To generate colorful HDR images, event-guided HDR reconstructions [13, 14, 48] take both LDR image and events as inputs, complementarily learning to predict color and intensity in over-/under-exposed areas, as demonstrated in Figure 1 (e). They can reconstruct HDR information, e.g., the cloud in the green box, and provide colorful results. Although regression models are adopted to build the relationship between the predicted pixel values and its neighboring pixels, those methods cannot recover proper intensity, especially when the LDR image has large over-/under-exposed areas. Because absolute HDR intensity (events only contain thresholded radiance changes) and color information are still missing in these regions, it is difficult for regression models to appropriately hallucinate their values." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.612 + ], + "angle": 0, + "content": "Diffusion models [16, 37-39] are emerging generative models that show great potential for generating plausible and high-quality images. Even if for ill-posed problems, they can provide visually pleasing results thanks to the expressive power of diffusion priors. Diffusion models have also been adopted for image restoration [23, 43, 49]. Recent attempts, e.g., Sagiri [22], have shown the potential of diffusion priors for HDR reconstruction by providing more visually realistic details in over-exposed areas, e.g., the cloud in the sky of Figure 1 (f). However, the generated HDR images still cannot appropriately keep the fidelity of the original scene, and additional distortion are brought into well-exposed regions, as shown by the red box in Figure 1 (f)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To reconstruct visually pleasing and faithful HDR images with less distortion, we introduce events to provide differential HDR intensity and adopt conditional diffusion models [50] to compensate for the ill-posed nature of hallucinating missing information and constrain the generative model to be consistent with the original scene. Specially, we focus on solving the following two key problems via employing diffusion models in event-guided HDR reconstruction: 1) Condition extraction: Since events and LDR image are two different modalities of visual representation, directly adopting them as conditions is non-compatible for the conditional diffusion models to conduct effective fusion. Inspired by the modality alignment of HDRev [48], we introduce an event-image encoder and a pyramid fusion module to fuse events and LDR image in different resolutions and provide conditions to guide the generation process. 2) Details preservation: Applying diffusion models to HDR reconstruction [22, 23] suffers from fine-grained detail distortion, as shown in Figure 1 (f). Such artifacts become" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.212 + ], + "angle": 0, + "content": "more obvious when LDR images are high quality and have large well-exposed areas. To refine the fine-grained details and make them consistent with LDR images and events, we design a refinement module with histogram-based structure loss. The proposed method is validated on both synthetic and real data and outperforms existing methods by providing fidelity-preserving HDR reconstruction, as illustrated in Figure 1 (g), with the following technical contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.214, + 0.905, + 0.242 + ], + "angle": 0, + "content": "- integrating events and the conditional diffusion models to recover missing information faithfully;" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.244, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- adopting the pretrained event-image encoder and pyramid fusion to effectively apply conditions; and" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.274, + 0.905, + 0.303 + ], + "angle": 0, + "content": "- designing the refinement module and histogram-based structure loss to further strengthen fine-grained details." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.214, + 0.905, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.32, + 0.661, + 0.335 + ], + "angle": 0, + "content": "2. Related works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.905, + 0.505 + ], + "angle": 0, + "content": "Image-based HDR reconstruction Many approaches leverage the capabilities of neural networks to directly hallucinate HDR information from training data, such as CNNs [5, 6, 9], Generative Adversarial Networks (GANs) [40], and diffusion models [7, 22]. Liu et al. [24] incorporates the LDR image formulation pipeline into its hallucination process to perform step-by-step prediction. For a more comprehensive view, we refer to GTA-HDR [2]. Since information is unavoidably missing in over-/underexposed areas, the above methods only rely on hallucination, which may not correspond to real scenes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.509, + 0.905, + 0.75 + ], + "angle": 0, + "content": "Event-based HDR reconstruction Benefiting from the HDR property of event cameras, previous works reconstruct HDR images from events with recurrent neural networks [32, 33, 53] or GANs [28, 41]. Since events only record logarithmic irradiance changes, it's difficult to reconstruct color images. To reconstruct color HDR images, Han et al. [13, 14] adopted a hybrid camera system to capture events and frame, enabling event-guided HDR reconstruction. Yang et al. [48] treated events and frame as different modalities and proposed multi-modalities alignment and fusion to provide HDR videos. Other methods [26, 35] incorporate events and bracketed exposure for HDR reconstruction, while long-exposed images bring blurry artifacts. The proposed method takes events and LDR image as input, incorporating diffusion models to compensate for realistic and faithful information in over-/under-exposed areas." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.751, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Generative priors Generative Priors capture the underlying data distribution, enabling the generated images to follow real-world image distribution. GAN [10] has shown its potential to leverage generative priors for image restoration [20, 21, 30]. Recently, diffusion models [16, 37-39] have widely adopted as effective generative priors for producing realistic images [29, 34, 42, 50], which also demonstrate strong potential for image restorations [23, 43, 49], specifically, for HDR image reconstruction from bracketed exposure [12, 17, 46]. There are some attempts [11, 22] to" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "11788" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.28, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 2. Pipeline of the proposed method consists of three parts: diffusion model and denoising process (green part), event-image conditioning (blue part), and detail refinement (orange part). The blue part fuses LDR image and events to generate injected features for Denoiser as described in Section 3.2. The green part adopts the injected features as conditions to estimate noise from latent \\( z_{t} \\) step by step with fixed pretrained parameters. To refine the distortion of denoised images, the refinement module in the orange part, whose architecture is shown on the right, is proposed to leverage the HDR features and denoised latent \\( z_{0} \\). Considering the uncertainty of diffusion results in color and brightness, we introduce structure loss, which adopts adaptive Ground Truth (GT) as the supervision target, as described in Section 3.3. After refinement, we obtain a plausible and realistic refined image without distortion." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.482, + 0.43 + ], + "angle": 0, + "content": "take advantage of diffusion models to generate HDR images from single LDR image, while it cannot provide faithful reconstruction due to the inevitable dynamic clipping." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.443, + 0.265, + 0.459 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.468, + 0.483, + 0.681 + ], + "angle": 0, + "content": "Section 3.1 introduces basic concepts about diffusion models and event cameras. The pipeline of the proposed method is illustrated in Figure 2. The green part is the pretrained latent diffusion model, which performs denoising in latent space step by step with the injected feature from the blue part. The blue part is the proposed event-guided conditioning and generation process, which adopts an event-image encoder and a pyramid fusion to utilize events and the LDR image efficiently and effectively, as described in Section 3.2. The orange part is the fine-grained detail enhancement based on the Variant Auto Encoder (VAE) decoder, which learns to refine fine-grained details by the proposed histogram-based structure loss as described in Section 3.3. Training details are described in Section 3.4." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.689, + 0.221, + 0.705 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.71, + 0.484, + 0.833 + ], + "angle": 0, + "content": "Diffusion models Diffusion models [16, 34, 38] can generate realistic high-quality images by diffusing and denoising processes. Latent diffusion models [34] are a type of diffusion models that perform those processes on latent space. During training, the latent diffusion models project the clean image into the latent space with a VAE encoder as \\( z_0 \\), then diffuse \\( z_0 \\) by adding Gaussian noise for \\( n \\) steps to get the noisy latent \\( z_n \\). Adding noise in one step is as:" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.843, + 0.483, + 0.861 + ], + "angle": 0, + "content": "\\[\nz _ {n} = \\sqrt {a _ {n}} z _ {n - 1} + \\sqrt {1 - a _ {n}} \\epsilon_ {n - 1}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(n \\in [1, N]\\) means the \\(n\\)-th diffusion step, \\(N\\) is the total diffusion steps, \\(\\epsilon_n \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})\\) is the added noise, \\(a_n\\) is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.416 + ], + "angle": 0, + "content": "the pre-defined parameter of the noise scheduler. Defining \\(\\bar{a}_n = \\prod_{i=1}^n a_i\\), the closed form equation of Eq. (1) [16] is:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.421, + 0.905, + 0.438 + ], + "angle": 0, + "content": "\\[\nz _ {n} = \\sqrt {\\bar {a} _ {n}} z _ {0} + \\sqrt {1 - \\bar {a} _ {n}} \\epsilon , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.551 + ], + "angle": 0, + "content": "where \\(\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The denoising process trains a denoiser \\(\\epsilon_{\\theta}\\) to estimate the noise \\(\\epsilon\\) usually by minimizing Mean Square Error (MSE) \\(\\mathcal{L}_{n,z_0,\\epsilon} = \\| \\epsilon -\\epsilon_{\\theta}(z_n,n)\\|\\). As the variable \\(z_{n}\\) will be an approximately standard Gaussian distribution when \\(n\\) is large enough, during inference, the denoising process often starts from an i.i.d. noise \\(z_{N}\\) and generates a clean image \\(z_{0}\\) using \\(\\epsilon_{\\theta}\\) over \\(N\\) steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.644 + ], + "angle": 0, + "content": "Event triggering and stacking Event cameras trigger an event \\((t,x,y,p)\\) when illuminance changes of pixel \\((x,y)\\) in the logarithm domain over a predefined threshold \\(\\eta\\) in time \\(t\\), where the polarity \\(p\\in \\{-1,1\\}\\) indicates the decrease and increase of the illuminance. Mathematically, an event is generated when the following inequation occurs:" + }, + { + "type": "equation", + "bbox": [ + 0.597, + 0.651, + 0.905, + 0.668 + ], + "angle": 0, + "content": "\\[\n\\left\\| \\log I _ {x, y} (t) - \\log I _ {x, y} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\eta , \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.673, + 0.907, + 0.81 + ], + "angle": 0, + "content": "where \\( I_{x,y}(t) \\) is the illuminance at time \\( t \\), \\( I_{x,y}(t_{\\mathrm{ref}}) \\) is the reference illuminance level, \\( t_\\mathrm{ref} \\) is the last event triggered time of pixel \\( (x,y) \\), \\( \\eta \\) is a pre-defined event threshold. We convert the stream-like events to tensors using voxel grid [52], which encodes temporal information in a C-channel 3D volume by discretizing event timestamps into \\( C \\) temporal bins. Following Rebecq et al. [32, 33], \\( k \\)-th event \\( (t_k,x_k,y_k,p_k) \\) distributes its polarity \\( p_k \\) to the two closest temporal bins related to its normalized timestamp by:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.816, + 0.905, + 0.848 + ], + "angle": 0, + "content": "\\[\nE _ {j} ^ {x, y} = \\sum_ {x _ {k} = x, y _ {k} = y} p _ {k} \\max (0, 1 - | \\tilde {t} _ {k} - j |), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.854, + 0.906, + 0.902 + ], + "angle": 0, + "content": "where \\(\\tilde{t}_k = \\frac{C - 1}{\\Delta t} (t_k - t_0)\\) is the normalized timestamp, \\(\\Delta_t = \\max_k(t_k) - t_0\\) is the time span of events, \\(t_0\\) is the start timestamp, and \\(j \\in [0, C - 1]\\) is the index of the bin." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11789" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.34, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.34, + 0.089, + 0.576, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.576, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.228, + 0.908, + 0.286 + ], + "angle": 0, + "content": "Figure 3. Pipeline of the event-guided conditioning process, with comparison to conditioning using HDRev [48] restored image. Directly concatenating events and LDR image cannot fully utilize events, thus we introduce an event-image encoder to fuse them as shown in the orange part. The HDRev [48] restored images by decoding HDR features can serve as input to the embedding module, as shown in the yellow part. The proposed pyramid fusion module, as shown in the blue part, fuses different levels HDR features to provide conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.291, + 0.461, + 0.307 + ], + "angle": 0, + "content": "3.2. Event-guided Conditioning and Generation" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.313, + 0.484, + 0.508 + ], + "angle": 0, + "content": "To hallucinate missing information for HDR reconstruction, previous methods mainly come in two ways: adopting better hallucination models (e.g., diffusion models [22]) or introducing HDR sensors (e.g., event cameras [13, 14, 48]). Diffusion models generate high-quality and realistic images by modeling real-world image distributions with diffusion priors. Event cameras provide differential HDR intensity, making events well-suited as conditions for diffusion models to generate more accurate details in over-/under-exposed areas, thereby compensating the dynamic range of LDR images. Thus we propose to leverage both event cameras and diffusion models to hallucinate realistic HDR images to be more consistent with real scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.509, + 0.484, + 0.765 + ], + "angle": 0, + "content": "To generate images corresponding to conditions, conditional models [18, 50] are proposed to generate features injected into specific layers of the well-trained denoiser. ControlNet [50] is one of the widelyconditional models for image restoration [23, 43, 44], which generates and injects features for each down-sample and middle layers of the denoiser. As illustrated in the interaction between the blue and green parts of Figure 2, we perform feature injection similar to ControlNet. In the green part, we utilize Stable Diffusion [34] V1.5 as the pretrained denoiser \\(\\epsilon_{\\theta}\\) to denoise from the initial latent \\(z_{N}\\) step by step. The generated image is obtained by decoding the estimated initial latent \\(\\hat{z}_0\\) with VAE decoder. With an image-like input, ControlNet employs an embedding module to generate conditions, then adopts a control module to obtain injected features. However, it is challenging for the embedding module to effectively fuse LDR image with events and extract useful information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.843 + ], + "angle": 0, + "content": "Events record temporal intensity changes, making them distinct from LDR images. Directly concatenating stacked events \\(E\\) and LDR images \\(I_{\\mathrm{LDR}}\\) together as the input for the embedding module \\(\\mathcal{T}_{\\mathrm{con}}\\) introduces HDR information, where the conditioning process is:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.848, + 0.483, + 0.866 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {C o n C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {c o n}} \\left(\\left[ I _ {\\text {L D R}}, E \\right]\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.87, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[.,] denotes concatenation. But it cannot fully utilize HDR information encoded in events, e.g., decoration and cable" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.488 + ], + "angle": 0, + "content": "in over-exposed areas cannot be accurately reconstructed in Figure 4 (d). Therefore, a specifically designed HDR information extraction model is needed. HDRev [48] performs event-guided HDR reconstruction through representation alignment and feature fusion, which addresses the modality gap between events and LDR images. It performs representation alignment through pre-training to obtain modality-specific encoders as shown in the orange part of Figure 3. The LDR image and events are encoded into modality-specific features, represented by green features (LDR image) and yellow features (events), respectively. Those two features are fused through the fusion module to obtain HDR features. The restoration process can be written into:" + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.499, + 0.907, + 0.515 + ], + "angle": 0, + "content": "\\[\nI _ {\\text {r e s t o r e d}} = \\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.524, + 0.906, + 0.631 + ], + "angle": 0, + "content": "where \\(\\mathcal{B}\\) and \\(\\mathcal{H}\\) are the decoder and encoder in HDRev [48]. The restoration process utilizes events to recover more details, as shown by Figure 4 (e), which offers a better reconstruction of the decoration and cable than (d). The intuitive way to better utilize events for conditioning is adopting \\(I_{\\mathrm{restored}}\\) as input for the embedding module \\(\\mathcal{T}_{\\mathrm{rest}}\\) as shown in the yellow part of Figure 3. The conditioning process1 is:" + }, + { + "type": "equation", + "bbox": [ + 0.564, + 0.641, + 0.907, + 0.657 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} _ {\\text {R e s t C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {r e s t}} \\left(\\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\text {L D R}}, E\\right)\\right)\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.906, + 0.743 + ], + "angle": 0, + "content": "Diffusion models improve image quality as shown by the improved shape and color of the table in Figure 4 (f) compared to (e). Adopting \\( I_{\\mathrm{restored}} \\) as conditions (\"RestCond\") better utilizes events and provides more details than \"Con-Cond\", e.g., the decoration in Figure 4 (f) is better than (d)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.744, + 0.907, + 0.879 + ], + "angle": 0, + "content": "Although HDRev is an encoder-decoder architecture, the fusion between LDR image and events is performed only in the encoder. Applying \\( I_{\\mathrm{restored}} \\) as conditions not only suffers from a decoder-encoder redundant computation \\( \\mathcal{T}_{\\mathrm{rest}}(\\mathcal{B}(\\cdot)) \\) as shown in the yellow part of Figure 3, but also fails to reconstruct details from events, e.g., the shape of table in the red box of Figure 4, and is difficult to adjust the proper brightness of over-exposed areas, e.g., the cloud in Figure 1 (e). The embedding module \\( \\mathcal{T}_{\\mathrm{rest}} \\) may misunderstand" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.843, + 0.901 + ], + "angle": 0, + "content": "1 More details are provided in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11790" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.067, + 0.205, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.126, + 0.192, + 0.174, + 0.204 + ], + "angle": 0, + "content": "(a) LDR" + }, + { + "type": "image", + "bbox": [ + 0.211, + 0.068, + 0.322, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.193, + 0.298, + 0.204 + ], + "angle": 0, + "content": "(b) Events" + }, + { + "type": "image", + "bbox": [ + 0.329, + 0.068, + 0.439, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.334, + 0.193, + 0.428, + 0.204 + ], + "angle": 0, + "content": "(c) GroundTruth" + }, + { + "type": "image", + "bbox": [ + 0.445, + 0.068, + 0.555, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.193, + 0.538, + 0.204 + ], + "angle": 0, + "content": "(d) ConCond" + }, + { + "type": "image", + "bbox": [ + 0.561, + 0.068, + 0.671, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.579, + 0.193, + 0.653, + 0.204 + ], + "angle": 0, + "content": "(e) Restored" + }, + { + "type": "image", + "bbox": [ + 0.678, + 0.068, + 0.788, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.695, + 0.193, + 0.771, + 0.204 + ], + "angle": 0, + "content": "(f) RestCond" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.068, + 0.904, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.811, + 0.193, + 0.888, + 0.204 + ], + "angle": 0, + "content": "(g) Proposed" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.206, + 0.907, + 0.263 + ], + "angle": 0, + "content": "Figure 4. Comparison with different conditioning ways: directly concatenating events and LDR image (\"ConCond\") and HDRev [48] restored image (\"RestCond\"). \"ConCond\" (d) cannot fully utilize events in over-exposed areas to reconstruct cable. Illustrated by \"Rest-Cond\" (f), the unsatisfactory results of decoration and table in \"Restored\" (e) failed to provide HDR information for conditioning in over-exposed areas. The proposed conditioning module can restore missing information as depicted in (g)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.27, + 0.483, + 0.3 + ], + "angle": 0, + "content": "these unsatisfactory areas as shown by the shape of the decoration in the red box of Figure 4 (f)." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.3, + 0.484, + 0.557 + ], + "angle": 0, + "content": "To mitigate issues of poor brightness adjustment in HDRev, and leverage its advantage in event-image fusion, we adopt HDR features from each layer of its event-image encoder to provide conditions. Directly adopting HDR features as conditions is not only more efficient, because of removing the redundant decoder-encoder process, but also more effective, because of avoiding introducing unsatisfactory reconstruction from the decoder \\(\\mathcal{B}\\). However, the HDR features are pyramidal, in which different levels of visual information are encoded separately. To fuse different level HDR features, we propose a pyramid fusion module \\(\\mathcal{P}\\) as shown in the blue part of Figure 3, which fuses different level features with down-sample layers and feature concatenation to provide conditions. More accurate details and colors can be recovered with it as shown in Figure 4 (g). The proposed conditioning process \\(\\mathcal{E}\\) consists of an event-image encoder and pyramid fusion module, denoted by:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.564, + 0.483, + 0.581 + ], + "angle": 0, + "content": "\\[\n\\mathcal {E} \\left(I _ {\\mathrm {L D R}}, E\\right) = \\mathcal {P} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.586, + 0.484, + 0.663 + ], + "angle": 0, + "content": "Then the predicted noise is \\(\\epsilon_{\\theta}(z_n,n,\\mathcal{C}(\\mathcal{E}(I_{\\mathrm{LDR}},E))\\), where \\(\\mathcal{C}\\) is the control module derived from ControlNet [50]. With the pretrained denoiser \\(\\epsilon_{\\theta}\\), we are able to train the control module \\(\\mathcal{C}\\) and conditioning process \\(\\mathcal{E}\\) to hallucinate HDR images by minimizing the noise loss:" + }, + { + "type": "equation", + "bbox": [ + 0.133, + 0.668, + 0.483, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {n, z _ {0}, \\epsilon_ {n}} = \\| \\epsilon_ {n} - \\epsilon_ {\\theta} (z _ {n}, n, \\mathcal {C} (\\mathcal {E} (I _ {\\mathrm {L D R}}, E))) \\| ^ {2}. \\tag {9}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.691, + 0.371, + 0.707 + ], + "angle": 0, + "content": "3.3. Fine-grained Detail Refinement" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.712, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Although diffusion models can reconstruct realistic HDR images, it suffers from distortion, especially for fine-grained details as shown in Figure 1 (f) and Figure 5 \"W/o Refinement\". Therefore, we aim to introduce information from events and LDR image to refine those distortion. Typically, we adopt HDR features from the event-image encoder for refinement. As shown in the orange part of Figure 2, our refinement module is a decoder \\(\\mathcal{D}\\) in which the up-sample layers are initialized from the VAE decoder. It fuses HDR features \\(\\mathcal{H}(I_{\\mathrm{LDR}},E)\\) with latent \\(z_0\\) to provide the refined HDR image \\(H\\):" + }, + { + "type": "equation", + "bbox": [ + 0.201, + 0.885, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nH = \\mathcal {D} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right), z _ {0}\\right). \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.27, + 0.907, + 0.481 + ], + "angle": 0, + "content": "By training the refinement module with the Ground Truth (GT), we can refine some distortion as shown in Figure 5 \"W/o Structure\". Due to the inconsistent color between GT and diffusion results as illustrated in Figure 5, applying GT as a supervision target brings color artifacts similar to regression-based methods [24, 48] as depicted in the red box of Figure 5 \"W/o Structure\". Those artifacts come from the discrepancy between the deterministic GT and the uncertainty in diffusion results during the training process of refinement. Additionally, it also depicts that the details in diffusion results have been wiped (the road curb in the red box) after this refinement. Because it is easier for the refinement to minimize the total loss by reducing the brightness and color difference than the detail difference." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.909, + 0.664 + ], + "angle": 0, + "content": "Meanwhile, although the diffusion results suffer from distortion, they already have visually pleasant brightness and color consistent with LDR images. Therefore, we tend to maintain the brightness and color properties of diffusion results and only focus on detail refinement by introducing local histogram matching [45] to adjust the GT to serve as a new supervision target. Local histogram matching [45] is able to adjust the pixel intensity distribution to match the histogram of a target image for each local area, which is suitable for adjusting only color and brightness, and maintaining details of GT. Adopting local histogram matching to adjust ground truth image \\(O\\) to diffusion result \\(H_{\\mathrm{diff}}\\) is as:" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.675, + 0.905, + 0.693 + ], + "angle": 0, + "content": "\\[\nO _ {\\mathrm {a d j}} = \\operatorname {H i s t} (O, H _ {\\text {d i f f}}). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.704, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Our histogram-based structure loss adopts the adjusted image \\(O_{\\mathrm{adj}}\\) as ground truth and is composed of two losses. The first one is MSE loss:" + }, + { + "type": "equation", + "bbox": [ + 0.602, + 0.761, + 0.905, + 0.78 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M S E}} (H, O _ {\\mathrm {a d j}}) = \\| H - O _ {\\mathrm {a d j}} \\| ^ {2}, \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.791, + 0.907, + 0.822 + ], + "angle": 0, + "content": "and the perceptual loss \\(\\mathcal{L}_{perc}\\) based on the feature maps extracted by the pretrained VGG-16 [36] network:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.833, + 0.905, + 0.877 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l} \\| \\phi_ {l} (H) - \\phi_ {l} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2} + \\| \\mathcal {G} _ {l} ^ {\\phi} (H) - \\mathcal {G} _ {l} ^ {\\phi} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "page_footnote", + "bbox": [ + 0.531, + 0.887, + 0.892, + 0.901 + ], + "angle": 0, + "content": "2Diverse diffusion results are shown in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11791" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.205, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.208, + 0.089, + 0.321, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.326, + 0.089, + 0.438, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.444, + 0.089, + 0.554, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.089, + 0.672, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.677, + 0.089, + 0.788, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.794, + 0.089, + 0.905, + 0.226 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.235, + 0.907, + 0.32 + ], + "angle": 0, + "content": "Figure 5. Validation for the effectiveness of the structure loss. With only conditioning and diffusion process, the results of \"W/o Refinement\" suffer from distortion as shown in the green box. Training the refinement module with ground truth, some distortions are refined, but introducing unnatural saturation adjustment and unsatisfactory detail prediction as shown in the red box of \"W/o Structure\". We introduce local histogram matching [45] to adjust the brightness and color of ground truth to diffusion results, making the refinement module focus on refining the details. The adapted ground truth is shown in \"Adapted GT\". As shown in \"Proposed\", by adopting adapted ground truth as a supervision target, the refinement module with structure loss refines distortion with more natural colors and details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.483, + 0.365 + ], + "angle": 0, + "content": "where \\(\\phi_l\\) is the extracted feature from \\(l\\)-th layer of VGG-16, \\(\\mathcal{G}_l^\\phi\\) is the Gram matrix of \\(\\phi_l\\). Our structure loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.38, + 0.482, + 0.396 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {s t r u c t}} = \\alpha \\mathcal {L} _ {\\mathrm {M S E}} + \\beta \\mathcal {L} _ {\\text {p e r c}}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.487 + ], + "angle": 0, + "content": "where \\(\\alpha = 0.01, \\beta = 0.001\\) are the balance weights of different terms. Using \\(O_{\\mathrm{adj}}\\) as the supervision target enables the refinement module to focus on fine-grained detail refinement as illustrated in Figure 5 \"Complete\", in which the building is undistorted and the road curb is natural." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.5, + 0.25, + 0.516 + ], + "angle": 0, + "content": "3.4. Training details" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.484, + 0.721 + ], + "angle": 0, + "content": "As a large number of paired events, LDR image, and GT is needed for training, we adopt 733 HDR images collected by Yang et al. [48] and follow the same data generation process to generate training and testing data with resolution (512, 512). We first simulate 733 HDR videos with random global motion by generating random camera motion trajectories [3]. We utilize the event simulator [31] and virtual camera [6] to simulate the events and LDR images, respectively. The dataset generated from 733 images is separated into training and testing following the setting of Yang et al. [48], while 663 for training and 70 for testing. During training, we randomly over- or under-exposed \\(20\\%\\) to \\(50\\%\\) pixels of the HDR images to generate LDR images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Other implementation details. The proposed method is implemented with the Pytorch framework and runs on a single NVIDIA GeForce RTX 3090 GPU. For optimization, we use the ADAM optimizer [19] with a linear decay learning rate scheduler starting from step 0 and an initial learning rate of \\(\\gamma = 10^{-5}\\). DDIM noise scheduler is adopted both in training and sampling, with training steps \\(N = 1000\\) and linear-scaled beta schedule. We adopt \\(C = 5\\) same to HDRev [48] to apply their pretrain model. During sampling, we adopt classifier-free guidance with guidance scale \\(u = 1.5\\) and inference steps \\(n = 9\\) to obtain our results." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.331, + 0.907, + 0.373 + ], + "angle": 0, + "content": "Table 1. Quantitative evaluation of synthetic data on dataset collected by Yang et al. [48]. The red metric shows the best performance. \\(\\uparrow (\\downarrow)\\) means higher (lower) is better." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.377, + 0.907, + 0.501 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
Liu et al. [24]18.350.7710.27615.3378.414.02
EventHDR [53]11.040.3340.44723.44182.114.58
Sagiri [22]12.500.4530.41417.6783.465.35
HDRev [48]14.050.6190.23817.6946.233.88
NeurImg [14]18.530.6210.33817.42105.074.04
Dille et al. [4]19.730.8200.2439.7676.834.22
Ours25.670.9260.0996.0127.093.86
" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.514, + 0.646, + 0.531 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.646 + ], + "angle": 0, + "content": "The comparison with existing methods including four categories: event-based EventHDR [53], single-image CNN-based Liu et al. [24] and Dille et al. [4], single-image diffusion-based Sagiri [22], and event-guided NeurImg-HDR [14] and HDRev [48]. Experiments are conducted both on synthetic data and real data, as described in Section 4.1 and Section 4.2 respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.909, + 0.772 + ], + "angle": 0, + "content": "Metrics For fidelity preserving, we adopt reference metrics including peak signal-to-noise ratio (PSNR), structural similarity (SSIM), the perceptual error with learned perceptual image patch similarity (LPIPS) [51], the Frechet Inception Distance (FID) [15], and the Color Difference Evaluation 2000 (CIEDE) [25], respectively. A non-reference metric named the Natural Image Quality Evaluator (NIQE) [27] is adopted to evaluate the image quality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.777, + 0.776, + 0.793 + ], + "angle": 0, + "content": "4.1. Evaluation on Synthetic Data" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "The quantitative evaluation is reported in Table 1. The proposed method outperforms the previous methods in terms of fidelity and structure preservation as shown by reference metrics. Also, ours image quality outperforms other methods, as indicated by NIQE [27]. The qualitative evaluation is illustrated in Figure 6. EventHDR [53] reconstructs faithful grayscale HDR images without color. Liu et al. [24]" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11792" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.092, + 0.904, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.234, + 0.298, + 0.762, + 0.312 + ], + "angle": 0, + "content": "Figure 6. Qualitative evaluation of synthetic data on dataset collected by Yang et al. [48]." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.319, + 0.905, + 0.546 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.325, + 0.547, + 0.671, + 0.56 + ], + "angle": 0, + "content": "Figure 7. Qualitative evaluation of real data on DSEC [8]." + }, + { + "type": "table_caption", + "bbox": [ + 0.115, + 0.568, + 0.882, + 0.583 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation (NIQE [27]↓/MANIQA [47]↑/QualiClip [1]↑) of real data. Blue metrics indicate the second-best." + }, + { + "type": "table", + "bbox": [ + 0.12, + 0.588, + 0.875, + 0.649 + ], + "angle": 0, + "content": "
DatasetLiu et al. [24]EventHDR [53]Sagiri [22]HDRev [48]NeurImg [14]Dille et al.Ours
DSEC [8](3.65/0.466)(4.00/0.270)(5.87/0.617)(3.41/0.402)(3.80/0.570)(3.63/0.467)(3.12/0.693)
HES-HDR [14](5.67/0.335)(5.52/0.203)(6.40/0.576)(5.01/0.322)(4.61/0.442)(5.72/0.367)(4.81/0.530)
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.656, + 0.483, + 0.807 + ], + "angle": 0, + "content": "maintains details from LDR images but is hard to compensate for missing information. Sagiri [22] adopts diffusion models to provide high-quality images but fails to preserve consistency with inputs. NeurImg [14] has difficulty combining events with LDR image to predict over-exposed areas and has noise-like artifacts. HDRev [48] recovers HDR information, while it is difficult to deal with brightness and color adjustment in over-/under-exposed areas. Our method provides natural results and outperforms other methods both on missing information generation and fidelity preservation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.815, + 0.317, + 0.83 + ], + "angle": 0, + "content": "4.2. Evaluation on Real Data" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.834, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Comparison on DSEC dataset DSEC [8] is a stereo event camera dataset for driving scenarios. We choose parts3 of data recording HDR scenes as our testing data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.822 + ], + "angle": 0, + "content": "The quantitative evaluation using non-reference metrics is shown in Table 2, indicating that the proposed method achieves higher image quality than others. The qualitative evaluation of the DSEC dataset is depicted in Figure 7. EventHDR [53] fails to restore grayscale results. It's challenging for Liu et al. [24] and Sagiri [22] to hallucinate missing information. NeurImg [14] and HDRev [48] have problems utilizing the HDR information provided by events. The proposed method not only provides realistic and faithful results compared with other methods but also exhibits generation ability even with fewer events." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Comparison on HES-HDR dataset HES-HDR [14] is a hybrid event and spike HDR dataset. Its hybrid event dataset is suitable for our task. Although the non-reference metrics are not the best, we show better recovery ability as shown by qualitative evaluation3. The qualitative evalua" + }, + { + "type": "page_footnote", + "bbox": [ + 0.108, + 0.887, + 0.42, + 0.901 + ], + "angle": 0, + "content": "3More details are provided in the supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.904, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.316, + 0.688, + 0.331 + ], + "angle": 0, + "content": "Figure 8. Qualitative evaluation of real data on HES-HDR [14]." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.335, + 0.483, + 0.391 + ], + "angle": 0, + "content": "Table 3. Ablation studies of synthetic data on the dataset collected by Yang et al. [48]. \"Proposed*\" is calculated with adjusted GT \\( O_{\\mathrm{adj}} \\). The red metric shows the best performance among different ablation parts. \\( \\uparrow (\\downarrow) \\) means higher (lower) is better." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.396, + 0.485, + 0.525 + ], + "angle": 0, + "content": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
ConCond22.600.7490.2007.4848.144.23
RestCond20.790.7250.2479.3361.744.20
W/o Pyramid21.620.7280.2238.6746.993.89
W/o Refinement23.290.770.1707.2938.133.63
W/o structure26.650.9330.0935.6725.903.90
Proposed25.670.9260.0996.0127.093.86
Proposed*29.770.9390.0803.7922.623.86
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.53, + 0.484, + 0.592 + ], + "angle": 0, + "content": "tion of the HES-HDR dataset is depicted in Figure 8, which shows similar conclusions as DSEC [8]. Besides, even with unnatural color-shifted LDR image, our results have a more natural appearance and better fidelity." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.245, + 0.622 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Our ablation study includes two parts, different conditioning processes, and different refinement targets. The quantitative evaluations of different conditioning processes are shown in the upper part of Table 3. \"ConCond\" and \"Rest-Cond\" are as described in Section 3.2. \"W/o Pyramid\" represents removing the pyramid fusion module and adopting the lowest HDR features as conditions. The proposed one (\"W/o Refinement\") outperforms others. The qualitative evaluations in Figure 4 show that the proposed one can better utilize events and LDR information to provide faithful and realistic images. The quantitative evaluations of refinement and structure loss are depicted in the lower parts of Table 3. The ablation of removing the refinement module is shown by \"W/o Refinement\". Compared with diffusion-only method \"W/o Refinement\", the refinement module improves the fidelity in terms of supervised metrics. Employing GT as a supervision target (\"W/o structure\") makes the reference metrics a little better than \"Proposed\". However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.337, + 0.907, + 0.459 + ], + "angle": 0, + "content": "due to the different supervision targets, directly comparing the proposed method with GT cannot show our structure-preserving capacity. Therefore, we include \"Proposed*\", which is calculated with the adjusted GT \\( O_{\\mathrm{adj}} \\). The image quality reflected by NIQE of \"Proposed\" is better than \"W/o structure\". As shown by Figure 5, more natural and higher-quality images without distortion can be generated with the proposed refinement module." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.47, + 0.634, + 0.485 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.907, + 0.673 + ], + "angle": 0, + "content": "In this paper, we present a fidelity-preserving HDR reconstruction method based on diffusion models. Our method consists of two parts, event-guided conditioning generation and fine-grained detail refinement. We utilize an event-image encoder and a pyramid fusion module to efficiently and effectively fuse events and images, providing HDR conditions to guide the diffusion process and generate missing information. A refinement module with histogram-based structure loss is proposed to tackle distortion without changing the brightness and contrast. Experiments on both synthetic and real data demonstrate the generation and fidelity preservation ability of our method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.749 + ], + "angle": 0, + "content": "Limitations. Since diffusion models have uncertainty in the generation process, it is difficult to apply our method for video generation, for the challenge of maintaining consistency over consecutive frames." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.762, + 0.673, + 0.778 + ], + "angle": 0, + "content": "Acknowledgement" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.9 + ], + "angle": 0, + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resource." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11794" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.169 + ], + "angle": 0, + "content": "[1] Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-aware image-text alignment for opinion-unaware image quality assessment. arXiv preprint arXiv:2403.11176, 2024. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.228 + ], + "angle": 0, + "content": "[2] Hrishav Bakul Barua, Kalin Stefanov, KokSheik Wong, Abhinav Dhall, and Ganesh Krishnasamy. GTA-HDR: A large-scale synthetic dataset for HDR image reconstruction. arXiv preprint arXiv:2403.17837, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.231, + 0.483, + 0.27 + ], + "angle": 0, + "content": "[3] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE TIP, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.273, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[4] Sebastian Dille, Chris Careaga, and Yaqiz Aksoy. Intrinsic single-image hdr reconstruction. In ECCV, 2024. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.483, + 0.357 + ], + "angle": 0, + "content": "[5] Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafal K Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Jun Mitani. Deep reverse tone mapping. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[7] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[8] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[9] Michael Gharbi, Jiawen Chen, Jonathan Barron, Samuel Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.575, + 0.482, + 0.628 + ], + "angle": 0, + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.632, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[11] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.688, + 0.482, + 0.742 + ], + "angle": 0, + "content": "[12] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.746, + 0.482, + 0.8 + ], + "angle": 0, + "content": "[13] Jin Han, Chu Zhou, Peiqi Duan, Yehui Tang, Chang Xu, Chao Xu, Tiejun Huang, and Boxin Shi. Neuromorphic camera guided high dynamic range imaging. In CVPR, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[14] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 4, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.119 + ], + "angle": 0, + "content": "two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.15, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[17] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for hdr deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.206, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In Proceedings of the European Conference on Computer Vision (ECCV). ECCV, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.275, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.369 + ], + "angle": 0, + "content": "[20] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.373, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.443, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[22] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arxiv, 2024. 1, 2, 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.498, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[23] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-BIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.554, + 0.905, + 0.608 + ], + "angle": 0, + "content": "[24] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.61, + 0.905, + 0.718 + ], + "angle": 0, + "content": "[25] M Ronnier Luo, Guihua Cui, and Bryan Rigg. The development of the cie 2000 colour-difference formula: Ciede2000. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Français de la Couleur, 26(5):340-350, 2001. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.721, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[26] Nico Messikommer, Stamatios Georgoulis, Daniel Gehrig, Stepan Tulyakov, Julius Erbach, Alfredo Bochicchio, Yuanyou Li, and Davide Scaramuzza. Multi-bracket high dynamic range imaging with event cameras. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.79, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[27] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[28] Mohammad Mostafavi, Lin Wang, and Kuk-Jin Yoon. Learning to reconstruct hdr images from events, with applications to depth and flow prediction. IJCV, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11795" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.149 + ], + "angle": 0, + "content": "[30] Yohan Poirier-Ginter and Jean-François Lalonde. Robust unsupervised stylegan image restoration. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.151, + 0.482, + 0.192 + ], + "angle": 0, + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: An open event camera simulator. In Proc. of Conference on Robotics Learning, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.194, + 0.483, + 0.234 + ], + "angle": 0, + "content": "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.236, + 0.483, + 0.277 + ], + "angle": 0, + "content": "[33] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.279, + 0.482, + 0.32 + ], + "angle": 0, + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.483, + 0.363 + ], + "angle": 0, + "content": "[35] Richard Shaw, Sibi Catley-Chandar, Ales Leonardis, and Eduardo Pérez-Pellitero. HDR reconstruction from bracketed exposures and events. In BMVC, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.365, + 0.483, + 0.406 + ], + "angle": 0, + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.483, + 0.463 + ], + "angle": 0, + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.464, + 0.483, + 0.504 + ], + "angle": 0, + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.483, + 0.56 + ], + "angle": 0, + "content": "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[40] Chao Wang, Ana Serrano, Xingang Pan, Bin Chen, Karol Myszkowski, Hans-Peter Seidel, Christian Theobalt, and Thomas Leimkuhler. GlowGAN: Unsupervised learning of HDR images from LDR images in the wild. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.483, + 0.688 + ], + "angle": 0, + "content": "[41] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.69, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[42] Zejia Weng, Xitong Yang, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Genrec: Unifying video generation and recognition with diffusion models. arXiv preprint arXiv:2408.15241, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.747, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[43] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.829 + ], + "angle": 0, + "content": "[44] Wenhan Wu, Xian Liu, and Jie Chen. Image restoration with controlnet. IEEE Transactions on Image Processing, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[45] Garima Yadav, Saurabh Maheshwari, and Anjali Agarwal. Contrast limited adaptive histogram equalization based enhancement for real time video system. In 2014 International Conference on Advances in Computing, Communications and Informatics (ICACCI), 2014. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[46] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[47] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Mania: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[48] Yixin Yang, Jin Han, Jinxiu Liang, Imari Sato, and Boxin Shi. Learning event guided high dynamic range video reconstruction. In CVPR, 2023. 1, 2, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.405, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.905, + 0.489 + ], + "angle": 0, + "content": "[53] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, 2021. 1, 2, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.489 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11796" + } + ] +] \ No newline at end of file diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_origin.pdf b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7ecba965019901306c9b9c72b8726696e24dea91 --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/1d6e20a9-59f5-4c36-8c14-b470c2a5ccd6_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:33b78ce2ea8db56767ed773f31a25b2aee74e2b74a5259c974b1b0d42589d7f1 +size 8168553 diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/full.md b/2025/Event-guided HDR Reconstruction with Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..60569759a220535087e3c79d059637cc1103dc7b --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/full.md @@ -0,0 +1,365 @@ +# Event-guided HDR Reconstruction with Diffusion Priors + +Yixin Yang $^{1,2,\dagger}$ Jiawei Zhang $^{3}$ Yang Zhang $^{1,2}$ + +Yunxuan Wei $^{3}$ Dongqing Zou $^{4}$ Jimmy S. Ren $^{3,5}$ Boxin Shi $^{1,2,*}$ + +1 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University + +$^{2}$ Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University + +$^{3}$ SenseTime Research $^{4}$ PBVR $^{5}$ Hong Kong Metropolitan University + +{yangyixin93, shiboxin}@pku.edu.cn, Github Page: github.com/YixinYang-00/HDRev-Diff + +# Abstract + +Events provide High Dynamic Range (HDR) intensity change which can guide Low Dynamic Range (LDR) image for HDR reconstruction. However, events only provide temporal intensity differences and it is still ill-posed in over-/under-exposed areas due to missing initial reference brightness and color information. With strong generation ability, diffusion models have shown their potential for tackling ill-posed problems. Therefore, we introduce conditional diffusion models to hallucinate missing information. Whereas, directly adopting events and LDR image as conditions is complicated for diffusion models to sufficiently utilize their information. Thus we introduce a pretrained events-image encoder tailored for HDR reconstruction and a pyramid fusion module to provide HDR conditions, which can be efficiently and effectively utilized by the diffusion model. Moreover, the generation results of diffusion models usually exhibit distortion, particularly for fine-grained details. To better preserve fidelity and suppress distortion, we propose a fine-grained detail recovery approach using a histogram-based structural loss. Experiments on real and synthetic data show the effectiveness of the proposed method in terms of both detail preservation and information hallucination. + +# 1. Introduction + +Extending the dynamic range of an image (usually with low dynamic range, LDR) to record the brightness of real scenes plausibly is particularly useful for daily photography and computer vision tasks. High dynamic range (HDR) imaging can be, in the most direct way, achieved by utilizing neural networks to hallucinate missing information in over-/under-exposed areas [4, 5, 22, 24], as shown in Figure 1 (a), which is challenging to faithfully recover + +![](images/62146720e129dfa968c35c47d423c59e5f4698ccedfafb5fdf21f9a824d6bd02.jpg) +(a) Dille et. al. + +![](images/8c51f95e84dc932d2802ce39488b0639c37b04c35a729cfbf67c117216778dd0.jpg) +(b) Events + +![](images/6f2c90e6df82210e9bd3da228d4f921e98a1810a649abd06611c19768d0947ea.jpg) +(c) EventHDR + +![](images/1ae2c101566f5800d0e01250561b958a7c4bd48811ea929a68dd66440aa727a0.jpg) +(d) LDR image + +![](images/9e79e0012e1eaa40b5ac7c9a54f1fec3426ac3f4b01ec6d8e95ee81e3e1f684c.jpg) + +![](images/a8d7b7f26b188735f773ee416cfae1762946492602decfa1771b4d41dba7fe87.jpg) +(e) HDRev + +![](images/dfd498d74bd0f46b0354755a13961313ae0a3d4b1bc598eb7c55aa2a43b8e4e1.jpg) +(f) Sagiri +Figure 1. HDR image can be recovered by only using a single LDR image (d) as input (e.g., Dille et al. [4]), which is a difficult problem as shown in (a). With only events (b) as input, event-to-image reconstruction (e.g., EventHDR [53]) (c) recovers an HDR scene without color. Event-guided HDR methods (e.g., HDRev [48]) (e) take both events and the LDR image as input, which recovers colorful HDR scenes with moderate details. Diffusion-based methods (e.g., Sagiri [22]) (f) take the LDR image (and optionally the text prompt) as input showing visually pleasing results but may have challenges in maintaining consistent contents with the original scene. The proposed method (g) takes events and the LDR image as input and leverages the diffusion priors to recover a plausible and realistic HDR image with rich details and less distortion. scene details without any auxiliary information. Neuromorphic cameras, such as event cameras, are proven to be helpful in providing additional details for HDR purpose [28, 32, 53]. Event cameras capture intensity changes + +![](images/ce99c953bdb05d0ee04ced2ae08a814ced006d2ba1e7a2af30c8bb62749a0928.jpg) +(g) Ours + +$\Delta$ : Events +$\Delta$ :Colorful results +$\Delta$ : More realistic +$\Delta$ : Less distortion +: CNN models +Diffusion models + +by triggering events asynchronously, which inherently contains HDR information (e.g., 120 dB for DAVIS340), and events can be directly converted to HDR images, as demonstrated by EventHDR [53] in Figure 1 (c). The event-to-image reconstruction [28, 32, 53] can recover details in over-/under-exposed regions, but fail to provide colorful results due to lacking color information. + +To generate colorful HDR images, event-guided HDR reconstructions [13, 14, 48] take both LDR image and events as inputs, complementarily learning to predict color and intensity in over-/under-exposed areas, as demonstrated in Figure 1 (e). They can reconstruct HDR information, e.g., the cloud in the green box, and provide colorful results. Although regression models are adopted to build the relationship between the predicted pixel values and its neighboring pixels, those methods cannot recover proper intensity, especially when the LDR image has large over-/under-exposed areas. Because absolute HDR intensity (events only contain thresholded radiance changes) and color information are still missing in these regions, it is difficult for regression models to appropriately hallucinate their values. + +Diffusion models [16, 37-39] are emerging generative models that show great potential for generating plausible and high-quality images. Even if for ill-posed problems, they can provide visually pleasing results thanks to the expressive power of diffusion priors. Diffusion models have also been adopted for image restoration [23, 43, 49]. Recent attempts, e.g., Sagiri [22], have shown the potential of diffusion priors for HDR reconstruction by providing more visually realistic details in over-exposed areas, e.g., the cloud in the sky of Figure 1 (f). However, the generated HDR images still cannot appropriately keep the fidelity of the original scene, and additional distortion are brought into well-exposed regions, as shown by the red box in Figure 1 (f). + +To reconstruct visually pleasing and faithful HDR images with less distortion, we introduce events to provide differential HDR intensity and adopt conditional diffusion models [50] to compensate for the ill-posed nature of hallucinating missing information and constrain the generative model to be consistent with the original scene. Specially, we focus on solving the following two key problems via employing diffusion models in event-guided HDR reconstruction: 1) Condition extraction: Since events and LDR image are two different modalities of visual representation, directly adopting them as conditions is non-compatible for the conditional diffusion models to conduct effective fusion. Inspired by the modality alignment of HDRev [48], we introduce an event-image encoder and a pyramid fusion module to fuse events and LDR image in different resolutions and provide conditions to guide the generation process. 2) Details preservation: Applying diffusion models to HDR reconstruction [22, 23] suffers from fine-grained detail distortion, as shown in Figure 1 (f). Such artifacts become + +more obvious when LDR images are high quality and have large well-exposed areas. To refine the fine-grained details and make them consistent with LDR images and events, we design a refinement module with histogram-based structure loss. The proposed method is validated on both synthetic and real data and outperforms existing methods by providing fidelity-preserving HDR reconstruction, as illustrated in Figure 1 (g), with the following technical contributions: + +- integrating events and the conditional diffusion models to recover missing information faithfully; +- adopting the pretrained event-image encoder and pyramid fusion to effectively apply conditions; and +- designing the refinement module and histogram-based structure loss to further strengthen fine-grained details. + +# 2. Related works + +Image-based HDR reconstruction Many approaches leverage the capabilities of neural networks to directly hallucinate HDR information from training data, such as CNNs [5, 6, 9], Generative Adversarial Networks (GANs) [40], and diffusion models [7, 22]. Liu et al. [24] incorporates the LDR image formulation pipeline into its hallucination process to perform step-by-step prediction. For a more comprehensive view, we refer to GTA-HDR [2]. Since information is unavoidably missing in over-/underexposed areas, the above methods only rely on hallucination, which may not correspond to real scenes. + +Event-based HDR reconstruction Benefiting from the HDR property of event cameras, previous works reconstruct HDR images from events with recurrent neural networks [32, 33, 53] or GANs [28, 41]. Since events only record logarithmic irradiance changes, it's difficult to reconstruct color images. To reconstruct color HDR images, Han et al. [13, 14] adopted a hybrid camera system to capture events and frame, enabling event-guided HDR reconstruction. Yang et al. [48] treated events and frame as different modalities and proposed multi-modalities alignment and fusion to provide HDR videos. Other methods [26, 35] incorporate events and bracketed exposure for HDR reconstruction, while long-exposed images bring blurry artifacts. The proposed method takes events and LDR image as input, incorporating diffusion models to compensate for realistic and faithful information in over-/under-exposed areas. + +Generative priors Generative Priors capture the underlying data distribution, enabling the generated images to follow real-world image distribution. GAN [10] has shown its potential to leverage generative priors for image restoration [20, 21, 30]. Recently, diffusion models [16, 37-39] have widely adopted as effective generative priors for producing realistic images [29, 34, 42, 50], which also demonstrate strong potential for image restorations [23, 43, 49], specifically, for HDR image reconstruction from bracketed exposure [12, 17, 46]. There are some attempts [11, 22] to + +![](images/48d481cd855603f94ca7c593645d5fe45d625f9a99def7ac4b0b1f6b7bb25e1c.jpg) +Figure 2. Pipeline of the proposed method consists of three parts: diffusion model and denoising process (green part), event-image conditioning (blue part), and detail refinement (orange part). The blue part fuses LDR image and events to generate injected features for Denoiser as described in Section 3.2. The green part adopts the injected features as conditions to estimate noise from latent $z_{t}$ step by step with fixed pretrained parameters. To refine the distortion of denoised images, the refinement module in the orange part, whose architecture is shown on the right, is proposed to leverage the HDR features and denoised latent $z_{0}$ . Considering the uncertainty of diffusion results in color and brightness, we introduce structure loss, which adopts adaptive Ground Truth (GT) as the supervision target, as described in Section 3.3. After refinement, we obtain a plausible and realistic refined image without distortion. + +take advantage of diffusion models to generate HDR images from single LDR image, while it cannot provide faithful reconstruction due to the inevitable dynamic clipping. + +# 3. Proposed Method + +Section 3.1 introduces basic concepts about diffusion models and event cameras. The pipeline of the proposed method is illustrated in Figure 2. The green part is the pretrained latent diffusion model, which performs denoising in latent space step by step with the injected feature from the blue part. The blue part is the proposed event-guided conditioning and generation process, which adopts an event-image encoder and a pyramid fusion to utilize events and the LDR image efficiently and effectively, as described in Section 3.2. The orange part is the fine-grained detail enhancement based on the Variant Auto Encoder (VAE) decoder, which learns to refine fine-grained details by the proposed histogram-based structure loss as described in Section 3.3. Training details are described in Section 3.4. + +# 3.1. Preliminary + +Diffusion models Diffusion models [16, 34, 38] can generate realistic high-quality images by diffusing and denoising processes. Latent diffusion models [34] are a type of diffusion models that perform those processes on latent space. During training, the latent diffusion models project the clean image into the latent space with a VAE encoder as $z_0$ , then diffuse $z_0$ by adding Gaussian noise for $n$ steps to get the noisy latent $z_n$ . Adding noise in one step is as: + +$$ +z _ {n} = \sqrt {a _ {n}} z _ {n - 1} + \sqrt {1 - a _ {n}} \epsilon_ {n - 1}, \tag {1} +$$ + +where $n \in [1, N]$ means the $n$ -th diffusion step, $N$ is the total diffusion steps, $\epsilon_n \sim \mathcal{N}(\mathbf{0}, \mathbf{I})$ is the added noise, $a_n$ is + +the pre-defined parameter of the noise scheduler. Defining $\bar{a}_n = \prod_{i=1}^n a_i$ , the closed form equation of Eq. (1) [16] is: + +$$ +z _ {n} = \sqrt {\bar {a} _ {n}} z _ {0} + \sqrt {1 - \bar {a} _ {n}} \epsilon , \tag {2} +$$ + +where $\epsilon \sim \mathcal{N}(\mathbf{0},\mathbf{I})$ . The denoising process trains a denoiser $\epsilon_{\theta}$ to estimate the noise $\epsilon$ usually by minimizing Mean Square Error (MSE) $\mathcal{L}_{n,z_0,\epsilon} = \| \epsilon -\epsilon_{\theta}(z_n,n)\|$ . As the variable $z_{n}$ will be an approximately standard Gaussian distribution when $n$ is large enough, during inference, the denoising process often starts from an i.i.d. noise $z_{N}$ and generates a clean image $z_{0}$ using $\epsilon_{\theta}$ over $N$ steps. + +Event triggering and stacking Event cameras trigger an event $(t,x,y,p)$ when illuminance changes of pixel $(x,y)$ in the logarithm domain over a predefined threshold $\eta$ in time $t$ , where the polarity $p\in \{-1,1\}$ indicates the decrease and increase of the illuminance. Mathematically, an event is generated when the following inequation occurs: + +$$ +\left\| \log I _ {x, y} (t) - \log I _ {x, y} \left(t _ {\text {r e f}}\right) \right\| \geq \eta , \tag {3} +$$ + +where $I_{x,y}(t)$ is the illuminance at time $t$ , $I_{x,y}(t_{\mathrm{ref}})$ is the reference illuminance level, $t_\mathrm{ref}$ is the last event triggered time of pixel $(x,y)$ , $\eta$ is a pre-defined event threshold. We convert the stream-like events to tensors using voxel grid [52], which encodes temporal information in a C-channel 3D volume by discretizing event timestamps into $C$ temporal bins. Following Rebecq et al. [32, 33], $k$ -th event $(t_k,x_k,y_k,p_k)$ distributes its polarity $p_k$ to the two closest temporal bins related to its normalized timestamp by: + +$$ +E _ {j} ^ {x, y} = \sum_ {x _ {k} = x, y _ {k} = y} p _ {k} \max (0, 1 - | \tilde {t} _ {k} - j |), \tag {4} +$$ + +where $\tilde{t}_k = \frac{C - 1}{\Delta t} (t_k - t_0)$ is the normalized timestamp, $\Delta_t = \max_k(t_k) - t_0$ is the time span of events, $t_0$ is the start timestamp, and $j \in [0, C - 1]$ is the index of the bin. + +![](images/26d7d1d70d68d9dedeb95de0f48359d962b474bc969b0c535093b34a0c8c15cc.jpg) +Figure 3. Pipeline of the event-guided conditioning process, with comparison to conditioning using HDRev [48] restored image. Directly concatenating events and LDR image cannot fully utilize events, thus we introduce an event-image encoder to fuse them as shown in the orange part. The HDRev [48] restored images by decoding HDR features can serve as input to the embedding module, as shown in the yellow part. The proposed pyramid fusion module, as shown in the blue part, fuses different levels HDR features to provide conditions. + +![](images/b044f0e8017dfb15c9ff581463b50135748688fc6798c14fa2b0279ffb810a06.jpg) + +![](images/8bf7ba3d618c03b8f712d7598c448f854c0f5a1140c77866026f4ca6a04cff98.jpg) + +# 3.2. Event-guided Conditioning and Generation + +To hallucinate missing information for HDR reconstruction, previous methods mainly come in two ways: adopting better hallucination models (e.g., diffusion models [22]) or introducing HDR sensors (e.g., event cameras [13, 14, 48]). Diffusion models generate high-quality and realistic images by modeling real-world image distributions with diffusion priors. Event cameras provide differential HDR intensity, making events well-suited as conditions for diffusion models to generate more accurate details in over-/under-exposed areas, thereby compensating the dynamic range of LDR images. Thus we propose to leverage both event cameras and diffusion models to hallucinate realistic HDR images to be more consistent with real scenes. + +To generate images corresponding to conditions, conditional models [18, 50] are proposed to generate features injected into specific layers of the well-trained denoiser. ControlNet [50] is one of the widelyconditional models for image restoration [23, 43, 44], which generates and injects features for each down-sample and middle layers of the denoiser. As illustrated in the interaction between the blue and green parts of Figure 2, we perform feature injection similar to ControlNet. In the green part, we utilize Stable Diffusion [34] V1.5 as the pretrained denoiser $\epsilon_{\theta}$ to denoise from the initial latent $z_{N}$ step by step. The generated image is obtained by decoding the estimated initial latent $\hat{z}_0$ with VAE decoder. With an image-like input, ControlNet employs an embedding module to generate conditions, then adopts a control module to obtain injected features. However, it is challenging for the embedding module to effectively fuse LDR image with events and extract useful information. + +Events record temporal intensity changes, making them distinct from LDR images. Directly concatenating stacked events $E$ and LDR images $I_{\mathrm{LDR}}$ together as the input for the embedding module $\mathcal{T}_{\mathrm{con}}$ introduces HDR information, where the conditioning process is: + +$$ +\mathcal {E} _ {\text {C o n C o n d}} \left(I _ {\text {L D R}}, E\right) = \mathcal {T} _ {\text {c o n}} \left(\left[ I _ {\text {L D R}}, E \right]\right), \tag {5} +$$ + +[.,] denotes concatenation. But it cannot fully utilize HDR information encoded in events, e.g., decoration and cable + +in over-exposed areas cannot be accurately reconstructed in Figure 4 (d). Therefore, a specifically designed HDR information extraction model is needed. HDRev [48] performs event-guided HDR reconstruction through representation alignment and feature fusion, which addresses the modality gap between events and LDR images. It performs representation alignment through pre-training to obtain modality-specific encoders as shown in the orange part of Figure 3. The LDR image and events are encoded into modality-specific features, represented by green features (LDR image) and yellow features (events), respectively. Those two features are fused through the fusion module to obtain HDR features. The restoration process can be written into: + +$$ +I _ {\text {r e s t o r e d}} = \mathcal {B} \left(\mathcal {H} \left(I _ {\mathrm {L D R}}, E\right)\right), \tag {6} +$$ + +where $\mathcal{B}$ and $\mathcal{H}$ are the decoder and encoder in HDRev [48]. The restoration process utilizes events to recover more details, as shown by Figure 4 (e), which offers a better reconstruction of the decoration and cable than (d). The intuitive way to better utilize events for conditioning is adopting $I_{\mathrm{restored}}$ as input for the embedding module $\mathcal{T}_{\mathrm{rest}}$ as shown in the yellow part of Figure 3. The conditioning process1 is: + +$$ +\mathcal {E} _ {\text {R e s t C o n d}} \left(I _ {\text {L D R}}, E\right) = \mathcal {T} _ {\text {r e s t}} \left(\mathcal {B} \left(\mathcal {H} \left(I _ {\text {L D R}}, E\right)\right)\right). \tag {7} +$$ + +Diffusion models improve image quality as shown by the improved shape and color of the table in Figure 4 (f) compared to (e). Adopting $I_{\mathrm{restored}}$ as conditions ("RestCond") better utilizes events and provides more details than "Con-Cond", e.g., the decoration in Figure 4 (f) is better than (d). + +Although HDRev is an encoder-decoder architecture, the fusion between LDR image and events is performed only in the encoder. Applying $I_{\mathrm{restored}}$ as conditions not only suffers from a decoder-encoder redundant computation $\mathcal{T}_{\mathrm{rest}}(\mathcal{B}(\cdot))$ as shown in the yellow part of Figure 3, but also fails to reconstruct details from events, e.g., the shape of table in the red box of Figure 4, and is difficult to adjust the proper brightness of over-exposed areas, e.g., the cloud in Figure 1 (e). The embedding module $\mathcal{T}_{\mathrm{rest}}$ may misunderstand + +![](images/fbe39092ac18e63fb7979b084c7aa0bd69decec4bbd5a69fc9070d9d7e63707e.jpg) +(a) LDR + +![](images/de5214b870f828a3a944b400bfdade9305f5cfb4659b4df235384bf0022ea3dc.jpg) +(b) Events + +![](images/693b024506a71d8b6f956f9215074ea3e6ba8cec76ec9038e176069b3f5ae773.jpg) +(c) GroundTruth +Figure 4. Comparison with different conditioning ways: directly concatenating events and LDR image ("ConCond") and HDRev [48] restored image ("RestCond"). "ConCond" (d) cannot fully utilize events in over-exposed areas to reconstruct cable. Illustrated by "Rest-Cond" (f), the unsatisfactory results of decoration and table in "Restored" (e) failed to provide HDR information for conditioning in over-exposed areas. The proposed conditioning module can restore missing information as depicted in (g). + +![](images/1be280b534db4d0a78dc333062b31f958dd19452ec93a405299693334ef80f20.jpg) +(d) ConCond + +![](images/2f741ba38a4d6d50b3f018e32a3fd7564e8bc0beaeb1e54220ff8b30aa1155b4.jpg) +(e) Restored + +![](images/02580f47034e2d670432cce1c7136d2190f2b1c5e96f5b4a1b47eab3ef544d82.jpg) +(f) RestCond + +![](images/9e70a1d9dba1fa16735cc8f97e08d05f4953dabd41cb91b72545e91219bf06a5.jpg) +(g) Proposed + +these unsatisfactory areas as shown by the shape of the decoration in the red box of Figure 4 (f). + +To mitigate issues of poor brightness adjustment in HDRev, and leverage its advantage in event-image fusion, we adopt HDR features from each layer of its event-image encoder to provide conditions. Directly adopting HDR features as conditions is not only more efficient, because of removing the redundant decoder-encoder process, but also more effective, because of avoiding introducing unsatisfactory reconstruction from the decoder $\mathcal{B}$ . However, the HDR features are pyramidal, in which different levels of visual information are encoded separately. To fuse different level HDR features, we propose a pyramid fusion module $\mathcal{P}$ as shown in the blue part of Figure 3, which fuses different level features with down-sample layers and feature concatenation to provide conditions. More accurate details and colors can be recovered with it as shown in Figure 4 (g). The proposed conditioning process $\mathcal{E}$ consists of an event-image encoder and pyramid fusion module, denoted by: + +$$ +\mathcal {E} \left(I _ {\mathrm {L D R}}, E\right) = \mathcal {P} \left(\mathcal {H} \left(I _ {\mathrm {L D R}}, E\right)\right). \tag {8} +$$ + +Then the predicted noise is $\epsilon_{\theta}(z_n,n,\mathcal{C}(\mathcal{E}(I_{\mathrm{LDR}},E))$ , where $\mathcal{C}$ is the control module derived from ControlNet [50]. With the pretrained denoiser $\epsilon_{\theta}$ , we are able to train the control module $\mathcal{C}$ and conditioning process $\mathcal{E}$ to hallucinate HDR images by minimizing the noise loss: + +$$ +\mathcal {L} _ {n, z _ {0}, \epsilon_ {n}} = \| \epsilon_ {n} - \epsilon_ {\theta} (z _ {n}, n, \mathcal {C} (\mathcal {E} (I _ {\mathrm {L D R}}, E))) \| ^ {2}. \tag {9} +$$ + +# 3.3. Fine-grained Detail Refinement + +Although diffusion models can reconstruct realistic HDR images, it suffers from distortion, especially for fine-grained details as shown in Figure 1 (f) and Figure 5 "W/o Refinement". Therefore, we aim to introduce information from events and LDR image to refine those distortion. Typically, we adopt HDR features from the event-image encoder for refinement. As shown in the orange part of Figure 2, our refinement module is a decoder $\mathcal{D}$ in which the up-sample layers are initialized from the VAE decoder. It fuses HDR features $\mathcal{H}(I_{\mathrm{LDR}},E)$ with latent $z_0$ to provide the refined HDR image $H$ : + +$$ +H = \mathcal {D} \left(\mathcal {H} \left(I _ {\mathrm {L D R}}, E\right), z _ {0}\right). \tag {10} +$$ + +By training the refinement module with the Ground Truth (GT), we can refine some distortion as shown in Figure 5 "W/o Structure". Due to the inconsistent color between GT and diffusion results as illustrated in Figure 5, applying GT as a supervision target brings color artifacts similar to regression-based methods [24, 48] as depicted in the red box of Figure 5 "W/o Structure". Those artifacts come from the discrepancy between the deterministic GT and the uncertainty in diffusion results during the training process of refinement. Additionally, it also depicts that the details in diffusion results have been wiped (the road curb in the red box) after this refinement. Because it is easier for the refinement to minimize the total loss by reducing the brightness and color difference than the detail difference. + +Meanwhile, although the diffusion results suffer from distortion, they already have visually pleasant brightness and color consistent with LDR images. Therefore, we tend to maintain the brightness and color properties of diffusion results and only focus on detail refinement by introducing local histogram matching [45] to adjust the GT to serve as a new supervision target. Local histogram matching [45] is able to adjust the pixel intensity distribution to match the histogram of a target image for each local area, which is suitable for adjusting only color and brightness, and maintaining details of GT. Adopting local histogram matching to adjust ground truth image $O$ to diffusion result $H_{\mathrm{diff}}$ is as: + +$$ +O _ {\mathrm {a d j}} = \operatorname {H i s t} (O, H _ {\text {d i f f}}). \tag {11} +$$ + +Our histogram-based structure loss adopts the adjusted image $O_{\mathrm{adj}}$ as ground truth and is composed of two losses. The first one is MSE loss: + +$$ +\mathcal {L} _ {\mathrm {M S E}} (H, O _ {\mathrm {a d j}}) = \| H - O _ {\mathrm {a d j}} \| ^ {2}, \tag {12} +$$ + +and the perceptual loss $\mathcal{L}_{perc}$ based on the feature maps extracted by the pretrained VGG-16 [36] network: + +$$ +\mathcal {L} _ {\text {p e r c}} = \sum_ {l} \| \phi_ {l} (H) - \phi_ {l} \left(O _ {\text {a d j}}\right) \| ^ {2} + \| \mathcal {G} _ {l} ^ {\phi} (H) - \mathcal {G} _ {l} ^ {\phi} \left(O _ {\text {a d j}}\right) \| ^ {2}, \tag {13} +$$ + +![](images/091bd60993080a16a587d3cd8e95327b29cb7070d8899032bf3e24a0d1b0cab6.jpg) +Figure 5. Validation for the effectiveness of the structure loss. With only conditioning and diffusion process, the results of "W/o Refinement" suffer from distortion as shown in the green box. Training the refinement module with ground truth, some distortions are refined, but introducing unnatural saturation adjustment and unsatisfactory detail prediction as shown in the red box of "W/o Structure". We introduce local histogram matching [45] to adjust the brightness and color of ground truth to diffusion results, making the refinement module focus on refining the details. The adapted ground truth is shown in "Adapted GT". As shown in "Proposed", by adopting adapted ground truth as a supervision target, the refinement module with structure loss refines distortion with more natural colors and details. + +![](images/cbb780887490dc6c0d0ba6ad31d1b4bd61311ee38169c07ad948e49afecf553e.jpg) + +![](images/2c8e0d3deaa7378e30eb5b74d25ae0885c4f85e96648f82b41d861bdda21a1d9.jpg) + +![](images/765b5887378ad2204b9be39ccbea1de9ac1c1c9e370b3451465e5eb39d1aa5c2.jpg) + +![](images/f957d898bf9dc599bcdd8e9216c0ab098ecce1d032a55024b188210e832dd08a.jpg) + +![](images/8918109efd6f8d94ac290a21a5d650ab71a3b36a73bf67bdbd5656d45009e362.jpg) + +![](images/15dfe72c48d5cc6e9f93f560bb9449fd27012baa0ca7c4dc18b2a4709e8fd4a3.jpg) + +where $\phi_l$ is the extracted feature from $l$ -th layer of VGG-16, $\mathcal{G}_l^\phi$ is the Gram matrix of $\phi_l$ . Our structure loss is: + +$$ +\mathcal {L} _ {\text {s t r u c t}} = \alpha \mathcal {L} _ {\mathrm {M S E}} + \beta \mathcal {L} _ {\text {p e r c}}, \tag {14} +$$ + +where $\alpha = 0.01, \beta = 0.001$ are the balance weights of different terms. Using $O_{\mathrm{adj}}$ as the supervision target enables the refinement module to focus on fine-grained detail refinement as illustrated in Figure 5 "Complete", in which the building is undistorted and the road curb is natural. + +# 3.4. Training details + +As a large number of paired events, LDR image, and GT is needed for training, we adopt 733 HDR images collected by Yang et al. [48] and follow the same data generation process to generate training and testing data with resolution (512, 512). We first simulate 733 HDR videos with random global motion by generating random camera motion trajectories [3]. We utilize the event simulator [31] and virtual camera [6] to simulate the events and LDR images, respectively. The dataset generated from 733 images is separated into training and testing following the setting of Yang et al. [48], while 663 for training and 70 for testing. During training, we randomly over- or under-exposed $20\%$ to $50\%$ pixels of the HDR images to generate LDR images. + +Other implementation details. The proposed method is implemented with the Pytorch framework and runs on a single NVIDIA GeForce RTX 3090 GPU. For optimization, we use the ADAM optimizer [19] with a linear decay learning rate scheduler starting from step 0 and an initial learning rate of $\gamma = 10^{-5}$ . DDIM noise scheduler is adopted both in training and sampling, with training steps $N = 1000$ and linear-scaled beta schedule. We adopt $C = 5$ same to HDRev [48] to apply their pretrain model. During sampling, we adopt classifier-free guidance with guidance scale $u = 1.5$ and inference steps $n = 9$ to obtain our results. + +Table 1. Quantitative evaluation of synthetic data on dataset collected by Yang et al. [48]. The red metric shows the best performance. $\uparrow (\downarrow)$ means higher (lower) is better. + +
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
Liu et al. [24]18.350.7710.27615.3378.414.02
EventHDR [53]11.040.3340.44723.44182.114.58
Sagiri [22]12.500.4530.41417.6783.465.35
HDRev [48]14.050.6190.23817.6946.233.88
NeurImg [14]18.530.6210.33817.42105.074.04
Dille et al. [4]19.730.8200.2439.7676.834.22
Ours25.670.9260.0996.0127.093.86
+ +# 4. Experiments + +The comparison with existing methods including four categories: event-based EventHDR [53], single-image CNN-based Liu et al. [24] and Dille et al. [4], single-image diffusion-based Sagiri [22], and event-guided NeurImg-HDR [14] and HDRev [48]. Experiments are conducted both on synthetic data and real data, as described in Section 4.1 and Section 4.2 respectively. + +Metrics For fidelity preserving, we adopt reference metrics including peak signal-to-noise ratio (PSNR), structural similarity (SSIM), the perceptual error with learned perceptual image patch similarity (LPIPS) [51], the Frechet Inception Distance (FID) [15], and the Color Difference Evaluation 2000 (CIEDE) [25], respectively. A non-reference metric named the Natural Image Quality Evaluator (NIQE) [27] is adopted to evaluate the image quality. + +# 4.1. Evaluation on Synthetic Data + +The quantitative evaluation is reported in Table 1. The proposed method outperforms the previous methods in terms of fidelity and structure preservation as shown by reference metrics. Also, ours image quality outperforms other methods, as indicated by NIQE [27]. The qualitative evaluation is illustrated in Figure 6. EventHDR [53] reconstructs faithful grayscale HDR images without color. Liu et al. [24] + +![](images/cce0b5cd0446b2cbf19bee4561fb0576067408b861fd8430c1f37cd9d9bc97bd.jpg) +Figure 6. Qualitative evaluation of synthetic data on dataset collected by Yang et al. [48]. + +![](images/a80edf8c56f24607ea4705587b770068af85c464680f34370ce212d7326d3ee5.jpg) +Figure 7. Qualitative evaluation of real data on DSEC [8]. + +Table 2. Quantitative evaluation (NIQE [27]↓/MANIQA [47]↑/QualiClip [1]↑) of real data. Blue metrics indicate the second-best. + +
DatasetLiu et al. [24]EventHDR [53]Sagiri [22]HDRev [48]NeurImg [14]Dille et al.Ours
DSEC [8](3.65/0.466)(4.00/0.270)(5.87/0.617)(3.41/0.402)(3.80/0.570)(3.63/0.467)(3.12/0.693)
HES-HDR [14](5.67/0.335)(5.52/0.203)(6.40/0.576)(5.01/0.322)(4.61/0.442)(5.72/0.367)(4.81/0.530)
+ +maintains details from LDR images but is hard to compensate for missing information. Sagiri [22] adopts diffusion models to provide high-quality images but fails to preserve consistency with inputs. NeurImg [14] has difficulty combining events with LDR image to predict over-exposed areas and has noise-like artifacts. HDRev [48] recovers HDR information, while it is difficult to deal with brightness and color adjustment in over-/under-exposed areas. Our method provides natural results and outperforms other methods both on missing information generation and fidelity preservation. + +# 4.2. Evaluation on Real Data + +Comparison on DSEC dataset DSEC [8] is a stereo event camera dataset for driving scenarios. We choose parts3 of data recording HDR scenes as our testing data. + +The quantitative evaluation using non-reference metrics is shown in Table 2, indicating that the proposed method achieves higher image quality than others. The qualitative evaluation of the DSEC dataset is depicted in Figure 7. EventHDR [53] fails to restore grayscale results. It's challenging for Liu et al. [24] and Sagiri [22] to hallucinate missing information. NeurImg [14] and HDRev [48] have problems utilizing the HDR information provided by events. The proposed method not only provides realistic and faithful results compared with other methods but also exhibits generation ability even with fewer events. + +Comparison on HES-HDR dataset HES-HDR [14] is a hybrid event and spike HDR dataset. Its hybrid event dataset is suitable for our task. Although the non-reference metrics are not the best, we show better recovery ability as shown by qualitative evaluation3. The qualitative evalua + +![](images/a8bb1686632dcfa21009aa04b4ed909e9cb103dd165faea8e19686e631e4ffce.jpg) +Figure 8. Qualitative evaluation of real data on HES-HDR [14]. + +Table 3. Ablation studies of synthetic data on the dataset collected by Yang et al. [48]. "Proposed*" is calculated with adjusted GT $O_{\mathrm{adj}}$ . The red metric shows the best performance among different ablation parts. $\uparrow (\downarrow)$ means higher (lower) is better. + +
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
ConCond22.600.7490.2007.4848.144.23
RestCond20.790.7250.2479.3361.744.20
W/o Pyramid21.620.7280.2238.6746.993.89
W/o Refinement23.290.770.1707.2938.133.63
W/o structure26.650.9330.0935.6725.903.90
Proposed25.670.9260.0996.0127.093.86
Proposed*29.770.9390.0803.7922.623.86
+ +tion of the HES-HDR dataset is depicted in Figure 8, which shows similar conclusions as DSEC [8]. Besides, even with unnatural color-shifted LDR image, our results have a more natural appearance and better fidelity. + +# 4.3. Ablation Study + +Our ablation study includes two parts, different conditioning processes, and different refinement targets. The quantitative evaluations of different conditioning processes are shown in the upper part of Table 3. "ConCond" and "Rest-Cond" are as described in Section 3.2. "W/o Pyramid" represents removing the pyramid fusion module and adopting the lowest HDR features as conditions. The proposed one ("W/o Refinement") outperforms others. The qualitative evaluations in Figure 4 show that the proposed one can better utilize events and LDR information to provide faithful and realistic images. The quantitative evaluations of refinement and structure loss are depicted in the lower parts of Table 3. The ablation of removing the refinement module is shown by "W/o Refinement". Compared with diffusion-only method "W/o Refinement", the refinement module improves the fidelity in terms of supervised metrics. Employing GT as a supervision target ("W/o structure") makes the reference metrics a little better than "Proposed". However, + +due to the different supervision targets, directly comparing the proposed method with GT cannot show our structure-preserving capacity. Therefore, we include "Proposed*", which is calculated with the adjusted GT $O_{\mathrm{adj}}$ . The image quality reflected by NIQE of "Proposed" is better than "W/o structure". As shown by Figure 5, more natural and higher-quality images without distortion can be generated with the proposed refinement module. + +# 5. Conclusion + +In this paper, we present a fidelity-preserving HDR reconstruction method based on diffusion models. Our method consists of two parts, event-guided conditioning generation and fine-grained detail refinement. We utilize an event-image encoder and a pyramid fusion module to efficiently and effectively fuse events and images, providing HDR conditions to guide the diffusion process and generate missing information. A refinement module with histogram-based structure loss is proposed to tackle distortion without changing the brightness and contrast. Experiments on both synthetic and real data demonstrate the generation and fidelity preservation ability of our method. + +Limitations. Since diffusion models have uncertainty in the generation process, it is difficult to apply our method for video generation, for the challenge of maintaining consistency over consecutive frames. + +# Acknowledgement + +This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resource. + +# References + +[1] Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-aware image-text alignment for opinion-unaware image quality assessment. arXiv preprint arXiv:2403.11176, 2024. 7 +[2] Hrishav Bakul Barua, Kalin Stefanov, KokSheik Wong, Abhinav Dhall, and Ganesh Krishnasamy. GTA-HDR: A large-scale synthetic dataset for HDR image reconstruction. arXiv preprint arXiv:2403.17837, 2024. 2 +[3] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE TIP, 2012. 6 +[4] Sebastian Dille, Chris Careaga, and Yaqiz Aksoy. Intrinsic single-image hdr reconstruction. In ECCV, 2024. 1, 6 +[5] Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafal K Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 1, 2 +[6] Yuki Endo, Yoshihiro Kanamori, and Jun Mitani. Deep reverse tone mapping. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 2, 6 +[7] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In CVPR, 2023. 2 +[8] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), 2021. 7, 8 +[9] Michael Gharbi, Jiawen Chen, Jonathan Barron, Samuel Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2017. 2 +[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2 +[11] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024. 2 +[12] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2 +[13] Jin Han, Chu Zhou, Peiqi Duan, Yehui Tang, Chang Xu, Chao Xu, Tiejun Huang, and Boxin Shi. Neuromorphic camera guided high dynamic range imaging. In CVPR, 2020. 2, 4 +[14] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 4, 6, 7, 8 +[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a + +two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3 +[17] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for hdr deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In Proceedings of the European Conference on Computer Vision (ECCV). ECCV, 2024. 4 +[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[20] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In CVPR, 2018. 2 +[21] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2 +[22] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arxiv, 2024. 1, 2, 4, 6, 7 +[23] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-BIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2024. 2, 4 +[24] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 1, 2, 5, 6, 7 +[25] M Ronnier Luo, Guihua Cui, and Bryan Rigg. The development of the cie 2000 colour-difference formula: Ciede2000. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Français de la Couleur, 26(5):340-350, 2001. 6 +[26] Nico Messikommer, Stamatios Georgoulis, Daniel Gehrig, Stepan Tulyakov, Julius Erbach, Alfredo Bochicchio, Yuanyou Li, and Davide Scaramuzza. Multi-bracket high dynamic range imaging with event cameras. In CVPR, 2022. 2 +[27] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6, 7 +[28] Mohammad Mostafavi, Lin Wang, and Kuk-Jin Yoon. Learning to reconstruct hdr images from events, with applications to depth and flow prediction. IJCV, 2021. 1, 2 +[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and + +Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2 +[30] Yohan Poirier-Ginter and Jean-François Lalonde. Robust unsupervised stylegan image restoration. In CVPR, 2023. 2 +[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: An open event camera simulator. In Proc. of Conference on Robotics Learning, 2018. 6 +[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 1, 2, 3 +[33] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 2, 3 +[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4 +[35] Richard Shaw, Sibi Catley-Chandar, Ales Leonardis, and Eduardo Pérez-Pellitero. HDR reconstruction from bracketed exposures and events. In BMVC, 2022. 2 +[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5 +[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 2 +[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2 +[40] Chao Wang, Ana Serrano, Xingang Pan, Bin Chen, Karol Myszkowski, Hans-Peter Seidel, Christian Theobalt, and Thomas Leimkuhler. GlowGAN: Unsupervised learning of HDR images from LDR images in the wild. In ICCV, 2023. 2 +[41] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In CVPR, 2019. 2 +[42] Zejia Weng, Xitong Yang, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Genrec: Unifying video generation and recognition with diffusion models. arXiv preprint arXiv:2408.15241, 2024. 2 +[43] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2, 4 +[44] Wenhan Wu, Xian Liu, and Jie Chen. Image restoration with controlnet. IEEE Transactions on Image Processing, 2023. 4 +[45] Garima Yadav, Saurabh Maheshwari, and Anjali Agarwal. Contrast limited adaptive histogram equalization based enhancement for real time video system. In 2014 International Conference on Advances in Computing, Communications and Informatics (ICACCI), 2014. 5, 6 + +[46] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2 +[47] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Mania: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 7 +[48] Yixin Yang, Jin Han, Jinxiu Liang, Imari Sato, and Boxin Shi. Learning event guided high dynamic range video reconstruction. In CVPR, 2023. 1, 2, 4, 5, 6, 7, 8 +[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2 +[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5 +[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6 +[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 3 +[53] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, 2021. 1, 2, 6, 7 \ No newline at end of file diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/images.zip b/2025/Event-guided HDR Reconstruction with Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..0e5fd757f9c9f70232c06ccfebba7bbe518bb194 --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b64d7643db6da0226bdb45c25d6bb42cd6da49df0526c82d0d50b6412e9852c4 +size 918011 diff --git a/2025/Event-guided HDR Reconstruction with Diffusion Priors/layout.json b/2025/Event-guided HDR Reconstruction with Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..e349c39f3fe95e388af19323e93901c69108cc4f --- /dev/null +++ b/2025/Event-guided HDR Reconstruction with Diffusion Priors/layout.json @@ -0,0 +1,10039 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "spans": [ + { + "bbox": [ + 130, + 103, + 481, + 121 + ], + "type": "text", + "content": "Event-guided HDR Reconstruction with Diffusion Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "spans": [ + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "text", + "content": "Yixin Yang" + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "inline_equation", + "content": "^{1,2,\\dagger}" + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "text", + "content": " Jiawei Zhang" + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "text", + "content": " Yang Zhang" + }, + { + "bbox": [ + 187, + 142, + 416, + 157 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "spans": [ + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "text", + "content": "Yunxuan Wei" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "text", + "content": " Dongqing Zou" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "text", + "content": " Jimmy S. Ren" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "inline_equation", + "content": "^{3,5}" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 146, + 157, + 462, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 71, + 171, + 539, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 171, + 539, + 186 + ], + "spans": [ + { + "bbox": [ + 71, + 171, + 539, + 186 + ], + "type": "text", + "content": "1 State Key Lab of Multimedia Info. Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 70, + 186, + 539, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 186, + 539, + 199 + ], + "spans": [ + { + "bbox": [ + 70, + 186, + 539, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 70, + 186, + 539, + 199 + ], + "type": "text", + "content": " Nat'l Eng. Research Ctr. of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "spans": [ + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "text", + "content": " SenseTime Research " + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "text", + "content": "PBVR " + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 135, + 199, + 474, + 213 + ], + "type": "text", + "content": "Hong Kong Metropolitan University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 215, + 494, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 215, + 494, + 228 + ], + "spans": [ + { + "bbox": [ + 115, + 215, + 494, + 228 + ], + "type": "text", + "content": "{yangyixin93, shiboxin}@pku.edu.cn, Github Page: github.com/YixinYang-00/HDRev-Diff" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 270, + 296, + 546 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 296, + 546 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 296, + 546 + ], + "type": "text", + "content": "Events provide High Dynamic Range (HDR) intensity change which can guide Low Dynamic Range (LDR) image for HDR reconstruction. However, events only provide temporal intensity differences and it is still ill-posed in over-/under-exposed areas due to missing initial reference brightness and color information. With strong generation ability, diffusion models have shown their potential for tackling ill-posed problems. Therefore, we introduce conditional diffusion models to hallucinate missing information. Whereas, directly adopting events and LDR image as conditions is complicated for diffusion models to sufficiently utilize their information. Thus we introduce a pretrained events-image encoder tailored for HDR reconstruction and a pyramid fusion module to provide HDR conditions, which can be efficiently and effectively utilized by the diffusion model. Moreover, the generation results of diffusion models usually exhibit distortion, particularly for fine-grained details. To better preserve fidelity and suppress distortion, we propose a fine-grained detail recovery approach using a histogram-based structural loss. Experiments on real and synthetic data show the effectiveness of the proposed method in terms of both detail preservation and information hallucination." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 568, + 135, + 581 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 135, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 135, + 581 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 589, + 296, + 685 + ], + "type": "text", + "content": "Extending the dynamic range of an image (usually with low dynamic range, LDR) to record the brightness of real scenes plausibly is particularly useful for daily photography and computer vision tasks. High dynamic range (HDR) imaging can be, in the most direct way, achieved by utilizing neural networks to hallucinate missing information in over-/under-exposed areas [4, 5, 22, 24], as shown in Figure 1 (a), which is challenging to faithfully recover" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 244, + 391, + 319 + ], + "blocks": [ + { + "bbox": [ + 317, + 244, + 391, + 319 + ], + "lines": [ + { + "bbox": [ + 317, + 244, + 391, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 244, + 391, + 319 + ], + "type": "image", + "image_path": "62146720e129dfa968c35c47d423c59e5f4698ccedfafb5fdf21f9a824d6bd02.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 319, + 380, + 330 + ], + "lines": [ + { + "bbox": [ + 328, + 319, + 380, + 330 + ], + "spans": [ + { + "bbox": [ + 328, + 319, + 380, + 330 + ], + "type": "text", + "content": "(a) Dille et. al." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 392, + 244, + 468, + 319 + ], + "blocks": [ + { + "bbox": [ + 392, + 244, + 468, + 319 + ], + "lines": [ + { + "bbox": [ + 392, + 244, + 468, + 319 + ], + "spans": [ + { + "bbox": [ + 392, + 244, + 468, + 319 + ], + "type": "image", + "image_path": "8c51f95e84dc932d2802ce39488b0639c37b04c35a729cfbf67c117216778dd0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 410, + 319, + 450, + 330 + ], + "lines": [ + { + "bbox": [ + 410, + 319, + 450, + 330 + ], + "spans": [ + { + "bbox": [ + 410, + 319, + 450, + 330 + ], + "type": "text", + "content": "(b) Events" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 479, + 244, + 553, + 319 + ], + "blocks": [ + { + "bbox": [ + 479, + 244, + 553, + 319 + ], + "lines": [ + { + "bbox": [ + 479, + 244, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 479, + 244, + 553, + 319 + ], + "type": "image", + "image_path": "6f2c90e6df82210e9bd3da228d4f921e98a1810a649abd06611c19768d0947ea.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 488, + 319, + 541, + 330 + ], + "lines": [ + { + "bbox": [ + 488, + 319, + 541, + 330 + ], + "spans": [ + { + "bbox": [ + 488, + 319, + 541, + 330 + ], + "type": "text", + "content": "(c) EventHDR" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 317, + 333, + 391, + 411 + ], + "blocks": [ + { + "bbox": [ + 317, + 333, + 391, + 411 + ], + "lines": [ + { + "bbox": [ + 317, + 333, + 391, + 411 + ], + "spans": [ + { + "bbox": [ + 317, + 333, + 391, + 411 + ], + "type": "image", + "image_path": "1ae2c101566f5800d0e01250561b958a7c4bd48811ea929a68dd66440aa727a0.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 412, + 382, + 423 + ], + "lines": [ + { + "bbox": [ + 326, + 412, + 382, + 423 + ], + "spans": [ + { + "bbox": [ + 326, + 412, + 382, + 423 + ], + "type": "text", + "content": "(d) LDR image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 392, + 333, + 476, + 429 + ], + "blocks": [ + { + "bbox": [ + 392, + 333, + 476, + 429 + ], + "lines": [ + { + "bbox": [ + 392, + 333, + 476, + 429 + ], + "spans": [ + { + "bbox": [ + 392, + 333, + 476, + 429 + ], + "type": "image", + "image_path": "9e79e0012e1eaa40b5ac7c9a54f1fec3426ac3f4b01ec6d8e95ee81e3e1f684c.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 479, + 333, + 553, + 411 + ], + "blocks": [ + { + "bbox": [ + 479, + 333, + 553, + 411 + ], + "lines": [ + { + "bbox": [ + 479, + 333, + 553, + 411 + ], + "spans": [ + { + "bbox": [ + 479, + 333, + 553, + 411 + ], + "type": "image", + "image_path": "a8d7b7f26b188735f773ee416cfae1762946492602decfa1771b4d41dba7fe87.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 412, + 536, + 423 + ], + "lines": [ + { + "bbox": [ + 495, + 412, + 536, + 423 + ], + "spans": [ + { + "bbox": [ + 495, + 412, + 536, + 423 + ], + "type": "text", + "content": "(e) HDRev" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 317, + 429, + 391, + 503 + ], + "blocks": [ + { + "bbox": [ + 317, + 429, + 391, + 503 + ], + "lines": [ + { + "bbox": [ + 317, + 429, + 391, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 429, + 391, + 503 + ], + "type": "image", + "image_path": "dfd498d74bd0f46b0354755a13961313ae0a3d4b1bc598eb7c55aa2a43b8e4e1.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 505, + 371, + 516 + ], + "lines": [ + { + "bbox": [ + 337, + 505, + 371, + 516 + ], + "spans": [ + { + "bbox": [ + 337, + 505, + 371, + 516 + ], + "type": "text", + "content": "(f) Sagiri" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "Figure 1. HDR image can be recovered by only using a single LDR image (d) as input (e.g., Dille et al. [4]), which is a difficult problem as shown in (a). With only events (b) as input, event-to-image reconstruction (e.g., EventHDR [53]) (c) recovers an HDR scene without color. Event-guided HDR methods (e.g., HDRev [48]) (e) take both events and the LDR image as input, which recovers colorful HDR scenes with moderate details. Diffusion-based methods (e.g., Sagiri [22]) (f) take the LDR image (and optionally the text prompt) as input showing visually pleasing results but may have challenges in maintaining consistent contents with the original scene. The proposed method (g) takes events and the LDR image as input and leverages the diffusion priors to recover a plausible and realistic HDR image with rich details and less distortion. scene details without any auxiliary information. Neuromorphic cameras, such as event cameras, are proven to be helpful in providing additional details for HDR purpose [28, 32, 53]. Event cameras capture intensity changes" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 392, + 429, + 468, + 503 + ], + "blocks": [ + { + "bbox": [ + 392, + 429, + 468, + 503 + ], + "lines": [ + { + "bbox": [ + 392, + 429, + 468, + 503 + ], + "spans": [ + { + "bbox": [ + 392, + 429, + 468, + 503 + ], + "type": "image", + "image_path": "ce99c953bdb05d0ee04ced2ae08a814ced006d2ba1e7a2af30c8bb62749a0928.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 414, + 505, + 446, + 516 + ], + "lines": [ + { + "bbox": [ + 414, + 505, + 446, + 516 + ], + "spans": [ + { + "bbox": [ + 414, + 505, + 446, + 516 + ], + "type": "text", + "content": "(g) Ours" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "bbox": [ + 475, + 437, + 551, + 497 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 475, + 437, + 515, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 437, + 515, + 447 + ], + "spans": [ + { + "bbox": [ + 475, + 437, + 515, + 447 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 475, + 437, + 515, + 447 + ], + "type": "text", + "content": " : Events" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 475, + 448, + 544, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 448, + 544, + 456 + ], + "spans": [ + { + "bbox": [ + 475, + 448, + 544, + 456 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 475, + 448, + 544, + 456 + ], + "type": "text", + "content": " :Colorful results" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 475, + 457, + 539, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 457, + 539, + 467 + ], + "spans": [ + { + "bbox": [ + 475, + 457, + 539, + 467 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 475, + 457, + 539, + 467 + ], + "type": "text", + "content": " : More realistic" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 475, + 468, + 543, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 468, + 543, + 477 + ], + "spans": [ + { + "bbox": [ + 475, + 468, + 543, + 477 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 475, + 468, + 543, + 477 + ], + "type": "text", + "content": " : Less distortion" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 476, + 478, + 536, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 476, + 478, + 536, + 487 + ], + "spans": [ + { + "bbox": [ + 476, + 478, + 536, + 487 + ], + "type": "text", + "content": ": CNN models" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 475, + 487, + 551, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 475, + 487, + 551, + 497 + ], + "spans": [ + { + "bbox": [ + 475, + 487, + 551, + 497 + ], + "type": "text", + "content": "Diffusion models" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "spans": [ + { + "bbox": [ + 66, + 693, + 259, + 703 + ], + "type": "text", + "content": "This work is done during Yixin's internship at SenseTime." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "* Corresponding author." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11787" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 156 + ], + "type": "text", + "content": "by triggering events asynchronously, which inherently contains HDR information (e.g., 120 dB for DAVIS340), and events can be directly converted to HDR images, as demonstrated by EventHDR [53] in Figure 1 (c). The event-to-image reconstruction [28, 32, 53] can recover details in over-/under-exposed regions, but fail to provide colorful results due to lacking color information." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 158, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 326 + ], + "type": "text", + "content": "To generate colorful HDR images, event-guided HDR reconstructions [13, 14, 48] take both LDR image and events as inputs, complementarily learning to predict color and intensity in over-/under-exposed areas, as demonstrated in Figure 1 (e). They can reconstruct HDR information, e.g., the cloud in the green box, and provide colorful results. Although regression models are adopted to build the relationship between the predicted pixel values and its neighboring pixels, those methods cannot recover proper intensity, especially when the LDR image has large over-/under-exposed areas. Because absolute HDR intensity (events only contain thresholded radiance changes) and color information are still missing in these regions, it is difficult for regression models to appropriately hallucinate their values." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 484 + ], + "type": "text", + "content": "Diffusion models [16, 37-39] are emerging generative models that show great potential for generating plausible and high-quality images. Even if for ill-posed problems, they can provide visually pleasing results thanks to the expressive power of diffusion priors. Diffusion models have also been adopted for image restoration [23, 43, 49]. Recent attempts, e.g., Sagiri [22], have shown the potential of diffusion priors for HDR reconstruction by providing more visually realistic details in over-exposed areas, e.g., the cloud in the sky of Figure 1 (f). However, the generated HDR images still cannot appropriately keep the fidelity of the original scene, and additional distortion are brought into well-exposed regions, as shown by the red box in Figure 1 (f)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 295, + 713 + ], + "type": "text", + "content": "To reconstruct visually pleasing and faithful HDR images with less distortion, we introduce events to provide differential HDR intensity and adopt conditional diffusion models [50] to compensate for the ill-posed nature of hallucinating missing information and constrain the generative model to be consistent with the original scene. Specially, we focus on solving the following two key problems via employing diffusion models in event-guided HDR reconstruction: 1) Condition extraction: Since events and LDR image are two different modalities of visual representation, directly adopting them as conditions is non-compatible for the conditional diffusion models to conduct effective fusion. Inspired by the modality alignment of HDRev [48], we introduce an event-image encoder and a pyramid fusion module to fuse events and LDR image in different resolutions and provide conditions to guide the generation process. 2) Details preservation: Applying diffusion models to HDR reconstruction [22, 23] suffers from fine-grained detail distortion, as shown in Figure 1 (f). Such artifacts become" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 167 + ], + "type": "text", + "content": "more obvious when LDR images are high quality and have large well-exposed areas. To refine the fine-grained details and make them consistent with LDR images and events, we design a refinement module with histogram-based structure loss. The proposed method is validated on both synthetic and real data and outperforms existing methods by providing fidelity-preserving HDR reconstruction, as illustrated in Figure 1 (g), with the following technical contributions:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 169, + 553, + 239 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 169, + 553, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 169, + 553, + 191 + ], + "spans": [ + { + "bbox": [ + 314, + 169, + 553, + 191 + ], + "type": "text", + "content": "- integrating events and the conditional diffusion models to recover missing information faithfully;" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 216 + ], + "type": "text", + "content": "- adopting the pretrained event-image encoder and pyramid fusion to effectively apply conditions; and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 239 + ], + "type": "text", + "content": "- designing the refinement module and histogram-based structure loss to further strengthen fine-grained details." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 253, + 404, + 265 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 253, + 404, + 265 + ], + "spans": [ + { + "bbox": [ + 314, + 253, + 404, + 265 + ], + "type": "text", + "content": "2. Related works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 268, + 553, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 553, + 399 + ], + "type": "text", + "content": "Image-based HDR reconstruction Many approaches leverage the capabilities of neural networks to directly hallucinate HDR information from training data, such as CNNs [5, 6, 9], Generative Adversarial Networks (GANs) [40], and diffusion models [7, 22]. Liu et al. [24] incorporates the LDR image formulation pipeline into its hallucination process to perform step-by-step prediction. For a more comprehensive view, we refer to GTA-HDR [2]. Since information is unavoidably missing in over-/underexposed areas, the above methods only rely on hallucination, which may not correspond to real scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "content": "Event-based HDR reconstruction Benefiting from the HDR property of event cameras, previous works reconstruct HDR images from events with recurrent neural networks [32, 33, 53] or GANs [28, 41]. Since events only record logarithmic irradiance changes, it's difficult to reconstruct color images. To reconstruct color HDR images, Han et al. [13, 14] adopted a hybrid camera system to capture events and frame, enabling event-guided HDR reconstruction. Yang et al. [48] treated events and frame as different modalities and proposed multi-modalities alignment and fusion to provide HDR videos. Other methods [26, 35] incorporate events and bracketed exposure for HDR reconstruction, while long-exposed images bring blurry artifacts. The proposed method takes events and LDR image as input, incorporating diffusion models to compensate for realistic and faithful information in over-/under-exposed areas." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 713 + ], + "type": "text", + "content": "Generative priors Generative Priors capture the underlying data distribution, enabling the generated images to follow real-world image distribution. GAN [10] has shown its potential to leverage generative priors for image restoration [20, 21, 30]. Recently, diffusion models [16, 37-39] have widely adopted as effective generative priors for producing realistic images [29, 34, 42, 50], which also demonstrate strong potential for image restorations [23, 43, 49], specifically, for HDR image reconstruction from bracketed exposure [12, 17, 46]. There are some attempts [11, 22] to" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11788" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 220 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 220 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 220 + ], + "type": "image", + "image_path": "48d481cd855603f94ca7c593645d5fe45d625f9a99def7ac4b0b1f6b7bb25e1c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "type": "text", + "content": "Figure 2. Pipeline of the proposed method consists of three parts: diffusion model and denoising process (green part), event-image conditioning (blue part), and detail refinement (orange part). The blue part fuses LDR image and events to generate injected features for Denoiser as described in Section 3.2. The green part adopts the injected features as conditions to estimate noise from latent " + }, + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "type": "inline_equation", + "content": "z_{t}" + }, + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "type": "text", + "content": " step by step with fixed pretrained parameters. To refine the distortion of denoised images, the refinement module in the orange part, whose architecture is shown on the right, is proposed to leverage the HDR features and denoised latent " + }, + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 54, + 221, + 555, + 299 + ], + "type": "text", + "content": ". Considering the uncertainty of diffusion results in color and brightness, we introduce structure loss, which adopts adaptive Ground Truth (GT) as the supervision target, as described in Section 3.3. After refinement, we obtain a plausible and realistic refined image without distortion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 304, + 294, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 294, + 340 + ], + "type": "text", + "content": "take advantage of diffusion models to generate HDR images from single LDR image, while it cannot provide faithful reconstruction due to the inevitable dynamic clipping." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 350, + 162, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 162, + 363 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 162, + 363 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 295, + 539 + ], + "type": "text", + "content": "Section 3.1 introduces basic concepts about diffusion models and event cameras. The pipeline of the proposed method is illustrated in Figure 2. The green part is the pretrained latent diffusion model, which performs denoising in latent space step by step with the injected feature from the blue part. The blue part is the proposed event-guided conditioning and generation process, which adopts an event-image encoder and a pyramid fusion to utilize events and the LDR image efficiently and effectively, as described in Section 3.2. The orange part is the fine-grained detail enhancement based on the Variant Auto Encoder (VAE) decoder, which learns to refine fine-grained details by the proposed histogram-based structure loss as described in Section 3.3. Training details are described in Section 3.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 135, + 558 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "content": "Diffusion models Diffusion models [16, 34, 38] can generate realistic high-quality images by diffusing and denoising processes. Latent diffusion models [34] are a type of diffusion models that perform those processes on latent space. During training, the latent diffusion models project the clean image into the latent space with a VAE encoder as " + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "content": ", then diffuse " + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "content": " by adding Gaussian noise for " + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "content": " steps to get the noisy latent " + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "inline_equation", + "content": "z_n" + }, + { + "bbox": [ + 55, + 562, + 296, + 659 + ], + "type": "text", + "content": ". Adding noise in one step is as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "spans": [ + { + "bbox": [ + 107, + 667, + 295, + 681 + ], + "type": "interline_equation", + "content": "z _ {n} = \\sqrt {a _ {n}} z _ {n - 1} + \\sqrt {1 - a _ {n}} \\epsilon_ {n - 1}, \\tag {1}", + "image_path": "65441916c27e5519585187d13ed5a40950f67ed6acde140cc73863da4b0114f0.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "n \\in [1, N]" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " means the " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "-th diffusion step, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the total diffusion steps, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_n \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{I})" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is the added noise, " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "a_n" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "text", + "content": "the pre-defined parameter of the noise scheduler. Defining " + }, + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "inline_equation", + "content": "\\bar{a}_n = \\prod_{i=1}^n a_i" + }, + { + "bbox": [ + 313, + 304, + 555, + 329 + ], + "type": "text", + "content": ", the closed form equation of Eq. (1) [16] is:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 379, + 333, + 553, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 333, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 379, + 333, + 553, + 346 + ], + "type": "interline_equation", + "content": "z _ {n} = \\sqrt {\\bar {a} _ {n}} z _ {0} + \\sqrt {1 - \\bar {a} _ {n}} \\epsilon , \\tag {2}", + "image_path": "1a9924d6f62414dc7534c354c1419c4256adddbfd59ec21b5812185fca2c7ffc.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": ". The denoising process trains a denoiser " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " to estimate the noise " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " usually by minimizing Mean Square Error (MSE) " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{n,z_0,\\epsilon} = \\| \\epsilon -\\epsilon_{\\theta}(z_n,n)\\|" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": ". As the variable " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "z_{n}" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " will be an approximately standard Gaussian distribution when " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " is large enough, during inference, the denoising process often starts from an i.i.d. noise " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "z_{N}" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " and generates a clean image " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "z_{0}" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " using " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " over " + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 351, + 555, + 436 + ], + "type": "text", + "content": " steps." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": "Event triggering and stacking Event cameras trigger an event " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "(t,x,y,p)" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " when illuminance changes of pixel " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " in the logarithm domain over a predefined threshold " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " in time " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": ", where the polarity " + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "inline_equation", + "content": "p\\in \\{-1,1\\}" + }, + { + "bbox": [ + 313, + 437, + 554, + 510 + ], + "type": "text", + "content": " indicates the decrease and increase of the illuminance. Mathematically, an event is generated when the following inequation occurs:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 365, + 515, + 553, + 529 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 365, + 515, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 365, + 515, + 553, + 529 + ], + "type": "interline_equation", + "content": "\\left\\| \\log I _ {x, y} (t) - \\log I _ {x, y} \\left(t _ {\\text {r e f}}\\right) \\right\\| \\geq \\eta , \\tag {3}", + "image_path": "2c23a42f6f72c26f737422257fcddafe813a40cf589b2e4a4079f21f65fde64f.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "I_{x,y}(t)" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " is the illuminance at time " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "I_{x,y}(t_{\\mathrm{ref}})" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " is the reference illuminance level, " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "t_\\mathrm{ref}" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " is the last event triggered time of pixel " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " is a pre-defined event threshold. We convert the stream-like events to tensors using voxel grid [52], which encodes temporal information in a C-channel 3D volume by discretizing event timestamps into " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " temporal bins. Following Rebecq et al. [32, 33], " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": "-th event " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "(t_k,x_k,y_k,p_k)" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " distributes its polarity " + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "inline_equation", + "content": "p_k" + }, + { + "bbox": [ + 313, + 533, + 555, + 641 + ], + "type": "text", + "content": " to the two closest temporal bins related to its normalized timestamp by:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 345, + 646, + 553, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 646, + 553, + 671 + ], + "spans": [ + { + "bbox": [ + 345, + 646, + 553, + 671 + ], + "type": "interline_equation", + "content": "E _ {j} ^ {x, y} = \\sum_ {x _ {k} = x, y _ {k} = y} p _ {k} \\max (0, 1 - | \\tilde {t} _ {k} - j |), \\tag {4}", + "image_path": "0bf34083e8ac47e5e7e46b22f751bc5f728f13268c100d50d0917e6dd6603995.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{t}_k = \\frac{C - 1}{\\Delta t} (t_k - t_0)" + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "content": " is the normalized timestamp, " + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\Delta_t = \\max_k(t_k) - t_0" + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "content": " is the time span of events, " + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "inline_equation", + "content": "t_0" + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "content": " is the start timestamp, and " + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "inline_equation", + "content": "j \\in [0, C - 1]" + }, + { + "bbox": [ + 313, + 676, + 554, + 714 + ], + "type": "text", + "content": " is the index of the bin." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11789" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 208, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 208, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 208, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 208, + 178 + ], + "type": "image", + "image_path": "26d7d1d70d68d9dedeb95de0f48359d962b474bc969b0c535093b34a0c8c15cc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 180, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 54, + 180, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 555, + 226 + ], + "type": "text", + "content": "Figure 3. Pipeline of the event-guided conditioning process, with comparison to conditioning using HDRev [48] restored image. Directly concatenating events and LDR image cannot fully utilize events, thus we introduce an event-image encoder to fuse them as shown in the orange part. The HDRev [48] restored images by decoding HDR features can serve as input to the embedding module, as shown in the yellow part. The proposed pyramid fusion module, as shown in the blue part, fuses different levels HDR features to provide conditions." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 208, + 70, + 352, + 178 + ], + "blocks": [ + { + "bbox": [ + 208, + 70, + 352, + 178 + ], + "lines": [ + { + "bbox": [ + 208, + 70, + 352, + 178 + ], + "spans": [ + { + "bbox": [ + 208, + 70, + 352, + 178 + ], + "type": "image", + "image_path": "b044f0e8017dfb15c9ff581463b50135748688fc6798c14fa2b0279ffb810a06.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 352, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 352, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 352, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 352, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "8bf7ba3d618c03b8f712d7598c448f854c0f5a1140c77866026f4ca6a04cff98.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 230, + 282, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 230, + 282, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 230, + 282, + 243 + ], + "type": "text", + "content": "3.2. Event-guided Conditioning and Generation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 247, + 296, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 296, + 402 + ], + "type": "text", + "content": "To hallucinate missing information for HDR reconstruction, previous methods mainly come in two ways: adopting better hallucination models (e.g., diffusion models [22]) or introducing HDR sensors (e.g., event cameras [13, 14, 48]). Diffusion models generate high-quality and realistic images by modeling real-world image distributions with diffusion priors. Event cameras provide differential HDR intensity, making events well-suited as conditions for diffusion models to generate more accurate details in over-/under-exposed areas, thereby compensating the dynamic range of LDR images. Thus we propose to leverage both event cameras and diffusion models to hallucinate realistic HDR images to be more consistent with real scenes." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "text", + "content": "To generate images corresponding to conditions, conditional models [18, 50] are proposed to generate features injected into specific layers of the well-trained denoiser. ControlNet [50] is one of the widelyconditional models for image restoration [23, 43, 44], which generates and injects features for each down-sample and middle layers of the denoiser. As illustrated in the interaction between the blue and green parts of Figure 2, we perform feature injection similar to ControlNet. In the green part, we utilize Stable Diffusion [34] V1.5 as the pretrained denoiser " + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "text", + "content": " to denoise from the initial latent " + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "inline_equation", + "content": "z_{N}" + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "text", + "content": " step by step. The generated image is obtained by decoding the estimated initial latent " + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 403, + 296, + 605 + ], + "type": "text", + "content": " with VAE decoder. With an image-like input, ControlNet employs an embedding module to generate conditions, then adopts a control module to obtain injected features. However, it is challenging for the embedding module to effectively fuse LDR image with events and extract useful information." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "text", + "content": "Events record temporal intensity changes, making them distinct from LDR images. Directly concatenating stacked events " + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "text", + "content": " and LDR images " + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LDR}}" + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "text", + "content": " together as the input for the embedding module " + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{con}}" + }, + { + "bbox": [ + 55, + 606, + 296, + 667 + ], + "type": "text", + "content": " introduces HDR information, where the conditioning process is:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 99, + 671, + 295, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 671, + 295, + 685 + ], + "spans": [ + { + "bbox": [ + 99, + 671, + 295, + 685 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {C o n C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {c o n}} \\left(\\left[ I _ {\\text {L D R}}, E \\right]\\right), \\tag {5}", + "image_path": "42f53f74802c816d21930b6b182ae56f2cae3e5fc512f89db6b1487338c9a277.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "[.,] denotes concatenation. But it cannot fully utilize HDR information encoded in events, e.g., decoration and cable" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 386 + ], + "type": "text", + "content": "in over-exposed areas cannot be accurately reconstructed in Figure 4 (d). Therefore, a specifically designed HDR information extraction model is needed. HDRev [48] performs event-guided HDR reconstruction through representation alignment and feature fusion, which addresses the modality gap between events and LDR images. It performs representation alignment through pre-training to obtain modality-specific encoders as shown in the orange part of Figure 3. The LDR image and events are encoded into modality-specific features, represented by green features (LDR image) and yellow features (events), respectively. Those two features are fused through the fusion module to obtain HDR features. The restoration process can be written into:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 379, + 395, + 555, + 407 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 395, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 379, + 395, + 555, + 407 + ], + "type": "interline_equation", + "content": "I _ {\\text {r e s t o r e d}} = \\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right), \\tag {6}", + "image_path": "90dc322b5cb086020c7bacb824e61e93b5f3175ca7c324cda9bcc700b4afecb2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " are the decoder and encoder in HDRev [48]. The restoration process utilizes events to recover more details, as shown by Figure 4 (e), which offers a better reconstruction of the decoration and cable than (d). The intuitive way to better utilize events for conditioning is adopting " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{restored}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " as input for the embedding module " + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{rest}}" + }, + { + "bbox": [ + 313, + 415, + 554, + 499 + ], + "type": "text", + "content": " as shown in the yellow part of Figure 3. The conditioning process1 is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 345, + 507, + 555, + 520 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 507, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 345, + 507, + 555, + 520 + ], + "type": "interline_equation", + "content": "\\mathcal {E} _ {\\text {R e s t C o n d}} \\left(I _ {\\text {L D R}}, E\\right) = \\mathcal {T} _ {\\text {r e s t}} \\left(\\mathcal {B} \\left(\\mathcal {H} \\left(I _ {\\text {L D R}}, E\\right)\\right)\\right). \\tag {7}", + "image_path": "7482a2beb7d33dfaf9fe7905118615648c1588bb780c73d845134b9f4a8b755a.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 529, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 588 + ], + "type": "text", + "content": "Diffusion models improve image quality as shown by the improved shape and color of the table in Figure 4 (f) compared to (e). Adopting " + }, + { + "bbox": [ + 313, + 529, + 554, + 588 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{restored}}" + }, + { + "bbox": [ + 313, + 529, + 554, + 588 + ], + "type": "text", + "content": " as conditions (\"RestCond\") better utilizes events and provides more details than \"Con-Cond\", e.g., the decoration in Figure 4 (f) is better than (d)." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "content": "Although HDRev is an encoder-decoder architecture, the fusion between LDR image and events is performed only in the encoder. Applying " + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{restored}}" + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "content": " as conditions not only suffers from a decoder-encoder redundant computation " + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{rest}}(\\mathcal{B}(\\cdot))" + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "content": " as shown in the yellow part of Figure 3, but also fails to reconstruct details from events, e.g., the shape of table in the red box of Figure 4, and is difficult to adjust the proper brightness of over-exposed areas, e.g., the cloud in Figure 1 (e). The embedding module " + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_{\\mathrm{rest}}" + }, + { + "bbox": [ + 313, + 589, + 555, + 696 + ], + "type": "text", + "content": " may misunderstand" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 515, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 515, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 515, + 713 + ], + "type": "text", + "content": "1 More details are provided in the supplementary materials." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11790" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 53, + 125, + 151 + ], + "blocks": [ + { + "bbox": [ + 58, + 53, + 125, + 151 + ], + "lines": [ + { + "bbox": [ + 58, + 53, + 125, + 151 + ], + "spans": [ + { + "bbox": [ + 58, + 53, + 125, + 151 + ], + "type": "image", + "image_path": "fbe39092ac18e63fb7979b084c7aa0bd69decec4bbd5a69fc9070d9d7e63707e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 77, + 152, + 106, + 161 + ], + "lines": [ + { + "bbox": [ + 77, + 152, + 106, + 161 + ], + "spans": [ + { + "bbox": [ + 77, + 152, + 106, + 161 + ], + "type": "text", + "content": "(a) LDR" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 129, + 53, + 197, + 151 + ], + "blocks": [ + { + "bbox": [ + 129, + 53, + 197, + 151 + ], + "lines": [ + { + "bbox": [ + 129, + 53, + 197, + 151 + ], + "spans": [ + { + "bbox": [ + 129, + 53, + 197, + 151 + ], + "type": "image", + "image_path": "de5214b870f828a3a944b400bfdade9305f5cfb4659b4df235384bf0022ea3dc.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 152, + 182, + 161 + ], + "lines": [ + { + "bbox": [ + 143, + 152, + 182, + 161 + ], + "spans": [ + { + "bbox": [ + 143, + 152, + 182, + 161 + ], + "type": "text", + "content": "(b) Events" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 201, + 53, + 268, + 151 + ], + "blocks": [ + { + "bbox": [ + 201, + 53, + 268, + 151 + ], + "lines": [ + { + "bbox": [ + 201, + 53, + 268, + 151 + ], + "spans": [ + { + "bbox": [ + 201, + 53, + 268, + 151 + ], + "type": "image", + "image_path": "693b024506a71d8b6f956f9215074ea3e6ba8cec76ec9038e176069b3f5ae773.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 204, + 152, + 261, + 161 + ], + "lines": [ + { + "bbox": [ + 204, + 152, + 261, + 161 + ], + "spans": [ + { + "bbox": [ + 204, + 152, + 261, + 161 + ], + "type": "text", + "content": "(c) GroundTruth" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 163, + 555, + 208 + ], + "lines": [ + { + "bbox": [ + 55, + 163, + 555, + 208 + ], + "spans": [ + { + "bbox": [ + 55, + 163, + 555, + 208 + ], + "type": "text", + "content": "Figure 4. Comparison with different conditioning ways: directly concatenating events and LDR image (\"ConCond\") and HDRev [48] restored image (\"RestCond\"). \"ConCond\" (d) cannot fully utilize events in over-exposed areas to reconstruct cable. Illustrated by \"Rest-Cond\" (f), the unsatisfactory results of decoration and table in \"Restored\" (e) failed to provide HDR information for conditioning in over-exposed areas. The proposed conditioning module can restore missing information as depicted in (g)." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 272, + 53, + 339, + 151 + ], + "blocks": [ + { + "bbox": [ + 272, + 53, + 339, + 151 + ], + "lines": [ + { + "bbox": [ + 272, + 53, + 339, + 151 + ], + "spans": [ + { + "bbox": [ + 272, + 53, + 339, + 151 + ], + "type": "image", + "image_path": "1be280b534db4d0a78dc333062b31f958dd19452ec93a405299693334ef80f20.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 152, + 329, + 161 + ], + "lines": [ + { + "bbox": [ + 282, + 152, + 329, + 161 + ], + "spans": [ + { + "bbox": [ + 282, + 152, + 329, + 161 + ], + "type": "text", + "content": "(d) ConCond" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 343, + 53, + 410, + 151 + ], + "blocks": [ + { + "bbox": [ + 343, + 53, + 410, + 151 + ], + "lines": [ + { + "bbox": [ + 343, + 53, + 410, + 151 + ], + "spans": [ + { + "bbox": [ + 343, + 53, + 410, + 151 + ], + "type": "image", + "image_path": "2f741ba38a4d6d50b3f018e32a3fd7564e8bc0beaeb1e54220ff8b30aa1155b4.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 354, + 152, + 399, + 161 + ], + "lines": [ + { + "bbox": [ + 354, + 152, + 399, + 161 + ], + "spans": [ + { + "bbox": [ + 354, + 152, + 399, + 161 + ], + "type": "text", + "content": "(e) Restored" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 414, + 53, + 482, + 151 + ], + "blocks": [ + { + "bbox": [ + 414, + 53, + 482, + 151 + ], + "lines": [ + { + "bbox": [ + 414, + 53, + 482, + 151 + ], + "spans": [ + { + "bbox": [ + 414, + 53, + 482, + 151 + ], + "type": "image", + "image_path": "02580f47034e2d670432cce1c7136d2190f2b1c5e96f5b4a1b47eab3ef544d82.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 425, + 152, + 471, + 161 + ], + "lines": [ + { + "bbox": [ + 425, + 152, + 471, + 161 + ], + "spans": [ + { + "bbox": [ + 425, + 152, + 471, + 161 + ], + "type": "text", + "content": "(f) RestCond" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 486, + 53, + 553, + 151 + ], + "blocks": [ + { + "bbox": [ + 486, + 53, + 553, + 151 + ], + "lines": [ + { + "bbox": [ + 486, + 53, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 486, + 53, + 553, + 151 + ], + "type": "image", + "image_path": "9e70a1d9dba1fa16735cc8f97e08d05f4953dabd41cb91b72545e91219bf06a5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 152, + 543, + 161 + ], + "lines": [ + { + "bbox": [ + 496, + 152, + 543, + 161 + ], + "spans": [ + { + "bbox": [ + 496, + 152, + 543, + 161 + ], + "type": "text", + "content": "(g) Proposed" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 54, + 213, + 295, + 237 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 213, + 295, + 237 + ], + "spans": [ + { + "bbox": [ + 54, + 213, + 295, + 237 + ], + "type": "text", + "content": "these unsatisfactory areas as shown by the shape of the decoration in the red box of Figure 4 (f)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "text", + "content": "To mitigate issues of poor brightness adjustment in HDRev, and leverage its advantage in event-image fusion, we adopt HDR features from each layer of its event-image encoder to provide conditions. Directly adopting HDR features as conditions is not only more efficient, because of removing the redundant decoder-encoder process, but also more effective, because of avoiding introducing unsatisfactory reconstruction from the decoder " + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "text", + "content": ". However, the HDR features are pyramidal, in which different levels of visual information are encoded separately. To fuse different level HDR features, we propose a pyramid fusion module " + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "text", + "content": " as shown in the blue part of Figure 3, which fuses different level features with down-sample layers and feature concatenation to provide conditions. More accurate details and colors can be recovered with it as shown in Figure 4 (g). The proposed conditioning process " + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 237, + 296, + 441 + ], + "type": "text", + "content": " consists of an event-image encoder and pyramid fusion module, denoted by:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 111, + 446, + 295, + 460 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 446, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 111, + 446, + 295, + 460 + ], + "type": "interline_equation", + "content": "\\mathcal {E} \\left(I _ {\\mathrm {L D R}}, E\\right) = \\mathcal {P} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right)\\right). \\tag {8}", + "image_path": "785fde143b7cfc39abd90387e4e2703d3cf6cf95223d46e37c2c78522381198e.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": "Then the predicted noise is " + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}(z_n,n,\\mathcal{C}(\\mathcal{E}(I_{\\mathrm{LDR}},E))" + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": " is the control module derived from ControlNet [50]. With the pretrained denoiser " + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": ", we are able to train the control module " + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{C}" + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": " and conditioning process " + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 464, + 296, + 525 + ], + "type": "text", + "content": " to hallucinate HDR images by minimizing the noise loss:" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 81, + 529, + 295, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 81, + 529, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 81, + 529, + 295, + 543 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {n, z _ {0}, \\epsilon_ {n}} = \\| \\epsilon_ {n} - \\epsilon_ {\\theta} (z _ {n}, n, \\mathcal {C} (\\mathcal {E} (I _ {\\mathrm {L D R}}, E))) \\| ^ {2}. \\tag {9}", + "image_path": "36b59f10750ffc2b8275f478dbdb04ae045d19ca7443bd45181c1da1c13d0abf.jpg" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 547, + 227, + 559 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 547, + 227, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 547, + 227, + 559 + ], + "type": "text", + "content": "3.3. Fine-grained Detail Refinement" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": "Although diffusion models can reconstruct realistic HDR images, it suffers from distortion, especially for fine-grained details as shown in Figure 1 (f) and Figure 5 \"W/o Refinement\". Therefore, we aim to introduce information from events and LDR image to refine those distortion. Typically, we adopt HDR features from the event-image encoder for refinement. As shown in the orange part of Figure 2, our refinement module is a decoder " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": " in which the up-sample layers are initialized from the VAE decoder. It fuses HDR features " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{H}(I_{\\mathrm{LDR}},E)" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": " with latent " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "z_0" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": " to provide the refined HDR image " + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 563, + 296, + 696 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 123, + 700, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 700, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 123, + 700, + 295, + 714 + ], + "type": "interline_equation", + "content": "H = \\mathcal {D} \\left(\\mathcal {H} \\left(I _ {\\mathrm {L D R}}, E\\right), z _ {0}\\right). \\tag {10}", + "image_path": "e323d7dbb8d8093f9c1f407702c16a91ea5a9e95f1cdd4e98c257366eb3803b1.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "By training the refinement module with the Ground Truth (GT), we can refine some distortion as shown in Figure 5 \"W/o Structure\". Due to the inconsistent color between GT and diffusion results as illustrated in Figure 5, applying GT as a supervision target brings color artifacts similar to regression-based methods [24, 48] as depicted in the red box of Figure 5 \"W/o Structure\". Those artifacts come from the discrepancy between the deterministic GT and the uncertainty in diffusion results during the training process of refinement. Additionally, it also depicts that the details in diffusion results have been wiped (the road curb in the red box) after this refinement. Because it is easier for the refinement to minimize the total loss by reducing the brightness and color difference than the detail difference." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "text", + "content": "Meanwhile, although the diffusion results suffer from distortion, they already have visually pleasant brightness and color consistent with LDR images. Therefore, we tend to maintain the brightness and color properties of diffusion results and only focus on detail refinement by introducing local histogram matching [45] to adjust the GT to serve as a new supervision target. Local histogram matching [45] is able to adjust the pixel intensity distribution to match the histogram of a target image for each local area, which is suitable for adjusting only color and brightness, and maintaining details of GT. Adopting local histogram matching to adjust ground truth image " + }, + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "text", + "content": " to diffusion result " + }, + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{diff}}" + }, + { + "bbox": [ + 313, + 380, + 556, + 525 + ], + "type": "text", + "content": " is as:" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 388, + 534, + 553, + 548 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 534, + 553, + 548 + ], + "spans": [ + { + "bbox": [ + 388, + 534, + 553, + 548 + ], + "type": "interline_equation", + "content": "O _ {\\mathrm {a d j}} = \\operatorname {H i s t} (O, H _ {\\text {d i f f}}). \\tag {11}", + "image_path": "72d94fbada3e5615414e10503e71d7b63d09e5ab9ae45c808154fafa82309ba0.jpg" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 313, + 557, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 593 + ], + "type": "text", + "content": "Our histogram-based structure loss adopts the adjusted image " + }, + { + "bbox": [ + 313, + 557, + 555, + 593 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{adj}}" + }, + { + "bbox": [ + 313, + 557, + 555, + 593 + ], + "type": "text", + "content": " as ground truth and is composed of two losses. The first one is MSE loss:" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 368, + 602, + 553, + 617 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 368, + 602, + 553, + 617 + ], + "spans": [ + { + "bbox": [ + 368, + 602, + 553, + 617 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M S E}} (H, O _ {\\mathrm {a d j}}) = \\| H - O _ {\\mathrm {a d j}} \\| ^ {2}, \\tag {12}", + "image_path": "b3889fa30decb243181e7167ea9a9f4bcd0be88e027824f8d6a938edc00c58a9.jpg" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 626, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 626, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 626, + 555, + 651 + ], + "type": "text", + "content": "and the perceptual loss " + }, + { + "bbox": [ + 313, + 626, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{perc}" + }, + { + "bbox": [ + 313, + 626, + 555, + 651 + ], + "type": "text", + "content": " based on the feature maps extracted by the pretrained VGG-16 [36] network:" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 314, + 659, + 553, + 694 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 553, + 694 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 553, + 694 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p e r c}} = \\sum_ {l} \\| \\phi_ {l} (H) - \\phi_ {l} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2} + \\| \\mathcal {G} _ {l} ^ {\\phi} (H) - \\mathcal {G} _ {l} ^ {\\phi} \\left(O _ {\\text {a d j}}\\right) \\| ^ {2}, \\tag {13}", + "image_path": "61756fe9f47fd3c43d9d3d2a22378a08968f311fcff32bc84c82478b195895f9.jpg" + } + ] + } + ], + "index": 29 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "spans": [ + { + "bbox": [ + 324, + 702, + 545, + 713 + ], + "type": "text", + "content": "2Diverse diffusion results are shown in the supplementary materials." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11791" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 125, + 178 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 125, + 178 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 125, + 178 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 125, + 178 + ], + "type": "image", + "image_path": "091bd60993080a16a587d3cd8e95327b29cb7070d8899032bf3e24a0d1b0cab6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 186, + 555, + 253 + ], + "lines": [ + { + "bbox": [ + 55, + 186, + 555, + 253 + ], + "spans": [ + { + "bbox": [ + 55, + 186, + 555, + 253 + ], + "type": "text", + "content": "Figure 5. Validation for the effectiveness of the structure loss. With only conditioning and diffusion process, the results of \"W/o Refinement\" suffer from distortion as shown in the green box. Training the refinement module with ground truth, some distortions are refined, but introducing unnatural saturation adjustment and unsatisfactory detail prediction as shown in the red box of \"W/o Structure\". We introduce local histogram matching [45] to adjust the brightness and color of ground truth to diffusion results, making the refinement module focus on refining the details. The adapted ground truth is shown in \"Adapted GT\". As shown in \"Proposed\", by adopting adapted ground truth as a supervision target, the refinement module with structure loss refines distortion with more natural colors and details." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 70, + 196, + 178 + ], + "blocks": [ + { + "bbox": [ + 127, + 70, + 196, + 178 + ], + "lines": [ + { + "bbox": [ + 127, + 70, + 196, + 178 + ], + "spans": [ + { + "bbox": [ + 127, + 70, + 196, + 178 + ], + "type": "image", + "image_path": "cbb780887490dc6c0d0ba6ad31d1b4bd61311ee38169c07ad948e49afecf553e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 199, + 70, + 268, + 178 + ], + "blocks": [ + { + "bbox": [ + 199, + 70, + 268, + 178 + ], + "lines": [ + { + "bbox": [ + 199, + 70, + 268, + 178 + ], + "spans": [ + { + "bbox": [ + 199, + 70, + 268, + 178 + ], + "type": "image", + "image_path": "2c8e0d3deaa7378e30eb5b74d25ae0885c4f85e96648f82b41d861bdda21a1d9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 271, + 70, + 339, + 178 + ], + "blocks": [ + { + "bbox": [ + 271, + 70, + 339, + 178 + ], + "lines": [ + { + "bbox": [ + 271, + 70, + 339, + 178 + ], + "spans": [ + { + "bbox": [ + 271, + 70, + 339, + 178 + ], + "type": "image", + "image_path": "765b5887378ad2204b9be39ccbea1de9ac1c1c9e370b3451465e5eb39d1aa5c2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 342, + 70, + 411, + 178 + ], + "blocks": [ + { + "bbox": [ + 342, + 70, + 411, + 178 + ], + "lines": [ + { + "bbox": [ + 342, + 70, + 411, + 178 + ], + "spans": [ + { + "bbox": [ + 342, + 70, + 411, + 178 + ], + "type": "image", + "image_path": "f957d898bf9dc599bcdd8e9216c0ab098ecce1d032a55024b188210e832dd08a.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 414, + 70, + 482, + 178 + ], + "blocks": [ + { + "bbox": [ + 414, + 70, + 482, + 178 + ], + "lines": [ + { + "bbox": [ + 414, + 70, + 482, + 178 + ], + "spans": [ + { + "bbox": [ + 414, + 70, + 482, + 178 + ], + "type": "image", + "image_path": "8918109efd6f8d94ac290a21a5d650ab71a3b36a73bf67bdbd5656d45009e362.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 485, + 70, + 553, + 178 + ], + "blocks": [ + { + "bbox": [ + 485, + 70, + 553, + 178 + ], + "lines": [ + { + "bbox": [ + 485, + 70, + 553, + 178 + ], + "spans": [ + { + "bbox": [ + 485, + 70, + 553, + 178 + ], + "type": "image", + "image_path": "15dfe72c48d5cc6e9f93f560bb9449fd27012baa0ca7c4dc18b2a4709e8fd4a3.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": " is the extracted feature from " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": "-th layer of VGG-16, " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_l^\\phi" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": " is the Gram matrix of " + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "inline_equation", + "content": "\\phi_l" + }, + { + "bbox": [ + 55, + 263, + 295, + 289 + ], + "type": "text", + "content": ". Our structure loss is:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 121, + 300, + 294, + 313 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 300, + 294, + 313 + ], + "spans": [ + { + "bbox": [ + 121, + 300, + 294, + 313 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {s t r u c t}} = \\alpha \\mathcal {L} _ {\\mathrm {M S E}} + \\beta \\mathcal {L} _ {\\text {p e r c}}, \\tag {14}", + "image_path": "2eb2469e459e1a970a3bed7da4573dbb50a11f2c37816e0cd494183c3540bd74.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "inline_equation", + "content": "\\alpha = 0.01, \\beta = 0.001" + }, + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "text", + "content": " are the balance weights of different terms. Using " + }, + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{adj}}" + }, + { + "bbox": [ + 55, + 324, + 296, + 385 + ], + "type": "text", + "content": " as the supervision target enables the refinement module to focus on fine-grained detail refinement as illustrated in Figure 5 \"Complete\", in which the building is undistorted and the road curb is natural." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 396, + 153, + 408 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 396, + 153, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 396, + 153, + 408 + ], + "type": "text", + "content": "3.4. Training details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "text", + "content": "As a large number of paired events, LDR image, and GT is needed for training, we adopt 733 HDR images collected by Yang et al. [48] and follow the same data generation process to generate training and testing data with resolution (512, 512). We first simulate 733 HDR videos with random global motion by generating random camera motion trajectories [3]. We utilize the event simulator [31] and virtual camera [6] to simulate the events and LDR images, respectively. The dataset generated from 733 images is separated into training and testing following the setting of Yang et al. [48], while 663 for training and 70 for testing. During training, we randomly over- or under-exposed " + }, + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 54, + 415, + 296, + 571 + ], + "type": "text", + "content": " pixels of the HDR images to generate LDR images." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": "Other implementation details. The proposed method is implemented with the Pytorch framework and runs on a single NVIDIA GeForce RTX 3090 GPU. For optimization, we use the ADAM optimizer [19] with a linear decay learning rate scheduler starting from step 0 and an initial learning rate of " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\gamma = 10^{-5}" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": ". DDIM noise scheduler is adopted both in training and sampling, with training steps " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "N = 1000" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and linear-scaled beta schedule. We adopt " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "C = 5" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " same to HDRev [48] to apply their pretrain model. During sampling, we adopt classifier-free guidance with guidance scale " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "u = 1.5" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " and inference steps " + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "inline_equation", + "content": "n = 9" + }, + { + "bbox": [ + 55, + 582, + 296, + 713 + ], + "type": "text", + "content": " to obtain our results." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 298, + 555, + 396 + ], + "blocks": [ + { + "bbox": [ + 313, + 262, + 555, + 295 + ], + "lines": [ + { + "bbox": [ + 313, + 262, + 555, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 262, + 555, + 295 + ], + "type": "text", + "content": "Table 1. Quantitative evaluation of synthetic data on dataset collected by Yang et al. [48]. The red metric shows the best performance. " + }, + { + "bbox": [ + 313, + 262, + 555, + 295 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 313, + 262, + 555, + 295 + ], + "type": "text", + "content": " means higher (lower) is better." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 298, + 555, + 396 + ], + "lines": [ + { + "bbox": [ + 315, + 298, + 555, + 396 + ], + "spans": [ + { + "bbox": [ + 315, + 298, + 555, + 396 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
Liu et al. [24]18.350.7710.27615.3378.414.02
EventHDR [53]11.040.3340.44723.44182.114.58
Sagiri [22]12.500.4530.41417.6783.465.35
HDRev [48]14.050.6190.23817.6946.233.88
NeurImg [14]18.530.6210.33817.42105.074.04
Dille et al. [4]19.730.8200.2439.7676.834.22
Ours25.670.9260.0996.0127.093.86
", + "image_path": "352c6f914f42a1c5db480dbae694a975523d22d1c2006a5186e910a8b7c7c0e0.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 407, + 395, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 395, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 395, + 420 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 427, + 555, + 511 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 511 + ], + "type": "text", + "content": "The comparison with existing methods including four categories: event-based EventHDR [53], single-image CNN-based Liu et al. [24] and Dille et al. [4], single-image diffusion-based Sagiri [22], and event-guided NeurImg-HDR [14] and HDRev [48]. Experiments are conducted both on synthetic data and real data, as described in Section 4.1 and Section 4.2 respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 514, + 556, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 556, + 611 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 556, + 611 + ], + "type": "text", + "content": "Metrics For fidelity preserving, we adopt reference metrics including peak signal-to-noise ratio (PSNR), structural similarity (SSIM), the perceptual error with learned perceptual image patch similarity (LPIPS) [51], the Frechet Inception Distance (FID) [15], and the Color Difference Evaluation 2000 (CIEDE) [25], respectively. A non-reference metric named the Natural Image Quality Evaluator (NIQE) [27] is adopted to evaluate the image quality." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 615, + 474, + 628 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 615, + 474, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 615, + 474, + 628 + ], + "type": "text", + "content": "4.1. Evaluation on Synthetic Data" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "The quantitative evaluation is reported in Table 1. The proposed method outperforms the previous methods in terms of fidelity and structure preservation as shown by reference metrics. Also, ours image quality outperforms other methods, as indicated by NIQE [27]. The qualitative evaluation is illustrated in Figure 6. EventHDR [53] reconstructs faithful grayscale HDR images without color. Liu et al. [24]" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "11792" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 72, + 553, + 230 + ], + "blocks": [ + { + "bbox": [ + 59, + 72, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 59, + 72, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 59, + 72, + 553, + 230 + ], + "type": "image", + "image_path": "cce0b5cd0446b2cbf19bee4561fb0576067408b861fd8430c1f37cd9d9bc97bd.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 236, + 466, + 247 + ], + "lines": [ + { + "bbox": [ + 143, + 236, + 466, + 247 + ], + "spans": [ + { + "bbox": [ + 143, + 236, + 466, + 247 + ], + "type": "text", + "content": "Figure 6. Qualitative evaluation of synthetic data on dataset collected by Yang et al. [48]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 252, + 553, + 432 + ], + "blocks": [ + { + "bbox": [ + 57, + 252, + 553, + 432 + ], + "lines": [ + { + "bbox": [ + 57, + 252, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 57, + 252, + 553, + 432 + ], + "type": "image", + "image_path": "a80edf8c56f24607ea4705587b770068af85c464680f34370ce212d7326d3ee5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 198, + 433, + 410, + 443 + ], + "lines": [ + { + "bbox": [ + 198, + 433, + 410, + 443 + ], + "spans": [ + { + "bbox": [ + 198, + 433, + 410, + 443 + ], + "type": "text", + "content": "Figure 7. Qualitative evaluation of real data on DSEC [8]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 73, + 465, + 535, + 514 + ], + "blocks": [ + { + "bbox": [ + 70, + 449, + 539, + 461 + ], + "lines": [ + { + "bbox": [ + 70, + 449, + 539, + 461 + ], + "spans": [ + { + "bbox": [ + 70, + 449, + 539, + 461 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation (NIQE [27]↓/MANIQA [47]↑/QualiClip [1]↑) of real data. Blue metrics indicate the second-best." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 73, + 465, + 535, + 514 + ], + "lines": [ + { + "bbox": [ + 73, + 465, + 535, + 514 + ], + "spans": [ + { + "bbox": [ + 73, + 465, + 535, + 514 + ], + "type": "table", + "html": "
DatasetLiu et al. [24]EventHDR [53]Sagiri [22]HDRev [48]NeurImg [14]Dille et al.Ours
DSEC [8](3.65/0.466)(4.00/0.270)(5.87/0.617)(3.41/0.402)(3.80/0.570)(3.63/0.467)(3.12/0.693)
HES-HDR [14](5.67/0.335)(5.52/0.203)(6.40/0.576)(5.01/0.322)(4.61/0.442)(5.72/0.367)(4.81/0.530)
", + "image_path": "046e149a28427118415d145e4390d9bec1c5e18e71c9dd5f906953cf0bf80783.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 54, + 519, + 295, + 639 + ], + "type": "text", + "content": "maintains details from LDR images but is hard to compensate for missing information. Sagiri [22] adopts diffusion models to provide high-quality images but fails to preserve consistency with inputs. NeurImg [14] has difficulty combining events with LDR image to predict over-exposed areas and has noise-like artifacts. HDRev [48] recovers HDR information, while it is difficult to deal with brightness and color adjustment in over-/under-exposed areas. Our method provides natural results and outperforms other methods both on missing information generation and fidelity preservation." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 645, + 194, + 657 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 645, + 194, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 194, + 657 + ], + "type": "text", + "content": "4.2. Evaluation on Real Data" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 296, + 696 + ], + "type": "text", + "content": "Comparison on DSEC dataset DSEC [8] is a stereo event camera dataset for driving scenarios. We choose parts3 of data recording HDR scenes as our testing data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 651 + ], + "type": "text", + "content": "The quantitative evaluation using non-reference metrics is shown in Table 2, indicating that the proposed method achieves higher image quality than others. The qualitative evaluation of the DSEC dataset is depicted in Figure 7. EventHDR [53] fails to restore grayscale results. It's challenging for Liu et al. [24] and Sagiri [22] to hallucinate missing information. NeurImg [14] and HDRev [48] have problems utilizing the HDR information provided by events. The proposed method not only provides realistic and faithful results compared with other methods but also exhibits generation ability even with fewer events." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 556, + 715 + ], + "type": "text", + "content": "Comparison on HES-HDR dataset HES-HDR [14] is a hybrid event and spike HDR dataset. Its hybrid event dataset is suitable for our task. Although the non-reference metrics are not the best, we show better recovery ability as shown by qualitative evaluation3. The qualitative evalua" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 257, + 713 + ], + "type": "text", + "content": "3More details are provided in the supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11793" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 553, + 249 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 553, + 249 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 553, + 249 + ], + "type": "image", + "image_path": "a8bb1686632dcfa21009aa04b4ed909e9cb103dd165faea8e19686e631e4ffce.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 250, + 421, + 262 + ], + "lines": [ + { + "bbox": [ + 188, + 250, + 421, + 262 + ], + "spans": [ + { + "bbox": [ + 188, + 250, + 421, + 262 + ], + "type": "text", + "content": "Figure 8. Qualitative evaluation of real data on HES-HDR [14]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 56, + 313, + 296, + 415 + ], + "blocks": [ + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "type": "text", + "content": "Table 3. Ablation studies of synthetic data on the dataset collected by Yang et al. [48]. \"Proposed*\" is calculated with adjusted GT " + }, + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{adj}}" + }, + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "type": "text", + "content": ". The red metric shows the best performance among different ablation parts. " + }, + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "type": "inline_equation", + "content": "\\uparrow (\\downarrow)" + }, + { + "bbox": [ + 55, + 265, + 295, + 309 + ], + "type": "text", + "content": " means higher (lower) is better." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 313, + 296, + 415 + ], + "lines": [ + { + "bbox": [ + 56, + 313, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 296, + 415 + ], + "type": "table", + "html": "
PSNR↑SSIM↑LPIPS↓CIEDE↓FID↓NIQE↓
ConCond22.600.7490.2007.4848.144.23
RestCond20.790.7250.2479.3361.744.20
W/o Pyramid21.620.7280.2238.6746.993.89
W/o Refinement23.290.770.1707.2938.133.63
W/o structure26.650.9330.0935.6725.903.90
Proposed25.670.9260.0996.0127.093.86
Proposed*29.770.9390.0803.7922.623.86
", + "image_path": "d8b6cac3fe629e9bf18b060588a7a0af793f9e574a78fd0ff9a8bd3a08cef6aa.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 419, + 296, + 468 + ], + "type": "text", + "content": "tion of the HES-HDR dataset is depicted in Figure 8, which shows similar conclusions as DSEC [8]. Besides, even with unnatural color-shifted LDR image, our results have a more natural appearance and better fidelity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 149, + 492 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 713 + ], + "type": "text", + "content": "Our ablation study includes two parts, different conditioning processes, and different refinement targets. The quantitative evaluations of different conditioning processes are shown in the upper part of Table 3. \"ConCond\" and \"Rest-Cond\" are as described in Section 3.2. \"W/o Pyramid\" represents removing the pyramid fusion module and adopting the lowest HDR features as conditions. The proposed one (\"W/o Refinement\") outperforms others. The qualitative evaluations in Figure 4 show that the proposed one can better utilize events and LDR information to provide faithful and realistic images. The quantitative evaluations of refinement and structure loss are depicted in the lower parts of Table 3. The ablation of removing the refinement module is shown by \"W/o Refinement\". Compared with diffusion-only method \"W/o Refinement\", the refinement module improves the fidelity in terms of supervised metrics. Employing GT as a supervision target (\"W/o structure\") makes the reference metrics a little better than \"Proposed\". However," + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": "due to the different supervision targets, directly comparing the proposed method with GT cannot show our structure-preserving capacity. Therefore, we include \"Proposed*\", which is calculated with the adjusted GT " + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "inline_equation", + "content": "O_{\\mathrm{adj}}" + }, + { + "bbox": [ + 313, + 266, + 555, + 363 + ], + "type": "text", + "content": ". The image quality reflected by NIQE of \"Proposed\" is better than \"W/o structure\". As shown by Figure 5, more natural and higher-quality images without distortion can be generated with the proposed refinement module." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 372, + 388, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 372, + 388, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 372, + 388, + 384 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 387, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 555, + 533 + ], + "type": "text", + "content": "In this paper, we present a fidelity-preserving HDR reconstruction method based on diffusion models. Our method consists of two parts, event-guided conditioning generation and fine-grained detail refinement. We utilize an event-image encoder and a pyramid fusion module to efficiently and effectively fuse events and images, providing HDR conditions to guide the diffusion process and generate missing information. A refinement module with histogram-based structure loss is proposed to tackle distortion without changing the brightness and contrast. Experiments on both synthetic and real data demonstrate the generation and fidelity preservation ability of our method." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 593 + ], + "type": "text", + "content": "Limitations. Since diffusion models have uncertainty in the generation process, it is difficult to apply our method for video generation, for the challenge of maintaining consistency over consecutive frames." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 603, + 411, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 411, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 411, + 616 + ], + "type": "text", + "content": "Acknowledgement" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 712 + ], + "type": "text", + "content": "This work was supported by National Natural Science Foundation of China (Grant No. 62088102, 62136001), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resource." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11794" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 133 + ], + "type": "text", + "content": "[1] Lorenzo Agnolucci, Leonardo Galteri, and Marco Bertini. Quality-aware image-text alignment for opinion-unaware image quality assessment. arXiv preprint arXiv:2403.11176, 2024. 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 180 + ], + "type": "text", + "content": "[2] Hrishav Bakul Barua, Kalin Stefanov, KokSheik Wong, Abhinav Dhall, and Ganesh Krishnasamy. GTA-HDR: A large-scale synthetic dataset for HDR image reconstruction. arXiv preprint arXiv:2403.17837, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "spans": [ + { + "bbox": [ + 62, + 182, + 295, + 213 + ], + "type": "text", + "content": "[3] Giacomo Boracchi and Alessandro Foi. Modeling the performance of image restoration from motion blur. IEEE TIP, 2012. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 216, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 216, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 216, + 294, + 237 + ], + "type": "text", + "content": "[4] Sebastian Dille, Chris Careaga, and Yaqiz Aksoy. Intrinsic single-image hdr reconstruction. In ECCV, 2024. 1, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 295, + 282 + ], + "type": "text", + "content": "[5] Gabriel Eilertsen, Joel Kronander, Gyorgy Denes, Rafal K Mantiuk, and Jonas Unger. HDR image reconstruction from a single exposure using deep CNNs. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 1, 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 317 + ], + "type": "text", + "content": "[6] Yuki Endo, Yoshihiro Kanamori, and Jun Mitani. Deep reverse tone mapping. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH Asia), 2017. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 294, + 361 + ], + "type": "text", + "content": "[7] Ben Fei, Zhaoyang Lyu, Liang Pan, Junzhe Zhang, Weidong Yang, Tianyue Luo, Bo Zhang, and Bo Dai. Generative diffusion prior for unified image restoration and enhancement. In CVPR, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "text", + "content": "[8] Mathias Gehrig, Mario Millhäuser, Daniel Gehrig, and Davide Scaramuzza. E-RAFT: Dense optical flow from event cameras. In International Conference on 3D Vision (3DV), 2021. 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 453 + ], + "type": "text", + "content": "[9] Michael Gharbi, Jiawen Chen, Jonathan Barron, Samuel Hasinoff, and Frédo Durand. Deep bilateral learning for real-time image enhancement. ACM Transactions on Graphics (Proc. of ACM SIGGRAPH), 2017. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 455, + 294, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 455, + 294, + 497 + ], + "spans": [ + { + "bbox": [ + 57, + 455, + 294, + 497 + ], + "type": "text", + "content": "[10] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. In NeurIPS, 2014. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 500, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 500, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 57, + 500, + 294, + 542 + ], + "type": "text", + "content": "[11] Abhishek Goswami, Aru Ranjan Singh, Francesco Banterle, Kurt Debattista, and Thomas Bashford-Rogers. Semantic aware diffusion inverse tone mapping. arXiv preprint arXiv:2405.15468, 2024. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 544, + 294, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 544, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 57, + 544, + 294, + 587 + ], + "type": "text", + "content": "[12] Yuanshen Guan, Ruikang Xu, Mingde Yao, Ruisheng Gao, Lizhi Wang, and Zhiwei Xiong. Diffusion-promoted HDR video reconstruction. arXiv preprint arXiv:2406.08204, 2024. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 294, + 633 + ], + "type": "text", + "content": "[13] Jin Han, Chu Zhou, Peiqi Duan, Yehui Tang, Chang Xu, Chao Xu, Tiejun Huang, and Boxin Shi. Neuromorphic camera guided high dynamic range imaging. In CVPR, 2020. 2, 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 294, + 689 + ], + "type": "text", + "content": "[14] Jin Han, Yixin Yang, Peiqi Duan, Chu Zhou, Lei Ma, Chao Xu, Tiejun Huang, Imari Sato, and Boxin Shi. Hybrid high dynamic range imaging fusing neuromorphic and conventional images. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023. 2, 4, 6, 7, 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[15] Martin Heusel, Hubert Ramsauer, Thomas Unterthiner, Bernhard Nessler, and Sepp Hochreiter. Gans trained by a" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 94 + ], + "type": "text", + "content": "two time-scale update rule converge to a local nash equilibrium. NeurIPS, 2017. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 117 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 161 + ], + "type": "text", + "content": "[17] Tao Hu, Qingsen Yan, Yuankai Qi, and Yanning Zhang. Generating content for hdr deghosting from frequency view. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 553, + 216 + ], + "type": "text", + "content": "[18] Xuan Ju, Xian Liu, Xintao Wang, Yuxuan Bian, Ying Shan, and Qiang Xu. Brushnet: A plug-and-play image inpainting model with decomposed dual-branch diffusion. In Proceedings of the European Conference on Computer Vision (ECCV). ECCV, 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 217, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 217, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 317, + 217, + 553, + 248 + ], + "type": "text", + "content": "[19] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 251, + 553, + 292 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 292 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 292 + ], + "type": "text", + "content": "[20] Orest Kupyn, Volodymyr Budzan, Mykola Mykhailych, Dmytro Mishkin, and Jiří Matas. Deblurgan: Blind motion deblurring using conditional adversarial networks. In CVPR, 2018. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 295, + 553, + 348 + ], + "type": "text", + "content": "[21] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 392 + ], + "type": "text", + "content": "[22] Baiang Li, Sizhuo Ma, Yanhong Zeng, Xiaogang Xu, Youqing Fang, Zhao Zhang, Jian Wang, and Kai Chen. Sagiri: Low dynamic range image enhancement with generative diffusion prior. arxiv, 2024. 1, 2, 4, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 317, + 394, + 553, + 437 + ], + "type": "text", + "content": "[23] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-BIR: Towards blind image restoration with generative diffusion prior. arXiv preprint arXiv:2308.15070, 2024. 2, 4" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "spans": [ + { + "bbox": [ + 317, + 438, + 553, + 481 + ], + "type": "text", + "content": "[24] Yu-Lun Liu, Wei-Sheng Lai, Yu-Sheng Chen, Yi-Lung Kao, Ming-Hsuan Yang, Yung-Yu Chuang, and Jia-Bin Huang. Single-image HDR reconstruction by learning to reverse the camera pipeline. In CVPR, 2020. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 483, + 553, + 568 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 483, + 553, + 568 + ], + "spans": [ + { + "bbox": [ + 317, + 483, + 553, + 568 + ], + "type": "text", + "content": "[25] M Ronnier Luo, Guihua Cui, and Bryan Rigg. The development of the cie 2000 colour-difference formula: Ciede2000. Color Research & Application: Endorsed by Inter-Society Color Council, The Colour Group (Great Britain), Canadian Society for Color, Color Science Association of Japan, Dutch Society for the Study of Color, The Swedish Colour Centre Foundation, Colour Society of Australia, Centre Français de la Couleur, 26(5):340-350, 2001. 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 571, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 571, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 317, + 571, + 553, + 622 + ], + "type": "text", + "content": "[26] Nico Messikommer, Stamatios Georgoulis, Daniel Gehrig, Stepan Tulyakov, Julius Erbach, Alfredo Bochicchio, Yuanyou Li, and Davide Scaramuzza. Multi-bracket high dynamic range imaging with event cameras. In CVPR, 2022. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 625, + 553, + 657 + ], + "type": "text", + "content": "[27] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 690 + ], + "type": "text", + "content": "[28] Mohammad Mostafavi, Lin Wang, and Kuk-Jin Yoon. Learning to reconstruct hdr images from events, with applications to depth and flow prediction. IJCV, 2021. 1, 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 692, + 553, + 713 + ], + "type": "text", + "content": "[29] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11795" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 95 + ], + "type": "text", + "content": "Robin Rombach. SDXL: Improving latent diffusion models for high-resolution image synthesis. In ICLR, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 118 + ], + "type": "text", + "content": "[30] Yohan Poirier-Ginter and Jean-François Lalonde. Robust unsupervised stylegan image restoration. In CVPR, 2023. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "spans": [ + { + "bbox": [ + 56, + 119, + 294, + 152 + ], + "type": "text", + "content": "[31] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: An open event camera simulator. In Proc. of Conference on Robotics Learning, 2018. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "spans": [ + { + "bbox": [ + 57, + 153, + 295, + 185 + ], + "type": "text", + "content": "[32] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. Events-to-video: Bringing modern computer vision to event cameras. In CVPR, 2019. 1, 2, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 219 + ], + "type": "text", + "content": "[33] Henri Rebecq, René Ranftl, Vladlen Koltun, and Davide Scaramuzza. High speed and high dynamic range video with an event camera. IEEE TPAMI, 2019. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 220, + 294, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 220, + 294, + 253 + ], + "spans": [ + { + "bbox": [ + 57, + 220, + 294, + 253 + ], + "type": "text", + "content": "[34] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 2, 3, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 295, + 287 + ], + "type": "text", + "content": "[35] Richard Shaw, Sibi Catley-Chandar, Ales Leonardis, and Eduardo Pérez-Pellitero. HDR reconstruction from bracketed exposures and events. In BMVC, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 289, + 295, + 321 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 289, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 57, + 289, + 295, + 321 + ], + "type": "text", + "content": "[36] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 322, + 295, + 366 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 295, + 366 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 295, + 366 + ], + "type": "text", + "content": "[37] Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 57, + 367, + 295, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 367, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 57, + 367, + 295, + 399 + ], + "type": "text", + "content": "[38] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 401, + 295, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 295, + 443 + ], + "type": "text", + "content": "[39] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 295, + 498 + ], + "type": "text", + "content": "[40] Chao Wang, Ana Serrano, Xingang Pan, Bin Chen, Karol Myszkowski, Hans-Peter Seidel, Christian Theobalt, and Thomas Leimkuhler. GlowGAN: Unsupervised learning of HDR images from LDR images in the wild. In ICCV, 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 295, + 544 + ], + "type": "text", + "content": "[41] Lin Wang, Yo-Sung Ho, Kuk-Jin Yoon, et al. Event-based high dynamic range image and very high frame rate video generation using conditional generative adversarial networks. In CVPR, 2019. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 546, + 295, + 589 + ], + "type": "text", + "content": "[42] Zejia Weng, Xitong Yang, Zhen Xing, Zuxuan Wu, and Yu-Gang Jiang. Genrec: Unifying video generation and recognition with diffusion models. arXiv preprint arXiv:2408.15241, 2024. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 591, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 591, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 591, + 295, + 624 + ], + "type": "text", + "content": "[43] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 656 + ], + "type": "text", + "content": "[44] Wenhan Wu, Xian Liu, and Jie Chen. Image restoration with controlnet. IEEE Transactions on Image Processing, 2023. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[45] Garima Yadav, Saurabh Maheshwari, and Anjali Agarwal. Contrast limited adaptive histogram equalization based enhancement for real time video system. In 2014 International Conference on Advances in Computing, Communications and Informatics (ICACCI), 2014. 5, 6" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 387 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[46] Qingsen Yan, Tao Hu, Yuan Sun, Hao Tang, Yu Zhu, Wei Dong, Luc Van Gool, and Yanning Zhang. Towards high-quality HDR dehosting with conditional diffusion models. IEEE Transactions on Circuits and Systems for Video Technology, 2023. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[47] Sidi Yang, Tianhe Wu, Shuwei Shi, Shanshan Lao, Yuan Gong, Mingdeng Cao, Jiahao Wang, and Yujiu Yang. Mania: Multi-dimension attention network for no-reference image quality assessment. In CVPR, 2022. 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 206 + ], + "type": "text", + "content": "[48] Yixin Yang, Jin Han, Jinxiu Liang, Imari Sato, and Boxin Shi. Learning event guided high dynamic range video reconstruction. In CVPR, 2023. 1, 2, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 251 + ], + "type": "text", + "content": "[49] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[50] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models, 2023. 2, 4, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 553, + 319 + ], + "type": "text", + "content": "[51] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 320, + 553, + 354 + ], + "type": "text", + "content": "[52] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In CVPR, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 387 + ], + "type": "text", + "content": "[53] Yunhao Zou, Yinqiang Zheng, Tsuyoshi Takatani, and Ying Fu. Learning to reconstruct high speed and high dynamic range videos from events. In CVPR, 2021. 1, 2, 6, 7" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11796" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_content_list.json b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e621c8569cfd84cf3ad114b42fb1b1a9d94e9f61 --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_content_list.json @@ -0,0 +1,1610 @@ +[ + { + "type": "text", + "text": "Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring", + "text_level": 1, + "bbox": [ + 114, + 128, + 880, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taewoo Kim* KAIST", + "bbox": [ + 328, + 203, + 439, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "intelpro@kaist.ac.kr", + "bbox": [ + 292, + 242, + 470, + 255 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Kuk-Jin Yoon KAIST", + "bbox": [ + 565, + 203, + 679, + 237 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "kjyoon@kaist.ac.kr", + "bbox": [ + 537, + 242, + 700, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In low-light environments, longer exposure times are commonly used to enhance image visibility; however, this inevitably leads to motion blur. Even with a long exposure time, videos captured in low-light environments still suffer from issues such as low visibility, low contrast, and color distortion. Additionally, the long exposure time results in videos with a low frame rate. Therefore, videos captured in low-light exhibit low visibility and motion blur, as well as low frame rates. To overcome these limitations, we propose a novel problem aimed at transforming motion-blurred, low-frame-rate videos with poor visibility in low-light environments into high-frame-rate videos while simultaneously enhancing their visibility. To tackle this challenge, we leverage the unique advantages of event cameras, which capture scene changes asynchronously, providing superior temporal resolution and a wider dynamic range compared to conventional frame-based cameras. These properties make event cameras particularly effective in reducing motion blur, compensating for low frame rates, and enhancing visibility in low-light conditions. To this end, we developed a hybrid camera system that integrates two RGB cameras and an event camera, capturing a dedicated dataset for this task and proposing novel network architectures to effectively address this problem. The project pages are available at https://sites.google.com/view/eledi-iccv.", + "bbox": [ + 91, + 323, + 483, + 715 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 220, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In our daily lives, low-light conditions are common and occur in various lighting scenarios, both during the day and at night. In such environments, a longer exposure time is generally required to capture the limited available light. However, this extended exposure time inevitably leads to motion blur caused by object movement or dynamic camera motion, while also reducing video frame rates. Additionally,", + "bbox": [ + 89, + 771, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "videos recorded under these conditions often suffer from reduced visibility due to insufficient lighting, along with significant image noise, even with a long exposure time. As a result, videos captured in low-light conditions typically exhibit low frame rates due to the long exposure time, as well as poor visibility and motion blur.", + "bbox": [ + 511, + 292, + 906, + 383 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "With advancements in deep learning, numerous efforts have focused on improving low-light conditions [3, 53, 54], aiming to suppress noise and enhance visibility in such environments. Furthermore, research has been conducted on removing motion blur artifacts through motion deblurring studies [6, 30, 31], as well as on Video Frame Interpolation (VFI) [13, 36, 38, 44], which converts low-frame-rate videos into high-frame-rate ones. While each of these areas has been explored independently, few studies address the combined challenges of multiple degradations in low-light conditions. Among existing works, relatively active research has focused on the problem of joint deblurring and interpolation [30, 63], which aims to restore high-frame-rate, sharp frames from motion-blurred ones. Additionally, some works [16, 65] have investigated integrating motion deblurring and low-light enhancement. However, no existing approach comprehensively tackles all three degradations simultaneously.", + "bbox": [ + 511, + 383, + 908, + 656 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To overcome these limitations, we propose a research problem that simultaneously addresses the challenges of low frame rate, motion blur, and low visibility in low-light conditions. Specifically, we aim to simultaneously enhance visibility, reduce motion blur, and increase the frame rate to improve overall video quality in low-light environments. A simple approach would involve cascading separate models for each task. However, such a cascaded manner suffers from error accumulation and inefficient joint optimization, often leading to suboptimal results. Thus, a more effective solution is needed to jointly optimize these interrelated tasks in an end-to-end manner. Moreover, solving these problems using only frame-based information is extremely challenging, particularly in cases of severe degradation, making it an ill-posed problem.", + "bbox": [ + 511, + 657, + 908, + 883 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To tackle this issue, we propose leveraging the ad", + "bbox": [ + 532, + 885, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 805, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Work done while at KAIST.", + "bbox": [ + 109, + 886, + 264, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "8524", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e089bde8fcf3380ea2e68be893f3a71e322a7ab12e70cbdeef422c3611020cc5.jpg", + "image_caption": [ + "Low-light blurred image" + ], + "image_footnote": [], + "bbox": [ + 94, + 87, + 379, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3295b74b9e3312293e6128ae5c0ba86d4c0d69ed462de1213c093ab8f7cdf20e.jpg", + "image_caption": [ + "t=5/8", + "Figure 1. Qualitative result of our method on the RELBFI datasets. Our task aims to restore sharp frames captured in a normal light environment within the shutter interval using low-light blurred images and the events." + ], + "image_footnote": [], + "bbox": [ + 379, + 88, + 509, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/3e6e47a1b59447b699604f4027a8a9377753aecd3de09424a1c0c03322ed6bad.jpg", + "image_caption": [ + "t=6/8" + ], + "image_footnote": [], + "bbox": [ + 509, + 88, + 638, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e8226a961a1f7a383a5fdf30df640b81c51223ec04eea5110da2adeaaf579ec.jpg", + "image_caption": [ + "t=7/8" + ], + "image_footnote": [], + "bbox": [ + 638, + 88, + 767, + 243 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/cd4136efb7b560ad34626348d90fcc243a287509fd6a507b046fdc9fd5ec8473.jpg", + "image_caption": [ + "t=8/8" + ], + "image_footnote": [], + "bbox": [ + 767, + 88, + 897, + 243 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "vantages of event cameras, which, inspired by neuromorphic vision systems, detect brightness changes with microsecond-level temporal resolution [8]. Additionally, their superior dynamic range, compared to conventional cameras, allows for high-speed motion capture even in low-light environments. As a result, event cameras are widely used in video enhancement tasks, including motion deblurring [40, 41], frame interpolation [13, 44], low-light enhancement [16, 20, 23], and joint deblurring and interpolation [41], effectively addressing issues like motion blur, low frame rates, low visibility.", + "bbox": [ + 88, + 313, + 480, + 479 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Therefore, we propose a novel task that leverages event cameras to address these multiple degradations in low-light conditions. However, since this novel task has never been addressed in previous works, it is necessary to build a new dataset. To overcome these challenges, we build a sophisticated hybrid camera system to capture a new RELBFI(Real-world Event-guided Low-light enhancement and Blur Frame Interpolation) dataset, consisting of real-world low-light blur videos with a low frame rate, low-light event data, and high-frame-rate normal light sharp videos (ground truth). We believe that these resources will be a valuable asset to the research community.", + "bbox": [ + 88, + 484, + 482, + 667 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Beyond proposing this dataset, we also introduce a novel unified end-to-end framework to tackle this challenging task. To achieve this, we design two key modules that play a crucial role in effectively addressing the problem. The first module focuses on fusing multi-temporal information from events and frames under low-light conditions. Given that video frames in such environments often suffer from significant noise and color distortion, this module aims to efficiently integrate multi-temporal and multi-modal information sources. For the second module, we introduce a mechanism to reinforce temporal dependencies among generated video frames during the generation of multiple frames within the shutter period. Since this process is computationally intensive, we designed a lightweight yet effective module to strengthen bidirectional dependencies between", + "bbox": [ + 89, + 674, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "frames, thereby enhancing overall efficiency.", + "bbox": [ + 511, + 313, + 810, + 328 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, the contributions of this paper are as follows: (I) We propose a new research problem that addresses multiple degradations, including noise, motion blur, low visibility, and low frame rate, which commonly occur in low-light conditions, while also converting videos captured at low frame rates into high frame rates. (II) We present a novel unified framework that effectively addresses this problem, demonstrating at least a 2 dB improvement in PSNR compared to state-of-the-art methods. (III) We introduce the first RELBFI dataset, built using a hybrid camera system.", + "bbox": [ + 511, + 329, + 906, + 494 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 511, + 506, + 661, + 522 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Motion Deblurring.", + "text_level": 1, + "bbox": [ + 511, + 531, + 700, + 547 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since the advent of deep learning, significant progress has been made in motion deblurring, further accelerated by specialized datasets such as GoPro [29], DVD [39], RealBlur [37], and BSD [62]. In frame-based deblurring, various architectures have been proposed, including multiscale [6], multi-stage [58, 60], transformer-based [42, 49, 59], and frequency-modulated [7, 17, 28] architectures. Video deblurring methods [19, 21, 22, 32, 61, 62] have also emerged. However, deblurring solely from frames remains challenging due to motion information loss, prompting event camera-based research. Early works explored recurrent architectures [10], followed by studies on cross-modal attention-based feature fusion [40], unknown exposure time [12], non-coaxial deblurring [5], frequency-based video deblurring [15], and attention-based cross-modal alignment [14] and spatio-temporal collaboration [56].", + "bbox": [ + 511, + 553, + 906, + 795 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Video Frame Interpolation.", + "text_level": 1, + "bbox": [ + 511, + 801, + 759, + 819 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Research on frame-based video interpolation [1, 2, 18, 26, 35, 36, 38] has made significant progress, but it remains challenging due to the absence of explicit motion information between frames. In contrast, event cameras, with their high temporal resolution, have gained popularity in", + "bbox": [ + 511, + 824, + 905, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "8525", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/62ac4327a9f769fc244d73516257e7bf7660aa3209d3b16900a4cdb575a06bba.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 98, + 90, + 377, + 303 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/62d3ea117d00ef5c5004792ba1d1afbd969da9840ffddb0369a649454a121e2f.jpg", + "image_caption": [ + "(b)", + "Figure 2. In the RELBFI dataset description, (a) shows a diagram of the hybrid camera system used to capture the dataset. (b) represents the temporal irradiance relationship for the high-fps RGB camera (used for recording sharp images) and the camera used for capturing blur. (c) presents sample images from the RELBFI dataset: from top to bottom, a low-light blur image, a brightness-adjusted version of the low-light blur image, and a ground truth image captured in normal light conditions." + ], + "image_footnote": [], + "bbox": [ + 388, + 97, + 612, + 300 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a9666428c0e8758c25f76efb13d49ed6540daa96ece2cd6de886fc03467e9f3c.jpg", + "image_caption": [ + "(c)" + ], + "image_footnote": [], + "bbox": [ + 633, + 97, + 875, + 292 + ], + "page_idx": 2 + }, + { + "type": "table", + "img_path": "images/bd4cdca7b02a1653db924ecd786c6160423930697573cf7232f64295d4119dfe.jpg", + "table_caption": [ + "Table 1. Comparison of HighREV and RELBFI Datasets. In the table, NL\" refers to images taken in normal-light environments, and \"LL\" refers to images taken in low-light environments. \"LFR\" refers to low frame rate, and \"HFR\" refers to high frame rate." + ], + "table_footnote": [], + "table_body": "
HighREV [41]RELBFI
LFR RGB InputNL blurLL blur
Event inputNL eventLL event
HFR RGB GTNL sharpNL sharp
Blur typeSyntheticReal
Event typeRealReal
#N of sharp images2858932629
#N of sequences3067
", + "bbox": [ + 109, + 458, + 457, + 584 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "video interpolation tasks. Early work like TimeLens [43] introduced event-based interpolation using a cascaded U-Net framework, while TimeLens++ [45] enhanced this by estimating non-linear motion with a motion spline estimator. Later studies, including A $^2$ OF [51] and the unsupervised TimeReplayer [9], further advanced the field. Recently, Kim et al. [13] introduced an event-image fusion VFI framework for inter-frame optical flow estimation, while the follow-up study, TimeLens-XL [27], presented a VFI framework with iterative flow updates.", + "bbox": [ + 89, + 606, + 483, + 758 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Low-light Enhancement.", + "text_level": 1, + "bbox": [ + 89, + 770, + 316, + 787 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Many frame-based low-light enhancement (LLE) studies focus on improving visibility and reducing noise using image information. These efforts have led to various approaches, including SNR-based transformers [53], Retinex-based methods [3, 52], gamma correction [47], implicit neural representations [55], diffusion models [48, 57], and other techniques [25, 54, 65]. Leveraging its high dynamic range,", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the event camera excels in capturing motion under low light, making it suitable for LLE. Liu et al. [24] used synthetic events for LLE due to dataset limitations, while Liang et al. [23] modeled multimodal coherence. Recently, Liang et al. [20] proposed an SNR-based framework and dataset for event-guided LLE, and Kim et al. [16] introduced a joint framework for event-guided LLE and deblurring.", + "bbox": [ + 511, + 398, + 906, + 506 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Joint deblurring and frame interpolation.", + "text_level": 1, + "bbox": [ + 511, + 515, + 867, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The joint task of deblurring and interpolation aims to simultaneously remove blur and generate high-frame-rate videos within a single framework. Jin et al. [11] first addressed this without event cameras using a deblur-then-interpolate approach. Recently, DeMFI [30] combined flow-guided attention and recursive boosting, while BiT [63] introduced a Swin Transformer-based solution. Event camera-based research includes EDI [33, 34], which used a physical model, REFID [41] leveraging a recurrent architecture, and Weng et al. [50], tackling blur frame interpolation under unknown exposure times. Unlike these works, we leverage the robust properties of event cameras in low-light conditions to jointly perform low-light enhancement, deblurring, and frame interpolation.", + "bbox": [ + 509, + 537, + 906, + 750 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. RELBFI Datasets", + "text_level": 1, + "bbox": [ + 511, + 762, + 687, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Comparisons with Other Datasets.", + "text_level": 1, + "bbox": [ + 511, + 787, + 813, + 803 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the advent of deep learning, various event- and frame-based video enhancement datasets [13, 16, 20, 63] have emerged. Some studies [41, 63] focus on motion deblurring and frame interpolation under normal lighting.", + "bbox": [ + 511, + 809, + 905, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Sun et al. [41] introduced a dataset for joint deblurring and interpolation using an event camera, where motion blur", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "8526", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4ae18f5f76e9fad75caf4e21d552afd2bf9a69ab8343a58e00334cd64b02f237.jpg", + "image_caption": [ + "Figure 3. Overall framework of the proposed method. \"TFG\" in the blue box denotes the Template Feature Generator block." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 883, + 361 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "was synthetically generated by merging adjacent frames. However, such synthetic datasets fail to capture real-world motion blur, as they overlook physical factors like noise variations from different exposure times.", + "bbox": [ + 88, + 401, + 480, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recently, Kim et al. [16] proposed the RELED dataset for simultaneous low-light enhancement and deblurring using an event camera. However, its low frame rate limits its effectiveness in converting low-frame-rate, low-light videos into high-frame-rate, normal-light videos.", + "bbox": [ + 88, + 460, + 482, + 536 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To date, no existing research or dataset has directly tackled the transformation of low-frame-rate, low-light, noisy, and motion-blurred inputs into high-frame-rate, noise-free videos under normal lighting by leveraging event data.", + "bbox": [ + 88, + 537, + 482, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. RELBFI dataset acquisitions.", + "text_level": 1, + "bbox": [ + 89, + 606, + 352, + 622 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Synthetic-based datasets struggle to capture the complexity of real-world degradations. Therefore, building a real-world dataset is both practical and challenging, especially when simultaneously capturing high-frame-rate sharp video, low-light blurred video, and low-light events.", + "bbox": [ + 88, + 628, + 482, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To address this, we build the RELBFI (Real-world Event-guided Low-light Video Enhancement and Blur Frame Interpolation) dataset using a hybrid camera system with beam splitters. As shown in Fig. 2(a), the system includes two RGB cameras (Blackfly BFS-U3-16S2C-CS) and an event camera (Propheesee Gen4). One RGB camera captures sharp images at 166fps with a 3ms exposure and a 3ms readout dead time, receiving $25\\%$ of the total light as it passes through the beam splitter twice without an ND filter. The other RGB camera records blurred images in low-light conditions with a 24ms exposure, eight times longer, and uses a 1/64 ND filter, receiving only $0.39\\%$ of the original light. As a result, each blurred image has eight times less", + "bbox": [ + 88, + 704, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "light intensity than a sharp image, resulting in low visibility.", + "bbox": [ + 511, + 401, + 903, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, to illustrate real-world frame acquisition [12, 41, 63], capturing blurred images involves exposure time and readout dead time, collectively termed the shutter interval [12]. To simulate readout dead time, we used a $24\\mathrm{ms}$ exposure followed by a $24\\mathrm{ms}$ dead time, yielding one low-light blurred image per $48\\mathrm{ms}$ (21fps) and eight sharp images under normal lighting (166fps). The event camera, like the blur-recording RGB camera, employs a 1/64 ND filter for low-light capture.", + "bbox": [ + 511, + 416, + 903, + 551 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Regarding calibration, all cameras in the hybrid system are mechanically fixed, but slight adjustments, such as rotation and baseline alignment, are required. To this end, we estimated the homography matrix for each camera and performed additional calibration. As shown in Fig. 2(a), the devices were temporally synchronized via hardware using an external trigger. Using this setup, we captured 67 urban scenes featuring dynamic elements like vehicles, moving objects, and flickering streetlights, along with rich textures. As shown in Tab. 1, RELBFI provides diverse sequences tailored for developing and evaluating methods in this challenging task. It serves as a practical resource to support joint low-light enhancement, deblurring, and interpolation in real-world scenarios.", + "bbox": [ + 511, + 551, + 905, + 763 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Proposed Methods", + "text_level": 1, + "bbox": [ + 511, + 777, + 694, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 803, + 624, + 816 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preliminary As explained in Sec. 3, the shutter interval consists of exposure time and readout deadtime. During the readout deadtime, events are recorded, but no blurred frames are captured. We define $N$ as the total number of sharp frames captured during a shutter interval $T$ (ms),", + "bbox": [ + 511, + 824, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "8527", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fa6c05dcf02dcfc6de3426008d6787ef8549aa1ba49258694ca98e334a571e74.jpg", + "image_caption": [ + "Figure 4. The MM-ISTFF module in the scale factor $s$ . In the figure, $\\mathcal{D}_s^{\\mathrm{fw}}$ and $\\mathcal{D}_s^{\\mathrm{bw}}$ represent the forward and backward dynamic filters, respectively, as defined in Eq.(1)." + ], + "image_footnote": [], + "bbox": [ + 94, + 85, + 478, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N = E + R$ , with $E$ and $R$ representing the number of the frames from exposure and readout deadtime, respectively. During a single shutter interval $t$ , one blurred frame $B^t$ is captured, and events are generated corresponding to both the exposure time and readout deadtime. The events $\\{e_i\\}_t$ during shutter interval $t$ are transformed into event voxels [66] with 32 bins.", + "bbox": [ + 89, + 332, + 483, + 438 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Overall framework The proposed framework is illustrated in Fig. 3. We utilize 3 blurred images and 3 corresponding event voxel grids from shutter intervals $t - 1$ , $t$ , and $t + 1$ . Given these inputs $\\{B^{t-1}, B^t, B^{t+1}\\}$ and $\\{E^{t-1}, E^t, E^{t+1}\\}$ , the network estimates $N$ normal light sharp frames at the shutter interval $t$ .", + "bbox": [ + 89, + 439, + 483, + 529 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The process consists of two main steps. First, the MM-ISTFF (Multi-Modal Inter-frame Spatio-Temporal Feature Fusion) module combines spatio-temporal event data with blurred video frames to improve video quality and reduce low-light noise, producing fused features $\\mathcal{F}(\\mathcal{K})_{s=0}^{t}$ . Next, the event voxel grid for $t$ is divided into $N$ temporal units $\\{E_0^t,\\ldots,E_{N-1}^t\\}$ , which are combined with $\\mathcal{F}(\\mathcal{K})_{s=0}^{t}$ to generate $N$ temporal template features $\\{\\mathcal{F}(\\mathcal{T})_0,\\ldots,\\mathcal{F}(\\mathcal{T})_{N-1}\\}$ . Next, the IS-TFF (Inter-Shutter Interval Temporal Feature Fusion) based-encoder module leverages temporal information within the shutter interval then uses these features to produce an enhanced feature map $\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}$ . Finally, the decoder generates $N$ normal light sharp frames $\\{S_0,\\ldots,S_{N-1}\\}$ for the $t$ shutter interval.", + "bbox": [ + 89, + 531, + 483, + 747 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Multi-Modal Inter-frame Spatio-Temporal Feature Fusion(MM-ISTFF) Module.", + "text_level": 1, + "bbox": [ + 89, + 756, + 482, + 787 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In low-light conditions, frame-based information is severely affected by noise and low visibility, while motion blur further exacerbates the degradation. In such scenarios, events also suffer from significant degradation. Therefore, it is crucial to fuse features by complementing the spatio-temporal visual information from images with the rich temporal details from events, effectively utilizing them", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/0af3b3d77c22ae4440a0b502ecba8985cebc7facb9b922c2a4be71ede8a6d973.jpg", + "image_caption": [ + "Figure 5. Illustration of the IS-TFF module. The $s$ after the forward and backward units indicates the scale factor." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 901, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for video enhancement. To address this, we propose the MM-ISTFF module, designed to mine discriminative temporal information between frames and events across different shutter intervals. For this, we extract multi-scale multi-modal features $\\{\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(B)_s^{t + 1}\\}$ and $\\{\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t,\\mathcal{F}(E)_s^{t + 1}\\}$ from the low-light blurred and event voxel inputs. As shown in Fig.3, we aim to leverage these multi-modal inputs to extract discriminative spatio-temporal information across multiple visual scales. To achieve this, we use past and future information from the current shutter interval $t$ . For simplicity, only the process for past information is described, as the same applies to future information. As shown in Fig.4, to integrate spatiotemporal information from past multi-modal data, we first apply a conv. layer with LeakyReLU to the concatenated features $[\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t]$ , generating the feature $\\mathbf{V}_s^{\\mathrm{fw}}$ . To extract discriminative spatio-temporal information, pixel-wise spatial dynamic filters [64] are applied to obtain content-adaptive features. Here, to estimate a dynamic filter for each pixel, we use $f_{s}^{fw} = f_{\\mathrm{filter}}^{\\mathrm{fw}}(\\mathbf{V}_{s}^{\\mathrm{fw}})$ to generate dynamic filters, which are then applied to the original feature $\\mathbf{V}_s^{\\mathrm{fw}}$ , where $f_{\\mathrm{filter}}$ We apply dynamic filter as follows:", + "bbox": [ + 511, + 358, + 906, + 705 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} = f _ {s} ^ {\\mathrm {f w}} \\otimes \\mathbf {V} _ {s} ^ {\\mathrm {f w}} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 718, + 906, + 738 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $f_{s}^{\\mathrm{fw}}$ represents the estimated filter, and $\\otimes$ denotes the dynamic filter application. This process extracts content-adaptive information, but in low-light conditions, selective regional features are needed due to noise and degradation. Thus, we apply region-wise attention for feature filtering.", + "bbox": [ + 511, + 742, + 905, + 819 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathbf {V}} _ {s} ^ {f w} = \\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} \\odot \\sigma (\\mathrm {L R e L U} ((W _ {3} (\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}})))) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 575, + 825, + 906, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\sigma$ denotes the sigmoid operation, and $W_{3}$ and LReLU represent the $3\\times 3$ conv. layer and the LeakyReLU operation, respectively. Similarly, future multi-modal informa", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "8528", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "tion is utilized to obtain $\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}}$ . These two features $(\\tilde{\\mathbf{V}}_s^{\\mathrm{fw}}$ and $\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}})$ are combined via channel-wise concatenation followed by a ResBlock to produce the final feature $\\mathcal{F}(\\mathcal{K})_s^t$ . This process is repeated for $s = \\{0,1,2\\}$ , with $\\mathcal{F}(\\mathcal{K})_{s=0}^{t}$ serving as input to the IS-TFF module.", + "bbox": [ + 89, + 89, + 480, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Inter-Shutter Interval Temporal Feature Fusion(IS-TFF) Module.", + "text_level": 1, + "bbox": [ + 89, + 180, + 482, + 212 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Template Feature Generation After the MM-ISTFF module, we obtain a spatio-temporally fused feature $\\mathcal{F}(\\mathcal{K})_{s=0}^{t}$ centered at shutter interval $t$ . To estimate $N$ sharp images within the shutter interval, template features capturing the characteristics of each timestep must be generated for all $N$ timesteps. To address this, we first divide the event voxel grid into $N$ temporal units, each with an equal time interval $\\Delta t_{unit} = T / N$ . This gives us temporally divided unit event voxels $E_k^t$ , which are then encoded to generate temporal event features. Therefore, we obtain the temporally divided unit event voxel $\\{E_k^t\\}$ with temporal index $k \\in \\{0, \\dots, N-1\\}$ . Subsequently, each event voxel unit is encoded to generate temporal event features $\\{\\mathcal{F}(E_k^t) \\in \\mathbb{R}^{C \\times H \\times W}\\}_k$ . This process is repeated from $k = 0$ to $N-1$ . Next, we define a zero tensor $Z_k \\in \\mathbb{R}^{N \\times C \\times H \\times W}$ and populate it with temporal unit event features for each timestep:", + "bbox": [ + 89, + 219, + 483, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nZ _ {k} [ k,:,:,: ] = \\mathcal {F} \\left(E _ {k} ^ {t}\\right), \\quad \\forall k \\in \\{0, 1, \\dots , N - 1 \\}.\n$$\n", + "text_format": "latex", + "bbox": [ + 120, + 474, + 450, + 492 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This tensor $Z_{k}$ contains unit event voxel features $\\mathcal{F}(E_k^t)$ filled only at the $k$ -th position along the temporal axis, with the rest initialized to zero. As a result, it encapsulates information about which timestep is being interpolated and the motion information required for the interpolation. Afterward, for $k = \\{0,\\dots ,N - 1\\}$ , we concatenate the $Z_{k}$ tensor with $\\mathcal{F}(\\mathcal{K})_{s = 0}^{t}$ , pass them through a Residual blocks, and generate the template features $\\{\\mathcal{F}(\\mathcal{T})_k\\}$ for $k$ timesteps.", + "bbox": [ + 89, + 505, + 483, + 626 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "IS-TFF based Bidirectional Encoder Module. Through the template feature generator, we obtain the feature $\\mathcal{F}(\\mathcal{T})_k$ for each timestep $k$ , capturing the essential information to interpolate the normal light sharp frame $S_{k}$ . By using these features, in this module, we aim to enhance the temporal dependency for timestep $k$ within the shutter interval $t$ using multi-scale temporal information.", + "bbox": [ + 89, + 627, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "While existing methods [4, 46] enhance temporal dependency within frames, they are computationally intensive, especially when applied to multiple timesteps at various scales. To address this, we propose a lightweight yet effective temporal feature fusion method within the shutter interval, which enables information sharing across the temporal dimension. As shown in Fig. 3, the IS-TFF-based encoder modules operate bidirectionally at each visual scale $(s = 0,1,2)$ . For clarity, we describe the process at scale $s$ in the forward direction; the same applies to the backward direction and other scales.", + "bbox": [ + 89, + 734, + 483, + 898 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As shown in Fig. 5, we first extract the features through $\\mathcal{F}(\\mathcal{T})_{k,s}^{fw}$ processed at the previous scale, passing them through a conv. layer to transform them into $\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}$ . For the fusion of timesteps $k$ and $k - 1$ in forward direction, the fusion of $\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}$ and $\\mathcal{G}(\\mathcal{T})_{s,k - 1}^{\\mathrm{fw}}$ begins by applying a conv. layer with LeakyReLU to their channel-wise concatenation, producing the feature $\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}$ . Then, this feature is split along the channel dimension into two parts, $\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}$ and $\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k - 1}$ . After mixing the adjacent features through concatenation and the first conv. layer, we analyze their temporally significant parts through the channel-wise dimension. For this, we first apply AAP (adaptive average pooling) to compress the spatial dimensions, as follows: $\\mathcal{P}(\\mathcal{G})_{s,k}^{\\mathrm{fw}} = \\mathrm{AAP}(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}})$ . Next, we apply attention [58] to the feature at each timestep $k$ and $k - 1$ as follows:", + "bbox": [ + 511, + 90, + 906, + 325 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = \\tilde {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} \\cdot \\sigma \\left(W _ {3} \\left(\\operatorname {R e L U} \\left(W _ {3} \\left(\\mathcal {P} (\\mathcal {G}) _ {s, k} ^ {\\mathrm {f w}}\\right)\\right)\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 333, + 903, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{3}$ denotes a $3 \\times 3$ conv. layer, ReLU represents the rectified linear unit activation function, and $\\sigma$ is the sigmoid activation function. Similarly, the feature $\\bar{\\mathcal{G}}(\\mathcal{T})_{s,k-1}^{\\mathrm{fw}}$ for the $k-1$ timestep is obtained in the same manner. Finally, we fuse the $k-1$ and $k$ timestep features as follows:", + "bbox": [ + 511, + 366, + 905, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}}\\right) + W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k - 1} ^ {\\mathrm {f w}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 446, + 903, + 467 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $W_{1}$ denotes a $1\\times 1$ conv. layer, we pass $\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}$ to the next IS-TFF encoder block with downsampling. This process is repeated for $s = \\{0,1,2\\}$ , producing outputs at three scales. This process is also repeated through a separate backward encoder, resulting in a set of feature maps $\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\}$ from the forward encoder pass and $\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{bw}}\\}$ from the backward encoder pass at timestep $k$ . We then fuse the forward and backward features at scale $s$ as follows.", + "bbox": [ + 511, + 474, + 905, + 595 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {D} (\\mathcal {T}) _ {s, k} = \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} + \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {b w}} \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 602, + 609, + 903, + 628 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "With this approach, we generate a multi-scale feature map set $\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}$ , where $s\\in \\{0,1,2\\}$ , that reinforces rich temporal information within the shutter interval $t$ .", + "bbox": [ + 511, + 631, + 905, + 676 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.4. Decoders", + "text_level": 1, + "bbox": [ + 511, + 684, + 622, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "After the IS-TFF based encoding stage, we decode the feature map for each timestep $k$ . For this, we apply two separate decoders, one for exposure time(generating $E$ frames) and another for readout-dead time(generating $R$ frames), share the same structure but have different weights. Consequently, we simultaneously estimate $N$ sharp frames $\\{S_0, \\ldots, S_{N-1}\\}$ .", + "bbox": [ + 511, + 705, + 905, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 823, + 645, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Datasets and Baseline Settings", + "text_level": 1, + "bbox": [ + 511, + 848, + 782, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To the best of our knowledge, no prior work has jointly tackled low-light video enhancement, frame interpolation,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "8529", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5b419bea1be519ed944f6d97033635dc291c34f3fe38fa4a9f45425fb39207a3.jpg", + "table_caption": [ + "Table 2. Quantitative evaluation results for the tasks of joint low-light enhancement, deblurring, and frame interpolation on the RELBFI dataset. Bold and underline denote the best and second-best performance, respectively. The symbol $\\dagger$ denotes event-guided methods. The number preceding the parameter corresponds to the deblurring method, while the number following it corresponds to the video frame interpolation methods. The runtime is defined as the time required to generate one or eight $1280 \\times 896$ normal-light sharp images, and was measured using an RTX 3090 GPU." + ], + "table_footnote": [], + "table_body": "
Method\nDeblur+VFI / JointPSNR↑SSIM↑Params(MB)↓Runtime
1×[s]↓8×[s]↓
LEDNet [65] + TimeLens† [44]24.420.8017.41 + 72.30.3622.894
LEDNet [65] + CBMNet† [44]24.550.8027.41 + 15.40.9017.204
EFNet† [40] + TimeLens† [44]27.040.8438.47 + 72.30.3773.018
EFNet† [40] + CBMNet† [44]27.420.8528.47 + 15.40.9167.328
ELEDNet† [16] + TimeLens† [44]28.650.86312.8 + 72.30.4863.892
ELEDNet† [16] + CBMNet† [44]29.280.87712.8 + 15.41.0258.202
DEMFI [30]24.220.7947.411.2019.606
BiT [63]26.560.83111.30.2341.875
REFID† [41]29.490.87615.90.2071.656
Ours†31.620.9029.40.1481.184
", + "bbox": [ + 124, + 162, + 870, + 359 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2510371be0efdf9e4f01fab5c96b6a6b50263d651d5eda71000c54a215c2dc32.jpg", + "table_caption": [ + "Table 3. Quantitative evaluation results for the joint low-light enhancement and deblurring tasks on the RELBFI datasets." + ], + "table_footnote": [], + "table_body": "
MethodsPSNRSSIMParams(MB)
LEDNet [65]24.850.8077.41
EFNet [40]27.630.8628.47
DEMFI [30]24.150.8057.41
BiT [63]24.690.81211.3
REFID [41]29.080.87815.9
ELEDNet [16]30.480.89912.8
Ours31.650.9079.4
", + "bbox": [ + 94, + 401, + 478, + 541 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and deblurring. Therefore, we established our own baselines and datasets for comprehensive benchmarking. Since no existing dataset includes high-speed ground-truth frames captured under normal lighting, low-frame-rate low-light blurry video frames, and low-light events, we conducted all experiments using our RELBFI dataset. The RELBFI dataset consists of 67 sequences, with 47 used for training and 20 reserved for testing. We established two types of baselines. The first is a cascaded approach, where deblurring and video frame interpolation methods are applied sequentially. For joint low-light enhancement and deblurring, we utilized LEDNet [65] (Frame-based), ELEDNet [16] (Event-based), and EFNet [40] (Event-based). These methods were trained for the same number of epochs as our network. For video frame interpolation (VFI), we used widely adopted event-based methods, Timelens [44] and CBMNet [13]. As these methods typically take longer to converge compared to our deblurring network, we trained them for additional epochs to achieve sufficient convergence. Finally, we evaluated state-of-the-art joint deblurring and frame interpolation methods, including DEMFI [30] (Frame-based), BiT [63] (Frame-based), and REFID [41] (Event-based). These methods were also", + "bbox": [ + 88, + 551, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/23a42d7d3afe8301ad107ec79cc9b6cb4b72e9d62df41d90c2782ed69caec610.jpg", + "table_caption": [ + "Table 4. Ablation study on the RELBFI datasets." + ], + "table_footnote": [], + "table_body": "
MethodsMM-ISTFFIS-TFF(fw)IS-TFF(bw)PSNRs
Ver.127.07
Ver.2✓30.81
Ver.3✓✓31.35
Ver.4✓✓✓31.62
", + "bbox": [ + 519, + 387, + 895, + 474 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ba58381fda17cbbf28d99a2f936d1f53c0e9b27487c1873f39236989da76bb2b.jpg", + "table_caption": [ + "Table 5. Performance effectiveness analysis of the MM-ISTFF module. The (-) symbol indicates that training was unstable and did not converge." + ], + "table_footnote": [], + "table_body": "
Methodsw/o Eq.(1)w/o Eq.(2)w/o BiOurs
PSNRs30.84(-)31.0431.62
", + "bbox": [ + 517, + 529, + 898, + 565 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3b796851b7ce5dd512703903aca24a5d6394b8270412f08f1abac831dad4e6b9.jpg", + "table_caption": [ + "Table 6. Network component analysis of the dual decoder module." + ], + "table_footnote": [], + "table_body": "
MethodsSingle weight-sharing decoderOurs
PSNRs31.1631.62
", + "bbox": [ + 519, + 589, + 895, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "trained for the same number of epochs as our network.", + "bbox": [ + 511, + 642, + 872, + 657 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Experimental results on the RELBFI datasets", + "text_level": 1, + "bbox": [ + 511, + 667, + 898, + 683 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We have provided the results of the quantitative evaluation on the RELBFI dataset in Tab.3(Joint LLE and deblurring) and Tab.2(Joint LLE and deblurring and interpolation).", + "bbox": [ + 511, + 688, + 906, + 734 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Joint LLE and deblurring and interpolation In Tab.2, we evaluated two groups. Cascaded solutions, where joint deblurring and LLE methods are followed by video frame interpolation (VFI), showed significant performance degradation in frame-based methods like LEDNet [65]. Performance improved with event-based networks such as EFNet [40] and ELEDNet [16], but these cascaded solutions still had a PSNR gap of 2.34 dB to 4.58 dB compared to our network. For joint-based solutions, frame-based methods like DEMFI [30] and BiT [63] showed a PSNR difference of 5.06 dB to 7.40 dB, while compared to REFID [41],", + "bbox": [ + 511, + 734, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "8530", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae680d372260006a55bc1ec40bdcb59114c0e7772538d45f39440404bc3ffebe.jpg", + "image_caption": [ + "Figure 6. Qualitative results on the RELBFI dataset. From left to right, the images show (a) input low-light blurry images with brightness adjustment, (b) low-light events, (c) BiT [63], (d) REFID [41], (e) Ours, and (f) ground-truth normal-light sharp images." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 898, + 368 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "the state-of-the-art joint event-based method, our network outperformed by 2.13 dB with a large margin. From the perspective of computational cost, our method achieves the best performance with fewer computations and reasonable parameters compared to all state-of-the-art methods.", + "bbox": [ + 88, + 417, + 480, + 492 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Joint LLE and deblurring In Tab.3, we conducted an evaluation of the joint low-light enhancement and deblurring method, excluding interpolation. When compared to ELEDNet, the best state-of-the-art method, it showed a performance difference of 1.17 dB. Comparisons with the other frame-based and event-based methods revealed differences ranging from a maximum of 7.5 dB to a minimum of 2.57 dB. This also evidences the robustness of the proposed method in both deblurring and low-light enhancement (LLE) tasks.", + "bbox": [ + 88, + 493, + 482, + 643 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 652, + 243, + 667 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "All ablation experiments were conducted on the RELBFI dataset, with networks trained for 200 epochs. The results are summarized in Tab. 4. For the baseline model, the blur input feature $\\mathcal{F}(B)^t$ at the $t$ -shutter interval was used as input for the IS-TFF module-based networks (stage 2) instead of $\\mathcal{F}(\\mathcal{K})_{s=0}^t$ from the MM-ISTFF module.", + "bbox": [ + 89, + 672, + 482, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "MM-ISTFF module The MM-ISTFF module extracts discriminative spatio-temporal information from multi-modal data. As shown in Tab. 4, Ver. 2 outperforms Ver. 1 by $3.74\\mathrm{dB}$ . In Tab. 5, we analyze MM-ISTFF components. Removing the pixel-wise filter from Eq. (1) caused a 0.78 dB drop, while removing regional feature selection from Eq. (2) led to unstable training and non-convergence, underscoring its importance for region-wise selection of degraded multi-modal data. Limiting temporal fusion to current and", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "future parts (third row) resulted in a $0.50\\mathrm{dB}$ drop, highlighting the benefits of bidirectional fusion.", + "bbox": [ + 511, + 417, + 903, + 448 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "IS-TFF module We validated the effectiveness of the IS-TFF module. Comparing Ver. 2 and Ver. 3 in Tab. 4, incorporating IS-TFF in the forward direction improved performance by $0.5\\mathrm{dB}$ . Adding the IS-TFF module in the backward direction (Ver. 3 vs. Ver. 4) yielded an additional $0.33\\mathrm{dB}$ gain, resulting in a total improvement of $0.83\\mathrm{dB}$ .", + "bbox": [ + 511, + 450, + 905, + 542 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The importance of the separate dual decoders In the decoder following the IS-TFF encoder, separate decoders with different weights are used for frames from the exposure time and the readout deadtime. As shown in Tab. 6, this improves performance $(+0.46\\mathrm{dB})$ by applying different decoding methods—blur decomposition or interpolation—tailored to the exposure and readout periods, respectively.", + "bbox": [ + 511, + 544, + 906, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusions", + "text_level": 1, + "bbox": [ + 511, + 679, + 638, + 695 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper addresses, for the first time, the novel and previously unexplored problem of unified event-guided joint low-light enhancement, deblurring, and frame interpolation. To solve this task, we propose an integrated framework that jointly handles these sub-tasks with high precision, achieving state-of-the-art performance in both accuracy and computational efficiency. We expect our work to offer initial insights that may support future research in event-guided video restoration.", + "bbox": [ + 511, + 700, + 906, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. NRF-2022R1A2B5B03002636).", + "bbox": [ + 511, + 839, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "8531", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[2] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 2", + "[3] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 1, 3", + "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 6", + "[5] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12492-12503, 2023. 2", + "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2", + "[7] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12345-12354, 2023. 2", + "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2", + "[9] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 3", + "[10] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2", + "[11] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019. 3", + "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos." + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 4", + "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 3, 7", + "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. arXiv preprint arXiv:2408.14930, 2024. 2", + "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2", + "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 1, 2, 3, 4, 7", + "[17] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2", + "[18] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2", + "[19] Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial-temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9822-9832, 2023. 2", + "[20] Guoqiang Liang, Kanghao Chen, Hangyu Li, Yunfan Lu, and Lin Wang. Towards robust event-guided low-light image enhancement: A large-scale real-world event-image dataset and novel approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23-33, 2024. 2, 3", + "[21] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2", + "[22] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2", + "[23] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2, 3", + "[24] Lin Liu, Junfeng An, Jianzhuang Liu, Shanxin Yuan, Xi-angyu Chen, Wengang Zhou, Houqiang Li, Yan Feng Wang," + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "8532", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Qi Tian. Low-light video enhancement with synthetic event guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1692-1700, 2023. 3", + "[25] Yunlong Liu, Tao Huang, Weisheng Dong, Fangfang Wu, Xin Li, and Guangming Shi. Low-light image enhancement with multi-stage residue quantization and brightness-aware attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12140-12149, 2023. 3", + "[26] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[27] Yongrui Ma, Shi Guo, Yutian Chen, Tianfan Xue, and Jinwei Gu. Timelens-xl: Real-time event-based video frame interpolation with large motion. In European Conference on Computer Vision, pages 178-194. Springer, 2025. 3", + "[28] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1905-1913, 2023. 2", + "[29] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 2", + "[30] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision (ECCV), 2022. 1, 3, 7", + "[31] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1", + "[32] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22191-22200, 2023. 2", + "[33] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3", + "[34] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3", + "[35] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 2", + "[36] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 2", + "[38] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2", + "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 2", + "[40] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2, 7", + "[41] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. arXiv preprint arXiv:2301.05191, 2023. 2, 3, 4, 7, 8", + "[42] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 2", + "[43] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 3", + "[44] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 7", + "[45] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-based frame interpolation with nonlinear parametric flow and multi-scale fusion. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3", + "[46] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 6", + "[47] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13128-13137, 2023. 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "8533", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 3", + "[49] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 2", + "[50] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 3", + "[51] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3", + "[52] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2022. 3", + "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 1, 3", + "[54] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 1, 3", + "[55] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 3", + "[56] Wen Yang, Jinjian Wu, Jupo Ma, Leida Li, and Guangming Shi. Motion deblurring via spatial-temporal collaboration of frames and events. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6531-6539, 2024. 2", + "[57] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 3", + "[58] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6", + "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2", + "[60] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for im" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "age deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2", + "[61] Huicong Zhang, Haozhe Xie, and Hongxun Yao. Blur-aware spatio-temporal sparse transformer for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2673-2681, 2024. 2", + "[62] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 191-207. Springer, 2020. 2", + "[63] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 1, 3, 4, 7, 8", + "[64] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 5", + "[65] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 3, 7", + "[66] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 5" + ], + "bbox": [ + 516, + 92, + 906, + 487 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "8534", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_model.json b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf514db366c5377c189dbeb7bcad9a13d5143dcf --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_model.json @@ -0,0 +1,2422 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.807, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.116, + 0.13, + 0.882, + 0.177 + ], + "angle": 0, + "content": "Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring" + }, + { + "type": "text", + "bbox": [ + 0.329, + 0.204, + 0.441, + 0.238 + ], + "angle": 0, + "content": "Taewoo Kim* KAIST" + }, + { + "type": "text", + "bbox": [ + 0.293, + 0.243, + 0.471, + 0.256 + ], + "angle": 0, + "content": "intelpro@kaist.ac.kr" + }, + { + "type": "text", + "bbox": [ + 0.566, + 0.204, + 0.68, + 0.238 + ], + "angle": 0, + "content": "Kuk-Jin Yoon KAIST" + }, + { + "type": "text", + "bbox": [ + 0.539, + 0.243, + 0.702, + 0.257 + ], + "angle": 0, + "content": "kjyoon@kaist.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.324, + 0.485, + 0.716 + ], + "angle": 0, + "content": "In low-light environments, longer exposure times are commonly used to enhance image visibility; however, this inevitably leads to motion blur. Even with a long exposure time, videos captured in low-light environments still suffer from issues such as low visibility, low contrast, and color distortion. Additionally, the long exposure time results in videos with a low frame rate. Therefore, videos captured in low-light exhibit low visibility and motion blur, as well as low frame rates. To overcome these limitations, we propose a novel problem aimed at transforming motion-blurred, low-frame-rate videos with poor visibility in low-light environments into high-frame-rate videos while simultaneously enhancing their visibility. To tackle this challenge, we leverage the unique advantages of event cameras, which capture scene changes asynchronously, providing superior temporal resolution and a wider dynamic range compared to conventional frame-based cameras. These properties make event cameras particularly effective in reducing motion blur, compensating for low frame rates, and enhancing visibility in low-light conditions. To this end, we developed a hybrid camera system that integrates two RGB cameras and an event camera, capturing a dedicated dataset for this task and proposing novel network architectures to effectively address this problem. The project pages are available at https://sites.google.com/view/eledi-iccv." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.746, + 0.222, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.483, + 0.879 + ], + "angle": 0, + "content": "In our daily lives, low-light conditions are common and occur in various lighting scenarios, both during the day and at night. In such environments, a longer exposure time is generally required to capture the limited available light. However, this extended exposure time inevitably leads to motion blur caused by object movement or dynamic camera motion, while also reducing video frame rates. Additionally," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.293, + 0.907, + 0.384 + ], + "angle": 0, + "content": "videos recorded under these conditions often suffer from reduced visibility due to insufficient lighting, along with significant image noise, even with a long exposure time. As a result, videos captured in low-light conditions typically exhibit low frame rates due to the long exposure time, as well as poor visibility and motion blur." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.909, + 0.657 + ], + "angle": 0, + "content": "With advancements in deep learning, numerous efforts have focused on improving low-light conditions [3, 53, 54], aiming to suppress noise and enhance visibility in such environments. Furthermore, research has been conducted on removing motion blur artifacts through motion deblurring studies [6, 30, 31], as well as on Video Frame Interpolation (VFI) [13, 36, 38, 44], which converts low-frame-rate videos into high-frame-rate ones. While each of these areas has been explored independently, few studies address the combined challenges of multiple degradations in low-light conditions. Among existing works, relatively active research has focused on the problem of joint deblurring and interpolation [30, 63], which aims to restore high-frame-rate, sharp frames from motion-blurred ones. Additionally, some works [16, 65] have investigated integrating motion deblurring and low-light enhancement. However, no existing approach comprehensively tackles all three degradations simultaneously." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.658, + 0.909, + 0.885 + ], + "angle": 0, + "content": "To overcome these limitations, we propose a research problem that simultaneously addresses the challenges of low frame rate, motion blur, and low visibility in low-light conditions. Specifically, we aim to simultaneously enhance visibility, reduce motion blur, and increase the frame rate to improve overall video quality in low-light environments. A simple approach would involve cascading separate models for each task. However, such a cascaded manner suffers from error accumulation and inefficient joint optimization, often leading to suboptimal results. Thus, a more effective solution is needed to jointly optimize these interrelated tasks in an end-to-end manner. Moreover, solving these problems using only frame-based information is extremely challenging, particularly in cases of severe degradation, making it an ill-posed problem." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.904, + 0.902 + ], + "angle": 0, + "content": "To tackle this issue, we propose leveraging the ad" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.266, + 0.9 + ], + "angle": 0, + "content": "*Work done while at KAIST." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8524" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.38, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.165, + 0.246, + 0.297, + 0.258 + ], + "angle": 0, + "content": "Low-light blurred image" + }, + { + "type": "image", + "bbox": [ + 0.38, + 0.089, + 0.51, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.428, + 0.246, + 0.466, + 0.258 + ], + "angle": 0, + "content": "t=5/8" + }, + { + "type": "image", + "bbox": [ + 0.51, + 0.089, + 0.639, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.554, + 0.246, + 0.592, + 0.257 + ], + "angle": 0, + "content": "t=6/8" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.089, + 0.768, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.683, + 0.246, + 0.721, + 0.258 + ], + "angle": 0, + "content": "t=7/8" + }, + { + "type": "image", + "bbox": [ + 0.768, + 0.089, + 0.898, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.815, + 0.246, + 0.854, + 0.257 + ], + "angle": 0, + "content": "t=8/8" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.266, + 0.908, + 0.294 + ], + "angle": 0, + "content": "Figure 1. Qualitative result of our method on the RELBFI datasets. Our task aims to restore sharp frames captured in a normal light environment within the shutter interval using low-light blurred images and the events." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.314, + 0.482, + 0.48 + ], + "angle": 0, + "content": "vantages of event cameras, which, inspired by neuromorphic vision systems, detect brightness changes with microsecond-level temporal resolution [8]. Additionally, their superior dynamic range, compared to conventional cameras, allows for high-speed motion capture even in low-light environments. As a result, event cameras are widely used in video enhancement tasks, including motion deblurring [40, 41], frame interpolation [13, 44], low-light enhancement [16, 20, 23], and joint deblurring and interpolation [41], effectively addressing issues like motion blur, low frame rates, low visibility." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.486, + 0.483, + 0.669 + ], + "angle": 0, + "content": "Therefore, we propose a novel task that leverages event cameras to address these multiple degradations in low-light conditions. However, since this novel task has never been addressed in previous works, it is necessary to build a new dataset. To overcome these challenges, we build a sophisticated hybrid camera system to capture a new RELBFI(Real-world Event-guided Low-light enhancement and Blur Frame Interpolation) dataset, consisting of real-world low-light blur videos with a low frame rate, low-light event data, and high-frame-rate normal light sharp videos (ground truth). We believe that these resources will be a valuable asset to the research community." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Beyond proposing this dataset, we also introduce a novel unified end-to-end framework to tackle this challenging task. To achieve this, we design two key modules that play a crucial role in effectively addressing the problem. The first module focuses on fusing multi-temporal information from events and frames under low-light conditions. Given that video frames in such environments often suffer from significant noise and color distortion, this module aims to efficiently integrate multi-temporal and multi-modal information sources. For the second module, we introduce a mechanism to reinforce temporal dependencies among generated video frames during the generation of multiple frames within the shutter period. Since this process is computationally intensive, we designed a lightweight yet effective module to strengthen bidirectional dependencies between" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.314, + 0.811, + 0.329 + ], + "angle": 0, + "content": "frames, thereby enhancing overall efficiency." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.495 + ], + "angle": 0, + "content": "To summarize, the contributions of this paper are as follows: (I) We propose a new research problem that addresses multiple degradations, including noise, motion blur, low visibility, and low frame rate, which commonly occur in low-light conditions, while also converting videos captured at low frame rates into high frame rates. (II) We present a novel unified framework that effectively addresses this problem, demonstrating at least a 2 dB improvement in PSNR compared to state-of-the-art methods. (III) We introduce the first RELBFI dataset, built using a hybrid camera system." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.507, + 0.663, + 0.523 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.702, + 0.548 + ], + "angle": 0, + "content": "2.1. Motion Deblurring." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.907, + 0.796 + ], + "angle": 0, + "content": "Since the advent of deep learning, significant progress has been made in motion deblurring, further accelerated by specialized datasets such as GoPro [29], DVD [39], RealBlur [37], and BSD [62]. In frame-based deblurring, various architectures have been proposed, including multiscale [6], multi-stage [58, 60], transformer-based [42, 49, 59], and frequency-modulated [7, 17, 28] architectures. Video deblurring methods [19, 21, 22, 32, 61, 62] have also emerged. However, deblurring solely from frames remains challenging due to motion information loss, prompting event camera-based research. Early works explored recurrent architectures [10], followed by studies on cross-modal attention-based feature fusion [40], unknown exposure time [12], non-coaxial deblurring [5], frequency-based video deblurring [15], and attention-based cross-modal alignment [14] and spatio-temporal collaboration [56]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.761, + 0.82 + ], + "angle": 0, + "content": "2.2. Video Frame Interpolation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Research on frame-based video interpolation [1, 2, 18, 26, 35, 36, 38] has made significant progress, but it remains challenging due to the absence of explicit motion information between frames. In contrast, event cameras, with their high temporal resolution, have gained popularity in" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8525" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.092, + 0.378, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.24, + 0.306, + 0.259, + 0.319 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.39, + 0.098, + 0.614, + 0.301 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.485, + 0.306, + 0.504, + 0.319 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image", + "bbox": [ + 0.635, + 0.098, + 0.876, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.306, + 0.76, + 0.318 + ], + "angle": 0, + "content": "(c)" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.324, + 0.906, + 0.382 + ], + "angle": 0, + "content": "Figure 2. In the RELBFI dataset description, (a) shows a diagram of the hybrid camera system used to capture the dataset. (b) represents the temporal irradiance relationship for the high-fps RGB camera (used for recording sharp images) and the camera used for capturing blur. (c) presents sample images from the RELBFI dataset: from top to bottom, a low-light blur image, a brightness-adjusted version of the low-light blur image, and a ground truth image captured in normal light conditions." + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.398, + 0.483, + 0.455 + ], + "angle": 0, + "content": "Table 1. Comparison of HighREV and RELBFI Datasets. In the table, NL\" refers to images taken in normal-light environments, and \"LL\" refers to images taken in low-light environments. \"LFR\" refers to low frame rate, and \"HFR\" refers to high frame rate." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.459, + 0.458, + 0.585 + ], + "angle": 0, + "content": "
HighREV [41]RELBFI
LFR RGB InputNL blurLL blur
Event inputNL eventLL event
HFR RGB GTNL sharpNL sharp
Blur typeSyntheticReal
Event typeRealReal
#N of sharp images2858932629
#N of sequences3067
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.484, + 0.759 + ], + "angle": 0, + "content": "video interpolation tasks. Early work like TimeLens [43] introduced event-based interpolation using a cascaded U-Net framework, while TimeLens++ [45] enhanced this by estimating non-linear motion with a motion spline estimator. Later studies, including A\\(^2\\)OF [51] and the unsupervised TimeReplayer [9], further advanced the field. Recently, Kim et al. [13] introduced an event-image fusion VFI framework for inter-frame optical flow estimation, while the follow-up study, TimeLens-XL [27], presented a VFI framework with iterative flow updates." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.771, + 0.317, + 0.789 + ], + "angle": 0, + "content": "2.3. Low-light Enhancement." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Many frame-based low-light enhancement (LLE) studies focus on improving visibility and reducing noise using image information. These efforts have led to various approaches, including SNR-based transformers [53], Retinex-based methods [3, 52], gamma correction [47], implicit neural representations [55], diffusion models [48, 57], and other techniques [25, 54, 65]. Leveraging its high dynamic range," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.507 + ], + "angle": 0, + "content": "the event camera excels in capturing motion under low light, making it suitable for LLE. Liu et al. [24] used synthetic events for LLE due to dataset limitations, while Liang et al. [23] modeled multimodal coherence. Recently, Liang et al. [20] proposed an SNR-based framework and dataset for event-guided LLE, and Kim et al. [16] introduced a joint framework for event-guided LLE and deblurring." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.516, + 0.868, + 0.532 + ], + "angle": 0, + "content": "2.4. Joint deblurring and frame interpolation." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.538, + 0.907, + 0.75 + ], + "angle": 0, + "content": "The joint task of deblurring and interpolation aims to simultaneously remove blur and generate high-frame-rate videos within a single framework. Jin et al. [11] first addressed this without event cameras using a deblur-then-interpolate approach. Recently, DeMFI [30] combined flow-guided attention and recursive boosting, while BiT [63] introduced a Swin Transformer-based solution. Event camera-based research includes EDI [33, 34], which used a physical model, REFID [41] leveraging a recurrent architecture, and Weng et al. [50], tackling blur frame interpolation under unknown exposure times. Unlike these works, we leverage the robust properties of event cameras in low-light conditions to jointly perform low-light enhancement, deblurring, and frame interpolation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.763, + 0.688, + 0.779 + ], + "angle": 0, + "content": "3. RELBFI Datasets" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.815, + 0.804 + ], + "angle": 0, + "content": "3.1. Comparisons with Other Datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.906, + 0.87 + ], + "angle": 0, + "content": "Since the advent of deep learning, various event- and frame-based video enhancement datasets [13, 16, 20, 63] have emerged. Some studies [41, 63] focus on motion deblurring and frame interpolation under normal lighting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Sun et al. [41] introduced a dataset for joint deblurring and interpolation using an event camera, where motion blur" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8526" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.885, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.367, + 0.866, + 0.382 + ], + "angle": 0, + "content": "Figure 3. Overall framework of the proposed method. \"TFG\" in the blue box denotes the Template Feature Generator block." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.482, + 0.461 + ], + "angle": 0, + "content": "was synthetically generated by merging adjacent frames. However, such synthetic datasets fail to capture real-world motion blur, as they overlook physical factors like noise variations from different exposure times." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.462, + 0.483, + 0.537 + ], + "angle": 0, + "content": "Recently, Kim et al. [16] proposed the RELED dataset for simultaneous low-light enhancement and deblurring using an event camera. However, its low frame rate limits its effectiveness in converting low-frame-rate, low-light videos into high-frame-rate, normal-light videos." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.538, + 0.483, + 0.599 + ], + "angle": 0, + "content": "To date, no existing research or dataset has directly tackled the transformation of low-frame-rate, low-light, noisy, and motion-blurred inputs into high-frame-rate, noise-free videos under normal lighting by leveraging event data." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.607, + 0.354, + 0.623 + ], + "angle": 0, + "content": "3.2. RELBFI dataset acquisitions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.629, + 0.483, + 0.704 + ], + "angle": 0, + "content": "Synthetic-based datasets struggle to capture the complexity of real-world degradations. Therefore, building a real-world dataset is both practical and challenging, especially when simultaneously capturing high-frame-rate sharp video, low-light blurred video, and low-light events." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address this, we build the RELBFI (Real-world Event-guided Low-light Video Enhancement and Blur Frame Interpolation) dataset using a hybrid camera system with beam splitters. As shown in Fig. 2(a), the system includes two RGB cameras (Blackfly BFS-U3-16S2C-CS) and an event camera (Propheesee Gen4). One RGB camera captures sharp images at 166fps with a 3ms exposure and a 3ms readout dead time, receiving \\(25\\%\\) of the total light as it passes through the beam splitter twice without an ND filter. The other RGB camera records blurred images in low-light conditions with a 24ms exposure, eight times longer, and uses a 1/64 ND filter, receiving only \\(0.39\\%\\) of the original light. As a result, each blurred image has eight times less" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.402, + 0.905, + 0.417 + ], + "angle": 0, + "content": "light intensity than a sharp image, resulting in low visibility." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.905, + 0.552 + ], + "angle": 0, + "content": "Furthermore, to illustrate real-world frame acquisition [12, 41, 63], capturing blurred images involves exposure time and readout dead time, collectively termed the shutter interval [12]. To simulate readout dead time, we used a \\(24\\mathrm{ms}\\) exposure followed by a \\(24\\mathrm{ms}\\) dead time, yielding one low-light blurred image per \\(48\\mathrm{ms}\\) (21fps) and eight sharp images under normal lighting (166fps). The event camera, like the blur-recording RGB camera, employs a 1/64 ND filter for low-light capture." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.765 + ], + "angle": 0, + "content": "Regarding calibration, all cameras in the hybrid system are mechanically fixed, but slight adjustments, such as rotation and baseline alignment, are required. To this end, we estimated the homography matrix for each camera and performed additional calibration. As shown in Fig. 2(a), the devices were temporally synchronized via hardware using an external trigger. Using this setup, we captured 67 urban scenes featuring dynamic elements like vehicles, moving objects, and flickering streetlights, along with rich textures. As shown in Tab. 1, RELBFI provides diverse sequences tailored for developing and evaluating methods in this challenging task. It serves as a practical resource to support joint low-light enhancement, deblurring, and interpolation in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.779, + 0.695, + 0.795 + ], + "angle": 0, + "content": "4. Proposed Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.804, + 0.625, + 0.818 + ], + "angle": 0, + "content": "4.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Preliminary As explained in Sec. 3, the shutter interval consists of exposure time and readout deadtime. During the readout deadtime, events are recorded, but no blurred frames are captured. We define \\(N\\) as the total number of sharp frames captured during a shutter interval \\(T\\) (ms)," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8527" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.087, + 0.48, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.273, + 0.483, + 0.318 + ], + "angle": 0, + "content": "Figure 4. The MM-ISTFF module in the scale factor \\( s \\). In the figure, \\( \\mathcal{D}_s^{\\mathrm{fw}} \\) and \\( \\mathcal{D}_s^{\\mathrm{bw}} \\) represent the forward and backward dynamic filters, respectively, as defined in Eq.(1)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.333, + 0.484, + 0.439 + ], + "angle": 0, + "content": "where \\(N = E + R\\), with \\(E\\) and \\(R\\) representing the number of the frames from exposure and readout deadtime, respectively. During a single shutter interval \\(t\\), one blurred frame \\(B^t\\) is captured, and events are generated corresponding to both the exposure time and readout deadtime. The events \\(\\{e_i\\}_t\\) during shutter interval \\(t\\) are transformed into event voxels [66] with 32 bins." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.53 + ], + "angle": 0, + "content": "Overall framework The proposed framework is illustrated in Fig. 3. We utilize 3 blurred images and 3 corresponding event voxel grids from shutter intervals \\( t - 1 \\), \\( t \\), and \\( t + 1 \\). Given these inputs \\( \\{B^{t-1}, B^t, B^{t+1}\\} \\) and \\( \\{E^{t-1}, E^t, E^{t+1}\\} \\), the network estimates \\( N \\) normal light sharp frames at the shutter interval \\( t \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.748 + ], + "angle": 0, + "content": "The process consists of two main steps. First, the MM-ISTFF (Multi-Modal Inter-frame Spatio-Temporal Feature Fusion) module combines spatio-temporal event data with blurred video frames to improve video quality and reduce low-light noise, producing fused features \\(\\mathcal{F}(\\mathcal{K})_{s=0}^{t}\\). Next, the event voxel grid for \\(t\\) is divided into \\(N\\) temporal units \\(\\{E_0^t,\\ldots,E_{N-1}^t\\}\\), which are combined with \\(\\mathcal{F}(\\mathcal{K})_{s=0}^{t}\\) to generate \\(N\\) temporal template features \\(\\{\\mathcal{F}(\\mathcal{T})_0,\\ldots,\\mathcal{F}(\\mathcal{T})_{N-1}\\}\\). Next, the IS-TFF (Inter-Shutter Interval Temporal Feature Fusion) based-encoder module leverages temporal information within the shutter interval then uses these features to produce an enhanced feature map \\(\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}\\). Finally, the decoder generates \\(N\\) normal light sharp frames \\(\\{S_0,\\ldots,S_{N-1}\\}\\) for the \\(t\\) shutter interval." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.757, + 0.483, + 0.789 + ], + "angle": 0, + "content": "4.2. Multi-Modal Inter-frame Spatio-Temporal Feature Fusion(MM-ISTFF) Module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In low-light conditions, frame-based information is severely affected by noise and low visibility, while motion blur further exacerbates the degradation. In such scenarios, events also suffer from significant degradation. Therefore, it is crucial to fuse features by complementing the spatio-temporal visual information from images with the rich temporal details from events, effectively utilizing them" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.903, + 0.313 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.318, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Figure 5. Illustration of the IS-TFF module. The \\(s\\) after the forward and backward units indicates the scale factor." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.907, + 0.706 + ], + "angle": 0, + "content": "for video enhancement. To address this, we propose the MM-ISTFF module, designed to mine discriminative temporal information between frames and events across different shutter intervals. For this, we extract multi-scale multi-modal features \\(\\{\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(B)_s^{t + 1}\\}\\) and \\(\\{\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t,\\mathcal{F}(E)_s^{t + 1}\\}\\) from the low-light blurred and event voxel inputs. As shown in Fig.3, we aim to leverage these multi-modal inputs to extract discriminative spatio-temporal information across multiple visual scales. To achieve this, we use past and future information from the current shutter interval \\(t\\). For simplicity, only the process for past information is described, as the same applies to future information. As shown in Fig.4, to integrate spatiotemporal information from past multi-modal data, we first apply a conv. layer with LeakyReLU to the concatenated features \\([\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t]\\), generating the feature \\(\\mathbf{V}_s^{\\mathrm{fw}}\\). To extract discriminative spatio-temporal information, pixel-wise spatial dynamic filters [64] are applied to obtain content-adaptive features. Here, to estimate a dynamic filter for each pixel, we use \\(f_{s}^{fw} = f_{\\mathrm{filter}}^{\\mathrm{fw}}(\\mathbf{V}_{s}^{\\mathrm{fw}})\\) to generate dynamic filters, which are then applied to the original feature \\(\\mathbf{V}_s^{\\mathrm{fw}}\\), where \\(f_{\\mathrm{filter}}\\) We apply dynamic filter as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.719, + 0.907, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} = f _ {s} ^ {\\mathrm {f w}} \\otimes \\mathbf {V} _ {s} ^ {\\mathrm {f w}} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.743, + 0.906, + 0.82 + ], + "angle": 0, + "content": "where \\( f_{s}^{\\mathrm{fw}} \\) represents the estimated filter, and \\( \\otimes \\) denotes the dynamic filter application. This process extracts content-adaptive information, but in low-light conditions, selective regional features are needed due to noise and degradation. Thus, we apply region-wise attention for feature filtering." + }, + { + "type": "equation", + "bbox": [ + 0.576, + 0.827, + 0.907, + 0.848 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathbf {V}} _ {s} ^ {f w} = \\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} \\odot \\sigma (\\mathrm {L R e L U} ((W _ {3} (\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}})))) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "where \\(\\sigma\\) denotes the sigmoid operation, and \\(W_{3}\\) and LReLU represent the \\(3\\times 3\\) conv. layer and the LeakyReLU operation, respectively. Similarly, future multi-modal informa" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8528" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.09, + 0.482, + 0.169 + ], + "angle": 0, + "content": "tion is utilized to obtain \\(\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}}\\). These two features \\((\\tilde{\\mathbf{V}}_s^{\\mathrm{fw}}\\) and \\(\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}})\\) are combined via channel-wise concatenation followed by a ResBlock to produce the final feature \\(\\mathcal{F}(\\mathcal{K})_s^t\\). This process is repeated for \\(s = \\{0,1,2\\}\\), with \\(\\mathcal{F}(\\mathcal{K})_{s=0}^{t}\\) serving as input to the IS-TFF module." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.181, + 0.483, + 0.213 + ], + "angle": 0, + "content": "4.3. Inter-Shutter Interval Temporal Feature Fusion(IS-TFF) Module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.22, + 0.484, + 0.462 + ], + "angle": 0, + "content": "Template Feature Generation After the MM-ISTFF module, we obtain a spatio-temporally fused feature \\(\\mathcal{F}(\\mathcal{K})_{s=0}^{t}\\) centered at shutter interval \\(t\\). To estimate \\(N\\) sharp images within the shutter interval, template features capturing the characteristics of each timestep must be generated for all \\(N\\) timesteps. To address this, we first divide the event voxel grid into \\(N\\) temporal units, each with an equal time interval \\(\\Delta t_{unit} = T / N\\). This gives us temporally divided unit event voxels \\(E_k^t\\), which are then encoded to generate temporal event features. Therefore, we obtain the temporally divided unit event voxel \\(\\{E_k^t\\}\\) with temporal index \\(k \\in \\{0, \\dots, N-1\\}\\). Subsequently, each event voxel unit is encoded to generate temporal event features \\(\\{\\mathcal{F}(E_k^t) \\in \\mathbb{R}^{C \\times H \\times W}\\}_k\\). This process is repeated from \\(k = 0\\) to \\(N-1\\). Next, we define a zero tensor \\(Z_k \\in \\mathbb{R}^{N \\times C \\times H \\times W}\\) and populate it with temporal unit event features for each timestep:" + }, + { + "type": "equation", + "bbox": [ + 0.122, + 0.475, + 0.452, + 0.493 + ], + "angle": 0, + "content": "\\[\nZ _ {k} [ k,:,:,: ] = \\mathcal {F} \\left(E _ {k} ^ {t}\\right), \\quad \\forall k \\in \\{0, 1, \\dots , N - 1 \\}.\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.506, + 0.484, + 0.627 + ], + "angle": 0, + "content": "This tensor \\(Z_{k}\\) contains unit event voxel features \\(\\mathcal{F}(E_k^t)\\) filled only at the \\(k\\)-th position along the temporal axis, with the rest initialized to zero. As a result, it encapsulates information about which timestep is being interpolated and the motion information required for the interpolation. Afterward, for \\(k = \\{0,\\dots ,N - 1\\}\\), we concatenate the \\(Z_{k}\\) tensor with \\(\\mathcal{F}(\\mathcal{K})_{s = 0}^{t}\\), pass them through a Residual blocks, and generate the template features \\(\\{\\mathcal{F}(\\mathcal{T})_k\\}\\) for \\(k\\) timesteps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.734 + ], + "angle": 0, + "content": "IS-TFF based Bidirectional Encoder Module. Through the template feature generator, we obtain the feature \\(\\mathcal{F}(\\mathcal{T})_k\\) for each timestep \\(k\\), capturing the essential information to interpolate the normal light sharp frame \\(S_{k}\\). By using these features, in this module, we aim to enhance the temporal dependency for timestep \\(k\\) within the shutter interval \\(t\\) using multi-scale temporal information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.9 + ], + "angle": 0, + "content": "While existing methods [4, 46] enhance temporal dependency within frames, they are computationally intensive, especially when applied to multiple timesteps at various scales. To address this, we propose a lightweight yet effective temporal feature fusion method within the shutter interval, which enables information sharing across the temporal dimension. As shown in Fig. 3, the IS-TFF-based encoder modules operate bidirectionally at each visual scale \\((s = 0,1,2)\\). For clarity, we describe the process at scale \\(s\\) in the forward direction; the same applies to the backward direction and other scales." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.327 + ], + "angle": 0, + "content": "As shown in Fig. 5, we first extract the features through \\(\\mathcal{F}(\\mathcal{T})_{k,s}^{fw}\\) processed at the previous scale, passing them through a conv. layer to transform them into \\(\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\). For the fusion of timesteps \\(k\\) and \\(k - 1\\) in forward direction, the fusion of \\(\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\) and \\(\\mathcal{G}(\\mathcal{T})_{s,k - 1}^{\\mathrm{fw}}\\) begins by applying a conv. layer with LeakyReLU to their channel-wise concatenation, producing the feature \\(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}\\). Then, this feature is split along the channel dimension into two parts, \\(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}\\) and \\(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k - 1}\\). After mixing the adjacent features through concatenation and the first conv. layer, we analyze their temporally significant parts through the channel-wise dimension. For this, we first apply AAP (adaptive average pooling) to compress the spatial dimensions, as follows: \\(\\mathcal{P}(\\mathcal{G})_{s,k}^{\\mathrm{fw}} = \\mathrm{AAP}(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}})\\). Next, we apply attention [58] to the feature at each timestep \\(k\\) and \\(k - 1\\) as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.334, + 0.905, + 0.367 + ], + "angle": 0, + "content": "\\[\n\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = \\tilde {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} \\cdot \\sigma \\left(W _ {3} \\left(\\operatorname {R e L U} \\left(W _ {3} \\left(\\mathcal {P} (\\mathcal {G}) _ {s, k} ^ {\\mathrm {f w}}\\right)\\right)\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.367, + 0.906, + 0.442 + ], + "angle": 0, + "content": "where \\( W_{3} \\) denotes a \\( 3 \\times 3 \\) conv. layer, ReLU represents the rectified linear unit activation function, and \\( \\sigma \\) is the sigmoid activation function. Similarly, the feature \\( \\bar{\\mathcal{G}}(\\mathcal{T})_{s,k-1}^{\\mathrm{fw}} \\) for the \\( k-1 \\) timestep is obtained in the same manner. Finally, we fuse the \\( k-1 \\) and \\( k \\) timestep features as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.448, + 0.905, + 0.468 + ], + "angle": 0, + "content": "\\[\n\\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}}\\right) + W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k - 1} ^ {\\mathrm {f w}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.475, + 0.906, + 0.596 + ], + "angle": 0, + "content": "where \\(W_{1}\\) denotes a \\(1\\times 1\\) conv. layer, we pass \\(\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\) to the next IS-TFF encoder block with downsampling. This process is repeated for \\(s = \\{0,1,2\\}\\), producing outputs at three scales. This process is also repeated through a separate backward encoder, resulting in a set of feature maps \\(\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\}\\) from the forward encoder pass and \\(\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{bw}}\\}\\) from the backward encoder pass at timestep \\(k\\). We then fuse the forward and backward features at scale \\(s\\) as follows." + }, + { + "type": "equation", + "bbox": [ + 0.603, + 0.61, + 0.905, + 0.629 + ], + "angle": 0, + "content": "\\[\n\\mathcal {D} (\\mathcal {T}) _ {s, k} = \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} + \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {b w}} \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.632, + 0.906, + 0.677 + ], + "angle": 0, + "content": "With this approach, we generate a multi-scale feature map set \\(\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}\\), where \\(s\\in \\{0,1,2\\}\\), that reinforces rich temporal information within the shutter interval \\(t\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.685, + 0.623, + 0.699 + ], + "angle": 0, + "content": "4.4. Decoders" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.707, + 0.906, + 0.813 + ], + "angle": 0, + "content": "After the IS-TFF based encoding stage, we decode the feature map for each timestep \\( k \\). For this, we apply two separate decoders, one for exposure time(generating \\( E \\) frames) and another for readout-dead time(generating \\( R \\) frames), share the same structure but have different weights. Consequently, we simultaneously estimate \\( N \\) sharp frames \\( \\{S_0, \\ldots, S_{N-1}\\} \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.824, + 0.646, + 0.841 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.783, + 0.866 + ], + "angle": 0, + "content": "5.1. Datasets and Baseline Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "To the best of our knowledge, no prior work has jointly tackled low-light video enhancement, frame interpolation," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8529" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.159 + ], + "angle": 0, + "content": "Table 2. Quantitative evaluation results for the tasks of joint low-light enhancement, deblurring, and frame interpolation on the RELBFI dataset. Bold and underline denote the best and second-best performance, respectively. The symbol \\(\\dagger\\) denotes event-guided methods. The number preceding the parameter corresponds to the deblurring method, while the number following it corresponds to the video frame interpolation methods. The runtime is defined as the time required to generate one or eight \\(1280 \\times 896\\) normal-light sharp images, and was measured using an RTX 3090 GPU." + }, + { + "type": "table", + "bbox": [ + 0.125, + 0.163, + 0.872, + 0.361 + ], + "angle": 0, + "content": "
Method\nDeblur+VFI / JointPSNR↑SSIM↑Params(MB)↓Runtime
1×[s]↓8×[s]↓
LEDNet [65] + TimeLens† [44]24.420.8017.41 + 72.30.3622.894
LEDNet [65] + CBMNet† [44]24.550.8027.41 + 15.40.9017.204
EFNet† [40] + TimeLens† [44]27.040.8438.47 + 72.30.3773.018
EFNet† [40] + CBMNet† [44]27.420.8528.47 + 15.40.9167.328
ELEDNet† [16] + TimeLens† [44]28.650.86312.8 + 72.30.4863.892
ELEDNet† [16] + CBMNet† [44]29.280.87712.8 + 15.41.0258.202
DEMFI [30]24.220.7947.411.2019.606
BiT [63]26.560.83111.30.2341.875
REFID† [41]29.490.87615.90.2071.656
Ours†31.620.9029.40.1481.184
" + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.369, + 0.483, + 0.398 + ], + "angle": 0, + "content": "Table 3. Quantitative evaluation results for the joint low-light enhancement and deblurring tasks on the RELBFI datasets." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.402, + 0.479, + 0.542 + ], + "angle": 0, + "content": "
MethodsPSNRSSIMParams(MB)
LEDNet [65]24.850.8077.41
EFNet [40]27.630.8628.47
DEMFI [30]24.150.8057.41
BiT [63]24.690.81211.3
REFID [41]29.080.87815.9
ELEDNet [16]30.480.89912.8
Ours31.650.9079.4
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.553, + 0.486, + 0.902 + ], + "angle": 0, + "content": "and deblurring. Therefore, we established our own baselines and datasets for comprehensive benchmarking. Since no existing dataset includes high-speed ground-truth frames captured under normal lighting, low-frame-rate low-light blurry video frames, and low-light events, we conducted all experiments using our RELBFI dataset. The RELBFI dataset consists of 67 sequences, with 47 used for training and 20 reserved for testing. We established two types of baselines. The first is a cascaded approach, where deblurring and video frame interpolation methods are applied sequentially. For joint low-light enhancement and deblurring, we utilized LEDNet [65] (Frame-based), ELEDNet [16] (Event-based), and EFNet [40] (Event-based). These methods were trained for the same number of epochs as our network. For video frame interpolation (VFI), we used widely adopted event-based methods, Timelens [44] and CBMNet [13]. As these methods typically take longer to converge compared to our deblurring network, we trained them for additional epochs to achieve sufficient convergence. Finally, we evaluated state-of-the-art joint deblurring and frame interpolation methods, including DEMFI [30] (Frame-based), BiT [63] (Frame-based), and REFID [41] (Event-based). These methods were also" + }, + { + "type": "table_caption", + "bbox": [ + 0.565, + 0.37, + 0.855, + 0.384 + ], + "angle": 0, + "content": "Table 4. Ablation study on the RELBFI datasets." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.388, + 0.897, + 0.476 + ], + "angle": 0, + "content": "
MethodsMM-ISTFFIS-TFF(fw)IS-TFF(bw)PSNRs
Ver.127.07
Ver.2✓30.81
Ver.3✓✓31.35
Ver.4✓✓✓31.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.484, + 0.907, + 0.526 + ], + "angle": 0, + "content": "Table 5. Performance effectiveness analysis of the MM-ISTFF module. The (-) symbol indicates that training was unstable and did not converge." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.53, + 0.899, + 0.566 + ], + "angle": 0, + "content": "
Methodsw/o Eq.(1)w/o Eq.(2)w/o BiOurs
PSNRs30.84(-)31.0431.62
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.572, + 0.904, + 0.586 + ], + "angle": 0, + "content": "Table 6. Network component analysis of the dual decoder module." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.59, + 0.897, + 0.625 + ], + "angle": 0, + "content": "
MethodsSingle weight-sharing decoderOurs
PSNRs31.1631.62
" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.643, + 0.874, + 0.658 + ], + "angle": 0, + "content": "trained for the same number of epochs as our network." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.668, + 0.9, + 0.684 + ], + "angle": 0, + "content": "5.2. Experimental results on the RELBFI datasets" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.907, + 0.735 + ], + "angle": 0, + "content": "We have provided the results of the quantitative evaluation on the RELBFI dataset in Tab.3(Joint LLE and deblurring) and Tab.2(Joint LLE and deblurring and interpolation)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Joint LLE and deblurring and interpolation In Tab.2, we evaluated two groups. Cascaded solutions, where joint deblurring and LLE methods are followed by video frame interpolation (VFI), showed significant performance degradation in frame-based methods like LEDNet [65]. Performance improved with event-based networks such as EFNet [40] and ELEDNet [16], but these cascaded solutions still had a PSNR gap of 2.34 dB to 4.58 dB compared to our network. For joint-based solutions, frame-based methods like DEMFI [30] and BiT [63] showed a PSNR difference of 5.06 dB to 7.40 dB, while compared to REFID [41]," + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8530" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.9, + 0.369 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.37, + 0.908, + 0.4 + ], + "angle": 0, + "content": "Figure 6. Qualitative results on the RELBFI dataset. From left to right, the images show (a) input low-light blurry images with brightness adjustment, (b) low-light events, (c) BiT [63], (d) REFID [41], (e) Ours, and (f) ground-truth normal-light sharp images." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.419, + 0.482, + 0.493 + ], + "angle": 0, + "content": "the state-of-the-art joint event-based method, our network outperformed by 2.13 dB with a large margin. From the perspective of computational cost, our method achieves the best performance with fewer computations and reasonable parameters compared to all state-of-the-art methods." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.494, + 0.483, + 0.645 + ], + "angle": 0, + "content": "Joint LLE and deblurring In Tab.3, we conducted an evaluation of the joint low-light enhancement and deblurring method, excluding interpolation. When compared to ELEDNet, the best state-of-the-art method, it showed a performance difference of 1.17 dB. Comparisons with the other frame-based and event-based methods revealed differences ranging from a maximum of 7.5 dB to a minimum of 2.57 dB. This also evidences the robustness of the proposed method in both deblurring and low-light enhancement (LLE) tasks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.653, + 0.245, + 0.669 + ], + "angle": 0, + "content": "5.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.765 + ], + "angle": 0, + "content": "All ablation experiments were conducted on the RELBFI dataset, with networks trained for 200 epochs. The results are summarized in Tab. 4. For the baseline model, the blur input feature \\(\\mathcal{F}(B)^t\\) at the \\(t\\)-shutter interval was used as input for the IS-TFF module-based networks (stage 2) instead of \\(\\mathcal{F}(\\mathcal{K})_{s=0}^t\\) from the MM-ISTFF module." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "MM-ISTFF module The MM-ISTFF module extracts discriminative spatio-temporal information from multi-modal data. As shown in Tab. 4, Ver. 2 outperforms Ver. 1 by \\(3.74\\mathrm{dB}\\). In Tab. 5, we analyze MM-ISTFF components. Removing the pixel-wise filter from Eq. (1) caused a 0.78 dB drop, while removing regional feature selection from Eq. (2) led to unstable training and non-convergence, underscoring its importance for region-wise selection of degraded multi-modal data. Limiting temporal fusion to current and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.905, + 0.449 + ], + "angle": 0, + "content": "future parts (third row) resulted in a \\(0.50\\mathrm{dB}\\) drop, highlighting the benefits of bidirectional fusion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.451, + 0.906, + 0.543 + ], + "angle": 0, + "content": "IS-TFF module We validated the effectiveness of the IS-TFF module. Comparing Ver. 2 and Ver. 3 in Tab. 4, incorporating IS-TFF in the forward direction improved performance by \\(0.5\\mathrm{dB}\\). Adding the IS-TFF module in the backward direction (Ver. 3 vs. Ver. 4) yielded an additional \\(0.33\\mathrm{dB}\\) gain, resulting in a total improvement of \\(0.83\\mathrm{dB}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.667 + ], + "angle": 0, + "content": "The importance of the separate dual decoders In the decoder following the IS-TFF encoder, separate decoders with different weights are used for frames from the exposure time and the readout deadtime. As shown in Tab. 6, this improves performance \\((+0.46\\mathrm{dB})\\) by applying different decoding methods—blur decomposition or interpolation—tailored to the exposure and readout periods, respectively." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.68, + 0.64, + 0.696 + ], + "angle": 0, + "content": "6. Conclusions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.702, + 0.907, + 0.837 + ], + "angle": 0, + "content": "This paper addresses, for the first time, the novel and previously unexplored problem of unified event-guided joint low-light enhancement, deblurring, and frame interpolation. To solve this task, we propose an integrated framework that jointly handles these sub-tasks with high precision, achieving state-of-the-art performance in both accuracy and computational efficiency. We expect our work to offer initial insights that may support future research in event-guided video restoration." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. NRF-2022R1A2B5B03002636)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "8531" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.243 + ], + "angle": 0, + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.483, + 0.314 + ], + "angle": 0, + "content": "[3] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.316, + 0.484, + 0.385 + ], + "angle": 0, + "content": "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.387, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[5] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12492-12503, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.459, + 0.483, + 0.527 + ], + "angle": 0, + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.53, + 0.483, + 0.598 + ], + "angle": 0, + "content": "[7] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12345-12354, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.602, + 0.483, + 0.671 + ], + "angle": 0, + "content": "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.673, + 0.484, + 0.757 + ], + "angle": 0, + "content": "[9] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.758, + 0.483, + 0.814 + ], + "angle": 0, + "content": "[10] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.816, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[11] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos." + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. arXiv preprint arXiv:2408.14930, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.314 + ], + "angle": 0, + "content": "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.372, + 0.905, + 0.44 + ], + "angle": 0, + "content": "[17] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.442, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[18] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.593 + ], + "angle": 0, + "content": "[19] Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial-temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9822-9832, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.595, + 0.905, + 0.677 + ], + "angle": 0, + "content": "[20] Guoqiang Liang, Kanghao Chen, Hangyu Li, Yunfan Lu, and Lin Wang. Towards robust event-guided low-light image enhancement: A large-scale real-world event-image dataset and novel approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23-33, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[21] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.734, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[22] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Lin Liu, Junfeng An, Jianzhuang Liu, Shanxin Yuan, Xi-angyu Chen, Wengang Zhou, Houqiang Li, Yan Feng Wang," + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8532" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.135 + ], + "angle": 0, + "content": "and Qi Tian. Low-light video enhancement with synthetic event guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1692-1700, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.137, + 0.483, + 0.219 + ], + "angle": 0, + "content": "[25] Yunlong Liu, Tao Huang, Weisheng Dong, Fangfang Wu, Xin Li, and Guangming Shi. Low-light image enhancement with multi-stage residue quantization and brightness-aware attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12140-12149, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.223, + 0.482, + 0.277 + ], + "angle": 0, + "content": "[26] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.281, + 0.482, + 0.336 + ], + "angle": 0, + "content": "[27] Yongrui Ma, Shi Guo, Yutian Chen, Tianfan Xue, and Jinwei Gu. Timelens-xl: Real-time event-based video frame interpolation with large motion. In European Conference on Computer Vision, pages 178-194. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.339, + 0.482, + 0.395 + ], + "angle": 0, + "content": "[28] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1905-1913, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.398, + 0.482, + 0.464 + ], + "angle": 0, + "content": "[29] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[30] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision (ECCV), 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.527, + 0.482, + 0.595 + ], + "angle": 0, + "content": "[31] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.599, + 0.482, + 0.668 + ], + "angle": 0, + "content": "[32] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22191-22200, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.671, + 0.482, + 0.74 + ], + "angle": 0, + "content": "[33] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.743, + 0.482, + 0.798 + ], + "angle": 0, + "content": "[34] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.801, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[35] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[36] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.174 + ], + "angle": 0, + "content": "[37] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.23 + ], + "angle": 0, + "content": "[38] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.233, + 0.905, + 0.301 + ], + "angle": 0, + "content": "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.371 + ], + "angle": 0, + "content": "[40] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.438 + ], + "angle": 0, + "content": "[41] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. arXiv preprint arXiv:2301.05191, 2023. 2, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[42] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.498, + 0.905, + 0.579 + ], + "angle": 0, + "content": "[43] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.663 + ], + "angle": 0, + "content": "[44] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.665, + 0.905, + 0.746 + ], + "angle": 0, + "content": "[45] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-based frame interpolation with nonlinear parametric flow and multi-scale fusion. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[46] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[47] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13128-13137, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "8533" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[48] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[49] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.234, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[50] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[51] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[52] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.482, + 0.489 + ], + "angle": 0, + "content": "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[54] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[55] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[56] Wen Yang, Jinjian Wu, Jupo Ma, Leida Li, and Guangming Shi. Motion deblurring via spatial-temporal collaboration of frames and events. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6531-6539, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[57] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.76, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[58] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.817, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[60] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for im" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "age deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[61] Huicong Zhang, Haozhe Xie, and Hongxun Yao. Blur-aware spatio-temporal sparse transformer for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2673-2681, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.248 + ], + "angle": 0, + "content": "[62] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 191-207. Springer, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[63] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 1, 3, 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[64] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[65] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[66] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 5" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.488 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "8534" + } + ] +] \ No newline at end of file diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_origin.pdf b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..02206148730b8f8f3a9288b42205dd95a6b00bca --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/57e7702d-451a-45d5-82e6-7003cf61c7f2_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e2e68e6c6e1c2a6b80af4b5165efba693d40df91ccea7794ea06dd88864fbcd +size 26322767 diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/full.md b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/full.md new file mode 100644 index 0000000000000000000000000000000000000000..133b8ee449022c484d7ebf4f85946c7f6cf0d6ea --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/full.md @@ -0,0 +1,331 @@ +# Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring + +Taewoo Kim* KAIST + +intelpro@kaist.ac.kr + +Kuk-Jin Yoon KAIST + +kjyoon@kaist.ac.kr + +# Abstract + +In low-light environments, longer exposure times are commonly used to enhance image visibility; however, this inevitably leads to motion blur. Even with a long exposure time, videos captured in low-light environments still suffer from issues such as low visibility, low contrast, and color distortion. Additionally, the long exposure time results in videos with a low frame rate. Therefore, videos captured in low-light exhibit low visibility and motion blur, as well as low frame rates. To overcome these limitations, we propose a novel problem aimed at transforming motion-blurred, low-frame-rate videos with poor visibility in low-light environments into high-frame-rate videos while simultaneously enhancing their visibility. To tackle this challenge, we leverage the unique advantages of event cameras, which capture scene changes asynchronously, providing superior temporal resolution and a wider dynamic range compared to conventional frame-based cameras. These properties make event cameras particularly effective in reducing motion blur, compensating for low frame rates, and enhancing visibility in low-light conditions. To this end, we developed a hybrid camera system that integrates two RGB cameras and an event camera, capturing a dedicated dataset for this task and proposing novel network architectures to effectively address this problem. The project pages are available at https://sites.google.com/view/eledi-iccv. + +# 1. Introduction + +In our daily lives, low-light conditions are common and occur in various lighting scenarios, both during the day and at night. In such environments, a longer exposure time is generally required to capture the limited available light. However, this extended exposure time inevitably leads to motion blur caused by object movement or dynamic camera motion, while also reducing video frame rates. Additionally, + +videos recorded under these conditions often suffer from reduced visibility due to insufficient lighting, along with significant image noise, even with a long exposure time. As a result, videos captured in low-light conditions typically exhibit low frame rates due to the long exposure time, as well as poor visibility and motion blur. + +With advancements in deep learning, numerous efforts have focused on improving low-light conditions [3, 53, 54], aiming to suppress noise and enhance visibility in such environments. Furthermore, research has been conducted on removing motion blur artifacts through motion deblurring studies [6, 30, 31], as well as on Video Frame Interpolation (VFI) [13, 36, 38, 44], which converts low-frame-rate videos into high-frame-rate ones. While each of these areas has been explored independently, few studies address the combined challenges of multiple degradations in low-light conditions. Among existing works, relatively active research has focused on the problem of joint deblurring and interpolation [30, 63], which aims to restore high-frame-rate, sharp frames from motion-blurred ones. Additionally, some works [16, 65] have investigated integrating motion deblurring and low-light enhancement. However, no existing approach comprehensively tackles all three degradations simultaneously. + +To overcome these limitations, we propose a research problem that simultaneously addresses the challenges of low frame rate, motion blur, and low visibility in low-light conditions. Specifically, we aim to simultaneously enhance visibility, reduce motion blur, and increase the frame rate to improve overall video quality in low-light environments. A simple approach would involve cascading separate models for each task. However, such a cascaded manner suffers from error accumulation and inefficient joint optimization, often leading to suboptimal results. Thus, a more effective solution is needed to jointly optimize these interrelated tasks in an end-to-end manner. Moreover, solving these problems using only frame-based information is extremely challenging, particularly in cases of severe degradation, making it an ill-posed problem. + +To tackle this issue, we propose leveraging the ad + +![](images/e089bde8fcf3380ea2e68be893f3a71e322a7ab12e70cbdeef422c3611020cc5.jpg) +Low-light blurred image + +![](images/3295b74b9e3312293e6128ae5c0ba86d4c0d69ed462de1213c093ab8f7cdf20e.jpg) +t=5/8 +Figure 1. Qualitative result of our method on the RELBFI datasets. Our task aims to restore sharp frames captured in a normal light environment within the shutter interval using low-light blurred images and the events. + +![](images/3e6e47a1b59447b699604f4027a8a9377753aecd3de09424a1c0c03322ed6bad.jpg) +t=6/8 + +![](images/5e8226a961a1f7a383a5fdf30df640b81c51223ec04eea5110da2adeaaf579ec.jpg) +t=7/8 + +![](images/cd4136efb7b560ad34626348d90fcc243a287509fd6a507b046fdc9fd5ec8473.jpg) +t=8/8 + +vantages of event cameras, which, inspired by neuromorphic vision systems, detect brightness changes with microsecond-level temporal resolution [8]. Additionally, their superior dynamic range, compared to conventional cameras, allows for high-speed motion capture even in low-light environments. As a result, event cameras are widely used in video enhancement tasks, including motion deblurring [40, 41], frame interpolation [13, 44], low-light enhancement [16, 20, 23], and joint deblurring and interpolation [41], effectively addressing issues like motion blur, low frame rates, low visibility. + +Therefore, we propose a novel task that leverages event cameras to address these multiple degradations in low-light conditions. However, since this novel task has never been addressed in previous works, it is necessary to build a new dataset. To overcome these challenges, we build a sophisticated hybrid camera system to capture a new RELBFI(Real-world Event-guided Low-light enhancement and Blur Frame Interpolation) dataset, consisting of real-world low-light blur videos with a low frame rate, low-light event data, and high-frame-rate normal light sharp videos (ground truth). We believe that these resources will be a valuable asset to the research community. + +Beyond proposing this dataset, we also introduce a novel unified end-to-end framework to tackle this challenging task. To achieve this, we design two key modules that play a crucial role in effectively addressing the problem. The first module focuses on fusing multi-temporal information from events and frames under low-light conditions. Given that video frames in such environments often suffer from significant noise and color distortion, this module aims to efficiently integrate multi-temporal and multi-modal information sources. For the second module, we introduce a mechanism to reinforce temporal dependencies among generated video frames during the generation of multiple frames within the shutter period. Since this process is computationally intensive, we designed a lightweight yet effective module to strengthen bidirectional dependencies between + +frames, thereby enhancing overall efficiency. + +To summarize, the contributions of this paper are as follows: (I) We propose a new research problem that addresses multiple degradations, including noise, motion blur, low visibility, and low frame rate, which commonly occur in low-light conditions, while also converting videos captured at low frame rates into high frame rates. (II) We present a novel unified framework that effectively addresses this problem, demonstrating at least a 2 dB improvement in PSNR compared to state-of-the-art methods. (III) We introduce the first RELBFI dataset, built using a hybrid camera system. + +# 2. Related Works + +# 2.1. Motion Deblurring. + +Since the advent of deep learning, significant progress has been made in motion deblurring, further accelerated by specialized datasets such as GoPro [29], DVD [39], RealBlur [37], and BSD [62]. In frame-based deblurring, various architectures have been proposed, including multiscale [6], multi-stage [58, 60], transformer-based [42, 49, 59], and frequency-modulated [7, 17, 28] architectures. Video deblurring methods [19, 21, 22, 32, 61, 62] have also emerged. However, deblurring solely from frames remains challenging due to motion information loss, prompting event camera-based research. Early works explored recurrent architectures [10], followed by studies on cross-modal attention-based feature fusion [40], unknown exposure time [12], non-coaxial deblurring [5], frequency-based video deblurring [15], and attention-based cross-modal alignment [14] and spatio-temporal collaboration [56]. + +# 2.2. Video Frame Interpolation. + +Research on frame-based video interpolation [1, 2, 18, 26, 35, 36, 38] has made significant progress, but it remains challenging due to the absence of explicit motion information between frames. In contrast, event cameras, with their high temporal resolution, have gained popularity in + +![](images/62ac4327a9f769fc244d73516257e7bf7660aa3209d3b16900a4cdb575a06bba.jpg) +(a) + +![](images/62d3ea117d00ef5c5004792ba1d1afbd969da9840ffddb0369a649454a121e2f.jpg) +(b) +Figure 2. In the RELBFI dataset description, (a) shows a diagram of the hybrid camera system used to capture the dataset. (b) represents the temporal irradiance relationship for the high-fps RGB camera (used for recording sharp images) and the camera used for capturing blur. (c) presents sample images from the RELBFI dataset: from top to bottom, a low-light blur image, a brightness-adjusted version of the low-light blur image, and a ground truth image captured in normal light conditions. + +![](images/a9666428c0e8758c25f76efb13d49ed6540daa96ece2cd6de886fc03467e9f3c.jpg) +(c) + +Table 1. Comparison of HighREV and RELBFI Datasets. In the table, NL" refers to images taken in normal-light environments, and "LL" refers to images taken in low-light environments. "LFR" refers to low frame rate, and "HFR" refers to high frame rate. + +
HighREV [41]RELBFI
LFR RGB InputNL blurLL blur
Event inputNL eventLL event
HFR RGB GTNL sharpNL sharp
Blur typeSyntheticReal
Event typeRealReal
#N of sharp images2858932629
#N of sequences3067
+ +video interpolation tasks. Early work like TimeLens [43] introduced event-based interpolation using a cascaded U-Net framework, while TimeLens++ [45] enhanced this by estimating non-linear motion with a motion spline estimator. Later studies, including A $^2$ OF [51] and the unsupervised TimeReplayer [9], further advanced the field. Recently, Kim et al. [13] introduced an event-image fusion VFI framework for inter-frame optical flow estimation, while the follow-up study, TimeLens-XL [27], presented a VFI framework with iterative flow updates. + +# 2.3. Low-light Enhancement. + +Many frame-based low-light enhancement (LLE) studies focus on improving visibility and reducing noise using image information. These efforts have led to various approaches, including SNR-based transformers [53], Retinex-based methods [3, 52], gamma correction [47], implicit neural representations [55], diffusion models [48, 57], and other techniques [25, 54, 65]. Leveraging its high dynamic range, + +the event camera excels in capturing motion under low light, making it suitable for LLE. Liu et al. [24] used synthetic events for LLE due to dataset limitations, while Liang et al. [23] modeled multimodal coherence. Recently, Liang et al. [20] proposed an SNR-based framework and dataset for event-guided LLE, and Kim et al. [16] introduced a joint framework for event-guided LLE and deblurring. + +# 2.4. Joint deblurring and frame interpolation. + +The joint task of deblurring and interpolation aims to simultaneously remove blur and generate high-frame-rate videos within a single framework. Jin et al. [11] first addressed this without event cameras using a deblur-then-interpolate approach. Recently, DeMFI [30] combined flow-guided attention and recursive boosting, while BiT [63] introduced a Swin Transformer-based solution. Event camera-based research includes EDI [33, 34], which used a physical model, REFID [41] leveraging a recurrent architecture, and Weng et al. [50], tackling blur frame interpolation under unknown exposure times. Unlike these works, we leverage the robust properties of event cameras in low-light conditions to jointly perform low-light enhancement, deblurring, and frame interpolation. + +# 3. RELBFI Datasets + +# 3.1. Comparisons with Other Datasets. + +Since the advent of deep learning, various event- and frame-based video enhancement datasets [13, 16, 20, 63] have emerged. Some studies [41, 63] focus on motion deblurring and frame interpolation under normal lighting. + +Sun et al. [41] introduced a dataset for joint deblurring and interpolation using an event camera, where motion blur + +![](images/4ae18f5f76e9fad75caf4e21d552afd2bf9a69ab8343a58e00334cd64b02f237.jpg) +Figure 3. Overall framework of the proposed method. "TFG" in the blue box denotes the Template Feature Generator block. + +was synthetically generated by merging adjacent frames. However, such synthetic datasets fail to capture real-world motion blur, as they overlook physical factors like noise variations from different exposure times. + +Recently, Kim et al. [16] proposed the RELED dataset for simultaneous low-light enhancement and deblurring using an event camera. However, its low frame rate limits its effectiveness in converting low-frame-rate, low-light videos into high-frame-rate, normal-light videos. + +To date, no existing research or dataset has directly tackled the transformation of low-frame-rate, low-light, noisy, and motion-blurred inputs into high-frame-rate, noise-free videos under normal lighting by leveraging event data. + +# 3.2. RELBFI dataset acquisitions. + +Synthetic-based datasets struggle to capture the complexity of real-world degradations. Therefore, building a real-world dataset is both practical and challenging, especially when simultaneously capturing high-frame-rate sharp video, low-light blurred video, and low-light events. + +To address this, we build the RELBFI (Real-world Event-guided Low-light Video Enhancement and Blur Frame Interpolation) dataset using a hybrid camera system with beam splitters. As shown in Fig. 2(a), the system includes two RGB cameras (Blackfly BFS-U3-16S2C-CS) and an event camera (Propheesee Gen4). One RGB camera captures sharp images at 166fps with a 3ms exposure and a 3ms readout dead time, receiving $25\%$ of the total light as it passes through the beam splitter twice without an ND filter. The other RGB camera records blurred images in low-light conditions with a 24ms exposure, eight times longer, and uses a 1/64 ND filter, receiving only $0.39\%$ of the original light. As a result, each blurred image has eight times less + +light intensity than a sharp image, resulting in low visibility. + +Furthermore, to illustrate real-world frame acquisition [12, 41, 63], capturing blurred images involves exposure time and readout dead time, collectively termed the shutter interval [12]. To simulate readout dead time, we used a $24\mathrm{ms}$ exposure followed by a $24\mathrm{ms}$ dead time, yielding one low-light blurred image per $48\mathrm{ms}$ (21fps) and eight sharp images under normal lighting (166fps). The event camera, like the blur-recording RGB camera, employs a 1/64 ND filter for low-light capture. + +Regarding calibration, all cameras in the hybrid system are mechanically fixed, but slight adjustments, such as rotation and baseline alignment, are required. To this end, we estimated the homography matrix for each camera and performed additional calibration. As shown in Fig. 2(a), the devices were temporally synchronized via hardware using an external trigger. Using this setup, we captured 67 urban scenes featuring dynamic elements like vehicles, moving objects, and flickering streetlights, along with rich textures. As shown in Tab. 1, RELBFI provides diverse sequences tailored for developing and evaluating methods in this challenging task. It serves as a practical resource to support joint low-light enhancement, deblurring, and interpolation in real-world scenarios. + +# 4. Proposed Methods + +# 4.1. Overview + +Preliminary As explained in Sec. 3, the shutter interval consists of exposure time and readout deadtime. During the readout deadtime, events are recorded, but no blurred frames are captured. We define $N$ as the total number of sharp frames captured during a shutter interval $T$ (ms), + +![](images/fa6c05dcf02dcfc6de3426008d6787ef8549aa1ba49258694ca98e334a571e74.jpg) +Figure 4. The MM-ISTFF module in the scale factor $s$ . In the figure, $\mathcal{D}_s^{\mathrm{fw}}$ and $\mathcal{D}_s^{\mathrm{bw}}$ represent the forward and backward dynamic filters, respectively, as defined in Eq.(1). + +where $N = E + R$ , with $E$ and $R$ representing the number of the frames from exposure and readout deadtime, respectively. During a single shutter interval $t$ , one blurred frame $B^t$ is captured, and events are generated corresponding to both the exposure time and readout deadtime. The events $\{e_i\}_t$ during shutter interval $t$ are transformed into event voxels [66] with 32 bins. + +Overall framework The proposed framework is illustrated in Fig. 3. We utilize 3 blurred images and 3 corresponding event voxel grids from shutter intervals $t - 1$ , $t$ , and $t + 1$ . Given these inputs $\{B^{t-1}, B^t, B^{t+1}\}$ and $\{E^{t-1}, E^t, E^{t+1}\}$ , the network estimates $N$ normal light sharp frames at the shutter interval $t$ . + +The process consists of two main steps. First, the MM-ISTFF (Multi-Modal Inter-frame Spatio-Temporal Feature Fusion) module combines spatio-temporal event data with blurred video frames to improve video quality and reduce low-light noise, producing fused features $\mathcal{F}(\mathcal{K})_{s=0}^{t}$ . Next, the event voxel grid for $t$ is divided into $N$ temporal units $\{E_0^t,\ldots,E_{N-1}^t\}$ , which are combined with $\mathcal{F}(\mathcal{K})_{s=0}^{t}$ to generate $N$ temporal template features $\{\mathcal{F}(\mathcal{T})_0,\ldots,\mathcal{F}(\mathcal{T})_{N-1}\}$ . Next, the IS-TFF (Inter-Shutter Interval Temporal Feature Fusion) based-encoder module leverages temporal information within the shutter interval then uses these features to produce an enhanced feature map $\{\mathcal{D}(\mathcal{T})_{s,k}\}$ . Finally, the decoder generates $N$ normal light sharp frames $\{S_0,\ldots,S_{N-1}\}$ for the $t$ shutter interval. + +# 4.2. Multi-Modal Inter-frame Spatio-Temporal Feature Fusion(MM-ISTFF) Module. + +In low-light conditions, frame-based information is severely affected by noise and low visibility, while motion blur further exacerbates the degradation. In such scenarios, events also suffer from significant degradation. Therefore, it is crucial to fuse features by complementing the spatio-temporal visual information from images with the rich temporal details from events, effectively utilizing them + +![](images/0af3b3d77c22ae4440a0b502ecba8985cebc7facb9b922c2a4be71ede8a6d973.jpg) +Figure 5. Illustration of the IS-TFF module. The $s$ after the forward and backward units indicates the scale factor. + +for video enhancement. To address this, we propose the MM-ISTFF module, designed to mine discriminative temporal information between frames and events across different shutter intervals. For this, we extract multi-scale multi-modal features $\{\mathcal{F}(B)_s^{t - 1},\mathcal{F}(B)_s^t,\mathcal{F}(B)_s^{t + 1}\}$ and $\{\mathcal{F}(E)_s^{t - 1},\mathcal{F}(E)_s^t,\mathcal{F}(E)_s^{t + 1}\}$ from the low-light blurred and event voxel inputs. As shown in Fig.3, we aim to leverage these multi-modal inputs to extract discriminative spatio-temporal information across multiple visual scales. To achieve this, we use past and future information from the current shutter interval $t$ . For simplicity, only the process for past information is described, as the same applies to future information. As shown in Fig.4, to integrate spatiotemporal information from past multi-modal data, we first apply a conv. layer with LeakyReLU to the concatenated features $[\mathcal{F}(B)_s^{t - 1},\mathcal{F}(B)_s^t,\mathcal{F}(E)_s^{t - 1},\mathcal{F}(E)_s^t]$ , generating the feature $\mathbf{V}_s^{\mathrm{fw}}$ . To extract discriminative spatio-temporal information, pixel-wise spatial dynamic filters [64] are applied to obtain content-adaptive features. Here, to estimate a dynamic filter for each pixel, we use $f_{s}^{fw} = f_{\mathrm{filter}}^{\mathrm{fw}}(\mathbf{V}_{s}^{\mathrm{fw}})$ to generate dynamic filters, which are then applied to the original feature $\mathbf{V}_s^{\mathrm{fw}}$ , where $f_{\mathrm{filter}}$ We apply dynamic filter as follows: + +$$ +\hat {\mathbf {V}} _ {s} ^ {\mathrm {f w}} = f _ {s} ^ {\mathrm {f w}} \otimes \mathbf {V} _ {s} ^ {\mathrm {f w}} \tag {1} +$$ + +where $f_{s}^{\mathrm{fw}}$ represents the estimated filter, and $\otimes$ denotes the dynamic filter application. This process extracts content-adaptive information, but in low-light conditions, selective regional features are needed due to noise and degradation. Thus, we apply region-wise attention for feature filtering. + +$$ +\tilde {\mathbf {V}} _ {s} ^ {f w} = \hat {\mathbf {V}} _ {s} ^ {\mathrm {f w}} \odot \sigma (\mathrm {L R e L U} ((W _ {3} (\hat {\mathbf {V}} _ {s} ^ {\mathrm {f w}})))) \tag {2} +$$ + +where $\sigma$ denotes the sigmoid operation, and $W_{3}$ and LReLU represent the $3\times 3$ conv. layer and the LeakyReLU operation, respectively. Similarly, future multi-modal informa + +tion is utilized to obtain $\tilde{\mathbf{V}}_s^{\mathrm{bw}}$ . These two features $(\tilde{\mathbf{V}}_s^{\mathrm{fw}}$ and $\tilde{\mathbf{V}}_s^{\mathrm{bw}})$ are combined via channel-wise concatenation followed by a ResBlock to produce the final feature $\mathcal{F}(\mathcal{K})_s^t$ . This process is repeated for $s = \{0,1,2\}$ , with $\mathcal{F}(\mathcal{K})_{s=0}^{t}$ serving as input to the IS-TFF module. + +# 4.3. Inter-Shutter Interval Temporal Feature Fusion(IS-TFF) Module. + +Template Feature Generation After the MM-ISTFF module, we obtain a spatio-temporally fused feature $\mathcal{F}(\mathcal{K})_{s=0}^{t}$ centered at shutter interval $t$ . To estimate $N$ sharp images within the shutter interval, template features capturing the characteristics of each timestep must be generated for all $N$ timesteps. To address this, we first divide the event voxel grid into $N$ temporal units, each with an equal time interval $\Delta t_{unit} = T / N$ . This gives us temporally divided unit event voxels $E_k^t$ , which are then encoded to generate temporal event features. Therefore, we obtain the temporally divided unit event voxel $\{E_k^t\}$ with temporal index $k \in \{0, \dots, N-1\}$ . Subsequently, each event voxel unit is encoded to generate temporal event features $\{\mathcal{F}(E_k^t) \in \mathbb{R}^{C \times H \times W}\}_k$ . This process is repeated from $k = 0$ to $N-1$ . Next, we define a zero tensor $Z_k \in \mathbb{R}^{N \times C \times H \times W}$ and populate it with temporal unit event features for each timestep: + +$$ +Z _ {k} [ k,:,:,: ] = \mathcal {F} \left(E _ {k} ^ {t}\right), \quad \forall k \in \{0, 1, \dots , N - 1 \}. +$$ + +This tensor $Z_{k}$ contains unit event voxel features $\mathcal{F}(E_k^t)$ filled only at the $k$ -th position along the temporal axis, with the rest initialized to zero. As a result, it encapsulates information about which timestep is being interpolated and the motion information required for the interpolation. Afterward, for $k = \{0,\dots ,N - 1\}$ , we concatenate the $Z_{k}$ tensor with $\mathcal{F}(\mathcal{K})_{s = 0}^{t}$ , pass them through a Residual blocks, and generate the template features $\{\mathcal{F}(\mathcal{T})_k\}$ for $k$ timesteps. + +IS-TFF based Bidirectional Encoder Module. Through the template feature generator, we obtain the feature $\mathcal{F}(\mathcal{T})_k$ for each timestep $k$ , capturing the essential information to interpolate the normal light sharp frame $S_{k}$ . By using these features, in this module, we aim to enhance the temporal dependency for timestep $k$ within the shutter interval $t$ using multi-scale temporal information. + +While existing methods [4, 46] enhance temporal dependency within frames, they are computationally intensive, especially when applied to multiple timesteps at various scales. To address this, we propose a lightweight yet effective temporal feature fusion method within the shutter interval, which enables information sharing across the temporal dimension. As shown in Fig. 3, the IS-TFF-based encoder modules operate bidirectionally at each visual scale $(s = 0,1,2)$ . For clarity, we describe the process at scale $s$ in the forward direction; the same applies to the backward direction and other scales. + +As shown in Fig. 5, we first extract the features through $\mathcal{F}(\mathcal{T})_{k,s}^{fw}$ processed at the previous scale, passing them through a conv. layer to transform them into $\mathcal{G}(\mathcal{T})_{s,k}^{\mathrm{fw}}$ . For the fusion of timesteps $k$ and $k - 1$ in forward direction, the fusion of $\mathcal{G}(\mathcal{T})_{s,k}^{\mathrm{fw}}$ and $\mathcal{G}(\mathcal{T})_{s,k - 1}^{\mathrm{fw}}$ begins by applying a conv. layer with LeakyReLU to their channel-wise concatenation, producing the feature $\tilde{\mathcal{G}} (\mathcal{T})_{s,k}$ . Then, this feature is split along the channel dimension into two parts, $\tilde{\mathcal{G}} (\mathcal{T})_{s,k}$ and $\tilde{\mathcal{G}} (\mathcal{T})_{s,k - 1}$ . After mixing the adjacent features through concatenation and the first conv. layer, we analyze their temporally significant parts through the channel-wise dimension. For this, we first apply AAP (adaptive average pooling) to compress the spatial dimensions, as follows: $\mathcal{P}(\mathcal{G})_{s,k}^{\mathrm{fw}} = \mathrm{AAP}(\tilde{\mathcal{G}} (\mathcal{T})_{s,k}^{\mathrm{fw}})$ . Next, we apply attention [58] to the feature at each timestep $k$ and $k - 1$ as follows: + +$$ +\bar {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {f w}} = \tilde {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {f w}} \cdot \sigma \left(W _ {3} \left(\operatorname {R e L U} \left(W _ {3} \left(\mathcal {P} (\mathcal {G}) _ {s, k} ^ {\mathrm {f w}}\right)\right)\right)\right) \tag {3} +$$ + +where $W_{3}$ denotes a $3 \times 3$ conv. layer, ReLU represents the rectified linear unit activation function, and $\sigma$ is the sigmoid activation function. Similarly, the feature $\bar{\mathcal{G}}(\mathcal{T})_{s,k-1}^{\mathrm{fw}}$ for the $k-1$ timestep is obtained in the same manner. Finally, we fuse the $k-1$ and $k$ timestep features as follows: + +$$ +\hat {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {f w}} = W _ {1} \left(\bar {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {f w}}\right) + W _ {1} \left(\bar {\mathcal {G}} (\mathcal {T}) _ {s, k - 1} ^ {\mathrm {f w}}\right) \tag {4} +$$ + +where $W_{1}$ denotes a $1\times 1$ conv. layer, we pass $\hat{\mathcal{G}} (\mathcal{T})_{s,k}^{\mathrm{fw}}$ to the next IS-TFF encoder block with downsampling. This process is repeated for $s = \{0,1,2\}$ , producing outputs at three scales. This process is also repeated through a separate backward encoder, resulting in a set of feature maps $\{\hat{\mathcal{G}} (\mathcal{T})_{s,k}^{\mathrm{fw}}\}$ from the forward encoder pass and $\{\hat{\mathcal{G}} (\mathcal{T})_{s,k}^{\mathrm{bw}}\}$ from the backward encoder pass at timestep $k$ . We then fuse the forward and backward features at scale $s$ as follows. + +$$ +\mathcal {D} (\mathcal {T}) _ {s, k} = \hat {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {f w}} + \hat {\mathcal {G}} (\mathcal {T}) _ {s, k} ^ {\mathrm {b w}} \tag {5} +$$ + +With this approach, we generate a multi-scale feature map set $\{\mathcal{D}(\mathcal{T})_{s,k}\}$ , where $s\in \{0,1,2\}$ , that reinforces rich temporal information within the shutter interval $t$ . + +# 4.4. Decoders + +After the IS-TFF based encoding stage, we decode the feature map for each timestep $k$ . For this, we apply two separate decoders, one for exposure time(generating $E$ frames) and another for readout-dead time(generating $R$ frames), share the same structure but have different weights. Consequently, we simultaneously estimate $N$ sharp frames $\{S_0, \ldots, S_{N-1}\}$ . + +# 5. Experiments + +# 5.1. Datasets and Baseline Settings + +To the best of our knowledge, no prior work has jointly tackled low-light video enhancement, frame interpolation, + +Table 2. Quantitative evaluation results for the tasks of joint low-light enhancement, deblurring, and frame interpolation on the RELBFI dataset. Bold and underline denote the best and second-best performance, respectively. The symbol $\dagger$ denotes event-guided methods. The number preceding the parameter corresponds to the deblurring method, while the number following it corresponds to the video frame interpolation methods. The runtime is defined as the time required to generate one or eight $1280 \times 896$ normal-light sharp images, and was measured using an RTX 3090 GPU. + +
Method +Deblur+VFI / JointPSNR↑SSIM↑Params(MB)↓Runtime
1×[s]↓8×[s]↓
LEDNet [65] + TimeLens† [44]24.420.8017.41 + 72.30.3622.894
LEDNet [65] + CBMNet† [44]24.550.8027.41 + 15.40.9017.204
EFNet† [40] + TimeLens† [44]27.040.8438.47 + 72.30.3773.018
EFNet† [40] + CBMNet† [44]27.420.8528.47 + 15.40.9167.328
ELEDNet† [16] + TimeLens† [44]28.650.86312.8 + 72.30.4863.892
ELEDNet† [16] + CBMNet† [44]29.280.87712.8 + 15.41.0258.202
DEMFI [30]24.220.7947.411.2019.606
BiT [63]26.560.83111.30.2341.875
REFID† [41]29.490.87615.90.2071.656
Ours†31.620.9029.40.1481.184
+ +Table 3. Quantitative evaluation results for the joint low-light enhancement and deblurring tasks on the RELBFI datasets. + +
MethodsPSNRSSIMParams(MB)
LEDNet [65]24.850.8077.41
EFNet [40]27.630.8628.47
DEMFI [30]24.150.8057.41
BiT [63]24.690.81211.3
REFID [41]29.080.87815.9
ELEDNet [16]30.480.89912.8
Ours31.650.9079.4
+ +and deblurring. Therefore, we established our own baselines and datasets for comprehensive benchmarking. Since no existing dataset includes high-speed ground-truth frames captured under normal lighting, low-frame-rate low-light blurry video frames, and low-light events, we conducted all experiments using our RELBFI dataset. The RELBFI dataset consists of 67 sequences, with 47 used for training and 20 reserved for testing. We established two types of baselines. The first is a cascaded approach, where deblurring and video frame interpolation methods are applied sequentially. For joint low-light enhancement and deblurring, we utilized LEDNet [65] (Frame-based), ELEDNet [16] (Event-based), and EFNet [40] (Event-based). These methods were trained for the same number of epochs as our network. For video frame interpolation (VFI), we used widely adopted event-based methods, Timelens [44] and CBMNet [13]. As these methods typically take longer to converge compared to our deblurring network, we trained them for additional epochs to achieve sufficient convergence. Finally, we evaluated state-of-the-art joint deblurring and frame interpolation methods, including DEMFI [30] (Frame-based), BiT [63] (Frame-based), and REFID [41] (Event-based). These methods were also + +Table 4. Ablation study on the RELBFI datasets. + +
MethodsMM-ISTFFIS-TFF(fw)IS-TFF(bw)PSNRs
Ver.127.07
Ver.2✓30.81
Ver.3✓✓31.35
Ver.4✓✓✓31.62
+ +Table 5. Performance effectiveness analysis of the MM-ISTFF module. The (-) symbol indicates that training was unstable and did not converge. + +
Methodsw/o Eq.(1)w/o Eq.(2)w/o BiOurs
PSNRs30.84(-)31.0431.62
+ +Table 6. Network component analysis of the dual decoder module. + +
MethodsSingle weight-sharing decoderOurs
PSNRs31.1631.62
+ +trained for the same number of epochs as our network. + +# 5.2. Experimental results on the RELBFI datasets + +We have provided the results of the quantitative evaluation on the RELBFI dataset in Tab.3(Joint LLE and deblurring) and Tab.2(Joint LLE and deblurring and interpolation). + +Joint LLE and deblurring and interpolation In Tab.2, we evaluated two groups. Cascaded solutions, where joint deblurring and LLE methods are followed by video frame interpolation (VFI), showed significant performance degradation in frame-based methods like LEDNet [65]. Performance improved with event-based networks such as EFNet [40] and ELEDNet [16], but these cascaded solutions still had a PSNR gap of 2.34 dB to 4.58 dB compared to our network. For joint-based solutions, frame-based methods like DEMFI [30] and BiT [63] showed a PSNR difference of 5.06 dB to 7.40 dB, while compared to REFID [41], + +![](images/ae680d372260006a55bc1ec40bdcb59114c0e7772538d45f39440404bc3ffebe.jpg) +Figure 6. Qualitative results on the RELBFI dataset. From left to right, the images show (a) input low-light blurry images with brightness adjustment, (b) low-light events, (c) BiT [63], (d) REFID [41], (e) Ours, and (f) ground-truth normal-light sharp images. + +the state-of-the-art joint event-based method, our network outperformed by 2.13 dB with a large margin. From the perspective of computational cost, our method achieves the best performance with fewer computations and reasonable parameters compared to all state-of-the-art methods. + +Joint LLE and deblurring In Tab.3, we conducted an evaluation of the joint low-light enhancement and deblurring method, excluding interpolation. When compared to ELEDNet, the best state-of-the-art method, it showed a performance difference of 1.17 dB. Comparisons with the other frame-based and event-based methods revealed differences ranging from a maximum of 7.5 dB to a minimum of 2.57 dB. This also evidences the robustness of the proposed method in both deblurring and low-light enhancement (LLE) tasks. + +# 5.3. Ablation Study + +All ablation experiments were conducted on the RELBFI dataset, with networks trained for 200 epochs. The results are summarized in Tab. 4. For the baseline model, the blur input feature $\mathcal{F}(B)^t$ at the $t$ -shutter interval was used as input for the IS-TFF module-based networks (stage 2) instead of $\mathcal{F}(\mathcal{K})_{s=0}^t$ from the MM-ISTFF module. + +MM-ISTFF module The MM-ISTFF module extracts discriminative spatio-temporal information from multi-modal data. As shown in Tab. 4, Ver. 2 outperforms Ver. 1 by $3.74\mathrm{dB}$ . In Tab. 5, we analyze MM-ISTFF components. Removing the pixel-wise filter from Eq. (1) caused a 0.78 dB drop, while removing regional feature selection from Eq. (2) led to unstable training and non-convergence, underscoring its importance for region-wise selection of degraded multi-modal data. Limiting temporal fusion to current and + +future parts (third row) resulted in a $0.50\mathrm{dB}$ drop, highlighting the benefits of bidirectional fusion. + +IS-TFF module We validated the effectiveness of the IS-TFF module. Comparing Ver. 2 and Ver. 3 in Tab. 4, incorporating IS-TFF in the forward direction improved performance by $0.5\mathrm{dB}$ . Adding the IS-TFF module in the backward direction (Ver. 3 vs. Ver. 4) yielded an additional $0.33\mathrm{dB}$ gain, resulting in a total improvement of $0.83\mathrm{dB}$ . + +The importance of the separate dual decoders In the decoder following the IS-TFF encoder, separate decoders with different weights are used for frames from the exposure time and the readout deadtime. As shown in Tab. 6, this improves performance $(+0.46\mathrm{dB})$ by applying different decoding methods—blur decomposition or interpolation—tailored to the exposure and readout periods, respectively. + +# 6. Conclusions + +This paper addresses, for the first time, the novel and previously unexplored problem of unified event-guided joint low-light enhancement, deblurring, and frame interpolation. To solve this task, we propose an integrated framework that jointly handles these sub-tasks with high precision, achieving state-of-the-art performance in both accuracy and computational efficiency. We expect our work to offer initial insights that may support future research in event-guided video restoration. + +Acknowledgement. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. NRF-2022R1A2B5B03002636). + +# References + +[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2 +[2] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 2 +[3] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 1, 3 +[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 6 +[5] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12492-12503, 2023. 2 +[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2 +[7] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12345-12354, 2023. 2 +[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2 +[9] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 3 +[10] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2 +[11] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019. 3 +[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos. + +In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 4 +[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 3, 7 +[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. arXiv preprint arXiv:2408.14930, 2024. 2 +[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2 +[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 1, 2, 3, 4, 7 +[17] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2 +[18] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2 +[19] Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial-temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9822-9832, 2023. 2 +[20] Guoqiang Liang, Kanghao Chen, Hangyu Li, Yunfan Lu, and Lin Wang. Towards robust event-guided low-light image enhancement: A large-scale real-world event-image dataset and novel approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23-33, 2024. 2, 3 +[21] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2 +[22] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2 +[23] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2, 3 +[24] Lin Liu, Junfeng An, Jianzhuang Liu, Shanxin Yuan, Xi-angyu Chen, Wengang Zhou, Houqiang Li, Yan Feng Wang, + +and Qi Tian. Low-light video enhancement with synthetic event guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1692-1700, 2023. 3 +[25] Yunlong Liu, Tao Huang, Weisheng Dong, Fangfang Wu, Xin Li, and Guangming Shi. Low-light image enhancement with multi-stage residue quantization and brightness-aware attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12140-12149, 2023. 3 +[26] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[27] Yongrui Ma, Shi Guo, Yutian Chen, Tianfan Xue, and Jinwei Gu. Timelens-xl: Real-time event-based video frame interpolation with large motion. In European Conference on Computer Vision, pages 178-194. Springer, 2025. 3 +[28] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1905-1913, 2023. 2 +[29] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 2 +[30] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision (ECCV), 2022. 1, 3, 7 +[31] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1 +[32] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22191-22200, 2023. 2 +[33] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3 +[34] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3 +[35] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 2 +[36] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2 + +[37] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 2 +[38] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2 +[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 2 +[40] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2, 7 +[41] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. arXiv preprint arXiv:2301.05191, 2023. 2, 3, 4, 7, 8 +[42] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 2 +[43] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 3 +[44] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 7 +[45] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-based frame interpolation with nonlinear parametric flow and multi-scale fusion. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3 +[46] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 6 +[47] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13128-13137, 2023. 3 + +[48] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 3 +[49] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 2 +[50] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 3 +[51] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3 +[52] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2022. 3 +[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 1, 3 +[54] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 1, 3 +[55] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 3 +[56] Wen Yang, Jinjian Wu, Jupo Ma, Leida Li, and Guangming Shi. Motion deblurring via spatial-temporal collaboration of frames and events. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6531-6539, 2024. 2 +[57] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 3 +[58] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6 +[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2 +[60] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for im + +age deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2 +[61] Huicong Zhang, Haozhe Xie, and Hongxun Yao. Blur-aware spatio-temporal sparse transformer for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2673-2681, 2024. 2 +[62] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 191-207. Springer, 2020. 2 +[63] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 1, 3, 4, 7, 8 +[64] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 5 +[65] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 3, 7 +[66] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 5 \ No newline at end of file diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/images.zip b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..25a3dbdc76e8e48608c5ec5a15f6a7d400edb5ae --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:51c3c3991ffa429b48e8611029df5db1e883d33ad00424ad86f9dd7a0caa4e0e +size 610387 diff --git a/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/layout.json b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..29f541ba82248ecf97f6f934ccdd3d809a055c24 --- /dev/null +++ b/2025/Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring/layout.json @@ -0,0 +1,10035 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 70, + 102, + 539, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 102, + 539, + 140 + ], + "spans": [ + { + "bbox": [ + 70, + 102, + 539, + 140 + ], + "type": "text", + "content": "Event-guided Unified Framework for Low-light Video Enhancement, Frame Interpolation, and Deblurring" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 201, + 161, + 269, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 201, + 161, + 269, + 188 + ], + "spans": [ + { + "bbox": [ + 201, + 161, + 269, + 188 + ], + "type": "text", + "content": "Taewoo Kim* KAIST" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 179, + 192, + 288, + 202 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 179, + 192, + 288, + 202 + ], + "spans": [ + { + "bbox": [ + 179, + 192, + 288, + 202 + ], + "type": "text", + "content": "intelpro@kaist.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 346, + 161, + 416, + 188 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 346, + 161, + 416, + 188 + ], + "spans": [ + { + "bbox": [ + 346, + 161, + 416, + 188 + ], + "type": "text", + "content": "Kuk-Jin Yoon KAIST" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 329, + 192, + 429, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 329, + 192, + 429, + 203 + ], + "spans": [ + { + "bbox": [ + 329, + 192, + 429, + 203 + ], + "type": "text", + "content": "kjyoon@kaist.ac.kr" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 256, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 256, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 256, + 296, + 567 + ], + "type": "text", + "content": "In low-light environments, longer exposure times are commonly used to enhance image visibility; however, this inevitably leads to motion blur. Even with a long exposure time, videos captured in low-light environments still suffer from issues such as low visibility, low contrast, and color distortion. Additionally, the long exposure time results in videos with a low frame rate. Therefore, videos captured in low-light exhibit low visibility and motion blur, as well as low frame rates. To overcome these limitations, we propose a novel problem aimed at transforming motion-blurred, low-frame-rate videos with poor visibility in low-light environments into high-frame-rate videos while simultaneously enhancing their visibility. To tackle this challenge, we leverage the unique advantages of event cameras, which capture scene changes asynchronously, providing superior temporal resolution and a wider dynamic range compared to conventional frame-based cameras. These properties make event cameras particularly effective in reducing motion blur, compensating for low frame rates, and enhancing visibility in low-light conditions. To this end, we developed a hybrid camera system that integrates two RGB cameras and an event camera, capturing a dedicated dataset for this task and proposing novel network architectures to effectively address this problem. The project pages are available at https://sites.google.com/view/eledi-iccv." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 135, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 295, + 696 + ], + "type": "text", + "content": "In our daily lives, low-light conditions are common and occur in various lighting scenarios, both during the day and at night. In such environments, a longer exposure time is generally required to capture the limited available light. However, this extended exposure time inevitably leads to motion blur caused by object movement or dynamic camera motion, while also reducing video frame rates. Additionally," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 555, + 304 + ], + "type": "text", + "content": "videos recorded under these conditions often suffer from reduced visibility due to insufficient lighting, along with significant image noise, even with a long exposure time. As a result, videos captured in low-light conditions typically exhibit low frame rates due to the long exposure time, as well as poor visibility and motion blur." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 304, + 556, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 556, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 556, + 520 + ], + "type": "text", + "content": "With advancements in deep learning, numerous efforts have focused on improving low-light conditions [3, 53, 54], aiming to suppress noise and enhance visibility in such environments. Furthermore, research has been conducted on removing motion blur artifacts through motion deblurring studies [6, 30, 31], as well as on Video Frame Interpolation (VFI) [13, 36, 38, 44], which converts low-frame-rate videos into high-frame-rate ones. While each of these areas has been explored independently, few studies address the combined challenges of multiple degradations in low-light conditions. Among existing works, relatively active research has focused on the problem of joint deblurring and interpolation [30, 63], which aims to restore high-frame-rate, sharp frames from motion-blurred ones. Additionally, some works [16, 65] have investigated integrating motion deblurring and low-light enhancement. However, no existing approach comprehensively tackles all three degradations simultaneously." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 700 + ], + "type": "text", + "content": "To overcome these limitations, we propose a research problem that simultaneously addresses the challenges of low frame rate, motion blur, and low visibility in low-light conditions. Specifically, we aim to simultaneously enhance visibility, reduce motion blur, and increase the frame rate to improve overall video quality in low-light environments. A simple approach would involve cascading separate models for each task. However, such a cascaded manner suffers from error accumulation and inefficient joint optimization, often leading to suboptimal results. Thus, a more effective solution is needed to jointly optimize these interrelated tasks in an end-to-end manner. Moreover, solving these problems using only frame-based information is extremely challenging, particularly in cases of severe degradation, making it an ill-posed problem." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 714 + ], + "type": "text", + "content": "To tackle this issue, we propose leveraging the ad" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 493, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 162, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 162, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 162, + 712 + ], + "type": "text", + "content": "*Work done while at KAIST." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8524" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 232, + 193 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 232, + 193 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 232, + 193 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 232, + 193 + ], + "type": "image", + "image_path": "e089bde8fcf3380ea2e68be893f3a71e322a7ab12e70cbdeef422c3611020cc5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 100, + 194, + 181, + 204 + ], + "lines": [ + { + "bbox": [ + 100, + 194, + 181, + 204 + ], + "spans": [ + { + "bbox": [ + 100, + 194, + 181, + 204 + ], + "type": "text", + "content": "Low-light blurred image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 232, + 70, + 312, + 193 + ], + "blocks": [ + { + "bbox": [ + 232, + 70, + 312, + 193 + ], + "lines": [ + { + "bbox": [ + 232, + 70, + 312, + 193 + ], + "spans": [ + { + "bbox": [ + 232, + 70, + 312, + 193 + ], + "type": "image", + "image_path": "3295b74b9e3312293e6128ae5c0ba86d4c0d69ed462de1213c093ab8f7cdf20e.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 261, + 194, + 285, + 204 + ], + "lines": [ + { + "bbox": [ + 261, + 194, + 285, + 204 + ], + "spans": [ + { + "bbox": [ + 261, + 194, + 285, + 204 + ], + "type": "text", + "content": "t=5/8" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 210, + 555, + 232 + ], + "lines": [ + { + "bbox": [ + 55, + 210, + 555, + 232 + ], + "spans": [ + { + "bbox": [ + 55, + 210, + 555, + 232 + ], + "type": "text", + "content": "Figure 1. Qualitative result of our method on the RELBFI datasets. Our task aims to restore sharp frames captured in a normal light environment within the shutter interval using low-light blurred images and the events." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 312, + 70, + 391, + 193 + ], + "blocks": [ + { + "bbox": [ + 312, + 70, + 391, + 193 + ], + "lines": [ + { + "bbox": [ + 312, + 70, + 391, + 193 + ], + "spans": [ + { + "bbox": [ + 312, + 70, + 391, + 193 + ], + "type": "image", + "image_path": "3e6e47a1b59447b699604f4027a8a9377753aecd3de09424a1c0c03322ed6bad.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 194, + 362, + 203 + ], + "lines": [ + { + "bbox": [ + 339, + 194, + 362, + 203 + ], + "spans": [ + { + "bbox": [ + 339, + 194, + 362, + 203 + ], + "type": "text", + "content": "t=6/8" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 391, + 70, + 470, + 193 + ], + "blocks": [ + { + "bbox": [ + 391, + 70, + 470, + 193 + ], + "lines": [ + { + "bbox": [ + 391, + 70, + 470, + 193 + ], + "spans": [ + { + "bbox": [ + 391, + 70, + 470, + 193 + ], + "type": "image", + "image_path": "5e8226a961a1f7a383a5fdf30df640b81c51223ec04eea5110da2adeaaf579ec.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 417, + 194, + 441, + 204 + ], + "lines": [ + { + "bbox": [ + 417, + 194, + 441, + 204 + ], + "spans": [ + { + "bbox": [ + 417, + 194, + 441, + 204 + ], + "type": "text", + "content": "t=7/8" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 470, + 70, + 549, + 193 + ], + "blocks": [ + { + "bbox": [ + 470, + 70, + 549, + 193 + ], + "lines": [ + { + "bbox": [ + 470, + 70, + 549, + 193 + ], + "spans": [ + { + "bbox": [ + 470, + 70, + 549, + 193 + ], + "type": "image", + "image_path": "cd4136efb7b560ad34626348d90fcc243a287509fd6a507b046fdc9fd5ec8473.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 498, + 194, + 522, + 203 + ], + "lines": [ + { + "bbox": [ + 498, + 194, + 522, + 203 + ], + "spans": [ + { + "bbox": [ + 498, + 194, + 522, + 203 + ], + "type": "text", + "content": "t=8/8" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "spans": [ + { + "bbox": [ + 54, + 248, + 294, + 380 + ], + "type": "text", + "content": "vantages of event cameras, which, inspired by neuromorphic vision systems, detect brightness changes with microsecond-level temporal resolution [8]. Additionally, their superior dynamic range, compared to conventional cameras, allows for high-speed motion capture even in low-light environments. As a result, event cameras are widely used in video enhancement tasks, including motion deblurring [40, 41], frame interpolation [13, 44], low-light enhancement [16, 20, 23], and joint deblurring and interpolation [41], effectively addressing issues like motion blur, low frame rates, low visibility." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 54, + 384, + 295, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 384, + 295, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 384, + 295, + 529 + ], + "type": "text", + "content": "Therefore, we propose a novel task that leverages event cameras to address these multiple degradations in low-light conditions. However, since this novel task has never been addressed in previous works, it is necessary to build a new dataset. To overcome these challenges, we build a sophisticated hybrid camera system to capture a new RELBFI(Real-world Event-guided Low-light enhancement and Blur Frame Interpolation) dataset, consisting of real-world low-light blur videos with a low frame rate, low-light event data, and high-frame-rate normal light sharp videos (ground truth). We believe that these resources will be a valuable asset to the research community." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 713 + ], + "type": "text", + "content": "Beyond proposing this dataset, we also introduce a novel unified end-to-end framework to tackle this challenging task. To achieve this, we design two key modules that play a crucial role in effectively addressing the problem. The first module focuses on fusing multi-temporal information from events and frames under low-light conditions. Given that video frames in such environments often suffer from significant noise and color distortion, this module aims to efficiently integrate multi-temporal and multi-modal information sources. For the second module, we introduce a mechanism to reinforce temporal dependencies among generated video frames during the generation of multiple frames within the shutter period. Since this process is computationally intensive, we designed a lightweight yet effective module to strengthen bidirectional dependencies between" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 248, + 496, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 496, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 496, + 260 + ], + "type": "text", + "content": "frames, thereby enhancing overall efficiency." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 261, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 392 + ], + "type": "text", + "content": "To summarize, the contributions of this paper are as follows: (I) We propose a new research problem that addresses multiple degradations, including noise, motion blur, low visibility, and low frame rate, which commonly occur in low-light conditions, while also converting videos captured at low frame rates into high frame rates. (II) We present a novel unified framework that effectively addresses this problem, demonstrating at least a 2 dB improvement in PSNR compared to state-of-the-art methods. (III) We introduce the first RELBFI dataset, built using a hybrid camera system." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 401, + 405, + 414 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 405, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 405, + 414 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 421, + 429, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 429, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 429, + 434 + ], + "type": "text", + "content": "2.1. Motion Deblurring." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 438, + 555, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 555, + 630 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 555, + 630 + ], + "type": "text", + "content": "Since the advent of deep learning, significant progress has been made in motion deblurring, further accelerated by specialized datasets such as GoPro [29], DVD [39], RealBlur [37], and BSD [62]. In frame-based deblurring, various architectures have been proposed, including multiscale [6], multi-stage [58, 60], transformer-based [42, 49, 59], and frequency-modulated [7, 17, 28] architectures. Video deblurring methods [19, 21, 22, 32, 61, 62] have also emerged. However, deblurring solely from frames remains challenging due to motion information loss, prompting event camera-based research. Early works explored recurrent architectures [10], followed by studies on cross-modal attention-based feature fusion [40], unknown exposure time [12], non-coaxial deblurring [5], frequency-based video deblurring [15], and attention-based cross-modal alignment [14] and spatio-temporal collaboration [56]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 635, + 465, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 465, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 465, + 649 + ], + "type": "text", + "content": "2.2. Video Frame Interpolation." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 714 + ], + "type": "text", + "content": "Research on frame-based video interpolation [1, 2, 18, 26, 35, 36, 38] has made significant progress, but it remains challenging due to the absence of explicit motion information between frames. In contrast, event cameras, with their high temporal resolution, have gained popularity in" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8525" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 72, + 231, + 240 + ], + "blocks": [ + { + "bbox": [ + 60, + 72, + 231, + 240 + ], + "lines": [ + { + "bbox": [ + 60, + 72, + 231, + 240 + ], + "spans": [ + { + "bbox": [ + 60, + 72, + 231, + 240 + ], + "type": "image", + "image_path": "62ac4327a9f769fc244d73516257e7bf7660aa3209d3b16900a4cdb575a06bba.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 146, + 242, + 158, + 252 + ], + "lines": [ + { + "bbox": [ + 146, + 242, + 158, + 252 + ], + "spans": [ + { + "bbox": [ + 146, + 242, + 158, + 252 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 238, + 77, + 375, + 238 + ], + "blocks": [ + { + "bbox": [ + 238, + 77, + 375, + 238 + ], + "lines": [ + { + "bbox": [ + 238, + 77, + 375, + 238 + ], + "spans": [ + { + "bbox": [ + 238, + 77, + 375, + 238 + ], + "type": "image", + "image_path": "62d3ea117d00ef5c5004792ba1d1afbd969da9840ffddb0369a649454a121e2f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 296, + 242, + 308, + 252 + ], + "lines": [ + { + "bbox": [ + 296, + 242, + 308, + 252 + ], + "spans": [ + { + "bbox": [ + 296, + 242, + 308, + 252 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 256, + 554, + 302 + ], + "lines": [ + { + "bbox": [ + 54, + 256, + 554, + 302 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 554, + 302 + ], + "type": "text", + "content": "Figure 2. In the RELBFI dataset description, (a) shows a diagram of the hybrid camera system used to capture the dataset. (b) represents the temporal irradiance relationship for the high-fps RGB camera (used for recording sharp images) and the camera used for capturing blur. (c) presents sample images from the RELBFI dataset: from top to bottom, a low-light blur image, a brightness-adjusted version of the low-light blur image, and a ground truth image captured in normal light conditions." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 388, + 77, + 536, + 232 + ], + "blocks": [ + { + "bbox": [ + 388, + 77, + 536, + 232 + ], + "lines": [ + { + "bbox": [ + 388, + 77, + 536, + 232 + ], + "spans": [ + { + "bbox": [ + 388, + 77, + 536, + 232 + ], + "type": "image", + "image_path": "a9666428c0e8758c25f76efb13d49ed6540daa96ece2cd6de886fc03467e9f3c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 242, + 465, + 251 + ], + "lines": [ + { + "bbox": [ + 453, + 242, + 465, + 251 + ], + "spans": [ + { + "bbox": [ + 453, + 242, + 465, + 251 + ], + "type": "text", + "content": "(c)" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 67, + 363, + 280, + 463 + ], + "blocks": [ + { + "bbox": [ + 55, + 315, + 295, + 360 + ], + "lines": [ + { + "bbox": [ + 55, + 315, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 55, + 315, + 295, + 360 + ], + "type": "text", + "content": "Table 1. Comparison of HighREV and RELBFI Datasets. In the table, NL\" refers to images taken in normal-light environments, and \"LL\" refers to images taken in low-light environments. \"LFR\" refers to low frame rate, and \"HFR\" refers to high frame rate." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 67, + 363, + 280, + 463 + ], + "lines": [ + { + "bbox": [ + 67, + 363, + 280, + 463 + ], + "spans": [ + { + "bbox": [ + 67, + 363, + 280, + 463 + ], + "type": "table", + "html": "
HighREV [41]RELBFI
LFR RGB InputNL blurLL blur
Event inputNL eventLL event
HFR RGB GTNL sharpNL sharp
Blur typeSyntheticReal
Event typeRealReal
#N of sharp images2858932629
#N of sequences3067
", + "image_path": "bd4cdca7b02a1653db924ecd786c6160423930697573cf7232f64295d4119dfe.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": "video interpolation tasks. Early work like TimeLens [43] introduced event-based interpolation using a cascaded U-Net framework, while TimeLens++ [45] enhanced this by estimating non-linear motion with a motion spline estimator. Later studies, including A" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 55, + 480, + 296, + 601 + ], + "type": "text", + "content": "OF [51] and the unsupervised TimeReplayer [9], further advanced the field. Recently, Kim et al. [13] introduced an event-image fusion VFI framework for inter-frame optical flow estimation, while the follow-up study, TimeLens-XL [27], presented a VFI framework with iterative flow updates." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 610, + 194, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 194, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 194, + 624 + ], + "type": "text", + "content": "2.3. Low-light Enhancement." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Many frame-based low-light enhancement (LLE) studies focus on improving visibility and reducing noise using image information. These efforts have led to various approaches, including SNR-based transformers [53], Retinex-based methods [3, 52], gamma correction [47], implicit neural representations [55], diffusion models [48, 57], and other techniques [25, 54, 65]. Leveraging its high dynamic range," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 555, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 401 + ], + "type": "text", + "content": "the event camera excels in capturing motion under low light, making it suitable for LLE. Liu et al. [24] used synthetic events for LLE due to dataset limitations, while Liang et al. [23] modeled multimodal coherence. Recently, Liang et al. [20] proposed an SNR-based framework and dataset for event-guided LLE, and Kim et al. [16] introduced a joint framework for event-guided LLE and deblurring." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 408, + 531, + 421 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 408, + 531, + 421 + ], + "spans": [ + { + "bbox": [ + 313, + 408, + 531, + 421 + ], + "type": "text", + "content": "2.4. Joint deblurring and frame interpolation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 312, + 426, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 426, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 312, + 426, + 555, + 594 + ], + "type": "text", + "content": "The joint task of deblurring and interpolation aims to simultaneously remove blur and generate high-frame-rate videos within a single framework. Jin et al. [11] first addressed this without event cameras using a deblur-then-interpolate approach. Recently, DeMFI [30] combined flow-guided attention and recursive boosting, while BiT [63] introduced a Swin Transformer-based solution. Event camera-based research includes EDI [33, 34], which used a physical model, REFID [41] leveraging a recurrent architecture, and Weng et al. [50], tackling blur frame interpolation under unknown exposure times. Unlike these works, we leverage the robust properties of event cameras in low-light conditions to jointly perform low-light enhancement, deblurring, and frame interpolation." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 604, + 421, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 604, + 421, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 604, + 421, + 616 + ], + "type": "text", + "content": "3. RELBFI Datasets" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 624, + 498, + 636 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 498, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 498, + 636 + ], + "type": "text", + "content": "3.1. Comparisons with Other Datasets." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 554, + 689 + ], + "type": "text", + "content": "Since the advent of deep learning, various event- and frame-based video enhancement datasets [13, 16, 20, 63] have emerged. Some studies [41, 63] focus on motion deblurring and frame interpolation under normal lighting." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Sun et al. [41] introduced a dataset for joint deblurring and interpolation using an event camera, where motion blur" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8526" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 541, + 286 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 541, + 286 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 541, + 286 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 541, + 286 + ], + "type": "image", + "image_path": "4ae18f5f76e9fad75caf4e21d552afd2bf9a69ab8343a58e00334cd64b02f237.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 290, + 529, + 302 + ], + "lines": [ + { + "bbox": [ + 79, + 290, + 529, + 302 + ], + "spans": [ + { + "bbox": [ + 79, + 290, + 529, + 302 + ], + "type": "text", + "content": "Figure 3. Overall framework of the proposed method. \"TFG\" in the blue box denotes the Template Feature Generator block." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 318, + 294, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 294, + 365 + ], + "type": "text", + "content": "was synthetically generated by merging adjacent frames. However, such synthetic datasets fail to capture real-world motion blur, as they overlook physical factors like noise variations from different exposure times." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 365, + 295, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 365, + 295, + 425 + ], + "spans": [ + { + "bbox": [ + 54, + 365, + 295, + 425 + ], + "type": "text", + "content": "Recently, Kim et al. [16] proposed the RELED dataset for simultaneous low-light enhancement and deblurring using an event camera. However, its low frame rate limits its effectiveness in converting low-frame-rate, low-light videos into high-frame-rate, normal-light videos." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "spans": [ + { + "bbox": [ + 54, + 426, + 295, + 474 + ], + "type": "text", + "content": "To date, no existing research or dataset has directly tackled the transformation of low-frame-rate, low-light, noisy, and motion-blurred inputs into high-frame-rate, noise-free videos under normal lighting by leveraging event data." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 480, + 216, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 216, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 216, + 493 + ], + "type": "text", + "content": "3.2. RELBFI dataset acquisitions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 498, + 295, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 498, + 295, + 557 + ], + "spans": [ + { + "bbox": [ + 54, + 498, + 295, + 557 + ], + "type": "text", + "content": "Synthetic-based datasets struggle to capture the complexity of real-world degradations. Therefore, building a real-world dataset is both practical and challenging, especially when simultaneously capturing high-frame-rate sharp video, low-light blurred video, and low-light events." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "content": "To address this, we build the RELBFI (Real-world Event-guided Low-light Video Enhancement and Blur Frame Interpolation) dataset using a hybrid camera system with beam splitters. As shown in Fig. 2(a), the system includes two RGB cameras (Blackfly BFS-U3-16S2C-CS) and an event camera (Propheesee Gen4). One RGB camera captures sharp images at 166fps with a 3ms exposure and a 3ms readout dead time, receiving " + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "25\\%" + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "content": " of the total light as it passes through the beam splitter twice without an ND filter. The other RGB camera records blurred images in low-light conditions with a 24ms exposure, eight times longer, and uses a 1/64 ND filter, receiving only " + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "inline_equation", + "content": "0.39\\%" + }, + { + "bbox": [ + 54, + 558, + 295, + 714 + ], + "type": "text", + "content": " of the original light. As a result, each blurred image has eight times less" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 318, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 553, + 330 + ], + "type": "text", + "content": "light intensity than a sharp image, resulting in low visibility." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "text", + "content": "Furthermore, to illustrate real-world frame acquisition [12, 41, 63], capturing blurred images involves exposure time and readout dead time, collectively termed the shutter interval [12]. To simulate readout dead time, we used a " + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "inline_equation", + "content": "24\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "text", + "content": " exposure followed by a " + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "inline_equation", + "content": "24\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "text", + "content": " dead time, yielding one low-light blurred image per " + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "inline_equation", + "content": "48\\mathrm{ms}" + }, + { + "bbox": [ + 313, + 330, + 553, + 437 + ], + "type": "text", + "content": " (21fps) and eight sharp images under normal lighting (166fps). The event camera, like the blur-recording RGB camera, employs a 1/64 ND filter for low-light capture." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 437, + 554, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 605 + ], + "type": "text", + "content": "Regarding calibration, all cameras in the hybrid system are mechanically fixed, but slight adjustments, such as rotation and baseline alignment, are required. To this end, we estimated the homography matrix for each camera and performed additional calibration. As shown in Fig. 2(a), the devices were temporally synchronized via hardware using an external trigger. Using this setup, we captured 67 urban scenes featuring dynamic elements like vehicles, moving objects, and flickering streetlights, along with rich textures. As shown in Tab. 1, RELBFI provides diverse sequences tailored for developing and evaluating methods in this challenging task. It serves as a practical resource to support joint low-light enhancement, deblurring, and interpolation in real-world scenarios." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 616, + 425, + 629 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 425, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 425, + 629 + ], + "type": "text", + "content": "4. Proposed Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 636, + 382, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 636, + 382, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 636, + 382, + 647 + ], + "type": "text", + "content": "4.1. Overview" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": "Preliminary As explained in Sec. 3, the shutter interval consists of exposure time and readout deadtime. During the readout deadtime, events are recorded, but no blurred frames are captured. We define " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " as the total number of sharp frames captured during a shutter interval " + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 653, + 554, + 713 + ], + "type": "text", + "content": " (ms)," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8527" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 68, + 293, + 210 + ], + "blocks": [ + { + "bbox": [ + 58, + 68, + 293, + 210 + ], + "lines": [ + { + "bbox": [ + 58, + 68, + 293, + 210 + ], + "spans": [ + { + "bbox": [ + 58, + 68, + 293, + 210 + ], + "type": "image", + "image_path": "fa6c05dcf02dcfc6de3426008d6787ef8549aa1ba49258694ca98e334a571e74.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "lines": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": "Figure 4. The MM-ISTFF module in the scale factor " + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": ". In the figure, " + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s^{\\mathrm{fw}}" + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_s^{\\mathrm{bw}}" + }, + { + "bbox": [ + 55, + 216, + 295, + 251 + ], + "type": "text", + "content": " represent the forward and backward dynamic filters, respectively, as defined in Eq.(1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "N = E + R" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": " representing the number of the frames from exposure and readout deadtime, respectively. During a single shutter interval " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": ", one blurred frame " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "B^t" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": " is captured, and events are generated corresponding to both the exposure time and readout deadtime. The events " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "\\{e_i\\}_t" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": " during shutter interval " + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 263, + 296, + 347 + ], + "type": "text", + "content": " are transformed into event voxels [66] with 32 bins." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": "Overall framework The proposed framework is illustrated in Fig. 3. We utilize 3 blurred images and 3 corresponding event voxel grids from shutter intervals " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "t - 1" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "t + 1" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": ". Given these inputs " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\{B^{t-1}, B^t, B^{t+1}\\}" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "\\{E^{t-1}, E^t, E^{t+1}\\}" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": ", the network estimates " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": " normal light sharp frames at the shutter interval " + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 348, + 296, + 419 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": "The process consists of two main steps. First, the MM-ISTFF (Multi-Modal Inter-frame Spatio-Temporal Feature Fusion) module combines spatio-temporal event data with blurred video frames to improve video quality and reduce low-light noise, producing fused features " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s=0}^{t}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": ". Next, the event voxel grid for " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " is divided into " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " temporal units " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\{E_0^t,\\ldots,E_{N-1}^t\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": ", which are combined with " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s=0}^{t}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " to generate " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " temporal template features " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(\\mathcal{T})_0,\\ldots,\\mathcal{F}(\\mathcal{T})_{N-1}\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": ". Next, the IS-TFF (Inter-Shutter Interval Temporal Feature Fusion) based-encoder module leverages temporal information within the shutter interval then uses these features to produce an enhanced feature map " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": ". Finally, the decoder generates " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " normal light sharp frames " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "\\{S_0,\\ldots,S_{N-1}\\}" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 421, + 296, + 592 + ], + "type": "text", + "content": " shutter interval." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 295, + 624 + ], + "type": "text", + "content": "4.2. Multi-Modal Inter-frame Spatio-Temporal Feature Fusion(MM-ISTFF) Module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "In low-light conditions, frame-based information is severely affected by noise and low visibility, while motion blur further exacerbates the degradation. In such scenarios, events also suffer from significant degradation. Therefore, it is crucial to fuse features by complementing the spatio-temporal visual information from images with the rich temporal details from events, effectively utilizing them" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 552, + 247 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 552, + 247 + ], + "type": "image", + "image_path": "0af3b3d77c22ae4440a0b502ecba8985cebc7facb9b922c2a4be71ede8a6d973.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 251, + 553, + 275 + ], + "lines": [ + { + "bbox": [ + 313, + 251, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 251, + 553, + 275 + ], + "type": "text", + "content": "Figure 5. Illustration of the IS-TFF module. The " + }, + { + "bbox": [ + 313, + 251, + 553, + 275 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 251, + 553, + 275 + ], + "type": "text", + "content": " after the forward and backward units indicates the scale factor." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": "for video enhancement. To address this, we propose the MM-ISTFF module, designed to mine discriminative temporal information between frames and events across different shutter intervals. For this, we extract multi-scale multi-modal features " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(B)_s^{t + 1}\\}" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t,\\mathcal{F}(E)_s^{t + 1}\\}" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": " from the low-light blurred and event voxel inputs. As shown in Fig.3, we aim to leverage these multi-modal inputs to extract discriminative spatio-temporal information across multiple visual scales. To achieve this, we use past and future information from the current shutter interval " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": ". For simplicity, only the process for past information is described, as the same applies to future information. As shown in Fig.4, to integrate spatiotemporal information from past multi-modal data, we first apply a conv. layer with LeakyReLU to the concatenated features " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "[\\mathcal{F}(B)_s^{t - 1},\\mathcal{F}(B)_s^t,\\mathcal{F}(E)_s^{t - 1},\\mathcal{F}(E)_s^t]" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": ", generating the feature " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": ". To extract discriminative spatio-temporal information, pixel-wise spatial dynamic filters [64] are applied to obtain content-adaptive features. Here, to estimate a dynamic filter for each pixel, we use " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "f_{s}^{fw} = f_{\\mathrm{filter}}^{\\mathrm{fw}}(\\mathbf{V}_{s}^{\\mathrm{fw}})" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": " to generate dynamic filters, which are then applied to the original feature " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "\\mathbf{V}_s^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{filter}}" + }, + { + "bbox": [ + 313, + 284, + 555, + 559 + ], + "type": "text", + "content": " We apply dynamic filter as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 395, + 569, + 555, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 569, + 555, + 585 + ], + "spans": [ + { + "bbox": [ + 395, + 569, + 555, + 585 + ], + "type": "interline_equation", + "content": "\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} = f _ {s} ^ {\\mathrm {f w}} \\otimes \\mathbf {V} _ {s} ^ {\\mathrm {f w}} \\tag {1}", + "image_path": "df530a818c1fcb1df4d28e68234f39da2b0e2c5c9fc6c71d476e743f43619656.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "inline_equation", + "content": "f_{s}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "text", + "content": " represents the estimated filter, and " + }, + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 313, + 588, + 554, + 649 + ], + "type": "text", + "content": " denotes the dynamic filter application. This process extracts content-adaptive information, but in low-light conditions, selective regional features are needed due to noise and degradation. Thus, we apply region-wise attention for feature filtering." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 352, + 654, + 555, + 671 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 352, + 654, + 555, + 671 + ], + "spans": [ + { + "bbox": [ + 352, + 654, + 555, + 671 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathbf {V}} _ {s} ^ {f w} = \\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}} \\odot \\sigma (\\mathrm {L R e L U} ((W _ {3} (\\hat {\\mathbf {V}} _ {s} ^ {\\mathrm {f w}})))) \\tag {2}", + "image_path": "1a75d98ddb101bb3a73199a19032b5b70f84ed15ec4da1615ec2890d597b6331.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " denotes the sigmoid operation, and " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "W_{3}" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " and LReLU represent the " + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "inline_equation", + "content": "3\\times 3" + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": " conv. layer and the LeakyReLU operation, respectively. Similarly, future multi-modal informa" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8528" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": "tion is utilized to obtain " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": ". These two features " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "(\\tilde{\\mathbf{V}}_s^{\\mathrm{fw}}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{V}}_s^{\\mathrm{bw}})" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " are combined via channel-wise concatenation followed by a ResBlock to produce the final feature " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_s^t" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": ". This process is repeated for " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "s = \\{0,1,2\\}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": ", with " + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s=0}^{t}" + }, + { + "bbox": [ + 55, + 71, + 294, + 133 + ], + "type": "text", + "content": " serving as input to the IS-TFF module." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 143, + 295, + 168 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 295, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 295, + 168 + ], + "type": "text", + "content": "4.3. Inter-Shutter Interval Temporal Feature Fusion(IS-TFF) Module." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": "Template Feature Generation After the MM-ISTFF module, we obtain a spatio-temporally fused feature " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s=0}^{t}" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " centered at shutter interval " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ". To estimate " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " sharp images within the shutter interval, template features capturing the characteristics of each timestep must be generated for all " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " timesteps. To address this, we first divide the event voxel grid into " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " temporal units, each with an equal time interval " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\Delta t_{unit} = T / N" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ". This gives us temporally divided unit event voxels " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "E_k^t" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ", which are then encoded to generate temporal event features. Therefore, we obtain the temporally divided unit event voxel " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\{E_k^t\\}" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " with temporal index " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "k \\in \\{0, \\dots, N-1\\}" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ". Subsequently, each event voxel unit is encoded to generate temporal event features " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(E_k^t) \\in \\mathbb{R}^{C \\times H \\times W}\\}_k" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ". This process is repeated from " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "k = 0" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "N-1" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": ". Next, we define a zero tensor " + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "inline_equation", + "content": "Z_k \\in \\mathbb{R}^{N \\times C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 174, + 296, + 365 + ], + "type": "text", + "content": " and populate it with temporal unit event features for each timestep:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 74, + 376, + 276, + 390 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 376, + 276, + 390 + ], + "spans": [ + { + "bbox": [ + 74, + 376, + 276, + 390 + ], + "type": "interline_equation", + "content": "Z _ {k} [ k,:,:,: ] = \\mathcal {F} \\left(E _ {k} ^ {t}\\right), \\quad \\forall k \\in \\{0, 1, \\dots , N - 1 \\}.", + "image_path": "0da75bb6bfafc86d73b4a9267d3bc28caf6b2cdb52a52adb5436decef116cd34.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "spans": [ + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "This tensor " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "Z_{k}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " contains unit event voxel features " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(E_k^t)" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " filled only at the " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": "-th position along the temporal axis, with the rest initialized to zero. As a result, it encapsulates information about which timestep is being interpolated and the motion information required for the interpolation. Afterward, for " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "k = \\{0,\\dots ,N - 1\\}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": ", we concatenate the " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "Z_{k}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " tensor with " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s = 0}^{t}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": ", pass them through a Residual blocks, and generate the template features " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{F}(\\mathcal{T})_k\\}" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 400, + 296, + 496 + ], + "type": "text", + "content": " timesteps." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": "IS-TFF based Bidirectional Encoder Module. Through the template feature generator, we obtain the feature " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{T})_k" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " for each timestep " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ", capturing the essential information to interpolate the normal light sharp frame " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "S_{k}" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": ". By using these features, in this module, we aim to enhance the temporal dependency for timestep " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " within the shutter interval " + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 497, + 296, + 581 + ], + "type": "text", + "content": " using multi-scale temporal information." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": "While existing methods [4, 46] enhance temporal dependency within frames, they are computationally intensive, especially when applied to multiple timesteps at various scales. To address this, we propose a lightweight yet effective temporal feature fusion method within the shutter interval, which enables information sharing across the temporal dimension. As shown in Fig. 3, the IS-TFF-based encoder modules operate bidirectionally at each visual scale " + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "inline_equation", + "content": "(s = 0,1,2)" + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": ". For clarity, we describe the process at scale " + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 582, + 296, + 712 + ], + "type": "text", + "content": " in the forward direction; the same applies to the backward direction and other scales." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": "As shown in Fig. 5, we first extract the features through " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{T})_{k,s}^{fw}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " processed at the previous scale, passing them through a conv. layer to transform them into " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": ". For the fusion of timesteps " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " in forward direction, the fusion of " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{T})_{s,k}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{G}(\\mathcal{T})_{s,k - 1}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " begins by applying a conv. layer with LeakyReLU to their channel-wise concatenation, producing the feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": ". Then, this feature is split along the channel dimension into two parts, " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k - 1}" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": ". After mixing the adjacent features through concatenation and the first conv. layer, we analyze their temporally significant parts through the channel-wise dimension. For this, we first apply AAP (adaptive average pooling) to compress the spatial dimensions, as follows: " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{P}(\\mathcal{G})_{s,k}^{\\mathrm{fw}} = \\mathrm{AAP}(\\tilde{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}})" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": ". Next, we apply attention [58] to the feature at each timestep " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "inline_equation", + "content": "k - 1" + }, + { + "bbox": [ + 313, + 72, + 555, + 258 + ], + "type": "text", + "content": " as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 264, + 553, + 290 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 264, + 553, + 290 + ], + "spans": [ + { + "bbox": [ + 321, + 264, + 553, + 290 + ], + "type": "interline_equation", + "content": "\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = \\tilde {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} \\cdot \\sigma \\left(W _ {3} \\left(\\operatorname {R e L U} \\left(W _ {3} \\left(\\mathcal {P} (\\mathcal {G}) _ {s, k} ^ {\\mathrm {f w}}\\right)\\right)\\right)\\right) \\tag {3}", + "image_path": "8f94dc202667d404b1f6fabf395b6844e84ebf1218903b48dec0853ebfbce1ac.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "W_{3}" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " conv. layer, ReLU represents the rectified linear unit activation function, and " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " is the sigmoid activation function. Similarly, the feature " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{G}}(\\mathcal{T})_{s,k-1}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "k-1" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " timestep is obtained in the same manner. Finally, we fuse the " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "k-1" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 290, + 554, + 350 + ], + "type": "text", + "content": " timestep features as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 342, + 354, + 553, + 370 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 354, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 342, + 354, + 553, + 370 + ], + "type": "interline_equation", + "content": "\\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} = W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}}\\right) + W _ {1} \\left(\\bar {\\mathcal {G}} (\\mathcal {T}) _ {s, k - 1} ^ {\\mathrm {f w}}\\right) \\tag {4}", + "image_path": "add006b380f4e480b1c8e05b860c7cb3f59b3e4f73fe1391c310c3dbb20c337f.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "W_{1}" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " denotes a " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "1\\times 1" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " conv. layer, we pass " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " to the next IS-TFF encoder block with downsampling. This process is repeated for " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "s = \\{0,1,2\\}" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": ", producing outputs at three scales. This process is also repeated through a separate backward encoder, resulting in a set of feature maps " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{fw}}\\}" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " from the forward encoder pass and " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "\\{\\hat{\\mathcal{G}} (\\mathcal{T})_{s,k}^{\\mathrm{bw}}\\}" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " from the backward encoder pass at timestep " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": ". We then fuse the forward and backward features at scale " + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 313, + 376, + 554, + 472 + ], + "type": "text", + "content": " as follows." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 369, + 483, + 553, + 498 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 369, + 483, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 369, + 483, + 553, + 498 + ], + "type": "interline_equation", + "content": "\\mathcal {D} (\\mathcal {T}) _ {s, k} = \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {f w}} + \\hat {\\mathcal {G}} (\\mathcal {T}) _ {s, k} ^ {\\mathrm {b w}} \\tag {5}", + "image_path": "85f9e42b91c2bc316fcbe54b57ad12409eaf7414318d2fa2d8a47f582e3a7342.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": "With this approach, we generate a multi-scale feature map set " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}(\\mathcal{T})_{s,k}\\}" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "s\\in \\{0,1,2\\}" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": ", that reinforces rich temporal information within the shutter interval " + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 500, + 554, + 536 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 542, + 381, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 381, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 381, + 553 + ], + "type": "text", + "content": "4.4. Decoders" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": "After the IS-TFF based encoding stage, we decode the feature map for each timestep " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": ". For this, we apply two separate decoders, one for exposure time(generating " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "E" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": " frames) and another for readout-dead time(generating " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "R" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": " frames), share the same structure but have different weights. Consequently, we simultaneously estimate " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": " sharp frames " + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "inline_equation", + "content": "\\{S_0, \\ldots, S_{N-1}\\}" + }, + { + "bbox": [ + 313, + 559, + 554, + 643 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 652, + 395, + 666 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 672, + 479, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 479, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 479, + 685 + ], + "type": "text", + "content": "5.1. Datasets and Baseline Settings" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "To the best of our knowledge, no prior work has jointly tackled low-light video enhancement, frame interpolation," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8529" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 76, + 129, + 533, + 285 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": "Table 2. Quantitative evaluation results for the tasks of joint low-light enhancement, deblurring, and frame interpolation on the RELBFI dataset. Bold and underline denote the best and second-best performance, respectively. The symbol " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " denotes event-guided methods. The number preceding the parameter corresponds to the deblurring method, while the number following it corresponds to the video frame interpolation methods. The runtime is defined as the time required to generate one or eight " + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "inline_equation", + "content": "1280 \\times 896" + }, + { + "bbox": [ + 55, + 70, + 555, + 125 + ], + "type": "text", + "content": " normal-light sharp images, and was measured using an RTX 3090 GPU." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 76, + 129, + 533, + 285 + ], + "lines": [ + { + "bbox": [ + 76, + 129, + 533, + 285 + ], + "spans": [ + { + "bbox": [ + 76, + 129, + 533, + 285 + ], + "type": "table", + "html": "
Method\nDeblur+VFI / JointPSNR↑SSIM↑Params(MB)↓Runtime
1×[s]↓8×[s]↓
LEDNet [65] + TimeLens† [44]24.420.8017.41 + 72.30.3622.894
LEDNet [65] + CBMNet† [44]24.550.8027.41 + 15.40.9017.204
EFNet† [40] + TimeLens† [44]27.040.8438.47 + 72.30.3773.018
EFNet† [40] + CBMNet† [44]27.420.8528.47 + 15.40.9167.328
ELEDNet† [16] + TimeLens† [44]28.650.86312.8 + 72.30.4863.892
ELEDNet† [16] + CBMNet† [44]29.280.87712.8 + 15.41.0258.202
DEMFI [30]24.220.7947.411.2019.606
BiT [63]26.560.83111.30.2341.875
REFID† [41]29.490.87615.90.2071.656
Ours†31.620.9029.40.1481.184
", + "image_path": "5b419bea1be519ed944f6d97033635dc291c34f3fe38fa4a9f45425fb39207a3.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 318, + 293, + 429 + ], + "blocks": [ + { + "bbox": [ + 55, + 292, + 295, + 315 + ], + "lines": [ + { + "bbox": [ + 55, + 292, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 55, + 292, + 295, + 315 + ], + "type": "text", + "content": "Table 3. Quantitative evaluation results for the joint low-light enhancement and deblurring tasks on the RELBFI datasets." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 318, + 293, + 429 + ], + "lines": [ + { + "bbox": [ + 58, + 318, + 293, + 429 + ], + "spans": [ + { + "bbox": [ + 58, + 318, + 293, + 429 + ], + "type": "table", + "html": "
MethodsPSNRSSIMParams(MB)
LEDNet [65]24.850.8077.41
EFNet [40]27.630.8628.47
DEMFI [30]24.150.8057.41
BiT [63]24.690.81211.3
REFID [41]29.080.87815.9
ELEDNet [16]30.480.89912.8
Ours31.650.9079.4
", + "image_path": "2510371be0efdf9e4f01fab5c96b6a6b50263d651d5eda71000c54a215c2dc32.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 437, + 297, + 714 + ], + "type": "text", + "content": "and deblurring. Therefore, we established our own baselines and datasets for comprehensive benchmarking. Since no existing dataset includes high-speed ground-truth frames captured under normal lighting, low-frame-rate low-light blurry video frames, and low-light events, we conducted all experiments using our RELBFI dataset. The RELBFI dataset consists of 67 sequences, with 47 used for training and 20 reserved for testing. We established two types of baselines. The first is a cascaded approach, where deblurring and video frame interpolation methods are applied sequentially. For joint low-light enhancement and deblurring, we utilized LEDNet [65] (Frame-based), ELEDNet [16] (Event-based), and EFNet [40] (Event-based). These methods were trained for the same number of epochs as our network. For video frame interpolation (VFI), we used widely adopted event-based methods, Timelens [44] and CBMNet [13]. As these methods typically take longer to converge compared to our deblurring network, we trained them for additional epochs to achieve sufficient convergence. Finally, we evaluated state-of-the-art joint deblurring and frame interpolation methods, including DEMFI [30] (Frame-based), BiT [63] (Frame-based), and REFID [41] (Event-based). These methods were also" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 318, + 307, + 548, + 376 + ], + "blocks": [ + { + "bbox": [ + 345, + 293, + 523, + 304 + ], + "lines": [ + { + "bbox": [ + 345, + 293, + 523, + 304 + ], + "spans": [ + { + "bbox": [ + 345, + 293, + 523, + 304 + ], + "type": "text", + "content": "Table 4. Ablation study on the RELBFI datasets." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 307, + 548, + 376 + ], + "lines": [ + { + "bbox": [ + 318, + 307, + 548, + 376 + ], + "spans": [ + { + "bbox": [ + 318, + 307, + 548, + 376 + ], + "type": "table", + "html": "
MethodsMM-ISTFFIS-TFF(fw)IS-TFF(bw)PSNRs
Ver.127.07
Ver.2✓30.81
Ver.3✓✓31.35
Ver.4✓✓✓31.62
", + "image_path": "23a42d7d3afe8301ad107ec79cc9b6cb4b72e9d62df41d90c2782ed69caec610.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 419, + 550, + 448 + ], + "blocks": [ + { + "bbox": [ + 313, + 383, + 555, + 416 + ], + "lines": [ + { + "bbox": [ + 313, + 383, + 555, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 383, + 555, + 416 + ], + "type": "text", + "content": "Table 5. Performance effectiveness analysis of the MM-ISTFF module. The (-) symbol indicates that training was unstable and did not converge." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 419, + 550, + 448 + ], + "lines": [ + { + "bbox": [ + 317, + 419, + 550, + 448 + ], + "spans": [ + { + "bbox": [ + 317, + 419, + 550, + 448 + ], + "type": "table", + "html": "
Methodsw/o Eq.(1)w/o Eq.(2)w/o BiOurs
PSNRs30.84(-)31.0431.62
", + "image_path": "ba58381fda17cbbf28d99a2f936d1f53c0e9b27487c1873f39236989da76bb2b.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 467, + 548, + 495 + ], + "blocks": [ + { + "bbox": [ + 314, + 453, + 553, + 464 + ], + "lines": [ + { + "bbox": [ + 314, + 453, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 314, + 453, + 553, + 464 + ], + "type": "text", + "content": "Table 6. Network component analysis of the dual decoder module." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 467, + 548, + 495 + ], + "lines": [ + { + "bbox": [ + 318, + 467, + 548, + 495 + ], + "spans": [ + { + "bbox": [ + 318, + 467, + 548, + 495 + ], + "type": "table", + "html": "
MethodsSingle weight-sharing decoderOurs
PSNRs31.1631.62
", + "image_path": "3b796851b7ce5dd512703903aca24a5d6394b8270412f08f1abac831dad4e6b9.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 509, + 534, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 534, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 534, + 521 + ], + "type": "text", + "content": "trained for the same number of epochs as our network." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 529, + 550, + 541 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 550, + 541 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 550, + 541 + ], + "type": "text", + "content": "5.2. Experimental results on the RELBFI datasets" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 555, + 582 + ], + "type": "text", + "content": "We have provided the results of the quantitative evaluation on the RELBFI dataset in Tab.3(Joint LLE and deblurring) and Tab.2(Joint LLE and deblurring and interpolation)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 714 + ], + "type": "text", + "content": "Joint LLE and deblurring and interpolation In Tab.2, we evaluated two groups. Cascaded solutions, where joint deblurring and LLE methods are followed by video frame interpolation (VFI), showed significant performance degradation in frame-based methods like LEDNet [65]. Performance improved with event-based networks such as EFNet [40] and ELEDNet [16], but these cascaded solutions still had a PSNR gap of 2.34 dB to 4.58 dB compared to our network. For joint-based solutions, frame-based methods like DEMFI [30] and BiT [63] showed a PSNR difference of 5.06 dB to 7.40 dB, while compared to REFID [41]," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "8530" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 550, + 292 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 550, + 292 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 550, + 292 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 550, + 292 + ], + "type": "image", + "image_path": "ae680d372260006a55bc1ec40bdcb59114c0e7772538d45f39440404bc3ffebe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "lines": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 555, + 316 + ], + "type": "text", + "content": "Figure 6. Qualitative results on the RELBFI dataset. From left to right, the images show (a) input low-light blurry images with brightness adjustment, (b) low-light events, (c) BiT [63], (d) REFID [41], (e) Ours, and (f) ground-truth normal-light sharp images." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 331, + 294, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 331, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 54, + 331, + 294, + 390 + ], + "type": "text", + "content": "the state-of-the-art joint event-based method, our network outperformed by 2.13 dB with a large margin. From the perspective of computational cost, our method achieves the best performance with fewer computations and reasonable parameters compared to all state-of-the-art methods." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 391, + 295, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 391, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 54, + 391, + 295, + 510 + ], + "type": "text", + "content": "Joint LLE and deblurring In Tab.3, we conducted an evaluation of the joint low-light enhancement and deblurring method, excluding interpolation. When compared to ELEDNet, the best state-of-the-art method, it showed a performance difference of 1.17 dB. Comparisons with the other frame-based and event-based methods revealed differences ranging from a maximum of 7.5 dB to a minimum of 2.57 dB. This also evidences the robustness of the proposed method in both deblurring and low-light enhancement (LLE) tasks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 149, + 529 + ], + "type": "text", + "content": "5.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "All ablation experiments were conducted on the RELBFI dataset, with networks trained for 200 epochs. The results are summarized in Tab. 4. For the baseline model, the blur input feature " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(B)^t" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": " at the " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": "-shutter interval was used as input for the IS-TFF module-based networks (stage 2) instead of " + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\mathcal{K})_{s=0}^t" + }, + { + "bbox": [ + 55, + 533, + 295, + 605 + ], + "type": "text", + "content": " from the MM-ISTFF module." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "MM-ISTFF module The MM-ISTFF module extracts discriminative spatio-temporal information from multi-modal data. As shown in Tab. 4, Ver. 2 outperforms Ver. 1 by " + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "inline_equation", + "content": "3.74\\mathrm{dB}" + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": ". In Tab. 5, we analyze MM-ISTFF components. Removing the pixel-wise filter from Eq. (1) caused a 0.78 dB drop, while removing regional feature selection from Eq. (2) led to unstable training and non-convergence, underscoring its importance for region-wise selection of degraded multi-modal data. Limiting temporal fusion to current and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "content": "future parts (third row) resulted in a " + }, + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "inline_equation", + "content": "0.50\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 331, + 553, + 355 + ], + "type": "text", + "content": " drop, highlighting the benefits of bidirectional fusion." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "text", + "content": "IS-TFF module We validated the effectiveness of the IS-TFF module. Comparing Ver. 2 and Ver. 3 in Tab. 4, incorporating IS-TFF in the forward direction improved performance by " + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "inline_equation", + "content": "0.5\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "text", + "content": ". Adding the IS-TFF module in the backward direction (Ver. 3 vs. Ver. 4) yielded an additional " + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "inline_equation", + "content": "0.33\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "text", + "content": " gain, resulting in a total improvement of " + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "inline_equation", + "content": "0.83\\mathrm{dB}" + }, + { + "bbox": [ + 313, + 357, + 554, + 430 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 431, + 555, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 528 + ], + "type": "text", + "content": "The importance of the separate dual decoders In the decoder following the IS-TFF encoder, separate decoders with different weights are used for frames from the exposure time and the readout deadtime. As shown in Tab. 6, this improves performance " + }, + { + "bbox": [ + 313, + 431, + 555, + 528 + ], + "type": "inline_equation", + "content": "(+0.46\\mathrm{dB})" + }, + { + "bbox": [ + 313, + 431, + 555, + 528 + ], + "type": "text", + "content": " by applying different decoding methods—blur decomposition or interpolation—tailored to the exposure and readout periods, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 391, + 551 + ], + "type": "text", + "content": "6. Conclusions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 555, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 555, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 555, + 555, + 662 + ], + "type": "text", + "content": "This paper addresses, for the first time, the novel and previously unexplored problem of unified event-guided joint low-light enhancement, deblurring, and frame interpolation. To solve this task, we propose an integrated framework that jointly handles these sub-tasks with high precision, achieving state-of-the-art performance in both accuracy and computational efficiency. We expect our work to offer initial insights that may support future research in event-guided video restoration." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korean government (MSIT) (No. NRF-2022R1A2B5B03002636)." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "8531" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Wenbo Bao, Wei-Sheng Lai, Chao Ma, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Depth-aware video frame interpolation. In IEEE Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 192 + ], + "type": "text", + "content": "[2] Wenbo Bao, Wei-Sheng Lai, Xiaoyun Zhang, Zhiyong Gao, and Ming-Hsuan Yang. Memc-net: Motion estimation and motion compensation driven neural network for video interpolation and enhancement. IEEE transactions on pattern analysis and machine intelligence, 43(3):933-948, 2019. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 295, + 248 + ], + "type": "text", + "content": "[3] Yuanhao Cai, Hao Bian, Jing Lin, Haoqian Wang, Radu Timofte, and Yulun Zhang. Retinexformer: One-stage retina-based transformer for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12504-12513, 2023. 1, 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 250, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 250, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 250, + 296, + 304 + ], + "type": "text", + "content": "[4] Kelvin C.K. Chan, Shangchen Zhou, Xiangyu Xu, and Chen Change Loy. Basicvsr++: Improving video superresolution with enhanced propagation and alignment. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 5972-5981, 2022. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 62, + 306, + 295, + 360 + ], + "type": "text", + "content": "[5] Hoonhee Cho, Yuhwan Jeong, Taewoo Kim, and Kuk-Jin Yoon. Non-coaxial event-guided motion deblurring with spatial alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), pages 12492-12503, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 363, + 295, + 417 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 363, + 295, + 417 + ], + "spans": [ + { + "bbox": [ + 62, + 363, + 295, + 417 + ], + "type": "text", + "content": "[6] Sung-Jin Cho, Seo-Won Ji, Jun-Pyo Hong, Seung-Won Jung, and Sung-Jea Ko. Rethinking coarse-to-fine approach in single image deblurring. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4641-4650, 2021. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 419, + 295, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 419, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 62, + 419, + 295, + 473 + ], + "type": "text", + "content": "[7] Jiangxin Dong, Jinshan Pan, Zhongbao Yang, and Jinhui Tang. Multi-scale residual low-pass filter network for image deblurring. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12345-12354, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 62, + 476, + 295, + 531 + ], + "type": "text", + "content": "[8] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE transactions on pattern analysis and machine intelligence, 44(1):154-180, 2020. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 533, + 296, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 533, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 62, + 533, + 296, + 599 + ], + "type": "text", + "content": "[9] Weihua He, Kaichao You, Zhendong Qiao, Xu Jia, Ziyang Zhang, Wenhui Wang, Huchuan Lu, Yaoyuan Wang, and Jianxing Liao. Timereplayer: Unlocking the potential of event cameras for video interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17804-17813, 2022. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 600, + 295, + 644 + ], + "type": "text", + "content": "[10] Zhe Jiang, Yu Zhang, Dongqing Zou, Jimmy Ren, Jiancheng Lv, and Yebin Liu. Learning event-based motion deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3320-3329, 2020. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 646, + 296, + 689 + ], + "type": "text", + "content": "[11] Meiguang Jin, Zhe Hu, and Paolo Favaro. Learning to extract flawless slow motion from blurry videos. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8112-8121, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 713 + ], + "type": "text", + "content": "[12] Taewoo Kim, Jeongmin Lee, Lin Wang, and Kuk-Jin Yoon. Event-guided deblurring of unknown exposure time videos." + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "In European Conference on Computer Vision, pages 519-538. Springer, 2022. 2, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[13] Taewoo Kim, Yujeong Chae, Hyun-Kurl Jang, and Kuk-Jin Yoon. Event-based video frame interpolation with cross-modal asymmetric bidirectional motion fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 18032-18042, 2023. 1, 2, 3, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 194 + ], + "type": "text", + "content": "[14] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Cmta: Cross-modal temporal alignment for event-guided video deblurring. arXiv preprint arXiv:2408.14930, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 248 + ], + "type": "text", + "content": "[15] Taewoo Kim, Hoonhee Cho, and Kuk-Jin Yoon. Frequency-aware event-based video deblurring for real-world motion blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 24966-24976, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 553, + 293 + ], + "type": "text", + "content": "[16] Taewoo Kim, Jaeseok Jeong, Hoonhee Cho, Yuhwan Jeong, and Kuk-Jin Yoon. Towards real-world event-guided low-light video enhancement and deblurring. arXiv preprint arXiv:2408.14916, 2024. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "spans": [ + { + "bbox": [ + 317, + 294, + 553, + 348 + ], + "type": "text", + "content": "[17] Lingshun Kong, Jiangxin Dong, Jianjun Ge, Mingqiang Li, and Jinshan Pan. Efficient frequency domain-based transformers for high-quality image deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5886-5895, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 350, + 553, + 404 + ], + "type": "text", + "content": "[18] Hyeongmin Lee, Taeoh Kim, Tae-young Chung, Daehyun Pak, Yuseok Ban, and Sangyoun Lee. Adacof: Adaptive collaboration of flows for video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5316-5325, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 469 + ], + "type": "text", + "content": "[19] Dasong Li, Xiaoyu Shi, Yi Zhang, Ka Chun Cheung, Simon See, Xiaogang Wang, Hongwei Qin, and Hongsheng Li. A simple baseline for video restoration with grouped spatial-temporal shift. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9822-9832, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 471, + 553, + 536 + ], + "type": "text", + "content": "[20] Guoqiang Liang, Kanghao Chen, Hangyu Li, Yunfan Lu, and Lin Wang. Towards robust event-guided low-light image enhancement: A large-scale real-world event-image dataset and novel approach. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23-33, 2024. 2, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 580 + ], + "type": "text", + "content": "[21] Jingyun Liang, Jiezhang Cao, Yuchen Fan, Kai Zhang, Rakesh Ranjan, Yawei Li, Radu Timofte, and Luc Van Gool. Vrt: A video restoration transformer. arXiv preprint arXiv:2201.12288, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 317, + 581, + 553, + 635 + ], + "type": "text", + "content": "[22] Jingyun Liang, Yuchen Fan, Xiaoyu Xiang, Rakesh Ranjan, Eddy Ilg, Simon Green, Jiezhang Cao, Kai Zhang, Radu Timofte, and Luc V Gool. Recurrent video restoration transformer with guided deformable attention. Advances in Neural Information Processing Systems, 35:378-393, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 689 + ], + "type": "text", + "content": "[23] Jinxiu Liang, Yixin Yang, Boyu Li, Peiqi Duan, Yong Xu, and Boxin Shi. Coherent event guided low-light video enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10615-10625, 2023. 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[24] Lin Liu, Junfeng An, Jianzhuang Liu, Shanxin Yuan, Xi-angyu Chen, Wengang Zhou, Houqiang Li, Yan Feng Wang," + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "8532" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 106 + ], + "type": "text", + "content": "and Qi Tian. Low-light video enhancement with synthetic event guidance. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1692-1700, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 108, + 295, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 108, + 295, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 108, + 295, + 173 + ], + "type": "text", + "content": "[25] Yunlong Liu, Tao Huang, Weisheng Dong, Fangfang Wu, Xin Li, and Guangming Shi. Low-light image enhancement with multi-stage residue quantization and brightness-aware attention. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12140-12149, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "spans": [ + { + "bbox": [ + 56, + 176, + 294, + 219 + ], + "type": "text", + "content": "[26] Huaijia Lin Jiangbo Lu Liying Lu, Ruizheng Wu and Jiaya Jia. Video frame interpolation with transformer. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "spans": [ + { + "bbox": [ + 56, + 222, + 294, + 266 + ], + "type": "text", + "content": "[27] Yongrui Ma, Shi Guo, Yutian Chen, Tianfan Xue, and Jinwei Gu. Timelens-xl: Real-time event-based video frame interpolation with large motion. In European Conference on Computer Vision, pages 178-194. Springer, 2025. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 56, + 268, + 294, + 312 + ], + "type": "text", + "content": "[28] Xintian Mao, Yiming Liu, Fengze Liu, Qingli Li, Wei Shen, and Yan Wang. Intriguing findings of frequency selection for image deblurring. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1905-1913, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 315, + 294, + 367 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 315, + 294, + 367 + ], + "spans": [ + { + "bbox": [ + 56, + 315, + 294, + 367 + ], + "type": "text", + "content": "[29] Seungjun Nah, Tae Hyun Kim, and Kyoung Mu Lee. Deep multi-scale convolutional neural network for dynamic scene deblurring. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3883-3891, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[30] Jihyong Oh and Munchurl Kim. Demfi: Deep joint deblurring and multi-frame interpolation with flow-guided attentive correlation and recursive boosting. In European Conference on Computer Vision (ECCV), 2022. 1, 3, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 56, + 417, + 294, + 471 + ], + "type": "text", + "content": "[31] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 474, + 294, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 474, + 294, + 529 + ], + "spans": [ + { + "bbox": [ + 56, + 474, + 294, + 529 + ], + "type": "text", + "content": "[32] Jinshan Pan, Boming Xu, Jiangxin Dong, Jianjun Ge, and Jinhui Tang. Deep discriminative spatial and temporal network for efficient video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22191-22200, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 56, + 531, + 294, + 586 + ], + "type": "text", + "content": "[33] Liyuan Pan, Cedric Scheerlinck, Xin Yu, Richard Hartley, Miaomiao Liu, and Yuchao Dai. Bringing a blurry frame alive at high frame-rate with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6820-6829, 2019. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 588, + 294, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 588, + 294, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 588, + 294, + 632 + ], + "type": "text", + "content": "[34] Liyuan Pan, Richard Hartley, Cedric Scheerlinck, Miaomiao Liu, Xin Yu, and Yuchao Dai. High frame rate video reconstruction based on an event camera. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2020. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 634, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 634, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 634, + 294, + 677 + ], + "type": "text", + "content": "[35] Junheum Park, Keunsoo Ko, Chul Lee, and Chang-Su Kim. Bmbc: Bilateral motion estimation with bilateral cost volume for video interpolation. In European Conference on Computer Vision, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[36] Junheum Park, Chul Lee, and Chang-Su Kim. Asymmetric bilateral motion estimation for video frame interpolation. In International Conference on Computer Vision, 2021. 1, 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 25, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 137 + ], + "type": "text", + "content": "[37] Jaesung Rim, Haeyun Lee, Jucheol Won, and Sunghyun Cho. Real-world blur dataset for learning and benchmarking deblurring algorithms. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part XXV 16, pages 184-201. Springer, 2020. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 182 + ], + "type": "text", + "content": "[38] Hyeonjun Sim, Jihyong Oh, and Munchurl Kim. Xvfi: extreme video frame interpolation. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2021. 1, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "spans": [ + { + "bbox": [ + 316, + 184, + 553, + 238 + ], + "type": "text", + "content": "[39] Shuochen Su, Maurizio Delbracio, Jue Wang, Guillermo Sapiro, Wolfgang Heidrich, and Oliver Wang. Deep video deblurring for hand-held cameras. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1279-1288, 2017. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 293 + ], + "type": "text", + "content": "[40] Lei Sun, Christos Sakaridis, Jingyun Liang, Qi Jiang, Kailun Yang, Peng Sun, Yaozu Ye, Kaiwei Wang, and Luc Van Gool. Event-based fusion for motion deblurring with cross-modal attention. In European conference on computer vision, pages 412-428. Springer, 2022. 2, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 294, + 553, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 346 + ], + "type": "text", + "content": "[41] Lei Sun, Christos Sakaridis, Jingyun Liang, Peng Sun, Jiezhang Cao, Kai Zhang, Qi Jiang, Kaiwei Wang, and Luc Van Gool. Event-based frame interpolation with ad-hoc deblurring. arXiv preprint arXiv:2301.05191, 2023. 2, 3, 4, 7, 8" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 392 + ], + "type": "text", + "content": "[42] Fu-Jen Tsai, Yan-Tsung Peng, Yen-Yu Lin, Chung-Chi Tsai, and Chia-Wen Lin. Stripformer: Strip transformer for fast image deblurring. In European conference on computer vision, pages 146-162. Springer, 2022. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 394, + 553, + 458 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 394, + 553, + 458 + ], + "spans": [ + { + "bbox": [ + 316, + 394, + 553, + 458 + ], + "type": "text", + "content": "[43] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16155-16164, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 525 + ], + "type": "text", + "content": "[44] Stepan Tulyakov, Daniel Gehrig, Stamatios Georgoulis, Julius Erbach, Mathias Gehrig, Yuanyou Li, and Davide Scaramuzza. Time lens: Event-based video frame interpolation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16155-16164, 2021. 1, 2, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 526, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 526, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 316, + 526, + 553, + 590 + ], + "type": "text", + "content": "[45] Stepan Tulyakov, Alfredo Bochicchio, Daniel Gehrig, Stamatios Georgoulis, Yuanyou Li, and Davide Scaramuzza. Time Lens++: Event-based frame interpolation with nonlinear parametric flow and multi-scale fusion. IEEE Conference on Computer Vision and Pattern Recognition, 2022. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 646 + ], + "type": "text", + "content": "[46] Xintao Wang, Kelvin C.K. Chan, Ke Yu, Chao Dong, and Chen Change Loy. Edvr: Video restoration with enhanced deformable convolutional networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2019. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 712 + ], + "type": "text", + "content": "[47] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 13128-13137, 2023. 3" + } + ] + } + ], + "index": 24 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8533" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[48] Yufei Wang, Yi Yu, Wenhan Yang, Lanqing Guo, Lap-Pui Chau, Alex C Kot, and Bihan Wen. Exposurediffusion: Learning to expose for low-light image enhancement. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12438-12448, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 184 + ], + "type": "text", + "content": "[49] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 17683-17693, 2022. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 185, + 294, + 229 + ], + "type": "text", + "content": "[50] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based blurry frame interpolation under blind exposure. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1588-1598, 2023. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 285 + ], + "type": "text", + "content": "[51] Song Wu, Kaichao You, Weihua He, Chen Yang, Yang Tian, Yaoyuan Wang, Ziyang Zhang, and Jianxing Liao. Video interpolation by event-driven anisotropic adjustment of optical flow. In Proceedings of the European Conference on Computer Vision (ECCV), 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 342 + ], + "type": "text", + "content": "[52] Wenhui Wu, Jian Weng, Pingping Zhang, Xu Wang, Henhan Yang, and Jianmin Jiang. Uretinex-net: Retinex-based deep unfolding network for low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901-5910, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 294, + 387 + ], + "type": "text", + "content": "[53] Xiaogang Xu, Ruixing Wang, Chi-Wing Fu, and Jiaya Jia. Snr-aware low-light image enhancement. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 17714-17724, 2022. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 441 + ], + "type": "text", + "content": "[54] Xiaogang Xu, Ruixing Wang, and Jiangbo Lu. Low-light image enhancement via structure modeling and guidance. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9893-9903, 2023. 1, 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 498 + ], + "type": "text", + "content": "[55] Shuzhou Yang, Moxuan Ding, Yanmin Wu, Zihan Li, and Jian Zhang. Implicit neural representation for cooperative low-light image enhancement. In Proceedings of the IEEE/CVF international conference on computer vision, pages 12918-12927, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 544 + ], + "type": "text", + "content": "[56] Wen Yang, Jinjian Wu, Jupo Ma, Leida Li, and Guangming Shi. Motion deblurring via spatial-temporal collaboration of frames and events. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6531-6539, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 600 + ], + "type": "text", + "content": "[57] Xunpeng Yi, Han Xu, Hao Zhang, Linfeng Tang, and Jiayi Ma. Diff-retinex: Rethinking low-light image enhancement with a generative diffusion model. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12302-12311, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 56, + 601, + 294, + 644 + ], + "type": "text", + "content": "[58] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, Ming-Hsuan Yang, and Ling Shao. Multi-stage progressive image restoration. In CVPR, 2021. 2, 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[59] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[60] Hongguang Zhang, Yuchao Dai, Hongdong Li, and Piotr Koniusz. Deep stacked hierarchical multi-patch network for im" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 386 + ], + "type": "list", + "angle": 0, + "index": 21, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "age deblurring. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2019. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 140 + ], + "type": "text", + "content": "[61] Huicong Zhang, Haozhe Xie, and Hongxun Yao. Blur-aware spatio-temporal sparse transformer for video deblurring. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2673-2681, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 196 + ], + "type": "text", + "content": "[62] Zhihang Zhong, Ye Gao, Yinqiang Zheng, and Bo Zheng. Efficient spatio-temporal recurrent neural network for video deblurring. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part VI 16, pages 191-207. Springer, 2020. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 251 + ], + "type": "text", + "content": "[63] Zhihang Zhong, Mingdeng Cao, Xiang Ji, Yinqiang Zheng, and Imari Sato. Blur interpolation transformer for real-world motion from blur. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5713-5723, 2023. 1, 3, 4, 7, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[64] Jingkai Zhou, Varun Jampani, Zhixiong Pi, Qiong Liu, and Ming-Hsuan Yang. Decoupled dynamic filter networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6647-6656, 2021. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 329 + ], + "type": "text", + "content": "[65] Shangchen Zhou, Chongyi Li, and Chen Change Loy. Lednet: Joint low-light enhancement and deblurring in the dark. In ECCV, 2022. 1, 3, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 331, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 386 + ], + "type": "text", + "content": "[66] Alex Zihao Zhu, Liangzhe Yuan, Kenneth Chaney, and Kostas Daniilidis. Unsupervised event-based learning of optical flow, depth, and egomotion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 989-997, 2019. 5" + } + ] + } + ], + "index": 20 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "8534" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_content_list.json b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..2a19a004a00ed368cd3e0fa6de5eb47c989c45bb --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_content_list.json @@ -0,0 +1,1641 @@ +[ + { + "type": "text", + "text": "EventUPS: Uncalibrated Photometric Stereo Using an Event Camera", + "text_level": 1, + "bbox": [ + 150, + 130, + 846, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jinxiu Liang $^{1,2,3\\#}$ Bohan Yu $^{1,2\\#}$ Siqi Yang $^{1,2,4}$ Haotian Zhuang $^{5}$ Jieji Ren $^{6}$ Peiqi Duan $^{1,2}$ Boxin Shi $^{1,2*}$", + "bbox": [ + 187, + 178, + 805, + 215 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "$^{1}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University", + "$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + ], + "bbox": [ + 169, + 217, + 826, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ National Institute of Informatics $^{4}$ Institute for Artificial Intelligence, Peking University", + "bbox": [ + 228, + 252, + 766, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "5 Tsinghua University 6 Shanghai Jiaotong University", + "bbox": [ + 333, + 270, + 663, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cssherryliang@gmail.com,{ybhl998,yousiki,duanqi0001,shiboxin}@pku.edu.cn,", + "bbox": [ + 176, + 289, + 816, + 304 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "zht23@mails.tsinghua.edu.cn, jiejiren@sjtu.edu.cn", + "bbox": [ + 281, + 306, + 705, + 321 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 356, + 326, + 372 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We present EventUPS, the first uncalibrated photometric stereo (UPS) method using an event camera—a neuromorphic sensor that asynchronously detects brightness changes with microsecond resolution. Traditional frame-based UPS methods are hindered by high bandwidth demands and limited use in dynamic scenes. These methods require dense image correspondence under varying illumination and are incompatible with the fundamentally different sensing paradigm of event data. Our approach introduces three key innovations: an augmented null space formulation that directly relates each event to joint constraints on surface normals and lighting, naturally handling ambient illumination; a continuous parameterization of time-varying illumination that connects asynchronous events to synchronized lighting estimation; and a lighting fixture with known relative geometry that reduces ambiguity to a convex-concave uncertainty. We validate EventUPS using a custom-built LED lighting system. Experimental results show that our method achieves accuracy surpassing its frame-based counterpart while requiring only $5\\%$ of the data bandwidth.", + "bbox": [ + 89, + 388, + 483, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 717, + 220, + 732 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Photometric stereo (PS) recovers shape from images under varying illumination [14, 42]. For decades, it has been a cornerstone for high-fidelity 3D reconstruction, enabling detailed surface normal recovery for applications from cultural heritage to industrial inspection. The uncalibrated variant of PS (UPS)—where lighting directions are unknown—offers particular practical appeal by eliminating the need for precise calibration apparatus, facilitating deployment in uncontrolled environments. However, this flex", + "bbox": [ + 89, + 742, + 482, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8c8e1aa3ad56055e81c2442bed480b9badb7735ed929e17df868a34836401bf3.jpg", + "image_caption": [ + "(a) Our Setup" + ], + "image_footnote": [], + "bbox": [ + 521, + 358, + 707, + 523 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/8665700ed919e6d969f51d133ae577ead7a336e4a1905dfb03347e5d8b1da8b2.jpg", + "image_caption": [ + "(c) Our Results" + ], + "image_footnote": [], + "bbox": [ + 718, + 358, + 893, + 438 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d0e6dcd1faceb1911e8d55be8ad528db4faa8cf1057a411a1041f4d7e128fddd.jpg", + "image_caption": [ + "(b) Captured events", + "Figure 1. We propose an uncalibrated photometric stereo method designed for event cameras, EventUPS. Our approach employs (a) a portable lighting system with known relative geometry (e.g., dual LED rings) but an unknown global pose, removing the need for precise calibration. From (b) sparse, asynchronous events triggered by sequentially activated lights, it jointly recovers (c) dense surface normals and the time-varying lighting trajectory1." + ], + "image_footnote": [], + "bbox": [ + 720, + 453, + 893, + 530 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ibility incurs significant costs: conventional uncalibrated PS methods demand dense correspondences across multiple images under varying illumination, imposing high data bandwidth and rendering them ill-suited for dynamic scenes or resource-constrained systems.", + "bbox": [ + 511, + 681, + 906, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meanwhile, event cameras represent a paradigm shift in visual sensing [10]. Instead of capturing intensity frames at fixed intervals, these bio-inspired sensors asynchronously detect per-pixel logarithmic brightness changes. This provides microsecond temporal resolution, an exceptional dynamic range (140dB vs. 60dB), and drastically reduced data", + "bbox": [ + 511, + 758, + 908, + 849 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Equal contributions. *Corresponding author.", + "bbox": [ + 102, + 886, + 354, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ Est. Lights: The estimated time-varying light trajectory $(x, y$ components on the unit sphere), with numbers indicating temporal sequence, colors denoting relative intensity (yellow higher, green lower).", + "bbox": [ + 511, + 862, + 906, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7516", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "rates—all critical for platforms like mobile robots. These unique attributes have transformed high-speed tasks such as SLAM [39] and object tracking [7].", + "bbox": [ + 89, + 90, + 480, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in event-based PS—exemplified by the EventPS [45]—have shown promise by leveraging these unique capabilities. By using a calibrated, continuously rotating light source, it computes surface normals directly from brightness change events, naturally accommodating rapid variations in surface orientation even in dynamic scenes. This approach dramatically reduces both latency and data redundancy compared to frame-based methods. However, EventPS [45] requires precise lighting calibration and darkroom conditions, which are challenging outside labs. In many real-world scenarios—such as robotic deployments, motion capture, or industrial settings—maintaining calibrated lighting and negligible ambient illumination becomes impractical due to environmental constraints and operational demands.", + "bbox": [ + 89, + 138, + 482, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A natural question arises: Can established frame-based UPS approaches adapt directly to event cameras? Our analysis reveals fundamental incompatibilities: i) Inability to handle ambient illumination. Frame-based PS methods often address ambient illumination by capturing a dark frame without designed lighting. Unfortunately, this strategy is not applicable to events as they are generated asynchronously by logarithmic brightness changes—encoding brightness ratios rather than absolute intensities. ii) Absence of synchronous correspondence. Non-learning frame-based UPS [2, 46] typically performs matrix factorization to decompose observed intensities into surface normals and lighting directions, establishing constraints through pixel correspondences under shared illumination conditions. However, when applied to events—which correspond to different lighting directions and occur asynchronously—the establishment of such correspondences becomes challenging. iii) Inapplicability of conventional ambiguity resolution. Existing UPS resolves the remaining ambiguity by leveraging effective priors on albedo characteristics [28], which are not applicable to events since the event generation equation is inherently invariant to albedo.", + "bbox": [ + 89, + 369, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we present EventUPS—the first uncalibrated photometric stereo method using an event camera. Our approach rethinks PS for the event domain, leveraging the unique properties of event cameras to recover detailed surface geometry while simultaneously estimating the time-varying illumination, all without explicit lighting calibration. The core of our contribution is a theoretical framework that formulates each event as a constraint in the null space of an augmented normal vector, enabling joint recovery of surface normals and lighting directions despite the asynchronous nature of events and the presence of ambient illumination. To address the incompatibilities, we make the following technical contributions:", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- An augmented null-space formulation able to handle ambient illumination by modeling it as an unknown variable within event-based constraints on normals and lighting.", + "- A continuous parameterization of the time-varying illumination that establishes correspondence by linking asynchronous events to a shared lighting trajectory.", + "- A geometric UPS ambiguity resolution technique applicable to event data, which uses a lighting fixture with known relative geometry but an unknown global pose." + ], + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Along with the proposed EventUPS framework, we develop a joint optimization framework that effectively integrates integrability constraints with our event-based null space formulation. To validate our approach in real scenes, we build a hardware prototype around an LED-based lighting system, which implements the lighting configurations with known geometry of dual ring (see Figure 1) and trefoil curve patterns. Experiments on semi-real and real data demonstrate that our method achieves accuracy outperforming the leading frame-based non-learning UPS [28] while requiring only $5\\%$ of the data bandwidth.", + "bbox": [ + 511, + 227, + 906, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 513, + 406, + 663, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PS and event-based vision have advanced independently for decades. Our work lies at their intersection, specifically addressing the challenges of UPS with event cameras. Here, we review developments in both domains that inform ours.", + "bbox": [ + 511, + 431, + 905, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Uncalibrated Photometric Stereo", + "text_level": 1, + "bbox": [ + 511, + 501, + 803, + 516 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "PS estimates surface normals from images captured under varying illumination directions [15, 42]. While classical approaches assume known lighting, the uncalibrated setting—where lighting is unknown—presents greater challenges but offers superior practical advantages. For comprehensive reviews, see [30, 33, 34, 41].", + "bbox": [ + 511, + 523, + 905, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "A fundamental challenge in UPS is the ambiguity in simultaneously recovering lighting and surface normals. This was first identified as a general $3 \\times 3$ linear ambiguity [14], later reduced to the 3-parameter generalized bas-relief (GBR) ambiguity by imposing integrability constraints [2, 46]. Resolving the ambiguity typically relies on physical cues such as specular reflections [8, 9, 11, 38, 43], statistical priors on albedo [1, 21, 28, 32], shadows [36], interreflections [3], or perspective effects [27]. The light source configuration is also critical; strategic arrangements like ring lights with multi-view information can provide additional constraints to resolve ambiguity without strong surface assumptions [48]. Recent learning-based methods have shown promising results [4, 5, 20], but they often require extensive training data and can struggle with generalization.", + "bbox": [ + 511, + 614, + 905, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Accelerating PS acquisition remains an active research area, with solutions like high-speed cameras [18, 22, 40] or multi-tap CMOS [44] with synchronized illumination, though these dramatically increase bandwidth. Single-shot", + "bbox": [ + 511, + 840, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7517", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "multi-spectral systems [19, 26, 35, 37] offer an alternative but face challenges like limited spectral bands, crosstalk, and intensity inconsistencies across bands introduce additional challenges [12, 16, 17].", + "bbox": [ + 89, + 90, + 483, + 151 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Photometric Methods with Event Cameras", + "text_level": 1, + "bbox": [ + 89, + 157, + 455, + 174 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With their high temporal resolution, wide dynamic range, and asynchronous operation, event cameras are powerful tools for photometric measurements. These advantages make them particularly suitable for capturing dynamic lighting effects crucial for photometric applications. Their microsecond resolution is especially valuable for active lighting techniques. MC3D [23] pioneered event-based structured light for high-speed 3D scanning, while ESL [24] improved noise robustness by exploiting spatial correlations. Researchers have explored diverse illumination patterns for specific tasks, such as capturing events during light activation to reconstruct iso-contours [13] and address intensity-distance ambiguity by using a diffuse sphere [6], estimating surface normals using rotating polarizers [25], and separating direct/global illumination components [47].", + "bbox": [ + 89, + 180, + 483, + 407 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Most relevant to our work, EFPS-Net [31] fused sparse events with RGB images for normal estimation, while EventPS [45] pioneered purely event-based PS. EventPS showed that the differential nature of event cameras aligns perfectly with PS requirements, but it assumes calibrated lighting and cannot handle ambient illumination. The sparse, asynchronous nature of event data presents unique challenges in uncalibrated settings, particularly in establishing pixel correspondences under consistent lighting. This fundamentally different sensing paradigm makes direct application of non-learning UPS methods [14, 46] infeasible. Our work, EventUPS, is the first to solve these challenges.", + "bbox": [ + 89, + 407, + 483, + 590 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 599, + 181, + 614 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 625, + 294, + 638 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image formulation. We consider a scene captured by an orthographic projection camera2. The surface is assumed to be Lambertian and is illuminated by a combination of a distant, time-varying directional light source and constant ambient illumination. For a pixel $\\pmb{x}$ in the image plane $\\Omega \\subset \\mathbb{R}^2$ , the reflected radiance $I_{\\pmb{x}}(t)$ at time $t$ is given by:", + "bbox": [ + 89, + 646, + 483, + 737 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\boldsymbol {x}} (t) = \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t) \\right] + b _ {\\boldsymbol {x}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 743, + 482, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pmb{n}_{\\pmb{x}} \\in \\mathbb{R}^{3}$ is the unit surface normal ( $\\| \\pmb{n}_{\\pmb{x}} \\|_{2} = 1$ ), $a_{x} > 0$ is the diffuse albedo, and $b_{x} \\geq 0$ is the ambient illumination term. The lighting is described by the vector $\\pmb{s}(t) \\in \\mathbb{R}^{3}$ with intensity $\\| \\pmb{s}(t) \\|^{2}$ and unit direction $\\hat{\\pmb{s}}(t) = \\pmb{s}(t) / \\| \\pmb{s}(t) \\|$ . The max operator models attached shadows, where the dot product becomes negative.", + "bbox": [ + 89, + 765, + 483, + 856 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/078f49e4491f40c7ab347cc4ce11d522b3f79b9a17527a8320725f73e0ea33d8.jpg", + "image_caption": [ + "Figure 2. The null space vectors and estimated normals: (a) Lambertian sphere under ideal conditions, (b) non-Lambertian sphere with real events. EventUPS models ambient and albedo effects via a residual term, enhancing robustness." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 901, + 239 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event generation. Event cameras are bio-inspired sensors that asynchronously report logarithmic changes in brightness. An event $(\\pmb{x}, t_{\\pmb{x}}^{(k)}, p_{\\pmb{x}}^{(k)})$ is triggered at pixel $\\pmb{x}$ when the logarithmic change in brightness exceeds a predetermined contrast threshold $\\theta > 0$ :", + "bbox": [ + 511, + 329, + 906, + 405 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right)\\right) - \\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right)\\right) \\geq p _ {\\boldsymbol {x}} ^ {(k)} \\theta , \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 412, + 906, + 431 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $t_{\\pmb{x}}^{(k)}$ denotes the timestamp of the $k$ -th event at pixel $\\pmb{x}$ , and $p_{\\pmb{x}}^{(k)} \\in \\{+1, -1\\}$ indicates whether the brightness increased or decreased. The relationship between consecutive brightness measurements can be approximated as:", + "bbox": [ + 511, + 440, + 905, + 505 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\approx I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right) \\cdot \\exp \\left(p _ {\\boldsymbol {x}} ^ {(k)} \\theta\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 586, + 511, + 905, + 531 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For notational simplicity, we define $\\tilde{\\theta}_{\\pmb{x}}^{(k)} = \\exp (p_{\\pmb{x}}^{(k)}\\theta)$", + "bbox": [ + 511, + 539, + 877, + 558 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "UPS with events. Unlike prior event-based works [45] that assume calibrated lighting, we aim to jointly estimate the per-pixel surface normals $\\pmb{n}_{\\pmb{x}}$ and the continuous, time-varying lighting function $s(t)$ from only the event stream $\\mathbb{E} = \\{(x,t_x^{(k)},p_x^{(k)})\\}_{x\\in \\Omega}$ across all pixels $x\\in \\Omega$ . This inverse problem is formulated as the minimization of the joint loss function that measures consistency between the observed events and the image formation model:", + "bbox": [ + 511, + 573, + 906, + 699 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {a r g m i n} _ {\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\} _ {\\boldsymbol {x} \\in \\Omega}, \\boldsymbol {s} (t)} \\mathcal {L} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {s} (t); \\mathbb {E}\\right). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 705, + 906, + 724 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Estimating Normal and Lighting", + "text_level": 1, + "bbox": [ + 511, + 729, + 844, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The augmented null-space formulation. Conventional photometric stereo often handles ambient light by subtracting a dark frame. This is infeasible for event-based methods where measurements are relative and asynchronous. We propose an alternative by substituting the image model Eq. (1) into the event model Eq. (3):", + "bbox": [ + 511, + 752, + 906, + 844 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k)}) \\right] + b _ {\\boldsymbol {x}} \\\\ \\approx \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} \\cdot (\\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k - 1)}) \\right] + b _ {\\boldsymbol {x}}). \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 563, + 849, + 854, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_footnote", + "text": "$^{2}$ Orthographic projection is a widely used assumption in PS [33] when the object-to-camera distance (approximately $20~\\mathrm{cm}$ in our setup) is significantly larger than object depth variations (approximately $3\\mathrm{cm}$ ).", + "bbox": [ + 89, + 862, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7518", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For most observed events, the surface point is illuminated at both timestamps (i.e., it is not in an attached shadow). This allows us to remove the max operator. The albedo $a_{\\pmb{x}}$ can be eliminated by dividing it out, as it appears on both sides. This makes our formulation inherently albedo-invariant. Rearranging the terms yields a linear constraint on the normal $n_{\\pmb{x}}$ and the \"residual\" pixel-wise constant ambient-albedo ratio $r_{\\pmb{x}} = b_{\\pmb{x}} / a_{\\pmb{x}} \\geq 0$ :", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pmb {n} _ {\\pmb {x}} \\cdot (\\pmb {s} (t _ {\\pmb {x}} ^ {(k)}) - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)} \\pmb {s} (t _ {\\pmb {x}} ^ {(k - 1)})) + r _ {\\pmb {x}} (1 - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)}) = 0. \\quad (6)\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 223, + 482, + 244 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Shadow events, which occur when one of the terms under the max operator becomes zero, are relatively rare and can be identified through temporal consistency checks. When they do occur, they provide constraints at shadow boundaries that can be incorporated into our framework.", + "bbox": [ + 89, + 256, + 483, + 330 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $\\boldsymbol{z}_{\\boldsymbol{x}}^{(k)} = \\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k)}) - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)}\\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k-1)})$ . We define an augmented normal vector $\\tilde{\\boldsymbol{n}}_{\\boldsymbol{x}} = [\\boldsymbol{n}_{\\boldsymbol{x}}^{\\top}, r_{\\boldsymbol{x}}]^{\\top} \\in \\mathbb{R}^{4}$ , which is constant for each pixel $\\boldsymbol{x}$ . For each event, we define a corresponding augmented null space vector $\\tilde{\\boldsymbol{z}}_{\\boldsymbol{x}}^{(k)} = [\\boldsymbol{z}_{\\boldsymbol{x}}^{(k)\\top}, (1 - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)})]^{\\top}$ . This recasts Eq. (6) into an geometric relationship:", + "bbox": [ + 89, + 330, + 483, + 415 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {z} _ {\\boldsymbol {x}} ^ {(k) \\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} = 0. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 430, + 482, + 450 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This means the augmented normal $\\tilde{n}_{\\pmb{x}}$ must lie in the null space of the vectors $\\{\\tilde{z}_x^{(k)}\\}$ . Please see Fig. 2 for an illustration. With at least three linearly independent null space vectors from events at a pixel, we can determine the augmented normal $\\tilde{n}_{\\pmb{x}}$ up to a scale factor. By stacking all augmented null space vectors $\\tilde{z}_x^{(k)}$ for a pixel $\\pmb{x}$ into a matrix $Z_{\\pmb{x}} \\in \\mathbb{R}^{4 \\times K_{\\pmb{x}}}$ where $K_{\\pmb{x}}$ is the number of events at that pixel and the $k$ -th row is $\\tilde{z}_x^{(k)\\top}$ , we solve for the normal by finding the vector that minimizes the projection:", + "bbox": [ + 89, + 458, + 483, + 599 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} ^ {*} = \\underset {\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}}} {\\operatorname {a r g m i n}} \\| Z _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2} \\quad \\text {s u b j e c t t o} \\quad \\| \\boldsymbol {n} _ {\\boldsymbol {x}} \\| _ {2} = 1. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 609, + 482, + 638 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This constrained optimization problem has an elegant solution via Singular Value Decomposition (SVD) of the matrix $Z_{x}Z_{x}^{\\top}\\in \\mathbb{R}^{4\\times 4}$ . The solution is the eigenvector corresponding to the smallest eigenvalue, appropriately scaled to ensure the unit normal constraint.", + "bbox": [ + 89, + 650, + 483, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Continuous lighting parameterization. In the uncalibrated setting, the lighting direction $s(t)$ is unknown and varies continuously over time. The asynchronous nature of events makes it challenging to establish correspondences across pixels under identical lighting conditions, which is a prerequisite for non-learning uncalibrated PS methods. To address this challenge, we model $s(t)$ as a B-spline curve. This provides a smooth, local, and efficient parameterization that is ideal for asynchronous data. By selecting $M$ control points $\\{s(t_j)\\}_{j=0}^{M-1}$ at uniform times $t_j = t_0 + j\\Delta t$ ,", + "bbox": [ + 89, + 750, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/bd0cbfd15a6849f221219272691ded0ed08711f9cc03010a9d5a6f5c9ec9ff91.jpg", + "image_caption": [ + "Figure 3. Continuous illumination model. We represent the time-varying lighting $s(t)$ as a B-spline controlled by a set of points. This continuous representation links asynchronous events to a globally synchronized lighting estimate, shown here for a rotating ring pattern projected onto a 2D plane." + ], + "image_footnote": [], + "bbox": [ + 517, + 89, + 901, + 311 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the lighting at any time $t$ is a linear combination:", + "bbox": [ + 511, + 414, + 836, + 429 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} (t) = \\sum_ {j = 0} ^ {M - 1} C _ {j, l} (t) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 435, + 905, + 478 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\{C_{j,l}(t)\\}$ are the B-spline basis functions of order $l$ . An important property of this formulation is that the value of $s(t)$ for $t_j \\leq t < t_{j+1}$ depends only on $l$ consecutive control points $s(t_j), s(t_{j+1}), \\ldots, s(t_{j+l-1})$ , providing locality in the temporal domain. This links the null space vector $z_x^{(k)}$ directly to the control points:", + "bbox": [ + 511, + 484, + 905, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {z} _ {\\boldsymbol {x}} ^ {(k)} = \\sum_ {j = 0} ^ {M - 1} \\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 614, + 583, + 905, + 626 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where", + "bbox": [ + 513, + 632, + 558, + 643 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) = C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) - \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 650, + 903, + 671 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "This expression depends only on the observed event times-stamps $t_{\\pmb{x}}^{(k)}$ at pixel $\\pmb{x}$ , providing a crucial bridge between pixel-specific asynchronous events and a global, continuous lighting model (see Fig. 3). In matrix form, we stack the control points $\\{s(t_j)\\}_{j = 0}^{M - 1}$ into a matrix $\\pmb {S}\\in \\mathbb{R}^{3\\times M}$ and the time-dependent coefficients $\\widetilde{C}_{j,l}(t_x^{(k)})$ into a matrix $C_x\\in \\mathbb{R}^{M\\times K_x}$ . The augmented measurement matrix for a pixel $\\pmb{x}$ with $\\tilde{\\theta}_{\\pmb{x}} = [(1 - \\tilde{\\theta}_{\\pmb{x}}^{(1)}),\\dots ,(1 - \\tilde{\\theta}_{\\pmb{x}}^{(K_{\\pmb{x}})})]\\in \\mathbb{R}^{K_{\\pmb{x}}}$ becomes:", + "bbox": [ + 511, + 676, + 905, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} = \\left[ \\begin{array}{c} \\boldsymbol {S C} _ {\\boldsymbol {x}} \\\\ \\tilde {\\boldsymbol {\\theta}} _ {\\boldsymbol {x}} ^ {\\top} \\end{array} \\right] \\in \\mathbb {R} ^ {4 \\times K _ {\\boldsymbol {x}}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 622, + 816, + 903, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The data term becomes:", + "bbox": [ + 511, + 852, + 674, + 864 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) = \\sum_ {\\boldsymbol {x} \\in \\Omega} \\| \\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 871, + 903, + 904 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7519", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/325400be889e930192c3e7bdbe447f281b9a19e5cb733bc62cdce102cdef7083.jpg", + "image_caption": [ + "Figure 4. Resolving Ambiguity. With ground-truth lighting (a), the normal is recovered accurately. Assuming incorrect lighting transformations (b-e), estimating normals without constraints leads to large errors. The strong correlation between normal error and integrability cost validates the effectiveness of enforcing integrability with a known lighting fixture to resolve ambiguity." + ], + "image_footnote": [], + "bbox": [ + 99, + 88, + 477, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This bilinear structure naturally lends itself to an alternating optimization scheme.", + "bbox": [ + 89, + 483, + 483, + 513 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Resolving Ambiguity", + "text_level": 1, + "bbox": [ + 89, + 523, + 290, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Integrability. UPS suffers from a fundamental ambiguity where multiple combinations of surface normals and lighting directions can produce identical observations. Physical constraints, particularly the integrability of surface normals, require that the normals be derivable from the gradient of an underlying surface. For non-learning frame-based UPS, this constraint reduces the general linear ambiguity to the GBR transformation [2, 46], which maps the surface $z(x)$ to $\\lambda z(x) + \\mu x + \\nu y$ , where $\\lambda \\neq 0$ . For event-based UPS, imposing integrability similarly reduces the ambiguity to the GBR transformation, as the constraint acts on the normal field independently of the observation modality. Specifically, we use an integrability term. It ensures the normal field $\\{n_x\\}$ corresponds to a valid, continuous surface by penalizing the curl of the corresponding gradient field:", + "bbox": [ + 89, + 546, + 483, + 773 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {i n t}} = \\sum_ {\\boldsymbol {x}} \\left\\| \\frac {\\partial}{\\partial y} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {x}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) - \\frac {\\partial}{\\partial x} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {y}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) \\right\\| ^ {2}. \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 784, + 483, + 823 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Constraints from lighting fixture configurations. To further resolve this ambiguity, we leverage a portable lighting system with a known 3D structure undergoing unknown 6-DoF transformations. The system consists of light", + "bbox": [ + 89, + 839, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/20c7063fc4f21b3b9552161a5a18e45d8564564300f2cade2a00bdad1604bbc2.jpg", + "image_caption": [ + "Figure 5. The lighting configurations employed in our experiments. We built physical prototypes for the \"dual ring\" and the \"trefoil curve\" patterns using programmable LEDs, shown both statically (left) and as dynamic trajectories captured with long exposure (right). We also test on the standard \"DiLiGenT\" pattern." + ], + "image_footnote": [], + "bbox": [ + 519, + 88, + 668, + 263 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/51cbbcc237ea51eed65a0686540afb07cd3b022da0264c56696493a1cd669a66.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 674, + 93, + 782, + 265 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5051bb9b9bce01fad6359a87acdbe00756aa6aaf991b439d2f4d502deb4a624b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 789, + 93, + 897, + 265 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "sources at known positions $\\pmb{p}_i$ (for $i = 1,2,\\dots,N$ ) in a local reference frame. Assuming the camera is at the origin, the world positions become $\\pmb{p}_i' = \\pmb{R}\\pmb{p}_i + \\pmb{t}$ under a rigid transformation with orthogonal matrix $\\pmb{R} \\in O(3)$ and translation vector $\\pmb{t} \\in \\mathbb{R}^3$ . The light vector for the $i$ -th source is $s_i = u_i p_i'$ with unknown intensity $u_i > 0$ . A GBR transformation matrix $\\pmb{G}$ acts on the recovered normals, and an inverse transformation $\\pmb{G}^{-\\top}$ acts on the light vectors. The observed lighting $s_i'$ is related to the true lighting $s_i$ by $s_i' = \\pmb{G}^{-\\top}s_i$ . Combining these, we obtain a relationship between the known local positions $\\pmb{p}_i$ and the estimated, transformed light vectors $s_i'$ :", + "bbox": [ + 511, + 375, + 906, + 556 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {s} _ {i} ^ {\\prime} = u _ {i} \\boldsymbol {G} ^ {- \\top} (\\boldsymbol {R p} _ {i} + \\boldsymbol {t}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 565, + 906, + 584 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "This system of equations constrains the possible transformations. As formally proven in the supplement, these constraints reduce the GBR ambiguity to a discrete set. Specifically, for a set of at least seven light positions in general position (i.e., not all lying on any nontrivial quadric surface of a specified form passing through the origin), under physical constraints, the ambiguity can be resolved up to the classic convex/concave reflection. This eliminates the need for albedo or reflectance priors. An illustration of integrability's role in resolving positional ambiguity in our setup is shown in Fig. 4.", + "bbox": [ + 511, + 595, + 906, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Lighting configurations. For event cameras that asynchronously detect brightness changes with high temporal resolution, we sequentially activate each light (e.g., on millisecond timescales) to generate an event stream, with timestamps correlating to lighting changes. We designed two physical patterns (see Fig. 5) that satisfy the geometric requirements: a coaxial dual-ring consisting of LEDs arranged on two concentric rings of different radii, separated", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7520", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "by a known axial distance; and a trefoil curve with a moving ring of lights along a non-planar, asymmetric 3D trefoil knot trajectory. Both configurations meet the conditions for ambiguity resolution. Details of the theorem, proof, and lighting configurations are provided in the supplement.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Optimization", + "text_level": 1, + "bbox": [ + 89, + 178, + 230, + 194 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We solve for normals and lighting via alternating optimization, minimizing a joint loss function:", + "bbox": [ + 89, + 200, + 483, + 231 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\text {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) + \\lambda_ {\\text {i n t}} \\mathcal {L} _ {\\text {i n t}} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}\\right), \\tag {16}\n$$\n", + "text_format": "latex", + "bbox": [ + 156, + 244, + 482, + 262 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where the weight $\\lambda_{\\mathrm{int}}$ balances data fidelity Eq. (13) against surface smoothness Eq. (14). The optimization proceeds by alternating between two main steps until convergence:", + "bbox": [ + 89, + 275, + 483, + 320 + ], + "page_idx": 5 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Normal update: With the lighting control points $S$ fixed, the problem decouples per pixel. We solve for each augmented normal $\\tilde{n}_x$ from its observation matrix $\\tilde{Z}_x\\tilde{Z}_x^\\top$ . The integrability term is also incorporated.", + "- Lighting update: With the normals $\\{\\tilde{n}_x\\}$ fixed, the data term $\\mathcal{L}_{\\mathrm{data}}$ becomes a quadratic function of the lighting control points $S$ . This step reduces to solving a global linear least-squares system to update $S$ ." + ], + "bbox": [ + 89, + 321, + 483, + 441 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We alternate between updating the normals $\\{\\tilde{n}_x\\}$ and the lighting parameters $S$ until convergence. This framework robustly recovers accurate surface normals from event data, even with ambient light and without calibration.", + "bbox": [ + 89, + 441, + 483, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 89, + 518, + 215, + 536 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 544, + 307, + 560 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our EventUPS implementation uses a hybrid CPU/GPU approach: event pre-processing is GPU-accelerated with OpenCL and Rust bindings, while the PyTorch-based optimization framework handles the core algorithm. For the B-spline lighting representation, we use cubic splines (order $l = 4$ ) with 31 control points for each $360^{\\circ}$ rotation cycle unless otherwise specified. The integrability weight is $\\lambda_{\\mathrm{int}} = 0.1$ for all real data and $\\lambda_{\\mathrm{int}} = 0.01$ for all synthetic data.", + "bbox": [ + 89, + 566, + 483, + 700 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We developed a custom lighting system using programmable IC LEDs. For the \"dual ring\" configuration, we mounted concentric LED rings (WS2818B protocol, $800\\mathrm{kHz}$ fixed rate) with different radii sharing a common axis. The inner ring contains 40 LEDs, while the outer ring has 60 LEDs. One complete scanning cycle takes approximately 4 seconds. For the \"trefoil curve\" configuration, we attached an LED ring (APA102, up to $20\\mathrm{MHz}$ SPI clock frequency, 24 LEDs) to a rotating mechanism. The LEDs switch at a frequency three times faster than the rotation mechanism, creating a trefoil curve pattern. One complete cycle takes approximately 1/3 second, limited by the rotation speed of the mechanism.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/db7bcc0bfe49a7187d2a07e06385f813efe2ca2122a05e558a66b5759e0645eb.jpg", + "image_caption": [ + "Figure 6. Performance vs. data rate comparison. This plot shows the average MAE of EventUPS against several frame-based methods on the DiLiGenT-Ev dataset as a function of input frames." + ], + "image_footnote": [], + "bbox": [ + 553, + 87, + 867, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We captured data using a Propheesee EVK4 HD camera with an IMX636 event sensor. The contrast sensitivity threshold biases were set to $-20$ and the dead time bias to $-20$ , resulting in an approximately refractory period of $580~\\mu \\mathrm{s}$ between consecutive events at a pixel.", + "bbox": [ + 511, + 290, + 906, + 366 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Frame-based UPS methods", + "text_level": 1, + "bbox": [ + 511, + 375, + 897, + 391 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Data. To facilitate comparison with frame-based methods, we created a semi-real dataset named DiLiGenT-Ev using images from the DiLiGenT dataset [34]. This procedure, which follows the protocol of EventPS [45], allows for a direct comparison between event-based and frame-based techniques on identical underlying scenes and lighting conditions. Specifically, we selected images captured under the outermost rectangular ring of light directions and converted them to event streams using an event simulator [29].", + "bbox": [ + 511, + 397, + 906, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Results. We benchmark EventUPS against state-of-the-art frame-based UPS methods, including the top-performing non-learning method PF14 [28] from the original DiLiGenT benchmark, and several deep learning-based approaches: CW20 [5], CH19 [4], and LL22 [20]. We evaluate accuracy (Mean Angular Error, MAE) and data efficiency. For data rate calculations, we use Total Events $\\times$ 16 bits/event (encoded in the Prophesee EVT 3.0 format) for event-based methods, and Frames (lighting directions) $\\times$ Resolution $\\times$ 8 bits/pixel $\\times$ 3 exposures (bracketing for the limited dynamic range) for frame-based methods.", + "bbox": [ + 511, + 553, + 905, + 718 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Fig. 6 illustrates the trade-off between data rate and accuracy for frame-based methods—using more images improves accuracy but increases data requirements. Table 1 provides the detailed quantitative breakdown of comparisons with PF14 and CW20. In sharp contrast, EventUPS (marked by a star) demonstrates remarkable data efficiency. EventUPS achieves an MAE of $14.14^{\\circ}$ , a result that is not only superior to the best performance of the non-learning baseline PF14 $(17.16^{\\circ}$ MAE) but does so using a data volume equivalent to roughly two frames. Our method operates at just $1.6\\%$ of the data rate of the 56-frame deep learning setup and less than $5\\%$ of the data rate of PF14. This ef", + "bbox": [ + 511, + 719, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7521", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/958ad1c7fd193ae4441f54ba85d66730e5ea2deae80aa9123fbb527576c6baca.jpg", + "image_caption": [ + "Figure 7. Qualitative results on the DiLiGenT-Ev dataset. Our uncalibrated method produces reconstructions superior to the non-learning, frame-based PF14 [28] and comparable to the calibrated EventPS [45], but without requiring lighting calibration. The \"Light\" columns visualize the estimated trajectories ( $x, y$ components on the unit sphere), with colors denoting relative intensity.", + "Table 1. Quantitative comparison (MAE in degrees) on the DiLiGenT-Ev dataset. Our method outperforms the frame-based method PF14 [28] (best-performing non-learning method in the original DiLiGenT dataset [34]) while using less than $5\\%$ data rate. We include results from a learning-based method CW20 [5] and the only event-based method EventPS [45] (a calibrated baseline of ours, upper bound) as reference. The last column shows the percentage of data rate compared to CW20 with the best performance. The last row indicates the number (#) of events per round for each data." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 903, + 287 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8f54c7e0a05305f6806bc98dad42983bf45e6c8e5e74d1b270c52f0509b5b997.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Uncali-bratedNon-learningInputMethodBALLBUDACATCOWGOLETHARVESTPOT1POT2READERAverageData Rate
✓✘6 framesCW205.3112.109.279.6410.9720.999.197.6113.5010.9511%
✓✘56 framesCW204.9610.237.206.429.4718.557.436.5712.179.22100%
✓✓56 framesPF145.2314.1010.0319.3228.6330.659.8915.7920.8217.1633%
✓✓EventsOurs8.8612.2110.7818.5812.9723.328.7015.1316.7114.141.6%
✘✓EventsEventPS3.4513.227.3223.3313.5429.808.3412.4216.1414.171.6%
# Events470 k306 k354 k458 k200 k327 k246 k334 k393 k343 k/
", + "bbox": [ + 106, + 438, + 893, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ficiency stems from the event camera's inherent ability to capture only meaningful brightness changes.", + "bbox": [ + 89, + 609, + 482, + 641 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fig. 7 shows the visual comparison between different non-learning methods, that is, the frame-based UPS method PF14, the proposed EventUPS, and the calibrated event-based method EventPS. EventUPS reconstructs clean and detailed normals that closely match the ground truth, and accurately recovers the circular lighting trajectory. The normals from PF14 appear visibly noisier and less defined, especially for the challenging HARVEST object, aligning with the error metrics in the table. This demonstrates our method's ability to robustly disentangle geometry, lighting, and non-ideal effects from sparse event data alone. EventUPS's superiority over non-learning, frame-based PF14 shows particular strength on objects with complex peculiarities and textures like HARVEST (23.32° vs. 30.65°) and READER (16.71° vs. 20.82°). Most notably, our uncalibrated method's performance is on par with, and even slightly surpasses, the average performance of the fully cal", + "bbox": [ + 88, + 643, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ibrated event-based method EventPS (14.17° MAE). For objects like HARVEST and BUDDHA, our method is more accurate. We attribute this to our model's explicit residual term, which effectively isolates and accounts for non-ideal, real-world effects like ambient illumination.", + "bbox": [ + 511, + 609, + 906, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Evaluation on Real Camera", + "text_level": 1, + "bbox": [ + 511, + 694, + 764, + 709 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Data. To demonstrate the practical applicability of our method, we fabricated five physical objects with ground truth normal maps and captured real events using our custom lighting systems. These data have varying characteristics: simple geometry (BALL), spatially-varying albedo (BALLICCV), moderate geometric details (BUNNY), and complex details (HORSE, TIGER).", + "bbox": [ + 511, + 715, + 906, + 821 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results. Fig. 8 shows results on our captured real dataset, which presents additional challenges including sensor noise, non-idealities in event generation. Our method produces high-quality normal reconstructions with clear", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7522", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b5f50a3b4db4cd2b8af8a548a1351b6b8e885e734861507b7b46ad50ca48eff1.jpg", + "image_caption": [ + "Figure 8. Results on real data. EventUPS successfully recovers detailed normals despite sensor noise and non-ideal lighting (light trajectories are visualized as in Fig. 1)." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 486, + 411 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "preservation of fine details, achieving an average MAE of $16.58^{\\circ}$ across all objects. The spherical object BALLICCV with \"ICCV\" text highlights an important property: the estimated normal map is invariant to ambient-albedo ratio variations, as the text is barely visible in both the event and the reconstructed normal map. The BUNNY and TIGER objects demonstrate EventUPS's ability to recover intricate geometric details. We observe slightly higher angular errors near object boundaries, which we attribute to near-light effects (the distance between the big ring light and the object is approximately $10\\mathrm{cm}$ ) and imperfect lighting alignment. Nevertheless, the overall quality of reconstruction is high. The estimated lighting configurations accurately reflect the dual ring structure used during capture.", + "bbox": [ + 88, + 489, + 486, + 702 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Robustness to lighting patterns. To evaluate the robustness of our method to different lighting configurations, we captured the same objects using both the \"dual ring\" and \"trefoil curve\" lighting patterns. As shown in Fig. 9, EventUPS produces consistent normal reconstructions regardless of the lighting pattern used. This demonstrates that our approach effectively recovers both surface normals and lighting directions without requiring a specific lighting arrangement, providing flexibility for different practical scenarios. The trefoil pattern yields better results for objects with complex geometry due to its more varied lighting directions, while the dual ring configuration provides more uniform", + "bbox": [ + 89, + 719, + 486, + 902 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c664ac65b1101568e15329ab71443f7e5f53bd3fd0a1ab01c4796c8f343d8e1e.jpg", + "image_caption": [ + "Figure 9. Performance across illumination patterns. EventUPS delivers consistent normal and lighting estimates for both the (a) dual-ring and (b) trefoil curve patterns, showcasing its robustness." + ], + "image_footnote": [], + "bbox": [ + 517, + 87, + 903, + 268 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "coverage for smoother objects.", + "bbox": [ + 511, + 344, + 720, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 372, + 761, + 388 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have presented EventUPS, the first method to solve UPS using an event camera. By introducing a novel augmented null space formulation, a continuous lighting model, and leveraging known lighting fixture geometry, we resolve the fundamental challenges of applying UPS to the event domain. Our experiments show that EventUPS achieves accuracy comparable to or better than its traditional frame-based counterpart while requiring only a fraction (5%) of the data bandwidth. Our method is robust to ambient illumination and different lighting configurations, making it practical for deployment in less controlled environments.", + "bbox": [ + 511, + 397, + 906, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While our approach represents an advance in uncalibrated event-based PS, several avenues for future work remain. First, extending our framework to handle non-Lambertian surfaces would enhance its applicability in real-world scenarios with complex materials. Second, integrating learning-based components could further improve robustness to noise and outliers while maintaining the physical interpretability of our current model. Finally, exploring applications in dynamic scene reconstruction, where both the camera and objects may be in motion, represents an exciting direction for future research.", + "bbox": [ + 511, + 564, + 908, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgment", + "text_level": 1, + "bbox": [ + 513, + 753, + 666, + 771 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work is supported by National Natural Science Foundation of China (Grant Nos. 62136001, 62302019, 624B2006, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources.", + "bbox": [ + 511, + 779, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7523", + "bbox": [ + 482, + 944, + 514, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Neil G. Alldrin, Satya P. Mallick, and David J. Kriegman. Resolving the Generalized Bas-Relief Ambiguity by Entropy Minimization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2", + "[2] Peter N. Belhumeur, David J. Kriegman, and Alan L. Yuille. The Bas-Relief Ambiguity. Int. J. Comput. Vis., 35(1):33-44, 1999. 2, 5", + "[3] M.K. Chandraker, F. Kahl, and D.J. Kriegman. Reflections on the generalized bas-relief ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2005. 2", + "[4] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. Self-Calibrating Deep Photometric Stereo Networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2, 6", + "[5] Guanying Chen, Michael Waechter, Boxin Shi, Kwan-Yee K. Wong, and Yasuyuki Matsushita. What Is Learned in Deep Uncalibrated Photometric Stereo? In Proc. of Eur. Conf. Comput. Vis., 2020. 2, 6, 7, 5", + "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor Lighting Estimation Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[7] Jason Chui, Simon Klenk, and Daniel Cremers. Event-Based Feature Tracking in Continuous Time with Sliding Window Optimization. arxiv preprint arxiv:2107.04536, 2021. 2", + "[8] O. Drbohlav and M. Chaniler. Can two specular pixels calibrate photometric stereo? In Proc. of Int. Conf. Comput. Vis., 2005. 2", + "[9] Ondrej Drbohlav and Radim Šára. Specularities Reduce Ambiguity of Uncalibrated Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2002. 2", + "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1", + "[11] Georgiades. Incorporating the Torrance and Sparrow model of reflectance in uncalibrated photometric stereo. In Proc. of Int. Conf. Comput. Vis., 2003. 2", + "[12] Heng Guo, Fumio Okura, Boxin Shi, Takuya Funatomi, Yasuhiro Mukaigawa, and Yasuyuki Matsushita. Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3", + "[13] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-Fidelity Event-Radiance Recovery Via Transient Event Frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[14] Hideki Hayakawa. Photometric stereo under a light source with arbitrary motion. JOSA A, 11:3079-3089, 1994. 1, 2, 3", + "[15] Micah K Johnson and Edward H Adelson. Shape estimation in natural illumination. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2011. 2, 4", + "[16] Yakun Ju, Xinghui Dong, Yingyu Wang, Lin Qi, and Junyu Dong. A dual-cue network for multispectral photometric stereo. Pattern Recogn., 2020. 3" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[17] Dongyeop Kang, Yu Jin Jang, and Sangchul Won. Development of an inspection system for planar steel surface using multispectral photometric stereo. Optic. Eng., 2013. 3", + "[18] Hyeongwoo Kim, Bennett Wilburn, and Moshe Ben-Ezra. Photometric Stereo for Dynamic Surface Orientations. In Proc. of Eur. Conf. Comput. Vis., 2010. 2", + "[19] Leonid L Kontsevich, AP Petrov, and IS Vergelskaya. Reconstruction of shape from shading in color images. J. Opt. Soc. Am. A, 1994. 3", + "[20] Junxuan Li and Hongdong Li. Self-calibrating Photometric Stereo by Neural Inverse Rendering. In Proc. of Eur. Conf. Comput. Vis., 2022, 2, 6", + "[21] Feng Lu, Yasuyuki Matsushita, Imari Sato, Takahiro Okabe, and Yoichi Sato. Uncalibrated Photometric Stereo for Unknown Isotropic Reflectances. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2", + "[22] Tom Malzbender, Bennett Wilburn, Dan Gelb, and Bill Ambrisco. Surface Enhancement Using Real-Time Photometric Stereo and reflectance transformation. In Proc. of Eurographics Symp. Render. Tech., 2006. 2", + "[23] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. MC3D: Motion contrast 3D scanning. In Proc. of Int. Conf. Comput. Photogr., 2015. 3", + "[24] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. ESL: Event-based structured light. In Proc. of Int. Conf. 3D Vis., 2021. 3", + "[25] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3", + "[26] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE Comput. Graphics Appl., 2014. 3", + "[27] Thoma Papadhimitri and Paolo Favaro. A New Perspective on Uncalibrated Photometric Stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2", + "[28] Thoma Papadhinitri and Paolo Favaro. A Closed-Form, Consistent and Robust Solution to Uncalibrated Photometric Stereo Via Local Diffuse Reflectance Maxima. Int. J. Comput. Vis., 107(2):139-154, 2014. 2, 6, 7, 5", + "[29] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In Proc. of Conf. on Robotics Learning, 2018. 6, 4", + "[30] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. DiLiGenT10 $^2$ : A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2", + "[31] Wonjeong Ryoo, Giljoo Nam, Jae-Sang Hyun, and Sangpil Kim. Event fusion photometric stereo network. *Neural Networks*, 2023. 3", + "[32] Boxin Shi, Yasuyuki Matsushita, Yichen Wei, Chao Xu, and Ping Tan. Self-calibrating photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2010. 2", + "[33] Boxin Shi, Zhe Wu, Zhipeng Mo, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7524", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[34] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2, 6, 7", + "[35] Melvyn L. Smith and Lyndon N. Smith. Dynamic photometric stereo - a new technique for moving surface analysis. Image Vision Comput., 2005. 3", + "[36] Kalyan Sunkavalli, Todd Zickler, and Hanspeter Pfister. Visibility Subspaces: Uncalibrated Photometric Stereo with Shadows. In Proc. of Eur. Conf. Comput. Vis., 2010. 2", + "[37] Tsuyoshi Takatani, Yasuyuki Matsushita, Stephen Lin, Yasuhiro Mukaigawa, and Yasushi Yagi. Enhanced photometric stereo with multispectral images. In Proc. of Int. Conf. on Mach. Vis. Appl., 2013. 3", + "[38] Ping Tan, Satya P. Mallick, Long Quan, David J. Kriegman, and Todd Zickler. Isotropy, Reciprocity and the Generalized Bas-Relief Ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2", + "[39] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and imu for robust visual slam in HDR and high-speed scenarios. IEEE Robot. Autom. Lett., 2018. 2", + "[40] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2", + "[41] Feishi Wang, Jieji Ren, Heng Guo, Mingjun Ren, and Boxin Shi. DiLiGenT-Pi: Photometric stereo for planar surfaces with rich details-benchmark dataset and beyond. In Proc. of Int. Conf. Comput. Vis., 2023. 2", + "[42] Robert J. Woodham. Photometric method for determining surface orientation from multiple images. Opt. Eng., 1980, 1, 2", + "[43] Zhe Wu and Ping Tan. Calibrating Photometric Stereo by Holistic Reflectance Symmetry Analysis. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2", + "[44] Takuya Yoda, Hajime Nagahara, Rin-ichiro Taniguchi, Kei-ichiro Kagawa, Keita Yasutomi, and Shoji Kawahito. The dynamic photometric stereo method using a multi-tap cmos image sensor. Sensors-basel., 2018. 2", + "[45] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-Time Photometric Stereo Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 5", + "[46] A.L. Yuille, D. Snow, R. Epstein, and P.N. Belhumeur. Determining Generative Models of Objects Under Varying Illumination: Shape and Albedo from Multiple Images Using SVD and Integrability. Int. J. Comput. Vis., 35(3):203-222, 1999. 2, 3, 5", + "[47] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided Direct and Global Components Separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3", + "[48] Zhenglong Zhou and Ping Tan. Ring-Light Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + ], + "bbox": [ + 91, + 92, + 480, + 869 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7525", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_model.json b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_model.json new file mode 100644 index 0000000000000000000000000000000000000000..11e6b27a816af89e28f8cde066a54952250dacd5 --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_model.json @@ -0,0 +1,2255 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.151, + 0.131, + 0.848, + 0.154 + ], + "angle": 0, + "content": "EventUPS: Uncalibrated Photometric Stereo Using an Event Camera" + }, + { + "type": "text", + "bbox": [ + 0.189, + 0.179, + 0.807, + 0.217 + ], + "angle": 0, + "content": "Jinxiu Liang\\(^{1,2,3\\#}\\) Bohan Yu\\(^{1,2\\#}\\) Siqi Yang\\(^{1,2,4}\\) Haotian Zhuang\\(^{5}\\) Jieji Ren\\(^{6}\\) Peiqi Duan\\(^{1,2}\\) Boxin Shi\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.218, + 0.827, + 0.235 + ], + "angle": 0, + "content": "\\(^{1}\\) State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.235, + 0.827, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\) National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + }, + { + "type": "list", + "bbox": [ + 0.17, + 0.218, + 0.827, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.23, + 0.253, + 0.767, + 0.269 + ], + "angle": 0, + "content": "\\(^{3}\\) National Institute of Informatics \\(^{4}\\) Institute for Artificial Intelligence, Peking University" + }, + { + "type": "text", + "bbox": [ + 0.334, + 0.271, + 0.664, + 0.288 + ], + "angle": 0, + "content": "5 Tsinghua University 6 Shanghai Jiaotong University" + }, + { + "type": "text", + "bbox": [ + 0.177, + 0.29, + 0.817, + 0.305 + ], + "angle": 0, + "content": "cssherryliang@gmail.com,{ybhl998,yousiki,duanqi0001,shiboxin}@pku.edu.cn," + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.308, + 0.707, + 0.322 + ], + "angle": 0, + "content": "zht23@mails.tsinghua.edu.cn, jiejiren@sjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.357, + 0.327, + 0.373 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.389, + 0.485, + 0.692 + ], + "angle": 0, + "content": "We present EventUPS, the first uncalibrated photometric stereo (UPS) method using an event camera—a neuromorphic sensor that asynchronously detects brightness changes with microsecond resolution. Traditional frame-based UPS methods are hindered by high bandwidth demands and limited use in dynamic scenes. These methods require dense image correspondence under varying illumination and are incompatible with the fundamentally different sensing paradigm of event data. Our approach introduces three key innovations: an augmented null space formulation that directly relates each event to joint constraints on surface normals and lighting, naturally handling ambient illumination; a continuous parameterization of time-varying illumination that connects asynchronous events to synchronized lighting estimation; and a lighting fixture with known relative geometry that reduces ambiguity to a convex-concave uncertainty. We validate EventUPS using a custom-built LED lighting system. Experimental results show that our method achieves accuracy surpassing its frame-based counterpart while requiring only \\(5\\%\\) of the data bandwidth." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.718, + 0.222, + 0.733 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.743, + 0.483, + 0.88 + ], + "angle": 0, + "content": "Photometric stereo (PS) recovers shape from images under varying illumination [14, 42]. For decades, it has been a cornerstone for high-fidelity 3D reconstruction, enabling detailed surface normal recovery for applications from cultural heritage to industrial inspection. The uncalibrated variant of PS (UPS)—where lighting directions are unknown—offers particular practical appeal by eliminating the need for precise calibration apparatus, facilitating deployment in uncontrolled environments. However, this flex" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.359, + 0.709, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.585, + 0.531, + 0.662, + 0.544 + ], + "angle": 0, + "content": "(a) Our Setup" + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.359, + 0.895, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.441, + 0.842, + 0.452 + ], + "angle": 0, + "content": "(c) Our Results" + }, + { + "type": "image", + "bbox": [ + 0.722, + 0.454, + 0.894, + 0.531 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.749, + 0.531, + 0.859, + 0.543 + ], + "angle": 0, + "content": "(b) Captured events" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.558, + 0.907, + 0.656 + ], + "angle": 0, + "content": "Figure 1. We propose an uncalibrated photometric stereo method designed for event cameras, EventUPS. Our approach employs (a) a portable lighting system with known relative geometry (e.g., dual LED rings) but an unknown global pose, removing the need for precise calibration. From (b) sparse, asynchronous events triggered by sequentially activated lights, it jointly recovers (c) dense surface normals and the time-varying lighting trajectory1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.759 + ], + "angle": 0, + "content": "ibility incurs significant costs: conventional uncalibrated PS methods demand dense correspondences across multiple images under varying illumination, imposing high data bandwidth and rendering them ill-suited for dynamic scenes or resource-constrained systems." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.76, + 0.909, + 0.851 + ], + "angle": 0, + "content": "Meanwhile, event cameras represent a paradigm shift in visual sensing [10]. Instead of capturing intensity frames at fixed intervals, these bio-inspired sensors asynchronously detect per-pixel logarithmic brightness changes. This provides microsecond temporal resolution, an exceptional dynamic range (140dB vs. 60dB), and drastically reduced data" + }, + { + "type": "page_footnote", + "bbox": [ + 0.104, + 0.887, + 0.356, + 0.901 + ], + "angle": 0, + "content": "Equal contributions. *Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.513, + 0.863, + 0.907, + 0.901 + ], + "angle": 0, + "content": "\\(^{1}\\)Est. Lights: The estimated time-varying light trajectory \\((x, y\\) components on the unit sphere), with numbers indicating temporal sequence, colors denoting relative intensity (yellow higher, green lower)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7516" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.137 + ], + "angle": 0, + "content": "rates—all critical for platforms like mobile robots. These unique attributes have transformed high-speed tasks such as SLAM [39] and object tracking [7]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.483, + 0.366 + ], + "angle": 0, + "content": "Recent advances in event-based PS—exemplified by the EventPS [45]—have shown promise by leveraging these unique capabilities. By using a calibrated, continuously rotating light source, it computes surface normals directly from brightness change events, naturally accommodating rapid variations in surface orientation even in dynamic scenes. This approach dramatically reduces both latency and data redundancy compared to frame-based methods. However, EventPS [45] requires precise lighting calibration and darkroom conditions, which are challenging outside labs. In many real-world scenarios—such as robotic deployments, motion capture, or industrial settings—maintaining calibrated lighting and negligible ambient illumination becomes impractical due to environmental constraints and operational demands." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.702 + ], + "angle": 0, + "content": "A natural question arises: Can established frame-based UPS approaches adapt directly to event cameras? Our analysis reveals fundamental incompatibilities: i) Inability to handle ambient illumination. Frame-based PS methods often address ambient illumination by capturing a dark frame without designed lighting. Unfortunately, this strategy is not applicable to events as they are generated asynchronously by logarithmic brightness changes—encoding brightness ratios rather than absolute intensities. ii) Absence of synchronous correspondence. Non-learning frame-based UPS [2, 46] typically performs matrix factorization to decompose observed intensities into surface normals and lighting directions, establishing constraints through pixel correspondences under shared illumination conditions. However, when applied to events—which correspond to different lighting directions and occur asynchronously—the establishment of such correspondences becomes challenging. iii) Inapplicability of conventional ambiguity resolution. Existing UPS resolves the remaining ambiguity by leveraging effective priors on albedo characteristics [28], which are not applicable to events since the event generation equation is inherently invariant to albedo." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In this paper, we present EventUPS—the first uncalibrated photometric stereo method using an event camera. Our approach rethinks PS for the event domain, leveraging the unique properties of event cameras to recover detailed surface geometry while simultaneously estimating the time-varying illumination, all without explicit lighting calibration. The core of our contribution is a theoretical framework that formulates each event as a constraint in the null space of an augmented normal vector, enabling joint recovery of surface normals and lighting directions despite the asynchronous nature of events and the presence of ambient illumination. To address the incompatibilities, we make the following technical contributions:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "- An augmented null-space formulation able to handle ambient illumination by modeling it as an unknown variable within event-based constraints on normals and lighting." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.138, + 0.905, + 0.182 + ], + "angle": 0, + "content": "- A continuous parameterization of the time-varying illumination that establishes correspondence by linking asynchronous events to a shared lighting trajectory." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.183, + 0.905, + 0.227 + ], + "angle": 0, + "content": "- A geometric UPS ambiguity resolution technique applicable to event data, which uses a lighting fixture with known relative geometry but an unknown global pose." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.907, + 0.394 + ], + "angle": 0, + "content": "Along with the proposed EventUPS framework, we develop a joint optimization framework that effectively integrates integrability constraints with our event-based null space formulation. To validate our approach in real scenes, we build a hardware prototype around an LED-based lighting system, which implements the lighting configurations with known geometry of dual ring (see Figure 1) and trefoil curve patterns. Experiments on semi-real and real data demonstrate that our method achieves accuracy outperforming the leading frame-based non-learning UPS [28] while requiring only \\(5\\%\\) of the data bandwidth." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.407, + 0.664, + 0.423 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.906, + 0.493 + ], + "angle": 0, + "content": "PS and event-based vision have advanced independently for decades. Our work lies at their intersection, specifically addressing the challenges of UPS with event cameras. Here, we review developments in both domains that inform ours." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.805, + 0.517 + ], + "angle": 0, + "content": "2.1. Uncalibrated Photometric Stereo" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.906, + 0.614 + ], + "angle": 0, + "content": "PS estimates surface normals from images captured under varying illumination directions [15, 42]. While classical approaches assume known lighting, the uncalibrated setting—where lighting is unknown—presents greater challenges but offers superior practical advantages. For comprehensive reviews, see [30, 33, 34, 41]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.615, + 0.906, + 0.84 + ], + "angle": 0, + "content": "A fundamental challenge in UPS is the ambiguity in simultaneously recovering lighting and surface normals. This was first identified as a general \\(3 \\times 3\\) linear ambiguity [14], later reduced to the 3-parameter generalized bas-relief (GBR) ambiguity by imposing integrability constraints [2, 46]. Resolving the ambiguity typically relies on physical cues such as specular reflections [8, 9, 11, 38, 43], statistical priors on albedo [1, 21, 28, 32], shadows [36], interreflections [3], or perspective effects [27]. The light source configuration is also critical; strategic arrangements like ring lights with multi-view information can provide additional constraints to resolve ambiguity without strong surface assumptions [48]. Recent learning-based methods have shown promising results [4, 5, 20], but they often require extensive training data and can struggle with generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Accelerating PS acquisition remains an active research area, with solutions like high-speed cameras [18, 22, 40] or multi-tap CMOS [44] with synchronized illumination, though these dramatically increase bandwidth. Single-shot" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7517" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.152 + ], + "angle": 0, + "content": "multi-spectral systems [19, 26, 35, 37] offer an alternative but face challenges like limited spectral bands, crosstalk, and intensity inconsistencies across bands introduce additional challenges [12, 16, 17]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.159, + 0.456, + 0.175 + ], + "angle": 0, + "content": "2.2. Photometric Methods with Event Cameras" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.181, + 0.485, + 0.408 + ], + "angle": 0, + "content": "With their high temporal resolution, wide dynamic range, and asynchronous operation, event cameras are powerful tools for photometric measurements. These advantages make them particularly suitable for capturing dynamic lighting effects crucial for photometric applications. Their microsecond resolution is especially valuable for active lighting techniques. MC3D [23] pioneered event-based structured light for high-speed 3D scanning, while ESL [24] improved noise robustness by exploiting spatial correlations. Researchers have explored diverse illumination patterns for specific tasks, such as capturing events during light activation to reconstruct iso-contours [13] and address intensity-distance ambiguity by using a diffuse sphere [6], estimating surface normals using rotating polarizers [25], and separating direct/global illumination components [47]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.408, + 0.484, + 0.591 + ], + "angle": 0, + "content": "Most relevant to our work, EFPS-Net [31] fused sparse events with RGB images for normal estimation, while EventPS [45] pioneered purely event-based PS. EventPS showed that the differential nature of event cameras aligns perfectly with PS requirements, but it assumes calibrated lighting and cannot handle ambient illumination. The sparse, asynchronous nature of event data presents unique challenges in uncalibrated settings, particularly in establishing pixel correspondences under consistent lighting. This fundamentally different sensing paradigm makes direct application of non-learning UPS methods [14, 46] infeasible. Our work, EventUPS, is the first to solve these challenges." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.6, + 0.182, + 0.615 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.625, + 0.295, + 0.64 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.647, + 0.484, + 0.738 + ], + "angle": 0, + "content": "Image formulation. We consider a scene captured by an orthographic projection camera2. The surface is assumed to be Lambertian and is illuminated by a combination of a distant, time-varying directional light source and constant ambient illumination. For a pixel \\( \\pmb{x} \\) in the image plane \\( \\Omega \\subset \\mathbb{R}^2 \\), the reflected radiance \\( I_{\\pmb{x}}(t) \\) at time \\( t \\) is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.744, + 0.483, + 0.761 + ], + "angle": 0, + "content": "\\[\nI _ {\\boldsymbol {x}} (t) = \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t) \\right] + b _ {\\boldsymbol {x}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.766, + 0.484, + 0.857 + ], + "angle": 0, + "content": "where \\(\\pmb{n}_{\\pmb{x}} \\in \\mathbb{R}^{3}\\) is the unit surface normal (\\(\\| \\pmb{n}_{\\pmb{x}} \\|_{2} = 1\\)), \\(a_{x} > 0\\) is the diffuse albedo, and \\(b_{x} \\geq 0\\) is the ambient illumination term. The lighting is described by the vector \\(\\pmb{s}(t) \\in \\mathbb{R}^{3}\\) with intensity \\(\\| \\pmb{s}(t) \\|^{2}\\) and unit direction \\(\\hat{\\pmb{s}}(t) = \\pmb{s}(t) / \\| \\pmb{s}(t) \\|\\). The max operator models attached shadows, where the dot product becomes negative." + }, + { + "type": "page_footnote", + "bbox": [ + 0.09, + 0.863, + 0.484, + 0.902 + ], + "angle": 0, + "content": "\\(^{2}\\)Orthographic projection is a widely used assumption in PS [33] when the object-to-camera distance (approximately \\(20~\\mathrm{cm}\\) in our setup) is significantly larger than object depth variations (approximately \\(3\\mathrm{cm}\\))." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.903, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.25, + 0.907, + 0.306 + ], + "angle": 0, + "content": "Figure 2. The null space vectors and estimated normals: (a) Lambertian sphere under ideal conditions, (b) non-Lambertian sphere with real events. EventUPS models ambient and albedo effects via a residual term, enhancing robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.33, + 0.907, + 0.406 + ], + "angle": 0, + "content": "Event generation. Event cameras are bio-inspired sensors that asynchronously report logarithmic changes in brightness. An event \\((\\pmb{x}, t_{\\pmb{x}}^{(k)}, p_{\\pmb{x}}^{(k)})\\) is triggered at pixel \\(\\pmb{x}\\) when the logarithmic change in brightness exceeds a predetermined contrast threshold \\(\\theta > 0\\):" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.414, + 0.907, + 0.433 + ], + "angle": 0, + "content": "\\[\n\\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right)\\right) - \\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right)\\right) \\geq p _ {\\boldsymbol {x}} ^ {(k)} \\theta , \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.441, + 0.906, + 0.506 + ], + "angle": 0, + "content": "where \\(t_{\\pmb{x}}^{(k)}\\) denotes the timestamp of the \\(k\\)-th event at pixel \\(\\pmb{x}\\), and \\(p_{\\pmb{x}}^{(k)} \\in \\{+1, -1\\}\\) indicates whether the brightness increased or decreased. The relationship between consecutive brightness measurements can be approximated as:" + }, + { + "type": "equation", + "bbox": [ + 0.588, + 0.512, + 0.906, + 0.532 + ], + "angle": 0, + "content": "\\[\nI _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\approx I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right) \\cdot \\exp \\left(p _ {\\boldsymbol {x}} ^ {(k)} \\theta\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.54, + 0.879, + 0.559 + ], + "angle": 0, + "content": "For notational simplicity, we define \\(\\tilde{\\theta}_{\\pmb{x}}^{(k)} = \\exp (p_{\\pmb{x}}^{(k)}\\theta)\\)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.574, + 0.907, + 0.7 + ], + "angle": 0, + "content": "UPS with events. Unlike prior event-based works [45] that assume calibrated lighting, we aim to jointly estimate the per-pixel surface normals \\( \\pmb{n}_{\\pmb{x}} \\) and the continuous, time-varying lighting function \\( s(t) \\) from only the event stream \\( \\mathbb{E} = \\{(x,t_x^{(k)},p_x^{(k)})\\}_{x\\in \\Omega} \\) across all pixels \\( x\\in \\Omega \\). This inverse problem is formulated as the minimization of the joint loss function that measures consistency between the observed events and the image formation model:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.707, + 0.907, + 0.726 + ], + "angle": 0, + "content": "\\[\n\\operatorname {a r g m i n} _ {\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\} _ {\\boldsymbol {x} \\in \\Omega}, \\boldsymbol {s} (t)} \\mathcal {L} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {s} (t); \\mathbb {E}\\right). \\tag {4}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.731, + 0.845, + 0.748 + ], + "angle": 0, + "content": "3.2. Joint Estimating Normal and Lighting" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.753, + 0.907, + 0.845 + ], + "angle": 0, + "content": "The augmented null-space formulation. Conventional photometric stereo often handles ambient light by subtracting a dark frame. This is infeasible for event-based methods where measurements are relative and asynchronous. We propose an alternative by substituting the image model Eq. (1) into the event model Eq. (3):" + }, + { + "type": "equation", + "bbox": [ + 0.565, + 0.851, + 0.855, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k)}) \\right] + b _ {\\boldsymbol {x}} \\\\ \\approx \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} \\cdot (\\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k - 1)}) \\right] + b _ {\\boldsymbol {x}}). \\\\ \\end{array}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7518" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "For most observed events, the surface point is illuminated at both timestamps (i.e., it is not in an attached shadow). This allows us to remove the max operator. The albedo \\( a_{\\pmb{x}} \\) can be eliminated by dividing it out, as it appears on both sides. This makes our formulation inherently albedo-invariant. Rearranging the terms yields a linear constraint on the normal \\( n_{\\pmb{x}} \\) and the \"residual\" pixel-wise constant ambient-albedo ratio \\( r_{\\pmb{x}} = b_{\\pmb{x}} / a_{\\pmb{x}} \\geq 0 \\):" + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.224, + 0.483, + 0.246 + ], + "angle": 0, + "content": "\\[\n\\pmb {n} _ {\\pmb {x}} \\cdot (\\pmb {s} (t _ {\\pmb {x}} ^ {(k)}) - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)} \\pmb {s} (t _ {\\pmb {x}} ^ {(k - 1)})) + r _ {\\pmb {x}} (1 - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)}) = 0. \\quad (6)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.257, + 0.484, + 0.331 + ], + "angle": 0, + "content": "Shadow events, which occur when one of the terms under the max operator becomes zero, are relatively rare and can be identified through temporal consistency checks. When they do occur, they provide constraints at shadow boundaries that can be incorporated into our framework." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.332, + 0.484, + 0.416 + ], + "angle": 0, + "content": "Let \\( \\boldsymbol{z}_{\\boldsymbol{x}}^{(k)} = \\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k)}) - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)}\\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k-1)}) \\). We define an augmented normal vector \\( \\tilde{\\boldsymbol{n}}_{\\boldsymbol{x}} = [\\boldsymbol{n}_{\\boldsymbol{x}}^{\\top}, r_{\\boldsymbol{x}}]^{\\top} \\in \\mathbb{R}^{4} \\), which is constant for each pixel \\( \\boldsymbol{x} \\). For each event, we define a corresponding augmented null space vector \\( \\tilde{\\boldsymbol{z}}_{\\boldsymbol{x}}^{(k)} = [\\boldsymbol{z}_{\\boldsymbol{x}}^{(k)\\top}, (1 - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)})]^{\\top} \\). This recasts Eq. (6) into an geometric relationship:" + }, + { + "type": "equation", + "bbox": [ + 0.238, + 0.431, + 0.483, + 0.451 + ], + "angle": 0, + "content": "\\[\n\\tilde {z} _ {\\boldsymbol {x}} ^ {(k) \\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} = 0. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.6 + ], + "angle": 0, + "content": "This means the augmented normal \\(\\tilde{n}_{\\pmb{x}}\\) must lie in the null space of the vectors \\(\\{\\tilde{z}_x^{(k)}\\}\\). Please see Fig. 2 for an illustration. With at least three linearly independent null space vectors from events at a pixel, we can determine the augmented normal \\(\\tilde{n}_{\\pmb{x}}\\) up to a scale factor. By stacking all augmented null space vectors \\(\\tilde{z}_x^{(k)}\\) for a pixel \\(\\pmb{x}\\) into a matrix \\(Z_{\\pmb{x}} \\in \\mathbb{R}^{4 \\times K_{\\pmb{x}}}\\) where \\(K_{\\pmb{x}}\\) is the number of events at that pixel and the \\(k\\)-th row is \\(\\tilde{z}_x^{(k)\\top}\\), we solve for the normal by finding the vector that minimizes the projection:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.611, + 0.483, + 0.639 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} ^ {*} = \\underset {\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}}} {\\operatorname {a r g m i n}} \\| Z _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2} \\quad \\text {s u b j e c t t o} \\quad \\| \\boldsymbol {n} _ {\\boldsymbol {x}} \\| _ {2} = 1. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.651, + 0.484, + 0.725 + ], + "angle": 0, + "content": "This constrained optimization problem has an elegant solution via Singular Value Decomposition (SVD) of the matrix \\( Z_{x}Z_{x}^{\\top}\\in \\mathbb{R}^{4\\times 4} \\). The solution is the eigenvector corresponding to the smallest eigenvalue, appropriately scaled to ensure the unit normal constraint." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Continuous lighting parameterization. In the uncalibrated setting, the lighting direction \\( s(t) \\) is unknown and varies continuously over time. The asynchronous nature of events makes it challenging to establish correspondences across pixels under identical lighting conditions, which is a prerequisite for non-learning uncalibrated PS methods. To address this challenge, we model \\( s(t) \\) as a B-spline curve. This provides a smooth, local, and efficient parameterization that is ideal for asynchronous data. By selecting \\( M \\) control points \\( \\{s(t_j)\\}_{j=0}^{M-1} \\) at uniform times \\( t_j = t_0 + j\\Delta t \\)," + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.09, + 0.902, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.323, + 0.907, + 0.393 + ], + "angle": 0, + "content": "Figure 3. Continuous illumination model. We represent the time-varying lighting \\( s(t) \\) as a B-spline controlled by a set of points. This continuous representation links asynchronous events to a globally synchronized lighting estimate, shown here for a rotating ring pattern projected onto a 2D plane." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.415, + 0.838, + 0.43 + ], + "angle": 0, + "content": "the lighting at any time \\(t\\) is a linear combination:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.436, + 0.906, + 0.479 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} (t) = \\sum_ {j = 0} ^ {M - 1} C _ {j, l} (t) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.485, + 0.906, + 0.578 + ], + "angle": 0, + "content": "where \\(\\{C_{j,l}(t)\\}\\) are the B-spline basis functions of order \\(l\\). An important property of this formulation is that the value of \\(s(t)\\) for \\(t_j \\leq t < t_{j+1}\\) depends only on \\(l\\) consecutive control points \\(s(t_j), s(t_{j+1}), \\ldots, s(t_{j+l-1})\\), providing locality in the temporal domain. This links the null space vector \\(z_x^{(k)}\\) directly to the control points:" + }, + { + "type": "equation", + "bbox": [ + 0.615, + 0.584, + 0.906, + 0.627 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {z} _ {\\boldsymbol {x}} ^ {(k)} = \\sum_ {j = 0} ^ {M - 1} \\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.633, + 0.56, + 0.645 + ], + "angle": 0, + "content": "where" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.651, + 0.905, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) = C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) - \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.677, + 0.906, + 0.818 + ], + "angle": 0, + "content": "This expression depends only on the observed event times-stamps \\(t_{\\pmb{x}}^{(k)}\\) at pixel \\(\\pmb{x}\\), providing a crucial bridge between pixel-specific asynchronous events and a global, continuous lighting model (see Fig. 3). In matrix form, we stack the control points \\(\\{s(t_j)\\}_{j = 0}^{M - 1}\\) into a matrix \\(\\pmb {S}\\in \\mathbb{R}^{3\\times M}\\) and the time-dependent coefficients \\(\\widetilde{C}_{j,l}(t_x^{(k)})\\) into a matrix \\(C_x\\in \\mathbb{R}^{M\\times K_x}\\). The augmented measurement matrix for a pixel \\(\\pmb{x}\\) with \\(\\tilde{\\theta}_{\\pmb{x}} = [(1 - \\tilde{\\theta}_{\\pmb{x}}^{(1)}),\\dots ,(1 - \\tilde{\\theta}_{\\pmb{x}}^{(K_{\\pmb{x}})})]\\in \\mathbb{R}^{K_{\\pmb{x}}}\\) becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.623, + 0.818, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} = \\left[ \\begin{array}{c} \\boldsymbol {S C} _ {\\boldsymbol {x}} \\\\ \\tilde {\\boldsymbol {\\theta}} _ {\\boldsymbol {x}} ^ {\\top} \\end{array} \\right] \\in \\mathbb {R} ^ {4 \\times K _ {\\boldsymbol {x}}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.853, + 0.675, + 0.866 + ], + "angle": 0, + "content": "The data term becomes:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.872, + 0.905, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) = \\sum_ {\\boldsymbol {x} \\in \\Omega} \\| \\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2}. \\tag {13}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7519" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.1, + 0.089, + 0.478, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.461 + ], + "angle": 0, + "content": "Figure 4. Resolving Ambiguity. With ground-truth lighting (a), the normal is recovered accurately. Assuming incorrect lighting transformations (b-e), estimating normals without constraints leads to large errors. The strong correlation between normal error and integrability cost validates the effectiveness of enforcing integrability with a known lighting fixture to resolve ambiguity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.484, + 0.484, + 0.515 + ], + "angle": 0, + "content": "This bilinear structure naturally lends itself to an alternating optimization scheme." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.525, + 0.291, + 0.542 + ], + "angle": 0, + "content": "3.3. Resolving Ambiguity" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.547, + 0.484, + 0.775 + ], + "angle": 0, + "content": "Integrability. UPS suffers from a fundamental ambiguity where multiple combinations of surface normals and lighting directions can produce identical observations. Physical constraints, particularly the integrability of surface normals, require that the normals be derivable from the gradient of an underlying surface. For non-learning frame-based UPS, this constraint reduces the general linear ambiguity to the GBR transformation [2, 46], which maps the surface \\( z(x) \\) to \\( \\lambda z(x) + \\mu x + \\nu y \\), where \\( \\lambda \\neq 0 \\). For event-based UPS, imposing integrability similarly reduces the ambiguity to the GBR transformation, as the constraint acts on the normal field independently of the observation modality. Specifically, we use an integrability term. It ensures the normal field \\( \\{n_x\\} \\) corresponds to a valid, continuous surface by penalizing the curl of the corresponding gradient field:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.785, + 0.484, + 0.824 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {i n t}} = \\sum_ {\\boldsymbol {x}} \\left\\| \\frac {\\partial}{\\partial y} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {x}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) - \\frac {\\partial}{\\partial x} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {y}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) \\right\\| ^ {2}. \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Constraints from lighting fixture configurations. To further resolve this ambiguity, we leverage a portable lighting system with a known 3D structure undergoing unknown 6-DoF transformations. The system consists of light" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.089, + 0.669, + 0.265 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.675, + 0.094, + 0.783, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.79, + 0.094, + 0.898, + 0.266 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.279, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Figure 5. The lighting configurations employed in our experiments. We built physical prototypes for the \"dual ring\" and the \"trefoil curve\" patterns using programmable LEDs, shown both statically (left) and as dynamic trajectories captured with long exposure (right). We also test on the standard \"DiLiGenT\" pattern." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.557 + ], + "angle": 0, + "content": "sources at known positions \\( \\pmb{p}_i \\) (for \\( i = 1,2,\\dots,N \\)) in a local reference frame. Assuming the camera is at the origin, the world positions become \\( \\pmb{p}_i' = \\pmb{R}\\pmb{p}_i + \\pmb{t} \\) under a rigid transformation with orthogonal matrix \\( \\pmb{R} \\in O(3) \\) and translation vector \\( \\pmb{t} \\in \\mathbb{R}^3 \\). The light vector for the \\( i \\)-th source is \\( s_i = u_i p_i' \\) with unknown intensity \\( u_i > 0 \\). A GBR transformation matrix \\( \\pmb{G} \\) acts on the recovered normals, and an inverse transformation \\( \\pmb{G}^{-\\top} \\) acts on the light vectors. The observed lighting \\( s_i' \\) is related to the true lighting \\( s_i \\) by \\( s_i' = \\pmb{G}^{-\\top}s_i \\). Combining these, we obtain a relationship between the known local positions \\( \\pmb{p}_i \\) and the estimated, transformed light vectors \\( s_i' \\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.566, + 0.907, + 0.585 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {s} _ {i} ^ {\\prime} = u _ {i} \\boldsymbol {G} ^ {- \\top} (\\boldsymbol {R p} _ {i} + \\boldsymbol {t}). \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.763 + ], + "angle": 0, + "content": "This system of equations constrains the possible transformations. As formally proven in the supplement, these constraints reduce the GBR ambiguity to a discrete set. Specifically, for a set of at least seven light positions in general position (i.e., not all lying on any nontrivial quadric surface of a specified form passing through the origin), under physical constraints, the ambiguity can be resolved up to the classic convex/concave reflection. This eliminates the need for albedo or reflectance priors. An illustration of integrability's role in resolving positional ambiguity in our setup is shown in Fig. 4." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Lighting configurations. For event cameras that asynchronously detect brightness changes with high temporal resolution, we sequentially activate each light (e.g., on millisecond timescales) to generate an event stream, with timestamps correlating to lighting changes. We designed two physical patterns (see Fig. 5) that satisfy the geometric requirements: a coaxial dual-ring consisting of LEDs arranged on two concentric rings of different radii, separated" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7520" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "by a known axial distance; and a trefoil curve with a moving ring of lights along a non-planar, asymmetric 3D trefoil knot trajectory. Both configurations meet the conditions for ambiguity resolution. Details of the theorem, proof, and lighting configurations are provided in the supplement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.179, + 0.231, + 0.195 + ], + "angle": 0, + "content": "3.4. Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.202, + 0.484, + 0.232 + ], + "angle": 0, + "content": "We solve for normals and lighting via alternating optimization, minimizing a joint loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.157, + 0.246, + 0.483, + 0.263 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\text {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) + \\lambda_ {\\text {i n t}} \\mathcal {L} _ {\\text {i n t}} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}\\right), \\tag {16}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.321 + ], + "angle": 0, + "content": "where the weight \\(\\lambda_{\\mathrm{int}}\\) balances data fidelity Eq. (13) against surface smoothness Eq. (14). The optimization proceeds by alternating between two main steps until convergence:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.322, + 0.483, + 0.381 + ], + "angle": 0, + "content": "- Normal update: With the lighting control points \\( S \\) fixed, the problem decouples per pixel. We solve for each augmented normal \\( \\tilde{n}_x \\) from its observation matrix \\( \\tilde{Z}_x\\tilde{Z}_x^\\top \\). The integrability term is also incorporated." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.382, + 0.484, + 0.442 + ], + "angle": 0, + "content": "- Lighting update: With the normals \\(\\{\\tilde{n}_x\\}\\) fixed, the data term \\(\\mathcal{L}_{\\mathrm{data}}\\) becomes a quadratic function of the lighting control points \\(S\\). This step reduces to solving a global linear least-squares system to update \\(S\\)." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.322, + 0.484, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.443, + 0.484, + 0.504 + ], + "angle": 0, + "content": "We alternate between updating the normals \\(\\{\\tilde{n}_x\\}\\) and the lighting parameters \\(S\\) until convergence. This framework robustly recovers accurate surface normals from event data, even with ambient light and without calibration." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.52, + 0.216, + 0.537 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.545, + 0.308, + 0.561 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.702 + ], + "angle": 0, + "content": "Our EventUPS implementation uses a hybrid CPU/GPU approach: event pre-processing is GPU-accelerated with OpenCL and Rust bindings, while the PyTorch-based optimization framework handles the core algorithm. For the B-spline lighting representation, we use cubic splines (order \\( l = 4 \\)) with 31 control points for each \\( 360^{\\circ} \\) rotation cycle unless otherwise specified. The integrability weight is \\( \\lambda_{\\mathrm{int}} = 0.1 \\) for all real data and \\( \\lambda_{\\mathrm{int}} = 0.01 \\) for all synthetic data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We developed a custom lighting system using programmable IC LEDs. For the \"dual ring\" configuration, we mounted concentric LED rings (WS2818B protocol, \\(800\\mathrm{kHz}\\) fixed rate) with different radii sharing a common axis. The inner ring contains 40 LEDs, while the outer ring has 60 LEDs. One complete scanning cycle takes approximately 4 seconds. For the \"trefoil curve\" configuration, we attached an LED ring (APA102, up to \\(20\\mathrm{MHz}\\) SPI clock frequency, 24 LEDs) to a rotating mechanism. The LEDs switch at a frequency three times faster than the rotation mechanism, creating a trefoil curve pattern. One complete cycle takes approximately 1/3 second, limited by the rotation speed of the mechanism." + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.088, + 0.868, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.222, + 0.906, + 0.264 + ], + "angle": 0, + "content": "Figure 6. Performance vs. data rate comparison. This plot shows the average MAE of EventUPS against several frame-based methods on the DiLiGenT-Ev dataset as a function of input frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.291, + 0.907, + 0.367 + ], + "angle": 0, + "content": "We captured data using a Propheesee EVK4 HD camera with an IMX636 event sensor. The contrast sensitivity threshold biases were set to \\(-20\\) and the dead time bias to \\(-20\\), resulting in an approximately refractory period of \\(580~\\mu \\mathrm{s}\\) between consecutive events at a pixel." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.376, + 0.898, + 0.392 + ], + "angle": 0, + "content": "4.2. Comparison with Frame-based UPS methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Data. To facilitate comparison with frame-based methods, we created a semi-real dataset named DiLiGenT-Ev using images from the DiLiGenT dataset [34]. This procedure, which follows the protocol of EventPS [45], allows for a direct comparison between event-based and frame-based techniques on identical underlying scenes and lighting conditions. Specifically, we selected images captured under the outermost rectangular ring of light directions and converted them to event streams using an event simulator [29]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.719 + ], + "angle": 0, + "content": "Results. We benchmark EventUPS against state-of-the-art frame-based UPS methods, including the top-performing non-learning method PF14 [28] from the original DiLiGenT benchmark, and several deep learning-based approaches: CW20 [5], CH19 [4], and LL22 [20]. We evaluate accuracy (Mean Angular Error, MAE) and data efficiency. For data rate calculations, we use Total Events \\(\\times\\) 16 bits/event (encoded in the Prophesee EVT 3.0 format) for event-based methods, and Frames (lighting directions) \\(\\times\\) Resolution \\(\\times\\) 8 bits/pixel \\(\\times\\) 3 exposures (bracketing for the limited dynamic range) for frame-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Fig. 6 illustrates the trade-off between data rate and accuracy for frame-based methods—using more images improves accuracy but increases data requirements. Table 1 provides the detailed quantitative breakdown of comparisons with PF14 and CW20. In sharp contrast, EventUPS (marked by a star) demonstrates remarkable data efficiency. EventUPS achieves an MAE of \\(14.14^{\\circ}\\), a result that is not only superior to the best performance of the non-learning baseline PF14 \\((17.16^{\\circ}\\) MAE) but does so using a data volume equivalent to roughly two frames. Our method operates at just \\(1.6\\%\\) of the data rate of the 56-frame deep learning setup and less than \\(5\\%\\) of the data rate of PF14. This ef" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "7521" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.905, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.298, + 0.908, + 0.342 + ], + "angle": 0, + "content": "Figure 7. Qualitative results on the DiLiGenT-Ev dataset. Our uncalibrated method produces reconstructions superior to the non-learning, frame-based PF14 [28] and comparable to the calibrated EventPS [45], but without requiring lighting calibration. The \"Light\" columns visualize the estimated trajectories (\\(x, y\\) components on the unit sphere), with colors denoting relative intensity." + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.354, + 0.908, + 0.424 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison (MAE in degrees) on the DiLiGenT-Ev dataset. Our method outperforms the frame-based method PF14 [28] (best-performing non-learning method in the original DiLiGenT dataset [34]) while using less than \\(5\\%\\) data rate. We include results from a learning-based method CW20 [5] and the only event-based method EventPS [45] (a calibrated baseline of ours, upper bound) as reference. The last column shows the percentage of data rate compared to CW20 with the best performance. The last row indicates the number (#) of events per round for each data." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.439, + 0.894, + 0.584 + ], + "angle": 0, + "content": "
Uncali-bratedNon-learningInputMethodBALLBUDACATCOWGOLETHARVESTPOT1POT2READERAverageData Rate
✓✘6 framesCW205.3112.109.279.6410.9720.999.197.6113.5010.9511%
✓✘56 framesCW204.9610.237.206.429.4718.557.436.5712.179.22100%
✓✓56 framesPF145.2314.1010.0319.3228.6330.659.8915.7920.8217.1633%
✓✓EventsOurs8.8612.2110.7818.5812.9723.328.7015.1316.7114.141.6%
✘✓EventsEventPS3.4513.227.3223.3313.5429.808.3412.4216.1414.171.6%
# Events470 k306 k354 k458 k200 k327 k246 k334 k393 k343 k/
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.642 + ], + "angle": 0, + "content": "ficiency stems from the event camera's inherent ability to capture only meaningful brightness changes." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.644, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Fig. 7 shows the visual comparison between different non-learning methods, that is, the frame-based UPS method PF14, the proposed EventUPS, and the calibrated event-based method EventPS. EventUPS reconstructs clean and detailed normals that closely match the ground truth, and accurately recovers the circular lighting trajectory. The normals from PF14 appear visibly noisier and less defined, especially for the challenging HARVEST object, aligning with the error metrics in the table. This demonstrates our method's ability to robustly disentangle geometry, lighting, and non-ideal effects from sparse event data alone. EventUPS's superiority over non-learning, frame-based PF14 shows particular strength on objects with complex peculiarities and textures like HARVEST (23.32° vs. 30.65°) and READER (16.71° vs. 20.82°). Most notably, our uncalibrated method's performance is on par with, and even slightly surpasses, the average performance of the fully cal" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.611, + 0.907, + 0.687 + ], + "angle": 0, + "content": "ibrated event-based method EventPS (14.17° MAE). For objects like HARVEST and BUDDHA, our method is more accurate. We attribute this to our model's explicit residual term, which effectively isolates and accounts for non-ideal, real-world effects like ambient illumination." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.766, + 0.71 + ], + "angle": 0, + "content": "4.3. Evaluation on Real Camera" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.907, + 0.823 + ], + "angle": 0, + "content": "Data. To demonstrate the practical applicability of our method, we fabricated five physical objects with ground truth normal maps and captured real events using our custom lighting systems. These data have varying characteristics: simple geometry (BALL), spatially-varying albedo (BALLICCV), moderate geometric details (BUNNY), and complex details (HORSE, TIGER)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Results. Fig. 8 shows results on our captured real dataset, which presents additional challenges including sensor noise, non-idealities in event generation. Our method produces high-quality normal reconstructions with clear" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7522" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.487, + 0.412 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.422, + 0.484, + 0.466 + ], + "angle": 0, + "content": "Figure 8. Results on real data. EventUPS successfully recovers detailed normals despite sensor noise and non-ideal lighting (light trajectories are visualized as in Fig. 1)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.49, + 0.487, + 0.703 + ], + "angle": 0, + "content": "preservation of fine details, achieving an average MAE of \\(16.58^{\\circ}\\) across all objects. The spherical object BALLICCV with \"ICCV\" text highlights an important property: the estimated normal map is invariant to ambient-albedo ratio variations, as the text is barely visible in both the event and the reconstructed normal map. The BUNNY and TIGER objects demonstrate EventUPS's ability to recover intricate geometric details. We observe slightly higher angular errors near object boundaries, which we attribute to near-light effects (the distance between the big ring light and the object is approximately \\(10\\mathrm{cm}\\)) and imperfect lighting alignment. Nevertheless, the overall quality of reconstruction is high. The estimated lighting configurations accurately reflect the dual ring structure used during capture." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.487, + 0.903 + ], + "angle": 0, + "content": "Robustness to lighting patterns. To evaluate the robustness of our method to different lighting configurations, we captured the same objects using both the \"dual ring\" and \"trefoil curve\" lighting patterns. As shown in Fig. 9, EventUPS produces consistent normal reconstructions regardless of the lighting pattern used. This demonstrates that our approach effectively recovers both surface normals and lighting directions without requiring a specific lighting arrangement, providing flexibility for different practical scenarios. The trefoil pattern yields better results for objects with complex geometry due to its more varied lighting directions, while the dual ring configuration provides more uniform" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.088, + 0.905, + 0.269 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.278, + 0.908, + 0.322 + ], + "angle": 0, + "content": "Figure 9. Performance across illumination patterns. EventUPS delivers consistent normal and lighting estimates for both the (a) dual-ring and (b) trefoil curve patterns, showcasing its robustness." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.345, + 0.722, + 0.362 + ], + "angle": 0, + "content": "coverage for smoother objects." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.373, + 0.763, + 0.389 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.398, + 0.907, + 0.565 + ], + "angle": 0, + "content": "We have presented EventUPS, the first method to solve UPS using an event camera. By introducing a novel augmented null space formulation, a continuous lighting model, and leveraging known lighting fixture geometry, we resolve the fundamental challenges of applying UPS to the event domain. Our experiments show that EventUPS achieves accuracy comparable to or better than its traditional frame-based counterpart while requiring only a fraction (5%) of the data bandwidth. Our method is robust to ambient illumination and different lighting configurations, making it practical for deployment in less controlled environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.565, + 0.909, + 0.732 + ], + "angle": 0, + "content": "While our approach represents an advance in uncalibrated event-based PS, several avenues for future work remain. First, extending our framework to handle non-Lambertian surfaces would enhance its applicability in real-world scenarios with complex materials. Second, integrating learning-based components could further improve robustness to noise and outliers while maintaining the physical interpretability of our current model. Finally, exploring applications in dynamic scene reconstruction, where both the camera and objects may be in motion, represents an exciting direction for future research." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.755, + 0.667, + 0.772 + ], + "angle": 0, + "content": "Acknowledgment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This work is supported by National Natural Science Foundation of China (Grant Nos. 62136001, 62302019, 624B2006, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.958 + ], + "angle": 0, + "content": "7523" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.17 + ], + "angle": 0, + "content": "[1] Neil G. Alldrin, Satya P. Mallick, and David J. Kriegman. Resolving the Generalized Bas-Relief Ambiguity by Entropy Minimization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.171, + 0.484, + 0.212 + ], + "angle": 0, + "content": "[2] Peter N. Belhumeur, David J. Kriegman, and Alan L. Yuille. The Bas-Relief Ambiguity. Int. J. Comput. Vis., 35(1):33-44, 1999. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.213, + 0.484, + 0.256 + ], + "angle": 0, + "content": "[3] M.K. Chandraker, F. Kahl, and D.J. Kriegman. Reflections on the generalized bas-relief ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.257, + 0.483, + 0.311 + ], + "angle": 0, + "content": "[4] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. Self-Calibrating Deep Photometric Stereo Networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.312, + 0.483, + 0.367 + ], + "angle": 0, + "content": "[5] Guanying Chen, Michael Waechter, Boxin Shi, Kwan-Yee K. Wong, and Yasuyuki Matsushita. What Is Learned in Deep Uncalibrated Photometric Stereo? In Proc. of Eur. Conf. Comput. Vis., 2020. 2, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.368, + 0.483, + 0.422 + ], + "angle": 0, + "content": "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor Lighting Estimation Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.424, + 0.483, + 0.465 + ], + "angle": 0, + "content": "[7] Jason Chui, Simon Klenk, and Daniel Cremers. Event-Based Feature Tracking in Continuous Time with Sliding Window Optimization. arxiv preprint arxiv:2107.04536, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.466, + 0.483, + 0.507 + ], + "angle": 0, + "content": "[8] O. Drbohlav and M. Chaniler. Can two specular pixels calibrate photometric stereo? In Proc. of Int. Conf. Comput. Vis., 2005. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.508, + 0.483, + 0.55 + ], + "angle": 0, + "content": "[9] Ondrej Drbohlav and Radim Šára. Specularities Reduce Ambiguity of Uncalibrated Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2002. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.551, + 0.483, + 0.619 + ], + "angle": 0, + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.621, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[11] Georgiades. Incorporating the Torrance and Sparrow model of reflectance in uncalibrated photometric stereo. In Proc. of Int. Conf. Comput. Vis., 2003. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.663, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[12] Heng Guo, Fumio Okura, Boxin Shi, Takuya Funatomi, Yasuhiro Mukaigawa, and Yasuyuki Matsushita. Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.733, + 0.483, + 0.788 + ], + "angle": 0, + "content": "[13] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-Fidelity Event-Radiance Recovery Via Transient Event Frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.789, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[14] Hideki Hayakawa. Photometric stereo under a light source with arbitrary motion. JOSA A, 11:3079-3089, 1994. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[15] Micah K Johnson and Edward H Adelson. Shape estimation in natural illumination. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2011. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.859, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[16] Yakun Ju, Xinghui Dong, Yingyu Wang, Lin Qi, and Junyu Dong. A dual-cue network for multispectral photometric stereo. Pattern Recogn., 2020. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.135 + ], + "angle": 0, + "content": "[17] Dongyeop Kang, Yu Jin Jang, and Sangchul Won. Development of an inspection system for planar steel surface using multispectral photometric stereo. Optic. Eng., 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[18] Hyeongwoo Kim, Bennett Wilburn, and Moshe Ben-Ezra. Photometric Stereo for Dynamic Surface Orientations. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.177, + 0.906, + 0.217 + ], + "angle": 0, + "content": "[19] Leonid L Kontsevich, AP Petrov, and IS Vergelskaya. Reconstruction of shape from shading in color images. J. Opt. Soc. Am. A, 1994. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.218, + 0.906, + 0.259 + ], + "angle": 0, + "content": "[20] Junxuan Li and Hongdong Li. Self-calibrating Photometric Stereo by Neural Inverse Rendering. In Proc. of Eur. Conf. Comput. Vis., 2022, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.315 + ], + "angle": 0, + "content": "[21] Feng Lu, Yasuyuki Matsushita, Imari Sato, Takahiro Okabe, and Yoichi Sato. Uncalibrated Photometric Stereo for Unknown Isotropic Reflectances. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.371 + ], + "angle": 0, + "content": "[22] Tom Malzbender, Bennett Wilburn, Dan Gelb, and Bill Ambrisco. Surface Enhancement Using Real-Time Photometric Stereo and reflectance transformation. In Proc. of Eurographics Symp. Render. Tech., 2006. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.906, + 0.413 + ], + "angle": 0, + "content": "[23] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. MC3D: Motion contrast 3D scanning. In Proc. of Int. Conf. Comput. Photogr., 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.906, + 0.455 + ], + "angle": 0, + "content": "[24] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. ESL: Event-based structured light. In Proc. of Int. Conf. 3D Vis., 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.906, + 0.511 + ], + "angle": 0, + "content": "[25] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.906, + 0.553 + ], + "angle": 0, + "content": "[26] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE Comput. Graphics Appl., 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.906, + 0.594 + ], + "angle": 0, + "content": "[27] Thoma Papadhimitri and Paolo Favaro. A New Perspective on Uncalibrated Photometric Stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.906, + 0.65 + ], + "angle": 0, + "content": "[28] Thoma Papadhinitri and Paolo Favaro. A Closed-Form, Consistent and Robust Solution to Uncalibrated Photometric Stereo Via Local Diffuse Reflectance Maxima. Int. J. Comput. Vis., 107(2):139-154, 2014. 2, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.65, + 0.906, + 0.692 + ], + "angle": 0, + "content": "[29] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In Proc. of Conf. on Robotics Learning, 2018. 6, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.906, + 0.76 + ], + "angle": 0, + "content": "[30] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. DiLiGenT10\\(^2\\): A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.761, + 0.906, + 0.802 + ], + "angle": 0, + "content": "[31] Wonjeong Ryoo, Giljoo Nam, Jae-Sang Hyun, and Sangpil Kim. Event fusion photometric stereo network. *Neural Networks*, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[32] Boxin Shi, Yasuyuki Matsushita, Yichen Wei, Chao Xu, and Ping Tan. Self-calibrating photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[33] Boxin Shi, Zhe Wu, Zhipeng Mo, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7524" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[34] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.19 + ], + "angle": 0, + "content": "[35] Melvyn L. Smith and Lyndon N. Smith. Dynamic photometric stereo - a new technique for moving surface analysis. Image Vision Comput., 2005. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[36] Kalyan Sunkavalli, Todd Zickler, and Hanspeter Pfister. Visibility Subspaces: Uncalibrated Photometric Stereo with Shadows. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[37] Tsuyoshi Takatani, Yasuyuki Matsushita, Stephen Lin, Yasuhiro Mukaigawa, and Yasushi Yagi. Enhanced photometric stereo with multispectral images. In Proc. of Int. Conf. on Mach. Vis. Appl., 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.346 + ], + "angle": 0, + "content": "[38] Ping Tan, Satya P. Mallick, Long Quan, David J. Kriegman, and Todd Zickler. Isotropy, Reciprocity and the Generalized Bas-Relief Ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.348, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[39] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and imu for robust visual slam in HDR and high-speed scenarios. IEEE Robot. Autom. Lett., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[40] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[41] Feishi Wang, Jieji Ren, Heng Guo, Mingjun Ren, and Boxin Shi. DiLiGenT-Pi: Photometric stereo for planar surfaces with rich details-benchmark dataset and beyond. In Proc. of Int. Conf. Comput. Vis., 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.557 + ], + "angle": 0, + "content": "[42] Robert J. Woodham. Photometric method for determining surface orientation from multiple images. Opt. Eng., 1980, 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.561, + 0.482, + 0.602 + ], + "angle": 0, + "content": "[43] Zhe Wu and Ping Tan. Calibrating Photometric Stereo by Holistic Reflectance Symmetry Analysis. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[44] Takuya Yoda, Hajime Nagahara, Rin-ichiro Taniguchi, Kei-ichiro Kagawa, Keita Yasutomi, and Shoji Kawahito. The dynamic photometric stereo method using a multi-tap cmos image sensor. Sensors-basel., 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.714 + ], + "angle": 0, + "content": "[45] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-Time Photometric Stereo Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.784 + ], + "angle": 0, + "content": "[46] A.L. Yuille, D. Snow, R. Epstein, and P.N. Belhumeur. Determining Generative Models of Objects Under Varying Illumination: Shape and Albedo from Multiple Images Using SVD and Integrability. Int. J. Comput. Vis., 35(3):203-222, 1999. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.787, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[47] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided Direct and Global Components Separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[48] Zhenglong Zhou and Ping Tan. Ring-Light Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.871 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "7525" + } + ] +] \ No newline at end of file diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_origin.pdf b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..0f06c518d76c05c7cefb96b57ef49ac999f614f2 --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/c1fa3d5f-f9b9-4099-a85c-b7d3f0d0b412_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c96577525fb73b5ee1e44bdea854c4e207f75d9a8f09bc528690a5add1b52510 +size 2563129 diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/full.md b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b9535e4276025c7e8bcc9d974d9b8c1af4010cec --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/full.md @@ -0,0 +1,333 @@ +# EventUPS: Uncalibrated Photometric Stereo Using an Event Camera + +Jinxiu Liang $^{1,2,3\#}$ Bohan Yu $^{1,2\#}$ Siqi Yang $^{1,2,4}$ Haotian Zhuang $^{5}$ Jieji Ren $^{6}$ Peiqi Duan $^{1,2}$ Boxin Shi $^{1,2*}$ + +$^{1}$ State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University +$^{2}$ National Engineering Research Center of Visual Technology, School of Computer Science, Peking University + +$^{3}$ National Institute of Informatics $^{4}$ Institute for Artificial Intelligence, Peking University + +5 Tsinghua University 6 Shanghai Jiaotong University + +cssherryliang@gmail.com,{ybhl998,yousiki,duanqi0001,shiboxin}@pku.edu.cn, + +zht23@mails.tsinghua.edu.cn, jiejiren@sjtu.edu.cn + +# Abstract + +We present EventUPS, the first uncalibrated photometric stereo (UPS) method using an event camera—a neuromorphic sensor that asynchronously detects brightness changes with microsecond resolution. Traditional frame-based UPS methods are hindered by high bandwidth demands and limited use in dynamic scenes. These methods require dense image correspondence under varying illumination and are incompatible with the fundamentally different sensing paradigm of event data. Our approach introduces three key innovations: an augmented null space formulation that directly relates each event to joint constraints on surface normals and lighting, naturally handling ambient illumination; a continuous parameterization of time-varying illumination that connects asynchronous events to synchronized lighting estimation; and a lighting fixture with known relative geometry that reduces ambiguity to a convex-concave uncertainty. We validate EventUPS using a custom-built LED lighting system. Experimental results show that our method achieves accuracy surpassing its frame-based counterpart while requiring only $5\%$ of the data bandwidth. + +# 1. Introduction + +Photometric stereo (PS) recovers shape from images under varying illumination [14, 42]. For decades, it has been a cornerstone for high-fidelity 3D reconstruction, enabling detailed surface normal recovery for applications from cultural heritage to industrial inspection. The uncalibrated variant of PS (UPS)—where lighting directions are unknown—offers particular practical appeal by eliminating the need for precise calibration apparatus, facilitating deployment in uncontrolled environments. However, this flex + +![](images/8c8e1aa3ad56055e81c2442bed480b9badb7735ed929e17df868a34836401bf3.jpg) +(a) Our Setup + +![](images/8665700ed919e6d969f51d133ae577ead7a336e4a1905dfb03347e5d8b1da8b2.jpg) +(c) Our Results + +![](images/d0e6dcd1faceb1911e8d55be8ad528db4faa8cf1057a411a1041f4d7e128fddd.jpg) +(b) Captured events +Figure 1. We propose an uncalibrated photometric stereo method designed for event cameras, EventUPS. Our approach employs (a) a portable lighting system with known relative geometry (e.g., dual LED rings) but an unknown global pose, removing the need for precise calibration. From (b) sparse, asynchronous events triggered by sequentially activated lights, it jointly recovers (c) dense surface normals and the time-varying lighting trajectory1. + +ibility incurs significant costs: conventional uncalibrated PS methods demand dense correspondences across multiple images under varying illumination, imposing high data bandwidth and rendering them ill-suited for dynamic scenes or resource-constrained systems. + +Meanwhile, event cameras represent a paradigm shift in visual sensing [10]. Instead of capturing intensity frames at fixed intervals, these bio-inspired sensors asynchronously detect per-pixel logarithmic brightness changes. This provides microsecond temporal resolution, an exceptional dynamic range (140dB vs. 60dB), and drastically reduced data + +rates—all critical for platforms like mobile robots. These unique attributes have transformed high-speed tasks such as SLAM [39] and object tracking [7]. + +Recent advances in event-based PS—exemplified by the EventPS [45]—have shown promise by leveraging these unique capabilities. By using a calibrated, continuously rotating light source, it computes surface normals directly from brightness change events, naturally accommodating rapid variations in surface orientation even in dynamic scenes. This approach dramatically reduces both latency and data redundancy compared to frame-based methods. However, EventPS [45] requires precise lighting calibration and darkroom conditions, which are challenging outside labs. In many real-world scenarios—such as robotic deployments, motion capture, or industrial settings—maintaining calibrated lighting and negligible ambient illumination becomes impractical due to environmental constraints and operational demands. + +A natural question arises: Can established frame-based UPS approaches adapt directly to event cameras? Our analysis reveals fundamental incompatibilities: i) Inability to handle ambient illumination. Frame-based PS methods often address ambient illumination by capturing a dark frame without designed lighting. Unfortunately, this strategy is not applicable to events as they are generated asynchronously by logarithmic brightness changes—encoding brightness ratios rather than absolute intensities. ii) Absence of synchronous correspondence. Non-learning frame-based UPS [2, 46] typically performs matrix factorization to decompose observed intensities into surface normals and lighting directions, establishing constraints through pixel correspondences under shared illumination conditions. However, when applied to events—which correspond to different lighting directions and occur asynchronously—the establishment of such correspondences becomes challenging. iii) Inapplicability of conventional ambiguity resolution. Existing UPS resolves the remaining ambiguity by leveraging effective priors on albedo characteristics [28], which are not applicable to events since the event generation equation is inherently invariant to albedo. + +In this paper, we present EventUPS—the first uncalibrated photometric stereo method using an event camera. Our approach rethinks PS for the event domain, leveraging the unique properties of event cameras to recover detailed surface geometry while simultaneously estimating the time-varying illumination, all without explicit lighting calibration. The core of our contribution is a theoretical framework that formulates each event as a constraint in the null space of an augmented normal vector, enabling joint recovery of surface normals and lighting directions despite the asynchronous nature of events and the presence of ambient illumination. To address the incompatibilities, we make the following technical contributions: + +- An augmented null-space formulation able to handle ambient illumination by modeling it as an unknown variable within event-based constraints on normals and lighting. +- A continuous parameterization of the time-varying illumination that establishes correspondence by linking asynchronous events to a shared lighting trajectory. +- A geometric UPS ambiguity resolution technique applicable to event data, which uses a lighting fixture with known relative geometry but an unknown global pose. + +Along with the proposed EventUPS framework, we develop a joint optimization framework that effectively integrates integrability constraints with our event-based null space formulation. To validate our approach in real scenes, we build a hardware prototype around an LED-based lighting system, which implements the lighting configurations with known geometry of dual ring (see Figure 1) and trefoil curve patterns. Experiments on semi-real and real data demonstrate that our method achieves accuracy outperforming the leading frame-based non-learning UPS [28] while requiring only $5\%$ of the data bandwidth. + +# 2. Related Works + +PS and event-based vision have advanced independently for decades. Our work lies at their intersection, specifically addressing the challenges of UPS with event cameras. Here, we review developments in both domains that inform ours. + +# 2.1. Uncalibrated Photometric Stereo + +PS estimates surface normals from images captured under varying illumination directions [15, 42]. While classical approaches assume known lighting, the uncalibrated setting—where lighting is unknown—presents greater challenges but offers superior practical advantages. For comprehensive reviews, see [30, 33, 34, 41]. + +A fundamental challenge in UPS is the ambiguity in simultaneously recovering lighting and surface normals. This was first identified as a general $3 \times 3$ linear ambiguity [14], later reduced to the 3-parameter generalized bas-relief (GBR) ambiguity by imposing integrability constraints [2, 46]. Resolving the ambiguity typically relies on physical cues such as specular reflections [8, 9, 11, 38, 43], statistical priors on albedo [1, 21, 28, 32], shadows [36], interreflections [3], or perspective effects [27]. The light source configuration is also critical; strategic arrangements like ring lights with multi-view information can provide additional constraints to resolve ambiguity without strong surface assumptions [48]. Recent learning-based methods have shown promising results [4, 5, 20], but they often require extensive training data and can struggle with generalization. + +Accelerating PS acquisition remains an active research area, with solutions like high-speed cameras [18, 22, 40] or multi-tap CMOS [44] with synchronized illumination, though these dramatically increase bandwidth. Single-shot + +multi-spectral systems [19, 26, 35, 37] offer an alternative but face challenges like limited spectral bands, crosstalk, and intensity inconsistencies across bands introduce additional challenges [12, 16, 17]. + +# 2.2. Photometric Methods with Event Cameras + +With their high temporal resolution, wide dynamic range, and asynchronous operation, event cameras are powerful tools for photometric measurements. These advantages make them particularly suitable for capturing dynamic lighting effects crucial for photometric applications. Their microsecond resolution is especially valuable for active lighting techniques. MC3D [23] pioneered event-based structured light for high-speed 3D scanning, while ESL [24] improved noise robustness by exploiting spatial correlations. Researchers have explored diverse illumination patterns for specific tasks, such as capturing events during light activation to reconstruct iso-contours [13] and address intensity-distance ambiguity by using a diffuse sphere [6], estimating surface normals using rotating polarizers [25], and separating direct/global illumination components [47]. + +Most relevant to our work, EFPS-Net [31] fused sparse events with RGB images for normal estimation, while EventPS [45] pioneered purely event-based PS. EventPS showed that the differential nature of event cameras aligns perfectly with PS requirements, but it assumes calibrated lighting and cannot handle ambient illumination. The sparse, asynchronous nature of event data presents unique challenges in uncalibrated settings, particularly in establishing pixel correspondences under consistent lighting. This fundamentally different sensing paradigm makes direct application of non-learning UPS methods [14, 46] infeasible. Our work, EventUPS, is the first to solve these challenges. + +# 3. Method + +# 3.1. Problem Formulation + +Image formulation. We consider a scene captured by an orthographic projection camera2. The surface is assumed to be Lambertian and is illuminated by a combination of a distant, time-varying directional light source and constant ambient illumination. For a pixel $\pmb{x}$ in the image plane $\Omega \subset \mathbb{R}^2$ , the reflected radiance $I_{\pmb{x}}(t)$ at time $t$ is given by: + +$$ +I _ {\boldsymbol {x}} (t) = \max \left[ 0, a _ {\boldsymbol {x}} \boldsymbol {n} _ {\boldsymbol {x}} \cdot \boldsymbol {s} (t) \right] + b _ {\boldsymbol {x}}, \tag {1} +$$ + +where $\pmb{n}_{\pmb{x}} \in \mathbb{R}^{3}$ is the unit surface normal ( $\| \pmb{n}_{\pmb{x}} \|_{2} = 1$ ), $a_{x} > 0$ is the diffuse albedo, and $b_{x} \geq 0$ is the ambient illumination term. The lighting is described by the vector $\pmb{s}(t) \in \mathbb{R}^{3}$ with intensity $\| \pmb{s}(t) \|^{2}$ and unit direction $\hat{\pmb{s}}(t) = \pmb{s}(t) / \| \pmb{s}(t) \|$ . The max operator models attached shadows, where the dot product becomes negative. + +![](images/078f49e4491f40c7ab347cc4ce11d522b3f79b9a17527a8320725f73e0ea33d8.jpg) +Figure 2. The null space vectors and estimated normals: (a) Lambertian sphere under ideal conditions, (b) non-Lambertian sphere with real events. EventUPS models ambient and albedo effects via a residual term, enhancing robustness. + +Event generation. Event cameras are bio-inspired sensors that asynchronously report logarithmic changes in brightness. An event $(\pmb{x}, t_{\pmb{x}}^{(k)}, p_{\pmb{x}}^{(k)})$ is triggered at pixel $\pmb{x}$ when the logarithmic change in brightness exceeds a predetermined contrast threshold $\theta > 0$ : + +$$ +\log \left(I _ {\boldsymbol {x}} \left(t _ {\boldsymbol {x}} ^ {(k)}\right)\right) - \log \left(I _ {\boldsymbol {x}} \left(t _ {\boldsymbol {x}} ^ {(k - 1)}\right)\right) \geq p _ {\boldsymbol {x}} ^ {(k)} \theta , \tag {2} +$$ + +where $t_{\pmb{x}}^{(k)}$ denotes the timestamp of the $k$ -th event at pixel $\pmb{x}$ , and $p_{\pmb{x}}^{(k)} \in \{+1, -1\}$ indicates whether the brightness increased or decreased. The relationship between consecutive brightness measurements can be approximated as: + +$$ +I _ {\boldsymbol {x}} \left(t _ {\boldsymbol {x}} ^ {(k)}\right) \approx I _ {\boldsymbol {x}} \left(t _ {\boldsymbol {x}} ^ {(k - 1)}\right) \cdot \exp \left(p _ {\boldsymbol {x}} ^ {(k)} \theta\right). \tag {3} +$$ + +For notational simplicity, we define $\tilde{\theta}_{\pmb{x}}^{(k)} = \exp (p_{\pmb{x}}^{(k)}\theta)$ + +UPS with events. Unlike prior event-based works [45] that assume calibrated lighting, we aim to jointly estimate the per-pixel surface normals $\pmb{n}_{\pmb{x}}$ and the continuous, time-varying lighting function $s(t)$ from only the event stream $\mathbb{E} = \{(x,t_x^{(k)},p_x^{(k)})\}_{x\in \Omega}$ across all pixels $x\in \Omega$ . This inverse problem is formulated as the minimization of the joint loss function that measures consistency between the observed events and the image formation model: + +$$ +\operatorname {a r g m i n} _ {\left\{\boldsymbol {n} _ {\boldsymbol {x}} \right\} _ {\boldsymbol {x} \in \Omega}, \boldsymbol {s} (t)} \mathcal {L} \left(\left\{\boldsymbol {n} _ {\boldsymbol {x}} \right\}, \boldsymbol {s} (t); \mathbb {E}\right). \tag {4} +$$ + +# 3.2. Joint Estimating Normal and Lighting + +The augmented null-space formulation. Conventional photometric stereo often handles ambient light by subtracting a dark frame. This is infeasible for event-based methods where measurements are relative and asynchronous. We propose an alternative by substituting the image model Eq. (1) into the event model Eq. (3): + +$$ +\begin{array}{l} \max \left[ 0, a _ {\boldsymbol {x}} \boldsymbol {n} _ {\boldsymbol {x}} \cdot \boldsymbol {s} (t _ {\boldsymbol {x}} ^ {(k)}) \right] + b _ {\boldsymbol {x}} \\ \approx \tilde {\theta} _ {\boldsymbol {x}} ^ {(k)} \cdot (\max \left[ 0, a _ {\boldsymbol {x}} \boldsymbol {n} _ {\boldsymbol {x}} \cdot \boldsymbol {s} (t _ {\boldsymbol {x}} ^ {(k - 1)}) \right] + b _ {\boldsymbol {x}}). \\ \end{array} +$$ + +For most observed events, the surface point is illuminated at both timestamps (i.e., it is not in an attached shadow). This allows us to remove the max operator. The albedo $a_{\pmb{x}}$ can be eliminated by dividing it out, as it appears on both sides. This makes our formulation inherently albedo-invariant. Rearranging the terms yields a linear constraint on the normal $n_{\pmb{x}}$ and the "residual" pixel-wise constant ambient-albedo ratio $r_{\pmb{x}} = b_{\pmb{x}} / a_{\pmb{x}} \geq 0$ : + +$$ +\pmb {n} _ {\pmb {x}} \cdot (\pmb {s} (t _ {\pmb {x}} ^ {(k)}) - \tilde {\theta} _ {\pmb {x}} ^ {(k)} \pmb {s} (t _ {\pmb {x}} ^ {(k - 1)})) + r _ {\pmb {x}} (1 - \tilde {\theta} _ {\pmb {x}} ^ {(k)}) = 0. \quad (6) +$$ + +Shadow events, which occur when one of the terms under the max operator becomes zero, are relatively rare and can be identified through temporal consistency checks. When they do occur, they provide constraints at shadow boundaries that can be incorporated into our framework. + +Let $\boldsymbol{z}_{\boldsymbol{x}}^{(k)} = \boldsymbol{s}(t_{\boldsymbol{x}}^{(k)}) - \tilde{\theta}_{\boldsymbol{x}}^{(k)}\boldsymbol{s}(t_{\boldsymbol{x}}^{(k-1)})$ . We define an augmented normal vector $\tilde{\boldsymbol{n}}_{\boldsymbol{x}} = [\boldsymbol{n}_{\boldsymbol{x}}^{\top}, r_{\boldsymbol{x}}]^{\top} \in \mathbb{R}^{4}$ , which is constant for each pixel $\boldsymbol{x}$ . For each event, we define a corresponding augmented null space vector $\tilde{\boldsymbol{z}}_{\boldsymbol{x}}^{(k)} = [\boldsymbol{z}_{\boldsymbol{x}}^{(k)\top}, (1 - \tilde{\theta}_{\boldsymbol{x}}^{(k)})]^{\top}$ . This recasts Eq. (6) into an geometric relationship: + +$$ +\tilde {z} _ {\boldsymbol {x}} ^ {(k) \top} \tilde {\boldsymbol {n}} _ {\boldsymbol {x}} = 0. \tag {7} +$$ + +This means the augmented normal $\tilde{n}_{\pmb{x}}$ must lie in the null space of the vectors $\{\tilde{z}_x^{(k)}\}$ . Please see Fig. 2 for an illustration. With at least three linearly independent null space vectors from events at a pixel, we can determine the augmented normal $\tilde{n}_{\pmb{x}}$ up to a scale factor. By stacking all augmented null space vectors $\tilde{z}_x^{(k)}$ for a pixel $\pmb{x}$ into a matrix $Z_{\pmb{x}} \in \mathbb{R}^{4 \times K_{\pmb{x}}}$ where $K_{\pmb{x}}$ is the number of events at that pixel and the $k$ -th row is $\tilde{z}_x^{(k)\top}$ , we solve for the normal by finding the vector that minimizes the projection: + +$$ +\tilde {\boldsymbol {n}} _ {\boldsymbol {x}} ^ {*} = \underset {\tilde {\boldsymbol {n}} _ {\boldsymbol {x}}} {\operatorname {a r g m i n}} \| Z _ {\boldsymbol {x}} ^ {\top} \tilde {\boldsymbol {n}} _ {\boldsymbol {x}} \| _ {2} ^ {2} \quad \text {s u b j e c t t o} \quad \| \boldsymbol {n} _ {\boldsymbol {x}} \| _ {2} = 1. \tag {8} +$$ + +This constrained optimization problem has an elegant solution via Singular Value Decomposition (SVD) of the matrix $Z_{x}Z_{x}^{\top}\in \mathbb{R}^{4\times 4}$ . The solution is the eigenvector corresponding to the smallest eigenvalue, appropriately scaled to ensure the unit normal constraint. + +Continuous lighting parameterization. In the uncalibrated setting, the lighting direction $s(t)$ is unknown and varies continuously over time. The asynchronous nature of events makes it challenging to establish correspondences across pixels under identical lighting conditions, which is a prerequisite for non-learning uncalibrated PS methods. To address this challenge, we model $s(t)$ as a B-spline curve. This provides a smooth, local, and efficient parameterization that is ideal for asynchronous data. By selecting $M$ control points $\{s(t_j)\}_{j=0}^{M-1}$ at uniform times $t_j = t_0 + j\Delta t$ , + +![](images/bd0cbfd15a6849f221219272691ded0ed08711f9cc03010a9d5a6f5c9ec9ff91.jpg) +Figure 3. Continuous illumination model. We represent the time-varying lighting $s(t)$ as a B-spline controlled by a set of points. This continuous representation links asynchronous events to a globally synchronized lighting estimate, shown here for a rotating ring pattern projected onto a 2D plane. + +the lighting at any time $t$ is a linear combination: + +$$ +\boldsymbol {s} (t) = \sum_ {j = 0} ^ {M - 1} C _ {j, l} (t) \boldsymbol {s} \left(t _ {j}\right), \tag {9} +$$ + +where $\{C_{j,l}(t)\}$ are the B-spline basis functions of order $l$ . An important property of this formulation is that the value of $s(t)$ for $t_j \leq t < t_{j+1}$ depends only on $l$ consecutive control points $s(t_j), s(t_{j+1}), \ldots, s(t_{j+l-1})$ , providing locality in the temporal domain. This links the null space vector $z_x^{(k)}$ directly to the control points: + +$$ +\boldsymbol {z} _ {\boldsymbol {x}} ^ {(k)} = \sum_ {j = 0} ^ {M - 1} \widetilde {C} _ {j, l} \left(t _ {\boldsymbol {x}} ^ {(k)}\right) \boldsymbol {s} \left(t _ {j}\right), \tag {10} +$$ + +where + +$$ +\widetilde {C} _ {j, l} \left(t _ {\boldsymbol {x}} ^ {(k)}\right) = C _ {j, l} \left(t _ {\boldsymbol {x}} ^ {(k)}\right) - \tilde {\theta} _ {\boldsymbol {x}} ^ {(k)} C _ {j, l} \left(t _ {\boldsymbol {x}} ^ {(k - 1)}\right), \tag {11} +$$ + +This expression depends only on the observed event times-stamps $t_{\pmb{x}}^{(k)}$ at pixel $\pmb{x}$ , providing a crucial bridge between pixel-specific asynchronous events and a global, continuous lighting model (see Fig. 3). In matrix form, we stack the control points $\{s(t_j)\}_{j = 0}^{M - 1}$ into a matrix $\pmb {S}\in \mathbb{R}^{3\times M}$ and the time-dependent coefficients $\widetilde{C}_{j,l}(t_x^{(k)})$ into a matrix $C_x\in \mathbb{R}^{M\times K_x}$ . The augmented measurement matrix for a pixel $\pmb{x}$ with $\tilde{\theta}_{\pmb{x}} = [(1 - \tilde{\theta}_{\pmb{x}}^{(1)}),\dots ,(1 - \tilde{\theta}_{\pmb{x}}^{(K_{\pmb{x}})})]\in \mathbb{R}^{K_{\pmb{x}}}$ becomes: + +$$ +\tilde {\boldsymbol {Z}} _ {\boldsymbol {x}} = \left[ \begin{array}{c} \boldsymbol {S C} _ {\boldsymbol {x}} \\ \tilde {\boldsymbol {\theta}} _ {\boldsymbol {x}} ^ {\top} \end{array} \right] \in \mathbb {R} ^ {4 \times K _ {\boldsymbol {x}}}. \tag {12} +$$ + +The data term becomes: + +$$ +\mathcal {L} _ {\mathrm {d a t a}} \left(\left\{\tilde {\boldsymbol {n}} _ {\boldsymbol {x}} \right\}, \boldsymbol {S}\right) = \sum_ {\boldsymbol {x} \in \Omega} \| \tilde {\boldsymbol {Z}} _ {\boldsymbol {x}} ^ {\top} \tilde {\boldsymbol {n}} _ {\boldsymbol {x}} \| _ {2} ^ {2}. \tag {13} +$$ + +![](images/325400be889e930192c3e7bdbe447f281b9a19e5cb733bc62cdce102cdef7083.jpg) +Figure 4. Resolving Ambiguity. With ground-truth lighting (a), the normal is recovered accurately. Assuming incorrect lighting transformations (b-e), estimating normals without constraints leads to large errors. The strong correlation between normal error and integrability cost validates the effectiveness of enforcing integrability with a known lighting fixture to resolve ambiguity. + +This bilinear structure naturally lends itself to an alternating optimization scheme. + +# 3.3. Resolving Ambiguity + +Integrability. UPS suffers from a fundamental ambiguity where multiple combinations of surface normals and lighting directions can produce identical observations. Physical constraints, particularly the integrability of surface normals, require that the normals be derivable from the gradient of an underlying surface. For non-learning frame-based UPS, this constraint reduces the general linear ambiguity to the GBR transformation [2, 46], which maps the surface $z(x)$ to $\lambda z(x) + \mu x + \nu y$ , where $\lambda \neq 0$ . For event-based UPS, imposing integrability similarly reduces the ambiguity to the GBR transformation, as the constraint acts on the normal field independently of the observation modality. Specifically, we use an integrability term. It ensures the normal field $\{n_x\}$ corresponds to a valid, continuous surface by penalizing the curl of the corresponding gradient field: + +$$ +\mathcal {L} _ {\text {i n t}} = \sum_ {\boldsymbol {x}} \left\| \frac {\partial}{\partial y} \left(\frac {\left(\boldsymbol {n} _ {\boldsymbol {x}}\right) _ {x}}{\left(\boldsymbol {n} _ {\boldsymbol {x}}\right) _ {z}}\right) - \frac {\partial}{\partial x} \left(\frac {\left(\boldsymbol {n} _ {\boldsymbol {x}}\right) _ {y}}{\left(\boldsymbol {n} _ {\boldsymbol {x}}\right) _ {z}}\right) \right\| ^ {2}. \tag {14} +$$ + +Constraints from lighting fixture configurations. To further resolve this ambiguity, we leverage a portable lighting system with a known 3D structure undergoing unknown 6-DoF transformations. The system consists of light + +![](images/20c7063fc4f21b3b9552161a5a18e45d8564564300f2cade2a00bdad1604bbc2.jpg) +Figure 5. The lighting configurations employed in our experiments. We built physical prototypes for the "dual ring" and the "trefoil curve" patterns using programmable LEDs, shown both statically (left) and as dynamic trajectories captured with long exposure (right). We also test on the standard "DiLiGenT" pattern. + +![](images/51cbbcc237ea51eed65a0686540afb07cd3b022da0264c56696493a1cd669a66.jpg) + +![](images/5051bb9b9bce01fad6359a87acdbe00756aa6aaf991b439d2f4d502deb4a624b.jpg) + +sources at known positions $\pmb{p}_i$ (for $i = 1,2,\dots,N$ ) in a local reference frame. Assuming the camera is at the origin, the world positions become $\pmb{p}_i' = \pmb{R}\pmb{p}_i + \pmb{t}$ under a rigid transformation with orthogonal matrix $\pmb{R} \in O(3)$ and translation vector $\pmb{t} \in \mathbb{R}^3$ . The light vector for the $i$ -th source is $s_i = u_i p_i'$ with unknown intensity $u_i > 0$ . A GBR transformation matrix $\pmb{G}$ acts on the recovered normals, and an inverse transformation $\pmb{G}^{-\top}$ acts on the light vectors. The observed lighting $s_i'$ is related to the true lighting $s_i$ by $s_i' = \pmb{G}^{-\top}s_i$ . Combining these, we obtain a relationship between the known local positions $\pmb{p}_i$ and the estimated, transformed light vectors $s_i'$ : + +$$ +\boldsymbol {s} _ {i} ^ {\prime} = u _ {i} \boldsymbol {G} ^ {- \top} (\boldsymbol {R p} _ {i} + \boldsymbol {t}). \tag {15} +$$ + +This system of equations constrains the possible transformations. As formally proven in the supplement, these constraints reduce the GBR ambiguity to a discrete set. Specifically, for a set of at least seven light positions in general position (i.e., not all lying on any nontrivial quadric surface of a specified form passing through the origin), under physical constraints, the ambiguity can be resolved up to the classic convex/concave reflection. This eliminates the need for albedo or reflectance priors. An illustration of integrability's role in resolving positional ambiguity in our setup is shown in Fig. 4. + +Lighting configurations. For event cameras that asynchronously detect brightness changes with high temporal resolution, we sequentially activate each light (e.g., on millisecond timescales) to generate an event stream, with timestamps correlating to lighting changes. We designed two physical patterns (see Fig. 5) that satisfy the geometric requirements: a coaxial dual-ring consisting of LEDs arranged on two concentric rings of different radii, separated + +by a known axial distance; and a trefoil curve with a moving ring of lights along a non-planar, asymmetric 3D trefoil knot trajectory. Both configurations meet the conditions for ambiguity resolution. Details of the theorem, proof, and lighting configurations are provided in the supplement. + +# 3.4. Optimization + +We solve for normals and lighting via alternating optimization, minimizing a joint loss function: + +$$ +\mathcal {L} = \mathcal {L} _ {\text {d a t a}} \left(\left\{\tilde {\boldsymbol {n}} _ {\boldsymbol {x}} \right\}, \boldsymbol {S}\right) + \lambda_ {\text {i n t}} \mathcal {L} _ {\text {i n t}} \left(\left\{\boldsymbol {n} _ {\boldsymbol {x}} \right\}\right), \tag {16} +$$ + +where the weight $\lambda_{\mathrm{int}}$ balances data fidelity Eq. (13) against surface smoothness Eq. (14). The optimization proceeds by alternating between two main steps until convergence: + +- Normal update: With the lighting control points $S$ fixed, the problem decouples per pixel. We solve for each augmented normal $\tilde{n}_x$ from its observation matrix $\tilde{Z}_x\tilde{Z}_x^\top$ . The integrability term is also incorporated. +- Lighting update: With the normals $\{\tilde{n}_x\}$ fixed, the data term $\mathcal{L}_{\mathrm{data}}$ becomes a quadratic function of the lighting control points $S$ . This step reduces to solving a global linear least-squares system to update $S$ . + +We alternate between updating the normals $\{\tilde{n}_x\}$ and the lighting parameters $S$ until convergence. This framework robustly recovers accurate surface normals from event data, even with ambient light and without calibration. + +# 4. Experiment + +# 4.1. Implementation Details + +Our EventUPS implementation uses a hybrid CPU/GPU approach: event pre-processing is GPU-accelerated with OpenCL and Rust bindings, while the PyTorch-based optimization framework handles the core algorithm. For the B-spline lighting representation, we use cubic splines (order $l = 4$ ) with 31 control points for each $360^{\circ}$ rotation cycle unless otherwise specified. The integrability weight is $\lambda_{\mathrm{int}} = 0.1$ for all real data and $\lambda_{\mathrm{int}} = 0.01$ for all synthetic data. + +We developed a custom lighting system using programmable IC LEDs. For the "dual ring" configuration, we mounted concentric LED rings (WS2818B protocol, $800\mathrm{kHz}$ fixed rate) with different radii sharing a common axis. The inner ring contains 40 LEDs, while the outer ring has 60 LEDs. One complete scanning cycle takes approximately 4 seconds. For the "trefoil curve" configuration, we attached an LED ring (APA102, up to $20\mathrm{MHz}$ SPI clock frequency, 24 LEDs) to a rotating mechanism. The LEDs switch at a frequency three times faster than the rotation mechanism, creating a trefoil curve pattern. One complete cycle takes approximately 1/3 second, limited by the rotation speed of the mechanism. + +![](images/db7bcc0bfe49a7187d2a07e06385f813efe2ca2122a05e558a66b5759e0645eb.jpg) +Figure 6. Performance vs. data rate comparison. This plot shows the average MAE of EventUPS against several frame-based methods on the DiLiGenT-Ev dataset as a function of input frames. + +We captured data using a Propheesee EVK4 HD camera with an IMX636 event sensor. The contrast sensitivity threshold biases were set to $-20$ and the dead time bias to $-20$ , resulting in an approximately refractory period of $580~\mu \mathrm{s}$ between consecutive events at a pixel. + +# 4.2. Comparison with Frame-based UPS methods + +Data. To facilitate comparison with frame-based methods, we created a semi-real dataset named DiLiGenT-Ev using images from the DiLiGenT dataset [34]. This procedure, which follows the protocol of EventPS [45], allows for a direct comparison between event-based and frame-based techniques on identical underlying scenes and lighting conditions. Specifically, we selected images captured under the outermost rectangular ring of light directions and converted them to event streams using an event simulator [29]. + +Results. We benchmark EventUPS against state-of-the-art frame-based UPS methods, including the top-performing non-learning method PF14 [28] from the original DiLiGenT benchmark, and several deep learning-based approaches: CW20 [5], CH19 [4], and LL22 [20]. We evaluate accuracy (Mean Angular Error, MAE) and data efficiency. For data rate calculations, we use Total Events $\times$ 16 bits/event (encoded in the Prophesee EVT 3.0 format) for event-based methods, and Frames (lighting directions) $\times$ Resolution $\times$ 8 bits/pixel $\times$ 3 exposures (bracketing for the limited dynamic range) for frame-based methods. + +Fig. 6 illustrates the trade-off between data rate and accuracy for frame-based methods—using more images improves accuracy but increases data requirements. Table 1 provides the detailed quantitative breakdown of comparisons with PF14 and CW20. In sharp contrast, EventUPS (marked by a star) demonstrates remarkable data efficiency. EventUPS achieves an MAE of $14.14^{\circ}$ , a result that is not only superior to the best performance of the non-learning baseline PF14 $(17.16^{\circ}$ MAE) but does so using a data volume equivalent to roughly two frames. Our method operates at just $1.6\%$ of the data rate of the 56-frame deep learning setup and less than $5\%$ of the data rate of PF14. This ef + +![](images/958ad1c7fd193ae4441f54ba85d66730e5ea2deae80aa9123fbb527576c6baca.jpg) +Figure 7. Qualitative results on the DiLiGenT-Ev dataset. Our uncalibrated method produces reconstructions superior to the non-learning, frame-based PF14 [28] and comparable to the calibrated EventPS [45], but without requiring lighting calibration. The "Light" columns visualize the estimated trajectories ( $x, y$ components on the unit sphere), with colors denoting relative intensity. +Table 1. Quantitative comparison (MAE in degrees) on the DiLiGenT-Ev dataset. Our method outperforms the frame-based method PF14 [28] (best-performing non-learning method in the original DiLiGenT dataset [34]) while using less than $5\%$ data rate. We include results from a learning-based method CW20 [5] and the only event-based method EventPS [45] (a calibrated baseline of ours, upper bound) as reference. The last column shows the percentage of data rate compared to CW20 with the best performance. The last row indicates the number (#) of events per round for each data. + +
Uncali-bratedNon-learningInputMethodBALLBUDACATCOWGOLETHARVESTPOT1POT2READERAverageData Rate
✓✘6 framesCW205.3112.109.279.6410.9720.999.197.6113.5010.9511%
✓✘56 framesCW204.9610.237.206.429.4718.557.436.5712.179.22100%
✓✓56 framesPF145.2314.1010.0319.3228.6330.659.8915.7920.8217.1633%
✓✓EventsOurs8.8612.2110.7818.5812.9723.328.7015.1316.7114.141.6%
✘✓EventsEventPS3.4513.227.3223.3313.5429.808.3412.4216.1414.171.6%
# Events470 k306 k354 k458 k200 k327 k246 k334 k393 k343 k/
+ +ficiency stems from the event camera's inherent ability to capture only meaningful brightness changes. + +Fig. 7 shows the visual comparison between different non-learning methods, that is, the frame-based UPS method PF14, the proposed EventUPS, and the calibrated event-based method EventPS. EventUPS reconstructs clean and detailed normals that closely match the ground truth, and accurately recovers the circular lighting trajectory. The normals from PF14 appear visibly noisier and less defined, especially for the challenging HARVEST object, aligning with the error metrics in the table. This demonstrates our method's ability to robustly disentangle geometry, lighting, and non-ideal effects from sparse event data alone. EventUPS's superiority over non-learning, frame-based PF14 shows particular strength on objects with complex peculiarities and textures like HARVEST (23.32° vs. 30.65°) and READER (16.71° vs. 20.82°). Most notably, our uncalibrated method's performance is on par with, and even slightly surpasses, the average performance of the fully cal + +ibrated event-based method EventPS (14.17° MAE). For objects like HARVEST and BUDDHA, our method is more accurate. We attribute this to our model's explicit residual term, which effectively isolates and accounts for non-ideal, real-world effects like ambient illumination. + +# 4.3. Evaluation on Real Camera + +Data. To demonstrate the practical applicability of our method, we fabricated five physical objects with ground truth normal maps and captured real events using our custom lighting systems. These data have varying characteristics: simple geometry (BALL), spatially-varying albedo (BALLICCV), moderate geometric details (BUNNY), and complex details (HORSE, TIGER). + +Results. Fig. 8 shows results on our captured real dataset, which presents additional challenges including sensor noise, non-idealities in event generation. Our method produces high-quality normal reconstructions with clear + +![](images/b5f50a3b4db4cd2b8af8a548a1351b6b8e885e734861507b7b46ad50ca48eff1.jpg) +Figure 8. Results on real data. EventUPS successfully recovers detailed normals despite sensor noise and non-ideal lighting (light trajectories are visualized as in Fig. 1). + +preservation of fine details, achieving an average MAE of $16.58^{\circ}$ across all objects. The spherical object BALLICCV with "ICCV" text highlights an important property: the estimated normal map is invariant to ambient-albedo ratio variations, as the text is barely visible in both the event and the reconstructed normal map. The BUNNY and TIGER objects demonstrate EventUPS's ability to recover intricate geometric details. We observe slightly higher angular errors near object boundaries, which we attribute to near-light effects (the distance between the big ring light and the object is approximately $10\mathrm{cm}$ ) and imperfect lighting alignment. Nevertheless, the overall quality of reconstruction is high. The estimated lighting configurations accurately reflect the dual ring structure used during capture. + +Robustness to lighting patterns. To evaluate the robustness of our method to different lighting configurations, we captured the same objects using both the "dual ring" and "trefoil curve" lighting patterns. As shown in Fig. 9, EventUPS produces consistent normal reconstructions regardless of the lighting pattern used. This demonstrates that our approach effectively recovers both surface normals and lighting directions without requiring a specific lighting arrangement, providing flexibility for different practical scenarios. The trefoil pattern yields better results for objects with complex geometry due to its more varied lighting directions, while the dual ring configuration provides more uniform + +![](images/c664ac65b1101568e15329ab71443f7e5f53bd3fd0a1ab01c4796c8f343d8e1e.jpg) +Figure 9. Performance across illumination patterns. EventUPS delivers consistent normal and lighting estimates for both the (a) dual-ring and (b) trefoil curve patterns, showcasing its robustness. + +coverage for smoother objects. + +# 5. Conclusion and Discussion + +We have presented EventUPS, the first method to solve UPS using an event camera. By introducing a novel augmented null space formulation, a continuous lighting model, and leveraging known lighting fixture geometry, we resolve the fundamental challenges of applying UPS to the event domain. Our experiments show that EventUPS achieves accuracy comparable to or better than its traditional frame-based counterpart while requiring only a fraction (5%) of the data bandwidth. Our method is robust to ambient illumination and different lighting configurations, making it practical for deployment in less controlled environments. + +While our approach represents an advance in uncalibrated event-based PS, several avenues for future work remain. First, extending our framework to handle non-Lambertian surfaces would enhance its applicability in real-world scenarios with complex materials. Second, integrating learning-based components could further improve robustness to noise and outliers while maintaining the physical interpretability of our current model. Finally, exploring applications in dynamic scene reconstruction, where both the camera and objects may be in motion, represents an exciting direction for future research. + +# Acknowledgment + +This work is supported by National Natural Science Foundation of China (Grant Nos. 62136001, 62302019, 624B2006, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources. + +# References + +[1] Neil G. Alldrin, Satya P. Mallick, and David J. Kriegman. Resolving the Generalized Bas-Relief Ambiguity by Entropy Minimization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2 +[2] Peter N. Belhumeur, David J. Kriegman, and Alan L. Yuille. The Bas-Relief Ambiguity. Int. J. Comput. Vis., 35(1):33-44, 1999. 2, 5 +[3] M.K. Chandraker, F. Kahl, and D.J. Kriegman. Reflections on the generalized bas-relief ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2005. 2 +[4] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. Self-Calibrating Deep Photometric Stereo Networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2, 6 +[5] Guanying Chen, Michael Waechter, Boxin Shi, Kwan-Yee K. Wong, and Yasuyuki Matsushita. What Is Learned in Deep Uncalibrated Photometric Stereo? In Proc. of Eur. Conf. Comput. Vis., 2020. 2, 6, 7, 5 +[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor Lighting Estimation Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[7] Jason Chui, Simon Klenk, and Daniel Cremers. Event-Based Feature Tracking in Continuous Time with Sliding Window Optimization. arxiv preprint arxiv:2107.04536, 2021. 2 +[8] O. Drbohlav and M. Chaniler. Can two specular pixels calibrate photometric stereo? In Proc. of Int. Conf. Comput. Vis., 2005. 2 +[9] Ondrej Drbohlav and Radim Šára. Specularities Reduce Ambiguity of Uncalibrated Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2002. 2 +[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1 +[11] Georgiades. Incorporating the Torrance and Sparrow model of reflectance in uncalibrated photometric stereo. In Proc. of Int. Conf. Comput. Vis., 2003. 2 +[12] Heng Guo, Fumio Okura, Boxin Shi, Takuya Funatomi, Yasuhiro Mukaigawa, and Yasuyuki Matsushita. Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3 +[13] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-Fidelity Event-Radiance Recovery Via Transient Event Frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[14] Hideki Hayakawa. Photometric stereo under a light source with arbitrary motion. JOSA A, 11:3079-3089, 1994. 1, 2, 3 +[15] Micah K Johnson and Edward H Adelson. Shape estimation in natural illumination. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2011. 2, 4 +[16] Yakun Ju, Xinghui Dong, Yingyu Wang, Lin Qi, and Junyu Dong. A dual-cue network for multispectral photometric stereo. Pattern Recogn., 2020. 3 + +[17] Dongyeop Kang, Yu Jin Jang, and Sangchul Won. Development of an inspection system for planar steel surface using multispectral photometric stereo. Optic. Eng., 2013. 3 +[18] Hyeongwoo Kim, Bennett Wilburn, and Moshe Ben-Ezra. Photometric Stereo for Dynamic Surface Orientations. In Proc. of Eur. Conf. Comput. Vis., 2010. 2 +[19] Leonid L Kontsevich, AP Petrov, and IS Vergelskaya. Reconstruction of shape from shading in color images. J. Opt. Soc. Am. A, 1994. 3 +[20] Junxuan Li and Hongdong Li. Self-calibrating Photometric Stereo by Neural Inverse Rendering. In Proc. of Eur. Conf. Comput. Vis., 2022, 2, 6 +[21] Feng Lu, Yasuyuki Matsushita, Imari Sato, Takahiro Okabe, and Yoichi Sato. Uncalibrated Photometric Stereo for Unknown Isotropic Reflectances. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2 +[22] Tom Malzbender, Bennett Wilburn, Dan Gelb, and Bill Ambrisco. Surface Enhancement Using Real-Time Photometric Stereo and reflectance transformation. In Proc. of Eurographics Symp. Render. Tech., 2006. 2 +[23] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. MC3D: Motion contrast 3D scanning. In Proc. of Int. Conf. Comput. Photogr., 2015. 3 +[24] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. ESL: Event-based structured light. In Proc. of Int. Conf. 3D Vis., 2021. 3 +[25] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3 +[26] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE Comput. Graphics Appl., 2014. 3 +[27] Thoma Papadhimitri and Paolo Favaro. A New Perspective on Uncalibrated Photometric Stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2 +[28] Thoma Papadhinitri and Paolo Favaro. A Closed-Form, Consistent and Robust Solution to Uncalibrated Photometric Stereo Via Local Diffuse Reflectance Maxima. Int. J. Comput. Vis., 107(2):139-154, 2014. 2, 6, 7, 5 +[29] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In Proc. of Conf. on Robotics Learning, 2018. 6, 4 +[30] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. DiLiGenT10 $^2$ : A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2 +[31] Wonjeong Ryoo, Giljoo Nam, Jae-Sang Hyun, and Sangpil Kim. Event fusion photometric stereo network. *Neural Networks*, 2023. 3 +[32] Boxin Shi, Yasuyuki Matsushita, Yichen Wei, Chao Xu, and Ping Tan. Self-calibrating photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2010. 2 +[33] Boxin Shi, Zhe Wu, Zhipeng Mo, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3 + +[34] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2, 6, 7 +[35] Melvyn L. Smith and Lyndon N. Smith. Dynamic photometric stereo - a new technique for moving surface analysis. Image Vision Comput., 2005. 3 +[36] Kalyan Sunkavalli, Todd Zickler, and Hanspeter Pfister. Visibility Subspaces: Uncalibrated Photometric Stereo with Shadows. In Proc. of Eur. Conf. Comput. Vis., 2010. 2 +[37] Tsuyoshi Takatani, Yasuyuki Matsushita, Stephen Lin, Yasuhiro Mukaigawa, and Yasushi Yagi. Enhanced photometric stereo with multispectral images. In Proc. of Int. Conf. on Mach. Vis. Appl., 2013. 3 +[38] Ping Tan, Satya P. Mallick, Long Quan, David J. Kriegman, and Todd Zickler. Isotropy, Reciprocity and the Generalized Bas-Relief Ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2 +[39] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and imu for robust visual slam in HDR and high-speed scenarios. IEEE Robot. Autom. Lett., 2018. 2 +[40] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2 +[41] Feishi Wang, Jieji Ren, Heng Guo, Mingjun Ren, and Boxin Shi. DiLiGenT-Pi: Photometric stereo for planar surfaces with rich details-benchmark dataset and beyond. In Proc. of Int. Conf. Comput. Vis., 2023. 2 +[42] Robert J. Woodham. Photometric method for determining surface orientation from multiple images. Opt. Eng., 1980, 1, 2 +[43] Zhe Wu and Ping Tan. Calibrating Photometric Stereo by Holistic Reflectance Symmetry Analysis. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2 +[44] Takuya Yoda, Hajime Nagahara, Rin-ichiro Taniguchi, Kei-ichiro Kagawa, Keita Yasutomi, and Shoji Kawahito. The dynamic photometric stereo method using a multi-tap cmos image sensor. Sensors-basel., 2018. 2 +[45] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-Time Photometric Stereo Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 5 +[46] A.L. Yuille, D. Snow, R. Epstein, and P.N. Belhumeur. Determining Generative Models of Objects Under Varying Illumination: Shape and Albedo from Multiple Images Using SVD and Integrability. Int. J. Comput. Vis., 35(3):203-222, 1999. 2, 3, 5 +[47] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided Direct and Global Components Separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3 +[48] Zhenglong Zhou and Ping Tan. Ring-Light Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2010. 2 \ No newline at end of file diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/images.zip b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..b1e9bd40afde68a4c5fe85db76fc9b3eafd9917e --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:24a5efb93170fe610962f8d0a5dfedb7b0f353a99f5d843c63186b48c31d4316 +size 555111 diff --git a/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/layout.json b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3662913c3acdea1688c26c4f14ee40f2207f83be --- /dev/null +++ b/2025/EventUPS_ Uncalibrated Photometric Stereo Using an Event Camera/layout.json @@ -0,0 +1,9456 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "spans": [ + { + "bbox": [ + 92, + 103, + 518, + 121 + ], + "type": "text", + "content": "EventUPS: Uncalibrated Photometric Stereo Using an Event Camera" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "spans": [ + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": "Jinxiu Liang" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,3\\#}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Bohan Yu" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1,2\\#}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Siqi Yang" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1,2,4}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Haotian Zhuang" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Jieji Ren" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Peiqi Duan" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "text", + "content": " Boxin Shi" + }, + { + "bbox": [ + 115, + 141, + 493, + 171 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 172, + 506, + 199 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 104, + 172, + 506, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 172, + 506, + 186 + ], + "spans": [ + { + "bbox": [ + 104, + 172, + 506, + 186 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 104, + 172, + 506, + 186 + ], + "type": "text", + "content": " State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 104, + 186, + 506, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 186, + 506, + 199 + ], + "spans": [ + { + "bbox": [ + 104, + 186, + 506, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 104, + 186, + 506, + 199 + ], + "type": "text", + "content": " National Engineering Research Center of Visual Technology, School of Computer Science, Peking University" + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "spans": [ + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "content": " National Institute of Informatics " + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 140, + 200, + 469, + 213 + ], + "type": "text", + "content": " Institute for Artificial Intelligence, Peking University" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 204, + 214, + 406, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 204, + 214, + 406, + 228 + ], + "spans": [ + { + "bbox": [ + 204, + 214, + 406, + 228 + ], + "type": "text", + "content": "5 Tsinghua University 6 Shanghai Jiaotong University" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 108, + 229, + 500, + 241 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 229, + 500, + 241 + ], + "spans": [ + { + "bbox": [ + 108, + 229, + 500, + 241 + ], + "type": "text", + "content": "cssherryliang@gmail.com,{ybhl998,yousiki,duanqi0001,shiboxin}@pku.edu.cn," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 172, + 243, + 432, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 243, + 432, + 255 + ], + "spans": [ + { + "bbox": [ + 172, + 243, + 432, + 255 + ], + "type": "text", + "content": "zht23@mails.tsinghua.edu.cn, jiejiren@sjtu.edu.cn" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "spans": [ + { + "bbox": [ + 151, + 282, + 200, + 295 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 308, + 296, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 308, + 296, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 296, + 548 + ], + "type": "text", + "content": "We present EventUPS, the first uncalibrated photometric stereo (UPS) method using an event camera—a neuromorphic sensor that asynchronously detects brightness changes with microsecond resolution. Traditional frame-based UPS methods are hindered by high bandwidth demands and limited use in dynamic scenes. These methods require dense image correspondence under varying illumination and are incompatible with the fundamentally different sensing paradigm of event data. Our approach introduces three key innovations: an augmented null space formulation that directly relates each event to joint constraints on surface normals and lighting, naturally handling ambient illumination; a continuous parameterization of time-varying illumination that connects asynchronous events to synchronized lighting estimation; and a lighting fixture with known relative geometry that reduces ambiguity to a convex-concave uncertainty. We validate EventUPS using a custom-built LED lighting system. Experimental results show that our method achieves accuracy surpassing its frame-based counterpart while requiring only " + }, + { + "bbox": [ + 55, + 308, + 296, + 548 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 308, + 296, + 548 + ], + "type": "text", + "content": " of the data bandwidth." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "spans": [ + { + "bbox": [ + 55, + 568, + 135, + 580 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 295, + 696 + ], + "type": "text", + "content": "Photometric stereo (PS) recovers shape from images under varying illumination [14, 42]. For decades, it has been a cornerstone for high-fidelity 3D reconstruction, enabling detailed surface normal recovery for applications from cultural heritage to industrial inspection. The uncalibrated variant of PS (UPS)—where lighting directions are unknown—offers particular practical appeal by eliminating the need for precise calibration apparatus, facilitating deployment in uncontrolled environments. However, this flex" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 319, + 284, + 433, + 415 + ], + "blocks": [ + { + "bbox": [ + 319, + 284, + 433, + 415 + ], + "lines": [ + { + "bbox": [ + 319, + 284, + 433, + 415 + ], + "spans": [ + { + "bbox": [ + 319, + 284, + 433, + 415 + ], + "type": "image", + "image_path": "8c8e1aa3ad56055e81c2442bed480b9badb7735ed929e17df868a34836401bf3.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 358, + 420, + 405, + 430 + ], + "lines": [ + { + "bbox": [ + 358, + 420, + 405, + 430 + ], + "spans": [ + { + "bbox": [ + 358, + 420, + 405, + 430 + ], + "type": "text", + "content": "(a) Our Setup" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 440, + 284, + 547, + 347 + ], + "blocks": [ + { + "bbox": [ + 440, + 284, + 547, + 347 + ], + "lines": [ + { + "bbox": [ + 440, + 284, + 547, + 347 + ], + "spans": [ + { + "bbox": [ + 440, + 284, + 547, + 347 + ], + "type": "image", + "image_path": "8665700ed919e6d969f51d133ae577ead7a336e4a1905dfb03347e5d8b1da8b2.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 349, + 515, + 357 + ], + "lines": [ + { + "bbox": [ + 463, + 349, + 515, + 357 + ], + "spans": [ + { + "bbox": [ + 463, + 349, + 515, + 357 + ], + "type": "text", + "content": "(c) Our Results" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 441, + 359, + 547, + 420 + ], + "blocks": [ + { + "bbox": [ + 441, + 359, + 547, + 420 + ], + "lines": [ + { + "bbox": [ + 441, + 359, + 547, + 420 + ], + "spans": [ + { + "bbox": [ + 441, + 359, + 547, + 420 + ], + "type": "image", + "image_path": "d0e6dcd1faceb1911e8d55be8ad528db4faa8cf1057a411a1041f4d7e128fddd.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 458, + 420, + 525, + 430 + ], + "lines": [ + { + "bbox": [ + 458, + 420, + 525, + 430 + ], + "spans": [ + { + "bbox": [ + 458, + 420, + 525, + 430 + ], + "type": "text", + "content": "(b) Captured events" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 441, + 555, + 519 + ], + "lines": [ + { + "bbox": [ + 313, + 441, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 555, + 519 + ], + "type": "text", + "content": "Figure 1. We propose an uncalibrated photometric stereo method designed for event cameras, EventUPS. Our approach employs (a) a portable lighting system with known relative geometry (e.g., dual LED rings) but an unknown global pose, removing the need for precise calibration. From (b) sparse, asynchronous events triggered by sequentially activated lights, it jointly recovers (c) dense surface normals and the time-varying lighting trajectory1." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 601 + ], + "type": "text", + "content": "ibility incurs significant costs: conventional uncalibrated PS methods demand dense correspondences across multiple images under varying illumination, imposing high data bandwidth and rendering them ill-suited for dynamic scenes or resource-constrained systems." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 601, + 556, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 601, + 556, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 601, + 556, + 673 + ], + "type": "text", + "content": "Meanwhile, event cameras represent a paradigm shift in visual sensing [10]. Instead of capturing intensity frames at fixed intervals, these bio-inspired sensors asynchronously detect per-pixel logarithmic brightness changes. This provides microsecond temporal resolution, an exceptional dynamic range (140dB vs. 60dB), and drastically reduced data" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 702, + 217, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 702, + 217, + 713 + ], + "spans": [ + { + "bbox": [ + 63, + 702, + 217, + 713 + ], + "type": "text", + "content": "Equal contributions. *Corresponding author." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": "Est. Lights: The estimated time-varying light trajectory " + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "inline_equation", + "content": "(x, y" + }, + { + "bbox": [ + 313, + 683, + 555, + 713 + ], + "type": "text", + "content": " components on the unit sphere), with numbers indicating temporal sequence, colors denoting relative intensity (yellow higher, green lower)." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7516" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 108 + ], + "type": "text", + "content": "rates—all critical for platforms like mobile robots. These unique attributes have transformed high-speed tasks such as SLAM [39] and object tracking [7]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 289 + ], + "type": "text", + "content": "Recent advances in event-based PS—exemplified by the EventPS [45]—have shown promise by leveraging these unique capabilities. By using a calibrated, continuously rotating light source, it computes surface normals directly from brightness change events, naturally accommodating rapid variations in surface orientation even in dynamic scenes. This approach dramatically reduces both latency and data redundancy compared to frame-based methods. However, EventPS [45] requires precise lighting calibration and darkroom conditions, which are challenging outside labs. In many real-world scenarios—such as robotic deployments, motion capture, or industrial settings—maintaining calibrated lighting and negligible ambient illumination becomes impractical due to environmental constraints and operational demands." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 293, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 555 + ], + "type": "text", + "content": "A natural question arises: Can established frame-based UPS approaches adapt directly to event cameras? Our analysis reveals fundamental incompatibilities: i) Inability to handle ambient illumination. Frame-based PS methods often address ambient illumination by capturing a dark frame without designed lighting. Unfortunately, this strategy is not applicable to events as they are generated asynchronously by logarithmic brightness changes—encoding brightness ratios rather than absolute intensities. ii) Absence of synchronous correspondence. Non-learning frame-based UPS [2, 46] typically performs matrix factorization to decompose observed intensities into surface normals and lighting directions, establishing constraints through pixel correspondences under shared illumination conditions. However, when applied to events—which correspond to different lighting directions and occur asynchronously—the establishment of such correspondences becomes challenging. iii) Inapplicability of conventional ambiguity resolution. Existing UPS resolves the remaining ambiguity by leveraging effective priors on albedo characteristics [28], which are not applicable to events since the event generation equation is inherently invariant to albedo." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "In this paper, we present EventUPS—the first uncalibrated photometric stereo method using an event camera. Our approach rethinks PS for the event domain, leveraging the unique properties of event cameras to recover detailed surface geometry while simultaneously estimating the time-varying illumination, all without explicit lighting calibration. The core of our contribution is a theoretical framework that formulates each event as a constraint in the null space of an augmented normal vector, enabling joint recovery of surface normals and lighting directions despite the asynchronous nature of events and the presence of ambient illumination. To address the incompatibilities, we make the following technical contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 553, + 108 + ], + "type": "text", + "content": "- An augmented null-space formulation able to handle ambient illumination by modeling it as an unknown variable within event-based constraints on normals and lighting." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 109, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 144 + ], + "type": "text", + "content": "- A continuous parameterization of the time-varying illumination that establishes correspondence by linking asynchronous events to a shared lighting trajectory." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 179 + ], + "type": "text", + "content": "- A geometric UPS ambiguity resolution technique applicable to event data, which uses a lighting fixture with known relative geometry but an unknown global pose." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": "Along with the proposed EventUPS framework, we develop a joint optimization framework that effectively integrates integrability constraints with our event-based null space formulation. To validate our approach in real scenes, we build a hardware prototype around an LED-based lighting system, which implements the lighting configurations with known geometry of dual ring (see Figure 1) and trefoil curve patterns. Experiments on semi-real and real data demonstrate that our method achieves accuracy outperforming the leading frame-based non-learning UPS [28] while requiring only " + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 180, + 555, + 312 + ], + "type": "text", + "content": " of the data bandwidth." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 322, + 406, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 322, + 406, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 322, + 406, + 335 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 342, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 390 + ], + "type": "text", + "content": "PS and event-based vision have advanced independently for decades. Our work lies at their intersection, specifically addressing the challenges of UPS with event cameras. Here, we review developments in both domains that inform ours." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 397, + 492, + 409 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 492, + 409 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 492, + 409 + ], + "type": "text", + "content": "2.1. Uncalibrated Photometric Stereo" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 554, + 486 + ], + "type": "text", + "content": "PS estimates surface normals from images captured under varying illumination directions [15, 42]. While classical approaches assume known lighting, the uncalibrated setting—where lighting is unknown—presents greater challenges but offers superior practical advantages. For comprehensive reviews, see [30, 33, 34, 41]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 487, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 554, + 665 + ], + "type": "text", + "content": "A fundamental challenge in UPS is the ambiguity in simultaneously recovering lighting and surface normals. This was first identified as a general " + }, + { + "bbox": [ + 313, + 487, + 554, + 665 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 487, + 554, + 665 + ], + "type": "text", + "content": " linear ambiguity [14], later reduced to the 3-parameter generalized bas-relief (GBR) ambiguity by imposing integrability constraints [2, 46]. Resolving the ambiguity typically relies on physical cues such as specular reflections [8, 9, 11, 38, 43], statistical priors on albedo [1, 21, 28, 32], shadows [36], interreflections [3], or perspective effects [27]. The light source configuration is also critical; strategic arrangements like ring lights with multi-view information can provide additional constraints to resolve ambiguity without strong surface assumptions [48]. Recent learning-based methods have shown promising results [4, 5, 20], but they often require extensive training data and can struggle with generalization." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 713 + ], + "type": "text", + "content": "Accelerating PS acquisition remains an active research area, with solutions like high-speed cameras [18, 22, 40] or multi-tap CMOS [44] with synchronized illumination, though these dramatically increase bandwidth. Single-shot" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7517" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 120 + ], + "type": "text", + "content": "multi-spectral systems [19, 26, 35, 37] offer an alternative but face challenges like limited spectral bands, crosstalk, and intensity inconsistencies across bands introduce additional challenges [12, 16, 17]." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 125, + 279, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 125, + 279, + 138 + ], + "spans": [ + { + "bbox": [ + 55, + 125, + 279, + 138 + ], + "type": "text", + "content": "2.2. Photometric Methods with Event Cameras" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 143, + 296, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 143, + 296, + 323 + ], + "spans": [ + { + "bbox": [ + 55, + 143, + 296, + 323 + ], + "type": "text", + "content": "With their high temporal resolution, wide dynamic range, and asynchronous operation, event cameras are powerful tools for photometric measurements. These advantages make them particularly suitable for capturing dynamic lighting effects crucial for photometric applications. Their microsecond resolution is especially valuable for active lighting techniques. MC3D [23] pioneered event-based structured light for high-speed 3D scanning, while ESL [24] improved noise robustness by exploiting spatial correlations. Researchers have explored diverse illumination patterns for specific tasks, such as capturing events during light activation to reconstruct iso-contours [13] and address intensity-distance ambiguity by using a diffuse sphere [6], estimating surface normals using rotating polarizers [25], and separating direct/global illumination components [47]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 323, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 468 + ], + "type": "text", + "content": "Most relevant to our work, EFPS-Net [31] fused sparse events with RGB images for normal estimation, while EventPS [45] pioneered purely event-based PS. EventPS showed that the differential nature of event cameras aligns perfectly with PS requirements, but it assumes calibrated lighting and cannot handle ambient illumination. The sparse, asynchronous nature of event data presents unique challenges in uncalibrated settings, particularly in establishing pixel correspondences under consistent lighting. This fundamentally different sensing paradigm makes direct application of non-learning UPS methods [14, 46] infeasible. Our work, EventUPS, is the first to solve these challenges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 475, + 111, + 487 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 475, + 111, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 475, + 111, + 487 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 495, + 180, + 506 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 495, + 180, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 495, + 180, + 506 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "spans": [ + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": "Image formulation. We consider a scene captured by an orthographic projection camera2. The surface is assumed to be Lambertian and is illuminated by a combination of a distant, time-varying directional light source and constant ambient illumination. For a pixel " + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": " in the image plane " + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": ", the reflected radiance " + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "inline_equation", + "content": "I_{\\pmb{x}}(t)" + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": " at time " + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 512, + 296, + 584 + ], + "type": "text", + "content": " is given by:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 102, + 589, + 295, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 589, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 102, + 589, + 295, + 602 + ], + "type": "interline_equation", + "content": "I _ {\\boldsymbol {x}} (t) = \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t) \\right] + b _ {\\boldsymbol {x}}, \\tag {1}", + "image_path": "440b6fb20e8def99f1ee4c600026a9866890be006a62c8fffe2ca1636fce8cf7.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "spans": [ + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\pmb{n}_{\\pmb{x}} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": " is the unit surface normal (" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\| \\pmb{n}_{\\pmb{x}} \\|_{2} = 1" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "a_{x} > 0" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": " is the diffuse albedo, and " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "b_{x} \\geq 0" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": " is the ambient illumination term. The lighting is described by the vector " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\pmb{s}(t) \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": " with intensity " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\| \\pmb{s}(t) \\|^{2}" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": " and unit direction " + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "inline_equation", + "content": "\\hat{\\pmb{s}}(t) = \\pmb{s}(t) / \\| \\pmb{s}(t) \\|" + }, + { + "bbox": [ + 55, + 606, + 296, + 678 + ], + "type": "text", + "content": ". The max operator models attached shadows, where the dot product becomes negative." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 552, + 190 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 552, + 190 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 552, + 190 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 552, + 190 + ], + "type": "image", + "image_path": "078f49e4491f40c7ab347cc4ce11d522b3f79b9a17527a8320725f73e0ea33d8.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 198, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 198, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 555, + 242 + ], + "type": "text", + "content": "Figure 2. The null space vectors and estimated normals: (a) Lambertian sphere under ideal conditions, (b) non-Lambertian sphere with real events. EventUPS models ambient and albedo effects via a residual term, enhancing robustness." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "spans": [ + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": "Event generation. Event cameras are bio-inspired sensors that asynchronously report logarithmic changes in brightness. An event " + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "inline_equation", + "content": "(\\pmb{x}, t_{\\pmb{x}}^{(k)}, p_{\\pmb{x}}^{(k)})" + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": " is triggered at pixel " + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": " when the logarithmic change in brightness exceeds a predetermined contrast threshold " + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "inline_equation", + "content": "\\theta > 0" + }, + { + "bbox": [ + 313, + 261, + 555, + 321 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 349, + 327, + 555, + 342 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 327, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 349, + 327, + 555, + 342 + ], + "type": "interline_equation", + "content": "\\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right)\\right) - \\log \\left(I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right)\\right) \\geq p _ {\\boldsymbol {x}} ^ {(k)} \\theta , \\tag {2}", + "image_path": "116d88571fac4f1bdf93706284d41d3e9d4733845df625c39222639206406c33.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "inline_equation", + "content": "t_{\\pmb{x}}^{(k)}" + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "content": " denotes the timestamp of the " + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "content": "-th event at pixel " + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "inline_equation", + "content": "p_{\\pmb{x}}^{(k)} \\in \\{+1, -1\\}" + }, + { + "bbox": [ + 313, + 349, + 554, + 400 + ], + "type": "text", + "content": " indicates whether the brightness increased or decreased. The relationship between consecutive brightness measurements can be approximated as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 359, + 405, + 554, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 405, + 554, + 421 + ], + "spans": [ + { + "bbox": [ + 359, + 405, + 554, + 421 + ], + "type": "interline_equation", + "content": "I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\approx I _ {\\boldsymbol {x}} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right) \\cdot \\exp \\left(p _ {\\boldsymbol {x}} ^ {(k)} \\theta\\right). \\tag {3}", + "image_path": "03c11a90385fc7164123c9ce34b650c4aa311a222555e59a03ef9afff97ea734.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 427, + 537, + 442 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 427, + 537, + 442 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 537, + 442 + ], + "type": "text", + "content": "For notational simplicity, we define " + }, + { + "bbox": [ + 313, + 427, + 537, + 442 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_{\\pmb{x}}^{(k)} = \\exp (p_{\\pmb{x}}^{(k)}\\theta)" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "content": "UPS with events. Unlike prior event-based works [45] that assume calibrated lighting, we aim to jointly estimate the per-pixel surface normals " + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "inline_equation", + "content": "\\pmb{n}_{\\pmb{x}}" + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "content": " and the continuous, time-varying lighting function " + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "content": " from only the event stream " + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "inline_equation", + "content": "\\mathbb{E} = \\{(x,t_x^{(k)},p_x^{(k)})\\}_{x\\in \\Omega}" + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "content": " across all pixels " + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "inline_equation", + "content": "x\\in \\Omega" + }, + { + "bbox": [ + 313, + 454, + 555, + 554 + ], + "type": "text", + "content": ". This inverse problem is formulated as the minimization of the joint loss function that measures consistency between the observed events and the image formation model:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 356, + 559, + 555, + 574 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 559, + 555, + 574 + ], + "spans": [ + { + "bbox": [ + 356, + 559, + 555, + 574 + ], + "type": "interline_equation", + "content": "\\operatorname {a r g m i n} _ {\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\} _ {\\boldsymbol {x} \\in \\Omega}, \\boldsymbol {s} (t)} \\mathcal {L} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {s} (t); \\mathbb {E}\\right). \\tag {4}", + "image_path": "9b52328f6ae8e636240e5f1d8235341ae362865008acb9932fd161bce77a0ffa.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 578, + 517, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 578, + 517, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 578, + 517, + 592 + ], + "type": "text", + "content": "3.2. Joint Estimating Normal and Lighting" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 596, + 555, + 669 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 596, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 313, + 596, + 555, + 669 + ], + "type": "text", + "content": "The augmented null-space formulation. Conventional photometric stereo often handles ambient light by subtracting a dark frame. This is infeasible for event-based methods where measurements are relative and asynchronous. We propose an alternative by substituting the image model Eq. (1) into the event model Eq. (3):" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 345, + 673, + 523, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 345, + 673, + 523, + 716 + ], + "spans": [ + { + "bbox": [ + 345, + 673, + 523, + 716 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k)}) \\right] + b _ {\\boldsymbol {x}} \\\\ \\approx \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} \\cdot (\\max \\left[ 0, a _ {\\boldsymbol {x}} \\boldsymbol {n} _ {\\boldsymbol {x}} \\cdot \\boldsymbol {s} (t _ {\\boldsymbol {x}} ^ {(k - 1)}) \\right] + b _ {\\boldsymbol {x}}). \\\\ \\end{array}", + "image_path": "23b65d3d4a32968273896ead3689be01251e255edfa509f851ee7490096eda8b.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": "Orthographic projection is a widely used assumption in PS [33] when the object-to-camera distance (approximately " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "20~\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": " in our setup) is significantly larger than object depth variations (approximately " + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "inline_equation", + "content": "3\\mathrm{cm}" + }, + { + "bbox": [ + 55, + 683, + 296, + 714 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7518" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "For most observed events, the surface point is illuminated at both timestamps (i.e., it is not in an attached shadow). This allows us to remove the max operator. The albedo " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "a_{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " can be eliminated by dividing it out, as it appears on both sides. This makes our formulation inherently albedo-invariant. Rearranging the terms yields a linear constraint on the normal " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "n_{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " and the \"residual\" pixel-wise constant ambient-albedo ratio " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "r_{\\pmb{x}} = b_{\\pmb{x}} / a_{\\pmb{x}} \\geq 0" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 65, + 177, + 295, + 194 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 177, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 65, + 177, + 295, + 194 + ], + "type": "interline_equation", + "content": "\\pmb {n} _ {\\pmb {x}} \\cdot (\\pmb {s} (t _ {\\pmb {x}} ^ {(k)}) - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)} \\pmb {s} (t _ {\\pmb {x}} ^ {(k - 1)})) + r _ {\\pmb {x}} (1 - \\tilde {\\theta} _ {\\pmb {x}} ^ {(k)}) = 0. \\quad (6)", + "image_path": "45b8f163649b57615226d7fb6f2b78f066cb7ead8ee6aecf701fecb971ebf733.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 203, + 296, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 203, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 203, + 296, + 262 + ], + "type": "text", + "content": "Shadow events, which occur when one of the terms under the max operator becomes zero, are relatively rare and can be identified through temporal consistency checks. When they do occur, they provide constraints at shadow boundaries that can be incorporated into our framework." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "spans": [ + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\boldsymbol{z}_{\\boldsymbol{x}}^{(k)} = \\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k)}) - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)}\\boldsymbol{s}(t_{\\boldsymbol{x}}^{(k-1)})" + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "content": ". We define an augmented normal vector " + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{n}}_{\\boldsymbol{x}} = [\\boldsymbol{n}_{\\boldsymbol{x}}^{\\top}, r_{\\boldsymbol{x}}]^{\\top} \\in \\mathbb{R}^{4}" + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "content": ", which is constant for each pixel " + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\boldsymbol{x}" + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "content": ". For each event, we define a corresponding augmented null space vector " + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "inline_equation", + "content": "\\tilde{\\boldsymbol{z}}_{\\boldsymbol{x}}^{(k)} = [\\boldsymbol{z}_{\\boldsymbol{x}}^{(k)\\top}, (1 - \\tilde{\\theta}_{\\boldsymbol{x}}^{(k)})]^{\\top}" + }, + { + "bbox": [ + 55, + 262, + 296, + 329 + ], + "type": "text", + "content": ". This recasts Eq. (6) into an geometric relationship:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 145, + 341, + 295, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 341, + 295, + 357 + ], + "spans": [ + { + "bbox": [ + 145, + 341, + 295, + 357 + ], + "type": "interline_equation", + "content": "\\tilde {z} _ {\\boldsymbol {x}} ^ {(k) \\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} = 0. \\tag {7}", + "image_path": "8f5c6d2bada47af9c696c4fd3b066997ec7c22b00ed94c3e46b539be0c36f0fc.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": "This means the augmented normal " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{n}_{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " must lie in the null space of the vectors " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\{\\tilde{z}_x^{(k)}\\}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": ". Please see Fig. 2 for an illustration. With at least three linearly independent null space vectors from events at a pixel, we can determine the augmented normal " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{n}_{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " up to a scale factor. By stacking all augmented null space vectors " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{z}_x^{(k)}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " for a pixel " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " into a matrix " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "Z_{\\pmb{x}} \\in \\mathbb{R}^{4 \\times K_{\\pmb{x}}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "K_{\\pmb{x}}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": " is the number of events at that pixel and the " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": "-th row is " + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "inline_equation", + "content": "\\tilde{z}_x^{(k)\\top}" + }, + { + "bbox": [ + 55, + 363, + 296, + 475 + ], + "type": "text", + "content": ", we solve for the normal by finding the vector that minimizes the projection:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 64, + 483, + 295, + 506 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 483, + 295, + 506 + ], + "spans": [ + { + "bbox": [ + 64, + 483, + 295, + 506 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} ^ {*} = \\underset {\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}}} {\\operatorname {a r g m i n}} \\| Z _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2} \\quad \\text {s u b j e c t t o} \\quad \\| \\boldsymbol {n} _ {\\boldsymbol {x}} \\| _ {2} = 1. \\tag {8}", + "image_path": "cfe532d313db03ac9b3848c615842712bfd09d495f6e73b6afd35910f3ed0131.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 515, + 296, + 574 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 515, + 296, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 515, + 296, + 574 + ], + "type": "text", + "content": "This constrained optimization problem has an elegant solution via Singular Value Decomposition (SVD) of the matrix " + }, + { + "bbox": [ + 55, + 515, + 296, + 574 + ], + "type": "inline_equation", + "content": "Z_{x}Z_{x}^{\\top}\\in \\mathbb{R}^{4\\times 4}" + }, + { + "bbox": [ + 55, + 515, + 296, + 574 + ], + "type": "text", + "content": ". The solution is the eigenvector corresponding to the smallest eigenvalue, appropriately scaled to ensure the unit normal constraint." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "Continuous lighting parameterization. In the uncalibrated setting, the lighting direction " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " is unknown and varies continuously over time. The asynchronous nature of events makes it challenging to establish correspondences across pixels under identical lighting conditions, which is a prerequisite for non-learning uncalibrated PS methods. To address this challenge, we model " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " as a B-spline curve. This provides a smooth, local, and efficient parameterization that is ideal for asynchronous data. By selecting " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " control points " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{s(t_j)\\}_{j=0}^{M-1}" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": " at uniform times " + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "inline_equation", + "content": "t_j = t_0 + j\\Delta t" + }, + { + "bbox": [ + 55, + 594, + 296, + 715 + ], + "type": "text", + "content": "," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 71, + 552, + 247 + ], + "blocks": [ + { + "bbox": [ + 317, + 71, + 552, + 247 + ], + "lines": [ + { + "bbox": [ + 317, + 71, + 552, + 247 + ], + "spans": [ + { + "bbox": [ + 317, + 71, + 552, + 247 + ], + "type": "image", + "image_path": "bd0cbfd15a6849f221219272691ded0ed08711f9cc03010a9d5a6f5c9ec9ff91.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 255, + 555, + 311 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 311 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 311 + ], + "type": "text", + "content": "Figure 3. Continuous illumination model. We represent the time-varying lighting " + }, + { + "bbox": [ + 313, + 255, + 555, + 311 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 255, + 555, + 311 + ], + "type": "text", + "content": " as a B-spline controlled by a set of points. This continuous representation links asynchronous events to a globally synchronized lighting estimate, shown here for a rotating ring pattern projected onto a 2D plane." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 328, + 512, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 328, + 512, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 328, + 512, + 340 + ], + "type": "text", + "content": "the lighting at any time " + }, + { + "bbox": [ + 313, + 328, + 512, + 340 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 328, + 512, + 340 + ], + "type": "text", + "content": " is a linear combination:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 382, + 345, + 554, + 379 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 345, + 554, + 379 + ], + "spans": [ + { + "bbox": [ + 382, + 345, + 554, + 379 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} (t) = \\sum_ {j = 0} ^ {M - 1} C _ {j, l} (t) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {9}", + "image_path": "6827778c56b5ffec4ff52ee88a45368f9c670edf73dca1a13f59f70dc6d9443b.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "\\{C_{j,l}(t)\\}" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": " are the B-spline basis functions of order " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": ". An important property of this formulation is that the value of " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "s(t)" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "t_j \\leq t < t_{j+1}" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": " depends only on " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": " consecutive control points " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "s(t_j), s(t_{j+1}), \\ldots, s(t_{j+l-1})" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": ", providing locality in the temporal domain. This links the null space vector " + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "inline_equation", + "content": "z_x^{(k)}" + }, + { + "bbox": [ + 313, + 384, + 554, + 457 + ], + "type": "text", + "content": " directly to the control points:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 376, + 462, + 554, + 496 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 376, + 462, + 554, + 496 + ], + "spans": [ + { + "bbox": [ + 376, + 462, + 554, + 496 + ], + "type": "interline_equation", + "content": "\\boldsymbol {z} _ {\\boldsymbol {x}} ^ {(k)} = \\sum_ {j = 0} ^ {M - 1} \\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) \\boldsymbol {s} \\left(t _ {j}\\right), \\tag {10}", + "image_path": "0a371c932f7c7b927525f8ac0f2474e22de6e4be57d586522cb33be879737f74.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 501, + 342, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 342, + 510 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 342, + 510 + ], + "type": "text", + "content": "where" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 347, + 515, + 553, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 515, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 347, + 515, + 553, + 532 + ], + "type": "interline_equation", + "content": "\\widetilde {C} _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) = C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k)}\\right) - \\tilde {\\theta} _ {\\boldsymbol {x}} ^ {(k)} C _ {j, l} \\left(t _ {\\boldsymbol {x}} ^ {(k - 1)}\\right), \\tag {11}", + "image_path": "28056e9fad1e5d0e8b3393c9ba5eec132468f819a38f2f9eaf132793cf8e3c31.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": "This expression depends only on the observed event times-stamps " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "t_{\\pmb{x}}^{(k)}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " at pixel " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": ", providing a crucial bridge between pixel-specific asynchronous events and a global, continuous lighting model (see Fig. 3). In matrix form, we stack the control points " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\{s(t_j)\\}_{j = 0}^{M - 1}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " into a matrix " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\pmb {S}\\in \\mathbb{R}^{3\\times M}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " and the time-dependent coefficients " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\widetilde{C}_{j,l}(t_x^{(k)})" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " into a matrix " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "C_x\\in \\mathbb{R}^{M\\times K_x}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": ". The augmented measurement matrix for a pixel " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\pmb{x}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "inline_equation", + "content": "\\tilde{\\theta}_{\\pmb{x}} = [(1 - \\tilde{\\theta}_{\\pmb{x}}^{(1)}),\\dots ,(1 - \\tilde{\\theta}_{\\pmb{x}}^{(K_{\\pmb{x}})})]\\in \\mathbb{R}^{K_{\\pmb{x}}}" + }, + { + "bbox": [ + 313, + 536, + 554, + 647 + ], + "type": "text", + "content": " becomes:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 381, + 647, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 381, + 647, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 381, + 647, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} = \\left[ \\begin{array}{c} \\boldsymbol {S C} _ {\\boldsymbol {x}} \\\\ \\tilde {\\boldsymbol {\\theta}} _ {\\boldsymbol {x}} ^ {\\top} \\end{array} \\right] \\in \\mathbb {R} ^ {4 \\times K _ {\\boldsymbol {x}}}. \\tag {12}", + "image_path": "585b9336232c69f2905074d7d93119bc3f1a1e486477da756de848b63ca22e41.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 675, + 413, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 675, + 413, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 675, + 413, + 685 + ], + "type": "text", + "content": "The data term becomes:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 364, + 690, + 553, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 690, + 553, + 716 + ], + "spans": [ + { + "bbox": [ + 364, + 690, + 553, + 716 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) = \\sum_ {\\boldsymbol {x} \\in \\Omega} \\| \\tilde {\\boldsymbol {Z}} _ {\\boldsymbol {x}} ^ {\\top} \\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\| _ {2} ^ {2}. \\tag {13}", + "image_path": "e010ca48c6ccbd1d46080429d9e74d776e20182c51834928df6c8b7373d53844.jpg" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7519" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 61, + 70, + 292, + 287 + ], + "blocks": [ + { + "bbox": [ + 61, + 70, + 292, + 287 + ], + "lines": [ + { + "bbox": [ + 61, + 70, + 292, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 70, + 292, + 287 + ], + "type": "image", + "image_path": "325400be889e930192c3e7bdbe447f281b9a19e5cb733bc62cdce102cdef7083.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 295, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 365 + ], + "type": "text", + "content": "Figure 4. Resolving Ambiguity. With ground-truth lighting (a), the normal is recovered accurately. Assuming incorrect lighting transformations (b-e), estimating normals without constraints leads to large errors. The strong correlation between normal error and integrability cost validates the effectiveness of enforcing integrability with a known lighting fixture to resolve ambiguity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 55, + 383, + 296, + 407 + ], + "type": "text", + "content": "This bilinear structure naturally lends itself to an alternating optimization scheme." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 178, + 429 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 178, + 429 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 178, + 429 + ], + "type": "text", + "content": "3.3. Resolving Ambiguity" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": "Integrability. UPS suffers from a fundamental ambiguity where multiple combinations of surface normals and lighting directions can produce identical observations. Physical constraints, particularly the integrability of surface normals, require that the normals be derivable from the gradient of an underlying surface. For non-learning frame-based UPS, this constraint reduces the general linear ambiguity to the GBR transformation [2, 46], which maps the surface " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "z(x)" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\lambda z(x) + \\mu x + \\nu y" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\lambda \\neq 0" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": ". For event-based UPS, imposing integrability similarly reduces the ambiguity to the GBR transformation, as the constraint acts on the normal field independently of the observation modality. Specifically, we use an integrability term. It ensures the normal field " + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "inline_equation", + "content": "\\{n_x\\}" + }, + { + "bbox": [ + 55, + 433, + 296, + 613 + ], + "type": "text", + "content": " corresponds to a valid, continuous surface by penalizing the curl of the corresponding gradient field:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 69, + 621, + 296, + 652 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 621, + 296, + 652 + ], + "spans": [ + { + "bbox": [ + 69, + 621, + 296, + 652 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {i n t}} = \\sum_ {\\boldsymbol {x}} \\left\\| \\frac {\\partial}{\\partial y} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {x}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) - \\frac {\\partial}{\\partial x} \\left(\\frac {\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {y}}{\\left(\\boldsymbol {n} _ {\\boldsymbol {x}}\\right) _ {z}}\\right) \\right\\| ^ {2}. \\tag {14}", + "image_path": "11d4f6463445fb663733d9c917fe37e7874c02ad3e3f005e19f4bdd5942917eb.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 715 + ], + "type": "text", + "content": "Constraints from lighting fixture configurations. To further resolve this ambiguity, we leverage a portable lighting system with a known 3D structure undergoing unknown 6-DoF transformations. The system consists of light" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 318, + 70, + 409, + 209 + ], + "blocks": [ + { + "bbox": [ + 318, + 70, + 409, + 209 + ], + "lines": [ + { + "bbox": [ + 318, + 70, + 409, + 209 + ], + "spans": [ + { + "bbox": [ + 318, + 70, + 409, + 209 + ], + "type": "image", + "image_path": "20c7063fc4f21b3b9552161a5a18e45d8564564300f2cade2a00bdad1604bbc2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 220, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 277 + ], + "type": "text", + "content": "Figure 5. The lighting configurations employed in our experiments. We built physical prototypes for the \"dual ring\" and the \"trefoil curve\" patterns using programmable LEDs, shown both statically (left) and as dynamic trajectories captured with long exposure (right). We also test on the standard \"DiLiGenT\" pattern." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 413, + 74, + 479, + 210 + ], + "blocks": [ + { + "bbox": [ + 413, + 74, + 479, + 210 + ], + "lines": [ + { + "bbox": [ + 413, + 74, + 479, + 210 + ], + "spans": [ + { + "bbox": [ + 413, + 74, + 479, + 210 + ], + "type": "image", + "image_path": "51cbbcc237ea51eed65a0686540afb07cd3b022da0264c56696493a1cd669a66.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 483, + 74, + 549, + 210 + ], + "blocks": [ + { + "bbox": [ + 483, + 74, + 549, + 210 + ], + "lines": [ + { + "bbox": [ + 483, + 74, + 549, + 210 + ], + "spans": [ + { + "bbox": [ + 483, + 74, + 549, + 210 + ], + "type": "image", + "image_path": "5051bb9b9bce01fad6359a87acdbe00756aa6aaf991b439d2f4d502deb4a624b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": "sources at known positions " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " (for " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "i = 1,2,\\dots,N" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ") in a local reference frame. Assuming the camera is at the origin, the world positions become " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i' = \\pmb{R}\\pmb{p}_i + \\pmb{t}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " under a rigid transformation with orthogonal matrix " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{R} \\in O(3)" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " and translation vector " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{t} \\in \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ". The light vector for the " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": "-th source is " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "s_i = u_i p_i'" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " with unknown intensity " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "u_i > 0" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ". A GBR transformation matrix " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{G}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " acts on the recovered normals, and an inverse transformation " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{G}^{-\\top}" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " acts on the light vectors. The observed lighting " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "s_i'" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " is related to the true lighting " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " by " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "s_i' = \\pmb{G}^{-\\top}s_i" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ". Combining these, we obtain a relationship between the known local positions " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "\\pmb{p}_i" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": " and the estimated, transformed light vectors " + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "inline_equation", + "content": "s_i'" + }, + { + "bbox": [ + 313, + 297, + 555, + 441 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 383, + 448, + 555, + 463 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 448, + 555, + 463 + ], + "spans": [ + { + "bbox": [ + 383, + 448, + 555, + 463 + ], + "type": "interline_equation", + "content": "\\boldsymbol {s} _ {i} ^ {\\prime} = u _ {i} \\boldsymbol {G} ^ {- \\top} (\\boldsymbol {R p} _ {i} + \\boldsymbol {t}). \\tag {15}", + "image_path": "807ac29d2b4dbf81d1a86767aac10a87c69591fc2129d61236bae1dfdf78a6d4.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 555, + 604 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 604 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 604 + ], + "type": "text", + "content": "This system of equations constrains the possible transformations. As formally proven in the supplement, these constraints reduce the GBR ambiguity to a discrete set. Specifically, for a set of at least seven light positions in general position (i.e., not all lying on any nontrivial quadric surface of a specified form passing through the origin), under physical constraints, the ambiguity can be resolved up to the classic convex/concave reflection. This eliminates the need for albedo or reflectance priors. An illustration of integrability's role in resolving positional ambiguity in our setup is shown in Fig. 4." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "Lighting configurations. For event cameras that asynchronously detect brightness changes with high temporal resolution, we sequentially activate each light (e.g., on millisecond timescales) to generate an event stream, with timestamps correlating to lighting changes. We designed two physical patterns (see Fig. 5) that satisfy the geometric requirements: a coaxial dual-ring consisting of LEDs arranged on two concentric rings of different radii, separated" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7520" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "by a known axial distance; and a trefoil curve with a moving ring of lights along a non-planar, asymmetric 3D trefoil knot trajectory. Both configurations meet the conditions for ambiguity resolution. Details of the theorem, proof, and lighting configurations are provided in the supplement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 141, + 141, + 154 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 141, + 154 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 141, + 154 + ], + "type": "text", + "content": "3.4. Optimization" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 159, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 296, + 183 + ], + "type": "text", + "content": "We solve for normals and lighting via alternating optimization, minimizing a joint loss function:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 96, + 194, + 295, + 208 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 194, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 96, + 194, + 295, + 208 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\text {d a t a}} \\left(\\left\\{\\tilde {\\boldsymbol {n}} _ {\\boldsymbol {x}} \\right\\}, \\boldsymbol {S}\\right) + \\lambda_ {\\text {i n t}} \\mathcal {L} _ {\\text {i n t}} \\left(\\left\\{\\boldsymbol {n} _ {\\boldsymbol {x}} \\right\\}\\right), \\tag {16}", + "image_path": "f6d905f6f6eab76f3b9b6764cd9e2942172a6c01c414153003e756a6d0d7da5a.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 218, + 296, + 254 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 254 + ], + "type": "text", + "content": "where the weight " + }, + { + "bbox": [ + 55, + 218, + 296, + 254 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{int}}" + }, + { + "bbox": [ + 55, + 218, + 296, + 254 + ], + "type": "text", + "content": " balances data fidelity Eq. (13) against surface smoothness Eq. (14). The optimization proceeds by alternating between two main steps until convergence:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 255, + 296, + 350 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "text", + "content": "- Normal update: With the lighting control points " + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "text", + "content": " fixed, the problem decouples per pixel. We solve for each augmented normal " + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tilde{n}_x" + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "text", + "content": " from its observation matrix " + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "inline_equation", + "content": "\\tilde{Z}_x\\tilde{Z}_x^\\top" + }, + { + "bbox": [ + 55, + 255, + 295, + 301 + ], + "type": "text", + "content": ". The integrability term is also incorporated." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "content": "- Lighting update: With the normals " + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\{\\tilde{n}_x\\}" + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "content": " fixed, the data term " + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{data}}" + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "content": " becomes a quadratic function of the lighting control points " + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "content": ". This step reduces to solving a global linear least-squares system to update " + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 56, + 302, + 296, + 350 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": "We alternate between updating the normals " + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "inline_equation", + "content": "\\{\\tilde{n}_x\\}" + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": " and the lighting parameters " + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 350, + 296, + 399 + ], + "type": "text", + "content": " until convergence. This framework robustly recovers accurate surface normals from event data, even with ambient light and without calibration." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 411, + 132, + 425 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 132, + 425 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 132, + 425 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 431, + 188, + 444 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 431, + 188, + 444 + ], + "spans": [ + { + "bbox": [ + 55, + 431, + 188, + 444 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "content": "Our EventUPS implementation uses a hybrid CPU/GPU approach: event pre-processing is GPU-accelerated with OpenCL and Rust bindings, while the PyTorch-based optimization framework handles the core algorithm. For the B-spline lighting representation, we use cubic splines (order " + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "inline_equation", + "content": "l = 4" + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "content": ") with 31 control points for each " + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "content": " rotation cycle unless otherwise specified. The integrability weight is " + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{int}} = 0.1" + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "content": " for all real data and " + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{int}} = 0.01" + }, + { + "bbox": [ + 55, + 449, + 296, + 555 + ], + "type": "text", + "content": " for all synthetic data." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "We developed a custom lighting system using programmable IC LEDs. For the \"dual ring\" configuration, we mounted concentric LED rings (WS2818B protocol, " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "800\\mathrm{kHz}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " fixed rate) with different radii sharing a common axis. The inner ring contains 40 LEDs, while the outer ring has 60 LEDs. One complete scanning cycle takes approximately 4 seconds. For the \"trefoil curve\" configuration, we attached an LED ring (APA102, up to " + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "inline_equation", + "content": "20\\mathrm{MHz}" + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": " SPI clock frequency, 24 LEDs) to a rotating mechanism. The LEDs switch at a frequency three times faster than the rotation mechanism, creating a trefoil curve pattern. One complete cycle takes approximately 1/3 second, limited by the rotation speed of the mechanism." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 339, + 69, + 531, + 167 + ], + "blocks": [ + { + "bbox": [ + 339, + 69, + 531, + 167 + ], + "lines": [ + { + "bbox": [ + 339, + 69, + 531, + 167 + ], + "spans": [ + { + "bbox": [ + 339, + 69, + 531, + 167 + ], + "type": "image", + "image_path": "db7bcc0bfe49a7187d2a07e06385f813efe2ca2122a05e558a66b5759e0645eb.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 554, + 209 + ], + "type": "text", + "content": "Figure 6. Performance vs. data rate comparison. This plot shows the average MAE of EventUPS against several frame-based methods on the DiLiGenT-Ev dataset as a function of input frames." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "text", + "content": "We captured data using a Propheesee EVK4 HD camera with an IMX636 event sensor. The contrast sensitivity threshold biases were set to " + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "inline_equation", + "content": "-20" + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "text", + "content": " and the dead time bias to " + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "inline_equation", + "content": "-20" + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "text", + "content": ", resulting in an approximately refractory period of " + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "inline_equation", + "content": "580~\\mu \\mathrm{s}" + }, + { + "bbox": [ + 313, + 230, + 555, + 290 + ], + "type": "text", + "content": " between consecutive events at a pixel." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 297, + 549, + 310 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 549, + 310 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 549, + 310 + ], + "type": "text", + "content": "4.2. Comparison with Frame-based UPS methods" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 315, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 423 + ], + "type": "text", + "content": "Data. To facilitate comparison with frame-based methods, we created a semi-real dataset named DiLiGenT-Ev using images from the DiLiGenT dataset [34]. This procedure, which follows the protocol of EventPS [45], allows for a direct comparison between event-based and frame-based techniques on identical underlying scenes and lighting conditions. Specifically, we selected images captured under the outermost rectangular ring of light directions and converted them to event streams using an event simulator [29]." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "content": "Results. We benchmark EventUPS against state-of-the-art frame-based UPS methods, including the top-performing non-learning method PF14 [28] from the original DiLiGenT benchmark, and several deep learning-based approaches: CW20 [5], CH19 [4], and LL22 [20]. We evaluate accuracy (Mean Angular Error, MAE) and data efficiency. For data rate calculations, we use Total Events " + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "content": " 16 bits/event (encoded in the Prophesee EVT 3.0 format) for event-based methods, and Frames (lighting directions) " + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "content": " Resolution " + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "content": " 8 bits/pixel " + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "inline_equation", + "content": "\\times" + }, + { + "bbox": [ + 313, + 438, + 554, + 569 + ], + "type": "text", + "content": " 3 exposures (bracketing for the limited dynamic range) for frame-based methods." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": "Fig. 6 illustrates the trade-off between data rate and accuracy for frame-based methods—using more images improves accuracy but increases data requirements. Table 1 provides the detailed quantitative breakdown of comparisons with PF14 and CW20. In sharp contrast, EventUPS (marked by a star) demonstrates remarkable data efficiency. EventUPS achieves an MAE of " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "14.14^{\\circ}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": ", a result that is not only superior to the best performance of the non-learning baseline PF14 " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "(17.16^{\\circ}" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " MAE) but does so using a data volume equivalent to roughly two frames. Our method operates at just " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "1.6\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of the data rate of the 56-frame deep learning setup and less than " + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 570, + 555, + 713 + ], + "type": "text", + "content": " of the data rate of PF14. This ef" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "7521" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 553, + 228 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 553, + 228 + ], + "type": "image", + "image_path": "958ad1c7fd193ae4441f54ba85d66730e5ea2deae80aa9123fbb527576c6baca.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "lines": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "type": "text", + "content": "Figure 7. Qualitative results on the DiLiGenT-Ev dataset. Our uncalibrated method produces reconstructions superior to the non-learning, frame-based PF14 [28] and comparable to the calibrated EventPS [45], but without requiring lighting calibration. The \"Light\" columns visualize the estimated trajectories (" + }, + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 55, + 236, + 555, + 270 + ], + "type": "text", + "content": " components on the unit sphere), with colors denoting relative intensity." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 280, + 555, + 335 + ], + "lines": [ + { + "bbox": [ + 54, + 280, + 555, + 335 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 555, + 335 + ], + "type": "text", + "content": "Table 1. Quantitative comparison (MAE in degrees) on the DiLiGenT-Ev dataset. Our method outperforms the frame-based method PF14 [28] (best-performing non-learning method in the original DiLiGenT dataset [34]) while using less than " + }, + { + "bbox": [ + 54, + 280, + 555, + 335 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 54, + 280, + 555, + 335 + ], + "type": "text", + "content": " data rate. We include results from a learning-based method CW20 [5] and the only event-based method EventPS [45] (a calibrated baseline of ours, upper bound) as reference. The last column shows the percentage of data rate compared to CW20 with the best performance. The last row indicates the number (#) of events per round for each data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 65, + 347, + 547, + 462 + ], + "blocks": [ + { + "bbox": [ + 65, + 347, + 547, + 462 + ], + "lines": [ + { + "bbox": [ + 65, + 347, + 547, + 462 + ], + "spans": [ + { + "bbox": [ + 65, + 347, + 547, + 462 + ], + "type": "table", + "html": "
Uncali-bratedNon-learningInputMethodBALLBUDACATCOWGOLETHARVESTPOT1POT2READERAverageData Rate
✓✘6 framesCW205.3112.109.279.6410.9720.999.197.6113.5010.9511%
✓✘56 framesCW204.9610.237.206.429.4718.557.436.5712.179.22100%
✓✓56 framesPF145.2314.1010.0319.3228.6330.659.8915.7920.8217.1633%
✓✓EventsOurs8.8612.2110.7818.5812.9723.328.7015.1316.7114.141.6%
✘✓EventsEventPS3.4513.227.3223.3313.5429.808.3412.4216.1414.171.6%
# Events470 k306 k354 k458 k200 k327 k246 k334 k393 k343 k/
", + "image_path": "8f54c7e0a05305f6806bc98dad42983bf45e6c8e5e74d1b270c52f0509b5b997.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 483, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 508 + ], + "type": "text", + "content": "ficiency stems from the event camera's inherent ability to capture only meaningful brightness changes." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 510, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 510, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 510, + 295, + 714 + ], + "type": "text", + "content": "Fig. 7 shows the visual comparison between different non-learning methods, that is, the frame-based UPS method PF14, the proposed EventUPS, and the calibrated event-based method EventPS. EventUPS reconstructs clean and detailed normals that closely match the ground truth, and accurately recovers the circular lighting trajectory. The normals from PF14 appear visibly noisier and less defined, especially for the challenging HARVEST object, aligning with the error metrics in the table. This demonstrates our method's ability to robustly disentangle geometry, lighting, and non-ideal effects from sparse event data alone. EventUPS's superiority over non-learning, frame-based PF14 shows particular strength on objects with complex peculiarities and textures like HARVEST (23.32° vs. 30.65°) and READER (16.71° vs. 20.82°). Most notably, our uncalibrated method's performance is on par with, and even slightly surpasses, the average performance of the fully cal" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 544 + ], + "type": "text", + "content": "ibrated event-based method EventPS (14.17° MAE). For objects like HARVEST and BUDDHA, our method is more accurate. We attribute this to our model's explicit residual term, which effectively isolates and accounts for non-ideal, real-world effects like ambient illumination." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 468, + 562 + ], + "type": "text", + "content": "4.3. Evaluation on Real Camera" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 555, + 651 + ], + "type": "text", + "content": "Data. To demonstrate the practical applicability of our method, we fabricated five physical objects with ground truth normal maps and captured real events using our custom lighting systems. These data have varying characteristics: simple geometry (BALL), spatially-varying albedo (BALLICCV), moderate geometric details (BUNNY), and complex details (HORSE, TIGER)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Results. Fig. 8 shows results on our captured real dataset, which presents additional challenges including sensor noise, non-idealities in event generation. Our method produces high-quality normal reconstructions with clear" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7522" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 298, + 326 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 298, + 326 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 298, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 298, + 326 + ], + "type": "image", + "image_path": "b5f50a3b4db4cd2b8af8a548a1351b6b8e885e734861507b7b46ad50ca48eff1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 334, + 296, + 369 + ], + "lines": [ + { + "bbox": [ + 55, + 334, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 334, + 296, + 369 + ], + "type": "text", + "content": "Figure 8. Results on real data. EventUPS successfully recovers detailed normals despite sensor noise and non-ideal lighting (light trajectories are visualized as in Fig. 1)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "text", + "content": "preservation of fine details, achieving an average MAE of " + }, + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "inline_equation", + "content": "16.58^{\\circ}" + }, + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "text", + "content": " across all objects. The spherical object BALLICCV with \"ICCV\" text highlights an important property: the estimated normal map is invariant to ambient-albedo ratio variations, as the text is barely visible in both the event and the reconstructed normal map. The BUNNY and TIGER objects demonstrate EventUPS's ability to recover intricate geometric details. We observe slightly higher angular errors near object boundaries, which we attribute to near-light effects (the distance between the big ring light and the object is approximately " + }, + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "inline_equation", + "content": "10\\mathrm{cm}" + }, + { + "bbox": [ + 54, + 388, + 298, + 556 + ], + "type": "text", + "content": ") and imperfect lighting alignment. Nevertheless, the overall quality of reconstruction is high. The estimated lighting configurations accurately reflect the dual ring structure used during capture." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 298, + 715 + ], + "type": "text", + "content": "Robustness to lighting patterns. To evaluate the robustness of our method to different lighting configurations, we captured the same objects using both the \"dual ring\" and \"trefoil curve\" lighting patterns. As shown in Fig. 9, EventUPS produces consistent normal reconstructions regardless of the lighting pattern used. This demonstrates that our approach effectively recovers both surface normals and lighting directions without requiring a specific lighting arrangement, providing flexibility for different practical scenarios. The trefoil pattern yields better results for objects with complex geometry due to its more varied lighting directions, while the dual ring configuration provides more uniform" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 69, + 553, + 213 + ], + "blocks": [ + { + "bbox": [ + 317, + 69, + 553, + 213 + ], + "lines": [ + { + "bbox": [ + 317, + 69, + 553, + 213 + ], + "spans": [ + { + "bbox": [ + 317, + 69, + 553, + 213 + ], + "type": "image", + "image_path": "c664ac65b1101568e15329ab71443f7e5f53bd3fd0a1ab01c4796c8f343d8e1e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 220, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 220, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 220, + 555, + 255 + ], + "type": "text", + "content": "Figure 9. Performance across illumination patterns. EventUPS delivers consistent normal and lighting estimates for both the (a) dual-ring and (b) trefoil curve patterns, showcasing its robustness." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 273, + 441, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 273, + 441, + 286 + ], + "spans": [ + { + "bbox": [ + 313, + 273, + 441, + 286 + ], + "type": "text", + "content": "coverage for smoother objects." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 295, + 466, + 308 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 295, + 466, + 308 + ], + "spans": [ + { + "bbox": [ + 313, + 295, + 466, + 308 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 315, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 315, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 315, + 555, + 447 + ], + "type": "text", + "content": "We have presented EventUPS, the first method to solve UPS using an event camera. By introducing a novel augmented null space formulation, a continuous lighting model, and leveraging known lighting fixture geometry, we resolve the fundamental challenges of applying UPS to the event domain. Our experiments show that EventUPS achieves accuracy comparable to or better than its traditional frame-based counterpart while requiring only a fraction (5%) of the data bandwidth. Our method is robust to ambient illumination and different lighting configurations, making it practical for deployment in less controlled environments." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 447, + 556, + 579 + ], + "type": "text", + "content": "While our approach represents an advance in uncalibrated event-based PS, several avenues for future work remain. First, extending our framework to handle non-Lambertian surfaces would enhance its applicability in real-world scenarios with complex materials. Second, integrating learning-based components could further improve robustness to noise and outliers while maintaining the physical interpretability of our current model. Finally, exploring applications in dynamic scene reconstruction, where both the camera and objects may be in motion, represents an exciting direction for future research." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 597, + 408, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 597, + 408, + 611 + ], + "spans": [ + { + "bbox": [ + 314, + 597, + 408, + 611 + ], + "type": "text", + "content": "Acknowledgment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 715 + ], + "type": "text", + "content": "This work is supported by National Natural Science Foundation of China (Grant Nos. 62136001, 62302019, 624B2006, 62088102), Beijing Natural Science Foundation (Grant No. L233024), and Beijing Municipal Science & Technology Commission, Administrative Commission of Zhongguancun Science Park (Grant No. Z241100003524012). PKU-affiliated authors thank openbayes.com for providing computing resources." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 758 + ], + "type": "text", + "content": "7523" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 134 + ], + "type": "text", + "content": "[1] Neil G. Alldrin, Satya P. Mallick, and David J. Kriegman. Resolving the Generalized Bas-Relief Ambiguity by Entropy Minimization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 135, + 296, + 167 + ], + "type": "text", + "content": "[2] Peter N. Belhumeur, David J. Kriegman, and Alan L. Yuille. The Bas-Relief Ambiguity. Int. J. Comput. Vis., 35(1):33-44, 1999. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 168, + 296, + 202 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 168, + 296, + 202 + ], + "spans": [ + { + "bbox": [ + 62, + 168, + 296, + 202 + ], + "type": "text", + "content": "[3] M.K. Chandraker, F. Kahl, and D.J. Kriegman. Reflections on the generalized bas-relief ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2005. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "spans": [ + { + "bbox": [ + 62, + 203, + 295, + 246 + ], + "type": "text", + "content": "[4] Guanying Chen, Kai Han, Boxin Shi, Yasuyuki Matsushita, and Kwan-Yee K. Wong. Self-Calibrating Deep Photometric Stereo Networks. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2019. 2, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "spans": [ + { + "bbox": [ + 62, + 247, + 295, + 290 + ], + "type": "text", + "content": "[5] Guanying Chen, Michael Waechter, Boxin Shi, Kwan-Yee K. Wong, and Yasuyuki Matsushita. What Is Learned in Deep Uncalibrated Photometric Stereo? In Proc. of Eur. Conf. Comput. Vis., 2020. 2, 6, 7, 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "spans": [ + { + "bbox": [ + 62, + 291, + 295, + 334 + ], + "type": "text", + "content": "[6] Zehao Chen, Qian Zheng, Peisong Niu, Huajin Tang, and Gang Pan. Indoor Lighting Estimation Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "spans": [ + { + "bbox": [ + 62, + 335, + 295, + 368 + ], + "type": "text", + "content": "[7] Jason Chui, Simon Klenk, and Daniel Cremers. Event-Based Feature Tracking in Continuous Time with Sliding Window Optimization. arxiv preprint arxiv:2107.04536, 2021. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 295, + 401 + ], + "type": "text", + "content": "[8] O. Drbohlav and M. Chaniler. Can two specular pixels calibrate photometric stereo? In Proc. of Int. Conf. Comput. Vis., 2005. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "spans": [ + { + "bbox": [ + 62, + 402, + 295, + 435 + ], + "type": "text", + "content": "[9] Ondrej Drbohlav and Radim Šára. Specularities Reduce Ambiguity of Uncalibrated Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2002. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 436, + 295, + 490 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 436, + 295, + 490 + ], + "spans": [ + { + "bbox": [ + 57, + 436, + 295, + 490 + ], + "type": "text", + "content": "[10] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Trans. Pattern Anal. Mach. Intell., 2020. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 491, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 491, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 57, + 491, + 296, + 524 + ], + "type": "text", + "content": "[11] Georgiades. Incorporating the Torrance and Sparrow model of reflectance in uncalibrated photometric stereo. In Proc. of Int. Conf. Comput. Vis., 2003. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 57, + 525, + 295, + 579 + ], + "type": "text", + "content": "[12] Heng Guo, Fumio Okura, Boxin Shi, Takuya Funatomi, Yasuhiro Mukaigawa, and Yasuyuki Matsushita. Multispectral photometric stereo for spatially-varying spectral reflectances: A well posed problem? In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 57, + 580, + 295, + 624 + ], + "type": "text", + "content": "[13] Jin Han, Yuta Asano, Boxin Shi, Yinqiang Zheng, and Imari Sato. High-Fidelity Event-Radiance Recovery Via Transient Event Frequency. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 624, + 295, + 646 + ], + "type": "text", + "content": "[14] Hideki Hayakawa. Photometric stereo under a light source with arbitrary motion. JOSA A, 11:3079-3089, 1994. 1, 2, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 679 + ], + "type": "text", + "content": "[15] Micah K Johnson and Edward H Adelson. Shape estimation in natural illumination. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2011. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 680, + 295, + 713 + ], + "type": "text", + "content": "[16] Yakun Ju, Xinghui Dong, Yingyu Wang, Lin Qi, and Junyu Dong. A dual-cue network for multispectral photometric stereo. Pattern Recogn., 2020. 3" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 106 + ], + "type": "text", + "content": "[17] Dongyeop Kang, Yu Jin Jang, and Sangchul Won. Development of an inspection system for planar steel surface using multispectral photometric stereo. Optic. Eng., 2013. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 554, + 139 + ], + "type": "text", + "content": "[18] Hyeongwoo Kim, Bennett Wilburn, and Moshe Ben-Ezra. Photometric Stereo for Dynamic Surface Orientations. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 140, + 554, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 140, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 317, + 140, + 554, + 171 + ], + "type": "text", + "content": "[19] Leonid L Kontsevich, AP Petrov, and IS Vergelskaya. Reconstruction of shape from shading in color images. J. Opt. Soc. Am. A, 1994. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 172, + 554, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 172, + 554, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 172, + 554, + 205 + ], + "type": "text", + "content": "[20] Junxuan Li and Hongdong Li. Self-calibrating Photometric Stereo by Neural Inverse Rendering. In Proc. of Eur. Conf. Comput. Vis., 2022, 2, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 205, + 554, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 249 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 249 + ], + "type": "text", + "content": "[21] Feng Lu, Yasuyuki Matsushita, Imari Sato, Takahiro Okabe, and Yoichi Sato. Uncalibrated Photometric Stereo for Unknown Isotropic Reflectances. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 293 + ], + "type": "text", + "content": "[22] Tom Malzbender, Bennett Wilburn, Dan Gelb, and Bill Ambrisco. Surface Enhancement Using Real-Time Photometric Stereo and reflectance transformation. In Proc. of Eurographics Symp. Render. Tech., 2006. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 294, + 554, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 554, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 554, + 327 + ], + "type": "text", + "content": "[23] Nathan Matsuda, Oliver Cossairt, and Mohit Gupta. MC3D: Motion contrast 3D scanning. In Proc. of Int. Conf. Comput. Photogr., 2015. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 554, + 360 + ], + "type": "text", + "content": "[24] Manasi Muglikar, Guillermo Gallego, and Davide Scaramuzza. ESL: Event-based structured light. In Proc. of Int. Conf. 3D Vis., 2021. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 554, + 404 + ], + "type": "text", + "content": "[25] Manasi Muglikar, Leonard Bauersfeld, Diederik Paul Moeys, and Davide Scaramuzza. Event-based shape from polarization. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 404, + 554, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 554, + 437 + ], + "type": "text", + "content": "[26] Giljoo Nam and Min H Kim. Multispectral photometric stereo for acquiring high-fidelity surface normals. IEEE Comput. Graphics Appl., 2014. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 438, + 554, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 554, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 554, + 470 + ], + "type": "text", + "content": "[27] Thoma Papadhimitri and Paolo Favaro. A New Perspective on Uncalibrated Photometric Stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 554, + 514 + ], + "type": "text", + "content": "[28] Thoma Papadhinitri and Paolo Favaro. A Closed-Form, Consistent and Robust Solution to Uncalibrated Photometric Stereo Via Local Diffuse Reflectance Maxima. Int. J. Comput. Vis., 107(2):139-154, 2014. 2, 6, 7, 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 514, + 554, + 548 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 554, + 548 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 554, + 548 + ], + "type": "text", + "content": "[29] Henri Rebecq, Daniel Gehrig, and Davide Scaramuzza. ESIM: an open event camera simulator. In Proc. of Conf. on Robotics Learning, 2018. 6, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "text", + "content": "[30] Jieji Ren, Feishi Wang, Jiahao Zhang, Qian Zheng, Mingjun Ren, and Boxin Shi. DiLiGenT10" + }, + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 317, + 548, + 554, + 601 + ], + "type": "text", + "content": ": A photometric stereo benchmark dataset with controlled shape and material variation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2022. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 602, + 554, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 602, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 602, + 554, + 635 + ], + "type": "text", + "content": "[31] Wonjeong Ryoo, Giljoo Nam, Jae-Sang Hyun, and Sangpil Kim. Event fusion photometric stereo network. *Neural Networks*, 2023. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 554, + 669 + ], + "type": "text", + "content": "[32] Boxin Shi, Yasuyuki Matsushita, Yichen Wei, Chao Xu, and Ping Tan. Self-calibrating photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2010. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[33] Boxin Shi, Zhe Wu, Zhipeng Mo, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2016. 2, 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7524" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 689 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[34] Boxin Shi, Zhipeng Mo, Zhe Wu, Dinglong Duan, Sai-Kit Yeung, and Ping Tan. A benchmark dataset and evaluation for non-lambertian and uncalibrated photometric stereo. IEEE Trans. Pattern Anal. Mach. Intell., 2019. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 150 + ], + "type": "text", + "content": "[35] Melvyn L. Smith and Lyndon N. Smith. Dynamic photometric stereo - a new technique for moving surface analysis. Image Vision Comput., 2005. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[36] Kalyan Sunkavalli, Todd Zickler, and Hanspeter Pfister. Visibility Subspaces: Uncalibrated Photometric Stereo with Shadows. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[37] Tsuyoshi Takatani, Yasuyuki Matsushita, Stephen Lin, Yasuhiro Mukaigawa, and Yasushi Yagi. Enhanced photometric stereo with multispectral images. In Proc. of Int. Conf. on Mach. Vis. Appl., 2013. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 274 + ], + "type": "text", + "content": "[38] Ping Tan, Satya P. Mallick, Long Quan, David J. Kriegman, and Todd Zickler. Isotropy, Reciprocity and the Generalized Bas-Relief Ambiguity. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2007. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 275, + 294, + 319 + ], + "type": "text", + "content": "[39] Antoni Rosinol Vidal, Henri Rebecq, Timo Horstschaefer, and Davide Scaramuzza. Ultimate SLAM? combining events, images, and imu for robust visual slam in HDR and high-speed scenarios. IEEE Robot. Autom. Lett., 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 294, + 363 + ], + "type": "text", + "content": "[40] Daniel Vlasic, Pieter Peers, Ilya Baran, Paul Debevec, Jovan Popovic, Szymon Rusinkiewicz, and Wojciech Matusik. Dynamic shape capture using multi-view photometric stereo. ACM Trans. Graph., 28(5):174, 2009. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 408 + ], + "type": "text", + "content": "[41] Feishi Wang, Jieji Ren, Heng Guo, Mingjun Ren, and Boxin Shi. DiLiGenT-Pi: Photometric stereo for planar surfaces with rich details-benchmark dataset and beyond. In Proc. of Int. Conf. Comput. Vis., 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 410, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 441 + ], + "type": "text", + "content": "[42] Robert J. Woodham. Photometric method for determining surface orientation from multiple images. Opt. Eng., 1980, 1, 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "spans": [ + { + "bbox": [ + 56, + 444, + 294, + 476 + ], + "type": "text", + "content": "[43] Zhe Wu and Ping Tan. Calibrating Photometric Stereo by Holistic Reflectance Symmetry Analysis. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2013. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 294, + 521 + ], + "type": "text", + "content": "[44] Takuya Yoda, Hajime Nagahara, Rin-ichiro Taniguchi, Kei-ichiro Kagawa, Keita Yasutomi, and Shoji Kawahito. The dynamic photometric stereo method using a multi-tap cmos image sensor. Sensors-basel., 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 565 + ], + "type": "text", + "content": "[45] Bohan Yu, Jieji Ren, Jin Han, Feishi Wang, Jinxiu Liang, and Boxin Shi. EventPS: Real-Time Photometric Stereo Using an Event Camera. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 2, 3, 6, 7, 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 620 + ], + "type": "text", + "content": "[46] A.L. Yuille, D. Snow, R. Epstein, and P.N. Belhumeur. Determining Generative Models of Objects Under Varying Illumination: Shape and Albedo from Multiple Images Using SVD and Integrability. Int. J. Comput. Vis., 35(3):203-222, 1999. 2, 3, 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 623, + 294, + 666 + ], + "type": "text", + "content": "[47] Xinyu Zhou, Peiqi Duan, Boyu Li, Chu Zhou, Chao Xu, and Boxin Shi. EvDiG: Event-guided Direct and Global Components Separation. In Proc. of IEEE Conf. Comput. Vis. Pattern Recog., 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 689 + ], + "type": "text", + "content": "[48] Zhenglong Zhou and Ping Tan. Ring-Light Photometric Stereo. In Proc. of Eur. Conf. Comput. Vis., 2010. 2" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "7525" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_content_list.json b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..8a27d3ae2a6de7e16124f81f52b9ed2b055472e2 --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_content_list.json @@ -0,0 +1,1426 @@ +[ + { + "type": "text", + "text": "Everything is a Video: Unifying Modalities through Next-Frame Prediction", + "text_level": 1, + "bbox": [ + 120, + 130, + 875, + 151 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "G. Thomas Hudson", + "bbox": [ + 169, + 180, + 330, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dean Slack", + "bbox": [ + 364, + 181, + 457, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Thomas Winterbottom", + "bbox": [ + 493, + 181, + 673, + 196 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jamie Sterling", + "bbox": [ + 707, + 181, + 825, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chenghao Xiao", + "bbox": [ + 281, + 199, + 408, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Junjie Shentu", + "bbox": [ + 442, + 199, + 553, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Noura Al Moubayed", + "bbox": [ + 586, + 199, + 754, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Department of Computer Sciences, Durham University", + "bbox": [ + 281, + 233, + 714, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "g.t.hudson@durham.ac.uk", + "bbox": [ + 393, + 255, + 601, + 268 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa0cbccea2015ca31ff32b5b015614b53a703ee11e6c36da55fec0b265aba43f.jpg", + "image_caption": [ + "Figure 1. Sample outputs from our next frame prediction model across multiple modalities. The model receives frames to the left of the red line as input and everything right of the red line has been generated by the model." + ], + "image_footnote": [], + "bbox": [ + 161, + 277, + 834, + 463 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 267, + 508, + 346, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Multimodal learning, which involves integrating information from various modalities such as text, images, audio, and video, is pivotal for numerous complex tasks like visual question answering, cross-modal retrieval, and caption generation. Traditional approaches rely on modality-specific encoders and late fusion techniques, which can hinder flexibility when adapting to new tasks or modalities. To address these limitations, we introduce a novel framework that extends the concept of task reformulation beyond natural language processing (NLP) to multimodal learning. We reformulate diverse multimodal tasks into a unified next-frame prediction problem, allowing a single model to handle different modalities without modality-specific components. This method treats all inputs and outputs as sequential frames in a video, enabling seamless integration of modalities and effective knowledge transfer across tasks. Our approach is evaluated on a range of tasks, including text-to-text, image-to-text, video-to-video, video-to-text, and audio-to-text, demonstrating the model's ability to generalize across modalities. We show that task reformulation can significantly simplify multimodal model design across various tasks, laying the groundwork for more generalized multimodal foundation models.", + "bbox": [ + 107, + 540, + 503, + 888 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 514, + 508, + 643, + 523 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Many tasks we seek to solve are multimodal. The ability to process and integrate information from multiple modalities such as text, image, audio, and video, have proven crucial for a range of complex tasks, including cross-modal retrieval [31], visual question answering [20, 33, 41], and caption generation [1]. The ability to understand and reason about information from different sensory inputs simultaneously mirrors human perception and cognition, making multimodal models an essential component in advancing artificial intelligence towards more generalized and versatile capabilities.", + "bbox": [ + 511, + 536, + 906, + 700 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The predominant approach to building multimodal models involves using separate encoders for each modality. For example, in vision-and-language models, a convolutional neural network (CNN) or a vision transformer (ViT) [10] might be used to encode images, while a recurrent neural network (RNN) or transformer is used to encode text. These modality-specific encoders are then fused at a later stage, typically via concatenation or attention mechanisms, to enable the model to perform joint reasoning over the encoded representations. While effective, this paradigm has inherent limitations. It necessitates careful design choices for each modality and often struggles with scalability and flexibility, especially when extending to new combinations of modali", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "22004", + "bbox": [ + 478, + 945, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ties or tasks not seen during training.", + "bbox": [ + 89, + 90, + 334, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel approach centered around task reformulation. Task reformulation is a well-established paradigm in natural language processing (NLP) that facilitates multitask learning by transforming diverse tasks into a common format [21]. A prominent example is prompt-based learning, where NLP tasks are reimagined as instructions or prompts to a large language model (LLM). This paradigm allows a wide range of NLP tasks, such as translation, summarization, and question answering, to be represented uniformly in a chat message format [23]. By doing so, it leverages the generalization capabilities of large pre-trained models, enabling them to perform multiple tasks with minimal task-specific customization.", + "bbox": [ + 89, + 106, + 480, + 316 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we extend the concept of task reformulation beyond NLP to the realm of multimodal learning. Specifically, we propose a framework in which tasks across various modalities are reformulated as a unified next-frame prediction problem. This approach simplifies the design of multimodal models by enabling them to handle diverse input types — text, images, audio, or video using a single, coherent mechanism. By representing all tasks as next-frame prediction, we create a shared interface for the model to learn and reason about information across modalities. This not only facilitates the integration of new modalities with minimal effort but also enhances the model's ability to transfer knowledge across different tasks and domains. Ultimately, just as LLMs function as foundation models for NLP, we intend this framing to be a foundation model across all modalities.", + "bbox": [ + 89, + 316, + 482, + 556 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions of this paper are as follows:", + "bbox": [ + 107, + 559, + 450, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a task-reformulation method and show how to represent a range of multimodal tasks as next-frame prediction.", + "- We explore how this reformulation allows a single transformer-based model to solve text, image, video, and audio processing tasks without any modality-specific encoders.", + "- Experimental results show our model achieves comparable performance to single-task models which haven't been pretrained on additional data." + ], + "bbox": [ + 91, + 575, + 478, + 723 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 89, + 739, + 227, + 753 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Learning With the rise of transformers, LLMs, and very large scale pretraining, new paradigms for harmonising different modalities have been pushed to the forefront of multimodal learning. UNITER [7] uses a single universal transformer to process both image and text inputs into a join embedding space, with image representations pretended to the beginning of the input embeddings to the transformer. However, the image inputs are first processed by an RNN to obtain an embedding such that it is", + "bbox": [ + 89, + 763, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "compatible with the textual inputs. Our framework in contrast has no need to preprocess input modalities differently as they are unified into a single visual input space theoretically capable of fully representing all multimodal interactions. Baevski et al. [2] propose data2vec, a framework for instilling features from multiple modalities into a single latent representation. Where data2vec seeks to process raw text, images, and audio using the same method, our work here reformulates the task inputs from multiple modalities into the same visual input medium. UniT [14], a multimodal multitask transformer uses separate encoders for each modality followed by a shared decoder. Our work uses a single visual encoder for our unified visual input paradigm. The most powerful state-of-the-art multimodal models such as FLAVA [27] and GPT-4 vision [23] integrate vision inputs via a patchwise process pipeline through a visual transformer fully integrated into the training pipeline. Both the architecture design and scale at which these foundation models are trained yield revolutionary performance at a variety of open-ended multimodal tasks. Nonetheless, though these foundation model are the closest that benchmarks have practically come to seamlessly integrating images with visual inputs, fundamentally the input modalities are handled differently by the model. Our work here bridges this remaining gap in processing of separate modalities. We reformulate typically spura-visual tasks into a visual input domain with no loss of information i.e. text and audio can be fully represented in images. This reformulation allows us to —for the first time—explore the generative pretraining capacity of multimodal information in one truly unified representation.", + "bbox": [ + 511, + 90, + 903, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task Reformulation Task reformulation involves allowing a single model to solve multiple tasks be converting them into a single 'supertask'. This technique is common in NLP where a range of tasks such as sentiment classification, and coreference resolution have been reformulated as span-extraction [16], and question answering [21] among others.", + "bbox": [ + 511, + 578, + 903, + 681 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The recent trend of prompt-based learning uses language modelling on prompt-answer sequences as the 'supertask' [23]. In this paradigm, the model is trained to follow an instruction describing the task and provide a response. In this way, any NLP task can be reformulated as the language modelling 'supertask'. Similar ideas have been explored in computer vision, notably the Transframer [22] which aimed to unify visual tasks. In our work, we extend this concept, exploring how next-frame prediction can be used as a 'supertask' to unite multiple modalities.", + "bbox": [ + 511, + 684, + 903, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Next-frame Prediction CNN-based deep learning architectures for the autoregressive generation of future video frames have been steadily improving over the last decade", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "22005", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "[24, 26, 30, 38]. Transformer-based models trained at scale such as VideoGPT [35] have substantially improved the performance of quality and fidelity of long and short term future frame predictions. Modern diffusion-based frame prediction models are now able to generate photorealistic outputs [13]. The aim of our work is not a new model architecture. Our model is inspired from state-of-the-art architectures to serve as an appropriately powerful benchmark for our analysis on our benchmark datasets.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Visual Representation of Text Inspired by the success of unsupervised next-word prediction with language models, learning the next pixel of images was proposed [6] and shown able to learn image representations. Learning textual semantics with vision models, by rendering texts into images, has drawn more and more attention [12, 25, 29] due to the drawbacks of tokenization in traditional language models and limitations in cross-lingual transferability. Xiao et al. [34] show that vision models pretrained on rendered images have stronger robustness on typo and word-order shuffling perturbations, and the representations of them display better isotropy on out-of-distribution (OOD) languages. In this work, we extend this approach when reformulating NLP tasks to video as well as other modalities which output text.", + "bbox": [ + 89, + 250, + 482, + 474 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 491, + 181, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our proposed framework consists of two key components: (1) Methods of reformulating a diverse range of input and output modalities into the single task of next-frame prediction; (2) A pure transformer-based model architecture. We introduce these components in the following sections.", + "bbox": [ + 89, + 517, + 482, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Reformulation", + "text_level": 1, + "bbox": [ + 89, + 604, + 238, + 619 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The datasets we use are carefully selected to cover a diverse range of input and output modalities, illustrating the versatility of our approach in reformulating various tasks as next-frame prediction problems. Our selection spans tasks from text and image classification to more complex audio, video, and multimodal tasks, enabling us to evaluate the model's ability to generalize across different data types and task requirements. By converting each task into a uniform format—a $64 \\times 64$ RGB video sequence—we create a consistent framework where the model treats every input and output as sequential frames, simplifying the multimodal learning process.", + "bbox": [ + 89, + 627, + 482, + 808 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "For each task, we insert a separator token $(|)$ between the input and output frames to clearly delineate where the input ends and the prediction begins, guiding the model in distinguishing between these two parts of the sequence.", + "bbox": [ + 89, + 809, + 482, + 869 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In tasks involving textual data, we use a simple tokenizer (splitting on spaces and also on punctuation) to break", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "down the text into individual tokens, which we then render as video frames in the sequence. Each token is represented in a consistent format: a fixed-width font scaled to fill the $64 \\times 64$ frame, ensuring clarity and compatibility with our video-based input structure. This approach allows the model to read and predict text as it would any other frame, effectively bypassing the need for traditional text-based tokenization while integrating text seamlessly with other modalities. By converting each token to a visual format, we enable cross-modal knowledge transfer, allowing the model to process text, images, and other modalities through a shared, frame-based learning paradigm.", + "bbox": [ + 511, + 90, + 903, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Text-to-Text The SST2 (Stanford Sentiment Treebank 2) dataset is a widely-used benchmark for sentiment classification, consisting of thousands of movie review excerpts labelled with binary sentiment labels (positive or negative) [28]. We use the text-encoding method described above, rendering each token as a video frame using a fixed-width font (Figure 2).", + "bbox": [ + 511, + 290, + 903, + 397 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/857b51dd8cf3d79d1619fea16da761d4bdd4874d67aa49984f8644ee41160899.jpg", + "image_caption": [ + "Figure 2. Truncated example of the SST2 sentiment dataset rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 535, + 409, + 885, + 449 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Image-to-Text As a simple test of image recognition ability, we employ the CIFAR-10 dataset. CIFAR-10 is a benchmark for image classification and consists of 60,000 color images, each with dimensions of $32 \\times 32$ pixels [17]. These images are equally divided into ten distinct, mutually exclusive classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck.", + "bbox": [ + 511, + 523, + 903, + 628 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/208eee69929332ad58dec6c1f51a1510e52fa25e71973ce147570a935609e136.jpg", + "image_caption": [ + "Figure 3. Truncated example of the CIFAR10 image classification dataset rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 611, + 641, + 807, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In our reformulation approach, each CIFAR-10 image is resized to $64 \\times 64$ pixels to fit our standardized input frame size. Following the resized image, we insert a separator frame, which visually signifies the end of the input and the beginning of the output sequence, followed by a frame containing the class label text rendered as an image (Figure 3).", + "bbox": [ + 511, + 761, + 903, + 852 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Video-to-Text We utilize the TinyVIRAT action classification dataset to rigorously assess the model's capacity for", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "22006", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "understanding video content [9]. The TinyVIRAT dataset comprises of 7,663 training and 5,166 testing examples, each annotated with one or more of 26 distinct action labels that describe the activities depicted within low-resolution video sequences. This dataset is particularly noteworthy for its multi-label nature, allowing for nuanced action recognition; for instance, a video sequence might feature a man walking while carrying a backpack, which would be simultaneously labelled with both \"walking\" and \"carrying\" actions. This characteristic enables the evaluation of the model's ability to recognize and differentiate between overlapping actions occurring in dynamic environments.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To reformulate the task for next-frame prediction, we reformulate each video sequence by first presenting the input video frames, followed by a separator token, and concluding with a comma-separated list of action labels that correspond to the activities depicted in the video. (Figure 4).", + "bbox": [ + 89, + 273, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d327b0bd614e4a5ceef757a7e35dd786cf82e0b950ca02cd4998d2f339dc653a.jpg", + "image_caption": [ + "Figure 4. Truncated example of the Tinyvirat video classification dataset rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 91, + 363, + 483, + 397 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video+Text-to-Text To explore the integration of multiple modalities within a single input, we leverage the CLEVRER (CoLlision Events for Video REpresentation and Reasoning) dataset, which is designed for video question answering (VQA) [37]. This dataset comprises synthetic video sequences depicting rendered 3D objects interacting through collisions, movements, and occlusions within a simple, controlled environment. Each video is paired with questions that require the model to understand and reason about these visual events. The questions in CLEVRER span various reasoning tasks, including descriptive, explanatory, and predictive questions about the events in the video, such as identifying specific objects, understanding object interactions, and predicting future states.", + "bbox": [ + 89, + 491, + 483, + 702 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, we represent each question and corresponding video as a unified sequence of frames. Each input sequence begins with frames of the CLEVRER video sequence subsampled to 4 frames. This is followed by the question text, separator token, and answer text encoded in the manner described previously (Figure 5). Similarly, we use the frame QA task from TGIF-QA dataset[15], containing animated gifs with QA pairs.", + "bbox": [ + 89, + 704, + 483, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To keep the sequence length manageable, we exclude complex counterfactual questions and multiple-choice answers from the dataset (only using the 'descriptive' subset), as these would require additional frames and could lead to overly long sequences that complicate training.", + "bbox": [ + 89, + 825, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cfa90a4f0fa41cb12efc8e8c2e93e7fa898a267af929e40de053e1cfc73e0576.jpg", + "image_caption": [ + "Figure 5. Truncated example of the CLEVRR task rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 514, + 88, + 905, + 118 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Video-to-Video We employ two tasks to test video-to-video performance. Firstly, we convert the TinyVIRAT videos into greyscale and test the model's ability to colourise the output. The sequence consists of the greyscale-converted video sequence, the separator frame, then the original RGB video sequence (Figure 6).", + "bbox": [ + 511, + 188, + 906, + 279 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/b1fc20fc80ac61fd4f40f2f050872c0149be33646840806e3bf55f0ffa733c8c.jpg", + "image_caption": [ + "Figure 6. Truncated example of the colorization task rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 532, + 297, + 885, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we test the model's ability to perform object tracking. The Large-scale Single Object Tracking dataset (LaSOT) consists of 1,550 video sequences hand-annotated with bounding boxes following an object of interest [11]. We reformulate this task by converting it into a sequence which consists of the first frame of the video with the bounding box overlayed, the rest of the sequence, the separator frame, and finally the full sequence including the overlayed bounding box (Figure 7).", + "bbox": [ + 511, + 398, + 906, + 534 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d69b9847d3f28d7759b7472286a44efed2fc8f7c77d86d624d982b4d8e451b66.jpg", + "image_caption": [ + "Figure 7. Truncated example of the LaSOT object tracking task rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 532, + 551, + 885, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Audio-to-Text To demonstrate the ability to process audio data, we utilize the AudioMNIST dataset, which contains 30,000 audio recordings of spoken digits ranging from zero to nine [3]. These recordings are spoken by multiple speakers with varying accents and intonations, providing a diverse dataset that challenges the model to recognize and classify spoken language across different vocal characteristics.", + "bbox": [ + 511, + 686, + 905, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In our framework, each audio sample is preprocessed into a spectrogram, which is used at the first frame in the video sequence, representing the audio content visually for consistency with other modalities. Following the spectrogram, a separator token frame is added, and the label digit is represented in the final frame as text (Figure 8).", + "bbox": [ + 511, + 810, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "22007", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/39f555cdaf3142592099cfb4a1ad63de0c0043d8fdd139fd4873e89d91444f9b.jpg", + "image_caption": [ + "Figure 8. Example of the AudioMNIST audio classification dataset rendered as a video. Each square is a frame of the video sequence." + ], + "image_footnote": [], + "bbox": [ + 189, + 88, + 387, + 140 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2. Video Prediction Model", + "text_level": 1, + "bbox": [ + 89, + 219, + 308, + 234 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We perform all tasks using a pure transformer-based model architecture. Building upon the successes of the ViT for image classification and TimesFormer [4] for video understanding, we modify these architectures to accommodate for spatiotemporal modelling of video sequences and autoregressive video prediction. We further adapt the model to progressively process patches at smaller resolutions, similar to that of U-net style models, such that adjacent input patches are linearly merged to reduce the effective image resolution at each block on the encoding side. This design choice allows us to train end-to-end models efficiently without resorting to pretrained encoder/decoder models. As illustrated in Fig. 9, partitioning into patches is followed by a series of local space-time transformed blocks separated by patch merge operations, with a global space-time transformer block operating on the reduced resolution patch-based representations. This has the benefit of capturing multiscale spatial features, while reducing the computational complexity of the global self-attention operations due to a smaller number of image patches. The reverse process is applied on the decoding side of the model to produce a predicted next image frame for each input frame. During inference, the final predicted frame is concatenated to the input sequence in an autoregressive manner. We outline the key model components and considerations below. Our model uses 41M parameters (for comparison, Transframer [22] uses 470–829M parameters).", + "bbox": [ + 91, + 243, + 483, + 650 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Patch Representation A core feature of the model is in the partitioning of input video frames into non-overlapping patches, similar to ViT. These patches are processed using local and global spatiotemporal self-attention mechanisms as described in [4], with spatial self-attention focusing on intra-frame patch relationships and temporal self-attention handling inter-frame causal dependencies. Temporal causality is enforced by masking future patches during training, ensuring that only past and present information is available for prediction. The model takes as input a video sequence $\\mathcal{X} = \\{x_{t}\\}_{t = 1}^{T}$ , such that each $x_{t} \\in \\mathbb{R}^{C \\times H \\times W}$ is a $C$ -channel image frame of height $H$ and width $W$ at timestep $t$ for $t = 1,2,\\ldots,T$ . Frames are partitioned into $N$ equally-sized non-overlapping patches, and linearly embedded to a 1D representation of size $K$ (we set $K$ to 512", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/8a2b77de0fb96c9b90813dc4f9ffef359029a1c4cb4cf3467a2f9d1938920873.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskInput LenTarget LenBatch Size
Text Classification21116
Image Classification21128
Video Classification11116
Audio Classification21256
Video QA21116
Object Tracking21208
Video Colorization111016
", + "bbox": [ + 522, + 88, + 901, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Training configurations for each task", + "bbox": [ + 571, + 223, + 846, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "throughout). This embedding is combined with learnable spatial and temporal positional encodings to form the input sequence for the transformer. Spatial attention is applied between image patches within each frame independently, while causal temporal attention is applied across timesteps at the same spatial patch location.", + "bbox": [ + 511, + 265, + 906, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Details", + "text_level": 1, + "bbox": [ + 511, + 368, + 671, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train our model on each dataset independently, with no language model or image pretraining, in order to determine the feasibility of our task reformulation approach with no latent performance benefits from transferring existing learned representations. We train end-to-end with a Multi-Scale Structural Similarity Index Measure (SSIM) loss [32], with a constant learning rate of $3 \\times 10^{-4}$ , and a batch size of 8-32 depending on the task context length. We train with the AdamW [19] optimizer using default parameters, and set dropout to 0.1 for all layers except the final output layer. All models are evaluated on checkpoints corresponding to the best SSIM validation performance, with all models trained on a single NVIDIA A100 for a maximum of 7 GPU days. The input/output lengths and batch sizes vary per task, and are detailed in Table 1.", + "bbox": [ + 511, + 391, + 906, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 633, + 643, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We train on each task in turn in a single-task manner (we leave training on all the tasks jointly in a multitask setting as future work). For tasks which output text, we perform OCR on the output frames using the open-source tesseract OCR engine1. For tasks which have a fixed vocabulary (e.g. classification tasks), the OCR text is matched to the closest vocabulary word to minimise any OCR-related errors. For classification tasks, the resulting text is measured using F1-score and Accuracy.", + "bbox": [ + 511, + 660, + 906, + 796 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The object tracking task is evaluated by extracting the bounding box in the outputted frames, then comparing these to the labelled boxes using Intersection over Union (IoU).", + "bbox": [ + 511, + 797, + 905, + 843 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For video colorization, we evaluate both colorization quality and temporal consistency. This dual evaluation is", + "bbox": [ + 511, + 844, + 905, + 875 + ], + "page_idx": 4 + }, + { + "type": "page_footnote", + "text": "1https://github.com/tesseract-ocr/tesseract", + "bbox": [ + 531, + 887, + 870, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "22008", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/39e907a353976dd30fc2df5f811125c3af6de5ab05ea6a7a482c1eb0f8d724bb.jpg", + "image_caption": [ + "Figure 9. Illustration of the Transformer-based video prediction model used in this work. $T$ timestep video image frames are partitioned into $N$ non-overlapping patches and linearly embedded before being processed by a series of local and global space-time transformer blocks. Local blocks are restricted to local spatial and temporal attention windows, separated by non-overlapping patch merge/unmerge operations with a patch window size of 2. Global blocks operate on a smaller effective image resolution (e.g., for an input of $64 \\times 64$ and a patch size of $8 \\times 8$ , the effective resolution at the global blocks is $16 \\times 16$ ). Skip connections (dashed lines) are used between corresponding patch merge/unmerge operations to preserve information such as static background features." + ], + "image_footnote": [], + "bbox": [ + 174, + 87, + 823, + 218 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/2cf86d690bccee7b8ba94b5d2b73a36665e900b551f45360e3d4a54a60bf22bb.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TaskDatasetModelPSNR ↑SSIM ↑F1 ↑Acc ↑IoU ↑CDC ↓Colorfulness ↑
Text ClassificationSST-2Ours41.898.776.875.5---
BERT-base---81.2---
Vega v1 [40]†--91.3----
Image ClassificationCIFAR-10Ours45.795.989.189.1---
ViT-base--71.371.3---
Resnet-101--90.090.0---
PCANet [5]--77.177.1---
ViT-huge [10]†--99.599.5---
Video ClassificationTinyVIRATOurs50.197.374.1 (30.4*)60.4---
ResNet50†--29.1*----
WideResNet†--32.6*----
Audio ClassificationaudioMNISTOurs50.998.996.997.1---
AlexNet [3]---95.8---
Video QACLEVREROurs25.779.852.452.5--
LSTM [37]†---34.7---
+ CNN [37]†---51.8---
TGIF-QAOurs42.098.553.252.3---
LSTM + Attn [15]---51.9---
Object TrackingLaSOTOurs35.798.7--0.63--
Dai et al. [8]----0.60--
Video ColorizationTinyVIRATOurs42.499.7---0.0273.1
Zhang et al. [39]-----0.0282.2
Yang et al. [36]-----0.0372.0
", + "bbox": [ + 109, + 324, + 887, + 705 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Performance metrics across various tasks using our proposed next-frame prediction framework. Metrics include Peak Signa-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), F1-score and Accuracy (F1, Acc), Intersection over Union (IoU), Color Distribution Consistency index (CDC), and Colorfulness, which are each tailored to evaluate specific task outputs. *macro F1-score to allow direct comparison with baseline models. †indicates models with pretraining on other datasets.", + "bbox": [ + 88, + 715, + 906, + 773 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "crucial, as a model could appear consistent by merely reproducing the grayscale input. We use PSNR and the colorfulness metric from Liu et al. to assess color diversity. Temporal consistency is measured using the Color Distribution Consistency (CDC) index [18], which calculates the Jensen-Shannon (JS) divergence of color distributions", + "bbox": [ + 88, + 797, + 480, + 888 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "across frames:", + "bbox": [ + 513, + 799, + 612, + 811 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\nC D C _ {t} = \\frac {1}{3 \\times (N - t)} \\sum_ {c \\in \\{r, g, b \\}} \\sum_ {i = 1} ^ {N - t} J S (P _ {c} (I ^ {i}), P _ {c} (I ^ {I + t}))\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 832, + 903, + 877 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $P_{c}(I^{i})$ is the normalized probability distribution", + "bbox": [ + 532, + 883, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "22009", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "of channel $c$ in frame $I^i$ of $N$ total frames. Overall consistency is computed as:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nC D C = \\frac {1}{3} \\left(C D C _ {1} + C D C _ {2} + C D C _ {4}\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 133, + 421, + 162 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Results", + "text_level": 1, + "bbox": [ + 89, + 175, + 176, + 191 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results across each task are presented in Table 2, while Figure 1 illustrates several sample outputs produced by our model, offering a qualitative perspective on its performance.", + "bbox": [ + 89, + 200, + 482, + 246 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the SST-2 text classification task, our model achieves an F1-score of 76.8. Although this is lower than the current state-of-the-art performance of 91.3, that was achieved through the use of a \"Bag of Tricks\" methodology involving extensive pretraining and complex finetuning processes [40]. Our model is simply trained on the SST-2 dataset, with no use of pretrained weights or advanced external embeddings, making it comparable to other models used without pretraining (See the BERT-base result). Some portion of this lower score can also be attributed to truncating any inputs over 20 tokens (limiting the evaluation to shorter inputs 20 tokens or fewer increases the F1-score to 80.0).", + "bbox": [ + 88, + 247, + 482, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the CIFAR-10 image classification task, our model attains an accuracy of 89.1. While this score is lower than fine-tuned ViT models leveraging pretrained weights, which reach an impressive accuracy of 99.5 [10], our model performs significantly better than simpler baseline models, such as PCANet, which achieves 77.1 accuracy [5]. This performance suggests that our reformulation approach is able to perform image classification tasks without any pretraining, surpassing older, simpler architectures while trailing behind state-of-the-art models that benefit from extensive pretraining.", + "bbox": [ + 89, + 429, + 482, + 595 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In multi-label image classification on the TinyVIRAT dataset, our approach successfully outputs multiple labels per instance. When compared to baselines, our model surpasses the ResNet50 baseline (29.1) in performance but does not reach the level of the WideResNet model (32.6) [9].", + "bbox": [ + 89, + 595, + 482, + 686 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the AudioMNIST audio classification task, the model achieves an accuracy of 97.1, surpassing the AlexNet baseline score of 95.82 from the original dataset paper [3]. This notable improvement demonstrates the model's efficacy in audio-based classification tasks. Analysis of misclassifications reveals that the most frequent error involves confusion between the spoken digits \"four\" and \"five,\" suggesting areas for potential refinement in distinguishing phonetically similar sounds.", + "bbox": [ + 89, + 686, + 482, + 823 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For the CLEVRR VQA (Visual Question Answering) dataset, our model achieves an accuracy of 52.5 on descriptive questions. This performance surpasses both the LSTM (34.7) and LSTM+CNN (51.8) baselines, illustrating our model's ability to understand the relationship between", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "visual and textual information without separate encoder-decoder models [37]. It is important to note that due to hardware limitations, and the long question lengths included in the CLEVRER dataset, we subsampled the video to only include 4 frames of the moving scene, while also reducing the resolution to $64 \\times 64$ , significantly limiting the accuracy achievable by our approach.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the LASOT object tracking task, the model reaches an intersection over union (IoU) of 0.63, demonstrating a consistent ability to track objects throughout video sequences. Analysis of tracking outputs reveals that the model effectively follows the object of interest across frames with accurately drawn bounding boxes. However, tracking accuracy diminishes slightly toward the end of longer sequences, with the borders of the overlaid bounding box eroding, likely due to autoregressive pixel-level error propagation.", + "bbox": [ + 511, + 198, + 903, + 333 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "On the colorization task, our model shows improved color diversity in the generated outputs compared to the ground truth (73.1 color diversity for our model versus 70.6 in the original dataset). However, it exhibits less consistency in color application across the video sequence (0.0169 consistency for our model versus 0.00522 in the original dataset). This trade-off between diversity and consistency suggests that while our approach introduces a more vibrant color palette, further refinement could improve temporal coherence, particularly for applications requiring uniform color continuity across frames.", + "bbox": [ + 511, + 334, + 903, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ch", + "bbox": [ + 532, + 502, + 553, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.1. Attention Maps", + "text_level": 1, + "bbox": [ + 511, + 527, + 669, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In order to examine what is being learned by our model, we visualize internal patch-wise attention scores both spatially (within a frame) and temporally (between frames) in Figure 10.", + "bbox": [ + 511, + 550, + 903, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the TinyVIRAT video classification task, the model pays most attention to the object/person performing the action, as well as frames and pairs of frames which indicate the class. In the first example (top row), spatial attention focuses on image patch locations containing the edges of the car which suggest relative movement, and temporal attention is highest for the final two frames in which the vehicle moves position, indicating the label \"vehicle moving\" should be output.", + "bbox": [ + 511, + 612, + 903, + 747 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "For colorization, the greatest attention is paid to the first frames of the input sequence, which we suggest allows the model to produce consistent coloring across the whole sequence (more attention is also paid to the first frame of the output sequence, presumably to maintain color consistency). Spatially, the model clearly attends to key moving objects and areas of distinct color.", + "bbox": [ + 511, + 750, + 903, + 853 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In the CLEVRR examples, we see how the model is able to focus on question frames which contains key words needed to produce the output, such as \"color\" and \"metal\".", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "22010", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3a566e7b18f2f85f0f1f45f399715a17a43b7f97d653300fbfad4a414d61a5ab.jpg", + "image_caption": [ + "Figure 10. Attention visualisations showing where the model attends to both spatially (indicated by light/dark areas overlaid on the frame), and temporally (indicated by a colour scale above the sequence). Spatial attention scores are calculated by taking the patch-wise contribution per frame, averaged over all layers. Temporal attention scores are calculated by taking the temporal attention weights for each input frame when generating the target frames, averaged over all global space-time attention layers." + ], + "image_footnote": [], + "bbox": [ + 205, + 70, + 794, + 425 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Within the moving object scene frames, we see how the model spatially attends to objects and their trajectory paths, correctly identifying objects such as \"the metal cylinder\" and \"the sphere\".", + "bbox": [ + 89, + 493, + 482, + 554 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Lastly, for sentiment text classification, we observe a similar pattern to attention maps in LLMs, where the model pays more attention to words with a strong emotive sentiment. In our model we see frames containing the words 'nightmare', 'painful', and 'dumb' exhibiting high temporal attention weightings, as they are strong indicators of the overall sentiment of the review.", + "bbox": [ + 89, + 554, + 482, + 659 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Limitations and Future work", + "text_level": 1, + "bbox": [ + 89, + 667, + 352, + 681 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The aim of this work is not necessarily to surpass state-of-the-art performance on any of the tasks, but to show that unifying these modalities via reformulation is possible, without heavily compromising on performance relative to task-specific baselines. Future work should aim to scale this up, improving performance and tackling more complex tasks. We suggest that this can be done by mirroring the improvements gained from pretraining in NLP and computer vision by first training the model on unstructured tasks from each modality. Tasks such as language modelling and image captioning can be naturally reformulated as next-frame-prediction and included in our paradigm. This is especially important as a large portion of the current training time on tasks such as classification and question answering is ded", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "icated to learning to output words rather than solving the task itself. Beginning with a checkpoint already trained on language modelling would mitigate this.", + "bbox": [ + 511, + 493, + 906, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 563, + 633, + 579 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In the age of large foundation models, the trend is towards larger, more unified, fully end-to-end models without task-specific components. We introduced a novel multimodal learning framework that reformulates various multimodal tasks into a unified next-frame prediction paradigm. This approach addresses critical limitations in current multimodal model designs, which often require modality-specific encoders and are limited in scalability and flexibility when adapting to new tasks. By unifying diverse multimodal tasks under a single framework, our model can handle text, image, audio, and video inputs without modality-specific components, significantly simplifying the design and training process. We show it is possible to train a simple next-frame prediction model to solve these tasks, with many of them matching the performance of comparable models trained without pretraining data.", + "bbox": [ + 509, + 593, + 906, + 835 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ultimately, our work lays the groundwork for building more generalized and efficient multimodal foundation models, towards systems that can process information from diverse modalities in a unified and scalable way.", + "bbox": [ + 511, + 839, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "22011", + "bbox": [ + 478, + 944, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Lakshita Agarwal and Bindu Verma. From methods to datasets: A survey on image-caption generators. Multimedia Tools and Applications, 83:28077-28123, 2023. 1", + "[2] Alexei Baevski, Wei-Ning Hsu, Quantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. ArXiv, abs/2202.03555, 2022. 2", + "[3] Soren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian Lapuschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute, 361(1):418-428, 2024. 4, 6, 7", + "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5", + "[5] Tsung-Han Chan, Kui Jia, Shenghua Gao, Jiwen Lu, Zinan Zeng, and Yi Ma. Pcanet: A simple deep learning baseline for image classification? IEEE transactions on image processing, 24(12):5017-5032, 2015. 6, 7", + "[6] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3", + "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Learning universal image-text representations. ArXiv, abs/1909.11740, 2019. 2", + "[8] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 6", + "[9] Ugur Demir, Yogesh S Rawat, and Mubarak Shah. Tinyvirat: Low-resolution video action recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 7387-7394. IEEE, 2021. 4, 7", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 6, 7", + "[11] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. 4", + "[12] Tianyu Gao, Zirui Wang, Adithya Bhaskar, and Danqi Chen. Improving language understanding from screenshots. arXiv preprint arXiv:2402.14073, 2024. 3", + "[13] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 3", + "[14] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. 2021 IEEE/CVF" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "International Conference on Computer Vision (ICCV), pages 1419-1429, 2021. 2", + "[15] Yunseok Jang, Yale Song, Chris Dongjoo Kim, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Video Question Answering with Spatio-Temporal Reasoning. IJCV, 2019. 4, 6", + "[16] Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. Unifying question answering, text classification, and regression via span extraction. arXiv preprint arXiv:1904.09286, 2019. 2", + "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3", + "[18] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. Computational Visual Media, 10(2):375–395, 2024. 6", + "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5", + "[20] Siyu Lu, Mingzhe Liu, Lirong Yin, Zhengtong Yin, Xuan Liu, and Wenfeng Zheng. The multi-modal fusion in visual question answering: a review of attention mechanisms. PeerJ Computer Science, 9:e1400, 2023. 1", + "[21] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018. 2", + "[22] Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 2, 5", + "[23] OpenAI. Gpt-4 technical report, 2024. 2", + "[24] Marc'Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014. 3", + "[25] Phillip Rust, Jonas F Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Eliott. Language modelling with pixels. arXiv preprint arXiv:2207.06991, 2022. 3", + "[26] Min seok Seo, Hakjin Lee, Do-Yeon Kim, and Junghoon Seo. Implicit stacked autoregressive model for video prediction. ArXiv, abs/2303.07849, 2023. 3", + "[27] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 2", + "[28] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642, 2013. 3", + "[29] Yintao Tai, Xiyang Liao, Alessandro Suglia, and Antonio Vergari. Pixar: Auto-regressive language modeling in pixel space. arXiv preprint arXiv:2401.03321, 2024. 3" + ], + "bbox": [ + 517, + 93, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "22012", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Joost R. van Amersfoort, Anitha Kannan, Marc'Aurelio Ranzato, Arthur Szlam, Du Tran, and Soumith Chintala. Transformation-based models of video sequences. ArXiv, abs/1701.08435, 2017. 3", + "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8296–8305, 2024. 1", + "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[33] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1", + "[34] Chenghao Xiao, Zhuoxu Huang, Danlu Chen, G Thomas Hudson, Yizhi Li, Haoran Duan, Chenghua Lin, Jie Fu, Jungong Han, and Noura Al Moubayed. Pixel sentence representation learning. arXiv preprint arXiv:2402.08183, 2024. 3", + "[35] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers, 2021. 3", + "[36] Yixin Yang, Jiangxin Dong, Jinhui Tang, and Pan Jinshan. Colormapnet: A memory-based deep spatial-temporal feature propagation network for video colorization. In ECCV, 2024. 6", + "[37] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 4, 6, 7", + "[38] M. Akin Yilmaz and Ahmet Murat Tekalp. Dfpn: Deformable frame prediction network. 2021 IEEE International Conference on Image Processing (ICIP), pages 1944-1948, 2021. 3", + "[39] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 6", + "[40] Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint arXiv:2302.09268, 2023. 6, 7", + "[41] Yeyun Zou and Qiyu Xie. A survey on vqa: Datasets and approaches. In 2020 2nd International Conference on Information Technology and Computer Application (ITCA), pages 289-297. IEEE, 2020. 1" + ], + "bbox": [ + 91, + 92, + 482, + 768 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "22013", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_model.json b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_model.json new file mode 100644 index 0000000000000000000000000000000000000000..68946c9876bd52b19e4d039be2a12d28e2f1b626 --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_model.json @@ -0,0 +1,1936 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.121, + 0.131, + 0.877, + 0.152 + ], + "angle": 0, + "content": "Everything is a Video: Unifying Modalities through Next-Frame Prediction" + }, + { + "type": "text", + "bbox": [ + 0.171, + 0.181, + 0.331, + 0.198 + ], + "angle": 0, + "content": "G. Thomas Hudson" + }, + { + "type": "text", + "bbox": [ + 0.365, + 0.182, + 0.458, + 0.198 + ], + "angle": 0, + "content": "Dean Slack" + }, + { + "type": "text", + "bbox": [ + 0.494, + 0.182, + 0.674, + 0.198 + ], + "angle": 0, + "content": "Thomas Winterbottom" + }, + { + "type": "text", + "bbox": [ + 0.708, + 0.182, + 0.826, + 0.199 + ], + "angle": 0, + "content": "Jamie Sterling" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.2, + 0.409, + 0.216 + ], + "angle": 0, + "content": "Chenghao Xiao" + }, + { + "type": "text", + "bbox": [ + 0.444, + 0.2, + 0.554, + 0.217 + ], + "angle": 0, + "content": "Junjie Shentu" + }, + { + "type": "text", + "bbox": [ + 0.588, + 0.2, + 0.755, + 0.217 + ], + "angle": 0, + "content": "Noura Al Moubayed" + }, + { + "type": "text", + "bbox": [ + 0.282, + 0.234, + 0.715, + 0.253 + ], + "angle": 0, + "content": "Department of Computer Sciences, Durham University" + }, + { + "type": "text", + "bbox": [ + 0.395, + 0.256, + 0.602, + 0.269 + ], + "angle": 0, + "content": "g.t.hudson@durham.ac.uk" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.279, + 0.836, + 0.464 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.468, + 0.908, + 0.497 + ], + "angle": 0, + "content": "Figure 1. Sample outputs from our next frame prediction model across multiple modalities. The model receives frames to the left of the red line as input and everything right of the red line has been generated by the model." + }, + { + "type": "title", + "bbox": [ + 0.269, + 0.51, + 0.347, + 0.525 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.541, + 0.504, + 0.89 + ], + "angle": 0, + "content": "Multimodal learning, which involves integrating information from various modalities such as text, images, audio, and video, is pivotal for numerous complex tasks like visual question answering, cross-modal retrieval, and caption generation. Traditional approaches rely on modality-specific encoders and late fusion techniques, which can hinder flexibility when adapting to new tasks or modalities. To address these limitations, we introduce a novel framework that extends the concept of task reformulation beyond natural language processing (NLP) to multimodal learning. We reformulate diverse multimodal tasks into a unified next-frame prediction problem, allowing a single model to handle different modalities without modality-specific components. This method treats all inputs and outputs as sequential frames in a video, enabling seamless integration of modalities and effective knowledge transfer across tasks. Our approach is evaluated on a range of tasks, including text-to-text, image-to-text, video-to-video, video-to-text, and audio-to-text, demonstrating the model's ability to generalize across modalities. We show that task reformulation can significantly simplify multimodal model design across various tasks, laying the groundwork for more generalized multimodal foundation models." + }, + { + "type": "title", + "bbox": [ + 0.515, + 0.51, + 0.645, + 0.525 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.907, + 0.702 + ], + "angle": 0, + "content": "Many tasks we seek to solve are multimodal. The ability to process and integrate information from multiple modalities such as text, image, audio, and video, have proven crucial for a range of complex tasks, including cross-modal retrieval [31], visual question answering [20, 33, 41], and caption generation [1]. The ability to understand and reason about information from different sensory inputs simultaneously mirrors human perception and cognition, making multimodal models an essential component in advancing artificial intelligence towards more generalized and versatile capabilities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The predominant approach to building multimodal models involves using separate encoders for each modality. For example, in vision-and-language models, a convolutional neural network (CNN) or a vision transformer (ViT) [10] might be used to encode images, while a recurrent neural network (RNN) or transformer is used to encode text. These modality-specific encoders are then fused at a later stage, typically via concatenation or attention mechanisms, to enable the model to perform joint reasoning over the encoded representations. While effective, this paradigm has inherent limitations. It necessitates careful design choices for each modality and often struggles with scalability and flexibility, especially when extending to new combinations of modali" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22004" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.336, + 0.106 + ], + "angle": 0, + "content": "ties or tasks not seen during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.107, + 0.482, + 0.317 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel approach centered around task reformulation. Task reformulation is a well-established paradigm in natural language processing (NLP) that facilitates multitask learning by transforming diverse tasks into a common format [21]. A prominent example is prompt-based learning, where NLP tasks are reimagined as instructions or prompts to a large language model (LLM). This paradigm allows a wide range of NLP tasks, such as translation, summarization, and question answering, to be represented uniformly in a chat message format [23]. By doing so, it leverages the generalization capabilities of large pre-trained models, enabling them to perform multiple tasks with minimal task-specific customization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.483, + 0.558 + ], + "angle": 0, + "content": "In this work, we extend the concept of task reformulation beyond NLP to the realm of multimodal learning. Specifically, we propose a framework in which tasks across various modalities are reformulated as a unified next-frame prediction problem. This approach simplifies the design of multimodal models by enabling them to handle diverse input types — text, images, audio, or video using a single, coherent mechanism. By representing all tasks as next-frame prediction, we create a shared interface for the model to learn and reason about information across modalities. This not only facilitates the integration of new modalities with minimal effort but also enhances the model's ability to transfer knowledge across different tasks and domains. Ultimately, just as LLMs function as foundation models for NLP, we intend this framing to be a foundation model across all modalities." + }, + { + "type": "text", + "bbox": [ + 0.109, + 0.56, + 0.451, + 0.574 + ], + "angle": 0, + "content": "The main contributions of this paper are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.577, + 0.48, + 0.619 + ], + "angle": 0, + "content": "- We propose a task-reformulation method and show how to represent a range of multimodal tasks as next-frame prediction." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.622, + 0.48, + 0.679 + ], + "angle": 0, + "content": "- We explore how this reformulation allows a single transformer-based model to solve text, image, video, and audio processing tasks without any modality-specific encoders." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.683, + 0.48, + 0.724 + ], + "angle": 0, + "content": "- Experimental results show our model achieves comparable performance to single-task models which haven't been pretrained on additional data." + }, + { + "type": "list", + "bbox": [ + 0.092, + 0.577, + 0.48, + 0.724 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.74, + 0.228, + 0.755 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Multimodal Learning With the rise of transformers, LLMs, and very large scale pretraining, new paradigms for harmonising different modalities have been pushed to the forefront of multimodal learning. UNITER [7] uses a single universal transformer to process both image and text inputs into a join embedding space, with image representations pretended to the beginning of the input embeddings to the transformer. However, the image inputs are first processed by an RNN to obtain an embedding such that it is" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.56 + ], + "angle": 0, + "content": "compatible with the textual inputs. Our framework in contrast has no need to preprocess input modalities differently as they are unified into a single visual input space theoretically capable of fully representing all multimodal interactions. Baevski et al. [2] propose data2vec, a framework for instilling features from multiple modalities into a single latent representation. Where data2vec seeks to process raw text, images, and audio using the same method, our work here reformulates the task inputs from multiple modalities into the same visual input medium. UniT [14], a multimodal multitask transformer uses separate encoders for each modality followed by a shared decoder. Our work uses a single visual encoder for our unified visual input paradigm. The most powerful state-of-the-art multimodal models such as FLAVA [27] and GPT-4 vision [23] integrate vision inputs via a patchwise process pipeline through a visual transformer fully integrated into the training pipeline. Both the architecture design and scale at which these foundation models are trained yield revolutionary performance at a variety of open-ended multimodal tasks. Nonetheless, though these foundation model are the closest that benchmarks have practically come to seamlessly integrating images with visual inputs, fundamentally the input modalities are handled differently by the model. Our work here bridges this remaining gap in processing of separate modalities. We reformulate typically spura-visual tasks into a visual input domain with no loss of information i.e. text and audio can be fully represented in images. This reformulation allows us to —for the first time—explore the generative pretraining capacity of multimodal information in one truly unified representation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.579, + 0.905, + 0.683 + ], + "angle": 0, + "content": "Task Reformulation Task reformulation involves allowing a single model to solve multiple tasks be converting them into a single 'supertask'. This technique is common in NLP where a range of tasks such as sentiment classification, and coreference resolution have been reformulated as span-extraction [16], and question answering [21] among others." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.685, + 0.905, + 0.835 + ], + "angle": 0, + "content": "The recent trend of prompt-based learning uses language modelling on prompt-answer sequences as the 'supertask' [23]. In this paradigm, the model is trained to follow an instruction describing the task and provide a response. In this way, any NLP task can be reformulated as the language modelling 'supertask'. Similar ideas have been explored in computer vision, notably the Transframer [22] which aimed to unify visual tasks. In our work, we extend this concept, exploring how next-frame prediction can be used as a 'supertask' to unite multiple modalities." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Next-frame Prediction CNN-based deep learning architectures for the autoregressive generation of future video frames have been steadily improving over the last decade" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "22005" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "[24, 26, 30, 38]. Transformer-based models trained at scale such as VideoGPT [35] have substantially improved the performance of quality and fidelity of long and short term future frame predictions. Modern diffusion-based frame prediction models are now able to generate photorealistic outputs [13]. The aim of our work is not a new model architecture. Our model is inspired from state-of-the-art architectures to serve as an appropriately powerful benchmark for our analysis on our benchmark datasets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.25, + 0.483, + 0.476 + ], + "angle": 0, + "content": "Visual Representation of Text Inspired by the success of unsupervised next-word prediction with language models, learning the next pixel of images was proposed [6] and shown able to learn image representations. Learning textual semantics with vision models, by rendering texts into images, has drawn more and more attention [12, 25, 29] due to the drawbacks of tokenization in traditional language models and limitations in cross-lingual transferability. Xiao et al. [34] show that vision models pretrained on rendered images have stronger robustness on typo and word-order shuffling perturbations, and the representations of them display better isotropy on out-of-distribution (OOD) languages. In this work, we extend this approach when reformulating NLP tasks to video as well as other modalities which output text." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.492, + 0.182, + 0.508 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.483, + 0.595 + ], + "angle": 0, + "content": "Our proposed framework consists of two key components: (1) Methods of reformulating a diverse range of input and output modalities into the single task of next-frame prediction; (2) A pure transformer-based model architecture. We introduce these components in the following sections." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.24, + 0.62 + ], + "angle": 0, + "content": "3.1. Reformulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.483, + 0.809 + ], + "angle": 0, + "content": "The datasets we use are carefully selected to cover a diverse range of input and output modalities, illustrating the versatility of our approach in reformulating various tasks as next-frame prediction problems. Our selection spans tasks from text and image classification to more complex audio, video, and multimodal tasks, enabling us to evaluate the model's ability to generalize across different data types and task requirements. By converting each task into a uniform format—a \\(64 \\times 64\\) RGB video sequence—we create a consistent framework where the model treats every input and output as sequential frames, simplifying the multimodal learning process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.483, + 0.87 + ], + "angle": 0, + "content": "For each task, we insert a separator token \\((|)\\) between the input and output frames to clearly delineate where the input ends and the prediction begins, guiding the model in distinguishing between these two parts of the sequence." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "In tasks involving textual data, we use a simple tokenizer (splitting on spaces and also on punctuation) to break" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.273 + ], + "angle": 0, + "content": "down the text into individual tokens, which we then render as video frames in the sequence. Each token is represented in a consistent format: a fixed-width font scaled to fill the \\(64 \\times 64\\) frame, ensuring clarity and compatibility with our video-based input structure. This approach allows the model to read and predict text as it would any other frame, effectively bypassing the need for traditional text-based tokenization while integrating text seamlessly with other modalities. By converting each token to a visual format, we enable cross-modal knowledge transfer, allowing the model to process text, images, and other modalities through a shared, frame-based learning paradigm." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.291, + 0.905, + 0.398 + ], + "angle": 0, + "content": "Text-to-Text The SST2 (Stanford Sentiment Treebank 2) dataset is a widely-used benchmark for sentiment classification, consisting of thousands of movie review excerpts labelled with binary sentiment labels (positive or negative) [28]. We use the text-encoding method described above, rendering each token as a video frame using a fixed-width font (Figure 2)." + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.41, + 0.887, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.46, + 0.905, + 0.489 + ], + "angle": 0, + "content": "Figure 2. Truncated example of the SST2 sentiment dataset rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.63 + ], + "angle": 0, + "content": "Image-to-Text As a simple test of image recognition ability, we employ the CIFAR-10 dataset. CIFAR-10 is a benchmark for image classification and consists of 60,000 color images, each with dimensions of \\(32 \\times 32\\) pixels [17]. These images are equally divided into ten distinct, mutually exclusive classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck." + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.642, + 0.808, + 0.692 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.704, + 0.905, + 0.744 + ], + "angle": 0, + "content": "Figure 3. Truncated example of the CIFAR10 image classification dataset rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.762, + 0.905, + 0.853 + ], + "angle": 0, + "content": "In our reformulation approach, each CIFAR-10 image is resized to \\(64 \\times 64\\) pixels to fit our standardized input frame size. Following the resized image, we insert a separator frame, which visually signifies the end of the input and the beginning of the output sequence, followed by a frame containing the class label text rendered as an image (Figure 3)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Video-to-Text We utilize the TinyVIRAT action classification dataset to rigorously assess the model's capacity for" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22006" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "understanding video content [9]. The TinyVIRAT dataset comprises of 7,663 training and 5,166 testing examples, each annotated with one or more of 26 distinct action labels that describe the activities depicted within low-resolution video sequences. This dataset is particularly noteworthy for its multi-label nature, allowing for nuanced action recognition; for instance, a video sequence might feature a man walking while carrying a backpack, which would be simultaneously labelled with both \"walking\" and \"carrying\" actions. This characteristic enables the evaluation of the model's ability to recognize and differentiate between overlapping actions occurring in dynamic environments." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.274, + 0.485, + 0.351 + ], + "angle": 0, + "content": "To reformulate the task for next-frame prediction, we reformulate each video sequence by first presenting the input video frames, followed by a separator token, and concluding with a comma-separated list of action labels that correspond to the activities depicted in the video. (Figure 4)." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.364, + 0.484, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.409, + 0.484, + 0.451 + ], + "angle": 0, + "content": "Figure 4. Truncated example of the Tinyvirat video classification dataset rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.703 + ], + "angle": 0, + "content": "Video+Text-to-Text To explore the integration of multiple modalities within a single input, we leverage the CLEVRER (CoLlision Events for Video REpresentation and Reasoning) dataset, which is designed for video question answering (VQA) [37]. This dataset comprises synthetic video sequences depicting rendered 3D objects interacting through collisions, movements, and occlusions within a simple, controlled environment. Each video is paired with questions that require the model to understand and reason about these visual events. The questions in CLEVRER span various reasoning tasks, including descriptive, explanatory, and predictive questions about the events in the video, such as identifying specific objects, understanding object interactions, and predicting future states." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.825 + ], + "angle": 0, + "content": "In our framework, we represent each question and corresponding video as a unified sequence of frames. Each input sequence begins with frames of the CLEVRER video sequence subsampled to 4 frames. This is followed by the question text, separator token, and answer text encoded in the manner described previously (Figure 5). Similarly, we use the frame QA task from TGIF-QA dataset[15], containing animated gifs with QA pairs." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To keep the sequence length manageable, we exclude complex counterfactual questions and multiple-choice answers from the dataset (only using the 'descriptive' subset), as these would require additional frames and could lead to overly long sequences that complicate training." + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.089, + 0.906, + 0.119 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.129, + 0.907, + 0.158 + ], + "angle": 0, + "content": "Figure 5. Truncated example of the CLEVRR task rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.189, + 0.907, + 0.28 + ], + "angle": 0, + "content": "Video-to-Video We employ two tasks to test video-to-video performance. Firstly, we convert the TinyVIRAT videos into greyscale and test the model's ability to colourise the output. The sequence consists of the greyscale-converted video sequence, the separator frame, then the original RGB video sequence (Figure 6)." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.298, + 0.887, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.349, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Figure 6. Truncated example of the colorization task rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.399, + 0.907, + 0.535 + ], + "angle": 0, + "content": "Additionally, we test the model's ability to perform object tracking. The Large-scale Single Object Tracking dataset (LaSOT) consists of 1,550 video sequences hand-annotated with bounding boxes following an object of interest [11]. We reformulate this task by converting it into a sequence which consists of the first frame of the video with the bounding box overlayed, the rest of the sequence, the separator frame, and finally the full sequence including the overlayed bounding box (Figure 7)." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.552, + 0.887, + 0.596 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.606, + 0.906, + 0.636 + ], + "angle": 0, + "content": "Figure 7. Truncated example of the LaSOT object tracking task rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.687, + 0.906, + 0.807 + ], + "angle": 0, + "content": "Audio-to-Text To demonstrate the ability to process audio data, we utilize the AudioMNIST dataset, which contains 30,000 audio recordings of spoken digits ranging from zero to nine [3]. These recordings are spoken by multiple speakers with varying accents and intonations, providing a diverse dataset that challenges the model to recognize and classify spoken language across different vocal characteristics." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In our framework, each audio sample is preprocessed into a spectrogram, which is used at the first frame in the video sequence, representing the audio content visually for consistency with other modalities. Following the spectrogram, a separator token frame is added, and the label digit is represented in the final frame as text (Figure 8)." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22007" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.19, + 0.089, + 0.388, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.15, + 0.483, + 0.193 + ], + "angle": 0, + "content": "Figure 8. Example of the AudioMNIST audio classification dataset rendered as a video. Each square is a frame of the video sequence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.22, + 0.31, + 0.236 + ], + "angle": 0, + "content": "3.2. Video Prediction Model" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.244, + 0.485, + 0.651 + ], + "angle": 0, + "content": "We perform all tasks using a pure transformer-based model architecture. Building upon the successes of the ViT for image classification and TimesFormer [4] for video understanding, we modify these architectures to accommodate for spatiotemporal modelling of video sequences and autoregressive video prediction. We further adapt the model to progressively process patches at smaller resolutions, similar to that of U-net style models, such that adjacent input patches are linearly merged to reduce the effective image resolution at each block on the encoding side. This design choice allows us to train end-to-end models efficiently without resorting to pretrained encoder/decoder models. As illustrated in Fig. 9, partitioning into patches is followed by a series of local space-time transformed blocks separated by patch merge operations, with a global space-time transformer block operating on the reduced resolution patch-based representations. This has the benefit of capturing multiscale spatial features, while reducing the computational complexity of the global self-attention operations due to a smaller number of image patches. The reverse process is applied on the decoding side of the model to produce a predicted next image frame for each input frame. During inference, the final predicted frame is concatenated to the input sequence in an autoregressive manner. We outline the key model components and considerations below. Our model uses 41M parameters (for comparison, Transframer [22] uses 470–829M parameters)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Patch Representation A core feature of the model is in the partitioning of input video frames into non-overlapping patches, similar to ViT. These patches are processed using local and global spatiotemporal self-attention mechanisms as described in [4], with spatial self-attention focusing on intra-frame patch relationships and temporal self-attention handling inter-frame causal dependencies. Temporal causality is enforced by masking future patches during training, ensuring that only past and present information is available for prediction. The model takes as input a video sequence \\(\\mathcal{X} = \\{x_{t}\\}_{t = 1}^{T}\\), such that each \\(x_{t} \\in \\mathbb{R}^{C \\times H \\times W}\\) is a \\(C\\)-channel image frame of height \\(H\\) and width \\(W\\) at timestep \\(t\\) for \\(t = 1,2,\\ldots,T\\). Frames are partitioned into \\(N\\) equally-sized non-overlapping patches, and linearly embedded to a 1D representation of size \\(K\\) (we set \\(K\\) to 512" + }, + { + "type": "table", + "bbox": [ + 0.524, + 0.089, + 0.903, + 0.214 + ], + "angle": 0, + "content": "
TaskInput LenTarget LenBatch Size
Text Classification21116
Image Classification21128
Video Classification11116
Audio Classification21256
Video QA21116
Object Tracking21208
Video Colorization111016
" + }, + { + "type": "table_caption", + "bbox": [ + 0.573, + 0.224, + 0.847, + 0.238 + ], + "angle": 0, + "content": "Table 1. Training configurations for each task" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.266, + 0.907, + 0.357 + ], + "angle": 0, + "content": "throughout). This embedding is combined with learnable spatial and temporal positional encodings to form the input sequence for the transformer. Spatial attention is applied between image patches within each frame independently, while causal temporal attention is applied across timesteps at the same spatial patch location." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.369, + 0.673, + 0.385 + ], + "angle": 0, + "content": "3.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.392, + 0.907, + 0.619 + ], + "angle": 0, + "content": "We train our model on each dataset independently, with no language model or image pretraining, in order to determine the feasibility of our task reformulation approach with no latent performance benefits from transferring existing learned representations. We train end-to-end with a Multi-Scale Structural Similarity Index Measure (SSIM) loss [32], with a constant learning rate of \\( 3 \\times 10^{-4} \\), and a batch size of 8-32 depending on the task context length. We train with the AdamW [19] optimizer using default parameters, and set dropout to 0.1 for all layers except the final output layer. All models are evaluated on checkpoints corresponding to the best SSIM validation performance, with all models trained on a single NVIDIA A100 for a maximum of 7 GPU days. The input/output lengths and batch sizes vary per task, and are detailed in Table 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.635, + 0.645, + 0.652 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.907, + 0.797 + ], + "angle": 0, + "content": "We train on each task in turn in a single-task manner (we leave training on all the tasks jointly in a multitask setting as future work). For tasks which output text, we perform OCR on the output frames using the open-source tesseract OCR engine1. For tasks which have a fixed vocabulary (e.g. classification tasks), the OCR text is matched to the closest vocabulary word to minimise any OCR-related errors. For classification tasks, the resulting text is measured using F1-score and Accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.798, + 0.906, + 0.844 + ], + "angle": 0, + "content": "The object tracking task is evaluated by extracting the bounding box in the outputted frames, then comparing these to the labelled boxes using Intersection over Union (IoU)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.845, + 0.906, + 0.875 + ], + "angle": 0, + "content": "For video colorization, we evaluate both colorization quality and temporal consistency. This dual evaluation is" + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.888, + 0.871, + 0.9 + ], + "angle": 0, + "content": "1https://github.com/tesseract-ocr/tesseract" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22008" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.175, + 0.088, + 0.825, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.228, + 0.908, + 0.314 + ], + "angle": 0, + "content": "Figure 9. Illustration of the Transformer-based video prediction model used in this work. \\( T \\) timestep video image frames are partitioned into \\( N \\) non-overlapping patches and linearly embedded before being processed by a series of local and global space-time transformer blocks. Local blocks are restricted to local spatial and temporal attention windows, separated by non-overlapping patch merge/unmerge operations with a patch window size of 2. Global blocks operate on a smaller effective image resolution (e.g., for an input of \\( 64 \\times 64 \\) and a patch size of \\( 8 \\times 8 \\), the effective resolution at the global blocks is \\( 16 \\times 16 \\)). Skip connections (dashed lines) are used between corresponding patch merge/unmerge operations to preserve information such as static background features." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.325, + 0.888, + 0.707 + ], + "angle": 0, + "content": "
TaskDatasetModelPSNR ↑SSIM ↑F1 ↑Acc ↑IoU ↑CDC ↓Colorfulness ↑
Text ClassificationSST-2Ours41.898.776.875.5---
BERT-base---81.2---
Vega v1 [40]†--91.3----
Image ClassificationCIFAR-10Ours45.795.989.189.1---
ViT-base--71.371.3---
Resnet-101--90.090.0---
PCANet [5]--77.177.1---
ViT-huge [10]†--99.599.5---
Video ClassificationTinyVIRATOurs50.197.374.1 (30.4*)60.4---
ResNet50†--29.1*----
WideResNet†--32.6*----
Audio ClassificationaudioMNISTOurs50.998.996.997.1---
AlexNet [3]---95.8---
Video QACLEVREROurs25.779.852.452.5--
LSTM [37]†---34.7---
+ CNN [37]†---51.8---
TGIF-QAOurs42.098.553.252.3---
LSTM + Attn [15]---51.9---
Object TrackingLaSOTOurs35.798.7--0.63--
Dai et al. [8]----0.60--
Video ColorizationTinyVIRATOurs42.499.7---0.0273.1
Zhang et al. [39]-----0.0282.2
Yang et al. [36]-----0.0372.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.716, + 0.908, + 0.775 + ], + "angle": 0, + "content": "Table 2. Performance metrics across various tasks using our proposed next-frame prediction framework. Metrics include Peak Signa-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), F1-score and Accuracy (F1, Acc), Intersection over Union (IoU), Color Distribution Consistency index (CDC), and Colorfulness, which are each tailored to evaluate specific task outputs. *macro F1-score to allow direct comparison with baseline models. †indicates models with pretraining on other datasets." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.799, + 0.482, + 0.89 + ], + "angle": 0, + "content": "crucial, as a model could appear consistent by merely reproducing the grayscale input. We use PSNR and the colorfulness metric from Liu et al. to assess color diversity. Temporal consistency is measured using the Color Distribution Consistency (CDC) index [18], which calculates the Jensen-Shannon (JS) divergence of color distributions" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.8, + 0.613, + 0.813 + ], + "angle": 0, + "content": "across frames:" + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.833, + 0.905, + 0.878 + ], + "angle": 0, + "content": "\\[\nC D C _ {t} = \\frac {1}{3 \\times (N - t)} \\sum_ {c \\in \\{r, g, b \\}} \\sum_ {i = 1} ^ {N - t} J S (P _ {c} (I ^ {i}), P _ {c} (I ^ {I + t}))\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.885, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\(P_{c}(I^{i})\\) is the normalized probability distribution" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22009" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.483, + 0.123 + ], + "angle": 0, + "content": "of channel \\(c\\) in frame \\(I^i\\) of \\(N\\) total frames. Overall consistency is computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.134, + 0.422, + 0.164 + ], + "angle": 0, + "content": "\\[\nC D C = \\frac {1}{3} \\left(C D C _ {1} + C D C _ {2} + C D C _ {4}\\right)\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.176, + 0.178, + 0.192 + ], + "angle": 0, + "content": "5. Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.201, + 0.483, + 0.247 + ], + "angle": 0, + "content": "The results across each task are presented in Table 2, while Figure 1 illustrates several sample outputs produced by our model, offering a qualitative perspective on its performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.248, + 0.483, + 0.429 + ], + "angle": 0, + "content": "For the SST-2 text classification task, our model achieves an F1-score of 76.8. Although this is lower than the current state-of-the-art performance of 91.3, that was achieved through the use of a \"Bag of Tricks\" methodology involving extensive pretraining and complex finetuning processes [40]. Our model is simply trained on the SST-2 dataset, with no use of pretrained weights or advanced external embeddings, making it comparable to other models used without pretraining (See the BERT-base result). Some portion of this lower score can also be attributed to truncating any inputs over 20 tokens (limiting the evaluation to shorter inputs 20 tokens or fewer increases the F1-score to 80.0)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.43, + 0.483, + 0.596 + ], + "angle": 0, + "content": "On the CIFAR-10 image classification task, our model attains an accuracy of 89.1. While this score is lower than fine-tuned ViT models leveraging pretrained weights, which reach an impressive accuracy of 99.5 [10], our model performs significantly better than simpler baseline models, such as PCANet, which achieves 77.1 accuracy [5]. This performance suggests that our reformulation approach is able to perform image classification tasks without any pretraining, surpassing older, simpler architectures while trailing behind state-of-the-art models that benefit from extensive pretraining." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.597, + 0.483, + 0.687 + ], + "angle": 0, + "content": "In multi-label image classification on the TinyVIRAT dataset, our approach successfully outputs multiple labels per instance. When compared to baselines, our model surpasses the ResNet50 baseline (29.1) in performance but does not reach the level of the WideResNet model (32.6) [9]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.824 + ], + "angle": 0, + "content": "On the AudioMNIST audio classification task, the model achieves an accuracy of 97.1, surpassing the AlexNet baseline score of 95.82 from the original dataset paper [3]. This notable improvement demonstrates the model's efficacy in audio-based classification tasks. Analysis of misclassifications reveals that the most frequent error involves confusion between the spoken digits \"four\" and \"five,\" suggesting areas for potential refinement in distinguishing phonetically similar sounds." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "For the CLEVRR VQA (Visual Question Answering) dataset, our model achieves an accuracy of 52.5 on descriptive questions. This performance surpasses both the LSTM (34.7) and LSTM+CNN (51.8) baselines, illustrating our model's ability to understand the relationship between" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "visual and textual information without separate encoder-decoder models [37]. It is important to note that due to hardware limitations, and the long question lengths included in the CLEVRER dataset, we subsampled the video to only include 4 frames of the moving scene, while also reducing the resolution to \\(64 \\times 64\\), significantly limiting the accuracy achievable by our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.199, + 0.905, + 0.334 + ], + "angle": 0, + "content": "In the LASOT object tracking task, the model reaches an intersection over union (IoU) of 0.63, demonstrating a consistent ability to track objects throughout video sequences. Analysis of tracking outputs reveals that the model effectively follows the object of interest across frames with accurately drawn bounding boxes. However, tracking accuracy diminishes slightly toward the end of longer sequences, with the borders of the overlaid bounding box eroding, likely due to autoregressive pixel-level error propagation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.335, + 0.905, + 0.501 + ], + "angle": 0, + "content": "On the colorization task, our model shows improved color diversity in the generated outputs compared to the ground truth (73.1 color diversity for our model versus 70.6 in the original dataset). However, it exhibits less consistency in color application across the video sequence (0.0169 consistency for our model versus 0.00522 in the original dataset). This trade-off between diversity and consistency suggests that while our approach introduces a more vibrant color palette, further refinement could improve temporal coherence, particularly for applications requiring uniform color continuity across frames." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.503, + 0.555, + 0.516 + ], + "angle": 0, + "content": "ch" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.528, + 0.671, + 0.545 + ], + "angle": 0, + "content": "5.1. Attention Maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.905, + 0.61 + ], + "angle": 0, + "content": "In order to examine what is being learned by our model, we visualize internal patch-wise attention scores both spatially (within a frame) and temporally (between frames) in Figure 10." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.905, + 0.748 + ], + "angle": 0, + "content": "In the TinyVIRAT video classification task, the model pays most attention to the object/person performing the action, as well as frames and pairs of frames which indicate the class. In the first example (top row), spatial attention focuses on image patch locations containing the edges of the car which suggest relative movement, and temporal attention is highest for the final two frames in which the vehicle moves position, indicating the label \"vehicle moving\" should be output." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.905, + 0.854 + ], + "angle": 0, + "content": "For colorization, the greatest attention is paid to the first frames of the input sequence, which we suggest allows the model to produce consistent coloring across the whole sequence (more attention is also paid to the first frame of the output sequence, presumably to maintain color consistency). Spatially, the model clearly attends to key moving objects and areas of distinct color." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In the CLEVRR examples, we see how the model is able to focus on question frames which contains key words needed to produce the output, such as \"color\" and \"metal\"." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "22010" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.207, + 0.071, + 0.795, + 0.426 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.43, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Figure 10. Attention visualisations showing where the model attends to both spatially (indicated by light/dark areas overlaid on the frame), and temporally (indicated by a colour scale above the sequence). Spatial attention scores are calculated by taking the patch-wise contribution per frame, averaged over all layers. Temporal attention scores are calculated by taking the temporal attention weights for each input frame when generating the target frames, averaged over all global space-time attention layers." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.483, + 0.555 + ], + "angle": 0, + "content": "Within the moving object scene frames, we see how the model spatially attends to objects and their trajectory paths, correctly identifying objects such as \"the metal cylinder\" and \"the sphere\"." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.483, + 0.66 + ], + "angle": 0, + "content": "Lastly, for sentiment text classification, we observe a similar pattern to attention maps in LLMs, where the model pays more attention to words with a strong emotive sentiment. In our model we see frames containing the words 'nightmare', 'painful', and 'dumb' exhibiting high temporal attention weightings, as they are strong indicators of the overall sentiment of the review." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.668, + 0.353, + 0.683 + ], + "angle": 0, + "content": "5.2. Limitations and Future work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "The aim of this work is not necessarily to surpass state-of-the-art performance on any of the tasks, but to show that unifying these modalities via reformulation is possible, without heavily compromising on performance relative to task-specific baselines. Future work should aim to scale this up, improving performance and tackling more complex tasks. We suggest that this can be done by mirroring the improvements gained from pretraining in NLP and computer vision by first training the model on unstructured tasks from each modality. Tasks such as language modelling and image captioning can be naturally reformulated as next-frame-prediction and included in our paradigm. This is especially important as a large portion of the current training time on tasks such as classification and question answering is ded" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.494, + 0.907, + 0.54 + ], + "angle": 0, + "content": "icated to learning to output words rather than solving the task itself. Beginning with a checkpoint already trained on language modelling would mitigate this." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.564, + 0.634, + 0.58 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.594, + 0.907, + 0.837 + ], + "angle": 0, + "content": "In the age of large foundation models, the trend is towards larger, more unified, fully end-to-end models without task-specific components. We introduced a novel multimodal learning framework that reformulates various multimodal tasks into a unified next-frame prediction paradigm. This approach addresses critical limitations in current multimodal model designs, which often require modality-specific encoders and are limited in scalability and flexibility when adapting to new tasks. By unifying diverse multimodal tasks under a single framework, our model can handle text, image, audio, and video inputs without modality-specific components, significantly simplifying the design and training process. We show it is possible to train a simple next-frame prediction model to solve these tasks, with many of them matching the performance of comparable models trained without pretraining data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Ultimately, our work lays the groundwork for building more generalized and efficient multimodal foundation models, towards systems that can process information from diverse modalities in a unified and scalable way." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.958 + ], + "angle": 0, + "content": "22011" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.116, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Lakshita Agarwal and Bindu Verma. From methods to datasets: A survey on image-caption generators. Multimedia Tools and Applications, 83:28077-28123, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.483, + 0.212 + ], + "angle": 0, + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Quantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. ArXiv, abs/2202.03555, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.215, + 0.484, + 0.282 + ], + "angle": 0, + "content": "[3] Soren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian Lapuschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute, 361(1):418-428, 2024. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.284, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.482, + 0.381 + ], + "angle": 0, + "content": "[5] Tsung-Han Chan, Kui Jia, Shenghua Gao, Jiwen Lu, Zinan Zeng, and Yi Ma. Pcanet: A simple deep learning baseline for image classification? IEEE transactions on image processing, 24(12):5017-5032, 2015. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.383, + 0.482, + 0.437 + ], + "angle": 0, + "content": "[6] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.439, + 0.482, + 0.493 + ], + "angle": 0, + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Learning universal image-text representations. ArXiv, abs/1909.11740, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.496, + 0.482, + 0.536 + ], + "angle": 0, + "content": "[8] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.538, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[9] Ugur Demir, Yogesh S Rawat, and Mubarak Shah. Tinyvirat: Low-resolution video action recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 7387-7394. IEEE, 2021. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.594, + 0.482, + 0.689 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.691, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[11] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.776, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[12] Tianyu Gao, Zirui Wang, Adithya Bhaskar, and Danqi Chen. Improving language understanding from screenshots. arXiv preprint arXiv:2402.14073, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[13] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. 2021 IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.116, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.12 + ], + "angle": 0, + "content": "International Conference on Computer Vision (ICCV), pages 1419-1429, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[15] Yunseok Jang, Yale Song, Chris Dongjoo Kim, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Video Question Answering with Spatio-Temporal Reasoning. IJCV, 2019. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.167, + 0.905, + 0.221 + ], + "angle": 0, + "content": "[16] Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. Unifying question answering, text classification, and regression via span extraction. arXiv preprint arXiv:1904.09286, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.255, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[18] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. Computational Visual Media, 10(2):375–395, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.326, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.369, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[20] Siyu Lu, Mingzhe Liu, Lirong Yin, Zhengtong Yin, Xuan Liu, and Wenfeng Zheng. The multi-modal fusion in visual question answering: a review of attention mechanisms. PeerJ Computer Science, 9:e1400, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.427, + 0.905, + 0.481 + ], + "angle": 0, + "content": "[21] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[22] Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.542, + 0.79, + 0.556 + ], + "angle": 0, + "content": "[23] OpenAI. Gpt-4 technical report, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.558, + 0.905, + 0.613 + ], + "angle": 0, + "content": "[24] Marc'Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.616, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[25] Phillip Rust, Jonas F Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Eliott. Language modelling with pixels. arXiv preprint arXiv:2207.06991, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.673, + 0.905, + 0.713 + ], + "angle": 0, + "content": "[26] Min seok Seo, Hakjin Lee, Do-Yeon Kim, and Junghoon Seo. Implicit stacked autoregressive model for video prediction. ArXiv, abs/2303.07849, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.717, + 0.905, + 0.771 + ], + "angle": 0, + "content": "[27] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.774, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[28] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Yintao Tai, Xiyang Liao, Alessandro Suglia, and Antonio Vergari. Pixar: Auto-regressive language modeling in pixel space. arXiv preprint arXiv:2401.03321, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22012" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[30] Joost R. van Amersfoort, Anitha Kannan, Marc'Aurelio Ranzato, Arthur Szlam, Du Tran, and Soumith Chintala. Transformation-based models of video sequences. ArXiv, abs/1701.08435, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.217 + ], + "angle": 0, + "content": "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8296–8305, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.33 + ], + "angle": 0, + "content": "[33] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.4 + ], + "angle": 0, + "content": "[34] Chenghao Xiao, Zhuoxu Huang, Danlu Chen, G Thomas Hudson, Yizhi Li, Haoran Duan, Chenghua Lin, Jie Fu, Jungong Han, and Noura Al Moubayed. Pixel sentence representation learning. arXiv preprint arXiv:2402.08183, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.443 + ], + "angle": 0, + "content": "[35] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.5 + ], + "angle": 0, + "content": "[36] Yixin Yang, Jiangxin Dong, Jinhui Tang, and Pan Jinshan. Colormapnet: A memory-based deep spatial-temporal feature propagation network for video colorization. In ECCV, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[37] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 4, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[38] M. Akin Yilmaz and Ahmet Murat Tekalp. Dfpn: Deformable frame prediction network. 2021 IEEE International Conference on Image Processing (ICIP), pages 1944-1948, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.616, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[39] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[40] Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint arXiv:2302.09268, 2023. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[41] Yeyun Zou and Qiyu Xie. A survey on vqa: Datasets and approaches. In 2020 2nd International Conference on Information Technology and Computer Application (ITCA), pages 289-297. IEEE, 2020. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.77 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "22013" + } + ] +] \ No newline at end of file diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_origin.pdf b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d82a7e03d77c709a6c31891ce70880e3ebfead9a --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/992f5003-7f3c-4061-aa0e-ae51513cab64_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9cd13e92acadc436def1e0c3f8db94277c693cb8423dd09438f5863ede84db05 +size 1561726 diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/full.md b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..4c58e0c7a4ca9fb4d2194fe3f6c382ec973a05df --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/full.md @@ -0,0 +1,266 @@ +# Everything is a Video: Unifying Modalities through Next-Frame Prediction + +G. Thomas Hudson + +Dean Slack + +Thomas Winterbottom + +Jamie Sterling + +Chenghao Xiao + +Junjie Shentu + +Noura Al Moubayed + +Department of Computer Sciences, Durham University + +g.t.hudson@durham.ac.uk + +![](images/fa0cbccea2015ca31ff32b5b015614b53a703ee11e6c36da55fec0b265aba43f.jpg) +Figure 1. Sample outputs from our next frame prediction model across multiple modalities. The model receives frames to the left of the red line as input and everything right of the red line has been generated by the model. + +# Abstract + +Multimodal learning, which involves integrating information from various modalities such as text, images, audio, and video, is pivotal for numerous complex tasks like visual question answering, cross-modal retrieval, and caption generation. Traditional approaches rely on modality-specific encoders and late fusion techniques, which can hinder flexibility when adapting to new tasks or modalities. To address these limitations, we introduce a novel framework that extends the concept of task reformulation beyond natural language processing (NLP) to multimodal learning. We reformulate diverse multimodal tasks into a unified next-frame prediction problem, allowing a single model to handle different modalities without modality-specific components. This method treats all inputs and outputs as sequential frames in a video, enabling seamless integration of modalities and effective knowledge transfer across tasks. Our approach is evaluated on a range of tasks, including text-to-text, image-to-text, video-to-video, video-to-text, and audio-to-text, demonstrating the model's ability to generalize across modalities. We show that task reformulation can significantly simplify multimodal model design across various tasks, laying the groundwork for more generalized multimodal foundation models. + +# 1. Introduction + +Many tasks we seek to solve are multimodal. The ability to process and integrate information from multiple modalities such as text, image, audio, and video, have proven crucial for a range of complex tasks, including cross-modal retrieval [31], visual question answering [20, 33, 41], and caption generation [1]. The ability to understand and reason about information from different sensory inputs simultaneously mirrors human perception and cognition, making multimodal models an essential component in advancing artificial intelligence towards more generalized and versatile capabilities. + +The predominant approach to building multimodal models involves using separate encoders for each modality. For example, in vision-and-language models, a convolutional neural network (CNN) or a vision transformer (ViT) [10] might be used to encode images, while a recurrent neural network (RNN) or transformer is used to encode text. These modality-specific encoders are then fused at a later stage, typically via concatenation or attention mechanisms, to enable the model to perform joint reasoning over the encoded representations. While effective, this paradigm has inherent limitations. It necessitates careful design choices for each modality and often struggles with scalability and flexibility, especially when extending to new combinations of modali + +ties or tasks not seen during training. + +To address these challenges, we propose a novel approach centered around task reformulation. Task reformulation is a well-established paradigm in natural language processing (NLP) that facilitates multitask learning by transforming diverse tasks into a common format [21]. A prominent example is prompt-based learning, where NLP tasks are reimagined as instructions or prompts to a large language model (LLM). This paradigm allows a wide range of NLP tasks, such as translation, summarization, and question answering, to be represented uniformly in a chat message format [23]. By doing so, it leverages the generalization capabilities of large pre-trained models, enabling them to perform multiple tasks with minimal task-specific customization. + +In this work, we extend the concept of task reformulation beyond NLP to the realm of multimodal learning. Specifically, we propose a framework in which tasks across various modalities are reformulated as a unified next-frame prediction problem. This approach simplifies the design of multimodal models by enabling them to handle diverse input types — text, images, audio, or video using a single, coherent mechanism. By representing all tasks as next-frame prediction, we create a shared interface for the model to learn and reason about information across modalities. This not only facilitates the integration of new modalities with minimal effort but also enhances the model's ability to transfer knowledge across different tasks and domains. Ultimately, just as LLMs function as foundation models for NLP, we intend this framing to be a foundation model across all modalities. + +The main contributions of this paper are as follows: + +- We propose a task-reformulation method and show how to represent a range of multimodal tasks as next-frame prediction. +- We explore how this reformulation allows a single transformer-based model to solve text, image, video, and audio processing tasks without any modality-specific encoders. +- Experimental results show our model achieves comparable performance to single-task models which haven't been pretrained on additional data. + +# 2. Related work + +Multimodal Learning With the rise of transformers, LLMs, and very large scale pretraining, new paradigms for harmonising different modalities have been pushed to the forefront of multimodal learning. UNITER [7] uses a single universal transformer to process both image and text inputs into a join embedding space, with image representations pretended to the beginning of the input embeddings to the transformer. However, the image inputs are first processed by an RNN to obtain an embedding such that it is + +compatible with the textual inputs. Our framework in contrast has no need to preprocess input modalities differently as they are unified into a single visual input space theoretically capable of fully representing all multimodal interactions. Baevski et al. [2] propose data2vec, a framework for instilling features from multiple modalities into a single latent representation. Where data2vec seeks to process raw text, images, and audio using the same method, our work here reformulates the task inputs from multiple modalities into the same visual input medium. UniT [14], a multimodal multitask transformer uses separate encoders for each modality followed by a shared decoder. Our work uses a single visual encoder for our unified visual input paradigm. The most powerful state-of-the-art multimodal models such as FLAVA [27] and GPT-4 vision [23] integrate vision inputs via a patchwise process pipeline through a visual transformer fully integrated into the training pipeline. Both the architecture design and scale at which these foundation models are trained yield revolutionary performance at a variety of open-ended multimodal tasks. Nonetheless, though these foundation model are the closest that benchmarks have practically come to seamlessly integrating images with visual inputs, fundamentally the input modalities are handled differently by the model. Our work here bridges this remaining gap in processing of separate modalities. We reformulate typically spura-visual tasks into a visual input domain with no loss of information i.e. text and audio can be fully represented in images. This reformulation allows us to —for the first time—explore the generative pretraining capacity of multimodal information in one truly unified representation. + +Task Reformulation Task reformulation involves allowing a single model to solve multiple tasks be converting them into a single 'supertask'. This technique is common in NLP where a range of tasks such as sentiment classification, and coreference resolution have been reformulated as span-extraction [16], and question answering [21] among others. + +The recent trend of prompt-based learning uses language modelling on prompt-answer sequences as the 'supertask' [23]. In this paradigm, the model is trained to follow an instruction describing the task and provide a response. In this way, any NLP task can be reformulated as the language modelling 'supertask'. Similar ideas have been explored in computer vision, notably the Transframer [22] which aimed to unify visual tasks. In our work, we extend this concept, exploring how next-frame prediction can be used as a 'supertask' to unite multiple modalities. + +Next-frame Prediction CNN-based deep learning architectures for the autoregressive generation of future video frames have been steadily improving over the last decade + +[24, 26, 30, 38]. Transformer-based models trained at scale such as VideoGPT [35] have substantially improved the performance of quality and fidelity of long and short term future frame predictions. Modern diffusion-based frame prediction models are now able to generate photorealistic outputs [13]. The aim of our work is not a new model architecture. Our model is inspired from state-of-the-art architectures to serve as an appropriately powerful benchmark for our analysis on our benchmark datasets. + +Visual Representation of Text Inspired by the success of unsupervised next-word prediction with language models, learning the next pixel of images was proposed [6] and shown able to learn image representations. Learning textual semantics with vision models, by rendering texts into images, has drawn more and more attention [12, 25, 29] due to the drawbacks of tokenization in traditional language models and limitations in cross-lingual transferability. Xiao et al. [34] show that vision models pretrained on rendered images have stronger robustness on typo and word-order shuffling perturbations, and the representations of them display better isotropy on out-of-distribution (OOD) languages. In this work, we extend this approach when reformulating NLP tasks to video as well as other modalities which output text. + +# 3. Method + +Our proposed framework consists of two key components: (1) Methods of reformulating a diverse range of input and output modalities into the single task of next-frame prediction; (2) A pure transformer-based model architecture. We introduce these components in the following sections. + +# 3.1. Reformulation + +The datasets we use are carefully selected to cover a diverse range of input and output modalities, illustrating the versatility of our approach in reformulating various tasks as next-frame prediction problems. Our selection spans tasks from text and image classification to more complex audio, video, and multimodal tasks, enabling us to evaluate the model's ability to generalize across different data types and task requirements. By converting each task into a uniform format—a $64 \times 64$ RGB video sequence—we create a consistent framework where the model treats every input and output as sequential frames, simplifying the multimodal learning process. + +For each task, we insert a separator token $(|)$ between the input and output frames to clearly delineate where the input ends and the prediction begins, guiding the model in distinguishing between these two parts of the sequence. + +In tasks involving textual data, we use a simple tokenizer (splitting on spaces and also on punctuation) to break + +down the text into individual tokens, which we then render as video frames in the sequence. Each token is represented in a consistent format: a fixed-width font scaled to fill the $64 \times 64$ frame, ensuring clarity and compatibility with our video-based input structure. This approach allows the model to read and predict text as it would any other frame, effectively bypassing the need for traditional text-based tokenization while integrating text seamlessly with other modalities. By converting each token to a visual format, we enable cross-modal knowledge transfer, allowing the model to process text, images, and other modalities through a shared, frame-based learning paradigm. + +Text-to-Text The SST2 (Stanford Sentiment Treebank 2) dataset is a widely-used benchmark for sentiment classification, consisting of thousands of movie review excerpts labelled with binary sentiment labels (positive or negative) [28]. We use the text-encoding method described above, rendering each token as a video frame using a fixed-width font (Figure 2). + +![](images/857b51dd8cf3d79d1619fea16da761d4bdd4874d67aa49984f8644ee41160899.jpg) +Figure 2. Truncated example of the SST2 sentiment dataset rendered as a video. Each square is a frame of the video sequence. + +Image-to-Text As a simple test of image recognition ability, we employ the CIFAR-10 dataset. CIFAR-10 is a benchmark for image classification and consists of 60,000 color images, each with dimensions of $32 \times 32$ pixels [17]. These images are equally divided into ten distinct, mutually exclusive classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck. + +![](images/208eee69929332ad58dec6c1f51a1510e52fa25e71973ce147570a935609e136.jpg) +Figure 3. Truncated example of the CIFAR10 image classification dataset rendered as a video. Each square is a frame of the video sequence. + +In our reformulation approach, each CIFAR-10 image is resized to $64 \times 64$ pixels to fit our standardized input frame size. Following the resized image, we insert a separator frame, which visually signifies the end of the input and the beginning of the output sequence, followed by a frame containing the class label text rendered as an image (Figure 3). + +Video-to-Text We utilize the TinyVIRAT action classification dataset to rigorously assess the model's capacity for + +understanding video content [9]. The TinyVIRAT dataset comprises of 7,663 training and 5,166 testing examples, each annotated with one or more of 26 distinct action labels that describe the activities depicted within low-resolution video sequences. This dataset is particularly noteworthy for its multi-label nature, allowing for nuanced action recognition; for instance, a video sequence might feature a man walking while carrying a backpack, which would be simultaneously labelled with both "walking" and "carrying" actions. This characteristic enables the evaluation of the model's ability to recognize and differentiate between overlapping actions occurring in dynamic environments. + +To reformulate the task for next-frame prediction, we reformulate each video sequence by first presenting the input video frames, followed by a separator token, and concluding with a comma-separated list of action labels that correspond to the activities depicted in the video. (Figure 4). + +![](images/d327b0bd614e4a5ceef757a7e35dd786cf82e0b950ca02cd4998d2f339dc653a.jpg) +Figure 4. Truncated example of the Tinyvirat video classification dataset rendered as a video. Each square is a frame of the video sequence. + +Video+Text-to-Text To explore the integration of multiple modalities within a single input, we leverage the CLEVRER (CoLlision Events for Video REpresentation and Reasoning) dataset, which is designed for video question answering (VQA) [37]. This dataset comprises synthetic video sequences depicting rendered 3D objects interacting through collisions, movements, and occlusions within a simple, controlled environment. Each video is paired with questions that require the model to understand and reason about these visual events. The questions in CLEVRER span various reasoning tasks, including descriptive, explanatory, and predictive questions about the events in the video, such as identifying specific objects, understanding object interactions, and predicting future states. + +In our framework, we represent each question and corresponding video as a unified sequence of frames. Each input sequence begins with frames of the CLEVRER video sequence subsampled to 4 frames. This is followed by the question text, separator token, and answer text encoded in the manner described previously (Figure 5). Similarly, we use the frame QA task from TGIF-QA dataset[15], containing animated gifs with QA pairs. + +To keep the sequence length manageable, we exclude complex counterfactual questions and multiple-choice answers from the dataset (only using the 'descriptive' subset), as these would require additional frames and could lead to overly long sequences that complicate training. + +![](images/cfa90a4f0fa41cb12efc8e8c2e93e7fa898a267af929e40de053e1cfc73e0576.jpg) +Figure 5. Truncated example of the CLEVRR task rendered as a video. Each square is a frame of the video sequence. + +Video-to-Video We employ two tasks to test video-to-video performance. Firstly, we convert the TinyVIRAT videos into greyscale and test the model's ability to colourise the output. The sequence consists of the greyscale-converted video sequence, the separator frame, then the original RGB video sequence (Figure 6). + +![](images/b1fc20fc80ac61fd4f40f2f050872c0149be33646840806e3bf55f0ffa733c8c.jpg) +Figure 6. Truncated example of the colorization task rendered as a video. Each square is a frame of the video sequence. + +Additionally, we test the model's ability to perform object tracking. The Large-scale Single Object Tracking dataset (LaSOT) consists of 1,550 video sequences hand-annotated with bounding boxes following an object of interest [11]. We reformulate this task by converting it into a sequence which consists of the first frame of the video with the bounding box overlayed, the rest of the sequence, the separator frame, and finally the full sequence including the overlayed bounding box (Figure 7). + +![](images/d69b9847d3f28d7759b7472286a44efed2fc8f7c77d86d624d982b4d8e451b66.jpg) +Figure 7. Truncated example of the LaSOT object tracking task rendered as a video. Each square is a frame of the video sequence. + +Audio-to-Text To demonstrate the ability to process audio data, we utilize the AudioMNIST dataset, which contains 30,000 audio recordings of spoken digits ranging from zero to nine [3]. These recordings are spoken by multiple speakers with varying accents and intonations, providing a diverse dataset that challenges the model to recognize and classify spoken language across different vocal characteristics. + +In our framework, each audio sample is preprocessed into a spectrogram, which is used at the first frame in the video sequence, representing the audio content visually for consistency with other modalities. Following the spectrogram, a separator token frame is added, and the label digit is represented in the final frame as text (Figure 8). + +![](images/39f555cdaf3142592099cfb4a1ad63de0c0043d8fdd139fd4873e89d91444f9b.jpg) +Figure 8. Example of the AudioMNIST audio classification dataset rendered as a video. Each square is a frame of the video sequence. + +# 3.2. Video Prediction Model + +We perform all tasks using a pure transformer-based model architecture. Building upon the successes of the ViT for image classification and TimesFormer [4] for video understanding, we modify these architectures to accommodate for spatiotemporal modelling of video sequences and autoregressive video prediction. We further adapt the model to progressively process patches at smaller resolutions, similar to that of U-net style models, such that adjacent input patches are linearly merged to reduce the effective image resolution at each block on the encoding side. This design choice allows us to train end-to-end models efficiently without resorting to pretrained encoder/decoder models. As illustrated in Fig. 9, partitioning into patches is followed by a series of local space-time transformed blocks separated by patch merge operations, with a global space-time transformer block operating on the reduced resolution patch-based representations. This has the benefit of capturing multiscale spatial features, while reducing the computational complexity of the global self-attention operations due to a smaller number of image patches. The reverse process is applied on the decoding side of the model to produce a predicted next image frame for each input frame. During inference, the final predicted frame is concatenated to the input sequence in an autoregressive manner. We outline the key model components and considerations below. Our model uses 41M parameters (for comparison, Transframer [22] uses 470–829M parameters). + +Patch Representation A core feature of the model is in the partitioning of input video frames into non-overlapping patches, similar to ViT. These patches are processed using local and global spatiotemporal self-attention mechanisms as described in [4], with spatial self-attention focusing on intra-frame patch relationships and temporal self-attention handling inter-frame causal dependencies. Temporal causality is enforced by masking future patches during training, ensuring that only past and present information is available for prediction. The model takes as input a video sequence $\mathcal{X} = \{x_{t}\}_{t = 1}^{T}$ , such that each $x_{t} \in \mathbb{R}^{C \times H \times W}$ is a $C$ -channel image frame of height $H$ and width $W$ at timestep $t$ for $t = 1,2,\ldots,T$ . Frames are partitioned into $N$ equally-sized non-overlapping patches, and linearly embedded to a 1D representation of size $K$ (we set $K$ to 512 + +
TaskInput LenTarget LenBatch Size
Text Classification21116
Image Classification21128
Video Classification11116
Audio Classification21256
Video QA21116
Object Tracking21208
Video Colorization111016
+ +Table 1. Training configurations for each task + +throughout). This embedding is combined with learnable spatial and temporal positional encodings to form the input sequence for the transformer. Spatial attention is applied between image patches within each frame independently, while causal temporal attention is applied across timesteps at the same spatial patch location. + +# 3.3. Training Details + +We train our model on each dataset independently, with no language model or image pretraining, in order to determine the feasibility of our task reformulation approach with no latent performance benefits from transferring existing learned representations. We train end-to-end with a Multi-Scale Structural Similarity Index Measure (SSIM) loss [32], with a constant learning rate of $3 \times 10^{-4}$ , and a batch size of 8-32 depending on the task context length. We train with the AdamW [19] optimizer using default parameters, and set dropout to 0.1 for all layers except the final output layer. All models are evaluated on checkpoints corresponding to the best SSIM validation performance, with all models trained on a single NVIDIA A100 for a maximum of 7 GPU days. The input/output lengths and batch sizes vary per task, and are detailed in Table 1. + +# 4. Experiments + +We train on each task in turn in a single-task manner (we leave training on all the tasks jointly in a multitask setting as future work). For tasks which output text, we perform OCR on the output frames using the open-source tesseract OCR engine1. For tasks which have a fixed vocabulary (e.g. classification tasks), the OCR text is matched to the closest vocabulary word to minimise any OCR-related errors. For classification tasks, the resulting text is measured using F1-score and Accuracy. + +The object tracking task is evaluated by extracting the bounding box in the outputted frames, then comparing these to the labelled boxes using Intersection over Union (IoU). + +For video colorization, we evaluate both colorization quality and temporal consistency. This dual evaluation is + +![](images/39e907a353976dd30fc2df5f811125c3af6de5ab05ea6a7a482c1eb0f8d724bb.jpg) +Figure 9. Illustration of the Transformer-based video prediction model used in this work. $T$ timestep video image frames are partitioned into $N$ non-overlapping patches and linearly embedded before being processed by a series of local and global space-time transformer blocks. Local blocks are restricted to local spatial and temporal attention windows, separated by non-overlapping patch merge/unmerge operations with a patch window size of 2. Global blocks operate on a smaller effective image resolution (e.g., for an input of $64 \times 64$ and a patch size of $8 \times 8$ , the effective resolution at the global blocks is $16 \times 16$ ). Skip connections (dashed lines) are used between corresponding patch merge/unmerge operations to preserve information such as static background features. + +
TaskDatasetModelPSNR ↑SSIM ↑F1 ↑Acc ↑IoU ↑CDC ↓Colorfulness ↑
Text ClassificationSST-2Ours41.898.776.875.5---
BERT-base---81.2---
Vega v1 [40]†--91.3----
Image ClassificationCIFAR-10Ours45.795.989.189.1---
ViT-base--71.371.3---
Resnet-101--90.090.0---
PCANet [5]--77.177.1---
ViT-huge [10]†--99.599.5---
Video ClassificationTinyVIRATOurs50.197.374.1 (30.4*)60.4---
ResNet50†--29.1*----
WideResNet†--32.6*----
Audio ClassificationaudioMNISTOurs50.998.996.997.1---
AlexNet [3]---95.8---
Video QACLEVREROurs25.779.852.452.5--
LSTM [37]†---34.7---
+ CNN [37]†---51.8---
TGIF-QAOurs42.098.553.252.3---
LSTM + Attn [15]---51.9---
Object TrackingLaSOTOurs35.798.7--0.63--
Dai et al. [8]----0.60--
Video ColorizationTinyVIRATOurs42.499.7---0.0273.1
Zhang et al. [39]-----0.0282.2
Yang et al. [36]-----0.0372.0
+ +Table 2. Performance metrics across various tasks using our proposed next-frame prediction framework. Metrics include Peak Signa-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), F1-score and Accuracy (F1, Acc), Intersection over Union (IoU), Color Distribution Consistency index (CDC), and Colorfulness, which are each tailored to evaluate specific task outputs. *macro F1-score to allow direct comparison with baseline models. †indicates models with pretraining on other datasets. + +crucial, as a model could appear consistent by merely reproducing the grayscale input. We use PSNR and the colorfulness metric from Liu et al. to assess color diversity. Temporal consistency is measured using the Color Distribution Consistency (CDC) index [18], which calculates the Jensen-Shannon (JS) divergence of color distributions + +across frames: + +$$ +C D C _ {t} = \frac {1}{3 \times (N - t)} \sum_ {c \in \{r, g, b \}} \sum_ {i = 1} ^ {N - t} J S (P _ {c} (I ^ {i}), P _ {c} (I ^ {I + t})) +$$ + +where $P_{c}(I^{i})$ is the normalized probability distribution + +of channel $c$ in frame $I^i$ of $N$ total frames. Overall consistency is computed as: + +$$ +C D C = \frac {1}{3} \left(C D C _ {1} + C D C _ {2} + C D C _ {4}\right) +$$ + +# 5. Results + +The results across each task are presented in Table 2, while Figure 1 illustrates several sample outputs produced by our model, offering a qualitative perspective on its performance. + +For the SST-2 text classification task, our model achieves an F1-score of 76.8. Although this is lower than the current state-of-the-art performance of 91.3, that was achieved through the use of a "Bag of Tricks" methodology involving extensive pretraining and complex finetuning processes [40]. Our model is simply trained on the SST-2 dataset, with no use of pretrained weights or advanced external embeddings, making it comparable to other models used without pretraining (See the BERT-base result). Some portion of this lower score can also be attributed to truncating any inputs over 20 tokens (limiting the evaluation to shorter inputs 20 tokens or fewer increases the F1-score to 80.0). + +On the CIFAR-10 image classification task, our model attains an accuracy of 89.1. While this score is lower than fine-tuned ViT models leveraging pretrained weights, which reach an impressive accuracy of 99.5 [10], our model performs significantly better than simpler baseline models, such as PCANet, which achieves 77.1 accuracy [5]. This performance suggests that our reformulation approach is able to perform image classification tasks without any pretraining, surpassing older, simpler architectures while trailing behind state-of-the-art models that benefit from extensive pretraining. + +In multi-label image classification on the TinyVIRAT dataset, our approach successfully outputs multiple labels per instance. When compared to baselines, our model surpasses the ResNet50 baseline (29.1) in performance but does not reach the level of the WideResNet model (32.6) [9]. + +On the AudioMNIST audio classification task, the model achieves an accuracy of 97.1, surpassing the AlexNet baseline score of 95.82 from the original dataset paper [3]. This notable improvement demonstrates the model's efficacy in audio-based classification tasks. Analysis of misclassifications reveals that the most frequent error involves confusion between the spoken digits "four" and "five," suggesting areas for potential refinement in distinguishing phonetically similar sounds. + +For the CLEVRR VQA (Visual Question Answering) dataset, our model achieves an accuracy of 52.5 on descriptive questions. This performance surpasses both the LSTM (34.7) and LSTM+CNN (51.8) baselines, illustrating our model's ability to understand the relationship between + +visual and textual information without separate encoder-decoder models [37]. It is important to note that due to hardware limitations, and the long question lengths included in the CLEVRER dataset, we subsampled the video to only include 4 frames of the moving scene, while also reducing the resolution to $64 \times 64$ , significantly limiting the accuracy achievable by our approach. + +In the LASOT object tracking task, the model reaches an intersection over union (IoU) of 0.63, demonstrating a consistent ability to track objects throughout video sequences. Analysis of tracking outputs reveals that the model effectively follows the object of interest across frames with accurately drawn bounding boxes. However, tracking accuracy diminishes slightly toward the end of longer sequences, with the borders of the overlaid bounding box eroding, likely due to autoregressive pixel-level error propagation. + +On the colorization task, our model shows improved color diversity in the generated outputs compared to the ground truth (73.1 color diversity for our model versus 70.6 in the original dataset). However, it exhibits less consistency in color application across the video sequence (0.0169 consistency for our model versus 0.00522 in the original dataset). This trade-off between diversity and consistency suggests that while our approach introduces a more vibrant color palette, further refinement could improve temporal coherence, particularly for applications requiring uniform color continuity across frames. + +ch + +# 5.1. Attention Maps + +In order to examine what is being learned by our model, we visualize internal patch-wise attention scores both spatially (within a frame) and temporally (between frames) in Figure 10. + +In the TinyVIRAT video classification task, the model pays most attention to the object/person performing the action, as well as frames and pairs of frames which indicate the class. In the first example (top row), spatial attention focuses on image patch locations containing the edges of the car which suggest relative movement, and temporal attention is highest for the final two frames in which the vehicle moves position, indicating the label "vehicle moving" should be output. + +For colorization, the greatest attention is paid to the first frames of the input sequence, which we suggest allows the model to produce consistent coloring across the whole sequence (more attention is also paid to the first frame of the output sequence, presumably to maintain color consistency). Spatially, the model clearly attends to key moving objects and areas of distinct color. + +In the CLEVRR examples, we see how the model is able to focus on question frames which contains key words needed to produce the output, such as "color" and "metal". + +![](images/3a566e7b18f2f85f0f1f45f399715a17a43b7f97d653300fbfad4a414d61a5ab.jpg) +Figure 10. Attention visualisations showing where the model attends to both spatially (indicated by light/dark areas overlaid on the frame), and temporally (indicated by a colour scale above the sequence). Spatial attention scores are calculated by taking the patch-wise contribution per frame, averaged over all layers. Temporal attention scores are calculated by taking the temporal attention weights for each input frame when generating the target frames, averaged over all global space-time attention layers. + +Within the moving object scene frames, we see how the model spatially attends to objects and their trajectory paths, correctly identifying objects such as "the metal cylinder" and "the sphere". + +Lastly, for sentiment text classification, we observe a similar pattern to attention maps in LLMs, where the model pays more attention to words with a strong emotive sentiment. In our model we see frames containing the words 'nightmare', 'painful', and 'dumb' exhibiting high temporal attention weightings, as they are strong indicators of the overall sentiment of the review. + +# 5.2. Limitations and Future work + +The aim of this work is not necessarily to surpass state-of-the-art performance on any of the tasks, but to show that unifying these modalities via reformulation is possible, without heavily compromising on performance relative to task-specific baselines. Future work should aim to scale this up, improving performance and tackling more complex tasks. We suggest that this can be done by mirroring the improvements gained from pretraining in NLP and computer vision by first training the model on unstructured tasks from each modality. Tasks such as language modelling and image captioning can be naturally reformulated as next-frame-prediction and included in our paradigm. This is especially important as a large portion of the current training time on tasks such as classification and question answering is ded + +icated to learning to output words rather than solving the task itself. Beginning with a checkpoint already trained on language modelling would mitigate this. + +# 6. Conclusion + +In the age of large foundation models, the trend is towards larger, more unified, fully end-to-end models without task-specific components. We introduced a novel multimodal learning framework that reformulates various multimodal tasks into a unified next-frame prediction paradigm. This approach addresses critical limitations in current multimodal model designs, which often require modality-specific encoders and are limited in scalability and flexibility when adapting to new tasks. By unifying diverse multimodal tasks under a single framework, our model can handle text, image, audio, and video inputs without modality-specific components, significantly simplifying the design and training process. We show it is possible to train a simple next-frame prediction model to solve these tasks, with many of them matching the performance of comparable models trained without pretraining data. + +Ultimately, our work lays the groundwork for building more generalized and efficient multimodal foundation models, towards systems that can process information from diverse modalities in a unified and scalable way. + +# References + +[1] Lakshita Agarwal and Bindu Verma. From methods to datasets: A survey on image-caption generators. Multimedia Tools and Applications, 83:28077-28123, 2023. 1 +[2] Alexei Baevski, Wei-Ning Hsu, Quantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. ArXiv, abs/2202.03555, 2022. 2 +[3] Soren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian Lapuschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute, 361(1):418-428, 2024. 4, 6, 7 +[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5 +[5] Tsung-Han Chan, Kui Jia, Shenghua Gao, Jiwen Lu, Zinan Zeng, and Yi Ma. Pcanet: A simple deep learning baseline for image classification? IEEE transactions on image processing, 24(12):5017-5032, 2015. 6, 7 +[6] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3 +[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Learning universal image-text representations. ArXiv, abs/1909.11740, 2019. 2 +[8] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 6 +[9] Ugur Demir, Yogesh S Rawat, and Mubarak Shah. Tinyvirat: Low-resolution video action recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 7387-7394. IEEE, 2021. 4, 7 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 6, 7 +[11] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. 4 +[12] Tianyu Gao, Zirui Wang, Adithya Bhaskar, and Danqi Chen. Improving language understanding from screenshots. arXiv preprint arXiv:2402.14073, 2024. 3 +[13] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 3 +[14] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. 2021 IEEE/CVF + +International Conference on Computer Vision (ICCV), pages 1419-1429, 2021. 2 +[15] Yunseok Jang, Yale Song, Chris Dongjoo Kim, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Video Question Answering with Spatio-Temporal Reasoning. IJCV, 2019. 4, 6 +[16] Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. Unifying question answering, text classification, and regression via span extraction. arXiv preprint arXiv:1904.09286, 2019. 2 +[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3 +[18] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. Computational Visual Media, 10(2):375–395, 2024. 6 +[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5 +[20] Siyu Lu, Mingzhe Liu, Lirong Yin, Zhengtong Yin, Xuan Liu, and Wenfeng Zheng. The multi-modal fusion in visual question answering: a review of attention mechanisms. PeerJ Computer Science, 9:e1400, 2023. 1 +[21] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018. 2 +[22] Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 2, 5 +[23] OpenAI. Gpt-4 technical report, 2024. 2 +[24] Marc'Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014. 3 +[25] Phillip Rust, Jonas F Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Eliott. Language modelling with pixels. arXiv preprint arXiv:2207.06991, 2022. 3 +[26] Min seok Seo, Hakjin Lee, Do-Yeon Kim, and Junghoon Seo. Implicit stacked autoregressive model for video prediction. ArXiv, abs/2303.07849, 2023. 3 +[27] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 2 +[28] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642, 2013. 3 +[29] Yintao Tai, Xiyang Liao, Alessandro Suglia, and Antonio Vergari. Pixar: Auto-regressive language modeling in pixel space. arXiv preprint arXiv:2401.03321, 2024. 3 + +[30] Joost R. van Amersfoort, Anitha Kannan, Marc'Aurelio Ranzato, Arthur Szlam, Du Tran, and Soumith Chintala. Transformation-based models of video sequences. ArXiv, abs/1701.08435, 2017. 3 +[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8296–8305, 2024. 1 +[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[33] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1 +[34] Chenghao Xiao, Zhuoxu Huang, Danlu Chen, G Thomas Hudson, Yizhi Li, Haoran Duan, Chenghua Lin, Jie Fu, Jungong Han, and Noura Al Moubayed. Pixel sentence representation learning. arXiv preprint arXiv:2402.08183, 2024. 3 +[35] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers, 2021. 3 +[36] Yixin Yang, Jiangxin Dong, Jinhui Tang, and Pan Jinshan. Colormapnet: A memory-based deep spatial-temporal feature propagation network for video colorization. In ECCV, 2024. 6 +[37] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 4, 6, 7 +[38] M. Akin Yilmaz and Ahmet Murat Tekalp. Dfpn: Deformable frame prediction network. 2021 IEEE International Conference on Image Processing (ICIP), pages 1944-1948, 2021. 3 +[39] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 6 +[40] Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint arXiv:2302.09268, 2023. 6, 7 +[41] Yeyun Zou and Qiyu Xie. A survey on vqa: Datasets and approaches. In 2020 2nd International Conference on Information Technology and Computer Application (ITCA), pages 289-297. IEEE, 2020. 1 \ No newline at end of file diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/images.zip b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..709ead55f35f510e9d9da6bf79aea9648cfa7d5e --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:45d2fecb9621431a88822a07ba4e32c8b528ba7483aeda31bf41bb28f5389fef +size 407145 diff --git a/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/layout.json b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..7b1cbf4e87efc0cc65e00fa7b14c6a23c0d31cbf --- /dev/null +++ b/2025/Everything is a Video_ Unifying Modalities through Next-Frame Prediction/layout.json @@ -0,0 +1,6450 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 74, + 103, + 536, + 120 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 74, + 103, + 536, + 120 + ], + "spans": [ + { + "bbox": [ + 74, + 103, + 536, + 120 + ], + "type": "text", + "content": "Everything is a Video: Unifying Modalities through Next-Frame Prediction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 104, + 143, + 202, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 143, + 202, + 156 + ], + "spans": [ + { + "bbox": [ + 104, + 143, + 202, + 156 + ], + "type": "text", + "content": "G. Thomas Hudson" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 223, + 144, + 280, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 223, + 144, + 280, + 156 + ], + "spans": [ + { + "bbox": [ + 223, + 144, + 280, + 156 + ], + "type": "text", + "content": "Dean Slack" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 302, + 144, + 412, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 302, + 144, + 412, + 156 + ], + "spans": [ + { + "bbox": [ + 302, + 144, + 412, + 156 + ], + "type": "text", + "content": "Thomas Winterbottom" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 433, + 144, + 505, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 433, + 144, + 505, + 157 + ], + "spans": [ + { + "bbox": [ + 433, + 144, + 505, + 157 + ], + "type": "text", + "content": "Jamie Sterling" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 172, + 158, + 250, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 158, + 250, + 171 + ], + "spans": [ + { + "bbox": [ + 172, + 158, + 250, + 171 + ], + "type": "text", + "content": "Chenghao Xiao" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 271, + 158, + 339, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 271, + 158, + 339, + 171 + ], + "spans": [ + { + "bbox": [ + 271, + 158, + 339, + 171 + ], + "type": "text", + "content": "Junjie Shentu" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 359, + 158, + 462, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 359, + 158, + 462, + 171 + ], + "spans": [ + { + "bbox": [ + 359, + 158, + 462, + 171 + ], + "type": "text", + "content": "Noura Al Moubayed" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 172, + 185, + 437, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 172, + 185, + 437, + 200 + ], + "spans": [ + { + "bbox": [ + 172, + 185, + 437, + 200 + ], + "type": "text", + "content": "Department of Computer Sciences, Durham University" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 241, + 202, + 368, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 241, + 202, + 368, + 213 + ], + "spans": [ + { + "bbox": [ + 241, + 202, + 368, + 213 + ], + "type": "text", + "content": "g.t.hudson@durham.ac.uk" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 99, + 220, + 511, + 367 + ], + "blocks": [ + { + "bbox": [ + 99, + 220, + 511, + 367 + ], + "lines": [ + { + "bbox": [ + 99, + 220, + 511, + 367 + ], + "spans": [ + { + "bbox": [ + 99, + 220, + 511, + 367 + ], + "type": "image", + "image_path": "fa0cbccea2015ca31ff32b5b015614b53a703ee11e6c36da55fec0b265aba43f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 370, + 555, + 393 + ], + "lines": [ + { + "bbox": [ + 54, + 370, + 555, + 393 + ], + "spans": [ + { + "bbox": [ + 54, + 370, + 555, + 393 + ], + "type": "text", + "content": "Figure 1. Sample outputs from our next frame prediction model across multiple modalities. The model receives frames to the left of the red line as input and everything right of the red line has been generated by the model." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 164, + 403, + 212, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 164, + 403, + 212, + 415 + ], + "spans": [ + { + "bbox": [ + 164, + 403, + 212, + 415 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 66, + 428, + 308, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 428, + 308, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 428, + 308, + 704 + ], + "type": "text", + "content": "Multimodal learning, which involves integrating information from various modalities such as text, images, audio, and video, is pivotal for numerous complex tasks like visual question answering, cross-modal retrieval, and caption generation. Traditional approaches rely on modality-specific encoders and late fusion techniques, which can hinder flexibility when adapting to new tasks or modalities. To address these limitations, we introduce a novel framework that extends the concept of task reformulation beyond natural language processing (NLP) to multimodal learning. We reformulate diverse multimodal tasks into a unified next-frame prediction problem, allowing a single model to handle different modalities without modality-specific components. This method treats all inputs and outputs as sequential frames in a video, enabling seamless integration of modalities and effective knowledge transfer across tasks. Our approach is evaluated on a range of tasks, including text-to-text, image-to-text, video-to-video, video-to-text, and audio-to-text, demonstrating the model's ability to generalize across modalities. We show that task reformulation can significantly simplify multimodal model design across various tasks, laying the groundwork for more generalized multimodal foundation models." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "spans": [ + { + "bbox": [ + 315, + 403, + 394, + 415 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 555, + 555 + ], + "type": "text", + "content": "Many tasks we seek to solve are multimodal. The ability to process and integrate information from multiple modalities such as text, image, audio, and video, have proven crucial for a range of complex tasks, including cross-modal retrieval [31], visual question answering [20, 33, 41], and caption generation [1]. The ability to understand and reason about information from different sensory inputs simultaneously mirrors human perception and cognition, making multimodal models an essential component in advancing artificial intelligence towards more generalized and versatile capabilities." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "The predominant approach to building multimodal models involves using separate encoders for each modality. For example, in vision-and-language models, a convolutional neural network (CNN) or a vision transformer (ViT) [10] might be used to encode images, while a recurrent neural network (RNN) or transformer is used to encode text. These modality-specific encoders are then fused at a later stage, typically via concatenation or attention mechanisms, to enable the model to perform joint reasoning over the encoded representations. While effective, this paradigm has inherent limitations. It necessitates careful design choices for each modality and often struggles with scalability and flexibility, especially when extending to new combinations of modali" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 758 + ], + "type": "text", + "content": "22004" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 205, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 205, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 205, + 83 + ], + "type": "text", + "content": "ties or tasks not seen during training." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 84, + 294, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 84, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 55, + 84, + 294, + 251 + ], + "type": "text", + "content": "To address these challenges, we propose a novel approach centered around task reformulation. Task reformulation is a well-established paradigm in natural language processing (NLP) that facilitates multitask learning by transforming diverse tasks into a common format [21]. A prominent example is prompt-based learning, where NLP tasks are reimagined as instructions or prompts to a large language model (LLM). This paradigm allows a wide range of NLP tasks, such as translation, summarization, and question answering, to be represented uniformly in a chat message format [23]. By doing so, it leverages the generalization capabilities of large pre-trained models, enabling them to perform multiple tasks with minimal task-specific customization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 251, + 295, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 441 + ], + "type": "text", + "content": "In this work, we extend the concept of task reformulation beyond NLP to the realm of multimodal learning. Specifically, we propose a framework in which tasks across various modalities are reformulated as a unified next-frame prediction problem. This approach simplifies the design of multimodal models by enabling them to handle diverse input types — text, images, audio, or video using a single, coherent mechanism. By representing all tasks as next-frame prediction, we create a shared interface for the model to learn and reason about information across modalities. This not only facilitates the integration of new modalities with minimal effort but also enhances the model's ability to transfer knowledge across different tasks and domains. Ultimately, just as LLMs function as foundation models for NLP, we intend this framing to be a foundation model across all modalities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 66, + 443, + 276, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 443, + 276, + 454 + ], + "spans": [ + { + "bbox": [ + 66, + 443, + 276, + 454 + ], + "type": "text", + "content": "The main contributions of this paper are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 456, + 293, + 573 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 56, + 456, + 293, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 293, + 490 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 293, + 490 + ], + "type": "text", + "content": "- We propose a task-reformulation method and show how to represent a range of multimodal tasks as next-frame prediction." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 492, + 293, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 293, + 537 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 293, + 537 + ], + "type": "text", + "content": "- We explore how this reformulation allows a single transformer-based model to solve text, image, video, and audio processing tasks without any modality-specific encoders." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 540, + 293, + 573 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 540, + 293, + 573 + ], + "spans": [ + { + "bbox": [ + 56, + 540, + 293, + 573 + ], + "type": "text", + "content": "- Experimental results show our model achieves comparable performance to single-task models which haven't been pretrained on additional data." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 586, + 139, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 586, + 139, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 586, + 139, + 597 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 713 + ], + "type": "text", + "content": "Multimodal Learning With the rise of transformers, LLMs, and very large scale pretraining, new paradigms for harmonising different modalities have been pushed to the forefront of multimodal learning. UNITER [7] uses a single universal transformer to process both image and text inputs into a join embedding space, with image representations pretended to the beginning of the input embeddings to the transformer. However, the image inputs are first processed by an RNN to obtain an embedding such that it is" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 443 + ], + "type": "text", + "content": "compatible with the textual inputs. Our framework in contrast has no need to preprocess input modalities differently as they are unified into a single visual input space theoretically capable of fully representing all multimodal interactions. Baevski et al. [2] propose data2vec, a framework for instilling features from multiple modalities into a single latent representation. Where data2vec seeks to process raw text, images, and audio using the same method, our work here reformulates the task inputs from multiple modalities into the same visual input medium. UniT [14], a multimodal multitask transformer uses separate encoders for each modality followed by a shared decoder. Our work uses a single visual encoder for our unified visual input paradigm. The most powerful state-of-the-art multimodal models such as FLAVA [27] and GPT-4 vision [23] integrate vision inputs via a patchwise process pipeline through a visual transformer fully integrated into the training pipeline. Both the architecture design and scale at which these foundation models are trained yield revolutionary performance at a variety of open-ended multimodal tasks. Nonetheless, though these foundation model are the closest that benchmarks have practically come to seamlessly integrating images with visual inputs, fundamentally the input modalities are handled differently by the model. Our work here bridges this remaining gap in processing of separate modalities. We reformulate typically spura-visual tasks into a visual input domain with no loss of information i.e. text and audio can be fully represented in images. This reformulation allows us to —for the first time—explore the generative pretraining capacity of multimodal information in one truly unified representation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 458, + 553, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 458, + 553, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 458, + 553, + 540 + ], + "type": "text", + "content": "Task Reformulation Task reformulation involves allowing a single model to solve multiple tasks be converting them into a single 'supertask'. This technique is common in NLP where a range of tasks such as sentiment classification, and coreference resolution have been reformulated as span-extraction [16], and question answering [21] among others." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 542, + 553, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 542, + 553, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 542, + 553, + 661 + ], + "type": "text", + "content": "The recent trend of prompt-based learning uses language modelling on prompt-answer sequences as the 'supertask' [23]. In this paradigm, the model is trained to follow an instruction describing the task and provide a response. In this way, any NLP task can be reformulated as the language modelling 'supertask'. Similar ideas have been explored in computer vision, notably the Transframer [22] which aimed to unify visual tasks. In our work, we extend this concept, exploring how next-frame prediction can be used as a 'supertask' to unite multiple modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "Next-frame Prediction CNN-based deep learning architectures for the autoregressive generation of future video frames have been steadily improving over the last decade" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22005" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "[24, 26, 30, 38]. Transformer-based models trained at scale such as VideoGPT [35] have substantially improved the performance of quality and fidelity of long and short term future frame predictions. Modern diffusion-based frame prediction models are now able to generate photorealistic outputs [13]. The aim of our work is not a new model architecture. Our model is inspired from state-of-the-art architectures to serve as an appropriately powerful benchmark for our analysis on our benchmark datasets." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 198, + 295, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 198, + 295, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 198, + 295, + 376 + ], + "type": "text", + "content": "Visual Representation of Text Inspired by the success of unsupervised next-word prediction with language models, learning the next pixel of images was proposed [6] and shown able to learn image representations. Learning textual semantics with vision models, by rendering texts into images, has drawn more and more attention [12, 25, 29] due to the drawbacks of tokenization in traditional language models and limitations in cross-lingual transferability. Xiao et al. [34] show that vision models pretrained on rendered images have stronger robustness on typo and word-order shuffling perturbations, and the representations of them display better isotropy on out-of-distribution (OOD) languages. In this work, we extend this approach when reformulating NLP tasks to video as well as other modalities which output text." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 389, + 111, + 402 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 111, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 111, + 402 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 410, + 295, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 295, + 471 + ], + "type": "text", + "content": "Our proposed framework consists of two key components: (1) Methods of reformulating a diverse range of input and output modalities into the single task of next-frame prediction; (2) A pure transformer-based model architecture. We introduce these components in the following sections." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 479, + 146, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 146, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 146, + 491 + ], + "type": "text", + "content": "3.1. Reformulation" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "content": "The datasets we use are carefully selected to cover a diverse range of input and output modalities, illustrating the versatility of our approach in reformulating various tasks as next-frame prediction problems. Our selection spans tasks from text and image classification to more complex audio, video, and multimodal tasks, enabling us to evaluate the model's ability to generalize across different data types and task requirements. By converting each task into a uniform format—a " + }, + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 497, + 295, + 640 + ], + "type": "text", + "content": " RGB video sequence—we create a consistent framework where the model treats every input and output as sequential frames, simplifying the multimodal learning process." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": "For each task, we insert a separator token " + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "inline_equation", + "content": "(|)" + }, + { + "bbox": [ + 55, + 641, + 295, + 689 + ], + "type": "text", + "content": " between the input and output frames to clearly delineate where the input ends and the prediction begins, guiding the model in distinguishing between these two parts of the sequence." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "In tasks involving textual data, we use a simple tokenizer (splitting on spaces and also on punctuation) to break" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": "down the text into individual tokens, which we then render as video frames in the sequence. Each token is represented in a consistent format: a fixed-width font scaled to fill the " + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 72, + 553, + 216 + ], + "type": "text", + "content": " frame, ensuring clarity and compatibility with our video-based input structure. This approach allows the model to read and predict text as it would any other frame, effectively bypassing the need for traditional text-based tokenization while integrating text seamlessly with other modalities. By converting each token to a visual format, we enable cross-modal knowledge transfer, allowing the model to process text, images, and other modalities through a shared, frame-based learning paradigm." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 230, + 553, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 230, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 230, + 553, + 315 + ], + "type": "text", + "content": "Text-to-Text The SST2 (Stanford Sentiment Treebank 2) dataset is a widely-used benchmark for sentiment classification, consisting of thousands of movie review excerpts labelled with binary sentiment labels (positive or negative) [28]. We use the text-encoding method described above, rendering each token as a video frame using a fixed-width font (Figure 2)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 328, + 324, + 542, + 356 + ], + "blocks": [ + { + "bbox": [ + 328, + 324, + 542, + 356 + ], + "lines": [ + { + "bbox": [ + 328, + 324, + 542, + 356 + ], + "spans": [ + { + "bbox": [ + 328, + 324, + 542, + 356 + ], + "type": "image", + "image_path": "857b51dd8cf3d79d1619fea16da761d4bdd4874d67aa49984f8644ee41160899.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 364, + 553, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 364, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 553, + 387 + ], + "type": "text", + "content": "Figure 2. Truncated example of the SST2 sentiment dataset rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 415, + 553, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 498 + ], + "type": "text", + "content": "Image-to-Text As a simple test of image recognition ability, we employ the CIFAR-10 dataset. CIFAR-10 is a benchmark for image classification and consists of 60,000 color images, each with dimensions of " + }, + { + "bbox": [ + 313, + 415, + 553, + 498 + ], + "type": "inline_equation", + "content": "32 \\times 32" + }, + { + "bbox": [ + 313, + 415, + 553, + 498 + ], + "type": "text", + "content": " pixels [17]. These images are equally divided into ten distinct, mutually exclusive classes: airplane, automobile, bird, cat, deer, dog, frog, horse, ship, and truck." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 374, + 508, + 494, + 548 + ], + "blocks": [ + { + "bbox": [ + 374, + 508, + 494, + 548 + ], + "lines": [ + { + "bbox": [ + 374, + 508, + 494, + 548 + ], + "spans": [ + { + "bbox": [ + 374, + 508, + 494, + 548 + ], + "type": "image", + "image_path": "208eee69929332ad58dec6c1f51a1510e52fa25e71973ce147570a935609e136.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 557, + 553, + 589 + ], + "lines": [ + { + "bbox": [ + 313, + 557, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 553, + 589 + ], + "type": "text", + "content": "Figure 3. Truncated example of the CIFAR10 image classification dataset rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "text", + "content": "In our reformulation approach, each CIFAR-10 image is resized to " + }, + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 603, + 553, + 675 + ], + "type": "text", + "content": " pixels to fit our standardized input frame size. Following the resized image, we insert a separator frame, which visually signifies the end of the input and the beginning of the output sequence, followed by a frame containing the class label text rendered as an image (Figure 3)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "Video-to-Text We utilize the TinyVIRAT action classification dataset to rigorously assess the model's capacity for" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22006" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "understanding video content [9]. The TinyVIRAT dataset comprises of 7,663 training and 5,166 testing examples, each annotated with one or more of 26 distinct action labels that describe the activities depicted within low-resolution video sequences. This dataset is particularly noteworthy for its multi-label nature, allowing for nuanced action recognition; for instance, a video sequence might feature a man walking while carrying a backpack, which would be simultaneously labelled with both \"walking\" and \"carrying\" actions. This characteristic enables the evaluation of the model's ability to recognize and differentiate between overlapping actions occurring in dynamic environments." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 296, + 277 + ], + "type": "text", + "content": "To reformulate the task for next-frame prediction, we reformulate each video sequence by first presenting the input video frames, followed by a separator token, and concluding with a comma-separated list of action labels that correspond to the activities depicted in the video. (Figure 4)." + } + ] + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 288, + 296, + 315 + ], + "blocks": [ + { + "bbox": [ + 56, + 288, + 296, + 315 + ], + "lines": [ + { + "bbox": [ + 56, + 288, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 56, + 288, + 296, + 315 + ], + "type": "image", + "image_path": "d327b0bd614e4a5ceef757a7e35dd786cf82e0b950ca02cd4998d2f339dc653a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "lines": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 323, + 296, + 357 + ], + "type": "text", + "content": "Figure 4. Truncated example of the Tinyvirat video classification dataset rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 556 + ], + "type": "text", + "content": "Video+Text-to-Text To explore the integration of multiple modalities within a single input, we leverage the CLEVRER (CoLlision Events for Video REpresentation and Reasoning) dataset, which is designed for video question answering (VQA) [37]. This dataset comprises synthetic video sequences depicting rendered 3D objects interacting through collisions, movements, and occlusions within a simple, controlled environment. Each video is paired with questions that require the model to understand and reason about these visual events. The questions in CLEVRER span various reasoning tasks, including descriptive, explanatory, and predictive questions about the events in the video, such as identifying specific objects, understanding object interactions, and predicting future states." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 653 + ], + "type": "text", + "content": "In our framework, we represent each question and corresponding video as a unified sequence of frames. Each input sequence begins with frames of the CLEVRER video sequence subsampled to 4 frames. This is followed by the question text, separator token, and answer text encoded in the manner described previously (Figure 5). Similarly, we use the frame QA task from TGIF-QA dataset[15], containing animated gifs with QA pairs." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 715 + ], + "type": "text", + "content": "To keep the sequence length manageable, we exclude complex counterfactual questions and multiple-choice answers from the dataset (only using the 'descriptive' subset), as these would require additional frames and could lead to overly long sequences that complicate training." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 315, + 70, + 554, + 94 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 554, + 94 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 94 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 94 + ], + "type": "image", + "image_path": "cfa90a4f0fa41cb12efc8e8c2e93e7fa898a267af929e40de053e1cfc73e0576.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 102, + 555, + 125 + ], + "lines": [ + { + "bbox": [ + 313, + 102, + 555, + 125 + ], + "spans": [ + { + "bbox": [ + 313, + 102, + 555, + 125 + ], + "type": "text", + "content": "Figure 5. Truncated example of the CLEVRR task rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 149, + 555, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 149, + 555, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 149, + 555, + 221 + ], + "type": "text", + "content": "Video-to-Video We employ two tasks to test video-to-video performance. Firstly, we convert the TinyVIRAT videos into greyscale and test the model's ability to colourise the output. The sequence consists of the greyscale-converted video sequence, the separator frame, then the original RGB video sequence (Figure 6)." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 236, + 542, + 268 + ], + "blocks": [ + { + "bbox": [ + 326, + 236, + 542, + 268 + ], + "lines": [ + { + "bbox": [ + 326, + 236, + 542, + 268 + ], + "spans": [ + { + "bbox": [ + 326, + 236, + 542, + 268 + ], + "type": "image", + "image_path": "b1fc20fc80ac61fd4f40f2f050872c0149be33646840806e3bf55f0ffa733c8c.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 276, + 555, + 299 + ], + "lines": [ + { + "bbox": [ + 313, + 276, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 555, + 299 + ], + "type": "text", + "content": "Figure 6. Truncated example of the colorization task rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 316, + 555, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 423 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 423 + ], + "type": "text", + "content": "Additionally, we test the model's ability to perform object tracking. The Large-scale Single Object Tracking dataset (LaSOT) consists of 1,550 video sequences hand-annotated with bounding boxes following an object of interest [11]. We reformulate this task by converting it into a sequence which consists of the first frame of the video with the bounding box overlayed, the rest of the sequence, the separator frame, and finally the full sequence including the overlayed bounding box (Figure 7)." + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 326, + 437, + 542, + 472 + ], + "blocks": [ + { + "bbox": [ + 326, + 437, + 542, + 472 + ], + "lines": [ + { + "bbox": [ + 326, + 437, + 542, + 472 + ], + "spans": [ + { + "bbox": [ + 326, + 437, + 542, + 472 + ], + "type": "image", + "image_path": "d69b9847d3f28d7759b7472286a44efed2fc8f7c77d86d624d982b4d8e451b66.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 503 + ], + "type": "text", + "content": "Figure 7. Truncated example of the LaSOT object tracking task rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 544, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 544, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 544, + 554, + 639 + ], + "type": "text", + "content": "Audio-to-Text To demonstrate the ability to process audio data, we utilize the AudioMNIST dataset, which contains 30,000 audio recordings of spoken digits ranging from zero to nine [3]. These recordings are spoken by multiple speakers with varying accents and intonations, providing a diverse dataset that challenges the model to recognize and classify spoken language across different vocal characteristics." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 714 + ], + "type": "text", + "content": "In our framework, each audio sample is preprocessed into a spectrogram, which is used at the first frame in the video sequence, representing the audio content visually for consistency with other modalities. Following the spectrogram, a separator token frame is added, and the label digit is represented in the final frame as text (Figure 8)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22007" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 116, + 70, + 237, + 111 + ], + "blocks": [ + { + "bbox": [ + 116, + 70, + 237, + 111 + ], + "lines": [ + { + "bbox": [ + 116, + 70, + 237, + 111 + ], + "spans": [ + { + "bbox": [ + 116, + 70, + 237, + 111 + ], + "type": "image", + "image_path": "39f555cdaf3142592099cfb4a1ad63de0c0043d8fdd139fd4873e89d91444f9b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 118, + 295, + 152 + ], + "lines": [ + { + "bbox": [ + 55, + 118, + 295, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 118, + 295, + 152 + ], + "type": "text", + "content": "Figure 8. Example of the AudioMNIST audio classification dataset rendered as a video. Each square is a frame of the video sequence." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 174, + 189, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 174, + 189, + 186 + ], + "spans": [ + { + "bbox": [ + 55, + 174, + 189, + 186 + ], + "type": "text", + "content": "3.2. Video Prediction Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 193, + 296, + 515 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 296, + 515 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 296, + 515 + ], + "type": "text", + "content": "We perform all tasks using a pure transformer-based model architecture. Building upon the successes of the ViT for image classification and TimesFormer [4] for video understanding, we modify these architectures to accommodate for spatiotemporal modelling of video sequences and autoregressive video prediction. We further adapt the model to progressively process patches at smaller resolutions, similar to that of U-net style models, such that adjacent input patches are linearly merged to reduce the effective image resolution at each block on the encoding side. This design choice allows us to train end-to-end models efficiently without resorting to pretrained encoder/decoder models. As illustrated in Fig. 9, partitioning into patches is followed by a series of local space-time transformed blocks separated by patch merge operations, with a global space-time transformer block operating on the reduced resolution patch-based representations. This has the benefit of capturing multiscale spatial features, while reducing the computational complexity of the global self-attention operations due to a smaller number of image patches. The reverse process is applied on the decoding side of the model to produce a predicted next image frame for each input frame. During inference, the final predicted frame is concatenated to the input sequence in an autoregressive manner. We outline the key model components and considerations below. Our model uses 41M parameters (for comparison, Transframer [22] uses 470–829M parameters)." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Patch Representation A core feature of the model is in the partitioning of input video frames into non-overlapping patches, similar to ViT. These patches are processed using local and global spatiotemporal self-attention mechanisms as described in [4], with spatial self-attention focusing on intra-frame patch relationships and temporal self-attention handling inter-frame causal dependencies. Temporal causality is enforced by masking future patches during training, ensuring that only past and present information is available for prediction. The model takes as input a video sequence " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{X} = \\{x_{t}\\}_{t = 1}^{T}" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ", such that each " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "x_{t} \\in \\mathbb{R}^{C \\times H \\times W}" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "-channel image frame of height " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " and width " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "t = 1,2,\\ldots,T" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": ". Frames are partitioned into " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " equally-sized non-overlapping patches, and linearly embedded to a 1D representation of size " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " (we set " + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": " to 512" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 320, + 70, + 552, + 169 + ], + "blocks": [ + { + "bbox": [ + 320, + 70, + 552, + 169 + ], + "lines": [ + { + "bbox": [ + 320, + 70, + 552, + 169 + ], + "spans": [ + { + "bbox": [ + 320, + 70, + 552, + 169 + ], + "type": "table", + "html": "
TaskInput LenTarget LenBatch Size
Text Classification21116
Image Classification21128
Video Classification11116
Audio Classification21256
Video QA21116
Object Tracking21208
Video Colorization111016
", + "image_path": "8a2b77de0fb96c9b90813dc4f9ffef359029a1c4cb4cf3467a2f9d1938920873.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 350, + 177, + 518, + 188 + ], + "lines": [ + { + "bbox": [ + 350, + 177, + 518, + 188 + ], + "spans": [ + { + "bbox": [ + 350, + 177, + 518, + 188 + ], + "type": "text", + "content": "Table 1. Training configurations for each task" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 210, + 555, + 282 + ], + "type": "text", + "content": "throughout). This embedding is combined with learnable spatial and temporal positional encodings to form the input sequence for the transformer. Spatial attention is applied between image patches within each frame independently, while causal temporal attention is applied across timesteps at the same spatial patch location." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 292, + 411, + 304 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 292, + 411, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 292, + 411, + 304 + ], + "type": "text", + "content": "3.3. Training Details" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 310, + 555, + 490 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 310, + 555, + 490 + ], + "spans": [ + { + "bbox": [ + 313, + 310, + 555, + 490 + ], + "type": "text", + "content": "We train our model on each dataset independently, with no language model or image pretraining, in order to determine the feasibility of our task reformulation approach with no latent performance benefits from transferring existing learned representations. We train end-to-end with a Multi-Scale Structural Similarity Index Measure (SSIM) loss [32], with a constant learning rate of " + }, + { + "bbox": [ + 313, + 310, + 555, + 490 + ], + "type": "inline_equation", + "content": "3 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 310, + 555, + 490 + ], + "type": "text", + "content": ", and a batch size of 8-32 depending on the task context length. We train with the AdamW [19] optimizer using default parameters, and set dropout to 0.1 for all layers except the final output layer. All models are evaluated on checkpoints corresponding to the best SSIM validation performance, with all models trained on a single NVIDIA A100 for a maximum of 7 GPU days. The input/output lengths and batch sizes vary per task, and are detailed in Table 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 502, + 394, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 394, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 394, + 516 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 523, + 555, + 631 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 631 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 631 + ], + "type": "text", + "content": "We train on each task in turn in a single-task manner (we leave training on all the tasks jointly in a multitask setting as future work). For tasks which output text, we perform OCR on the output frames using the open-source tesseract OCR engine1. For tasks which have a fixed vocabulary (e.g. classification tasks), the OCR text is matched to the closest vocabulary word to minimise any OCR-related errors. For classification tasks, the resulting text is measured using F1-score and Accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 632, + 554, + 668 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 668 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 668 + ], + "type": "text", + "content": "The object tracking task is evaluated by extracting the bounding box in the outputted frames, then comparing these to the labelled boxes using Intersection over Union (IoU)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 669, + 554, + 693 + ], + "type": "text", + "content": "For video colorization, we evaluate both colorization quality and temporal consistency. This dual evaluation is" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 703, + 533, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 703, + 533, + 712 + ], + "spans": [ + { + "bbox": [ + 325, + 703, + 533, + 712 + ], + "type": "text", + "content": "1https://github.com/tesseract-ocr/tesseract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22008" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 107, + 69, + 504, + 173 + ], + "blocks": [ + { + "bbox": [ + 107, + 69, + 504, + 173 + ], + "lines": [ + { + "bbox": [ + 107, + 69, + 504, + 173 + ], + "spans": [ + { + "bbox": [ + 107, + 69, + 504, + 173 + ], + "type": "image", + "image_path": "39e907a353976dd30fc2df5f811125c3af6de5ab05ea6a7a482c1eb0f8d724bb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "lines": [ + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "spans": [ + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": "Figure 9. Illustration of the Transformer-based video prediction model used in this work. " + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": " timestep video image frames are partitioned into " + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": " non-overlapping patches and linearly embedded before being processed by a series of local and global space-time transformer blocks. Local blocks are restricted to local spatial and temporal attention windows, separated by non-overlapping patch merge/unmerge operations with a patch window size of 2. Global blocks operate on a smaller effective image resolution (e.g., for an input of " + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": " and a patch size of " + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": ", the effective resolution at the global blocks is " + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 54, + 180, + 555, + 248 + ], + "type": "text", + "content": "). Skip connections (dashed lines) are used between corresponding patch merge/unmerge operations to preserve information such as static background features." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 67, + 257, + 543, + 559 + ], + "blocks": [ + { + "bbox": [ + 67, + 257, + 543, + 559 + ], + "lines": [ + { + "bbox": [ + 67, + 257, + 543, + 559 + ], + "spans": [ + { + "bbox": [ + 67, + 257, + 543, + 559 + ], + "type": "table", + "html": "
TaskDatasetModelPSNR ↑SSIM ↑F1 ↑Acc ↑IoU ↑CDC ↓Colorfulness ↑
Text ClassificationSST-2Ours41.898.776.875.5---
BERT-base---81.2---
Vega v1 [40]†--91.3----
Image ClassificationCIFAR-10Ours45.795.989.189.1---
ViT-base--71.371.3---
Resnet-101--90.090.0---
PCANet [5]--77.177.1---
ViT-huge [10]†--99.599.5---
Video ClassificationTinyVIRATOurs50.197.374.1 (30.4*)60.4---
ResNet50†--29.1*----
WideResNet†--32.6*----
Audio ClassificationaudioMNISTOurs50.998.996.997.1---
AlexNet [3]---95.8---
Video QACLEVREROurs25.779.852.452.5--
LSTM [37]†---34.7---
+ CNN [37]†---51.8---
TGIF-QAOurs42.098.553.252.3---
LSTM + Attn [15]---51.9---
Object TrackingLaSOTOurs35.798.7--0.63--
Dai et al. [8]----0.60--
Video ColorizationTinyVIRATOurs42.499.7---0.0273.1
Zhang et al. [39]-----0.0282.2
Yang et al. [36]-----0.0372.0
", + "image_path": "2cf86d690bccee7b8ba94b5d2b73a36665e900b551f45360e3d4a54a60bf22bb.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 567, + 555, + 613 + ], + "lines": [ + { + "bbox": [ + 54, + 567, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 54, + 567, + 555, + 613 + ], + "type": "text", + "content": "Table 2. Performance metrics across various tasks using our proposed next-frame prediction framework. Metrics include Peak Signa-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), F1-score and Accuracy (F1, Acc), Intersection over Union (IoU), Color Distribution Consistency index (CDC), and Colorfulness, which are each tailored to evaluate specific task outputs. *macro F1-score to allow direct comparison with baseline models. †indicates models with pretraining on other datasets." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 632, + 294, + 704 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 632, + 294, + 704 + ], + "spans": [ + { + "bbox": [ + 54, + 632, + 294, + 704 + ], + "type": "text", + "content": "crucial, as a model could appear consistent by merely reproducing the grayscale input. We use PSNR and the colorfulness metric from Liu et al. to assess color diversity. Temporal consistency is measured using the Color Distribution Consistency (CDC) index [18], which calculates the Jensen-Shannon (JS) divergence of color distributions" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 633, + 375, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 633, + 375, + 643 + ], + "spans": [ + { + "bbox": [ + 314, + 633, + 375, + 643 + ], + "type": "text", + "content": "across frames:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 659, + 553, + 695 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 553, + 695 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 553, + 695 + ], + "type": "interline_equation", + "content": "C D C _ {t} = \\frac {1}{3 \\times (N - t)} \\sum_ {c \\in \\{r, g, b \\}} \\sum_ {i = 1} ^ {N - t} J S (P _ {c} (I ^ {i}), P _ {c} (I ^ {I + t}))", + "image_path": "a6051f42e40f8720a53e68f35918c8a953f714ab53d962a74568ab1808e8dafe.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 700, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 700, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 700, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 326, + 700, + 555, + 714 + ], + "type": "inline_equation", + "content": "P_{c}(I^{i})" + }, + { + "bbox": [ + 326, + 700, + 555, + 714 + ], + "type": "text", + "content": " is the normalized probability distribution" + } + ] + } + ], + "index": 7 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22009" + } + ] + } + ], + "index": 8 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "of channel " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in frame " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "I^i" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " of " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " total frames. Overall consistency is computed as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 106, + 258, + 129 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 106, + 258, + 129 + ], + "spans": [ + { + "bbox": [ + 93, + 106, + 258, + 129 + ], + "type": "interline_equation", + "content": "C D C = \\frac {1}{3} \\left(C D C _ {1} + C D C _ {2} + C D C _ {4}\\right)", + "image_path": "06ae0fea0ad1165ba81164088c2d07bb2d57fc1e82a4365e3f1350720896ffc5.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 139, + 108, + 152 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 139, + 108, + 152 + ], + "spans": [ + { + "bbox": [ + 55, + 139, + 108, + 152 + ], + "type": "text", + "content": "5. Results" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 159, + 295, + 195 + ], + "type": "text", + "content": "The results across each task are presented in Table 2, while Figure 1 illustrates several sample outputs produced by our model, offering a qualitative perspective on its performance." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 196, + 295, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 196, + 295, + 339 + ], + "spans": [ + { + "bbox": [ + 54, + 196, + 295, + 339 + ], + "type": "text", + "content": "For the SST-2 text classification task, our model achieves an F1-score of 76.8. Although this is lower than the current state-of-the-art performance of 91.3, that was achieved through the use of a \"Bag of Tricks\" methodology involving extensive pretraining and complex finetuning processes [40]. Our model is simply trained on the SST-2 dataset, with no use of pretrained weights or advanced external embeddings, making it comparable to other models used without pretraining (See the BERT-base result). Some portion of this lower score can also be attributed to truncating any inputs over 20 tokens (limiting the evaluation to shorter inputs 20 tokens or fewer increases the F1-score to 80.0)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 340, + 295, + 472 + ], + "type": "text", + "content": "On the CIFAR-10 image classification task, our model attains an accuracy of 89.1. While this score is lower than fine-tuned ViT models leveraging pretrained weights, which reach an impressive accuracy of 99.5 [10], our model performs significantly better than simpler baseline models, such as PCANet, which achieves 77.1 accuracy [5]. This performance suggests that our reformulation approach is able to perform image classification tasks without any pretraining, surpassing older, simpler architectures while trailing behind state-of-the-art models that benefit from extensive pretraining." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 295, + 544 + ], + "type": "text", + "content": "In multi-label image classification on the TinyVIRAT dataset, our approach successfully outputs multiple labels per instance. When compared to baselines, our model surpasses the ResNet50 baseline (29.1) in performance but does not reach the level of the WideResNet model (32.6) [9]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 652 + ], + "type": "text", + "content": "On the AudioMNIST audio classification task, the model achieves an accuracy of 97.1, surpassing the AlexNet baseline score of 95.82 from the original dataset paper [3]. This notable improvement demonstrates the model's efficacy in audio-based classification tasks. Analysis of misclassifications reveals that the most frequent error involves confusion between the spoken digits \"four\" and \"five,\" suggesting areas for potential refinement in distinguishing phonetically similar sounds." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "For the CLEVRR VQA (Visual Question Answering) dataset, our model achieves an accuracy of 52.5 on descriptive questions. This performance surpasses both the LSTM (34.7) and LSTM+CNN (51.8) baselines, illustrating our model's ability to understand the relationship between" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "visual and textual information without separate encoder-decoder models [37]. It is important to note that due to hardware limitations, and the long question lengths included in the CLEVRER dataset, we subsampled the video to only include 4 frames of the moving scene, while also reducing the resolution to " + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": ", significantly limiting the accuracy achievable by our approach." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 264 + ], + "type": "text", + "content": "In the LASOT object tracking task, the model reaches an intersection over union (IoU) of 0.63, demonstrating a consistent ability to track objects throughout video sequences. Analysis of tracking outputs reveals that the model effectively follows the object of interest across frames with accurately drawn bounding boxes. However, tracking accuracy diminishes slightly toward the end of longer sequences, with the borders of the overlaid bounding box eroding, likely due to autoregressive pixel-level error propagation." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 265, + 553, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 265, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 313, + 265, + 553, + 396 + ], + "type": "text", + "content": "On the colorization task, our model shows improved color diversity in the generated outputs compared to the ground truth (73.1 color diversity for our model versus 70.6 in the original dataset). However, it exhibits less consistency in color application across the video sequence (0.0169 consistency for our model versus 0.00522 in the original dataset). This trade-off between diversity and consistency suggests that while our approach introduces a more vibrant color palette, further refinement could improve temporal coherence, particularly for applications requiring uniform color continuity across frames." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 398, + 339, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 398, + 339, + 408 + ], + "spans": [ + { + "bbox": [ + 326, + 398, + 339, + 408 + ], + "type": "text", + "content": "ch" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 418, + 410, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 410, + 431 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 410, + 431 + ], + "type": "text", + "content": "5.1. Attention Maps" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 436, + 553, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 553, + 483 + ], + "type": "text", + "content": "In order to examine what is being learned by our model, we visualize internal patch-wise attention scores both spatially (within a frame) and temporally (between frames) in Figure 10." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 553, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 553, + 592 + ], + "type": "text", + "content": "In the TinyVIRAT video classification task, the model pays most attention to the object/person performing the action, as well as frames and pairs of frames which indicate the class. In the first example (top row), spatial attention focuses on image patch locations containing the edges of the car which suggest relative movement, and temporal attention is highest for the final two frames in which the vehicle moves position, indicating the label \"vehicle moving\" should be output." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 594, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 553, + 676 + ], + "type": "text", + "content": "For colorization, the greatest attention is paid to the first frames of the input sequence, which we suggest allows the model to produce consistent coloring across the whole sequence (more attention is also paid to the first frame of the output sequence, presumably to maintain color consistency). Spatially, the model clearly attends to key moving objects and areas of distinct color." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "In the CLEVRR examples, we see how the model is able to focus on question frames which contains key words needed to produce the output, such as \"color\" and \"metal\"." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "22010" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 126, + 56, + 486, + 337 + ], + "blocks": [ + { + "bbox": [ + 126, + 56, + 486, + 337 + ], + "lines": [ + { + "bbox": [ + 126, + 56, + 486, + 337 + ], + "spans": [ + { + "bbox": [ + 126, + 56, + 486, + 337 + ], + "type": "image", + "image_path": "3a566e7b18f2f85f0f1f45f399715a17a43b7f97d653300fbfad4a414d61a5ab.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 340, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 54, + 340, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 340, + 555, + 386 + ], + "type": "text", + "content": "Figure 10. Attention visualisations showing where the model attends to both spatially (indicated by light/dark areas overlaid on the frame), and temporally (indicated by a colour scale above the sequence). Spatial attention scores are calculated by taking the patch-wise contribution per frame, averaged over all layers. Temporal attention scores are calculated by taking the temporal attention weights for each input frame when generating the target frames, averaged over all global space-time attention layers." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 391, + 295, + 439 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 295, + 439 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 295, + 439 + ], + "type": "text", + "content": "Within the moving object scene frames, we see how the model spatially attends to objects and their trajectory paths, correctly identifying objects such as \"the metal cylinder\" and \"the sphere\"." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 295, + 522 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 295, + 522 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 295, + 522 + ], + "type": "text", + "content": "Lastly, for sentiment text classification, we observe a similar pattern to attention maps in LLMs, where the model pays more attention to words with a strong emotive sentiment. In our model we see frames containing the words 'nightmare', 'painful', and 'dumb' exhibiting high temporal attention weightings, as they are strong indicators of the overall sentiment of the review." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 529, + 216, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 216, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 216, + 540 + ], + "type": "text", + "content": "5.2. Limitations and Future work" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "The aim of this work is not necessarily to surpass state-of-the-art performance on any of the tasks, but to show that unifying these modalities via reformulation is possible, without heavily compromising on performance relative to task-specific baselines. Future work should aim to scale this up, improving performance and tackling more complex tasks. We suggest that this can be done by mirroring the improvements gained from pretraining in NLP and computer vision by first training the model on unstructured tasks from each modality. Tasks such as language modelling and image captioning can be naturally reformulated as next-frame-prediction and included in our paradigm. This is especially important as a large portion of the current training time on tasks such as classification and question answering is ded" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 391, + 555, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 555, + 427 + ], + "type": "text", + "content": "icated to learning to output words rather than solving the task itself. Beginning with a checkpoint already trained on language modelling would mitigate this." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 446, + 388, + 459 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 388, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 388, + 459 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 470, + 555, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 470, + 555, + 662 + ], + "spans": [ + { + "bbox": [ + 312, + 470, + 555, + 662 + ], + "type": "text", + "content": "In the age of large foundation models, the trend is towards larger, more unified, fully end-to-end models without task-specific components. We introduced a novel multimodal learning framework that reformulates various multimodal tasks into a unified next-frame prediction paradigm. This approach addresses critical limitations in current multimodal model designs, which often require modality-specific encoders and are limited in scalability and flexibility when adapting to new tasks. By unifying diverse multimodal tasks under a single framework, our model can handle text, image, audio, and video inputs without modality-specific components, significantly simplifying the design and training process. We show it is possible to train a simple next-frame prediction model to solve these tasks, with many of them matching the performance of comparable models trained without pretraining data." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 715 + ], + "type": "text", + "content": "Ultimately, our work lays the groundwork for building more generalized and efficient multimodal foundation models, towards systems that can process information from diverse modalities in a unified and scalable way." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 758 + ], + "type": "text", + "content": "22011" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Lakshita Agarwal and Bindu Verma. From methods to datasets: A survey on image-caption generators. Multimedia Tools and Applications, 83:28077-28123, 2023. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 295, + 167 + ], + "type": "text", + "content": "[2] Alexei Baevski, Wei-Ning Hsu, Quantong Xu, Arun Babu, Jiatao Gu, and Michael Auli. data2vec: A general framework for self-supervised learning in speech, vision and language. ArXiv, abs/2202.03555, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 170, + 296, + 223 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 170, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 62, + 170, + 296, + 223 + ], + "type": "text", + "content": "[3] Soren Becker, Johanna Vielhaben, Marcel Ackermann, Klaus-Robert Müller, Sebastian Lapuschkin, and Wojciech Samek. Audiomnist: Exploring explainable artificial intelligence for audio analysis on a simple benchmark. Journal of the Franklin Institute, 361(1):418-428, 2024. 4, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 224, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 224, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 224, + 295, + 257 + ], + "type": "text", + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In ICML, page 4, 2021. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 294, + 301 + ], + "type": "text", + "content": "[5] Tsung-Han Chan, Kui Jia, Shenghua Gao, Jiwen Lu, Zinan Zeng, and Yi Ma. Pcanet: A simple deep learning baseline for image classification? IEEE transactions on image processing, 24(12):5017-5032, 2015. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 62, + 303, + 294, + 346 + ], + "type": "text", + "content": "[6] Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, and Ilya Sutskever. Generative pretraining from pixels. In International conference on machine learning, pages 1691-1703. PMLR, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "spans": [ + { + "bbox": [ + 62, + 347, + 294, + 390 + ], + "type": "text", + "content": "[7] Yen-Chun Chen, Linjie Li, Licheng Yu, Ahmed El Kholy, Faisal Ahmed, Zhe Gan, Yu Cheng, and Jingjing Liu. Uniter: Learning universal image-text representations. ArXiv, abs/1909.11740, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 392, + 294, + 424 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 294, + 424 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 294, + 424 + ], + "type": "text", + "content": "[8] Kenan Dai, Yunhua Zhang, Dong Wang, Jianhua Li, Huchuan Lu, and Xiaoyun Yang. High-performance long-term tracking with meta-updater. In CVPR, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 426, + 294, + 468 + ], + "type": "text", + "content": "[9] Ugur Demir, Yogesh S Rawat, and Mubarak Shah. Tinyvirat: Low-resolution video action recognition. In 2020 25th international conference on pattern recognition (ICPR), pages 7387-7394. IEEE, 2021. 4, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 470, + 294, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 470, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 57, + 470, + 294, + 545 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 1, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 547, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 547, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 57, + 547, + 294, + 612 + ], + "type": "text", + "content": "[11] Heng Fan, Liting Lin, Fan Yang, Peng Chu, Ge Deng, Sijia Yu, Hexin Bai, Yong Xu, Chunyuan Liao, and Haibin Ling. Lasot: A high-quality benchmark for large-scale single object tracking. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5374-5383, 2019. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 614, + 294, + 646 + ], + "type": "text", + "content": "[12] Tianyu Gao, Zirui Wang, Adithya Bhaskar, and Danqi Chen. Improving language understanding from screenshots. arXiv preprint arXiv:2402.14073, 2024. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 690 + ], + "type": "text", + "content": "[13] Agrim Gupta, Lijun Yu, Kihyuk Sohn, Xiuye Gu, Meera Hahn, Fei-Fei Li, Irfan Essa, Lu Jiang, and José Lezama. Photorealistic video generation with diffusion models. In European Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 713 + ], + "type": "text", + "content": "[14] Ronghang Hu and Amanpreet Singh. Unit: Multimodal multitask learning with a unified transformer. 2021 IEEE/CVF" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "International Conference on Computer Vision (ICCV), pages 1419-1429, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 129 + ], + "type": "text", + "content": "[15] Yunseok Jang, Yale Song, Chris Dongjoo Kim, Youngjae Yu, Youngjin Kim, and Gunhee Kim. Video Question Answering with Spatio-Temporal Reasoning. IJCV, 2019. 4, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 317, + 132, + 553, + 175 + ], + "type": "text", + "content": "[16] Nitish Shirish Keskar, Bryan McCann, Caiming Xiong, and Richard Socher. Unifying question answering, text classification, and regression via span extraction. arXiv preprint arXiv:1904.09286, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 177, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 553, + 199 + ], + "type": "text", + "content": "[17] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 201, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 201, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 201, + 553, + 255 + ], + "type": "text", + "content": "[18] Yihao Liu, Hengyuan Zhao, Kelvin CK Chan, Xintao Wang, Chen Change Loy, Yu Qiao, and Chao Dong. Temporally consistent video colorization with deep feature propagation and self-regularization learning. Computational Visual Media, 10(2):375–395, 2024. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 258, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 258, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 317, + 258, + 553, + 289 + ], + "type": "text", + "content": "[19] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, 2017. 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 317, + 292, + 553, + 335 + ], + "type": "text", + "content": "[20] Siyu Lu, Mingzhe Liu, Lirong Yin, Zhengtong Yin, Xuan Liu, and Wenfeng Zheng. The multi-modal fusion in visual question answering: a review of attention mechanisms. PeerJ Computer Science, 9:e1400, 2023. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 317, + 338, + 553, + 380 + ], + "type": "text", + "content": "[21] Bryan McCann, Nitish Shirish Keskar, Caiming Xiong, and Richard Socher. The natural language decathlon: Multitask learning as question answering. arXiv preprint arXiv:1806.08730, 2018. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 383, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 427 + ], + "type": "text", + "content": "[22] Charlie Nash, Joao Carreira, Jacob Walker, Iain Barr, Andrew Jaegle, Mateusz Malinowski, and Peter Battaglia. Transframer: Arbitrary frame prediction with generative models. arXiv preprint arXiv:2203.09494, 2022. 2, 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 429, + 483, + 440 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 429, + 483, + 440 + ], + "spans": [ + { + "bbox": [ + 317, + 429, + 483, + 440 + ], + "type": "text", + "content": "[23] OpenAI. Gpt-4 technical report, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 441, + 553, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 441, + 553, + 485 + ], + "spans": [ + { + "bbox": [ + 317, + 441, + 553, + 485 + ], + "type": "text", + "content": "[24] Marc'Aurelio Ranzato, Arthur Szlam, Joan Bruna, Michael Mathieu, Ronan Collobert, and Sumit Chopra. Video (language) modeling: a baseline for generative models of natural videos. ArXiv, abs/1412.6604, 2014. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 487, + 553, + 529 + ], + "type": "text", + "content": "[25] Phillip Rust, Jonas F Lotz, Emanuele Bugliarello, Elizabeth Salesky, Miryam de Lhoneux, and Desmond Eliott. Language modelling with pixels. arXiv preprint arXiv:2207.06991, 2022. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 533, + 553, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 533, + 553, + 564 + ], + "spans": [ + { + "bbox": [ + 317, + 533, + 553, + 564 + ], + "type": "text", + "content": "[26] Min seok Seo, Hakjin Lee, Do-Yeon Kim, and Junghoon Seo. Implicit stacked autoregressive model for video prediction. ArXiv, abs/2303.07849, 2023. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "spans": [ + { + "bbox": [ + 317, + 567, + 553, + 610 + ], + "type": "text", + "content": "[27] Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. FLAVA: A foundational language and vision alignment model. CoRR, abs/2112.04482, 2021. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 678 + ], + "type": "text", + "content": "[28] Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In Proceedings of the 2013 conference on empirical methods in natural language processing, pages 1631-1642, 2013. 3" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 553, + 713 + ], + "type": "text", + "content": "[29] Yintao Tai, Xiyang Liao, Alessandro Suglia, and Antonio Vergari. Pixar: Auto-regressive language modeling in pixel space. arXiv preprint arXiv:2401.03321, 2024. 3" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "22012" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 609 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[30] Joost R. van Amersfoort, Anitha Kannan, Marc'Aurelio Ranzato, Arthur Szlam, Du Tran, and Soumith Chintala. Transformation-based models of video sequences. ArXiv, abs/1701.08435, 2017. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 171 + ], + "type": "text", + "content": "[31] Yabing Wang, Le Wang, Qiang Zhou, Zhibin Wang, Hao Li, Gang Hua, and Wei Tang. Multimodal llm enhanced cross-lingual cross-modal retrieval. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 8296–8305, 2024. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 216 + ], + "type": "text", + "content": "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 261 + ], + "type": "text", + "content": "[33] Qi Wu, Damien Teney, Peng Wang, Chunhua Shen, Anthony Dick, and Anton Van Den Hengel. Visual question answering: A survey of methods and datasets. Computer Vision and Image Understanding, 163:21-40, 2017. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 316 + ], + "type": "text", + "content": "[34] Chenghao Xiao, Zhuoxu Huang, Danlu Chen, G Thomas Hudson, Yizhi Li, Haoran Duan, Chenghua Lin, Jie Fu, Jungong Han, and Noura Al Moubayed. Pixel sentence representation learning. arXiv preprint arXiv:2402.08183, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 350 + ], + "type": "text", + "content": "[35] Wilson Yan, Yunzhi Zhang, Pieter Abbeel, and Aravind Srinivas. Videogpt: Video generation using vq-vae and transformers, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 396 + ], + "type": "text", + "content": "[36] Yixin Yang, Jiangxin Dong, Jinhui Tang, and Pan Jinshan. Colormapnet: A memory-based deep spatial-temporal feature propagation network for video colorization. In ECCV, 2024. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 441 + ], + "type": "text", + "content": "[37] Kexin Yi, Chuang Gan, Yunzhu Li, Pushmeet Kohli, Jiajun Wu, Antonio Torralba, and Joshua B Tenenbaum. Clevrer: Collision events for video representation and reasoning. arXiv preprint arXiv:1910.01442, 2019. 4, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 485 + ], + "type": "text", + "content": "[38] M. Akin Yilmaz and Ahmet Murat Tekalp. Dfpn: Deformable frame prediction network. 2021 IEEE International Conference on Image Processing (ICIP), pages 1944-1948, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 487, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 509 + ], + "type": "text", + "content": "[39] Richard Zhang, Phillip Isola, and Alexei A Efros. Colorful image colorization. In ECCV, 2016. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 294, + 563 + ], + "type": "text", + "content": "[40] Qihuang Zhong, Liang Ding, Keqin Peng, Juhua Liu, Bo Du, Li Shen, Yibing Zhan, and Dacheng Tao. Bag of tricks for effective language model pretraining and downstream adaptation: A case study on glue. arXiv preprint arXiv:2302.09268, 2023. 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 609 + ], + "type": "text", + "content": "[41] Yeyun Zou and Qiyu Xie. A survey on vqa: Datasets and approaches. In 2020 2nd International Conference on Information Technology and Computer Application (ITCA), pages 289-297. IEEE, 2020. 1" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "22013" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_content_list.json b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..98431c29135d136b612f241fd0b36bcded21d2d2 --- /dev/null +++ b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_content_list.json @@ -0,0 +1,1515 @@ +[ + { + "type": "text", + "text": "Evidential Knowledge Distillation", + "text_level": 1, + "bbox": [ + 326, + 130, + 671, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Liangyu Xiang $^{1,2}$ Junyu Gao $^{1,2\\dagger}$ Changsheng Xu $^{1,2,3}$", + "bbox": [ + 243, + 178, + 750, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)", + "bbox": [ + 192, + 198, + 805, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)", + "bbox": [ + 156, + 233, + 841, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Peng Cheng Laboratory, ShenZhen, China", + "bbox": [ + 326, + 251, + 671, + 268 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "xiangliangyu2023@ia.ac.cn; {junyu.gao, csxu}@nlpr.ia.ac.cn", + "bbox": [ + 238, + 271, + 751, + 287 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 321, + 326, + 338 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Existing logit-based knowledge distillation methods typically employ singularly deterministic categorical distributions, which eliminates the inherent uncertainty in network predictions and thereby limiting the effective transfer of knowledge. To address this limitation, we introduce distribution-based probabilistic modeling as a more comprehensive representation of network knowledge. Specifically, we regard the categorical distribution as a random variable and leverage deep neural networks to predict its distribution, representing it as an evidential second-order distribution. Based on the second-oder modeling, we propose Evidential Knowledge Distillation (EKD) which distills both the expectation of the teacher distribution and the distribution itself into the student. The expectation captures the macroscopic characteristics of the distribution, while the distribution itself conveys microscopic information about the classification boundaries. Additionally, we theoretically show that EKD's distillation objective provides an upper bound on the student's expected risk when treating the teacher's predictions as ground truth. Extensive experiments on several standard benchmarks across various teacher-student network pairs highlight the effectiveness and superior performance of EKD. Our code is available at https://github.com/lyxiang-casia/EKD.", + "bbox": [ + 88, + 353, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 744, + 223, + 760 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "As deep neural networks (DNNs) continue to advance, their performance has rapidly improved, yet their increasing network size and training costs pose challenges, especially in resource-constrained and real-time processing scenarios [62, 89]. Consequently, logit-based knowledge distillation (KD) [30, 88], which aims to compress the predictive capabilities of complex models into lightweight ones, has", + "bbox": [ + 89, + 770, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e9c82d45add299cc7332d4345c442a3f96c84aec0be55c85a4b295232af45d9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 319, + 908, + 459 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/48d12f289c6be655a348d29dd4f14be951fbedc2e92ea5139810ad62e2ea7dbd.jpg", + "image_caption": [ + "(a) Vanilla Knowledge Distillation", + "(b) Evidential Knowledge Distillation", + "Figure 1. Illustrations of vanilla KD and our EKD. (a) Vanilla KD deterministically converts the logits into categorical distributions through the softmax function. Each categorical distribution corresponds to a point on the simplex, with each categorical probability representing the barycentric coordinates of that point. Vanilla KD is achieved by minimizing the differences between the teacher's and student's barycentric coordinates. (b) EKD treats the categorical distribution as a random variable following a Dirichlet distribution, parameterized by the logits. The heatmap depicts the probability density of the Dirichlet distribution across the simplex. Moreover, EKD includes distillation of both the expectation of the second-order distribution, and the second-order distribution itself." + ], + "image_footnote": [], + "bbox": [ + 514, + 472, + 908, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "gained increasing popularity. KD transfers knowledge by guiding the predictions of a smaller network to align with those of a larger network during the training.", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Current knowledge distillation methods largely adhere to the vanilla paradigm [30] where only the categorical distri", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 886, + 236, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "2814", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "butions are aligned. As illustrated in Figure 1a, the logits of a complex network (teacher network) and a lightweight network (student network) are firstly converted into categorical probabilities using a softmax function. Knowledge is then transferred through the process of minimizing the KL divergence between the two categorical distributions. Enhancements to this distillation process, such as decoupling the categorical distributions [88], smoothing the distribution with multiple levels [45], or adopting sample-wise distribution smoothness [69], have demonstrated notable improvements in the student network's predictive performance. Despite these advancements, all such methods are constrained by a fundamental limitation in their probabilistic modeling, which assumes that categorical probabilities are singularly deterministic [12, 14]. Specifically, vanilla KD implicitly assumes that for any given sample, its probability belonging to a particular class is deterministic and can be approximated by a deep neural network. In the case of a three-class classification problem, as shown in Figure 1a, this assumption is akin to identifying a single point on a 2D simplex that best represents the entire simplex. However, in practice, networks trained on finite data with limited capability exhibit inherent uncertainty in their predictions for specific samples [39, 47]. For example, networks with different initial weights may yield different predictions for the same test samples. Consequently, this singularly deterministic probabilistic modeling creates a bottleneck in the effective representation of diverse and uncertain knowledge.", + "bbox": [ + 93, + 90, + 480, + 512 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome the limitations of singularly deterministic probabilistic modeling, we employ a distribution-based probabilistic modeling, treating the categorical distribution as a random variable governed by a second-order distribution. Following Subjective Logic [35] and Dempster-Shafer Theory of Evidence [81], the Dirichlet distribution, a conjugate prior of the categorical likelihood, instantiates this second-order distribution, facilitating the success of evidential deep learning across multiple domains [5, 16, 25]. However, recent studies employing distribution-based probabilistic modeling are largely confined to macro-level training within a single network, where the second-order distribution is integrated into a single categorical distribution during the network's training process [8, 11, 64]. The second-order distribution inherently provides micro-level information across all possible categorical distributions, offering advantages in cross-network information propagation. Therefore, the second-order distribution serves as a bridge to enable knowledge transfer at both the macro and micro levels, supported by PAC-Bayesain theory [19]. As illustrated in Figure 1b, the logits output by each network are converted into a set of Dirichlet parameters, forming respective second-order distributions. Essentially, the second-order distribution assigns varying weights to all possible first-order categorical distributions,", + "bbox": [ + 93, + 515, + 480, + 890 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "where the non-uniformity of these weights reflects the network's capacity for classification. Building on second-order predictions, we propose the Evidential Knowledge Distillation (EKD), integrating both macro and micro perspectives. At the macro level, we reduce each second-order distribution to their first-order categorical distributions via an expectation operation in order to ensure that the centroids of their Dirichlet distributions coincide. By representing the entire simplex with a single point, the student is encouraged to focus more on the overall density distribution of the teacher at this stage, which leads to the proportions of student's outputs across categories aligning with those of the teacher. At the micro level, we replace the first-order distribution with the second-order distribution to enable the transfer of more granular classification information. As a complement to proportional alignment, this step refines the magnitude in the student network's outputs. Ultimately, the student model benefits from both the teacher's overall characteristics and its detailed classification structure.", + "bbox": [ + 517, + 90, + 903, + 376 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To summarize, Our main contribution are as follows:", + "bbox": [ + 535, + 378, + 880, + 392 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We reformulate network predictions in knowledge distillation through distribution-based probabilistic modeling. Building on this reformulation, we propose EKD, which distills the teacher's second-order distribution into the student model across both macro and micro levels.", + "- With the help of PAC-Bayesian theory [19], we prove that the distillation objective in EKD serves as an upper bound on the expected risk of the student network when treating the teacher network's outputs as ground truth.", + "- Extensive experiments are conducted on several standard benchmarks, encompassing a wide range of teacher and student models. Without bells and whistles, EKD achieves performance improvement of up to $3.61\\%$ over KD across various teacher-student network combinations." + ], + "bbox": [ + 516, + 393, + 903, + 603 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 618, + 653, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Knowledge Distillation Knowledge distillation enhances a simpler student network using insights from a complex teacher network. Existing methods fall into three categories: logit-based [2, 30, 44, 45, 54, 85, 88], feature-based [33, 56, 57, 70, 73], and hybrid approaches [22, 31, 34, 58, 80, 82]. Unlike feature-based methods, which demand architecture-specific designs, logit-based methods offer wider applicability by relying solely on network outputs [4, 53, 84]. Introduced by Hinton and Geoffrey [30], logit-based distillation converts teacher and student logs into categorical distributions aligned via KL divergence. Subsequent efforts have refined this by adjusting the temperature parameter dynamically, based on training stages [45], teacher-student gaps [21], and output variance [69], or by decomposing distributions [88] and leveraging intermediate networks [31, 54].", + "bbox": [ + 517, + 643, + 903, + 883 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The aforementioned traditional distillation methods", + "bbox": [ + 535, + 886, + 903, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "2815", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "share a common limitation in their reliance on point estimates for model parameters or categorical distributions, whereas Bayesian distillation [13, 37, 50, 74, 77] employs distributional estimation, which enables the student network to inherit the teacher's strengths in uncertainty estimation. Initial explorations [24, 37, 46, 51] adopt teacher ensembles to approximate the distribution of teacher model parameters and utilize the vanilla KD [30] for distillation. Subsequent extensions continue to advance by improving the student network architecture [20, 50, 61, 72, 76], optimization objectives [74], and sampling methods [13]. Additionally, some approaches introduce the utilization of quantized uncertainty during training to enable more fine-grained knowledge transfer [28, 66, 87]. While these methods introduce uncertainty into the model parameters and the resultant predictive distributions, they remain limited by their reliance on a finite number of ensembles to approximate the distributions. Consequently, we propose to adopt a continuous probabilistic modeling framework for both the teacher and the student, while leveraging more advanced distillation techniques. Unlike Bayesian distillation, which performs uncertainty-related distillation, we implicitly express the uncertainty of the predictive distribution through second-order probabilistic modeling and focus on enhancing the performance of the student network.", + "bbox": [ + 89, + 90, + 483, + 467 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidential Deep Learning Based on Subjective Logic (SL) [35] and Dempster-Shafer Theory (DST) [65], Sensoy et al. [64] initially propose a framework in which evidence collected by a deep neural network for each class is used to parameterize a second-order Dirichlet distribution for network predictions. This concept of evidence collection allows evidential networks to express \"I don't know\", showing advantages in uncertainty estimation. Since then, numerous studies have advanced and expanded upon Evidential Deep Learning (EDL) from both theoretical and application-oriented perspectives [17]. Theoretical investigations have primarily focused on refining the evidence collection process [11, 15, 18, 40, 55], delving into different training strategies [5, 26, 43] and extending evidential theory to regression networks [1, 49, 52]. Moreover, EDL has been widely applied in computer vision tasks, including action localization [5-7, 16], stereo matching [75], anomaly detection [68], and multi-view classification [25].", + "bbox": [ + 89, + 468, + 483, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 753, + 181, + 768 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first present the fundamental framework of our Evidential Knowledge Distillation (EKD), which includes both macroscopic first-order and microscopic second-order distribution distillation. Theoretically, we utilize PAC-Bayesian theory to demonstrate that the optimization objective of EKD effectively represents an upper bound on the student's expected risk. At the end of this section, we illustrate the complementary effect of the two", + "bbox": [ + 89, + 779, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "distillation strategies with a concrete toy case.", + "bbox": [ + 513, + 90, + 816, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 513, + 117, + 653, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Evidential Deep Learning. In accordance with the formulation of evidential deep learning [8, 9, 64], for a classification task involving $K$ classes, the categorical probability vector $p$ is treated as a random variable governed by a conjugate prior Dirichlet distribution. To define this distribution's parameters, an evidential activation function $\\sigma$ which we designate as the exponential function (exp), is employed to transform the network's output logits $z$ into a non-negative evidence vector $e = \\sigma(z)$ . This evidence vector is subsequently combined with a prior weight $\\lambda$ to yield the $K$ -dimensional Dirichlet parameters $\\alpha$ which fully specify a Dirichlet distribution $Dir(\\alpha)$ [8]. In terms of optimization, the widely adopted cross-entropy loss is computed by integrating over the categorical probabilities, guided by second-order the Dirichlet distribution.", + "bbox": [ + 511, + 138, + 906, + 364 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {E D L - c e} = \\int \\left[ \\sum_ {i = 1} ^ {K} - y _ {i} \\log \\left(p _ {i}\\right) \\right] D i r (\\boldsymbol {p} | \\boldsymbol {\\alpha}) d \\boldsymbol {p} \\\\ = \\sum_ {i = 1} ^ {K} y _ {i} (\\psi (\\alpha_ {0}) - \\psi (\\alpha_ {i})) \\tag {1} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 549, + 367, + 905, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Here, $y_{i}$ denotes the ground truth, $\\alpha_0 = \\sum_{i=1}^{K} (\\alpha_i)$ represents the Drichlet strength, and $\\psi$ refers to the digamma function. To support the acquisition of evidence, we employ Eq. (1), rather than the conventional cross-entropy loss, across all network training stages, including both the training of the teacher networks and student networks during the knowledge distillation process.", + "bbox": [ + 511, + 454, + 905, + 561 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Knowledge Distillation. Existing logit-based knowledge distillation methods rely on categorical distribution predictions derived from the softmax function. Specifically, for a given input sample $x$ , the network produces logits $z$ , which are then transformed into a categorical distribution via the softmax operation, denoted as $p = \\mathrm{softmax}(z)$ . Knowledge transfer is accomplished by minimizing the Kullback-Leibler divergence(KL) divergence between the teacher's and student's predicted distributions, $p^T$ and $p^S$ :", + "bbox": [ + 511, + 561, + 905, + 696 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {K D} = K L \\left(\\boldsymbol {p} ^ {T} \\| \\boldsymbol {p} ^ {S}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 696, + 903, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The teacher's predictions serve as soft labels, replacing the traditional one-hot labels, capturing inter-class similarities.", + "bbox": [ + 511, + 715, + 903, + 746 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Evidential Knowledge Distillation", + "text_level": 1, + "bbox": [ + 513, + 757, + 808, + 773 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Previous knowledge distillation methods have generally followed a unified paradigm, where the network's output logits are transformed into categorical probabilities via the softmax function, and the student network is guided to align its categorical distribution with that of the teacher network. While these approaches have yielded notable success, a fundamental limitation persists: they treat categorical probabilities as singularly deterministic values. This assumption", + "bbox": [ + 511, + 779, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "2816", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "overlooks the uncertainty in network predictions, thus constraining the sharing of more granular information. As a more robust alternative, we predict the distribution of categorical distributions using neural networks, specifically modeling it as a second-order Dirichlet distribution. More importantly, we align both the first-order distribution, obtained by integrating the second-order distribution, and the second-order distribution itself.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Firstly, we aim to extract the overall characteristics of the second-order distribution to capture the general distribution trend of probability density across the simplex, as well as to achieve alignment between networks in terms of final predicted classes. Specifically, given the second-order distributions predicted by the teacher and student, $Dir(\\pmb{\\alpha}^T)$ and $Dir(\\pmb{\\alpha}^S)$ , we compute their expectations where all possible categorical distributions are summed under the weights specified by the second-order distribution:", + "bbox": [ + 89, + 210, + 480, + 347 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} ^ {T} = \\mathbb {E} _ {\\boldsymbol {p} ^ {T} \\sim D i r (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\hat {\\boldsymbol {p}} ^ {S} = \\mathbb {E} _ {D i r (\\boldsymbol {p} ^ {S} \\sim \\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ] \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 347, + 482, + 367 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The expectation above, serving as the final predictive outcome for both teacher and student, is equivalent to the centroid of the second-order distribution, enabling the transfer of comprehensive probability density information via first-order distillation at a macro level.", + "bbox": [ + 89, + 367, + 482, + 441 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {1 s t} = K L (\\hat {\\boldsymbol {p}} ^ {T} | | \\hat {\\boldsymbol {p}} ^ {S}) = \\sum_ {i = 1} ^ {K} \\frac {\\alpha_ {i} ^ {T}}{\\alpha_ {0} ^ {T}} l o g \\left(\\frac {\\alpha_ {i} ^ {T} \\alpha_ {0} ^ {S}}{\\alpha_ {0} ^ {T} \\alpha_ {i} ^ {S}}\\right) \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 441, + 482, + 481 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Formally, optimizing Eq. (4) involves adjusting the proportion of each class's Dirichlet parameter relative to the total, which significantly enhances the alignment between the student's and teacher's predicted classes. In terms of formulation, Eq. (4) is very similar to vanilla KD, yet it is emphasized that vanilla KD can be regarded as a special case of first-order distillation when $\\sigma = \\exp(\\cdot)$ and $\\lambda = 0$ . Proof is provided in the Supplementary Material.", + "bbox": [ + 89, + 481, + 482, + 601 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The first-order distillation objective above only provides macroscopic guidance for the student's second-order distribution, failing to accurately and comprehensively represent the weighting of each categorical distribution. In other words, given a fixed teacher distribution, the student Dirichlet distribution that minimizes Eq. (4) is not unique. Therefore, we propose a microscopic second-order distillation as a complement to first-order distillation, aiming to provide fine-grained teacher supervision signals across the entire simplex for the student distribution. Specifically, we achieve this by applying the KL divergence directly between the two Dirichlet distributions:", + "bbox": [ + 89, + 602, + 482, + 782 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {2 n d} = K L (D i r (\\boldsymbol {p} ^ {T} | \\boldsymbol {\\alpha} ^ {T}) | | D i r (\\boldsymbol {p} ^ {S} | \\boldsymbol {\\alpha} ^ {S})) \\\\ = l o g \\frac {\\Gamma (\\alpha_ {0} ^ {T})}{\\Gamma (\\alpha_ {0} ^ {S})} - \\sum_ {i = 1} ^ {K} l o g \\frac {\\Gamma (\\alpha_ {i} ^ {T})}{\\Gamma (\\alpha_ {i} ^ {S})} \\\\ + \\sum_ {i = 1} ^ {K} \\left(\\alpha_ {i} ^ {T} - \\alpha_ {i} ^ {S}\\right) \\left(\\psi \\left(\\alpha_ {i} ^ {T}\\right) - \\psi \\left(\\alpha_ {0} ^ {T}\\right)\\right) \\tag {5} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 782, + 482, + 891 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\Gamma(\\cdot)$ denotes the Gamma function. The detailed derivation for Eq. (5) is provided in Supplementary Material. Since Eq. (5) considers the whole $(K - 1)$ dimensional simplex, overflow may occur [50]; thus, we apply the $\\log(1 + x)$ function to smooth the network outputs. While the above Eq. (4) aims to optimize the proportional relationships between classes, Eq. (5) here places greater emphasis on aligning the magnitudes within each class. Here, the alignment within the class means that the gradients of a loss function with respect to the output for a particular class do not conflict with the gradients of the output for other classes. This meticulous numerical alignment supplies each point on the $(K - 1)$ -dimensional simplex of the student's second-order distribution with the corresponding teacher label, ensuring a precise mapping of knowledge across the entire distribution space. By doing so, it strives to faithfully replicate the teacher's second-order predictive nuances at a granular, micro-level, capturing the fine-grained patterns and uncertainties inherent in the teacher model's outputs.", + "bbox": [ + 511, + 90, + 903, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "By combining these two complementary optimization objectives, one optimizing proportional relationships at the macro level and the other aligning numerical values at the micro level, we arrive at the final EKD distillation:", + "bbox": [ + 511, + 378, + 903, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {E K D} = \\mathcal {L} _ {1 s t} + \\gamma \\mathcal {L} _ {2 n d}, \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 440, + 903, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\gamma$ represents the coefficient that adjusts the balance between the first-order and second-order distillation.", + "bbox": [ + 511, + 458, + 903, + 486 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Theoretical Analysis", + "text_level": 1, + "bbox": [ + 513, + 503, + 705, + 520 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we provide theoretical support for Eq. (6) using PAC-Bayesian Theory [19]. While PAC-Bayesian theory has been explored in the context of the Dirichlet distribution [26], its application remains confined to single-network optimization, neglecting the relationship among multiple networks. Therefore, we aim to estimate the alignment of different networks across the entire data distribution based on their alignment on training samples.", + "bbox": [ + 511, + 526, + 903, + 647 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Let $x$ denote a certain sample and $y$ denote its corresponding label. Under the PAC setting, samples are drawn from a fixed but unknown distribution $\\mathcal{D}$ . A classifier $h$ maps $x$ to $y$ . Moreover, we define a classifier $h$ correspond to a point on the $(K - 1)$ -dimensional simplex, i.e., a categorical distribution $p$ . Essentially, the Dirichlet distribution made by each network serves as the posterior distribution of a set of classifiers. By weighting the set of classifiers with its posterior distribution, we obtain the final majority vote classifier (or Bayesain classifier):", + "bbox": [ + 511, + 648, + 905, + 797 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {B} ^ {T} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 797, + 903, + 816 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {B} ^ {S} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ], \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 819, + 903, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{B}^{T}$ and $\\pmb{B}^{S}$ denote the final classification results of the teacher and student classifiers, respectively.", + "bbox": [ + 511, + 839, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In knowledge distillation, the student treats the teacher's predictions as soft labels. Below, we define the student's", + "bbox": [ + 511, + 869, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "2817", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "misclassification risk using cross-entropy, which is equivalent to KL divergence from an optimization perspective.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nR ^ {S} = \\mathbb {E} _ {x \\sim \\mathcal {D}} \\left[ - \\boldsymbol {B} ^ {T} \\log \\boldsymbol {B} ^ {S} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 119, + 482, + 140 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {R} ^ {S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ - \\boldsymbol {B} _ {i} ^ {T} \\log \\boldsymbol {B} _ {i} ^ {S} \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 143, + 482, + 183 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $R^S$ represents the expected risk of the student over the entire data distribution $\\mathcal{D}$ and $\\hat{R}^S$ denotes the empirical risk on the training set $\\mathcal{D}_t$ containing $N$ training samples.", + "bbox": [ + 89, + 185, + 483, + 231 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Theorem 1. For any $\\delta \\in (0,1]$ and $\\gamma \\in \\mathbb{R}^{+}$ , $R^S$ is bounded above by the sum of $\\hat{R}^S$ , $\\sum_{1}^{N}KL(Dir(\\pmb{\\alpha}_i^T))||Dir(\\pmb{\\alpha}_i^S))$ , and a constant related to $\\delta$ , denoted as $C(\\delta)$ :", + "bbox": [ + 89, + 238, + 483, + 286 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} P _ {\\mathcal {D} _ {t} \\sim \\mathcal {D} ^ {N}} (R ^ {S} \\leq \\hat {R} ^ {S} + \\frac {\\gamma}{N} \\sum_ {i = 1} ^ {N} K L (D i r (\\pmb {\\alpha} _ {i} ^ {T}) | | D i r (\\pmb {\\alpha} _ {i} ^ {S})) \\\\ + C (\\delta)) \\geq 1 - \\delta . \\tag {11} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 287, + 480, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Detailed proof and definition of $C(\\delta)$ can be found in Supplementary Material. Theorem 1 indicates that the upper bound on the student's excepted risk over the unknown data distribution $\\mathcal{D}$ consists of two components: the alignment of the first-order distribution with the true labels (teacher labels) and the similarity between the predicted second-order distribution and the true second-order distribution (teacher distribution). Clearly, this upper bound indicates that the empirical risk optimization represented by Eq. (2) does not guarantee a reduction in the student's expected risk, suggesting that the transfer of the teacher's generalization performance is both inefficient and incomplete. In contrast, EKD directly adopts this upper bound as optimization objectives, refining the formulation of empirical risk while incorporating second-order distribution alignment between the student and teacher, thereby striving to minimize the student's expected risk.", + "bbox": [ + 88, + 371, + 483, + 628 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Toy Case", + "text_level": 1, + "bbox": [ + 89, + 637, + 197, + 652 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Figure 2 demonstrates the complementary roles of first-order and second-order distillation through a three-class example. Initially, the student Dirichlet parameters differ significantly from those of the teacher, both in relative proportion and in absolute values. The left side of Figure 2 illustrates the essence of first-order distillation, which minimizes the distance between the expectations of two Dirichlet distributions over the $(K - 1)$ dimensional simplex measured by KL divergence. However, the expectation of the Dirichlet distribution depends solely on the proportion of each class's Dirichlet parameter relative to the overall sum, which can lead to discrepancies in the specific values of the student's output compared to the teacher's. Although this discrepancy does not affect the final class prediction, it causes the student's second-order distribution to differ significantly from the teacher's. To address this, we use", + "bbox": [ + 89, + 659, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/648f6e6a352fa9bba6eba19e4d2833bd02ba36a33432a16289f374833fcf0545.jpg", + "image_caption": [ + "Figure 2. A toy case illustrating the complementary roles of the two distillation losses in EKD. Given the Dirichlet parameters output by teacher and student networks, the left panel shows the results when only first-order distillation is used to optimize the proportion of each category. The right panel presents the outcome when only second-order distillation is applied for intra-class numerical alignment. The center panel presents the results of combining both objectives, which yields the optimal performance." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 905, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "second-order distillation as a complement to first-order distillation, aiming to align the Dirichlet parameter values. In contrast to first-order distillation, second-order distribution alignment encourages consistency of parameters within the same class across different networks, disregarding the relative magnitudes among different classes. As a result, we combine these complementary distillation objectives to achieve a comprehensive alignment of the second-order distribution. The quantitative effects of the two types of distillation will be discussed in detail in Section 4.2", + "bbox": [ + 511, + 523, + 906, + 676 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 694, + 645, + 710 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Our experiments are primarily conducted on two standard datasets including CIFAR100 [38] and ImageNet [60]. The CIFAR-100 dataset [38] consists of 60k 32x32 color images divided into 100 classes, with each class containing 600 images. It is split into a training set of 50k images and a validation set of 10k images. ImageNet [60] is a large-scale visual database comprising 1,000 categories that cover a diverse range of objects and scenes. It includes approximately 1.28 million labeled images for training and an additional 50,000 images for validation.", + "bbox": [ + 511, + 719, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We adopt the experimental settings established in prior studies [10, 69, 88]. For the", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "2818", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/7c6ebc9551dddf7a941ac0cd60408c70d8cc2bf2512deac321d9f10e15c670b6.jpg", + "table_caption": [ + "Table 1. Results on the CIFAR-100 validation set. The teacher and student networks share the same architecture. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. $\\Delta$ denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + ], + "table_footnote": [], + "table_body": "
Experiment Group1234567
TeacherArchitectureResNet32×4VGG13WRN-40-2WRN-40-2ResNet56ResNet110ResNet110
Softmax79.4274.6475.6175.6172.3474.3174.31
EDL79.5374.9675.4175.4173.0974.6774.67
StudentArchitectureResNet8×4VGG8WRN-40-1WRN-16-2ResNet20ResNet32ResNet20
Softmax72.5070.3671.9873.2669.0671.1469.06
EDL72.7770.6771.8473.3669.2270.969.22
FeatureRKD[CVPR'19] [56]71.9071.4872.2273.3569.6171.8269.25
CRD[ICLR'20] [70]75.5173.9474.1475.4871.1673.4871.46
ReviewKD[CVPR'21] [10]75.6374.8475.0976.1271.8973.8971.34
SimKD[CVPR'22] [3]78.0874.8974.5375.5371.0573.9271.06
CAT-KD[CVPR'23] [23]76.9174.6574.8275.6071.6273.6271.37
TopKD[ICML'24] [36]75.4074.0174.4375.7571.5873.7771.47
LogitKD[NeurIPS'14] [30]73.3372.9873.5474.9270.6673.0870.67
Δ0.832.621.561.661.61.941.61
DKD[CVPR'22] [88]76.3274.6874.8176.2471.9774.1171.06
Δ3.824.322.832.982.912.972.0
CTKD[AAAI'23] [45]73.3973.5273.9375.4571.1973.5270.99
Δ0.893.161.952.192.132.381.93
Logit_Stand[CVPR'24] [69]76.6274.3674.3776.1171.4374.1771.48
Δ4.124.002.392.852.373.032.42
SDD[CVPR'24] [48]75.92*73.90*74.26*75.98*70.72*73.90*71.43*
Δ3.42*3.54*2.28*2.72*1.66*2.76*2.37*
TeKAP[ICLR'25] [31]74.7973.8*73.8075.2171.7173.50*71.00*
Δ2.293.44*1.821.952.652.36*1.94*
EKD(Ours)77.2174.7174.4376.1571.4873.6871.51
Δ4.444.042.592.792.262.782.29
", + "bbox": [ + 91, + 146, + 919, + 467 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8462bab17d4b7f13be597859e7b8ee18fea8c62db8761bc706e94543e9440645.jpg", + "table_caption": [ + "Table 3. The top-1 and top-5 accuracy (\\%) on the ImageNet validation set [60]. The best and second best results are emphasized in bold and underlined. “*” indicates our reproduced results." + ], + "table_footnote": [], + "table_body": "
TeacherArchitectureResNet34ResNet50
Accuracytop-1top-5top-1top-5
Softmax73.3191.4276.1692.86
EDL73.8391.6976.1092.98
StudentArchitectureResNet18MNV1
Accuracytop-1top-5top-1top-5
Softmax69.7589.0768.8788.76
EDL70.2889.5869.9289.46
FeatureCRD [70]71.1790.1371.3790.41
ReviewKD [10]71.6190.5172.5691.00
SimKD [3]71.5990.4872.2590.86
CAT-KD [23]71.2690.4572.2491.13
LogitKD [30]71.0390.0570.5089.80
DKD [88]71.7090.4172.0591.05
CTKD [45]71.3890.2771.1690.11
Logit_Stand [69]71.4290.2972.1890.80
SDD [48]71.4490.0572.15*90.79*
EKD(Ours)71.7390.6972.5491.20
", + "bbox": [ + 99, + 532, + 475, + 750 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "experiments on CIFAR-100, we use the SGD optimizer with 240 epochs. The learning rate is set to 0.01 for MobileNets [32, 63] and ShuffleNets [86] and 0.05 for other networks, including ResNets [27], WRNs [83] and VGGs [67], with a decay factor of 0.1 applied at epochs 150, 180, and 210. Regarding hyperparameter configuration, we set $\\gamma$ in Eq. (6) to 1 by default, requiring no additional hyperparameter tuning. All results represent the mean of five", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "independent runs. Additional experimental details are provided in the Supplementary Material.", + "bbox": [ + 511, + 491, + 906, + 521 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. To benchmark EKD, we incorporate five state-of-the-art logit-based distillation methods: DKD (CVPR'22) [88], CTKD (AAAI'23) [45], Logit_Stand (CVPR'24) [69], SDD (CVPR'24) [48] and TeKAP (ICLR'25) [31]. Additionally, a selection of feature-based distillation methods [3, 10, 23, 56, 59, 70] is included as references to facilitate a broader contextual analysis.", + "bbox": [ + 511, + 521, + 906, + 627 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Main Results", + "text_level": 1, + "bbox": [ + 511, + 637, + 653, + 652 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Performance comparison on CIFAR100. We conduct experiments comparing EKD with other distillation methods under various teacher-student pair settings. Table 1 and Table 2 present the accuracy results for settings where the teacher and student have identical and distinct architectures. As presented in Table 1, EKD demonstrates favourable performance improvements over KD, ranging from $0.66\\%$ to $3.61\\%$ . In particular, EKD demonstrates its strongest performance in Groups 1, 2, and 7, while achieving comparable results to DKD [88] in Groups 3-6. The underlying reason for this phenomenon lies in the performance gap between the teacher and student. In groups 1, 2, and 7, the average performance gap between the teacher and student is $5.51\\%$ , whereas it is $3.31\\%$ in groups 3-6. Furthermore, we observe that pairs with a larger performance gap tend to have a higher distillation loss, indicating a greater opti", + "bbox": [ + 511, + 657, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "2819", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b3ed2da6ce7d5eccf2c40b89df48afd6bc646c926e55c70264322d24c9551638.jpg", + "table_caption": [ + "Table 2. Results on the CIFAR-100 validation set. The teacher and student networks have different architectures. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. $\\Delta$ denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + ], + "table_footnote": [], + "table_body": "
Experiment Group1234567
TeacherArchitectureResNet32×4ResNet32×4ResNet32×4WRN-40-2WRN-40-2VGG13ResNet50
Softmax79.4279.4279.4275.6175.6174.6479.34
EDL79.5379.5379.5375.4175.4175.1179.69
StudentArchitectureSHN-V2WRN-16-2WRN-40-2ResNet8×4MN-V2MN-V2MN-V2
Softmax71.8273.2675.6172.5064.6064.6064.60
EDL72.7373.3675.4172.7765.3965.3965.39
FeatureRKD[CVPR'19] [56]73.2174.8677.8275.2669.2764.5264.43
CRD[ICLR'20] [70]75.6575.6578.1575.2470.2869.7369.11
ReviewKD[CVPR'21] [10]77.7876.1178.9674.3471.2870.3769.89
SimKD[CVPR'22] [3]78.3977.1779.2975.2970.1069.4469.97
CAT-KD[CVPR'23] [23]78.4176.9778.5975.3870.2469.1371.36
TopKD[ICML'24] [36]76.3374.34*76.18*75.83*69.54*69.32*70.11*
LogitKD[NeurIPS'14] [30]74.4574.9077.7073.9768.3667.3767.35
Δ2.631.642.091.473.762.772.75
DKD[CVPR'22] [88]77.0775.7078.4675.5669.2869.7170.35
Δ5.252.442.853.064.685.115.75
CTKD[AAAI'23] [45]75.3774.5777.6674.6168.3468.5068.67
Δ3.551.312.052.113.743.904.07
Logit_Stand[CVPR'24] [69]75.5675.2677.9277.1169.2368.6169.02
Δ3.742.002.314.614.634.014.42
SDD[CVPR'24] [48]76.88*75.92*78.54*74.83*70.14*69.06*69.31*
Δ5.06*2.66*2.93*2.33*5.54*4.46*4.71*
TeKAP[ICLR'25] [31]75.4375.26*76.93*74.99*68.65*68.27*68.75*
Δ3.612.00*1.32*2.49*4.05*3.67*4.15*
EKD(Ours)77.6576.5178.9176.7569.2069.9470.38
Δ4.923.153.503.983.814.554.99
", + "bbox": [ + 91, + 146, + 928, + 467 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/cc3aca791d747bf599cef20da1782b80841666fd194c4aba5930849571f89913.jpg", + "table_caption": [ + "Table 4. Ablation study for each components in EKD. The \"MSE\" refers to aligning the magnitude of model outputs with mean square error. The \"Accuracy\" denotes the Top-1 accuracy (\\%) on the CIFAR100 validation set. ResNet32×4 and ResNet8×4 are adopted as teacher and student." + ], + "table_footnote": [], + "table_body": "
IndexDistillation ObjectiveAccuracy
\\( \\mathcal{L}_{KD} \\)\\( \\mathcal{L}_{1st} \\)\\( \\mathcal{L}_{2nd} \\)MSE
1✓74.45
2✓73.80
3✓76.42
4✓76.10
5✓✓77.21
6✓✓76.83
7✓✓76.34
8✓✓76.13
", + "bbox": [ + 133, + 561, + 444, + 691 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "mization space during distillation. Conversely, for teacher-student pairs with closer performance, their distillation loss is generally smaller, leading EKD to encounter local minima during optimization. For the distillation of distinct networks shown in Table 2, EKD achieves the highest performance in half of the experimental groups. When using MobileNet [63] as the student network, the student trained solely with EDL achieves a $0.79\\%$ higher performance than the one trained with softmax alone. However, the result for EKD is only comparable to that of other methods. We attribute this phenomenon to the inherent performance ceiling of the student network due to its limited capacity.", + "bbox": [ + 89, + 719, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Performance comparison on ImageNet. Table 3 presents the performance of EKD on the large-scale ImageNet dataset. EKD markedly outperforms other logit-based methods in the ResNet50-MNV1 experiment and achieves results comparable to top methods in the ResNet34-ResNet18 experiment. The modest gains in the latter can also be attributed to the narrow performance gap between the teacher and student.", + "bbox": [ + 511, + 491, + 906, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 622, + 679, + 636 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To elucidate the superiority of the two alignment strategies implemented in EKD and their alignment with theoretical underpinnings, the ablation study presented in Table 4 evaluates various distillation objectives on student model performance. In addressing alignment with respect to the proportional distribution across output classes, we investigate two methodologies: the conventional vanilla KD approach presented in Eq. (2), which aligns softmax-based categorical distributions, and $\\mathcal{L}_{1st}$ in Eq. (4). Comparative analyses between Index 1 and Index 2, Index 5 and Index 7, and Index 6 and Index 8 consistently demonstrate that $\\mathcal{L}_{1st}$ surpasses vanilla KD. This finding underscores that EKD's first-order alignment serves as a robust generalization of vanilla KD, effectively incorporating the alignment of categorical distributions by leveraging the centroids of second-order Dirichlet distributions.", + "bbox": [ + 511, + 643, + 906, + 883 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In addition to the second-order distillation, $\\mathcal{L}_{2nd}$ , shown", + "bbox": [ + 531, + 885, + 906, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "2820", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ee2b7413342f96963665e650abd61aa82ca0a0a2399e4dd9d95c3511450cbf28.jpg", + "table_caption": [ + "Table 5. The Top-1 accuracy $(\\%)$ of combining EKD with feature distillation-based methods on CIFAR-100." + ], + "table_footnote": [], + "table_body": "
TeacherArchitectureVGG13ResNet32x4
Accuracy74.9679.53
StudentArchitectureVGG8SHN-V2
Accuracy70.6772.73
LogitEKD74.7177.65
FeatureRKD [56]71.4873.21
CAT-KD [23]74.6578.41
Logit+FeatureEKD+RKD [56]74.99↑78.08↑
EKD+CAT-KD [23]74.91↑78.46↑
", + "bbox": [ + 106, + 119, + 470, + 244 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "in Eq. (5), another intuitive approach to aligning model output magnitudes is through mean squared error (MSE). Consequently, we compare $\\mathcal{L}_{2nd}$ and MSE in Table 4. Comparisons across Index 3 and Index 4, Index 5 and Index 6, and Index 7 and Index 8 reveal that $\\mathcal{L}_{2nd}$ consistently outperforms MSE. These empirical results corroborate the theoretical conclusions delineated in Section 3.3. It is suggested that $\\mathcal{L}_{2nd}$ , being part of the upper bound on the student's expected risk, directly optimizes the student's generalization performance, whereas the MSE objective does so indirectly.", + "bbox": [ + 89, + 275, + 483, + 428 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Further Remarks", + "text_level": 1, + "bbox": [ + 89, + 440, + 264, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations. Figure 3 presents the t-SNE visualizations of the laerned feature representations and, in parallel, the differences in logits between the teacher and student networks, where the teacher is implemented as ResNet $32 \\times 4$ and the student as ResNet $8 \\times 4$ . As shown in the first row, EKD achieves a clear improvement over vanilla KD baseline with respect to class separability, leading to more distinctly clustered feature distributions. Furthermore, the logits produced under EKD exhibit a noticeably stronger alignment with those of the teacher network, indicating a closer match in predictive behavior compared to the results obtained with vanilla KD.", + "bbox": [ + 89, + 462, + 483, + 642 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Compatibility with feature-based methods. Table 5 presents the results of integrating EKD with several representative feature-based distillation methods [23, 56]. For feature-based approaches, the introduction of EKD enhances their performance, demonstrating strong compatibility, particularly with methods that originally exhibit lower performance [56]. Conversely, these methods can further compensate for the limitation of the EKD approach in expressing knowledge through logits.", + "bbox": [ + 89, + 643, + 483, + 779 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Distilling ViT. Transformer-based models are gaining increasing popularity, and logit-based distillation methods can be easily applied to these models. Therefore, we test the distillation effectiveness of EKD on the ViT model, as presented in Table 6. Following the configuration in [41, 42], we use ResNet56 as the teacher model and four ViT variants, DeiT-Ti [71], PiT-Ti [29], PVT-Ti [78], and PVTv2 [79], as the student models. EKD achieves an aver-", + "bbox": [ + 89, + 780, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2a7f338f5bb870264b0798680b7e76ebaa01936905edbd95875230ec0d3c5c14.jpg", + "image_caption": [ + "(a) Vanilla KD" + ], + "image_footnote": [], + "bbox": [ + 535, + 95, + 687, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2594fa2eed89f487982a5fde7f4bb9a2dfa2f501723ccda403b4d9ca248de348.jpg", + "image_caption": [ + "(b) EKD" + ], + "image_footnote": [], + "bbox": [ + 718, + 95, + 870, + 212 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/dbe4e60df6dcc6342f43e51ad612ee43a24b0c462fce1c020766092fcf2b4486.jpg", + "image_caption": [ + "(c) Vanilla KD" + ], + "image_footnote": [], + "bbox": [ + 534, + 229, + 699, + 334 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3a3fa36847db0e6259a3d13b6b8998a4aecef64ecdf6b4e2814fea8ab6df344f.jpg", + "image_caption": [ + "(d) EKD", + "Figure 3. Visualization of features and logits output. The first row shows the t-SNE of features learned by vanilla KD and EKD. The second row illustrates the difference of correlation matrices of logits between the student and teacher." + ], + "image_footnote": [], + "bbox": [ + 718, + 229, + 883, + 335 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/34e4f10813de0ce0107a57a32831453f06b9607532f5f128a12be563ebe5f399.jpg", + "table_caption": [ + "Table 6. Top-1 accuracy of various ViT student models on CIFAR100. Teacher model is ResNet56." + ], + "table_footnote": [], + "table_body": "
ArchitectureSizeEDLSoftmaxKD [30]AutoKD [42]EKD(Ours)
DeiT-Ti [71]5M64.7765.0873.2578.5578.64
PiT-Ti [29]5M73.4873.5875.4778.7679.33
PVT-Ti [78]13M69.0269.2273.6078.4378.74
PVTv2 [79]3M76.6877.4478.8179.3779.80
", + "bbox": [ + 514, + 455, + 910, + 546 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "age performance improvement of $3.75\\%$ over KD [30] and an average improvement of $0.25\\%$ over AutoKD [42].", + "bbox": [ + 511, + 571, + 906, + 603 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 618, + 633, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper reinterprets vanilla knowledge distillation (KD) through a distribution-based probabilistic framework, employing evidential second-order distributions to effectively capture predictive uncertainty and provide a comprehensive knowledge representation. This pioneering methodology underpins Evidential Knowledge Distillation (EKD), facilitating knowledge transfer across macro and micro levels. At the macro level, aligning the expectation of the second-order distribution, which reflects its global characteristics, enhances optimization of inter-class proportions in the student's output. At the micro level, aligning the second-order distribution aligns the magnitudes of the student's output. These complementary mechanisms ensure a robust distillation process. Through PAC-Bayesian theory, EKD optimizes the upper bound of the student's expected risk. Extensive experiments underscore EKD's pronounced advantage over prevailing logit-based distillation techniques.", + "bbox": [ + 509, + 643, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "2821", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported in part by the National Key Research and Development Plan of China under Grant 2023YFC3310700, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62472422, and U2333215, and in part by Beijing Natural Science Foundation under Grant 4242051.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 219, + 187, + 234 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. NIPS, 33:14927-14937, 2020. 3", + "[2] Defang Chen, Jian-Ping Mei, Can Wang, Yan Feng, and Chun Chen. Online knowledge distillation with diverse peers. In AAAI, pages 3430-3437, 2020. 2", + "[3] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, pages 11933-11942, 2022. 6, 7", + "[4] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. NIPS, 30, 2017. 2", + "[5] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV, pages 192-208. Springer, 2022. 2, 3", + "[6] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Cascade evidential learning for open-world weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14741-14750, 2023.", + "[7] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Uncertainty-aware dual-evidential learning for weakly-supervised temporal action localization. IEEE transactions on pattern analysis and machine intelligence, 45(12):15896-15911, 2023. 3", + "[8] Mengyuan Chen, Junyu Gao, and Changsheng Xu. R-edl: Relaxing nonessential settings of evidential deep learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3", + "[9] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Revisiting essential and nonessential settings of evidential deep learning. arXiv preprint arXiv:2410.00393, 2024. 3", + "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, pages 5008-5017, 2021. 5, 6, 7", + "[11] Danruo Deng, Guangyong Chen, Yang Yu, Furui Liu, and Pheng-Ann Heng. Uncertainty estimation by fisher information-based evidential deep learning. In ICML, pages 7596–7616. PMLR, 2023. 2, 3", + "[12] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2", + "[13] Luyang Fang, Yongkai Chen, Wenxuan Zhong, and Ping Ma. Bayesian knowledge distillation: A bayesian perspec" + ], + "bbox": [ + 93, + 244, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "tive of distillation with uncertainty quantification. In Forty-first ICML, 2024. 3", + "[14] Yarin Gal et al. Uncertainty in deep learning. 2016. 2", + "[15] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18827-18836, 2023. 3", + "[16] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Vectorized evidential learning for weakly-supervised temporal action localization. PAMI, 2023. 2, 3", + "[17] Junyu Gao, Mengyuan Chen, Liangyu Xiang, and Changsheng Xu. A comprehensive survey on evidential deep learning and its applications. arXiv preprint arXiv:2409.04720, 2024.3", + "[18] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Learning probabilistic presence-absence evidence for weakly-supervised audio-visual event perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3", + "[19] Pascal Germain, Alexandre Lacasse, François Laviolette, and Mario Marchand. Pac-bayesian learning of linear classifiers. In ICML, pages 353-360, 2009. 2, 4", + "[20] Hongji Guo, Hanjing Wang, and Qiang Ji. Bayesian evidential deep learning for online action detection. In ECCV, pages 283-301. Springer, 2025. 3", + "[21] Jia Guo. Reducing the teacher-student gap via adaptive temperatures. 2022. 2", + "[22] Yong Guo, Shulian Zhang, Haolin Pan, Jing Liu, Yulun Zhang, and Jian Chen. Gap preserving distillation by building bidirectional mappings with a dynamic teacher. arXiv preprint arXiv:2410.04140, 2024. 2", + "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, pages 11868-11877, 2023. 6, 7, 8", + "[24] Corina Gurau, Alex Bewley, and Ingmar Posner. Dropout distillation for efficiently estimating model confidence. arXiv preprint arXiv:1809.10562, 2018. 3", + "[25] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 2, 3", + "[26] Manuel Haussmann, Sebastian Gerwinn, and Melih Kandemir. Bayesian evidential deep learning with pac regularization. arXiv preprint arXiv:1906.00816, 2019. 3, 4", + "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6", + "[28] Mina Hemmatian, Ali Shahzadi, and Saeed Mozaffari. Uncertainty-based knowledge distillation for bayesian deep neural network compression. International Journal of Approximate Reasoning, 175:109301, 2024. 3", + "[29] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In CVPR, pages 11936-11945, 2021. 8", + "[30] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 3, 6, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "2822", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[31] Md Imtiaz Hossain, Sharmen Akhter, Choong Seon Hong, and Eui-Nam Huh. Single teacher, multiple perspectives: Teacher knowledge augmentation for enhanced knowledge distillation. In ICLR, 2025. 2, 6, 7", + "[32] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6", + "[33] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. NIPS, 35:33716-33727, 2022. 2", + "[34] Yaomin Huang, Zaomin Yan, Chaomin Shen, Faming Fang, and Guixu Zhang. Harmonizing knowledge transfer in neural network with unified distillation. In ECCV, pages 58-74. Springer, 2024. 2", + "[35] Audun Jøsang. Subjective logic. Springer, 2016. 2, 3", + "[36] Jungeun Kim, Junwon You, Dongjin Lee, Ha Young Kim, and Jae-Hun Jung. Do topological characteristics help in knowledge distillation? In ICML, 2024. 6, 7", + "[37] Anoop Korattikara Balan, Vivek Rathod, Kevin P Murphy, and Max Welling. Bayesian dark knowledge. NIPS, 28, 2015. 3", + "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5", + "[39] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. NIPS, 30, 2017. 2", + "[40] Changbin Li, Kangshuo Li, Yuzhe Ou, Lance M Kaplan, Audun Jøsang, Jin-Hee Cho, Dong Hyun Jeong, and Feng Chen. Hyper evidential deep learning to quantify composite classification uncertainty. arXiv preprint arXiv:2404.10980, 2024.3", + "[41] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, pages 110-127. Springer, 2022. 8", + "[42] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In CVPR, pages 17413-17424, 2023. 8", + "[43] Xue Li, Wei Shen, and Denis Charles. Tedl: A two-stage evidential deep learning method for classification uncertainty quantification. arXiv preprint arXiv:2209.05522, 2022. 3", + "[44] Zheng Li, Ying Huang, Defang Chen, Tianren Luo, Ning Cai, and Zhigeng Pan. Online knowledge distillation via multi-branch diversity enhancement. In Proceedings of the Asian conference on computer vision, 2020. 2", + "[45] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 6, 7", + "[46] Jakob Lindqvist, Amanda Olmin, Fredrik Lindsten, and Lennart Svensson. A general framework for ensemble distribution distillation. In 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2020. 3", + "[47] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[48] Shicai Wei Chunbo Luo Yang Luo. Scale decoupled distillation. arXiv preprint arXiv:2403.13512, 2024. 6, 7", + "[49] Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions. NIPS, 34:6881-6893, 2021. 3", + "[50] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 3, 4", + "[51] Zelda E Mariet, Rodolphe Jenatton, Florian Wenzel, and Dustin Tran. Distilling ensembles improves uncertainty estimates. In Third symposium on advances in approximate bayesian inference, 2021. 3", + "[52] Nis Meinert and Alexander Lavin. Multivariate deep evidential regression. arXiv preprint arXiv:2104.06135, 2021. 3", + "[53] Zhong Meng, Jinyu Li, Yong Zhao, and Yifan Gong. Conditional teacher-student learning. In ICASSP, pages 6445-6449. IEEE, 2019. 2", + "[54] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, pages 5191-5198, 2020. 2", + "[55] Deep Shankar Pandey and Qi Yu. Learn to accumulate evidence from all training samples: theory and practice. In ICML, pages 26963-26989. PMLR, 2023. 3", + "[56] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, pages 3967-3976, 2019. 2, 6, 7, 8", + "[57] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, pages 5007-5016, 2019. 2", + "[58] Bo Peng, Zhen Fang, Guangquan Zhang, and Jie Lu. Knowledge distillation with auxiliary variable. In ICML, 2024. 2", + "[59] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 6", + "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 5, 6", + "[61] Max Ryabinin, Andrey Malinin, and Mark Gales. Scaling ensemble distribution distillation to many classes with proxy targets. NIPS, 34:6023-6035, 2021. 3", + "[62] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference. In HPEC, pages 1-9. IEEE, 2023. 1", + "[63] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6, 7", + "[64] Murat Sensoy, Lance Kaplan, and Melih Kandemir. *Evidential deep learning to quantify classification uncertainty.* NIPS, 31, 2018. 2, 3" + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "2823", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[65] Glenn Shafer. Dempster-shafer theory. Encyclopedia of artificial intelligence, 1:330-331, 1992. 3", + "[66] Yichen Shen, Zhilu Zhang, Mert R Sabuncu, and Lin Sun. Real-time uncertainty estimation in computer vision via uncertainty-aware distribution distillation. In WACV, pages 707-716, 2021. 3", + "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[68] Che Sun, Yunde Jia, and Yuwei Wu. Evidential reasoning for video anomaly detection. In ACMMM, pages 2106-2114, 2022. 3", + "[69] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, pages 15731-15740, 2024. 2, 5, 6, 7", + "[70] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 2, 6, 7", + "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 8", + "[72] Linh Tran, Bastiaan S Veeling, Kevin Roth, Jakub Swiatkowski, Joshua V Dillon, Jasper Snoek, Stephan Mandt, Tim Salimans, Sebastian Nowozin, and Rodolphe Jenatton. Hydra: Preserving ensemble diversity for model distillation. arXiv preprint arXiv:2001.04694, 2020. 3", + "[73] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In CVPR, pages 1365-1374, 2019. 2", + "[74] Meet Vadera, Brian Jalaian, and Benjamin Marlin. Generalized bayesian posterior expectation distillation for deep neural networks. In Conference on Uncertainty in Artificial Intelligence, pages 719-728. PMLR, 2020. 3", + "[75] Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun Zhou, and Edwin Hancock. Uncertainty estimation for stereo matching based on evidential deep learning. pattern recognition, 124:108498, 2022. 3", + "[76] Hanjing Wang and Qiang Ji. Beyond dirichlet-based models: when bayesian neural networks meet evidential deep learning. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 3", + "[77] Kuan-Chieh Wang, Paul Vicol, James Lucas, Li Gu, Roger Grosse, and Richard Zemel. Adversarial distillation of bayesian neural network posteriors. In ICML, pages 5190-5199. PMLR, 2018. 3", + "[78] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, pages 568-578, 2021. 8", + "[79] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 8", + "[80] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, pages 588-604. Springer, 2020. 2" + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[81] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions. Springer, 2008. 2", + "[82] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. ICLR (ICLR), 2021. 2", + "[83] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6", + "[84] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In CVPR, pages 3517-3526, 2019. 2", + "[85] Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Jialu Liu, Michael Bendersky, Marc Najork, and Chao Zhang. Do not blindly imitate the teacher: Using perturbed loss for knowledge distillation. arXiv preprint arXiv:2305.05010, 2023. 2", + "[86] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 6", + "[87] Yuan Zhang, Weihua Chen, Yichen Lu, Tao Huang, Xiuyu Sun, and Jian Cao. Avatar knowledge distillation: self-ensemble teacher paradigm with uncertainty. In ACMMM, pages 5272-5280, 2023. 3", + "[88] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, pages 11953-11962, 2022. 1, 2, 5, 6, 7", + "[89] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In CVPR, pages 16965-16974, 2024. 1" + ], + "bbox": [ + 516, + 90, + 903, + 503 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "2824", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_model.json b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5ca4b908e84fe1dc32506cab668a322dc11d8f84 --- /dev/null +++ b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_model.json @@ -0,0 +1,2532 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.328, + 0.131, + 0.672, + 0.154 + ], + "angle": 0, + "content": "Evidential Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.245, + 0.179, + 0.751, + 0.2 + ], + "angle": 0, + "content": "Liangyu Xiang\\(^{1,2}\\) Junyu Gao\\(^{1,2\\dagger}\\) Changsheng Xu\\(^{1,2,3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.194, + 0.199, + 0.806, + 0.234 + ], + "angle": 0, + "content": "\\(^{1}\\)State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)" + }, + { + "type": "text", + "bbox": [ + 0.157, + 0.234, + 0.842, + 0.252 + ], + "angle": 0, + "content": "\\(^{2}\\)School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + }, + { + "type": "text", + "bbox": [ + 0.328, + 0.252, + 0.672, + 0.27 + ], + "angle": 0, + "content": "\\(^{3}\\)Peng Cheng Laboratory, ShenZhen, China" + }, + { + "type": "text", + "bbox": [ + 0.239, + 0.272, + 0.753, + 0.288 + ], + "angle": 0, + "content": "xiangliangyu2023@ia.ac.cn; {junyu.gao, csxu}@nlpr.ia.ac.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.322, + 0.328, + 0.339 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.354, + 0.486, + 0.718 + ], + "angle": 0, + "content": "Existing logit-based knowledge distillation methods typically employ singularly deterministic categorical distributions, which eliminates the inherent uncertainty in network predictions and thereby limiting the effective transfer of knowledge. To address this limitation, we introduce distribution-based probabilistic modeling as a more comprehensive representation of network knowledge. Specifically, we regard the categorical distribution as a random variable and leverage deep neural networks to predict its distribution, representing it as an evidential second-order distribution. Based on the second-oder modeling, we propose Evidential Knowledge Distillation (EKD) which distills both the expectation of the teacher distribution and the distribution itself into the student. The expectation captures the macroscopic characteristics of the distribution, while the distribution itself conveys microscopic information about the classification boundaries. Additionally, we theoretically show that EKD's distillation objective provides an upper bound on the student's expected risk when treating the teacher's predictions as ground truth. Extensive experiments on several standard benchmarks across various teacher-student network pairs highlight the effectiveness and superior performance of EKD. Our code is available at https://github.com/lyxiang-casia/EKD." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.746, + 0.224, + 0.761 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.771, + 0.485, + 0.879 + ], + "angle": 0, + "content": "As deep neural networks (DNNs) continue to advance, their performance has rapidly improved, yet their increasing network size and training costs pose challenges, especially in resource-constrained and real-time processing scenarios [62, 89]. Consequently, logit-based knowledge distillation (KD) [30, 88], which aims to compress the predictive capabilities of complex models into lightweight ones, has" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.32, + 0.909, + 0.46 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.619, + 0.461, + 0.802, + 0.473 + ], + "angle": 0, + "content": "(a) Vanilla Knowledge Distillation" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.473, + 0.909, + 0.614 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.61, + 0.616, + 0.812, + 0.629 + ], + "angle": 0, + "content": "(b) Evidential Knowledge Distillation" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.631, + 0.907, + 0.798 + ], + "angle": 0, + "content": "Figure 1. Illustrations of vanilla KD and our EKD. (a) Vanilla KD deterministically converts the logits into categorical distributions through the softmax function. Each categorical distribution corresponds to a point on the simplex, with each categorical probability representing the barycentric coordinates of that point. Vanilla KD is achieved by minimizing the differences between the teacher's and student's barycentric coordinates. (b) EKD treats the categorical distribution as a random variable following a Dirichlet distribution, parameterized by the logits. The heatmap depicts the probability density of the Dirichlet distribution across the simplex. Moreover, EKD includes distillation of both the expectation of the second-order distribution, and the second-order distribution itself." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.871 + ], + "angle": 0, + "content": "gained increasing popularity. KD transfers knowledge by guiding the predictions of a smaller network to align with those of a larger network during the training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Current knowledge distillation methods largely adhere to the vanilla paradigm [30] where only the categorical distri" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.237, + 0.901 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2814" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.513 + ], + "angle": 0, + "content": "butions are aligned. As illustrated in Figure 1a, the logits of a complex network (teacher network) and a lightweight network (student network) are firstly converted into categorical probabilities using a softmax function. Knowledge is then transferred through the process of minimizing the KL divergence between the two categorical distributions. Enhancements to this distillation process, such as decoupling the categorical distributions [88], smoothing the distribution with multiple levels [45], or adopting sample-wise distribution smoothness [69], have demonstrated notable improvements in the student network's predictive performance. Despite these advancements, all such methods are constrained by a fundamental limitation in their probabilistic modeling, which assumes that categorical probabilities are singularly deterministic [12, 14]. Specifically, vanilla KD implicitly assumes that for any given sample, its probability belonging to a particular class is deterministic and can be approximated by a deep neural network. In the case of a three-class classification problem, as shown in Figure 1a, this assumption is akin to identifying a single point on a 2D simplex that best represents the entire simplex. However, in practice, networks trained on finite data with limited capability exhibit inherent uncertainty in their predictions for specific samples [39, 47]. For example, networks with different initial weights may yield different predictions for the same test samples. Consequently, this singularly deterministic probabilistic modeling creates a bottleneck in the effective representation of diverse and uncertain knowledge." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.516, + 0.482, + 0.891 + ], + "angle": 0, + "content": "To overcome the limitations of singularly deterministic probabilistic modeling, we employ a distribution-based probabilistic modeling, treating the categorical distribution as a random variable governed by a second-order distribution. Following Subjective Logic [35] and Dempster-Shafer Theory of Evidence [81], the Dirichlet distribution, a conjugate prior of the categorical likelihood, instantiates this second-order distribution, facilitating the success of evidential deep learning across multiple domains [5, 16, 25]. However, recent studies employing distribution-based probabilistic modeling are largely confined to macro-level training within a single network, where the second-order distribution is integrated into a single categorical distribution during the network's training process [8, 11, 64]. The second-order distribution inherently provides micro-level information across all possible categorical distributions, offering advantages in cross-network information propagation. Therefore, the second-order distribution serves as a bridge to enable knowledge transfer at both the macro and micro levels, supported by PAC-Bayesain theory [19]. As illustrated in Figure 1b, the logits output by each network are converted into a set of Dirichlet parameters, forming respective second-order distributions. Essentially, the second-order distribution assigns varying weights to all possible first-order categorical distributions," + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.377 + ], + "angle": 0, + "content": "where the non-uniformity of these weights reflects the network's capacity for classification. Building on second-order predictions, we propose the Evidential Knowledge Distillation (EKD), integrating both macro and micro perspectives. At the macro level, we reduce each second-order distribution to their first-order categorical distributions via an expectation operation in order to ensure that the centroids of their Dirichlet distributions coincide. By representing the entire simplex with a single point, the student is encouraged to focus more on the overall density distribution of the teacher at this stage, which leads to the proportions of student's outputs across categories aligning with those of the teacher. At the micro level, we replace the first-order distribution with the second-order distribution to enable the transfer of more granular classification information. As a complement to proportional alignment, this step refines the magnitude in the student network's outputs. Ultimately, the student model benefits from both the teacher's overall characteristics and its detailed classification structure." + }, + { + "type": "text", + "bbox": [ + 0.536, + 0.379, + 0.882, + 0.393 + ], + "angle": 0, + "content": "To summarize, Our main contribution are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.394, + 0.905, + 0.468 + ], + "angle": 0, + "content": "- We reformulate network predictions in knowledge distillation through distribution-based probabilistic modeling. Building on this reformulation, we propose EKD, which distills the teacher's second-order distribution into the student model across both macro and micro levels." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.529 + ], + "angle": 0, + "content": "- With the help of PAC-Bayesian theory [19], we prove that the distillation objective in EKD serves as an upper bound on the expected risk of the student network when treating the teacher network's outputs as ground truth." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.53, + 0.905, + 0.604 + ], + "angle": 0, + "content": "- Extensive experiments are conducted on several standard benchmarks, encompassing a wide range of teacher and student models. Without bells and whistles, EKD achieves performance improvement of up to \\(3.61\\%\\) over KD across various teacher-student network combinations." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.394, + 0.905, + 0.604 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.619, + 0.654, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.518, + 0.645, + 0.905, + 0.885 + ], + "angle": 0, + "content": "Knowledge Distillation Knowledge distillation enhances a simpler student network using insights from a complex teacher network. Existing methods fall into three categories: logit-based [2, 30, 44, 45, 54, 85, 88], feature-based [33, 56, 57, 70, 73], and hybrid approaches [22, 31, 34, 58, 80, 82]. Unlike feature-based methods, which demand architecture-specific designs, logit-based methods offer wider applicability by relying solely on network outputs [4, 53, 84]. Introduced by Hinton and Geoffrey [30], logit-based distillation converts teacher and student logs into categorical distributions aligned via KL divergence. Subsequent efforts have refined this by adjusting the temperature parameter dynamically, based on training stages [45], teacher-student gaps [21], and output variance [69], or by decomposing distributions [88] and leveraging intermediate networks [31, 54]." + }, + { + "type": "text", + "bbox": [ + 0.537, + 0.887, + 0.905, + 0.9 + ], + "angle": 0, + "content": "The aforementioned traditional distillation methods" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.957 + ], + "angle": 0, + "content": "2815" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.468 + ], + "angle": 0, + "content": "share a common limitation in their reliance on point estimates for model parameters or categorical distributions, whereas Bayesian distillation [13, 37, 50, 74, 77] employs distributional estimation, which enables the student network to inherit the teacher's strengths in uncertainty estimation. Initial explorations [24, 37, 46, 51] adopt teacher ensembles to approximate the distribution of teacher model parameters and utilize the vanilla KD [30] for distillation. Subsequent extensions continue to advance by improving the student network architecture [20, 50, 61, 72, 76], optimization objectives [74], and sampling methods [13]. Additionally, some approaches introduce the utilization of quantized uncertainty during training to enable more fine-grained knowledge transfer [28, 66, 87]. While these methods introduce uncertainty into the model parameters and the resultant predictive distributions, they remain limited by their reliance on a finite number of ensembles to approximate the distributions. Consequently, we propose to adopt a continuous probabilistic modeling framework for both the teacher and the student, while leveraging more advanced distillation techniques. Unlike Bayesian distillation, which performs uncertainty-related distillation, we implicitly express the uncertainty of the predictive distribution through second-order probabilistic modeling and focus on enhancing the performance of the student network." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.469, + 0.484, + 0.742 + ], + "angle": 0, + "content": "Evidential Deep Learning Based on Subjective Logic (SL) [35] and Dempster-Shafer Theory (DST) [65], Sensoy et al. [64] initially propose a framework in which evidence collected by a deep neural network for each class is used to parameterize a second-order Dirichlet distribution for network predictions. This concept of evidence collection allows evidential networks to express \"I don't know\", showing advantages in uncertainty estimation. Since then, numerous studies have advanced and expanded upon Evidential Deep Learning (EDL) from both theoretical and application-oriented perspectives [17]. Theoretical investigations have primarily focused on refining the evidence collection process [11, 15, 18, 40, 55], delving into different training strategies [5, 26, 43] and extending evidential theory to regression networks [1, 49, 52]. Moreover, EDL has been widely applied in computer vision tasks, including action localization [5-7, 16], stereo matching [75], anomaly detection [68], and multi-view classification [25]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.183, + 0.77 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.484, + 0.901 + ], + "angle": 0, + "content": "In this section, we first present the fundamental framework of our Evidential Knowledge Distillation (EKD), which includes both macroscopic first-order and microscopic second-order distribution distillation. Theoretically, we utilize PAC-Bayesian theory to demonstrate that the optimization objective of EKD effectively represents an upper bound on the student's expected risk. At the end of this section, we illustrate the complementary effect of the two" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.818, + 0.107 + ], + "angle": 0, + "content": "distillation strategies with a concrete toy case." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.118, + 0.655, + 0.133 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.14, + 0.907, + 0.366 + ], + "angle": 0, + "content": "Evidential Deep Learning. In accordance with the formulation of evidential deep learning [8, 9, 64], for a classification task involving \\( K \\) classes, the categorical probability vector \\( p \\) is treated as a random variable governed by a conjugate prior Dirichlet distribution. To define this distribution's parameters, an evidential activation function \\( \\sigma \\) which we designate as the exponential function (exp), is employed to transform the network's output logits \\( z \\) into a non-negative evidence vector \\( e = \\sigma(z) \\). This evidence vector is subsequently combined with a prior weight \\( \\lambda \\) to yield the \\( K \\)-dimensional Dirichlet parameters \\( \\alpha \\) which fully specify a Dirichlet distribution \\( Dir(\\alpha) \\) [8]. In terms of optimization, the widely adopted cross-entropy loss is computed by integrating over the categorical probabilities, guided by second-order the Dirichlet distribution." + }, + { + "type": "equation", + "bbox": [ + 0.55, + 0.368, + 0.906, + 0.452 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {E D L - c e} = \\int \\left[ \\sum_ {i = 1} ^ {K} - y _ {i} \\log \\left(p _ {i}\\right) \\right] D i r (\\boldsymbol {p} | \\boldsymbol {\\alpha}) d \\boldsymbol {p} \\\\ = \\sum_ {i = 1} ^ {K} y _ {i} (\\psi (\\alpha_ {0}) - \\psi (\\alpha_ {i})) \\tag {1} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.455, + 0.906, + 0.562 + ], + "angle": 0, + "content": "Here, \\( y_{i} \\) denotes the ground truth, \\( \\alpha_0 = \\sum_{i=1}^{K} (\\alpha_i) \\) represents the Drichlet strength, and \\( \\psi \\) refers to the digamma function. To support the acquisition of evidence, we employ Eq. (1), rather than the conventional cross-entropy loss, across all network training stages, including both the training of the teacher networks and student networks during the knowledge distillation process." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.562, + 0.906, + 0.698 + ], + "angle": 0, + "content": "Knowledge Distillation. Existing logit-based knowledge distillation methods rely on categorical distribution predictions derived from the softmax function. Specifically, for a given input sample \\( x \\), the network produces logits \\( z \\), which are then transformed into a categorical distribution via the softmax operation, denoted as \\( p = \\mathrm{softmax}(z) \\). Knowledge transfer is accomplished by minimizing the Kullback-Leibler divergence(KL) divergence between the teacher's and student's predicted distributions, \\( p^T \\) and \\( p^S \\):" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.698, + 0.905, + 0.716 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {K D} = K L \\left(\\boldsymbol {p} ^ {T} \\| \\boldsymbol {p} ^ {S}\\right) \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.717, + 0.904, + 0.747 + ], + "angle": 0, + "content": "The teacher's predictions serve as soft labels, replacing the traditional one-hot labels, capturing inter-class similarities." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.758, + 0.809, + 0.774 + ], + "angle": 0, + "content": "3.2. Evidential Knowledge Distillation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Previous knowledge distillation methods have generally followed a unified paradigm, where the network's output logits are transformed into categorical probabilities via the softmax function, and the student network is guided to align its categorical distribution with that of the teacher network. While these approaches have yielded notable success, a fundamental limitation persists: they treat categorical probabilities as singularly deterministic values. This assumption" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2816" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "overlooks the uncertainty in network predictions, thus constraining the sharing of more granular information. As a more robust alternative, we predict the distribution of categorical distributions using neural networks, specifically modeling it as a second-order Dirichlet distribution. More importantly, we align both the first-order distribution, obtained by integrating the second-order distribution, and the second-order distribution itself." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.212, + 0.482, + 0.348 + ], + "angle": 0, + "content": "Firstly, we aim to extract the overall characteristics of the second-order distribution to capture the general distribution trend of probability density across the simplex, as well as to achieve alignment between networks in terms of final predicted classes. Specifically, given the second-order distributions predicted by the teacher and student, \\(Dir(\\pmb{\\alpha}^T)\\) and \\(Dir(\\pmb{\\alpha}^S)\\), we compute their expectations where all possible categorical distributions are summed under the weights specified by the second-order distribution:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.348, + 0.483, + 0.368 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} ^ {T} = \\mathbb {E} _ {\\boldsymbol {p} ^ {T} \\sim D i r (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\hat {\\boldsymbol {p}} ^ {S} = \\mathbb {E} _ {D i r (\\boldsymbol {p} ^ {S} \\sim \\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ] \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.483, + 0.442 + ], + "angle": 0, + "content": "The expectation above, serving as the final predictive outcome for both teacher and student, is equivalent to the centroid of the second-order distribution, enabling the transfer of comprehensive probability density information via first-order distillation at a macro level." + }, + { + "type": "equation", + "bbox": [ + 0.14, + 0.443, + 0.483, + 0.482 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {1 s t} = K L (\\hat {\\boldsymbol {p}} ^ {T} | | \\hat {\\boldsymbol {p}} ^ {S}) = \\sum_ {i = 1} ^ {K} \\frac {\\alpha_ {i} ^ {T}}{\\alpha_ {0} ^ {T}} l o g \\left(\\frac {\\alpha_ {i} ^ {T} \\alpha_ {0} ^ {S}}{\\alpha_ {0} ^ {T} \\alpha_ {i} ^ {S}}\\right) \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.482, + 0.483, + 0.602 + ], + "angle": 0, + "content": "Formally, optimizing Eq. (4) involves adjusting the proportion of each class's Dirichlet parameter relative to the total, which significantly enhances the alignment between the student's and teacher's predicted classes. In terms of formulation, Eq. (4) is very similar to vanilla KD, yet it is emphasized that vanilla KD can be regarded as a special case of first-order distillation when \\(\\sigma = \\exp(\\cdot)\\) and \\(\\lambda = 0\\). Proof is provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.603, + 0.483, + 0.783 + ], + "angle": 0, + "content": "The first-order distillation objective above only provides macroscopic guidance for the student's second-order distribution, failing to accurately and comprehensively represent the weighting of each categorical distribution. In other words, given a fixed teacher distribution, the student Dirichlet distribution that minimizes Eq. (4) is not unique. Therefore, we propose a microscopic second-order distillation as a complement to first-order distillation, aiming to provide fine-grained teacher supervision signals across the entire simplex for the student distribution. Specifically, we achieve this by applying the KL divergence directly between the two Dirichlet distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.784, + 0.483, + 0.892 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {2 n d} = K L (D i r (\\boldsymbol {p} ^ {T} | \\boldsymbol {\\alpha} ^ {T}) | | D i r (\\boldsymbol {p} ^ {S} | \\boldsymbol {\\alpha} ^ {S})) \\\\ = l o g \\frac {\\Gamma (\\alpha_ {0} ^ {T})}{\\Gamma (\\alpha_ {0} ^ {S})} - \\sum_ {i = 1} ^ {K} l o g \\frac {\\Gamma (\\alpha_ {i} ^ {T})}{\\Gamma (\\alpha_ {i} ^ {S})} \\\\ + \\sum_ {i = 1} ^ {K} \\left(\\alpha_ {i} ^ {T} - \\alpha_ {i} ^ {S}\\right) \\left(\\psi \\left(\\alpha_ {i} ^ {T}\\right) - \\psi \\left(\\alpha_ {0} ^ {T}\\right)\\right) \\tag {5} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.379 + ], + "angle": 0, + "content": "Here, \\(\\Gamma(\\cdot)\\) denotes the Gamma function. The detailed derivation for Eq. (5) is provided in Supplementary Material. Since Eq. (5) considers the whole \\((K - 1)\\) dimensional simplex, overflow may occur [50]; thus, we apply the \\(\\log(1 + x)\\) function to smooth the network outputs. While the above Eq. (4) aims to optimize the proportional relationships between classes, Eq. (5) here places greater emphasis on aligning the magnitudes within each class. Here, the alignment within the class means that the gradients of a loss function with respect to the output for a particular class do not conflict with the gradients of the output for other classes. This meticulous numerical alignment supplies each point on the \\((K - 1)\\)-dimensional simplex of the student's second-order distribution with the corresponding teacher label, ensuring a precise mapping of knowledge across the entire distribution space. By doing so, it strives to faithfully replicate the teacher's second-order predictive nuances at a granular, micro-level, capturing the fine-grained patterns and uncertainties inherent in the teacher model's outputs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.379, + 0.905, + 0.438 + ], + "angle": 0, + "content": "By combining these two complementary optimization objectives, one optimizing proportional relationships at the macro level and the other aligning numerical values at the micro level, we arrive at the final EKD distillation:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.441, + 0.905, + 0.457 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {E K D} = \\mathcal {L} _ {1 s t} + \\gamma \\mathcal {L} _ {2 n d}, \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.487 + ], + "angle": 0, + "content": "where \\(\\gamma\\) represents the coefficient that adjusts the balance between the first-order and second-order distillation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.505, + 0.707, + 0.521 + ], + "angle": 0, + "content": "3.3. Theoretical Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.527, + 0.905, + 0.648 + ], + "angle": 0, + "content": "In this section, we provide theoretical support for Eq. (6) using PAC-Bayesian Theory [19]. While PAC-Bayesian theory has been explored in the context of the Dirichlet distribution [26], its application remains confined to single-network optimization, neglecting the relationship among multiple networks. Therefore, we aim to estimate the alignment of different networks across the entire data distribution based on their alignment on training samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.649, + 0.906, + 0.799 + ], + "angle": 0, + "content": "Let \\( x \\) denote a certain sample and \\( y \\) denote its corresponding label. Under the PAC setting, samples are drawn from a fixed but unknown distribution \\( \\mathcal{D} \\). A classifier \\( h \\) maps \\( x \\) to \\( y \\). Moreover, we define a classifier \\( h \\) correspond to a point on the \\( (K - 1) \\)-dimensional simplex, i.e., a categorical distribution \\( p \\). Essentially, the Dirichlet distribution made by each network serves as the posterior distribution of a set of classifiers. By weighting the set of classifiers with its posterior distribution, we obtain the final majority vote classifier (or Bayesain classifier):" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.799, + 0.905, + 0.818 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {B} ^ {T} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\tag {7}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.82, + 0.905, + 0.839 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {B} ^ {S} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ], \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.84, + 0.905, + 0.871 + ], + "angle": 0, + "content": "where \\(\\pmb{B}^{T}\\) and \\(\\pmb{B}^{S}\\) denote the final classification results of the teacher and student classifiers, respectively." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.901 + ], + "angle": 0, + "content": "In knowledge distillation, the student treats the teacher's predictions as soft labels. Below, we define the student's" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2817" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "misclassification risk using cross-entropy, which is equivalent to KL divergence from an optimization perspective." + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.121, + 0.483, + 0.141 + ], + "angle": 0, + "content": "\\[\nR ^ {S} = \\mathbb {E} _ {x \\sim \\mathcal {D}} \\left[ - \\boldsymbol {B} ^ {T} \\log \\boldsymbol {B} ^ {S} \\right], \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.144, + 0.483, + 0.184 + ], + "angle": 0, + "content": "\\[\n\\hat {R} ^ {S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ - \\boldsymbol {B} _ {i} ^ {T} \\log \\boldsymbol {B} _ {i} ^ {S} \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.484, + 0.232 + ], + "angle": 0, + "content": "where \\( R^S \\) represents the expected risk of the student over the entire data distribution \\( \\mathcal{D} \\) and \\( \\hat{R}^S \\) denotes the empirical risk on the training set \\( \\mathcal{D}_t \\) containing \\( N \\) training samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.239, + 0.484, + 0.287 + ], + "angle": 0, + "content": "Theorem 1. For any \\(\\delta \\in (0,1]\\) and \\(\\gamma \\in \\mathbb{R}^{+}\\), \\(R^S\\) is bounded above by the sum of \\(\\hat{R}^S\\), \\(\\sum_{1}^{N}KL(Dir(\\pmb{\\alpha}_i^T))||Dir(\\pmb{\\alpha}_i^S))\\), and a constant related to \\(\\delta\\), denoted as \\(C(\\delta)\\):" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.288, + 0.482, + 0.363 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} P _ {\\mathcal {D} _ {t} \\sim \\mathcal {D} ^ {N}} (R ^ {S} \\leq \\hat {R} ^ {S} + \\frac {\\gamma}{N} \\sum_ {i = 1} ^ {N} K L (D i r (\\pmb {\\alpha} _ {i} ^ {T}) | | D i r (\\pmb {\\alpha} _ {i} ^ {S})) \\\\ + C (\\delta)) \\geq 1 - \\delta . \\tag {11} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.372, + 0.484, + 0.63 + ], + "angle": 0, + "content": "Detailed proof and definition of \\( C(\\delta) \\) can be found in Supplementary Material. Theorem 1 indicates that the upper bound on the student's excepted risk over the unknown data distribution \\( \\mathcal{D} \\) consists of two components: the alignment of the first-order distribution with the true labels (teacher labels) and the similarity between the predicted second-order distribution and the true second-order distribution (teacher distribution). Clearly, this upper bound indicates that the empirical risk optimization represented by Eq. (2) does not guarantee a reduction in the student's expected risk, suggesting that the transfer of the teacher's generalization performance is both inefficient and incomplete. In contrast, EKD directly adopts this upper bound as optimization objectives, refining the formulation of empirical risk while incorporating second-order distribution alignment between the student and teacher, thereby striving to minimize the student's expected risk." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.638, + 0.199, + 0.654 + ], + "angle": 0, + "content": "3.4. Toy Case" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Figure 2 demonstrates the complementary roles of first-order and second-order distillation through a three-class example. Initially, the student Dirichlet parameters differ significantly from those of the teacher, both in relative proportion and in absolute values. The left side of Figure 2 illustrates the essence of first-order distillation, which minimizes the distance between the expectations of two Dirichlet distributions over the \\((K - 1)\\) dimensional simplex measured by KL divergence. However, the expectation of the Dirichlet distribution depends solely on the proportion of each class's Dirichlet parameter relative to the overall sum, which can lead to discrepancies in the specific values of the student's output compared to the teacher's. Although this discrepancy does not affect the final class prediction, it causes the student's second-order distribution to differ significantly from the teacher's. To address this, we use" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.906, + 0.374 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.383, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Figure 2. A toy case illustrating the complementary roles of the two distillation losses in EKD. Given the Dirichlet parameters output by teacher and student networks, the left panel shows the results when only first-order distillation is used to optimize the proportion of each category. The right panel presents the outcome when only second-order distillation is applied for intra-class numerical alignment. The center panel presents the results of combining both objectives, which yields the optimal performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.525, + 0.907, + 0.678 + ], + "angle": 0, + "content": "second-order distillation as a complement to first-order distillation, aiming to align the Dirichlet parameter values. In contrast to first-order distillation, second-order distribution alignment encourages consistency of parameters within the same class across different networks, disregarding the relative magnitudes among different classes. As a result, we combine these complementary distillation objectives to achieve a comprehensive alignment of the second-order distribution. The quantitative effects of the two types of distillation will be discussed in detail in Section 4.2" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.695, + 0.646, + 0.712 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Datasets. Our experiments are primarily conducted on two standard datasets including CIFAR100 [38] and ImageNet [60]. The CIFAR-100 dataset [38] consists of 60k 32x32 color images divided into 100 classes, with each class containing 600 images. It is split into a training set of 50k images and a validation set of 10k images. ImageNet [60] is a large-scale visual database comprising 1,000 categories that cover a diverse range of objects and scenes. It includes approximately 1.28 million labeled images for training and an additional 50,000 images for validation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Implementation Details. We adopt the experimental settings established in prior studies [10, 69, 88]. For the" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2818" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 1. Results on the CIFAR-100 validation set. The teacher and student networks share the same architecture. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. \\(\\Delta\\) denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.147, + 0.92, + 0.468 + ], + "angle": 0, + "content": "
Experiment Group1234567
TeacherArchitectureResNet32×4VGG13WRN-40-2WRN-40-2ResNet56ResNet110ResNet110
Softmax79.4274.6475.6175.6172.3474.3174.31
EDL79.5374.9675.4175.4173.0974.6774.67
StudentArchitectureResNet8×4VGG8WRN-40-1WRN-16-2ResNet20ResNet32ResNet20
Softmax72.5070.3671.9873.2669.0671.1469.06
EDL72.7770.6771.8473.3669.2270.969.22
FeatureRKD[CVPR'19] [56]71.9071.4872.2273.3569.6171.8269.25
CRD[ICLR'20] [70]75.5173.9474.1475.4871.1673.4871.46
ReviewKD[CVPR'21] [10]75.6374.8475.0976.1271.8973.8971.34
SimKD[CVPR'22] [3]78.0874.8974.5375.5371.0573.9271.06
CAT-KD[CVPR'23] [23]76.9174.6574.8275.6071.6273.6271.37
TopKD[ICML'24] [36]75.4074.0174.4375.7571.5873.7771.47
LogitKD[NeurIPS'14] [30]73.3372.9873.5474.9270.6673.0870.67
Δ0.832.621.561.661.61.941.61
DKD[CVPR'22] [88]76.3274.6874.8176.2471.9774.1171.06
Δ3.824.322.832.982.912.972.0
CTKD[AAAI'23] [45]73.3973.5273.9375.4571.1973.5270.99
Δ0.893.161.952.192.132.381.93
Logit_Stand[CVPR'24] [69]76.6274.3674.3776.1171.4374.1771.48
Δ4.124.002.392.852.373.032.42
SDD[CVPR'24] [48]75.92*73.90*74.26*75.98*70.72*73.90*71.43*
Δ3.42*3.54*2.28*2.72*1.66*2.76*2.37*
TeKAP[ICLR'25] [31]74.7973.8*73.8075.2171.7173.50*71.00*
Δ2.293.44*1.821.952.652.36*1.94*
EKD(Ours)77.2174.7174.4376.1571.4873.6871.51
Δ4.444.042.592.792.262.782.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.49, + 0.483, + 0.531 + ], + "angle": 0, + "content": "Table 3. The top-1 and top-5 accuracy (\\%) on the ImageNet validation set [60]. The best and second best results are emphasized in bold and underlined. “*” indicates our reproduced results." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.533, + 0.477, + 0.751 + ], + "angle": 0, + "content": "
TeacherArchitectureResNet34ResNet50
Accuracytop-1top-5top-1top-5
Softmax73.3191.4276.1692.86
EDL73.8391.6976.1092.98
StudentArchitectureResNet18MNV1
Accuracytop-1top-5top-1top-5
Softmax69.7589.0768.8788.76
EDL70.2889.5869.9289.46
FeatureCRD [70]71.1790.1371.3790.41
ReviewKD [10]71.6190.5172.5691.00
SimKD [3]71.5990.4872.2590.86
CAT-KD [23]71.2690.4572.2491.13
LogitKD [30]71.0390.0570.5089.80
DKD [88]71.7090.4172.0591.05
CTKD [45]71.3890.2771.1690.11
Logit_Stand [69]71.4290.2972.1890.80
SDD [48]71.4490.0572.15*90.79*
EKD(Ours)71.7390.6972.5491.20
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "experiments on CIFAR-100, we use the SGD optimizer with 240 epochs. The learning rate is set to 0.01 for MobileNets [32, 63] and ShuffleNets [86] and 0.05 for other networks, including ResNets [27], WRNs [83] and VGGs [67], with a decay factor of 0.1 applied at epochs 150, 180, and 210. Regarding hyperparameter configuration, we set \\(\\gamma\\) in Eq. (6) to 1 by default, requiring no additional hyperparameter tuning. All results represent the mean of five" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.492, + 0.907, + 0.522 + ], + "angle": 0, + "content": "independent runs. Additional experimental details are provided in the Supplementary Material." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.522, + 0.907, + 0.628 + ], + "angle": 0, + "content": "Baselines. To benchmark EKD, we incorporate five state-of-the-art logit-based distillation methods: DKD (CVPR'22) [88], CTKD (AAAI'23) [45], Logit_Stand (CVPR'24) [69], SDD (CVPR'24) [48] and TeKAP (ICLR'25) [31]. Additionally, a selection of feature-based distillation methods [3, 10, 23, 56, 59, 70] is included as references to facilitate a broader contextual analysis." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.655, + 0.653 + ], + "angle": 0, + "content": "4.1. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Performance comparison on CIFAR100. We conduct experiments comparing EKD with other distillation methods under various teacher-student pair settings. Table 1 and Table 2 present the accuracy results for settings where the teacher and student have identical and distinct architectures. As presented in Table 1, EKD demonstrates favourable performance improvements over KD, ranging from \\(0.66\\%\\) to \\(3.61\\%\\). In particular, EKD demonstrates its strongest performance in Groups 1, 2, and 7, while achieving comparable results to DKD [88] in Groups 3-6. The underlying reason for this phenomenon lies in the performance gap between the teacher and student. In groups 1, 2, and 7, the average performance gap between the teacher and student is \\(5.51\\%\\), whereas it is \\(3.31\\%\\) in groups 3-6. Furthermore, we observe that pairs with a larger performance gap tend to have a higher distillation loss, indicating a greater opti" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2819" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 2. Results on the CIFAR-100 validation set. The teacher and student networks have different architectures. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. \\(\\Delta\\) denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.147, + 0.929, + 0.468 + ], + "angle": 0, + "content": "
Experiment Group1234567
TeacherArchitectureResNet32×4ResNet32×4ResNet32×4WRN-40-2WRN-40-2VGG13ResNet50
Softmax79.4279.4279.4275.6175.6174.6479.34
EDL79.5379.5379.5375.4175.4175.1179.69
StudentArchitectureSHN-V2WRN-16-2WRN-40-2ResNet8×4MN-V2MN-V2MN-V2
Softmax71.8273.2675.6172.5064.6064.6064.60
EDL72.7373.3675.4172.7765.3965.3965.39
FeatureRKD[CVPR'19] [56]73.2174.8677.8275.2669.2764.5264.43
CRD[ICLR'20] [70]75.6575.6578.1575.2470.2869.7369.11
ReviewKD[CVPR'21] [10]77.7876.1178.9674.3471.2870.3769.89
SimKD[CVPR'22] [3]78.3977.1779.2975.2970.1069.4469.97
CAT-KD[CVPR'23] [23]78.4176.9778.5975.3870.2469.1371.36
TopKD[ICML'24] [36]76.3374.34*76.18*75.83*69.54*69.32*70.11*
LogitKD[NeurIPS'14] [30]74.4574.9077.7073.9768.3667.3767.35
Δ2.631.642.091.473.762.772.75
DKD[CVPR'22] [88]77.0775.7078.4675.5669.2869.7170.35
Δ5.252.442.853.064.685.115.75
CTKD[AAAI'23] [45]75.3774.5777.6674.6168.3468.5068.67
Δ3.551.312.052.113.743.904.07
Logit_Stand[CVPR'24] [69]75.5675.2677.9277.1169.2368.6169.02
Δ3.742.002.314.614.634.014.42
SDD[CVPR'24] [48]76.88*75.92*78.54*74.83*70.14*69.06*69.31*
Δ5.06*2.66*2.93*2.33*5.54*4.46*4.71*
TeKAP[ICLR'25] [31]75.4375.26*76.93*74.99*68.65*68.27*68.75*
Δ3.612.00*1.32*2.49*4.05*3.67*4.15*
EKD(Ours)77.6576.5178.9176.7569.2069.9470.38
Δ4.923.153.503.983.814.554.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.56 + ], + "angle": 0, + "content": "Table 4. Ablation study for each components in EKD. The \"MSE\" refers to aligning the magnitude of model outputs with mean square error. The \"Accuracy\" denotes the Top-1 accuracy (\\%) on the CIFAR100 validation set. ResNet32×4 and ResNet8×4 are adopted as teacher and student." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.562, + 0.445, + 0.692 + ], + "angle": 0, + "content": "
IndexDistillation ObjectiveAccuracy
\\( \\mathcal{L}_{KD} \\)\\( \\mathcal{L}_{1st} \\)\\( \\mathcal{L}_{2nd} \\)MSE
1✓74.45
2✓73.80
3✓76.42
4✓76.10
5✓✓77.21
6✓✓76.83
7✓✓76.34
8✓✓76.13
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.901 + ], + "angle": 0, + "content": "mization space during distillation. Conversely, for teacher-student pairs with closer performance, their distillation loss is generally smaller, leading EKD to encounter local minima during optimization. For the distillation of distinct networks shown in Table 2, EKD achieves the highest performance in half of the experimental groups. When using MobileNet [63] as the student network, the student trained solely with EDL achieves a \\(0.79\\%\\) higher performance than the one trained with softmax alone. However, the result for EKD is only comparable to that of other methods. We attribute this phenomenon to the inherent performance ceiling of the student network due to its limited capacity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.492, + 0.907, + 0.613 + ], + "angle": 0, + "content": "Performance comparison on ImageNet. Table 3 presents the performance of EKD on the large-scale ImageNet dataset. EKD markedly outperforms other logit-based methods in the ResNet50-MNV1 experiment and achieves results comparable to top methods in the ResNet34-ResNet18 experiment. The modest gains in the latter can also be attributed to the narrow performance gap between the teacher and student." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.623, + 0.68, + 0.637 + ], + "angle": 0, + "content": "4.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.884 + ], + "angle": 0, + "content": "To elucidate the superiority of the two alignment strategies implemented in EKD and their alignment with theoretical underpinnings, the ablation study presented in Table 4 evaluates various distillation objectives on student model performance. In addressing alignment with respect to the proportional distribution across output classes, we investigate two methodologies: the conventional vanilla KD approach presented in Eq. (2), which aligns softmax-based categorical distributions, and \\(\\mathcal{L}_{1st}\\) in Eq. (4). Comparative analyses between Index 1 and Index 2, Index 5 and Index 7, and Index 6 and Index 8 consistently demonstrate that \\(\\mathcal{L}_{1st}\\) surpasses vanilla KD. This finding underscores that EKD's first-order alignment serves as a robust generalization of vanilla KD, effectively incorporating the alignment of categorical distributions by leveraging the centroids of second-order Dirichlet distributions." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "In addition to the second-order distillation, \\(\\mathcal{L}_{2nd}\\), shown" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2820" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.09, + 0.483, + 0.117 + ], + "angle": 0, + "content": "Table 5. The Top-1 accuracy \\((\\%)\\) of combining EKD with feature distillation-based methods on CIFAR-100." + }, + { + "type": "table", + "bbox": [ + 0.107, + 0.12, + 0.472, + 0.246 + ], + "angle": 0, + "content": "
TeacherArchitectureVGG13ResNet32x4
Accuracy74.9679.53
StudentArchitectureVGG8SHN-V2
Accuracy70.6772.73
LogitEKD74.7177.65
FeatureRKD [56]71.4873.21
CAT-KD [23]74.6578.41
Logit+FeatureEKD+RKD [56]74.99↑78.08↑
EKD+CAT-KD [23]74.91↑78.46↑
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.276, + 0.484, + 0.429 + ], + "angle": 0, + "content": "in Eq. (5), another intuitive approach to aligning model output magnitudes is through mean squared error (MSE). Consequently, we compare \\(\\mathcal{L}_{2nd}\\) and MSE in Table 4. Comparisons across Index 3 and Index 4, Index 5 and Index 6, and Index 7 and Index 8 reveal that \\(\\mathcal{L}_{2nd}\\) consistently outperforms MSE. These empirical results corroborate the theoretical conclusions delineated in Section 3.3. It is suggested that \\(\\mathcal{L}_{2nd}\\), being part of the upper bound on the student's expected risk, directly optimizes the student's generalization performance, whereas the MSE objective does so indirectly." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.441, + 0.265, + 0.456 + ], + "angle": 0, + "content": "4.3. Further Remarks" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.463, + 0.484, + 0.643 + ], + "angle": 0, + "content": "Visualizations. Figure 3 presents the t-SNE visualizations of the laerned feature representations and, in parallel, the differences in logits between the teacher and student networks, where the teacher is implemented as ResNet \\(32 \\times 4\\) and the student as ResNet \\(8 \\times 4\\). As shown in the first row, EKD achieves a clear improvement over vanilla KD baseline with respect to class separability, leading to more distinctly clustered feature distributions. Furthermore, the logits produced under EKD exhibit a noticeably stronger alignment with those of the teacher network, indicating a closer match in predictive behavior compared to the results obtained with vanilla KD." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.645, + 0.484, + 0.78 + ], + "angle": 0, + "content": "Compatibility with feature-based methods. Table 5 presents the results of integrating EKD with several representative feature-based distillation methods [23, 56]. For feature-based approaches, the introduction of EKD enhances their performance, demonstrating strong compatibility, particularly with methods that originally exhibit lower performance [56]. Conversely, these methods can further compensate for the limitation of the EKD approach in expressing knowledge through logits." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Distilling ViT. Transformer-based models are gaining increasing popularity, and logit-based distillation methods can be easily applied to these models. Therefore, we test the distillation effectiveness of EKD on the ViT model, as presented in Table 6. Following the configuration in [41, 42], we use ResNet56 as the teacher model and four ViT variants, DeiT-Ti [71], PiT-Ti [29], PVT-Ti [78], and PVTv2 [79], as the student models. EKD achieves an aver-" + }, + { + "type": "image", + "bbox": [ + 0.536, + 0.097, + 0.689, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.567, + 0.215, + 0.652, + 0.228 + ], + "angle": 0, + "content": "(a) Vanilla KD" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.097, + 0.872, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.767, + 0.215, + 0.819, + 0.228 + ], + "angle": 0, + "content": "(b) EKD" + }, + { + "type": "image", + "bbox": [ + 0.535, + 0.231, + 0.7, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.562, + 0.339, + 0.646, + 0.351 + ], + "angle": 0, + "content": "(c) Vanilla KD" + }, + { + "type": "image", + "bbox": [ + 0.719, + 0.231, + 0.884, + 0.337 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.772, + 0.339, + 0.824, + 0.351 + ], + "angle": 0, + "content": "(d) EKD" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.355, + 0.908, + 0.41 + ], + "angle": 0, + "content": "Figure 3. Visualization of features and logits output. The first row shows the t-SNE of features learned by vanilla KD and EKD. The second row illustrates the difference of correlation matrices of logits between the student and teacher." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.427, + 0.906, + 0.454 + ], + "angle": 0, + "content": "Table 6. Top-1 accuracy of various ViT student models on CIFAR100. Teacher model is ResNet56." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.457, + 0.911, + 0.547 + ], + "angle": 0, + "content": "
ArchitectureSizeEDLSoftmaxKD [30]AutoKD [42]EKD(Ours)
DeiT-Ti [71]5M64.7765.0873.2578.5578.64
PiT-Ti [29]5M73.4873.5875.4778.7679.33
PVT-Ti [78]13M69.0269.2273.6078.4378.74
PVTv2 [79]3M76.6877.4478.8179.3779.80
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.907, + 0.604 + ], + "angle": 0, + "content": "age performance improvement of \\(3.75\\%\\) over KD [30] and an average improvement of \\(0.25\\%\\) over AutoKD [42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.619, + 0.634, + 0.635 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This paper reinterprets vanilla knowledge distillation (KD) through a distribution-based probabilistic framework, employing evidential second-order distributions to effectively capture predictive uncertainty and provide a comprehensive knowledge representation. This pioneering methodology underpins Evidential Knowledge Distillation (EKD), facilitating knowledge transfer across macro and micro levels. At the macro level, aligning the expectation of the second-order distribution, which reflects its global characteristics, enhances optimization of inter-class proportions in the student's output. At the micro level, aligning the second-order distribution aligns the magnitudes of the student's output. These complementary mechanisms ensure a robust distillation process. Through PAC-Bayesian theory, EKD optimizes the upper bound of the student's expected risk. Extensive experiments underscore EKD's pronounced advantage over prevailing logit-based distillation techniques." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "2821" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported in part by the National Key Research and Development Plan of China under Grant 2023YFC3310700, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62472422, and U2333215, and in part by Beijing Natural Science Foundation under Grant 4242051." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.22, + 0.188, + 0.236 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.286 + ], + "angle": 0, + "content": "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. NIPS, 33:14927-14937, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.289, + 0.484, + 0.33 + ], + "angle": 0, + "content": "[2] Defang Chen, Jian-Ping Mei, Can Wang, Yan Feng, and Chun Chen. Online knowledge distillation with diverse peers. In AAAI, pages 3430-3437, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.332, + 0.483, + 0.386 + ], + "angle": 0, + "content": "[3] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, pages 11933-11942, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.389, + 0.482, + 0.43 + ], + "angle": 0, + "content": "[4] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. NIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.432, + 0.482, + 0.486 + ], + "angle": 0, + "content": "[5] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV, pages 192-208. Springer, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.489, + 0.482, + 0.558 + ], + "angle": 0, + "content": "[6] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Cascade evidential learning for open-world weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14741-14750, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.56, + 0.482, + 0.627 + ], + "angle": 0, + "content": "[7] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Uncertainty-aware dual-evidential learning for weakly-supervised temporal action localization. IEEE transactions on pattern analysis and machine intelligence, 45(12):15896-15911, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.631, + 0.482, + 0.685 + ], + "angle": 0, + "content": "[8] Mengyuan Chen, Junyu Gao, and Changsheng Xu. R-edl: Relaxing nonessential settings of evidential deep learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.482, + 0.729 + ], + "angle": 0, + "content": "[9] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Revisiting essential and nonessential settings of evidential deep learning. arXiv preprint arXiv:2410.00393, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.731, + 0.482, + 0.771 + ], + "angle": 0, + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, pages 5008-5017, 2021. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.828 + ], + "angle": 0, + "content": "[11] Danruo Deng, Guangyong Chen, Yang Yu, Furui Liu, and Pheng-Ann Heng. Uncertainty estimation by fisher information-based evidential deep learning. In ICML, pages 7596–7616. PMLR, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.831, + 0.482, + 0.872 + ], + "angle": 0, + "content": "[12] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[13] Luyang Fang, Yongkai Chen, Wenxuan Zhong, and Ping Ma. Bayesian knowledge distillation: A bayesian perspec" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.121 + ], + "angle": 0, + "content": "tive of distillation with uncertainty quantification. In Forty-first ICML, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.868, + 0.137 + ], + "angle": 0, + "content": "[14] Yarin Gal et al. Uncertainty in deep learning. 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.14, + 0.907, + 0.207 + ], + "angle": 0, + "content": "[15] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18827-18836, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.21, + 0.906, + 0.25 + ], + "angle": 0, + "content": "[16] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Vectorized evidential learning for weakly-supervised temporal action localization. PAMI, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.306 + ], + "angle": 0, + "content": "[17] Junyu Gao, Mengyuan Chen, Liangyu Xiang, and Changsheng Xu. A comprehensive survey on evidential deep learning and its applications. arXiv preprint arXiv:2409.04720, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.311, + 0.905, + 0.366 + ], + "angle": 0, + "content": "[18] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Learning probabilistic presence-absence evidence for weakly-supervised audio-visual event perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.369, + 0.905, + 0.409 + ], + "angle": 0, + "content": "[19] Pascal Germain, Alexandre Lacasse, François Laviolette, and Mario Marchand. Pac-bayesian learning of linear classifiers. In ICML, pages 353-360, 2009. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[20] Hongji Guo, Hanjing Wang, and Qiang Ji. Bayesian evidential deep learning for online action detection. In ECCV, pages 283-301. Springer, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[21] Jia Guo. Reducing the teacher-student gap via adaptive temperatures. 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.485, + 0.905, + 0.54 + ], + "angle": 0, + "content": "[22] Yong Guo, Shulian Zhang, Haolin Pan, Jing Liu, Yulun Zhang, and Jian Chen. Gap preserving distillation by building bidirectional mappings with a dynamic teacher. arXiv preprint arXiv:2410.04140, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.542, + 0.905, + 0.583 + ], + "angle": 0, + "content": "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, pages 11868-11877, 2023. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.586, + 0.905, + 0.627 + ], + "angle": 0, + "content": "[24] Corina Gurau, Alex Bewley, and Ingmar Posner. Dropout distillation for efficiently estimating model confidence. arXiv preprint arXiv:1809.10562, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.629, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[25] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.659, + 0.905, + 0.699 + ], + "angle": 0, + "content": "[26] Manuel Haussmann, Sebastian Gerwinn, and Melih Kandemir. Bayesian evidential deep learning with pac regularization. arXiv preprint arXiv:1906.00816, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.702, + 0.905, + 0.743 + ], + "angle": 0, + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[28] Mina Hemmatian, Ali Shahzadi, and Saeed Mozaffari. Uncertainty-based knowledge distillation for bayesian deep neural network compression. International Journal of Approximate Reasoning, 175:109301, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.802, + 0.905, + 0.856 + ], + "angle": 0, + "content": "[29] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In CVPR, pages 11936-11945, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[30] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 3, 6, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "2822" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[31] Md Imtiaz Hossain, Sharmen Akhter, Choong Seon Hong, and Eui-Nam Huh. Single teacher, multiple perspectives: Teacher knowledge augmentation for enhanced knowledge distillation. In ICLR, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[32] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[33] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. NIPS, 35:33716-33727, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.29 + ], + "angle": 0, + "content": "[34] Yaomin Huang, Zaomin Yan, Chaomin Shen, Faming Fang, and Guixu Zhang. Harmonizing knowledge transfer in neural network with unified distillation. In ECCV, pages 58-74. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.441, + 0.306 + ], + "angle": 0, + "content": "[35] Audun Jøsang. Subjective logic. Springer, 2016. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[36] Jungeun Kim, Junwon You, Dongjin Lee, Ha Young Kim, and Jae-Hun Jung. Do topological characteristics help in knowledge distillation? In ICML, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[37] Anoop Korattikara Balan, Vivek Rathod, Kevin P Murphy, and Max Welling. Bayesian dark knowledge. NIPS, 28, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.392, + 0.482, + 0.42 + ], + "angle": 0, + "content": "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.462 + ], + "angle": 0, + "content": "[39] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. NIPS, 30, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.464, + 0.482, + 0.532 + ], + "angle": 0, + "content": "[40] Changbin Li, Kangshuo Li, Yuzhe Ou, Lance M Kaplan, Audun Jøsang, Jin-Hee Cho, Dong Hyun Jeong, and Feng Chen. Hyper evidential deep learning to quantify composite classification uncertainty. arXiv preprint arXiv:2404.10980, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[41] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, pages 110-127. Springer, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.59, + 0.482, + 0.632 + ], + "angle": 0, + "content": "[42] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In CVPR, pages 17413-17424, 2023. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.675 + ], + "angle": 0, + "content": "[43] Xue Li, Wei Shen, and Denis Charles. Tedl: A two-stage evidential deep learning method for classification uncertainty quantification. arXiv preprint arXiv:2209.05522, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[44] Zheng Li, Ying Huang, Defang Chen, Tianren Luo, Ning Cai, and Zhigeng Pan. Online knowledge distillation via multi-branch diversity enhancement. In Proceedings of the Asian conference on computer vision, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[45] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.858 + ], + "angle": 0, + "content": "[46] Jakob Lindqvist, Amanda Olmin, Fredrik Lindsten, and Lennart Svensson. A general framework for ensemble distribution distillation. In 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[47] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[48] Shicai Wei Chunbo Luo Yang Luo. Scale decoupled distillation. arXiv preprint arXiv:2403.13512, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[49] Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions. NIPS, 34:6881-6893, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[50] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[51] Zelda E Mariet, Rodolphe Jenatton, Florian Wenzel, and Dustin Tran. Distilling ensembles improves uncertainty estimates. In Third symposium on advances in approximate bayesian inference, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[52] Nis Meinert and Alexander Lavin. Multivariate deep evidential regression. arXiv preprint arXiv:2104.06135, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[53] Zhong Meng, Jinyu Li, Yong Zhao, and Yifan Gong. Conditional teacher-student learning. In ICASSP, pages 6445-6449. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[54] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, pages 5191-5198, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.441 + ], + "angle": 0, + "content": "[55] Deep Shankar Pandey and Qi Yu. Learn to accumulate evidence from all training samples: theory and practice. In ICML, pages 26963-26989. PMLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.442, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[56] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, pages 3967-3976, 2019. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.537 + ], + "angle": 0, + "content": "[57] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, pages 5007-5016, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[58] Bo Peng, Zhen Fang, Guangquan Zhang, and Jie Lu. Knowledge distillation with auxiliary variable. In ICML, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[59] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[61] Max Ryabinin, Andrey Malinin, and Mark Gales. Scaling ensemble distribution distillation to many classes with proxy targets. NIPS, 34:6023-6035, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.905, + 0.803 + ], + "angle": 0, + "content": "[62] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference. In HPEC, pages 1-9. IEEE, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[63] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[64] Murat Sensoy, Lance Kaplan, and Melih Kandemir. *Evidential deep learning to quantify classification uncertainty.* NIPS, 31, 2018. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2823" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.12 + ], + "angle": 0, + "content": "[65] Glenn Shafer. Dempster-shafer theory. Encyclopedia of artificial intelligence, 1:330-331, 1992. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.177 + ], + "angle": 0, + "content": "[66] Yichen Shen, Zhilu Zhang, Mert R Sabuncu, and Lin Sun. Real-time uncertainty estimation in computer vision via uncertainty-aware distribution distillation. In WACV, pages 707-716, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.484, + 0.22 + ], + "angle": 0, + "content": "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.222, + 0.484, + 0.261 + ], + "angle": 0, + "content": "[68] Che Sun, Yunde Jia, and Yuwei Wu. Evidential reasoning for video anomaly detection. In ACMMM, pages 2106-2114, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.305 + ], + "angle": 0, + "content": "[69] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, pages 15731-15740, 2024. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.307, + 0.484, + 0.348 + ], + "angle": 0, + "content": "[70] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.35, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.407, + 0.484, + 0.475 + ], + "angle": 0, + "content": "[72] Linh Tran, Bastiaan S Veeling, Kevin Roth, Jakub Swiatkowski, Joshua V Dillon, Jasper Snoek, Stephan Mandt, Tim Salimans, Sebastian Nowozin, and Rodolphe Jenatton. Hydra: Preserving ensemble diversity for model distillation. arXiv preprint arXiv:2001.04694, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.477, + 0.484, + 0.505 + ], + "angle": 0, + "content": "[73] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In CVPR, pages 1365-1374, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.507, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[74] Meet Vadera, Brian Jalaian, and Benjamin Marlin. Generalized bayesian posterior expectation distillation for deep neural networks. In Conference on Uncertainty in Artificial Intelligence, pages 719-728. PMLR, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.618 + ], + "angle": 0, + "content": "[75] Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun Zhou, and Edwin Hancock. Uncertainty estimation for stereo matching based on evidential deep learning. pattern recognition, 124:108498, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.62, + 0.484, + 0.674 + ], + "angle": 0, + "content": "[76] Hanjing Wang and Qiang Ji. Beyond dirichlet-based models: when bayesian neural networks meet evidential deep learning. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.484, + 0.731 + ], + "angle": 0, + "content": "[77] Kuan-Chieh Wang, Paul Vicol, James Lucas, Li Gu, Roger Grosse, and Richard Zemel. Adversarial distillation of bayesian neural network posteriors. In ICML, pages 5190-5199. PMLR, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.484, + 0.8 + ], + "angle": 0, + "content": "[78] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, pages 568-578, 2021. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.803, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[79] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[80] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, pages 588-604. Springer, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[81] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions. Springer, 2008. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[82] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. ICLR (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[83] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.235 + ], + "angle": 0, + "content": "[84] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In CVPR, pages 3517-3526, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.237, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[85] Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Jialu Liu, Michael Bendersky, Marc Najork, and Chao Zhang. Do not blindly imitate the teacher: Using perturbed loss for knowledge distillation. arXiv preprint arXiv:2305.05010, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.294, + 0.905, + 0.346 + ], + "angle": 0, + "content": "[86] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.405 + ], + "angle": 0, + "content": "[87] Yuan Zhang, Weihua Chen, Yichen Lu, Tao Huang, Xiuyu Sun, and Jian Cao. Avatar knowledge distillation: self-ensemble teacher paradigm with uncertainty. In ACMMM, pages 5272-5280, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.407, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[88] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, pages 11953-11962, 2022. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[89] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In CVPR, pages 16965-16974, 2024. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "2824" + } + ] +] \ No newline at end of file diff --git a/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_origin.pdf b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..b04e8c6bb7d3965ed7b35d92c74f1b8612f191f6 --- /dev/null +++ b/2025/Evidential Knowledge Distillation/c547b9b0-9570-47d5-bf3e-877b877331fc_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:a237b1e7bbf4b59bca3ba5dd734803605ad600543b5fd33fa98323315c999c89 +size 2650568 diff --git a/2025/Evidential Knowledge Distillation/full.md b/2025/Evidential Knowledge Distillation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..685f6f653df1231e3640f6144f156c43b03acf9e --- /dev/null +++ b/2025/Evidential Knowledge Distillation/full.md @@ -0,0 +1,340 @@ +# Evidential Knowledge Distillation + +Liangyu Xiang $^{1,2}$ Junyu Gao $^{1,2\dagger}$ Changsheng Xu $^{1,2,3}$ + +$^{1}$ State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA) + +$^{2}$ School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS) + +$^{3}$ Peng Cheng Laboratory, ShenZhen, China + +xiangliangyu2023@ia.ac.cn; {junyu.gao, csxu}@nlpr.ia.ac.cn + +# Abstract + +Existing logit-based knowledge distillation methods typically employ singularly deterministic categorical distributions, which eliminates the inherent uncertainty in network predictions and thereby limiting the effective transfer of knowledge. To address this limitation, we introduce distribution-based probabilistic modeling as a more comprehensive representation of network knowledge. Specifically, we regard the categorical distribution as a random variable and leverage deep neural networks to predict its distribution, representing it as an evidential second-order distribution. Based on the second-oder modeling, we propose Evidential Knowledge Distillation (EKD) which distills both the expectation of the teacher distribution and the distribution itself into the student. The expectation captures the macroscopic characteristics of the distribution, while the distribution itself conveys microscopic information about the classification boundaries. Additionally, we theoretically show that EKD's distillation objective provides an upper bound on the student's expected risk when treating the teacher's predictions as ground truth. Extensive experiments on several standard benchmarks across various teacher-student network pairs highlight the effectiveness and superior performance of EKD. Our code is available at https://github.com/lyxiang-casia/EKD. + +# 1. Introduction + +As deep neural networks (DNNs) continue to advance, their performance has rapidly improved, yet their increasing network size and training costs pose challenges, especially in resource-constrained and real-time processing scenarios [62, 89]. Consequently, logit-based knowledge distillation (KD) [30, 88], which aims to compress the predictive capabilities of complex models into lightweight ones, has + +![](images/e9c82d45add299cc7332d4345c442a3f96c84aec0be55c85a4b295232af45d9d.jpg) + +![](images/48d12f289c6be655a348d29dd4f14be951fbedc2e92ea5139810ad62e2ea7dbd.jpg) +(a) Vanilla Knowledge Distillation +(b) Evidential Knowledge Distillation +Figure 1. Illustrations of vanilla KD and our EKD. (a) Vanilla KD deterministically converts the logits into categorical distributions through the softmax function. Each categorical distribution corresponds to a point on the simplex, with each categorical probability representing the barycentric coordinates of that point. Vanilla KD is achieved by minimizing the differences between the teacher's and student's barycentric coordinates. (b) EKD treats the categorical distribution as a random variable following a Dirichlet distribution, parameterized by the logits. The heatmap depicts the probability density of the Dirichlet distribution across the simplex. Moreover, EKD includes distillation of both the expectation of the second-order distribution, and the second-order distribution itself. + +gained increasing popularity. KD transfers knowledge by guiding the predictions of a smaller network to align with those of a larger network during the training. + +Current knowledge distillation methods largely adhere to the vanilla paradigm [30] where only the categorical distri + +butions are aligned. As illustrated in Figure 1a, the logits of a complex network (teacher network) and a lightweight network (student network) are firstly converted into categorical probabilities using a softmax function. Knowledge is then transferred through the process of minimizing the KL divergence between the two categorical distributions. Enhancements to this distillation process, such as decoupling the categorical distributions [88], smoothing the distribution with multiple levels [45], or adopting sample-wise distribution smoothness [69], have demonstrated notable improvements in the student network's predictive performance. Despite these advancements, all such methods are constrained by a fundamental limitation in their probabilistic modeling, which assumes that categorical probabilities are singularly deterministic [12, 14]. Specifically, vanilla KD implicitly assumes that for any given sample, its probability belonging to a particular class is deterministic and can be approximated by a deep neural network. In the case of a three-class classification problem, as shown in Figure 1a, this assumption is akin to identifying a single point on a 2D simplex that best represents the entire simplex. However, in practice, networks trained on finite data with limited capability exhibit inherent uncertainty in their predictions for specific samples [39, 47]. For example, networks with different initial weights may yield different predictions for the same test samples. Consequently, this singularly deterministic probabilistic modeling creates a bottleneck in the effective representation of diverse and uncertain knowledge. + +To overcome the limitations of singularly deterministic probabilistic modeling, we employ a distribution-based probabilistic modeling, treating the categorical distribution as a random variable governed by a second-order distribution. Following Subjective Logic [35] and Dempster-Shafer Theory of Evidence [81], the Dirichlet distribution, a conjugate prior of the categorical likelihood, instantiates this second-order distribution, facilitating the success of evidential deep learning across multiple domains [5, 16, 25]. However, recent studies employing distribution-based probabilistic modeling are largely confined to macro-level training within a single network, where the second-order distribution is integrated into a single categorical distribution during the network's training process [8, 11, 64]. The second-order distribution inherently provides micro-level information across all possible categorical distributions, offering advantages in cross-network information propagation. Therefore, the second-order distribution serves as a bridge to enable knowledge transfer at both the macro and micro levels, supported by PAC-Bayesain theory [19]. As illustrated in Figure 1b, the logits output by each network are converted into a set of Dirichlet parameters, forming respective second-order distributions. Essentially, the second-order distribution assigns varying weights to all possible first-order categorical distributions, + +where the non-uniformity of these weights reflects the network's capacity for classification. Building on second-order predictions, we propose the Evidential Knowledge Distillation (EKD), integrating both macro and micro perspectives. At the macro level, we reduce each second-order distribution to their first-order categorical distributions via an expectation operation in order to ensure that the centroids of their Dirichlet distributions coincide. By representing the entire simplex with a single point, the student is encouraged to focus more on the overall density distribution of the teacher at this stage, which leads to the proportions of student's outputs across categories aligning with those of the teacher. At the micro level, we replace the first-order distribution with the second-order distribution to enable the transfer of more granular classification information. As a complement to proportional alignment, this step refines the magnitude in the student network's outputs. Ultimately, the student model benefits from both the teacher's overall characteristics and its detailed classification structure. + +To summarize, Our main contribution are as follows: + +- We reformulate network predictions in knowledge distillation through distribution-based probabilistic modeling. Building on this reformulation, we propose EKD, which distills the teacher's second-order distribution into the student model across both macro and micro levels. +- With the help of PAC-Bayesian theory [19], we prove that the distillation objective in EKD serves as an upper bound on the expected risk of the student network when treating the teacher network's outputs as ground truth. +- Extensive experiments are conducted on several standard benchmarks, encompassing a wide range of teacher and student models. Without bells and whistles, EKD achieves performance improvement of up to $3.61\%$ over KD across various teacher-student network combinations. + +# 2. Related Work + +Knowledge Distillation Knowledge distillation enhances a simpler student network using insights from a complex teacher network. Existing methods fall into three categories: logit-based [2, 30, 44, 45, 54, 85, 88], feature-based [33, 56, 57, 70, 73], and hybrid approaches [22, 31, 34, 58, 80, 82]. Unlike feature-based methods, which demand architecture-specific designs, logit-based methods offer wider applicability by relying solely on network outputs [4, 53, 84]. Introduced by Hinton and Geoffrey [30], logit-based distillation converts teacher and student logs into categorical distributions aligned via KL divergence. Subsequent efforts have refined this by adjusting the temperature parameter dynamically, based on training stages [45], teacher-student gaps [21], and output variance [69], or by decomposing distributions [88] and leveraging intermediate networks [31, 54]. + +The aforementioned traditional distillation methods + +share a common limitation in their reliance on point estimates for model parameters or categorical distributions, whereas Bayesian distillation [13, 37, 50, 74, 77] employs distributional estimation, which enables the student network to inherit the teacher's strengths in uncertainty estimation. Initial explorations [24, 37, 46, 51] adopt teacher ensembles to approximate the distribution of teacher model parameters and utilize the vanilla KD [30] for distillation. Subsequent extensions continue to advance by improving the student network architecture [20, 50, 61, 72, 76], optimization objectives [74], and sampling methods [13]. Additionally, some approaches introduce the utilization of quantized uncertainty during training to enable more fine-grained knowledge transfer [28, 66, 87]. While these methods introduce uncertainty into the model parameters and the resultant predictive distributions, they remain limited by their reliance on a finite number of ensembles to approximate the distributions. Consequently, we propose to adopt a continuous probabilistic modeling framework for both the teacher and the student, while leveraging more advanced distillation techniques. Unlike Bayesian distillation, which performs uncertainty-related distillation, we implicitly express the uncertainty of the predictive distribution through second-order probabilistic modeling and focus on enhancing the performance of the student network. + +Evidential Deep Learning Based on Subjective Logic (SL) [35] and Dempster-Shafer Theory (DST) [65], Sensoy et al. [64] initially propose a framework in which evidence collected by a deep neural network for each class is used to parameterize a second-order Dirichlet distribution for network predictions. This concept of evidence collection allows evidential networks to express "I don't know", showing advantages in uncertainty estimation. Since then, numerous studies have advanced and expanded upon Evidential Deep Learning (EDL) from both theoretical and application-oriented perspectives [17]. Theoretical investigations have primarily focused on refining the evidence collection process [11, 15, 18, 40, 55], delving into different training strategies [5, 26, 43] and extending evidential theory to regression networks [1, 49, 52]. Moreover, EDL has been widely applied in computer vision tasks, including action localization [5-7, 16], stereo matching [75], anomaly detection [68], and multi-view classification [25]. + +# 3. Method + +In this section, we first present the fundamental framework of our Evidential Knowledge Distillation (EKD), which includes both macroscopic first-order and microscopic second-order distribution distillation. Theoretically, we utilize PAC-Bayesian theory to demonstrate that the optimization objective of EKD effectively represents an upper bound on the student's expected risk. At the end of this section, we illustrate the complementary effect of the two + +distillation strategies with a concrete toy case. + +# 3.1. Preliminaries + +Evidential Deep Learning. In accordance with the formulation of evidential deep learning [8, 9, 64], for a classification task involving $K$ classes, the categorical probability vector $p$ is treated as a random variable governed by a conjugate prior Dirichlet distribution. To define this distribution's parameters, an evidential activation function $\sigma$ which we designate as the exponential function (exp), is employed to transform the network's output logits $z$ into a non-negative evidence vector $e = \sigma(z)$ . This evidence vector is subsequently combined with a prior weight $\lambda$ to yield the $K$ -dimensional Dirichlet parameters $\alpha$ which fully specify a Dirichlet distribution $Dir(\alpha)$ [8]. In terms of optimization, the widely adopted cross-entropy loss is computed by integrating over the categorical probabilities, guided by second-order the Dirichlet distribution. + +$$ +\begin{array}{l} \mathcal {L} _ {E D L - c e} = \int \left[ \sum_ {i = 1} ^ {K} - y _ {i} \log \left(p _ {i}\right) \right] D i r (\boldsymbol {p} | \boldsymbol {\alpha}) d \boldsymbol {p} \\ = \sum_ {i = 1} ^ {K} y _ {i} (\psi (\alpha_ {0}) - \psi (\alpha_ {i})) \tag {1} \\ \end{array} +$$ + +Here, $y_{i}$ denotes the ground truth, $\alpha_0 = \sum_{i=1}^{K} (\alpha_i)$ represents the Drichlet strength, and $\psi$ refers to the digamma function. To support the acquisition of evidence, we employ Eq. (1), rather than the conventional cross-entropy loss, across all network training stages, including both the training of the teacher networks and student networks during the knowledge distillation process. + +Knowledge Distillation. Existing logit-based knowledge distillation methods rely on categorical distribution predictions derived from the softmax function. Specifically, for a given input sample $x$ , the network produces logits $z$ , which are then transformed into a categorical distribution via the softmax operation, denoted as $p = \mathrm{softmax}(z)$ . Knowledge transfer is accomplished by minimizing the Kullback-Leibler divergence(KL) divergence between the teacher's and student's predicted distributions, $p^T$ and $p^S$ : + +$$ +\mathcal {L} _ {K D} = K L \left(\boldsymbol {p} ^ {T} \| \boldsymbol {p} ^ {S}\right) \tag {2} +$$ + +The teacher's predictions serve as soft labels, replacing the traditional one-hot labels, capturing inter-class similarities. + +# 3.2. Evidential Knowledge Distillation + +Previous knowledge distillation methods have generally followed a unified paradigm, where the network's output logits are transformed into categorical probabilities via the softmax function, and the student network is guided to align its categorical distribution with that of the teacher network. While these approaches have yielded notable success, a fundamental limitation persists: they treat categorical probabilities as singularly deterministic values. This assumption + +overlooks the uncertainty in network predictions, thus constraining the sharing of more granular information. As a more robust alternative, we predict the distribution of categorical distributions using neural networks, specifically modeling it as a second-order Dirichlet distribution. More importantly, we align both the first-order distribution, obtained by integrating the second-order distribution, and the second-order distribution itself. + +Firstly, we aim to extract the overall characteristics of the second-order distribution to capture the general distribution trend of probability density across the simplex, as well as to achieve alignment between networks in terms of final predicted classes. Specifically, given the second-order distributions predicted by the teacher and student, $Dir(\pmb{\alpha}^T)$ and $Dir(\pmb{\alpha}^S)$ , we compute their expectations where all possible categorical distributions are summed under the weights specified by the second-order distribution: + +$$ +\hat {\boldsymbol {p}} ^ {T} = \mathbb {E} _ {\boldsymbol {p} ^ {T} \sim D i r (\boldsymbol {\alpha} ^ {T})} [ \boldsymbol {p} ^ {T} ], \hat {\boldsymbol {p}} ^ {S} = \mathbb {E} _ {D i r (\boldsymbol {p} ^ {S} \sim \boldsymbol {\alpha} ^ {S})} [ \boldsymbol {p} ^ {S} ] \tag {3} +$$ + +The expectation above, serving as the final predictive outcome for both teacher and student, is equivalent to the centroid of the second-order distribution, enabling the transfer of comprehensive probability density information via first-order distillation at a macro level. + +$$ +\mathcal {L} _ {1 s t} = K L (\hat {\boldsymbol {p}} ^ {T} | | \hat {\boldsymbol {p}} ^ {S}) = \sum_ {i = 1} ^ {K} \frac {\alpha_ {i} ^ {T}}{\alpha_ {0} ^ {T}} l o g \left(\frac {\alpha_ {i} ^ {T} \alpha_ {0} ^ {S}}{\alpha_ {0} ^ {T} \alpha_ {i} ^ {S}}\right) \tag {4} +$$ + +Formally, optimizing Eq. (4) involves adjusting the proportion of each class's Dirichlet parameter relative to the total, which significantly enhances the alignment between the student's and teacher's predicted classes. In terms of formulation, Eq. (4) is very similar to vanilla KD, yet it is emphasized that vanilla KD can be regarded as a special case of first-order distillation when $\sigma = \exp(\cdot)$ and $\lambda = 0$ . Proof is provided in the Supplementary Material. + +The first-order distillation objective above only provides macroscopic guidance for the student's second-order distribution, failing to accurately and comprehensively represent the weighting of each categorical distribution. In other words, given a fixed teacher distribution, the student Dirichlet distribution that minimizes Eq. (4) is not unique. Therefore, we propose a microscopic second-order distillation as a complement to first-order distillation, aiming to provide fine-grained teacher supervision signals across the entire simplex for the student distribution. Specifically, we achieve this by applying the KL divergence directly between the two Dirichlet distributions: + +$$ +\begin{array}{l} \mathcal {L} _ {2 n d} = K L (D i r (\boldsymbol {p} ^ {T} | \boldsymbol {\alpha} ^ {T}) | | D i r (\boldsymbol {p} ^ {S} | \boldsymbol {\alpha} ^ {S})) \\ = l o g \frac {\Gamma (\alpha_ {0} ^ {T})}{\Gamma (\alpha_ {0} ^ {S})} - \sum_ {i = 1} ^ {K} l o g \frac {\Gamma (\alpha_ {i} ^ {T})}{\Gamma (\alpha_ {i} ^ {S})} \\ + \sum_ {i = 1} ^ {K} \left(\alpha_ {i} ^ {T} - \alpha_ {i} ^ {S}\right) \left(\psi \left(\alpha_ {i} ^ {T}\right) - \psi \left(\alpha_ {0} ^ {T}\right)\right) \tag {5} \\ \end{array} +$$ + +Here, $\Gamma(\cdot)$ denotes the Gamma function. The detailed derivation for Eq. (5) is provided in Supplementary Material. Since Eq. (5) considers the whole $(K - 1)$ dimensional simplex, overflow may occur [50]; thus, we apply the $\log(1 + x)$ function to smooth the network outputs. While the above Eq. (4) aims to optimize the proportional relationships between classes, Eq. (5) here places greater emphasis on aligning the magnitudes within each class. Here, the alignment within the class means that the gradients of a loss function with respect to the output for a particular class do not conflict with the gradients of the output for other classes. This meticulous numerical alignment supplies each point on the $(K - 1)$ -dimensional simplex of the student's second-order distribution with the corresponding teacher label, ensuring a precise mapping of knowledge across the entire distribution space. By doing so, it strives to faithfully replicate the teacher's second-order predictive nuances at a granular, micro-level, capturing the fine-grained patterns and uncertainties inherent in the teacher model's outputs. + +By combining these two complementary optimization objectives, one optimizing proportional relationships at the macro level and the other aligning numerical values at the micro level, we arrive at the final EKD distillation: + +$$ +\mathcal {L} _ {E K D} = \mathcal {L} _ {1 s t} + \gamma \mathcal {L} _ {2 n d}, \tag {6} +$$ + +where $\gamma$ represents the coefficient that adjusts the balance between the first-order and second-order distillation. + +# 3.3. Theoretical Analysis + +In this section, we provide theoretical support for Eq. (6) using PAC-Bayesian Theory [19]. While PAC-Bayesian theory has been explored in the context of the Dirichlet distribution [26], its application remains confined to single-network optimization, neglecting the relationship among multiple networks. Therefore, we aim to estimate the alignment of different networks across the entire data distribution based on their alignment on training samples. + +Let $x$ denote a certain sample and $y$ denote its corresponding label. Under the PAC setting, samples are drawn from a fixed but unknown distribution $\mathcal{D}$ . A classifier $h$ maps $x$ to $y$ . Moreover, we define a classifier $h$ correspond to a point on the $(K - 1)$ -dimensional simplex, i.e., a categorical distribution $p$ . Essentially, the Dirichlet distribution made by each network serves as the posterior distribution of a set of classifiers. By weighting the set of classifiers with its posterior distribution, we obtain the final majority vote classifier (or Bayesain classifier): + +$$ +\boldsymbol {B} ^ {T} = \mathbb {E} _ {\operatorname {D i r} (\boldsymbol {\alpha} ^ {T})} [ \boldsymbol {p} ^ {T} ], \tag {7} +$$ + +$$ +\boldsymbol {B} ^ {S} = \mathbb {E} _ {\operatorname {D i r} (\boldsymbol {\alpha} ^ {S})} [ \boldsymbol {p} ^ {S} ], \tag {8} +$$ + +where $\pmb{B}^{T}$ and $\pmb{B}^{S}$ denote the final classification results of the teacher and student classifiers, respectively. + +In knowledge distillation, the student treats the teacher's predictions as soft labels. Below, we define the student's + +misclassification risk using cross-entropy, which is equivalent to KL divergence from an optimization perspective. + +$$ +R ^ {S} = \mathbb {E} _ {x \sim \mathcal {D}} \left[ - \boldsymbol {B} ^ {T} \log \boldsymbol {B} ^ {S} \right], \tag {9} +$$ + +$$ +\hat {R} ^ {S} = \frac {1}{N} \sum_ {i = 1} ^ {N} \left[ - \boldsymbol {B} _ {i} ^ {T} \log \boldsymbol {B} _ {i} ^ {S} \right]. \tag {10} +$$ + +where $R^S$ represents the expected risk of the student over the entire data distribution $\mathcal{D}$ and $\hat{R}^S$ denotes the empirical risk on the training set $\mathcal{D}_t$ containing $N$ training samples. + +Theorem 1. For any $\delta \in (0,1]$ and $\gamma \in \mathbb{R}^{+}$ , $R^S$ is bounded above by the sum of $\hat{R}^S$ , $\sum_{1}^{N}KL(Dir(\pmb{\alpha}_i^T))||Dir(\pmb{\alpha}_i^S))$ , and a constant related to $\delta$ , denoted as $C(\delta)$ : + +$$ +\begin{array}{l} P _ {\mathcal {D} _ {t} \sim \mathcal {D} ^ {N}} (R ^ {S} \leq \hat {R} ^ {S} + \frac {\gamma}{N} \sum_ {i = 1} ^ {N} K L (D i r (\pmb {\alpha} _ {i} ^ {T}) | | D i r (\pmb {\alpha} _ {i} ^ {S})) \\ + C (\delta)) \geq 1 - \delta . \tag {11} \\ \end{array} +$$ + +Detailed proof and definition of $C(\delta)$ can be found in Supplementary Material. Theorem 1 indicates that the upper bound on the student's excepted risk over the unknown data distribution $\mathcal{D}$ consists of two components: the alignment of the first-order distribution with the true labels (teacher labels) and the similarity between the predicted second-order distribution and the true second-order distribution (teacher distribution). Clearly, this upper bound indicates that the empirical risk optimization represented by Eq. (2) does not guarantee a reduction in the student's expected risk, suggesting that the transfer of the teacher's generalization performance is both inefficient and incomplete. In contrast, EKD directly adopts this upper bound as optimization objectives, refining the formulation of empirical risk while incorporating second-order distribution alignment between the student and teacher, thereby striving to minimize the student's expected risk. + +# 3.4. Toy Case + +Figure 2 demonstrates the complementary roles of first-order and second-order distillation through a three-class example. Initially, the student Dirichlet parameters differ significantly from those of the teacher, both in relative proportion and in absolute values. The left side of Figure 2 illustrates the essence of first-order distillation, which minimizes the distance between the expectations of two Dirichlet distributions over the $(K - 1)$ dimensional simplex measured by KL divergence. However, the expectation of the Dirichlet distribution depends solely on the proportion of each class's Dirichlet parameter relative to the overall sum, which can lead to discrepancies in the specific values of the student's output compared to the teacher's. Although this discrepancy does not affect the final class prediction, it causes the student's second-order distribution to differ significantly from the teacher's. To address this, we use + +![](images/648f6e6a352fa9bba6eba19e4d2833bd02ba36a33432a16289f374833fcf0545.jpg) +Figure 2. A toy case illustrating the complementary roles of the two distillation losses in EKD. Given the Dirichlet parameters output by teacher and student networks, the left panel shows the results when only first-order distillation is used to optimize the proportion of each category. The right panel presents the outcome when only second-order distillation is applied for intra-class numerical alignment. The center panel presents the results of combining both objectives, which yields the optimal performance. + +second-order distillation as a complement to first-order distillation, aiming to align the Dirichlet parameter values. In contrast to first-order distillation, second-order distribution alignment encourages consistency of parameters within the same class across different networks, disregarding the relative magnitudes among different classes. As a result, we combine these complementary distillation objectives to achieve a comprehensive alignment of the second-order distribution. The quantitative effects of the two types of distillation will be discussed in detail in Section 4.2 + +# 4. Experiments + +Datasets. Our experiments are primarily conducted on two standard datasets including CIFAR100 [38] and ImageNet [60]. The CIFAR-100 dataset [38] consists of 60k 32x32 color images divided into 100 classes, with each class containing 600 images. It is split into a training set of 50k images and a validation set of 10k images. ImageNet [60] is a large-scale visual database comprising 1,000 categories that cover a diverse range of objects and scenes. It includes approximately 1.28 million labeled images for training and an additional 50,000 images for validation. + +Implementation Details. We adopt the experimental settings established in prior studies [10, 69, 88]. For the + +Table 1. Results on the CIFAR-100 validation set. The teacher and student networks share the same architecture. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. $\Delta$ denotes the improvement of the distilled student model compared to the independently trained student model. Note that "Softmax" indicates that the model uses the vanilla KD probabilistic model, while "EDL" indicates that the model are trained by Eq. (1). \*\* indicates our reproduced results. + +
Experiment Group1234567
TeacherArchitectureResNet32×4VGG13WRN-40-2WRN-40-2ResNet56ResNet110ResNet110
Softmax79.4274.6475.6175.6172.3474.3174.31
EDL79.5374.9675.4175.4173.0974.6774.67
StudentArchitectureResNet8×4VGG8WRN-40-1WRN-16-2ResNet20ResNet32ResNet20
Softmax72.5070.3671.9873.2669.0671.1469.06
EDL72.7770.6771.8473.3669.2270.969.22
FeatureRKD[CVPR'19] [56]71.9071.4872.2273.3569.6171.8269.25
CRD[ICLR'20] [70]75.5173.9474.1475.4871.1673.4871.46
ReviewKD[CVPR'21] [10]75.6374.8475.0976.1271.8973.8971.34
SimKD[CVPR'22] [3]78.0874.8974.5375.5371.0573.9271.06
CAT-KD[CVPR'23] [23]76.9174.6574.8275.6071.6273.6271.37
TopKD[ICML'24] [36]75.4074.0174.4375.7571.5873.7771.47
LogitKD[NeurIPS'14] [30]73.3372.9873.5474.9270.6673.0870.67
Δ0.832.621.561.661.61.941.61
DKD[CVPR'22] [88]76.3274.6874.8176.2471.9774.1171.06
Δ3.824.322.832.982.912.972.0
CTKD[AAAI'23] [45]73.3973.5273.9375.4571.1973.5270.99
Δ0.893.161.952.192.132.381.93
Logit_Stand[CVPR'24] [69]76.6274.3674.3776.1171.4374.1771.48
Δ4.124.002.392.852.373.032.42
SDD[CVPR'24] [48]75.92*73.90*74.26*75.98*70.72*73.90*71.43*
Δ3.42*3.54*2.28*2.72*1.66*2.76*2.37*
TeKAP[ICLR'25] [31]74.7973.8*73.8075.2171.7173.50*71.00*
Δ2.293.44*1.821.952.652.36*1.94*
EKD(Ours)77.2174.7174.4376.1571.4873.6871.51
Δ4.444.042.592.792.262.782.29
+ +Table 3. The top-1 and top-5 accuracy (\%) on the ImageNet validation set [60]. The best and second best results are emphasized in bold and underlined. “*” indicates our reproduced results. + +
TeacherArchitectureResNet34ResNet50
Accuracytop-1top-5top-1top-5
Softmax73.3191.4276.1692.86
EDL73.8391.6976.1092.98
StudentArchitectureResNet18MNV1
Accuracytop-1top-5top-1top-5
Softmax69.7589.0768.8788.76
EDL70.2889.5869.9289.46
FeatureCRD [70]71.1790.1371.3790.41
ReviewKD [10]71.6190.5172.5691.00
SimKD [3]71.5990.4872.2590.86
CAT-KD [23]71.2690.4572.2491.13
LogitKD [30]71.0390.0570.5089.80
DKD [88]71.7090.4172.0591.05
CTKD [45]71.3890.2771.1690.11
Logit_Stand [69]71.4290.2972.1890.80
SDD [48]71.4490.0572.15*90.79*
EKD(Ours)71.7390.6972.5491.20
+ +experiments on CIFAR-100, we use the SGD optimizer with 240 epochs. The learning rate is set to 0.01 for MobileNets [32, 63] and ShuffleNets [86] and 0.05 for other networks, including ResNets [27], WRNs [83] and VGGs [67], with a decay factor of 0.1 applied at epochs 150, 180, and 210. Regarding hyperparameter configuration, we set $\gamma$ in Eq. (6) to 1 by default, requiring no additional hyperparameter tuning. All results represent the mean of five + +independent runs. Additional experimental details are provided in the Supplementary Material. + +Baselines. To benchmark EKD, we incorporate five state-of-the-art logit-based distillation methods: DKD (CVPR'22) [88], CTKD (AAAI'23) [45], Logit_Stand (CVPR'24) [69], SDD (CVPR'24) [48] and TeKAP (ICLR'25) [31]. Additionally, a selection of feature-based distillation methods [3, 10, 23, 56, 59, 70] is included as references to facilitate a broader contextual analysis. + +# 4.1. Main Results + +Performance comparison on CIFAR100. We conduct experiments comparing EKD with other distillation methods under various teacher-student pair settings. Table 1 and Table 2 present the accuracy results for settings where the teacher and student have identical and distinct architectures. As presented in Table 1, EKD demonstrates favourable performance improvements over KD, ranging from $0.66\%$ to $3.61\%$ . In particular, EKD demonstrates its strongest performance in Groups 1, 2, and 7, while achieving comparable results to DKD [88] in Groups 3-6. The underlying reason for this phenomenon lies in the performance gap between the teacher and student. In groups 1, 2, and 7, the average performance gap between the teacher and student is $5.51\%$ , whereas it is $3.31\%$ in groups 3-6. Furthermore, we observe that pairs with a larger performance gap tend to have a higher distillation loss, indicating a greater opti + +Table 2. Results on the CIFAR-100 validation set. The teacher and student networks have different architectures. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. $\Delta$ denotes the improvement of the distilled student model compared to the independently trained student model. Note that "Softmax" indicates that the model uses the vanilla KD probabilistic model, while "EDL" indicates that the model are trained by Eq. (1). \*\* indicates our reproduced results. + +
Experiment Group1234567
TeacherArchitectureResNet32×4ResNet32×4ResNet32×4WRN-40-2WRN-40-2VGG13ResNet50
Softmax79.4279.4279.4275.6175.6174.6479.34
EDL79.5379.5379.5375.4175.4175.1179.69
StudentArchitectureSHN-V2WRN-16-2WRN-40-2ResNet8×4MN-V2MN-V2MN-V2
Softmax71.8273.2675.6172.5064.6064.6064.60
EDL72.7373.3675.4172.7765.3965.3965.39
FeatureRKD[CVPR'19] [56]73.2174.8677.8275.2669.2764.5264.43
CRD[ICLR'20] [70]75.6575.6578.1575.2470.2869.7369.11
ReviewKD[CVPR'21] [10]77.7876.1178.9674.3471.2870.3769.89
SimKD[CVPR'22] [3]78.3977.1779.2975.2970.1069.4469.97
CAT-KD[CVPR'23] [23]78.4176.9778.5975.3870.2469.1371.36
TopKD[ICML'24] [36]76.3374.34*76.18*75.83*69.54*69.32*70.11*
LogitKD[NeurIPS'14] [30]74.4574.9077.7073.9768.3667.3767.35
Δ2.631.642.091.473.762.772.75
DKD[CVPR'22] [88]77.0775.7078.4675.5669.2869.7170.35
Δ5.252.442.853.064.685.115.75
CTKD[AAAI'23] [45]75.3774.5777.6674.6168.3468.5068.67
Δ3.551.312.052.113.743.904.07
Logit_Stand[CVPR'24] [69]75.5675.2677.9277.1169.2368.6169.02
Δ3.742.002.314.614.634.014.42
SDD[CVPR'24] [48]76.88*75.92*78.54*74.83*70.14*69.06*69.31*
Δ5.06*2.66*2.93*2.33*5.54*4.46*4.71*
TeKAP[ICLR'25] [31]75.4375.26*76.93*74.99*68.65*68.27*68.75*
Δ3.612.00*1.32*2.49*4.05*3.67*4.15*
EKD(Ours)77.6576.5178.9176.7569.2069.9470.38
Δ4.923.153.503.983.814.554.99
+ +Table 4. Ablation study for each components in EKD. The "MSE" refers to aligning the magnitude of model outputs with mean square error. The "Accuracy" denotes the Top-1 accuracy (\%) on the CIFAR100 validation set. ResNet32×4 and ResNet8×4 are adopted as teacher and student. + +
IndexDistillation ObjectiveAccuracy
\( \mathcal{L}_{KD} \)\( \mathcal{L}_{1st} \)\( \mathcal{L}_{2nd} \)MSE
1✓74.45
2✓73.80
3✓76.42
4✓76.10
5✓✓77.21
6✓✓76.83
7✓✓76.34
8✓✓76.13
+ +mization space during distillation. Conversely, for teacher-student pairs with closer performance, their distillation loss is generally smaller, leading EKD to encounter local minima during optimization. For the distillation of distinct networks shown in Table 2, EKD achieves the highest performance in half of the experimental groups. When using MobileNet [63] as the student network, the student trained solely with EDL achieves a $0.79\%$ higher performance than the one trained with softmax alone. However, the result for EKD is only comparable to that of other methods. We attribute this phenomenon to the inherent performance ceiling of the student network due to its limited capacity. + +Performance comparison on ImageNet. Table 3 presents the performance of EKD on the large-scale ImageNet dataset. EKD markedly outperforms other logit-based methods in the ResNet50-MNV1 experiment and achieves results comparable to top methods in the ResNet34-ResNet18 experiment. The modest gains in the latter can also be attributed to the narrow performance gap between the teacher and student. + +# 4.2. Ablation Studies + +To elucidate the superiority of the two alignment strategies implemented in EKD and their alignment with theoretical underpinnings, the ablation study presented in Table 4 evaluates various distillation objectives on student model performance. In addressing alignment with respect to the proportional distribution across output classes, we investigate two methodologies: the conventional vanilla KD approach presented in Eq. (2), which aligns softmax-based categorical distributions, and $\mathcal{L}_{1st}$ in Eq. (4). Comparative analyses between Index 1 and Index 2, Index 5 and Index 7, and Index 6 and Index 8 consistently demonstrate that $\mathcal{L}_{1st}$ surpasses vanilla KD. This finding underscores that EKD's first-order alignment serves as a robust generalization of vanilla KD, effectively incorporating the alignment of categorical distributions by leveraging the centroids of second-order Dirichlet distributions. + +In addition to the second-order distillation, $\mathcal{L}_{2nd}$ , shown + +Table 5. The Top-1 accuracy $(\%)$ of combining EKD with feature distillation-based methods on CIFAR-100. + +
TeacherArchitectureVGG13ResNet32x4
Accuracy74.9679.53
StudentArchitectureVGG8SHN-V2
Accuracy70.6772.73
LogitEKD74.7177.65
FeatureRKD [56]71.4873.21
CAT-KD [23]74.6578.41
Logit+FeatureEKD+RKD [56]74.99↑78.08↑
EKD+CAT-KD [23]74.91↑78.46↑
+ +in Eq. (5), another intuitive approach to aligning model output magnitudes is through mean squared error (MSE). Consequently, we compare $\mathcal{L}_{2nd}$ and MSE in Table 4. Comparisons across Index 3 and Index 4, Index 5 and Index 6, and Index 7 and Index 8 reveal that $\mathcal{L}_{2nd}$ consistently outperforms MSE. These empirical results corroborate the theoretical conclusions delineated in Section 3.3. It is suggested that $\mathcal{L}_{2nd}$ , being part of the upper bound on the student's expected risk, directly optimizes the student's generalization performance, whereas the MSE objective does so indirectly. + +# 4.3. Further Remarks + +Visualizations. Figure 3 presents the t-SNE visualizations of the laerned feature representations and, in parallel, the differences in logits between the teacher and student networks, where the teacher is implemented as ResNet $32 \times 4$ and the student as ResNet $8 \times 4$ . As shown in the first row, EKD achieves a clear improvement over vanilla KD baseline with respect to class separability, leading to more distinctly clustered feature distributions. Furthermore, the logits produced under EKD exhibit a noticeably stronger alignment with those of the teacher network, indicating a closer match in predictive behavior compared to the results obtained with vanilla KD. + +Compatibility with feature-based methods. Table 5 presents the results of integrating EKD with several representative feature-based distillation methods [23, 56]. For feature-based approaches, the introduction of EKD enhances their performance, demonstrating strong compatibility, particularly with methods that originally exhibit lower performance [56]. Conversely, these methods can further compensate for the limitation of the EKD approach in expressing knowledge through logits. + +Distilling ViT. Transformer-based models are gaining increasing popularity, and logit-based distillation methods can be easily applied to these models. Therefore, we test the distillation effectiveness of EKD on the ViT model, as presented in Table 6. Following the configuration in [41, 42], we use ResNet56 as the teacher model and four ViT variants, DeiT-Ti [71], PiT-Ti [29], PVT-Ti [78], and PVTv2 [79], as the student models. EKD achieves an aver- + +![](images/2a7f338f5bb870264b0798680b7e76ebaa01936905edbd95875230ec0d3c5c14.jpg) +(a) Vanilla KD + +![](images/2594fa2eed89f487982a5fde7f4bb9a2dfa2f501723ccda403b4d9ca248de348.jpg) +(b) EKD + +![](images/dbe4e60df6dcc6342f43e51ad612ee43a24b0c462fce1c020766092fcf2b4486.jpg) +(c) Vanilla KD + +![](images/3a3fa36847db0e6259a3d13b6b8998a4aecef64ecdf6b4e2814fea8ab6df344f.jpg) +(d) EKD +Figure 3. Visualization of features and logits output. The first row shows the t-SNE of features learned by vanilla KD and EKD. The second row illustrates the difference of correlation matrices of logits between the student and teacher. + +Table 6. Top-1 accuracy of various ViT student models on CIFAR100. Teacher model is ResNet56. + +
ArchitectureSizeEDLSoftmaxKD [30]AutoKD [42]EKD(Ours)
DeiT-Ti [71]5M64.7765.0873.2578.5578.64
PiT-Ti [29]5M73.4873.5875.4778.7679.33
PVT-Ti [78]13M69.0269.2273.6078.4378.74
PVTv2 [79]3M76.6877.4478.8179.3779.80
+ +age performance improvement of $3.75\%$ over KD [30] and an average improvement of $0.25\%$ over AutoKD [42]. + +# 5. Conclusion + +This paper reinterprets vanilla knowledge distillation (KD) through a distribution-based probabilistic framework, employing evidential second-order distributions to effectively capture predictive uncertainty and provide a comprehensive knowledge representation. This pioneering methodology underpins Evidential Knowledge Distillation (EKD), facilitating knowledge transfer across macro and micro levels. At the macro level, aligning the expectation of the second-order distribution, which reflects its global characteristics, enhances optimization of inter-class proportions in the student's output. At the micro level, aligning the second-order distribution aligns the magnitudes of the student's output. These complementary mechanisms ensure a robust distillation process. Through PAC-Bayesian theory, EKD optimizes the upper bound of the student's expected risk. Extensive experiments underscore EKD's pronounced advantage over prevailing logit-based distillation techniques. + +# Acknowledgements + +This work was supported in part by the National Key Research and Development Plan of China under Grant 2023YFC3310700, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62472422, and U2333215, and in part by Beijing Natural Science Foundation under Grant 4242051. + +# References + +[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. NIPS, 33:14927-14937, 2020. 3 +[2] Defang Chen, Jian-Ping Mei, Can Wang, Yan Feng, and Chun Chen. Online knowledge distillation with diverse peers. In AAAI, pages 3430-3437, 2020. 2 +[3] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, pages 11933-11942, 2022. 6, 7 +[4] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. NIPS, 30, 2017. 2 +[5] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV, pages 192-208. Springer, 2022. 2, 3 +[6] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Cascade evidential learning for open-world weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14741-14750, 2023. +[7] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Uncertainty-aware dual-evidential learning for weakly-supervised temporal action localization. IEEE transactions on pattern analysis and machine intelligence, 45(12):15896-15911, 2023. 3 +[8] Mengyuan Chen, Junyu Gao, and Changsheng Xu. R-edl: Relaxing nonessential settings of evidential deep learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3 +[9] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Revisiting essential and nonessential settings of evidential deep learning. arXiv preprint arXiv:2410.00393, 2024. 3 +[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, pages 5008-5017, 2021. 5, 6, 7 +[11] Danruo Deng, Guangyong Chen, Yang Yu, Furui Liu, and Pheng-Ann Heng. Uncertainty estimation by fisher information-based evidential deep learning. In ICML, pages 7596–7616. PMLR, 2023. 2, 3 +[12] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2 +[13] Luyang Fang, Yongkai Chen, Wenxuan Zhong, and Ping Ma. Bayesian knowledge distillation: A bayesian perspec + +tive of distillation with uncertainty quantification. In Forty-first ICML, 2024. 3 +[14] Yarin Gal et al. Uncertainty in deep learning. 2016. 2 +[15] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18827-18836, 2023. 3 +[16] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Vectorized evidential learning for weakly-supervised temporal action localization. PAMI, 2023. 2, 3 +[17] Junyu Gao, Mengyuan Chen, Liangyu Xiang, and Changsheng Xu. A comprehensive survey on evidential deep learning and its applications. arXiv preprint arXiv:2409.04720, 2024.3 +[18] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Learning probabilistic presence-absence evidence for weakly-supervised audio-visual event perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3 +[19] Pascal Germain, Alexandre Lacasse, François Laviolette, and Mario Marchand. Pac-bayesian learning of linear classifiers. In ICML, pages 353-360, 2009. 2, 4 +[20] Hongji Guo, Hanjing Wang, and Qiang Ji. Bayesian evidential deep learning for online action detection. In ECCV, pages 283-301. Springer, 2025. 3 +[21] Jia Guo. Reducing the teacher-student gap via adaptive temperatures. 2022. 2 +[22] Yong Guo, Shulian Zhang, Haolin Pan, Jing Liu, Yulun Zhang, and Jian Chen. Gap preserving distillation by building bidirectional mappings with a dynamic teacher. arXiv preprint arXiv:2410.04140, 2024. 2 +[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, pages 11868-11877, 2023. 6, 7, 8 +[24] Corina Gurau, Alex Bewley, and Ingmar Posner. Dropout distillation for efficiently estimating model confidence. arXiv preprint arXiv:1809.10562, 2018. 3 +[25] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 2, 3 +[26] Manuel Haussmann, Sebastian Gerwinn, and Melih Kandemir. Bayesian evidential deep learning with pac regularization. arXiv preprint arXiv:1906.00816, 2019. 3, 4 +[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6 +[28] Mina Hemmatian, Ali Shahzadi, and Saeed Mozaffari. Uncertainty-based knowledge distillation for bayesian deep neural network compression. International Journal of Approximate Reasoning, 175:109301, 2024. 3 +[29] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In CVPR, pages 11936-11945, 2021. 8 +[30] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 3, 6, 7, 8 + +[31] Md Imtiaz Hossain, Sharmen Akhter, Choong Seon Hong, and Eui-Nam Huh. Single teacher, multiple perspectives: Teacher knowledge augmentation for enhanced knowledge distillation. In ICLR, 2025. 2, 6, 7 +[32] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6 +[33] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. NIPS, 35:33716-33727, 2022. 2 +[34] Yaomin Huang, Zaomin Yan, Chaomin Shen, Faming Fang, and Guixu Zhang. Harmonizing knowledge transfer in neural network with unified distillation. In ECCV, pages 58-74. Springer, 2024. 2 +[35] Audun Jøsang. Subjective logic. Springer, 2016. 2, 3 +[36] Jungeun Kim, Junwon You, Dongjin Lee, Ha Young Kim, and Jae-Hun Jung. Do topological characteristics help in knowledge distillation? In ICML, 2024. 6, 7 +[37] Anoop Korattikara Balan, Vivek Rathod, Kevin P Murphy, and Max Welling. Bayesian dark knowledge. NIPS, 28, 2015. 3 +[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5 +[39] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. NIPS, 30, 2017. 2 +[40] Changbin Li, Kangshuo Li, Yuzhe Ou, Lance M Kaplan, Audun Jøsang, Jin-Hee Cho, Dong Hyun Jeong, and Feng Chen. Hyper evidential deep learning to quantify composite classification uncertainty. arXiv preprint arXiv:2404.10980, 2024.3 +[41] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, pages 110-127. Springer, 2022. 8 +[42] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In CVPR, pages 17413-17424, 2023. 8 +[43] Xue Li, Wei Shen, and Denis Charles. Tedl: A two-stage evidential deep learning method for classification uncertainty quantification. arXiv preprint arXiv:2209.05522, 2022. 3 +[44] Zheng Li, Ying Huang, Defang Chen, Tianren Luo, Ning Cai, and Zhigeng Pan. Online knowledge distillation via multi-branch diversity enhancement. In Proceedings of the Asian conference on computer vision, 2020. 2 +[45] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 6, 7 +[46] Jakob Lindqvist, Amanda Olmin, Fredrik Lindsten, and Lennart Svensson. A general framework for ensemble distribution distillation. In 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2020. 3 +[47] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 2 + +[48] Shicai Wei Chunbo Luo Yang Luo. Scale decoupled distillation. arXiv preprint arXiv:2403.13512, 2024. 6, 7 +[49] Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions. NIPS, 34:6881-6893, 2021. 3 +[50] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 3, 4 +[51] Zelda E Mariet, Rodolphe Jenatton, Florian Wenzel, and Dustin Tran. Distilling ensembles improves uncertainty estimates. In Third symposium on advances in approximate bayesian inference, 2021. 3 +[52] Nis Meinert and Alexander Lavin. Multivariate deep evidential regression. arXiv preprint arXiv:2104.06135, 2021. 3 +[53] Zhong Meng, Jinyu Li, Yong Zhao, and Yifan Gong. Conditional teacher-student learning. In ICASSP, pages 6445-6449. IEEE, 2019. 2 +[54] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, pages 5191-5198, 2020. 2 +[55] Deep Shankar Pandey and Qi Yu. Learn to accumulate evidence from all training samples: theory and practice. In ICML, pages 26963-26989. PMLR, 2023. 3 +[56] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, pages 3967-3976, 2019. 2, 6, 7, 8 +[57] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, pages 5007-5016, 2019. 2 +[58] Bo Peng, Zhen Fang, Guangquan Zhang, and Jie Lu. Knowledge distillation with auxiliary variable. In ICML, 2024. 2 +[59] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 6 +[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 5, 6 +[61] Max Ryabinin, Andrey Malinin, and Mark Gales. Scaling ensemble distribution distillation to many classes with proxy targets. NIPS, 34:6023-6035, 2021. 3 +[62] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference. In HPEC, pages 1-9. IEEE, 2023. 1 +[63] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6, 7 +[64] Murat Sensoy, Lance Kaplan, and Melih Kandemir. *Evidential deep learning to quantify classification uncertainty.* NIPS, 31, 2018. 2, 3 + +[65] Glenn Shafer. Dempster-shafer theory. Encyclopedia of artificial intelligence, 1:330-331, 1992. 3 +[66] Yichen Shen, Zhilu Zhang, Mert R Sabuncu, and Lin Sun. Real-time uncertainty estimation in computer vision via uncertainty-aware distribution distillation. In WACV, pages 707-716, 2021. 3 +[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[68] Che Sun, Yunde Jia, and Yuwei Wu. Evidential reasoning for video anomaly detection. In ACMMM, pages 2106-2114, 2022. 3 +[69] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, pages 15731-15740, 2024. 2, 5, 6, 7 +[70] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 2, 6, 7 +[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 8 +[72] Linh Tran, Bastiaan S Veeling, Kevin Roth, Jakub Swiatkowski, Joshua V Dillon, Jasper Snoek, Stephan Mandt, Tim Salimans, Sebastian Nowozin, and Rodolphe Jenatton. Hydra: Preserving ensemble diversity for model distillation. arXiv preprint arXiv:2001.04694, 2020. 3 +[73] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In CVPR, pages 1365-1374, 2019. 2 +[74] Meet Vadera, Brian Jalaian, and Benjamin Marlin. Generalized bayesian posterior expectation distillation for deep neural networks. In Conference on Uncertainty in Artificial Intelligence, pages 719-728. PMLR, 2020. 3 +[75] Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun Zhou, and Edwin Hancock. Uncertainty estimation for stereo matching based on evidential deep learning. pattern recognition, 124:108498, 2022. 3 +[76] Hanjing Wang and Qiang Ji. Beyond dirichlet-based models: when bayesian neural networks meet evidential deep learning. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 3 +[77] Kuan-Chieh Wang, Paul Vicol, James Lucas, Li Gu, Roger Grosse, and Richard Zemel. Adversarial distillation of bayesian neural network posteriors. In ICML, pages 5190-5199. PMLR, 2018. 3 +[78] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, pages 568-578, 2021. 8 +[79] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 8 +[80] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, pages 588-604. Springer, 2020. 2 + +[81] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions. Springer, 2008. 2 +[82] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. ICLR (ICLR), 2021. 2 +[83] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6 +[84] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In CVPR, pages 3517-3526, 2019. 2 +[85] Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Jialu Liu, Michael Bendersky, Marc Najork, and Chao Zhang. Do not blindly imitate the teacher: Using perturbed loss for knowledge distillation. arXiv preprint arXiv:2305.05010, 2023. 2 +[86] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 6 +[87] Yuan Zhang, Weihua Chen, Yichen Lu, Tao Huang, Xiuyu Sun, and Jian Cao. Avatar knowledge distillation: self-ensemble teacher paradigm with uncertainty. In ACMMM, pages 5272-5280, 2023. 3 +[88] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, pages 11953-11962, 2022. 1, 2, 5, 6, 7 +[89] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In CVPR, pages 16965-16974, 2024. 1 \ No newline at end of file diff --git a/2025/Evidential Knowledge Distillation/images.zip b/2025/Evidential Knowledge Distillation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4eb0f89e6586707ccfc25a44dda1e0ecfae5af2f --- /dev/null +++ b/2025/Evidential Knowledge Distillation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8e7077d854b6e89f2e3cc685cd2fa48c983b04383d6bfc734ae18a648961992a +size 636454 diff --git a/2025/Evidential Knowledge Distillation/layout.json b/2025/Evidential Knowledge Distillation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..68fa9108f943237ebaa44cbec5da21ad82813671 --- /dev/null +++ b/2025/Evidential Knowledge Distillation/layout.json @@ -0,0 +1,9211 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 200, + 103, + 411, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 103, + 411, + 121 + ], + "spans": [ + { + "bbox": [ + 200, + 103, + 411, + 121 + ], + "type": "text", + "content": "Evidential Knowledge Distillation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "spans": [ + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": "Liangyu Xiang" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Junyu Gao" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1,2\\dagger}" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "text", + "content": " Changsheng Xu" + }, + { + "bbox": [ + 149, + 141, + 459, + 158 + ], + "type": "inline_equation", + "content": "^{1,2,3}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 118, + 157, + 493, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 118, + 157, + 493, + 185 + ], + "spans": [ + { + "bbox": [ + 118, + 157, + 493, + 185 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 118, + 157, + 493, + 185 + ], + "type": "text", + "content": "State Key Laboratory of Multimodal Artificial Intelligence Systems (MAIS), Institute of Automation, Chinese Academy of Sciences (CASIA)" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 96, + 185, + 515, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 96, + 185, + 515, + 199 + ], + "spans": [ + { + "bbox": [ + 96, + 185, + 515, + 199 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 96, + 185, + 515, + 199 + ], + "type": "text", + "content": "School of Artificial Intelligence, University of Chinese Academy of Sciences (UCAS)" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 200, + 199, + 411, + 213 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 200, + 199, + 411, + 213 + ], + "spans": [ + { + "bbox": [ + 200, + 199, + 411, + 213 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 200, + 199, + 411, + 213 + ], + "type": "text", + "content": "Peng Cheng Laboratory, ShenZhen, China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 146, + 215, + 460, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 215, + 460, + 228 + ], + "spans": [ + { + "bbox": [ + 146, + 215, + 460, + 228 + ], + "type": "text", + "content": "xiangliangyu2023@ia.ac.cn; {junyu.gao, csxu}@nlpr.ia.ac.cn" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "spans": [ + { + "bbox": [ + 151, + 255, + 200, + 268 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 280, + 297, + 568 + ], + "type": "text", + "content": "Existing logit-based knowledge distillation methods typically employ singularly deterministic categorical distributions, which eliminates the inherent uncertainty in network predictions and thereby limiting the effective transfer of knowledge. To address this limitation, we introduce distribution-based probabilistic modeling as a more comprehensive representation of network knowledge. Specifically, we regard the categorical distribution as a random variable and leverage deep neural networks to predict its distribution, representing it as an evidential second-order distribution. Based on the second-oder modeling, we propose Evidential Knowledge Distillation (EKD) which distills both the expectation of the teacher distribution and the distribution itself into the student. The expectation captures the macroscopic characteristics of the distribution, while the distribution itself conveys microscopic information about the classification boundaries. Additionally, we theoretically show that EKD's distillation objective provides an upper bound on the student's expected risk when treating the teacher's predictions as ground truth. Extensive experiments on several standard benchmarks across various teacher-student network pairs highlight the effectiveness and superior performance of EKD. Our code is available at https://github.com/lyxiang-casia/EKD." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 590, + 137, + 602 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 137, + 602 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 137, + 602 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 610, + 296, + 696 + ], + "type": "text", + "content": "As deep neural networks (DNNs) continue to advance, their performance has rapidly improved, yet their increasing network size and training costs pose challenges, especially in resource-constrained and real-time processing scenarios [62, 89]. Consequently, logit-based knowledge distillation (KD) [30, 88], which aims to compress the predictive capabilities of complex models into lightweight ones, has" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 315, + 253, + 556, + 364 + ], + "blocks": [ + { + "bbox": [ + 315, + 253, + 556, + 364 + ], + "lines": [ + { + "bbox": [ + 315, + 253, + 556, + 364 + ], + "spans": [ + { + "bbox": [ + 315, + 253, + 556, + 364 + ], + "type": "image", + "image_path": "e9c82d45add299cc7332d4345c442a3f96c84aec0be55c85a4b295232af45d9d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 315, + 374, + 556, + 486 + ], + "blocks": [ + { + "bbox": [ + 378, + 365, + 490, + 374 + ], + "lines": [ + { + "bbox": [ + 378, + 365, + 490, + 374 + ], + "spans": [ + { + "bbox": [ + 378, + 365, + 490, + 374 + ], + "type": "text", + "content": "(a) Vanilla Knowledge Distillation" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 374, + 556, + 486 + ], + "lines": [ + { + "bbox": [ + 315, + 374, + 556, + 486 + ], + "spans": [ + { + "bbox": [ + 315, + 374, + 556, + 486 + ], + "type": "image", + "image_path": "48d12f289c6be655a348d29dd4f14be951fbedc2e92ea5139810ad62e2ea7dbd.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 373, + 487, + 496, + 498 + ], + "lines": [ + { + "bbox": [ + 373, + 487, + 496, + 498 + ], + "spans": [ + { + "bbox": [ + 373, + 487, + 496, + 498 + ], + "type": "text", + "content": "(b) Evidential Knowledge Distillation" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 499, + 555, + 632 + ], + "lines": [ + { + "bbox": [ + 313, + 499, + 555, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 555, + 632 + ], + "type": "text", + "content": "Figure 1. Illustrations of vanilla KD and our EKD. (a) Vanilla KD deterministically converts the logits into categorical distributions through the softmax function. Each categorical distribution corresponds to a point on the simplex, with each categorical probability representing the barycentric coordinates of that point. Vanilla KD is achieved by minimizing the differences between the teacher's and student's barycentric coordinates. (b) EKD treats the categorical distribution as a random variable following a Dirichlet distribution, parameterized by the logits. The heatmap depicts the probability density of the Dirichlet distribution across the simplex. Moreover, EKD includes distillation of both the expectation of the second-order distribution, and the second-order distribution itself." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "gained increasing popularity. KD transfers knowledge by guiding the predictions of a smaller network to align with those of a larger network during the training." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "Current knowledge distillation methods largely adhere to the vanilla paradigm [30] where only the categorical distri" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 145, + 713 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2814" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 406 + ], + "type": "text", + "content": "butions are aligned. As illustrated in Figure 1a, the logits of a complex network (teacher network) and a lightweight network (student network) are firstly converted into categorical probabilities using a softmax function. Knowledge is then transferred through the process of minimizing the KL divergence between the two categorical distributions. Enhancements to this distillation process, such as decoupling the categorical distributions [88], smoothing the distribution with multiple levels [45], or adopting sample-wise distribution smoothness [69], have demonstrated notable improvements in the student network's predictive performance. Despite these advancements, all such methods are constrained by a fundamental limitation in their probabilistic modeling, which assumes that categorical probabilities are singularly deterministic [12, 14]. Specifically, vanilla KD implicitly assumes that for any given sample, its probability belonging to a particular class is deterministic and can be approximated by a deep neural network. In the case of a three-class classification problem, as shown in Figure 1a, this assumption is akin to identifying a single point on a 2D simplex that best represents the entire simplex. However, in practice, networks trained on finite data with limited capability exhibit inherent uncertainty in their predictions for specific samples [39, 47]. For example, networks with different initial weights may yield different predictions for the same test samples. Consequently, this singularly deterministic probabilistic modeling creates a bottleneck in the effective representation of diverse and uncertain knowledge." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 408, + 294, + 705 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 408, + 294, + 705 + ], + "spans": [ + { + "bbox": [ + 57, + 408, + 294, + 705 + ], + "type": "text", + "content": "To overcome the limitations of singularly deterministic probabilistic modeling, we employ a distribution-based probabilistic modeling, treating the categorical distribution as a random variable governed by a second-order distribution. Following Subjective Logic [35] and Dempster-Shafer Theory of Evidence [81], the Dirichlet distribution, a conjugate prior of the categorical likelihood, instantiates this second-order distribution, facilitating the success of evidential deep learning across multiple domains [5, 16, 25]. However, recent studies employing distribution-based probabilistic modeling are largely confined to macro-level training within a single network, where the second-order distribution is integrated into a single categorical distribution during the network's training process [8, 11, 64]. The second-order distribution inherently provides micro-level information across all possible categorical distributions, offering advantages in cross-network information propagation. Therefore, the second-order distribution serves as a bridge to enable knowledge transfer at both the macro and micro levels, supported by PAC-Bayesain theory [19]. As illustrated in Figure 1b, the logits output by each network are converted into a set of Dirichlet parameters, forming respective second-order distributions. Essentially, the second-order distribution assigns varying weights to all possible first-order categorical distributions," + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 317, + 72, + 553, + 298 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 298 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 298 + ], + "type": "text", + "content": "where the non-uniformity of these weights reflects the network's capacity for classification. Building on second-order predictions, we propose the Evidential Knowledge Distillation (EKD), integrating both macro and micro perspectives. At the macro level, we reduce each second-order distribution to their first-order categorical distributions via an expectation operation in order to ensure that the centroids of their Dirichlet distributions coincide. By representing the entire simplex with a single point, the student is encouraged to focus more on the overall density distribution of the teacher at this stage, which leads to the proportions of student's outputs across categories aligning with those of the teacher. At the micro level, we replace the first-order distribution with the second-order distribution to enable the transfer of more granular classification information. As a complement to proportional alignment, this step refines the magnitude in the student network's outputs. Ultimately, the student model benefits from both the teacher's overall characteristics and its detailed classification structure." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 328, + 300, + 539, + 311 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 300, + 539, + 311 + ], + "spans": [ + { + "bbox": [ + 328, + 300, + 539, + 311 + ], + "type": "text", + "content": "To summarize, Our main contribution are as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 316, + 312, + 553, + 478 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 316, + 312, + 553, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 553, + 370 + ], + "type": "text", + "content": "- We reformulate network predictions in knowledge distillation through distribution-based probabilistic modeling. Building on this reformulation, we propose EKD, which distills the teacher's second-order distribution into the student model across both macro and micro levels." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 316, + 372, + 553, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 418 + ], + "type": "text", + "content": "- With the help of PAC-Bayesian theory [19], we prove that the distillation objective in EKD serves as an upper bound on the expected risk of the student network when treating the teacher network's outputs as ground truth." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 316, + 419, + 553, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 419, + 553, + 478 + ], + "spans": [ + { + "bbox": [ + 316, + 419, + 553, + 478 + ], + "type": "text", + "content": "- Extensive experiments are conducted on several standard benchmarks, encompassing a wide range of teacher and student models. Without bells and whistles, EKD achieves performance improvement of up to " + }, + { + "bbox": [ + 316, + 419, + 553, + 478 + ], + "type": "inline_equation", + "content": "3.61\\%" + }, + { + "bbox": [ + 316, + 419, + 553, + 478 + ], + "type": "text", + "content": " over KD across various teacher-student network combinations." + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 490, + 400, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 490, + 400, + 502 + ], + "spans": [ + { + "bbox": [ + 316, + 490, + 400, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 317, + 510, + 553, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 510, + 553, + 700 + ], + "spans": [ + { + "bbox": [ + 317, + 510, + 553, + 700 + ], + "type": "text", + "content": "Knowledge Distillation Knowledge distillation enhances a simpler student network using insights from a complex teacher network. Existing methods fall into three categories: logit-based [2, 30, 44, 45, 54, 85, 88], feature-based [33, 56, 57, 70, 73], and hybrid approaches [22, 31, 34, 58, 80, 82]. Unlike feature-based methods, which demand architecture-specific designs, logit-based methods offer wider applicability by relying solely on network outputs [4, 53, 84]. Introduced by Hinton and Geoffrey [30], logit-based distillation converts teacher and student logs into categorical distributions aligned via KL divergence. Subsequent efforts have refined this by adjusting the temperature parameter dynamically, based on training stages [45], teacher-student gaps [21], and output variance [69], or by decomposing distributions [88] and leveraging intermediate networks [31, 54]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 328, + 702, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 328, + 702, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 328, + 702, + 553, + 712 + ], + "type": "text", + "content": "The aforementioned traditional distillation methods" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "2815" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 370 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 370 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 370 + ], + "type": "text", + "content": "share a common limitation in their reliance on point estimates for model parameters or categorical distributions, whereas Bayesian distillation [13, 37, 50, 74, 77] employs distributional estimation, which enables the student network to inherit the teacher's strengths in uncertainty estimation. Initial explorations [24, 37, 46, 51] adopt teacher ensembles to approximate the distribution of teacher model parameters and utilize the vanilla KD [30] for distillation. Subsequent extensions continue to advance by improving the student network architecture [20, 50, 61, 72, 76], optimization objectives [74], and sampling methods [13]. Additionally, some approaches introduce the utilization of quantized uncertainty during training to enable more fine-grained knowledge transfer [28, 66, 87]. While these methods introduce uncertainty into the model parameters and the resultant predictive distributions, they remain limited by their reliance on a finite number of ensembles to approximate the distributions. Consequently, we propose to adopt a continuous probabilistic modeling framework for both the teacher and the student, while leveraging more advanced distillation techniques. Unlike Bayesian distillation, which performs uncertainty-related distillation, we implicitly express the uncertainty of the predictive distribution through second-order probabilistic modeling and focus on enhancing the performance of the student network." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 371, + 296, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 296, + 587 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 296, + 587 + ], + "type": "text", + "content": "Evidential Deep Learning Based on Subjective Logic (SL) [35] and Dempster-Shafer Theory (DST) [65], Sensoy et al. [64] initially propose a framework in which evidence collected by a deep neural network for each class is used to parameterize a second-order Dirichlet distribution for network predictions. This concept of evidence collection allows evidential networks to express \"I don't know\", showing advantages in uncertainty estimation. Since then, numerous studies have advanced and expanded upon Evidential Deep Learning (EDL) from both theoretical and application-oriented perspectives [17]. Theoretical investigations have primarily focused on refining the evidence collection process [11, 15, 18, 40, 55], delving into different training strategies [5, 26, 43] and extending evidential theory to regression networks [1, 49, 52]. Moreover, EDL has been widely applied in computer vision tasks, including action localization [5-7, 16], stereo matching [75], anomaly detection [68], and multi-view classification [25]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 597, + 111, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 111, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 111, + 609 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 296, + 713 + ], + "type": "text", + "content": "In this section, we first present the fundamental framework of our Evidential Knowledge Distillation (EKD), which includes both macroscopic first-order and microscopic second-order distribution distillation. Theoretically, we utilize PAC-Bayesian theory to demonstrate that the optimization objective of EKD effectively represents an upper bound on the student's expected risk. At the end of this section, we illustrate the complementary effect of the two" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 72, + 500, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 500, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 500, + 84 + ], + "type": "text", + "content": "distillation strategies with a concrete toy case." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 93, + 400, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 93, + 400, + 105 + ], + "spans": [ + { + "bbox": [ + 314, + 93, + 400, + 105 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": "Evidential Deep Learning. In accordance with the formulation of evidential deep learning [8, 9, 64], for a classification task involving " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " classes, the categorical probability vector " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " is treated as a random variable governed by a conjugate prior Dirichlet distribution. To define this distribution's parameters, an evidential activation function " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\sigma" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " which we designate as the exponential function (exp), is employed to transform the network's output logits " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " into a non-negative evidence vector " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "e = \\sigma(z)" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": ". This evidence vector is subsequently combined with a prior weight " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " to yield the " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": "-dimensional Dirichlet parameters " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " which fully specify a Dirichlet distribution " + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "inline_equation", + "content": "Dir(\\alpha)" + }, + { + "bbox": [ + 313, + 110, + 555, + 289 + ], + "type": "text", + "content": " [8]. In terms of optimization, the widely adopted cross-entropy loss is computed by integrating over the categorical probabilities, guided by second-order the Dirichlet distribution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 336, + 291, + 554, + 357 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 336, + 291, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 336, + 291, + 554, + 357 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {E D L - c e} = \\int \\left[ \\sum_ {i = 1} ^ {K} - y _ {i} \\log \\left(p _ {i}\\right) \\right] D i r (\\boldsymbol {p} | \\boldsymbol {\\alpha}) d \\boldsymbol {p} \\\\ = \\sum_ {i = 1} ^ {K} y _ {i} (\\psi (\\alpha_ {0}) - \\psi (\\alpha_ {i})) \\tag {1} \\\\ \\end{array}", + "image_path": "4c196fc81bbf5025f46770375d4d38bef0cfe927ce5135983a7002674b7da67b.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "inline_equation", + "content": "y_{i}" + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "text", + "content": " denotes the ground truth, " + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\alpha_0 = \\sum_{i=1}^{K} (\\alpha_i)" + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "text", + "content": " represents the Drichlet strength, and " + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "inline_equation", + "content": "\\psi" + }, + { + "bbox": [ + 313, + 360, + 554, + 445 + ], + "type": "text", + "content": " refers to the digamma function. To support the acquisition of evidence, we employ Eq. (1), rather than the conventional cross-entropy loss, across all network training stages, including both the training of the teacher networks and student networks during the knowledge distillation process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": "Knowledge Distillation. Existing logit-based knowledge distillation methods rely on categorical distribution predictions derived from the softmax function. Specifically, for a given input sample " + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": ", the network produces logits " + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": ", which are then transformed into a categorical distribution via the softmax operation, denoted as " + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "inline_equation", + "content": "p = \\mathrm{softmax}(z)" + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": ". Knowledge transfer is accomplished by minimizing the Kullback-Leibler divergence(KL) divergence between the teacher's and student's predicted distributions, " + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "inline_equation", + "content": "p^T" + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "inline_equation", + "content": "p^S" + }, + { + "bbox": [ + 313, + 445, + 554, + 552 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 389, + 552, + 553, + 567 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 552, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 389, + 552, + 553, + 567 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {K D} = K L \\left(\\boldsymbol {p} ^ {T} \\| \\boldsymbol {p} ^ {S}\\right) \\tag {2}", + "image_path": "a0a99064538699a562fa1fe95b21ccad57c89eaa13294c8480573ef2228827ce.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 553, + 591 + ], + "type": "text", + "content": "The teacher's predictions serve as soft labels, replacing the traditional one-hot labels, capturing inter-class similarities." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 600, + 495, + 613 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 600, + 495, + 613 + ], + "spans": [ + { + "bbox": [ + 314, + 600, + 495, + 613 + ], + "type": "text", + "content": "3.2. Evidential Knowledge Distillation" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 713 + ], + "type": "text", + "content": "Previous knowledge distillation methods have generally followed a unified paradigm, where the network's output logits are transformed into categorical probabilities via the softmax function, and the student network is guided to align its categorical distribution with that of the teacher network. While these approaches have yielded notable success, a fundamental limitation persists: they treat categorical probabilities as singularly deterministic values. This assumption" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2816" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "overlooks the uncertainty in network predictions, thus constraining the sharing of more granular information. As a more robust alternative, we predict the distribution of categorical distributions using neural networks, specifically modeling it as a second-order Dirichlet distribution. More importantly, we align both the first-order distribution, obtained by integrating the second-order distribution, and the second-order distribution itself." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "text", + "content": "Firstly, we aim to extract the overall characteristics of the second-order distribution to capture the general distribution trend of probability density across the simplex, as well as to achieve alignment between networks in terms of final predicted classes. Specifically, given the second-order distributions predicted by the teacher and student, " + }, + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "inline_equation", + "content": "Dir(\\pmb{\\alpha}^T)" + }, + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "inline_equation", + "content": "Dir(\\pmb{\\alpha}^S)" + }, + { + "bbox": [ + 55, + 167, + 294, + 275 + ], + "type": "text", + "content": ", we compute their expectations where all possible categorical distributions are summed under the weights specified by the second-order distribution:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 275, + 295, + 291 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 275, + 295, + 291 + ], + "spans": [ + { + "bbox": [ + 67, + 275, + 295, + 291 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} ^ {T} = \\mathbb {E} _ {\\boldsymbol {p} ^ {T} \\sim D i r (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\hat {\\boldsymbol {p}} ^ {S} = \\mathbb {E} _ {D i r (\\boldsymbol {p} ^ {S} \\sim \\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ] \\tag {3}", + "image_path": "75fc794a653452293b522c6d41be20d12200437668873d9a4e811be1aa9e48ba.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 295, + 350 + ], + "type": "text", + "content": "The expectation above, serving as the final predictive outcome for both teacher and student, is equivalent to the centroid of the second-order distribution, enabling the transfer of comprehensive probability density information via first-order distillation at a macro level." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 350, + 295, + 381 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 350, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 85, + 350, + 295, + 381 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {1 s t} = K L (\\hat {\\boldsymbol {p}} ^ {T} | | \\hat {\\boldsymbol {p}} ^ {S}) = \\sum_ {i = 1} ^ {K} \\frac {\\alpha_ {i} ^ {T}}{\\alpha_ {0} ^ {T}} l o g \\left(\\frac {\\alpha_ {i} ^ {T} \\alpha_ {0} ^ {S}}{\\alpha_ {0} ^ {T} \\alpha_ {i} ^ {S}}\\right) \\tag {4}", + "image_path": "41589a8d13124bb3089d7e4323307a4d4db37a869b4ba4a521a92da1c658317c.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "text", + "content": "Formally, optimizing Eq. (4) involves adjusting the proportion of each class's Dirichlet parameter relative to the total, which significantly enhances the alignment between the student's and teacher's predicted classes. In terms of formulation, Eq. (4) is very similar to vanilla KD, yet it is emphasized that vanilla KD can be regarded as a special case of first-order distillation when " + }, + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\sigma = \\exp(\\cdot)" + }, + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "inline_equation", + "content": "\\lambda = 0" + }, + { + "bbox": [ + 55, + 381, + 295, + 476 + ], + "type": "text", + "content": ". Proof is provided in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 477, + 295, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 477, + 295, + 620 + ], + "spans": [ + { + "bbox": [ + 55, + 477, + 295, + 620 + ], + "type": "text", + "content": "The first-order distillation objective above only provides macroscopic guidance for the student's second-order distribution, failing to accurately and comprehensively represent the weighting of each categorical distribution. In other words, given a fixed teacher distribution, the student Dirichlet distribution that minimizes Eq. (4) is not unique. Therefore, we propose a microscopic second-order distillation as a complement to first-order distillation, aiming to provide fine-grained teacher supervision signals across the entire simplex for the student distribution. Specifically, we achieve this by applying the KL divergence directly between the two Dirichlet distributions:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 89, + 620, + 295, + 706 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 620, + 295, + 706 + ], + "spans": [ + { + "bbox": [ + 89, + 620, + 295, + 706 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {2 n d} = K L (D i r (\\boldsymbol {p} ^ {T} | \\boldsymbol {\\alpha} ^ {T}) | | D i r (\\boldsymbol {p} ^ {S} | \\boldsymbol {\\alpha} ^ {S})) \\\\ = l o g \\frac {\\Gamma (\\alpha_ {0} ^ {T})}{\\Gamma (\\alpha_ {0} ^ {S})} - \\sum_ {i = 1} ^ {K} l o g \\frac {\\Gamma (\\alpha_ {i} ^ {T})}{\\Gamma (\\alpha_ {i} ^ {S})} \\\\ + \\sum_ {i = 1} ^ {K} \\left(\\alpha_ {i} ^ {T} - \\alpha_ {i} ^ {S}\\right) \\left(\\psi \\left(\\alpha_ {i} ^ {T}\\right) - \\psi \\left(\\alpha_ {0} ^ {T}\\right)\\right) \\tag {5} \\\\ \\end{array}", + "image_path": "8510bc7ed2dc858824784e62eb820788e6d1c35868dd9963548062269808c4e1.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\Gamma(\\cdot)" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " denotes the Gamma function. The detailed derivation for Eq. (5) is provided in Supplementary Material. Since Eq. (5) considers the whole " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " dimensional simplex, overflow may occur [50]; thus, we apply the " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "\\log(1 + x)" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": " function to smooth the network outputs. While the above Eq. (4) aims to optimize the proportional relationships between classes, Eq. (5) here places greater emphasis on aligning the magnitudes within each class. Here, the alignment within the class means that the gradients of a loss function with respect to the output for a particular class do not conflict with the gradients of the output for other classes. This meticulous numerical alignment supplies each point on the " + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 313, + 72, + 553, + 300 + ], + "type": "text", + "content": "-dimensional simplex of the student's second-order distribution with the corresponding teacher label, ensuring a precise mapping of knowledge across the entire distribution space. By doing so, it strives to faithfully replicate the teacher's second-order predictive nuances at a granular, micro-level, capturing the fine-grained patterns and uncertainties inherent in the teacher model's outputs." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 300, + 553, + 346 + ], + "type": "text", + "content": "By combining these two complementary optimization objectives, one optimizing proportional relationships at the macro level and the other aligning numerical values at the micro level, we arrive at the final EKD distillation:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 382, + 349, + 553, + 361 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 349, + 553, + 361 + ], + "spans": [ + { + "bbox": [ + 382, + 349, + 553, + 361 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {E K D} = \\mathcal {L} _ {1 s t} + \\gamma \\mathcal {L} _ {2 n d}, \\tag {6}", + "image_path": "1dba7a887a8a660449a218f911ff2051f478aff02b74d3c117256f696e7cd4c4.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 313, + 363, + 553, + 385 + ], + "type": "text", + "content": " represents the coefficient that adjusts the balance between the first-order and second-order distillation." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 399, + 432, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 399, + 432, + 412 + ], + "spans": [ + { + "bbox": [ + 314, + 399, + 432, + 412 + ], + "type": "text", + "content": "3.3. Theoretical Analysis" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 417, + 553, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 417, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 417, + 553, + 513 + ], + "type": "text", + "content": "In this section, we provide theoretical support for Eq. (6) using PAC-Bayesian Theory [19]. While PAC-Bayesian theory has been explored in the context of the Dirichlet distribution [26], its application remains confined to single-network optimization, neglecting the relationship among multiple networks. Therefore, we aim to estimate the alignment of different networks across the entire data distribution based on their alignment on training samples." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": " denote a certain sample and " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": " denote its corresponding label. Under the PAC setting, samples are drawn from a fixed but unknown distribution " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": ". A classifier " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": " maps " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": ". Moreover, we define a classifier " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "h" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": " correspond to a point on the " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": "-dimensional simplex, i.e., a categorical distribution " + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 514, + 554, + 632 + ], + "type": "text", + "content": ". Essentially, the Dirichlet distribution made by each network serves as the posterior distribution of a set of classifiers. By weighting the set of classifiers with its posterior distribution, we obtain the final majority vote classifier (or Bayesain classifier):" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 389, + 632, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 632, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 389, + 632, + 553, + 647 + ], + "type": "interline_equation", + "content": "\\boldsymbol {B} ^ {T} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {T})} [ \\boldsymbol {p} ^ {T} ], \\tag {7}", + "image_path": "86e96701b2f04411724e713a79bf597f6e526fa9461d8bbd8892393379123fbb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 649, + 553, + 664 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 649, + 553, + 664 + ], + "spans": [ + { + "bbox": [ + 389, + 649, + 553, + 664 + ], + "type": "interline_equation", + "content": "\\boldsymbol {B} ^ {S} = \\mathbb {E} _ {\\operatorname {D i r} (\\boldsymbol {\\alpha} ^ {S})} [ \\boldsymbol {p} ^ {S} ], \\tag {8}", + "image_path": "ed4957676e6a14918c45e1e5b5c7821bf86e6caadbc4cfa14c7695fb03c68ba4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{B}^{T}" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\pmb{B}^{S}" + }, + { + "bbox": [ + 313, + 665, + 553, + 689 + ], + "type": "text", + "content": " denote the final classification results of the teacher and student classifiers, respectively." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 713 + ], + "type": "text", + "content": "In knowledge distillation, the student treats the teacher's predictions as soft labels. Below, we define the student's" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2817" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "misclassification risk using cross-entropy, which is equivalent to KL divergence from an optimization perspective." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 111, + 95, + 295, + 111 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 95, + 295, + 111 + ], + "spans": [ + { + "bbox": [ + 111, + 95, + 295, + 111 + ], + "type": "interline_equation", + "content": "R ^ {S} = \\mathbb {E} _ {x \\sim \\mathcal {D}} \\left[ - \\boldsymbol {B} ^ {T} \\log \\boldsymbol {B} ^ {S} \\right], \\tag {9}", + "image_path": "0f1b0ee5e6c4277481d5e8fb27e9d21b2365e4d4bf44e9135cced3e09bab9735.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 111, + 114, + 295, + 145 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 114, + 295, + 145 + ], + "spans": [ + { + "bbox": [ + 111, + 114, + 295, + 145 + ], + "type": "interline_equation", + "content": "\\hat {R} ^ {S} = \\frac {1}{N} \\sum_ {i = 1} ^ {N} \\left[ - \\boldsymbol {B} _ {i} ^ {T} \\log \\boldsymbol {B} _ {i} ^ {S} \\right]. \\tag {10}", + "image_path": "4bba2a5a05a6e748c513839353831a68b5c70bfc166818c7c8392d4e0d41ff9a.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "inline_equation", + "content": "R^S" + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": " represents the expected risk of the student over the entire data distribution " + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\hat{R}^S" + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": " denotes the empirical risk on the training set " + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t" + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": " containing " + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 147, + 296, + 183 + ], + "type": "text", + "content": " training samples." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": "Theorem 1. For any " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\delta \\in (0,1]" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\gamma \\in \\mathbb{R}^{+}" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "R^S" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": " is bounded above by the sum of " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\hat{R}^S" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\sum_{1}^{N}KL(Dir(\\pmb{\\alpha}_i^T))||Dir(\\pmb{\\alpha}_i^S))" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": ", and a constant related to " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": ", denoted as " + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "inline_equation", + "content": "C(\\delta)" + }, + { + "bbox": [ + 55, + 189, + 296, + 227 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 228, + 294, + 287 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 294, + 287 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} P _ {\\mathcal {D} _ {t} \\sim \\mathcal {D} ^ {N}} (R ^ {S} \\leq \\hat {R} ^ {S} + \\frac {\\gamma}{N} \\sum_ {i = 1} ^ {N} K L (D i r (\\pmb {\\alpha} _ {i} ^ {T}) | | D i r (\\pmb {\\alpha} _ {i} ^ {S})) \\\\ + C (\\delta)) \\geq 1 - \\delta . \\tag {11} \\\\ \\end{array}", + "image_path": "6e221830c88058d30f244ab87a801737be738c54a1e2c0bb7dbe4bbc3d9abf4e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "spans": [ + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "text", + "content": "Detailed proof and definition of " + }, + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "inline_equation", + "content": "C(\\delta)" + }, + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "text", + "content": " can be found in Supplementary Material. Theorem 1 indicates that the upper bound on the student's excepted risk over the unknown data distribution " + }, + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 54, + 294, + 296, + 498 + ], + "type": "text", + "content": " consists of two components: the alignment of the first-order distribution with the true labels (teacher labels) and the similarity between the predicted second-order distribution and the true second-order distribution (teacher distribution). Clearly, this upper bound indicates that the empirical risk optimization represented by Eq. (2) does not guarantee a reduction in the student's expected risk, suggesting that the transfer of the teacher's generalization performance is both inefficient and incomplete. In contrast, EKD directly adopts this upper bound as optimization objectives, refining the formulation of empirical risk while incorporating second-order distribution alignment between the student and teacher, thereby striving to minimize the student's expected risk." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 505, + 121, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 505, + 121, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 505, + 121, + 517 + ], + "type": "text", + "content": "3.4. Toy Case" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": "Figure 2 demonstrates the complementary roles of first-order and second-order distillation through a three-class example. Initially, the student Dirichlet parameters differ significantly from those of the teacher, both in relative proportion and in absolute values. The left side of Figure 2 illustrates the essence of first-order distillation, which minimizes the distance between the expectations of two Dirichlet distributions over the " + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "inline_equation", + "content": "(K - 1)" + }, + { + "bbox": [ + 55, + 522, + 296, + 713 + ], + "type": "text", + "content": " dimensional simplex measured by KL divergence. However, the expectation of the Dirichlet distribution depends solely on the proportion of each class's Dirichlet parameter relative to the overall sum, which can lead to discrepancies in the specific values of the student's output compared to the teacher's. Although this discrepancy does not affect the final class prediction, it causes the student's second-order distribution to differ significantly from the teacher's. To address this, we use" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 554, + 296 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 554, + 296 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 554, + 296 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 554, + 296 + ], + "type": "image", + "image_path": "648f6e6a352fa9bba6eba19e4d2833bd02ba36a33432a16289f374833fcf0545.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 303, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 313, + 303, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 555, + 392 + ], + "type": "text", + "content": "Figure 2. A toy case illustrating the complementary roles of the two distillation losses in EKD. Given the Dirichlet parameters output by teacher and student networks, the left panel shows the results when only first-order distillation is used to optimize the proportion of each category. The right panel presents the outcome when only second-order distillation is applied for intra-class numerical alignment. The center panel presents the results of combining both objectives, which yields the optimal performance." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 555, + 536 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 536 + ], + "type": "text", + "content": "second-order distillation as a complement to first-order distillation, aiming to align the Dirichlet parameter values. In contrast to first-order distillation, second-order distribution alignment encourages consistency of parameters within the same class across different networks, disregarding the relative magnitudes among different classes. As a result, we combine these complementary distillation objectives to achieve a comprehensive alignment of the second-order distribution. The quantitative effects of the two types of distillation will be discussed in detail in Section 4.2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 550, + 395, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 395, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 395, + 563 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": "Datasets. Our experiments are primarily conducted on two standard datasets including CIFAR100 [38] and ImageNet [60]. The CIFAR-100 dataset [38] consists of 60k 32x32 color images divided into 100 classes, with each class containing 600 images. It is split into a training set of 50k images and a validation set of 10k images. ImageNet [60] is a large-scale visual database comprising 1,000 categories that cover a diverse range of objects and scenes. It includes approximately 1.28 million labeled images for training and an additional 50,000 images for validation." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Implementation Details. We adopt the experimental settings established in prior studies [10, 69, 88]. For the" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2818" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 116, + 563, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 1. Results on the CIFAR-100 validation set. The teacher and student networks share the same architecture. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 116, + 563, + 370 + ], + "lines": [ + { + "bbox": [ + 56, + 116, + 563, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 563, + 370 + ], + "type": "table", + "html": "
Experiment Group1234567
TeacherArchitectureResNet32×4VGG13WRN-40-2WRN-40-2ResNet56ResNet110ResNet110
Softmax79.4274.6475.6175.6172.3474.3174.31
EDL79.5374.9675.4175.4173.0974.6774.67
StudentArchitectureResNet8×4VGG8WRN-40-1WRN-16-2ResNet20ResNet32ResNet20
Softmax72.5070.3671.9873.2669.0671.1469.06
EDL72.7770.6771.8473.3669.2270.969.22
FeatureRKD[CVPR'19] [56]71.9071.4872.2273.3569.6171.8269.25
CRD[ICLR'20] [70]75.5173.9474.1475.4871.1673.4871.46
ReviewKD[CVPR'21] [10]75.6374.8475.0976.1271.8973.8971.34
SimKD[CVPR'22] [3]78.0874.8974.5375.5371.0573.9271.06
CAT-KD[CVPR'23] [23]76.9174.6574.8275.6071.6273.6271.37
TopKD[ICML'24] [36]75.4074.0174.4375.7571.5873.7771.47
LogitKD[NeurIPS'14] [30]73.3372.9873.5474.9270.6673.0870.67
Δ0.832.621.561.661.61.941.61
DKD[CVPR'22] [88]76.3274.6874.8176.2471.9774.1171.06
Δ3.824.322.832.982.912.972.0
CTKD[AAAI'23] [45]73.3973.5273.9375.4571.1973.5270.99
Δ0.893.161.952.192.132.381.93
Logit_Stand[CVPR'24] [69]76.6274.3674.3776.1171.4374.1771.48
Δ4.124.002.392.852.373.032.42
SDD[CVPR'24] [48]75.92*73.90*74.26*75.98*70.72*73.90*71.43*
Δ3.42*3.54*2.28*2.72*1.66*2.76*2.37*
TeKAP[ICLR'25] [31]74.7973.8*73.8075.2171.7173.50*71.00*
Δ2.293.44*1.821.952.652.36*1.94*
EKD(Ours)77.2174.7174.4376.1571.4873.6871.51
Δ4.444.042.592.792.262.782.29
", + "image_path": "7c6ebc9551dddf7a941ac0cd60408c70d8cc2bf2512deac321d9f10e15c670b6.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 422, + 291, + 594 + ], + "blocks": [ + { + "bbox": [ + 55, + 388, + 295, + 420 + ], + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 420 + ], + "type": "text", + "content": "Table 3. The top-1 and top-5 accuracy (\\%) on the ImageNet validation set [60]. The best and second best results are emphasized in bold and underlined. “*” indicates our reproduced results." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 422, + 291, + 594 + ], + "lines": [ + { + "bbox": [ + 61, + 422, + 291, + 594 + ], + "spans": [ + { + "bbox": [ + 61, + 422, + 291, + 594 + ], + "type": "table", + "html": "
TeacherArchitectureResNet34ResNet50
Accuracytop-1top-5top-1top-5
Softmax73.3191.4276.1692.86
EDL73.8391.6976.1092.98
StudentArchitectureResNet18MNV1
Accuracytop-1top-5top-1top-5
Softmax69.7589.0768.8788.76
EDL70.2889.5869.9289.46
FeatureCRD [70]71.1790.1371.3790.41
ReviewKD [10]71.6190.5172.5691.00
SimKD [3]71.5990.4872.2590.86
CAT-KD [23]71.2690.4572.2491.13
LogitKD [30]71.0390.0570.5089.80
DKD [88]71.7090.4172.0591.05
CTKD [45]71.3890.2771.1690.11
Logit_Stand [69]71.4290.2972.1890.80
SDD [48]71.4490.0572.15*90.79*
EKD(Ours)71.7390.6972.5491.20
", + "image_path": "8462bab17d4b7f13be597859e7b8ee18fea8c62db8761bc706e94543e9440645.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "experiments on CIFAR-100, we use the SGD optimizer with 240 epochs. The learning rate is set to 0.01 for MobileNets [32, 63] and ShuffleNets [86] and 0.05 for other networks, including ResNets [27], WRNs [83] and VGGs [67], with a decay factor of 0.1 applied at epochs 150, 180, and 210. Regarding hyperparameter configuration, we set " + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\gamma" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": " in Eq. (6) to 1 by default, requiring no additional hyperparameter tuning. All results represent the mean of five" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 389, + 555, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 413 + ], + "type": "text", + "content": "independent runs. Additional experimental details are provided in the Supplementary Material." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 413, + 555, + 497 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 497 + ], + "type": "text", + "content": "Baselines. To benchmark EKD, we incorporate five state-of-the-art logit-based distillation methods: DKD (CVPR'22) [88], CTKD (AAAI'23) [45], Logit_Stand (CVPR'24) [69], SDD (CVPR'24) [48] and TeKAP (ICLR'25) [31]. Additionally, a selection of feature-based distillation methods [3, 10, 23, 56, 59, 70] is included as references to facilitate a broader contextual analysis." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 505, + 400, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 400, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 400, + 517 + ], + "type": "text", + "content": "4.1. Main Results" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": "Performance comparison on CIFAR100. We conduct experiments comparing EKD with other distillation methods under various teacher-student pair settings. Table 1 and Table 2 present the accuracy results for settings where the teacher and student have identical and distinct architectures. As presented in Table 1, EKD demonstrates favourable performance improvements over KD, ranging from " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "0.66\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.61\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ". In particular, EKD demonstrates its strongest performance in Groups 1, 2, and 7, while achieving comparable results to DKD [88] in Groups 3-6. The underlying reason for this phenomenon lies in the performance gap between the teacher and student. In groups 1, 2, and 7, the average performance gap between the teacher and student is " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "5.51\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": ", whereas it is " + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 313, + 521, + 556, + 715 + ], + "type": "text", + "content": " in groups 3-6. Furthermore, we observe that pairs with a larger performance gap tend to have a higher distillation loss, indicating a greater opti" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2819" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 116, + 568, + 370 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 2. Results on the CIFAR-100 validation set. The teacher and student networks have different architectures. The best and second-best results among logits-based methods are highlighted in bold and underlined, respectively. " + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "inline_equation", + "content": "\\Delta" + }, + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": " denotes the improvement of the distilled student model compared to the independently trained student model. Note that \"Softmax\" indicates that the model uses the vanilla KD probabilistic model, while \"EDL\" indicates that the model are trained by Eq. (1). \\*\\* indicates our reproduced results." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 116, + 568, + 370 + ], + "lines": [ + { + "bbox": [ + 56, + 116, + 568, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 116, + 568, + 370 + ], + "type": "table", + "html": "
Experiment Group1234567
TeacherArchitectureResNet32×4ResNet32×4ResNet32×4WRN-40-2WRN-40-2VGG13ResNet50
Softmax79.4279.4279.4275.6175.6174.6479.34
EDL79.5379.5379.5375.4175.4175.1179.69
StudentArchitectureSHN-V2WRN-16-2WRN-40-2ResNet8×4MN-V2MN-V2MN-V2
Softmax71.8273.2675.6172.5064.6064.6064.60
EDL72.7373.3675.4172.7765.3965.3965.39
FeatureRKD[CVPR'19] [56]73.2174.8677.8275.2669.2764.5264.43
CRD[ICLR'20] [70]75.6575.6578.1575.2470.2869.7369.11
ReviewKD[CVPR'21] [10]77.7876.1178.9674.3471.2870.3769.89
SimKD[CVPR'22] [3]78.3977.1779.2975.2970.1069.4469.97
CAT-KD[CVPR'23] [23]78.4176.9778.5975.3870.2469.1371.36
TopKD[ICML'24] [36]76.3374.34*76.18*75.83*69.54*69.32*70.11*
LogitKD[NeurIPS'14] [30]74.4574.9077.7073.9768.3667.3767.35
Δ2.631.642.091.473.762.772.75
DKD[CVPR'22] [88]77.0775.7078.4675.5669.2869.7170.35
Δ5.252.442.853.064.685.115.75
CTKD[AAAI'23] [45]75.3774.5777.6674.6168.3468.5068.67
Δ3.551.312.052.113.743.904.07
Logit_Stand[CVPR'24] [69]75.5675.2677.9277.1169.2368.6169.02
Δ3.742.002.314.614.634.014.42
SDD[CVPR'24] [48]76.88*75.92*78.54*74.83*70.14*69.06*69.31*
Δ5.06*2.66*2.93*2.33*5.54*4.46*4.71*
TeKAP[ICLR'25] [31]75.4375.26*76.93*74.99*68.65*68.27*68.75*
Δ3.612.00*1.32*2.49*4.05*3.67*4.15*
EKD(Ours)77.6576.5178.9176.7569.2069.9470.38
Δ4.923.153.503.983.814.554.99
", + "image_path": "b3ed2da6ce7d5eccf2c40b89df48afd6bc646c926e55c70264322d24c9551638.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 82, + 445, + 272, + 548 + ], + "blocks": [ + { + "bbox": [ + 55, + 387, + 295, + 443 + ], + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 443 + ], + "type": "text", + "content": "Table 4. Ablation study for each components in EKD. The \"MSE\" refers to aligning the magnitude of model outputs with mean square error. The \"Accuracy\" denotes the Top-1 accuracy (\\%) on the CIFAR100 validation set. ResNet32×4 and ResNet8×4 are adopted as teacher and student." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 445, + 272, + 548 + ], + "lines": [ + { + "bbox": [ + 82, + 445, + 272, + 548 + ], + "spans": [ + { + "bbox": [ + 82, + 445, + 272, + 548 + ], + "type": "table", + "html": "
IndexDistillation ObjectiveAccuracy
\\( \\mathcal{L}_{KD} \\)\\( \\mathcal{L}_{1st} \\)\\( \\mathcal{L}_{2nd} \\)MSE
1✓74.45
2✓73.80
3✓76.42
4✓76.10
5✓✓77.21
6✓✓76.83
7✓✓76.34
8✓✓76.13
", + "image_path": "cc3aca791d747bf599cef20da1782b80841666fd194c4aba5930849571f89913.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": "mization space during distillation. Conversely, for teacher-student pairs with closer performance, their distillation loss is generally smaller, leading EKD to encounter local minima during optimization. For the distillation of distinct networks shown in Table 2, EKD achieves the highest performance in half of the experimental groups. When using MobileNet [63] as the student network, the student trained solely with EDL achieves a " + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "inline_equation", + "content": "0.79\\%" + }, + { + "bbox": [ + 55, + 570, + 296, + 713 + ], + "type": "text", + "content": " higher performance than the one trained with softmax alone. However, the result for EKD is only comparable to that of other methods. We attribute this phenomenon to the inherent performance ceiling of the student network due to its limited capacity." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 313, + 389, + 555, + 485 + ], + "type": "text", + "content": "Performance comparison on ImageNet. Table 3 presents the performance of EKD on the large-scale ImageNet dataset. EKD markedly outperforms other logit-based methods in the ResNet50-MNV1 experiment and achieves results comparable to top methods in the ResNet34-ResNet18 experiment. The modest gains in the latter can also be attributed to the narrow performance gap between the teacher and student." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 493, + 416, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 416, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 416, + 504 + ], + "type": "text", + "content": "4.2. Ablation Studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "text", + "content": "To elucidate the superiority of the two alignment strategies implemented in EKD and their alignment with theoretical underpinnings, the ablation study presented in Table 4 evaluates various distillation objectives on student model performance. In addressing alignment with respect to the proportional distribution across output classes, we investigate two methodologies: the conventional vanilla KD approach presented in Eq. (2), which aligns softmax-based categorical distributions, and " + }, + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1st}" + }, + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "text", + "content": " in Eq. (4). Comparative analyses between Index 1 and Index 2, Index 5 and Index 7, and Index 6 and Index 8 consistently demonstrate that " + }, + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{1st}" + }, + { + "bbox": [ + 313, + 510, + 555, + 700 + ], + "type": "text", + "content": " surpasses vanilla KD. This finding underscores that EKD's first-order alignment serves as a robust generalization of vanilla KD, effectively incorporating the alignment of categorical distributions by leveraging the centroids of second-order Dirichlet distributions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": "In addition to the second-order distillation, " + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 325, + 701, + 555, + 714 + ], + "type": "text", + "content": ", shown" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2820" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 95, + 288, + 194 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": "Table 5. The Top-1 accuracy " + }, + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "inline_equation", + "content": "(\\%)" + }, + { + "bbox": [ + 55, + 71, + 295, + 92 + ], + "type": "text", + "content": " of combining EKD with feature distillation-based methods on CIFAR-100." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 65, + 95, + 288, + 194 + ], + "lines": [ + { + "bbox": [ + 65, + 95, + 288, + 194 + ], + "spans": [ + { + "bbox": [ + 65, + 95, + 288, + 194 + ], + "type": "table", + "html": "
TeacherArchitectureVGG13ResNet32x4
Accuracy74.9679.53
StudentArchitectureVGG8SHN-V2
Accuracy70.6772.73
LogitEKD74.7177.65
FeatureRKD [56]71.4873.21
CAT-KD [23]74.6578.41
Logit+FeatureEKD+RKD [56]74.99↑78.08↑
EKD+CAT-KD [23]74.91↑78.46↑
", + "image_path": "ee2b7413342f96963665e650abd61aa82ca0a0a2399e4dd9d95c3511450cbf28.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "text", + "content": "in Eq. (5), another intuitive approach to aligning model output magnitudes is through mean squared error (MSE). Consequently, we compare " + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "text", + "content": " and MSE in Table 4. Comparisons across Index 3 and Index 4, Index 5 and Index 6, and Index 7 and Index 8 reveal that " + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "text", + "content": " consistently outperforms MSE. These empirical results corroborate the theoretical conclusions delineated in Section 3.3. It is suggested that " + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{2nd}" + }, + { + "bbox": [ + 55, + 218, + 296, + 339 + ], + "type": "text", + "content": ", being part of the upper bound on the student's expected risk, directly optimizes the student's generalization performance, whereas the MSE objective does so indirectly." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 349, + 162, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 349, + 162, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 349, + 162, + 361 + ], + "type": "text", + "content": "4.3. Further Remarks" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "text", + "content": "Visualizations. Figure 3 presents the t-SNE visualizations of the laerned feature representations and, in parallel, the differences in logits between the teacher and student networks, where the teacher is implemented as ResNet " + }, + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "inline_equation", + "content": "32 \\times 4" + }, + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "text", + "content": " and the student as ResNet " + }, + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "inline_equation", + "content": "8 \\times 4" + }, + { + "bbox": [ + 55, + 366, + 296, + 509 + ], + "type": "text", + "content": ". As shown in the first row, EKD achieves a clear improvement over vanilla KD baseline with respect to class separability, leading to more distinctly clustered feature distributions. Furthermore, the logits produced under EKD exhibit a noticeably stronger alignment with those of the teacher network, indicating a closer match in predictive behavior compared to the results obtained with vanilla KD." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 296, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 617 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 617 + ], + "type": "text", + "content": "Compatibility with feature-based methods. Table 5 presents the results of integrating EKD with several representative feature-based distillation methods [23, 56]. For feature-based approaches, the introduction of EKD enhances their performance, demonstrating strong compatibility, particularly with methods that originally exhibit lower performance [56]. Conversely, these methods can further compensate for the limitation of the EKD approach in expressing knowledge through logits." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 713 + ], + "type": "text", + "content": "Distilling ViT. Transformer-based models are gaining increasing popularity, and logit-based distillation methods can be easily applied to these models. Therefore, we test the distillation effectiveness of EKD on the ViT model, as presented in Table 6. Following the configuration in [41, 42], we use ResNet56 as the teacher model and four ViT variants, DeiT-Ti [71], PiT-Ti [29], PVT-Ti [78], and PVTv2 [79], as the student models. EKD achieves an aver-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 328, + 76, + 421, + 168 + ], + "blocks": [ + { + "bbox": [ + 328, + 76, + 421, + 168 + ], + "lines": [ + { + "bbox": [ + 328, + 76, + 421, + 168 + ], + "spans": [ + { + "bbox": [ + 328, + 76, + 421, + 168 + ], + "type": "image", + "image_path": "2a7f338f5bb870264b0798680b7e76ebaa01936905edbd95875230ec0d3c5c14.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 347, + 170, + 399, + 180 + ], + "lines": [ + { + "bbox": [ + 347, + 170, + 399, + 180 + ], + "spans": [ + { + "bbox": [ + 347, + 170, + 399, + 180 + ], + "type": "text", + "content": "(a) Vanilla KD" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 440, + 76, + 533, + 168 + ], + "blocks": [ + { + "bbox": [ + 440, + 76, + 533, + 168 + ], + "lines": [ + { + "bbox": [ + 440, + 76, + 533, + 168 + ], + "spans": [ + { + "bbox": [ + 440, + 76, + 533, + 168 + ], + "type": "image", + "image_path": "2594fa2eed89f487982a5fde7f4bb9a2dfa2f501723ccda403b4d9ca248de348.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 469, + 170, + 501, + 180 + ], + "lines": [ + { + "bbox": [ + 469, + 170, + 501, + 180 + ], + "spans": [ + { + "bbox": [ + 469, + 170, + 501, + 180 + ], + "type": "text", + "content": "(b) EKD" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 327, + 182, + 428, + 265 + ], + "blocks": [ + { + "bbox": [ + 327, + 182, + 428, + 265 + ], + "lines": [ + { + "bbox": [ + 327, + 182, + 428, + 265 + ], + "spans": [ + { + "bbox": [ + 327, + 182, + 428, + 265 + ], + "type": "image", + "image_path": "dbe4e60df6dcc6342f43e51ad612ee43a24b0c462fce1c020766092fcf2b4486.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "lines": [ + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "spans": [ + { + "bbox": [ + 343, + 268, + 395, + 277 + ], + "type": "text", + "content": "(c) Vanilla KD" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 440, + 182, + 541, + 266 + ], + "blocks": [ + { + "bbox": [ + 440, + 182, + 541, + 266 + ], + "lines": [ + { + "bbox": [ + 440, + 182, + 541, + 266 + ], + "spans": [ + { + "bbox": [ + 440, + 182, + 541, + 266 + ], + "type": "image", + "image_path": "3a3fa36847db0e6259a3d13b6b8998a4aecef64ecdf6b4e2814fea8ab6df344f.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 472, + 268, + 504, + 277 + ], + "lines": [ + { + "bbox": [ + 472, + 268, + 504, + 277 + ], + "spans": [ + { + "bbox": [ + 472, + 268, + 504, + 277 + ], + "type": "text", + "content": "(d) EKD" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 281, + 555, + 324 + ], + "lines": [ + { + "bbox": [ + 313, + 281, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 281, + 555, + 324 + ], + "type": "text", + "content": "Figure 3. Visualization of features and logits output. The first row shows the t-SNE of features learned by vanilla KD and EKD. The second row illustrates the difference of correlation matrices of logits between the student and teacher." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 361, + 557, + 433 + ], + "blocks": [ + { + "bbox": [ + 313, + 338, + 554, + 359 + ], + "lines": [ + { + "bbox": [ + 313, + 338, + 554, + 359 + ], + "spans": [ + { + "bbox": [ + 313, + 338, + 554, + 359 + ], + "type": "text", + "content": "Table 6. Top-1 accuracy of various ViT student models on CIFAR100. Teacher model is ResNet56." + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 361, + 557, + 433 + ], + "lines": [ + { + "bbox": [ + 315, + 361, + 557, + 433 + ], + "spans": [ + { + "bbox": [ + 315, + 361, + 557, + 433 + ], + "type": "table", + "html": "
ArchitectureSizeEDLSoftmaxKD [30]AutoKD [42]EKD(Ours)
DeiT-Ti [71]5M64.7765.0873.2578.5578.64
PiT-Ti [29]5M73.4873.5875.4778.7679.33
PVT-Ti [78]13M69.0269.2273.6078.4378.74
PVTv2 [79]3M76.6877.4478.8179.3779.80
", + "image_path": "34e4f10813de0ce0107a57a32831453f06b9607532f5f128a12be563ebe5f399.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "table_body" + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "text", + "content": "age performance improvement of " + }, + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "inline_equation", + "content": "3.75\\%" + }, + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "text", + "content": " over KD [30] and an average improvement of " + }, + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 313, + 453, + 555, + 478 + ], + "type": "text", + "content": " over AutoKD [42]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 388, + 502 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 715 + ], + "type": "text", + "content": "This paper reinterprets vanilla knowledge distillation (KD) through a distribution-based probabilistic framework, employing evidential second-order distributions to effectively capture predictive uncertainty and provide a comprehensive knowledge representation. This pioneering methodology underpins Evidential Knowledge Distillation (EKD), facilitating knowledge transfer across macro and micro levels. At the macro level, aligning the expectation of the second-order distribution, which reflects its global characteristics, enhances optimization of inter-class proportions in the student's output. At the micro level, aligning the second-order distribution aligns the magnitudes of the student's output. These complementary mechanisms ensure a robust distillation process. Through PAC-Bayesian theory, EKD optimizes the upper bound of the student's expected risk. Extensive experiments underscore EKD's pronounced advantage over prevailing logit-based distillation techniques." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "2821" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported in part by the National Key Research and Development Plan of China under Grant 2023YFC3310700, in part by the National Natural Science Foundation of China under Grants 62036012, U21B2044, 62236008, 62472422, and U2333215, and in part by Beijing Natural Science Foundation under Grant 4242051." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "spans": [ + { + "bbox": [ + 56, + 174, + 115, + 186 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 226 + ], + "type": "text", + "content": "[1] Alexander Amini, Wilko Schwarting, Ava Soleimany, and Daniela Rus. Deep evidential regression. NIPS, 33:14927-14937, 2020. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "spans": [ + { + "bbox": [ + 61, + 228, + 296, + 261 + ], + "type": "text", + "content": "[2] Defang Chen, Jian-Ping Mei, Can Wang, Yan Feng, and Chun Chen. Online knowledge distillation with diverse peers. In AAAI, pages 3430-3437, 2020. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 262, + 295, + 305 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 262, + 295, + 305 + ], + "spans": [ + { + "bbox": [ + 62, + 262, + 295, + 305 + ], + "type": "text", + "content": "[3] Defang Chen, Jian-Ping Mei, Hailin Zhang, Can Wang, Yan Feng, and Chun Chen. Knowledge distillation with the reused teacher classifier. In CVPR, pages 11933-11942, 2022. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 308, + 294, + 340 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 308, + 294, + 340 + ], + "spans": [ + { + "bbox": [ + 62, + 308, + 294, + 340 + ], + "type": "text", + "content": "[4] Guobin Chen, Wongun Choi, Xiang Yu, Tony Han, and Manmohan Chandraker. Learning efficient object detection models with knowledge distillation. NIPS, 30, 2017. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 342, + 294, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 342, + 294, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 342, + 294, + 384 + ], + "type": "text", + "content": "[5] Mengyuan Chen, Junyu Gao, Shicai Yang, and Changsheng Xu. Dual-evidential learning for weakly-supervised temporal action localization. In ECCV, pages 192-208. Springer, 2022. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "spans": [ + { + "bbox": [ + 62, + 387, + 294, + 441 + ], + "type": "text", + "content": "[6] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Cascade evidential learning for open-world weakly-supervised temporal action localization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 14741-14750, 2023." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "spans": [ + { + "bbox": [ + 62, + 443, + 294, + 496 + ], + "type": "text", + "content": "[7] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Uncertainty-aware dual-evidential learning for weakly-supervised temporal action localization. IEEE transactions on pattern analysis and machine intelligence, 45(12):15896-15911, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 499, + 294, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 499, + 294, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 499, + 294, + 542 + ], + "type": "text", + "content": "[8] Mengyuan Chen, Junyu Gao, and Changsheng Xu. R-edl: Relaxing nonessential settings of evidential deep learning. In The Twelfth International Conference on Learning Representations, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 294, + 577 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 294, + 577 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 294, + 577 + ], + "type": "text", + "content": "[9] Mengyuan Chen, Junyu Gao, and Changsheng Xu. Revisiting essential and nonessential settings of evidential deep learning. arXiv preprint arXiv:2410.00393, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 294, + 610 + ], + "type": "text", + "content": "[10] Pengguang Chen, Shu Liu, Hengshuang Zhao, and Jiaya Jia. Distilling knowledge via knowledge review. In CVPR, pages 5008-5017, 2021. 5, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 655 + ], + "type": "text", + "content": "[11] Danruo Deng, Guangyong Chen, Yang Yu, Furui Liu, and Pheng-Ann Heng. Uncertainty estimation by fisher information-based evidential deep learning. In ICML, pages 7596–7616. PMLR, 2023. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 690 + ], + "type": "text", + "content": "[12] Thomas G Dietterich. Ensemble methods in machine learning. In International workshop on multiple classifier systems, pages 1-15. Springer, 2000. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 294, + 714 + ], + "type": "text", + "content": "[13] Luyang Fang, Yongkai Chen, Wenxuan Zhong, and Ping Ma. Bayesian knowledge distillation: A bayesian perspec" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "tive of distillation with uncertainty quantification. In Forty-first ICML, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 98, + 531, + 108 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 531, + 108 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 531, + 108 + ], + "type": "text", + "content": "[14] Yarin Gal et al. Uncertainty in deep learning. 2016. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 110, + 555, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 110, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 110, + 555, + 163 + ], + "type": "text", + "content": "[15] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Collecting cross-modal presence-absence evidence for weakly-supervised audio-visual event perception. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 18827-18836, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 166, + 554, + 198 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 166, + 554, + 198 + ], + "spans": [ + { + "bbox": [ + 317, + 166, + 554, + 198 + ], + "type": "text", + "content": "[16] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Vectorized evidential learning for weakly-supervised temporal action localization. PAMI, 2023. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 201, + 553, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 242 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 242 + ], + "type": "text", + "content": "[17] Junyu Gao, Mengyuan Chen, Liangyu Xiang, and Changsheng Xu. A comprehensive survey on evidential deep learning and its applications. arXiv preprint arXiv:2409.04720, 2024.3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 316, + 246, + 553, + 289 + ], + "type": "text", + "content": "[18] Junyu Gao, Mengyuan Chen, and Changsheng Xu. Learning probabilistic presence-absence evidence for weakly-supervised audio-visual event perception. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 292, + 553, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 292, + 553, + 323 + ], + "spans": [ + { + "bbox": [ + 316, + 292, + 553, + 323 + ], + "type": "text", + "content": "[19] Pascal Germain, Alexandre Lacasse, François Laviolette, and Mario Marchand. Pac-bayesian learning of linear classifiers. In ICML, pages 353-360, 2009. 2, 4" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 326, + 553, + 358 + ], + "type": "text", + "content": "[20] Hongji Guo, Hanjing Wang, and Qiang Ji. Bayesian evidential deep learning for online action detection. In ECCV, pages 283-301. Springer, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 360, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 381 + ], + "type": "text", + "content": "[21] Jia Guo. Reducing the teacher-student gap via adaptive temperatures. 2022. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 384, + 553, + 427 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 384, + 553, + 427 + ], + "spans": [ + { + "bbox": [ + 317, + 384, + 553, + 427 + ], + "type": "text", + "content": "[22] Yong Guo, Shulian Zhang, Haolin Pan, Jing Liu, Yulun Zhang, and Jian Chen. Gap preserving distillation by building bidirectional mappings with a dynamic teacher. arXiv preprint arXiv:2410.04140, 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "spans": [ + { + "bbox": [ + 316, + 429, + 553, + 461 + ], + "type": "text", + "content": "[23] Ziyao Guo, Haonan Yan, Hui Li, and Xiaodong Lin. Class attention transfer based knowledge distillation. In CVPR, pages 11868-11877, 2023. 6, 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 496 + ], + "type": "text", + "content": "[24] Corina Gurau, Alex Bewley, and Ingmar Posner. Dropout distillation for efficiently estimating model confidence. arXiv preprint arXiv:1809.10562, 2018. 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 519 + ], + "type": "text", + "content": "[25] Zongbo Han, Changqing Zhang, Huazhu Fu, and Joey Tianyi Zhou. Trusted multi-view classification. In ICLR, 2020. 2, 3" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 553 + ], + "type": "text", + "content": "[26] Manuel Haussmann, Sebastian Gerwinn, and Melih Kandemir. Bayesian evidential deep learning with pac regularization. arXiv preprint arXiv:1906.00816, 2019. 3, 4" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "spans": [ + { + "bbox": [ + 317, + 555, + 553, + 588 + ], + "type": "text", + "content": "[27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2016. 6" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 633 + ], + "type": "text", + "content": "[28] Mina Hemmatian, Ali Shahzadi, and Saeed Mozaffari. Uncertainty-based knowledge distillation for bayesian deep neural network compression. International Journal of Approximate Reasoning, 175:109301, 2024. 3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 677 + ], + "type": "text", + "content": "[29] Byeongho Heo, Sangdoo Yun, Dongyoon Han, Sanghyuk Chun, Junsuk Choe, and Seong Joon Oh. Rethinking spatial dimensions of vision transformers. In CVPR, pages 11936-11945, 2021. 8" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[30] Geoffrey Hinton. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 1, 2, 3, 6, 7, 8" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2822" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 17, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[31] Md Imtiaz Hossain, Sharmen Akhter, Choong Seon Hong, and Eui-Nam Huh. Single teacher, multiple perspectives: Teacher knowledge augmentation for enhanced knowledge distillation. In ICLR, 2025. 2, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 150 + ], + "type": "text", + "content": "[32] Andrew G Howard. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861, 2017. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 184 + ], + "type": "text", + "content": "[33] Tao Huang, Shan You, Fei Wang, Chen Qian, and Chang Xu. Knowledge distillation from a stronger teacher. NIPS, 35:33716-33727, 2022. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 229 + ], + "type": "text", + "content": "[34] Yaomin Huang, Zaomin Yan, Chaomin Shen, Faming Fang, and Guixu Zhang. Harmonizing knowledge transfer in neural network with unified distillation. In ECCV, pages 58-74. Springer, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 231, + 269, + 242 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 269, + 242 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 269, + 242 + ], + "type": "text", + "content": "[35] Audun Jøsang. Subjective logic. Springer, 2016. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 294, + 275 + ], + "type": "text", + "content": "[36] Jungeun Kim, Junwon You, Dongjin Lee, Ha Young Kim, and Jae-Hun Jung. Do topological characteristics help in knowledge distillation? In ICML, 2024. 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 308 + ], + "type": "text", + "content": "[37] Anoop Korattikara Balan, Vivek Rathod, Kevin P Murphy, and Max Welling. Bayesian dark knowledge. NIPS, 28, 2015. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 310, + 294, + 332 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 310, + 294, + 332 + ], + "spans": [ + { + "bbox": [ + 56, + 310, + 294, + 332 + ], + "type": "text", + "content": "[38] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. 2009. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 365 + ], + "type": "text", + "content": "[39] Balaji Lakshminarayanan, Alexander Pritzel, and Charles Blundell. Simple and scalable predictive uncertainty estimation using deep ensembles. NIPS, 30, 2017. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 367, + 294, + 421 + ], + "type": "text", + "content": "[40] Changbin Li, Kangshuo Li, Yuzhe Ou, Lance M Kaplan, Audun Jøsang, Jin-Hee Cho, Dong Hyun Jeong, and Feng Chen. Hyper evidential deep learning to quantify composite classification uncertainty. arXiv preprint arXiv:2404.10980, 2024.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 464 + ], + "type": "text", + "content": "[41] Kehan Li, Runyi Yu, Zhennan Wang, Li Yuan, Guoli Song, and Jie Chen. Locality guidance for improving vision transformers on tiny datasets. In ECCV, pages 110-127. Springer, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "spans": [ + { + "bbox": [ + 56, + 467, + 294, + 500 + ], + "type": "text", + "content": "[42] Lujun Li, Peijie Dong, Zimian Wei, and Ya Yang. Automated knowledge distillation via monte carlo tree search. In CVPR, pages 17413-17424, 2023. 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 501, + 294, + 534 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 534 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 534 + ], + "type": "text", + "content": "[43] Xue Li, Wei Shen, and Denis Charles. Tedl: A two-stage evidential deep learning method for classification uncertainty quantification. arXiv preprint arXiv:2209.05522, 2022. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 579 + ], + "type": "text", + "content": "[44] Zheng Li, Ying Huang, Defang Chen, Tianren Luo, Ning Cai, and Zhigeng Pan. Online knowledge distillation via multi-branch diversity enhancement. In Proceedings of the Asian conference on computer vision, 2020. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[45] Zheng Li, Xiang Li, Lingfeng Yang, Borui Zhao, Renjie Song, Lei Luo, Jun Li, and Jian Yang. Curriculum temperature for knowledge distillation. In AAAI, pages 1504-1512, 2023. 2, 6, 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 679 + ], + "type": "text", + "content": "[46] Jakob Lindqvist, Amanda Olmin, Fredrik Lindsten, and Lennart Svensson. A general framework for ensemble distribution distillation. In 2020 IEEE 30th International Workshop on Machine Learning for Signal Processing (MLSP), pages 1-6. IEEE, 2020. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[47] Michal Lukasik, Srinadh Bhojanapalli, Aditya Menon, and Sanjiv Kumar. Does label smoothing mitigate label noise? In ICML, pages 6448-6458. PMLR, 2020. 2" + } + ] + } + ], + "index": 16 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 95 + ], + "type": "text", + "content": "[48] Shicai Wei Chunbo Luo Yang Luo. Scale decoupled distillation. arXiv preprint arXiv:2403.13512, 2024. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[49] Huan Ma, Zongbo Han, Changqing Zhang, Huazhu Fu, Joey Tianyi Zhou, and Qinghua Hu. Trustworthy multimodal regression with mixture of normal-inverse gamma distributions. NIPS, 34:6881-6893, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 171 + ], + "type": "text", + "content": "[50] Andrey Malinin, Bruno Mlodozeniec, and Mark Gales. Ensemble distribution distillation. arXiv preprint arXiv:1905.00076, 2019. 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 216 + ], + "type": "text", + "content": "[51] Zelda E Mariet, Rodolphe Jenatton, Florian Wenzel, and Dustin Tran. Distilling ensembles improves uncertainty estimates. In Third symposium on advances in approximate bayesian inference, 2021. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 239 + ], + "type": "text", + "content": "[52] Nis Meinert and Alexander Lavin. Multivariate deep evidential regression. arXiv preprint arXiv:2104.06135, 2021. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 271 + ], + "type": "text", + "content": "[53] Zhong Meng, Jinyu Li, Yong Zhao, and Yifan Gong. Conditional teacher-student learning. In ICASSP, pages 6445-6449. IEEE, 2019. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 316 + ], + "type": "text", + "content": "[54] Seyed Iman Mirzadeh, Mehrdad Farajtabar, Ang Li, Nir Levine, Akihiro Matsukawa, and Hassan Ghasemzadeh. Improved knowledge distillation via teacher assistant. In AAAI, pages 5191-5198, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 317, + 553, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 349 + ], + "type": "text", + "content": "[55] Deep Shankar Pandey and Qi Yu. Learn to accumulate evidence from all training samples: theory and practice. In ICML, pages 26963-26989. PMLR, 2023. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 350, + 553, + 381 + ], + "type": "text", + "content": "[56] Wonpyo Park, Dongju Kim, Yan Lu, and Minsu Cho. Relational knowledge distillation. In CVPR, pages 3967-3976, 2019. 2, 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 425 + ], + "type": "text", + "content": "[57] Baoyun Peng, Xiao Jin, Jiaheng Liu, Dongsheng Li, Yichao Wu, Yu Liu, Shunfeng Zhou, and Zhaoning Zhang. Correlation congruence for knowledge distillation. In CVPR, pages 5007-5016, 2019. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 426, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 553, + 449 + ], + "type": "text", + "content": "[58] Bo Peng, Zhen Fang, Guangquan Zhang, and Jie Lu. Knowledge distillation with auxiliary variable. In ICML, 2024. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 491 + ], + "type": "text", + "content": "[59] Adriana Romero, Nicolas Ballas, Samira Ebrahimi Kahou, Antoine Chassang, Carlo Gatta, and Yoshua Bengio. Fitnets: Hints for thin deep nets. arXiv preprint arXiv:1412.6550, 2014. 6" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[60] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, et al. Imagenet large scale visual recognition challenge. IJCV, 115:211-252, 2015. 5, 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 548, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 580 + ], + "type": "text", + "content": "[61] Max Ryabinin, Andrey Malinin, and Mark Gales. Scaling ensemble distribution distillation to many classes with proxy targets. NIPS, 34:6023-6035, 2021. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 553, + 635 + ], + "type": "text", + "content": "[62] Siddharth Samsi, Dan Zhao, Joseph McDonald, Baolin Li, Adam Michaleas, Michael Jones, William Bergeron, Jeremy Kepner, Devesh Tiwari, and Vijay Gadepally. From words to watts: Benchmarking the energy costs of large language model inference. In HPEC, pages 1-9. IEEE, 2023. 1" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 679 + ], + "type": "text", + "content": "[63] Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zh-moginov, and Liang-Chieh Chen. *Mobilenetv2: Inverted residuals and linear bottlenecks*. In *CVPR*, pages 4510–4520, 2018. 6, 7" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 680, + 553, + 713 + ], + "type": "text", + "content": "[64] Murat Sensoy, Lance Kaplan, and Melih Kandemir. *Evidential deep learning to quantify classification uncertainty.* NIPS, 31, 2018. 2, 3" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2823" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 296, + 95 + ], + "type": "text", + "content": "[65] Glenn Shafer. Dempster-shafer theory. Encyclopedia of artificial intelligence, 1:330-331, 1992. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 140 + ], + "type": "text", + "content": "[66] Yichen Shen, Zhilu Zhang, Mert R Sabuncu, and Lin Sun. Real-time uncertainty estimation in computer vision via uncertainty-aware distribution distillation. In WACV, pages 707-716, 2021. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 296, + 174 + ], + "type": "text", + "content": "[67] Karen Simonyan and Andrew Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 175, + 296, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 296, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 296, + 206 + ], + "type": "text", + "content": "[68] Che Sun, Yunde Jia, and Yuwei Wu. Evidential reasoning for video anomaly detection. In ACMMM, pages 2106-2114, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 296, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 241 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 241 + ], + "type": "text", + "content": "[69] Shangquan Sun, Wenqi Ren, Jingzhi Li, Rui Wang, and Xiaochun Cao. Logit standardization in knowledge distillation. In CVPR, pages 15731-15740, 2024. 2, 5, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 243, + 296, + 275 + ], + "type": "text", + "content": "[70] Yonglong Tian, Dilip Krishnan, and Phillip Isola. Contrastive representation distillation. arXiv preprint arXiv:1910.10699, 2019. 2, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 56, + 277, + 296, + 320 + ], + "type": "text", + "content": "[71] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Hervé Jégou. Training data-efficient image transformers & distillation through attention. In ICML, pages 10347-10357. PMLR, 2021. 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 322, + 296, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 322, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 322, + 296, + 376 + ], + "type": "text", + "content": "[72] Linh Tran, Bastiaan S Veeling, Kevin Roth, Jakub Swiatkowski, Joshua V Dillon, Jasper Snoek, Stephan Mandt, Tim Salimans, Sebastian Nowozin, and Rodolphe Jenatton. Hydra: Preserving ensemble diversity for model distillation. arXiv preprint arXiv:2001.04694, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 377, + 296, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 377, + 296, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 377, + 296, + 399 + ], + "type": "text", + "content": "[73] Frederick Tung and Greg Mori. Similarity-preserving knowledge distillation. In CVPR, pages 1365-1374, 2019. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 401, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 401, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 401, + 296, + 443 + ], + "type": "text", + "content": "[74] Meet Vadera, Brian Jalaian, and Benjamin Marlin. Generalized bayesian posterior expectation distillation for deep neural networks. In Conference on Uncertainty in Artificial Intelligence, pages 719-728. PMLR, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 445, + 296, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 489 + ], + "type": "text", + "content": "[75] Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun Zhou, and Edwin Hancock. Uncertainty estimation for stereo matching based on evidential deep learning. pattern recognition, 124:108498, 2022. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 491, + 296, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 491, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 491, + 296, + 533 + ], + "type": "text", + "content": "[76] Hanjing Wang and Qiang Ji. Beyond dirichlet-based models: when bayesian neural networks meet evidential deep learning. In The 40th Conference on Uncertainty in Artificial Intelligence, 2024. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 296, + 578 + ], + "type": "text", + "content": "[77] Kuan-Chieh Wang, Paul Vicol, James Lucas, Li Gu, Roger Grosse, and Richard Zemel. Adversarial distillation of bayesian neural network posteriors. In ICML, pages 5190-5199. PMLR, 2018. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 296, + 633 + ], + "type": "text", + "content": "[78] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions. In CVPR, pages 568-578, 2021. 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 679 + ], + "type": "text", + "content": "[79] Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, and Ling Shao. Pvt v2: Improved baselines with pyramid vision transformer. Computational Visual Media, 8(3):415-424, 2022. 8" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 296, + 713 + ], + "type": "text", + "content": "[80] Guodong Xu, Ziwei Liu, Xiaoxiao Li, and Chen Change Loy. Knowledge distillation meets self-supervision. In ECCV, pages 588-604. Springer, 2020. 2" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 399 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 105 + ], + "type": "text", + "content": "[81] Ronald R Yager and Liping Liu. Classic works of the Dempster-Shafer theory of belief functions. Springer, 2008. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 140 + ], + "type": "text", + "content": "[82] Jing Yang, Brais Martinez, Adrian Bulat, Georgios Tzimiropoulos, et al. Knowledge distillation via softmax regression representation learning. ICLR (ICLR), 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 141, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 162 + ], + "type": "text", + "content": "[83] Sergey Zagoruyko. Wide residual networks. arXiv preprint arXiv:1605.07146, 2016. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "text", + "content": "[84] Feng Zhang, Xiatian Zhu, and Mao Ye. Fast human pose estimation. In CVPR, pages 3517-3526, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 187, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 187, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 187, + 553, + 230 + ], + "type": "text", + "content": "[85] Rongzhi Zhang, Jiaming Shen, Tianqi Liu, Jialu Liu, Michael Bendersky, Marc Najork, and Chao Zhang. Do not blindly imitate the teacher: Using perturbed loss for knowledge distillation. arXiv preprint arXiv:2305.05010, 2023. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 232, + 553, + 274 + ], + "type": "text", + "content": "[86] Xiangyu Zhang, Xinyu Zhou, Mengxiao Lin, and Jian Sun. Shufflenet: An extremely efficient convolutional neural network for mobile devices. In CVPR, pages 6848-6856, 2018. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 320 + ], + "type": "text", + "content": "[87] Yuan Zhang, Weihua Chen, Yichen Lu, Tao Huang, Xiuyu Sun, and Jian Cao. Avatar knowledge distillation: self-ensemble teacher paradigm with uncertainty. In ACMMM, pages 5272-5280, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 322, + 553, + 354 + ], + "type": "text", + "content": "[88] Borui Zhao, Quan Cui, Renjie Song, Yiyu Qiu, and Jiajun Liang. Decoupled knowledge distillation. In CVPR, pages 11953-11962, 2022. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "text", + "content": "[89] Yian Zhao, Wenyu Lv, Shangliang Xu, Jinman Wei, Guanzhong Wang, Qingqing Dang, Yi Liu, and Jie Chen. Detrs beat yolos on real-time object detection. In CVPR, pages 16965-16974, 2024. 1" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "2824" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_content_list.json b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..5db9d4684d158d26c9c93b8c0f014456fb946ef5 --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_content_list.json @@ -0,0 +1,1571 @@ +[ + { + "type": "text", + "text": "EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment", + "text_level": 1, + "bbox": [ + 235, + 130, + 759, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yufei Zhu $^{1,\\ast}$ , Yiming Zhong $^{1,\\ast}$ , Zemin Yang $^{1}$ , Peishan Cong $^{1}$ , Jingyi Yu $^{1}$ , Xinge Zhu $^{2,\\dagger}$ , Yuexin Ma $^{1,\\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ The Chinese University of Hong Kong https://evolvinggrasp.github.io/", + "bbox": [ + 230, + 202, + 764, + 273 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d4655f9aa5ef7b2023e42f5999c7209ff12a76e6328a513bdf7203923f33cfcf.jpg", + "image_caption": [ + "Figure 1. The left part illustrates EvolvingGrasp, an approach akin to evolution, where it enables the model to learn from experience and iteratively refine its grasping strategy. The right part demonstrates its efficiency and effectiveness." + ], + "image_footnote": [], + "bbox": [ + 96, + 313, + 545, + 578 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/46c6c0a21e9df52e097391058e722c6db16a390b7bac5b7b1c3bfd065838c04e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 558, + 316, + 901, + 578 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 619, + 326, + 636 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dexterous robotic hands often struggle to generalize effectively in complex environments due to models trained on low-diversity data. However, the real world presents an inherently unbounded range of scenarios. A natural solution is to enable robots learning from experience in complex environments—an approach akin to evolution, where systems improve through learning from both failures and successes. Motivated by this, we propose EvolvingGrasp, an evolutionary grasp generation method that continuously enhances grasping performance through efficient preference alignment. Specifically, we introduce Handpose-wise", + "bbox": [ + 88, + 656, + 482, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Preference Optimization (HPO), which allows the model to continuously align with preferences from both positive and negative feedback while progressively refining its grasping strategies. To further enhance efficiency and reliability during online adjustments, we incorporate a Physics-aware Consistency Model within HPO, which accelerates inference, reduces the number of timesteps needed for preference fine-tuning, and ensures physical plausibility throughout the process. Our results validate that EvolvingGrasp enables evolutionary grasp generation, ensuring robust, physically feasible, and preference-aligned grasping in both simulation and real scenarios.", + "bbox": [ + 511, + 622, + 906, + 803 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 840, + 645, + 857 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dexterous robotic grasping has made significant strides in embodied manipulation, enabling more adaptive and pre", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{1}$ * Equal contribution.", + "bbox": [ + 107, + 838, + 230, + 851 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$^{2\\dagger}$ Corresponding author. This work was supported by NSFC (No.62206173), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (KLIP-HuMaCo).", + "bbox": [ + 91, + 851, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "11665", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "cise interactions compared to traditional grippers. Existing grasping methods [5, 9, 15, 17-19, 29, 31-33, 42, 48] generally fall into two categories: optimization-based approaches [5, 9, 29, 32], which refine hand poses to achieve force-closure states, learning-based approaches [26, 40, 41, 43], which directly map object features to grasp poses through regression, probabilistic modeling, and generative-based methods [12, 13, 24, 25, 42], which utilize diffusion model to estimate the distribution of hand poses. Recent advances, such as DexGrasp Anything [49], have further introduced physics-based constraints to improve grasp feasibility. However, a fundamental limitation persists—limited generalization. These methods, trained on limited datasets, struggle to adapt to complex environments. This challenge is exacerbated by an inherent property of the real world: its unbounded diversity. The vast range of object shapes, materials, and environmental conditions makes it impractical to predefine an exhaustive set of grasping strategies. Without the ability to adapt in deployment, grasping models remain constrained, failing to handle variations beyond their training distribution. To overcome this, a natural approach is to enable evolutionary grasp generation, where the system learns from experience (i.e., both failures and successes), refining its grasping strategy through iterative improvements based on real-world interactions. This process not only enhances generalization but also allows for preference alignment, ensuring that grasping behaviors adapt to task-specific requirements and user-defined preferences. However, achieving efficient evolutionary refinement is non-trivial, as many existing learning-based approaches rely on slow, compute-intensive updates, particularly in diffusion-based models that require numerous iterative steps and physics simulations.", + "bbox": [ + 93, + 90, + 480, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose EvolvingGrasp, an evolutionary grasp generation framework that efficiently refines grasp strategies through preference alignment while maintaining physical plausibility. At its core, we introduce Handpose-wise Preference Optimization (HPO), a novel method that reformulates preference alignment [2, 37, 38, 47, 50] as a posterior probability optimization problem, encouraging generated grasps to converge toward preferred distributions while diverging from non-preferred ones. Notably, the proposed HPO is an extension of Direct Preference Optimization (DPO) [30], where it is also, to the best of our knowledge, the first to incorporate the DPO into the dexterous grasp. To further improve efficiency, we integrate HPO into a Physics-Aware Consistency Model (including two parts, i.e., Physics-Aware Distillation for training and Physics-Aware Sampling for inference), which pretrains a diffusion model and distills it into a lightweight, few-step sampling model. This enables both rapid inference and efficient preference fine-tuning, significantly reducing the number of required sampling steps and optimization iteraa", + "bbox": [ + 93, + 599, + 480, + 898 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "tions. Additionally, we introduce three physics-aware constraints to ensure the stability, realism, and feasibility of generated grasp poses—surface pulling force to maintain stable contact, external penetration repulsion force to prevent object penetration, and self-penetration repulsion force to avoid inter-finger collisions.", + "bbox": [ + 516, + 90, + 903, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments across four benchmark datasets demonstrate that EvolvingGrasp achieves state-of-the-art results, significantly improving grasp success rate, sampling efficiency, and physical plausibility, with 30x speedup over existing methods, demonstrating robust generalization across simulated and real-world benchmarks. Our contributions can be summarized as follows:", + "bbox": [ + 516, + 183, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce EvolvingGrasp, an efficient evolutionary grasp generation framework that enables iterative refinement, addressing the challenge of generalizing to diverse and unstructured real-world environments.", + "- Efficient preference alignment is achieved through Handpose-wise Preference Optimization (HPO), which formulates grasp adaptation as a posterior probability optimization problem, enabling the model to iteratively converge toward preferred grasp distributions.", + "- We propose a Physics-Aware Consistency Model (PCM) that accelerates preference alignment by reducing sampling steps while enforcing geometric consistency and physical feasibility through structured constraints.", + "- Extensive experiments across four benchmark datasets demonstrate that our method achieves state-of-the-art grasp success rates, physical plausibility, and sampling efficiency. Furthermore, it enables real-time grasp generation with minimal computational overhead, achieving comparable performance to gradient-based methods." + ], + "bbox": [ + 513, + 291, + 903, + 577 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 516, + 595, + 653, + 611 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Dexterous Grasping", + "text_level": 1, + "bbox": [ + 516, + 622, + 702, + 637 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Dexterous grasp generation aims to produce high-quality grasping poses for robotic hands. Recent works can be categorized into regression-based [17, 18, 31] and generation-based methods [12, 13, 24-26, 40, 41]. Regression-based methods directly predict grasping parameters from the input object, but they often suffer from mode collapse issues that limits output diversity. Generation-based methods, though capable of producing varied solutions, face efficiency challenges. SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] require multiple sampling steps to generate diverse grasping poses, and DexGrasp Anything [49] additionally incurs computational overhead by calculating physical constraint losses and performing gradient updates at each sampling iteration. However, existing methods cannot generate grasping poses aligned with human habits due to their lack of evolutionary adaptation and preference alignment. They also struggle to balance physical plausibility", + "bbox": [ + 514, + 643, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "11666", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/57c83978e95fa440daea1bee7421ad3f6a20fc8a0f44a19eaabb6e9f4333761e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 91, + 88, + 903, + 292 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/887f1e1bf94a34a3990de5441bba2b5321063f5966e68198a3406d5e2caf7aaa.jpg", + "image_caption": [ + "Figure 2. Overview of EvolvingGrasp. The evolutionary process begins with the human preference guidance, where Handpose-wise Preference Optimization (HPO, highlighted in the green rectangle) is employed to facilitate preference alignment. These grasp poses are generated by the Physics-Aware Consistency Model (shown in blue rectangles), including Sampling and Distillation mechanism, to ensure the sampling efficiency and the physical plausibility. In this way, EvolvingGrasp, an efficient evolutionary grasp generation framework is proposed to enable the grasp model iteratively converge toward preferred distributions." + ], + "image_footnote": [], + "bbox": [ + 93, + 296, + 903, + 406 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "with computational efficiency.", + "bbox": [ + 89, + 512, + 294, + 527 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Accelerating Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 539, + 359, + 556 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Accelerating dexterous grasping is vital for enhancing real-time adaptability. A series of diffusion-based acceleration methods have emerged, making real-time and efficient generation possible. DDIM [34] accelerates inference by reformulating the stochastic denoising process into a deterministic ODE solver. The DPM-Solver [22, 23] series further designs efficient high-order ODE solvers, achieving comparable quality with only $10 \\sim 20$ steps. However, their performance degrades significantly when sampling steps are reduced to $2 \\sim 4$ . Consistency works [14, 21] support few-step sampling while preserving quality. CTM [14] extends CM to reduce cumulative errors and discretization inaccuracies. sCMs [21] unifies flow matching and diffusion frameworks to avoid discretization errors and hyperparameter tuning. Therefore, we employ the consistency model to enable efficient evolutionary refinement, reducing inference timesteps while maintaining result quality.", + "bbox": [ + 88, + 561, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Preference Finetuning for Diffusion Models", + "text_level": 1, + "bbox": [ + 89, + 830, + 460, + 848 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preference alignment play a crucial role in enabling generative models to learn from user feedback, optimize generation strategies, and progressively improve output qual", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ity. Some researchers utilize it into diffusion models to better align with human preferences in image generation area. These methods can be categorized into two types, finetuning with reward model [1, 4, 6, 8] and reward-free finetuning [10, 38, 45]. The former includes DDPO [1] and DPOK [8] which treat the denoising process of diffusion models as an MDP and finetune using multiple reward models. Finetuning without reward model includes D3PO [45] and Diffusion-DPO [38], which extend the theoretical framework of DPO [30] to multi-step MDPs. These approaches learn an optimal reward model and then use it to refine the sampling strategy, making them a relatively more cost-effective alternative. However, RL-based fine-tuning for diffusion models still requires backpropagation at every sampled timestep, which is time-consuming. Applying this approach directly to grasping tasks is impractical for real-world applications. Therefore, we propose a faster preference alignment fine-tuning approach that reduces backpropagation steps.", + "bbox": [ + 511, + 511, + 906, + 800 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 511, + 819, + 648, + 835 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 511, + 845, + 624, + 861 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To achieve efficient evolutionary grasp generation that aligns with human preferences while maintaining both ef", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "11667", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ficiency and physical plausibility, we propose Evolving-Grasp, as illustrated in Fig. 2. The evolutionary process begins with preference alignment, where we introduce Handpose-Wise Preference Optimization (HPO) (Section 3.3). HPO formulates grasp adaptation as a posterior probability optimization problem, allowing the model to iteratively converge toward preferred grasp distributions. However, directly applying HPO suffers from slow sampling and inefficiencies in preference alignment, limiting its practicality for real-time grasp refinement. To address these challenges, we incorporate a Physics-Aware Consistency Model (PCM) (Section 3.4), which leverages a consistency model framework to enhance efficiency by reducing the number of required sampling steps. While this improves inference speed, naive consistency-based sampling may still generate physically implausible grasp poses. To ensure geometric consistency and physical feasibility, PCM integrates a physics-aware distillation and sampling mechanism, which enforces structured physical constraints on the generated poses.", + "bbox": [ + 89, + 90, + 483, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 402, + 294, + 419 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given the object point cloud representation $O \\in \\mathbb{R}^{N \\times 3}$ , our goal is to generate dexterous grasp poses with high success rate and low penetration from the posterior distribution $P(x \\mid O)$ , where $x = \\{x_{i}\\}_{i=1}^{n}$ . Specifically, the pose parameters contain three parts, joint angles of the hand $\\theta_{h} \\in \\mathbb{R}^{24}$ , global translation $T_{global} \\in \\mathbb{R}^{3}$ , and global rotation $R_{global} \\in SO(3)$ .", + "bbox": [ + 89, + 425, + 483, + 532 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Given ground truth samples from the data distribution $\\pi(x_0)$ , the noise schedule weight $\\alpha$ , the goal of diffusion models is to fit the GT data distribution. The objective of training diffusion model is as follows:", + "bbox": [ + 89, + 532, + 483, + 594 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbb {E} _ {t \\sim [ 0, T ], x _ {0}, \\epsilon \\sim N (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, O) \\| ^ {2} \\right] \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 145, + 604, + 482, + 632 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon$ . Given the noise related parameter $\\sigma_{t}$ and the mean from $x_{t}$ to $x_{t-1}$ , the reverse process of diffusion model is as follows:", + "bbox": [ + 89, + 643, + 482, + 688 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right) \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 700, + 482, + 720 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Handpose-wise Preference Optimization", + "text_level": 1, + "bbox": [ + 89, + 742, + 436, + 758 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To achieve preference alignment in the evolutionary adaptation process, we introduce the HPO, an extension of direct preference optimization (DPO). In DPO, the fundamental assumption is that we have access to data generated by the model, with human annotators providing the corresponding preferences. Specifically, for a given object $O$ , we observe pairs of grasp poses $x_0^w$ and $x_0^l$ , where the set $x_0^w$ is preferred over $x_0^l$ . These preferences can be modeled using the Bradley-Terry model, which expresses the probability", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of one sample being preferred over another as:", + "bbox": [ + 511, + 90, + 821, + 106 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np _ {B T} \\left(x _ {0} ^ {w} \\succ x _ {0} ^ {l}\\right) = \\sigma \\left(r \\left(c, x _ {0} ^ {w}\\right) - r \\left(c, x _ {0} ^ {l}\\right)\\right) \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 117, + 906, + 137 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The training objective of DPO is to maximize the likelihood of the observed preferences, which can be formulated as a binary classification problem. Specifically, the whole model is finetuned to minimize the following loss function:", + "bbox": [ + 511, + 148, + 906, + 210 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL _ {\\mathrm {B T}} (\\theta) = - \\mathbb {E} _ {\\left(x _ {0} ^ {w}, x _ {0} ^ {l}\\right) \\sim \\mathcal {D}} \\log \\sigma \\left( \\begin{array}{c} \\beta \\mathbb {E} \\\\ x _ {1: T} ^ {w} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {w} \\mid x _ {0} ^ {w}\\right) \\\\ x _ {1: T} ^ {l} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {l} \\mid x _ {0} ^ {l}\\right) \\end{array} \\right. \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 222, + 905, + 276 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left. \\left[ \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {w}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {l}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {l}\\right)} \\right]\\right)\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 277, + 867, + 316 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{D}$ is the paired-wise grasp pose dataset. Following [38], we can get the upper bound of Eq. (4) as:", + "bbox": [ + 511, + 328, + 906, + 359 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL_{BT}(\\theta)\\leq -\\begin{array}{c}\\beta \\mathbb{E}\\\\ \\big(x^{w}_{0},x^{l}_{0}\\big)\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{w}_{n - 1,n}\\sim \\pi_{\\theta}\\big(x^{w}_{n - 1,n}|x^{w}_{0}\\big) \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 369, + 890, + 402 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\log \\sigma \\left(\\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}\\right) \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 402, + 903, + 457 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pi_{\\theta}\\left(x_{n - 1} \\mid x_n\\right)$ is the probability of sampling $x_{n - 1}$ from $x_{n}$ , which can be computed by:", + "bbox": [ + 511, + 458, + 903, + 488 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\pi_ {\\theta} \\left(x _ {n - 1} \\mid x _ {n}\\right) = \\frac {1}{\\sqrt {2 \\pi} \\sigma_ {n}} \\exp \\left(- \\frac {\\left(x _ {n} - \\mu_ {\\theta} \\left(x _ {n} , n\\right)\\right) ^ {2}}{2 \\sigma_ {n} ^ {2}}\\right) \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 521, + 500, + 905, + 535 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As far as we know, HPO is the first to integrate DPO into grasp pose generation, from which we extend DPO to a more flexible form. In HPO, there is no requirement to maintain an equal number of preferred and non-preferred grasp poses, which enables more adaptive preference learning. The objective of HPO is introduced as follows:", + "bbox": [ + 511, + 546, + 906, + 637 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal{L}_{HPO} = \\operatorname *{\\beta}\\limits_{\\substack{x^{i}_{0}\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{i}_{n - 1,n}\\sim \\pi_{\\theta}\\left(x^{i}_{n - 1,n}|x^{i}_{0}\\right)}}\\log \\sigma\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 647, + 893, + 676 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\left(\\sum_ {i = 1} ^ {N _ {s u c}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)} - \\sum_ {j = 1} ^ {N _ {f a i l}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}\\right) \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 678, + 903, + 734 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $N_{\\text{succ}}$ and $N_{\\text{fail}}$ represent the number of preferred and non-preferred poses. HPO optimizes grasp pose generation by quantifying the probabilistic divergence between successful grasps (preferred samples) and failed grasps (non-preferred samples), thereby driving the model towards human-preferred behaviors. This process inherently involves the dynamic adjustment and maximization of reward signals: preferred grasp poses are reinforced due to their higher probability, while non-preferred poses are suppressed, guiding the model to progressively discard ineffective strategies.", + "bbox": [ + 511, + 734, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "11668", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preferred grasp selection can be conducted through either simulation-based evaluation or human-in-the-loop selection. In the simulation-based approach, poses that achieve success across all six directional evaluations are classified as preferred samples, while the remaining poses are treated as negative samples. Alternatively, in the human selection process, grasp poses that align with human intuition and habitual preferences are designated as preferred samples, while the others are considered non-preferred. Finally, we finetune the whole model via LoRA [11] using Eq. (7) to align the model with preferences.", + "bbox": [ + 89, + 90, + 483, + 257 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Physics-Aware Consistency Model", + "text_level": 1, + "bbox": [ + 89, + 266, + 388, + 282 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.1. Consistency Model", + "text_level": 1, + "bbox": [ + 89, + 287, + 269, + 301 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Directly applying HPO faces two major efficiency challenges: generating poses requires over hundreds of timesteps per inference, and preference fine-tuning demands a large number of backpropagation steps. To address these inefficiencies, we adopt the consistency model framework to accelerate both sampling process and preference alignment. The core idea of the consistency model is to learn a mapping from any point along the ODE trajectory back to its starting point, which corresponds to the data distribution. Given a Probability Flow ODE trajectory [35] $\\{x_{\\tau_n}\\}_{\\tau_n\\in [0,T]}$ , a consistency model $f_{\\theta}$ is defined as:", + "bbox": [ + 89, + 306, + 483, + 474 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta}: \\left(x _ {\\tau_ {n}}, \\tau_ {n}, O\\right) \\mapsto x _ {0} \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 487, + 480, + 503 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $x_0$ is the initial point of the trajectory, $O$ is the observation as condition. Due to space limitations, we omitted condition $O$ in other parts of the paper, except for the appendix. The self-consistency property ensures that:", + "bbox": [ + 89, + 513, + 483, + 575 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta^ {\\prime}} \\left(x _ {\\tau_ {n} ^ {\\prime}}, \\tau_ {n} ^ {\\prime}\\right) \\quad \\forall \\tau_ {n}, \\tau_ {n} ^ {\\prime} \\in [ 0, T ] \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 584, + 482, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The consistency model needs to meet a key boundary condition: when $t = 0$ , the model output should be the input itself, that is, $f_{\\theta}(x_0,0,O) = x_0$ . This can be achieved as the following way:", + "bbox": [ + 89, + 612, + 483, + 672 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = c _ {\\text {s k i p}} \\left(\\tau_ {n}\\right) x _ {\\tau_ {n}} + c _ {\\text {o u t}} \\left(\\tau_ {n}\\right) F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 104, + 683, + 482, + 700 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $c_{skip}(t)$ and $c_{out}(t)$ are differentiable functions that satisfy $c_{skip}(\\epsilon) = 1$ and $c_{out}(\\epsilon) = 0$ . $F_{\\theta}(x,t)$ denotes a deep neural network that predicts $\\hat{x}_0$ .", + "bbox": [ + 89, + 710, + 483, + 756 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nF _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {\\tau_ {n}}}} \\left(x _ {\\tau_ {n}} - \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n}}} \\epsilon_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right)\\right) \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 106, + 765, + 480, + 809 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Consistency distillation leverages pre-trained diffusion model to condense multi-step sampling into a more efficient few-step inference process. The process involves generating pairs of adjacent points on the PFODE trajectory using numerical ODE solvers and minimizing the difference between the model's outputs for these pairs. The loss function", + "bbox": [ + 89, + 810, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "for consistency distillation is defined as:", + "bbox": [ + 513, + 90, + 779, + 104 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {C D} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), f _ {\\theta^ {\\prime}} \\left(\\hat {x} _ {\\tau_ {n - 1}} ^ {*}, \\tau_ {n - 1}\\right)\\right) \\right] \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 531, + 116, + 903, + 142 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $O$ is omitted for simplicity, $\\hat{x}_{\\tau_{n-1}}^*$ is computed using a numerical ODE solver, $d(\\cdot, \\cdot)$ is a $L_2$ distance metric, $\\theta$ and $\\theta'$ represent the online network and target network parameters. In addition, the parameters of $\\theta'$ are updated by the exponential moving average (EMA) of the parameters of $\\theta$ . $\\hat{x}_{\\tau_{n-1}}^*$ can be computed as:", + "bbox": [ + 511, + 152, + 906, + 247 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {\\tau_ {n - 1}} ^ {*} \\leftarrow \\sqrt {\\bar {\\alpha} _ {\\tau_ {n - 1}}} F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n - 1}}} \\epsilon \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 257, + 903, + 277 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During sampling, we utilize the consistency function to directly generate the final sample and the quality of the generated sample can be enhanced through an iterative process that alternates between denoising and injecting noise. Given the sequence of timesteps $S \\in \\{\\tau_i \\mid \\tau_0 = 0, \\tau_{N-1} = T, \\tau_i < \\tau_{i+1} \\text{ for } i = 0, 1, \\dots, N-1\\}$ , the adding-noise process can be formulated as:", + "bbox": [ + 511, + 287, + 911, + 393 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {x} _ {\\tau_ {n - 1}} = \\mu_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sigma_ {\\tau_ {n}} \\epsilon \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 611, + 405, + 903, + 422 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mu_{\\theta} = \\sqrt{\\bar{\\alpha}_{\\tau_{n - 1}}} f_{\\theta}(x_{\\tau_n},\\tau_n)$ $\\sigma_{\\tau_n} = \\sqrt{1 - \\bar{\\alpha}_{\\tau_{n - 1}}}$ Subsequently, we conduct the prediction of the final sample utilizing the trained consistency function once more.", + "bbox": [ + 511, + 433, + 903, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4.2. Physics-Aware Distillation and Sampling", + "text_level": 1, + "bbox": [ + 511, + 501, + 843, + 517 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Although adopting the consistency model improves sampling and preference fine-tuning efficiency, it still generates physically implausible poses. To address this, we introduce a Physics-Aware Distillation and Sampling paradigm that enforces physical constraints during the distillation process of predicting $\\hat{x}_0$ while ensuring that the sampling trajectories adhere to specific constraints. Following [49], we incorporate three physics-aware objectives, with the distillation objective of the consistency model formulated as follows:", + "bbox": [ + 511, + 520, + 906, + 656 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {P A D} = \\mathcal {L} _ {C D} + \\sum_ {i = 1} ^ {m} \\alpha_ {i} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 527, + 667, + 903, + 705 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $L_{PA_i}\\left(F_\\theta(x_t, t), \\epsilon_\\theta\\right)$ is the $i^{th}$ physical constraint loss and $m = 3$ denotes three constraints, i.e., Surface Pulling Force, External-penetration Repulsion Force, Self-Penetration Repulsion Force respectively [49]. These constraints guarantee grasping feasibility and maintain geometric accuracy in finger-to-object and inter-finger interactions. $\\alpha_i$ is the corresponding weight parameter. We first train a diffusion model for grasp pose generation as a teacher model. Then, we utilize $\\mathcal{L}_{\\mathrm{PAD}}$ to distill the teacher model into a student model.", + "bbox": [ + 511, + 719, + 906, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "During sampling, the consistency model estimates a clearer hand pose $x_0$ based on the current noisy hand", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "11669", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e1628acfa9f5a828e02c518a23a04c4075c5e239503df446d5f57fe47f10145.jpg", + "image_caption": [ + "Figure 3. Mean grasping performance in terms of success rate and penetration of randomly selected 6 objects with the finetuning epoch increasing during inference optimization." + ], + "image_footnote": [], + "bbox": [ + 91, + 85, + 490, + 220 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ecba72db4560fdacdd30b98d254c24d670062dc788dd5e4e07ad8c504a333d26.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 491, + 85, + 905, + 220 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pose $x_{\\tau_n}$ and object information $O$ by consistency function $f_{\\theta}(x_{\\tau_n}, \\tau_n, O)$ . Subsequently, physical constraints are applied to steer the sampling process, making it closer to a physically feasible grasping pose. Following [3, 7, 46], the gradient of the constraint loss is used to modify the mean from $x_t$ to $x_{t-1}$ :", + "bbox": [ + 89, + 268, + 483, + 358 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\hat {\\mu} _ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) = \\mu_ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) \\\\ + \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {16} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 114, + 368, + 480, + 426 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\gamma_{i}$ is the weight parameter corresponding to each physical constraint. Moreover, we can derive a new mapping from noise to data as follows:", + "bbox": [ + 89, + 436, + 483, + 481 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\frac {1}{\\sqrt {\\alpha_ {\\tau_ {n - 1}}}} \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\tag {17}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 492, + 480, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By employing the physics-aware consistency model, we derive a novel preference alignment objective based on the new sampling path. The detailed derivation is provided in Appendix B. As a result, our method can efficiently generate higher-quality poses during the evolutionary process.", + "bbox": [ + 89, + 545, + 483, + 619 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 633, + 223, + 648 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We first simply present experimental setup which includes datasets, evaluation metrics and baselines in Sec. 4.3. More details of setup including datasets and implementation details are in Appendix C. Then we provide the quantitative and qualitative results of EvolvingGrasp in Sec. 4.2, followed by ablation study about different modules and hyperparameter analyses in Sec. 4.3. Next, we demonstrate the evolutionary improvement of EvolvingGrasp through an experiment starting from a suboptimal model trained on a degraded dataset in Sec. 4.4. Finally, We will demonstrate EvolvingGrasp in the real-world deployment in Sec. 4.5.", + "bbox": [ + 89, + 657, + 482, + 824 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 89, + 833, + 281, + 849 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct experiment on four datasets, including DexGraspNet [39], Multidex [16], Realdex [20], Dex-GRAB [36] respectively.", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metrics. We use four metrics to evaluate the grasping performance. Success rate Suc.6 measures the proportion of grasping poses where the object's displacement does not exceed $2\\mathrm{cm}$ in all six axial directions $(\\pm X, \\pm Y, \\pm Z)$ , evaluating multi-directional stability. Suc.1 measures the proportion where displacement does not exceed $2\\mathrm{cm}$ in at least one direction, assessing single-direction stability. Pen. indicates the maximum penetration depth (mm) between the hand and the object, with lower values suggesting more physically plausible grasps. Above metrics are calculated in the IssacGym simulator [27] with settings consistent with SceneDiffuser [12]. For efficiency, Time refers to the computational time required to generate grasping poses for a batch of objects.", + "bbox": [ + 511, + 268, + 906, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. Some generative-based methods such as UniDexGrasp [44], GraspTTA [13], SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] are compared on four benchmark datasets.", + "bbox": [ + 511, + 481, + 906, + 542 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Main Results", + "text_level": 1, + "bbox": [ + 511, + 556, + 651, + 571 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To validate the effectiveness of our method in continuously enhancing grasping performance, we conduct quantitative evaluations on six randomly selected objects from each dataset, measuring both the Suc.6 metric and the degree of penetration. The results are illustrated in Fig. 3, indicating that as the number of inferences increases, the Suc.6 metric steadily improves. Although penetration exhibits some fluctuations due to increased contact area between the hand and the object, it still follows a downward trend. Additionally, we present qualitative results in Fig. 4, demonstrating how our method integrates human preferences during inference. This enables the selection of more favorable poses and progressively refines grasping poses to better align with human preferences through iteration. For instance, early generated poses may obstruct the camera lens, but after several rounds of preference-based finetuning, the model learns to produce poses that avoid blocking the lens.", + "bbox": [ + 511, + 580, + 906, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct the experiment with comparing with other dexterous grasp generation methods in Tab. 1. Compared to other generative-based approaches, our method excels in generating high-quality, physically plausible grasping", + "bbox": [ + 511, + 839, + 906, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "11670", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/597b86aab741ae7722c4dd4ff1de62686462c91d9b7737a4316546686b60b18f.jpg", + "image_caption": [ + "Figure 4. Evolution of robotic grasp preferences during efficient feedback-driven finetuning across 10 epochs. Top row illustrates the adjustment from hand occlusion to clear nozzle visibility. Middle row demonstrates the transition from lens obstruction to an unobstructed camera view. The bottom row shows the evolution from a top-down grasping approach to a bottom-up one, while simultaneously mitigating physical impacts." + ], + "image_footnote": [], + "bbox": [ + 94, + 89, + 906, + 362 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "poses while significantly reducing computational time. It is worth highlighting that the proposed method achieves 30x speedup (32s~38s v.s. 0.73s~2.7s) compared to these SO-TAs. After fine-tuning, our approach achieves superior performance by leveraging its evolutionary capability through iterative refinement. We also evaluate the trade-off between efficiency and grasp quality during preference alignment with different numbers of inference steps. While increasing the steps improves performance, it inevitably incurs additional computational costs. Moreover, even without guidance during sampling, we can achieve comparable results in real time.", + "bbox": [ + 88, + 443, + 485, + 625 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 636, + 243, + 652 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We investigate the impact of different modules in the physics-aware consistency model, Physical constraints Guidance in Distillation and Sampling from which we dubbed as PGD and PGS. We also consider the performance of HPO with and without physical guidance during sampling on test split of the multidex dataset. The results presented in Tab. 2 demonstrate the key role of physical constraints and preference finetuning to successful grasping pose generation during preference alignment.", + "bbox": [ + 88, + 657, + 482, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "HPO not only effectively aligns grasping poses with human preferences at the individual object level but also enhances performance across the entire dataset. During inference across the entire dataset, we collect both successful and failed grasping samples for each object, utilizing Eq. (5) to perform lightweight finetuning of the entire model. Fig. 5 illustrates the relationship between EvolvingGrasp's", + "bbox": [ + 88, + 795, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/854e1d31947e04fe8280bd1a5bb6660829ac6d7277e90de68576bcd22d4642a6.jpg", + "image_caption": [ + "Figure 5. The relationship between the success rate of Evolving-Grasp and the number of generated samples." + ], + "image_footnote": [], + "bbox": [ + 542, + 440, + 877, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Suc.6 metric and the number of generated samples, showing that our method continuously enhances grasping performance as more samples are generated. Further ablation studies on the impact of different hyperparameters are provided in Appendix D.2.", + "bbox": [ + 511, + 679, + 906, + 755 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Training from a Degraded Dataset", + "text_level": 1, + "bbox": [ + 511, + 770, + 815, + 787 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In this subsection, we explore the feasibility of enhancing model performance through feedback-driven finetuning for single-object grasping tasks, particularly when trained on suboptimal data. To begin with, we randomly add the noise on the hand pose parameter of the original multidex dataset. This perturbation creates a degraded dataset, which we then use to train and distill a suboptimal model. During the in-", + "bbox": [ + 511, + 794, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "11671", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/d0a83e3ae4ced5a9941bf31816311e5658ca82bce6e59a69889ffe5589126b58.jpg", + "table_caption": [ + "Table 1. Grasping performance in terms of Suc.6, Suc.1, and Pen. comparison across different methods and datasets. \"Step\" refers to inference timestep. Bold values highlight the best results and underlined values indicate the runner-up." + ], + "table_footnote": [], + "table_body": "
Dataset MethodDexGraspNetMultiDexRealDexDexGRABTime ↓
Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓
UniDexGrasp [44]33.970.131.921.647.513.527.159.439.020.855.837.40.46±0.11
GraspTTA [13]18.667.824.530.362.819.013.346.440.114.451.051.410.41±0.32
SceneDiffuser [12]26.666.931.069.885.614.621.756.142.039.185.041.13.41±0.13
UGG [24]46.979.025.255.393.410.332.763.434.442.790.633.238.34±2.31
DexGrasp Any. [49]53.690.421.572.296.39.634.671.223.156.591.828.632.91±1.34
Ours w/o HPO (2-step)60.891.019.265.697.515.241.975.419.552.293.925.10.73±0.03
Ours (2-step)62.490.519.365.997.215.344.077.819.753.392.524.30.73±0.03
Ours w/o HPO (4-step)63.893.017.475.397.113.151.682.920.555.696.023.81.41±0.07
Ours (4-step)65.292.717.276.898.413.050.682.520.357.795.223.71.41±0.07
Ours w/o HPO (8-step)65.293.516.275.698.712.263.686.621.956.896.823.12.71±0.08
Ours (8-step)65.492.315.980.398.712.364.489.121.860.896.422.32.71±0.08
Real-time (2-step)55.290.520.063.795.013.846.578.921.448.993.324.80.06±0.01
Real-time (4-step)59.990.619.864.396.511.658.287.021.155.495.424.10.10±0.02
", + "bbox": [ + 93, + 128, + 906, + 364 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/02a7b28a68d870b975b5f3909b9d41d1f51efcf7455782c2cbc9f2054d426a47.jpg", + "table_caption": [ + "Table 2. Ablation study on incorporating physical constraints during both training and sampling stages and the LLM module. The evaluation is conducted on Multidex when timestep is 4." + ], + "table_footnote": [], + "table_body": "
CMPGDPGSHPOSuc.6 ↑Suc.1 ↑Pen. ↓Time ↓
a✓60.094.614.00.10±0.02
b✓✓64.396.512.50.10±0.02
c✓✓66.295.614.91.41±0.07
d✓✓✓67.596.511.90.10±0.02
e✓✓✓75.397.113.11.41±0.07
f✓✓✓✓76.898.413.01.41±0.07
", + "bbox": [ + 94, + 436, + 480, + 527 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/347377672a72be5f4464839c7b1b72eadd8df24e41a723651f7f4cf354a5cc67.jpg", + "image_caption": [ + "Figure 6. The variation of the mean grasping success rate of the evolving model for randomly selected objects in the multidex dataset with increasing finetuning epochs. The blue dashed line represents the grasping success rate under the original model." + ], + "image_footnote": [], + "bbox": [ + 119, + 531, + 454, + 696 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ference phase of this suboptimal model, we collect successfully grasping poses as positive samples and all other poses as negative samples. Fig. 6 illustrates the change in the average grasping success rate of the suboptimal model on randomly sampled objects from the multidex dataset as the number of fine-tuning epochs increases. The results indicate that we continuously improve the success rate during the evolutionary adaptation process and ultimately outper", + "bbox": [ + 89, + 780, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/515f9ebd32d356ba4083810aa5534e7907e440e62cbb961b1800640966305eb4.jpg", + "image_caption": [ + "Figure 7. Real-world deployment on Shadow Hand." + ], + "image_footnote": [], + "bbox": [ + 516, + 383, + 906, + 453 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "form the accuracy of the original model.", + "bbox": [ + 513, + 491, + 781, + 507 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Real World Deployment", + "text_level": 1, + "bbox": [ + 511, + 523, + 736, + 540 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To verify that EvolvingGrasp can continuously enhance grasping performance and align with human preferences in real world, we deploy our model on a real ShadowHand robot, as shown in Fig. 7. The pre-grasping motion trajectory is generated based on RealDex [20]. Real-world experiments demonstrate that our method achieves successful grasping through several efficient preference finetuning when the initial grasp attempt fails. Additional video demonstrations can be found in supplementary materials.", + "bbox": [ + 511, + 549, + 906, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 705, + 633, + 722 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We propose EvolvingGrasp, an evolutionary grasp generation method through efficient preference alignment. HPO is introduced to allow the model to continuously align the performance with preference signals. We design the Physics-Aware Consistency Model to achieve both rapid inference and efficient preference finetuning while maintaining physical plausibility. Extensive experiments across four benchmarks demonstrate that our method achieves state-of-the-art results. Furthermore, we deploy our model on a real ShadowHand robot to validate its evolutionary capability in real-world scenarios.", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "11672", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3", + "[2] Yuxin Chen, Devesh K Jha, Masayoshi Tomizuka, and Diego Romeres. Fdpp: Fine-tune diffusion policy with human preference. arXiv preprint arXiv:2501.08259, 2025. 2", + "[3] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 6", + "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3", + "[5] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2", + "[6] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, pages 7423-7433, 2024. 3", + "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6", + "[8] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858-79885, 2023. 3", + "[9] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, pages 2290-2295. IEEE, 1992. 2", + "[10] Ayano Hiranaka, Shang-Fu Chen, Chieh-Hsin Lai, Dongjun Kim, Naoki Murata, Takashi Shibuya, Wei-Hsiang Liao, Shao-Hua Sun, and Yuki Mitsufuji. Human-feedback efficient reinforcement learning for online diffusion model finetuning. arXiv preprint arXiv:2410.05116, 2024. 3", + "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5, 12", + "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, pages 16750-16761, 2023. 2, 6, 8", + "[13] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In ICCV, pages 11107-11116, 2021. 2, 6, 8", + "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 3" + ], + "bbox": [ + 93, + 114, + 482, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 2", + "[16] Puhao Li, Tengyu Liu, Yuyang Li, Yiran Geng, Yixin Zhu, Yaodong Yang, and Siyuan Huang. Gendexgrasp: Generalizable dexterous grasping. In IEEE International Conference on Robotics and Automation), pages 8068-8074. IEEE, 2023. 6", + "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Generating grasp poses for a high-dof gripper using neural networks. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1518-1525. IEEE, 2019. 2", + "[18] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2", + "[19] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2", + "[20] Yumeng Liu, Yaxun Yang, Youzhuo Wang, Xiaofei Wu, Ji-amin Wang, Yichen Yao, Soren Schwertfeger, Sibei Yang, Wenping Wang, Jingyi Yu, et al. Realdex: Towards human-like grasping for robotic dexterous hand. arXiv preprint arXiv:2402.13853, 2024. 6, 8", + "[21] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 3", + "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3", + "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3", + "[24] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. Ugg: Unified generative grasping. arXiv preprint arXiv:2311.16917, 2023. 2, 6, 8", + "[25] Jens Lundell, Enric Corona, Tran Nguyen Le, Francesco Verdoja, Philippe Weinzaepfel, Grégory Rogez, Francesc Moreno-Noguer, and Ville Kyrki. Multi-fingan: Generative coarse-to-fine sampling of multi-finger grasps. In IEEE International Conference on Robotics and Automation, pages 4495-4501. IEEE, 2021. 2", + "[26] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 6(4):6899-6906, 2021. 2", + "[27] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 6" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "11673", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13", + "[29] Jean Ponce, Steve Sullivan, J-D Boissonnat, and J-P Merlet. On characterizing and computing three-and four-finger force-closure grasps of polyhedral objects. In IEEE International Conference on Robotics and Automation, pages 821-827. IEEE, 1993. 2", + "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 3", + "[31] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2", + "[32] Carlos Rosales, Raul Suárez, Marco Gabiccini, and Antonio Bicchi. On the synthesis of feasible and prehensile robotic grasps. In IEEE International Conference on Robotics and Automation, pages 550-556. IEEE, 2012. 2", + "[33] Yanming Shao and Chenxi Xiao. Bimanual grasp synthesis for dexterous robot hands. IEEE Robotics and Automation Letters, 2024. 2", + "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3", + "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 5", + "[36] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6", + "[37] Xiaofeng Tan, Hongsong Wang, Xin Geng, and Pan Zhou. Sopo: Text-to-motion generation using semi-online preference optimization. arXiv preprint arXiv:2412.05095, 2024. 2", + "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 2, 3, 4", + "[39] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In IEEE International Conference on Robotics and Automation, pages 11359-11366. IEEE, 2023. 6", + "[40] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 7(2):1659-1666, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. arXiv preprint arXiv:2405.19291, 2024. 2", + "[42] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. IEEE Robotics and Automation Letters, 2024. 2", + "[43] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In CVPR, pages 17933-17942, 2024. 2", + "[44] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In CVPR, pages 4737-4746, 2023. 6, 8", + "[45] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, pages 8941-8951, 2024. 3", + "[46] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. arXiv preprint arXiv:2402.03201, 2024. 6", + "[47] Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen, Joel Jang, Yi Li, Chaoqi Wang, Mingyu Ding, Dieter Fox, and Huaxiu Yao. Grape: Generalizing robot policy via preference alignment. arXiv preprint arXiv:2411.19309, 2024. 2", + "[48] Zhengshen Zhang, Lei Zhou, Chenchen Liu, Zhiyang Liu, Chengran Yuan, Sheng Guo, Ruiteng Zhao, Marcelo H Ang Jr, and Francis EH Tay. Dexgrasp-diffusion: Diffusion-based unified functional grasp synthesis method for multideterous robotic hands. arXiv preprint arXiv:2407.09899, 2024. 2", + "[49] Yiming Zhong, Qi Jiang, Jingyi Yu, and Yuexin Ma. Dexterity: Towards universal robotic dexterous grasping with physics awareness. In CVPR, 2025. 2, 5, 6, 8, 13", + "[50] Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, and Tat-Seng Chua. Dreamdpo: Aligning text-to-3d generation with human preferences via direct preference optimization. arXiv preprint arXiv:2502.04370, 2025. 2" + ], + "bbox": [ + 514, + 92, + 903, + 684 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "11674", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_model.json b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bc8630ea841690f985eee0cdbbcce218109f335f --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_model.json @@ -0,0 +1,2145 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.236, + 0.131, + 0.761, + 0.176 + ], + "angle": 0, + "content": "EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment" + }, + { + "type": "text", + "bbox": [ + 0.232, + 0.203, + 0.766, + 0.275 + ], + "angle": 0, + "content": "Yufei Zhu\\(^{1,\\ast}\\), Yiming Zhong\\(^{1,\\ast}\\), Zemin Yang\\(^{1}\\), Peishan Cong\\(^{1}\\), Jingyi Yu\\(^{1}\\), Xinge Zhu\\(^{2,\\dagger}\\), Yuexin Ma\\(^{1,\\dagger}\\) \n\\(^{1}\\) ShanghaiTech University \\(^{2}\\) The Chinese University of Hong Kong https://evolvinggrasp.github.io/" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.314, + 0.546, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.56, + 0.318, + 0.902, + 0.579 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.581, + 0.908, + 0.611 + ], + "angle": 0, + "content": "Figure 1. The left part illustrates EvolvingGrasp, an approach akin to evolution, where it enables the model to learn from experience and iteratively refine its grasping strategy. The right part demonstrates its efficiency and effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.621, + 0.327, + 0.637 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.657, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Dexterous robotic hands often struggle to generalize effectively in complex environments due to models trained on low-diversity data. However, the real world presents an inherently unbounded range of scenarios. A natural solution is to enable robots learning from experience in complex environments—an approach akin to evolution, where systems improve through learning from both failures and successes. Motivated by this, we propose EvolvingGrasp, an evolutionary grasp generation method that continuously enhances grasping performance through efficient preference alignment. Specifically, we introduce Handpose-wise" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.623, + 0.907, + 0.804 + ], + "angle": 0, + "content": "Preference Optimization (HPO), which allows the model to continuously align with preferences from both positive and negative feedback while progressively refining its grasping strategies. To further enhance efficiency and reliability during online adjustments, we incorporate a Physics-aware Consistency Model within HPO, which accelerates inference, reduces the number of timesteps needed for preference fine-tuning, and ensures physical plausibility throughout the process. Our results validate that EvolvingGrasp enables evolutionary grasp generation, ensuring robust, physically feasible, and preference-aligned grasping in both simulation and real scenarios." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.842, + 0.646, + 0.858 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Dexterous robotic grasping has made significant strides in embodied manipulation, enabling more adaptive and pre" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.839, + 0.231, + 0.852 + ], + "angle": 0, + "content": "\\(^{1}\\) * Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.852, + 0.483, + 0.9 + ], + "angle": 0, + "content": "\\(^{2\\dagger}\\) Corresponding author. This work was supported by NSFC (No.62206173), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (KLIP-HuMaCo)." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.839, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11665" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.094, + 0.092, + 0.482, + 0.59 + ], + "angle": 0, + "content": "cise interactions compared to traditional grippers. Existing grasping methods [5, 9, 15, 17-19, 29, 31-33, 42, 48] generally fall into two categories: optimization-based approaches [5, 9, 29, 32], which refine hand poses to achieve force-closure states, learning-based approaches [26, 40, 41, 43], which directly map object features to grasp poses through regression, probabilistic modeling, and generative-based methods [12, 13, 24, 25, 42], which utilize diffusion model to estimate the distribution of hand poses. Recent advances, such as DexGrasp Anything [49], have further introduced physics-based constraints to improve grasp feasibility. However, a fundamental limitation persists—limited generalization. These methods, trained on limited datasets, struggle to adapt to complex environments. This challenge is exacerbated by an inherent property of the real world: its unbounded diversity. The vast range of object shapes, materials, and environmental conditions makes it impractical to predefine an exhaustive set of grasping strategies. Without the ability to adapt in deployment, grasping models remain constrained, failing to handle variations beyond their training distribution. To overcome this, a natural approach is to enable evolutionary grasp generation, where the system learns from experience (i.e., both failures and successes), refining its grasping strategy through iterative improvements based on real-world interactions. This process not only enhances generalization but also allows for preference alignment, ensuring that grasping behaviors adapt to task-specific requirements and user-defined preferences. However, achieving efficient evolutionary refinement is non-trivial, as many existing learning-based approaches rely on slow, compute-intensive updates, particularly in diffusion-based models that require numerous iterative steps and physics simulations." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.6, + 0.482, + 0.9 + ], + "angle": 0, + "content": "To address these challenges, we propose EvolvingGrasp, an evolutionary grasp generation framework that efficiently refines grasp strategies through preference alignment while maintaining physical plausibility. At its core, we introduce Handpose-wise Preference Optimization (HPO), a novel method that reformulates preference alignment [2, 37, 38, 47, 50] as a posterior probability optimization problem, encouraging generated grasps to converge toward preferred distributions while diverging from non-preferred ones. Notably, the proposed HPO is an extension of Direct Preference Optimization (DPO) [30], where it is also, to the best of our knowledge, the first to incorporate the DPO into the dexterous grasp. To further improve efficiency, we integrate HPO into a Physics-Aware Consistency Model (including two parts, i.e., Physics-Aware Distillation for training and Physics-Aware Sampling for inference), which pretrains a diffusion model and distills it into a lightweight, few-step sampling model. This enables both rapid inference and efficient preference fine-tuning, significantly reducing the number of required sampling steps and optimization iteraa" + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.182 + ], + "angle": 0, + "content": "tions. Additionally, we introduce three physics-aware constraints to ensure the stability, realism, and feasibility of generated grasp poses—surface pulling force to maintain stable contact, external penetration repulsion force to prevent object penetration, and self-penetration repulsion force to avoid inter-finger collisions." + }, + { + "type": "text", + "bbox": [ + 0.517, + 0.184, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Extensive experiments across four benchmark datasets demonstrate that EvolvingGrasp achieves state-of-the-art results, significantly improving grasp success rate, sampling efficiency, and physical plausibility, with 30x speedup over existing methods, demonstrating robust generalization across simulated and real-world benchmarks. Our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.292, + 0.904, + 0.351 + ], + "angle": 0, + "content": "- We introduce EvolvingGrasp, an efficient evolutionary grasp generation framework that enables iterative refinement, addressing the challenge of generalizing to diverse and unstructured real-world environments." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.353, + 0.905, + 0.427 + ], + "angle": 0, + "content": "- Efficient preference alignment is achieved through Handpose-wise Preference Optimization (HPO), which formulates grasp adaptation as a posterior probability optimization problem, enabling the model to iteratively converge toward preferred grasp distributions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.428, + 0.905, + 0.488 + ], + "angle": 0, + "content": "- We propose a Physics-Aware Consistency Model (PCM) that accelerates preference alignment by reducing sampling steps while enforcing geometric consistency and physical feasibility through structured constraints." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.489, + 0.905, + 0.578 + ], + "angle": 0, + "content": "- Extensive experiments across four benchmark datasets demonstrate that our method achieves state-of-the-art grasp success rates, physical plausibility, and sampling efficiency. Furthermore, it enables real-time grasp generation with minimal computational overhead, achieving comparable performance to gradient-based methods." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.292, + 0.905, + 0.578 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.597, + 0.654, + 0.612 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.517, + 0.623, + 0.703, + 0.638 + ], + "angle": 0, + "content": "2.1. Dexterous Grasping" + }, + { + "type": "text", + "bbox": [ + 0.516, + 0.645, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Dexterous grasp generation aims to produce high-quality grasping poses for robotic hands. Recent works can be categorized into regression-based [17, 18, 31] and generation-based methods [12, 13, 24-26, 40, 41]. Regression-based methods directly predict grasping parameters from the input object, but they often suffer from mode collapse issues that limits output diversity. Generation-based methods, though capable of producing varied solutions, face efficiency challenges. SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] require multiple sampling steps to generate diverse grasping poses, and DexGrasp Anything [49] additionally incurs computational overhead by calculating physical constraint losses and performing gradient updates at each sampling iteration. However, existing methods cannot generate grasping poses aligned with human habits due to their lack of evolutionary adaptation and preference alignment. They also struggle to balance physical plausibility" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11666" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.089, + 0.905, + 0.294 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.297, + 0.905, + 0.407 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.416, + 0.908, + 0.488 + ], + "angle": 0, + "content": "Figure 2. Overview of EvolvingGrasp. The evolutionary process begins with the human preference guidance, where Handpose-wise Preference Optimization (HPO, highlighted in the green rectangle) is employed to facilitate preference alignment. These grasp poses are generated by the Physics-Aware Consistency Model (shown in blue rectangles), including Sampling and Distillation mechanism, to ensure the sampling efficiency and the physical plausibility. In this way, EvolvingGrasp, an efficient evolutionary grasp generation framework is proposed to enable the grasp model iteratively converge toward preferred distributions." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.513, + 0.295, + 0.528 + ], + "angle": 0, + "content": "with computational efficiency." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.361, + 0.557 + ], + "angle": 0, + "content": "2.2. Accelerating Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.563, + 0.484, + 0.822 + ], + "angle": 0, + "content": "Accelerating dexterous grasping is vital for enhancing real-time adaptability. A series of diffusion-based acceleration methods have emerged, making real-time and efficient generation possible. DDIM [34] accelerates inference by reformulating the stochastic denoising process into a deterministic ODE solver. The DPM-Solver [22, 23] series further designs efficient high-order ODE solvers, achieving comparable quality with only \\(10 \\sim 20\\) steps. However, their performance degrades significantly when sampling steps are reduced to \\(2 \\sim 4\\). Consistency works [14, 21] support few-step sampling while preserving quality. CTM [14] extends CM to reduce cumulative errors and discretization inaccuracies. sCMs [21] unifies flow matching and diffusion frameworks to avoid discretization errors and hyperparameter tuning. Therefore, we employ the consistency model to enable efficient evolutionary refinement, reducing inference timesteps while maintaining result quality." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.832, + 0.462, + 0.849 + ], + "angle": 0, + "content": "2.3. Preference Finetuning for Diffusion Models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Preference alignment play a crucial role in enabling generative models to learn from user feedback, optimize generation strategies, and progressively improve output qual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.512, + 0.907, + 0.801 + ], + "angle": 0, + "content": "ity. Some researchers utilize it into diffusion models to better align with human preferences in image generation area. These methods can be categorized into two types, finetuning with reward model [1, 4, 6, 8] and reward-free finetuning [10, 38, 45]. The former includes DDPO [1] and DPOK [8] which treat the denoising process of diffusion models as an MDP and finetune using multiple reward models. Finetuning without reward model includes D3PO [45] and Diffusion-DPO [38], which extend the theoretical framework of DPO [30] to multi-step MDPs. These approaches learn an optimal reward model and then use it to refine the sampling strategy, making them a relatively more cost-effective alternative. However, RL-based fine-tuning for diffusion models still requires backpropagation at every sampled timestep, which is time-consuming. Applying this approach directly to grasping tasks is impractical for real-world applications. Therefore, we propose a faster preference alignment fine-tuning approach that reduces backpropagation steps." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.82, + 0.649, + 0.837 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.847, + 0.625, + 0.862 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "To achieve efficient evolutionary grasp generation that aligns with human preferences while maintaining both ef" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11667" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.395 + ], + "angle": 0, + "content": "ficiency and physical plausibility, we propose Evolving-Grasp, as illustrated in Fig. 2. The evolutionary process begins with preference alignment, where we introduce Handpose-Wise Preference Optimization (HPO) (Section 3.3). HPO formulates grasp adaptation as a posterior probability optimization problem, allowing the model to iteratively converge toward preferred grasp distributions. However, directly applying HPO suffers from slow sampling and inefficiencies in preference alignment, limiting its practicality for real-time grasp refinement. To address these challenges, we incorporate a Physics-Aware Consistency Model (PCM) (Section 3.4), which leverages a consistency model framework to enhance efficiency by reducing the number of required sampling steps. While this improves inference speed, naive consistency-based sampling may still generate physically implausible grasp poses. To ensure geometric consistency and physical feasibility, PCM integrates a physics-aware distillation and sampling mechanism, which enforces structured physical constraints on the generated poses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.404, + 0.295, + 0.42 + ], + "angle": 0, + "content": "3.2. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.426, + 0.484, + 0.534 + ], + "angle": 0, + "content": "Given the object point cloud representation \\( O \\in \\mathbb{R}^{N \\times 3} \\), our goal is to generate dexterous grasp poses with high success rate and low penetration from the posterior distribution \\( P(x \\mid O) \\), where \\( x = \\{x_{i}\\}_{i=1}^{n} \\). Specifically, the pose parameters contain three parts, joint angles of the hand \\( \\theta_{h} \\in \\mathbb{R}^{24} \\), global translation \\( T_{global} \\in \\mathbb{R}^{3} \\), and global rotation \\( R_{global} \\in SO(3) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.485, + 0.595 + ], + "angle": 0, + "content": "Given ground truth samples from the data distribution \\(\\pi(x_0)\\), the noise schedule weight \\(\\alpha\\), the goal of diffusion models is to fit the GT data distribution. The objective of training diffusion model is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.146, + 0.606, + 0.483, + 0.633 + ], + "angle": 0, + "content": "\\[\n\\mathbb {E} _ {t \\sim [ 0, T ], x _ {0}, \\epsilon \\sim N (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, O) \\| ^ {2} \\right] \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.689 + ], + "angle": 0, + "content": "where \\( x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon \\). Given the noise related parameter \\( \\sigma_{t} \\) and the mean from \\( x_{t} \\) to \\( x_{t-1} \\), the reverse process of diffusion model is as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.144, + 0.702, + 0.483, + 0.721 + ], + "angle": 0, + "content": "\\[\np _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right) \\tag {2}\n\\]" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.437, + 0.76 + ], + "angle": 0, + "content": "3.3. Handpose-wise Preference Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.903 + ], + "angle": 0, + "content": "To achieve preference alignment in the evolutionary adaptation process, we introduce the HPO, an extension of direct preference optimization (DPO). In DPO, the fundamental assumption is that we have access to data generated by the model, with human annotators providing the corresponding preferences. Specifically, for a given object \\( O \\), we observe pairs of grasp poses \\( x_0^w \\) and \\( x_0^l \\), where the set \\( x_0^w \\) is preferred over \\( x_0^l \\). These preferences can be modeled using the Bradley-Terry model, which expresses the probability" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.823, + 0.107 + ], + "angle": 0, + "content": "of one sample being preferred over another as:" + }, + { + "type": "equation", + "bbox": [ + 0.562, + 0.118, + 0.907, + 0.138 + ], + "angle": 0, + "content": "\\[\np _ {B T} \\left(x _ {0} ^ {w} \\succ x _ {0} ^ {l}\\right) = \\sigma \\left(r \\left(c, x _ {0} ^ {w}\\right) - r \\left(c, x _ {0} ^ {l}\\right)\\right) \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.15, + 0.907, + 0.211 + ], + "angle": 0, + "content": "The training objective of DPO is to maximize the likelihood of the observed preferences, which can be formulated as a binary classification problem. Specifically, the whole model is finetuned to minimize the following loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.223, + 0.906, + 0.277 + ], + "angle": 0, + "content": "\\[\nL _ {\\mathrm {B T}} (\\theta) = - \\mathbb {E} _ {\\left(x _ {0} ^ {w}, x _ {0} ^ {l}\\right) \\sim \\mathcal {D}} \\log \\sigma \\left( \\begin{array}{c} \\beta \\mathbb {E} \\\\ x _ {1: T} ^ {w} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {w} \\mid x _ {0} ^ {w}\\right) \\\\ x _ {1: T} ^ {l} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {l} \\mid x _ {0} ^ {l}\\right) \\end{array} \\right. \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.279, + 0.868, + 0.318 + ], + "angle": 0, + "content": "\\[\n\\left. \\left[ \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {w}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {l}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {l}\\right)} \\right]\\right)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.329, + 0.907, + 0.36 + ], + "angle": 0, + "content": "where \\(\\mathcal{D}\\) is the paired-wise grasp pose dataset. Following [38], we can get the upper bound of Eq. (4) as:" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.371, + 0.892, + 0.403 + ], + "angle": 0, + "content": "\\[\nL_{BT}(\\theta)\\leq -\\begin{array}{c}\\beta \\mathbb{E}\\\\ \\big(x^{w}_{0},x^{l}_{0}\\big)\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{w}_{n - 1,n}\\sim \\pi_{\\theta}\\big(x^{w}_{n - 1,n}|x^{w}_{0}\\big) \\end{array}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.404, + 0.905, + 0.458 + ], + "angle": 0, + "content": "\\[\n\\log \\sigma \\left(\\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}\\right) \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.905, + 0.489 + ], + "angle": 0, + "content": "where \\(\\pi_{\\theta}\\left(x_{n - 1} \\mid x_n\\right)\\) is the probability of sampling \\(x_{n - 1}\\) from \\(x_{n}\\), which can be computed by:" + }, + { + "type": "equation", + "bbox": [ + 0.522, + 0.501, + 0.906, + 0.536 + ], + "angle": 0, + "content": "\\[\n\\pi_ {\\theta} \\left(x _ {n - 1} \\mid x _ {n}\\right) = \\frac {1}{\\sqrt {2 \\pi} \\sigma_ {n}} \\exp \\left(- \\frac {\\left(x _ {n} - \\mu_ {\\theta} \\left(x _ {n} , n\\right)\\right) ^ {2}}{2 \\sigma_ {n} ^ {2}}\\right) \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.907, + 0.638 + ], + "angle": 0, + "content": "As far as we know, HPO is the first to integrate DPO into grasp pose generation, from which we extend DPO to a more flexible form. In HPO, there is no requirement to maintain an equal number of preferred and non-preferred grasp poses, which enables more adaptive preference learning. The objective of HPO is introduced as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.648, + 0.895, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal{L}_{HPO} = \\operatorname *{\\beta}\\limits_{\\substack{x^{i}_{0}\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{i}_{n - 1,n}\\sim \\pi_{\\theta}\\left(x^{i}_{n - 1,n}|x^{i}_{0}\\right)}}\\log \\sigma\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.528, + 0.679, + 0.905, + 0.735 + ], + "angle": 0, + "content": "\\[\n\\left(\\sum_ {i = 1} ^ {N _ {s u c}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)} - \\sum_ {j = 1} ^ {N _ {f a i l}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}\\right) \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.902 + ], + "angle": 0, + "content": "where \\( N_{\\text{succ}} \\) and \\( N_{\\text{fail}} \\) represent the number of preferred and non-preferred poses. HPO optimizes grasp pose generation by quantifying the probabilistic divergence between successful grasps (preferred samples) and failed grasps (non-preferred samples), thereby driving the model towards human-preferred behaviors. This process inherently involves the dynamic adjustment and maximization of reward signals: preferred grasp poses are reinforced due to their higher probability, while non-preferred poses are suppressed, guiding the model to progressively discard ineffective strategies." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11668" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.258 + ], + "angle": 0, + "content": "Preferred grasp selection can be conducted through either simulation-based evaluation or human-in-the-loop selection. In the simulation-based approach, poses that achieve success across all six directional evaluations are classified as preferred samples, while the remaining poses are treated as negative samples. Alternatively, in the human selection process, grasp poses that align with human intuition and habitual preferences are designated as preferred samples, while the others are considered non-preferred. Finally, we finetune the whole model via LoRA [11] using Eq. (7) to align the model with preferences." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.267, + 0.39, + 0.283 + ], + "angle": 0, + "content": "3.4. Physics-Aware Consistency Model" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.289, + 0.27, + 0.303 + ], + "angle": 0, + "content": "3.4.1. Consistency Model" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.308, + 0.484, + 0.476 + ], + "angle": 0, + "content": "Directly applying HPO faces two major efficiency challenges: generating poses requires over hundreds of timesteps per inference, and preference fine-tuning demands a large number of backpropagation steps. To address these inefficiencies, we adopt the consistency model framework to accelerate both sampling process and preference alignment. The core idea of the consistency model is to learn a mapping from any point along the ODE trajectory back to its starting point, which corresponds to the data distribution. Given a Probability Flow ODE trajectory [35] \\(\\{x_{\\tau_n}\\}_{\\tau_n\\in [0,T]}\\), a consistency model \\(f_{\\theta}\\) is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.209, + 0.488, + 0.482, + 0.504 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta}: \\left(x _ {\\tau_ {n}}, \\tau_ {n}, O\\right) \\mapsto x _ {0} \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.515, + 0.484, + 0.576 + ], + "angle": 0, + "content": "where \\( x_0 \\) is the initial point of the trajectory, \\( O \\) is the observation as condition. Due to space limitations, we omitted condition \\( O \\) in other parts of the paper, except for the appendix. The self-consistency property ensures that:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.585, + 0.483, + 0.603 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta^ {\\prime}} \\left(x _ {\\tau_ {n} ^ {\\prime}}, \\tau_ {n} ^ {\\prime}\\right) \\quad \\forall \\tau_ {n}, \\tau_ {n} ^ {\\prime} \\in [ 0, T ] \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.613, + 0.484, + 0.673 + ], + "angle": 0, + "content": "The consistency model needs to meet a key boundary condition: when \\( t = 0 \\), the model output should be the input itself, that is, \\( f_{\\theta}(x_0,0,O) = x_0 \\). This can be achieved as the following way:" + }, + { + "type": "equation", + "bbox": [ + 0.105, + 0.684, + 0.483, + 0.701 + ], + "angle": 0, + "content": "\\[\nf _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = c _ {\\text {s k i p}} \\left(\\tau_ {n}\\right) x _ {\\tau_ {n}} + c _ {\\text {o u t}} \\left(\\tau_ {n}\\right) F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.711, + 0.484, + 0.757 + ], + "angle": 0, + "content": "where \\( c_{skip}(t) \\) and \\( c_{out}(t) \\) are differentiable functions that satisfy \\( c_{skip}(\\epsilon) = 1 \\) and \\( c_{out}(\\epsilon) = 0 \\). \\( F_{\\theta}(x,t) \\) denotes a deep neural network that predicts \\( \\hat{x}_0 \\)." + }, + { + "type": "equation", + "bbox": [ + 0.107, + 0.766, + 0.482, + 0.81 + ], + "angle": 0, + "content": "\\[\nF _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {\\tau_ {n}}}} \\left(x _ {\\tau_ {n}} - \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n}}} \\epsilon_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right)\\right) \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Consistency distillation leverages pre-trained diffusion model to condense multi-step sampling into a more efficient few-step inference process. The process involves generating pairs of adjacent points on the PFODE trajectory using numerical ODE solvers and minimizing the difference between the model's outputs for these pairs. The loss function" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.092, + 0.78, + 0.106 + ], + "angle": 0, + "content": "for consistency distillation is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.532, + 0.117, + 0.905, + 0.143 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {C D} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), f _ {\\theta^ {\\prime}} \\left(\\hat {x} _ {\\tau_ {n - 1}} ^ {*}, \\tau_ {n - 1}\\right)\\right) \\right] \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.153, + 0.907, + 0.248 + ], + "angle": 0, + "content": "where \\(O\\) is omitted for simplicity, \\(\\hat{x}_{\\tau_{n-1}}^*\\) is computed using a numerical ODE solver, \\(d(\\cdot, \\cdot)\\) is a \\(L_2\\) distance metric, \\(\\theta\\) and \\(\\theta'\\) represent the online network and target network parameters. In addition, the parameters of \\(\\theta'\\) are updated by the exponential moving average (EMA) of the parameters of \\(\\theta\\). \\(\\hat{x}_{\\tau_{n-1}}^*\\) can be computed as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.258, + 0.905, + 0.278 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {\\tau_ {n - 1}} ^ {*} \\leftarrow \\sqrt {\\bar {\\alpha} _ {\\tau_ {n - 1}}} F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n - 1}}} \\epsilon \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.912, + 0.394 + ], + "angle": 0, + "content": "During sampling, we utilize the consistency function to directly generate the final sample and the quality of the generated sample can be enhanced through an iterative process that alternates between denoising and injecting noise. Given the sequence of timesteps \\( S \\in \\{\\tau_i \\mid \\tau_0 = 0, \\tau_{N-1} = T, \\tau_i < \\tau_{i+1} \\text{ for } i = 0, 1, \\dots, N-1\\} \\), the adding-noise process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.612, + 0.406, + 0.905, + 0.423 + ], + "angle": 0, + "content": "\\[\n\\hat {x} _ {\\tau_ {n - 1}} = \\mu_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sigma_ {\\tau_ {n}} \\epsilon \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.434, + 0.905, + 0.48 + ], + "angle": 0, + "content": "where \\(\\mu_{\\theta} = \\sqrt{\\bar{\\alpha}_{\\tau_{n - 1}}} f_{\\theta}(x_{\\tau_n},\\tau_n)\\) \\(\\sigma_{\\tau_n} = \\sqrt{1 - \\bar{\\alpha}_{\\tau_{n - 1}}}\\) Subsequently, we conduct the prediction of the final sample utilizing the trained consistency function once more." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.844, + 0.518 + ], + "angle": 0, + "content": "3.4.2. Physics-Aware Distillation and Sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.907, + 0.657 + ], + "angle": 0, + "content": "Although adopting the consistency model improves sampling and preference fine-tuning efficiency, it still generates physically implausible poses. To address this, we introduce a Physics-Aware Distillation and Sampling paradigm that enforces physical constraints during the distillation process of predicting \\(\\hat{x}_0\\) while ensuring that the sampling trajectories adhere to specific constraints. Following [49], we incorporate three physics-aware objectives, with the distillation objective of the consistency model formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.529, + 0.668, + 0.905, + 0.707 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {P A D} = \\mathcal {L} _ {C D} + \\sum_ {i = 1} ^ {m} \\alpha_ {i} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.907, + 0.87 + ], + "angle": 0, + "content": "where \\(L_{PA_i}\\left(F_\\theta(x_t, t), \\epsilon_\\theta\\right)\\) is the \\(i^{th}\\) physical constraint loss and \\(m = 3\\) denotes three constraints, i.e., Surface Pulling Force, External-penetration Repulsion Force, Self-Penetration Repulsion Force respectively [49]. These constraints guarantee grasping feasibility and maintain geometric accuracy in finger-to-object and inter-finger interactions. \\(\\alpha_i\\) is the corresponding weight parameter. We first train a diffusion model for grasp pose generation as a teacher model. Then, we utilize \\(\\mathcal{L}_{\\mathrm{PAD}}\\) to distill the teacher model into a student model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "During sampling, the consistency model estimates a clearer hand pose \\( x_0 \\) based on the current noisy hand" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11669" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.087, + 0.491, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.493, + 0.087, + 0.906, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.231, + 0.908, + 0.26 + ], + "angle": 0, + "content": "Figure 3. Mean grasping performance in terms of success rate and penetration of randomly selected 6 objects with the finetuning epoch increasing during inference optimization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.269, + 0.484, + 0.359 + ], + "angle": 0, + "content": "pose \\( x_{\\tau_n} \\) and object information \\( O \\) by consistency function \\( f_{\\theta}(x_{\\tau_n}, \\tau_n, O) \\). Subsequently, physical constraints are applied to steer the sampling process, making it closer to a physically feasible grasping pose. Following [3, 7, 46], the gradient of the constraint loss is used to modify the mean from \\( x_t \\) to \\( x_{t-1} \\):" + }, + { + "type": "equation", + "bbox": [ + 0.116, + 0.369, + 0.482, + 0.428 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\hat {\\mu} _ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) = \\mu_ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) \\\\ + \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {16} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.438, + 0.484, + 0.482 + ], + "angle": 0, + "content": "where \\(\\gamma_{i}\\) is the weight parameter corresponding to each physical constraint. Moreover, we can derive a new mapping from noise to data as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.101, + 0.493, + 0.482, + 0.544 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\frac {1}{\\sqrt {\\alpha_ {\\tau_ {n - 1}}}} \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\tag {17}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.546, + 0.484, + 0.621 + ], + "angle": 0, + "content": "By employing the physics-aware consistency model, we derive a novel preference alignment objective based on the new sampling path. The detailed derivation is provided in Appendix B. As a result, our method can efficiently generate higher-quality poses during the evolutionary process." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.634, + 0.224, + 0.65 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.659, + 0.483, + 0.825 + ], + "angle": 0, + "content": "We first simply present experimental setup which includes datasets, evaluation metrics and baselines in Sec. 4.3. More details of setup including datasets and implementation details are in Appendix C. Then we provide the quantitative and qualitative results of EvolvingGrasp in Sec. 4.2, followed by ablation study about different modules and hyperparameter analyses in Sec. 4.3. Next, we demonstrate the evolutionary improvement of EvolvingGrasp through an experiment starting from a suboptimal model trained on a degraded dataset in Sec. 4.4. Finally, We will demonstrate EvolvingGrasp in the real-world deployment in Sec. 4.5." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.834, + 0.282, + 0.85 + ], + "angle": 0, + "content": "4.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Datasets. We conduct experiment on four datasets, including DexGraspNet [39], Multidex [16], Realdex [20], Dex-GRAB [36] respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.907, + 0.48 + ], + "angle": 0, + "content": "Evaluation Metrics. We use four metrics to evaluate the grasping performance. Success rate Suc.6 measures the proportion of grasping poses where the object's displacement does not exceed \\(2\\mathrm{cm}\\) in all six axial directions \\((\\pm X, \\pm Y, \\pm Z)\\), evaluating multi-directional stability. Suc.1 measures the proportion where displacement does not exceed \\(2\\mathrm{cm}\\) in at least one direction, assessing single-direction stability. Pen. indicates the maximum penetration depth (mm) between the hand and the object, with lower values suggesting more physically plausible grasps. Above metrics are calculated in the IssacGym simulator [27] with settings consistent with SceneDiffuser [12]. For efficiency, Time refers to the computational time required to generate grasping poses for a batch of objects." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.482, + 0.907, + 0.543 + ], + "angle": 0, + "content": "Baselines. Some generative-based methods such as UniDexGrasp [44], GraspTTA [13], SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] are compared on four benchmark datasets." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.558, + 0.653, + 0.573 + ], + "angle": 0, + "content": "4.2. Main Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.582, + 0.907, + 0.839 + ], + "angle": 0, + "content": "To validate the effectiveness of our method in continuously enhancing grasping performance, we conduct quantitative evaluations on six randomly selected objects from each dataset, measuring both the Suc.6 metric and the degree of penetration. The results are illustrated in Fig. 3, indicating that as the number of inferences increases, the Suc.6 metric steadily improves. Although penetration exhibits some fluctuations due to increased contact area between the hand and the object, it still follows a downward trend. Additionally, we present qualitative results in Fig. 4, demonstrating how our method integrates human preferences during inference. This enables the selection of more favorable poses and progressively refines grasping poses to better align with human preferences through iteration. For instance, early generated poses may obstruct the camera lens, but after several rounds of preference-based finetuning, the model learns to produce poses that avoid blocking the lens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.907, + 0.902 + ], + "angle": 0, + "content": "We conduct the experiment with comparing with other dexterous grasp generation methods in Tab. 1. Compared to other generative-based approaches, our method excels in generating high-quality, physically plausible grasping" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11670" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.09, + 0.907, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.373, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Evolution of robotic grasp preferences during efficient feedback-driven finetuning across 10 epochs. Top row illustrates the adjustment from hand occlusion to clear nozzle visibility. Middle row demonstrates the transition from lens obstruction to an unobstructed camera view. The bottom row shows the evolution from a top-down grasping approach to a bottom-up one, while simultaneously mitigating physical impacts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.444, + 0.486, + 0.626 + ], + "angle": 0, + "content": "poses while significantly reducing computational time. It is worth highlighting that the proposed method achieves 30x speedup (32s~38s v.s. 0.73s~2.7s) compared to these SO-TAs. After fine-tuning, our approach achieves superior performance by leveraging its evolutionary capability through iterative refinement. We also evaluate the trade-off between efficiency and grasp quality during preference alignment with different numbers of inference steps. While increasing the steps improves performance, it inevitably incurs additional computational costs. Moreover, even without guidance during sampling, we can achieve comparable results in real time." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.637, + 0.245, + 0.654 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.659, + 0.483, + 0.795 + ], + "angle": 0, + "content": "We investigate the impact of different modules in the physics-aware consistency model, Physical constraints Guidance in Distillation and Sampling from which we dubbed as PGD and PGS. We also consider the performance of HPO with and without physical guidance during sampling on test split of the multidex dataset. The results presented in Tab. 2 demonstrate the key role of physical constraints and preference finetuning to successful grasping pose generation during preference alignment." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.796, + 0.483, + 0.902 + ], + "angle": 0, + "content": "HPO not only effectively aligns grasping poses with human preferences at the individual object level but also enhances performance across the entire dataset. During inference across the entire dataset, we collect both successful and failed grasping samples for each object, utilizing Eq. (5) to perform lightweight finetuning of the entire model. Fig. 5 illustrates the relationship between EvolvingGrasp's" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.441, + 0.878, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.633, + 0.907, + 0.663 + ], + "angle": 0, + "content": "Figure 5. The relationship between the success rate of Evolving-Grasp and the number of generated samples." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.68, + 0.907, + 0.756 + ], + "angle": 0, + "content": "Suc.6 metric and the number of generated samples, showing that our method continuously enhances grasping performance as more samples are generated. Further ablation studies on the impact of different hyperparameters are provided in Appendix D.2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.771, + 0.816, + 0.789 + ], + "angle": 0, + "content": "4.4. Training from a Degraded Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this subsection, we explore the feasibility of enhancing model performance through feedback-driven finetuning for single-object grasping tasks, particularly when trained on suboptimal data. To begin with, we randomly add the noise on the hand pose parameter of the original multidex dataset. This perturbation creates a degraded dataset, which we then use to train and distill a suboptimal model. During the in-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "11671" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 1. Grasping performance in terms of Suc.6, Suc.1, and Pen. comparison across different methods and datasets. \"Step\" refers to inference timestep. Bold values highlight the best results and underlined values indicate the runner-up." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.13, + 0.907, + 0.365 + ], + "angle": 0, + "content": "
Dataset MethodDexGraspNetMultiDexRealDexDexGRABTime ↓
Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓
UniDexGrasp [44]33.970.131.921.647.513.527.159.439.020.855.837.40.46±0.11
GraspTTA [13]18.667.824.530.362.819.013.346.440.114.451.051.410.41±0.32
SceneDiffuser [12]26.666.931.069.885.614.621.756.142.039.185.041.13.41±0.13
UGG [24]46.979.025.255.393.410.332.763.434.442.790.633.238.34±2.31
DexGrasp Any. [49]53.690.421.572.296.39.634.671.223.156.591.828.632.91±1.34
Ours w/o HPO (2-step)60.891.019.265.697.515.241.975.419.552.293.925.10.73±0.03
Ours (2-step)62.490.519.365.997.215.344.077.819.753.392.524.30.73±0.03
Ours w/o HPO (4-step)63.893.017.475.397.113.151.682.920.555.696.023.81.41±0.07
Ours (4-step)65.292.717.276.898.413.050.682.520.357.795.223.71.41±0.07
Ours w/o HPO (8-step)65.293.516.275.698.712.263.686.621.956.896.823.12.71±0.08
Ours (8-step)65.492.315.980.398.712.364.489.121.860.896.422.32.71±0.08
Real-time (2-step)55.290.520.063.795.013.846.578.921.448.993.324.80.06±0.01
Real-time (4-step)59.990.619.864.396.511.658.287.021.155.495.424.10.10±0.02
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.383, + 0.486, + 0.426 + ], + "angle": 0, + "content": "Table 2. Ablation study on incorporating physical constraints during both training and sampling stages and the LLM module. The evaluation is conducted on Multidex when timestep is 4." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.437, + 0.482, + 0.528 + ], + "angle": 0, + "content": "
CMPGDPGSHPOSuc.6 ↑Suc.1 ↑Pen. ↓Time ↓
a✓60.094.614.00.10±0.02
b✓✓64.396.512.50.10±0.02
c✓✓66.295.614.91.41±0.07
d✓✓✓67.596.511.90.10±0.02
e✓✓✓75.397.113.11.41±0.07
f✓✓✓✓76.898.413.01.41±0.07
" + }, + { + "type": "image", + "bbox": [ + 0.12, + 0.532, + 0.455, + 0.698 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.708, + 0.486, + 0.765 + ], + "angle": 0, + "content": "Figure 6. The variation of the mean grasping success rate of the evolving model for randomly selected objects in the multidex dataset with increasing finetuning epochs. The blue dashed line represents the grasping success rate under the original model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.485, + 0.902 + ], + "angle": 0, + "content": "ference phase of this suboptimal model, we collect successfully grasping poses as positive samples and all other poses as negative samples. Fig. 6 illustrates the change in the average grasping success rate of the suboptimal model on randomly sampled objects from the multidex dataset as the number of fine-tuning epochs increases. The results indicate that we continuously improve the success rate during the evolutionary adaptation process and ultimately outper" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.384, + 0.907, + 0.454 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.555, + 0.463, + 0.866, + 0.478 + ], + "angle": 0, + "content": "Figure 7. Real-world deployment on Shadow Hand." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.492, + 0.782, + 0.508 + ], + "angle": 0, + "content": "form the accuracy of the original model." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.525, + 0.738, + 0.541 + ], + "angle": 0, + "content": "4.5. Real World Deployment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.55, + 0.907, + 0.687 + ], + "angle": 0, + "content": "To verify that EvolvingGrasp can continuously enhance grasping performance and align with human preferences in real world, we deploy our model on a real ShadowHand robot, as shown in Fig. 7. The pre-grasping motion trajectory is generated based on RealDex [20]. Real-world experiments demonstrate that our method achieves successful grasping through several efficient preference finetuning when the initial grasp attempt fails. Additional video demonstrations can be found in supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.707, + 0.634, + 0.723 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "We propose EvolvingGrasp, an evolutionary grasp generation method through efficient preference alignment. HPO is introduced to allow the model to continuously align the performance with preference signals. We design the Physics-Aware Consistency Model to achieve both rapid inference and efficient preference finetuning while maintaining physical plausibility. Extensive experiments across four benchmarks demonstrate that our method achieves state-of-the-art results. Furthermore, we deploy our model on a real ShadowHand robot to validate its evolutionary capability in real-world scenarios." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "11672" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.482, + 0.157 + ], + "angle": 0, + "content": "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.159, + 0.482, + 0.2 + ], + "angle": 0, + "content": "[2] Yuxin Chen, Devesh K Jha, Masayoshi Tomizuka, and Diego Romeres. Fdpp: Fine-tune diffusion policy with human preference. arXiv preprint arXiv:2501.08259, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.202, + 0.483, + 0.257 + ], + "angle": 0, + "content": "[3] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.482, + 0.3 + ], + "angle": 0, + "content": "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.303, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[5] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.36, + 0.482, + 0.415 + ], + "angle": 0, + "content": "[6] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, pages 7423-7433, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.418, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.461, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[8] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858-79885, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[9] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, pages 2290-2295. IEEE, 1992. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.589, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[10] Ayano Hiranaka, Shang-Fu Chen, Chieh-Hsin Lai, Dongjun Kim, Naoki Murata, Takashi Shibuya, Wei-Hsiang Liao, Shao-Hua Sun, and Yuki Mitsufuji. Human-feedback efficient reinforcement learning for online diffusion model finetuning. arXiv preprint arXiv:2410.05116, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5, 12" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.718, + 0.482, + 0.772 + ], + "angle": 0, + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, pages 16750-16761, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[13] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In ICCV, pages 11107-11116, 2021. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.832, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.161 + ], + "angle": 0, + "content": "[15] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.164, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[16] Puhao Li, Tengyu Liu, Yuyang Li, Yiran Geng, Yixin Zhu, Yaodong Yang, and Siyuan Huang. Gendexgrasp: Generalizable dexterous grasping. In IEEE International Conference on Robotics and Automation), pages 8068-8074. IEEE, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.235, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Generating grasp poses for a high-dof gripper using neural networks. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1518-1525. IEEE, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[18] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.349, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[19] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.487 + ], + "angle": 0, + "content": "[20] Yumeng Liu, Yaxun Yang, Youzhuo Wang, Xiaofei Wu, Ji-amin Wang, Yichen Yao, Soren Schwertfeger, Sibei Yang, Wenping Wang, Jingyi Yu, et al. Realdex: Towards human-like grasping for robotic dexterous hand. arXiv preprint arXiv:2402.13853, 2024. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.49, + 0.905, + 0.531 + ], + "angle": 0, + "content": "[21] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.533, + 0.905, + 0.601 + ], + "angle": 0, + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.658 + ], + "angle": 0, + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.904, + 0.702 + ], + "angle": 0, + "content": "[24] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. Ugg: Unified generative grasping. arXiv preprint arXiv:2311.16917, 2023. 2, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.704, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[25] Jens Lundell, Enric Corona, Tran Nguyen Le, Francesco Verdoja, Philippe Weinzaepfel, Grégory Rogez, Francesc Moreno-Noguer, and Ville Kyrki. Multi-fingan: Generative coarse-to-fine sampling of multi-finger grasps. In IEEE International Conference on Robotics and Automation, pages 4495-4501. IEEE, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[26] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 6(4):6899-6906, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 6" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "11673" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.174 + ], + "angle": 0, + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Jean Ponce, Steve Sullivan, J-D Boissonnat, and J-P Merlet. On characterizing and computing three-and four-finger force-closure grasps of polyhedral objects. In IEEE International Conference on Robotics and Automation, pages 821-827. IEEE, 1993. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[31] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] Carlos Rosales, Raul Suárez, Marco Gabiccini, and Antonio Bicchi. On the synthesis of feasible and prehensile robotic grasps. In IEEE International Conference on Robotics and Automation, pages 550-556. IEEE, 2012. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[33] Yanming Shao and Chenxi Xiao. Bimanual grasp synthesis for dexterous robot hands. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.506, + 0.482, + 0.56 + ], + "angle": 0, + "content": "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.482, + 0.631 + ], + "angle": 0, + "content": "[36] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.634, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[37] Xiaofeng Tan, Hongsong Wang, Xin Geng, and Pan Zhou. Sopo: Text-to-motion generation using semi-online preference optimization. arXiv preprint arXiv:2412.05095, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.482, + 0.758 + ], + "angle": 0, + "content": "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.761, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[39] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In IEEE International Conference on Robotics and Automation, pages 11359-11366. IEEE, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[40] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 7(2):1659-1666, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. arXiv preprint arXiv:2405.19291, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[42] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. IEEE Robotics and Automation Letters, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[43] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In CVPR, pages 17933-17942, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.33 + ], + "angle": 0, + "content": "[44] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In CVPR, pages 4737-4746, 2023. 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.387 + ], + "angle": 0, + "content": "[45] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, pages 8941-8951, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.444 + ], + "angle": 0, + "content": "[46] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. arXiv preprint arXiv:2402.03201, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.447, + 0.905, + 0.501 + ], + "angle": 0, + "content": "[47] Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen, Joel Jang, Yi Li, Chaoqi Wang, Mingyu Ding, Dieter Fox, and Huaxiu Yao. Grape: Generalizing robot policy via preference alignment. arXiv preprint arXiv:2411.19309, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.905, + 0.584 + ], + "angle": 0, + "content": "[48] Zhengshen Zhang, Lei Zhou, Chenchen Liu, Zhiyang Liu, Chengran Yuan, Sheng Guo, Ruiteng Zhao, Marcelo H Ang Jr, and Francis EH Tay. Dexgrasp-diffusion: Diffusion-based unified functional grasp synthesis method for multideterous robotic hands. arXiv preprint arXiv:2407.09899, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.628 + ], + "angle": 0, + "content": "[49] Yiming Zhong, Qi Jiang, Jingyi Yu, and Yuexin Ma. Dexterity: Towards universal robotic dexterous grasping with physics awareness. In CVPR, 2025. 2, 5, 6, 8, 13" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.63, + 0.905, + 0.685 + ], + "angle": 0, + "content": "[50] Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, and Tat-Seng Chua. Dreamdpo: Aligning text-to-3d generation with human preferences via direct preference optimization. arXiv preprint arXiv:2502.04370, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.905, + 0.685 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "11674" + } + ] +] \ No newline at end of file diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_origin.pdf b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e674f7c32bf93abe20533bceed90cf3db714049d --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/a5b9e976-b527-49c5-a219-2287871511f8_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:afdaf7fef5aef9ff6c6864799d2aee959f2db81aaa1ea46250ad661e0e396a79 +size 22115737 diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/full.md b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d0b6c1b3a238790f285adf6d2c9a98363fe0232e --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/full.md @@ -0,0 +1,329 @@ +# EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment + +Yufei Zhu $^{1,\ast}$ , Yiming Zhong $^{1,\ast}$ , Zemin Yang $^{1}$ , Peishan Cong $^{1}$ , Jingyi Yu $^{1}$ , Xinge Zhu $^{2,\dagger}$ , Yuexin Ma $^{1,\dagger}$ $^{1}$ ShanghaiTech University $^{2}$ The Chinese University of Hong Kong https://evolvinggrasp.github.io/ + +![](images/d4655f9aa5ef7b2023e42f5999c7209ff12a76e6328a513bdf7203923f33cfcf.jpg) +Figure 1. The left part illustrates EvolvingGrasp, an approach akin to evolution, where it enables the model to learn from experience and iteratively refine its grasping strategy. The right part demonstrates its efficiency and effectiveness. + +![](images/46c6c0a21e9df52e097391058e722c6db16a390b7bac5b7b1c3bfd065838c04e.jpg) + +# Abstract + +Dexterous robotic hands often struggle to generalize effectively in complex environments due to models trained on low-diversity data. However, the real world presents an inherently unbounded range of scenarios. A natural solution is to enable robots learning from experience in complex environments—an approach akin to evolution, where systems improve through learning from both failures and successes. Motivated by this, we propose EvolvingGrasp, an evolutionary grasp generation method that continuously enhances grasping performance through efficient preference alignment. Specifically, we introduce Handpose-wise + +Preference Optimization (HPO), which allows the model to continuously align with preferences from both positive and negative feedback while progressively refining its grasping strategies. To further enhance efficiency and reliability during online adjustments, we incorporate a Physics-aware Consistency Model within HPO, which accelerates inference, reduces the number of timesteps needed for preference fine-tuning, and ensures physical plausibility throughout the process. Our results validate that EvolvingGrasp enables evolutionary grasp generation, ensuring robust, physically feasible, and preference-aligned grasping in both simulation and real scenarios. + +# 1. Introduction + +Dexterous robotic grasping has made significant strides in embodied manipulation, enabling more adaptive and pre + +cise interactions compared to traditional grippers. Existing grasping methods [5, 9, 15, 17-19, 29, 31-33, 42, 48] generally fall into two categories: optimization-based approaches [5, 9, 29, 32], which refine hand poses to achieve force-closure states, learning-based approaches [26, 40, 41, 43], which directly map object features to grasp poses through regression, probabilistic modeling, and generative-based methods [12, 13, 24, 25, 42], which utilize diffusion model to estimate the distribution of hand poses. Recent advances, such as DexGrasp Anything [49], have further introduced physics-based constraints to improve grasp feasibility. However, a fundamental limitation persists—limited generalization. These methods, trained on limited datasets, struggle to adapt to complex environments. This challenge is exacerbated by an inherent property of the real world: its unbounded diversity. The vast range of object shapes, materials, and environmental conditions makes it impractical to predefine an exhaustive set of grasping strategies. Without the ability to adapt in deployment, grasping models remain constrained, failing to handle variations beyond their training distribution. To overcome this, a natural approach is to enable evolutionary grasp generation, where the system learns from experience (i.e., both failures and successes), refining its grasping strategy through iterative improvements based on real-world interactions. This process not only enhances generalization but also allows for preference alignment, ensuring that grasping behaviors adapt to task-specific requirements and user-defined preferences. However, achieving efficient evolutionary refinement is non-trivial, as many existing learning-based approaches rely on slow, compute-intensive updates, particularly in diffusion-based models that require numerous iterative steps and physics simulations. + +To address these challenges, we propose EvolvingGrasp, an evolutionary grasp generation framework that efficiently refines grasp strategies through preference alignment while maintaining physical plausibility. At its core, we introduce Handpose-wise Preference Optimization (HPO), a novel method that reformulates preference alignment [2, 37, 38, 47, 50] as a posterior probability optimization problem, encouraging generated grasps to converge toward preferred distributions while diverging from non-preferred ones. Notably, the proposed HPO is an extension of Direct Preference Optimization (DPO) [30], where it is also, to the best of our knowledge, the first to incorporate the DPO into the dexterous grasp. To further improve efficiency, we integrate HPO into a Physics-Aware Consistency Model (including two parts, i.e., Physics-Aware Distillation for training and Physics-Aware Sampling for inference), which pretrains a diffusion model and distills it into a lightweight, few-step sampling model. This enables both rapid inference and efficient preference fine-tuning, significantly reducing the number of required sampling steps and optimization iteraa + +tions. Additionally, we introduce three physics-aware constraints to ensure the stability, realism, and feasibility of generated grasp poses—surface pulling force to maintain stable contact, external penetration repulsion force to prevent object penetration, and self-penetration repulsion force to avoid inter-finger collisions. + +Extensive experiments across four benchmark datasets demonstrate that EvolvingGrasp achieves state-of-the-art results, significantly improving grasp success rate, sampling efficiency, and physical plausibility, with 30x speedup over existing methods, demonstrating robust generalization across simulated and real-world benchmarks. Our contributions can be summarized as follows: + +- We introduce EvolvingGrasp, an efficient evolutionary grasp generation framework that enables iterative refinement, addressing the challenge of generalizing to diverse and unstructured real-world environments. +- Efficient preference alignment is achieved through Handpose-wise Preference Optimization (HPO), which formulates grasp adaptation as a posterior probability optimization problem, enabling the model to iteratively converge toward preferred grasp distributions. +- We propose a Physics-Aware Consistency Model (PCM) that accelerates preference alignment by reducing sampling steps while enforcing geometric consistency and physical feasibility through structured constraints. +- Extensive experiments across four benchmark datasets demonstrate that our method achieves state-of-the-art grasp success rates, physical plausibility, and sampling efficiency. Furthermore, it enables real-time grasp generation with minimal computational overhead, achieving comparable performance to gradient-based methods. + +# 2. Related Work + +# 2.1. Dexterous Grasping + +Dexterous grasp generation aims to produce high-quality grasping poses for robotic hands. Recent works can be categorized into regression-based [17, 18, 31] and generation-based methods [12, 13, 24-26, 40, 41]. Regression-based methods directly predict grasping parameters from the input object, but they often suffer from mode collapse issues that limits output diversity. Generation-based methods, though capable of producing varied solutions, face efficiency challenges. SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] require multiple sampling steps to generate diverse grasping poses, and DexGrasp Anything [49] additionally incurs computational overhead by calculating physical constraint losses and performing gradient updates at each sampling iteration. However, existing methods cannot generate grasping poses aligned with human habits due to their lack of evolutionary adaptation and preference alignment. They also struggle to balance physical plausibility + +![](images/57c83978e95fa440daea1bee7421ad3f6a20fc8a0f44a19eaabb6e9f4333761e.jpg) + +![](images/887f1e1bf94a34a3990de5441bba2b5321063f5966e68198a3406d5e2caf7aaa.jpg) +Figure 2. Overview of EvolvingGrasp. The evolutionary process begins with the human preference guidance, where Handpose-wise Preference Optimization (HPO, highlighted in the green rectangle) is employed to facilitate preference alignment. These grasp poses are generated by the Physics-Aware Consistency Model (shown in blue rectangles), including Sampling and Distillation mechanism, to ensure the sampling efficiency and the physical plausibility. In this way, EvolvingGrasp, an efficient evolutionary grasp generation framework is proposed to enable the grasp model iteratively converge toward preferred distributions. + +with computational efficiency. + +# 2.2. Accelerating Diffusion Models + +Accelerating dexterous grasping is vital for enhancing real-time adaptability. A series of diffusion-based acceleration methods have emerged, making real-time and efficient generation possible. DDIM [34] accelerates inference by reformulating the stochastic denoising process into a deterministic ODE solver. The DPM-Solver [22, 23] series further designs efficient high-order ODE solvers, achieving comparable quality with only $10 \sim 20$ steps. However, their performance degrades significantly when sampling steps are reduced to $2 \sim 4$ . Consistency works [14, 21] support few-step sampling while preserving quality. CTM [14] extends CM to reduce cumulative errors and discretization inaccuracies. sCMs [21] unifies flow matching and diffusion frameworks to avoid discretization errors and hyperparameter tuning. Therefore, we employ the consistency model to enable efficient evolutionary refinement, reducing inference timesteps while maintaining result quality. + +# 2.3. Preference Finetuning for Diffusion Models + +Preference alignment play a crucial role in enabling generative models to learn from user feedback, optimize generation strategies, and progressively improve output qual + +ity. Some researchers utilize it into diffusion models to better align with human preferences in image generation area. These methods can be categorized into two types, finetuning with reward model [1, 4, 6, 8] and reward-free finetuning [10, 38, 45]. The former includes DDPO [1] and DPOK [8] which treat the denoising process of diffusion models as an MDP and finetune using multiple reward models. Finetuning without reward model includes D3PO [45] and Diffusion-DPO [38], which extend the theoretical framework of DPO [30] to multi-step MDPs. These approaches learn an optimal reward model and then use it to refine the sampling strategy, making them a relatively more cost-effective alternative. However, RL-based fine-tuning for diffusion models still requires backpropagation at every sampled timestep, which is time-consuming. Applying this approach directly to grasping tasks is impractical for real-world applications. Therefore, we propose a faster preference alignment fine-tuning approach that reduces backpropagation steps. + +# 3. Methodology + +# 3.1. Overview + +To achieve efficient evolutionary grasp generation that aligns with human preferences while maintaining both ef + +ficiency and physical plausibility, we propose Evolving-Grasp, as illustrated in Fig. 2. The evolutionary process begins with preference alignment, where we introduce Handpose-Wise Preference Optimization (HPO) (Section 3.3). HPO formulates grasp adaptation as a posterior probability optimization problem, allowing the model to iteratively converge toward preferred grasp distributions. However, directly applying HPO suffers from slow sampling and inefficiencies in preference alignment, limiting its practicality for real-time grasp refinement. To address these challenges, we incorporate a Physics-Aware Consistency Model (PCM) (Section 3.4), which leverages a consistency model framework to enhance efficiency by reducing the number of required sampling steps. While this improves inference speed, naive consistency-based sampling may still generate physically implausible grasp poses. To ensure geometric consistency and physical feasibility, PCM integrates a physics-aware distillation and sampling mechanism, which enforces structured physical constraints on the generated poses. + +# 3.2. Problem Formulation + +Given the object point cloud representation $O \in \mathbb{R}^{N \times 3}$ , our goal is to generate dexterous grasp poses with high success rate and low penetration from the posterior distribution $P(x \mid O)$ , where $x = \{x_{i}\}_{i=1}^{n}$ . Specifically, the pose parameters contain three parts, joint angles of the hand $\theta_{h} \in \mathbb{R}^{24}$ , global translation $T_{global} \in \mathbb{R}^{3}$ , and global rotation $R_{global} \in SO(3)$ . + +Given ground truth samples from the data distribution $\pi(x_0)$ , the noise schedule weight $\alpha$ , the goal of diffusion models is to fit the GT data distribution. The objective of training diffusion model is as follows: + +$$ +\mathbb {E} _ {t \sim [ 0, T ], x _ {0}, \epsilon \sim N (0, I)} \left[ \| \epsilon - \epsilon_ {\theta} (x _ {t}, t, O) \| ^ {2} \right] \tag {1} +$$ + +where $x_{t} = \sqrt{\bar{\alpha}_{t}} x_{0} + \sqrt{1 - \bar{\alpha}_{t}}\epsilon$ . Given the noise related parameter $\sigma_{t}$ and the mean from $x_{t}$ to $x_{t-1}$ , the reverse process of diffusion model is as follows: + +$$ +p _ {\theta} \left(x _ {t - 1} \mid x _ {t}\right) = \mathcal {N} \left(x _ {t - 1}; \mu_ {\theta} \left(x _ {t}, t\right), \sigma_ {t} ^ {2} \mathrm {I}\right) \tag {2} +$$ + +# 3.3. Handpose-wise Preference Optimization + +To achieve preference alignment in the evolutionary adaptation process, we introduce the HPO, an extension of direct preference optimization (DPO). In DPO, the fundamental assumption is that we have access to data generated by the model, with human annotators providing the corresponding preferences. Specifically, for a given object $O$ , we observe pairs of grasp poses $x_0^w$ and $x_0^l$ , where the set $x_0^w$ is preferred over $x_0^l$ . These preferences can be modeled using the Bradley-Terry model, which expresses the probability + +of one sample being preferred over another as: + +$$ +p _ {B T} \left(x _ {0} ^ {w} \succ x _ {0} ^ {l}\right) = \sigma \left(r \left(c, x _ {0} ^ {w}\right) - r \left(c, x _ {0} ^ {l}\right)\right) \tag {3} +$$ + +The training objective of DPO is to maximize the likelihood of the observed preferences, which can be formulated as a binary classification problem. Specifically, the whole model is finetuned to minimize the following loss function: + +$$ +L _ {\mathrm {B T}} (\theta) = - \mathbb {E} _ {\left(x _ {0} ^ {w}, x _ {0} ^ {l}\right) \sim \mathcal {D}} \log \sigma \left( \begin{array}{c} \beta \mathbb {E} \\ x _ {1: T} ^ {w} \sim \pi_ {\theta} \left(x _ {1: T} ^ {w} \mid x _ {0} ^ {w}\right) \\ x _ {1: T} ^ {l} \sim \pi_ {\theta} \left(x _ {1: T} ^ {l} \mid x _ {0} ^ {l}\right) \end{array} \right. \tag {4} +$$ + +$$ +\left. \left[ \log \frac {\pi_ {\theta} \left(x _ {0 : T} ^ {w}\right)}{\pi_ {\mathrm {r e f}} \left(x _ {0 : T} ^ {w}\right)} - \log \frac {\pi_ {\theta} \left(x _ {0 : T} ^ {l}\right)}{\pi_ {\mathrm {r e f}} \left(x _ {0 : T} ^ {l}\right)} \right]\right) +$$ + +where $\mathcal{D}$ is the paired-wise grasp pose dataset. Following [38], we can get the upper bound of Eq. (4) as: + +$$ +L_{BT}(\theta)\leq -\begin{array}{c}\beta \mathbb{E}\\ \big(x^{w}_{0},x^{l}_{0}\big)\sim \mathcal{D},n\sim \mathcal{U}(1,N),x^{w}_{n - 1,n}\sim \pi_{\theta}\big(x^{w}_{n - 1,n}|x^{w}_{0}\big) \end{array} +$$ + +$$ +\log \sigma \left(\log \frac {\pi_ {\theta} \left(x _ {n - 1} ^ {w} \mid x _ {n} ^ {w}\right)}{\pi_ {\text {r e f}} \left(x _ {n - 1} ^ {w} \mid x _ {n} ^ {w}\right)} - \log \frac {\pi_ {\theta} \left(x _ {n - 1} ^ {l} \mid x _ {n} ^ {l}\right)}{\pi_ {\text {r e f}} \left(x _ {n - 1} ^ {l} \mid x _ {n} ^ {l}\right)}\right) \tag {5} +$$ + +where $\pi_{\theta}\left(x_{n - 1} \mid x_n\right)$ is the probability of sampling $x_{n - 1}$ from $x_{n}$ , which can be computed by: + +$$ +\pi_ {\theta} \left(x _ {n - 1} \mid x _ {n}\right) = \frac {1}{\sqrt {2 \pi} \sigma_ {n}} \exp \left(- \frac {\left(x _ {n} - \mu_ {\theta} \left(x _ {n} , n\right)\right) ^ {2}}{2 \sigma_ {n} ^ {2}}\right) \tag {6} +$$ + +As far as we know, HPO is the first to integrate DPO into grasp pose generation, from which we extend DPO to a more flexible form. In HPO, there is no requirement to maintain an equal number of preferred and non-preferred grasp poses, which enables more adaptive preference learning. The objective of HPO is introduced as follows: + +$$ +\mathcal{L}_{HPO} = \operatorname *{\beta}\limits_{\substack{x^{i}_{0}\sim \mathcal{D},n\sim \mathcal{U}(1,N),x^{i}_{n - 1,n}\sim \pi_{\theta}\left(x^{i}_{n - 1,n}|x^{i}_{0}\right)}}\log \sigma +$$ + +$$ +\left(\sum_ {i = 1} ^ {N _ {s u c}} \log \frac {\pi_ {\theta} \left(x _ {n - 1} ^ {i} \mid x _ {n} ^ {i}\right)}{\pi_ {\text {r e f}} \left(x _ {n - 1} ^ {i} \mid x _ {n} ^ {i}\right)} - \sum_ {j = 1} ^ {N _ {f a i l}} \log \frac {\pi_ {\theta} \left(x _ {n - 1} ^ {j} \mid x _ {n} ^ {j}\right)}{\pi_ {\text {r e f}} \left(x _ {n - 1} ^ {j} \mid x _ {n} ^ {j}\right)}\right) \tag {7} +$$ + +where $N_{\text{succ}}$ and $N_{\text{fail}}$ represent the number of preferred and non-preferred poses. HPO optimizes grasp pose generation by quantifying the probabilistic divergence between successful grasps (preferred samples) and failed grasps (non-preferred samples), thereby driving the model towards human-preferred behaviors. This process inherently involves the dynamic adjustment and maximization of reward signals: preferred grasp poses are reinforced due to their higher probability, while non-preferred poses are suppressed, guiding the model to progressively discard ineffective strategies. + +Preferred grasp selection can be conducted through either simulation-based evaluation or human-in-the-loop selection. In the simulation-based approach, poses that achieve success across all six directional evaluations are classified as preferred samples, while the remaining poses are treated as negative samples. Alternatively, in the human selection process, grasp poses that align with human intuition and habitual preferences are designated as preferred samples, while the others are considered non-preferred. Finally, we finetune the whole model via LoRA [11] using Eq. (7) to align the model with preferences. + +# 3.4. Physics-Aware Consistency Model + +# 3.4.1. Consistency Model + +Directly applying HPO faces two major efficiency challenges: generating poses requires over hundreds of timesteps per inference, and preference fine-tuning demands a large number of backpropagation steps. To address these inefficiencies, we adopt the consistency model framework to accelerate both sampling process and preference alignment. The core idea of the consistency model is to learn a mapping from any point along the ODE trajectory back to its starting point, which corresponds to the data distribution. Given a Probability Flow ODE trajectory [35] $\{x_{\tau_n}\}_{\tau_n\in [0,T]}$ , a consistency model $f_{\theta}$ is defined as: + +$$ +f _ {\theta}: \left(x _ {\tau_ {n}}, \tau_ {n}, O\right) \mapsto x _ {0} \tag {8} +$$ + +where $x_0$ is the initial point of the trajectory, $O$ is the observation as condition. Due to space limitations, we omitted condition $O$ in other parts of the paper, except for the appendix. The self-consistency property ensures that: + +$$ +f _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) = f _ {\theta^ {\prime}} \left(x _ {\tau_ {n} ^ {\prime}}, \tau_ {n} ^ {\prime}\right) \quad \forall \tau_ {n}, \tau_ {n} ^ {\prime} \in [ 0, T ] \tag {9} +$$ + +The consistency model needs to meet a key boundary condition: when $t = 0$ , the model output should be the input itself, that is, $f_{\theta}(x_0,0,O) = x_0$ . This can be achieved as the following way: + +$$ +f _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) = c _ {\text {s k i p}} \left(\tau_ {n}\right) x _ {\tau_ {n}} + c _ {\text {o u t}} \left(\tau_ {n}\right) F _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) \tag {10} +$$ + +where $c_{skip}(t)$ and $c_{out}(t)$ are differentiable functions that satisfy $c_{skip}(\epsilon) = 1$ and $c_{out}(\epsilon) = 0$ . $F_{\theta}(x,t)$ denotes a deep neural network that predicts $\hat{x}_0$ . + +$$ +F _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) = \frac {1}{\sqrt {\bar {\alpha} _ {\tau_ {n}}}} \left(x _ {\tau_ {n}} - \sqrt {1 - \bar {\alpha} _ {\tau_ {n}}} \epsilon_ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right)\right) \tag {11} +$$ + +Consistency distillation leverages pre-trained diffusion model to condense multi-step sampling into a more efficient few-step inference process. The process involves generating pairs of adjacent points on the PFODE trajectory using numerical ODE solvers and minimizing the difference between the model's outputs for these pairs. The loss function + +for consistency distillation is defined as: + +$$ +\mathcal {L} _ {C D} = \mathbb {E} \left[ d \left(f _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right), f _ {\theta^ {\prime}} \left(\hat {x} _ {\tau_ {n - 1}} ^ {*}, \tau_ {n - 1}\right)\right) \right] \tag {12} +$$ + +where $O$ is omitted for simplicity, $\hat{x}_{\tau_{n-1}}^*$ is computed using a numerical ODE solver, $d(\cdot, \cdot)$ is a $L_2$ distance metric, $\theta$ and $\theta'$ represent the online network and target network parameters. In addition, the parameters of $\theta'$ are updated by the exponential moving average (EMA) of the parameters of $\theta$ . $\hat{x}_{\tau_{n-1}}^*$ can be computed as: + +$$ +\hat {x} _ {\tau_ {n - 1}} ^ {*} \leftarrow \sqrt {\bar {\alpha} _ {\tau_ {n - 1}}} F _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) + \sqrt {1 - \bar {\alpha} _ {\tau_ {n - 1}}} \epsilon \tag {13} +$$ + +During sampling, we utilize the consistency function to directly generate the final sample and the quality of the generated sample can be enhanced through an iterative process that alternates between denoising and injecting noise. Given the sequence of timesteps $S \in \{\tau_i \mid \tau_0 = 0, \tau_{N-1} = T, \tau_i < \tau_{i+1} \text{ for } i = 0, 1, \dots, N-1\}$ , the adding-noise process can be formulated as: + +$$ +\hat {x} _ {\tau_ {n - 1}} = \mu_ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) + \sigma_ {\tau_ {n}} \epsilon \tag {14} +$$ + +where $\mu_{\theta} = \sqrt{\bar{\alpha}_{\tau_{n - 1}}} f_{\theta}(x_{\tau_n},\tau_n)$ $\sigma_{\tau_n} = \sqrt{1 - \bar{\alpha}_{\tau_{n - 1}}}$ Subsequently, we conduct the prediction of the final sample utilizing the trained consistency function once more. + +# 3.4.2. Physics-Aware Distillation and Sampling + +Although adopting the consistency model improves sampling and preference fine-tuning efficiency, it still generates physically implausible poses. To address this, we introduce a Physics-Aware Distillation and Sampling paradigm that enforces physical constraints during the distillation process of predicting $\hat{x}_0$ while ensuring that the sampling trajectories adhere to specific constraints. Following [49], we incorporate three physics-aware objectives, with the distillation objective of the consistency model formulated as follows: + +$$ +\mathcal {L} _ {P A D} = \mathcal {L} _ {C D} + \sum_ {i = 1} ^ {m} \alpha_ {i} L _ {P A _ {i}} \left(F _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right), \epsilon_ {\theta}\right) \tag {15} +$$ + +where $L_{PA_i}\left(F_\theta(x_t, t), \epsilon_\theta\right)$ is the $i^{th}$ physical constraint loss and $m = 3$ denotes three constraints, i.e., Surface Pulling Force, External-penetration Repulsion Force, Self-Penetration Repulsion Force respectively [49]. These constraints guarantee grasping feasibility and maintain geometric accuracy in finger-to-object and inter-finger interactions. $\alpha_i$ is the corresponding weight parameter. We first train a diffusion model for grasp pose generation as a teacher model. Then, we utilize $\mathcal{L}_{\mathrm{PAD}}$ to distill the teacher model into a student model. + +During sampling, the consistency model estimates a clearer hand pose $x_0$ based on the current noisy hand + +![](images/6e1628acfa9f5a828e02c518a23a04c4075c5e239503df446d5f57fe47f10145.jpg) +Figure 3. Mean grasping performance in terms of success rate and penetration of randomly selected 6 objects with the finetuning epoch increasing during inference optimization. + +![](images/ecba72db4560fdacdd30b98d254c24d670062dc788dd5e4e07ad8c504a333d26.jpg) + +pose $x_{\tau_n}$ and object information $O$ by consistency function $f_{\theta}(x_{\tau_n}, \tau_n, O)$ . Subsequently, physical constraints are applied to steer the sampling process, making it closer to a physically feasible grasping pose. Following [3, 7, 46], the gradient of the constraint loss is used to modify the mean from $x_t$ to $x_{t-1}$ : + +$$ +\begin{array}{l} \hat {\mu} _ {\theta} (x _ {\tau_ {n}}, \tau_ {n}) = \mu_ {\theta} (x _ {\tau_ {n}}, \tau_ {n}) \\ + \sum_ {i = 1} ^ {m} \gamma_ {i} \nabla_ {x _ {\tau_ {n}}} L _ {P A _ {i}} \left(F _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right), \epsilon_ {\theta}\right) \tag {16} \\ \end{array} +$$ + +where $\gamma_{i}$ is the weight parameter corresponding to each physical constraint. Moreover, we can derive a new mapping from noise to data as follows: + +$$ +\hat {f} _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) = f _ {\theta} \left(x _ {\tau_ {n}}, \tau_ {n}\right) + \frac {1}{\sqrt {\alpha_ {\tau_ {n - 1}}}} \sum_ {i = 1} ^ {m} \gamma_ {i} \nabla_ {x _ {\tau_ {n}}} L _ {P A _ {i}} \tag {17} +$$ + +By employing the physics-aware consistency model, we derive a novel preference alignment objective based on the new sampling path. The detailed derivation is provided in Appendix B. As a result, our method can efficiently generate higher-quality poses during the evolutionary process. + +# 4. Experiments + +We first simply present experimental setup which includes datasets, evaluation metrics and baselines in Sec. 4.3. More details of setup including datasets and implementation details are in Appendix C. Then we provide the quantitative and qualitative results of EvolvingGrasp in Sec. 4.2, followed by ablation study about different modules and hyperparameter analyses in Sec. 4.3. Next, we demonstrate the evolutionary improvement of EvolvingGrasp through an experiment starting from a suboptimal model trained on a degraded dataset in Sec. 4.4. Finally, We will demonstrate EvolvingGrasp in the real-world deployment in Sec. 4.5. + +# 4.1. Experimental Setup + +Datasets. We conduct experiment on four datasets, including DexGraspNet [39], Multidex [16], Realdex [20], Dex-GRAB [36] respectively. + +Evaluation Metrics. We use four metrics to evaluate the grasping performance. Success rate Suc.6 measures the proportion of grasping poses where the object's displacement does not exceed $2\mathrm{cm}$ in all six axial directions $(\pm X, \pm Y, \pm Z)$ , evaluating multi-directional stability. Suc.1 measures the proportion where displacement does not exceed $2\mathrm{cm}$ in at least one direction, assessing single-direction stability. Pen. indicates the maximum penetration depth (mm) between the hand and the object, with lower values suggesting more physically plausible grasps. Above metrics are calculated in the IssacGym simulator [27] with settings consistent with SceneDiffuser [12]. For efficiency, Time refers to the computational time required to generate grasping poses for a batch of objects. + +Baselines. Some generative-based methods such as UniDexGrasp [44], GraspTTA [13], SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] are compared on four benchmark datasets. + +# 4.2. Main Results + +To validate the effectiveness of our method in continuously enhancing grasping performance, we conduct quantitative evaluations on six randomly selected objects from each dataset, measuring both the Suc.6 metric and the degree of penetration. The results are illustrated in Fig. 3, indicating that as the number of inferences increases, the Suc.6 metric steadily improves. Although penetration exhibits some fluctuations due to increased contact area between the hand and the object, it still follows a downward trend. Additionally, we present qualitative results in Fig. 4, demonstrating how our method integrates human preferences during inference. This enables the selection of more favorable poses and progressively refines grasping poses to better align with human preferences through iteration. For instance, early generated poses may obstruct the camera lens, but after several rounds of preference-based finetuning, the model learns to produce poses that avoid blocking the lens. + +We conduct the experiment with comparing with other dexterous grasp generation methods in Tab. 1. Compared to other generative-based approaches, our method excels in generating high-quality, physically plausible grasping + +![](images/597b86aab741ae7722c4dd4ff1de62686462c91d9b7737a4316546686b60b18f.jpg) +Figure 4. Evolution of robotic grasp preferences during efficient feedback-driven finetuning across 10 epochs. Top row illustrates the adjustment from hand occlusion to clear nozzle visibility. Middle row demonstrates the transition from lens obstruction to an unobstructed camera view. The bottom row shows the evolution from a top-down grasping approach to a bottom-up one, while simultaneously mitigating physical impacts. + +poses while significantly reducing computational time. It is worth highlighting that the proposed method achieves 30x speedup (32s~38s v.s. 0.73s~2.7s) compared to these SO-TAs. After fine-tuning, our approach achieves superior performance by leveraging its evolutionary capability through iterative refinement. We also evaluate the trade-off between efficiency and grasp quality during preference alignment with different numbers of inference steps. While increasing the steps improves performance, it inevitably incurs additional computational costs. Moreover, even without guidance during sampling, we can achieve comparable results in real time. + +# 4.3. Ablation Study + +We investigate the impact of different modules in the physics-aware consistency model, Physical constraints Guidance in Distillation and Sampling from which we dubbed as PGD and PGS. We also consider the performance of HPO with and without physical guidance during sampling on test split of the multidex dataset. The results presented in Tab. 2 demonstrate the key role of physical constraints and preference finetuning to successful grasping pose generation during preference alignment. + +HPO not only effectively aligns grasping poses with human preferences at the individual object level but also enhances performance across the entire dataset. During inference across the entire dataset, we collect both successful and failed grasping samples for each object, utilizing Eq. (5) to perform lightweight finetuning of the entire model. Fig. 5 illustrates the relationship between EvolvingGrasp's + +![](images/854e1d31947e04fe8280bd1a5bb6660829ac6d7277e90de68576bcd22d4642a6.jpg) +Figure 5. The relationship between the success rate of Evolving-Grasp and the number of generated samples. + +Suc.6 metric and the number of generated samples, showing that our method continuously enhances grasping performance as more samples are generated. Further ablation studies on the impact of different hyperparameters are provided in Appendix D.2. + +# 4.4. Training from a Degraded Dataset + +In this subsection, we explore the feasibility of enhancing model performance through feedback-driven finetuning for single-object grasping tasks, particularly when trained on suboptimal data. To begin with, we randomly add the noise on the hand pose parameter of the original multidex dataset. This perturbation creates a degraded dataset, which we then use to train and distill a suboptimal model. During the in- + +Table 1. Grasping performance in terms of Suc.6, Suc.1, and Pen. comparison across different methods and datasets. "Step" refers to inference timestep. Bold values highlight the best results and underlined values indicate the runner-up. + +
Dataset MethodDexGraspNetMultiDexRealDexDexGRABTime ↓
Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓
UniDexGrasp [44]33.970.131.921.647.513.527.159.439.020.855.837.40.46±0.11
GraspTTA [13]18.667.824.530.362.819.013.346.440.114.451.051.410.41±0.32
SceneDiffuser [12]26.666.931.069.885.614.621.756.142.039.185.041.13.41±0.13
UGG [24]46.979.025.255.393.410.332.763.434.442.790.633.238.34±2.31
DexGrasp Any. [49]53.690.421.572.296.39.634.671.223.156.591.828.632.91±1.34
Ours w/o HPO (2-step)60.891.019.265.697.515.241.975.419.552.293.925.10.73±0.03
Ours (2-step)62.490.519.365.997.215.344.077.819.753.392.524.30.73±0.03
Ours w/o HPO (4-step)63.893.017.475.397.113.151.682.920.555.696.023.81.41±0.07
Ours (4-step)65.292.717.276.898.413.050.682.520.357.795.223.71.41±0.07
Ours w/o HPO (8-step)65.293.516.275.698.712.263.686.621.956.896.823.12.71±0.08
Ours (8-step)65.492.315.980.398.712.364.489.121.860.896.422.32.71±0.08
Real-time (2-step)55.290.520.063.795.013.846.578.921.448.993.324.80.06±0.01
Real-time (4-step)59.990.619.864.396.511.658.287.021.155.495.424.10.10±0.02
+ +Table 2. Ablation study on incorporating physical constraints during both training and sampling stages and the LLM module. The evaluation is conducted on Multidex when timestep is 4. + +
CMPGDPGSHPOSuc.6 ↑Suc.1 ↑Pen. ↓Time ↓
a✓60.094.614.00.10±0.02
b✓✓64.396.512.50.10±0.02
c✓✓66.295.614.91.41±0.07
d✓✓✓67.596.511.90.10±0.02
e✓✓✓75.397.113.11.41±0.07
f✓✓✓✓76.898.413.01.41±0.07
+ +![](images/347377672a72be5f4464839c7b1b72eadd8df24e41a723651f7f4cf354a5cc67.jpg) +Figure 6. The variation of the mean grasping success rate of the evolving model for randomly selected objects in the multidex dataset with increasing finetuning epochs. The blue dashed line represents the grasping success rate under the original model. + +ference phase of this suboptimal model, we collect successfully grasping poses as positive samples and all other poses as negative samples. Fig. 6 illustrates the change in the average grasping success rate of the suboptimal model on randomly sampled objects from the multidex dataset as the number of fine-tuning epochs increases. The results indicate that we continuously improve the success rate during the evolutionary adaptation process and ultimately outper + +![](images/515f9ebd32d356ba4083810aa5534e7907e440e62cbb961b1800640966305eb4.jpg) +Figure 7. Real-world deployment on Shadow Hand. + +form the accuracy of the original model. + +# 4.5. Real World Deployment + +To verify that EvolvingGrasp can continuously enhance grasping performance and align with human preferences in real world, we deploy our model on a real ShadowHand robot, as shown in Fig. 7. The pre-grasping motion trajectory is generated based on RealDex [20]. Real-world experiments demonstrate that our method achieves successful grasping through several efficient preference finetuning when the initial grasp attempt fails. Additional video demonstrations can be found in supplementary materials. + +# 5. Conclusion + +We propose EvolvingGrasp, an evolutionary grasp generation method through efficient preference alignment. HPO is introduced to allow the model to continuously align the performance with preference signals. We design the Physics-Aware Consistency Model to achieve both rapid inference and efficient preference finetuning while maintaining physical plausibility. Extensive experiments across four benchmarks demonstrate that our method achieves state-of-the-art results. Furthermore, we deploy our model on a real ShadowHand robot to validate its evolutionary capability in real-world scenarios. + +# References + +[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3 +[2] Yuxin Chen, Devesh K Jha, Masayoshi Tomizuka, and Diego Romeres. Fdpp: Fine-tune diffusion policy with human preference. arXiv preprint arXiv:2501.08259, 2025. 2 +[3] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 6 +[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3 +[5] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2 +[6] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, pages 7423-7433, 2024. 3 +[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6 +[8] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858-79885, 2023. 3 +[9] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, pages 2290-2295. IEEE, 1992. 2 +[10] Ayano Hiranaka, Shang-Fu Chen, Chieh-Hsin Lai, Dongjun Kim, Naoki Murata, Takashi Shibuya, Wei-Hsiang Liao, Shao-Hua Sun, and Yuki Mitsufuji. Human-feedback efficient reinforcement learning for online diffusion model finetuning. arXiv preprint arXiv:2410.05116, 2024. 3 +[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5, 12 +[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, pages 16750-16761, 2023. 2, 6, 8 +[13] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In ICCV, pages 11107-11116, 2021. 2, 6, 8 +[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 3 + +[15] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 2 +[16] Puhao Li, Tengyu Liu, Yuyang Li, Yiran Geng, Yixin Zhu, Yaodong Yang, and Siyuan Huang. Gendexgrasp: Generalizable dexterous grasping. In IEEE International Conference on Robotics and Automation), pages 8068-8074. IEEE, 2023. 6 +[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Generating grasp poses for a high-dof gripper using neural networks. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1518-1525. IEEE, 2019. 2 +[18] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2 +[19] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2 +[20] Yumeng Liu, Yaxun Yang, Youzhuo Wang, Xiaofei Wu, Ji-amin Wang, Yichen Yao, Soren Schwertfeger, Sibei Yang, Wenping Wang, Jingyi Yu, et al. Realdex: Towards human-like grasping for robotic dexterous hand. arXiv preprint arXiv:2402.13853, 2024. 6, 8 +[21] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 3 +[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3 +[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3 +[24] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. Ugg: Unified generative grasping. arXiv preprint arXiv:2311.16917, 2023. 2, 6, 8 +[25] Jens Lundell, Enric Corona, Tran Nguyen Le, Francesco Verdoja, Philippe Weinzaepfel, Grégory Rogez, Francesc Moreno-Noguer, and Ville Kyrki. Multi-fingan: Generative coarse-to-fine sampling of multi-finger grasps. In IEEE International Conference on Robotics and Automation, pages 4495-4501. IEEE, 2021. 2 +[26] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 6(4):6899-6906, 2021. 2 +[27] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 6 + +[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13 +[29] Jean Ponce, Steve Sullivan, J-D Boissonnat, and J-P Merlet. On characterizing and computing three-and four-finger force-closure grasps of polyhedral objects. In IEEE International Conference on Robotics and Automation, pages 821-827. IEEE, 1993. 2 +[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 3 +[31] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2 +[32] Carlos Rosales, Raul Suárez, Marco Gabiccini, and Antonio Bicchi. On the synthesis of feasible and prehensile robotic grasps. In IEEE International Conference on Robotics and Automation, pages 550-556. IEEE, 2012. 2 +[33] Yanming Shao and Chenxi Xiao. Bimanual grasp synthesis for dexterous robot hands. IEEE Robotics and Automation Letters, 2024. 2 +[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3 +[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 5 +[36] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6 +[37] Xiaofeng Tan, Hongsong Wang, Xin Geng, and Pan Zhou. Sopo: Text-to-motion generation using semi-online preference optimization. arXiv preprint arXiv:2412.05095, 2024. 2 +[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 2, 3, 4 +[39] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In IEEE International Conference on Robotics and Automation, pages 11359-11366. IEEE, 2023. 6 +[40] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 7(2):1659-1666, 2022. 2 + +[41] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. arXiv preprint arXiv:2405.19291, 2024. 2 +[42] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. IEEE Robotics and Automation Letters, 2024. 2 +[43] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In CVPR, pages 17933-17942, 2024. 2 +[44] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In CVPR, pages 4737-4746, 2023. 6, 8 +[45] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, pages 8941-8951, 2024. 3 +[46] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. arXiv preprint arXiv:2402.03201, 2024. 6 +[47] Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen, Joel Jang, Yi Li, Chaoqi Wang, Mingyu Ding, Dieter Fox, and Huaxiu Yao. Grape: Generalizing robot policy via preference alignment. arXiv preprint arXiv:2411.19309, 2024. 2 +[48] Zhengshen Zhang, Lei Zhou, Chenchen Liu, Zhiyang Liu, Chengran Yuan, Sheng Guo, Ruiteng Zhao, Marcelo H Ang Jr, and Francis EH Tay. Dexgrasp-diffusion: Diffusion-based unified functional grasp synthesis method for multideterous robotic hands. arXiv preprint arXiv:2407.09899, 2024. 2 +[49] Yiming Zhong, Qi Jiang, Jingyi Yu, and Yuexin Ma. Dexterity: Towards universal robotic dexterous grasping with physics awareness. In CVPR, 2025. 2, 5, 6, 8, 13 +[50] Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, and Tat-Seng Chua. Dreamdpo: Aligning text-to-3d generation with human preferences via direct preference optimization. arXiv preprint arXiv:2502.04370, 2025. 2 \ No newline at end of file diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/images.zip b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..f722f271c584e34697ea1a9b7f248d2de56cc75f --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:4793f71feeac4e991b1c64037bd44b20e6e1b6d490b9e83d209f7d0959819e61 +size 740257 diff --git a/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/layout.json b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..76ac8c86fa77f17c944487809bd44e6531c233cd --- /dev/null +++ b/2025/EvolvingGrasp_ Evolutionary Grasp Generation via Efficient Preference Alignment/layout.json @@ -0,0 +1,8023 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 144, + 103, + 465, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 144, + 103, + 465, + 139 + ], + "spans": [ + { + "bbox": [ + 144, + 103, + 465, + 139 + ], + "type": "text", + "content": "EvolvingGrasp: Evolutionary Grasp Generation via Efficient Preference Alignment" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "spans": [ + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": "Yufei Zhu" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Yiming Zhong" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\ast}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Zemin Yang" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Peishan Cong" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Jingyi Yu" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Xinge Zhu" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{2,\\dagger}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": ", Yuexin Ma" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1,\\dagger}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": " ShanghaiTech University " + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 141, + 160, + 468, + 217 + ], + "type": "text", + "content": " The Chinese University of Hong Kong https://evolvinggrasp.github.io/" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 248, + 334, + 458 + ], + "blocks": [ + { + "bbox": [ + 59, + 248, + 334, + 458 + ], + "lines": [ + { + "bbox": [ + 59, + 248, + 334, + 458 + ], + "spans": [ + { + "bbox": [ + 59, + 248, + 334, + 458 + ], + "type": "image", + "image_path": "d4655f9aa5ef7b2023e42f5999c7209ff12a76e6328a513bdf7203923f33cfcf.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 460, + 555, + 483 + ], + "lines": [ + { + "bbox": [ + 54, + 460, + 555, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 460, + 555, + 483 + ], + "type": "text", + "content": "Figure 1. The left part illustrates EvolvingGrasp, an approach akin to evolution, where it enables the model to learn from experience and iteratively refine its grasping strategy. The right part demonstrates its efficiency and effectiveness." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 342, + 251, + 552, + 458 + ], + "blocks": [ + { + "bbox": [ + 342, + 251, + 552, + 458 + ], + "lines": [ + { + "bbox": [ + 342, + 251, + 552, + 458 + ], + "spans": [ + { + "bbox": [ + 342, + 251, + 552, + 458 + ], + "type": "image", + "image_path": "46c6c0a21e9df52e097391058e722c6db16a390b7bac5b7b1c3bfd065838c04e.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 491, + 200, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 491, + 200, + 504 + ], + "spans": [ + { + "bbox": [ + 151, + 491, + 200, + 504 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 520, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 520, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 54, + 520, + 295, + 653 + ], + "type": "text", + "content": "Dexterous robotic hands often struggle to generalize effectively in complex environments due to models trained on low-diversity data. However, the real world presents an inherently unbounded range of scenarios. A natural solution is to enable robots learning from experience in complex environments—an approach akin to evolution, where systems improve through learning from both failures and successes. Motivated by this, we propose EvolvingGrasp, an evolutionary grasp generation method that continuously enhances grasping performance through efficient preference alignment. Specifically, we introduce Handpose-wise" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 493, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 493, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 493, + 555, + 636 + ], + "type": "text", + "content": "Preference Optimization (HPO), which allows the model to continuously align with preferences from both positive and negative feedback while progressively refining its grasping strategies. To further enhance efficiency and reliability during online adjustments, we incorporate a Physics-aware Consistency Model within HPO, which accelerates inference, reduces the number of timesteps needed for preference fine-tuning, and ensures physical plausibility throughout the process. Our results validate that EvolvingGrasp enables evolutionary grasp generation, ensuring robust, physically feasible, and preference-aligned grasping in both simulation and real scenarios." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "spans": [ + { + "bbox": [ + 314, + 666, + 395, + 679 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Dexterous robotic grasping has made significant strides in embodied manipulation, enabling more adaptive and pre" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 664, + 141, + 674 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 664, + 141, + 674 + ], + "spans": [ + { + "bbox": [ + 66, + 664, + 141, + 674 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 66, + 664, + 141, + 674 + ], + "type": "text", + "content": " * Equal contribution." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 674, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 674, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 674, + 295, + 712 + ], + "type": "inline_equation", + "content": "^{2\\dagger}" + }, + { + "bbox": [ + 56, + 674, + 295, + 712 + ], + "type": "text", + "content": " Corresponding author. This work was supported by NSFC (No.62206173), Shanghai Frontiers Science Center of Human-centered Artificial Intelligence (ShangHAI), MoE Key Laboratory of Intelligent Perception and Human-Machine Collaboration (KLIP-HuMaCo)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "11665" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 57, + 72, + 294, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 72, + 294, + 467 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 294, + 467 + ], + "type": "text", + "content": "cise interactions compared to traditional grippers. Existing grasping methods [5, 9, 15, 17-19, 29, 31-33, 42, 48] generally fall into two categories: optimization-based approaches [5, 9, 29, 32], which refine hand poses to achieve force-closure states, learning-based approaches [26, 40, 41, 43], which directly map object features to grasp poses through regression, probabilistic modeling, and generative-based methods [12, 13, 24, 25, 42], which utilize diffusion model to estimate the distribution of hand poses. Recent advances, such as DexGrasp Anything [49], have further introduced physics-based constraints to improve grasp feasibility. However, a fundamental limitation persists—limited generalization. These methods, trained on limited datasets, struggle to adapt to complex environments. This challenge is exacerbated by an inherent property of the real world: its unbounded diversity. The vast range of object shapes, materials, and environmental conditions makes it impractical to predefine an exhaustive set of grasping strategies. Without the ability to adapt in deployment, grasping models remain constrained, failing to handle variations beyond their training distribution. To overcome this, a natural approach is to enable evolutionary grasp generation, where the system learns from experience (i.e., both failures and successes), refining its grasping strategy through iterative improvements based on real-world interactions. This process not only enhances generalization but also allows for preference alignment, ensuring that grasping behaviors adapt to task-specific requirements and user-defined preferences. However, achieving efficient evolutionary refinement is non-trivial, as many existing learning-based approaches rely on slow, compute-intensive updates, particularly in diffusion-based models that require numerous iterative steps and physics simulations." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 475, + 294, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 475, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 475, + 294, + 712 + ], + "type": "text", + "content": "To address these challenges, we propose EvolvingGrasp, an evolutionary grasp generation framework that efficiently refines grasp strategies through preference alignment while maintaining physical plausibility. At its core, we introduce Handpose-wise Preference Optimization (HPO), a novel method that reformulates preference alignment [2, 37, 38, 47, 50] as a posterior probability optimization problem, encouraging generated grasps to converge toward preferred distributions while diverging from non-preferred ones. Notably, the proposed HPO is an extension of Direct Preference Optimization (DPO) [30], where it is also, to the best of our knowledge, the first to incorporate the DPO into the dexterous grasp. To further improve efficiency, we integrate HPO into a Physics-Aware Consistency Model (including two parts, i.e., Physics-Aware Distillation for training and Physics-Aware Sampling for inference), which pretrains a diffusion model and distills it into a lightweight, few-step sampling model. This enables both rapid inference and efficient preference fine-tuning, significantly reducing the number of required sampling steps and optimization iteraa" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 72, + 553, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 144 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 144 + ], + "type": "text", + "content": "tions. Additionally, we introduce three physics-aware constraints to ensure the stability, realism, and feasibility of generated grasp poses—surface pulling force to maintain stable contact, external penetration repulsion force to prevent object penetration, and self-penetration repulsion force to avoid inter-finger collisions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 316, + 145, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 145, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 145, + 553, + 227 + ], + "type": "text", + "content": "Extensive experiments across four benchmark datasets demonstrate that EvolvingGrasp achieves state-of-the-art results, significantly improving grasp success rate, sampling efficiency, and physical plausibility, with 30x speedup over existing methods, demonstrating robust generalization across simulated and real-world benchmarks. Our contributions can be summarized as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 314, + 231, + 553, + 457 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 231, + 553, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 231, + 553, + 277 + ], + "spans": [ + { + "bbox": [ + 314, + 231, + 553, + 277 + ], + "type": "text", + "content": "- We introduce EvolvingGrasp, an efficient evolutionary grasp generation framework that enables iterative refinement, addressing the challenge of generalizing to diverse and unstructured real-world environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 279, + 553, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 279, + 553, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 279, + 553, + 338 + ], + "type": "text", + "content": "- Efficient preference alignment is achieved through Handpose-wise Preference Optimization (HPO), which formulates grasp adaptation as a posterior probability optimization problem, enabling the model to iteratively converge toward preferred grasp distributions." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 338, + 553, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 338, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 314, + 338, + 553, + 386 + ], + "type": "text", + "content": "- We propose a Physics-Aware Consistency Model (PCM) that accelerates preference alignment by reducing sampling steps while enforcing geometric consistency and physical feasibility through structured constraints." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 387, + 553, + 457 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 387, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 314, + 387, + 553, + 457 + ], + "type": "text", + "content": "- Extensive experiments across four benchmark datasets demonstrate that our method achieves state-of-the-art grasp success rates, physical plausibility, and sampling efficiency. Furthermore, it enables real-time grasp generation with minimal computational overhead, achieving comparable performance to gradient-based methods." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 316, + 472, + 400, + 484 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 472, + 400, + 484 + ], + "spans": [ + { + "bbox": [ + 316, + 472, + 400, + 484 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 316, + 493, + 430, + 505 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 430, + 505 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 430, + 505 + ], + "type": "text", + "content": "2.1. Dexterous Grasping" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 315, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 510, + 553, + 713 + ], + "type": "text", + "content": "Dexterous grasp generation aims to produce high-quality grasping poses for robotic hands. Recent works can be categorized into regression-based [17, 18, 31] and generation-based methods [12, 13, 24-26, 40, 41]. Regression-based methods directly predict grasping parameters from the input object, but they often suffer from mode collapse issues that limits output diversity. Generation-based methods, though capable of producing varied solutions, face efficiency challenges. SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] require multiple sampling steps to generate diverse grasping poses, and DexGrasp Anything [49] additionally incurs computational overhead by calculating physical constraint losses and performing gradient updates at each sampling iteration. However, existing methods cannot generate grasping poses aligned with human habits due to their lack of evolutionary adaptation and preference alignment. They also struggle to balance physical plausibility" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "11666" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 70, + 553, + 232 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 553, + 232 + ], + "type": "image", + "image_path": "57c83978e95fa440daea1bee7421ad3f6a20fc8a0f44a19eaabb6e9f4333761e.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 235, + 553, + 322 + ], + "blocks": [ + { + "bbox": [ + 57, + 235, + 553, + 322 + ], + "lines": [ + { + "bbox": [ + 57, + 235, + 553, + 322 + ], + "spans": [ + { + "bbox": [ + 57, + 235, + 553, + 322 + ], + "type": "image", + "image_path": "887f1e1bf94a34a3990de5441bba2b5321063f5966e68198a3406d5e2caf7aaa.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 329, + 555, + 386 + ], + "lines": [ + { + "bbox": [ + 54, + 329, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 54, + 329, + 555, + 386 + ], + "type": "text", + "content": "Figure 2. Overview of EvolvingGrasp. The evolutionary process begins with the human preference guidance, where Handpose-wise Preference Optimization (HPO, highlighted in the green rectangle) is employed to facilitate preference alignment. These grasp poses are generated by the Physics-Aware Consistency Model (shown in blue rectangles), including Sampling and Distillation mechanism, to ensure the sampling efficiency and the physical plausibility. In this way, EvolvingGrasp, an efficient evolutionary grasp generation framework is proposed to enable the grasp model iteratively converge toward preferred distributions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 406, + 180, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 406, + 180, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 180, + 418 + ], + "type": "text", + "content": "with computational efficiency." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 220, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 220, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 220, + 441 + ], + "type": "text", + "content": "2.2. Accelerating Diffusion Models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "text", + "content": "Accelerating dexterous grasping is vital for enhancing real-time adaptability. A series of diffusion-based acceleration methods have emerged, making real-time and efficient generation possible. DDIM [34] accelerates inference by reformulating the stochastic denoising process into a deterministic ODE solver. The DPM-Solver [22, 23] series further designs efficient high-order ODE solvers, achieving comparable quality with only " + }, + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "inline_equation", + "content": "10 \\sim 20" + }, + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "text", + "content": " steps. However, their performance degrades significantly when sampling steps are reduced to " + }, + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "inline_equation", + "content": "2 \\sim 4" + }, + { + "bbox": [ + 54, + 445, + 296, + 651 + ], + "type": "text", + "content": ". Consistency works [14, 21] support few-step sampling while preserving quality. CTM [14] extends CM to reduce cumulative errors and discretization inaccuracies. sCMs [21] unifies flow matching and diffusion frameworks to avoid discretization errors and hyperparameter tuning. Therefore, we employ the consistency model to enable efficient evolutionary refinement, reducing inference timesteps while maintaining result quality." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 658, + 282, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 282, + 672 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 282, + 672 + ], + "type": "text", + "content": "2.3. Preference Finetuning for Diffusion Models" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Preference alignment play a crucial role in enabling generative models to learn from user feedback, optimize generation strategies, and progressively improve output qual" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 405, + 555, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 405, + 555, + 634 + ], + "spans": [ + { + "bbox": [ + 313, + 405, + 555, + 634 + ], + "type": "text", + "content": "ity. Some researchers utilize it into diffusion models to better align with human preferences in image generation area. These methods can be categorized into two types, finetuning with reward model [1, 4, 6, 8] and reward-free finetuning [10, 38, 45]. The former includes DDPO [1] and DPOK [8] which treat the denoising process of diffusion models as an MDP and finetune using multiple reward models. Finetuning without reward model includes D3PO [45] and Diffusion-DPO [38], which extend the theoretical framework of DPO [30] to multi-step MDPs. These approaches learn an optimal reward model and then use it to refine the sampling strategy, making them a relatively more cost-effective alternative. However, RL-based fine-tuning for diffusion models still requires backpropagation at every sampled timestep, which is time-consuming. Applying this approach directly to grasping tasks is impractical for real-world applications. Therefore, we propose a faster preference alignment fine-tuning approach that reduces backpropagation steps." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 649, + 397, + 662 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 649, + 397, + 662 + ], + "spans": [ + { + "bbox": [ + 313, + 649, + 397, + 662 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "spans": [ + { + "bbox": [ + 313, + 670, + 382, + 682 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "To achieve efficient evolutionary grasp generation that aligns with human preferences while maintaining both ef" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11667" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 312 + ], + "type": "text", + "content": "ficiency and physical plausibility, we propose Evolving-Grasp, as illustrated in Fig. 2. The evolutionary process begins with preference alignment, where we introduce Handpose-Wise Preference Optimization (HPO) (Section 3.3). HPO formulates grasp adaptation as a posterior probability optimization problem, allowing the model to iteratively converge toward preferred grasp distributions. However, directly applying HPO suffers from slow sampling and inefficiencies in preference alignment, limiting its practicality for real-time grasp refinement. To address these challenges, we incorporate a Physics-Aware Consistency Model (PCM) (Section 3.4), which leverages a consistency model framework to enhance efficiency by reducing the number of required sampling steps. While this improves inference speed, naive consistency-based sampling may still generate physically implausible grasp poses. To ensure geometric consistency and physical feasibility, PCM integrates a physics-aware distillation and sampling mechanism, which enforces structured physical constraints on the generated poses." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 319, + 180, + 332 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 180, + 332 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 180, + 332 + ], + "type": "text", + "content": "3.2. Problem Formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": "Given the object point cloud representation " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "O \\in \\mathbb{R}^{N \\times 3}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": ", our goal is to generate dexterous grasp poses with high success rate and low penetration from the posterior distribution " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "P(x \\mid O)" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "x = \\{x_{i}\\}_{i=1}^{n}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": ". Specifically, the pose parameters contain three parts, joint angles of the hand " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "\\theta_{h} \\in \\mathbb{R}^{24}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": ", global translation " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "T_{global} \\in \\mathbb{R}^{3}" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": ", and global rotation " + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "inline_equation", + "content": "R_{global} \\in SO(3)" + }, + { + "bbox": [ + 55, + 337, + 296, + 422 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": "Given ground truth samples from the data distribution " + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\pi(x_0)" + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": ", the noise schedule weight " + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 422, + 296, + 471 + ], + "type": "text", + "content": ", the goal of diffusion models is to fit the GT data distribution. The objective of training diffusion model is as follows:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 89, + 479, + 295, + 501 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 89, + 479, + 295, + 501 + ], + "spans": [ + { + "bbox": [ + 89, + 479, + 295, + 501 + ], + "type": "interline_equation", + "content": "\\mathbb {E} _ {t \\sim [ 0, T ], x _ {0}, \\epsilon \\sim N (0, I)} \\left[ \\| \\epsilon - \\epsilon_ {\\theta} (x _ {t}, t, O) \\| ^ {2} \\right] \\tag {1}", + "image_path": "93b6fca7552238270d7d892c5c421de85bb994bdeb9c91e0b57f9f993db537da.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "x_{t} = \\sqrt{\\bar{\\alpha}_{t}} x_{0} + \\sqrt{1 - \\bar{\\alpha}_{t}}\\epsilon" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": ". Given the noise related parameter " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\sigma_{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": " and the mean from " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 55, + 510, + 295, + 545 + ], + "type": "text", + "content": ", the reverse process of diffusion model is as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 88, + 555, + 295, + 571 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 555, + 295, + 571 + ], + "spans": [ + { + "bbox": [ + 88, + 555, + 295, + 571 + ], + "type": "interline_equation", + "content": "p _ {\\theta} \\left(x _ {t - 1} \\mid x _ {t}\\right) = \\mathcal {N} \\left(x _ {t - 1}; \\mu_ {\\theta} \\left(x _ {t}, t\\right), \\sigma_ {t} ^ {2} \\mathrm {I}\\right) \\tag {2}", + "image_path": "5af074449d0d41d54b1b9b7eaea944f4b1061e1ca707db3077f11164910ecbf2.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 588, + 267, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 267, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 267, + 601 + ], + "type": "text", + "content": "3.3. Handpose-wise Preference Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "To achieve preference alignment in the evolutionary adaptation process, we introduce the HPO, an extension of direct preference optimization (DPO). In DPO, the fundamental assumption is that we have access to data generated by the model, with human annotators providing the corresponding preferences. Specifically, for a given object " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", we observe pairs of grasp poses " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_0^w" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_0^l" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ", where the set " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_0^w" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " is preferred over " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_0^l" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": ". These preferences can be modeled using the Bradley-Terry model, which expresses the probability" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 503, + 84 + ], + "type": "text", + "content": "of one sample being preferred over another as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 343, + 93, + 555, + 109 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 93, + 555, + 109 + ], + "spans": [ + { + "bbox": [ + 343, + 93, + 555, + 109 + ], + "type": "interline_equation", + "content": "p _ {B T} \\left(x _ {0} ^ {w} \\succ x _ {0} ^ {l}\\right) = \\sigma \\left(r \\left(c, x _ {0} ^ {w}\\right) - r \\left(c, x _ {0} ^ {l}\\right)\\right) \\tag {3}", + "image_path": "aaca98c62e0430b1b627754a7054d9f113a3934f6c2020c01e6a3389448649d0.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 118, + 555, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 555, + 167 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 555, + 167 + ], + "type": "text", + "content": "The training objective of DPO is to maximize the likelihood of the observed preferences, which can be formulated as a binary classification problem. Specifically, the whole model is finetuned to minimize the following loss function:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 323, + 176, + 554, + 219 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 176, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 323, + 176, + 554, + 219 + ], + "type": "interline_equation", + "content": "L _ {\\mathrm {B T}} (\\theta) = - \\mathbb {E} _ {\\left(x _ {0} ^ {w}, x _ {0} ^ {l}\\right) \\sim \\mathcal {D}} \\log \\sigma \\left( \\begin{array}{c} \\beta \\mathbb {E} \\\\ x _ {1: T} ^ {w} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {w} \\mid x _ {0} ^ {w}\\right) \\\\ x _ {1: T} ^ {l} \\sim \\pi_ {\\theta} \\left(x _ {1: T} ^ {l} \\mid x _ {0} ^ {l}\\right) \\end{array} \\right. \\tag {4}", + "image_path": "c99024427bfb9394e2d728812eb0d99d3c2818aacba02796d54e8a3b10636a88.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 382, + 220, + 531, + 251 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 220, + 531, + 251 + ], + "spans": [ + { + "bbox": [ + 382, + 220, + 531, + 251 + ], + "type": "interline_equation", + "content": "\\left. \\left[ \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {w}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {0 : T} ^ {l}\\right)}{\\pi_ {\\mathrm {r e f}} \\left(x _ {0 : T} ^ {l}\\right)} \\right]\\right)", + "image_path": "885d65619075399af85ff70bc6470b9af912e8f73a0d0979fe4b66f97b6d9fbe.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 260, + 555, + 285 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 555, + 285 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 260, + 555, + 285 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 313, + 260, + 555, + 285 + ], + "type": "text", + "content": " is the paired-wise grasp pose dataset. Following [38], we can get the upper bound of Eq. (4) as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 321, + 293, + 545, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 293, + 545, + 319 + ], + "spans": [ + { + "bbox": [ + 321, + 293, + 545, + 319 + ], + "type": "interline_equation", + "content": "L_{BT}(\\theta)\\leq -\\begin{array}{c}\\beta \\mathbb{E}\\\\ \\big(x^{w}_{0},x^{l}_{0}\\big)\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{w}_{n - 1,n}\\sim \\pi_{\\theta}\\big(x^{w}_{n - 1,n}|x^{w}_{0}\\big) \\end{array}", + "image_path": "0f4433ee385367db2f86ed63dac7c622effbc36d39c7f0847dbfc77236270416.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 319, + 553, + 362 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 319, + 553, + 362 + ], + "spans": [ + { + "bbox": [ + 322, + 319, + 553, + 362 + ], + "type": "interline_equation", + "content": "\\log \\sigma \\left(\\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {w} \\mid x _ {n} ^ {w}\\right)} - \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {l} \\mid x _ {n} ^ {l}\\right)}\\right) \\tag {5}", + "image_path": "ee86fb0c33cc52a1cc274599921f1aed648de2a0646dd888829d88ce15af1525.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "inline_equation", + "content": "\\pi_{\\theta}\\left(x_{n - 1} \\mid x_n\\right)" + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": " is the probability of sampling " + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "inline_equation", + "content": "x_{n - 1}" + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "inline_equation", + "content": "x_{n}" + }, + { + "bbox": [ + 313, + 363, + 553, + 387 + ], + "type": "text", + "content": ", which can be computed by:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 319, + 396, + 554, + 424 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 319, + 396, + 554, + 424 + ], + "spans": [ + { + "bbox": [ + 319, + 396, + 554, + 424 + ], + "type": "interline_equation", + "content": "\\pi_ {\\theta} \\left(x _ {n - 1} \\mid x _ {n}\\right) = \\frac {1}{\\sqrt {2 \\pi} \\sigma_ {n}} \\exp \\left(- \\frac {\\left(x _ {n} - \\mu_ {\\theta} \\left(x _ {n} , n\\right)\\right) ^ {2}}{2 \\sigma_ {n} ^ {2}}\\right) \\tag {6}", + "image_path": "5b4f2ae5207f35aecb10ff26452494f66c4e90ceed2832e20d997cce89d0b76c.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 433, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 555, + 505 + ], + "type": "text", + "content": "As far as we know, HPO is the first to integrate DPO into grasp pose generation, from which we extend DPO to a more flexible form. In HPO, there is no requirement to maintain an equal number of preferred and non-preferred grasp poses, which enables more adaptive preference learning. The objective of HPO is introduced as follows:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 347, + 513, + 547, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 513, + 547, + 536 + ], + "spans": [ + { + "bbox": [ + 347, + 513, + 547, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal{L}_{HPO} = \\operatorname *{\\beta}\\limits_{\\substack{x^{i}_{0}\\sim \\mathcal{D},n\\sim \\mathcal{U}(1,N),x^{i}_{n - 1,n}\\sim \\pi_{\\theta}\\left(x^{i}_{n - 1,n}|x^{i}_{0}\\right)}}\\log \\sigma", + "image_path": "276f60455ec604b4f3ef1e88088599c09b85b9f1074d016c51883621fc09bd80.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 537, + 553, + 582 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 537, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 323, + 537, + 553, + 582 + ], + "type": "interline_equation", + "content": "\\left(\\sum_ {i = 1} ^ {N _ {s u c}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {i} \\mid x _ {n} ^ {i}\\right)} - \\sum_ {j = 1} ^ {N _ {f a i l}} \\log \\frac {\\pi_ {\\theta} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}{\\pi_ {\\text {r e f}} \\left(x _ {n - 1} ^ {j} \\mid x _ {n} ^ {j}\\right)}\\right) \\tag {7}", + "image_path": "f849aad02b05eea534d064cf1ce82e8aba9c2aa3407fffae08fc4b76742bac55.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{\\text{succ}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "inline_equation", + "content": "N_{\\text{fail}}" + }, + { + "bbox": [ + 313, + 582, + 555, + 714 + ], + "type": "text", + "content": " represent the number of preferred and non-preferred poses. HPO optimizes grasp pose generation by quantifying the probabilistic divergence between successful grasps (preferred samples) and failed grasps (non-preferred samples), thereby driving the model towards human-preferred behaviors. This process inherently involves the dynamic adjustment and maximization of reward signals: preferred grasp poses are reinforced due to their higher probability, while non-preferred poses are suppressed, guiding the model to progressively discard ineffective strategies." + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11668" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 204 + ], + "type": "text", + "content": "Preferred grasp selection can be conducted through either simulation-based evaluation or human-in-the-loop selection. In the simulation-based approach, poses that achieve success across all six directional evaluations are classified as preferred samples, while the remaining poses are treated as negative samples. Alternatively, in the human selection process, grasp poses that align with human intuition and habitual preferences are designated as preferred samples, while the others are considered non-preferred. Finally, we finetune the whole model via LoRA [11] using Eq. (7) to align the model with preferences." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 211, + 238, + 224 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 238, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 238, + 224 + ], + "type": "text", + "content": "3.4. Physics-Aware Consistency Model" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 228, + 165, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 228, + 165, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 228, + 165, + 239 + ], + "type": "text", + "content": "3.4.1. Consistency Model" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "spans": [ + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "text", + "content": "Directly applying HPO faces two major efficiency challenges: generating poses requires over hundreds of timesteps per inference, and preference fine-tuning demands a large number of backpropagation steps. To address these inefficiencies, we adopt the consistency model framework to accelerate both sampling process and preference alignment. The core idea of the consistency model is to learn a mapping from any point along the ODE trajectory back to its starting point, which corresponds to the data distribution. Given a Probability Flow ODE trajectory [35] " + }, + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "inline_equation", + "content": "\\{x_{\\tau_n}\\}_{\\tau_n\\in [0,T]}" + }, + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "text", + "content": ", a consistency model " + }, + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "inline_equation", + "content": "f_{\\theta}" + }, + { + "bbox": [ + 55, + 243, + 296, + 376 + ], + "type": "text", + "content": " is defined as:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 386, + 294, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 386, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 127, + 386, + 294, + 399 + ], + "type": "interline_equation", + "content": "f _ {\\theta}: \\left(x _ {\\tau_ {n}}, \\tau_ {n}, O\\right) \\mapsto x _ {0} \\tag {8}", + "image_path": "5e8d202eecfd1fae00cd4c205fd9bafbf1f55ef11adf2269177a92d0913d43b8.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " is the initial point of the trajectory, " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " is the observation as condition. Due to space limitations, we omitted condition " + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 407, + 296, + 456 + ], + "type": "text", + "content": " in other parts of the paper, except for the appendix. The self-consistency property ensures that:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 79, + 463, + 295, + 477 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 463, + 295, + 477 + ], + "spans": [ + { + "bbox": [ + 79, + 463, + 295, + 477 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta^ {\\prime}} \\left(x _ {\\tau_ {n} ^ {\\prime}}, \\tau_ {n} ^ {\\prime}\\right) \\quad \\forall \\tau_ {n}, \\tau_ {n} ^ {\\prime} \\in [ 0, T ] \\tag {9}", + "image_path": "6a2626e3f79a6f70f729a7c527bcd7c035fb0c85fe7216d8744b2826465fc87f.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "text", + "content": "The consistency model needs to meet a key boundary condition: when " + }, + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "inline_equation", + "content": "t = 0" + }, + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "text", + "content": ", the model output should be the input itself, that is, " + }, + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_0,0,O) = x_0" + }, + { + "bbox": [ + 55, + 485, + 296, + 533 + ], + "type": "text", + "content": ". This can be achieved as the following way:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 541, + 295, + 555 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 541, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 64, + 541, + 295, + 555 + ], + "type": "interline_equation", + "content": "f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = c _ {\\text {s k i p}} \\left(\\tau_ {n}\\right) x _ {\\tau_ {n}} + c _ {\\text {o u t}} \\left(\\tau_ {n}\\right) F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) \\tag {10}", + "image_path": "fbf7f808a65d923cd769deac01b1c7ce8fa77642f56a5d6b359c1b6f08aa23ec.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "spans": [ + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "c_{skip}(t)" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "c_{out}(t)" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": " are differentiable functions that satisfy " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "c_{skip}(\\epsilon) = 1" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "c_{out}(\\epsilon) = 0" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "F_{\\theta}(x,t)" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": " denotes a deep neural network that predicts " + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 55, + 563, + 296, + 599 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 65, + 606, + 294, + 641 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 606, + 294, + 641 + ], + "spans": [ + { + "bbox": [ + 65, + 606, + 294, + 641 + ], + "type": "interline_equation", + "content": "F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = \\frac {1}{\\sqrt {\\bar {\\alpha} _ {\\tau_ {n}}}} \\left(x _ {\\tau_ {n}} - \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n}}} \\epsilon_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right)\\right) \\tag {11}", + "image_path": "aefd6a12e109a34073227cc797656476367b90c005a3c2d8c2d632edc55fc0d3.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 296, + 713 + ], + "type": "text", + "content": "Consistency distillation leverages pre-trained diffusion model to condense multi-step sampling into a more efficient few-step inference process. The process involves generating pairs of adjacent points on the PFODE trajectory using numerical ODE solvers and minimizing the difference between the model's outputs for these pairs. The loss function" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 314, + 72, + 477, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 477, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 477, + 83 + ], + "type": "text", + "content": "for consistency distillation is defined as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 325, + 92, + 553, + 113 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 92, + 553, + 113 + ], + "spans": [ + { + "bbox": [ + 325, + 92, + 553, + 113 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {C D} = \\mathbb {E} \\left[ d \\left(f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), f _ {\\theta^ {\\prime}} \\left(\\hat {x} _ {\\tau_ {n - 1}} ^ {*}, \\tau_ {n - 1}\\right)\\right) \\right] \\tag {12}", + "image_path": "e9f2c8ab3ad72e7b54eb4e9847cf84aa4b2f978b60c470fe8d0c0e3cd2a2357c.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " is omitted for simplicity, " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\tau_{n-1}}^*" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " is computed using a numerical ODE solver, " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "d(\\cdot, \\cdot)" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " is a " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "L_2" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " distance metric, " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " represent the online network and target network parameters. In addition, the parameters of " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\theta'" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " are updated by the exponential moving average (EMA) of the parameters of " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "inline_equation", + "content": "\\hat{x}_{\\tau_{n-1}}^*" + }, + { + "bbox": [ + 313, + 121, + 555, + 196 + ], + "type": "text", + "content": " can be computed as:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 204, + 553, + 220 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 204, + 553, + 220 + ], + "spans": [ + { + "bbox": [ + 326, + 204, + 553, + 220 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {\\tau_ {n - 1}} ^ {*} \\leftarrow \\sqrt {\\bar {\\alpha} _ {\\tau_ {n - 1}}} F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sqrt {1 - \\bar {\\alpha} _ {\\tau_ {n - 1}}} \\epsilon \\tag {13}", + "image_path": "858eba8ae544a5f2198648aaaf42e02cbb9e24ded0f2e999413fa6638e9981eb.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 228, + 558, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 558, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 558, + 312 + ], + "type": "text", + "content": "During sampling, we utilize the consistency function to directly generate the final sample and the quality of the generated sample can be enhanced through an iterative process that alternates between denoising and injecting noise. Given the sequence of timesteps " + }, + { + "bbox": [ + 313, + 228, + 558, + 312 + ], + "type": "inline_equation", + "content": "S \\in \\{\\tau_i \\mid \\tau_0 = 0, \\tau_{N-1} = T, \\tau_i < \\tau_{i+1} \\text{ for } i = 0, 1, \\dots, N-1\\}" + }, + { + "bbox": [ + 313, + 228, + 558, + 312 + ], + "type": "text", + "content": ", the adding-noise process can be formulated as:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 374, + 321, + 553, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 374, + 321, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 374, + 321, + 553, + 335 + ], + "type": "interline_equation", + "content": "\\hat {x} _ {\\tau_ {n - 1}} = \\mu_ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\sigma_ {\\tau_ {n}} \\epsilon \\tag {14}", + "image_path": "5d827e07eccf75da32cf5bce39cb4406f25231d2e4a10d7fd81904ea5f192d7a.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\mu_{\\theta} = \\sqrt{\\bar{\\alpha}_{\\tau_{n - 1}}} f_{\\theta}(x_{\\tau_n},\\tau_n)" + }, + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "type": "inline_equation", + "content": "\\sigma_{\\tau_n} = \\sqrt{1 - \\bar{\\alpha}_{\\tau_{n - 1}}}" + }, + { + "bbox": [ + 313, + 343, + 553, + 380 + ], + "type": "text", + "content": " Subsequently, we conduct the prediction of the final sample utilizing the trained consistency function once more." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 397, + 516, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 516, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 516, + 410 + ], + "type": "text", + "content": "3.4.2. Physics-Aware Distillation and Sampling" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "text", + "content": "Although adopting the consistency model improves sampling and preference fine-tuning efficiency, it still generates physically implausible poses. To address this, we introduce a Physics-Aware Distillation and Sampling paradigm that enforces physical constraints during the distillation process of predicting " + }, + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "inline_equation", + "content": "\\hat{x}_0" + }, + { + "bbox": [ + 313, + 412, + 555, + 520 + ], + "type": "text", + "content": " while ensuring that the sampling trajectories adhere to specific constraints. Following [49], we incorporate three physics-aware objectives, with the distillation objective of the consistency model formulated as follows:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 323, + 529, + 553, + 559 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 529, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 323, + 529, + 553, + 559 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {P A D} = \\mathcal {L} _ {C D} + \\sum_ {i = 1} ^ {m} \\alpha_ {i} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {15}", + "image_path": "c4f8b3393d1c291ca96ad6343570cf5786aff0b839721eeec29ab34c03cd6875.jpg" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "inline_equation", + "content": "L_{PA_i}\\left(F_\\theta(x_t, t), \\epsilon_\\theta\\right)" + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "inline_equation", + "content": "i^{th}" + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": " physical constraint loss and " + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "inline_equation", + "content": "m = 3" + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": " denotes three constraints, i.e., Surface Pulling Force, External-penetration Repulsion Force, Self-Penetration Repulsion Force respectively [49]. These constraints guarantee grasping feasibility and maintain geometric accuracy in finger-to-object and inter-finger interactions. " + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\alpha_i" + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": " is the corresponding weight parameter. We first train a diffusion model for grasp pose generation as a teacher model. Then, we utilize " + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{PAD}}" + }, + { + "bbox": [ + 313, + 570, + 555, + 689 + ], + "type": "text", + "content": " to distill the teacher model into a student model." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "During sampling, the consistency model estimates a clearer hand pose " + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "inline_equation", + "content": "x_0" + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": " based on the current noisy hand" + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11669" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 68, + 300, + 175 + ], + "blocks": [ + { + "bbox": [ + 56, + 68, + 300, + 175 + ], + "lines": [ + { + "bbox": [ + 56, + 68, + 300, + 175 + ], + "spans": [ + { + "bbox": [ + 56, + 68, + 300, + 175 + ], + "type": "image", + "image_path": "6e1628acfa9f5a828e02c518a23a04c4075c5e239503df446d5f57fe47f10145.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "lines": [ + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 182, + 555, + 205 + ], + "type": "text", + "content": "Figure 3. Mean grasping performance in terms of success rate and penetration of randomly selected 6 objects with the finetuning epoch increasing during inference optimization." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 301, + 68, + 554, + 175 + ], + "blocks": [ + { + "bbox": [ + 301, + 68, + 554, + 175 + ], + "lines": [ + { + "bbox": [ + 301, + 68, + 554, + 175 + ], + "spans": [ + { + "bbox": [ + 301, + 68, + 554, + 175 + ], + "type": "image", + "image_path": "ecba72db4560fdacdd30b98d254c24d670062dc788dd5e4e07ad8c504a333d26.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": "pose " + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "inline_equation", + "content": "x_{\\tau_n}" + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": " and object information " + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "inline_equation", + "content": "O" + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": " by consistency function " + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "inline_equation", + "content": "f_{\\theta}(x_{\\tau_n}, \\tau_n, O)" + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": ". Subsequently, physical constraints are applied to steer the sampling process, making it closer to a physically feasible grasping pose. Following [3, 7, 46], the gradient of the constraint loss is used to modify the mean from " + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "inline_equation", + "content": "x_t" + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "inline_equation", + "content": "x_{t-1}" + }, + { + "bbox": [ + 55, + 213, + 296, + 284 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 70, + 292, + 294, + 338 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 70, + 292, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 70, + 292, + 294, + 338 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\hat {\\mu} _ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) = \\mu_ {\\theta} (x _ {\\tau_ {n}}, \\tau_ {n}) \\\\ + \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\left(F _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right), \\epsilon_ {\\theta}\\right) \\tag {16} \\\\ \\end{array}", + "image_path": "dcd594d3aff54f91c6467a0b1f12c00ed316a7b9a9ef9c8691d08297fd39c426.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "inline_equation", + "content": "\\gamma_{i}" + }, + { + "bbox": [ + 55, + 346, + 296, + 381 + ], + "type": "text", + "content": " is the weight parameter corresponding to each physical constraint. Moreover, we can derive a new mapping from noise to data as follows:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 61, + 390, + 294, + 430 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 390, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 61, + 390, + 294, + 430 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) = f _ {\\theta} \\left(x _ {\\tau_ {n}}, \\tau_ {n}\\right) + \\frac {1}{\\sqrt {\\alpha_ {\\tau_ {n - 1}}}} \\sum_ {i = 1} ^ {m} \\gamma_ {i} \\nabla_ {x _ {\\tau_ {n}}} L _ {P A _ {i}} \\tag {17}", + "image_path": "2d1b594210ebd73d1d1f0be8fee0711b314fcfd099811646fa28985c171c1caa.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 432, + 296, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 432, + 296, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 432, + 296, + 491 + ], + "type": "text", + "content": "By employing the physics-aware consistency model, we derive a novel preference alignment objective based on the new sampling path. The detailed derivation is provided in Appendix B. As a result, our method can efficiently generate higher-quality poses during the evolutionary process." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 137, + 514 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 653 + ], + "type": "text", + "content": "We first simply present experimental setup which includes datasets, evaluation metrics and baselines in Sec. 4.3. More details of setup including datasets and implementation details are in Appendix C. Then we provide the quantitative and qualitative results of EvolvingGrasp in Sec. 4.2, followed by ablation study about different modules and hyperparameter analyses in Sec. 4.3. Next, we demonstrate the evolutionary improvement of EvolvingGrasp through an experiment starting from a suboptimal model trained on a degraded dataset in Sec. 4.4. Finally, We will demonstrate EvolvingGrasp in the real-world deployment in Sec. 4.5." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 660, + 172, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 660, + 172, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 660, + 172, + 673 + ], + "type": "text", + "content": "4.1. Experimental Setup" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Datasets. We conduct experiment on four datasets, including DexGraspNet [39], Multidex [16], Realdex [20], Dex-GRAB [36] respectively." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": "Evaluation Metrics. We use four metrics to evaluate the grasping performance. Success rate Suc.6 measures the proportion of grasping poses where the object's displacement does not exceed " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " in all six axial directions " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "(\\pm X, \\pm Y, \\pm Z)" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": ", evaluating multi-directional stability. Suc.1 measures the proportion where displacement does not exceed " + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 213, + 555, + 380 + ], + "type": "text", + "content": " in at least one direction, assessing single-direction stability. Pen. indicates the maximum penetration depth (mm) between the hand and the object, with lower values suggesting more physically plausible grasps. Above metrics are calculated in the IssacGym simulator [27] with settings consistent with SceneDiffuser [12]. For efficiency, Time refers to the computational time required to generate grasping poses for a batch of objects." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 381, + 555, + 430 + ], + "type": "text", + "content": "Baselines. Some generative-based methods such as UniDexGrasp [44], GraspTTA [13], SceneDiffuser [12], UGG [24], and DexGrasp Anything [49] are compared on four benchmark datasets." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 441, + 399, + 453 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 399, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 399, + 453 + ], + "type": "text", + "content": "4.2. Main Results" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 460, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 555, + 664 + ], + "type": "text", + "content": "To validate the effectiveness of our method in continuously enhancing grasping performance, we conduct quantitative evaluations on six randomly selected objects from each dataset, measuring both the Suc.6 metric and the degree of penetration. The results are illustrated in Fig. 3, indicating that as the number of inferences increases, the Suc.6 metric steadily improves. Although penetration exhibits some fluctuations due to increased contact area between the hand and the object, it still follows a downward trend. Additionally, we present qualitative results in Fig. 4, demonstrating how our method integrates human preferences during inference. This enables the selection of more favorable poses and progressively refines grasping poses to better align with human preferences through iteration. For instance, early generated poses may obstruct the camera lens, but after several rounds of preference-based finetuning, the model learns to produce poses that avoid blocking the lens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 555, + 714 + ], + "type": "text", + "content": "We conduct the experiment with comparing with other dexterous grasp generation methods in Tab. 1. Compared to other generative-based approaches, our method excels in generating high-quality, physically plausible grasping" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11670" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 71, + 555, + 287 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 555, + 287 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 555, + 287 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 555, + 287 + ], + "type": "image", + "image_path": "597b86aab741ae7722c4dd4ff1de62686462c91d9b7737a4316546686b60b18f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 54, + 295, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Evolution of robotic grasp preferences during efficient feedback-driven finetuning across 10 epochs. Top row illustrates the adjustment from hand occlusion to clear nozzle visibility. Middle row demonstrates the transition from lens obstruction to an unobstructed camera view. The bottom row shows the evolution from a top-down grasping approach to a bottom-up one, while simultaneously mitigating physical impacts." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 351, + 297, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 351, + 297, + 495 + ], + "spans": [ + { + "bbox": [ + 54, + 351, + 297, + 495 + ], + "type": "text", + "content": "poses while significantly reducing computational time. It is worth highlighting that the proposed method achieves 30x speedup (32s~38s v.s. 0.73s~2.7s) compared to these SO-TAs. After fine-tuning, our approach achieves superior performance by leveraging its evolutionary capability through iterative refinement. We also evaluate the trade-off between efficiency and grasp quality during preference alignment with different numbers of inference steps. While increasing the steps improves performance, it inevitably incurs additional computational costs. Moreover, even without guidance during sampling, we can achieve comparable results in real time." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 504, + 149, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 504, + 149, + 517 + ], + "spans": [ + { + "bbox": [ + 55, + 504, + 149, + 517 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 521, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 521, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 54, + 521, + 295, + 629 + ], + "type": "text", + "content": "We investigate the impact of different modules in the physics-aware consistency model, Physical constraints Guidance in Distillation and Sampling from which we dubbed as PGD and PGS. We also consider the performance of HPO with and without physical guidance during sampling on test split of the multidex dataset. The results presented in Tab. 2 demonstrate the key role of physical constraints and preference finetuning to successful grasping pose generation during preference alignment." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 630, + 295, + 714 + ], + "type": "text", + "content": "HPO not only effectively aligns grasping poses with human preferences at the individual object level but also enhances performance across the entire dataset. During inference across the entire dataset, we collect both successful and failed grasping samples for each object, utilizing Eq. (5) to perform lightweight finetuning of the entire model. Fig. 5 illustrates the relationship between EvolvingGrasp's" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 332, + 349, + 537, + 494 + ], + "blocks": [ + { + "bbox": [ + 332, + 349, + 537, + 494 + ], + "lines": [ + { + "bbox": [ + 332, + 349, + 537, + 494 + ], + "spans": [ + { + "bbox": [ + 332, + 349, + 537, + 494 + ], + "type": "image", + "image_path": "854e1d31947e04fe8280bd1a5bb6660829ac6d7277e90de68576bcd22d4642a6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "lines": [ + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 501, + 555, + 525 + ], + "type": "text", + "content": "Figure 5. The relationship between the success rate of Evolving-Grasp and the number of generated samples." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 313, + 538, + 555, + 598 + ], + "type": "text", + "content": "Suc.6 metric and the number of generated samples, showing that our method continuously enhances grasping performance as more samples are generated. Further ablation studies on the impact of different hyperparameters are provided in Appendix D.2." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 610, + 499, + 624 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 610, + 499, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 610, + 499, + 624 + ], + "type": "text", + "content": "4.4. Training from a Degraded Dataset" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 556, + 715 + ], + "type": "text", + "content": "In this subsection, we explore the feasibility of enhancing model performance through feedback-driven finetuning for single-object grasping tasks, particularly when trained on suboptimal data. To begin with, we randomly add the noise on the hand pose parameter of the original multidex dataset. This perturbation creates a degraded dataset, which we then use to train and distill a suboptimal model. During the in-" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11671" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 102, + 555, + 289 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Grasping performance in terms of Suc.6, Suc.1, and Pen. comparison across different methods and datasets. \"Step\" refers to inference timestep. Bold values highlight the best results and underlined values indicate the runner-up." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 102, + 555, + 289 + ], + "lines": [ + { + "bbox": [ + 57, + 102, + 555, + 289 + ], + "spans": [ + { + "bbox": [ + 57, + 102, + 555, + 289 + ], + "type": "table", + "html": "
Dataset MethodDexGraspNetMultiDexRealDexDexGRABTime ↓
Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓Suc.6 ↑Suc.1 ↑Pen. ↓
UniDexGrasp [44]33.970.131.921.647.513.527.159.439.020.855.837.40.46±0.11
GraspTTA [13]18.667.824.530.362.819.013.346.440.114.451.051.410.41±0.32
SceneDiffuser [12]26.666.931.069.885.614.621.756.142.039.185.041.13.41±0.13
UGG [24]46.979.025.255.393.410.332.763.434.442.790.633.238.34±2.31
DexGrasp Any. [49]53.690.421.572.296.39.634.671.223.156.591.828.632.91±1.34
Ours w/o HPO (2-step)60.891.019.265.697.515.241.975.419.552.293.925.10.73±0.03
Ours (2-step)62.490.519.365.997.215.344.077.819.753.392.524.30.73±0.03
Ours w/o HPO (4-step)63.893.017.475.397.113.151.682.920.555.696.023.81.41±0.07
Ours (4-step)65.292.717.276.898.413.050.682.520.357.795.223.71.41±0.07
Ours w/o HPO (8-step)65.293.516.275.698.712.263.686.621.956.896.823.12.71±0.08
Ours (8-step)65.492.315.980.398.712.364.489.121.860.896.422.32.71±0.08
Real-time (2-step)55.290.520.063.795.013.846.578.921.448.993.324.80.06±0.01
Real-time (4-step)59.990.619.864.396.511.658.287.021.155.495.424.10.10±0.02
", + "image_path": "d0a83e3ae4ced5a9941bf31816311e5658ca82bce6e59a69889ffe5589126b58.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 58, + 346, + 294, + 418 + ], + "blocks": [ + { + "bbox": [ + 55, + 303, + 297, + 337 + ], + "lines": [ + { + "bbox": [ + 55, + 303, + 297, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 303, + 297, + 337 + ], + "type": "text", + "content": "Table 2. Ablation study on incorporating physical constraints during both training and sampling stages and the LLM module. The evaluation is conducted on Multidex when timestep is 4." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 346, + 294, + 418 + ], + "lines": [ + { + "bbox": [ + 58, + 346, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 58, + 346, + 294, + 418 + ], + "type": "table", + "html": "
CMPGDPGSHPOSuc.6 ↑Suc.1 ↑Pen. ↓Time ↓
a✓60.094.614.00.10±0.02
b✓✓64.396.512.50.10±0.02
c✓✓66.295.614.91.41±0.07
d✓✓✓67.596.511.90.10±0.02
e✓✓✓75.397.113.11.41±0.07
f✓✓✓✓76.898.413.01.41±0.07
", + "image_path": "02a7b28a68d870b975b5f3909b9d41d1f51efcf7455782c2cbc9f2054d426a47.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 73, + 421, + 278, + 552 + ], + "blocks": [ + { + "bbox": [ + 73, + 421, + 278, + 552 + ], + "lines": [ + { + "bbox": [ + 73, + 421, + 278, + 552 + ], + "spans": [ + { + "bbox": [ + 73, + 421, + 278, + 552 + ], + "type": "image", + "image_path": "347377672a72be5f4464839c7b1b72eadd8df24e41a723651f7f4cf354a5cc67.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 560, + 297, + 605 + ], + "lines": [ + { + "bbox": [ + 55, + 560, + 297, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 560, + 297, + 605 + ], + "type": "text", + "content": "Figure 6. The variation of the mean grasping success rate of the evolving model for randomly selected objects in the multidex dataset with increasing finetuning epochs. The blue dashed line represents the grasping success rate under the original model." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 296, + 714 + ], + "type": "text", + "content": "ference phase of this suboptimal model, we collect successfully grasping poses as positive samples and all other poses as negative samples. Fig. 6 illustrates the change in the average grasping success rate of the suboptimal model on randomly sampled objects from the multidex dataset as the number of fine-tuning epochs increases. The results indicate that we continuously improve the success rate during the evolutionary adaptation process and ultimately outper" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 304, + 555, + 359 + ], + "blocks": [ + { + "bbox": [ + 316, + 304, + 555, + 359 + ], + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 359 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 359 + ], + "type": "image", + "image_path": "515f9ebd32d356ba4083810aa5534e7907e440e62cbb961b1800640966305eb4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 339, + 366, + 529, + 378 + ], + "lines": [ + { + "bbox": [ + 339, + 366, + 529, + 378 + ], + "spans": [ + { + "bbox": [ + 339, + 366, + 529, + 378 + ], + "type": "text", + "content": "Figure 7. Real-world deployment on Shadow Hand." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 389, + 478, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 389, + 478, + 402 + ], + "spans": [ + { + "bbox": [ + 314, + 389, + 478, + 402 + ], + "type": "text", + "content": "form the accuracy of the original model." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 415, + 451, + 428 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 451, + 428 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 451, + 428 + ], + "type": "text", + "content": "4.5. Real World Deployment" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 435, + 555, + 544 + ], + "type": "text", + "content": "To verify that EvolvingGrasp can continuously enhance grasping performance and align with human preferences in real world, we deploy our model on a real ShadowHand robot, as shown in Fig. 7. The pre-grasping motion trajectory is generated based on RealDex [20]. Real-world experiments demonstrate that our method achieves successful grasping through several efficient preference finetuning when the initial grasp attempt fails. Additional video demonstrations can be found in supplementary materials." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 559, + 388, + 572 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 559, + 388, + 572 + ], + "spans": [ + { + "bbox": [ + 313, + 559, + 388, + 572 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "We propose EvolvingGrasp, an evolutionary grasp generation method through efficient preference alignment. HPO is introduced to allow the model to continuously align the performance with preference signals. We design the Physics-Aware Consistency Model to achieve both rapid inference and efficient preference finetuning while maintaining physical plausibility. Extensive experiments across four benchmarks demonstrate that our method achieves state-of-the-art results. Furthermore, we deploy our model on a real ShadowHand robot to validate its evolutionary capability in real-world scenarios." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "11672" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 294, + 124 + ], + "type": "text", + "content": "[1] Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. arXiv preprint arXiv:2305.13301, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 125, + 294, + 158 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 125, + 294, + 158 + ], + "spans": [ + { + "bbox": [ + 61, + 125, + 294, + 158 + ], + "type": "text", + "content": "[2] Yuxin Chen, Devesh K Jha, Masayoshi Tomizuka, and Diego Romeres. Fdpp: Fine-tune diffusion policy with human preference. arXiv preprint arXiv:2501.08259, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "spans": [ + { + "bbox": [ + 62, + 159, + 295, + 203 + ], + "type": "text", + "content": "[3] Hyungjin Chung, Jeongsol Kim, Michael T McCann, Marc L Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. arXiv preprint arXiv:2209.14687, 2022. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 294, + 237 + ], + "type": "text", + "content": "[4] Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable rewards. arXiv preprint arXiv:2309.17400, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 62, + 239, + 294, + 282 + ], + "type": "text", + "content": "[5] Hongkai Dai, Anirudha Majumdar, and Russ Tedrake. Synthesis and optimization of force closure grasps via sequential semidefinite programming. Robotics Research: Volume 1, pages 285-305, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 285, + 294, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 294, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 294, + 328 + ], + "type": "text", + "content": "[6] Fei Deng, Qifei Wang, Wei Wei, Tingbo Hou, and Matthias Grundmann. Prdp: Proximal reward difference prediction for large-scale reward finetuning of diffusion models. In CVPR, pages 7423-7433, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 331, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 331, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 331, + 294, + 363 + ], + "type": "text", + "content": "[7] Prafulla Dhariwal and Alexander Nichol. Diffusion models beat gans on image synthesis. Advances in neural information processing systems, 34:8780-8794, 2021. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 365, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 365, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 365, + 294, + 430 + ], + "type": "text", + "content": "[8] Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Dpok: Reinforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Systems, 36:79858-79885, 2023. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 432, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 294, + 464 + ], + "type": "text", + "content": "[9] Carlo Ferrari, John Canny, et al. Planning optimal grasps. In IEEE International Conference on Robotics and Automation, pages 2290-2295. IEEE, 1992. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 466, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 466, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 57, + 466, + 294, + 521 + ], + "type": "text", + "content": "[10] Ayano Hiranaka, Shang-Fu Chen, Chieh-Hsin Lai, Dongjun Kim, Naoki Murata, Takashi Shibuya, Wei-Hsiang Liao, Shao-Hua Sun, and Yuki Mitsufuji. Human-feedback efficient reinforcement learning for online diffusion model finetuning. arXiv preprint arXiv:2410.05116, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 57, + 523, + 294, + 566 + ], + "type": "text", + "content": "[11] Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 5, 12" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 568, + 294, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 568, + 294, + 611 + ], + "spans": [ + { + "bbox": [ + 57, + 568, + 294, + 611 + ], + "type": "text", + "content": "[12] Siyuan Huang, Zan Wang, Puhao Li, Baoxiong Jia, Tengyu Liu, Yixin Zhu, Wei Liang, and Song-Chun Zhu. Diffusion-based generation, optimization, and planning in 3d scenes. In CVPR, pages 16750-16761, 2023. 2, 6, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 656 + ], + "type": "text", + "content": "[13] Hanwen Jiang, Shaowei Liu, Jiashun Wang, and Xiaolong Wang. Hand-object contact consistency reasoning for human grasps generation. In ICCV, pages 11107-11116, 2021. 2, 6, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 658, + 294, + 713 + ], + "type": "text", + "content": "[14] Dongjun Kim, Chieh-Hsin Lai, Wei-Hsiang Liao, Naoki Murata, Yuhta Takida, Toshimitsu Uesaka, Yutong He, Yuki Mitsufuji, and Stefano Ermon. Consistency trajectory models: Learning probability flow ode trajectory of diffusion. arXiv preprint arXiv:2310.02279, 2023. 3" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 127 + ], + "type": "text", + "content": "[15] Haosheng Li, Weixin Mao, Weipeng Deng, Chenyu Meng, Haoqiang Fan, Tiancai Wang, Ping Tan, Hongan Wang, and Xiaoming Deng. Multi-grasplm: A multimodal llm for multi-hand semantic guided grasp generation. arXiv preprint arXiv:2412.08468, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 553, + 183 + ], + "type": "text", + "content": "[16] Puhao Li, Tengyu Liu, Yuyang Li, Yiran Geng, Yixin Zhu, Yaodong Yang, and Siyuan Huang. Gendexgrasp: Generalizable dexterous grasping. In IEEE International Conference on Robotics and Automation), pages 8068-8074. IEEE, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 239 + ], + "type": "text", + "content": "[17] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Generating grasp poses for a high-dof gripper using neural networks. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1518-1525. IEEE, 2019. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[18] Min Liu, Zherong Pan, Kai Xu, Kanishka Ganguly, and Dinesh Manocha. Deep differentiable grasp planner for high-dof grippers. arXiv preprint arXiv:2002.01530, 2020. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 276, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 276, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 317, + 276, + 553, + 329 + ], + "type": "text", + "content": "[19] Tengyu Liu, Zeyu Liu, Ziyuan Jiao, Yixin Zhu, and Song-Chun Zhu. Synthesizing diverse and physically stable grasps with arbitrary hand structures using differentiable force closure estimator. IEEE Robotics and Automation Letters, 7(1): 470-477, 2021. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 385 + ], + "type": "text", + "content": "[20] Yumeng Liu, Yaxun Yang, Youzhuo Wang, Xiaofei Wu, Ji-amin Wang, Yichen Yao, Soren Schwertfeger, Sibei Yang, Wenping Wang, Jingyi Yu, et al. Realdex: Towards human-like grasping for robotic dexterous hand. arXiv preprint arXiv:2402.13853, 2024. 6, 8" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "spans": [ + { + "bbox": [ + 316, + 388, + 553, + 420 + ], + "type": "text", + "content": "[21] Cheng Lu and Yang Song. Simplifying, stabilizing and scaling continuous-time consistency models. arXiv preprint arXiv:2410.11081, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "spans": [ + { + "bbox": [ + 317, + 422, + 553, + 475 + ], + "type": "text", + "content": "[22] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. Advances in Neural Information Processing Systems, 35:5775-5787, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 521 + ], + "type": "text", + "content": "[23] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 555 + ], + "type": "text", + "content": "[24] Jiaxin Lu, Hao Kang, Haoxiang Li, Bo Liu, Yiding Yang, Qixing Huang, and Gang Hua. Ugg: Unified generative grasping. arXiv preprint arXiv:2311.16917, 2023. 2, 6, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 557, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 557, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 557, + 553, + 621 + ], + "type": "text", + "content": "[25] Jens Lundell, Enric Corona, Tran Nguyen Le, Francesco Verdoja, Philippe Weinzaepfel, Grégory Rogez, Francesc Moreno-Noguer, and Ville Kyrki. Multi-fingan: Generative coarse-to-fine sampling of multi-finger grasps. In IEEE International Conference on Robotics and Automation, pages 4495-4501. IEEE, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 656 + ], + "type": "text", + "content": "[26] Jens Lundell, Francesco Verdoja, and Ville Kyrki. Ddgc: Generative deep dexterous grasping in clutter. IEEE Robotics and Automation Letters, 6(4):6899-6906, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[27] Viktor Makoviychuk, Lukasz Wawrzyniak, Yunrong Guo, Michelle Lu, Kier Storey, Miles Macklin, David Hoeller, Nikita Rudin, Arthur Allshire, Ankur Handa, et al. Isaac gym: High performancegpu-based physics simulation for robot learning. arXiv preprint arXiv:2108.10470, 2021. 6" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "11673" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 137 + ], + "type": "text", + "content": "[28] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 13" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 195 + ], + "type": "text", + "content": "[29] Jean Ponce, Steve Sullivan, J-D Boissonnat, and J-P Merlet. On characterizing and computing three-and four-finger force-closure grasps of polyhedral objects. In IEEE International Conference on Robotics and Automation, pages 821-827. IEEE, 1993. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[30] Rafael Rafailov, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36:53728-53741, 2023. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[31] Javier Romero, Dimitrios Tzionas, and Michael J Black. Embodied hands: Modeling and capturing hands and bodies together. arXiv preprint arXiv:2201.02610, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[32] Carlos Rosales, Raul Suárez, Marco Gabiccini, and Antonio Bicchi. On the synthesis of feasible and prehensile robotic grasps. In IEEE International Conference on Robotics and Automation, pages 550-556. IEEE, 2012. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 364 + ], + "type": "text", + "content": "[33] Yanming Shao and Chenxi Xiao. Bimanual grasp synthesis for dexterous robot hands. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[34] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 400, + 294, + 443 + ], + "type": "text", + "content": "[35] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 499 + ], + "type": "text", + "content": "[36] Omid Taheri, Nima Ghorbani, Michael J Black, and Dimitrios Tzionas. Grab: A dataset of whole-body human grasping of objects. In Computer Vision-ECCV 2020: 16th European Conference, Glasgow, UK, August 23-28, 2020, Proceedings, Part IV 16, pages 581-600. Springer, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 502, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 502, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 502, + 294, + 544 + ], + "type": "text", + "content": "[37] Xiaofeng Tan, Hongsong Wang, Xin Geng, and Pan Zhou. Sopo: Text-to-motion generation using semi-online preference optimization. arXiv preprint arXiv:2412.05095, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 547, + 294, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 294, + 600 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 294, + 600 + ], + "type": "text", + "content": "[38] Bram Wallace, Meihua Dang, Rafael Rafailov, Linqi Zhou, Aaron Lou, Senthil Purushwalkam, Stefano Ermon, Caiming Xiong, Shafiq Joty, and Nikhil Naik. Diffusion model alignment using direct preference optimization. In CVPR, pages 8228-8238, 2024. 2, 3, 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 602, + 294, + 667 + ], + "type": "text", + "content": "[39] Ruicheng Wang, Jialiang Zhang, Jiayi Chen, Yinzhen Xu, Puhao Li, Tengyu Liu, and He Wang. Dexgraspnet: A large-scale robotic dexterous grasp dataset for general objects based on simulation. In IEEE International Conference on Robotics and Automation, pages 11359-11366. IEEE, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Wei Wei, Daheng Li, Peng Wang, Yiming Li, Wanyi Li, Yongkang Luo, and Jun Zhong. Dvgg: Deep variational grasp generation for dextrous manipulation. IEEE Robotics and Automation Letters, 7(2):1659-1666, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 553, + 542 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 553, + 116 + ], + "type": "text", + "content": "[41] Yi-Lin Wei, Jian-Jian Jiang, Chengyi Xing, Xian-Tuo Tan, Xiao-Ming Wu, Hao Li, Mark Cutkosky, and Wei-Shi Zheng. Grasp as you say: Language-guided dexterous grasp generation. arXiv preprint arXiv:2405.19291, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 160 + ], + "type": "text", + "content": "[42] Zehang Weng, Haofei Lu, Danica Kragic, and Jens Lundell. Dexdiffuser: Generating dexterous grasps with diffusion models. IEEE Robotics and Automation Letters, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 195 + ], + "type": "text", + "content": "[43] Guo-Hao Xu, Yi-Lin Wei, Dian Zheng, Xiao-Ming Wu, and Wei-Shi Zheng. Dexterous grasp transformer. In CVPR, pages 17933-17942, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 261 + ], + "type": "text", + "content": "[44] Yinzhen Xu, Weikang Wan, Jialiang Zhang, Haoran Liu, Zikang Shan, Hao Shen, Ruicheng Wang, Haoran Geng, Yijia Weng, Jiayi Chen, et al. Unidxexgrasp: Universal robotic dexterous grasping via learning diverse proposal generation and goal-conditioned policy. In CVPR, pages 4737-4746, 2023. 6, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 306 + ], + "type": "text", + "content": "[45] Kai Yang, Jian Tao, Jiafei Lyu, Chunjiang Ge, Jiaxin Chen, Weihan Shen, Xiaolong Zhu, and Xiu Li. Using human feedback to fine-tune diffusion models without any reward model. In CVPR, pages 8941-8951, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 351 + ], + "type": "text", + "content": "[46] Lingxiao Yang, Shutong Ding, Yifan Cai, Jingyi Yu, Jingya Wang, and Ye Shi. Guidance with spherical gaussian constraint for conditional diffusion. arXiv preprint arXiv:2402.03201, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 553, + 396 + ], + "type": "text", + "content": "[47] Zijian Zhang, Kaiyuan Zheng, Zhaorun Chen, Joel Jang, Yi Li, Chaoqi Wang, Mingyu Ding, Dieter Fox, and Huaxiu Yao. Grape: Generalizing robot policy via preference alignment. arXiv preprint arXiv:2411.19309, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 553, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 462 + ], + "type": "text", + "content": "[48] Zhengshen Zhang, Lei Zhou, Chenchen Liu, Zhiyang Liu, Chengran Yuan, Sheng Guo, Ruiteng Zhao, Marcelo H Ang Jr, and Francis EH Tay. Dexgrasp-diffusion: Diffusion-based unified functional grasp synthesis method for multideterous robotic hands. arXiv preprint arXiv:2407.09899, 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 497 + ], + "type": "text", + "content": "[49] Yiming Zhong, Qi Jiang, Jingyi Yu, and Yuexin Ma. Dexterity: Towards universal robotic dexterous grasping with physics awareness. In CVPR, 2025. 2, 5, 6, 8, 13" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "spans": [ + { + "bbox": [ + 316, + 498, + 553, + 542 + ], + "type": "text", + "content": "[50] Zhenglin Zhou, Xiaobo Xia, Fan Ma, Hehe Fan, Yi Yang, and Tat-Seng Chua. Dreamdpo: Aligning text-to-3d generation with human preferences via direct preference optimization. arXiv preprint arXiv:2502.04370, 2025. 2" + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "11674" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_content_list.json b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..24bacb62b79cf69f1c849aa788a26da114d417d8 --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_content_list.json @@ -0,0 +1,1492 @@ +[ + { + "type": "text", + "text": "ExCap3D: Expressive 3D Scene Understanding via Object Captioning with Varying Detail", + "text_level": 1, + "bbox": [ + 147, + 128, + 849, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Chandan Yeshwanth", + "bbox": [ + 246, + 203, + 413, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Dávid Rozenberszki", + "bbox": [ + 446, + 203, + 609, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Angela Dai", + "bbox": [ + 650, + 203, + 743, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Technical University of Munich", + "bbox": [ + 369, + 228, + 622, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://cy94.github.io/excap3d/", + "bbox": [ + 310, + 253, + 679, + 271 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0991bec62ee5f861c0a6a8add2c523d099186590c1871e0e0e4ff5cb0943fc4c.jpg", + "image_caption": [ + "Figure 1. ExCap3D produces descriptions at multiple level of details for objects in an input 3D scan. For each detected object, we generate both object- and part-level of detail captions. The generated captions are consistent across the two levels and contain rich details." + ], + "image_footnote": [], + "bbox": [ + 93, + 309, + 903, + 468 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 520, + 326, + 535 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Generating text descriptions of objects in 3D indoor scenes is an important building block of embodied understanding. Existing methods describe objects at a single level of detail and do not capture fine-grained details of the parts of objects. In order to produce varying levels of detail capturing both coarse object-level information and detailed part-level descriptions, we propose the task of expressive 3D captioning. Given an input 3D scene, the task is to describe objects at multiple levels of detail: a high-level object description, and a low-level description of the properties of its parts. To produce such captions, we present ExCap3D, an expressive 3D captioning model which takes as input a 3D scan, and for each detected object in the scan, generates a fine-grained collective description of the parts of the object, along with an object-level description conditioned on the part-level description. We design ExCap3D to encourage consistency between the multiple levels of descriptions. To enable this task, we generated the ExCap3D Dataset by leveraging a visual-language model (VLM) for multi-view captioning. The ExCap3D Dataset contains captions on the ScanNet++ dataset with varying levels of detail, comprising 190k text descriptions of 34k 3D objects in 947 indoor scenes. Our experiments show that the object- and", + "bbox": [ + 88, + 553, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "part-level details generated by ExCap3D are more expressive than those produced by state-of-the-art methods, with a CIDEr score improvement of $17\\%$ and $124\\%$ for object- and part-level details respectively. Our code, dataset and models will be made publicly available.", + "bbox": [ + 511, + 521, + 908, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 611, + 643, + 626 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "3D indoor scene understanding is central to applications such as augmented reality (AR), virtual reality (VR), and robotics. While detecting 3D objects and their semantics is important for these applications, natural language descriptions of objects enable more wide-ranging possibilities – they can encode more complex information efficiently and, importantly, enable more natural interactions between embodied agents as well as humans and the scene. To produce a representation that facilities a diverse array of applications, such language descriptions must also encompass varying levels of detail: object-level descriptions for coarse identification in the scene, and part-level details for precise manipulation and fine-grained understanding. For instance, a robotic assistant would only need a coarse, high-level description of the objects in a room (e.g., 'a recliner chair') in order to navigate effectively in it, but it would need fine-grained descriptions of the objects (e.g., 'recliner chair with a high padded back, wooden armrests, a soft cushion and an", + "bbox": [ + 511, + 628, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21699", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adjustable footrest') to provide assistive AI descriptions for accessibility aid or to describe the scene to others.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current 3D captioning models have made significant progress in describing objects in indoor scenes by detecting objects and generating their individual descriptions at a single level of detail. These methods are trained only for captioning [10, 15, 32] or for multiple tasks [11, 55, 59, 60]. These methods can effectively describe coarse object information in the global context of the scene, but preclude detailed perception of the scene as they make limited use of fine-grained 3D inputs and cannot model object-part hierarchy. They fail to describe highly localized properties of object parts, such as appearance, material, and function. In order to enable more complex perception of a scene requiring these details, we propose to make use of fine-grained 3D inputs and explicitly model object-part hierarchy. We describe objects at multiple levels: an object-level of detail that captures its most prominent aspects (semantic class; possibly size or overall appearance), and a finer-grained part-level of detail that describes the details of elements composing the object (different materials, color, and functions). These two descriptions are complementary to each other and enable different applications depending on the level of detail required. Examples of such multilevel descriptions are shown in Fig. 1 and Fig. 2.", + "bbox": [ + 91, + 125, + 483, + 472 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To generate such multilevel captions in 3D scans, we introduce ExCap3D, a 3D object captioning model that produces expressive captions at both object- and part-levels of detail using two captioning heads. Examples of the captions generated by our method are shown in Fig. 1. The generated multilevel descriptions are complementary to each other and describe different aspects of the objects, while maintaining a few shared elements for consistency and coherence. The part-level caption generated by ExCap3D collectively describes the different parts of an object. Knowing the descriptions of the parts can enable our model to generate more consistent captions describing the object as a whole. Hence, ExCap3D first produces the part-level details and then uses these to inform the object-level captioner. To further reduce inconsistencies between the two levels of detail in terms of the object semantics and text content, ExCap3D employs additional consistency losses during training.", + "bbox": [ + 89, + 474, + 482, + 746 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To enable this multilevel captioning task, we provide the ExCap3D Dataset, which contains 190k object and partlevel natural language captions of 34k 3D objects from the 947 scenes of the ScanNet++ dataset. We automatically generate object captions for training, using a pretrained vision-language model (VLM) on high-resolution DSLR images along with ground truth semantic information from the ScanNet++ dataset [52]. We employ multi-view aggregation to robustly capture key object- and part-level details. VLMs perform strongly on 2D tasks, but", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d3a94ac01a17ef33178809c2442c6109ed8ed8d9e6019b4cbb18d069ddafed88.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 517, + 90, + 566, + 140 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/61347f251bbea1099252ff1792990875aca0fab4d8f536a18f6dcd5773a1ba1a.jpg", + "image_caption": [ + "a brown cushion chair with arms. it is located close to the wall.", + "Q: What is next to the orange armchair? \nA: wall \nA: another orange armchair \nA: coffee table" + ], + "image_footnote": [], + "bbox": [ + 645, + 90, + 774, + 140 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/49f2febe57af4f2c95bc88c3a0e2756e939228e6fc6fe63b98f904873dea2149.jpg", + "image_caption": [ + "ExCap3D", + "Object-level: A three-seater leather couch with dark, smooth, glossy leather, curved and straight edges. \nPart-Level: The black leather sofa has a smooth, glossy surface with visible stitching, a curved armrest, and cushioned seat, with a tufted backrest made of multiple leather panels stitched together, and sturdy structure with visible stitching along the edges.", + "Figure 2. Descriptions of objects in existing datasets such as Scan2Cap and ScanQA largely contain relations between objects, and limited local details at a single level. In contrast, the ExCap3D Dataset describes both the object as a whole and as a sum of its parts." + ], + "image_footnote": [], + "bbox": [ + 517, + 152, + 643, + 199 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "struggle with 3D orientation [47] and distances [8]. Hence, our dataset contains large-scale and high-quality object- and part-level captions, in contrast to Scan2Cap [15] and Nr3D [2] with smaller-scale manually annotated spatial relations. A comparison of the multilevel captions from the ExCap3D Dataset with the single-level descriptions from the existing Scan2Cap and ScanQA [3] is shown in Fig. 2. The ExCap3D Dataset describes objects with significantly more detail, while providing two complementary multilevel descriptions of each object.", + "bbox": [ + 511, + 311, + 903, + 462 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We show that ExCap3D produces richer and more consistent object- and part-level captions compared to the state-of-the-art captioning methods, originally designed to generate only a single level description. To summarize, our main contributions are:", + "bbox": [ + 511, + 463, + 903, + 537 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- an expressive 3D captioning model that generates captions of objects in 3D indoor scenes at multiple levels of detail: fine-grained part-level details and object-level details", + "- modeling object-level detail as a sum of its parts' unique appearance and functional attributes by generating object-level captions conditioned on part-level descriptions, using the cross-attention mechanism of the object captioner and the language model's intermediate hidden states", + "- regularization of our joint 3D captioning approach through semantic and textual-similarity consistency losses that maintain consistency of the object-level and part-level predicted captions", + "- the ExCap3D Dataset: a large-scale, precise and challenging dataset of 190k descriptive object- and part-level captions of 34k objects in 3D indoor scenes built on the ScanNet++ dataset with fine-grained texture, material, and functional descriptions" + ], + "bbox": [ + 513, + 539, + 903, + 810 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 823, + 653, + 838 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. 2D Image Captioning", + "text_level": 1, + "bbox": [ + 511, + 848, + 714, + 864 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2D image captioning has been extensively studied in computer vision. DenseCap [29] popularized the task of dense", + "bbox": [ + 511, + 869, + 903, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21700", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "captioning of images, jointly localizing objects, and generating their short descriptions. More recently, COCONutPanCap [19] and Pix2Cap-COCO [53] combined the tasks of fine-grained instance segmentation and detailed captioning. FlexCap [22] and FineCaption [26] can produce descriptions at varying levels of detail given the image and a localized object as input. In contrast, we combine the tasks of detection and multilevel captioning and describe detected objects and their parts.", + "bbox": [ + 89, + 90, + 480, + 227 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. 3D Scene Captioning", + "text_level": 1, + "bbox": [ + 89, + 255, + 287, + 272 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Scan2Cap [15] introduced the task of 3D dense captioning and a method for context-aware captioning of 3D objects, operating on ScanNet [18] scenes. Since then, several methods have been developed to effectively solve this task by extracting object features and modeling their relations with nearby objects [7, 9, 10, 28, 31, 49, 54]. These methods can accurately describe objects at the scene-level through relations with neighboring objects. Following the trends of utilizing foundation models in 3D reasoning, Cap3D [38] and follow-up works [24, 39] describe isolated 3D objects through multi-view aggregation of VLM-generated captions. More recent works have jointly trained on multiple tasks such as 3D grounding, question answering, and dense captioning by training custom language models from scratch [16, 58] or by fine-tuning pretrained LLMs [11, 25, 27, 50, 57, 60]. However, these works focus on high-level, proximity-based relations between objects at the scene-level, underutilize local features to produce limited object-level details and no part-level details We tackle a similarly inspired problem, and leverage the high-fidelity geometry and RGB captures from ScanNet++ [52] to enable complex, multilevel joint captioning. Moreover, we demonstrate that our joint captioning approach improves both levels of detail.", + "bbox": [ + 91, + 284, + 483, + 647 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Part Understanding of 3D Objects in Scenes", + "text_level": 1, + "bbox": [ + 89, + 675, + 465, + 691 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Several methods have focused on part-level understanding of 3D scenes by only relying on explicit geometry [5, 6], images and radiance fields [4, 30] or by using both modalities and mesh-based oversegmentation [46]. A related method for 3D shapes rather than scenes was proposed by Chen et. al. [13], which produces part-level segmentation maps and text descriptions of individual 3D objects at a single, fine-grained level of detail. This produces promising detailed captions, but is designed for isolated objects rather than complex 3D scenes. In contrast, we jointly detect and caption objects in large 3D scenes at multiple levels of detail, ensuring consistency and enabling greater expressivity for downstream tasks.", + "bbox": [ + 89, + 704, + 482, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 89, + 604, + 106 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Problem Formulation The task of expressive captioning is to detect the objects in a 3D scene and describe each object at different levels of detail: object- and part-levels. Given an input 3D scene in the form of a 3D mesh $M = (\\mathcal{V},\\mathcal{F})$ with vertices $\\mathcal{V}\\in \\mathbb{R}^{N_v\\times 3}$ and triangular faces $f_{i}\\in \\mathcal{F}$ connecting the vertices, we aim to detect the 3D object instances $o_i = (b_i,s_i,c_{\\mathrm{obj},i},c_{\\mathrm{part},i})$ in the scene, characterized by the objects' bounding boxes $b_{i}$ , semantic class $s_i$ , and $c_{\\mathrm{obj},i}$ and $c_{\\mathrm{part},i}$ describing the object in natural language at object- and part-levels of detail, respectively.", + "bbox": [ + 511, + 114, + 906, + 267 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Overview of ExCap3D First, ExCap3D performs instance segmentation and instance-aware query generation to produce queries containing rich per-object information. Next, the object- and part-level captioning models with shared information generate two separate text captions, conditioned on individual instance-aware queries and the corresponding features from the instance segmentation backbone. Finally, semantic and whole-text consistency losses are applied to the hidden states of the captioning models to regularize the consistency of the predicted captions.", + "bbox": [ + 511, + 282, + 906, + 435 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. 3D Instance Segmentation", + "text_level": 1, + "bbox": [ + 511, + 441, + 751, + 458 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given the mesh of the 3D scene $M$ as input, the instance segmentation model produces a fixed number of refined instance-aware queries $Q_{r}$ corresponding to the $N_{q}$ objects detected in the scene, and their binary instance masks on the voxelized 3D input $I\\in \\{0,1\\}^{N_q\\times N_{\\mathrm{vox}}}$ . The instance queries are a compact representation of the 3D features of the objects as well as their semantics. The instance segmentation model also produces dense features $F\\in \\mathbb{R}^{N_{\\mathrm{vox}}\\times D}$ on the voxels. The refined and instance-aware queries $Q_{r}$ and 3D voxel features $F$ are used to inform the per-instance captioning in the next step.", + "bbox": [ + 511, + 463, + 908, + 628 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The input mesh is first voxelized into a voxel grid $V \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times 6}$ containing the voxel coordinates and RGB colors. We use the Mask3D [45] transformer-based model for instance segmentation. It consists of a 3D sparse convolutional UNet backbone to encode the voxelized 3D scene, which outputs features $F \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times D}$ over the input voxels. A parallel mask module takes as input query vectors $Q \\in \\mathbb{R}^{N_q \\times D_q}$ denoting the $N_q$ instances in the scene. The queries are refined iteratively through several transformer encoder layers, with cross-attention to the 3D backbone features. The output instance-aware query vectors contain rich representations of the object instances in the 3D scene and are used to predict the instance masks $I \\in \\{0,1\\}^{N_q \\times N_{\\mathrm{vox}}}$ and semantic class of each object.", + "bbox": [ + 511, + 630, + 908, + 840 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Joint Captioning with Information Sharing", + "text_level": 1, + "bbox": [ + 511, + 848, + 880, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ExCap3D contains two separate captioners for the object- and part-levels of details. These are transformer-based lan", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21701", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/411fdae6ebb29bf9fbb66498e538d4530434426e1f9ec170b0befa1a2ea82bef.jpg", + "image_caption": [ + "Figure 3. Overview of our ExCap3D captioning method. We predict 3D instances in the input 3D scene using Mask3D [45], then predict object-level and part-levels of detail using two separate captioning heads. The object-level captions are further constrained by the part-level captioner's hidden states. Semantic- and textual consistency losses are applied to ensure the overall consistency of both predicted captions." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 867, + 305 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "guage models $\\Psi_{\\mathrm{obj}}$ and $\\Psi_{\\mathrm{part}}$ that produce the object-level caption $c_{\\mathrm{obj},o}$ , part-level caption $c_{\\mathrm{part},o}$ . They are conditioned on the instance-aware queries $Q_{r}$ and the dense voxel features $F$ , and are trained for next-token prediction. The language model captioners take as input a sequence of previously predicted tokens $t_1 \\ldots t_{i-1}$ and decode the next token $t_i$ autoregressively as", + "bbox": [ + 88, + 383, + 483, + 491 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(t _ {i} \\mid t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(t _ {1} \\dots t _ {i - 1}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 166, + 506, + 403, + 522 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Psi$ is the language model. We provide two sources of rich instance-level information to the captioner: the instance-aware refined query output by the instance-segmentation model, and voxel features from 3D UNet backbone $F$ . The refined query vector $Q_{r,o}$ corresponding to an object $o$ is first projected with a linear layer to generate the caption-aware query $Q_{c,o} \\in \\mathbb{R}^{D_{\\mathrm{caption}}}$ for that object", + "bbox": [ + 88, + 529, + 483, + 638 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nQ _ {c, o} = \\Phi_ {\\mathrm {q u e r y}} (Q _ {r, o}),\n$$\n", + "text_format": "latex", + "bbox": [ + 210, + 648, + 362, + 666 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{\\mathrm{caption}}$ is the embedding dimension of the language model. $Q_{c,o}$ contains localized information required to generate the caption for an object, such as its semantic class and its high-level features, and is used as the initial token to start the caption generation as", + "bbox": [ + 88, + 676, + 483, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np \\left(t _ {i} \\mid Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right).\n$$\n", + "text_format": "latex", + "bbox": [ + 130, + 781, + 441, + 797 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Additionally, we inject fine-grained information by using the captioner's cross-attention layers to attend to fine-grained contextual features from the 3D backbone $F$ corresponding to the object being captioned, after a linear projection to $D_{\\mathrm{caption}}$ . To reduce the computational complexity of the cross-attention operation, we aggregate features within", + "bbox": [ + 88, + 809, + 483, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "precomputed segments [23] on the mesh surface, giving the context features $S_{o} \\in \\mathbb{R}^{n_{s,o} \\times D_{\\mathrm{caption}}}$ , where $n_{s,o}$ is the number of segments in $I_{o}$ , the predicted mask for object $o$ . The cross-attention operation is $A = \\operatorname{softmax}(qk^T / \\sqrt{d})v$ where queries $q$ and values $k$ are the hidden states of $\\Psi$ , and the keys $k$ are $S_{o}$ . The object and part captioners use context features $S_{\\mathrm{obj},o}$ and $S_{\\mathrm{part},o}$ respectively.", + "bbox": [ + 511, + 383, + 905, + 489 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Importantly, to enable the object-level captioner to be informed by the part-level captioner, we make use of the last layer hidden states $[h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},i}]$ of the part-level captioner $\\Psi_{\\mathrm{part}}$ . We use a linear projection", + "bbox": [ + 511, + 489, + 905, + 551 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nH _ {\\text {p a r t}, o} = \\Phi_ {\\text {h i d d e n}} ([ h _ {\\text {p a r t}, 1} \\dots h _ {\\text {p a r t}, i} ])\n$$\n", + "text_format": "latex", + "bbox": [ + 578, + 565, + 839, + 583 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and then concatenate $H_{\\mathrm{part},o}$ to the contextual features of the object captioner $S_{\\mathrm{obj},o}$ for object $o$ . The object-level caption for object $o$ is finally informed by the caption-aware query $Q_{c,o}$ as well as features $[H_{\\mathrm{part},o};S_{\\mathrm{obj},o}]$ . Since different objects in a batch have different numbers of predicted segments, we zero-pad $S_{\\mathrm{obj},o}$ to the maximum context length in the batch.", + "bbox": [ + 511, + 589, + 905, + 694 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The two caption models are trained with cross-entropy (CE) loss against the object- and part-level ground truth (GT) captions respectively, giving the loss", + "bbox": [ + 511, + 694, + 905, + 742 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c a p t i o n}} = \\mathcal {L} _ {\\text {o b j - c a p t i o n}} + \\mathcal {L} _ {\\text {p a r t - c a p t i o n}}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 755, + 844, + 772 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The outputs of this stage are the object-level caption $c_{\\mathrm{obj},o}$ , part-level caption $c_{\\mathrm{part},o}$ and the corresponding language model hidden states $h_{\\mathrm{obj}} = [h_{\\mathrm{obj},1}\\dots h_{\\mathrm{obj},i}]$ and $h_{\\mathrm{part}} = [h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},j}]$ for each object $o$ .", + "bbox": [ + 511, + 777, + 905, + 842 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Semantic and Textual Consistency Losses", + "text_level": 1, + "bbox": [ + 511, + 848, + 867, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "While the object-level of detail is informed by the part-level of detail, supervising these two captioners with different GT", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21702", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "captions may lead to inconsistencies in the predicted captions due to the different GT distributions of the two levels of detail. Hence, we use additional constraints in the form of consistency losses over the predicted captions. We encourage semantic consistency for both captions to refer to the same object semantics, such as its fine-grained object class and material, and textual consistency for the captions to be similar to each other. Since consistency losses cannot be applied directly on the generated text due to the text tokenizer being non-differentiable, we instead use the language model hidden states $h_{\\mathrm{obj}}$ and $h_{\\mathrm{part}}$ from the last captioner layers as a proxy.", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic Consistency To promote both levels of captions to contain similar semantic information, we use a semantic consistency loss. We do this by projecting the hidden states $h_{\\mathrm{obj}}$ and $h_{\\mathrm{part}}$ into a lower dimension to retain only meaningful semantic information over the whole caption. We then pass these sequences independently through object- and part-level caption projection transformers $\\Phi_{\\mathrm{obj,sem}}$ and $\\Phi_{\\mathrm{part,sem}}$ to obtain new embedded sequences, and these aggregate into single feature vectors representing the overall semantics of the predicted captions $h_{\\mathrm{obj,sem}} \\in \\mathbb{R}^{D_{ff}}$ and $h_{\\mathrm{part,sem}} \\in \\mathbb{R}^{D_{ff}}$ respectively, where $D_{ff}$ is the feed-forward dimension of $\\Phi$ . These are again independently projected into $N_{\\mathrm{sem}}$ fine-grained latent classes using linear layers to get class probabilities $sem_{\\mathrm{obj}}$ and $sem_{\\mathrm{part}}$ . The semantic consistency loss is a symmetric sum [14] of cross-entropy (CE) losses", + "bbox": [ + 89, + 285, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{r} \\mathcal {L} _ {\\mathrm {s e m a n t i c}} = C E (s e m _ {\\mathrm {o b j}}, S G (s e m _ {\\mathrm {p a r t}})) + \\\\ C E (s e m _ {\\mathrm {p a r t}}, S G (s e m _ {\\mathrm {o b j}})), \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 556, + 434, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $SG$ is the stop-gradient operator.", + "bbox": [ + 89, + 604, + 354, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Textual Consistency The object- and part-level captions are expected to slightly overlap, since the salient features of an object are reflected in its object-level caption. We encourage this similarity through the similarity of the whole sequences of object-level and part-level captions. Similar to the semantic consistency loss, we use projectors $\\Phi_{\\mathrm{obj},\\mathrm{text}}$ and $\\Phi_{\\mathrm{part},\\mathrm{text}}$ to project the language model hidden states into $h_{\\mathrm{obj},\\mathrm{text}}$ and $h_{\\mathrm{part},\\mathrm{text}}$ , and aggregate over the sequence of tokens to get single feature vectors $\\overline{h}_{\\mathrm{obj},\\mathrm{text}}$ and $\\overline{h}_{\\mathrm{obj},\\mathrm{part}}$ . The text-consistency loss is the distance $d(.,.)$ between these two feature vectors", + "bbox": [ + 89, + 632, + 483, + 797 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {t e x t u a l}} = d (\\overline {{h}} _ {\\mathrm {o b j , t e x t}}, \\overline {{h}} _ {\\mathrm {p a r t , t e x t}}).\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 814, + 400, + 833 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss used to optimize ExCap3D is a weighted sum of the individual losses", + "bbox": [ + 89, + 840, + 482, + 871 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = w _ {1} \\mathcal {L} _ {\\text {c a p t i o n}} + w _ {2} \\mathcal {L} _ {\\text {s e m a n t i c}} + w _ {3} \\mathcal {L} _ {\\text {t e x t u a l}}\n$$\n", + "text_format": "latex", + "bbox": [ + 133, + 885, + 437, + 901 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/06fc61d8856315f8a3e6feebb09466c0397b22eddf02d7ba7615639d65749500.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Dataset#Desc#Cls#ObjsLoD
NR3D [2]41k765.8ks + o
Multi3DRefer [56]55k26510.5ks + o
ScanQA [3]35k3709.5ks + o
Scan2Cap [15]46k2659.9ks + o
Ours190k2k34.7ko + p
", + "bbox": [ + 526, + 88, + 895, + 196 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1. Comparison of the ExCap3D Dataset with existing datasets that describe objects in 3D indoor scenes. We compare the number of descriptions (#Desc), number of unique object classes described (#Cls), number of objects described (#Objs) and level of detail (LoD, s = scene, o = object, p = part). Our descriptions have greater object detail and introduce part-level details, while scaling up the number of descriptions, and semantic classes and objects described.", + "bbox": [ + 511, + 205, + 905, + 316 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details Predicted instances are matched with GT instances using Hungarian matching [34], and the predicted captions are supervised by the captions of the matched instances. We train both the instance segmentation and captioning stages with a batch size of 6 scenes. In the captioning stage, one batch that is input to the captioner contains all objects from scenes in the batch.", + "bbox": [ + 511, + 332, + 905, + 439 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. ExCap3D Dataset", + "text_level": 1, + "bbox": [ + 511, + 452, + 689, + 470 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Existing datasets for captioning objects in 3D indoor scenes contain captions at a single level of detail, that combines both object properties such as color and texture, as well as spatial and functional relations to other objects in the scene. To enable the task of captioning at both the object- and part-levels of detail, we created the ExCap3D Dataset based on the ScanNet++ [52] dataset of 3D indoor scans. The data is generated in a scalable and fully automated manner by leveraging a VLM that takes an RGB image crop of an object and a text prompt as input, and outputs a description of the object. The objects to be captioned are identified using the ground truth semantic instance annotations provided by ScanNet++. We prompt the VLM with text prompts to generate specific information about an object's color, texture, material and similar properties, and leverage multiview reasoning to obtain reliable final caption generation for ScanNet++. A comparison of the ExCap3D Dataset with existing object_caption and other 3D object-text datasets is shown in Tab. 1.", + "bbox": [ + 511, + 478, + 906, + 765 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Object Level of Detail Captioning", + "text_level": 1, + "bbox": [ + 511, + 772, + 808, + 789 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We utilize the ground truth 3D instance and semantic annotations in the ScanNet++ dataset to generate the ExCap3D Dataset. We project the 3D annotations onto the high-resolution DSLR images that are registered with the 3D scan, to obtain a corresponding 2D bounding box of each object in each DSLR image. We use this 2D mask to crop a rectangular region containing the object from the image,", + "bbox": [ + 511, + 794, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21703", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a2b3e4384a1a5d9a5f77f6e0664af3c55f80bba9c00984515ea93d97b166a31c.jpg", + "image_caption": [ + "Figure 4. Generation of ExCap3D Dataset. We use the ground truth 3D semantics in ScanNet++, project them onto multiview DSLR images and obtain descriptions of the image crops $c_{i,j}$ using a VLM. For parts crops, we use pseudo-ground truth from MaskClustering. Finally, we summarize the captions from different views using an LLM." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 480, + 181 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "after expanding the bounding box dimensions by $10\\%$ . To maximize the RGB information that is input to the VLM, we pick the top 3 images that cover at least $10\\%$ the mesh vertices corresponding to the object mask. The VLM generates $k = 5$ descriptions for each view of the object, giving outputs $c_{i,j}$ from the $i^{\\mathrm{th}}$ view. We use an instruction-tuned large language model (LLM) to obtain the $j^{th}$ object-level caption $\\hat{c}_{\\mathrm{obj},j}$ by combining the $j^{th}$ caption from each view.", + "bbox": [ + 89, + 297, + 483, + 420 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Part Level of Detail Captioning", + "text_level": 1, + "bbox": [ + 89, + 426, + 369, + 443 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We produce pseudo-ground truth 3D part segmentation data using a modified version of MaskClustering [51] that favours smaller segments over larger ones, along with SAM [33]. SAM is used to generate initial 2D segments on the DSLR images. MaskClustering combines these 2D segments to produce coherent 3D part segments. We associate these 3D parts with the objects in ScanNet++ based on an intersection-over-union (IoU) threshold. The rest of the part captioning process largely follows that of object captioning. We generate one caption $p_i$ in each view, for each part, and then combine the descriptions of all the parts into a single part-level caption $\\hat{c}_{\\mathrm{part}}$ for each object using the LLM. While the pseudo-masks at the part-level obtained from MaskClustering can detect parts of long-tail and small objects in the ScanNet++ data, these masks are not as reliable as the manually annotated object GT masks in ScanNet++. Hence, we aggregate the captions of different parts into a single caption containing overall part-level details for an object.", + "bbox": [ + 89, + 449, + 483, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details We use Llava 1.6 [37] (7B parameters) with a Mistral backbone as the VLM, and instruction-tuned Llama 3.1 [21] (8B parameters) for the LLM. To ensure the quality of the captions for evaluation, we manually removed incorrect captions in the test set. Due to the robust multiview aggregation of captions, $93.21\\%$ of the captions were found to be accurate.", + "bbox": [ + 89, + 734, + 483, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 845, + 223, + 862 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate caption generation on the 84 instance class categories of the ScanNet++ benchmark [52]. We use the", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "train/val/test splits of the semantic tasks of ScanNet++.", + "bbox": [ + 511, + 90, + 877, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Evaluation Metrics", + "text_level": 1, + "bbox": [ + 511, + 117, + 697, + 132 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We adopt the commonly used CIDEr [44], ROUGE [36], METEOR [20] and BLUE-4 [41] caption evaluation metrics used by other 3D captioning methods [10, 15, 27, 58, 60]. Metrics are reported separately for the object- and part-level details.", + "bbox": [ + 511, + 140, + 905, + 215 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Baselines", + "text_level": 1, + "bbox": [ + 511, + 228, + 622, + 242 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method against six state-of-the-art captioning methods. Cap3D [38] captions individual objects using pretrained BLIP2 [35] and GPT-4 [1]. LLaVA-3D [57] aggregates multiview 2D features and uses these as input to an LLM, and is evaluated zero-shot from a pretrained model. D3Net [9] uses a PointGroup sparse convolutional backbone with embedded object relations, and a GRU for text generation. Vote2Cap-DETR [10] tokenizes the 3D input, and generates captions using GPT2 [42]. Vote2Cap-DETR++ [12] additionally uses decoupled detection and caption queries. PQ3D [60] uses promptable queries for regions of interest, and generates captions with T5-Small [43].", + "bbox": [ + 511, + 251, + 906, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train on ScanNet++ [52] for bounding box detection (Vote2Cap and D3Net) or instance segmentation (PQ3D and ExCap3D). Since the baselines were designed for single level of detail captioning, we train them separately on the object- and part-level captions, while ExCap3D is trained jointly on both levels of detail.", + "bbox": [ + 511, + 448, + 905, + 539 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5140244823f4f24060d36679e14805b2c9ba6fd5558414533c669021dea6f29e.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodObject-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]5.57.67.70.0
D3Net [9]6.75.46.70.4
LLaVA-3D [57]0.06.016.30.0
Vote2Cap-DETR [10]13.312.917.21.0
Vote2Cap-DETR++ [12]10.611.915.71.1
PQ3D [60]27.911.612.61.9
ExCap3D (Ours)32.916.718.01.2
", + "bbox": [ + 514, + 551, + 901, + 676 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5598622341d9cccaacbd2b61aeb5e4dd79135a357e3e87ae041fe8674d754b0d.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodPart-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]2.111.49.80.0
D3Net [9]7.77.67.31.3
LLaVA-3D [57]0.613.015.50.0
Vote2Cap-DETR [10]7.720.321.73.9
Vote2Cap-DETR++ [12]19.519.919.83.8
PQ3D [60]14.516.015.34.3
ExCap3D (Ours)32.821.921.14.8
", + "bbox": [ + 516, + 679, + 901, + 800 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparison of object-level and part-level captioning performance between ExCap3D and baselines. The second-best result is underlined. ExCap3D improves CIDEr scores significantly through joint captioning and consistency losses. All metrics are evaluated at an IoU threshold of 0.5", + "bbox": [ + 511, + 810, + 905, + 878 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21704", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9a79e919983cbbb110090eddc8107076f497abe89e22bcc416a097b6c7ee5129.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Input 3D Scene\nDetected 3D Objectwhite, rectangular, smooth, glossy, unk*white, rectangular, smooth, glossy, unk*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with white, smooth, glossy surface, made of wood or laminate, metal*
D3Net\n(Chen et al. 21)Object-levelwhite, smooth, possibly plastic or metal*a white, smooth, possibly plastic or metal*a black, curved backrest with a smooth texture, attached to a black metal frame with a smooth texture.a white, smooth, possibly plastic or metal*
Part-level
Vote2Cap-DETR\n(Chen et al. 23)Object-levelBlack, rectangular, smooth, glossy, plastic or metal computer tower. Black, smooth, plastic or metal, rectangular, cylindrical, or cylWhite, rectangular cabinet with a smooth, glossy finish, made of wood or laminate,like composite, flat top, and a handle. Dark woodA modern chair with a curved backrest, smooth glossy finish, dark fabric, and a chrome metal frame.. Modern, vibrant red, smooth,Dark-colored, leather-like, smooth, glossy, curved, L-shaped sofa. L-shaped sofa with a sturdy frame. Black,
Part-levelA black, rectangular trash can with a smooth, possibly plastic or metal surface, a black, cylindrical trash can with a smooth, possibly plastic surface,*A rectangular file cabinet with a flat top and bottom, made of light-colored wood with a smooth finish and visible grain pattern. It has a flat top surface and a slightly curved top edge. A rectangular drawer with a flat top and bottomThe chair has a curved, dark-colored backrest with a smooth, glossy finish, a curved, dark-colored backrest with a smooth texture, and a metal frame. The backrest is a curved, smooth, glossy piece of darkThe sofa has a curved, padded backrest with a smooth, plush texture, a curved, padded armrest with a smooth, glossy finish.* The backrest is a curved
PQ3D\n(Zhu et al. 24)Object-levelblack, rectangular, smooth, glossy, plastic or metal*white, rectangular, smooth, glossy, or metal.*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with a smooth, glossy, glossy surface, and a light-colored, possibly wood or wood-like material
Part-levela black, rectangular computer tower with a flat top and bottom, curved top, and a flat top, made of a smooth, possibly plastic or metal materialthe cabinet has a light-colored, smooth finish, possibly wood or composite, with a flat top and bottom*, a slightly curved top edge, and a slightly angled frontthe chair has a dark-colored, smooth, possibly plastic or metal backrest with a curved shape and a uniform texture.the sofa has a curved, cushioned backrest with a smooth, glossy finish, a curved, padded armrest with a smooth, glossy finish*
ExCap3D\n(Ours)Object-levelBlack, glossy, rectangular computer tower with a smooth textureWooden cabinet with light brown color, smooth texture, and rectangular shape.Wooden chair with light brown color, smooth texture, and rectangular shape.Black leather sofa with smooth texture and rectangular shape with rounded corners.
Part-levelA black computer tower with a textured surface, vents, and a rectangular drive bay with a textured surface.The cabinet has a light-colored, smooth finish, possibly made of wood or a wood-like composite, with a flat top and bottom, a series of drawers with metal handles.The chair has a light brown, smooth, curved backrest made of wood or wood-like composite, attached to a metal frame.The sofa has a black leather backrest with a smooth, leather-like texture, a curved armrest with a smooth, possibly leather-like material, and a rectangular cushion with a smooth texture.
Ground TruthObject-levelBlack, rectangular computer tower with a flat top, smooth, polished surface, and a tapered base.Light-colored, rectangular, boxy cabinet with smooth, glossy, synthetic surface.A chair with a curved backrest, flat seat, smooth, light brown, wooden surface, and metal legs.Dark-colored leather sofa with curved shape, plush cushions, and smooth texture.
Part-levelA black computer tower with a black power supply unit attached, both with smooth, glossy finishes.The cabinet has a light brown, smooth wooden surface, glass front, three wooden drawers with silver, curved handles, and a dark-colored, rectangular door.A curved, light-colored wood or wood-like composite backrest with smooth or matte finish, supported by a metal frame.The sofa has a curved, padded armrest, a cushion, and a backrest, all upholstered in dark, rich brown or black leather with a smooth, shiny texture.
", + "bbox": [ + 109, + 87, + 885, + 844 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Figure 5. Qualitative evaluation on ScanNet++ [52] scenes, in comparison with D3Net [9], Vote2CAP-DETR [10], and PQ3D [60]. Predicted text from methods are denoted by color to indicate correct and incorrect generated phrases compared to the ground truth. Our method produces more consistent and detailed captions at both the object- and part-levels of detail. * indicates truncated output where the underlined phrase was predicted repeatedly.", + "bbox": [ + 89, + 859, + 906, + 916 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21705", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Comparison to State of the Art", + "text_level": 1, + "bbox": [ + 89, + 90, + 364, + 106 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We compare with state-of-the-art captioning methods in Tab. 2. ExCap3D outperforms all baselines on the CIDEr score, which aligns the most closely with human perception, improving upon the highest performing baseline PQ3D by a margin of $17\\%$ and $124\\%$ for object- and part-level details respectively. We also outperform the baselines on the ROUGE and METEOR scores of object-level details, and ROUGE scores of part-level details. Our joint prediction of captions at different levels of detail enables more robust, accurate descriptions.", + "bbox": [ + 88, + 112, + 480, + 263 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Qualitative comparisons are shown in Fig. 5. ExCap3D produces complete and detailed descriptions at both the object- and part-levels of detail, while the baseline methods often omit details in both levels. Cap3D's BLIP-2 captioner can produce limited details, and is easily distracted by background objects. LLaVA-3D produces general captions with few object-specific details, and has limited object localization due to a single location prompt as input. D3Net uses a limited GLOVE embedding of text which cannot produce detailed output descriptions. Vote2Cap-DETR, Vote2Cap-DETR++ and PQ3D interpret 3D scenes more coarsely, which can result in hallucination of incorrect fine-grained details.", + "bbox": [ + 91, + 265, + 483, + 460 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.4. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 472, + 254, + 487 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We ablate the different components of our method and show the results in Tab. 3. As a baseline, we use two independent captioning heads and train with only $\\mathcal{L}_{\\mathrm{caption}}$ . Each component of our method is added separately to the baseline to show its efficacy.", + "bbox": [ + 89, + 494, + 482, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Enforcing multilevel consistency improves caption quality. Semantic consistency primarily improves part-level details while textual consistency primarily improves object-level details as shown in Tab. 3, which demonstrates the complementary nature of these two losses. As the object-level captions typically contain a stronger notion of global object semantics compared to the part-level caption, the latter improves in quality with semantic consistency. The object-level captions also receive fine-grained cues through textual similarity from the part-level details which helps to refine the global description of the object.", + "bbox": [ + 89, + 583, + 482, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Part-object captioner information sharing improves performance Adding part $\\rightarrow$ object information sharing by conditioning object captions on the part-level of detail improves both levels of captions as shown in Tab. 3. This is because object details are improved when they are informed by part captions which contain aggregated details that can produce a better global description of the object. At the same time, the part-level of detail improves as it is guided to produce consistent descriptions of parts that can be useful for object-level captioning.", + "bbox": [ + 89, + 761, + 482, + 912 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/793f587b2326b2508e2857fc7a65c5df7954faa12b5afa3245041484fc1af2c8.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodObject-LevelPart-level
Separate Models29.717.9
w/ Semantic Consistency30.223.4
w/ Textual Consistency32.218.6
w/ Part → Obj. Inf. Sharing34.824.0
Full Model (Ours)32.932.8
", + "bbox": [ + 514, + 88, + 908, + 196 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3. Ablation study of the different components of our method. Performance is reported using CIDEr for object-level and part-level captions. Each ablation is applied independently to the baseline with separate models.", + "bbox": [ + 511, + 205, + 903, + 262 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Consistency and information sharing are complementary to each other The semantic and textual consistency losses act independently of the part $\\rightarrow$ object information sharing mechanism, and hence our full model improves significantly over the individual components, particularly in the part-level of detail as shown in Tab. 3. Intuitively, this is because the consistency losses model only two kinds of consistency over lower-dimension embeddings, while sharing language model hidden states enables the object-level captioner to directly condition on the internal high-dimension representation of the part-level captions with fine-grained features.", + "bbox": [ + 511, + 287, + 906, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations While our approach generates expressive captioning for 3D scans, various limitations remain. We use two captioning models, with one conditioned on the hidden states of the other through cross-attention. This could limit the amount of information shared between the captioners. Our method relies on a sparse convolutional backbone [17] which is limited to a resolution of about $2\\mathrm{cm}$ . This could restrict captioning performance on small or thin objects. Similarly, the class imbalance in the ScanNet++ dataset could restrict performance on long-tail classes.", + "bbox": [ + 511, + 476, + 908, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 640, + 633, + 656 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We introduced ExCap3D, a method for expressive dense captioning of 3D scenes at multiple object- and part-levels of detail. We showed that describing 3D objects as the sum of their parts while maintaining consistency between the multiple levels of descriptions leads to more accurate descriptions aligned with human perception. We hope that our method and the accompanying ExCap3D Dataset open up new possibilities for a holistic, expressive understanding of 3D scans.", + "bbox": [ + 511, + 665, + 906, + 801 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 815, + 671, + 830 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This project was supported by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding.\"", + "bbox": [ + 511, + 839, + 903, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21706", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6", + "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 5", + "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 5", + "[4] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 3", + "[5] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3", + "[6] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3", + "[7] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16464-16473, 2022. 3", + "[8] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2", + "[9] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In European Conference on Computer Vision, pages 487-505. Springer, 2022. 3, 6, 7", + "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11124-11133, 2023. 2, 3, 6, 7", + "[11] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF" + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 3", + "[12] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7331-7347, 2024. 6", + "[13] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3", + "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 5", + "[15] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 2, 3, 5, 6", + "[16] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 18109-18119, 2023. 3", + "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 8", + "[18] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3, 1", + "[19] Xueqing Deng, Qihang Yu, Ali Athar, Chenglin Yang, Linjie Yang, Xiaojie Jin, Xiaohui Shen, and Liang-Chieh Chen. Coconut-pancap: Joint panoptic segmentation and grounded captions for fine-grained understanding and generation. arXiv preprint arXiv:2502.02589, 2025. 3", + "[20] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014. 6, 3", + "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6", + "[22] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. 2024. 3", + "[23] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4, 1, 3" + ], + "bbox": [ + 517, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21707", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[24] Yunhao Ge, Xiaohui Zeng, Jacob Samuel Huffman, Tsung-Yi Lin, Ming-Yu Liu, and Yin Cui. Visual fact checker: enabling high-fidelity detailed caption generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14033-14042, 2024. 3", + "[25] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3", + "[26] Hang Hua, Qing Liu, Lingzhi Zhang, Jing Shi, Zhifei Zhang, Yilin Wang, Jianming Zhang, and Jiebo Luo. Finecaption: Compositional image captioning focusing on wherever you want at any granularity. arXiv preprint arXiv:2411.15411, 2024.3", + "[27] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6", + "[28] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In European Conference on Computer Vision, pages 528-545. Springer, 2022. 3", + "[29] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2", + "[30] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3", + "[31] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 3", + "[32] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 2", + "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 6, 3", + "[34] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5", + "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 6", + "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + ], + "bbox": [ + 91, + 90, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.6", + "[38] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36: 75307-75337, 2023. 3, 6", + "[39] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In European Conference on Computer Vision, pages 180-197. Springer, 2024. 3", + "[40] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 3", + "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6, 3", + "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6", + "[43] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6", + "[44] Gabriel Oliveira dos Santos, Esther Luna Colombini, and Sandra Avila. Cider-r: Robust consensus-based image description evaluation. arXiv preprint arXiv:2109.13701, 2021. 6, 3", + "[45] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3, 4", + "[46] Ayca Takmaz, Alexandros Delitzas, Robert W Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3d: Hierarchical open-vocabulary 3d segmentation. IEEE Robotics and Automation Letters, 2025. 3", + "[47] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2", + "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 3", + "[49] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 3", + "[50] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 3", + "[51] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for" + ], + "bbox": [ + 516, + 90, + 906, + 902 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21708", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 6, 3", + "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 3, 5, 6, 7", + "[53] Zuyao You, Junke Wang, Lingyu Kong, Bo He, and Zuxuan Wu. Pix2cap-coco: Advancing visual comprehension via pixel-level captioning. arXiv preprint arXiv:2501.13893, 2025. 3", + "[54] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3", + "[55] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15459-15469, 2024. 2", + "[56] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 5", + "[57] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 3, 6", + "[58] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 3, 6", + "[59] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2", + "[60] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188–206. Springer, 2024. 2, 3, 6, 7" + ], + "bbox": [ + 91, + 90, + 482, + 726 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21709", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_model.json b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_model.json new file mode 100644 index 0000000000000000000000000000000000000000..ce0f3329953100f329afb77f15d257177f9cd268 --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.148, + 0.13, + 0.851, + 0.177 + ], + "angle": 0, + "content": "ExCap3D: Expressive 3D Scene Understanding via Object Captioning with Varying Detail" + }, + { + "type": "text", + "bbox": [ + 0.248, + 0.204, + 0.414, + 0.221 + ], + "angle": 0, + "content": "Chandan Yeshwanth" + }, + { + "type": "text", + "bbox": [ + 0.447, + 0.204, + 0.611, + 0.221 + ], + "angle": 0, + "content": "Dávid Rozenberszki" + }, + { + "type": "text", + "bbox": [ + 0.651, + 0.204, + 0.745, + 0.222 + ], + "angle": 0, + "content": "Angela Dai" + }, + { + "type": "text", + "bbox": [ + 0.37, + 0.229, + 0.623, + 0.246 + ], + "angle": 0, + "content": "Technical University of Munich" + }, + { + "type": "text", + "bbox": [ + 0.312, + 0.255, + 0.68, + 0.272 + ], + "angle": 0, + "content": "https://cy94.github.io/excap3d/" + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.31, + 0.905, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.48, + 0.908, + 0.508 + ], + "angle": 0, + "content": "Figure 1. ExCap3D produces descriptions at multiple level of details for objects in an input 3D scan. For each detected object, we generate both object- and part-level of detail captions. The generated captions are consistent across the two levels and contain rich details." + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.521, + 0.327, + 0.536 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Generating text descriptions of objects in 3D indoor scenes is an important building block of embodied understanding. Existing methods describe objects at a single level of detail and do not capture fine-grained details of the parts of objects. In order to produce varying levels of detail capturing both coarse object-level information and detailed part-level descriptions, we propose the task of expressive 3D captioning. Given an input 3D scene, the task is to describe objects at multiple levels of detail: a high-level object description, and a low-level description of the properties of its parts. To produce such captions, we present ExCap3D, an expressive 3D captioning model which takes as input a 3D scan, and for each detected object in the scan, generates a fine-grained collective description of the parts of the object, along with an object-level description conditioned on the part-level description. We design ExCap3D to encourage consistency between the multiple levels of descriptions. To enable this task, we generated the ExCap3D Dataset by leveraging a visual-language model (VLM) for multi-view captioning. The ExCap3D Dataset contains captions on the ScanNet++ dataset with varying levels of detail, comprising 190k text descriptions of 34k 3D objects in 947 indoor scenes. Our experiments show that the object- and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.909, + 0.598 + ], + "angle": 0, + "content": "part-level details generated by ExCap3D are more expressive than those produced by state-of-the-art methods, with a CIDEr score improvement of \\(17\\%\\) and \\(124\\%\\) for object- and part-level details respectively. Our code, dataset and models will be made publicly available." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.612, + 0.645, + 0.627 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.909, + 0.903 + ], + "angle": 0, + "content": "3D indoor scene understanding is central to applications such as augmented reality (AR), virtual reality (VR), and robotics. While detecting 3D objects and their semantics is important for these applications, natural language descriptions of objects enable more wide-ranging possibilities – they can encode more complex information efficiently and, importantly, enable more natural interactions between embodied agents as well as humans and the scene. To produce a representation that facilities a diverse array of applications, such language descriptions must also encompass varying levels of detail: object-level descriptions for coarse identification in the scene, and part-level details for precise manipulation and fine-grained understanding. For instance, a robotic assistant would only need a coarse, high-level description of the objects in a room (e.g., 'a recliner chair') in order to navigate effectively in it, but it would need fine-grained descriptions of the objects (e.g., 'recliner chair with a high padded back, wooden armrests, a soft cushion and an" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21699" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "adjustable footrest') to provide assistive AI descriptions for accessibility aid or to describe the scene to others." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.125, + 0.484, + 0.473 + ], + "angle": 0, + "content": "Current 3D captioning models have made significant progress in describing objects in indoor scenes by detecting objects and generating their individual descriptions at a single level of detail. These methods are trained only for captioning [10, 15, 32] or for multiple tasks [11, 55, 59, 60]. These methods can effectively describe coarse object information in the global context of the scene, but preclude detailed perception of the scene as they make limited use of fine-grained 3D inputs and cannot model object-part hierarchy. They fail to describe highly localized properties of object parts, such as appearance, material, and function. In order to enable more complex perception of a scene requiring these details, we propose to make use of fine-grained 3D inputs and explicitly model object-part hierarchy. We describe objects at multiple levels: an object-level of detail that captures its most prominent aspects (semantic class; possibly size or overall appearance), and a finer-grained part-level of detail that describes the details of elements composing the object (different materials, color, and functions). These two descriptions are complementary to each other and enable different applications depending on the level of detail required. Examples of such multilevel descriptions are shown in Fig. 1 and Fig. 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.476, + 0.483, + 0.747 + ], + "angle": 0, + "content": "To generate such multilevel captions in 3D scans, we introduce ExCap3D, a 3D object captioning model that produces expressive captions at both object- and part-levels of detail using two captioning heads. Examples of the captions generated by our method are shown in Fig. 1. The generated multilevel descriptions are complementary to each other and describe different aspects of the objects, while maintaining a few shared elements for consistency and coherence. The part-level caption generated by ExCap3D collectively describes the different parts of an object. Knowing the descriptions of the parts can enable our model to generate more consistent captions describing the object as a whole. Hence, ExCap3D first produces the part-level details and then uses these to inform the object-level captioner. To further reduce inconsistencies between the two levels of detail in terms of the object semantics and text content, ExCap3D employs additional consistency losses during training." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To enable this multilevel captioning task, we provide the ExCap3D Dataset, which contains 190k object and partlevel natural language captions of 34k 3D objects from the 947 scenes of the ScanNet++ dataset. We automatically generate object captions for training, using a pretrained vision-language model (VLM) on high-resolution DSLR images along with ground truth semantic information from the ScanNet++ dataset [52]. We employ multi-view aggregation to robustly capture key object- and part-level details. VLMs perform strongly on 2D tasks, but" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.091, + 0.568, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.582, + 0.102, + 0.631, + 0.14 + ], + "angle": 0, + "content": "a brown cushion chair with arms. it is located close to the wall." + }, + { + "type": "image", + "bbox": [ + 0.647, + 0.091, + 0.775, + 0.141 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.101, + 0.882, + 0.141 + ], + "angle": 0, + "content": "Q: What is next to the orange armchair? \nA: wall \nA: another orange armchair \nA: coffee table" + }, + { + "type": "image_caption", + "bbox": [ + 0.565, + 0.145, + 0.6, + 0.153 + ], + "angle": 0, + "content": "ExCap3D" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.154, + 0.645, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.655, + 0.154, + 0.887, + 0.202 + ], + "angle": 0, + "content": "Object-level: A three-seater leather couch with dark, smooth, glossy leather, curved and straight edges. \nPart-Level: The black leather sofa has a smooth, glossy surface with visible stitching, a curved armrest, and cushioned seat, with a tufted backrest made of multiple leather panels stitched together, and sturdy structure with visible stitching along the edges." + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.217, + 0.907, + 0.287 + ], + "angle": 0, + "content": "Figure 2. Descriptions of objects in existing datasets such as Scan2Cap and ScanQA largely contain relations between objects, and limited local details at a single level. In contrast, the ExCap3D Dataset describes both the object as a whole and as a sum of its parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.313, + 0.905, + 0.463 + ], + "angle": 0, + "content": "struggle with 3D orientation [47] and distances [8]. Hence, our dataset contains large-scale and high-quality object- and part-level captions, in contrast to Scan2Cap [15] and Nr3D [2] with smaller-scale manually annotated spatial relations. A comparison of the multilevel captions from the ExCap3D Dataset with the single-level descriptions from the existing Scan2Cap and ScanQA [3] is shown in Fig. 2. The ExCap3D Dataset describes objects with significantly more detail, while providing two complementary multilevel descriptions of each object." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.905, + 0.538 + ], + "angle": 0, + "content": "We show that ExCap3D produces richer and more consistent object- and part-level captions compared to the state-of-the-art captioning methods, originally designed to generate only a single level description. To summarize, our main contributions are:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.54, + 0.905, + 0.598 + ], + "angle": 0, + "content": "- an expressive 3D captioning model that generates captions of objects in 3D indoor scenes at multiple levels of detail: fine-grained part-level details and object-level details" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.6, + 0.905, + 0.674 + ], + "angle": 0, + "content": "- modeling object-level detail as a sum of its parts' unique appearance and functional attributes by generating object-level captions conditioned on part-level descriptions, using the cross-attention mechanism of the object captioner and the language model's intermediate hidden states" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.675, + 0.905, + 0.735 + ], + "angle": 0, + "content": "- regularization of our joint 3D captioning approach through semantic and textual-similarity consistency losses that maintain consistency of the object-level and part-level predicted captions" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.736, + 0.905, + 0.811 + ], + "angle": 0, + "content": "- the ExCap3D Dataset: a large-scale, precise and challenging dataset of 190k descriptive object- and part-level captions of 34k objects in 3D indoor scenes built on the ScanNet++ dataset with fine-grained texture, material, and functional descriptions" + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.54, + 0.905, + 0.811 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.824, + 0.655, + 0.839 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.715, + 0.866 + ], + "angle": 0, + "content": "2.1. 2D Image Captioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.905, + 0.902 + ], + "angle": 0, + "content": "2D image captioning has been extensively studied in computer vision. DenseCap [29] popularized the task of dense" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21700" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.228 + ], + "angle": 0, + "content": "captioning of images, jointly localizing objects, and generating their short descriptions. More recently, COCONutPanCap [19] and Pix2Cap-COCO [53] combined the tasks of fine-grained instance segmentation and detailed captioning. FlexCap [22] and FineCaption [26] can produce descriptions at varying levels of detail given the image and a localized object as input. In contrast, we combine the tasks of detection and multilevel captioning and describe detected objects and their parts." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.256, + 0.289, + 0.273 + ], + "angle": 0, + "content": "2.2. 3D Scene Captioning" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.285, + 0.485, + 0.648 + ], + "angle": 0, + "content": "Scan2Cap [15] introduced the task of 3D dense captioning and a method for context-aware captioning of 3D objects, operating on ScanNet [18] scenes. Since then, several methods have been developed to effectively solve this task by extracting object features and modeling their relations with nearby objects [7, 9, 10, 28, 31, 49, 54]. These methods can accurately describe objects at the scene-level through relations with neighboring objects. Following the trends of utilizing foundation models in 3D reasoning, Cap3D [38] and follow-up works [24, 39] describe isolated 3D objects through multi-view aggregation of VLM-generated captions. More recent works have jointly trained on multiple tasks such as 3D grounding, question answering, and dense captioning by training custom language models from scratch [16, 58] or by fine-tuning pretrained LLMs [11, 25, 27, 50, 57, 60]. However, these works focus on high-level, proximity-based relations between objects at the scene-level, underutilize local features to produce limited object-level details and no part-level details We tackle a similarly inspired problem, and leverage the high-fidelity geometry and RGB captures from ScanNet++ [52] to enable complex, multilevel joint captioning. Moreover, we demonstrate that our joint captioning approach improves both levels of detail." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.676, + 0.466, + 0.693 + ], + "angle": 0, + "content": "2.3. Part Understanding of 3D Objects in Scenes" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Several methods have focused on part-level understanding of 3D scenes by only relying on explicit geometry [5, 6], images and radiance fields [4, 30] or by using both modalities and mesh-based oversegmentation [46]. A related method for 3D shapes rather than scenes was proposed by Chen et. al. [13], which produces part-level segmentation maps and text descriptions of individual 3D objects at a single, fine-grained level of detail. This produces promising detailed captions, but is designed for isolated objects rather than complex 3D scenes. In contrast, we jointly detect and caption objects in large 3D scenes at multiple levels of detail, ensuring consistency and enabling greater expressivity for downstream tasks." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.605, + 0.107 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.115, + 0.907, + 0.268 + ], + "angle": 0, + "content": "Problem Formulation The task of expressive captioning is to detect the objects in a 3D scene and describe each object at different levels of detail: object- and part-levels. Given an input 3D scene in the form of a 3D mesh \\( M = (\\mathcal{V},\\mathcal{F}) \\) with vertices \\( \\mathcal{V}\\in \\mathbb{R}^{N_v\\times 3} \\) and triangular faces \\( f_{i}\\in \\mathcal{F} \\) connecting the vertices, we aim to detect the 3D object instances \\( o_i = (b_i,s_i,c_{\\mathrm{obj},i},c_{\\mathrm{part},i}) \\) in the scene, characterized by the objects' bounding boxes \\( b_{i} \\), semantic class \\( s_i \\), and \\( c_{\\mathrm{obj},i} \\) and \\( c_{\\mathrm{part},i} \\) describing the object in natural language at object- and part-levels of detail, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.284, + 0.907, + 0.436 + ], + "angle": 0, + "content": "Overview of ExCap3D First, ExCap3D performs instance segmentation and instance-aware query generation to produce queries containing rich per-object information. Next, the object- and part-level captioning models with shared information generate two separate text captions, conditioned on individual instance-aware queries and the corresponding features from the instance segmentation backbone. Finally, semantic and whole-text consistency losses are applied to the hidden states of the captioning models to regularize the consistency of the predicted captions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.442, + 0.753, + 0.459 + ], + "angle": 0, + "content": "3.1. 3D Instance Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.464, + 0.909, + 0.63 + ], + "angle": 0, + "content": "Given the mesh of the 3D scene \\(M\\) as input, the instance segmentation model produces a fixed number of refined instance-aware queries \\(Q_{r}\\) corresponding to the \\(N_{q}\\) objects detected in the scene, and their binary instance masks on the voxelized 3D input \\(I\\in \\{0,1\\}^{N_q\\times N_{\\mathrm{vox}}}\\). The instance queries are a compact representation of the 3D features of the objects as well as their semantics. The instance segmentation model also produces dense features \\(F\\in \\mathbb{R}^{N_{\\mathrm{vox}}\\times D}\\) on the voxels. The refined and instance-aware queries \\(Q_{r}\\) and 3D voxel features \\(F\\) are used to inform the per-instance captioning in the next step." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.631, + 0.909, + 0.842 + ], + "angle": 0, + "content": "The input mesh is first voxelized into a voxel grid \\(V \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times 6}\\) containing the voxel coordinates and RGB colors. We use the Mask3D [45] transformer-based model for instance segmentation. It consists of a 3D sparse convolutional UNet backbone to encode the voxelized 3D scene, which outputs features \\(F \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times D}\\) over the input voxels. A parallel mask module takes as input query vectors \\(Q \\in \\mathbb{R}^{N_q \\times D_q}\\) denoting the \\(N_q\\) instances in the scene. The queries are refined iteratively through several transformer encoder layers, with cross-attention to the 3D backbone features. The output instance-aware query vectors contain rich representations of the object instances in the 3D scene and are used to predict the instance masks \\(I \\in \\{0,1\\}^{N_q \\times N_{\\mathrm{vox}}}\\) and semantic class of each object." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.882, + 0.866 + ], + "angle": 0, + "content": "3.2. Joint Captioning with Information Sharing" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "ExCap3D contains two separate captioners for the object- and part-levels of details. These are transformer-based lan" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21701" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.868, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.316, + 0.908, + 0.361 + ], + "angle": 0, + "content": "Figure 3. Overview of our ExCap3D captioning method. We predict 3D instances in the input 3D scene using Mask3D [45], then predict object-level and part-levels of detail using two separate captioning heads. The object-level captions are further constrained by the part-level captioner's hidden states. Semantic- and textual consistency losses are applied to ensure the overall consistency of both predicted captions." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.385, + 0.484, + 0.492 + ], + "angle": 0, + "content": "guage models \\(\\Psi_{\\mathrm{obj}}\\) and \\(\\Psi_{\\mathrm{part}}\\) that produce the object-level caption \\(c_{\\mathrm{obj},o}\\), part-level caption \\(c_{\\mathrm{part},o}\\). They are conditioned on the instance-aware queries \\(Q_{r}\\) and the dense voxel features \\(F\\), and are trained for next-token prediction. The language model captioners take as input a sequence of previously predicted tokens \\(t_1 \\ldots t_{i-1}\\) and decode the next token \\(t_i\\) autoregressively as" + }, + { + "type": "equation", + "bbox": [ + 0.168, + 0.507, + 0.405, + 0.523 + ], + "angle": 0, + "content": "\\[\np \\left(t _ {i} \\mid t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(t _ {1} \\dots t _ {i - 1}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.484, + 0.639 + ], + "angle": 0, + "content": "where \\(\\Psi\\) is the language model. We provide two sources of rich instance-level information to the captioner: the instance-aware refined query output by the instance-segmentation model, and voxel features from 3D UNet backbone \\(F\\). The refined query vector \\(Q_{r,o}\\) corresponding to an object \\(o\\) is first projected with a linear layer to generate the caption-aware query \\(Q_{c,o} \\in \\mathbb{R}^{D_{\\mathrm{caption}}}\\) for that object" + }, + { + "type": "equation", + "bbox": [ + 0.212, + 0.65, + 0.363, + 0.667 + ], + "angle": 0, + "content": "\\[\nQ _ {c, o} = \\Phi_ {\\mathrm {q u e r y}} (Q _ {r, o}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.678, + 0.484, + 0.754 + ], + "angle": 0, + "content": "where \\( D_{\\mathrm{caption}} \\) is the embedding dimension of the language model. \\( Q_{c,o} \\) contains localized information required to generate the caption for an object, such as its semantic class and its high-level features, and is used as the initial token to start the caption generation as" + }, + { + "type": "equation", + "bbox": [ + 0.131, + 0.782, + 0.442, + 0.799 + ], + "angle": 0, + "content": "\\[\np \\left(t _ {i} \\mid Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Additionally, we inject fine-grained information by using the captioner's cross-attention layers to attend to fine-grained contextual features from the 3D backbone \\( F \\) corresponding to the object being captioned, after a linear projection to \\( D_{\\mathrm{caption}} \\). To reduce the computational complexity of the cross-attention operation, we aggregate features within" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.906, + 0.491 + ], + "angle": 0, + "content": "precomputed segments [23] on the mesh surface, giving the context features \\( S_{o} \\in \\mathbb{R}^{n_{s,o} \\times D_{\\mathrm{caption}}} \\), where \\( n_{s,o} \\) is the number of segments in \\( I_{o} \\), the predicted mask for object \\( o \\). The cross-attention operation is \\( A = \\operatorname{softmax}(qk^T / \\sqrt{d})v \\) where queries \\( q \\) and values \\( k \\) are the hidden states of \\( \\Psi \\), and the keys \\( k \\) are \\( S_{o} \\). The object and part captioners use context features \\( S_{\\mathrm{obj},o} \\) and \\( S_{\\mathrm{part},o} \\) respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.906, + 0.552 + ], + "angle": 0, + "content": "Importantly, to enable the object-level captioner to be informed by the part-level captioner, we make use of the last layer hidden states \\([h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},i}]\\) of the part-level captioner \\(\\Psi_{\\mathrm{part}}\\). We use a linear projection" + }, + { + "type": "equation", + "bbox": [ + 0.58, + 0.566, + 0.84, + 0.584 + ], + "angle": 0, + "content": "\\[\nH _ {\\text {p a r t}, o} = \\Phi_ {\\text {h i d d e n}} ([ h _ {\\text {p a r t}, 1} \\dots h _ {\\text {p a r t}, i} ])\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.59, + 0.906, + 0.695 + ], + "angle": 0, + "content": "and then concatenate \\( H_{\\mathrm{part},o} \\) to the contextual features of the object captioner \\( S_{\\mathrm{obj},o} \\) for object \\( o \\). The object-level caption for object \\( o \\) is finally informed by the caption-aware query \\( Q_{c,o} \\) as well as features \\( [H_{\\mathrm{part},o};S_{\\mathrm{obj},o}] \\). Since different objects in a batch have different numbers of predicted segments, we zero-pad \\( S_{\\mathrm{obj},o} \\) to the maximum context length in the batch." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.695, + 0.906, + 0.743 + ], + "angle": 0, + "content": "The two caption models are trained with cross-entropy (CE) loss against the object- and part-level ground truth (GT) captions respectively, giving the loss" + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.756, + 0.845, + 0.773 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c a p t i o n}} = \\mathcal {L} _ {\\text {o b j - c a p t i o n}} + \\mathcal {L} _ {\\text {p a r t - c a p t i o n}}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.779, + 0.906, + 0.843 + ], + "angle": 0, + "content": "The outputs of this stage are the object-level caption \\( c_{\\mathrm{obj},o} \\), part-level caption \\( c_{\\mathrm{part},o} \\) and the corresponding language model hidden states \\( h_{\\mathrm{obj}} = [h_{\\mathrm{obj},1}\\dots h_{\\mathrm{obj},i}] \\) and \\( h_{\\mathrm{part}} = [h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},j}] \\) for each object \\( o \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.868, + 0.866 + ], + "angle": 0, + "content": "3.3. Semantic and Textual Consistency Losses" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "While the object-level of detail is informed by the part-level of detail, supervising these two captioners with different GT" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21702" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "captions may lead to inconsistencies in the predicted captions due to the different GT distributions of the two levels of detail. Hence, we use additional constraints in the form of consistency losses over the predicted captions. We encourage semantic consistency for both captions to refer to the same object semantics, such as its fine-grained object class and material, and textual consistency for the captions to be similar to each other. Since consistency losses cannot be applied directly on the generated text due to the text tokenizer being non-differentiable, we instead use the language model hidden states \\( h_{\\mathrm{obj}} \\) and \\( h_{\\mathrm{part}} \\) from the last captioner layers as a proxy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.485, + 0.529 + ], + "angle": 0, + "content": "Semantic Consistency To promote both levels of captions to contain similar semantic information, we use a semantic consistency loss. We do this by projecting the hidden states \\( h_{\\mathrm{obj}} \\) and \\( h_{\\mathrm{part}} \\) into a lower dimension to retain only meaningful semantic information over the whole caption. We then pass these sequences independently through object- and part-level caption projection transformers \\( \\Phi_{\\mathrm{obj,sem}} \\) and \\( \\Phi_{\\mathrm{part,sem}} \\) to obtain new embedded sequences, and these aggregate into single feature vectors representing the overall semantics of the predicted captions \\( h_{\\mathrm{obj,sem}} \\in \\mathbb{R}^{D_{ff}} \\) and \\( h_{\\mathrm{part,sem}} \\in \\mathbb{R}^{D_{ff}} \\) respectively, where \\( D_{ff} \\) is the feed-forward dimension of \\( \\Phi \\). These are again independently projected into \\( N_{\\mathrm{sem}} \\) fine-grained latent classes using linear layers to get class probabilities \\( sem_{\\mathrm{obj}} \\) and \\( sem_{\\mathrm{part}} \\). The semantic consistency loss is a symmetric sum [14] of cross-entropy (CE) losses" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.557, + 0.435, + 0.594 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{r} \\mathcal {L} _ {\\mathrm {s e m a n t i c}} = C E (s e m _ {\\mathrm {o b j}}, S G (s e m _ {\\mathrm {p a r t}})) + \\\\ C E (s e m _ {\\mathrm {p a r t}}, S G (s e m _ {\\mathrm {o b j}})), \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.605, + 0.356, + 0.621 + ], + "angle": 0, + "content": "where \\( SG \\) is the stop-gradient operator." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.633, + 0.484, + 0.799 + ], + "angle": 0, + "content": "Textual Consistency The object- and part-level captions are expected to slightly overlap, since the salient features of an object are reflected in its object-level caption. We encourage this similarity through the similarity of the whole sequences of object-level and part-level captions. Similar to the semantic consistency loss, we use projectors \\(\\Phi_{\\mathrm{obj},\\mathrm{text}}\\) and \\(\\Phi_{\\mathrm{part},\\mathrm{text}}\\) to project the language model hidden states into \\(h_{\\mathrm{obj},\\mathrm{text}}\\) and \\(h_{\\mathrm{part},\\mathrm{text}}\\), and aggregate over the sequence of tokens to get single feature vectors \\(\\overline{h}_{\\mathrm{obj},\\mathrm{text}}\\) and \\(\\overline{h}_{\\mathrm{obj},\\mathrm{part}}\\). The text-consistency loss is the distance \\(d(.,.)\\) between these two feature vectors" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.815, + 0.401, + 0.834 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {t e x t u a l}} = d (\\overline {{h}} _ {\\mathrm {o b j , t e x t}}, \\overline {{h}} _ {\\mathrm {p a r t , t e x t}}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.842, + 0.483, + 0.872 + ], + "angle": 0, + "content": "The final loss used to optimize ExCap3D is a weighted sum of the individual losses" + }, + { + "type": "equation", + "bbox": [ + 0.135, + 0.886, + 0.438, + 0.902 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = w _ {1} \\mathcal {L} _ {\\text {c a p t i o n}} + w _ {2} \\mathcal {L} _ {\\text {s e m a n t i c}} + w _ {3} \\mathcal {L} _ {\\text {t e x t u a l}}\n\\]" + }, + { + "type": "table", + "bbox": [ + 0.527, + 0.089, + 0.896, + 0.198 + ], + "angle": 0, + "content": "
Dataset#Desc#Cls#ObjsLoD
NR3D [2]41k765.8ks + o
Multi3DRefer [56]55k26510.5ks + o
ScanQA [3]35k3709.5ks + o
Scan2Cap [15]46k2659.9ks + o
Ours190k2k34.7ko + p
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.207, + 0.906, + 0.318 + ], + "angle": 0, + "content": "Table 1. Comparison of the ExCap3D Dataset with existing datasets that describe objects in 3D indoor scenes. We compare the number of descriptions (#Desc), number of unique object classes described (#Cls), number of objects described (#Objs) and level of detail (LoD, s = scene, o = object, p = part). Our descriptions have greater object detail and introduce part-level details, while scaling up the number of descriptions, and semantic classes and objects described." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.906, + 0.44 + ], + "angle": 0, + "content": "Implementation Details Predicted instances are matched with GT instances using Hungarian matching [34], and the predicted captions are supervised by the captions of the matched instances. We train both the instance segmentation and captioning stages with a batch size of 6 scenes. In the captioning stage, one batch that is input to the captioner contains all objects from scenes in the batch." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.453, + 0.691, + 0.471 + ], + "angle": 0, + "content": "4. ExCap3D Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.479, + 0.907, + 0.766 + ], + "angle": 0, + "content": "Existing datasets for captioning objects in 3D indoor scenes contain captions at a single level of detail, that combines both object properties such as color and texture, as well as spatial and functional relations to other objects in the scene. To enable the task of captioning at both the object- and part-levels of detail, we created the ExCap3D Dataset based on the ScanNet++ [52] dataset of 3D indoor scans. The data is generated in a scalable and fully automated manner by leveraging a VLM that takes an RGB image crop of an object and a text prompt as input, and outputs a description of the object. The objects to be captioned are identified using the ground truth semantic instance annotations provided by ScanNet++. We prompt the VLM with text prompts to generate specific information about an object's color, texture, material and similar properties, and leverage multiview reasoning to obtain reliable final caption generation for ScanNet++. A comparison of the ExCap3D Dataset with existing object_caption and other 3D object-text datasets is shown in Tab. 1." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.773, + 0.809, + 0.79 + ], + "angle": 0, + "content": "4.1. Object Level of Detail Captioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.795, + 0.906, + 0.903 + ], + "angle": 0, + "content": "We utilize the ground truth 3D instance and semantic annotations in the ScanNet++ dataset to generate the ExCap3D Dataset. We project the 3D annotations onto the high-resolution DSLR images that are registered with the 3D scan, to obtain a corresponding 2D bounding box of each object in each DSLR image. We use this 2D mask to crop a rectangular region containing the object from the image," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21703" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.482, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.189, + 0.485, + 0.274 + ], + "angle": 0, + "content": "Figure 4. Generation of ExCap3D Dataset. We use the ground truth 3D semantics in ScanNet++, project them onto multiview DSLR images and obtain descriptions of the image crops \\( c_{i,j} \\) using a VLM. For parts crops, we use pseudo-ground truth from MaskClustering. Finally, we summarize the captions from different views using an LLM." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.421 + ], + "angle": 0, + "content": "after expanding the bounding box dimensions by \\(10\\%\\). To maximize the RGB information that is input to the VLM, we pick the top 3 images that cover at least \\(10\\%\\) the mesh vertices corresponding to the object mask. The VLM generates \\(k = 5\\) descriptions for each view of the object, giving outputs \\(c_{i,j}\\) from the \\(i^{\\mathrm{th}}\\) view. We use an instruction-tuned large language model (LLM) to obtain the \\(j^{th}\\) object-level caption \\(\\hat{c}_{\\mathrm{obj},j}\\) by combining the \\(j^{th}\\) caption from each view." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.427, + 0.37, + 0.444 + ], + "angle": 0, + "content": "4.2. Part Level of Detail Captioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.45, + 0.484, + 0.735 + ], + "angle": 0, + "content": "We produce pseudo-ground truth 3D part segmentation data using a modified version of MaskClustering [51] that favours smaller segments over larger ones, along with SAM [33]. SAM is used to generate initial 2D segments on the DSLR images. MaskClustering combines these 2D segments to produce coherent 3D part segments. We associate these 3D parts with the objects in ScanNet++ based on an intersection-over-union (IoU) threshold. The rest of the part captioning process largely follows that of object captioning. We generate one caption \\( p_i \\) in each view, for each part, and then combine the descriptions of all the parts into a single part-level caption \\( \\hat{c}_{\\mathrm{part}} \\) for each object using the LLM. While the pseudo-masks at the part-level obtained from MaskClustering can detect parts of long-tail and small objects in the ScanNet++ data, these masks are not as reliable as the manually annotated object GT masks in ScanNet++. Hence, we aggregate the captions of different parts into a single caption containing overall part-level details for an object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.842 + ], + "angle": 0, + "content": "Implementation Details We use Llava 1.6 [37] (7B parameters) with a Mistral backbone as the VLM, and instruction-tuned Llama 3.1 [21] (8B parameters) for the LLM. To ensure the quality of the captions for evaluation, we manually removed incorrect captions in the test set. Due to the robust multiview aggregation of captions, \\(93.21\\%\\) of the captions were found to be accurate." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.846, + 0.224, + 0.863 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "We evaluate caption generation on the 84 instance class categories of the ScanNet++ benchmark [52]. We use the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.878, + 0.107 + ], + "angle": 0, + "content": "train/val/test splits of the semantic tasks of ScanNet++." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.118, + 0.698, + 0.133 + ], + "angle": 0, + "content": "5.1. Evaluation Metrics" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.141, + 0.906, + 0.216 + ], + "angle": 0, + "content": "We adopt the commonly used CIDEr [44], ROUGE [36], METEOR [20] and BLUE-4 [41] caption evaluation metrics used by other 3D captioning methods [10, 15, 27, 58, 60]. Metrics are reported separately for the object- and part-level details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.229, + 0.623, + 0.243 + ], + "angle": 0, + "content": "5.2. Baselines" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.252, + 0.907, + 0.447 + ], + "angle": 0, + "content": "We compare our method against six state-of-the-art captioning methods. Cap3D [38] captions individual objects using pretrained BLIP2 [35] and GPT-4 [1]. LLaVA-3D [57] aggregates multiview 2D features and uses these as input to an LLM, and is evaluated zero-shot from a pretrained model. D3Net [9] uses a PointGroup sparse convolutional backbone with embedded object relations, and a GRU for text generation. Vote2Cap-DETR [10] tokenizes the 3D input, and generates captions using GPT2 [42]. Vote2Cap-DETR++ [12] additionally uses decoupled detection and caption queries. PQ3D [60] uses promptable queries for regions of interest, and generates captions with T5-Small [43]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.906, + 0.54 + ], + "angle": 0, + "content": "We train on ScanNet++ [52] for bounding box detection (Vote2Cap and D3Net) or instance segmentation (PQ3D and ExCap3D). Since the baselines were designed for single level of detail captioning, we train them separately on the object- and part-level captions, while ExCap3D is trained jointly on both levels of detail." + }, + { + "type": "table", + "bbox": [ + 0.516, + 0.553, + 0.902, + 0.678 + ], + "angle": 0, + "content": "
MethodObject-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]5.57.67.70.0
D3Net [9]6.75.46.70.4
LLaVA-3D [57]0.06.016.30.0
Vote2Cap-DETR [10]13.312.917.21.0
Vote2Cap-DETR++ [12]10.611.915.71.1
PQ3D [60]27.911.612.61.9
ExCap3D (Ours)32.916.718.01.2
" + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.68, + 0.902, + 0.801 + ], + "angle": 0, + "content": "
MethodPart-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]2.111.49.80.0
D3Net [9]7.77.67.31.3
LLaVA-3D [57]0.613.015.50.0
Vote2Cap-DETR [10]7.720.321.73.9
Vote2Cap-DETR++ [12]19.519.919.83.8
PQ3D [60]14.516.015.34.3
ExCap3D (Ours)32.821.921.14.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.88 + ], + "angle": 0, + "content": "Table 2. Comparison of object-level and part-level captioning performance between ExCap3D and baselines. The second-best result is underlined. ExCap3D improves CIDEr scores significantly through joint captioning and consistency losses. All metrics are evaluated at an IoU threshold of 0.5" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21704" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.111, + 0.088, + 0.887, + 0.845 + ], + "angle": 0, + "content": "
Input 3D Scene\nDetected 3D Objectwhite, rectangular, smooth, glossy, unk*white, rectangular, smooth, glossy, unk*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with white, smooth, glossy surface, made of wood or laminate, metal*
D3Net\n(Chen et al. 21)Object-levelwhite, smooth, possibly plastic or metal*a white, smooth, possibly plastic or metal*a black, curved backrest with a smooth texture, attached to a black metal frame with a smooth texture.a white, smooth, possibly plastic or metal*
Part-level
Vote2Cap-DETR\n(Chen et al. 23)Object-levelBlack, rectangular, smooth, glossy, plastic or metal computer tower. Black, smooth, plastic or metal, rectangular, cylindrical, or cylWhite, rectangular cabinet with a smooth, glossy finish, made of wood or laminate,like composite, flat top, and a handle. Dark woodA modern chair with a curved backrest, smooth glossy finish, dark fabric, and a chrome metal frame.. Modern, vibrant red, smooth,Dark-colored, leather-like, smooth, glossy, curved, L-shaped sofa. L-shaped sofa with a sturdy frame. Black,
Part-levelA black, rectangular trash can with a smooth, possibly plastic or metal surface, a black, cylindrical trash can with a smooth, possibly plastic surface,*A rectangular file cabinet with a flat top and bottom, made of light-colored wood with a smooth finish and visible grain pattern. It has a flat top surface and a slightly curved top edge. A rectangular drawer with a flat top and bottomThe chair has a curved, dark-colored backrest with a smooth, glossy finish, a curved, dark-colored backrest with a smooth texture, and a metal frame. The backrest is a curved, smooth, glossy piece of darkThe sofa has a curved, padded backrest with a smooth, plush texture, a curved, padded armrest with a smooth, glossy finish.* The backrest is a curved
PQ3D\n(Zhu et al. 24)Object-levelblack, rectangular, smooth, glossy, plastic or metal*white, rectangular, smooth, glossy, or metal.*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with a smooth, glossy, glossy surface, and a light-colored, possibly wood or wood-like material
Part-levela black, rectangular computer tower with a flat top and bottom, curved top, and a flat top, made of a smooth, possibly plastic or metal materialthe cabinet has a light-colored, smooth finish, possibly wood or composite, with a flat top and bottom*, a slightly curved top edge, and a slightly angled frontthe chair has a dark-colored, smooth, possibly plastic or metal backrest with a curved shape and a uniform texture.the sofa has a curved, cushioned backrest with a smooth, glossy finish, a curved, padded armrest with a smooth, glossy finish*
ExCap3D\n(Ours)Object-levelBlack, glossy, rectangular computer tower with a smooth textureWooden cabinet with light brown color, smooth texture, and rectangular shape.Wooden chair with light brown color, smooth texture, and rectangular shape.Black leather sofa with smooth texture and rectangular shape with rounded corners.
Part-levelA black computer tower with a textured surface, vents, and a rectangular drive bay with a textured surface.The cabinet has a light-colored, smooth finish, possibly made of wood or a wood-like composite, with a flat top and bottom, a series of drawers with metal handles.The chair has a light brown, smooth, curved backrest made of wood or wood-like composite, attached to a metal frame.The sofa has a black leather backrest with a smooth, leather-like texture, a curved armrest with a smooth, possibly leather-like material, and a rectangular cushion with a smooth texture.
Ground TruthObject-levelBlack, rectangular computer tower with a flat top, smooth, polished surface, and a tapered base.Light-colored, rectangular, boxy cabinet with smooth, glossy, synthetic surface.A chair with a curved backrest, flat seat, smooth, light brown, wooden surface, and metal legs.Dark-colored leather sofa with curved shape, plush cushions, and smooth texture.
Part-levelA black computer tower with a black power supply unit attached, both with smooth, glossy finishes.The cabinet has a light brown, smooth wooden surface, glass front, three wooden drawers with silver, curved handles, and a dark-colored, rectangular door.A curved, light-colored wood or wood-like composite backrest with smooth or matte finish, supported by a metal frame.The sofa has a curved, padded armrest, a cushion, and a backrest, all upholstered in dark, rich brown or black leather with a smooth, shiny texture.
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.86, + 0.907, + 0.917 + ], + "angle": 0, + "content": "Figure 5. Qualitative evaluation on ScanNet++ [52] scenes, in comparison with D3Net [9], Vote2CAP-DETR [10], and PQ3D [60]. Predicted text from methods are denoted by color to indicate correct and incorrect generated phrases compared to the ground truth. Our method produces more consistent and detailed captions at both the object- and part-levels of detail. * indicates truncated output where the underlined phrase was predicted repeatedly." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21705" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.365, + 0.107 + ], + "angle": 0, + "content": "5.3. Comparison to State of the Art" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.113, + 0.482, + 0.265 + ], + "angle": 0, + "content": "We compare with state-of-the-art captioning methods in Tab. 2. ExCap3D outperforms all baselines on the CIDEr score, which aligns the most closely with human perception, improving upon the highest performing baseline PQ3D by a margin of \\(17\\%\\) and \\(124\\%\\) for object- and part-level details respectively. We also outperform the baselines on the ROUGE and METEOR scores of object-level details, and ROUGE scores of part-level details. Our joint prediction of captions at different levels of detail enables more robust, accurate descriptions." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.266, + 0.485, + 0.462 + ], + "angle": 0, + "content": "Qualitative comparisons are shown in Fig. 5. ExCap3D produces complete and detailed descriptions at both the object- and part-levels of detail, while the baseline methods often omit details in both levels. Cap3D's BLIP-2 captioner can produce limited details, and is easily distracted by background objects. LLaVA-3D produces general captions with few object-specific details, and has limited object localization due to a single location prompt as input. D3Net uses a limited GLOVE embedding of text which cannot produce detailed output descriptions. Vote2Cap-DETR, Vote2Cap-DETR++ and PQ3D interpret 3D scenes more coarsely, which can result in hallucination of incorrect fine-grained details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.473, + 0.255, + 0.488 + ], + "angle": 0, + "content": "5.4. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.496, + 0.483, + 0.573 + ], + "angle": 0, + "content": "We ablate the different components of our method and show the results in Tab. 3. As a baseline, we use two independent captioning heads and train with only \\(\\mathcal{L}_{\\mathrm{caption}}\\). Each component of our method is added separately to the baseline to show its efficacy." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.751 + ], + "angle": 0, + "content": "Enforcing multilevel consistency improves caption quality. Semantic consistency primarily improves part-level details while textual consistency primarily improves object-level details as shown in Tab. 3, which demonstrates the complementary nature of these two losses. As the object-level captions typically contain a stronger notion of global object semantics compared to the part-level caption, the latter improves in quality with semantic consistency. The object-level captions also receive fine-grained cues through textual similarity from the part-level details which helps to refine the global description of the object." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.762, + 0.483, + 0.913 + ], + "angle": 0, + "content": "Part-object captioner information sharing improves performance Adding part \\(\\rightarrow\\) object information sharing by conditioning object captions on the part-level of detail improves both levels of captions as shown in Tab. 3. This is because object details are improved when they are informed by part captions which contain aggregated details that can produce a better global description of the object. At the same time, the part-level of detail improves as it is guided to produce consistent descriptions of parts that can be useful for object-level captioning." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.909, + 0.198 + ], + "angle": 0, + "content": "
MethodObject-LevelPart-level
Separate Models29.717.9
w/ Semantic Consistency30.223.4
w/ Textual Consistency32.218.6
w/ Part → Obj. Inf. Sharing34.824.0
Full Model (Ours)32.932.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.206, + 0.905, + 0.263 + ], + "angle": 0, + "content": "Table 3. Ablation study of the different components of our method. Performance is reported using CIDEr for object-level and part-level captions. Each ablation is applied independently to the baseline with separate models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.907, + 0.469 + ], + "angle": 0, + "content": "Consistency and information sharing are complementary to each other The semantic and textual consistency losses act independently of the part \\(\\rightarrow\\) object information sharing mechanism, and hence our full model improves significantly over the individual components, particularly in the part-level of detail as shown in Tab. 3. Intuitively, this is because the consistency losses model only two kinds of consistency over lower-dimension embeddings, while sharing language model hidden states enables the object-level captioner to directly condition on the internal high-dimension representation of the part-level captions with fine-grained features." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.477, + 0.909, + 0.629 + ], + "angle": 0, + "content": "Limitations While our approach generates expressive captioning for 3D scans, various limitations remain. We use two captioning models, with one conditioned on the hidden states of the other through cross-attention. This could limit the amount of information shared between the captioners. Our method relies on a sparse convolutional backbone [17] which is limited to a resolution of about \\(2\\mathrm{cm}\\). This could restrict captioning performance on small or thin objects. Similarly, the class imbalance in the ScanNet++ dataset could restrict performance on long-tail classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.641, + 0.634, + 0.657 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.666, + 0.907, + 0.803 + ], + "angle": 0, + "content": "We introduced ExCap3D, a method for expressive dense captioning of 3D scenes at multiple object- and part-levels of detail. We showed that describing 3D objects as the sum of their parts while maintaining consistency between the multiple levels of descriptions leads to more accurate descriptions aligned with human perception. We hope that our method and the accompanying ExCap3D Dataset open up new possibilities for a holistic, expressive understanding of 3D scans." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.816, + 0.673, + 0.832 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.903 + ], + "angle": 0, + "content": "This project was supported by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding.\"" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21706" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.257 + ], + "angle": 0, + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.259, + 0.483, + 0.327 + ], + "angle": 0, + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.33, + 0.483, + 0.398 + ], + "angle": 0, + "content": "[4] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.402, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[5] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.473, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[6] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.546, + 0.483, + 0.614 + ], + "angle": 0, + "content": "[7] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16464-16473, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.483, + 0.699 + ], + "angle": 0, + "content": "[8] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.483, + 0.77 + ], + "angle": 0, + "content": "[9] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In European Conference on Computer Vision, pages 487-505. Springer, 2022. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11124-11133, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[12] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7331-7347, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[13] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.288, + 0.905, + 0.342 + ], + "angle": 0, + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.344, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[15] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[16] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 18109-18119, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.484, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[18] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[19] Xueqing Deng, Qihang Yu, Ali Athar, Chenglin Yang, Linjie Yang, Xiaojie Jin, Xiaohui Shen, and Liang-Chieh Chen. Coconut-pancap: Joint panoptic segmentation and grounded captions for fine-grained understanding and generation. arXiv preprint arXiv:2502.02589, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.692, + 0.905, + 0.748 + ], + "angle": 0, + "content": "[20] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.749, + 0.905, + 0.815 + ], + "angle": 0, + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[22] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.86, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[23] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4, 1, 3" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21707" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.161 + ], + "angle": 0, + "content": "[24] Yunhao Ge, Xiaohui Zeng, Jacob Samuel Huffman, Tsung-Yi Lin, Ming-Yu Liu, and Yin Cui. Visual fact checker: enabling high-fidelity detailed caption generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14033-14042, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.485, + 0.232 + ], + "angle": 0, + "content": "[25] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.484, + 0.3 + ], + "angle": 0, + "content": "[26] Hang Hua, Qing Liu, Lingzhi Zhang, Jing Shi, Zhifei Zhang, Yilin Wang, Jianming Zhang, and Jiebo Luo. Finecaption: Compositional image captioning focusing on wherever you want at any granularity. arXiv preprint arXiv:2411.15411, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.301, + 0.484, + 0.384 + ], + "angle": 0, + "content": "[27] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.484, + 0.442 + ], + "angle": 0, + "content": "[28] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In European Conference on Computer Vision, pages 528-545. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.497 + ], + "angle": 0, + "content": "[29] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.485, + 0.566 + ], + "angle": 0, + "content": "[30] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.622 + ], + "angle": 0, + "content": "[31] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[32] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.748 + ], + "angle": 0, + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[34] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.135 + ], + "angle": 0, + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[38] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36: 75307-75337, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.907, + 0.247 + ], + "angle": 0, + "content": "[39] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In European Conference on Computer Vision, pages 180-197. Springer, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.906, + 0.277 + ], + "angle": 0, + "content": "[40] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.906, + 0.391 + ], + "angle": 0, + "content": "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.906, + 0.461 + ], + "angle": 0, + "content": "[43] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.462, + 0.906, + 0.517 + ], + "angle": 0, + "content": "[44] Gabriel Oliveira dos Santos, Esther Luna Colombini, and Sandra Avila. Cider-r: Robust consensus-based image description evaluation. arXiv preprint arXiv:2109.13701, 2021. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.518, + 0.907, + 0.589 + ], + "angle": 0, + "content": "[45] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.59, + 0.906, + 0.645 + ], + "angle": 0, + "content": "[46] Ayca Takmaz, Alexandros Delitzas, Robert W Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3d: Hierarchical open-vocabulary 3d segmentation. IEEE Robotics and Automation Letters, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.646, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[47] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.717, + 0.907, + 0.773 + ], + "angle": 0, + "content": "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.774, + 0.906, + 0.816 + ], + "angle": 0, + "content": "[49] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.817, + 0.906, + 0.871 + ], + "angle": 0, + "content": "[50] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.903 + ], + "angle": 0, + "content": "[51] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.907, + 0.903 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21708" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.124, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 6, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[53] Zuyao You, Junke Wang, Lingyu Kong, Bo He, and Zuxuan Wu. Pix2cap-coco: Advancing visual comprehension via pixel-level captioning. arXiv preprint arXiv:2501.13893, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.331 + ], + "angle": 0, + "content": "[54] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.333, + 0.482, + 0.401 + ], + "angle": 0, + "content": "[55] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15459-15469, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.403, + 0.482, + 0.459 + ], + "angle": 0, + "content": "[56] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[57] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[58] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.587, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[59] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[60] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188–206. Springer, 2024. 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.727 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "21709" + } + ] +] \ No newline at end of file diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_origin.pdf b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..927e6fd2399cd4b43990fc1a2bf4f07b5b0c4549 --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/6db99a4e-cf3e-43fb-908c-7fbd63bfc3bd_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ad9516281b0f5dc8166440fad198d3f3db6c698fbcfd90d7735eaf9bfdcc3c19 +size 6086673 diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/full.md b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/full.md new file mode 100644 index 0000000000000000000000000000000000000000..9375be59397b4dd526bde5d2167a2ce9ab5ea880 --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/full.md @@ -0,0 +1,310 @@ +# ExCap3D: Expressive 3D Scene Understanding via Object Captioning with Varying Detail + +Chandan Yeshwanth + +Dávid Rozenberszki + +Angela Dai + +Technical University of Munich + +https://cy94.github.io/excap3d/ + +![](images/0991bec62ee5f861c0a6a8add2c523d099186590c1871e0e0e4ff5cb0943fc4c.jpg) +Figure 1. ExCap3D produces descriptions at multiple level of details for objects in an input 3D scan. For each detected object, we generate both object- and part-level of detail captions. The generated captions are consistent across the two levels and contain rich details. + +# Abstract + +Generating text descriptions of objects in 3D indoor scenes is an important building block of embodied understanding. Existing methods describe objects at a single level of detail and do not capture fine-grained details of the parts of objects. In order to produce varying levels of detail capturing both coarse object-level information and detailed part-level descriptions, we propose the task of expressive 3D captioning. Given an input 3D scene, the task is to describe objects at multiple levels of detail: a high-level object description, and a low-level description of the properties of its parts. To produce such captions, we present ExCap3D, an expressive 3D captioning model which takes as input a 3D scan, and for each detected object in the scan, generates a fine-grained collective description of the parts of the object, along with an object-level description conditioned on the part-level description. We design ExCap3D to encourage consistency between the multiple levels of descriptions. To enable this task, we generated the ExCap3D Dataset by leveraging a visual-language model (VLM) for multi-view captioning. The ExCap3D Dataset contains captions on the ScanNet++ dataset with varying levels of detail, comprising 190k text descriptions of 34k 3D objects in 947 indoor scenes. Our experiments show that the object- and + +part-level details generated by ExCap3D are more expressive than those produced by state-of-the-art methods, with a CIDEr score improvement of $17\%$ and $124\%$ for object- and part-level details respectively. Our code, dataset and models will be made publicly available. + +# 1. Introduction + +3D indoor scene understanding is central to applications such as augmented reality (AR), virtual reality (VR), and robotics. While detecting 3D objects and their semantics is important for these applications, natural language descriptions of objects enable more wide-ranging possibilities – they can encode more complex information efficiently and, importantly, enable more natural interactions between embodied agents as well as humans and the scene. To produce a representation that facilities a diverse array of applications, such language descriptions must also encompass varying levels of detail: object-level descriptions for coarse identification in the scene, and part-level details for precise manipulation and fine-grained understanding. For instance, a robotic assistant would only need a coarse, high-level description of the objects in a room (e.g., 'a recliner chair') in order to navigate effectively in it, but it would need fine-grained descriptions of the objects (e.g., 'recliner chair with a high padded back, wooden armrests, a soft cushion and an + +adjustable footrest') to provide assistive AI descriptions for accessibility aid or to describe the scene to others. + +Current 3D captioning models have made significant progress in describing objects in indoor scenes by detecting objects and generating their individual descriptions at a single level of detail. These methods are trained only for captioning [10, 15, 32] or for multiple tasks [11, 55, 59, 60]. These methods can effectively describe coarse object information in the global context of the scene, but preclude detailed perception of the scene as they make limited use of fine-grained 3D inputs and cannot model object-part hierarchy. They fail to describe highly localized properties of object parts, such as appearance, material, and function. In order to enable more complex perception of a scene requiring these details, we propose to make use of fine-grained 3D inputs and explicitly model object-part hierarchy. We describe objects at multiple levels: an object-level of detail that captures its most prominent aspects (semantic class; possibly size or overall appearance), and a finer-grained part-level of detail that describes the details of elements composing the object (different materials, color, and functions). These two descriptions are complementary to each other and enable different applications depending on the level of detail required. Examples of such multilevel descriptions are shown in Fig. 1 and Fig. 2. + +To generate such multilevel captions in 3D scans, we introduce ExCap3D, a 3D object captioning model that produces expressive captions at both object- and part-levels of detail using two captioning heads. Examples of the captions generated by our method are shown in Fig. 1. The generated multilevel descriptions are complementary to each other and describe different aspects of the objects, while maintaining a few shared elements for consistency and coherence. The part-level caption generated by ExCap3D collectively describes the different parts of an object. Knowing the descriptions of the parts can enable our model to generate more consistent captions describing the object as a whole. Hence, ExCap3D first produces the part-level details and then uses these to inform the object-level captioner. To further reduce inconsistencies between the two levels of detail in terms of the object semantics and text content, ExCap3D employs additional consistency losses during training. + +To enable this multilevel captioning task, we provide the ExCap3D Dataset, which contains 190k object and partlevel natural language captions of 34k 3D objects from the 947 scenes of the ScanNet++ dataset. We automatically generate object captions for training, using a pretrained vision-language model (VLM) on high-resolution DSLR images along with ground truth semantic information from the ScanNet++ dataset [52]. We employ multi-view aggregation to robustly capture key object- and part-level details. VLMs perform strongly on 2D tasks, but + +![](images/d3a94ac01a17ef33178809c2442c6109ed8ed8d9e6019b4cbb18d069ddafed88.jpg) + +![](images/61347f251bbea1099252ff1792990875aca0fab4d8f536a18f6dcd5773a1ba1a.jpg) +a brown cushion chair with arms. it is located close to the wall. +Q: What is next to the orange armchair? +A: wall +A: another orange armchair +A: coffee table + +![](images/49f2febe57af4f2c95bc88c3a0e2756e939228e6fc6fe63b98f904873dea2149.jpg) +ExCap3D +Object-level: A three-seater leather couch with dark, smooth, glossy leather, curved and straight edges. +Part-Level: The black leather sofa has a smooth, glossy surface with visible stitching, a curved armrest, and cushioned seat, with a tufted backrest made of multiple leather panels stitched together, and sturdy structure with visible stitching along the edges. +Figure 2. Descriptions of objects in existing datasets such as Scan2Cap and ScanQA largely contain relations between objects, and limited local details at a single level. In contrast, the ExCap3D Dataset describes both the object as a whole and as a sum of its parts. + +struggle with 3D orientation [47] and distances [8]. Hence, our dataset contains large-scale and high-quality object- and part-level captions, in contrast to Scan2Cap [15] and Nr3D [2] with smaller-scale manually annotated spatial relations. A comparison of the multilevel captions from the ExCap3D Dataset with the single-level descriptions from the existing Scan2Cap and ScanQA [3] is shown in Fig. 2. The ExCap3D Dataset describes objects with significantly more detail, while providing two complementary multilevel descriptions of each object. + +We show that ExCap3D produces richer and more consistent object- and part-level captions compared to the state-of-the-art captioning methods, originally designed to generate only a single level description. To summarize, our main contributions are: + +- an expressive 3D captioning model that generates captions of objects in 3D indoor scenes at multiple levels of detail: fine-grained part-level details and object-level details +- modeling object-level detail as a sum of its parts' unique appearance and functional attributes by generating object-level captions conditioned on part-level descriptions, using the cross-attention mechanism of the object captioner and the language model's intermediate hidden states +- regularization of our joint 3D captioning approach through semantic and textual-similarity consistency losses that maintain consistency of the object-level and part-level predicted captions +- the ExCap3D Dataset: a large-scale, precise and challenging dataset of 190k descriptive object- and part-level captions of 34k objects in 3D indoor scenes built on the ScanNet++ dataset with fine-grained texture, material, and functional descriptions + +# 2. Related Work + +# 2.1. 2D Image Captioning + +2D image captioning has been extensively studied in computer vision. DenseCap [29] popularized the task of dense + +captioning of images, jointly localizing objects, and generating their short descriptions. More recently, COCONutPanCap [19] and Pix2Cap-COCO [53] combined the tasks of fine-grained instance segmentation and detailed captioning. FlexCap [22] and FineCaption [26] can produce descriptions at varying levels of detail given the image and a localized object as input. In contrast, we combine the tasks of detection and multilevel captioning and describe detected objects and their parts. + +# 2.2. 3D Scene Captioning + +Scan2Cap [15] introduced the task of 3D dense captioning and a method for context-aware captioning of 3D objects, operating on ScanNet [18] scenes. Since then, several methods have been developed to effectively solve this task by extracting object features and modeling their relations with nearby objects [7, 9, 10, 28, 31, 49, 54]. These methods can accurately describe objects at the scene-level through relations with neighboring objects. Following the trends of utilizing foundation models in 3D reasoning, Cap3D [38] and follow-up works [24, 39] describe isolated 3D objects through multi-view aggregation of VLM-generated captions. More recent works have jointly trained on multiple tasks such as 3D grounding, question answering, and dense captioning by training custom language models from scratch [16, 58] or by fine-tuning pretrained LLMs [11, 25, 27, 50, 57, 60]. However, these works focus on high-level, proximity-based relations between objects at the scene-level, underutilize local features to produce limited object-level details and no part-level details We tackle a similarly inspired problem, and leverage the high-fidelity geometry and RGB captures from ScanNet++ [52] to enable complex, multilevel joint captioning. Moreover, we demonstrate that our joint captioning approach improves both levels of detail. + +# 2.3. Part Understanding of 3D Objects in Scenes + +Several methods have focused on part-level understanding of 3D scenes by only relying on explicit geometry [5, 6], images and radiance fields [4, 30] or by using both modalities and mesh-based oversegmentation [46]. A related method for 3D shapes rather than scenes was proposed by Chen et. al. [13], which produces part-level segmentation maps and text descriptions of individual 3D objects at a single, fine-grained level of detail. This produces promising detailed captions, but is designed for isolated objects rather than complex 3D scenes. In contrast, we jointly detect and caption objects in large 3D scenes at multiple levels of detail, ensuring consistency and enabling greater expressivity for downstream tasks. + +# 3. Method + +Problem Formulation The task of expressive captioning is to detect the objects in a 3D scene and describe each object at different levels of detail: object- and part-levels. Given an input 3D scene in the form of a 3D mesh $M = (\mathcal{V},\mathcal{F})$ with vertices $\mathcal{V}\in \mathbb{R}^{N_v\times 3}$ and triangular faces $f_{i}\in \mathcal{F}$ connecting the vertices, we aim to detect the 3D object instances $o_i = (b_i,s_i,c_{\mathrm{obj},i},c_{\mathrm{part},i})$ in the scene, characterized by the objects' bounding boxes $b_{i}$ , semantic class $s_i$ , and $c_{\mathrm{obj},i}$ and $c_{\mathrm{part},i}$ describing the object in natural language at object- and part-levels of detail, respectively. + +Overview of ExCap3D First, ExCap3D performs instance segmentation and instance-aware query generation to produce queries containing rich per-object information. Next, the object- and part-level captioning models with shared information generate two separate text captions, conditioned on individual instance-aware queries and the corresponding features from the instance segmentation backbone. Finally, semantic and whole-text consistency losses are applied to the hidden states of the captioning models to regularize the consistency of the predicted captions. + +# 3.1. 3D Instance Segmentation + +Given the mesh of the 3D scene $M$ as input, the instance segmentation model produces a fixed number of refined instance-aware queries $Q_{r}$ corresponding to the $N_{q}$ objects detected in the scene, and their binary instance masks on the voxelized 3D input $I\in \{0,1\}^{N_q\times N_{\mathrm{vox}}}$ . The instance queries are a compact representation of the 3D features of the objects as well as their semantics. The instance segmentation model also produces dense features $F\in \mathbb{R}^{N_{\mathrm{vox}}\times D}$ on the voxels. The refined and instance-aware queries $Q_{r}$ and 3D voxel features $F$ are used to inform the per-instance captioning in the next step. + +The input mesh is first voxelized into a voxel grid $V \in \mathbb{R}^{N_{\mathrm{vox}} \times 6}$ containing the voxel coordinates and RGB colors. We use the Mask3D [45] transformer-based model for instance segmentation. It consists of a 3D sparse convolutional UNet backbone to encode the voxelized 3D scene, which outputs features $F \in \mathbb{R}^{N_{\mathrm{vox}} \times D}$ over the input voxels. A parallel mask module takes as input query vectors $Q \in \mathbb{R}^{N_q \times D_q}$ denoting the $N_q$ instances in the scene. The queries are refined iteratively through several transformer encoder layers, with cross-attention to the 3D backbone features. The output instance-aware query vectors contain rich representations of the object instances in the 3D scene and are used to predict the instance masks $I \in \{0,1\}^{N_q \times N_{\mathrm{vox}}}$ and semantic class of each object. + +# 3.2. Joint Captioning with Information Sharing + +ExCap3D contains two separate captioners for the object- and part-levels of details. These are transformer-based lan + +![](images/411fdae6ebb29bf9fbb66498e538d4530434426e1f9ec170b0befa1a2ea82bef.jpg) +Figure 3. Overview of our ExCap3D captioning method. We predict 3D instances in the input 3D scene using Mask3D [45], then predict object-level and part-levels of detail using two separate captioning heads. The object-level captions are further constrained by the part-level captioner's hidden states. Semantic- and textual consistency losses are applied to ensure the overall consistency of both predicted captions. + +guage models $\Psi_{\mathrm{obj}}$ and $\Psi_{\mathrm{part}}$ that produce the object-level caption $c_{\mathrm{obj},o}$ , part-level caption $c_{\mathrm{part},o}$ . They are conditioned on the instance-aware queries $Q_{r}$ and the dense voxel features $F$ , and are trained for next-token prediction. The language model captioners take as input a sequence of previously predicted tokens $t_1 \ldots t_{i-1}$ and decode the next token $t_i$ autoregressively as + +$$ +p \left(t _ {i} \mid t _ {1} \dots t _ {i - 1}\right) = \Psi \left(t _ {1} \dots t _ {i - 1}\right), +$$ + +where $\Psi$ is the language model. We provide two sources of rich instance-level information to the captioner: the instance-aware refined query output by the instance-segmentation model, and voxel features from 3D UNet backbone $F$ . The refined query vector $Q_{r,o}$ corresponding to an object $o$ is first projected with a linear layer to generate the caption-aware query $Q_{c,o} \in \mathbb{R}^{D_{\mathrm{caption}}}$ for that object + +$$ +Q _ {c, o} = \Phi_ {\mathrm {q u e r y}} (Q _ {r, o}), +$$ + +where $D_{\mathrm{caption}}$ is the embedding dimension of the language model. $Q_{c,o}$ contains localized information required to generate the caption for an object, such as its semantic class and its high-level features, and is used as the initial token to start the caption generation as + +$$ +p \left(t _ {i} \mid Q _ {c, o}, t _ {1} \dots t _ {i - 1}\right) = \Psi \left(Q _ {c, o}, t _ {1} \dots t _ {i - 1}\right). +$$ + +Additionally, we inject fine-grained information by using the captioner's cross-attention layers to attend to fine-grained contextual features from the 3D backbone $F$ corresponding to the object being captioned, after a linear projection to $D_{\mathrm{caption}}$ . To reduce the computational complexity of the cross-attention operation, we aggregate features within + +precomputed segments [23] on the mesh surface, giving the context features $S_{o} \in \mathbb{R}^{n_{s,o} \times D_{\mathrm{caption}}}$ , where $n_{s,o}$ is the number of segments in $I_{o}$ , the predicted mask for object $o$ . The cross-attention operation is $A = \operatorname{softmax}(qk^T / \sqrt{d})v$ where queries $q$ and values $k$ are the hidden states of $\Psi$ , and the keys $k$ are $S_{o}$ . The object and part captioners use context features $S_{\mathrm{obj},o}$ and $S_{\mathrm{part},o}$ respectively. + +Importantly, to enable the object-level captioner to be informed by the part-level captioner, we make use of the last layer hidden states $[h_{\mathrm{part},1}\dots h_{\mathrm{part},i}]$ of the part-level captioner $\Psi_{\mathrm{part}}$ . We use a linear projection + +$$ +H _ {\text {p a r t}, o} = \Phi_ {\text {h i d d e n}} ([ h _ {\text {p a r t}, 1} \dots h _ {\text {p a r t}, i} ]) +$$ + +and then concatenate $H_{\mathrm{part},o}$ to the contextual features of the object captioner $S_{\mathrm{obj},o}$ for object $o$ . The object-level caption for object $o$ is finally informed by the caption-aware query $Q_{c,o}$ as well as features $[H_{\mathrm{part},o};S_{\mathrm{obj},o}]$ . Since different objects in a batch have different numbers of predicted segments, we zero-pad $S_{\mathrm{obj},o}$ to the maximum context length in the batch. + +The two caption models are trained with cross-entropy (CE) loss against the object- and part-level ground truth (GT) captions respectively, giving the loss + +$$ +\mathcal {L} _ {\text {c a p t i o n}} = \mathcal {L} _ {\text {o b j - c a p t i o n}} + \mathcal {L} _ {\text {p a r t - c a p t i o n}} +$$ + +The outputs of this stage are the object-level caption $c_{\mathrm{obj},o}$ , part-level caption $c_{\mathrm{part},o}$ and the corresponding language model hidden states $h_{\mathrm{obj}} = [h_{\mathrm{obj},1}\dots h_{\mathrm{obj},i}]$ and $h_{\mathrm{part}} = [h_{\mathrm{part},1}\dots h_{\mathrm{part},j}]$ for each object $o$ . + +# 3.3. Semantic and Textual Consistency Losses + +While the object-level of detail is informed by the part-level of detail, supervising these two captioners with different GT + +captions may lead to inconsistencies in the predicted captions due to the different GT distributions of the two levels of detail. Hence, we use additional constraints in the form of consistency losses over the predicted captions. We encourage semantic consistency for both captions to refer to the same object semantics, such as its fine-grained object class and material, and textual consistency for the captions to be similar to each other. Since consistency losses cannot be applied directly on the generated text due to the text tokenizer being non-differentiable, we instead use the language model hidden states $h_{\mathrm{obj}}$ and $h_{\mathrm{part}}$ from the last captioner layers as a proxy. + +Semantic Consistency To promote both levels of captions to contain similar semantic information, we use a semantic consistency loss. We do this by projecting the hidden states $h_{\mathrm{obj}}$ and $h_{\mathrm{part}}$ into a lower dimension to retain only meaningful semantic information over the whole caption. We then pass these sequences independently through object- and part-level caption projection transformers $\Phi_{\mathrm{obj,sem}}$ and $\Phi_{\mathrm{part,sem}}$ to obtain new embedded sequences, and these aggregate into single feature vectors representing the overall semantics of the predicted captions $h_{\mathrm{obj,sem}} \in \mathbb{R}^{D_{ff}}$ and $h_{\mathrm{part,sem}} \in \mathbb{R}^{D_{ff}}$ respectively, where $D_{ff}$ is the feed-forward dimension of $\Phi$ . These are again independently projected into $N_{\mathrm{sem}}$ fine-grained latent classes using linear layers to get class probabilities $sem_{\mathrm{obj}}$ and $sem_{\mathrm{part}}$ . The semantic consistency loss is a symmetric sum [14] of cross-entropy (CE) losses + +$$ +\begin{array}{r} \mathcal {L} _ {\mathrm {s e m a n t i c}} = C E (s e m _ {\mathrm {o b j}}, S G (s e m _ {\mathrm {p a r t}})) + \\ C E (s e m _ {\mathrm {p a r t}}, S G (s e m _ {\mathrm {o b j}})), \end{array} +$$ + +where $SG$ is the stop-gradient operator. + +Textual Consistency The object- and part-level captions are expected to slightly overlap, since the salient features of an object are reflected in its object-level caption. We encourage this similarity through the similarity of the whole sequences of object-level and part-level captions. Similar to the semantic consistency loss, we use projectors $\Phi_{\mathrm{obj},\mathrm{text}}$ and $\Phi_{\mathrm{part},\mathrm{text}}$ to project the language model hidden states into $h_{\mathrm{obj},\mathrm{text}}$ and $h_{\mathrm{part},\mathrm{text}}$ , and aggregate over the sequence of tokens to get single feature vectors $\overline{h}_{\mathrm{obj},\mathrm{text}}$ and $\overline{h}_{\mathrm{obj},\mathrm{part}}$ . The text-consistency loss is the distance $d(.,.)$ between these two feature vectors + +$$ +\mathcal {L} _ {\mathrm {t e x t u a l}} = d (\overline {{h}} _ {\mathrm {o b j , t e x t}}, \overline {{h}} _ {\mathrm {p a r t , t e x t}}). +$$ + +The final loss used to optimize ExCap3D is a weighted sum of the individual losses + +$$ +\mathcal {L} = w _ {1} \mathcal {L} _ {\text {c a p t i o n}} + w _ {2} \mathcal {L} _ {\text {s e m a n t i c}} + w _ {3} \mathcal {L} _ {\text {t e x t u a l}} +$$ + +
Dataset#Desc#Cls#ObjsLoD
NR3D [2]41k765.8ks + o
Multi3DRefer [56]55k26510.5ks + o
ScanQA [3]35k3709.5ks + o
Scan2Cap [15]46k2659.9ks + o
Ours190k2k34.7ko + p
+ +Table 1. Comparison of the ExCap3D Dataset with existing datasets that describe objects in 3D indoor scenes. We compare the number of descriptions (#Desc), number of unique object classes described (#Cls), number of objects described (#Objs) and level of detail (LoD, s = scene, o = object, p = part). Our descriptions have greater object detail and introduce part-level details, while scaling up the number of descriptions, and semantic classes and objects described. + +Implementation Details Predicted instances are matched with GT instances using Hungarian matching [34], and the predicted captions are supervised by the captions of the matched instances. We train both the instance segmentation and captioning stages with a batch size of 6 scenes. In the captioning stage, one batch that is input to the captioner contains all objects from scenes in the batch. + +# 4. ExCap3D Dataset + +Existing datasets for captioning objects in 3D indoor scenes contain captions at a single level of detail, that combines both object properties such as color and texture, as well as spatial and functional relations to other objects in the scene. To enable the task of captioning at both the object- and part-levels of detail, we created the ExCap3D Dataset based on the ScanNet++ [52] dataset of 3D indoor scans. The data is generated in a scalable and fully automated manner by leveraging a VLM that takes an RGB image crop of an object and a text prompt as input, and outputs a description of the object. The objects to be captioned are identified using the ground truth semantic instance annotations provided by ScanNet++. We prompt the VLM with text prompts to generate specific information about an object's color, texture, material and similar properties, and leverage multiview reasoning to obtain reliable final caption generation for ScanNet++. A comparison of the ExCap3D Dataset with existing object_caption and other 3D object-text datasets is shown in Tab. 1. + +# 4.1. Object Level of Detail Captioning + +We utilize the ground truth 3D instance and semantic annotations in the ScanNet++ dataset to generate the ExCap3D Dataset. We project the 3D annotations onto the high-resolution DSLR images that are registered with the 3D scan, to obtain a corresponding 2D bounding box of each object in each DSLR image. We use this 2D mask to crop a rectangular region containing the object from the image, + +![](images/a2b3e4384a1a5d9a5f77f6e0664af3c55f80bba9c00984515ea93d97b166a31c.jpg) +Figure 4. Generation of ExCap3D Dataset. We use the ground truth 3D semantics in ScanNet++, project them onto multiview DSLR images and obtain descriptions of the image crops $c_{i,j}$ using a VLM. For parts crops, we use pseudo-ground truth from MaskClustering. Finally, we summarize the captions from different views using an LLM. + +after expanding the bounding box dimensions by $10\%$ . To maximize the RGB information that is input to the VLM, we pick the top 3 images that cover at least $10\%$ the mesh vertices corresponding to the object mask. The VLM generates $k = 5$ descriptions for each view of the object, giving outputs $c_{i,j}$ from the $i^{\mathrm{th}}$ view. We use an instruction-tuned large language model (LLM) to obtain the $j^{th}$ object-level caption $\hat{c}_{\mathrm{obj},j}$ by combining the $j^{th}$ caption from each view. + +# 4.2. Part Level of Detail Captioning + +We produce pseudo-ground truth 3D part segmentation data using a modified version of MaskClustering [51] that favours smaller segments over larger ones, along with SAM [33]. SAM is used to generate initial 2D segments on the DSLR images. MaskClustering combines these 2D segments to produce coherent 3D part segments. We associate these 3D parts with the objects in ScanNet++ based on an intersection-over-union (IoU) threshold. The rest of the part captioning process largely follows that of object captioning. We generate one caption $p_i$ in each view, for each part, and then combine the descriptions of all the parts into a single part-level caption $\hat{c}_{\mathrm{part}}$ for each object using the LLM. While the pseudo-masks at the part-level obtained from MaskClustering can detect parts of long-tail and small objects in the ScanNet++ data, these masks are not as reliable as the manually annotated object GT masks in ScanNet++. Hence, we aggregate the captions of different parts into a single caption containing overall part-level details for an object. + +Implementation Details We use Llava 1.6 [37] (7B parameters) with a Mistral backbone as the VLM, and instruction-tuned Llama 3.1 [21] (8B parameters) for the LLM. To ensure the quality of the captions for evaluation, we manually removed incorrect captions in the test set. Due to the robust multiview aggregation of captions, $93.21\%$ of the captions were found to be accurate. + +# 5. Experiments + +We evaluate caption generation on the 84 instance class categories of the ScanNet++ benchmark [52]. We use the + +train/val/test splits of the semantic tasks of ScanNet++. + +# 5.1. Evaluation Metrics + +We adopt the commonly used CIDEr [44], ROUGE [36], METEOR [20] and BLUE-4 [41] caption evaluation metrics used by other 3D captioning methods [10, 15, 27, 58, 60]. Metrics are reported separately for the object- and part-level details. + +# 5.2. Baselines + +We compare our method against six state-of-the-art captioning methods. Cap3D [38] captions individual objects using pretrained BLIP2 [35] and GPT-4 [1]. LLaVA-3D [57] aggregates multiview 2D features and uses these as input to an LLM, and is evaluated zero-shot from a pretrained model. D3Net [9] uses a PointGroup sparse convolutional backbone with embedded object relations, and a GRU for text generation. Vote2Cap-DETR [10] tokenizes the 3D input, and generates captions using GPT2 [42]. Vote2Cap-DETR++ [12] additionally uses decoupled detection and caption queries. PQ3D [60] uses promptable queries for regions of interest, and generates captions with T5-Small [43]. + +We train on ScanNet++ [52] for bounding box detection (Vote2Cap and D3Net) or instance segmentation (PQ3D and ExCap3D). Since the baselines were designed for single level of detail captioning, we train them separately on the object- and part-level captions, while ExCap3D is trained jointly on both levels of detail. + +
MethodObject-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]5.57.67.70.0
D3Net [9]6.75.46.70.4
LLaVA-3D [57]0.06.016.30.0
Vote2Cap-DETR [10]13.312.917.21.0
Vote2Cap-DETR++ [12]10.611.915.71.1
PQ3D [60]27.911.612.61.9
ExCap3D (Ours)32.916.718.01.2
+ +
MethodPart-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]2.111.49.80.0
D3Net [9]7.77.67.31.3
LLaVA-3D [57]0.613.015.50.0
Vote2Cap-DETR [10]7.720.321.73.9
Vote2Cap-DETR++ [12]19.519.919.83.8
PQ3D [60]14.516.015.34.3
ExCap3D (Ours)32.821.921.14.8
+ +Table 2. Comparison of object-level and part-level captioning performance between ExCap3D and baselines. The second-best result is underlined. ExCap3D improves CIDEr scores significantly through joint captioning and consistency losses. All metrics are evaluated at an IoU threshold of 0.5 + +
Input 3D Scene +Detected 3D Objectwhite, rectangular, smooth, glossy, unk*white, rectangular, smooth, glossy, unk*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with white, smooth, glossy surface, made of wood or laminate, metal*
D3Net +(Chen et al. 21)Object-levelwhite, smooth, possibly plastic or metal*a white, smooth, possibly plastic or metal*a black, curved backrest with a smooth texture, attached to a black metal frame with a smooth texture.a white, smooth, possibly plastic or metal*
Part-level
Vote2Cap-DETR +(Chen et al. 23)Object-levelBlack, rectangular, smooth, glossy, plastic or metal computer tower. Black, smooth, plastic or metal, rectangular, cylindrical, or cylWhite, rectangular cabinet with a smooth, glossy finish, made of wood or laminate,like composite, flat top, and a handle. Dark woodA modern chair with a curved backrest, smooth glossy finish, dark fabric, and a chrome metal frame.. Modern, vibrant red, smooth,Dark-colored, leather-like, smooth, glossy, curved, L-shaped sofa. L-shaped sofa with a sturdy frame. Black,
Part-levelA black, rectangular trash can with a smooth, possibly plastic or metal surface, a black, cylindrical trash can with a smooth, possibly plastic surface,*A rectangular file cabinet with a flat top and bottom, made of light-colored wood with a smooth finish and visible grain pattern. It has a flat top surface and a slightly curved top edge. A rectangular drawer with a flat top and bottomThe chair has a curved, dark-colored backrest with a smooth, glossy finish, a curved, dark-colored backrest with a smooth texture, and a metal frame. The backrest is a curved, smooth, glossy piece of darkThe sofa has a curved, padded backrest with a smooth, plush texture, a curved, padded armrest with a smooth, glossy finish.* The backrest is a curved
PQ3D +(Zhu et al. 24)Object-levelblack, rectangular, smooth, glossy, plastic or metal*white, rectangular, smooth, glossy, or metal.*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with a smooth, glossy, glossy surface, and a light-colored, possibly wood or wood-like material
Part-levela black, rectangular computer tower with a flat top and bottom, curved top, and a flat top, made of a smooth, possibly plastic or metal materialthe cabinet has a light-colored, smooth finish, possibly wood or composite, with a flat top and bottom*, a slightly curved top edge, and a slightly angled frontthe chair has a dark-colored, smooth, possibly plastic or metal backrest with a curved shape and a uniform texture.the sofa has a curved, cushioned backrest with a smooth, glossy finish, a curved, padded armrest with a smooth, glossy finish*
ExCap3D +(Ours)Object-levelBlack, glossy, rectangular computer tower with a smooth textureWooden cabinet with light brown color, smooth texture, and rectangular shape.Wooden chair with light brown color, smooth texture, and rectangular shape.Black leather sofa with smooth texture and rectangular shape with rounded corners.
Part-levelA black computer tower with a textured surface, vents, and a rectangular drive bay with a textured surface.The cabinet has a light-colored, smooth finish, possibly made of wood or a wood-like composite, with a flat top and bottom, a series of drawers with metal handles.The chair has a light brown, smooth, curved backrest made of wood or wood-like composite, attached to a metal frame.The sofa has a black leather backrest with a smooth, leather-like texture, a curved armrest with a smooth, possibly leather-like material, and a rectangular cushion with a smooth texture.
Ground TruthObject-levelBlack, rectangular computer tower with a flat top, smooth, polished surface, and a tapered base.Light-colored, rectangular, boxy cabinet with smooth, glossy, synthetic surface.A chair with a curved backrest, flat seat, smooth, light brown, wooden surface, and metal legs.Dark-colored leather sofa with curved shape, plush cushions, and smooth texture.
Part-levelA black computer tower with a black power supply unit attached, both with smooth, glossy finishes.The cabinet has a light brown, smooth wooden surface, glass front, three wooden drawers with silver, curved handles, and a dark-colored, rectangular door.A curved, light-colored wood or wood-like composite backrest with smooth or matte finish, supported by a metal frame.The sofa has a curved, padded armrest, a cushion, and a backrest, all upholstered in dark, rich brown or black leather with a smooth, shiny texture.
+ +Figure 5. Qualitative evaluation on ScanNet++ [52] scenes, in comparison with D3Net [9], Vote2CAP-DETR [10], and PQ3D [60]. Predicted text from methods are denoted by color to indicate correct and incorrect generated phrases compared to the ground truth. Our method produces more consistent and detailed captions at both the object- and part-levels of detail. * indicates truncated output where the underlined phrase was predicted repeatedly. + +# 5.3. Comparison to State of the Art + +We compare with state-of-the-art captioning methods in Tab. 2. ExCap3D outperforms all baselines on the CIDEr score, which aligns the most closely with human perception, improving upon the highest performing baseline PQ3D by a margin of $17\%$ and $124\%$ for object- and part-level details respectively. We also outperform the baselines on the ROUGE and METEOR scores of object-level details, and ROUGE scores of part-level details. Our joint prediction of captions at different levels of detail enables more robust, accurate descriptions. + +Qualitative comparisons are shown in Fig. 5. ExCap3D produces complete and detailed descriptions at both the object- and part-levels of detail, while the baseline methods often omit details in both levels. Cap3D's BLIP-2 captioner can produce limited details, and is easily distracted by background objects. LLaVA-3D produces general captions with few object-specific details, and has limited object localization due to a single location prompt as input. D3Net uses a limited GLOVE embedding of text which cannot produce detailed output descriptions. Vote2Cap-DETR, Vote2Cap-DETR++ and PQ3D interpret 3D scenes more coarsely, which can result in hallucination of incorrect fine-grained details. + +# 5.4. Ablation Studies + +We ablate the different components of our method and show the results in Tab. 3. As a baseline, we use two independent captioning heads and train with only $\mathcal{L}_{\mathrm{caption}}$ . Each component of our method is added separately to the baseline to show its efficacy. + +Enforcing multilevel consistency improves caption quality. Semantic consistency primarily improves part-level details while textual consistency primarily improves object-level details as shown in Tab. 3, which demonstrates the complementary nature of these two losses. As the object-level captions typically contain a stronger notion of global object semantics compared to the part-level caption, the latter improves in quality with semantic consistency. The object-level captions also receive fine-grained cues through textual similarity from the part-level details which helps to refine the global description of the object. + +Part-object captioner information sharing improves performance Adding part $\rightarrow$ object information sharing by conditioning object captions on the part-level of detail improves both levels of captions as shown in Tab. 3. This is because object details are improved when they are informed by part captions which contain aggregated details that can produce a better global description of the object. At the same time, the part-level of detail improves as it is guided to produce consistent descriptions of parts that can be useful for object-level captioning. + +
MethodObject-LevelPart-level
Separate Models29.717.9
w/ Semantic Consistency30.223.4
w/ Textual Consistency32.218.6
w/ Part → Obj. Inf. Sharing34.824.0
Full Model (Ours)32.932.8
+ +Table 3. Ablation study of the different components of our method. Performance is reported using CIDEr for object-level and part-level captions. Each ablation is applied independently to the baseline with separate models. + +Consistency and information sharing are complementary to each other The semantic and textual consistency losses act independently of the part $\rightarrow$ object information sharing mechanism, and hence our full model improves significantly over the individual components, particularly in the part-level of detail as shown in Tab. 3. Intuitively, this is because the consistency losses model only two kinds of consistency over lower-dimension embeddings, while sharing language model hidden states enables the object-level captioner to directly condition on the internal high-dimension representation of the part-level captions with fine-grained features. + +Limitations While our approach generates expressive captioning for 3D scans, various limitations remain. We use two captioning models, with one conditioned on the hidden states of the other through cross-attention. This could limit the amount of information shared between the captioners. Our method relies on a sparse convolutional backbone [17] which is limited to a resolution of about $2\mathrm{cm}$ . This could restrict captioning performance on small or thin objects. Similarly, the class imbalance in the ScanNet++ dataset could restrict performance on long-tail classes. + +# 6. Conclusion + +We introduced ExCap3D, a method for expressive dense captioning of 3D scenes at multiple object- and part-levels of detail. We showed that describing 3D objects as the sum of their parts while maintaining consistency between the multiple levels of descriptions leads to more accurate descriptions aligned with human perception. We hope that our method and the accompanying ExCap3D Dataset open up new possibilities for a holistic, expressive understanding of 3D scans. + +# Acknowledgments + +This project was supported by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant "Learning How to Interact with Scenes through Part-Based Understanding." + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6 +[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 5 +[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 5 +[4] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 3 +[5] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3 +[6] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3 +[7] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16464-16473, 2022. 3 +[8] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2 +[9] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In European Conference on Computer Vision, pages 487-505. Springer, 2022. 3, 6, 7 +[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11124-11133, 2023. 2, 3, 6, 7 +[11] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF + +Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 3 +[12] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7331-7347, 2024. 6 +[13] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3 +[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 5 +[15] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 2, 3, 5, 6 +[16] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 18109-18119, 2023. 3 +[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 8 +[18] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3, 1 +[19] Xueqing Deng, Qihang Yu, Ali Athar, Chenglin Yang, Linjie Yang, Xiaojie Jin, Xiaohui Shen, and Liang-Chieh Chen. Coconut-pancap: Joint panoptic segmentation and grounded captions for fine-grained understanding and generation. arXiv preprint arXiv:2502.02589, 2025. 3 +[20] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014. 6, 3 +[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6 +[22] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. 2024. 3 +[23] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4, 1, 3 + +[24] Yunhao Ge, Xiaohui Zeng, Jacob Samuel Huffman, Tsung-Yi Lin, Ming-Yu Liu, and Yin Cui. Visual fact checker: enabling high-fidelity detailed caption generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14033-14042, 2024. 3 +[25] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3 +[26] Hang Hua, Qing Liu, Lingzhi Zhang, Jing Shi, Zhifei Zhang, Yilin Wang, Jianming Zhang, and Jiebo Luo. Finecaption: Compositional image captioning focusing on wherever you want at any granularity. arXiv preprint arXiv:2411.15411, 2024.3 +[27] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6 +[28] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In European Conference on Computer Vision, pages 528-545. Springer, 2022. 3 +[29] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2 +[30] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3 +[31] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 3 +[32] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 2 +[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 6, 3 +[34] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5 +[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 6 +[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6 + +[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.6 +[38] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36: 75307-75337, 2023. 3, 6 +[39] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In European Conference on Computer Vision, pages 180-197. Springer, 2024. 3 +[40] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 3 +[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6, 3 +[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6 +[43] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6 +[44] Gabriel Oliveira dos Santos, Esther Luna Colombini, and Sandra Avila. Cider-r: Robust consensus-based image description evaluation. arXiv preprint arXiv:2109.13701, 2021. 6, 3 +[45] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3, 4 +[46] Ayca Takmaz, Alexandros Delitzas, Robert W Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3d: Hierarchical open-vocabulary 3d segmentation. IEEE Robotics and Automation Letters, 2025. 3 +[47] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2 +[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 3 +[49] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 3 +[50] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 3 +[51] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for + +open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 6, 3 +[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 3, 5, 6, 7 +[53] Zuyao You, Junke Wang, Lingyu Kong, Bo He, and Zuxuan Wu. Pix2cap-coco: Advancing visual comprehension via pixel-level captioning. arXiv preprint arXiv:2501.13893, 2025. 3 +[54] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3 +[55] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15459-15469, 2024. 2 +[56] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 5 +[57] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 3, 6 +[58] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 3, 6 +[59] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2 +[60] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188–206. Springer, 2024. 2, 3, 6, 7 \ No newline at end of file diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/images.zip b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..bfe40649c66d731eb49441cdadf0ef501387f21f --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fca0bcbe26918b437bcaee10d86f0c150b7b9c6483ecd1866443b95782f57960 +size 691944 diff --git a/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/layout.json b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..a305d78a08d2543532866dc2a0082873d3cc751e --- /dev/null +++ b/2025/ExCap3D_ Expressive 3D Scene Understanding via Object Captioning with Varying Detail/layout.json @@ -0,0 +1,8706 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 90, + 102, + 520, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 90, + 102, + 520, + 140 + ], + "spans": [ + { + "bbox": [ + 90, + 102, + 520, + 140 + ], + "type": "text", + "content": "ExCap3D: Expressive 3D Scene Understanding via Object Captioning with Varying Detail" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 151, + 161, + 253, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 161, + 253, + 175 + ], + "spans": [ + { + "bbox": [ + 151, + 161, + 253, + 175 + ], + "type": "text", + "content": "Chandan Yeshwanth" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 273, + 161, + 373, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 273, + 161, + 373, + 175 + ], + "spans": [ + { + "bbox": [ + 273, + 161, + 373, + 175 + ], + "type": "text", + "content": "Dávid Rozenberszki" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 398, + 161, + 455, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 398, + 161, + 455, + 175 + ], + "spans": [ + { + "bbox": [ + 398, + 161, + 455, + 175 + ], + "type": "text", + "content": "Angela Dai" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 226, + 181, + 381, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 226, + 181, + 381, + 194 + ], + "spans": [ + { + "bbox": [ + 226, + 181, + 381, + 194 + ], + "type": "text", + "content": "Technical University of Munich" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 190, + 201, + 416, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 190, + 201, + 416, + 215 + ], + "spans": [ + { + "bbox": [ + 190, + 201, + 416, + 215 + ], + "type": "text", + "content": "https://cy94.github.io/excap3d/" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 57, + 245, + 553, + 371 + ], + "blocks": [ + { + "bbox": [ + 57, + 245, + 553, + 371 + ], + "lines": [ + { + "bbox": [ + 57, + 245, + 553, + 371 + ], + "spans": [ + { + "bbox": [ + 57, + 245, + 553, + 371 + ], + "type": "image", + "image_path": "0991bec62ee5f861c0a6a8add2c523d099186590c1871e0e0e4ff5cb0943fc4c.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 380, + 555, + 402 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 555, + 402 + ], + "type": "text", + "content": "Figure 1. ExCap3D produces descriptions at multiple level of details for objects in an input 3D scan. For each detected object, we generate both object- and part-level of detail captions. The generated captions are consistent across the two levels and contain rich details." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "spans": [ + { + "bbox": [ + 151, + 412, + 200, + 424 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 297, + 714 + ], + "type": "text", + "content": "Generating text descriptions of objects in 3D indoor scenes is an important building block of embodied understanding. Existing methods describe objects at a single level of detail and do not capture fine-grained details of the parts of objects. In order to produce varying levels of detail capturing both coarse object-level information and detailed part-level descriptions, we propose the task of expressive 3D captioning. Given an input 3D scene, the task is to describe objects at multiple levels of detail: a high-level object description, and a low-level description of the properties of its parts. To produce such captions, we present ExCap3D, an expressive 3D captioning model which takes as input a 3D scan, and for each detected object in the scan, generates a fine-grained collective description of the parts of the object, along with an object-level description conditioned on the part-level description. We design ExCap3D to encourage consistency between the multiple levels of descriptions. To enable this task, we generated the ExCap3D Dataset by leveraging a visual-language model (VLM) for multi-view captioning. The ExCap3D Dataset contains captions on the ScanNet++ dataset with varying levels of detail, comprising 190k text descriptions of 34k 3D objects in 947 indoor scenes. Our experiments show that the object- and" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "text", + "content": "part-level details generated by ExCap3D are more expressive than those produced by state-of-the-art methods, with a CIDEr score improvement of " + }, + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "inline_equation", + "content": "124\\%" + }, + { + "bbox": [ + 313, + 413, + 556, + 473 + ], + "type": "text", + "content": " for object- and part-level details respectively. Our code, dataset and models will be made publicly available." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 484, + 394, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 484, + 394, + 496 + ], + "spans": [ + { + "bbox": [ + 314, + 484, + 394, + 496 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 556, + 715 + ], + "type": "text", + "content": "3D indoor scene understanding is central to applications such as augmented reality (AR), virtual reality (VR), and robotics. While detecting 3D objects and their semantics is important for these applications, natural language descriptions of objects enable more wide-ranging possibilities – they can encode more complex information efficiently and, importantly, enable more natural interactions between embodied agents as well as humans and the scene. To produce a representation that facilities a diverse array of applications, such language descriptions must also encompass varying levels of detail: object-level descriptions for coarse identification in the scene, and part-level details for precise manipulation and fine-grained understanding. For instance, a robotic assistant would only need a coarse, high-level description of the objects in a room (e.g., 'a recliner chair') in order to navigate effectively in it, but it would need fine-grained descriptions of the objects (e.g., 'recliner chair with a high padded back, wooden armrests, a soft cushion and an" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21699" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "adjustable footrest') to provide assistive AI descriptions for accessibility aid or to describe the scene to others." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 99, + 296, + 374 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 99, + 296, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 99, + 296, + 374 + ], + "type": "text", + "content": "Current 3D captioning models have made significant progress in describing objects in indoor scenes by detecting objects and generating their individual descriptions at a single level of detail. These methods are trained only for captioning [10, 15, 32] or for multiple tasks [11, 55, 59, 60]. These methods can effectively describe coarse object information in the global context of the scene, but preclude detailed perception of the scene as they make limited use of fine-grained 3D inputs and cannot model object-part hierarchy. They fail to describe highly localized properties of object parts, such as appearance, material, and function. In order to enable more complex perception of a scene requiring these details, we propose to make use of fine-grained 3D inputs and explicitly model object-part hierarchy. We describe objects at multiple levels: an object-level of detail that captures its most prominent aspects (semantic class; possibly size or overall appearance), and a finer-grained part-level of detail that describes the details of elements composing the object (different materials, color, and functions). These two descriptions are complementary to each other and enable different applications depending on the level of detail required. Examples of such multilevel descriptions are shown in Fig. 1 and Fig. 2." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 376, + 295, + 591 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 376, + 295, + 591 + ], + "spans": [ + { + "bbox": [ + 55, + 376, + 295, + 591 + ], + "type": "text", + "content": "To generate such multilevel captions in 3D scans, we introduce ExCap3D, a 3D object captioning model that produces expressive captions at both object- and part-levels of detail using two captioning heads. Examples of the captions generated by our method are shown in Fig. 1. The generated multilevel descriptions are complementary to each other and describe different aspects of the objects, while maintaining a few shared elements for consistency and coherence. The part-level caption generated by ExCap3D collectively describes the different parts of an object. Knowing the descriptions of the parts can enable our model to generate more consistent captions describing the object as a whole. Hence, ExCap3D first produces the part-level details and then uses these to inform the object-level captioner. To further reduce inconsistencies between the two levels of detail in terms of the object semantics and text content, ExCap3D employs additional consistency losses during training." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "To enable this multilevel captioning task, we provide the ExCap3D Dataset, which contains 190k object and partlevel natural language captions of 34k 3D objects from the 947 scenes of the ScanNet++ dataset. We automatically generate object captions for training, using a pretrained vision-language model (VLM) on high-resolution DSLR images along with ground truth semantic information from the ScanNet++ dataset [52]. We employ multi-view aggregation to robustly capture key object- and part-level details. VLMs perform strongly on 2D tasks, but" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 347, + 111 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 347, + 111 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 347, + 111 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 347, + 111 + ], + "type": "image", + "image_path": "d3a94ac01a17ef33178809c2442c6109ed8ed8d9e6019b4cbb18d069ddafed88.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 395, + 72, + 474, + 111 + ], + "blocks": [ + { + "bbox": [ + 356, + 80, + 386, + 110 + ], + "lines": [ + { + "bbox": [ + 356, + 80, + 386, + 110 + ], + "spans": [ + { + "bbox": [ + 356, + 80, + 386, + 110 + ], + "type": "text", + "content": "a brown cushion chair with arms. it is located close to the wall." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 395, + 72, + 474, + 111 + ], + "lines": [ + { + "bbox": [ + 395, + 72, + 474, + 111 + ], + "spans": [ + { + "bbox": [ + 395, + 72, + 474, + 111 + ], + "type": "image", + "image_path": "61347f251bbea1099252ff1792990875aca0fab4d8f536a18f6dcd5773a1ba1a.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 79, + 539, + 111 + ], + "lines": [ + { + "bbox": [ + 479, + 79, + 539, + 111 + ], + "spans": [ + { + "bbox": [ + 479, + 79, + 539, + 111 + ], + "type": "text", + "content": "Q: What is next to the orange armchair? \nA: wall \nA: another orange armchair \nA: coffee table" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 317, + 121, + 394, + 158 + ], + "blocks": [ + { + "bbox": [ + 345, + 114, + 367, + 121 + ], + "lines": [ + { + "bbox": [ + 345, + 114, + 367, + 121 + ], + "spans": [ + { + "bbox": [ + 345, + 114, + 367, + 121 + ], + "type": "text", + "content": "ExCap3D" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 121, + 394, + 158 + ], + "lines": [ + { + "bbox": [ + 317, + 121, + 394, + 158 + ], + "spans": [ + { + "bbox": [ + 317, + 121, + 394, + 158 + ], + "type": "image", + "image_path": "49f2febe57af4f2c95bc88c3a0e2756e939228e6fc6fe63b98f904873dea2149.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 400, + 121, + 542, + 159 + ], + "lines": [ + { + "bbox": [ + 400, + 121, + 542, + 159 + ], + "spans": [ + { + "bbox": [ + 400, + 121, + 542, + 159 + ], + "type": "text", + "content": "Object-level: A three-seater leather couch with dark, smooth, glossy leather, curved and straight edges. \nPart-Level: The black leather sofa has a smooth, glossy surface with visible stitching, a curved armrest, and cushioned seat, with a tufted backrest made of multiple leather panels stitched together, and sturdy structure with visible stitching along the edges." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "lines": [ + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 171, + 555, + 227 + ], + "type": "text", + "content": "Figure 2. Descriptions of objects in existing datasets such as Scan2Cap and ScanQA largely contain relations between objects, and limited local details at a single level. In contrast, the ExCap3D Dataset describes both the object as a whole and as a sum of its parts." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 247, + 553, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 366 + ], + "type": "text", + "content": "struggle with 3D orientation [47] and distances [8]. Hence, our dataset contains large-scale and high-quality object- and part-level captions, in contrast to Scan2Cap [15] and Nr3D [2] with smaller-scale manually annotated spatial relations. A comparison of the multilevel captions from the ExCap3D Dataset with the single-level descriptions from the existing Scan2Cap and ScanQA [3] is shown in Fig. 2. The ExCap3D Dataset describes objects with significantly more detail, while providing two complementary multilevel descriptions of each object." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 367, + 553, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 553, + 426 + ], + "type": "text", + "content": "We show that ExCap3D produces richer and more consistent object- and part-level captions compared to the state-of-the-art captioning methods, originally designed to generate only a single level description. To summarize, our main contributions are:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 427, + 553, + 642 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 314, + 427, + 553, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 427, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 314, + 427, + 553, + 473 + ], + "type": "text", + "content": "- an expressive 3D captioning model that generates captions of objects in 3D indoor scenes at multiple levels of detail: fine-grained part-level details and object-level details" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 475, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 475, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 314, + 475, + 553, + 533 + ], + "type": "text", + "content": "- modeling object-level detail as a sum of its parts' unique appearance and functional attributes by generating object-level captions conditioned on part-level descriptions, using the cross-attention mechanism of the object captioner and the language model's intermediate hidden states" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 314, + 534, + 553, + 582 + ], + "type": "text", + "content": "- regularization of our joint 3D captioning approach through semantic and textual-similarity consistency losses that maintain consistency of the object-level and part-level predicted captions" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 582, + 553, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 582, + 553, + 642 + ], + "spans": [ + { + "bbox": [ + 314, + 582, + 553, + 642 + ], + "type": "text", + "content": "- the ExCap3D Dataset: a large-scale, precise and challenging dataset of 190k descriptive object- and part-level captions of 34k objects in 3D indoor scenes built on the ScanNet++ dataset with fine-grained texture, material, and functional descriptions" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 652, + 400, + 664 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 652, + 400, + 664 + ], + "spans": [ + { + "bbox": [ + 314, + 652, + 400, + 664 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 672, + 437, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 437, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 437, + 685 + ], + "type": "text", + "content": "2.1. 2D Image Captioning" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 714 + ], + "type": "text", + "content": "2D image captioning has been extensively studied in computer vision. DenseCap [29] popularized the task of dense" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21700" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 180 + ], + "type": "text", + "content": "captioning of images, jointly localizing objects, and generating their short descriptions. More recently, COCONutPanCap [19] and Pix2Cap-COCO [53] combined the tasks of fine-grained instance segmentation and detailed captioning. FlexCap [22] and FineCaption [26] can produce descriptions at varying levels of detail given the image and a localized object as input. In contrast, we combine the tasks of detection and multilevel captioning and describe detected objects and their parts." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 202, + 176, + 216 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 176, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 176, + 216 + ], + "type": "text", + "content": "2.2. 3D Scene Captioning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 225, + 296, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 225, + 296, + 513 + ], + "spans": [ + { + "bbox": [ + 56, + 225, + 296, + 513 + ], + "type": "text", + "content": "Scan2Cap [15] introduced the task of 3D dense captioning and a method for context-aware captioning of 3D objects, operating on ScanNet [18] scenes. Since then, several methods have been developed to effectively solve this task by extracting object features and modeling their relations with nearby objects [7, 9, 10, 28, 31, 49, 54]. These methods can accurately describe objects at the scene-level through relations with neighboring objects. Following the trends of utilizing foundation models in 3D reasoning, Cap3D [38] and follow-up works [24, 39] describe isolated 3D objects through multi-view aggregation of VLM-generated captions. More recent works have jointly trained on multiple tasks such as 3D grounding, question answering, and dense captioning by training custom language models from scratch [16, 58] or by fine-tuning pretrained LLMs [11, 25, 27, 50, 57, 60]. However, these works focus on high-level, proximity-based relations between objects at the scene-level, underutilize local features to produce limited object-level details and no part-level details We tackle a similarly inspired problem, and leverage the high-fidelity geometry and RGB captures from ScanNet++ [52] to enable complex, multilevel joint captioning. Moreover, we demonstrate that our joint captioning approach improves both levels of detail." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 535, + 285, + 548 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 535, + 285, + 548 + ], + "spans": [ + { + "bbox": [ + 55, + 535, + 285, + 548 + ], + "type": "text", + "content": "2.3. Part Understanding of 3D Objects in Scenes" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 713 + ], + "type": "text", + "content": "Several methods have focused on part-level understanding of 3D scenes by only relying on explicit geometry [5, 6], images and radiance fields [4, 30] or by using both modalities and mesh-based oversegmentation [46]. A related method for 3D shapes rather than scenes was proposed by Chen et. al. [13], which produces part-level segmentation maps and text descriptions of individual 3D objects at a single, fine-grained level of detail. This produces promising detailed captions, but is designed for isolated objects rather than complex 3D scenes. In contrast, we jointly detect and caption objects in large 3D scenes at multiple levels of detail, ensuring consistency and enabling greater expressivity for downstream tasks." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 370, + 84 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "spans": [ + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": "Problem Formulation The task of expressive captioning is to detect the objects in a 3D scene and describe each object at different levels of detail: object- and part-levels. Given an input 3D scene in the form of a 3D mesh " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "M = (\\mathcal{V},\\mathcal{F})" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " with vertices " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "\\mathcal{V}\\in \\mathbb{R}^{N_v\\times 3}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " and triangular faces " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "f_{i}\\in \\mathcal{F}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " connecting the vertices, we aim to detect the 3D object instances " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "o_i = (b_i,s_i,c_{\\mathrm{obj},i},c_{\\mathrm{part},i})" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " in the scene, characterized by the objects' bounding boxes " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "b_{i}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": ", semantic class " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "s_i" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{obj},i}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{part},i}" + }, + { + "bbox": [ + 313, + 91, + 555, + 212 + ], + "type": "text", + "content": " describing the object in natural language at object- and part-levels of detail, respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 224, + 555, + 345 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 224, + 555, + 345 + ], + "spans": [ + { + "bbox": [ + 313, + 224, + 555, + 345 + ], + "type": "text", + "content": "Overview of ExCap3D First, ExCap3D performs instance segmentation and instance-aware query generation to produce queries containing rich per-object information. Next, the object- and part-level captioning models with shared information generate two separate text captions, conditioned on individual instance-aware queries and the corresponding features from the instance segmentation backbone. Finally, semantic and whole-text consistency losses are applied to the hidden states of the captioning models to regularize the consistency of the predicted captions." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 350, + 460, + 363 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 350, + 460, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 350, + 460, + 363 + ], + "type": "text", + "content": "3.1. 3D Instance Segmentation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": "Given the mesh of the 3D scene " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " as input, the instance segmentation model produces a fixed number of refined instance-aware queries " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "Q_{r}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " corresponding to the " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "N_{q}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " objects detected in the scene, and their binary instance masks on the voxelized 3D input " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "I\\in \\{0,1\\}^{N_q\\times N_{\\mathrm{vox}}}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": ". The instance queries are a compact representation of the 3D features of the objects as well as their semantics. The instance segmentation model also produces dense features " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "F\\in \\mathbb{R}^{N_{\\mathrm{vox}}\\times D}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " on the voxels. The refined and instance-aware queries " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "Q_{r}" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " and 3D voxel features " + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 313, + 367, + 556, + 498 + ], + "type": "text", + "content": " are used to inform the per-instance captioning in the next step." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": "The input mesh is first voxelized into a voxel grid " + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "inline_equation", + "content": "V \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times 6}" + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": " containing the voxel coordinates and RGB colors. We use the Mask3D [45] transformer-based model for instance segmentation. It consists of a 3D sparse convolutional UNet backbone to encode the voxelized 3D scene, which outputs features " + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "inline_equation", + "content": "F \\in \\mathbb{R}^{N_{\\mathrm{vox}} \\times D}" + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": " over the input voxels. A parallel mask module takes as input query vectors " + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "inline_equation", + "content": "Q \\in \\mathbb{R}^{N_q \\times D_q}" + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": " denoting the " + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "inline_equation", + "content": "N_q" + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": " instances in the scene. The queries are refined iteratively through several transformer encoder layers, with cross-attention to the 3D backbone features. The output instance-aware query vectors contain rich representations of the object instances in the 3D scene and are used to predict the instance masks " + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "inline_equation", + "content": "I \\in \\{0,1\\}^{N_q \\times N_{\\mathrm{vox}}}" + }, + { + "bbox": [ + 313, + 499, + 556, + 666 + ], + "type": "text", + "content": " and semantic class of each object." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 672, + 539, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 539, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 539, + 685 + ], + "type": "text", + "content": "3.2. Joint Captioning with Information Sharing" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "ExCap3D contains two separate captioners for the object- and part-levels of details. These are transformer-based lan" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21701" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 531, + 242 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 531, + 242 + ], + "type": "image", + "image_path": "411fdae6ebb29bf9fbb66498e538d4530434426e1f9ec170b0befa1a2ea82bef.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "lines": [ + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 54, + 250, + 555, + 285 + ], + "type": "text", + "content": "Figure 3. Overview of our ExCap3D captioning method. We predict 3D instances in the input 3D scene using Mask3D [45], then predict object-level and part-levels of detail using two separate captioning heads. The object-level captions are further constrained by the part-level captioner's hidden states. Semantic- and textual consistency losses are applied to ensure the overall consistency of both predicted captions." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": "guage models " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{obj}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{part}}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": " that produce the object-level caption " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{obj},o}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": ", part-level caption " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{part},o}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": ". They are conditioned on the instance-aware queries " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "Q_{r}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": " and the dense voxel features " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": ", and are trained for next-token prediction. The language model captioners take as input a sequence of previously predicted tokens " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "t_1 \\ldots t_{i-1}" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": " and decode the next token " + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 54, + 304, + 296, + 389 + ], + "type": "text", + "content": " autoregressively as" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 102, + 401, + 247, + 414 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 401, + 247, + 414 + ], + "spans": [ + { + "bbox": [ + 102, + 401, + 247, + 414 + ], + "type": "interline_equation", + "content": "p \\left(t _ {i} \\mid t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(t _ {1} \\dots t _ {i - 1}\\right),", + "image_path": "1e4df0363ac328ee30f9169ba8e8c0eb8d84b5090626dcfeffa3c2972699849c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": " is the language model. We provide two sources of rich instance-level information to the captioner: the instance-aware refined query output by the instance-segmentation model, and voxel features from 3D UNet backbone " + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": ". The refined query vector " + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "inline_equation", + "content": "Q_{r,o}" + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": " corresponding to an object " + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": " is first projected with a linear layer to generate the caption-aware query " + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "inline_equation", + "content": "Q_{c,o} \\in \\mathbb{R}^{D_{\\mathrm{caption}}}" + }, + { + "bbox": [ + 54, + 419, + 296, + 506 + ], + "type": "text", + "content": " for that object" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 129, + 514, + 222, + 528 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 129, + 514, + 222, + 528 + ], + "spans": [ + { + "bbox": [ + 129, + 514, + 222, + 528 + ], + "type": "interline_equation", + "content": "Q _ {c, o} = \\Phi_ {\\mathrm {q u e r y}} (Q _ {r, o}),", + "image_path": "b0a4abce6501c3b91fce61f6891c95378045cf43f3730d7e2447fca68b3f7653.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{caption}}" + }, + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "text", + "content": " is the embedding dimension of the language model. " + }, + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "inline_equation", + "content": "Q_{c,o}" + }, + { + "bbox": [ + 54, + 536, + 296, + 597 + ], + "type": "text", + "content": " contains localized information required to generate the caption for an object, such as its semantic class and its high-level features, and is used as the initial token to start the caption generation as" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 80, + 619, + 270, + 632 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 80, + 619, + 270, + 632 + ], + "spans": [ + { + "bbox": [ + 80, + 619, + 270, + 632 + ], + "type": "interline_equation", + "content": "p \\left(t _ {i} \\mid Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right) = \\Psi \\left(Q _ {c, o}, t _ {1} \\dots t _ {i - 1}\\right).", + "image_path": "5523bf657b168922f51e66971078b180aced7a1f2800053e0072a92c4d5c3d28.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "text", + "content": "Additionally, we inject fine-grained information by using the captioner's cross-attention layers to attend to fine-grained contextual features from the 3D backbone " + }, + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "text", + "content": " corresponding to the object being captioned, after a linear projection to " + }, + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "inline_equation", + "content": "D_{\\mathrm{caption}}" + }, + { + "bbox": [ + 54, + 641, + 296, + 715 + ], + "type": "text", + "content": ". To reduce the computational complexity of the cross-attention operation, we aggregate features within" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": "precomputed segments [23] on the mesh surface, giving the context features " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "S_{o} \\in \\mathbb{R}^{n_{s,o} \\times D_{\\mathrm{caption}}}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "n_{s,o}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " is the number of segments in " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "I_{o}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ", the predicted mask for object " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ". The cross-attention operation is " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "A = \\operatorname{softmax}(qk^T / \\sqrt{d})v" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " where queries " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " and values " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " are the hidden states of " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "\\Psi" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ", and the keys " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " are " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "S_{o}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": ". The object and part captioners use context features " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{obj},o}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{part},o}" + }, + { + "bbox": [ + 313, + 304, + 554, + 388 + ], + "type": "text", + "content": " respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": "Importantly, to enable the object-level captioner to be informed by the part-level captioner, we make use of the last layer hidden states " + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "inline_equation", + "content": "[h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},i}]" + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": " of the part-level captioner " + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "inline_equation", + "content": "\\Psi_{\\mathrm{part}}" + }, + { + "bbox": [ + 313, + 388, + 554, + 437 + ], + "type": "text", + "content": ". We use a linear projection" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 354, + 448, + 514, + 462 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 354, + 448, + 514, + 462 + ], + "spans": [ + { + "bbox": [ + 354, + 448, + 514, + 462 + ], + "type": "interline_equation", + "content": "H _ {\\text {p a r t}, o} = \\Phi_ {\\text {h i d d e n}} ([ h _ {\\text {p a r t}, 1} \\dots h _ {\\text {p a r t}, i} ])", + "image_path": "32cb78f85dd90b15978cbd79ed9cdf2be75645e48371bbfdd4345ed198c48ae1.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "spans": [ + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": "and then concatenate " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "H_{\\mathrm{part},o}" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": " to the contextual features of the object captioner " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{obj},o}" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": " for object " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": ". The object-level caption for object " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": " is finally informed by the caption-aware query " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "Q_{c,o}" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": " as well as features " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "[H_{\\mathrm{part},o};S_{\\mathrm{obj},o}]" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": ". Since different objects in a batch have different numbers of predicted segments, we zero-pad " + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "inline_equation", + "content": "S_{\\mathrm{obj},o}" + }, + { + "bbox": [ + 313, + 467, + 554, + 550 + ], + "type": "text", + "content": " to the maximum context length in the batch." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 550, + 554, + 588 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 550, + 554, + 588 + ], + "spans": [ + { + "bbox": [ + 313, + 550, + 554, + 588 + ], + "type": "text", + "content": "The two caption models are trained with cross-entropy (CE) loss against the object- and part-level ground truth (GT) captions respectively, giving the loss" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 351, + 598, + 517, + 612 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 598, + 517, + 612 + ], + "spans": [ + { + "bbox": [ + 351, + 598, + 517, + 612 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c a p t i o n}} = \\mathcal {L} _ {\\text {o b j - c a p t i o n}} + \\mathcal {L} _ {\\text {p a r t - c a p t i o n}}", + "image_path": "1d0db58834ea95080c7802dabbb09acf231b7cf6eb0fbb7efa5d26dae1aae774.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "spans": [ + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": "The outputs of this stage are the object-level caption " + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{obj},o}" + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": ", part-level caption " + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "inline_equation", + "content": "c_{\\mathrm{part},o}" + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": " and the corresponding language model hidden states " + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj}} = [h_{\\mathrm{obj},1}\\dots h_{\\mathrm{obj},i}]" + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{part}} = [h_{\\mathrm{part},1}\\dots h_{\\mathrm{part},j}]" + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": " for each object " + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 313, + 616, + 554, + 667 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 672, + 531, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 531, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 531, + 685 + ], + "type": "text", + "content": "3.3. Semantic and Textual Consistency Losses" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "While the object-level of detail is informed by the part-level of detail, supervising these two captioners with different GT" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21702" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "captions may lead to inconsistencies in the predicted captions due to the different GT distributions of the two levels of detail. Hence, we use additional constraints in the form of consistency losses over the predicted captions. We encourage semantic consistency for both captions to refer to the same object semantics, such as its fine-grained object class and material, and textual consistency for the captions to be similar to each other. Since consistency losses cannot be applied directly on the generated text due to the text tokenizer being non-differentiable, we instead use the language model hidden states " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{part}}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " from the last captioner layers as a proxy." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": "Semantic Consistency To promote both levels of captions to contain similar semantic information, we use a semantic consistency loss. We do this by projecting the hidden states " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{part}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " into a lower dimension to retain only meaningful semantic information over the whole caption. We then pass these sequences independently through object- and part-level caption projection transformers " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{obj,sem}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{part,sem}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " to obtain new embedded sequences, and these aggregate into single feature vectors representing the overall semantics of the predicted captions " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj,sem}} \\in \\mathbb{R}^{D_{ff}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{part,sem}} \\in \\mathbb{R}^{D_{ff}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " respectively, where " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "D_{ff}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " is the feed-forward dimension of " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": ". These are again independently projected into " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "N_{\\mathrm{sem}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " fine-grained latent classes using linear layers to get class probabilities " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "sem_{\\mathrm{obj}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "inline_equation", + "content": "sem_{\\mathrm{part}}" + }, + { + "bbox": [ + 55, + 226, + 296, + 418 + ], + "type": "text", + "content": ". The semantic consistency loss is a symmetric sum [14] of cross-entropy (CE) losses" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 86, + 441, + 266, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 441, + 266, + 470 + ], + "spans": [ + { + "bbox": [ + 86, + 441, + 266, + 470 + ], + "type": "interline_equation", + "content": "\\begin{array}{r} \\mathcal {L} _ {\\mathrm {s e m a n t i c}} = C E (s e m _ {\\mathrm {o b j}}, S G (s e m _ {\\mathrm {p a r t}})) + \\\\ C E (s e m _ {\\mathrm {p a r t}}, S G (s e m _ {\\mathrm {o b j}})), \\end{array}", + "image_path": "0263876024f6468391f142253e266c531dbf3c1b44c23aefd5cba71d7da941b9.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 479, + 217, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 217, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 217, + 491 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 479, + 217, + 491 + ], + "type": "inline_equation", + "content": "SG" + }, + { + "bbox": [ + 55, + 479, + 217, + 491 + ], + "type": "text", + "content": " is the stop-gradient operator." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": "Textual Consistency The object- and part-level captions are expected to slightly overlap, since the salient features of an object are reflected in its object-level caption. We encourage this similarity through the similarity of the whole sequences of object-level and part-level captions. Similar to the semantic consistency loss, we use projectors " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{obj},\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\Phi_{\\mathrm{part},\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": " to project the language model hidden states into " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{obj},\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "h_{\\mathrm{part},\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": ", and aggregate over the sequence of tokens to get single feature vectors " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\overline{h}_{\\mathrm{obj},\\mathrm{text}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "\\overline{h}_{\\mathrm{obj},\\mathrm{part}}" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": ". The text-consistency loss is the distance " + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "inline_equation", + "content": "d(.,.)" + }, + { + "bbox": [ + 55, + 501, + 296, + 632 + ], + "type": "text", + "content": " between these two feature vectors" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 105, + 645, + 245, + 660 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 645, + 245, + 660 + ], + "spans": [ + { + "bbox": [ + 105, + 645, + 245, + 660 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {t e x t u a l}} = d (\\overline {{h}} _ {\\mathrm {o b j , t e x t}}, \\overline {{h}} _ {\\mathrm {p a r t , t e x t}}).", + "image_path": "41039518446bc9807387ba6fd180eedcfc2da9ba1feebf6f6c39ecbb9550f560.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 666, + 295, + 690 + ], + "type": "text", + "content": "The final loss used to optimize ExCap3D is a weighted sum of the individual losses" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 82, + 701, + 268, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 82, + 701, + 268, + 714 + ], + "spans": [ + { + "bbox": [ + 82, + 701, + 268, + 714 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = w _ {1} \\mathcal {L} _ {\\text {c a p t i o n}} + w _ {2} \\mathcal {L} _ {\\text {s e m a n t i c}} + w _ {3} \\mathcal {L} _ {\\text {t e x t u a l}}", + "image_path": "87aa5f9984d3115e589dd0cb5e9d6205d295d7fe60abf6fb84bb76a663af19ef.jpg" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 322, + 70, + 548, + 156 + ], + "blocks": [ + { + "bbox": [ + 322, + 70, + 548, + 156 + ], + "lines": [ + { + "bbox": [ + 322, + 70, + 548, + 156 + ], + "spans": [ + { + "bbox": [ + 322, + 70, + 548, + 156 + ], + "type": "table", + "html": "
Dataset#Desc#Cls#ObjsLoD
NR3D [2]41k765.8ks + o
Multi3DRefer [56]55k26510.5ks + o
ScanQA [3]35k3709.5ks + o
Scan2Cap [15]46k2659.9ks + o
Ours190k2k34.7ko + p
", + "image_path": "06fc61d8856315f8a3e6feebb09466c0397b22eddf02d7ba7615639d65749500.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 163, + 554, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 554, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 554, + 251 + ], + "type": "text", + "content": "Table 1. Comparison of the ExCap3D Dataset with existing datasets that describe objects in 3D indoor scenes. We compare the number of descriptions (#Desc), number of unique object classes described (#Cls), number of objects described (#Objs) and level of detail (LoD, s = scene, o = object, p = part). Our descriptions have greater object detail and introduce part-level details, while scaling up the number of descriptions, and semantic classes and objects described." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 263, + 554, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 554, + 348 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 554, + 348 + ], + "type": "text", + "content": "Implementation Details Predicted instances are matched with GT instances using Hungarian matching [34], and the predicted captions are supervised by the captions of the matched instances. We train both the instance segmentation and captioning stages with a batch size of 6 scenes. In the captioning stage, one batch that is input to the captioner contains all objects from scenes in the batch." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 358, + 422, + 373 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 358, + 422, + 373 + ], + "spans": [ + { + "bbox": [ + 313, + 358, + 422, + 373 + ], + "type": "text", + "content": "4. ExCap3D Dataset" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 379, + 555, + 606 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 555, + 606 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 555, + 606 + ], + "type": "text", + "content": "Existing datasets for captioning objects in 3D indoor scenes contain captions at a single level of detail, that combines both object properties such as color and texture, as well as spatial and functional relations to other objects in the scene. To enable the task of captioning at both the object- and part-levels of detail, we created the ExCap3D Dataset based on the ScanNet++ [52] dataset of 3D indoor scans. The data is generated in a scalable and fully automated manner by leveraging a VLM that takes an RGB image crop of an object and a text prompt as input, and outputs a description of the object. The objects to be captioned are identified using the ground truth semantic instance annotations provided by ScanNet++. We prompt the VLM with text prompts to generate specific information about an object's color, texture, material and similar properties, and leverage multiview reasoning to obtain reliable final caption generation for ScanNet++. A comparison of the ExCap3D Dataset with existing object_caption and other 3D object-text datasets is shown in Tab. 1." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 612, + 495, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 612, + 495, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 612, + 495, + 625 + ], + "type": "text", + "content": "4.1. Object Level of Detail Captioning" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 629, + 554, + 715 + ], + "type": "text", + "content": "We utilize the ground truth 3D instance and semantic annotations in the ScanNet++ dataset to generate the ExCap3D Dataset. We project the 3D annotations onto the high-resolution DSLR images that are registered with the 3D scan, to obtain a corresponding 2D bounding box of each object in each DSLR image. We use this 2D mask to crop a rectangular region containing the object from the image," + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21703" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 294, + 144 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 294, + 144 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 294, + 144 + ], + "type": "image", + "image_path": "a2b3e4384a1a5d9a5f77f6e0664af3c55f80bba9c00984515ea93d97b166a31c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 149, + 296, + 217 + ], + "lines": [ + { + "bbox": [ + 55, + 149, + 296, + 217 + ], + "spans": [ + { + "bbox": [ + 55, + 149, + 296, + 217 + ], + "type": "text", + "content": "Figure 4. Generation of ExCap3D Dataset. We use the ground truth 3D semantics in ScanNet++, project them onto multiview DSLR images and obtain descriptions of the image crops " + }, + { + "bbox": [ + 55, + 149, + 296, + 217 + ], + "type": "inline_equation", + "content": "c_{i,j}" + }, + { + "bbox": [ + 55, + 149, + 296, + 217 + ], + "type": "text", + "content": " using a VLM. For parts crops, we use pseudo-ground truth from MaskClustering. Finally, we summarize the captions from different views using an LLM." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": "after expanding the bounding box dimensions by " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": ". To maximize the RGB information that is input to the VLM, we pick the top 3 images that cover at least " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "10\\%" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " the mesh vertices corresponding to the object mask. The VLM generates " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " descriptions for each view of the object, giving outputs " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "c_{i,j}" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " from the " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "i^{\\mathrm{th}}" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " view. We use an instruction-tuned large language model (LLM) to obtain the " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " object-level caption " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "\\hat{c}_{\\mathrm{obj},j}" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " by combining the " + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "inline_equation", + "content": "j^{th}" + }, + { + "bbox": [ + 55, + 236, + 296, + 333 + ], + "type": "text", + "content": " caption from each view." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 338, + 226, + 351 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 226, + 351 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 226, + 351 + ], + "type": "text", + "content": "4.2. Part Level of Detail Captioning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "text", + "content": "We produce pseudo-ground truth 3D part segmentation data using a modified version of MaskClustering [51] that favours smaller segments over larger ones, along with SAM [33]. SAM is used to generate initial 2D segments on the DSLR images. MaskClustering combines these 2D segments to produce coherent 3D part segments. We associate these 3D parts with the objects in ScanNet++ based on an intersection-over-union (IoU) threshold. The rest of the part captioning process largely follows that of object captioning. We generate one caption " + }, + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "text", + "content": " in each view, for each part, and then combine the descriptions of all the parts into a single part-level caption " + }, + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "inline_equation", + "content": "\\hat{c}_{\\mathrm{part}}" + }, + { + "bbox": [ + 55, + 356, + 296, + 582 + ], + "type": "text", + "content": " for each object using the LLM. While the pseudo-masks at the part-level obtained from MaskClustering can detect parts of long-tail and small objects in the ScanNet++ data, these masks are not as reliable as the manually annotated object GT masks in ScanNet++. Hence, we aggregate the captions of different parts into a single caption containing overall part-level details for an object." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": "Implementation Details We use Llava 1.6 [37] (7B parameters) with a Mistral backbone as the VLM, and instruction-tuned Llama 3.1 [21] (8B parameters) for the LLM. To ensure the quality of the captions for evaluation, we manually removed incorrect captions in the test set. Due to the robust multiview aggregation of captions, " + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "inline_equation", + "content": "93.21\\%" + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": " of the captions were found to be accurate." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 670, + 137, + 683 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "We evaluate caption generation on the 84 instance class categories of the ScanNet++ benchmark [52]. We use the" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 537, + 84 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 537, + 84 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 537, + 84 + ], + "type": "text", + "content": "train/val/test splits of the semantic tasks of ScanNet++." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 93, + 427, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 93, + 427, + 105 + ], + "spans": [ + { + "bbox": [ + 313, + 93, + 427, + 105 + ], + "type": "text", + "content": "5.1. Evaluation Metrics" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 111, + 554, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 111, + 554, + 171 + ], + "spans": [ + { + "bbox": [ + 313, + 111, + 554, + 171 + ], + "type": "text", + "content": "We adopt the commonly used CIDEr [44], ROUGE [36], METEOR [20] and BLUE-4 [41] caption evaluation metrics used by other 3D captioning methods [10, 15, 27, 58, 60]. Metrics are reported separately for the object- and part-level details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 181, + 381, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 181, + 381, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 181, + 381, + 192 + ], + "type": "text", + "content": "5.2. Baselines" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 199, + 555, + 354 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 199, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 313, + 199, + 555, + 354 + ], + "type": "text", + "content": "We compare our method against six state-of-the-art captioning methods. Cap3D [38] captions individual objects using pretrained BLIP2 [35] and GPT-4 [1]. LLaVA-3D [57] aggregates multiview 2D features and uses these as input to an LLM, and is evaluated zero-shot from a pretrained model. D3Net [9] uses a PointGroup sparse convolutional backbone with embedded object relations, and a GRU for text generation. Vote2Cap-DETR [10] tokenizes the 3D input, and generates captions using GPT2 [42]. Vote2Cap-DETR++ [12] additionally uses decoupled detection and caption queries. PQ3D [60] uses promptable queries for regions of interest, and generates captions with T5-Small [43]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 355, + 554, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 554, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 554, + 427 + ], + "type": "text", + "content": "We train on ScanNet++ [52] for bounding box detection (Vote2Cap and D3Net) or instance segmentation (PQ3D and ExCap3D). Since the baselines were designed for single level of detail captioning, we train them separately on the object- and part-level captions, while ExCap3D is trained jointly on both levels of detail." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 315, + 437, + 552, + 536 + ], + "blocks": [ + { + "bbox": [ + 315, + 437, + 552, + 536 + ], + "lines": [ + { + "bbox": [ + 315, + 437, + 552, + 536 + ], + "spans": [ + { + "bbox": [ + 315, + 437, + 552, + 536 + ], + "type": "table", + "html": "
MethodObject-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]5.57.67.70.0
D3Net [9]6.75.46.70.4
LLaVA-3D [57]0.06.016.30.0
Vote2Cap-DETR [10]13.312.917.21.0
Vote2Cap-DETR++ [12]10.611.915.71.1
PQ3D [60]27.911.612.61.9
ExCap3D (Ours)32.916.718.01.2
", + "image_path": "5140244823f4f24060d36679e14805b2c9ba6fd5558414533c669021dea6f29e.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "type": "table", + "bbox": [ + 316, + 538, + 552, + 634 + ], + "blocks": [ + { + "bbox": [ + 316, + 538, + 552, + 634 + ], + "lines": [ + { + "bbox": [ + 316, + 538, + 552, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 538, + 552, + 634 + ], + "type": "table", + "html": "
MethodPart-level Details
CIDErROUGEMETEORBLEU-4
Cap3D [38]2.111.49.80.0
D3Net [9]7.77.67.31.3
LLaVA-3D [57]0.613.015.50.0
Vote2Cap-DETR [10]7.720.321.73.9
Vote2Cap-DETR++ [12]19.519.919.83.8
PQ3D [60]14.516.015.34.3
ExCap3D (Ours)32.821.921.14.8
", + "image_path": "5598622341d9cccaacbd2b61aeb5e4dd79135a357e3e87ae041fe8674d754b0d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "table_body" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 554, + 696 + ], + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 696 + ], + "type": "text", + "content": "Table 2. Comparison of object-level and part-level captioning performance between ExCap3D and baselines. The second-best result is underlined. ExCap3D improves CIDEr scores significantly through joint captioning and consistency losses. All metrics are evaluated at an IoU threshold of 0.5" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21704" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 67, + 69, + 542, + 669 + ], + "blocks": [ + { + "bbox": [ + 67, + 69, + 542, + 669 + ], + "lines": [ + { + "bbox": [ + 67, + 69, + 542, + 669 + ], + "spans": [ + { + "bbox": [ + 67, + 69, + 542, + 669 + ], + "type": "table", + "html": "
Input 3D Scene\nDetected 3D Objectwhite, rectangular, smooth, glossy, unk*white, rectangular, smooth, glossy, unk*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with white, smooth, glossy surface, made of wood or laminate, metal*
D3Net\n(Chen et al. 21)Object-levelwhite, smooth, possibly plastic or metal*a white, smooth, possibly plastic or metal*a black, curved backrest with a smooth texture, attached to a black metal frame with a smooth texture.a white, smooth, possibly plastic or metal*
Part-level
Vote2Cap-DETR\n(Chen et al. 23)Object-levelBlack, rectangular, smooth, glossy, plastic or metal computer tower. Black, smooth, plastic or metal, rectangular, cylindrical, or cylWhite, rectangular cabinet with a smooth, glossy finish, made of wood or laminate,like composite, flat top, and a handle. Dark woodA modern chair with a curved backrest, smooth glossy finish, dark fabric, and a chrome metal frame.. Modern, vibrant red, smooth,Dark-colored, leather-like, smooth, glossy, curved, L-shaped sofa. L-shaped sofa with a sturdy frame. Black,
Part-levelA black, rectangular trash can with a smooth, possibly plastic or metal surface, a black, cylindrical trash can with a smooth, possibly plastic surface,*A rectangular file cabinet with a flat top and bottom, made of light-colored wood with a smooth finish and visible grain pattern. It has a flat top surface and a slightly curved top edge. A rectangular drawer with a flat top and bottomThe chair has a curved, dark-colored backrest with a smooth, glossy finish, a curved, dark-colored backrest with a smooth texture, and a metal frame. The backrest is a curved, smooth, glossy piece of darkThe sofa has a curved, padded backrest with a smooth, plush texture, a curved, padded armrest with a smooth, glossy finish.* The backrest is a curved
PQ3D\n(Zhu et al. 24)Object-levelblack, rectangular, smooth, glossy, plastic or metal*white, rectangular, smooth, glossy, or metal.*wooden chair with light brown color, smooth texture, and curved backrest.rectangular table with a smooth, glossy, glossy surface, and a light-colored, possibly wood or wood-like material
Part-levela black, rectangular computer tower with a flat top and bottom, curved top, and a flat top, made of a smooth, possibly plastic or metal materialthe cabinet has a light-colored, smooth finish, possibly wood or composite, with a flat top and bottom*, a slightly curved top edge, and a slightly angled frontthe chair has a dark-colored, smooth, possibly plastic or metal backrest with a curved shape and a uniform texture.the sofa has a curved, cushioned backrest with a smooth, glossy finish, a curved, padded armrest with a smooth, glossy finish*
ExCap3D\n(Ours)Object-levelBlack, glossy, rectangular computer tower with a smooth textureWooden cabinet with light brown color, smooth texture, and rectangular shape.Wooden chair with light brown color, smooth texture, and rectangular shape.Black leather sofa with smooth texture and rectangular shape with rounded corners.
Part-levelA black computer tower with a textured surface, vents, and a rectangular drive bay with a textured surface.The cabinet has a light-colored, smooth finish, possibly made of wood or a wood-like composite, with a flat top and bottom, a series of drawers with metal handles.The chair has a light brown, smooth, curved backrest made of wood or wood-like composite, attached to a metal frame.The sofa has a black leather backrest with a smooth, leather-like texture, a curved armrest with a smooth, possibly leather-like material, and a rectangular cushion with a smooth texture.
Ground TruthObject-levelBlack, rectangular computer tower with a flat top, smooth, polished surface, and a tapered base.Light-colored, rectangular, boxy cabinet with smooth, glossy, synthetic surface.A chair with a curved backrest, flat seat, smooth, light brown, wooden surface, and metal legs.Dark-colored leather sofa with curved shape, plush cushions, and smooth texture.
Part-levelA black computer tower with a black power supply unit attached, both with smooth, glossy finishes.The cabinet has a light brown, smooth wooden surface, glass front, three wooden drawers with silver, curved handles, and a dark-colored, rectangular door.A curved, light-colored wood or wood-like composite backrest with smooth or matte finish, supported by a metal frame.The sofa has a curved, padded armrest, a cushion, and a backrest, all upholstered in dark, rich brown or black leather with a smooth, shiny texture.
", + "image_path": "9a79e919983cbbb110090eddc8107076f497abe89e22bcc416a097b6c7ee5129.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 681, + 555, + 726 + ], + "lines": [ + { + "bbox": [ + 55, + 681, + 555, + 726 + ], + "spans": [ + { + "bbox": [ + 55, + 681, + 555, + 726 + ], + "type": "text", + "content": "Figure 5. Qualitative evaluation on ScanNet++ [52] scenes, in comparison with D3Net [9], Vote2CAP-DETR [10], and PQ3D [60]. Predicted text from methods are denoted by color to indicate correct and incorrect generated phrases compared to the ground truth. Our method produces more consistent and detailed captions at both the object- and part-levels of detail. * indicates truncated output where the underlined phrase was predicted repeatedly." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21705" + } + ] + } + ], + "index": 2 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 223, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 223, + 84 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 223, + 84 + ], + "type": "text", + "content": "5.3. Comparison to State of the Art" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "spans": [ + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "text", + "content": "We compare with state-of-the-art captioning methods in Tab. 2. ExCap3D outperforms all baselines on the CIDEr score, which aligns the most closely with human perception, improving upon the highest performing baseline PQ3D by a margin of " + }, + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "inline_equation", + "content": "124\\%" + }, + { + "bbox": [ + 54, + 89, + 294, + 209 + ], + "type": "text", + "content": " for object- and part-level details respectively. We also outperform the baselines on the ROUGE and METEOR scores of object-level details, and ROUGE scores of part-level details. Our joint prediction of captions at different levels of detail enables more robust, accurate descriptions." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 210, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 210, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 210, + 296, + 365 + ], + "type": "text", + "content": "Qualitative comparisons are shown in Fig. 5. ExCap3D produces complete and detailed descriptions at both the object- and part-levels of detail, while the baseline methods often omit details in both levels. Cap3D's BLIP-2 captioner can produce limited details, and is easily distracted by background objects. LLaVA-3D produces general captions with few object-specific details, and has limited object localization due to a single location prompt as input. D3Net uses a limited GLOVE embedding of text which cannot produce detailed output descriptions. Vote2Cap-DETR, Vote2Cap-DETR++ and PQ3D interpret 3D scenes more coarsely, which can result in hallucination of incorrect fine-grained details." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 374, + 156, + 386 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 374, + 156, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 374, + 156, + 386 + ], + "type": "text", + "content": "5.4. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 392, + 295, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 392, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 392, + 295, + 453 + ], + "type": "text", + "content": "We ablate the different components of our method and show the results in Tab. 3. As a baseline, we use two independent captioning heads and train with only " + }, + { + "bbox": [ + 55, + 392, + 295, + 453 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{caption}}" + }, + { + "bbox": [ + 55, + 392, + 295, + 453 + ], + "type": "text", + "content": ". Each component of our method is added separately to the baseline to show its efficacy." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 594 + ], + "type": "text", + "content": "Enforcing multilevel consistency improves caption quality. Semantic consistency primarily improves part-level details while textual consistency primarily improves object-level details as shown in Tab. 3, which demonstrates the complementary nature of these two losses. As the object-level captions typically contain a stronger notion of global object semantics compared to the part-level caption, the latter improves in quality with semantic consistency. The object-level captions also receive fine-grained cues through textual similarity from the part-level details which helps to refine the global description of the object." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 603, + 295, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 603, + 295, + 723 + ], + "spans": [ + { + "bbox": [ + 55, + 603, + 295, + 723 + ], + "type": "text", + "content": "Part-object captioner information sharing improves performance Adding part " + }, + { + "bbox": [ + 55, + 603, + 295, + 723 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 603, + 295, + 723 + ], + "type": "text", + "content": " object information sharing by conditioning object captions on the part-level of detail improves both levels of captions as shown in Tab. 3. This is because object details are improved when they are informed by part captions which contain aggregated details that can produce a better global description of the object. At the same time, the part-level of detail improves as it is guided to produce consistent descriptions of parts that can be useful for object-level captioning." + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 556, + 156 + ], + "blocks": [ + { + "bbox": [ + 315, + 70, + 556, + 156 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 556, + 156 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 556, + 156 + ], + "type": "table", + "html": "
MethodObject-LevelPart-level
Separate Models29.717.9
w/ Semantic Consistency30.223.4
w/ Textual Consistency32.218.6
w/ Part → Obj. Inf. Sharing34.824.0
Full Model (Ours)32.932.8
", + "image_path": "793f587b2326b2508e2857fc7a65c5df7954faa12b5afa3245041484fc1af2c8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "lines": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 313, + 163, + 553, + 208 + ], + "type": "text", + "content": "Table 3. Ablation study of the different components of our method. Performance is reported using CIDEr for object-level and part-level captions. Each ablation is applied independently to the baseline with separate models." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 228, + 555, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 555, + 371 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 555, + 371 + ], + "type": "text", + "content": "Consistency and information sharing are complementary to each other The semantic and textual consistency losses act independently of the part " + }, + { + "bbox": [ + 313, + 228, + 555, + 371 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 313, + 228, + 555, + 371 + ], + "type": "text", + "content": " object information sharing mechanism, and hence our full model improves significantly over the individual components, particularly in the part-level of detail as shown in Tab. 3. Intuitively, this is because the consistency losses model only two kinds of consistency over lower-dimension embeddings, while sharing language model hidden states enables the object-level captioner to directly condition on the internal high-dimension representation of the part-level captions with fine-grained features." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 377, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 556, + 498 + ], + "type": "text", + "content": "Limitations While our approach generates expressive captioning for 3D scans, various limitations remain. We use two captioning models, with one conditioned on the hidden states of the other through cross-attention. This could limit the amount of information shared between the captioners. Our method relies on a sparse convolutional backbone [17] which is limited to a resolution of about " + }, + { + "bbox": [ + 313, + 377, + 556, + 498 + ], + "type": "inline_equation", + "content": "2\\mathrm{cm}" + }, + { + "bbox": [ + 313, + 377, + 556, + 498 + ], + "type": "text", + "content": ". This could restrict captioning performance on small or thin objects. Similarly, the class imbalance in the ScanNet++ dataset could restrict performance on long-tail classes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 507, + 388, + 520 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 527, + 555, + 635 + ], + "type": "text", + "content": "We introduced ExCap3D, a method for expressive dense captioning of 3D scenes at multiple object- and part-levels of detail. We showed that describing 3D objects as the sum of their parts while maintaining consistency between the multiple levels of descriptions leads to more accurate descriptions aligned with human perception. We hope that our method and the accompanying ExCap3D Dataset open up new possibilities for a holistic, expressive understanding of 3D scans." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 646, + 411, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 646, + 411, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 646, + 411, + 658 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 715 + ], + "type": "text", + "content": "This project was supported by the ERC Starting Grant SpatialSem (101076253), and the German Research Foundation (DFG) Grant \"Learning How to Interact with Scenes through Part-Based Understanding.\"" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21706" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 203 + ], + "type": "text", + "content": "[2] Panos Achlioptas, Ahmed Abdelreheem, Fei Xia, Mohamed Elhoseiny, and Leonidas J. Guibas. ReferIt3D: Neural listeners for fine-grained 3d object identification in real-world scenes. In 16th European Conference on Computer Vision (ECCV), 2020. 2, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "spans": [ + { + "bbox": [ + 62, + 205, + 295, + 258 + ], + "type": "text", + "content": "[3] Daichi Azuma, Taiki Miyanishi, Shuhei Kurita, and Motoaki Kawanabe. Scanqa: 3d question answering for spatial scene understanding. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19129-19139, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 261, + 295, + 315 + ], + "type": "text", + "content": "[4] Yash Bhalgat, Iro Laina, João F Henriques, Andrew Zisserman, and Andrea Vedaldi. N2f2: Hierarchical scene understanding with nested neural feature fields. In European Conference on Computer Vision, pages 197-214. Springer, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 318, + 295, + 372 + ], + "type": "text", + "content": "[5] Aleksei Bokhovkin and Angela Dai. Neural part priors: Learning to optimize part-based object completion in rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9032-9042, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 374, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 374, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 374, + 295, + 430 + ], + "type": "text", + "content": "[6] Alexey Bokhovkin, Vladislav Ishimtsev, Emil Bogomolov, Denis Zorin, Alexey Artemov, Evgeny Burnaev, and Angela Dai. Towards part-based understanding of rgb-d scans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7484-7494, 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "spans": [ + { + "bbox": [ + 62, + 432, + 295, + 486 + ], + "type": "text", + "content": "[7] Daigang Cai, Lichen Zhao, Jing Zhang, Lu Sheng, and Dong Xu. 3djcg: A unified framework for joint dense captioning and visual grounding on 3d point clouds. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16464-16473, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 295, + 553 + ], + "type": "text", + "content": "[8] Boyuan Chen, Zhuo Xu, Sean Kirmani, Brain Ichter, Dorsa Sadigh, Leonidas Guibas, and Fei Xia. Spatialvlm: Endowing vision-language models with spatial reasoning capabilities. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14455-14465, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 295, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 295, + 609 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 295, + 609 + ], + "type": "text", + "content": "[9] Dave Zhenyu Chen, Qirui Wu, Matthias Nießner, and Angel X Chang. D3net: A unified speaker-listener architecture for 3d dense captioning and visual grounding. In European Conference on Computer Vision, pages 487-505. Springer, 2022. 3, 6, 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 295, + 667 + ], + "type": "text", + "content": "[10] Sijin Chen, Hongyuan Zhu, Xin Chen, Yinjie Lei, Gang Yu, and Tao Chen. End-to-end 3d dense captioning with vote2cap-detr. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11124-11133, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Sijin Chen, Xin Chen, Chi Zhang, Mingsheng Li, Gang Yu, Hao Fei, Hongyuan Zhu, Jiayuan Fan, and Tao Chen. L13da: Visual interactive instruction tuning for omni-3d understanding reasoning and planning. In Proceedings of the IEEE/CVF" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "Conference on Computer Vision and Pattern Recognition, pages 26428-26438, 2024. 2, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 96, + 553, + 160 + ], + "type": "text", + "content": "[12] Sijin Chen, Hongyuan Zhu, Mingsheng Li, Xin Chen, Peng Guo, Yinjie Lei, Gang Yu, Taihao Li, and Tao Chen. Vote2cap-detr++: Decoupling localization and describing for end-to-end 3d dense captioning. IEEE Transactions on Pattern Analysis and Machine Intelligence, 46(11):7331-7347, 2024. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 227 + ], + "type": "text", + "content": "[13] Tianrun Chen, Chunan Yu, Jing Li, Jianqi Zhang, Lanyun Zhu, Deyi Ji, Yong Zhang, Ying Zang, Zejian Li, and Lingyun Sun. Reasoning3d-grounding and reasoning in 3d: Fine-grained zero-shot open-vocabulary 3d reasoning part segmentation via large vision-language models. arXiv preprint arXiv:2405.19326, 2024. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 317, + 228, + 553, + 270 + ], + "type": "text", + "content": "[14] Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 15750-15758, 2021. 5" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 317, + 272, + 553, + 326 + ], + "type": "text", + "content": "[15] Zhenyu Chen, Ali Gholami, Matthias Nießner, and Angel X Chang. Scan2cap: Context-aware dense captioning in rgb-d scans. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3193-3203, 2021. 2, 3, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 317, + 327, + 553, + 381 + ], + "type": "text", + "content": "[16] Zhenyu Chen, Ronghang Hu, Xinlei Chen, Matthias Nießner, and Angel X Chang. Unit3d: A unified transformer for 3d dense captioning and visual grounding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 18109-18119, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 317, + 383, + 553, + 435 + ], + "type": "text", + "content": "[17] Christopher Choy, JunYoung Gwak, and Silvio Savarese. 4d spatio-temporal convnets: Minkowski convolutional neural networks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3075-3084, 2019. 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 437, + 553, + 491 + ], + "type": "text", + "content": "[18] Angela Dai, Angel X Chang, Manolis Savva, Maciej Halber, Thomas Funkhouser, and Matthias Nießner. Scannet: Richly-annotated 3d reconstructions of indoor scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5828-5839, 2017. 3, 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 553, + 547 + ], + "type": "text", + "content": "[19] Xueqing Deng, Qihang Yu, Ali Athar, Chenglin Yang, Linjie Yang, Xiaojie Jin, Xiaohui Shen, and Liang-Chieh Chen. Coconut-pancap: Joint panoptic segmentation and grounded captions for fine-grained understanding and generation. arXiv preprint arXiv:2502.02589, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "spans": [ + { + "bbox": [ + 317, + 548, + 553, + 592 + ], + "type": "text", + "content": "[20] Michael Denkowski and Alon Lavie. Meteor universal: Language specific translation evaluation for any target language. In Proceedings of the ninth workshop on statistical machine translation, pages 376-380, 2014. 6, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "spans": [ + { + "bbox": [ + 317, + 593, + 553, + 645 + ], + "type": "text", + "content": "[21] Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024.6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 317, + 647, + 553, + 680 + ], + "type": "text", + "content": "[22] Debidatta Dwibedi, Vidhi Jain, Jonathan Tompson, Andrew Zisserman, and Yusuf Aytar. Flexcap: Generating rich, localized, and flexible captions in images. 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 681, + 555, + 713 + ], + "type": "text", + "content": "[23] Pedro F Felzenszwalb and Daniel P Huttenlocher. Efficient graph-based image segmentation. International journal of computer vision, 59:167-181, 2004. 4, 1, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21707" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 127 + ], + "type": "text", + "content": "[24] Yunhao Ge, Xiaohui Zeng, Jacob Samuel Huffman, Tsung-Yi Lin, Ming-Yu Liu, and Yin Cui. Visual fact checker: enabling high-fidelity detailed caption generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14033-14042, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 296, + 183 + ], + "type": "text", + "content": "[25] Yining Hong, Haoyu Zhen, Peihao Chen, Shuhong Zheng, Yilun Du, Zhenfang Chen, and Chuang Gan. 3d-llm: Injecting the 3d world into large language models. Advances in Neural Information Processing Systems, 36:20482-20494, 2023. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 237 + ], + "type": "text", + "content": "[26] Hang Hua, Qing Liu, Lingzhi Zhang, Jing Shi, Zhifei Zhang, Yilin Wang, Jianming Zhang, and Jiebo Luo. Finecaption: Compositional image captioning focusing on wherever you want at any granularity. arXiv preprint arXiv:2411.15411, 2024.3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 238, + 296, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 238, + 296, + 304 + ], + "spans": [ + { + "bbox": [ + 56, + 238, + 296, + 304 + ], + "type": "text", + "content": "[27] Haifeng Huang, Yilun Chen, Zehan Wang, Rongjie Huang, Runsen Xu, Tai Wang, Luping Liu, Xize Cheng, Yang Zhao, Jiangmiao Pang, et al. Chat-scene: Bridging 3d scene and large language models with object identifiers. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024. 3, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 296, + 350 + ], + "type": "text", + "content": "[28] Yang Jiao, Shaoxiang Chen, Zequn Jie, Jingjing Chen, Lin Ma, and Yu-Gang Jiang. More: Multi-order relation mining for dense captioning in 3d scenes. In European Conference on Computer Vision, pages 528-545. Springer, 2022. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 296, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 393 + ], + "type": "text", + "content": "[29] Justin Johnson, Andrej Karpathy, and Li Fei-Fei. Densecap: Fully convolutional localization networks for dense captioning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4565–4574, 2016. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 296, + 448 + ], + "type": "text", + "content": "[30] Chung Min Kim, Mingxuan Wu, Justin Kerr, Ken Goldberg, Matthew Tancik, and Angjoo Kanazawa. Garfield: Group anything with radiance fields. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21530-21539, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 492 + ], + "type": "text", + "content": "[31] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 296, + 536 + ], + "type": "text", + "content": "[32] Minjung Kim, Hyung Suk Lim, Seung Hwan Kim, Soonyoung Lee, Bumsoo Kim, and Gunhee Kim. See it all: Contextualized late aggregation for 3d dense captioning. arXiv preprint arXiv:2408.07648, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 296, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 592 + ], + "type": "text", + "content": "[33] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015-4026, 2023. 6, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 296, + 624 + ], + "type": "text", + "content": "[34] Harold W Kuhn. The hungarian method for the assignment problem. Naval research logistics quarterly, 2(1-2):83-97, 1955. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 679 + ], + "type": "text", + "content": "[35] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730–19742. PMLR, 2023. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[36] Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 555, + 715 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 106 + ], + "type": "text", + "content": "[37] Haotian Liu, Chunyuan Li, Yuheng Li, Bo Li, Yuanhan Zhang, Sheng Shen, and Yong Jae Lee. Llava next: Improved reasoning,OCR, and world knowledge, 2024.6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 555, + 150 + ], + "type": "text", + "content": "[38] Tiange Luo, Chris Rockwell, Honglak Lee, and Justin Johnson. Scalable 3d captioning with pretrained models. Advances in Neural Information Processing Systems, 36: 75307-75337, 2023. 3, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 555, + 195 + ], + "type": "text", + "content": "[39] Tiange Luo, Justin Johnson, and Honglak Lee. View selection for 3d captioning via diffusion ranking. In European Conference on Computer Vision, pages 180-197. Springer, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 554, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 554, + 219 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 554, + 219 + ], + "type": "text", + "content": "[40] George A Miller. Wordnet: a lexical database for english. Communications of the ACM, 38(11):39-41, 1995. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "text", + "content": "[41] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th annual meeting of the Association for Computational Linguistics, pages 311-318, 2002. 6, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 554, + 309 + ], + "type": "text", + "content": "[42] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, Ilya Sutskever, et al. Language models are unsupervised multitask learners. OpenAI blog, 1(8):9, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 554, + 365 + ], + "type": "text", + "content": "[43] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 21(140):1-67, 2020. 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 554, + 409 + ], + "type": "text", + "content": "[44] Gabriel Oliveira dos Santos, Esther Luna Colombini, and Sandra Avila. Cider-r: Robust consensus-based image description evaluation. arXiv preprint arXiv:2109.13701, 2021. 6, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 410, + 555, + 466 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 410, + 555, + 466 + ], + "spans": [ + { + "bbox": [ + 316, + 410, + 555, + 466 + ], + "type": "text", + "content": "[45] Jonas Schult, Francis Engelmann, Alexander Hermans, Or Litany, Siyu Tang, and Bastian Leibe. Mask3d: Mask transformer for 3d semantic instance segmentation. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 8216-8223. IEEE, 2023. 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 467, + 554, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 467, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 467, + 554, + 510 + ], + "type": "text", + "content": "[46] Ayca Takmaz, Alexandros Delitzas, Robert W Sumner, Francis Engelmann, Johanna Wald, and Federico Tombari. Search3d: Hierarchical open-vocabulary 3d segmentation. IEEE Robotics and Automation Letters, 2025. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 511, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 511, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 511, + 555, + 567 + ], + "type": "text", + "content": "[47] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9568-9578, 2024. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 567, + 555, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 567, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 567, + 555, + 612 + ], + "type": "text", + "content": "[48] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 4566-4575, 2015. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 554, + 646 + ], + "type": "text", + "content": "[49] Heng Wang, Chaoyi Zhang, Jianhui Yu, and Weidong Cai. Spatiality-guided transformer for 3d dense captioning on point clouds. arXiv preprint arXiv:2204.10688, 2022. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 554, + 689 + ], + "type": "text", + "content": "[50] Zehan Wang, Haifeng Huang, Yang Zhao, Ziang Zhang, and Zhou Zhao. Chat-3d: Data-efficiently tuning large language model for universal dialogue of 3d scenes. arXiv preprint arXiv:2308.08769, 2023. 3" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 715 + ], + "type": "text", + "content": "[51] Mi Yan, Jiazhao Zhang, Yan Zhu, and He Wang. Maskclustering: View consensus based mask graph clustering for" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21708" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 575 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 75, + 72, + 294, + 106 + ], + "type": "text", + "content": "open-vocabulary 3d instance segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28274-28284, 2024. 6, 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[52] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the International Conference on Computer Vision (ICCV), 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[53] Zuyao You, Junke Wang, Lingyu Kong, Bo He, and Zuxuan Wu. Pix2cap-coco: Advancing visual comprehension via pixel-level captioning. arXiv preprint arXiv:2501.13893, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 262 + ], + "type": "text", + "content": "[54] Zhihao Yuan, Xu Yan, Yinghong Liao, Yao Guo, Guanbin Li, Shuguang Cui, and Zhen Li. X-trans2cap: Cross-modal knowledge transfer using transformer for 3d dense captioning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8563-8573, 2022. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "spans": [ + { + "bbox": [ + 56, + 263, + 294, + 317 + ], + "type": "text", + "content": "[55] Jiawei Zhang, Chejian Xu, and Bo Li. Chatscene: Knowledge-enabled safety-critical scenario generation for autonomous vehicles. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15459-15469, 2024. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "spans": [ + { + "bbox": [ + 56, + 319, + 294, + 363 + ], + "type": "text", + "content": "[56] Yiming Zhang, ZeMing Gong, and Angel X Chang. Multi3drefer: Grounding text description to multiple 3d objects. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15225-15236, 2023. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 364, + 294, + 407 + ], + "type": "text", + "content": "[57] Chenming Zhu, Tai Wang, Wenwei Zhang, Jiangmiao Pang, and Xihui Liu. Llava-3d: A simple yet effective pathway to empowering lmms with 3d-awareness. arXiv preprint arXiv:2409.18125, 2024. 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 463 + ], + "type": "text", + "content": "[58] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 3, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 464, + 294, + 518 + ], + "type": "text", + "content": "[59] Ziyu Zhu, Xiaojian Ma, Yixin Chen, Zhidong Deng, Siyuan Huang, and Qing Li. 3d-vista: Pre-trained transformer for 3d vision and text alignment. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2911-2921, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 575 + ], + "type": "text", + "content": "[60] Ziyu Zhu, Zhuofan Zhang, Xiaojian Ma, Xuesong Niu, Yixin Chen, Baoxiong Jia, Zhidong Deng, Siyuan Huang, and Qing Li. Unifying 3d vision-language understanding via promptable queries. In European Conference on Computer Vision, pages 188–206. Springer, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21709" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_content_list.json b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..533af3cf3344a313f99d5b72b9150e531d807927 --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_content_list.json @@ -0,0 +1,1681 @@ +[ + { + "type": "text", + "text": "Explaining Human Preferences via Metrics for Structured 3D Reconstruction", + "text_level": 1, + "bbox": [ + 107, + 130, + 888, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jack Langerman $^{1*}$ Denys Rozumnyi $^{2,3\\dagger}$ Yuzhong Huang $^{4}$ Dmytro Mishkin $^{3,4}$", + "bbox": [ + 215, + 179, + 852, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Independent Researcher, USA $^{2}$ ETH Zurich, Switzerland", + "bbox": [ + 300, + 198, + 772, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{3}$ Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic $^{4}$ Hover Inc., USA", + "bbox": [ + 99, + 215, + 970, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jack@jackml.com rozumden@gmail.com yuzhong.huang@hover.to dmytro.mishkin@hover.to", + "bbox": [ + 166, + 236, + 900, + 251 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"What cannot be measured cannot be improved\" while likely never uttered by Lord Kelvin, summarizes effectively the driving force behind this work. This paper presents a detailed discussion of automated metrics for evaluating structured 3D reconstructions. Pitfalls of each metric are discussed, and an analysis through the lens of expert 3D modelers' preferences is presented. A set of systematic \"unit tests\" are proposed to empirically verify desirable properties, and context aware recommendations regarding which metric to use depending on application are provided. Finally, a learned metric distilled from human expert judgments is proposed and analyzed. The source code is available at https://github.com/s23dr/wireframe-metrics-iccv2025.", + "bbox": [ + 89, + 316, + 483, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 542, + 220, + 558 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Benchmarks have been key drivers of progress in computer vision; the canonical example is certainly ImageNet [31], but prominent examples abound beyond image classification: object tracking [21, 23], image retrieval [4, 30], image matching [20], 6D pose estimation [15], optical flow [27], etc. Benchmarks have three main components - the data, the protocol, and the metrics. While the data is the single most important component, progress is hard without being able to answer the question, \"progress on what?\" Good metrics are the quantitative answer to this question. While metrics do not need to be perfect, their gradient should point progress in the right direction.", + "bbox": [ + 88, + 566, + 482, + 748 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We consider an area of structured and semi-structured reconstruction, which has recently gained popularity [1] [19] [5]. Given a set or sequence of sensory data, such as ground images [22], satellite images [9], or aerial LiDAR [33], the goal is to produce a wireframe or a CAD model of a building or other structure. The modeling output is presented as a spatial graph with vertices (such as roof apex point, etc.) and edges (ridge line, etc.). Several", + "bbox": [ + 88, + 750, + 482, + 869 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6b489c11183e473453f10fd566e8f66ccbe32cb5b7926e6fd5b6071c22d101b8.jpg", + "image_caption": [ + "Figure 1. A motivating example for this work. While humans tend to sort the wireframes from best to worst in the presented order, popular metrics (defined in Sec 3) sort them differently, sometimes completely inverting the order. Top, left to right: GT - ground truth wireframe, WF1 - wireframe with edges split into several segments, maintaining geometrical and topological accuracy, WF2 - wireframe with missing vertices and edges, WF3 - wireframe with only one correct vertex. Bottom: distances between GT and respective wireframe. Numbers that change sorting are in red." + ], + "image_footnote": [], + "bbox": [ + 516, + 284, + 859, + 527 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "datasets have been recently proposed [22, 33] for the task. The issue is that hardly a pair of publications in the area use the same metric to evaluate quantitative results. Some use recognition metrics such as precision and recall on vertices and edges [25] or more enhanced versions such as Structured Average Precision [26, 36]. Others opt for graph-based metrics, such as the Wireframe Edit Distance (WED) [8, 24, 25]. Finally, some methods treat the problem similarly to point cloud registration and report Chamfer Distance (CD) [10, 16, 17]. Other related fields, like structure from motion, use downstream metrics, such as image generation quality [6] and camera pose accuracy [20]. One aspect of the difficulty is related to the fact that structured reconstructions often have different goals. For instance, one purpose of the reconstructed wireframes is to represent", + "bbox": [ + 511, + 672, + 906, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Now at Apple.", + "bbox": [ + 107, + 876, + 194, + 888 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Now at Meta.", + "bbox": [ + 109, + 888, + 189, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "26944", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "building plans and answer questions such as \"what is the area of the bedroom?\" Within this formulation, a black-box model, e.g. visual-and-language model (VLM), which takes an image as an input and outputs a correct estimate, would be a perfect match. On the other hand, a floorplan or a blueprint has value for record-keeping, planning, and other applications that cannot be easily replaced with a black-box model.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Finally, many existing metrics, while useful, often fail to deliver value in practice when comparing two imperfect estimations, and designing metrics that effectively compare a pair of \"very good\" and a pair of \"very bad\" solutions at the same time is also non-trivial.", + "bbox": [ + 89, + 212, + 482, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, there is evidence [22] that some metrics can be \"hacked\" or exploited in such a way that obviously bad solutions have better scores than flawed but ultimately quite reasonable solutions. Such examples include a number of corner cases, which can cause existing metrics to become useless in practical scenarios. For instance, in the case where long edges are split into smaller, yet perfectly collinear, pieces, e.g. Fig. 1, most commonly used metrics, such as edge F1, vertex F1, and Wireframe Edit Distance, fail completely and prefer much worse solutions.", + "bbox": [ + 89, + 289, + 483, + 439 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This paper makes the following contributions:", + "bbox": [ + 109, + 441, + 415, + 455 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(1) Measure the perceived quality of reconstructions by human domain experts, and infer a global ranking of all structured reconstructions. This ranking is then compared to the ranking given by all metrics.", + "(2) Show how well existing metrics agree with human preferences and how much they correlate with each other.", + "(3) Propose a set of \"unit-tests\" for testing the properties of the metrics.", + "(4) Introduce a simple learned metric, which correlates well with human judgment.", + "(5) Make recommendations of which metrics to use depending on the use case." + ], + "bbox": [ + 89, + 460, + 482, + 641 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. What do we actually need from wireframe comparison metrics?", + "text_level": 1, + "bbox": [ + 89, + 657, + 482, + 693 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We consider semi-automated 3D modeling as a target task with enough generality to drive progress that can smoothly scale to fully-automated modeling, but is still tractable and reliable for commercial applications today.", + "bbox": [ + 89, + 702, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For this task, the wireframe representation is estimated from some \"raw\" inputs such as images or a LiDAR point cloud and then transferred to human experts to (1) approve it, (2) correct it and then approve, (3) reject and create the model from scratch manually.", + "bbox": [ + 89, + 763, + 482, + 839 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We argue that such task formulation creates an implicit usefulness ranking over reconstructions (and thereby over reconstruction methods). Otherwise, without considering human involvement, everything becomes binary - either the", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2f3d971d3ac4ce16064fb86898c58dfa79a16035def0accfe11f3227edbf8cd8.jpg", + "image_caption": [ + "Figure 2. Wireframe ranking interface for human annotators." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 890, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "model is good enough to be used (e.g. for 3D printing, measurement extraction), or it is not.", + "bbox": [ + 511, + 325, + 903, + 354 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "For this reason, we consider the following experimental setup to benchmark wireframe comparison metrics. A pool of professional 3D modelers, whose everyday job is creating CAD-like models from raw data such as images, is asked to rank pairs of wireframes. An example of the ranking setup is shown in Fig. 2. All wireframes are superimposed with ground truth models, and annotators are provided tools allowing them to translate, scale, and rotate the wireframes in 3D.", + "bbox": [ + 511, + 357, + 905, + 492 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The wireframes are drawn from two pools, as described below. We then evaluate how the existing metrics agree with the judgment of professional human 3D modeling experts.", + "bbox": [ + 511, + 494, + 905, + 541 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pool1- $S^23DR$ . We acquire a representative set of $S^23DR$ challenge entries [22] as well as a PC2WF [24] baseline. These wireframes were algorithmically (and with the help of deep-learning models) reconstructed from multiview inputs with the goal of minimizing a variant of WED. We include the top-10 entries with team names used as identifiers. The ground truth models were created by human experts and have undergone significant validation. The input data were captured by users on mobile phones in North America.", + "bbox": [ + 511, + 545, + 905, + 696 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pool2 — Corrupted ground truth. We apply one of the following operations on the ground-truth wireframes from Pool 1 – examples are shown in Fig. 3:", + "bbox": [ + 511, + 702, + 905, + 747 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- (deform\\_{low, medium, high}) Split the edge into several edges and perturb the positions of the vertices without breaking the topology.", + "- (perturb\\_{low, medium, high}) Split the vertex into several ones. Each of the new vertices is randomly shifted from the ground truth. If the original vertex was connected to multiple neighbors, randomly decide which of the new vertices is connected to which neighbors.", + "- (add\\_{low, medium, high}) Randomly add wrong edges to the model." + ], + "bbox": [ + 513, + 750, + 905, + 898 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "26945", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/2e5b3dde039e5a06620d9d5709c9db432835955cd98a4f6ea88a06d78ed79353.jpg", + "image_caption": [ + "Figure 3. Examples of corrupted ground truth wireframes, used for wireframe ranking. Left to right: GT, deformed edges (deform_medium), vertex duplication and random movement (perturb_medium), edge addition (add_low), edge deletion(remove_low)." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 480, + 236 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "- (remove\\_{low, medium, high}) Randomly delete some of the vertices and all the edges connected to them.", + "bbox": [ + 89, + 324, + 483, + 356 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unit-tests & Desired properties of dissimilarity scores: In addition to being aligned with human judgment, we also propose a set of \"unit-tests\" for the metrics, which we believe are reasonable, and check if the dissimilarity scores satisfy these requirements. We design tests for the formal properties of mathematical metrics as well as additional properties relevant to evaluating dissimilarity in structured reconstruction tasks: For example, if wrong edges $E_{1}$ and $E_{2}$ are added to the GT model, the metric should score the resulting wireframe lower than if $E_{1}$ or $E_{2}$ are added separately.", + "bbox": [ + 89, + 359, + 483, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Identity of Indiscernibles: This property ensures that identical inputs receive a dissimilarity score of zero, indicating perfect similarity. For any reconstruction $x$ , a metric $d$ satisfies this property if $d(x,x) = 0$ .", + "bbox": [ + 89, + 531, + 483, + 594 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Symmetry: A symmetric metric produces the same dissimilarity score regardless of the order of the inputs. For reconstructions $x$ and $y$ , a metric satisfies symmetry if $d(x, y) = d(y, x)$ .", + "bbox": [ + 89, + 597, + 483, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Triangle Inequality: The triangle inequality ensures that for any three reconstructions $x, y$ , and $z$ , the dissimilarity between $x$ and $z$ is less than or equal to the sum of dissimilarities between $x$ and $y$ , and $y$ and $z$ . This relationship is expressed as $d(x, z) \\leq d(x, y) + d(y, z)$ .", + "bbox": [ + 89, + 662, + 483, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monotonicity: This property describes how the dissimilarity score behaves when components (such as vertices or edges) are removed from a reconstruction. A metric satisfies monotonicity if the dissimilarity score does not increase when wrong vertices or edges are deleted. Similarly, the dissimilarity must not increase when correct vertices or edges are added.", + "bbox": [ + 89, + 744, + 483, + 851 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Quasi-proportionality: This property holds when the metric changes smoothly under perturbations. This is evaluated by moving random vertices with small increments and", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "checking the variance of the differences in the score. We use the following perturbations to simulate better or worse reconstructions: (i) remove correct edges from the ground truth wireframe; (ii) add wrong edges to the ground truth wireframe; (iii) disconnect ground truth edges; (iv) remove correct vertices; (v) move ground truth vertices to the wrong location. For every perturbation, we apply it 10 times and declare an example monotonic if it is strictly increasing (or decreasing as appropriate) for those continuous 10 perturbations.", + "bbox": [ + 511, + 90, + 906, + 242 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Metrics", + "text_level": 1, + "bbox": [ + 513, + 253, + 604, + 270 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The following metrics are considered:", + "bbox": [ + 511, + 279, + 767, + 295 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "WED - Wireframe Edit Distance was proposed by Liu et al. [24] as an extension of the Graph Edit Distance (GED) [32]. GED quantifies the distance between two graphs as the minimum number of elementary operations (inserting and deleting edges and vertices) required to transform one graph into another. WED extends this to wireframes (graphs with node positions and edge lengths) and proposes a cheap approximation to the NP-Hard problem of computing the optimal sequence of edits. Concretely, an assignment is first computed between the predicted and ground-truth vertices, and a cost is paid proportional to the distance between matched vertices. Next, unmatched vertices are deleted, and missing vertices are inserted (paying a cost proportional to the number of inserted/deleted vertices). Finally, given the vertex assignments, missing edges are inserted and extra edges deleted, paying a cost proportional to their length. In order to use WED, one needs to decide on the cost of insertion and deletion of the vertices, as well as the order of operations, and method of computing vertex/edge assignment. WED was used to determine the winner in the Building3D [33] and $S^2 3DR$ [22] CVPR Challenges.", + "bbox": [ + 511, + 297, + 906, + 630 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ECD - Edge Chamfer Distance. ECD is commonly used in structured reconstruction papers [10, 16, 17]. We consider a family of chamfer-like metrics between two point sets $A$ and $B$ sampled from wireframe edges. The general form is:", + "bbox": [ + 511, + 633, + 906, + 709 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nd (A, B) := \\inf _ {\\pi_ {A B}: A \\rightarrow B} \\mathbb {E} _ {a \\in A} [ f (a, \\pi_ {A B} (a)) ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 717, + 906, + 741 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\pi_{AB}$ represents an assignment from elements in $A$ to elements in $B$ , and $f$ is typically an $\\ell_p$ norm of the difference between the inputs. Different constraints on $\\pi_{AB}$ yield different metrics:", + "bbox": [ + 511, + 748, + 905, + 806 + ], + "page_idx": 2 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- The classical chamfer distance corresponds to $\\pi_{AB}(a) = \\arg \\min_{b\\in B}f(a,b)$ , i.e. nearest neighbor matching.", + "- The most constrained version requires $\\pi_{AB}$ to be a bijective matching, which can be computed via the Hungarian algorithm and is equivalent to the Earth Mover's Distance when $f$ is the $\\ell_p$ norm." + ], + "bbox": [ + 511, + 810, + 903, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "26946", + "bbox": [ + 478, + 945, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f4bfc0a23c27ee08617745a209476d2699bdf48e4855aae9fd8b3e53671206f8.jpg", + "image_caption": [ + "Figure 4. Annotator agreement (all pairs). Left to right: annotator agreement with each other, the learned metric, handcrafted metrics, and VLMs. Annotators background: A-K - 3D modellers, Des[0-2] - designers, CV[0-3] - computer vision engineers. Best zoom-in." + ], + "image_footnote": [], + "bbox": [ + 143, + 88, + 854, + 359 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Length Weighted Spectral Graph Distance - SD incorporates both topological and geometric information by framing graph (wireframe) distance in terms of distances between the spectra of weighted graph Laplacians. We measure the spectral distance using the 2-Wasserstein metric between the eigenvalue distributions:", + "bbox": [ + 89, + 416, + 483, + 506 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nS D \\left(G _ {1}, G _ {2}\\right) := W _ {2} \\left(\\lambda \\left(L _ {1}\\right), \\lambda \\left(L _ {2}\\right)\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 518, + 482, + 537 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\lambda (L)$ denotes the spectrum of the Laplacian $L$ . For a graph $G = (V,E)$ , the weighted graph Laplacian is defined:", + "bbox": [ + 89, + 549, + 483, + 580 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nL := D - A \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 241, + 595, + 480, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D$ is the weighted degree matrix $(|V|\\times |V|$ diagonal matrix with each diagonal entry containing the sum of the lengths of edges incident to that vertex), and $A$ is the weighted adjacency matrix $(|V|\\times |V|$ with $A_{ij} = \\left\\| \\mathrm{coord}(V_i) - \\mathrm{coord}(V_j)\\right\\| _2$ if $(i,j)\\in E$ and 0 otherwise).", + "bbox": [ + 89, + 619, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "**Corner and Edge Metrics.** We also compute precision, recall, and F1 scores for both corners and edges. For corners, we consider a prediction correct if it lies within a distance threshold of a ground truth corner. For edges, we use the Hausdorff distance between line segments to determine matches. These metrics provide an intuitive measure of the topological accuracy of the predicted wireframes.", + "bbox": [ + 89, + 699, + 483, + 806 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Hausdorff Distance measures the maximum of minimal distances between two sets of points. For wireframes, we sample points along the edges and compute the Hausdorff distance between these point sets, providing a measure of geometric similarity that considers both corner positions and edge geometry.", + "bbox": [ + 89, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Intersection over Union is a popular metric in a wide range of fields, e.g. segmentation, tracking, object detection. However, it is rarely used to assess the quality of wireframe reconstructions. We extend the definition of the wireframe as a set of cylinders with a fixed radius (the only hyperparameter of this metric) and define the metric as an IoU between two sets of cylinders, given by two wireframe reconstructions that need to be compared. An approximation via point sampling is considered: sample random points from both sets of cylinders and compute the average number of times when the point falls inside of both sets of cylinders. The Jaccard distance is reported between two sets.", + "bbox": [ + 511, + 416, + 906, + 598 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Visual-and-language models could potentially be used for this task. We consider 4o, o1 [13], Grok 2 [34], qwen-2.5 [18], pixtral 12b [2], claude 3.5 and 3.7 [3], gemini 2.0 and gemini 2.0-flash [12] models via OpenRouter [28]. The prompts are provided in the supplementary.", + "bbox": [ + 511, + 604, + 906, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learned metric. We also explore how to distill the human annotations directly into a metric. To this end, we propose learning a metric with transfer learning. First, reconstruction and ground-truth wireframes are plotted in 3D and rendered from a canonical viewpoint (denoted $r_i$ ). Then, DiNOv2 [29] features are extracted from those renderings and an MLP-based regression head is then trained to regress scores based on the extracted features $(g(r_i))$ . A Bradley-Terry [7] probability model is assumed and pairwise annotations are used to supervise the training by minimizing a binary cross-entropy loss with a batch size of 16. 10-fold cross validation splits the data such that the sets of ground truth structures and reconstruction methods used in the training and test sets are disjoint. We observe average", + "bbox": [ + 511, + 688, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "26947", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/cd16ca1a1682d7cb9fd311975cf52469089e2c142b4000ca4570147193ff0f0c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 114, + 90, + 352, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/9a30fc9d39fb192c1fb636582b4b3691271fa64242e14fb78ead7d8c7d4c9159.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 90, + 599, + 157 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2cc93e652f50491651771a5f51639c840e72b6ab17c35baa0a4e04175e6810c1.jpg", + "image_caption": [ + "Figure 5. Probability of selecting wrong \"winner\" depending on number of raters (left), individual accuracy (center), win rate (right)." + ], + "image_footnote": [], + "bbox": [ + 604, + 90, + 901, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "accuracy across the folds of $76\\%$ where a prediction is considered correct if $g(r_{\\mathrm{winner}}) > g(r_{\\mathrm{loser}})$ .", + "bbox": [ + 89, + 181, + 482, + 214 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 226, + 223, + 243 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Human wireframe ranking and its consistency. To determine which of the metrics under consideration are most appropriate, we employ three groups of people to provide pseudo-ground-truth rankings of the solutions. The first and biggest group is made up of human 3D modeling experts who professionally create CAD models of objects from photos. The second group is computer vision researchers, and finally, the third group is people who do not work with 3D modeling in their daily lives (designers). There are 11, 4, and 3 people in those groups respectively.", + "bbox": [ + 89, + 255, + 483, + 406 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Annotators were shown pairs of reconstructions of the same structure, and asked to specify which reconstruction most closely matched the superimposed corresponding wireframe. An example of the user interface is shown in Fig. 2. Annotators were able to zoom, pan, and rotate, allowing them to examine the solutions from all sides if needed; the viewpoint of both solutions is synchronized to ease the comparison. All 27 methods are compared exhaustively by every rater, meaning $\\binom{27}{2} = 351$ method pairs per house $\\times 10$ houses $= 3510$ pairs for every rater. Most raters rate a few more pairs because of the self-consistency checks. See the Suppl. for the additional information.", + "bbox": [ + 89, + 406, + 483, + 588 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Rater Reliability. We quantify rater reliability using two complementary methods: self-consistency and correctness on synthetic samples.", + "bbox": [ + 89, + 590, + 483, + 637 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Correctness on Synthetic Samples. We introduced synthetic pairs of wireframes with known ground-truth rankings based on systematically applied \"corruptions.\" Each corruption type featured \"Low,\" \"Medium,\" and \"High\" severity levels. We treat the \"low\" vs \"high\" per each corruption type to be obvious enough that if annotators rank them differently, it can be treated as a labeling error, e.g. because of wrong clicks. Average rater accuracy on these \"easy\" pairs is $98.3\\%$ .", + "bbox": [ + 89, + 640, + 482, + 775 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Self-consistency. We assessed intra-rater reliability by measuring how consistently annotators rated repeated pairs of wireframes, occasionally reversing pair order to mitigate order biases. These repeated evaluations constituted only a small subset of the total ratings (self-consistency checks are performed with $5\\%$ probability). The average self-consistency score is $89.4\\%$ , which could be partially attributed to the labeling mistakes and partially to the chang-", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "ing preferences during the annotation process, which we experienced ourselves.", + "bbox": [ + 511, + 181, + 903, + 213 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Do we label enough? Under the assumption that there is some latent \"correct\" winner in any given pair, and $N$ raters each independently select this \"correct\" winner with probability $p$ , we can compute the probability that the majority is \"wrong\" (see Fig. 5). Therefore, the estimated panel error rate per pair is $\\approx 1\\%$ for 11 (expert) raters, and $0.25\\%$ for 17 (all) raters (assuming a significant $20\\%$ individual error rate). We also analyze the stability of our results and present adequacy analysis for the number of raters, comparisons, and houses. For each, we sweep a range of subsample sizes and resample 500 times at each size. For each subsample of a given size, we compute ranking implied by the win rates for that subset and rank correlation (Kendall $\\tau$ ) between the subset ranks and the rankings using the full dataset. We then construct a $95\\%$ Bootstrap CI across the 500 iterates for $\\tau$ . The minimum number of raters/comparisons/houses needed for $\\tau \\geq 0.95$ : comparisons $\\geq 3350$ , houses $\\geq 4$ , raters $\\geq 8$ (in all cases we have more than that).", + "bbox": [ + 511, + 223, + 906, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Finding agreement. We compute an agreement score for each pair for annotators using the following simple rule: the same ranking gets 1 point, decisive ranking vs \"equal\" gets 0.5 points, and the opposite ranking gets zero.", + "bbox": [ + 511, + 505, + 905, + 566 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The agreement table between human annotators, metrics, and VLMs is shown in Fig. 4. The agreement table with all \"equal\" rankings excluded, is shown in the supplementary.", + "bbox": [ + 511, + 573, + 905, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Observation 4.1", + "text_level": 1, + "bbox": [ + 526, + 641, + 640, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "When accounting for ties, there is a moderate global consensus among the human annotators; however, they form two distinct clusters that do not seem to depend on the annotator's background. One group assigns more weights to the edge accuracy – correlated with edge F1 and Jaccard distance, and the second – vertex accuracy, correlated with corner F1 score.", + "bbox": [ + 524, + 662, + 893, + 768 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The average agreement score is around $80\\%$ across all annotators, but within clusters, it increases to $85\\%$ . When annotators are decisive (select one of the reconstructions as clearly better rather than selecting \"equal\"), then the average agreement increases to $91\\%$ , and no clusters are observed. Agreement with the metrics (middle part of Fig. 4) suggests an explanation of preferences.", + "bbox": [ + 511, + 795, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "26948", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e6843ed5aef9bb1c8b830636bbcb71b9eb16400f24f1be86d9c4ac2962d41752.jpg", + "image_caption": [ + "Figure 6. Metric ranking by agreement with group in average. Left to right: the first group of raters with more attention to vertices, the second group of raters with more attention to edges, computer vision engineers, and designers." + ], + "image_footnote": [], + "bbox": [ + 107, + 88, + 300, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/0bc83e8298cab77cd938ad2f41165a811bedf6e8fc0255ad29ed655fc1935197.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 305, + 88, + 496, + 167 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d55c4ce7d380ca9cda75de73148c2c50da3599ab1b4f5cf6d151685e81b27794.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 504, + 88, + 696, + 166 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87a7ae90818a204364d3f5f1d6413dd6ce96f8aef66c8b1b5bb8ee88cd12cb4e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 700, + 88, + 895, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 4.2", + "text_level": 1, + "bbox": [ + 102, + 213, + 218, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human annotators pay more attention to correct parts of the reconstruction than the incorrect parts. Regardless of whether edges or vertices are considered, recall metrics agree more with human preferences than precision ones.", + "bbox": [ + 101, + 234, + 473, + 309 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cluster 1 (raters A-G, CV, Des1-2) mostly correlates with corner-based metrics, such as corner recall and corner F1-score, whereas Cluster 2 (raters H-K, Des0) correlates more with edge-based metrics, such as edge recall, edge F1-score and Jaccard distance. The second difference between clusters is that Cluster 2 is more likely to give an \"equal\" score for the low-quality reconstruction, whereas Cluster 1 tried to rank reconstruction more decisively.", + "bbox": [ + 89, + 316, + 483, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The metrics rankings w.r.t. different groups of raters is shown in Fig. 8.", + "bbox": [ + 89, + 438, + 483, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 4.3", + "text_level": 1, + "bbox": [ + 102, + 472, + 218, + 486 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The average agreement with human preferences of the top handcrafted metrics does not vary significantly. WED-based scores correlate with annotators the least.", + "bbox": [ + 101, + 494, + 470, + 539 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The supplementary material shows the full agreement table. Furthermore, the ranking setup changes the human preferences for different metrics; for example, Jaccard distance performs much better for the decisive pairs compared to all other metrics. WED with pre-registration, which was used in the $S^23DR$ challenge, shows the worst correlation and is just slightly better than random chance. Other flavors of the WED metric, namely WED_mnn and WED_AP (used in Building3D Challenge), perform better but still worse than the rest. Graph structure metrics are in the middle. We also consider the agreement with VLMs. Despite initial success with individual examples, in our tests they did not perform meaningfully better than chance.", + "bbox": [ + 89, + 547, + 483, + 743 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 4.4", + "text_level": 1, + "bbox": [ + 102, + 753, + 218, + 767 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "VLMs do not show significant agreement with human preferences in the wireframe ranking. The only exceptions are OpenAI models, as well as Grok2; yet those are only slightly better than chance. VLMs agree the most with the WED metrics family.", + "bbox": [ + 101, + 775, + 470, + 851 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Finding the best reconstruction. To determine if there is a single \"quality\" factor which explains the human ex", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f0b61420cc352e3b4d1eb7b264dafbe0d3785aef2ef6c9250ad1c81b4cfe7e76.jpg", + "image_caption": [ + "Figure 7. Metric ranking by agreement with VLMs" + ], + "image_footnote": [], + "bbox": [ + 553, + 210, + 864, + 349 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "pert judgments, we employ three distinct approaches to map these pairwise comparisons to a single ranking of the methods: simple win rates, a Bradley-Terry probability model, and a factor analysis based approach. While the methods differ, they point to highly concordant conclusions.", + "bbox": [ + 511, + 380, + 906, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Observation 4.5", + "text_level": 1, + "bbox": [ + 526, + 467, + 640, + 479 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Human raters tend to rank equally all solutions that are below some quality threshold. Having no solution often ranks better, compared to a totally wrong reconstruction.", + "bbox": [ + 524, + 488, + 893, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Simple Win Rate The first approach is similar to chess scoring – first a win-count table is computed. For each rated pair, each method receives 1 point for a win, 0.5 for a tie, and 0 for a loss. This scoring does not break ties but distributes the points evenly. Results with selected reconstruction methods are shown in Fig. 8. Consistent with the metric-human agreements, recall plays a more important role, and the wireframes with perfect recall – \"add_low\" and \"add_med\" are among the leaders in all groups. In other words, extra (erroneous) edges are considered less of an issue when compared to missing an edge or a vertex – the \"remove_*\" family. One possible explanation resides in the information contained in the reconstruction: if all correct edges are present, we may identify and remove any erroneous ones. However, in many cases, inferring the exact position of an edge or vertex is not possible given the reconstruction alone if key information is missing.", + "bbox": [ + 511, + 551, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The next set of solutions contains roughly correct but slightly noisy reconstructions - \"perturb_*\" and \"deform*\" family. One of the best solutions from the S23DR Challenge ranked better than highly deformed ground truth but worse than the less invasive corruptions.", + "bbox": [ + 511, + 810, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The rest of the reconstructions get almost equal scores", + "bbox": [ + 532, + 885, + 903, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "26949", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "due to the high proportion of draws and lack of wins among themselves.", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Bradley-Terry Model We have also modeled the quality of each solution using a Bradley-Terry (BT) [7, 11, 14, 35] preference model on the expressed preferences of the annotators (using the BT-Abilities as scores). The Bradley-Terry model defines the probability that item $i$ is preferred over item $j$ as:", + "bbox": [ + 89, + 126, + 483, + 215 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP (i > j) = \\frac {a _ {i}}{a _ {i} + a _ {j}}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 212, + 218, + 482, + 248 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $a_{i}, a_{j}$ are positive real numbers representing the latent strength of each item.", + "bbox": [ + 89, + 257, + 482, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Following standard practice, we reparameterize these latent strengths as exponentials of real-valued parameters $\\theta_{i}$ , giving $a_{i} = e^{\\theta_{i}}$ . This yields:", + "bbox": [ + 89, + 289, + 483, + 334 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\nP (i > j) = \\frac {e ^ {\\theta_ {i}}}{e ^ {\\theta_ {i}} + e ^ {\\theta_ {j}}} = \\frac {1}{1 + e ^ {- (\\theta_ {i} - \\theta_ {j})}} = \\sigma \\left(\\theta_ {i} - \\theta_ {j}\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 347, + 482, + 393 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $\\sigma (x)$ is the sigmoid function $\\sigma (x) = \\frac{1}{1 + e^{-x}}$", + "bbox": [ + 89, + 410, + 429, + 428 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "This formulation is further generalized by introducing a scale parameter $s$ and offset $o$ :", + "bbox": [ + 89, + 428, + 483, + 455 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\np _ {i j} = \\sigma \\left(\\frac {\\theta_ {i} - \\theta_ {j}}{s} + o\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 470, + 482, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "With $s = 1$ and $o = 0$ , this reduces to the standard Bradley-Terry formulation. Alternatively, setting $s = 400$ and $o = 800$ yields the Elo scoring system familiar to chess players.", + "bbox": [ + 89, + 516, + 483, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To estimate the latent abilities $\\theta$ , we initialize each $\\theta_{i}$ by sampling from an independent Gaussian distribution and then iteratively minimize the expectation of the following binary cross-entropy loss using stochastic gradient descent (SGD) with the Adam optimizer:", + "bbox": [ + 89, + 563, + 483, + 638 + ], + "page_idx": 6 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathbb {E} _ {(i, j)} \\left[ - y \\log \\left(p _ {i j}\\right) - (1 - y) \\log \\left(1 - p _ {i j}\\right) \\right], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 112, + 652, + 482, + 671 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "where $y = 1$ if item $i$ was chosen over item $j$ , and $y = 0$ otherwise.", + "bbox": [ + 89, + 684, + 483, + 713 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Factor Analysis To investigate whether the data reflect a single underlying dimension of quality, we additionally peruse a factor analysis based approach. We form the methods-by-raters table $M$ such that $M_{kl}$ is the rate at which rater $k$ chose method $l$ when they saw it. We hypothesize the empirical log-odds of these win-rates (rate $l$ wins according to $k$ ), $\\eta = \\log \\frac{M}{1 - M}$ possess a low-rank structure (rank one in the ideal case); this would indicate a single dominant factor (\"quality\") governing outcomes. We apply singular value decomposition (SVD) to factorize $\\eta = U\\Sigma V^T$ and extract the first left singular vector of $\\eta$ containing the estimated quality scores.", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/feba01b554919a65425f707e7a0bdeda35c8f4461c22fd67071e371b35ff61f6.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEmpirical Win RateImplied Win Rate (BT)Implied Win Rate (Elo)BT AbilityElo ScoreQuality Factor
add_low0.890.890.892.7919370.03
add_med0.860.860.862.4717690.02
perturb_med0.850.850.852.3317390.02
add_high0.820.820.822.041604-0.02
perturb_low0.790.790.791.791510-0.01
remove_low0.790.780.791.711498-0.02
perturb_high0.670.670.680.831144-0.09
deform_med0.670.670.660.831107-0.09
deform_low0.660.660.660.781094-0.10
remove_high0.650.650.650.671077-0.10
remove_med0.630.640.640.601027-0.11
kc920.510.510.51-0.25698-0.18
Siromanec0.500.500.49-0.34645-0.19
deform_high0.500.500.50-0.34669-0.19
maximivashechkin0.390.390.39-1.01386-0.27
rozumden0.380.380.38-1.10373-0.28
kcml0.350.350.35-1.28291-0.26
rozumden0.340.340.33-1.35236-0.26
Ana-Geneva0.320.320.32-1.47221-0.25
pc2wf_retrain0.290.290.30-1.65157-0.23
Yurii0.290.290.29-1.63146-0.25
snuggler0.250.250.24-1.9215-0.25
baseline0.250.250.25-1.9119-0.25
Hunter-X0.220.220.22-2.10-43-0.25
TUM0.220.220.22-2.13-48-0.26
Fudan EDLAB0.210.210.21-2.18-76-0.26
pc2wf_pretrained0.200.200.20-2.28-117-0.25
", + "bbox": [ + 516, + 89, + 916, + 397 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Win rates for selected wireframe, according to the pairwise human annotations and estimated Elo.", + "bbox": [ + 513, + 398, + 906, + 426 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Observation 4.6", + "text_level": 1, + "bbox": [ + 526, + 448, + 640, + 462 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We find a Kendall correlation coefficient $>0.7$ between the rankings implied by SVD and those implied by BT. This lends additional evidence to the hypothesis that there is a true \"quality\" factor driving the raters' views. The result is shown in Table 1.", + "bbox": [ + 524, + 469, + 893, + 544 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Metric properties and \"unit tests\". We define a range of properties a good metric should have, implement tests for these properties, and report the results for all metrics. We use a dataset of 128 ground truth wireframes, which we disturb or alter and check the behavior of each metric. We report the percentage of wireframes where the property was valid for a given metric. Results are shown in Table 2. Most of the metrics pass the triangle inequality test, the identity of indiscernibles, and symmetry. None of the metrics is perfectly monotonic w.r.t. wireframe changes, which makes sense; for example, the precision metrics are insensitive to the number of predicted vertices/edges as long as each predicted element is aligned with an element of the ground truth. Hausdorff and WED with pre-registration pass the fewest tests, which is in line with their poor performance w.r.t. human preferences. Conversely - corner F1, edge F1, and Jaccard perform well and align well with the preferences of annotators. The spectral L1 distance and Chamfer edge distance are exceptions - the spectral distance scores well on properties, but not human alignment, and Chamfer one - vice versa.", + "bbox": [ + 511, + 563, + 906, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Additional considerations. After speaking to the human", + "bbox": [ + 513, + 885, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "26950", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f238762e1203070789b3fbd94d1ab6e9c8db6254fd8b9149a3906f5ee2b6feeb.jpg", + "image_caption": [ + "Figure 8. Methods scores according to different groups." + ], + "image_footnote": [], + "bbox": [ + 106, + 85, + 893, + 188 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/002363555356aeda9edf2a05284ddb96f85979b9b3bceb4258a5d9a6417964c9.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
TestCornerEdgeWEDSpectralIoUHausdorffChamfer
PrecRecF1offsetPrecRecF1preregMNNnearestAPL1L2Jaccarddist.edge
Monotonic
Monotonic (wrong edges)0.020.010.021.001.000.001.000.000.000.000.980.630.140.010.090.00
Monotonic (deform/split)1.000.861.000.651.000.981.000.740.790.670.600.000.060.160.050.40
Monotonic (moving vertex)1.001.001.000.981.001.001.000.991.000.991.000.990.990.450.991.00
Monotonic (disconnect edges)0.520.000.500.310.000.000.000.261.001.001.000.000.011.000.000.02
Monotonic (delete vertices)0.001.001.001.000.001.001.000.751.001.001.000.900.330.670.140.99
Monotonic (delete edges)0.000.000.001.000.001.001.001.001.001.001.001.000.311.000.070.83
Identity
Identity of indiscernibles1.001.001.001.001.001.001.000.021.001.001.01.001.001.000.000.00
Near identity of indiscernibles1.001.001.001.001.001.001.001.001.001.001.01.001.001.000.000.00
Symmetry
Symmetry (0 mean, weighted)1.001.001.000.881.001.001.000.440.440.440.440.440.901.000.841.00
Near symmetry (0 mean, weighted)1.001.001.000.871.001.001.000.580.540.540.450.900.921.001.001.00
Symmetry (shift, weighted)1.001.001.001.001.001.001.000.880.990.920.980.991.001.000.821.00
Near symmetry (shift, weighted)1.001.001.001.001.001.001.001.001.000.920.991.001.001.001.001.00
Quasi-proportionality
Quasi-proportionality (shift, far)0.000.000.000.000.000.000.001.000.000.000.001.000.720.050.000.00
Quasi-proportionality (shift, close)0.000.000.000.000.000.000.001.000.160.160.911.000.720.060.000.00
Triangle ineq
Triangle ineq. (rand other)0.991.001.001.001.000.981.001.001.000.990.950.920.971.001.000.96
Triangle ineq. (add noise)1.001.001.000.911.001.001.000.840.810.831.000.900.691.001.001.00
Triangle ineq. (del1/del2)1.001.001.001.001.001.001.000.900.980.981.001.000.661.001.000.43
Pass count11/1711/1712/1711/1712/1713/1714/178/1710/1710/1713/1713/178/1711/176/178/17
", + "bbox": [ + 127, + 208, + 870, + 513 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Properties \"unit-test\" results for metrics, percentage of tests passed. The test is passed if the result is $\\geq {90}\\%$ (black).", + "bbox": [ + 125, + 525, + 867, + 539 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "annotators, we note the following. First, the 3D reconstruction experts are considering how the estimated reconstruction could help them in 3D modeling. There are two main approaches to using such help. The first (in case of noisy reconstructions) is to use the wireframe as a guide to create their own clean reconstruction. The second, if the reconstruction is already close enough, is to fix the wireframe to produce the final result. The Wireframe Edit Distance is designed to estimate how costly it would be to modify the predicted wireframe to match the ground truth. The issue with it is the set of operations – WED only considers vertex/edge deletion/insertion and vertex movement. In practice, one could fit a single edge to multiple noisy ones, bulk-delete a lot of wrong edges or vertices, and apply a rigid transform on the whole model. All of these operations are commonly used in 3D editing software, and WED could benefit from them.", + "bbox": [ + 88, + 553, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finally, if the wireframe is totally wrong, then it is better to have no reconstruction at all and start from scratch. In this \"low quality regime\", most human annotators see no difference between reconstructions if both are wrong.", + "bbox": [ + 89, + 810, + 482, + 869 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Considering usage in competition and benchmarks, we would recommend the use of F1-score, despite the fact that", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "recall-based metrics are more in line with human preferences, as it is less easy to game. For example, a dense grid of vertices and edges could score perfectly on recall but be useless in practice and score poorly on precision-based metrics.", + "bbox": [ + 511, + 553, + 905, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 637, + 633, + 652 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We have studied how human preferences in structured reconstruction evaluation are explained via a wide range of metrics. We show that human preferences can be learned from a small number of examples by transferring from pretrained models which can subsequently be used to score unseen reconstructions. However, we conclude that additional study is warranted prior to relying on such learned metrics as the sole adjudication mechanism for competitions (especially those with strong incentives), because of the potential for reward hacking, gradient-based adversarial attacks, and the like. Based on our study, we recommend using a combination of edge-based (edge F1 or Jaccard score) and corner-based metrics (F1) for benchmarks and competitions. They better explain human preferences in ranking structured reconstructions than more complex and fragile graph-based metrics such as WED or spectral distances.", + "bbox": [ + 509, + 659, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "26951", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements We thank Sherwin Dale Sotto, Rosalie Cellacay, John Carlo Castro, Johsua Vernon Awog, Romeo Sapitanan, Francis Angel Viloria, Joenard Sarmiento, Danica B. Mercado, Zander Hechanova, Michael James Dulay, Joshua Muñoz, Jonah Franz R. De Leon, Viktor Rusov, Olha Mishkina, Yihe Wang and Tyler Nguyen for their help with wireframe annotations. We also thank Tolga Birdal, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, and Ilke Demir for helpful discussions and dataset work at the early stage of the paper. Dmytro Mishkin is supported by MPO 60273/24/21300/21000 CEDMO 2.0 NPO project.", + "bbox": [ + 89, + 90, + 485, + 258 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 275, + 187, + 290 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] 1st workshop on urban scene modeling: Where vision meets photogrammetry and graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1", + "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amelie Heliou, Paul Jacob, Albert Q. Jiang, Kartik Khandelwal, Timothee Lacroix, Guillaume Lample, Diego Las Casas, Thibaut Lavril, Teven Le Scao, Andy Lo, William Marshall, Louis Martin, Arthur Mensch, Pavankumar Muddireddy, Valera Nemychnikova, Marie Pellat, Patrick Von Platen, Nikhil Raghuraman, Baptiste Rozière, Alexandre Sablayrolles, Lucile Saulnier, Romain Sauvestre, Wendy Shang, Roman Soletskyi, Lawrence Stewart, Pierre Stock, Joachim Studnia, Sandeep Subramanian, Sagar Vaze, Thomas Wang, and Sophia Yang. Pixtral 12b, 2024. 4", + "[3] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 2024. 4", + "[4] Andre Araujo, Bingyi Cao, Cam Askew, Jack Sim, Maggie, Tobias Weyand, and Will Cukierski. Google landmark retrieval 2021. https://kaggle.com/competitions/landmark-retrieval-2021, 2021. Kaggle. 1", + "[5] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. Scenescript: Reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision (ECCV), 2024. 1", + "[6] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Positioning of image collections via incremental learning of a relocalizer. In ECCV, 2024. 1", + "[7] Ralph A. Bradley and Milton E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4, 7", + "[8] Li Cao, Yike Xu, Jianwei Guo, and Xiaoping Liu. Wireframenet: A novel method for wireframe generation from" + ], + "bbox": [ + 99, + 300, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "point cloud. Computers and Graphics, 115:226-235, 2023. 1", + "[9] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1", + "[10] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2727-2735, 2023. 1, 3", + "[11] Arpad E. Elo. The Rating of Chessplayers, Past and Present. Arco Publishing, 1978. Accessed: 2025-03-06. 7", + "[12] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 4", + "[13] OpenAI et al. Openai o1 system card, 2024. 4", + "[14] Mark E Glickman. Introductory note to 1928, 2013. 7", + "[15] Tomás Hodań, Martin Sundermeyer, Yann Labbé, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1", + "[16] Perpetual Hope Akwensi, Akshay Bharadwaj, and Ruisheng Wang. Apc2mesh: Bridging the gap from occluded building facades to full 3d models. ISPRS Journal of Photogrammetry and Remote Sensing, 211:438-451, 2024. 1, 3", + "[17] Shangfeng Huang, Ruisheng Wang, Bo Guo, and Hongxin Yang. Pbwr: Parametric-building-wireframe reconstruction from aerial lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27778-27787, 2024. 1, 3", + "[18] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024. 4", + "[19] Apple Inc. Roomplan: Create 3d floor plans with iphone and ipad. https://developer.apple.com/augmented-reality/roomplan/, 2022.1", + "[20] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 1", + "[21] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomas Vojir, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 1", + "[22] Jack Langerman, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, Ilke Demir, Dmytro Mishkin, and Tolga Birdal. S23dr competition at 1st workshop on urban scene modeling @ cvpr 2024. https://huggingface.co/usm3d, 2024. 1, 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "26952", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] L. Leal-Taixe, A. Milan, I. Reid, S. Roth, and K. Schindler. MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942 [cs], 2015. arXiv: 1504.01942. 1", + "[24] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 1, 2, 3", + "[25] Yicheng Luo, Jing Ren, Xuefei Zhe, Di Kang, Yajing Xu, Peter Wonka, and Linchao Bao. Lc2wf: learning to construct 3d building wireframes from 3d line clouds. In Proceedings of the British Machine Vision Conference (BMVC), 2022. 1", + "[26] Wenchao Ma, Bin Tan, Nan Xue, Tianfu Wu, Xianwei Zheng, and Gui-Song Xia. How-3d: Holistic 3d wireframe perception from a single image. In International Conference on 3D Vision, 2022. 1", + "[27] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1", + "[28] OpenRouter. Openrouter: Unified api for ai models, 2024. Accessed: 2025-03-07. 4", + "[29] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4", + "[30] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1", + "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 1", + "[32] Alberto Sanfeliu and King-Sun Fu. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13(3):353-362, 1983. 3", + "[33] Ruisheng Wang, Shangfeng Huang, and Hongxin Yang. Building3D: An Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds. In ICCV, pages 20019-20029, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 3", + "[34] xAI. Grok-2: Advancements in ai language modeling. 2024. Accessed: 2025-03-07. 4", + "[35] Ernst Zermelo. Die berechnung der turnier-ergebnisse als ein maximumproblem der wahrscheinlichkeitsrechnung. Mathematische Zeitschrift, 29:436-460, 1929. 7", + "[36] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In ICCV 2019, 2019. 1" + ], + "bbox": [ + 91, + 90, + 480, + 882 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "26953", + "bbox": [ + 478, + 945, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_model.json b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_model.json new file mode 100644 index 0000000000000000000000000000000000000000..b49c8a890908a5a2a6edba76536d99597300d0a9 --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_model.json @@ -0,0 +1,2189 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.109, + 0.131, + 0.89, + 0.154 + ], + "angle": 0, + "content": "Explaining Human Preferences via Metrics for Structured 3D Reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.18, + 0.854, + 0.199 + ], + "angle": 0, + "content": "Jack Langerman\\(^{1*}\\) Denys Rozumnyi\\(^{2,3\\dagger}\\) Yuzhong Huang\\(^{4}\\) Dmytro Mishkin\\(^{3,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.301, + 0.199, + 0.774, + 0.216 + ], + "angle": 0, + "content": "\\(^{1}\\)Independent Researcher, USA \\(^{2}\\)ETH Zurich, Switzerland" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.216, + 0.972, + 0.235 + ], + "angle": 0, + "content": "\\(^{3}\\)Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic \\(^{4}\\)Hover Inc., USA" + }, + { + "type": "text", + "bbox": [ + 0.167, + 0.237, + 0.901, + 0.252 + ], + "angle": 0, + "content": "jack@jackml.com rozumden@gmail.com yuzhong.huang@hover.to dmytro.mishkin@hover.to" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.528 + ], + "angle": 0, + "content": "\"What cannot be measured cannot be improved\" while likely never uttered by Lord Kelvin, summarizes effectively the driving force behind this work. This paper presents a detailed discussion of automated metrics for evaluating structured 3D reconstructions. Pitfalls of each metric are discussed, and an analysis through the lens of expert 3D modelers' preferences is presented. A set of systematic \"unit tests\" are proposed to empirically verify desirable properties, and context aware recommendations regarding which metric to use depending on application are provided. Finally, a learned metric distilled from human expert judgments is proposed and analyzed. The source code is available at https://github.com/s23dr/wireframe-metrics-iccv2025." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.544, + 0.222, + 0.559 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.568, + 0.483, + 0.749 + ], + "angle": 0, + "content": "Benchmarks have been key drivers of progress in computer vision; the canonical example is certainly ImageNet [31], but prominent examples abound beyond image classification: object tracking [21, 23], image retrieval [4, 30], image matching [20], 6D pose estimation [15], optical flow [27], etc. Benchmarks have three main components - the data, the protocol, and the metrics. While the data is the single most important component, progress is hard without being able to answer the question, \"progress on what?\" Good metrics are the quantitative answer to this question. While metrics do not need to be perfect, their gradient should point progress in the right direction." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.75, + 0.483, + 0.871 + ], + "angle": 0, + "content": "We consider an area of structured and semi-structured reconstruction, which has recently gained popularity [1] [19] [5]. Given a set or sequence of sensory data, such as ground images [22], satellite images [9], or aerial LiDAR [33], the goal is to produce a wireframe or a CAD model of a building or other structure. The modeling output is presented as a spatial graph with vertices (such as roof apex point, etc.) and edges (ridge line, etc.). Several" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.285, + 0.86, + 0.528 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.535, + 0.907, + 0.661 + ], + "angle": 0, + "content": "Figure 1. A motivating example for this work. While humans tend to sort the wireframes from best to worst in the presented order, popular metrics (defined in Sec 3) sort them differently, sometimes completely inverting the order. Top, left to right: GT - ground truth wireframe, WF1 - wireframe with edges split into several segments, maintaining geometrical and topological accuracy, WF2 - wireframe with missing vertices and edges, WF3 - wireframe with only one correct vertex. Bottom: distances between GT and respective wireframe. Numbers that change sorting are in red." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.907, + 0.903 + ], + "angle": 0, + "content": "datasets have been recently proposed [22, 33] for the task. The issue is that hardly a pair of publications in the area use the same metric to evaluate quantitative results. Some use recognition metrics such as precision and recall on vertices and edges [25] or more enhanced versions such as Structured Average Precision [26, 36]. Others opt for graph-based metrics, such as the Wireframe Edit Distance (WED) [8, 24, 25]. Finally, some methods treat the problem similarly to point cloud registration and report Chamfer Distance (CD) [10, 16, 17]. Other related fields, like structure from motion, use downstream metrics, such as image generation quality [6] and camera pose accuracy [20]. One aspect of the difficulty is related to the fact that structured reconstructions often have different goals. For instance, one purpose of the reconstructed wireframes is to represent" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.877, + 0.196, + 0.889 + ], + "angle": 0, + "content": "*Now at Apple." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.889, + 0.19, + 0.9 + ], + "angle": 0, + "content": "† Now at Meta." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.877, + 0.196, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26944" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "building plans and answer questions such as \"what is the area of the bedroom?\" Within this formulation, a black-box model, e.g. visual-and-language model (VLM), which takes an image as an input and outputs a correct estimate, would be a perfect match. On the other hand, a floorplan or a blueprint has value for record-keeping, planning, and other applications that cannot be easily replaced with a black-box model." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.213, + 0.483, + 0.288 + ], + "angle": 0, + "content": "Finally, many existing metrics, while useful, often fail to deliver value in practice when comparing two imperfect estimations, and designing metrics that effectively compare a pair of \"very good\" and a pair of \"very bad\" solutions at the same time is also non-trivial." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.44 + ], + "angle": 0, + "content": "Moreover, there is evidence [22] that some metrics can be \"hacked\" or exploited in such a way that obviously bad solutions have better scores than flawed but ultimately quite reasonable solutions. Such examples include a number of corner cases, which can cause existing metrics to become useless in practical scenarios. For instance, in the case where long edges are split into smaller, yet perfectly collinear, pieces, e.g. Fig. 1, most commonly used metrics, such as edge F1, vertex F1, and Wireframe Edit Distance, fail completely and prefer much worse solutions." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.442, + 0.416, + 0.456 + ], + "angle": 0, + "content": "This paper makes the following contributions:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.461, + 0.483, + 0.521 + ], + "angle": 0, + "content": "(1) Measure the perceived quality of reconstructions by human domain experts, and infer a global ranking of all structured reconstructions. This ranking is then compared to the ranking given by all metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.522, + 0.483, + 0.551 + ], + "angle": 0, + "content": "(2) Show how well existing metrics agree with human preferences and how much they correlate with each other." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.552, + 0.483, + 0.581 + ], + "angle": 0, + "content": "(3) Propose a set of \"unit-tests\" for testing the properties of the metrics." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.582, + 0.483, + 0.612 + ], + "angle": 0, + "content": "(4) Introduce a simple learned metric, which correlates well with human judgment." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.612, + 0.483, + 0.642 + ], + "angle": 0, + "content": "(5) Make recommendations of which metrics to use depending on the use case." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.461, + 0.483, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.659, + 0.483, + 0.694 + ], + "angle": 0, + "content": "2. What do we actually need from wireframe comparison metrics?" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.703, + 0.483, + 0.764 + ], + "angle": 0, + "content": "We consider semi-automated 3D modeling as a target task with enough generality to drive progress that can smoothly scale to fully-automated modeling, but is still tractable and reliable for commercial applications today." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.84 + ], + "angle": 0, + "content": "For this task, the wireframe representation is estimated from some \"raw\" inputs such as images or a LiDAR point cloud and then transferred to human experts to (1) approve it, (2) correct it and then approve, (3) reject and create the model from scratch manually." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We argue that such task formulation creates an implicit usefulness ranking over reconstructions (and thereby over reconstruction methods). Otherwise, without considering human involvement, everything becomes binary - either the" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.092, + 0.892, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.528, + 0.301, + 0.89, + 0.315 + ], + "angle": 0, + "content": "Figure 2. Wireframe ranking interface for human annotators." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.327, + 0.905, + 0.356 + ], + "angle": 0, + "content": "model is good enough to be used (e.g. for 3D printing, measurement extraction), or it is not." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.906, + 0.493 + ], + "angle": 0, + "content": "For this reason, we consider the following experimental setup to benchmark wireframe comparison metrics. A pool of professional 3D modelers, whose everyday job is creating CAD-like models from raw data such as images, is asked to rank pairs of wireframes. An example of the ranking setup is shown in Fig. 2. All wireframes are superimposed with ground truth models, and annotators are provided tools allowing them to translate, scale, and rotate the wireframes in 3D." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.496, + 0.906, + 0.542 + ], + "angle": 0, + "content": "The wireframes are drawn from two pools, as described below. We then evaluate how the existing metrics agree with the judgment of professional human 3D modeling experts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.546, + 0.906, + 0.697 + ], + "angle": 0, + "content": "Pool1- \\( S^23DR \\). We acquire a representative set of \\( S^23DR \\) challenge entries [22] as well as a PC2WF [24] baseline. These wireframes were algorithmically (and with the help of deep-learning models) reconstructed from multiview inputs with the goal of minimizing a variant of WED. We include the top-10 entries with team names used as identifiers. The ground truth models were created by human experts and have undergone significant validation. The input data were captured by users on mobile phones in North America." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.703, + 0.906, + 0.748 + ], + "angle": 0, + "content": "Pool2 — Corrupted ground truth. We apply one of the following operations on the ground-truth wireframes from Pool 1 – examples are shown in Fig. 3:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.75, + 0.905, + 0.796 + ], + "angle": 0, + "content": "- (deform\\_{low, medium, high}) Split the edge into several edges and perturb the positions of the vertices without breaking the topology." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.797, + 0.906, + 0.87 + ], + "angle": 0, + "content": "- (perturb\\_{low, medium, high}) Split the vertex into several ones. Each of the new vertices is randomly shifted from the ground truth. If the original vertex was connected to multiple neighbors, randomly decide which of the new vertices is connected to which neighbors." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "- (add\\_{low, medium, high}) Randomly add wrong edges to the model." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.75, + 0.906, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26945" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.482, + 0.237 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.485, + 0.314 + ], + "angle": 0, + "content": "Figure 3. Examples of corrupted ground truth wireframes, used for wireframe ranking. Left to right: GT, deformed edges (deform_medium), vertex duplication and random movement (perturb_medium), edge addition (add_low), edge deletion(remove_low)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.325, + 0.484, + 0.357 + ], + "angle": 0, + "content": "- (remove\\_{low, medium, high}) Randomly delete some of the vertices and all the edges connected to them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.527 + ], + "angle": 0, + "content": "Unit-tests & Desired properties of dissimilarity scores: In addition to being aligned with human judgment, we also propose a set of \"unit-tests\" for the metrics, which we believe are reasonable, and check if the dissimilarity scores satisfy these requirements. We design tests for the formal properties of mathematical metrics as well as additional properties relevant to evaluating dissimilarity in structured reconstruction tasks: For example, if wrong edges \\( E_{1} \\) and \\( E_{2} \\) are added to the GT model, the metric should score the resulting wireframe lower than if \\( E_{1} \\) or \\( E_{2} \\) are added separately." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.532, + 0.484, + 0.595 + ], + "angle": 0, + "content": "Identity of Indiscernibles: This property ensures that identical inputs receive a dissimilarity score of zero, indicating perfect similarity. For any reconstruction \\( x \\), a metric \\( d \\) satisfies this property if \\( d(x,x) = 0 \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.661 + ], + "angle": 0, + "content": "Symmetry: A symmetric metric produces the same dissimilarity score regardless of the order of the inputs. For reconstructions \\( x \\) and \\( y \\), a metric satisfies symmetry if \\( d(x, y) = d(y, x) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.664, + 0.484, + 0.741 + ], + "angle": 0, + "content": "Triangle Inequality: The triangle inequality ensures that for any three reconstructions \\( x, y \\), and \\( z \\), the dissimilarity between \\( x \\) and \\( z \\) is less than or equal to the sum of dissimilarities between \\( x \\) and \\( y \\), and \\( y \\) and \\( z \\). This relationship is expressed as \\( d(x, z) \\leq d(x, y) + d(y, z) \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.745, + 0.484, + 0.852 + ], + "angle": 0, + "content": "Monotonicity: This property describes how the dissimilarity score behaves when components (such as vertices or edges) are removed from a reconstruction. A metric satisfies monotonicity if the dissimilarity score does not increase when wrong vertices or edges are deleted. Similarly, the dissimilarity must not increase when correct vertices or edges are added." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Quasi-proportionality: This property holds when the metric changes smoothly under perturbations. This is evaluated by moving random vertices with small increments and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.243 + ], + "angle": 0, + "content": "checking the variance of the differences in the score. We use the following perturbations to simulate better or worse reconstructions: (i) remove correct edges from the ground truth wireframe; (ii) add wrong edges to the ground truth wireframe; (iii) disconnect ground truth edges; (iv) remove correct vertices; (v) move ground truth vertices to the wrong location. For every perturbation, we apply it 10 times and declare an example monotonic if it is strictly increasing (or decreasing as appropriate) for those continuous 10 perturbations." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.255, + 0.605, + 0.271 + ], + "angle": 0, + "content": "3. Metrics" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.28, + 0.768, + 0.296 + ], + "angle": 0, + "content": "The following metrics are considered:" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.299, + 0.907, + 0.631 + ], + "angle": 0, + "content": "WED - Wireframe Edit Distance was proposed by Liu et al. [24] as an extension of the Graph Edit Distance (GED) [32]. GED quantifies the distance between two graphs as the minimum number of elementary operations (inserting and deleting edges and vertices) required to transform one graph into another. WED extends this to wireframes (graphs with node positions and edge lengths) and proposes a cheap approximation to the NP-Hard problem of computing the optimal sequence of edits. Concretely, an assignment is first computed between the predicted and ground-truth vertices, and a cost is paid proportional to the distance between matched vertices. Next, unmatched vertices are deleted, and missing vertices are inserted (paying a cost proportional to the number of inserted/deleted vertices). Finally, given the vertex assignments, missing edges are inserted and extra edges deleted, paying a cost proportional to their length. In order to use WED, one needs to decide on the cost of insertion and deletion of the vertices, as well as the order of operations, and method of computing vertex/edge assignment. WED was used to determine the winner in the Building3D [33] and \\( S^2 3DR \\) [22] CVPR Challenges." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.635, + 0.907, + 0.71 + ], + "angle": 0, + "content": "ECD - Edge Chamfer Distance. ECD is commonly used in structured reconstruction papers [10, 16, 17]. We consider a family of chamfer-like metrics between two point sets \\( A \\) and \\( B \\) sampled from wireframe edges. The general form is:" + }, + { + "type": "equation", + "bbox": [ + 0.559, + 0.718, + 0.907, + 0.742 + ], + "angle": 0, + "content": "\\[\nd (A, B) := \\inf _ {\\pi_ {A B}: A \\rightarrow B} \\mathbb {E} _ {a \\in A} [ f (a, \\pi_ {A B} (a)) ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.749, + 0.906, + 0.808 + ], + "angle": 0, + "content": "where \\(\\pi_{AB}\\) represents an assignment from elements in \\(A\\) to elements in \\(B\\), and \\(f\\) is typically an \\(\\ell_p\\) norm of the difference between the inputs. Different constraints on \\(\\pi_{AB}\\) yield different metrics:" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- The classical chamfer distance corresponds to \\(\\pi_{AB}(a) = \\arg \\min_{b\\in B}f(a,b)\\), i.e. nearest neighbor matching." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- The most constrained version requires \\(\\pi_{AB}\\) to be a bijective matching, which can be computed via the Hungarian algorithm and is equivalent to the Earth Mover's Distance when \\(f\\) is the \\(\\ell_p\\) norm." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.811, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.946, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26946" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.144, + 0.089, + 0.856, + 0.361 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.908, + 0.404 + ], + "angle": 0, + "content": "Figure 4. Annotator agreement (all pairs). Left to right: annotator agreement with each other, the learned metric, handcrafted metrics, and VLMs. Annotators background: A-K - 3D modellers, Des[0-2] - designers, CV[0-3] - computer vision engineers. Best zoom-in." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.507 + ], + "angle": 0, + "content": "Length Weighted Spectral Graph Distance - SD incorporates both topological and geometric information by framing graph (wireframe) distance in terms of distances between the spectra of weighted graph Laplacians. We measure the spectral distance using the 2-Wasserstein metric between the eigenvalue distributions:" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.52, + 0.483, + 0.539 + ], + "angle": 0, + "content": "\\[\nS D \\left(G _ {1}, G _ {2}\\right) := W _ {2} \\left(\\lambda \\left(L _ {1}\\right), \\lambda \\left(L _ {2}\\right)\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.55, + 0.484, + 0.581 + ], + "angle": 0, + "content": "where \\(\\lambda (L)\\) denotes the spectrum of the Laplacian \\(L\\). For a graph \\(G = (V,E)\\), the weighted graph Laplacian is defined:" + }, + { + "type": "equation", + "bbox": [ + 0.243, + 0.597, + 0.482, + 0.612 + ], + "angle": 0, + "content": "\\[\nL := D - A \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.62, + 0.484, + 0.698 + ], + "angle": 0, + "content": "where \\(D\\) is the weighted degree matrix \\((|V|\\times |V|\\) diagonal matrix with each diagonal entry containing the sum of the lengths of edges incident to that vertex), and \\(A\\) is the weighted adjacency matrix \\((|V|\\times |V|\\) with \\(A_{ij} = \\left\\| \\mathrm{coord}(V_i) - \\mathrm{coord}(V_j)\\right\\| _2\\) if \\((i,j)\\in E\\) and 0 otherwise)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.484, + 0.807 + ], + "angle": 0, + "content": "**Corner and Edge Metrics.** We also compute precision, recall, and F1 scores for both corners and edges. For corners, we consider a prediction correct if it lies within a distance threshold of a ground truth corner. For edges, we use the Hausdorff distance between line segments to determine matches. These metrics provide an intuitive measure of the topological accuracy of the predicted wireframes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Hausdorff Distance measures the maximum of minimal distances between two sets of points. For wireframes, we sample points along the edges and compute the Hausdorff distance between these point sets, providing a measure of geometric similarity that considers both corner positions and edge geometry." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.417, + 0.907, + 0.599 + ], + "angle": 0, + "content": "Intersection over Union is a popular metric in a wide range of fields, e.g. segmentation, tracking, object detection. However, it is rarely used to assess the quality of wireframe reconstructions. We extend the definition of the wireframe as a set of cylinders with a fixed radius (the only hyperparameter of this metric) and define the metric as an IoU between two sets of cylinders, given by two wireframe reconstructions that need to be compared. An approximation via point sampling is considered: sample random points from both sets of cylinders and compute the average number of times when the point falls inside of both sets of cylinders. The Jaccard distance is reported between two sets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.907, + 0.682 + ], + "angle": 0, + "content": "Visual-and-language models could potentially be used for this task. We consider 4o, o1 [13], Grok 2 [34], qwen-2.5 [18], pixtral 12b [2], claude 3.5 and 3.7 [3], gemini 2.0 and gemini 2.0-flash [12] models via OpenRouter [28]. The prompts are provided in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Learned metric. We also explore how to distill the human annotations directly into a metric. To this end, we propose learning a metric with transfer learning. First, reconstruction and ground-truth wireframes are plotted in 3D and rendered from a canonical viewpoint (denoted \\( r_i \\)). Then, DiNOv2 [29] features are extracted from those renderings and an MLP-based regression head is then trained to regress scores based on the extracted features \\( (g(r_i)) \\). A Bradley-Terry [7] probability model is assumed and pairwise annotations are used to supervise the training by minimizing a binary cross-entropy loss with a batch size of 16. 10-fold cross validation splits the data such that the sets of ground truth structures and reconstruction methods used in the training and test sets are disjoint. We observe average" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "26947" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.091, + 0.353, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.091, + 0.601, + 0.158 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.605, + 0.091, + 0.903, + 0.16 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.16, + 0.89, + 0.175 + ], + "angle": 0, + "content": "Figure 5. Probability of selecting wrong \"winner\" depending on number of raters (left), individual accuracy (center), win rate (right)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.215 + ], + "angle": 0, + "content": "accuracy across the folds of \\(76\\%\\) where a prediction is considered correct if \\(g(r_{\\mathrm{winner}}) > g(r_{\\mathrm{loser}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.227, + 0.224, + 0.244 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.256, + 0.484, + 0.407 + ], + "angle": 0, + "content": "Human wireframe ranking and its consistency. To determine which of the metrics under consideration are most appropriate, we employ three groups of people to provide pseudo-ground-truth rankings of the solutions. The first and biggest group is made up of human 3D modeling experts who professionally create CAD models of objects from photos. The second group is computer vision researchers, and finally, the third group is people who do not work with 3D modeling in their daily lives (designers). There are 11, 4, and 3 people in those groups respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.484, + 0.589 + ], + "angle": 0, + "content": "Annotators were shown pairs of reconstructions of the same structure, and asked to specify which reconstruction most closely matched the superimposed corresponding wireframe. An example of the user interface is shown in Fig. 2. Annotators were able to zoom, pan, and rotate, allowing them to examine the solutions from all sides if needed; the viewpoint of both solutions is synchronized to ease the comparison. All 27 methods are compared exhaustively by every rater, meaning \\(\\binom{27}{2} = 351\\) method pairs per house \\(\\times 10\\) houses \\(= 3510\\) pairs for every rater. Most raters rate a few more pairs because of the self-consistency checks. See the Suppl. for the additional information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.592, + 0.484, + 0.638 + ], + "angle": 0, + "content": "Rater Reliability. We quantify rater reliability using two complementary methods: self-consistency and correctness on synthetic samples." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.641, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Correctness on Synthetic Samples. We introduced synthetic pairs of wireframes with known ground-truth rankings based on systematically applied \"corruptions.\" Each corruption type featured \"Low,\" \"Medium,\" and \"High\" severity levels. We treat the \"low\" vs \"high\" per each corruption type to be obvious enough that if annotators rank them differently, it can be treated as a labeling error, e.g. because of wrong clicks. Average rater accuracy on these \"easy\" pairs is \\(98.3\\%\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Self-consistency. We assessed intra-rater reliability by measuring how consistently annotators rated repeated pairs of wireframes, occasionally reversing pair order to mitigate order biases. These repeated evaluations constituted only a small subset of the total ratings (self-consistency checks are performed with \\(5\\%\\) probability). The average self-consistency score is \\(89.4\\%\\), which could be partially attributed to the labeling mistakes and partially to the chang-" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.183, + 0.905, + 0.214 + ], + "angle": 0, + "content": "ing preferences during the annotation process, which we experienced ourselves." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.224, + 0.907, + 0.495 + ], + "angle": 0, + "content": "Do we label enough? Under the assumption that there is some latent \"correct\" winner in any given pair, and \\( N \\) raters each independently select this \"correct\" winner with probability \\( p \\), we can compute the probability that the majority is \"wrong\" (see Fig. 5). Therefore, the estimated panel error rate per pair is \\( \\approx 1\\% \\) for 11 (expert) raters, and \\( 0.25\\% \\) for 17 (all) raters (assuming a significant \\( 20\\% \\) individual error rate). We also analyze the stability of our results and present adequacy analysis for the number of raters, comparisons, and houses. For each, we sweep a range of subsample sizes and resample 500 times at each size. For each subsample of a given size, we compute ranking implied by the win rates for that subset and rank correlation (Kendall \\( \\tau \\)) between the subset ranks and the rankings using the full dataset. We then construct a \\( 95\\% \\) Bootstrap CI across the 500 iterates for \\( \\tau \\). The minimum number of raters/comparisons/houses needed for \\( \\tau \\geq 0.95 \\): comparisons \\( \\geq 3350 \\), houses \\( \\geq 4 \\), raters \\( \\geq 8 \\) (in all cases we have more than that)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.506, + 0.906, + 0.567 + ], + "angle": 0, + "content": "Finding agreement. We compute an agreement score for each pair for annotators using the following simple rule: the same ranking gets 1 point, decisive ranking vs \"equal\" gets 0.5 points, and the opposite ranking gets zero." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.574, + 0.906, + 0.62 + ], + "angle": 0, + "content": "The agreement table between human annotators, metrics, and VLMs is shown in Fig. 4. The agreement table with all \"equal\" rankings excluded, is shown in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.642, + 0.642, + 0.656 + ], + "angle": 0, + "content": "Observation 4.1" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.664, + 0.895, + 0.769 + ], + "angle": 0, + "content": "When accounting for ties, there is a moderate global consensus among the human annotators; however, they form two distinct clusters that do not seem to depend on the annotator's background. One group assigns more weights to the edge accuracy – correlated with edge F1 and Jaccard distance, and the second – vertex accuracy, correlated with corner F1 score." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The average agreement score is around \\(80\\%\\) across all annotators, but within clusters, it increases to \\(85\\%\\). When annotators are decisive (select one of the reconstructions as clearly better rather than selecting \"equal\"), then the average agreement increases to \\(91\\%\\), and no clusters are observed. Agreement with the metrics (middle part of Fig. 4) suggests an explanation of preferences." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26948" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.109, + 0.089, + 0.301, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.306, + 0.089, + 0.498, + 0.168 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.505, + 0.089, + 0.697, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.701, + 0.089, + 0.897, + 0.167 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.177, + 0.908, + 0.207 + ], + "angle": 0, + "content": "Figure 6. Metric ranking by agreement with group in average. Left to right: the first group of raters with more attention to vertices, the second group of raters with more attention to edges, computer vision engineers, and designers." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.214, + 0.22, + 0.228 + ], + "angle": 0, + "content": "Observation 4.2" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.235, + 0.474, + 0.31 + ], + "angle": 0, + "content": "Human annotators pay more attention to correct parts of the reconstruction than the incorrect parts. Regardless of whether edges or vertices are considered, recall metrics agree more with human preferences than precision ones." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.318, + 0.484, + 0.438 + ], + "angle": 0, + "content": "Cluster 1 (raters A-G, CV, Des1-2) mostly correlates with corner-based metrics, such as corner recall and corner F1-score, whereas Cluster 2 (raters H-K, Des0) correlates more with edge-based metrics, such as edge recall, edge F1-score and Jaccard distance. The second difference between clusters is that Cluster 2 is more likely to give an \"equal\" score for the low-quality reconstruction, whereas Cluster 1 tried to rank reconstruction more decisively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.439, + 0.484, + 0.469 + ], + "angle": 0, + "content": "The metrics rankings w.r.t. different groups of raters is shown in Fig. 8." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.473, + 0.22, + 0.487 + ], + "angle": 0, + "content": "Observation 4.3" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.495, + 0.472, + 0.54 + ], + "angle": 0, + "content": "The average agreement with human preferences of the top handcrafted metrics does not vary significantly. WED-based scores correlate with annotators the least." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.548, + 0.484, + 0.744 + ], + "angle": 0, + "content": "The supplementary material shows the full agreement table. Furthermore, the ranking setup changes the human preferences for different metrics; for example, Jaccard distance performs much better for the decisive pairs compared to all other metrics. WED with pre-registration, which was used in the \\( S^23DR \\) challenge, shows the worst correlation and is just slightly better than random chance. Other flavors of the WED metric, namely WED_mnn and WED_AP (used in Building3D Challenge), perform better but still worse than the rest. Graph structure metrics are in the middle. We also consider the agreement with VLMs. Despite initial success with individual examples, in our tests they did not perform meaningfully better than chance." + }, + { + "type": "title", + "bbox": [ + 0.104, + 0.755, + 0.22, + 0.768 + ], + "angle": 0, + "content": "Observation 4.4" + }, + { + "type": "text", + "bbox": [ + 0.102, + 0.776, + 0.472, + 0.852 + ], + "angle": 0, + "content": "VLMs do not show significant agreement with human preferences in the wireframe ranking. The only exceptions are OpenAI models, as well as Grok2; yet those are only slightly better than chance. VLMs agree the most with the WED metrics family." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Finding the best reconstruction. To determine if there is a single \"quality\" factor which explains the human ex" + }, + { + "type": "image", + "bbox": [ + 0.554, + 0.212, + 0.865, + 0.35 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.557, + 0.353, + 0.863, + 0.367 + ], + "angle": 0, + "content": "Figure 7. Metric ranking by agreement with VLMs" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.381, + 0.907, + 0.457 + ], + "angle": 0, + "content": "pert judgments, we employ three distinct approaches to map these pairwise comparisons to a single ranking of the methods: simple win rates, a Bradley-Terry probability model, and a factor analysis based approach. While the methods differ, they point to highly concordant conclusions." + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.468, + 0.642, + 0.481 + ], + "angle": 0, + "content": "Observation 4.5" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.489, + 0.895, + 0.535 + ], + "angle": 0, + "content": "Human raters tend to rank equally all solutions that are below some quality threshold. Having no solution often ranks better, compared to a totally wrong reconstruction." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.81 + ], + "angle": 0, + "content": "Simple Win Rate The first approach is similar to chess scoring – first a win-count table is computed. For each rated pair, each method receives 1 point for a win, 0.5 for a tie, and 0 for a loss. This scoring does not break ties but distributes the points evenly. Results with selected reconstruction methods are shown in Fig. 8. Consistent with the metric-human agreements, recall plays a more important role, and the wireframes with perfect recall – \"add_low\" and \"add_med\" are among the leaders in all groups. In other words, extra (erroneous) edges are considered less of an issue when compared to missing an edge or a vertex – the \"remove_*\" family. One possible explanation resides in the information contained in the reconstruction: if all correct edges are present, we may identify and remove any erroneous ones. However, in many cases, inferring the exact position of an edge or vertex is not possible given the reconstruction alone if key information is missing." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.906, + 0.885 + ], + "angle": 0, + "content": "The next set of solutions contains roughly correct but slightly noisy reconstructions - \"perturb_*\" and \"deform*\" family. One of the best solutions from the S23DR Challenge ranked better than highly deformed ground truth but worse than the less invasive corruptions." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The rest of the reconstructions get almost equal scores" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26949" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "due to the high proportion of draws and lack of wins among themselves." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.127, + 0.484, + 0.217 + ], + "angle": 0, + "content": "Bradley-Terry Model We have also modeled the quality of each solution using a Bradley-Terry (BT) [7, 11, 14, 35] preference model on the expressed preferences of the annotators (using the BT-Abilities as scores). The Bradley-Terry model defines the probability that item \\( i \\) is preferred over item \\( j \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.214, + 0.219, + 0.483, + 0.249 + ], + "angle": 0, + "content": "\\[\nP (i > j) = \\frac {a _ {i}}{a _ {i} + a _ {j}}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.258, + 0.483, + 0.287 + ], + "angle": 0, + "content": "where \\(a_{i}, a_{j}\\) are positive real numbers representing the latent strength of each item." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.335 + ], + "angle": 0, + "content": "Following standard practice, we reparameterize these latent strengths as exponentials of real-valued parameters \\(\\theta_{i}\\), giving \\(a_{i} = e^{\\theta_{i}}\\). This yields:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.348, + 0.483, + 0.395 + ], + "angle": 0, + "content": "\\[\nP (i > j) = \\frac {e ^ {\\theta_ {i}}}{e ^ {\\theta_ {i}} + e ^ {\\theta_ {j}}} = \\frac {1}{1 + e ^ {- (\\theta_ {i} - \\theta_ {j})}} = \\sigma \\left(\\theta_ {i} - \\theta_ {j}\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.411, + 0.431, + 0.429 + ], + "angle": 0, + "content": "where \\(\\sigma (x)\\) is the sigmoid function \\(\\sigma (x) = \\frac{1}{1 + e^{-x}}\\)" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.484, + 0.457 + ], + "angle": 0, + "content": "This formulation is further generalized by introducing a scale parameter \\( s \\) and offset \\( o \\):" + }, + { + "type": "equation", + "bbox": [ + 0.199, + 0.471, + 0.483, + 0.505 + ], + "angle": 0, + "content": "\\[\np _ {i j} = \\sigma \\left(\\frac {\\theta_ {i} - \\theta_ {j}}{s} + o\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.517, + 0.484, + 0.563 + ], + "angle": 0, + "content": "With \\( s = 1 \\) and \\( o = 0 \\), this reduces to the standard Bradley-Terry formulation. Alternatively, setting \\( s = 400 \\) and \\( o = 800 \\) yields the Elo scoring system familiar to chess players." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.484, + 0.64 + ], + "angle": 0, + "content": "To estimate the latent abilities \\(\\theta\\), we initialize each \\(\\theta_{i}\\) by sampling from an independent Gaussian distribution and then iteratively minimize the expectation of the following binary cross-entropy loss using stochastic gradient descent (SGD) with the Adam optimizer:" + }, + { + "type": "equation", + "bbox": [ + 0.114, + 0.654, + 0.483, + 0.672 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathbb {E} _ {(i, j)} \\left[ - y \\log \\left(p _ {i j}\\right) - (1 - y) \\log \\left(1 - p _ {i j}\\right) \\right], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.484, + 0.714 + ], + "angle": 0, + "content": "where \\( y = 1 \\) if item \\( i \\) was chosen over item \\( j \\), and \\( y = 0 \\) otherwise." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Factor Analysis To investigate whether the data reflect a single underlying dimension of quality, we additionally peruse a factor analysis based approach. We form the methods-by-raters table \\( M \\) such that \\( M_{kl} \\) is the rate at which rater \\( k \\) chose method \\( l \\) when they saw it. We hypothesize the empirical log-odds of these win-rates (rate \\( l \\) wins according to \\( k \\)), \\( \\eta = \\log \\frac{M}{1 - M} \\) possess a low-rank structure (rank one in the ideal case); this would indicate a single dominant factor (\"quality\") governing outcomes. We apply singular value decomposition (SVD) to factorize \\( \\eta = U\\Sigma V^T \\) and extract the first left singular vector of \\( \\eta \\) containing the estimated quality scores." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.09, + 0.918, + 0.398 + ], + "angle": 0, + "content": "
MethodEmpirical Win RateImplied Win Rate (BT)Implied Win Rate (Elo)BT AbilityElo ScoreQuality Factor
add_low0.890.890.892.7919370.03
add_med0.860.860.862.4717690.02
perturb_med0.850.850.852.3317390.02
add_high0.820.820.822.041604-0.02
perturb_low0.790.790.791.791510-0.01
remove_low0.790.780.791.711498-0.02
perturb_high0.670.670.680.831144-0.09
deform_med0.670.670.660.831107-0.09
deform_low0.660.660.660.781094-0.10
remove_high0.650.650.650.671077-0.10
remove_med0.630.640.640.601027-0.11
kc920.510.510.51-0.25698-0.18
Siromanec0.500.500.49-0.34645-0.19
deform_high0.500.500.50-0.34669-0.19
maximivashechkin0.390.390.39-1.01386-0.27
rozumden0.380.380.38-1.10373-0.28
kcml0.350.350.35-1.28291-0.26
rozumden0.340.340.33-1.35236-0.26
Ana-Geneva0.320.320.32-1.47221-0.25
pc2wf_retrain0.290.290.30-1.65157-0.23
Yurii0.290.290.29-1.63146-0.25
snuggler0.250.250.24-1.9215-0.25
baseline0.250.250.25-1.9119-0.25
Hunter-X0.220.220.22-2.10-43-0.25
TUM0.220.220.22-2.13-48-0.26
Fudan EDLAB0.210.210.21-2.18-76-0.26
pc2wf_pretrained0.200.200.20-2.28-117-0.25
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.4, + 0.907, + 0.428 + ], + "angle": 0, + "content": "Table 1. Win rates for selected wireframe, according to the pairwise human annotations and estimated Elo." + }, + { + "type": "title", + "bbox": [ + 0.527, + 0.449, + 0.642, + 0.463 + ], + "angle": 0, + "content": "Observation 4.6" + }, + { + "type": "text", + "bbox": [ + 0.525, + 0.47, + 0.895, + 0.545 + ], + "angle": 0, + "content": "We find a Kendall correlation coefficient \\(>0.7\\) between the rankings implied by SVD and those implied by BT. This lends additional evidence to the hypothesis that there is a true \"quality\" factor driving the raters' views. The result is shown in Table 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.88 + ], + "angle": 0, + "content": "Metric properties and \"unit tests\". We define a range of properties a good metric should have, implement tests for these properties, and report the results for all metrics. We use a dataset of 128 ground truth wireframes, which we disturb or alter and check the behavior of each metric. We report the percentage of wireframes where the property was valid for a given metric. Results are shown in Table 2. Most of the metrics pass the triangle inequality test, the identity of indiscernibles, and symmetry. None of the metrics is perfectly monotonic w.r.t. wireframe changes, which makes sense; for example, the precision metrics are insensitive to the number of predicted vertices/edges as long as each predicted element is aligned with an element of the ground truth. Hausdorff and WED with pre-registration pass the fewest tests, which is in line with their poor performance w.r.t. human preferences. Conversely - corner F1, edge F1, and Jaccard perform well and align well with the preferences of annotators. The spectral L1 distance and Chamfer edge distance are exceptions - the spectral distance scores well on properties, but not human alignment, and Chamfer one - vice versa." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Additional considerations. After speaking to the human" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "26950" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.087, + 0.895, + 0.189 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.332, + 0.194, + 0.665, + 0.208 + ], + "angle": 0, + "content": "Figure 8. Methods scores according to different groups." + }, + { + "type": "table", + "bbox": [ + 0.128, + 0.209, + 0.871, + 0.515 + ], + "angle": 0, + "content": "
TestCornerEdgeWEDSpectralIoUHausdorffChamfer
PrecRecF1offsetPrecRecF1preregMNNnearestAPL1L2Jaccarddist.edge
Monotonic
Monotonic (wrong edges)0.020.010.021.001.000.001.000.000.000.000.980.630.140.010.090.00
Monotonic (deform/split)1.000.861.000.651.000.981.000.740.790.670.600.000.060.160.050.40
Monotonic (moving vertex)1.001.001.000.981.001.001.000.991.000.991.000.990.990.450.991.00
Monotonic (disconnect edges)0.520.000.500.310.000.000.000.261.001.001.000.000.011.000.000.02
Monotonic (delete vertices)0.001.001.001.000.001.001.000.751.001.001.000.900.330.670.140.99
Monotonic (delete edges)0.000.000.001.000.001.001.001.001.001.001.001.000.311.000.070.83
Identity
Identity of indiscernibles1.001.001.001.001.001.001.000.021.001.001.01.001.001.000.000.00
Near identity of indiscernibles1.001.001.001.001.001.001.001.001.001.001.01.001.001.000.000.00
Symmetry
Symmetry (0 mean, weighted)1.001.001.000.881.001.001.000.440.440.440.440.440.901.000.841.00
Near symmetry (0 mean, weighted)1.001.001.000.871.001.001.000.580.540.540.450.900.921.001.001.00
Symmetry (shift, weighted)1.001.001.001.001.001.001.000.880.990.920.980.991.001.000.821.00
Near symmetry (shift, weighted)1.001.001.001.001.001.001.001.001.000.920.991.001.001.001.001.00
Quasi-proportionality
Quasi-proportionality (shift, far)0.000.000.000.000.000.000.001.000.000.000.001.000.720.050.000.00
Quasi-proportionality (shift, close)0.000.000.000.000.000.000.001.000.160.160.911.000.720.060.000.00
Triangle ineq
Triangle ineq. (rand other)0.991.001.001.001.000.981.001.001.000.990.950.920.971.001.000.96
Triangle ineq. (add noise)1.001.001.000.911.001.001.000.840.810.831.000.900.691.001.001.00
Triangle ineq. (del1/del2)1.001.001.001.001.001.001.000.900.980.981.001.000.661.001.000.43
Pass count11/1711/1712/1711/1712/1713/1714/178/1710/1710/1713/1713/178/1711/176/178/17
" + }, + { + "type": "table_caption", + "bbox": [ + 0.127, + 0.526, + 0.868, + 0.54 + ], + "angle": 0, + "content": "Table 2. Properties \"unit-test\" results for metrics, percentage of tests passed. The test is passed if the result is \\( \\geq {90}\\% \\) (black)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.554, + 0.483, + 0.809 + ], + "angle": 0, + "content": "annotators, we note the following. First, the 3D reconstruction experts are considering how the estimated reconstruction could help them in 3D modeling. There are two main approaches to using such help. The first (in case of noisy reconstructions) is to use the wireframe as a guide to create their own clean reconstruction. The second, if the reconstruction is already close enough, is to fix the wireframe to produce the final result. The Wireframe Edit Distance is designed to estimate how costly it would be to modify the predicted wireframe to match the ground truth. The issue with it is the set of operations – WED only considers vertex/edge deletion/insertion and vertex movement. In practice, one could fit a single edge to multiple noisy ones, bulk-delete a lot of wrong edges or vertices, and apply a rigid transform on the whole model. All of these operations are commonly used in 3D editing software, and WED could benefit from them." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.87 + ], + "angle": 0, + "content": "Finally, if the wireframe is totally wrong, then it is better to have no reconstruction at all and start from scratch. In this \"low quality regime\", most human annotators see no difference between reconstructions if both are wrong." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Considering usage in competition and benchmarks, we would recommend the use of F1-score, despite the fact that" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.554, + 0.906, + 0.629 + ], + "angle": 0, + "content": "recall-based metrics are more in line with human preferences, as it is less easy to game. For example, a dense grid of vertices and edges could score perfectly on recall but be useless in practice and score poorly on precision-based metrics." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.638, + 0.634, + 0.654 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.66, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We have studied how human preferences in structured reconstruction evaluation are explained via a wide range of metrics. We show that human preferences can be learned from a small number of examples by transferring from pretrained models which can subsequently be used to score unseen reconstructions. However, we conclude that additional study is warranted prior to relying on such learned metrics as the sole adjudication mechanism for competitions (especially those with strong incentives), because of the potential for reward hacking, gradient-based adversarial attacks, and the like. Based on our study, we recommend using a combination of edge-based (edge F1 or Jaccard score) and corner-based metrics (F1) for benchmarks and competitions. They better explain human preferences in ranking structured reconstructions than more complex and fragile graph-based metrics such as WED or spectral distances." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "26951" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.486, + 0.26 + ], + "angle": 0, + "content": "Acknowledgements We thank Sherwin Dale Sotto, Rosalie Cellacay, John Carlo Castro, Johsua Vernon Awog, Romeo Sapitanan, Francis Angel Viloria, Joenard Sarmiento, Danica B. Mercado, Zander Hechanova, Michael James Dulay, Joshua Muñoz, Jonah Franz R. De Leon, Viktor Rusov, Olha Mishkina, Yihe Wang and Tyler Nguyen for their help with wireframe annotations. We also thank Tolga Birdal, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, and Ilke Demir for helpful discussions and dataset work at the early stage of the paper. Dmytro Mishkin is supported by MPO 60273/24/21300/21000 CEDMO 2.0 NPO project." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.276, + 0.188, + 0.291 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.301, + 0.484, + 0.356 + ], + "angle": 0, + "content": "[1] 1st workshop on urban scene modeling: Where vision meets photogrammetry and graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.359, + 0.484, + 0.552 + ], + "angle": 0, + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amelie Heliou, Paul Jacob, Albert Q. Jiang, Kartik Khandelwal, Timothee Lacroix, Guillaume Lample, Diego Las Casas, Thibaut Lavril, Teven Le Scao, Andy Lo, William Marshall, Louis Martin, Arthur Mensch, Pavankumar Muddireddy, Valera Nemychnikova, Marie Pellat, Patrick Von Platen, Nikhil Raghuraman, Baptiste Rozière, Alexandre Sablayrolles, Lucile Saulnier, Romain Sauvestre, Wendy Shang, Roman Soletskyi, Lawrence Stewart, Pierre Stock, Joachim Studnia, Sandeep Subramanian, Sagar Vaze, Thomas Wang, and Sophia Yang. Pixtral 12b, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.556, + 0.482, + 0.582 + ], + "angle": 0, + "content": "[3] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.586, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[4] Andre Araujo, Bingyi Cao, Cam Askew, Jack Sim, Maggie, Tobias Weyand, and Will Cukierski. Google landmark retrieval 2021. https://kaggle.com/competitions/landmark-retrieval-2021, 2021. Kaggle. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.657, + 0.482, + 0.754 + ], + "angle": 0, + "content": "[5] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. Scenescript: Reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision (ECCV), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.757, + 0.482, + 0.825 + ], + "angle": 0, + "content": "[6] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Positioning of image collections via incremental learning of a relocalizer. In ECCV, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.829, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[7] Ralph A. Bradley and Milton E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[8] Li Cao, Yike Xu, Jianwei Guo, and Xiaoping Liu. Wireframenet: A novel method for wireframe generation from" + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.301, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "point cloud. Computers and Graphics, 115:226-235, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.524, + 0.123, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[9] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.181, + 0.905, + 0.263 + ], + "angle": 0, + "content": "[10] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2727-2735, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.266, + 0.904, + 0.293 + ], + "angle": 0, + "content": "[11] Arpad E. Elo. The Rating of Chessplayers, Past and Present. Arco Publishing, 1978. Accessed: 2025-03-06. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.296, + 0.905, + 0.323 + ], + "angle": 0, + "content": "[12] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.326, + 0.823, + 0.34 + ], + "angle": 0, + "content": "[13] OpenAI et al. Openai o1 system card, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.342, + 0.869, + 0.356 + ], + "angle": 0, + "content": "[14] Mark E Glickman. Introductory note to 1928, 2013. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.359, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[15] Tomás Hodań, Martin Sundermeyer, Yann Labbé, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.458, + 0.905, + 0.513 + ], + "angle": 0, + "content": "[16] Perpetual Hope Akwensi, Akshay Bharadwaj, and Ruisheng Wang. Apc2mesh: Bridging the gap from occluded building facades to full 3d models. ISPRS Journal of Photogrammetry and Remote Sensing, 211:438-451, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.516, + 0.907, + 0.584 + ], + "angle": 0, + "content": "[17] Shangfeng Huang, Ruisheng Wang, Bo Guo, and Hongxin Yang. Pbwr: Parametric-building-wireframe reconstruction from aerial lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27778-27787, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.587, + 0.905, + 0.641 + ], + "angle": 0, + "content": "[18] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.645, + 0.905, + 0.686 + ], + "angle": 0, + "content": "[19] Apple Inc. Roomplan: Create 3d floor plans with iphone and ipad. https://developer.apple.com/augmented-reality/roomplan/, 2022.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.689, + 0.905, + 0.744 + ], + "angle": 0, + "content": "[20] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.747, + 0.905, + 0.828 + ], + "angle": 0, + "content": "[21] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomas Vojir, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[22] Jack Langerman, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, Ilke Demir, Dmytro Mishkin, and Tolga Birdal. S23dr competition at 1st workshop on urban scene modeling @ cvpr 2024. https://huggingface.co/usm3d, 2024. 1, 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26952" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[23] L. Leal-Taixe, A. Milan, I. Reid, S. Roth, and K. Schindler. MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942 [cs], 2015. arXiv: 1504.01942. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[24] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[25] Yicheng Luo, Jing Ren, Xuefei Zhe, Di Kang, Yajing Xu, Peter Wonka, and Linchao Bao. Lc2wf: learning to construct 3d building wireframes from 3d line clouds. In Proceedings of the British Machine Vision Conference (BMVC), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[26] Wenchao Ma, Bin Tan, Nan Xue, Tianfu Wu, Xianwei Zheng, and Gui-Song Xia. How-3d: Holistic 3d wireframe perception from a single image. In International Conference on 3D Vision, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.36 + ], + "angle": 0, + "content": "[27] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.388 + ], + "angle": 0, + "content": "[28] OpenRouter. Openrouter: Unified api for ai models, 2024. Accessed: 2025-03-07. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.515 + ], + "angle": 0, + "content": "[29] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.517, + 0.482, + 0.571 + ], + "angle": 0, + "content": "[30] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.574, + 0.482, + 0.655 + ], + "angle": 0, + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.657, + 0.482, + 0.712 + ], + "angle": 0, + "content": "[32] Alberto Sanfeliu and King-Sun Fu. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13(3):353-362, 1983. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.715, + 0.482, + 0.783 + ], + "angle": 0, + "content": "[33] Ruisheng Wang, Shangfeng Huang, and Hongxin Yang. Building3D: An Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds. In ICCV, pages 20019-20029, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.785, + 0.482, + 0.811 + ], + "angle": 0, + "content": "[34] xAI. Grok-2: Advancements in ai language modeling. 2024. Accessed: 2025-03-07. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.814, + 0.482, + 0.854 + ], + "angle": 0, + "content": "[35] Ernst Zermelo. Die berechnung der turnier-ergebnisse als ein maximumproblem der wahrscheinlichkeitsrechnung. Mathematische Zeitschrift, 29:436-460, 1929. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.856, + 0.482, + 0.883 + ], + "angle": 0, + "content": "[36] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In ICCV 2019, 2019. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.883 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "26953" + } + ] +] \ No newline at end of file diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_origin.pdf b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..cb2fa994899a9e7eebc1f2fbb9ab5f2784fc650d --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/daa34d93-214e-4447-9b72-957738436c31_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:249f7c264022b79cbfb2c23aa641de1dce9d4c0674090370c3cde45fa60642b5 +size 584395 diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/full.md b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/full.md new file mode 100644 index 0000000000000000000000000000000000000000..055b85804130d1fa4c0c87d0b9aca918b32a7137 --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/full.md @@ -0,0 +1,316 @@ +# Explaining Human Preferences via Metrics for Structured 3D Reconstruction + +Jack Langerman $^{1*}$ Denys Rozumnyi $^{2,3\dagger}$ Yuzhong Huang $^{4}$ Dmytro Mishkin $^{3,4}$ + +$^{1}$ Independent Researcher, USA $^{2}$ ETH Zurich, Switzerland + +$^{3}$ Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic $^{4}$ Hover Inc., USA + +jack@jackml.com rozumden@gmail.com yuzhong.huang@hover.to dmytro.mishkin@hover.to + +# Abstract + +"What cannot be measured cannot be improved" while likely never uttered by Lord Kelvin, summarizes effectively the driving force behind this work. This paper presents a detailed discussion of automated metrics for evaluating structured 3D reconstructions. Pitfalls of each metric are discussed, and an analysis through the lens of expert 3D modelers' preferences is presented. A set of systematic "unit tests" are proposed to empirically verify desirable properties, and context aware recommendations regarding which metric to use depending on application are provided. Finally, a learned metric distilled from human expert judgments is proposed and analyzed. The source code is available at https://github.com/s23dr/wireframe-metrics-iccv2025. + +# 1. Introduction + +Benchmarks have been key drivers of progress in computer vision; the canonical example is certainly ImageNet [31], but prominent examples abound beyond image classification: object tracking [21, 23], image retrieval [4, 30], image matching [20], 6D pose estimation [15], optical flow [27], etc. Benchmarks have three main components - the data, the protocol, and the metrics. While the data is the single most important component, progress is hard without being able to answer the question, "progress on what?" Good metrics are the quantitative answer to this question. While metrics do not need to be perfect, their gradient should point progress in the right direction. + +We consider an area of structured and semi-structured reconstruction, which has recently gained popularity [1] [19] [5]. Given a set or sequence of sensory data, such as ground images [22], satellite images [9], or aerial LiDAR [33], the goal is to produce a wireframe or a CAD model of a building or other structure. The modeling output is presented as a spatial graph with vertices (such as roof apex point, etc.) and edges (ridge line, etc.). Several + +![](images/6b489c11183e473453f10fd566e8f66ccbe32cb5b7926e6fd5b6071c22d101b8.jpg) +Figure 1. A motivating example for this work. While humans tend to sort the wireframes from best to worst in the presented order, popular metrics (defined in Sec 3) sort them differently, sometimes completely inverting the order. Top, left to right: GT - ground truth wireframe, WF1 - wireframe with edges split into several segments, maintaining geometrical and topological accuracy, WF2 - wireframe with missing vertices and edges, WF3 - wireframe with only one correct vertex. Bottom: distances between GT and respective wireframe. Numbers that change sorting are in red. + +datasets have been recently proposed [22, 33] for the task. The issue is that hardly a pair of publications in the area use the same metric to evaluate quantitative results. Some use recognition metrics such as precision and recall on vertices and edges [25] or more enhanced versions such as Structured Average Precision [26, 36]. Others opt for graph-based metrics, such as the Wireframe Edit Distance (WED) [8, 24, 25]. Finally, some methods treat the problem similarly to point cloud registration and report Chamfer Distance (CD) [10, 16, 17]. Other related fields, like structure from motion, use downstream metrics, such as image generation quality [6] and camera pose accuracy [20]. One aspect of the difficulty is related to the fact that structured reconstructions often have different goals. For instance, one purpose of the reconstructed wireframes is to represent + +building plans and answer questions such as "what is the area of the bedroom?" Within this formulation, a black-box model, e.g. visual-and-language model (VLM), which takes an image as an input and outputs a correct estimate, would be a perfect match. On the other hand, a floorplan or a blueprint has value for record-keeping, planning, and other applications that cannot be easily replaced with a black-box model. + +Finally, many existing metrics, while useful, often fail to deliver value in practice when comparing two imperfect estimations, and designing metrics that effectively compare a pair of "very good" and a pair of "very bad" solutions at the same time is also non-trivial. + +Moreover, there is evidence [22] that some metrics can be "hacked" or exploited in such a way that obviously bad solutions have better scores than flawed but ultimately quite reasonable solutions. Such examples include a number of corner cases, which can cause existing metrics to become useless in practical scenarios. For instance, in the case where long edges are split into smaller, yet perfectly collinear, pieces, e.g. Fig. 1, most commonly used metrics, such as edge F1, vertex F1, and Wireframe Edit Distance, fail completely and prefer much worse solutions. + +This paper makes the following contributions: + +(1) Measure the perceived quality of reconstructions by human domain experts, and infer a global ranking of all structured reconstructions. This ranking is then compared to the ranking given by all metrics. +(2) Show how well existing metrics agree with human preferences and how much they correlate with each other. +(3) Propose a set of "unit-tests" for testing the properties of the metrics. +(4) Introduce a simple learned metric, which correlates well with human judgment. +(5) Make recommendations of which metrics to use depending on the use case. + +# 2. What do we actually need from wireframe comparison metrics? + +We consider semi-automated 3D modeling as a target task with enough generality to drive progress that can smoothly scale to fully-automated modeling, but is still tractable and reliable for commercial applications today. + +For this task, the wireframe representation is estimated from some "raw" inputs such as images or a LiDAR point cloud and then transferred to human experts to (1) approve it, (2) correct it and then approve, (3) reject and create the model from scratch manually. + +We argue that such task formulation creates an implicit usefulness ranking over reconstructions (and thereby over reconstruction methods). Otherwise, without considering human involvement, everything becomes binary - either the + +![](images/2f3d971d3ac4ce16064fb86898c58dfa79a16035def0accfe11f3227edbf8cd8.jpg) +Figure 2. Wireframe ranking interface for human annotators. + +model is good enough to be used (e.g. for 3D printing, measurement extraction), or it is not. + +For this reason, we consider the following experimental setup to benchmark wireframe comparison metrics. A pool of professional 3D modelers, whose everyday job is creating CAD-like models from raw data such as images, is asked to rank pairs of wireframes. An example of the ranking setup is shown in Fig. 2. All wireframes are superimposed with ground truth models, and annotators are provided tools allowing them to translate, scale, and rotate the wireframes in 3D. + +The wireframes are drawn from two pools, as described below. We then evaluate how the existing metrics agree with the judgment of professional human 3D modeling experts. + +Pool1- $S^23DR$ . We acquire a representative set of $S^23DR$ challenge entries [22] as well as a PC2WF [24] baseline. These wireframes were algorithmically (and with the help of deep-learning models) reconstructed from multiview inputs with the goal of minimizing a variant of WED. We include the top-10 entries with team names used as identifiers. The ground truth models were created by human experts and have undergone significant validation. The input data were captured by users on mobile phones in North America. + +Pool2 — Corrupted ground truth. We apply one of the following operations on the ground-truth wireframes from Pool 1 – examples are shown in Fig. 3: + +- (deform\_{low, medium, high}) Split the edge into several edges and perturb the positions of the vertices without breaking the topology. +- (perturb\_{low, medium, high}) Split the vertex into several ones. Each of the new vertices is randomly shifted from the ground truth. If the original vertex was connected to multiple neighbors, randomly decide which of the new vertices is connected to which neighbors. +- (add\_{low, medium, high}) Randomly add wrong edges to the model. + +![](images/2e5b3dde039e5a06620d9d5709c9db432835955cd98a4f6ea88a06d78ed79353.jpg) +Figure 3. Examples of corrupted ground truth wireframes, used for wireframe ranking. Left to right: GT, deformed edges (deform_medium), vertex duplication and random movement (perturb_medium), edge addition (add_low), edge deletion(remove_low). + +- (remove\_{low, medium, high}) Randomly delete some of the vertices and all the edges connected to them. + +Unit-tests & Desired properties of dissimilarity scores: In addition to being aligned with human judgment, we also propose a set of "unit-tests" for the metrics, which we believe are reasonable, and check if the dissimilarity scores satisfy these requirements. We design tests for the formal properties of mathematical metrics as well as additional properties relevant to evaluating dissimilarity in structured reconstruction tasks: For example, if wrong edges $E_{1}$ and $E_{2}$ are added to the GT model, the metric should score the resulting wireframe lower than if $E_{1}$ or $E_{2}$ are added separately. + +Identity of Indiscernibles: This property ensures that identical inputs receive a dissimilarity score of zero, indicating perfect similarity. For any reconstruction $x$ , a metric $d$ satisfies this property if $d(x,x) = 0$ . + +Symmetry: A symmetric metric produces the same dissimilarity score regardless of the order of the inputs. For reconstructions $x$ and $y$ , a metric satisfies symmetry if $d(x, y) = d(y, x)$ . + +Triangle Inequality: The triangle inequality ensures that for any three reconstructions $x, y$ , and $z$ , the dissimilarity between $x$ and $z$ is less than or equal to the sum of dissimilarities between $x$ and $y$ , and $y$ and $z$ . This relationship is expressed as $d(x, z) \leq d(x, y) + d(y, z)$ . + +Monotonicity: This property describes how the dissimilarity score behaves when components (such as vertices or edges) are removed from a reconstruction. A metric satisfies monotonicity if the dissimilarity score does not increase when wrong vertices or edges are deleted. Similarly, the dissimilarity must not increase when correct vertices or edges are added. + +Quasi-proportionality: This property holds when the metric changes smoothly under perturbations. This is evaluated by moving random vertices with small increments and + +checking the variance of the differences in the score. We use the following perturbations to simulate better or worse reconstructions: (i) remove correct edges from the ground truth wireframe; (ii) add wrong edges to the ground truth wireframe; (iii) disconnect ground truth edges; (iv) remove correct vertices; (v) move ground truth vertices to the wrong location. For every perturbation, we apply it 10 times and declare an example monotonic if it is strictly increasing (or decreasing as appropriate) for those continuous 10 perturbations. + +# 3. Metrics + +The following metrics are considered: + +WED - Wireframe Edit Distance was proposed by Liu et al. [24] as an extension of the Graph Edit Distance (GED) [32]. GED quantifies the distance between two graphs as the minimum number of elementary operations (inserting and deleting edges and vertices) required to transform one graph into another. WED extends this to wireframes (graphs with node positions and edge lengths) and proposes a cheap approximation to the NP-Hard problem of computing the optimal sequence of edits. Concretely, an assignment is first computed between the predicted and ground-truth vertices, and a cost is paid proportional to the distance between matched vertices. Next, unmatched vertices are deleted, and missing vertices are inserted (paying a cost proportional to the number of inserted/deleted vertices). Finally, given the vertex assignments, missing edges are inserted and extra edges deleted, paying a cost proportional to their length. In order to use WED, one needs to decide on the cost of insertion and deletion of the vertices, as well as the order of operations, and method of computing vertex/edge assignment. WED was used to determine the winner in the Building3D [33] and $S^2 3DR$ [22] CVPR Challenges. + +ECD - Edge Chamfer Distance. ECD is commonly used in structured reconstruction papers [10, 16, 17]. We consider a family of chamfer-like metrics between two point sets $A$ and $B$ sampled from wireframe edges. The general form is: + +$$ +d (A, B) := \inf _ {\pi_ {A B}: A \rightarrow B} \mathbb {E} _ {a \in A} [ f (a, \pi_ {A B} (a)) ], \tag {1} +$$ + +where $\pi_{AB}$ represents an assignment from elements in $A$ to elements in $B$ , and $f$ is typically an $\ell_p$ norm of the difference between the inputs. Different constraints on $\pi_{AB}$ yield different metrics: + +- The classical chamfer distance corresponds to $\pi_{AB}(a) = \arg \min_{b\in B}f(a,b)$ , i.e. nearest neighbor matching. +- The most constrained version requires $\pi_{AB}$ to be a bijective matching, which can be computed via the Hungarian algorithm and is equivalent to the Earth Mover's Distance when $f$ is the $\ell_p$ norm. + +![](images/f4bfc0a23c27ee08617745a209476d2699bdf48e4855aae9fd8b3e53671206f8.jpg) +Figure 4. Annotator agreement (all pairs). Left to right: annotator agreement with each other, the learned metric, handcrafted metrics, and VLMs. Annotators background: A-K - 3D modellers, Des[0-2] - designers, CV[0-3] - computer vision engineers. Best zoom-in. + +Length Weighted Spectral Graph Distance - SD incorporates both topological and geometric information by framing graph (wireframe) distance in terms of distances between the spectra of weighted graph Laplacians. We measure the spectral distance using the 2-Wasserstein metric between the eigenvalue distributions: + +$$ +S D \left(G _ {1}, G _ {2}\right) := W _ {2} \left(\lambda \left(L _ {1}\right), \lambda \left(L _ {2}\right)\right), \tag {2} +$$ + +where $\lambda (L)$ denotes the spectrum of the Laplacian $L$ . For a graph $G = (V,E)$ , the weighted graph Laplacian is defined: + +$$ +L := D - A \tag {3} +$$ + +where $D$ is the weighted degree matrix $(|V|\times |V|$ diagonal matrix with each diagonal entry containing the sum of the lengths of edges incident to that vertex), and $A$ is the weighted adjacency matrix $(|V|\times |V|$ with $A_{ij} = \left\| \mathrm{coord}(V_i) - \mathrm{coord}(V_j)\right\| _2$ if $(i,j)\in E$ and 0 otherwise). + +**Corner and Edge Metrics.** We also compute precision, recall, and F1 scores for both corners and edges. For corners, we consider a prediction correct if it lies within a distance threshold of a ground truth corner. For edges, we use the Hausdorff distance between line segments to determine matches. These metrics provide an intuitive measure of the topological accuracy of the predicted wireframes. + +Hausdorff Distance measures the maximum of minimal distances between two sets of points. For wireframes, we sample points along the edges and compute the Hausdorff distance between these point sets, providing a measure of geometric similarity that considers both corner positions and edge geometry. + +Intersection over Union is a popular metric in a wide range of fields, e.g. segmentation, tracking, object detection. However, it is rarely used to assess the quality of wireframe reconstructions. We extend the definition of the wireframe as a set of cylinders with a fixed radius (the only hyperparameter of this metric) and define the metric as an IoU between two sets of cylinders, given by two wireframe reconstructions that need to be compared. An approximation via point sampling is considered: sample random points from both sets of cylinders and compute the average number of times when the point falls inside of both sets of cylinders. The Jaccard distance is reported between two sets. + +Visual-and-language models could potentially be used for this task. We consider 4o, o1 [13], Grok 2 [34], qwen-2.5 [18], pixtral 12b [2], claude 3.5 and 3.7 [3], gemini 2.0 and gemini 2.0-flash [12] models via OpenRouter [28]. The prompts are provided in the supplementary. + +Learned metric. We also explore how to distill the human annotations directly into a metric. To this end, we propose learning a metric with transfer learning. First, reconstruction and ground-truth wireframes are plotted in 3D and rendered from a canonical viewpoint (denoted $r_i$ ). Then, DiNOv2 [29] features are extracted from those renderings and an MLP-based regression head is then trained to regress scores based on the extracted features $(g(r_i))$ . A Bradley-Terry [7] probability model is assumed and pairwise annotations are used to supervise the training by minimizing a binary cross-entropy loss with a batch size of 16. 10-fold cross validation splits the data such that the sets of ground truth structures and reconstruction methods used in the training and test sets are disjoint. We observe average + +![](images/cd16ca1a1682d7cb9fd311975cf52469089e2c142b4000ca4570147193ff0f0c.jpg) + +![](images/9a30fc9d39fb192c1fb636582b4b3691271fa64242e14fb78ead7d8c7d4c9159.jpg) + +![](images/2cc93e652f50491651771a5f51639c840e72b6ab17c35baa0a4e04175e6810c1.jpg) +Figure 5. Probability of selecting wrong "winner" depending on number of raters (left), individual accuracy (center), win rate (right). + +accuracy across the folds of $76\%$ where a prediction is considered correct if $g(r_{\mathrm{winner}}) > g(r_{\mathrm{loser}})$ . + +# 4. Experiments + +Human wireframe ranking and its consistency. To determine which of the metrics under consideration are most appropriate, we employ three groups of people to provide pseudo-ground-truth rankings of the solutions. The first and biggest group is made up of human 3D modeling experts who professionally create CAD models of objects from photos. The second group is computer vision researchers, and finally, the third group is people who do not work with 3D modeling in their daily lives (designers). There are 11, 4, and 3 people in those groups respectively. + +Annotators were shown pairs of reconstructions of the same structure, and asked to specify which reconstruction most closely matched the superimposed corresponding wireframe. An example of the user interface is shown in Fig. 2. Annotators were able to zoom, pan, and rotate, allowing them to examine the solutions from all sides if needed; the viewpoint of both solutions is synchronized to ease the comparison. All 27 methods are compared exhaustively by every rater, meaning $\binom{27}{2} = 351$ method pairs per house $\times 10$ houses $= 3510$ pairs for every rater. Most raters rate a few more pairs because of the self-consistency checks. See the Suppl. for the additional information. + +Rater Reliability. We quantify rater reliability using two complementary methods: self-consistency and correctness on synthetic samples. + +Correctness on Synthetic Samples. We introduced synthetic pairs of wireframes with known ground-truth rankings based on systematically applied "corruptions." Each corruption type featured "Low," "Medium," and "High" severity levels. We treat the "low" vs "high" per each corruption type to be obvious enough that if annotators rank them differently, it can be treated as a labeling error, e.g. because of wrong clicks. Average rater accuracy on these "easy" pairs is $98.3\%$ . + +Self-consistency. We assessed intra-rater reliability by measuring how consistently annotators rated repeated pairs of wireframes, occasionally reversing pair order to mitigate order biases. These repeated evaluations constituted only a small subset of the total ratings (self-consistency checks are performed with $5\%$ probability). The average self-consistency score is $89.4\%$ , which could be partially attributed to the labeling mistakes and partially to the chang- + +ing preferences during the annotation process, which we experienced ourselves. + +Do we label enough? Under the assumption that there is some latent "correct" winner in any given pair, and $N$ raters each independently select this "correct" winner with probability $p$ , we can compute the probability that the majority is "wrong" (see Fig. 5). Therefore, the estimated panel error rate per pair is $\approx 1\%$ for 11 (expert) raters, and $0.25\%$ for 17 (all) raters (assuming a significant $20\%$ individual error rate). We also analyze the stability of our results and present adequacy analysis for the number of raters, comparisons, and houses. For each, we sweep a range of subsample sizes and resample 500 times at each size. For each subsample of a given size, we compute ranking implied by the win rates for that subset and rank correlation (Kendall $\tau$ ) between the subset ranks and the rankings using the full dataset. We then construct a $95\%$ Bootstrap CI across the 500 iterates for $\tau$ . The minimum number of raters/comparisons/houses needed for $\tau \geq 0.95$ : comparisons $\geq 3350$ , houses $\geq 4$ , raters $\geq 8$ (in all cases we have more than that). + +Finding agreement. We compute an agreement score for each pair for annotators using the following simple rule: the same ranking gets 1 point, decisive ranking vs "equal" gets 0.5 points, and the opposite ranking gets zero. + +The agreement table between human annotators, metrics, and VLMs is shown in Fig. 4. The agreement table with all "equal" rankings excluded, is shown in the supplementary. + +# Observation 4.1 + +When accounting for ties, there is a moderate global consensus among the human annotators; however, they form two distinct clusters that do not seem to depend on the annotator's background. One group assigns more weights to the edge accuracy – correlated with edge F1 and Jaccard distance, and the second – vertex accuracy, correlated with corner F1 score. + +The average agreement score is around $80\%$ across all annotators, but within clusters, it increases to $85\%$ . When annotators are decisive (select one of the reconstructions as clearly better rather than selecting "equal"), then the average agreement increases to $91\%$ , and no clusters are observed. Agreement with the metrics (middle part of Fig. 4) suggests an explanation of preferences. + +![](images/e6843ed5aef9bb1c8b830636bbcb71b9eb16400f24f1be86d9c4ac2962d41752.jpg) +Figure 6. Metric ranking by agreement with group in average. Left to right: the first group of raters with more attention to vertices, the second group of raters with more attention to edges, computer vision engineers, and designers. + +![](images/0bc83e8298cab77cd938ad2f41165a811bedf6e8fc0255ad29ed655fc1935197.jpg) + +![](images/d55c4ce7d380ca9cda75de73148c2c50da3599ab1b4f5cf6d151685e81b27794.jpg) + +![](images/87a7ae90818a204364d3f5f1d6413dd6ce96f8aef66c8b1b5bb8ee88cd12cb4e.jpg) + +# Observation 4.2 + +Human annotators pay more attention to correct parts of the reconstruction than the incorrect parts. Regardless of whether edges or vertices are considered, recall metrics agree more with human preferences than precision ones. + +Cluster 1 (raters A-G, CV, Des1-2) mostly correlates with corner-based metrics, such as corner recall and corner F1-score, whereas Cluster 2 (raters H-K, Des0) correlates more with edge-based metrics, such as edge recall, edge F1-score and Jaccard distance. The second difference between clusters is that Cluster 2 is more likely to give an "equal" score for the low-quality reconstruction, whereas Cluster 1 tried to rank reconstruction more decisively. + +The metrics rankings w.r.t. different groups of raters is shown in Fig. 8. + +# Observation 4.3 + +The average agreement with human preferences of the top handcrafted metrics does not vary significantly. WED-based scores correlate with annotators the least. + +The supplementary material shows the full agreement table. Furthermore, the ranking setup changes the human preferences for different metrics; for example, Jaccard distance performs much better for the decisive pairs compared to all other metrics. WED with pre-registration, which was used in the $S^23DR$ challenge, shows the worst correlation and is just slightly better than random chance. Other flavors of the WED metric, namely WED_mnn and WED_AP (used in Building3D Challenge), perform better but still worse than the rest. Graph structure metrics are in the middle. We also consider the agreement with VLMs. Despite initial success with individual examples, in our tests they did not perform meaningfully better than chance. + +# Observation 4.4 + +VLMs do not show significant agreement with human preferences in the wireframe ranking. The only exceptions are OpenAI models, as well as Grok2; yet those are only slightly better than chance. VLMs agree the most with the WED metrics family. + +Finding the best reconstruction. To determine if there is a single "quality" factor which explains the human ex + +![](images/f0b61420cc352e3b4d1eb7b264dafbe0d3785aef2ef6c9250ad1c81b4cfe7e76.jpg) +Figure 7. Metric ranking by agreement with VLMs + +pert judgments, we employ three distinct approaches to map these pairwise comparisons to a single ranking of the methods: simple win rates, a Bradley-Terry probability model, and a factor analysis based approach. While the methods differ, they point to highly concordant conclusions. + +# Observation 4.5 + +Human raters tend to rank equally all solutions that are below some quality threshold. Having no solution often ranks better, compared to a totally wrong reconstruction. + +Simple Win Rate The first approach is similar to chess scoring – first a win-count table is computed. For each rated pair, each method receives 1 point for a win, 0.5 for a tie, and 0 for a loss. This scoring does not break ties but distributes the points evenly. Results with selected reconstruction methods are shown in Fig. 8. Consistent with the metric-human agreements, recall plays a more important role, and the wireframes with perfect recall – "add_low" and "add_med" are among the leaders in all groups. In other words, extra (erroneous) edges are considered less of an issue when compared to missing an edge or a vertex – the "remove_*" family. One possible explanation resides in the information contained in the reconstruction: if all correct edges are present, we may identify and remove any erroneous ones. However, in many cases, inferring the exact position of an edge or vertex is not possible given the reconstruction alone if key information is missing. + +The next set of solutions contains roughly correct but slightly noisy reconstructions - "perturb_*" and "deform*" family. One of the best solutions from the S23DR Challenge ranked better than highly deformed ground truth but worse than the less invasive corruptions. + +The rest of the reconstructions get almost equal scores + +due to the high proportion of draws and lack of wins among themselves. + +Bradley-Terry Model We have also modeled the quality of each solution using a Bradley-Terry (BT) [7, 11, 14, 35] preference model on the expressed preferences of the annotators (using the BT-Abilities as scores). The Bradley-Terry model defines the probability that item $i$ is preferred over item $j$ as: + +$$ +P (i > j) = \frac {a _ {i}}{a _ {i} + a _ {j}}, \tag {4} +$$ + +where $a_{i}, a_{j}$ are positive real numbers representing the latent strength of each item. + +Following standard practice, we reparameterize these latent strengths as exponentials of real-valued parameters $\theta_{i}$ , giving $a_{i} = e^{\theta_{i}}$ . This yields: + +$$ +P (i > j) = \frac {e ^ {\theta_ {i}}}{e ^ {\theta_ {i}} + e ^ {\theta_ {j}}} = \frac {1}{1 + e ^ {- (\theta_ {i} - \theta_ {j})}} = \sigma \left(\theta_ {i} - \theta_ {j}\right), \tag {5} +$$ + +where $\sigma (x)$ is the sigmoid function $\sigma (x) = \frac{1}{1 + e^{-x}}$ + +This formulation is further generalized by introducing a scale parameter $s$ and offset $o$ : + +$$ +p _ {i j} = \sigma \left(\frac {\theta_ {i} - \theta_ {j}}{s} + o\right). \tag {6} +$$ + +With $s = 1$ and $o = 0$ , this reduces to the standard Bradley-Terry formulation. Alternatively, setting $s = 400$ and $o = 800$ yields the Elo scoring system familiar to chess players. + +To estimate the latent abilities $\theta$ , we initialize each $\theta_{i}$ by sampling from an independent Gaussian distribution and then iteratively minimize the expectation of the following binary cross-entropy loss using stochastic gradient descent (SGD) with the Adam optimizer: + +$$ +\mathcal {L} = \mathbb {E} _ {(i, j)} \left[ - y \log \left(p _ {i j}\right) - (1 - y) \log \left(1 - p _ {i j}\right) \right], \tag {7} +$$ + +where $y = 1$ if item $i$ was chosen over item $j$ , and $y = 0$ otherwise. + +Factor Analysis To investigate whether the data reflect a single underlying dimension of quality, we additionally peruse a factor analysis based approach. We form the methods-by-raters table $M$ such that $M_{kl}$ is the rate at which rater $k$ chose method $l$ when they saw it. We hypothesize the empirical log-odds of these win-rates (rate $l$ wins according to $k$ ), $\eta = \log \frac{M}{1 - M}$ possess a low-rank structure (rank one in the ideal case); this would indicate a single dominant factor ("quality") governing outcomes. We apply singular value decomposition (SVD) to factorize $\eta = U\Sigma V^T$ and extract the first left singular vector of $\eta$ containing the estimated quality scores. + +
MethodEmpirical Win RateImplied Win Rate (BT)Implied Win Rate (Elo)BT AbilityElo ScoreQuality Factor
add_low0.890.890.892.7919370.03
add_med0.860.860.862.4717690.02
perturb_med0.850.850.852.3317390.02
add_high0.820.820.822.041604-0.02
perturb_low0.790.790.791.791510-0.01
remove_low0.790.780.791.711498-0.02
perturb_high0.670.670.680.831144-0.09
deform_med0.670.670.660.831107-0.09
deform_low0.660.660.660.781094-0.10
remove_high0.650.650.650.671077-0.10
remove_med0.630.640.640.601027-0.11
kc920.510.510.51-0.25698-0.18
Siromanec0.500.500.49-0.34645-0.19
deform_high0.500.500.50-0.34669-0.19
maximivashechkin0.390.390.39-1.01386-0.27
rozumden0.380.380.38-1.10373-0.28
kcml0.350.350.35-1.28291-0.26
rozumden0.340.340.33-1.35236-0.26
Ana-Geneva0.320.320.32-1.47221-0.25
pc2wf_retrain0.290.290.30-1.65157-0.23
Yurii0.290.290.29-1.63146-0.25
snuggler0.250.250.24-1.9215-0.25
baseline0.250.250.25-1.9119-0.25
Hunter-X0.220.220.22-2.10-43-0.25
TUM0.220.220.22-2.13-48-0.26
Fudan EDLAB0.210.210.21-2.18-76-0.26
pc2wf_pretrained0.200.200.20-2.28-117-0.25
+ +Table 1. Win rates for selected wireframe, according to the pairwise human annotations and estimated Elo. + +# Observation 4.6 + +We find a Kendall correlation coefficient $>0.7$ between the rankings implied by SVD and those implied by BT. This lends additional evidence to the hypothesis that there is a true "quality" factor driving the raters' views. The result is shown in Table 1. + +Metric properties and "unit tests". We define a range of properties a good metric should have, implement tests for these properties, and report the results for all metrics. We use a dataset of 128 ground truth wireframes, which we disturb or alter and check the behavior of each metric. We report the percentage of wireframes where the property was valid for a given metric. Results are shown in Table 2. Most of the metrics pass the triangle inequality test, the identity of indiscernibles, and symmetry. None of the metrics is perfectly monotonic w.r.t. wireframe changes, which makes sense; for example, the precision metrics are insensitive to the number of predicted vertices/edges as long as each predicted element is aligned with an element of the ground truth. Hausdorff and WED with pre-registration pass the fewest tests, which is in line with their poor performance w.r.t. human preferences. Conversely - corner F1, edge F1, and Jaccard perform well and align well with the preferences of annotators. The spectral L1 distance and Chamfer edge distance are exceptions - the spectral distance scores well on properties, but not human alignment, and Chamfer one - vice versa. + +Additional considerations. After speaking to the human + +![](images/f238762e1203070789b3fbd94d1ab6e9c8db6254fd8b9149a3906f5ee2b6feeb.jpg) +Figure 8. Methods scores according to different groups. + +
TestCornerEdgeWEDSpectralIoUHausdorffChamfer
PrecRecF1offsetPrecRecF1preregMNNnearestAPL1L2Jaccarddist.edge
Monotonic
Monotonic (wrong edges)0.020.010.021.001.000.001.000.000.000.000.980.630.140.010.090.00
Monotonic (deform/split)1.000.861.000.651.000.981.000.740.790.670.600.000.060.160.050.40
Monotonic (moving vertex)1.001.001.000.981.001.001.000.991.000.991.000.990.990.450.991.00
Monotonic (disconnect edges)0.520.000.500.310.000.000.000.261.001.001.000.000.011.000.000.02
Monotonic (delete vertices)0.001.001.001.000.001.001.000.751.001.001.000.900.330.670.140.99
Monotonic (delete edges)0.000.000.001.000.001.001.001.001.001.001.001.000.311.000.070.83
Identity
Identity of indiscernibles1.001.001.001.001.001.001.000.021.001.001.01.001.001.000.000.00
Near identity of indiscernibles1.001.001.001.001.001.001.001.001.001.001.01.001.001.000.000.00
Symmetry
Symmetry (0 mean, weighted)1.001.001.000.881.001.001.000.440.440.440.440.440.901.000.841.00
Near symmetry (0 mean, weighted)1.001.001.000.871.001.001.000.580.540.540.450.900.921.001.001.00
Symmetry (shift, weighted)1.001.001.001.001.001.001.000.880.990.920.980.991.001.000.821.00
Near symmetry (shift, weighted)1.001.001.001.001.001.001.001.001.000.920.991.001.001.001.001.00
Quasi-proportionality
Quasi-proportionality (shift, far)0.000.000.000.000.000.000.001.000.000.000.001.000.720.050.000.00
Quasi-proportionality (shift, close)0.000.000.000.000.000.000.001.000.160.160.911.000.720.060.000.00
Triangle ineq
Triangle ineq. (rand other)0.991.001.001.001.000.981.001.001.000.990.950.920.971.001.000.96
Triangle ineq. (add noise)1.001.001.000.911.001.001.000.840.810.831.000.900.691.001.001.00
Triangle ineq. (del1/del2)1.001.001.001.001.001.001.000.900.980.981.001.000.661.001.000.43
Pass count11/1711/1712/1711/1712/1713/1714/178/1710/1710/1713/1713/178/1711/176/178/17
+ +Table 2. Properties "unit-test" results for metrics, percentage of tests passed. The test is passed if the result is $\geq {90}\%$ (black). + +annotators, we note the following. First, the 3D reconstruction experts are considering how the estimated reconstruction could help them in 3D modeling. There are two main approaches to using such help. The first (in case of noisy reconstructions) is to use the wireframe as a guide to create their own clean reconstruction. The second, if the reconstruction is already close enough, is to fix the wireframe to produce the final result. The Wireframe Edit Distance is designed to estimate how costly it would be to modify the predicted wireframe to match the ground truth. The issue with it is the set of operations – WED only considers vertex/edge deletion/insertion and vertex movement. In practice, one could fit a single edge to multiple noisy ones, bulk-delete a lot of wrong edges or vertices, and apply a rigid transform on the whole model. All of these operations are commonly used in 3D editing software, and WED could benefit from them. + +Finally, if the wireframe is totally wrong, then it is better to have no reconstruction at all and start from scratch. In this "low quality regime", most human annotators see no difference between reconstructions if both are wrong. + +Considering usage in competition and benchmarks, we would recommend the use of F1-score, despite the fact that + +recall-based metrics are more in line with human preferences, as it is less easy to game. For example, a dense grid of vertices and edges could score perfectly on recall but be useless in practice and score poorly on precision-based metrics. + +# 5. Conclusion + +We have studied how human preferences in structured reconstruction evaluation are explained via a wide range of metrics. We show that human preferences can be learned from a small number of examples by transferring from pretrained models which can subsequently be used to score unseen reconstructions. However, we conclude that additional study is warranted prior to relying on such learned metrics as the sole adjudication mechanism for competitions (especially those with strong incentives), because of the potential for reward hacking, gradient-based adversarial attacks, and the like. Based on our study, we recommend using a combination of edge-based (edge F1 or Jaccard score) and corner-based metrics (F1) for benchmarks and competitions. They better explain human preferences in ranking structured reconstructions than more complex and fragile graph-based metrics such as WED or spectral distances. + +Acknowledgements We thank Sherwin Dale Sotto, Rosalie Cellacay, John Carlo Castro, Johsua Vernon Awog, Romeo Sapitanan, Francis Angel Viloria, Joenard Sarmiento, Danica B. Mercado, Zander Hechanova, Michael James Dulay, Joshua Muñoz, Jonah Franz R. De Leon, Viktor Rusov, Olha Mishkina, Yihe Wang and Tyler Nguyen for their help with wireframe annotations. We also thank Tolga Birdal, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, and Ilke Demir for helpful discussions and dataset work at the early stage of the paper. Dmytro Mishkin is supported by MPO 60273/24/21300/21000 CEDMO 2.0 NPO project. + +# References + +[1] 1st workshop on urban scene modeling: Where vision meets photogrammetry and graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1 +[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amelie Heliou, Paul Jacob, Albert Q. Jiang, Kartik Khandelwal, Timothee Lacroix, Guillaume Lample, Diego Las Casas, Thibaut Lavril, Teven Le Scao, Andy Lo, William Marshall, Louis Martin, Arthur Mensch, Pavankumar Muddireddy, Valera Nemychnikova, Marie Pellat, Patrick Von Platen, Nikhil Raghuraman, Baptiste Rozière, Alexandre Sablayrolles, Lucile Saulnier, Romain Sauvestre, Wendy Shang, Roman Soletskyi, Lawrence Stewart, Pierre Stock, Joachim Studnia, Sandeep Subramanian, Sagar Vaze, Thomas Wang, and Sophia Yang. Pixtral 12b, 2024. 4 +[3] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 2024. 4 +[4] Andre Araujo, Bingyi Cao, Cam Askew, Jack Sim, Maggie, Tobias Weyand, and Will Cukierski. Google landmark retrieval 2021. https://kaggle.com/competitions/landmark-retrieval-2021, 2021. Kaggle. 1 +[5] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. Scenescript: Reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision (ECCV), 2024. 1 +[6] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Positioning of image collections via incremental learning of a relocalizer. In ECCV, 2024. 1 +[7] Ralph A. Bradley and Milton E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4, 7 +[8] Li Cao, Yike Xu, Jianwei Guo, and Xiaoping Liu. Wireframenet: A novel method for wireframe generation from + +point cloud. Computers and Graphics, 115:226-235, 2023. 1 +[9] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1 +[10] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2727-2735, 2023. 1, 3 +[11] Arpad E. Elo. The Rating of Chessplayers, Past and Present. Arco Publishing, 1978. Accessed: 2025-03-06. 7 +[12] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 4 +[13] OpenAI et al. Openai o1 system card, 2024. 4 +[14] Mark E Glickman. Introductory note to 1928, 2013. 7 +[15] Tomás Hodań, Martin Sundermeyer, Yann Labbé, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1 +[16] Perpetual Hope Akwensi, Akshay Bharadwaj, and Ruisheng Wang. Apc2mesh: Bridging the gap from occluded building facades to full 3d models. ISPRS Journal of Photogrammetry and Remote Sensing, 211:438-451, 2024. 1, 3 +[17] Shangfeng Huang, Ruisheng Wang, Bo Guo, and Hongxin Yang. Pbwr: Parametric-building-wireframe reconstruction from aerial lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27778-27787, 2024. 1, 3 +[18] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024. 4 +[19] Apple Inc. Roomplan: Create 3d floor plans with iphone and ipad. https://developer.apple.com/augmented-reality/roomplan/, 2022.1 +[20] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 1 +[21] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomas Vojir, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 1 +[22] Jack Langerman, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, Ilke Demir, Dmytro Mishkin, and Tolga Birdal. S23dr competition at 1st workshop on urban scene modeling @ cvpr 2024. https://huggingface.co/usm3d, 2024. 1, 2, 3 + +[23] L. Leal-Taixe, A. Milan, I. Reid, S. Roth, and K. Schindler. MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942 [cs], 2015. arXiv: 1504.01942. 1 +[24] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 1, 2, 3 +[25] Yicheng Luo, Jing Ren, Xuefei Zhe, Di Kang, Yajing Xu, Peter Wonka, and Linchao Bao. Lc2wf: learning to construct 3d building wireframes from 3d line clouds. In Proceedings of the British Machine Vision Conference (BMVC), 2022. 1 +[26] Wenchao Ma, Bin Tan, Nan Xue, Tianfu Wu, Xianwei Zheng, and Gui-Song Xia. How-3d: Holistic 3d wireframe perception from a single image. In International Conference on 3D Vision, 2022. 1 +[27] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1 +[28] OpenRouter. Openrouter: Unified api for ai models, 2024. Accessed: 2025-03-07. 4 +[29] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4 +[30] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1 +[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 1 +[32] Alberto Sanfeliu and King-Sun Fu. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13(3):353-362, 1983. 3 +[33] Ruisheng Wang, Shangfeng Huang, and Hongxin Yang. Building3D: An Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds. In ICCV, pages 20019-20029, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 3 +[34] xAI. Grok-2: Advancements in ai language modeling. 2024. Accessed: 2025-03-07. 4 +[35] Ernst Zermelo. Die berechnung der turnier-ergebnisse als ein maximumproblem der wahrscheinlichkeitsrechnung. Mathematische Zeitschrift, 29:436-460, 1929. 7 +[36] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In ICCV 2019, 2019. 1 \ No newline at end of file diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/images.zip b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2ff77d6d4fed73313f0b6cb98d8ef6f4dc476be4 --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:61b50d0b25d22531c4d71d8e520b6b3860a9fe9d1a513d05f08cb39dada934bb +size 750483 diff --git a/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/layout.json b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..563e967039715693e9b84dfa018ab459effe58f5 --- /dev/null +++ b/2025/Explaining Human Preferences via Metrics for Structured 3D Reconstruction/layout.json @@ -0,0 +1,8759 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "spans": [ + { + "bbox": [ + 66, + 103, + 544, + 121 + ], + "type": "text", + "content": "Explaining Human Preferences via Metrics for Structured 3D Reconstruction" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "spans": [ + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "text", + "content": "Jack Langerman" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "text", + "content": " Denys Rozumnyi" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{2,3\\dagger}" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "text", + "content": " Yuzhong Huang" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "text", + "content": " Dmytro Mishkin" + }, + { + "bbox": [ + 132, + 142, + 522, + 157 + ], + "type": "inline_equation", + "content": "^{3,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "spans": [ + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "type": "text", + "content": "Independent Researcher, USA " + }, + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 184, + 157, + 473, + 171 + ], + "type": "text", + "content": "ETH Zurich, Switzerland" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "spans": [ + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "type": "text", + "content": "Faculty of Electrical Engineering, Czech Technical University in Prague, Czech Republic " + }, + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 61, + 171, + 594, + 186 + ], + "type": "text", + "content": "Hover Inc., USA" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 102, + 187, + 551, + 199 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 102, + 187, + 551, + 199 + ], + "spans": [ + { + "bbox": [ + 102, + 187, + 551, + 199 + ], + "type": "text", + "content": "jack@jackml.com rozumden@gmail.com yuzhong.huang@hover.to dmytro.mishkin@hover.to" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 251, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 418 + ], + "type": "text", + "content": "\"What cannot be measured cannot be improved\" while likely never uttered by Lord Kelvin, summarizes effectively the driving force behind this work. This paper presents a detailed discussion of automated metrics for evaluating structured 3D reconstructions. Pitfalls of each metric are discussed, and an analysis through the lens of expert 3D modelers' preferences is presented. A set of systematic \"unit tests\" are proposed to empirically verify desirable properties, and context aware recommendations regarding which metric to use depending on application are provided. Finally, a learned metric distilled from human expert judgments is proposed and analyzed. The source code is available at https://github.com/s23dr/wireframe-metrics-iccv2025." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 135, + 442 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 54, + 449, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 449, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 54, + 449, + 295, + 593 + ], + "type": "text", + "content": "Benchmarks have been key drivers of progress in computer vision; the canonical example is certainly ImageNet [31], but prominent examples abound beyond image classification: object tracking [21, 23], image retrieval [4, 30], image matching [20], 6D pose estimation [15], optical flow [27], etc. Benchmarks have three main components - the data, the protocol, and the metrics. While the data is the single most important component, progress is hard without being able to answer the question, \"progress on what?\" Good metrics are the quantitative answer to this question. While metrics do not need to be perfect, their gradient should point progress in the right direction." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 54, + 594, + 295, + 689 + ], + "type": "text", + "content": "We consider an area of structured and semi-structured reconstruction, which has recently gained popularity [1] [19] [5]. Given a set or sequence of sensory data, such as ground images [22], satellite images [9], or aerial LiDAR [33], the goal is to produce a wireframe or a CAD model of a building or other structure. The modeling output is presented as a spatial graph with vertices (such as roof apex point, etc.) and edges (ridge line, etc.). Several" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 316, + 225, + 526, + 418 + ], + "blocks": [ + { + "bbox": [ + 316, + 225, + 526, + 418 + ], + "lines": [ + { + "bbox": [ + 316, + 225, + 526, + 418 + ], + "spans": [ + { + "bbox": [ + 316, + 225, + 526, + 418 + ], + "type": "image", + "image_path": "6b489c11183e473453f10fd566e8f66ccbe32cb5b7926e6fd5b6071c22d101b8.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 423, + 555, + 523 + ], + "lines": [ + { + "bbox": [ + 313, + 423, + 555, + 523 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 555, + 523 + ], + "type": "text", + "content": "Figure 1. A motivating example for this work. While humans tend to sort the wireframes from best to worst in the presented order, popular metrics (defined in Sec 3) sort them differently, sometimes completely inverting the order. Top, left to right: GT - ground truth wireframe, WF1 - wireframe with edges split into several segments, maintaining geometrical and topological accuracy, WF2 - wireframe with missing vertices and edges, WF3 - wireframe with only one correct vertex. Bottom: distances between GT and respective wireframe. Numbers that change sorting are in red." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 555, + 715 + ], + "type": "text", + "content": "datasets have been recently proposed [22, 33] for the task. The issue is that hardly a pair of publications in the area use the same metric to evaluate quantitative results. Some use recognition metrics such as precision and recall on vertices and edges [25] or more enhanced versions such as Structured Average Precision [26, 36]. Others opt for graph-based metrics, such as the Wireframe Edit Distance (WED) [8, 24, 25]. Finally, some methods treat the problem similarly to point cloud registration and report Chamfer Distance (CD) [10, 16, 17]. Other related fields, like structure from motion, use downstream metrics, such as image generation quality [6] and camera pose accuracy [20]. One aspect of the difficulty is related to the fact that structured reconstructions often have different goals. For instance, one purpose of the reconstructed wireframes is to represent" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 694, + 119, + 704 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 694, + 119, + 704 + ], + "spans": [ + { + "bbox": [ + 66, + 694, + 119, + 704 + ], + "type": "text", + "content": "*Now at Apple." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 704, + 116, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 704, + 116, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 704, + 116, + 712 + ], + "type": "text", + "content": "† Now at Meta." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26944" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "building plans and answer questions such as \"what is the area of the bedroom?\" Within this formulation, a black-box model, e.g. visual-and-language model (VLM), which takes an image as an input and outputs a correct estimate, would be a perfect match. On the other hand, a floorplan or a blueprint has value for record-keeping, planning, and other applications that cannot be easily replaced with a black-box model." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 168, + 295, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 168, + 295, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 168, + 295, + 228 + ], + "type": "text", + "content": "Finally, many existing metrics, while useful, often fail to deliver value in practice when comparing two imperfect estimations, and designing metrics that effectively compare a pair of \"very good\" and a pair of \"very bad\" solutions at the same time is also non-trivial." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 229, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 348 + ], + "type": "text", + "content": "Moreover, there is evidence [22] that some metrics can be \"hacked\" or exploited in such a way that obviously bad solutions have better scores than flawed but ultimately quite reasonable solutions. Such examples include a number of corner cases, which can cause existing metrics to become useless in practical scenarios. For instance, in the case where long edges are split into smaller, yet perfectly collinear, pieces, e.g. Fig. 1, most commonly used metrics, such as edge F1, vertex F1, and Wireframe Edit Distance, fail completely and prefer much worse solutions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 350, + 254, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 350, + 254, + 361 + ], + "spans": [ + { + "bbox": [ + 67, + 350, + 254, + 361 + ], + "type": "text", + "content": "This paper makes the following contributions:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 365, + 295, + 508 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 55, + 365, + 295, + 412 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 365, + 295, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 365, + 295, + 412 + ], + "type": "text", + "content": "(1) Measure the perceived quality of reconstructions by human domain experts, and infer a global ranking of all structured reconstructions. This ranking is then compared to the ranking given by all metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 436 + ], + "type": "text", + "content": "(2) Show how well existing metrics agree with human preferences and how much they correlate with each other." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 295, + 460 + ], + "type": "text", + "content": "(3) Propose a set of \"unit-tests\" for testing the properties of the metrics." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 484 + ], + "type": "text", + "content": "(4) Introduce a simple learned metric, which correlates well with human judgment." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "spans": [ + { + "bbox": [ + 55, + 484, + 295, + 508 + ], + "type": "text", + "content": "(5) Make recommendations of which metrics to use depending on the use case." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 55, + 521, + 295, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 549 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 549 + ], + "type": "text", + "content": "2. What do we actually need from wireframe comparison metrics?" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 556, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 556, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 556, + 295, + 605 + ], + "type": "text", + "content": "We consider semi-automated 3D modeling as a target task with enough generality to drive progress that can smoothly scale to fully-automated modeling, but is still tractable and reliable for commercial applications today." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 665 + ], + "type": "text", + "content": "For this task, the wireframe representation is estimated from some \"raw\" inputs such as images or a LiDAR point cloud and then transferred to human experts to (1) approve it, (2) correct it and then approve, (3) reject and create the model from scratch manually." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "We argue that such task formulation creates an implicit usefulness ranking over reconstructions (and thereby over reconstruction methods). Otherwise, without considering human involvement, everything becomes binary - either the" + } + ] + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 545, + 228 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 545, + 228 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 545, + 228 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 545, + 228 + ], + "type": "image", + "image_path": "2f3d971d3ac4ce16064fb86898c58dfa79a16035def0accfe11f3227edbf8cd8.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 323, + 238, + 544, + 249 + ], + "lines": [ + { + "bbox": [ + 323, + 238, + 544, + 249 + ], + "spans": [ + { + "bbox": [ + 323, + 238, + 544, + 249 + ], + "type": "text", + "content": "Figure 2. Wireframe ranking interface for human annotators." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 553, + 281 + ], + "type": "text", + "content": "model is good enough to be used (e.g. for 3D printing, measurement extraction), or it is not." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 283, + 554, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 554, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 554, + 390 + ], + "type": "text", + "content": "For this reason, we consider the following experimental setup to benchmark wireframe comparison metrics. A pool of professional 3D modelers, whose everyday job is creating CAD-like models from raw data such as images, is asked to rank pairs of wireframes. An example of the ranking setup is shown in Fig. 2. All wireframes are superimposed with ground truth models, and annotators are provided tools allowing them to translate, scale, and rotate the wireframes in 3D." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 392, + 554, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 554, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 554, + 429 + ], + "type": "text", + "content": "The wireframes are drawn from two pools, as described below. We then evaluate how the existing metrics agree with the judgment of professional human 3D modeling experts." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "text", + "content": "Pool1- " + }, + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "inline_equation", + "content": "S^23DR" + }, + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "text", + "content": ". We acquire a representative set of " + }, + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "inline_equation", + "content": "S^23DR" + }, + { + "bbox": [ + 313, + 432, + 554, + 552 + ], + "type": "text", + "content": " challenge entries [22] as well as a PC2WF [24] baseline. These wireframes were algorithmically (and with the help of deep-learning models) reconstructed from multiview inputs with the goal of minimizing a variant of WED. We include the top-10 entries with team names used as identifiers. The ground truth models were created by human experts and have undergone significant validation. The input data were captured by users on mobile phones in North America." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 556, + 554, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 556, + 554, + 592 + ], + "spans": [ + { + "bbox": [ + 313, + 556, + 554, + 592 + ], + "type": "text", + "content": "Pool2 — Corrupted ground truth. We apply one of the following operations on the ground-truth wireframes from Pool 1 – examples are shown in Fig. 3:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 314, + 594, + 554, + 712 + ], + "type": "list", + "angle": 0, + "index": 24, + "blocks": [ + { + "bbox": [ + 314, + 594, + 553, + 630 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 594, + 553, + 630 + ], + "spans": [ + { + "bbox": [ + 314, + 594, + 553, + 630 + ], + "type": "text", + "content": "- (deform\\_{low, medium, high}) Split the edge into several edges and perturb the positions of the vertices without breaking the topology." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 631, + 554, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 631, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 631, + 554, + 689 + ], + "type": "text", + "content": "- (perturb\\_{low, medium, high}) Split the vertex into several ones. Each of the new vertices is randomly shifted from the ground truth. If the original vertex was connected to multiple neighbors, randomly decide which of the new vertices is connected to which neighbors." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 314, + 689, + 553, + 712 + ], + "type": "text", + "content": "- (add\\_{low, medium, high}) Randomly add wrong edges to the model." + } + ] + } + ], + "index": 23 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26945" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 294, + 187 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 294, + 187 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 294, + 187 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 294, + 187 + ], + "type": "image", + "image_path": "2e5b3dde039e5a06620d9d5709c9db432835955cd98a4f6ea88a06d78ed79353.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 193, + 296, + 248 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 296, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 296, + 248 + ], + "type": "text", + "content": "Figure 3. Examples of corrupted ground truth wireframes, used for wireframe ranking. Left to right: GT, deformed edges (deform_medium), vertex duplication and random movement (perturb_medium), edge addition (add_low), edge deletion(remove_low)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 257, + 296, + 282 + ], + "type": "text", + "content": "- (remove\\_{low, medium, high}) Randomly delete some of the vertices and all the edges connected to them." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "content": "Unit-tests & Desired properties of dissimilarity scores: In addition to being aligned with human judgment, we also propose a set of \"unit-tests\" for the metrics, which we believe are reasonable, and check if the dissimilarity scores satisfy these requirements. We design tests for the formal properties of mathematical metrics as well as additional properties relevant to evaluating dissimilarity in structured reconstruction tasks: For example, if wrong edges " + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "content": " are added to the GT model, the metric should score the resulting wireframe lower than if " + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "inline_equation", + "content": "E_{1}" + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "content": " or " + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "inline_equation", + "content": "E_{2}" + }, + { + "bbox": [ + 55, + 285, + 296, + 417 + ], + "type": "text", + "content": " are added separately." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "text", + "content": "Identity of Indiscernibles: This property ensures that identical inputs receive a dissimilarity score of zero, indicating perfect similarity. For any reconstruction " + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "text", + "content": ", a metric " + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "text", + "content": " satisfies this property if " + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "inline_equation", + "content": "d(x,x) = 0" + }, + { + "bbox": [ + 55, + 421, + 296, + 471 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "text", + "content": "Symmetry: A symmetric metric produces the same dissimilarity score regardless of the order of the inputs. For reconstructions " + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "text", + "content": ", a metric satisfies symmetry if " + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "inline_equation", + "content": "d(x, y) = d(y, x)" + }, + { + "bbox": [ + 55, + 473, + 296, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": "Triangle Inequality: The triangle inequality ensures that for any three reconstructions " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "x, y" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": ", the dissimilarity between " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": " is less than or equal to the sum of dissimilarities between " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": ". This relationship is expressed as " + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "inline_equation", + "content": "d(x, z) \\leq d(x, y) + d(y, z)" + }, + { + "bbox": [ + 55, + 525, + 296, + 586 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 296, + 674 + ], + "type": "text", + "content": "Monotonicity: This property describes how the dissimilarity score behaves when components (such as vertices or edges) are removed from a reconstruction. A metric satisfies monotonicity if the dissimilarity score does not increase when wrong vertices or edges are deleted. Similarly, the dissimilarity must not increase when correct vertices or edges are added." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Quasi-proportionality: This property holds when the metric changes smoothly under perturbations. This is evaluated by moving random vertices with small increments and" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 192 + ], + "type": "text", + "content": "checking the variance of the differences in the score. We use the following perturbations to simulate better or worse reconstructions: (i) remove correct edges from the ground truth wireframe; (ii) add wrong edges to the ground truth wireframe; (iii) disconnect ground truth edges; (iv) remove correct vertices; (v) move ground truth vertices to the wrong location. For every perturbation, we apply it 10 times and declare an example monotonic if it is strictly increasing (or decreasing as appropriate) for those continuous 10 perturbations." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 201, + 370, + 214 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 201, + 370, + 214 + ], + "spans": [ + { + "bbox": [ + 314, + 201, + 370, + 214 + ], + "type": "text", + "content": "3. Metrics" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 221, + 470, + 234 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 221, + 470, + 234 + ], + "spans": [ + { + "bbox": [ + 313, + 221, + 470, + 234 + ], + "type": "text", + "content": "The following metrics are considered:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 236, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 236, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 236, + 555, + 499 + ], + "type": "text", + "content": "WED - Wireframe Edit Distance was proposed by Liu et al. [24] as an extension of the Graph Edit Distance (GED) [32]. GED quantifies the distance between two graphs as the minimum number of elementary operations (inserting and deleting edges and vertices) required to transform one graph into another. WED extends this to wireframes (graphs with node positions and edge lengths) and proposes a cheap approximation to the NP-Hard problem of computing the optimal sequence of edits. Concretely, an assignment is first computed between the predicted and ground-truth vertices, and a cost is paid proportional to the distance between matched vertices. Next, unmatched vertices are deleted, and missing vertices are inserted (paying a cost proportional to the number of inserted/deleted vertices). Finally, given the vertex assignments, missing edges are inserted and extra edges deleted, paying a cost proportional to their length. In order to use WED, one needs to decide on the cost of insertion and deletion of the vertices, as well as the order of operations, and method of computing vertex/edge assignment. WED was used to determine the winner in the Building3D [33] and " + }, + { + "bbox": [ + 313, + 236, + 555, + 499 + ], + "type": "inline_equation", + "content": "S^2 3DR" + }, + { + "bbox": [ + 313, + 236, + 555, + 499 + ], + "type": "text", + "content": " [22] CVPR Challenges." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "spans": [ + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": "ECD - Edge Chamfer Distance. ECD is commonly used in structured reconstruction papers [10, 16, 17]. We consider a family of chamfer-like metrics between two point sets " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 502, + 555, + 562 + ], + "type": "text", + "content": " sampled from wireframe edges. The general form is:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 342, + 568, + 555, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 568, + 555, + 587 + ], + "spans": [ + { + "bbox": [ + 342, + 568, + 555, + 587 + ], + "type": "interline_equation", + "content": "d (A, B) := \\inf _ {\\pi_ {A B}: A \\rightarrow B} \\mathbb {E} _ {a \\in A} [ f (a, \\pi_ {A B} (a)) ], \\tag {1}", + "image_path": "f15c53ba54a392bb11aa27dcea315a11bd7734c9715a279c2721f653916b9c16.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{AB}" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": " represents an assignment from elements in " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": " to elements in " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": " is typically an " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": " norm of the difference between the inputs. Different constraints on " + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "inline_equation", + "content": "\\pi_{AB}" + }, + { + "bbox": [ + 313, + 593, + 554, + 639 + ], + "type": "text", + "content": " yield different metrics:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 642, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": "- The classical chamfer distance corresponds to " + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "inline_equation", + "content": "\\pi_{AB}(a) = \\arg \\min_{b\\in B}f(a,b)" + }, + { + "bbox": [ + 313, + 642, + 553, + 666 + ], + "type": "text", + "content": ", i.e. nearest neighbor matching." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": "- The most constrained version requires " + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\pi_{AB}" + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": " to be a bijective matching, which can be computed via the Hungarian algorithm and is equivalent to the Earth Mover's Distance when " + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": " is the " + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "inline_equation", + "content": "\\ell_p" + }, + { + "bbox": [ + 313, + 666, + 553, + 714 + ], + "type": "text", + "content": " norm." + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 318, + 757 + ], + "type": "text", + "content": "26946" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 70, + 523, + 285 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 523, + 285 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 523, + 285 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 523, + 285 + ], + "type": "image", + "image_path": "f4bfc0a23c27ee08617745a209476d2699bdf48e4855aae9fd8b3e53671206f8.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 555, + 319 + ], + "type": "text", + "content": "Figure 4. Annotator agreement (all pairs). Left to right: annotator agreement with each other, the learned metric, handcrafted metrics, and VLMs. Annotators background: A-K - 3D modellers, Des[0-2] - designers, CV[0-3] - computer vision engineers. Best zoom-in." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 401 + ], + "type": "text", + "content": "Length Weighted Spectral Graph Distance - SD incorporates both topological and geometric information by framing graph (wireframe) distance in terms of distances between the spectra of weighted graph Laplacians. We measure the spectral distance using the 2-Wasserstein metric between the eigenvalue distributions:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 99, + 411, + 295, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 411, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 99, + 411, + 295, + 426 + ], + "type": "interline_equation", + "content": "S D \\left(G _ {1}, G _ {2}\\right) := W _ {2} \\left(\\lambda \\left(L _ {1}\\right), \\lambda \\left(L _ {2}\\right)\\right), \\tag {2}", + "image_path": "d3b5b1de50f74240e637a81c756cf3ec24d6621368477d309015b0a0ce10a8e3.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "inline_equation", + "content": "\\lambda (L)" + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "text", + "content": " denotes the spectrum of the Laplacian " + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "text", + "content": ". For a graph " + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "inline_equation", + "content": "G = (V,E)" + }, + { + "bbox": [ + 55, + 435, + 296, + 460 + ], + "type": "text", + "content": ", the weighted graph Laplacian is defined:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 148, + 472, + 294, + 484 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 148, + 472, + 294, + 484 + ], + "spans": [ + { + "bbox": [ + 148, + 472, + 294, + 484 + ], + "type": "interline_equation", + "content": "L := D - A \\tag {3}", + "image_path": "6eadfd26bf4e5497a4017452b8d20e850cf3f8020c61cd06915275e6283e9799.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " is the weighted degree matrix " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "(|V|\\times |V|" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " diagonal matrix with each diagonal entry containing the sum of the lengths of edges incident to that vertex), and " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "A" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " is the weighted adjacency matrix " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "(|V|\\times |V|" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "A_{ij} = \\left\\| \\mathrm{coord}(V_i) - \\mathrm{coord}(V_j)\\right\\| _2" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " if " + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "inline_equation", + "content": "(i,j)\\in E" + }, + { + "bbox": [ + 55, + 491, + 296, + 552 + ], + "type": "text", + "content": " and 0 otherwise)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 296, + 639 + ], + "type": "text", + "content": "**Corner and Edge Metrics.** We also compute precision, recall, and F1 scores for both corners and edges. For corners, we consider a prediction correct if it lies within a distance threshold of a ground truth corner. For edges, we use the Hausdorff distance between line segments to determine matches. These metrics provide an intuitive measure of the topological accuracy of the predicted wireframes." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 641, + 296, + 714 + ], + "type": "text", + "content": "Hausdorff Distance measures the maximum of minimal distances between two sets of points. For wireframes, we sample points along the edges and compute the Hausdorff distance between these point sets, providing a measure of geometric similarity that considers both corner positions and edge geometry." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 330, + 555, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 330, + 555, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 330, + 555, + 474 + ], + "type": "text", + "content": "Intersection over Union is a popular metric in a wide range of fields, e.g. segmentation, tracking, object detection. However, it is rarely used to assess the quality of wireframe reconstructions. We extend the definition of the wireframe as a set of cylinders with a fixed radius (the only hyperparameter of this metric) and define the metric as an IoU between two sets of cylinders, given by two wireframe reconstructions that need to be compared. An approximation via point sampling is considered: sample random points from both sets of cylinders and compute the average number of times when the point falls inside of both sets of cylinders. The Jaccard distance is reported between two sets." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 479, + 555, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 555, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 555, + 540 + ], + "type": "text", + "content": "Visual-and-language models could potentially be used for this task. We consider 4o, o1 [13], Grok 2 [34], qwen-2.5 [18], pixtral 12b [2], claude 3.5 and 3.7 [3], gemini 2.0 and gemini 2.0-flash [12] models via OpenRouter [28]. The prompts are provided in the supplementary." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "Learned metric. We also explore how to distill the human annotations directly into a metric. To this end, we propose learning a metric with transfer learning. First, reconstruction and ground-truth wireframes are plotted in 3D and rendered from a canonical viewpoint (denoted " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "r_i" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": "). Then, DiNOv2 [29] features are extracted from those renderings and an MLP-based regression head is then trained to regress scores based on the extracted features " + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "inline_equation", + "content": "(g(r_i))" + }, + { + "bbox": [ + 313, + 545, + 556, + 715 + ], + "type": "text", + "content": ". A Bradley-Terry [7] probability model is assumed and pairwise annotations are used to supervise the training by minimizing a binary cross-entropy loss with a batch size of 16. 10-fold cross validation splits the data such that the sets of ground truth structures and reconstruction methods used in the training and test sets are disjoint. We observe average" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "26947" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 72, + 216, + 125 + ], + "blocks": [ + { + "bbox": [ + 70, + 72, + 216, + 125 + ], + "lines": [ + { + "bbox": [ + 70, + 72, + 216, + 125 + ], + "spans": [ + { + "bbox": [ + 70, + 72, + 216, + 125 + ], + "type": "image", + "image_path": "cd16ca1a1682d7cb9fd311975cf52469089e2c142b4000ca4570147193ff0f0c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 220, + 72, + 367, + 125 + ], + "blocks": [ + { + "bbox": [ + 220, + 72, + 367, + 125 + ], + "lines": [ + { + "bbox": [ + 220, + 72, + 367, + 125 + ], + "spans": [ + { + "bbox": [ + 220, + 72, + 367, + 125 + ], + "type": "image", + "image_path": "9a30fc9d39fb192c1fb636582b4b3691271fa64242e14fb78ead7d8c7d4c9159.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 370, + 72, + 552, + 126 + ], + "blocks": [ + { + "bbox": [ + 370, + 72, + 552, + 126 + ], + "lines": [ + { + "bbox": [ + 370, + 72, + 552, + 126 + ], + "spans": [ + { + "bbox": [ + 370, + 72, + 552, + 126 + ], + "type": "image", + "image_path": "2cc93e652f50491651771a5f51639c840e72b6ab17c35baa0a4e04175e6810c1.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 126, + 544, + 138 + ], + "lines": [ + { + "bbox": [ + 64, + 126, + 544, + 138 + ], + "spans": [ + { + "bbox": [ + 64, + 126, + 544, + 138 + ], + "type": "text", + "content": "Figure 5. Probability of selecting wrong \"winner\" depending on number of raters (left), individual accuracy (center), win rate (right)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "text", + "content": "accuracy across the folds of " + }, + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "inline_equation", + "content": "76\\%" + }, + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "text", + "content": " where a prediction is considered correct if " + }, + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "inline_equation", + "content": "g(r_{\\mathrm{winner}}) > g(r_{\\mathrm{loser}})" + }, + { + "bbox": [ + 55, + 144, + 295, + 170 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 179, + 137, + 193 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 179, + 137, + 193 + ], + "spans": [ + { + "bbox": [ + 55, + 179, + 137, + 193 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 202, + 296, + 322 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 202, + 296, + 322 + ], + "spans": [ + { + "bbox": [ + 55, + 202, + 296, + 322 + ], + "type": "text", + "content": "Human wireframe ranking and its consistency. To determine which of the metrics under consideration are most appropriate, we employ three groups of people to provide pseudo-ground-truth rankings of the solutions. The first and biggest group is made up of human 3D modeling experts who professionally create CAD models of objects from photos. The second group is computer vision researchers, and finally, the third group is people who do not work with 3D modeling in their daily lives (designers). There are 11, 4, and 3 people in those groups respectively." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": "Annotators were shown pairs of reconstructions of the same structure, and asked to specify which reconstruction most closely matched the superimposed corresponding wireframe. An example of the user interface is shown in Fig. 2. Annotators were able to zoom, pan, and rotate, allowing them to examine the solutions from all sides if needed; the viewpoint of both solutions is synchronized to ease the comparison. All 27 methods are compared exhaustively by every rater, meaning " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\binom{27}{2} = 351" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": " method pairs per house " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "\\times 10" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": " houses " + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "inline_equation", + "content": "= 3510" + }, + { + "bbox": [ + 55, + 322, + 296, + 466 + ], + "type": "text", + "content": " pairs for every rater. Most raters rate a few more pairs because of the self-consistency checks. See the Suppl. for the additional information." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 468, + 296, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 505 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 505 + ], + "type": "text", + "content": "Rater Reliability. We quantify rater reliability using two complementary methods: self-consistency and correctness on synthetic samples." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 507, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 507, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 507, + 295, + 614 + ], + "type": "text", + "content": "Correctness on Synthetic Samples. We introduced synthetic pairs of wireframes with known ground-truth rankings based on systematically applied \"corruptions.\" Each corruption type featured \"Low,\" \"Medium,\" and \"High\" severity levels. We treat the \"low\" vs \"high\" per each corruption type to be obvious enough that if annotators rank them differently, it can be treated as a labeling error, e.g. because of wrong clicks. Average rater accuracy on these \"easy\" pairs is " + }, + { + "bbox": [ + 55, + 507, + 295, + 614 + ], + "type": "inline_equation", + "content": "98.3\\%" + }, + { + "bbox": [ + 55, + 507, + 295, + 614 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Self-consistency. We assessed intra-rater reliability by measuring how consistently annotators rated repeated pairs of wireframes, occasionally reversing pair order to mitigate order biases. These repeated evaluations constituted only a small subset of the total ratings (self-consistency checks are performed with " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": " probability). The average self-consistency score is " + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "inline_equation", + "content": "89.4\\%" + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": ", which could be partially attributed to the labeling mistakes and partially to the chang-" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 144, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 144, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 313, + 144, + 553, + 169 + ], + "type": "text", + "content": "ing preferences during the annotation process, which we experienced ourselves." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": "Do we label enough? Under the assumption that there is some latent \"correct\" winner in any given pair, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " raters each independently select this \"correct\" winner with probability " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ", we can compute the probability that the majority is \"wrong\" (see Fig. 5). Therefore, the estimated panel error rate per pair is " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\approx 1\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " for 11 (expert) raters, and " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "0.25\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " for 17 (all) raters (assuming a significant " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " individual error rate). We also analyze the stability of our results and present adequacy analysis for the number of raters, comparisons, and houses. For each, we sweep a range of subsample sizes and resample 500 times at each size. For each subsample of a given size, we compute ranking implied by the win rates for that subset and rank correlation (Kendall " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ") between the subset ranks and the rankings using the full dataset. We then construct a " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " Bootstrap CI across the 500 iterates for " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ". The minimum number of raters/comparisons/houses needed for " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\tau \\geq 0.95" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ": comparisons " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\geq 3350" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ", houses " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\geq 4" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": ", raters " + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "inline_equation", + "content": "\\geq 8" + }, + { + "bbox": [ + 313, + 177, + 555, + 392 + ], + "type": "text", + "content": " (in all cases we have more than that)." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 400, + 554, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 400, + 554, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 400, + 554, + 449 + ], + "type": "text", + "content": "Finding agreement. We compute an agreement score for each pair for annotators using the following simple rule: the same ranking gets 1 point, decisive ranking vs \"equal\" gets 0.5 points, and the opposite ranking gets zero." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 454, + 554, + 491 + ], + "type": "text", + "content": "The agreement table between human annotators, metrics, and VLMs is shown in Fig. 4. The agreement table with all \"equal\" rankings excluded, is shown in the supplementary." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 322, + 508, + 392, + 519 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 508, + 392, + 519 + ], + "spans": [ + { + "bbox": [ + 322, + 508, + 392, + 519 + ], + "type": "text", + "content": "Observation 4.1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 525, + 547, + 609 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 525, + 547, + 609 + ], + "spans": [ + { + "bbox": [ + 321, + 525, + 547, + 609 + ], + "type": "text", + "content": "When accounting for ties, there is a moderate global consensus among the human annotators; however, they form two distinct clusters that do not seem to depend on the annotator's background. One group assigns more weights to the edge accuracy – correlated with edge F1 and Jaccard distance, and the second – vertex accuracy, correlated with corner F1 score." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": "The average agreement score is around " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "80\\%" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": " across all annotators, but within clusters, it increases to " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "85\\%" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ". When annotators are decisive (select one of the reconstructions as clearly better rather than selecting \"equal\"), then the average agreement increases to " + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "inline_equation", + "content": "91\\%" + }, + { + "bbox": [ + 313, + 630, + 554, + 714 + ], + "type": "text", + "content": ", and no clusters are observed. Agreement with the metrics (middle part of Fig. 4) suggests an explanation of preferences." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26948" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 66, + 70, + 184, + 133 + ], + "blocks": [ + { + "bbox": [ + 66, + 70, + 184, + 133 + ], + "lines": [ + { + "bbox": [ + 66, + 70, + 184, + 133 + ], + "spans": [ + { + "bbox": [ + 66, + 70, + 184, + 133 + ], + "type": "image", + "image_path": "e6843ed5aef9bb1c8b830636bbcb71b9eb16400f24f1be86d9c4ac2962d41752.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 140, + 555, + 163 + ], + "type": "text", + "content": "Figure 6. Metric ranking by agreement with group in average. Left to right: the first group of raters with more attention to vertices, the second group of raters with more attention to edges, computer vision engineers, and designers." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 187, + 70, + 304, + 133 + ], + "blocks": [ + { + "bbox": [ + 187, + 70, + 304, + 133 + ], + "lines": [ + { + "bbox": [ + 187, + 70, + 304, + 133 + ], + "spans": [ + { + "bbox": [ + 187, + 70, + 304, + 133 + ], + "type": "image", + "image_path": "0bc83e8298cab77cd938ad2f41165a811bedf6e8fc0255ad29ed655fc1935197.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 309, + 70, + 426, + 132 + ], + "blocks": [ + { + "bbox": [ + 309, + 70, + 426, + 132 + ], + "lines": [ + { + "bbox": [ + 309, + 70, + 426, + 132 + ], + "spans": [ + { + "bbox": [ + 309, + 70, + 426, + 132 + ], + "type": "image", + "image_path": "d55c4ce7d380ca9cda75de73148c2c50da3599ab1b4f5cf6d151685e81b27794.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 429, + 70, + 548, + 132 + ], + "blocks": [ + { + "bbox": [ + 429, + 70, + 548, + 132 + ], + "lines": [ + { + "bbox": [ + 429, + 70, + 548, + 132 + ], + "spans": [ + { + "bbox": [ + 429, + 70, + 548, + 132 + ], + "type": "image", + "image_path": "87a7ae90818a204364d3f5f1d6413dd6ce96f8aef66c8b1b5bb8ee88cd12cb4e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 63, + 169, + 134, + 180 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 169, + 134, + 180 + ], + "spans": [ + { + "bbox": [ + 63, + 169, + 134, + 180 + ], + "type": "text", + "content": "Observation 4.2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 186, + 290, + 245 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 186, + 290, + 245 + ], + "spans": [ + { + "bbox": [ + 62, + 186, + 290, + 245 + ], + "type": "text", + "content": "Human annotators pay more attention to correct parts of the reconstruction than the incorrect parts. Regardless of whether edges or vertices are considered, recall metrics agree more with human preferences than precision ones." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 251, + 296, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 296, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 296, + 346 + ], + "type": "text", + "content": "Cluster 1 (raters A-G, CV, Des1-2) mostly correlates with corner-based metrics, such as corner recall and corner F1-score, whereas Cluster 2 (raters H-K, Des0) correlates more with edge-based metrics, such as edge recall, edge F1-score and Jaccard distance. The second difference between clusters is that Cluster 2 is more likely to give an \"equal\" score for the low-quality reconstruction, whereas Cluster 1 tried to rank reconstruction more decisively." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 347, + 296, + 371 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 347, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 296, + 371 + ], + "type": "text", + "content": "The metrics rankings w.r.t. different groups of raters is shown in Fig. 8." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 63, + 374, + 134, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 374, + 134, + 385 + ], + "spans": [ + { + "bbox": [ + 63, + 374, + 134, + 385 + ], + "type": "text", + "content": "Observation 4.3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 392, + 288, + 427 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 392, + 288, + 427 + ], + "spans": [ + { + "bbox": [ + 62, + 392, + 288, + 427 + ], + "type": "text", + "content": "The average agreement with human preferences of the top handcrafted metrics does not vary significantly. WED-based scores correlate with annotators the least." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "text", + "content": "The supplementary material shows the full agreement table. Furthermore, the ranking setup changes the human preferences for different metrics; for example, Jaccard distance performs much better for the decisive pairs compared to all other metrics. WED with pre-registration, which was used in the " + }, + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "inline_equation", + "content": "S^23DR" + }, + { + "bbox": [ + 55, + 434, + 296, + 589 + ], + "type": "text", + "content": " challenge, shows the worst correlation and is just slightly better than random chance. Other flavors of the WED metric, namely WED_mnn and WED_AP (used in Building3D Challenge), perform better but still worse than the rest. Graph structure metrics are in the middle. We also consider the agreement with VLMs. Despite initial success with individual examples, in our tests they did not perform meaningfully better than chance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 63, + 597, + 134, + 608 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 597, + 134, + 608 + ], + "spans": [ + { + "bbox": [ + 63, + 597, + 134, + 608 + ], + "type": "text", + "content": "Observation 4.4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 62, + 614, + 288, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 614, + 288, + 674 + ], + "spans": [ + { + "bbox": [ + 62, + 614, + 288, + 674 + ], + "type": "text", + "content": "VLMs do not show significant agreement with human preferences in the wireframe ranking. The only exceptions are OpenAI models, as well as Grok2; yet those are only slightly better than chance. VLMs agree the most with the WED metrics family." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Finding the best reconstruction. To determine if there is a single \"quality\" factor which explains the human ex" + } + ] + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 339, + 167, + 529, + 277 + ], + "blocks": [ + { + "bbox": [ + 339, + 167, + 529, + 277 + ], + "lines": [ + { + "bbox": [ + 339, + 167, + 529, + 277 + ], + "spans": [ + { + "bbox": [ + 339, + 167, + 529, + 277 + ], + "type": "image", + "image_path": "f0b61420cc352e3b4d1eb7b264dafbe0d3785aef2ef6c9250ad1c81b4cfe7e76.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 340, + 279, + 528, + 290 + ], + "lines": [ + { + "bbox": [ + 340, + 279, + 528, + 290 + ], + "spans": [ + { + "bbox": [ + 340, + 279, + 528, + 290 + ], + "type": "text", + "content": "Figure 7. Metric ranking by agreement with VLMs" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 301, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 555, + 361 + ], + "type": "text", + "content": "pert judgments, we employ three distinct approaches to map these pairwise comparisons to a single ranking of the methods: simple win rates, a Bradley-Terry probability model, and a factor analysis based approach. While the methods differ, they point to highly concordant conclusions." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 322, + 370, + 392, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 370, + 392, + 380 + ], + "spans": [ + { + "bbox": [ + 322, + 370, + 392, + 380 + ], + "type": "text", + "content": "Observation 4.5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 321, + 387, + 547, + 423 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 387, + 547, + 423 + ], + "spans": [ + { + "bbox": [ + 321, + 387, + 547, + 423 + ], + "type": "text", + "content": "Human raters tend to rank equally all solutions that are below some quality threshold. Having no solution often ranks better, compared to a totally wrong reconstruction." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 437, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 641 + ], + "type": "text", + "content": "Simple Win Rate The first approach is similar to chess scoring – first a win-count table is computed. For each rated pair, each method receives 1 point for a win, 0.5 for a tie, and 0 for a loss. This scoring does not break ties but distributes the points evenly. Results with selected reconstruction methods are shown in Fig. 8. Consistent with the metric-human agreements, recall plays a more important role, and the wireframes with perfect recall – \"add_low\" and \"add_med\" are among the leaders in all groups. In other words, extra (erroneous) edges are considered less of an issue when compared to missing an edge or a vertex – the \"remove_*\" family. One possible explanation resides in the information contained in the reconstruction: if all correct edges are present, we may identify and remove any erroneous ones. However, in many cases, inferring the exact position of an edge or vertex is not possible given the reconstruction alone if key information is missing." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 554, + 700 + ], + "type": "text", + "content": "The next set of solutions contains roughly correct but slightly noisy reconstructions - \"perturb_*\" and \"deform*\" family. One of the best solutions from the S23DR Challenge ranked better than highly deformed ground truth but worse than the less invasive corruptions." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 553, + 713 + ], + "type": "text", + "content": "The rest of the reconstructions get almost equal scores" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26949" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "due to the high proportion of draws and lack of wins among themselves." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "text", + "content": "Bradley-Terry Model We have also modeled the quality of each solution using a Bradley-Terry (BT) [7, 11, 14, 35] preference model on the expressed preferences of the annotators (using the BT-Abilities as scores). The Bradley-Terry model defines the probability that item " + }, + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "text", + "content": " is preferred over item " + }, + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 100, + 296, + 171 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 130, + 173, + 295, + 197 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 173, + 295, + 197 + ], + "spans": [ + { + "bbox": [ + 130, + 173, + 295, + 197 + ], + "type": "interline_equation", + "content": "P (i > j) = \\frac {a _ {i}}{a _ {i} + a _ {j}}, \\tag {4}", + "image_path": "b0ff6064ebb3f12c4d859e55f2e31ff769438d9487aab228755655dfeb23f15f.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "spans": [ + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "type": "inline_equation", + "content": "a_{i}, a_{j}" + }, + { + "bbox": [ + 55, + 204, + 295, + 227 + ], + "type": "text", + "content": " are positive real numbers representing the latent strength of each item." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "content": "Following standard practice, we reparameterize these latent strengths as exponentials of real-valued parameters " + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "content": ", giving " + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "inline_equation", + "content": "a_{i} = e^{\\theta_{i}}" + }, + { + "bbox": [ + 55, + 229, + 296, + 265 + ], + "type": "text", + "content": ". This yields:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 63, + 275, + 295, + 312 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 275, + 295, + 312 + ], + "spans": [ + { + "bbox": [ + 63, + 275, + 295, + 312 + ], + "type": "interline_equation", + "content": "P (i > j) = \\frac {e ^ {\\theta_ {i}}}{e ^ {\\theta_ {i}} + e ^ {\\theta_ {j}}} = \\frac {1}{1 + e ^ {- (\\theta_ {i} - \\theta_ {j})}} = \\sigma \\left(\\theta_ {i} - \\theta_ {j}\\right), \\tag {5}", + "image_path": "db841298ca5e8d51525e9796ab9f5b9181f2a78c0f9f024d3c83b4aa41ab9cf1.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "spans": [ + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "type": "inline_equation", + "content": "\\sigma (x)" + }, + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "type": "text", + "content": " is the sigmoid function " + }, + { + "bbox": [ + 55, + 325, + 263, + 339 + ], + "type": "inline_equation", + "content": "\\sigma (x) = \\frac{1}{1 + e^{-x}}" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "text", + "content": "This formulation is further generalized by introducing a scale parameter " + }, + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "text", + "content": " and offset " + }, + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "inline_equation", + "content": "o" + }, + { + "bbox": [ + 55, + 339, + 296, + 361 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 121, + 373, + 295, + 399 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 373, + 295, + 399 + ], + "spans": [ + { + "bbox": [ + 121, + 373, + 295, + 399 + ], + "type": "interline_equation", + "content": "p _ {i j} = \\sigma \\left(\\frac {\\theta_ {i} - \\theta_ {j}}{s} + o\\right). \\tag {6}", + "image_path": "3441b7f920b9db0e37e462026fe28d5987bc4e131af17695f7e5a5639f3e1723.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": "With " + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "inline_equation", + "content": "s = 1" + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "inline_equation", + "content": "o = 0" + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": ", this reduces to the standard Bradley-Terry formulation. Alternatively, setting " + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "inline_equation", + "content": "s = 400" + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "inline_equation", + "content": "o = 800" + }, + { + "bbox": [ + 55, + 409, + 296, + 445 + ], + "type": "text", + "content": " yields the Elo scoring system familiar to chess players." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "text", + "content": "To estimate the latent abilities " + }, + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "text", + "content": ", we initialize each " + }, + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "inline_equation", + "content": "\\theta_{i}" + }, + { + "bbox": [ + 55, + 446, + 296, + 506 + ], + "type": "text", + "content": " by sampling from an independent Gaussian distribution and then iteratively minimize the expectation of the following binary cross-entropy loss using stochastic gradient descent (SGD) with the Adam optimizer:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 69, + 517, + 295, + 532 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 517, + 295, + 532 + ], + "spans": [ + { + "bbox": [ + 69, + 517, + 295, + 532 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathbb {E} _ {(i, j)} \\left[ - y \\log \\left(p _ {i j}\\right) - (1 - y) \\log \\left(1 - p _ {i j}\\right) \\right], \\tag {7}", + "image_path": "3f4b74f64b958585800ca92ad6e1e6d19a56ce6350d74c31c0b2af0250340853.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "inline_equation", + "content": "y = 1" + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "content": " if item " + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "content": " was chosen over item " + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "inline_equation", + "content": "y = 0" + }, + { + "bbox": [ + 55, + 542, + 296, + 565 + ], + "type": "text", + "content": " otherwise." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Factor Analysis To investigate whether the data reflect a single underlying dimension of quality, we additionally peruse a factor analysis based approach. We form the methods-by-raters table " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "M_{kl}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " is the rate at which rater " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " chose method " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " when they saw it. We hypothesize the empirical log-odds of these win-rates (rate " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " wins according to " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "), " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta = \\log \\frac{M}{1 - M}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " possess a low-rank structure (rank one in the ideal case); this would indicate a single dominant factor (\"quality\") governing outcomes. We apply singular value decomposition (SVD) to factorize " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta = U\\Sigma V^T" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and extract the first left singular vector of " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " containing the estimated quality scores." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 316, + 71, + 561, + 315 + ], + "blocks": [ + { + "bbox": [ + 316, + 71, + 561, + 315 + ], + "lines": [ + { + "bbox": [ + 316, + 71, + 561, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 71, + 561, + 315 + ], + "type": "table", + "html": "
MethodEmpirical Win RateImplied Win Rate (BT)Implied Win Rate (Elo)BT AbilityElo ScoreQuality Factor
add_low0.890.890.892.7919370.03
add_med0.860.860.862.4717690.02
perturb_med0.850.850.852.3317390.02
add_high0.820.820.822.041604-0.02
perturb_low0.790.790.791.791510-0.01
remove_low0.790.780.791.711498-0.02
perturb_high0.670.670.680.831144-0.09
deform_med0.670.670.660.831107-0.09
deform_low0.660.660.660.781094-0.10
remove_high0.650.650.650.671077-0.10
remove_med0.630.640.640.601027-0.11
kc920.510.510.51-0.25698-0.18
Siromanec0.500.500.49-0.34645-0.19
deform_high0.500.500.50-0.34669-0.19
maximivashechkin0.390.390.39-1.01386-0.27
rozumden0.380.380.38-1.10373-0.28
kcml0.350.350.35-1.28291-0.26
rozumden0.340.340.33-1.35236-0.26
Ana-Geneva0.320.320.32-1.47221-0.25
pc2wf_retrain0.290.290.30-1.65157-0.23
Yurii0.290.290.29-1.63146-0.25
snuggler0.250.250.24-1.9215-0.25
baseline0.250.250.25-1.9119-0.25
Hunter-X0.220.220.22-2.10-43-0.25
TUM0.220.220.22-2.13-48-0.26
Fudan EDLAB0.210.210.21-2.18-76-0.26
pc2wf_pretrained0.200.200.20-2.28-117-0.25
", + "image_path": "feba01b554919a65425f707e7a0bdeda35c8f4461c22fd67071e371b35ff61f6.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 316, + 555, + 338 + ], + "lines": [ + { + "bbox": [ + 314, + 316, + 555, + 338 + ], + "spans": [ + { + "bbox": [ + 314, + 316, + 555, + 338 + ], + "type": "text", + "content": "Table 1. Win rates for selected wireframe, according to the pairwise human annotations and estimated Elo." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 322, + 355, + 392, + 366 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 355, + 392, + 366 + ], + "spans": [ + { + "bbox": [ + 322, + 355, + 392, + 366 + ], + "type": "text", + "content": "Observation 4.6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 321, + 372, + 547, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 372, + 547, + 431 + ], + "spans": [ + { + "bbox": [ + 321, + 372, + 547, + 431 + ], + "type": "text", + "content": "We find a Kendall correlation coefficient " + }, + { + "bbox": [ + 321, + 372, + 547, + 431 + ], + "type": "inline_equation", + "content": ">0.7" + }, + { + "bbox": [ + 321, + 372, + 547, + 431 + ], + "type": "text", + "content": " between the rankings implied by SVD and those implied by BT. This lends additional evidence to the hypothesis that there is a true \"quality\" factor driving the raters' views. The result is shown in Table 1." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 446, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 696 + ], + "type": "text", + "content": "Metric properties and \"unit tests\". We define a range of properties a good metric should have, implement tests for these properties, and report the results for all metrics. We use a dataset of 128 ground truth wireframes, which we disturb or alter and check the behavior of each metric. We report the percentage of wireframes where the property was valid for a given metric. Results are shown in Table 2. Most of the metrics pass the triangle inequality test, the identity of indiscernibles, and symmetry. None of the metrics is perfectly monotonic w.r.t. wireframe changes, which makes sense; for example, the precision metrics are insensitive to the number of predicted vertices/edges as long as each predicted element is aligned with an element of the ground truth. Hausdorff and WED with pre-registration pass the fewest tests, which is in line with their poor performance w.r.t. human preferences. Conversely - corner F1, edge F1, and Jaccard perform well and align well with the preferences of annotators. The spectral L1 distance and Chamfer edge distance are exceptions - the spectral distance scores well on properties, but not human alignment, and Chamfer one - vice versa." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 555, + 713 + ], + "type": "text", + "content": "Additional considerations. After speaking to the human" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "26950" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 68, + 547, + 149 + ], + "blocks": [ + { + "bbox": [ + 65, + 68, + 547, + 149 + ], + "lines": [ + { + "bbox": [ + 65, + 68, + 547, + 149 + ], + "spans": [ + { + "bbox": [ + 65, + 68, + 547, + 149 + ], + "type": "image", + "image_path": "f238762e1203070789b3fbd94d1ab6e9c8db6254fd8b9149a3906f5ee2b6feeb.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 203, + 153, + 406, + 164 + ], + "lines": [ + { + "bbox": [ + 203, + 153, + 406, + 164 + ], + "spans": [ + { + "bbox": [ + 203, + 153, + 406, + 164 + ], + "type": "text", + "content": "Figure 8. Methods scores according to different groups." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 78, + 165, + 533, + 407 + ], + "blocks": [ + { + "bbox": [ + 78, + 165, + 533, + 407 + ], + "lines": [ + { + "bbox": [ + 78, + 165, + 533, + 407 + ], + "spans": [ + { + "bbox": [ + 78, + 165, + 533, + 407 + ], + "type": "table", + "html": "
TestCornerEdgeWEDSpectralIoUHausdorffChamfer
PrecRecF1offsetPrecRecF1preregMNNnearestAPL1L2Jaccarddist.edge
Monotonic
Monotonic (wrong edges)0.020.010.021.001.000.001.000.000.000.000.980.630.140.010.090.00
Monotonic (deform/split)1.000.861.000.651.000.981.000.740.790.670.600.000.060.160.050.40
Monotonic (moving vertex)1.001.001.000.981.001.001.000.991.000.991.000.990.990.450.991.00
Monotonic (disconnect edges)0.520.000.500.310.000.000.000.261.001.001.000.000.011.000.000.02
Monotonic (delete vertices)0.001.001.001.000.001.001.000.751.001.001.000.900.330.670.140.99
Monotonic (delete edges)0.000.000.001.000.001.001.001.001.001.001.001.000.311.000.070.83
Identity
Identity of indiscernibles1.001.001.001.001.001.001.000.021.001.001.01.001.001.000.000.00
Near identity of indiscernibles1.001.001.001.001.001.001.001.001.001.001.01.001.001.000.000.00
Symmetry
Symmetry (0 mean, weighted)1.001.001.000.881.001.001.000.440.440.440.440.440.901.000.841.00
Near symmetry (0 mean, weighted)1.001.001.000.871.001.001.000.580.540.540.450.900.921.001.001.00
Symmetry (shift, weighted)1.001.001.001.001.001.001.000.880.990.920.980.991.001.000.821.00
Near symmetry (shift, weighted)1.001.001.001.001.001.001.001.001.000.920.991.001.001.001.001.00
Quasi-proportionality
Quasi-proportionality (shift, far)0.000.000.000.000.000.000.001.000.000.000.001.000.720.050.000.00
Quasi-proportionality (shift, close)0.000.000.000.000.000.000.001.000.160.160.911.000.720.060.000.00
Triangle ineq
Triangle ineq. (rand other)0.991.001.001.001.000.981.001.001.000.990.950.920.971.001.000.96
Triangle ineq. (add noise)1.001.001.000.911.001.001.000.840.810.831.000.900.691.001.001.00
Triangle ineq. (del1/del2)1.001.001.001.001.001.001.000.900.980.981.001.000.661.001.000.43
Pass count11/1711/1712/1711/1712/1713/1714/178/1710/1710/1713/1713/178/1711/176/178/17
", + "image_path": "002363555356aeda9edf2a05284ddb96f85979b9b3bceb4258a5d9a6417964c9.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 77, + 416, + 531, + 427 + ], + "lines": [ + { + "bbox": [ + 77, + 416, + 531, + 427 + ], + "spans": [ + { + "bbox": [ + 77, + 416, + 531, + 427 + ], + "type": "text", + "content": "Table 2. Properties \"unit-test\" results for metrics, percentage of tests passed. The test is passed if the result is " + }, + { + "bbox": [ + 77, + 416, + 531, + 427 + ], + "type": "inline_equation", + "content": "\\geq {90}\\%" + }, + { + "bbox": [ + 77, + 416, + 531, + 427 + ], + "type": "text", + "content": " (black)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 438, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 438, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 438, + 295, + 640 + ], + "type": "text", + "content": "annotators, we note the following. First, the 3D reconstruction experts are considering how the estimated reconstruction could help them in 3D modeling. There are two main approaches to using such help. The first (in case of noisy reconstructions) is to use the wireframe as a guide to create their own clean reconstruction. The second, if the reconstruction is already close enough, is to fix the wireframe to produce the final result. The Wireframe Edit Distance is designed to estimate how costly it would be to modify the predicted wireframe to match the ground truth. The issue with it is the set of operations – WED only considers vertex/edge deletion/insertion and vertex movement. In practice, one could fit a single edge to multiple noisy ones, bulk-delete a lot of wrong edges or vertices, and apply a rigid transform on the whole model. All of these operations are commonly used in 3D editing software, and WED could benefit from them." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 689 + ], + "type": "text", + "content": "Finally, if the wireframe is totally wrong, then it is better to have no reconstruction at all and start from scratch. In this \"low quality regime\", most human annotators see no difference between reconstructions if both are wrong." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "Considering usage in competition and benchmarks, we would recommend the use of F1-score, despite the fact that" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 438, + 554, + 498 + ], + "type": "text", + "content": "recall-based metrics are more in line with human preferences, as it is less easy to game. For example, a dense grid of vertices and edges could score perfectly on recall but be useless in practice and score poorly on precision-based metrics." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 388, + 517 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 312, + 522, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 522, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 312, + 522, + 555, + 713 + ], + "type": "text", + "content": "We have studied how human preferences in structured reconstruction evaluation are explained via a wide range of metrics. We show that human preferences can be learned from a small number of examples by transferring from pretrained models which can subsequently be used to score unseen reconstructions. However, we conclude that additional study is warranted prior to relying on such learned metrics as the sole adjudication mechanism for competitions (especially those with strong incentives), because of the potential for reward hacking, gradient-based adversarial attacks, and the like. Based on our study, we recommend using a combination of edge-based (edge F1 or Jaccard score) and corner-based metrics (F1) for benchmarks and competitions. They better explain human preferences in ranking structured reconstructions than more complex and fragile graph-based metrics such as WED or spectral distances." + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "26951" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 205 + ], + "type": "text", + "content": "Acknowledgements We thank Sherwin Dale Sotto, Rosalie Cellacay, John Carlo Castro, Johsua Vernon Awog, Romeo Sapitanan, Francis Angel Viloria, Joenard Sarmiento, Danica B. Mercado, Zander Hechanova, Michael James Dulay, Joshua Muñoz, Jonah Franz R. De Leon, Viktor Rusov, Olha Mishkina, Yihe Wang and Tyler Nguyen for their help with wireframe annotations. We also thank Tolga Birdal, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, and Ilke Demir for helpful discussions and dataset work at the early stage of the paper. Dmytro Mishkin is supported by MPO 60273/24/21300/21000 CEDMO 2.0 NPO project." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 218, + 115, + 230 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 218, + 115, + 230 + ], + "spans": [ + { + "bbox": [ + 56, + 218, + 115, + 230 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 238, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 61, + 238, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 238, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 61, + 238, + 296, + 281 + ], + "type": "text", + "content": "[1] 1st workshop on urban scene modeling: Where vision meets photogrammetry and graphics. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 284, + 296, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 284, + 296, + 437 + ], + "spans": [ + { + "bbox": [ + 61, + 284, + 296, + 437 + ], + "type": "text", + "content": "[2] Pravesh Agrawal, Szymon Antoniak, Emma Bou Hanna, Baptiste Bout, Devendra Chaplot, Jessica Chudnovsky, Diogo Costa, Baudouin De Monicault, Saurabh Garg, Theophile Gervet, Soham Ghosh, Amelie Heliou, Paul Jacob, Albert Q. Jiang, Kartik Khandelwal, Timothee Lacroix, Guillaume Lample, Diego Las Casas, Thibaut Lavril, Teven Le Scao, Andy Lo, William Marshall, Louis Martin, Arthur Mensch, Pavankumar Muddireddy, Valera Nemychnikova, Marie Pellat, Patrick Von Platen, Nikhil Raghuraman, Baptiste Rozière, Alexandre Sablayrolles, Lucile Saulnier, Romain Sauvestre, Wendy Shang, Roman Soletskyi, Lawrence Stewart, Pierre Stock, Joachim Studnia, Sandeep Subramanian, Sagar Vaze, Thomas Wang, and Sophia Yang. Pixtral 12b, 2024. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 440, + 294, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 440, + 294, + 460 + ], + "spans": [ + { + "bbox": [ + 62, + 440, + 294, + 460 + ], + "type": "text", + "content": "[3] AI Anthropic. Claude 3.5 sonnet model card addendum. Claude-3.5 Model Card, 2024. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 62, + 464, + 294, + 518 + ], + "type": "text", + "content": "[4] Andre Araujo, Bingyi Cao, Cam Askew, Jack Sim, Maggie, Tobias Weyand, and Will Cukierski. Google landmark retrieval 2021. https://kaggle.com/competitions/landmark-retrieval-2021, 2021. Kaggle. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 520, + 294, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 520, + 294, + 597 + ], + "spans": [ + { + "bbox": [ + 62, + 520, + 294, + 597 + ], + "type": "text", + "content": "[5] Armen Avetisyan, Christopher Xie, Henry Howard-Jenkins, Tsun-Yi Yang, Samir Aroudj, Suvam Patra, Fuyang Zhang, Duncan Frost, Luke Holland, Campbell Orme, Jakob Engel, Edward Miller, Richard Newcombe, and Vasileios Balntas. Scenescript: Reconstructing scenes with an autoregressive structured language model. In European Conference on Computer Vision (ECCV), 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 599, + 294, + 653 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 599, + 294, + 653 + ], + "spans": [ + { + "bbox": [ + 62, + 599, + 294, + 653 + ], + "type": "text", + "content": "[6] Eric Brachmann, Jamie Wynn, Shuai Chen, Tommaso Cavallari, Aron Monszpart, Daniyar Turmukhambetov, and Victor Adrian Prisacariu. Scene coordinate reconstruction: Positioning of image collections via incremental learning of a relocalizer. In ECCV, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 656, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 656, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 656, + 294, + 689 + ], + "type": "text", + "content": "[7] Ralph A. Bradley and Milton E. Terry. Rank analysis of incomplete block designs: I. the method of paired comparisons. Biometrika, 39(3/4):324-345, 1952. 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 691, + 294, + 713 + ], + "type": "text", + "content": "[8] Li Cao, Yike Xu, Jianwei Guo, and Xiaoping Liu. Wireframenet: A novel method for wireframe generation from" + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "point cloud. Computers and Graphics, 115:226-235, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 320, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 320, + 97, + 553, + 140 + ], + "type": "text", + "content": "[9] Jiacheng Chen, Yiming Qian, and Yasutaka Furukawa. Heat: Holistic edge attention transformer for structured reconstruction. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 316, + 143, + 553, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 143, + 553, + 208 + ], + "spans": [ + { + "bbox": [ + 316, + 143, + 553, + 208 + ], + "type": "text", + "content": "[10] Kseniya Cherenkova, Elena Dupont, Anis Kacem, Ilya Arzhannikov, Gleb Gusev, and Djamila Aouada. Sepicnet: Sharp edges recovery by parametric inference of curves in 3d shapes. In 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 2727-2735, 2023. 1, 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 210, + 553, + 232 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 210, + 553, + 232 + ], + "spans": [ + { + "bbox": [ + 316, + 210, + 553, + 232 + ], + "type": "text", + "content": "[11] Arpad E. Elo. The Rating of Chessplayers, Past and Present. Arco Publishing, 1978. Accessed: 2025-03-06. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 234, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 234, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 316, + 234, + 553, + 255 + ], + "type": "text", + "content": "[12] Gemini Team et al. Gemini: A family of highly capable multimodal models, 2024. 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 258, + 503, + 269 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 258, + 503, + 269 + ], + "spans": [ + { + "bbox": [ + 316, + 258, + 503, + 269 + ], + "type": "text", + "content": "[13] OpenAI et al. Openai o1 system card, 2024. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 270, + 531, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 270, + 531, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 270, + 531, + 281 + ], + "type": "text", + "content": "[14] Mark E Glickman. Introductory note to 1928, 2013. 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 284, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 284, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 317, + 284, + 553, + 360 + ], + "type": "text", + "content": "[15] Tomás Hodań, Martin Sundermeyer, Yann Labbé, Van Nguyen Nguyen, Gu Wang, Eric Brachmann, Bertram Drost, Vincent Lepetit, Carsten Rother, and Jiri Matas. BOP challenge 2023 on detection, segmentation and pose estimation of seen and unseen rigid objects. Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2024. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 362, + 553, + 406 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 362, + 553, + 406 + ], + "spans": [ + { + "bbox": [ + 316, + 362, + 553, + 406 + ], + "type": "text", + "content": "[16] Perpetual Hope Akwensi, Akshay Bharadwaj, and Ruisheng Wang. Apc2mesh: Bridging the gap from occluded building facades to full 3d models. ISPRS Journal of Photogrammetry and Remote Sensing, 211:438-451, 2024. 1, 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 316, + 408, + 555, + 462 + ], + "type": "text", + "content": "[17] Shangfeng Huang, Ruisheng Wang, Bo Guo, and Hongxin Yang. Pbwr: Parametric-building-wireframe reconstruction from aerial lidar point clouds. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 27778-27787, 2024. 1, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "spans": [ + { + "bbox": [ + 316, + 464, + 553, + 507 + ], + "type": "text", + "content": "[18] Binyuan Hui, Jian Yang, Zeyu Cui, Jiaxi Yang, Dayiheng Liu, Lei Zhang, Tianyu Liu, Jiajun Zhang, Bowen Yu, Kai Dang, et al. Qwen2. 5-coder technical report. arXiv preprint arXiv:2409.12186, 2024. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "spans": [ + { + "bbox": [ + 316, + 510, + 553, + 543 + ], + "type": "text", + "content": "[19] Apple Inc. Roomplan: Create 3d floor plans with iphone and ipad. https://developer.apple.com/augmented-reality/roomplan/, 2022.1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 317, + 545, + 553, + 589 + ], + "type": "text", + "content": "[20] Yuhe Jin, Dmytro Mishkin, Anastasiia Mishchuk, Jiri Matas, Pascal Fua, Kwang Moo Yi, and Eduard Trulls. Image Matching across Wide Baselines: From Paper to Practice. International Journal of Computer Vision, 2020. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 591, + 553, + 655 + ], + "type": "text", + "content": "[21] Matej Kristan, Jiri Matas, Aleš Leonardis, Tomas Vojir, Roman Pflugfelder, Gustavo Fernandez, Georg Nebehay, Fatih Porikli, and Luka Čehovin. A novel performance evaluation methodology for single-target trackers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38(11):2137-2155, 2016. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[22] Jack Langerman, Caner Korkmaz, Hanzhi Chen, Daoyi Gao, Ilke Demir, Dmytro Mishkin, and Tolga Birdal. S23dr competition at 1st workshop on urban scene modeling @ cvpr 2024. https://huggingface.co/usm3d, 2024. 1, 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26952" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 699 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 115 + ], + "type": "text", + "content": "[23] L. Leal-Taixe, A. Milan, I. Reid, S. Roth, and K. Schindler. MOTChallenge 2015: Towards a benchmark for multi-target tracking. arXiv:1504.01942 [cs], 2015. arXiv: 1504.01942. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 161 + ], + "type": "text", + "content": "[24] Yujia Liu, Stefano D'Aronco, Konrad Schindler, and Jan Dirk Wegner. Pc2wf: 3d wireframe reconstruction from raw point clouds. In International Conference on Learning Representations, 2021. 1, 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[25] Yicheng Luo, Jing Ren, Xuefei Zhe, Di Kang, Yajing Xu, Peter Wonka, and Linchao Bao. Lc2wf: learning to construct 3d building wireframes from 3d line clouds. In Proceedings of the British Machine Vision Conference (BMVC), 2022. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 251 + ], + "type": "text", + "content": "[26] Wenchao Ma, Bin Tan, Nan Xue, Tianfu Wu, Xianwei Zheng, and Gui-Song Xia. How-3d: Holistic 3d wireframe perception from a single image. In International Conference on 3D Vision, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 285 + ], + "type": "text", + "content": "[27] Moritz Menze and Andreas Geiger. Object scene flow for autonomous vehicles. In Conference on Computer Vision and Pattern Recognition (CVPR), 2015. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 286, + 294, + 307 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 307 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 307 + ], + "type": "text", + "content": "[28] OpenRouter. Openrouter: Unified api for ai models, 2024. Accessed: 2025-03-07. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 309, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 407 + ], + "type": "text", + "content": "[29] Maxime Oquab, Timothée Darcet, Theo Moutakanni, Huy V. Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Russell Howes, Po-Yao Huang, Hu Xu, Vasu Sharma, Shang-Wen Li, Wojciech Galuba, Mike Rabbat, Mido Assran, Nicolas Ballas, Gabriel Synnaeve, Ishan Misra, Herve Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, and Piotr Bojanowski. Dinov2: Learning robust visual features without supervision, 2023. 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "spans": [ + { + "bbox": [ + 56, + 409, + 294, + 452 + ], + "type": "text", + "content": "[30] Filip Radenović, Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Revisiting oxford and paris: Large-scale image retrieval benchmarking. In CVPR, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 454, + 294, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 454, + 294, + 518 + ], + "spans": [ + { + "bbox": [ + 56, + 454, + 294, + 518 + ], + "type": "text", + "content": "[31] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet Large Scale Visual Recognition Challenge. International Journal of Computer Vision (IJCV), 115 (3):211-252, 2015. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "spans": [ + { + "bbox": [ + 56, + 520, + 294, + 563 + ], + "type": "text", + "content": "[32] Alberto Sanfeliu and King-Sun Fu. A distance measure between attributed relational graphs for pattern recognition. IEEE Transactions on Systems, Man, and Cybernetics, SMC-13(3):353-362, 1983. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 566, + 294, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 566, + 294, + 620 + ], + "spans": [ + { + "bbox": [ + 56, + 566, + 294, + 620 + ], + "type": "text", + "content": "[33] Ruisheng Wang, Shangfeng Huang, and Hongxin Yang. Building3D: An Urban-Scale Dataset and Benchmarks for Learning Roof Structures from Point Clouds. In ICCV, pages 20019-20029, Los Alamitos, CA, USA, 2023. IEEE Computer Society. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 621, + 294, + 642 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 621, + 294, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 621, + 294, + 642 + ], + "type": "text", + "content": "[34] xAI. Grok-2: Advancements in ai language modeling. 2024. Accessed: 2025-03-07. 4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 644, + 294, + 676 + ], + "type": "text", + "content": "[35] Ernst Zermelo. Die berechnung der turnier-ergebnisse als ein maximumproblem der wahrscheinlichkeitsrechnung. Mathematische Zeitschrift, 29:436-460, 1929. 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 677, + 294, + 699 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 677, + 294, + 699 + ], + "spans": [ + { + "bbox": [ + 56, + 677, + 294, + 699 + ], + "type": "text", + "content": "[36] Yichao Zhou, Haozhi Qi, and Yi Ma. End-to-end wireframe parsing. In ICCV 2019, 2019. 1" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 749, + 317, + 757 + ], + "type": "text", + "content": "26953" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_content_list.json b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..ef400f42b28962985e32458cfafb81c12a4d7ea7 --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_content_list.json @@ -0,0 +1,2060 @@ +[ + { + "type": "text", + "text": "Exploiting Diffusion Prior for Task-driven Image Restoration", + "text_level": 1, + "bbox": [ + 189, + 130, + 810, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jaeha Kim1 Junghun Oh1 Kyoung Mu Lee1,2", + "bbox": [ + 303, + 178, + 692, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Dept. of ECE&ASRI, $^{2}$ IPAI, Seoul National University, Korea", + "bbox": [ + 248, + 198, + 750, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jhkim97s2@gmail.com, {dh6dh, kyoungmu}@snu.ac.kr", + "bbox": [ + 282, + 218, + 709, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 268, + 326, + 284 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Task-driven image restoration (TDIR) has recently emerged to address performance drops in high-level vision tasks caused by low-quality (LQ) inputs. Previous TDIR methods struggle to handle practical scenarios in which images are degraded by multiple complex factors, leaving minimal clues for restoration. This motivates us to leverage the diffusion prior, one of the most powerful natural image priors. However, while the diffusion prior can help generate visually plausible results, using it to restore task-relevant details remains challenging, even when combined with recent TDIR methods. To address this, we propose EDTR, which effectively harnesses the power of diffusion prior to restore task-relevant details. Specifically, we propose directly leveraging useful clues from LQ images in the diffusion process by generating from pixel-error-based pre-restored LQ images with mild noise added. Moreover, we employ a small number of denoising steps to prevent the generation of redundant details that dilute crucial task-related information. We demonstrate that our method effectively utilizes diffusion prior for TDIR, significantly enhancing task performance and visual quality across diverse tasks with multiple complex degradations.", + "bbox": [ + 88, + 301, + 485, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 662, + 220, + 679 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image degradation often occurs in the real world due to various factors, such as transmission loss, limited camera performance, or poor shooting conditions. In such cases, key high-frequency details can be lost, resulting in significant performance drops in high-level vision tasks, such as image classification, semantic segmentation, and object detection.", + "bbox": [ + 89, + 686, + 482, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Image restoration (IR) can be a promising solution to improve the performance of high-level vision tasks. It involves recovering missing details in low-quality (LQ) images using a learned natural image prior. However, recent studies [8, 20, 39, 57] show that the straightforward application of IR methods as a pre-processing step is not very effective in restoring essential information for high-level vision tasks. This has led to the emergence of Task-driven Image", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2ae9dcb11dad0ad03823d3fb5961357209071030c7809e3614d7cfedd296956d.jpg", + "image_caption": [ + "(a) Low-quality" + ], + "image_footnote": [], + "bbox": [ + 514, + 334, + 616, + 449 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/affdd27bc038ead7af827914a4992b257db76fb56c349a2fa73c2d1ec99aa948.jpg", + "image_caption": [ + "(b) DiffBIR [30]" + ], + "image_footnote": [], + "bbox": [ + 625, + 267, + 759, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5dd5d79915e0cdfb5e37028eca6b4ea3092a227cb30376a693a009f95991ac46.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 767, + 267, + 900, + 383 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ee7b2292e5a0eb27f08c8320dd77d9c34413216ce584f630516d50d59a57f497.jpg", + "image_caption": [ + "(d) EDTR (Ours)" + ], + "image_footnote": [], + "bbox": [ + 625, + 398, + 759, + 515 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/74da3437fe14e26767f8bd707d20663872a57513534260d2dd4204e304ac946d.jpg", + "image_caption": [ + "(c) SR4IR [20] $+\\mathrm{SD}$ [41]", + "(e) High-quality", + "Figure 1. Effect of diffusion prior on image classification performance. While incorporating the StableDiffusion [41] produces visually plausible results as (b), it lacks task-relevant details (e.g., eye shape), even when combined with the previous TDIR method as (c), leading to incorrect classification. In contrast, EDTR successfully restores details as (d), resulting in correct classification." + ], + "image_footnote": [], + "bbox": [ + 767, + 398, + 900, + 515 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Restoration (TDIR), which aims to improve visual quality in ways that benefit high-level vision tasks.", + "bbox": [ + 511, + 641, + 903, + 671 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Previous TDIR methods address each type of degradation separately, e.g., super-resolution [20], denoising [31]. This restricts their practical applicability, as real-world images often exhibit multiple degradations simultaneously. To address this, we tackle a more realistic scenario by jointly handling multiple complex degradations. However, we observe that previous TDIR methods achieve limited performance gains in such scenarios, as they struggle to utilize the few available clues in degraded images to restore task-relevant details. This highlights the need for a strong natural image prior for the TDIR problem.", + "bbox": [ + 509, + 672, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we leverage the prior knowledge from the pre-trained StableDiffusion (SD) [41], which provides one of the most powerful natural image priors. However, we observe that while the conventional SD-based IR method [30]", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10151", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "can generate visually pleasing results, its straightforward adoption fails to restore task-relevant details, leading to incorrect predictions, as shown in Figure 1b. More importantly, we observe that SD still struggles to restore task-relevant details even when combined with a recent state-of-the-art TDIR method [20], as shown in Figure 1c. In other words, while diffusion prior is powerful, it must be properly adapted, as its standard form fails to improve TDIR. This issue arises because restoring task-relevant details is inherently challenging in the standard diffusion process used in conventional SD-based IR methods [30, 52, 59-61]. We identify two key reasons: (1) Although the LQ image is used as a condition for SD, generating from pure noise hinders the direct utilization of useful clues in LQ images, making it difficult to restore task-relevant details. (2) Long-step denoising, around 50, introduces redundant details that dilute crucial task-related information for TDIR.", + "bbox": [ + 89, + 90, + 480, + 345 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, we propose EDTR, Exploiting Diffusion prior for Task-driven Image Restoration, the TDIR method adopting the diffusion prior in ways that directly leverage its power to restore task-relevant details in complex degradation scenarios. The primary component of EDTR is threefold: (1) We directly incorporate LQ images into the diffusion process. Since SD was not originally trained to handle complex degradations, directly feeding LQ images into the diffusion process is ineffective. Therefore, we first prerestore the LQ images using a pixel-error-based IR method, then generate from these images with mild noise added, rather than generating from pure noise, a process referred to as pre-restoration and partial diffusion. (2) We apply a short-step denoising (e.g., 1, 4) during inference. We claim that increasing the number of steps (e.g., 50) reduces task performance by introducing redundant details that dilute crucial details for TDIR. (3) We introduce tailored training loss terms to guide diffusion prior to restore task-relevant details. These loss terms are essential for achieving visually appealing results, even in a one-step denoising.", + "bbox": [ + 91, + 348, + 482, + 651 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We validate our method in three high-level vision tasks: image classification, semantic segmentation, and object detection. These tasks involve complex degradation scenarios, including a mix of downsampling, blur, noise, and JPEG artifacts. Experimental results demonstrate that EDTR effectively harnesses the power of diffusion prior, significantly improving task performance (e.g., about $+50\\%$ improvement in detection) while producing visually pleasing results, as shown in Figure 1d, and generalizes well to real-world images. We summarize our contributions as:", + "bbox": [ + 89, + 651, + 480, + 801 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose EDTR, the TDIR method that directly leverages a powerful pre-trained diffusion prior.", + "- We propose tailored designs for exploiting the diffusion prior to effectively restore task-relevant details.", + "- We significantly improve task performance while producing visually pleasing results under complex degradation." + ], + "bbox": [ + 89, + 804, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 89, + 653, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "General image restoration. Since the pioneering SRCNN [10], numerous deep neural network-based methods [5, 9, 13, 19, 24, 28, 38, 53, 65, 66] have emerged in the field of image restoration (IR). Vision transformers [11]-based IR methods [3, 7, 55, 62], such as SwinIR [27], have also demonstrated strong performance. Recently, IR studies [30, 34, 52, 60, 61] have focused on leveraging the large pre-trained diffusion models, primarily StableDiffusion (SD) [41], which provides a powerful natural image prior essential for addressing the ill-posedness of the IR problem. StableSR [52] and DiffBIR [30] introduce trainable modules alongside a frozen SD backbone, similar to ControlNet [63]. More recent works [35, 49, 59, 61] have focused on fully harnessing SD's generative capabilities by designing carefully crafted input conditions.", + "bbox": [ + 511, + 109, + 903, + 335 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Task-driven image restoration. Taking a different perspective, recent studies [1, 2, 33, 44, 50, 57] have focused on the emerging field of Task-driven Image Restoration (TDIR), which aims to improve high-level vision task performance when handling LQ images. AOD-Net [25] suggests that using dehazing as a preprocessing step and applying object detection can enhance detection performance in foggy conditions. However, Pei et al. [39] show that applying IR methods as a simple preprocessing step does not significantly improve detection performance, with similar findings reported in the super-resolution field [8].", + "bbox": [ + 511, + 340, + 903, + 506 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address this, recent TDIR methods [14, 18, 26, 31, 32, 45, 67] have started training the IR network in combination with the task networks. TDSR [14] trains the IR network using both pixel-error loss and detection loss from a pre-trained object detection network, with a dedicated training schedule. RSRSSN [67] proposes an end-to-end trainable IR and detection networks with Feature Maps Multi-box loss. SR4IR [20] introduces a joint training framework for TDIR with a task-driven perceptual loss to guide the IR network in learning task-relevant features. More recently, UniRestore [4] introduces the diffusion prior into TDIR; however, its diffusion-based training is independent of taskdriven training, performed in two stages, limiting the ability to fully leverage the diffusion prior for TDIR. We instead propose integrating the diffusion prior directly within taskdriven training to better restore task-relevant details.", + "bbox": [ + 511, + 508, + 903, + 750 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Proposed Method", + "text_level": 1, + "bbox": [ + 511, + 760, + 686, + 776 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Problem definition. Let $I_{\\mathrm{HQ}}$ and $I_{\\mathrm{LQ}}$ denote the high-quality (HQ) and low-quality (LQ) images, respectively, with paired labels $y$ from corresponding high-level vision task, e.g., class label in the image classification. The goal of Task-driven Image Restoration (TDIR) is to restore LQ images in a way that enhances the subsequent task performance while also generating visually appealing results. Formally, given a trainable image restoration (IR) model $\\mathcal{R}$", + "bbox": [ + 511, + 779, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10152", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/f57cde187a8d7762035bb6f76d259f66e627592b7f8110e6e12380623877883b.jpg", + "image_caption": [ + "(a) The architecture of the proposed EDTR with high-level vision task network." + ], + "image_footnote": [], + "bbox": [ + 101, + 88, + 897, + 220 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/64199ac137ae30ea48a96aea7337f52d532af6260f13845ebed59b76b071ab8d.jpg", + "image_caption": [ + "(b) The joint training procedure of EDTR and high-level vision task network.", + "Figure 2. The overall framework of the proposed method. $I_{\\mathrm{EDTR,train}}$ is the EDTR restored image when EDTR is in training mode." + ], + "image_footnote": [], + "bbox": [ + 125, + 239, + 869, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "and a trainable task network $\\mathcal{H}$ , our aim is to maximize $\\mathbf{M}_{\\mathrm{H}}(\\mathcal{H}(\\mathcal{R}(I_{\\mathrm{LQ}})),y)$ , where $\\mathbf{M}_{\\mathrm{H}}$ is the evaluation metric for the task, e.g., the accuracy for classification, while also achieving visually pleasing restored results $\\mathcal{R}(I_{\\mathrm{LQ}})$ .", + "bbox": [ + 89, + 434, + 483, + 496 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary: Conventional SD-based IR", + "text_level": 1, + "bbox": [ + 89, + 503, + 434, + 520 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Unlike pure generation, IR requires generating images that are faithful to the given LQ images. To achieve this, previous StableDiffusion (SD)-based IR methods [30, 49, 52, 60, 61] commonly use LQ images as a condition for SD through ControlNet [63]. Note that, for better efficiency and more stable training, SD applies the diffusion process in the latent space rather than the image space, following the encoder and decoder structure of the VAE [21].", + "bbox": [ + 89, + 526, + 483, + 648 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During training, given an image pair $(I_{\\mathrm{HQ}}, I_{\\mathrm{LQ}})$ , a VAE encoder $\\mathcal{E}$ , and Gaussian noise with variance $\\beta_t \\in (0,1)$ , $I_{\\mathrm{HQ}}$ is first diffused according to DDPM [17] as follows:", + "bbox": [ + 89, + 648, + 483, + 695 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t} = q \\left(z _ {t} \\mid z _ {0} = \\mathcal {E} \\left(I _ {\\mathrm {H Q}}\\right)\\right) \\tag {1} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {E} (I _ {\\mathrm {H Q}}) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 181, + 700, + 480, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $q(z_{t}|z_{t - 1})\\coloneqq \\mathcal{N}(\\sqrt{1 - \\beta_{t}} z_{t - 1},\\beta_{t}\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t},$ $\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}$ , and $\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})$ . The timestep $t$ is randomly selected from $[1,T]$ , where $T$ is the total number of timesteps. Then the trainable module, usually ControlNet, is trained to predict the injected noise as follows:", + "bbox": [ + 89, + 752, + 483, + 829 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(I _ {\\mathrm {L Q}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 184, + 837, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\epsilon_{\\theta}$ is a denoising network combined with ControlNet. Note that $I_{\\mathrm{LQ}}$ is used only as condition input for ControlNet.", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "During inference, denoising starts from a pure noise. It gradually moves toward less noisy images by iteratively alternating between denoising and adding noise, using sampling methods [17, 37, 46]. After approximately 50 steps, the process obtains the predicted latent $\\hat{z}_0$ , which is then mapped back to the image space using the VAE decoder.", + "bbox": [ + 511, + 434, + 905, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2.EDTR Architecture", + "text_level": 1, + "bbox": [ + 511, + 531, + 702, + 545 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pre-restoration and partial diffusion. As noted earlier, conventional SD-based IR methods begin the denoising process from pure noise during the inference. Although $I_{\\mathrm{LQ}}$ are used as a condition through ControlNet, its behavior heavily depends on SD's generative capacity. This poses a challenge for TDIR, as it hinders the use of valuable clues in LQ images, making it difficult to restore task-relevant details.", + "bbox": [ + 511, + 550, + 905, + 655 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To address this, we directly incorporate $I_{\\mathrm{LQ}}$ into the diffusion process, i.e., Equation (1). However, this approach poses two key challenges: (1) Since SD was not originally trained on $I_{\\mathrm{LQ}}$ , which include complex degradations such as JPEG artifacts, directly incorporating $I_{\\mathrm{LQ}}$ into the diffusion process is ineffective at leveraging the power of SD. (2) When $I_{\\mathrm{LQ}}$ is fed directly into SD, degradation artifacts can be interpreted as generation cues, producing tasks-irrelevant details. Instead, we first pre-restore $I_{\\mathrm{LQ}}$ using a pixel-error-based IR method $\\mathcal{R}_{\\mathrm{pix}}$ , such as SwinIR [27]:", + "bbox": [ + 511, + 655, + 906, + 806 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {p r e - r e s}} = \\mathcal {E} \\left(\\mathcal {R} _ {\\text {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right). \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 814, + 903, + 832 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Although introducing strong noise in the diffusion process enables the generation of additional information by heavily relying on the diffusion prior, preserving valuable clues in the LQ image is more crucial for the TDIR prob", + "bbox": [ + 511, + 840, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10153", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/f7b05482ebbed1a55492bb1aad1097269bbc8e765741bb2638b1c064a27b46be.jpg", + "image_caption": [ + "Figure 3. Illustration of pre-restoration and partial diffusion." + ], + "image_footnote": [], + "bbox": [ + 91, + 87, + 483, + 169 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "lem. Thus, we propose adding mild noise to the pre-restored images and starting generation from it. We denote this technique as partial diffusion, which can be formally defined as:", + "bbox": [ + 89, + 191, + 483, + 239 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} z _ {t, \\text {p a r t i a l}} = q \\left(z _ {t \\leq t _ {p}} \\mid z _ {0} = z _ {\\text {p r e - r e s}}\\right) \\tag {4} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} z _ {\\text {p r e - r e s}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 243, + 482, + 281 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $z_{t,\\text{partial}}$ is the partial diffusion result and $t_p$ is a fixed hyper-parameter with a relatively smaller value than $T$ . We empirically set $(t_p, T) = (200, 1000)$ . Figure 3 illustrates the proposed diffusion process.", + "bbox": [ + 89, + 291, + 483, + 353 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Short-step denoising. As noted in the preliminary section, conventional SD-based IR methods employ long-step denoising (e.g., 50). However, in the context of the TDIR problem, we find that a large number of denoising steps negatively impact on task performance. Although each step introduces details through the diffusion prior, repeatedly adding such details can generate redundant content, thereby diluting crucial task-relevant information. Hence, we adopt a short-step denoising (e.g., 1, 4) during inference. Specifically, for a given number of denoising steps $n$ , we denoise $z_{t_p,\\text{partial}}$ by passing it through $n$ timesteps: $\\mathcal{T} = [t_p, \\lfloor t_p / n \\cdot (n - 1) \\rfloor, \\dots, \\lfloor t_p / n \\rfloor]$ , following [37], and obtain the final denoising result $z_{\\text{diff-res}}$ . In the one-step denosing case (i.e., $n = 1$ ), this process simplifies to:", + "bbox": [ + 89, + 356, + 483, + 568 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nz _ {\\text {d i f f - r e s}} = \\frac {z _ {t _ {p} , \\text {p a r t i a l}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {p}}} \\epsilon_ {\\theta} \\left(z _ {t _ {p} , \\text {p a r t i a l}}, t _ {p} , z _ {\\text {p r e - r e s}}\\right)}{\\sqrt {\\bar {\\alpha} _ {t _ {p}}}}. \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 102, + 575, + 482, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Note that conventional SD-based IR methods [30, 52] struggle to generate visually pleasing results in short-step denoising scenarios [54, 58]. However, we find that optimizing loss functions, specifically designed to restore task-relevant details as discussed in Section 3.3, enables the achievement of visually pleasing results, even in one-step denoising.", + "bbox": [ + 89, + 628, + 483, + 719 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Figure 4 compares the overall IR procedure. Unlike conventional methods, which start from pure noise and require around 50 denoising steps, we use the input LQ image with pre-restoration and partial diffusion, followed by short-step denoising. This approach directly leverages useful clues in LQ images to restore task-relevant details.", + "bbox": [ + 89, + 719, + 483, + 810 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Decoding into RGB space. After the denoising process, we transform the $z_{\\mathrm{diff-res}}$ back into the image space using the VAE decoder $\\mathcal{D}$ . We further apply Wavelet color correction [52] to achieve a faithful reconstruction as follows:", + "bbox": [ + 89, + 815, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathrm {E D T R}} = \\mathbf {H} \\left(\\mathcal {D} \\left(z _ {\\text {d i f f - r e s}}\\right)\\right) + \\mathbf {L} \\left(\\mathcal {R} _ {\\operatorname {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 882, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fe2f5071225bb7a808144a9c4292d608701251a497b41de27000a09f2fe255a2.jpg", + "image_caption": [ + "(b) Pre-restoration, partial diffusion, and short-step denosing.", + "Figure 4. The comparison of IR procedure. We omit the encoding, decoding, and conditioning for simplicity in visualization." + ], + "image_footnote": [], + "bbox": [ + 535, + 88, + 885, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I_{\\mathrm{EDTR}}$ is the final restored result from our EDTR. The $\\mathbf{H}(\\cdot)$ and $\\mathbf{L}(\\cdot)$ correspond to the high-frequency and low-frequency components extracted by Wavelet transform. Figure 2a illustrates the overall architecture of our EDTR.", + "bbox": [ + 511, + 294, + 905, + 356 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Training Framework", + "text_level": 1, + "bbox": [ + 511, + 364, + 712, + 380 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "High-level feature loss. The goal of TDIR is to restore task-relevant details to enhance task performance. However, the noise prediction loss in Equation (2), which is conventionally used for training diffusion models, is not designed for this purpose. Hence, we introduce high-level feature (HLF) loss, which can guide the model to harness the diffusion prior for restoring task-relevant details.", + "bbox": [ + 511, + 382, + 905, + 487 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "HLF loss measures the feature distance between EDTR restored results and HQ images, following the previous TDIR method [20]. Specifically, we compute the HLF loss using the feature space of two task networks: the task network $\\mathcal{H}$ , which is optimized during training, and another task network $\\mathcal{H}_{\\mathrm{HQ}}$ , pre-trained on HQ images, to capture complementary task-relevant information (see Supplementary Section S5). Formally, the HLF loss is defined as:", + "bbox": [ + 511, + 488, + 905, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {H L F}} = \\frac {1}{2} \\left(\\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R}, \\text {t r a i n}}\\right) - \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1} \\right. \\tag {7} \\\\ + \\left\\| \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {E D T R , t r a i n}}) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {H Q}}) \\right\\| _ {1}), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 611, + 903, + 669 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{H}^f$ and $\\mathcal{H}_{\\mathrm{HQ}}^f$ are intermediate feature spaces of $\\mathcal{H}$ and $\\mathcal{H}_{\\mathrm{HQ}}$ , respectively. Note that $I_{\\mathrm{EDTR,train}}$ in Equation (7) represents the restored image from the one-step denoising result of $z_{t,\\mathrm{partial}}$ , where $t$ is randomly sampled from $\\mathcal{T}$ during training, following the conventional diffusion loss in Equation (2), regardless of denoising steps $n$ during inference (see Supplementary Algorithm S1 for details). The HLF loss updates only EDTR and does not affect $\\mathcal{H}$ .", + "bbox": [ + 511, + 671, + 905, + 794 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Task loss. While training the proposed EDTR to restore task-relevant details, we also train the task networks using the restored outputs. We utilize both $I_{\\mathrm{EDTR}}$ and HQ images to stabilize training and introduce additional regularization. Formally, task loss $\\mathcal{L}_{\\mathrm{task}}$ for the task network is defined as:", + "bbox": [ + 511, + 797, + 905, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t a s k}} = f _ {\\text {t a s k}} \\left(\\mathcal {H} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right), y\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 599, + 883, + 903, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10154", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $f_{\\mathrm{task}}$ is the standard loss function for each task, e.g., cross-entropy loss for classification, $I_{\\mathrm{EDTR}} \\oplus_{\\mathrm{HQ}}$ represents batch-wise concatenated images, consisting of a half-batch from $I_{\\mathrm{EDTR}}$ and a half-batch from $I_{\\mathrm{HQ}}$ , and $y$ are corresponding task labels. Note that $\\mathcal{L}_{\\mathrm{task}}$ updates only the task network $\\mathcal{H}$ and does not update EDTR.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Feature matching loss. To align the feature space of task network $\\mathcal{H}$ with $\\mathcal{H}_{\\mathrm{HQ}}$ , which is well-trained to utilize high-frequency details present in HQ images, we further introduce feature matching (FM) loss, defined as:", + "bbox": [ + 89, + 184, + 482, + 246 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {F M}} = \\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1}. \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 143, + 250, + 482, + 275 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For clarity, note that $I_{\\mathrm{EDTR}}$ in Equations (8) and (9) is the $n$ step denoising result, different from the one-step result $I_{\\mathrm{EDTR,train}}$ in Equation (7). The $\\mathcal{L}_{\\mathrm{FM}}$ can be interpreted as a form of cross-quality knowledge distillation [16, 48] at the feature level. Similar to the task loss, $\\mathcal{L}_{\\mathrm{FM}}$ only updates $\\mathcal{H}$ .", + "bbox": [ + 89, + 280, + 482, + 356 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Joint training in an alternating manner. Since SD was originally trained to operate in the image space, backpropagating the task loss, such as cross-entropy loss that is not image-based, causes instability to the EDTR. To avoid this, we train both the EDTR and the task network $\\mathcal{H}$ using each corresponding training loss in an alternating manner [20], as illustrated in Figure 2b. Specifically, in each training iteration, we first update the EDTR using the HLF loss, then update $\\mathcal{H}$ using a combination of the task loss and the FM loss, which can be formally written as follows:", + "bbox": [ + 89, + 359, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {E D T R}}} \\mathcal {L} _ {\\mathrm {H L F}}, \\text {t h e n} \\min _ {\\boldsymbol {\\theta} _ {\\mathcal {H}}} \\mathcal {L} _ {\\text {t a s k}} + \\alpha \\mathcal {L} _ {\\mathrm {F M}}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 513, + 480, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\theta_{\\mathrm{EDTR}}$ and $\\theta_{\\mathcal{H}}$ are the trainable parameters of the EDTR and task network $\\mathcal{H}$ , and $\\alpha$ is the hyper-parameter.", + "bbox": [ + 89, + 542, + 482, + 574 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 585, + 223, + 602 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We validate our EDTR on three representative high-level vision tasks: image classification, semantic segmentation, and object detection, under scenarios involving image degradation. Note that the final prediction $\\hat{y}$ for each task is obtained as: $\\hat{y} = \\mathcal{H}(\\mathrm{EDTR}(I_{\\mathrm{LQ}}))$ , where $\\mathcal{H}$ is the trained task network and $I_{\\mathrm{LQ}}$ is the input low-quality (LQ) images.", + "bbox": [ + 89, + 606, + 482, + 698 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We use three high-level vision tasks with representative datasets: CUB200 [51] for classification, and PASCAL VOC2012 [12] for segmentation and detection. We construct two degraded LQ image sets: 1) Mixture-A: downsampling combined with JPEG compression, and 2) Mixture-B: a combination of downsampling, blur, noise, and compression, following the CodeFormer [68] degradation. Details on degradation types are in the second column of Table 1 and Section S1 in the supplementary materials.", + "bbox": [ + 89, + 700, + 482, + 837 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation metrics. We use representative evaluation metrics for each task: top-1 accuracy (Acc) for classification, mean Intersection over Union (mIoU) for segmentation, and mean Average Precision (mAP) for detection. The", + "bbox": [ + 89, + 839, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "mAP score is measured following the standard COCO [29] evaluation setting. We use two no-reference metrics to assess visual appeal: NIQE [36] and Q-Align [56], where the latter leveraging large vision-language models. We also report the standard PSNR for comparison.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network architecture. We use StableDiffusion-2.1 [41] (SD) and ControlNet [63] for EDTR. ControlNet and the VAE decoder of SD are set as trainable modules. We also incorporate the pixel-error-based SwinIR [27] model as the pre-restoration network. For the task network, we employ representative architecture for each task: ResNet [15] for classification, DeepLabV3 [6] for segmentation, and FasterRCNN [40] for detection. The task network $\\mathcal{H}$ shares the same architecture as $\\mathcal{H}_{\\mathrm{HQ}}$ (the network pre-trained at HQ), and the weights of $\\mathcal{H}$ are initialized from $\\mathcal{H}_{\\mathrm{HQ}}$ .", + "bbox": [ + 511, + 172, + 906, + 325 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental Results", + "text_level": 1, + "bbox": [ + 511, + 335, + 714, + 351 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Table 1 and Figure 5 present quantitative and qualitative results, respectively, across various types of degradation in image classification, semantic segmentation, and object detection. Note that in Table 1, 'Oracle', 'No restoration', and 'SwinIR' refer to the performance when the task network is trained and tested directly on HQ, LQ, and pixel-error-based SwinIR restored images, respectively. EDTR- $n$ step refers to the EDTR model with $n$ denoising steps.", + "bbox": [ + 511, + 358, + 905, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Image classification. Table 1 shows that the proposed EDTR significantly outperforms previous methods under both types of degradation, achieving classification accuracy improvements of $+2.9\\%$ and $+5.4\\%$ over the previous state-of-the-art task-driven image restoration (TDIR) method, SR4IR [20]. Performance gains increase as the degradation becomes more severe, highlighting our method's efficacy in practical scenarios. Furthermore, EDTR achieves the highest scores in image quality assessments such as NIQE and Q-Align, indicating its effectiveness in producing visually pleasing results. However, despite the superior task performance, EDTR does not achieve the highest PSNR score, suggesting that PSNR is a less relevant metric in TDIR. As presented in the first row of Figure 5, our EDTR, leveraging a strong natural image prior, successfully restores detailed wing patterns—an important area for classification as shown by the Grad-CAM [43] in the upper-left—and correctly predicts the label, whereas the previous method fails to restore these details and misclassifies the image.", + "bbox": [ + 511, + 484, + 906, + 772 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Semantic segmentation. The performance improvement in segmentation is more notable than in classification, with mIoU gains of $+9.1\\%$ and $+11.9\\%$ over the previous state-of-the-art. Although our EDTR achieves the second-best score in NIQE, it consistently achieves the highest Q-Align score, demonstrating its capability to generate perceptually realistic images. In addition, unlike classification, increasing the number of denoising steps from 1 to 4 leads to a task", + "bbox": [ + 511, + 780, + 908, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10155", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/b585e2f5222664d1b7c19c12a8682bcdbcde556a1c409865b46f6ab9f9973955.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
IR MethodsDegradation typeImage classificationSemantic segmentationObject detection
Task Acc↑ (%)Visual qualityTask mIoU↑ (%)Visual qualityTask mAP↑ (%)Visual quality
NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑
OracleNone82.55.373.69+inf67.04.714.02+inf36.93.703.95+inf
No restoration60.814.081.0323.7247.413.671.1622.4314.514.541.0421.08
SwinIR [27]70.09.972.5525.2156.28.662.5423.4721.08.962.2221.70
TDSR [14]Mixture-A:66.17.532.5124.9951.37.672.3723.4116.26.432.1321.62
RSRSSN [67]SR(s=8)+69.27.322.7524.0453.06.972.5823.0622.36.592.1021.40
SR4IR [20]JPEG(q=75)71.55.413.3724.0756.24.803.1622.8522.35.383.0020.95
EDTR-1 step (Ours)74.44.683.5323.6364.15.303.5922.3331.24.623.6820.60
EDTR-4 step (Ours)74.14.263.8122.4665.35.173.9121.4833.44.194.0119.89
No restoration47.615.111.0823.0540.214.841.1922.100.015.241.0720.86
SwinIR [27]Mixture-B:60.710.262.0725.0450.49.472.3023.6722.19.601.9222.20
TDSR [14]SR(s[1,16]) +56.47.292.2024.6846.08.072.1323.5514.66.361.9421.97
RSRSSN [67]Blur(σ[0,8]) +61.47.352.3523.3548.27.722.3623.0123.16.811.8621.68
SR4IR [20]Noise(σ[0,10]) +63.46.083.1123.6251.04.723.1522.7122.64.523.1420.97
EDTR-1 step (Ours)JPEG(q[50,100])68.84.753.4823.0360.46.363.5522.0430.64.483.6420.70
EDTR-4 step (Ours)68.44.463.4923.2262.95.803.7422.2431.94.124.0220.29
", + "bbox": [ + 93, + 89, + 903, + 303 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 1. Performance of high-level vision tasks and quality of restored image results under various degradation types. We indicate the best and second-best results with bold and underlined formatting, respectively.", + "bbox": [ + 89, + 303, + 906, + 333 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4b7f455994bc8490a864b5bc4db556f84a078427738621a178ea936056825340.jpg", + "image_caption": [ + "pred: [Blue Grosbeak]" + ], + "image_footnote": [], + "bbox": [ + 93, + 345, + 294, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/49ffc14a44bb4a43fc9c07016fd3798afb5ad78362f3b6ea791635c0a69b0931.jpg", + "image_caption": [ + "pred: [Blue Grosbeak] X" + ], + "image_footnote": [], + "bbox": [ + 295, + 345, + 496, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/28d8ad8fe15c2a50b2aae041616803a6a82812eb31db92e5818ae24c682928d5.jpg", + "image_caption": [ + "pred: [Indigo Bunting]" + ], + "image_footnote": [], + "bbox": [ + 500, + 345, + 700, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/25a63364286dd88b5fc466012482d3a0b9045e78d6b0dd431ce4ea7107460f5f.jpg", + "image_caption": [ + "pred: [Indigo Bunting]" + ], + "image_footnote": [], + "bbox": [ + 702, + 345, + 903, + 476 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8cded3d279935b22e7a947771095d5638b0efb764ffde57ad242c3c669f0f830.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 502, + 194, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/76ea47ecb4662490a7e3acef484ab8410f486d67af0b5c36c6e579f50b4863c5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 571, + 194, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c70905a90aa2b735178675c5e59f88055f54b1594d763e78cc0cab5869950d19.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 503, + 295, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d2b34b8fae38ef6ad016c45be54f5d9d6c2bc3e6686bb0e9092492a43d11cab9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 196, + 570, + 295, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ad9405d657b8fc6f5d3d89cc2bd714aa89e5813c8d0908bd085919a02adbe906.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 503, + 398, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3d9be70e7ecdf95a7abee7dc90eb266fa778051cede40d0b743e1f4bc8ee5814.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 297, + 570, + 397, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/4ceae1e77ea961a6651d17ef64d4dc1ff1ee07741ac447369f0d3df56c5305c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 503, + 498, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cb2c7f41b010ebaea61063f7c7ec2a27ad457cd7adb63a00585483fc68857a2b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 398, + 570, + 496, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3284048523fe6c6f2f53c7a0f036623de4fa4d6ac00955a255ab3e8de5018cc7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 503, + 601, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1e12c188fec29898a01cb6d29a17a6b9813ed545ceb16e2cd20edd3a1ae5a3bb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 570, + 599, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/02e2f232754a517f8a8424d33ba2499eefe8d4a0a74df7deb52957b1a2415e36.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 503, + 700, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8a65090f74b838e951aed696aa56b10e945c23b3bd5a31fb6434f972e5789513.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 601, + 570, + 700, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e0ff27934d4c0e00b7c18ea3364c9e5a21a540423adcdaf22876cc4cec9cef91.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 503, + 803, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/bce4efbb93b19de2a1a1946bb2911920290d7755bf9b447532028dd85f3c353b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 570, + 802, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/74c0a4ee6dfdcc8a12d0552a6972ac3883328170843aa818ea993d36b6d65727.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 503, + 903, + 570 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/cd397e4b0dee1dd57047c52f8f5b73c5d30810a0580c02184b2eeb438c96068a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 570, + 903, + 647 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/6e123f70a3cf4531b42709a319ba1bfe17a9437d7e1616ccd4a82557e9a74413.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 657, + 294, + 734 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/1aba62c84f3fd0fb7c6f9c795fe7cf3a790b50ac92b0ddd29320f69c03bdc567.jpg", + "image_caption": [ + "(a) LQ (No restoration)" + ], + "image_footnote": [], + "bbox": [ + 94, + 734, + 294, + 799 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9207c3fde75b4530a3a66da74ef0c53b27476652da8a1193295b226b25e27220.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 295, + 657, + 496, + 734 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ee2fea1af0b2212d5cc7fd9a59ad1bcc3b36b64a716ea7d8a8aeb713e5c495f8.jpg", + "image_caption": [ + "(b) SR4IR [20]" + ], + "image_footnote": [], + "bbox": [ + 295, + 734, + 496, + 799 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ceca33532dc97d8f5641d07ce682334443a966fd43e50f9af7f805845a8ef7ca.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 657, + 700, + 734 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/42e5ee08054fb7a6d4f4318949c69523152dec3e2a54b9a78267e98e0acfbc65.jpg", + "image_caption": [ + "(c)EDTR(Ours)", + "Figure 5. Visualization of images and diverse high-level vision task results on degraded LQ (Mixture-B) images. The first, second, and third rows represent the results of image classification, semantic segmentation, and object detection, respectively. We show the restored images and the corresponding predicted labels. For visualization, we use the EDTR-1 step model for classification and the EDTR-4 step model for segmentation and detection. For more visualization results, please refer to Section S12 of our supplementary materials." + ], + "image_footnote": [], + "bbox": [ + 500, + 734, + 700, + 799 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3357a641f72a1e5da0ab1a2095bd7cc520bf32df7247eba302f5cb84c5f0b978.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 702, + 657, + 903, + 734 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/d6e7c306f0c47135aa46e503d73079b89693563996c40c27a1958624a5a7f6f1.jpg", + "image_caption": [ + "(d) HQ (Oracle)" + ], + "image_footnote": [], + "bbox": [ + 702, + 734, + 903, + 799 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10156", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5548bb3deab6c7bf8e1a41cf3327be2d1f66dc522270d2501d8740bcf1659a4f.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ExpSDTraining lossDiffusion processAcc↑ (%)fd↓Q-Align↑
HLFFMPre-restorationPartial diffusion# of denoising steps
(1)✓X✓XX5064.30.6413.87
(2)✓✓✓XX5062.90.7542.37
(3)✓✓✓XX163.50.5793.18
(4)✓✓✓X✓166.60.5303.38
(5)✓X✓✓✓164.50.7192.18
(6)✓✓X✓✓167.60.5073.48
(7)X✓✓✓✓164.20.5413.20
Ours✓✓✓✓✓168.80.5073.48
", + "bbox": [ + 94, + 89, + 903, + 238 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Ablation study on the impact of each component on EDTR classification performance.", + "bbox": [ + 200, + 239, + 795, + 253 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance improvement. In the second row of Figure 5, the LQ image is severely degraded, causing the details in the upper-left part of the bike wheel to be almost lost. While the previous TDIR method fails to restore the missing details, our method effectively restores them by leveraging a strong natural image prior, correctly recognizing that they belong to the wheel. As a result, EDTR produces a segmentation prediction that correctly maps the upper-left part of the bike wheel, demonstrating its effectiveness.", + "bbox": [ + 88, + 266, + 482, + 402 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Object detection. Our EDTR demonstrates a surprisingly significant performance gain over the previous state-of-the-art in the Mixture-A setting, increasing mAP from $22.3\\%$ to $33.4\\%$ —an improvement of approximately $50\\%$ . As shown in the third row of Figure 5, EDTR correctly restores the shape of the sofa by leveraging the powerful diffusion prior, successfully detecting the sofa in the room, whereas the previous method fails. These results demonstrate that our method consistently achieves superior task performance and visual quality under severe degradation.", + "bbox": [ + 89, + 409, + 483, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Ablation studies", + "text_level": 1, + "bbox": [ + 89, + 571, + 251, + 585 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "All experiments use the Mixture- $B$ degradation setting.", + "bbox": [ + 89, + 593, + 452, + 609 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2.1. Effect of proposed components.", + "text_level": 1, + "bbox": [ + 89, + 613, + 356, + 628 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2 shows the classification performance of our EDTR according to the use of each component. Specifically, when a proposed component is not used, i.e., $X$ , we follow the conventional SD-based IR method setting, as described in Section 3.1. Except for Exp-(7), we adopt the FM loss as a default, since it relates to task network training rather than the IR method. In addition, to assess how well task-relevant details are restored, we measure the distance in the feature space of the task network $\\mathcal{H}_{\\mathrm{HQ}}$ , which is pre-trained on HQ images. We denote this fidelity metric as feature distance, defined as $f_{d} = \\left\\| \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{EDTR}}) - \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{HQ}})\\right\\|_{1}$ .", + "bbox": [ + 89, + 633, + 482, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Exp-(1) shows performance without any of the proposed components, i.e., conventional SD-based IR method [30]. Compared to EDTR performance (last row), although it achieves a more perceptually realistic restoration, as indicated by the higher Q-Align score, its large $f_{d}$ value suggests that task-relevant details are not well restored, resulting in a $4.5\\%$ worse classification accuracy.", + "bbox": [ + 89, + 806, + 483, + 912 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d33e4b6da70d19f40fe84df9876c9d0421ecdac5c3641095990d2893fec3646d.jpg", + "image_caption": [ + "(a) LQ" + ], + "image_footnote": [], + "bbox": [ + 516, + 263, + 612, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ae631f535949326764b912ca8f520abbf462487acb770ba5aeb90df2191fa7e8.jpg", + "image_caption": [ + "(b) Exp-(2)" + ], + "image_footnote": [], + "bbox": [ + 614, + 265, + 707, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/47d220f2689a8a3ff4604beae42dcac058213c83e7c718cf9cff4c97d34b86f3.jpg", + "image_caption": [ + "(c)EDTR(Ours)", + "Figure 6. Visualization comparison in classification." + ], + "image_footnote": [], + "bbox": [ + 710, + 265, + 805, + 339 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e212bf1ba0e3581f1e588acaca487471e3e85f240287a0929948a4d35caf4ad1.jpg", + "image_caption": [ + "(d) HQ" + ], + "image_footnote": [], + "bbox": [ + 808, + 265, + 903, + 339 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of pre-restoration & partial diffusion. Exp-(2) shows performance without our design choice for the diffusion process, i.e., SD is directly applied to TDIR. Although it incorporates a loss for restoring task-relevant details, its task performance is even worse than that of Exp-(1). Figure 6 illustrates that Exp-(2) leads to excessive generation and produces non-existent content, in contrast to the results of EDTR, ultimately failing to improve task performance. This highlights the necessity of a tailored diffusion process design for TDIR. Exp-(3) shows that even with short-step denoising, EDTR without partial diffusion does not perform well for TDIR, demonstrating the importance of partial diffusion. Exp-(4) further shows that the pre-restoration process is also essential for achieving both strong task performance and good visual quality. For performance regarding the pre-restoration network, see Supplementary Section S3.", + "bbox": [ + 511, + 375, + 906, + 616 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of HLF & FM losses. Exp-(5) and Exp-(6) show EDTR performance without HLF and FM losses, respectively. In both cases, task performance drops significantly, highlighting the importance of each training loss term. Interestingly, Exp-(5) results in a notably lower Q-Align score, indicating that the HLF loss is crucial for achieving visually pleasing results, and it enables the generation of such results even in the challenging one-step denoising scenario. This finding is consistent with previous research [47], as the HLF loss shares similarities with LPIPS [64] in measuring distance in feature space. In addition, incorporating feature spaces from two networks in the HLF loss (as in Equation (7)) yields a $+1.0\\%$ gain compared to using only one (see Supplementary Section S5 for details).", + "bbox": [ + 511, + 621, + 908, + 834 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effect of SD. Exp-(7) shows the performance of EDTR without pre-trained SD weights. Specifically, we remove the frozen denoising network (purple module in Figure 2a); thus, the total number of trainable parameters remains", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10157", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nearly the same. Compared to this setting, EDTR largely reduces the $f_{d}$ , resulting in a significant $4.6\\%$ increase in classification accuracy. These results highlight the crucial role of the diffusion prior in restoring task-relevant details.", + "bbox": [ + 89, + 90, + 482, + 151 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.2.2. Effect of the number of denoising steps.", + "text_level": 1, + "bbox": [ + 89, + 159, + 413, + 175 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 3 shows the performance of EDTR with varying numbers of denoising steps during inference. As shown in the table, a larger number of denoising steps (e.g., 50) leads to decreased task performance. This supports our claim that repeatedly introducing details is not beneficial for TDIR and may even dilute crucial task-relevant information. In addition, we observe that short-step denoising ( $n = 4$ ) is sufficient to produce visually pleasing results, as indicated by its favorable Q-Align score—further confirming the effectiveness of the HLF loss in guiding perceptual quality.", + "bbox": [ + 89, + 178, + 482, + 330 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Generalizability of EDTR", + "text_level": 1, + "bbox": [ + 89, + 338, + 326, + 356 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Performance of EDTR across various tasks. In Table 4, we present the performance of EDTR-1 step models, each trained separately on a specific task. Note that the diagonal (blue) values represent the performance of the original EDTR on its respective task. Since the original EDTR is independently trained for a single task, it performs well on its target task but less effectively on others. To demonstrate that our EDTR can be extended for broader applicability, we additionally train a single model on all three high-level vision tasks simultaneously, referring to this model as EDTR_All-in-One. The last row in the table presents the results. We observe that this model achieves performance comparable to, or even better than, the original EDTR across all three tasks, highlighting its potential for generalization to a variety of applications.", + "bbox": [ + 89, + 361, + 482, + 589 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Generalization toward unseen datasets. Table 5 shows the classification performance on unseen datasets during training. Specifically, we train the IR models using TinyImageNet $^{1}$ [23] and evaluate them on CUB200 [51] and StanfordCars [22] datasets, which were not seen during training. When evaluating on these unseen datasets, the task networks trained on HQ images corresponding to each dataset are used. The results show that our EDTR significantly outperforms other methods on these unseen datasets, demonstrating the robustness of our approach.", + "bbox": [ + 89, + 592, + 482, + 743 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Results on real-world images. Figure 7 presents real-world object detection results and restored image visualizations. We compare (a) object detection without restoration and (b) EDTR-4 step trained on Mixture- $B$ degradation. Due to complex real-world degradation, detection without restoration fails to identify the person in the image. In contrast, EDTR effectively restores key high-frequency details,", + "bbox": [ + 89, + 747, + 482, + 854 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/917d7a1c84b1ab980b6b238e943048902b39fc5ea66fdb611472d2d809611544.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
# of denoising steps4304050
Classification (Acc↑ / Q-Align↑)68.4 / 3.4967.7 / 3.4667.1 / 3.4666.8 / 3.45
Segmentation (mIoU↑ / Q-Align↑)62.9 / 3.7462.4 / 3.5462.4 / 3.4661.8 / 3.43
Detection (mAP↑ / Q-Align↑)31.9 / 4.0231.8 / 3.9731.8 / 3.9931.6 / 3.94
", + "bbox": [ + 517, + 90, + 901, + 138 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e71a61cce196b0f80fbc233f0924c53464c55ffc255c2acb08f3a5dd746fc4c9.jpg", + "table_caption": [ + "Table 3. Effect of the number of denoising steps in EDTR." + ], + "table_footnote": [], + "table_body": "
Eval task Used modelClassification (Acc↑)Segmentation (mIoU↑)Detection (mAP↑)
EDTR_Classification68.848.714.3
EDTR_Segmentation56.860.424.8
EDTR_Detection59.358.930.6
EDTR_All-in-One68.461.030.4
", + "bbox": [ + 517, + 162, + 901, + 233 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e056ece91dfd62a1e745648c547d10b62af152ad382cf92860451133a1afc781.jpg", + "table_caption": [ + "Table 4. Performance of EDTR models across various tasks. EDTR_Task represents the EDTR model optimized for each task." + ], + "table_footnote": [], + "table_body": "
Dataset IR methodsSeen \nTiny-ImageNet† [23]Unseen
CUB200 [51]StanfordCars [22]
Oracle83.782.590.4
No restoration61.210.112.1
SwinIR [27]68.021.838.7
SR4IR [20]68.943.346.1
EDTR-1 step (Ours)75.152.567.0
EDTR-4 step (Ours)76.153.066.4
", + "bbox": [ + 517, + 271, + 901, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Classification on seen and unseen datasets.", + "bbox": [ + 545, + 364, + 867, + 376 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/99652a1c9d21ebdc2f0c179346df90c069af9df990824ac5c6b49e2cd15f75f4.jpg", + "image_caption": [ + "(a) LQ (No restoration)" + ], + "image_footnote": [], + "bbox": [ + 516, + 383, + 705, + 516 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/badd37325bea15074e57bfef96b598f343cf59a6673bbcad4f838ff1f9d58085.jpg", + "image_caption": [ + "(b)EDTR(Ours)", + "Figure 7. Real-world object detection results and visualization of the restored image. The image is from \"56.png\" in the real-world image set RealPhoto60 [61]." + ], + "image_footnote": [], + "bbox": [ + 715, + 383, + 905, + 516 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "such as clothing wrinkles on a person, enabling successful detection. These results demonstrate that EDTR can handle complex real-world degradations and detect objects in degraded images by effectively leveraging the powerful diffusion prior. For more visualization results, please refer to Section S12 in our supplementary materials.", + "bbox": [ + 511, + 582, + 905, + 674 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 681, + 633, + 698 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper proposes EDTR, a novel task-driven image restoration (TDIR) method that directly leverages a pretrained diffusion prior to enhance high-level vision task performance. We observe that while incorporating the diffusion prior produces visually pleasing results, restoring task-relevant details remains challenging. To address this, we propose tailored designs for exploiting the diffusion prior in TDIR. Our method significantly improves the performance of various tasks in complex image degradation scenarios while achieving visually appealing results, demonstrating its broad applicability and effectiveness. We anticipate that our approach will provide new insights into leveraging diffusion priors to enhance task performance.", + "bbox": [ + 511, + 704, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1Since the original Tiny-ImageNet [23] has a resolution of 64x64, which is too small for IR problem, we resampled from ImageNet [42] with the same classes and number of images to create Tiny-ImageNet†.", + "bbox": [ + 89, + 862, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10158", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], and the Industrial Technology Alchemist Project [No. RS-2024-00432410] funded by MOTIE, Korea.", + "bbox": [ + 89, + 90, + 485, + 176 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 204, + 187, + 219 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In CVPR, 2018. 2", + "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, 2018. 2", + "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2", + "[4] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 2", + "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2", + "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2016. 5", + "[7] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. In CVPR, 2023. 2", + "[8] Dengxin Dai, Yujuan Wang, Yuhua Chen, and Luc Van Gool. Is image super-resolution helpful for other vision tasks? In WACV, 2016. 1, 2", + "[9] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In CVPR, 2019. 2", + "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 2", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2", + "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 5", + "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018. 2", + "[14] Muhammad Haris, Shakhnarovich Greg, and Ukita Norimichi. Task-driven super resolution: Object detection in low-resolution images. In ICONIP, 2021. 2, 6" + ], + "bbox": [ + 93, + 229, + 480, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5", + "[16] Geoffrey Hinton. Distilling the knowledge in a neural network. In NIPS Workshop, 2014. 5", + "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[18] Shih-Chia Huang, Trung-Hieu Le, and Da-Wei Jaw. Dsnet: Joint semantic learning for object detection in inclement weather conditions. TPAMI, 2020. 2", + "[19] Jiwon Kim, Jung Kown Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2", + "[20] Jaeha Kim, Junghun Oh, and Kyoung Mu Lee. Beyond image super-resolution for image recognition with task-driven perceptual loss. In CVPR, 2024. 1, 2, 4, 5, 6, 8", + "[21] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 3", + "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, 2013. 8", + "[23] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. https://api_semanticscholar.org/ CorpusID:16664790, 2015.8", + "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2", + "[25] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In ICCV, 2017. 2", + "[26] Chengyang Li, Heng Zhou, Yang Liu, Caidong Yang, Yongqiang Xie, Zhongbo Li, and Liping Zhu. Detection-friendly dehazing: Object detection in real-world hazy scenes. TPAMI, 2023. 2", + "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV, 2021. 2, 3, 5, 6, 8", + "[28] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 2", + "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 5", + "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. In ECCV, 2024. 1, 2, 3, 4, 7", + "[31] Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, and Thomas S Huang. When image denoising meets high-level vision tasks: A deep learning approach. In *IJCAI*, 2018. 1, 2", + "[32] Ding Liu, Bihan Wen, Jianbo Jiao, Xianming Liu, Zhangyang Wang, and Thomas S Huang. Connecting image denoising and high-level vision tasks via deep learning. TIP, 2020. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10159", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[33] Wenyu Liu, Gaofeng Ren, Runsheng Yu, Shi Guo, Jianke Zhu, and Lei Zhang. Image-adaptive yolo for object detection in adverse weather conditions. In AAAI, 2022. 2", + "[34] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In CVPR, 2024. 2", + "[35] Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image superresolution. In CVPR, 2025. 2", + "[36] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 2012. 5", + "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4", + "[38] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In ECCV, 2020. 2", + "[39] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In ECCV, 2018. 1, 2", + "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, 2015. 5", + "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5", + "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 2015. 8", + "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5", + "[44] Jacob Shermeyer and Adam Van Etten. The effects of superresolution on object detection performance in satellite imagery. In CVPR Workshops, 2019. 2", + "[45] Taeyoung Son, Juwon Kang, Namyup Kim, Sunghyun Cho, and Suha Kwak. Urie: Universal image enhancement for visual recognition in the wild. In ECCV, 2020. 2", + "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[47] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 7", + "[48] Jong-Chyi Su and Subhransu Maji. Adapting models to signal degradation using distillation. In BMVC, 2017. 5", + "[49] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2, 3", + "[50] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. In NeurIPS, 2022. 2" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 8", + "[52] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 2, 3, 4", + "[53] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: enhanced super-resolution generative adversarial networks. In ECCV Workshops, 2018. 2", + "[54] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 4", + "[55] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 2", + "[56] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In ICML, 2024. 5", + "[57] Jiawei Wu and Zhi Jin. Unsupervised variational translator for bridging image restoration and high-level vision tasks. In ECCV, 2024. 1, 2", + "[58] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 4", + "[59] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2", + "[60] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024. 2, 3", + "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2, 3, 8", + "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2", + "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5", + "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7", + "[65] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 2", + "[66] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 2" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10160", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[67] Xiaotong Zhao, Wei Li, Yifan Zhang, and Zhiyong Feng. Residual super-resolution single shot network for low-resolution object detection. IEEE Access, 2018. 2, 6", + "[68] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022." + ], + "bbox": [ + 89, + 90, + 482, + 176 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5", + "bbox": [ + 125, + 178, + 133, + 188 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10161", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_model.json b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_model.json new file mode 100644 index 0000000000000000000000000000000000000000..961eb71ced3e8dd25ab867ebc4d72919918a5cb0 --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_model.json @@ -0,0 +1,2939 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.19, + 0.131, + 0.812, + 0.154 + ], + "angle": 0, + "content": "Exploiting Diffusion Prior for Task-driven Image Restoration" + }, + { + "type": "text", + "bbox": [ + 0.305, + 0.179, + 0.693, + 0.199 + ], + "angle": 0, + "content": "Jaeha Kim1 Junghun Oh1 Kyoung Mu Lee1,2" + }, + { + "type": "text", + "bbox": [ + 0.249, + 0.199, + 0.75, + 0.218 + ], + "angle": 0, + "content": "\\(^{1}\\)Dept. of ECE&ASRI, \\(^{2}\\)IPAI, Seoul National University, Korea" + }, + { + "type": "text", + "bbox": [ + 0.284, + 0.219, + 0.71, + 0.234 + ], + "angle": 0, + "content": "jhkim97s2@gmail.com, {dh6dh, kyoungmu}@snu.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.269, + 0.327, + 0.285 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.302, + 0.486, + 0.635 + ], + "angle": 0, + "content": "Task-driven image restoration (TDIR) has recently emerged to address performance drops in high-level vision tasks caused by low-quality (LQ) inputs. Previous TDIR methods struggle to handle practical scenarios in which images are degraded by multiple complex factors, leaving minimal clues for restoration. This motivates us to leverage the diffusion prior, one of the most powerful natural image priors. However, while the diffusion prior can help generate visually plausible results, using it to restore task-relevant details remains challenging, even when combined with recent TDIR methods. To address this, we propose EDTR, which effectively harnesses the power of diffusion prior to restore task-relevant details. Specifically, we propose directly leveraging useful clues from LQ images in the diffusion process by generating from pixel-error-based pre-restored LQ images with mild noise added. Moreover, we employ a small number of denoising steps to prevent the generation of redundant details that dilute crucial task-related information. We demonstrate that our method effectively utilizes diffusion prior for TDIR, significantly enhancing task performance and visual quality across diverse tasks with multiple complex degradations." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.664, + 0.222, + 0.68 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.688, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Image degradation often occurs in the real world due to various factors, such as transmission loss, limited camera performance, or poor shooting conditions. In such cases, key high-frequency details can be lost, resulting in significant performance drops in high-level vision tasks, such as image classification, semantic segmentation, and object detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Image restoration (IR) can be a promising solution to improve the performance of high-level vision tasks. It involves recovering missing details in low-quality (LQ) images using a learned natural image prior. However, recent studies [8, 20, 39, 57] show that the straightforward application of IR methods as a pre-processing step is not very effective in restoring essential information for high-level vision tasks. This has led to the emergence of Task-driven Image" + }, + { + "type": "image", + "bbox": [ + 0.516, + 0.335, + 0.617, + 0.45 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.524, + 0.452, + 0.61, + 0.465 + ], + "angle": 0, + "content": "(a) Low-quality" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.268, + 0.761, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.649, + 0.386, + 0.739, + 0.398 + ], + "angle": 0, + "content": "(b) DiffBIR [30]" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.268, + 0.901, + 0.384 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.769, + 0.386, + 0.9, + 0.399 + ], + "angle": 0, + "content": "(c) SR4IR [20] \\(+\\mathrm{SD}\\) [41]" + }, + { + "type": "image", + "bbox": [ + 0.627, + 0.399, + 0.761, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.645, + 0.519, + 0.743, + 0.531 + ], + "angle": 0, + "content": "(d) EDTR (Ours)" + }, + { + "type": "image", + "bbox": [ + 0.769, + 0.399, + 0.901, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.791, + 0.519, + 0.878, + 0.532 + ], + "angle": 0, + "content": "(e) High-quality" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.54, + 0.907, + 0.624 + ], + "angle": 0, + "content": "Figure 1. Effect of diffusion prior on image classification performance. While incorporating the StableDiffusion [41] produces visually plausible results as (b), it lacks task-relevant details (e.g., eye shape), even when combined with the previous TDIR method as (c), leading to incorrect classification. In contrast, EDTR successfully restores details as (d), resulting in correct classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.642, + 0.905, + 0.672 + ], + "angle": 0, + "content": "Restoration (TDIR), which aims to improve visual quality in ways that benefit high-level vision tasks." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.673, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Previous TDIR methods address each type of degradation separately, e.g., super-resolution [20], denoising [31]. This restricts their practical applicability, as real-world images often exhibit multiple degradations simultaneously. To address this, we tackle a more realistic scenario by jointly handling multiple complex degradations. However, we observe that previous TDIR methods achieve limited performance gains in such scenarios, as they struggle to utilize the few available clues in degraded images to restore task-relevant details. This highlights the need for a strong natural image prior for the TDIR problem." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In this paper, we leverage the prior knowledge from the pre-trained StableDiffusion (SD) [41], which provides one of the most powerful natural image priors. However, we observe that while the conventional SD-based IR method [30]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10151" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.347 + ], + "angle": 0, + "content": "can generate visually pleasing results, its straightforward adoption fails to restore task-relevant details, leading to incorrect predictions, as shown in Figure 1b. More importantly, we observe that SD still struggles to restore task-relevant details even when combined with a recent state-of-the-art TDIR method [20], as shown in Figure 1c. In other words, while diffusion prior is powerful, it must be properly adapted, as its standard form fails to improve TDIR. This issue arises because restoring task-relevant details is inherently challenging in the standard diffusion process used in conventional SD-based IR methods [30, 52, 59-61]. We identify two key reasons: (1) Although the LQ image is used as a condition for SD, generating from pure noise hinders the direct utilization of useful clues in LQ images, making it difficult to restore task-relevant details. (2) Long-step denoising, around 50, introduces redundant details that dilute crucial task-related information for TDIR." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.349, + 0.483, + 0.652 + ], + "angle": 0, + "content": "To address this, we propose EDTR, Exploiting Diffusion prior for Task-driven Image Restoration, the TDIR method adopting the diffusion prior in ways that directly leverage its power to restore task-relevant details in complex degradation scenarios. The primary component of EDTR is threefold: (1) We directly incorporate LQ images into the diffusion process. Since SD was not originally trained to handle complex degradations, directly feeding LQ images into the diffusion process is ineffective. Therefore, we first prerestore the LQ images using a pixel-error-based IR method, then generate from these images with mild noise added, rather than generating from pure noise, a process referred to as pre-restoration and partial diffusion. (2) We apply a short-step denoising (e.g., 1, 4) during inference. We claim that increasing the number of steps (e.g., 50) reduces task performance by introducing redundant details that dilute crucial details for TDIR. (3) We introduce tailored training loss terms to guide diffusion prior to restore task-relevant details. These loss terms are essential for achieving visually appealing results, even in a one-step denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.652, + 0.482, + 0.803 + ], + "angle": 0, + "content": "We validate our method in three high-level vision tasks: image classification, semantic segmentation, and object detection. These tasks involve complex degradation scenarios, including a mix of downsampling, blur, noise, and JPEG artifacts. Experimental results demonstrate that EDTR effectively harnesses the power of diffusion prior, significantly improving task performance (e.g., about \\(+50\\%\\) improvement in detection) while producing visually pleasing results, as shown in Figure 1d, and generalizes well to real-world images. We summarize our contributions as:" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.805, + 0.483, + 0.835 + ], + "angle": 0, + "content": "- We propose EDTR, the TDIR method that directly leverages a powerful pre-trained diffusion prior." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.838, + 0.483, + 0.868 + ], + "angle": 0, + "content": "- We propose tailored designs for exploiting the diffusion prior to effectively restore task-relevant details." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.872, + 0.482, + 0.902 + ], + "angle": 0, + "content": "- We significantly improve task performance while producing visually pleasing results under complex degradation." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.805, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.655, + 0.105 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.11, + 0.905, + 0.336 + ], + "angle": 0, + "content": "General image restoration. Since the pioneering SRCNN [10], numerous deep neural network-based methods [5, 9, 13, 19, 24, 28, 38, 53, 65, 66] have emerged in the field of image restoration (IR). Vision transformers [11]-based IR methods [3, 7, 55, 62], such as SwinIR [27], have also demonstrated strong performance. Recently, IR studies [30, 34, 52, 60, 61] have focused on leveraging the large pre-trained diffusion models, primarily StableDiffusion (SD) [41], which provides a powerful natural image prior essential for addressing the ill-posedness of the IR problem. StableSR [52] and DiffBIR [30] introduce trainable modules alongside a frozen SD backbone, similar to ControlNet [63]. More recent works [35, 49, 59, 61] have focused on fully harnessing SD's generative capabilities by designing carefully crafted input conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.341, + 0.905, + 0.507 + ], + "angle": 0, + "content": "Task-driven image restoration. Taking a different perspective, recent studies [1, 2, 33, 44, 50, 57] have focused on the emerging field of Task-driven Image Restoration (TDIR), which aims to improve high-level vision task performance when handling LQ images. AOD-Net [25] suggests that using dehazing as a preprocessing step and applying object detection can enhance detection performance in foggy conditions. However, Pei et al. [39] show that applying IR methods as a simple preprocessing step does not significantly improve detection performance, with similar findings reported in the super-resolution field [8]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.509, + 0.905, + 0.75 + ], + "angle": 0, + "content": "To address this, recent TDIR methods [14, 18, 26, 31, 32, 45, 67] have started training the IR network in combination with the task networks. TDSR [14] trains the IR network using both pixel-error loss and detection loss from a pre-trained object detection network, with a dedicated training schedule. RSRSSN [67] proposes an end-to-end trainable IR and detection networks with Feature Maps Multi-box loss. SR4IR [20] introduces a joint training framework for TDIR with a task-driven perceptual loss to guide the IR network in learning task-relevant features. More recently, UniRestore [4] introduces the diffusion prior into TDIR; however, its diffusion-based training is independent of taskdriven training, performed in two stages, limiting the ability to fully leverage the diffusion prior for TDIR. We instead propose integrating the diffusion prior directly within taskdriven training to better restore task-relevant details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.761, + 0.687, + 0.777 + ], + "angle": 0, + "content": "3. Proposed Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Problem definition. Let \\( I_{\\mathrm{HQ}} \\) and \\( I_{\\mathrm{LQ}} \\) denote the high-quality (HQ) and low-quality (LQ) images, respectively, with paired labels \\( y \\) from corresponding high-level vision task, e.g., class label in the image classification. The goal of Task-driven Image Restoration (TDIR) is to restore LQ images in a way that enhances the subsequent task performance while also generating visually appealing results. Formally, given a trainable image restoration (IR) model \\( \\mathcal{R} \\)" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10152" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.102, + 0.089, + 0.898, + 0.222 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.289, + 0.225, + 0.707, + 0.238 + ], + "angle": 0, + "content": "(a) The architecture of the proposed EDTR with high-level vision task network." + }, + { + "type": "image", + "bbox": [ + 0.127, + 0.241, + 0.87, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.398, + 0.701, + 0.411 + ], + "angle": 0, + "content": "(b) The joint training procedure of EDTR and high-level vision task network." + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.412, + 0.89, + 0.427 + ], + "angle": 0, + "content": "Figure 2. The overall framework of the proposed method. \\( I_{\\mathrm{EDTR,train}} \\) is the EDTR restored image when EDTR is in training mode." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.497 + ], + "angle": 0, + "content": "and a trainable task network \\(\\mathcal{H}\\), our aim is to maximize \\(\\mathbf{M}_{\\mathrm{H}}(\\mathcal{H}(\\mathcal{R}(I_{\\mathrm{LQ}})),y)\\), where \\(\\mathbf{M}_{\\mathrm{H}}\\) is the evaluation metric for the task, e.g., the accuracy for classification, while also achieving visually pleasing restored results \\(\\mathcal{R}(I_{\\mathrm{LQ}})\\)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.505, + 0.436, + 0.521 + ], + "angle": 0, + "content": "3.1. Preliminary: Conventional SD-based IR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.649 + ], + "angle": 0, + "content": "Unlike pure generation, IR requires generating images that are faithful to the given LQ images. To achieve this, previous StableDiffusion (SD)-based IR methods [30, 49, 52, 60, 61] commonly use LQ images as a condition for SD through ControlNet [63]. Note that, for better efficiency and more stable training, SD applies the diffusion process in the latent space rather than the image space, following the encoder and decoder structure of the VAE [21]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.65, + 0.484, + 0.696 + ], + "angle": 0, + "content": "During training, given an image pair \\((I_{\\mathrm{HQ}}, I_{\\mathrm{LQ}})\\), a VAE encoder \\(\\mathcal{E}\\), and Gaussian noise with variance \\(\\beta_t \\in (0,1)\\), \\(I_{\\mathrm{HQ}}\\) is first diffused according to DDPM [17] as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.182, + 0.702, + 0.482, + 0.74 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t} = q \\left(z _ {t} \\mid z _ {0} = \\mathcal {E} \\left(I _ {\\mathrm {H Q}}\\right)\\right) \\tag {1} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {E} (I _ {\\mathrm {H Q}}) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.753, + 0.484, + 0.83 + ], + "angle": 0, + "content": "where \\(q(z_{t}|z_{t - 1})\\coloneqq \\mathcal{N}(\\sqrt{1 - \\beta_{t}} z_{t - 1},\\beta_{t}\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t},\\) \\(\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}\\) , and \\(\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})\\) . The timestep \\(t\\) is randomly selected from \\([1,T]\\) , where \\(T\\) is the total number of timesteps. Then the trainable module, usually ControlNet, is trained to predict the injected noise as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.186, + 0.838, + 0.483, + 0.856 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(I _ {\\mathrm {L Q}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "where \\(\\epsilon_{\\theta}\\) is a denoising network combined with ControlNet. Note that \\(I_{\\mathrm{LQ}}\\) is used only as condition input for ControlNet." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.435, + 0.906, + 0.527 + ], + "angle": 0, + "content": "During inference, denoising starts from a pure noise. It gradually moves toward less noisy images by iteratively alternating between denoising and adding noise, using sampling methods [17, 37, 46]. After approximately 50 steps, the process obtains the predicted latent \\(\\hat{z}_0\\), which is then mapped back to the image space using the VAE decoder." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.532, + 0.704, + 0.546 + ], + "angle": 0, + "content": "3.2.EDTR Architecture" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.551, + 0.906, + 0.656 + ], + "angle": 0, + "content": "Pre-restoration and partial diffusion. As noted earlier, conventional SD-based IR methods begin the denoising process from pure noise during the inference. Although \\( I_{\\mathrm{LQ}} \\) are used as a condition through ControlNet, its behavior heavily depends on SD's generative capacity. This poses a challenge for TDIR, as it hinders the use of valuable clues in LQ images, making it difficult to restore task-relevant details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.656, + 0.907, + 0.807 + ], + "angle": 0, + "content": "To address this, we directly incorporate \\( I_{\\mathrm{LQ}} \\) into the diffusion process, i.e., Equation (1). However, this approach poses two key challenges: (1) Since SD was not originally trained on \\( I_{\\mathrm{LQ}} \\), which include complex degradations such as JPEG artifacts, directly incorporating \\( I_{\\mathrm{LQ}} \\) into the diffusion process is ineffective at leveraging the power of SD. (2) When \\( I_{\\mathrm{LQ}} \\) is fed directly into SD, degradation artifacts can be interpreted as generation cues, producing tasks-irrelevant details. Instead, we first pre-restore \\( I_{\\mathrm{LQ}} \\) using a pixel-error-based IR method \\( \\mathcal{R}_{\\mathrm{pix}} \\), such as SwinIR [27]:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.815, + 0.905, + 0.833 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {p r e - r e s}} = \\mathcal {E} \\left(\\mathcal {R} _ {\\text {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right). \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Although introducing strong noise in the diffusion process enables the generation of additional information by heavily relying on the diffusion prior, preserving valuable clues in the LQ image is more crucial for the TDIR prob" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10153" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.088, + 0.485, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.094, + 0.17, + 0.48, + 0.185 + ], + "angle": 0, + "content": "Figure 3. Illustration of pre-restoration and partial diffusion." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.193, + 0.484, + 0.24 + ], + "angle": 0, + "content": "lem. Thus, we propose adding mild noise to the pre-restored images and starting generation from it. We denote this technique as partial diffusion, which can be formally defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.244, + 0.483, + 0.282 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} z _ {t, \\text {p a r t i a l}} = q \\left(z _ {t \\leq t _ {p}} \\mid z _ {0} = z _ {\\text {p r e - r e s}}\\right) \\tag {4} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} z _ {\\text {p r e - r e s}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.292, + 0.484, + 0.354 + ], + "angle": 0, + "content": "where \\( z_{t,\\text{partial}} \\) is the partial diffusion result and \\( t_p \\) is a fixed hyper-parameter with a relatively smaller value than \\( T \\). We empirically set \\( (t_p, T) = (200, 1000) \\). Figure 3 illustrates the proposed diffusion process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.357, + 0.484, + 0.569 + ], + "angle": 0, + "content": "Short-step denoising. As noted in the preliminary section, conventional SD-based IR methods employ long-step denoising (e.g., 50). However, in the context of the TDIR problem, we find that a large number of denoising steps negatively impact on task performance. Although each step introduces details through the diffusion prior, repeatedly adding such details can generate redundant content, thereby diluting crucial task-relevant information. Hence, we adopt a short-step denoising (e.g., 1, 4) during inference. Specifically, for a given number of denoising steps \\( n \\), we denoise \\( z_{t_p,\\text{partial}} \\) by passing it through \\( n \\) timesteps: \\( \\mathcal{T} = [t_p, \\lfloor t_p / n \\cdot (n - 1) \\rfloor, \\dots, \\lfloor t_p / n \\rfloor] \\), following [37], and obtain the final denoising result \\( z_{\\text{diff-res}} \\). In the one-step denosing case (i.e., \\( n = 1 \\)), this process simplifies to:" + }, + { + "type": "equation", + "bbox": [ + 0.104, + 0.576, + 0.483, + 0.628 + ], + "angle": 0, + "content": "\\[\nz _ {\\text {d i f f - r e s}} = \\frac {z _ {t _ {p} , \\text {p a r t i a l}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {p}}} \\epsilon_ {\\theta} \\left(z _ {t _ {p} , \\text {p a r t i a l}}, t _ {p} , z _ {\\text {p r e - r e s}}\\right)}{\\sqrt {\\bar {\\alpha} _ {t _ {p}}}}. \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.63, + 0.484, + 0.72 + ], + "angle": 0, + "content": "Note that conventional SD-based IR methods [30, 52] struggle to generate visually pleasing results in short-step denoising scenarios [54, 58]. However, we find that optimizing loss functions, specifically designed to restore task-relevant details as discussed in Section 3.3, enables the achievement of visually pleasing results, even in one-step denoising." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.811 + ], + "angle": 0, + "content": "Figure 4 compares the overall IR procedure. Unlike conventional methods, which start from pure noise and require around 50 denoising steps, we use the input LQ image with pre-restoration and partial diffusion, followed by short-step denoising. This approach directly leverages useful clues in LQ images to restore task-relevant details." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.816, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Decoding into RGB space. After the denoising process, we transform the \\( z_{\\mathrm{diff-res}} \\) back into the image space using the VAE decoder \\( \\mathcal{D} \\). We further apply Wavelet color correction [52] to achieve a faithful reconstruction as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.883, + 0.483, + 0.901 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathrm {E D T R}} = \\mathbf {H} \\left(\\mathcal {D} \\left(z _ {\\text {d i f f - r e s}}\\right)\\right) + \\mathbf {L} \\left(\\mathcal {R} _ {\\operatorname {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.537, + 0.089, + 0.887, + 0.24 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.242, + 0.871, + 0.256 + ], + "angle": 0, + "content": "(b) Pre-restoration, partial diffusion, and short-step denosing." + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.26, + 0.905, + 0.288 + ], + "angle": 0, + "content": "Figure 4. The comparison of IR procedure. We omit the encoding, decoding, and conditioning for simplicity in visualization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.295, + 0.906, + 0.357 + ], + "angle": 0, + "content": "where \\(I_{\\mathrm{EDTR}}\\) is the final restored result from our EDTR. The \\(\\mathbf{H}(\\cdot)\\) and \\(\\mathbf{L}(\\cdot)\\) correspond to the high-frequency and low-frequency components extracted by Wavelet transform. Figure 2a illustrates the overall architecture of our EDTR." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.365, + 0.714, + 0.381 + ], + "angle": 0, + "content": "3.3. Training Framework" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.383, + 0.906, + 0.488 + ], + "angle": 0, + "content": "High-level feature loss. The goal of TDIR is to restore task-relevant details to enhance task performance. However, the noise prediction loss in Equation (2), which is conventionally used for training diffusion models, is not designed for this purpose. Hence, we introduce high-level feature (HLF) loss, which can guide the model to harness the diffusion prior for restoring task-relevant details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.489, + 0.906, + 0.61 + ], + "angle": 0, + "content": "HLF loss measures the feature distance between EDTR restored results and HQ images, following the previous TDIR method [20]. Specifically, we compute the HLF loss using the feature space of two task networks: the task network \\(\\mathcal{H}\\), which is optimized during training, and another task network \\(\\mathcal{H}_{\\mathrm{HQ}}\\), pre-trained on HQ images, to capture complementary task-relevant information (see Supplementary Section S5). Formally, the HLF loss is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.612, + 0.905, + 0.67 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {H L F}} = \\frac {1}{2} \\left(\\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R}, \\text {t r a i n}}\\right) - \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1} \\right. \\tag {7} \\\\ + \\left\\| \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {E D T R , t r a i n}}) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {H Q}}) \\right\\| _ {1}), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.672, + 0.906, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathcal{H}^f\\) and \\(\\mathcal{H}_{\\mathrm{HQ}}^f\\) are intermediate feature spaces of \\(\\mathcal{H}\\) and \\(\\mathcal{H}_{\\mathrm{HQ}}\\), respectively. Note that \\(I_{\\mathrm{EDTR,train}}\\) in Equation (7) represents the restored image from the one-step denoising result of \\(z_{t,\\mathrm{partial}}\\), where \\(t\\) is randomly sampled from \\(\\mathcal{T}\\) during training, following the conventional diffusion loss in Equation (2), regardless of denoising steps \\(n\\) during inference (see Supplementary Algorithm S1 for details). The HLF loss updates only EDTR and does not affect \\(\\mathcal{H}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.799, + 0.906, + 0.875 + ], + "angle": 0, + "content": "Task loss. While training the proposed EDTR to restore task-relevant details, we also train the task networks using the restored outputs. We utilize both \\( I_{\\mathrm{EDTR}} \\) and HQ images to stabilize training and introduce additional regularization. Formally, task loss \\( \\mathcal{L}_{\\mathrm{task}} \\) for the task network is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.601, + 0.884, + 0.905, + 0.901 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t a s k}} = f _ {\\text {t a s k}} \\left(\\mathcal {H} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right), y\\right), \\tag {8}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10154" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.182 + ], + "angle": 0, + "content": "where \\( f_{\\mathrm{task}} \\) is the standard loss function for each task, e.g., cross-entropy loss for classification, \\( I_{\\mathrm{EDTR}} \\oplus_{\\mathrm{HQ}} \\) represents batch-wise concatenated images, consisting of a half-batch from \\( I_{\\mathrm{EDTR}} \\) and a half-batch from \\( I_{\\mathrm{HQ}} \\), and \\( y \\) are corresponding task labels. Note that \\( \\mathcal{L}_{\\mathrm{task}} \\) updates only the task network \\( \\mathcal{H} \\) and does not update EDTR." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.185, + 0.483, + 0.247 + ], + "angle": 0, + "content": "Feature matching loss. To align the feature space of task network \\(\\mathcal{H}\\) with \\(\\mathcal{H}_{\\mathrm{HQ}}\\), which is well-trained to utilize high-frequency details present in HQ images, we further introduce feature matching (FM) loss, defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.145, + 0.25, + 0.483, + 0.276 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {F M}} = \\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1}. \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.281, + 0.483, + 0.357 + ], + "angle": 0, + "content": "For clarity, note that \\( I_{\\mathrm{EDTR}} \\) in Equations (8) and (9) is the \\( n \\) step denoising result, different from the one-step result \\( I_{\\mathrm{EDTR,train}} \\) in Equation (7). The \\( \\mathcal{L}_{\\mathrm{FM}} \\) can be interpreted as a form of cross-quality knowledge distillation [16, 48] at the feature level. Similar to the task loss, \\( \\mathcal{L}_{\\mathrm{FM}} \\) only updates \\( \\mathcal{H} \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.361, + 0.484, + 0.512 + ], + "angle": 0, + "content": "Joint training in an alternating manner. Since SD was originally trained to operate in the image space, backpropagating the task loss, such as cross-entropy loss that is not image-based, causes instability to the EDTR. To avoid this, we train both the EDTR and the task network \\(\\mathcal{H}\\) using each corresponding training loss in an alternating manner [20], as illustrated in Figure 2b. Specifically, in each training iteration, we first update the EDTR using the HLF loss, then update \\(\\mathcal{H}\\) using a combination of the task loss and the FM loss, which can be formally written as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.155, + 0.515, + 0.482, + 0.539 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {E D T R}}} \\mathcal {L} _ {\\mathrm {H L F}}, \\text {t h e n} \\min _ {\\boldsymbol {\\theta} _ {\\mathcal {H}}} \\mathcal {L} _ {\\text {t a s k}} + \\alpha \\mathcal {L} _ {\\mathrm {F M}}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.544, + 0.483, + 0.575 + ], + "angle": 0, + "content": "where \\(\\theta_{\\mathrm{EDTR}}\\) and \\(\\theta_{\\mathcal{H}}\\) are the trainable parameters of the EDTR and task network \\(\\mathcal{H}\\), and \\(\\alpha\\) is the hyper-parameter." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.586, + 0.224, + 0.603 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.607, + 0.483, + 0.699 + ], + "angle": 0, + "content": "We validate our EDTR on three representative high-level vision tasks: image classification, semantic segmentation, and object detection, under scenarios involving image degradation. Note that the final prediction \\(\\hat{y}\\) for each task is obtained as: \\(\\hat{y} = \\mathcal{H}(\\mathrm{EDTR}(I_{\\mathrm{LQ}}))\\), where \\(\\mathcal{H}\\) is the trained task network and \\(I_{\\mathrm{LQ}}\\) is the input low-quality (LQ) images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.702, + 0.483, + 0.838 + ], + "angle": 0, + "content": "Datasets. We use three high-level vision tasks with representative datasets: CUB200 [51] for classification, and PASCAL VOC2012 [12] for segmentation and detection. We construct two degraded LQ image sets: 1) Mixture-A: downsampling combined with JPEG compression, and 2) Mixture-B: a combination of downsampling, blur, noise, and compression, following the CodeFormer [68] degradation. Details on degradation types are in the second column of Table 1 and Section S1 in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Evaluation metrics. We use representative evaluation metrics for each task: top-1 accuracy (Acc) for classification, mean Intersection over Union (mIoU) for segmentation, and mean Average Precision (mAP) for detection. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.168 + ], + "angle": 0, + "content": "mAP score is measured following the standard COCO [29] evaluation setting. We use two no-reference metrics to assess visual appeal: NIQE [36] and Q-Align [56], where the latter leveraging large vision-language models. We also report the standard PSNR for comparison." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.174, + 0.907, + 0.326 + ], + "angle": 0, + "content": "Network architecture. We use StableDiffusion-2.1 [41] (SD) and ControlNet [63] for EDTR. ControlNet and the VAE decoder of SD are set as trainable modules. We also incorporate the pixel-error-based SwinIR [27] model as the pre-restoration network. For the task network, we employ representative architecture for each task: ResNet [15] for classification, DeepLabV3 [6] for segmentation, and FasterRCNN [40] for detection. The task network \\(\\mathcal{H}\\) shares the same architecture as \\(\\mathcal{H}_{\\mathrm{HQ}}\\) (the network pre-trained at HQ), and the weights of \\(\\mathcal{H}\\) are initialized from \\(\\mathcal{H}_{\\mathrm{HQ}}\\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.336, + 0.715, + 0.352 + ], + "angle": 0, + "content": "4.1. Experimental Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.359, + 0.906, + 0.48 + ], + "angle": 0, + "content": "Table 1 and Figure 5 present quantitative and qualitative results, respectively, across various types of degradation in image classification, semantic segmentation, and object detection. Note that in Table 1, 'Oracle', 'No restoration', and 'SwinIR' refer to the performance when the task network is trained and tested directly on HQ, LQ, and pixel-error-based SwinIR restored images, respectively. EDTR-\\(n\\) step refers to the EDTR model with \\(n\\) denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.486, + 0.907, + 0.773 + ], + "angle": 0, + "content": "Image classification. Table 1 shows that the proposed EDTR significantly outperforms previous methods under both types of degradation, achieving classification accuracy improvements of \\(+2.9\\%\\) and \\(+5.4\\%\\) over the previous state-of-the-art task-driven image restoration (TDIR) method, SR4IR [20]. Performance gains increase as the degradation becomes more severe, highlighting our method's efficacy in practical scenarios. Furthermore, EDTR achieves the highest scores in image quality assessments such as NIQE and Q-Align, indicating its effectiveness in producing visually pleasing results. However, despite the superior task performance, EDTR does not achieve the highest PSNR score, suggesting that PSNR is a less relevant metric in TDIR. As presented in the first row of Figure 5, our EDTR, leveraging a strong natural image prior, successfully restores detailed wing patterns—an important area for classification as shown by the Grad-CAM [43] in the upper-left—and correctly predicts the label, whereas the previous method fails to restore these details and misclassifies the image." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Semantic segmentation. The performance improvement in segmentation is more notable than in classification, with mIoU gains of \\(+9.1\\%\\) and \\(+11.9\\%\\) over the previous state-of-the-art. Although our EDTR achieves the second-best score in NIQE, it consistently achieves the highest Q-Align score, demonstrating its capability to generate perceptually realistic images. In addition, unlike classification, increasing the number of denoising steps from 1 to 4 leads to a task" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10155" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.09, + 0.905, + 0.304 + ], + "angle": 0, + "content": "
IR MethodsDegradation typeImage classificationSemantic segmentationObject detection
Task Acc↑ (%)Visual qualityTask mIoU↑ (%)Visual qualityTask mAP↑ (%)Visual quality
NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑
OracleNone82.55.373.69+inf67.04.714.02+inf36.93.703.95+inf
No restoration60.814.081.0323.7247.413.671.1622.4314.514.541.0421.08
SwinIR [27]70.09.972.5525.2156.28.662.5423.4721.08.962.2221.70
TDSR [14]Mixture-A:66.17.532.5124.9951.37.672.3723.4116.26.432.1321.62
RSRSSN [67]SR(s=8)+69.27.322.7524.0453.06.972.5823.0622.36.592.1021.40
SR4IR [20]JPEG(q=75)71.55.413.3724.0756.24.803.1622.8522.35.383.0020.95
EDTR-1 step (Ours)74.44.683.5323.6364.15.303.5922.3331.24.623.6820.60
EDTR-4 step (Ours)74.14.263.8122.4665.35.173.9121.4833.44.194.0119.89
No restoration47.615.111.0823.0540.214.841.1922.100.015.241.0720.86
SwinIR [27]Mixture-B:60.710.262.0725.0450.49.472.3023.6722.19.601.9222.20
TDSR [14]SR(s[1,16]) +56.47.292.2024.6846.08.072.1323.5514.66.361.9421.97
RSRSSN [67]Blur(σ[0,8]) +61.47.352.3523.3548.27.722.3623.0123.16.811.8621.68
SR4IR [20]Noise(σ[0,10]) +63.46.083.1123.6251.04.723.1522.7122.64.523.1420.97
EDTR-1 step (Ours)JPEG(q[50,100])68.84.753.4823.0360.46.363.5522.0430.64.483.6420.70
EDTR-4 step (Ours)68.44.463.4923.2262.95.803.7422.2431.94.124.0220.29
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.304, + 0.907, + 0.334 + ], + "angle": 0, + "content": "Table 1. Performance of high-level vision tasks and quality of restored image results under various degradation types. We indicate the best and second-best results with bold and underlined formatting, respectively." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.347, + 0.295, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.105, + 0.479, + 0.264, + 0.493 + ], + "angle": 0, + "content": "pred: [Blue Grosbeak]" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.347, + 0.498, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.308, + 0.479, + 0.495, + 0.492 + ], + "angle": 0, + "content": "pred: [Blue Grosbeak] X" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.347, + 0.702, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.507, + 0.479, + 0.672, + 0.492 + ], + "angle": 0, + "content": "pred: [Indigo Bunting]" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.347, + 0.905, + 0.477 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.708, + 0.479, + 0.899, + 0.492 + ], + "angle": 0, + "content": "pred: [Indigo Bunting]" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.503, + 0.196, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.572, + 0.195, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.504, + 0.296, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.197, + 0.571, + 0.296, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.504, + 0.399, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.298, + 0.571, + 0.398, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.504, + 0.499, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.4, + 0.571, + 0.498, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.504, + 0.602, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.501, + 0.571, + 0.6, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.504, + 0.702, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.571, + 0.702, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.504, + 0.805, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.704, + 0.571, + 0.803, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.504, + 0.904, + 0.571 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.571, + 0.904, + 0.648 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.658, + 0.295, + 0.736 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.736, + 0.295, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.14, + 0.802, + 0.249, + 0.813 + ], + "angle": 0, + "content": "(a) LQ (No restoration)" + }, + { + "type": "image", + "bbox": [ + 0.296, + 0.658, + 0.498, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.297, + 0.735, + 0.498, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.362, + 0.802, + 0.434, + 0.813 + ], + "angle": 0, + "content": "(b) SR4IR [20]" + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.658, + 0.702, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.735, + 0.701, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.802, + 0.643, + 0.813 + ], + "angle": 0, + "content": "(c)EDTR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.658, + 0.905, + 0.735 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.703, + 0.735, + 0.904, + 0.8 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.766, + 0.802, + 0.842, + 0.813 + ], + "angle": 0, + "content": "(d) HQ (Oracle)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.815, + 0.907, + 0.871 + ], + "angle": 0, + "content": "Figure 5. Visualization of images and diverse high-level vision task results on degraded LQ (Mixture-B) images. The first, second, and third rows represent the results of image classification, semantic segmentation, and object detection, respectively. We show the restored images and the corresponding predicted labels. For visualization, we use the EDTR-1 step model for classification and the EDTR-4 step model for segmentation and detection. For more visualization results, please refer to Section S12 of our supplementary materials." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10156" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.09, + 0.905, + 0.239 + ], + "angle": 0, + "content": "
ExpSDTraining lossDiffusion processAcc↑ (%)fd↓Q-Align↑
HLFFMPre-restorationPartial diffusion# of denoising steps
(1)✓X✓XX5064.30.6413.87
(2)✓✓✓XX5062.90.7542.37
(3)✓✓✓XX163.50.5793.18
(4)✓✓✓X✓166.60.5303.38
(5)✓X✓✓✓164.50.7192.18
(6)✓✓X✓✓167.60.5073.48
(7)X✓✓✓✓164.20.5413.20
Ours✓✓✓✓✓168.80.5073.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.201, + 0.241, + 0.796, + 0.255 + ], + "angle": 0, + "content": "Table 2. Ablation study on the impact of each component on EDTR classification performance." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.267, + 0.483, + 0.404 + ], + "angle": 0, + "content": "performance improvement. In the second row of Figure 5, the LQ image is severely degraded, causing the details in the upper-left part of the bike wheel to be almost lost. While the previous TDIR method fails to restore the missing details, our method effectively restores them by leveraging a strong natural image prior, correctly recognizing that they belong to the wheel. As a result, EDTR produces a segmentation prediction that correctly maps the upper-left part of the bike wheel, demonstrating its effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.41, + 0.484, + 0.562 + ], + "angle": 0, + "content": "Object detection. Our EDTR demonstrates a surprisingly significant performance gain over the previous state-of-the-art in the Mixture-A setting, increasing mAP from \\(22.3\\%\\) to \\(33.4\\%\\) —an improvement of approximately \\(50\\%\\). As shown in the third row of Figure 5, EDTR correctly restores the shape of the sofa by leveraging the powerful diffusion prior, successfully detecting the sofa in the room, whereas the previous method fails. These results demonstrate that our method consistently achieves superior task performance and visual quality under severe degradation." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.572, + 0.253, + 0.586 + ], + "angle": 0, + "content": "4.2. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.594, + 0.454, + 0.61 + ], + "angle": 0, + "content": "All experiments use the Mixture- \\(B\\) degradation setting." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.614, + 0.357, + 0.629 + ], + "angle": 0, + "content": "4.2.1. Effect of proposed components." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.634, + 0.483, + 0.808 + ], + "angle": 0, + "content": "Table 2 shows the classification performance of our EDTR according to the use of each component. Specifically, when a proposed component is not used, i.e., \\(X\\), we follow the conventional SD-based IR method setting, as described in Section 3.1. Except for Exp-(7), we adopt the FM loss as a default, since it relates to task network training rather than the IR method. In addition, to assess how well task-relevant details are restored, we measure the distance in the feature space of the task network \\(\\mathcal{H}_{\\mathrm{HQ}}\\), which is pre-trained on HQ images. We denote this fidelity metric as feature distance, defined as \\(f_{d} = \\left\\| \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{EDTR}}) - \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{HQ}})\\right\\|_{1}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.808, + 0.484, + 0.913 + ], + "angle": 0, + "content": "Exp-(1) shows performance without any of the proposed components, i.e., conventional SD-based IR method [30]. Compared to EDTR performance (last row), although it achieves a more perceptually realistic restoration, as indicated by the higher Q-Align score, its large \\( f_{d} \\) value suggests that task-relevant details are not well restored, resulting in a \\( 4.5\\% \\) worse classification accuracy." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.265, + 0.613, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.545, + 0.341, + 0.583, + 0.353 + ], + "angle": 0, + "content": "(a) LQ" + }, + { + "type": "image", + "bbox": [ + 0.615, + 0.266, + 0.709, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.341, + 0.692, + 0.353 + ], + "angle": 0, + "content": "(b) Exp-(2)" + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.266, + 0.807, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.712, + 0.341, + 0.807, + 0.353 + ], + "angle": 0, + "content": "(c)EDTR(Ours)" + }, + { + "type": "image", + "bbox": [ + 0.81, + 0.266, + 0.905, + 0.34 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.838, + 0.341, + 0.877, + 0.353 + ], + "angle": 0, + "content": "(d) HQ" + }, + { + "type": "image_caption", + "bbox": [ + 0.547, + 0.355, + 0.872, + 0.368 + ], + "angle": 0, + "content": "Figure 6. Visualization comparison in classification." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.617 + ], + "angle": 0, + "content": "Effect of pre-restoration & partial diffusion. Exp-(2) shows performance without our design choice for the diffusion process, i.e., SD is directly applied to TDIR. Although it incorporates a loss for restoring task-relevant details, its task performance is even worse than that of Exp-(1). Figure 6 illustrates that Exp-(2) leads to excessive generation and produces non-existent content, in contrast to the results of EDTR, ultimately failing to improve task performance. This highlights the necessity of a tailored diffusion process design for TDIR. Exp-(3) shows that even with short-step denoising, EDTR without partial diffusion does not perform well for TDIR, demonstrating the importance of partial diffusion. Exp-(4) further shows that the pre-restoration process is also essential for achieving both strong task performance and good visual quality. For performance regarding the pre-restoration network, see Supplementary Section S3." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.622, + 0.909, + 0.835 + ], + "angle": 0, + "content": "Effect of HLF & FM losses. Exp-(5) and Exp-(6) show EDTR performance without HLF and FM losses, respectively. In both cases, task performance drops significantly, highlighting the importance of each training loss term. Interestingly, Exp-(5) results in a notably lower Q-Align score, indicating that the HLF loss is crucial for achieving visually pleasing results, and it enables the generation of such results even in the challenging one-step denoising scenario. This finding is consistent with previous research [47], as the HLF loss shares similarities with LPIPS [64] in measuring distance in feature space. In addition, incorporating feature spaces from two networks in the HLF loss (as in Equation (7)) yields a \\(+1.0\\%\\) gain compared to using only one (see Supplementary Section S5 for details)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Effect of SD. Exp-(7) shows the performance of EDTR without pre-trained SD weights. Specifically, we remove the frozen denoising network (purple module in Figure 2a); thus, the total number of trainable parameters remains" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10157" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.152 + ], + "angle": 0, + "content": "nearly the same. Compared to this setting, EDTR largely reduces the \\( f_{d} \\), resulting in a significant \\( 4.6\\% \\) increase in classification accuracy. These results highlight the crucial role of the diffusion prior in restoring task-relevant details." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.16, + 0.414, + 0.176 + ], + "angle": 0, + "content": "4.2.2. Effect of the number of denoising steps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.331 + ], + "angle": 0, + "content": "Table 3 shows the performance of EDTR with varying numbers of denoising steps during inference. As shown in the table, a larger number of denoising steps (e.g., 50) leads to decreased task performance. This supports our claim that repeatedly introducing details is not beneficial for TDIR and may even dilute crucial task-relevant information. In addition, we observe that short-step denoising (\\(n = 4\\)) is sufficient to produce visually pleasing results, as indicated by its favorable Q-Align score—further confirming the effectiveness of the HLF loss in guiding perceptual quality." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.339, + 0.327, + 0.357 + ], + "angle": 0, + "content": "4.3. Generalizability of EDTR" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.362, + 0.483, + 0.59 + ], + "angle": 0, + "content": "Performance of EDTR across various tasks. In Table 4, we present the performance of EDTR-1 step models, each trained separately on a specific task. Note that the diagonal (blue) values represent the performance of the original EDTR on its respective task. Since the original EDTR is independently trained for a single task, it performs well on its target task but less effectively on others. To demonstrate that our EDTR can be extended for broader applicability, we additionally train a single model on all three high-level vision tasks simultaneously, referring to this model as EDTR_All-in-One. The last row in the table presents the results. We observe that this model achieves performance comparable to, or even better than, the original EDTR across all three tasks, highlighting its potential for generalization to a variety of applications." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.593, + 0.483, + 0.744 + ], + "angle": 0, + "content": "Generalization toward unseen datasets. Table 5 shows the classification performance on unseen datasets during training. Specifically, we train the IR models using TinyImageNet\\(^{1}\\) [23] and evaluate them on CUB200 [51] and StanfordCars [22] datasets, which were not seen during training. When evaluating on these unseen datasets, the task networks trained on HQ images corresponding to each dataset are used. The results show that our EDTR significantly outperforms other methods on these unseen datasets, demonstrating the robustness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.748, + 0.483, + 0.855 + ], + "angle": 0, + "content": "Results on real-world images. Figure 7 presents real-world object detection results and restored image visualizations. We compare (a) object detection without restoration and (b) EDTR-4 step trained on Mixture-\\(B\\) degradation. Due to complex real-world degradation, detection without restoration fails to identify the person in the image. In contrast, EDTR effectively restores key high-frequency details," + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.091, + 0.903, + 0.14 + ], + "angle": 0, + "content": "
# of denoising steps4304050
Classification (Acc↑ / Q-Align↑)68.4 / 3.4967.7 / 3.4667.1 / 3.4666.8 / 3.45
Segmentation (mIoU↑ / Q-Align↑)62.9 / 3.7462.4 / 3.5462.4 / 3.4661.8 / 3.43
Detection (mAP↑ / Q-Align↑)31.9 / 4.0231.8 / 3.9731.8 / 3.9931.6 / 3.94
" + }, + { + "type": "table_caption", + "bbox": [ + 0.526, + 0.143, + 0.892, + 0.157 + ], + "angle": 0, + "content": "Table 3. Effect of the number of denoising steps in EDTR." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.163, + 0.903, + 0.234 + ], + "angle": 0, + "content": "
Eval task Used modelClassification (Acc↑)Segmentation (mIoU↑)Detection (mAP↑)
EDTR_Classification68.848.714.3
EDTR_Segmentation56.860.424.8
EDTR_Detection59.358.930.6
EDTR_All-in-One68.461.030.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.514, + 0.237, + 0.905, + 0.265 + ], + "angle": 0, + "content": "Table 4. Performance of EDTR models across various tasks. EDTR_Task represents the EDTR model optimized for each task." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.272, + 0.903, + 0.362 + ], + "angle": 0, + "content": "
Dataset IR methodsSeen \nTiny-ImageNet† [23]Unseen
CUB200 [51]StanfordCars [22]
Oracle83.782.590.4
No restoration61.210.112.1
SwinIR [27]68.021.838.7
SR4IR [20]68.943.346.1
EDTR-1 step (Ours)75.152.567.0
EDTR-4 step (Ours)76.153.066.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.365, + 0.868, + 0.377 + ], + "angle": 0, + "content": "Table 5. Classification on seen and unseen datasets." + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.384, + 0.706, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.549, + 0.52, + 0.672, + 0.531 + ], + "angle": 0, + "content": "(a) LQ (No restoration)" + }, + { + "type": "image", + "bbox": [ + 0.717, + 0.384, + 0.906, + 0.517 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.762, + 0.519, + 0.86, + 0.531 + ], + "angle": 0, + "content": "(b)EDTR(Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.533, + 0.906, + 0.575 + ], + "angle": 0, + "content": "Figure 7. Real-world object detection results and visualization of the restored image. The image is from \"56.png\" in the real-world image set RealPhoto60 [61]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.906, + 0.675 + ], + "angle": 0, + "content": "such as clothing wrinkles on a person, enabling successful detection. These results demonstrate that EDTR can handle complex real-world degradations and detect objects in degraded images by effectively leveraging the powerful diffusion prior. For more visualization results, please refer to Section S12 in our supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.683, + 0.634, + 0.699 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.901 + ], + "angle": 0, + "content": "This paper proposes EDTR, a novel task-driven image restoration (TDIR) method that directly leverages a pretrained diffusion prior to enhance high-level vision task performance. We observe that while incorporating the diffusion prior produces visually pleasing results, restoring task-relevant details remains challenging. To address this, we propose tailored designs for exploiting the diffusion prior in TDIR. Our method significantly improves the performance of various tasks in complex image degradation scenarios while achieving visually appealing results, demonstrating its broad applicability and effectiveness. We anticipate that our approach will provide new insights into leveraging diffusion priors to enhance task performance." + }, + { + "type": "page_footnote", + "bbox": [ + 0.091, + 0.863, + 0.483, + 0.901 + ], + "angle": 0, + "content": "1Since the original Tiny-ImageNet [23] has a resolution of 64x64, which is too small for IR problem, we resampled from ImageNet [42] with the same classes and number of images to create Tiny-ImageNet†." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10158" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.177 + ], + "angle": 0, + "content": "Acknowledgments. This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], and the Industrial Technology Alchemist Project [No. RS-2024-00432410] funded by MOTIE, Korea." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.205, + 0.188, + 0.22 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.23, + 0.482, + 0.271 + ], + "angle": 0, + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.273, + 0.482, + 0.314 + ], + "angle": 0, + "content": "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.317, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.373, + 0.482, + 0.427 + ], + "angle": 0, + "content": "[4] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.43, + 0.482, + 0.457 + ], + "angle": 0, + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.482, + 0.514 + ], + "angle": 0, + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.482, + 0.57 + ], + "angle": 0, + "content": "[7] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[8] Dengxin Dai, Yujuan Wang, Yuhua Chen, and Luc Van Gool. Is image super-resolution helpful for other vision tasks? In WACV, 2016. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[9] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In CVPR, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.66, + 0.482, + 0.7 + ], + "angle": 0, + "content": "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.703, + 0.482, + 0.77 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.774, + 0.482, + 0.814 + ], + "angle": 0, + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.817, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[14] Muhammad Haris, Shakhnarovich Greg, and Ukita Norimichi. Task-driven super resolution: Object detection in low-resolution images. In ICONIP, 2021. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.23, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.136, + 0.905, + 0.163 + ], + "angle": 0, + "content": "[16] Geoffrey Hinton. Distilling the knowledge in a neural network. In NIPS Workshop, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.165, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.233 + ], + "angle": 0, + "content": "[18] Shih-Chia Huang, Trung-Hieu Le, and Da-Wei Jaw. Dsnet: Joint semantic learning for object detection in inclement weather conditions. TPAMI, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.236, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[19] Jiwon Kim, Jung Kown Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.279, + 0.905, + 0.32 + ], + "angle": 0, + "content": "[20] Jaeha Kim, Junghun Oh, and Kyoung Mu Lee. Beyond image super-resolution for image recognition with task-driven perceptual loss. In CVPR, 2024. 1, 2, 4, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.905, + 0.348 + ], + "angle": 0, + "content": "[21] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.35, + 0.905, + 0.391 + ], + "angle": 0, + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, 2013. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.394, + 0.905, + 0.433 + ], + "angle": 0, + "content": "[23] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. https://api_semanticscholar.org/ CorpusID:16664790, 2015.8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.436, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.506, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[25] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.603 + ], + "angle": 0, + "content": "[26] Chengyang Li, Heng Zhou, Yang Liu, Caidong Yang, Yongqiang Xie, Zhongbo Li, and Liping Zhu. Detection-friendly dehazing: Object detection in real-world hazy scenes. TPAMI, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.606, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV, 2021. 2, 3, 5, 6, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[28] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.69, + 0.905, + 0.745 + ], + "angle": 0, + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.748, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. In ECCV, 2024. 1, 2, 3, 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.804, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[31] Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, and Thomas S Huang. When image denoising meets high-level vision tasks: A deep learning approach. In *IJCAI*, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[32] Ding Liu, Bihan Wen, Jianbo Jiao, Xianming Liu, Zhangyang Wang, and Thomas S Huang. Connecting image denoising and high-level vision tasks via deep learning. TIP, 2020. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10159" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[33] Wenyu Liu, Gaofeng Ren, Runsheng Yu, Shi Guo, Jianke Zhu, and Lei Zhang. Image-adaptive yolo for object detection in adverse weather conditions. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.176 + ], + "angle": 0, + "content": "[34] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.233 + ], + "angle": 0, + "content": "[35] Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image superresolution. In CVPR, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.482, + 0.274 + ], + "angle": 0, + "content": "[36] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[38] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[39] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In ECCV, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.419, + 0.482, + 0.46 + ], + "angle": 0, + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.482, + 0.503 + ], + "angle": 0, + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 2015. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.632, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[44] Jacob Shermeyer and Adam Van Etten. The effects of superresolution on object detection performance in satellite imagery. In CVPR Workshops, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.675, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[45] Taeyoung Son, Juwon Kang, Namyup Kim, Sunghyun Cho, and Suha Kwak. Urie: Universal image enhancement for visual recognition in the wild. In ECCV, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.717, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.746, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[47] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.802 + ], + "angle": 0, + "content": "[48] Jong-Chyi Su and Subhransu Maji. Adapting models to signal degradation using distillation. In BMVC, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.857 + ], + "angle": 0, + "content": "[49] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[50] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. In NeurIPS, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.148 + ], + "angle": 0, + "content": "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[52] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[53] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: enhanced super-resolution generative adversarial networks. In ECCV Workshops, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[54] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.303, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[55] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.426 + ], + "angle": 0, + "content": "[56] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In ICML, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.428, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[57] Jiawei Wu and Zhi Jin. Unsupervised variational translator for bridging image restoration and high-level vision tasks. In ECCV, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.904, + 0.51 + ], + "angle": 0, + "content": "[58] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[59] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.554, + 0.905, + 0.607 + ], + "angle": 0, + "content": "[60] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.609, + 0.905, + 0.676 + ], + "angle": 0, + "content": "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.678, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[65] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[66] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10160" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.091, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[67] Xiaotong Zhao, Wei Li, Yifan Zhang, and Zhiyong Feng. Residual super-resolution single shot network for low-resolution object detection. IEEE Access, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[68] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022." + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.127, + 0.179, + 0.135, + 0.189 + ], + "angle": 0, + "content": "5" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10161" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_origin.pdf b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7f8d45ccc3e82628fa312ab85973a96e6e1df705 --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/1dbcf4df-da9a-471c-b4a5-1d3b5d46842e_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:36b96eb9c7d298761e1edac354ce04f51672129c1d7e0245b10d0ac24a14f404 +size 11561267 diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/full.md b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/full.md new file mode 100644 index 0000000000000000000000000000000000000000..978402ab109c3ee52114d601ff54ec0ef7547d36 --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/full.md @@ -0,0 +1,414 @@ +# Exploiting Diffusion Prior for Task-driven Image Restoration + +Jaeha Kim1 Junghun Oh1 Kyoung Mu Lee1,2 + +$^{1}$ Dept. of ECE&ASRI, $^{2}$ IPAI, Seoul National University, Korea + +jhkim97s2@gmail.com, {dh6dh, kyoungmu}@snu.ac.kr + +# Abstract + +Task-driven image restoration (TDIR) has recently emerged to address performance drops in high-level vision tasks caused by low-quality (LQ) inputs. Previous TDIR methods struggle to handle practical scenarios in which images are degraded by multiple complex factors, leaving minimal clues for restoration. This motivates us to leverage the diffusion prior, one of the most powerful natural image priors. However, while the diffusion prior can help generate visually plausible results, using it to restore task-relevant details remains challenging, even when combined with recent TDIR methods. To address this, we propose EDTR, which effectively harnesses the power of diffusion prior to restore task-relevant details. Specifically, we propose directly leveraging useful clues from LQ images in the diffusion process by generating from pixel-error-based pre-restored LQ images with mild noise added. Moreover, we employ a small number of denoising steps to prevent the generation of redundant details that dilute crucial task-related information. We demonstrate that our method effectively utilizes diffusion prior for TDIR, significantly enhancing task performance and visual quality across diverse tasks with multiple complex degradations. + +# 1. Introduction + +Image degradation often occurs in the real world due to various factors, such as transmission loss, limited camera performance, or poor shooting conditions. In such cases, key high-frequency details can be lost, resulting in significant performance drops in high-level vision tasks, such as image classification, semantic segmentation, and object detection. + +Image restoration (IR) can be a promising solution to improve the performance of high-level vision tasks. It involves recovering missing details in low-quality (LQ) images using a learned natural image prior. However, recent studies [8, 20, 39, 57] show that the straightforward application of IR methods as a pre-processing step is not very effective in restoring essential information for high-level vision tasks. This has led to the emergence of Task-driven Image + +![](images/2ae9dcb11dad0ad03823d3fb5961357209071030c7809e3614d7cfedd296956d.jpg) +(a) Low-quality + +![](images/affdd27bc038ead7af827914a4992b257db76fb56c349a2fa73c2d1ec99aa948.jpg) +(b) DiffBIR [30] + +![](images/5dd5d79915e0cdfb5e37028eca6b4ea3092a227cb30376a693a009f95991ac46.jpg) + +![](images/ee7b2292e5a0eb27f08c8320dd77d9c34413216ce584f630516d50d59a57f497.jpg) +(d) EDTR (Ours) + +![](images/74da3437fe14e26767f8bd707d20663872a57513534260d2dd4204e304ac946d.jpg) +(c) SR4IR [20] $+\mathrm{SD}$ [41] +(e) High-quality +Figure 1. Effect of diffusion prior on image classification performance. While incorporating the StableDiffusion [41] produces visually plausible results as (b), it lacks task-relevant details (e.g., eye shape), even when combined with the previous TDIR method as (c), leading to incorrect classification. In contrast, EDTR successfully restores details as (d), resulting in correct classification. + +Restoration (TDIR), which aims to improve visual quality in ways that benefit high-level vision tasks. + +Previous TDIR methods address each type of degradation separately, e.g., super-resolution [20], denoising [31]. This restricts their practical applicability, as real-world images often exhibit multiple degradations simultaneously. To address this, we tackle a more realistic scenario by jointly handling multiple complex degradations. However, we observe that previous TDIR methods achieve limited performance gains in such scenarios, as they struggle to utilize the few available clues in degraded images to restore task-relevant details. This highlights the need for a strong natural image prior for the TDIR problem. + +In this paper, we leverage the prior knowledge from the pre-trained StableDiffusion (SD) [41], which provides one of the most powerful natural image priors. However, we observe that while the conventional SD-based IR method [30] + +can generate visually pleasing results, its straightforward adoption fails to restore task-relevant details, leading to incorrect predictions, as shown in Figure 1b. More importantly, we observe that SD still struggles to restore task-relevant details even when combined with a recent state-of-the-art TDIR method [20], as shown in Figure 1c. In other words, while diffusion prior is powerful, it must be properly adapted, as its standard form fails to improve TDIR. This issue arises because restoring task-relevant details is inherently challenging in the standard diffusion process used in conventional SD-based IR methods [30, 52, 59-61]. We identify two key reasons: (1) Although the LQ image is used as a condition for SD, generating from pure noise hinders the direct utilization of useful clues in LQ images, making it difficult to restore task-relevant details. (2) Long-step denoising, around 50, introduces redundant details that dilute crucial task-related information for TDIR. + +To address this, we propose EDTR, Exploiting Diffusion prior for Task-driven Image Restoration, the TDIR method adopting the diffusion prior in ways that directly leverage its power to restore task-relevant details in complex degradation scenarios. The primary component of EDTR is threefold: (1) We directly incorporate LQ images into the diffusion process. Since SD was not originally trained to handle complex degradations, directly feeding LQ images into the diffusion process is ineffective. Therefore, we first prerestore the LQ images using a pixel-error-based IR method, then generate from these images with mild noise added, rather than generating from pure noise, a process referred to as pre-restoration and partial diffusion. (2) We apply a short-step denoising (e.g., 1, 4) during inference. We claim that increasing the number of steps (e.g., 50) reduces task performance by introducing redundant details that dilute crucial details for TDIR. (3) We introduce tailored training loss terms to guide diffusion prior to restore task-relevant details. These loss terms are essential for achieving visually appealing results, even in a one-step denoising. + +We validate our method in three high-level vision tasks: image classification, semantic segmentation, and object detection. These tasks involve complex degradation scenarios, including a mix of downsampling, blur, noise, and JPEG artifacts. Experimental results demonstrate that EDTR effectively harnesses the power of diffusion prior, significantly improving task performance (e.g., about $+50\%$ improvement in detection) while producing visually pleasing results, as shown in Figure 1d, and generalizes well to real-world images. We summarize our contributions as: + +- We propose EDTR, the TDIR method that directly leverages a powerful pre-trained diffusion prior. +- We propose tailored designs for exploiting the diffusion prior to effectively restore task-relevant details. +- We significantly improve task performance while producing visually pleasing results under complex degradation. + +# 2. Related Work + +General image restoration. Since the pioneering SRCNN [10], numerous deep neural network-based methods [5, 9, 13, 19, 24, 28, 38, 53, 65, 66] have emerged in the field of image restoration (IR). Vision transformers [11]-based IR methods [3, 7, 55, 62], such as SwinIR [27], have also demonstrated strong performance. Recently, IR studies [30, 34, 52, 60, 61] have focused on leveraging the large pre-trained diffusion models, primarily StableDiffusion (SD) [41], which provides a powerful natural image prior essential for addressing the ill-posedness of the IR problem. StableSR [52] and DiffBIR [30] introduce trainable modules alongside a frozen SD backbone, similar to ControlNet [63]. More recent works [35, 49, 59, 61] have focused on fully harnessing SD's generative capabilities by designing carefully crafted input conditions. + +Task-driven image restoration. Taking a different perspective, recent studies [1, 2, 33, 44, 50, 57] have focused on the emerging field of Task-driven Image Restoration (TDIR), which aims to improve high-level vision task performance when handling LQ images. AOD-Net [25] suggests that using dehazing as a preprocessing step and applying object detection can enhance detection performance in foggy conditions. However, Pei et al. [39] show that applying IR methods as a simple preprocessing step does not significantly improve detection performance, with similar findings reported in the super-resolution field [8]. + +To address this, recent TDIR methods [14, 18, 26, 31, 32, 45, 67] have started training the IR network in combination with the task networks. TDSR [14] trains the IR network using both pixel-error loss and detection loss from a pre-trained object detection network, with a dedicated training schedule. RSRSSN [67] proposes an end-to-end trainable IR and detection networks with Feature Maps Multi-box loss. SR4IR [20] introduces a joint training framework for TDIR with a task-driven perceptual loss to guide the IR network in learning task-relevant features. More recently, UniRestore [4] introduces the diffusion prior into TDIR; however, its diffusion-based training is independent of taskdriven training, performed in two stages, limiting the ability to fully leverage the diffusion prior for TDIR. We instead propose integrating the diffusion prior directly within taskdriven training to better restore task-relevant details. + +# 3. Proposed Method + +Problem definition. Let $I_{\mathrm{HQ}}$ and $I_{\mathrm{LQ}}$ denote the high-quality (HQ) and low-quality (LQ) images, respectively, with paired labels $y$ from corresponding high-level vision task, e.g., class label in the image classification. The goal of Task-driven Image Restoration (TDIR) is to restore LQ images in a way that enhances the subsequent task performance while also generating visually appealing results. Formally, given a trainable image restoration (IR) model $\mathcal{R}$ + +![](images/f57cde187a8d7762035bb6f76d259f66e627592b7f8110e6e12380623877883b.jpg) +(a) The architecture of the proposed EDTR with high-level vision task network. + +![](images/64199ac137ae30ea48a96aea7337f52d532af6260f13845ebed59b76b071ab8d.jpg) +(b) The joint training procedure of EDTR and high-level vision task network. +Figure 2. The overall framework of the proposed method. $I_{\mathrm{EDTR,train}}$ is the EDTR restored image when EDTR is in training mode. + +and a trainable task network $\mathcal{H}$ , our aim is to maximize $\mathbf{M}_{\mathrm{H}}(\mathcal{H}(\mathcal{R}(I_{\mathrm{LQ}})),y)$ , where $\mathbf{M}_{\mathrm{H}}$ is the evaluation metric for the task, e.g., the accuracy for classification, while also achieving visually pleasing restored results $\mathcal{R}(I_{\mathrm{LQ}})$ . + +# 3.1. Preliminary: Conventional SD-based IR + +Unlike pure generation, IR requires generating images that are faithful to the given LQ images. To achieve this, previous StableDiffusion (SD)-based IR methods [30, 49, 52, 60, 61] commonly use LQ images as a condition for SD through ControlNet [63]. Note that, for better efficiency and more stable training, SD applies the diffusion process in the latent space rather than the image space, following the encoder and decoder structure of the VAE [21]. + +During training, given an image pair $(I_{\mathrm{HQ}}, I_{\mathrm{LQ}})$ , a VAE encoder $\mathcal{E}$ , and Gaussian noise with variance $\beta_t \in (0,1)$ , $I_{\mathrm{HQ}}$ is first diffused according to DDPM [17] as follows: + +$$ +\begin{array}{l} z _ {t} = q \left(z _ {t} \mid z _ {0} = \mathcal {E} \left(I _ {\mathrm {H Q}}\right)\right) \tag {1} \\ = \sqrt {\bar {\alpha} _ {t}} \mathcal {E} (I _ {\mathrm {H Q}}) + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \\ \end{array} +$$ + +where $q(z_{t}|z_{t - 1})\coloneqq \mathcal{N}(\sqrt{1 - \beta_{t}} z_{t - 1},\beta_{t}\mathbf{I}),\alpha_{t} = 1 - \beta_{t},$ $\bar{\alpha}_{t} = \prod_{i = 1}^{t}\alpha_{i}$ , and $\epsilon \sim \mathcal{N}(0,\mathbf{I})$ . The timestep $t$ is randomly selected from $[1,T]$ , where $T$ is the total number of timesteps. Then the trainable module, usually ControlNet, is trained to predict the injected noise as follows: + +$$ +\mathcal {L} _ {\epsilon} = \left\| \epsilon - \epsilon_ {\theta} \left(z _ {t}, t, \mathcal {E} \left(I _ {\mathrm {L Q}}\right)\right) \right\| _ {2} ^ {2}, \tag {2} +$$ + +where $\epsilon_{\theta}$ is a denoising network combined with ControlNet. Note that $I_{\mathrm{LQ}}$ is used only as condition input for ControlNet. + +During inference, denoising starts from a pure noise. It gradually moves toward less noisy images by iteratively alternating between denoising and adding noise, using sampling methods [17, 37, 46]. After approximately 50 steps, the process obtains the predicted latent $\hat{z}_0$ , which is then mapped back to the image space using the VAE decoder. + +# 3.2.EDTR Architecture + +Pre-restoration and partial diffusion. As noted earlier, conventional SD-based IR methods begin the denoising process from pure noise during the inference. Although $I_{\mathrm{LQ}}$ are used as a condition through ControlNet, its behavior heavily depends on SD's generative capacity. This poses a challenge for TDIR, as it hinders the use of valuable clues in LQ images, making it difficult to restore task-relevant details. + +To address this, we directly incorporate $I_{\mathrm{LQ}}$ into the diffusion process, i.e., Equation (1). However, this approach poses two key challenges: (1) Since SD was not originally trained on $I_{\mathrm{LQ}}$ , which include complex degradations such as JPEG artifacts, directly incorporating $I_{\mathrm{LQ}}$ into the diffusion process is ineffective at leveraging the power of SD. (2) When $I_{\mathrm{LQ}}$ is fed directly into SD, degradation artifacts can be interpreted as generation cues, producing tasks-irrelevant details. Instead, we first pre-restore $I_{\mathrm{LQ}}$ using a pixel-error-based IR method $\mathcal{R}_{\mathrm{pix}}$ , such as SwinIR [27]: + +$$ +z _ {\text {p r e - r e s}} = \mathcal {E} \left(\mathcal {R} _ {\text {p i x}} \left(I _ {\mathrm {L Q}}\right)\right). \tag {3} +$$ + +Although introducing strong noise in the diffusion process enables the generation of additional information by heavily relying on the diffusion prior, preserving valuable clues in the LQ image is more crucial for the TDIR prob + +![](images/f7b05482ebbed1a55492bb1aad1097269bbc8e765741bb2638b1c064a27b46be.jpg) +Figure 3. Illustration of pre-restoration and partial diffusion. + +lem. Thus, we propose adding mild noise to the pre-restored images and starting generation from it. We denote this technique as partial diffusion, which can be formally defined as: + +$$ +\begin{array}{l} z _ {t, \text {p a r t i a l}} = q \left(z _ {t \leq t _ {p}} \mid z _ {0} = z _ {\text {p r e - r e s}}\right) \tag {4} \\ = \sqrt {\bar {\alpha} _ {t}} z _ {\text {p r e - r e s}} + \sqrt {1 - \bar {\alpha} _ {t}} \epsilon , \\ \end{array} +$$ + +where $z_{t,\text{partial}}$ is the partial diffusion result and $t_p$ is a fixed hyper-parameter with a relatively smaller value than $T$ . We empirically set $(t_p, T) = (200, 1000)$ . Figure 3 illustrates the proposed diffusion process. + +Short-step denoising. As noted in the preliminary section, conventional SD-based IR methods employ long-step denoising (e.g., 50). However, in the context of the TDIR problem, we find that a large number of denoising steps negatively impact on task performance. Although each step introduces details through the diffusion prior, repeatedly adding such details can generate redundant content, thereby diluting crucial task-relevant information. Hence, we adopt a short-step denoising (e.g., 1, 4) during inference. Specifically, for a given number of denoising steps $n$ , we denoise $z_{t_p,\text{partial}}$ by passing it through $n$ timesteps: $\mathcal{T} = [t_p, \lfloor t_p / n \cdot (n - 1) \rfloor, \dots, \lfloor t_p / n \rfloor]$ , following [37], and obtain the final denoising result $z_{\text{diff-res}}$ . In the one-step denosing case (i.e., $n = 1$ ), this process simplifies to: + +$$ +z _ {\text {d i f f - r e s}} = \frac {z _ {t _ {p} , \text {p a r t i a l}} - \sqrt {1 - \bar {\alpha} _ {t _ {p}}} \epsilon_ {\theta} \left(z _ {t _ {p} , \text {p a r t i a l}}, t _ {p} , z _ {\text {p r e - r e s}}\right)}{\sqrt {\bar {\alpha} _ {t _ {p}}}}. \tag {5} +$$ + +Note that conventional SD-based IR methods [30, 52] struggle to generate visually pleasing results in short-step denoising scenarios [54, 58]. However, we find that optimizing loss functions, specifically designed to restore task-relevant details as discussed in Section 3.3, enables the achievement of visually pleasing results, even in one-step denoising. + +Figure 4 compares the overall IR procedure. Unlike conventional methods, which start from pure noise and require around 50 denoising steps, we use the input LQ image with pre-restoration and partial diffusion, followed by short-step denoising. This approach directly leverages useful clues in LQ images to restore task-relevant details. + +Decoding into RGB space. After the denoising process, we transform the $z_{\mathrm{diff-res}}$ back into the image space using the VAE decoder $\mathcal{D}$ . We further apply Wavelet color correction [52] to achieve a faithful reconstruction as follows: + +$$ +I _ {\mathrm {E D T R}} = \mathbf {H} \left(\mathcal {D} \left(z _ {\text {d i f f - r e s}}\right)\right) + \mathbf {L} \left(\mathcal {R} _ {\operatorname {p i x}} \left(I _ {\mathrm {L Q}}\right)\right), \tag {6} +$$ + +![](images/fe2f5071225bb7a808144a9c4292d608701251a497b41de27000a09f2fe255a2.jpg) +(b) Pre-restoration, partial diffusion, and short-step denosing. +Figure 4. The comparison of IR procedure. We omit the encoding, decoding, and conditioning for simplicity in visualization. + +where $I_{\mathrm{EDTR}}$ is the final restored result from our EDTR. The $\mathbf{H}(\cdot)$ and $\mathbf{L}(\cdot)$ correspond to the high-frequency and low-frequency components extracted by Wavelet transform. Figure 2a illustrates the overall architecture of our EDTR. + +# 3.3. Training Framework + +High-level feature loss. The goal of TDIR is to restore task-relevant details to enhance task performance. However, the noise prediction loss in Equation (2), which is conventionally used for training diffusion models, is not designed for this purpose. Hence, we introduce high-level feature (HLF) loss, which can guide the model to harness the diffusion prior for restoring task-relevant details. + +HLF loss measures the feature distance between EDTR restored results and HQ images, following the previous TDIR method [20]. Specifically, we compute the HLF loss using the feature space of two task networks: the task network $\mathcal{H}$ , which is optimized during training, and another task network $\mathcal{H}_{\mathrm{HQ}}$ , pre-trained on HQ images, to capture complementary task-relevant information (see Supplementary Section S5). Formally, the HLF loss is defined as: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {H L F}} = \frac {1}{2} \left(\left\| \mathcal {H} ^ {f} \left(I _ {\mathrm {E D T R}, \text {t r a i n}}\right) - \mathcal {H} ^ {f} \left(I _ {\mathrm {H Q}}\right) \right\| _ {1} \right. \tag {7} \\ + \left\| \mathcal {H} _ {\mathrm {H Q}} ^ {f} (I _ {\mathrm {E D T R , t r a i n}}) - \mathcal {H} _ {\mathrm {H Q}} ^ {f} (I _ {\mathrm {H Q}}) \right\| _ {1}), \\ \end{array} +$$ + +where $\mathcal{H}^f$ and $\mathcal{H}_{\mathrm{HQ}}^f$ are intermediate feature spaces of $\mathcal{H}$ and $\mathcal{H}_{\mathrm{HQ}}$ , respectively. Note that $I_{\mathrm{EDTR,train}}$ in Equation (7) represents the restored image from the one-step denoising result of $z_{t,\mathrm{partial}}$ , where $t$ is randomly sampled from $\mathcal{T}$ during training, following the conventional diffusion loss in Equation (2), regardless of denoising steps $n$ during inference (see Supplementary Algorithm S1 for details). The HLF loss updates only EDTR and does not affect $\mathcal{H}$ . + +Task loss. While training the proposed EDTR to restore task-relevant details, we also train the task networks using the restored outputs. We utilize both $I_{\mathrm{EDTR}}$ and HQ images to stabilize training and introduce additional regularization. Formally, task loss $\mathcal{L}_{\mathrm{task}}$ for the task network is defined as: + +$$ +\mathcal {L} _ {\text {t a s k}} = f _ {\text {t a s k}} \left(\mathcal {H} \left(I _ {\mathrm {E D T R} \oplus \mathrm {H Q}}\right), y\right), \tag {8} +$$ + +where $f_{\mathrm{task}}$ is the standard loss function for each task, e.g., cross-entropy loss for classification, $I_{\mathrm{EDTR}} \oplus_{\mathrm{HQ}}$ represents batch-wise concatenated images, consisting of a half-batch from $I_{\mathrm{EDTR}}$ and a half-batch from $I_{\mathrm{HQ}}$ , and $y$ are corresponding task labels. Note that $\mathcal{L}_{\mathrm{task}}$ updates only the task network $\mathcal{H}$ and does not update EDTR. + +Feature matching loss. To align the feature space of task network $\mathcal{H}$ with $\mathcal{H}_{\mathrm{HQ}}$ , which is well-trained to utilize high-frequency details present in HQ images, we further introduce feature matching (FM) loss, defined as: + +$$ +\mathcal {L} _ {\mathrm {F M}} = \left\| \mathcal {H} ^ {f} \left(I _ {\mathrm {E D T R} \oplus \mathrm {H Q}}\right) - \mathcal {H} _ {\mathrm {H Q}} ^ {f} \left(I _ {\mathrm {H Q}}\right) \right\| _ {1}. \tag {9} +$$ + +For clarity, note that $I_{\mathrm{EDTR}}$ in Equations (8) and (9) is the $n$ step denoising result, different from the one-step result $I_{\mathrm{EDTR,train}}$ in Equation (7). The $\mathcal{L}_{\mathrm{FM}}$ can be interpreted as a form of cross-quality knowledge distillation [16, 48] at the feature level. Similar to the task loss, $\mathcal{L}_{\mathrm{FM}}$ only updates $\mathcal{H}$ . + +Joint training in an alternating manner. Since SD was originally trained to operate in the image space, backpropagating the task loss, such as cross-entropy loss that is not image-based, causes instability to the EDTR. To avoid this, we train both the EDTR and the task network $\mathcal{H}$ using each corresponding training loss in an alternating manner [20], as illustrated in Figure 2b. Specifically, in each training iteration, we first update the EDTR using the HLF loss, then update $\mathcal{H}$ using a combination of the task loss and the FM loss, which can be formally written as follows: + +$$ +\min _ {\boldsymbol {\theta} _ {\mathrm {E D T R}}} \mathcal {L} _ {\mathrm {H L F}}, \text {t h e n} \min _ {\boldsymbol {\theta} _ {\mathcal {H}}} \mathcal {L} _ {\text {t a s k}} + \alpha \mathcal {L} _ {\mathrm {F M}}, \tag {10} +$$ + +where $\theta_{\mathrm{EDTR}}$ and $\theta_{\mathcal{H}}$ are the trainable parameters of the EDTR and task network $\mathcal{H}$ , and $\alpha$ is the hyper-parameter. + +# 4. Experiments + +We validate our EDTR on three representative high-level vision tasks: image classification, semantic segmentation, and object detection, under scenarios involving image degradation. Note that the final prediction $\hat{y}$ for each task is obtained as: $\hat{y} = \mathcal{H}(\mathrm{EDTR}(I_{\mathrm{LQ}}))$ , where $\mathcal{H}$ is the trained task network and $I_{\mathrm{LQ}}$ is the input low-quality (LQ) images. + +Datasets. We use three high-level vision tasks with representative datasets: CUB200 [51] for classification, and PASCAL VOC2012 [12] for segmentation and detection. We construct two degraded LQ image sets: 1) Mixture-A: downsampling combined with JPEG compression, and 2) Mixture-B: a combination of downsampling, blur, noise, and compression, following the CodeFormer [68] degradation. Details on degradation types are in the second column of Table 1 and Section S1 in the supplementary materials. + +Evaluation metrics. We use representative evaluation metrics for each task: top-1 accuracy (Acc) for classification, mean Intersection over Union (mIoU) for segmentation, and mean Average Precision (mAP) for detection. The + +mAP score is measured following the standard COCO [29] evaluation setting. We use two no-reference metrics to assess visual appeal: NIQE [36] and Q-Align [56], where the latter leveraging large vision-language models. We also report the standard PSNR for comparison. + +Network architecture. We use StableDiffusion-2.1 [41] (SD) and ControlNet [63] for EDTR. ControlNet and the VAE decoder of SD are set as trainable modules. We also incorporate the pixel-error-based SwinIR [27] model as the pre-restoration network. For the task network, we employ representative architecture for each task: ResNet [15] for classification, DeepLabV3 [6] for segmentation, and FasterRCNN [40] for detection. The task network $\mathcal{H}$ shares the same architecture as $\mathcal{H}_{\mathrm{HQ}}$ (the network pre-trained at HQ), and the weights of $\mathcal{H}$ are initialized from $\mathcal{H}_{\mathrm{HQ}}$ . + +# 4.1. Experimental Results + +Table 1 and Figure 5 present quantitative and qualitative results, respectively, across various types of degradation in image classification, semantic segmentation, and object detection. Note that in Table 1, 'Oracle', 'No restoration', and 'SwinIR' refer to the performance when the task network is trained and tested directly on HQ, LQ, and pixel-error-based SwinIR restored images, respectively. EDTR- $n$ step refers to the EDTR model with $n$ denoising steps. + +Image classification. Table 1 shows that the proposed EDTR significantly outperforms previous methods under both types of degradation, achieving classification accuracy improvements of $+2.9\%$ and $+5.4\%$ over the previous state-of-the-art task-driven image restoration (TDIR) method, SR4IR [20]. Performance gains increase as the degradation becomes more severe, highlighting our method's efficacy in practical scenarios. Furthermore, EDTR achieves the highest scores in image quality assessments such as NIQE and Q-Align, indicating its effectiveness in producing visually pleasing results. However, despite the superior task performance, EDTR does not achieve the highest PSNR score, suggesting that PSNR is a less relevant metric in TDIR. As presented in the first row of Figure 5, our EDTR, leveraging a strong natural image prior, successfully restores detailed wing patterns—an important area for classification as shown by the Grad-CAM [43] in the upper-left—and correctly predicts the label, whereas the previous method fails to restore these details and misclassifies the image. + +Semantic segmentation. The performance improvement in segmentation is more notable than in classification, with mIoU gains of $+9.1\%$ and $+11.9\%$ over the previous state-of-the-art. Although our EDTR achieves the second-best score in NIQE, it consistently achieves the highest Q-Align score, demonstrating its capability to generate perceptually realistic images. In addition, unlike classification, increasing the number of denoising steps from 1 to 4 leads to a task + +
IR MethodsDegradation typeImage classificationSemantic segmentationObject detection
Task Acc↑ (%)Visual qualityTask mIoU↑ (%)Visual qualityTask mAP↑ (%)Visual quality
NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑
OracleNone82.55.373.69+inf67.04.714.02+inf36.93.703.95+inf
No restoration60.814.081.0323.7247.413.671.1622.4314.514.541.0421.08
SwinIR [27]70.09.972.5525.2156.28.662.5423.4721.08.962.2221.70
TDSR [14]Mixture-A:66.17.532.5124.9951.37.672.3723.4116.26.432.1321.62
RSRSSN [67]SR(s=8)+69.27.322.7524.0453.06.972.5823.0622.36.592.1021.40
SR4IR [20]JPEG(q=75)71.55.413.3724.0756.24.803.1622.8522.35.383.0020.95
EDTR-1 step (Ours)74.44.683.5323.6364.15.303.5922.3331.24.623.6820.60
EDTR-4 step (Ours)74.14.263.8122.4665.35.173.9121.4833.44.194.0119.89
No restoration47.615.111.0823.0540.214.841.1922.100.015.241.0720.86
SwinIR [27]Mixture-B:60.710.262.0725.0450.49.472.3023.6722.19.601.9222.20
TDSR [14]SR(s[1,16]) +56.47.292.2024.6846.08.072.1323.5514.66.361.9421.97
RSRSSN [67]Blur(σ[0,8]) +61.47.352.3523.3548.27.722.3623.0123.16.811.8621.68
SR4IR [20]Noise(σ[0,10]) +63.46.083.1123.6251.04.723.1522.7122.64.523.1420.97
EDTR-1 step (Ours)JPEG(q[50,100])68.84.753.4823.0360.46.363.5522.0430.64.483.6420.70
EDTR-4 step (Ours)68.44.463.4923.2262.95.803.7422.2431.94.124.0220.29
+ +Table 1. Performance of high-level vision tasks and quality of restored image results under various degradation types. We indicate the best and second-best results with bold and underlined formatting, respectively. + +![](images/4b7f455994bc8490a864b5bc4db556f84a078427738621a178ea936056825340.jpg) +pred: [Blue Grosbeak] + +![](images/49ffc14a44bb4a43fc9c07016fd3798afb5ad78362f3b6ea791635c0a69b0931.jpg) +pred: [Blue Grosbeak] X + +![](images/28d8ad8fe15c2a50b2aae041616803a6a82812eb31db92e5818ae24c682928d5.jpg) +pred: [Indigo Bunting] + +![](images/25a63364286dd88b5fc466012482d3a0b9045e78d6b0dd431ce4ea7107460f5f.jpg) +pred: [Indigo Bunting] + +![](images/8cded3d279935b22e7a947771095d5638b0efb764ffde57ad242c3c669f0f830.jpg) + +![](images/76ea47ecb4662490a7e3acef484ab8410f486d67af0b5c36c6e579f50b4863c5.jpg) + +![](images/c70905a90aa2b735178675c5e59f88055f54b1594d763e78cc0cab5869950d19.jpg) + +![](images/d2b34b8fae38ef6ad016c45be54f5d9d6c2bc3e6686bb0e9092492a43d11cab9.jpg) + +![](images/ad9405d657b8fc6f5d3d89cc2bd714aa89e5813c8d0908bd085919a02adbe906.jpg) + +![](images/3d9be70e7ecdf95a7abee7dc90eb266fa778051cede40d0b743e1f4bc8ee5814.jpg) + +![](images/4ceae1e77ea961a6651d17ef64d4dc1ff1ee07741ac447369f0d3df56c5305c3.jpg) + +![](images/cb2c7f41b010ebaea61063f7c7ec2a27ad457cd7adb63a00585483fc68857a2b.jpg) + +![](images/3284048523fe6c6f2f53c7a0f036623de4fa4d6ac00955a255ab3e8de5018cc7.jpg) + +![](images/1e12c188fec29898a01cb6d29a17a6b9813ed545ceb16e2cd20edd3a1ae5a3bb.jpg) + +![](images/02e2f232754a517f8a8424d33ba2499eefe8d4a0a74df7deb52957b1a2415e36.jpg) + +![](images/8a65090f74b838e951aed696aa56b10e945c23b3bd5a31fb6434f972e5789513.jpg) + +![](images/e0ff27934d4c0e00b7c18ea3364c9e5a21a540423adcdaf22876cc4cec9cef91.jpg) + +![](images/bce4efbb93b19de2a1a1946bb2911920290d7755bf9b447532028dd85f3c353b.jpg) + +![](images/74c0a4ee6dfdcc8a12d0552a6972ac3883328170843aa818ea993d36b6d65727.jpg) + +![](images/cd397e4b0dee1dd57047c52f8f5b73c5d30810a0580c02184b2eeb438c96068a.jpg) + +![](images/6e123f70a3cf4531b42709a319ba1bfe17a9437d7e1616ccd4a82557e9a74413.jpg) + +![](images/1aba62c84f3fd0fb7c6f9c795fe7cf3a790b50ac92b0ddd29320f69c03bdc567.jpg) +(a) LQ (No restoration) + +![](images/9207c3fde75b4530a3a66da74ef0c53b27476652da8a1193295b226b25e27220.jpg) + +![](images/ee2fea1af0b2212d5cc7fd9a59ad1bcc3b36b64a716ea7d8a8aeb713e5c495f8.jpg) +(b) SR4IR [20] + +![](images/ceca33532dc97d8f5641d07ce682334443a966fd43e50f9af7f805845a8ef7ca.jpg) + +![](images/42e5ee08054fb7a6d4f4318949c69523152dec3e2a54b9a78267e98e0acfbc65.jpg) +(c)EDTR(Ours) +Figure 5. Visualization of images and diverse high-level vision task results on degraded LQ (Mixture-B) images. The first, second, and third rows represent the results of image classification, semantic segmentation, and object detection, respectively. We show the restored images and the corresponding predicted labels. For visualization, we use the EDTR-1 step model for classification and the EDTR-4 step model for segmentation and detection. For more visualization results, please refer to Section S12 of our supplementary materials. + +![](images/3357a641f72a1e5da0ab1a2095bd7cc520bf32df7247eba302f5cb84c5f0b978.jpg) + +![](images/d6e7c306f0c47135aa46e503d73079b89693563996c40c27a1958624a5a7f6f1.jpg) +(d) HQ (Oracle) + +
ExpSDTraining lossDiffusion processAcc↑ (%)fd↓Q-Align↑
HLFFMPre-restorationPartial diffusion# of denoising steps
(1)✓X✓XX5064.30.6413.87
(2)✓✓✓XX5062.90.7542.37
(3)✓✓✓XX163.50.5793.18
(4)✓✓✓X✓166.60.5303.38
(5)✓X✓✓✓164.50.7192.18
(6)✓✓X✓✓167.60.5073.48
(7)X✓✓✓✓164.20.5413.20
Ours✓✓✓✓✓168.80.5073.48
+ +Table 2. Ablation study on the impact of each component on EDTR classification performance. + +performance improvement. In the second row of Figure 5, the LQ image is severely degraded, causing the details in the upper-left part of the bike wheel to be almost lost. While the previous TDIR method fails to restore the missing details, our method effectively restores them by leveraging a strong natural image prior, correctly recognizing that they belong to the wheel. As a result, EDTR produces a segmentation prediction that correctly maps the upper-left part of the bike wheel, demonstrating its effectiveness. + +Object detection. Our EDTR demonstrates a surprisingly significant performance gain over the previous state-of-the-art in the Mixture-A setting, increasing mAP from $22.3\%$ to $33.4\%$ —an improvement of approximately $50\%$ . As shown in the third row of Figure 5, EDTR correctly restores the shape of the sofa by leveraging the powerful diffusion prior, successfully detecting the sofa in the room, whereas the previous method fails. These results demonstrate that our method consistently achieves superior task performance and visual quality under severe degradation. + +# 4.2. Ablation studies + +All experiments use the Mixture- $B$ degradation setting. + +# 4.2.1. Effect of proposed components. + +Table 2 shows the classification performance of our EDTR according to the use of each component. Specifically, when a proposed component is not used, i.e., $X$ , we follow the conventional SD-based IR method setting, as described in Section 3.1. Except for Exp-(7), we adopt the FM loss as a default, since it relates to task network training rather than the IR method. In addition, to assess how well task-relevant details are restored, we measure the distance in the feature space of the task network $\mathcal{H}_{\mathrm{HQ}}$ , which is pre-trained on HQ images. We denote this fidelity metric as feature distance, defined as $f_{d} = \left\| \mathcal{H}_{\mathrm{HQ}}^{f}(I_{\mathrm{EDTR}}) - \mathcal{H}_{\mathrm{HQ}}^{f}(I_{\mathrm{HQ}})\right\|_{1}$ . + +Exp-(1) shows performance without any of the proposed components, i.e., conventional SD-based IR method [30]. Compared to EDTR performance (last row), although it achieves a more perceptually realistic restoration, as indicated by the higher Q-Align score, its large $f_{d}$ value suggests that task-relevant details are not well restored, resulting in a $4.5\%$ worse classification accuracy. + +![](images/d33e4b6da70d19f40fe84df9876c9d0421ecdac5c3641095990d2893fec3646d.jpg) +(a) LQ + +![](images/ae631f535949326764b912ca8f520abbf462487acb770ba5aeb90df2191fa7e8.jpg) +(b) Exp-(2) + +![](images/47d220f2689a8a3ff4604beae42dcac058213c83e7c718cf9cff4c97d34b86f3.jpg) +(c)EDTR(Ours) +Figure 6. Visualization comparison in classification. + +![](images/e212bf1ba0e3581f1e588acaca487471e3e85f240287a0929948a4d35caf4ad1.jpg) +(d) HQ + +Effect of pre-restoration & partial diffusion. Exp-(2) shows performance without our design choice for the diffusion process, i.e., SD is directly applied to TDIR. Although it incorporates a loss for restoring task-relevant details, its task performance is even worse than that of Exp-(1). Figure 6 illustrates that Exp-(2) leads to excessive generation and produces non-existent content, in contrast to the results of EDTR, ultimately failing to improve task performance. This highlights the necessity of a tailored diffusion process design for TDIR. Exp-(3) shows that even with short-step denoising, EDTR without partial diffusion does not perform well for TDIR, demonstrating the importance of partial diffusion. Exp-(4) further shows that the pre-restoration process is also essential for achieving both strong task performance and good visual quality. For performance regarding the pre-restoration network, see Supplementary Section S3. + +Effect of HLF & FM losses. Exp-(5) and Exp-(6) show EDTR performance without HLF and FM losses, respectively. In both cases, task performance drops significantly, highlighting the importance of each training loss term. Interestingly, Exp-(5) results in a notably lower Q-Align score, indicating that the HLF loss is crucial for achieving visually pleasing results, and it enables the generation of such results even in the challenging one-step denoising scenario. This finding is consistent with previous research [47], as the HLF loss shares similarities with LPIPS [64] in measuring distance in feature space. In addition, incorporating feature spaces from two networks in the HLF loss (as in Equation (7)) yields a $+1.0\%$ gain compared to using only one (see Supplementary Section S5 for details). + +Effect of SD. Exp-(7) shows the performance of EDTR without pre-trained SD weights. Specifically, we remove the frozen denoising network (purple module in Figure 2a); thus, the total number of trainable parameters remains + +nearly the same. Compared to this setting, EDTR largely reduces the $f_{d}$ , resulting in a significant $4.6\%$ increase in classification accuracy. These results highlight the crucial role of the diffusion prior in restoring task-relevant details. + +# 4.2.2. Effect of the number of denoising steps. + +Table 3 shows the performance of EDTR with varying numbers of denoising steps during inference. As shown in the table, a larger number of denoising steps (e.g., 50) leads to decreased task performance. This supports our claim that repeatedly introducing details is not beneficial for TDIR and may even dilute crucial task-relevant information. In addition, we observe that short-step denoising ( $n = 4$ ) is sufficient to produce visually pleasing results, as indicated by its favorable Q-Align score—further confirming the effectiveness of the HLF loss in guiding perceptual quality. + +# 4.3. Generalizability of EDTR + +Performance of EDTR across various tasks. In Table 4, we present the performance of EDTR-1 step models, each trained separately on a specific task. Note that the diagonal (blue) values represent the performance of the original EDTR on its respective task. Since the original EDTR is independently trained for a single task, it performs well on its target task but less effectively on others. To demonstrate that our EDTR can be extended for broader applicability, we additionally train a single model on all three high-level vision tasks simultaneously, referring to this model as EDTR_All-in-One. The last row in the table presents the results. We observe that this model achieves performance comparable to, or even better than, the original EDTR across all three tasks, highlighting its potential for generalization to a variety of applications. + +Generalization toward unseen datasets. Table 5 shows the classification performance on unseen datasets during training. Specifically, we train the IR models using TinyImageNet $^{1}$ [23] and evaluate them on CUB200 [51] and StanfordCars [22] datasets, which were not seen during training. When evaluating on these unseen datasets, the task networks trained on HQ images corresponding to each dataset are used. The results show that our EDTR significantly outperforms other methods on these unseen datasets, demonstrating the robustness of our approach. + +Results on real-world images. Figure 7 presents real-world object detection results and restored image visualizations. We compare (a) object detection without restoration and (b) EDTR-4 step trained on Mixture- $B$ degradation. Due to complex real-world degradation, detection without restoration fails to identify the person in the image. In contrast, EDTR effectively restores key high-frequency details, + +
# of denoising steps4304050
Classification (Acc↑ / Q-Align↑)68.4 / 3.4967.7 / 3.4667.1 / 3.4666.8 / 3.45
Segmentation (mIoU↑ / Q-Align↑)62.9 / 3.7462.4 / 3.5462.4 / 3.4661.8 / 3.43
Detection (mAP↑ / Q-Align↑)31.9 / 4.0231.8 / 3.9731.8 / 3.9931.6 / 3.94
+ +Table 3. Effect of the number of denoising steps in EDTR. + +
Eval task Used modelClassification (Acc↑)Segmentation (mIoU↑)Detection (mAP↑)
EDTR_Classification68.848.714.3
EDTR_Segmentation56.860.424.8
EDTR_Detection59.358.930.6
EDTR_All-in-One68.461.030.4
+ +Table 4. Performance of EDTR models across various tasks. EDTR_Task represents the EDTR model optimized for each task. + +
Dataset IR methodsSeen +Tiny-ImageNet† [23]Unseen
CUB200 [51]StanfordCars [22]
Oracle83.782.590.4
No restoration61.210.112.1
SwinIR [27]68.021.838.7
SR4IR [20]68.943.346.1
EDTR-1 step (Ours)75.152.567.0
EDTR-4 step (Ours)76.153.066.4
+ +Table 5. Classification on seen and unseen datasets. + +![](images/99652a1c9d21ebdc2f0c179346df90c069af9df990824ac5c6b49e2cd15f75f4.jpg) +(a) LQ (No restoration) + +![](images/badd37325bea15074e57bfef96b598f343cf59a6673bbcad4f838ff1f9d58085.jpg) +(b)EDTR(Ours) +Figure 7. Real-world object detection results and visualization of the restored image. The image is from "56.png" in the real-world image set RealPhoto60 [61]. + +such as clothing wrinkles on a person, enabling successful detection. These results demonstrate that EDTR can handle complex real-world degradations and detect objects in degraded images by effectively leveraging the powerful diffusion prior. For more visualization results, please refer to Section S12 in our supplementary materials. + +# 5. Conclusion + +This paper proposes EDTR, a novel task-driven image restoration (TDIR) method that directly leverages a pretrained diffusion prior to enhance high-level vision task performance. We observe that while incorporating the diffusion prior produces visually pleasing results, restoring task-relevant details remains challenging. To address this, we propose tailored designs for exploiting the diffusion prior in TDIR. Our method significantly improves the performance of various tasks in complex image degradation scenarios while achieving visually appealing results, demonstrating its broad applicability and effectiveness. We anticipate that our approach will provide new insights into leveraging diffusion priors to enhance task performance. + +Acknowledgments. This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], and the Industrial Technology Alchemist Project [No. RS-2024-00432410] funded by MOTIE, Korea. + +# References + +[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In CVPR, 2018. 2 +[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, 2018. 2 +[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2 +[4] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 2 +[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2 +[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2016. 5 +[7] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. In CVPR, 2023. 2 +[8] Dengxin Dai, Yujuan Wang, Yuhua Chen, and Luc Van Gool. Is image super-resolution helpful for other vision tasks? In WACV, 2016. 1, 2 +[9] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In CVPR, 2019. 2 +[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 2 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2 +[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 5 +[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018. 2 +[14] Muhammad Haris, Shakhnarovich Greg, and Ukita Norimichi. Task-driven super resolution: Object detection in low-resolution images. In ICONIP, 2021. 2, 6 + +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5 +[16] Geoffrey Hinton. Distilling the knowledge in a neural network. In NIPS Workshop, 2014. 5 +[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[18] Shih-Chia Huang, Trung-Hieu Le, and Da-Wei Jaw. Dsnet: Joint semantic learning for object detection in inclement weather conditions. TPAMI, 2020. 2 +[19] Jiwon Kim, Jung Kown Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2 +[20] Jaeha Kim, Junghun Oh, and Kyoung Mu Lee. Beyond image super-resolution for image recognition with task-driven perceptual loss. In CVPR, 2024. 1, 2, 4, 5, 6, 8 +[21] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 3 +[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, 2013. 8 +[23] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. https://api_semanticscholar.org/ CorpusID:16664790, 2015.8 +[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2 +[25] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In ICCV, 2017. 2 +[26] Chengyang Li, Heng Zhou, Yang Liu, Caidong Yang, Yongqiang Xie, Zhongbo Li, and Liping Zhu. Detection-friendly dehazing: Object detection in real-world hazy scenes. TPAMI, 2023. 2 +[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV, 2021. 2, 3, 5, 6, 8 +[28] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 2 +[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 5 +[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. In ECCV, 2024. 1, 2, 3, 4, 7 +[31] Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, and Thomas S Huang. When image denoising meets high-level vision tasks: A deep learning approach. In *IJCAI*, 2018. 1, 2 +[32] Ding Liu, Bihan Wen, Jianbo Jiao, Xianming Liu, Zhangyang Wang, and Thomas S Huang. Connecting image denoising and high-level vision tasks via deep learning. TIP, 2020. 2 + +[33] Wenyu Liu, Gaofeng Ren, Runsheng Yu, Shi Guo, Jianke Zhu, and Lei Zhang. Image-adaptive yolo for object detection in adverse weather conditions. In AAAI, 2022. 2 +[34] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In CVPR, 2024. 2 +[35] Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image superresolution. In CVPR, 2025. 2 +[36] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a "completely blind" image quality analyzer. SPL, 2012. 5 +[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4 +[38] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In ECCV, 2020. 2 +[39] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In ECCV, 2018. 1, 2 +[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, 2015. 5 +[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5 +[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 2015. 8 +[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5 +[44] Jacob Shermeyer and Adam Van Etten. The effects of superresolution on object detection performance in satellite imagery. In CVPR Workshops, 2019. 2 +[45] Taeyoung Son, Juwon Kang, Namyup Kim, Sunghyun Cho, and Suha Kwak. Urie: Universal image enhancement for visual recognition in the wild. In ECCV, 2020. 2 +[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[47] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 7 +[48] Jong-Chyi Su and Subhransu Maji. Adapting models to signal degradation using distillation. In BMVC, 2017. 5 +[49] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2, 3 +[50] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. In NeurIPS, 2022. 2 + +[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 8 +[52] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 2, 3, 4 +[53] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: enhanced super-resolution generative adversarial networks. In ECCV Workshops, 2018. 2 +[54] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 4 +[55] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 2 +[56] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In ICML, 2024. 5 +[57] Jiawei Wu and Zhi Jin. Unsupervised variational translator for bridging image restoration and high-level vision tasks. In ECCV, 2024. 1, 2 +[58] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 4 +[59] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2 +[60] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024. 2, 3 +[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2, 3, 8 +[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2 +[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5 +[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7 +[65] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 2 +[66] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 2 + +[67] Xiaotong Zhao, Wei Li, Yifan Zhang, and Zhiyong Feng. Residual super-resolution single shot network for low-resolution object detection. IEEE Access, 2018. 2, 6 +[68] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022. \ No newline at end of file diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/images.zip b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3d68c43788c2ef62b01f0288946291b670ad6c42 --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b6ff366456074b660a1c0374698d74138e8d149e794fe27ceb24684ec68cee25 +size 801482 diff --git a/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/layout.json b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6f9dd2f7af26807cbda3e7c4a0db0b023720899f --- /dev/null +++ b/2025/Exploiting Diffusion Prior for Task-driven Image Restoration/layout.json @@ -0,0 +1,11559 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "spans": [ + { + "bbox": [ + 116, + 103, + 496, + 121 + ], + "type": "text", + "content": "Exploiting Diffusion Prior for Task-driven Image Restoration" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 186, + 141, + 424, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 186, + 141, + 424, + 157 + ], + "spans": [ + { + "bbox": [ + 186, + 141, + 424, + 157 + ], + "type": "text", + "content": "Jaeha Kim1 Junghun Oh1 Kyoung Mu Lee1,2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "spans": [ + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "type": "text", + "content": "Dept. of ECE&ASRI, " + }, + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 152, + 157, + 459, + 172 + ], + "type": "text", + "content": "IPAI, Seoul National University, Korea" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 173, + 173, + 434, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 173, + 173, + 434, + 185 + ], + "spans": [ + { + "bbox": [ + 173, + 173, + 434, + 185 + ], + "type": "text", + "content": "jhkim97s2@gmail.com, {dh6dh, kyoungmu}@snu.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "spans": [ + { + "bbox": [ + 152, + 213, + 200, + 225 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 239, + 297, + 502 + ], + "type": "text", + "content": "Task-driven image restoration (TDIR) has recently emerged to address performance drops in high-level vision tasks caused by low-quality (LQ) inputs. Previous TDIR methods struggle to handle practical scenarios in which images are degraded by multiple complex factors, leaving minimal clues for restoration. This motivates us to leverage the diffusion prior, one of the most powerful natural image priors. However, while the diffusion prior can help generate visually plausible results, using it to restore task-relevant details remains challenging, even when combined with recent TDIR methods. To address this, we propose EDTR, which effectively harnesses the power of diffusion prior to restore task-relevant details. Specifically, we propose directly leveraging useful clues from LQ images in the diffusion process by generating from pixel-error-based pre-restored LQ images with mild noise added. Moreover, we employ a small number of denoising steps to prevent the generation of redundant details that dilute crucial task-related information. We demonstrate that our method effectively utilizes diffusion prior for TDIR, significantly enhancing task performance and visual quality across diverse tasks with multiple complex degradations." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 135, + 538 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 616 + ], + "type": "text", + "content": "Image degradation often occurs in the real world due to various factors, such as transmission loss, limited camera performance, or poor shooting conditions. In such cases, key high-frequency details can be lost, resulting in significant performance drops in high-level vision tasks, such as image classification, semantic segmentation, and object detection." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Image restoration (IR) can be a promising solution to improve the performance of high-level vision tasks. It involves recovering missing details in low-quality (LQ) images using a learned natural image prior. However, recent studies [8, 20, 39, 57] show that the straightforward application of IR methods as a pre-processing step is not very effective in restoring essential information for high-level vision tasks. This has led to the emergence of Task-driven Image" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 315, + 265, + 377, + 356 + ], + "blocks": [ + { + "bbox": [ + 315, + 265, + 377, + 356 + ], + "lines": [ + { + "bbox": [ + 315, + 265, + 377, + 356 + ], + "spans": [ + { + "bbox": [ + 315, + 265, + 377, + 356 + ], + "type": "image", + "image_path": "2ae9dcb11dad0ad03823d3fb5961357209071030c7809e3614d7cfedd296956d.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 320, + 357, + 373, + 368 + ], + "lines": [ + { + "bbox": [ + 320, + 357, + 373, + 368 + ], + "spans": [ + { + "bbox": [ + 320, + 357, + 373, + 368 + ], + "type": "text", + "content": "(a) Low-quality" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 383, + 212, + 465, + 304 + ], + "blocks": [ + { + "bbox": [ + 383, + 212, + 465, + 304 + ], + "lines": [ + { + "bbox": [ + 383, + 212, + 465, + 304 + ], + "spans": [ + { + "bbox": [ + 383, + 212, + 465, + 304 + ], + "type": "image", + "image_path": "affdd27bc038ead7af827914a4992b257db76fb56c349a2fa73c2d1ec99aa948.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 397, + 305, + 452, + 315 + ], + "lines": [ + { + "bbox": [ + 397, + 305, + 452, + 315 + ], + "spans": [ + { + "bbox": [ + 397, + 305, + 452, + 315 + ], + "type": "text", + "content": "(b) DiffBIR [30]" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 470, + 212, + 551, + 304 + ], + "blocks": [ + { + "bbox": [ + 470, + 212, + 551, + 304 + ], + "lines": [ + { + "bbox": [ + 470, + 212, + 551, + 304 + ], + "spans": [ + { + "bbox": [ + 470, + 212, + 551, + 304 + ], + "type": "image", + "image_path": "5dd5d79915e0cdfb5e37028eca6b4ea3092a227cb30376a693a009f95991ac46.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 383, + 316, + 465, + 408 + ], + "blocks": [ + { + "bbox": [ + 383, + 316, + 465, + 408 + ], + "lines": [ + { + "bbox": [ + 383, + 316, + 465, + 408 + ], + "spans": [ + { + "bbox": [ + 383, + 316, + 465, + 408 + ], + "type": "image", + "image_path": "ee7b2292e5a0eb27f08c8320dd77d9c34413216ce584f630516d50d59a57f497.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 411, + 454, + 420 + ], + "lines": [ + { + "bbox": [ + 394, + 411, + 454, + 420 + ], + "spans": [ + { + "bbox": [ + 394, + 411, + 454, + 420 + ], + "type": "text", + "content": "(d) EDTR (Ours)" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 470, + 316, + 551, + 408 + ], + "blocks": [ + { + "bbox": [ + 470, + 305, + 550, + 316 + ], + "lines": [ + { + "bbox": [ + 470, + 305, + 550, + 316 + ], + "spans": [ + { + "bbox": [ + 470, + 305, + 550, + 316 + ], + "type": "text", + "content": "(c) SR4IR [20] " + }, + { + "bbox": [ + 470, + 305, + 550, + 316 + ], + "type": "inline_equation", + "content": "+\\mathrm{SD}" + }, + { + "bbox": [ + 470, + 305, + 550, + 316 + ], + "type": "text", + "content": " [41]" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 470, + 316, + 551, + 408 + ], + "lines": [ + { + "bbox": [ + 470, + 316, + 551, + 408 + ], + "spans": [ + { + "bbox": [ + 470, + 316, + 551, + 408 + ], + "type": "image", + "image_path": "74da3437fe14e26767f8bd707d20663872a57513534260d2dd4204e304ac946d.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 484, + 411, + 537, + 421 + ], + "lines": [ + { + "bbox": [ + 484, + 411, + 537, + 421 + ], + "spans": [ + { + "bbox": [ + 484, + 411, + 537, + 421 + ], + "type": "text", + "content": "(e) High-quality" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 427, + 555, + 494 + ], + "lines": [ + { + "bbox": [ + 313, + 427, + 555, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 427, + 555, + 494 + ], + "type": "text", + "content": "Figure 1. Effect of diffusion prior on image classification performance. While incorporating the StableDiffusion [41] produces visually plausible results as (b), it lacks task-relevant details (e.g., eye shape), even when combined with the previous TDIR method as (c), leading to incorrect classification. In contrast, EDTR successfully restores details as (d), resulting in correct classification." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 508, + 553, + 532 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 508, + 553, + 532 + ], + "spans": [ + { + "bbox": [ + 313, + 508, + 553, + 532 + ], + "type": "text", + "content": "Restoration (TDIR), which aims to improve visual quality in ways that benefit high-level vision tasks." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 533, + 555, + 664 + ], + "type": "text", + "content": "Previous TDIR methods address each type of degradation separately, e.g., super-resolution [20], denoising [31]. This restricts their practical applicability, as real-world images often exhibit multiple degradations simultaneously. To address this, we tackle a more realistic scenario by jointly handling multiple complex degradations. However, we observe that previous TDIR methods achieve limited performance gains in such scenarios, as they struggle to utilize the few available clues in degraded images to restore task-relevant details. This highlights the need for a strong natural image prior for the TDIR problem." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "In this paper, we leverage the prior knowledge from the pre-trained StableDiffusion (SD) [41], which provides one of the most powerful natural image priors. However, we observe that while the conventional SD-based IR method [30]" + } + ] + } + ], + "index": 24 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10151" + } + ] + } + ], + "index": 25 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 274 + ], + "type": "text", + "content": "can generate visually pleasing results, its straightforward adoption fails to restore task-relevant details, leading to incorrect predictions, as shown in Figure 1b. More importantly, we observe that SD still struggles to restore task-relevant details even when combined with a recent state-of-the-art TDIR method [20], as shown in Figure 1c. In other words, while diffusion prior is powerful, it must be properly adapted, as its standard form fails to improve TDIR. This issue arises because restoring task-relevant details is inherently challenging in the standard diffusion process used in conventional SD-based IR methods [30, 52, 59-61]. We identify two key reasons: (1) Although the LQ image is used as a condition for SD, generating from pure noise hinders the direct utilization of useful clues in LQ images, making it difficult to restore task-relevant details. (2) Long-step denoising, around 50, introduces redundant details that dilute crucial task-related information for TDIR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 276, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 295, + 516 + ], + "type": "text", + "content": "To address this, we propose EDTR, Exploiting Diffusion prior for Task-driven Image Restoration, the TDIR method adopting the diffusion prior in ways that directly leverage its power to restore task-relevant details in complex degradation scenarios. The primary component of EDTR is threefold: (1) We directly incorporate LQ images into the diffusion process. Since SD was not originally trained to handle complex degradations, directly feeding LQ images into the diffusion process is ineffective. Therefore, we first prerestore the LQ images using a pixel-error-based IR method, then generate from these images with mild noise added, rather than generating from pure noise, a process referred to as pre-restoration and partial diffusion. (2) We apply a short-step denoising (e.g., 1, 4) during inference. We claim that increasing the number of steps (e.g., 50) reduces task performance by introducing redundant details that dilute crucial details for TDIR. (3) We introduce tailored training loss terms to guide diffusion prior to restore task-relevant details. These loss terms are essential for achieving visually appealing results, even in a one-step denoising." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 516, + 294, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 516, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 55, + 516, + 294, + 635 + ], + "type": "text", + "content": "We validate our method in three high-level vision tasks: image classification, semantic segmentation, and object detection. These tasks involve complex degradation scenarios, including a mix of downsampling, blur, noise, and JPEG artifacts. Experimental results demonstrate that EDTR effectively harnesses the power of diffusion prior, significantly improving task performance (e.g., about " + }, + { + "bbox": [ + 55, + 516, + 294, + 635 + ], + "type": "inline_equation", + "content": "+50\\%" + }, + { + "bbox": [ + 55, + 516, + 294, + 635 + ], + "type": "text", + "content": " improvement in detection) while producing visually pleasing results, as shown in Figure 1d, and generalizes well to real-world images. We summarize our contributions as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 637, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 6, + "blocks": [ + { + "bbox": [ + 55, + 637, + 295, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 637, + 295, + 661 + ], + "spans": [ + { + "bbox": [ + 55, + 637, + 295, + 661 + ], + "type": "text", + "content": "- We propose EDTR, the TDIR method that directly leverages a powerful pre-trained diffusion prior." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 663, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 663, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 663, + 295, + 687 + ], + "type": "text", + "content": "- We propose tailored designs for exploiting the diffusion prior to effectively restore task-relevant details." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 690, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 690, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 690, + 294, + 714 + ], + "type": "text", + "content": "- We significantly improve task performance while producing visually pleasing results under complex degradation." + } + ] + } + ], + "index": 5 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 400, + 83 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 87, + 553, + 266 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 87, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 87, + 553, + 266 + ], + "type": "text", + "content": "General image restoration. Since the pioneering SRCNN [10], numerous deep neural network-based methods [5, 9, 13, 19, 24, 28, 38, 53, 65, 66] have emerged in the field of image restoration (IR). Vision transformers [11]-based IR methods [3, 7, 55, 62], such as SwinIR [27], have also demonstrated strong performance. Recently, IR studies [30, 34, 52, 60, 61] have focused on leveraging the large pre-trained diffusion models, primarily StableDiffusion (SD) [41], which provides a powerful natural image prior essential for addressing the ill-posedness of the IR problem. StableSR [52] and DiffBIR [30] introduce trainable modules alongside a frozen SD backbone, similar to ControlNet [63]. More recent works [35, 49, 59, 61] have focused on fully harnessing SD's generative capabilities by designing carefully crafted input conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 553, + 401 + ], + "type": "text", + "content": "Task-driven image restoration. Taking a different perspective, recent studies [1, 2, 33, 44, 50, 57] have focused on the emerging field of Task-driven Image Restoration (TDIR), which aims to improve high-level vision task performance when handling LQ images. AOD-Net [25] suggests that using dehazing as a preprocessing step and applying object detection can enhance detection performance in foggy conditions. However, Pei et al. [39] show that applying IR methods as a simple preprocessing step does not significantly improve detection performance, with similar findings reported in the super-resolution field [8]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 553, + 594 + ], + "type": "text", + "content": "To address this, recent TDIR methods [14, 18, 26, 31, 32, 45, 67] have started training the IR network in combination with the task networks. TDSR [14] trains the IR network using both pixel-error loss and detection loss from a pre-trained object detection network, with a dedicated training schedule. RSRSSN [67] proposes an end-to-end trainable IR and detection networks with Feature Maps Multi-box loss. SR4IR [20] introduces a joint training framework for TDIR with a task-driven perceptual loss to guide the IR network in learning task-relevant features. More recently, UniRestore [4] introduces the diffusion prior into TDIR; however, its diffusion-based training is independent of taskdriven training, performed in two stages, limiting the ability to fully leverage the diffusion prior for TDIR. We instead propose integrating the diffusion prior directly within taskdriven training to better restore task-relevant details." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 602, + 420, + 615 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 602, + 420, + 615 + ], + "spans": [ + { + "bbox": [ + 313, + 602, + 420, + 615 + ], + "type": "text", + "content": "3. Proposed Method" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": "Problem definition. Let " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " denote the high-quality (HQ) and low-quality (LQ) images, respectively, with paired labels " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "text", + "content": " from corresponding high-level vision task, e.g., class label in the image classification. The goal of Task-driven Image Restoration (TDIR) is to restore LQ images in a way that enhances the subsequent task performance while also generating visually appealing results. Formally, given a trainable image restoration (IR) model " + }, + { + "bbox": [ + 313, + 617, + 553, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{R}" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10152" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 62, + 70, + 549, + 175 + ], + "blocks": [ + { + "bbox": [ + 62, + 70, + 549, + 175 + ], + "lines": [ + { + "bbox": [ + 62, + 70, + 549, + 175 + ], + "spans": [ + { + "bbox": [ + 62, + 70, + 549, + 175 + ], + "type": "image", + "image_path": "f57cde187a8d7762035bb6f76d259f66e627592b7f8110e6e12380623877883b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 176, + 178, + 432, + 188 + ], + "lines": [ + { + "bbox": [ + 176, + 178, + 432, + 188 + ], + "spans": [ + { + "bbox": [ + 176, + 178, + 432, + 188 + ], + "type": "text", + "content": "(a) The architecture of the proposed EDTR with high-level vision task network." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 77, + 190, + 532, + 312 + ], + "blocks": [ + { + "bbox": [ + 77, + 190, + 532, + 312 + ], + "lines": [ + { + "bbox": [ + 77, + 190, + 532, + 312 + ], + "spans": [ + { + "bbox": [ + 77, + 190, + 532, + 312 + ], + "type": "image", + "image_path": "64199ac137ae30ea48a96aea7337f52d532af6260f13845ebed59b76b071ab8d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 315, + 429, + 325 + ], + "lines": [ + { + "bbox": [ + 180, + 315, + 429, + 325 + ], + "spans": [ + { + "bbox": [ + 180, + 315, + 429, + 325 + ], + "type": "text", + "content": "(b) The joint training procedure of EDTR and high-level vision task network." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 64, + 326, + 544, + 338 + ], + "lines": [ + { + "bbox": [ + 64, + 326, + 544, + 338 + ], + "spans": [ + { + "bbox": [ + 64, + 326, + 544, + 338 + ], + "type": "text", + "content": "Figure 2. The overall framework of the proposed method. " + }, + { + "bbox": [ + 64, + 326, + 544, + 338 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR,train}}" + }, + { + "bbox": [ + 64, + 326, + 544, + 338 + ], + "type": "text", + "content": " is the EDTR restored image when EDTR is in training mode." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "content": "and a trainable task network " + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "content": ", our aim is to maximize " + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{H}}(\\mathcal{H}(\\mathcal{R}(I_{\\mathrm{LQ}})),y)" + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mathbf{M}_{\\mathrm{H}}" + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "content": " is the evaluation metric for the task, e.g., the accuracy for classification, while also achieving visually pleasing restored results " + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "inline_equation", + "content": "\\mathcal{R}(I_{\\mathrm{LQ}})" + }, + { + "bbox": [ + 55, + 344, + 296, + 393 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 399, + 266, + 412 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 266, + 412 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 266, + 412 + ], + "type": "text", + "content": "3.1. Preliminary: Conventional SD-based IR" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 417, + 296, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 514 + ], + "type": "text", + "content": "Unlike pure generation, IR requires generating images that are faithful to the given LQ images. To achieve this, previous StableDiffusion (SD)-based IR methods [30, 49, 52, 60, 61] commonly use LQ images as a condition for SD through ControlNet [63]. Note that, for better efficiency and more stable training, SD applies the diffusion process in the latent space rather than the image space, following the encoder and decoder structure of the VAE [21]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": "During training, given an image pair " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "(I_{\\mathrm{HQ}}, I_{\\mathrm{LQ}})" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": ", a VAE encoder " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": ", and Gaussian noise with variance " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "\\beta_t \\in (0,1)" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 55, + 514, + 296, + 551 + ], + "type": "text", + "content": " is first diffused according to DDPM [17] as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 111, + 555, + 294, + 586 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 111, + 555, + 294, + 586 + ], + "spans": [ + { + "bbox": [ + 111, + 555, + 294, + 586 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t} = q \\left(z _ {t} \\mid z _ {0} = \\mathcal {E} \\left(I _ {\\mathrm {H Q}}\\right)\\right) \\tag {1} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} \\mathcal {E} (I _ {\\mathrm {H Q}}) + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}", + "image_path": "edf5dc121aa4c0eb9ab67b3cf88f4f3d82494955fb10b2648213c1e9664217f7.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "q(z_{t}|z_{t - 1})\\coloneqq \\mathcal{N}(\\sqrt{1 - \\beta_{t}} z_{t - 1},\\beta_{t}\\mathbf{I}),\\alpha_{t} = 1 - \\beta_{t}," + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{t} = \\prod_{i = 1}^{t}\\alpha_{i}" + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "\\epsilon \\sim \\mathcal{N}(0,\\mathbf{I})" + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": " . The timestep " + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": " is randomly selected from " + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "[1,T]" + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": " , where " + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 596, + 296, + 657 + ], + "type": "text", + "content": " is the total number of timesteps. Then the trainable module, usually ControlNet, is trained to predict the injected noise as follows:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 113, + 663, + 295, + 677 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 663, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 113, + 663, + 295, + 677 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\epsilon} = \\left\\| \\epsilon - \\epsilon_ {\\theta} \\left(z _ {t}, t, \\mathcal {E} \\left(I _ {\\mathrm {L Q}}\\right)\\right) \\right\\| _ {2} ^ {2}, \\tag {2}", + "image_path": "0334ecf7fef12a3a4a78e4e48bdeb8780d365d8a75adabbf7c3e8a45390e1f15.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\epsilon_{\\theta}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is a denoising network combined with ControlNet. Note that " + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": " is used only as condition input for ControlNet." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 344, + 554, + 417 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 344, + 554, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 344, + 554, + 417 + ], + "type": "text", + "content": "During inference, denoising starts from a pure noise. It gradually moves toward less noisy images by iteratively alternating between denoising and adding noise, using sampling methods [17, 37, 46]. After approximately 50 steps, the process obtains the predicted latent " + }, + { + "bbox": [ + 313, + 344, + 554, + 417 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 313, + 344, + 554, + 417 + ], + "type": "text", + "content": ", which is then mapped back to the image space using the VAE decoder." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 421, + 430, + 432 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 421, + 430, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 421, + 430, + 432 + ], + "type": "text", + "content": "3.2.EDTR Architecture" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 436, + 554, + 519 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 436, + 554, + 519 + ], + "spans": [ + { + "bbox": [ + 313, + 436, + 554, + 519 + ], + "type": "text", + "content": "Pre-restoration and partial diffusion. As noted earlier, conventional SD-based IR methods begin the denoising process from pure noise during the inference. Although " + }, + { + "bbox": [ + 313, + 436, + 554, + 519 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 436, + 554, + 519 + ], + "type": "text", + "content": " are used as a condition through ControlNet, its behavior heavily depends on SD's generative capacity. This poses a challenge for TDIR, as it hinders the use of valuable clues in LQ images, making it difficult to restore task-relevant details." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": "To address this, we directly incorporate " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": " into the diffusion process, i.e., Equation (1). However, this approach poses two key challenges: (1) Since SD was not originally trained on " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": ", which include complex degradations such as JPEG artifacts, directly incorporating " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": " into the diffusion process is ineffective at leveraging the power of SD. (2) When " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": " is fed directly into SD, degradation artifacts can be interpreted as generation cues, producing tasks-irrelevant details. Instead, we first pre-restore " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": " using a pixel-error-based IR method " + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{R}_{\\mathrm{pix}}" + }, + { + "bbox": [ + 313, + 519, + 555, + 639 + ], + "type": "text", + "content": ", such as SwinIR [27]:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 386, + 645, + 553, + 659 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 645, + 553, + 659 + ], + "spans": [ + { + "bbox": [ + 386, + 645, + 553, + 659 + ], + "type": "interline_equation", + "content": "z _ {\\text {p r e - r e s}} = \\mathcal {E} \\left(\\mathcal {R} _ {\\text {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right). \\tag {3}", + "image_path": "f4ec2e11afcbe280e57d80b2e7678c9caada320bdc401d1b73af38d1380622e0.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 554, + 714 + ], + "type": "text", + "content": "Although introducing strong noise in the diffusion process enables the generation of additional information by heavily relying on the diffusion prior, preserving valuable clues in the LQ image is more crucial for the TDIR prob" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10153" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 69, + 296, + 134 + ], + "blocks": [ + { + "bbox": [ + 56, + 69, + 296, + 134 + ], + "lines": [ + { + "bbox": [ + 56, + 69, + 296, + 134 + ], + "spans": [ + { + "bbox": [ + 56, + 69, + 296, + 134 + ], + "type": "image", + "image_path": "f7b05482ebbed1a55492bb1aad1097269bbc8e765741bb2638b1c064a27b46be.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 57, + 134, + 293, + 146 + ], + "lines": [ + { + "bbox": [ + 57, + 134, + 293, + 146 + ], + "spans": [ + { + "bbox": [ + 57, + 134, + 293, + 146 + ], + "type": "text", + "content": "Figure 3. Illustration of pre-restoration and partial diffusion." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 55, + 152, + 296, + 190 + ], + "type": "text", + "content": "lem. Thus, we propose adding mild noise to the pre-restored images and starting generation from it. We denote this technique as partial diffusion, which can be formally defined as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 103, + 193, + 295, + 223 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 193, + 295, + 223 + ], + "spans": [ + { + "bbox": [ + 103, + 193, + 295, + 223 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} z _ {t, \\text {p a r t i a l}} = q \\left(z _ {t \\leq t _ {p}} \\mid z _ {0} = z _ {\\text {p r e - r e s}}\\right) \\tag {4} \\\\ = \\sqrt {\\bar {\\alpha} _ {t}} z _ {\\text {p r e - r e s}} + \\sqrt {1 - \\bar {\\alpha} _ {t}} \\epsilon , \\\\ \\end{array}", + "image_path": "69d89b62860198a8e60b55eb6577399ae085a522a5721b1ca2d02595f7fe04f8.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "spans": [ + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "inline_equation", + "content": "z_{t,\\text{partial}}" + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "content": " is the partial diffusion result and " + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "inline_equation", + "content": "t_p" + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "content": " is a fixed hyper-parameter with a relatively smaller value than " + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "content": ". We empirically set " + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "inline_equation", + "content": "(t_p, T) = (200, 1000)" + }, + { + "bbox": [ + 55, + 231, + 296, + 280 + ], + "type": "text", + "content": ". Figure 3 illustrates the proposed diffusion process." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": "Short-step denoising. As noted in the preliminary section, conventional SD-based IR methods employ long-step denoising (e.g., 50). However, in the context of the TDIR problem, we find that a large number of denoising steps negatively impact on task performance. Although each step introduces details through the diffusion prior, repeatedly adding such details can generate redundant content, thereby diluting crucial task-relevant information. Hence, we adopt a short-step denoising (e.g., 1, 4) during inference. Specifically, for a given number of denoising steps " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": ", we denoise " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "z_{t_p,\\text{partial}}" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": " by passing it through " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": " timesteps: " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = [t_p, \\lfloor t_p / n \\cdot (n - 1) \\rfloor, \\dots, \\lfloor t_p / n \\rfloor]" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": ", following [37], and obtain the final denoising result " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "z_{\\text{diff-res}}" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": ". In the one-step denosing case (i.e., " + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "inline_equation", + "content": "n = 1" + }, + { + "bbox": [ + 55, + 282, + 296, + 450 + ], + "type": "text", + "content": "), this process simplifies to:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 63, + 456, + 295, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 456, + 295, + 497 + ], + "spans": [ + { + "bbox": [ + 63, + 456, + 295, + 497 + ], + "type": "interline_equation", + "content": "z _ {\\text {d i f f - r e s}} = \\frac {z _ {t _ {p} , \\text {p a r t i a l}} - \\sqrt {1 - \\bar {\\alpha} _ {t _ {p}}} \\epsilon_ {\\theta} \\left(z _ {t _ {p} , \\text {p a r t i a l}}, t _ {p} , z _ {\\text {p r e - r e s}}\\right)}{\\sqrt {\\bar {\\alpha} _ {t _ {p}}}}. \\tag {5}", + "image_path": "00b0e03b06a2b245575c7476b164019bdb226f4eea5f1e0701febd93d62a0d30.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 570 + ], + "type": "text", + "content": "Note that conventional SD-based IR methods [30, 52] struggle to generate visually pleasing results in short-step denoising scenarios [54, 58]. However, we find that optimizing loss functions, specifically designed to restore task-relevant details as discussed in Section 3.3, enables the achievement of visually pleasing results, even in one-step denoising." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 570, + 296, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 642 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 642 + ], + "type": "text", + "content": "Figure 4 compares the overall IR procedure. Unlike conventional methods, which start from pure noise and require around 50 denoising steps, we use the input LQ image with pre-restoration and partial diffusion, followed by short-step denoising. This approach directly leverages useful clues in LQ images to restore task-relevant details." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": "Decoding into RGB space. After the denoising process, we transform the " + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "inline_equation", + "content": "z_{\\mathrm{diff-res}}" + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": " back into the image space using the VAE decoder " + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{D}" + }, + { + "bbox": [ + 55, + 646, + 296, + 693 + ], + "type": "text", + "content": ". We further apply Wavelet color correction [52] to achieve a faithful reconstruction as follows:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 91, + 699, + 295, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 699, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 91, + 699, + 295, + 713 + ], + "type": "interline_equation", + "content": "I _ {\\mathrm {E D T R}} = \\mathbf {H} \\left(\\mathcal {D} \\left(z _ {\\text {d i f f - r e s}}\\right)\\right) + \\mathbf {L} \\left(\\mathcal {R} _ {\\operatorname {p i x}} \\left(I _ {\\mathrm {L Q}}\\right)\\right), \\tag {6}", + "image_path": "7efcc9601c581597c1dc576d9566a7ea34cfa7df2897b020435b69382691ef82.jpg" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 328, + 70, + 542, + 190 + ], + "blocks": [ + { + "bbox": [ + 328, + 70, + 542, + 190 + ], + "lines": [ + { + "bbox": [ + 328, + 70, + 542, + 190 + ], + "spans": [ + { + "bbox": [ + 328, + 70, + 542, + 190 + ], + "type": "image", + "image_path": "fe2f5071225bb7a808144a9c4292d608701251a497b41de27000a09f2fe255a2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 334, + 191, + 533, + 202 + ], + "lines": [ + { + "bbox": [ + 334, + 191, + 533, + 202 + ], + "spans": [ + { + "bbox": [ + 334, + 191, + 533, + 202 + ], + "type": "text", + "content": "(b) Pre-restoration, partial diffusion, and short-step denosing." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 228 + ], + "type": "text", + "content": "Figure 4. The comparison of IR procedure. We omit the encoding, decoding, and conditioning for simplicity in visualization." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR}}" + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "text", + "content": " is the final restored result from our EDTR. The " + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{H}(\\cdot)" + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "inline_equation", + "content": "\\mathbf{L}(\\cdot)" + }, + { + "bbox": [ + 313, + 233, + 554, + 282 + ], + "type": "text", + "content": " correspond to the high-frequency and low-frequency components extracted by Wavelet transform. Figure 2a illustrates the overall architecture of our EDTR." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 289, + 436, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 436, + 301 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 436, + 301 + ], + "type": "text", + "content": "3.3. Training Framework" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 303, + 554, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 303, + 554, + 386 + ], + "spans": [ + { + "bbox": [ + 313, + 303, + 554, + 386 + ], + "type": "text", + "content": "High-level feature loss. The goal of TDIR is to restore task-relevant details to enhance task performance. However, the noise prediction loss in Equation (2), which is conventionally used for training diffusion models, is not designed for this purpose. Hence, we introduce high-level feature (HLF) loss, which can guide the model to harness the diffusion prior for restoring task-relevant details." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "text", + "content": "HLF loss measures the feature distance between EDTR restored results and HQ images, following the previous TDIR method [20]. Specifically, we compute the HLF loss using the feature space of two task networks: the task network " + }, + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "text", + "content": ", which is optimized during training, and another task network " + }, + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 313, + 387, + 554, + 483 + ], + "type": "text", + "content": ", pre-trained on HQ images, to capture complementary task-relevant information (see Supplementary Section S5). Formally, the HLF loss is defined as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 338, + 484, + 553, + 530 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 484, + 553, + 530 + ], + "spans": [ + { + "bbox": [ + 338, + 484, + 553, + 530 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {H L F}} = \\frac {1}{2} \\left(\\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R}, \\text {t r a i n}}\\right) - \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1} \\right. \\tag {7} \\\\ + \\left\\| \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {E D T R , t r a i n}}) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} (I _ {\\mathrm {H Q}}) \\right\\| _ {1}), \\\\ \\end{array}", + "image_path": "596d36abf3750b3012788bdc805aa8694046b77cacf2153d2b6f0feeb290aa35.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}^f" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}^f" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " are intermediate feature spaces of " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": ", respectively. Note that " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR,train}}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " in Equation (7) represents the restored image from the one-step denoising result of " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "z_{t,\\mathrm{partial}}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " is randomly sampled from " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " during training, following the conventional diffusion loss in Equation (2), regardless of denoising steps " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": " during inference (see Supplementary Algorithm S1 for details). The HLF loss updates only EDTR and does not affect " + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 532, + 554, + 629 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": "Task loss. While training the proposed EDTR to restore task-relevant details, we also train the task networks using the restored outputs. We utilize both " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR}}" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": " and HQ images to stabilize training and introduce additional regularization. Formally, task loss " + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 313, + 632, + 554, + 693 + ], + "type": "text", + "content": " for the task network is defined as:" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 367, + 700, + 553, + 713 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 367, + 700, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 367, + 700, + 553, + 713 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t a s k}} = f _ {\\text {t a s k}} \\left(\\mathcal {H} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right), y\\right), \\tag {8}", + "image_path": "f81b1da3d2b6e0c70e3a9b841fdb8e036732dc35dc768f3c53c3f21572a3ed96.jpg" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10154" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{task}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " is the standard loss function for each task, e.g., cross-entropy loss for classification, " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR}} \\oplus_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " represents batch-wise concatenated images, consisting of a half-batch from " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and a half-batch from " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "y" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " are corresponding task labels. Note that " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " updates only the task network " + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": " and does not update EDTR." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "spans": [ + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "text", + "content": "Feature matching loss. To align the feature space of task network " + }, + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "text", + "content": " with " + }, + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 55, + 146, + 295, + 195 + ], + "type": "text", + "content": ", which is well-trained to utilize high-frequency details present in HQ images, we further introduce feature matching (FM) loss, defined as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 88, + 198, + 295, + 218 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 88, + 198, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 88, + 198, + 295, + 218 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {F M}} = \\left\\| \\mathcal {H} ^ {f} \\left(I _ {\\mathrm {E D T R} \\oplus \\mathrm {H Q}}\\right) - \\mathcal {H} _ {\\mathrm {H Q}} ^ {f} \\left(I _ {\\mathrm {H Q}}\\right) \\right\\| _ {1}. \\tag {9}", + "image_path": "8d72e62ef0a8115eb3e77eb34b73c4811ff19eedfd272217a53df7a026cc4508.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": "For clarity, note that " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR}}" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": " in Equations (8) and (9) is the " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": " step denoising result, different from the one-step result " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{EDTR,train}}" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": " in Equation (7). The " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FM}}" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": " can be interpreted as a form of cross-quality knowledge distillation [16, 48] at the feature level. Similar to the task loss, " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{FM}}" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": " only updates " + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 222, + 295, + 282 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "text", + "content": "Joint training in an alternating manner. Since SD was originally trained to operate in the image space, backpropagating the task loss, such as cross-entropy loss that is not image-based, causes instability to the EDTR. To avoid this, we train both the EDTR and the task network " + }, + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "text", + "content": " using each corresponding training loss in an alternating manner [20], as illustrated in Figure 2b. Specifically, in each training iteration, we first update the EDTR using the HLF loss, then update " + }, + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 285, + 296, + 405 + ], + "type": "text", + "content": " using a combination of the task loss and the FM loss, which can be formally written as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 94, + 407, + 294, + 426 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 407, + 294, + 426 + ], + "spans": [ + { + "bbox": [ + 94, + 407, + 294, + 426 + ], + "type": "interline_equation", + "content": "\\min _ {\\boldsymbol {\\theta} _ {\\mathrm {E D T R}}} \\mathcal {L} _ {\\mathrm {H L F}}, \\text {t h e n} \\min _ {\\boldsymbol {\\theta} _ {\\mathcal {H}}} \\mathcal {L} _ {\\text {t a s k}} + \\alpha \\mathcal {L} _ {\\mathrm {F M}}, \\tag {10}", + "image_path": "78cdfad9136c3098eefd017ae1d0a339def9849357cfe11721a44d7edebe730e.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathrm{EDTR}}" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\theta_{\\mathcal{H}}" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": " are the trainable parameters of the EDTR and task network " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 430, + 295, + 455 + ], + "type": "text", + "content": " is the hyper-parameter." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 464, + 137, + 477 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 464, + 137, + 477 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 137, + 477 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "content": "We validate our EDTR on three representative high-level vision tasks: image classification, semantic segmentation, and object detection, under scenarios involving image degradation. Note that the final prediction " + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\hat{y}" + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "content": " for each task is obtained as: " + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\hat{y} = \\mathcal{H}(\\mathrm{EDTR}(I_{\\mathrm{LQ}}))" + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "content": " is the trained task network and " + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{LQ}}" + }, + { + "bbox": [ + 55, + 480, + 295, + 553 + ], + "type": "text", + "content": " is the input low-quality (LQ) images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 555, + 295, + 663 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 555, + 295, + 663 + ], + "spans": [ + { + "bbox": [ + 55, + 555, + 295, + 663 + ], + "type": "text", + "content": "Datasets. We use three high-level vision tasks with representative datasets: CUB200 [51] for classification, and PASCAL VOC2012 [12] for segmentation and detection. We construct two degraded LQ image sets: 1) Mixture-A: downsampling combined with JPEG compression, and 2) Mixture-B: a combination of downsampling, blur, noise, and compression, following the CodeFormer [68] degradation. Details on degradation types are in the second column of Table 1 and Section S1 in the supplementary materials." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 713 + ], + "type": "text", + "content": "Evaluation metrics. We use representative evaluation metrics for each task: top-1 accuracy (Acc) for classification, mean Intersection over Union (mIoU) for segmentation, and mean Average Precision (mAP) for detection. The" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "mAP score is measured following the standard COCO [29] evaluation setting. We use two no-reference metrics to assess visual appeal: NIQE [36] and Q-Align [56], where the latter leveraging large vision-language models. We also report the standard PSNR for comparison." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "content": "Network architecture. We use StableDiffusion-2.1 [41] (SD) and ControlNet [63] for EDTR. ControlNet and the VAE decoder of SD are set as trainable modules. We also incorporate the pixel-error-based SwinIR [27] model as the pre-restoration network. For the task network, we employ representative architecture for each task: ResNet [15] for classification, DeepLabV3 [6] for segmentation, and FasterRCNN [40] for detection. The task network " + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "content": " shares the same architecture as " + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "content": " (the network pre-trained at HQ), and the weights of " + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "content": " are initialized from " + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 313, + 137, + 555, + 258 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 266, + 437, + 278 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 437, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 437, + 278 + ], + "type": "text", + "content": "4.1. Experimental Results" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "content": "Table 1 and Figure 5 present quantitative and qualitative results, respectively, across various types of degradation in image classification, semantic segmentation, and object detection. Note that in Table 1, 'Oracle', 'No restoration', and 'SwinIR' refer to the performance when the task network is trained and tested directly on HQ, LQ, and pixel-error-based SwinIR restored images, respectively. EDTR-" + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "content": " step refers to the EDTR model with " + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 284, + 554, + 380 + ], + "type": "text", + "content": " denoising steps." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "spans": [ + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": "Image classification. Table 1 shows that the proposed EDTR significantly outperforms previous methods under both types of degradation, achieving classification accuracy improvements of " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "+2.9\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "inline_equation", + "content": "+5.4\\%" + }, + { + "bbox": [ + 313, + 384, + 555, + 612 + ], + "type": "text", + "content": " over the previous state-of-the-art task-driven image restoration (TDIR) method, SR4IR [20]. Performance gains increase as the degradation becomes more severe, highlighting our method's efficacy in practical scenarios. Furthermore, EDTR achieves the highest scores in image quality assessments such as NIQE and Q-Align, indicating its effectiveness in producing visually pleasing results. However, despite the superior task performance, EDTR does not achieve the highest PSNR score, suggesting that PSNR is a less relevant metric in TDIR. As presented in the first row of Figure 5, our EDTR, leveraging a strong natural image prior, successfully restores detailed wing patterns—an important area for classification as shown by the Grad-CAM [43] in the upper-left—and correctly predicts the label, whereas the previous method fails to restore these details and misclassifies the image." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": "Semantic segmentation. The performance improvement in segmentation is more notable than in classification, with mIoU gains of " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "+9.1\\%" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "inline_equation", + "content": "+11.9\\%" + }, + { + "bbox": [ + 313, + 618, + 556, + 714 + ], + "type": "text", + "content": " over the previous state-of-the-art. Although our EDTR achieves the second-best score in NIQE, it consistently achieves the highest Q-Align score, demonstrating its capability to generate perceptually realistic images. In addition, unlike classification, increasing the number of denoising steps from 1 to 4 leads to a task" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10155" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 71, + 553, + 240 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 553, + 240 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 553, + 240 + ], + "type": "table", + "html": "
IR MethodsDegradation typeImage classificationSemantic segmentationObject detection
Task Acc↑ (%)Visual qualityTask mIoU↑ (%)Visual qualityTask mAP↑ (%)Visual quality
NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑NIQE↓Q-Align↑PSNR↑
OracleNone82.55.373.69+inf67.04.714.02+inf36.93.703.95+inf
No restoration60.814.081.0323.7247.413.671.1622.4314.514.541.0421.08
SwinIR [27]70.09.972.5525.2156.28.662.5423.4721.08.962.2221.70
TDSR [14]Mixture-A:66.17.532.5124.9951.37.672.3723.4116.26.432.1321.62
RSRSSN [67]SR(s=8)+69.27.322.7524.0453.06.972.5823.0622.36.592.1021.40
SR4IR [20]JPEG(q=75)71.55.413.3724.0756.24.803.1622.8522.35.383.0020.95
EDTR-1 step (Ours)74.44.683.5323.6364.15.303.5922.3331.24.623.6820.60
EDTR-4 step (Ours)74.14.263.8122.4665.35.173.9121.4833.44.194.0119.89
No restoration47.615.111.0823.0540.214.841.1922.100.015.241.0720.86
SwinIR [27]Mixture-B:60.710.262.0725.0450.49.472.3023.6722.19.601.9222.20
TDSR [14]SR(s[1,16]) +56.47.292.2024.6846.08.072.1323.5514.66.361.9421.97
RSRSSN [67]Blur(σ[0,8]) +61.47.352.3523.3548.27.722.3623.0123.16.811.8621.68
SR4IR [20]Noise(σ[0,10]) +63.46.083.1123.6251.04.723.1522.7122.64.523.1420.97
EDTR-1 step (Ours)JPEG(q[50,100])68.84.753.4823.0360.46.363.5522.0430.64.483.6420.70
EDTR-4 step (Ours)68.44.463.4923.2262.95.803.7422.2431.94.124.0220.29
", + "image_path": "b585e2f5222664d1b7c19c12a8682bcdbcde556a1c409865b46f6ab9f9973955.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 55, + 240, + 555, + 264 + ], + "type": "text", + "content": "Table 1. Performance of high-level vision tasks and quality of restored image results under various degradation types. We indicate the best and second-best results with bold and underlined formatting, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 57, + 274, + 180, + 377 + ], + "blocks": [ + { + "bbox": [ + 57, + 274, + 180, + 377 + ], + "lines": [ + { + "bbox": [ + 57, + 274, + 180, + 377 + ], + "spans": [ + { + "bbox": [ + 57, + 274, + 180, + 377 + ], + "type": "image", + "image_path": "4b7f455994bc8490a864b5bc4db556f84a078427738621a178ea936056825340.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 379, + 161, + 390 + ], + "lines": [ + { + "bbox": [ + 64, + 379, + 161, + 390 + ], + "spans": [ + { + "bbox": [ + 64, + 379, + 161, + 390 + ], + "type": "text", + "content": "pred: [Blue Grosbeak]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 181, + 274, + 304, + 377 + ], + "blocks": [ + { + "bbox": [ + 181, + 274, + 304, + 377 + ], + "lines": [ + { + "bbox": [ + 181, + 274, + 304, + 377 + ], + "spans": [ + { + "bbox": [ + 181, + 274, + 304, + 377 + ], + "type": "image", + "image_path": "49ffc14a44bb4a43fc9c07016fd3798afb5ad78362f3b6ea791635c0a69b0931.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 188, + 379, + 302, + 389 + ], + "lines": [ + { + "bbox": [ + 188, + 379, + 302, + 389 + ], + "spans": [ + { + "bbox": [ + 188, + 379, + 302, + 389 + ], + "type": "text", + "content": "pred: [Blue Grosbeak] X" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 306, + 274, + 429, + 377 + ], + "blocks": [ + { + "bbox": [ + 306, + 274, + 429, + 377 + ], + "lines": [ + { + "bbox": [ + 306, + 274, + 429, + 377 + ], + "spans": [ + { + "bbox": [ + 306, + 274, + 429, + 377 + ], + "type": "image", + "image_path": "28d8ad8fe15c2a50b2aae041616803a6a82812eb31db92e5818ae24c682928d5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 310, + 379, + 411, + 389 + ], + "lines": [ + { + "bbox": [ + 310, + 379, + 411, + 389 + ], + "spans": [ + { + "bbox": [ + 310, + 379, + 411, + 389 + ], + "type": "text", + "content": "pred: [Indigo Bunting]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 430, + 274, + 553, + 377 + ], + "blocks": [ + { + "bbox": [ + 430, + 274, + 553, + 377 + ], + "lines": [ + { + "bbox": [ + 430, + 274, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 430, + 274, + 553, + 377 + ], + "type": "image", + "image_path": "25a63364286dd88b5fc466012482d3a0b9045e78d6b0dd431ce4ea7107460f5f.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 379, + 550, + 389 + ], + "lines": [ + { + "bbox": [ + 433, + 379, + 550, + 389 + ], + "spans": [ + { + "bbox": [ + 433, + 379, + 550, + 389 + ], + "type": "text", + "content": "pred: [Indigo Bunting]" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 58, + 398, + 119, + 452 + ], + "blocks": [ + { + "bbox": [ + 58, + 398, + 119, + 452 + ], + "lines": [ + { + "bbox": [ + 58, + 398, + 119, + 452 + ], + "spans": [ + { + "bbox": [ + 58, + 398, + 119, + 452 + ], + "type": "image", + "image_path": "8cded3d279935b22e7a947771095d5638b0efb764ffde57ad242c3c669f0f830.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 453, + 119, + 513 + ], + "blocks": [ + { + "bbox": [ + 58, + 453, + 119, + 513 + ], + "lines": [ + { + "bbox": [ + 58, + 453, + 119, + 513 + ], + "spans": [ + { + "bbox": [ + 58, + 453, + 119, + 513 + ], + "type": "image", + "image_path": "76ea47ecb4662490a7e3acef484ab8410f486d67af0b5c36c6e579f50b4863c5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 120, + 399, + 181, + 452 + ], + "blocks": [ + { + "bbox": [ + 120, + 399, + 181, + 452 + ], + "lines": [ + { + "bbox": [ + 120, + 399, + 181, + 452 + ], + "spans": [ + { + "bbox": [ + 120, + 399, + 181, + 452 + ], + "type": "image", + "image_path": "c70905a90aa2b735178675c5e59f88055f54b1594d763e78cc0cab5869950d19.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 120, + 452, + 181, + 513 + ], + "blocks": [ + { + "bbox": [ + 120, + 452, + 181, + 513 + ], + "lines": [ + { + "bbox": [ + 120, + 452, + 181, + 513 + ], + "spans": [ + { + "bbox": [ + 120, + 452, + 181, + 513 + ], + "type": "image", + "image_path": "d2b34b8fae38ef6ad016c45be54f5d9d6c2bc3e6686bb0e9092492a43d11cab9.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 181, + 399, + 244, + 452 + ], + "blocks": [ + { + "bbox": [ + 181, + 399, + 244, + 452 + ], + "lines": [ + { + "bbox": [ + 181, + 399, + 244, + 452 + ], + "spans": [ + { + "bbox": [ + 181, + 399, + 244, + 452 + ], + "type": "image", + "image_path": "ad9405d657b8fc6f5d3d89cc2bd714aa89e5813c8d0908bd085919a02adbe906.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 182, + 452, + 243, + 513 + ], + "blocks": [ + { + "bbox": [ + 182, + 452, + 243, + 513 + ], + "lines": [ + { + "bbox": [ + 182, + 452, + 243, + 513 + ], + "spans": [ + { + "bbox": [ + 182, + 452, + 243, + 513 + ], + "type": "image", + "image_path": "3d9be70e7ecdf95a7abee7dc90eb266fa778051cede40d0b743e1f4bc8ee5814.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 244, + 399, + 305, + 452 + ], + "blocks": [ + { + "bbox": [ + 244, + 399, + 305, + 452 + ], + "lines": [ + { + "bbox": [ + 244, + 399, + 305, + 452 + ], + "spans": [ + { + "bbox": [ + 244, + 399, + 305, + 452 + ], + "type": "image", + "image_path": "4ceae1e77ea961a6651d17ef64d4dc1ff1ee07741ac447369f0d3df56c5305c3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 244, + 452, + 304, + 513 + ], + "blocks": [ + { + "bbox": [ + 244, + 452, + 304, + 513 + ], + "lines": [ + { + "bbox": [ + 244, + 452, + 304, + 513 + ], + "spans": [ + { + "bbox": [ + 244, + 452, + 304, + 513 + ], + "type": "image", + "image_path": "cb2c7f41b010ebaea61063f7c7ec2a27ad457cd7adb63a00585483fc68857a2b.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 306, + 399, + 368, + 452 + ], + "blocks": [ + { + "bbox": [ + 306, + 399, + 368, + 452 + ], + "lines": [ + { + "bbox": [ + 306, + 399, + 368, + 452 + ], + "spans": [ + { + "bbox": [ + 306, + 399, + 368, + 452 + ], + "type": "image", + "image_path": "3284048523fe6c6f2f53c7a0f036623de4fa4d6ac00955a255ab3e8de5018cc7.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 306, + 452, + 367, + 513 + ], + "blocks": [ + { + "bbox": [ + 306, + 452, + 367, + 513 + ], + "lines": [ + { + "bbox": [ + 306, + 452, + 367, + 513 + ], + "spans": [ + { + "bbox": [ + 306, + 452, + 367, + 513 + ], + "type": "image", + "image_path": "1e12c188fec29898a01cb6d29a17a6b9813ed545ceb16e2cd20edd3a1ae5a3bb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 368, + 399, + 429, + 452 + ], + "blocks": [ + { + "bbox": [ + 368, + 399, + 429, + 452 + ], + "lines": [ + { + "bbox": [ + 368, + 399, + 429, + 452 + ], + "spans": [ + { + "bbox": [ + 368, + 399, + 429, + 452 + ], + "type": "image", + "image_path": "02e2f232754a517f8a8424d33ba2499eefe8d4a0a74df7deb52957b1a2415e36.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 368, + 452, + 429, + 513 + ], + "blocks": [ + { + "bbox": [ + 368, + 452, + 429, + 513 + ], + "lines": [ + { + "bbox": [ + 368, + 452, + 429, + 513 + ], + "spans": [ + { + "bbox": [ + 368, + 452, + 429, + 513 + ], + "type": "image", + "image_path": "8a65090f74b838e951aed696aa56b10e945c23b3bd5a31fb6434f972e5789513.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 430, + 399, + 492, + 452 + ], + "blocks": [ + { + "bbox": [ + 430, + 399, + 492, + 452 + ], + "lines": [ + { + "bbox": [ + 430, + 399, + 492, + 452 + ], + "spans": [ + { + "bbox": [ + 430, + 399, + 492, + 452 + ], + "type": "image", + "image_path": "e0ff27934d4c0e00b7c18ea3364c9e5a21a540423adcdaf22876cc4cec9cef91.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 430, + 452, + 491, + 513 + ], + "blocks": [ + { + "bbox": [ + 430, + 452, + 491, + 513 + ], + "lines": [ + { + "bbox": [ + 430, + 452, + 491, + 513 + ], + "spans": [ + { + "bbox": [ + 430, + 452, + 491, + 513 + ], + "type": "image", + "image_path": "bce4efbb93b19de2a1a1946bb2911920290d7755bf9b447532028dd85f3c353b.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 493, + 399, + 553, + 452 + ], + "blocks": [ + { + "bbox": [ + 493, + 399, + 553, + 452 + ], + "lines": [ + { + "bbox": [ + 493, + 399, + 553, + 452 + ], + "spans": [ + { + "bbox": [ + 493, + 399, + 553, + 452 + ], + "type": "image", + "image_path": "74c0a4ee6dfdcc8a12d0552a6972ac3883328170843aa818ea993d36b6d65727.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 493, + 452, + 553, + 513 + ], + "blocks": [ + { + "bbox": [ + 493, + 452, + 553, + 513 + ], + "lines": [ + { + "bbox": [ + 493, + 452, + 553, + 513 + ], + "spans": [ + { + "bbox": [ + 493, + 452, + 553, + 513 + ], + "type": "image", + "image_path": "cd397e4b0dee1dd57047c52f8f5b73c5d30810a0580c02184b2eeb438c96068a.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 58, + 521, + 180, + 582 + ], + "blocks": [ + { + "bbox": [ + 58, + 521, + 180, + 582 + ], + "lines": [ + { + "bbox": [ + 58, + 521, + 180, + 582 + ], + "spans": [ + { + "bbox": [ + 58, + 521, + 180, + 582 + ], + "type": "image", + "image_path": "6e123f70a3cf4531b42709a319ba1bfe17a9437d7e1616ccd4a82557e9a74413.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 58, + 582, + 180, + 633 + ], + "blocks": [ + { + "bbox": [ + 58, + 582, + 180, + 633 + ], + "lines": [ + { + "bbox": [ + 58, + 582, + 180, + 633 + ], + "spans": [ + { + "bbox": [ + 58, + 582, + 180, + 633 + ], + "type": "image", + "image_path": "1aba62c84f3fd0fb7c6f9c795fe7cf3a790b50ac92b0ddd29320f69c03bdc567.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 635, + 152, + 643 + ], + "lines": [ + { + "bbox": [ + 85, + 635, + 152, + 643 + ], + "spans": [ + { + "bbox": [ + 85, + 635, + 152, + 643 + ], + "type": "text", + "content": "(a) LQ (No restoration)" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 181, + 521, + 304, + 582 + ], + "blocks": [ + { + "bbox": [ + 181, + 521, + 304, + 582 + ], + "lines": [ + { + "bbox": [ + 181, + 521, + 304, + 582 + ], + "spans": [ + { + "bbox": [ + 181, + 521, + 304, + 582 + ], + "type": "image", + "image_path": "9207c3fde75b4530a3a66da74ef0c53b27476652da8a1193295b226b25e27220.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 181, + 582, + 304, + 633 + ], + "blocks": [ + { + "bbox": [ + 181, + 582, + 304, + 633 + ], + "lines": [ + { + "bbox": [ + 181, + 582, + 304, + 633 + ], + "spans": [ + { + "bbox": [ + 181, + 582, + 304, + 633 + ], + "type": "image", + "image_path": "ee2fea1af0b2212d5cc7fd9a59ad1bcc3b36b64a716ea7d8a8aeb713e5c495f8.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 221, + 635, + 265, + 643 + ], + "lines": [ + { + "bbox": [ + 221, + 635, + 265, + 643 + ], + "spans": [ + { + "bbox": [ + 221, + 635, + 265, + 643 + ], + "type": "text", + "content": "(b) SR4IR [20]" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 306, + 521, + 429, + 582 + ], + "blocks": [ + { + "bbox": [ + 306, + 521, + 429, + 582 + ], + "lines": [ + { + "bbox": [ + 306, + 521, + 429, + 582 + ], + "spans": [ + { + "bbox": [ + 306, + 521, + 429, + 582 + ], + "type": "image", + "image_path": "ceca33532dc97d8f5641d07ce682334443a966fd43e50f9af7f805845a8ef7ca.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + } + ], + "index": 32 + }, + { + "type": "image", + "bbox": [ + 306, + 582, + 429, + 633 + ], + "blocks": [ + { + "bbox": [ + 306, + 582, + 429, + 633 + ], + "lines": [ + { + "bbox": [ + 306, + 582, + 429, + 633 + ], + "spans": [ + { + "bbox": [ + 306, + 582, + 429, + 633 + ], + "type": "image", + "image_path": "42e5ee08054fb7a6d4f4318949c69523152dec3e2a54b9a78267e98e0acfbc65.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 341, + 635, + 393, + 643 + ], + "lines": [ + { + "bbox": [ + 341, + 635, + 393, + 643 + ], + "spans": [ + { + "bbox": [ + 341, + 635, + 393, + 643 + ], + "type": "text", + "content": "(c)EDTR(Ours)" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 645, + 555, + 689 + ], + "lines": [ + { + "bbox": [ + 55, + 645, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 645, + 555, + 689 + ], + "type": "text", + "content": "Figure 5. Visualization of images and diverse high-level vision task results on degraded LQ (Mixture-B) images. The first, second, and third rows represent the results of image classification, semantic segmentation, and object detection, respectively. We show the restored images and the corresponding predicted labels. For visualization, we use the EDTR-1 step model for classification and the EDTR-4 step model for segmentation and detection. For more visualization results, please refer to Section S12 of our supplementary materials." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_caption" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 430, + 521, + 553, + 582 + ], + "blocks": [ + { + "bbox": [ + 430, + 521, + 553, + 582 + ], + "lines": [ + { + "bbox": [ + 430, + 521, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 430, + 521, + 553, + 582 + ], + "type": "image", + "image_path": "3357a641f72a1e5da0ab1a2095bd7cc520bf32df7247eba302f5cb84c5f0b978.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 430, + 582, + 553, + 633 + ], + "blocks": [ + { + "bbox": [ + 430, + 582, + 553, + 633 + ], + "lines": [ + { + "bbox": [ + 430, + 582, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 430, + 582, + 553, + 633 + ], + "type": "image", + "image_path": "d6e7c306f0c47135aa46e503d73079b89693563996c40c27a1958624a5a7f6f1.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 468, + 635, + 515, + 643 + ], + "lines": [ + { + "bbox": [ + 468, + 635, + 515, + 643 + ], + "spans": [ + { + "bbox": [ + 468, + 635, + 515, + 643 + ], + "type": "text", + "content": "(d) HQ (Oracle)" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10156" + } + ] + } + ], + "index": 39 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 71, + 553, + 189 + ], + "blocks": [ + { + "bbox": [ + 58, + 71, + 553, + 189 + ], + "lines": [ + { + "bbox": [ + 58, + 71, + 553, + 189 + ], + "spans": [ + { + "bbox": [ + 58, + 71, + 553, + 189 + ], + "type": "table", + "html": "
ExpSDTraining lossDiffusion processAcc↑ (%)fd↓Q-Align↑
HLFFMPre-restorationPartial diffusion# of denoising steps
(1)✓X✓XX5064.30.6413.87
(2)✓✓✓XX5062.90.7542.37
(3)✓✓✓XX163.50.5793.18
(4)✓✓✓X✓166.60.5303.38
(5)✓X✓✓✓164.50.7192.18
(6)✓✓X✓✓167.60.5073.48
(7)X✓✓✓✓164.20.5413.20
Ours✓✓✓✓✓168.80.5073.48
", + "image_path": "5548bb3deab6c7bf8e1a41cf3327be2d1f66dc522270d2501d8740bcf1659a4f.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 123, + 190, + 487, + 201 + ], + "lines": [ + { + "bbox": [ + 123, + 190, + 487, + 201 + ], + "spans": [ + { + "bbox": [ + 123, + 190, + 487, + 201 + ], + "type": "text", + "content": "Table 2. Ablation study on the impact of each component on EDTR classification performance." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 54, + 211, + 295, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 211, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 54, + 211, + 295, + 319 + ], + "type": "text", + "content": "performance improvement. In the second row of Figure 5, the LQ image is severely degraded, causing the details in the upper-left part of the bike wheel to be almost lost. While the previous TDIR method fails to restore the missing details, our method effectively restores them by leveraging a strong natural image prior, correctly recognizing that they belong to the wheel. As a result, EDTR produces a segmentation prediction that correctly maps the upper-left part of the bike wheel, demonstrating its effectiveness." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "text", + "content": "Object detection. Our EDTR demonstrates a surprisingly significant performance gain over the previous state-of-the-art in the Mixture-A setting, increasing mAP from " + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "inline_equation", + "content": "22.3\\%" + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "inline_equation", + "content": "33.4\\%" + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "text", + "content": " —an improvement of approximately " + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 324, + 296, + 445 + ], + "type": "text", + "content": ". As shown in the third row of Figure 5, EDTR correctly restores the shape of the sofa by leveraging the powerful diffusion prior, successfully detecting the sofa in the room, whereas the previous method fails. These results demonstrate that our method consistently achieves superior task performance and visual quality under severe degradation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 453, + 154, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 154, + 464 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 154, + 464 + ], + "type": "text", + "content": "4.2. Ablation studies" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 470, + 277, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 470, + 277, + 483 + ], + "spans": [ + { + "bbox": [ + 55, + 470, + 277, + 483 + ], + "type": "text", + "content": "All experiments use the Mixture- " + }, + { + "bbox": [ + 55, + 470, + 277, + 483 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 470, + 277, + 483 + ], + "type": "text", + "content": " degradation setting." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 486, + 218, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 218, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 218, + 498 + ], + "type": "text", + "content": "4.2.1. Effect of proposed components." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "text", + "content": "Table 2 shows the classification performance of our EDTR according to the use of each component. Specifically, when a proposed component is not used, i.e., " + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "inline_equation", + "content": "X" + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "text", + "content": ", we follow the conventional SD-based IR method setting, as described in Section 3.1. Except for Exp-(7), we adopt the FM loss as a default, since it relates to task network training rather than the IR method. In addition, to assess how well task-relevant details are restored, we measure the distance in the feature space of the task network " + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "inline_equation", + "content": "\\mathcal{H}_{\\mathrm{HQ}}" + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "text", + "content": ", which is pre-trained on HQ images. We denote this fidelity metric as feature distance, defined as " + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "inline_equation", + "content": "f_{d} = \\left\\| \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{EDTR}}) - \\mathcal{H}_{\\mathrm{HQ}}^{f}(I_{\\mathrm{HQ}})\\right\\|_{1}" + }, + { + "bbox": [ + 55, + 502, + 295, + 639 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "text", + "content": "Exp-(1) shows performance without any of the proposed components, i.e., conventional SD-based IR method [30]. Compared to EDTR performance (last row), although it achieves a more perceptually realistic restoration, as indicated by the higher Q-Align score, its large " + }, + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "text", + "content": " value suggests that task-relevant details are not well restored, resulting in a " + }, + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "inline_equation", + "content": "4.5\\%" + }, + { + "bbox": [ + 55, + 639, + 296, + 723 + ], + "type": "text", + "content": " worse classification accuracy." + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 209, + 375, + 269 + ], + "blocks": [ + { + "bbox": [ + 316, + 209, + 375, + 269 + ], + "lines": [ + { + "bbox": [ + 316, + 209, + 375, + 269 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 375, + 269 + ], + "type": "image", + "image_path": "d33e4b6da70d19f40fe84df9876c9d0421ecdac5c3641095990d2893fec3646d.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 333, + 270, + 356, + 279 + ], + "lines": [ + { + "bbox": [ + 333, + 270, + 356, + 279 + ], + "spans": [ + { + "bbox": [ + 333, + 270, + 356, + 279 + ], + "type": "text", + "content": "(a) LQ" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 376, + 210, + 433, + 269 + ], + "blocks": [ + { + "bbox": [ + 376, + 210, + 433, + 269 + ], + "lines": [ + { + "bbox": [ + 376, + 210, + 433, + 269 + ], + "spans": [ + { + "bbox": [ + 376, + 210, + 433, + 269 + ], + "type": "image", + "image_path": "ae631f535949326764b912ca8f520abbf462487acb770ba5aeb90df2191fa7e8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 270, + 423, + 279 + ], + "lines": [ + { + "bbox": [ + 386, + 270, + 423, + 279 + ], + "spans": [ + { + "bbox": [ + 386, + 270, + 423, + 279 + ], + "type": "text", + "content": "(b) Exp-(2)" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 435, + 210, + 493, + 269 + ], + "blocks": [ + { + "bbox": [ + 435, + 210, + 493, + 269 + ], + "lines": [ + { + "bbox": [ + 435, + 210, + 493, + 269 + ], + "spans": [ + { + "bbox": [ + 435, + 210, + 493, + 269 + ], + "type": "image", + "image_path": "47d220f2689a8a3ff4604beae42dcac058213c83e7c718cf9cff4c97d34b86f3.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 435, + 270, + 493, + 279 + ], + "lines": [ + { + "bbox": [ + 435, + 270, + 493, + 279 + ], + "spans": [ + { + "bbox": [ + 435, + 270, + 493, + 279 + ], + "type": "text", + "content": "(c)EDTR(Ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 334, + 281, + 533, + 291 + ], + "lines": [ + { + "bbox": [ + 334, + 281, + 533, + 291 + ], + "spans": [ + { + "bbox": [ + 334, + 281, + 533, + 291 + ], + "type": "text", + "content": "Figure 6. Visualization comparison in classification." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 495, + 210, + 553, + 269 + ], + "blocks": [ + { + "bbox": [ + 495, + 210, + 553, + 269 + ], + "lines": [ + { + "bbox": [ + 495, + 210, + 553, + 269 + ], + "spans": [ + { + "bbox": [ + 495, + 210, + 553, + 269 + ], + "type": "image", + "image_path": "e212bf1ba0e3581f1e588acaca487471e3e85f240287a0929948a4d35caf4ad1.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 512, + 270, + 536, + 279 + ], + "lines": [ + { + "bbox": [ + 512, + 270, + 536, + 279 + ], + "spans": [ + { + "bbox": [ + 512, + 270, + 536, + 279 + ], + "type": "text", + "content": "(d) HQ" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 297, + 555, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 488 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 488 + ], + "type": "text", + "content": "Effect of pre-restoration & partial diffusion. Exp-(2) shows performance without our design choice for the diffusion process, i.e., SD is directly applied to TDIR. Although it incorporates a loss for restoring task-relevant details, its task performance is even worse than that of Exp-(1). Figure 6 illustrates that Exp-(2) leads to excessive generation and produces non-existent content, in contrast to the results of EDTR, ultimately failing to improve task performance. This highlights the necessity of a tailored diffusion process design for TDIR. Exp-(3) shows that even with short-step denoising, EDTR without partial diffusion does not perform well for TDIR, demonstrating the importance of partial diffusion. Exp-(4) further shows that the pre-restoration process is also essential for achieving both strong task performance and good visual quality. For performance regarding the pre-restoration network, see Supplementary Section S3." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 492, + 556, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 492, + 556, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 556, + 661 + ], + "type": "text", + "content": "Effect of HLF & FM losses. Exp-(5) and Exp-(6) show EDTR performance without HLF and FM losses, respectively. In both cases, task performance drops significantly, highlighting the importance of each training loss term. Interestingly, Exp-(5) results in a notably lower Q-Align score, indicating that the HLF loss is crucial for achieving visually pleasing results, and it enables the generation of such results even in the challenging one-step denoising scenario. This finding is consistent with previous research [47], as the HLF loss shares similarities with LPIPS [64] in measuring distance in feature space. In addition, incorporating feature spaces from two networks in the HLF loss (as in Equation (7)) yields a " + }, + { + "bbox": [ + 313, + 492, + 556, + 661 + ], + "type": "inline_equation", + "content": "+1.0\\%" + }, + { + "bbox": [ + 313, + 492, + 556, + 661 + ], + "type": "text", + "content": " gain compared to using only one (see Supplementary Section S5 for details)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Effect of SD. Exp-(7) shows the performance of EDTR without pre-trained SD weights. Specifically, we remove the frozen denoising network (purple module in Figure 2a); thus, the total number of trainable parameters remains" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10157" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": "nearly the same. Compared to this setting, EDTR largely reduces the " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "f_{d}" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": ", resulting in a significant " + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "inline_equation", + "content": "4.6\\%" + }, + { + "bbox": [ + 55, + 72, + 295, + 120 + ], + "type": "text", + "content": " increase in classification accuracy. These results highlight the crucial role of the diffusion prior in restoring task-relevant details." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 126, + 253, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 253, + 139 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 253, + 139 + ], + "type": "text", + "content": "4.2.2. Effect of the number of denoising steps." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 141, + 295, + 262 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 262 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 262 + ], + "type": "text", + "content": "Table 3 shows the performance of EDTR with varying numbers of denoising steps during inference. As shown in the table, a larger number of denoising steps (e.g., 50) leads to decreased task performance. This supports our claim that repeatedly introducing details is not beneficial for TDIR and may even dilute crucial task-relevant information. In addition, we observe that short-step denoising (" + }, + { + "bbox": [ + 55, + 141, + 295, + 262 + ], + "type": "inline_equation", + "content": "n = 4" + }, + { + "bbox": [ + 55, + 141, + 295, + 262 + ], + "type": "text", + "content": ") is sufficient to produce visually pleasing results, as indicated by its favorable Q-Align score—further confirming the effectiveness of the HLF loss in guiding perceptual quality." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 268, + 200, + 282 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 268, + 200, + 282 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 200, + 282 + ], + "type": "text", + "content": "4.3. Generalizability of EDTR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 286, + 295, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 286, + 295, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 286, + 295, + 467 + ], + "type": "text", + "content": "Performance of EDTR across various tasks. In Table 4, we present the performance of EDTR-1 step models, each trained separately on a specific task. Note that the diagonal (blue) values represent the performance of the original EDTR on its respective task. Since the original EDTR is independently trained for a single task, it performs well on its target task but less effectively on others. To demonstrate that our EDTR can be extended for broader applicability, we additionally train a single model on all three high-level vision tasks simultaneously, referring to this model as EDTR_All-in-One. The last row in the table presents the results. We observe that this model achieves performance comparable to, or even better than, the original EDTR across all three tasks, highlighting its potential for generalization to a variety of applications." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": "Generalization toward unseen datasets. Table 5 shows the classification performance on unseen datasets during training. Specifically, we train the IR models using TinyImageNet" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 55, + 469, + 295, + 589 + ], + "type": "text", + "content": " [23] and evaluate them on CUB200 [51] and StanfordCars [22] datasets, which were not seen during training. When evaluating on these unseen datasets, the task networks trained on HQ images corresponding to each dataset are used. The results show that our EDTR significantly outperforms other methods on these unseen datasets, demonstrating the robustness of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 592, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 592, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 592, + 295, + 677 + ], + "type": "text", + "content": "Results on real-world images. Figure 7 presents real-world object detection results and restored image visualizations. We compare (a) object detection without restoration and (b) EDTR-4 step trained on Mixture-" + }, + { + "bbox": [ + 55, + 592, + 295, + 677 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 592, + 295, + 677 + ], + "type": "text", + "content": " degradation. Due to complex real-world degradation, detection without restoration fails to identify the person in the image. In contrast, EDTR effectively restores key high-frequency details," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 72, + 552, + 110 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 552, + 110 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 552, + 110 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 552, + 110 + ], + "type": "table", + "html": "
# of denoising steps4304050
Classification (Acc↑ / Q-Align↑)68.4 / 3.4967.7 / 3.4667.1 / 3.4666.8 / 3.45
Segmentation (mIoU↑ / Q-Align↑)62.9 / 3.7462.4 / 3.5462.4 / 3.4661.8 / 3.43
Detection (mAP↑ / Q-Align↑)31.9 / 4.0231.8 / 3.9731.8 / 3.9931.6 / 3.94
", + "image_path": "917d7a1c84b1ab980b6b238e943048902b39fc5ea66fdb611472d2d809611544.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 317, + 129, + 552, + 185 + ], + "blocks": [ + { + "bbox": [ + 321, + 113, + 545, + 124 + ], + "lines": [ + { + "bbox": [ + 321, + 113, + 545, + 124 + ], + "spans": [ + { + "bbox": [ + 321, + 113, + 545, + 124 + ], + "type": "text", + "content": "Table 3. Effect of the number of denoising steps in EDTR." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 129, + 552, + 185 + ], + "lines": [ + { + "bbox": [ + 317, + 129, + 552, + 185 + ], + "spans": [ + { + "bbox": [ + 317, + 129, + 552, + 185 + ], + "type": "table", + "html": "
Eval task Used modelClassification (Acc↑)Segmentation (mIoU↑)Detection (mAP↑)
EDTR_Classification68.848.714.3
EDTR_Segmentation56.860.424.8
EDTR_Detection59.358.930.6
EDTR_All-in-One68.461.030.4
", + "image_path": "e71a61cce196b0f80fbc233f0924c53464c55ffc255c2acb08f3a5dd746fc4c9.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 317, + 215, + 552, + 286 + ], + "blocks": [ + { + "bbox": [ + 314, + 187, + 553, + 209 + ], + "lines": [ + { + "bbox": [ + 314, + 187, + 553, + 209 + ], + "spans": [ + { + "bbox": [ + 314, + 187, + 553, + 209 + ], + "type": "text", + "content": "Table 4. Performance of EDTR models across various tasks. EDTR_Task represents the EDTR model optimized for each task." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 215, + 552, + 286 + ], + "lines": [ + { + "bbox": [ + 317, + 215, + 552, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 215, + 552, + 286 + ], + "type": "table", + "html": "
Dataset IR methodsSeen \nTiny-ImageNet† [23]Unseen
CUB200 [51]StanfordCars [22]
Oracle83.782.590.4
No restoration61.210.112.1
SwinIR [27]68.021.838.7
SR4IR [20]68.943.346.1
EDTR-1 step (Ours)75.152.567.0
EDTR-4 step (Ours)76.153.066.4
", + "image_path": "e056ece91dfd62a1e745648c547d10b62af152ad382cf92860451133a1afc781.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_body" + } + ], + "index": 11 + }, + { + "bbox": [ + 334, + 289, + 531, + 298 + ], + "lines": [ + { + "bbox": [ + 334, + 289, + 531, + 298 + ], + "spans": [ + { + "bbox": [ + 334, + 289, + 531, + 298 + ], + "type": "text", + "content": "Table 5. Classification on seen and unseen datasets." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 316, + 304, + 432, + 409 + ], + "blocks": [ + { + "bbox": [ + 316, + 304, + 432, + 409 + ], + "lines": [ + { + "bbox": [ + 316, + 304, + 432, + 409 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 432, + 409 + ], + "type": "image", + "image_path": "99652a1c9d21ebdc2f0c179346df90c069af9df990824ac5c6b49e2cd15f75f4.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 335, + 411, + 411, + 420 + ], + "lines": [ + { + "bbox": [ + 335, + 411, + 411, + 420 + ], + "spans": [ + { + "bbox": [ + 335, + 411, + 411, + 420 + ], + "type": "text", + "content": "(a) LQ (No restoration)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 438, + 304, + 554, + 409 + ], + "blocks": [ + { + "bbox": [ + 438, + 304, + 554, + 409 + ], + "lines": [ + { + "bbox": [ + 438, + 304, + 554, + 409 + ], + "spans": [ + { + "bbox": [ + 438, + 304, + 554, + 409 + ], + "type": "image", + "image_path": "badd37325bea15074e57bfef96b598f343cf59a6673bbcad4f838ff1f9d58085.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 466, + 411, + 526, + 420 + ], + "lines": [ + { + "bbox": [ + 466, + 411, + 526, + 420 + ], + "spans": [ + { + "bbox": [ + 466, + 411, + 526, + 420 + ], + "type": "text", + "content": "(b)EDTR(Ours)" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 422, + 554, + 455 + ], + "lines": [ + { + "bbox": [ + 313, + 422, + 554, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 422, + 554, + 455 + ], + "type": "text", + "content": "Figure 7. Real-world object detection results and visualization of the restored image. The image is from \"56.png\" in the real-world image set RealPhoto60 [61]." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 554, + 534 + ], + "type": "text", + "content": "such as clothing wrinkles on a person, enabling successful detection. These results demonstrate that EDTR can handle complex real-world degradations and detect objects in degraded images by effectively leveraging the powerful diffusion prior. For more visualization results, please refer to Section S12 in our supplementary materials." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 388, + 553 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 713 + ], + "type": "text", + "content": "This paper proposes EDTR, a novel task-driven image restoration (TDIR) method that directly leverages a pretrained diffusion prior to enhance high-level vision task performance. We observe that while incorporating the diffusion prior produces visually pleasing results, restoring task-relevant details remains challenging. To address this, we propose tailored designs for exploiting the diffusion prior in TDIR. Our method significantly improves the performance of various tasks in complex image degradation scenarios while achieving visually appealing results, demonstrating its broad applicability and effectiveness. We anticipate that our approach will provide new insights into leveraging diffusion priors to enhance task performance." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 683, + 295, + 713 + ], + "type": "text", + "content": "1Since the original Tiny-ImageNet [23] has a resolution of 64x64, which is too small for IR problem, we resampled from ImageNet [42] with the same classes and number of images to create Tiny-ImageNet†." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10158" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 140 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 140 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 140 + ], + "type": "text", + "content": "Acknowledgments. This work was supported in part by the IITP grants [No.2021-0-01343, Artificial Intelligence Graduate School Program (Seoul National University), No. 2021-0-02068, and No.2023-0-00156], and the Industrial Technology Alchemist Project [No. RS-2024-00432410] funded by MOTIE, Korea." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 162, + 115, + 174 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 115, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 115, + 174 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 182, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 182, + 294, + 214 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 182, + 294, + 214 + ], + "spans": [ + { + "bbox": [ + 61, + 182, + 294, + 214 + ], + "type": "text", + "content": "[1] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Finding tiny faces in the wild with generative adversarial network. In CVPR, 2018. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 216, + 294, + 248 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 216, + 294, + 248 + ], + "spans": [ + { + "bbox": [ + 61, + 216, + 294, + 248 + ], + "type": "text", + "content": "[2] Yancheng Bai, Yongqiang Zhang, Mingli Ding, and Bernard Ghanem. Sod-mtgan: Small object detection via multi-task generative adversarial network. In ECCV, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 251, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 251, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 61, + 251, + 294, + 293 + ], + "type": "text", + "content": "[3] Hanting Chen, Yunhe Wang, Tianyu Guo, Chang Xu, Yiping Deng, Zhenhua Liu, Siwei Ma, Chunjing Xu, Chao Xu, and Wen Gao. Pre-trained image processing transformer. In CVPR, 2021. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "spans": [ + { + "bbox": [ + 62, + 295, + 294, + 338 + ], + "type": "text", + "content": "[4] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 340, + 294, + 361 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 340, + 294, + 361 + ], + "spans": [ + { + "bbox": [ + 62, + 340, + 294, + 361 + ], + "type": "text", + "content": "[5] Liangyu Chen, Xiaojie Chu, Xiangyu Zhang, and Jian Sun. Simple baselines for image restoration. In ECCV, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 294, + 407 + ], + "type": "text", + "content": "[6] Liang-Chieh Chen, George Papandreou, Iasonas Kokkinos, Kevin Murphy, and Alan L. Yuille. Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolution, and fully connected crfs. TPAMI, 2016. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 294, + 451 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 294, + 451 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 294, + 451 + ], + "type": "text", + "content": "[7] Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. Hat: Hybrid attention transformer for image restoration. In CVPR, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 294, + 485 + ], + "type": "text", + "content": "[8] Dengxin Dai, Yujuan Wang, Yuhua Chen, and Luc Van Gool. Is image super-resolution helpful for other vision tasks? In WACV, 2016. 1, 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 294, + 520 + ], + "type": "text", + "content": "[9] Tao Dai, Jianrui Cai, Yongbing Zhang, Shu-Tao Xia, and Lei Zhang. Second-order attention network for single image super-resolution. In CVPR, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "spans": [ + { + "bbox": [ + 57, + 522, + 294, + 554 + ], + "type": "text", + "content": "[10] Chao Dong, Chen Change Loy, Kaiming He, and Xiaou Tang. Image super-resolution using deep convolutional networks. TPAMI, 2016. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 57, + 556, + 294, + 609 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 294, + 644 + ], + "type": "text", + "content": "[12] Mark Everingham, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The pascal visual object classes (voc) challenge. IJCV, 2010. 5" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 294, + 678 + ], + "type": "text", + "content": "[13] Muhammad Haris, Gregory Shakhnarovich, and Norimichi Ukita. Deep back-projection networks for super-resolution. In CVPR, 2018. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 294, + 712 + ], + "type": "text", + "content": "[14] Muhammad Haris, Shakhnarovich Greg, and Ukita Norimichi. Task-driven super resolution: Object detection in low-resolution images. In ICONIP, 2021. 2, 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 105 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2016. 5" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 107, + 553, + 129 + ], + "type": "text", + "content": "[16] Geoffrey Hinton. Distilling the knowledge in a neural network. In NIPS Workshop, 2014. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 130, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 130, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 130, + 553, + 151 + ], + "type": "text", + "content": "[17] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 153, + 553, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 184 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 184 + ], + "type": "text", + "content": "[18] Shih-Chia Huang, Trung-Hieu Le, and Da-Wei Jaw. Dsnet: Joint semantic learning for object detection in inclement weather conditions. TPAMI, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 186, + 553, + 217 + ], + "type": "text", + "content": "[19] Jiwon Kim, Jung Kown Lee, and Kyoung Mu Lee. Accurate image super-resolution using very deep convolutional networks. In CVPR, 2016. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 253 + ], + "type": "text", + "content": "[20] Jaeha Kim, Junghun Oh, and Kyoung Mu Lee. Beyond image super-resolution for image recognition with task-driven perceptual loss. In CVPR, 2024. 1, 2, 4, 5, 6, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 275 + ], + "type": "text", + "content": "[21] Diederik P Kingma, Max Welling, et al. Auto-encoding variational bayes. In ICLR, 2014. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "spans": [ + { + "bbox": [ + 316, + 277, + 553, + 309 + ], + "type": "text", + "content": "[22] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV workshops, 2013. 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 312, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 312, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 312, + 553, + 342 + ], + "type": "text", + "content": "[23] Yann Le and Xuan Yang. Tiny imagenet visual recognition challenge. https://api_semanticscholar.org/ CorpusID:16664790, 2015.8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 345, + 553, + 399 + ], + "type": "text", + "content": "[24] Christian Ledig, Lucas Theis, Ferenc Huszar, Jose Caballero, Andrew Cunningham, Alejandro Acosta, Andrew Aitken, Alykhan Tejani, Johannes Totz, Zehan Wang, and Wenzhe Shi. Photo-realistic single image super-resolution using a generative adversarial network. In CVPR, 2017. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 400, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 400, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 400, + 553, + 431 + ], + "type": "text", + "content": "[25] Boyi Li, Xiulian Peng, Zhangyang Wang, Jizheng Xu, and Dan Feng. Aod-net: All-in-one dehazing network. In ICCV, 2017. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 477 + ], + "type": "text", + "content": "[26] Chengyang Li, Heng Zhou, Yang Liu, Caidong Yang, Yongqiang Xie, Zhongbo Li, and Liping Zhu. Detection-friendly dehazing: Object detection in real-world hazy scenes. TPAMI, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 479, + 553, + 511 + ], + "type": "text", + "content": "[27] Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. SwinIR: Image restoration using swin transformer. In ICCV, 2021. 2, 3, 5, 6, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 514, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 544 + ], + "type": "text", + "content": "[28] Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Young Mu Lee. Enhanced deep residual networks for single image super-resolution. In CVPR Workshops, 2017. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "spans": [ + { + "bbox": [ + 317, + 546, + 553, + 590 + ], + "type": "text", + "content": "[29] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dolkar, and C Lawrence Zitnick. Microsoft COCO: common objects in context. In ECCV, 2014. 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 317, + 592, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 592, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 592, + 553, + 634 + ], + "type": "text", + "content": "[30] Xinqi Lin, Jingwen He, Ziyan Chen, Zhaoyang Lyu, Bo Dai, Fanghua Yu, Wanli Ouyang, Yu Qiao, and Chao Dong. Diff-bir: Towards blind image restoration with generative diffusion prior. In ECCV, 2024. 1, 2, 3, 4, 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 636, + 553, + 669 + ], + "type": "text", + "content": "[31] Ding Liu, Bihan Wen, Xianming Liu, Zhangyang Wang, and Thomas S Huang. When image denoising meets high-level vision tasks: A deep learning approach. In *IJCAI*, 2018. 1, 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[32] Ding Liu, Bihan Wen, Jianbo Jiao, Xianming Liu, Zhangyang Wang, and Thomas S Huang. Connecting image denoising and high-level vision tasks via deep learning. TIP, 2020. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10159" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[33] Wenyu Liu, Gaofeng Ren, Runsheng Yu, Shi Guo, Jianke Zhu, and Lei Zhang. Image-adaptive yolo for object detection in adverse weather conditions. In AAAI, 2022. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 139 + ], + "type": "text", + "content": "[34] Yuhao Liu, Zhanghan Ke, Fang Liu, Nanxuan Zhao, and Rynson WH Lau. Diff-plugin: Revitalizing details for diffusion-based low-level tasks. In CVPR, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 184 + ], + "type": "text", + "content": "[35] Kangfu Mei, Hossein Talebi, Mojtaba Ardakani, Vishal M Patel, Peyman Milanfar, and Mauricio Delbracio. The power of context: How multimodality improves image superresolution. In CVPR, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 294, + 217 + ], + "type": "text", + "content": "[36] Anish Mittal, Rajiv Soundararajan, and Alan C Bovik. Making a \"completely blind\" image quality analyzer. SPL, 2012. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 251 + ], + "type": "text", + "content": "[37] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, 2021. 3, 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[38] Ben Niu, Weilei Wen, Wenqi Ren, Xiangde Zhang, Lianping Yang, Shuzhen Wang, Kaihao Zhang, Xiaochun Cao, and Haifeng Shen. Single image super-resolution via a holistic attention network. In ECCV, 2020. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 294, + 330 + ], + "type": "text", + "content": "[39] Yanting Pei, Yaping Huang, Qi Zou, Yuhang Lu, and Song Wang. Does haze removal help cnn-based image classification? In ECCV, 2018. 1, 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "spans": [ + { + "bbox": [ + 56, + 331, + 294, + 364 + ], + "type": "text", + "content": "[40] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun. Faster r-cnn: Towards real-time object detection with region proposal networks. In NIPS, 2015. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 294, + 398 + ], + "type": "text", + "content": "[41] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2022. 1, 2, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 454 + ], + "type": "text", + "content": "[42] Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, Sanjeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael Bernstein, Alexander C. Berg, and Li Fei-Fei. ImageNet large scale visual recognition challenge. IJCV, 2015. 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 498 + ], + "type": "text", + "content": "[43] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-cam: Visual explanations from deep networks via gradient-based localization. In ICCV, 2017. 5" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 500, + 294, + 533 + ], + "type": "text", + "content": "[44] Jacob Shermeyer and Adam Van Etten. The effects of superresolution on object detection performance in satellite imagery. In CVPR Workshops, 2019. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 534, + 294, + 566 + ], + "type": "text", + "content": "[45] Taeyoung Son, Juwon Kang, Namyup Kim, Sunghyun Cho, and Suha Kwak. Urie: Universal image enhancement for visual recognition in the wild. In ECCV, 2020. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 567, + 294, + 589 + ], + "type": "text", + "content": "[46] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 612 + ], + "type": "text", + "content": "[47] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In ICML, 2023. 7" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 635 + ], + "type": "text", + "content": "[48] Jong-Chyi Su and Subhransu Maji. Adapting models to signal degradation using distillation. In BMVC, 2017. 5" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 678 + ], + "type": "text", + "content": "[49] Haoze Sun, Wenbo Li, Jianzhuang Liu, Haoyu Chen, Renjing Pei, Xueyi Zou, Youliang Yan, and Yujiu Yang. Coser: Bridging image and language for cognitive super-resolution. In CVPR, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 712 + ], + "type": "text", + "content": "[50] Shangquan Sun, Wenqi Ren, Tao Wang, and Xiaochun Cao. Rethinking image restoration for object detection. In NeurIPS, 2022. 2" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 35, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 117 + ], + "type": "text", + "content": "[51] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The caltech-ucsd birds-200-2011 dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 5, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 150 + ], + "type": "text", + "content": "[52] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 2, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[53] Xintao Wang, Ke Yu, Shixiang Wu, Jinjin Gu, Yihao Liu, Chao Dong, Yu Qiao, and Chen Change Loy. ESRGAN: enhanced super-resolution generative adversarial networks. In ECCV Workshops, 2018. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 239 + ], + "type": "text", + "content": "[54] Yufei Wang, Wenhan Yang, Xinyuan Chen, Yaohui Wang, Lanqing Guo, Lap-Pui Chau, Ziwei Liu, Yu Qiao, Alex C Kot, and Bihan Wen. Sinsr: diffusion-based image superresolution in a single step. In CVPR, 2024. 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 281 + ], + "type": "text", + "content": "[55] Zhendong Wang, Xiaodong Cun, Jianmin Bao, Wengang Zhou, Jianzhuang Liu, and Houqiang Li. Uformer: A general u-shaped transformer for image restoration. In CVPR, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 337 + ], + "type": "text", + "content": "[56] Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Chunyi Li, Liang Liao, Annan Wang, Erli Zhang, Wenxiu Sun, Qiong Yan, Xiongkuo Min, Guangtai Zhai, and Weisi Lin. Q-align: Teaching Imms for visual scoring via discrete text-defined levels. In ICML, 2024. 5" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 370 + ], + "type": "text", + "content": "[57] Jiawei Wu and Zhi Jin. Unsupervised variational translator for bridging image restoration and high-level vision tasks. In ECCV, 2024. 1, 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 403 + ], + "type": "text", + "content": "[58] Rongyuan Wu, Lingchen Sun, Zhiyuan Ma, and Lei Zhang. One-step effective diffusion network for real-world image super-resolution. NeurIPS, 2024. 4" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 437 + ], + "type": "text", + "content": "[59] Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. Seesr: Towards semantics-aware real-world image super-resolution. In CVPR, 2024. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "spans": [ + { + "bbox": [ + 316, + 438, + 553, + 480 + ], + "type": "text", + "content": "[60] Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "spans": [ + { + "bbox": [ + 316, + 482, + 553, + 535 + ], + "type": "text", + "content": "[61] Fanghua Yu, Jinjin Gu, Zheyuan Li, Jinfan Hu, Xiangtao Kong, Xintao Wang, Jingwen He, Yu Qiao, and Chao Dong. Scaling up to excellence: Practicing model scaling for photorealistic image restoration in the wild. In CVPR, 2024. 2, 3, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 536, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 580 + ], + "type": "text", + "content": "[62] Syed Waqas Zamir, Aditya Arora, Salman Khan, Munawar Hayat, Fahad Shahbaz Khan, and Ming-Hsuan Yang. Restormer: Efficient transformer for high-resolution image restoration. In CVPR, 2022. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[63] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 2, 3, 5" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 647 + ], + "type": "text", + "content": "[64] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In CVPR, 2018. 7" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 680 + ], + "type": "text", + "content": "[65] Yulun Zhang, Kunpeng Li, Kai Li, Lichen Wang, Bineng Zhong, and Yun Fu. Image super-resolution using very deep residual channel attention networks. In ECCV, 2018. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 712 + ], + "type": "text", + "content": "[66] Yulun Zhang, Yapeng Tian, Yu Kong, Bineng Zhong, and Yun Fu. Residual dense network for image super-resolution. In CVPR, 2018. 2" + } + ] + } + ], + "index": 34 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10160" + } + ] + } + ], + "index": 36 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 106 + ], + "type": "text", + "content": "[67] Xiaotong Zhao, Wei Li, Yifan Zhang, and Zhiyong Feng. Residual super-resolution single shot network for low-resolution object detection. IEEE Access, 2018. 2, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 140 + ], + "type": "text", + "content": "[68] Shangchen Zhou, Kelvin Chan, Chongyi Li, and Chen Change Loy. Towards robust blind face restoration with codebook lookup transformer. In NeurIPS, 2022." + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 77, + 141, + 82, + 149 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 141, + 82, + 149 + ], + "spans": [ + { + "bbox": [ + 77, + 141, + 82, + 149 + ], + "type": "text", + "content": "5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10161" + } + ] + } + ], + "index": 4 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_content_list.json b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..260bb87f08582de2a93e593f2acaa32e5e143ba7 --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_content_list.json @@ -0,0 +1,1424 @@ +[ + { + "type": "text", + "text": "Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation", + "text_level": 1, + "bbox": [ + 106, + 128, + 893, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Seogkyu Jeon $^{1}$ Kibeom Hong $^{2*}$ Hyeran Byun $^{1*}$", + "bbox": [ + 295, + 202, + 709, + 222 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Yonsei University $^{2}$ Sookmyung Women's University", + "bbox": [ + 269, + 227, + 727, + 247 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "jone9312@yonsei.ac.kr", + "bbox": [ + 403, + 250, + 591, + 263 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 297, + 326, + 314 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent domain generalized semantic segmentation (DGSS) studies have achieved notable improvements by distilling semantic knowledge from Vision-Language Models (VLMs). However, they overlook the semantic misalignment between visual and textual contexts, which arises due to the rigidity of a fixed context prompt learned on a single source domain. To this end, we present a novel domain generalization framework for semantic segmentation, namely Domain-aware Prompt-driven Masked Transformer (DPM-Former). Firstly, we introduce domain-aware prompt learning to facilitate semantic alignment between visual and textual cues. To capture various domain-specific properties with a single source dataset, we propose domain-aware contrastive learning along with the texture perturbation that diversifies the observable domains. Lastly, to establish a framework resilient against diverse environmental changes, we have proposed the domain-robust consistency learning which guides the model to minimize discrepancies of prediction from original and the augmented images. Through experiments and analyses, we demonstrate the superiority of the proposed framework, which establishes a new state-of-the-art on various DGSS benchmarks.", + "bbox": [ + 88, + 332, + 485, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 696, + 220, + 710 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Over the decades, semantic segmentation has made remarkable progress, now being able to precisely classify each pixel in an image into categories. However, one of the shadows that lies in these advancements is that models often exhibit inconsistent and degraded performance when deployed in various real-world environments. To this end, the task of Domain Generalized Semantic Segmentation (DGSS) has arisen to overcome the discrepancy between the training and test domains, i.e., domain shift. DGSS has progressed through diverse approaches, such as feature whiten-", + "bbox": [ + 89, + 722, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4d41545c8df7c1ea7e385bccc12f726a643c721caffa29a89c4052f18bfd71b5.jpg", + "image_caption": [ + "Fixed context prompts [31, 36]" + ], + "image_footnote": [], + "bbox": [ + 524, + 297, + 736, + 424 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ddfc0a3c25bed2213266027f02423d85c1a347d28ca61f056f3a2181431b441d.jpg", + "image_caption": [ + "DPMFormer (Ours)", + "Figure 1. Motivation of DPMFormer. Using a fixed context prompts [31, 36] tend to retain source domain properties, causing contextual misalignment with the target domain. On the other hand, DPMFormer translates domain properties of the input image into context prompts, enhancing semantic alignments." + ], + "image_footnote": [], + "bbox": [ + 738, + 297, + 895, + 424 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "ing [6, 37-39] and domain randomization [4, 14, 19, 24, 27, 28, 56, 59, 61, 62]. Furthermore, some studies [3, 10] exploited multi-scale object queries with a transformer-based architecture, i.e., Mask2Former [5].", + "bbox": [ + 511, + 551, + 906, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Despite these advancements, learning domain-robust representations solely from the single source domain remains a significant hurdle to performance improvement. Recently, several studies [13, 21, 36, 55] employed Vision-Language Models (VLMs) [41, 45] owing to their semantic knowledge learned from diverse large-scale text-image datasets. For this, pioneering VLM-based DGSS works [13, 21, 55] have adopted their pre-trained visual encoder for initialization and fine-tuning [21, 55]. In addition, TQDM [36] has introduced a framework that constructs object queries from the textual descriptions, leveraging semantic concepts from linguistic expressions.", + "bbox": [ + 511, + 613, + 908, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yet, they still have limitations in fully utilizing textual knowledge to improve domain generalizability. Concretely, the context prompts that decorates textual descriptions of each category are either a predefined template [31] (e.g., 'a photo of') or a single learnable text embedding [36]. However, we contend that fixed context prompts have limited capability for target domain images due to the discrepancy", + "bbox": [ + 511, + 795, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "20791", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd2224252574a5e01732501e2ef9a84ce30f390bbe6082512c594b780621a4c4.jpg", + "image_caption": [ + "(a) Visualization of textual prompts" + ], + "image_footnote": [], + "bbox": [ + 133, + 89, + 524, + 277 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/6875ebcbc8727141610ab081774bbaa924c05cf3f83ded44d4d4fb0193976df5.jpg", + "image_caption": [ + "(b) Results on BDD100K [58]", + "Figure 2. PCA visualization of textual prompts (left) and qualitative results on various environments (e.g., Day, Dawn, Night) in BDD100K [58] (right). The models are trained on GTAV [42] with the CLIP-pretrained backbone (ViT-B) [41]. A fixed single-context prompt lacks flexibility in adapting to various domain shifts due to its rigidity. In contrast, our framework utilizes domain-specific properties from input images as context prompts, enhancing semantic alignment between text and images. As a result, as shown in (b), our approach exhibits improved robustness across diverse environments." + ], + "image_footnote": [], + "bbox": [ + 531, + 90, + 864, + 277 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "between visual and textual contexts as depicted in Fig. 1. Firstly, a handcrafted template inherently encodes the characteristics of a specific domain which restricts its generalizability to other unseen domains. Secondly, prompt optimization is prone to source domain overfitting, especially in the single-source setting. Consequently, both approaches may enlarge the gap between textual and visual contexts, leading to suboptimal performances on target domains.", + "bbox": [ + 88, + 401, + 480, + 521 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this perspective, we point out that the semantic form of the textual representation for an object category should be changed with respect to the input visual context in order to strengthen their semantic correspondence. For instance, in the real-world driving scene at night of Fig. 2 (b), target categories (e.g., car and sky) possess distinct textures from those of synthetic daytime images. Hence, it would be more appropriate to decorate 'a car' with the textual prompt 'at night in the real-world' which is reflecting the domain characteristics. As a result, the modified text feature will include details such as darkened car exteriors and light reflections, enhancing the semantic alignment. Nevertheless, it is challenging to design domain-adaptive prompts in the DGSS setting where the training dataset covers only a single domain. Moreover, this setting hinders the visual pipeline from learning robust image features against domain shifts.", + "bbox": [ + 88, + 523, + 482, + 763 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we introduce a novel framework, namely Domain-aware Prompt-driven Masked Transformer (DPMFormer), focusing on two key aspects: (1) Leveraging domain-specific properties of the input image (domain-awareness) and (2) Generating accurate outputs on images with dissimilar domain characteristics (domain-robustness). To cultivate the domain-awareness, we propose a novel domain-aware prompt learning, which translates the domain-specific properties of an input image into con", + "bbox": [ + 88, + 763, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "text prompts via an auxiliary network. In addition, to obtain diverse domain properties in the single-source setting, we apply texture perturbations to synthesize novel domain images. Furthermore, exploiting both source and novel domain images, we propose a domain-aware contrastive loss to ensure that the derived prompts effectively capture domain-specific properties of the input image. This loss encourages the anchor context prompts to be distinguishable from those of different domain characteristics, while being closer to those from the anchor domain. With the proposed domain-aware prompt, the model accurately identifies target classes while being aware of the input domain.", + "bbox": [ + 511, + 401, + 906, + 583 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Moreover, we strive to provide better domain robustness guidances to the visual encoder and decoders. We carefully organize the texture perturbations with structure-preserving image transformations, ensuring that the original visual context remains intact. Instead of simply reusing the original ground truths for novel domain images, we introduce domain-robust consistency loss to guarantee reliable predictions under severe domain shifts. The loss consists of class consistency and mask consistency losses, which penalize discrepancies between class and mask predictions of given image pairs, respectively. Furthermore, domain-robust consistency losses are applied at every layer of the transformer decoder, preventing discrepancies in earlier layers from propagating to later parts.", + "bbox": [ + 511, + 590, + 908, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Through evaluations on various DGSS benchmarks, we demonstrate the superiority of the proposed framework. Notably, our framework achieves state-of-the-art semantic segmentation performance across domain generalization scenarios. Additionally, ablation studies and detailed analyses validate the effectiveness of each component.", + "bbox": [ + 511, + 809, + 908, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "20792", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 89, + 240, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.1. Domain Generalized Semantic Segmentation", + "text_level": 1, + "bbox": [ + 89, + 116, + 470, + 132 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS) aims to learn domain-invariant representations that generalize robustly to various unseen target domains. The task assumes a single-source setting, where only one dataset is available during training. Unlike domain adaptation [8, 49, 66] and test-time domain adaptation [53, 54, 57], access to target domains is strictly prohibited, making DGSS more challenging. Previous studies approached DGSS in two ways primarily: feature whitening and normalization [6, 37-39], and domain randomization approaches [4, 14, 19, 24, 27, 28, 37, 38, 56, 59, 61, 62].", + "bbox": [ + 89, + 138, + 482, + 306 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Feature whitening and normalization approaches [6, 37-39] mainly focus on removing features which are variant to domain shifts. Exploiting the characteristics of instance whitening [30] and instance normalization [50] which can effectively remove texture and style from the input image, these approach leverages those operations in between the backbone module to minimize the effect from domain and texture changes. Representatively, RobustNet [6] introduced instance selective whitening module that finds gram matrix components sensitive to photometric changes and minimize their changes. However, due to the natural difficulty in disentangling domain-specific and domain-invariant features, these approaches show limited performance gains.", + "bbox": [ + 89, + 308, + 482, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Domain randomization approaches [4, 14, 19, 24, 27, 28, 56, 59, 61, 62] augment novel domains from the source domain by modifying either the images or their features through various methods, e.g. affine transformations [14, 20, 28, 56, 61, 62], image translation [59, 65], frequency decomposition [4, 19], and photometric transformations [24]. Synthesized samples increase the domain diversity of the training dataset, reducing the domain gap between the learned representation and the test data. Moreover, most of the generative DGSS methods adopt context-preserving transformations to compute the output discrepancy between the original sample and the augmented one. Representatively, SHADE [61] proposed a style consistency loss to encourage model to learn invariant pixel-level semantic information by minimizing the Jenson-Shannon Divergence (JSD) between the output predictions of original and augmented images.", + "bbox": [ + 89, + 521, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Meanwhile, several studies [3, 10] have built DGSS upon Mask2Former [5], that leverages attention mechanism [51] renowned for robustness against domain shifts [16, 17, 46]. Mask2former [5] involves a transformer decoder that exploits object query features to group pixels of same objects or categories. Based on this, HGFormer [10] first proposed a hierarchical framework that groups pixels to form part-level masks for complementing whole-level pixel group-", + "bbox": [ + 89, + 779, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "ing procedure. Similarly, CMFormer [3] utilizes downsampled features additionally via feature fusion which are more domain-invariant than the original features.", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Language-driven Domain Generalized Semantic Segmentation", + "text_level": 1, + "bbox": [ + 511, + 148, + 903, + 180 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies [13, 21, 22, 36, 55] has investigated to exploit Vision Language Models (VLMs) [9, 23, 33, 40, 41, 45, 47] for DGSS, owing to its powerful generalizability learned from large-scale datasets of image-text pairs. VLTseg [21] employs the image encoder of VLMs for initializing backbone parameters, and fine-tune the network with the segmentation objective function. FAMix [13] proposes to yield class-specific novel styles by concatenating random style description and class names, then performs style randomization by locally mixing the source and the novel styles. DAP [22] exploits text encoder to distill the semantic textual knowledge of target categories to the visual backbone. TQDM [36] points out deficient use of language information in aforementioned works, and introduces a textual-query driven framework for DGSS. They utilize the textual description of each categories for initializing input query features for transformer decoder, and also for computing the text-to-pixel attention to enhance semantic clarity of pixel features. Despite their advancements, we argue that two points are overlooked: (1) the domain discrepancy between the learned textual prompt and the domain-specific property of the input image (2) lack of domain robustness guidance. To mitigate these limitations, we carefully refine the textual queries with the captured domain-specific property of the input image, and augment novel domain images to improve domain-awareness of the model as well as prediction consistency learning.", + "bbox": [ + 511, + 186, + 906, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methods", + "text_level": 1, + "bbox": [ + 511, + 609, + 612, + 626 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 636, + 653, + 651 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We design our DPMFormer based on mask classification architecture, i.e., Mask2Former [5]. Mask2Former consists of an image encoder $ENC_I$ , pixel decoder $DEC_{\\mathrm{pix}}$ and transformer decoder $DEC_{\\mathrm{tr}}$ . The input RGB image $x \\in \\mathbb{R}^{3 \\times H \\times W}$ is first fed to the image encoder for feature extraction, and then converted to pixel-wise features $z \\in \\mathbb{R}^{D \\times H \\times W}$ through the pixel decoder. The transformer decoder iteratively refines $N$ object queries $q \\in \\mathbb{R}^{N \\times L}$ with multi-scale image features from the pixel decoder. Thereafter, refined object queries are projected into mask embeddings $m \\in \\mathbb{R}^{N \\times D}$ to generate pixel-level mask prediction via dot product as $\\hat{y}^{\\mathrm{mask}} = m \\cdot z$ , $\\hat{y}^{\\mathrm{mask}} \\in \\mathbb{R}^{N \\times H \\times W}$ . The category label of each object query $q$ is predicted with a linear classifier, i.e., $c_q \\in \\mathbb{R}^K$ , $c \\in \\mathbb{R}^{N \\times K}$ . The final prediction result is derived by the matrix multiplication between the pixel-level mask prediction and class prediction", + "bbox": [ + 511, + 657, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "20793", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/3f416f43c54e57657e6fdc211311c1c3745cf26c4ef7318bcde770565afa1646.jpg", + "image_caption": [ + "Figure 3. Illustration of DPMFormer. We use Mask2Former [5] based architecture which consists of a backbone image encoder $(ENC_I)$ , a pixel decoder $(DEC_{\\mathrm{pix}})$ , a transformer decoder $(DEC_{\\mathrm{tr}})$ , and a text encoder $(ENC_T)$ . During training, we synthesize images with a novel domain style via texture perturbation. Both images are incorporated to compose a batch and exploited for learning domain-awareness (Sec. 3.2.1) and domain-robustness (Sec. 3.2.2)." + ], + "image_footnote": [], + "bbox": [ + 93, + 85, + 911, + 314 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of queries as $\\hat{y} = c^{\\top}\\cdot \\hat{y}^{\\mathrm{mask}},\\hat{y}\\in \\mathbb{R}^{K\\times H\\times W}$", + "bbox": [ + 89, + 406, + 387, + 424 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Furthermore, inspired by [36], we employ a text encoder $ENC_{T}$ for initializing object queries from textual descriptions. Pre-trained vision language mdoels (VLMs), e.g., CLIP [41], is exploited to initialize both $ENC_{I}$ and $ENC_{T}$ parameters, and the parameters of $ENC_{T}$ remain frozen during training to preserve learned linguistic knowledge. The text encoder $ENC_{T}$ produces a text feature $t_k \\in \\mathbb{R}^C$ corresponding to each class label $\\mathrm{class}_k$ with a learnable context prompt $p$ , i.e., $t_k = ENC_{T}([p, \\{\\mathrm{class}_{k}\\}])$ . Then, obtained text embeddings $t = \\{t_k\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times C}$ are passed to a multi-layer perceptron (MLP) to derive initial textual object queries $q_t^0$ . Also, the text-to-pixel attention mechanism is established in a pixel decoder layer to enhance pixel semantic clarity enhancement.", + "bbox": [ + 88, + 424, + 483, + 636 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. DPMFormer", + "text_level": 1, + "bbox": [ + 89, + 648, + 227, + 664 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 3, Domain-aware Prompt-driven Masked Transformer (DPMFormer), aims to cultivate two core aspects of domain generalization for the model: domain-awareness and domain-robustness. For domain-awareness, we propose the domain-aware context prompt learning which translates domain-specific property of the input image into the text embedding for semantic alignment between textual and visual features. And for domain-robustness, we encourage the model to generate accurate outputs against textural changes via consistency learning.", + "bbox": [ + 89, + 672, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Texture perturbation. In order to effectively guide the model with diverse domain characteristics, we stylize the training dataset to generate an auxiliary domain dataset. Following RobustNet [6], we adopt photometric transformations—comprising strong color jittering, gaussian blur", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "and noise injection—for their simplicity as well as content-preserving property. The generated image $x'$ is combined with its original image $x$ to form a batch for training.", + "bbox": [ + 511, + 409, + 906, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2.1. Domain-aware context prompt learning", + "text_level": 1, + "bbox": [ + 511, + 459, + 836, + 474 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Textual query generation and prompt learning [64] plays a crucial role in enabling the model to leverage the linguistic semantics of each class previously learned by the VLM. Although the learned context prompt is beneficial for performance improvement, it is optimized solely within the source domain without direct consideration of domain shifts. This single context prompt may yield strong results on several domain shift scenarios where target domain has similar domain characteristics with the source. However, in cases of severe domain shift, the performance gain may be limited due to the contextual mismatch between the target domain image and the learned prompt. For example, if the prompt learning is conducted on a dataset containing only sunny day images, the learned prompt is could be misaligned when encountering rainy night scenes. As such, the semantic misalignment should be addressed in order to fully utilize rich semantic knowledge of pre-trained VLM. To this end, we propose to generate domain-aware context prompt that extracts domain-specific properties from the input image as a context prompt.", + "bbox": [ + 511, + 477, + 906, + 779 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To obtain domain-aware context prompt from the input image $x$ , we design an auxiliary network $h_{\\theta}(\\cdot)$ that takes visual feature as an input and generates a domain-specific prompt embedding $\\pi_x = h_\\theta (\\hat{F} (x))$ , where $\\hat{F} (x)$ denotes a visual feature extracted from a frozen visual backbone of CLIP [41]. We use the class token as a visual feature $\\hat{F} (x)$ for its global representation [11]. Nextly, the obtained domain-specific prompt embedding $\\pi_{x}$ is in", + "bbox": [ + 511, + 780, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "20794", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tegrated with the context prompt embedding $p$ through addition, i.e., $p_x = p + \\pi_x$ , then concatenated with text embeddings of classes to generate domain-aware textual features as $t_{x,k} = ENC_T([p_x,\\{\\mathrm{class}_k\\}])$ . To encourage derived textual features to include domain-specific information of the input image, we introduce a novel domain-aware contrastive learning framework with the original and augmented images. The loss function is depicted as follows:", + "bbox": [ + 89, + 90, + 483, + 212 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {c o n t r a}} = - \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} \\log \\frac {\\sum_ {j \\in \\mathcal {P} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}{\\sum_ {j \\in \\mathcal {P} _ {i} \\cup \\mathcal {N} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 101, + 237, + 482, + 291 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{sim}(\\cdot)$ means similarity metric and $\\tau$ is a temperature parameter. $B$ denotes the batch size of original images, $\\mathcal{P}_i$ and $\\mathcal{N}_i$ are positive sets and negative sets of the $i$ -th image, respectively. The positive set $\\mathcal{P}_i$ is composed of the indices of samples having same domain characteristics with the anchor $i$ whereas the others belong to negative set $\\mathcal{N}_i$ . For example, when an anchor $i$ is of original source domain images, other original source domain images belongs to $\\mathcal{P}_i$ while all augmented images are included in $\\mathcal{N}_i$ . In case of an augmented image as an anchor, $\\mathcal{P}_i = \\{i\\}$ and $\\mathcal{N}_i = \\{1,\\dots,2B\\} \\setminus i$ . With the proposed loss, we encourage $h_\\theta$ to capture domain-specific property of the image and reflect it to the output text feature $t_k$ . Also, the final domain-aware text feature is guided by the task loss $\\mathcal{L}_{seg}$ to be aligned with the original image and to minimize segmentation errors. We note that the proposed loss considers the domain information unlike CoCoOp [63] which treats $\\mathcal{P}_i$ as the text feature corresponding to a specific image feature while the remaining images forming the negative set. We provide comparative analysis with CoCoOp in Sec.4.4.2.", + "bbox": [ + 91, + 291, + 483, + 593 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.2.2. Domain-robust consistency learning", + "text_level": 1, + "bbox": [ + 89, + 603, + 387, + 618 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To further enhance the domain robustness, we encourage model to generate persistent predictions in the domain shift scenario. To enhance prediction consistency, we induce the model to minimize the prediction discrepancy in terms of the mask and class label as follows:", + "bbox": [ + 89, + 622, + 483, + 696 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c o n s}} = \\sum_ {s = 1} ^ {S} \\lambda_ {\\mathrm {m c}} \\cdot \\mathcal {L} _ {\\mathrm {m c}} \\left(\\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}, \\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}\\right) + \\lambda_ {\\mathrm {c c}} \\cdot \\mathcal {L} _ {\\mathrm {c c}} \\left(\\hat {c} _ {q _ {i}, \\mathrm {s}}, \\hat {c} _ {q _ {i} ^ {\\prime}, \\mathrm {s}}\\right). \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 708, + 483, + 762 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "$S$ denotes the number of transformer blocks in the transformer decoder, and $\\mathcal{L}_{\\mathrm{mc}}$ and $\\mathcal{L}_{\\mathrm{cc}}$ represent the mask and class consistency losses, respectively. To compute the loss at the $s$ -th transformer decoder block, we obtain $\\{\\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}, \\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}\\}$ and $\\{\\hat{c}_{q_i,s}, \\hat{c}_{q_i',s}\\}$ which are pairs of mask predictions and class predictions of $i$ -th query $q_i$ from the original and augmented image pair $\\{x, x'\\}$ , respectively. We employ binary cross entropy and Jensen-Shannon divergence as a discrepancy measure of the mask consistency", + "bbox": [ + 89, + 762, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "loss $(\\mathcal{L}_{mc})$ and class consistency loss $(\\mathcal{L}_{cc})$ , respectively. With the help of domain-aware context prompts and above losses, our model learns to predict not only accurately but also consistently in various domain shift scenarios during training.", + "bbox": [ + 511, + 90, + 905, + 167 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Overall Loss Functions", + "text_level": 1, + "bbox": [ + 511, + 172, + 728, + 186 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The overall loss of our framework is a weighted sum of the task loss $\\mathcal{L}_{\\mathrm{seg}}$ , VLM regularization loss $\\mathcal{L}_{\\mathrm{reg}}$ , domain-aware contrastive loss $\\mathcal{L}_{\\mathrm{contra}}$ , and the consistency loss $\\mathcal{L}_{\\mathrm{cons}}$ .", + "bbox": [ + 511, + 195, + 905, + 242 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c o n t r a}} \\mathcal {L} _ {\\text {c o n t r a}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 247, + 906, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}$ are constant weighting factors. We note that $\\mathcal{L}_{\\mathrm{seg}}$ and $\\mathcal{L}_{\\mathrm{cons}}$ are calculated with all queries and its predictions from every block of the transformer decoder. We provide details of the baseline losses $(\\mathcal{L}_{\\mathrm{seg}},\\mathcal{L}_{\\mathrm{reg}})$ in the supplementary.", + "bbox": [ + 511, + 268, + 905, + 345 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 356, + 645, + 372 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 380, + 730, + 395 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. We validate DPMFormer on synthetic-to-real and real-to-real scenarios in the single-source setting.", + "bbox": [ + 511, + 402, + 905, + 431 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Synthetic datasets. GTAV [42] is a representative synthetic dataset which consists of 24,966 images with a resolution of $1914 \\times 1052$ . The training split contains 12,403 images, while the validation and test set are of 6,382 and 6,181 images, respectively. SYNTHIA [43] dataset provides 6,580 images for training and 2,820 images for validation respectively, with the image resolution at $1280 \\times 760$ .", + "bbox": [ + 511, + 431, + 905, + 537 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Real-world datasets. Cityscapes [7] is a dataset collected from the real environment. The resolution of each image is $2048 \\times 1024$ , and the population of training and validation split is 2,975 and 500, respectively. BDD100K [58] includes 7,000 training images and 1,000 validation images with the resolution of $1280 \\times 720$ . Mapillary [34] is composed of images with diverse resolution, where the training and validation size is 18,000 and 2,000, respectively.", + "bbox": [ + 511, + 537, + 905, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Network architecture. Our approach leverages the vision transformer-based models as backbones, initialized with either the CLIP [41] (ViT-B) or EVA02-CLIP [45] (EVA02-L) model. The CLIP backbone is configured with a patch size of 16, while the EVA02-CLIP backbone uses a patch size of 14. For the pixel and transformer decoder, aforementioned, we employ a mask classification architecture [5], consisting of $N = 9$ layers with masked attention mechanisms. Additionally, we design the auxiliary network $h_{\\theta}$ for domain-aware context prompt generation as a shallow multi-layer structure (BatchNorm-Linear-ReLU-Linear).", + "bbox": [ + 511, + 659, + 905, + 824 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training and evaluation. To optimize DPMFormer, we employ an AdamW [32] where the learning rate is set as $1 \\times 10^{-5}$ and $1 \\times 10^{-4}$ for synthetic and real training datasets, respectively. Following previous transformer-based studies [18, 36], we apply linear warm-up [15] for", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "20795", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/c770aa7ca80db43f06de4e70ddd3b4b7fbe26f6451586baee7200147ef582a8c.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Models (GTAV)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10145.3341.1840.7742.43
WildNet [28]ResNet-10145.7941.7347.0844.87
SHADE [61]ResNet-10146.6643.6645.5045.27
TLDR [27]ResNet-10147.5844.8848.8047.09
FAMix* [13]ResNet-10149.4746.4051.9749.28
SHADE [61]MiT-B553.2748.1954.9952.15
IBAFormer [44]MiT-B556.3449.7658.2654.79
VLTSeg* [21]ViT-B47.5045.7054.3049.17
TQDM* [36]ViT-B57.5047.6659.7654.97
DPMFormer* (ours)ViT-B59.0051.8063.6258.14
VLTSeg** [21]EVA02-L65.6058.4066.5063.50
Rein** [55]EVA02-L65.3060.5064.9063.60
Rein†ViT-L66.4060.4066.1064.30
TQDM** [36]EVA02-L68.8859.1870.1066.05
DPMFormer** (ours)EVA02-L70.0860.4870.6667.07
", + "bbox": [ + 94, + 88, + 478, + 273 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8b1f81c2c2731508663b6303b86d01dfb15f6d5617f5132d922d3bda5f6a273d.jpg", + "table_caption": [ + "Table 1. Comparison with the state-of-the-art DGSS methods on synthetic-to-real scenario with GTAV [42] as a source. We note that *, **, and † denotes the models initialized with pretrained CLIP [41], EVA02-CLIP [45], and DINOv2 [35], respectively. The best and the second best performances are highlighted with bold and underline, respectively." + ], + "table_footnote": [], + "table_body": "
Models (Synthia)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10140.8735.9837.2638.04
TLDR [27]ResNet-10142.6035.4637.4638.51
IBAFormer [44]MiT-B550.9244.6650.5848.72
VLTSeg** [21]EVA02-L56.8050.5054.5053.93
TQDM** [36]EVA02-L57.9952.4354.8755.10
DPMFormer** (ours)EVA02-L58.9254.3960.0857.80
", + "bbox": [ + 94, + 383, + 478, + 468 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "initial 1,500 iterations and rare class sampling [18]. The optimizer settings are identical for both CLIP and EVA02-CLIP backbone models. We set the total training iterations and the batch size as 20,000 and 8, respectively. Weighting factors $\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}$ are set as 1, 1, 10. We crop the input image to have a resolution of $512\\times 512$ . The texture perturbation is only applied during training. We use mean Intersection over Union (mIoU) [12] to quantitatively evaluate the results following the convention.", + "bbox": [ + 88, + 534, + 482, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Quantitative Results", + "text_level": 1, + "bbox": [ + 89, + 681, + 284, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Synthetic-to-real. As shown in Tab. 1, we achieve state-of-the-art in every target domain with both backbones. Notably, with the ViT-B backbone initialized with pretrained CLIP [41], DPMFormer consistently surpasses previous state-of-the-art [36] by $3.17\\%$ on average mIoU among target domains. In detail, on Cityscapes which contains mostly daytime real images, the domain robustness empowered by consistency learning assists DPMFormer to cope with the domain gap caused by visual realism, improving the state-of-the-art performance by $1.5\\%$ . Meanwhile on BDD100K and Mapillary which have higher environmental variation in terms of weather, time and location, the domain-aware prompt generation enables the model to adaptively leverage", + "bbox": [ + 89, + 703, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b1d97389faca534b3600cea308a80115e5f378469b3ee5ce8f8f5743c4011ce6.jpg", + "table_caption": [ + "Table 2. Comparison with the state-of-the-art DGSS methods where SYNTHIA [43] is set as a source dataset." + ], + "table_footnote": [], + "table_body": "
Models (Cityscapes)BackboneBDDMapillaryAvg.
SAN-SAW [39]ResNet-10154.7361.2742.43
WildNet [28]ResNet-10147.0141.7344.87
SHADE [61]ResNet-10150.9543.6645.27
TQDM* [36]ViT-B50.5465.7458.14
DPMFormer* (ours)ViT-B54.8167.7261.27
HGFormer [10]EVA02-L61.5072.1066.80
VLTSeg** [21]EVA02-L64.4076.4070.40
Rein** [55]EVA02-L64.1069.5066.80
Rein†ViT-L65.0072.3068.65
TQDM [36]**EVA02-L64.7276.1570.44
DPMFormer** (ours)EVA02-L64.276.6770.44
", + "bbox": [ + 517, + 88, + 903, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 3. Comparison with the state-of-the-arts trained with Cityscapes [7] on the real-to-real scenario.", + "bbox": [ + 511, + 267, + 903, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "the textual knowledge for segmentation, impressively escalating the performance by $4.14\\%$ , and $3.17\\%$ , respectively. Moreover, even with the larger backbone [45], we mark the highest average performance of $67.07\\%$ . In Cityscapes and Mapillary dataset, we outperform TQDM [36] by $1.2\\%$ and $0.56\\%$ respectively, while nearly reaching the score of Rein [55] in BDD. The overall results demonstrate the superiority of DPMFormer, emphasizing the efficacy of domain-awareness as well as domain-robustness.", + "bbox": [ + 511, + 321, + 906, + 458 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Tab. 2, we also present synthetic-to-real results where SYNTHIA [43] is set as a source domain. Again, DPM-Former outperforms all competitors by a large margin, achieving the state-of-the-art performance in every target domain. Remarkably, our framework surpasses the previous state-of-the-art [36] by an average of $2.7\\%$ mIoU across target domains. In particular, the performance on Mapillary [34] improves significantly by $5.21\\%$ . These results demonstrate the effectiveness of DPM-Former in addressing the domain gap in terms of texture and perspective.", + "bbox": [ + 511, + 459, + 908, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Real-to-real. Furthermore, as shown in Tab.3, DPMFormer records the highest average mIoU with both backbones. With the CLIP-pretrained backbone, we significantly boost the state-of-the-art performance by $4.27\\%$ and $1.98\\%$ on BDD and Mapillary, respectively. These results indicate that enhanced domain robustness supports the model in predicting consistently under environmental changes, while the domain-aware prompts facilitate the semantic alignment on the unseen domain. As a result, DPMFormer successfully achieves an average performance gain of $3.13\\%$ over the previous state-of-the-art. Equipped with the EVA-CLIP [45] pretrained backbone, we attain the highest mIoU of $76.67\\%$ on Mapillary, while performing slightly lower on BDD. Overall, the average performance is $70.44\\%$ , which is comparable to TQDM [36].", + "bbox": [ + 511, + 611, + 908, + 838 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Qualitative Results", + "text_level": 1, + "bbox": [ + 511, + 847, + 697, + 864 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In Fig. 4, we qualitatively compare our method with FAMix [13] and TQDM [36] on the synthetic-to-real sce", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "20796", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/87cd7074766a0a9379133f1848d6d5a51b55df0de2bce5ca0d3133ae9c034bf1.jpg", + "image_caption": [ + "Figure 4. Qualitative comparison on synthetic-to-real scenario with the CLIP-pretrained backbone (ViT-B). The training source domain is set as GTAV [42] while the target domains are BDD100K [58] and Mapillary [34]. The overall result shows that DPMFormer accomplishes precise segmentation with the images of strong illumination contrast as well as confusing textures." + ], + "image_footnote": [], + "bbox": [ + 112, + 85, + 890, + 369 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "nario with the pretrained CLIP backbone (ViT-B). As depicted in the results on BDD100K, both competitors struggle to discriminate accurately under environments having confusing textures or large variations of illumination. In particular, the road in the first-row image has a texture similar to the exterior of a car, leading TQDM to a mismatch between the visual and textual features. In addition, FAMix shows sensitivity to the shades, mislabeling the road as a sidewalk. On the contrary, our method effectively distinguishes the road from other categories by reflecting the domain-specific properties of real-world roads in clear weather. In the case of the second row image with the high illumination contrast, both FAMix and TQDM fail to notice sidewalks in the image due to their low brightness. On the other hand, DPMFormer carries out precise predictions under severe photometric changes owing to the domain robustness acquired from handling diverse texture changes.", + "bbox": [ + 88, + 449, + 485, + 708 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "As observed with the samples from the Mapillary (the third and fourth rows), FAMix and TQDM lack discriminability between the road and the sidewalk. To be specific with the first image (third row), these classes appear to have a similar color and texture, resulting in misclassifications. Meanwhile, the domain-specific properties from the image transfer the semantic knowledge of the appearance of real-world sidewalk to the model, DPMFormer accurately distinguish between the road and the sidewalk. With the second image, TQDM [36] shows vulnerability against small textural changes on the sidewalk (e.g., shades of the poles). Contrarily, DPMformer effectively performs segmentation ow", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ing to its domain-robustness as well as the obtained domain-aware textual queries.", + "bbox": [ + 511, + 449, + 906, + 479 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Analysis", + "text_level": 1, + "bbox": [ + 511, + 487, + 617, + 503 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.1. Ablation studies", + "text_level": 1, + "bbox": [ + 511, + 508, + 671, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To demonstrate the effectiveness of the components in DPMFormer, i.e., texture perturbation, domain-robust consistency learning $(\\mathcal{L}_{cons})$ , and domain-aware context prompt learning $(\\mathcal{L}_{contra})$ , we conduct ablation studies with the CLIP pretrained backbone on synthetic-to-real scenario. As presented in Tab. 4, every component contributes to the performance gain. Specifically, the texture perturbation enlarges the observable domain during training, boosts the average performance by $0.65\\%$ . In addition with the $\\mathcal{L}_{cons}$ , the model successfully equips the robustness against domain-shift and enhances the prediction accuracy on unseen target domains, increasing the average mIoU by $0.77\\%$ . Furthermore, $\\mathcal{L}_{contra}$ allows the model to acquire domain-specific property from the input image in the form of textual prompt embedding. Consequently, the model exploits both visual and textual cues properly aligned for the target domain, considerably elevate the average mIoU by $1.99\\%$ . Combining all components together, DPMFormer effectively learns both domain-awareness and domain-robustness to generate accurate results consistently on unseen target domains.", + "bbox": [ + 511, + 527, + 908, + 830 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4.2. Comparison with prompt learning methods", + "text_level": 1, + "bbox": [ + 511, + 835, + 862, + 852 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In Tab. 5, we compare our domain-aware context prompt learning $(\\mathcal{L}_{\\mathrm{contra}})$ with previous prompt learning methods [25, 26, 63] to validate the efficacy for domain gen", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "20797", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e6654b6caca6741142c6ab05d2b59a2c96c8f63a543388e6f73a4383a98bf22.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Models (GTAV)CityscapesBDDMapillaryAvg.
Baseline57.547.6659.7654.97
+ Perturbation57.0448.1960.9155.38
+ Lcons58.2249.3960.8456.15
+ Lcontra59.051.863.6258.14
", + "bbox": [ + 91, + 88, + 482, + 172 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/81442a0034263349e47936b3c2e9035b557ce3b03bd18944350598372fe43f34.jpg", + "image_caption": [ + "Figure 5. Qualitative results on diverse styles, i.e., Minimalist, Pop Art, Bauhaus, and Cubism. The models are trained on GTAV with CLIP pretrained backbone (ViT-B)." + ], + "image_footnote": [], + "bbox": [ + 93, + 257, + 483, + 393 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "deralization. Given the textual features $t = ENC_{T}(p_{x})$ and visual features $v = \\hat{F}(x)$ , the contrastive loss of CoCoOp [63] encourages the similarity between the anchor text feature and its corresponding visual feature to be higher than others. As observed in the second row, CoCoOp marginally improves the performance of prompt generator $h_{\\theta}$ , indicating that the generated feature becomes instancespecific which is less generalizable to other domains. Moreover, additionally assigning the visual feature from augmented sample as positives $(\\mathrm{CoCoOp}^{+})$ resulted in negligible performance gain, since the derived prompt contains domain-invariant information rather than domain-specific properties. Furthermore, although both MaPLe [25] and PromptSRC [26] employ multi-modal prompts, their performance gains are modest due to the limited generalizability of fixed prompts. On the other hand, our domain-aware contrastive loss design enable the model to proficiently capture domain-specific property from the image which are more helpful for the semantic alignment between the visual features and the textual knowledge. From the fifth to the eighth row, we verify $\\mathcal{L}_{\\text{contra}}$ with different similarity calculation targets for $\\sin(\\cdot)$ , i.e., text features $t$ , text and visual features $(t, v)$ , and (3) output context embeddings $\\pi$ . The results confirm that computing the domain-aware contrastive loss empowers the domain generalizability especially when computed with the context embeddings $\\pi$ which can provide direct domain guidance to $h_{\\theta}$ .", + "bbox": [ + 88, + 492, + 486, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/2dfc67f6bb833a13710922897e632a3fe0702ec45951f6e1f2d3890732b6109d.jpg", + "table_caption": [ + "Table 4. Ablation study of proposed components. The models are trained on GTAV with CLIP pretrained backbone (ViT-B). The best performance for each column is highlighted in bold." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{contra}} \\) (GTAV)CityscapesBDDMapillaryAvg.
-57.6549.6361.1056.13
CoCoOp [63]57.8449.9161.3356.36
\\( \\mathrm{CoCoOp}^{+} \\)57.6050.0460.9656.20
MaPLe [25]57.8750.1261.0456.34
PromptSRC [26]58.1049.7362.5156.78
Ours (sim(t,t))58.1750.1861.3756.57
Ours (sim(t,v))59.4950.3462.1857.34
Ours (sim(π,π))59.0051.8063.6258.14
", + "bbox": [ + 514, + 88, + 905, + 234 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 5. Comparison with prompt learning methods [25, 26, 63] in synthetic-to-real scenario. - denotes DPMFormer without $\\mathcal{L}_{\\mathrm{contra}}$ . $\\mathrm{CoCoOp}^{+}$ indicates the modified loss that additionally include the feature from augmented sample of anchor as positives. $t$ , $v$ , and $\\pi$ refers to the text feature, visual feature, and context embedding.", + "bbox": [ + 511, + 243, + 906, + 315 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4.3. Qualitative results on diverse styles", + "text_level": 1, + "bbox": [ + 511, + 338, + 805, + 354 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To further verify the domain generalization capability of DPMFormer, we compare our model with TQDM [36] on images with diverse artistic styles generated by $\\mathrm{ChatGPT}^1$ . As shown in Fig. 5, DPMFormer correctly predicts objects and their surroundings even in severe domain shift scenarios, e.g., Cubism. With modern art styles (i.e., Minimalist, Pop Art, and Bauhaus), TQDM perplexes among road, sidewalk, and terrain due to their textural changes. Notably, DPMFormer produces more reliable results owing to the enhanced robustness against texture variations that are learned from the texture perturbation and consistency learning. In case of the scene in cubism style, TQDM mispredicts the person on the left side because of its visual similarity with the person in the traffic sign. On the other hand, our method reflects the cubism style to the textual object queries, making an accurate classification with the same instance.", + "bbox": [ + 511, + 357, + 908, + 599 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 611, + 633, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we presented DPMFormer, a novel domain generalization framework for semantic segmentation. To address the limited generalizability of fixed context prompt learned from a single source dataset, we introduced a novel domain-aware prompt learning which can reflect domain-specific properties of the input image into textual prompts. The proposed component enhanced the semantic alignment between visual and textual cues, assisting the model to fully leverage the abundant semantic knowledge of VLMs. Moreover, to empower domain-robustness, we simulated various domain shifts via texture perturbations, and provided consistency guidance to the model by minimizing prediction discrepancies between original and augmented images. Through extensive experiments and analyses, we demonstrated the effectiveness of DPMFormer, achieving state-of-the-art performance on various benchmarks.", + "bbox": [ + 511, + 637, + 908, + 878 + ], + "page_idx": 7 + }, + { + "type": "page_footnote", + "text": "1 https://chat.openai.com", + "bbox": [ + 531, + 886, + 665, + 901 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "20798", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments. This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364). This research was supported by Sookmyung Women's University Research Grants (1-2403-2034).", + "bbox": [ + 89, + 90, + 485, + 167 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 181, + 187, + 196 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2", + "[2] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 1, 2", + "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 819-827, 2024. 1, 3", + "[4] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19288-19300, 2023. 1, 3", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 3, 4, 5", + "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 1, 3, 4", + "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 1, 3, 7, 8", + "[8] Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. Domain adaptation in computer vision applications, pages 1-35, 2017. 3", + "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3", + "[10] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15413-15423, 2023. 1, 3, 6", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + ], + "bbox": [ + 93, + 205, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4", + "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 6", + "[13] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23428-23437, 2024. 1, 3, 6", + "[14] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023. 1, 3", + "[15] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5", + "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3, 1", + "[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 3", + "[18] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 5, 6", + "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 1, 3", + "[20] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 3", + "[21] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. arXiv preprint arXiv:2312.02021, 2023. 1, 3, 6", + "[22] Xinyue Huo, Lingxi Xie, Hengtong Hu, Wengang Zhou, Houqiang Li, and Qi Tian. Domain-agnostic priors for semantic segmentation under unsupervised domain adaptation and domain generalization. International Journal of Computer Vision, 132(9):3954-3976, 2024. 3", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "20799", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "conference on machine learning, pages 4904-4916. PMLR, 2021.3", + "[24] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18993-19003, 2023. 1, 3", + "[25] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113-19122, 2023. 7, 8", + "[26] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15190-15200, 2023. 7, 8", + "[27] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 677-687, 2023. 1, 3, 6", + "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9936-9946, 2022. 1, 3, 6", + "[29] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5543-5551, 2017. 1", + "[30] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3", + "[31] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 1", + "[32] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5", + "[33] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3", + "[34] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 5, 6, 7, 3, 8", + "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. European Conference on Computer Vision, 2024. 1, 3, 4, 5, 6, 7, 8, 2", + "[37] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 3", + "[38] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1863-1871, 2019. 3", + "[39] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2594-2605, 2022. 1, 3, 6", + "[40] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for zero-shot transfer learning. Neurocomputing, 555: 126658, 2023. 3", + "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6", + "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 2, 5, 6, 7, 3, 4", + "[43] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 5, 6", + "[44] Qiyu Sun, Huilin Chen, Meng Zheng, Ziyan Wu, Michael Felsberg, and Yang Tang. Ibaformer: Intra-batch attention transformer for domain generalized semantic segmentation. arXiv preprint arXiv:2309.06282, 2023. 6", + "[45] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 3, 5, 6", + "[46] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3", + "[47] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3", + "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 1", + "[49] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker." + ], + "bbox": [ + 516, + 90, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "20800", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 3", + "[50] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 3", + "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3", + "[52] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 1, 2", + "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In The Nineth International Conference on Learning Representations (ICLR 2021). OpenReview, 2021. 3", + "[54] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3", + "[55] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28619-28630, 2024. 1, 3, 6", + "[56] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 1, 3", + "[57] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8978-8987, 2021. 3", + "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 2, 5, 7, 3, 4, 6", + "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 1, 3", + "[60] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2" + ], + "bbox": [ + 91, + 90, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[61] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In European conference on computer vision, pages 535-552. Springer, 2022. 1, 3, 6", + "[62] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. Advances in neural information processing systems, 35:338-350, 2022. 1, 3", + "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 5, 7, 8, 2", + "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4", + "[65] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3", + "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + ], + "bbox": [ + 516, + 92, + 903, + 484 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "20801", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_model.json b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_model.json new file mode 100644 index 0000000000000000000000000000000000000000..8652cbccc1d2107294a2313fd3dcbc9f04dae49b --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_model.json @@ -0,0 +1,2180 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.107, + 0.13, + 0.895, + 0.178 + ], + "angle": 0, + "content": "Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.296, + 0.203, + 0.71, + 0.223 + ], + "angle": 0, + "content": "Seogkyu Jeon\\(^{1}\\) Kibeom Hong\\(^{2*}\\) Hyeran Byun\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.271, + 0.228, + 0.728, + 0.248 + ], + "angle": 0, + "content": "\\(^{1}\\)Yonsei University \\(^{2}\\)Sookmyung Women's University" + }, + { + "type": "text", + "bbox": [ + 0.404, + 0.25, + 0.593, + 0.264 + ], + "angle": 0, + "content": "jone9312@yonsei.ac.kr" + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.299, + 0.327, + 0.315 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.333, + 0.486, + 0.665 + ], + "angle": 0, + "content": "Recent domain generalized semantic segmentation (DGSS) studies have achieved notable improvements by distilling semantic knowledge from Vision-Language Models (VLMs). However, they overlook the semantic misalignment between visual and textual contexts, which arises due to the rigidity of a fixed context prompt learned on a single source domain. To this end, we present a novel domain generalization framework for semantic segmentation, namely Domain-aware Prompt-driven Masked Transformer (DPM-Former). Firstly, we introduce domain-aware prompt learning to facilitate semantic alignment between visual and textual cues. To capture various domain-specific properties with a single source dataset, we propose domain-aware contrastive learning along with the texture perturbation that diversifies the observable domains. Lastly, to establish a framework resilient against diverse environmental changes, we have proposed the domain-robust consistency learning which guides the model to minimize discrepancies of prediction from original and the augmented images. Through experiments and analyses, we demonstrate the superiority of the proposed framework, which establishes a new state-of-the-art on various DGSS benchmarks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.697, + 0.222, + 0.712 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.723, + 0.485, + 0.875 + ], + "angle": 0, + "content": "Over the decades, semantic segmentation has made remarkable progress, now being able to precisely classify each pixel in an image into categories. However, one of the shadows that lies in these advancements is that models often exhibit inconsistent and degraded performance when deployed in various real-world environments. To this end, the task of Domain Generalized Semantic Segmentation (DGSS) has arisen to overcome the discrepancy between the training and test domains, i.e., domain shift. DGSS has progressed through diverse approaches, such as feature whiten-" + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.298, + 0.737, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.578, + 0.433, + 0.704, + 0.444 + ], + "angle": 0, + "content": "Fixed context prompts [31, 36]" + }, + { + "type": "image", + "bbox": [ + 0.74, + 0.298, + 0.897, + 0.425 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.774, + 0.433, + 0.855, + 0.443 + ], + "angle": 0, + "content": "DPMFormer (Ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.457, + 0.908, + 0.527 + ], + "angle": 0, + "content": "Figure 1. Motivation of DPMFormer. Using a fixed context prompts [31, 36] tend to retain source domain properties, causing contextual misalignment with the target domain. On the other hand, DPMFormer translates domain properties of the input image into context prompts, enhancing semantic alignments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.614 + ], + "angle": 0, + "content": "ing [6, 37-39] and domain randomization [4, 14, 19, 24, 27, 28, 56, 59, 61, 62]. Furthermore, some studies [3, 10] exploited multi-scale object queries with a transformer-based architecture, i.e., Mask2Former [5]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.795 + ], + "angle": 0, + "content": "Despite these advancements, learning domain-robust representations solely from the single source domain remains a significant hurdle to performance improvement. Recently, several studies [13, 21, 36, 55] employed Vision-Language Models (VLMs) [41, 45] owing to their semantic knowledge learned from diverse large-scale text-image datasets. For this, pioneering VLM-based DGSS works [13, 21, 55] have adopted their pre-trained visual encoder for initialization and fine-tuning [21, 55]. In addition, TQDM [36] has introduced a framework that constructs object queries from the textual descriptions, leveraging semantic concepts from linguistic expressions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Yet, they still have limitations in fully utilizing textual knowledge to improve domain generalizability. Concretely, the context prompts that decorates textual descriptions of each category are either a predefined template [31] (e.g., 'a photo of') or a single learnable text embedding [36]. However, we contend that fixed context prompts have limited capability for target domain images due to the discrepancy" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20791" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.09, + 0.526, + 0.278 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.237, + 0.281, + 0.413, + 0.294 + ], + "angle": 0, + "content": "(a) Visualization of textual prompts" + }, + { + "type": "image", + "bbox": [ + 0.532, + 0.091, + 0.866, + 0.279 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.644, + 0.281, + 0.794, + 0.294 + ], + "angle": 0, + "content": "(b) Results on BDD100K [58]" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.305, + 0.908, + 0.377 + ], + "angle": 0, + "content": "Figure 2. PCA visualization of textual prompts (left) and qualitative results on various environments (e.g., Day, Dawn, Night) in BDD100K [58] (right). The models are trained on GTAV [42] with the CLIP-pretrained backbone (ViT-B) [41]. A fixed single-context prompt lacks flexibility in adapting to various domain shifts due to its rigidity. In contrast, our framework utilizes domain-specific properties from input images as context prompts, enhancing semantic alignment between text and images. As a result, as shown in (b), our approach exhibits improved robustness across diverse environments." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.402, + 0.482, + 0.522 + ], + "angle": 0, + "content": "between visual and textual contexts as depicted in Fig. 1. Firstly, a handcrafted template inherently encodes the characteristics of a specific domain which restricts its generalizability to other unseen domains. Secondly, prompt optimization is prone to source domain overfitting, especially in the single-source setting. Consequently, both approaches may enlarge the gap between textual and visual contexts, leading to suboptimal performances on target domains." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.524, + 0.483, + 0.765 + ], + "angle": 0, + "content": "In this perspective, we point out that the semantic form of the textual representation for an object category should be changed with respect to the input visual context in order to strengthen their semantic correspondence. For instance, in the real-world driving scene at night of Fig. 2 (b), target categories (e.g., car and sky) possess distinct textures from those of synthetic daytime images. Hence, it would be more appropriate to decorate 'a car' with the textual prompt 'at night in the real-world' which is reflecting the domain characteristics. As a result, the modified text feature will include details such as darkened car exteriors and light reflections, enhancing the semantic alignment. Nevertheless, it is challenging to design domain-adaptive prompts in the DGSS setting where the training dataset covers only a single domain. Moreover, this setting hinders the visual pipeline from learning robust image features against domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "To address these challenges, we introduce a novel framework, namely Domain-aware Prompt-driven Masked Transformer (DPMFormer), focusing on two key aspects: (1) Leveraging domain-specific properties of the input image (domain-awareness) and (2) Generating accurate outputs on images with dissimilar domain characteristics (domain-robustness). To cultivate the domain-awareness, we propose a novel domain-aware prompt learning, which translates the domain-specific properties of an input image into con" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.584 + ], + "angle": 0, + "content": "text prompts via an auxiliary network. In addition, to obtain diverse domain properties in the single-source setting, we apply texture perturbations to synthesize novel domain images. Furthermore, exploiting both source and novel domain images, we propose a domain-aware contrastive loss to ensure that the derived prompts effectively capture domain-specific properties of the input image. This loss encourages the anchor context prompts to be distinguishable from those of different domain characteristics, while being closer to those from the anchor domain. With the proposed domain-aware prompt, the model accurately identifies target classes while being aware of the input domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.591, + 0.909, + 0.803 + ], + "angle": 0, + "content": "Moreover, we strive to provide better domain robustness guidances to the visual encoder and decoders. We carefully organize the texture perturbations with structure-preserving image transformations, ensuring that the original visual context remains intact. Instead of simply reusing the original ground truths for novel domain images, we introduce domain-robust consistency loss to guarantee reliable predictions under severe domain shifts. The loss consists of class consistency and mask consistency losses, which penalize discrepancies between class and mask predictions of given image pairs, respectively. Furthermore, domain-robust consistency losses are applied at every layer of the transformer decoder, preventing discrepancies in earlier layers from propagating to later parts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Through evaluations on various DGSS benchmarks, we demonstrate the superiority of the proposed framework. Notably, our framework achieves state-of-the-art semantic segmentation performance across domain generalization scenarios. Additionally, ablation studies and detailed analyses validate the effectiveness of each component." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20792" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.241, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.117, + 0.471, + 0.133 + ], + "angle": 0, + "content": "2.1. Domain Generalized Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.14, + 0.483, + 0.307 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to learn domain-invariant representations that generalize robustly to various unseen target domains. The task assumes a single-source setting, where only one dataset is available during training. Unlike domain adaptation [8, 49, 66] and test-time domain adaptation [53, 54, 57], access to target domains is strictly prohibited, making DGSS more challenging. Previous studies approached DGSS in two ways primarily: feature whitening and normalization [6, 37-39], and domain randomization approaches [4, 14, 19, 24, 27, 28, 37, 38, 56, 59, 61, 62]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.309, + 0.483, + 0.52 + ], + "angle": 0, + "content": "Feature whitening and normalization approaches [6, 37-39] mainly focus on removing features which are variant to domain shifts. Exploiting the characteristics of instance whitening [30] and instance normalization [50] which can effectively remove texture and style from the input image, these approach leverages those operations in between the backbone module to minimize the effect from domain and texture changes. Representatively, RobustNet [6] introduced instance selective whitening module that finds gram matrix components sensitive to photometric changes and minimize their changes. However, due to the natural difficulty in disentangling domain-specific and domain-invariant features, these approaches show limited performance gains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Domain randomization approaches [4, 14, 19, 24, 27, 28, 56, 59, 61, 62] augment novel domains from the source domain by modifying either the images or their features through various methods, e.g. affine transformations [14, 20, 28, 56, 61, 62], image translation [59, 65], frequency decomposition [4, 19], and photometric transformations [24]. Synthesized samples increase the domain diversity of the training dataset, reducing the domain gap between the learned representation and the test data. Moreover, most of the generative DGSS methods adopt context-preserving transformations to compute the output discrepancy between the original sample and the augmented one. Representatively, SHADE [61] proposed a style consistency loss to encourage model to learn invariant pixel-level semantic information by minimizing the Jenson-Shannon Divergence (JSD) between the output predictions of original and augmented images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.903 + ], + "angle": 0, + "content": "Meanwhile, several studies [3, 10] have built DGSS upon Mask2Former [5], that leverages attention mechanism [51] renowned for robustness against domain shifts [16, 17, 46]. Mask2former [5] involves a transformer decoder that exploits object query features to group pixels of same objects or categories. Based on this, HGFormer [10] first proposed a hierarchical framework that groups pixels to form part-level masks for complementing whole-level pixel group-" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "ing procedure. Similarly, CMFormer [3] utilizes downsampled features additionally via feature fusion which are more domain-invariant than the original features." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.149, + 0.905, + 0.181 + ], + "angle": 0, + "content": "2.2. Language-driven Domain Generalized Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.187, + 0.907, + 0.596 + ], + "angle": 0, + "content": "Recent studies [13, 21, 22, 36, 55] has investigated to exploit Vision Language Models (VLMs) [9, 23, 33, 40, 41, 45, 47] for DGSS, owing to its powerful generalizability learned from large-scale datasets of image-text pairs. VLTseg [21] employs the image encoder of VLMs for initializing backbone parameters, and fine-tune the network with the segmentation objective function. FAMix [13] proposes to yield class-specific novel styles by concatenating random style description and class names, then performs style randomization by locally mixing the source and the novel styles. DAP [22] exploits text encoder to distill the semantic textual knowledge of target categories to the visual backbone. TQDM [36] points out deficient use of language information in aforementioned works, and introduces a textual-query driven framework for DGSS. They utilize the textual description of each categories for initializing input query features for transformer decoder, and also for computing the text-to-pixel attention to enhance semantic clarity of pixel features. Despite their advancements, we argue that two points are overlooked: (1) the domain discrepancy between the learned textual prompt and the domain-specific property of the input image (2) lack of domain robustness guidance. To mitigate these limitations, we carefully refine the textual queries with the captured domain-specific property of the input image, and augment novel domain images to improve domain-awareness of the model as well as prediction consistency learning." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.611, + 0.613, + 0.627 + ], + "angle": 0, + "content": "3. Methods" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.637, + 0.655, + 0.652 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We design our DPMFormer based on mask classification architecture, i.e., Mask2Former [5]. Mask2Former consists of an image encoder \\(ENC_I\\), pixel decoder \\(DEC_{\\mathrm{pix}}\\) and transformer decoder \\(DEC_{\\mathrm{tr}}\\). The input RGB image \\(x \\in \\mathbb{R}^{3 \\times H \\times W}\\) is first fed to the image encoder for feature extraction, and then converted to pixel-wise features \\(z \\in \\mathbb{R}^{D \\times H \\times W}\\) through the pixel decoder. The transformer decoder iteratively refines \\(N\\) object queries \\(q \\in \\mathbb{R}^{N \\times L}\\) with multi-scale image features from the pixel decoder. Thereafter, refined object queries are projected into mask embeddings \\(m \\in \\mathbb{R}^{N \\times D}\\) to generate pixel-level mask prediction via dot product as \\(\\hat{y}^{\\mathrm{mask}} = m \\cdot z\\), \\(\\hat{y}^{\\mathrm{mask}} \\in \\mathbb{R}^{N \\times H \\times W}\\). The category label of each object query \\(q\\) is predicted with a linear classifier, i.e., \\(c_q \\in \\mathbb{R}^K\\), \\(c \\in \\mathbb{R}^{N \\times K}\\). The final prediction result is derived by the matrix multiplication between the pixel-level mask prediction and class prediction" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20793" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.087, + 0.912, + 0.315 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.326, + 0.908, + 0.385 + ], + "angle": 0, + "content": "Figure 3. Illustration of DPMFormer. We use Mask2Former [5] based architecture which consists of a backbone image encoder \\((ENC_I)\\), a pixel decoder \\((DEC_{\\mathrm{pix}})\\), a transformer decoder \\((DEC_{\\mathrm{tr}})\\), and a text encoder \\((ENC_T)\\). During training, we synthesize images with a novel domain style via texture perturbation. Both images are incorporated to compose a batch and exploited for learning domain-awareness (Sec. 3.2.1) and domain-robustness (Sec. 3.2.2)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.407, + 0.388, + 0.425 + ], + "angle": 0, + "content": "of queries as \\(\\hat{y} = c^{\\top}\\cdot \\hat{y}^{\\mathrm{mask}},\\hat{y}\\in \\mathbb{R}^{K\\times H\\times W}\\)" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.484, + 0.637 + ], + "angle": 0, + "content": "Furthermore, inspired by [36], we employ a text encoder \\(ENC_{T}\\) for initializing object queries from textual descriptions. Pre-trained vision language mdoels (VLMs), e.g., CLIP [41], is exploited to initialize both \\(ENC_{I}\\) and \\(ENC_{T}\\) parameters, and the parameters of \\(ENC_{T}\\) remain frozen during training to preserve learned linguistic knowledge. The text encoder \\(ENC_{T}\\) produces a text feature \\(t_k \\in \\mathbb{R}^C\\) corresponding to each class label \\(\\mathrm{class}_k\\) with a learnable context prompt \\(p\\), i.e., \\(t_k = ENC_{T}([p, \\{\\mathrm{class}_{k}\\}])\\). Then, obtained text embeddings \\(t = \\{t_k\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times C}\\) are passed to a multi-layer perceptron (MLP) to derive initial textual object queries \\(q_t^0\\). Also, the text-to-pixel attention mechanism is established in a pixel decoder layer to enhance pixel semantic clarity enhancement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.65, + 0.228, + 0.665 + ], + "angle": 0, + "content": "3.2. DPMFormer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.824 + ], + "angle": 0, + "content": "As illustrated in Fig. 3, Domain-aware Prompt-driven Masked Transformer (DPMFormer), aims to cultivate two core aspects of domain generalization for the model: domain-awareness and domain-robustness. For domain-awareness, we propose the domain-aware context prompt learning which translates domain-specific property of the input image into the text embedding for semantic alignment between textual and visual features. And for domain-robustness, we encourage the model to generate accurate outputs against textural changes via consistency learning." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Texture perturbation. In order to effectively guide the model with diverse domain characteristics, we stylize the training dataset to generate an auxiliary domain dataset. Following RobustNet [6], we adopt photometric transformations—comprising strong color jittering, gaussian blur" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.456 + ], + "angle": 0, + "content": "and noise injection—for their simplicity as well as content-preserving property. The generated image \\( x' \\) is combined with its original image \\( x \\) to form a batch for training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.46, + 0.837, + 0.475 + ], + "angle": 0, + "content": "3.2.1. Domain-aware context prompt learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.78 + ], + "angle": 0, + "content": "Textual query generation and prompt learning [64] plays a crucial role in enabling the model to leverage the linguistic semantics of each class previously learned by the VLM. Although the learned context prompt is beneficial for performance improvement, it is optimized solely within the source domain without direct consideration of domain shifts. This single context prompt may yield strong results on several domain shift scenarios where target domain has similar domain characteristics with the source. However, in cases of severe domain shift, the performance gain may be limited due to the contextual mismatch between the target domain image and the learned prompt. For example, if the prompt learning is conducted on a dataset containing only sunny day images, the learned prompt is could be misaligned when encountering rainy night scenes. As such, the semantic misalignment should be addressed in order to fully utilize rich semantic knowledge of pre-trained VLM. To this end, we propose to generate domain-aware context prompt that extracts domain-specific properties from the input image as a context prompt." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.781, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To obtain domain-aware context prompt from the input image \\( x \\), we design an auxiliary network \\( h_{\\theta}(\\cdot) \\) that takes visual feature as an input and generates a domain-specific prompt embedding \\( \\pi_x = h_\\theta (\\hat{F} (x)) \\), where \\( \\hat{F} (x) \\) denotes a visual feature extracted from a frozen visual backbone of CLIP [41]. We use the class token as a visual feature \\( \\hat{F} (x) \\) for its global representation [11]. Nextly, the obtained domain-specific prompt embedding \\( \\pi_{x} \\) is in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20794" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.213 + ], + "angle": 0, + "content": "tegrated with the context prompt embedding \\( p \\) through addition, i.e., \\( p_x = p + \\pi_x \\), then concatenated with text embeddings of classes to generate domain-aware textual features as \\( t_{x,k} = ENC_T([p_x,\\{\\mathrm{class}_k\\}]) \\). To encourage derived textual features to include domain-specific information of the input image, we introduce a novel domain-aware contrastive learning framework with the original and augmented images. The loss function is depicted as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.102, + 0.238, + 0.483, + 0.292 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {c o n t r a}} = - \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} \\log \\frac {\\sum_ {j \\in \\mathcal {P} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}{\\sum_ {j \\in \\mathcal {P} _ {i} \\cup \\mathcal {N} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.292, + 0.485, + 0.594 + ], + "angle": 0, + "content": "where \\(\\mathrm{sim}(\\cdot)\\) means similarity metric and \\(\\tau\\) is a temperature parameter. \\(B\\) denotes the batch size of original images, \\(\\mathcal{P}_i\\) and \\(\\mathcal{N}_i\\) are positive sets and negative sets of the \\(i\\)-th image, respectively. The positive set \\(\\mathcal{P}_i\\) is composed of the indices of samples having same domain characteristics with the anchor \\(i\\) whereas the others belong to negative set \\(\\mathcal{N}_i\\). For example, when an anchor \\(i\\) is of original source domain images, other original source domain images belongs to \\(\\mathcal{P}_i\\) while all augmented images are included in \\(\\mathcal{N}_i\\). In case of an augmented image as an anchor, \\(\\mathcal{P}_i = \\{i\\}\\) and \\(\\mathcal{N}_i = \\{1,\\dots,2B\\} \\setminus i\\). With the proposed loss, we encourage \\(h_\\theta\\) to capture domain-specific property of the image and reflect it to the output text feature \\(t_k\\). Also, the final domain-aware text feature is guided by the task loss \\(\\mathcal{L}_{seg}\\) to be aligned with the original image and to minimize segmentation errors. We note that the proposed loss considers the domain information unlike CoCoOp [63] which treats \\(\\mathcal{P}_i\\) as the text feature corresponding to a specific image feature while the remaining images forming the negative set. We provide comparative analysis with CoCoOp in Sec.4.4.2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.604, + 0.388, + 0.619 + ], + "angle": 0, + "content": "3.2.2. Domain-robust consistency learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.623, + 0.485, + 0.698 + ], + "angle": 0, + "content": "To further enhance the domain robustness, we encourage model to generate persistent predictions in the domain shift scenario. To enhance prediction consistency, we induce the model to minimize the prediction discrepancy in terms of the mask and class label as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.763 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c o n s}} = \\sum_ {s = 1} ^ {S} \\lambda_ {\\mathrm {m c}} \\cdot \\mathcal {L} _ {\\mathrm {m c}} \\left(\\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}, \\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}\\right) + \\lambda_ {\\mathrm {c c}} \\cdot \\mathcal {L} _ {\\mathrm {c c}} \\left(\\hat {c} _ {q _ {i}, \\mathrm {s}}, \\hat {c} _ {q _ {i} ^ {\\prime}, \\mathrm {s}}\\right). \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.763, + 0.485, + 0.902 + ], + "angle": 0, + "content": "\\(S\\) denotes the number of transformer blocks in the transformer decoder, and \\(\\mathcal{L}_{\\mathrm{mc}}\\) and \\(\\mathcal{L}_{\\mathrm{cc}}\\) represent the mask and class consistency losses, respectively. To compute the loss at the \\(s\\)-th transformer decoder block, we obtain \\(\\{\\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}, \\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}\\}\\) and \\(\\{\\hat{c}_{q_i,s}, \\hat{c}_{q_i',s}\\}\\) which are pairs of mask predictions and class predictions of \\(i\\)-th query \\(q_i\\) from the original and augmented image pair \\(\\{x, x'\\}\\), respectively. We employ binary cross entropy and Jensen-Shannon divergence as a discrepancy measure of the mask consistency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.168 + ], + "angle": 0, + "content": "loss \\((\\mathcal{L}_{mc})\\) and class consistency loss \\((\\mathcal{L}_{cc})\\), respectively. With the help of domain-aware context prompts and above losses, our model learns to predict not only accurately but also consistently in various domain shift scenarios during training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.174, + 0.729, + 0.188 + ], + "angle": 0, + "content": "3.3. Overall Loss Functions" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.196, + 0.906, + 0.243 + ], + "angle": 0, + "content": "The overall loss of our framework is a weighted sum of the task loss \\(\\mathcal{L}_{\\mathrm{seg}}\\), VLM regularization loss \\(\\mathcal{L}_{\\mathrm{reg}}\\), domain-aware contrastive loss \\(\\mathcal{L}_{\\mathrm{contra}}\\), and the consistency loss \\(\\mathcal{L}_{\\mathrm{cons}}\\)." + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.248, + 0.907, + 0.265 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c o n t r a}} \\mathcal {L} _ {\\text {c o n t r a}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.269, + 0.906, + 0.346 + ], + "angle": 0, + "content": "where \\(\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}\\) are constant weighting factors. We note that \\(\\mathcal{L}_{\\mathrm{seg}}\\) and \\(\\mathcal{L}_{\\mathrm{cons}}\\) are calculated with all queries and its predictions from every block of the transformer decoder. We provide details of the baseline losses \\((\\mathcal{L}_{\\mathrm{seg}},\\mathcal{L}_{\\mathrm{reg}})\\) in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.357, + 0.646, + 0.373 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.381, + 0.731, + 0.396 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.403, + 0.906, + 0.433 + ], + "angle": 0, + "content": "Datasets. We validate DPMFormer on synthetic-to-real and real-to-real scenarios in the single-source setting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.433, + 0.906, + 0.538 + ], + "angle": 0, + "content": "Synthetic datasets. GTAV [42] is a representative synthetic dataset which consists of 24,966 images with a resolution of \\(1914 \\times 1052\\). The training split contains 12,403 images, while the validation and test set are of 6,382 and 6,181 images, respectively. SYNTHIA [43] dataset provides 6,580 images for training and 2,820 images for validation respectively, with the image resolution at \\(1280 \\times 760\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.659 + ], + "angle": 0, + "content": "Real-world datasets. Cityscapes [7] is a dataset collected from the real environment. The resolution of each image is \\(2048 \\times 1024\\), and the population of training and validation split is 2,975 and 500, respectively. BDD100K [58] includes 7,000 training images and 1,000 validation images with the resolution of \\(1280 \\times 720\\). Mapillary [34] is composed of images with diverse resolution, where the training and validation size is 18,000 and 2,000, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.66, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Network architecture. Our approach leverages the vision transformer-based models as backbones, initialized with either the CLIP [41] (ViT-B) or EVA02-CLIP [45] (EVA02-L) model. The CLIP backbone is configured with a patch size of 16, while the EVA02-CLIP backbone uses a patch size of 14. For the pixel and transformer decoder, aforementioned, we employ a mask classification architecture [5], consisting of \\(N = 9\\) layers with masked attention mechanisms. Additionally, we design the auxiliary network \\(h_{\\theta}\\) for domain-aware context prompt generation as a shallow multi-layer structure (BatchNorm-Linear-ReLU-Linear)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Training and evaluation. To optimize DPMFormer, we employ an AdamW [32] where the learning rate is set as \\(1 \\times 10^{-5}\\) and \\(1 \\times 10^{-4}\\) for synthetic and real training datasets, respectively. Following previous transformer-based studies [18, 36], we apply linear warm-up [15] for" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "20795" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.095, + 0.089, + 0.48, + 0.274 + ], + "angle": 0, + "content": "
Models (GTAV)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10145.3341.1840.7742.43
WildNet [28]ResNet-10145.7941.7347.0844.87
SHADE [61]ResNet-10146.6643.6645.5045.27
TLDR [27]ResNet-10147.5844.8848.8047.09
FAMix* [13]ResNet-10149.4746.4051.9749.28
SHADE [61]MiT-B553.2748.1954.9952.15
IBAFormer [44]MiT-B556.3449.7658.2654.79
VLTSeg* [21]ViT-B47.5045.7054.3049.17
TQDM* [36]ViT-B57.5047.6659.7654.97
DPMFormer* (ours)ViT-B59.0051.8063.6258.14
VLTSeg** [21]EVA02-L65.6058.4066.5063.50
Rein** [55]EVA02-L65.3060.5064.9063.60
Rein†ViT-L66.4060.4066.1064.30
TQDM** [36]EVA02-L68.8859.1870.1066.05
DPMFormer** (ours)EVA02-L70.0860.4870.6667.07
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.286, + 0.483, + 0.371 + ], + "angle": 0, + "content": "Table 1. Comparison with the state-of-the-art DGSS methods on synthetic-to-real scenario with GTAV [42] as a source. We note that *, **, and † denotes the models initialized with pretrained CLIP [41], EVA02-CLIP [45], and DINOv2 [35], respectively. The best and the second best performances are highlighted with bold and underline, respectively." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.384, + 0.48, + 0.469 + ], + "angle": 0, + "content": "
Models (Synthia)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10140.8735.9837.2638.04
TLDR [27]ResNet-10142.6035.4637.4638.51
IBAFormer [44]MiT-B550.9244.6650.5848.72
VLTSeg** [21]EVA02-L56.8050.5054.5053.93
TQDM** [36]EVA02-L57.9952.4354.8755.10
DPMFormer** (ours)EVA02-L58.9254.3960.0857.80
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.48, + 0.483, + 0.509 + ], + "angle": 0, + "content": "Table 2. Comparison with the state-of-the-art DGSS methods where SYNTHIA [43] is set as a source dataset." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.535, + 0.483, + 0.673 + ], + "angle": 0, + "content": "initial 1,500 iterations and rare class sampling [18]. The optimizer settings are identical for both CLIP and EVA02-CLIP backbone models. We set the total training iterations and the batch size as 20,000 and 8, respectively. Weighting factors \\(\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}\\) are set as 1, 1, 10. We crop the input image to have a resolution of \\(512\\times 512\\). The texture perturbation is only applied during training. We use mean Intersection over Union (mIoU) [12] to quantitatively evaluate the results following the convention." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.682, + 0.285, + 0.699 + ], + "angle": 0, + "content": "4.2. Quantitative Results" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.704, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Synthetic-to-real. As shown in Tab. 1, we achieve state-of-the-art in every target domain with both backbones. Notably, with the ViT-B backbone initialized with pretrained CLIP [41], DPMFormer consistently surpasses previous state-of-the-art [36] by \\(3.17\\%\\) on average mIoU among target domains. In detail, on Cityscapes which contains mostly daytime real images, the domain robustness empowered by consistency learning assists DPMFormer to cope with the domain gap caused by visual realism, improving the state-of-the-art performance by \\(1.5\\%\\). Meanwhile on BDD100K and Mapillary which have higher environmental variation in terms of weather, time and location, the domain-aware prompt generation enables the model to adaptively leverage" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.089, + 0.905, + 0.258 + ], + "angle": 0, + "content": "
Models (Cityscapes)BackboneBDDMapillaryAvg.
SAN-SAW [39]ResNet-10154.7361.2742.43
WildNet [28]ResNet-10147.0141.7344.87
SHADE [61]ResNet-10150.9543.6645.27
TQDM* [36]ViT-B50.5465.7458.14
DPMFormer* (ours)ViT-B54.8167.7261.27
HGFormer [10]EVA02-L61.5072.1066.80
VLTSeg** [21]EVA02-L64.4076.4070.40
Rein** [55]EVA02-L64.1069.5066.80
Rein†ViT-L65.0072.3068.65
TQDM [36]**EVA02-L64.7276.1570.44
DPMFormer** (ours)EVA02-L64.276.6770.44
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.268, + 0.905, + 0.297 + ], + "angle": 0, + "content": "Table 3. Comparison with the state-of-the-arts trained with Cityscapes [7] on the real-to-real scenario." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.459 + ], + "angle": 0, + "content": "the textual knowledge for segmentation, impressively escalating the performance by \\(4.14\\%\\), and \\(3.17\\%\\), respectively. Moreover, even with the larger backbone [45], we mark the highest average performance of \\(67.07\\%\\). In Cityscapes and Mapillary dataset, we outperform TQDM [36] by \\(1.2\\%\\) and \\(0.56\\%\\) respectively, while nearly reaching the score of Rein [55] in BDD. The overall results demonstrate the superiority of DPMFormer, emphasizing the efficacy of domain-awareness as well as domain-robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.909, + 0.612 + ], + "angle": 0, + "content": "In Tab. 2, we also present synthetic-to-real results where SYNTHIA [43] is set as a source domain. Again, DPM-Former outperforms all competitors by a large margin, achieving the state-of-the-art performance in every target domain. Remarkably, our framework surpasses the previous state-of-the-art [36] by an average of \\(2.7\\%\\) mIoU across target domains. In particular, the performance on Mapillary [34] improves significantly by \\(5.21\\%\\). These results demonstrate the effectiveness of DPM-Former in addressing the domain gap in terms of texture and perspective." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Real-to-real. Furthermore, as shown in Tab.3, DPMFormer records the highest average mIoU with both backbones. With the CLIP-pretrained backbone, we significantly boost the state-of-the-art performance by \\(4.27\\%\\) and \\(1.98\\%\\) on BDD and Mapillary, respectively. These results indicate that enhanced domain robustness supports the model in predicting consistently under environmental changes, while the domain-aware prompts facilitate the semantic alignment on the unseen domain. As a result, DPMFormer successfully achieves an average performance gain of \\(3.13\\%\\) over the previous state-of-the-art. Equipped with the EVA-CLIP [45] pretrained backbone, we attain the highest mIoU of \\(76.67\\%\\) on Mapillary, while performing slightly lower on BDD. Overall, the average performance is \\(70.44\\%\\), which is comparable to TQDM [36]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.848, + 0.698, + 0.865 + ], + "angle": 0, + "content": "4.3. Qualitative Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In Fig. 4, we qualitatively compare our method with FAMix [13] and TQDM [36] on the synthetic-to-real sce" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20796" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.087, + 0.891, + 0.37 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.381, + 0.908, + 0.425 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison on synthetic-to-real scenario with the CLIP-pretrained backbone (ViT-B). The training source domain is set as GTAV [42] while the target domains are BDD100K [58] and Mapillary [34]. The overall result shows that DPMFormer accomplishes precise segmentation with the images of strong illumination contrast as well as confusing textures." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.45, + 0.486, + 0.709 + ], + "angle": 0, + "content": "nario with the pretrained CLIP backbone (ViT-B). As depicted in the results on BDD100K, both competitors struggle to discriminate accurately under environments having confusing textures or large variations of illumination. In particular, the road in the first-row image has a texture similar to the exterior of a car, leading TQDM to a mismatch between the visual and textual features. In addition, FAMix shows sensitivity to the shades, mislabeling the road as a sidewalk. On the contrary, our method effectively distinguishes the road from other categories by reflecting the domain-specific properties of real-world roads in clear weather. In the case of the second row image with the high illumination contrast, both FAMix and TQDM fail to notice sidewalks in the image due to their low brightness. On the other hand, DPMFormer carries out precise predictions under severe photometric changes owing to the domain robustness acquired from handling diverse texture changes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "As observed with the samples from the Mapillary (the third and fourth rows), FAMix and TQDM lack discriminability between the road and the sidewalk. To be specific with the first image (third row), these classes appear to have a similar color and texture, resulting in misclassifications. Meanwhile, the domain-specific properties from the image transfer the semantic knowledge of the appearance of real-world sidewalk to the model, DPMFormer accurately distinguish between the road and the sidewalk. With the second image, TQDM [36] shows vulnerability against small textural changes on the sidewalk (e.g., shades of the poles). Contrarily, DPMformer effectively performs segmentation ow" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.45, + 0.907, + 0.481 + ], + "angle": 0, + "content": "ing to its domain-robustness as well as the obtained domain-aware textual queries." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.488, + 0.618, + 0.505 + ], + "angle": 0, + "content": "4.4. Analysis" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.51, + 0.673, + 0.525 + ], + "angle": 0, + "content": "4.4.1. Ablation studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.529, + 0.909, + 0.832 + ], + "angle": 0, + "content": "To demonstrate the effectiveness of the components in DPMFormer, i.e., texture perturbation, domain-robust consistency learning \\((\\mathcal{L}_{cons})\\), and domain-aware context prompt learning \\((\\mathcal{L}_{contra})\\), we conduct ablation studies with the CLIP pretrained backbone on synthetic-to-real scenario. As presented in Tab. 4, every component contributes to the performance gain. Specifically, the texture perturbation enlarges the observable domain during training, boosts the average performance by \\(0.65\\%\\). In addition with the \\(\\mathcal{L}_{cons}\\), the model successfully equips the robustness against domain-shift and enhances the prediction accuracy on unseen target domains, increasing the average mIoU by \\(0.77\\%\\). Furthermore, \\(\\mathcal{L}_{contra}\\) allows the model to acquire domain-specific property from the input image in the form of textual prompt embedding. Consequently, the model exploits both visual and textual cues properly aligned for the target domain, considerably elevate the average mIoU by \\(1.99\\%\\). Combining all components together, DPMFormer effectively learns both domain-awareness and domain-robustness to generate accurate results consistently on unseen target domains." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.837, + 0.863, + 0.853 + ], + "angle": 0, + "content": "4.4.2. Comparison with prompt learning methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.908, + 0.903 + ], + "angle": 0, + "content": "In Tab. 5, we compare our domain-aware context prompt learning \\((\\mathcal{L}_{\\mathrm{contra}})\\) with previous prompt learning methods [25, 26, 63] to validate the efficacy for domain gen" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20797" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.173 + ], + "angle": 0, + "content": "
Models (GTAV)CityscapesBDDMapillaryAvg.
Baseline57.547.6659.7654.97
+ Perturbation57.0448.1960.9155.38
+ Lcons58.2249.3960.8456.15
+ Lcontra59.051.863.6258.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.183, + 0.483, + 0.227 + ], + "angle": 0, + "content": "Table 4. Ablation study of proposed components. The models are trained on GTAV with CLIP pretrained backbone (ViT-B). The best performance for each column is highlighted in bold." + }, + { + "type": "image", + "bbox": [ + 0.094, + 0.258, + 0.485, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.404, + 0.484, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Qualitative results on diverse styles, i.e., Minimalist, Pop Art, Bauhaus, and Cubism. The models are trained on GTAV with CLIP pretrained backbone (ViT-B)." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.493, + 0.487, + 0.902 + ], + "angle": 0, + "content": "deralization. Given the textual features \\( t = ENC_{T}(p_{x}) \\) and visual features \\( v = \\hat{F}(x) \\), the contrastive loss of CoCoOp [63] encourages the similarity between the anchor text feature and its corresponding visual feature to be higher than others. As observed in the second row, CoCoOp marginally improves the performance of prompt generator \\( h_{\\theta} \\), indicating that the generated feature becomes instancespecific which is less generalizable to other domains. Moreover, additionally assigning the visual feature from augmented sample as positives \\( (\\mathrm{CoCoOp}^{+}) \\) resulted in negligible performance gain, since the derived prompt contains domain-invariant information rather than domain-specific properties. Furthermore, although both MaPLe [25] and PromptSRC [26] employ multi-modal prompts, their performance gains are modest due to the limited generalizability of fixed prompts. On the other hand, our domain-aware contrastive loss design enable the model to proficiently capture domain-specific property from the image which are more helpful for the semantic alignment between the visual features and the textual knowledge. From the fifth to the eighth row, we verify \\( \\mathcal{L}_{\\text{contra}} \\) with different similarity calculation targets for \\( \\sin(\\cdot) \\), i.e., text features \\( t \\), text and visual features \\( (t, v) \\), and (3) output context embeddings \\( \\pi \\). The results confirm that computing the domain-aware contrastive loss empowers the domain generalizability especially when computed with the context embeddings \\( \\pi \\) which can provide direct domain guidance to \\( h_{\\theta} \\)." + }, + { + "type": "table", + "bbox": [ + 0.515, + 0.089, + 0.906, + 0.235 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{contra}} \\) (GTAV)CityscapesBDDMapillaryAvg.
-57.6549.6361.1056.13
CoCoOp [63]57.8449.9161.3356.36
\\( \\mathrm{CoCoOp}^{+} \\)57.6050.0460.9656.20
MaPLe [25]57.8750.1261.0456.34
PromptSRC [26]58.1049.7362.5156.78
Ours (sim(t,t))58.1750.1861.3756.57
Ours (sim(t,v))59.4950.3462.1857.34
Ours (sim(π,π))59.0051.8063.6258.14
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.244, + 0.907, + 0.316 + ], + "angle": 0, + "content": "Table 5. Comparison with prompt learning methods [25, 26, 63] in synthetic-to-real scenario. - denotes DPMFormer without \\(\\mathcal{L}_{\\mathrm{contra}}\\). \\(\\mathrm{CoCoOp}^{+}\\) indicates the modified loss that additionally include the feature from augmented sample of anchor as positives. \\(t\\), \\(v\\), and \\(\\pi\\) refers to the text feature, visual feature, and context embedding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.339, + 0.806, + 0.355 + ], + "angle": 0, + "content": "4.4.3. Qualitative results on diverse styles" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.358, + 0.909, + 0.6 + ], + "angle": 0, + "content": "To further verify the domain generalization capability of DPMFormer, we compare our model with TQDM [36] on images with diverse artistic styles generated by \\(\\mathrm{ChatGPT}^1\\). As shown in Fig. 5, DPMFormer correctly predicts objects and their surroundings even in severe domain shift scenarios, e.g., Cubism. With modern art styles (i.e., Minimalist, Pop Art, and Bauhaus), TQDM perplexes among road, sidewalk, and terrain due to their textural changes. Notably, DPMFormer produces more reliable results owing to the enhanced robustness against texture variations that are learned from the texture perturbation and consistency learning. In case of the scene in cubism style, TQDM mispredicts the person on the left side because of its visual similarity with the person in the traffic sign. On the other hand, our method reflects the cubism style to the textual object queries, making an accurate classification with the same instance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.612, + 0.634, + 0.629 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.638, + 0.909, + 0.879 + ], + "angle": 0, + "content": "In this paper, we presented DPMFormer, a novel domain generalization framework for semantic segmentation. To address the limited generalizability of fixed context prompt learned from a single source dataset, we introduced a novel domain-aware prompt learning which can reflect domain-specific properties of the input image into textual prompts. The proposed component enhanced the semantic alignment between visual and textual cues, assisting the model to fully leverage the abundant semantic knowledge of VLMs. Moreover, to empower domain-robustness, we simulated various domain shifts via texture perturbations, and provided consistency guidance to the model by minimizing prediction discrepancies between original and augmented images. Through extensive experiments and analyses, we demonstrated the effectiveness of DPMFormer, achieving state-of-the-art performance on various benchmarks." + }, + { + "type": "page_footnote", + "bbox": [ + 0.532, + 0.887, + 0.666, + 0.902 + ], + "angle": 0, + "content": "1 https://chat.openai.com" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "20798" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.168 + ], + "angle": 0, + "content": "Acknowledgments. This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364). This research was supported by Sookmyung Women's University Research Grants (1-2403-2034)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.182, + 0.188, + 0.198 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.207, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[1] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.264, + 0.484, + 0.317 + ], + "angle": 0, + "content": "[2] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.321, + 0.484, + 0.376 + ], + "angle": 0, + "content": "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 819-827, 2024. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.379, + 0.484, + 0.46 + ], + "angle": 0, + "content": "[4] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19288-19300, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.463, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.484, + 0.616 + ], + "angle": 0, + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 1, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.618, + 0.484, + 0.701 + ], + "angle": 0, + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 1, 3, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.703, + 0.484, + 0.744 + ], + "angle": 0, + "content": "[8] Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. Domain adaptation in computer vision applications, pages 1-35, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.484, + 0.801 + ], + "angle": 0, + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[10] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15413-15423, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.207, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.147 + ], + "angle": 0, + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.907, + 0.216 + ], + "angle": 0, + "content": "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.219, + 0.907, + 0.287 + ], + "angle": 0, + "content": "[13] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23428-23437, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.289, + 0.906, + 0.357 + ], + "angle": 0, + "content": "[14] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.359, + 0.906, + 0.385 + ], + "angle": 0, + "content": "[15] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.387, + 0.906, + 0.427 + ], + "angle": 0, + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.429, + 0.906, + 0.51 + ], + "angle": 0, + "content": "[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[18] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.582, + 0.906, + 0.649 + ], + "angle": 0, + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.651, + 0.906, + 0.705 + ], + "angle": 0, + "content": "[20] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.707, + 0.906, + 0.774 + ], + "angle": 0, + "content": "[21] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. arXiv preprint arXiv:2312.02021, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.906, + 0.845 + ], + "angle": 0, + "content": "[22] Xinyue Huo, Lingxi Xie, Hengtong Hu, Wengang Zhou, Houqiang Li, and Qi Tian. Domain-agnostic priors for semantic segmentation under unsupervised domain adaptation and domain generalization. International Journal of Computer Vision, 132(9):3954-3976, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.847, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20799" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.092, + 0.482, + 0.119 + ], + "angle": 0, + "content": "conference on machine learning, pages 4904-4916. PMLR, 2021.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.483, + 0.188 + ], + "angle": 0, + "content": "[24] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18993-19003, 2023. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.483, + 0.26 + ], + "angle": 0, + "content": "[25] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113-19122, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.261, + 0.483, + 0.343 + ], + "angle": 0, + "content": "[26] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15190-15200, 2023. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.483, + 0.412 + ], + "angle": 0, + "content": "[27] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 677-687, 2023. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.483, + 0.482 + ], + "angle": 0, + "content": "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9936-9946, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.483, + 0.483, + 0.538 + ], + "angle": 0, + "content": "[29] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5543-5551, 2017. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[30] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.678 + ], + "angle": 0, + "content": "[31] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.483, + 0.706 + ], + "angle": 0, + "content": "[32] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.707, + 0.483, + 0.762 + ], + "angle": 0, + "content": "[33] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.762, + 0.483, + 0.831 + ], + "angle": 0, + "content": "[34] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 5, 6, 7, 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.149 + ], + "angle": 0, + "content": "[36] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. European Conference on Computer Vision, 2024. 1, 3, 4, 5, 6, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.149, + 0.905, + 0.206 + ], + "angle": 0, + "content": "[37] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[38] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1863-1871, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[39] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2594-2605, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.333, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[40] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for zero-shot transfer learning. Neurocomputing, 555: 126658, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.404, + 0.905, + 0.488 + ], + "angle": 0, + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.489, + 0.905, + 0.572 + ], + "angle": 0, + "content": "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 2, 5, 6, 7, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.573, + 0.905, + 0.656 + ], + "angle": 0, + "content": "[43] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.658, + 0.905, + 0.714 + ], + "angle": 0, + "content": "[44] Qiyu Sun, Huilin Chen, Meng Zheng, Ziyan Wu, Michael Felsberg, and Yang Tang. Ibaformer: Intra-batch attention transformer for domain generalized semantic segmentation. arXiv preprint arXiv:2309.06282, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.714, + 0.905, + 0.757 + ], + "angle": 0, + "content": "[45] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.786 + ], + "angle": 0, + "content": "[46] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.786, + 0.905, + 0.829 + ], + "angle": 0, + "content": "[47] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "20800" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.146 + ], + "angle": 0, + "content": "Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[50] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.221, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.251, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[52] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.322, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In The Nineth International Conference on Learning Representations (ICLR 2021). OpenReview, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.392, + 0.483, + 0.447 + ], + "angle": 0, + "content": "[54] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.45, + 0.482, + 0.545 + ], + "angle": 0, + "content": "[55] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28619-28630, 2024. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.549, + 0.482, + 0.603 + ], + "angle": 0, + "content": "[56] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.606, + 0.482, + 0.66 + ], + "angle": 0, + "content": "[57] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8978-8987, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.663, + 0.482, + 0.744 + ], + "angle": 0, + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 2, 5, 7, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.841 + ], + "angle": 0, + "content": "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[60] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[61] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In European conference on computer vision, pages 535-552. Springer, 2022. 1, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.219 + ], + "angle": 0, + "content": "[62] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. Advances in neural information processing systems, 35:338-350, 2022. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.221, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 5, 7, 8, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.291, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.905, + 0.415 + ], + "angle": 0, + "content": "[65] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.905, + 0.485 + ], + "angle": 0, + "content": "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "20801" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_origin.pdf b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..30707c0f1dcd58c5492e94ae6329c13ddeb4b65e --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/0339a1b5-4ddb-43d4-a904-de29094899c3_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9b2516ae003fa96646b6b9a2c04eac2c73b6770f3f45dae298f3e7de26f15a0b +size 1482988 diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/full.md b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..96b5f4cac504ee5c669d49665b7b7306087394a9 --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/full.md @@ -0,0 +1,283 @@ +# Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation + +Seogkyu Jeon $^{1}$ Kibeom Hong $^{2*}$ Hyeran Byun $^{1*}$ + +$^{1}$ Yonsei University $^{2}$ Sookmyung Women's University + +jone9312@yonsei.ac.kr + +# Abstract + +Recent domain generalized semantic segmentation (DGSS) studies have achieved notable improvements by distilling semantic knowledge from Vision-Language Models (VLMs). However, they overlook the semantic misalignment between visual and textual contexts, which arises due to the rigidity of a fixed context prompt learned on a single source domain. To this end, we present a novel domain generalization framework for semantic segmentation, namely Domain-aware Prompt-driven Masked Transformer (DPM-Former). Firstly, we introduce domain-aware prompt learning to facilitate semantic alignment between visual and textual cues. To capture various domain-specific properties with a single source dataset, we propose domain-aware contrastive learning along with the texture perturbation that diversifies the observable domains. Lastly, to establish a framework resilient against diverse environmental changes, we have proposed the domain-robust consistency learning which guides the model to minimize discrepancies of prediction from original and the augmented images. Through experiments and analyses, we demonstrate the superiority of the proposed framework, which establishes a new state-of-the-art on various DGSS benchmarks. + +# 1. Introduction + +Over the decades, semantic segmentation has made remarkable progress, now being able to precisely classify each pixel in an image into categories. However, one of the shadows that lies in these advancements is that models often exhibit inconsistent and degraded performance when deployed in various real-world environments. To this end, the task of Domain Generalized Semantic Segmentation (DGSS) has arisen to overcome the discrepancy between the training and test domains, i.e., domain shift. DGSS has progressed through diverse approaches, such as feature whiten- + +![](images/4d41545c8df7c1ea7e385bccc12f726a643c721caffa29a89c4052f18bfd71b5.jpg) +Fixed context prompts [31, 36] + +![](images/ddfc0a3c25bed2213266027f02423d85c1a347d28ca61f056f3a2181431b441d.jpg) +DPMFormer (Ours) +Figure 1. Motivation of DPMFormer. Using a fixed context prompts [31, 36] tend to retain source domain properties, causing contextual misalignment with the target domain. On the other hand, DPMFormer translates domain properties of the input image into context prompts, enhancing semantic alignments. + +ing [6, 37-39] and domain randomization [4, 14, 19, 24, 27, 28, 56, 59, 61, 62]. Furthermore, some studies [3, 10] exploited multi-scale object queries with a transformer-based architecture, i.e., Mask2Former [5]. + +Despite these advancements, learning domain-robust representations solely from the single source domain remains a significant hurdle to performance improvement. Recently, several studies [13, 21, 36, 55] employed Vision-Language Models (VLMs) [41, 45] owing to their semantic knowledge learned from diverse large-scale text-image datasets. For this, pioneering VLM-based DGSS works [13, 21, 55] have adopted their pre-trained visual encoder for initialization and fine-tuning [21, 55]. In addition, TQDM [36] has introduced a framework that constructs object queries from the textual descriptions, leveraging semantic concepts from linguistic expressions. + +Yet, they still have limitations in fully utilizing textual knowledge to improve domain generalizability. Concretely, the context prompts that decorates textual descriptions of each category are either a predefined template [31] (e.g., 'a photo of') or a single learnable text embedding [36]. However, we contend that fixed context prompts have limited capability for target domain images due to the discrepancy + +![](images/fd2224252574a5e01732501e2ef9a84ce30f390bbe6082512c594b780621a4c4.jpg) +(a) Visualization of textual prompts + +![](images/6875ebcbc8727141610ab081774bbaa924c05cf3f83ded44d4d4fb0193976df5.jpg) +(b) Results on BDD100K [58] +Figure 2. PCA visualization of textual prompts (left) and qualitative results on various environments (e.g., Day, Dawn, Night) in BDD100K [58] (right). The models are trained on GTAV [42] with the CLIP-pretrained backbone (ViT-B) [41]. A fixed single-context prompt lacks flexibility in adapting to various domain shifts due to its rigidity. In contrast, our framework utilizes domain-specific properties from input images as context prompts, enhancing semantic alignment between text and images. As a result, as shown in (b), our approach exhibits improved robustness across diverse environments. + +between visual and textual contexts as depicted in Fig. 1. Firstly, a handcrafted template inherently encodes the characteristics of a specific domain which restricts its generalizability to other unseen domains. Secondly, prompt optimization is prone to source domain overfitting, especially in the single-source setting. Consequently, both approaches may enlarge the gap between textual and visual contexts, leading to suboptimal performances on target domains. + +In this perspective, we point out that the semantic form of the textual representation for an object category should be changed with respect to the input visual context in order to strengthen their semantic correspondence. For instance, in the real-world driving scene at night of Fig. 2 (b), target categories (e.g., car and sky) possess distinct textures from those of synthetic daytime images. Hence, it would be more appropriate to decorate 'a car' with the textual prompt 'at night in the real-world' which is reflecting the domain characteristics. As a result, the modified text feature will include details such as darkened car exteriors and light reflections, enhancing the semantic alignment. Nevertheless, it is challenging to design domain-adaptive prompts in the DGSS setting where the training dataset covers only a single domain. Moreover, this setting hinders the visual pipeline from learning robust image features against domain shifts. + +To address these challenges, we introduce a novel framework, namely Domain-aware Prompt-driven Masked Transformer (DPMFormer), focusing on two key aspects: (1) Leveraging domain-specific properties of the input image (domain-awareness) and (2) Generating accurate outputs on images with dissimilar domain characteristics (domain-robustness). To cultivate the domain-awareness, we propose a novel domain-aware prompt learning, which translates the domain-specific properties of an input image into con + +text prompts via an auxiliary network. In addition, to obtain diverse domain properties in the single-source setting, we apply texture perturbations to synthesize novel domain images. Furthermore, exploiting both source and novel domain images, we propose a domain-aware contrastive loss to ensure that the derived prompts effectively capture domain-specific properties of the input image. This loss encourages the anchor context prompts to be distinguishable from those of different domain characteristics, while being closer to those from the anchor domain. With the proposed domain-aware prompt, the model accurately identifies target classes while being aware of the input domain. + +Moreover, we strive to provide better domain robustness guidances to the visual encoder and decoders. We carefully organize the texture perturbations with structure-preserving image transformations, ensuring that the original visual context remains intact. Instead of simply reusing the original ground truths for novel domain images, we introduce domain-robust consistency loss to guarantee reliable predictions under severe domain shifts. The loss consists of class consistency and mask consistency losses, which penalize discrepancies between class and mask predictions of given image pairs, respectively. Furthermore, domain-robust consistency losses are applied at every layer of the transformer decoder, preventing discrepancies in earlier layers from propagating to later parts. + +Through evaluations on various DGSS benchmarks, we demonstrate the superiority of the proposed framework. Notably, our framework achieves state-of-the-art semantic segmentation performance across domain generalization scenarios. Additionally, ablation studies and detailed analyses validate the effectiveness of each component. + +# 2. Related Works + +# 2.1. Domain Generalized Semantic Segmentation + +Domain Generalized Semantic Segmentation (DGSS) aims to learn domain-invariant representations that generalize robustly to various unseen target domains. The task assumes a single-source setting, where only one dataset is available during training. Unlike domain adaptation [8, 49, 66] and test-time domain adaptation [53, 54, 57], access to target domains is strictly prohibited, making DGSS more challenging. Previous studies approached DGSS in two ways primarily: feature whitening and normalization [6, 37-39], and domain randomization approaches [4, 14, 19, 24, 27, 28, 37, 38, 56, 59, 61, 62]. + +Feature whitening and normalization approaches [6, 37-39] mainly focus on removing features which are variant to domain shifts. Exploiting the characteristics of instance whitening [30] and instance normalization [50] which can effectively remove texture and style from the input image, these approach leverages those operations in between the backbone module to minimize the effect from domain and texture changes. Representatively, RobustNet [6] introduced instance selective whitening module that finds gram matrix components sensitive to photometric changes and minimize their changes. However, due to the natural difficulty in disentangling domain-specific and domain-invariant features, these approaches show limited performance gains. + +Domain randomization approaches [4, 14, 19, 24, 27, 28, 56, 59, 61, 62] augment novel domains from the source domain by modifying either the images or their features through various methods, e.g. affine transformations [14, 20, 28, 56, 61, 62], image translation [59, 65], frequency decomposition [4, 19], and photometric transformations [24]. Synthesized samples increase the domain diversity of the training dataset, reducing the domain gap between the learned representation and the test data. Moreover, most of the generative DGSS methods adopt context-preserving transformations to compute the output discrepancy between the original sample and the augmented one. Representatively, SHADE [61] proposed a style consistency loss to encourage model to learn invariant pixel-level semantic information by minimizing the Jenson-Shannon Divergence (JSD) between the output predictions of original and augmented images. + +Meanwhile, several studies [3, 10] have built DGSS upon Mask2Former [5], that leverages attention mechanism [51] renowned for robustness against domain shifts [16, 17, 46]. Mask2former [5] involves a transformer decoder that exploits object query features to group pixels of same objects or categories. Based on this, HGFormer [10] first proposed a hierarchical framework that groups pixels to form part-level masks for complementing whole-level pixel group- + +ing procedure. Similarly, CMFormer [3] utilizes downsampled features additionally via feature fusion which are more domain-invariant than the original features. + +# 2.2. Language-driven Domain Generalized Semantic Segmentation + +Recent studies [13, 21, 22, 36, 55] has investigated to exploit Vision Language Models (VLMs) [9, 23, 33, 40, 41, 45, 47] for DGSS, owing to its powerful generalizability learned from large-scale datasets of image-text pairs. VLTseg [21] employs the image encoder of VLMs for initializing backbone parameters, and fine-tune the network with the segmentation objective function. FAMix [13] proposes to yield class-specific novel styles by concatenating random style description and class names, then performs style randomization by locally mixing the source and the novel styles. DAP [22] exploits text encoder to distill the semantic textual knowledge of target categories to the visual backbone. TQDM [36] points out deficient use of language information in aforementioned works, and introduces a textual-query driven framework for DGSS. They utilize the textual description of each categories for initializing input query features for transformer decoder, and also for computing the text-to-pixel attention to enhance semantic clarity of pixel features. Despite their advancements, we argue that two points are overlooked: (1) the domain discrepancy between the learned textual prompt and the domain-specific property of the input image (2) lack of domain robustness guidance. To mitigate these limitations, we carefully refine the textual queries with the captured domain-specific property of the input image, and augment novel domain images to improve domain-awareness of the model as well as prediction consistency learning. + +# 3. Methods + +# 3.1. Preliminaries + +We design our DPMFormer based on mask classification architecture, i.e., Mask2Former [5]. Mask2Former consists of an image encoder $ENC_I$ , pixel decoder $DEC_{\mathrm{pix}}$ and transformer decoder $DEC_{\mathrm{tr}}$ . The input RGB image $x \in \mathbb{R}^{3 \times H \times W}$ is first fed to the image encoder for feature extraction, and then converted to pixel-wise features $z \in \mathbb{R}^{D \times H \times W}$ through the pixel decoder. The transformer decoder iteratively refines $N$ object queries $q \in \mathbb{R}^{N \times L}$ with multi-scale image features from the pixel decoder. Thereafter, refined object queries are projected into mask embeddings $m \in \mathbb{R}^{N \times D}$ to generate pixel-level mask prediction via dot product as $\hat{y}^{\mathrm{mask}} = m \cdot z$ , $\hat{y}^{\mathrm{mask}} \in \mathbb{R}^{N \times H \times W}$ . The category label of each object query $q$ is predicted with a linear classifier, i.e., $c_q \in \mathbb{R}^K$ , $c \in \mathbb{R}^{N \times K}$ . The final prediction result is derived by the matrix multiplication between the pixel-level mask prediction and class prediction + +![](images/3f416f43c54e57657e6fdc211311c1c3745cf26c4ef7318bcde770565afa1646.jpg) +Figure 3. Illustration of DPMFormer. We use Mask2Former [5] based architecture which consists of a backbone image encoder $(ENC_I)$ , a pixel decoder $(DEC_{\mathrm{pix}})$ , a transformer decoder $(DEC_{\mathrm{tr}})$ , and a text encoder $(ENC_T)$ . During training, we synthesize images with a novel domain style via texture perturbation. Both images are incorporated to compose a batch and exploited for learning domain-awareness (Sec. 3.2.1) and domain-robustness (Sec. 3.2.2). + +of queries as $\hat{y} = c^{\top}\cdot \hat{y}^{\mathrm{mask}},\hat{y}\in \mathbb{R}^{K\times H\times W}$ + +Furthermore, inspired by [36], we employ a text encoder $ENC_{T}$ for initializing object queries from textual descriptions. Pre-trained vision language mdoels (VLMs), e.g., CLIP [41], is exploited to initialize both $ENC_{I}$ and $ENC_{T}$ parameters, and the parameters of $ENC_{T}$ remain frozen during training to preserve learned linguistic knowledge. The text encoder $ENC_{T}$ produces a text feature $t_k \in \mathbb{R}^C$ corresponding to each class label $\mathrm{class}_k$ with a learnable context prompt $p$ , i.e., $t_k = ENC_{T}([p, \{\mathrm{class}_{k}\}])$ . Then, obtained text embeddings $t = \{t_k\}_{k=1}^{K} \in \mathbb{R}^{K \times C}$ are passed to a multi-layer perceptron (MLP) to derive initial textual object queries $q_t^0$ . Also, the text-to-pixel attention mechanism is established in a pixel decoder layer to enhance pixel semantic clarity enhancement. + +# 3.2. DPMFormer + +As illustrated in Fig. 3, Domain-aware Prompt-driven Masked Transformer (DPMFormer), aims to cultivate two core aspects of domain generalization for the model: domain-awareness and domain-robustness. For domain-awareness, we propose the domain-aware context prompt learning which translates domain-specific property of the input image into the text embedding for semantic alignment between textual and visual features. And for domain-robustness, we encourage the model to generate accurate outputs against textural changes via consistency learning. + +Texture perturbation. In order to effectively guide the model with diverse domain characteristics, we stylize the training dataset to generate an auxiliary domain dataset. Following RobustNet [6], we adopt photometric transformations—comprising strong color jittering, gaussian blur + +and noise injection—for their simplicity as well as content-preserving property. The generated image $x'$ is combined with its original image $x$ to form a batch for training. + +# 3.2.1. Domain-aware context prompt learning + +Textual query generation and prompt learning [64] plays a crucial role in enabling the model to leverage the linguistic semantics of each class previously learned by the VLM. Although the learned context prompt is beneficial for performance improvement, it is optimized solely within the source domain without direct consideration of domain shifts. This single context prompt may yield strong results on several domain shift scenarios where target domain has similar domain characteristics with the source. However, in cases of severe domain shift, the performance gain may be limited due to the contextual mismatch between the target domain image and the learned prompt. For example, if the prompt learning is conducted on a dataset containing only sunny day images, the learned prompt is could be misaligned when encountering rainy night scenes. As such, the semantic misalignment should be addressed in order to fully utilize rich semantic knowledge of pre-trained VLM. To this end, we propose to generate domain-aware context prompt that extracts domain-specific properties from the input image as a context prompt. + +To obtain domain-aware context prompt from the input image $x$ , we design an auxiliary network $h_{\theta}(\cdot)$ that takes visual feature as an input and generates a domain-specific prompt embedding $\pi_x = h_\theta (\hat{F} (x))$ , where $\hat{F} (x)$ denotes a visual feature extracted from a frozen visual backbone of CLIP [41]. We use the class token as a visual feature $\hat{F} (x)$ for its global representation [11]. Nextly, the obtained domain-specific prompt embedding $\pi_{x}$ is in + +tegrated with the context prompt embedding $p$ through addition, i.e., $p_x = p + \pi_x$ , then concatenated with text embeddings of classes to generate domain-aware textual features as $t_{x,k} = ENC_T([p_x,\{\mathrm{class}_k\}])$ . To encourage derived textual features to include domain-specific information of the input image, we introduce a novel domain-aware contrastive learning framework with the original and augmented images. The loss function is depicted as follows: + +$$ +\mathcal {L} _ {\text {c o n t r a}} = - \frac {1}{2 B} \sum_ {i = 1} ^ {2 B} \log \frac {\sum_ {j \in \mathcal {P} _ {i}} \exp \sin \left(\pi_ {i} , \pi_ {j}\right) / \tau}{\sum_ {j \in \mathcal {P} _ {i} \cup \mathcal {N} _ {i}} \exp \sin \left(\pi_ {i} , \pi_ {j}\right) / \tau}, \tag {1} +$$ + +where $\mathrm{sim}(\cdot)$ means similarity metric and $\tau$ is a temperature parameter. $B$ denotes the batch size of original images, $\mathcal{P}_i$ and $\mathcal{N}_i$ are positive sets and negative sets of the $i$ -th image, respectively. The positive set $\mathcal{P}_i$ is composed of the indices of samples having same domain characteristics with the anchor $i$ whereas the others belong to negative set $\mathcal{N}_i$ . For example, when an anchor $i$ is of original source domain images, other original source domain images belongs to $\mathcal{P}_i$ while all augmented images are included in $\mathcal{N}_i$ . In case of an augmented image as an anchor, $\mathcal{P}_i = \{i\}$ and $\mathcal{N}_i = \{1,\dots,2B\} \setminus i$ . With the proposed loss, we encourage $h_\theta$ to capture domain-specific property of the image and reflect it to the output text feature $t_k$ . Also, the final domain-aware text feature is guided by the task loss $\mathcal{L}_{seg}$ to be aligned with the original image and to minimize segmentation errors. We note that the proposed loss considers the domain information unlike CoCoOp [63] which treats $\mathcal{P}_i$ as the text feature corresponding to a specific image feature while the remaining images forming the negative set. We provide comparative analysis with CoCoOp in Sec.4.4.2. + +# 3.2.2. Domain-robust consistency learning + +To further enhance the domain robustness, we encourage model to generate persistent predictions in the domain shift scenario. To enhance prediction consistency, we induce the model to minimize the prediction discrepancy in terms of the mask and class label as follows: + +$$ +\mathcal {L} _ {\mathrm {c o n s}} = \sum_ {s = 1} ^ {S} \lambda_ {\mathrm {m c}} \cdot \mathcal {L} _ {\mathrm {m c}} \left(\hat {y} _ {\mathrm {s}} ^ {\text {m a s k}}, \hat {y} _ {\mathrm {s}} ^ {\text {m a s k}}\right) + \lambda_ {\mathrm {c c}} \cdot \mathcal {L} _ {\mathrm {c c}} \left(\hat {c} _ {q _ {i}, \mathrm {s}}, \hat {c} _ {q _ {i} ^ {\prime}, \mathrm {s}}\right). \tag {2} +$$ + +$S$ denotes the number of transformer blocks in the transformer decoder, and $\mathcal{L}_{\mathrm{mc}}$ and $\mathcal{L}_{\mathrm{cc}}$ represent the mask and class consistency losses, respectively. To compute the loss at the $s$ -th transformer decoder block, we obtain $\{\hat{y}_{\mathrm{s}}^{\mathrm{mask}}, \hat{y}_{\mathrm{s}}^{\mathrm{mask}}\}$ and $\{\hat{c}_{q_i,s}, \hat{c}_{q_i',s}\}$ which are pairs of mask predictions and class predictions of $i$ -th query $q_i$ from the original and augmented image pair $\{x, x'\}$ , respectively. We employ binary cross entropy and Jensen-Shannon divergence as a discrepancy measure of the mask consistency + +loss $(\mathcal{L}_{mc})$ and class consistency loss $(\mathcal{L}_{cc})$ , respectively. With the help of domain-aware context prompts and above losses, our model learns to predict not only accurately but also consistently in various domain shift scenarios during training. + +# 3.3. Overall Loss Functions + +The overall loss of our framework is a weighted sum of the task loss $\mathcal{L}_{\mathrm{seg}}$ , VLM regularization loss $\mathcal{L}_{\mathrm{reg}}$ , domain-aware contrastive loss $\mathcal{L}_{\mathrm{contra}}$ , and the consistency loss $\mathcal{L}_{\mathrm{cons}}$ . + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} _ {\text {s e g}} + \lambda_ {\text {r e g}} \mathcal {L} _ {\text {r e g}} + \lambda_ {\text {c o n t r a}} \mathcal {L} _ {\text {c o n t r a}} + \lambda_ {\text {c o n s}} \mathcal {L} _ {\text {c o n s}}, \tag {3} +$$ + +where $\{\lambda_{\mathrm{reg}},\lambda_{\mathrm{contra}},\lambda_{\mathrm{cons}}\}$ are constant weighting factors. We note that $\mathcal{L}_{\mathrm{seg}}$ and $\mathcal{L}_{\mathrm{cons}}$ are calculated with all queries and its predictions from every block of the transformer decoder. We provide details of the baseline losses $(\mathcal{L}_{\mathrm{seg}},\mathcal{L}_{\mathrm{reg}})$ in the supplementary. + +# 4. Experiments + +# 4.1. Implementation Details + +Datasets. We validate DPMFormer on synthetic-to-real and real-to-real scenarios in the single-source setting. + +Synthetic datasets. GTAV [42] is a representative synthetic dataset which consists of 24,966 images with a resolution of $1914 \times 1052$ . The training split contains 12,403 images, while the validation and test set are of 6,382 and 6,181 images, respectively. SYNTHIA [43] dataset provides 6,580 images for training and 2,820 images for validation respectively, with the image resolution at $1280 \times 760$ . + +Real-world datasets. Cityscapes [7] is a dataset collected from the real environment. The resolution of each image is $2048 \times 1024$ , and the population of training and validation split is 2,975 and 500, respectively. BDD100K [58] includes 7,000 training images and 1,000 validation images with the resolution of $1280 \times 720$ . Mapillary [34] is composed of images with diverse resolution, where the training and validation size is 18,000 and 2,000, respectively. + +Network architecture. Our approach leverages the vision transformer-based models as backbones, initialized with either the CLIP [41] (ViT-B) or EVA02-CLIP [45] (EVA02-L) model. The CLIP backbone is configured with a patch size of 16, while the EVA02-CLIP backbone uses a patch size of 14. For the pixel and transformer decoder, aforementioned, we employ a mask classification architecture [5], consisting of $N = 9$ layers with masked attention mechanisms. Additionally, we design the auxiliary network $h_{\theta}$ for domain-aware context prompt generation as a shallow multi-layer structure (BatchNorm-Linear-ReLU-Linear). + +Training and evaluation. To optimize DPMFormer, we employ an AdamW [32] where the learning rate is set as $1 \times 10^{-5}$ and $1 \times 10^{-4}$ for synthetic and real training datasets, respectively. Following previous transformer-based studies [18, 36], we apply linear warm-up [15] for + +
Models (GTAV)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10145.3341.1840.7742.43
WildNet [28]ResNet-10145.7941.7347.0844.87
SHADE [61]ResNet-10146.6643.6645.5045.27
TLDR [27]ResNet-10147.5844.8848.8047.09
FAMix* [13]ResNet-10149.4746.4051.9749.28
SHADE [61]MiT-B553.2748.1954.9952.15
IBAFormer [44]MiT-B556.3449.7658.2654.79
VLTSeg* [21]ViT-B47.5045.7054.3049.17
TQDM* [36]ViT-B57.5047.6659.7654.97
DPMFormer* (ours)ViT-B59.0051.8063.6258.14
VLTSeg** [21]EVA02-L65.6058.4066.5063.50
Rein** [55]EVA02-L65.3060.5064.9063.60
Rein†ViT-L66.4060.4066.1064.30
TQDM** [36]EVA02-L68.8859.1870.1066.05
DPMFormer** (ours)EVA02-L70.0860.4870.6667.07
+ +Table 1. Comparison with the state-of-the-art DGSS methods on synthetic-to-real scenario with GTAV [42] as a source. We note that *, **, and † denotes the models initialized with pretrained CLIP [41], EVA02-CLIP [45], and DINOv2 [35], respectively. The best and the second best performances are highlighted with bold and underline, respectively. + +
Models (Synthia)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10140.8735.9837.2638.04
TLDR [27]ResNet-10142.6035.4637.4638.51
IBAFormer [44]MiT-B550.9244.6650.5848.72
VLTSeg** [21]EVA02-L56.8050.5054.5053.93
TQDM** [36]EVA02-L57.9952.4354.8755.10
DPMFormer** (ours)EVA02-L58.9254.3960.0857.80
+ +initial 1,500 iterations and rare class sampling [18]. The optimizer settings are identical for both CLIP and EVA02-CLIP backbone models. We set the total training iterations and the batch size as 20,000 and 8, respectively. Weighting factors $\{\lambda_{\mathrm{reg}},\lambda_{\mathrm{contra}},\lambda_{\mathrm{cons}}\}$ are set as 1, 1, 10. We crop the input image to have a resolution of $512\times 512$ . The texture perturbation is only applied during training. We use mean Intersection over Union (mIoU) [12] to quantitatively evaluate the results following the convention. + +# 4.2. Quantitative Results + +Synthetic-to-real. As shown in Tab. 1, we achieve state-of-the-art in every target domain with both backbones. Notably, with the ViT-B backbone initialized with pretrained CLIP [41], DPMFormer consistently surpasses previous state-of-the-art [36] by $3.17\%$ on average mIoU among target domains. In detail, on Cityscapes which contains mostly daytime real images, the domain robustness empowered by consistency learning assists DPMFormer to cope with the domain gap caused by visual realism, improving the state-of-the-art performance by $1.5\%$ . Meanwhile on BDD100K and Mapillary which have higher environmental variation in terms of weather, time and location, the domain-aware prompt generation enables the model to adaptively leverage + +Table 2. Comparison with the state-of-the-art DGSS methods where SYNTHIA [43] is set as a source dataset. + +
Models (Cityscapes)BackboneBDDMapillaryAvg.
SAN-SAW [39]ResNet-10154.7361.2742.43
WildNet [28]ResNet-10147.0141.7344.87
SHADE [61]ResNet-10150.9543.6645.27
TQDM* [36]ViT-B50.5465.7458.14
DPMFormer* (ours)ViT-B54.8167.7261.27
HGFormer [10]EVA02-L61.5072.1066.80
VLTSeg** [21]EVA02-L64.4076.4070.40
Rein** [55]EVA02-L64.1069.5066.80
Rein†ViT-L65.0072.3068.65
TQDM [36]**EVA02-L64.7276.1570.44
DPMFormer** (ours)EVA02-L64.276.6770.44
+ +Table 3. Comparison with the state-of-the-arts trained with Cityscapes [7] on the real-to-real scenario. + +the textual knowledge for segmentation, impressively escalating the performance by $4.14\%$ , and $3.17\%$ , respectively. Moreover, even with the larger backbone [45], we mark the highest average performance of $67.07\%$ . In Cityscapes and Mapillary dataset, we outperform TQDM [36] by $1.2\%$ and $0.56\%$ respectively, while nearly reaching the score of Rein [55] in BDD. The overall results demonstrate the superiority of DPMFormer, emphasizing the efficacy of domain-awareness as well as domain-robustness. + +In Tab. 2, we also present synthetic-to-real results where SYNTHIA [43] is set as a source domain. Again, DPM-Former outperforms all competitors by a large margin, achieving the state-of-the-art performance in every target domain. Remarkably, our framework surpasses the previous state-of-the-art [36] by an average of $2.7\%$ mIoU across target domains. In particular, the performance on Mapillary [34] improves significantly by $5.21\%$ . These results demonstrate the effectiveness of DPM-Former in addressing the domain gap in terms of texture and perspective. + +Real-to-real. Furthermore, as shown in Tab.3, DPMFormer records the highest average mIoU with both backbones. With the CLIP-pretrained backbone, we significantly boost the state-of-the-art performance by $4.27\%$ and $1.98\%$ on BDD and Mapillary, respectively. These results indicate that enhanced domain robustness supports the model in predicting consistently under environmental changes, while the domain-aware prompts facilitate the semantic alignment on the unseen domain. As a result, DPMFormer successfully achieves an average performance gain of $3.13\%$ over the previous state-of-the-art. Equipped with the EVA-CLIP [45] pretrained backbone, we attain the highest mIoU of $76.67\%$ on Mapillary, while performing slightly lower on BDD. Overall, the average performance is $70.44\%$ , which is comparable to TQDM [36]. + +# 4.3. Qualitative Results + +In Fig. 4, we qualitatively compare our method with FAMix [13] and TQDM [36] on the synthetic-to-real sce + +![](images/87cd7074766a0a9379133f1848d6d5a51b55df0de2bce5ca0d3133ae9c034bf1.jpg) +Figure 4. Qualitative comparison on synthetic-to-real scenario with the CLIP-pretrained backbone (ViT-B). The training source domain is set as GTAV [42] while the target domains are BDD100K [58] and Mapillary [34]. The overall result shows that DPMFormer accomplishes precise segmentation with the images of strong illumination contrast as well as confusing textures. + +nario with the pretrained CLIP backbone (ViT-B). As depicted in the results on BDD100K, both competitors struggle to discriminate accurately under environments having confusing textures or large variations of illumination. In particular, the road in the first-row image has a texture similar to the exterior of a car, leading TQDM to a mismatch between the visual and textual features. In addition, FAMix shows sensitivity to the shades, mislabeling the road as a sidewalk. On the contrary, our method effectively distinguishes the road from other categories by reflecting the domain-specific properties of real-world roads in clear weather. In the case of the second row image with the high illumination contrast, both FAMix and TQDM fail to notice sidewalks in the image due to their low brightness. On the other hand, DPMFormer carries out precise predictions under severe photometric changes owing to the domain robustness acquired from handling diverse texture changes. + +As observed with the samples from the Mapillary (the third and fourth rows), FAMix and TQDM lack discriminability between the road and the sidewalk. To be specific with the first image (third row), these classes appear to have a similar color and texture, resulting in misclassifications. Meanwhile, the domain-specific properties from the image transfer the semantic knowledge of the appearance of real-world sidewalk to the model, DPMFormer accurately distinguish between the road and the sidewalk. With the second image, TQDM [36] shows vulnerability against small textural changes on the sidewalk (e.g., shades of the poles). Contrarily, DPMformer effectively performs segmentation ow + +ing to its domain-robustness as well as the obtained domain-aware textual queries. + +# 4.4. Analysis + +# 4.4.1. Ablation studies + +To demonstrate the effectiveness of the components in DPMFormer, i.e., texture perturbation, domain-robust consistency learning $(\mathcal{L}_{cons})$ , and domain-aware context prompt learning $(\mathcal{L}_{contra})$ , we conduct ablation studies with the CLIP pretrained backbone on synthetic-to-real scenario. As presented in Tab. 4, every component contributes to the performance gain. Specifically, the texture perturbation enlarges the observable domain during training, boosts the average performance by $0.65\%$ . In addition with the $\mathcal{L}_{cons}$ , the model successfully equips the robustness against domain-shift and enhances the prediction accuracy on unseen target domains, increasing the average mIoU by $0.77\%$ . Furthermore, $\mathcal{L}_{contra}$ allows the model to acquire domain-specific property from the input image in the form of textual prompt embedding. Consequently, the model exploits both visual and textual cues properly aligned for the target domain, considerably elevate the average mIoU by $1.99\%$ . Combining all components together, DPMFormer effectively learns both domain-awareness and domain-robustness to generate accurate results consistently on unseen target domains. + +# 4.4.2. Comparison with prompt learning methods + +In Tab. 5, we compare our domain-aware context prompt learning $(\mathcal{L}_{\mathrm{contra}})$ with previous prompt learning methods [25, 26, 63] to validate the efficacy for domain gen + +
Models (GTAV)CityscapesBDDMapillaryAvg.
Baseline57.547.6659.7654.97
+ Perturbation57.0448.1960.9155.38
+ Lcons58.2249.3960.8456.15
+ Lcontra59.051.863.6258.14
+ +![](images/81442a0034263349e47936b3c2e9035b557ce3b03bd18944350598372fe43f34.jpg) +Figure 5. Qualitative results on diverse styles, i.e., Minimalist, Pop Art, Bauhaus, and Cubism. The models are trained on GTAV with CLIP pretrained backbone (ViT-B). + +deralization. Given the textual features $t = ENC_{T}(p_{x})$ and visual features $v = \hat{F}(x)$ , the contrastive loss of CoCoOp [63] encourages the similarity between the anchor text feature and its corresponding visual feature to be higher than others. As observed in the second row, CoCoOp marginally improves the performance of prompt generator $h_{\theta}$ , indicating that the generated feature becomes instancespecific which is less generalizable to other domains. Moreover, additionally assigning the visual feature from augmented sample as positives $(\mathrm{CoCoOp}^{+})$ resulted in negligible performance gain, since the derived prompt contains domain-invariant information rather than domain-specific properties. Furthermore, although both MaPLe [25] and PromptSRC [26] employ multi-modal prompts, their performance gains are modest due to the limited generalizability of fixed prompts. On the other hand, our domain-aware contrastive loss design enable the model to proficiently capture domain-specific property from the image which are more helpful for the semantic alignment between the visual features and the textual knowledge. From the fifth to the eighth row, we verify $\mathcal{L}_{\text{contra}}$ with different similarity calculation targets for $\sin(\cdot)$ , i.e., text features $t$ , text and visual features $(t, v)$ , and (3) output context embeddings $\pi$ . The results confirm that computing the domain-aware contrastive loss empowers the domain generalizability especially when computed with the context embeddings $\pi$ which can provide direct domain guidance to $h_{\theta}$ . + +Table 4. Ablation study of proposed components. The models are trained on GTAV with CLIP pretrained backbone (ViT-B). The best performance for each column is highlighted in bold. + +
\( \mathcal{L}_{\text{contra}} \) (GTAV)CityscapesBDDMapillaryAvg.
-57.6549.6361.1056.13
CoCoOp [63]57.8449.9161.3356.36
\( \mathrm{CoCoOp}^{+} \)57.6050.0460.9656.20
MaPLe [25]57.8750.1261.0456.34
PromptSRC [26]58.1049.7362.5156.78
Ours (sim(t,t))58.1750.1861.3756.57
Ours (sim(t,v))59.4950.3462.1857.34
Ours (sim(π,π))59.0051.8063.6258.14
+ +Table 5. Comparison with prompt learning methods [25, 26, 63] in synthetic-to-real scenario. - denotes DPMFormer without $\mathcal{L}_{\mathrm{contra}}$ . $\mathrm{CoCoOp}^{+}$ indicates the modified loss that additionally include the feature from augmented sample of anchor as positives. $t$ , $v$ , and $\pi$ refers to the text feature, visual feature, and context embedding. + +# 4.4.3. Qualitative results on diverse styles + +To further verify the domain generalization capability of DPMFormer, we compare our model with TQDM [36] on images with diverse artistic styles generated by $\mathrm{ChatGPT}^1$ . As shown in Fig. 5, DPMFormer correctly predicts objects and their surroundings even in severe domain shift scenarios, e.g., Cubism. With modern art styles (i.e., Minimalist, Pop Art, and Bauhaus), TQDM perplexes among road, sidewalk, and terrain due to their textural changes. Notably, DPMFormer produces more reliable results owing to the enhanced robustness against texture variations that are learned from the texture perturbation and consistency learning. In case of the scene in cubism style, TQDM mispredicts the person on the left side because of its visual similarity with the person in the traffic sign. On the other hand, our method reflects the cubism style to the textual object queries, making an accurate classification with the same instance. + +# 5. Conclusion + +In this paper, we presented DPMFormer, a novel domain generalization framework for semantic segmentation. To address the limited generalizability of fixed context prompt learned from a single source dataset, we introduced a novel domain-aware prompt learning which can reflect domain-specific properties of the input image into textual prompts. The proposed component enhanced the semantic alignment between visual and textual cues, assisting the model to fully leverage the abundant semantic knowledge of VLMs. Moreover, to empower domain-robustness, we simulated various domain shifts via texture perturbations, and provided consistency guidance to the model by minimizing prediction discrepancies between original and augmented images. Through extensive experiments and analyses, we demonstrated the effectiveness of DPMFormer, achieving state-of-the-art performance on various benchmarks. + +Acknowledgments. This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364). This research was supported by Sookmyung Women's University Research Grants (1-2403-2034). + +# References + +[1] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2 +[2] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 1, 2 +[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 819-827, 2024. 1, 3 +[4] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19288-19300, 2023. 1, 3 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 3, 4, 5 +[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 1, 3, 4 +[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 1, 3, 7, 8 +[8] Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. Domain adaptation in computer vision applications, pages 1-35, 2017. 3 +[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3 +[10] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15413-15423, 2023. 1, 3, 6 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, + +Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4 +[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 6 +[13] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23428-23437, 2024. 1, 3, 6 +[14] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023. 1, 3 +[15] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5 +[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3, 1 +[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 3 +[18] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 5, 6 +[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 1, 3 +[20] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 3 +[21] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. arXiv preprint arXiv:2312.02021, 2023. 1, 3, 6 +[22] Xinyue Huo, Lingxi Xie, Hengtong Hu, Wengang Zhou, Houqiang Li, and Qi Tian. Domain-agnostic priors for semantic segmentation under unsupervised domain adaptation and domain generalization. International Journal of Computer Vision, 132(9):3954-3976, 2024. 3 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International + +conference on machine learning, pages 4904-4916. PMLR, 2021.3 +[24] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18993-19003, 2023. 1, 3 +[25] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113-19122, 2023. 7, 8 +[26] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15190-15200, 2023. 7, 8 +[27] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 677-687, 2023. 1, 3, 6 +[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9936-9946, 2022. 1, 3, 6 +[29] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5543-5551, 2017. 1 +[30] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3 +[31] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 1 +[32] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5 +[33] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3 +[34] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 5, 6, 7, 3, 8 +[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6 + +[36] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. European Conference on Computer Vision, 2024. 1, 3, 4, 5, 6, 7, 8, 2 +[37] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 3 +[38] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1863-1871, 2019. 3 +[39] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2594-2605, 2022. 1, 3, 6 +[40] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for zero-shot transfer learning. Neurocomputing, 555: 126658, 2023. 3 +[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6 +[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 2, 5, 6, 7, 3, 4 +[43] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 5, 6 +[44] Qiyu Sun, Huilin Chen, Meng Zheng, Ziyan Wu, Michael Felsberg, and Yang Tang. Ibaformer: Intra-batch attention transformer for domain generalized semantic segmentation. arXiv preprint arXiv:2309.06282, 2023. 6 +[45] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 3, 5, 6 +[46] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3 +[47] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3 +[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 1 +[49] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker. + +Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 3 +[50] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 3 +[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3 +[52] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 1, 2 +[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In The Nineth International Conference on Learning Representations (ICLR 2021). OpenReview, 2021. 3 +[54] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3 +[55] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28619-28630, 2024. 1, 3, 6 +[56] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 1, 3 +[57] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8978-8987, 2021. 3 +[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 2, 5, 7, 3, 4, 6 +[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 1, 3 +[60] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2 + +[61] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In European conference on computer vision, pages 535-552. Springer, 2022. 1, 3, 6 +[62] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. Advances in neural information processing systems, 35:338-350, 2022. 1, 3 +[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 5, 7, 8, 2 +[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4 +[65] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3 +[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3 \ No newline at end of file diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/images.zip b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..fab382baaae576cd20433044ed7dd611b04b2396 --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fe9ca00b6756f0506fe62b4b94216f1aff74925419bd7c9c08a0ceaa9ed8a85e +size 566905 diff --git a/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/layout.json b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..758faa28802dd667000eb8a854acd244c311bd34 --- /dev/null +++ b/2025/Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation/layout.json @@ -0,0 +1,9213 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "spans": [ + { + "bbox": [ + 65, + 102, + 547, + 140 + ], + "type": "text", + "content": "Exploiting Domain Properties in Language-Driven Domain Generalization for Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "spans": [ + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "text", + "content": "Seogkyu Jeon" + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "text", + "content": " Kibeom Hong" + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "text", + "content": " Hyeran Byun" + }, + { + "bbox": [ + 181, + 160, + 434, + 176 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "spans": [ + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "type": "text", + "content": "Yonsei University " + }, + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 165, + 180, + 445, + 196 + ], + "type": "text", + "content": "Sookmyung Women's University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 247, + 198, + 362, + 209 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 247, + 198, + 362, + 209 + ], + "spans": [ + { + "bbox": [ + 247, + 198, + 362, + 209 + ], + "type": "text", + "content": "jone9312@yonsei.ac.kr" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 152, + 236, + 200, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 236, + 200, + 249 + ], + "spans": [ + { + "bbox": [ + 152, + 236, + 200, + 249 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 263, + 297, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 263, + 297, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 263, + 297, + 526 + ], + "type": "text", + "content": "Recent domain generalized semantic segmentation (DGSS) studies have achieved notable improvements by distilling semantic knowledge from Vision-Language Models (VLMs). However, they overlook the semantic misalignment between visual and textual contexts, which arises due to the rigidity of a fixed context prompt learned on a single source domain. To this end, we present a novel domain generalization framework for semantic segmentation, namely Domain-aware Prompt-driven Masked Transformer (DPM-Former). Firstly, we introduce domain-aware prompt learning to facilitate semantic alignment between visual and textual cues. To capture various domain-specific properties with a single source dataset, we propose domain-aware contrastive learning along with the texture perturbation that diversifies the observable domains. Lastly, to establish a framework resilient against diverse environmental changes, we have proposed the domain-robust consistency learning which guides the model to minimize discrepancies of prediction from original and the augmented images. Through experiments and analyses, we demonstrate the superiority of the proposed framework, which establishes a new state-of-the-art on various DGSS benchmarks." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 552, + 135, + 563 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 572, + 296, + 693 + ], + "type": "text", + "content": "Over the decades, semantic segmentation has made remarkable progress, now being able to precisely classify each pixel in an image into categories. However, one of the shadows that lies in these advancements is that models often exhibit inconsistent and degraded performance when deployed in various real-world environments. To this end, the task of Domain Generalized Semantic Segmentation (DGSS) has arisen to overcome the discrepancy between the training and test domains, i.e., domain shift. DGSS has progressed through diverse approaches, such as feature whiten-" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 321, + 236, + 451, + 336 + ], + "blocks": [ + { + "bbox": [ + 321, + 236, + 451, + 336 + ], + "lines": [ + { + "bbox": [ + 321, + 236, + 451, + 336 + ], + "spans": [ + { + "bbox": [ + 321, + 236, + 451, + 336 + ], + "type": "image", + "image_path": "4d41545c8df7c1ea7e385bccc12f726a643c721caffa29a89c4052f18bfd71b5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 353, + 342, + 430, + 351 + ], + "lines": [ + { + "bbox": [ + 353, + 342, + 430, + 351 + ], + "spans": [ + { + "bbox": [ + 353, + 342, + 430, + 351 + ], + "type": "text", + "content": "Fixed context prompts [31, 36]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 452, + 236, + 548, + 336 + ], + "blocks": [ + { + "bbox": [ + 452, + 236, + 548, + 336 + ], + "lines": [ + { + "bbox": [ + 452, + 236, + 548, + 336 + ], + "spans": [ + { + "bbox": [ + 452, + 236, + 548, + 336 + ], + "type": "image", + "image_path": "ddfc0a3c25bed2213266027f02423d85c1a347d28ca61f056f3a2181431b441d.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 342, + 523, + 350 + ], + "lines": [ + { + "bbox": [ + 473, + 342, + 523, + 350 + ], + "spans": [ + { + "bbox": [ + 473, + 342, + 523, + 350 + ], + "type": "text", + "content": "DPMFormer (Ours)" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 361, + 555, + 417 + ], + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 417 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 417 + ], + "type": "text", + "content": "Figure 1. Motivation of DPMFormer. Using a fixed context prompts [31, 36] tend to retain source domain properties, causing contextual misalignment with the target domain. On the other hand, DPMFormer translates domain properties of the input image into context prompts, enhancing semantic alignments." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 555, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 486 + ], + "type": "text", + "content": "ing [6, 37-39] and domain randomization [4, 14, 19, 24, 27, 28, 56, 59, 61, 62]. Furthermore, some studies [3, 10] exploited multi-scale object queries with a transformer-based architecture, i.e., Mask2Former [5]." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 629 + ], + "type": "text", + "content": "Despite these advancements, learning domain-robust representations solely from the single source domain remains a significant hurdle to performance improvement. Recently, several studies [13, 21, 36, 55] employed Vision-Language Models (VLMs) [41, 45] owing to their semantic knowledge learned from diverse large-scale text-image datasets. For this, pioneering VLM-based DGSS works [13, 21, 55] have adopted their pre-trained visual encoder for initialization and fine-tuning [21, 55]. In addition, TQDM [36] has introduced a framework that constructs object queries from the textual descriptions, leveraging semantic concepts from linguistic expressions." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 556, + 715 + ], + "type": "text", + "content": "Yet, they still have limitations in fully utilizing textual knowledge to improve domain generalizability. Concretely, the context prompts that decorates textual descriptions of each category are either a predefined template [31] (e.g., 'a photo of') or a single learnable text embedding [36]. However, we contend that fixed context prompts have limited capability for target domain images due to the discrepancy" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 712 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20791" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 71, + 321, + 220 + ], + "blocks": [ + { + "bbox": [ + 82, + 71, + 321, + 220 + ], + "lines": [ + { + "bbox": [ + 82, + 71, + 321, + 220 + ], + "spans": [ + { + "bbox": [ + 82, + 71, + 321, + 220 + ], + "type": "image", + "image_path": "fd2224252574a5e01732501e2ef9a84ce30f390bbe6082512c594b780621a4c4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 145, + 222, + 252, + 232 + ], + "lines": [ + { + "bbox": [ + 145, + 222, + 252, + 232 + ], + "spans": [ + { + "bbox": [ + 145, + 222, + 252, + 232 + ], + "type": "text", + "content": "(a) Visualization of textual prompts" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 325, + 72, + 529, + 220 + ], + "blocks": [ + { + "bbox": [ + 325, + 72, + 529, + 220 + ], + "lines": [ + { + "bbox": [ + 325, + 72, + 529, + 220 + ], + "spans": [ + { + "bbox": [ + 325, + 72, + 529, + 220 + ], + "type": "image", + "image_path": "6875ebcbc8727141610ab081774bbaa924c05cf3f83ded44d4d4fb0193976df5.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 394, + 222, + 485, + 232 + ], + "lines": [ + { + "bbox": [ + 394, + 222, + 485, + 232 + ], + "spans": [ + { + "bbox": [ + 394, + 222, + 485, + 232 + ], + "type": "text", + "content": "(b) Results on BDD100K [58]" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "lines": [ + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 555, + 298 + ], + "type": "text", + "content": "Figure 2. PCA visualization of textual prompts (left) and qualitative results on various environments (e.g., Day, Dawn, Night) in BDD100K [58] (right). The models are trained on GTAV [42] with the CLIP-pretrained backbone (ViT-B) [41]. A fixed single-context prompt lacks flexibility in adapting to various domain shifts due to its rigidity. In contrast, our framework utilizes domain-specific properties from input images as context prompts, enhancing semantic alignment between text and images. As a result, as shown in (b), our approach exhibits improved robustness across diverse environments." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 318, + 294, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 318, + 294, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 318, + 294, + 413 + ], + "type": "text", + "content": "between visual and textual contexts as depicted in Fig. 1. Firstly, a handcrafted template inherently encodes the characteristics of a specific domain which restricts its generalizability to other unseen domains. Secondly, prompt optimization is prone to source domain overfitting, especially in the single-source setting. Consequently, both approaches may enlarge the gap between textual and visual contexts, leading to suboptimal performances on target domains." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 415, + 295, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 415, + 295, + 605 + ], + "spans": [ + { + "bbox": [ + 54, + 415, + 295, + 605 + ], + "type": "text", + "content": "In this perspective, we point out that the semantic form of the textual representation for an object category should be changed with respect to the input visual context in order to strengthen their semantic correspondence. For instance, in the real-world driving scene at night of Fig. 2 (b), target categories (e.g., car and sky) possess distinct textures from those of synthetic daytime images. Hence, it would be more appropriate to decorate 'a car' with the textual prompt 'at night in the real-world' which is reflecting the domain characteristics. As a result, the modified text feature will include details such as darkened car exteriors and light reflections, enhancing the semantic alignment. Nevertheless, it is challenging to design domain-adaptive prompts in the DGSS setting where the training dataset covers only a single domain. Moreover, this setting hinders the visual pipeline from learning robust image features against domain shifts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 605, + 295, + 714 + ], + "type": "text", + "content": "To address these challenges, we introduce a novel framework, namely Domain-aware Prompt-driven Masked Transformer (DPMFormer), focusing on two key aspects: (1) Leveraging domain-specific properties of the input image (domain-awareness) and (2) Generating accurate outputs on images with dissimilar domain characteristics (domain-robustness). To cultivate the domain-awareness, we propose a novel domain-aware prompt learning, which translates the domain-specific properties of an input image into con" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 318, + 555, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 462 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 462 + ], + "type": "text", + "content": "text prompts via an auxiliary network. In addition, to obtain diverse domain properties in the single-source setting, we apply texture perturbations to synthesize novel domain images. Furthermore, exploiting both source and novel domain images, we propose a domain-aware contrastive loss to ensure that the derived prompts effectively capture domain-specific properties of the input image. This loss encourages the anchor context prompts to be distinguishable from those of different domain characteristics, while being closer to those from the anchor domain. With the proposed domain-aware prompt, the model accurately identifies target classes while being aware of the input domain." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 556, + 635 + ], + "type": "text", + "content": "Moreover, we strive to provide better domain robustness guidances to the visual encoder and decoders. We carefully organize the texture perturbations with structure-preserving image transformations, ensuring that the original visual context remains intact. Instead of simply reusing the original ground truths for novel domain images, we introduce domain-robust consistency loss to guarantee reliable predictions under severe domain shifts. The loss consists of class consistency and mask consistency losses, which penalize discrepancies between class and mask predictions of given image pairs, respectively. Furthermore, domain-robust consistency losses are applied at every layer of the transformer decoder, preventing discrepancies in earlier layers from propagating to later parts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 715 + ], + "type": "text", + "content": "Through evaluations on various DGSS benchmarks, we demonstrate the superiority of the proposed framework. Notably, our framework achieves state-of-the-art semantic segmentation performance across domain generalization scenarios. Additionally, ablation studies and detailed analyses validate the effectiveness of each component." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20792" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 92, + 288, + 105 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 92, + 288, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 92, + 288, + 105 + ], + "type": "text", + "content": "2.1. Domain Generalized Semantic Segmentation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 110, + 295, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 110, + 295, + 243 + ], + "spans": [ + { + "bbox": [ + 55, + 110, + 295, + 243 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS) aims to learn domain-invariant representations that generalize robustly to various unseen target domains. The task assumes a single-source setting, where only one dataset is available during training. Unlike domain adaptation [8, 49, 66] and test-time domain adaptation [53, 54, 57], access to target domains is strictly prohibited, making DGSS more challenging. Previous studies approached DGSS in two ways primarily: feature whitening and normalization [6, 37-39], and domain randomization approaches [4, 14, 19, 24, 27, 28, 37, 38, 56, 59, 61, 62]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 244, + 295, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 244, + 295, + 411 + ], + "spans": [ + { + "bbox": [ + 55, + 244, + 295, + 411 + ], + "type": "text", + "content": "Feature whitening and normalization approaches [6, 37-39] mainly focus on removing features which are variant to domain shifts. Exploiting the characteristics of instance whitening [30] and instance normalization [50] which can effectively remove texture and style from the input image, these approach leverages those operations in between the backbone module to minimize the effect from domain and texture changes. Representatively, RobustNet [6] introduced instance selective whitening module that finds gram matrix components sensitive to photometric changes and minimize their changes. However, due to the natural difficulty in disentangling domain-specific and domain-invariant features, these approaches show limited performance gains." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 616 + ], + "type": "text", + "content": "Domain randomization approaches [4, 14, 19, 24, 27, 28, 56, 59, 61, 62] augment novel domains from the source domain by modifying either the images or their features through various methods, e.g. affine transformations [14, 20, 28, 56, 61, 62], image translation [59, 65], frequency decomposition [4, 19], and photometric transformations [24]. Synthesized samples increase the domain diversity of the training dataset, reducing the domain gap between the learned representation and the test data. Moreover, most of the generative DGSS methods adopt context-preserving transformations to compute the output discrepancy between the original sample and the augmented one. Representatively, SHADE [61] proposed a style consistency loss to encourage model to learn invariant pixel-level semantic information by minimizing the Jenson-Shannon Divergence (JSD) between the output predictions of original and augmented images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 715 + ], + "type": "text", + "content": "Meanwhile, several studies [3, 10] have built DGSS upon Mask2Former [5], that leverages attention mechanism [51] renowned for robustness against domain shifts [16, 17, 46]. Mask2former [5] involves a transformer decoder that exploits object query features to group pixels of same objects or categories. Based on this, HGFormer [10] first proposed a hierarchical framework that groups pixels to form part-level masks for complementing whole-level pixel group-" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "ing procedure. Similarly, CMFormer [3] utilizes downsampled features additionally via feature fusion which are more domain-invariant than the original features." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 118, + 553, + 143 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 118, + 553, + 143 + ], + "spans": [ + { + "bbox": [ + 313, + 118, + 553, + 143 + ], + "type": "text", + "content": "2.2. Language-driven Domain Generalized Semantic Segmentation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 148, + 555, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 148, + 555, + 472 + ], + "spans": [ + { + "bbox": [ + 313, + 148, + 555, + 472 + ], + "type": "text", + "content": "Recent studies [13, 21, 22, 36, 55] has investigated to exploit Vision Language Models (VLMs) [9, 23, 33, 40, 41, 45, 47] for DGSS, owing to its powerful generalizability learned from large-scale datasets of image-text pairs. VLTseg [21] employs the image encoder of VLMs for initializing backbone parameters, and fine-tune the network with the segmentation objective function. FAMix [13] proposes to yield class-specific novel styles by concatenating random style description and class names, then performs style randomization by locally mixing the source and the novel styles. DAP [22] exploits text encoder to distill the semantic textual knowledge of target categories to the visual backbone. TQDM [36] points out deficient use of language information in aforementioned works, and introduces a textual-query driven framework for DGSS. They utilize the textual description of each categories for initializing input query features for transformer decoder, and also for computing the text-to-pixel attention to enhance semantic clarity of pixel features. Despite their advancements, we argue that two points are overlooked: (1) the domain discrepancy between the learned textual prompt and the domain-specific property of the input image (2) lack of domain robustness guidance. To mitigate these limitations, we carefully refine the textual queries with the captured domain-specific property of the input image, and augment novel domain images to improve domain-awareness of the model as well as prediction consistency learning." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 483, + 375, + 496 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 483, + 375, + 496 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 375, + 496 + ], + "type": "text", + "content": "3. Methods" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 400, + 516 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": "We design our DPMFormer based on mask classification architecture, i.e., Mask2Former [5]. Mask2Former consists of an image encoder " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "ENC_I" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ", pixel decoder " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "DEC_{\\mathrm{pix}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " and transformer decoder " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "DEC_{\\mathrm{tr}}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ". The input RGB image " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "x \\in \\mathbb{R}^{3 \\times H \\times W}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " is first fed to the image encoder for feature extraction, and then converted to pixel-wise features " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "z \\in \\mathbb{R}^{D \\times H \\times W}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " through the pixel decoder. The transformer decoder iteratively refines " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " object queries " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "q \\in \\mathbb{R}^{N \\times L}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " with multi-scale image features from the pixel decoder. Thereafter, refined object queries are projected into mask embeddings " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "m \\in \\mathbb{R}^{N \\times D}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " to generate pixel-level mask prediction via dot product as " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}^{\\mathrm{mask}} = m \\cdot z" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "\\hat{y}^{\\mathrm{mask}} \\in \\mathbb{R}^{N \\times H \\times W}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ". The category label of each object query " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "q" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": " is predicted with a linear classifier, i.e., " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "c_q \\in \\mathbb{R}^K" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "inline_equation", + "content": "c \\in \\mathbb{R}^{N \\times K}" + }, + { + "bbox": [ + 313, + 521, + 555, + 715 + ], + "type": "text", + "content": ". The final prediction result is derived by the matrix multiplication between the pixel-level mask prediction and class prediction" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20793" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 68, + 558, + 249 + ], + "blocks": [ + { + "bbox": [ + 57, + 68, + 558, + 249 + ], + "lines": [ + { + "bbox": [ + 57, + 68, + 558, + 249 + ], + "spans": [ + { + "bbox": [ + 57, + 68, + 558, + 249 + ], + "type": "image", + "image_path": "3f416f43c54e57657e6fdc211311c1c3745cf26c4ef7318bcde770565afa1646.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": "Figure 3. Illustration of DPMFormer. We use Mask2Former [5] based architecture which consists of a backbone image encoder " + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "inline_equation", + "content": "(ENC_I)" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": ", a pixel decoder " + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "inline_equation", + "content": "(DEC_{\\mathrm{pix}})" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": ", a transformer decoder " + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "inline_equation", + "content": "(DEC_{\\mathrm{tr}})" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": ", and a text encoder " + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "inline_equation", + "content": "(ENC_T)" + }, + { + "bbox": [ + 54, + 258, + 555, + 304 + ], + "type": "text", + "content": ". During training, we synthesize images with a novel domain style via texture perturbation. Both images are incorporated to compose a batch and exploited for learning domain-awareness (Sec. 3.2.1) and domain-robustness (Sec. 3.2.2)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 322, + 237, + 336 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 322, + 237, + 336 + ], + "spans": [ + { + "bbox": [ + 55, + 322, + 237, + 336 + ], + "type": "text", + "content": "of queries as " + }, + { + "bbox": [ + 55, + 322, + 237, + 336 + ], + "type": "inline_equation", + "content": "\\hat{y} = c^{\\top}\\cdot \\hat{y}^{\\mathrm{mask}},\\hat{y}\\in \\mathbb{R}^{K\\times H\\times W}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": "Furthermore, inspired by [36], we employ a text encoder " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "ENC_{T}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " for initializing object queries from textual descriptions. Pre-trained vision language mdoels (VLMs), e.g., CLIP [41], is exploited to initialize both " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "ENC_{I}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "ENC_{T}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " parameters, and the parameters of " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "ENC_{T}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " remain frozen during training to preserve learned linguistic knowledge. The text encoder " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "ENC_{T}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " produces a text feature " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "t_k \\in \\mathbb{R}^C" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " corresponding to each class label " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "\\mathrm{class}_k" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " with a learnable context prompt " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": ", i.e., " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "t_k = ENC_{T}([p, \\{\\mathrm{class}_{k}\\}])" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": ". Then, obtained text embeddings " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "t = \\{t_k\\}_{k=1}^{K} \\in \\mathbb{R}^{K \\times C}" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": " are passed to a multi-layer perceptron (MLP) to derive initial textual object queries " + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "inline_equation", + "content": "q_t^0" + }, + { + "bbox": [ + 54, + 336, + 296, + 504 + ], + "type": "text", + "content": ". Also, the text-to-pixel attention mechanism is established in a pixel decoder layer to enhance pixel semantic clarity enhancement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 514, + 139, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 139, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 139, + 526 + ], + "type": "text", + "content": "3.2. DPMFormer" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 652 + ], + "type": "text", + "content": "As illustrated in Fig. 3, Domain-aware Prompt-driven Masked Transformer (DPMFormer), aims to cultivate two core aspects of domain generalization for the model: domain-awareness and domain-robustness. For domain-awareness, we propose the domain-aware context prompt learning which translates domain-specific property of the input image into the text embedding for semantic alignment between textual and visual features. And for domain-robustness, we encourage the model to generate accurate outputs against textural changes via consistency learning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Texture perturbation. In order to effectively guide the model with diverse domain characteristics, we stylize the training dataset to generate an auxiliary domain dataset. Following RobustNet [6], we adopt photometric transformations—comprising strong color jittering, gaussian blur" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "text", + "content": "and noise injection—for their simplicity as well as content-preserving property. The generated image " + }, + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "inline_equation", + "content": "x'" + }, + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "text", + "content": " is combined with its original image " + }, + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 324, + 555, + 361 + ], + "type": "text", + "content": " to form a batch for training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 364, + 512, + 376 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 512, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 512, + 376 + ], + "type": "text", + "content": "3.2.1. Domain-aware context prompt learning" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 378, + 555, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 617 + ], + "type": "text", + "content": "Textual query generation and prompt learning [64] plays a crucial role in enabling the model to leverage the linguistic semantics of each class previously learned by the VLM. Although the learned context prompt is beneficial for performance improvement, it is optimized solely within the source domain without direct consideration of domain shifts. This single context prompt may yield strong results on several domain shift scenarios where target domain has similar domain characteristics with the source. However, in cases of severe domain shift, the performance gain may be limited due to the contextual mismatch between the target domain image and the learned prompt. For example, if the prompt learning is conducted on a dataset containing only sunny day images, the learned prompt is could be misaligned when encountering rainy night scenes. As such, the semantic misalignment should be addressed in order to fully utilize rich semantic knowledge of pre-trained VLM. To this end, we propose to generate domain-aware context prompt that extracts domain-specific properties from the input image as a context prompt." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": "To obtain domain-aware context prompt from the input image " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": ", we design an auxiliary network " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "h_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": " that takes visual feature as an input and generates a domain-specific prompt embedding " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_x = h_\\theta (\\hat{F} (x))" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{F} (x)" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": " denotes a visual feature extracted from a frozen visual backbone of CLIP [41]. We use the class token as a visual feature " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\hat{F} (x)" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": " for its global representation [11]. Nextly, the obtained domain-specific prompt embedding " + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\pi_{x}" + }, + { + "bbox": [ + 313, + 618, + 556, + 715 + ], + "type": "text", + "content": " is in" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20794" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": "tegrated with the context prompt embedding " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "p" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": " through addition, i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "p_x = p + \\pi_x" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ", then concatenated with text embeddings of classes to generate domain-aware textual features as " + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "inline_equation", + "content": "t_{x,k} = ENC_T([p_x,\\{\\mathrm{class}_k\\}])" + }, + { + "bbox": [ + 55, + 72, + 296, + 168 + ], + "type": "text", + "content": ". To encourage derived textual features to include domain-specific information of the input image, we introduce a novel domain-aware contrastive learning framework with the original and augmented images. The loss function is depicted as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 62, + 188, + 295, + 231 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 188, + 295, + 231 + ], + "spans": [ + { + "bbox": [ + 62, + 188, + 295, + 231 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {c o n t r a}} = - \\frac {1}{2 B} \\sum_ {i = 1} ^ {2 B} \\log \\frac {\\sum_ {j \\in \\mathcal {P} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}{\\sum_ {j \\in \\mathcal {P} _ {i} \\cup \\mathcal {N} _ {i}} \\exp \\sin \\left(\\pi_ {i} , \\pi_ {j}\\right) / \\tau}, \\tag {1}", + "image_path": "4543cb1f01db81bfcfc17f401b6123ee303b472e740169fc7932793079eb8ffe.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\cdot)" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " means similarity metric and " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " is a temperature parameter. " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " denotes the batch size of original images, " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " are positive sets and negative sets of the " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": "-th image, respectively. The positive set " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " is composed of the indices of samples having same domain characteristics with the anchor " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " whereas the others belong to negative set " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": ". For example, when an anchor " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " is of original source domain images, other original source domain images belongs to " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " while all augmented images are included in " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": ". In case of an augmented image as an anchor, " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i = \\{i\\}" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{N}_i = \\{1,\\dots,2B\\} \\setminus i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": ". With the proposed loss, we encourage " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "h_\\theta" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " to capture domain-specific property of the image and reflect it to the output text feature " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "t_k" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": ". Also, the final domain-aware text feature is guided by the task loss " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{seg}" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " to be aligned with the original image and to minimize segmentation errors. We note that the proposed loss considers the domain information unlike CoCoOp [63] which treats " + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_i" + }, + { + "bbox": [ + 56, + 231, + 296, + 470 + ], + "type": "text", + "content": " as the text feature corresponding to a specific image feature while the remaining images forming the negative set. We provide comparative analysis with CoCoOp in Sec.4.4.2." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 478, + 237, + 490 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 237, + 490 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 237, + 490 + ], + "type": "text", + "content": "3.2.2. Domain-robust consistency learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 493, + 296, + 552 + ], + "type": "text", + "content": "To further enhance the domain robustness, we encourage model to generate persistent predictions in the domain shift scenario. To enhance prediction consistency, we induce the model to minimize the prediction discrepancy in terms of the mask and class label as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 561, + 296, + 604 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 604 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c o n s}} = \\sum_ {s = 1} ^ {S} \\lambda_ {\\mathrm {m c}} \\cdot \\mathcal {L} _ {\\mathrm {m c}} \\left(\\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}, \\hat {y} _ {\\mathrm {s}} ^ {\\text {m a s k}}\\right) + \\lambda_ {\\mathrm {c c}} \\cdot \\mathcal {L} _ {\\mathrm {c c}} \\left(\\hat {c} _ {q _ {i}, \\mathrm {s}}, \\hat {c} _ {q _ {i} ^ {\\prime}, \\mathrm {s}}\\right). \\tag {2}", + "image_path": "b9de224495594ff89155597dc05bb44f4fc3a8b8437e60e11a607d1e39e294ae.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " denotes the number of transformer blocks in the transformer decoder, and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{mc}}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cc}}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " represent the mask and class consistency losses, respectively. To compute the loss at the " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "s" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": "-th transformer decoder block, we obtain " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}, \\hat{y}_{\\mathrm{s}}^{\\mathrm{mask}}\\}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{\\hat{c}_{q_i,s}, \\hat{c}_{q_i',s}\\}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " which are pairs of mask predictions and class predictions of " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": "-th query " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "q_i" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": " from the original and augmented image pair " + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\{x, x'\\}" + }, + { + "bbox": [ + 55, + 604, + 296, + 714 + ], + "type": "text", + "content": ", respectively. We employ binary cross entropy and Jensen-Shannon divergence as a discrepancy measure of the mask consistency" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": "loss " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{mc})" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": " and class consistency loss " + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{cc})" + }, + { + "bbox": [ + 313, + 72, + 554, + 133 + ], + "type": "text", + "content": ", respectively. With the help of domain-aware context prompts and above losses, our model learns to predict not only accurately but also consistently in various domain shift scenarios during training." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 137, + 446, + 148 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 137, + 446, + 148 + ], + "spans": [ + { + "bbox": [ + 313, + 137, + 446, + 148 + ], + "type": "text", + "content": "3.3. Overall Loss Functions" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "spans": [ + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "content": "The overall loss of our framework is a weighted sum of the task loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "content": ", VLM regularization loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{reg}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "content": ", domain-aware contrastive loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contra}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "content": ", and the consistency loss " + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cons}}" + }, + { + "bbox": [ + 313, + 155, + 554, + 192 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 322, + 196, + 555, + 209 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 196, + 555, + 209 + ], + "spans": [ + { + "bbox": [ + 322, + 196, + 555, + 209 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} _ {\\text {s e g}} + \\lambda_ {\\text {r e g}} \\mathcal {L} _ {\\text {r e g}} + \\lambda_ {\\text {c o n t r a}} \\mathcal {L} _ {\\text {c o n t r a}} + \\lambda_ {\\text {c o n s}} \\mathcal {L} _ {\\text {c o n s}}, \\tag {3}", + "image_path": "5e3b2d7054e859ac1f9343e1ba15e9ebe1c045fdb6187e55bff185d984bb8799.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "inline_equation", + "content": "\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}" + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": " are constant weighting factors. We note that " + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{seg}}" + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{cons}}" + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": " are calculated with all queries and its predictions from every block of the transformer decoder. We provide details of the baseline losses " + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{seg}},\\mathcal{L}_{\\mathrm{reg}})" + }, + { + "bbox": [ + 313, + 213, + 554, + 274 + ], + "type": "text", + "content": " in the supplementary." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 282, + 395, + 295 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 282, + 395, + 295 + ], + "spans": [ + { + "bbox": [ + 313, + 282, + 395, + 295 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 301, + 447, + 313 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 301, + 447, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 301, + 447, + 313 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 319, + 554, + 342 + ], + "type": "text", + "content": "Datasets. We validate DPMFormer on synthetic-to-real and real-to-real scenarios in the single-source setting." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": "Synthetic datasets. GTAV [42] is a representative synthetic dataset which consists of 24,966 images with a resolution of " + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "inline_equation", + "content": "1914 \\times 1052" + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": ". The training split contains 12,403 images, while the validation and test set are of 6,382 and 6,181 images, respectively. SYNTHIA [43] dataset provides 6,580 images for training and 2,820 images for validation respectively, with the image resolution at " + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "inline_equation", + "content": "1280 \\times 760" + }, + { + "bbox": [ + 313, + 342, + 554, + 426 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": "Real-world datasets. Cityscapes [7] is a dataset collected from the real environment. The resolution of each image is " + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": ", and the population of training and validation split is 2,975 and 500, respectively. BDD100K [58] includes 7,000 training images and 1,000 validation images with the resolution of " + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "inline_equation", + "content": "1280 \\times 720" + }, + { + "bbox": [ + 313, + 426, + 554, + 521 + ], + "type": "text", + "content": ". Mapillary [34] is composed of images with diverse resolution, where the training and validation size is 18,000 and 2,000, respectively." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "text", + "content": "Network architecture. Our approach leverages the vision transformer-based models as backbones, initialized with either the CLIP [41] (ViT-B) or EVA02-CLIP [45] (EVA02-L) model. The CLIP backbone is configured with a patch size of 16, while the EVA02-CLIP backbone uses a patch size of 14. For the pixel and transformer decoder, aforementioned, we employ a mask classification architecture [5], consisting of " + }, + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "inline_equation", + "content": "N = 9" + }, + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "text", + "content": " layers with masked attention mechanisms. Additionally, we design the auxiliary network " + }, + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 313, + 522, + 554, + 653 + ], + "type": "text", + "content": " for domain-aware context prompt generation as a shallow multi-layer structure (BatchNorm-Linear-ReLU-Linear)." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "Training and evaluation. To optimize DPMFormer, we employ an AdamW [32] where the learning rate is set as " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " for synthetic and real training datasets, respectively. Following previous transformer-based studies [18, 36], we apply linear warm-up [15] for" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20795" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 58, + 70, + 293, + 217 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 293, + 217 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 293, + 217 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 293, + 217 + ], + "type": "table", + "html": "
Models (GTAV)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10145.3341.1840.7742.43
WildNet [28]ResNet-10145.7941.7347.0844.87
SHADE [61]ResNet-10146.6643.6645.5045.27
TLDR [27]ResNet-10147.5844.8848.8047.09
FAMix* [13]ResNet-10149.4746.4051.9749.28
SHADE [61]MiT-B553.2748.1954.9952.15
IBAFormer [44]MiT-B556.3449.7658.2654.79
VLTSeg* [21]ViT-B47.5045.7054.3049.17
TQDM* [36]ViT-B57.5047.6659.7654.97
DPMFormer* (ours)ViT-B59.0051.8063.6258.14
VLTSeg** [21]EVA02-L65.6058.4066.5063.50
Rein** [55]EVA02-L65.3060.5064.9063.60
Rein†ViT-L66.4060.4066.1064.30
TQDM** [36]EVA02-L68.8859.1870.1066.05
DPMFormer** (ours)EVA02-L70.0860.4870.6667.07
", + "image_path": "c770aa7ca80db43f06de4e70ddd3b4b7fbe26f6451586baee7200147ef582a8c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 58, + 304, + 293, + 371 + ], + "blocks": [ + { + "bbox": [ + 55, + 226, + 295, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 226, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 295, + 293 + ], + "type": "text", + "content": "Table 1. Comparison with the state-of-the-art DGSS methods on synthetic-to-real scenario with GTAV [42] as a source. We note that *, **, and † denotes the models initialized with pretrained CLIP [41], EVA02-CLIP [45], and DINOv2 [35], respectively. The best and the second best performances are highlighted with bold and underline, respectively." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 58, + 304, + 293, + 371 + ], + "lines": [ + { + "bbox": [ + 58, + 304, + 293, + 371 + ], + "spans": [ + { + "bbox": [ + 58, + 304, + 293, + 371 + ], + "type": "table", + "html": "
Models (Synthia)BackboneCityscapesBDDMapillaryAvg.
SAN-SAW [39]ResNet-10140.8735.9837.2638.04
TLDR [27]ResNet-10142.6035.4637.4638.51
IBAFormer [44]MiT-B550.9244.6650.5848.72
VLTSeg** [21]EVA02-L56.8050.5054.5053.93
TQDM** [36]EVA02-L57.9952.4354.8755.10
DPMFormer** (ours)EVA02-L58.9254.3960.0857.80
", + "image_path": "8b1f81c2c2731508663b6303b86d01dfb15f6d5617f5132d922d3bda5f6a273d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "text", + "content": "initial 1,500 iterations and rare class sampling [18]. The optimizer settings are identical for both CLIP and EVA02-CLIP backbone models. We set the total training iterations and the batch size as 20,000 and 8, respectively. Weighting factors " + }, + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "inline_equation", + "content": "\\{\\lambda_{\\mathrm{reg}},\\lambda_{\\mathrm{contra}},\\lambda_{\\mathrm{cons}}\\}" + }, + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "text", + "content": " are set as 1, 1, 10. We crop the input image to have a resolution of " + }, + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "inline_equation", + "content": "512\\times 512" + }, + { + "bbox": [ + 54, + 423, + 295, + 533 + ], + "type": "text", + "content": ". The texture perturbation is only applied during training. We use mean Intersection over Union (mIoU) [12] to quantitatively evaluate the results following the convention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 540, + 174, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 174, + 553 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 174, + 553 + ], + "type": "text", + "content": "4.2. Quantitative Results" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": "Synthetic-to-real. As shown in Tab. 1, we achieve state-of-the-art in every target domain with both backbones. Notably, with the ViT-B backbone initialized with pretrained CLIP [41], DPMFormer consistently surpasses previous state-of-the-art [36] by " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": " on average mIoU among target domains. In detail, on Cityscapes which contains mostly daytime real images, the domain robustness empowered by consistency learning assists DPMFormer to cope with the domain gap caused by visual realism, improving the state-of-the-art performance by " + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "inline_equation", + "content": "1.5\\%" + }, + { + "bbox": [ + 55, + 557, + 295, + 714 + ], + "type": "text", + "content": ". Meanwhile on BDD100K and Mapillary which have higher environmental variation in terms of weather, time and location, the domain-aware prompt generation enables the model to adaptively leverage" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 317, + 70, + 553, + 204 + ], + "blocks": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "lines": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 55, + 380, + 295, + 403 + ], + "type": "text", + "content": "Table 2. Comparison with the state-of-the-art DGSS methods where SYNTHIA [43] is set as a source dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 70, + 553, + 204 + ], + "lines": [ + { + "bbox": [ + 317, + 70, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 317, + 70, + 553, + 204 + ], + "type": "table", + "html": "
Models (Cityscapes)BackboneBDDMapillaryAvg.
SAN-SAW [39]ResNet-10154.7361.2742.43
WildNet [28]ResNet-10147.0141.7344.87
SHADE [61]ResNet-10150.9543.6645.27
TQDM* [36]ViT-B50.5465.7458.14
DPMFormer* (ours)ViT-B54.8167.7261.27
HGFormer [10]EVA02-L61.5072.1066.80
VLTSeg** [21]EVA02-L64.4076.4070.40
Rein** [55]EVA02-L64.1069.5066.80
Rein†ViT-L65.0072.3068.65
TQDM [36]**EVA02-L64.7276.1570.44
DPMFormer** (ours)EVA02-L64.276.6770.44
", + "image_path": "b1d97389faca534b3600cea308a80115e5f378469b3ee5ce8f8f5743c4011ce6.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 553, + 235 + ], + "type": "text", + "content": "Table 3. Comparison with the state-of-the-arts trained with Cityscapes [7] on the real-to-real scenario." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": "the textual knowledge for segmentation, impressively escalating the performance by " + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "inline_equation", + "content": "4.14\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "inline_equation", + "content": "3.17\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": ", respectively. Moreover, even with the larger backbone [45], we mark the highest average performance of " + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "inline_equation", + "content": "67.07\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": ". In Cityscapes and Mapillary dataset, we outperform TQDM [36] by " + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "inline_equation", + "content": "1.2\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "inline_equation", + "content": "0.56\\%" + }, + { + "bbox": [ + 313, + 255, + 555, + 363 + ], + "type": "text", + "content": " respectively, while nearly reaching the score of Rein [55] in BDD. The overall results demonstrate the superiority of DPMFormer, emphasizing the efficacy of domain-awareness as well as domain-robustness." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "text", + "content": "In Tab. 2, we also present synthetic-to-real results where SYNTHIA [43] is set as a source domain. Again, DPM-Former outperforms all competitors by a large margin, achieving the state-of-the-art performance in every target domain. Remarkably, our framework surpasses the previous state-of-the-art [36] by an average of " + }, + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "inline_equation", + "content": "2.7\\%" + }, + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "text", + "content": " mIoU across target domains. In particular, the performance on Mapillary [34] improves significantly by " + }, + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "inline_equation", + "content": "5.21\\%" + }, + { + "bbox": [ + 313, + 364, + 556, + 484 + ], + "type": "text", + "content": ". These results demonstrate the effectiveness of DPM-Former in addressing the domain gap in terms of texture and perspective." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": "Real-to-real. Furthermore, as shown in Tab.3, DPMFormer records the highest average mIoU with both backbones. With the CLIP-pretrained backbone, we significantly boost the state-of-the-art performance by " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "4.27\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "1.98\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": " on BDD and Mapillary, respectively. These results indicate that enhanced domain robustness supports the model in predicting consistently under environmental changes, while the domain-aware prompts facilitate the semantic alignment on the unseen domain. As a result, DPMFormer successfully achieves an average performance gain of " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "3.13\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": " over the previous state-of-the-art. Equipped with the EVA-CLIP [45] pretrained backbone, we attain the highest mIoU of " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "76.67\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": " on Mapillary, while performing slightly lower on BDD. Overall, the average performance is " + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "inline_equation", + "content": "70.44\\%" + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": ", which is comparable to TQDM [36]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 671, + 427, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 671, + 427, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 671, + 427, + 685 + ], + "type": "text", + "content": "4.3. Qualitative Results" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "In Fig. 4, we qualitatively compare our method with FAMix [13] and TQDM [36] on the synthetic-to-real sce" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20796" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 68, + 545, + 293 + ], + "blocks": [ + { + "bbox": [ + 69, + 68, + 545, + 293 + ], + "lines": [ + { + "bbox": [ + 69, + 68, + 545, + 293 + ], + "spans": [ + { + "bbox": [ + 69, + 68, + 545, + 293 + ], + "type": "image", + "image_path": "87cd7074766a0a9379133f1848d6d5a51b55df0de2bce5ca0d3133ae9c034bf1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 301, + 555, + 336 + ], + "lines": [ + { + "bbox": [ + 54, + 301, + 555, + 336 + ], + "spans": [ + { + "bbox": [ + 54, + 301, + 555, + 336 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison on synthetic-to-real scenario with the CLIP-pretrained backbone (ViT-B). The training source domain is set as GTAV [42] while the target domains are BDD100K [58] and Mapillary [34]. The overall result shows that DPMFormer accomplishes precise segmentation with the images of strong illumination contrast as well as confusing textures." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 356, + 297, + 561 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 356, + 297, + 561 + ], + "spans": [ + { + "bbox": [ + 54, + 356, + 297, + 561 + ], + "type": "text", + "content": "nario with the pretrained CLIP backbone (ViT-B). As depicted in the results on BDD100K, both competitors struggle to discriminate accurately under environments having confusing textures or large variations of illumination. In particular, the road in the first-row image has a texture similar to the exterior of a car, leading TQDM to a mismatch between the visual and textual features. In addition, FAMix shows sensitivity to the shades, mislabeling the road as a sidewalk. On the contrary, our method effectively distinguishes the road from other categories by reflecting the domain-specific properties of real-world roads in clear weather. In the case of the second row image with the high illumination contrast, both FAMix and TQDM fail to notice sidewalks in the image due to their low brightness. On the other hand, DPMFormer carries out precise predictions under severe photometric changes owing to the domain robustness acquired from handling diverse texture changes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "As observed with the samples from the Mapillary (the third and fourth rows), FAMix and TQDM lack discriminability between the road and the sidewalk. To be specific with the first image (third row), these classes appear to have a similar color and texture, resulting in misclassifications. Meanwhile, the domain-specific properties from the image transfer the semantic knowledge of the appearance of real-world sidewalk to the model, DPMFormer accurately distinguish between the road and the sidewalk. With the second image, TQDM [36] shows vulnerability against small textural changes on the sidewalk (e.g., shades of the poles). Contrarily, DPMformer effectively performs segmentation ow" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 356, + 555, + 380 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 356, + 555, + 380 + ], + "spans": [ + { + "bbox": [ + 313, + 356, + 555, + 380 + ], + "type": "text", + "content": "ing to its domain-robustness as well as the obtained domain-aware textual queries." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 386, + 378, + 399 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 378, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 378, + 399 + ], + "type": "text", + "content": "4.4. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 403, + 411, + 415 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 403, + 411, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 411, + 415 + ], + "type": "text", + "content": "4.4.1. Ablation studies" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": "To demonstrate the effectiveness of the components in DPMFormer, i.e., texture perturbation, domain-robust consistency learning " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{cons})" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ", and domain-aware context prompt learning " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{contra})" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ", we conduct ablation studies with the CLIP pretrained backbone on synthetic-to-real scenario. As presented in Tab. 4, every component contributes to the performance gain. Specifically, the texture perturbation enlarges the observable domain during training, boosts the average performance by " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "0.65\\%" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ". In addition with the " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{cons}" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ", the model successfully equips the robustness against domain-shift and enhances the prediction accuracy on unseen target domains, increasing the average mIoU by " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "0.77\\%" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ". Furthermore, " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{contra}" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": " allows the model to acquire domain-specific property from the input image in the form of textual prompt embedding. Consequently, the model exploits both visual and textual cues properly aligned for the target domain, considerably elevate the average mIoU by " + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "inline_equation", + "content": "1.99\\%" + }, + { + "bbox": [ + 313, + 418, + 556, + 658 + ], + "type": "text", + "content": ". Combining all components together, DPMFormer effectively learns both domain-awareness and domain-robustness to generate accurate results consistently on unseen target domains." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 662, + 528, + 675 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 662, + 528, + 675 + ], + "spans": [ + { + "bbox": [ + 313, + 662, + 528, + 675 + ], + "type": "text", + "content": "4.4.2. Comparison with prompt learning methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "In Tab. 5, we compare our domain-aware context prompt learning " + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "inline_equation", + "content": "(\\mathcal{L}_{\\mathrm{contra}})" + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": " with previous prompt learning methods [25, 26, 63] to validate the efficacy for domain gen" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20797" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 70, + 295, + 137 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 137 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 137 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 137 + ], + "type": "table", + "html": "
Models (GTAV)CityscapesBDDMapillaryAvg.
Baseline57.547.6659.7654.97
+ Perturbation57.0448.1960.9155.38
+ Lcons58.2249.3960.8456.15
+ Lcontra59.051.863.6258.14
", + "image_path": "9e6654b6caca6741142c6ab05d2b59a2c96c8f63a543388e6f73a4383a98bf22.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 57, + 204, + 296, + 312 + ], + "blocks": [ + { + "bbox": [ + 57, + 204, + 296, + 312 + ], + "lines": [ + { + "bbox": [ + 57, + 204, + 296, + 312 + ], + "spans": [ + { + "bbox": [ + 57, + 204, + 296, + 312 + ], + "type": "image", + "image_path": "81442a0034263349e47936b3c2e9035b557ce3b03bd18944350598372fe43f34.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "lines": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 296, + 354 + ], + "type": "text", + "content": "Figure 5. Qualitative results on diverse styles, i.e., Minimalist, Pop Art, Bauhaus, and Cubism. The models are trained on GTAV with CLIP pretrained backbone (ViT-B)." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": "deralization. Given the textual features " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "t = ENC_{T}(p_{x})" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": " and visual features " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "v = \\hat{F}(x)" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ", the contrastive loss of CoCoOp [63] encourages the similarity between the anchor text feature and its corresponding visual feature to be higher than others. As observed in the second row, CoCoOp marginally improves the performance of prompt generator " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ", indicating that the generated feature becomes instancespecific which is less generalizable to other domains. Moreover, additionally assigning the visual feature from augmented sample as positives " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "(\\mathrm{CoCoOp}^{+})" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": " resulted in negligible performance gain, since the derived prompt contains domain-invariant information rather than domain-specific properties. Furthermore, although both MaPLe [25] and PromptSRC [26] employ multi-modal prompts, their performance gains are modest due to the limited generalizability of fixed prompts. On the other hand, our domain-aware contrastive loss design enable the model to proficiently capture domain-specific property from the image which are more helpful for the semantic alignment between the visual features and the textual knowledge. From the fifth to the eighth row, we verify " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\text{contra}}" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": " with different similarity calculation targets for " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\sin(\\cdot)" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ", i.e., text features " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ", text and visual features " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "(t, v)" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ", and (3) output context embeddings " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": ". The results confirm that computing the domain-aware contrastive loss empowers the domain generalizability especially when computed with the context embeddings " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": " which can provide direct domain guidance to " + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "inline_equation", + "content": "h_{\\theta}" + }, + { + "bbox": [ + 54, + 390, + 298, + 714 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 315, + 70, + 554, + 186 + ], + "blocks": [ + { + "bbox": [ + 55, + 144, + 295, + 179 + ], + "lines": [ + { + "bbox": [ + 55, + 144, + 295, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 144, + 295, + 179 + ], + "type": "text", + "content": "Table 4. Ablation study of proposed components. The models are trained on GTAV with CLIP pretrained backbone (ViT-B). The best performance for each column is highlighted in bold." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 315, + 70, + 554, + 186 + ], + "lines": [ + { + "bbox": [ + 315, + 70, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 315, + 70, + 554, + 186 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{contra}} \\) (GTAV)CityscapesBDDMapillaryAvg.
-57.6549.6361.1056.13
CoCoOp [63]57.8449.9161.3356.36
\\( \\mathrm{CoCoOp}^{+} \\)57.6050.0460.9656.20
MaPLe [25]57.8750.1261.0456.34
PromptSRC [26]58.1049.7362.5156.78
Ours (sim(t,t))58.1750.1861.3756.57
Ours (sim(t,v))59.4950.3462.1857.34
Ours (sim(π,π))59.0051.8063.6258.14
", + "image_path": "2dfc67f6bb833a13710922897e632a3fe0702ec45951f6e1f2d3890732b6109d.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "lines": [ + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": "Table 5. Comparison with prompt learning methods [25, 26, 63] in synthetic-to-real scenario. - denotes DPMFormer without " + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{contra}}" + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\mathrm{CoCoOp}^{+}" + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": " indicates the modified loss that additionally include the feature from augmented sample of anchor as positives. " + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 313, + 193, + 555, + 250 + ], + "type": "text", + "content": " refers to the text feature, visual feature, and context embedding." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 268, + 493, + 281 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 493, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 493, + 281 + ], + "type": "text", + "content": "4.4.3. Qualitative results on diverse styles" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 283, + 556, + 475 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 283, + 556, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 283, + 556, + 475 + ], + "type": "text", + "content": "To further verify the domain generalization capability of DPMFormer, we compare our model with TQDM [36] on images with diverse artistic styles generated by " + }, + { + "bbox": [ + 313, + 283, + 556, + 475 + ], + "type": "inline_equation", + "content": "\\mathrm{ChatGPT}^1" + }, + { + "bbox": [ + 313, + 283, + 556, + 475 + ], + "type": "text", + "content": ". As shown in Fig. 5, DPMFormer correctly predicts objects and their surroundings even in severe domain shift scenarios, e.g., Cubism. With modern art styles (i.e., Minimalist, Pop Art, and Bauhaus), TQDM perplexes among road, sidewalk, and terrain due to their textural changes. Notably, DPMFormer produces more reliable results owing to the enhanced robustness against texture variations that are learned from the texture perturbation and consistency learning. In case of the scene in cubism style, TQDM mispredicts the person on the left side because of its visual similarity with the person in the traffic sign. On the other hand, our method reflects the cubism style to the textual object queries, making an accurate classification with the same instance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 484, + 388, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 388, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 388, + 498 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 505, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 556, + 696 + ], + "type": "text", + "content": "In this paper, we presented DPMFormer, a novel domain generalization framework for semantic segmentation. To address the limited generalizability of fixed context prompt learned from a single source dataset, we introduced a novel domain-aware prompt learning which can reflect domain-specific properties of the input image into textual prompts. The proposed component enhanced the semantic alignment between visual and textual cues, assisting the model to fully leverage the abundant semantic knowledge of VLMs. Moreover, to empower domain-robustness, we simulated various domain shifts via texture perturbations, and provided consistency guidance to the model by minimizing prediction discrepancies between original and augmented images. Through extensive experiments and analyses, we demonstrated the effectiveness of DPMFormer, achieving state-of-the-art performance on various benchmarks." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 325, + 702, + 407, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 702, + 407, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 702, + 407, + 714 + ], + "type": "text", + "content": "1 https://chat.openai.com" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "20798" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 133 + ], + "type": "text", + "content": "Acknowledgments. This project was supported by the National Research Foundation of Korea grant funded by the Korea government (MSIT) (No. 2022R1A2B5B02001467; RS-2024-00346364). This research was supported by Sookmyung Women's University Research Grants (1-2403-2034)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "spans": [ + { + "bbox": [ + 56, + 144, + 115, + 156 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 163, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 61, + 163, + 296, + 207 + ], + "type": "text", + "content": "[1] Shuanghao Bai, Yuedi Zhang, Wanqi Zhou, Zhirong Luan, and Badong Chen. Soft prompt generation for domain generalization. In European Conference on Computer Vision, pages 434-450. Springer, 2025. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "spans": [ + { + "bbox": [ + 61, + 209, + 296, + 251 + ], + "type": "text", + "content": "[2] Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pages 456-473, 2018. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 254, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 254, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 254, + 296, + 297 + ], + "type": "text", + "content": "[3] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 819-827, 2024. 1, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 300, + 296, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 300, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 62, + 300, + 296, + 364 + ], + "type": "text", + "content": "[4] Prithvijit Chattopadhyay, Kartik Sarangmath, Vivek Vijaykumar, and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain generalization. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 19288-19300, 2023. 1, 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 366, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 366, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 366, + 296, + 421 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 1, 3, 4, 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 487 + ], + "type": "text", + "content": "[6] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11580-11590, 2021. 1, 3, 4" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 489, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 489, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 489, + 296, + 555 + ], + "type": "text", + "content": "[7] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 1, 3, 7, 8" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 556, + 296, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 556, + 296, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 556, + 296, + 589 + ], + "type": "text", + "content": "[8] Gabriela Csurka. A comprehensive survey on domain adaptation for visual applications. Domain adaptation in computer vision applications, pages 1-35, 2017. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 296, + 634 + ], + "type": "text", + "content": "[9] Karan Desai and Justin Johnson. Virtex: Learning visual representations from textual annotations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11162-11173, 2021. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 296, + 689 + ], + "type": "text", + "content": "[10] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15413-15423, 2023. 1, 3, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 296, + 713 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner," + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 116 + ], + "type": "text", + "content": "Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR, 2021. 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 555, + 171 + ], + "type": "text", + "content": "[12] Mark Everingham, SM Ali Eslami, Luc Van Gool, Christopher KI Williams, John Winn, and Andrew Zisserman. The Pascal visual object classes challenge: A retrospective. International journal of computer vision, 111:98-136, 2015. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 173, + 555, + 227 + ], + "type": "text", + "content": "[13] Mohammad Fahes, Tuan-Hung Vu, Andrei Bursuc, Patrick Pérez, and Raoul de Charette. A simple recipe for language-guided domain generalized segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23428-23437, 2024. 1, 3, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 554, + 282 + ], + "type": "text", + "content": "[14] Qi Fan, Mattia Segu, Yu-Wing Tai, Fisher Yu, Chi-Keung Tang, Bernt Schiele, and Dengxin Dai. Towards robust object detection invariant to real-world domain shifts. In The Eleventh International Conference on Learning Representations (ICLR 2023). OpenReview, 2023. 1, 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 284, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 284, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 284, + 554, + 304 + ], + "type": "text", + "content": "[15] P Goyal. Accurate, large minibatch sg d: training imagenet in 1 hour. arXiv preprint arXiv:1706.02677, 2017. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 306, + 554, + 338 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 306, + 554, + 338 + ], + "spans": [ + { + "bbox": [ + 316, + 306, + 554, + 338 + ], + "type": "text", + "content": "[16] Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261, 2019. 3, 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 339, + 554, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 339, + 554, + 403 + ], + "spans": [ + { + "bbox": [ + 317, + 339, + 554, + 403 + ], + "type": "text", + "content": "[17] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8340-8349, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 555, + 459 + ], + "type": "text", + "content": "[18] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9924-9935, 2022. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 317, + 460, + 554, + 514 + ], + "type": "text", + "content": "[19] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6891-6902, 2021. 1, 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 554, + 558 + ], + "type": "text", + "content": "[20] Xun Huang and Serge Belongie. Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE International Conference on Computer Vision, pages 1501-1510, 2017. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "spans": [ + { + "bbox": [ + 317, + 559, + 554, + 613 + ], + "type": "text", + "content": "[21] Christoph Hummer, Manuel Schwonberg, Liangwei Zhong, Hu Cao, Alois Knoll, and Hanno Gottschalk. Vltseg: Simple transfer of clip-based vision-language representations for domain generalized semantic segmentation. arXiv preprint arXiv:2312.02021, 2023. 1, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 614, + 554, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 554, + 669 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 554, + 669 + ], + "type": "text", + "content": "[22] Xinyue Huo, Lingxi Xie, Hengtong Hu, Wengang Zhou, Houqiang Li, and Qi Tian. Domain-agnostic priors for semantic segmentation under unsupervised domain adaptation and domain generalization. International Journal of Computer Vision, 132(9):3954-3976, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 554, + 713 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20799" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "spans": [ + { + "bbox": [ + 77, + 72, + 294, + 94 + ], + "type": "text", + "content": "conference on machine learning, pages 4904-4916. PMLR, 2021.3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 295, + 148 + ], + "type": "text", + "content": "[24] Xueying Jiang, Jiaxing Huang, Sheng Jin, and Shijian Lu. Domain generalization via balancing training difficulty and model capability. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 18993-19003, 2023. 1, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 295, + 205 + ], + "type": "text", + "content": "[25] Muhammad Uzair Khattak, Hanoona Rasheed, Muhammad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113-19122, 2023. 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 206, + 295, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 206, + 295, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 206, + 295, + 271 + ], + "type": "text", + "content": "[26] Muhammad Uzair Khattak, Syed Talal Wasim, Muzammal Naseer, Salman Khan, Ming-Hsuan Yang, and Fahad Shahbaz Khan. Self-regulating prompts: Foundational model adaptation without forgetting. In Proceedings of the IEEE/CVF international conference on computer vision, pages 15190-15200, 2023. 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 295, + 326 + ], + "type": "text", + "content": "[27] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Texture learning domain randomization for domain generalized segmentation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 677-687, 2023. 1, 3, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 295, + 381 + ], + "type": "text", + "content": "[28] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9936-9946, 2022. 1, 3, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "spans": [ + { + "bbox": [ + 56, + 382, + 295, + 426 + ], + "type": "text", + "content": "[29] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Deeper, broader and artier domain generalization. 2017 IEEE International Conference on Computer Vision (ICCV), pages 5543-5551, 2017. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 295, + 470 + ], + "type": "text", + "content": "[30] Yijun Li, Chen Fang, Jimei Yang, Zhaowen Wang, Xin Lu, and Ming-Hsuan Yang. Universal style transfer via feature transforms. Advances in neural information processing systems, 30, 2017. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 536 + ], + "type": "text", + "content": "[31] Yuqi Lin, Minghao Chen, Wenxiao Wang, Boxi Wu, Ke Li, Binbin Lin, Haifeng Liu, and Xiaofei He. Clip is also an efficient segmenter: A text-driven approach for weakly supervised semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15305-15314, 2023. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 537, + 295, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 295, + 559 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 295, + 559 + ], + "type": "text", + "content": "[32] I Loshchilov. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 295, + 603 + ], + "type": "text", + "content": "[33] Jiasen Lu, Dhruv Batra, Devi Parikh, and Stefan Lee. Vilbert: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks. Advances in neural information processing systems, 32, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 603, + 295, + 658 + ], + "type": "text", + "content": "[34] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulo, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international conference on computer vision, pages 4990-4999, 2017. 5, 6, 7, 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 713 + ], + "type": "text", + "content": "[35] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. 6" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 118 + ], + "type": "text", + "content": "[36] Byeonghyun Pak, Byeongju Woo, Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. Textual query-driven mask transformer for domain generalized segmentation. European Conference on Computer Vision, 2024. 1, 3, 4, 5, 6, 7, 8, 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "text", + "content": "[37] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In Proceedings of the European Conference on Computer Vision (ECCV), pages 464-479, 2018. 1, 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[38] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In Proceedings of the IEEE/CVF international conference on computer vision, pages 1863-1871, 2019. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 553, + 262 + ], + "type": "text", + "content": "[39] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2594-2605, 2022. 1, 3, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 263, + 553, + 319 + ], + "type": "text", + "content": "[40] Hieu Pham, Zihang Dai, Golnaz Ghiasi, Kenji Kawaguchi, Hanxiao Liu, Adams Wei Yu, Jiahui Yu, Yi-Ting Chen, Minh-Thang Luong, Yonghui Wu, et al. Combined scaling for zero-shot transfer learning. Neurocomputing, 555: 126658, 2023. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 319, + 553, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 319, + 553, + 386 + ], + "spans": [ + { + "bbox": [ + 316, + 319, + 553, + 386 + ], + "type": "text", + "content": "[41] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International conference on machine learning, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 387, + 553, + 453 + ], + "type": "text", + "content": "[42] Stephan R Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part II 14, pages 102-118. Springer, 2016. 2, 5, 6, 7, 3, 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 453, + 553, + 519 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 453, + 553, + 519 + ], + "spans": [ + { + "bbox": [ + 316, + 453, + 553, + 519 + ], + "type": "text", + "content": "[43] German Ros, Laura Sellart, Joanna Materzynska, David Vazquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3234-3243, 2016. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 553, + 565 + ], + "type": "text", + "content": "[44] Qiyu Sun, Huilin Chen, Meng Zheng, Ziyan Wu, Michael Felsberg, and Yang Tang. Ibaformer: Intra-batch attention transformer for domain generalized semantic segmentation. arXiv preprint arXiv:2309.06282, 2023. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 565, + 553, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 565, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 316, + 565, + 553, + 599 + ], + "type": "text", + "content": "[45] Quan Sun, Yuxin Fang, Ledell Wu, Xinlong Wang, and Yue Cao. Eva-clip: Improved training techniques for clip at scale. arXiv preprint arXiv:2303.15389, 2023. 1, 3, 5, 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 600, + 553, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 622 + ], + "type": "text", + "content": "[46] C Szegedy. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 622, + 553, + 656 + ], + "type": "text", + "content": "[47] Hao Tan and Mohit Bansal. Lxmert: Learning cross-modality encoder representations from transformers. arXiv preprint arXiv:1908.07490, 2019. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 553, + 689 + ], + "type": "text", + "content": "[48] Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pages 1521-1528. IEEE, 2011. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 553, + 713 + ], + "type": "text", + "content": "[49] Yi-Hsuan Tsai, Wei-Chih Hung, Samuel Schulter, Kiyuk Sohn, Ming-Hsuan Yang, and Manmohan Chandraker." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "20800" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 115 + ], + "type": "text", + "content": "Learning to adapt structured output space for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 7472-7481, 2018. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 173 + ], + "type": "text", + "content": "[50] Dmitry Ulyanov, Andrea Vedaldi, and Victor Lempitsky. Improved texture networks: Maximizing quality and diversity in feed-forward stylization and texture synthesis. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 6924-6932, 2017. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 57, + 175, + 295, + 196 + ], + "type": "text", + "content": "[51] A Vaswani. Attention is all you need. Advances in Neural Information Processing Systems, 2017. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 252 + ], + "type": "text", + "content": "[52] Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 5018-5027, 2017. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 255, + 294, + 308 + ], + "type": "text", + "content": "[53] Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In The Nineth International Conference on Learning Representations (ICLR 2021). OpenReview, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 57, + 310, + 295, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 310, + 295, + 354 + ], + "spans": [ + { + "bbox": [ + 57, + 310, + 295, + 354 + ], + "type": "text", + "content": "[54] Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7201-7211, 2022. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 356, + 294, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 356, + 294, + 431 + ], + "spans": [ + { + "bbox": [ + 56, + 356, + 294, + 431 + ], + "type": "text", + "content": "[55] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 28619-28630, 2024. 1, 3, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 57, + 434, + 294, + 477 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 434, + 294, + 477 + ], + "spans": [ + { + "bbox": [ + 57, + 434, + 294, + 477 + ], + "type": "text", + "content": "[56] Zhenyao Wu, Xinyi Wu, Xiaoping Zhang, Lili Ju, and Song Wang. Siamdoge: Domain generalizable semantic segmentation using siamese network. In European Conference on Computer Vision, pages 603-620. Springer, 2022. 1, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 57, + 479, + 294, + 522 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 479, + 294, + 522 + ], + "spans": [ + { + "bbox": [ + 57, + 479, + 294, + 522 + ], + "type": "text", + "content": "[57] Shiqi Yang, Yaxing Wang, Joost Van De Weijer, Luis Herranz, and Shangling Jui. Generalized source-free domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 8978-8987, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 589 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 589 + ], + "type": "text", + "content": "[58] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2636-2645, 2020. 2, 5, 7, 3, 4, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 666 + ], + "type": "text", + "content": "[59] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2100-2110, 2019. 1, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 294, + 713 + ], + "type": "text", + "content": "[60] Xin Zhang, Shixiang Shane Gu, Yutaka Matsuo, and Yusuke Iwasawa. Domain prompt learning for efficiently adapting clip to unseen domains. Transactions of the Japanese Society for Artificial Intelligence, 38(6):B-MC2_1, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 384 + ], + "type": "list", + "angle": 0, + "index": 19, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 127 + ], + "type": "text", + "content": "[61] Yuyang Zhao, Zhun Zhong, Na Zhao, Nicu Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In European conference on computer vision, pages 535-552. Springer, 2022. 1, 3, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 173 + ], + "type": "text", + "content": "[62] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. Advances in neural information processing systems, 35:338-350, 2022. 1, 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 228 + ], + "type": "text", + "content": "[63] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 16816-16825, 2022. 5, 7, 8, 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 230, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 230, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 553, + 271 + ], + "type": "text", + "content": "[64] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. International Journal of Computer Vision, 130(9):2337-2348, 2022. 4" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 553, + 328 + ], + "type": "text", + "content": "[65] Jun-Yan Zhu, Taesung Park, Phillip Isola, and Alexei A Efros. Unpaired image-to-image translation using cycle-consistent adversarial networks. In Proceedings of the IEEE international conference on computer vision, pages 2223-2232, 2017. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 553, + 384 + ], + "type": "text", + "content": "[66] Yang Zou, Zhiding Yu, BVK Kumar, and Jinsong Wang. Unsupervised domain adaptation for semantic segmentation via class-balanced self-training. In Proceedings of the European conference on computer vision (ECCV), pages 289-305, 2018. 3" + } + ] + } + ], + "index": 18 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "20801" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_content_list.json b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c7a2e2e8d230a1af9ac4fa967586439b23f97ff1 --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_content_list.json @@ -0,0 +1,1553 @@ +[ + { + "type": "text", + "text": "Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition", + "text_level": 1, + "bbox": [ + 101, + 128, + 895, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Meiqi Cao $^{1}$ , Xiangbo Shu $^{1*}$ , Xin Jiang $^{1}$ , Rui Yan $^{1,2}$ , Yazhou Yao $^{1}$ , Jinhui Tang $^{3}$ $^{1}$ Nanjing University of Science and Technology \n $^{2}$ Nanjing University \n $^{3}$ Nanjing Forestry University", + "bbox": [ + 174, + 202, + 821, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{cmq123,shuxb,xinjiang,ruiyan,yazhou.yao}@njust.edu.cn,tangjh@njfu.edu.cn", + "bbox": [ + 169, + 276, + 821, + 292 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 325, + 326, + 342 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "While event cameras excel in capturing microsecond temporal dynamics, they suffer from sparse spatial representations compared to traditional RGB data. Thus, multimodal event-based action recognition approaches aim to synergize complementary strengths by independently extracting and integrating paired RGB-Event features. However, this paradigm inevitably introduces additional data acquisition costs, while eroding the inherent privacy advantages of event-based sensing. Drawing inspiration from event-to-image reconstruction, texture-enriched visual representation directly reconstructed from asynchronous event streams is a promising solution. In response, we propose an Enhanced Multimodal Perceptual (EMP) framework that hierarchically explores multimodal cues (e.g., edges and textures) from raw event streams through two synergistic innovations spanning representation to feature levels. Specifically, we introduce Cross-Modal Frequency Enhancer (CFE) that leverages complementary frequency characteristics between reconstructed frames and stacked frames to refine event representations. Furthermore, to achieve unified feature encoding across modalities, we develop High-Frequency Guided Selector (HGS) for semantic consistency token selection guided by dynamic edge features while suppressing redundant multimodal information interference adaptively. Extensive experiments on four benchmark datasets demonstrate the superior effectiveness of our proposed framework. The code is available at https://github.com/caomq123/EMP.", + "bbox": [ + 88, + 358, + 485, + 782 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 808, + 220, + 824 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Traditional frame-based cameras often encounter performance limitations when dealing with rapid motion or complex illumination conditions [22, 25, 42, 43]. In contrast,", + "bbox": [ + 89, + 834, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/73c0dbf6f4e9f26cec48fdfdbf116f63d7b0b172c725f2d93740f3c296eb4649.jpg", + "image_caption": [ + "Figure 1. (a) Previous multi-modal input approaches independently extract features from paired RGB-Event modalities, both introducing excessive RGB data dependencies and disrupting the privacy-preserving property unique to event cameras. (b) Our approach takes a lightful reconstruct network (Firenet [40]) to substitute RGB data while narrowing the modality divergence with events, thereby enabling enhanced unified multimodal perception." + ], + "image_footnote": [], + "bbox": [ + 529, + 329, + 890, + 415 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "event-driven neuromorphic visual sensors, also known as event cameras [13, 15], operate by asynchronously capturing per-pixel intensity changes only when a significant light change occurs. This unique functionality endows event cameras with several prominent advantages, including high dynamic range, low power consumption, and high temporal resolution, which enable them to excel in challenging scenarios such as fast motion, extreme lighting conditions, and privacy-sensitive applications. These attributes have led to their widespread adoption in various computer vision tasks, including object detection and tracking [16, 26, 52], motion deblurring [5, 27], and image reconstruction [31, 40, 53]. Notably, the inherent sensitivity of event cameras to dynamic scenes provides a unique advantage in capturing motion-related actions. Leveraging these strengths, the task of Event-based Action Recognition (EAR) has emerged, aiming to effectively extract action features from event streams to enhance recognition performance.", + "bbox": [ + 511, + 563, + 906, + 849 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Currently, the EAR methods faces significant challenges in effectively utilizing the spatially sparse and asynchronous event streams. Traditional single-modal input approaches", + "bbox": [ + 511, + 854, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author", + "bbox": [ + 109, + 887, + 233, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5969", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "adopt two primary strategies: (i) Processing asynchronous event points through point-based encoders (GCNs [11, 54]) or spiking neural networks (SNNs [3, 8, 37]), and (ii) Applying CNN- [32, 44], Transformer- [7, 57] or Mamba- [38, 49] based visual encoders to frames stacked from asynchronous event streams. While point-based methods accommodate the asynchronous nature of events effectively, they exhibit suboptimal recognition performance. Simultaneously, frame-based methods suffer from the lack of critical texture information due to the focus on dynamic edge capture inherent in event cameras.", + "bbox": [ + 89, + 90, + 480, + 257 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In recent years, there has been a notable shift toward multi-modal input strategies [9, 24, 48, 50], leveraging diverse data sources to improve feature extraction robustness for practical applications. As depicted in Fig. 1(a), conventional multi-modal methods typically process RGB images and event streams in parallel through independent encoders, followed by late fusion. Although such approaches improve robustness by aggregating complementary features, two critical limitations persist: (i) The incorporation of RGB images not only introduces additional data acquisition costs but also compromises the intrinsic privacy-preserving nature of event cameras by capturing detailed individual visuals. (ii) Independent encoders disregarding inter-modal correlations introduce redundancy that critically undermines the acquisition of discriminative action features.", + "bbox": [ + 89, + 260, + 482, + 487 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address the challenges above, we seek a privacy-preserving alternative to RGB data that eliminates additional data dependency and narrows the representational gap between surrogate modal with event modal. As shown in Fig. 1(b), inspired by the event-to-image [40] techniques, we employ a lightweight event reconstruction network to generate texture-enriched surrogate frames directly from event streams. This approach effectively emulates RGB semantics while filtering privacy-sensitive details. Building upon this, we introduce a unified multi-modal encoding architecture that harmonizes edge-focused event representations with texture-augmented surrogates, preserving modality-specific discriminative strengths while suppressing cross-modal redundant information interference.", + "bbox": [ + 89, + 489, + 482, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Formally, we present an Enhanced Multi-modal Perceptual (EMP) framework designed to hierarchically exploit intrinsic multi-modal cues from the representation level to the feature level. Specifically, our framework consists of two components: Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS). The CFE leverages the distinct frequency traits inherent in different modalities, utilizing the high-frequency information (e.g., edges) from stacked event frames to enhance the reconstructed event frames including mid-frequency information (e.g., texture). This process effectively refines the event representations for better feature extraction. In Addition, considering the discrepancy between modalities,", + "bbox": [ + 89, + 704, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "the HGS adaptively captures semantic consistency tokens across different modalities, thereby realizing unified encoding of multimodal features.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The main contributions are summarized as follows,", + "bbox": [ + 532, + 137, + 872, + 151 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- A privacy-preserving paradigm inspired by event-to-image generates texture-enriched surrogate frames as alternatives to RGB images, eliminating auxiliary data acquisition while bridging the modality gap with events.", + "- A novel Enhanced-Multimodal Perceptual (EMP) framework that hierarchically explores intrinsic multi-modal cues from raw event streams through dual-stage innovations at representation and feature levels.", + "- A multi-modal unified perception strategy equipped with collaborative Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enables progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level." + ], + "bbox": [ + 511, + 152, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 364, + 653, + 380 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Single-modal Event-based Action Recognition", + "text_level": 1, + "bbox": [ + 511, + 390, + 900, + 406 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Existing single-modal approaches for event-based action recognition (EAR) focus on extracting discriminative features from event streams [8, 52]. These methods leverage the sparse spatial nature of event data by stacking asynchronous event streams into compact frames, which are processed using powerful frame-specific architectures, e.g. convolutional neural networks (CNNs) [32, 44] or visual transformers (ViTs) [7, 57]. For instance, EventTransAct [10] introduce video transformers and event contrastive learning to extract fine-grained features from the event stream. Similarly, ExACT [57] incorporate rich semantic information to effectively handle event frames from a cross-modal perspective. Beyond frame-based methods, some works [37, 54] explore networks tailored to the unique properties of asynchronous events, such as spiking neural networks (SNNs) and graph convolutional networks (GCNs). For example, SpikePoint [37] proposed an end-to-end SNN architecture that processes sparse event streams by simultaneously extracting global and local features. VMvGCN [54] introduced a novel volumetric multi-view framework to capture key spatial and temporal information from event streams. However, these approaches primarily fail to fully exploit the complementary infomation inherent in raw event streams. In contrast, our proposed Enhanced Multi-modal Perceptual framework synergistically integrates features from different modalities.", + "bbox": [ + 511, + 412, + 906, + 806 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Multi-modal Event-based Action Recognition", + "text_level": 1, + "bbox": [ + 511, + 816, + 897, + 834 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances in robust representation for multi-modal data garner significant attention across various visual domains [7, 48]. For EAR, previous research explore two primary paradigms for multi-modal input: (i) Integrating", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5970", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/64cc22384c3bf98a7248cf3e92627ebd89f8fce5cbe56a5a2eb9a92ab9ca81c9.jpg", + "image_caption": [ + "Figure 2. The overall framework of our proposed EMP. First, complementary event modalities (stacked and reconstructed frames) undergo frequency-domain refinement via the Cross-modal Frequency Enhancer (CFE), which amplifies edge dynamics while preserving textural in reconstructed counterparts. After that, during the unified visual encoding process of CLIP, the High-frequency Guided Selector (HGS) is employed to capture semantic-consistency tokens across modalities, which effectively eliminates spatiotemporally redundant features. Finally, event features and text features derived from the text/vision encoder are contrasted for activity recognition." + ], + "image_footnote": [], + "bbox": [ + 114, + 97, + 883, + 375 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "different representations of event data (e.g., event frames, event points, and event voxels); and (ii) Fusing RGB images with event data. For instance, AGCN [19] introduce a dual point-voxel absorbing graph representation to exploit complementary information between event points and event voxels. EventCrab [6] propose a framework that synergistically processes event frames and event points balancing efficiency and effectiveness. However, while these methods fully leverage the rich temporal information within event streams, they overlook the inherent lack of texture information in events. To address this limitation, SAFE [24] leverages pre-trained large-scale vision-language models to fuse semantic labels, RGB frames, and event streams, improving multi-modal recognition accuracy. Similarly, sstformer [48] formally propose to recognize patterns by fusing RGB frames and event streams simultaneously. Despite these advances, existing methods require additional resources for RGB images and fail to fully exploit the consistency between multi-modal data. In contrast, our proposed framework eliminates dependency on RGB while explicitly addressing the impact of irrelevant information on multimodal feature selection. Our approach provides a promising solution for practical deployment in real-world scenarios.", + "bbox": [ + 91, + 486, + 483, + 834 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Token Selection in Transformer", + "text_level": 1, + "bbox": [ + 89, + 847, + 370, + 862 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "With the widespread adoption of Transformer across various domains [6, 20, 21, 34, 35], token selection has gar", + "bbox": [ + 89, + 869, + 482, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nered increasing attention for its ability to leverage inherent redundancies between image tokens to focus on key objects and reduce computational costs [12, 36, 47, 55]. This is particularly crucial in visual tasks [36], where the extraction of fine-grained features is essential for visual understanding. DynamicViT [18, 36] inserts prediction modules into Transformer blocks to predict and select tokens with higher information content. TransFG [18] directly employs attention mechanisms to locate important tokens, demonstrating strong performance in fine-grained classification. Building on this, STA [12] proposes a framework that simultaneously considers temporal redundancy and semantic importance for selecting spatiotemporal tokens. Magic [55] extends the token selection to multi-modal tasks, enabling the model to dynamically choose object-centric tokens based on different input modalities. Unlike these methods, we propose a more flexible multi-modal token selection scheme that utilizes high-frequency modal-guided strategies to capture semantic consistency tokens, thereby enhancing the robustness of our approach.", + "bbox": [ + 511, + 486, + 906, + 789 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 805, + 604, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Joint Formulation", + "text_level": 1, + "bbox": [ + 513, + 832, + 689, + 847 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We introduce a jointly formulated framework for event-to-image reconstruction and CLIP-driven action recognition, without the need for additionally captured RGB images.", + "bbox": [ + 511, + 854, + 906, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5971", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Event-to-Image Reconstruction. Considering the event stream $\\pmb{E} = \\{(u_i, p_i, t_i)\\}_{i \\in (0, N)}$ , where $u_i$ denotes the pixel position, $p_i$ indicates the polarity, $t_i$ denotes the time of each event point, and $N$ is the number of event points. We first generate two complementary event representations from raw asynchronous event streams: (i) Edge-dominant stacked frames $I_s$ obtained by temporally aggregating events; (ii) Texture-enriched reconstructed frames $I_r$ obtained by the image reconstruction network $H$ as:", + "bbox": [ + 89, + 90, + 483, + 226 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {\\mathrm {r}} = \\mathrm {H} (\\boldsymbol {E}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 243, + 241, + 482, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\pmb{I}_{\\mathrm{s}},\\pmb{I}_{\\mathrm{r}}\\in \\mathbb{R}^{T\\times H\\times W\\times 3}$", + "bbox": [ + 89, + 268, + 282, + 285 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "CLIP-driven Action Recognition. In terms of $C$ action classes, we follow [57] to adopt two different text prompts: hand-crafted prompt \"A series of photos recording action for [class]\" and learnable prompt \"[V1], [V2], ..., [Vn], [class]\", where [class] represents the class name and $[V_n]$ is the random initialized parameter. The CLIP-based text encoder [56] first encodes the two prompts to obtain D-dimensional text features $\\pmb{f}_1^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}$ and $\\pmb{f}_2^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}$ , and the final text feature $\\pmb{f}^{t}\\in \\mathbb{R}^{C\\times D}$ is obtained by averaging $\\pmb{f}_1^{\\mathrm{t}}$ and $\\pmb{f}_2^{\\mathrm{t}}$ . Meanwhile, we use the CLIP-based vision encoder [56] for event feature extraction, fully exploiting the potential of Visual Language Models (VLMs) in event-based recognition. The stacked frames $\\pmb{I}_{\\mathrm{s}}$ and reconstructed frames $\\pmb{I}_{\\mathrm{r}}$ after being enhanced for event representation, are fed into the vision encoder to obtain the event feature $\\pmb{f}^{\\mathrm{e}}\\in \\mathbb{R}^{1\\times D}$ . Eventually, the class probability $p\\in \\mathbb{R}^{1\\times C}$ for the action recognition are obtained as,", + "bbox": [ + 89, + 286, + 483, + 544 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\np = \\operatorname {s o f t m a x} _ {(1)} \\left(\\boldsymbol {f} ^ {\\mathrm {e}} \\left(\\boldsymbol {f} ^ {\\mathrm {t}}\\right) ^ {\\top}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 556, + 482, + 582 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nc = \\operatorname {a r g m a x} (p),\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 579, + 316, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{softmax}(\\cdot)$ denotes the softmax function. By taking the category of maximum probability, we obtain the final action prediction class $c$ for the event stream.", + "bbox": [ + 89, + 607, + 483, + 654 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Enhanced-Multimodal Perceptual (EMP)", + "text_level": 1, + "bbox": [ + 89, + 665, + 444, + 681 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To preserve diverse information within and across modalities while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. Fig. 2 presents the architecture of EMP. It is equipped with Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enabling progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level.", + "bbox": [ + 89, + 688, + 482, + 823 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Cross-modal Frequency Enhancer (CFE). Event frames from distinct modalities demonstrate complementary spectral characteristics [55]. Specifically, stacked frames from asynchronous event streams predominantly focus on capturing high-frequency edge dynamics of transient actions,", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "whereas reconstructed frames preserve mid-frequency texture inherent to activity semantics. Both modalities are crucial for event action recognition, yet effectively harnessing their complementary strengths remains a critical challenge. The direct combination of these heterogeneous modalities yields suboptimal solutions because of their inherent discrepancies. Thus, we propose a Cross-modal Frequency Enhancer (CFE) that strategically leverages frequency traits across modalities to refine event representation. It enables the model to focus on more comprehensive features through frequency-domain enhancement, thereby optimizing the synergy between modality-specific information.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Whitin the CFE, we first transform stacked frames $I_{\\mathrm{s}}$ and reconstructed frames $I_{\\mathrm{r}}$ into the frequency domain via 3D Fast Fourier Transform (FFT) [29] as:", + "bbox": [ + 511, + 272, + 903, + 316 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}), \\\\ \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}) \\end{array} \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 325, + 903, + 353 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {r}}).\n$$\n", + "text_format": "latex", + "bbox": [ + 653, + 349, + 764, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The spectral conversion enables the explicit decomposition of modality-specific frequency components, where edge-dominant stacked frames predominantly occupy high-frequency bands while texture-abundant reconstructed frames exhibit concentrated energy in mid-frequency regions. The frequency-aware representation establishes a unified foundation for cross-modal information enhancement. Subsequently, to selectively amplify edge-dynamic information in stacked frames from the frequency spectrum, we apply a sigmoid layer to $I_{\\mathrm{s}}^{\\mathrm{ftt}}$ , generating a frequency filter A. The proposed filtering is calculated as follows:", + "bbox": [ + 511, + 375, + 906, + 541 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {A} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f t t}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 651, + 549, + 903, + 577 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t} ^ {\\prime}} = \\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} \\odot \\mathrm {A},\n$$\n", + "text_format": "latex", + "bbox": [ + 637, + 571, + 748, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\odot$ denotes element-wise multiplication. Finally, we transfer the $I_{\\mathrm{r}}^{\\mathrm{fft}'}$ in the frequency space into the spatial domain to acquire enhanced reconstructed frames $I_{\\mathrm{r}}'$ using an inverse Fast Fourier Transform (iFFT) [29] as:", + "bbox": [ + 511, + 601, + 906, + 664 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\prime} = \\mathrm {i F F T} \\left(\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\text {f f t}} ^ {\\prime}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 647, + 674, + 903, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "High-frequency Guided Selector (HGS). Given the inherent semantic coherence across modalities derived from the same event stream, existing methods [48] that sequentially extract and aggregate distinct modal features may introduce redundant feature interference. Building on cross-modal frequency enhancement, we further propose a high-frequency guided selection strategy to adaptively prioritize inter-modal critical regions that jointly characterize event semantics. Specifically, both stacked frames and enhanced reconstructed frames are processed through the shared vision encoder for visual feature learning. During event feature encoding, the [cls] token inherently aggregates global semantic representations through self-attention", + "bbox": [ + 511, + 704, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5972", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "mechanisms [17]. We leverage high-frequency [cls] token from stacked frames to steer the attention mechanisms of reconstructed frames, enabling selective focus on cross-modal consistent semantic regions while suppressing modality-specific interference.", + "bbox": [ + 89, + 90, + 480, + 165 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For stacked frames $I_{\\mathrm{s}}$ and enhanced reconstructed frames $I_{\\mathrm{r}}^{\\prime}$ , we first obtain the corresponding non-overlapping patches $f_{\\mathrm{s}}, f_{\\mathrm{r}} \\in \\mathbb{R}^{T \\times (N + 1) \\times D}$ including the special [cls] token $f_{\\mathrm{s}}^{\\mathrm{cls}}, f_{\\mathrm{r}}^{\\mathrm{cls}} \\in \\mathbb{R}^{1 \\times D}$ [2] with D-dimension through a linear projection, where $N = H \\times W / P^2$ and $(P, P)$ is the size of the patch. Between $l$ -th and $(l + 1)$ -th layer of the vision encoder, we deploy a cross-attention mechanism to facilitate inter-modal interaction. Here, the reconstructed tokens serve as query feature $Q$ , while the stacked tokens are transformed into key and value features, i.e. $K$ and $V$ . The detailed procedure can be formulated as:", + "bbox": [ + 89, + 167, + 482, + 332 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} Q = W _ {\\mathrm {q}} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}, K = W _ {\\mathrm {k}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, V = W _ {\\mathrm {v}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, \\\\ \\boldsymbol {f} _ {\\mathrm {s}} ^ {l} = \\operatorname {C r o s s A t t e n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V, \\tag {6} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 339, + 480, + 396 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $W_{\\mathrm{q}}, W_{\\mathrm{k}}, W_{\\mathrm{v}} \\in \\mathbb{R}^{D \\times D}$ are the weights of the linear transformations corresponding to $Q, K, V$ , respectively.", + "bbox": [ + 89, + 402, + 482, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Subsequently, we extract the [cls] token $f_{\\mathrm{s}}^{\\mathrm{cls}}$ that encapsulates global semantics from the stacked features $f_{\\mathrm{s}}^{l}$ and compute its semantic affinity with the patch tokens in the augmented reconstructed features $f_{\\mathrm{r}}^{l}$ via the cosine similarity. It can be formulated as:", + "bbox": [ + 89, + 435, + 483, + 510 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\sin \\left(\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}}, \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}\\right) = \\frac {\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} {} ^ {\\top} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}}{\\| \\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} \\| \\| \\boldsymbol {f} _ {\\mathrm {r}} ^ {l} \\|}. \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 186, + 518, + 482, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "By averaging the aforementioned similarity matrix $\\mathrm{sim}(\\pmb{f}_{\\mathrm{s}}^{\\mathrm{cls}},\\bar{\\pmb{f}}_{\\mathrm{s}}^{l})$ along the temporal dimension $T$ , the final attention score matrix $B\\in \\mathbb{R}^{T\\times N}$ is obtained. To maintain semantic consistency between the two modalities, we use $B$ to incorporate dynamic edge features, which guide the reconstruction branch in adaptively selecting tokens. The selection is formalized as:", + "bbox": [ + 89, + 564, + 483, + 667 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nS = \\left\\{\\operatorname {T o p} _ {K} (B) \\right\\}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 223, + 678, + 482, + 696 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathrm{Top}_K$ is a function that sorts scores in a set in descending order and outputs the indices corresponding to the $K$ largest values, $K$ is obtained by $\\mu \\cdot N$ and $\\mu$ is a selection parameter. Using index set $S$ , we select the corresponding critical tokens from $\\pmb{f}_{\\mathrm{r}}^{l}$ and merge them with the $\\pmb{f}_{\\mathrm{r}}^{\\mathrm{cls}}$ token before inputting them into the $(l + 1)$ -th visual encoder.", + "bbox": [ + 89, + 704, + 482, + 794 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Training Objectives", + "text_level": 1, + "bbox": [ + 89, + 801, + 279, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Modality Consistency Loss. Our model takes the edge-stacked frames and the texture-reconstructed frames as input, which are crucial for capturing the discriminative features of the event stream. The HGS we proposed above focuses on critical token selection guided by high-frequency", + "bbox": [ + 89, + 824, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "modality, based on the principle that different modalities maintain semantic consistency. Therefore, we introduce a loss to reduce the discrepancy between the two features of the same event in the feature space. The semantic consistency cross modalities $(f_{\\mathrm{s}} / f_{\\mathrm{r}})$ are constrained by the Mean Square Error (MSE) loss:", + "bbox": [ + 511, + 90, + 905, + 181 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\boldsymbol {f} _ {\\mathrm {s}} ^ {i} - \\boldsymbol {f} _ {\\mathrm {r}} ^ {i} \\right| \\right| _ {2} ^ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 194, + 905, + 237 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $N$ is the number of tokens.", + "bbox": [ + 511, + 247, + 740, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Contractive Recognition Loss. The event-text consistency $\\mathcal{L}_{\\mathrm{r}}$ is constrained by the contrastive loss between the event feature and the text feature as follows:", + "bbox": [ + 511, + 263, + 905, + 306 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {i} ^ {\\mathrm {t}} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {j} ^ {\\mathrm {t}} / \\tau\\right)}, \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 318, + 905, + 357 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tau$ is the temperature coefficient.", + "bbox": [ + 511, + 369, + 767, + 383 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The final loss is composed of the modality consistency loss $\\mathcal{L}_{\\mathrm{c}}$ and the contractive recognition loss $\\mathcal{L}_{\\mathrm{r}}$ as", + "bbox": [ + 511, + 383, + 905, + 415 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda \\mathcal {L} _ {\\mathrm {r}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 655, + 429, + 903, + 445 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda$ denotes the parameter used to balance the loss optimization.", + "bbox": [ + 511, + 458, + 905, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 503, + 645, + 521 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Datasets and Settings", + "text_level": 1, + "bbox": [ + 511, + 530, + 714, + 546 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. To evaluate the performance of our approach, we conduct research on three datasets, including PAF [30], SeAct [57], HARDVS [50] and DVS128 Gesture [1]. DVS Action, also known as PAF, contains 10 action categories with 450 recordings, conducted in an unoccupied office environment. SeAct is the first semantic-abundant dataset for event-text action recognition, containing 58 actions under four themes. HARDVS is currently the largest dataset for event-based action recognition, currently having 107,646 recordings for 300 action categories. Both of the above three datasets have an average time duration of 5 seconds with $346 \\times 260$ resolution. DVS128 Gesture is collected using a DAVIS128 camera with $128 \\times 128$ resolution, dividing into 11 distinct classes of gestures.", + "bbox": [ + 511, + 551, + 906, + 763 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Settings. The implementation of the overall framework is carried out on PyTorch in a Linux environment with two NVIDIA GeForce 4090 GPUs. We use the Adam [23] optimizer with the initial learning rate of $1 \\times 10^{-5}$ and weight decay of $2 \\times 10^{-4}$ . CosineAnnealingLR learning rate schedule is employed with a minimum learning rate of $1 \\times 10^{-6}$ . The representative lightweight FireNet [40] (only 38k parameters) for event-to-image reconstruction, which has low latency (1ms for $240 \\times 180$ ).", + "bbox": [ + 511, + 763, + 905, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5973", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/557ad002392a77660f1227f7f588b13ae9cd38207f36d1aa4c9e77786bf64d55.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodEvent RepresentationTop-1 Accuracy (%)
Frame-basedPoint-basedPAFSeActDVS128 Gesture
Motion SNN [28]-✓78.10-92.70
Slayer [41]-✓--93.64
MST [51]✓-88.21--
EV-ACT [14]✓-92.60--
EventTransAct [10]✓--57.8197.92
EvT [39]✓--61.3096.20
SpikePoint [37]-✓90.60-98.74
SpikMamba [8]✓-96.2871.0299.01
ExACT [57]✓-94.8367.2498.86
EventCrab [7]✓✓96.4972.4198.80
EventCLIP [56] + Recon✓-96.5571.0596.59
Ours✓-99.80(+3.31)75.00(+2.59)98.86(-0.15)
", + "bbox": [ + 143, + 88, + 857, + 310 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with SOTA Methods", + "text_level": 1, + "bbox": [ + 89, + 364, + 382, + 382 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Datasets without Paired RGB. As shown in Tab. 1, we compare the proposed method EMP with several representative event-based action recognition methods on three event-only datasets (PAF, SeAct, DVS128 Gesture), where paired RGB data is unavailable. All compared methods exclusively utilize event-driven representations, categorized into event frame-based and event point-based paradigms. As demonstrated in Tab. 1, point-based methods exhibit marginally inferior performance compared to frame-based approaches on the PAF and DVS128 Gesture datasets. This discrepancy stems from the latter's enhanced capability in extracting discriminative spatiotemporal features through visual architectural designs. EventCrab [7] achieves impressive results by collaboratively mining complementary information from both event points and frames, thereby leveraging multi-modal event representations. However, unlike the paradigm of primarily exploring the sparse spatial characteristics of event streams from EventCrab, our method introduces texture-rich reconstructed frames and holistically exploits intrinsic correlations and extrinsic specificities between event frames and reconstructed frames. This unified perceptual framework yields significantly improved recognition efficacy, attaining $3.31\\%$ (Top-1) and $2.59\\%$ (Top-1) accuracy gains on the PAF and SeAct datasets respectively, while maintaining competitive performance on the DVS128 Gesture dataset.", + "bbox": [ + 91, + 395, + 483, + 787 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Comparison on Datasets with Paired RGB. As shown in Tab. 2, we compare the proposed EMP with several representative EAR methods on the HARDVS dataset, which includes 300-class paired RGB-Event data. Comparative methods are categorized into two types: (i) those utilizing RGB modality and (ii) those solely relying on event streams. We can find that combining RGB frames with", + "bbox": [ + 89, + 794, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/ccd989ef29ce6ae73d1e0541ae8bcca9ca1e14fa41c3ce826de571be1589f435.jpg", + "table_caption": [ + "Table 1. Comparative performance for EAR on the PAF, SeAct and DVS128 Gesture datasets. The best results are in bold and the second-best ones are in underlined." + ], + "table_footnote": [], + "table_body": "
MethodRGBParam.Top-1 Accuracy (%)
R2Plus1D [46]X63.5M49.06
C3D [45]X147.2M49.94
TimeSformer [4]X121.2M50.77
ESTF [50]X46.1M51.22
ExACT [57]X155.1M90.10
ESTF* [50]✓76.8M49.93
SAFE [24]✓-50.17
C3D* [45]✓245.3M50.88
TimeSformer* [4]✓202.3M51.57
SSTFormer [48]✓336.4M52.97
OursX182.1M97.68
", + "bbox": [ + 517, + 363, + 905, + 534 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Comparative performance (\\%) on the HARDVS datasets. The best results are in **bold. (·)* denotes the combination between RGB frames and event streams.", + "bbox": [ + 511, + 544, + 906, + 585 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "event streams yields marginal improvements over the single modality-based version. For instance, C3D [45] achieves an increase from $49.94\\%$ to $50.88\\%$ (Top-1), while TimeSformer improves from $50.77\\%$ to $51.57\\%$ (Top-1). However, the significant semantic discrepancy between RGB and event modalities limits performance gains from superficial cross-modal fusion. Different from conventional fusion paradigms, EMP adopts the text-aligned adaptation strategy to bridge the modality gap, while leveraging frequency-domain characteristics to extract discriminative event features. The results in Tab. 2 validate the efficacy of our approach, demonstrating superior robustness in handling heterogeneous modality.", + "bbox": [ + 511, + 612, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 818, + 679, + 832 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of Each Component. As detailed in Tab. 3, we conduct an ablation study to evaluate the contributions of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS) within the pro", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5974", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/752cd90724c1a703f684f8db80ab3c86998aaae49ae0d1cdc955035dd456a677.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
CFEHGSSeAct (%)DVS128 Gesture (%)
XX71.2597.34
✓X72.4197.72
X✓73.2798.10
✓✓75.0098.86
", + "bbox": [ + 114, + 88, + 460, + 167 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/9e0cc1fa671869d0a34a75b473f36eacded77357cfb8a0dea6bd33ae0173df8a.jpg", + "table_caption": [ + "Table 3. Ablation study for effectiveness of each component in the proposed method EMP." + ], + "table_footnote": [], + "table_body": "
Selection ratioTop-1 (%)Top-5 (%)
0.270.6890.51
0.471.5592.24
0.675.0095.68
0.873.2793.96
1.072.4193.10
", + "bbox": [ + 143, + 220, + 434, + 314 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "posed EMP. The baseline model trained only on eventreconstructed frames without incorporating the CFE and HGS achieves $71.55\\%$ accuracy on SeAct and $97.34\\%$ on DVS128 Gesture. When adding the baseline with CFE, we can find the performance improvements of $0.86\\%$ on SeAct and $0.38\\%$ on DVS128 Gesture, respectively. These performance improvements demonstrate the effectiveness of cross-modal representation enhancement in the frequency domain. Furthermore, the introduction of HGS achieves dynamic token selection across modalities, resulting in a $1.72\\%$ improvement on SeAct and a $0.76\\%$ improvement on DVS128 Gesture. By integrating all components, our model achieves the optimal performance. These results validate the effectiveness of our EMP in different scenarios.", + "bbox": [ + 89, + 367, + 483, + 580 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Diagnose Studies", + "text_level": 1, + "bbox": [ + 89, + 604, + 256, + 619 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Different Token Selection Ratio. As illustrated in Tab. 4, we investigate the impact of varying selection thresholds $\\mu$ on the SeAct dataset, where $\\mu$ governs the proportion of semantically consistent tokens selected across modalities in the HGS framework. The experimental results reveal that the recognition accuracy improves as $\\mu$ increases, reaching its peak at $\\mu = 0.6$ , beyond which further threshold escalation correlates with performance degradation. This trend underscores the inherent divergence between modalities, suggesting that indiscriminate cross-modal fusion is suboptimal. Critically, calibrating $\\mu$ to an appropriate level optimally harnesses intermodal complementarity while suppressing shared redundancies.", + "bbox": [ + 89, + 628, + 482, + 824 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Different Selection Layer. As shown in Tab. 5, we investigate the impact of integrating token selection at distinct layers of the vision encoder on the SeAct dataset, with the optimal selection ratio fixed at $\\mu = 0.6$ (Tab. 4). We implement modality-guided token selection across indi", + "bbox": [ + 89, + 825, + 482, + 901 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/ce463d6cc28215eb1d610d8e4fdd506168cfd14d7049ce8caddcc6066f42d36a.jpg", + "table_caption": [ + "Table 4. Different selection ratio on the SeAct dataset." + ], + "table_footnote": [], + "table_body": "
Insertion LayerTop-1 (%)Top-5 (%)
[4]73.2793.96
[6]73.9692.41
[8]75.0095.68
[10]74.1393.10
", + "bbox": [ + 566, + 88, + 857, + 165 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 5. Different insertion layer of token selection strategy on the SeAct dataset.", + "bbox": [ + 511, + 176, + 903, + 203 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0b9b8ad542c1b4adb41fc6a2298fcaa12c0d35769d7fa6cf5f1ad4044b3e82e4.jpg", + "image_caption": [ + "Figure 3. Effect of different values of $\\lambda$ on SeAct and DVS 128 Gesture datasets." + ], + "image_footnote": [], + "bbox": [ + 586, + 229, + 823, + 388 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "vidual layers and observe that shallow and deep layers yield marginal performance gains. In contrast, layer 8 achieves peak Top-1, validating that appropriately positioned layers within vision encoders optimally encode cross-modal feature representations, thereby enabling effective modality-guided selection of semantically consistent tokens.", + "bbox": [ + 511, + 459, + 906, + 549 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Balance Between Consistency Loss and Recognition Loss. Our empirical validation substantiates that achieving equilibrium between modality consistency loss and contractive recognition (Eq. 11) loss constitutes a pivotal design criterion for robust multimodal learning. The contribution of $\\mathcal{L}_{\\mathrm{r}}$ is modified via $\\lambda$ . As shown in Fig. 3, the variation of $\\lambda$ affects the accuracy performance. The best performance is achieved with weights of 0.6, 0.4 on SeAct and DVS128 Gesture datasets respectively. It is worth noting that our findings highlight the potential benefits of leveraging consistency across modalities to enhance the representational capacity and recognition robustness of our EMP.", + "bbox": [ + 511, + 551, + 908, + 733 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.5. Qualitative Analysis", + "text_level": 1, + "bbox": [ + 511, + 742, + 705, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Visualization of Cross-modal Frequency Enhancer (CFE). To investigate the efficacy of the designed CFE, we visualize the four variants of event frames on the SeAct dataset: raw stacked frames, reconstructed frames, enhanced frames by sharpening operators, and enhanced frames by our CFE. As shown in fig. 4, conventional edge-enhancement techniques [33] yield marginal improvements over raw reconstructed frames. In contrast, event frames processed by CFE consistently highlight the critical regions", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5975", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/5a05f721c0097c87947867fdae625e46254ecd1a8ad66f8911cea2ebec49b5fe.jpg", + "image_caption": [ + "Figure 4. Visualization of event frames. \"Stack\" and \"Reconstruct\" denote stacked and reconstructed frames, respectively; \"Edge Enhance\" indicates enhanced reconstructed frames employing sharpening operators; \"CFE\" represents enhanced reconstructed frames by our Cross-modal Frequency Enhancer." + ], + "image_footnote": [], + "bbox": [ + 127, + 90, + 467, + 362 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "across both individual actions and interaction scenarios. Our qualitative analysis confirms that the proposed CFE selectively refines the high-frequency components of the reconstructed frames through spectral coordination with the edge dynamics of the stacked frames, thereby generating complementary event representations with synergistic spatial characteristics.", + "bbox": [ + 89, + 473, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Guidanced Selection. Fig. 5 visualizes the important tokens selected by the High-frequency Guided Selector (HGS). As shown in Fig. 5, the HGS demonstrates the adaptive selection of action-salient tokens across sequential event frames. This demonstrates the intrinsic consistency among multi-modal frames originating from the identical event stream, where HGS effectively identifies semantically coherent tokens while suppressing interference from spatiotemporal redundancies. The results highlight the ability of HGS to focus on critical regions of interest, thereby enhancing the discriminative power of EMP. This selection is crucial for improving the robustness and accuracy of action recognition in dynamic scenes.", + "bbox": [ + 89, + 580, + 482, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualization of Event Features Learned by EMP. As illustrated in Fig. 6, we visualize the feature distributions of our proposed EMP against the baseline on the SeAct dataset. Seven action classes are randomly selected to demonstrate inter-class feature separability. Empirical results demonstrate the capability of our EMP in achieving both intra-class compactness and inter-class dispersion, further substantiating its efficacy in preserving semantic dis", + "bbox": [ + 89, + 779, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/04aff77dad66b914df2ec9e5e3ff7b8500b256e54e47ba40471d68f0758ffdc0.jpg", + "image_caption": [ + "Figure 5. Visualization of cross-modal consistent tokens selected by the HGS, with unselected regions masked in white." + ], + "image_footnote": [], + "bbox": [ + 542, + 90, + 874, + 281 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4dfdd280301d7d5eb199bfadddda1c4fa644873a1f55848e0ab5ff538c7e00e4.jpg", + "image_caption": [ + "Figure 6. The t-SNE visualization of 7 randomly selected actions. Different color indicates different actions." + ], + "image_footnote": [], + "bbox": [ + 540, + 343, + 883, + 468 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tinctiveness across classes.", + "bbox": [ + 513, + 536, + 692, + 549 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion and Discussion", + "text_level": 1, + "bbox": [ + 511, + 565, + 761, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Conclusion. In this work, we propose a privacy-preserving paradigm that eliminates RGB dependencies by generating texture-enriched reconstructed frames directly from raw event streams, reducing multi-sensor acquisition costs while bridging the modality gap with event frames. To preserve diverse information across reconstructed and stacked frames while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. It consists of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), which enables the progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level to acquire discriminative event features. Extensive experimental results validate the effectiveness of the proposed method.", + "bbox": [ + 511, + 590, + 906, + 816 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Discussion. While our method achieves promising performance in EAR and demonstrates the potential for scalability, its ability to handle more challenging scenarios remains to be explored. Future work will focus on developing a robust multi-modal unified perceptual approach to cope with diverse and variable scenes.", + "bbox": [ + 511, + 816, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5976", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208), and the Nanjing University Integrated Research Platform of the Ministry of Education-Top Talents Program (Grant No. 2024300482).", + "bbox": [ + 89, + 92, + 485, + 161 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 174, + 187, + 189 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7243-7252, 2017. 5", + "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6836-6846, 2021. 5", + "[3] Sami Barchid, Benjamin Allaert, Amel Aissaoui, José Mennesson, and Chaabane C Djeraba. Spiking-fer: spiking neural network for facial expression recognition with event cameras. In Proceedings of the 20th International Conference on Content-based Multimedia Indexing, pages 1–7, 2023. 2", + "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning, page 4, 2021. 6", + "[5] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9286-9296, 2024. 1", + "[6] Meiqi Cao, Rui Yan, Xiangbo Shu, Jiachao Zhang, Jinpeng Wang, and Guo-Sen Xie. Mup: multi-granularity unified perception for panoramic activity recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7666-7675, 2023. 3", + "[7] Meiqi Cao, Xiangbo Shu, Jiachao Zhang, Rui Yan, Zechao Li, and Jinhui Tang. Eventcrab: Harnessing frame and point synergy for event-based action recognition and beyond. arXiv preprint arXiv:2411.18328, 2024. 2, 6", + "[8] Jiaqi Chen, Yan Yang, Shizhuo Deng, Da Teng, and Liyuan Pan. Spikmamba: When snn meets mamba in event-based human action recognition. In Proceedings of the 6th ACM International Conference on Multimedia in Asia, pages 1-8, 2024. 2, 6", + "[9] Lan Chen, Haoxiang Yang, Pengpeng Shao, Haoyu Song, Xiao Wang, Zhicheng Zhao, Yaowei Wang, and Yonghong Tian. Velora: A low-rank adaptation approach for efficient rgb-event based recognition. arXiv preprint arXiv:2412.20064, 2024. 2", + "[10] Tristan de Blegiers, Ishan Rajendrakumar Dave, Adeel Yousaf, and Mubarak Shah. Eventtransact: A video transformer-based framework for event-camera based action recognition. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1-7, 2023. 2, 6", + "[11] Yongjian Deng, Hao Chen, and Youfu Li. A dynamic gcn with cross-representation distillation for event-based learn" + ], + "bbox": [ + 93, + 199, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1492-1500, 2024. 2", + "[12] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 3", + "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 1", + "[14] Yue Gao, Jiaxuan Lu, Siqi Li, Nan Ma, Shaoyi Du, Yipeng Li, and Qionghai Dai. Action recognition and benchmark using event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14081-14097, 2023. 6", + "[15] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1", + "[16] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023. 1", + "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 5", + "[18] Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, and Changhu Wang. Transfg: A transformer architecture for fine-grained recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 852-860, 2022. 3", + "[19] Bo Jiang, Chengguo Yuan, Xiao Wang, Zhimin Bao, Lin Zhu, Yonghong Tian, and Jin Tang. Point-voxel absorbing graph representation learning for event stream based recognition. arXiv preprint arXiv:2306.05239, 2023. 3", + "[20] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2570-2578, 2024. 3", + "[21] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 3", + "[22] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1", + "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5", + "[24] Dong Li, Jiandong Jin, Yuhao Zhang, Yanlin Zhong, Yaoyang Wu, Lan Chen, Xiao Wang, and Bin Luo." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5977", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Semantic-aware frame-event fusion based pattern recognition via large vision-language models. arXiv preprint arXiv:2311.18592, 2023. 2, 3, 6", + "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1", + "[26] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 1", + "[27] Haoyue Liu, Shihan Peng, Lin Zhu, Yi Chang, Hanyu Zhou, and Luxin Yan. Seeing motion at nighttime with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25648-25658, 2024. 1", + "[28] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In International Joint Conference on Artificial Intelligence, pages 1743-1749, 2021. 6", + "[29] Shuangbiao Liu, Qian Wang, and Geng Liu. A versatile method of discrete convolution and fft (dc-fft) for contact analyses. Wear, 243(1-2):101-111, 2000. 4", + "[30] Shu Miao, Guang Chen, Xiangyu Ning, Yang Zi, Kejia Ren, Zhenshan Bing, and Alois Knoll. Neuromorphic vision datasets for pedestrian detection, action recognition, and fall detection. Frontiers in Neurorobotics, 13:38, 2019. 5", + "[31] Federico Paredes-Valles and Guido CHE De Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 1", + "[32] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 19935-19947, 2022. 2", + "[33] R Pushpavalli and G Sivaradje. Switching median filter for image enhancement. International Journal of Scientific & Engineering Research, 3(2):1-5, 2012. 7", + "[34] Hongyu Qu, Rui Yan, Xiangbo Shu, Hailiang Gao, Peng Huang, and Guo-Sen Xie. Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition. arXiv preprint arXiv:2405.02077, 2024. 3", + "[35] Hongyu Qu, Jianan Wei, Xiangbo Shu, and Wenguan Wang. Learning clustering-based prototypes for compositional zero-shot learning. In Proceedings of the International Conference on Learning Representations, 2025. 3", + "[36] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in Neural Information Processing Systems, 34:13937-13949, 2021. 3", + "[37] Hongwei Ren, Yue Zhou, Yulong Huang, Haotian Fu, Xiaopeng Lin, Jie Song, and Bojun Cheng. Spikepoint: An efficient point-based spiking neural network for event cam" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "eras action recognition. arXiv preprint arXiv:2310.07189, 2023. 2, 6", + "[38] Hongwei Ren, Yue Zhou, Jiadong Zhu, Xiaopeng Lin, Hao-tian Fu, Yulong Huang, Yuetong Fang, Fei Ma, Hao Yu, and Bojun Cheng. Rethinking efficient and effective point-based networks for event camera classification and regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2", + "[39] Alberto Sabater, Luis Montesano, and Ana C Murillo. Event transformer. a sparse-aware solution for efficient event data processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2677-2686, 2022. 6", + "[40] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 1, 2, 5", + "[41] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 31, 2018. 6", + "[42] Xiangbo Shu, Binqian Xu, Liyan Zhang, and Jinhui Tang. Multi-granularity anchor-contrastive representation learning for semi-supervised skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7559-7576, 2022. 1", + "[43] Xiangbo Shu, Jiawen Yang, Rui Yan, and Yan Song. Expansion-squeeze-excitation fusion network for elderly activity recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5281-5292, 2022. 1", + "[44] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3): 3200-3225, 2022. 2", + "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4489-4497, 2015. 6", + "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6", + "[47] Hongjie Wang, Bhishma Dedhia, and Niraj K Jha. Zero-prune: Zero-shot token pruning through leveraging of the attention graph in pre-trained transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16070-16079, 2024. 3", + "[48] Xiao Wang, Zongzhen Wu, Yao Rong, Lin Zhu, Bo Jiang, Jin Tang, and Yonghong Tian. Sstformer: Bridging spiking neural network and memory support transformer for frame-event based recognition. arXiv preprint arXiv:2308.04369, 2023. 2, 3, 4, 6", + "[49] Xiao Wang, Shiao Wang, Pengpeng Shao, Bo Jiang, Lin Zhu, and Yonghong Tian. Event stream based human action" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5978", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "recognition: a high-definition benchmark dataset and algorithms. arXiv preprint arXiv:2408.09764, 2024. 2", + "[50] Xiao Wang, Zongzhen Wu, Bo Jiang, Zhimin Bao, Lin Zhu, Guoqi Li, Yaowei Wang, and Yonghong Tian. Hardvs: Revisiting human activity recognition with dynamic vision sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5615-5623, 2024. 2, 5, 6", + "[51] Ziqing Wang, Yuetong Fang, Jiahang Cao, Qiang Zhang, Zhongrui Wang, and Renjing Xu. Masked spiking transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1761-1771, 2023. 6", + "[52] Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. In European Conference on Computer Vision, pages 310-328, 2024. 1, 2", + "[53] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 1", + "[54] Bochen Xie, Yongjian Deng, Zhanpeng Shao, Hai Liu, and Youfu Li. Vmv-gcn: Volumetric multi-view based graph cnn for event stream classification. IEEE Robotics and Automation Letters, 7(2):1976-1983, 2022. 2", + "[55] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17117-17126, 2024. 3, 4", + "[56] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Eventbind: Learning a unified representation to bind them all for event-based open-world understanding. In European Conference on Computer Vision, pages 477-494, 2024. 4, 6", + "[57] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024. 2, 4, 5, 6" + ], + "bbox": [ + 91, + 90, + 482, + 641 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5979", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_model.json b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_model.json new file mode 100644 index 0000000000000000000000000000000000000000..5717fc16f90c021206c332b09e7c211ea0a68616 --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_model.json @@ -0,0 +1,2224 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.896, + 0.176 + ], + "angle": 0, + "content": "Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.175, + 0.203, + 0.823, + 0.275 + ], + "angle": 0, + "content": "Meiqi Cao\\(^{1}\\), Xiangbo Shu\\(^{1*}\\), Xin Jiang\\(^{1}\\), Rui Yan\\(^{1,2}\\), Yazhou Yao\\(^{1}\\), Jinhui Tang\\(^{3}\\) \n\\(^{1}\\)Nanjing University of Science and Technology \n\\(^{2}\\)Nanjing University \n\\(^{3}\\)Nanjing Forestry University" + }, + { + "type": "text", + "bbox": [ + 0.17, + 0.277, + 0.823, + 0.293 + ], + "angle": 0, + "content": "{cmq123,shuxb,xinjiang,ruiyan,yazhou.yao}@njust.edu.cn,tangjh@njfu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.327, + 0.327, + 0.343 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.359, + 0.486, + 0.783 + ], + "angle": 0, + "content": "While event cameras excel in capturing microsecond temporal dynamics, they suffer from sparse spatial representations compared to traditional RGB data. Thus, multimodal event-based action recognition approaches aim to synergize complementary strengths by independently extracting and integrating paired RGB-Event features. However, this paradigm inevitably introduces additional data acquisition costs, while eroding the inherent privacy advantages of event-based sensing. Drawing inspiration from event-to-image reconstruction, texture-enriched visual representation directly reconstructed from asynchronous event streams is a promising solution. In response, we propose an Enhanced Multimodal Perceptual (EMP) framework that hierarchically explores multimodal cues (e.g., edges and textures) from raw event streams through two synergistic innovations spanning representation to feature levels. Specifically, we introduce Cross-Modal Frequency Enhancer (CFE) that leverages complementary frequency characteristics between reconstructed frames and stacked frames to refine event representations. Furthermore, to achieve unified feature encoding across modalities, we develop High-Frequency Guided Selector (HGS) for semantic consistency token selection guided by dynamic edge features while suppressing redundant multimodal information interference adaptively. Extensive experiments on four benchmark datasets demonstrate the superior effectiveness of our proposed framework. The code is available at https://github.com/caomq123/EMP." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.809, + 0.222, + 0.825 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.835, + 0.484, + 0.88 + ], + "angle": 0, + "content": "Traditional frame-based cameras often encounter performance limitations when dealing with rapid motion or complex illumination conditions [22, 25, 42, 43]. In contrast," + }, + { + "type": "image", + "bbox": [ + 0.53, + 0.33, + 0.892, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.43, + 0.908, + 0.529 + ], + "angle": 0, + "content": "Figure 1. (a) Previous multi-modal input approaches independently extract features from paired RGB-Event modalities, both introducing excessive RGB data dependencies and disrupting the privacy-preserving property unique to event cameras. (b) Our approach takes a lightful reconstruct network (Firenet [40]) to substitute RGB data while narrowing the modality divergence with events, thereby enabling enhanced unified multimodal perception." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.907, + 0.851 + ], + "angle": 0, + "content": "event-driven neuromorphic visual sensors, also known as event cameras [13, 15], operate by asynchronously capturing per-pixel intensity changes only when a significant light change occurs. This unique functionality endows event cameras with several prominent advantages, including high dynamic range, low power consumption, and high temporal resolution, which enable them to excel in challenging scenarios such as fast motion, extreme lighting conditions, and privacy-sensitive applications. These attributes have led to their widespread adoption in various computer vision tasks, including object detection and tracking [16, 26, 52], motion deblurring [5, 27], and image reconstruction [31, 40, 53]. Notably, the inherent sensitivity of event cameras to dynamic scenes provides a unique advantage in capturing motion-related actions. Leveraging these strengths, the task of Event-based Action Recognition (EAR) has emerged, aiming to effectively extract action features from event streams to enhance recognition performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Currently, the EAR methods faces significant challenges in effectively utilizing the spatially sparse and asynchronous event streams. Traditional single-modal input approaches" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.235, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5969" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.258 + ], + "angle": 0, + "content": "adopt two primary strategies: (i) Processing asynchronous event points through point-based encoders (GCNs [11, 54]) or spiking neural networks (SNNs [3, 8, 37]), and (ii) Applying CNN- [32, 44], Transformer- [7, 57] or Mamba- [38, 49] based visual encoders to frames stacked from asynchronous event streams. While point-based methods accommodate the asynchronous nature of events effectively, they exhibit suboptimal recognition performance. Simultaneously, frame-based methods suffer from the lack of critical texture information due to the focus on dynamic edge capture inherent in event cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.261, + 0.483, + 0.488 + ], + "angle": 0, + "content": "In recent years, there has been a notable shift toward multi-modal input strategies [9, 24, 48, 50], leveraging diverse data sources to improve feature extraction robustness for practical applications. As depicted in Fig. 1(a), conventional multi-modal methods typically process RGB images and event streams in parallel through independent encoders, followed by late fusion. Although such approaches improve robustness by aggregating complementary features, two critical limitations persist: (i) The incorporation of RGB images not only introduces additional data acquisition costs but also compromises the intrinsic privacy-preserving nature of event cameras by capturing detailed individual visuals. (ii) Independent encoders disregarding inter-modal correlations introduce redundancy that critically undermines the acquisition of discriminative action features." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.483, + 0.702 + ], + "angle": 0, + "content": "To address the challenges above, we seek a privacy-preserving alternative to RGB data that eliminates additional data dependency and narrows the representational gap between surrogate modal with event modal. As shown in Fig. 1(b), inspired by the event-to-image [40] techniques, we employ a lightweight event reconstruction network to generate texture-enriched surrogate frames directly from event streams. This approach effectively emulates RGB semantics while filtering privacy-sensitive details. Building upon this, we introduce a unified multi-modal encoding architecture that harmonizes edge-focused event representations with texture-augmented surrogates, preserving modality-specific discriminative strengths while suppressing cross-modal redundant information interference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Formally, we present an Enhanced Multi-modal Perceptual (EMP) framework designed to hierarchically exploit intrinsic multi-modal cues from the representation level to the feature level. Specifically, our framework consists of two components: Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS). The CFE leverages the distinct frequency traits inherent in different modalities, utilizing the high-frequency information (e.g., edges) from stacked event frames to enhance the reconstructed event frames including mid-frequency information (e.g., texture). This process effectively refines the event representations for better feature extraction. In Addition, considering the discrepancy between modalities," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "the HGS adaptively captures semantic consistency tokens across different modalities, thereby realizing unified encoding of multimodal features." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.138, + 0.874, + 0.152 + ], + "angle": 0, + "content": "The main contributions are summarized as follows," + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.154, + 0.905, + 0.214 + ], + "angle": 0, + "content": "- A privacy-preserving paradigm inspired by event-to-image generates texture-enriched surrogate frames as alternatives to RGB images, eliminating auxiliary data acquisition while bridging the modality gap with events." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.215, + 0.905, + 0.273 + ], + "angle": 0, + "content": "- A novel Enhanced-Multimodal Perceptual (EMP) framework that hierarchically explores intrinsic multi-modal cues from raw event streams through dual-stage innovations at representation and feature levels." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.275, + 0.905, + 0.349 + ], + "angle": 0, + "content": "- A multi-modal unified perception strategy equipped with collaborative Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enables progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.154, + 0.905, + 0.349 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.365, + 0.655, + 0.381 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.391, + 0.901, + 0.407 + ], + "angle": 0, + "content": "2.1. Single-modal Event-based Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.414, + 0.907, + 0.807 + ], + "angle": 0, + "content": "Existing single-modal approaches for event-based action recognition (EAR) focus on extracting discriminative features from event streams [8, 52]. These methods leverage the sparse spatial nature of event data by stacking asynchronous event streams into compact frames, which are processed using powerful frame-specific architectures, e.g. convolutional neural networks (CNNs) [32, 44] or visual transformers (ViTs) [7, 57]. For instance, EventTransAct [10] introduce video transformers and event contrastive learning to extract fine-grained features from the event stream. Similarly, ExACT [57] incorporate rich semantic information to effectively handle event frames from a cross-modal perspective. Beyond frame-based methods, some works [37, 54] explore networks tailored to the unique properties of asynchronous events, such as spiking neural networks (SNNs) and graph convolutional networks (GCNs). For example, SpikePoint [37] proposed an end-to-end SNN architecture that processes sparse event streams by simultaneously extracting global and local features. VMvGCN [54] introduced a novel volumetric multi-view framework to capture key spatial and temporal information from event streams. However, these approaches primarily fail to fully exploit the complementary infomation inherent in raw event streams. In contrast, our proposed Enhanced Multi-modal Perceptual framework synergistically integrates features from different modalities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.818, + 0.898, + 0.835 + ], + "angle": 0, + "content": "2.2. Multi-modal Event-based Action Recognition" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Recent advances in robust representation for multi-modal data garner significant attention across various visual domains [7, 48]. For EAR, previous research explore two primary paradigms for multi-modal input: (i) Integrating" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5970" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.116, + 0.098, + 0.885, + 0.376 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.39, + 0.908, + 0.462 + ], + "angle": 0, + "content": "Figure 2. The overall framework of our proposed EMP. First, complementary event modalities (stacked and reconstructed frames) undergo frequency-domain refinement via the Cross-modal Frequency Enhancer (CFE), which amplifies edge dynamics while preserving textural in reconstructed counterparts. After that, during the unified visual encoding process of CLIP, the High-frequency Guided Selector (HGS) is employed to capture semantic-consistency tokens across modalities, which effectively eliminates spatiotemporally redundant features. Finally, event features and text features derived from the text/vision encoder are contrasted for activity recognition." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.487, + 0.485, + 0.835 + ], + "angle": 0, + "content": "different representations of event data (e.g., event frames, event points, and event voxels); and (ii) Fusing RGB images with event data. For instance, AGCN [19] introduce a dual point-voxel absorbing graph representation to exploit complementary information between event points and event voxels. EventCrab [6] propose a framework that synergistically processes event frames and event points balancing efficiency and effectiveness. However, while these methods fully leverage the rich temporal information within event streams, they overlook the inherent lack of texture information in events. To address this limitation, SAFE [24] leverages pre-trained large-scale vision-language models to fuse semantic labels, RGB frames, and event streams, improving multi-modal recognition accuracy. Similarly, sstformer [48] formally propose to recognize patterns by fusing RGB frames and event streams simultaneously. Despite these advances, existing methods require additional resources for RGB images and fail to fully exploit the consistency between multi-modal data. In contrast, our proposed framework eliminates dependency on RGB while explicitly addressing the impact of irrelevant information on multimodal feature selection. Our approach provides a promising solution for practical deployment in real-world scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.371, + 0.863 + ], + "angle": 0, + "content": "2.3. Token Selection in Transformer" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.903 + ], + "angle": 0, + "content": "With the widespread adoption of Transformer across various domains [6, 20, 21, 34, 35], token selection has gar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.908, + 0.79 + ], + "angle": 0, + "content": "nered increasing attention for its ability to leverage inherent redundancies between image tokens to focus on key objects and reduce computational costs [12, 36, 47, 55]. This is particularly crucial in visual tasks [36], where the extraction of fine-grained features is essential for visual understanding. DynamicViT [18, 36] inserts prediction modules into Transformer blocks to predict and select tokens with higher information content. TransFG [18] directly employs attention mechanisms to locate important tokens, demonstrating strong performance in fine-grained classification. Building on this, STA [12] proposes a framework that simultaneously considers temporal redundancy and semantic importance for selecting spatiotemporal tokens. Magic [55] extends the token selection to multi-modal tasks, enabling the model to dynamically choose object-centric tokens based on different input modalities. Unlike these methods, we propose a more flexible multi-modal token selection scheme that utilizes high-frequency modal-guided strategies to capture semantic consistency tokens, thereby enhancing the robustness of our approach." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.806, + 0.605, + 0.822 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.833, + 0.691, + 0.848 + ], + "angle": 0, + "content": "3.1. Joint Formulation" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.903 + ], + "angle": 0, + "content": "We introduce a jointly formulated framework for event-to-image reconstruction and CLIP-driven action recognition, without the need for additionally captured RGB images." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5971" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.227 + ], + "angle": 0, + "content": "Event-to-Image Reconstruction. Considering the event stream \\( \\pmb{E} = \\{(u_i, p_i, t_i)\\}_{i \\in (0, N)} \\), where \\( u_i \\) denotes the pixel position, \\( p_i \\) indicates the polarity, \\( t_i \\) denotes the time of each event point, and \\( N \\) is the number of event points. We first generate two complementary event representations from raw asynchronous event streams: (i) Edge-dominant stacked frames \\( I_s \\) obtained by temporally aggregating events; (ii) Texture-enriched reconstructed frames \\( I_r \\) obtained by the image reconstruction network \\( H \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.245, + 0.242, + 0.483, + 0.259 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {\\mathrm {r}} = \\mathrm {H} (\\boldsymbol {E}), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.27, + 0.284, + 0.286 + ], + "angle": 0, + "content": "where \\(\\pmb{I}_{\\mathrm{s}},\\pmb{I}_{\\mathrm{r}}\\in \\mathbb{R}^{T\\times H\\times W\\times 3}\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.287, + 0.484, + 0.545 + ], + "angle": 0, + "content": "CLIP-driven Action Recognition. In terms of \\(C\\) action classes, we follow [57] to adopt two different text prompts: hand-crafted prompt \"A series of photos recording action for [class]\" and learnable prompt \"[V1], [V2], ..., [Vn], [class]\", where [class] represents the class name and \\([V_n]\\) is the random initialized parameter. The CLIP-based text encoder [56] first encodes the two prompts to obtain D-dimensional text features \\(\\pmb{f}_1^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}\\) and \\(\\pmb{f}_2^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}\\), and the final text feature \\(\\pmb{f}^{t}\\in \\mathbb{R}^{C\\times D}\\) is obtained by averaging \\(\\pmb{f}_1^{\\mathrm{t}}\\) and \\(\\pmb{f}_2^{\\mathrm{t}}\\). Meanwhile, we use the CLIP-based vision encoder [56] for event feature extraction, fully exploiting the potential of Visual Language Models (VLMs) in event-based recognition. The stacked frames \\(\\pmb{I}_{\\mathrm{s}}\\) and reconstructed frames \\(\\pmb{I}_{\\mathrm{r}}\\) after being enhanced for event representation, are fed into the vision encoder to obtain the event feature \\(\\pmb{f}^{\\mathrm{e}}\\in \\mathbb{R}^{1\\times D}\\). Eventually, the class probability \\(p\\in \\mathbb{R}^{1\\times C}\\) for the action recognition are obtained as," + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.557, + 0.483, + 0.583 + ], + "angle": 0, + "content": "\\[\np = \\operatorname {s o f t m a x} _ {(1)} \\left(\\boldsymbol {f} ^ {\\mathrm {e}} \\left(\\boldsymbol {f} ^ {\\mathrm {t}}\\right) ^ {\\top}\\right), \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.207, + 0.58, + 0.317, + 0.594 + ], + "angle": 0, + "content": "\\[\nc = \\operatorname {a r g m a x} (p),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.608, + 0.484, + 0.655 + ], + "angle": 0, + "content": "where \\(\\mathrm{softmax}(\\cdot)\\) denotes the softmax function. By taking the category of maximum probability, we obtain the final action prediction class \\(c\\) for the event stream." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.666, + 0.445, + 0.683 + ], + "angle": 0, + "content": "3.2. Enhanced-Multimodal Perceptual (EMP)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.824 + ], + "angle": 0, + "content": "To preserve diverse information within and across modalities while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. Fig. 2 presents the architecture of EMP. It is equipped with Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enabling progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Cross-modal Frequency Enhancer (CFE). Event frames from distinct modalities demonstrate complementary spectral characteristics [55]. Specifically, stacked frames from asynchronous event streams predominantly focus on capturing high-frequency edge dynamics of transient actions," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "whereas reconstructed frames preserve mid-frequency texture inherent to activity semantics. Both modalities are crucial for event action recognition, yet effectively harnessing their complementary strengths remains a critical challenge. The direct combination of these heterogeneous modalities yields suboptimal solutions because of their inherent discrepancies. Thus, we propose a Cross-modal Frequency Enhancer (CFE) that strategically leverages frequency traits across modalities to refine event representation. It enables the model to focus on more comprehensive features through frequency-domain enhancement, thereby optimizing the synergy between modality-specific information." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Whitin the CFE, we first transform stacked frames \\( I_{\\mathrm{s}} \\) and reconstructed frames \\( I_{\\mathrm{r}} \\) into the frequency domain via 3D Fast Fourier Transform (FFT) [29] as:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.327, + 0.905, + 0.354 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}), \\\\ \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}) \\end{array} \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.655, + 0.35, + 0.766, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {r}}).\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.375, + 0.907, + 0.542 + ], + "angle": 0, + "content": "The spectral conversion enables the explicit decomposition of modality-specific frequency components, where edge-dominant stacked frames predominantly occupy high-frequency bands while texture-abundant reconstructed frames exhibit concentrated energy in mid-frequency regions. The frequency-aware representation establishes a unified foundation for cross-modal information enhancement. Subsequently, to selectively amplify edge-dynamic information in stacked frames from the frequency spectrum, we apply a sigmoid layer to \\( I_{\\mathrm{s}}^{\\mathrm{ftt}} \\), generating a frequency filter A. The proposed filtering is calculated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.652, + 0.55, + 0.905, + 0.578 + ], + "angle": 0, + "content": "\\[\n\\mathrm {A} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f t t}}\\right), \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.638, + 0.572, + 0.749, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t} ^ {\\prime}} = \\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} \\odot \\mathrm {A},\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.602, + 0.907, + 0.665 + ], + "angle": 0, + "content": "where \\(\\odot\\) denotes element-wise multiplication. Finally, we transfer the \\(I_{\\mathrm{r}}^{\\mathrm{fft}'}\\) in the frequency space into the spatial domain to acquire enhanced reconstructed frames \\(I_{\\mathrm{r}}'\\) using an inverse Fast Fourier Transform (iFFT) [29] as:" + }, + { + "type": "equation", + "bbox": [ + 0.648, + 0.675, + 0.905, + 0.697 + ], + "angle": 0, + "content": "\\[\n\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\prime} = \\mathrm {i F F T} \\left(\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\text {f f t}} ^ {\\prime}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.902 + ], + "angle": 0, + "content": "High-frequency Guided Selector (HGS). Given the inherent semantic coherence across modalities derived from the same event stream, existing methods [48] that sequentially extract and aggregate distinct modal features may introduce redundant feature interference. Building on cross-modal frequency enhancement, we further propose a high-frequency guided selection strategy to adaptively prioritize inter-modal critical regions that jointly characterize event semantics. Specifically, both stacked frames and enhanced reconstructed frames are processed through the shared vision encoder for visual feature learning. During event feature encoding, the [cls] token inherently aggregates global semantic representations through self-attention" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5972" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.166 + ], + "angle": 0, + "content": "mechanisms [17]. We leverage high-frequency [cls] token from stacked frames to steer the attention mechanisms of reconstructed frames, enabling selective focus on cross-modal consistent semantic regions while suppressing modality-specific interference." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.483, + 0.333 + ], + "angle": 0, + "content": "For stacked frames \\( I_{\\mathrm{s}} \\) and enhanced reconstructed frames \\( I_{\\mathrm{r}}^{\\prime} \\), we first obtain the corresponding non-overlapping patches \\( f_{\\mathrm{s}}, f_{\\mathrm{r}} \\in \\mathbb{R}^{T \\times (N + 1) \\times D} \\) including the special [cls] token \\( f_{\\mathrm{s}}^{\\mathrm{cls}}, f_{\\mathrm{r}}^{\\mathrm{cls}} \\in \\mathbb{R}^{1 \\times D} \\) [2] with D-dimension through a linear projection, where \\( N = H \\times W / P^2 \\) and \\( (P, P) \\) is the size of the patch. Between \\( l \\)-th and \\( (l + 1) \\)-th layer of the vision encoder, we deploy a cross-attention mechanism to facilitate inter-modal interaction. Here, the reconstructed tokens serve as query feature \\( Q \\), while the stacked tokens are transformed into key and value features, i.e. \\( K \\) and \\( V \\). The detailed procedure can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.109, + 0.34, + 0.482, + 0.397 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} Q = W _ {\\mathrm {q}} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}, K = W _ {\\mathrm {k}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, V = W _ {\\mathrm {v}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, \\\\ \\boldsymbol {f} _ {\\mathrm {s}} ^ {l} = \\operatorname {C r o s s A t t e n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V, \\tag {6} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.404, + 0.483, + 0.435 + ], + "angle": 0, + "content": "where \\(W_{\\mathrm{q}}, W_{\\mathrm{k}}, W_{\\mathrm{v}} \\in \\mathbb{R}^{D \\times D}\\) are the weights of the linear transformations corresponding to \\(Q, K, V\\), respectively." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.436, + 0.484, + 0.511 + ], + "angle": 0, + "content": "Subsequently, we extract the [cls] token \\( f_{\\mathrm{s}}^{\\mathrm{cls}} \\) that encapsulates global semantics from the stacked features \\( f_{\\mathrm{s}}^{l} \\) and compute its semantic affinity with the patch tokens in the augmented reconstructed features \\( f_{\\mathrm{r}}^{l} \\) via the cosine similarity. It can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.187, + 0.519, + 0.483, + 0.557 + ], + "angle": 0, + "content": "\\[\n\\sin \\left(\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}}, \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}\\right) = \\frac {\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} {} ^ {\\top} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}}{\\| \\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} \\| \\| \\boldsymbol {f} _ {\\mathrm {r}} ^ {l} \\|}. \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.565, + 0.484, + 0.669 + ], + "angle": 0, + "content": "By averaging the aforementioned similarity matrix \\(\\mathrm{sim}(\\pmb{f}_{\\mathrm{s}}^{\\mathrm{cls}},\\bar{\\pmb{f}}_{\\mathrm{s}}^{l})\\) along the temporal dimension \\(T\\), the final attention score matrix \\(B\\in \\mathbb{R}^{T\\times N}\\) is obtained. To maintain semantic consistency between the two modalities, we use \\(B\\) to incorporate dynamic edge features, which guide the reconstruction branch in adaptively selecting tokens. The selection is formalized as:" + }, + { + "type": "equation", + "bbox": [ + 0.225, + 0.679, + 0.483, + 0.697 + ], + "angle": 0, + "content": "\\[\nS = \\left\\{\\operatorname {T o p} _ {K} (B) \\right\\}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.483, + 0.795 + ], + "angle": 0, + "content": "where \\(\\mathrm{Top}_K\\) is a function that sorts scores in a set in descending order and outputs the indices corresponding to the \\(K\\) largest values, \\(K\\) is obtained by \\(\\mu \\cdot N\\) and \\(\\mu\\) is a selection parameter. Using index set \\(S\\), we select the corresponding critical tokens from \\(\\pmb{f}_{\\mathrm{r}}^{l}\\) and merge them with the \\(\\pmb{f}_{\\mathrm{r}}^{\\mathrm{cls}}\\) token before inputting them into the \\((l + 1)\\)-th visual encoder." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.803, + 0.281, + 0.82 + ], + "angle": 0, + "content": "3.3. Training Objectives" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Modality Consistency Loss. Our model takes the edge-stacked frames and the texture-reconstructed frames as input, which are crucial for capturing the discriminative features of the event stream. The HGS we proposed above focuses on critical token selection guided by high-frequency" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.183 + ], + "angle": 0, + "content": "modality, based on the principle that different modalities maintain semantic consistency. Therefore, we introduce a loss to reduce the discrepancy between the two features of the same event in the feature space. The semantic consistency cross modalities \\((f_{\\mathrm{s}} / f_{\\mathrm{r}})\\) are constrained by the Mean Square Error (MSE) loss:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.195, + 0.906, + 0.238 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\boldsymbol {f} _ {\\mathrm {s}} ^ {i} - \\boldsymbol {f} _ {\\mathrm {r}} ^ {i} \\right| \\right| _ {2} ^ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.248, + 0.741, + 0.262 + ], + "angle": 0, + "content": "where \\(N\\) is the number of tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.265, + 0.906, + 0.308 + ], + "angle": 0, + "content": "Contractive Recognition Loss. The event-text consistency \\(\\mathcal{L}_{\\mathrm{r}}\\) is constrained by the contrastive loss between the event feature and the text feature as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.319, + 0.906, + 0.358 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {i} ^ {\\mathrm {t}} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {j} ^ {\\mathrm {t}} / \\tau\\right)}, \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.37, + 0.769, + 0.384 + ], + "angle": 0, + "content": "where \\(\\tau\\) is the temperature coefficient." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.385, + 0.906, + 0.416 + ], + "angle": 0, + "content": "The final loss is composed of the modality consistency loss \\(\\mathcal{L}_{\\mathrm{c}}\\) and the contractive recognition loss \\(\\mathcal{L}_{\\mathrm{r}}\\) as" + }, + { + "type": "equation", + "bbox": [ + 0.656, + 0.43, + 0.905, + 0.446 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda \\mathcal {L} _ {\\mathrm {r}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.906, + 0.487 + ], + "angle": 0, + "content": "where \\(\\lambda\\) denotes the parameter used to balance the loss optimization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.505, + 0.646, + 0.522 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.531, + 0.715, + 0.547 + ], + "angle": 0, + "content": "4.1. Datasets and Settings" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.907, + 0.764 + ], + "angle": 0, + "content": "Datasets. To evaluate the performance of our approach, we conduct research on three datasets, including PAF [30], SeAct [57], HARDVS [50] and DVS128 Gesture [1]. DVS Action, also known as PAF, contains 10 action categories with 450 recordings, conducted in an unoccupied office environment. SeAct is the first semantic-abundant dataset for event-text action recognition, containing 58 actions under four themes. HARDVS is currently the largest dataset for event-based action recognition, currently having 107,646 recordings for 300 action categories. Both of the above three datasets have an average time duration of 5 seconds with \\(346 \\times 260\\) resolution. DVS128 Gesture is collected using a DAVIS128 camera with \\(128 \\times 128\\) resolution, dividing into 11 distinct classes of gestures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Settings. The implementation of the overall framework is carried out on PyTorch in a Linux environment with two NVIDIA GeForce 4090 GPUs. We use the Adam [23] optimizer with the initial learning rate of \\(1 \\times 10^{-5}\\) and weight decay of \\(2 \\times 10^{-4}\\). CosineAnnealingLR learning rate schedule is employed with a minimum learning rate of \\(1 \\times 10^{-6}\\). The representative lightweight FireNet [40] (only 38k parameters) for event-to-image reconstruction, which has low latency (1ms for \\(240 \\times 180\\))." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5973" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.145, + 0.089, + 0.858, + 0.311 + ], + "angle": 0, + "content": "
MethodEvent RepresentationTop-1 Accuracy (%)
Frame-basedPoint-basedPAFSeActDVS128 Gesture
Motion SNN [28]-✓78.10-92.70
Slayer [41]-✓--93.64
MST [51]✓-88.21--
EV-ACT [14]✓-92.60--
EventTransAct [10]✓--57.8197.92
EvT [39]✓--61.3096.20
SpikePoint [37]-✓90.60-98.74
SpikMamba [8]✓-96.2871.0299.01
ExACT [57]✓-94.8367.2498.86
EventCrab [7]✓✓96.4972.4198.80
EventCLIP [56] + Recon✓-96.5571.0596.59
Ours✓-99.80(+3.31)75.00(+2.59)98.86(-0.15)
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.32, + 0.907, + 0.351 + ], + "angle": 0, + "content": "Table 1. Comparative performance for EAR on the PAF, SeAct and DVS128 Gesture datasets. The best results are in bold and the second-best ones are in underlined." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.366, + 0.383, + 0.383 + ], + "angle": 0, + "content": "4.2. Comparison with SOTA Methods" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.396, + 0.485, + 0.789 + ], + "angle": 0, + "content": "Comparison on Datasets without Paired RGB. As shown in Tab. 1, we compare the proposed method EMP with several representative event-based action recognition methods on three event-only datasets (PAF, SeAct, DVS128 Gesture), where paired RGB data is unavailable. All compared methods exclusively utilize event-driven representations, categorized into event frame-based and event point-based paradigms. As demonstrated in Tab. 1, point-based methods exhibit marginally inferior performance compared to frame-based approaches on the PAF and DVS128 Gesture datasets. This discrepancy stems from the latter's enhanced capability in extracting discriminative spatiotemporal features through visual architectural designs. EventCrab [7] achieves impressive results by collaboratively mining complementary information from both event points and frames, thereby leveraging multi-modal event representations. However, unlike the paradigm of primarily exploring the sparse spatial characteristics of event streams from EventCrab, our method introduces texture-rich reconstructed frames and holistically exploits intrinsic correlations and extrinsic specificities between event frames and reconstructed frames. This unified perceptual framework yields significantly improved recognition efficacy, attaining \\(3.31\\%\\) (Top-1) and \\(2.59\\%\\) (Top-1) accuracy gains on the PAF and SeAct datasets respectively, while maintaining competitive performance on the DVS128 Gesture dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Comparison on Datasets with Paired RGB. As shown in Tab. 2, we compare the proposed EMP with several representative EAR methods on the HARDVS dataset, which includes 300-class paired RGB-Event data. Comparative methods are categorized into two types: (i) those utilizing RGB modality and (ii) those solely relying on event streams. We can find that combining RGB frames with" + }, + { + "type": "table", + "bbox": [ + 0.519, + 0.364, + 0.906, + 0.535 + ], + "angle": 0, + "content": "
MethodRGBParam.Top-1 Accuracy (%)
R2Plus1D [46]X63.5M49.06
C3D [45]X147.2M49.94
TimeSformer [4]X121.2M50.77
ESTF [50]X46.1M51.22
ExACT [57]X155.1M90.10
ESTF* [50]✓76.8M49.93
SAFE [24]✓-50.17
C3D* [45]✓245.3M50.88
TimeSformer* [4]✓202.3M51.57
SSTFormer [48]✓336.4M52.97
OursX182.1M97.68
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.545, + 0.907, + 0.587 + ], + "angle": 0, + "content": "Table 2. Comparative performance (\\%) on the HARDVS datasets. The best results are in **bold. (·)* denotes the combination between RGB frames and event streams." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.907, + 0.81 + ], + "angle": 0, + "content": "event streams yields marginal improvements over the single modality-based version. For instance, C3D [45] achieves an increase from \\(49.94\\%\\) to \\(50.88\\%\\) (Top-1), while TimeSformer improves from \\(50.77\\%\\) to \\(51.57\\%\\) (Top-1). However, the significant semantic discrepancy between RGB and event modalities limits performance gains from superficial cross-modal fusion. Different from conventional fusion paradigms, EMP adopts the text-aligned adaptation strategy to bridge the modality gap, while leveraging frequency-domain characteristics to extract discriminative event features. The results in Tab. 2 validate the efficacy of our approach, demonstrating superior robustness in handling heterogeneous modality." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.68, + 0.833 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of Each Component. As detailed in Tab. 3, we conduct an ablation study to evaluate the contributions of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS) within the pro" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5974" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.116, + 0.089, + 0.462, + 0.168 + ], + "angle": 0, + "content": "
CFEHGSSeAct (%)DVS128 Gesture (%)
XX71.2597.34
✓X72.4197.72
X✓73.2798.10
✓✓75.0098.86
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.179, + 0.483, + 0.207 + ], + "angle": 0, + "content": "Table 3. Ablation study for effectiveness of each component in the proposed method EMP." + }, + { + "type": "table", + "bbox": [ + 0.145, + 0.222, + 0.435, + 0.315 + ], + "angle": 0, + "content": "
Selection ratioTop-1 (%)Top-5 (%)
0.270.6890.51
0.471.5592.24
0.675.0095.68
0.873.2793.96
1.072.4193.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.126, + 0.326, + 0.447, + 0.339 + ], + "angle": 0, + "content": "Table 4. Different selection ratio on the SeAct dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.368, + 0.484, + 0.581 + ], + "angle": 0, + "content": "posed EMP. The baseline model trained only on eventreconstructed frames without incorporating the CFE and HGS achieves \\(71.55\\%\\) accuracy on SeAct and \\(97.34\\%\\) on DVS128 Gesture. When adding the baseline with CFE, we can find the performance improvements of \\(0.86\\%\\) on SeAct and \\(0.38\\%\\) on DVS128 Gesture, respectively. These performance improvements demonstrate the effectiveness of cross-modal representation enhancement in the frequency domain. Furthermore, the introduction of HGS achieves dynamic token selection across modalities, resulting in a \\(1.72\\%\\) improvement on SeAct and a \\(0.76\\%\\) improvement on DVS128 Gesture. By integrating all components, our model achieves the optimal performance. These results validate the effectiveness of our EMP in different scenarios." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.606, + 0.258, + 0.621 + ], + "angle": 0, + "content": "4.4. Diagnose Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Impact of Different Token Selection Ratio. As illustrated in Tab. 4, we investigate the impact of varying selection thresholds \\(\\mu\\) on the SeAct dataset, where \\(\\mu\\) governs the proportion of semantically consistent tokens selected across modalities in the HGS framework. The experimental results reveal that the recognition accuracy improves as \\(\\mu\\) increases, reaching its peak at \\(\\mu = 0.6\\), beyond which further threshold escalation correlates with performance degradation. This trend underscores the inherent divergence between modalities, suggesting that indiscriminate cross-modal fusion is suboptimal. Critically, calibrating \\(\\mu\\) to an appropriate level optimally harnesses intermodal complementarity while suppressing shared redundancies." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Impact of Different Selection Layer. As shown in Tab. 5, we investigate the impact of integrating token selection at distinct layers of the vision encoder on the SeAct dataset, with the optimal selection ratio fixed at \\(\\mu = 0.6\\) (Tab. 4). We implement modality-guided token selection across indi" + }, + { + "type": "table", + "bbox": [ + 0.568, + 0.089, + 0.858, + 0.166 + ], + "angle": 0, + "content": "
Insertion LayerTop-1 (%)Top-5 (%)
[4]73.2793.96
[6]73.9692.41
[8]75.0095.68
[10]74.1393.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.177, + 0.905, + 0.204 + ], + "angle": 0, + "content": "Table 5. Different insertion layer of token selection strategy on the SeAct dataset." + }, + { + "type": "image", + "bbox": [ + 0.588, + 0.231, + 0.825, + 0.389 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.405, + 0.907, + 0.433 + ], + "angle": 0, + "content": "Figure 3. Effect of different values of \\(\\lambda\\) on SeAct and DVS 128 Gesture datasets." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.46, + 0.907, + 0.55 + ], + "angle": 0, + "content": "vidual layers and observe that shallow and deep layers yield marginal performance gains. In contrast, layer 8 achieves peak Top-1, validating that appropriately positioned layers within vision encoders optimally encode cross-modal feature representations, thereby enabling effective modality-guided selection of semantically consistent tokens." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Balance Between Consistency Loss and Recognition Loss. Our empirical validation substantiates that achieving equilibrium between modality consistency loss and contractive recognition (Eq. 11) loss constitutes a pivotal design criterion for robust multimodal learning. The contribution of \\(\\mathcal{L}_{\\mathrm{r}}\\) is modified via \\(\\lambda\\). As shown in Fig. 3, the variation of \\(\\lambda\\) affects the accuracy performance. The best performance is achieved with weights of 0.6, 0.4 on SeAct and DVS128 Gesture datasets respectively. It is worth noting that our findings highlight the potential benefits of leveraging consistency across modalities to enhance the representational capacity and recognition robustness of our EMP." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.707, + 0.759 + ], + "angle": 0, + "content": "4.5. Qualitative Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Visualization of Cross-modal Frequency Enhancer (CFE). To investigate the efficacy of the designed CFE, we visualize the four variants of event frames on the SeAct dataset: raw stacked frames, reconstructed frames, enhanced frames by sharpening operators, and enhanced frames by our CFE. As shown in fig. 4, conventional edge-enhancement techniques [33] yield marginal improvements over raw reconstructed frames. In contrast, event frames processed by CFE consistently highlight the critical regions" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5975" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.129, + 0.092, + 0.468, + 0.363 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.374, + 0.483, + 0.445 + ], + "angle": 0, + "content": "Figure 4. Visualization of event frames. \"Stack\" and \"Reconstruct\" denote stacked and reconstructed frames, respectively; \"Edge Enhance\" indicates enhanced reconstructed frames employing sharpening operators; \"CFE\" represents enhanced reconstructed frames by our Cross-modal Frequency Enhancer." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.483, + 0.581 + ], + "angle": 0, + "content": "across both individual actions and interaction scenarios. Our qualitative analysis confirms that the proposed CFE selectively refines the high-frequency components of the reconstructed frames through spectral coordination with the edge dynamics of the stacked frames, thereby generating complementary event representations with synergistic spatial characteristics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.582, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Visualization of Guidanced Selection. Fig. 5 visualizes the important tokens selected by the High-frequency Guided Selector (HGS). As shown in Fig. 5, the HGS demonstrates the adaptive selection of action-salient tokens across sequential event frames. This demonstrates the intrinsic consistency among multi-modal frames originating from the identical event stream, where HGS effectively identifies semantically coherent tokens while suppressing interference from spatiotemporal redundancies. The results highlight the ability of HGS to focus on critical regions of interest, thereby enhancing the discriminative power of EMP. This selection is crucial for improving the robustness and accuracy of action recognition in dynamic scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Visualization of Event Features Learned by EMP. As illustrated in Fig. 6, we visualize the feature distributions of our proposed EMP against the baseline on the SeAct dataset. Seven action classes are randomly selected to demonstrate inter-class feature separability. Empirical results demonstrate the capability of our EMP in achieving both intra-class compactness and inter-class dispersion, further substantiating its efficacy in preserving semantic dis" + }, + { + "type": "image", + "bbox": [ + 0.543, + 0.091, + 0.875, + 0.282 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.292, + 0.907, + 0.322 + ], + "angle": 0, + "content": "Figure 5. Visualization of cross-modal consistent tokens selected by the HGS, with unselected regions masked in white." + }, + { + "type": "image", + "bbox": [ + 0.541, + 0.344, + 0.885, + 0.469 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.481, + 0.905, + 0.51 + ], + "angle": 0, + "content": "Figure 6. The t-SNE visualization of 7 randomly selected actions. Different color indicates different actions." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.537, + 0.694, + 0.55 + ], + "angle": 0, + "content": "tinctiveness across classes." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.566, + 0.763, + 0.581 + ], + "angle": 0, + "content": "5. Conclusion and Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.592, + 0.907, + 0.817 + ], + "angle": 0, + "content": "Conclusion. In this work, we propose a privacy-preserving paradigm that eliminates RGB dependencies by generating texture-enriched reconstructed frames directly from raw event streams, reducing multi-sensor acquisition costs while bridging the modality gap with event frames. To preserve diverse information across reconstructed and stacked frames while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. It consists of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), which enables the progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level to acquire discriminative event features. Extensive experimental results validate the effectiveness of the proposed method." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.817, + 0.906, + 0.9 + ], + "angle": 0, + "content": "Discussion. While our method achieves promising performance in EAR and demonstrates the potential for scalability, its ability to handle more challenging scenarios remains to be explored. Future work will focus on developing a robust multi-modal unified perceptual approach to cope with diverse and variable scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5976" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.093, + 0.486, + 0.162 + ], + "angle": 0, + "content": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208), and the Nanjing University Integrated Research Platform of the Ministry of Education-Top Talents Program (Grant No. 2024300482)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.175, + 0.188, + 0.19 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.284 + ], + "angle": 0, + "content": "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7243-7252, 2017. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.285, + 0.484, + 0.352 + ], + "angle": 0, + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6836-6846, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.354, + 0.484, + 0.424 + ], + "angle": 0, + "content": "[3] Sami Barchid, Benjamin Allaert, Amel Aissaoui, José Mennesson, and Chaabane C Djeraba. Spiking-fer: spiking neural network for facial expression recognition with event cameras. In Proceedings of the 20th International Conference on Content-based Multimedia Indexing, pages 1–7, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.425, + 0.484, + 0.478 + ], + "angle": 0, + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning, page 4, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.484, + 0.535 + ], + "angle": 0, + "content": "[5] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9286-9296, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.484, + 0.606 + ], + "angle": 0, + "content": "[6] Meiqi Cao, Rui Yan, Xiangbo Shu, Jiachao Zhang, Jinpeng Wang, and Guo-Sen Xie. Mup: multi-granularity unified perception for panoramic activity recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7666-7675, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.607, + 0.484, + 0.662 + ], + "angle": 0, + "content": "[7] Meiqi Cao, Xiangbo Shu, Jiachao Zhang, Rui Yan, Zechao Li, and Jinhui Tang. Eventcrab: Harnessing frame and point synergy for event-based action recognition and beyond. arXiv preprint arXiv:2411.18328, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.663, + 0.484, + 0.732 + ], + "angle": 0, + "content": "[8] Jiaqi Chen, Yan Yang, Shizhuo Deng, Da Teng, and Liyuan Pan. Spikmamba: When snn meets mamba in event-based human action recognition. In Proceedings of the 6th ACM International Conference on Multimedia in Asia, pages 1-8, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.484, + 0.802 + ], + "angle": 0, + "content": "[9] Lan Chen, Haoxiang Yang, Pengpeng Shao, Haoyu Song, Xiao Wang, Zhicheng Zhao, Yaowei Wang, and Yonghong Tian. Velora: A low-rank adaptation approach for efficient rgb-event based recognition. arXiv preprint arXiv:2412.20064, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.484, + 0.873 + ], + "angle": 0, + "content": "[10] Tristan de Blegiers, Ishan Rajendrakumar Dave, Adeel Yousaf, and Mubarak Shah. Eventtransact: A video transformer-based framework for event-camera based action recognition. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1-7, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.874, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[11] Yongjian Deng, Hao Chen, and Youfu Li. A dynamic gcn with cross-representation distillation for event-based learn" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.2, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.907, + 0.121 + ], + "angle": 0, + "content": "ing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1492-1500, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.907, + 0.19 + ], + "angle": 0, + "content": "[12] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.192, + 0.906, + 0.262 + ], + "angle": 0, + "content": "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.906, + 0.319 + ], + "angle": 0, + "content": "[14] Yue Gao, Jiaxuan Lu, Siqi Li, Nan Ma, Shaoyi Du, Yipeng Li, and Qionghai Dai. Action recognition and benchmark using event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14081-14097, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.321, + 0.906, + 0.389 + ], + "angle": 0, + "content": "[15] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.906, + 0.447 + ], + "angle": 0, + "content": "[16] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.906, + 0.503 + ], + "angle": 0, + "content": "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.573 + ], + "angle": 0, + "content": "[18] Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, and Changhu Wang. Transfg: A transformer architecture for fine-grained recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 852-860, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.575, + 0.906, + 0.631 + ], + "angle": 0, + "content": "[19] Bo Jiang, Chengguo Yuan, Xiao Wang, Zhimin Bao, Lin Zhu, Yonghong Tian, and Jin Tang. Point-voxel absorbing graph representation learning for event stream based recognition. arXiv preprint arXiv:2306.05239, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.632, + 0.907, + 0.7 + ], + "angle": 0, + "content": "[20] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2570-2578, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.906, + 0.758 + ], + "angle": 0, + "content": "[21] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.76, + 0.906, + 0.828 + ], + "angle": 0, + "content": "[22] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.906, + 0.87 + ], + "angle": 0, + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.872, + 0.906, + 0.901 + ], + "angle": 0, + "content": "[24] Dong Li, Jiandong Jin, Yuhao Zhang, Yanlin Zhong, Yaoyang Wu, Lan Chen, Xiao Wang, and Bin Luo." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5977" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.483, + 0.134 + ], + "angle": 0, + "content": "Semantic-aware frame-event fusion based pattern recognition via large vision-language models. arXiv preprint arXiv:2311.18592, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[26] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.482, + 0.317 + ], + "angle": 0, + "content": "[27] Haoyue Liu, Shihan Peng, Lin Zhu, Yi Chang, Hanyu Zhou, and Luxin Yan. Seeing motion at nighttime with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25648-25658, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[28] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In International Joint Conference on Artificial Intelligence, pages 1743-1749, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.418 + ], + "angle": 0, + "content": "[29] Shuangbiao Liu, Qian Wang, and Geng Liu. A versatile method of discrete convolution and fft (dc-fft) for contact analyses. Wear, 243(1-2):101-111, 2000. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.421, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[30] Shu Miao, Guang Chen, Xiangyu Ning, Yang Zi, Kejia Ren, Zhenshan Bing, and Alois Knoll. Neuromorphic vision datasets for pedestrian detection, action recognition, and fall detection. Frontiers in Neurorobotics, 13:38, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.546 + ], + "angle": 0, + "content": "[31] Federico Paredes-Valles and Guido CHE De Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.483, + 0.63 + ], + "angle": 0, + "content": "[32] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 19935-19947, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[33] R Pushpavalli and G Sivaradje. Switching median filter for image enhancement. International Journal of Scientific & Engineering Research, 3(2):1-5, 2012. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[34] Hongyu Qu, Rui Yan, Xiangbo Shu, Hailiang Gao, Peng Huang, and Guo-Sen Xie. Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition. arXiv preprint arXiv:2405.02077, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[35] Hongyu Qu, Jianan Wei, Xiangbo Shu, and Wenguan Wang. Learning clustering-based prototypes for compositional zero-shot learning. In Proceedings of the International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[36] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in Neural Information Processing Systems, 34:13937-13949, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.86, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[37] Hongwei Ren, Yue Zhou, Yulong Huang, Haotian Fu, Xiaopeng Lin, Jie Song, and Bojun Cheng. Spikepoint: An efficient point-based spiking neural network for event cam" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "eras action recognition. arXiv preprint arXiv:2310.07189, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[38] Hongwei Ren, Yue Zhou, Jiadong Zhu, Xiaopeng Lin, Hao-tian Fu, Yulong Huang, Yuetong Fang, Fei Ma, Hao Yu, and Bojun Cheng. Rethinking efficient and effective point-based networks for event camera classification and regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[39] Alberto Sabater, Luis Montesano, and Ana C Murillo. Event transformer. a sparse-aware solution for efficient event data processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2677-2686, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.907, + 0.347 + ], + "angle": 0, + "content": "[40] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 1, 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[41] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 31, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[42] Xiangbo Shu, Binqian Xu, Liyan Zhang, and Jinhui Tang. Multi-granularity anchor-contrastive representation learning for semi-supervised skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7559-7576, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.904, + 0.518 + ], + "angle": 0, + "content": "[43] Xiangbo Shu, Jiawen Yang, Rui Yan, and Yan Song. Expansion-squeeze-excitation fusion network for elderly activity recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5281-5292, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[44] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3): 3200-3225, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.657 + ], + "angle": 0, + "content": "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4489-4497, 2015. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.661, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[47] Hongjie Wang, Bhishma Dedhia, and Niraj K Jha. Zero-prune: Zero-shot token pruning through leveraging of the attention graph in pre-trained transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16070-16079, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.803, + 0.905, + 0.87 + ], + "angle": 0, + "content": "[48] Xiao Wang, Zongzhen Wu, Yao Rong, Lin Zhu, Bo Jiang, Jin Tang, and Yonghong Tian. Sstformer: Bridging spiking neural network and memory support transformer for frame-event based recognition. arXiv preprint arXiv:2308.04369, 2023. 2, 3, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Xiao Wang, Shiao Wang, Pengpeng Shao, Bo Jiang, Lin Zhu, and Yonghong Tian. Event stream based human action" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5978" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "recognition: a high-definition benchmark dataset and algorithms. arXiv preprint arXiv:2408.09764, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.19 + ], + "angle": 0, + "content": "[50] Xiao Wang, Zongzhen Wu, Bo Jiang, Zhimin Bao, Lin Zhu, Guoqi Li, Yaowei Wang, and Yonghong Tian. Hardvs: Revisiting human activity recognition with dynamic vision sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5615-5623, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.192, + 0.482, + 0.248 + ], + "angle": 0, + "content": "[51] Ziqing Wang, Yuetong Fang, Jiahang Cao, Qiang Zhang, Zhongrui Wang, and Renjing Xu. Masked spiking transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1761-1771, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.249, + 0.482, + 0.318 + ], + "angle": 0, + "content": "[52] Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. In European Conference on Computer Vision, pages 310-328, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.375 + ], + "angle": 0, + "content": "[53] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.376, + 0.482, + 0.432 + ], + "angle": 0, + "content": "[54] Bochen Xie, Yongjian Deng, Zhanpeng Shao, Hai Liu, and Youfu Li. Vmv-gcn: Volumetric multi-view based graph cnn for event stream classification. IEEE Robotics and Automation Letters, 7(2):1976-1983, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.433, + 0.483, + 0.501 + ], + "angle": 0, + "content": "[55] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17117-17126, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.503, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[56] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Eventbind: Learning a unified representation to bind them all for event-based open-world understanding. In European Conference on Computer Vision, pages 477-494, 2024. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.56, + 0.482, + 0.642 + ], + "angle": 0, + "content": "[57] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024. 2, 4, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.642 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5979" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_origin.pdf b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..6e4501ab1900cf0ba3d94614de20210a72cd19df --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/ff885410-5312-499d-bf42-f2819adc9e95_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ce76209f804f9e56473631205f1d58b3de2e26074f3c82de229bdf0080a4d4b8 +size 2359999 diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/full.md b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/full.md new file mode 100644 index 0000000000000000000000000000000000000000..7ae7b726621624699d32b5e8e818508346073ca5 --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/full.md @@ -0,0 +1,325 @@ +# Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition + +Meiqi Cao $^{1}$ , Xiangbo Shu $^{1*}$ , Xin Jiang $^{1}$ , Rui Yan $^{1,2}$ , Yazhou Yao $^{1}$ , Jinhui Tang $^{3}$ $^{1}$ Nanjing University of Science and Technology + $^{2}$ Nanjing University + $^{3}$ Nanjing Forestry University + +{cmq123,shuxb,xinjiang,ruiyan,yazhou.yao}@njust.edu.cn,tangjh@njfu.edu.cn + +# Abstract + +While event cameras excel in capturing microsecond temporal dynamics, they suffer from sparse spatial representations compared to traditional RGB data. Thus, multimodal event-based action recognition approaches aim to synergize complementary strengths by independently extracting and integrating paired RGB-Event features. However, this paradigm inevitably introduces additional data acquisition costs, while eroding the inherent privacy advantages of event-based sensing. Drawing inspiration from event-to-image reconstruction, texture-enriched visual representation directly reconstructed from asynchronous event streams is a promising solution. In response, we propose an Enhanced Multimodal Perceptual (EMP) framework that hierarchically explores multimodal cues (e.g., edges and textures) from raw event streams through two synergistic innovations spanning representation to feature levels. Specifically, we introduce Cross-Modal Frequency Enhancer (CFE) that leverages complementary frequency characteristics between reconstructed frames and stacked frames to refine event representations. Furthermore, to achieve unified feature encoding across modalities, we develop High-Frequency Guided Selector (HGS) for semantic consistency token selection guided by dynamic edge features while suppressing redundant multimodal information interference adaptively. Extensive experiments on four benchmark datasets demonstrate the superior effectiveness of our proposed framework. The code is available at https://github.com/caomq123/EMP. + +# 1. Introduction + +Traditional frame-based cameras often encounter performance limitations when dealing with rapid motion or complex illumination conditions [22, 25, 42, 43]. In contrast, + +![](images/73c0dbf6f4e9f26cec48fdfdbf116f63d7b0b172c725f2d93740f3c296eb4649.jpg) +Figure 1. (a) Previous multi-modal input approaches independently extract features from paired RGB-Event modalities, both introducing excessive RGB data dependencies and disrupting the privacy-preserving property unique to event cameras. (b) Our approach takes a lightful reconstruct network (Firenet [40]) to substitute RGB data while narrowing the modality divergence with events, thereby enabling enhanced unified multimodal perception. + +event-driven neuromorphic visual sensors, also known as event cameras [13, 15], operate by asynchronously capturing per-pixel intensity changes only when a significant light change occurs. This unique functionality endows event cameras with several prominent advantages, including high dynamic range, low power consumption, and high temporal resolution, which enable them to excel in challenging scenarios such as fast motion, extreme lighting conditions, and privacy-sensitive applications. These attributes have led to their widespread adoption in various computer vision tasks, including object detection and tracking [16, 26, 52], motion deblurring [5, 27], and image reconstruction [31, 40, 53]. Notably, the inherent sensitivity of event cameras to dynamic scenes provides a unique advantage in capturing motion-related actions. Leveraging these strengths, the task of Event-based Action Recognition (EAR) has emerged, aiming to effectively extract action features from event streams to enhance recognition performance. + +Currently, the EAR methods faces significant challenges in effectively utilizing the spatially sparse and asynchronous event streams. Traditional single-modal input approaches + +adopt two primary strategies: (i) Processing asynchronous event points through point-based encoders (GCNs [11, 54]) or spiking neural networks (SNNs [3, 8, 37]), and (ii) Applying CNN- [32, 44], Transformer- [7, 57] or Mamba- [38, 49] based visual encoders to frames stacked from asynchronous event streams. While point-based methods accommodate the asynchronous nature of events effectively, they exhibit suboptimal recognition performance. Simultaneously, frame-based methods suffer from the lack of critical texture information due to the focus on dynamic edge capture inherent in event cameras. + +In recent years, there has been a notable shift toward multi-modal input strategies [9, 24, 48, 50], leveraging diverse data sources to improve feature extraction robustness for practical applications. As depicted in Fig. 1(a), conventional multi-modal methods typically process RGB images and event streams in parallel through independent encoders, followed by late fusion. Although such approaches improve robustness by aggregating complementary features, two critical limitations persist: (i) The incorporation of RGB images not only introduces additional data acquisition costs but also compromises the intrinsic privacy-preserving nature of event cameras by capturing detailed individual visuals. (ii) Independent encoders disregarding inter-modal correlations introduce redundancy that critically undermines the acquisition of discriminative action features. + +To address the challenges above, we seek a privacy-preserving alternative to RGB data that eliminates additional data dependency and narrows the representational gap between surrogate modal with event modal. As shown in Fig. 1(b), inspired by the event-to-image [40] techniques, we employ a lightweight event reconstruction network to generate texture-enriched surrogate frames directly from event streams. This approach effectively emulates RGB semantics while filtering privacy-sensitive details. Building upon this, we introduce a unified multi-modal encoding architecture that harmonizes edge-focused event representations with texture-augmented surrogates, preserving modality-specific discriminative strengths while suppressing cross-modal redundant information interference. + +Formally, we present an Enhanced Multi-modal Perceptual (EMP) framework designed to hierarchically exploit intrinsic multi-modal cues from the representation level to the feature level. Specifically, our framework consists of two components: Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS). The CFE leverages the distinct frequency traits inherent in different modalities, utilizing the high-frequency information (e.g., edges) from stacked event frames to enhance the reconstructed event frames including mid-frequency information (e.g., texture). This process effectively refines the event representations for better feature extraction. In Addition, considering the discrepancy between modalities, + +the HGS adaptively captures semantic consistency tokens across different modalities, thereby realizing unified encoding of multimodal features. + +The main contributions are summarized as follows, + +- A privacy-preserving paradigm inspired by event-to-image generates texture-enriched surrogate frames as alternatives to RGB images, eliminating auxiliary data acquisition while bridging the modality gap with events. +- A novel Enhanced-Multimodal Perceptual (EMP) framework that hierarchically explores intrinsic multi-modal cues from raw event streams through dual-stage innovations at representation and feature levels. +- A multi-modal unified perception strategy equipped with collaborative Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enables progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level. + +# 2. Related Work + +# 2.1. Single-modal Event-based Action Recognition + +Existing single-modal approaches for event-based action recognition (EAR) focus on extracting discriminative features from event streams [8, 52]. These methods leverage the sparse spatial nature of event data by stacking asynchronous event streams into compact frames, which are processed using powerful frame-specific architectures, e.g. convolutional neural networks (CNNs) [32, 44] or visual transformers (ViTs) [7, 57]. For instance, EventTransAct [10] introduce video transformers and event contrastive learning to extract fine-grained features from the event stream. Similarly, ExACT [57] incorporate rich semantic information to effectively handle event frames from a cross-modal perspective. Beyond frame-based methods, some works [37, 54] explore networks tailored to the unique properties of asynchronous events, such as spiking neural networks (SNNs) and graph convolutional networks (GCNs). For example, SpikePoint [37] proposed an end-to-end SNN architecture that processes sparse event streams by simultaneously extracting global and local features. VMvGCN [54] introduced a novel volumetric multi-view framework to capture key spatial and temporal information from event streams. However, these approaches primarily fail to fully exploit the complementary infomation inherent in raw event streams. In contrast, our proposed Enhanced Multi-modal Perceptual framework synergistically integrates features from different modalities. + +# 2.2. Multi-modal Event-based Action Recognition + +Recent advances in robust representation for multi-modal data garner significant attention across various visual domains [7, 48]. For EAR, previous research explore two primary paradigms for multi-modal input: (i) Integrating + +![](images/64cc22384c3bf98a7248cf3e92627ebd89f8fce5cbe56a5a2eb9a92ab9ca81c9.jpg) +Figure 2. The overall framework of our proposed EMP. First, complementary event modalities (stacked and reconstructed frames) undergo frequency-domain refinement via the Cross-modal Frequency Enhancer (CFE), which amplifies edge dynamics while preserving textural in reconstructed counterparts. After that, during the unified visual encoding process of CLIP, the High-frequency Guided Selector (HGS) is employed to capture semantic-consistency tokens across modalities, which effectively eliminates spatiotemporally redundant features. Finally, event features and text features derived from the text/vision encoder are contrasted for activity recognition. + +different representations of event data (e.g., event frames, event points, and event voxels); and (ii) Fusing RGB images with event data. For instance, AGCN [19] introduce a dual point-voxel absorbing graph representation to exploit complementary information between event points and event voxels. EventCrab [6] propose a framework that synergistically processes event frames and event points balancing efficiency and effectiveness. However, while these methods fully leverage the rich temporal information within event streams, they overlook the inherent lack of texture information in events. To address this limitation, SAFE [24] leverages pre-trained large-scale vision-language models to fuse semantic labels, RGB frames, and event streams, improving multi-modal recognition accuracy. Similarly, sstformer [48] formally propose to recognize patterns by fusing RGB frames and event streams simultaneously. Despite these advances, existing methods require additional resources for RGB images and fail to fully exploit the consistency between multi-modal data. In contrast, our proposed framework eliminates dependency on RGB while explicitly addressing the impact of irrelevant information on multimodal feature selection. Our approach provides a promising solution for practical deployment in real-world scenarios. + +# 2.3. Token Selection in Transformer + +With the widespread adoption of Transformer across various domains [6, 20, 21, 34, 35], token selection has gar + +nered increasing attention for its ability to leverage inherent redundancies between image tokens to focus on key objects and reduce computational costs [12, 36, 47, 55]. This is particularly crucial in visual tasks [36], where the extraction of fine-grained features is essential for visual understanding. DynamicViT [18, 36] inserts prediction modules into Transformer blocks to predict and select tokens with higher information content. TransFG [18] directly employs attention mechanisms to locate important tokens, demonstrating strong performance in fine-grained classification. Building on this, STA [12] proposes a framework that simultaneously considers temporal redundancy and semantic importance for selecting spatiotemporal tokens. Magic [55] extends the token selection to multi-modal tasks, enabling the model to dynamically choose object-centric tokens based on different input modalities. Unlike these methods, we propose a more flexible multi-modal token selection scheme that utilizes high-frequency modal-guided strategies to capture semantic consistency tokens, thereby enhancing the robustness of our approach. + +# 3. Method + +# 3.1. Joint Formulation + +We introduce a jointly formulated framework for event-to-image reconstruction and CLIP-driven action recognition, without the need for additionally captured RGB images. + +Event-to-Image Reconstruction. Considering the event stream $\pmb{E} = \{(u_i, p_i, t_i)\}_{i \in (0, N)}$ , where $u_i$ denotes the pixel position, $p_i$ indicates the polarity, $t_i$ denotes the time of each event point, and $N$ is the number of event points. We first generate two complementary event representations from raw asynchronous event streams: (i) Edge-dominant stacked frames $I_s$ obtained by temporally aggregating events; (ii) Texture-enriched reconstructed frames $I_r$ obtained by the image reconstruction network $H$ as: + +$$ +\boldsymbol {I} _ {\mathrm {r}} = \mathrm {H} (\boldsymbol {E}), \tag {1} +$$ + +where $\pmb{I}_{\mathrm{s}},\pmb{I}_{\mathrm{r}}\in \mathbb{R}^{T\times H\times W\times 3}$ + +CLIP-driven Action Recognition. In terms of $C$ action classes, we follow [57] to adopt two different text prompts: hand-crafted prompt "A series of photos recording action for [class]" and learnable prompt "[V1], [V2], ..., [Vn], [class]", where [class] represents the class name and $[V_n]$ is the random initialized parameter. The CLIP-based text encoder [56] first encodes the two prompts to obtain D-dimensional text features $\pmb{f}_1^{\mathrm{t}}\in \mathbb{R}^{C\times D}$ and $\pmb{f}_2^{\mathrm{t}}\in \mathbb{R}^{C\times D}$ , and the final text feature $\pmb{f}^{t}\in \mathbb{R}^{C\times D}$ is obtained by averaging $\pmb{f}_1^{\mathrm{t}}$ and $\pmb{f}_2^{\mathrm{t}}$ . Meanwhile, we use the CLIP-based vision encoder [56] for event feature extraction, fully exploiting the potential of Visual Language Models (VLMs) in event-based recognition. The stacked frames $\pmb{I}_{\mathrm{s}}$ and reconstructed frames $\pmb{I}_{\mathrm{r}}$ after being enhanced for event representation, are fed into the vision encoder to obtain the event feature $\pmb{f}^{\mathrm{e}}\in \mathbb{R}^{1\times D}$ . Eventually, the class probability $p\in \mathbb{R}^{1\times C}$ for the action recognition are obtained as, + +$$ +p = \operatorname {s o f t m a x} _ {(1)} \left(\boldsymbol {f} ^ {\mathrm {e}} \left(\boldsymbol {f} ^ {\mathrm {t}}\right) ^ {\top}\right), \tag {2} +$$ + +$$ +c = \operatorname {a r g m a x} (p), +$$ + +where $\mathrm{softmax}(\cdot)$ denotes the softmax function. By taking the category of maximum probability, we obtain the final action prediction class $c$ for the event stream. + +# 3.2. Enhanced-Multimodal Perceptual (EMP) + +To preserve diverse information within and across modalities while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. Fig. 2 presents the architecture of EMP. It is equipped with Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enabling progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level. + +Cross-modal Frequency Enhancer (CFE). Event frames from distinct modalities demonstrate complementary spectral characteristics [55]. Specifically, stacked frames from asynchronous event streams predominantly focus on capturing high-frequency edge dynamics of transient actions, + +whereas reconstructed frames preserve mid-frequency texture inherent to activity semantics. Both modalities are crucial for event action recognition, yet effectively harnessing their complementary strengths remains a critical challenge. The direct combination of these heterogeneous modalities yields suboptimal solutions because of their inherent discrepancies. Thus, we propose a Cross-modal Frequency Enhancer (CFE) that strategically leverages frequency traits across modalities to refine event representation. It enables the model to focus on more comprehensive features through frequency-domain enhancement, thereby optimizing the synergy between modality-specific information. + +Whitin the CFE, we first transform stacked frames $I_{\mathrm{s}}$ and reconstructed frames $I_{\mathrm{r}}$ into the frequency domain via 3D Fast Fourier Transform (FFT) [29] as: + +$$ +\begin{array}{l} \boldsymbol {I} _ {\mathrm {s}} ^ {\mathrm {f f t}} = \operatorname {F F T} (\boldsymbol {I} _ {\mathrm {s}}), \\ \boldsymbol {I} _ {\mathrm {s}} ^ {\mathrm {f f t}} = \operatorname {F F T} (\boldsymbol {I} _ {\mathrm {s}}) \end{array} \tag {3} +$$ + +$$ +\boldsymbol {I} _ {\mathrm {r}} ^ {\mathrm {f f t}} = \operatorname {F F T} (\boldsymbol {I} _ {\mathrm {r}}). +$$ + +The spectral conversion enables the explicit decomposition of modality-specific frequency components, where edge-dominant stacked frames predominantly occupy high-frequency bands while texture-abundant reconstructed frames exhibit concentrated energy in mid-frequency regions. The frequency-aware representation establishes a unified foundation for cross-modal information enhancement. Subsequently, to selectively amplify edge-dynamic information in stacked frames from the frequency spectrum, we apply a sigmoid layer to $I_{\mathrm{s}}^{\mathrm{ftt}}$ , generating a frequency filter A. The proposed filtering is calculated as follows: + +$$ +\mathrm {A} = \operatorname {s i g m o i d} \left(\boldsymbol {I} _ {\mathrm {s}} ^ {\mathrm {f t t}}\right), \tag {4} +$$ + +$$ +\boldsymbol {I} _ {\mathrm {r}} ^ {\mathrm {f f t} ^ {\prime}} = \boldsymbol {I} _ {\mathrm {r}} ^ {\mathrm {f f t}} \odot \mathrm {A}, +$$ + +where $\odot$ denotes element-wise multiplication. Finally, we transfer the $I_{\mathrm{r}}^{\mathrm{fft}'}$ in the frequency space into the spatial domain to acquire enhanced reconstructed frames $I_{\mathrm{r}}'$ using an inverse Fast Fourier Transform (iFFT) [29] as: + +$$ +\boldsymbol {I} _ {\mathrm {r}} ^ {\prime} = \mathrm {i F F T} \left(\boldsymbol {I} _ {\mathrm {r}} ^ {\text {f f t}} ^ {\prime}\right). \tag {5} +$$ + +High-frequency Guided Selector (HGS). Given the inherent semantic coherence across modalities derived from the same event stream, existing methods [48] that sequentially extract and aggregate distinct modal features may introduce redundant feature interference. Building on cross-modal frequency enhancement, we further propose a high-frequency guided selection strategy to adaptively prioritize inter-modal critical regions that jointly characterize event semantics. Specifically, both stacked frames and enhanced reconstructed frames are processed through the shared vision encoder for visual feature learning. During event feature encoding, the [cls] token inherently aggregates global semantic representations through self-attention + +mechanisms [17]. We leverage high-frequency [cls] token from stacked frames to steer the attention mechanisms of reconstructed frames, enabling selective focus on cross-modal consistent semantic regions while suppressing modality-specific interference. + +For stacked frames $I_{\mathrm{s}}$ and enhanced reconstructed frames $I_{\mathrm{r}}^{\prime}$ , we first obtain the corresponding non-overlapping patches $f_{\mathrm{s}}, f_{\mathrm{r}} \in \mathbb{R}^{T \times (N + 1) \times D}$ including the special [cls] token $f_{\mathrm{s}}^{\mathrm{cls}}, f_{\mathrm{r}}^{\mathrm{cls}} \in \mathbb{R}^{1 \times D}$ [2] with D-dimension through a linear projection, where $N = H \times W / P^2$ and $(P, P)$ is the size of the patch. Between $l$ -th and $(l + 1)$ -th layer of the vision encoder, we deploy a cross-attention mechanism to facilitate inter-modal interaction. Here, the reconstructed tokens serve as query feature $Q$ , while the stacked tokens are transformed into key and value features, i.e. $K$ and $V$ . The detailed procedure can be formulated as: + +$$ +\begin{array}{l} Q = W _ {\mathrm {q}} \boldsymbol {f} _ {\mathrm {r}} ^ {l}, K = W _ {\mathrm {k}} \boldsymbol {f} _ {\mathrm {s}} ^ {l}, V = W _ {\mathrm {v}} \boldsymbol {f} _ {\mathrm {s}} ^ {l}, \\ \boldsymbol {f} _ {\mathrm {s}} ^ {l} = \operatorname {C r o s s A t t e n} (Q, K, V) = \operatorname {s o f t m a x} \left(\frac {Q K ^ {\top}}{\sqrt {d}}\right) V, \tag {6} \\ \end{array} +$$ + +where $W_{\mathrm{q}}, W_{\mathrm{k}}, W_{\mathrm{v}} \in \mathbb{R}^{D \times D}$ are the weights of the linear transformations corresponding to $Q, K, V$ , respectively. + +Subsequently, we extract the [cls] token $f_{\mathrm{s}}^{\mathrm{cls}}$ that encapsulates global semantics from the stacked features $f_{\mathrm{s}}^{l}$ and compute its semantic affinity with the patch tokens in the augmented reconstructed features $f_{\mathrm{r}}^{l}$ via the cosine similarity. It can be formulated as: + +$$ +\sin \left(\boldsymbol {f} _ {\mathrm {s}} ^ {\text {c l s}}, \boldsymbol {f} _ {\mathrm {s}} ^ {l}\right) = \frac {\boldsymbol {f} _ {\mathrm {s}} ^ {\text {c l s}} {} ^ {\top} \boldsymbol {f} _ {\mathrm {r}} ^ {l}}{\| \boldsymbol {f} _ {\mathrm {s}} ^ {\text {c l s}} \| \| \boldsymbol {f} _ {\mathrm {r}} ^ {l} \|}. \tag {7} +$$ + +By averaging the aforementioned similarity matrix $\mathrm{sim}(\pmb{f}_{\mathrm{s}}^{\mathrm{cls}},\bar{\pmb{f}}_{\mathrm{s}}^{l})$ along the temporal dimension $T$ , the final attention score matrix $B\in \mathbb{R}^{T\times N}$ is obtained. To maintain semantic consistency between the two modalities, we use $B$ to incorporate dynamic edge features, which guide the reconstruction branch in adaptively selecting tokens. The selection is formalized as: + +$$ +S = \left\{\operatorname {T o p} _ {K} (B) \right\}, \tag {8} +$$ + +where $\mathrm{Top}_K$ is a function that sorts scores in a set in descending order and outputs the indices corresponding to the $K$ largest values, $K$ is obtained by $\mu \cdot N$ and $\mu$ is a selection parameter. Using index set $S$ , we select the corresponding critical tokens from $\pmb{f}_{\mathrm{r}}^{l}$ and merge them with the $\pmb{f}_{\mathrm{r}}^{\mathrm{cls}}$ token before inputting them into the $(l + 1)$ -th visual encoder. + +# 3.3. Training Objectives + +Modality Consistency Loss. Our model takes the edge-stacked frames and the texture-reconstructed frames as input, which are crucial for capturing the discriminative features of the event stream. The HGS we proposed above focuses on critical token selection guided by high-frequency + +modality, based on the principle that different modalities maintain semantic consistency. Therefore, we introduce a loss to reduce the discrepancy between the two features of the same event in the feature space. The semantic consistency cross modalities $(f_{\mathrm{s}} / f_{\mathrm{r}})$ are constrained by the Mean Square Error (MSE) loss: + +$$ +\mathcal {L} _ {\mathrm {c}} = \sum_ {i = 1} ^ {N} \left| \left| \boldsymbol {f} _ {\mathrm {s}} ^ {i} - \boldsymbol {f} _ {\mathrm {r}} ^ {i} \right| \right| _ {2} ^ {2}, \tag {9} +$$ + +where $N$ is the number of tokens. + +Contractive Recognition Loss. The event-text consistency $\mathcal{L}_{\mathrm{r}}$ is constrained by the contrastive loss between the event feature and the text feature as follows: + +$$ +\mathcal {L} _ {\mathrm {r}} = - \frac {1}{B} \sum_ {i = 1} ^ {B} \log \frac {\exp \left(\left(\boldsymbol {f} _ {i} ^ {\mathrm {e}}\right) ^ {\top} \cdot \boldsymbol {f} _ {i} ^ {\mathrm {t}} / \tau\right)}{\sum_ {j = 1} ^ {B} \exp \left(\left(\boldsymbol {f} _ {i} ^ {\mathrm {e}}\right) ^ {\top} \cdot \boldsymbol {f} _ {j} ^ {\mathrm {t}} / \tau\right)}, \tag {10} +$$ + +where $\tau$ is the temperature coefficient. + +The final loss is composed of the modality consistency loss $\mathcal{L}_{\mathrm{c}}$ and the contractive recognition loss $\mathcal{L}_{\mathrm{r}}$ as + +$$ +\mathcal {L} = \mathcal {L} _ {\mathrm {c}} + \lambda \mathcal {L} _ {\mathrm {r}}, \tag {11} +$$ + +where $\lambda$ denotes the parameter used to balance the loss optimization. + +# 4. Experiments + +# 4.1. Datasets and Settings + +Datasets. To evaluate the performance of our approach, we conduct research on three datasets, including PAF [30], SeAct [57], HARDVS [50] and DVS128 Gesture [1]. DVS Action, also known as PAF, contains 10 action categories with 450 recordings, conducted in an unoccupied office environment. SeAct is the first semantic-abundant dataset for event-text action recognition, containing 58 actions under four themes. HARDVS is currently the largest dataset for event-based action recognition, currently having 107,646 recordings for 300 action categories. Both of the above three datasets have an average time duration of 5 seconds with $346 \times 260$ resolution. DVS128 Gesture is collected using a DAVIS128 camera with $128 \times 128$ resolution, dividing into 11 distinct classes of gestures. + +Settings. The implementation of the overall framework is carried out on PyTorch in a Linux environment with two NVIDIA GeForce 4090 GPUs. We use the Adam [23] optimizer with the initial learning rate of $1 \times 10^{-5}$ and weight decay of $2 \times 10^{-4}$ . CosineAnnealingLR learning rate schedule is employed with a minimum learning rate of $1 \times 10^{-6}$ . The representative lightweight FireNet [40] (only 38k parameters) for event-to-image reconstruction, which has low latency (1ms for $240 \times 180$ ). + +
MethodEvent RepresentationTop-1 Accuracy (%)
Frame-basedPoint-basedPAFSeActDVS128 Gesture
Motion SNN [28]-✓78.10-92.70
Slayer [41]-✓--93.64
MST [51]✓-88.21--
EV-ACT [14]✓-92.60--
EventTransAct [10]✓--57.8197.92
EvT [39]✓--61.3096.20
SpikePoint [37]-✓90.60-98.74
SpikMamba [8]✓-96.2871.0299.01
ExACT [57]✓-94.8367.2498.86
EventCrab [7]✓✓96.4972.4198.80
EventCLIP [56] + Recon✓-96.5571.0596.59
Ours✓-99.80(+3.31)75.00(+2.59)98.86(-0.15)
+ +# 4.2. Comparison with SOTA Methods + +Comparison on Datasets without Paired RGB. As shown in Tab. 1, we compare the proposed method EMP with several representative event-based action recognition methods on three event-only datasets (PAF, SeAct, DVS128 Gesture), where paired RGB data is unavailable. All compared methods exclusively utilize event-driven representations, categorized into event frame-based and event point-based paradigms. As demonstrated in Tab. 1, point-based methods exhibit marginally inferior performance compared to frame-based approaches on the PAF and DVS128 Gesture datasets. This discrepancy stems from the latter's enhanced capability in extracting discriminative spatiotemporal features through visual architectural designs. EventCrab [7] achieves impressive results by collaboratively mining complementary information from both event points and frames, thereby leveraging multi-modal event representations. However, unlike the paradigm of primarily exploring the sparse spatial characteristics of event streams from EventCrab, our method introduces texture-rich reconstructed frames and holistically exploits intrinsic correlations and extrinsic specificities between event frames and reconstructed frames. This unified perceptual framework yields significantly improved recognition efficacy, attaining $3.31\%$ (Top-1) and $2.59\%$ (Top-1) accuracy gains on the PAF and SeAct datasets respectively, while maintaining competitive performance on the DVS128 Gesture dataset. + +Comparison on Datasets with Paired RGB. As shown in Tab. 2, we compare the proposed EMP with several representative EAR methods on the HARDVS dataset, which includes 300-class paired RGB-Event data. Comparative methods are categorized into two types: (i) those utilizing RGB modality and (ii) those solely relying on event streams. We can find that combining RGB frames with + +Table 1. Comparative performance for EAR on the PAF, SeAct and DVS128 Gesture datasets. The best results are in bold and the second-best ones are in underlined. + +
MethodRGBParam.Top-1 Accuracy (%)
R2Plus1D [46]X63.5M49.06
C3D [45]X147.2M49.94
TimeSformer [4]X121.2M50.77
ESTF [50]X46.1M51.22
ExACT [57]X155.1M90.10
ESTF* [50]✓76.8M49.93
SAFE [24]✓-50.17
C3D* [45]✓245.3M50.88
TimeSformer* [4]✓202.3M51.57
SSTFormer [48]✓336.4M52.97
OursX182.1M97.68
+ +Table 2. Comparative performance (\%) on the HARDVS datasets. The best results are in **bold. (·)* denotes the combination between RGB frames and event streams. + +event streams yields marginal improvements over the single modality-based version. For instance, C3D [45] achieves an increase from $49.94\%$ to $50.88\%$ (Top-1), while TimeSformer improves from $50.77\%$ to $51.57\%$ (Top-1). However, the significant semantic discrepancy between RGB and event modalities limits performance gains from superficial cross-modal fusion. Different from conventional fusion paradigms, EMP adopts the text-aligned adaptation strategy to bridge the modality gap, while leveraging frequency-domain characteristics to extract discriminative event features. The results in Tab. 2 validate the efficacy of our approach, demonstrating superior robustness in handling heterogeneous modality. + +# 4.3. Ablation Studies + +Effectiveness of Each Component. As detailed in Tab. 3, we conduct an ablation study to evaluate the contributions of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS) within the pro + +
CFEHGSSeAct (%)DVS128 Gesture (%)
XX71.2597.34
✓X72.4197.72
X✓73.2798.10
✓✓75.0098.86
+ +Table 3. Ablation study for effectiveness of each component in the proposed method EMP. + +
Selection ratioTop-1 (%)Top-5 (%)
0.270.6890.51
0.471.5592.24
0.675.0095.68
0.873.2793.96
1.072.4193.10
+ +posed EMP. The baseline model trained only on eventreconstructed frames without incorporating the CFE and HGS achieves $71.55\%$ accuracy on SeAct and $97.34\%$ on DVS128 Gesture. When adding the baseline with CFE, we can find the performance improvements of $0.86\%$ on SeAct and $0.38\%$ on DVS128 Gesture, respectively. These performance improvements demonstrate the effectiveness of cross-modal representation enhancement in the frequency domain. Furthermore, the introduction of HGS achieves dynamic token selection across modalities, resulting in a $1.72\%$ improvement on SeAct and a $0.76\%$ improvement on DVS128 Gesture. By integrating all components, our model achieves the optimal performance. These results validate the effectiveness of our EMP in different scenarios. + +# 4.4. Diagnose Studies + +Impact of Different Token Selection Ratio. As illustrated in Tab. 4, we investigate the impact of varying selection thresholds $\mu$ on the SeAct dataset, where $\mu$ governs the proportion of semantically consistent tokens selected across modalities in the HGS framework. The experimental results reveal that the recognition accuracy improves as $\mu$ increases, reaching its peak at $\mu = 0.6$ , beyond which further threshold escalation correlates with performance degradation. This trend underscores the inherent divergence between modalities, suggesting that indiscriminate cross-modal fusion is suboptimal. Critically, calibrating $\mu$ to an appropriate level optimally harnesses intermodal complementarity while suppressing shared redundancies. + +Impact of Different Selection Layer. As shown in Tab. 5, we investigate the impact of integrating token selection at distinct layers of the vision encoder on the SeAct dataset, with the optimal selection ratio fixed at $\mu = 0.6$ (Tab. 4). We implement modality-guided token selection across indi + +Table 4. Different selection ratio on the SeAct dataset. + +
Insertion LayerTop-1 (%)Top-5 (%)
[4]73.2793.96
[6]73.9692.41
[8]75.0095.68
[10]74.1393.10
+ +Table 5. Different insertion layer of token selection strategy on the SeAct dataset. + +![](images/0b9b8ad542c1b4adb41fc6a2298fcaa12c0d35769d7fa6cf5f1ad4044b3e82e4.jpg) +Figure 3. Effect of different values of $\lambda$ on SeAct and DVS 128 Gesture datasets. + +vidual layers and observe that shallow and deep layers yield marginal performance gains. In contrast, layer 8 achieves peak Top-1, validating that appropriately positioned layers within vision encoders optimally encode cross-modal feature representations, thereby enabling effective modality-guided selection of semantically consistent tokens. + +Balance Between Consistency Loss and Recognition Loss. Our empirical validation substantiates that achieving equilibrium between modality consistency loss and contractive recognition (Eq. 11) loss constitutes a pivotal design criterion for robust multimodal learning. The contribution of $\mathcal{L}_{\mathrm{r}}$ is modified via $\lambda$ . As shown in Fig. 3, the variation of $\lambda$ affects the accuracy performance. The best performance is achieved with weights of 0.6, 0.4 on SeAct and DVS128 Gesture datasets respectively. It is worth noting that our findings highlight the potential benefits of leveraging consistency across modalities to enhance the representational capacity and recognition robustness of our EMP. + +# 4.5. Qualitative Analysis + +Visualization of Cross-modal Frequency Enhancer (CFE). To investigate the efficacy of the designed CFE, we visualize the four variants of event frames on the SeAct dataset: raw stacked frames, reconstructed frames, enhanced frames by sharpening operators, and enhanced frames by our CFE. As shown in fig. 4, conventional edge-enhancement techniques [33] yield marginal improvements over raw reconstructed frames. In contrast, event frames processed by CFE consistently highlight the critical regions + +![](images/5a05f721c0097c87947867fdae625e46254ecd1a8ad66f8911cea2ebec49b5fe.jpg) +Figure 4. Visualization of event frames. "Stack" and "Reconstruct" denote stacked and reconstructed frames, respectively; "Edge Enhance" indicates enhanced reconstructed frames employing sharpening operators; "CFE" represents enhanced reconstructed frames by our Cross-modal Frequency Enhancer. + +across both individual actions and interaction scenarios. Our qualitative analysis confirms that the proposed CFE selectively refines the high-frequency components of the reconstructed frames through spectral coordination with the edge dynamics of the stacked frames, thereby generating complementary event representations with synergistic spatial characteristics. + +Visualization of Guidanced Selection. Fig. 5 visualizes the important tokens selected by the High-frequency Guided Selector (HGS). As shown in Fig. 5, the HGS demonstrates the adaptive selection of action-salient tokens across sequential event frames. This demonstrates the intrinsic consistency among multi-modal frames originating from the identical event stream, where HGS effectively identifies semantically coherent tokens while suppressing interference from spatiotemporal redundancies. The results highlight the ability of HGS to focus on critical regions of interest, thereby enhancing the discriminative power of EMP. This selection is crucial for improving the robustness and accuracy of action recognition in dynamic scenes. + +Visualization of Event Features Learned by EMP. As illustrated in Fig. 6, we visualize the feature distributions of our proposed EMP against the baseline on the SeAct dataset. Seven action classes are randomly selected to demonstrate inter-class feature separability. Empirical results demonstrate the capability of our EMP in achieving both intra-class compactness and inter-class dispersion, further substantiating its efficacy in preserving semantic dis + +![](images/04aff77dad66b914df2ec9e5e3ff7b8500b256e54e47ba40471d68f0758ffdc0.jpg) +Figure 5. Visualization of cross-modal consistent tokens selected by the HGS, with unselected regions masked in white. + +![](images/4dfdd280301d7d5eb199bfadddda1c4fa644873a1f55848e0ab5ff538c7e00e4.jpg) +Figure 6. The t-SNE visualization of 7 randomly selected actions. Different color indicates different actions. + +tinctiveness across classes. + +# 5. Conclusion and Discussion + +Conclusion. In this work, we propose a privacy-preserving paradigm that eliminates RGB dependencies by generating texture-enriched reconstructed frames directly from raw event streams, reducing multi-sensor acquisition costs while bridging the modality gap with event frames. To preserve diverse information across reconstructed and stacked frames while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. It consists of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), which enables the progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level to acquire discriminative event features. Extensive experimental results validate the effectiveness of the proposed method. + +Discussion. While our method achieves promising performance in EAR and demonstrates the potential for scalability, its ability to handle more challenging scenarios remains to be explored. Future work will focus on developing a robust multi-modal unified perceptual approach to cope with diverse and variable scenes. + +Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208), and the Nanjing University Integrated Research Platform of the Ministry of Education-Top Talents Program (Grant No. 2024300482). + +# References + +[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7243-7252, 2017. 5 +[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6836-6846, 2021. 5 +[3] Sami Barchid, Benjamin Allaert, Amel Aissaoui, José Mennesson, and Chaabane C Djeraba. Spiking-fer: spiking neural network for facial expression recognition with event cameras. In Proceedings of the 20th International Conference on Content-based Multimedia Indexing, pages 1–7, 2023. 2 +[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning, page 4, 2021. 6 +[5] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9286-9296, 2024. 1 +[6] Meiqi Cao, Rui Yan, Xiangbo Shu, Jiachao Zhang, Jinpeng Wang, and Guo-Sen Xie. Mup: multi-granularity unified perception for panoramic activity recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7666-7675, 2023. 3 +[7] Meiqi Cao, Xiangbo Shu, Jiachao Zhang, Rui Yan, Zechao Li, and Jinhui Tang. Eventcrab: Harnessing frame and point synergy for event-based action recognition and beyond. arXiv preprint arXiv:2411.18328, 2024. 2, 6 +[8] Jiaqi Chen, Yan Yang, Shizhuo Deng, Da Teng, and Liyuan Pan. Spikmamba: When snn meets mamba in event-based human action recognition. In Proceedings of the 6th ACM International Conference on Multimedia in Asia, pages 1-8, 2024. 2, 6 +[9] Lan Chen, Haoxiang Yang, Pengpeng Shao, Haoyu Song, Xiao Wang, Zhicheng Zhao, Yaowei Wang, and Yonghong Tian. Velora: A low-rank adaptation approach for efficient rgb-event based recognition. arXiv preprint arXiv:2412.20064, 2024. 2 +[10] Tristan de Blegiers, Ishan Rajendrakumar Dave, Adeel Yousaf, and Mubarak Shah. Eventtransact: A video transformer-based framework for event-camera based action recognition. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1-7, 2023. 2, 6 +[11] Yongjian Deng, Hao Chen, and Youfu Li. A dynamic gcn with cross-representation distillation for event-based learn + +ing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1492-1500, 2024. 2 +[12] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 3 +[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 1 +[14] Yue Gao, Jiaxuan Lu, Siqi Li, Nan Ma, Shaoyi Du, Yipeng Li, and Qionghai Dai. Action recognition and benchmark using event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14081-14097, 2023. 6 +[15] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1 +[16] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023. 1 +[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 5 +[18] Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, and Changhu Wang. Transfg: A transformer architecture for fine-grained recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 852-860, 2022. 3 +[19] Bo Jiang, Chengguo Yuan, Xiao Wang, Zhimin Bao, Lin Zhu, Yonghong Tian, and Jin Tang. Point-voxel absorbing graph representation learning for event stream based recognition. arXiv preprint arXiv:2306.05239, 2023. 3 +[20] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2570-2578, 2024. 3 +[21] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 3 +[22] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1 +[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5 +[24] Dong Li, Jiandong Jin, Yuhao Zhang, Yanlin Zhong, Yaoyang Wu, Lan Chen, Xiao Wang, and Bin Luo. + +Semantic-aware frame-event fusion based pattern recognition via large vision-language models. arXiv preprint arXiv:2311.18592, 2023. 2, 3, 6 +[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1 +[26] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 1 +[27] Haoyue Liu, Shihan Peng, Lin Zhu, Yi Chang, Hanyu Zhou, and Luxin Yan. Seeing motion at nighttime with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25648-25658, 2024. 1 +[28] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In International Joint Conference on Artificial Intelligence, pages 1743-1749, 2021. 6 +[29] Shuangbiao Liu, Qian Wang, and Geng Liu. A versatile method of discrete convolution and fft (dc-fft) for contact analyses. Wear, 243(1-2):101-111, 2000. 4 +[30] Shu Miao, Guang Chen, Xiangyu Ning, Yang Zi, Kejia Ren, Zhenshan Bing, and Alois Knoll. Neuromorphic vision datasets for pedestrian detection, action recognition, and fall detection. Frontiers in Neurorobotics, 13:38, 2019. 5 +[31] Federico Paredes-Valles and Guido CHE De Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 1 +[32] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 19935-19947, 2022. 2 +[33] R Pushpavalli and G Sivaradje. Switching median filter for image enhancement. International Journal of Scientific & Engineering Research, 3(2):1-5, 2012. 7 +[34] Hongyu Qu, Rui Yan, Xiangbo Shu, Hailiang Gao, Peng Huang, and Guo-Sen Xie. Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition. arXiv preprint arXiv:2405.02077, 2024. 3 +[35] Hongyu Qu, Jianan Wei, Xiangbo Shu, and Wenguan Wang. Learning clustering-based prototypes for compositional zero-shot learning. In Proceedings of the International Conference on Learning Representations, 2025. 3 +[36] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in Neural Information Processing Systems, 34:13937-13949, 2021. 3 +[37] Hongwei Ren, Yue Zhou, Yulong Huang, Haotian Fu, Xiaopeng Lin, Jie Song, and Bojun Cheng. Spikepoint: An efficient point-based spiking neural network for event cam + +eras action recognition. arXiv preprint arXiv:2310.07189, 2023. 2, 6 +[38] Hongwei Ren, Yue Zhou, Jiadong Zhu, Xiaopeng Lin, Hao-tian Fu, Yulong Huang, Yuetong Fang, Fei Ma, Hao Yu, and Bojun Cheng. Rethinking efficient and effective point-based networks for event camera classification and regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[39] Alberto Sabater, Luis Montesano, and Ana C Murillo. Event transformer. a sparse-aware solution for efficient event data processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2677-2686, 2022. 6 +[40] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 1, 2, 5 +[41] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 31, 2018. 6 +[42] Xiangbo Shu, Binqian Xu, Liyan Zhang, and Jinhui Tang. Multi-granularity anchor-contrastive representation learning for semi-supervised skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7559-7576, 2022. 1 +[43] Xiangbo Shu, Jiawen Yang, Rui Yan, and Yan Song. Expansion-squeeze-excitation fusion network for elderly activity recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5281-5292, 2022. 1 +[44] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3): 3200-3225, 2022. 2 +[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4489-4497, 2015. 6 +[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6 +[47] Hongjie Wang, Bhishma Dedhia, and Niraj K Jha. Zero-prune: Zero-shot token pruning through leveraging of the attention graph in pre-trained transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16070-16079, 2024. 3 +[48] Xiao Wang, Zongzhen Wu, Yao Rong, Lin Zhu, Bo Jiang, Jin Tang, and Yonghong Tian. Sstformer: Bridging spiking neural network and memory support transformer for frame-event based recognition. arXiv preprint arXiv:2308.04369, 2023. 2, 3, 4, 6 +[49] Xiao Wang, Shiao Wang, Pengpeng Shao, Bo Jiang, Lin Zhu, and Yonghong Tian. Event stream based human action + +recognition: a high-definition benchmark dataset and algorithms. arXiv preprint arXiv:2408.09764, 2024. 2 +[50] Xiao Wang, Zongzhen Wu, Bo Jiang, Zhimin Bao, Lin Zhu, Guoqi Li, Yaowei Wang, and Yonghong Tian. Hardvs: Revisiting human activity recognition with dynamic vision sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5615-5623, 2024. 2, 5, 6 +[51] Ziqing Wang, Yuetong Fang, Jiahang Cao, Qiang Zhang, Zhongrui Wang, and Renjing Xu. Masked spiking transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1761-1771, 2023. 6 +[52] Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. In European Conference on Computer Vision, pages 310-328, 2024. 1, 2 +[53] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 1 +[54] Bochen Xie, Yongjian Deng, Zhanpeng Shao, Hai Liu, and Youfu Li. Vmv-gcn: Volumetric multi-view based graph cnn for event stream classification. IEEE Robotics and Automation Letters, 7(2):1976-1983, 2022. 2 +[55] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17117-17126, 2024. 3, 4 +[56] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Eventbind: Learning a unified representation to bind them all for event-based open-world understanding. In European Conference on Computer Vision, pages 477-494, 2024. 4, 6 +[57] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024. 2, 4, 5, 6 \ No newline at end of file diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/images.zip b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..95e6e1c4ea9f6e4a1fa717922d9a8b356c6bf3f9 --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:3cb980747c32490c211daa81df932dd73075c5828e63d9129dedede41edf4cd4 +size 480760 diff --git a/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/layout.json b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..ddc99e978535f2a73d3943e425436fdf0349617b --- /dev/null +++ b/2025/Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition/layout.json @@ -0,0 +1,8764 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 548, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 548, + 139 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 548, + 139 + ], + "type": "text", + "content": "Exploiting Frequency Dynamics for Enhanced Multimodal Event-based Action Recognition" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "spans": [ + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": "Meiqi Cao" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": ", Xiangbo Shu" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": ", Xin Jiang" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": ", Rui Yan" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": ", Yazhou Yao" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": ", Jinhui Tang" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": "Nanjing University of Science and Technology \n" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": "Nanjing University \n" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 107, + 160, + 503, + 217 + ], + "type": "text", + "content": "Nanjing Forestry University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 104, + 219, + 503, + 232 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 104, + 219, + 503, + 232 + ], + "spans": [ + { + "bbox": [ + 104, + 219, + 503, + 232 + ], + "type": "text", + "content": "{cmq123,shuxb,xinjiang,ruiyan,yazhou.yao}@njust.edu.cn,tangjh@njfu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "spans": [ + { + "bbox": [ + 151, + 258, + 200, + 271 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 284, + 297, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 284, + 297, + 620 + ], + "spans": [ + { + "bbox": [ + 54, + 284, + 297, + 620 + ], + "type": "text", + "content": "While event cameras excel in capturing microsecond temporal dynamics, they suffer from sparse spatial representations compared to traditional RGB data. Thus, multimodal event-based action recognition approaches aim to synergize complementary strengths by independently extracting and integrating paired RGB-Event features. However, this paradigm inevitably introduces additional data acquisition costs, while eroding the inherent privacy advantages of event-based sensing. Drawing inspiration from event-to-image reconstruction, texture-enriched visual representation directly reconstructed from asynchronous event streams is a promising solution. In response, we propose an Enhanced Multimodal Perceptual (EMP) framework that hierarchically explores multimodal cues (e.g., edges and textures) from raw event streams through two synergistic innovations spanning representation to feature levels. Specifically, we introduce Cross-Modal Frequency Enhancer (CFE) that leverages complementary frequency characteristics between reconstructed frames and stacked frames to refine event representations. Furthermore, to achieve unified feature encoding across modalities, we develop High-Frequency Guided Selector (HGS) for semantic consistency token selection guided by dynamic edge features while suppressing redundant multimodal information interference adaptively. Extensive experiments on four benchmark datasets demonstrate the superior effectiveness of our proposed framework. The code is available at https://github.com/caomq123/EMP." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 640, + 135, + 653 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 640, + 135, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 640, + 135, + 653 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 296, + 696 + ], + "type": "text", + "content": "Traditional frame-based cameras often encounter performance limitations when dealing with rapid motion or complex illumination conditions [22, 25, 42, 43]. In contrast," + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 324, + 261, + 545, + 329 + ], + "blocks": [ + { + "bbox": [ + 324, + 261, + 545, + 329 + ], + "lines": [ + { + "bbox": [ + 324, + 261, + 545, + 329 + ], + "spans": [ + { + "bbox": [ + 324, + 261, + 545, + 329 + ], + "type": "image", + "image_path": "73c0dbf6f4e9f26cec48fdfdbf116f63d7b0b172c725f2d93740f3c296eb4649.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 340, + 555, + 418 + ], + "lines": [ + { + "bbox": [ + 313, + 340, + 555, + 418 + ], + "spans": [ + { + "bbox": [ + 313, + 340, + 555, + 418 + ], + "type": "text", + "content": "Figure 1. (a) Previous multi-modal input approaches independently extract features from paired RGB-Event modalities, both introducing excessive RGB data dependencies and disrupting the privacy-preserving property unique to event cameras. (b) Our approach takes a lightful reconstruct network (Firenet [40]) to substitute RGB data while narrowing the modality divergence with events, thereby enabling enhanced unified multimodal perception." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 446, + 555, + 673 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 555, + 673 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 555, + 673 + ], + "type": "text", + "content": "event-driven neuromorphic visual sensors, also known as event cameras [13, 15], operate by asynchronously capturing per-pixel intensity changes only when a significant light change occurs. This unique functionality endows event cameras with several prominent advantages, including high dynamic range, low power consumption, and high temporal resolution, which enable them to excel in challenging scenarios such as fast motion, extreme lighting conditions, and privacy-sensitive applications. These attributes have led to their widespread adoption in various computer vision tasks, including object detection and tracking [16, 26, 52], motion deblurring [5, 27], and image reconstruction [31, 40, 53]. Notably, the inherent sensitivity of event cameras to dynamic scenes provides a unique advantage in capturing motion-related actions. Leveraging these strengths, the task of Event-based Action Recognition (EAR) has emerged, aiming to effectively extract action features from event streams to enhance recognition performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 715 + ], + "type": "text", + "content": "Currently, the EAR methods faces significant challenges in effectively utilizing the spatially sparse and asynchronous event streams. Traditional single-modal input approaches" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 143, + 713 + ], + "type": "text", + "content": "*Corresponding author" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5969" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 204 + ], + "type": "text", + "content": "adopt two primary strategies: (i) Processing asynchronous event points through point-based encoders (GCNs [11, 54]) or spiking neural networks (SNNs [3, 8, 37]), and (ii) Applying CNN- [32, 44], Transformer- [7, 57] or Mamba- [38, 49] based visual encoders to frames stacked from asynchronous event streams. While point-based methods accommodate the asynchronous nature of events effectively, they exhibit suboptimal recognition performance. Simultaneously, frame-based methods suffer from the lack of critical texture information due to the focus on dynamic edge capture inherent in event cameras." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 206, + 295, + 386 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 206, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 55, + 206, + 295, + 386 + ], + "type": "text", + "content": "In recent years, there has been a notable shift toward multi-modal input strategies [9, 24, 48, 50], leveraging diverse data sources to improve feature extraction robustness for practical applications. As depicted in Fig. 1(a), conventional multi-modal methods typically process RGB images and event streams in parallel through independent encoders, followed by late fusion. Although such approaches improve robustness by aggregating complementary features, two critical limitations persist: (i) The incorporation of RGB images not only introduces additional data acquisition costs but also compromises the intrinsic privacy-preserving nature of event cameras by capturing detailed individual visuals. (ii) Independent encoders disregarding inter-modal correlations introduce redundancy that critically undermines the acquisition of discriminative action features." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 388, + 295, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 295, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 295, + 555 + ], + "type": "text", + "content": "To address the challenges above, we seek a privacy-preserving alternative to RGB data that eliminates additional data dependency and narrows the representational gap between surrogate modal with event modal. As shown in Fig. 1(b), inspired by the event-to-image [40] techniques, we employ a lightweight event reconstruction network to generate texture-enriched surrogate frames directly from event streams. This approach effectively emulates RGB semantics while filtering privacy-sensitive details. Building upon this, we introduce a unified multi-modal encoding architecture that harmonizes edge-focused event representations with texture-augmented surrogates, preserving modality-specific discriminative strengths while suppressing cross-modal redundant information interference." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 714 + ], + "type": "text", + "content": "Formally, we present an Enhanced Multi-modal Perceptual (EMP) framework designed to hierarchically exploit intrinsic multi-modal cues from the representation level to the feature level. Specifically, our framework consists of two components: Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS). The CFE leverages the distinct frequency traits inherent in different modalities, utilizing the high-frequency information (e.g., edges) from stacked event frames to enhance the reconstructed event frames including mid-frequency information (e.g., texture). This process effectively refines the event representations for better feature extraction. In Addition, considering the discrepancy between modalities," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "the HGS adaptively captures semantic consistency tokens across different modalities, thereby realizing unified encoding of multimodal features." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 326, + 109, + 534, + 120 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 109, + 534, + 120 + ], + "spans": [ + { + "bbox": [ + 326, + 109, + 534, + 120 + ], + "type": "text", + "content": "The main contributions are summarized as follows," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 121, + 553, + 276 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 314, + 121, + 553, + 169 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 121, + 553, + 169 + ], + "spans": [ + { + "bbox": [ + 314, + 121, + 553, + 169 + ], + "type": "text", + "content": "- A privacy-preserving paradigm inspired by event-to-image generates texture-enriched surrogate frames as alternatives to RGB images, eliminating auxiliary data acquisition while bridging the modality gap with events." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 170, + 553, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 170, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 170, + 553, + 216 + ], + "type": "text", + "content": "- A novel Enhanced-Multimodal Perceptual (EMP) framework that hierarchically explores intrinsic multi-modal cues from raw event streams through dual-stage innovations at representation and feature levels." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 217, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 553, + 276 + ], + "type": "text", + "content": "- A multi-modal unified perception strategy equipped with collaborative Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enables progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level." + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 289, + 400, + 301 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 309, + 551, + 322 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 551, + 322 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 551, + 322 + ], + "type": "text", + "content": "2.1. Single-modal Event-based Action Recognition" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 327, + 555, + 639 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 555, + 639 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 555, + 639 + ], + "type": "text", + "content": "Existing single-modal approaches for event-based action recognition (EAR) focus on extracting discriminative features from event streams [8, 52]. These methods leverage the sparse spatial nature of event data by stacking asynchronous event streams into compact frames, which are processed using powerful frame-specific architectures, e.g. convolutional neural networks (CNNs) [32, 44] or visual transformers (ViTs) [7, 57]. For instance, EventTransAct [10] introduce video transformers and event contrastive learning to extract fine-grained features from the event stream. Similarly, ExACT [57] incorporate rich semantic information to effectively handle event frames from a cross-modal perspective. Beyond frame-based methods, some works [37, 54] explore networks tailored to the unique properties of asynchronous events, such as spiking neural networks (SNNs) and graph convolutional networks (GCNs). For example, SpikePoint [37] proposed an end-to-end SNN architecture that processes sparse event streams by simultaneously extracting global and local features. VMvGCN [54] introduced a novel volumetric multi-view framework to capture key spatial and temporal information from event streams. However, these approaches primarily fail to fully exploit the complementary infomation inherent in raw event streams. In contrast, our proposed Enhanced Multi-modal Perceptual framework synergistically integrates features from different modalities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 647, + 549, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 549, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 549, + 661 + ], + "type": "text", + "content": "2.2. Multi-modal Event-based Action Recognition" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Recent advances in robust representation for multi-modal data garner significant attention across various visual domains [7, 48]. For EAR, previous research explore two primary paradigms for multi-modal input: (i) Integrating" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5970" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 77, + 541, + 297 + ], + "blocks": [ + { + "bbox": [ + 70, + 77, + 541, + 297 + ], + "lines": [ + { + "bbox": [ + 70, + 77, + 541, + 297 + ], + "spans": [ + { + "bbox": [ + 70, + 77, + 541, + 297 + ], + "type": "image", + "image_path": "64cc22384c3bf98a7248cf3e92627ebd89f8fce5cbe56a5a2eb9a92ab9ca81c9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 308, + 555, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 308, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 308, + 555, + 365 + ], + "type": "text", + "content": "Figure 2. The overall framework of our proposed EMP. First, complementary event modalities (stacked and reconstructed frames) undergo frequency-domain refinement via the Cross-modal Frequency Enhancer (CFE), which amplifies edge dynamics while preserving textural in reconstructed counterparts. After that, during the unified visual encoding process of CLIP, the High-frequency Guided Selector (HGS) is employed to capture semantic-consistency tokens across modalities, which effectively eliminates spatiotemporally redundant features. Finally, event features and text features derived from the text/vision encoder are contrasted for activity recognition." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 385, + 296, + 661 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 385, + 296, + 661 + ], + "spans": [ + { + "bbox": [ + 56, + 385, + 296, + 661 + ], + "type": "text", + "content": "different representations of event data (e.g., event frames, event points, and event voxels); and (ii) Fusing RGB images with event data. For instance, AGCN [19] introduce a dual point-voxel absorbing graph representation to exploit complementary information between event points and event voxels. EventCrab [6] propose a framework that synergistically processes event frames and event points balancing efficiency and effectiveness. However, while these methods fully leverage the rich temporal information within event streams, they overlook the inherent lack of texture information in events. To address this limitation, SAFE [24] leverages pre-trained large-scale vision-language models to fuse semantic labels, RGB frames, and event streams, improving multi-modal recognition accuracy. Similarly, sstformer [48] formally propose to recognize patterns by fusing RGB frames and event streams simultaneously. Despite these advances, existing methods require additional resources for RGB images and fail to fully exploit the consistency between multi-modal data. In contrast, our proposed framework eliminates dependency on RGB while explicitly addressing the impact of irrelevant information on multimodal feature selection. Our approach provides a promising solution for practical deployment in real-world scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 671, + 227, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 227, + 683 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 227, + 683 + ], + "type": "text", + "content": "2.3. Token Selection in Transformer" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 715 + ], + "type": "text", + "content": "With the widespread adoption of Transformer across various domains [6, 20, 21, 34, 35], token selection has gar" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 385, + 555, + 625 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 625 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 625 + ], + "type": "text", + "content": "nered increasing attention for its ability to leverage inherent redundancies between image tokens to focus on key objects and reduce computational costs [12, 36, 47, 55]. This is particularly crucial in visual tasks [36], where the extraction of fine-grained features is essential for visual understanding. DynamicViT [18, 36] inserts prediction modules into Transformer blocks to predict and select tokens with higher information content. TransFG [18] directly employs attention mechanisms to locate important tokens, demonstrating strong performance in fine-grained classification. Building on this, STA [12] proposes a framework that simultaneously considers temporal redundancy and semantic importance for selecting spatiotemporal tokens. Magic [55] extends the token selection to multi-modal tasks, enabling the model to dynamically choose object-centric tokens based on different input modalities. Unlike these methods, we propose a more flexible multi-modal token selection scheme that utilizes high-frequency modal-guided strategies to capture semantic consistency tokens, thereby enhancing the robustness of our approach." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 638, + 370, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 638, + 370, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 638, + 370, + 651 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 659, + 422, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 659, + 422, + 671 + ], + "spans": [ + { + "bbox": [ + 314, + 659, + 422, + 671 + ], + "type": "text", + "content": "3.1. Joint Formulation" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 715 + ], + "type": "text", + "content": "We introduce a jointly formulated framework for event-to-image reconstruction and CLIP-driven action recognition, without the need for additionally captured RGB images." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5971" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": "Event-to-Image Reconstruction. Considering the event stream " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "\\pmb{E} = \\{(u_i, p_i, t_i)\\}_{i \\in (0, N)}" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " denotes the pixel position, " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "p_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " indicates the polarity, " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "t_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " denotes the time of each event point, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " is the number of event points. We first generate two complementary event representations from raw asynchronous event streams: (i) Edge-dominant stacked frames " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "I_s" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " obtained by temporally aggregating events; (ii) Texture-enriched reconstructed frames " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "I_r" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " obtained by the image reconstruction network " + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "inline_equation", + "content": "H" + }, + { + "bbox": [ + 55, + 72, + 296, + 179 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 149, + 191, + 295, + 205 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 149, + 191, + 295, + 205 + ], + "spans": [ + { + "bbox": [ + 149, + 191, + 295, + 205 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {\\mathrm {r}} = \\mathrm {H} (\\boldsymbol {E}), \\tag {1}", + "image_path": "58c7d5f6b1cd227c27781774fced9b11a18f91428f9da37044ef7f41f5bd4722.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 213, + 173, + 226 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 213, + 173, + 226 + ], + "spans": [ + { + "bbox": [ + 55, + 213, + 173, + 226 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 213, + 173, + 226 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{\\mathrm{s}},\\pmb{I}_{\\mathrm{r}}\\in \\mathbb{R}^{T\\times H\\times W\\times 3}" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": "CLIP-driven Action Recognition. In terms of " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " action classes, we follow [57] to adopt two different text prompts: hand-crafted prompt \"A series of photos recording action for [class]\" and learnable prompt \"[V1], [V2], ..., [Vn], [class]\", where [class] represents the class name and " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "[V_n]" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " is the random initialized parameter. The CLIP-based text encoder [56] first encodes the two prompts to obtain D-dimensional text features " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2^{\\mathrm{t}}\\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": ", and the final text feature " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{t}\\in \\mathbb{R}^{C\\times D}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " is obtained by averaging " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}_1^{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}_2^{\\mathrm{t}}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": ". Meanwhile, we use the CLIP-based vision encoder [56] for event feature extraction, fully exploiting the potential of Visual Language Models (VLMs) in event-based recognition. The stacked frames " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " and reconstructed frames " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{I}_{\\mathrm{r}}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " after being enhanced for event representation, are fed into the vision encoder to obtain the event feature " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "\\pmb{f}^{\\mathrm{e}}\\in \\mathbb{R}^{1\\times D}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": ". Eventually, the class probability " + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "inline_equation", + "content": "p\\in \\mathbb{R}^{1\\times C}" + }, + { + "bbox": [ + 55, + 227, + 296, + 431 + ], + "type": "text", + "content": " for the action recognition are obtained as," + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 123, + 441, + 295, + 461 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 441, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 123, + 441, + 295, + 461 + ], + "type": "interline_equation", + "content": "p = \\operatorname {s o f t m a x} _ {(1)} \\left(\\boldsymbol {f} ^ {\\mathrm {e}} \\left(\\boldsymbol {f} ^ {\\mathrm {t}}\\right) ^ {\\top}\\right), \\tag {2}", + "image_path": "6c0fd1432a7e44bf250f6d55c6c8b1b9ef4ab38d8d7b8986c57389f80042239a.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 126, + 459, + 194, + 470 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 459, + 194, + 470 + ], + "spans": [ + { + "bbox": [ + 126, + 459, + 194, + 470 + ], + "type": "interline_equation", + "content": "c = \\operatorname {a r g m a x} (p),", + "image_path": "665503d78f2da717741f669178717a633ce582d11bea4d78945e8367a6f5ec04.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "\\mathrm{softmax}(\\cdot)" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": " denotes the softmax function. By taking the category of maximum probability, we obtain the final action prediction class " + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 481, + 296, + 518 + ], + "type": "text", + "content": " for the event stream." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 527, + 272, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 527, + 272, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 527, + 272, + 540 + ], + "type": "text", + "content": "3.2. Enhanced-Multimodal Perceptual (EMP)" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 652 + ], + "type": "text", + "content": "To preserve diverse information within and across modalities while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. Fig. 2 presents the architecture of EMP. It is equipped with Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), enabling progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Cross-modal Frequency Enhancer (CFE). Event frames from distinct modalities demonstrate complementary spectral characteristics [55]. Specifically, stacked frames from asynchronous event streams predominantly focus on capturing high-frequency edge dynamics of transient actions," + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "whereas reconstructed frames preserve mid-frequency texture inherent to activity semantics. Both modalities are crucial for event action recognition, yet effectively harnessing their complementary strengths remains a critical challenge. The direct combination of these heterogeneous modalities yields suboptimal solutions because of their inherent discrepancies. Thus, we propose a Cross-modal Frequency Enhancer (CFE) that strategically leverages frequency traits across modalities to refine event representation. It enables the model to focus on more comprehensive features through frequency-domain enhancement, thereby optimizing the synergy between modality-specific information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "content": "Whitin the CFE, we first transform stacked frames " + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{s}}" + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "content": " and reconstructed frames " + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 216, + 553, + 251 + ], + "type": "text", + "content": " into the frequency domain via 3D Fast Fourier Transform (FFT) [29] as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 399, + 258, + 553, + 280 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 258, + 553, + 280 + ], + "spans": [ + { + "bbox": [ + 399, + 258, + 553, + 280 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}), \\\\ \\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {s}}) \\end{array} \\tag {3}", + "image_path": "9d05286fe204eeeb71ac33b60f300fdf2150259c442bc2d9b90cd85d61e5ea35.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 400, + 277, + 468, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 400, + 277, + 468, + 289 + ], + "spans": [ + { + "bbox": [ + 400, + 277, + 468, + 289 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} = \\operatorname {F F T} (\\boldsymbol {I} _ {\\mathrm {r}}).", + "image_path": "0af36c50b81ccc3779c7cd5b94b7a120ceb89ef6ecc63120a9bd88b4512efeab.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 297, + 555, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 297, + 555, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 297, + 555, + 429 + ], + "type": "text", + "content": "The spectral conversion enables the explicit decomposition of modality-specific frequency components, where edge-dominant stacked frames predominantly occupy high-frequency bands while texture-abundant reconstructed frames exhibit concentrated energy in mid-frequency regions. The frequency-aware representation establishes a unified foundation for cross-modal information enhancement. Subsequently, to selectively amplify edge-dynamic information in stacked frames from the frequency spectrum, we apply a sigmoid layer to " + }, + { + "bbox": [ + 313, + 297, + 555, + 429 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{s}}^{\\mathrm{ftt}}" + }, + { + "bbox": [ + 313, + 297, + 555, + 429 + ], + "type": "text", + "content": ", generating a frequency filter A. The proposed filtering is calculated as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 399, + 435, + 553, + 457 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 399, + 435, + 553, + 457 + ], + "spans": [ + { + "bbox": [ + 399, + 435, + 553, + 457 + ], + "type": "interline_equation", + "content": "\\mathrm {A} = \\operatorname {s i g m o i d} \\left(\\boldsymbol {I} _ {\\mathrm {s}} ^ {\\mathrm {f t t}}\\right), \\tag {4}", + "image_path": "0ee277ab5755934905c1b6edbd5bdf4acb828e4dd2b2e2db02fb6ac05b279eb3.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 390, + 453, + 458, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 390, + 453, + 458, + 468 + ], + "spans": [ + { + "bbox": [ + 390, + 453, + 458, + 468 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t} ^ {\\prime}} = \\boldsymbol {I} _ {\\mathrm {r}} ^ {\\mathrm {f f t}} \\odot \\mathrm {A},", + "image_path": "ad39a5e4adf206504c3858c88a6554308c6c28ba6578c4e5e3818312fe70c813.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "spans": [ + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "text", + "content": " denotes element-wise multiplication. Finally, we transfer the " + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{r}}^{\\mathrm{fft}'}" + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "text", + "content": " in the frequency space into the spatial domain to acquire enhanced reconstructed frames " + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{r}}'" + }, + { + "bbox": [ + 313, + 476, + 555, + 526 + ], + "type": "text", + "content": " using an inverse Fast Fourier Transform (iFFT) [29] as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 396, + 534, + 553, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 396, + 534, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 396, + 534, + 553, + 552 + ], + "type": "interline_equation", + "content": "\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\prime} = \\mathrm {i F F T} \\left(\\boldsymbol {I} _ {\\mathrm {r}} ^ {\\text {f f t}} ^ {\\prime}\\right). \\tag {5}", + "image_path": "e3197155ef22cc762be860140ef1b221c55e812e29ddcf6b16783bc42e1b56b6.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 714 + ], + "type": "text", + "content": "High-frequency Guided Selector (HGS). Given the inherent semantic coherence across modalities derived from the same event stream, existing methods [48] that sequentially extract and aggregate distinct modal features may introduce redundant feature interference. Building on cross-modal frequency enhancement, we further propose a high-frequency guided selection strategy to adaptively prioritize inter-modal critical regions that jointly characterize event semantics. Specifically, both stacked frames and enhanced reconstructed frames are processed through the shared vision encoder for visual feature learning. During event feature encoding, the [cls] token inherently aggregates global semantic representations through self-attention" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5972" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 131 + ], + "type": "text", + "content": "mechanisms [17]. We leverage high-frequency [cls] token from stacked frames to steer the attention mechanisms of reconstructed frames, enabling selective focus on cross-modal consistent semantic regions while suppressing modality-specific interference." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": "For stacked frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{s}}" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " and enhanced reconstructed frames " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "I_{\\mathrm{r}}^{\\prime}" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": ", we first obtain the corresponding non-overlapping patches " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}, f_{\\mathrm{r}} \\in \\mathbb{R}^{T \\times (N + 1) \\times D}" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " including the special [cls] token " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}^{\\mathrm{cls}}, f_{\\mathrm{r}}^{\\mathrm{cls}} \\in \\mathbb{R}^{1 \\times D}" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " [2] with D-dimension through a linear projection, where " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "N = H \\times W / P^2" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "(P, P)" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " is the size of the patch. Between " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": "-th and " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "(l + 1)" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": "-th layer of the vision encoder, we deploy a cross-attention mechanism to facilitate inter-modal interaction. Here, the reconstructed tokens serve as query feature " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "Q" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": ", while the stacked tokens are transformed into key and value features, i.e. " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "inline_equation", + "content": "V" + }, + { + "bbox": [ + 55, + 133, + 295, + 263 + ], + "type": "text", + "content": ". The detailed procedure can be formulated as:" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 269, + 294, + 314 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 269, + 294, + 314 + ], + "spans": [ + { + "bbox": [ + 66, + 269, + 294, + 314 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} Q = W _ {\\mathrm {q}} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}, K = W _ {\\mathrm {k}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, V = W _ {\\mathrm {v}} \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}, \\\\ \\boldsymbol {f} _ {\\mathrm {s}} ^ {l} = \\operatorname {C r o s s A t t e n} (Q, K, V) = \\operatorname {s o f t m a x} \\left(\\frac {Q K ^ {\\top}}{\\sqrt {d}}\\right) V, \\tag {6} \\\\ \\end{array}", + "image_path": "63f84c6bff017fbc3ab4347bf9ba9378a64b0f1be8d51e2c16e987e406c5f547.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "spans": [ + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "inline_equation", + "content": "W_{\\mathrm{q}}, W_{\\mathrm{k}}, W_{\\mathrm{v}} \\in \\mathbb{R}^{D \\times D}" + }, + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "text", + "content": " are the weights of the linear transformations corresponding to " + }, + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "inline_equation", + "content": "Q, K, V" + }, + { + "bbox": [ + 55, + 319, + 295, + 344 + ], + "type": "text", + "content": ", respectively." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "spans": [ + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": "Subsequently, we extract the [cls] token " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " that encapsulates global semantics from the stacked features " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{s}}^{l}" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " and compute its semantic affinity with the patch tokens in the augmented reconstructed features " + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{r}}^{l}" + }, + { + "bbox": [ + 55, + 345, + 296, + 404 + ], + "type": "text", + "content": " via the cosine similarity. It can be formulated as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 114, + 411, + 295, + 441 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 114, + 411, + 295, + 441 + ], + "spans": [ + { + "bbox": [ + 114, + 411, + 295, + 441 + ], + "type": "interline_equation", + "content": "\\sin \\left(\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}}, \\boldsymbol {f} _ {\\mathrm {s}} ^ {l}\\right) = \\frac {\\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} {} ^ {\\top} \\boldsymbol {f} _ {\\mathrm {r}} ^ {l}}{\\| \\boldsymbol {f} _ {\\mathrm {s}} ^ {\\text {c l s}} \\| \\| \\boldsymbol {f} _ {\\mathrm {r}} ^ {l} \\|}. \\tag {7}", + "image_path": "87e8c02ce0d009eaa381d2e2dd37199d7e66725a777715cd7aae9d4af71350c6.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "spans": [ + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "content": "By averaging the aforementioned similarity matrix " + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{sim}(\\pmb{f}_{\\mathrm{s}}^{\\mathrm{cls}},\\bar{\\pmb{f}}_{\\mathrm{s}}^{l})" + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "content": " along the temporal dimension " + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "content": ", the final attention score matrix " + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "inline_equation", + "content": "B\\in \\mathbb{R}^{T\\times N}" + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "content": " is obtained. To maintain semantic consistency between the two modalities, we use " + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 55, + 447, + 296, + 529 + ], + "type": "text", + "content": " to incorporate dynamic edge features, which guide the reconstruction branch in adaptively selecting tokens. The selection is formalized as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 137, + 537, + 295, + 552 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 137, + 537, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 137, + 537, + 295, + 552 + ], + "type": "interline_equation", + "content": "S = \\left\\{\\operatorname {T o p} _ {K} (B) \\right\\}, \\tag {8}", + "image_path": "dbe680a8173be40b8808f91fd4e2aaa7ae5c5bfd7f31ec6396b329792acc2c43.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mathrm{Top}_K" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " is a function that sorts scores in a set in descending order and outputs the indices corresponding to the " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " largest values, " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " is obtained by " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mu \\cdot N" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " is a selection parameter. Using index set " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": ", we select the corresponding critical tokens from " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{r}}^{l}" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " and merge them with the " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "\\pmb{f}_{\\mathrm{r}}^{\\mathrm{cls}}" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": " token before inputting them into the " + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "inline_equation", + "content": "(l + 1)" + }, + { + "bbox": [ + 55, + 558, + 295, + 629 + ], + "type": "text", + "content": "-th visual encoder." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 635, + 171, + 649 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 635, + 171, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 635, + 171, + 649 + ], + "type": "text", + "content": "3.3. Training Objectives" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 296, + 714 + ], + "type": "text", + "content": "Modality Consistency Loss. Our model takes the edge-stacked frames and the texture-reconstructed frames as input, which are crucial for capturing the discriminative features of the event stream. The HGS we proposed above focuses on critical token selection guided by high-frequency" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": "modality, based on the principle that different modalities maintain semantic consistency. Therefore, we introduce a loss to reduce the discrepancy between the two features of the same event in the feature space. The semantic consistency cross modalities " + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "inline_equation", + "content": "(f_{\\mathrm{s}} / f_{\\mathrm{r}})" + }, + { + "bbox": [ + 313, + 72, + 554, + 144 + ], + "type": "text", + "content": " are constrained by the Mean Square Error (MSE) loss:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 387, + 154, + 554, + 188 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 154, + 554, + 188 + ], + "spans": [ + { + "bbox": [ + 387, + 154, + 554, + 188 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {c}} = \\sum_ {i = 1} ^ {N} \\left| \\left| \\boldsymbol {f} _ {\\mathrm {s}} ^ {i} - \\boldsymbol {f} _ {\\mathrm {r}} ^ {i} \\right| \\right| _ {2} ^ {2}, \\tag {9}", + "image_path": "609f64fa20018123c92bc15d9fdead751a740a6dba88a66c885df1c2238fa045.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 453, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 453, + 207 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 453, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 196, + 453, + 207 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 196, + 453, + 207 + ], + "type": "text", + "content": " is the number of tokens." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "spans": [ + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "text", + "content": "Contractive Recognition Loss. The event-text consistency " + }, + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 209, + 554, + 243 + ], + "type": "text", + "content": " is constrained by the contrastive loss between the event feature and the text feature as follows:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 252, + 554, + 283 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 252, + 554, + 283 + ], + "spans": [ + { + "bbox": [ + 326, + 252, + 554, + 283 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {r}} = - \\frac {1}{B} \\sum_ {i = 1} ^ {B} \\log \\frac {\\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {i} ^ {\\mathrm {t}} / \\tau\\right)}{\\sum_ {j = 1} ^ {B} \\exp \\left(\\left(\\boldsymbol {f} _ {i} ^ {\\mathrm {e}}\\right) ^ {\\top} \\cdot \\boldsymbol {f} _ {j} ^ {\\mathrm {t}} / \\tau\\right)}, \\tag {10}", + "image_path": "790bfff6518a1ae1c8216622afda714ca305257eee6ac7e7080f18697fe0bcf4.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 293, + 470, + 304 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 470, + 304 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 470, + 304 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 293, + 470, + 304 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 293, + 470, + 304 + ], + "type": "text", + "content": " is the temperature coefficient." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "text", + "content": "The final loss is composed of the modality consistency loss " + }, + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{c}}" + }, + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "text", + "content": " and the contractive recognition loss " + }, + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 304, + 554, + 329 + ], + "type": "text", + "content": " as" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 401, + 340, + 553, + 353 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 401, + 340, + 553, + 353 + ], + "spans": [ + { + "bbox": [ + 401, + 340, + 553, + 353 + ], + "type": "interline_equation", + "content": "\\mathcal {L} = \\mathcal {L} _ {\\mathrm {c}} + \\lambda \\mathcal {L} _ {\\mathrm {r}}, \\tag {11}", + "image_path": "f209cb71aa4dd1014664cb46d5afabfcaeac722304f11ba2fedfb6a6330894f9.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 363, + 554, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 554, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 554, + 385 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 363, + 554, + 385 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 363, + 554, + 385 + ], + "type": "text", + "content": " denotes the parameter used to balance the loss optimization." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 399, + 395, + 413 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 399, + 395, + 413 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 395, + 413 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 420, + 437, + 433 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 420, + 437, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 420, + 437, + 433 + ], + "type": "text", + "content": "4.1. Datasets and Settings" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": "Datasets. To evaluate the performance of our approach, we conduct research on three datasets, including PAF [30], SeAct [57], HARDVS [50] and DVS128 Gesture [1]. DVS Action, also known as PAF, contains 10 action categories with 450 recordings, conducted in an unoccupied office environment. SeAct is the first semantic-abundant dataset for event-text action recognition, containing 58 actions under four themes. HARDVS is currently the largest dataset for event-based action recognition, currently having 107,646 recordings for 300 action categories. Both of the above three datasets have an average time duration of 5 seconds with " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "346 \\times 260" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": " resolution. DVS128 Gesture is collected using a DAVIS128 camera with " + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 437, + 555, + 605 + ], + "type": "text", + "content": " resolution, dividing into 11 distinct classes of gestures." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": "Settings. The implementation of the overall framework is carried out on PyTorch in a Linux environment with two NVIDIA GeForce 4090 GPUs. We use the Adam [23] optimizer with the initial learning rate of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": " and weight decay of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "2 \\times 10^{-4}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ". CosineAnnealingLR learning rate schedule is employed with a minimum learning rate of " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-6}" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ". The representative lightweight FireNet [40] (only 38k parameters) for event-to-image reconstruction, which has low latency (1ms for " + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "inline_equation", + "content": "240 \\times 180" + }, + { + "bbox": [ + 313, + 605, + 554, + 713 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 23 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5973" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 88, + 70, + 525, + 246 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 525, + 246 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 525, + 246 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 525, + 246 + ], + "type": "table", + "html": "
MethodEvent RepresentationTop-1 Accuracy (%)
Frame-basedPoint-basedPAFSeActDVS128 Gesture
Motion SNN [28]-✓78.10-92.70
Slayer [41]-✓--93.64
MST [51]✓-88.21--
EV-ACT [14]✓-92.60--
EventTransAct [10]✓--57.8197.92
EvT [39]✓--61.3096.20
SpikePoint [37]-✓90.60-98.74
SpikMamba [8]✓-96.2871.0299.01
ExACT [57]✓-94.8367.2498.86
EventCrab [7]✓✓96.4972.4198.80
EventCLIP [56] + Recon✓-96.5571.0596.59
Ours✓-99.80(+3.31)75.00(+2.59)98.86(-0.15)
", + "image_path": "557ad002392a77660f1227f7f588b13ae9cd38207f36d1aa4c9e77786bf64d55.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 234, + 303 + ], + "type": "text", + "content": "4.2. Comparison with SOTA Methods" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "text", + "content": "Comparison on Datasets without Paired RGB. As shown in Tab. 1, we compare the proposed method EMP with several representative event-based action recognition methods on three event-only datasets (PAF, SeAct, DVS128 Gesture), where paired RGB data is unavailable. All compared methods exclusively utilize event-driven representations, categorized into event frame-based and event point-based paradigms. As demonstrated in Tab. 1, point-based methods exhibit marginally inferior performance compared to frame-based approaches on the PAF and DVS128 Gesture datasets. This discrepancy stems from the latter's enhanced capability in extracting discriminative spatiotemporal features through visual architectural designs. EventCrab [7] achieves impressive results by collaboratively mining complementary information from both event points and frames, thereby leveraging multi-modal event representations. However, unlike the paradigm of primarily exploring the sparse spatial characteristics of event streams from EventCrab, our method introduces texture-rich reconstructed frames and holistically exploits intrinsic correlations and extrinsic specificities between event frames and reconstructed frames. This unified perceptual framework yields significantly improved recognition efficacy, attaining " + }, + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "inline_equation", + "content": "3.31\\%" + }, + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "text", + "content": " (Top-1) and " + }, + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "inline_equation", + "content": "2.59\\%" + }, + { + "bbox": [ + 56, + 313, + 296, + 624 + ], + "type": "text", + "content": " (Top-1) accuracy gains on the PAF and SeAct datasets respectively, while maintaining competitive performance on the DVS128 Gesture dataset." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 295, + 713 + ], + "type": "text", + "content": "Comparison on Datasets with Paired RGB. As shown in Tab. 2, we compare the proposed EMP with several representative EAR methods on the HARDVS dataset, which includes 300-class paired RGB-Event data. Comparative methods are categorized into two types: (i) those utilizing RGB modality and (ii) those solely relying on event streams. We can find that combining RGB frames with" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 317, + 288, + 554, + 423 + ], + "blocks": [ + { + "bbox": [ + 55, + 253, + 555, + 277 + ], + "lines": [ + { + "bbox": [ + 55, + 253, + 555, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 253, + 555, + 277 + ], + "type": "text", + "content": "Table 1. Comparative performance for EAR on the PAF, SeAct and DVS128 Gesture datasets. The best results are in bold and the second-best ones are in underlined." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 317, + 288, + 554, + 423 + ], + "lines": [ + { + "bbox": [ + 317, + 288, + 554, + 423 + ], + "spans": [ + { + "bbox": [ + 317, + 288, + 554, + 423 + ], + "type": "table", + "html": "
MethodRGBParam.Top-1 Accuracy (%)
R2Plus1D [46]X63.5M49.06
C3D [45]X147.2M49.94
TimeSformer [4]X121.2M50.77
ESTF [50]X46.1M51.22
ExACT [57]X155.1M90.10
ESTF* [50]✓76.8M49.93
SAFE [24]✓-50.17
C3D* [45]✓245.3M50.88
TimeSformer* [4]✓202.3M51.57
SSTFormer [48]✓336.4M52.97
OursX182.1M97.68
", + "image_path": "ccd989ef29ce6ae73d1e0541ae8bcca9ca1e14fa41c3ce826de571be1589f435.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 464 + ], + "type": "text", + "content": "Table 2. Comparative performance (\\%) on the HARDVS datasets. The best results are in **bold. (·)* denotes the combination between RGB frames and event streams." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": "event streams yields marginal improvements over the single modality-based version. For instance, C3D [45] achieves an increase from " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "49.94\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "50.88\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " (Top-1), while TimeSformer improves from " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "50.77\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "inline_equation", + "content": "51.57\\%" + }, + { + "bbox": [ + 313, + 485, + 555, + 641 + ], + "type": "text", + "content": " (Top-1). However, the significant semantic discrepancy between RGB and event modalities limits performance gains from superficial cross-modal fusion. Different from conventional fusion paradigms, EMP adopts the text-aligned adaptation strategy to bridge the modality gap, while leveraging frequency-domain characteristics to extract discriminative event features. The results in Tab. 2 validate the efficacy of our approach, demonstrating superior robustness in handling heterogeneous modality." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 648, + 416, + 659 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 416, + 659 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 416, + 659 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Effectiveness of Each Component. As detailed in Tab. 3, we conduct an ablation study to evaluate the contributions of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS) within the pro" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5974" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 70, + 70, + 282, + 133 + ], + "blocks": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "lines": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "spans": [ + { + "bbox": [ + 70, + 70, + 282, + 133 + ], + "type": "table", + "html": "
CFEHGSSeAct (%)DVS128 Gesture (%)
XX71.2597.34
✓X72.4197.72
X✓73.2798.10
✓✓75.0098.86
", + "image_path": "752cd90724c1a703f684f8db80ab3c86998aaae49ae0d1cdc955035dd456a677.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 88, + 175, + 266, + 249 + ], + "blocks": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "lines": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 141, + 295, + 163 + ], + "type": "text", + "content": "Table 3. Ablation study for effectiveness of each component in the proposed method EMP." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 88, + 175, + 266, + 249 + ], + "lines": [ + { + "bbox": [ + 88, + 175, + 266, + 249 + ], + "spans": [ + { + "bbox": [ + 88, + 175, + 266, + 249 + ], + "type": "table", + "html": "
Selection ratioTop-1 (%)Top-5 (%)
0.270.6890.51
0.471.5592.24
0.675.0095.68
0.873.2793.96
1.072.4193.10
", + "image_path": "9e0cc1fa671869d0a34a75b473f36eacded77357cfb8a0dea6bd33ae0173df8a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": "posed EMP. The baseline model trained only on eventreconstructed frames without incorporating the CFE and HGS achieves " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "71.55\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " accuracy on SeAct and " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "97.34\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " on DVS128 Gesture. When adding the baseline with CFE, we can find the performance improvements of " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "0.86\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " on SeAct and " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "0.38\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " on DVS128 Gesture, respectively. These performance improvements demonstrate the effectiveness of cross-modal representation enhancement in the frequency domain. Furthermore, the introduction of HGS achieves dynamic token selection across modalities, resulting in a " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "1.72\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " improvement on SeAct and a " + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "inline_equation", + "content": "0.76\\%" + }, + { + "bbox": [ + 55, + 291, + 296, + 460 + ], + "type": "text", + "content": " improvement on DVS128 Gesture. By integrating all components, our model achieves the optimal performance. These results validate the effectiveness of our EMP in different scenarios." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 479, + 157, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 479, + 157, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 479, + 157, + 491 + ], + "type": "text", + "content": "4.4. Diagnose Studies" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": "Impact of Different Token Selection Ratio. As illustrated in Tab. 4, we investigate the impact of varying selection thresholds " + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": " on the SeAct dataset, where " + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": " governs the proportion of semantically consistent tokens selected across modalities in the HGS framework. The experimental results reveal that the recognition accuracy improves as " + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": " increases, reaching its peak at " + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu = 0.6" + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": ", beyond which further threshold escalation correlates with performance degradation. This trend underscores the inherent divergence between modalities, suggesting that indiscriminate cross-modal fusion is suboptimal. Critically, calibrating " + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\mu" + }, + { + "bbox": [ + 55, + 498, + 295, + 653 + ], + "type": "text", + "content": " to an appropriate level optimally harnesses intermodal complementarity while suppressing shared redundancies." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": "Impact of Different Selection Layer. As shown in Tab. 5, we investigate the impact of integrating token selection at distinct layers of the vision encoder on the SeAct dataset, with the optimal selection ratio fixed at " + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mu = 0.6" + }, + { + "bbox": [ + 55, + 654, + 295, + 714 + ], + "type": "text", + "content": " (Tab. 4). We implement modality-guided token selection across indi" + } + ] + } + ], + "index": 7 + }, + { + "type": "table", + "bbox": [ + 347, + 70, + 525, + 131 + ], + "blocks": [ + { + "bbox": [ + 77, + 258, + 273, + 268 + ], + "lines": [ + { + "bbox": [ + 77, + 258, + 273, + 268 + ], + "spans": [ + { + "bbox": [ + 77, + 258, + 273, + 268 + ], + "type": "text", + "content": "Table 4. Different selection ratio on the SeAct dataset." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 347, + 70, + 525, + 131 + ], + "lines": [ + { + "bbox": [ + 347, + 70, + 525, + 131 + ], + "spans": [ + { + "bbox": [ + 347, + 70, + 525, + 131 + ], + "type": "table", + "html": "
Insertion LayerTop-1 (%)Top-5 (%)
[4]73.2793.96
[6]73.9692.41
[8]75.0095.68
[10]74.1393.10
", + "image_path": "ce463d6cc28215eb1d610d8e4fdd506168cfd14d7049ce8caddcc6066f42d36a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 140, + 553, + 161 + ], + "lines": [ + { + "bbox": [ + 313, + 140, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 313, + 140, + 553, + 161 + ], + "type": "text", + "content": "Table 5. Different insertion layer of token selection strategy on the SeAct dataset." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 359, + 182, + 504, + 308 + ], + "blocks": [ + { + "bbox": [ + 359, + 182, + 504, + 308 + ], + "lines": [ + { + "bbox": [ + 359, + 182, + 504, + 308 + ], + "spans": [ + { + "bbox": [ + 359, + 182, + 504, + 308 + ], + "type": "image", + "image_path": "0b9b8ad542c1b4adb41fc6a2298fcaa12c0d35769d7fa6cf5f1ad4044b3e82e4.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 320, + 555, + 342 + ], + "lines": [ + { + "bbox": [ + 313, + 320, + 555, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 320, + 555, + 342 + ], + "type": "text", + "content": "Figure 3. Effect of different values of " + }, + { + "bbox": [ + 313, + 320, + 555, + 342 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 320, + 555, + 342 + ], + "type": "text", + "content": " on SeAct and DVS 128 Gesture datasets." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 364, + 555, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 364, + 555, + 435 + ], + "spans": [ + { + "bbox": [ + 313, + 364, + 555, + 435 + ], + "type": "text", + "content": "vidual layers and observe that shallow and deep layers yield marginal performance gains. In contrast, layer 8 achieves peak Top-1, validating that appropriately positioned layers within vision encoders optimally encode cross-modal feature representations, thereby enabling effective modality-guided selection of semantically consistent tokens." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": "Balance Between Consistency Loss and Recognition Loss. Our empirical validation substantiates that achieving equilibrium between modality consistency loss and contractive recognition (Eq. 11) loss constitutes a pivotal design criterion for robust multimodal learning. The contribution of " + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{r}}" + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": " is modified via " + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": ". As shown in Fig. 3, the variation of " + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "inline_equation", + "content": "\\lambda" + }, + { + "bbox": [ + 313, + 437, + 556, + 581 + ], + "type": "text", + "content": " affects the accuracy performance. The best performance is achieved with weights of 0.6, 0.4 on SeAct and DVS128 Gesture datasets respectively. It is worth noting that our findings highlight the potential benefits of leveraging consistency across modalities to enhance the representational capacity and recognition robustness of our EMP." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 432, + 601 + ], + "type": "text", + "content": "4.5. Qualitative Analysis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Visualization of Cross-modal Frequency Enhancer (CFE). To investigate the efficacy of the designed CFE, we visualize the four variants of event frames on the SeAct dataset: raw stacked frames, reconstructed frames, enhanced frames by sharpening operators, and enhanced frames by our CFE. As shown in fig. 4, conventional edge-enhancement techniques [33] yield marginal improvements over raw reconstructed frames. In contrast, event frames processed by CFE consistently highlight the critical regions" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5975" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 78, + 72, + 286, + 287 + ], + "blocks": [ + { + "bbox": [ + 78, + 72, + 286, + 287 + ], + "lines": [ + { + "bbox": [ + 78, + 72, + 286, + 287 + ], + "spans": [ + { + "bbox": [ + 78, + 72, + 286, + 287 + ], + "type": "image", + "image_path": "5a05f721c0097c87947867fdae625e46254ecd1a8ad66f8911cea2ebec49b5fe.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 296, + 295, + 352 + ], + "lines": [ + { + "bbox": [ + 55, + 296, + 295, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 296, + 295, + 352 + ], + "type": "text", + "content": "Figure 4. Visualization of event frames. \"Stack\" and \"Reconstruct\" denote stacked and reconstructed frames, respectively; \"Edge Enhance\" indicates enhanced reconstructed frames employing sharpening operators; \"CFE\" represents enhanced reconstructed frames by our Cross-modal Frequency Enhancer." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 375, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 295, + 460 + ], + "type": "text", + "content": "across both individual actions and interaction scenarios. Our qualitative analysis confirms that the proposed CFE selectively refines the high-frequency components of the reconstructed frames through spectral coordination with the edge dynamics of the stacked frames, thereby generating complementary event representations with synergistic spatial characteristics." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 460, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 616 + ], + "type": "text", + "content": "Visualization of Guidanced Selection. Fig. 5 visualizes the important tokens selected by the High-frequency Guided Selector (HGS). As shown in Fig. 5, the HGS demonstrates the adaptive selection of action-salient tokens across sequential event frames. This demonstrates the intrinsic consistency among multi-modal frames originating from the identical event stream, where HGS effectively identifies semantically coherent tokens while suppressing interference from spatiotemporal redundancies. The results highlight the ability of HGS to focus on critical regions of interest, thereby enhancing the discriminative power of EMP. This selection is crucial for improving the robustness and accuracy of action recognition in dynamic scenes." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 714 + ], + "type": "text", + "content": "Visualization of Event Features Learned by EMP. As illustrated in Fig. 6, we visualize the feature distributions of our proposed EMP against the baseline on the SeAct dataset. Seven action classes are randomly selected to demonstrate inter-class feature separability. Empirical results demonstrate the capability of our EMP in achieving both intra-class compactness and inter-class dispersion, further substantiating its efficacy in preserving semantic dis" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 332, + 72, + 535, + 223 + ], + "blocks": [ + { + "bbox": [ + 332, + 72, + 535, + 223 + ], + "lines": [ + { + "bbox": [ + 332, + 72, + 535, + 223 + ], + "spans": [ + { + "bbox": [ + 332, + 72, + 535, + 223 + ], + "type": "image", + "image_path": "04aff77dad66b914df2ec9e5e3ff7b8500b256e54e47ba40471d68f0758ffdc0.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 231, + 555, + 255 + ], + "lines": [ + { + "bbox": [ + 313, + 231, + 555, + 255 + ], + "spans": [ + { + "bbox": [ + 313, + 231, + 555, + 255 + ], + "type": "text", + "content": "Figure 5. Visualization of cross-modal consistent tokens selected by the HGS, with unselected regions masked in white." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 331, + 272, + 541, + 371 + ], + "blocks": [ + { + "bbox": [ + 331, + 272, + 541, + 371 + ], + "lines": [ + { + "bbox": [ + 331, + 272, + 541, + 371 + ], + "spans": [ + { + "bbox": [ + 331, + 272, + 541, + 371 + ], + "type": "image", + "image_path": "4dfdd280301d7d5eb199bfadddda1c4fa644873a1f55848e0ab5ff538c7e00e4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "lines": [ + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 553, + 403 + ], + "type": "text", + "content": "Figure 6. The t-SNE visualization of 7 randomly selected actions. Different color indicates different actions." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 425, + 424, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 425, + 424, + 435 + ], + "spans": [ + { + "bbox": [ + 314, + 425, + 424, + 435 + ], + "type": "text", + "content": "tinctiveness across classes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 448, + 466, + 460 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 466, + 460 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 466, + 460 + ], + "type": "text", + "content": "5. Conclusion and Discussion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 468, + 555, + 647 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 555, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 555, + 647 + ], + "type": "text", + "content": "Conclusion. In this work, we propose a privacy-preserving paradigm that eliminates RGB dependencies by generating texture-enriched reconstructed frames directly from raw event streams, reducing multi-sensor acquisition costs while bridging the modality gap with event frames. To preserve diverse information across reconstructed and stacked frames while eliminating the influence of irrelevant information, we propose our Enhanced Multimodal Perceptual (EMP) framework for robust multi-modal unified perception. It consists of Cross-modal Frequency Enhancer (CFE) and High-frequency Guided Selector (HGS), which enables the progressive exploitation of intrinsic multi-modal cues from the representation level to the feature level to acquire discriminative event features. Extensive experimental results validate the effectiveness of the proposed method." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 647, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 647, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 647, + 554, + 712 + ], + "type": "text", + "content": "Discussion. While our method achieves promising performance in EAR and demonstrates the potential for scalability, its ability to handle more challenging scenarios remains to be explored. Future work will focus on developing a robust multi-modal unified perceptual approach to cope with diverse and variable scenes." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5976" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 73, + 297, + 128 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 73, + 297, + 128 + ], + "spans": [ + { + "bbox": [ + 55, + 73, + 297, + 128 + ], + "type": "text", + "content": "Acknowledgements. The work is supported by the National Natural Science Foundation of China (Grant No. 62222207, 62427808, 62472208), and the Nanjing University Integrated Research Platform of the Ministry of Education-Top Talents Program (Grant No. 2024300482)." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 138, + 115, + 150 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 158, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 158, + 296, + 224 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 224 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 224 + ], + "type": "text", + "content": "[1] Arnon Amir, Brian Taba, David Berg, Timothy Melano, Jeffrey McKinstry, Carmelo Di Nolfo, Tapan Nayak, Alexander Andreopoulos, Guillaume Garreau, Marcela Mendoza, et al. A low power, fully event-based gesture recognition system. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 7243-7252, 2017. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 225, + 296, + 278 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 225, + 296, + 278 + ], + "spans": [ + { + "bbox": [ + 61, + 225, + 296, + 278 + ], + "type": "text", + "content": "[2] Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lucic, and Cordelia Schmid. Vivit: A video vision transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 6836-6846, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "spans": [ + { + "bbox": [ + 61, + 280, + 296, + 335 + ], + "type": "text", + "content": "[3] Sami Barchid, Benjamin Allaert, Amel Aissaoui, José Mennesson, and Chaabane C Djeraba. Spiking-fer: spiking neural network for facial expression recognition with event cameras. In Proceedings of the 20th International Conference on Content-based Multimedia Indexing, pages 1–7, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 336, + 296, + 378 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 336, + 296, + 378 + ], + "spans": [ + { + "bbox": [ + 62, + 336, + 296, + 378 + ], + "type": "text", + "content": "[4] Gedas Bertasius, Heng Wang, and Lorenzo Torresani. Is space-time attention all you need for video understanding? In International Conference on Machine Learning, page 4, 2021. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 296, + 423 + ], + "type": "text", + "content": "[5] Marco Cannici and Davide Scaramuzza. Mitigating motion blur in neural radiance fields with events and frames. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9286-9296, 2024. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 296, + 479 + ], + "type": "text", + "content": "[6] Meiqi Cao, Rui Yan, Xiangbo Shu, Jiachao Zhang, Jinpeng Wang, and Guo-Sen Xie. Mup: multi-granularity unified perception for panoramic activity recognition. In Proceedings of the 31st ACM International Conference on Multimedia, pages 7666-7675, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "spans": [ + { + "bbox": [ + 62, + 480, + 296, + 524 + ], + "type": "text", + "content": "[7] Meiqi Cao, Xiangbo Shu, Jiachao Zhang, Rui Yan, Zechao Li, and Jinhui Tang. Eventcrab: Harnessing frame and point synergy for event-based action recognition and beyond. arXiv preprint arXiv:2411.18328, 2024. 2, 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 525, + 296, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 525, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 525, + 296, + 579 + ], + "type": "text", + "content": "[8] Jiaqi Chen, Yan Yang, Shizhuo Deng, Da Teng, and Liyuan Pan. Spikmamba: When snn meets mamba in event-based human action recognition. In Proceedings of the 6th ACM International Conference on Multimedia in Asia, pages 1-8, 2024. 2, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 296, + 635 + ], + "type": "text", + "content": "[9] Lan Chen, Haoxiang Yang, Pengpeng Shao, Haoyu Song, Xiao Wang, Zhicheng Zhao, Yaowei Wang, and Yonghong Tian. Velora: A low-rank adaptation approach for efficient rgb-event based recognition. arXiv preprint arXiv:2412.20064, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 296, + 691 + ], + "type": "text", + "content": "[10] Tristan de Blegiers, Ishan Rajendrakumar Dave, Adeel Yousaf, and Mubarak Shah. Eventtransact: A video transformer-based framework for event-camera based action recognition. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 1-7, 2023. 2, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 692, + 296, + 713 + ], + "type": "text", + "content": "[11] Yongjian Deng, Hao Chen, and Youfu Li. A dynamic gcn with cross-representation distillation for event-based learn" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "ing. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1492-1500, 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 555, + 150 + ], + "type": "text", + "content": "[12] Shuangrui Ding, Peisen Zhao, Xiaopeng Zhang, Rui Qian, Hongkai Xiong, and Qi Tian. Prune spatio-temporal tokens by semantic-aware temporal accumulation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 16945-16956, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 554, + 207 + ], + "type": "text", + "content": "[13] Guillermo Gallego, Tobi Delbrück, Garrick Orchard, Chiara Bartolozzi, Brian Taba, Andrea Censi, Stefan Leutenegger, Andrew J Davison, Jörg Conradt, Kostas Daniilidis, et al. Event-based vision: A survey. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(1):154-180, 2020. 1" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 554, + 252 + ], + "type": "text", + "content": "[14] Yue Gao, Jiaxuan Lu, Siqi Li, Nan Ma, Shaoyi Du, Yipeng Li, and Qionghai Dai. Action recognition and benchmark using event cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(12):14081-14097, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 254, + 554, + 308 + ], + "type": "text", + "content": "[15] Daniel Gehrig, Mathias Gehrig, Javier Hidalgo-Carrió, and Davide Scaramuzza. Video to events: Recycling video datasets for event cameras. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3586-3595, 2020. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 554, + 354 + ], + "type": "text", + "content": "[16] Mathias Gehrig and Davide Scaramuzza. Recurrent vision transformers for object detection with event cameras. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 13884-13893, 2023. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 554, + 398 + ], + "type": "text", + "content": "[17] Kai Han, An Xiao, Enhua Wu, Jianyuan Guo, Chunjing Xu, and Yunhe Wang. Transformer in transformer. Advances in Neural Information Processing Systems, 34:15908-15919, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 453 + ], + "type": "text", + "content": "[18] Ju He, Jie-Neng Chen, Shuai Liu, Adam Kortylewski, Cheng Yang, Yutong Bai, and Changhu Wang. Transfg: A transformer architecture for fine-grained recognition. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 852-860, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 455, + 554, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 455, + 554, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 455, + 554, + 499 + ], + "type": "text", + "content": "[19] Bo Jiang, Chengguo Yuan, Xiao Wang, Zhimin Bao, Lin Zhu, Yonghong Tian, and Jin Tang. Point-voxel absorbing graph representation learning for event stream based recognition. arXiv preprint arXiv:2306.05239, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 500, + 555, + 554 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 500, + 555, + 554 + ], + "spans": [ + { + "bbox": [ + 316, + 500, + 555, + 554 + ], + "type": "text", + "content": "[20] Xin Jiang, Hao Tang, Junyao Gao, Xiaoyu Du, Shengfeng He, and Zechao Li. Delving into multimodal prompting for fine-grained visual classification. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2570-2578, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 554, + 600 + ], + "type": "text", + "content": "[21] Xin Jiang, Hao Tang, and Zechao Li. Global meets local: Dual activation hashing network for large-scale fine-grained image retrieval. IEEE Transactions on Knowledge and Data Engineering, 2024. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 601, + 554, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 601, + 554, + 655 + ], + "spans": [ + { + "bbox": [ + 316, + 601, + 554, + 655 + ], + "type": "text", + "content": "[22] Xin Jiang, Hao Tang, Rui Yan, Jinhui Tang, and Zechao Li. Dvf: Advancing robust and accurate fine-grained image retrieval with retrieval guidelines. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2379-2388, 2024. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 657, + 554, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 554, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 554, + 689 + ], + "type": "text", + "content": "[23] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.5" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 690, + 554, + 713 + ], + "type": "text", + "content": "[24] Dong Li, Jiandong Jin, Yuhao Zhang, Yanlin Zhong, Yaoyang Wu, Lan Chen, Xiao Wang, and Bin Luo." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5977" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 295, + 106 + ], + "type": "text", + "content": "Semantic-aware frame-event fusion based pattern recognition via large vision-language models. arXiv preprint arXiv:2311.18592, 2023. 2, 3, 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 295, + 150 + ], + "type": "text", + "content": "[25] Pengpeng Li, Xiangbo Shu, Chun-Mei Feng, Yifei Feng, Wangmeng Zuo, and Jinhui Tang. Surgical video workflow analysis via visual-language learning. npj Health Systems, 2 (1):5, 2025. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 195 + ], + "type": "text", + "content": "[26] Bingde Liu, Chang Xu, Wen Yang, Huai Yu, and Lei Yu. Motion robust high-speed light-weighted object detection with event camera. IEEE Transactions on Instrumentation and Measurement, 72:1-13, 2023. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 294, + 251 + ], + "type": "text", + "content": "[27] Haoyue Liu, Shihan Peng, Lin Zhu, Yi Chang, Hanyu Zhou, and Luxin Yan. Seeing motion at nighttime with an event camera. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 25648-25658, 2024. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 297 + ], + "type": "text", + "content": "[28] Qianhui Liu, Dong Xing, Huajin Tang, De Ma, and Gang Pan. Event-based action recognition using motion information and spiking neural networks. In International Joint Conference on Artificial Intelligence, pages 1743-1749, 2021. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 331 + ], + "type": "text", + "content": "[29] Shuangbiao Liu, Qian Wang, and Geng Liu. A versatile method of discrete convolution and fft (dc-fft) for contact analyses. Wear, 243(1-2):101-111, 2000. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 333, + 294, + 376 + ], + "type": "text", + "content": "[30] Shu Miao, Guang Chen, Xiangyu Ning, Yang Zi, Kejia Ren, Zhenshan Bing, and Alois Knoll. Neuromorphic vision datasets for pedestrian detection, action recognition, and fall detection. Frontiers in Neurorobotics, 13:38, 2019. 5" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 432 + ], + "type": "text", + "content": "[31] Federico Paredes-Valles and Guido CHE De Croon. Back to event basics: Self-supervised learning of image reconstruction for event cameras via photometric constancy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3446–3455, 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 434, + 295, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 295, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 295, + 498 + ], + "type": "text", + "content": "[32] Chiara Plizzari, Mirco Planamente, Gabriele Goletto, Marco Cannici, Emanuele Gusso, Matteo Matteucci, and Barbara Caputo. E2 (go) motion: Motion augmented event stream for egocentric action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 19935-19947, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 533 + ], + "type": "text", + "content": "[33] R Pushpavalli and G Sivaradje. Switching median filter for image enhancement. International Journal of Scientific & Engineering Research, 3(2):1-5, 2012. 7" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[34] Hongyu Qu, Rui Yan, Xiangbo Shu, Hailiang Gao, Peng Huang, and Guo-Sen Xie. Mvp-shot: Multi-velocity progressive-alignment framework for few-shot action recognition. arXiv preprint arXiv:2405.02077, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 624 + ], + "type": "text", + "content": "[35] Hongyu Qu, Jianan Wei, Xiangbo Shu, and Wenguan Wang. Learning clustering-based prototypes for compositional zero-shot learning. In Proceedings of the International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 677 + ], + "type": "text", + "content": "[36] Yongming Rao, Wenliang Zhao, Benlin Liu, Jiwen Lu, Jie Zhou, and Cho-Jui Hsieh. Dynamicvit: Efficient vision transformers with dynamic token sparsification. Advances in Neural Information Processing Systems, 34:13937-13949, 2021. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 681, + 294, + 713 + ], + "type": "text", + "content": "[37] Hongwei Ren, Yue Zhou, Yulong Huang, Haotian Fu, Xiaopeng Lin, Jie Song, and Bojun Cheng. Spikepoint: An efficient point-based spiking neural network for event cam" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "eras action recognition. arXiv preprint arXiv:2310.07189, 2023. 2, 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 162 + ], + "type": "text", + "content": "[38] Hongwei Ren, Yue Zhou, Jiadong Zhu, Xiaopeng Lin, Hao-tian Fu, Yulong Huang, Yuetong Fang, Fei Ma, Hao Yu, and Bojun Cheng. Rethinking efficient and effective point-based networks for event camera classification and regression. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "text", + "content": "[39] Alberto Sabater, Luis Montesano, and Ana C Murillo. Event transformer. a sparse-aware solution for efficient event data processing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2677-2686, 2022. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 555, + 274 + ], + "type": "text", + "content": "[40] Cedric Scheerlinck, Henri Rebecq, Daniel Gehrig, Nick Barnes, Robert Mahony, and Davide Scaramuzza. Fast image reconstruction with an event camera. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 156-163, 2020. 1, 2, 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[41] Sumit B Shrestha and Garrick Orchard. Slayer: Spike layer error reassignment in time. Advances in Neural Information Processing Systems, 31, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 364 + ], + "type": "text", + "content": "[42] Xiangbo Shu, Binqian Xu, Liyan Zhang, and Jinhui Tang. Multi-granularity anchor-contrastive representation learning for semi-supervised skeleton-based action recognition. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(6):7559-7576, 2022. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 410 + ], + "type": "text", + "content": "[43] Xiangbo Shu, Jiawen Yang, Rui Yan, and Yan Song. Expansion-squeeze-excitation fusion network for elderly activity recognition. IEEE Transactions on Circuits and Systems for Video Technology, 32(8):5281-5292, 2022. 1" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 412, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 464 + ], + "type": "text", + "content": "[44] Zehua Sun, Qiuhong Ke, Hossein Rahmani, Mohammed Bennamoun, Gang Wang, and Jun Liu. Human action recognition from various data modalities: A review. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3): 3200-3225, 2022. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 468, + 553, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 520 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 520 + ], + "type": "text", + "content": "[45] Du Tran, Lubomir Bourdev, Rob Fergus, Lorenzo Torresani, and Manohar Paluri. Learning spatiotemporal features with 3d convolutional networks. In Proceedings of the IEEE International Conference on Computer Vision, pages 4489-4497, 2015. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 523, + 553, + 578 + ], + "type": "text", + "content": "[46] Du Tran, Heng Wang, Lorenzo Torresani, Jamie Ray, Yann LeCun, and Manohar Paluri. A closer look at spatiotemporal convolutions for action recognition. In Proceedings of the IEEE/CVF conference on Computer Vision and Pattern Recognition, pages 6450-6459, 2018. 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 634 + ], + "type": "text", + "content": "[47] Hongjie Wang, Bhishma Dedhia, and Niraj K Jha. Zero-prune: Zero-shot token pruning through leveraging of the attention graph in pre-trained transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16070-16079, 2024. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 689 + ], + "type": "text", + "content": "[48] Xiao Wang, Zongzhen Wu, Yao Rong, Lin Zhu, Bo Jiang, Jin Tang, and Yonghong Tian. Sstformer: Bridging spiking neural network and memory support transformer for frame-event based recognition. arXiv preprint arXiv:2308.04369, 2023. 2, 3, 4, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[49] Xiao Wang, Shiao Wang, Pengpeng Shao, Bo Jiang, Lin Zhu, and Yonghong Tian. Event stream based human action" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5978" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 508 + ], + "type": "list", + "angle": 0, + "index": 9, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "recognition: a high-definition benchmark dataset and algorithms. arXiv preprint arXiv:2408.09764, 2024. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 150 + ], + "type": "text", + "content": "[50] Xiao Wang, Zongzhen Wu, Bo Jiang, Zhimin Bao, Lin Zhu, Guoqi Li, Yaowei Wang, and Yonghong Tian. Hardvs: Revisiting human activity recognition with dynamic vision sensors. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 5615-5623, 2024. 2, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 294, + 196 + ], + "type": "text", + "content": "[51] Ziqing Wang, Yuetong Fang, Jiahang Cao, Qiang Zhang, Zhongrui Wang, and Renjing Xu. Masked spiking transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 1761-1771, 2023. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "spans": [ + { + "bbox": [ + 56, + 197, + 294, + 251 + ], + "type": "text", + "content": "[52] Ziming Wang, Ziling Wang, Huaning Li, Lang Qin, Runhao Jiang, De Ma, and Huajin Tang. Eas-snn: End-to-end adaptive sampling and representation for event-based detection with recurrent spiking neural networks. In European Conference on Computer Vision, pages 310-328, 2024. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 297 + ], + "type": "text", + "content": "[53] Wenming Weng, Yueyi Zhang, and Zhiwei Xiong. Event-based video reconstruction using transformer. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2563-2572, 2021. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 342 + ], + "type": "text", + "content": "[54] Bochen Xie, Yongjian Deng, Zhanpeng Shao, Hai Liu, and Youfu Li. Vmv-gcn: Volumetric multi-view based graph cnn for event stream classification. IEEE Robotics and Automation Letters, 7(2):1976-1983, 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 342, + 295, + 396 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 295, + 396 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 295, + 396 + ], + "type": "text", + "content": "[55] Pingping Zhang, Yuhao Wang, Yang Liu, Zhengzheng Tu, and Huchuan Lu. Magic tokens: Select diverse tokens for multi-modal object re-identification. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17117-17126, 2024. 3, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 398, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 398, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 398, + 294, + 442 + ], + "type": "text", + "content": "[56] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Eventbind: Learning a unified representation to bind them all for event-based open-world understanding. In European Conference on Computer Vision, pages 477-494, 2024. 4, 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 443, + 294, + 508 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 443, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 443, + 294, + 508 + ], + "type": "text", + "content": "[57] Jiazhou Zhou, Xu Zheng, Yuanhuiyi Lyu, and Lin Wang. Exact: Language-guided conceptual reasoning and uncertainty estimation for event-based action recognition and more. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18633-18643, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 8 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5979" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_content_list.json b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..237a0f3ff9cd8f30b7cc0319385dfa54c81aafbb --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_content_list.json @@ -0,0 +1,1470 @@ +[ + { + "type": "text", + "text": "Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation", + "text_level": 1, + "bbox": [ + 122, + 128, + 875, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Tiankai Chen $^{1}$ , Yushu Li $^{23}$ , Adam Goodge $^{3}$ , Fei Teng $^{1}$ , Xulei Yang $^{3}$ , Tianrui Li $^{1}$ , Xun Xu $^{\\dagger 3}$ $^{1}$ Southwest Jiaotong University, $^{2}$ South China University of Technology \n $^{3}$ Institute for infocomm research(I $^{2}$ R), A*STAR", + "bbox": [ + 135, + 202, + 861, + 256 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Out-of-distribution (OOD) detection in 3D point cloud data remains a challenge, particularly in applications where safe and robust perception is critical. While existing OOD detection methods have shown progress for 2D image data, extending these to 3D environments involves unique obstacles. This paper introduces a training-free framework that leverages Vision-Language Models (VLMs) for effective OOD detection in 3D point clouds. By constructing a graph based on class prototypes and testing data, we exploit the data manifold structure to enhancing the effectiveness of VLMs for 3D OOD detection. We propose a novel Graph Score Propagation (GSP) method that incorporates prompt clustering and self-training negative prompting to improve OOD scoring with VLM. Our method is also adaptable to few-shot scenarios, providing options for practical applications. We demonstrate that GSP consistently outperforms state-of-the-art methods across synthetic and real-world datasets for 3D point cloud OOD detection.", + "bbox": [ + 89, + 323, + 485, + 597 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 628, + 220, + 643 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The detection of out-of-distribution (OOD) samples in 3D point cloud data is critical for the safe and robust deployment of deep learning models in real-world environments. For example, identifying unexpected objects is crucial for ensuring the safety of personnel, vehicles, and robots by preventing collisions [18, 39, 41]. Existing approaches to OOD detection in 3D point clouds often focus on training separate models for specific tasks [54]. These methods typically rely on abundant in-distribution (ID) training data to build models based on reconstruction [34], distance metrics [10], or discriminative strategies [15]. While these methods achieve strong results, they depend on substantial ID training data, which is often challenging to collect in many real-world applications and generalization may be", + "bbox": [ + 89, + 654, + 483, + 867 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd65801fa754b791ecc7ed43d2b422ae763933e9b0384fb59d3c4e4d8e329d58.jpg", + "image_caption": [ + "Figure 1. Exploiting the testing data manifold for zero-/few-shot OOD detection. ID prototypes propagate scores to test samples, with confident ones selected as pseudo negative prompts for refinement." + ], + "image_footnote": [], + "bbox": [ + 517, + 292, + 898, + 521 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "hindered by the limited diversity of ID training samples. These limitations highlight the need for alternative solutions that do not require extensive ID data collection.", + "bbox": [ + 511, + 595, + 906, + 641 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Vision-language models (VLMs) have emerged as powerful foundation models that bridge visual features with textual understanding [23, 28, 42, 58]. VLMs are trained to associate the content of visual inputs with their textual labels and they demonstrate strong zero-shot generalization to unseen visual content [42]. This capacity makes VLMs appealing for OOD detection without retraining [5, 36, 37, 51]. A common VLM-based OOD detection pipeline maps ID class concepts, such as class names with various templates, into a shared feature space using the text encoder. The visual samples, similarly projected using the visual encoder, are then scored by comparing their distances to the ID class concepts in this feature space [36].", + "bbox": [ + 509, + 642, + 908, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "This VLM-based OOD detection pipeline has shown competitive performance for 2D image detection, primarily due to the vast understanding of 2D VLMs from the massive-scale image-text pre-training datasets (e.g. 400M", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\dagger$ Correspondence to Xun Xu: xu_xun@i2ra.a-star.edu.sg, Code link: https://github.com/handsome999KK/GSP_OOD", + "bbox": [ + 102, + 878, + 457, + 904 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "28797", + "bbox": [ + 478, + 946, + 519, + 958 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "pairs for pre-training CLIP [42]). In contrast, VLMs for 3D point clouds, such as the state-of-the-art ULIP2 [55], are typically pre-trained on much more restricted datasets (e.g., 800k point cloud samples). This restricts their ability to generalize across unseen categories and distributional shifts (e.g., style changes), which reduces the effectiveness of similarity-based methods for 3D point cloud OOD detection.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose to leverage the structure of the testing data manifold to compensate for this deficiency and improve VLM adaptation for downstream tasks [31, 44], as illustrated in Fig. 1. Specifically, we construct a graph based on class prototypes (e.g., text prompts for ID classes) and testing data features, where prototype nodes are assigned labels that propagate to neighboring testing nodes. Given the limited training data, we hypothesize that exploiting the testing data manifold enhances the generalization capabilities of 3D point cloud VLMs. Our method, Graph Score Propagation (GSP), assigns a positive OOD score (e.g., +1) to ID class prototypes while all testing samples start with a neutral score (0). As scores propagate along the graph, the resultant scores indicate OOD likelihood.", + "bbox": [ + 89, + 213, + 480, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "This straightforward adoption of score propagation significantly improves upon existing distance-based methods, and we introduce two additional enhancements to further boost OOD detection accuracy. First, while conventional training-free approaches typically use a single class prototype by averaging text prompts [44, 61], we argue that capturing the prompt distribution through clustering better leverages prompt diversity. Second, negative prompting has shown strong performance in VLM-based OOD detection [30, 38, 48]. However, simply negating the description (e.g. add \"not\") is not very effective and learning negative prompts requires sophisticated optimization. Instead, we employ self-training to convert high-confidence OOD testing samples into pseudo-negative prompts and assign them a negative score (e.g., -1), for subsequent propagation iterations. Finally, our framework integrates seamlessly into traditional supervised OOD detection protocols by incorporating a few ID samples as positive prompts, further enhancing OOD detection under both zero-shot and few-shot protocols.", + "bbox": [ + 89, + 426, + 482, + 726 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are mainly summarized as follows.", + "bbox": [ + 109, + 729, + 460, + 744 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We employ a pre-trained vision-language model for 3D point cloud OOD detection and introduce Graph Score Propagation (GSP), a novel methodology that leverages the testing data manifold instead of solely relying on text-test feature distances.", + "- We enhance GSP with prompt clustering and self-training-based negative prompting. GSP is also compatible with both zero-shot and few-shot testing protocols.", + "- We validate the effectiveness of our method on synthetic and real-world 3D point cloud OOD detection bench" + ], + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "marks, consistently achieving superior performance over state-of-the-art VLM-based OOD detection methods.", + "bbox": [ + 527, + 90, + 906, + 119 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 135, + 653, + 150 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Out-of-Distribution Detection: Traditional OOD detection methods focused on softmax-based probability scoring [15]. Recently, attention has shifted to novel evaluation metrics [26, 47, 57, 62], such as differential network channel processing [57, 62] and non-parametric nearest-neighbor distances [47]. Other strategies involve training specific OOD discriminators, like DAL [52], which constructs an OOD distribution set, and MOOD [29], which uses masked image modeling. NECO [2] utilizes neural collapse geometry for OOD identification. Vision-language models have introduced new methods using text prompts to distinguish between in-distribution (ID) and OOD data [5, 30, 37, 38, 51]. LoCoOP [37] leverages CLIP's local features for OOD regularization, while Bai et al. [5] use CLIP to detect ID-like outliers, enhancing OOD detection and few-shot learning [59]. Training text encoders [51] and learnable prompt [30, 38] further increases the score gap between ID and OOD data. Recent research has expanded to include 3D datasets for point cloud OOD detection [1, 45]. Studies have shown that 3D OOD detection is particularly important for autonomous driving scenarios [18, 27].", + "bbox": [ + 511, + 160, + 903, + 477 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3D Vision-Language Models: Recent advancements in Vision-Language Models (VLMs) have facilitated innovative approaches for 3D understanding. Techniques utilizing CLIP [20, 35, 60] transform 3D point clouds into imagery data for extracting features aligned with text descriptions. More recently, methods like ULIP [54, 55] learn unified representations across images, text, and 3D point clouds, further improving zero-shot 3D classification. These approaches leverage the integration of text prompts and visual features which could better distinguish between indistribution and out-of-distribution data in 3D contexts.", + "bbox": [ + 511, + 478, + 903, + 643 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Label Propagation: Label propagation (LP) is a pivotal method in semi-supervised learning, leveraging limited labeled data to infer labels for unlabeled instances by exploiting data similarity across a graph [8, 14, 22, 40]. It assumes that labels vary smoothly across the graph, making it effective in transductive learning settings [22, 64]. Recent advancements integrate LP with graph neural networks to enhance its ability to separate ID from OOD data [6, 43, 53]. In the context of VLMs, LP employs text prompts as label representations, constructing graphs that propagate labels from text prototypes to test samples, thus improving zero-shot 2D classification [17, 31, 44]. Contrary to the existing attempts of combining LP with VLM, we tackle a one-class classification task. Only prototypes for in-distribution classes are available. We further propose a novel negative prompting technique via self-training and prompt clustering technique to better separate ID from OOD.", + "bbox": [ + 511, + 643, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "28798", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 89, + 225, + 107 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Problem Formulation", + "text_level": 1, + "bbox": [ + 89, + 114, + 294, + 130 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin by formally defining the task of VLM-based out-of-distribution (OOD) detection in 3D point clouds. Let $z^{f} = f(x;\\theta)$ and $z^{t} = g(t;\\phi)$ denote the encoded features of point cloud data $x$ and text $t$ , where $f(\\cdot ;\\theta)$ and $g(\\cdot ;\\phi)$ are the VLM visual and text encoders respectively. The task involves the encoded features of unlabeled test data $\\mathcal{D}_u = \\{f(x_i)\\}_{i = 1\\dots N_u}$ and, optionally, the features and labels of few-shot labeled data $\\mathcal{D}_l = \\{f(x_j),y_j\\}_{j = 1\\dots N_l}$ . The set of in-distribution (ID) classes $\\mathcal{Y}_{in} = \\{1,\\dots ,C_{in}\\}$ is assumed to be known. Following established practices for VLM adaptation [36], we represent each ID class with $N_{t}$ prompt templates $\\mathcal{T} = \\{t_j(y)\\}_{N_t}$ , e.g., \"This is a point cloud of $\\{\\mathrm{y}\\}$ \". \"A good point cloud of $\\{\\mathrm{y}\\}$ \". These textual prompts are encoded by the text encoder, creating prototypes for ID classes $\\mathcal{P}$ . Existing works often create prototypes by averaging the encoded prompts for each class, resulting in a prototype $z_c^p$ for the $c$ -th class, and the full set of prototypes is denoted $\\mathcal{P} = \\{z_c^p\\}_{c = 1\\dots C_{in}}$ . The OOD detection task is to determine whether a test sample $x_{i}\\in \\mathcal{D}_{u}$ belongs to an ID class, leveraging ID prototypes $\\mathcal{P}$ and, optionally, the few-shot labeled data $\\mathcal{D}_l$ .", + "bbox": [ + 89, + 137, + 483, + 455 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Preliminary of VLM: We briefly review the training objective of vision-language models (VLMs). The pre-training steps align the visual features with their associated text features [20, 54, 55], typically via a contrastive loss function:", + "bbox": [ + 89, + 455, + 483, + 517 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {v l m} = \\sum_ {i} \\log \\frac {\\exp \\left(z _ {i} ^ {f} z _ {i} ^ {t} / \\tau\\right)}{\\sum_ {j} \\exp \\left(z _ {i} ^ {f} z _ {j} ^ {t} / \\tau\\right)} \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 530, + 482, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "When a VLM is trained on a large amount of point cloud-text pairs, the cosine similarity of positive pairs (a point cloud sample and text prompt of its ground-truth class) is higher than that of negative pairs. Therefore, a pre-trained VLM provides a straightforward solution to 3D point cloud OOD detection by measuring the cosine similarity between textual descriptions of in-distribution classes and testing sample features. A high distance in the visual-text feature space suggests a greater likelihood of being OOD.", + "bbox": [ + 89, + 578, + 483, + 715 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. VLM for OOD Detection", + "text_level": 1, + "bbox": [ + 89, + 724, + 318, + 739 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vanilla OOD Detection with VLM: Pre-trained VLMs offer an effective approach to OOD detection by measuring the distance between visual samples and text labels in the feature space. For example, [36] employs the maximum normalized cosine similarity as the OOD score. We define a normalized score $S$ to indicate OOD-ness, where $S \\rightarrow 0$ suggests OOD and $S \\rightarrow 1$ suggests ID.", + "bbox": [ + 89, + 748, + 483, + 854 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS = \\max _ {c} \\frac {\\exp \\left(- d \\left(z ^ {f} , z _ {c} ^ {p}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(- d \\left(z ^ {f} , z _ {j} ^ {p}\\right) / \\tau\\right)} \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 867, + 482, + 905 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "It is assumed that in the feature space, OOD samples are further from the ID prototypes than ID samples. Subsequent research has further explored this approach by introducing techniques like negative prompts to better separate OOD from ID samples [24]. In this work, we are motivated by the challenge of measuring similarity/distance in the VLM feature space and propose to tackle the VLM OOD detection problem by exploiting the manifold of testing and, optionally, few-shot labeled data.", + "bbox": [ + 511, + 90, + 903, + 226 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Manifold Distance for OOD Detection: The vanilla measurement overlooks the complex manifold of testing data. A straightforward approach to consider testing data manifold for OOD detection is to replace the standard distance metric with a manifold distance, represented via a graph.", + "bbox": [ + 511, + 227, + 903, + 303 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a graph representing the data manifold $G = (E, V)$ with weight matrix $W \\in \\mathbb{R}^{(|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|) \\times (|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|)}$ , the manifold distance between two nodes on the graph is defined as the shortest path. The shortest path between arbitrary nodes can be obtained via Dijkstra's algorithm [13]. More concretely, this can be formulated as finding a path between a prototype $z^p \\in \\mathcal{P}$ or a few-shot labeled sample $z_k^f \\in \\mathcal{D}_l$ and a testing sample $z_j^f \\in \\mathcal{D}_u$ , consisting of edges $\\tilde{E}_{ij} = \\{e_k\\}$ on the graph, such that the sum of these edge weights is minimized. The manifold distance between two nodes is thus defined as $d(z_j, z_i) = \\min_{\\tilde{E}_{ij}} \\sum_{e_k \\in \\tilde{E}_{ij}} d(e_k)$ , where $d(e_k)$ denotes the distance for edge $e_k$ , e.g. the L2 distance between the encoded features.", + "bbox": [ + 511, + 304, + 905, + 518 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Our empirical results in Tab. 3 reveal that incorporating the manifold improves the effectiveness of OOD detection. However, there are two main disadvantages associated with distance-based OOD detection. First, the manifold distance is unnormalised, if directly using the reciprocal of manifold distance as OOD score the result may be biased by the relative scale. Additionally, it is non-trivial to integrate negative prompts within the distance-based OOD detection paradigm.", + "bbox": [ + 511, + 520, + 905, + 656 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. VLM OOD Detection as Score Propagation", + "text_level": 1, + "bbox": [ + 511, + 667, + 880, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of directly measuring the distance for OOD detection, we formulate OOD detection as a score propagation problem. Specifically, we assign a positive score $(+1)$ to ID prompts $\\mathcal{P}$ , while all testing samples are initially assigned a neutral score (0). Labels are then propagated from the positive nodes to all testing samples in an iterative manner. We follow a standard propagation rule [63] as shown in Eq. 3,", + "bbox": [ + 511, + 691, + 905, + 797 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS _ {t} = \\tilde {W} S _ {t - 1} + \\alpha S _ {0},\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 810, + 730, + 825 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\text {s . t .} \\tilde {W} = D ^ {- \\frac {1}{2}} W D ^ {- \\frac {1}{2}}, D = \\operatorname {d i a g} \\left(\\sum_ {j} W _ {i j}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 553, + 830, + 903, + 864 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nS_{0} = \\bigl[\\underbrace{1,\\cdots,1}_{|\\mathcal{P}| + |\\mathcal{D}_{l}| \\text{elements}},\\underbrace{0,\\cdots,0}_{|\\mathcal{D}_{u}| \\text{elements}}\\bigr],\n$$\n", + "text_format": "latex", + "bbox": [ + 588, + 867, + 815, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "28799", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/331e4e5e5dd294ddf6e887a4dbefd3b8fd4c66a668277220f12f5434e0979f36.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed framework for zero-/few-shot OOD detection via graph score propagation. The framework encodes ID category prompts and dataset samples using frozen text and 3D encoders, constructing a KNN graph. Initial score propagation generates pseudo prompts, which refine the graph before final score propagation computes OOD scores." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 897, + 239 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where each step of propagation updates the score for unlabeled testing samples as a weighted average of scores from adjacent nodes and initial scores $S_0$ until $T$ iterations are completed. Consequently, nodes that are closer to the prototypes on the manifold will obtain a higher score.", + "bbox": [ + 88, + 304, + 480, + 378 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Prompt Clustering: The above propagation strategy propagates labels from ID prototypes to testing samples on the graph. Thus, it is critical to design effective ID prototypes that capture the diversity of textual descriptions for ID classes. Contrary to existing practices that average the text prompt features across all prompt templates, we propose modeling the distribution of prompt templates with multiple representative ones. Given the prompt templates $\\mathcal{T}$ , we first initialize a pool of text prompts $\\mathcal{P}_c = \\{z_{c_j}^p = g(t_j(y_c))\\}_{j=1\\dots|\\mathcal{T}|}$ for the $c$ -th class. We then introduce a clustering step to reduce this set to $N_c$ representative prototypes $\\tilde{\\mathcal{P}}_c$ . This is achieved by applying K-means [21] clustering on $\\mathcal{P}_c$ , and the cluster centers are adopted as the ID prototypes:", + "bbox": [ + 89, + 378, + 483, + 592 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\mathcal {P}} _ {c} = \\arg \\min _ {\\{\\mu_ {k} \\}, \\gamma} \\sum_ {N _ {c}} \\sum_ {z _ {c j} ^ {p} \\in \\mathcal {P} _ {c}} \\gamma_ {j k} \\| z _ {c j} ^ {p} - \\mu_ {k} \\| ^ {2},\n$$\n", + "text_format": "latex", + "bbox": [ + 138, + 614, + 431, + 654 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l l} \\text {s . t .} & \\gamma \\in \\{0, 1 \\} ^ {| \\mathcal {T} | \\times N _ {c}}, \\sum_ {k} \\gamma_ {j k} = 1, \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 655, + 392, + 685 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mu_{k}$ are the cluster centers, and $\\gamma_{jk}$ indicates the assignment of each prompt feature $z_{cj}^{p}$ to its closest center. The new prototypes, $\\tilde{\\mathcal{P}} = \\tilde{\\mathcal{P}}_1\\cup \\tilde{\\mathcal{P}}_2\\cup \\dots \\tilde{\\mathcal{P}}_C$ , are used for score propagation as positive nodes.", + "bbox": [ + 89, + 686, + 482, + 748 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Negative Prompting via Self-Training: Negative prompts have been shown to substantially benefit OOD detection [30, 38, 48]. Incorporating negative prompts is particularly straightforward within the graph label propagation framework. We treat negative prompts as nodes $\\mathcal{P}^n$ with a label of $-1$ on the graph. During propagation, the negative labels compete with the positive labels (ID prompts) to influence the score of testing samples. Existing approach exploits ID data to learn negative prompt [48] which prohibits zero-shot inference. To address this, we propose a", + "bbox": [ + 89, + 750, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "strategy that generates negative prompts via self-training, as illustrated in Fig. 1. Specifically, we first conduct initial score propagation for $T$ iterations following Eq. 3, giving rise to initial propagated scores $S_{T}$ . We select the top $m\\%$ most confident scores from $S_{T}$ following Eq. 5 where $\\mathrm{top}(S,m\\%)$ indicates the top $m\\%$ value of $S$ .", + "bbox": [ + 511, + 304, + 906, + 395 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal{I}_{pos} = \\left\\{i\\mid S_{T}[i] > \\operatorname {top}(S_{T},m\\%)\\right\\} ,\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 406, + 815, + 422 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {I} _ {n e g} = \\left\\{i \\mid S _ {T} [ i ] < \\operatorname {top} \\left(S _ {T}, 1 - m \\%\\right) \\right\\} \\tag{5}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 422, + 903, + 443 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We label the selected samples as positive and negative prompts respectively and re-initialize the score $\\tilde{S}_0$ . For negative prompts, we assign scores of $-1$ to further increase the gap between ID and OOD samples. The final score is obtained by running Eq. 3 with $\\tilde{S}_0$ as initial scores.", + "bbox": [ + 511, + 446, + 906, + 522 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {S} _ {0} [ i ] = \\left\\{ \\begin{array}{l l} S _ {0} [ i ] & \\text {i f} i \\notin \\mathcal {I} _ {\\text {p o s}} \\& i \\notin \\mathcal {I} _ {\\text {n e g}}, \\\\ + 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {p o s}}, \\\\ - 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {n e g}}. \\end{array} \\right. \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 534, + 906, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Graph Construction: We now introduce the approach for constructing a graph for score propagation. As observed by [31, 44], the distance between text prompts and visual samples is generally larger than the distance between visual samples. This observation also holds for 3D point cloud VLM because the 3D VLMs are often finetuned from 2D VLMs [54, 55]. Therefore, we adopt a similar approach to construct the graph. We write the weight matrix $W$ in a blockwise form as shown in Eq. 7, where $W_{u} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{u}|}$ , $W_{up} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\tilde{\\mathcal{P}}|}$ and $W_{ul} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{l}|}$ . No connections are created between prototypes and few-shot labeled samples. All blocks are further constructed as a k-nearest neighbor (KNN) graph following [31, 44]. It is also computationally efficient to adapt an existing graph to new testing samples by incrementally adding the testing samples to the graph [31].", + "bbox": [ + 511, + 595, + 908, + 840 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nW = \\left[ \\begin{array}{c c c} \\mathbf {I} ^ {| \\tilde {\\mathcal {P}} | \\times | \\tilde {\\mathcal {P}} |} & \\mathbf {0} & W _ {u p} ^ {\\top} \\\\ \\mathbf {0} & \\mathbf {I} ^ {| \\mathcal {D} _ {l} | \\times | \\mathcal {D} _ {l} |} & W _ {u l} ^ {\\top} \\\\ W _ {u p} & W _ {u l} & W _ {u} \\end{array} \\right] \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 576, + 849, + 906, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "28800", + "bbox": [ + 478, + 946, + 519, + 957 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overall Algorithm: We present the overall algorithm for unified iterative label propagation for vision-language models in Algorithm 1. In Fig. 2, we show an overview of the proposed framework for zero-/few-shot OOD detection via graph propagation.", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [ + "Algorithm 1 Graph Score Propagation for OOD Detection" + ], + "code_body": "1: Input: training data $\\mathcal{D}_l$ , testing data $\\mathcal{D}_u$ and ID class prompt templates $\\mathcal{T}_c$ \n2: Output: Final OOD scores $\\tilde{S}_T$ \n3: for $\\mathcal{D}_l$ and $\\mathcal{T}_c$ , do \n Encode: $z_l = f(x_l)$ for training data and $\\mathcal{P}_c = \\{g(t_j(y_c))\\}$ for prompt templates. Prompt clustering: positive prototypes $\\tilde{\\mathcal{P}}$ (Eq. 4) \n4: end for \n5: for $\\mathcal{D}_u$ , do \n Encode: $z_i = f(x_i)$ for testing data \n Build graph to compute $W$ (Eq. 7) \n Initialize score $S_0^{(i)}$ (Eq. 3) \n Score propagate to compute $S_T$ (Eq. 3) \n Self-training to generate positive $\\mathcal{I}_{pos}$ and negative prompts $\\mathcal{I}_{neg}$ (Eq. 5) \n Updated initial scores $\\tilde{S}_0$ (Eq. 6) \n Score propagate to compute $\\tilde{S}_T$ (Eq. 3) \n6: end for \n7: return $\\tilde{S}_T$", + "bbox": [ + 99, + 199, + 483, + 541 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 566, + 223, + 583 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experimental setup", + "text_level": 1, + "bbox": [ + 89, + 590, + 277, + 607 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Dataset: In this work, we follow the 3DOS framework [1] to construct downstream datasets for evaluating our method. The 3DOS benchmark is based on two well-known 3D object datasets, ShapeNetCore [9] and ScanObjectNN [49]. ShapeNetCore includes 54 categories covering daily objects. ScanObjectNN is cropped out from ScanNet dataset featuring a real-world objects. It consists of 15 categories with 2,902 3D objects. To further validate our approach on real-world data, we introduce the Sydney Urban Object Dataset [11], nuScenes [7] and S3DIS [4]. Sydney Urban Object Dataset consists of LiDAR-scanned point clouds across 26 categories, totaling 631 unique instances. S3DIS dataset contains 6 large-scale indoor areas with 271 rooms annotated into 13 semantic categories. We cropped out individual objects using ground-truth segmentation masks, leading to 8931 objects. The nuScenes Dataset features point clouds captured by LiDAR sensor under autonomous driving scenario. We cropped out point clouds using ground-truth bounding boxes, leading to 2205", + "bbox": [ + 89, + 613, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "objects across 9 categories. Detailed visualization of each dataset is deferred to the supplementary.", + "bbox": [ + 511, + 90, + 906, + 121 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Protocol: Following the methodology outlined in 3DOS [1], we divide ShapeNetCore into three non-overlapping category sets, SN1, SN2, and SN3, referred to as the ShapeNetCore dataset. Each set contains 18 categories. For each OOD test scenario, one of these sets (SN1, SN2, or SN3) is treated as the known set of categories, while the remaining two sets are used as unknown categories, resulting in three distinct OOD test cases. Similarly, 3DOS defines three category sets, SR1, SR2, and SR3 on the ScanObjectNN dataset. Each SR-set includes 5 non-overlapping categories, with the other two sets acting as unknown categories in each scenario and call them as ScanObjNN. As for Sydney Urban Object Dataset, we categorize these objects into two groups: movable objects, e.g. \"truck\" \"pedestrian\" and non-movable objects, e.g. \"building\" and \"traffic sign\". Movable objects are treated as in-distribution (ID) classes, simulating a scenario in autonomous driving where object detection primarily focuses on movable objects. For consistency, we pre-process each point cloud sample by upsampling it to a uniform number of points. For S3DIS dataset, We retained instances containing more than 2048 original points, where each instance was randomly downsampled to retain exactly 2048 points. Then we set the foreground objects as ID and back-ground objects as OOD. For nuScences dataset, we have conducted preliminary screening on the original point cloud and retained instances with more than 200 original points and upsampling it to 1024 points. Then we define Four-wheeled vehicles (including car, bus, truck, construction_vehicle, trailer) as ID and the rest as OOD. And further details provided in the supplementary material.", + "bbox": [ + 511, + 123, + 906, + 592 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We use the pre-trained ULIP/ULIP2 [54, 55] model to extract point cloud features and text features. For the 3D encoder, we select PointBERT [56]. In constructing the KNN graph, we set a universal $\\mathrm{K}$ -value $(N_{k})$ of 10, connecting each text feature to its 10 nearest point cloud features. In Eq. 3, we set $\\alpha$ to 0.5 and the number of iterations $T$ to 5. The percentage of pseudo negative prompts, i.e. $m\\%$ in Sect. 3.3, is also set to $5\\%$ . Additionally, we use K-means clustering $(N_{c})$ to group the prompts for each category into 3 clusters.", + "bbox": [ + 511, + 594, + 908, + 744 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Competing Methods. We compare our method with several OOD detection approaches, including full-shot methods from 3DOS [1]. We also benchmark against recent zero-shot methods, such as DDCS [57], ZLaP[44], MCM [36], and NegLabel [24]. For zero-shot comparisons, all methods use the 3D encoder from [55]. Baseline results are also reported by directly using cosine similarity as the OOD score, with these methods referred to as ULIP and ULIP2, depending on the VLM used.", + "bbox": [ + 511, + 747, + 908, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate OOD detection perfor", + "bbox": [ + 511, + 885, + 903, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "28801", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/eedea123209e157080c5a34eb0412c095c69f1dc6364b96cf2538e5e981e5351.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodScanObjNNShapeNetCore
SR3AUROC↑FPR95↓SR2AUROC↑FPR95↓SR1AUROC↑FPR95↓AverageAUROC↑FPR95↓SN1AUROC↑FPR95↓SN2AUROC↑FPR95↓SN3AUROC↑FPR95↓AverageAUROC↑FPR95↓
Requiring training datasetMSP[15]88.167.380.684.073.780.380.877.274.382.880.078.189.752.281.371.0
MLS[50]89.453.883.473.176.475.383.067.472.080.883.964.189.840.581.961.8
ODIN[32]90.247.983.371.776.376.883.365.574.279.479.471.787.841.880.564.3
Energy[33]89.550.681.675.876.675.582.667.372.181.283.064.789.839.482.061.8
GradNorm[19]88.550.777.475.375.276.880.467.672.181.857.788.957.879.062.683.3
ReAct[46]90.348.982.475.875.477.682.767.473.779.489.652.195.027.286.152.9
NF88.047.780.668.275.681.481.465.881.572.571.178.091.049.681.266.7
OE+mixup[16]72.683.572.088.562.587.869.086.672.778.980.368.887.362.280.169.9
ARPL+CS[10]--------74.880.380.772.485.450.880.367.8
Cosine protoCE91.041.082.178.277.675.683.664.980.368.388.760.891.938.086.955.7
SupCon[25]--------80.975.583.568.285.145.183.262.9
SubArcface[12]87.161.378.976.973.781.479.973.279.081.282.960.389.132.883.758.1
GSP(1-shot)79.361.380.984.188.651.782.965.791.344.086.443.183.891.487.259.5
GSP(5-shot)80.754.890.147.289.140.286.647.495.619.294.622.197.012.595.717.9
GSP(full-shot)90.133.093.030.592.234.491.822.697.910.197.97.998.86.098.28.0
Zero-shotULIP[54]79.982.762.387.557.492.366.587.588.057.176.448.775.251.379.952.4
ULIP2[55]65.780.876.471.277.478.073.276.785.752.584.144.367.674.479.157.1
DDCS[57]63.390.562.480.064.496.263.488.961.186.766.886.577.276.268.483.1
MCM[36]52.092.362.391.381.380.465.288.085.151.683.246.166.475.878.257.8
NegLabel[24]65.785.566.891.367.488.266.688.360.687.880.678.688.048.376.471.6
ZLaP[44]70.181.767.084.767.883.668.383.388.252.872.366.477.490.079.369.7
GSP75.863.076.577.683.661.678.667.490.638.970.764.079.793.780.465.5
", + "bbox": [ + 93, + 88, + 906, + 335 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "mance using two widely adopted metrics: AUROC and FPR95. AUROC calculates the area under the receiver operating characteristic curve, with higher values indicating better performance. FPR95 measures the false positive rate at a true positive rate of $95\\%$ , with lower values reflecting improved performance.", + "bbox": [ + 89, + 391, + 482, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Results on OOD Detection", + "text_level": 1, + "bbox": [ + 89, + 497, + 328, + 511 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of ScanObjNN Results: From the results presented in Tab. 1, we derive several key insights. i) Our proposed method demonstrates highly competitive performance under the zero-shot protocol, achieving an improvement of over $5\\%$ in AUROC compared to ULIP2, which uses cosine distance directly as an OOD score. This substantial enhancement suggests that employing a manifold-based approach significantly enhances the differentiation between OOD samples and ID classes. ii) Our few-shot results are equally impressive, rivaling even fully supervised methodologies. Notably, with 5-shot training samples, our method attains an average AUROC of $86.6\\%$ , surpassing the best fully supervised method, Cosine proto, which scores $83.6\\%$ , by a margin of $3\\%$ . Furthermore, when integrating all available training samples into the graph, our approach achieves an outstanding average AUROC of $91.8\\%$ . iii) A noticeable $5\\%$ gap in average AUROC is observed between our zero-shot and 1-shot methods. This gap underscores the critical role of incorporating real in-distribution samples in effectively distinguishing OOD samples. Consequently, practitioners are advised to introduce at least a few ID samples for OOD detection, in addition to leveraging Vision-Language Models (VLMs).", + "bbox": [ + 89, + 521, + 482, + 867 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of ShapeNetCore Results: The results for ShapeNetCore, as shown in Tab. 1, lead us to several ob", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/e984d84464cd246816acd2246d20c361e32c5efd6dc619af1c8ed719db963861.jpg", + "table_caption": [ + "Table 1. Evaluation of OOD detection for ScanObjNN and ShapeNetCore datasets. Each column title for SR/SN indicates the chosen known class set and the rest sets of SR/SN are unknown." + ], + "table_footnote": [], + "table_body": "
MethodSydneynuScenesS3DIS
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
ULIP[54]63.481.753.293.353.195.1
ULIP2[55]66.283.764.293.062.592.3
DDCS[57]61.887.362.291.262.592.5
MCM[36]65.782.457.182.365.288.1
NegLabel[24]58.492.769.781.655.485.4
ZLaP[44]58.580.670.087.971.388.4
GSP72.568.980.880.777.694.0
", + "bbox": [ + 524, + 388, + 890, + 481 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Table 2. Zero-shot OOD detection results on the Sydney Urban Objects dataset, nuScenes dataset and S3DIS dataset.", + "bbox": [ + 511, + 487, + 906, + 513 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "servations. i) Unlike the findings from the ScanObjNN dataset, the performance gap between the leading baseline, ULIP, and our method under the zero-shot protocol is relatively narrow, approximately $0.5\\%$ . This modest improvement can be attributed to the substantial overlap between the downstream and pre-training datasets used in Vision-Language Models (VLMs). Specifically, ULIP/ULIP2 are pre-trained on ShapeNet, which shares significant commonalities with ShapeNetCore, thereby impacting OOD detection capabilities. Despite this, the slight enhancement achieved by our method underscores the utility of leveraging data manifolds to improve the differentiation of OOD samples. ii) A marked enhancement in performance is observed when few-shot samples are available. For instance, introducing 1-shot training samples results in nearly a $7\\%$ increase in average AUROC, while 5-shot samples yield an impressive $15\\%$ improvement. This highlights the critical role of few-shot learning in enhancing model performance.", + "bbox": [ + 511, + 537, + 906, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Analysis of Additional Real-World Datasets: The zero-shot OOD detection results for the Sydney Urban Objects dataset [11] and nuScenes Dataset [7], presented in Tab. 2, reinforce our previous conclusions. Our method consistently outperforms all competing approaches in zero-shot OOD detection, demonstrating the efficacy of utilizing test", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "28802", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing data manifolds, particularly when there is a substantial shift between the downstream task and the pre-training data. This finding emphasizes the robustness of our approach in adapting to significant distributional shifts in real-world applications. For S3DIS [4], GSP achieves the highest AUROC, outperforming all baselines by a significant margin (e.g., +12.4% over MCM). This suggests superior ability to distinguish between classes across all decision thresholds. However, GSP's FPR95 (94.0%) is slightly higher than ULIP (95.1%), indicating that while GSP excels in overall classification, it faces challenges in suppressing false positives when operating at a 95% true positive rate.", + "bbox": [ + 89, + 90, + 480, + 272 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Further Insights Synergizing All Results: By comparing the outcomes from the ScanObjNN, Sydney Urban Objects, S3DIS, nuScenes and ShapeNetCore datasets, we derive several overarching insights. i) There is a notable correlation between the distribution shift from pre-training to downstream datasets and the efficacy of manifold exploitation for OOD detection. This can be attributed to several factors. Firstly, when pre-training is conducted on datasets that closely resemble the downstream datasets, the contrastive training loss effectively aligns text prompts with visual features, resulting in a well-defined separation between ID and OOD samples. Conversely, in scenarios where the distribution shift is more substantial, VLMs may struggle to accurately quantify the distance between ID text prompts and visual samples. In such cases, leveraging the testing data manifold becomes essential to enhance the calibration of prompt-to-visual feature distances, thereby improving OOD detection performance.", + "bbox": [ + 91, + 276, + 483, + 547 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 571, + 243, + 589 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We begin by analyzing the performance of three Out-of-Distribution (OOD) detection methods: Cosine Similarity, Shortest Path, and Label Propagation. As shown in Tab. 3, we employ the ULIP2 backbone for extracting text and point cloud features and evaluate these methods for OOD scoring. The approach of directly computing the maximal normalized cosine similarity between text and point cloud features yields a baseline performance. Substituting cosine similarity with the shortest path metric results in a performance increment, albeit modest. It is noteworthy that the shortest path is computed using Dijkstra's algorithm, and the reciprocal of manifold distance serves as the OOD score. Lastly, we introduce a label propagation method to determine the OOD score. The experimental outcomes in Tab. 3 demonstrate that integrating graph-based information via label propagation significantly enhances the model's ability to differentiate between In-Distribution (ID) and OOD categories. This suggests that graph-based methods provide a more nuanced understanding of the feature space, leading to improved OOD detection capabilities.", + "bbox": [ + 89, + 598, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4a1fb485606fa542c9d329c4f6732b1ebf4c62337122c5e5d8c68627cf484394.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
OOD Det. Mthd.Prompt Clust.Neg. PromptAUROC↑FPR95↓
Cosine Dist.--73.276.7
Manifold Dist.--74.173.0
Score Prop.--76.472.6
Score Prop.✓-77.273.6
Score Prop.-✓77.371.4
Score Prop.✓✓78.667.4
", + "bbox": [ + 519, + 89, + 897, + 176 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 3. Ablation study on ScanObjNN dataset.", + "bbox": [ + 566, + 180, + 849, + 194 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/45c2ee79d7002bde5cc5bc02aebdc3488f57fab29dc063a33b492b1262781da5.jpg", + "image_caption": [ + "(a)" + ], + "image_footnote": [], + "bbox": [ + 521, + 203, + 705, + 287 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f1c1471fe75611a45cd261888034d8e51f3934b9cdf6e8d4f5b1c617e7a18762.jpg", + "image_caption": [ + "(b)", + "Figure 3. Few-shot OOD detection on the ScanObjNN dataset. (a) Results on AUROC metric. (b) Results on FPR95 metric." + ], + "image_footnote": [], + "bbox": [ + 712, + 203, + 895, + 286 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.4. Further Analysis", + "text_level": 1, + "bbox": [ + 511, + 363, + 679, + 378 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Few-shot OOD Detection: To evaluate the performance of label propagation in few-shot OOD detection, we incorporated a subset of training data from the 3DOS training dataset, treating them as in-distribution categories. The outcomes, presented in Fig. 3, show the effects of varying the number of training samples used in label propagation, with all samples randomly selected from the training dataset. The results indicate that as the number of training samples increases, OOD detection performance improves and eventually stabilizes.", + "bbox": [ + 511, + 386, + 903, + 536 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Analysis: To validate our method, we visualize sample predictions and scores from baseline method and our full GSP. As shown in Fig. 5, for easy-to-predict samples from the Real-world dataset, conventional distance-based methods perform well. However, in challenging cases (e.g., occlusions or missing parts), they often fail. In contrast, our method utilizes spatially proximate samples in feature space for score propagation, ensuring accurate OOD detection even in difficult scenarios. These qualitative results confirm the effectiveness and robustness of our approach.", + "bbox": [ + 511, + 537, + 905, + 703 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Impact of Prompt Clustering: To illustrate the benefits of prompt expansion via K-means clustering, we examined how varying the number of prompts affects OOD (Out-of-Distribution) detection. As depicted in Fig. 4 (a), we experimented with different values of $N_{c}$ in the K-means algorithm, where each $N_{c}$ value corresponds to a distinct number of clusters, and consequently, a different number of prompts. An increase in $N_{c}$ results in a greater number of distinct prompts.", + "bbox": [ + 511, + 704, + 905, + 839 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The results presented in the figure reveal that the quantity of prompts considerably impacts OOD detection performance. Specifically, when $N_{c} = 6$ , the AUROC surpasses $78.7\\%$ . However, further increasing $N_{c}$ can negatively af", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "28803", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/eb35f117bd9b1174e6783df8ad2831aa7144b96479f7b73ae896783e36cee4ea.jpg", + "image_caption": [ + "(a) $N_{k}$ Nearest Neighbors" + ], + "image_footnote": [], + "bbox": [ + 94, + 90, + 361, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/2b9530dee9a3d2a6ccd4cf28519da4f2600f1519fae7f6989d4b826006a3063d.jpg", + "image_caption": [ + "(b) Number of negative prompt $m$", + "Figure 4. Results from ScanObjNN Benchmark Experiments: (a) Evaluation of prompt clustering with different K-means cluster numbers. (b) Impact of varying negative prompts on label propagation. (c) Results of different number of K Neaerst Neighbors in KNN.", + "Figure 5. Visualization of some samples for ScanObjectNN and nuScenes Dataset with two methods predict scores. All scores are normalized to 0-1." + ], + "image_footnote": [], + "bbox": [ + 366, + 90, + 633, + 202 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/65800647109a20c84af0d895f5cef84f7b817989c9daad2c147d02885165d623.jpg", + "image_caption": [ + "(c) Cluster Number $N_{c}$" + ], + "image_footnote": [], + "bbox": [ + 635, + 90, + 905, + 202 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/93be5203fa709110b2f5b241f77a08875a8d27d48744f3372ac56618d10367b3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ScanObjectNN DatasetnuScenes Dataset
ID categoryEasy-to-predictHard-to-predictEasy-to-predictHard-to-predict
Distance-based score:0.6180.0930.6890.192
GSP score:0.4650.3570.3740.278
OOD category
Distance-based score:0.1740.3960.1480.431
GSP score:0.0720.0540.0740.127
", + "bbox": [ + 99, + 275, + 478, + 431 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fect the spatial alignment between text features and point cloud features. This phenomenon is likely due to the fact that typical prompts in Vision-Language Models (VLMs) are generally standardized, with only occasional outliers. Excessive clustering may isolate these outliers, leading to a decline in performance.", + "bbox": [ + 89, + 492, + 482, + 582 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In conventional VLM approaches, the mean text feature vector is typically calculated for each category (equivalent to setting $N_{c} = 1$ ), achieving an average AUROC of less than $78\\%$ . The findings in Fig. 4 (c) suggest that selecting an appropriate $N_{c}$ enhances the model's ability to differentiate OOD categories, underscoring the advantages of prompt clustering.", + "bbox": [ + 89, + 583, + 482, + 686 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Negative Prompting via Self-Training: We conducted experiments to evaluate the effectiveness of the proposed negative prompting. Using the ShapeNetCore dataset, we explored the impact of self-training for negative prompting by varying the percentage of samples tagged with pseudo labels, denoted as $m\\%$ as introduced in Section 3.3. In Fig. 4 (b), We observe a performance increase from $m = 1$ up to $m = 6$ , indicating that incorporating negative prompts through self-training is an effective strategy to leverage the manifold for score propagation. However, performance declines beyond $m = 6$ , likely due to the introduction of incorrect pseudo-negative prompts. This highlights the importance of mitigating the effects of confirmation bias [3] to prevent the model from being adversely", + "bbox": [ + 89, + 688, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/466e9cb93ca14792989b9e559a525c4f9eb545655fc2d86e26a8dea0cafeb329.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Modelload + DatasetloadText Feature ExtractPrompt ClusteringFeature ExtractBuild GraphScore PropagationTotal time
SR1552ms2212ms1128ms2780ms106ms8ms7786ms
SN1641ms2459ms1657ms25031ms671ms34ms31493ms
", + "bbox": [ + 517, + 275, + 924, + 334 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. The inference time of the GSP on SR and SN.", + "bbox": [ + 545, + 338, + 870, + 351 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "affected.", + "bbox": [ + 513, + 367, + 573, + 380 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of K-Nearest-Neighbor. We compare performance with different values of $N_{k}$ Nearest Neighbors. As we can see in Fig 4 (a), the results demonstrate that as the number of $N_{k}$ values increases, the OOD detection performance is tend to be stable. While $N_{k} = 16$ achieve the best performance and Less than 8 or larger than 18 get the low results.", + "bbox": [ + 511, + 383, + 906, + 474 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Cost: We profiled each step and test the inference time of GSP using a full GPU implementation in Tab. 4. Experiments are carried out on a single RTX3090 GPU and a i7-12700k CPU. As observed, the graph-based operations (Build Graph + Score Propagation) account for only around $2\\%$ of the total inference time. The additional computation cost introduced by graph operations is negligible compared with point cloud feature extraction.", + "bbox": [ + 511, + 476, + 906, + 597 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 616, + 633, + 632 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this work, we presented a novel training-free approach to 3D point cloud OOD detection that effectively leverages Vision-Language Models through Graph Score Propagation. By constructing a graph that incorporates both positive prototypes and pseudo-negative samples via self-training, our method captures the underlying structure of the testing data manifold. This approach allows for the propagation of informative OOD scores, thus surpassing existing distance-based techniques. Empirical results across multiple 3D point cloud datasets demonstrate the robustness and adaptability of our framework, particularly in zero- and few-shot settings. The GSP method addresses key limitations in prior OOD approaches by reducing dependency on extensive in-distribution training data while achieving high accuracy. Future work could explore further enhancements through dynamic graph updates, enabling real-time OOD detection in continuously evolving environments.", + "bbox": [ + 511, + 643, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "28804", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 209, + 187, + 224 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems, 2022. 2, 5", + "[2] Mouin Ben Ammar, Nacim Belkhir, Sebastian Popescu, Antoine Manzanera, and Gianni Franchi. Neco: Neural collapse based out-of-distribution detection. arXiv preprint arXiv:2310.06823, 2023. 2", + "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks, 2020. 8", + "[4] I Armeni, O Sener, Ar Zamir, H Jiang, I Brilakis, M Fischer, S Savarese, and IEEE. 3d semantic parsing of large-scale indoor spaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5, 7", + "[5] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2", + "[6] Tianyi Bao, Qitian Wu, Zetian Jiang, Yiting Chen, Jiawei Sun, and Junchi Yan. Graph out-of-distribution detection goes neighborhood shaping. In International Conference on Machine Learning. 2", + "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 5, 6", + "[8] Tianle Cai, Ruiqi Gao, Jason Lee, and Qi Lei. A theory of label propagation for subpopulation shift. In International Conference on Machine Learning, 2021. 2", + "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5", + "[10] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern analysis and Machine Intelligence, 2021. 1, 6", + "[11] Mark De Deuge, Alastair Quadros, Calvin Hung, and Bertrand Douillard. Unsupervised feature learning for classification of outdoor 3d scans. In Australasian Conference on Robotics and Automation, 2013. 5, 6" + ], + "bbox": [ + 94, + 234, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "This work was supported by the National Natural Science Foundation of China (No. U2468207), Sichuan Science and Technology Program (No. 2024NSFTD0036), and the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (No. M23L7b0021).", + "[12] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In European Conference on Computer Vision, 2020. 6", + "[13] Edsger W Dijkstra. A note on two problems in connexion with graphs. In Edsger Wybe Dijkstra: his life, work, and legacy. 2022. 3", + "[14] Shreyas Havaldar, Navodita Sharma, Shubhi Sareen, Karthikeyan Shanmugam, and Aravindan Raghuveer. Learning from label proportions: Bootstrapping supervised learners via belief propagation. arXiv preprint arXiv:2310.08056, 2023. 2", + "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 1, 2, 6", + "[16] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In International Conference on Learning Representations, 2019. 6", + "[17] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2", + "[18] Chengjie Huang, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, Krzysztof Czarnecki, et al. Out-of-distribution detection for lidar-based 3d object detectabel propagation for zero-shot classification with vision. In IEEE International Conference on Intelligent Transportation Systems, 2022. 1, 2", + "[19] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 2021. 6", + "[20] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In IEEE/CVF International Conference on Computer Vision, 2023. 2, 3", + "[21] Abiodun M. Ikatun, Absalom E. Ezugwu, Laith Abualigah, Belal Abuhaija, and Jia Heming. K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. Information science, 2023. 4", + "[22] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2", + "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 1", + "[24] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided ood detection with pretrained vision-language models. In International Conference on Learning Representations, 2024. 3, 5, 6" + ], + "bbox": [ + 94, + 92, + 903, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "28805", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 6", + "[26] Konstantin Kirchheim, Tim Gonschorek, and Frank Ortmeyer. Out-of-distribution detection with logical reasoning. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2", + "[27] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus C. J. Dietmayer. Revisiting out-of-distribution detection in lidar-based 3d object detection. IEEE Intelligent Vehicles Symposium, 2024. 2", + "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 1", + "[29] Jingyao Li, Pengguang Chen, Zexin He, Shaozuo Yu, Shu Liu, and Jiaya Jia. Rethinking out-of-distribution (ood) detection: Masked image modeling is all you need. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4", + "[31] Yushu Li, Yongyi Su, Adam Goodge, Kui Jia, and Xun Xu. Efficient and context-aware label propagation for zero/few-shot training-free adaptation of vision-language model. In International Conference on Learning Representations, 2025. 2, 4", + "[32] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 6", + "[33] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 2020. 6", + "[34] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In IEEE International Conference on Image Processing, 2021. 1", + "[35] Guofeng Mei, Luigi Riz, Yiming Wang, and Fabio Poiesi. Geometrically-driven aggregation for zero-shot 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2", + "[36] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In Advances in Neural Information Processing Systems, 2022. 1, 3, 5, 6", + "[37] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 2024. 1, 2", + "[38] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In International Conference on Learning Representations, 2024. 2, 4" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Julia Nitsch, Masha Itkina, Ransalu Senanayake, Juan Nieto, Max Schmidt, Roland Siegwart, Mykel J Kochenderfer, and Cesar Cadena. Out-of-distribution detection for automotive perception. In IEEE International Intelligent Transportation Systems Conference, 2021. 1", + "[40] Rattana Pukdee, Dylan Sam, Maria-Florina Balcan, and Pradeep Ravikumar. Label propagation with weak supervision. arXiv preprint arXiv:2210.03594, 2022. 2", + "[41] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 2022. 1", + "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2", + "[43] Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Gunnemann. Graph posterior network: Bayesian predictive uncertainty for node classification. Advances in Neural Information Processing Systems, 2021. 2", + "[44] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6", + "[45] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions. arXiv preprint arXiv:2201.12296, 2022. 2", + "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems, 2021. 6", + "[47] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, 2022. 2", + "[48] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4", + "[49] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019. 5", + "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 6", + "[51] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In IEEE/CVF International Conference on Computer Vision, 2023. 1, 2", + "[52] Qizhou Wang, Zhen Fang, Yonggang Zhang, Feng Liu, Yixuan Li, and Bo Han. Learning to augment distributions for out-of-distribution detection. Advances in Neural Information Processing Systems, 2023. 2" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "28806", + "bbox": [ + 478, + 946, + 519, + 957 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. Energy-based out-of-distribution detection for graph neural networks. ArXiv, 2023. 2", + "[54] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5, 6", + "[55] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 4, 5, 6", + "[56] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5", + "[57] Yue Yuan, Rundong He, Yicong Dong, Zhongyi Han, and Yilong Yin. Discriminability-driven channel selection for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5, 6", + "[58] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 1", + "[59] Ji Zhang, Jingkuan Song, Lianli Gao, Nicu Sebe, and Heng Tao Shen. Reliable few-shot learning under dual noises. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2", + "[60] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2", + "[61] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision. Springer, 2022. 2", + "[62] Zihan Zhang and Xiang Xiang. Decoupling maxlogit for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2", + "[63] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. ProQuest number: information to all users, 2002. 3", + "[64] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In International Conference on Machine Learning, 2003. 2" + ], + "bbox": [ + 91, + 92, + 480, + 825 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "28807", + "bbox": [ + 478, + 946, + 517, + 957 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_model.json b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..103d81b33e802ad24bffea2ab7d7a552aeb3c1cd --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_model.json @@ -0,0 +1,2213 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.123, + 0.13, + 0.877, + 0.177 + ], + "angle": 0, + "content": "Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation" + }, + { + "type": "text", + "bbox": [ + 0.136, + 0.203, + 0.862, + 0.257 + ], + "angle": 0, + "content": "Tiankai Chen\\(^{1}\\), Yushu Li\\(^{23}\\), Adam Goodge\\(^{3}\\), Fei Teng\\(^{1}\\), Xulei Yang\\(^{3}\\), Tianrui Li\\(^{1}\\), Xun Xu\\(^{\\dagger 3}\\) \n\\(^{1}\\)Southwest Jiaotong University, \\(^{2}\\)South China University of Technology \n\\(^{3}\\)Institute for infocomm research(I\\(^{2}\\)R), A*STAR" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.324, + 0.486, + 0.598 + ], + "angle": 0, + "content": "Out-of-distribution (OOD) detection in 3D point cloud data remains a challenge, particularly in applications where safe and robust perception is critical. While existing OOD detection methods have shown progress for 2D image data, extending these to 3D environments involves unique obstacles. This paper introduces a training-free framework that leverages Vision-Language Models (VLMs) for effective OOD detection in 3D point clouds. By constructing a graph based on class prototypes and testing data, we exploit the data manifold structure to enhancing the effectiveness of VLMs for 3D OOD detection. We propose a novel Graph Score Propagation (GSP) method that incorporates prompt clustering and self-training negative prompting to improve OOD scoring with VLM. Our method is also adaptable to few-shot scenarios, providing options for practical applications. We demonstrate that GSP consistently outperforms state-of-the-art methods across synthetic and real-world datasets for 3D point cloud OOD detection." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.629, + 0.222, + 0.645 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.655, + 0.485, + 0.868 + ], + "angle": 0, + "content": "The detection of out-of-distribution (OOD) samples in 3D point cloud data is critical for the safe and robust deployment of deep learning models in real-world environments. For example, identifying unexpected objects is crucial for ensuring the safety of personnel, vehicles, and robots by preventing collisions [18, 39, 41]. Existing approaches to OOD detection in 3D point clouds often focus on training separate models for specific tasks [54]. These methods typically rely on abundant in-distribution (ID) training data to build models based on reconstruction [34], distance metrics [10], or discriminative strategies [15]. While these methods achieve strong results, they depend on substantial ID training data, which is often challenging to collect in many real-world applications and generalization may be" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.293, + 0.9, + 0.522 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.525, + 0.907, + 0.58 + ], + "angle": 0, + "content": "Figure 1. Exploiting the testing data manifold for zero-/few-shot OOD detection. ID prototypes propagate scores to test samples, with confident ones selected as pseudo negative prompts for refinement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.596, + 0.907, + 0.642 + ], + "angle": 0, + "content": "hindered by the limited diversity of ID training samples. These limitations highlight the need for alternative solutions that do not require extensive ID data collection." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.643, + 0.909, + 0.839 + ], + "angle": 0, + "content": "Vision-language models (VLMs) have emerged as powerful foundation models that bridge visual features with textual understanding [23, 28, 42, 58]. VLMs are trained to associate the content of visual inputs with their textual labels and they demonstrate strong zero-shot generalization to unseen visual content [42]. This capacity makes VLMs appealing for OOD detection without retraining [5, 36, 37, 51]. A common VLM-based OOD detection pipeline maps ID class concepts, such as class names with various templates, into a shared feature space using the text encoder. The visual samples, similarly projected using the visual encoder, are then scored by comparing their distances to the ID class concepts in this feature space [36]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "This VLM-based OOD detection pipeline has shown competitive performance for 2D image detection, primarily due to the vast understanding of 2D VLMs from the massive-scale image-text pre-training datasets (e.g. 400M" + }, + { + "type": "page_footnote", + "bbox": [ + 0.104, + 0.88, + 0.458, + 0.905 + ], + "angle": 0, + "content": "\\(\\dagger\\) Correspondence to Xun Xu: xu_xun@i2ra.a-star.edu.sg, Code link: https://github.com/handsome999KK/GSP_OOD" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.947, + 0.521, + 0.959 + ], + "angle": 0, + "content": "28797" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.211 + ], + "angle": 0, + "content": "pairs for pre-training CLIP [42]). In contrast, VLMs for 3D point clouds, such as the state-of-the-art ULIP2 [55], are typically pre-trained on much more restricted datasets (e.g., 800k point cloud samples). This restricts their ability to generalize across unseen categories and distributional shifts (e.g., style changes), which reduces the effectiveness of similarity-based methods for 3D point cloud OOD detection." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.214, + 0.482, + 0.424 + ], + "angle": 0, + "content": "In this work, we propose to leverage the structure of the testing data manifold to compensate for this deficiency and improve VLM adaptation for downstream tasks [31, 44], as illustrated in Fig. 1. Specifically, we construct a graph based on class prototypes (e.g., text prompts for ID classes) and testing data features, where prototype nodes are assigned labels that propagate to neighboring testing nodes. Given the limited training data, we hypothesize that exploiting the testing data manifold enhances the generalization capabilities of 3D point cloud VLMs. Our method, Graph Score Propagation (GSP), assigns a positive OOD score (e.g., +1) to ID class prototypes while all testing samples start with a neutral score (0). As scores propagate along the graph, the resultant scores indicate OOD likelihood." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.427, + 0.483, + 0.727 + ], + "angle": 0, + "content": "This straightforward adoption of score propagation significantly improves upon existing distance-based methods, and we introduce two additional enhancements to further boost OOD detection accuracy. First, while conventional training-free approaches typically use a single class prototype by averaging text prompts [44, 61], we argue that capturing the prompt distribution through clustering better leverages prompt diversity. Second, negative prompting has shown strong performance in VLM-based OOD detection [30, 38, 48]. However, simply negating the description (e.g. add \"not\") is not very effective and learning negative prompts requires sophisticated optimization. Instead, we employ self-training to convert high-confidence OOD testing samples into pseudo-negative prompts and assign them a negative score (e.g., -1), for subsequent propagation iterations. Finally, our framework integrates seamlessly into traditional supervised OOD detection protocols by incorporating a few ID samples as positive prompts, further enhancing OOD detection under both zero-shot and few-shot protocols." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.73, + 0.462, + 0.745 + ], + "angle": 0, + "content": "Our contributions are mainly summarized as follows." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "- We employ a pre-trained vision-language model for 3D point cloud OOD detection and introduce Graph Score Propagation (GSP), a novel methodology that leverages the testing data manifold instead of solely relying on text-test feature distances." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.826, + 0.483, + 0.871 + ], + "angle": 0, + "content": "- We enhance GSP with prompt clustering and self-training-based negative prompting. GSP is also compatible with both zero-shot and few-shot testing protocols." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "- We validate the effectiveness of our method on synthetic and real-world 3D point cloud OOD detection bench" + }, + { + "type": "list", + "bbox": [ + 0.091, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.528, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "marks, consistently achieving superior performance over state-of-the-art VLM-based OOD detection methods." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.136, + 0.655, + 0.151 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.161, + 0.905, + 0.478 + ], + "angle": 0, + "content": "Out-of-Distribution Detection: Traditional OOD detection methods focused on softmax-based probability scoring [15]. Recently, attention has shifted to novel evaluation metrics [26, 47, 57, 62], such as differential network channel processing [57, 62] and non-parametric nearest-neighbor distances [47]. Other strategies involve training specific OOD discriminators, like DAL [52], which constructs an OOD distribution set, and MOOD [29], which uses masked image modeling. NECO [2] utilizes neural collapse geometry for OOD identification. Vision-language models have introduced new methods using text prompts to distinguish between in-distribution (ID) and OOD data [5, 30, 37, 38, 51]. LoCoOP [37] leverages CLIP's local features for OOD regularization, while Bai et al. [5] use CLIP to detect ID-like outliers, enhancing OOD detection and few-shot learning [59]. Training text encoders [51] and learnable prompt [30, 38] further increases the score gap between ID and OOD data. Recent research has expanded to include 3D datasets for point cloud OOD detection [1, 45]. Studies have shown that 3D OOD detection is particularly important for autonomous driving scenarios [18, 27]." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.479, + 0.905, + 0.644 + ], + "angle": 0, + "content": "3D Vision-Language Models: Recent advancements in Vision-Language Models (VLMs) have facilitated innovative approaches for 3D understanding. Techniques utilizing CLIP [20, 35, 60] transform 3D point clouds into imagery data for extracting features aligned with text descriptions. More recently, methods like ULIP [54, 55] learn unified representations across images, text, and 3D point clouds, further improving zero-shot 3D classification. These approaches leverage the integration of text prompts and visual features which could better distinguish between indistribution and out-of-distribution data in 3D contexts." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Label Propagation: Label propagation (LP) is a pivotal method in semi-supervised learning, leveraging limited labeled data to infer labels for unlabeled instances by exploiting data similarity across a graph [8, 14, 22, 40]. It assumes that labels vary smoothly across the graph, making it effective in transductive learning settings [22, 64]. Recent advancements integrate LP with graph neural networks to enhance its ability to separate ID from OOD data [6, 43, 53]. In the context of VLMs, LP employs text prompts as label representations, constructing graphs that propagate labels from text prototypes to test samples, thus improving zero-shot 2D classification [17, 31, 44]. Contrary to the existing attempts of combining LP with VLM, we tackle a one-class classification task. Only prototypes for in-distribution classes are available. We further propose a novel negative prompting technique via self-training and prompt clustering technique to better separate ID from OOD." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.958 + ], + "angle": 0, + "content": "28798" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.09, + 0.226, + 0.108 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.115, + 0.295, + 0.131 + ], + "angle": 0, + "content": "3.1. Problem Formulation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.138, + 0.484, + 0.456 + ], + "angle": 0, + "content": "We begin by formally defining the task of VLM-based out-of-distribution (OOD) detection in 3D point clouds. Let \\( z^{f} = f(x;\\theta) \\) and \\( z^{t} = g(t;\\phi) \\) denote the encoded features of point cloud data \\( x \\) and text \\( t \\), where \\( f(\\cdot ;\\theta) \\) and \\( g(\\cdot ;\\phi) \\) are the VLM visual and text encoders respectively. The task involves the encoded features of unlabeled test data \\( \\mathcal{D}_u = \\{f(x_i)\\}_{i = 1\\dots N_u} \\) and, optionally, the features and labels of few-shot labeled data \\( \\mathcal{D}_l = \\{f(x_j),y_j\\}_{j = 1\\dots N_l} \\). The set of in-distribution (ID) classes \\( \\mathcal{Y}_{in} = \\{1,\\dots ,C_{in}\\} \\) is assumed to be known. Following established practices for VLM adaptation [36], we represent each ID class with \\( N_{t} \\) prompt templates \\( \\mathcal{T} = \\{t_j(y)\\}_{N_t} \\), e.g., \"This is a point cloud of \\( \\{\\mathrm{y}\\} \\)\". \"A good point cloud of \\( \\{\\mathrm{y}\\} \\)\". These textual prompts are encoded by the text encoder, creating prototypes for ID classes \\( \\mathcal{P} \\). Existing works often create prototypes by averaging the encoded prompts for each class, resulting in a prototype \\( z_c^p \\) for the \\( c \\)-th class, and the full set of prototypes is denoted \\( \\mathcal{P} = \\{z_c^p\\}_{c = 1\\dots C_{in}} \\). The OOD detection task is to determine whether a test sample \\( x_{i}\\in \\mathcal{D}_{u} \\) belongs to an ID class, leveraging ID prototypes \\( \\mathcal{P} \\) and, optionally, the few-shot labeled data \\( \\mathcal{D}_l \\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.484, + 0.518 + ], + "angle": 0, + "content": "Preliminary of VLM: We briefly review the training objective of vision-language models (VLMs). The pre-training steps align the visual features with their associated text features [20, 54, 55], typically via a contrastive loss function:" + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.531, + 0.483, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {v l m} = \\sum_ {i} \\log \\frac {\\exp \\left(z _ {i} ^ {f} z _ {i} ^ {t} / \\tau\\right)}{\\sum_ {j} \\exp \\left(z _ {i} ^ {f} z _ {j} ^ {t} / \\tau\\right)} \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.579, + 0.484, + 0.716 + ], + "angle": 0, + "content": "When a VLM is trained on a large amount of point cloud-text pairs, the cosine similarity of positive pairs (a point cloud sample and text prompt of its ground-truth class) is higher than that of negative pairs. Therefore, a pre-trained VLM provides a straightforward solution to 3D point cloud OOD detection by measuring the cosine similarity between textual descriptions of in-distribution classes and testing sample features. A high distance in the visual-text feature space suggests a greater likelihood of being OOD." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.726, + 0.32, + 0.741 + ], + "angle": 0, + "content": "3.2. VLM for OOD Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.484, + 0.856 + ], + "angle": 0, + "content": "Vanilla OOD Detection with VLM: Pre-trained VLMs offer an effective approach to OOD detection by measuring the distance between visual samples and text labels in the feature space. For example, [36] employs the maximum normalized cosine similarity as the OOD score. We define a normalized score \\( S \\) to indicate OOD-ness, where \\( S \\rightarrow 0 \\) suggests OOD and \\( S \\rightarrow 1 \\) suggests ID." + }, + { + "type": "equation", + "bbox": [ + 0.172, + 0.868, + 0.483, + 0.906 + ], + "angle": 0, + "content": "\\[\nS = \\max _ {c} \\frac {\\exp \\left(- d \\left(z ^ {f} , z _ {c} ^ {p}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(- d \\left(z ^ {f} , z _ {j} ^ {p}\\right) / \\tau\\right)} \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.227 + ], + "angle": 0, + "content": "It is assumed that in the feature space, OOD samples are further from the ID prototypes than ID samples. Subsequent research has further explored this approach by introducing techniques like negative prompts to better separate OOD from ID samples [24]. In this work, we are motivated by the challenge of measuring similarity/distance in the VLM feature space and propose to tackle the VLM OOD detection problem by exploiting the manifold of testing and, optionally, few-shot labeled data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.228, + 0.905, + 0.304 + ], + "angle": 0, + "content": "Manifold Distance for OOD Detection: The vanilla measurement overlooks the complex manifold of testing data. A straightforward approach to consider testing data manifold for OOD detection is to replace the standard distance metric with a manifold distance, represented via a graph." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.906, + 0.519 + ], + "angle": 0, + "content": "Given a graph representing the data manifold \\( G = (E, V) \\) with weight matrix \\( W \\in \\mathbb{R}^{(|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|) \\times (|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|)} \\), the manifold distance between two nodes on the graph is defined as the shortest path. The shortest path between arbitrary nodes can be obtained via Dijkstra's algorithm [13]. More concretely, this can be formulated as finding a path between a prototype \\( z^p \\in \\mathcal{P} \\) or a few-shot labeled sample \\( z_k^f \\in \\mathcal{D}_l \\) and a testing sample \\( z_j^f \\in \\mathcal{D}_u \\), consisting of edges \\( \\tilde{E}_{ij} = \\{e_k\\} \\) on the graph, such that the sum of these edge weights is minimized. The manifold distance between two nodes is thus defined as \\( d(z_j, z_i) = \\min_{\\tilde{E}_{ij}} \\sum_{e_k \\in \\tilde{E}_{ij}} d(e_k) \\), where \\( d(e_k) \\) denotes the distance for edge \\( e_k \\), e.g. the L2 distance between the encoded features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.906, + 0.657 + ], + "angle": 0, + "content": "Our empirical results in Tab. 3 reveal that incorporating the manifold improves the effectiveness of OOD detection. However, there are two main disadvantages associated with distance-based OOD detection. First, the manifold distance is unnormalised, if directly using the reciprocal of manifold distance as OOD score the result may be biased by the relative scale. Additionally, it is non-trivial to integrate negative prompts within the distance-based OOD detection paradigm." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.669, + 0.882, + 0.685 + ], + "angle": 0, + "content": "3.3. VLM OOD Detection as Score Propagation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.692, + 0.906, + 0.799 + ], + "angle": 0, + "content": "Instead of directly measuring the distance for OOD detection, we formulate OOD detection as a score propagation problem. Specifically, we assign a positive score \\((+1)\\) to ID prompts \\(\\mathcal{P}\\), while all testing samples are initially assigned a neutral score (0). Labels are then propagated from the positive nodes to all testing samples in an iterative manner. We follow a standard propagation rule [63] as shown in Eq. 3," + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.811, + 0.732, + 0.827 + ], + "angle": 0, + "content": "\\[\nS _ {t} = \\tilde {W} S _ {t - 1} + \\alpha S _ {0},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.554, + 0.832, + 0.905, + 0.865 + ], + "angle": 0, + "content": "\\[\n\\text {s . t .} \\tilde {W} = D ^ {- \\frac {1}{2}} W D ^ {- \\frac {1}{2}}, D = \\operatorname {d i a g} \\left(\\sum_ {j} W _ {i j}\\right), \\tag {3}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.589, + 0.868, + 0.816, + 0.904 + ], + "angle": 0, + "content": "\\[\nS_{0} = \\bigl[\\underbrace{1,\\cdots,1}_{|\\mathcal{P}| + |\\mathcal{D}_{l}| \\text{elements}},\\underbrace{0,\\cdots,0}_{|\\mathcal{D}_{u}| \\text{elements}}\\bigr],\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.947, + 0.519, + 0.958 + ], + "angle": 0, + "content": "28799" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.089, + 0.898, + 0.241 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.247, + 0.908, + 0.29 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed framework for zero-/few-shot OOD detection via graph score propagation. The framework encodes ID category prompts and dataset samples using frozen text and 3D encoders, constructing a KNN graph. Initial score propagation generates pseudo prompts, which refine the graph before final score propagation computes OOD scores." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.305, + 0.482, + 0.38 + ], + "angle": 0, + "content": "where each step of propagation updates the score for unlabeled testing samples as a weighted average of scores from adjacent nodes and initial scores \\( S_0 \\) until \\( T \\) iterations are completed. Consequently, nodes that are closer to the prototypes on the manifold will obtain a higher score." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.593 + ], + "angle": 0, + "content": "Prompt Clustering: The above propagation strategy propagates labels from ID prototypes to testing samples on the graph. Thus, it is critical to design effective ID prototypes that capture the diversity of textual descriptions for ID classes. Contrary to existing practices that average the text prompt features across all prompt templates, we propose modeling the distribution of prompt templates with multiple representative ones. Given the prompt templates \\(\\mathcal{T}\\), we first initialize a pool of text prompts \\(\\mathcal{P}_c = \\{z_{c_j}^p = g(t_j(y_c))\\}_{j=1\\dots|\\mathcal{T}|}\\) for the \\(c\\)-th class. We then introduce a clustering step to reduce this set to \\(N_c\\) representative prototypes \\(\\tilde{\\mathcal{P}}_c\\). This is achieved by applying K-means [21] clustering on \\(\\mathcal{P}_c\\), and the cluster centers are adopted as the ID prototypes:" + }, + { + "type": "equation", + "bbox": [ + 0.139, + 0.615, + 0.432, + 0.655 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\mathcal {P}} _ {c} = \\arg \\min _ {\\{\\mu_ {k} \\}, \\gamma} \\sum_ {N _ {c}} \\sum_ {z _ {c j} ^ {p} \\in \\mathcal {P} _ {c}} \\gamma_ {j k} \\| z _ {c j} ^ {p} - \\mu_ {k} \\| ^ {2},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.141, + 0.656, + 0.393, + 0.686 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l l} \\text {s . t .} & \\gamma \\in \\{0, 1 \\} ^ {| \\mathcal {T} | \\times N _ {c}}, \\sum_ {k} \\gamma_ {j k} = 1, \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.687, + 0.483, + 0.749 + ], + "angle": 0, + "content": "where \\(\\mu_{k}\\) are the cluster centers, and \\(\\gamma_{jk}\\) indicates the assignment of each prompt feature \\(z_{cj}^{p}\\) to its closest center. The new prototypes, \\(\\tilde{\\mathcal{P}} = \\tilde{\\mathcal{P}}_1\\cup \\tilde{\\mathcal{P}}_2\\cup \\dots \\tilde{\\mathcal{P}}_C\\) , are used for score propagation as positive nodes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Negative Prompting via Self-Training: Negative prompts have been shown to substantially benefit OOD detection [30, 38, 48]. Incorporating negative prompts is particularly straightforward within the graph label propagation framework. We treat negative prompts as nodes \\(\\mathcal{P}^n\\) with a label of \\(-1\\) on the graph. During propagation, the negative labels compete with the positive labels (ID prompts) to influence the score of testing samples. Existing approach exploits ID data to learn negative prompt [48] which prohibits zero-shot inference. To address this, we propose a" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.305, + 0.907, + 0.396 + ], + "angle": 0, + "content": "strategy that generates negative prompts via self-training, as illustrated in Fig. 1. Specifically, we first conduct initial score propagation for \\( T \\) iterations following Eq. 3, giving rise to initial propagated scores \\( S_{T} \\). We select the top \\( m\\% \\) most confident scores from \\( S_{T} \\) following Eq. 5 where \\( \\mathrm{top}(S,m\\%) \\) indicates the top \\( m\\% \\) value of \\( S \\)." + }, + { + "type": "equation", + "bbox": [ + 0.574, + 0.407, + 0.816, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\mathcal{I}_{pos} = \\left\\{i\\mid S_{T}[i] > \\operatorname {top}(S_{T},m\\%)\\right\\} ,\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.424, + 0.905, + 0.444 + ], + "angle": 0, + "content": "\\[\n\\mathcal {I} _ {n e g} = \\left\\{i \\mid S _ {T} [ i ] < \\operatorname {top} \\left(S _ {T}, 1 - m \\%\\right) \\right\\} \\tag{5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.447, + 0.907, + 0.523 + ], + "angle": 0, + "content": "We label the selected samples as positive and negative prompts respectively and re-initialize the score \\(\\tilde{S}_0\\). For negative prompts, we assign scores of \\(-1\\) to further increase the gap between ID and OOD samples. The final score is obtained by running Eq. 3 with \\(\\tilde{S}_0\\) as initial scores." + }, + { + "type": "equation", + "bbox": [ + 0.572, + 0.535, + 0.907, + 0.591 + ], + "angle": 0, + "content": "\\[\n\\tilde {S} _ {0} [ i ] = \\left\\{ \\begin{array}{l l} S _ {0} [ i ] & \\text {i f} i \\notin \\mathcal {I} _ {\\text {p o s}} \\& i \\notin \\mathcal {I} _ {\\text {n e g}}, \\\\ + 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {p o s}}, \\\\ - 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {n e g}}. \\end{array} \\right. \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.909, + 0.841 + ], + "angle": 0, + "content": "Graph Construction: We now introduce the approach for constructing a graph for score propagation. As observed by [31, 44], the distance between text prompts and visual samples is generally larger than the distance between visual samples. This observation also holds for 3D point cloud VLM because the 3D VLMs are often finetuned from 2D VLMs [54, 55]. Therefore, we adopt a similar approach to construct the graph. We write the weight matrix \\(W\\) in a blockwise form as shown in Eq. 7, where \\(W_{u} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{u}|}\\), \\(W_{up} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\tilde{\\mathcal{P}}|}\\) and \\(W_{ul} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{l}|}\\). No connections are created between prototypes and few-shot labeled samples. All blocks are further constructed as a k-nearest neighbor (KNN) graph following [31, 44]. It is also computationally efficient to adapt an existing graph to new testing samples by incrementally adding the testing samples to the graph [31]." + }, + { + "type": "equation", + "bbox": [ + 0.577, + 0.85, + 0.907, + 0.902 + ], + "angle": 0, + "content": "\\[\nW = \\left[ \\begin{array}{c c c} \\mathbf {I} ^ {| \\tilde {\\mathcal {P}} | \\times | \\tilde {\\mathcal {P}} |} & \\mathbf {0} & W _ {u p} ^ {\\top} \\\\ \\mathbf {0} & \\mathbf {I} ^ {| \\mathcal {D} _ {l} | \\times | \\mathcal {D} _ {l} |} & W _ {u l} ^ {\\top} \\\\ W _ {u p} & W _ {u l} & W _ {u} \\end{array} \\right] \\tag {7}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.947, + 0.521, + 0.958 + ], + "angle": 0, + "content": "28800" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "Overall Algorithm: We present the overall algorithm for unified iterative label propagation for vision-language models in Algorithm 1. In Fig. 2, we show an overview of the proposed framework for zero-/few-shot OOD detection via graph propagation." + }, + { + "type": "code_caption", + "bbox": [ + 0.094, + 0.18, + 0.482, + 0.197 + ], + "angle": 0, + "content": "Algorithm 1 Graph Score Propagation for OOD Detection" + }, + { + "type": "algorithm", + "bbox": [ + 0.101, + 0.2, + 0.484, + 0.542 + ], + "angle": 0, + "content": "1: Input: training data \\(\\mathcal{D}_l\\), testing data \\(\\mathcal{D}_u\\) and ID class prompt templates \\(\\mathcal{T}_c\\) \n2: Output: Final OOD scores \\(\\tilde{S}_T\\) \n3: for \\(\\mathcal{D}_l\\) and \\(\\mathcal{T}_c\\), do \n Encode: \\(z_l = f(x_l)\\) for training data and \\(\\mathcal{P}_c = \\{g(t_j(y_c))\\}\\) for prompt templates. Prompt clustering: positive prototypes \\(\\tilde{\\mathcal{P}}\\) (Eq. 4) \n4: end for \n5: for \\(\\mathcal{D}_u\\), do \n Encode: \\(z_i = f(x_i)\\) for testing data \n Build graph to compute \\(W\\) (Eq. 7) \n Initialize score \\(S_0^{(i)}\\) (Eq. 3) \n Score propagate to compute \\(S_T\\) (Eq. 3) \n Self-training to generate positive \\(\\mathcal{I}_{pos}\\) and negative prompts \\(\\mathcal{I}_{neg}\\) (Eq. 5) \n Updated initial scores \\(\\tilde{S}_0\\) (Eq. 6) \n Score propagate to compute \\(\\tilde{S}_T\\) (Eq. 3) \n6: end for \n7: return \\(\\tilde{S}_T\\)" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.568, + 0.224, + 0.584 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.592, + 0.279, + 0.608 + ], + "angle": 0, + "content": "4.1. Experimental setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.614, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Dataset: In this work, we follow the 3DOS framework [1] to construct downstream datasets for evaluating our method. The 3DOS benchmark is based on two well-known 3D object datasets, ShapeNetCore [9] and ScanObjectNN [49]. ShapeNetCore includes 54 categories covering daily objects. ScanObjectNN is cropped out from ScanNet dataset featuring a real-world objects. It consists of 15 categories with 2,902 3D objects. To further validate our approach on real-world data, we introduce the Sydney Urban Object Dataset [11], nuScenes [7] and S3DIS [4]. Sydney Urban Object Dataset consists of LiDAR-scanned point clouds across 26 categories, totaling 631 unique instances. S3DIS dataset contains 6 large-scale indoor areas with 271 rooms annotated into 13 semantic categories. We cropped out individual objects using ground-truth segmentation masks, leading to 8931 objects. The nuScenes Dataset features point clouds captured by LiDAR sensor under autonomous driving scenario. We cropped out point clouds using ground-truth bounding boxes, leading to 2205" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.122 + ], + "angle": 0, + "content": "objects across 9 categories. Detailed visualization of each dataset is deferred to the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.124, + 0.907, + 0.593 + ], + "angle": 0, + "content": "Evaluation Protocol: Following the methodology outlined in 3DOS [1], we divide ShapeNetCore into three non-overlapping category sets, SN1, SN2, and SN3, referred to as the ShapeNetCore dataset. Each set contains 18 categories. For each OOD test scenario, one of these sets (SN1, SN2, or SN3) is treated as the known set of categories, while the remaining two sets are used as unknown categories, resulting in three distinct OOD test cases. Similarly, 3DOS defines three category sets, SR1, SR2, and SR3 on the ScanObjectNN dataset. Each SR-set includes 5 non-overlapping categories, with the other two sets acting as unknown categories in each scenario and call them as ScanObjNN. As for Sydney Urban Object Dataset, we categorize these objects into two groups: movable objects, e.g. \"truck\" \"pedestrian\" and non-movable objects, e.g. \"building\" and \"traffic sign\". Movable objects are treated as in-distribution (ID) classes, simulating a scenario in autonomous driving where object detection primarily focuses on movable objects. For consistency, we pre-process each point cloud sample by upsampling it to a uniform number of points. For S3DIS dataset, We retained instances containing more than 2048 original points, where each instance was randomly downsampled to retain exactly 2048 points. Then we set the foreground objects as ID and back-ground objects as OOD. For nuScences dataset, we have conducted preliminary screening on the original point cloud and retained instances with more than 200 original points and upsampling it to 1024 points. Then we define Four-wheeled vehicles (including car, bus, truck, construction_vehicle, trailer) as ID and the rest as OOD. And further details provided in the supplementary material." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.595, + 0.909, + 0.746 + ], + "angle": 0, + "content": "Implementation Details. We use the pre-trained ULIP/ULIP2 [54, 55] model to extract point cloud features and text features. For the 3D encoder, we select PointBERT [56]. In constructing the KNN graph, we set a universal \\( \\mathrm{K} \\)-value \\( (N_{k}) \\) of 10, connecting each text feature to its 10 nearest point cloud features. In Eq. 3, we set \\( \\alpha \\) to 0.5 and the number of iterations \\( T \\) to 5. The percentage of pseudo negative prompts, i.e. \\( m\\% \\) in Sect. 3.3, is also set to \\( 5\\% \\). Additionally, we use K-means clustering \\( (N_{c}) \\) to group the prompts for each category into 3 clusters." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.748, + 0.909, + 0.884 + ], + "angle": 0, + "content": "Competing Methods. We compare our method with several OOD detection approaches, including full-shot methods from 3DOS [1]. We also benchmark against recent zero-shot methods, such as DDCS [57], ZLaP[44], MCM [36], and NegLabel [24]. For zero-shot comparisons, all methods use the 3D encoder from [55]. Baseline results are also reported by directly using cosine similarity as the OOD score, with these methods referred to as ULIP and ULIP2, depending on the VLM used." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate OOD detection perfor" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.947, + 0.518, + 0.958 + ], + "angle": 0, + "content": "28801" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.094, + 0.089, + 0.907, + 0.337 + ], + "angle": 0, + "content": "
MethodScanObjNNShapeNetCore
SR3AUROC↑FPR95↓SR2AUROC↑FPR95↓SR1AUROC↑FPR95↓AverageAUROC↑FPR95↓SN1AUROC↑FPR95↓SN2AUROC↑FPR95↓SN3AUROC↑FPR95↓AverageAUROC↑FPR95↓
Requiring training datasetMSP[15]88.167.380.684.073.780.380.877.274.382.880.078.189.752.281.371.0
MLS[50]89.453.883.473.176.475.383.067.472.080.883.964.189.840.581.961.8
ODIN[32]90.247.983.371.776.376.883.365.574.279.479.471.787.841.880.564.3
Energy[33]89.550.681.675.876.675.582.667.372.181.283.064.789.839.482.061.8
GradNorm[19]88.550.777.475.375.276.880.467.672.181.857.788.957.879.062.683.3
ReAct[46]90.348.982.475.875.477.682.767.473.779.489.652.195.027.286.152.9
NF88.047.780.668.275.681.481.465.881.572.571.178.091.049.681.266.7
OE+mixup[16]72.683.572.088.562.587.869.086.672.778.980.368.887.362.280.169.9
ARPL+CS[10]--------74.880.380.772.485.450.880.367.8
Cosine protoCE91.041.082.178.277.675.683.664.980.368.388.760.891.938.086.955.7
SupCon[25]--------80.975.583.568.285.145.183.262.9
SubArcface[12]87.161.378.976.973.781.479.973.279.081.282.960.389.132.883.758.1
GSP(1-shot)79.361.380.984.188.651.782.965.791.344.086.443.183.891.487.259.5
GSP(5-shot)80.754.890.147.289.140.286.647.495.619.294.622.197.012.595.717.9
GSP(full-shot)90.133.093.030.592.234.491.822.697.910.197.97.998.86.098.28.0
Zero-shotULIP[54]79.982.762.387.557.492.366.587.588.057.176.448.775.251.379.952.4
ULIP2[55]65.780.876.471.277.478.073.276.785.752.584.144.367.674.479.157.1
DDCS[57]63.390.562.480.064.496.263.488.961.186.766.886.577.276.268.483.1
MCM[36]52.092.362.391.381.380.465.288.085.151.683.246.166.475.878.257.8
NegLabel[24]65.785.566.891.367.488.266.688.360.687.880.678.688.048.376.471.6
ZLaP[44]70.181.767.084.767.883.668.383.388.252.872.366.477.490.079.369.7
GSP75.863.076.577.683.661.678.667.490.638.970.764.079.793.780.465.5
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.344, + 0.907, + 0.371 + ], + "angle": 0, + "content": "Table 1. Evaluation of OOD detection for ScanObjNN and ShapeNetCore datasets. Each column title for SR/SN indicates the chosen known class set and the rest sets of SR/SN are unknown." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.392, + 0.483, + 0.482 + ], + "angle": 0, + "content": "mance using two widely adopted metrics: AUROC and FPR95. AUROC calculates the area under the receiver operating characteristic curve, with higher values indicating better performance. FPR95 measures the false positive rate at a true positive rate of \\(95\\%\\), with lower values reflecting improved performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.498, + 0.33, + 0.512 + ], + "angle": 0, + "content": "4.2. Results on OOD Detection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.522, + 0.483, + 0.868 + ], + "angle": 0, + "content": "Analysis of ScanObjNN Results: From the results presented in Tab. 1, we derive several key insights. i) Our proposed method demonstrates highly competitive performance under the zero-shot protocol, achieving an improvement of over \\(5\\%\\) in AUROC compared to ULIP2, which uses cosine distance directly as an OOD score. This substantial enhancement suggests that employing a manifold-based approach significantly enhances the differentiation between OOD samples and ID classes. ii) Our few-shot results are equally impressive, rivaling even fully supervised methodologies. Notably, with 5-shot training samples, our method attains an average AUROC of \\(86.6\\%\\), surpassing the best fully supervised method, Cosine proto, which scores \\(83.6\\%\\), by a margin of \\(3\\%\\). Furthermore, when integrating all available training samples into the graph, our approach achieves an outstanding average AUROC of \\(91.8\\%\\). iii) A noticeable \\(5\\%\\) gap in average AUROC is observed between our zero-shot and 1-shot methods. This gap underscores the critical role of incorporating real in-distribution samples in effectively distinguishing OOD samples. Consequently, practitioners are advised to introduce at least a few ID samples for OOD detection, in addition to leveraging Vision-Language Models (VLMs)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Analysis of ShapeNetCore Results: The results for ShapeNetCore, as shown in Tab. 1, lead us to several ob" + }, + { + "type": "table", + "bbox": [ + 0.526, + 0.389, + 0.892, + 0.482 + ], + "angle": 0, + "content": "
MethodSydneynuScenesS3DIS
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
ULIP[54]63.481.753.293.353.195.1
ULIP2[55]66.283.764.293.062.592.3
DDCS[57]61.887.362.291.262.592.5
MCM[36]65.782.457.182.365.288.1
NegLabel[24]58.492.769.781.655.485.4
ZLaP[44]58.580.670.087.971.388.4
GSP72.568.980.880.777.694.0
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.488, + 0.907, + 0.515 + ], + "angle": 0, + "content": "Table 2. Zero-shot OOD detection results on the Sydney Urban Objects dataset, nuScenes dataset and S3DIS dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.809 + ], + "angle": 0, + "content": "servations. i) Unlike the findings from the ScanObjNN dataset, the performance gap between the leading baseline, ULIP, and our method under the zero-shot protocol is relatively narrow, approximately \\(0.5\\%\\). This modest improvement can be attributed to the substantial overlap between the downstream and pre-training datasets used in Vision-Language Models (VLMs). Specifically, ULIP/ULIP2 are pre-trained on ShapeNet, which shares significant commonalities with ShapeNetCore, thereby impacting OOD detection capabilities. Despite this, the slight enhancement achieved by our method underscores the utility of leveraging data manifolds to improve the differentiation of OOD samples. ii) A marked enhancement in performance is observed when few-shot samples are available. For instance, introducing 1-shot training samples results in nearly a \\(7\\%\\) increase in average AUROC, while 5-shot samples yield an impressive \\(15\\%\\) improvement. This highlights the critical role of few-shot learning in enhancing model performance." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Analysis of Additional Real-World Datasets: The zero-shot OOD detection results for the Sydney Urban Objects dataset [11] and nuScenes Dataset [7], presented in Tab. 2, reinforce our previous conclusions. Our method consistently outperforms all competing approaches in zero-shot OOD detection, demonstrating the efficacy of utilizing test" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.519, + 0.958 + ], + "angle": 0, + "content": "28802" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.273 + ], + "angle": 0, + "content": "ing data manifolds, particularly when there is a substantial shift between the downstream task and the pre-training data. This finding emphasizes the robustness of our approach in adapting to significant distributional shifts in real-world applications. For S3DIS [4], GSP achieves the highest AUROC, outperforming all baselines by a significant margin (e.g., +12.4% over MCM). This suggests superior ability to distinguish between classes across all decision thresholds. However, GSP's FPR95 (94.0%) is slightly higher than ULIP (95.1%), indicating that while GSP excels in overall classification, it faces challenges in suppressing false positives when operating at a 95% true positive rate." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.277, + 0.485, + 0.549 + ], + "angle": 0, + "content": "Further Insights Synergizing All Results: By comparing the outcomes from the ScanObjNN, Sydney Urban Objects, S3DIS, nuScenes and ShapeNetCore datasets, we derive several overarching insights. i) There is a notable correlation between the distribution shift from pre-training to downstream datasets and the efficacy of manifold exploitation for OOD detection. This can be attributed to several factors. Firstly, when pre-training is conducted on datasets that closely resemble the downstream datasets, the contrastive training loss effectively aligns text prompts with visual features, resulting in a well-defined separation between ID and OOD samples. Conversely, in scenarios where the distribution shift is more substantial, VLMs may struggle to accurately quantify the distance between ID text prompts and visual samples. In such cases, leveraging the testing data manifold becomes essential to enhance the calibration of prompt-to-visual feature distances, thereby improving OOD detection performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.573, + 0.245, + 0.59 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.599, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We begin by analyzing the performance of three Out-of-Distribution (OOD) detection methods: Cosine Similarity, Shortest Path, and Label Propagation. As shown in Tab. 3, we employ the ULIP2 backbone for extracting text and point cloud features and evaluate these methods for OOD scoring. The approach of directly computing the maximal normalized cosine similarity between text and point cloud features yields a baseline performance. Substituting cosine similarity with the shortest path metric results in a performance increment, albeit modest. It is noteworthy that the shortest path is computed using Dijkstra's algorithm, and the reciprocal of manifold distance serves as the OOD score. Lastly, we introduce a label propagation method to determine the OOD score. The experimental outcomes in Tab. 3 demonstrate that integrating graph-based information via label propagation significantly enhances the model's ability to differentiate between In-Distribution (ID) and OOD categories. This suggests that graph-based methods provide a more nuanced understanding of the feature space, leading to improved OOD detection capabilities." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.09, + 0.898, + 0.177 + ], + "angle": 0, + "content": "
OOD Det. Mthd.Prompt Clust.Neg. PromptAUROC↑FPR95↓
Cosine Dist.--73.276.7
Manifold Dist.--74.173.0
Score Prop.--76.472.6
Score Prop.✓-77.273.6
Score Prop.-✓77.371.4
Score Prop.✓✓78.667.4
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.181, + 0.851, + 0.195 + ], + "angle": 0, + "content": "Table 3. Ablation study on ScanObjNN dataset." + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.204, + 0.707, + 0.289 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.607, + 0.29, + 0.623, + 0.301 + ], + "angle": 0, + "content": "(a)" + }, + { + "type": "image", + "bbox": [ + 0.714, + 0.204, + 0.896, + 0.287 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.81, + 0.289, + 0.828, + 0.301 + ], + "angle": 0, + "content": "(b)" + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.309, + 0.905, + 0.337 + ], + "angle": 0, + "content": "Figure 3. Few-shot OOD detection on the ScanObjNN dataset. (a) Results on AUROC metric. (b) Results on FPR95 metric." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.364, + 0.681, + 0.38 + ], + "angle": 0, + "content": "4.4. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.387, + 0.905, + 0.537 + ], + "angle": 0, + "content": "Few-shot OOD Detection: To evaluate the performance of label propagation in few-shot OOD detection, we incorporated a subset of training data from the 3DOS training dataset, treating them as in-distribution categories. The outcomes, presented in Fig. 3, show the effects of varying the number of training samples used in label propagation, with all samples randomly selected from the training dataset. The results indicate that as the number of training samples increases, OOD detection performance improves and eventually stabilizes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.906, + 0.704 + ], + "angle": 0, + "content": "Qualitative Analysis: To validate our method, we visualize sample predictions and scores from baseline method and our full GSP. As shown in Fig. 5, for easy-to-predict samples from the Real-world dataset, conventional distance-based methods perform well. However, in challenging cases (e.g., occlusions or missing parts), they often fail. In contrast, our method utilizes spatially proximate samples in feature space for score propagation, ensuring accurate OOD detection even in difficult scenarios. These qualitative results confirm the effectiveness and robustness of our approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.906, + 0.84 + ], + "angle": 0, + "content": "Impact of Prompt Clustering: To illustrate the benefits of prompt expansion via K-means clustering, we examined how varying the number of prompts affects OOD (Out-of-Distribution) detection. As depicted in Fig. 4 (a), we experimented with different values of \\( N_{c} \\) in the K-means algorithm, where each \\( N_{c} \\) value corresponds to a distinct number of clusters, and consequently, a different number of prompts. An increase in \\( N_{c} \\) results in a greater number of distinct prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "The results presented in the figure reveal that the quantity of prompts considerably impacts OOD detection performance. Specifically, when \\( N_{c} = 6 \\), the AUROC surpasses \\( 78.7\\% \\). However, further increasing \\( N_{c} \\) can negatively af" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.518, + 0.958 + ], + "angle": 0, + "content": "28803" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.092, + 0.362, + 0.202 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.205, + 0.308, + 0.218 + ], + "angle": 0, + "content": "(a) \\(N_{k}\\) Nearest Neighbors" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.092, + 0.634, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.409, + 0.205, + 0.619, + 0.218 + ], + "angle": 0, + "content": "(b) Number of negative prompt \\( m \\)" + }, + { + "type": "image", + "bbox": [ + 0.636, + 0.092, + 0.906, + 0.203 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.706, + 0.205, + 0.846, + 0.218 + ], + "angle": 0, + "content": "(c) Cluster Number \\(N_{c}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.905, + 0.251 + ], + "angle": 0, + "content": "Figure 4. Results from ScanObjNN Benchmark Experiments: (a) Evaluation of prompt clustering with different K-means cluster numbers. (b) Impact of varying negative prompts on label propagation. (c) Results of different number of K Neaerst Neighbors in KNN." + }, + { + "type": "table", + "bbox": [ + 0.101, + 0.276, + 0.48, + 0.432 + ], + "angle": 0, + "content": "
ScanObjectNN DatasetnuScenes Dataset
ID categoryEasy-to-predictHard-to-predictEasy-to-predictHard-to-predict
Distance-based score:0.6180.0930.6890.192
GSP score:0.4650.3570.3740.278
OOD category
Distance-based score:0.1740.3960.1480.431
GSP score:0.0720.0540.0740.127
" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.439, + 0.483, + 0.48 + ], + "angle": 0, + "content": "Figure 5. Visualization of some samples for ScanObjectNN and nuScenes Dataset with two methods predict scores. All scores are normalized to 0-1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.493, + 0.483, + 0.583 + ], + "angle": 0, + "content": "fect the spatial alignment between text features and point cloud features. This phenomenon is likely due to the fact that typical prompts in Vision-Language Models (VLMs) are generally standardized, with only occasional outliers. Excessive clustering may isolate these outliers, leading to a decline in performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.483, + 0.688 + ], + "angle": 0, + "content": "In conventional VLM approaches, the mean text feature vector is typically calculated for each category (equivalent to setting \\( N_{c} = 1 \\)), achieving an average AUROC of less than \\( 78\\% \\). The findings in Fig. 4 (c) suggest that selecting an appropriate \\( N_{c} \\) enhances the model's ability to differentiate OOD categories, underscoring the advantages of prompt clustering." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Effectiveness of Negative Prompting via Self-Training: We conducted experiments to evaluate the effectiveness of the proposed negative prompting. Using the ShapeNetCore dataset, we explored the impact of self-training for negative prompting by varying the percentage of samples tagged with pseudo labels, denoted as \\( m\\% \\) as introduced in Section 3.3. In Fig. 4 (b), We observe a performance increase from \\( m = 1 \\) up to \\( m = 6 \\), indicating that incorporating negative prompts through self-training is an effective strategy to leverage the manifold for score propagation. However, performance declines beyond \\( m = 6 \\), likely due to the introduction of incorrect pseudo-negative prompts. This highlights the importance of mitigating the effects of confirmation bias [3] to prevent the model from being adversely" + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.276, + 0.925, + 0.335 + ], + "angle": 0, + "content": "
Modelload + DatasetloadText Feature ExtractPrompt ClusteringFeature ExtractBuild GraphScore PropagationTotal time
SR1552ms2212ms1128ms2780ms106ms8ms7786ms
SN1641ms2459ms1657ms25031ms671ms34ms31493ms
" + }, + { + "type": "table_caption", + "bbox": [ + 0.547, + 0.339, + 0.872, + 0.352 + ], + "angle": 0, + "content": "Table 4. The inference time of the GSP on SR and SN." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.368, + 0.575, + 0.381 + ], + "angle": 0, + "content": "affected." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.385, + 0.907, + 0.476 + ], + "angle": 0, + "content": "Impact of K-Nearest-Neighbor. We compare performance with different values of \\( N_{k} \\) Nearest Neighbors. As we can see in Fig 4 (a), the results demonstrate that as the number of \\( N_{k} \\) values increases, the OOD detection performance is tend to be stable. While \\( N_{k} = 16 \\) achieve the best performance and Less than 8 or larger than 18 get the low results." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.477, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Computational Cost: We profiled each step and test the inference time of GSP using a full GPU implementation in Tab. 4. Experiments are carried out on a single RTX3090 GPU and a i7-12700k CPU. As observed, the graph-based operations (Build Graph + Score Propagation) account for only around \\(2\\%\\) of the total inference time. The additional computation cost introduced by graph operations is negligible compared with point cloud feature extraction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.617, + 0.634, + 0.633 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this work, we presented a novel training-free approach to 3D point cloud OOD detection that effectively leverages Vision-Language Models through Graph Score Propagation. By constructing a graph that incorporates both positive prototypes and pseudo-negative samples via self-training, our method captures the underlying structure of the testing data manifold. This approach allows for the propagation of informative OOD scores, thus surpassing existing distance-based techniques. Empirical results across multiple 3D point cloud datasets demonstrate the robustness and adaptability of our framework, particularly in zero- and few-shot settings. The GSP method addresses key limitations in prior OOD approaches by reducing dependency on extensive in-distribution training data while achieving high accuracy. Future work could explore further enhancements through dynamic graph updates, enabling real-time OOD detection in continuously evolving environments." + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.519, + 0.958 + ], + "angle": 0, + "content": "28804" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.114, + 0.484, + 0.186 + ], + "angle": 0, + "content": "This work was supported by the National Natural Science Foundation of China (No. U2468207), Sichuan Science and Technology Program (No. 2024NSFTD0036), and the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (No. M23L7b0021)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.21, + 0.188, + 0.226 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.235, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems, 2022. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.305, + 0.484, + 0.358 + ], + "angle": 0, + "content": "[2] Mouin Ben Ammar, Nacim Belkhir, Sebastian Popescu, Antoine Manzanera, and Gianni Franchi. Neco: Neural collapse based out-of-distribution detection. arXiv preprint arXiv:2310.06823, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.361, + 0.483, + 0.415 + ], + "angle": 0, + "content": "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks, 2020. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.417, + 0.483, + 0.47 + ], + "angle": 0, + "content": "[4] I Armeni, O Sener, Ar Zamir, H Jiang, I Brilakis, M Fischer, S Savarese, and IEEE. 3d semantic parsing of large-scale indoor spaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.472, + 0.484, + 0.539 + ], + "angle": 0, + "content": "[5] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.541, + 0.483, + 0.595 + ], + "angle": 0, + "content": "[6] Tianyi Bao, Qitian Wu, Zetian Jiang, Yiting Chen, Jiawei Sun, and Junchi Yan. Graph out-of-distribution detection goes neighborhood shaping. In International Conference on Machine Learning. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.597, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.68, + 0.484, + 0.72 + ], + "angle": 0, + "content": "[8] Tianle Cai, Ruiqi Gao, Jason Lee, and Qi Lei. A theory of label propagation for subpopulation shift. In International Conference on Machine Learning, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.722, + 0.483, + 0.789 + ], + "angle": 0, + "content": "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.096, + 0.791, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[10] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern analysis and Machine Intelligence, 2021. 1, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.095, + 0.847, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Mark De Deuge, Alastair Quadros, Calvin Hung, and Bertrand Douillard. Unsupervised feature learning for classification of outdoor 3d scans. In Australasian Conference on Robotics and Automation, 2013. 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.235, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[12] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In European Conference on Computer Vision, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.151, + 0.905, + 0.191 + ], + "angle": 0, + "content": "[13] Edsger W Dijkstra. A note on two problems in connexion with graphs. In Edsger Wybe Dijkstra: his life, work, and legacy. 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.194, + 0.905, + 0.261 + ], + "angle": 0, + "content": "[14] Shreyas Havaldar, Navodita Sharma, Shubhi Sareen, Karthikeyan Shanmugam, and Aravindan Raghuveer. Learning from label proportions: Bootstrapping supervised learners via belief propagation. arXiv preprint arXiv:2310.08056, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.265, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 1, 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.322, + 0.904, + 0.362 + ], + "angle": 0, + "content": "[16] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In International Conference on Learning Representations, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.365, + 0.904, + 0.433 + ], + "angle": 0, + "content": "[17] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.436, + 0.905, + 0.518 + ], + "angle": 0, + "content": "[18] Chengjie Huang, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, Krzysztof Czarnecki, et al. Out-of-distribution detection for lidar-based 3d object detectabel propagation for zero-shot classification with vision. In IEEE International Conference on Intelligent Transportation Systems, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.521, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[19] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.905, + 0.645 + ], + "angle": 0, + "content": "[20] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In IEEE/CVF International Conference on Computer Vision, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.702 + ], + "angle": 0, + "content": "[21] Abiodun M. Ikatun, Absalom E. Ezugwu, Laith Abualigah, Belal Abuhaija, and Jia Heming. K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. Information science, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.705, + 0.905, + 0.759 + ], + "angle": 0, + "content": "[22] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.833, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[24] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided ood detection with pretrained vision-language models. In International Conference on Learning Representations, 2024. 3, 5, 6" + }, + { + "type": "list", + "bbox": [ + 0.095, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.519, + 0.958 + ], + "angle": 0, + "content": "28805" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.148 + ], + "angle": 0, + "content": "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[26] Konstantin Kirchheim, Tim Gonschorek, and Frank Ortmeyer. Out-of-distribution detection with logical reasoning. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.262 + ], + "angle": 0, + "content": "[27] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus C. J. Dietmayer. Revisiting out-of-distribution detection in lidar-based 3d object detection. IEEE Intelligent Vehicles Symposium, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.321, + 0.482, + 0.389 + ], + "angle": 0, + "content": "[29] Jingyao Li, Pengguang Chen, Zexin He, Shaozuo Yu, Shu Liu, and Jiaya Jia. Rethinking out-of-distribution (ood) detection: Masked image modeling is all you need. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.391, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.448, + 0.482, + 0.516 + ], + "angle": 0, + "content": "[31] Yushu Li, Yongyi Su, Adam Goodge, Kui Jia, and Xun Xu. Efficient and context-aware label propagation for zero/few-shot training-free adaptation of vision-language model. In International Conference on Learning Representations, 2025. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.518, + 0.482, + 0.572 + ], + "angle": 0, + "content": "[32] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[33] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 2020. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.673 + ], + "angle": 0, + "content": "[34] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In IEEE International Conference on Image Processing, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.676, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[35] Guofeng Mei, Luigi Riz, Yiming Wang, and Fabio Poiesi. Geometrically-driven aggregation for zero-shot 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.787 + ], + "angle": 0, + "content": "[36] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In Advances in Neural Information Processing Systems, 2022. 1, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[37] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[38] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In International Conference on Learning Representations, 2024. 2, 4" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[39] Julia Nitsch, Masha Itkina, Ransalu Senanayake, Juan Nieto, Max Schmidt, Roland Siegwart, Mykel J Kochenderfer, and Cesar Cadena. Out-of-distribution detection for automotive perception. In IEEE International Intelligent Transportation Systems Conference, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[40] Rattana Pukdee, Dylan Sam, Maria-Florina Balcan, and Pradeep Ravikumar. Label propagation with weak supervision. arXiv preprint arXiv:2210.03594, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.236 + ], + "angle": 0, + "content": "[41] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.239, + 0.905, + 0.322 + ], + "angle": 0, + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.324, + 0.905, + 0.393 + ], + "angle": 0, + "content": "[43] Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Gunnemann. Graph posterior network: Bayesian predictive uncertainty for node classification. Advances in Neural Information Processing Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.396, + 0.905, + 0.452 + ], + "angle": 0, + "content": "[44] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.453, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[45] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions. arXiv preprint arXiv:2201.12296, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.512, + 0.905, + 0.553 + ], + "angle": 0, + "content": "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.556, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[47] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.655 + ], + "angle": 0, + "content": "[48] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.657, + 0.905, + 0.728 + ], + "angle": 0, + "content": "[49] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.73, + 0.905, + 0.784 + ], + "angle": 0, + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.787, + 0.905, + 0.842 + ], + "angle": 0, + "content": "[51] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In IEEE/CVF International Conference on Computer Vision, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[52] Qizhou Wang, Zhen Fang, Yonggang Zhang, Feng Liu, Yixuan Li, and Bo Han. Learning to augment distributions for out-of-distribution detection. Advances in Neural Information Processing Systems, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.52, + 0.958 + ], + "angle": 0, + "content": "28806" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[53] Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. Energy-based out-of-distribution detection for graph neural networks. ArXiv, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.219 + ], + "angle": 0, + "content": "[54] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[55] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.304, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[56] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.431 + ], + "angle": 0, + "content": "[57] Yue Yuan, Rundong He, Yicong Dong, Zhongyi Han, and Yilong Yin. Discriminability-driven channel selection for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.432, + 0.482, + 0.487 + ], + "angle": 0, + "content": "[58] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.488, + 0.482, + 0.543 + ], + "angle": 0, + "content": "[59] Ji Zhang, Jingkuan Song, Lianli Gao, Nicu Sebe, and Heng Tao Shen. Reliable few-shot learning under dual noises. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.545, + 0.482, + 0.613 + ], + "angle": 0, + "content": "[60] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.615, + 0.482, + 0.684 + ], + "angle": 0, + "content": "[61] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.686, + 0.482, + 0.727 + ], + "angle": 0, + "content": "[62] Zihan Zhang and Xiang Xiang. Decoupling maxlogit for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.728, + 0.482, + 0.769 + ], + "angle": 0, + "content": "[63] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. ProQuest number: information to all users, 2002. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.771, + 0.482, + 0.826 + ], + "angle": 0, + "content": "[64] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In International Conference on Machine Learning, 2003. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.826 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.947, + 0.519, + 0.958 + ], + "angle": 0, + "content": "28807" + } + ] +] \ No newline at end of file diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_origin.pdf b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..65b0cc5c637e045e98315918750df92741d2a16b --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/eecf87ed-7051-4e02-88a7-ac0885e7ef9b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:964e2d43ed9673a84c28d43caa4b906a991ec24f9cd46330e2df8aa77c272411 +size 686274 diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/full.md b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..b2e9f9c6bcfe845c5a98a85e00103e2a2a6322bc --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/full.md @@ -0,0 +1,321 @@ +# Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation + +Tiankai Chen $^{1}$ , Yushu Li $^{23}$ , Adam Goodge $^{3}$ , Fei Teng $^{1}$ , Xulei Yang $^{3}$ , Tianrui Li $^{1}$ , Xun Xu $^{\dagger 3}$ $^{1}$ Southwest Jiaotong University, $^{2}$ South China University of Technology + $^{3}$ Institute for infocomm research(I $^{2}$ R), A*STAR + +# Abstract + +Out-of-distribution (OOD) detection in 3D point cloud data remains a challenge, particularly in applications where safe and robust perception is critical. While existing OOD detection methods have shown progress for 2D image data, extending these to 3D environments involves unique obstacles. This paper introduces a training-free framework that leverages Vision-Language Models (VLMs) for effective OOD detection in 3D point clouds. By constructing a graph based on class prototypes and testing data, we exploit the data manifold structure to enhancing the effectiveness of VLMs for 3D OOD detection. We propose a novel Graph Score Propagation (GSP) method that incorporates prompt clustering and self-training negative prompting to improve OOD scoring with VLM. Our method is also adaptable to few-shot scenarios, providing options for practical applications. We demonstrate that GSP consistently outperforms state-of-the-art methods across synthetic and real-world datasets for 3D point cloud OOD detection. + +# 1. Introduction + +The detection of out-of-distribution (OOD) samples in 3D point cloud data is critical for the safe and robust deployment of deep learning models in real-world environments. For example, identifying unexpected objects is crucial for ensuring the safety of personnel, vehicles, and robots by preventing collisions [18, 39, 41]. Existing approaches to OOD detection in 3D point clouds often focus on training separate models for specific tasks [54]. These methods typically rely on abundant in-distribution (ID) training data to build models based on reconstruction [34], distance metrics [10], or discriminative strategies [15]. While these methods achieve strong results, they depend on substantial ID training data, which is often challenging to collect in many real-world applications and generalization may be + +![](images/fd65801fa754b791ecc7ed43d2b422ae763933e9b0384fb59d3c4e4d8e329d58.jpg) +Figure 1. Exploiting the testing data manifold for zero-/few-shot OOD detection. ID prototypes propagate scores to test samples, with confident ones selected as pseudo negative prompts for refinement. + +hindered by the limited diversity of ID training samples. These limitations highlight the need for alternative solutions that do not require extensive ID data collection. + +Vision-language models (VLMs) have emerged as powerful foundation models that bridge visual features with textual understanding [23, 28, 42, 58]. VLMs are trained to associate the content of visual inputs with their textual labels and they demonstrate strong zero-shot generalization to unseen visual content [42]. This capacity makes VLMs appealing for OOD detection without retraining [5, 36, 37, 51]. A common VLM-based OOD detection pipeline maps ID class concepts, such as class names with various templates, into a shared feature space using the text encoder. The visual samples, similarly projected using the visual encoder, are then scored by comparing their distances to the ID class concepts in this feature space [36]. + +This VLM-based OOD detection pipeline has shown competitive performance for 2D image detection, primarily due to the vast understanding of 2D VLMs from the massive-scale image-text pre-training datasets (e.g. 400M + +pairs for pre-training CLIP [42]). In contrast, VLMs for 3D point clouds, such as the state-of-the-art ULIP2 [55], are typically pre-trained on much more restricted datasets (e.g., 800k point cloud samples). This restricts their ability to generalize across unseen categories and distributional shifts (e.g., style changes), which reduces the effectiveness of similarity-based methods for 3D point cloud OOD detection. + +In this work, we propose to leverage the structure of the testing data manifold to compensate for this deficiency and improve VLM adaptation for downstream tasks [31, 44], as illustrated in Fig. 1. Specifically, we construct a graph based on class prototypes (e.g., text prompts for ID classes) and testing data features, where prototype nodes are assigned labels that propagate to neighboring testing nodes. Given the limited training data, we hypothesize that exploiting the testing data manifold enhances the generalization capabilities of 3D point cloud VLMs. Our method, Graph Score Propagation (GSP), assigns a positive OOD score (e.g., +1) to ID class prototypes while all testing samples start with a neutral score (0). As scores propagate along the graph, the resultant scores indicate OOD likelihood. + +This straightforward adoption of score propagation significantly improves upon existing distance-based methods, and we introduce two additional enhancements to further boost OOD detection accuracy. First, while conventional training-free approaches typically use a single class prototype by averaging text prompts [44, 61], we argue that capturing the prompt distribution through clustering better leverages prompt diversity. Second, negative prompting has shown strong performance in VLM-based OOD detection [30, 38, 48]. However, simply negating the description (e.g. add "not") is not very effective and learning negative prompts requires sophisticated optimization. Instead, we employ self-training to convert high-confidence OOD testing samples into pseudo-negative prompts and assign them a negative score (e.g., -1), for subsequent propagation iterations. Finally, our framework integrates seamlessly into traditional supervised OOD detection protocols by incorporating a few ID samples as positive prompts, further enhancing OOD detection under both zero-shot and few-shot protocols. + +Our contributions are mainly summarized as follows. + +- We employ a pre-trained vision-language model for 3D point cloud OOD detection and introduce Graph Score Propagation (GSP), a novel methodology that leverages the testing data manifold instead of solely relying on text-test feature distances. +- We enhance GSP with prompt clustering and self-training-based negative prompting. GSP is also compatible with both zero-shot and few-shot testing protocols. +- We validate the effectiveness of our method on synthetic and real-world 3D point cloud OOD detection bench + +marks, consistently achieving superior performance over state-of-the-art VLM-based OOD detection methods. + +# 2. Related Work + +Out-of-Distribution Detection: Traditional OOD detection methods focused on softmax-based probability scoring [15]. Recently, attention has shifted to novel evaluation metrics [26, 47, 57, 62], such as differential network channel processing [57, 62] and non-parametric nearest-neighbor distances [47]. Other strategies involve training specific OOD discriminators, like DAL [52], which constructs an OOD distribution set, and MOOD [29], which uses masked image modeling. NECO [2] utilizes neural collapse geometry for OOD identification. Vision-language models have introduced new methods using text prompts to distinguish between in-distribution (ID) and OOD data [5, 30, 37, 38, 51]. LoCoOP [37] leverages CLIP's local features for OOD regularization, while Bai et al. [5] use CLIP to detect ID-like outliers, enhancing OOD detection and few-shot learning [59]. Training text encoders [51] and learnable prompt [30, 38] further increases the score gap between ID and OOD data. Recent research has expanded to include 3D datasets for point cloud OOD detection [1, 45]. Studies have shown that 3D OOD detection is particularly important for autonomous driving scenarios [18, 27]. + +3D Vision-Language Models: Recent advancements in Vision-Language Models (VLMs) have facilitated innovative approaches for 3D understanding. Techniques utilizing CLIP [20, 35, 60] transform 3D point clouds into imagery data for extracting features aligned with text descriptions. More recently, methods like ULIP [54, 55] learn unified representations across images, text, and 3D point clouds, further improving zero-shot 3D classification. These approaches leverage the integration of text prompts and visual features which could better distinguish between indistribution and out-of-distribution data in 3D contexts. + +Label Propagation: Label propagation (LP) is a pivotal method in semi-supervised learning, leveraging limited labeled data to infer labels for unlabeled instances by exploiting data similarity across a graph [8, 14, 22, 40]. It assumes that labels vary smoothly across the graph, making it effective in transductive learning settings [22, 64]. Recent advancements integrate LP with graph neural networks to enhance its ability to separate ID from OOD data [6, 43, 53]. In the context of VLMs, LP employs text prompts as label representations, constructing graphs that propagate labels from text prototypes to test samples, thus improving zero-shot 2D classification [17, 31, 44]. Contrary to the existing attempts of combining LP with VLM, we tackle a one-class classification task. Only prototypes for in-distribution classes are available. We further propose a novel negative prompting technique via self-training and prompt clustering technique to better separate ID from OOD. + +# 3. Methodology + +# 3.1. Problem Formulation + +We begin by formally defining the task of VLM-based out-of-distribution (OOD) detection in 3D point clouds. Let $z^{f} = f(x;\theta)$ and $z^{t} = g(t;\phi)$ denote the encoded features of point cloud data $x$ and text $t$ , where $f(\cdot ;\theta)$ and $g(\cdot ;\phi)$ are the VLM visual and text encoders respectively. The task involves the encoded features of unlabeled test data $\mathcal{D}_u = \{f(x_i)\}_{i = 1\dots N_u}$ and, optionally, the features and labels of few-shot labeled data $\mathcal{D}_l = \{f(x_j),y_j\}_{j = 1\dots N_l}$ . The set of in-distribution (ID) classes $\mathcal{Y}_{in} = \{1,\dots ,C_{in}\}$ is assumed to be known. Following established practices for VLM adaptation [36], we represent each ID class with $N_{t}$ prompt templates $\mathcal{T} = \{t_j(y)\}_{N_t}$ , e.g., "This is a point cloud of $\{\mathrm{y}\}$ ". "A good point cloud of $\{\mathrm{y}\}$ ". These textual prompts are encoded by the text encoder, creating prototypes for ID classes $\mathcal{P}$ . Existing works often create prototypes by averaging the encoded prompts for each class, resulting in a prototype $z_c^p$ for the $c$ -th class, and the full set of prototypes is denoted $\mathcal{P} = \{z_c^p\}_{c = 1\dots C_{in}}$ . The OOD detection task is to determine whether a test sample $x_{i}\in \mathcal{D}_{u}$ belongs to an ID class, leveraging ID prototypes $\mathcal{P}$ and, optionally, the few-shot labeled data $\mathcal{D}_l$ . + +Preliminary of VLM: We briefly review the training objective of vision-language models (VLMs). The pre-training steps align the visual features with their associated text features [20, 54, 55], typically via a contrastive loss function: + +$$ +\mathcal {L} _ {v l m} = \sum_ {i} \log \frac {\exp \left(z _ {i} ^ {f} z _ {i} ^ {t} / \tau\right)}{\sum_ {j} \exp \left(z _ {i} ^ {f} z _ {j} ^ {t} / \tau\right)} \tag {1} +$$ + +When a VLM is trained on a large amount of point cloud-text pairs, the cosine similarity of positive pairs (a point cloud sample and text prompt of its ground-truth class) is higher than that of negative pairs. Therefore, a pre-trained VLM provides a straightforward solution to 3D point cloud OOD detection by measuring the cosine similarity between textual descriptions of in-distribution classes and testing sample features. A high distance in the visual-text feature space suggests a greater likelihood of being OOD. + +# 3.2. VLM for OOD Detection + +Vanilla OOD Detection with VLM: Pre-trained VLMs offer an effective approach to OOD detection by measuring the distance between visual samples and text labels in the feature space. For example, [36] employs the maximum normalized cosine similarity as the OOD score. We define a normalized score $S$ to indicate OOD-ness, where $S \rightarrow 0$ suggests OOD and $S \rightarrow 1$ suggests ID. + +$$ +S = \max _ {c} \frac {\exp \left(- d \left(z ^ {f} , z _ {c} ^ {p}\right) / \tau\right)}{\sum_ {j} \exp \left(- d \left(z ^ {f} , z _ {j} ^ {p}\right) / \tau\right)} \tag {2} +$$ + +It is assumed that in the feature space, OOD samples are further from the ID prototypes than ID samples. Subsequent research has further explored this approach by introducing techniques like negative prompts to better separate OOD from ID samples [24]. In this work, we are motivated by the challenge of measuring similarity/distance in the VLM feature space and propose to tackle the VLM OOD detection problem by exploiting the manifold of testing and, optionally, few-shot labeled data. + +Manifold Distance for OOD Detection: The vanilla measurement overlooks the complex manifold of testing data. A straightforward approach to consider testing data manifold for OOD detection is to replace the standard distance metric with a manifold distance, represented via a graph. + +Given a graph representing the data manifold $G = (E, V)$ with weight matrix $W \in \mathbb{R}^{(|\mathcal{P}| + |\mathcal{D}_l| + |\mathcal{D}_u|) \times (|\mathcal{P}| + |\mathcal{D}_l| + |\mathcal{D}_u|)}$ , the manifold distance between two nodes on the graph is defined as the shortest path. The shortest path between arbitrary nodes can be obtained via Dijkstra's algorithm [13]. More concretely, this can be formulated as finding a path between a prototype $z^p \in \mathcal{P}$ or a few-shot labeled sample $z_k^f \in \mathcal{D}_l$ and a testing sample $z_j^f \in \mathcal{D}_u$ , consisting of edges $\tilde{E}_{ij} = \{e_k\}$ on the graph, such that the sum of these edge weights is minimized. The manifold distance between two nodes is thus defined as $d(z_j, z_i) = \min_{\tilde{E}_{ij}} \sum_{e_k \in \tilde{E}_{ij}} d(e_k)$ , where $d(e_k)$ denotes the distance for edge $e_k$ , e.g. the L2 distance between the encoded features. + +Our empirical results in Tab. 3 reveal that incorporating the manifold improves the effectiveness of OOD detection. However, there are two main disadvantages associated with distance-based OOD detection. First, the manifold distance is unnormalised, if directly using the reciprocal of manifold distance as OOD score the result may be biased by the relative scale. Additionally, it is non-trivial to integrate negative prompts within the distance-based OOD detection paradigm. + +# 3.3. VLM OOD Detection as Score Propagation + +Instead of directly measuring the distance for OOD detection, we formulate OOD detection as a score propagation problem. Specifically, we assign a positive score $(+1)$ to ID prompts $\mathcal{P}$ , while all testing samples are initially assigned a neutral score (0). Labels are then propagated from the positive nodes to all testing samples in an iterative manner. We follow a standard propagation rule [63] as shown in Eq. 3, + +$$ +S _ {t} = \tilde {W} S _ {t - 1} + \alpha S _ {0}, +$$ + +$$ +\text {s . t .} \tilde {W} = D ^ {- \frac {1}{2}} W D ^ {- \frac {1}{2}}, D = \operatorname {d i a g} \left(\sum_ {j} W _ {i j}\right), \tag {3} +$$ + +$$ +S_{0} = \bigl[\underbrace{1,\cdots,1}_{|\mathcal{P}| + |\mathcal{D}_{l}| \text{elements}},\underbrace{0,\cdots,0}_{|\mathcal{D}_{u}| \text{elements}}\bigr], +$$ + +![](images/331e4e5e5dd294ddf6e887a4dbefd3b8fd4c66a668277220f12f5434e0979f36.jpg) +Figure 2. Overview of the proposed framework for zero-/few-shot OOD detection via graph score propagation. The framework encodes ID category prompts and dataset samples using frozen text and 3D encoders, constructing a KNN graph. Initial score propagation generates pseudo prompts, which refine the graph before final score propagation computes OOD scores. + +where each step of propagation updates the score for unlabeled testing samples as a weighted average of scores from adjacent nodes and initial scores $S_0$ until $T$ iterations are completed. Consequently, nodes that are closer to the prototypes on the manifold will obtain a higher score. + +Prompt Clustering: The above propagation strategy propagates labels from ID prototypes to testing samples on the graph. Thus, it is critical to design effective ID prototypes that capture the diversity of textual descriptions for ID classes. Contrary to existing practices that average the text prompt features across all prompt templates, we propose modeling the distribution of prompt templates with multiple representative ones. Given the prompt templates $\mathcal{T}$ , we first initialize a pool of text prompts $\mathcal{P}_c = \{z_{c_j}^p = g(t_j(y_c))\}_{j=1\dots|\mathcal{T}|}$ for the $c$ -th class. We then introduce a clustering step to reduce this set to $N_c$ representative prototypes $\tilde{\mathcal{P}}_c$ . This is achieved by applying K-means [21] clustering on $\mathcal{P}_c$ , and the cluster centers are adopted as the ID prototypes: + +$$ +\tilde {\mathcal {P}} _ {c} = \arg \min _ {\{\mu_ {k} \}, \gamma} \sum_ {N _ {c}} \sum_ {z _ {c j} ^ {p} \in \mathcal {P} _ {c}} \gamma_ {j k} \| z _ {c j} ^ {p} - \mu_ {k} \| ^ {2}, +$$ + +$$ +\begin{array}{l l} \text {s . t .} & \gamma \in \{0, 1 \} ^ {| \mathcal {T} | \times N _ {c}}, \sum_ {k} \gamma_ {j k} = 1, \end{array} +$$ + +where $\mu_{k}$ are the cluster centers, and $\gamma_{jk}$ indicates the assignment of each prompt feature $z_{cj}^{p}$ to its closest center. The new prototypes, $\tilde{\mathcal{P}} = \tilde{\mathcal{P}}_1\cup \tilde{\mathcal{P}}_2\cup \dots \tilde{\mathcal{P}}_C$ , are used for score propagation as positive nodes. + +Negative Prompting via Self-Training: Negative prompts have been shown to substantially benefit OOD detection [30, 38, 48]. Incorporating negative prompts is particularly straightforward within the graph label propagation framework. We treat negative prompts as nodes $\mathcal{P}^n$ with a label of $-1$ on the graph. During propagation, the negative labels compete with the positive labels (ID prompts) to influence the score of testing samples. Existing approach exploits ID data to learn negative prompt [48] which prohibits zero-shot inference. To address this, we propose a + +strategy that generates negative prompts via self-training, as illustrated in Fig. 1. Specifically, we first conduct initial score propagation for $T$ iterations following Eq. 3, giving rise to initial propagated scores $S_{T}$ . We select the top $m\%$ most confident scores from $S_{T}$ following Eq. 5 where $\mathrm{top}(S,m\%)$ indicates the top $m\%$ value of $S$ . + +$$ +\mathcal{I}_{pos} = \left\{i\mid S_{T}[i] > \operatorname {top}(S_{T},m\%)\right\} , +$$ + +$$ +\mathcal {I} _ {n e g} = \left\{i \mid S _ {T} [ i ] < \operatorname {top} \left(S _ {T}, 1 - m \%\right) \right\} \tag{5} +$$ + +We label the selected samples as positive and negative prompts respectively and re-initialize the score $\tilde{S}_0$ . For negative prompts, we assign scores of $-1$ to further increase the gap between ID and OOD samples. The final score is obtained by running Eq. 3 with $\tilde{S}_0$ as initial scores. + +$$ +\tilde {S} _ {0} [ i ] = \left\{ \begin{array}{l l} S _ {0} [ i ] & \text {i f} i \notin \mathcal {I} _ {\text {p o s}} \& i \notin \mathcal {I} _ {\text {n e g}}, \\ + 1 & \text {i f} i \in \mathcal {I} _ {\text {p o s}}, \\ - 1 & \text {i f} i \in \mathcal {I} _ {\text {n e g}}. \end{array} \right. \tag {6} +$$ + +Graph Construction: We now introduce the approach for constructing a graph for score propagation. As observed by [31, 44], the distance between text prompts and visual samples is generally larger than the distance between visual samples. This observation also holds for 3D point cloud VLM because the 3D VLMs are often finetuned from 2D VLMs [54, 55]. Therefore, we adopt a similar approach to construct the graph. We write the weight matrix $W$ in a blockwise form as shown in Eq. 7, where $W_{u} \in \mathbb{R}^{|\mathcal{D}_{u}| \times |\mathcal{D}_{u}|}$ , $W_{up} \in \mathbb{R}^{|\mathcal{D}_{u}| \times |\tilde{\mathcal{P}}|}$ and $W_{ul} \in \mathbb{R}^{|\mathcal{D}_{u}| \times |\mathcal{D}_{l}|}$ . No connections are created between prototypes and few-shot labeled samples. All blocks are further constructed as a k-nearest neighbor (KNN) graph following [31, 44]. It is also computationally efficient to adapt an existing graph to new testing samples by incrementally adding the testing samples to the graph [31]. + +$$ +W = \left[ \begin{array}{c c c} \mathbf {I} ^ {| \tilde {\mathcal {P}} | \times | \tilde {\mathcal {P}} |} & \mathbf {0} & W _ {u p} ^ {\top} \\ \mathbf {0} & \mathbf {I} ^ {| \mathcal {D} _ {l} | \times | \mathcal {D} _ {l} |} & W _ {u l} ^ {\top} \\ W _ {u p} & W _ {u l} & W _ {u} \end{array} \right] \tag {7} +$$ + +Overall Algorithm: We present the overall algorithm for unified iterative label propagation for vision-language models in Algorithm 1. In Fig. 2, we show an overview of the proposed framework for zero-/few-shot OOD detection via graph propagation. + +Algorithm 1 Graph Score Propagation for OOD Detection +1: Input: training data $\mathcal{D}_l$ , testing data $\mathcal{D}_u$ and ID class prompt templates $\mathcal{T}_c$ +2: Output: Final OOD scores $\tilde{S}_T$ +3: for $\mathcal{D}_l$ and $\mathcal{T}_c$ , do + Encode: $z_l = f(x_l)$ for training data and $\mathcal{P}_c = \{g(t_j(y_c))\}$ for prompt templates. Prompt clustering: positive prototypes $\tilde{\mathcal{P}}$ (Eq. 4) +4: end for +5: for $\mathcal{D}_u$ , do + Encode: $z_i = f(x_i)$ for testing data + Build graph to compute $W$ (Eq. 7) + Initialize score $S_0^{(i)}$ (Eq. 3) + Score propagate to compute $S_T$ (Eq. 3) + Self-training to generate positive $\mathcal{I}_{pos}$ and negative prompts $\mathcal{I}_{neg}$ (Eq. 5) + Updated initial scores $\tilde{S}_0$ (Eq. 6) + Score propagate to compute $\tilde{S}_T$ (Eq. 3) +6: end for +7: return $\tilde{S}_T$ + +# 4. Experiments + +# 4.1. Experimental setup + +Dataset: In this work, we follow the 3DOS framework [1] to construct downstream datasets for evaluating our method. The 3DOS benchmark is based on two well-known 3D object datasets, ShapeNetCore [9] and ScanObjectNN [49]. ShapeNetCore includes 54 categories covering daily objects. ScanObjectNN is cropped out from ScanNet dataset featuring a real-world objects. It consists of 15 categories with 2,902 3D objects. To further validate our approach on real-world data, we introduce the Sydney Urban Object Dataset [11], nuScenes [7] and S3DIS [4]. Sydney Urban Object Dataset consists of LiDAR-scanned point clouds across 26 categories, totaling 631 unique instances. S3DIS dataset contains 6 large-scale indoor areas with 271 rooms annotated into 13 semantic categories. We cropped out individual objects using ground-truth segmentation masks, leading to 8931 objects. The nuScenes Dataset features point clouds captured by LiDAR sensor under autonomous driving scenario. We cropped out point clouds using ground-truth bounding boxes, leading to 2205 + +objects across 9 categories. Detailed visualization of each dataset is deferred to the supplementary. + +Evaluation Protocol: Following the methodology outlined in 3DOS [1], we divide ShapeNetCore into three non-overlapping category sets, SN1, SN2, and SN3, referred to as the ShapeNetCore dataset. Each set contains 18 categories. For each OOD test scenario, one of these sets (SN1, SN2, or SN3) is treated as the known set of categories, while the remaining two sets are used as unknown categories, resulting in three distinct OOD test cases. Similarly, 3DOS defines three category sets, SR1, SR2, and SR3 on the ScanObjectNN dataset. Each SR-set includes 5 non-overlapping categories, with the other two sets acting as unknown categories in each scenario and call them as ScanObjNN. As for Sydney Urban Object Dataset, we categorize these objects into two groups: movable objects, e.g. "truck" "pedestrian" and non-movable objects, e.g. "building" and "traffic sign". Movable objects are treated as in-distribution (ID) classes, simulating a scenario in autonomous driving where object detection primarily focuses on movable objects. For consistency, we pre-process each point cloud sample by upsampling it to a uniform number of points. For S3DIS dataset, We retained instances containing more than 2048 original points, where each instance was randomly downsampled to retain exactly 2048 points. Then we set the foreground objects as ID and back-ground objects as OOD. For nuScences dataset, we have conducted preliminary screening on the original point cloud and retained instances with more than 200 original points and upsampling it to 1024 points. Then we define Four-wheeled vehicles (including car, bus, truck, construction_vehicle, trailer) as ID and the rest as OOD. And further details provided in the supplementary material. + +Implementation Details. We use the pre-trained ULIP/ULIP2 [54, 55] model to extract point cloud features and text features. For the 3D encoder, we select PointBERT [56]. In constructing the KNN graph, we set a universal $\mathrm{K}$ -value $(N_{k})$ of 10, connecting each text feature to its 10 nearest point cloud features. In Eq. 3, we set $\alpha$ to 0.5 and the number of iterations $T$ to 5. The percentage of pseudo negative prompts, i.e. $m\%$ in Sect. 3.3, is also set to $5\%$ . Additionally, we use K-means clustering $(N_{c})$ to group the prompts for each category into 3 clusters. + +Competing Methods. We compare our method with several OOD detection approaches, including full-shot methods from 3DOS [1]. We also benchmark against recent zero-shot methods, such as DDCS [57], ZLaP[44], MCM [36], and NegLabel [24]. For zero-shot comparisons, all methods use the 3D encoder from [55]. Baseline results are also reported by directly using cosine similarity as the OOD score, with these methods referred to as ULIP and ULIP2, depending on the VLM used. + +Evaluation Metrics. We evaluate OOD detection perfor + +
MethodScanObjNNShapeNetCore
SR3AUROC↑FPR95↓SR2AUROC↑FPR95↓SR1AUROC↑FPR95↓AverageAUROC↑FPR95↓SN1AUROC↑FPR95↓SN2AUROC↑FPR95↓SN3AUROC↑FPR95↓AverageAUROC↑FPR95↓
Requiring training datasetMSP[15]88.167.380.684.073.780.380.877.274.382.880.078.189.752.281.371.0
MLS[50]89.453.883.473.176.475.383.067.472.080.883.964.189.840.581.961.8
ODIN[32]90.247.983.371.776.376.883.365.574.279.479.471.787.841.880.564.3
Energy[33]89.550.681.675.876.675.582.667.372.181.283.064.789.839.482.061.8
GradNorm[19]88.550.777.475.375.276.880.467.672.181.857.788.957.879.062.683.3
ReAct[46]90.348.982.475.875.477.682.767.473.779.489.652.195.027.286.152.9
NF88.047.780.668.275.681.481.465.881.572.571.178.091.049.681.266.7
OE+mixup[16]72.683.572.088.562.587.869.086.672.778.980.368.887.362.280.169.9
ARPL+CS[10]--------74.880.380.772.485.450.880.367.8
Cosine protoCE91.041.082.178.277.675.683.664.980.368.388.760.891.938.086.955.7
SupCon[25]--------80.975.583.568.285.145.183.262.9
SubArcface[12]87.161.378.976.973.781.479.973.279.081.282.960.389.132.883.758.1
GSP(1-shot)79.361.380.984.188.651.782.965.791.344.086.443.183.891.487.259.5
GSP(5-shot)80.754.890.147.289.140.286.647.495.619.294.622.197.012.595.717.9
GSP(full-shot)90.133.093.030.592.234.491.822.697.910.197.97.998.86.098.28.0
Zero-shotULIP[54]79.982.762.387.557.492.366.587.588.057.176.448.775.251.379.952.4
ULIP2[55]65.780.876.471.277.478.073.276.785.752.584.144.367.674.479.157.1
DDCS[57]63.390.562.480.064.496.263.488.961.186.766.886.577.276.268.483.1
MCM[36]52.092.362.391.381.380.465.288.085.151.683.246.166.475.878.257.8
NegLabel[24]65.785.566.891.367.488.266.688.360.687.880.678.688.048.376.471.6
ZLaP[44]70.181.767.084.767.883.668.383.388.252.872.366.477.490.079.369.7
GSP75.863.076.577.683.661.678.667.490.638.970.764.079.793.780.465.5
+ +mance using two widely adopted metrics: AUROC and FPR95. AUROC calculates the area under the receiver operating characteristic curve, with higher values indicating better performance. FPR95 measures the false positive rate at a true positive rate of $95\%$ , with lower values reflecting improved performance. + +# 4.2. Results on OOD Detection + +Analysis of ScanObjNN Results: From the results presented in Tab. 1, we derive several key insights. i) Our proposed method demonstrates highly competitive performance under the zero-shot protocol, achieving an improvement of over $5\%$ in AUROC compared to ULIP2, which uses cosine distance directly as an OOD score. This substantial enhancement suggests that employing a manifold-based approach significantly enhances the differentiation between OOD samples and ID classes. ii) Our few-shot results are equally impressive, rivaling even fully supervised methodologies. Notably, with 5-shot training samples, our method attains an average AUROC of $86.6\%$ , surpassing the best fully supervised method, Cosine proto, which scores $83.6\%$ , by a margin of $3\%$ . Furthermore, when integrating all available training samples into the graph, our approach achieves an outstanding average AUROC of $91.8\%$ . iii) A noticeable $5\%$ gap in average AUROC is observed between our zero-shot and 1-shot methods. This gap underscores the critical role of incorporating real in-distribution samples in effectively distinguishing OOD samples. Consequently, practitioners are advised to introduce at least a few ID samples for OOD detection, in addition to leveraging Vision-Language Models (VLMs). + +Analysis of ShapeNetCore Results: The results for ShapeNetCore, as shown in Tab. 1, lead us to several ob + +Table 1. Evaluation of OOD detection for ScanObjNN and ShapeNetCore datasets. Each column title for SR/SN indicates the chosen known class set and the rest sets of SR/SN are unknown. + +
MethodSydneynuScenesS3DIS
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
ULIP[54]63.481.753.293.353.195.1
ULIP2[55]66.283.764.293.062.592.3
DDCS[57]61.887.362.291.262.592.5
MCM[36]65.782.457.182.365.288.1
NegLabel[24]58.492.769.781.655.485.4
ZLaP[44]58.580.670.087.971.388.4
GSP72.568.980.880.777.694.0
+ +Table 2. Zero-shot OOD detection results on the Sydney Urban Objects dataset, nuScenes dataset and S3DIS dataset. + +servations. i) Unlike the findings from the ScanObjNN dataset, the performance gap between the leading baseline, ULIP, and our method under the zero-shot protocol is relatively narrow, approximately $0.5\%$ . This modest improvement can be attributed to the substantial overlap between the downstream and pre-training datasets used in Vision-Language Models (VLMs). Specifically, ULIP/ULIP2 are pre-trained on ShapeNet, which shares significant commonalities with ShapeNetCore, thereby impacting OOD detection capabilities. Despite this, the slight enhancement achieved by our method underscores the utility of leveraging data manifolds to improve the differentiation of OOD samples. ii) A marked enhancement in performance is observed when few-shot samples are available. For instance, introducing 1-shot training samples results in nearly a $7\%$ increase in average AUROC, while 5-shot samples yield an impressive $15\%$ improvement. This highlights the critical role of few-shot learning in enhancing model performance. + +Analysis of Additional Real-World Datasets: The zero-shot OOD detection results for the Sydney Urban Objects dataset [11] and nuScenes Dataset [7], presented in Tab. 2, reinforce our previous conclusions. Our method consistently outperforms all competing approaches in zero-shot OOD detection, demonstrating the efficacy of utilizing test + +ing data manifolds, particularly when there is a substantial shift between the downstream task and the pre-training data. This finding emphasizes the robustness of our approach in adapting to significant distributional shifts in real-world applications. For S3DIS [4], GSP achieves the highest AUROC, outperforming all baselines by a significant margin (e.g., +12.4% over MCM). This suggests superior ability to distinguish between classes across all decision thresholds. However, GSP's FPR95 (94.0%) is slightly higher than ULIP (95.1%), indicating that while GSP excels in overall classification, it faces challenges in suppressing false positives when operating at a 95% true positive rate. + +Further Insights Synergizing All Results: By comparing the outcomes from the ScanObjNN, Sydney Urban Objects, S3DIS, nuScenes and ShapeNetCore datasets, we derive several overarching insights. i) There is a notable correlation between the distribution shift from pre-training to downstream datasets and the efficacy of manifold exploitation for OOD detection. This can be attributed to several factors. Firstly, when pre-training is conducted on datasets that closely resemble the downstream datasets, the contrastive training loss effectively aligns text prompts with visual features, resulting in a well-defined separation between ID and OOD samples. Conversely, in scenarios where the distribution shift is more substantial, VLMs may struggle to accurately quantify the distance between ID text prompts and visual samples. In such cases, leveraging the testing data manifold becomes essential to enhance the calibration of prompt-to-visual feature distances, thereby improving OOD detection performance. + +# 4.3. Ablation Study + +We begin by analyzing the performance of three Out-of-Distribution (OOD) detection methods: Cosine Similarity, Shortest Path, and Label Propagation. As shown in Tab. 3, we employ the ULIP2 backbone for extracting text and point cloud features and evaluate these methods for OOD scoring. The approach of directly computing the maximal normalized cosine similarity between text and point cloud features yields a baseline performance. Substituting cosine similarity with the shortest path metric results in a performance increment, albeit modest. It is noteworthy that the shortest path is computed using Dijkstra's algorithm, and the reciprocal of manifold distance serves as the OOD score. Lastly, we introduce a label propagation method to determine the OOD score. The experimental outcomes in Tab. 3 demonstrate that integrating graph-based information via label propagation significantly enhances the model's ability to differentiate between In-Distribution (ID) and OOD categories. This suggests that graph-based methods provide a more nuanced understanding of the feature space, leading to improved OOD detection capabilities. + +
OOD Det. Mthd.Prompt Clust.Neg. PromptAUROC↑FPR95↓
Cosine Dist.--73.276.7
Manifold Dist.--74.173.0
Score Prop.--76.472.6
Score Prop.✓-77.273.6
Score Prop.-✓77.371.4
Score Prop.✓✓78.667.4
+ +Table 3. Ablation study on ScanObjNN dataset. + +![](images/45c2ee79d7002bde5cc5bc02aebdc3488f57fab29dc063a33b492b1262781da5.jpg) +(a) + +![](images/f1c1471fe75611a45cd261888034d8e51f3934b9cdf6e8d4f5b1c617e7a18762.jpg) +(b) +Figure 3. Few-shot OOD detection on the ScanObjNN dataset. (a) Results on AUROC metric. (b) Results on FPR95 metric. + +# 4.4. Further Analysis + +Few-shot OOD Detection: To evaluate the performance of label propagation in few-shot OOD detection, we incorporated a subset of training data from the 3DOS training dataset, treating them as in-distribution categories. The outcomes, presented in Fig. 3, show the effects of varying the number of training samples used in label propagation, with all samples randomly selected from the training dataset. The results indicate that as the number of training samples increases, OOD detection performance improves and eventually stabilizes. + +Qualitative Analysis: To validate our method, we visualize sample predictions and scores from baseline method and our full GSP. As shown in Fig. 5, for easy-to-predict samples from the Real-world dataset, conventional distance-based methods perform well. However, in challenging cases (e.g., occlusions or missing parts), they often fail. In contrast, our method utilizes spatially proximate samples in feature space for score propagation, ensuring accurate OOD detection even in difficult scenarios. These qualitative results confirm the effectiveness and robustness of our approach. + +Impact of Prompt Clustering: To illustrate the benefits of prompt expansion via K-means clustering, we examined how varying the number of prompts affects OOD (Out-of-Distribution) detection. As depicted in Fig. 4 (a), we experimented with different values of $N_{c}$ in the K-means algorithm, where each $N_{c}$ value corresponds to a distinct number of clusters, and consequently, a different number of prompts. An increase in $N_{c}$ results in a greater number of distinct prompts. + +The results presented in the figure reveal that the quantity of prompts considerably impacts OOD detection performance. Specifically, when $N_{c} = 6$ , the AUROC surpasses $78.7\%$ . However, further increasing $N_{c}$ can negatively af + +![](images/eb35f117bd9b1174e6783df8ad2831aa7144b96479f7b73ae896783e36cee4ea.jpg) +(a) $N_{k}$ Nearest Neighbors + +![](images/2b9530dee9a3d2a6ccd4cf28519da4f2600f1519fae7f6989d4b826006a3063d.jpg) +(b) Number of negative prompt $m$ +Figure 4. Results from ScanObjNN Benchmark Experiments: (a) Evaluation of prompt clustering with different K-means cluster numbers. (b) Impact of varying negative prompts on label propagation. (c) Results of different number of K Neaerst Neighbors in KNN. +Figure 5. Visualization of some samples for ScanObjectNN and nuScenes Dataset with two methods predict scores. All scores are normalized to 0-1. + +![](images/65800647109a20c84af0d895f5cef84f7b817989c9daad2c147d02885165d623.jpg) +(c) Cluster Number $N_{c}$ + +
ScanObjectNN DatasetnuScenes Dataset
ID categoryEasy-to-predictHard-to-predictEasy-to-predictHard-to-predict
Distance-based score:0.6180.0930.6890.192
GSP score:0.4650.3570.3740.278
OOD category
Distance-based score:0.1740.3960.1480.431
GSP score:0.0720.0540.0740.127
+ +fect the spatial alignment between text features and point cloud features. This phenomenon is likely due to the fact that typical prompts in Vision-Language Models (VLMs) are generally standardized, with only occasional outliers. Excessive clustering may isolate these outliers, leading to a decline in performance. + +In conventional VLM approaches, the mean text feature vector is typically calculated for each category (equivalent to setting $N_{c} = 1$ ), achieving an average AUROC of less than $78\%$ . The findings in Fig. 4 (c) suggest that selecting an appropriate $N_{c}$ enhances the model's ability to differentiate OOD categories, underscoring the advantages of prompt clustering. + +Effectiveness of Negative Prompting via Self-Training: We conducted experiments to evaluate the effectiveness of the proposed negative prompting. Using the ShapeNetCore dataset, we explored the impact of self-training for negative prompting by varying the percentage of samples tagged with pseudo labels, denoted as $m\%$ as introduced in Section 3.3. In Fig. 4 (b), We observe a performance increase from $m = 1$ up to $m = 6$ , indicating that incorporating negative prompts through self-training is an effective strategy to leverage the manifold for score propagation. However, performance declines beyond $m = 6$ , likely due to the introduction of incorrect pseudo-negative prompts. This highlights the importance of mitigating the effects of confirmation bias [3] to prevent the model from being adversely + +
Modelload + DatasetloadText Feature ExtractPrompt ClusteringFeature ExtractBuild GraphScore PropagationTotal time
SR1552ms2212ms1128ms2780ms106ms8ms7786ms
SN1641ms2459ms1657ms25031ms671ms34ms31493ms
+ +Table 4. The inference time of the GSP on SR and SN. + +affected. + +Impact of K-Nearest-Neighbor. We compare performance with different values of $N_{k}$ Nearest Neighbors. As we can see in Fig 4 (a), the results demonstrate that as the number of $N_{k}$ values increases, the OOD detection performance is tend to be stable. While $N_{k} = 16$ achieve the best performance and Less than 8 or larger than 18 get the low results. + +Computational Cost: We profiled each step and test the inference time of GSP using a full GPU implementation in Tab. 4. Experiments are carried out on a single RTX3090 GPU and a i7-12700k CPU. As observed, the graph-based operations (Build Graph + Score Propagation) account for only around $2\%$ of the total inference time. The additional computation cost introduced by graph operations is negligible compared with point cloud feature extraction. + +# 5. Conclusion + +In this work, we presented a novel training-free approach to 3D point cloud OOD detection that effectively leverages Vision-Language Models through Graph Score Propagation. By constructing a graph that incorporates both positive prototypes and pseudo-negative samples via self-training, our method captures the underlying structure of the testing data manifold. This approach allows for the propagation of informative OOD scores, thus surpassing existing distance-based techniques. Empirical results across multiple 3D point cloud datasets demonstrate the robustness and adaptability of our framework, particularly in zero- and few-shot settings. The GSP method addresses key limitations in prior OOD approaches by reducing dependency on extensive in-distribution training data while achieving high accuracy. Future work could explore further enhancements through dynamic graph updates, enabling real-time OOD detection in continuously evolving environments. + +# Acknowledgments + +# References + +[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems, 2022. 2, 5 +[2] Mouin Ben Ammar, Nacim Belkhir, Sebastian Popescu, Antoine Manzanera, and Gianni Franchi. Neco: Neural collapse based out-of-distribution detection. arXiv preprint arXiv:2310.06823, 2023. 2 +[3] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks, 2020. 8 +[4] I Armeni, O Sener, Ar Zamir, H Jiang, I Brilakis, M Fischer, S Savarese, and IEEE. 3d semantic parsing of large-scale indoor spaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5, 7 +[5] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2 +[6] Tianyi Bao, Qitian Wu, Zetian Jiang, Yiting Chen, Jiawei Sun, and Junchi Yan. Graph out-of-distribution detection goes neighborhood shaping. In International Conference on Machine Learning. 2 +[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 5, 6 +[8] Tianle Cai, Ruiqi Gao, Jason Lee, and Qi Lei. A theory of label propagation for subpopulation shift. In International Conference on Machine Learning, 2021. 2 +[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5 +[10] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern analysis and Machine Intelligence, 2021. 1, 6 +[11] Mark De Deuge, Alastair Quadros, Calvin Hung, and Bertrand Douillard. Unsupervised feature learning for classification of outdoor 3d scans. In Australasian Conference on Robotics and Automation, 2013. 5, 6 + +This work was supported by the National Natural Science Foundation of China (No. U2468207), Sichuan Science and Technology Program (No. 2024NSFTD0036), and the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (No. M23L7b0021). +[12] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In European Conference on Computer Vision, 2020. 6 +[13] Edsger W Dijkstra. A note on two problems in connexion with graphs. In Edsger Wybe Dijkstra: his life, work, and legacy. 2022. 3 +[14] Shreyas Havaldar, Navodita Sharma, Shubhi Sareen, Karthikeyan Shanmugam, and Aravindan Raghuveer. Learning from label proportions: Bootstrapping supervised learners via belief propagation. arXiv preprint arXiv:2310.08056, 2023. 2 +[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 1, 2, 6 +[16] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In International Conference on Learning Representations, 2019. 6 +[17] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2 +[18] Chengjie Huang, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, Krzysztof Czarnecki, et al. Out-of-distribution detection for lidar-based 3d object detectabel propagation for zero-shot classification with vision. In IEEE International Conference on Intelligent Transportation Systems, 2022. 1, 2 +[19] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 2021. 6 +[20] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In IEEE/CVF International Conference on Computer Vision, 2023. 2, 3 +[21] Abiodun M. Ikatun, Absalom E. Ezugwu, Laith Abualigah, Belal Abuhaija, and Jia Heming. K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. Information science, 2023. 4 +[22] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2 +[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 1 +[24] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided ood detection with pretrained vision-language models. In International Conference on Learning Representations, 2024. 3, 5, 6 + +[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 6 +[26] Konstantin Kirchheim, Tim Gonschorek, and Frank Ortmeyer. Out-of-distribution detection with logical reasoning. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2 +[27] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus C. J. Dietmayer. Revisiting out-of-distribution detection in lidar-based 3d object detection. IEEE Intelligent Vehicles Symposium, 2024. 2 +[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 1 +[29] Jingyao Li, Pengguang Chen, Zexin He, Shaozuo Yu, Shu Liu, and Jiaya Jia. Rethinking out-of-distribution (ood) detection: Masked image modeling is all you need. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4 +[31] Yushu Li, Yongyi Su, Adam Goodge, Kui Jia, and Xun Xu. Efficient and context-aware label propagation for zero/few-shot training-free adaptation of vision-language model. In International Conference on Learning Representations, 2025. 2, 4 +[32] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 6 +[33] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 2020. 6 +[34] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In IEEE International Conference on Image Processing, 2021. 1 +[35] Guofeng Mei, Luigi Riz, Yiming Wang, and Fabio Poiesi. Geometrically-driven aggregation for zero-shot 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2 +[36] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In Advances in Neural Information Processing Systems, 2022. 1, 3, 5, 6 +[37] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 2024. 1, 2 +[38] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In International Conference on Learning Representations, 2024. 2, 4 + +[39] Julia Nitsch, Masha Itkina, Ransalu Senanayake, Juan Nieto, Max Schmidt, Roland Siegwart, Mykel J Kochenderfer, and Cesar Cadena. Out-of-distribution detection for automotive perception. In IEEE International Intelligent Transportation Systems Conference, 2021. 1 +[40] Rattana Pukdee, Dylan Sam, Maria-Florina Balcan, and Pradeep Ravikumar. Label propagation with weak supervision. arXiv preprint arXiv:2210.03594, 2022. 2 +[41] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 2022. 1 +[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2 +[43] Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Gunnemann. Graph posterior network: Bayesian predictive uncertainty for node classification. Advances in Neural Information Processing Systems, 2021. 2 +[44] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6 +[45] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions. arXiv preprint arXiv:2201.12296, 2022. 2 +[46] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems, 2021. 6 +[47] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, 2022. 2 +[48] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4 +[49] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019. 5 +[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 6 +[51] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In IEEE/CVF International Conference on Computer Vision, 2023. 1, 2 +[52] Qizhou Wang, Zhen Fang, Yonggang Zhang, Feng Liu, Yixuan Li, and Bo Han. Learning to augment distributions for out-of-distribution detection. Advances in Neural Information Processing Systems, 2023. 2 + +[53] Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. Energy-based out-of-distribution detection for graph neural networks. ArXiv, 2023. 2 +[54] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5, 6 +[55] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 4, 5, 6 +[56] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5 +[57] Yue Yuan, Rundong He, Yicong Dong, Zhongyi Han, and Yilong Yin. Discriminability-driven channel selection for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5, 6 +[58] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 1 +[59] Ji Zhang, Jingkuan Song, Lianli Gao, Nicu Sebe, and Heng Tao Shen. Reliable few-shot learning under dual noises. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2 +[60] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2 +[61] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision. Springer, 2022. 2 +[62] Zihan Zhang and Xiang Xiang. Decoupling maxlogit for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2 +[63] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. ProQuest number: information to all users, 2002. 3 +[64] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In International Conference on Machine Learning, 2003. 2 \ No newline at end of file diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/images.zip b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..2ecd38e3b6e2779a2cb2a7fa0ce7c0abd96d32e9 --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:f0b3b43feda6f965e7f36993126f92968b4b9361490c43213b0dea4d23243732 +size 491940 diff --git a/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/layout.json b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..0947ad52cd15f9d0083b3242c86d0fd6f85eda49 --- /dev/null +++ b/2025/Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation/layout.json @@ -0,0 +1,9303 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "spans": [ + { + "bbox": [ + 75, + 102, + 536, + 140 + ], + "type": "text", + "content": "Exploiting Vision Language Model for Training-Free 3D Point Cloud OOD Detection via Graph Score Propagation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "spans": [ + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": "Tiankai Chen" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Yushu Li" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{23}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Adam Goodge" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Fei Teng" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Xulei Yang" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Tianrui Li" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": ", Xun Xu" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{\\dagger 3}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": "Southwest Jiaotong University, " + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": "South China University of Technology \n" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": "Institute for infocomm research(I" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 83, + 160, + 527, + 203 + ], + "type": "text", + "content": "R), A*STAR" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 256, + 297, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 256, + 297, + 473 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 297, + 473 + ], + "type": "text", + "content": "Out-of-distribution (OOD) detection in 3D point cloud data remains a challenge, particularly in applications where safe and robust perception is critical. While existing OOD detection methods have shown progress for 2D image data, extending these to 3D environments involves unique obstacles. This paper introduces a training-free framework that leverages Vision-Language Models (VLMs) for effective OOD detection in 3D point clouds. By constructing a graph based on class prototypes and testing data, we exploit the data manifold structure to enhancing the effectiveness of VLMs for 3D OOD detection. We propose a novel Graph Score Propagation (GSP) method that incorporates prompt clustering and self-training negative prompting to improve OOD scoring with VLM. Our method is also adaptable to few-shot scenarios, providing options for practical applications. We demonstrate that GSP consistently outperforms state-of-the-art methods across synthetic and real-world datasets for 3D point cloud OOD detection." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 135, + 510 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 518, + 296, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 518, + 296, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 518, + 296, + 687 + ], + "type": "text", + "content": "The detection of out-of-distribution (OOD) samples in 3D point cloud data is critical for the safe and robust deployment of deep learning models in real-world environments. For example, identifying unexpected objects is crucial for ensuring the safety of personnel, vehicles, and robots by preventing collisions [18, 39, 41]. Existing approaches to OOD detection in 3D point clouds often focus on training separate models for specific tasks [54]. These methods typically rely on abundant in-distribution (ID) training data to build models based on reconstruction [34], distance metrics [10], or discriminative strategies [15]. While these methods achieve strong results, they depend on substantial ID training data, which is often challenging to collect in many real-world applications and generalization may be" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 317, + 232, + 550, + 413 + ], + "blocks": [ + { + "bbox": [ + 317, + 232, + 550, + 413 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 550, + 413 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 550, + 413 + ], + "type": "image", + "image_path": "fd65801fa754b791ecc7ed43d2b422ae763933e9b0384fb59d3c4e4d8e329d58.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 415, + 555, + 459 + ], + "lines": [ + { + "bbox": [ + 313, + 415, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 555, + 459 + ], + "type": "text", + "content": "Figure 1. Exploiting the testing data manifold for zero-/few-shot OOD detection. ID prototypes propagate scores to test samples, with confident ones selected as pseudo negative prompts for refinement." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 472, + 555, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 555, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 555, + 508 + ], + "type": "text", + "content": "hindered by the limited diversity of ID training samples. These limitations highlight the need for alternative solutions that do not require extensive ID data collection." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 509, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 509, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 509, + 556, + 664 + ], + "type": "text", + "content": "Vision-language models (VLMs) have emerged as powerful foundation models that bridge visual features with textual understanding [23, 28, 42, 58]. VLMs are trained to associate the content of visual inputs with their textual labels and they demonstrate strong zero-shot generalization to unseen visual content [42]. This capacity makes VLMs appealing for OOD detection without retraining [5, 36, 37, 51]. A common VLM-based OOD detection pipeline maps ID class concepts, such as class names with various templates, into a shared feature space using the text encoder. The visual samples, similarly projected using the visual encoder, are then scored by comparing their distances to the ID class concepts in this feature space [36]." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "This VLM-based OOD detection pipeline has shown competitive performance for 2D image detection, primarily due to the vast understanding of 2D VLMs from the massive-scale image-text pre-training datasets (e.g. 400M" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 63, + 696, + 280, + 716 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 696, + 280, + 716 + ], + "spans": [ + { + "bbox": [ + 63, + 696, + 280, + 716 + ], + "type": "inline_equation", + "content": "\\dagger" + }, + { + "bbox": [ + 63, + 696, + 280, + 716 + ], + "type": "text", + "content": " Correspondence to Xun Xu: xu_xun@i2ra.a-star.edu.sg, Code link: https://github.com/handsome999KK/GSP_OOD" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 759 + ], + "type": "text", + "content": "28797" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "pairs for pre-training CLIP [42]). In contrast, VLMs for 3D point clouds, such as the state-of-the-art ULIP2 [55], are typically pre-trained on much more restricted datasets (e.g., 800k point cloud samples). This restricts their ability to generalize across unseen categories and distributional shifts (e.g., style changes), which reduces the effectiveness of similarity-based methods for 3D point cloud OOD detection." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 55, + 169, + 294, + 335 + ], + "type": "text", + "content": "In this work, we propose to leverage the structure of the testing data manifold to compensate for this deficiency and improve VLM adaptation for downstream tasks [31, 44], as illustrated in Fig. 1. Specifically, we construct a graph based on class prototypes (e.g., text prompts for ID classes) and testing data features, where prototype nodes are assigned labels that propagate to neighboring testing nodes. Given the limited training data, we hypothesize that exploiting the testing data manifold enhances the generalization capabilities of 3D point cloud VLMs. Our method, Graph Score Propagation (GSP), assigns a positive OOD score (e.g., +1) to ID class prototypes while all testing samples start with a neutral score (0). As scores propagate along the graph, the resultant scores indicate OOD likelihood." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 338, + 295, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 338, + 295, + 575 + ], + "spans": [ + { + "bbox": [ + 55, + 338, + 295, + 575 + ], + "type": "text", + "content": "This straightforward adoption of score propagation significantly improves upon existing distance-based methods, and we introduce two additional enhancements to further boost OOD detection accuracy. First, while conventional training-free approaches typically use a single class prototype by averaging text prompts [44, 61], we argue that capturing the prompt distribution through clustering better leverages prompt diversity. Second, negative prompting has shown strong performance in VLM-based OOD detection [30, 38, 48]. However, simply negating the description (e.g. add \"not\") is not very effective and learning negative prompts requires sophisticated optimization. Instead, we employ self-training to convert high-confidence OOD testing samples into pseudo-negative prompts and assign them a negative score (e.g., -1), for subsequent propagation iterations. Finally, our framework integrates seamlessly into traditional supervised OOD detection protocols by incorporating a few ID samples as positive prompts, further enhancing OOD detection under both zero-shot and few-shot protocols." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 578, + 282, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 578, + 282, + 590 + ], + "spans": [ + { + "bbox": [ + 67, + 578, + 282, + 590 + ], + "type": "text", + "content": "Our contributions are mainly summarized as follows." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 7, + "blocks": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "- We employ a pre-trained vision-language model for 3D point cloud OOD detection and introduce Graph Score Propagation (GSP), a novel methodology that leverages the testing data manifold instead of solely relying on text-test feature distances." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 295, + 689 + ], + "type": "text", + "content": "- We enhance GSP with prompt clustering and self-training-based negative prompting. GSP is also compatible with both zero-shot and few-shot testing protocols." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "- We validate the effectiveness of our method on synthetic and real-world 3D point cloud OOD detection bench" + } + ] + } + ], + "index": 6 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 323, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 323, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 323, + 72, + 555, + 95 + ], + "type": "text", + "content": "marks, consistently achieving superior performance over state-of-the-art VLM-based OOD detection methods." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 107, + 400, + 119 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 107, + 400, + 119 + ], + "spans": [ + { + "bbox": [ + 314, + 107, + 400, + 119 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 127, + 553, + 378 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 127, + 553, + 378 + ], + "spans": [ + { + "bbox": [ + 313, + 127, + 553, + 378 + ], + "type": "text", + "content": "Out-of-Distribution Detection: Traditional OOD detection methods focused on softmax-based probability scoring [15]. Recently, attention has shifted to novel evaluation metrics [26, 47, 57, 62], such as differential network channel processing [57, 62] and non-parametric nearest-neighbor distances [47]. Other strategies involve training specific OOD discriminators, like DAL [52], which constructs an OOD distribution set, and MOOD [29], which uses masked image modeling. NECO [2] utilizes neural collapse geometry for OOD identification. Vision-language models have introduced new methods using text prompts to distinguish between in-distribution (ID) and OOD data [5, 30, 37, 38, 51]. LoCoOP [37] leverages CLIP's local features for OOD regularization, while Bai et al. [5] use CLIP to detect ID-like outliers, enhancing OOD detection and few-shot learning [59]. Training text encoders [51] and learnable prompt [30, 38] further increases the score gap between ID and OOD data. Recent research has expanded to include 3D datasets for point cloud OOD detection [1, 45]. Studies have shown that 3D OOD detection is particularly important for autonomous driving scenarios [18, 27]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 379, + 553, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 379, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 379, + 553, + 510 + ], + "type": "text", + "content": "3D Vision-Language Models: Recent advancements in Vision-Language Models (VLMs) have facilitated innovative approaches for 3D understanding. Techniques utilizing CLIP [20, 35, 60] transform 3D point clouds into imagery data for extracting features aligned with text descriptions. More recently, methods like ULIP [54, 55] learn unified representations across images, text, and 3D point clouds, further improving zero-shot 3D classification. These approaches leverage the integration of text prompts and visual features which could better distinguish between indistribution and out-of-distribution data in 3D contexts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 713 + ], + "type": "text", + "content": "Label Propagation: Label propagation (LP) is a pivotal method in semi-supervised learning, leveraging limited labeled data to infer labels for unlabeled instances by exploiting data similarity across a graph [8, 14, 22, 40]. It assumes that labels vary smoothly across the graph, making it effective in transductive learning settings [22, 64]. Recent advancements integrate LP with graph neural networks to enhance its ability to separate ID from OOD data [6, 43, 53]. In the context of VLMs, LP employs text prompts as label representations, constructing graphs that propagate labels from text prototypes to test samples, thus improving zero-shot 2D classification [17, 31, 44]. Contrary to the existing attempts of combining LP with VLM, we tackle a one-class classification task. Only prototypes for in-distribution classes are available. We further propose a novel negative prompting technique via self-training and prompt clustering technique to better separate ID from OOD." + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28798" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 138, + 85 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 180, + 103 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 180, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 180, + 103 + ], + "type": "text", + "content": "3.1. Problem Formulation" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": "We begin by formally defining the task of VLM-based out-of-distribution (OOD) detection in 3D point clouds. Let " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "z^{f} = f(x;\\theta)" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "z^{t} = g(t;\\phi)" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " denote the encoded features of point cloud data " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "x" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " and text " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "f(\\cdot ;\\theta)" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "g(\\cdot ;\\phi)" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " are the VLM visual and text encoders respectively. The task involves the encoded features of unlabeled test data " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u = \\{f(x_i)\\}_{i = 1\\dots N_u}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " and, optionally, the features and labels of few-shot labeled data " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l = \\{f(x_j),y_j\\}_{j = 1\\dots N_l}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": ". The set of in-distribution (ID) classes " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{Y}_{in} = \\{1,\\dots ,C_{in}\\}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " is assumed to be known. Following established practices for VLM adaptation [36], we represent each ID class with " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " prompt templates " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{T} = \\{t_j(y)\\}_{N_t}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": ", e.g., \"This is a point cloud of " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{y}\\}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": "\". \"A good point cloud of " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\{\\mathrm{y}\\}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": "\". These textual prompts are encoded by the text encoder, creating prototypes for ID classes " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": ". Existing works often create prototypes by averaging the encoded prompts for each class, resulting in a prototype " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "z_c^p" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": "-th class, and the full set of prototypes is denoted " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P} = \\{z_c^p\\}_{c = 1\\dots C_{in}}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": ". The OOD detection task is to determine whether a test sample " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "x_{i}\\in \\mathcal{D}_{u}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " belongs to an ID class, leveraging ID prototypes " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": " and, optionally, the few-shot labeled data " + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 55, + 109, + 296, + 361 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 296, + 410 + ], + "type": "text", + "content": "Preliminary of VLM: We briefly review the training objective of vision-language models (VLMs). The pre-training steps align the visual features with their associated text features [20, 54, 55], typically via a contrastive loss function:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 105, + 420, + 295, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 420, + 295, + 452 + ], + "spans": [ + { + "bbox": [ + 105, + 420, + 295, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {v l m} = \\sum_ {i} \\log \\frac {\\exp \\left(z _ {i} ^ {f} z _ {i} ^ {t} / \\tau\\right)}{\\sum_ {j} \\exp \\left(z _ {i} ^ {f} z _ {j} ^ {t} / \\tau\\right)} \\tag {1}", + "image_path": "05e25ff0cd8613acd3a3137c2cb5efdee5934d1cc1ff477ac8c86afc027bc0ca.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 567 + ], + "type": "text", + "content": "When a VLM is trained on a large amount of point cloud-text pairs, the cosine similarity of positive pairs (a point cloud sample and text prompt of its ground-truth class) is higher than that of negative pairs. Therefore, a pre-trained VLM provides a straightforward solution to 3D point cloud OOD detection by measuring the cosine similarity between textual descriptions of in-distribution classes and testing sample features. A high distance in the visual-text feature space suggests a greater likelihood of being OOD." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "spans": [ + { + "bbox": [ + 55, + 574, + 195, + 586 + ], + "type": "text", + "content": "3.2. VLM for OOD Detection" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "text", + "content": "Vanilla OOD Detection with VLM: Pre-trained VLMs offer an effective approach to OOD detection by measuring the distance between visual samples and text labels in the feature space. For example, [36] employs the maximum normalized cosine similarity as the OOD score. We define a normalized score " + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "text", + "content": " to indicate OOD-ness, where " + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "inline_equation", + "content": "S \\rightarrow 0" + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "text", + "content": " suggests OOD and " + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "inline_equation", + "content": "S \\rightarrow 1" + }, + { + "bbox": [ + 55, + 593, + 296, + 677 + ], + "type": "text", + "content": " suggests ID." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 105, + 687, + 295, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 687, + 295, + 717 + ], + "spans": [ + { + "bbox": [ + 105, + 687, + 295, + 717 + ], + "type": "interline_equation", + "content": "S = \\max _ {c} \\frac {\\exp \\left(- d \\left(z ^ {f} , z _ {c} ^ {p}\\right) / \\tau\\right)}{\\sum_ {j} \\exp \\left(- d \\left(z ^ {f} , z _ {j} ^ {p}\\right) / \\tau\\right)} \\tag {2}", + "image_path": "817ee5e897909741339534415807b1500960b282a8369621898854f0974e4174.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 179 + ], + "type": "text", + "content": "It is assumed that in the feature space, OOD samples are further from the ID prototypes than ID samples. Subsequent research has further explored this approach by introducing techniques like negative prompts to better separate OOD from ID samples [24]. In this work, we are motivated by the challenge of measuring similarity/distance in the VLM feature space and propose to tackle the VLM OOD detection problem by exploiting the manifold of testing and, optionally, few-shot labeled data." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 240 + ], + "type": "text", + "content": "Manifold Distance for OOD Detection: The vanilla measurement overlooks the complex manifold of testing data. A straightforward approach to consider testing data manifold for OOD detection is to replace the standard distance metric with a manifold distance, represented via a graph." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": "Given a graph representing the data manifold " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "G = (E, V)" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": " with weight matrix " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "W \\in \\mathbb{R}^{(|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|) \\times (|\\mathcal{P}| + |\\mathcal{D}_l| + |\\mathcal{D}_u|)}" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": ", the manifold distance between two nodes on the graph is defined as the shortest path. The shortest path between arbitrary nodes can be obtained via Dijkstra's algorithm [13]. More concretely, this can be formulated as finding a path between a prototype " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "z^p \\in \\mathcal{P}" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": " or a few-shot labeled sample " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "z_k^f \\in \\mathcal{D}_l" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": " and a testing sample " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "z_j^f \\in \\mathcal{D}_u" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": ", consisting of edges " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "\\tilde{E}_{ij} = \\{e_k\\}" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": " on the graph, such that the sum of these edge weights is minimized. The manifold distance between two nodes is thus defined as " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "d(z_j, z_i) = \\min_{\\tilde{E}_{ij}} \\sum_{e_k \\in \\tilde{E}_{ij}} d(e_k)" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "d(e_k)" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": " denotes the distance for edge " + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "inline_equation", + "content": "e_k" + }, + { + "bbox": [ + 313, + 241, + 554, + 411 + ], + "type": "text", + "content": ", e.g. the L2 distance between the encoded features." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 412, + 554, + 520 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 520 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 520 + ], + "type": "text", + "content": "Our empirical results in Tab. 3 reveal that incorporating the manifold improves the effectiveness of OOD detection. However, there are two main disadvantages associated with distance-based OOD detection. First, the manifold distance is unnormalised, if directly using the reciprocal of manifold distance as OOD score the result may be biased by the relative scale. Additionally, it is non-trivial to integrate negative prompts within the distance-based OOD detection paradigm." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 529, + 539, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 539, + 542 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 539, + 542 + ], + "type": "text", + "content": "3.3. VLM OOD Detection as Score Propagation" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "spans": [ + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "text", + "content": "Instead of directly measuring the distance for OOD detection, we formulate OOD detection as a score propagation problem. Specifically, we assign a positive score " + }, + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "inline_equation", + "content": "(+1)" + }, + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "text", + "content": " to ID prompts " + }, + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 313, + 548, + 554, + 632 + ], + "type": "text", + "content": ", while all testing samples are initially assigned a neutral score (0). Labels are then propagated from the positive nodes to all testing samples in an iterative manner. We follow a standard propagation rule [63] as shown in Eq. 3," + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 360, + 642, + 447, + 654 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 642, + 447, + 654 + ], + "spans": [ + { + "bbox": [ + 360, + 642, + 447, + 654 + ], + "type": "interline_equation", + "content": "S _ {t} = \\tilde {W} S _ {t - 1} + \\alpha S _ {0},", + "image_path": "9241be1447e7d333dd89ed24a5c084618cbd3389ad5645ed59a5524f821981b8.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 339, + 658, + 553, + 685 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 339, + 658, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 339, + 658, + 553, + 685 + ], + "type": "interline_equation", + "content": "\\text {s . t .} \\tilde {W} = D ^ {- \\frac {1}{2}} W D ^ {- \\frac {1}{2}}, D = \\operatorname {d i a g} \\left(\\sum_ {j} W _ {i j}\\right), \\tag {3}", + "image_path": "6526782cbf5b25bb9deb32fea0f6b1bdd82f85b1db933518201e43c1da61182c.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 360, + 687, + 499, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 360, + 687, + 499, + 715 + ], + "spans": [ + { + "bbox": [ + 360, + 687, + 499, + 715 + ], + "type": "interline_equation", + "content": "S_{0} = \\bigl[\\underbrace{1,\\cdots,1}_{|\\mathcal{P}| + |\\mathcal{D}_{l}| \\text{elements}},\\underbrace{0,\\cdots,0}_{|\\mathcal{D}_{u}| \\text{elements}}\\bigr],", + "image_path": "3678b2607c08d7cfd109abf844ba69b141e1e4cbfe59adfe8ddafe5d95a6be95.jpg" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28799" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 549, + 190 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 549, + 190 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 549, + 190 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 549, + 190 + ], + "type": "image", + "image_path": "331e4e5e5dd294ddf6e887a4dbefd3b8fd4c66a668277220f12f5434e0979f36.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 195, + 555, + 229 + ], + "lines": [ + { + "bbox": [ + 54, + 195, + 555, + 229 + ], + "spans": [ + { + "bbox": [ + 54, + 195, + 555, + 229 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed framework for zero-/few-shot OOD detection via graph score propagation. The framework encodes ID category prompts and dataset samples using frozen text and 3D encoders, constructing a KNN graph. Initial score propagation generates pseudo prompts, which refine the graph before final score propagation computes OOD scores." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "text", + "content": "where each step of propagation updates the score for unlabeled testing samples as a weighted average of scores from adjacent nodes and initial scores " + }, + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "inline_equation", + "content": "S_0" + }, + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "text", + "content": " until " + }, + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 54, + 241, + 294, + 300 + ], + "type": "text", + "content": " iterations are completed. Consequently, nodes that are closer to the prototypes on the manifold will obtain a higher score." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": "Prompt Clustering: The above propagation strategy propagates labels from ID prototypes to testing samples on the graph. Thus, it is critical to design effective ID prototypes that capture the diversity of textual descriptions for ID classes. Contrary to existing practices that average the text prompt features across all prompt templates, we propose modeling the distribution of prompt templates with multiple representative ones. Given the prompt templates " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{T}" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ", we first initialize a pool of text prompts " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c = \\{z_{c_j}^p = g(t_j(y_c))\\}_{j=1\\dots|\\mathcal{T}|}" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": " for the " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": "-th class. We then introduce a clustering step to reduce this set to " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "N_c" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": " representative prototypes " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}}_c" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ". This is achieved by applying K-means [21] clustering on " + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c" + }, + { + "bbox": [ + 55, + 300, + 296, + 469 + ], + "type": "text", + "content": ", and the cluster centers are adopted as the ID prototypes:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 85, + 487, + 264, + 518 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 487, + 264, + 518 + ], + "spans": [ + { + "bbox": [ + 85, + 487, + 264, + 518 + ], + "type": "interline_equation", + "content": "\\tilde {\\mathcal {P}} _ {c} = \\arg \\min _ {\\{\\mu_ {k} \\}, \\gamma} \\sum_ {N _ {c}} \\sum_ {z _ {c j} ^ {p} \\in \\mathcal {P} _ {c}} \\gamma_ {j k} \\| z _ {c j} ^ {p} - \\mu_ {k} \\| ^ {2},", + "image_path": "d426abd032c80c1748f57b5d3c901f78d1d625514c7bb922b275115a7b576d4f.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 519, + 240, + 543 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 519, + 240, + 543 + ], + "spans": [ + { + "bbox": [ + 86, + 519, + 240, + 543 + ], + "type": "interline_equation", + "content": "\\begin{array}{l l} \\text {s . t .} & \\gamma \\in \\{0, 1 \\} ^ {| \\mathcal {T} | \\times N _ {c}}, \\sum_ {k} \\gamma_ {j k} = 1, \\end{array}", + "image_path": "2567dfd7b63c84e8edb3ae5cab2416bf270b3a0abdd6d9d63fbe79fb6b60eb46.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\mu_{k}" + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": " are the cluster centers, and " + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\gamma_{jk}" + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": " indicates the assignment of each prompt feature " + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "inline_equation", + "content": "z_{cj}^{p}" + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": " to its closest center. The new prototypes, " + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}} = \\tilde{\\mathcal{P}}_1\\cup \\tilde{\\mathcal{P}}_2\\cup \\dots \\tilde{\\mathcal{P}}_C" + }, + { + "bbox": [ + 55, + 544, + 295, + 593 + ], + "type": "text", + "content": " , are used for score propagation as positive nodes." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": "Negative Prompting via Self-Training: Negative prompts have been shown to substantially benefit OOD detection [30, 38, 48]. Incorporating negative prompts is particularly straightforward within the graph label propagation framework. We treat negative prompts as nodes " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{P}^n" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " with a label of " + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 55, + 594, + 296, + 714 + ], + "type": "text", + "content": " on the graph. During propagation, the negative labels compete with the positive labels (ID prompts) to influence the score of testing samples. Existing approach exploits ID data to learn negative prompt [48] which prohibits zero-shot inference. To address this, we propose a" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": "strategy that generates negative prompts via self-training, as illustrated in Fig. 1. Specifically, we first conduct initial score propagation for " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": " iterations following Eq. 3, giving rise to initial propagated scores " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": ". We select the top " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "m\\%" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": " most confident scores from " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "S_{T}" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": " following Eq. 5 where " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "\\mathrm{top}(S,m\\%)" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": " indicates the top " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "m\\%" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": " value of " + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 241, + 555, + 313 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 351, + 322, + 499, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 322, + 499, + 335 + ], + "spans": [ + { + "bbox": [ + 351, + 322, + 499, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal{I}_{pos} = \\left\\{i\\mid S_{T}[i] > \\operatorname {top}(S_{T},m\\%)\\right\\} ,", + "image_path": "86a16f84ecf46727670cdc9291bdc5845581cef32957b404d20c76f98517d7f8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 353, + 335, + 553, + 351 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 335, + 553, + 351 + ], + "spans": [ + { + "bbox": [ + 353, + 335, + 553, + 351 + ], + "type": "interline_equation", + "content": "\\mathcal {I} _ {n e g} = \\left\\{i \\mid S _ {T} [ i ] < \\operatorname {top} \\left(S _ {T}, 1 - m \\%\\right) \\right\\} \\tag{5}", + "image_path": "5f5fbb3a7b8740e5b132d6321497021cf2909b9e328e937c6721f06d24b8ba02.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "spans": [ + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "text", + "content": "We label the selected samples as positive and negative prompts respectively and re-initialize the score " + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "inline_equation", + "content": "\\tilde{S}_0" + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "text", + "content": ". For negative prompts, we assign scores of " + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "inline_equation", + "content": "-1" + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "text", + "content": " to further increase the gap between ID and OOD samples. The final score is obtained by running Eq. 3 with " + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "inline_equation", + "content": "\\tilde{S}_0" + }, + { + "bbox": [ + 313, + 354, + 555, + 414 + ], + "type": "text", + "content": " as initial scores." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 350, + 423, + 555, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 423, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 350, + 423, + 555, + 468 + ], + "type": "interline_equation", + "content": "\\tilde {S} _ {0} [ i ] = \\left\\{ \\begin{array}{l l} S _ {0} [ i ] & \\text {i f} i \\notin \\mathcal {I} _ {\\text {p o s}} \\& i \\notin \\mathcal {I} _ {\\text {n e g}}, \\\\ + 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {p o s}}, \\\\ - 1 & \\text {i f} i \\in \\mathcal {I} _ {\\text {n e g}}. \\end{array} \\right. \\tag {6}", + "image_path": "03af2010752b78195489b1194dfc22aed9930610b297d9c7462ac7d231c9c739.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "content": "Graph Construction: We now introduce the approach for constructing a graph for score propagation. As observed by [31, 44], the distance between text prompts and visual samples is generally larger than the distance between visual samples. This observation also holds for 3D point cloud VLM because the 3D VLMs are often finetuned from 2D VLMs [54, 55]. Therefore, we adopt a similar approach to construct the graph. We write the weight matrix " + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "content": " in a blockwise form as shown in Eq. 7, where " + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "inline_equation", + "content": "W_{u} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{u}|}" + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "inline_equation", + "content": "W_{up} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\tilde{\\mathcal{P}}|}" + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "inline_equation", + "content": "W_{ul} \\in \\mathbb{R}^{|\\mathcal{D}_{u}| \\times |\\mathcal{D}_{l}|}" + }, + { + "bbox": [ + 313, + 472, + 556, + 666 + ], + "type": "text", + "content": ". No connections are created between prototypes and few-shot labeled samples. All blocks are further constructed as a k-nearest neighbor (KNN) graph following [31, 44]. It is also computationally efficient to adapt an existing graph to new testing samples by incrementally adding the testing samples to the graph [31]." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 353, + 673, + 555, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 353, + 673, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 353, + 673, + 555, + 714 + ], + "type": "interline_equation", + "content": "W = \\left[ \\begin{array}{c c c} \\mathbf {I} ^ {| \\tilde {\\mathcal {P}} | \\times | \\tilde {\\mathcal {P}} |} & \\mathbf {0} & W _ {u p} ^ {\\top} \\\\ \\mathbf {0} & \\mathbf {I} ^ {| \\mathcal {D} _ {l} | \\times | \\mathcal {D} _ {l} |} & W _ {u l} ^ {\\top} \\\\ W _ {u p} & W _ {u l} & W _ {u} \\end{array} \\right] \\tag {7}", + "image_path": "2902e6965503c9c68432b216a0a375f95aeb058024de9e4713b9e88a4eee9bed.jpg" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "type": "text", + "content": "28800" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "Overall Algorithm: We present the overall algorithm for unified iterative label propagation for vision-language models in Algorithm 1. In Fig. 2, we show an overview of the proposed framework for zero-/few-shot OOD detection via graph propagation." + } + ] + } + ], + "index": 0 + }, + { + "type": "code", + "bbox": [ + 61, + 158, + 296, + 429 + ], + "blocks": [ + { + "bbox": [ + 57, + 142, + 294, + 156 + ], + "lines": [ + { + "bbox": [ + 57, + 142, + 294, + 156 + ], + "spans": [ + { + "bbox": [ + 57, + 142, + 294, + 156 + ], + "type": "text", + "content": "Algorithm 1 Graph Score Propagation for OOD Detection" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "code_caption" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "lines": [ + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "spans": [ + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": "1: Input: training data " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": ", testing data " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " and ID class prompt templates " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " \n2: Output: Final OOD scores " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{S}_T" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " \n3: for " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_l" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{T}_c" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": ", do \n Encode: " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "z_l = f(x_l)" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " for training data and " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{P}_c = \\{g(t_j(y_c))\\}" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " for prompt templates. Prompt clustering: positive prototypes " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathcal{P}}" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 4) \n4: end for \n5: for " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_u" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": ", do \n Encode: " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "z_i = f(x_i)" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " for testing data \n Build graph to compute " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 7) \n Initialize score " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "S_0^{(i)}" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 3) \n Score propagate to compute " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "S_T" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 3) \n Self-training to generate positive " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{pos}" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " and negative prompts " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{I}_{neg}" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 5) \n Updated initial scores " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{S}_0" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 6) \n Score propagate to compute " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{S}_T" + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "text", + "content": " (Eq. 3) \n6: end for \n7: return " + }, + { + "bbox": [ + 61, + 158, + 296, + 429 + ], + "type": "inline_equation", + "content": "\\tilde{S}_T" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "code_body" + } + ], + "index": 2, + "sub_type": "algorithm" + }, + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 137, + 462 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 468, + 170, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 170, + 481 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 170, + 481 + ], + "type": "text", + "content": "4.1. Experimental setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 486, + 296, + 714 + ], + "type": "text", + "content": "Dataset: In this work, we follow the 3DOS framework [1] to construct downstream datasets for evaluating our method. The 3DOS benchmark is based on two well-known 3D object datasets, ShapeNetCore [9] and ScanObjectNN [49]. ShapeNetCore includes 54 categories covering daily objects. ScanObjectNN is cropped out from ScanNet dataset featuring a real-world objects. It consists of 15 categories with 2,902 3D objects. To further validate our approach on real-world data, we introduce the Sydney Urban Object Dataset [11], nuScenes [7] and S3DIS [4]. Sydney Urban Object Dataset consists of LiDAR-scanned point clouds across 26 categories, totaling 631 unique instances. S3DIS dataset contains 6 large-scale indoor areas with 271 rooms annotated into 13 semantic categories. We cropped out individual objects using ground-truth segmentation masks, leading to 8931 objects. The nuScenes Dataset features point clouds captured by LiDAR sensor under autonomous driving scenario. We cropped out point clouds using ground-truth bounding boxes, leading to 2205" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 96 + ], + "type": "text", + "content": "objects across 9 categories. Detailed visualization of each dataset is deferred to the supplementary." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 98, + 555, + 469 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 98, + 555, + 469 + ], + "spans": [ + { + "bbox": [ + 313, + 98, + 555, + 469 + ], + "type": "text", + "content": "Evaluation Protocol: Following the methodology outlined in 3DOS [1], we divide ShapeNetCore into three non-overlapping category sets, SN1, SN2, and SN3, referred to as the ShapeNetCore dataset. Each set contains 18 categories. For each OOD test scenario, one of these sets (SN1, SN2, or SN3) is treated as the known set of categories, while the remaining two sets are used as unknown categories, resulting in three distinct OOD test cases. Similarly, 3DOS defines three category sets, SR1, SR2, and SR3 on the ScanObjectNN dataset. Each SR-set includes 5 non-overlapping categories, with the other two sets acting as unknown categories in each scenario and call them as ScanObjNN. As for Sydney Urban Object Dataset, we categorize these objects into two groups: movable objects, e.g. \"truck\" \"pedestrian\" and non-movable objects, e.g. \"building\" and \"traffic sign\". Movable objects are treated as in-distribution (ID) classes, simulating a scenario in autonomous driving where object detection primarily focuses on movable objects. For consistency, we pre-process each point cloud sample by upsampling it to a uniform number of points. For S3DIS dataset, We retained instances containing more than 2048 original points, where each instance was randomly downsampled to retain exactly 2048 points. Then we set the foreground objects as ID and back-ground objects as OOD. For nuScences dataset, we have conducted preliminary screening on the original point cloud and retained instances with more than 200 original points and upsampling it to 1024 points. Then we define Four-wheeled vehicles (including car, bus, truck, construction_vehicle, trailer) as ID and the rest as OOD. And further details provided in the supplementary material." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "spans": [ + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": "Implementation Details. We use the pre-trained ULIP/ULIP2 [54, 55] model to extract point cloud features and text features. For the 3D encoder, we select PointBERT [56]. In constructing the KNN graph, we set a universal " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\mathrm{K}" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": "-value " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "(N_{k})" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " of 10, connecting each text feature to its 10 nearest point cloud features. In Eq. 3, we set " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " to 0.5 and the number of iterations " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " to 5. The percentage of pseudo negative prompts, i.e. " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "m\\%" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " in Sect. 3.3, is also set to " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": ". Additionally, we use K-means clustering " + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "inline_equation", + "content": "(N_{c})" + }, + { + "bbox": [ + 313, + 471, + 556, + 590 + ], + "type": "text", + "content": " to group the prompts for each category into 3 clusters." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 592, + 556, + 700 + ], + "type": "text", + "content": "Competing Methods. We compare our method with several OOD detection approaches, including full-shot methods from 3DOS [1]. We also benchmark against recent zero-shot methods, such as DDCS [57], ZLaP[44], MCM [36], and NegLabel [24]. For zero-shot comparisons, all methods use the 3D encoder from [55]. Baseline results are also reported by directly using cosine similarity as the OOD score, with these methods referred to as ULIP and ULIP2, depending on the VLM used." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 701, + 553, + 714 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate OOD detection perfor" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28801" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 57, + 70, + 555, + 266 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 555, + 266 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 555, + 266 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 555, + 266 + ], + "type": "table", + "html": "
MethodScanObjNNShapeNetCore
SR3AUROC↑FPR95↓SR2AUROC↑FPR95↓SR1AUROC↑FPR95↓AverageAUROC↑FPR95↓SN1AUROC↑FPR95↓SN2AUROC↑FPR95↓SN3AUROC↑FPR95↓AverageAUROC↑FPR95↓
Requiring training datasetMSP[15]88.167.380.684.073.780.380.877.274.382.880.078.189.752.281.371.0
MLS[50]89.453.883.473.176.475.383.067.472.080.883.964.189.840.581.961.8
ODIN[32]90.247.983.371.776.376.883.365.574.279.479.471.787.841.880.564.3
Energy[33]89.550.681.675.876.675.582.667.372.181.283.064.789.839.482.061.8
GradNorm[19]88.550.777.475.375.276.880.467.672.181.857.788.957.879.062.683.3
ReAct[46]90.348.982.475.875.477.682.767.473.779.489.652.195.027.286.152.9
NF88.047.780.668.275.681.481.465.881.572.571.178.091.049.681.266.7
OE+mixup[16]72.683.572.088.562.587.869.086.672.778.980.368.887.362.280.169.9
ARPL+CS[10]--------74.880.380.772.485.450.880.367.8
Cosine protoCE91.041.082.178.277.675.683.664.980.368.388.760.891.938.086.955.7
SupCon[25]--------80.975.583.568.285.145.183.262.9
SubArcface[12]87.161.378.976.973.781.479.973.279.081.282.960.389.132.883.758.1
GSP(1-shot)79.361.380.984.188.651.782.965.791.344.086.443.183.891.487.259.5
GSP(5-shot)80.754.890.147.289.140.286.647.495.619.294.622.197.012.595.717.9
GSP(full-shot)90.133.093.030.592.234.491.822.697.910.197.97.998.86.098.28.0
Zero-shotULIP[54]79.982.762.387.557.492.366.587.588.057.176.448.775.251.379.952.4
ULIP2[55]65.780.876.471.277.478.073.276.785.752.584.144.367.674.479.157.1
DDCS[57]63.390.562.480.064.496.263.488.961.186.766.886.577.276.268.483.1
MCM[36]52.092.362.391.381.380.465.288.085.151.683.246.166.475.878.257.8
NegLabel[24]65.785.566.891.367.488.266.688.360.687.880.678.688.048.376.471.6
ZLaP[44]70.181.767.084.767.883.668.383.388.252.872.366.477.490.079.369.7
GSP75.863.076.577.683.661.678.667.490.638.970.764.079.793.780.465.5
", + "image_path": "eedea123209e157080c5a34eb0412c095c69f1dc6364b96cf2538e5e981e5351.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 310, + 295, + 381 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 310, + 295, + 381 + ], + "spans": [ + { + "bbox": [ + 55, + 310, + 295, + 381 + ], + "type": "text", + "content": "mance using two widely adopted metrics: AUROC and FPR95. AUROC calculates the area under the receiver operating characteristic curve, with higher values indicating better performance. FPR95 measures the false positive rate at a true positive rate of " + }, + { + "bbox": [ + 55, + 310, + 295, + 381 + ], + "type": "inline_equation", + "content": "95\\%" + }, + { + "bbox": [ + 55, + 310, + 295, + 381 + ], + "type": "text", + "content": ", with lower values reflecting improved performance." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 394, + 201, + 405 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 394, + 201, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 394, + 201, + 405 + ], + "type": "text", + "content": "4.2. Results on OOD Detection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "spans": [ + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": "Analysis of ScanObjNN Results: From the results presented in Tab. 1, we derive several key insights. i) Our proposed method demonstrates highly competitive performance under the zero-shot protocol, achieving an improvement of over " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": " in AUROC compared to ULIP2, which uses cosine distance directly as an OOD score. This substantial enhancement suggests that employing a manifold-based approach significantly enhances the differentiation between OOD samples and ID classes. ii) Our few-shot results are equally impressive, rivaling even fully supervised methodologies. Notably, with 5-shot training samples, our method attains an average AUROC of " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "86.6\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": ", surpassing the best fully supervised method, Cosine proto, which scores " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "83.6\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": ", by a margin of " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "3\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": ". Furthermore, when integrating all available training samples into the graph, our approach achieves an outstanding average AUROC of " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "91.8\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": ". iii) A noticeable " + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "inline_equation", + "content": "5\\%" + }, + { + "bbox": [ + 55, + 413, + 295, + 687 + ], + "type": "text", + "content": " gap in average AUROC is observed between our zero-shot and 1-shot methods. This gap underscores the critical role of incorporating real in-distribution samples in effectively distinguishing OOD samples. Consequently, practitioners are advised to introduce at least a few ID samples for OOD detection, in addition to leveraging Vision-Language Models (VLMs)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "Analysis of ShapeNetCore Results: The results for ShapeNetCore, as shown in Tab. 1, lead us to several ob" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 321, + 308, + 545, + 381 + ], + "blocks": [ + { + "bbox": [ + 55, + 272, + 555, + 293 + ], + "lines": [ + { + "bbox": [ + 55, + 272, + 555, + 293 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 555, + 293 + ], + "type": "text", + "content": "Table 1. Evaluation of OOD detection for ScanObjNN and ShapeNetCore datasets. Each column title for SR/SN indicates the chosen known class set and the rest sets of SR/SN are unknown." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 321, + 308, + 545, + 381 + ], + "lines": [ + { + "bbox": [ + 321, + 308, + 545, + 381 + ], + "spans": [ + { + "bbox": [ + 321, + 308, + 545, + 381 + ], + "type": "table", + "html": "
MethodSydneynuScenesS3DIS
AUROC↑FPR95↓AUROC↑FPR95↓AUROC↑FPR95↓
ULIP[54]63.481.753.293.353.195.1
ULIP2[55]66.283.764.293.062.592.3
DDCS[57]61.887.362.291.262.592.5
MCM[36]65.782.457.182.365.288.1
NegLabel[24]58.492.769.781.655.485.4
ZLaP[44]58.580.670.087.971.388.4
GSP72.568.980.880.777.694.0
", + "image_path": "e984d84464cd246816acd2246d20c361e32c5efd6dc619af1c8ed719db963861.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 386, + 555, + 407 + ], + "lines": [ + { + "bbox": [ + 313, + 386, + 555, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 555, + 407 + ], + "type": "text", + "content": "Table 2. Zero-shot OOD detection results on the Sydney Urban Objects dataset, nuScenes dataset and S3DIS dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "content": "servations. i) Unlike the findings from the ScanObjNN dataset, the performance gap between the leading baseline, ULIP, and our method under the zero-shot protocol is relatively narrow, approximately " + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "inline_equation", + "content": "0.5\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "content": ". This modest improvement can be attributed to the substantial overlap between the downstream and pre-training datasets used in Vision-Language Models (VLMs). Specifically, ULIP/ULIP2 are pre-trained on ShapeNet, which shares significant commonalities with ShapeNetCore, thereby impacting OOD detection capabilities. Despite this, the slight enhancement achieved by our method underscores the utility of leveraging data manifolds to improve the differentiation of OOD samples. ii) A marked enhancement in performance is observed when few-shot samples are available. For instance, introducing 1-shot training samples results in nearly a " + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "inline_equation", + "content": "7\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "content": " increase in average AUROC, while 5-shot samples yield an impressive " + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "inline_equation", + "content": "15\\%" + }, + { + "bbox": [ + 313, + 426, + 555, + 640 + ], + "type": "text", + "content": " improvement. This highlights the critical role of few-shot learning in enhancing model performance." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Analysis of Additional Real-World Datasets: The zero-shot OOD detection results for the Sydney Urban Objects dataset [11] and nuScenes Dataset [7], presented in Tab. 2, reinforce our previous conclusions. Our method consistently outperforms all competing approaches in zero-shot OOD detection, demonstrating the efficacy of utilizing test" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28802" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 216 + ], + "type": "text", + "content": "ing data manifolds, particularly when there is a substantial shift between the downstream task and the pre-training data. This finding emphasizes the robustness of our approach in adapting to significant distributional shifts in real-world applications. For S3DIS [4], GSP achieves the highest AUROC, outperforming all baselines by a significant margin (e.g., +12.4% over MCM). This suggests superior ability to distinguish between classes across all decision thresholds. However, GSP's FPR95 (94.0%) is slightly higher than ULIP (95.1%), indicating that while GSP excels in overall classification, it faces challenges in suppressing false positives when operating at a 95% true positive rate." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 296, + 434 + ], + "type": "text", + "content": "Further Insights Synergizing All Results: By comparing the outcomes from the ScanObjNN, Sydney Urban Objects, S3DIS, nuScenes and ShapeNetCore datasets, we derive several overarching insights. i) There is a notable correlation between the distribution shift from pre-training to downstream datasets and the efficacy of manifold exploitation for OOD detection. This can be attributed to several factors. Firstly, when pre-training is conducted on datasets that closely resemble the downstream datasets, the contrastive training loss effectively aligns text prompts with visual features, resulting in a well-defined separation between ID and OOD samples. Conversely, in scenarios where the distribution shift is more substantial, VLMs may struggle to accurately quantify the distance between ID text prompts and visual samples. In such cases, leveraging the testing data manifold becomes essential to enhance the calibration of prompt-to-visual feature distances, thereby improving OOD detection performance." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 453, + 149, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 453, + 149, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 453, + 149, + 467 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 474, + 295, + 713 + ], + "type": "text", + "content": "We begin by analyzing the performance of three Out-of-Distribution (OOD) detection methods: Cosine Similarity, Shortest Path, and Label Propagation. As shown in Tab. 3, we employ the ULIP2 backbone for extracting text and point cloud features and evaluate these methods for OOD scoring. The approach of directly computing the maximal normalized cosine similarity between text and point cloud features yields a baseline performance. Substituting cosine similarity with the shortest path metric results in a performance increment, albeit modest. It is noteworthy that the shortest path is computed using Dijkstra's algorithm, and the reciprocal of manifold distance serves as the OOD score. Lastly, we introduce a label propagation method to determine the OOD score. The experimental outcomes in Tab. 3 demonstrate that integrating graph-based information via label propagation significantly enhances the model's ability to differentiate between In-Distribution (ID) and OOD categories. This suggests that graph-based methods provide a more nuanced understanding of the feature space, leading to improved OOD detection capabilities." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 318, + 71, + 549, + 140 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 549, + 140 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 549, + 140 + ], + "type": "table", + "html": "
OOD Det. Mthd.Prompt Clust.Neg. PromptAUROC↑FPR95↓
Cosine Dist.--73.276.7
Manifold Dist.--74.173.0
Score Prop.--76.472.6
Score Prop.✓-77.273.6
Score Prop.-✓77.371.4
Score Prop.✓✓78.667.4
", + "image_path": "4a1fb485606fa542c9d329c4f6732b1ebf4c62337122c5e5d8c68627cf484394.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 347, + 143, + 520, + 154 + ], + "lines": [ + { + "bbox": [ + 347, + 143, + 520, + 154 + ], + "spans": [ + { + "bbox": [ + 347, + 143, + 520, + 154 + ], + "type": "text", + "content": "Table 3. Ablation study on ScanObjNN dataset." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "text" + }, + { + "type": "image", + "bbox": [ + 319, + 161, + 432, + 228 + ], + "blocks": [ + { + "bbox": [ + 319, + 161, + 432, + 228 + ], + "lines": [ + { + "bbox": [ + 319, + 161, + 432, + 228 + ], + "spans": [ + { + "bbox": [ + 319, + 161, + 432, + 228 + ], + "type": "image", + "image_path": "45c2ee79d7002bde5cc5bc02aebdc3488f57fab29dc063a33b492b1262781da5.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 371, + 229, + 381, + 238 + ], + "lines": [ + { + "bbox": [ + 371, + 229, + 381, + 238 + ], + "spans": [ + { + "bbox": [ + 371, + 229, + 381, + 238 + ], + "type": "text", + "content": "(a)" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 436, + 161, + 548, + 227 + ], + "blocks": [ + { + "bbox": [ + 436, + 161, + 548, + 227 + ], + "lines": [ + { + "bbox": [ + 436, + 161, + 548, + 227 + ], + "spans": [ + { + "bbox": [ + 436, + 161, + 548, + 227 + ], + "type": "image", + "image_path": "f1c1471fe75611a45cd261888034d8e51f3934b9cdf6e8d4f5b1c617e7a18762.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 495, + 228, + 506, + 238 + ], + "lines": [ + { + "bbox": [ + 495, + 228, + 506, + 238 + ], + "spans": [ + { + "bbox": [ + 495, + 228, + 506, + 238 + ], + "type": "text", + "content": "(b)" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "lines": [ + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 313, + 244, + 553, + 266 + ], + "type": "text", + "content": "Figure 3. Few-shot OOD detection on the ScanObjNN dataset. (a) Results on AUROC metric. (b) Results on FPR95 metric." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 288, + 416, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 416, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 416, + 300 + ], + "type": "text", + "content": "4.4. Further Analysis" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 306, + 553, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 306, + 553, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 306, + 553, + 425 + ], + "type": "text", + "content": "Few-shot OOD Detection: To evaluate the performance of label propagation in few-shot OOD detection, we incorporated a subset of training data from the 3DOS training dataset, treating them as in-distribution categories. The outcomes, presented in Fig. 3, show the effects of varying the number of training samples used in label propagation, with all samples randomly selected from the training dataset. The results indicate that as the number of training samples increases, OOD detection performance improves and eventually stabilizes." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 557 + ], + "type": "text", + "content": "Qualitative Analysis: To validate our method, we visualize sample predictions and scores from baseline method and our full GSP. As shown in Fig. 5, for easy-to-predict samples from the Real-world dataset, conventional distance-based methods perform well. However, in challenging cases (e.g., occlusions or missing parts), they often fail. In contrast, our method utilizes spatially proximate samples in feature space for score propagation, ensuring accurate OOD detection even in difficult scenarios. These qualitative results confirm the effectiveness and robustness of our approach." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "content": "Impact of Prompt Clustering: To illustrate the benefits of prompt expansion via K-means clustering, we examined how varying the number of prompts affects OOD (Out-of-Distribution) detection. As depicted in Fig. 4 (a), we experimented with different values of " + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "content": " in the K-means algorithm, where each " + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "content": " value corresponds to a distinct number of clusters, and consequently, a different number of prompts. An increase in " + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 558, + 554, + 665 + ], + "type": "text", + "content": " results in a greater number of distinct prompts." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "The results presented in the figure reveal that the quantity of prompts considerably impacts OOD detection performance. Specifically, when " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "N_{c} = 6" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ", the AUROC surpasses " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "78.7\\%" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": ". However, further increasing " + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": " can negatively af" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28803" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 72, + 221, + 159 + ], + "blocks": [ + { + "bbox": [ + 58, + 72, + 221, + 159 + ], + "lines": [ + { + "bbox": [ + 58, + 72, + 221, + 159 + ], + "spans": [ + { + "bbox": [ + 58, + 72, + 221, + 159 + ], + "type": "image", + "image_path": "eb35f117bd9b1174e6783df8ad2831aa7144b96479f7b73ae896783e36cee4ea.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 92, + 162, + 188, + 172 + ], + "lines": [ + { + "bbox": [ + 92, + 162, + 188, + 172 + ], + "spans": [ + { + "bbox": [ + 92, + 162, + 188, + 172 + ], + "type": "text", + "content": "(a) " + }, + { + "bbox": [ + 92, + 162, + 188, + 172 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 92, + 162, + 188, + 172 + ], + "type": "text", + "content": " Nearest Neighbors" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 72, + 388, + 160 + ], + "blocks": [ + { + "bbox": [ + 224, + 72, + 388, + 160 + ], + "lines": [ + { + "bbox": [ + 224, + 72, + 388, + 160 + ], + "spans": [ + { + "bbox": [ + 224, + 72, + 388, + 160 + ], + "type": "image", + "image_path": "2b9530dee9a3d2a6ccd4cf28519da4f2600f1519fae7f6989d4b826006a3063d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 250, + 162, + 378, + 172 + ], + "lines": [ + { + "bbox": [ + 250, + 162, + 378, + 172 + ], + "spans": [ + { + "bbox": [ + 250, + 162, + 378, + 172 + ], + "type": "text", + "content": "(b) Number of negative prompt " + }, + { + "bbox": [ + 250, + 162, + 378, + 172 + ], + "type": "inline_equation", + "content": "m" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 553, + 198 + ], + "type": "text", + "content": "Figure 4. Results from ScanObjNN Benchmark Experiments: (a) Evaluation of prompt clustering with different K-means cluster numbers. (b) Impact of varying negative prompts on label propagation. (c) Results of different number of K Neaerst Neighbors in KNN." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 347, + 295, + 380 + ], + "lines": [ + { + "bbox": [ + 55, + 347, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 347, + 295, + 380 + ], + "type": "text", + "content": "Figure 5. Visualization of some samples for ScanObjectNN and nuScenes Dataset with two methods predict scores. All scores are normalized to 0-1." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 389, + 72, + 554, + 160 + ], + "blocks": [ + { + "bbox": [ + 389, + 72, + 554, + 160 + ], + "lines": [ + { + "bbox": [ + 389, + 72, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 389, + 72, + 554, + 160 + ], + "type": "image", + "image_path": "65800647109a20c84af0d895f5cef84f7b817989c9daad2c147d02885165d623.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 432, + 162, + 517, + 172 + ], + "lines": [ + { + "bbox": [ + 432, + 162, + 517, + 172 + ], + "spans": [ + { + "bbox": [ + 432, + 162, + 517, + 172 + ], + "type": "text", + "content": "(c) Cluster Number " + }, + { + "bbox": [ + 432, + 162, + 517, + 172 + ], + "type": "inline_equation", + "content": "N_{c}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 61, + 218, + 293, + 342 + ], + "blocks": [ + { + "bbox": [ + 61, + 218, + 293, + 342 + ], + "lines": [ + { + "bbox": [ + 61, + 218, + 293, + 342 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 293, + 342 + ], + "type": "table", + "html": "
ScanObjectNN DatasetnuScenes Dataset
ID categoryEasy-to-predictHard-to-predictEasy-to-predictHard-to-predict
Distance-based score:0.6180.0930.6890.192
GSP score:0.4650.3570.3740.278
OOD category
Distance-based score:0.1740.3960.1480.431
GSP score:0.0720.0540.0740.127
", + "image_path": "93be5203fa709110b2f5b241f77a08875a8d27d48744f3372ac56618d10367b3.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 390, + 295, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 390, + 295, + 461 + ], + "spans": [ + { + "bbox": [ + 55, + 390, + 295, + 461 + ], + "type": "text", + "content": "fect the spatial alignment between text features and point cloud features. This phenomenon is likely due to the fact that typical prompts in Vision-Language Models (VLMs) are generally standardized, with only occasional outliers. Excessive clustering may isolate these outliers, leading to a decline in performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "text", + "content": "In conventional VLM approaches, the mean text feature vector is typically calculated for each category (equivalent to setting " + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "inline_equation", + "content": "N_{c} = 1" + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "text", + "content": "), achieving an average AUROC of less than " + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "inline_equation", + "content": "78\\%" + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "text", + "content": ". The findings in Fig. 4 (c) suggest that selecting an appropriate " + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "inline_equation", + "content": "N_{c}" + }, + { + "bbox": [ + 55, + 462, + 295, + 544 + ], + "type": "text", + "content": " enhances the model's ability to differentiate OOD categories, underscoring the advantages of prompt clustering." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": "Effectiveness of Negative Prompting via Self-Training: We conducted experiments to evaluate the effectiveness of the proposed negative prompting. Using the ShapeNetCore dataset, we explored the impact of self-training for negative prompting by varying the percentage of samples tagged with pseudo labels, denoted as " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m\\%" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " as introduced in Section 3.3. In Fig. 4 (b), We observe a performance increase from " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 1" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": " up to " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", indicating that incorporating negative prompts through self-training is an effective strategy to leverage the manifold for score propagation. However, performance declines beyond " + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "inline_equation", + "content": "m = 6" + }, + { + "bbox": [ + 55, + 545, + 295, + 713 + ], + "type": "text", + "content": ", likely due to the introduction of incorrect pseudo-negative prompts. This highlights the importance of mitigating the effects of confirmation bias [3] to prevent the model from being adversely" + } + ] + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 317, + 218, + 566, + 265 + ], + "blocks": [ + { + "bbox": [ + 317, + 218, + 566, + 265 + ], + "lines": [ + { + "bbox": [ + 317, + 218, + 566, + 265 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 566, + 265 + ], + "type": "table", + "html": "
Modelload + DatasetloadText Feature ExtractPrompt ClusteringFeature ExtractBuild GraphScore PropagationTotal time
SR1552ms2212ms1128ms2780ms106ms8ms7786ms
SN1641ms2459ms1657ms25031ms671ms34ms31493ms
", + "image_path": "466e9cb93ca14792989b9e559a525c4f9eb545655fc2d86e26a8dea0cafeb329.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 334, + 268, + 533, + 278 + ], + "lines": [ + { + "bbox": [ + 334, + 268, + 533, + 278 + ], + "spans": [ + { + "bbox": [ + 334, + 268, + 533, + 278 + ], + "type": "text", + "content": "Table 4. The inference time of the GSP on SR and SN." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 314, + 291, + 351, + 301 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 291, + 351, + 301 + ], + "spans": [ + { + "bbox": [ + 314, + 291, + 351, + 301 + ], + "type": "text", + "content": "affected." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "text", + "content": "Impact of K-Nearest-Neighbor. We compare performance with different values of " + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "text", + "content": " Nearest Neighbors. As we can see in Fig 4 (a), the results demonstrate that as the number of " + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "inline_equation", + "content": "N_{k}" + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "text", + "content": " values increases, the OOD detection performance is tend to be stable. While " + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "inline_equation", + "content": "N_{k} = 16" + }, + { + "bbox": [ + 313, + 304, + 555, + 376 + ], + "type": "text", + "content": " achieve the best performance and Less than 8 or larger than 18 get the low results." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "content": "Computational Cost: We profiled each step and test the inference time of GSP using a full GPU implementation in Tab. 4. Experiments are carried out on a single RTX3090 GPU and a i7-12700k CPU. As observed, the graph-based operations (Build Graph + Score Propagation) account for only around " + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 377, + 555, + 473 + ], + "type": "text", + "content": " of the total inference time. The additional computation cost introduced by graph operations is negligible compared with point cloud feature extraction." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 488, + 388, + 501 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 488, + 388, + 501 + ], + "spans": [ + { + "bbox": [ + 314, + 488, + 388, + 501 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 713 + ], + "type": "text", + "content": "In this work, we presented a novel training-free approach to 3D point cloud OOD detection that effectively leverages Vision-Language Models through Graph Score Propagation. By constructing a graph that incorporates both positive prototypes and pseudo-negative samples via self-training, our method captures the underlying structure of the testing data manifold. This approach allows for the propagation of informative OOD scores, thus surpassing existing distance-based techniques. Empirical results across multiple 3D point cloud datasets demonstrate the robustness and adaptability of our framework, particularly in zero- and few-shot settings. The GSP method addresses key limitations in prior OOD approaches by reducing dependency on extensive in-distribution training data while achieving high accuracy. Future work could explore further enhancements through dynamic graph updates, enabling real-time OOD detection in continuously evolving environments." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28804" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "spans": [ + { + "bbox": [ + 56, + 166, + 115, + 178 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 58, + 186, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 186, + 296, + 240 + ], + "type": "text", + "content": "[1] Antonio Alliegro, Francesco Cappio Borlino, and Tatiana Tommasi. 3dos: Towards 3d open set learning-benchmarking and understanding semantic novelty detection on point clouds. Advances in Neural Information Processing Systems, 2022. 2, 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 241, + 296, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 241, + 296, + 283 + ], + "spans": [ + { + "bbox": [ + 61, + 241, + 296, + 283 + ], + "type": "text", + "content": "[2] Mouin Ben Ammar, Nacim Belkhir, Sebastian Popescu, Antoine Manzanera, and Gianni Franchi. Neco: Neural collapse based out-of-distribution detection. arXiv preprint arXiv:2310.06823, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "spans": [ + { + "bbox": [ + 62, + 285, + 295, + 328 + ], + "type": "text", + "content": "[3] Eric Arazo, Diego Ortego, Paul Albert, Noel E O'Connor, and Kevin McGuinness. Pseudo-labeling and confirmation bias in deep semi-supervised learning. In International joint conference on neural networks, 2020. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "spans": [ + { + "bbox": [ + 62, + 330, + 295, + 372 + ], + "type": "text", + "content": "[4] I Armeni, O Sener, Ar Zamir, H Jiang, I Brilakis, M Fischer, S Savarese, and IEEE. 3d semantic parsing of large-scale indoor spaces. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2016. 5, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "spans": [ + { + "bbox": [ + 62, + 373, + 296, + 426 + ], + "type": "text", + "content": "[5] Yichen Bai, Zongbo Han, Bing Cao, Xiaoheng Jiang, Qinghua Hu, and Changqing Zhang. Id-like prompt learning for few-shot out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 1, 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 428, + 295, + 471 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 428, + 295, + 471 + ], + "spans": [ + { + "bbox": [ + 62, + 428, + 295, + 471 + ], + "type": "text", + "content": "[6] Tianyi Bao, Qitian Wu, Zetian Jiang, Yiting Chen, Jiawei Sun, and Junchi Yan. Graph out-of-distribution detection goes neighborhood shaping. In International Conference on Machine Learning. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 472, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 472, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 62, + 472, + 296, + 536 + ], + "type": "text", + "content": "[7] Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh Vora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020. 5, 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 538, + 296, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 538, + 296, + 570 + ], + "spans": [ + { + "bbox": [ + 62, + 538, + 296, + 570 + ], + "type": "text", + "content": "[8] Tianle Cai, Ruiqi Gao, Jason Lee, and Qi Lei. A theory of label propagation for subpopulation shift. In International Conference on Machine Learning, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "spans": [ + { + "bbox": [ + 62, + 571, + 295, + 624 + ], + "type": "text", + "content": "[9] Angel X Chang, Thomas Funkhouser, Leonidas Guibas, Pat Hanrahan, Qixing Huang, Zimo Li, Silvio Savarese, Manolis Savva, Shuran Song, Hao Su, et al. Shapenet: An information-rich 3d model repository. arXiv preprint arXiv:1512.03012, 2015. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 58, + 626, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 626, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 58, + 626, + 295, + 669 + ], + "type": "text", + "content": "[10] Guangyao Chen, Peixi Peng, Xiangqian Wang, and Yonghong Tian. Adversarial reciprocal points learning for open set recognition. IEEE Transactions on Pattern analysis and Machine Intelligence, 2021. 1, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 58, + 670, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 58, + 670, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 58, + 670, + 295, + 712 + ], + "type": "text", + "content": "[11] Mark De Deuge, Alastair Quadros, Calvin Hung, and Bertrand Douillard. Unsupervised feature learning for classification of outdoor 3d scans. In Australasian Conference on Robotics and Automation, 2013. 5, 6" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 58, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 296, + 147 + ], + "type": "text", + "content": "This work was supported by the National Natural Science Foundation of China (No. U2468207), Sichuan Science and Technology Program (No. 2024NSFTD0036), and the Agency for Science, Technology and Research (A*STAR) under its MTC Programmatic Funds (No. M23L7b0021)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[12] Jiankang Deng, Jia Guo, Tongliang Liu, Mingming Gong, and Stefanos Zafeiriou. Sub-center arcface: Boosting face recognition by large-scale noisy web faces. In European Conference on Computer Vision, 2020. 6" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "spans": [ + { + "bbox": [ + 317, + 119, + 553, + 151 + ], + "type": "text", + "content": "[13] Edsger W Dijkstra. A note on two problems in connexion with graphs. In Edsger Wybe Dijkstra: his life, work, and legacy. 2022. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 153, + 553, + 206 + ], + "type": "text", + "content": "[14] Shreyas Havaldar, Navodita Sharma, Shubhi Sareen, Karthikeyan Shanmugam, and Aravindan Raghuveer. Learning from label proportions: Bootstrapping supervised learners via belief propagation. arXiv preprint arXiv:2310.08056, 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[15] Dan Hendrycks and Kevin Gimpel. A baseline for detecting misclassified and out-of-distribution examples in neural networks. In International Conference on Learning Representations, 2022. 1, 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 255, + 553, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 255, + 553, + 286 + ], + "spans": [ + { + "bbox": [ + 316, + 255, + 553, + 286 + ], + "type": "text", + "content": "[16] Dan Hendrycks, Mantas Mazeika, and Thomas Dietterich. Deep anomaly detection with outlier exposure. In International Conference on Learning Representations, 2019. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 289, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 289, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 289, + 553, + 342 + ], + "type": "text", + "content": "[17] Xuefeng Hu, Ke Zhang, Lu Xia, Albert Chen, Jiajia Luo, Yuyin Sun, Ken Wang, Nan Qiao, Xiao Zeng, Min Sun, et al. Reclip: Refine contrastive language image pre-training with source free domain adaptation. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 345, + 553, + 410 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 345, + 553, + 410 + ], + "spans": [ + { + "bbox": [ + 317, + 345, + 553, + 410 + ], + "type": "text", + "content": "[18] Chengjie Huang, Vahdat Abdelzad, Christopher Gus Mannes, Luke Rowe, Benjamin Therien, Rick Salay, Krzysztof Czarnecki, et al. Out-of-distribution detection for lidar-based 3d object detectabel propagation for zero-shot classification with vision. In IEEE International Conference on Intelligent Transportation Systems, 2022. 1, 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 412, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 412, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 412, + 553, + 453 + ], + "type": "text", + "content": "[19] Rui Huang, Andrew Geng, and Yixuan Li. On the importance of gradients for detecting distributional shifts in the wild. Advances in Neural Information Processing Systems, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 553, + 510 + ], + "type": "text", + "content": "[20] Tianyu Huang, Bowen Dong, Yunhan Yang, Xiaoshui Huang, Rynson WH Lau, Wanli Ouyang, and Wangmeng Zuo. Clip2point: Transfer clip to point cloud classification with image-depth pre-training. In IEEE/CVF International Conference on Computer Vision, 2023. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 555 + ], + "type": "text", + "content": "[21] Abiodun M. Ikatun, Absalom E. Ezugwu, Laith Abualigah, Belal Abuhaija, and Jia Heming. K-means clustering algorithms: A comprehensive review, variants analysis, and advances in the era of big data. Information science, 2023. 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 558, + 553, + 601 + ], + "type": "text", + "content": "[22] Ahmet Iscen, Giorgos Tolias, Yannis Avrithis, and Ondrej Chum. Label propagation for deep semi-supervised learning. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 553, + 657 + ], + "type": "text", + "content": "[23] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In International Conference on Machine Learning, 2021. 1" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 659, + 553, + 712 + ], + "type": "text", + "content": "[24] Xue Jiang, Feng Liu, Zhen Fang, Hong Chen, Tongliang Liu, Feng Zheng, and Bo Han. Negative label guided ood detection with pretrained vision-language models. In International Conference on Learning Representations, 2024. 3, 5, 6" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28805" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 117 + ], + "type": "text", + "content": "[25] Prannay Khosla, Piotr Teterwak, Chen Wang, Aaron Sarna, Yonglong Tian, Phillip Isola, Aaron Maschinot, Ce Liu, and Dilip Krishnan. Supervised contrastive learning. Advances in Neural Information Processing Systems, 2020. 6" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[26] Konstantin Kirchheim, Tim Gonschorek, and Frank Ortmeyer. Out-of-distribution detection with logical reasoning. In IEEE/CVF Winter Conference on Applications of Computer Vision, 2024. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 207 + ], + "type": "text", + "content": "[27] Michael Kösel, Marcel Schreiber, Michael Ulrich, Claudius Gläser, and Klaus C. J. Dietmayer. Revisiting out-of-distribution detection in lidar-based 3d object detection. IEEE Intelligent Vehicles Symposium, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 252 + ], + "type": "text", + "content": "[28] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In International Conference on Machine Learning, 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "spans": [ + { + "bbox": [ + 56, + 254, + 294, + 308 + ], + "type": "text", + "content": "[29] Jingyao Li, Pengguang Chen, Zexin He, Shaozuo Yu, Shu Liu, and Jiaya Jia. Rethinking out-of-distribution (ood) detection: Masked image modeling is all you need. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 309, + 294, + 353 + ], + "type": "text", + "content": "[30] Tianqi Li, Guansong Pang, Xiao Bai, Wenjun Miao, and Jin Zheng. Learning transferable negative prompts for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 354, + 294, + 408 + ], + "type": "text", + "content": "[31] Yushu Li, Yongyi Su, Adam Goodge, Kui Jia, and Xun Xu. Efficient and context-aware label propagation for zero/few-shot training-free adaptation of vision-language model. In International Conference on Learning Representations, 2025. 2, 4" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 410, + 294, + 453 + ], + "type": "text", + "content": "[32] Shiyu Liang, Yixuan Li, and R Srikant. Enhancing the reliability of out-of-distribution image detection in neural networks. In International Conference on Learning Representations, 2018. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[33] Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. Energy-based out-of-distribution detection. Advances in Neural Information Processing Systems, 2020. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 533 + ], + "type": "text", + "content": "[34] Mana Masuda, Ryo Hachiuma, Ryo Fujii, Hideo Saito, and Yusuke Sekikawa. Toward unsupervised 3d point cloud anomaly detection using variational autoencoder. In IEEE International Conference on Image Processing, 2021. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 535, + 294, + 578 + ], + "type": "text", + "content": "[35] Guofeng Mei, Luigi Riz, Yiming Wang, and Fabio Poiesi. Geometrically-driven aggregation for zero-shot 3d point cloud understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 623 + ], + "type": "text", + "content": "[36] Yifei Ming, Ziyang Cai, Jiuming Gu, Yiyou Sun, Wei Li, and Yixuan Li. Delving into out-of-distribution detection with vision-language representations. In Advances in Neural Information Processing Systems, 2022. 1, 3, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 667 + ], + "type": "text", + "content": "[37] Atsuyuki Miyai, Qing Yu, Go Irie, and Kiyoharu Aizawa. Locoop: Few-shot out-of-distribution detection via prompt learning. Advances in Neural Information Processing Systems, 2024. 1, 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 713 + ], + "type": "text", + "content": "[38] Jun Nie, Yonggang Zhang, Zhen Fang, Tongliang Liu, Bo Han, and Xinmei Tian. Out-of-distribution detection with negative prompts. In International Conference on Learning Representations, 2024. 2, 4" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 128 + ], + "type": "text", + "content": "[39] Julia Nitsch, Masha Itkina, Ransalu Senanayake, Juan Nieto, Max Schmidt, Roland Siegwart, Mykel J Kochenderfer, and Cesar Cadena. Out-of-distribution detection for automotive perception. In IEEE International Intelligent Transportation Systems Conference, 2021. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 162 + ], + "type": "text", + "content": "[40] Rattana Pukdee, Dylan Sam, Maria-Florina Balcan, and Pradeep Ravikumar. Label propagation with weak supervision. arXiv preprint arXiv:2210.03594, 2022. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 186 + ], + "type": "text", + "content": "[41] Rui Qian, Xin Lai, and Xirong Li. 3d object detection for autonomous driving: A survey. Pattern Recognition, 2022. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 189, + 553, + 255 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 189, + 553, + 255 + ], + "spans": [ + { + "bbox": [ + 317, + 189, + 553, + 255 + ], + "type": "text", + "content": "[42] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, 2021. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "spans": [ + { + "bbox": [ + 316, + 256, + 553, + 311 + ], + "type": "text", + "content": "[43] Maximilian Stadler, Bertrand Charpentier, Simon Geisler, Daniel Zügner, and Stephan Gunnemann. Graph posterior network: Bayesian predictive uncertainty for node classification. Advances in Neural Information Processing Systems, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 313, + 553, + 357 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 313, + 553, + 357 + ], + "spans": [ + { + "bbox": [ + 316, + 313, + 553, + 357 + ], + "type": "text", + "content": "[44] Vladan Stojnic, Yannis Kalantidis, and Giorgos Tolias. Label propagation for zero-shot classification with vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4, 5, 6" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 358, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 358, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 358, + 553, + 403 + ], + "type": "text", + "content": "[45] Jiachen Sun, Qingzhao Zhang, Bhavya Kailkhura, Zhiding Yu, Chaowei Xiao, and Z Morley Mao. Benchmarking robustness of 3d point cloud recognition against common corruptions. arXiv preprint arXiv:2201.12296, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 405, + 553, + 437 + ], + "type": "text", + "content": "[46] Yiyou Sun, Chuan Guo, and Yixuan Li. React: Out-of-distribution detection with rectified activations. Advances in Neural Information Processing Systems, 2021. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 440, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 440, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 440, + 553, + 473 + ], + "type": "text", + "content": "[47] Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. Out-of-distribution detection with deep nearest neighbors. In International Conference on Machine Learning, 2022. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 518 + ], + "type": "text", + "content": "[48] Xinyu Tian, Shu Zou, Zhaoyuan Yang, and Jing Zhang. Argue: Attribute-guided prompt tuning for vision-language models. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 520, + 553, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 520, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 316, + 520, + 553, + 576 + ], + "type": "text", + "content": "[49] Mikaela Angelina Uy, Quang-Hieu Pham, Binh-Son Hua, Thanh Nguyen, and Sai-Kit Yeung. Revisiting point cloud classification: A new benchmark dataset and classification model on real-world data. In IEEE/CVF International Conference on Computer Vision, 2019. 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "spans": [ + { + "bbox": [ + 316, + 578, + 553, + 620 + ], + "type": "text", + "content": "[50] Sagar Vaze, Kai Han, Andrea Vedaldi, and Andrew Zisserman. Open-set recognition: A good closed-set classifier is all you need. In International Conference on Learning Representations, 2022. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 316, + 623, + 553, + 666 + ], + "type": "text", + "content": "[51] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. Clipn for zero-shot ood detection: Teaching clip to say no. In IEEE/CVF International Conference on Computer Vision, 2023. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 713 + ], + "type": "text", + "content": "[52] Qizhou Wang, Zhen Fang, Yonggang Zhang, Feng Liu, Yixuan Li, and Bo Han. Learning to augment distributions for out-of-distribution detection. Advances in Neural Information Processing Systems, 2023. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 318, + 758 + ], + "type": "text", + "content": "28806" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 654 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 106 + ], + "type": "text", + "content": "[53] Qitian Wu, Yiting Chen, Chenxiao Yang, and Junchi Yan. Energy-based out-of-distribution detection for graph neural networks. ArXiv, 2023. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 173 + ], + "type": "text", + "content": "[54] Le Xue, Mingfei Gao, Chen Xing, Roberto Martin-Martin, Jiajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, and Silvio Savarese. Ulip: Learning a unified representation of language, images, and point clouds for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 1, 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 239 + ], + "type": "text", + "content": "[55] Le Xue, Ning Yu, Shu Zhang, Artemis Panagopoulou, Junnan Li, Roberto Martin-Martín, Jajun Wu, Caiming Xiong, Ran Xu, Juan Carlos Niebles, et al. Ulip-2: Towards scalable multimodal pre-training for 3d understanding. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3, 4, 5, 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 240, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 240, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 294, + 294 + ], + "type": "text", + "content": "[56] Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 341 + ], + "type": "text", + "content": "[57] Yue Yuan, Rundong He, Yicong Dong, Zhongyi Han, and Yilong Yin. Discriminability-driven channel selection for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 5, 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "spans": [ + { + "bbox": [ + 56, + 342, + 294, + 385 + ], + "type": "text", + "content": "[58] Ji Zhang, Shihan Wu, Lianli Gao, Heng Tao Shen, and Jingkuan Song. Dept: Decoupled prompt tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12924-12933, 2024. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 386, + 294, + 430 + ], + "type": "text", + "content": "[59] Ji Zhang, Jingkuan Song, Lianli Gao, Nicu Sebe, and Heng Tao Shen. Reliable few-shot learning under dual noises. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "spans": [ + { + "bbox": [ + 56, + 431, + 294, + 485 + ], + "type": "text", + "content": "[60] Renrui Zhang, Ziyu Guo, Wei Zhang, Kunchang Li, Xupeng Miao, Bin Cui, Yu Qiao, Peng Gao, and Hongsheng Li. Pointclip: Point cloud understanding by clip. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "spans": [ + { + "bbox": [ + 56, + 487, + 294, + 541 + ], + "type": "text", + "content": "[61] Renrui Zhang, Wei Zhang, Rongyao Fang, Peng Gao, Kunchang Li, Jifeng Dai, Yu Qiao, and Hongsheng Li. Tip-adapter: Training-free adaption of clip for few-shot classification. In European Conference on Computer Vision. Springer, 2022. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "spans": [ + { + "bbox": [ + 56, + 543, + 294, + 575 + ], + "type": "text", + "content": "[62] Zihan Zhang and Xiang Xiang. Decoupling maxlogit for out-of-distribution detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "spans": [ + { + "bbox": [ + 56, + 576, + 294, + 609 + ], + "type": "text", + "content": "[63] Xiaojin Zhu and Zoubin Ghahramani. Learning from labeled and unlabeled data with label propagation. ProQuest number: information to all users, 2002. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "spans": [ + { + "bbox": [ + 56, + 610, + 294, + 654 + ], + "type": "text", + "content": "[64] Xiaojin Zhu, Zoubin Ghahramani, and John D Lafferty. Semi-supervised learning using gaussian fields and harmonic functions. In International Conference on Machine Learning, 2003. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 750, + 317, + 758 + ], + "type": "text", + "content": "28807" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_content_list.json b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..46c22b8a8c628b574672b1bfdab44131bd992f4d --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_content_list.json @@ -0,0 +1,1579 @@ +[ + { + "type": "text", + "text": "ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors", + "text_level": 1, + "bbox": [ + 230, + 130, + 767, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Minsu Kim Subin Jeon In Cho Mijin Yoo Seon Joo Kim Yonsei University", + "bbox": [ + 236, + 203, + 759, + 242 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{minsul206, subinjeon, join, mijin6690, seonjookim}@yonsei.ac.kr", + "bbox": [ + 212, + 244, + 777, + 260 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/59bf2f231a0fad30377aff1a1467886b814f42c7b60d4a388afbc126f45bfaee.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 272, + 362, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ce604cabb669ad002c68b0c7cb7e2a28420e44df8a478225d17b0aa1378bb8dd.jpg", + "image_caption": [ + "(a) Exploration Results" + ], + "image_footnote": [], + "bbox": [ + 367, + 272, + 635, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/17b881f061e6c1beaaabe70dbc72fe4e9440d7d452729cc3228f4b4136582aff.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 638, + 272, + 905, + 455 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/f749aa8e21126ea38969b0f09c96cf28c0dd9c52d123a4ac8a4aad35199e4cab.jpg", + "image_caption": [ + "(b) Virtual Camera Sampling with Diffusion Prior" + ], + "image_footnote": [], + "bbox": [ + 99, + 470, + 478, + 563 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/ef91def92e136b83ce333261f0616179551751eb923b4b8a5521c957668a4e1e.jpg", + "image_caption": [ + "(c) Scene Reconstruction", + "Figure 1. (a) Explorable 3D scene reconstruction results. Our method renders photorealistic images from any arbitrary viewpoints. (b) Virtual cameras are sampled in regions with missing information, and pseudo-ground truth is generated using diffusion priors. (c) After optimizing 3D Gaussians with virtual viewpoint samples, our method enables real-time scene exploration." + ], + "image_footnote": [], + "bbox": [ + 555, + 470, + 875, + 566 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 643, + 326, + 659 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advances in novel view synthesis (NVS) have enabled real-time rendering with 3D Gaussian Splatting (3DGS). However, existing methods struggle with artifacts and missing regions when rendering from viewpoints that deviate from the training trajectory, limiting seamless scene exploration. To address this, we propose a 3DGS-based pipeline that generates additional training views to enhance reconstruction. We introduce an information-gain-driven virtual camera placement strategy to maximize scene coverage, followed by video diffusion priors to refine rendered results. Fine-tuning 3D Gaussians with these enhanced views significantly improves reconstruction quality. To evaluate our method, we present Wild-Explore, a benchmark designed for challenging scene exploration. Experiments", + "bbox": [ + 88, + 689, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "demonstrate that our approach outperforms existing 3DGS-based methods, enabling high-quality, artifact-free rendering from arbitrary viewpoints.", + "bbox": [ + 511, + 645, + 906, + 691 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 723, + 645, + 738 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The field of novel view synthesis (NVS) has seen significant advancements, enabling photorealistic rendering of scenes from previously unseen viewpoints. Much of this progress is attributed to NeRF [24] and its variants [1, 25]. This advancement has been further accelerated by recent 3D Gaussian Splatting (3DGS) [11], which enables high-quality rendering in real-time. The real-time capability of 3DGS broadens the applications of NVS in domains such as AR, VR, and gaming, leading users to expect the experience of freely exploring reconstructed scenes.", + "bbox": [ + 509, + 750, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "27042", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Unfortunately, such an experience is yet to be fully realized, as existing methods suffer from severe degradations in rendering quality when viewpoints deviate significantly from the input observations. These degradations include holes, oversized Gaussians, spiky surface artifacts, and spurious floaters, often arising from occlusions or limited viewpoint coverage (see Fig. 1 (a)). This limitation stems from missing information, since optimization-based approaches cannot synthesize contents beyond the observed data.", + "bbox": [ + 89, + 90, + 480, + 226 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We are interested in 3D scene reconstruction that can render more vivid and photorealistic images from arbitrary viewpoints, so that users can freely explore the reconstructed scenes. We term this task as \"explorable scene reconstruction\", which aims to reconstruct complete scenes within the target areas. To avoid generating arbitrary contents in regions far beyond the input observations—which would not reflect actual user experiences—we define the target areas as a bounding box defined by the input views.", + "bbox": [ + 89, + 228, + 480, + 364 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "One straightforward solution is to take thousands of images to completely cover the target scenes. However, collecting such dense data is time-consuming and labor-intensive, making it impractical for many real-world applications. Thus, we aim to reconstruct complete 3D scenes from input images that only partially cover the target scenes, which better reflects practical capturing scenarios.", + "bbox": [ + 89, + 366, + 480, + 470 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we introduce ExploreGS, a pipeline that enables explorable scene reconstruction using diffusion priors and 3DGS. ExploreGS leverages video diffusion priors to generate additional training views that complement missing information. These synthesized images at virtual viewpoints serve as pseudo ground truths to refine the initial, incomplete 3D Gaussians, thereby enabling complete, photorealistic rendering at arbitrary viewpoints.", + "bbox": [ + 89, + 473, + 480, + 594 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The key challenges of explorable scene reconstruction lie in determining the optimal placement of virtual viewpoints. Unlike previous works that assume evaluation trajectories are known [19, 33], virtual viewpoints in our pipeline need to be sampled to best complement missing information in poorly covered areas. We propose a novel view sampling strategy that strategically positions virtual cameras based on information gain. This approach identifies viewpoints that offer the most novel scene details from the initial 3DGS reconstruction as shown in Fig. 1 (b).", + "bbox": [ + 89, + 595, + 480, + 744 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To provide pseudo observations, we generate realistic images along sampled viewpoints. We introduce video diffusion model which converts renderings with degradations into clean images. Although it is motivated by 3DGS-Enhancer [21], we modify the training objectives to fit scene exploration and prepare larger dataset for training. We then finetune 3DGS on both virtual viewpoints and training viewpoints with uncertainty guidance.", + "bbox": [ + 89, + 748, + 480, + 868 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To validate our pipeline, we introduce Wild-Explore, a dataset designed to evaluate challenging exploration scenar", + "bbox": [ + 89, + 869, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "ios with diverse and distant viewpoints. We also assess our method on the Nerfbusters dataset [40], curated to better reflect scene exploration tasks. Experimental results show that our approach achieves the highest scores among 3DGS-based methods. Finally, an interactive exploration demo demonstrates that our method enables broader scene coverage with fewer artifacts compared to existing approaches.", + "bbox": [ + 511, + 90, + 903, + 196 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions can be organized as follows:", + "bbox": [ + 511, + 198, + 903, + 226 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a pipeline for explorable 3D scene reconstruction, which incorporates the real-time rendering of 3DGS, video diffusion priors to complete missing regions, and an uncertainty-based finetuning strategy.", + "- We propose a novel virtual camera sampling strategy to identify viewpoints that can maximize additional information gain, thereby effectively reducing artifacts and missing regions.", + "- We introduce Wild-Explore, a new benchmark dataset designed to evaluate challenging exploration scenarios." + ], + "bbox": [ + 511, + 228, + 903, + 378 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 395, + 650, + 410 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Novel view synthesis", + "text_level": 1, + "bbox": [ + 513, + 420, + 705, + 435 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Novel view synthesis aims to generate images of a scene from unseen viewpoints. NeRF [24] and its following works [1, 2] have revolutionized novel view synthesis by representing 3D scenes as continuous volumetric functions, significantly improving rendering quality yet still suffering from slow rendering speed. Recently, 3D Gaussian Splatting [11] has enabled real-time rendering through an explicit Gaussian ellipsoid representation and its differentiable splatting pipeline. Despite these advances, all these methods produce degraded results when the number of images or their view coverage are limited. To address these shortcomings, several works [15, 26, 35, 43] incorporate geometric constraints and depth regularization, leading to less artifacts and improving reconstruction fidelity under sparse-view conditions. However, they remain insufficient for reconstructing explorable scenes, as they lack the ability to fill missing information in the training views.", + "bbox": [ + 511, + 443, + 906, + 700 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Evaluating novel view synthesis", + "text_level": 1, + "bbox": [ + 511, + 710, + 790, + 727 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Previous novel view synthesis methods primarily focus on view interpolation, with LLFF [23], Blender synthetic dataset [24], TanksTemples [12] and MipNeRF360 [1] being widely used benchmarks. These datasets place test viewpoints between training viewpoints and thus are not suitable for evaluating performance under large viewpoint changes. To address this issue, Nerfbusters [40] introduces a new dataset with large viewpoint variations by capturing scenes from separate camera paths. Nevertheless, most of the test views in Nerfbusters lie behind the training views, which are largely unconstrained and primarily require gen-", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "27043", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "erating arbitrary contents rather than completing the underreconstructed target scenes.", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.3. Diffusion prior for 3D reconstruction", + "text_level": 1, + "bbox": [ + 89, + 131, + 413, + 148 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since optimization-based NVS methods cannot generate novel contents, recent works [20, 28, 30, 37, 38] have explored methods for exploiting generative diffusion priors. The vast majority of recent research [4-8, 13, 17, 21, 32, 36, 39] train multi-view diffusion models or fine-tune video diffusion models on large multi-view datasets [16, 29, 47] for generating 3D contents. Among these, 3DGS-Enhancer [21] fine-tunes a video diffusion model to enhance a sequence of 3D Gaussian splatting renderings to solve sparse-view reconstruction task. On the one hand, some works [19, 33] apply diffusion prior to address view extrapolation. ExtraNeRF [33] introduces scene-level finetuned image inpainting diffusion and depth-completion model. ViewExtrapolator [19] leverages SVD [3] in a training-free manner. While these methods are most closely related to our work, they assume known evaluation viewpoints and require running the diffusion model for every target view to render. Moreover, their diffusion models lack multiview consistency, leaving their ability of reconstructing explorable scenes unclear.", + "bbox": [ + 89, + 154, + 483, + 457 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.4. Next-view selection for data acquisition", + "text_level": 1, + "bbox": [ + 89, + 467, + 429, + 484 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A number of works studied methods to select the next viewpoint to acquire data when the number of input viewpoints are limited. Early works focus on active learning or information-based strategies for efficient data acquisition [9, 14, 27, 31, 34, 41]. In particular, FisherRF [9] leverages the Fisher information matrix to quantify the information gain or uncertainty of a given viewpoint, serving as a strong baseline for next-best-view selection. However, these methods primarily target single-object or forward-facing scenes, while our method demonstrates superior performance in complex scenes. Recent works extend this line of works to autonomous exploration [10, 18, 22, 45]. Instead of capturing additional inputs, ExploreGS samples virtual viewpoints from initially reconstructed scenes and exploits generative priors to synthesize images from that viewpoints.", + "bbox": [ + 89, + 489, + 483, + 717 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 731, + 181, + 747 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Overview", + "text_level": 1, + "bbox": [ + 89, + 756, + 200, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The overall pipeline of ExploreGS is illustrated in Fig. 2. Our pipeline follows three stages: (1) scene initialization, (2) pseudo observation generation with virtual camera sampling and a video diffusion prior, and (3) fine-tuning 3DGS using both the training and the generated observations sets.", + "bbox": [ + 89, + 777, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the first stage, we begin by optimizing 3D Gaussians, denoted as $\\mathcal{G}$ , to represent the 3D scene using the given training viewpoints $\\mathcal{V}^T = \\{V_i^T\\}_{i=1}^{N_T}$ where $N_T$ denotes the", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "number of training viewpoints. Here, each $V_{i}^{T}$ has known camera pose and ground truth image $I_{i}^{T}$ . Then, we determine the boundary of reconstructable scene based on the input observations and identify occupied regions for virtual viewpoints samplings.", + "bbox": [ + 511, + 90, + 903, + 167 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In the second stage, we sample virtual camera trajectories based on information gain of each candidate virtual viewpoint, as shown in Sec. 3.3. $N_{Tr}$ camera trajectories that maximizes the information gain are sampled, and each trajectory $Tr_{n}$ consists of $L$ progressively shifting virtual viewpoints: $Tr_{n} = \\{V_{l}^{G}\\}_{l=1}^{L}$ . Along these virtual trajectories, we render images to obtain $I_{i}^{G}$ and generate pseudo-observations $I_{i}^{P}$ from $I_{i}^{G}$ by enhancing them using the generative prior of a video diffusion model.", + "bbox": [ + 511, + 167, + 906, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Finally, we reconstruct complete and explorable 3D scenes by fine-tuning the initially optimized 3D Gaussians using both the training and virtual viewpoints, as described in Sec. 3.5. To mitigate the impact of inconsistencies between the generated and the original images during fine-tuning, we propose two techniques: (1) image-level confidence score, and (2) pixel-level confidence map.", + "bbox": [ + 511, + 303, + 905, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Scene initialization", + "text_level": 1, + "bbox": [ + 511, + 419, + 696, + 434 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Initial reconstruction. Our pipeline begins by optimizing the initial set of 3D Gaussians with the given training set within the 3DGS optimization framework [11]. Then, we estimate the target bounding box of the scene, which is a maximum boundary that can be reconstructed based on the given input images. As previously discussed, reconstructing content beyond this bounding box is highly challenging, as it lacks grounding in the input observations and increasingly resembles unconstrained content generation, which is beyond the scope of this work. This bounding box is computed by using the mesh extracted from the initial 3D Gaussians and the given cameras.", + "bbox": [ + 511, + 440, + 905, + 623 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Rasterization-based occupancy estimation Since renderings from cameras inside occupied regions (e.g. behind walls or inside solid objects such as an apple) offer redundant images, not relevant to the scene, they should be located in free space. To this end, we introduce a simple rasterization-based algorithm to construct the occupancy grid $\\mathcal{O} \\in \\mathbb{R}^{\\mathbb{S} \\times \\mathbb{S} \\times \\mathbb{S}}$ . We estimate the visibility of each 3D Gaussians using the Gaussian rasterizer by computing transmittance values from the training views. Among these, we average the top-3 visibility scores to obtain the final visibility estimate. A vertex is classified as occupied if the estimated visibility is below threshold $\\tau$ (we use 0.5); otherwise, it is considered free space.", + "bbox": [ + 511, + 642, + 906, + 838 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.3. Virtual view sampling", + "text_level": 1, + "bbox": [ + 511, + 848, + 720, + 864 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "After initializing the target scene, our method utilizes video diffusion priors to supplement the missing information from", + "bbox": [ + 511, + 869, + 905, + 902 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "27044", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/988f300a2f13c7b1e2d823127a654433cb37f4b76594f644c8b4b6638a5fa0e5.jpg", + "image_caption": [ + "(a) 3DGS training", + "(b) Virtual camera sampling & diffusion process", + "(c) Confidence-aware finetuning (Sec. 3.4)" + ], + "image_footnote": [], + "bbox": [ + 98, + 88, + 684, + 299 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/fc68156097f57e525983380453d7deda29ec136392e862142f2c57f3ae534d0b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 691, + 97, + 880, + 294 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/e2d02d3077b486d4718336ed42690db8b567db02a106d70bbf79dd31a1e48b23.jpg", + "image_caption": [ + "Figure 2. Overview of the proposed framework for scene exploration. (a) The scene is initially optimized using 3DGS on the given training viewpoints. (b) Based on the optimized 3DGS and training viewpoints, we generate virtual camera trajectories, and enhance the rendered views using our diffusion-based enhancement model. (c) Finally, the scene is further optimized using both the original training viewpoints and the newly generated virtual viewpoints.", + "Figure 3. (a) Viewpoint candidates for virtual camera viewpoint generation. (b) Information gain of each viewpoint. Simplified 2D examples of both are presented for clarity." + ], + "image_footnote": [], + "bbox": [ + 96, + 415, + 480, + 621 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "the input views and removes unwanted artifacts and missing regions. To effectively utilize diffusion priors, our sampling strategy selects virtual camera trajectories that can maximize the information gain of the scene.", + "bbox": [ + 89, + 688, + 482, + 747 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As illustrated in Fig. 2, our approach follows a best-first search strategy with depth-first characteristics to construct virtual trajectories. We begin by selecting $N_{Tr}$ training viewpoints at spatially uniform intervals as the initial states for each trajectory and gradually select the next viewpoint. At each step, the current viewpoint expands into $K$ candidate views. The one with the highest information gain becomes the next viewpoint in the trajectory, while the rest are stored for future consideration. Here, the current state refers to the most recently added camera in a trajectory.", + "bbox": [ + 89, + 750, + 482, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Candidate views. We define candidate views as the set of possible camera motion primitives, which determine the transitions from the current state. As shown in Fig. 3, our action space consists of 6 translational moves, 4 pure rotations of the camera, and 4 rotations where the camera orbits around the look-at point, totaling 14 actions. This choice ensures the virtual trajectory efficiently exploring the scene when compared to previous works just using going straight or turning around.", + "bbox": [ + 511, + 409, + 906, + 545 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "A candidate view is removed if it (1) lies outside the bounding box or in an occupied region, or (2) is close to the 3D Gaussians before fine-tuning, since rendering can be severely disrupted when the camera is placed near the occupied. We utilize the target scene bounding box and the occupancy grid from Sec. 3.2 to achieve this. These constraints prevent the trajectory from diverging away from the reconstructed scene or producing redundant pseudo-observations.", + "bbox": [ + 511, + 547, + 908, + 670 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Information gain. We aim to observe 3D Gaussians densely from diverse viewpoints, as sparse or limited-angle observations often lead to inaccurate shape and color estimation. To this end, we measure information gain based on view coverage. We employ a 2D map $\\mathcal{M} \\in \\mathbb{R}^{N_{\\mathcal{G}} \\times D}$ , where $N_{\\mathcal{G}}$ is the total number of Gaussians and $D$ is the number of discretized viewing directions around 3D Gaussians. The map $\\mathcal{M}$ stores binary values where each entry $\\mathcal{M}[j,k] = 1$ indicates that the $j$ -th Gaussian has been observed from the $k$ -th viewing direction and initialized with training viewpoints $\\mathcal{V}^T$ . At each step $t$ , we update the map as follows:", + "bbox": [ + 511, + 698, + 908, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {M} _ {t} [ j, k ] = 1, \\quad \\forall j \\in J, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 883, + 906, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "27045", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $J$ denotes the set of indices of 3D Gaussians visible from the candidate viewpoint and $k$ indicates the index assigned to the discretized view direction of the candidate viewpoint. This update is performed based on the previous map $\\mathcal{M}_{t-1}$ , which can be described as", + "bbox": [ + 89, + 90, + 483, + 167 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI G _ {t} = \\sum \\left(\\mathcal {M} _ {t} - \\mathcal {M} _ {t - 1}\\right), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 194, + 180, + 482, + 202 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $IG_{t}$ is the number of Gaussians that have either been newly observed or newly viewed from different directions compared to $\\mathcal{M}_{t - 1}$ . 3D Gaussians that are newly observed or seen from a different angle compared to $\\mathcal{M}_{t - 1}$ , ensuring that only newly gained information is counted. To avoid redundant overlaps among trajectories, $\\mathcal{M}$ is shared across among all trajectories $Tr$ .", + "bbox": [ + 89, + 215, + 483, + 321 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Termination. The search process terminates under two conditions. First, if no valid candidate views can be expanded from the current viewpoint while the number of cameras in the trajectory is still below the predefined length $L$ , the current branch is pruned and the search resumes from the next-best candidate view in the priority queue. Second, if the trajectory reaches the predefined length $L$ , the search concludes. The resulting trajectory is treated as a set of pseudo-observations, and then begins the searching process for the next trajectory $Tr_{i+1}$ . This process is repeated until $N_{Tr}$ trajectories are constructed. We provide visualizations of example results in the supplementary materials.", + "bbox": [ + 89, + 345, + 483, + 527 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Diffusion prior", + "text_level": 1, + "bbox": [ + 89, + 539, + 243, + 556 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Since renderings from virtual viewpoints often partially include well-reconstructed parts, it is desirable to leverage these visual cues. To this end, we train a video diffusion model to enhance degraded rendered images $I^{V}$ to be realistic images $I^{G}$ . Our model is built on [21] but introduces several key modifications. First, unlike [21], which adopts a frame interpolation setup, we address extreme view extrapolation in scene exploration by performing next-frame prediction using only the observation closest to the virtual camera as the sole visual cue. Second, to mitigate this ill-posed task, we condition the model on the scene text description and the ground truth image $I^{T}$ from the nearest training view to each pseudo-observation. These are injected into the cross-attention layers of the diffusion model. At last, we create a dataset specifically designed to model artifacts arising from extreme extrapolation. The details about the dataset and the training process for the diffusion model are provided in the supplementary materials.", + "bbox": [ + 89, + 561, + 483, + 835 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Finetuning", + "text_level": 1, + "bbox": [ + 89, + 847, + 212, + 863 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Previous work [21] observed that applying an image-based loss (e.g., L1 loss) to a rendered image from a virtual", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "camera and its pseudo ground truth often degrades well-reconstructed regions, as 3DGS is sensitive to minor inaccuracies from diffusion model. It mitigates this issue by using two confidence levels, where higher confidence values indicate fewer conflicts between the contents of training and virtual viewpoints, thereby guiding 3D Gaussians to receive stronger gradients. It adopts the image level confidence based on the distance between a virtual viewpoint and a training viewpoint and pixel level confidence based on the scale of 3D Gaussians.", + "bbox": [ + 511, + 90, + 903, + 239 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Higher image-level confidence is applied when virtual viewpoints far from training viewpoints since they have less conflicts between them. While effective, using distance alone often fails to account for pure camera rotations, as the distance remains unchanged even when the view frustum overlap between viewpoints decreases. As our generated trajectories involve both translations and rotations, this results in suboptimal supervision.", + "bbox": [ + 511, + 242, + 905, + 362 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To address this issue, we introduce a novel image-level confidence metric based on intersection-over-union (IoU) of Gaussians, denoted as $G$ -IOU. Our key insight is that a high view frustum overlaps between a training viewpoint and a virtual viewpoint implies many co-visible 3D Gaussians, leading to more conflicts. Using the Gaussian rasterizer, we efficiently extract the indices of visible Gaussians from each view, and define their intersection as the co-visible set:", + "bbox": [ + 511, + 363, + 905, + 484 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right) = \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {V}\\right) \\cap \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 491, + 906, + 512 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\mathcal{G}_{vis}(\\cdot)$ denotes the set of 3D Gaussians visible from a given viewpoint, and $\\nu_{ref}$ refers to the training viewpoint nearest to $\\nu^V$ . Based on this, $G$ -IOU is defined as", + "bbox": [ + 511, + 520, + 905, + 565 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nG I O U (\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}) = \\frac {\\left| \\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V} , \\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}{\\left| \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 573, + 903, + 609 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "and the image-level confidence is given by", + "bbox": [ + 511, + 616, + 795, + 631 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {\\mathrm {i m g}} = 1 - G \\mathrm {I O U} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 609, + 638, + 903, + 657 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "For the pixel level confidence, rather than solely relying on the scale of 3D Gaussians which still produce artifacts depending on the viewpoint, we define it as the perceptual distance [46] between the rendered image $I^{V}$ and the output of the diffusion model $I^{G}$ . Each pixel in the confidence map reflects its contribution to the overall perceptual difference. We upsample the pixel-aligned LPIPS map to the original image resolution and define pixel-level confidence as", + "bbox": [ + 511, + 667, + 905, + 787 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nU _ {\\text {p i x e l}} (x, y) = \\mathcal {U} (\\mathrm {L P I P S} \\left(I ^ {V}, I ^ {G}\\right)) (x, y), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 573, + 797, + 903, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $U_{\\mathrm{pixel}}(x,y)$ denotes the confidence value at pixel $(x,y)$ and $\\mathcal{U}$ is an upsampling operation to the original image resolution.", + "bbox": [ + 511, + 825, + 903, + 869 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Combining the above confidence, the loss for training views is defined as", + "bbox": [ + 511, + 869, + 903, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "27046", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/441cbd61c03ccda638c57383f59633b4019a369426d83303a11ce7ff32936151.jpg", + "image_caption": [ + "Training viewpoints Evaluation viewpoints Evaluation viewpoints (2nd)", + "Figure 4. Viewpoint difference comparison with previous 3D reconstruction dataset. Training and evaluation viewpoints are visualized." + ], + "image_footnote": [], + "bbox": [ + 94, + 106, + 480, + 334 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V} ^ {T}} = \\left\\| I ^ {T} - I ^ {V} \\right\\| _ {1} + \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(I ^ {T}, I ^ {V}\\right), \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 402, + 480, + 422 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "and the loss for virtual views is defined as", + "bbox": [ + 89, + 428, + 372, + 441 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathcal {V} ^ {V}} = U _ {i m g} \\cdot \\left(\\left| \\left| I ^ {G} - I ^ {V} \\right| \\right| _ {1} \\odot U _ {p i x e l} + \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I ^ {G}, I ^ {V}\\right)\\right). \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 450, + 483, + 483 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The $\\odot$ denotes the pixel-wise multiplication and $\\mathcal{L}_{\\mathrm{D - SSIM}}$ denotes the dissimilarity variant of the SSIM loss. This fine-tuning process is also illustrated in Fig. 2 (c).", + "bbox": [ + 89, + 484, + 483, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Evaluation dataset for scene exploration", + "text_level": 1, + "bbox": [ + 89, + 542, + 450, + 559 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Since existing datasets remain inadequate for evaluating scene exploration, as the test cameras are typically positioned within or close to the training trajectories, resulting in interpolation scenarios or only minor viewpoint variations. Although the Nerfbusters dataset [40] introduces extreme novel viewpoints via two separate camera trajectories, its test views are placed far behind the training views. Consequently, the task aligns more with generative outpainting rather than complete scene reconstruction. We introduce two datasets tailored for this task, with additional details provided in the supplementary materials.", + "bbox": [ + 89, + 566, + 483, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. WildExplore", + "text_level": 1, + "bbox": [ + 89, + 742, + 225, + 758 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To address the lack of an appropriate benchmark for scene exploration, we introduce WildExplore, a new dataset comprising four indoor and four outdoor scenes. As illustrated in Fig. 4, WildExplore is designed with two key properties: (1) it explicitly targets the exploration task by placing test viewpoints significantly distant from the training trajectories, ensuring large viewpoint transitions and broad scene coverage, and (2) the evaluation set is captured to ensure sufficient co-visible areas between training and test views,", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "minimizing the need for generating arbitrary contents, in contrast to Nerfbusters.", + "bbox": [ + 511, + 90, + 906, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Curated Nerfbusters", + "text_level": 1, + "bbox": [ + 511, + 130, + 709, + 145 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We curate the Nerfbusters dataset to better align with scene exploration objectives. We selects some scenes where the test views are placed far behind the training views and swap the train/test split, while leaving the others unchanged. This adjustment transforms the task into \"moving forward and closely examining the scene\", which better reflects realistic exploration scenarios. Our curated version includes seven scenes with swapped splits and two in their original form.", + "bbox": [ + 511, + 152, + 906, + 273 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 511, + 287, + 645, + 305 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Experimental Setup", + "text_level": 1, + "bbox": [ + 511, + 313, + 702, + 330 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation details. We first train 3DGS using the given training viewpoints, following the original setup in [11]. The number of virtual trajectories $N$ is set to 20 by default. The occupancy grid resolution $S$ is set to 64. At each step, we select the top 3 candidate views based on computed information gain. We set $D$ as 32 for discretized view direction. For the video diffusion model, we adopt DynamiCrafter [42] as the backbone. After generating the pseudo observation set, we fine-tune 3D Gaussians for 15K iterations, applying densification until 9K, with the same scheduling as 3DGS. Further details will be discussed in the supplements.", + "bbox": [ + 511, + 335, + 906, + 517 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Baselines. We evaluate our method in comparison to 3DGS-based methods, including 3DGS [11], 3DGS + Depth, DNGaussians [15], and ViewExtrapolator [19]. 3DGS + Depth applies additional depth regularization using monocular depth estimation [44], following the implementation from the official 3DGS repository [11]. DNGaussians [15] incorporates depth priors during partial optimization of the 3D Gaussian parameters. ViewExtrapolator leverages video diffusion but heuristically selects a single test camera for generating additional viewpoints due to the absence of a dedicated sampling strategy. In our experiments, we follow their protocol and choose the test camera farthest from the training set to encourage broader viewpoint coverage.", + "bbox": [ + 511, + 536, + 908, + 733 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Results", + "text_level": 1, + "bbox": [ + 511, + 742, + 607, + 757 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative comparison. Table 1 reports the quantitative results on the exploration task. To assess reconstruction quality, we measure PSNR, SSIM, and LPIPS [46]. The first row reports the result from 3DGS trained on both the training and test splits, serving as an upper bound. Since our method effectively removes artifacts and fills missing regions, as observed in the qualitative analysis, our model outperforms across all metrics, particularly in PSNR and SSIM.", + "bbox": [ + 511, + 763, + 906, + 898 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "27047", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8b4e5d921fcfe5dfb336ead3abf129ae68585315f5c66be4ae5fe963f8748428.jpg", + "image_caption": [ + "GT", + "3DGS", + "3DGS+Depth", + "DNGaussian", + "ViewExtrapolator", + "Ours", + "Figure 5. Qualitative comparison on the curated Nerfbusters dataset." + ], + "image_footnote": [], + "bbox": [ + 94, + 111, + 903, + 623 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative comparison. Fig. 5 and Fig. 6 show qualitative comparisons among our method and baseline methods. 3DGS [11] suffers from artifacts, and its variants with depth regularization also meet the same problem, as they lack the capability to fill missing information. Although ViewExtrapolator [19] utilizes a video diffusion prior and thereby shows competitive performance compared to 3DGS variants, it underperforms in challenging scenarios. Its heuristic camera sampling struggles to maximize information gain and results in limited viewpoint coverage, leading to lower visual quality when rendering from distant viewpoints. Its training-free diffusion prior also acts as a bottleneck. In contrast, our method fills missing regions and removes artifacts more effectively, producing images that align closely with the ground truth. See the supplementary materials for additional results.", + "bbox": [ + 89, + 652, + 482, + 893 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/24e08622fa8e69d4c1a60811a414fc87cbd6303d476990dec6950da232d11509.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodWildExploreCurated Nerfbusters
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS Pseudo GT30.370.9100.14429.340.9110.128
3DGS15.810.4870.41414.510.4170.454
3DGS + Depth15.780.4830.41315.000.4350.436
DNGaussians15.650.5040.45014.010.4420.525
ViewExtrapolator16.020.5100.44214.980.4500.455
Ours16.960.5340.41316.290.4790.434
", + "bbox": [ + 517, + 650, + 903, + 756 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparison.", + "bbox": [ + 607, + 762, + 810, + 776 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.3. Ablation study", + "text_level": 1, + "bbox": [ + 511, + 800, + 663, + 816 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Information gain. We validate the effectiveness of our information gain design by comparing it with alternative strategies, as shown in Table 2. One alternative samples the 3D space uniformly and evaluate view coverage based on the occupied vertices of an occupancy grid, as proposed in", + "bbox": [ + 511, + 824, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "27048", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87c4858e83b335710e494b0d46297f3a9ccb3428cc34b19b364781e2e137c723.jpg", + "image_caption": [ + "Figure 6. Qualitative comparison on the WildExplore dataset." + ], + "image_footnote": [], + "bbox": [ + 93, + 93, + 906, + 333 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/c4bb2f5fff4ed1f7fa02b39174673d37a9563ff2c115a8798fa229f621a5a6e4.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
Virtual View SamplingCurated Nerfbusters
Info. GainPSNR↑SSIM↑LPIPS↓
Grid view coverage [22]15.820.46750.4451
FisherRF [9]15.770.45930.4394
GS view coverage (Ours)16.290.47900.4336
", + "bbox": [ + 96, + 368, + 478, + 460 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/969a2e826341533e730d75a5884bdc3f444558648b412c3c8cfb2cb523f9af89.jpg", + "table_caption": [ + "Table 2. Comparisons on information gain design." + ], + "table_footnote": [], + "table_body": "
Method (K = 3)Garbage
PSNR↑SSIM↑LPIPS↓
3DGS14.420.4800.526
Ours with Bottom-K14.220.4750.521
Ours with Top-K15.450.5260.508
", + "bbox": [ + 133, + 491, + 439, + 568 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/e0ea9b5eff8efb347febdd71e32aef4e928d3ac57d2cd439268098762d745949.jpg", + "table_caption": [ + "Table 3. Ablation study on information gain. TopK vs BottomK" + ], + "table_footnote": [], + "table_body": "
FinetuningCurated Nerfbusters
Image levelPixel levelPSNR↑SSIM↑LPIPS↓
Distance [21]-15.000.4270.443
-Scale [21]16.180.4760.442
DistanceScale15.270.4320.440
--16.300.4760.447
-pixelLPIPS16.230.4780.441
GS-IOU-16.100.4730.439
GS-IOUpixelLPIPS16.290.4790.434
", + "bbox": [ + 102, + 602, + 468, + 738 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 4. Ablations study on finetuning methods.", + "bbox": [ + 142, + 742, + 428, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "[22]. Another follows typical active-learning strategies and select viewpoints with the highest uncertainty. We adopt the method proposed in [9] for comparison. As presented in the first and second rows, both show suboptimal results. Grid-based approach often fails to maximize information gain, as it includes the gain from free space, resulting in redundant viewpoint selections. For uncertainty-based method, we observe that it is sensitive to 3D Gaussians with large scales,", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "tending to chase them rather than addressing artifacts from viewpoint changes.", + "bbox": [ + 511, + 369, + 906, + 402 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Search strategy. We ablate our best first search strategy to validate its effectiveness. As shown in Table 3, the strategy selecting bottom K makes virtual cameras tend to stay near the trajectory of training viewpoints rather than deviating from it, leading to no performance gain.", + "bbox": [ + 511, + 419, + 906, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Finetuning methods. We validate our finetuning approach by benchmarking it against the methods proposed by [21], as shown in Table 4. The model without the image-level confidence produces worse results in terms of LPIPS, as generated images whose viewpoints closely match the training set negatively impact already well-reconstructed regions. Ignoring camera rotation further degrades performance, as observed in the second and fifth rows. Although both our image-level confidence map and scale based one are effective, our method leads to slight better performance.", + "bbox": [ + 511, + 512, + 908, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 676, + 633, + 693 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper presents a framework that enables high-quality rendering of the novel view in arbitrary viewpoints. Our method incorporates the video diffusion priors and the real-time rendering capability of 3DGS. It introduces a sampling strategy to effectively sample trajectories of virtual cameras for maximizing information gain.", + "bbox": [ + 511, + 703, + 906, + 792 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works. While our method incorporates video diffusion priors to fill the missing information, we expect that employing more strong diffusion models which can faithfully generate high frequency details may improve the overall performance. In addition, extending the scene bounding box to cover a large scale scene would be an interesting avenue for the future work.", + "bbox": [ + 511, + 792, + 908, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "27049", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220113, No. RS-2024-00457882), Artificial Intelligence Graduate School Program, Yonsei University, under Grant RS-2020-11201361.", + "bbox": [ + 89, + 113, + 485, + 200 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 224, + 187, + 242 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 1, 2", + "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields, 2023. 2", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 3", + "[4] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models, 2024. 3", + "[5] Jeong gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models, 2023.", + "[6] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation, 2024.", + "[7] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024.", + "[8] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2024. 3", + "[9] Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information. arXiv preprint arXiv:2311.17874, 2023. 3, 8", + "[10] Liren Jin, Xieyuanli Chen, Julius Rückin, and Marija Popovic. Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11305-11312, 2023. 3", + "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 6, 7" + ], + "bbox": [ + 93, + 250, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36:1-13, 2017. 2", + "[13] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J. Davison. Eschernet: A generative model for scalable view synthesis, 2024. 3", + "[14] Georgios Kopanas and George Drettakis. Improving NeRF Quality by Progressive Camera Placement for Free-Viewpoint Navigation. In Vision, Modeling, and Visualization. The Eurographics Association, 2023. 3", + "[15] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization, 2024. 2, 6", + "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, Xuanmao Li, Xingpeng Sun, Rohan Ashok, Aniruddha Mukherjee, Hao Kang, Xiangrui Kong, Gang Hua, Tianyi Zhang, Bedrich Benes, and Aniket Bera. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision, 2023. 3", + "[17] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model, 2024. 3", + "[18] Guangyi Liu, Wen Jiang, Boshu Lei, Vivek Pandey, Kostas Daniilidis, and Nader Motee. Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf, 2025. 3", + "[19] Kunhao Liu, Ling Shao, and Shijian Lu. Novel view extrapolation with video diffusion priors, 2024. 2, 3, 6, 7", + "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 3", + "[21] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors, 2024. 2, 3, 5, 8", + "[22] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra S Chaplot. Autonerf: Training implicit scene representations with autonomous agents. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13442-13449. IEEE, 2024. 3, 8", + "[23] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines, 2019. 2", + "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2", + "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1", + "[26] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "27050", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs, 2021. 2", + "[27] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Activenerf: Learning where to see with uncertainty estimation. In European Conference on Computer Vision, pages 230-246. Springer, 2022. 3", + "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion, 2022. 3", + "[29] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction, 2021. 3", + "[30] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. Zeronvs: Zero-shot 360-degree view synthesis from a single image, 2024. 3", + "[31] Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In 2021 International Conference on 3D Vision (3DV), pages 972-981. IEEE, 2021. 3", + "[32] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2024. 3", + "[33] Meng-Li Shih, Wei-Chiu Ma, Lorenzo Boyice, Aleksander Holynski, Forrester Cole, Brian Curless, and Janne Kontkanen. Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20385-20395, 2024. 2, 3", + "[34] Niko Sünderhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields, 2022. 3", + "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses, 2023. 2", + "[36] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 3", + "[37] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 3", + "[38] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 3", + "[39] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation, 2024. 3", + "[40] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs, 2023. 2, 6" + ], + "bbox": [ + 91, + 90, + 482, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Leo Lebrat. Nerf director: Revisiting view selection in neural volume rendering, 2024. 3", + "[42] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 6", + "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360deg sparseviewsynthesisusinggaussiansplatting,", + "[44] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 6", + "[45] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning, 2022. 3", + "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5, 6", + "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images, 2018. 3" + ], + "bbox": [ + 514, + 90, + 903, + 431 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "27051", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_model.json b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_model.json new file mode 100644 index 0000000000000000000000000000000000000000..74a6cec9e540a50d24cf78a916818c7ff14893a1 --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_model.json @@ -0,0 +1,2233 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.232, + 0.131, + 0.768, + 0.177 + ], + "angle": 0, + "content": "ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors" + }, + { + "type": "text", + "bbox": [ + 0.237, + 0.204, + 0.761, + 0.243 + ], + "angle": 0, + "content": "Minsu Kim Subin Jeon In Cho Mijin Yoo Seon Joo Kim Yonsei University" + }, + { + "type": "text", + "bbox": [ + 0.214, + 0.245, + 0.778, + 0.261 + ], + "angle": 0, + "content": "{minsul206, subinjeon, join, mijin6690, seonjookim}@yonsei.ac.kr" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.273, + 0.364, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.369, + 0.273, + 0.636, + 0.456 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.273, + 0.906, + 0.457 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.429, + 0.457, + 0.571, + 0.47 + ], + "angle": 0, + "content": "(a) Exploration Results" + }, + { + "type": "image", + "bbox": [ + 0.101, + 0.472, + 0.48, + 0.564 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.173, + 0.568, + 0.47, + 0.582 + ], + "angle": 0, + "content": "(b) Virtual Camera Sampling with Diffusion Prior" + }, + { + "type": "image", + "bbox": [ + 0.557, + 0.472, + 0.877, + 0.567 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.641, + 0.568, + 0.793, + 0.58 + ], + "angle": 0, + "content": "(c) Scene Reconstruction" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.59, + 0.908, + 0.634 + ], + "angle": 0, + "content": "Figure 1. (a) Explorable 3D scene reconstruction results. Our method renders photorealistic images from any arbitrary viewpoints. (b) Virtual cameras are sampled in regions with missing information, and pseudo-ground truth is generated using diffusion priors. (c) After optimizing 3D Gaussians with virtual viewpoint samples, our method enables real-time scene exploration." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.645, + 0.327, + 0.66 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.69, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Recent advances in novel view synthesis (NVS) have enabled real-time rendering with 3D Gaussian Splatting (3DGS). However, existing methods struggle with artifacts and missing regions when rendering from viewpoints that deviate from the training trajectory, limiting seamless scene exploration. To address this, we propose a 3DGS-based pipeline that generates additional training views to enhance reconstruction. We introduce an information-gain-driven virtual camera placement strategy to maximize scene coverage, followed by video diffusion priors to refine rendered results. Fine-tuning 3D Gaussians with these enhanced views significantly improves reconstruction quality. To evaluate our method, we present Wild-Explore, a benchmark designed for challenging scene exploration. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.907, + 0.693 + ], + "angle": 0, + "content": "demonstrate that our approach outperforms existing 3DGS-based methods, enabling high-quality, artifact-free rendering from arbitrary viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.724, + 0.646, + 0.739 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The field of novel view synthesis (NVS) has seen significant advancements, enabling photorealistic rendering of scenes from previously unseen viewpoints. Much of this progress is attributed to NeRF [24] and its variants [1, 25]. This advancement has been further accelerated by recent 3D Gaussian Splatting (3DGS) [11], which enables high-quality rendering in real-time. The real-time capability of 3DGS broadens the applications of NVS in domains such as AR, VR, and gaming, leading users to expect the experience of freely exploring reconstructed scenes." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27042" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.227 + ], + "angle": 0, + "content": "Unfortunately, such an experience is yet to be fully realized, as existing methods suffer from severe degradations in rendering quality when viewpoints deviate significantly from the input observations. These degradations include holes, oversized Gaussians, spiky surface artifacts, and spurious floaters, often arising from occlusions or limited viewpoint coverage (see Fig. 1 (a)). This limitation stems from missing information, since optimization-based approaches cannot synthesize contents beyond the observed data." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.229, + 0.482, + 0.365 + ], + "angle": 0, + "content": "We are interested in 3D scene reconstruction that can render more vivid and photorealistic images from arbitrary viewpoints, so that users can freely explore the reconstructed scenes. We term this task as \"explorable scene reconstruction\", which aims to reconstruct complete scenes within the target areas. To avoid generating arbitrary contents in regions far beyond the input observations—which would not reflect actual user experiences—we define the target areas as a bounding box defined by the input views." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.367, + 0.482, + 0.472 + ], + "angle": 0, + "content": "One straightforward solution is to take thousands of images to completely cover the target scenes. However, collecting such dense data is time-consuming and labor-intensive, making it impractical for many real-world applications. Thus, we aim to reconstruct complete 3D scenes from input images that only partially cover the target scenes, which better reflects practical capturing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.474, + 0.482, + 0.595 + ], + "angle": 0, + "content": "In this paper, we introduce ExploreGS, a pipeline that enables explorable scene reconstruction using diffusion priors and 3DGS. ExploreGS leverages video diffusion priors to generate additional training views that complement missing information. These synthesized images at virtual viewpoints serve as pseudo ground truths to refine the initial, incomplete 3D Gaussians, thereby enabling complete, photorealistic rendering at arbitrary viewpoints." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.482, + 0.746 + ], + "angle": 0, + "content": "The key challenges of explorable scene reconstruction lie in determining the optimal placement of virtual viewpoints. Unlike previous works that assume evaluation trajectories are known [19, 33], virtual viewpoints in our pipeline need to be sampled to best complement missing information in poorly covered areas. We propose a novel view sampling strategy that strategically positions virtual cameras based on information gain. This approach identifies viewpoints that offer the most novel scene details from the initial 3DGS reconstruction as shown in Fig. 1 (b)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.749, + 0.482, + 0.869 + ], + "angle": 0, + "content": "To provide pseudo observations, we generate realistic images along sampled viewpoints. We introduce video diffusion model which converts renderings with degradations into clean images. Although it is motivated by 3DGS-Enhancer [21], we modify the training objectives to fit scene exploration and prepare larger dataset for training. We then finetune 3DGS on both virtual viewpoints and training viewpoints with uncertainty guidance." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.482, + 0.901 + ], + "angle": 0, + "content": "To validate our pipeline, we introduce Wild-Explore, a dataset designed to evaluate challenging exploration scenar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.198 + ], + "angle": 0, + "content": "ios with diverse and distant viewpoints. We also assess our method on the Nerfbusters dataset [40], curated to better reflect scene exploration tasks. Experimental results show that our approach achieves the highest scores among 3DGS-based methods. Finally, an interactive exploration demo demonstrates that our method enables broader scene coverage with fewer artifacts compared to existing approaches." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.199, + 0.905, + 0.227 + ], + "angle": 0, + "content": "In summary, our contributions can be organized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.229, + 0.905, + 0.289 + ], + "angle": 0, + "content": "- We propose a pipeline for explorable 3D scene reconstruction, which incorporates the real-time rendering of 3DGS, video diffusion priors to complete missing regions, and an uncertainty-based finetuning strategy." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.29, + 0.905, + 0.349 + ], + "angle": 0, + "content": "- We propose a novel virtual camera sampling strategy to identify viewpoints that can maximize additional information gain, thereby effectively reducing artifacts and missing regions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.35, + 0.904, + 0.38 + ], + "angle": 0, + "content": "- We introduce Wild-Explore, a new benchmark dataset designed to evaluate challenging exploration scenarios." + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.229, + 0.905, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.396, + 0.651, + 0.411 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.421, + 0.706, + 0.436 + ], + "angle": 0, + "content": "2.1. Novel view synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.444, + 0.907, + 0.701 + ], + "angle": 0, + "content": "Novel view synthesis aims to generate images of a scene from unseen viewpoints. NeRF [24] and its following works [1, 2] have revolutionized novel view synthesis by representing 3D scenes as continuous volumetric functions, significantly improving rendering quality yet still suffering from slow rendering speed. Recently, 3D Gaussian Splatting [11] has enabled real-time rendering through an explicit Gaussian ellipsoid representation and its differentiable splatting pipeline. Despite these advances, all these methods produce degraded results when the number of images or their view coverage are limited. To address these shortcomings, several works [15, 26, 35, 43] incorporate geometric constraints and depth regularization, leading to less artifacts and improving reconstruction fidelity under sparse-view conditions. However, they remain insufficient for reconstructing explorable scenes, as they lack the ability to fill missing information in the training views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.712, + 0.792, + 0.728 + ], + "angle": 0, + "content": "2.2. Evaluating novel view synthesis" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Previous novel view synthesis methods primarily focus on view interpolation, with LLFF [23], Blender synthetic dataset [24], TanksTemples [12] and MipNeRF360 [1] being widely used benchmarks. These datasets place test viewpoints between training viewpoints and thus are not suitable for evaluating performance under large viewpoint changes. To address this issue, Nerfbusters [40] introduces a new dataset with large viewpoint variations by capturing scenes from separate camera paths. Nevertheless, most of the test views in Nerfbusters lie behind the training views, which are largely unconstrained and primarily require gen-" + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27043" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "erating arbitrary contents rather than completing the underreconstructed target scenes." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.132, + 0.414, + 0.149 + ], + "angle": 0, + "content": "2.3. Diffusion prior for 3D reconstruction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.155, + 0.485, + 0.458 + ], + "angle": 0, + "content": "Since optimization-based NVS methods cannot generate novel contents, recent works [20, 28, 30, 37, 38] have explored methods for exploiting generative diffusion priors. The vast majority of recent research [4-8, 13, 17, 21, 32, 36, 39] train multi-view diffusion models or fine-tune video diffusion models on large multi-view datasets [16, 29, 47] for generating 3D contents. Among these, 3DGS-Enhancer [21] fine-tunes a video diffusion model to enhance a sequence of 3D Gaussian splatting renderings to solve sparse-view reconstruction task. On the one hand, some works [19, 33] apply diffusion prior to address view extrapolation. ExtraNeRF [33] introduces scene-level finetuned image inpainting diffusion and depth-completion model. ViewExtrapolator [19] leverages SVD [3] in a training-free manner. While these methods are most closely related to our work, they assume known evaluation viewpoints and require running the diffusion model for every target view to render. Moreover, their diffusion models lack multiview consistency, leaving their ability of reconstructing explorable scenes unclear." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.468, + 0.43, + 0.485 + ], + "angle": 0, + "content": "2.4. Next-view selection for data acquisition" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.491, + 0.484, + 0.718 + ], + "angle": 0, + "content": "A number of works studied methods to select the next viewpoint to acquire data when the number of input viewpoints are limited. Early works focus on active learning or information-based strategies for efficient data acquisition [9, 14, 27, 31, 34, 41]. In particular, FisherRF [9] leverages the Fisher information matrix to quantify the information gain or uncertainty of a given viewpoint, serving as a strong baseline for next-best-view selection. However, these methods primarily target single-object or forward-facing scenes, while our method demonstrates superior performance in complex scenes. Recent works extend this line of works to autonomous exploration [10, 18, 22, 45]. Instead of capturing additional inputs, ExploreGS samples virtual viewpoints from initially reconstructed scenes and exploits generative priors to synthesize images from that viewpoints." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.732, + 0.183, + 0.748 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.757, + 0.201, + 0.772 + ], + "angle": 0, + "content": "3.1. Overview" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.779, + 0.483, + 0.856 + ], + "angle": 0, + "content": "The overall pipeline of ExploreGS is illustrated in Fig. 2. Our pipeline follows three stages: (1) scene initialization, (2) pseudo observation generation with virtual camera sampling and a video diffusion prior, and (3) fine-tuning 3DGS using both the training and the generated observations sets." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "In the first stage, we begin by optimizing 3D Gaussians, denoted as \\(\\mathcal{G}\\), to represent the 3D scene using the given training viewpoints \\(\\mathcal{V}^T = \\{V_i^T\\}_{i=1}^{N_T}\\) where \\(N_T\\) denotes the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.905, + 0.168 + ], + "angle": 0, + "content": "number of training viewpoints. Here, each \\( V_{i}^{T} \\) has known camera pose and ground truth image \\( I_{i}^{T} \\). Then, we determine the boundary of reconstructable scene based on the input observations and identify occupied regions for virtual viewpoints samplings." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.168, + 0.907, + 0.303 + ], + "angle": 0, + "content": "In the second stage, we sample virtual camera trajectories based on information gain of each candidate virtual viewpoint, as shown in Sec. 3.3. \\( N_{Tr} \\) camera trajectories that maximizes the information gain are sampled, and each trajectory \\( Tr_{n} \\) consists of \\( L \\) progressively shifting virtual viewpoints: \\( Tr_{n} = \\{V_{l}^{G}\\}_{l=1}^{L} \\). Along these virtual trajectories, we render images to obtain \\( I_{i}^{G} \\) and generate pseudo-observations \\( I_{i}^{P} \\) from \\( I_{i}^{G} \\) by enhancing them using the generative prior of a video diffusion model." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.304, + 0.906, + 0.41 + ], + "angle": 0, + "content": "Finally, we reconstruct complete and explorable 3D scenes by fine-tuning the initially optimized 3D Gaussians using both the training and virtual viewpoints, as described in Sec. 3.5. To mitigate the impact of inconsistencies between the generated and the original images during fine-tuning, we propose two techniques: (1) image-level confidence score, and (2) pixel-level confidence map." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.42, + 0.697, + 0.435 + ], + "angle": 0, + "content": "3.2. Scene initialization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.441, + 0.906, + 0.624 + ], + "angle": 0, + "content": "Initial reconstruction. Our pipeline begins by optimizing the initial set of 3D Gaussians with the given training set within the 3DGS optimization framework [11]. Then, we estimate the target bounding box of the scene, which is a maximum boundary that can be reconstructed based on the given input images. As previously discussed, reconstructing content beyond this bounding box is highly challenging, as it lacks grounding in the input observations and increasingly resembles unconstrained content generation, which is beyond the scope of this work. This bounding box is computed by using the mesh extracted from the initial 3D Gaussians and the given cameras." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Rasterization-based occupancy estimation Since renderings from cameras inside occupied regions (e.g. behind walls or inside solid objects such as an apple) offer redundant images, not relevant to the scene, they should be located in free space. To this end, we introduce a simple rasterization-based algorithm to construct the occupancy grid \\(\\mathcal{O} \\in \\mathbb{R}^{\\mathbb{S} \\times \\mathbb{S} \\times \\mathbb{S}}\\). We estimate the visibility of each 3D Gaussians using the Gaussian rasterizer by computing transmittance values from the training views. Among these, we average the top-3 visibility scores to obtain the final visibility estimate. A vertex is classified as occupied if the estimated visibility is below threshold \\(\\tau\\) (we use 0.5); otherwise, it is considered free space." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.722, + 0.866 + ], + "angle": 0, + "content": "3.3. Virtual view sampling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.903 + ], + "angle": 0, + "content": "After initializing the target scene, our method utilizes video diffusion priors to supplement the missing information from" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27044" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.089, + 0.685, + 0.3 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.118, + 0.304, + 0.207, + 0.317 + ], + "angle": 0, + "content": "(a) 3DGS training" + }, + { + "type": "image_caption", + "bbox": [ + 0.342, + 0.305, + 0.573, + 0.317 + ], + "angle": 0, + "content": "(b) Virtual camera sampling & diffusion process" + }, + { + "type": "image", + "bbox": [ + 0.692, + 0.098, + 0.882, + 0.295 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.692, + 0.304, + 0.898, + 0.316 + ], + "angle": 0, + "content": "(c) Confidence-aware finetuning (Sec. 3.4)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.333, + 0.907, + 0.391 + ], + "angle": 0, + "content": "Figure 2. Overview of the proposed framework for scene exploration. (a) The scene is initially optimized using 3DGS on the given training viewpoints. (b) Based on the optimized 3DGS and training viewpoints, we generate virtual camera trajectories, and enhance the rendered views using our diffusion-based enhancement model. (c) Finally, the scene is further optimized using both the original training viewpoints and the newly generated virtual viewpoints." + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.416, + 0.482, + 0.622 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.631, + 0.483, + 0.675 + ], + "angle": 0, + "content": "Figure 3. (a) Viewpoint candidates for virtual camera viewpoint generation. (b) Information gain of each viewpoint. Simplified 2D examples of both are presented for clarity." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.748 + ], + "angle": 0, + "content": "the input views and removes unwanted artifacts and missing regions. To effectively utilize diffusion priors, our sampling strategy selects virtual camera trajectories that can maximize the information gain of the scene." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.902 + ], + "angle": 0, + "content": "As illustrated in Fig. 2, our approach follows a best-first search strategy with depth-first characteristics to construct virtual trajectories. We begin by selecting \\( N_{Tr} \\) training viewpoints at spatially uniform intervals as the initial states for each trajectory and gradually select the next viewpoint. At each step, the current viewpoint expands into \\( K \\) candidate views. The one with the highest information gain becomes the next viewpoint in the trajectory, while the rest are stored for future consideration. Here, the current state refers to the most recently added camera in a trajectory." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.546 + ], + "angle": 0, + "content": "Candidate views. We define candidate views as the set of possible camera motion primitives, which determine the transitions from the current state. As shown in Fig. 3, our action space consists of 6 translational moves, 4 pure rotations of the camera, and 4 rotations where the camera orbits around the look-at point, totaling 14 actions. This choice ensures the virtual trajectory efficiently exploring the scene when compared to previous works just using going straight or turning around." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.548, + 0.909, + 0.671 + ], + "angle": 0, + "content": "A candidate view is removed if it (1) lies outside the bounding box or in an occupied region, or (2) is close to the 3D Gaussians before fine-tuning, since rendering can be severely disrupted when the camera is placed near the occupied. We utilize the target scene bounding box and the occupancy grid from Sec. 3.2 to achieve this. These constraints prevent the trajectory from diverging away from the reconstructed scene or producing redundant pseudo-observations." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.699, + 0.909, + 0.879 + ], + "angle": 0, + "content": "Information gain. We aim to observe 3D Gaussians densely from diverse viewpoints, as sparse or limited-angle observations often lead to inaccurate shape and color estimation. To this end, we measure information gain based on view coverage. We employ a 2D map \\(\\mathcal{M} \\in \\mathbb{R}^{N_{\\mathcal{G}} \\times D}\\), where \\(N_{\\mathcal{G}}\\) is the total number of Gaussians and \\(D\\) is the number of discretized viewing directions around 3D Gaussians. The map \\(\\mathcal{M}\\) stores binary values where each entry \\(\\mathcal{M}[j,k] = 1\\) indicates that the \\(j\\)-th Gaussian has been observed from the \\(k\\)-th viewing direction and initialized with training viewpoints \\(\\mathcal{V}^T\\). At each step \\(t\\), we update the map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.885, + 0.907, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\mathcal {M} _ {t} [ j, k ] = 1, \\quad \\forall j \\in J, \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27045" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.168 + ], + "angle": 0, + "content": "where \\(J\\) denotes the set of indices of 3D Gaussians visible from the candidate viewpoint and \\(k\\) indicates the index assigned to the discretized view direction of the candidate viewpoint. This update is performed based on the previous map \\(\\mathcal{M}_{t-1}\\), which can be described as" + }, + { + "type": "equation", + "bbox": [ + 0.196, + 0.181, + 0.483, + 0.203 + ], + "angle": 0, + "content": "\\[\nI G _ {t} = \\sum \\left(\\mathcal {M} _ {t} - \\mathcal {M} _ {t - 1}\\right), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.484, + 0.322 + ], + "angle": 0, + "content": "where \\(IG_{t}\\) is the number of Gaussians that have either been newly observed or newly viewed from different directions compared to \\(\\mathcal{M}_{t - 1}\\). 3D Gaussians that are newly observed or seen from a different angle compared to \\(\\mathcal{M}_{t - 1}\\), ensuring that only newly gained information is counted. To avoid redundant overlaps among trajectories, \\(\\mathcal{M}\\) is shared across among all trajectories \\(Tr\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.346, + 0.485, + 0.528 + ], + "angle": 0, + "content": "Termination. The search process terminates under two conditions. First, if no valid candidate views can be expanded from the current viewpoint while the number of cameras in the trajectory is still below the predefined length \\( L \\), the current branch is pruned and the search resumes from the next-best candidate view in the priority queue. Second, if the trajectory reaches the predefined length \\( L \\), the search concludes. The resulting trajectory is treated as a set of pseudo-observations, and then begins the searching process for the next trajectory \\( Tr_{i+1} \\). This process is repeated until \\( N_{Tr} \\) trajectories are constructed. We provide visualizations of example results in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.54, + 0.245, + 0.557 + ], + "angle": 0, + "content": "3.4. Diffusion prior" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.837 + ], + "angle": 0, + "content": "Since renderings from virtual viewpoints often partially include well-reconstructed parts, it is desirable to leverage these visual cues. To this end, we train a video diffusion model to enhance degraded rendered images \\( I^{V} \\) to be realistic images \\( I^{G} \\). Our model is built on [21] but introduces several key modifications. First, unlike [21], which adopts a frame interpolation setup, we address extreme view extrapolation in scene exploration by performing next-frame prediction using only the observation closest to the virtual camera as the sole visual cue. Second, to mitigate this ill-posed task, we condition the model on the scene text description and the ground truth image \\( I^{T} \\) from the nearest training view to each pseudo-observation. These are injected into the cross-attention layers of the diffusion model. At last, we create a dataset specifically designed to model artifacts arising from extreme extrapolation. The details about the dataset and the training process for the diffusion model are provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.214, + 0.864 + ], + "angle": 0, + "content": "3.5. Finetuning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Previous work [21] observed that applying an image-based loss (e.g., L1 loss) to a rendered image from a virtual" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.241 + ], + "angle": 0, + "content": "camera and its pseudo ground truth often degrades well-reconstructed regions, as 3DGS is sensitive to minor inaccuracies from diffusion model. It mitigates this issue by using two confidence levels, where higher confidence values indicate fewer conflicts between the contents of training and virtual viewpoints, thereby guiding 3D Gaussians to receive stronger gradients. It adopts the image level confidence based on the distance between a virtual viewpoint and a training viewpoint and pixel level confidence based on the scale of 3D Gaussians." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.243, + 0.906, + 0.363 + ], + "angle": 0, + "content": "Higher image-level confidence is applied when virtual viewpoints far from training viewpoints since they have less conflicts between them. While effective, using distance alone often fails to account for pure camera rotations, as the distance remains unchanged even when the view frustum overlap between viewpoints decreases. As our generated trajectories involve both translations and rotations, this results in suboptimal supervision." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.364, + 0.906, + 0.485 + ], + "angle": 0, + "content": "To address this issue, we introduce a novel image-level confidence metric based on intersection-over-union (IoU) of Gaussians, denoted as \\( G \\)-IOU. Our key insight is that a high view frustum overlaps between a training viewpoint and a virtual viewpoint implies many co-visible 3D Gaussians, leading to more conflicts. Using the Gaussian rasterizer, we efficiently extract the indices of visible Gaussians from each view, and define their intersection as the co-visible set:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.492, + 0.907, + 0.513 + ], + "angle": 0, + "content": "\\[\n\\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right) = \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {V}\\right) \\cap \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.521, + 0.906, + 0.566 + ], + "angle": 0, + "content": "where \\(\\mathcal{G}_{vis}(\\cdot)\\) denotes the set of 3D Gaussians visible from a given viewpoint, and \\(\\nu_{ref}\\) refers to the training viewpoint nearest to \\(\\nu^V\\). Based on this, \\(G\\)-IOU is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.574, + 0.905, + 0.61 + ], + "angle": 0, + "content": "\\[\nG I O U (\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}) = \\frac {\\left| \\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V} , \\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}{\\left| \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.617, + 0.797, + 0.632 + ], + "angle": 0, + "content": "and the image-level confidence is given by" + }, + { + "type": "equation", + "bbox": [ + 0.61, + 0.64, + 0.905, + 0.658 + ], + "angle": 0, + "content": "\\[\nU _ {\\mathrm {i m g}} = 1 - G \\mathrm {I O U} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.906, + 0.789 + ], + "angle": 0, + "content": "For the pixel level confidence, rather than solely relying on the scale of 3D Gaussians which still produce artifacts depending on the viewpoint, we define it as the perceptual distance [46] between the rendered image \\( I^{V} \\) and the output of the diffusion model \\( I^{G} \\). Each pixel in the confidence map reflects its contribution to the overall perceptual difference. We upsample the pixel-aligned LPIPS map to the original image resolution and define pixel-level confidence as" + }, + { + "type": "equation", + "bbox": [ + 0.575, + 0.798, + 0.905, + 0.817 + ], + "angle": 0, + "content": "\\[\nU _ {\\text {p i x e l}} (x, y) = \\mathcal {U} (\\mathrm {L P I P S} \\left(I ^ {V}, I ^ {G}\\right)) (x, y), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.826, + 0.905, + 0.87 + ], + "angle": 0, + "content": "where \\( U_{\\mathrm{pixel}}(x,y) \\) denotes the confidence value at pixel \\( (x,y) \\) and \\( \\mathcal{U} \\) is an upsampling operation to the original image resolution." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.905, + 0.9 + ], + "angle": 0, + "content": "Combining the above confidence, the loss for training views is defined as" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "27046" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.115, + 0.09, + 0.455, + 0.106 + ], + "angle": 0, + "content": "Training viewpoints Evaluation viewpoints Evaluation viewpoints (2nd)" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.107, + 0.482, + 0.335 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.339, + 0.483, + 0.38 + ], + "angle": 0, + "content": "Figure 4. Viewpoint difference comparison with previous 3D reconstruction dataset. Training and evaluation viewpoints are visualized." + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.404, + 0.482, + 0.424 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V} ^ {T}} = \\left\\| I ^ {T} - I ^ {V} \\right\\| _ {1} + \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(I ^ {T}, I ^ {V}\\right), \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.429, + 0.373, + 0.443 + ], + "angle": 0, + "content": "and the loss for virtual views is defined as" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.452, + 0.484, + 0.484 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathcal {V} ^ {V}} = U _ {i m g} \\cdot \\left(\\left| \\left| I ^ {G} - I ^ {V} \\right| \\right| _ {1} \\odot U _ {p i x e l} + \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I ^ {G}, I ^ {V}\\right)\\right). \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.485, + 0.484, + 0.531 + ], + "angle": 0, + "content": "The \\(\\odot\\) denotes the pixel-wise multiplication and \\(\\mathcal{L}_{\\mathrm{D - SSIM}}\\) denotes the dissimilarity variant of the SSIM loss. This fine-tuning process is also illustrated in Fig. 2 (c)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.543, + 0.451, + 0.56 + ], + "angle": 0, + "content": "4. Evaluation dataset for scene exploration" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.568, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Since existing datasets remain inadequate for evaluating scene exploration, as the test cameras are typically positioned within or close to the training trajectories, resulting in interpolation scenarios or only minor viewpoint variations. Although the Nerfbusters dataset [40] introduces extreme novel viewpoints via two separate camera trajectories, its test views are placed far behind the training views. Consequently, the task aligns more with generative outpainting rather than complete scene reconstruction. We introduce two datasets tailored for this task, with additional details provided in the supplementary materials." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.227, + 0.759 + ], + "angle": 0, + "content": "4.1. WildExplore" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "To address the lack of an appropriate benchmark for scene exploration, we introduce WildExplore, a new dataset comprising four indoor and four outdoor scenes. As illustrated in Fig. 4, WildExplore is designed with two key properties: (1) it explicitly targets the exploration task by placing test viewpoints significantly distant from the training trajectories, ensuring large viewpoint transitions and broad scene coverage, and (2) the evaluation set is captured to ensure sufficient co-visible areas between training and test views," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.121 + ], + "angle": 0, + "content": "minimizing the need for generating arbitrary contents, in contrast to Nerfbusters." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.131, + 0.71, + 0.146 + ], + "angle": 0, + "content": "4.2. Curated Nerfbusters" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.154, + 0.907, + 0.275 + ], + "angle": 0, + "content": "We curate the Nerfbusters dataset to better align with scene exploration objectives. We selects some scenes where the test views are placed far behind the training views and swap the train/test split, while leaving the others unchanged. This adjustment transforms the task into \"moving forward and closely examining the scene\", which better reflects realistic exploration scenarios. Our curated version includes seven scenes with swapped splits and two in their original form." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.289, + 0.646, + 0.306 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.314, + 0.704, + 0.331 + ], + "angle": 0, + "content": "5.1. Experimental Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.336, + 0.907, + 0.518 + ], + "angle": 0, + "content": "Implementation details. We first train 3DGS using the given training viewpoints, following the original setup in [11]. The number of virtual trajectories \\(N\\) is set to 20 by default. The occupancy grid resolution \\(S\\) is set to 64. At each step, we select the top 3 candidate views based on computed information gain. We set \\(D\\) as 32 for discretized view direction. For the video diffusion model, we adopt DynamiCrafter [42] as the backbone. After generating the pseudo observation set, we fine-tune 3D Gaussians for 15K iterations, applying densification until 9K, with the same scheduling as 3DGS. Further details will be discussed in the supplements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.537, + 0.909, + 0.734 + ], + "angle": 0, + "content": "Baselines. We evaluate our method in comparison to 3DGS-based methods, including 3DGS [11], 3DGS + Depth, DNGaussians [15], and ViewExtrapolator [19]. 3DGS + Depth applies additional depth regularization using monocular depth estimation [44], following the implementation from the official 3DGS repository [11]. DNGaussians [15] incorporates depth priors during partial optimization of the 3D Gaussian parameters. ViewExtrapolator leverages video diffusion but heuristically selects a single test camera for generating additional viewpoints due to the absence of a dedicated sampling strategy. In our experiments, we follow their protocol and choose the test camera farthest from the training set to encourage broader viewpoint coverage." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.608, + 0.758 + ], + "angle": 0, + "content": "5.2. Results" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Quantitative comparison. Table 1 reports the quantitative results on the exploration task. To assess reconstruction quality, we measure PSNR, SSIM, and LPIPS [46]. The first row reports the result from 3DGS trained on both the training and test splits, serving as an upper bound. Since our method effectively removes artifacts and fills missing regions, as observed in the qualitative analysis, our model outperforms across all metrics, particularly in PSNR and SSIM." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27047" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.151, + 0.097, + 0.176, + 0.109 + ], + "angle": 0, + "content": "GT" + }, + { + "type": "image_caption", + "bbox": [ + 0.27, + 0.097, + 0.314, + 0.11 + ], + "angle": 0, + "content": "3DGS" + }, + { + "type": "image_caption", + "bbox": [ + 0.382, + 0.097, + 0.477, + 0.111 + ], + "angle": 0, + "content": "3DGS+Depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.097, + 0.608, + 0.11 + ], + "angle": 0, + "content": "DNGaussian" + }, + { + "type": "image_caption", + "bbox": [ + 0.636, + 0.097, + 0.759, + 0.11 + ], + "angle": 0, + "content": "ViewExtrapolator" + }, + { + "type": "image_caption", + "bbox": [ + 0.813, + 0.097, + 0.848, + 0.11 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.095, + 0.112, + 0.904, + 0.624 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.295, + 0.628, + 0.7, + 0.642 + ], + "angle": 0, + "content": "Figure 5. Qualitative comparison on the curated Nerfbusters dataset." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.894 + ], + "angle": 0, + "content": "Qualitative comparison. Fig. 5 and Fig. 6 show qualitative comparisons among our method and baseline methods. 3DGS [11] suffers from artifacts, and its variants with depth regularization also meet the same problem, as they lack the capability to fill missing information. Although ViewExtrapolator [19] utilizes a video diffusion prior and thereby shows competitive performance compared to 3DGS variants, it underperforms in challenging scenarios. Its heuristic camera sampling struggles to maximize information gain and results in limited viewpoint coverage, leading to lower visual quality when rendering from distant viewpoints. Its training-free diffusion prior also acts as a bottleneck. In contrast, our method fills missing regions and removes artifacts more effectively, producing images that align closely with the ground truth. See the supplementary materials for additional results." + }, + { + "type": "table", + "bbox": [ + 0.518, + 0.651, + 0.905, + 0.757 + ], + "angle": 0, + "content": "
MethodWildExploreCurated Nerfbusters
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS Pseudo GT30.370.9100.14429.340.9110.128
3DGS15.810.4870.41414.510.4170.454
3DGS + Depth15.780.4830.41315.000.4350.436
DNGaussians15.650.5040.45014.010.4420.525
ViewExtrapolator16.020.5100.44214.980.4500.455
Ours16.960.5340.41316.290.4790.434
" + }, + { + "type": "table_caption", + "bbox": [ + 0.608, + 0.763, + 0.812, + 0.777 + ], + "angle": 0, + "content": "Table 1. Quantitative comparison." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.801, + 0.664, + 0.817 + ], + "angle": 0, + "content": "5.3. Ablation study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Information gain. We validate the effectiveness of our information gain design by comparing it with alternative strategies, as shown in Table 2. One alternative samples the 3D space uniformly and evaluate view coverage based on the occupied vertices of an occupancy grid, as proposed in" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "27048" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.094, + 0.907, + 0.334 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.315, + 0.336, + 0.68, + 0.351 + ], + "angle": 0, + "content": "Figure 6. Qualitative comparison on the WildExplore dataset." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.369, + 0.48, + 0.462 + ], + "angle": 0, + "content": "
Virtual View SamplingCurated Nerfbusters
Info. GainPSNR↑SSIM↑LPIPS↓
Grid view coverage [22]15.820.46750.4451
FisherRF [9]15.770.45930.4394
GS view coverage (Ours)16.290.47900.4336
" + }, + { + "type": "table_caption", + "bbox": [ + 0.138, + 0.468, + 0.435, + 0.482 + ], + "angle": 0, + "content": "Table 2. Comparisons on information gain design." + }, + { + "type": "table", + "bbox": [ + 0.134, + 0.492, + 0.44, + 0.569 + ], + "angle": 0, + "content": "
Method (K = 3)Garbage
PSNR↑SSIM↑LPIPS↓
3DGS14.420.4800.526
Ours with Bottom-K14.220.4750.521
Ours with Top-K15.450.5260.508
" + }, + { + "type": "table_caption", + "bbox": [ + 0.098, + 0.574, + 0.477, + 0.588 + ], + "angle": 0, + "content": "Table 3. Ablation study on information gain. TopK vs BottomK" + }, + { + "type": "table", + "bbox": [ + 0.104, + 0.603, + 0.469, + 0.739 + ], + "angle": 0, + "content": "
FinetuningCurated Nerfbusters
Image levelPixel levelPSNR↑SSIM↑LPIPS↓
Distance [21]-15.000.4270.443
-Scale [21]16.180.4760.442
DistanceScale15.270.4320.440
--16.300.4760.447
-pixelLPIPS16.230.4780.441
GS-IOU-16.100.4730.439
GS-IOUpixelLPIPS16.290.4790.434
" + }, + { + "type": "table_caption", + "bbox": [ + 0.143, + 0.743, + 0.429, + 0.757 + ], + "angle": 0, + "content": "Table 4. Ablations study on finetuning methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[22]. Another follows typical active-learning strategies and select viewpoints with the highest uncertainty. We adopt the method proposed in [9] for comparison. As presented in the first and second rows, both show suboptimal results. Grid-based approach often fails to maximize information gain, as it includes the gain from free space, resulting in redundant viewpoint selections. For uncertainty-based method, we observe that it is sensitive to 3D Gaussians with large scales," + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.371, + 0.907, + 0.403 + ], + "angle": 0, + "content": "tending to chase them rather than addressing artifacts from viewpoint changes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.907, + 0.496 + ], + "angle": 0, + "content": "Search strategy. We ablate our best first search strategy to validate its effectiveness. As shown in Table 3, the strategy selecting bottom K makes virtual cameras tend to stay near the trajectory of training viewpoints rather than deviating from it, leading to no performance gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.513, + 0.909, + 0.667 + ], + "angle": 0, + "content": "Finetuning methods. We validate our finetuning approach by benchmarking it against the methods proposed by [21], as shown in Table 4. The model without the image-level confidence produces worse results in terms of LPIPS, as generated images whose viewpoints closely match the training set negatively impact already well-reconstructed regions. Ignoring camera rotation further degrades performance, as observed in the second and fifth rows. Although both our image-level confidence map and scale based one are effective, our method leads to slight better performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.678, + 0.634, + 0.694 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.907, + 0.794 + ], + "angle": 0, + "content": "This paper presents a framework that enables high-quality rendering of the novel view in arbitrary viewpoints. Our method incorporates the video diffusion priors and the real-time rendering capability of 3DGS. It introduces a sampling strategy to effectively sample trajectories of virtual cameras for maximizing information gain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.794, + 0.909, + 0.9 + ], + "angle": 0, + "content": "Limitations and future works. While our method incorporates video diffusion priors to fill the missing information, we expect that employing more strong diffusion models which can faithfully generate high frequency details may improve the overall performance. In addition, extending the scene bounding box to cover a large scale scene would be an interesting avenue for the future work." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27049" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.114, + 0.486, + 0.201 + ], + "angle": 0, + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220113, No. RS-2024-00457882), Artificial Intelligence Graduate School Program, Yonsei University, under Grant RS-2020-11201361." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.226, + 0.188, + 0.243 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.251, + 0.486, + 0.32 + ], + "angle": 0, + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.322, + 0.484, + 0.363 + ], + "angle": 0, + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.365, + 0.484, + 0.434 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.436, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[4] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.492, + 0.484, + 0.533 + ], + "angle": 0, + "content": "[5] Jeong gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.535, + 0.484, + 0.577 + ], + "angle": 0, + "content": "[6] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.578, + 0.484, + 0.647 + ], + "angle": 0, + "content": "[7] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.649, + 0.484, + 0.716 + ], + "angle": 0, + "content": "[8] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.718, + 0.484, + 0.773 + ], + "angle": 0, + "content": "[9] Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information. arXiv preprint arXiv:2311.17874, 2023. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.775, + 0.484, + 0.844 + ], + "angle": 0, + "content": "[10] Liren Jin, Xieyuanli Chen, Julius Rückin, and Marija Popovic. Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11305-11312, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.484, + 0.9 + ], + "angle": 0, + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.251, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36:1-13, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.905, + 0.189 + ], + "angle": 0, + "content": "[13] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J. Davison. Eschernet: A generative model for scalable view synthesis, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Georgios Kopanas and George Drettakis. Improving NeRF Quality by Progressive Camera Placement for Free-Viewpoint Navigation. In Vision, Modeling, and Visualization. The Eurographics Association, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.3 + ], + "angle": 0, + "content": "[15] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.302, + 0.905, + 0.397 + ], + "angle": 0, + "content": "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, Xuanmao Li, Xingpeng Sun, Rohan Ashok, Aniruddha Mukherjee, Hao Kang, Xiangrui Kong, Gang Hua, Tianyi Zhang, Bedrich Benes, and Aniket Bera. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.4, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[17] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[18] Guangyi Liu, Wen Jiang, Boshu Lei, Vivek Pandey, Kostas Daniilidis, and Nader Motee. Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.538 + ], + "angle": 0, + "content": "[19] Kunhao Liu, Ling Shao, and Shijian Lu. Novel view extrapolation with video diffusion priors, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.58 + ], + "angle": 0, + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.581, + 0.905, + 0.622 + ], + "angle": 0, + "content": "[21] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors, 2024. 2, 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.704 + ], + "angle": 0, + "content": "[22] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra S Chaplot. Autonerf: Training implicit scene representations with autonomous agents. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13442-13449. IEEE, 2024. 3, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.905, + 0.762 + ], + "angle": 0, + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.763, + 0.905, + 0.816 + ], + "angle": 0, + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[26] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "27050" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.177 + ], + "angle": 0, + "content": "[27] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Activenerf: Learning where to see with uncertainty estimation. In European Conference on Computer Vision, pages 230-246. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.207 + ], + "angle": 0, + "content": "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.263 + ], + "angle": 0, + "content": "[29] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.265, + 0.482, + 0.332 + ], + "angle": 0, + "content": "[30] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. Zeronvs: Zero-shot 360-degree view synthesis from a single image, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.482, + 0.403 + ], + "angle": 0, + "content": "[31] Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In 2021 International Conference on 3D Vision (3DV), pages 972-981. IEEE, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[32] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[33] Meng-Li Shih, Wei-Chiu Ma, Lorenzo Boyice, Aleksander Holynski, Forrester Cole, Brian Curless, and Janne Kontkanen. Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20385-20395, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.534, + 0.482, + 0.574 + ], + "angle": 0, + "content": "[34] Niko Sünderhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.617 + ], + "angle": 0, + "content": "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.619, + 0.482, + 0.687 + ], + "angle": 0, + "content": "[36] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.482, + 0.731 + ], + "angle": 0, + "content": "[37] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.733, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[38] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.482, + 0.842 + ], + "angle": 0, + "content": "[39] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs, 2023. 2, 6" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[41] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Leo Lebrat. Nerf director: Revisiting view selection in neural volume rendering, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.515, + 0.149, + 0.905, + 0.205 + ], + "angle": 0, + "content": "[42] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360deg sparseviewsynthesisusinggaussiansplatting," + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.303 + ], + "angle": 0, + "content": "[44] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[45] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.389 + ], + "angle": 0, + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.391, + 0.905, + 0.432 + ], + "angle": 0, + "content": "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images, 2018. 3" + }, + { + "type": "list", + "bbox": [ + 0.515, + 0.092, + 0.905, + 0.432 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "27051" + } + ] +] \ No newline at end of file diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_origin.pdf b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..4a73b18184cc0ffa6dbc642a477be46c56529b38 --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/3e64bd2c-ab83-411a-816b-90e85430a7c9_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:9eb498b7a6831421642ebe8030c17d543ba47389408bcf59bdc146ab11a13c22 +size 3442964 diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/full.md b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..6a062b09f77d1d88bfc4cd36c5d6340234d7c675 --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/full.md @@ -0,0 +1,318 @@ +# ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors + +Minsu Kim Subin Jeon In Cho Mijin Yoo Seon Joo Kim Yonsei University + +{minsul206, subinjeon, join, mijin6690, seonjookim}@yonsei.ac.kr + +![](images/59bf2f231a0fad30377aff1a1467886b814f42c7b60d4a388afbc126f45bfaee.jpg) + +![](images/ce604cabb669ad002c68b0c7cb7e2a28420e44df8a478225d17b0aa1378bb8dd.jpg) +(a) Exploration Results + +![](images/17b881f061e6c1beaaabe70dbc72fe4e9440d7d452729cc3228f4b4136582aff.jpg) + +![](images/f749aa8e21126ea38969b0f09c96cf28c0dd9c52d123a4ac8a4aad35199e4cab.jpg) +(b) Virtual Camera Sampling with Diffusion Prior + +![](images/ef91def92e136b83ce333261f0616179551751eb923b4b8a5521c957668a4e1e.jpg) +(c) Scene Reconstruction +Figure 1. (a) Explorable 3D scene reconstruction results. Our method renders photorealistic images from any arbitrary viewpoints. (b) Virtual cameras are sampled in regions with missing information, and pseudo-ground truth is generated using diffusion priors. (c) After optimizing 3D Gaussians with virtual viewpoint samples, our method enables real-time scene exploration. + +# Abstract + +Recent advances in novel view synthesis (NVS) have enabled real-time rendering with 3D Gaussian Splatting (3DGS). However, existing methods struggle with artifacts and missing regions when rendering from viewpoints that deviate from the training trajectory, limiting seamless scene exploration. To address this, we propose a 3DGS-based pipeline that generates additional training views to enhance reconstruction. We introduce an information-gain-driven virtual camera placement strategy to maximize scene coverage, followed by video diffusion priors to refine rendered results. Fine-tuning 3D Gaussians with these enhanced views significantly improves reconstruction quality. To evaluate our method, we present Wild-Explore, a benchmark designed for challenging scene exploration. Experiments + +demonstrate that our approach outperforms existing 3DGS-based methods, enabling high-quality, artifact-free rendering from arbitrary viewpoints. + +# 1. Introduction + +The field of novel view synthesis (NVS) has seen significant advancements, enabling photorealistic rendering of scenes from previously unseen viewpoints. Much of this progress is attributed to NeRF [24] and its variants [1, 25]. This advancement has been further accelerated by recent 3D Gaussian Splatting (3DGS) [11], which enables high-quality rendering in real-time. The real-time capability of 3DGS broadens the applications of NVS in domains such as AR, VR, and gaming, leading users to expect the experience of freely exploring reconstructed scenes. + +Unfortunately, such an experience is yet to be fully realized, as existing methods suffer from severe degradations in rendering quality when viewpoints deviate significantly from the input observations. These degradations include holes, oversized Gaussians, spiky surface artifacts, and spurious floaters, often arising from occlusions or limited viewpoint coverage (see Fig. 1 (a)). This limitation stems from missing information, since optimization-based approaches cannot synthesize contents beyond the observed data. + +We are interested in 3D scene reconstruction that can render more vivid and photorealistic images from arbitrary viewpoints, so that users can freely explore the reconstructed scenes. We term this task as "explorable scene reconstruction", which aims to reconstruct complete scenes within the target areas. To avoid generating arbitrary contents in regions far beyond the input observations—which would not reflect actual user experiences—we define the target areas as a bounding box defined by the input views. + +One straightforward solution is to take thousands of images to completely cover the target scenes. However, collecting such dense data is time-consuming and labor-intensive, making it impractical for many real-world applications. Thus, we aim to reconstruct complete 3D scenes from input images that only partially cover the target scenes, which better reflects practical capturing scenarios. + +In this paper, we introduce ExploreGS, a pipeline that enables explorable scene reconstruction using diffusion priors and 3DGS. ExploreGS leverages video diffusion priors to generate additional training views that complement missing information. These synthesized images at virtual viewpoints serve as pseudo ground truths to refine the initial, incomplete 3D Gaussians, thereby enabling complete, photorealistic rendering at arbitrary viewpoints. + +The key challenges of explorable scene reconstruction lie in determining the optimal placement of virtual viewpoints. Unlike previous works that assume evaluation trajectories are known [19, 33], virtual viewpoints in our pipeline need to be sampled to best complement missing information in poorly covered areas. We propose a novel view sampling strategy that strategically positions virtual cameras based on information gain. This approach identifies viewpoints that offer the most novel scene details from the initial 3DGS reconstruction as shown in Fig. 1 (b). + +To provide pseudo observations, we generate realistic images along sampled viewpoints. We introduce video diffusion model which converts renderings with degradations into clean images. Although it is motivated by 3DGS-Enhancer [21], we modify the training objectives to fit scene exploration and prepare larger dataset for training. We then finetune 3DGS on both virtual viewpoints and training viewpoints with uncertainty guidance. + +To validate our pipeline, we introduce Wild-Explore, a dataset designed to evaluate challenging exploration scenar + +ios with diverse and distant viewpoints. We also assess our method on the Nerfbusters dataset [40], curated to better reflect scene exploration tasks. Experimental results show that our approach achieves the highest scores among 3DGS-based methods. Finally, an interactive exploration demo demonstrates that our method enables broader scene coverage with fewer artifacts compared to existing approaches. + +In summary, our contributions can be organized as follows: + +- We propose a pipeline for explorable 3D scene reconstruction, which incorporates the real-time rendering of 3DGS, video diffusion priors to complete missing regions, and an uncertainty-based finetuning strategy. +- We propose a novel virtual camera sampling strategy to identify viewpoints that can maximize additional information gain, thereby effectively reducing artifacts and missing regions. +- We introduce Wild-Explore, a new benchmark dataset designed to evaluate challenging exploration scenarios. + +# 2. Related work + +# 2.1. Novel view synthesis + +Novel view synthesis aims to generate images of a scene from unseen viewpoints. NeRF [24] and its following works [1, 2] have revolutionized novel view synthesis by representing 3D scenes as continuous volumetric functions, significantly improving rendering quality yet still suffering from slow rendering speed. Recently, 3D Gaussian Splatting [11] has enabled real-time rendering through an explicit Gaussian ellipsoid representation and its differentiable splatting pipeline. Despite these advances, all these methods produce degraded results when the number of images or their view coverage are limited. To address these shortcomings, several works [15, 26, 35, 43] incorporate geometric constraints and depth regularization, leading to less artifacts and improving reconstruction fidelity under sparse-view conditions. However, they remain insufficient for reconstructing explorable scenes, as they lack the ability to fill missing information in the training views. + +# 2.2. Evaluating novel view synthesis + +Previous novel view synthesis methods primarily focus on view interpolation, with LLFF [23], Blender synthetic dataset [24], TanksTemples [12] and MipNeRF360 [1] being widely used benchmarks. These datasets place test viewpoints between training viewpoints and thus are not suitable for evaluating performance under large viewpoint changes. To address this issue, Nerfbusters [40] introduces a new dataset with large viewpoint variations by capturing scenes from separate camera paths. Nevertheless, most of the test views in Nerfbusters lie behind the training views, which are largely unconstrained and primarily require gen- + +erating arbitrary contents rather than completing the underreconstructed target scenes. + +# 2.3. Diffusion prior for 3D reconstruction + +Since optimization-based NVS methods cannot generate novel contents, recent works [20, 28, 30, 37, 38] have explored methods for exploiting generative diffusion priors. The vast majority of recent research [4-8, 13, 17, 21, 32, 36, 39] train multi-view diffusion models or fine-tune video diffusion models on large multi-view datasets [16, 29, 47] for generating 3D contents. Among these, 3DGS-Enhancer [21] fine-tunes a video diffusion model to enhance a sequence of 3D Gaussian splatting renderings to solve sparse-view reconstruction task. On the one hand, some works [19, 33] apply diffusion prior to address view extrapolation. ExtraNeRF [33] introduces scene-level finetuned image inpainting diffusion and depth-completion model. ViewExtrapolator [19] leverages SVD [3] in a training-free manner. While these methods are most closely related to our work, they assume known evaluation viewpoints and require running the diffusion model for every target view to render. Moreover, their diffusion models lack multiview consistency, leaving their ability of reconstructing explorable scenes unclear. + +# 2.4. Next-view selection for data acquisition + +A number of works studied methods to select the next viewpoint to acquire data when the number of input viewpoints are limited. Early works focus on active learning or information-based strategies for efficient data acquisition [9, 14, 27, 31, 34, 41]. In particular, FisherRF [9] leverages the Fisher information matrix to quantify the information gain or uncertainty of a given viewpoint, serving as a strong baseline for next-best-view selection. However, these methods primarily target single-object or forward-facing scenes, while our method demonstrates superior performance in complex scenes. Recent works extend this line of works to autonomous exploration [10, 18, 22, 45]. Instead of capturing additional inputs, ExploreGS samples virtual viewpoints from initially reconstructed scenes and exploits generative priors to synthesize images from that viewpoints. + +# 3. Method + +# 3.1. Overview + +The overall pipeline of ExploreGS is illustrated in Fig. 2. Our pipeline follows three stages: (1) scene initialization, (2) pseudo observation generation with virtual camera sampling and a video diffusion prior, and (3) fine-tuning 3DGS using both the training and the generated observations sets. + +In the first stage, we begin by optimizing 3D Gaussians, denoted as $\mathcal{G}$ , to represent the 3D scene using the given training viewpoints $\mathcal{V}^T = \{V_i^T\}_{i=1}^{N_T}$ where $N_T$ denotes the + +number of training viewpoints. Here, each $V_{i}^{T}$ has known camera pose and ground truth image $I_{i}^{T}$ . Then, we determine the boundary of reconstructable scene based on the input observations and identify occupied regions for virtual viewpoints samplings. + +In the second stage, we sample virtual camera trajectories based on information gain of each candidate virtual viewpoint, as shown in Sec. 3.3. $N_{Tr}$ camera trajectories that maximizes the information gain are sampled, and each trajectory $Tr_{n}$ consists of $L$ progressively shifting virtual viewpoints: $Tr_{n} = \{V_{l}^{G}\}_{l=1}^{L}$ . Along these virtual trajectories, we render images to obtain $I_{i}^{G}$ and generate pseudo-observations $I_{i}^{P}$ from $I_{i}^{G}$ by enhancing them using the generative prior of a video diffusion model. + +Finally, we reconstruct complete and explorable 3D scenes by fine-tuning the initially optimized 3D Gaussians using both the training and virtual viewpoints, as described in Sec. 3.5. To mitigate the impact of inconsistencies between the generated and the original images during fine-tuning, we propose two techniques: (1) image-level confidence score, and (2) pixel-level confidence map. + +# 3.2. Scene initialization + +Initial reconstruction. Our pipeline begins by optimizing the initial set of 3D Gaussians with the given training set within the 3DGS optimization framework [11]. Then, we estimate the target bounding box of the scene, which is a maximum boundary that can be reconstructed based on the given input images. As previously discussed, reconstructing content beyond this bounding box is highly challenging, as it lacks grounding in the input observations and increasingly resembles unconstrained content generation, which is beyond the scope of this work. This bounding box is computed by using the mesh extracted from the initial 3D Gaussians and the given cameras. + +Rasterization-based occupancy estimation Since renderings from cameras inside occupied regions (e.g. behind walls or inside solid objects such as an apple) offer redundant images, not relevant to the scene, they should be located in free space. To this end, we introduce a simple rasterization-based algorithm to construct the occupancy grid $\mathcal{O} \in \mathbb{R}^{\mathbb{S} \times \mathbb{S} \times \mathbb{S}}$ . We estimate the visibility of each 3D Gaussians using the Gaussian rasterizer by computing transmittance values from the training views. Among these, we average the top-3 visibility scores to obtain the final visibility estimate. A vertex is classified as occupied if the estimated visibility is below threshold $\tau$ (we use 0.5); otherwise, it is considered free space. + +# 3.3. Virtual view sampling + +After initializing the target scene, our method utilizes video diffusion priors to supplement the missing information from + +![](images/988f300a2f13c7b1e2d823127a654433cb37f4b76594f644c8b4b6638a5fa0e5.jpg) +(a) 3DGS training +(b) Virtual camera sampling & diffusion process +(c) Confidence-aware finetuning (Sec. 3.4) + +![](images/fc68156097f57e525983380453d7deda29ec136392e862142f2c57f3ae534d0b.jpg) + +![](images/e2d02d3077b486d4718336ed42690db8b567db02a106d70bbf79dd31a1e48b23.jpg) +Figure 2. Overview of the proposed framework for scene exploration. (a) The scene is initially optimized using 3DGS on the given training viewpoints. (b) Based on the optimized 3DGS and training viewpoints, we generate virtual camera trajectories, and enhance the rendered views using our diffusion-based enhancement model. (c) Finally, the scene is further optimized using both the original training viewpoints and the newly generated virtual viewpoints. +Figure 3. (a) Viewpoint candidates for virtual camera viewpoint generation. (b) Information gain of each viewpoint. Simplified 2D examples of both are presented for clarity. + +the input views and removes unwanted artifacts and missing regions. To effectively utilize diffusion priors, our sampling strategy selects virtual camera trajectories that can maximize the information gain of the scene. + +As illustrated in Fig. 2, our approach follows a best-first search strategy with depth-first characteristics to construct virtual trajectories. We begin by selecting $N_{Tr}$ training viewpoints at spatially uniform intervals as the initial states for each trajectory and gradually select the next viewpoint. At each step, the current viewpoint expands into $K$ candidate views. The one with the highest information gain becomes the next viewpoint in the trajectory, while the rest are stored for future consideration. Here, the current state refers to the most recently added camera in a trajectory. + +Candidate views. We define candidate views as the set of possible camera motion primitives, which determine the transitions from the current state. As shown in Fig. 3, our action space consists of 6 translational moves, 4 pure rotations of the camera, and 4 rotations where the camera orbits around the look-at point, totaling 14 actions. This choice ensures the virtual trajectory efficiently exploring the scene when compared to previous works just using going straight or turning around. + +A candidate view is removed if it (1) lies outside the bounding box or in an occupied region, or (2) is close to the 3D Gaussians before fine-tuning, since rendering can be severely disrupted when the camera is placed near the occupied. We utilize the target scene bounding box and the occupancy grid from Sec. 3.2 to achieve this. These constraints prevent the trajectory from diverging away from the reconstructed scene or producing redundant pseudo-observations. + +Information gain. We aim to observe 3D Gaussians densely from diverse viewpoints, as sparse or limited-angle observations often lead to inaccurate shape and color estimation. To this end, we measure information gain based on view coverage. We employ a 2D map $\mathcal{M} \in \mathbb{R}^{N_{\mathcal{G}} \times D}$ , where $N_{\mathcal{G}}$ is the total number of Gaussians and $D$ is the number of discretized viewing directions around 3D Gaussians. The map $\mathcal{M}$ stores binary values where each entry $\mathcal{M}[j,k] = 1$ indicates that the $j$ -th Gaussian has been observed from the $k$ -th viewing direction and initialized with training viewpoints $\mathcal{V}^T$ . At each step $t$ , we update the map as follows: + +$$ +\mathcal {M} _ {t} [ j, k ] = 1, \quad \forall j \in J, \tag {1} +$$ + +where $J$ denotes the set of indices of 3D Gaussians visible from the candidate viewpoint and $k$ indicates the index assigned to the discretized view direction of the candidate viewpoint. This update is performed based on the previous map $\mathcal{M}_{t-1}$ , which can be described as + +$$ +I G _ {t} = \sum \left(\mathcal {M} _ {t} - \mathcal {M} _ {t - 1}\right), \tag {2} +$$ + +where $IG_{t}$ is the number of Gaussians that have either been newly observed or newly viewed from different directions compared to $\mathcal{M}_{t - 1}$ . 3D Gaussians that are newly observed or seen from a different angle compared to $\mathcal{M}_{t - 1}$ , ensuring that only newly gained information is counted. To avoid redundant overlaps among trajectories, $\mathcal{M}$ is shared across among all trajectories $Tr$ . + +Termination. The search process terminates under two conditions. First, if no valid candidate views can be expanded from the current viewpoint while the number of cameras in the trajectory is still below the predefined length $L$ , the current branch is pruned and the search resumes from the next-best candidate view in the priority queue. Second, if the trajectory reaches the predefined length $L$ , the search concludes. The resulting trajectory is treated as a set of pseudo-observations, and then begins the searching process for the next trajectory $Tr_{i+1}$ . This process is repeated until $N_{Tr}$ trajectories are constructed. We provide visualizations of example results in the supplementary materials. + +# 3.4. Diffusion prior + +Since renderings from virtual viewpoints often partially include well-reconstructed parts, it is desirable to leverage these visual cues. To this end, we train a video diffusion model to enhance degraded rendered images $I^{V}$ to be realistic images $I^{G}$ . Our model is built on [21] but introduces several key modifications. First, unlike [21], which adopts a frame interpolation setup, we address extreme view extrapolation in scene exploration by performing next-frame prediction using only the observation closest to the virtual camera as the sole visual cue. Second, to mitigate this ill-posed task, we condition the model on the scene text description and the ground truth image $I^{T}$ from the nearest training view to each pseudo-observation. These are injected into the cross-attention layers of the diffusion model. At last, we create a dataset specifically designed to model artifacts arising from extreme extrapolation. The details about the dataset and the training process for the diffusion model are provided in the supplementary materials. + +# 3.5. Finetuning + +Previous work [21] observed that applying an image-based loss (e.g., L1 loss) to a rendered image from a virtual + +camera and its pseudo ground truth often degrades well-reconstructed regions, as 3DGS is sensitive to minor inaccuracies from diffusion model. It mitigates this issue by using two confidence levels, where higher confidence values indicate fewer conflicts between the contents of training and virtual viewpoints, thereby guiding 3D Gaussians to receive stronger gradients. It adopts the image level confidence based on the distance between a virtual viewpoint and a training viewpoint and pixel level confidence based on the scale of 3D Gaussians. + +Higher image-level confidence is applied when virtual viewpoints far from training viewpoints since they have less conflicts between them. While effective, using distance alone often fails to account for pure camera rotations, as the distance remains unchanged even when the view frustum overlap between viewpoints decreases. As our generated trajectories involve both translations and rotations, this results in suboptimal supervision. + +To address this issue, we introduce a novel image-level confidence metric based on intersection-over-union (IoU) of Gaussians, denoted as $G$ -IOU. Our key insight is that a high view frustum overlaps between a training viewpoint and a virtual viewpoint implies many co-visible 3D Gaussians, leading to more conflicts. Using the Gaussian rasterizer, we efficiently extract the indices of visible Gaussians from each view, and define their intersection as the co-visible set: + +$$ +\mathcal {G} _ {\text {c o v i s}} \left(\mathcal {V} ^ {V}, \mathcal {V} ^ {\text {r e f}}\right) = \mathcal {G} _ {\text {v i s}} \left(\mathcal {V} ^ {V}\right) \cap \mathcal {G} _ {\text {v i s}} \left(\mathcal {V} ^ {\text {r e f}}\right), \tag {3} +$$ + +where $\mathcal{G}_{vis}(\cdot)$ denotes the set of 3D Gaussians visible from a given viewpoint, and $\nu_{ref}$ refers to the training viewpoint nearest to $\nu^V$ . Based on this, $G$ -IOU is defined as + +$$ +G I O U (\mathcal {V} ^ {V}, \mathcal {V} ^ {\text {r e f}}) = \frac {\left| \mathcal {G} _ {\text {c o v i s}} \left(\mathcal {V} ^ {V} , \mathcal {V} ^ {\text {r e f}}\right) \right|}{\left| \mathcal {G} _ {\text {v i s}} \left(\mathcal {V} ^ {\text {r e f}}\right) \right|}, \tag {4} +$$ + +and the image-level confidence is given by + +$$ +U _ {\mathrm {i m g}} = 1 - G \mathrm {I O U} \left(\mathcal {V} ^ {V}, \mathcal {V} ^ {\text {r e f}}\right). \tag {5} +$$ + +For the pixel level confidence, rather than solely relying on the scale of 3D Gaussians which still produce artifacts depending on the viewpoint, we define it as the perceptual distance [46] between the rendered image $I^{V}$ and the output of the diffusion model $I^{G}$ . Each pixel in the confidence map reflects its contribution to the overall perceptual difference. We upsample the pixel-aligned LPIPS map to the original image resolution and define pixel-level confidence as + +$$ +U _ {\text {p i x e l}} (x, y) = \mathcal {U} (\mathrm {L P I P S} \left(I ^ {V}, I ^ {G}\right)) (x, y), \tag {6} +$$ + +where $U_{\mathrm{pixel}}(x,y)$ denotes the confidence value at pixel $(x,y)$ and $\mathcal{U}$ is an upsampling operation to the original image resolution. + +Combining the above confidence, the loss for training views is defined as + +![](images/441cbd61c03ccda638c57383f59633b4019a369426d83303a11ce7ff32936151.jpg) +Training viewpoints Evaluation viewpoints Evaluation viewpoints (2nd) +Figure 4. Viewpoint difference comparison with previous 3D reconstruction dataset. Training and evaluation viewpoints are visualized. + +$$ +\mathcal {L} _ {\mathcal {V} ^ {T}} = \left\| I ^ {T} - I ^ {V} \right\| _ {1} + \mathcal {L} _ {\mathrm {D - S S I M}} \left(I ^ {T}, I ^ {V}\right), \tag {7} +$$ + +and the loss for virtual views is defined as + +$$ +\mathcal {L} _ {\mathcal {V} ^ {V}} = U _ {i m g} \cdot \left(\left| \left| I ^ {G} - I ^ {V} \right| \right| _ {1} \odot U _ {p i x e l} + \mathcal {L} _ {\mathrm {D} - \mathrm {S S I M}} \left(I ^ {G}, I ^ {V}\right)\right). \tag {8} +$$ + +The $\odot$ denotes the pixel-wise multiplication and $\mathcal{L}_{\mathrm{D - SSIM}}$ denotes the dissimilarity variant of the SSIM loss. This fine-tuning process is also illustrated in Fig. 2 (c). + +# 4. Evaluation dataset for scene exploration + +Since existing datasets remain inadequate for evaluating scene exploration, as the test cameras are typically positioned within or close to the training trajectories, resulting in interpolation scenarios or only minor viewpoint variations. Although the Nerfbusters dataset [40] introduces extreme novel viewpoints via two separate camera trajectories, its test views are placed far behind the training views. Consequently, the task aligns more with generative outpainting rather than complete scene reconstruction. We introduce two datasets tailored for this task, with additional details provided in the supplementary materials. + +# 4.1. WildExplore + +To address the lack of an appropriate benchmark for scene exploration, we introduce WildExplore, a new dataset comprising four indoor and four outdoor scenes. As illustrated in Fig. 4, WildExplore is designed with two key properties: (1) it explicitly targets the exploration task by placing test viewpoints significantly distant from the training trajectories, ensuring large viewpoint transitions and broad scene coverage, and (2) the evaluation set is captured to ensure sufficient co-visible areas between training and test views, + +minimizing the need for generating arbitrary contents, in contrast to Nerfbusters. + +# 4.2. Curated Nerfbusters + +We curate the Nerfbusters dataset to better align with scene exploration objectives. We selects some scenes where the test views are placed far behind the training views and swap the train/test split, while leaving the others unchanged. This adjustment transforms the task into "moving forward and closely examining the scene", which better reflects realistic exploration scenarios. Our curated version includes seven scenes with swapped splits and two in their original form. + +# 5. Experiments + +# 5.1. Experimental Setup + +Implementation details. We first train 3DGS using the given training viewpoints, following the original setup in [11]. The number of virtual trajectories $N$ is set to 20 by default. The occupancy grid resolution $S$ is set to 64. At each step, we select the top 3 candidate views based on computed information gain. We set $D$ as 32 for discretized view direction. For the video diffusion model, we adopt DynamiCrafter [42] as the backbone. After generating the pseudo observation set, we fine-tune 3D Gaussians for 15K iterations, applying densification until 9K, with the same scheduling as 3DGS. Further details will be discussed in the supplements. + +Baselines. We evaluate our method in comparison to 3DGS-based methods, including 3DGS [11], 3DGS + Depth, DNGaussians [15], and ViewExtrapolator [19]. 3DGS + Depth applies additional depth regularization using monocular depth estimation [44], following the implementation from the official 3DGS repository [11]. DNGaussians [15] incorporates depth priors during partial optimization of the 3D Gaussian parameters. ViewExtrapolator leverages video diffusion but heuristically selects a single test camera for generating additional viewpoints due to the absence of a dedicated sampling strategy. In our experiments, we follow their protocol and choose the test camera farthest from the training set to encourage broader viewpoint coverage. + +# 5.2. Results + +Quantitative comparison. Table 1 reports the quantitative results on the exploration task. To assess reconstruction quality, we measure PSNR, SSIM, and LPIPS [46]. The first row reports the result from 3DGS trained on both the training and test splits, serving as an upper bound. Since our method effectively removes artifacts and fills missing regions, as observed in the qualitative analysis, our model outperforms across all metrics, particularly in PSNR and SSIM. + +![](images/8b4e5d921fcfe5dfb336ead3abf129ae68585315f5c66be4ae5fe963f8748428.jpg) +GT +3DGS +3DGS+Depth +DNGaussian +ViewExtrapolator +Ours +Figure 5. Qualitative comparison on the curated Nerfbusters dataset. + +Qualitative comparison. Fig. 5 and Fig. 6 show qualitative comparisons among our method and baseline methods. 3DGS [11] suffers from artifacts, and its variants with depth regularization also meet the same problem, as they lack the capability to fill missing information. Although ViewExtrapolator [19] utilizes a video diffusion prior and thereby shows competitive performance compared to 3DGS variants, it underperforms in challenging scenarios. Its heuristic camera sampling struggles to maximize information gain and results in limited viewpoint coverage, leading to lower visual quality when rendering from distant viewpoints. Its training-free diffusion prior also acts as a bottleneck. In contrast, our method fills missing regions and removes artifacts more effectively, producing images that align closely with the ground truth. See the supplementary materials for additional results. + +
MethodWildExploreCurated Nerfbusters
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS Pseudo GT30.370.9100.14429.340.9110.128
3DGS15.810.4870.41414.510.4170.454
3DGS + Depth15.780.4830.41315.000.4350.436
DNGaussians15.650.5040.45014.010.4420.525
ViewExtrapolator16.020.5100.44214.980.4500.455
Ours16.960.5340.41316.290.4790.434
+ +Table 1. Quantitative comparison. + +# 5.3. Ablation study + +Information gain. We validate the effectiveness of our information gain design by comparing it with alternative strategies, as shown in Table 2. One alternative samples the 3D space uniformly and evaluate view coverage based on the occupied vertices of an occupancy grid, as proposed in + +![](images/87c4858e83b335710e494b0d46297f3a9ccb3428cc34b19b364781e2e137c723.jpg) +Figure 6. Qualitative comparison on the WildExplore dataset. + +
Virtual View SamplingCurated Nerfbusters
Info. GainPSNR↑SSIM↑LPIPS↓
Grid view coverage [22]15.820.46750.4451
FisherRF [9]15.770.45930.4394
GS view coverage (Ours)16.290.47900.4336
+ +Table 2. Comparisons on information gain design. + +
Method (K = 3)Garbage
PSNR↑SSIM↑LPIPS↓
3DGS14.420.4800.526
Ours with Bottom-K14.220.4750.521
Ours with Top-K15.450.5260.508
+ +Table 3. Ablation study on information gain. TopK vs BottomK + +
FinetuningCurated Nerfbusters
Image levelPixel levelPSNR↑SSIM↑LPIPS↓
Distance [21]-15.000.4270.443
-Scale [21]16.180.4760.442
DistanceScale15.270.4320.440
--16.300.4760.447
-pixelLPIPS16.230.4780.441
GS-IOU-16.100.4730.439
GS-IOUpixelLPIPS16.290.4790.434
+ +Table 4. Ablations study on finetuning methods. + +[22]. Another follows typical active-learning strategies and select viewpoints with the highest uncertainty. We adopt the method proposed in [9] for comparison. As presented in the first and second rows, both show suboptimal results. Grid-based approach often fails to maximize information gain, as it includes the gain from free space, resulting in redundant viewpoint selections. For uncertainty-based method, we observe that it is sensitive to 3D Gaussians with large scales, + +tending to chase them rather than addressing artifacts from viewpoint changes. + +Search strategy. We ablate our best first search strategy to validate its effectiveness. As shown in Table 3, the strategy selecting bottom K makes virtual cameras tend to stay near the trajectory of training viewpoints rather than deviating from it, leading to no performance gain. + +Finetuning methods. We validate our finetuning approach by benchmarking it against the methods proposed by [21], as shown in Table 4. The model without the image-level confidence produces worse results in terms of LPIPS, as generated images whose viewpoints closely match the training set negatively impact already well-reconstructed regions. Ignoring camera rotation further degrades performance, as observed in the second and fifth rows. Although both our image-level confidence map and scale based one are effective, our method leads to slight better performance. + +# 6. Conclusion + +This paper presents a framework that enables high-quality rendering of the novel view in arbitrary viewpoints. Our method incorporates the video diffusion priors and the real-time rendering capability of 3DGS. It introduces a sampling strategy to effectively sample trajectories of virtual cameras for maximizing information gain. + +Limitations and future works. While our method incorporates video diffusion priors to fill the missing information, we expect that employing more strong diffusion models which can faithfully generate high frequency details may improve the overall performance. In addition, extending the scene bounding box to cover a large scale scene would be an interesting avenue for the future work. + +# Acknowledgments + +This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220113, No. RS-2024-00457882), Artificial Intelligence Graduate School Program, Yonsei University, under Grant RS-2020-11201361. + +# References + +[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 1, 2 +[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields, 2023. 2 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 3 +[4] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models, 2024. 3 +[5] Jeong gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models, 2023. +[6] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation, 2024. +[7] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024. +[8] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2024. 3 +[9] Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information. arXiv preprint arXiv:2311.17874, 2023. 3, 8 +[10] Liren Jin, Xieyuanli Chen, Julius Rückin, and Marija Popovic. Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11305-11312, 2023. 3 +[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 6, 7 + +[12] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36:1-13, 2017. 2 +[13] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J. Davison. Eschernet: A generative model for scalable view synthesis, 2024. 3 +[14] Georgios Kopanas and George Drettakis. Improving NeRF Quality by Progressive Camera Placement for Free-Viewpoint Navigation. In Vision, Modeling, and Visualization. The Eurographics Association, 2023. 3 +[15] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization, 2024. 2, 6 +[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, Xuanmao Li, Xingpeng Sun, Rohan Ashok, Aniruddha Mukherjee, Hao Kang, Xiangrui Kong, Gang Hua, Tianyi Zhang, Bedrich Benes, and Aniket Bera. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision, 2023. 3 +[17] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model, 2024. 3 +[18] Guangyi Liu, Wen Jiang, Boshu Lei, Vivek Pandey, Kostas Daniilidis, and Nader Motee. Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf, 2025. 3 +[19] Kunhao Liu, Ling Shao, and Shijian Lu. Novel view extrapolation with video diffusion priors, 2024. 2, 3, 6, 7 +[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 3 +[21] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors, 2024. 2, 3, 5, 8 +[22] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra S Chaplot. Autonerf: Training implicit scene representations with autonomous agents. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13442-13449. IEEE, 2024. 3, 8 +[23] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines, 2019. 2 +[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2 +[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1 +[26] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan. + +Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs, 2021. 2 +[27] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Activenerf: Learning where to see with uncertainty estimation. In European Conference on Computer Vision, pages 230-246. Springer, 2022. 3 +[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion, 2022. 3 +[29] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction, 2021. 3 +[30] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. Zeronvs: Zero-shot 360-degree view synthesis from a single image, 2024. 3 +[31] Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In 2021 International Conference on 3D Vision (3DV), pages 972-981. IEEE, 2021. 3 +[32] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2024. 3 +[33] Meng-Li Shih, Wei-Chiu Ma, Lorenzo Boyice, Aleksander Holynski, Forrester Cole, Brian Curless, and Janne Kontkanen. Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20385-20395, 2024. 2, 3 +[34] Niko Sünderhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields, 2022. 3 +[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses, 2023. 2 +[36] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 3 +[37] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 3 +[38] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 3 +[39] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation, 2024. 3 +[40] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs, 2023. 2, 6 + +[41] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Leo Lebrat. Nerf director: Revisiting view selection in neural volume rendering, 2024. 3 +[42] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 6 +[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360deg sparseviewsynthesisusinggaussiansplatting, +[44] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 6 +[45] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning, 2022. 3 +[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5, 6 +[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images, 2018. 3 \ No newline at end of file diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/images.zip b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d232b4db4e8b30c253032d3187fc1e588babcd9 --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e2bebb6f8c2dddfce402621850a5682ac83bc3ca589ef5fa64724979cdc09b8d +size 868619 diff --git a/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/layout.json b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..4d3cf68aa00b2a70588327a35a00c04156811aec --- /dev/null +++ b/2025/ExploreGS_ Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors/layout.json @@ -0,0 +1,7946 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "spans": [ + { + "bbox": [ + 141, + 103, + 470, + 140 + ], + "type": "text", + "content": "ExploreGS: Explorable 3D Scene Reconstruction with Virtual Camera Samplings and Diffusion Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 145, + 161, + 465, + 192 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 161, + 465, + 192 + ], + "spans": [ + { + "bbox": [ + 145, + 161, + 465, + 192 + ], + "type": "text", + "content": "Minsu Kim Subin Jeon In Cho Mijin Yoo Seon Joo Kim Yonsei University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 130, + 194, + 476, + 206 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 130, + 194, + 476, + 206 + ], + "spans": [ + { + "bbox": [ + 130, + 194, + 476, + 206 + ], + "type": "text", + "content": "{minsul206, subinjeon, join, mijin6690, seonjookim}@yonsei.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 216, + 222, + 361 + ], + "blocks": [ + { + "bbox": [ + 59, + 216, + 222, + 361 + ], + "lines": [ + { + "bbox": [ + 59, + 216, + 222, + 361 + ], + "spans": [ + { + "bbox": [ + 59, + 216, + 222, + 361 + ], + "type": "image", + "image_path": "59bf2f231a0fad30377aff1a1467886b814f42c7b60d4a388afbc126f45bfaee.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 225, + 216, + 389, + 361 + ], + "blocks": [ + { + "bbox": [ + 225, + 216, + 389, + 361 + ], + "lines": [ + { + "bbox": [ + 225, + 216, + 389, + 361 + ], + "spans": [ + { + "bbox": [ + 225, + 216, + 389, + 361 + ], + "type": "image", + "image_path": "ce604cabb669ad002c68b0c7cb7e2a28420e44df8a478225d17b0aa1378bb8dd.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 262, + 361, + 349, + 372 + ], + "lines": [ + { + "bbox": [ + 262, + 361, + 349, + 372 + ], + "spans": [ + { + "bbox": [ + 262, + 361, + 349, + 372 + ], + "type": "text", + "content": "(a) Exploration Results" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 391, + 216, + 554, + 361 + ], + "blocks": [ + { + "bbox": [ + 391, + 216, + 554, + 361 + ], + "lines": [ + { + "bbox": [ + 391, + 216, + 554, + 361 + ], + "spans": [ + { + "bbox": [ + 391, + 216, + 554, + 361 + ], + "type": "image", + "image_path": "17b881f061e6c1beaaabe70dbc72fe4e9440d7d452729cc3228f4b4136582aff.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 61, + 373, + 293, + 446 + ], + "blocks": [ + { + "bbox": [ + 61, + 373, + 293, + 446 + ], + "lines": [ + { + "bbox": [ + 61, + 373, + 293, + 446 + ], + "spans": [ + { + "bbox": [ + 61, + 373, + 293, + 446 + ], + "type": "image", + "image_path": "f749aa8e21126ea38969b0f09c96cf28c0dd9c52d123a4ac8a4aad35199e4cab.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 105, + 449, + 287, + 460 + ], + "lines": [ + { + "bbox": [ + 105, + 449, + 287, + 460 + ], + "spans": [ + { + "bbox": [ + 105, + 449, + 287, + 460 + ], + "type": "text", + "content": "(b) Virtual Camera Sampling with Diffusion Prior" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 340, + 373, + 536, + 449 + ], + "blocks": [ + { + "bbox": [ + 340, + 373, + 536, + 449 + ], + "lines": [ + { + "bbox": [ + 340, + 373, + 536, + 449 + ], + "spans": [ + { + "bbox": [ + 340, + 373, + 536, + 449 + ], + "type": "image", + "image_path": "ef91def92e136b83ce333261f0616179551751eb923b4b8a5521c957668a4e1e.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 392, + 449, + 485, + 459 + ], + "lines": [ + { + "bbox": [ + 392, + 449, + 485, + 459 + ], + "spans": [ + { + "bbox": [ + 392, + 449, + 485, + 459 + ], + "type": "text", + "content": "(c) Scene Reconstruction" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "lines": [ + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 54, + 467, + 555, + 502 + ], + "type": "text", + "content": "Figure 1. (a) Explorable 3D scene reconstruction results. Our method renders photorealistic images from any arbitrary viewpoints. (b) Virtual cameras are sampled in regions with missing information, and pseudo-ground truth is generated using diffusion priors. (c) After optimizing 3D Gaussians with virtual viewpoint samples, our method enables real-time scene exploration." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "spans": [ + { + "bbox": [ + 152, + 510, + 200, + 522 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 546, + 297, + 714 + ], + "type": "text", + "content": "Recent advances in novel view synthesis (NVS) have enabled real-time rendering with 3D Gaussian Splatting (3DGS). However, existing methods struggle with artifacts and missing regions when rendering from viewpoints that deviate from the training trajectory, limiting seamless scene exploration. To address this, we propose a 3DGS-based pipeline that generates additional training views to enhance reconstruction. We introduce an information-gain-driven virtual camera placement strategy to maximize scene coverage, followed by video diffusion priors to refine rendered results. Fine-tuning 3D Gaussians with these enhanced views significantly improves reconstruction quality. To evaluate our method, we present Wild-Explore, a benchmark designed for challenging scene exploration. Experiments" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 511, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 555, + 548 + ], + "type": "text", + "content": "demonstrate that our approach outperforms existing 3DGS-based methods, enabling high-quality, artifact-free rendering from arbitrary viewpoints." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 314, + 573, + 395, + 585 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 573, + 395, + 585 + ], + "spans": [ + { + "bbox": [ + 314, + 573, + 395, + 585 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 556, + 715 + ], + "type": "text", + "content": "The field of novel view synthesis (NVS) has seen significant advancements, enabling photorealistic rendering of scenes from previously unseen viewpoints. Much of this progress is attributed to NeRF [24] and its variants [1, 25]. This advancement has been further accelerated by recent 3D Gaussian Splatting (3DGS) [11], which enables high-quality rendering in real-time. The real-time capability of 3DGS broadens the applications of NVS in domains such as AR, VR, and gaming, leading users to expect the experience of freely exploring reconstructed scenes." + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27042" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 179 + ], + "type": "text", + "content": "Unfortunately, such an experience is yet to be fully realized, as existing methods suffer from severe degradations in rendering quality when viewpoints deviate significantly from the input observations. These degradations include holes, oversized Gaussians, spiky surface artifacts, and spurious floaters, often arising from occlusions or limited viewpoint coverage (see Fig. 1 (a)). This limitation stems from missing information, since optimization-based approaches cannot synthesize contents beyond the observed data." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 181, + 294, + 289 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 181, + 294, + 289 + ], + "spans": [ + { + "bbox": [ + 55, + 181, + 294, + 289 + ], + "type": "text", + "content": "We are interested in 3D scene reconstruction that can render more vivid and photorealistic images from arbitrary viewpoints, so that users can freely explore the reconstructed scenes. We term this task as \"explorable scene reconstruction\", which aims to reconstruct complete scenes within the target areas. To avoid generating arbitrary contents in regions far beyond the input observations—which would not reflect actual user experiences—we define the target areas as a bounding box defined by the input views." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 290, + 294, + 373 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 290, + 294, + 373 + ], + "spans": [ + { + "bbox": [ + 55, + 290, + 294, + 373 + ], + "type": "text", + "content": "One straightforward solution is to take thousands of images to completely cover the target scenes. However, collecting such dense data is time-consuming and labor-intensive, making it impractical for many real-world applications. Thus, we aim to reconstruct complete 3D scenes from input images that only partially cover the target scenes, which better reflects practical capturing scenarios." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 375, + 294, + 471 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 375, + 294, + 471 + ], + "spans": [ + { + "bbox": [ + 55, + 375, + 294, + 471 + ], + "type": "text", + "content": "In this paper, we introduce ExploreGS, a pipeline that enables explorable scene reconstruction using diffusion priors and 3DGS. ExploreGS leverages video diffusion priors to generate additional training views that complement missing information. These synthesized images at virtual viewpoints serve as pseudo ground truths to refine the initial, incomplete 3D Gaussians, thereby enabling complete, photorealistic rendering at arbitrary viewpoints." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 472, + 294, + 590 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 294, + 590 + ], + "type": "text", + "content": "The key challenges of explorable scene reconstruction lie in determining the optimal placement of virtual viewpoints. Unlike previous works that assume evaluation trajectories are known [19, 33], virtual viewpoints in our pipeline need to be sampled to best complement missing information in poorly covered areas. We propose a novel view sampling strategy that strategically positions virtual cameras based on information gain. This approach identifies viewpoints that offer the most novel scene details from the initial 3DGS reconstruction as shown in Fig. 1 (b)." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 593, + 294, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 593, + 294, + 688 + ], + "spans": [ + { + "bbox": [ + 55, + 593, + 294, + 688 + ], + "type": "text", + "content": "To provide pseudo observations, we generate realistic images along sampled viewpoints. We introduce video diffusion model which converts renderings with degradations into clean images. Although it is motivated by 3DGS-Enhancer [21], we modify the training objectives to fit scene exploration and prepare larger dataset for training. We then finetune 3DGS on both virtual viewpoints and training viewpoints with uncertainty guidance." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 294, + 713 + ], + "type": "text", + "content": "To validate our pipeline, we introduce Wild-Explore, a dataset designed to evaluate challenging exploration scenar" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 156 + ], + "type": "text", + "content": "ios with diverse and distant viewpoints. We also assess our method on the Nerfbusters dataset [40], curated to better reflect scene exploration tasks. Experimental results show that our approach achieves the highest scores among 3DGS-based methods. Finally, an interactive exploration demo demonstrates that our method enables broader scene coverage with fewer artifacts compared to existing approaches." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 157, + 553, + 179 + ], + "type": "text", + "content": "In summary, our contributions can be organized as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 181, + 553, + 300 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 314, + 181, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 181, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 314, + 181, + 553, + 228 + ], + "type": "text", + "content": "- We propose a pipeline for explorable 3D scene reconstruction, which incorporates the real-time rendering of 3DGS, video diffusion priors to complete missing regions, and an uncertainty-based finetuning strategy." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 229, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 229, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 314, + 229, + 553, + 276 + ], + "type": "text", + "content": "- We propose a novel virtual camera sampling strategy to identify viewpoints that can maximize additional information gain, thereby effectively reducing artifacts and missing regions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 277, + 553, + 300 + ], + "type": "text", + "content": "- We introduce Wild-Explore, a new benchmark dataset designed to evaluate challenging exploration scenarios." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 313, + 398, + 325 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 313, + 398, + 325 + ], + "spans": [ + { + "bbox": [ + 314, + 313, + 398, + 325 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 333, + 432, + 345 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 333, + 432, + 345 + ], + "spans": [ + { + "bbox": [ + 314, + 333, + 432, + 345 + ], + "type": "text", + "content": "2.1. Novel view synthesis" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 351, + 555, + 555 + ], + "type": "text", + "content": "Novel view synthesis aims to generate images of a scene from unseen viewpoints. NeRF [24] and its following works [1, 2] have revolutionized novel view synthesis by representing 3D scenes as continuous volumetric functions, significantly improving rendering quality yet still suffering from slow rendering speed. Recently, 3D Gaussian Splatting [11] has enabled real-time rendering through an explicit Gaussian ellipsoid representation and its differentiable splatting pipeline. Despite these advances, all these methods produce degraded results when the number of images or their view coverage are limited. To address these shortcomings, several works [15, 26, 35, 43] incorporate geometric constraints and depth regularization, leading to less artifacts and improving reconstruction fidelity under sparse-view conditions. However, they remain insufficient for reconstructing explorable scenes, as they lack the ability to fill missing information in the training views." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 563, + 484, + 576 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 484, + 576 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 484, + 576 + ], + "type": "text", + "content": "2.2. Evaluating novel view synthesis" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Previous novel view synthesis methods primarily focus on view interpolation, with LLFF [23], Blender synthetic dataset [24], TanksTemples [12] and MipNeRF360 [1] being widely used benchmarks. These datasets place test viewpoints between training viewpoints and thus are not suitable for evaluating performance under large viewpoint changes. To address this issue, Nerfbusters [40] introduces a new dataset with large viewpoint variations by capturing scenes from separate camera paths. Nevertheless, most of the test views in Nerfbusters lie behind the training views, which are largely unconstrained and primarily require gen-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27043" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "erating arbitrary contents rather than completing the underreconstructed target scenes." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 104, + 253, + 118 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 104, + 253, + 118 + ], + "spans": [ + { + "bbox": [ + 55, + 104, + 253, + 118 + ], + "type": "text", + "content": "2.3. Diffusion prior for 3D reconstruction" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 122, + 296, + 362 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 122, + 296, + 362 + ], + "spans": [ + { + "bbox": [ + 55, + 122, + 296, + 362 + ], + "type": "text", + "content": "Since optimization-based NVS methods cannot generate novel contents, recent works [20, 28, 30, 37, 38] have explored methods for exploiting generative diffusion priors. The vast majority of recent research [4-8, 13, 17, 21, 32, 36, 39] train multi-view diffusion models or fine-tune video diffusion models on large multi-view datasets [16, 29, 47] for generating 3D contents. Among these, 3DGS-Enhancer [21] fine-tunes a video diffusion model to enhance a sequence of 3D Gaussian splatting renderings to solve sparse-view reconstruction task. On the one hand, some works [19, 33] apply diffusion prior to address view extrapolation. ExtraNeRF [33] introduces scene-level finetuned image inpainting diffusion and depth-completion model. ViewExtrapolator [19] leverages SVD [3] in a training-free manner. While these methods are most closely related to our work, they assume known evaluation viewpoints and require running the diffusion model for every target view to render. Moreover, their diffusion models lack multiview consistency, leaving their ability of reconstructing explorable scenes unclear." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 370, + 263, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 370, + 263, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 370, + 263, + 384 + ], + "type": "text", + "content": "2.4. Next-view selection for data acquisition" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 388, + 296, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 388, + 296, + 568 + ], + "spans": [ + { + "bbox": [ + 55, + 388, + 296, + 568 + ], + "type": "text", + "content": "A number of works studied methods to select the next viewpoint to acquire data when the number of input viewpoints are limited. Early works focus on active learning or information-based strategies for efficient data acquisition [9, 14, 27, 31, 34, 41]. In particular, FisherRF [9] leverages the Fisher information matrix to quantify the information gain or uncertainty of a given viewpoint, serving as a strong baseline for next-best-view selection. However, these methods primarily target single-object or forward-facing scenes, while our method demonstrates superior performance in complex scenes. Recent works extend this line of works to autonomous exploration [10, 18, 22, 45]. Instead of capturing additional inputs, ExploreGS samples virtual viewpoints from initially reconstructed scenes and exploits generative priors to synthesize images from that viewpoints." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 579, + 111, + 592 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 579, + 111, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 579, + 111, + 592 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 599, + 123, + 611 + ], + "type": "text", + "content": "3.1. Overview" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 295, + 677 + ], + "type": "text", + "content": "The overall pipeline of ExploreGS is illustrated in Fig. 2. Our pipeline follows three stages: (1) scene initialization, (2) pseudo observation generation with virtual camera sampling and a video diffusion prior, and (3) fine-tuning 3DGS using both the training and the generated observations sets." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "In the first stage, we begin by optimizing 3D Gaussians, denoted as " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{G}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": ", to represent the 3D scene using the given training viewpoints " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^T = \\{V_i^T\\}_{i=1}^{N_T}" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "inline_equation", + "content": "N_T" + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": " denotes the" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": "number of training viewpoints. Here, each " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "V_{i}^{T}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": " has known camera pose and ground truth image " + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "inline_equation", + "content": "I_{i}^{T}" + }, + { + "bbox": [ + 313, + 72, + 553, + 133 + ], + "type": "text", + "content": ". Then, we determine the boundary of reconstructable scene based on the input observations and identify occupied regions for virtual viewpoints samplings." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": "In the second stage, we sample virtual camera trajectories based on information gain of each candidate virtual viewpoint, as shown in Sec. 3.3. " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "N_{Tr}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " camera trajectories that maximizes the information gain are sampled, and each trajectory " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "Tr_{n}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " consists of " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " progressively shifting virtual viewpoints: " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "Tr_{n} = \\{V_{l}^{G}\\}_{l=1}^{L}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": ". Along these virtual trajectories, we render images to obtain " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "I_{i}^{G}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " and generate pseudo-observations " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "I_{i}^{P}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " from " + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "inline_equation", + "content": "I_{i}^{G}" + }, + { + "bbox": [ + 313, + 133, + 555, + 239 + ], + "type": "text", + "content": " by enhancing them using the generative prior of a video diffusion model." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 240, + 554, + 324 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 240, + 554, + 324 + ], + "spans": [ + { + "bbox": [ + 313, + 240, + 554, + 324 + ], + "type": "text", + "content": "Finally, we reconstruct complete and explorable 3D scenes by fine-tuning the initially optimized 3D Gaussians using both the training and virtual viewpoints, as described in Sec. 3.5. To mitigate the impact of inconsistencies between the generated and the original images during fine-tuning, we propose two techniques: (1) image-level confidence score, and (2) pixel-level confidence map." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 332, + 426, + 344 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 426, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 426, + 344 + ], + "type": "text", + "content": "3.2. Scene initialization" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 349, + 554, + 494 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 349, + 554, + 494 + ], + "spans": [ + { + "bbox": [ + 313, + 349, + 554, + 494 + ], + "type": "text", + "content": "Initial reconstruction. Our pipeline begins by optimizing the initial set of 3D Gaussians with the given training set within the 3DGS optimization framework [11]. Then, we estimate the target bounding box of the scene, which is a maximum boundary that can be reconstructed based on the given input images. As previously discussed, reconstructing content beyond this bounding box is highly challenging, as it lacks grounding in the input observations and increasingly resembles unconstrained content generation, which is beyond the scope of this work. This bounding box is computed by using the mesh extracted from the initial 3D Gaussians and the given cameras." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "text", + "content": "Rasterization-based occupancy estimation Since renderings from cameras inside occupied regions (e.g. behind walls or inside solid objects such as an apple) offer redundant images, not relevant to the scene, they should be located in free space. To this end, we introduce a simple rasterization-based algorithm to construct the occupancy grid " + }, + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{O} \\in \\mathbb{R}^{\\mathbb{S} \\times \\mathbb{S} \\times \\mathbb{S}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "text", + "content": ". We estimate the visibility of each 3D Gaussians using the Gaussian rasterizer by computing transmittance values from the training views. Among these, we average the top-3 visibility scores to obtain the final visibility estimate. A vertex is classified as occupied if the estimated visibility is below threshold " + }, + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 509, + 555, + 664 + ], + "type": "text", + "content": " (we use 0.5); otherwise, it is considered free space." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 441, + 685 + ], + "type": "text", + "content": "3.3. Virtual view sampling" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 715 + ], + "type": "text", + "content": "After initializing the target scene, our method utilizes video diffusion priors to supplement the missing information from" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27044" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 70, + 419, + 237 + ], + "blocks": [ + { + "bbox": [ + 60, + 70, + 419, + 237 + ], + "lines": [ + { + "bbox": [ + 60, + 70, + 419, + 237 + ], + "spans": [ + { + "bbox": [ + 60, + 70, + 419, + 237 + ], + "type": "image", + "image_path": "988f300a2f13c7b1e2d823127a654433cb37f4b76594f644c8b4b6638a5fa0e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 240, + 126, + 251 + ], + "lines": [ + { + "bbox": [ + 72, + 240, + 126, + 251 + ], + "spans": [ + { + "bbox": [ + 72, + 240, + 126, + 251 + ], + "type": "text", + "content": "(a) 3DGS training" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 209, + 241, + 350, + 251 + ], + "lines": [ + { + "bbox": [ + 209, + 241, + 350, + 251 + ], + "spans": [ + { + "bbox": [ + 209, + 241, + 350, + 251 + ], + "type": "text", + "content": "(b) Virtual camera sampling & diffusion process" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 423, + 240, + 549, + 250 + ], + "lines": [ + { + "bbox": [ + 423, + 240, + 549, + 250 + ], + "spans": [ + { + "bbox": [ + 423, + 240, + 549, + 250 + ], + "type": "text", + "content": "(c) Confidence-aware finetuning (Sec. 3.4)" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 423, + 77, + 539, + 233 + ], + "blocks": [ + { + "bbox": [ + 423, + 77, + 539, + 233 + ], + "lines": [ + { + "bbox": [ + 423, + 77, + 539, + 233 + ], + "spans": [ + { + "bbox": [ + 423, + 77, + 539, + 233 + ], + "type": "image", + "image_path": "fc68156097f57e525983380453d7deda29ec136392e862142f2c57f3ae534d0b.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 59, + 329, + 294, + 492 + ], + "blocks": [ + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "lines": [ + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 263, + 555, + 309 + ], + "type": "text", + "content": "Figure 2. Overview of the proposed framework for scene exploration. (a) The scene is initially optimized using 3DGS on the given training viewpoints. (b) Based on the optimized 3DGS and training viewpoints, we generate virtual camera trajectories, and enhance the rendered views using our diffusion-based enhancement model. (c) Finally, the scene is further optimized using both the original training viewpoints and the newly generated virtual viewpoints." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 329, + 294, + 492 + ], + "lines": [ + { + "bbox": [ + 59, + 329, + 294, + 492 + ], + "spans": [ + { + "bbox": [ + 59, + 329, + 294, + 492 + ], + "type": "image", + "image_path": "e2d02d3077b486d4718336ed42690db8b567db02a106d70bbf79dd31a1e48b23.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "lines": [ + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "spans": [ + { + "bbox": [ + 55, + 499, + 295, + 534 + ], + "type": "text", + "content": "Figure 3. (a) Viewpoint candidates for virtual camera viewpoint generation. (b) Information gain of each viewpoint. Simplified 2D examples of both are presented for clarity." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 592 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 592 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 592 + ], + "type": "text", + "content": "the input views and removes unwanted artifacts and missing regions. To effectively utilize diffusion priors, our sampling strategy selects virtual camera trajectories that can maximize the information gain of the scene." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": "As illustrated in Fig. 2, our approach follows a best-first search strategy with depth-first characteristics to construct virtual trajectories. We begin by selecting " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "N_{Tr}" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " training viewpoints at spatially uniform intervals as the initial states for each trajectory and gradually select the next viewpoint. At each step, the current viewpoint expands into " + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 594, + 295, + 714 + ], + "type": "text", + "content": " candidate views. The one with the highest information gain becomes the next viewpoint in the trajectory, while the rest are stored for future consideration. Here, the current state refers to the most recently added camera in a trajectory." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 432 + ], + "type": "text", + "content": "Candidate views. We define candidate views as the set of possible camera motion primitives, which determine the transitions from the current state. As shown in Fig. 3, our action space consists of 6 translational moves, 4 pure rotations of the camera, and 4 rotations where the camera orbits around the look-at point, totaling 14 actions. This choice ensures the virtual trajectory efficiently exploring the scene when compared to previous works just using going straight or turning around." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 434, + 556, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 556, + 531 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 556, + 531 + ], + "type": "text", + "content": "A candidate view is removed if it (1) lies outside the bounding box or in an occupied region, or (2) is close to the 3D Gaussians before fine-tuning, since rendering can be severely disrupted when the camera is placed near the occupied. We utilize the target scene bounding box and the occupancy grid from Sec. 3.2 to achieve this. These constraints prevent the trajectory from diverging away from the reconstructed scene or producing redundant pseudo-observations." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": "Information gain. We aim to observe 3D Gaussians densely from diverse viewpoints, as sparse or limited-angle observations often lead to inaccurate shape and color estimation. To this end, we measure information gain based on view coverage. We employ a 2D map " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M} \\in \\mathbb{R}^{N_{\\mathcal{G}} \\times D}" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "N_{\\mathcal{G}}" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": " is the total number of Gaussians and " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": " is the number of discretized viewing directions around 3D Gaussians. The map " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": " stores binary values where each entry " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{M}[j,k] = 1" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": " indicates that the " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": "-th Gaussian has been observed from the " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": "-th viewing direction and initialized with training viewpoints " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{V}^T" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": ". At each step " + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 313, + 553, + 556, + 696 + ], + "type": "text", + "content": ", we update the map as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 383, + 700, + 555, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 700, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 383, + 700, + 555, + 715 + ], + "type": "interline_equation", + "content": "\\mathcal {M} _ {t} [ j, k ] = 1, \\quad \\forall j \\in J, \\tag {1}", + "image_path": "766cb880db9461e3adc57a56d4b2d039103a9fb56f0e1d58993e35b9a7f56c83.jpg" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27045" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " denotes the set of indices of 3D Gaussians visible from the candidate viewpoint and " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": " indicates the index assigned to the discretized view direction of the candidate viewpoint. This update is performed based on the previous map " + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t-1}" + }, + { + "bbox": [ + 55, + 72, + 296, + 133 + ], + "type": "text", + "content": ", which can be described as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 119, + 143, + 295, + 160 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 119, + 143, + 295, + 160 + ], + "spans": [ + { + "bbox": [ + 119, + 143, + 295, + 160 + ], + "type": "interline_equation", + "content": "I G _ {t} = \\sum \\left(\\mathcal {M} _ {t} - \\mathcal {M} _ {t - 1}\\right), \\tag {2}", + "image_path": "8da43cad511f4573ae3918bc6ff299e233b7b43b79f52f531cf6a7f48f125fa9.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "inline_equation", + "content": "IG_{t}" + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": " is the number of Gaussians that have either been newly observed or newly viewed from different directions compared to " + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t - 1}" + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": ". 3D Gaussians that are newly observed or seen from a different angle compared to " + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{M}_{t - 1}" + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": ", ensuring that only newly gained information is counted. To avoid redundant overlaps among trajectories, " + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "inline_equation", + "content": "\\mathcal{M}" + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": " is shared across among all trajectories " + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "inline_equation", + "content": "Tr" + }, + { + "bbox": [ + 55, + 171, + 296, + 255 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "content": "Termination. The search process terminates under two conditions. First, if no valid candidate views can be expanded from the current viewpoint while the number of cameras in the trajectory is still below the predefined length " + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "content": ", the current branch is pruned and the search resumes from the next-best candidate view in the priority queue. Second, if the trajectory reaches the predefined length " + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "content": ", the search concludes. The resulting trajectory is treated as a set of pseudo-observations, and then begins the searching process for the next trajectory " + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "inline_equation", + "content": "Tr_{i+1}" + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "content": ". This process is repeated until " + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "inline_equation", + "content": "N_{Tr}" + }, + { + "bbox": [ + 55, + 274, + 296, + 418 + ], + "type": "text", + "content": " trajectories are constructed. We provide visualizations of example results in the supplementary materials." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 427, + 149, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 427, + 149, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 427, + 149, + 441 + ], + "type": "text", + "content": "3.4. Diffusion prior" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "text", + "content": "Since renderings from virtual viewpoints often partially include well-reconstructed parts, it is desirable to leverage these visual cues. To this end, we train a video diffusion model to enhance degraded rendered images " + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "inline_equation", + "content": "I^{V}" + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "text", + "content": " to be realistic images " + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "inline_equation", + "content": "I^{G}" + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "text", + "content": ". Our model is built on [21] but introduces several key modifications. First, unlike [21], which adopts a frame interpolation setup, we address extreme view extrapolation in scene exploration by performing next-frame prediction using only the observation closest to the virtual camera as the sole visual cue. Second, to mitigate this ill-posed task, we condition the model on the scene text description and the ground truth image " + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "inline_equation", + "content": "I^{T}" + }, + { + "bbox": [ + 55, + 445, + 296, + 662 + ], + "type": "text", + "content": " from the nearest training view to each pseudo-observation. These are injected into the cross-attention layers of the diffusion model. At last, we create a dataset specifically designed to model artifacts arising from extreme extrapolation. The details about the dataset and the training process for the diffusion model are provided in the supplementary materials." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 671, + 130, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 130, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 130, + 684 + ], + "type": "text", + "content": "3.5. Finetuning" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Previous work [21] observed that applying an image-based loss (e.g., L1 loss) to a rendered image from a virtual" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 190 + ], + "type": "text", + "content": "camera and its pseudo ground truth often degrades well-reconstructed regions, as 3DGS is sensitive to minor inaccuracies from diffusion model. It mitigates this issue by using two confidence levels, where higher confidence values indicate fewer conflicts between the contents of training and virtual viewpoints, thereby guiding 3D Gaussians to receive stronger gradients. It adopts the image level confidence based on the distance between a virtual viewpoint and a training viewpoint and pixel level confidence based on the scale of 3D Gaussians." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 192, + 554, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 192, + 554, + 287 + ], + "spans": [ + { + "bbox": [ + 313, + 192, + 554, + 287 + ], + "type": "text", + "content": "Higher image-level confidence is applied when virtual viewpoints far from training viewpoints since they have less conflicts between them. While effective, using distance alone often fails to account for pure camera rotations, as the distance remains unchanged even when the view frustum overlap between viewpoints decreases. As our generated trajectories involve both translations and rotations, this results in suboptimal supervision." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 288, + 554, + 384 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 288, + 554, + 384 + ], + "spans": [ + { + "bbox": [ + 313, + 288, + 554, + 384 + ], + "type": "text", + "content": "To address this issue, we introduce a novel image-level confidence metric based on intersection-over-union (IoU) of Gaussians, denoted as " + }, + { + "bbox": [ + 313, + 288, + 554, + 384 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 288, + 554, + 384 + ], + "type": "text", + "content": "-IOU. Our key insight is that a high view frustum overlaps between a training viewpoint and a virtual viewpoint implies many co-visible 3D Gaussians, leading to more conflicts. Using the Gaussian rasterizer, we efficiently extract the indices of visible Gaussians from each view, and define their intersection as the co-visible set:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 350, + 389, + 555, + 406 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 389, + 555, + 406 + ], + "spans": [ + { + "bbox": [ + 350, + 389, + 555, + 406 + ], + "type": "interline_equation", + "content": "\\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right) = \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {V}\\right) \\cap \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right), \\tag {3}", + "image_path": "dbaa1823d368d38d4e97180ed7e710fcdcdf497844fb27f8ad926c200f510e47.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\mathcal{G}_{vis}(\\cdot)" + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": " denotes the set of 3D Gaussians visible from a given viewpoint, and " + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\nu_{ref}" + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": " refers to the training viewpoint nearest to " + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "inline_equation", + "content": "\\nu^V" + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": ". Based on this, " + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "inline_equation", + "content": "G" + }, + { + "bbox": [ + 313, + 412, + 554, + 448 + ], + "type": "text", + "content": "-IOU is defined as" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 356, + 454, + 553, + 483 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 454, + 553, + 483 + ], + "spans": [ + { + "bbox": [ + 356, + 454, + 553, + 483 + ], + "type": "interline_equation", + "content": "G I O U (\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}) = \\frac {\\left| \\mathcal {G} _ {\\text {c o v i s}} \\left(\\mathcal {V} ^ {V} , \\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}{\\left| \\mathcal {G} _ {\\text {v i s}} \\left(\\mathcal {V} ^ {\\text {r e f}}\\right) \\right|}, \\tag {4}", + "image_path": "a2ebe9fc6a048af4a4040ce7022c4e141e2ab67f495837f246bda9beea356944.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 488, + 487, + 500 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 487, + 500 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 487, + 500 + ], + "type": "text", + "content": "and the image-level confidence is given by" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 373, + 506, + 553, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 373, + 506, + 553, + 521 + ], + "spans": [ + { + "bbox": [ + 373, + 506, + 553, + 521 + ], + "type": "interline_equation", + "content": "U _ {\\mathrm {i m g}} = 1 - G \\mathrm {I O U} \\left(\\mathcal {V} ^ {V}, \\mathcal {V} ^ {\\text {r e f}}\\right). \\tag {5}", + "image_path": "cdbf61d6492c26cd0471202bde6a34d14bc015d624c9a52bef5d7993817951e1.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "text", + "content": "For the pixel level confidence, rather than solely relying on the scale of 3D Gaussians which still produce artifacts depending on the viewpoint, we define it as the perceptual distance [46] between the rendered image " + }, + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "inline_equation", + "content": "I^{V}" + }, + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "text", + "content": " and the output of the diffusion model " + }, + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "inline_equation", + "content": "I^{G}" + }, + { + "bbox": [ + 313, + 529, + 554, + 624 + ], + "type": "text", + "content": ". Each pixel in the confidence map reflects its contribution to the overall perceptual difference. We upsample the pixel-aligned LPIPS map to the original image resolution and define pixel-level confidence as" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 351, + 632, + 553, + 647 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 351, + 632, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 351, + 632, + 553, + 647 + ], + "type": "interline_equation", + "content": "U _ {\\text {p i x e l}} (x, y) = \\mathcal {U} (\\mathrm {L P I P S} \\left(I ^ {V}, I ^ {G}\\right)) (x, y), \\tag {6}", + "image_path": "1ac9e804a950919a23cfc17e73506518a6203b5f593f2f54f66d70e2005828ba.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "U_{\\mathrm{pixel}}(x,y)" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": " denotes the confidence value at pixel " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathcal{U}" + }, + { + "bbox": [ + 313, + 654, + 553, + 689 + ], + "type": "text", + "content": " is an upsampling operation to the original image resolution." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 553, + 712 + ], + "type": "text", + "content": "Combining the above confidence, the loss for training views is defined as" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "27046" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 84, + 294, + 265 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 278, + 83 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 278, + 83 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 278, + 83 + ], + "type": "text", + "content": "Training viewpoints Evaluation viewpoints Evaluation viewpoints (2nd)" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 84, + 294, + 265 + ], + "lines": [ + { + "bbox": [ + 58, + 84, + 294, + 265 + ], + "spans": [ + { + "bbox": [ + 58, + 84, + 294, + 265 + ], + "type": "image", + "image_path": "441cbd61c03ccda638c57383f59633b4019a369426d83303a11ce7ff32936151.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 268, + 295, + 300 + ], + "lines": [ + { + "bbox": [ + 55, + 268, + 295, + 300 + ], + "spans": [ + { + "bbox": [ + 55, + 268, + 295, + 300 + ], + "type": "text", + "content": "Figure 4. Viewpoint difference comparison with previous 3D reconstruction dataset. Training and evaluation viewpoints are visualized." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 1 + }, + { + "bbox": [ + 91, + 319, + 294, + 335 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 319, + 294, + 335 + ], + "spans": [ + { + "bbox": [ + 91, + 319, + 294, + 335 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V} ^ {T}} = \\left\\| I ^ {T} - I ^ {V} \\right\\| _ {1} + \\mathcal {L} _ {\\mathrm {D - S S I M}} \\left(I ^ {T}, I ^ {V}\\right), \\tag {7}", + "image_path": "67ea7ff51813beef52306ff19c1fb16fcc03823d7f717274de5bb67d319df65e.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 339, + 228, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 339, + 228, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 339, + 228, + 350 + ], + "type": "text", + "content": "and the loss for virtual views is defined as" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 357, + 296, + 383 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 383 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 383 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathcal {V} ^ {V}} = U _ {i m g} \\cdot \\left(\\left| \\left| I ^ {G} - I ^ {V} \\right| \\right| _ {1} \\odot U _ {p i x e l} + \\mathcal {L} _ {\\mathrm {D} - \\mathrm {S S I M}} \\left(I ^ {G}, I ^ {V}\\right)\\right). \\tag {8}", + "image_path": "b4d99f5f2970770e427fdaad7044c316403383dcb35f54032203c8f153e6e187.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "text", + "content": "The " + }, + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "text", + "content": " denotes the pixel-wise multiplication and " + }, + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{D - SSIM}}" + }, + { + "bbox": [ + 55, + 384, + 296, + 420 + ], + "type": "text", + "content": " denotes the dissimilarity variant of the SSIM loss. This fine-tuning process is also illustrated in Fig. 2 (c)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 430, + 276, + 443 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 430, + 276, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 430, + 276, + 443 + ], + "type": "text", + "content": "4. Evaluation dataset for scene exploration" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 449, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 449, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 449, + 296, + 581 + ], + "type": "text", + "content": "Since existing datasets remain inadequate for evaluating scene exploration, as the test cameras are typically positioned within or close to the training trajectories, resulting in interpolation scenarios or only minor viewpoint variations. Although the Nerfbusters dataset [40] introduces extreme novel viewpoints via two separate camera trajectories, its test views are placed far behind the training views. Consequently, the task aligns more with generative outpainting rather than complete scene reconstruction. We introduce two datasets tailored for this task, with additional details provided in the supplementary materials." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 588, + 138, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 138, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 138, + 601 + ], + "type": "text", + "content": "4.1. WildExplore" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "To address the lack of an appropriate benchmark for scene exploration, we introduce WildExplore, a new dataset comprising four indoor and four outdoor scenes. As illustrated in Fig. 4, WildExplore is designed with two key properties: (1) it explicitly targets the exploration task by placing test viewpoints significantly distant from the training trajectories, ensuring large viewpoint transitions and broad scene coverage, and (2) the evaluation set is captured to ensure sufficient co-visible areas between training and test views," + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 95 + ], + "type": "text", + "content": "minimizing the need for generating arbitrary contents, in contrast to Nerfbusters." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 103, + 434, + 115 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 103, + 434, + 115 + ], + "spans": [ + { + "bbox": [ + 313, + 103, + 434, + 115 + ], + "type": "text", + "content": "4.2. Curated Nerfbusters" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 121, + 555, + 217 + ], + "type": "text", + "content": "We curate the Nerfbusters dataset to better align with scene exploration objectives. We selects some scenes where the test views are placed far behind the training views and swap the train/test split, while leaving the others unchanged. This adjustment transforms the task into \"moving forward and closely examining the scene\", which better reflects realistic exploration scenarios. Our curated version includes seven scenes with swapped splits and two in their original form." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 228, + 395, + 242 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 395, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 395, + 242 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 248, + 430, + 262 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 248, + 430, + 262 + ], + "spans": [ + { + "bbox": [ + 313, + 248, + 430, + 262 + ], + "type": "text", + "content": "5.1. Experimental Setup" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "text", + "content": "Implementation details. We first train 3DGS using the given training viewpoints, following the original setup in [11]. The number of virtual trajectories " + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "text", + "content": " is set to 20 by default. The occupancy grid resolution " + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "inline_equation", + "content": "S" + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "text", + "content": " is set to 64. At each step, we select the top 3 candidate views based on computed information gain. We set " + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 266, + 555, + 410 + ], + "type": "text", + "content": " as 32 for discretized view direction. For the video diffusion model, we adopt DynamiCrafter [42] as the backbone. After generating the pseudo observation set, we fine-tune 3D Gaussians for 15K iterations, applying densification until 9K, with the same scheduling as 3DGS. Further details will be discussed in the supplements." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 425, + 556, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 425, + 556, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 425, + 556, + 581 + ], + "type": "text", + "content": "Baselines. We evaluate our method in comparison to 3DGS-based methods, including 3DGS [11], 3DGS + Depth, DNGaussians [15], and ViewExtrapolator [19]. 3DGS + Depth applies additional depth regularization using monocular depth estimation [44], following the implementation from the official 3DGS repository [11]. DNGaussians [15] incorporates depth priors during partial optimization of the 3D Gaussian parameters. ViewExtrapolator leverages video diffusion but heuristically selects a single test camera for generating additional viewpoints due to the absence of a dedicated sampling strategy. In our experiments, we follow their protocol and choose the test camera farthest from the training set to encourage broader viewpoint coverage." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 588, + 372, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 372, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 372, + 600 + ], + "type": "text", + "content": "5.2. Results" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 712 + ], + "type": "text", + "content": "Quantitative comparison. Table 1 reports the quantitative results on the exploration task. To assess reconstruction quality, we measure PSNR, SSIM, and LPIPS [46]. The first row reports the result from 3DGS trained on both the training and test splits, serving as an upper bound. Since our method effectively removes artifacts and fills missing regions, as observed in the qualitative analysis, our model outperforms across all metrics, particularly in PSNR and SSIM." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27047" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 88, + 553, + 494 + ], + "blocks": [ + { + "bbox": [ + 92, + 76, + 107, + 86 + ], + "lines": [ + { + "bbox": [ + 92, + 76, + 107, + 86 + ], + "spans": [ + { + "bbox": [ + 92, + 76, + 107, + 86 + ], + "type": "text", + "content": "GT" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 165, + 76, + 192, + 87 + ], + "lines": [ + { + "bbox": [ + 165, + 76, + 192, + 87 + ], + "spans": [ + { + "bbox": [ + 165, + 76, + 192, + 87 + ], + "type": "text", + "content": "3DGS" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 233, + 76, + 291, + 87 + ], + "lines": [ + { + "bbox": [ + 233, + 76, + 291, + 87 + ], + "spans": [ + { + "bbox": [ + 233, + 76, + 291, + 87 + ], + "type": "text", + "content": "3DGS+Depth" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 318, + 76, + 372, + 87 + ], + "lines": [ + { + "bbox": [ + 318, + 76, + 372, + 87 + ], + "spans": [ + { + "bbox": [ + 318, + 76, + 372, + 87 + ], + "type": "text", + "content": "DNGaussian" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 389, + 76, + 464, + 87 + ], + "lines": [ + { + "bbox": [ + 389, + 76, + 464, + 87 + ], + "spans": [ + { + "bbox": [ + 389, + 76, + 464, + 87 + ], + "type": "text", + "content": "ViewExtrapolator" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 497, + 76, + 518, + 87 + ], + "lines": [ + { + "bbox": [ + 497, + 76, + 518, + 87 + ], + "spans": [ + { + "bbox": [ + 497, + 76, + 518, + 87 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 88, + 553, + 494 + ], + "lines": [ + { + "bbox": [ + 58, + 88, + 553, + 494 + ], + "spans": [ + { + "bbox": [ + 58, + 88, + 553, + 494 + ], + "type": "image", + "image_path": "8b4e5d921fcfe5dfb336ead3abf129ae68585315f5c66be4ae5fe963f8748428.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 180, + 497, + 428, + 508 + ], + "lines": [ + { + "bbox": [ + 180, + 497, + 428, + 508 + ], + "spans": [ + { + "bbox": [ + 180, + 497, + 428, + 508 + ], + "type": "text", + "content": "Figure 5. Qualitative comparison on the curated Nerfbusters dataset." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 517, + 295, + 708 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 708 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 708 + ], + "type": "text", + "content": "Qualitative comparison. Fig. 5 and Fig. 6 show qualitative comparisons among our method and baseline methods. 3DGS [11] suffers from artifacts, and its variants with depth regularization also meet the same problem, as they lack the capability to fill missing information. Although ViewExtrapolator [19] utilizes a video diffusion prior and thereby shows competitive performance compared to 3DGS variants, it underperforms in challenging scenarios. Its heuristic camera sampling struggles to maximize information gain and results in limited viewpoint coverage, leading to lower visual quality when rendering from distant viewpoints. Its training-free diffusion prior also acts as a bottleneck. In contrast, our method fills missing regions and removes artifacts more effectively, producing images that align closely with the ground truth. See the supplementary materials for additional results." + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 317, + 515, + 553, + 599 + ], + "blocks": [ + { + "bbox": [ + 317, + 515, + 553, + 599 + ], + "lines": [ + { + "bbox": [ + 317, + 515, + 553, + 599 + ], + "spans": [ + { + "bbox": [ + 317, + 515, + 553, + 599 + ], + "type": "table", + "html": "
MethodWildExploreCurated Nerfbusters
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
3DGS Pseudo GT30.370.9100.14429.340.9110.128
3DGS15.810.4870.41414.510.4170.454
3DGS + Depth15.780.4830.41315.000.4350.436
DNGaussians15.650.5040.45014.010.4420.525
ViewExtrapolator16.020.5100.44214.980.4500.455
Ours16.960.5340.41316.290.4790.434
", + "image_path": "24e08622fa8e69d4c1a60811a414fc87cbd6303d476990dec6950da232d11509.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_body" + } + ], + "index": 9 + }, + { + "bbox": [ + 372, + 604, + 496, + 615 + ], + "lines": [ + { + "bbox": [ + 372, + 604, + 496, + 615 + ], + "spans": [ + { + "bbox": [ + 372, + 604, + 496, + 615 + ], + "type": "text", + "content": "Table 1. Quantitative comparison." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 634, + 406, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 634, + 406, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 634, + 406, + 647 + ], + "type": "text", + "content": "5.3. Ablation study" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 555, + 715 + ], + "type": "text", + "content": "Information gain. We validate the effectiveness of our information gain design by comparing it with alternative strategies, as shown in Table 2. One alternative samples the 3D space uniformly and evaluate view coverage based on the occupied vertices of an occupancy grid, as proposed in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27048" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 74, + 555, + 264 + ], + "blocks": [ + { + "bbox": [ + 57, + 74, + 555, + 264 + ], + "lines": [ + { + "bbox": [ + 57, + 74, + 555, + 264 + ], + "spans": [ + { + "bbox": [ + 57, + 74, + 555, + 264 + ], + "type": "image", + "image_path": "87c4858e83b335710e494b0d46297f3a9ccb3428cc34b19b364781e2e137c723.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 192, + 266, + 416, + 277 + ], + "lines": [ + { + "bbox": [ + 192, + 266, + 416, + 277 + ], + "spans": [ + { + "bbox": [ + 192, + 266, + 416, + 277 + ], + "type": "text", + "content": "Figure 6. Qualitative comparison on the WildExplore dataset." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 59, + 292, + 293, + 365 + ], + "blocks": [ + { + "bbox": [ + 59, + 292, + 293, + 365 + ], + "lines": [ + { + "bbox": [ + 59, + 292, + 293, + 365 + ], + "spans": [ + { + "bbox": [ + 59, + 292, + 293, + 365 + ], + "type": "table", + "html": "
Virtual View SamplingCurated Nerfbusters
Info. GainPSNR↑SSIM↑LPIPS↓
Grid view coverage [22]15.820.46750.4451
FisherRF [9]15.770.45930.4394
GS view coverage (Ours)16.290.47900.4336
", + "image_path": "c4bb2f5fff4ed1f7fa02b39174673d37a9563ff2c115a8798fa229f621a5a6e4.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 82, + 389, + 269, + 450 + ], + "blocks": [ + { + "bbox": [ + 84, + 370, + 266, + 381 + ], + "lines": [ + { + "bbox": [ + 84, + 370, + 266, + 381 + ], + "spans": [ + { + "bbox": [ + 84, + 370, + 266, + 381 + ], + "type": "text", + "content": "Table 2. Comparisons on information gain design." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 82, + 389, + 269, + 450 + ], + "lines": [ + { + "bbox": [ + 82, + 389, + 269, + 450 + ], + "spans": [ + { + "bbox": [ + 82, + 389, + 269, + 450 + ], + "type": "table", + "html": "
Method (K = 3)Garbage
PSNR↑SSIM↑LPIPS↓
3DGS14.420.4800.526
Ours with Bottom-K14.220.4750.521
Ours with Top-K15.450.5260.508
", + "image_path": "969a2e826341533e730d75a5884bdc3f444558648b412c3c8cfb2cb523f9af89.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 63, + 477, + 287, + 585 + ], + "blocks": [ + { + "bbox": [ + 59, + 454, + 291, + 465 + ], + "lines": [ + { + "bbox": [ + 59, + 454, + 291, + 465 + ], + "spans": [ + { + "bbox": [ + 59, + 454, + 291, + 465 + ], + "type": "text", + "content": "Table 3. Ablation study on information gain. TopK vs BottomK" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 63, + 477, + 287, + 585 + ], + "lines": [ + { + "bbox": [ + 63, + 477, + 287, + 585 + ], + "spans": [ + { + "bbox": [ + 63, + 477, + 287, + 585 + ], + "type": "table", + "html": "
FinetuningCurated Nerfbusters
Image levelPixel levelPSNR↑SSIM↑LPIPS↓
Distance [21]-15.000.4270.443
-Scale [21]16.180.4760.442
DistanceScale15.270.4320.440
--16.300.4760.447
-pixelLPIPS16.230.4780.441
GS-IOU-16.100.4730.439
GS-IOUpixelLPIPS16.290.4790.434
", + "image_path": "e0ea9b5eff8efb347febdd71e32aef4e928d3ac57d2cd439268098762d745949.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 87, + 588, + 262, + 599 + ], + "lines": [ + { + "bbox": [ + 87, + 588, + 262, + 599 + ], + "spans": [ + { + "bbox": [ + 87, + 588, + 262, + 599 + ], + "type": "text", + "content": "Table 4. Ablations study on finetuning methods." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "[22]. Another follows typical active-learning strategies and select viewpoints with the highest uncertainty. We adopt the method proposed in [9] for comparison. As presented in the first and second rows, both show suboptimal results. Grid-based approach often fails to maximize information gain, as it includes the gain from free space, resulting in redundant viewpoint selections. For uncertainty-based method, we observe that it is sensitive to 3D Gaussians with large scales," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 293, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 293, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 293, + 555, + 319 + ], + "type": "text", + "content": "tending to chase them rather than addressing artifacts from viewpoint changes." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 555, + 392 + ], + "type": "text", + "content": "Search strategy. We ablate our best first search strategy to validate its effectiveness. As shown in Table 3, the strategy selecting bottom K makes virtual cameras tend to stay near the trajectory of training viewpoints rather than deviating from it, leading to no performance gain." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 406, + 556, + 528 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 406, + 556, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 406, + 556, + 528 + ], + "type": "text", + "content": "Finetuning methods. We validate our finetuning approach by benchmarking it against the methods proposed by [21], as shown in Table 4. The model without the image-level confidence produces worse results in terms of LPIPS, as generated images whose viewpoints closely match the training set negatively impact already well-reconstructed regions. Ignoring camera rotation further degrades performance, as observed in the second and fifth rows. Although both our image-level confidence map and scale based one are effective, our method leads to slight better performance." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 388, + 549 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 557, + 555, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 555, + 628 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 555, + 628 + ], + "type": "text", + "content": "This paper presents a framework that enables high-quality rendering of the novel view in arbitrary viewpoints. Our method incorporates the video diffusion priors and the real-time rendering capability of 3DGS. It introduces a sampling strategy to effectively sample trajectories of virtual cameras for maximizing information gain." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 628, + 556, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 556, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 556, + 712 + ], + "type": "text", + "content": "Limitations and future works. While our method incorporates video diffusion priors to fill the missing information, we expect that employing more strong diffusion models which can faithfully generate high frequency details may improve the overall performance. In addition, extending the scene bounding box to cover a large scale scene would be an interesting avenue for the future work." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27049" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 90, + 297, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 90, + 297, + 159 + ], + "spans": [ + { + "bbox": [ + 55, + 90, + 297, + 159 + ], + "type": "text", + "content": "This work was supported by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) (No. RS-2022-II220113, No. RS-2024-00457882), Artificial Intelligence Graduate School Program, Yonsei University, under Grant RS-2020-11201361." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "spans": [ + { + "bbox": [ + 56, + 178, + 115, + 192 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 198, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 198, + 297, + 253 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 198, + 297, + 253 + ], + "spans": [ + { + "bbox": [ + 61, + 198, + 297, + 253 + ], + "type": "text", + "content": "[1] Jonathan T Barron, Ben Mildenhall, Dor Verbin, Pratul P Srinivasan, and Peter Hedman. Mip-nerf 360: Unbounded anti-aliased neural radiance fields. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5470-5479, 2022. 1, 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 255, + 296, + 287 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 255, + 296, + 287 + ], + "spans": [ + { + "bbox": [ + 61, + 255, + 296, + 287 + ], + "type": "text", + "content": "[2] Jonathan T. Barron, Ben Mildenhall, Dor Verbin, Pratul P. Srinivasan, and Peter Hedman. Zip-nerf: Anti-aliased grid-based neural radiance fields, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 61, + 289, + 296, + 343 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, Varun Jampani, and Robin Rombach. Stable video diffusion: Scaling latent video diffusion models to large datasets, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 345, + 296, + 388 + ], + "type": "text", + "content": "[4] Ruiqi Gao, Aleksander Holynski, Philipp Henzler, Arthur Brussee, Ricardo Martin-Brualla, Pratul Srinivasan, Jonathan T. Barron, and Ben Poole. Cat3d: Create anything in 3d with multi-view diffusion models, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 389, + 296, + 422 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 389, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 62, + 389, + 296, + 422 + ], + "type": "text", + "content": "[5] Jeong gi Kwak, Erqun Dong, Yuhe Jin, Hanseok Ko, Shweta Mahajan, and Kwang Moo Yi. Vivid-1-to-3: Novel view synthesis with video diffusion models, 2023." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 423, + 296, + 456 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 423, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 62, + 423, + 296, + 456 + ], + "type": "text", + "content": "[6] Hao He, Yinghao Xu, Yuwei Guo, Gordon Wetzstein, Bo Dai, Hongsheng Li, and Ceyuan Yang. Cameractrol: Enabling camera control for text-to-video generation, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 457, + 296, + 512 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 457, + 296, + 512 + ], + "spans": [ + { + "bbox": [ + 62, + 457, + 296, + 512 + ], + "type": "text", + "content": "[7] Hanzhe Hu, Zhizhuo Zhou, Varun Jampani, and Shubham Tulsiani. Mvd-fusion: Single-view 3d via depth-consistent multi-view generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9698-9707, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 62, + 514, + 296, + 567 + ], + "type": "text", + "content": "[8] Zehuan Huang, Hao Wen, Junting Dong, Yaohui Wang, Yangguang Li, Xinyuan Chen, Yan-Pei Cao, Ding Liang, Yu Qiao, Bo Dai, and Lu Sheng. Epidiff: Enhancing multi-view synthesis via localized epipolar-constrained diffusion, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "spans": [ + { + "bbox": [ + 62, + 568, + 296, + 612 + ], + "type": "text", + "content": "[9] Wen Jiang, Boshu Lei, and Kostas Daniilidis. Fisherrf: Active view selection and uncertainty quantification for radiance fields using fisher information. arXiv preprint arXiv:2311.17874, 2023. 3, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "spans": [ + { + "bbox": [ + 57, + 613, + 296, + 668 + ], + "type": "text", + "content": "[10] Liren Jin, Xieyuanli Chen, Julius Rückin, and Marija Popovic. Neu-nbv: Next best view planning using uncertainty estimation in image-based neural rendering. In 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 11305-11312, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 712 + ], + "type": "text", + "content": "[11] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Arno Knapitsch, Jaesik Park, Qian-Yi Zhou, and Vladlen Koltun. Tanks and temples: benchmarking large-scale scene reconstruction. ACM Transactions on Graphics, 36:1-13, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 553, + 149 + ], + "type": "text", + "content": "[13] Xin Kong, Shikun Liu, Xiaoyang Lyu, Marwan Taher, Xiaojuan Qi, and Andrew J. Davison. Eschernet: A generative model for scalable view synthesis, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 151, + 553, + 194 + ], + "type": "text", + "content": "[14] Georgios Kopanas and George Drettakis. Improving NeRF Quality by Progressive Camera Placement for Free-Viewpoint Navigation. In Vision, Modeling, and Visualization. The Eurographics Association, 2023. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 237 + ], + "type": "text", + "content": "[15] Jiahe Li, Jiawei Zhang, Xiao Bai, Jin Zheng, Xin Ning, Jun Zhou, and Lin Gu. Dngaussian: Optimizing sparse-view 3d gaussian radiance fields with global-local depth normalization, 2024. 2, 6" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "spans": [ + { + "bbox": [ + 316, + 239, + 553, + 314 + ], + "type": "text", + "content": "[16] Lu Ling, Yichen Sheng, Zhi Tu, Wentian Zhao, Cheng Xin, Kun Wan, Lantao Yu, Qianyu Guo, Zixun Yu, Yawen Lu, Xuanmao Li, Xingpeng Sun, Rohan Ashok, Aniruddha Mukherjee, Hao Kang, Xiangrui Kong, Gang Hua, Tianyi Zhang, Bedrich Benes, and Aniket Bera. Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision, 2023. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 358 + ], + "type": "text", + "content": "[17] Fangfu Liu, Wenqiang Sun, Hanyang Wang, Yikai Wang, Haowen Sun, Junliang Ye, Jun Zhang, and Yueqi Duan. Reconx: Reconstruct any scene from sparse views with video diffusion model, 2024. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 403 + ], + "type": "text", + "content": "[18] Guangyi Liu, Wen Jiang, Boshu Lei, Vivek Pandey, Kostas Daniilidis, and Nader Motee. Beyond uncertainty: Risk-aware active view acquisition for safe robot navigation and 3d scene understanding with fisherrf, 2025. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "text", + "content": "[19] Kunhao Liu, Ling Shao, and Shijian Lu. Novel view extrapolation with video diffusion priors, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 459 + ], + "type": "text", + "content": "[20] Ruoshi Liu, Rundi Wu, Basile Van Hoorick, Pavel Tokmakov, Sergey Zakharov, and Carl Vondrick. Zero-1-to-3: Zero-shot one image to 3d object, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 460, + 553, + 492 + ], + "type": "text", + "content": "[21] Xi Liu, Chaoyi Zhou, and Siyu Huang. 3dgs-enhancer: Enhancing unbounded 3d gaussian splatting with view-consistent 2d diffusion priors, 2024. 2, 3, 5, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 557 + ], + "type": "text", + "content": "[22] Pierre Marza, Laetitia Matignon, Olivier Simonin, Dhruv Batra, Christian Wolf, and Devendra S Chaplot. Autonerf: Training implicit scene representations with autonomous agents. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 13442-13449. IEEE, 2024. 3, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 603 + ], + "type": "text", + "content": "[23] Ben Mildenhall, Pratul P. Srinivasan, Rodrigo Ortiz-Cayon, Nima Khademi Kalantari, Ravi Ramamoorthi, Ren Ng, and Abhishek Kar. Local light field fusion: Practical view synthesis with prescriptive sampling guidelines, 2019. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "spans": [ + { + "bbox": [ + 316, + 604, + 553, + 646 + ], + "type": "text", + "content": "[24] Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view synthesis. In ECCV, 2020. 1, 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[25] Thomas Müller, Alex Evans, Christoph Schied, and Alexander Keller. Instant neural graphics primitives with a multiresolution hash encoding. ACM transactions on graphics (TOG), 41(4):1-15, 2022. 1" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[26] Michael Niemeyer, Jonathan T. Barron, Ben Mildenhall, Mehdi S. M. Sajjadi, Andreas Geiger, and Noha Radwan." + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "27050" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "Regnerf: Regularizing neural radiance fields for view synthesis from sparse inputs, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 140 + ], + "type": "text", + "content": "[27] Xuran Pan, Zihang Lai, Shiji Song, and Gao Huang. Activenerf: Learning where to see with uncertainty estimation. In European Conference on Computer Vision, pages 230-246. Springer, 2022. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 163 + ], + "type": "text", + "content": "[28] Ben Poole, Ajay Jain, Jonathan T. Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion, 2022. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 208 + ], + "type": "text", + "content": "[29] Jeremy Reizenstein, Roman Shapovalov, Philipp Henzler, Luca Sbordone, Patrick Labatut, and David Novotny. Common objects in 3d: Large-scale learning and evaluation of real-life 3d category reconstruction, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 294, + 262 + ], + "type": "text", + "content": "[30] Kyle Sargent, Zizhang Li, Tanmay Shah, Charles Herrmann, Hong-Xing Yu, Yunzhi Zhang, Eric Ryan Chan, Dmitry Lagun, Li Fei-Fei, Deqing Sun, and Jiajun Wu. Zeronvs: Zero-shot 360-degree view synthesis from a single image, 2024. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 294, + 319 + ], + "type": "text", + "content": "[31] Jianxiong Shen, Adria Ruiz, Antonio Agudo, and Francesc Moreno-Noguer. Stochastic neural radiance fields: Quantifying uncertainty in implicit 3d representations. In 2021 International Conference on 3D Vision (3DV), pages 972-981. IEEE, 2021. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 353 + ], + "type": "text", + "content": "[32] Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d generation, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 355, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 295, + 421 + ], + "type": "text", + "content": "[33] Meng-Li Shih, Wei-Chiu Ma, Lorenzo Boyice, Aleksander Holynski, Forrester Cole, Brian Curless, and Janne Kontkanen. Extranerf: Visibility-aware view extrapolation of neural radiance fields with diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20385-20395, 2024. 2, 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "spans": [ + { + "bbox": [ + 56, + 422, + 294, + 454 + ], + "type": "text", + "content": "[34] Niko Sünderhauf, Jad Abou-Chakra, and Dimity Miller. Density-aware nerf ensembles: Quantifying predictive uncertainty in neural radiance fields, 2022. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 488 + ], + "type": "text", + "content": "[35] Prune Truong, Marie-Julie Rakotosaona, Fabian Manhardt, and Federico Tombari. Sparf: Neural radiance fields from sparse and noisy poses, 2023. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 490, + 294, + 544 + ], + "type": "text", + "content": "[36] Vikram Voleti, Chun-Han Yao, Mark Boss, Adam Letts, David Pankratz, Dmitry Tochilkin, Christian Laforte, Robin Rombach, and Varun Jampani. Sv3d: Novel multi-view synthesis and 3d generation from a single image using latent video diffusion, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 294, + 578 + ], + "type": "text", + "content": "[37] Haochen Wang, Xiaodan Du, Jiahao Li, Raymond A. Yeh, and Greg Shakhnarovich. Score jacobian chaining: Lifting pretrained 2d diffusion models for 3d generation, 2022. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 580, + 294, + 622 + ], + "type": "text", + "content": "[38] Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distillation, 2023. 3" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 294, + 666 + ], + "type": "text", + "content": "[39] Zhouxia Wang, Ziyang Yuan, Xintao Wang, Tianshui Chen, Menghan Xia, Ping Luo, and Ying Shan. Motionctrl: A unified and flexible motion controller for video generation, 2024. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[40] Frederik Warburg, Ethan Weber, Matthew Tancik, Aleksander Holynski, and Angjoo Kanazawa. Nerfbusters: Removing ghostly artifacts from casually captured nerfs, 2023. 2, 6" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 342 + ], + "type": "list", + "angle": 0, + "index": 23, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 116 + ], + "type": "text", + "content": "[41] Wenhui Xiao, Rodrigo Santa Cruz, David Ahmed-Aristizabal, Olivier Salvado, Clinton Fookes, and Leo Lebrat. Nerf director: Revisiting view selection in neural volume rendering, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "spans": [ + { + "bbox": [ + 315, + 118, + 553, + 162 + ], + "type": "text", + "content": "[42] Jinbo Xing, Menghan Xia, Yong Zhang, Haoxin Chen, Wangbo Yu, Hanyuan Liu, Xintao Wang, Tien-Tsin Wong, and Ying Shan. Dynamiccafter: Animating open-domain images with video diffusion priors, 2023. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[43] Haolin Xiong, Sairisheek Muttukuru, Rishi Upadhyay, Pradyumna Chari, and Achuta Kadambi. Sparsegs: Real-time 360deg sparseviewsynthesisusinggaussiansplatting," + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 209, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 239 + ], + "type": "text", + "content": "[44] Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything v2, 2024. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 553, + 274 + ], + "type": "text", + "content": "[45] Huangying Zhan, Jiyang Zheng, Yi Xu, Ian Reid, and Hamid Rezatofighi. Activermap: Radiance field for active mapping and planning, 2022. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 308 + ], + "type": "text", + "content": "[46] Richard Zhang, Phillip Isola, Alexei A. Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric, 2018. 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "spans": [ + { + "bbox": [ + 316, + 309, + 553, + 342 + ], + "type": "text", + "content": "[47] Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images, 2018. 3" + } + ] + } + ], + "index": 22 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "27051" + } + ] + } + ], + "index": 24 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_content_list.json b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c4bce8c1c7cf511d5345ef50557d5dfe7c65256f --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_content_list.json @@ -0,0 +1,1524 @@ +[ + { + "type": "text", + "text": "Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing", + "text_level": 1, + "bbox": [ + 246, + 130, + 751, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Joonghyuk Shin Alchan Hwang Yujin Kim Daneul Kim Jaesik Park Seoul National University", + "bbox": [ + 191, + 203, + 805, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{joonghyuk, alchan00, yujin.k, carpedkm, jaesik.park}@snu.ac.kr", + "bbox": [ + 217, + 241, + 774, + 257 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/586f68b5cedfe0bd0388233b015998856dc89a7b250960ca8ef09befd611a68c.jpg", + "image_caption": [ + "$\\mathrm{SD3}\\rightarrow \\mathrm{Flux}$" + ], + "image_footnote": [], + "bbox": [ + 96, + 306, + 352, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/712320bc04575f5da5148f541a8ba4b5e0ff37b211f266869e7d64527e757bff.jpg", + "image_caption": [ + "giraffe $\\rightarrow$ dragon" + ], + "image_footnote": [], + "bbox": [ + 374, + 306, + 627, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/0e6336a9884ab6b03336fe4c691f10e9a8044a5e8af9e3cbf0f8c17debc3fc6f.jpg", + "image_caption": [ + "frogs $\\rightarrow$ rabbits", + "ferret $\\rightarrow$ lion" + ], + "image_footnote": [], + "bbox": [ + 650, + 305, + 903, + 405 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/193575404c9756b1bbd8970e9ea93db2ad5a38ee8217abc12c561f33284e3e12.jpg", + "image_caption": [ + "'help me' $\\rightarrow$ 'Please don't kill me'" + ], + "image_footnote": [], + "bbox": [ + 96, + 416, + 238, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e13d1ee8f091335a98199df420dfc6ec468a43482a825fb608b986bd658b160d.jpg", + "image_caption": [ + "Figure 1. Our prompt-based editing examples on SD3 series and Flux.1, demonstrating high-quality results across various editing scenarios." + ], + "image_footnote": [], + "bbox": [ + 243, + 416, + 385, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/debaa94d5881c0ad10a2168e113198430a00cea96b0b33283a32a6ddab0a90bc.jpg", + "image_caption": [ + "pig $\\rightarrow$ whale" + ], + "image_footnote": [], + "bbox": [ + 387, + 416, + 527, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/e98a83cd068015c4f1c364d145ef582726591111e7547abd3850856cbffa9ae5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 529, + 416, + 717, + 542 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/6c269df0cc808f17a96040deae4eb2afb2a10317ee5746c5606642d5b69831eb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 718, + 416, + 903, + 542 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 580, + 326, + 595 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Transformer-based diffusion models have recently superseded traditional U-Net architectures, with multimodal diffusion transformers (MM-DiT) emerging as the dominant approach in state-of-the-art models like Stable Diffusion 3 and Flux.1. Previous approaches have relied on unidirectional cross-attention mechanisms, with information flowing from text embeddings to image latents. In contrast, MM-DiT introduces a unified attention mechanism that concatenates input projections from both modalities and performs a single full attention operation, allowing bidirectional information flow between text and image branches. This architectural shift presents significant challenges for existing editing techniques. In this paper, we systematically analyze MM-DiT's attention mechanism by decomposing attention matrices into four distinct blocks, revealing their inherent characteristics. Through these analyses, we propose a robust, prompt-based image editing method for MM-DiT that supports global to local edits across various MM-DiT variants, including few-step models. We believe our findings", + "bbox": [ + 86, + 613, + 485, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "bridge the gap between existing U-Net-based methods and emerging architectures, offering deeper insights into MM-DiT's behavioral patterns.", + "bbox": [ + 511, + 580, + 908, + 628 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 660, + 645, + 676 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent years have witnessed the emergence of diffusion models [21, 60] as the de-facto approach for visual generation, demonstrating remarkable capabilities in creating high-quality data from noise [15, 29, 39, 44, 49, 52, 54, 57]. While early implementations predominantly employed U-shaped networks [55], transformer-based architectures [7, 48] have recently emerged as the prevailing approach, with multimodal diffusion transformers (MM-DiT) showing particular promise in state-of-the-art models like Stable Diffusion 3 series [15] and Flux.1 [30].", + "bbox": [ + 509, + 686, + 906, + 838 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Motivated by MM-DiT's architectural differences and the challenges in directly applying existing U-Net-based editing methods, our work addresses a fundamental question: \"How can we leverage our understanding of U-Net", + "bbox": [ + 511, + 839, + 908, + 900 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "19492", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "attention mechanisms to interpret MM-DiT's behavior better and effectively transfer existing techniques to this new architecture?\" Given the significance of existing work built upon U-Net attention mechanisms [4, 5, 19, 33], answering this question is crucial for enabling existing methods to adapt to, and benefit from, emerging MM-DiT architectures.", + "bbox": [ + 89, + 90, + 480, + 195 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To this end, we systematically investigate MM-DiT's attention mechanisms through block matrix analysis, revealing key characteristics of its attention maps. As one of our key analytical insights, we observe an interesting phenomenon aligned with vision transformer [14] scaling laws [12]: as the model size increases, attention patterns become properly positioned but increasingly noisy, necessitating careful selection of attention layers for editing tasks. Building upon these analyses, we propose an efficient editing method at the architecture level that focuses on modifying image tokens in input projections. The proposed editing method can be jointly combined with inversion methods [56, 63] for editing real images as well.", + "bbox": [ + 89, + 198, + 480, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions are primarily architectural: (1) We provide the first systematic analysis of MM-DiT attention mechanisms, highlighting block-wise patterns and scaling behaviors. (2) Building upon these findings, we propose an efficient and robust editing method across diverse editing scenarios - from global stylistic changes to local edits such as text modification - and various MM-DiT variants, including few-step models (SD3-M, SD3.5-M, SD3.5-L, SD3.5-L-Turbo, and Flux.1-dev/schnell).", + "bbox": [ + 89, + 396, + 480, + 531 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 547, + 232, + 564 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Text-to-image diffusion models", + "text_level": 1, + "bbox": [ + 89, + 574, + 362, + 589 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Early text-to-image diffusion models mostly used U-shaped networks with residual blocks [55] incorporating pixelwise self-attention and text-image cross-attention layers, combined with downsampling and upsampling operations (Fig. 2). These models are typically trained with DDPM-style scheduling and noise prediction objective [49, 54].", + "bbox": [ + 89, + 597, + 480, + 688 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent advances introduced two significant shifts. First, architecturally, there has been a transition from U-Nets to transformer-based architectures like DiT [48], demonstrating superior scaling properties. The PixArt-alpha series [6-8] successfully adapted DiT for text conditional generation, maintaining a conventional cross-attention mechanism for text conditioning. Stable Diffusion 3 [15] subsequently introduced a pivotal shift with multimodal diffusion transformers (MM-DiT), using separate transformers for text and image modalities while concatenating their sequences for unified attention operations (Fig. 2). Second, formulation-wise, newer approaches have adopted v-prediction [58] with a rectified flow framework [32, 35] that connects data and noise distributions along straight paths, replacing the com", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "table", + "img_path": "images/b563a8be41d03824ae08c1205479b54ce604d3ff32391cf25221806bdf9bfb6e.jpg", + "table_caption": [ + "Table 1. Architectural details and inference times for $1024 \\times 1024$ image generation using 28 timesteps, using single A6000 GPU. Times are shown as optimized (left, using PyTorch SDPA kernels) vs. naive matrix multiplication (right). Text encoders: $\\mathrm{CLIP}^1$ refers to CLIP ViT-L/14 [50], $\\mathrm{CLIP}^2$ refers to OpenCLIP ViTG/14 [24], and T5 refers to T5-XXL [51]." + ], + "table_footnote": [], + "table_body": "
Arch. / ParamsText Enc.Attn TypeTime (s)
SD1U-Net / 860MCLIP1Self/Cross5.0/12.7
SDXLU-Net / 2.6BCLIP1,2Self/Cross5.6/7.9
SD3-MMM-DiT / 2BCLIP1,2 / T5Full7.1/20.5
SD3.5-MMM-DiT-X / 2.2BCLIP1,2 / T5Full/Self9.6/26.8
SD3.5-LMM-DiT / 8BCLIP1,2 / T5Full24.7/60.5
Flux.1-devMM-DiT* / 11.9BCLIP1 / T5Full27.8/36.8
", + "bbox": [ + 513, + 181, + 919, + 282 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "plex trajectories of traditional diffusion models. Together with increased model scales, these developments have significantly improved model performance. Parallel efforts have also accelerated iterative diffusion sampling via faster ODE solvers [37, 38], consistency distillation [40, 61], adversarial learning [28, 59, 66, 67], and trajectory straightening [32, 35, 36].", + "bbox": [ + 511, + 295, + 906, + 402 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Text-guided image editing", + "text_level": 1, + "bbox": [ + 511, + 409, + 750, + 426 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Text-guided image editing naturally extends text-to-image diffusion models. Prior work [18, 22, 33, 64] shows that self-attention governs spatial structure while cross-attention handles attribute-level semantics aligned with text. Prompt-to-Prompt (P2P) [19], a seminal approach, leverages this by directly transferring cross-attention maps for precise edits. Subsequent methods further manipulate self-attention mechanisms for better content-preserving edits [4, 25, 43, 70]. A key advantage of these attention-based methods is their ability to achieve precise local edits solely from textual prompts, eliminating the need for explicit spatial guidance or additional user input.", + "bbox": [ + 511, + 431, + 903, + 613 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another direction employs inversion techniques [13, 16, 23, 26, 56, 63], mapping real images to editable latent spaces. While inversion alone can lose fine details, combining it with attention-based methods significantly enhances localized editing [42]. Alternatively, methods like InstructPix2Pix [3] train dedicated models on synthetic data, though this approach incurs high computational costs and dataset preparation overheads [65].", + "bbox": [ + 511, + 613, + 903, + 734 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.3. MM-DiT variants", + "text_level": 1, + "bbox": [ + 511, + 742, + 687, + 757 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Since its introduction, MM-DiT architecture has seen several improvements. The initial dual-branch design in SD3 was motivated by the conceptual differences between text and image embeddings. This architecture was maintained in both SD3 and SD3.5-Large models, which contain 24 and 38 dual-branch blocks, respectively. Flux.1, scaled to 12B parameters, introduced a hybrid approach with 57 total blocks - 19 dual branch blocks followed by 38 single branch blocks (Fig. 2). The single branch variant processes", + "bbox": [ + 511, + 763, + 906, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "19493", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/5e64765a4f55ddaf639b4e1d4903f79d0681ca0528134398176b45032a3395e5.jpg", + "image_caption": [ + "Figure 2. Attention mechanisms across diffusion architectures. (a) U-Net with separate self-/cross- attention, (b) MM-DiT blocks with dual branches, (c) MM-DiT* single branch block used in later layers of Flux.1, and (d) MM-DiT-X dual attention block with explicit self-attention, used in early layers of SD3.5-M." + ], + "image_footnote": [], + "bbox": [ + 89, + 85, + 467, + 320 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "concatenated text and image embeddings as a single modality using unified weights (e.g., linear layers, attention parameters), reducing block parameters from 340M to 141M. The MM-DiT-X variant, later introduced in SD3.5-medium, added explicit self-attention operations exclusively for image tokens in the first 13 transformer blocks for better multi-resolution results and overall image coherence.", + "bbox": [ + 89, + 424, + 483, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Attention Analysis in MM-DiT", + "text_level": 1, + "bbox": [ + 89, + 551, + 375, + 569 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Block matrix formulation of MM-DiT attention", + "text_level": 1, + "bbox": [ + 89, + 578, + 483, + 595 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We begin the analysis with MM-DiT's attention structure under SD3-M's standard dimensions: $64 \\times 64$ image latents (flattened to 4096) and 333 text tokens (77 CLIP + 256 T5), with inner dimension 64. Omitting the batch and head dimensions for clarity, the input projections for image and text domains are denoted as $(\\mathbf{q}_i, \\mathbf{k}_i, \\mathbf{v}_i \\in \\mathbb{R}^{4096 \\times 64})$ and $(\\mathbf{q}_t, \\mathbf{k}_t, \\mathbf{v}_t \\in \\mathbb{R}^{333 \\times 64})$ . Then, query, key, and value projections are concatenated as shown below ( $\\mathbf{q}, \\mathbf{k}, \\mathbf{v} \\in \\mathbb{R}^{4429 \\times 64}$ ). It is worth noting that Flux.1 adopts an inverse concatenation order (text preceding image), which results in a different attention map.", + "bbox": [ + 89, + 603, + 483, + 771 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right], \\mathbf {k} = \\left[ \\begin{array}{l} \\mathbf {k} _ {i} \\\\ \\mathbf {k} _ {t} \\end{array} \\right], \\mathbf {v} = \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 171, + 784, + 483, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Using the standard scaled dot product attention formulation $\\text{softmax}\\left(\\frac{\\mathbf{q}\\mathbf{k}^T}{\\sqrt{d_k}}\\right)\\mathbf{v}$ , we can write the attention map $\\mathbf{q}\\mathbf{k}^T$ and its output representation as follows. Note that row-wise softmax normalization is applied to $\\mathbf{q}\\mathbf{k}^T$ in Eq. (2), before", + "bbox": [ + 89, + 834, + 485, + 902 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/7c710e26405805760b5aedf9c3f112a2d920ce22b9e18f928da065448a00ab17.jpg", + "image_caption": [ + "Figure 3. Visualizing the top-6 principal components extracted from I2I attention blocks, averaged across all timesteps and transformer blocks. Like U-Net's self-attention layers, these components reveal underlying spatial and geometric patterns." + ], + "image_footnote": [], + "bbox": [ + 519, + 90, + 905, + 186 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "multiplication with $\\mathbf{v}$ to obtain Eq. (3).", + "bbox": [ + 511, + 267, + 772, + 282 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} \\mathbf {k} ^ {T} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right] \\left[ \\begin{array}{l l} \\mathbf {k} _ {i} ^ {T} & \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\sim \\left[ \\begin{array}{l l} \\mathrm {I 2 I} & \\mathrm {T 2 I} \\\\ \\mathrm {I 2 T} & \\mathrm {T 2 T} \\end{array} \\right], \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 513, + 290, + 915, + 339 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {q} \\mathbf {k} ^ {T} \\mathbf {v} = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right] = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 514, + 342, + 905, + 391 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the upper and lower rows in Eq. (3) correspond to the resultant image and text latents, respectively, this notation choice naturally reflects the underlying behavior of concatenated attention. Within this generalized framework, our empirical findings suggest that I2I and T2I blocks can serve roles similar to the conventional self- and cross-attention patterns in U-Net architectures.", + "bbox": [ + 511, + 401, + 906, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Block-wise attention patterns", + "text_level": 1, + "bbox": [ + 511, + 516, + 774, + 532 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "I2I & T2T. I2I and T2T are unimodal portions of the attention map. As demonstrated by PCA analysis in Fig. 3, the I2I block is analogous to self-attention in U-Net architectures, capturing spatial layout and geometric information. In contrast, T2T blocks predominantly manifest as identity matrices, with attention weights notably weakening after the occurrence of meaningful tokens (e.g., EOS).", + "bbox": [ + 511, + 537, + 905, + 643 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "T2I & I2T. T2I and I2T blocks represent cross-modal interactions, where T2I influences subsequent image latents, and I2T affects text representations - with I2T being a novel addition not present in U-Net architectures. These blocks are essential for precise local editing as they encode token-specific image-region correspondences, generating binary masks for precise local blending. While both blocks can be utilized to obtain these masks, we find T2I more effective (Fig. 4). Specifically, due to the row-wise softmax operation in I2T blocks, where the rows correspond to image tokens, there is inherent competition among these tokens since their attention weights must sum to 1. This constraint dilutes the attention signals, limiting their ability to represent strong or broad influences. In contrast, T2I's structure allows multiple image regions to retain high attention values simultaneously. Additionally, for the SD3 model series using both CLIP and T5 encoders, we observe that T5-derived", + "bbox": [ + 511, + 643, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "19494", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/644646cbf21ca17501b775f4f75f830ed2d9c76e3a905ac29a4f3db7d6da5c56.jpg", + "image_caption": [ + "Figure 4. Comparison of Flux.1-dev's T2I and I2T attention blocks averaged across all timesteps and transformer blocks. T2I blocks show stronger localization due to their unconstrained attention distribution compared to I2T's row-wise competition, making them more effective source for localized editing." + ], + "image_footnote": [], + "bbox": [ + 91, + 89, + 480, + 223 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/194ae3a0fdc66c05ced6a36e4a70dcc6a00539c4e502b06dfd0dd142de2d9629.jpg", + "image_caption": [ + "Figure 5. Analysis of T2I attention blocks from SDXL and various MM-DiT models using the prompt \"House on a mountainside under a clear blue daylight sky\" (averaged across timesteps and transformer blocks). MM-DiT models show better localized attention compared to U-Net architectures, though attention noise increases with model scale. When editing \"blue daylight\" to \"red dawn\" by replacing entire attention blocks, MM-DiT's precise attention enables clean edits, while SDXL applies the change globally." + ], + "image_footnote": [], + "bbox": [ + 94, + 311, + 483, + 513 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "attention maps typically exhibit higher precision than CLIP.", + "bbox": [ + 89, + 657, + 482, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Based on these analyses, we observe that I2I blocks are crucial for preserving identity, and T2I blocks are effective in obtaining attention masks for precise targeted local editing. We also validate the relative importance of these four sub-blocks in Fig. A2, which highlights the effectiveness of the I2I block in preserving original image attributes. Furthermore, visualizations of T2I attention in Fig. A14, Fig. A15, Fig. A16, Fig. A17 suggest that, during full attention's modality mixing, rotary positional embeddings (RoPE) [62] and residual connections help each modality retain its distinctive characteristics at designated position throughout transformer blocks. This supports and justifies our block-based interpretation and editing. Due to space constraints, we present extensive additional findings in Appendix B, to which we refer readers for further details.", + "bbox": [ + 88, + 674, + 482, + 900 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/20250207dc019223b5a20ab40312e60b66eae532ff043e14ef184da6063f5c7b.jpg", + "image_caption": [ + "Figure 6. Visualization of Flux.1-dev's T2I block, averaged across timesteps, comparing masks from top-5 selected blocks (upper) versus all blocks (lower), both with and without smoothing. Block selection significantly reduces noise, while smoothing refines boundary transitions. The lower left displays Grounded SAM2's masks used as ground truth for top-5 block selection." + ], + "image_footnote": [], + "bbox": [ + 517, + 90, + 903, + 205 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Mitigating noisy attention maps", + "text_level": 1, + "bbox": [ + 511, + 328, + 795, + 344 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We compare cross-attention maps (T2I blocks for MM-DiTs) across various models in Fig. 5. While MM-DiT architectures exhibit significantly improved attention localization, we observe an interesting phenomenon: as MM-DiT model size grows, the attention maps become increasingly noisy. Directly employing these noisy, discontinuous attention maps as local blending masks introduce visible artifacts in generated images. This issue aligns with recent findings by [12], which report similar noise artifacts emerging in ViTs at larger model and dataset scales. While they propose to add a register (additional tokens) to remove these artifacts, it is impractical here due to the prohibitive cost of retraining large MM-DiT models. Thus, we propose two practical solutions: (1) selecting optimal transformer blocks that inherently produce more precise and less noisy attention maps, and (2) applying Gaussian smoothing to smooth mask boundaries and reduce artifacts.", + "bbox": [ + 511, + 352, + 906, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For systematic evaluation, we sampled 100 prompts from the PARTI prompts [68], a widely used prompt set for benchmarking, and we created ground truth masks using Grounded SAM2 [34, 53] for grounded segmentation tasks, as shown in Fig. 6. Intuitively, effective blocks should yield attention maps similar to the GT masks. Thus, using standard losses and measures (BCE, Soft mIoU, and MSE) against the binary GT masks, we selected the top-5 blocks for each model based on the average ranking. Blockwise metrics for Flux.1 are presented in Fig. 7, and specific rankings are provided in Tab. A2. We emphasize that these blocks are not prompt-specific, as they were calculated over 100 random prompts, and we consistently use the same five blocks throughout the paper. Local editing results with and without these selected blocks are presented in Fig. 8.", + "bbox": [ + 511, + 611, + 908, + 838 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Notably, applying Gaussian smoothing alters the block ranking in larger models, suggesting that some noisy blocks still contain meaningful information. In contrast, smaller models like SD3-2B, which inherently produce smoother", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "19495", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/58356bc12b89ef55c1b356719efa76f9dcb10d3fef9d5a3ff56a850bbd621342.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 88, + 223, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/40552c483fa6f7e86fc1ecaaf03b3c70f1cdc48b875dbea7bb187e28c94e5186.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 88, + 352, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/58b6675bab5ca2f98899159bcaab018bb2921014ca9535304e77fe197195f5f8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 88, + 480, + 212 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a637d3b3b2454f55fc3bcecc7c79e2f153766ad464605424ec5e8008eb69ed48.jpg", + "image_caption": [ + "Figure 7. Transformer block analysis of Flux.1-dev using Binary Cross Entropy Loss, Soft mIoU, and MSE, with Grounded SAM2 predictions as ground truth. Scores are shown without (upper) and with (lower) Gaussian smoothing, with all evaluation metrics showing notable improvements after smoothing. T2I shows higher accuracy than I2T across most blocks. Top-5 ranked T2I attention blocks are exclusively used for generating local edit masks. Additional model results are provided in Appendix C.1." + ], + "image_footnote": [], + "bbox": [ + 91, + 214, + 225, + 337 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4231d97e06240f648447c1d7ff2e5f678232a815dc14af09e01a15b367556056.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 228, + 214, + 354, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/36b9fe7a4b019eb29f4780af12375ac1fe651749a470e62d133ab0cbbe5d5851.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 361, + 214, + 483, + 338 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/62dbd1448bd0e8b41691f43b06bda3d766816657d1b1ac44482e309732c88dd8.jpg", + "image_caption": [ + "Figure 8. Local image editing results using masks derived from the top-5 transformer blocks with Gaussian smoothing. Employing only the top-5 blocks reduces noise in the attention masks, effectively preserving untouched regions (e.g., music notes). Applying Gaussian smoothing further reduces boundary artifacts." + ], + "image_footnote": [], + "bbox": [ + 91, + 468, + 480, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "attention maps, exhibit stable block rankings regardless of Gaussian smoothing. Unlike previous U-Net observations, where down-block features are consistently noisier than upblock ones, we observed no strong common trends across MM-DiT variants due to their distinct architectural differences. Nonetheless, very early and later blocks consistently produced noisier attention maps across all models. Thus, simply excluding these noisy blocks and using the identified top-5 blocks consistently led to robust results. Another practical advantage of selecting only the top-5 blocks is improved computational efficiency. Explicitly computing the full T2I attention map prohibits using optimized attention computations such as PyTorch's dedicated SDPA kernel [10, 11, 31, 47]. By limiting manual full attention computations to just the top-5 blocks, the remaining blocks can leverage these optimized kernels, significantly enhancing inference efficiency, as further discussed in Sec. 4.2.", + "bbox": [ + 89, + 643, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/eb89e1019389df6054e71642085ee3dab8c5de12eea31302d664a3a35baffafc.jpg", + "image_caption": [ + "Figure 9. Architectural illustration of our method. We replace target (editing) branch projections $\\mathbf{q}_i^{\\mathrm{tgt}}$ and $\\mathbf{k}_i^{\\mathrm{tgt}}$ with their source branch counterparts, $\\mathbf{q}_i^{\\mathrm{src}}$ and $\\mathbf{k}_i^{\\mathrm{src}}$ . For local blending, we store the T2I portion of unmodified attention maps from selected blocks in both branches during early timesteps. A binary mask is computed from the union of these attention maps after thresholding and used to blend the two latent images. Best viewed zoomed in." + ], + "image_footnote": [], + "bbox": [ + 568, + 90, + 854, + 551 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Editing via MM-DiT Attention", + "text_level": 1, + "bbox": [ + 511, + 670, + 795, + 686 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Attention map-based and input projection-based methods", + "text_level": 1, + "bbox": [ + 511, + 696, + 905, + 724 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our analysis reveals that MM-DiT attention components serve roles similar to their U-Net counterparts: I2I blocks preserve source attributes (self-attention-like), while T2I blocks localize text-image interactions (cross-attention-like). Inspired by previous studies, we demonstrate image editing by substituting the target branch's attention map with the source branch's during early denoising timesteps. For common rectified flow schedulers (Euler's method), replacing attention maps during the initial $20\\%$ of timesteps yields effective results.", + "bbox": [ + 511, + 733, + 906, + 883 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A key consideration is now determining which attention", + "bbox": [ + 532, + 885, + 906, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "19496", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/af54c4eba972f5600a0df8148d2223dc6c0c517d4bd262515543bd5256f24f58.jpg", + "image_caption": [ + "Figure 10. Effect of text misalignment on image editing using Flux.1-dev model. When editing with highly different prompts (full list in the Appendix), replacing only image projections $(\\mathbf{q}_i, \\mathbf{k}_i)$ maintains better stability compared to full attention map replacement (all $\\mathbf{q}, \\mathbf{k}$ )." + ], + "image_footnote": [], + "bbox": [ + 93, + 90, + 482, + 244 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/5decd7089e2f8a518b07d8476782361f626cb14635084a0b35da774730b0f930.jpg", + "image_caption": [ + "Figure 11. Comparison between replacing $\\mathbf{q}_i$ and $\\mathbf{k}_i$ projections and I2I block replacement. Results are highly similar, with projection-based methods being more computationally efficient." + ], + "image_footnote": [], + "bbox": [ + 96, + 330, + 483, + 438 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "components to replace. While substituting the entire attention map (equivalently, full $\\mathbf{q}$ , $\\mathbf{k}$ projections) is functional, it introduces text projection misalignment. Specifically, replacing the entire target attention map causes the text region to shift to the source branch's contexts ( $\\mathbf{q}_t^{\\mathrm{src}} \\times \\mathbf{k}_t^{\\mathrm{src}}$ ). When this altered attention is multiplied by the target branch's value matrix text tokens ( $\\mathbf{v}_t^{\\mathrm{tgt}}$ ), misalignment arises. With T5 embeddings, subtle prompt differences can significantly amplify these misalignments, leading to undesirable image shifts (illustrated in Fig. 10). This establishes a critical constraint: T2T regions must remain unmodified. In-depth explanation of this issue can be found in Appendix B.4.", + "bbox": [ + 88, + 507, + 482, + 688 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As a solution, we propose two alternative approaches: either selectively replacing only the I2I attention block, or specifically modifying just the image input projections $(\\mathbf{q}_i,$ $\\mathbf{k}_i)$ . Based on Eq. (2), these approaches slightly differ, as modifying $\\mathbf{q}_i$ and $\\mathbf{k}_i$ also influences T2I and I2T interactions. However, as shown in Fig. 11, results from both methods are nearly identical. Importantly, modifying at the input projection level $(\\mathbf{q}_i,\\mathbf{k}_i)$ is computationally more efficient since it enables optimized scaled dot product attention (SDPA) kernels. In contrast, replacing the I2I block disables SDPA, requiring manual attention calculation and careful switching before softmax to maintain proper attention distributions. This absence of SDPA can increase the computational time by up to $3\\times$ (Tab. 1).", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Algorithm 1: Prompt-based Image Editing with MM-DiT", + "text_level": 1, + "bbox": [ + 522, + 94, + 849, + 107 + ], + "page_idx": 5 + }, + { + "type": "code", + "sub_type": "algorithm", + "code_caption": [], + "code_body": "Parameters: Source prompt $P^{\\mathrm{src}}$ , Target prompt $P^{\\mathrm{tgt}}$ , Attention replace step $\\tau$ , Local Blending-{replace step $\\eta$ , threshold $\\theta$ } \n $x_{T}^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)$ ; $x_{T}^{\\mathrm{tgt}} \\gets x_{T}^{\\mathrm{src}}$ \nfor $t = T, T - 1, \\ldots, 1$ do \nif local blend and $t > \\eta$ then // Local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}, \\text{AtttnMap}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{src}} \\text{ if } t > \\tau\\right] \\\\\n\\text{Mask} \\gets (\\text{AtttnMap} > \\theta) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets x_{t-1}^{\\mathrm{tgt}} \\cdot \\text{Mask} + x_{t-1}^{\\mathrm{src}} \\cdot (1 - \\text{Mask})\n\\end{array}\n\\] \nelse // No local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{sucr}} \\text{ if } t > \\tau\\right]\n\\end{array}\n\\] \nreturn (Decode( $x_0^{\\mathrm{src}}$ ), Decode( $x_0^{\\mathrm{tgt}}$ ))", + "bbox": [ + 519, + 109, + 880, + 272 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To this end, we propose a new approach based on $\\mathbf{q}_i$ , $\\mathbf{k}_i$ replacement. Unlike prior U-Net-based attention modification methods that require a precise token mapping between source and target prompts, limiting their effectiveness when prompts differ significantly, our method can handle arbitrarily different prompt pairs. This approach effectively realizes our editing objective: maintaining base image similarity while incorporating edited prompt elements, a capability uniquely enabled by MM-DiT's full-attention architecture. By preserving the text region intact, our method eliminates the need for explicit token correspondence or specialized mappers between prompts, allowing robust performance across diverse editing scenarios.", + "bbox": [ + 511, + 305, + 906, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Local blending", + "text_level": 1, + "bbox": [ + 511, + 513, + 665, + 530 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Using insights from Sec. 3.3, we extract clear attention maps from the top-5 transformer blocks for precise local blending, ensuring only targeted regions change while preserving the base image elsewhere. We apply binary blending masks derived from these attention maps once per timestep after all blocks are processed, following P2P's method. This selective calculation allows optimized SDPA kernels in remaining blocks, maintaining inference speed similar to naive batch size-2 inference (SD3-M: 15.2s vs. 14.9s, Flux.1-dev: 55.9s vs. 53.7s, SD3.5-L: 50.1s vs. 47.5s on a single A6000 GPU).", + "bbox": [ + 511, + 537, + 906, + 702 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our complete pipeline is illustrated in Fig. 9, with detailed pseudocode provided in Algorithm 1. Among the three hyperparameters controlling our method, such as masking threshold $\\theta$ , attention replace timestep $\\tau$ , and local blending stop timestep $\\eta$ , we find that setting $\\tau$ to $20\\%$ and $\\eta$ to $50\\%$ (i.e., $\\tau = 0.8T$ , $\\eta = 0.5T$ ) of total timesteps typically achieves robust performance. Thus, the masking threshold $\\theta$ remains a primary parameter influencing edit granularity. As demonstrated in Fig. 12, higher thresholds effectively enable precise local edits, while lower thresholds allow broader modifications. Leveraging MM-DiT's precise attention localization and our attention map curation, the proposed local blending method significantly im", + "bbox": [ + 511, + 704, + 908, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "19497", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/9d1d616f17bc0fed6e7f584e6963029c7370686c208a3454f9b12d5310428db1.jpg", + "image_caption": [ + "Figure 12. Ablation results on local blending threshold. Fine edits like text modifications benefit from higher thresholds, whereas broader transformations like large object changes favor lower thresholds. Local blending was applied to the first $50\\%$ of timestep iterations." + ], + "image_footnote": [], + "bbox": [ + 138, + 92, + 859, + 305 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "proves edit precision compared to results obtained without attention curation (Fig. A7).", + "bbox": [ + 89, + 359, + 482, + 390 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Controlling edit strength via block selection", + "text_level": 1, + "bbox": [ + 89, + 398, + 460, + 416 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While injecting input projections or attention maps into all blocks is straightforward, we observe that later blocks have a stronger influence on the final output, producing edited images closely resembling the source. While this correlation aligns with our goal, it can pose issues for few-step models like Flux.1-schnell, where editing all blocks in a single step of a 4-step inference may yield outputs overly similar to the base image, regardless of the target prompt. To address this, we replace only the first 38 and 30 blocks for Flux.1-schnell and SD3.5-L-Turbo, respectively, while replacing every block for non-distilled models. Note that our two block-selection strategies (top-5 blocks for obtaining attention maps and blocks for $\\mathbf{q}_i$ , $\\mathbf{k}_i$ projections replacement) are not considered as hyperparameters and remain fixed. More details and ablations are available in Appendix C.2 and Fig. A8.", + "bbox": [ + 88, + 421, + 482, + 664 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5. Editing Real Images", + "text_level": 1, + "bbox": [ + 89, + 678, + 284, + 695 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "A common approach in existing real image editing methods [2, 9, 56, 63] is to invert the source image into an initial noisy latent and then generate the target image from this inverted latent using an editing prompt. Our editing technique, which operates via $\\mathbf{q}_i$ , $\\mathbf{k}_i$ replacement, functions in parallel with existing methods and can be readily applied in this scenario. Fig. 13 illustrates this process, and a preview of our results can be seen in Fig. 14. In Appendix A, we present further discussions, including quantitative and qualitative comparisons, using RF inversion as a baseline. Generally, the performance of real image editing methods depends heavily on inversion quality, as the initial latent affects background consistency and overall image quality.", + "bbox": [ + 89, + 704, + 483, + 901 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/fd726ff8904cf8f96cac1380fa8dfd6825a2c89e13db6ff08cea4fab8f5433d4.jpg", + "image_caption": [ + "Figure 13. Illustration of our real image editing method given an initial latent. Unlike the synthetic image editing case, the source branch is fixed as a straight conditional path from the initial noise to the source image, independent of model evaluation." + ], + "image_footnote": [], + "bbox": [ + 524, + 359, + 893, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "While this dependency also exists for our method, the impact of the initial latent is partially mitigated by continuously injecting input projections from the noisy source image and performing local blending to preserve unaffected regions. Consequently, our method can be used without inversion, starting from random noise, offering a trade-off between computational efficiency and performance.", + "bbox": [ + 511, + 604, + 905, + 710 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "6. Experiments", + "text_level": 1, + "bbox": [ + 511, + 724, + 645, + 741 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Due to the lack of dedicated architecture-level editing methods for MM-DiT, we validate our approach against two simple baselines: (1) direct generation with fixed seeds, and (2) prompt switching after $20\\%$ of timesteps, similar to SDEdit [41]. We randomly sample 60 diverse prompts from the PARTI dataset and use LLM [1] to generate edited prompts consisting of 30 involving simple object or word edits and 30 involving complex changes affecting multiple prompt components. We evaluate editing quality using LPIPS [69] and CLIPScore [20]. As these metrics have", + "bbox": [ + 509, + 750, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "19498", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7ebb9d1f4aa687fbfb9f7f7688896c4094b72cf0c0e19153296b8e37fc6d5370.jpg", + "image_caption": [ + "A Golden Retriever sitting on an autumn grass field" + ], + "image_footnote": [], + "bbox": [ + 93, + 89, + 483, + 433 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 2. Quantitative comparison with different baselines. While higher CLIP scores indicate better target prompt alignment, lower LPIPS scores do not necessarily reflect edit quality since minimal edits yield the lowest scores. Our method well balances source content preservation and target prompt matching, whereas fixed seed generation often produces entirely different images and prompt-change produces overly conservative modifications.", + "bbox": [ + 89, + 518, + 483, + 617 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1de2a55d7efbdbdf4218e47fc964679cff10a837a20f0eca949384bd6ae85524.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
ModelFixed SeedPrompt ChangeOurs
LPIPSCLIPLPIPSCLIPLPIPSCLIP
SD3-M0.5940.3770.3250.3440.3800.359
SD3.5-M0.5840.3800.3220.3490.4560.374
SD3.5-L0.5570.3880.3060.3630.4180.377
SD3.5-L-Turbo0.5090.3730.2250.3320.3880.363
Flux.1-dev0.5790.3590.2510.3110.3690.339
Flux.1-schnell0.5080.3800.1160.2890.2790.333
", + "bbox": [ + 91, + 628, + 485, + 747 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "inherent limitations in capturing various types of prompt-based editing quality, we also provide user study result in Fig. A9 and extensive qualitative results in Fig. 15, Fig. A11, and Fig. A12. All experiments use 28 timesteps, except for distilled models (SD3.5-L-Turbo and Flux.1-schnell), which use 4. To ensure fair comparisons and avoid hyperparameter tuning, we fix the attention replacement step to the first $20\\%$ of iterations and disable local blending for the results in Tab. 2 and Fig. 15, though our method", + "bbox": [ + 89, + 763, + 483, + 900 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/cd13a724b0f8fc7dc3a0d25833f4428f2d8a8760e444c4466f86b2fc61ebf099.jpg", + "image_caption": [ + "Figure 14. Real image editing comparison between our method and RF inversion. Our approach enables more precise local edits while preserving unchanged regions.", + "Figure 15. Qualitative comparison of our method with baselines. Our method consistently outperforms baseline approaches across diverse editing scenarios." + ], + "image_footnote": [], + "bbox": [ + 514, + 87, + 906, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "can incorporate blending when needed.", + "bbox": [ + 511, + 556, + 772, + 571 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "For real-image editing, we observe that editing performance depends more heavily on the inversion technique than the architectural-level editing method. Nonetheless, we report quantitative results on PIE-Bench [27] using RF inversion as our inversion technique and provide extensive qualitative results. Please refer to Appendix A.3 and A.4 for details. Our method provides additional controllability and improves the naive RF inversion baseline.", + "bbox": [ + 511, + 571, + 906, + 691 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 708, + 633, + 723 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work presents a comprehensive analysis of MM-DiT's attention mechanism, revealing several underlying insights, including scaling properties and text-image interactions. Based on these observations, we develop an efficient editing method tailored for MM-DiT architectures that achieves high-quality results across various scenarios, including few-step models and real image editing. We believe these findings help deepen the understanding of emerging architectures and contribute to advancements in controlled image generation and editing. Discussions on limitations and future directions can be found in Appendix F.", + "bbox": [ + 511, + 734, + 906, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "19499", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "8. Acknowledgments", + "text_level": 1, + "bbox": [ + 89, + 90, + 269, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work was supported by IITP grant (No. RS-2021-II211343, Artificial Intelligence Graduate School Program at Seoul National University) $(5\\%)$ , IITP grant (No. RS-2024-00509257, Global AI Frontier Lab) $(65\\%)$ , and NRF grant (No. RS-2024-00405857) $(30\\%)$ , funded by the Korea government (MSIT).", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Anthropic. Claude 3.5 Sonnet, 2023. Large language model developed by Anthropic. 7", + "[2] Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. Stable flow: Vital layers for training-free image editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 7, 9", + "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4", + "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 8", + "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2", + "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2", + "[7] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-alpha: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1", + "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-delta: Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 2", + "[9] Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. FluxSpace: Disentangled Semantic Editing in Rectified Flow Transformers. arXiv preprint arXiv:2412.09611, 2024. 7, 8", + "[10] Tri Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. In International Conference on Learning Representations (ICLR), 2024. 5", + "[11] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + ], + "bbox": [ + 93, + 246, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision Transformers Need Registers. In International Conference on Learning Representations (ICLR), 2024. 2, 4", + "[13] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing. arXiv preprint arXiv:2412.07517, 2024. 2, 8", + "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2", + "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 1, 2", + "[16] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. ReNoise: Real Image Inversion Through Iterative Noising. arXiv preprint arXiv:2403.14602, 2024. 2", + "[17] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4", + "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In International Conference on Learning Representations (ICLR), 2022. 2, 4", + "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. 7", + "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1", + "[22] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In IEEE International Conference on Computer Vision (ICCV), 2023. 2", + "[23] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 2", + "[25] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual Style Prompting with Swapping Self-Attention. arXiv preprint arXiv:2402.12974, 2024. 2" + ], + "bbox": [ + 516, + 92, + 905, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "19500", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[26] Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang, and Renjie Liao. Uniedit-flow: Unleashing inversion and editing in the era of flow models. arXiv preprint arXiv:2504.13109, 2025. 2", + "[27] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code. In International Conference on Learning Representations (ICLR), 2024. 8, 2, 4", + "[28] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling Diffusion Models into Conditional GANs. In European Conference on Computer Vision (ECCV), 2024. 2", + "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1", + "[30] Black Forest Labs. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.Accessed:November 2024.1", + "[31] Benjamin Lefaudeaux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xFormers: A modular and hackable Transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5", + "[32] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow Matching for Generative Modeling. In International Conference on Learning Representations (ICLR), 2023. 2", + "[33] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4", + "[35] Xingchao Liu, Chengyue Gong, et al. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In International Conference on Learning Representations (ICLR), 2023. 2, 1", + "[36] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In International Conference on Learning Representations (ICLR), 2024. 2", + "[37] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 2", + "[38] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[39] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 1", + "[40] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2", + "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 7, 2", + "[42] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2", + "[43] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), 2021. 1", + "[45] OpenAI. GPT-4o, 2024. Large language model developed by OpenAI. 3", + "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 4", + "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 5", + "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2", + "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1, 2", + "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2", + "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 2020. 2", + "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "19501", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[53] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4", + "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2", + "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 2", + "[56] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations. arXiv preprint arXiv:2410.10792, 2024. 2, 7, 1, 4", + "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1", + "[58] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In International Conference on Learning Representations (ICLR), 2022. 2", + "[59] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial Diffusion Distillation. arXiv e-prints, pages arXiv-2311, 2023. 2", + "[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 1", + "[61] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. In International Conference on Machine Learning (ICML), 2023. 2", + "[62] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4", + "[63] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 7, 8", + "[64] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2", + "[65] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. TurboEdit: Instant text-based" + ], + "bbox": [ + 91, + 90, + 480, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "image editing. In European Conference on Computer Vision (ECCV), 2024. 2", + "[66] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. arXiv preprint arXiv:2405.14867, 2024. 2", + "[67] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[68] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. Transactions on Machine Learning Research, 2022. 4", + "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7", + "[70] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation. arXiv preprint arXiv:2405.01434, 2024. 2" + ], + "bbox": [ + 516, + 92, + 903, + 430 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "19502", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_model.json b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bf81834ae7bcc7e2ddbee7a2dc31d6eb14c01872 --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_model.json @@ -0,0 +1,2367 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.247, + 0.131, + 0.753, + 0.177 + ], + "angle": 0, + "content": "Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing" + }, + { + "type": "text", + "bbox": [ + 0.192, + 0.204, + 0.807, + 0.24 + ], + "angle": 0, + "content": "Joonghyuk Shin Alchan Hwang Yujin Kim Daneul Kim Jaesik Park Seoul National University" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.242, + 0.775, + 0.258 + ], + "angle": 0, + "content": "{joonghyuk, alchan00, yujin.k, carpedkm, jaesik.park}@snu.ac.kr" + }, + { + "type": "image_caption", + "bbox": [ + 0.194, + 0.296, + 0.254, + 0.306 + ], + "angle": 0, + "content": "\\(\\mathrm{SD3}\\rightarrow \\mathrm{Flux}\\)" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.307, + 0.353, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.099, + 0.407, + 0.236, + 0.416 + ], + "angle": 0, + "content": "'help me' \\(\\rightarrow\\) 'Please don't kill me'" + }, + { + "type": "image_caption", + "bbox": [ + 0.461, + 0.297, + 0.54, + 0.307 + ], + "angle": 0, + "content": "giraffe \\(\\rightarrow\\) dragon" + }, + { + "type": "image", + "bbox": [ + 0.375, + 0.307, + 0.628, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.426, + 0.407, + 0.487, + 0.417 + ], + "angle": 0, + "content": "pig \\(\\rightarrow\\) whale" + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.297, + 0.814, + 0.306 + ], + "angle": 0, + "content": "frogs \\(\\rightarrow\\) rabbits" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.306, + 0.905, + 0.406 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.783, + 0.407, + 0.842, + 0.416 + ], + "angle": 0, + "content": "ferret \\(\\rightarrow\\) lion" + }, + { + "type": "image", + "bbox": [ + 0.098, + 0.417, + 0.24, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.244, + 0.417, + 0.386, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.388, + 0.417, + 0.529, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.531, + 0.417, + 0.718, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.72, + 0.417, + 0.905, + 0.543 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.553, + 0.906, + 0.569 + ], + "angle": 0, + "content": "Figure 1. Our prompt-based editing examples on SD3 series and Flux.1, demonstrating high-quality results across various editing scenarios." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.581, + 0.327, + 0.597 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.088, + 0.614, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Transformer-based diffusion models have recently superseded traditional U-Net architectures, with multimodal diffusion transformers (MM-DiT) emerging as the dominant approach in state-of-the-art models like Stable Diffusion 3 and Flux.1. Previous approaches have relied on unidirectional cross-attention mechanisms, with information flowing from text embeddings to image latents. In contrast, MM-DiT introduces a unified attention mechanism that concatenates input projections from both modalities and performs a single full attention operation, allowing bidirectional information flow between text and image branches. This architectural shift presents significant challenges for existing editing techniques. In this paper, we systematically analyze MM-DiT's attention mechanism by decomposing attention matrices into four distinct blocks, revealing their inherent characteristics. Through these analyses, we propose a robust, prompt-based image editing method for MM-DiT that supports global to local edits across various MM-DiT variants, including few-step models. We believe our findings" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.582, + 0.909, + 0.629 + ], + "angle": 0, + "content": "bridge the gap between existing U-Net-based methods and emerging architectures, offering deeper insights into MM-DiT's behavioral patterns." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.661, + 0.646, + 0.677 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.688, + 0.907, + 0.839 + ], + "angle": 0, + "content": "Recent years have witnessed the emergence of diffusion models [21, 60] as the de-facto approach for visual generation, demonstrating remarkable capabilities in creating high-quality data from noise [15, 29, 39, 44, 49, 52, 54, 57]. While early implementations predominantly employed U-shaped networks [55], transformer-based architectures [7, 48] have recently emerged as the prevailing approach, with multimodal diffusion transformers (MM-DiT) showing particular promise in state-of-the-art models like Stable Diffusion 3 series [15] and Flux.1 [30]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Motivated by MM-DiT's architectural differences and the challenges in directly applying existing U-Net-based editing methods, our work addresses a fundamental question: \"How can we leverage our understanding of U-Net" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19492" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.196 + ], + "angle": 0, + "content": "attention mechanisms to interpret MM-DiT's behavior better and effectively transfer existing techniques to this new architecture?\" Given the significance of existing work built upon U-Net attention mechanisms [4, 5, 19, 33], answering this question is crucial for enabling existing methods to adapt to, and benefit from, emerging MM-DiT architectures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.199, + 0.482, + 0.395 + ], + "angle": 0, + "content": "To this end, we systematically investigate MM-DiT's attention mechanisms through block matrix analysis, revealing key characteristics of its attention maps. As one of our key analytical insights, we observe an interesting phenomenon aligned with vision transformer [14] scaling laws [12]: as the model size increases, attention patterns become properly positioned but increasingly noisy, necessitating careful selection of attention layers for editing tasks. Building upon these analyses, we propose an efficient editing method at the architecture level that focuses on modifying image tokens in input projections. The proposed editing method can be jointly combined with inversion methods [56, 63] for editing real images as well." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.397, + 0.482, + 0.532 + ], + "angle": 0, + "content": "Our contributions are primarily architectural: (1) We provide the first systematic analysis of MM-DiT attention mechanisms, highlighting block-wise patterns and scaling behaviors. (2) Building upon these findings, we propose an efficient and robust editing method across diverse editing scenarios - from global stylistic changes to local edits such as text modification - and various MM-DiT variants, including few-step models (SD3-M, SD3.5-M, SD3.5-L, SD3.5-L-Turbo, and Flux.1-dev/schnell)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.549, + 0.233, + 0.565 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.575, + 0.364, + 0.59 + ], + "angle": 0, + "content": "2.1. Text-to-image diffusion models" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.482, + 0.689 + ], + "angle": 0, + "content": "Early text-to-image diffusion models mostly used U-shaped networks with residual blocks [55] incorporating pixelwise self-attention and text-image cross-attention layers, combined with downsampling and upsampling operations (Fig. 2). These models are typically trained with DDPM-style scheduling and noise prediction objective [49, 54]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Recent advances introduced two significant shifts. First, architecturally, there has been a transition from U-Nets to transformer-based architectures like DiT [48], demonstrating superior scaling properties. The PixArt-alpha series [6-8] successfully adapted DiT for text conditional generation, maintaining a conventional cross-attention mechanism for text conditioning. Stable Diffusion 3 [15] subsequently introduced a pivotal shift with multimodal diffusion transformers (MM-DiT), using separate transformers for text and image modalities while concatenating their sequences for unified attention operations (Fig. 2). Second, formulation-wise, newer approaches have adopted v-prediction [58] with a rectified flow framework [32, 35] that connects data and noise distributions along straight paths, replacing the com" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.173 + ], + "angle": 0, + "content": "Table 1. Architectural details and inference times for \\(1024 \\times 1024\\) image generation using 28 timesteps, using single A6000 GPU. Times are shown as optimized (left, using PyTorch SDPA kernels) vs. naive matrix multiplication (right). Text encoders: \\(\\mathrm{CLIP}^1\\) refers to CLIP ViT-L/14 [50], \\(\\mathrm{CLIP}^2\\) refers to OpenCLIP ViTG/14 [24], and T5 refers to T5-XXL [51]." + }, + { + "type": "table", + "bbox": [ + 0.514, + 0.182, + 0.92, + 0.283 + ], + "angle": 0, + "content": "
Arch. / ParamsText Enc.Attn TypeTime (s)
SD1U-Net / 860MCLIP1Self/Cross5.0/12.7
SDXLU-Net / 2.6BCLIP1,2Self/Cross5.6/7.9
SD3-MMM-DiT / 2BCLIP1,2 / T5Full7.1/20.5
SD3.5-MMM-DiT-X / 2.2BCLIP1,2 / T5Full/Self9.6/26.8
SD3.5-LMM-DiT / 8BCLIP1,2 / T5Full24.7/60.5
Flux.1-devMM-DiT* / 11.9BCLIP1 / T5Full27.8/36.8
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.403 + ], + "angle": 0, + "content": "plex trajectories of traditional diffusion models. Together with increased model scales, these developments have significantly improved model performance. Parallel efforts have also accelerated iterative diffusion sampling via faster ODE solvers [37, 38], consistency distillation [40, 61], adversarial learning [28, 59, 66, 67], and trajectory straightening [32, 35, 36]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.41, + 0.75, + 0.427 + ], + "angle": 0, + "content": "2.2. Text-guided image editing" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.432, + 0.905, + 0.614 + ], + "angle": 0, + "content": "Text-guided image editing naturally extends text-to-image diffusion models. Prior work [18, 22, 33, 64] shows that self-attention governs spatial structure while cross-attention handles attribute-level semantics aligned with text. Prompt-to-Prompt (P2P) [19], a seminal approach, leverages this by directly transferring cross-attention maps for precise edits. Subsequent methods further manipulate self-attention mechanisms for better content-preserving edits [4, 25, 43, 70]. A key advantage of these attention-based methods is their ability to achieve precise local edits solely from textual prompts, eliminating the need for explicit spatial guidance or additional user input." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.905, + 0.735 + ], + "angle": 0, + "content": "Another direction employs inversion techniques [13, 16, 23, 26, 56, 63], mapping real images to editable latent spaces. While inversion alone can lose fine details, combining it with attention-based methods significantly enhances localized editing [42]. Alternatively, methods like InstructPix2Pix [3] train dedicated models on synthetic data, though this approach incurs high computational costs and dataset preparation overheads [65]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.743, + 0.688, + 0.758 + ], + "angle": 0, + "content": "2.3. MM-DiT variants" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Since its introduction, MM-DiT architecture has seen several improvements. The initial dual-branch design in SD3 was motivated by the conceptual differences between text and image embeddings. This architecture was maintained in both SD3 and SD3.5-Large models, which contain 24 and 38 dual-branch blocks, respectively. Flux.1, scaled to 12B parameters, introduced a hybrid approach with 57 total blocks - 19 dual branch blocks followed by 38 single branch blocks (Fig. 2). The single branch variant processes" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19493" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.091, + 0.087, + 0.468, + 0.321 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.335, + 0.485, + 0.406 + ], + "angle": 0, + "content": "Figure 2. Attention mechanisms across diffusion architectures. (a) U-Net with separate self-/cross- attention, (b) MM-DiT blocks with dual branches, (c) MM-DiT* single branch block used in later layers of Flux.1, and (d) MM-DiT-X dual attention block with explicit self-attention, used in early layers of SD3.5-M." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.425, + 0.484, + 0.533 + ], + "angle": 0, + "content": "concatenated text and image embeddings as a single modality using unified weights (e.g., linear layers, attention parameters), reducing block parameters from 340M to 141M. The MM-DiT-X variant, later introduced in SD3.5-medium, added explicit self-attention operations exclusively for image tokens in the first 13 transformer blocks for better multi-resolution results and overall image coherence." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.552, + 0.377, + 0.57 + ], + "angle": 0, + "content": "3. Attention Analysis in MM-DiT" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.579, + 0.484, + 0.596 + ], + "angle": 0, + "content": "3.1. Block matrix formulation of MM-DiT attention" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.604, + 0.484, + 0.772 + ], + "angle": 0, + "content": "We begin the analysis with MM-DiT's attention structure under SD3-M's standard dimensions: \\(64 \\times 64\\) image latents (flattened to 4096) and 333 text tokens (77 CLIP + 256 T5), with inner dimension 64. Omitting the batch and head dimensions for clarity, the input projections for image and text domains are denoted as \\((\\mathbf{q}_i, \\mathbf{k}_i, \\mathbf{v}_i \\in \\mathbb{R}^{4096 \\times 64})\\) and \\((\\mathbf{q}_t, \\mathbf{k}_t, \\mathbf{v}_t \\in \\mathbb{R}^{333 \\times 64})\\). Then, query, key, and value projections are concatenated as shown below (\\(\\mathbf{q}, \\mathbf{k}, \\mathbf{v} \\in \\mathbb{R}^{4429 \\times 64}\\)). It is worth noting that Flux.1 adopts an inverse concatenation order (text preceding image), which results in a different attention map." + }, + { + "type": "equation", + "bbox": [ + 0.173, + 0.785, + 0.484, + 0.822 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right], \\mathbf {k} = \\left[ \\begin{array}{l} \\mathbf {k} _ {i} \\\\ \\mathbf {k} _ {t} \\end{array} \\right], \\mathbf {v} = \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.835, + 0.486, + 0.903 + ], + "angle": 0, + "content": "Using the standard scaled dot product attention formulation \\( \\text{softmax}\\left(\\frac{\\mathbf{q}\\mathbf{k}^T}{\\sqrt{d_k}}\\right)\\mathbf{v} \\), we can write the attention map \\( \\mathbf{q}\\mathbf{k}^T \\) and its output representation as follows. Note that row-wise softmax normalization is applied to \\( \\mathbf{q}\\mathbf{k}^T \\) in Eq. (2), before" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.091, + 0.906, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.197, + 0.907, + 0.254 + ], + "angle": 0, + "content": "Figure 3. Visualizing the top-6 principal components extracted from I2I attention blocks, averaged across all timesteps and transformer blocks. Like U-Net's self-attention layers, these components reveal underlying spatial and geometric patterns." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.268, + 0.774, + 0.284 + ], + "angle": 0, + "content": "multiplication with \\(\\mathbf{v}\\) to obtain Eq. (3)." + }, + { + "type": "equation", + "bbox": [ + 0.514, + 0.291, + 0.916, + 0.34 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} \\mathbf {k} ^ {T} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right] \\left[ \\begin{array}{l l} \\mathbf {k} _ {i} ^ {T} & \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\sim \\left[ \\begin{array}{l l} \\mathrm {I 2 I} & \\mathrm {T 2 I} \\\\ \\mathrm {I 2 T} & \\mathrm {T 2 T} \\end{array} \\right], \\tag {2}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.515, + 0.343, + 0.906, + 0.392 + ], + "angle": 0, + "content": "\\[\n\\mathbf {q} \\mathbf {k} ^ {T} \\mathbf {v} = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right] = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.402, + 0.907, + 0.508 + ], + "angle": 0, + "content": "Since the upper and lower rows in Eq. (3) correspond to the resultant image and text latents, respectively, this notation choice naturally reflects the underlying behavior of concatenated attention. Within this generalized framework, our empirical findings suggest that I2I and T2I blocks can serve roles similar to the conventional self- and cross-attention patterns in U-Net architectures." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.517, + 0.776, + 0.533 + ], + "angle": 0, + "content": "3.2. Block-wise attention patterns" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.539, + 0.906, + 0.644 + ], + "angle": 0, + "content": "I2I & T2T. I2I and T2T are unimodal portions of the attention map. As demonstrated by PCA analysis in Fig. 3, the I2I block is analogous to self-attention in U-Net architectures, capturing spatial layout and geometric information. In contrast, T2T blocks predominantly manifest as identity matrices, with attention weights notably weakening after the occurrence of meaningful tokens (e.g., EOS)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.901 + ], + "angle": 0, + "content": "T2I & I2T. T2I and I2T blocks represent cross-modal interactions, where T2I influences subsequent image latents, and I2T affects text representations - with I2T being a novel addition not present in U-Net architectures. These blocks are essential for precise local editing as they encode token-specific image-region correspondences, generating binary masks for precise local blending. While both blocks can be utilized to obtain these masks, we find T2I more effective (Fig. 4). Specifically, due to the row-wise softmax operation in I2T blocks, where the rows correspond to image tokens, there is inherent competition among these tokens since their attention weights must sum to 1. This constraint dilutes the attention signals, limiting their ability to represent strong or broad influences. In contrast, T2I's structure allows multiple image regions to retain high attention values simultaneously. Additionally, for the SD3 model series using both CLIP and T5 encoders, we observe that T5-derived" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "19494" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.092, + 0.09, + 0.482, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.233, + 0.483, + 0.304 + ], + "angle": 0, + "content": "Figure 4. Comparison of Flux.1-dev's T2I and I2T attention blocks averaged across all timesteps and transformer blocks. T2I blocks show stronger localization due to their unconstrained attention distribution compared to I2T's row-wise competition, making them more effective source for localized editing." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.312, + 0.485, + 0.515 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.527, + 0.484, + 0.64 + ], + "angle": 0, + "content": "Figure 5. Analysis of T2I attention blocks from SDXL and various MM-DiT models using the prompt \"House on a mountainside under a clear blue daylight sky\" (averaged across timesteps and transformer blocks). MM-DiT models show better localized attention compared to U-Net architectures, though attention noise increases with model scale. When editing \"blue daylight\" to \"red dawn\" by replacing entire attention blocks, MM-DiT's precise attention enables clean edits, while SDXL applies the change globally." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.658, + 0.483, + 0.673 + ], + "angle": 0, + "content": "attention maps typically exhibit higher precision than CLIP." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Based on these analyses, we observe that I2I blocks are crucial for preserving identity, and T2I blocks are effective in obtaining attention masks for precise targeted local editing. We also validate the relative importance of these four sub-blocks in Fig. A2, which highlights the effectiveness of the I2I block in preserving original image attributes. Furthermore, visualizations of T2I attention in Fig. A14, Fig. A15, Fig. A16, Fig. A17 suggest that, during full attention's modality mixing, rotary positional embeddings (RoPE) [62] and residual connections help each modality retain its distinctive characteristics at designated position throughout transformer blocks. This supports and justifies our block-based interpretation and editing. Due to space constraints, we present extensive additional findings in Appendix B, to which we refer readers for further details." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.092, + 0.905, + 0.207 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.221, + 0.908, + 0.306 + ], + "angle": 0, + "content": "Figure 6. Visualization of Flux.1-dev's T2I block, averaged across timesteps, comparing masks from top-5 selected blocks (upper) versus all blocks (lower), both with and without smoothing. Block selection significantly reduces noise, while smoothing refines boundary transitions. The lower left displays Grounded SAM2's masks used as ground truth for top-5 block selection." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.329, + 0.796, + 0.345 + ], + "angle": 0, + "content": "3.3. Mitigating noisy attention maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.907, + 0.609 + ], + "angle": 0, + "content": "We compare cross-attention maps (T2I blocks for MM-DiTs) across various models in Fig. 5. While MM-DiT architectures exhibit significantly improved attention localization, we observe an interesting phenomenon: as MM-DiT model size grows, the attention maps become increasingly noisy. Directly employing these noisy, discontinuous attention maps as local blending masks introduce visible artifacts in generated images. This issue aligns with recent findings by [12], which report similar noise artifacts emerging in ViTs at larger model and dataset scales. While they propose to add a register (additional tokens) to remove these artifacts, it is impractical here due to the prohibitive cost of retraining large MM-DiT models. Thus, we propose two practical solutions: (1) selecting optimal transformer blocks that inherently produce more precise and less noisy attention maps, and (2) applying Gaussian smoothing to smooth mask boundaries and reduce artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.909, + 0.839 + ], + "angle": 0, + "content": "For systematic evaluation, we sampled 100 prompts from the PARTI prompts [68], a widely used prompt set for benchmarking, and we created ground truth masks using Grounded SAM2 [34, 53] for grounded segmentation tasks, as shown in Fig. 6. Intuitively, effective blocks should yield attention maps similar to the GT masks. Thus, using standard losses and measures (BCE, Soft mIoU, and MSE) against the binary GT masks, we selected the top-5 blocks for each model based on the average ranking. Blockwise metrics for Flux.1 are presented in Fig. 7, and specific rankings are provided in Tab. A2. We emphasize that these blocks are not prompt-specific, as they were calculated over 100 random prompts, and we consistently use the same five blocks throughout the paper. Local editing results with and without these selected blocks are presented in Fig. 8." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Notably, applying Gaussian smoothing alters the block ranking in larger models, suggesting that some noisy blocks still contain meaningful information. In contrast, smaller models like SD3-2B, which inherently produce smoother" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19495" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.225, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.089, + 0.353, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.089, + 0.482, + 0.213 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.092, + 0.215, + 0.226, + 0.338 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.229, + 0.215, + 0.355, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.362, + 0.215, + 0.484, + 0.339 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.349, + 0.484, + 0.461 + ], + "angle": 0, + "content": "Figure 7. Transformer block analysis of Flux.1-dev using Binary Cross Entropy Loss, Soft mIoU, and MSE, with Grounded SAM2 predictions as ground truth. Scores are shown without (upper) and with (lower) Gaussian smoothing, with all evaluation metrics showing notable improvements after smoothing. T2I shows higher accuracy than I2T across most blocks. Top-5 ranked T2I attention blocks are exclusively used for generating local edit masks. Additional model results are provided in Appendix C.1." + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.549 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.561, + 0.484, + 0.631 + ], + "angle": 0, + "content": "Figure 8. Local image editing results using masks derived from the top-5 transformer blocks with Gaussian smoothing. Employing only the top-5 blocks reduces noise in the attention masks, effectively preserving untouched regions (e.g., music notes). Applying Gaussian smoothing further reduces boundary artifacts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.484, + 0.901 + ], + "angle": 0, + "content": "attention maps, exhibit stable block rankings regardless of Gaussian smoothing. Unlike previous U-Net observations, where down-block features are consistently noisier than upblock ones, we observed no strong common trends across MM-DiT variants due to their distinct architectural differences. Nonetheless, very early and later blocks consistently produced noisier attention maps across all models. Thus, simply excluding these noisy blocks and using the identified top-5 blocks consistently led to robust results. Another practical advantage of selecting only the top-5 blocks is improved computational efficiency. Explicitly computing the full T2I attention map prohibits using optimized attention computations such as PyTorch's dedicated SDPA kernel [10, 11, 31, 47]. By limiting manual full attention computations to just the top-5 blocks, the remaining blocks can leverage these optimized kernels, significantly enhancing inference efficiency, as further discussed in Sec. 4.2." + }, + { + "type": "image", + "bbox": [ + 0.569, + 0.091, + 0.856, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.658 + ], + "angle": 0, + "content": "Figure 9. Architectural illustration of our method. We replace target (editing) branch projections \\(\\mathbf{q}_i^{\\mathrm{tgt}}\\) and \\(\\mathbf{k}_i^{\\mathrm{tgt}}\\) with their source branch counterparts, \\(\\mathbf{q}_i^{\\mathrm{src}}\\) and \\(\\mathbf{k}_i^{\\mathrm{src}}\\). For local blending, we store the T2I portion of unmodified attention maps from selected blocks in both branches during early timesteps. A binary mask is computed from the union of these attention maps after thresholding and used to blend the two latent images. Best viewed zoomed in." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.671, + 0.796, + 0.687 + ], + "angle": 0, + "content": "4. Editing via MM-DiT Attention" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.697, + 0.906, + 0.726 + ], + "angle": 0, + "content": "4.1. Attention map-based and input projection-based methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.907, + 0.884 + ], + "angle": 0, + "content": "Our analysis reveals that MM-DiT attention components serve roles similar to their U-Net counterparts: I2I blocks preserve source attributes (self-attention-like), while T2I blocks localize text-image interactions (cross-attention-like). Inspired by previous studies, we demonstrate image editing by substituting the target branch's attention map with the source branch's during early denoising timesteps. For common rectified flow schedulers (Euler's method), replacing attention maps during the initial \\(20\\%\\) of timesteps yields effective results." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.886, + 0.907, + 0.901 + ], + "angle": 0, + "content": "A key consideration is now determining which attention" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19496" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.092, + 0.483, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.255, + 0.485, + 0.327 + ], + "angle": 0, + "content": "Figure 10. Effect of text misalignment on image editing using Flux.1-dev model. When editing with highly different prompts (full list in the Appendix), replacing only image projections \\((\\mathbf{q}_i, \\mathbf{k}_i)\\) maintains better stability compared to full attention map replacement (all \\(\\mathbf{q}, \\mathbf{k}\\))." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.332, + 0.484, + 0.439 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.449, + 0.484, + 0.492 + ], + "angle": 0, + "content": "Figure 11. Comparison between replacing \\(\\mathbf{q}_i\\) and \\(\\mathbf{k}_i\\) projections and I2I block replacement. Results are highly similar, with projection-based methods being more computationally efficient." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.508, + 0.483, + 0.689 + ], + "angle": 0, + "content": "components to replace. While substituting the entire attention map (equivalently, full \\(\\mathbf{q}\\), \\(\\mathbf{k}\\) projections) is functional, it introduces text projection misalignment. Specifically, replacing the entire target attention map causes the text region to shift to the source branch's contexts (\\(\\mathbf{q}_t^{\\mathrm{src}} \\times \\mathbf{k}_t^{\\mathrm{src}}\\)). When this altered attention is multiplied by the target branch's value matrix text tokens (\\(\\mathbf{v}_t^{\\mathrm{tgt}}\\)), misalignment arises. With T5 embeddings, subtle prompt differences can significantly amplify these misalignments, leading to undesirable image shifts (illustrated in Fig. 10). This establishes a critical constraint: T2T regions must remain unmodified. In-depth explanation of this issue can be found in Appendix B.4." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "As a solution, we propose two alternative approaches: either selectively replacing only the I2I attention block, or specifically modifying just the image input projections \\((\\mathbf{q}_i,\\) \\(\\mathbf{k}_i)\\). Based on Eq. (2), these approaches slightly differ, as modifying \\(\\mathbf{q}_i\\) and \\(\\mathbf{k}_i\\) also influences T2I and I2T interactions. However, as shown in Fig. 11, results from both methods are nearly identical. Importantly, modifying at the input projection level \\((\\mathbf{q}_i,\\mathbf{k}_i)\\) is computationally more efficient since it enables optimized scaled dot product attention (SDPA) kernels. In contrast, replacing the I2I block disables SDPA, requiring manual attention calculation and careful switching before softmax to maintain proper attention distributions. This absence of SDPA can increase the computational time by up to \\(3\\times\\) (Tab. 1)." + }, + { + "type": "title", + "bbox": [ + 0.523, + 0.095, + 0.85, + 0.108 + ], + "angle": 0, + "content": "Algorithm 1: Prompt-based Image Editing with MM-DiT" + }, + { + "type": "algorithm", + "bbox": [ + 0.52, + 0.111, + 0.882, + 0.273 + ], + "angle": 0, + "content": "Parameters: Source prompt \\(P^{\\mathrm{src}}\\), Target prompt \\(P^{\\mathrm{tgt}}\\), Attention replace step \\(\\tau\\), Local Blending-{replace step \\(\\eta\\), threshold \\(\\theta\\)} \n\\(x_{T}^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)\\); \\(x_{T}^{\\mathrm{tgt}} \\gets x_{T}^{\\mathrm{src}}\\) \nfor \\(t = T, T - 1, \\ldots, 1\\) do \nif local blend and \\(t > \\eta\\) then // Local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}, \\text{AtttnMap}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{src}} \\text{ if } t > \\tau\\right] \\\\\n\\text{Mask} \\gets (\\text{AtttnMap} > \\theta) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets x_{t-1}^{\\mathrm{tgt}} \\cdot \\text{Mask} + x_{t-1}^{\\mathrm{src}} \\cdot (1 - \\text{Mask})\n\\end{array}\n\\] \nelse // No local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{sucr}} \\text{ if } t > \\tau\\right]\n\\end{array}\n\\] \nreturn (Decode(\\(x_0^{\\mathrm{src}}\\)), Decode(\\(x_0^{\\mathrm{tgt}}\\)))" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.306, + 0.907, + 0.504 + ], + "angle": 0, + "content": "To this end, we propose a new approach based on \\(\\mathbf{q}_i\\), \\(\\mathbf{k}_i\\) replacement. Unlike prior U-Net-based attention modification methods that require a precise token mapping between source and target prompts, limiting their effectiveness when prompts differ significantly, our method can handle arbitrarily different prompt pairs. This approach effectively realizes our editing objective: maintaining base image similarity while incorporating edited prompt elements, a capability uniquely enabled by MM-DiT's full-attention architecture. By preserving the text region intact, our method eliminates the need for explicit token correspondence or specialized mappers between prompts, allowing robust performance across diverse editing scenarios." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.515, + 0.666, + 0.531 + ], + "angle": 0, + "content": "4.2. Local blending" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.538, + 0.907, + 0.703 + ], + "angle": 0, + "content": "Using insights from Sec. 3.3, we extract clear attention maps from the top-5 transformer blocks for precise local blending, ensuring only targeted regions change while preserving the base image elsewhere. We apply binary blending masks derived from these attention maps once per timestep after all blocks are processed, following P2P's method. This selective calculation allows optimized SDPA kernels in remaining blocks, maintaining inference speed similar to naive batch size-2 inference (SD3-M: 15.2s vs. 14.9s, Flux.1-dev: 55.9s vs. 53.7s, SD3.5-L: 50.1s vs. 47.5s on a single A6000 GPU)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Our complete pipeline is illustrated in Fig. 9, with detailed pseudocode provided in Algorithm 1. Among the three hyperparameters controlling our method, such as masking threshold \\(\\theta\\), attention replace timestep \\(\\tau\\), and local blending stop timestep \\(\\eta\\), we find that setting \\(\\tau\\) to \\(20\\%\\) and \\(\\eta\\) to \\(50\\%\\) (i.e., \\(\\tau = 0.8T\\), \\(\\eta = 0.5T\\)) of total timesteps typically achieves robust performance. Thus, the masking threshold \\(\\theta\\) remains a primary parameter influencing edit granularity. As demonstrated in Fig. 12, higher thresholds effectively enable precise local edits, while lower thresholds allow broader modifications. Leveraging MM-DiT's precise attention localization and our attention map curation, the proposed local blending method significantly im" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19497" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.139, + 0.093, + 0.861, + 0.306 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.908, + 0.345 + ], + "angle": 0, + "content": "Figure 12. Ablation results on local blending threshold. Fine edits like text modifications benefit from higher thresholds, whereas broader transformations like large object changes favor lower thresholds. Local blending was applied to the first \\(50\\%\\) of timestep iterations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.36, + 0.483, + 0.391 + ], + "angle": 0, + "content": "proves edit precision compared to results obtained without attention curation (Fig. A7)." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.4, + 0.462, + 0.417 + ], + "angle": 0, + "content": "4.3. Controlling edit strength via block selection" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.422, + 0.483, + 0.665 + ], + "angle": 0, + "content": "While injecting input projections or attention maps into all blocks is straightforward, we observe that later blocks have a stronger influence on the final output, producing edited images closely resembling the source. While this correlation aligns with our goal, it can pose issues for few-step models like Flux.1-schnell, where editing all blocks in a single step of a 4-step inference may yield outputs overly similar to the base image, regardless of the target prompt. To address this, we replace only the first 38 and 30 blocks for Flux.1-schnell and SD3.5-L-Turbo, respectively, while replacing every block for non-distilled models. Note that our two block-selection strategies (top-5 blocks for obtaining attention maps and blocks for \\(\\mathbf{q}_i\\), \\(\\mathbf{k}_i\\) projections replacement) are not considered as hyperparameters and remain fixed. More details and ablations are available in Appendix C.2 and Fig. A8." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.679, + 0.285, + 0.696 + ], + "angle": 0, + "content": "5. Editing Real Images" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.902 + ], + "angle": 0, + "content": "A common approach in existing real image editing methods [2, 9, 56, 63] is to invert the source image into an initial noisy latent and then generate the target image from this inverted latent using an editing prompt. Our editing technique, which operates via \\(\\mathbf{q}_i\\), \\(\\mathbf{k}_i\\) replacement, functions in parallel with existing methods and can be readily applied in this scenario. Fig. 13 illustrates this process, and a preview of our results can be seen in Fig. 14. In Appendix A, we present further discussions, including quantitative and qualitative comparisons, using RF inversion as a baseline. Generally, the performance of real image editing methods depends heavily on inversion quality, as the initial latent affects background consistency and overall image quality." + }, + { + "type": "image", + "bbox": [ + 0.526, + 0.361, + 0.895, + 0.525 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.539, + 0.907, + 0.595 + ], + "angle": 0, + "content": "Figure 13. Illustration of our real image editing method given an initial latent. Unlike the synthetic image editing case, the source branch is fixed as a straight conditional path from the initial noise to the source image, independent of model evaluation." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.606, + 0.906, + 0.712 + ], + "angle": 0, + "content": "While this dependency also exists for our method, the impact of the initial latent is partially mitigated by continuously injecting input projections from the noisy source image and performing local blending to preserve unaffected regions. Consequently, our method can be used without inversion, starting from random noise, offering a trade-off between computational efficiency and performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.725, + 0.646, + 0.742 + ], + "angle": 0, + "content": "6. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.75, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Due to the lack of dedicated architecture-level editing methods for MM-DiT, we validate our approach against two simple baselines: (1) direct generation with fixed seeds, and (2) prompt switching after \\(20\\%\\) of timesteps, similar to SDEdit [41]. We randomly sample 60 diverse prompts from the PARTI dataset and use LLM [1] to generate edited prompts consisting of 30 involving simple object or word edits and 30 involving complex changes affecting multiple prompt components. We evaluate editing quality using LPIPS [69] and CLIPScore [20]. As these metrics have" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19498" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.09, + 0.484, + 0.434 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.149, + 0.435, + 0.424, + 0.447 + ], + "angle": 0, + "content": "A Golden Retriever sitting on an autumn grass field" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.465, + 0.484, + 0.508 + ], + "angle": 0, + "content": "Figure 14. Real image editing comparison between our method and RF inversion. Our approach enables more precise local edits while preserving unchanged regions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.52, + 0.484, + 0.618 + ], + "angle": 0, + "content": "Table 2. Quantitative comparison with different baselines. While higher CLIP scores indicate better target prompt alignment, lower LPIPS scores do not necessarily reflect edit quality since minimal edits yield the lowest scores. Our method well balances source content preservation and target prompt matching, whereas fixed seed generation often produces entirely different images and prompt-change produces overly conservative modifications." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.629, + 0.486, + 0.748 + ], + "angle": 0, + "content": "
ModelFixed SeedPrompt ChangeOurs
LPIPSCLIPLPIPSCLIPLPIPSCLIP
SD3-M0.5940.3770.3250.3440.3800.359
SD3.5-M0.5840.3800.3220.3490.4560.374
SD3.5-L0.5570.3880.3060.3630.4180.377
SD3.5-L-Turbo0.5090.3730.2250.3320.3880.363
Flux.1-dev0.5790.3590.2510.3110.3690.339
Flux.1-schnell0.5080.3800.1160.2890.2790.333
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.901 + ], + "angle": 0, + "content": "inherent limitations in capturing various types of prompt-based editing quality, we also provide user study result in Fig. A9 and extensive qualitative results in Fig. 15, Fig. A11, and Fig. A12. All experiments use 28 timesteps, except for distilled models (SD3.5-L-Turbo and Flux.1-schnell), which use 4. To ensure fair comparisons and avoid hyperparameter tuning, we fix the attention replacement step to the first \\(20\\%\\) of iterations and disable local blending for the results in Tab. 2 and Fig. 15, though our method" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.088, + 0.907, + 0.495 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.505, + 0.907, + 0.547 + ], + "angle": 0, + "content": "Figure 15. Qualitative comparison of our method with baselines. Our method consistently outperforms baseline approaches across diverse editing scenarios." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.557, + 0.774, + 0.572 + ], + "angle": 0, + "content": "can incorporate blending when needed." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.573, + 0.907, + 0.693 + ], + "angle": 0, + "content": "For real-image editing, we observe that editing performance depends more heavily on the inversion technique than the architectural-level editing method. Nonetheless, we report quantitative results on PIE-Bench [27] using RF inversion as our inversion technique and provide extensive qualitative results. Please refer to Appendix A.3 and A.4 for details. Our method provides additional controllability and improves the naive RF inversion baseline." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.709, + 0.634, + 0.724 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.903 + ], + "angle": 0, + "content": "This work presents a comprehensive analysis of MM-DiT's attention mechanism, revealing several underlying insights, including scaling properties and text-image interactions. Based on these observations, we develop an efficient editing method tailored for MM-DiT architectures that achieves high-quality results across various scenarios, including few-step models and real image editing. We believe these findings help deepen the understanding of emerging architectures and contribute to advancements in controlled image generation and editing. Discussions on limitations and future directions can be found in Appendix F." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "19499" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.091, + 0.091, + 0.27, + 0.108 + ], + "angle": 0, + "content": "8. Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "This work was supported by IITP grant (No. RS-2021-II211343, Artificial Intelligence Graduate School Program at Seoul National University) \\((5\\%)\\), IITP grant (No. RS-2024-00509257, Global AI Frontier Lab) \\((65\\%)\\), and NRF grant (No. RS-2024-00405857) \\((30\\%)\\), funded by the Korea government (MSIT)." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.222, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.247, + 0.483, + 0.274 + ], + "angle": 0, + "content": "[1] Anthropic. Claude 3.5 Sonnet, 2023. Large language model developed by Anthropic. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.277, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. Stable flow: Vital layers for training-free image editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 7, 9" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.348, + 0.483, + 0.403 + ], + "angle": 0, + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.405, + 0.483, + 0.473 + ], + "angle": 0, + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.476, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.483, + 0.686 + ], + "angle": 0, + "content": "[7] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-alpha: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.483, + 0.744 + ], + "angle": 0, + "content": "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-delta: Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.746, + 0.483, + 0.8 + ], + "angle": 0, + "content": "[9] Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. FluxSpace: Disentangled Semantic Editing in Rectified Flow Transformers. arXiv preprint arXiv:2412.09611, 2024. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.843 + ], + "angle": 0, + "content": "[10] Tri Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. In International Conference on Learning Representations (ICLR), 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.846, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[11] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.247, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.147 + ], + "angle": 0, + "content": "[12] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision Transformers Need Registers. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.149, + 0.906, + 0.203 + ], + "angle": 0, + "content": "[13] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing. arXiv preprint arXiv:2412.07517, 2024. 2, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.247, + 0.906, + 0.328 + ], + "angle": 0, + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.906, + 0.384 + ], + "angle": 0, + "content": "[16] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. ReNoise: Real Image Inversion Through Iterative Noising. arXiv preprint arXiv:2403.14602, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.386, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[17] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.456, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.511, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In International Conference on Learning Representations (ICLR), 2022. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[22] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In IEEE International Conference on Computer Vision (ICCV), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.906, + 0.789 + ], + "angle": 0, + "content": "[23] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.79, + 0.905, + 0.859 + ], + "angle": 0, + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[25] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual Style Prompting with Swapping Self-Attention. arXiv preprint arXiv:2402.12974, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19500" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.146 + ], + "angle": 0, + "content": "[26] Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang, and Renjie Liao. Uniedit-flow: Unleashing inversion and editing in the era of flow models. arXiv preprint arXiv:2504.13109, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.148, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[27] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code. In International Conference on Learning Representations (ICLR), 2024. 8, 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.273 + ], + "angle": 0, + "content": "[28] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling Diffusion Models into Conditional GANs. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.371 + ], + "angle": 0, + "content": "[30] Black Forest Labs. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.Accessed:November 2024.1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[31] Benjamin Lefaudeaux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xFormers: A modular and hackable Transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.524 + ], + "angle": 0, + "content": "[32] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow Matching for Generative Modeling. In International Conference on Learning Representations (ICLR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.525, + 0.482, + 0.594 + ], + "angle": 0, + "content": "[33] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.482, + 0.663 + ], + "angle": 0, + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.664, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[35] Xingchao Liu, Chengyue Gong, et al. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In International Conference on Learning Representations (ICLR), 2023. 2, 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[36] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In International Conference on Learning Representations (ICLR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[37] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[38] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[39] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[40] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 7, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.305 + ], + "angle": 0, + "content": "[42] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.307, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[43] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.377, + 0.905, + 0.418 + ], + "angle": 0, + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.905, + 0.448 + ], + "angle": 0, + "content": "[45] OpenAI. GPT-4o, 2024. Large language model developed by OpenAI. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.49 + ], + "angle": 0, + "content": "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.492, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.577, + 0.905, + 0.618 + ], + "angle": 0, + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.62, + 0.905, + 0.689 + ], + "angle": 0, + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "19501" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.189 + ], + "angle": 0, + "content": "[53] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.19, + 0.482, + 0.258 + ], + "angle": 0, + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.342 + ], + "angle": 0, + "content": "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.482, + 0.411 + ], + "angle": 0, + "content": "[56] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations. arXiv preprint arXiv:2410.10792, 2024. 2, 7, 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.495 + ], + "angle": 0, + "content": "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.537 + ], + "angle": 0, + "content": "[58] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In International Conference on Learning Representations (ICLR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.539, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[59] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial Diffusion Distillation. arXiv e-prints, pages arXiv-2311, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.649 + ], + "angle": 0, + "content": "[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.691 + ], + "angle": 0, + "content": "[61] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. In International Conference on Machine Learning (ICML), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[62] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.803 + ], + "angle": 0, + "content": "[63] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[64] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[65] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. TurboEdit: Instant text-based" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "image editing. In European Conference on Computer Vision (ECCV), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.178 + ], + "angle": 0, + "content": "[66] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.179, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[67] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.318 + ], + "angle": 0, + "content": "[68] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. Transactions on Machine Learning Research, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.375 + ], + "angle": 0, + "content": "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.376, + 0.905, + 0.431 + ], + "angle": 0, + "content": "[70] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation. arXiv preprint arXiv:2405.01434, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.431 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "19502" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_origin.pdf b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..d790d711e93b1e37d394f7a6e44c20263269f66f --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/db432361-cc2c-467e-9c54-baa03e561a43_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:e0d6b643a8b26b9549f5c61e4b610a8a5eac2cfb60b17bae55fddd465d148c86 +size 17739476 diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/full.md b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/full.md new file mode 100644 index 0000000000000000000000000000000000000000..d84b477af2b6895d7c44db3ade3d096c62a77420 --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/full.md @@ -0,0 +1,325 @@ +# Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing + +Joonghyuk Shin Alchan Hwang Yujin Kim Daneul Kim Jaesik Park Seoul National University + +{joonghyuk, alchan00, yujin.k, carpedkm, jaesik.park}@snu.ac.kr + +![](images/586f68b5cedfe0bd0388233b015998856dc89a7b250960ca8ef09befd611a68c.jpg) +$\mathrm{SD3}\rightarrow \mathrm{Flux}$ + +![](images/712320bc04575f5da5148f541a8ba4b5e0ff37b211f266869e7d64527e757bff.jpg) +giraffe $\rightarrow$ dragon + +![](images/0e6336a9884ab6b03336fe4c691f10e9a8044a5e8af9e3cbf0f8c17debc3fc6f.jpg) +frogs $\rightarrow$ rabbits +ferret $\rightarrow$ lion + +![](images/193575404c9756b1bbd8970e9ea93db2ad5a38ee8217abc12c561f33284e3e12.jpg) +'help me' $\rightarrow$ 'Please don't kill me' + +![](images/e13d1ee8f091335a98199df420dfc6ec468a43482a825fb608b986bd658b160d.jpg) +Figure 1. Our prompt-based editing examples on SD3 series and Flux.1, demonstrating high-quality results across various editing scenarios. + +![](images/debaa94d5881c0ad10a2168e113198430a00cea96b0b33283a32a6ddab0a90bc.jpg) +pig $\rightarrow$ whale + +![](images/e98a83cd068015c4f1c364d145ef582726591111e7547abd3850856cbffa9ae5.jpg) + +![](images/6c269df0cc808f17a96040deae4eb2afb2a10317ee5746c5606642d5b69831eb.jpg) + +# Abstract + +Transformer-based diffusion models have recently superseded traditional U-Net architectures, with multimodal diffusion transformers (MM-DiT) emerging as the dominant approach in state-of-the-art models like Stable Diffusion 3 and Flux.1. Previous approaches have relied on unidirectional cross-attention mechanisms, with information flowing from text embeddings to image latents. In contrast, MM-DiT introduces a unified attention mechanism that concatenates input projections from both modalities and performs a single full attention operation, allowing bidirectional information flow between text and image branches. This architectural shift presents significant challenges for existing editing techniques. In this paper, we systematically analyze MM-DiT's attention mechanism by decomposing attention matrices into four distinct blocks, revealing their inherent characteristics. Through these analyses, we propose a robust, prompt-based image editing method for MM-DiT that supports global to local edits across various MM-DiT variants, including few-step models. We believe our findings + +bridge the gap between existing U-Net-based methods and emerging architectures, offering deeper insights into MM-DiT's behavioral patterns. + +# 1. Introduction + +Recent years have witnessed the emergence of diffusion models [21, 60] as the de-facto approach for visual generation, demonstrating remarkable capabilities in creating high-quality data from noise [15, 29, 39, 44, 49, 52, 54, 57]. While early implementations predominantly employed U-shaped networks [55], transformer-based architectures [7, 48] have recently emerged as the prevailing approach, with multimodal diffusion transformers (MM-DiT) showing particular promise in state-of-the-art models like Stable Diffusion 3 series [15] and Flux.1 [30]. + +Motivated by MM-DiT's architectural differences and the challenges in directly applying existing U-Net-based editing methods, our work addresses a fundamental question: "How can we leverage our understanding of U-Net + +attention mechanisms to interpret MM-DiT's behavior better and effectively transfer existing techniques to this new architecture?" Given the significance of existing work built upon U-Net attention mechanisms [4, 5, 19, 33], answering this question is crucial for enabling existing methods to adapt to, and benefit from, emerging MM-DiT architectures. + +To this end, we systematically investigate MM-DiT's attention mechanisms through block matrix analysis, revealing key characteristics of its attention maps. As one of our key analytical insights, we observe an interesting phenomenon aligned with vision transformer [14] scaling laws [12]: as the model size increases, attention patterns become properly positioned but increasingly noisy, necessitating careful selection of attention layers for editing tasks. Building upon these analyses, we propose an efficient editing method at the architecture level that focuses on modifying image tokens in input projections. The proposed editing method can be jointly combined with inversion methods [56, 63] for editing real images as well. + +Our contributions are primarily architectural: (1) We provide the first systematic analysis of MM-DiT attention mechanisms, highlighting block-wise patterns and scaling behaviors. (2) Building upon these findings, we propose an efficient and robust editing method across diverse editing scenarios - from global stylistic changes to local edits such as text modification - and various MM-DiT variants, including few-step models (SD3-M, SD3.5-M, SD3.5-L, SD3.5-L-Turbo, and Flux.1-dev/schnell). + +# 2. Related Work + +# 2.1. Text-to-image diffusion models + +Early text-to-image diffusion models mostly used U-shaped networks with residual blocks [55] incorporating pixelwise self-attention and text-image cross-attention layers, combined with downsampling and upsampling operations (Fig. 2). These models are typically trained with DDPM-style scheduling and noise prediction objective [49, 54]. + +Recent advances introduced two significant shifts. First, architecturally, there has been a transition from U-Nets to transformer-based architectures like DiT [48], demonstrating superior scaling properties. The PixArt-alpha series [6-8] successfully adapted DiT for text conditional generation, maintaining a conventional cross-attention mechanism for text conditioning. Stable Diffusion 3 [15] subsequently introduced a pivotal shift with multimodal diffusion transformers (MM-DiT), using separate transformers for text and image modalities while concatenating their sequences for unified attention operations (Fig. 2). Second, formulation-wise, newer approaches have adopted v-prediction [58] with a rectified flow framework [32, 35] that connects data and noise distributions along straight paths, replacing the com + +Table 1. Architectural details and inference times for $1024 \times 1024$ image generation using 28 timesteps, using single A6000 GPU. Times are shown as optimized (left, using PyTorch SDPA kernels) vs. naive matrix multiplication (right). Text encoders: $\mathrm{CLIP}^1$ refers to CLIP ViT-L/14 [50], $\mathrm{CLIP}^2$ refers to OpenCLIP ViTG/14 [24], and T5 refers to T5-XXL [51]. + +
Arch. / ParamsText Enc.Attn TypeTime (s)
SD1U-Net / 860MCLIP1Self/Cross5.0/12.7
SDXLU-Net / 2.6BCLIP1,2Self/Cross5.6/7.9
SD3-MMM-DiT / 2BCLIP1,2 / T5Full7.1/20.5
SD3.5-MMM-DiT-X / 2.2BCLIP1,2 / T5Full/Self9.6/26.8
SD3.5-LMM-DiT / 8BCLIP1,2 / T5Full24.7/60.5
Flux.1-devMM-DiT* / 11.9BCLIP1 / T5Full27.8/36.8
+ +plex trajectories of traditional diffusion models. Together with increased model scales, these developments have significantly improved model performance. Parallel efforts have also accelerated iterative diffusion sampling via faster ODE solvers [37, 38], consistency distillation [40, 61], adversarial learning [28, 59, 66, 67], and trajectory straightening [32, 35, 36]. + +# 2.2. Text-guided image editing + +Text-guided image editing naturally extends text-to-image diffusion models. Prior work [18, 22, 33, 64] shows that self-attention governs spatial structure while cross-attention handles attribute-level semantics aligned with text. Prompt-to-Prompt (P2P) [19], a seminal approach, leverages this by directly transferring cross-attention maps for precise edits. Subsequent methods further manipulate self-attention mechanisms for better content-preserving edits [4, 25, 43, 70]. A key advantage of these attention-based methods is their ability to achieve precise local edits solely from textual prompts, eliminating the need for explicit spatial guidance or additional user input. + +Another direction employs inversion techniques [13, 16, 23, 26, 56, 63], mapping real images to editable latent spaces. While inversion alone can lose fine details, combining it with attention-based methods significantly enhances localized editing [42]. Alternatively, methods like InstructPix2Pix [3] train dedicated models on synthetic data, though this approach incurs high computational costs and dataset preparation overheads [65]. + +# 2.3. MM-DiT variants + +Since its introduction, MM-DiT architecture has seen several improvements. The initial dual-branch design in SD3 was motivated by the conceptual differences between text and image embeddings. This architecture was maintained in both SD3 and SD3.5-Large models, which contain 24 and 38 dual-branch blocks, respectively. Flux.1, scaled to 12B parameters, introduced a hybrid approach with 57 total blocks - 19 dual branch blocks followed by 38 single branch blocks (Fig. 2). The single branch variant processes + +![](images/5e64765a4f55ddaf639b4e1d4903f79d0681ca0528134398176b45032a3395e5.jpg) +Figure 2. Attention mechanisms across diffusion architectures. (a) U-Net with separate self-/cross- attention, (b) MM-DiT blocks with dual branches, (c) MM-DiT* single branch block used in later layers of Flux.1, and (d) MM-DiT-X dual attention block with explicit self-attention, used in early layers of SD3.5-M. + +concatenated text and image embeddings as a single modality using unified weights (e.g., linear layers, attention parameters), reducing block parameters from 340M to 141M. The MM-DiT-X variant, later introduced in SD3.5-medium, added explicit self-attention operations exclusively for image tokens in the first 13 transformer blocks for better multi-resolution results and overall image coherence. + +# 3. Attention Analysis in MM-DiT + +# 3.1. Block matrix formulation of MM-DiT attention + +We begin the analysis with MM-DiT's attention structure under SD3-M's standard dimensions: $64 \times 64$ image latents (flattened to 4096) and 333 text tokens (77 CLIP + 256 T5), with inner dimension 64. Omitting the batch and head dimensions for clarity, the input projections for image and text domains are denoted as $(\mathbf{q}_i, \mathbf{k}_i, \mathbf{v}_i \in \mathbb{R}^{4096 \times 64})$ and $(\mathbf{q}_t, \mathbf{k}_t, \mathbf{v}_t \in \mathbb{R}^{333 \times 64})$ . Then, query, key, and value projections are concatenated as shown below ( $\mathbf{q}, \mathbf{k}, \mathbf{v} \in \mathbb{R}^{4429 \times 64}$ ). It is worth noting that Flux.1 adopts an inverse concatenation order (text preceding image), which results in a different attention map. + +$$ +\mathbf {q} = \left[ \begin{array}{l} \mathbf {q} _ {i} \\ \mathbf {q} _ {t} \end{array} \right], \mathbf {k} = \left[ \begin{array}{l} \mathbf {k} _ {i} \\ \mathbf {k} _ {t} \end{array} \right], \mathbf {v} = \left[ \begin{array}{l} \mathbf {v} _ {i} \\ \mathbf {v} _ {t} \end{array} \right]. \tag {1} +$$ + +Using the standard scaled dot product attention formulation $\text{softmax}\left(\frac{\mathbf{q}\mathbf{k}^T}{\sqrt{d_k}}\right)\mathbf{v}$ , we can write the attention map $\mathbf{q}\mathbf{k}^T$ and its output representation as follows. Note that row-wise softmax normalization is applied to $\mathbf{q}\mathbf{k}^T$ in Eq. (2), before + +![](images/7c710e26405805760b5aedf9c3f112a2d920ce22b9e18f928da065448a00ab17.jpg) +Figure 3. Visualizing the top-6 principal components extracted from I2I attention blocks, averaged across all timesteps and transformer blocks. Like U-Net's self-attention layers, these components reveal underlying spatial and geometric patterns. + +multiplication with $\mathbf{v}$ to obtain Eq. (3). + +$$ +\mathbf {q} \mathbf {k} ^ {T} = \left[ \begin{array}{l} \mathbf {q} _ {i} \\ \mathbf {q} _ {t} \end{array} \right] \left[ \begin{array}{l l} \mathbf {k} _ {i} ^ {T} & \mathbf {k} _ {t} ^ {T} \end{array} \right] = \left[ \begin{array}{l l} \mathbf {q} _ {i} \mathbf {k} _ {i} ^ {T} & \mathbf {q} _ {i} \mathbf {k} _ {t} ^ {T} \\ \mathbf {q} _ {t} \mathbf {k} _ {i} ^ {T} & \mathbf {q} _ {t} \mathbf {k} _ {t} ^ {T} \end{array} \right] \sim \left[ \begin{array}{l l} \mathrm {I 2 I} & \mathrm {T 2 I} \\ \mathrm {I 2 T} & \mathrm {T 2 T} \end{array} \right], \tag {2} +$$ + +$$ +\mathbf {q} \mathbf {k} ^ {T} \mathbf {v} = \left[ \begin{array}{l l} \mathbf {q} _ {i} \mathbf {k} _ {i} ^ {T} & \mathbf {q} _ {i} \mathbf {k} _ {t} ^ {T} \\ \mathbf {q} _ {t} \mathbf {k} _ {i} ^ {T} & \mathbf {q} _ {t} \mathbf {k} _ {t} ^ {T} \end{array} \right] \left[ \begin{array}{l} \mathbf {v} _ {i} \\ \mathbf {v} _ {t} \end{array} \right] = \left[ \begin{array}{l} \mathbf {q} _ {i} \mathbf {k} _ {i} ^ {T} \mathbf {v} _ {i} + \mathbf {q} _ {i} \mathbf {k} _ {t} ^ {T} \mathbf {v} _ {t} \\ \mathbf {q} _ {t} \mathbf {k} _ {i} ^ {T} \mathbf {v} _ {i} + \mathbf {q} _ {t} \mathbf {k} _ {t} ^ {T} \mathbf {v} _ {t} \end{array} \right]. \tag {3} +$$ + +Since the upper and lower rows in Eq. (3) correspond to the resultant image and text latents, respectively, this notation choice naturally reflects the underlying behavior of concatenated attention. Within this generalized framework, our empirical findings suggest that I2I and T2I blocks can serve roles similar to the conventional self- and cross-attention patterns in U-Net architectures. + +# 3.2. Block-wise attention patterns + +I2I & T2T. I2I and T2T are unimodal portions of the attention map. As demonstrated by PCA analysis in Fig. 3, the I2I block is analogous to self-attention in U-Net architectures, capturing spatial layout and geometric information. In contrast, T2T blocks predominantly manifest as identity matrices, with attention weights notably weakening after the occurrence of meaningful tokens (e.g., EOS). + +T2I & I2T. T2I and I2T blocks represent cross-modal interactions, where T2I influences subsequent image latents, and I2T affects text representations - with I2T being a novel addition not present in U-Net architectures. These blocks are essential for precise local editing as they encode token-specific image-region correspondences, generating binary masks for precise local blending. While both blocks can be utilized to obtain these masks, we find T2I more effective (Fig. 4). Specifically, due to the row-wise softmax operation in I2T blocks, where the rows correspond to image tokens, there is inherent competition among these tokens since their attention weights must sum to 1. This constraint dilutes the attention signals, limiting their ability to represent strong or broad influences. In contrast, T2I's structure allows multiple image regions to retain high attention values simultaneously. Additionally, for the SD3 model series using both CLIP and T5 encoders, we observe that T5-derived + +![](images/644646cbf21ca17501b775f4f75f830ed2d9c76e3a905ac29a4f3db7d6da5c56.jpg) +Figure 4. Comparison of Flux.1-dev's T2I and I2T attention blocks averaged across all timesteps and transformer blocks. T2I blocks show stronger localization due to their unconstrained attention distribution compared to I2T's row-wise competition, making them more effective source for localized editing. + +![](images/194ae3a0fdc66c05ced6a36e4a70dcc6a00539c4e502b06dfd0dd142de2d9629.jpg) +Figure 5. Analysis of T2I attention blocks from SDXL and various MM-DiT models using the prompt "House on a mountainside under a clear blue daylight sky" (averaged across timesteps and transformer blocks). MM-DiT models show better localized attention compared to U-Net architectures, though attention noise increases with model scale. When editing "blue daylight" to "red dawn" by replacing entire attention blocks, MM-DiT's precise attention enables clean edits, while SDXL applies the change globally. + +attention maps typically exhibit higher precision than CLIP. + +Based on these analyses, we observe that I2I blocks are crucial for preserving identity, and T2I blocks are effective in obtaining attention masks for precise targeted local editing. We also validate the relative importance of these four sub-blocks in Fig. A2, which highlights the effectiveness of the I2I block in preserving original image attributes. Furthermore, visualizations of T2I attention in Fig. A14, Fig. A15, Fig. A16, Fig. A17 suggest that, during full attention's modality mixing, rotary positional embeddings (RoPE) [62] and residual connections help each modality retain its distinctive characteristics at designated position throughout transformer blocks. This supports and justifies our block-based interpretation and editing. Due to space constraints, we present extensive additional findings in Appendix B, to which we refer readers for further details. + +![](images/20250207dc019223b5a20ab40312e60b66eae532ff043e14ef184da6063f5c7b.jpg) +Figure 6. Visualization of Flux.1-dev's T2I block, averaged across timesteps, comparing masks from top-5 selected blocks (upper) versus all blocks (lower), both with and without smoothing. Block selection significantly reduces noise, while smoothing refines boundary transitions. The lower left displays Grounded SAM2's masks used as ground truth for top-5 block selection. + +# 3.3. Mitigating noisy attention maps + +We compare cross-attention maps (T2I blocks for MM-DiTs) across various models in Fig. 5. While MM-DiT architectures exhibit significantly improved attention localization, we observe an interesting phenomenon: as MM-DiT model size grows, the attention maps become increasingly noisy. Directly employing these noisy, discontinuous attention maps as local blending masks introduce visible artifacts in generated images. This issue aligns with recent findings by [12], which report similar noise artifacts emerging in ViTs at larger model and dataset scales. While they propose to add a register (additional tokens) to remove these artifacts, it is impractical here due to the prohibitive cost of retraining large MM-DiT models. Thus, we propose two practical solutions: (1) selecting optimal transformer blocks that inherently produce more precise and less noisy attention maps, and (2) applying Gaussian smoothing to smooth mask boundaries and reduce artifacts. + +For systematic evaluation, we sampled 100 prompts from the PARTI prompts [68], a widely used prompt set for benchmarking, and we created ground truth masks using Grounded SAM2 [34, 53] for grounded segmentation tasks, as shown in Fig. 6. Intuitively, effective blocks should yield attention maps similar to the GT masks. Thus, using standard losses and measures (BCE, Soft mIoU, and MSE) against the binary GT masks, we selected the top-5 blocks for each model based on the average ranking. Blockwise metrics for Flux.1 are presented in Fig. 7, and specific rankings are provided in Tab. A2. We emphasize that these blocks are not prompt-specific, as they were calculated over 100 random prompts, and we consistently use the same five blocks throughout the paper. Local editing results with and without these selected blocks are presented in Fig. 8. + +Notably, applying Gaussian smoothing alters the block ranking in larger models, suggesting that some noisy blocks still contain meaningful information. In contrast, smaller models like SD3-2B, which inherently produce smoother + +![](images/58356bc12b89ef55c1b356719efa76f9dcb10d3fef9d5a3ff56a850bbd621342.jpg) + +![](images/40552c483fa6f7e86fc1ecaaf03b3c70f1cdc48b875dbea7bb187e28c94e5186.jpg) + +![](images/58b6675bab5ca2f98899159bcaab018bb2921014ca9535304e77fe197195f5f8.jpg) + +![](images/a637d3b3b2454f55fc3bcecc7c79e2f153766ad464605424ec5e8008eb69ed48.jpg) +Figure 7. Transformer block analysis of Flux.1-dev using Binary Cross Entropy Loss, Soft mIoU, and MSE, with Grounded SAM2 predictions as ground truth. Scores are shown without (upper) and with (lower) Gaussian smoothing, with all evaluation metrics showing notable improvements after smoothing. T2I shows higher accuracy than I2T across most blocks. Top-5 ranked T2I attention blocks are exclusively used for generating local edit masks. Additional model results are provided in Appendix C.1. + +![](images/4231d97e06240f648447c1d7ff2e5f678232a815dc14af09e01a15b367556056.jpg) + +![](images/36b9fe7a4b019eb29f4780af12375ac1fe651749a470e62d133ab0cbbe5d5851.jpg) + +![](images/62dbd1448bd0e8b41691f43b06bda3d766816657d1b1ac44482e309732c88dd8.jpg) +Figure 8. Local image editing results using masks derived from the top-5 transformer blocks with Gaussian smoothing. Employing only the top-5 blocks reduces noise in the attention masks, effectively preserving untouched regions (e.g., music notes). Applying Gaussian smoothing further reduces boundary artifacts. + +attention maps, exhibit stable block rankings regardless of Gaussian smoothing. Unlike previous U-Net observations, where down-block features are consistently noisier than upblock ones, we observed no strong common trends across MM-DiT variants due to their distinct architectural differences. Nonetheless, very early and later blocks consistently produced noisier attention maps across all models. Thus, simply excluding these noisy blocks and using the identified top-5 blocks consistently led to robust results. Another practical advantage of selecting only the top-5 blocks is improved computational efficiency. Explicitly computing the full T2I attention map prohibits using optimized attention computations such as PyTorch's dedicated SDPA kernel [10, 11, 31, 47]. By limiting manual full attention computations to just the top-5 blocks, the remaining blocks can leverage these optimized kernels, significantly enhancing inference efficiency, as further discussed in Sec. 4.2. + +![](images/eb89e1019389df6054e71642085ee3dab8c5de12eea31302d664a3a35baffafc.jpg) +Figure 9. Architectural illustration of our method. We replace target (editing) branch projections $\mathbf{q}_i^{\mathrm{tgt}}$ and $\mathbf{k}_i^{\mathrm{tgt}}$ with their source branch counterparts, $\mathbf{q}_i^{\mathrm{src}}$ and $\mathbf{k}_i^{\mathrm{src}}$ . For local blending, we store the T2I portion of unmodified attention maps from selected blocks in both branches during early timesteps. A binary mask is computed from the union of these attention maps after thresholding and used to blend the two latent images. Best viewed zoomed in. + +# 4. Editing via MM-DiT Attention + +# 4.1. Attention map-based and input projection-based methods + +Our analysis reveals that MM-DiT attention components serve roles similar to their U-Net counterparts: I2I blocks preserve source attributes (self-attention-like), while T2I blocks localize text-image interactions (cross-attention-like). Inspired by previous studies, we demonstrate image editing by substituting the target branch's attention map with the source branch's during early denoising timesteps. For common rectified flow schedulers (Euler's method), replacing attention maps during the initial $20\%$ of timesteps yields effective results. + +A key consideration is now determining which attention + +![](images/af54c4eba972f5600a0df8148d2223dc6c0c517d4bd262515543bd5256f24f58.jpg) +Figure 10. Effect of text misalignment on image editing using Flux.1-dev model. When editing with highly different prompts (full list in the Appendix), replacing only image projections $(\mathbf{q}_i, \mathbf{k}_i)$ maintains better stability compared to full attention map replacement (all $\mathbf{q}, \mathbf{k}$ ). + +![](images/5decd7089e2f8a518b07d8476782361f626cb14635084a0b35da774730b0f930.jpg) +Figure 11. Comparison between replacing $\mathbf{q}_i$ and $\mathbf{k}_i$ projections and I2I block replacement. Results are highly similar, with projection-based methods being more computationally efficient. + +components to replace. While substituting the entire attention map (equivalently, full $\mathbf{q}$ , $\mathbf{k}$ projections) is functional, it introduces text projection misalignment. Specifically, replacing the entire target attention map causes the text region to shift to the source branch's contexts ( $\mathbf{q}_t^{\mathrm{src}} \times \mathbf{k}_t^{\mathrm{src}}$ ). When this altered attention is multiplied by the target branch's value matrix text tokens ( $\mathbf{v}_t^{\mathrm{tgt}}$ ), misalignment arises. With T5 embeddings, subtle prompt differences can significantly amplify these misalignments, leading to undesirable image shifts (illustrated in Fig. 10). This establishes a critical constraint: T2T regions must remain unmodified. In-depth explanation of this issue can be found in Appendix B.4. + +As a solution, we propose two alternative approaches: either selectively replacing only the I2I attention block, or specifically modifying just the image input projections $(\mathbf{q}_i,$ $\mathbf{k}_i)$ . Based on Eq. (2), these approaches slightly differ, as modifying $\mathbf{q}_i$ and $\mathbf{k}_i$ also influences T2I and I2T interactions. However, as shown in Fig. 11, results from both methods are nearly identical. Importantly, modifying at the input projection level $(\mathbf{q}_i,\mathbf{k}_i)$ is computationally more efficient since it enables optimized scaled dot product attention (SDPA) kernels. In contrast, replacing the I2I block disables SDPA, requiring manual attention calculation and careful switching before softmax to maintain proper attention distributions. This absence of SDPA can increase the computational time by up to $3\times$ (Tab. 1). + +# Algorithm 1: Prompt-based Image Editing with MM-DiT + +Parameters: Source prompt $P^{\mathrm{src}}$ , Target prompt $P^{\mathrm{tgt}}$ , Attention replace step $\tau$ , Local Blending-{replace step $\eta$ , threshold $\theta$ } + $x_{T}^{\mathrm{src}} \sim \mathcal{N}(0, I)$ ; $x_{T}^{\mathrm{tgt}} \gets x_{T}^{\mathrm{src}}$ +for $t = T, T - 1, \ldots, 1$ do +if local blend and $t > \eta$ then // Local blend +\[ +\begin{array}{l} +x_{t-1}^{\mathrm{src}}, (\{q, k\}_{i,t}^{\mathrm{src}}, \text{AtttnMap}) \leftarrow \mathrm{DM}(x_{t}^{\mathrm{src}}, P^{\mathrm{src}}, t) \\ +x_{t-1}^{\mathrm{tgt}} \gets \mathrm{DM}(x_{t}^{\mathrm{tgt}}, P^{\mathrm{tgt}}, t) \left[\{q, k\}_{i,t}^{\mathrm{tgt}} \gets \{q, k\}_{i,t}^{\mathrm{src}} \text{ if } t > \tau\right] \\ +\text{Mask} \gets (\text{AtttnMap} > \theta) \\ +x_{t-1}^{\mathrm{tgt}} \gets x_{t-1}^{\mathrm{tgt}} \cdot \text{Mask} + x_{t-1}^{\mathrm{src}} \cdot (1 - \text{Mask}) +\end{array} +\] +else // No local blend +\[ +\begin{array}{l} +x_{t-1}^{\mathrm{src}}, (\{q, k\}_{i,t}^{\mathrm{src}}) \leftarrow \mathrm{DM}(x_{t}^{\mathrm{src}}, P^{\mathrm{src}}, t) \\ +x_{t-1}^{\mathrm{tgt}} \gets \mathrm{DM}(x_{t}^{\mathrm{tgt}}, P^{\mathrm{tgt}}, t) \left[\{q, k\}_{i,t}^{\mathrm{tgt}} \gets \{q, k\}_{i,t}^{\mathrm{sucr}} \text{ if } t > \tau\right] +\end{array} +\] +return (Decode( $x_0^{\mathrm{src}}$ ), Decode( $x_0^{\mathrm{tgt}}$ )) + +To this end, we propose a new approach based on $\mathbf{q}_i$ , $\mathbf{k}_i$ replacement. Unlike prior U-Net-based attention modification methods that require a precise token mapping between source and target prompts, limiting their effectiveness when prompts differ significantly, our method can handle arbitrarily different prompt pairs. This approach effectively realizes our editing objective: maintaining base image similarity while incorporating edited prompt elements, a capability uniquely enabled by MM-DiT's full-attention architecture. By preserving the text region intact, our method eliminates the need for explicit token correspondence or specialized mappers between prompts, allowing robust performance across diverse editing scenarios. + +# 4.2. Local blending + +Using insights from Sec. 3.3, we extract clear attention maps from the top-5 transformer blocks for precise local blending, ensuring only targeted regions change while preserving the base image elsewhere. We apply binary blending masks derived from these attention maps once per timestep after all blocks are processed, following P2P's method. This selective calculation allows optimized SDPA kernels in remaining blocks, maintaining inference speed similar to naive batch size-2 inference (SD3-M: 15.2s vs. 14.9s, Flux.1-dev: 55.9s vs. 53.7s, SD3.5-L: 50.1s vs. 47.5s on a single A6000 GPU). + +Our complete pipeline is illustrated in Fig. 9, with detailed pseudocode provided in Algorithm 1. Among the three hyperparameters controlling our method, such as masking threshold $\theta$ , attention replace timestep $\tau$ , and local blending stop timestep $\eta$ , we find that setting $\tau$ to $20\%$ and $\eta$ to $50\%$ (i.e., $\tau = 0.8T$ , $\eta = 0.5T$ ) of total timesteps typically achieves robust performance. Thus, the masking threshold $\theta$ remains a primary parameter influencing edit granularity. As demonstrated in Fig. 12, higher thresholds effectively enable precise local edits, while lower thresholds allow broader modifications. Leveraging MM-DiT's precise attention localization and our attention map curation, the proposed local blending method significantly im + +![](images/9d1d616f17bc0fed6e7f584e6963029c7370686c208a3454f9b12d5310428db1.jpg) +Figure 12. Ablation results on local blending threshold. Fine edits like text modifications benefit from higher thresholds, whereas broader transformations like large object changes favor lower thresholds. Local blending was applied to the first $50\%$ of timestep iterations. + +proves edit precision compared to results obtained without attention curation (Fig. A7). + +# 4.3. Controlling edit strength via block selection + +While injecting input projections or attention maps into all blocks is straightforward, we observe that later blocks have a stronger influence on the final output, producing edited images closely resembling the source. While this correlation aligns with our goal, it can pose issues for few-step models like Flux.1-schnell, where editing all blocks in a single step of a 4-step inference may yield outputs overly similar to the base image, regardless of the target prompt. To address this, we replace only the first 38 and 30 blocks for Flux.1-schnell and SD3.5-L-Turbo, respectively, while replacing every block for non-distilled models. Note that our two block-selection strategies (top-5 blocks for obtaining attention maps and blocks for $\mathbf{q}_i$ , $\mathbf{k}_i$ projections replacement) are not considered as hyperparameters and remain fixed. More details and ablations are available in Appendix C.2 and Fig. A8. + +# 5. Editing Real Images + +A common approach in existing real image editing methods [2, 9, 56, 63] is to invert the source image into an initial noisy latent and then generate the target image from this inverted latent using an editing prompt. Our editing technique, which operates via $\mathbf{q}_i$ , $\mathbf{k}_i$ replacement, functions in parallel with existing methods and can be readily applied in this scenario. Fig. 13 illustrates this process, and a preview of our results can be seen in Fig. 14. In Appendix A, we present further discussions, including quantitative and qualitative comparisons, using RF inversion as a baseline. Generally, the performance of real image editing methods depends heavily on inversion quality, as the initial latent affects background consistency and overall image quality. + +![](images/fd726ff8904cf8f96cac1380fa8dfd6825a2c89e13db6ff08cea4fab8f5433d4.jpg) +Figure 13. Illustration of our real image editing method given an initial latent. Unlike the synthetic image editing case, the source branch is fixed as a straight conditional path from the initial noise to the source image, independent of model evaluation. + +While this dependency also exists for our method, the impact of the initial latent is partially mitigated by continuously injecting input projections from the noisy source image and performing local blending to preserve unaffected regions. Consequently, our method can be used without inversion, starting from random noise, offering a trade-off between computational efficiency and performance. + +# 6. Experiments + +Due to the lack of dedicated architecture-level editing methods for MM-DiT, we validate our approach against two simple baselines: (1) direct generation with fixed seeds, and (2) prompt switching after $20\%$ of timesteps, similar to SDEdit [41]. We randomly sample 60 diverse prompts from the PARTI dataset and use LLM [1] to generate edited prompts consisting of 30 involving simple object or word edits and 30 involving complex changes affecting multiple prompt components. We evaluate editing quality using LPIPS [69] and CLIPScore [20]. As these metrics have + +![](images/7ebb9d1f4aa687fbfb9f7f7688896c4094b72cf0c0e19153296b8e37fc6d5370.jpg) +A Golden Retriever sitting on an autumn grass field + +Table 2. Quantitative comparison with different baselines. While higher CLIP scores indicate better target prompt alignment, lower LPIPS scores do not necessarily reflect edit quality since minimal edits yield the lowest scores. Our method well balances source content preservation and target prompt matching, whereas fixed seed generation often produces entirely different images and prompt-change produces overly conservative modifications. + +
ModelFixed SeedPrompt ChangeOurs
LPIPSCLIPLPIPSCLIPLPIPSCLIP
SD3-M0.5940.3770.3250.3440.3800.359
SD3.5-M0.5840.3800.3220.3490.4560.374
SD3.5-L0.5570.3880.3060.3630.4180.377
SD3.5-L-Turbo0.5090.3730.2250.3320.3880.363
Flux.1-dev0.5790.3590.2510.3110.3690.339
Flux.1-schnell0.5080.3800.1160.2890.2790.333
+ +inherent limitations in capturing various types of prompt-based editing quality, we also provide user study result in Fig. A9 and extensive qualitative results in Fig. 15, Fig. A11, and Fig. A12. All experiments use 28 timesteps, except for distilled models (SD3.5-L-Turbo and Flux.1-schnell), which use 4. To ensure fair comparisons and avoid hyperparameter tuning, we fix the attention replacement step to the first $20\%$ of iterations and disable local blending for the results in Tab. 2 and Fig. 15, though our method + +![](images/cd13a724b0f8fc7dc3a0d25833f4428f2d8a8760e444c4466f86b2fc61ebf099.jpg) +Figure 14. Real image editing comparison between our method and RF inversion. Our approach enables more precise local edits while preserving unchanged regions. +Figure 15. Qualitative comparison of our method with baselines. Our method consistently outperforms baseline approaches across diverse editing scenarios. + +can incorporate blending when needed. + +For real-image editing, we observe that editing performance depends more heavily on the inversion technique than the architectural-level editing method. Nonetheless, we report quantitative results on PIE-Bench [27] using RF inversion as our inversion technique and provide extensive qualitative results. Please refer to Appendix A.3 and A.4 for details. Our method provides additional controllability and improves the naive RF inversion baseline. + +# 7. Conclusion + +This work presents a comprehensive analysis of MM-DiT's attention mechanism, revealing several underlying insights, including scaling properties and text-image interactions. Based on these observations, we develop an efficient editing method tailored for MM-DiT architectures that achieves high-quality results across various scenarios, including few-step models and real image editing. We believe these findings help deepen the understanding of emerging architectures and contribute to advancements in controlled image generation and editing. Discussions on limitations and future directions can be found in Appendix F. + +# 8. Acknowledgments + +This work was supported by IITP grant (No. RS-2021-II211343, Artificial Intelligence Graduate School Program at Seoul National University) $(5\%)$ , IITP grant (No. RS-2024-00509257, Global AI Frontier Lab) $(65\%)$ , and NRF grant (No. RS-2024-00405857) $(30\%)$ , funded by the Korea government (MSIT). + +# References + +[1] Anthropic. Claude 3.5 Sonnet, 2023. Large language model developed by Anthropic. 7 +[2] Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. Stable flow: Vital layers for training-free image editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 7, 9 +[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4 +[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 8 +[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2 +[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2 +[7] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-alpha: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1 +[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-delta: Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 2 +[9] Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. FluxSpace: Disentangled Semantic Editing in Rectified Flow Transformers. arXiv preprint arXiv:2412.09611, 2024. 7, 8 +[10] Tri Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. In International Conference on Learning Representations (ICLR), 2024. 5 +[11] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5 + +[12] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision Transformers Need Registers. In International Conference on Learning Representations (ICLR), 2024. 2, 4 +[13] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing. arXiv preprint arXiv:2412.07517, 2024. 2, 8 +[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2 +[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 1, 2 +[16] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. ReNoise: Real Image Inversion Through Iterative Noising. arXiv preprint arXiv:2403.14602, 2024. 2 +[17] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4 +[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In International Conference on Learning Representations (ICLR), 2022. 2, 4 +[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. 7 +[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1 +[22] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In IEEE International Conference on Computer Vision (ICCV), 2023. 2 +[23] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 2 +[25] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual Style Prompting with Swapping Self-Attention. arXiv preprint arXiv:2402.12974, 2024. 2 + +[26] Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang, and Renjie Liao. Uniedit-flow: Unleashing inversion and editing in the era of flow models. arXiv preprint arXiv:2504.13109, 2025. 2 +[27] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code. In International Conference on Learning Representations (ICLR), 2024. 8, 2, 4 +[28] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling Diffusion Models into Conditional GANs. In European Conference on Computer Vision (ECCV), 2024. 2 +[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1 +[30] Black Forest Labs. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.Accessed:November 2024.1 +[31] Benjamin Lefaudeaux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xFormers: A modular and hackable Transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5 +[32] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow Matching for Generative Modeling. In International Conference on Learning Representations (ICLR), 2023. 2 +[33] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4 +[35] Xingchao Liu, Chengyue Gong, et al. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In International Conference on Learning Representations (ICLR), 2023. 2, 1 +[36] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In International Conference on Learning Representations (ICLR), 2024. 2 +[37] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 2 +[38] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2 + +[39] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 1 +[40] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2 +[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 7, 2 +[42] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2 +[43] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), 2021. 1 +[45] OpenAI. GPT-4o, 2024. Large language model developed by OpenAI. 3 +[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 4 +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 5 +[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2 +[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1, 2 +[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2 +[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 2020. 2 +[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1 + +[53] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4 +[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2 +[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 2 +[56] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations. arXiv preprint arXiv:2410.10792, 2024. 2, 7, 1, 4 +[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1 +[58] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In International Conference on Learning Representations (ICLR), 2022. 2 +[59] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial Diffusion Distillation. arXiv e-prints, pages arXiv-2311, 2023. 2 +[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 1 +[61] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. In International Conference on Machine Learning (ICML), 2023. 2 +[62] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4 +[63] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 7, 8 +[64] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2 +[65] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. TurboEdit: Instant text-based + +image editing. In European Conference on Computer Vision (ECCV), 2024. 2 +[66] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. arXiv preprint arXiv:2405.14867, 2024. 2 +[67] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[68] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. Transactions on Machine Learning Research, 2022. 4 +[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7 +[70] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation. arXiv preprint arXiv:2405.01434, 2024. 2 \ No newline at end of file diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/images.zip b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..3fa9b903db06f0bfc5e921088b8534886d235e7b --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:416533a681da9b37c014e4d2bb2719734a55411c34cd138df2d822a3ecd5c46f +size 1036278 diff --git a/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/layout.json b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1b365cbca4a6187deb403c8cf04550e2ded76fb5 --- /dev/null +++ b/2025/Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing/layout.json @@ -0,0 +1,8758 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 151, + 103, + 460, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 103, + 460, + 140 + ], + "spans": [ + { + "bbox": [ + 151, + 103, + 460, + 140 + ], + "type": "text", + "content": "Exploring Multimodal Diffusion Transformers for Enhanced Prompt-based Image Editing" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 117, + 161, + 493, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 117, + 161, + 493, + 190 + ], + "spans": [ + { + "bbox": [ + 117, + 161, + 493, + 190 + ], + "type": "text", + "content": "Joonghyuk Shin Alchan Hwang Yujin Kim Daneul Kim Jaesik Park Seoul National University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "spans": [ + { + "bbox": [ + 133, + 191, + 474, + 204 + ], + "type": "text", + "content": "{joonghyuk, alchan00, yujin.k, carpedkm, jaesik.park}@snu.ac.kr" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 243, + 216, + 321 + ], + "blocks": [ + { + "bbox": [ + 118, + 234, + 155, + 242 + ], + "lines": [ + { + "bbox": [ + 118, + 234, + 155, + 242 + ], + "spans": [ + { + "bbox": [ + 118, + 234, + 155, + 242 + ], + "type": "inline_equation", + "content": "\\mathrm{SD3}\\rightarrow \\mathrm{Flux}" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 243, + 216, + 321 + ], + "lines": [ + { + "bbox": [ + 59, + 243, + 216, + 321 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 216, + 321 + ], + "type": "image", + "image_path": "586f68b5cedfe0bd0388233b015998856dc89a7b250960ca8ef09befd611a68c.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 229, + 243, + 384, + 321 + ], + "blocks": [ + { + "bbox": [ + 282, + 235, + 330, + 243 + ], + "lines": [ + { + "bbox": [ + 282, + 235, + 330, + 243 + ], + "spans": [ + { + "bbox": [ + 282, + 235, + 330, + 243 + ], + "type": "text", + "content": "giraffe " + }, + { + "bbox": [ + 282, + 235, + 330, + 243 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 282, + 235, + 330, + 243 + ], + "type": "text", + "content": " dragon" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 229, + 243, + 384, + 321 + ], + "lines": [ + { + "bbox": [ + 229, + 243, + 384, + 321 + ], + "spans": [ + { + "bbox": [ + 229, + 243, + 384, + 321 + ], + "type": "image", + "image_path": "712320bc04575f5da5148f541a8ba4b5e0ff37b211f266869e7d64527e757bff.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 398, + 242, + 553, + 321 + ], + "blocks": [ + { + "bbox": [ + 453, + 235, + 498, + 242 + ], + "lines": [ + { + "bbox": [ + 453, + 235, + 498, + 242 + ], + "spans": [ + { + "bbox": [ + 453, + 235, + 498, + 242 + ], + "type": "text", + "content": "frogs " + }, + { + "bbox": [ + 453, + 235, + 498, + 242 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 453, + 235, + 498, + 242 + ], + "type": "text", + "content": " rabbits" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 398, + 242, + 553, + 321 + ], + "lines": [ + { + "bbox": [ + 398, + 242, + 553, + 321 + ], + "spans": [ + { + "bbox": [ + 398, + 242, + 553, + 321 + ], + "type": "image", + "image_path": "0e6336a9884ab6b03336fe4c691f10e9a8044a5e8af9e3cbf0f8c17debc3fc6f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 479, + 322, + 515, + 329 + ], + "lines": [ + { + "bbox": [ + 479, + 322, + 515, + 329 + ], + "spans": [ + { + "bbox": [ + 479, + 322, + 515, + 329 + ], + "type": "text", + "content": "ferret " + }, + { + "bbox": [ + 479, + 322, + 515, + 329 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 479, + 322, + 515, + 329 + ], + "type": "text", + "content": " lion" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 59, + 330, + 146, + 430 + ], + "blocks": [ + { + "bbox": [ + 60, + 322, + 144, + 329 + ], + "lines": [ + { + "bbox": [ + 60, + 322, + 144, + 329 + ], + "spans": [ + { + "bbox": [ + 60, + 322, + 144, + 329 + ], + "type": "text", + "content": "'help me' " + }, + { + "bbox": [ + 60, + 322, + 144, + 329 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 60, + 322, + 144, + 329 + ], + "type": "text", + "content": " 'Please don't kill me'" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 330, + 146, + 430 + ], + "lines": [ + { + "bbox": [ + 59, + 330, + 146, + 430 + ], + "spans": [ + { + "bbox": [ + 59, + 330, + 146, + 430 + ], + "type": "image", + "image_path": "193575404c9756b1bbd8970e9ea93db2ad5a38ee8217abc12c561f33284e3e12.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 149, + 330, + 236, + 430 + ], + "blocks": [ + { + "bbox": [ + 149, + 330, + 236, + 430 + ], + "lines": [ + { + "bbox": [ + 149, + 330, + 236, + 430 + ], + "spans": [ + { + "bbox": [ + 149, + 330, + 236, + 430 + ], + "type": "image", + "image_path": "e13d1ee8f091335a98199df420dfc6ec468a43482a825fb608b986bd658b160d.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 437, + 554, + 450 + ], + "lines": [ + { + "bbox": [ + 55, + 437, + 554, + 450 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 554, + 450 + ], + "type": "text", + "content": "Figure 1. Our prompt-based editing examples on SD3 series and Flux.1, demonstrating high-quality results across various editing scenarios." + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 237, + 330, + 323, + 430 + ], + "blocks": [ + { + "bbox": [ + 260, + 322, + 298, + 330 + ], + "lines": [ + { + "bbox": [ + 260, + 322, + 298, + 330 + ], + "spans": [ + { + "bbox": [ + 260, + 322, + 298, + 330 + ], + "type": "text", + "content": "pig " + }, + { + "bbox": [ + 260, + 322, + 298, + 330 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 260, + 322, + 298, + 330 + ], + "type": "text", + "content": " whale" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 237, + 330, + 323, + 430 + ], + "lines": [ + { + "bbox": [ + 237, + 330, + 323, + 430 + ], + "spans": [ + { + "bbox": [ + 237, + 330, + 323, + 430 + ], + "type": "image", + "image_path": "debaa94d5881c0ad10a2168e113198430a00cea96b0b33283a32a6ddab0a90bc.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 324, + 330, + 439, + 430 + ], + "blocks": [ + { + "bbox": [ + 324, + 330, + 439, + 430 + ], + "lines": [ + { + "bbox": [ + 324, + 330, + 439, + 430 + ], + "spans": [ + { + "bbox": [ + 324, + 330, + 439, + 430 + ], + "type": "image", + "image_path": "e98a83cd068015c4f1c364d145ef582726591111e7547abd3850856cbffa9ae5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 440, + 330, + 553, + 430 + ], + "blocks": [ + { + "bbox": [ + 440, + 330, + 553, + 430 + ], + "lines": [ + { + "bbox": [ + 440, + 330, + 553, + 430 + ], + "spans": [ + { + "bbox": [ + 440, + 330, + 553, + 430 + ], + "type": "image", + "image_path": "6c269df0cc808f17a96040deae4eb2afb2a10317ee5746c5606642d5b69831eb.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "spans": [ + { + "bbox": [ + 152, + 460, + 200, + 472 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 53, + 486, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 53, + 486, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 53, + 486, + 297, + 714 + ], + "type": "text", + "content": "Transformer-based diffusion models have recently superseded traditional U-Net architectures, with multimodal diffusion transformers (MM-DiT) emerging as the dominant approach in state-of-the-art models like Stable Diffusion 3 and Flux.1. Previous approaches have relied on unidirectional cross-attention mechanisms, with information flowing from text embeddings to image latents. In contrast, MM-DiT introduces a unified attention mechanism that concatenates input projections from both modalities and performs a single full attention operation, allowing bidirectional information flow between text and image branches. This architectural shift presents significant challenges for existing editing techniques. In this paper, we systematically analyze MM-DiT's attention mechanism by decomposing attention matrices into four distinct blocks, revealing their inherent characteristics. Through these analyses, we propose a robust, prompt-based image editing method for MM-DiT that supports global to local edits across various MM-DiT variants, including few-step models. We believe our findings" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 460, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 556, + 498 + ], + "type": "text", + "content": "bridge the gap between existing U-Net-based methods and emerging architectures, offering deeper insights into MM-DiT's behavioral patterns." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 314, + 523, + 395, + 536 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 523, + 395, + 536 + ], + "spans": [ + { + "bbox": [ + 314, + 523, + 395, + 536 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "spans": [ + { + "bbox": [ + 312, + 544, + 555, + 664 + ], + "type": "text", + "content": "Recent years have witnessed the emergence of diffusion models [21, 60] as the de-facto approach for visual generation, demonstrating remarkable capabilities in creating high-quality data from noise [15, 29, 39, 44, 49, 52, 54, 57]. While early implementations predominantly employed U-shaped networks [55], transformer-based architectures [7, 48] have recently emerged as the prevailing approach, with multimodal diffusion transformers (MM-DiT) showing particular promise in state-of-the-art models like Stable Diffusion 3 series [15] and Flux.1 [30]." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 713 + ], + "type": "text", + "content": "Motivated by MM-DiT's architectural differences and the challenges in directly applying existing U-Net-based editing methods, our work addresses a fundamental question: \"How can we leverage our understanding of U-Net" + } + ] + } + ], + "index": 25 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19492" + } + ] + } + ], + "index": 26 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 155 + ], + "type": "text", + "content": "attention mechanisms to interpret MM-DiT's behavior better and effectively transfer existing techniques to this new architecture?\" Given the significance of existing work built upon U-Net attention mechanisms [4, 5, 19, 33], answering this question is crucial for enabling existing methods to adapt to, and benefit from, emerging MM-DiT architectures." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 157, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 157, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 157, + 294, + 312 + ], + "type": "text", + "content": "To this end, we systematically investigate MM-DiT's attention mechanisms through block matrix analysis, revealing key characteristics of its attention maps. As one of our key analytical insights, we observe an interesting phenomenon aligned with vision transformer [14] scaling laws [12]: as the model size increases, attention patterns become properly positioned but increasingly noisy, necessitating careful selection of attention layers for editing tasks. Building upon these analyses, we propose an efficient editing method at the architecture level that focuses on modifying image tokens in input projections. The proposed editing method can be jointly combined with inversion methods [56, 63] for editing real images as well." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 314, + 294, + 421 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 294, + 421 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 294, + 421 + ], + "type": "text", + "content": "Our contributions are primarily architectural: (1) We provide the first systematic analysis of MM-DiT attention mechanisms, highlighting block-wise patterns and scaling behaviors. (2) Building upon these findings, we propose an efficient and robust editing method across diverse editing scenarios - from global stylistic changes to local edits such as text modification - and various MM-DiT variants, including few-step models (SD3-M, SD3.5-M, SD3.5-L, SD3.5-L-Turbo, and Flux.1-dev/schnell)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 434, + 142, + 447 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 434, + 142, + 447 + ], + "spans": [ + { + "bbox": [ + 55, + 434, + 142, + 447 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 455, + 222, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 455, + 222, + 467 + ], + "spans": [ + { + "bbox": [ + 55, + 455, + 222, + 467 + ], + "type": "text", + "content": "2.1. Text-to-image diffusion models" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 473, + 294, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 294, + 545 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 294, + 545 + ], + "type": "text", + "content": "Early text-to-image diffusion models mostly used U-shaped networks with residual blocks [55] incorporating pixelwise self-attention and text-image cross-attention layers, combined with downsampling and upsampling operations (Fig. 2). These models are typically trained with DDPM-style scheduling and noise prediction objective [49, 54]." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Recent advances introduced two significant shifts. First, architecturally, there has been a transition from U-Nets to transformer-based architectures like DiT [48], demonstrating superior scaling properties. The PixArt-alpha series [6-8] successfully adapted DiT for text conditional generation, maintaining a conventional cross-attention mechanism for text conditioning. Stable Diffusion 3 [15] subsequently introduced a pivotal shift with multimodal diffusion transformers (MM-DiT), using separate transformers for text and image modalities while concatenating their sequences for unified attention operations (Fig. 2). Second, formulation-wise, newer approaches have adopted v-prediction [58] with a rectified flow framework [32, 35] that connects data and noise distributions along straight paths, replacing the com" + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 314, + 144, + 563, + 224 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": "Table 1. Architectural details and inference times for " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "1024 \\times 1024" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " image generation using 28 timesteps, using single A6000 GPU. Times are shown as optimized (left, using PyTorch SDPA kernels) vs. naive matrix multiplication (right). Text encoders: " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^1" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " refers to CLIP ViT-L/14 [50], " + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "inline_equation", + "content": "\\mathrm{CLIP}^2" + }, + { + "bbox": [ + 313, + 71, + 553, + 137 + ], + "type": "text", + "content": " refers to OpenCLIP ViTG/14 [24], and T5 refers to T5-XXL [51]." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 314, + 144, + 563, + 224 + ], + "lines": [ + { + "bbox": [ + 314, + 144, + 563, + 224 + ], + "spans": [ + { + "bbox": [ + 314, + 144, + 563, + 224 + ], + "type": "table", + "html": "
Arch. / ParamsText Enc.Attn TypeTime (s)
SD1U-Net / 860MCLIP1Self/Cross5.0/12.7
SDXLU-Net / 2.6BCLIP1,2Self/Cross5.6/7.9
SD3-MMM-DiT / 2BCLIP1,2 / T5Full7.1/20.5
SD3.5-MMM-DiT-X / 2.2BCLIP1,2 / T5Full/Self9.6/26.8
SD3.5-LMM-DiT / 8BCLIP1,2 / T5Full24.7/60.5
Flux.1-devMM-DiT* / 11.9BCLIP1 / T5Full27.8/36.8
", + "image_path": "b563a8be41d03824ae08c1205479b54ce604d3ff32391cf25221806bdf9bfb6e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 234, + 555, + 319 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 319 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 319 + ], + "type": "text", + "content": "plex trajectories of traditional diffusion models. Together with increased model scales, these developments have significantly improved model performance. Parallel efforts have also accelerated iterative diffusion sampling via faster ODE solvers [37, 38], consistency distillation [40, 61], adversarial learning [28, 59, 66, 67], and trajectory straightening [32, 35, 36]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 324, + 459, + 338 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 459, + 338 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 459, + 338 + ], + "type": "text", + "content": "2.2. Text-guided image editing" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 342, + 553, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 342, + 553, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 342, + 553, + 486 + ], + "type": "text", + "content": "Text-guided image editing naturally extends text-to-image diffusion models. Prior work [18, 22, 33, 64] shows that self-attention governs spatial structure while cross-attention handles attribute-level semantics aligned with text. Prompt-to-Prompt (P2P) [19], a seminal approach, leverages this by directly transferring cross-attention maps for precise edits. Subsequent methods further manipulate self-attention mechanisms for better content-preserving edits [4, 25, 43, 70]. A key advantage of these attention-based methods is their ability to achieve precise local edits solely from textual prompts, eliminating the need for explicit spatial guidance or additional user input." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 486, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 553, + 582 + ], + "type": "text", + "content": "Another direction employs inversion techniques [13, 16, 23, 26, 56, 63], mapping real images to editable latent spaces. While inversion alone can lose fine details, combining it with attention-based methods significantly enhances localized editing [42]. Alternatively, methods like InstructPix2Pix [3] train dedicated models on synthetic data, though this approach incurs high computational costs and dataset preparation overheads [65]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 588, + 421, + 600 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 588, + 421, + 600 + ], + "spans": [ + { + "bbox": [ + 313, + 588, + 421, + 600 + ], + "type": "text", + "content": "2.3. MM-DiT variants" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 715 + ], + "type": "text", + "content": "Since its introduction, MM-DiT architecture has seen several improvements. The initial dual-branch design in SD3 was motivated by the conceptual differences between text and image embeddings. This architecture was maintained in both SD3 and SD3.5-Large models, which contain 24 and 38 dual-branch blocks, respectively. Flux.1, scaled to 12B parameters, introduced a hybrid approach with 57 total blocks - 19 dual branch blocks followed by 38 single branch blocks (Fig. 2). The single branch variant processes" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19493" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 55, + 68, + 286, + 254 + ], + "blocks": [ + { + "bbox": [ + 55, + 68, + 286, + 254 + ], + "lines": [ + { + "bbox": [ + 55, + 68, + 286, + 254 + ], + "spans": [ + { + "bbox": [ + 55, + 68, + 286, + 254 + ], + "type": "image", + "image_path": "5e64765a4f55ddaf639b4e1d4903f79d0681ca0528134398176b45032a3395e5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 265, + 296, + 321 + ], + "lines": [ + { + "bbox": [ + 55, + 265, + 296, + 321 + ], + "spans": [ + { + "bbox": [ + 55, + 265, + 296, + 321 + ], + "type": "text", + "content": "Figure 2. Attention mechanisms across diffusion architectures. (a) U-Net with separate self-/cross- attention, (b) MM-DiT blocks with dual branches, (c) MM-DiT* single branch block used in later layers of Flux.1, and (d) MM-DiT-X dual attention block with explicit self-attention, used in early layers of SD3.5-M." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 336, + 296, + 422 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 336, + 296, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 336, + 296, + 422 + ], + "type": "text", + "content": "concatenated text and image embeddings as a single modality using unified weights (e.g., linear layers, attention parameters), reducing block parameters from 340M to 141M. The MM-DiT-X variant, later introduced in SD3.5-medium, added explicit self-attention operations exclusively for image tokens in the first 13 transformer blocks for better multi-resolution results and overall image coherence." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 437, + 230, + 451 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 437, + 230, + 451 + ], + "spans": [ + { + "bbox": [ + 55, + 437, + 230, + 451 + ], + "type": "text", + "content": "3. Attention Analysis in MM-DiT" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 458, + 296, + 472 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 458, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 458, + 296, + 472 + ], + "type": "text", + "content": "3.1. Block matrix formulation of MM-DiT attention" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "spans": [ + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "content": "We begin the analysis with MM-DiT's attention structure under SD3-M's standard dimensions: " + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "inline_equation", + "content": "64 \\times 64" + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "content": " image latents (flattened to 4096) and 333 text tokens (77 CLIP + 256 T5), with inner dimension 64. Omitting the batch and head dimensions for clarity, the input projections for image and text domains are denoted as " + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "inline_equation", + "content": "(\\mathbf{q}_i, \\mathbf{k}_i, \\mathbf{v}_i \\in \\mathbb{R}^{4096 \\times 64})" + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "inline_equation", + "content": "(\\mathbf{q}_t, \\mathbf{k}_t, \\mathbf{v}_t \\in \\mathbb{R}^{333 \\times 64})" + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "content": ". Then, query, key, and value projections are concatenated as shown below (" + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "inline_equation", + "content": "\\mathbf{q}, \\mathbf{k}, \\mathbf{v} \\in \\mathbb{R}^{4429 \\times 64}" + }, + { + "bbox": [ + 55, + 478, + 296, + 611 + ], + "type": "text", + "content": "). It is worth noting that Flux.1 adopts an inverse concatenation order (text preceding image), which results in a different attention map." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 105, + 621, + 296, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 105, + 621, + 296, + 651 + ], + "spans": [ + { + "bbox": [ + 105, + 621, + 296, + 651 + ], + "type": "interline_equation", + "content": "\\mathbf {q} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right], \\mathbf {k} = \\left[ \\begin{array}{l} \\mathbf {k} _ {i} \\\\ \\mathbf {k} _ {t} \\end{array} \\right], \\mathbf {v} = \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {1}", + "image_path": "0ded79c3c4fcaf8b9622ec0e2b5369f681c0d4b557219d3067d3e5afd46f0aea.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "text", + "content": "Using the standard scaled dot product attention formulation " + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\text{softmax}\\left(\\frac{\\mathbf{q}\\mathbf{k}^T}{\\sqrt{d_k}}\\right)\\mathbf{v}" + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "text", + "content": ", we can write the attention map " + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\mathbf{k}^T" + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "text", + "content": " and its output representation as follows. Note that row-wise softmax normalization is applied to " + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "inline_equation", + "content": "\\mathbf{q}\\mathbf{k}^T" + }, + { + "bbox": [ + 55, + 661, + 297, + 715 + ], + "type": "text", + "content": " in Eq. (2), before" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 318, + 72, + 554, + 148 + ], + "blocks": [ + { + "bbox": [ + 318, + 72, + 554, + 148 + ], + "lines": [ + { + "bbox": [ + 318, + 72, + 554, + 148 + ], + "spans": [ + { + "bbox": [ + 318, + 72, + 554, + 148 + ], + "type": "image", + "image_path": "7c710e26405805760b5aedf9c3f112a2d920ce22b9e18f928da065448a00ab17.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 156, + 555, + 201 + ], + "lines": [ + { + "bbox": [ + 313, + 156, + 555, + 201 + ], + "spans": [ + { + "bbox": [ + 313, + 156, + 555, + 201 + ], + "type": "text", + "content": "Figure 3. Visualizing the top-6 principal components extracted from I2I attention blocks, averaged across all timesteps and transformer blocks. Like U-Net's self-attention layers, these components reveal underlying spatial and geometric patterns." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 212, + 473, + 224 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 212, + 473, + 224 + ], + "spans": [ + { + "bbox": [ + 313, + 212, + 473, + 224 + ], + "type": "text", + "content": "multiplication with " + }, + { + "bbox": [ + 313, + 212, + 473, + 224 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 212, + 473, + 224 + ], + "type": "text", + "content": " to obtain Eq. (3)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 230, + 560, + 269 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 230, + 560, + 269 + ], + "spans": [ + { + "bbox": [ + 314, + 230, + 560, + 269 + ], + "type": "interline_equation", + "content": "\\mathbf {q} \\mathbf {k} ^ {T} = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\\\ \\mathbf {q} _ {t} \\end{array} \\right] \\left[ \\begin{array}{l l} \\mathbf {k} _ {i} ^ {T} & \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\sim \\left[ \\begin{array}{l l} \\mathrm {I 2 I} & \\mathrm {T 2 I} \\\\ \\mathrm {I 2 T} & \\mathrm {T 2 T} \\end{array} \\right], \\tag {2}", + "image_path": "4f14b44df16bb851de7908358017f41e7e9bb400b6da59e68363c1c06112de3f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 315, + 271, + 554, + 310 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 271, + 554, + 310 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 554, + 310 + ], + "type": "interline_equation", + "content": "\\mathbf {q} \\mathbf {k} ^ {T} \\mathbf {v} = \\left[ \\begin{array}{l l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} & \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\end{array} \\right] \\left[ \\begin{array}{l} \\mathbf {v} _ {i} \\\\ \\mathbf {v} _ {t} \\end{array} \\right] = \\left[ \\begin{array}{l} \\mathbf {q} _ {i} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {i} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\\\ \\mathbf {q} _ {t} \\mathbf {k} _ {i} ^ {T} \\mathbf {v} _ {i} + \\mathbf {q} _ {t} \\mathbf {k} _ {t} ^ {T} \\mathbf {v} _ {t} \\end{array} \\right]. \\tag {3}", + "image_path": "476c4baa4096ef457acf9373b8fff072c143f5e2e3449e600e973044ca0bb745.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 318, + 555, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 318, + 555, + 402 + ], + "spans": [ + { + "bbox": [ + 313, + 318, + 555, + 402 + ], + "type": "text", + "content": "Since the upper and lower rows in Eq. (3) correspond to the resultant image and text latents, respectively, this notation choice naturally reflects the underlying behavior of concatenated attention. Within this generalized framework, our empirical findings suggest that I2I and T2I blocks can serve roles similar to the conventional self- and cross-attention patterns in U-Net architectures." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 409, + 474, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 409, + 474, + 422 + ], + "spans": [ + { + "bbox": [ + 313, + 409, + 474, + 422 + ], + "type": "text", + "content": "3.2. Block-wise attention patterns" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 554, + 510 + ], + "type": "text", + "content": "I2I & T2T. I2I and T2T are unimodal portions of the attention map. As demonstrated by PCA analysis in Fig. 3, the I2I block is analogous to self-attention in U-Net architectures, capturing spatial layout and geometric information. In contrast, T2T blocks predominantly manifest as identity matrices, with attention weights notably weakening after the occurrence of meaningful tokens (e.g., EOS)." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 713 + ], + "type": "text", + "content": "T2I & I2T. T2I and I2T blocks represent cross-modal interactions, where T2I influences subsequent image latents, and I2T affects text representations - with I2T being a novel addition not present in U-Net architectures. These blocks are essential for precise local editing as they encode token-specific image-region correspondences, generating binary masks for precise local blending. While both blocks can be utilized to obtain these masks, we find T2I more effective (Fig. 4). Specifically, due to the row-wise softmax operation in I2T blocks, where the rows correspond to image tokens, there is inherent competition among these tokens since their attention weights must sum to 1. This constraint dilutes the attention signals, limiting their ability to represent strong or broad influences. In contrast, T2I's structure allows multiple image regions to retain high attention values simultaneously. Additionally, for the SD3 model series using both CLIP and T5 encoders, we observe that T5-derived" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "19494" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 56, + 71, + 294, + 177 + ], + "blocks": [ + { + "bbox": [ + 56, + 71, + 294, + 177 + ], + "lines": [ + { + "bbox": [ + 56, + 71, + 294, + 177 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 294, + 177 + ], + "type": "image", + "image_path": "644646cbf21ca17501b775f4f75f830ed2d9c76e3a905ac29a4f3db7d6da5c56.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 184, + 295, + 240 + ], + "lines": [ + { + "bbox": [ + 55, + 184, + 295, + 240 + ], + "spans": [ + { + "bbox": [ + 55, + 184, + 295, + 240 + ], + "type": "text", + "content": "Figure 4. Comparison of Flux.1-dev's T2I and I2T attention blocks averaged across all timesteps and transformer blocks. T2I blocks show stronger localization due to their unconstrained attention distribution compared to I2T's row-wise competition, making them more effective source for localized editing." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 247, + 296, + 407 + ], + "blocks": [ + { + "bbox": [ + 58, + 247, + 296, + 407 + ], + "lines": [ + { + "bbox": [ + 58, + 247, + 296, + 407 + ], + "spans": [ + { + "bbox": [ + 58, + 247, + 296, + 407 + ], + "type": "image", + "image_path": "194ae3a0fdc66c05ced6a36e4a70dcc6a00539c4e502b06dfd0dd142de2d9629.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 417, + 296, + 506 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 296, + 506 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 296, + 506 + ], + "type": "text", + "content": "Figure 5. Analysis of T2I attention blocks from SDXL and various MM-DiT models using the prompt \"House on a mountainside under a clear blue daylight sky\" (averaged across timesteps and transformer blocks). MM-DiT models show better localized attention compared to U-Net architectures, though attention noise increases with model scale. When editing \"blue daylight\" to \"red dawn\" by replacing entire attention blocks, MM-DiT's precise attention enables clean edits, while SDXL applies the change globally." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 521, + 295, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 521, + 295, + 533 + ], + "spans": [ + { + "bbox": [ + 55, + 521, + 295, + 533 + ], + "type": "text", + "content": "attention maps typically exhibit higher precision than CLIP." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 295, + 713 + ], + "type": "text", + "content": "Based on these analyses, we observe that I2I blocks are crucial for preserving identity, and T2I blocks are effective in obtaining attention masks for precise targeted local editing. We also validate the relative importance of these four sub-blocks in Fig. A2, which highlights the effectiveness of the I2I block in preserving original image attributes. Furthermore, visualizations of T2I attention in Fig. A14, Fig. A15, Fig. A16, Fig. A17 suggest that, during full attention's modality mixing, rotary positional embeddings (RoPE) [62] and residual connections help each modality retain its distinctive characteristics at designated position throughout transformer blocks. This supports and justifies our block-based interpretation and editing. Due to space constraints, we present extensive additional findings in Appendix B, to which we refer readers for further details." + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 317, + 72, + 553, + 163 + ], + "blocks": [ + { + "bbox": [ + 317, + 72, + 553, + 163 + ], + "lines": [ + { + "bbox": [ + 317, + 72, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 317, + 72, + 553, + 163 + ], + "type": "image", + "image_path": "20250207dc019223b5a20ab40312e60b66eae532ff043e14ef184da6063f5c7b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 175, + 555, + 242 + ], + "lines": [ + { + "bbox": [ + 313, + 175, + 555, + 242 + ], + "spans": [ + { + "bbox": [ + 313, + 175, + 555, + 242 + ], + "type": "text", + "content": "Figure 6. Visualization of Flux.1-dev's T2I block, averaged across timesteps, comparing masks from top-5 selected blocks (upper) versus all blocks (lower), both with and without smoothing. Block selection significantly reduces noise, while smoothing refines boundary transitions. The lower left displays Grounded SAM2's masks used as ground truth for top-5 block selection." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 260, + 487, + 273 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 487, + 273 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 487, + 273 + ], + "type": "text", + "content": "3.3. Mitigating noisy attention maps" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 279, + 555, + 482 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 555, + 482 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 555, + 482 + ], + "type": "text", + "content": "We compare cross-attention maps (T2I blocks for MM-DiTs) across various models in Fig. 5. While MM-DiT architectures exhibit significantly improved attention localization, we observe an interesting phenomenon: as MM-DiT model size grows, the attention maps become increasingly noisy. Directly employing these noisy, discontinuous attention maps as local blending masks introduce visible artifacts in generated images. This issue aligns with recent findings by [12], which report similar noise artifacts emerging in ViTs at larger model and dataset scales. While they propose to add a register (additional tokens) to remove these artifacts, it is impractical here due to the prohibitive cost of retraining large MM-DiT models. Thus, we propose two practical solutions: (1) selecting optimal transformer blocks that inherently produce more precise and less noisy attention maps, and (2) applying Gaussian smoothing to smooth mask boundaries and reduce artifacts." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 556, + 664 + ], + "type": "text", + "content": "For systematic evaluation, we sampled 100 prompts from the PARTI prompts [68], a widely used prompt set for benchmarking, and we created ground truth masks using Grounded SAM2 [34, 53] for grounded segmentation tasks, as shown in Fig. 6. Intuitively, effective blocks should yield attention maps similar to the GT masks. Thus, using standard losses and measures (BCE, Soft mIoU, and MSE) against the binary GT masks, we selected the top-5 blocks for each model based on the average ranking. Blockwise metrics for Flux.1 are presented in Fig. 7, and specific rankings are provided in Tab. A2. We emphasize that these blocks are not prompt-specific, as they were calculated over 100 random prompts, and we consistently use the same five blocks throughout the paper. Local editing results with and without these selected blocks are presented in Fig. 8." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Notably, applying Gaussian smoothing alters the block ranking in larger models, suggesting that some noisy blocks still contain meaningful information. In contrast, smaller models like SD3-2B, which inherently produce smoother" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19495" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 137, + 168 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 137, + 168 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 137, + 168 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 137, + 168 + ], + "type": "image", + "image_path": "58356bc12b89ef55c1b356719efa76f9dcb10d3fef9d5a3ff56a850bbd621342.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 139, + 70, + 216, + 168 + ], + "blocks": [ + { + "bbox": [ + 139, + 70, + 216, + 168 + ], + "lines": [ + { + "bbox": [ + 139, + 70, + 216, + 168 + ], + "spans": [ + { + "bbox": [ + 139, + 70, + 216, + 168 + ], + "type": "image", + "image_path": "40552c483fa6f7e86fc1ecaaf03b3c70f1cdc48b875dbea7bb187e28c94e5186.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 220, + 70, + 294, + 168 + ], + "blocks": [ + { + "bbox": [ + 220, + 70, + 294, + 168 + ], + "lines": [ + { + "bbox": [ + 220, + 70, + 294, + 168 + ], + "spans": [ + { + "bbox": [ + 220, + 70, + 294, + 168 + ], + "type": "image", + "image_path": "58b6675bab5ca2f98899159bcaab018bb2921014ca9535304e77fe197195f5f8.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 56, + 170, + 138, + 267 + ], + "blocks": [ + { + "bbox": [ + 56, + 170, + 138, + 267 + ], + "lines": [ + { + "bbox": [ + 56, + 170, + 138, + 267 + ], + "spans": [ + { + "bbox": [ + 56, + 170, + 138, + 267 + ], + "type": "image", + "image_path": "a637d3b3b2454f55fc3bcecc7c79e2f153766ad464605424ec5e8008eb69ed48.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 276, + 296, + 365 + ], + "lines": [ + { + "bbox": [ + 55, + 276, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 276, + 296, + 365 + ], + "type": "text", + "content": "Figure 7. Transformer block analysis of Flux.1-dev using Binary Cross Entropy Loss, Soft mIoU, and MSE, with Grounded SAM2 predictions as ground truth. Scores are shown without (upper) and with (lower) Gaussian smoothing, with all evaluation metrics showing notable improvements after smoothing. T2I shows higher accuracy than I2T across most blocks. Top-5 ranked T2I attention blocks are exclusively used for generating local edit masks. Additional model results are provided in Appendix C.1." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 140, + 170, + 217, + 268 + ], + "blocks": [ + { + "bbox": [ + 140, + 170, + 217, + 268 + ], + "lines": [ + { + "bbox": [ + 140, + 170, + 217, + 268 + ], + "spans": [ + { + "bbox": [ + 140, + 170, + 217, + 268 + ], + "type": "image", + "image_path": "4231d97e06240f648447c1d7ff2e5f678232a815dc14af09e01a15b367556056.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 221, + 170, + 296, + 268 + ], + "blocks": [ + { + "bbox": [ + 221, + 170, + 296, + 268 + ], + "lines": [ + { + "bbox": [ + 221, + 170, + 296, + 268 + ], + "spans": [ + { + "bbox": [ + 221, + 170, + 296, + 268 + ], + "type": "image", + "image_path": "36b9fe7a4b019eb29f4780af12375ac1fe651749a470e62d133ab0cbbe5d5851.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 56, + 371, + 294, + 434 + ], + "blocks": [ + { + "bbox": [ + 56, + 371, + 294, + 434 + ], + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 434 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 434 + ], + "type": "image", + "image_path": "62dbd1448bd0e8b41691f43b06bda3d766816657d1b1ac44482e309732c88dd8.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 444, + 296, + 499 + ], + "lines": [ + { + "bbox": [ + 55, + 444, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 444, + 296, + 499 + ], + "type": "text", + "content": "Figure 8. Local image editing results using masks derived from the top-5 transformer blocks with Gaussian smoothing. Employing only the top-5 blocks reduces noise in the attention masks, effectively preserving untouched regions (e.g., music notes). Applying Gaussian smoothing further reduces boundary artifacts." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 296, + 713 + ], + "type": "text", + "content": "attention maps, exhibit stable block rankings regardless of Gaussian smoothing. Unlike previous U-Net observations, where down-block features are consistently noisier than upblock ones, we observed no strong common trends across MM-DiT variants due to their distinct architectural differences. Nonetheless, very early and later blocks consistently produced noisier attention maps across all models. Thus, simply excluding these noisy blocks and using the identified top-5 blocks consistently led to robust results. Another practical advantage of selecting only the top-5 blocks is improved computational efficiency. Explicitly computing the full T2I attention map prohibits using optimized attention computations such as PyTorch's dedicated SDPA kernel [10, 11, 31, 47]. By limiting manual full attention computations to just the top-5 blocks, the remaining blocks can leverage these optimized kernels, significantly enhancing inference efficiency, as further discussed in Sec. 4.2." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 348, + 72, + 523, + 437 + ], + "blocks": [ + { + "bbox": [ + 348, + 72, + 523, + 437 + ], + "lines": [ + { + "bbox": [ + 348, + 72, + 523, + 437 + ], + "spans": [ + { + "bbox": [ + 348, + 72, + 523, + 437 + ], + "type": "image", + "image_path": "eb89e1019389df6054e71642085ee3dab8c5de12eea31302d664a3a35baffafc.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "text", + "content": "Figure 9. Architectural illustration of our method. We replace target (editing) branch projections " + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "text", + "content": " with their source branch counterparts, " + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i^{\\mathrm{src}}" + }, + { + "bbox": [ + 313, + 444, + 555, + 521 + ], + "type": "text", + "content": ". For local blending, we store the T2I portion of unmodified attention maps from selected blocks in both branches during early timesteps. A binary mask is computed from the union of these attention maps after thresholding and used to blend the two latent images. Best viewed zoomed in." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 531, + 487, + 544 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 531, + 487, + 544 + ], + "spans": [ + { + "bbox": [ + 313, + 531, + 487, + 544 + ], + "type": "text", + "content": "4. Editing via MM-DiT Attention" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 552, + 554, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 552, + 554, + 574 + ], + "spans": [ + { + "bbox": [ + 313, + 552, + 554, + 574 + ], + "type": "text", + "content": "4.1. Attention map-based and input projection-based methods" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "content": "Our analysis reveals that MM-DiT attention components serve roles similar to their U-Net counterparts: I2I blocks preserve source attributes (self-attention-like), while T2I blocks localize text-image interactions (cross-attention-like). Inspired by previous studies, we demonstrate image editing by substituting the target branch's attention map with the source branch's during early denoising timesteps. For common rectified flow schedulers (Euler's method), replacing attention maps during the initial " + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 581, + 555, + 700 + ], + "type": "text", + "content": " of timesteps yields effective results." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 713 + ], + "type": "text", + "content": "A key consideration is now determining which attention" + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19496" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 72, + 295, + 194 + ], + "blocks": [ + { + "bbox": [ + 57, + 72, + 295, + 194 + ], + "lines": [ + { + "bbox": [ + 57, + 72, + 295, + 194 + ], + "spans": [ + { + "bbox": [ + 57, + 72, + 295, + 194 + ], + "type": "image", + "image_path": "af54c4eba972f5600a0df8148d2223dc6c0c517d4bd262515543bd5256f24f58.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "lines": [ + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "text", + "content": "Figure 10. Effect of text misalignment on image editing using Flux.1-dev model. When editing with highly different prompts (full list in the Appendix), replacing only image projections " + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "inline_equation", + "content": "(\\mathbf{q}_i, \\mathbf{k}_i)" + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "text", + "content": " maintains better stability compared to full attention map replacement (all " + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "inline_equation", + "content": "\\mathbf{q}, \\mathbf{k}" + }, + { + "bbox": [ + 55, + 201, + 296, + 258 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 59, + 262, + 296, + 347 + ], + "blocks": [ + { + "bbox": [ + 59, + 262, + 296, + 347 + ], + "lines": [ + { + "bbox": [ + 59, + 262, + 296, + 347 + ], + "spans": [ + { + "bbox": [ + 59, + 262, + 296, + 347 + ], + "type": "image", + "image_path": "5decd7089e2f8a518b07d8476782361f626cb14635084a0b35da774730b0f930.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "lines": [ + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "spans": [ + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "type": "text", + "content": "Figure 11. Comparison between replacing " + }, + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 55, + 355, + 296, + 389 + ], + "type": "text", + "content": " projections and I2I block replacement. Results are highly similar, with projection-based methods being more computationally efficient." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "spans": [ + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "content": "components to replace. While substituting the entire attention map (equivalently, full " + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{q}" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{k}" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "content": " projections) is functional, it introduces text projection misalignment. Specifically, replacing the entire target attention map causes the text region to shift to the source branch's contexts (" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_t^{\\mathrm{src}} \\times \\mathbf{k}_t^{\\mathrm{src}}" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "content": "). When this altered attention is multiplied by the target branch's value matrix text tokens (" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 54, + 402, + 295, + 545 + ], + "type": "text", + "content": "), misalignment arises. With T5 embeddings, subtle prompt differences can significantly amplify these misalignments, leading to undesirable image shifts (illustrated in Fig. 10). This establishes a critical constraint: T2T regions must remain unmodified. In-depth explanation of this issue can be found in Appendix B.4." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "As a solution, we propose two alternative approaches: either selectively replacing only the I2I attention block, or specifically modifying just the image input projections " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{q}_i," + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i)" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": ". Based on Eq. (2), these approaches slightly differ, as modifying " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " also influences T2I and I2T interactions. However, as shown in Fig. 11, results from both methods are nearly identical. Importantly, modifying at the input projection level " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "(\\mathbf{q}_i,\\mathbf{k}_i)" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " is computationally more efficient since it enables optimized scaled dot product attention (SDPA) kernels. In contrast, replacing the I2I block disables SDPA, requiring manual attention calculation and careful switching before softmax to maintain proper attention distributions. This absence of SDPA can increase the computational time by up to " + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "inline_equation", + "content": "3\\times" + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": " (Tab. 1)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 320, + 75, + 520, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 75, + 520, + 85 + ], + "spans": [ + { + "bbox": [ + 320, + 75, + 520, + 85 + ], + "type": "text", + "content": "Algorithm 1: Prompt-based Image Editing with MM-DiT" + } + ] + } + ], + "index": 6 + }, + { + "type": "code", + "bbox": [ + 318, + 87, + 539, + 216 + ], + "blocks": [ + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "lines": [ + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "spans": [ + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": "Parameters: Source prompt " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{src}}" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": ", Target prompt " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "P^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": ", Attention replace step " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": ", Local Blending-{replace step " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": ", threshold " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": "} \n" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "x_{T}^{\\mathrm{src}} \\sim \\mathcal{N}(0, I)" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": "; " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "x_{T}^{\\mathrm{tgt}} \\gets x_{T}^{\\mathrm{src}}" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": " \nfor " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "t = T, T - 1, \\ldots, 1" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": " do \nif local blend and " + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "t > \\eta" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": " then // Local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}, \\text{AtttnMap}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{src}} \\text{ if } t > \\tau\\right] \\\\\n\\text{Mask} \\gets (\\text{AtttnMap} > \\theta) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets x_{t-1}^{\\mathrm{tgt}} \\cdot \\text{Mask} + x_{t-1}^{\\mathrm{src}} \\cdot (1 - \\text{Mask})\n\\end{array}\n\\] \nelse // No local blend \n\\[\n\\begin{array}{l}\nx_{t-1}^{\\mathrm{src}}, (\\{q, k\\}_{i,t}^{\\mathrm{src}}) \\leftarrow \\mathrm{DM}(x_{t}^{\\mathrm{src}}, P^{\\mathrm{src}}, t) \\\\\nx_{t-1}^{\\mathrm{tgt}} \\gets \\mathrm{DM}(x_{t}^{\\mathrm{tgt}}, P^{\\mathrm{tgt}}, t) \\left[\\{q, k\\}_{i,t}^{\\mathrm{tgt}} \\gets \\{q, k\\}_{i,t}^{\\mathrm{sucr}} \\text{ if } t > \\tau\\right]\n\\end{array}\n\\] \nreturn (Decode(" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{src}}" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": "), Decode(" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "inline_equation", + "content": "x_0^{\\mathrm{tgt}}" + }, + { + "bbox": [ + 318, + 87, + 539, + 216 + ], + "type": "text", + "content": "))" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "code_body" + } + ], + "index": 7, + "sub_type": "algorithm" + }, + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "text", + "content": "To this end, we propose a new approach based on " + }, + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 313, + 242, + 555, + 399 + ], + "type": "text", + "content": " replacement. Unlike prior U-Net-based attention modification methods that require a precise token mapping between source and target prompts, limiting their effectiveness when prompts differ significantly, our method can handle arbitrarily different prompt pairs. This approach effectively realizes our editing objective: maintaining base image similarity while incorporating edited prompt elements, a capability uniquely enabled by MM-DiT's full-attention architecture. By preserving the text region intact, our method eliminates the need for explicit token correspondence or specialized mappers between prompts, allowing robust performance across diverse editing scenarios." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 407, + 407, + 420 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 407, + 420 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 407, + 420 + ], + "type": "text", + "content": "4.2. Local blending" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 426, + 555, + 556 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 556 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 556 + ], + "type": "text", + "content": "Using insights from Sec. 3.3, we extract clear attention maps from the top-5 transformer blocks for precise local blending, ensuring only targeted regions change while preserving the base image elsewhere. We apply binary blending masks derived from these attention maps once per timestep after all blocks are processed, following P2P's method. This selective calculation allows optimized SDPA kernels in remaining blocks, maintaining inference speed similar to naive batch size-2 inference (SD3-M: 15.2s vs. 14.9s, Flux.1-dev: 55.9s vs. 53.7s, SD3.5-L: 50.1s vs. 47.5s on a single A6000 GPU)." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": "Our complete pipeline is illustrated in Fig. 9, with detailed pseudocode provided in Algorithm 1. Among the three hyperparameters controlling our method, such as masking threshold " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", attention replace timestep " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", and local blending stop timestep " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", we find that setting " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " (i.e., " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\tau = 0.8T" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\eta = 0.5T" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": ") of total timesteps typically achieves robust performance. Thus, the masking threshold " + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 558, + 556, + 715 + ], + "type": "text", + "content": " remains a primary parameter influencing edit granularity. As demonstrated in Fig. 12, higher thresholds effectively enable precise local edits, while lower thresholds allow broader modifications. Leveraging MM-DiT's precise attention localization and our attention map curation, the proposed local blending method significantly im" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19497" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 85, + 73, + 526, + 242 + ], + "blocks": [ + { + "bbox": [ + 85, + 73, + 526, + 242 + ], + "lines": [ + { + "bbox": [ + 85, + 73, + 526, + 242 + ], + "spans": [ + { + "bbox": [ + 85, + 73, + 526, + 242 + ], + "type": "image", + "image_path": "9d1d616f17bc0fed6e7f584e6963029c7370686c208a3454f9b12d5310428db1.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": "Figure 12. Ablation results on local blending threshold. Fine edits like text modifications benefit from higher thresholds, whereas broader transformations like large object changes favor lower thresholds. Local blending was applied to the first " + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "inline_equation", + "content": "50\\%" + }, + { + "bbox": [ + 55, + 249, + 555, + 273 + ], + "type": "text", + "content": " of timestep iterations." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 285, + 295, + 309 + ], + "type": "text", + "content": "proves edit precision compared to results obtained without attention curation (Fig. A7)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 316, + 282, + 330 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 282, + 330 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 282, + 330 + ], + "type": "text", + "content": "4.3. Controlling edit strength via block selection" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "spans": [ + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "text", + "content": "While injecting input projections or attention maps into all blocks is straightforward, we observe that later blocks have a stronger influence on the final output, producing edited images closely resembling the source. While this correlation aligns with our goal, it can pose issues for few-step models like Flux.1-schnell, where editing all blocks in a single step of a 4-step inference may yield outputs overly similar to the base image, regardless of the target prompt. To address this, we replace only the first 38 and 30 blocks for Flux.1-schnell and SD3.5-L-Turbo, respectively, while replacing every block for non-distilled models. Note that our two block-selection strategies (top-5 blocks for obtaining attention maps and blocks for " + }, + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 54, + 334, + 295, + 526 + ], + "type": "text", + "content": " projections replacement) are not considered as hyperparameters and remain fixed. More details and ablations are available in Appendix C.2 and Fig. A8." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 537, + 174, + 551 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 537, + 174, + 551 + ], + "spans": [ + { + "bbox": [ + 55, + 537, + 174, + 551 + ], + "type": "text", + "content": "5. Editing Real Images" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": "A common approach in existing real image editing methods [2, 9, 56, 63] is to invert the source image into an initial noisy latent and then generate the target image from this inverted latent using an editing prompt. Our editing technique, which operates via " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{q}_i" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{k}_i" + }, + { + "bbox": [ + 55, + 558, + 296, + 714 + ], + "type": "text", + "content": " replacement, functions in parallel with existing methods and can be readily applied in this scenario. Fig. 13 illustrates this process, and a preview of our results can be seen in Fig. 14. In Appendix A, we present further discussions, including quantitative and qualitative comparisons, using RF inversion as a baseline. Generally, the performance of real image editing methods depends heavily on inversion quality, as the initial latent affects background consistency and overall image quality." + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 321, + 285, + 547, + 415 + ], + "blocks": [ + { + "bbox": [ + 321, + 285, + 547, + 415 + ], + "lines": [ + { + "bbox": [ + 321, + 285, + 547, + 415 + ], + "spans": [ + { + "bbox": [ + 321, + 285, + 547, + 415 + ], + "type": "image", + "image_path": "fd726ff8904cf8f96cac1380fa8dfd6825a2c89e13db6ff08cea4fab8f5433d4.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "lines": [ + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "spans": [ + { + "bbox": [ + 313, + 426, + 555, + 471 + ], + "type": "text", + "content": "Figure 13. Illustration of our real image editing method given an initial latent. Unlike the synthetic image editing case, the source branch is fixed as a straight conditional path from the initial noise to the source image, independent of model evaluation." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 479, + 554, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 554, + 563 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 554, + 563 + ], + "type": "text", + "content": "While this dependency also exists for our method, the impact of the initial latent is partially mitigated by continuously injecting input projections from the noisy source image and performing local blending to preserve unaffected regions. Consequently, our method can be used without inversion, starting from random noise, offering a trade-off between computational efficiency and performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 574, + 395, + 587 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 574, + 395, + 587 + ], + "spans": [ + { + "bbox": [ + 313, + 574, + 395, + 587 + ], + "type": "text", + "content": "6. Experiments" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": "Due to the lack of dedicated architecture-level editing methods for MM-DiT, we validate our approach against two simple baselines: (1) direct generation with fixed seeds, and (2) prompt switching after " + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 312, + 594, + 555, + 715 + ], + "type": "text", + "content": " of timesteps, similar to SDEdit [41]. We randomly sample 60 diverse prompts from the PARTI dataset and use LLM [1] to generate edited prompts consisting of 30 involving simple object or word edits and 30 involving complex changes affecting multiple prompt components. We evaluate editing quality using LPIPS [69] and CLIPScore [20]. As these metrics have" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19498" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 71, + 296, + 343 + ], + "blocks": [ + { + "bbox": [ + 57, + 71, + 296, + 343 + ], + "lines": [ + { + "bbox": [ + 57, + 71, + 296, + 343 + ], + "spans": [ + { + "bbox": [ + 57, + 71, + 296, + 343 + ], + "type": "image", + "image_path": "7ebb9d1f4aa687fbfb9f7f7688896c4094b72cf0c0e19153296b8e37fc6d5370.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 91, + 344, + 259, + 354 + ], + "lines": [ + { + "bbox": [ + 91, + 344, + 259, + 354 + ], + "spans": [ + { + "bbox": [ + 91, + 344, + 259, + 354 + ], + "type": "text", + "content": "A Golden Retriever sitting on an autumn grass field" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 411, + 296, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 411, + 296, + 489 + ], + "spans": [ + { + "bbox": [ + 55, + 411, + 296, + 489 + ], + "type": "text", + "content": "Table 2. Quantitative comparison with different baselines. While higher CLIP scores indicate better target prompt alignment, lower LPIPS scores do not necessarily reflect edit quality since minimal edits yield the lowest scores. Our method well balances source content preservation and target prompt matching, whereas fixed seed generation often produces entirely different images and prompt-change produces overly conservative modifications." + } + ] + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 498, + 297, + 592 + ], + "blocks": [ + { + "bbox": [ + 56, + 498, + 297, + 592 + ], + "lines": [ + { + "bbox": [ + 56, + 498, + 297, + 592 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 297, + 592 + ], + "type": "table", + "html": "
ModelFixed SeedPrompt ChangeOurs
LPIPSCLIPLPIPSCLIPLPIPSCLIP
SD3-M0.5940.3770.3250.3440.3800.359
SD3.5-M0.5840.3800.3220.3490.4560.374
SD3.5-L0.5570.3880.3060.3630.4180.377
SD3.5-L-Turbo0.5090.3730.2250.3320.3880.363
Flux.1-dev0.5790.3590.2510.3110.3690.339
Flux.1-schnell0.5080.3800.1160.2890.2790.333
", + "image_path": "1de2a55d7efbdbdf4218e47fc964679cff10a837a20f0eca949384bd6ae85524.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": "inherent limitations in capturing various types of prompt-based editing quality, we also provide user study result in Fig. A9 and extensive qualitative results in Fig. 15, Fig. A11, and Fig. A12. All experiments use 28 timesteps, except for distilled models (SD3.5-L-Turbo and Flux.1-schnell), which use 4. To ensure fair comparisons and avoid hyperparameter tuning, we fix the attention replacement step to the first " + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "inline_equation", + "content": "20\\%" + }, + { + "bbox": [ + 55, + 605, + 296, + 713 + ], + "type": "text", + "content": " of iterations and disable local blending for the results in Tab. 2 and Fig. 15, though our method" + } + ] + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 315, + 69, + 555, + 392 + ], + "blocks": [ + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "lines": [ + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "spans": [ + { + "bbox": [ + 55, + 368, + 296, + 402 + ], + "type": "text", + "content": "Figure 14. Real image editing comparison between our method and RF inversion. Our approach enables more precise local edits while preserving unchanged regions." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 315, + 69, + 555, + 392 + ], + "lines": [ + { + "bbox": [ + 315, + 69, + 555, + 392 + ], + "spans": [ + { + "bbox": [ + 315, + 69, + 555, + 392 + ], + "type": "image", + "image_path": "cd13a724b0f8fc7dc3a0d25833f4428f2d8a8760e444c4466f86b2fc61ebf099.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 399, + 555, + 433 + ], + "lines": [ + { + "bbox": [ + 313, + 399, + 555, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 399, + 555, + 433 + ], + "type": "text", + "content": "Figure 15. Qualitative comparison of our method with baselines. Our method consistently outperforms baseline approaches across diverse editing scenarios." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 441, + 473, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 441, + 473, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 441, + 473, + 453 + ], + "type": "text", + "content": "can incorporate blending when needed." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 555, + 548 + ], + "type": "text", + "content": "For real-image editing, we observe that editing performance depends more heavily on the inversion technique than the architectural-level editing method. Nonetheless, we report quantitative results on PIE-Bench [27] using RF inversion as our inversion technique and provide extensive qualitative results. Please refer to Appendix A.3 and A.4 for details. Our method provides additional controllability and improves the naive RF inversion baseline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "spans": [ + { + "bbox": [ + 313, + 561, + 388, + 573 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 715 + ], + "type": "text", + "content": "This work presents a comprehensive analysis of MM-DiT's attention mechanism, revealing several underlying insights, including scaling properties and text-image interactions. Based on these observations, we develop an efficient editing method tailored for MM-DiT architectures that achieves high-quality results across various scenarios, including few-step models and real image editing. We believe these findings help deepen the understanding of emerging architectures and contribute to advancements in controlled image generation and editing. Discussions on limitations and future directions can be found in Appendix F." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19499" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 165, + 85 + ], + "type": "text", + "content": "8. Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "This work was supported by IITP grant (No. RS-2021-II211343, Artificial Intelligence Graduate School Program at Seoul National University) " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "(5\\%)" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ", IITP grant (No. RS-2024-00509257, Global AI Frontier Lab) " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "(65\\%)" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ", and NRF grant (No. RS-2024-00405857) " + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "inline_equation", + "content": "(30\\%)" + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": ", funded by the Korea government (MSIT)." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 195, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 61, + 195, + 295, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 195, + 295, + 217 + ], + "spans": [ + { + "bbox": [ + 61, + 195, + 295, + 217 + ], + "type": "text", + "content": "[1] Anthropic. Claude 3.5 Sonnet, 2023. Large language model developed by Anthropic. 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 219, + 296, + 273 + ], + "type": "text", + "content": "[2] Omri Avrahami, Or Patashnik, Ohad Fried, Egor Nemchinov, Kfir Aberman, Dani Lischinski, and Daniel Cohen-Or. Stable flow: Vital layers for training-free image editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 7, 9" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 62, + 275, + 295, + 319 + ], + "type": "text", + "content": "[3] Tim Brooks, Aleksander Holynski, and Alexei A Efros. Instructpix2pix: Learning to follow image editing instructions. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2, 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "spans": [ + { + "bbox": [ + 62, + 320, + 295, + 374 + ], + "type": "text", + "content": "[4] Mingdeng Cao, Xintao Wang, Zhongang Qi, Ying Shan, Xiaohu Qie, and Yinqiang Zheng. Masactrl: Tuning-free mutual self-attention control for consistent image synthesis and editing. In IEEE International Conference on Computer Vision (ICCV), 2023. 2, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 376, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 376, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 376, + 296, + 421 + ], + "type": "text", + "content": "[5] Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention-based semantic guidance for text-to-image diffusion models. ACM Transactions on Graphics (TOG), 42(4):1-10, 2023. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "text", + "content": "[6] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-sigma: Weak-to-strong training of diffusion transformer for 4k text-to-image generation. arXiv preprint arXiv:2403.04692, 2024. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 295, + 543 + ], + "type": "text", + "content": "[7] Junsong Chen, YU Jincheng, GE Chongjian, Lewei Yao, Enze Xie, Zhongdao Wang, James Kwok, Ping Luo, Huchuan Lu, and Zhenguo Li. PixArt-alpha: Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 544, + 295, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 295, + 589 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 295, + 589 + ], + "type": "text", + "content": "[8] Junsong Chen, Yue Wu, Simian Luo, Enze Xie, Sayak Paul, Ping Luo, Hang Zhao, and Zhenguo Li. Pixart-delta: Fast and controllable image generation with latent consistency models. arXiv preprint arXiv:2401.05252, 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 590, + 295, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 590, + 295, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 590, + 295, + 633 + ], + "type": "text", + "content": "[9] Yusuf Dalva, Kavana Venkatesh, and Pinar Yanardag. FluxSpace: Disentangled Semantic Editing in Rectified Flow Transformers. arXiv preprint arXiv:2412.09611, 2024. 7, 8" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 667 + ], + "type": "text", + "content": "[10] Tri Dao. FlashAttention-2: Faster Attention with Better Parallelism and Work Partitioning. In International Conference on Learning Representations (ICLR), 2024. 5" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 295, + 713 + ], + "type": "text", + "content": "[11] Tri Dao, Daniel Y. Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 5" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 554, + 713 + ], + "type": "list", + "angle": 0, + "index": 29, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 116 + ], + "type": "text", + "content": "[12] Timothee Darcet, Maxime Aquab, Julien Mairal, and Piotr Bojanowski. Vision Transformers Need Registers. In International Conference on Learning Representations (ICLR), 2024. 2, 4" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "spans": [ + { + "bbox": [ + 317, + 118, + 554, + 160 + ], + "type": "text", + "content": "[13] Yingying Deng, Xiangyu He, Changwang Mei, Peisong Wang, and Fan Tang. FireFlow: Fast Inversion of Rectified Flow for Image Semantic Editing. arXiv preprint arXiv:2412.07517, 2024. 2, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 194 + ], + "type": "text", + "content": "[14] Alexey Dosovitskiy. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 195, + 554, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 195, + 554, + 259 + ], + "spans": [ + { + "bbox": [ + 317, + 195, + 554, + 259 + ], + "type": "text", + "content": "[15] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas Müller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In International Conference on Machine Learning (ICML), 2024. 1, 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 554, + 304 + ], + "type": "text", + "content": "[16] Daniel Garibi, Or Patashnik, Andrey Voynov, Hadar Averbuch-Elor, and Daniel Cohen-Or. ReNoise: Real Image Inversion Through Iterative Noising. arXiv preprint arXiv:2403.14602, 2024. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 305, + 553, + 360 + ], + "type": "text", + "content": "[17] Zigang Geng, Binxin Yang, Tiankai Hang, Chen Li, Shuyang Gu, Ting Zhang, Jianmin Bao, Zheng Zhang, Houqiang Li, Han Hu, et al. Instructdiffusion: A generalist modeling interface for vision tasks. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 316, + 361, + 553, + 404 + ], + "type": "text", + "content": "[18] Qin Guo and Tianwei Lin. Focus on your instruction: Fine-grained and multi-instruction image editing by attention modulation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 404, + 553, + 449 + ], + "type": "text", + "content": "[19] Amir Hertz, Ron Mokady, Jay Tenenbaum, Kfir Aberman, Yael Pritch, and Daniel Cohen-or. Prompt-to-Prompt Image Editing with Cross-Attention Control. In International Conference on Learning Representations (ICLR), 2022. 2, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 503 + ], + "type": "text", + "content": "[20] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 2021. 7" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[21] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Conference on Neural Information Processing Systems (NeurIPS), 2020. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[22] Susung Hong, Gyuseong Lee, Wooseok Jang, and Seunglyong Kim. Improving sample quality of diffusion models using self-attention guidance. In IEEE International Conference on Computer Vision (ICCV), 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 554, + 624 + ], + "type": "text", + "content": "[23] Inbar Huberman-Spiegelglas, Vladimir Kulikov, and Tomer Michaeli. An edit friendly ddpm noise space: Inversion and manipulations. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "spans": [ + { + "bbox": [ + 316, + 625, + 553, + 680 + ], + "type": "text", + "content": "[24] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[25] Jaeseok Jeong, Junho Kim, Yunjey Choi, Gayoung Lee, and Youngjung Uh. Visual Style Prompting with Swapping Self-Attention. arXiv preprint arXiv:2402.12974, 2024. 2" + } + ] + } + ], + "index": 28 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19500" + } + ] + } + ], + "index": 30 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 115 + ], + "type": "text", + "content": "[26] Guanlong Jiao, Biqing Huang, Kuan-Chieh Wang, and Renjie Liao. Uniedit-flow: Unleashing inversion and editing in the era of flow models. arXiv preprint arXiv:2504.13109, 2025. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 117, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 117, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 117, + 294, + 161 + ], + "type": "text", + "content": "[27] Xuan Ju, Ailing Zeng, Yuxuan Bian, Shaoteng Liu, and Qiang Xu. PnP Inversion: Boosting Diffusion-based Editing with 3 Lines of Code. In International Conference on Learning Representations (ICLR), 2024. 8, 2, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 216 + ], + "type": "text", + "content": "[28] Minguk Kang, Richard Zhang, Connelly Barnes, Sylvain Paris, Suha Kwak, Jaesik Park, Eli Shechtman, Jun-Yan Zhu, and Taesung Park. Distilling Diffusion Models into Conditional GANs. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 294, + 259 + ], + "type": "text", + "content": "[29] Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 293 + ], + "type": "text", + "content": "[30] Black Forest Labs. Flux.1. https://blackforestlabs.ai/announcing-black-forest-labs/,2024.Accessed:November 2024.1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 370 + ], + "type": "text", + "content": "[31] Benjamin Lefaudeaux, Francisco Massa, Diana Liskovich, Wenhan Xiong, Vittorio Caggiano, Sean Naren, Min Xu, Jieru Hu, Marta Tintore, Susan Zhang, Patrick Labatut, Daniel Haziza, Luca Wehrstedt, Jeremy Reizenstein, and Grigory Sizov. xFormers: A modular and hackable Transformer modelling library. https://github.com/facebookresearch/xformers, 2022.5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 415 + ], + "type": "text", + "content": "[32] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matthew Le. Flow Matching for Generative Modeling. In International Conference on Learning Representations (ICLR), 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 294, + 470 + ], + "type": "text", + "content": "[33] Bingyan Liu, Chengyu Wang, Tingfeng Cao, Kui Jia, and Jun Huang. Towards Understanding Cross and Self-Attention in Stable Diffusion for Text-Guided Image Editing. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 294, + 525 + ], + "type": "text", + "content": "[34] Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Chunyuan Li, Jianwei Yang, Hang Su, Jun Zhu, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. arXiv preprint arXiv:2303.05499, 2023. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 525, + 294, + 569 + ], + "type": "text", + "content": "[35] Xingchao Liu, Chengyue Gong, et al. Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow. In International Conference on Learning Representations (ICLR), 2023. 2, 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[36] Xingchao Liu, Xiwen Zhang, Jianzhu Ma, Jian Peng, et al. Instaflow: One step is enough for high-quality diffusion-based text-to-image generation. In International Conference on Learning Representations (ICLR), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[37] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver: A fast ode solver for diffusion probabilistic model sampling in around 10 steps. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[38] Cheng Lu, Yuhao Zhou, Fan Bao, Jianfei Chen, Chongxuan Li, and Jun Zhu. Dpm-solver++: Fast solver for guided sampling of diffusion probabilistic models. arXiv preprint arXiv:2211.01095, 2022. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[39] Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[40] Simian Luo, Yiqin Tan, Suraj Patil, Daniel Gu, Patrick von Platen, Apolinário Passos, Longbo Huang, Jian Li, and Hang Zhao. Lcm-lora: A universal stable-diffusion acceleration module. arXiv preprint arXiv:2311.05556, 2023. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[41] Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided Image Synthesis and Editing with Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 7, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 241 + ], + "type": "text", + "content": "[42] Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 243, + 553, + 297 + ], + "type": "text", + "content": "[43] Jisu Nam, Heesu Kim, DongJae Lee, Siyoon Jin, Seungryong Kim, and Seunggyu Chang. Dreamcatcher: Appearance matching self-attention for semantically-consistent text-to-image personalization. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 298, + 553, + 331 + ], + "type": "text", + "content": "[44] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In International Conference on Machine Learning (ICML), 2021. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 332, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 553, + 354 + ], + "type": "text", + "content": "[45] OpenAI. GPT-4o, 2024. Large language model developed by OpenAI. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 388 + ], + "type": "text", + "content": "[46] Gaurav Parmar, Krishna Kumar Singh, Richard Zhang, Yijun Li, Jingwan Lu, and Jun-Yan Zhu. Zero-shot image-to-image translation. In ACM SIGGRAPH, 2023. 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 389, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 389, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 389, + 553, + 455 + ], + "type": "text", + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In Conference on Neural Information Processing Systems (NeurIPS), 2019. 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 456, + 553, + 489 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 489 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 489 + ], + "type": "text", + "content": "[48] William Peebles and Saining Xie. Scalable diffusion models with transformers. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 1, 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 491, + 553, + 545 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 491, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 316, + 491, + 553, + 545 + ], + "type": "text", + "content": "[49] Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas Müller, Joe Penna, and Robin Rombach. SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. In International Conference on Learning Representations (ICLR), 2024. 1, 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 546, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 612 + ], + "type": "text", + "content": "[50] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning (ICML), 2021. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 613, + 553, + 667 + ], + "type": "text", + "content": "[51] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. Journal of machine learning research, 2020. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[52] Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. Hierarchical text-conditional image generation with clip latents. arXiv preprint arXiv:2204.06125, 1 (2):3, 2022. 1" + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "19501" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 149 + ], + "type": "text", + "content": "[53] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman Rädle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junting Pan, Kalyan Vasudev Alwala, Nicolas Carion, ChaoYuan Wu, Ross Girshick, Piotr Dólár, and Christoph Feichtenhofer. SAM 2: Segment Anything in Images and Videos. arXiv preprint arXiv:2408.00714, 2024. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "spans": [ + { + "bbox": [ + 56, + 150, + 294, + 204 + ], + "type": "text", + "content": "[54] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 1, 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 270 + ], + "type": "text", + "content": "[55] Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted intervention-MICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part III 18, pages 234-241. Springer, 2015. 1, 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 294, + 325 + ], + "type": "text", + "content": "[56] Litu Rout, Yujia Chen, Nataniel Ruiz, Constantine Caramanis, Sanjay Shakkottai, and Wen-Sheng Chu. Semantic Image Inversion and Editing using Rectified Stochastic Differential Equations. arXiv preprint arXiv:2410.10792, 2024. 2, 7, 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 392 + ], + "type": "text", + "content": "[57] Chitwan Sahara, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al. Photorealistic text-to-image diffusion models with deep language understanding. In Conference on Neural Information Processing Systems (NeurIPS), 2022. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 393, + 294, + 425 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 425 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 425 + ], + "type": "text", + "content": "[58] Tim Salimans and Jonathan Ho. Progressive Distillation for Fast Sampling of Diffusion Models. In International Conference on Learning Representations (ICLR), 2022. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 426, + 294, + 459 + ], + "type": "text", + "content": "[59] Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. Adversarial Diffusion Distillation. arXiv e-prints, pages arXiv-2311, 2023. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 514 + ], + "type": "text", + "content": "[60] Yang Song, Jascha Sohl-Dickstein, Diederik P Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-Based Generative Modeling through Stochastic Differential Equations. In International Conference on Learning Representations (ICLR), 2021. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 514, + 294, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 547 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 547 + ], + "type": "text", + "content": "[61] Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency Models. In International Conference on Machine Learning (ICML), 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 590 + ], + "type": "text", + "content": "[62] Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 635 + ], + "type": "text", + "content": "[63] Jiangshan Wang, Junfu Pu, Zhongang Qi, Jiayi Guo, Yue Ma, Nisha Huang, Yuxin Chen, Xiu Li, and Ying Shan. Taming rectified flow for inversion and editing. arXiv preprint arXiv:2411.04746, 2024. 2, 7, 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 294, + 689 + ], + "type": "text", + "content": "[64] Kai Wang, Fei Yang, Shiqi Yang, Muhammad Atif Butt, and Joost van de Weijer. Dynamic prompt learning: Addressing cross-attention leakage for text-based image editing. In Conference on Neural Information Processing Systems (NeurIPS), 2023. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 713 + ], + "type": "text", + "content": "[65] Zongze Wu, Nicholas Kolkin, Jonathan Brandt, Richard Zhang, and Eli Shechtman. TurboEdit: Instant text-based" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 341 + ], + "type": "list", + "angle": 0, + "index": 20, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "image editing. In European Conference on Computer Vision (ECCV), 2024. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 140 + ], + "type": "text", + "content": "[66] Tianwei Yin, Michael Gharbi, Taesung Park, Richard Zhang, Eli Shechtman, Fredo Durand, and William T Freeman. Improved Distribution Matching Distillation for Fast Image Synthesis. arXiv preprint arXiv:2405.14867, 2024. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 141, + 553, + 195 + ], + "type": "text", + "content": "[67] Tianwei Yin, Michael Gharbi, Richard Zhang, Eli Shechtman, Fredo Durand, William T Freeman, and Taesung Park. One-step diffusion with distribution matching distillation. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[68] Jiahui Yu, Yuanzhong Xu, Jing Yu Koh, Thang Luong, Gunjan Baid, Zirui Wang, Vijay Vasudevan, Alexander Ku, Yinfei Yang, Burcu Karagol Ayan, et al. Scaling Autoregressive Models for Content-Rich Text-to-Image Generation. Transactions on Machine Learning Research, 2022. 4" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 297 + ], + "type": "text", + "content": "[69] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "spans": [ + { + "bbox": [ + 316, + 297, + 553, + 341 + ], + "type": "text", + "content": "[70] Yupeng Zhou, Daquan Zhou, Ming-Ming Cheng, Jiashi Feng, and Qibin Hou. StoryDiffusion: Consistent Self-Attention for Long-Range Image and Video Generation. arXiv preprint arXiv:2405.01434, 2024. 2" + } + ] + } + ], + "index": 19 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "19502" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_content_list.json b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..193e4d801bbbc73c085b05817f17d3e45525b9b4 --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_content_list.json @@ -0,0 +1,2259 @@ +[ + { + "type": "text", + "text": "Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation", + "text_level": 1, + "bbox": [ + 101, + 128, + 897, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "I-Hsiang Chen $^{1,2*,\\dagger}$ Hua-En Chang $^{1*}$ Wei-Ting Chen $^{3}$ Jenq-Neng Hwang $^{2}$ Sy-Yen Kuo $^{1,4}$", + "bbox": [ + 109, + 200, + 885, + 223 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ National Taiwan University $^{2}$ University of Washington $^{3}$ Microsoft $^{4}$ Chang Gung University", + "bbox": [ + 109, + 237, + 883, + 258 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/09a34eda5f6c4fe065d992174737b340116711ec76ef6d676ea86718baf40cf9.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 96, + 306, + 225, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/9fc9d664f8f71143c10bc4a5a310ce3883de73fb8b322b157ec49ada266d79d7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 96, + 366, + 223, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fa5ed57ff3e71523a9139e5e6a776588803ef8a89286046b9bfc81e9a3357704.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 96, + 439, + 223, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c552f6f053db62a4452791fa0779bfda6ee78d20ec52c81e9b3da0b00c110554.jpg", + "image_caption": [ + "(a) Qualitative comparison between PDAF and its backbones", + "Figure 1. The proposed PDAF enhances existing segmentation backbones (DeepLabV3Plus [7] and Mask2Former [10]) by leveraging latent domain prior modeling as a compensation mechanism, yielding consistent improvements on unseen target domains such as BDD100K [67]." + ], + "image_footnote": [], + "bbox": [ + 96, + 500, + 223, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/1ccf65a3e979c1efa3f87f4a80c775ccf6abaffe7cdd4205df920a765d7c43a7.jpg", + "image_caption": [ + "DeepLabV3Plus" + ], + "image_footnote": [], + "bbox": [ + 227, + 306, + 356, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/fd2b5b3668205a89ecab4283c519de64ca492e9b4754935639eea3d3fc944012.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 366, + 356, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d5a937519db47f9b13822931eb3beb981eb3162b56b91af2175d6acd9b0b8818.jpg", + "image_caption": [ + "Mask2Former" + ], + "image_footnote": [], + "bbox": [ + 227, + 439, + 356, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/81a198799bc201b612de3fcf01effc14e864f5427cc2cb7c0676be2d9268eff3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 227, + 500, + 356, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/164e3eb26b432fe92f43a126670cb49175598f495392ce375bf3ae46bd4ef25e.jpg", + "image_caption": [ + "PDAF" + ], + "image_footnote": [], + "bbox": [ + 359, + 306, + 488, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5fd74c7ea111b540e9617995901819f4984a0d784310a7f8420e67043579e499.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 366, + 488, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/92841ec651ba45a3669e930584f5f9733e806c1048485f2e3c729cf29945d525.jpg", + "image_caption": [ + "PDAF" + ], + "image_footnote": [], + "bbox": [ + 359, + 439, + 490, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/673bee209971741d0e944c3dce10903d6b2f930684fef2d00e51fb6f44e57ef4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 359, + 500, + 490, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a55e8a41870fb6d49f656d7aa262bfb4693910d392cce831e4342d56a0e207fc.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 493, + 305, + 622, + 366 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/bb4bffc791e9ab59c5451e3d92bb594f0cae9926cfb49009948c0e05e2208695.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 366, + 620, + 426 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/cf3791240a2ed224c10be1b25c91c245291c9efc75d381ef4cf1e9c9c84f50d4.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 493, + 439, + 620, + 500 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/36c8a7eb6ae3a6ddd6562d7ba8e0928b0e47c3b0b3af2225746ea0374390dfb4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 493, + 500, + 620, + 560 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/7b917f6661e98c0f6acf641fe67c6d21206000df8e99ff2d6e2adad19cba6c8f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 295, + 833, + 364 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/04b5fe741d99e0b12de6641b8a9984d634e738379b0d0920e9d5d237e432fbb7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 656, + 367, + 897, + 489 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/d812e8e0d1942657943ec70329e413a0e8483150ccf91f0f92bf72594643b33b.jpg", + "image_caption": [ + "(b) Overview of PDAF framework." + ], + "image_footnote": [], + "bbox": [ + 656, + 489, + 831, + 554 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 648, + 326, + 664 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS) is a critical yet challenging task, as domain shifts in unseen environments can severely compromise model performance. While recent studies enhance feature alignment by projecting features into the source domain, they often neglect intrinsic latent domain priors, leading to suboptimal results. In this paper, we introduce PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. PDAF introduces a Latent Domain Prior", + "bbox": [ + 89, + 684, + 485, + 835 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "(LDP) to capture domain shifts and uses this prior as a conditioning factor to align both source and unseen target domains. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired source and pseudotarget images to simulate latent domain shifts, enabling LDP modeling. The framework comprises three modules: the Latent Prior Extractor (LPE) predicts the LDP by supervising domain shifts; the Domain Compensation Module (DCM) adjusts feature representations to mitigate domain shifts; and the Diffusion Prior Estimator (DPE) leverages a diffusion process to estimate the LDP without requiring paired samples. This design enables PDAF to iteratively model domain shifts, progressively refining feature representations to enhance generalization under complex target conditions. Extensive experiments validate the effectiveness of PDAF across diverse and challenging urban scenes.", + "bbox": [ + 511, + 636, + 908, + 878 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "Project page: https://pdfa-f.iccv.github.io", + "bbox": [ + 112, + 851, + 403, + 864 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "* Indicates equal contribution.", + "bbox": [ + 116, + 864, + 276, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† This work was conducted while I-Hsiang Chen was an intern at University of Washington.", + "bbox": [ + 91, + 875, + 482, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "21755", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 89, + 222, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation (DGSS) remains a fundamental challenge in computer vision, primarily due to inevitable domain shifts encountered in unseen environments [51, 69]. Segmentation models trained on a source domain frequently degrade on the unseen target domains [2, 8, 37]. This challenge is particularly critical in real-world applications such as autonomous driving [3, 23] and robotic systems [43, 44], where accurate semantic segmentation is essential for ensuring operational safety. Factors such as lighting variations, weather changes, and other underrepresented conditions in training data exacerbate the generalization problem. Addressing the domain shifts thus remains a key obstacle in developing robust and reliable semantic segmentation models.", + "bbox": [ + 89, + 114, + 480, + 325 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Within the field of DGSS, two primary methodologies have emerged: data augmentation-based approaches and domain-invariant representation learning-based approaches. Data augmentation-based methods enhance the diversity of training data by introducing synthetic variations, thereby exposing models to a broader spectrum of potential domain shifts and improving their generalization ability [24, 75, 76]. Nevertheless, a crucial limitation of data augmentation-based methods is their significant dependence on auxiliary domains or generation models [18, 29, 39]. Conversely, domain-invariant representation learning methods aim to train models that extract consistent features across different domains, effectively mitigating the impact of domain shift [2, 9, 16, 19, 35, 36, 41, 56]. However, the inherent entanglement between style and content poses a significant challenge, often resulting in the loss of critical semantic information during feature decomposition [1, 26, 66]. Recent works achieve feature alignment by projecting features into a constrained feature space [1, 26, 66], but they often overlook the intrinsic properties of latent domain priors, leading to suboptimal performance [58, 61].", + "bbox": [ + 91, + 327, + 483, + 657 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To overcome this limitation, we propose PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. Specifically, PDAF introduces a Latent Domain Prior (LDP), which models domain-specific variations and guides feature alignment between source and unseen target domains, as illustrated in Figure 1. Specifically, PDAF integrates into a pre-trained segmentation model and leverages paired source and pseudotarget images to simulate domain shifts. This process enables LDP modeling, which serves as a foundation for robust domain alignment. To leverage these priors effectively, PDAF first utilizes the Latent Prior Extractor (LPE) to capture cross-domain relationships between the source and pseudo-target samples, generating an optimal LDP that encodes domain-specific variations. Building upon this, the", + "bbox": [ + 89, + 659, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Compensation Module (DCM) conditions feature representations on the extracted LDP, ensuring that domain priors guide the adaptation process to mitigate domain shifts while preserving task-relevant information. However, the direct estimation of LDP from explicit source-target pairs may be infeasible in real-world scenarios; to overcome this, the Diffusion Prior Estimator (DPE) employs probabilistic diffusion modeling to infer LDP without requiring paired samples, enhancing the model's adaptability to unseen domains. Through this progressive refinement, PDAF not only aligns features across domains but also maintains critical variations essential for generalization, leading to improved segmentation performance in complex environments.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Extensive experiments on benchmark datasets validate the effectiveness of PDAF, demonstrating that probabilistic modeling of LDP successfully enhances the domain generalization of existing segmentation networks.", + "bbox": [ + 511, + 287, + 903, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our main contributions are as follows:", + "bbox": [ + 532, + 348, + 787, + 361 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce PDAF, a novel probabilistic diffusion alignment framework that leverages probabilistic diffusion modeling of latent domain priors (LDP) to enhance the robustness of existing segmentation models against unseen domain shifts.", + "- We propose three components for PDAF: the latent prior extractor, the domain compensation module, and the diffusion prior estimator. These modules collectively estimate and leverage the LDP for robust feature alignment between source and unseen target domains.", + "- Extensive experiments across diverse datasets and degradation scenarios demonstrate that our method effectively improves the domain generalization of two widely adopted segmentation models (e.g., DeepLabV3Plus [7] and Mask2Former [10]), achieving state-of-the-art performance." + ], + "bbox": [ + 532, + 363, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 618, + 653, + 633 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Domain Generalized Semantic Segmentation. Domain Generalized Semantic Segmentation (DGSS) enhances model robustness to unseen domains without requiring target data. Existing approaches fall into data augmentation-based methods and domain-invariant representation learning-based methods.", + "bbox": [ + 511, + 643, + 903, + 733 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Data augmentation-based methods improve generalization by synthesizing or randomizing styles. DRPC [68] enforces pyramid consistency across stylized versions using real images as style references. GTR-LTR [47] replaces source styles with artistic paintings for global and local texture variation, while WildNet [34] enhances feature-level representations with ImageNet [13] as an external data source. Domain-invariant representation learning-based methods remove domain-specific information via normalization and whitening. IBN-Net [45] integrates Instance Normalization (IN) and Batch Normalization (BN) [27],", + "bbox": [ + 511, + 734, + 906, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "21756", + "bbox": [ + 478, + 944, + 519, + 957 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "while Switchable Whitening (SW) [46] adaptively merges whitening and standardization. ISW [11] and DIRL [65] reduce style-sensitive covariance components, whereas SHADE [75] and SAN-SAW [48] introduce style hallucination and semantic-aware alignment. HGFormer [15] and CMFormer [5] leverage self-attention's visual grouping to handle style variations, demonstrating the strengths of transformer-based models.", + "bbox": [ + 89, + 90, + 480, + 210 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Recent studies reveal that excessive style regularization can weaken content discrimination. SPC [26] projects target styles into a source-style space, while DPCL [66] applies multi-level contrastive learning on projected source features. BlindNet [1] integrates covariance alignment and semantic consistency contrastive learning to reduce style sensitivity in the encoder while enhancing decoder robustness. In addition, some methods explore Vision Foundation Models (VFMs) to improve robustness and transferability in semantic segmentation [4, 62, 72, 73]. Despite progress, DGSS methods focus on feature alignment and augmentation, leaving domain prior modeling underexplored. To bridge this gap, we introduce probabilistic diffusion modeling, enabling structured domain prior estimation and improved generalization to unseen domains.", + "bbox": [ + 89, + 215, + 482, + 441 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models. Diffusion models (DMs) [22] have recently gained significant attention for their outstanding performance in generative modeling. By parameterizing a Markov chain and optimizing the variational lower bound on the likelihood function, DMs effectively capture complex data distributions and produce high-quality samples[22, 30, 33, 59, 70, 71, 74], outperforming conventional frameworks like GANs [14]. To reduce computational overhead, LDM [53] and Wüstchen [50] conduct the diffusion process in latent space while retaining high-quality outputs. In addition to this, several studies have leveraged DMs to encode prior knowledge that facilitate downstream tasks. For instance, DiffIR [64] introduces a diffusion process applied to a compact image prior representation, enhancing both efficiency and stability under challenging conditions, while CDFormer [38] jointly reconstructs content and degraded representations to improve textural fidelity, and UniRestore [6] further bridges perceptual and task-oriented image restoration by leveraging diffusion priors. These advancements highlight the role of diffusion-based methods in capturing complex data distribution, enhancing robustness and adaptability in generative modeling", + "bbox": [ + 89, + 444, + 482, + 777 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Diffusion Models for Domain Generalization. Recent works explore diffusion models for domain generalization in semantic segmentation. DatasetDM [63] synthesizes diverse image-annotation pairs, while DGIn-Style [29] employs semantic guidance to produce consistently paired image-label samples. Gong et al. [20] adopt prompt randomization with pretrained diffusion features, and DIFF [28] integrates sampling and fusion from pre", + "bbox": [ + 89, + 780, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "trained diffusion models to learn universal feature representations. These methods demonstrate the effectiveness of external diffusion-based models in enhancing domain generalization, but they typically incur substantial computational overhead. By contrast, our approach adopts an efficient probabilistic diffusion modeling framework to explicitly model latent domain priors, capturing diverse domain variations without imposing excessive resource demands. Moreover, this design seamlessly integrates with existing segmentation network, enabling robust out-of-distribution performance with minimal additional overhead.", + "bbox": [ + 511, + 90, + 903, + 257 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 270, + 604, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We first introduce the Latent Domain Prior (LDP) Modeling (Sec. 3.1), where we formulate domain generalized semantic segmentation (DGSS) as a probabilistic learning problem and derive the variational inference framework for LDP estimation. Based on this, we develop the Probabilistic Diffusion Alignment Framework (Sec. 3.2), which leverages LDPs to model domain shifts and guide the existing segmentation model in adapting feature representations. Finally, we describe Model Optimization (Sec. 3.3), detailing the training objectives and loss functions designed to enhance robust domain generalization.", + "bbox": [ + 511, + 295, + 906, + 460 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Latent Domain Prior Modeling", + "text_level": 1, + "bbox": [ + 511, + 470, + 790, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We propose a probabilistic modeling framework that explicitly integrates LDP to model domain shifts in the DGSS task. By formulating LDP within a variational inference framework, our method enables principled estimation of domain prior that serves as conditioning variables, thereby guiding segmentation models to generalize more effectively on unseen target domains.", + "bbox": [ + 511, + 492, + 905, + 598 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Probabilistic Problem Formulation. To generalize across domain shifts, we cast DGSS as a probabilistic learning problem. We introduce a Latent Domain Prior (LDP) as a latent variable $z$ to capture domain-specific variations that are not directly observable. Given the target domain distribution $p(x_{t},y_{t})$ , where $y_{t}$ denotes the ground truth, we infer $z$ solely from the available target-domain image $x_{t}$ . Consequently, the prediction function of PDAF is formulated as:", + "bbox": [ + 511, + 598, + 905, + 719 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\np _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) = \\int p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) p _ {\\phi} \\left(z \\mid x _ {t}\\right) d z, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 729, + 906, + 760 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\theta$ and $\\phi$ represent the parameters of the segmentation model and the network designed to estimate $z$ (i.e., LDP), respectively. By incorporating LDP, PDAF regularizes feature adaptation, consequently mitigating domain shifts and enhancing model robustness.", + "bbox": [ + 511, + 763, + 905, + 839 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Variational Inference for LDP Estimation. To facilitate tractable inference of LDP, we introduce a variational posterior $q_{\\varphi}(z|x_t,x_s)$ to approximate the true posterior distribution $p(z|x_t,x_s)$ , where $\\varphi$ represents the parameters", + "bbox": [ + 511, + 839, + 905, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "21757", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/a7408cf0f9b1ee702682731fafed0759b7d3fa7eef7fffd1c251e877422905bc.jpg", + "image_caption": [ + "Figure 2. Overall architecture of PDAF. PDAF augments a pre-trained segmentation network by introducing LDP modeling to enhance its domain generalization. The LPE learns the optimal LDP by modeling cross-domain relationships between source and pseudo-target domains, while DPE is employed to estimate LDP using only target inputs. Finally, the DCM enhances segmentation network with LDP guidance, refining feature alignment and improving domain generalization." + ], + "image_footnote": [], + "bbox": [ + 106, + 92, + 877, + 393 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "of the inference network and $x_{s}$ indicates the source domain counterpart of target domain image $x_{t}$ . The pair of $x_{s}$ and $x_{t}$ preserves semantic consistency while exhibiting domain-specific variations. By leveraging variational inference, we construct a principled optimization framework that estimates LDP to enhance domain alignment. The Evidence Lower Bound (ELBO) of the predictive function is formulated as:", + "bbox": [ + 88, + 489, + 483, + 609 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\log p _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) \\geq \\mathbb {E} _ {q _ {\\varphi} \\left(z \\mid x _ {t}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) \\right] \\tag {2} \\\\ - \\mathbb {K L} \\left[ q _ {\\varphi} (z | x _ {t}, x _ {s}) | | p _ {\\phi} (z | x _ {t}) \\right]. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 129, + 619, + 482, + 657 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The variational posterior $q_{\\varphi}(z|x_t,x_s)$ explicitly encodes cross-domain feature correlations from source-target pairs, enabling precise modeling of latent domain variation critical for effective generalization. The KL divergence term encourages the learned posterior to approximate the prior, ensuring structured regularization. A detailed derivation of ELBO is provided in the supplementary material.", + "bbox": [ + 88, + 667, + 483, + 776 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Probabilistic Diffusion Alignment Framework", + "text_level": 1, + "bbox": [ + 89, + 784, + 480, + 800 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Overview of the Framework. We introduce the Probabilistic Diffusion Alignment Framework (PDAF), which instantiates LDP modeling to obtain domain shift, thereby enhancing generalization performance of segmentation models. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired images generated", + "bbox": [ + 88, + 809, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "through data augmentation, effectively simulating domain shifts. This process preserves semantic consistency while injecting domain-specific variations, thus facilitating the principled ELBO approximation. The optimization function of PDAF is described as follows:", + "bbox": [ + 511, + 489, + 906, + 565 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {P D A F}} = - \\mathbb {E} _ {q _ {\\varphi} \\left(z ^ {\\prime} \\mid x _ {t ^ {\\prime}}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t ^ {\\prime}} \\mid x _ {t ^ {\\prime}}, z ^ {\\prime}\\right) \\right] \\\\ + \\mathbb {K L} \\left(q _ {\\varphi} \\left(z ^ {\\prime} | x _ {t ^ {\\prime}}, x _ {s}\\right) | | p _ {\\phi} \\left(z ^ {\\prime} | x _ {t}\\right)\\right), \\tag {3} \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 560, + 570, + 903, + 608 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $x_{t'}$ represents pseudo-target images generated from source images $x_s$ via data augmentation, while $y_{t'}$ is the corresponding ground truth. $z'$ denotes the LDP, inferred through cross-domain relationships between source and pseudo-target domain samples. The prediction model $p_{\\theta}(.)$ , variational posterior $q_{\\varphi}(.)$ and prior $p_{\\phi}(.)$ jointly optimize segmentation accuracy while modeling domain shifts by maximizing likelihood and minimizing KL divergence, respectively.", + "bbox": [ + 511, + 613, + 906, + 750 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Corresponding to the optimization function in Eq. 3, PDAF introduces three components: (i) a Latent Prior Extractor (LPE), which serves as $q_{\\varphi}(.)$ and infers the optimal LDP from source and pseudo-target pairs; (ii) a Domain Compensation Module (DCM), integrated into a pretrained segmentation network to implement the prediction model $p_{\\theta}(.)$ , which leverages the extracted LDP to refine feature representations, thereby mitigating domain shifts and improving segmentation accuracy; and (iii) a Diffusion Prior Estimator (DPE), implementing $p_{\\phi}(.)$ through probabilistic", + "bbox": [ + 511, + 750, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "21758", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/93dc5a1cfb1bea9e7d80a33f89adc6a69bdf0661c46140c3c34e2c0571bdaa34.jpg", + "image_caption": [ + "(a) Details of LPE" + ], + "image_footnote": [], + "bbox": [ + 98, + 97, + 357, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/054ffc7bd43f082a8bf16f571b7a5c7806895e7e37b6ef7a2ef99d30ae01037b.jpg", + "image_caption": [ + "(b) Details of DCM", + "Figure 3. Schematic diagrams of (a) the Latent Prior Extractor, (b) the Domain Compensation Module and (c) the Diffusion Prior Estimator." + ], + "image_footnote": [], + "bbox": [ + 372, + 95, + 630, + 193 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e8a729b231389f50f222bfb3619eda4a414d5500bfd1c94813761632b2e8eb6d.jpg", + "image_caption": [ + "(c) Details of DPE" + ], + "image_footnote": [], + "bbox": [ + 643, + 97, + 895, + 193 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "diffusion modeling to estimate LDP without paired source and target data. Notably, PDAF is compatible with any existing segmentation model, allowing frozen pre-trained encoders $E_{\\vartheta}$ and decoders $D_{\\vartheta}$ to discover LDP, and then finetuning the target network $E_{\\theta}$ and $D_{\\theta}$ for DGSS. The overall architecture of PDAF is illustrated in Figure 2.", + "bbox": [ + 88, + 280, + 480, + 371 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Latent Prior Extractor. The LPE is designed to estimate the optimal LDP by supervising cross-domain feature relationships, providing effective guidance for feature alignment. As illustrated in Figure 3a, the LPE first concatenates source features $h_{\\vartheta,s}$ and pseudo-target features $h_{\\vartheta,t'}$ , then utilizes residual blocks to model cross-domain feature relationships. The extracted features are then processed through two projection layers to derive the probabilistic modeling parameters, namely the mean $(\\mu)$ and variance $(\\sigma)$ . To ensure stable and effective training, we constrain the optimal LDP $\\tilde{z}$ to follow a standard normal distribution [32], enhancing latent space regularization and enabling more robust prior estimation. The procedure can be formulated as:", + "bbox": [ + 89, + 376, + 483, + 571 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} h _ {\\vartheta , s} = E _ {\\vartheta} \\left(x _ {s}\\right), h _ {\\vartheta , t ^ {\\prime}} = E _ {\\vartheta} \\left(x _ {t ^ {\\prime}}\\right), \\\\ \\mu , \\sigma = F _ {\\varphi} \\left(h _ {\\vartheta , s}, h _ {\\vartheta , t ^ {\\prime}}\\right), \\tag {4} \\\\ \\tilde {z} = \\mu + \\epsilon \\sigma , \\epsilon \\sim \\mathcal {N} (0, I), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 173, + 588, + 480, + 641 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $F_{\\varphi}$ indicates LPE parameterized by $\\varphi$ . The dimension of LDP is $\\tilde{z} \\in \\mathbb{R}^{c' \\times h \\times w}$ , where $c'$ denotes a predefined number of the channel of LDP and $h, w$ correspond to the spatial dimensions of the feature maps.", + "bbox": [ + 89, + 652, + 482, + 715 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Domain Compensation Module. The DCM enhances segmentation models for domain generalization. Drawing inspiration from Spatial Feature Transformation (SFT) [60], DCM leverages LDP as a compensation mechanism, providing domain-aware modulation to improve feature alignment. As illustrated in Figure 3b, a straightforward approach involves projecting LDP onto affine transformation parameters using dedicated projection layers $(F_{\\Psi})$ , yielding scale $(\\tilde{\\gamma})$ and shift $(\\tilde{\\beta})$ parameters. These parameters are then used to rescale and shift the feature representations, effectively normalizing domain discrepancies across spatial regions. The refined features $\\tilde{h}_{\\theta ,t^{\\prime}}$ are subsequently pro", + "bbox": [ + 89, + 719, + 483, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "cessed by the segmentation head $D_{\\theta}$ to generate accurate segmentation maps $\\tilde{y}_{t'}$ . The process can be formulated as:", + "bbox": [ + 511, + 281, + 905, + 311 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {\\gamma}, \\tilde {\\beta} = F _ {\\Psi} (\\tilde {z}),\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 325, + 735, + 340 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {h} _ {\\theta , t ^ {\\prime}} = \\tilde {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\tilde {\\beta}, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 345, + 903, + 364 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\tilde {y} _ {t ^ {\\prime}} = D _ {\\theta} \\left(\\tilde {h} _ {\\vartheta , t ^ {\\prime}}\\right),\n$$\n", + "text_format": "latex", + "bbox": [ + 645, + 368, + 759, + 385 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\odot$ and $\\oplus$ denote element-wise multiplication and element-wise addition, respectively. The dimensions of the affine transformation parameters correspond to the feature map, with $\\tilde{\\gamma},\\tilde{\\beta}\\in \\mathbb{R}^{c\\times h\\times w}$ .", + "bbox": [ + 511, + 398, + 905, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Diffusion Prior Estimator. Diffusion Models (DMs) have demonstrated their effectiveness in sampling from complex target domains, significantly improving various computer vision tasks as evidenced in previous studies [38, 50, 53]. In DPE, we leverage probabilistic diffusion modeling to estimate the LDP, enabling arbitrary target domain and accurate prior estimation for domain alignment. As illustrated in Figure 3c, we first apply the forward diffusion process to an optimal LDP $\\tilde{z}$ , iteratively adding Gaussian noise to obtain an intermediate latent state $z_{T}$ :", + "bbox": [ + 511, + 464, + 906, + 614 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nq \\left(z _ {T} | \\tilde {z}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} \\tilde {z}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 580, + 628, + 903, + 646 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\alpha_{T} = 1 - \\beta_{T}$ and $\\bar{\\alpha}_{T} = \\prod_{i=1}^{T} \\alpha_{i}$ . For the reverse diffusion process, the DPE initializes from $\\hat{z}_{T}$ and performs T-step denoising process to estimate the LDP $\\hat{z}_{0}$ , which is formulated as:", + "bbox": [ + 511, + 660, + 905, + 720 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {T - 1} = \\frac {1}{\\sqrt {\\alpha_ {T}}} \\left(\\hat {z} _ {T} - \\epsilon \\frac {1 - \\alpha_ {T}}{\\sqrt {1 - \\bar {\\alpha} _ {T}}}\\right). \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 594, + 732, + 903, + 766 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, to estimate LDP for target images, we condition the DPE on the extracted target features $h_{\\theta ,t'}$ . DPE $(F_{\\phi})$ then performs a denoising operation at timestep $T$ , represented as $F_{\\phi}(\\hat{z}_{T},T,h_{\\theta ,t'})$ . To accommodate the reduced channel dimension of LDP, we employ accelerated diffusion-based optimization [57] for efficient and rapid inference. Moreover, this design allows for the joint training of DPE and the segmentation head $D_{\\theta}$ , thereby optimizing", + "bbox": [ + 511, + 779, + 906, + 901 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "21759", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "LDP directly for the segmentation task. This process is defined as follows:", + "bbox": [ + 89, + 90, + 482, + 119 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {z} _ {0} = F _ {\\phi} (\\hat {z} _ {T}, T, h _ {\\theta , t ^ {\\prime}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 220, + 128, + 369, + 146 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\gamma}, \\hat {\\beta} = F _ {\\Psi} (\\hat {z} _ {0}), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 207, + 148, + 482, + 176 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {h} _ {\\theta , t ^ {\\prime}} = \\hat {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\hat {\\beta},\n$$\n", + "text_format": "latex", + "bbox": [ + 205, + 170, + 354, + 189 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {y} _ {t ^ {\\prime}} = D _ {\\theta} (\\hat {h} _ {\\vartheta , t ^ {\\prime}}),\n$$\n", + "text_format": "latex", + "bbox": [ + 215, + 193, + 330, + 210 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{y}_{t'}$ represents the predicted segmentation maps conditions on the estimated LDP $\\hat{z}_0$ . The dimensions of the affine transformation parameters correspond to the feature map, with $\\hat{\\gamma}, \\hat{\\beta} \\in \\mathbb{R}^{c \\times h \\times w}$ .", + "bbox": [ + 89, + 220, + 483, + 281 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Inference. During inference, only target domain images $x_{t}$ are available, so the variational posterior is omitted. Instead, we extract target features $h_{\\theta ,t}$ using the segmentation backbone $E_{\\theta}(\\cdot)$ as the conditioning input and sample a Gaussian noise vector $\\hat{z}_T$ . The DPE then iteratively refines this noise over $T$ denoising steps to produce the estimated LDP $\\hat{z}_0$ . Finally, the DCM leverages the estimated LDP to refine the feature representations, which are used for robust segmentation prediction.", + "bbox": [ + 89, + 285, + 483, + 421 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.3. Model Optimization", + "text_level": 1, + "bbox": [ + 89, + 430, + 282, + 446 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We train our framework by optimizing the likelihood estimation and enforcing posterior regularization to enhance domain generalization, as defined in Eq. (3). To achieve maximum likelihood estimation, we employ task-specific loss functions $\\mathcal{L}_{\\mathrm{task}}$ tailored to different segmentation models. Specifically, we use weighted cross-entropy loss for DeepLabV3+ [7] and focal loss for Mask2Former [10]. To further enhance feature consistency, we introduce a semantic consistency loss $\\mathcal{L}_{\\mathrm{sc}}$ that measures the discrepancy between the predictions of the source and pseudo-target images. This loss enforces semantic alignment in feature space and accelerates model convergence:", + "bbox": [ + 89, + 452, + 483, + 633 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {s c}} = \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\tilde {y} _ {t ^ {\\prime}} \\right\\| _ {2} + \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\hat {y} _ {t ^ {\\prime}} \\right\\| _ {2}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 142, + 643, + 482, + 662 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $f_{\\vartheta}$ denotes the frozen segmentation network $E_{\\vartheta}(.)$ and $D_{\\vartheta}(.)$ , $\\tilde{y}_{t'}$ and $\\hat{y}_{t'}$ represent the predictions from Eq. (5) and Eq. (8), respectively. To minimize the KL divergence and enforce posterior regularization, we adopt a prior constraint loss $\\mathcal{L}_{\\mathrm{prior}}$ to align the outputs of LPE and DPE, thereby implicitly reducing the distributional discrepancy between the posterior and the prior:", + "bbox": [ + 89, + 671, + 483, + 777 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {p r i o r}} = \\| \\hat {z} _ {0} - \\tilde {z} \\| _ {2}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 218, + 789, + 482, + 806 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\hat{z}_0$ and $\\tilde{z}$ indicate the optimal LDP and estimated LDP from Eq. (4) and Eq. (8), respectively. The overall loss is:", + "bbox": [ + 89, + 816, + 483, + 847 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {\\mathrm {s c}} \\mathcal {L} _ {\\mathrm {s c}} + \\lambda_ {\\text {p r i o r}} \\mathcal {L} _ {\\text {p r i o r}}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 858, + 482, + 875 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "where $\\lambda_{\\mathrm{task}}$ , $\\lambda_{\\mathrm{sc}}$ , and $\\lambda_{\\mathrm{prior}}$ represents the scaling factors.", + "bbox": [ + 89, + 885, + 460, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Implementation Details", + "text_level": 1, + "bbox": [ + 513, + 89, + 735, + 106 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Datasets", + "text_level": 1, + "bbox": [ + 513, + 114, + 616, + 128 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We conduct experiments on five urban scene semantic segmentation datasets that share 19 common scene categories.", + "bbox": [ + 511, + 136, + 903, + 166 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Cityscapes (C) [12] consists of high-resolution images captured from 50 distinct cities, where we exclusively use the finely annotated subset containing 2,975 training and 500 validation images.", + "bbox": [ + 511, + 167, + 903, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "BDD-100K (B) [67] provides 7,000 training images and 1,000 validation images collected from various locations conditions.", + "bbox": [ + 511, + 227, + 903, + 271 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Mapillary (M) [42] is a large-scale dataset containing 25,000 high-resolution street-view images sourced globally. GTA5 (G) [52] includes 24,966 images simulated using the Grand Theft Auto V game engine, divided into 12,403 training, 6,382 validation, and 6,181 testing samples.", + "bbox": [ + 511, + 272, + 903, + 348 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "SYNTHIA (S) [54] consists of 9,400 photorealistic synthetic images, including 2,830 for testing.", + "bbox": [ + 511, + 348, + 903, + 378 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Evaluation Protocols", + "text_level": 1, + "bbox": [ + 513, + 388, + 709, + 402 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Following existing DGSS benchmarks [1, 11, 15, 45, 46, 49], we adopt a single-source domain generalization setting, where one dataset is used for training (source domain), and the remaining four datasets are treated as unseen target domains for evaluation. For a fair comparison, all experiments are built on DeepLabV3Plus [7] (with ResNet-50 [21]) and Mask2Former [10] (with Swin-T and Swin-L [40]). We use the mean intersection over the union (mIoU) [17] to measure the performance of segmentation.", + "bbox": [ + 511, + 410, + 905, + 546 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Training Details", + "text_level": 1, + "bbox": [ + 513, + 560, + 673, + 575 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To facilitate a fair evaluation, we exclusively employ Photometric Augmentation as the data augmentation strategy to generate pseudo-target domain samples. The model is trained with Adam Optimizer [31] with an initial learning rate of $1 \\times 10^{-5}$ , momentum parameters $(\\beta_{1},\\beta_{2}) = (0.9,0.999)$ , and a batch size of 4 for 100 epochs. The channel of LDP is set to $c' = 4$ . In DPE, we adopt a total diffusion timestep of $T = 4$ , with the noise variance $\\beta_{t}$ linearly increasing from $\\beta_{1} = 0.1$ to $\\beta_{T} = 0.99$ . The loss coefficients are adjusted to $(\\lambda_{\\mathrm{task}},\\lambda_{\\mathrm{sc}},\\lambda_{\\mathrm{prior}}) = (0.5,0.5,1.0)$ , balancing the contributions of different loss terms. All experiments are implemented in PyTorch and conducted on a single NVIDIA RTX 4090 GPU. More details of the implementation and hyperparameters analysis are provided in the Supplementary Material.", + "bbox": [ + 511, + 582, + 906, + 809 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 513, + 823, + 645, + 839 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Comparison with State-of-the-arts", + "text_level": 1, + "bbox": [ + 511, + 848, + 813, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "To access the effectiveness of PDAF, we compare our approach against a comprehensive set of state-of-the-art meth", + "bbox": [ + 511, + 869, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "21760", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/f3fc0319fb6e225f02f4cb33b4c699e2ca28d1e079e01165c94f1395e9b208c3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 87, + 194, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7d163fe516cfbe098b12324d14ec783f727dee55adcd7326740f1c62e0d8ae9d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 94, + 156, + 194, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/165c7e4eef114c0eb381cd64c25aa2a144536e4dfd5382ee83e7486be19aff84.jpg", + "image_caption": [ + "Input" + ], + "image_footnote": [], + "bbox": [ + 94, + 224, + 194, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b2660d3a756f451534a98804057300856c424c0c2667a911b881219344688ba3.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 88, + 299, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a1e5e81f42ac5523e1a2b48932c8ced41f546414eb152a0553d8607dad422c22.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 197, + 156, + 297, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0a3dfe1443b9652ca045eb7d75b02eef45d29391636c82cdffee6291507cf607.jpg", + "image_caption": [ + "DeepLabV3Plus [7]" + ], + "image_footnote": [], + "bbox": [ + 197, + 224, + 299, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f8a960ec99f3b14a4f911c4e6d25b9d0645bb81677247ca7faeb42e1969e0ba.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 88, + 401, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/19e597b728a49a813bfd76748a2ff58996ff97b047c5593fa85cdf0440c88908.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 302, + 156, + 401, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2f53b884f22fba0c6a87b5c3b9fe8c566f050df8e709256db53c93e49b4a85a7.jpg", + "image_caption": [ + "ISW [11]" + ], + "image_footnote": [], + "bbox": [ + 302, + 224, + 401, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/562139c18de1ab8f1bed582fad83f1f88eb0faaff0659106a763b66f242c95d5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 88, + 501, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ed2d78f2b6b74e645b4b3a688bd37f10354064915aab39660226ce0405c50b6d.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 401, + 156, + 501, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7106d2d10e74c9140b485a7d656e585a5dea84283c55394f8acc70d648ceaa62.jpg", + "image_caption": [ + "SHADE [75]", + "Figure 4. Qualitative comparison between DGSS methods trained on GTAV (G) [52] using DeepLabV3Plus [7] with ResNet-50 [21]. The first, second, and third rows show the predictions for Cityscapes (C) [12], BDD-100K (B) [67], and Mapillary (M) [42], respectively." + ], + "image_footnote": [], + "bbox": [ + 401, + 224, + 501, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a1088d99bfe9667cec19e2e700240293e0ed35a079ad70362ecf43c14d0534c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 88, + 602, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/327e9b3ee22ccb2177af0b192b53920e204b035251f40fcd4ec46adfdccbf248.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 503, + 156, + 602, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/4d7be68d2f9913c97b84a58107b44f1147e586a00bd76dd439c42318fcd4c0ad.jpg", + "image_caption": [ + "WildNet [34]" + ], + "image_footnote": [], + "bbox": [ + 503, + 224, + 602, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/299c53624a2e292d5470ebadf3995512803b23b9e28abe14677d7f1f6a8c9d8a.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 88, + 702, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f68322b4936bbe68b29d9fc4a6b4e5d6d3d0ee7daa6612046dc2a9d042f61cf7.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 602, + 156, + 702, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/e4dd0f0daeb0c5e63f477d5c82b73e1d9c2b5319d122d86d60a4e038888249bf.jpg", + "image_caption": [ + "BlindNet [1]" + ], + "image_footnote": [], + "bbox": [ + 602, + 224, + 702, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/03ed4bbfe0227a35fb998b549d91ae0457fba1f65823d7f6bd9efa1360371e32.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 88, + 803, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/9a372b15018a9575633c2f9d738dda67deb7f5bf490d02fe9953887193a22eb5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 704, + 156, + 803, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/01bb4ee1a8fa13ba1cfbda9a377008786cb85755bc24e5949030fec56a42f499.jpg", + "image_caption": [ + "PDAF" + ], + "image_footnote": [], + "bbox": [ + 704, + 224, + 803, + 290 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/438216f6227b4b0cf09247f11276d9d72b594effa57324bc9c41f387057cd51c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 88, + 903, + 155 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7bf49b21301531a8bdf53d7671b5637f06f4d2005f9f35a92dba010e1f0e1973.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 805, + 156, + 903, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/2810a97e5384679a90d4e0594ffefec88c46f4d64cf16a503cc88a5716f8fba6.jpg", + "image_caption": [ + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 805, + 224, + 903, + 290 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/56f45b6371d1e7f25f35323918153a50ccb862ad30e3255afac7831b667815f3.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneTrained on Cityscapes (C)
BMGSAvg.
DeepLabV3Plus [7]Res5044.9651.6842.5523.2940.62
IBN [45]48.5657.0445.0626.1444.20
IW [46]48.4955.8244.8726.1043.82
Itemorm [25]49.2356.2645.7325.9844.30
DRPC [68]49.8656.3445.6226.5844.60
ISW [11]50.7358.6445.0026.2045.14
GTR [47]50.7557.1645.7926.4745.04
DIRL [65]51.80-46.5226.50-
SHADE [75]50.9560.6748.6127.6246.96
SAW [48]52.9559.8147.2828.3247.09
WildNet [34]50.9458.7947.0127.9546.17
DPCL [66]50.9758.5946.0025.8545.34
BlindNet [1]51.8460.1847.9728.5147.13
PDAF53.5062.9350.5430.6849.41
Mask2Former [10]51.3065.3050.6034.0050.30
HGFformer [15]Swin-T53.4066.9051.3033.6051.30
PDAF56.4066.4055.2034.8053.20
Mask2Former [10]60.1072.2057.8042.4058.13
HGFformer [15]Swin-L61.5072.1059.4041.3058.58
CMFormer [5]62.6073.6060.7043.0059.98
PDAF63.0074.1063.2044.0061.08
", + "bbox": [ + 109, + 369, + 467, + 618 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "ods including both CNN-based techniques [1, 7, 11, 25, 26, 34, 45-48, 65, 66, 68, 75, 76] and Mask2Former-based approaches [5, 10, 15]. We adhere to standard DGSS evaluation protocols and report results directly cited from prior works [1, 5, 11, 46, 48].", + "bbox": [ + 89, + 699, + 482, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Cityscapes Source Domain. Table 1 shows that PDAF achieves the best overall performance across four unseen datasets, demonstrating strong generalization to diverse domains. PDAF effectively transfers from real-world datasets (C) [12] to target domains with distinct visual distributions, such as game-rendered synthetic scenes (G) [52] and photo-realistic simulations (S) [54]. Notably, PDAF consistently improves performance across CNN-based and", + "bbox": [ + 89, + 779, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/508dd63bfe76a9f96acf8ba00d2328afef5acb26975ad2e906aaa0557c24e4ba.jpg", + "table_caption": [ + "Table 1. Comparison with existing methods trained on Cityscapes (C) [12]. The best and second-best mIoU scores are blue and magenta." + ], + "table_footnote": [], + "table_body": "
MethodTrained on GTAV (G)
CBMSAvg.
DeepLabV3Plus [7]28.9525.1428.1826.2327.13
IBN [45]33.8532.3037.7527.9032.95
IW [46]29.9127.4829.7127.6128.68
Iternorm [25]31.8132.7033.8827.0731.37
DRPC [68]37.4232.1434.1228.0632.94
ISW [11]36.5835.2040.3328.3035.10
GTR [47]37.5333.7534.5228.1733.49
DIRL [65]41.0439.1541.60--
SHADE [75]44.6539.2843.34--
SAW [48]39.7537.3441.8630.7937.44
WildNet [34]44.6238.4246.0931.3440.12
AdvStyle [76]39.6235.5437.00--
SPC [26]44.1040.4645.51--
DPCL [66]44.7440.5946.3330.8140.62
BlindNet [1]45.7241.3247.0831.3941.38
PDAF45.4043.5649.4532.9842.85
", + "bbox": [ + 535, + 369, + 885, + 599 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Comparison with existing DGSS methods using a CNN-based network trained on GTAV (G) [52].", + "bbox": [ + 511, + 609, + 906, + 638 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Mask2Former-based methods, highlighting its flexibility in enhancing existing segmentation networks.", + "bbox": [ + 511, + 666, + 903, + 696 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "GTAV Source Domain. Table 2 demonstrates the effectiveness of PDAF in bridging the domain gap between synthetic and real-world datasets. Unlike existing DGSS methods that rely primarily on feature normalization, PDAF explicitly models latent domain priors, providing structured guidance for feature alignment and enhancing domain generalization. As shown in Figure 4, qualitative comparisons indicate that PDAF effectively preserves detailed structural information—such as road boundaries and object contours—resulting in more precise and consistent segmentation outcomes.", + "bbox": [ + 511, + 699, + 906, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Clear to Corruptions. We evaluate PDAF on ACDC [55], which introduces four real-world adverse conditions (rain,", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "21761", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/4e8b5728f79b9cde1cc64dd29b8e56124570eee358a353a3d89498f8f55e9829.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MethodBackboneTrained on Cityscape (C)
FoggyNightRainSnowAvg.
IBN [45]63.8021.2050.4049.6043.70
Itemorm [25]63.3023.8050.1049.9045.30
IW [46]Res5062.4021.8052.4047.6046.60
ISW [11]64.3024.3056.0049.8048.10
PDAF68.2830.1459.2353.5852.81
Mask2Former [10]56.4039.1058.9058.2053.15
HGFormer [15]Swin-T58.5043.3062.0058.3055.53
PDAF73.7547.7568.1262.2562.97
Mask2Former [10]69.1053.1068.3065.2063.93
HGFormer [15]Swin-L69.9052.7072.0068.6065.80
PDAF80.7255.1273.1371.4370.10
", + "bbox": [ + 106, + 88, + 470, + 238 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/aa80e0164b820852f7369f3711a772f2cf38e3c1267e808954b59b3ff47d006b.jpg", + "table_caption": [ + "Table 3. Comparison with existing methods evaluated on ACDC [55]." + ], + "table_footnote": [], + "table_body": "
MethodsBMGSAvg.
Baseline44.9651.6842.5523.2940.62
PDAF w/o LPE49.5957.6346.3427.1945.19
PDAF w/o DCM51.5160.7949.4229.5547.82
PDAF w/o DPE51.9860.3349.3828.8647.64
PDAF53.5062.9350.5430.6849.41
", + "bbox": [ + 140, + 290, + 434, + 363 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/1111b4f5692d14c77f47be2cac94e84076e8713e44461386868d09b63b03dbde.jpg", + "table_caption": [ + "Table 4. Effectiveness of proposed modules in PDAF." + ], + "table_footnote": [], + "table_body": "
\\( \\mathcal{L}_{\\text{task}} \\)\\( \\mathcal{L}_{\\text{sc}} \\)\\( \\mathcal{L}_{\\text{prior}} \\)BMGSAvg.
✓--49.7658.7347.3227.5145.83
✓✓-51.2859.5848.3827.9446.80
✓-✓53.3662.1149.6930.1048.82
✓✓✓53.5062.9350.5430.6849.41
", + "bbox": [ + 135, + 402, + 439, + 465 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "fog, night and snow). We employ DeepLabV3Plus [7] and Mask2Former [10] as the segmentation network and train our method on Cityscapes [12]. As shown in Table 3, PDAF achieves consistent improvement under all adversarial conditions, demonstrating that LDP effectively compensates for degraded features caused by image corruptions and enhances robustness in varied environments.", + "bbox": [ + 89, + 517, + 483, + 621 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.2. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 632, + 254, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "To validate the effectiveness of the proposed modules, we perform an ablation study. All experiments are conducted using DeepLabV3Plus [7] with ResNet-50 [21] as the segmentation network. The model is trained on Cityscapes (C) [12] and evaluated on BDD-100K (B) [67], Mapillary (M) [42], GTAV (G) [52], and SYNTHIA (S) [54] to assess generalization across diverse domain shifts.", + "bbox": [ + 89, + 655, + 482, + 760 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Effectiveness of Proposed Modules. We consider several experimental configurations: (i) Baseline: vanilla DeepLabV3Plus [7] trained without PDAF; (ii) PDAF w/o LPE: where variational posterior is removed, resulting in LDP estimation without cross-domain guidance; (iii) PDAF w/o DCM: where LDP is directly added to the original features without employing DCM; (iv) PDAF w/o DPE: where DPE is replaced by a single-source LPE, omitting the diffusion process; and (v) PDAF: which integrates all proposed", + "bbox": [ + 89, + 763, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/729addfbddbdff75fa2083a9320fad2169336eed4c5336d9e34d12d4092fa4d5.jpg", + "table_caption": [ + "Table 5. Analysis on the proposed loss functions." + ], + "table_footnote": [], + "table_body": "
Number of c'1416642561024
Avg. of mIoU48.7449.4148.6648.3546.1445.69
", + "bbox": [ + 539, + 89, + 880, + 121 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/61fd6feec8e15941998b6bcc414e90cbf160fa22435a2ef597b9edfd3c6af50e.jpg", + "table_caption": [ + "Table 6. Ablation study on the number of channels for LDP." + ], + "table_footnote": [], + "table_body": "
Number of Steps12481632
Avg. of mIoU48.8849.1249.4149.4349.4649.48
", + "bbox": [ + 531, + 160, + 887, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Table 7. Ablation study on the number of timestep in DPE.", + "bbox": [ + 524, + 203, + 893, + 217 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "modules. As shown in Table 4, each module enhances segmentation performance, with LPE playing a pivotal role in posterior regularization for structured LDP modeling.", + "bbox": [ + 511, + 244, + 903, + 290 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Loss Functions. Table 5 demonstrates that each loss term contributes to improved performance. The prior constraint loss, $\\mathcal{L}_{\\mathrm{prior}}$ , enforces consistency between the LDPs estimated by LPE and DPE, enabling DPE to accurately infer LDP solely from target samples. Additionally, the semantic consistency loss, $\\mathcal{L}_{\\mathrm{sc}}$ , enhances feature alignment by constraining predictions from both source and pseudo-target samples, further bolstering domain generalization.", + "bbox": [ + 511, + 292, + 905, + 429 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of LDP Channel Dimensions. LDP serve as a variational prior for modeling domain shifts while guiding feature alignment in DGSS. The channel dimension of LDP is crucial for capturing domain-specific variations. As shown in Table 6, a channel dimension of 4 yields the best overall performance. In contrast, higher channel dimensions introduce feature redundancy and increase the risk of overfitting, which limits further performance improvements.", + "bbox": [ + 511, + 433, + 905, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of steps for DPE. DPE iteratively refines LDP through the denoising process, facilitating precise alignment across diverse target domains. As shown in Table 7, performance gradually improves with an increasing number of diffusion steps. To achieve an optimal trade-off between efficiency and accuracy, we set the diffusion timestep to 4.", + "bbox": [ + 511, + 558, + 905, + 648 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 662, + 633, + 679 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This paper introduces PDAF, a probabilistic diffusion alignment framework designed for DGSS. By explicitly modeling latent domain prior (LDP), PDAF captures domain-specific variations to facilitate robust feature alignment across unseen target domains. To achieve this, PDAF involves a latent prior extractor to infer cross-domain relationships, a domain compensation module to condition segmentation features on LDP and a diffusion prior estimator to estimate LDP through a denoising process, allowing robust generalization to unseen domains. Extensive experiments demonstrate the effectiveness of PDAF, improving the generalization of existing segmentation models across diverse environments, highlighting the robustness of LDP modeling in domain generalized semantic segmentation.", + "bbox": [ + 511, + 688, + 905, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "21762", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "We thank to National Center for High-performance Computing (NCHC) for providing computational and storage resources. This research was supported by Taiwan's National Science and Technology Council under Grant NSTC 111-2221-E-002-136-MY3 and NSTC 114-2221-E-002-067-MY3.", + "bbox": [ + 89, + 114, + 483, + 205 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 220, + 187, + 236 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Woojin Ahn, Geun Yeong Yang, Hyun Duck Choi, and Myo Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3, 6, 7", + "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. In NeurIPS, 2018. 2", + "[3] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on robot learning, 2023. 2", + "[4] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3", + "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 3, 7", + "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 3", + "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2, 6, 7, 8", + "[8] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Hua-En Chang, Jian-Jiun Ding, and Sy-Yen Kuo. Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning. In ECCV, 2022. 2", + "[9] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In AAAI, 2022. 2", + "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2021. 1, 2, 6, 7, 8", + "[11] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 3, 6, 7, 8", + "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + ], + "bbox": [ + 93, + 244, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "dataset for semantic urban scene understanding. In CVPR, 2016. 6, 7, 8", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2", + "[14] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3", + "[15] Jian Ding, Nan Xue, Guisong Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 6, 7, 8", + "[16] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 2", + "[17] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2014. 6", + "[18] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2", + "[19] Muhammad Ghifary, W. Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In ICCV, 2015. 2", + "[20] Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Mangas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3", + "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2015. 6, 7, 8", + "[22] Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3", + "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2", + "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2", + "[25] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, 2019. 7, 8", + "[26] Wei Huang, Chang Wen Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 2, 3, 7", + "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2", + "[28] Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. arXiv preprint arXiv:2406.00777, 2024.3", + "[29] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "21763", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR, 2024. 3", + "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6", + "[32] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2013. 5", + "[33] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 3", + "[34] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 2, 7", + "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2017. 2", + "[36] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex Chichung Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 2", + "[37] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2", + "[38] Qingguo Liu, Chenyi Zhuang, Pan Gao, and Jie Qin. Cdformer: when degradation prediction embraces diffusion model for blind image super-resolution. In CVPR, 2024. 3, 5", + "[39] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 2", + "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6", + "[41] Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, 2017. 2", + "[42] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6, 7, 8", + "[43] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In AAAI, 2021. 2", + "[44] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In Int. Conf. on Intel. Robots and Systems, 2021. 2", + "[45] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 2, 6, 7, 8", + "[46] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 3, 6, 7, 8", + "[47] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liua. Global and local texture randomization for" + ], + "bbox": [ + 91, + 90, + 485, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "synthetic-to-real semantic segmentation. IEEE TIP, 2021. 2,7", + "[48] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 3, 7", + "[49] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 6", + "[50] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637v2, 2023. 3, 5", + "[51] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In CVPR, 2024. 2", + "[52] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 6, 7, 8", + "[53] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 3, 5", + "[54] Germán Ros, Laura Sellart, Joanna Materzynska, David Vázquez, and Antonio M. López. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 6, 7, 8", + "[55] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 7, 8", + "[56] Seonguk Seo, Yumin Suh, Dongwan Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, 2019. 2", + "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5", + "[58] Naveen Venkat, Jogendra Nath Kundu, Durgesh Singh, Ambareesh Revanur, et al. Your classifier can secretly suffice multi-source domain adaptation. NIPS, 2020. 2", + "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3", + "[60] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5", + "[61] Zongbin Wang, Bin Pan, Shiyu Shen, Tianyang Shi, and Zhenwei Shi. Domain generalization guided by large-scale pre-trained priors. arXiv preprint arXiv:2406.05628, 2024. 2", + "[62] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3", + "[63] Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. NeurIPS, 2023. 3" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "21764", + "bbox": [ + 478, + 944, + 519, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[64] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 3", + "[65] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3, 7", + "[66] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 2, 3, 7", + "[67] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2018. 1, 6, 7, 8", + "[68] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto L. Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 2, 7", + "[69] Haojie Zhang, Yongyi Su, Xun Xu, and Kui Jia. Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation. In CVPR, 2024. 2", + "[70] Juntao Zhang, Yuehuai Liu, Yu-Wing Tai, and Chi-Keung Tang. C3net: Compound conditioned control for multimodal content generation. In CVPR, 2024. 3", + "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3", + "[72] Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In CVPR, 2025. 3", + "[73] Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher-guided robust tuning of vision foundation models for domain generalized segmentation. In CVPR, 2025. 3", + "[74] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3", + "[75] Yuyang Zhao, Zhun Zhong, Na Zhao, N. Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 2, 3, 7", + "[76] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and N. Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 2, 7" + ], + "bbox": [ + 91, + 90, + 480, + 785 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "21765", + "bbox": [ + 478, + 944, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_model.json b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_model.json new file mode 100644 index 0000000000000000000000000000000000000000..0d41adb2e0bb9db4faeb9768440c85b6a0ec6ee7 --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_model.json @@ -0,0 +1,3214 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.102, + 0.13, + 0.898, + 0.177 + ], + "angle": 0, + "content": "Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.202, + 0.886, + 0.224 + ], + "angle": 0, + "content": "I-Hsiang Chen\\(^{1,2*,\\dagger}\\) Hua-En Chang\\(^{1*}\\) Wei-Ting Chen\\(^{3}\\) Jenq-Neng Hwang\\(^{2}\\) Sy-Yen Kuo\\(^{1,4}\\)" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.238, + 0.884, + 0.259 + ], + "angle": 0, + "content": "\\(^{1}\\)National Taiwan University \\(^{2}\\)University of Washington \\(^{3}\\)Microsoft \\(^{4}\\)Chang Gung University" + }, + { + "type": "image_caption", + "bbox": [ + 0.145, + 0.296, + 0.176, + 0.307 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.307, + 0.226, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.367, + 0.225, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.146, + 0.429, + 0.176, + 0.44 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.44, + 0.225, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.501, + 0.225, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.249, + 0.296, + 0.338, + 0.307 + ], + "angle": 0, + "content": "DeepLabV3Plus" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.307, + 0.357, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.367, + 0.357, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.255, + 0.429, + 0.332, + 0.44 + ], + "angle": 0, + "content": "Mask2Former" + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.44, + 0.357, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.228, + 0.501, + 0.357, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.296, + 0.444, + 0.306 + ], + "angle": 0, + "content": "PDAF" + }, + { + "type": "image", + "bbox": [ + 0.36, + 0.307, + 0.49, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.367, + 0.49, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.408, + 0.429, + 0.444, + 0.44 + ], + "angle": 0, + "content": "PDAF" + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.44, + 0.491, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.361, + 0.501, + 0.491, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.296, + 0.594, + 0.306 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.306, + 0.623, + 0.367 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.367, + 0.622, + 0.427 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.521, + 0.429, + 0.594, + 0.44 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.44, + 0.622, + 0.5 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.494, + 0.501, + 0.622, + 0.561 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.191, + 0.566, + 0.512, + 0.579 + ], + "angle": 0, + "content": "(a) Qualitative comparison between PDAF and its backbones" + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.296, + 0.834, + 0.366 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.368, + 0.898, + 0.49 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.658, + 0.491, + 0.832, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.685, + 0.566, + 0.871, + 0.578 + ], + "angle": 0, + "content": "(b) Overview of PDAF framework." + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.591, + 0.908, + 0.621 + ], + "angle": 0, + "content": "Figure 1. The proposed PDAF enhances existing segmentation backbones (DeepLabV3Plus [7] and Mask2Former [10]) by leveraging latent domain prior modeling as a compensation mechanism, yielding consistent improvements on unseen target domains such as BDD100K [67]." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.65, + 0.327, + 0.665 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.685, + 0.486, + 0.836 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS) is a critical yet challenging task, as domain shifts in unseen environments can severely compromise model performance. While recent studies enhance feature alignment by projecting features into the source domain, they often neglect intrinsic latent domain priors, leading to suboptimal results. In this paper, we introduce PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. PDAF introduces a Latent Domain Prior" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.909, + 0.879 + ], + "angle": 0, + "content": "(LDP) to capture domain shifts and uses this prior as a conditioning factor to align both source and unseen target domains. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired source and pseudotarget images to simulate latent domain shifts, enabling LDP modeling. The framework comprises three modules: the Latent Prior Extractor (LPE) predicts the LDP by supervising domain shifts; the Domain Compensation Module (DCM) adjusts feature representations to mitigate domain shifts; and the Diffusion Prior Estimator (DPE) leverages a diffusion process to estimate the LDP without requiring paired samples. This design enables PDAF to iteratively model domain shifts, progressively refining feature representations to enhance generalization under complex target conditions. Extensive experiments validate the effectiveness of PDAF across diverse and challenging urban scenes." + }, + { + "type": "page_footnote", + "bbox": [ + 0.114, + 0.852, + 0.404, + 0.865 + ], + "angle": 0, + "content": "Project page: https://pdfa-f.iccv.github.io" + }, + { + "type": "page_footnote", + "bbox": [ + 0.117, + 0.865, + 0.277, + 0.876 + ], + "angle": 0, + "content": "* Indicates equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.093, + 0.876, + 0.483, + 0.9 + ], + "angle": 0, + "content": "† This work was conducted while I-Hsiang Chen was an intern at University of Washington." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.852, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21755" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.223, + 0.106 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.482, + 0.326 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation (DGSS) remains a fundamental challenge in computer vision, primarily due to inevitable domain shifts encountered in unseen environments [51, 69]. Segmentation models trained on a source domain frequently degrade on the unseen target domains [2, 8, 37]. This challenge is particularly critical in real-world applications such as autonomous driving [3, 23] and robotic systems [43, 44], where accurate semantic segmentation is essential for ensuring operational safety. Factors such as lighting variations, weather changes, and other underrepresented conditions in training data exacerbate the generalization problem. Addressing the domain shifts thus remains a key obstacle in developing robust and reliable semantic segmentation models." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.328, + 0.485, + 0.658 + ], + "angle": 0, + "content": "Within the field of DGSS, two primary methodologies have emerged: data augmentation-based approaches and domain-invariant representation learning-based approaches. Data augmentation-based methods enhance the diversity of training data by introducing synthetic variations, thereby exposing models to a broader spectrum of potential domain shifts and improving their generalization ability [24, 75, 76]. Nevertheless, a crucial limitation of data augmentation-based methods is their significant dependence on auxiliary domains or generation models [18, 29, 39]. Conversely, domain-invariant representation learning methods aim to train models that extract consistent features across different domains, effectively mitigating the impact of domain shift [2, 9, 16, 19, 35, 36, 41, 56]. However, the inherent entanglement between style and content poses a significant challenge, often resulting in the loss of critical semantic information during feature decomposition [1, 26, 66]. Recent works achieve feature alignment by projecting features into a constrained feature space [1, 26, 66], but they often overlook the intrinsic properties of latent domain priors, leading to suboptimal performance [58, 61]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To overcome this limitation, we propose PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. Specifically, PDAF introduces a Latent Domain Prior (LDP), which models domain-specific variations and guides feature alignment between source and unseen target domains, as illustrated in Figure 1. Specifically, PDAF integrates into a pre-trained segmentation model and leverages paired source and pseudotarget images to simulate domain shifts. This process enables LDP modeling, which serves as a foundation for robust domain alignment. To leverage these priors effectively, PDAF first utilizes the Latent Prior Extractor (LPE) to capture cross-domain relationships between the source and pseudo-target samples, generating an optimal LDP that encodes domain-specific variations. Building upon this, the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "Domain Compensation Module (DCM) conditions feature representations on the extracted LDP, ensuring that domain priors guide the adaptation process to mitigate domain shifts while preserving task-relevant information. However, the direct estimation of LDP from explicit source-target pairs may be infeasible in real-world scenarios; to overcome this, the Diffusion Prior Estimator (DPE) employs probabilistic diffusion modeling to infer LDP without requiring paired samples, enhancing the model's adaptability to unseen domains. Through this progressive refinement, PDAF not only aligns features across domains but also maintains critical variations essential for generalization, leading to improved segmentation performance in complex environments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.288, + 0.905, + 0.348 + ], + "angle": 0, + "content": "Extensive experiments on benchmark datasets validate the effectiveness of PDAF, demonstrating that probabilistic modeling of LDP successfully enhances the domain generalization of existing segmentation networks." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.349, + 0.788, + 0.362 + ], + "angle": 0, + "content": "Our main contributions are as follows:" + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.364, + 0.905, + 0.438 + ], + "angle": 0, + "content": "- We introduce PDAF, a novel probabilistic diffusion alignment framework that leverages probabilistic diffusion modeling of latent domain priors (LDP) to enhance the robustness of existing segmentation models against unseen domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.44, + 0.905, + 0.514 + ], + "angle": 0, + "content": "- We propose three components for PDAF: the latent prior extractor, the domain compensation module, and the diffusion prior estimator. These modules collectively estimate and leverage the LDP for robust feature alignment between source and unseen target domains." + }, + { + "type": "text", + "bbox": [ + 0.533, + 0.515, + 0.905, + 0.605 + ], + "angle": 0, + "content": "- Extensive experiments across diverse datasets and degradation scenarios demonstrate that our method effectively improves the domain generalization of two widely adopted segmentation models (e.g., DeepLabV3Plus [7] and Mask2Former [10]), achieving state-of-the-art performance." + }, + { + "type": "list", + "bbox": [ + 0.533, + 0.364, + 0.905, + 0.605 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.655, + 0.635 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.645, + 0.905, + 0.734 + ], + "angle": 0, + "content": "Domain Generalized Semantic Segmentation. Domain Generalized Semantic Segmentation (DGSS) enhances model robustness to unseen domains without requiring target data. Existing approaches fall into data augmentation-based methods and domain-invariant representation learning-based methods." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Data augmentation-based methods improve generalization by synthesizing or randomizing styles. DRPC [68] enforces pyramid consistency across stylized versions using real images as style references. GTR-LTR [47] replaces source styles with artistic paintings for global and local texture variation, while WildNet [34] enhances feature-level representations with ImageNet [13] as an external data source. Domain-invariant representation learning-based methods remove domain-specific information via normalization and whitening. IBN-Net [45] integrates Instance Normalization (IN) and Batch Normalization (BN) [27]," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.958 + ], + "angle": 0, + "content": "21756" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.212 + ], + "angle": 0, + "content": "while Switchable Whitening (SW) [46] adaptively merges whitening and standardization. ISW [11] and DIRL [65] reduce style-sensitive covariance components, whereas SHADE [75] and SAN-SAW [48] introduce style hallucination and semantic-aware alignment. HGFormer [15] and CMFormer [5] leverage self-attention's visual grouping to handle style variations, demonstrating the strengths of transformer-based models." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.216, + 0.483, + 0.442 + ], + "angle": 0, + "content": "Recent studies reveal that excessive style regularization can weaken content discrimination. SPC [26] projects target styles into a source-style space, while DPCL [66] applies multi-level contrastive learning on projected source features. BlindNet [1] integrates covariance alignment and semantic consistency contrastive learning to reduce style sensitivity in the encoder while enhancing decoder robustness. In addition, some methods explore Vision Foundation Models (VFMs) to improve robustness and transferability in semantic segmentation [4, 62, 72, 73]. Despite progress, DGSS methods focus on feature alignment and augmentation, leaving domain prior modeling underexplored. To bridge this gap, we introduce probabilistic diffusion modeling, enabling structured domain prior estimation and improved generalization to unseen domains." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.445, + 0.483, + 0.779 + ], + "angle": 0, + "content": "Diffusion Models. Diffusion models (DMs) [22] have recently gained significant attention for their outstanding performance in generative modeling. By parameterizing a Markov chain and optimizing the variational lower bound on the likelihood function, DMs effectively capture complex data distributions and produce high-quality samples[22, 30, 33, 59, 70, 71, 74], outperforming conventional frameworks like GANs [14]. To reduce computational overhead, LDM [53] and Wüstchen [50] conduct the diffusion process in latent space while retaining high-quality outputs. In addition to this, several studies have leveraged DMs to encode prior knowledge that facilitate downstream tasks. For instance, DiffIR [64] introduces a diffusion process applied to a compact image prior representation, enhancing both efficiency and stability under challenging conditions, while CDFormer [38] jointly reconstructs content and degraded representations to improve textural fidelity, and UniRestore [6] further bridges perceptual and task-oriented image restoration by leveraging diffusion priors. These advancements highlight the role of diffusion-based methods in capturing complex data distribution, enhancing robustness and adaptability in generative modeling" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Diffusion Models for Domain Generalization. Recent works explore diffusion models for domain generalization in semantic segmentation. DatasetDM [63] synthesizes diverse image-annotation pairs, while DGIn-Style [29] employs semantic guidance to produce consistently paired image-label samples. Gong et al. [20] adopt prompt randomization with pretrained diffusion features, and DIFF [28] integrates sampling and fusion from pre" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.258 + ], + "angle": 0, + "content": "trained diffusion models to learn universal feature representations. These methods demonstrate the effectiveness of external diffusion-based models in enhancing domain generalization, but they typically incur substantial computational overhead. By contrast, our approach adopts an efficient probabilistic diffusion modeling framework to explicitly model latent domain priors, capturing diverse domain variations without imposing excessive resource demands. Moreover, this design seamlessly integrates with existing segmentation network, enabling robust out-of-distribution performance with minimal additional overhead." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.271, + 0.605, + 0.287 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.296, + 0.907, + 0.462 + ], + "angle": 0, + "content": "We first introduce the Latent Domain Prior (LDP) Modeling (Sec. 3.1), where we formulate domain generalized semantic segmentation (DGSS) as a probabilistic learning problem and derive the variational inference framework for LDP estimation. Based on this, we develop the Probabilistic Diffusion Alignment Framework (Sec. 3.2), which leverages LDPs to model domain shifts and guide the existing segmentation model in adapting feature representations. Finally, we describe Model Optimization (Sec. 3.3), detailing the training objectives and loss functions designed to enhance robust domain generalization." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.471, + 0.792, + 0.487 + ], + "angle": 0, + "content": "3.1. Latent Domain Prior Modeling" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.493, + 0.906, + 0.599 + ], + "angle": 0, + "content": "We propose a probabilistic modeling framework that explicitly integrates LDP to model domain shifts in the DGSS task. By formulating LDP within a variational inference framework, our method enables principled estimation of domain prior that serves as conditioning variables, thereby guiding segmentation models to generalize more effectively on unseen target domains." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.906, + 0.72 + ], + "angle": 0, + "content": "Probabilistic Problem Formulation. To generalize across domain shifts, we cast DGSS as a probabilistic learning problem. We introduce a Latent Domain Prior (LDP) as a latent variable \\( z \\) to capture domain-specific variations that are not directly observable. Given the target domain distribution \\( p(x_{t},y_{t}) \\), where \\( y_{t} \\) denotes the ground truth, we infer \\( z \\) solely from the available target-domain image \\( x_{t} \\). Consequently, the prediction function of PDAF is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.73, + 0.907, + 0.761 + ], + "angle": 0, + "content": "\\[\np _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) = \\int p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) p _ {\\phi} \\left(z \\mid x _ {t}\\right) d z, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.906, + 0.84 + ], + "angle": 0, + "content": "where \\(\\theta\\) and \\(\\phi\\) represent the parameters of the segmentation model and the network designed to estimate \\(z\\) (i.e., LDP), respectively. By incorporating LDP, PDAF regularizes feature adaptation, consequently mitigating domain shifts and enhancing model robustness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.902 + ], + "angle": 0, + "content": "Variational Inference for LDP Estimation. To facilitate tractable inference of LDP, we introduce a variational posterior \\( q_{\\varphi}(z|x_t,x_s) \\) to approximate the true posterior distribution \\( p(z|x_t,x_s) \\), where \\( \\varphi \\) represents the parameters" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21757" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.107, + 0.093, + 0.878, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.409, + 0.908, + 0.466 + ], + "angle": 0, + "content": "Figure 2. Overall architecture of PDAF. PDAF augments a pre-trained segmentation network by introducing LDP modeling to enhance its domain generalization. The LPE learns the optimal LDP by modeling cross-domain relationships between source and pseudo-target domains, while DPE is employed to estimate LDP using only target inputs. Finally, the DCM enhances segmentation network with LDP guidance, refining feature alignment and improving domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.491, + 0.484, + 0.611 + ], + "angle": 0, + "content": "of the inference network and \\( x_{s} \\) indicates the source domain counterpart of target domain image \\( x_{t} \\). The pair of \\( x_{s} \\) and \\( x_{t} \\) preserves semantic consistency while exhibiting domain-specific variations. By leveraging variational inference, we construct a principled optimization framework that estimates LDP to enhance domain alignment. The Evidence Lower Bound (ELBO) of the predictive function is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.13, + 0.621, + 0.483, + 0.658 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\log p _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) \\geq \\mathbb {E} _ {q _ {\\varphi} \\left(z \\mid x _ {t}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) \\right] \\tag {2} \\\\ - \\mathbb {K L} \\left[ q _ {\\varphi} (z | x _ {t}, x _ {s}) | | p _ {\\phi} (z | x _ {t}) \\right]. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.669, + 0.484, + 0.777 + ], + "angle": 0, + "content": "The variational posterior \\( q_{\\varphi}(z|x_t,x_s) \\) explicitly encodes cross-domain feature correlations from source-target pairs, enabling precise modeling of latent domain variation critical for effective generalization. The KL divergence term encourages the learned posterior to approximate the prior, ensuring structured regularization. A detailed derivation of ELBO is provided in the supplementary material." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.785, + 0.481, + 0.801 + ], + "angle": 0, + "content": "3.2. Probabilistic Diffusion Alignment Framework" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.81, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Overview of the Framework. We introduce the Probabilistic Diffusion Alignment Framework (PDAF), which instantiates LDP modeling to obtain domain shift, thereby enhancing generalization performance of segmentation models. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired images generated" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.491, + 0.907, + 0.566 + ], + "angle": 0, + "content": "through data augmentation, effectively simulating domain shifts. This process preserves semantic consistency while injecting domain-specific variations, thus facilitating the principled ELBO approximation. The optimization function of PDAF is described as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.561, + 0.571, + 0.905, + 0.609 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {\\mathrm {P D A F}} = - \\mathbb {E} _ {q _ {\\varphi} \\left(z ^ {\\prime} \\mid x _ {t ^ {\\prime}}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t ^ {\\prime}} \\mid x _ {t ^ {\\prime}}, z ^ {\\prime}\\right) \\right] \\\\ + \\mathbb {K L} \\left(q _ {\\varphi} \\left(z ^ {\\prime} | x _ {t ^ {\\prime}}, x _ {s}\\right) | | p _ {\\phi} \\left(z ^ {\\prime} | x _ {t}\\right)\\right), \\tag {3} \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.907, + 0.75 + ], + "angle": 0, + "content": "where \\( x_{t'} \\) represents pseudo-target images generated from source images \\( x_s \\) via data augmentation, while \\( y_{t'} \\) is the corresponding ground truth. \\( z' \\) denotes the LDP, inferred through cross-domain relationships between source and pseudo-target domain samples. The prediction model \\( p_{\\theta}(.) \\), variational posterior \\( q_{\\varphi}(.) \\) and prior \\( p_{\\phi}(.) \\) jointly optimize segmentation accuracy while modeling domain shifts by maximizing likelihood and minimizing KL divergence, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Corresponding to the optimization function in Eq. 3, PDAF introduces three components: (i) a Latent Prior Extractor (LPE), which serves as \\( q_{\\varphi}(.) \\) and infers the optimal LDP from source and pseudo-target pairs; (ii) a Domain Compensation Module (DCM), integrated into a pretrained segmentation network to implement the prediction model \\( p_{\\theta}(.) \\), which leverages the extracted LDP to refine feature representations, thereby mitigating domain shifts and improving segmentation accuracy; and (iii) a Diffusion Prior Estimator (DPE), implementing \\( p_{\\phi}(.) \\) through probabilistic" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21758" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.098, + 0.359, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.202, + 0.278, + 0.214 + ], + "angle": 0, + "content": "(a) Details of LPE" + }, + { + "type": "image", + "bbox": [ + 0.373, + 0.097, + 0.631, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.446, + 0.202, + 0.551, + 0.214 + ], + "angle": 0, + "content": "(b) Details of DCM" + }, + { + "type": "image", + "bbox": [ + 0.645, + 0.098, + 0.896, + 0.194 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.721, + 0.202, + 0.82, + 0.214 + ], + "angle": 0, + "content": "(c) Details of DPE" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.226, + 0.907, + 0.253 + ], + "angle": 0, + "content": "Figure 3. Schematic diagrams of (a) the Latent Prior Extractor, (b) the Domain Compensation Module and (c) the Diffusion Prior Estimator." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.281, + 0.482, + 0.372 + ], + "angle": 0, + "content": "diffusion modeling to estimate LDP without paired source and target data. Notably, PDAF is compatible with any existing segmentation model, allowing frozen pre-trained encoders \\( E_{\\vartheta} \\) and decoders \\( D_{\\vartheta} \\) to discover LDP, and then finetuning the target network \\( E_{\\theta} \\) and \\( D_{\\theta} \\) for DGSS. The overall architecture of PDAF is illustrated in Figure 2." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.377, + 0.484, + 0.573 + ], + "angle": 0, + "content": "Latent Prior Extractor. The LPE is designed to estimate the optimal LDP by supervising cross-domain feature relationships, providing effective guidance for feature alignment. As illustrated in Figure 3a, the LPE first concatenates source features \\( h_{\\vartheta,s} \\) and pseudo-target features \\( h_{\\vartheta,t'} \\), then utilizes residual blocks to model cross-domain feature relationships. The extracted features are then processed through two projection layers to derive the probabilistic modeling parameters, namely the mean \\( (\\mu) \\) and variance \\( (\\sigma) \\). To ensure stable and effective training, we constrain the optimal LDP \\( \\tilde{z} \\) to follow a standard normal distribution [32], enhancing latent space regularization and enabling more robust prior estimation. The procedure can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.174, + 0.589, + 0.482, + 0.642 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} h _ {\\vartheta , s} = E _ {\\vartheta} \\left(x _ {s}\\right), h _ {\\vartheta , t ^ {\\prime}} = E _ {\\vartheta} \\left(x _ {t ^ {\\prime}}\\right), \\\\ \\mu , \\sigma = F _ {\\varphi} \\left(h _ {\\vartheta , s}, h _ {\\vartheta , t ^ {\\prime}}\\right), \\tag {4} \\\\ \\tilde {z} = \\mu + \\epsilon \\sigma , \\epsilon \\sim \\mathcal {N} (0, I), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.716 + ], + "angle": 0, + "content": "where \\(F_{\\varphi}\\) indicates LPE parameterized by \\(\\varphi\\). The dimension of LDP is \\(\\tilde{z} \\in \\mathbb{R}^{c' \\times h \\times w}\\), where \\(c'\\) denotes a predefined number of the channel of LDP and \\(h, w\\) correspond to the spatial dimensions of the feature maps." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Domain Compensation Module. The DCM enhances segmentation models for domain generalization. Drawing inspiration from Spatial Feature Transformation (SFT) [60], DCM leverages LDP as a compensation mechanism, providing domain-aware modulation to improve feature alignment. As illustrated in Figure 3b, a straightforward approach involves projecting LDP onto affine transformation parameters using dedicated projection layers \\((F_{\\Psi})\\), yielding scale \\((\\tilde{\\gamma})\\) and shift \\((\\tilde{\\beta})\\) parameters. These parameters are then used to rescale and shift the feature representations, effectively normalizing domain discrepancies across spatial regions. The refined features \\(\\tilde{h}_{\\theta ,t^{\\prime}}\\) are subsequently pro" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.282, + 0.906, + 0.312 + ], + "angle": 0, + "content": "cessed by the segmentation head \\(D_{\\theta}\\) to generate accurate segmentation maps \\(\\tilde{y}_{t'}\\). The process can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.636, + 0.326, + 0.736, + 0.342 + ], + "angle": 0, + "content": "\\[\n\\tilde {\\gamma}, \\tilde {\\beta} = F _ {\\Psi} (\\tilde {z}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.635, + 0.346, + 0.905, + 0.365 + ], + "angle": 0, + "content": "\\[\n\\tilde {h} _ {\\theta , t ^ {\\prime}} = \\tilde {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\tilde {\\beta}, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.647, + 0.369, + 0.76, + 0.386 + ], + "angle": 0, + "content": "\\[\n\\tilde {y} _ {t ^ {\\prime}} = D _ {\\theta} \\left(\\tilde {h} _ {\\vartheta , t ^ {\\prime}}\\right),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.461 + ], + "angle": 0, + "content": "where \\(\\odot\\) and \\(\\oplus\\) denote element-wise multiplication and element-wise addition, respectively. The dimensions of the affine transformation parameters correspond to the feature map, with \\(\\tilde{\\gamma},\\tilde{\\beta}\\in \\mathbb{R}^{c\\times h\\times w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.907, + 0.616 + ], + "angle": 0, + "content": "Diffusion Prior Estimator. Diffusion Models (DMs) have demonstrated their effectiveness in sampling from complex target domains, significantly improving various computer vision tasks as evidenced in previous studies [38, 50, 53]. In DPE, we leverage probabilistic diffusion modeling to estimate the LDP, enabling arbitrary target domain and accurate prior estimation for domain alignment. As illustrated in Figure 3c, we first apply the forward diffusion process to an optimal LDP \\(\\tilde{z}\\), iteratively adding Gaussian noise to obtain an intermediate latent state \\(z_{T}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.581, + 0.63, + 0.905, + 0.647 + ], + "angle": 0, + "content": "\\[\nq \\left(z _ {T} | \\tilde {z}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} \\tilde {z}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.661, + 0.906, + 0.722 + ], + "angle": 0, + "content": "where \\(\\alpha_{T} = 1 - \\beta_{T}\\) and \\(\\bar{\\alpha}_{T} = \\prod_{i=1}^{T} \\alpha_{i}\\). For the reverse diffusion process, the DPE initializes from \\(\\hat{z}_{T}\\) and performs T-step denoising process to estimate the LDP \\(\\hat{z}_{0}\\), which is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.595, + 0.733, + 0.905, + 0.767 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {T - 1} = \\frac {1}{\\sqrt {\\alpha_ {T}}} \\left(\\hat {z} _ {T} - \\epsilon \\frac {1 - \\alpha_ {T}}{\\sqrt {1 - \\bar {\\alpha} _ {T}}}\\right). \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.907, + 0.902 + ], + "angle": 0, + "content": "Specifically, to estimate LDP for target images, we condition the DPE on the extracted target features \\( h_{\\theta ,t'} \\). DPE \\( (F_{\\phi}) \\) then performs a denoising operation at timestep \\( T \\), represented as \\( F_{\\phi}(\\hat{z}_{T},T,h_{\\theta ,t'}) \\). To accommodate the reduced channel dimension of LDP, we employ accelerated diffusion-based optimization [57] for efficient and rapid inference. Moreover, this design allows for the joint training of DPE and the segmentation head \\( D_{\\theta} \\), thereby optimizing" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21759" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.121 + ], + "angle": 0, + "content": "LDP directly for the segmentation task. This process is defined as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.221, + 0.13, + 0.37, + 0.147 + ], + "angle": 0, + "content": "\\[\n\\hat {z} _ {0} = F _ {\\phi} (\\hat {z} _ {T}, T, h _ {\\theta , t ^ {\\prime}}),\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.208, + 0.15, + 0.483, + 0.177 + ], + "angle": 0, + "content": "\\[\n\\hat {\\gamma}, \\hat {\\beta} = F _ {\\Psi} (\\hat {z} _ {0}), \\tag {8}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.206, + 0.171, + 0.355, + 0.19 + ], + "angle": 0, + "content": "\\[\n\\hat {h} _ {\\theta , t ^ {\\prime}} = \\hat {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\hat {\\beta},\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.217, + 0.194, + 0.331, + 0.212 + ], + "angle": 0, + "content": "\\[\n\\hat {y} _ {t ^ {\\prime}} = D _ {\\theta} (\\hat {h} _ {\\vartheta , t ^ {\\prime}}),\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.222, + 0.484, + 0.282 + ], + "angle": 0, + "content": "where \\(\\hat{y}_{t'}\\) represents the predicted segmentation maps conditions on the estimated LDP \\(\\hat{z}_0\\). The dimensions of the affine transformation parameters correspond to the feature map, with \\(\\hat{\\gamma}, \\hat{\\beta} \\in \\mathbb{R}^{c \\times h \\times w}\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.286, + 0.484, + 0.422 + ], + "angle": 0, + "content": "Inference. During inference, only target domain images \\( x_{t} \\) are available, so the variational posterior is omitted. Instead, we extract target features \\( h_{\\theta ,t} \\) using the segmentation backbone \\( E_{\\theta}(\\cdot) \\) as the conditioning input and sample a Gaussian noise vector \\( \\hat{z}_T \\). The DPE then iteratively refines this noise over \\( T \\) denoising steps to produce the estimated LDP \\( \\hat{z}_0 \\). Finally, the DCM leverages the estimated LDP to refine the feature representations, which are used for robust segmentation prediction." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.431, + 0.284, + 0.447 + ], + "angle": 0, + "content": "3.3. Model Optimization" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.453, + 0.484, + 0.635 + ], + "angle": 0, + "content": "We train our framework by optimizing the likelihood estimation and enforcing posterior regularization to enhance domain generalization, as defined in Eq. (3). To achieve maximum likelihood estimation, we employ task-specific loss functions \\(\\mathcal{L}_{\\mathrm{task}}\\) tailored to different segmentation models. Specifically, we use weighted cross-entropy loss for DeepLabV3+ [7] and focal loss for Mask2Former [10]. To further enhance feature consistency, we introduce a semantic consistency loss \\(\\mathcal{L}_{\\mathrm{sc}}\\) that measures the discrepancy between the predictions of the source and pseudo-target images. This loss enforces semantic alignment in feature space and accelerates model convergence:" + }, + { + "type": "equation", + "bbox": [ + 0.143, + 0.645, + 0.483, + 0.663 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {s c}} = \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\tilde {y} _ {t ^ {\\prime}} \\right\\| _ {2} + \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\hat {y} _ {t ^ {\\prime}} \\right\\| _ {2}, \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.779 + ], + "angle": 0, + "content": "where \\( f_{\\vartheta} \\) denotes the frozen segmentation network \\( E_{\\vartheta}(.) \\) and \\( D_{\\vartheta}(.) \\), \\( \\tilde{y}_{t'} \\) and \\( \\hat{y}_{t'} \\) represent the predictions from Eq. (5) and Eq. (8), respectively. To minimize the KL divergence and enforce posterior regularization, we adopt a prior constraint loss \\( \\mathcal{L}_{\\mathrm{prior}} \\) to align the outputs of LPE and DPE, thereby implicitly reducing the distributional discrepancy between the posterior and the prior:" + }, + { + "type": "equation", + "bbox": [ + 0.22, + 0.79, + 0.483, + 0.807 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {p r i o r}} = \\| \\hat {z} _ {0} - \\tilde {z} \\| _ {2}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.817, + 0.484, + 0.848 + ], + "angle": 0, + "content": "where \\(\\hat{z}_0\\) and \\(\\tilde{z}\\) indicate the optimal LDP and estimated LDP from Eq. (4) and Eq. (8), respectively. The overall loss is:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.859, + 0.483, + 0.876 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {\\mathrm {s c}} \\mathcal {L} _ {\\mathrm {s c}} + \\lambda_ {\\text {p r i o r}} \\mathcal {L} _ {\\text {p r i o r}}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.462, + 0.902 + ], + "angle": 0, + "content": "where \\(\\lambda_{\\mathrm{task}}\\), \\(\\lambda_{\\mathrm{sc}}\\), and \\(\\lambda_{\\mathrm{prior}}\\) represents the scaling factors." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.09, + 0.736, + 0.107 + ], + "angle": 0, + "content": "4. Implementation Details" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.115, + 0.617, + 0.13 + ], + "angle": 0, + "content": "4.1. Datasets" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.137, + 0.905, + 0.167 + ], + "angle": 0, + "content": "We conduct experiments on five urban scene semantic segmentation datasets that share 19 common scene categories." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.168, + 0.905, + 0.228 + ], + "angle": 0, + "content": "Cityscapes (C) [12] consists of high-resolution images captured from 50 distinct cities, where we exclusively use the finely annotated subset containing 2,975 training and 500 validation images." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.905, + 0.272 + ], + "angle": 0, + "content": "BDD-100K (B) [67] provides 7,000 training images and 1,000 validation images collected from various locations conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.273, + 0.905, + 0.349 + ], + "angle": 0, + "content": "Mapillary (M) [42] is a large-scale dataset containing 25,000 high-resolution street-view images sourced globally. GTA5 (G) [52] includes 24,966 images simulated using the Grand Theft Auto V game engine, divided into 12,403 training, 6,382 validation, and 6,181 testing samples." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.349, + 0.905, + 0.38 + ], + "angle": 0, + "content": "SYNTHIA (S) [54] consists of 9,400 photorealistic synthetic images, including 2,830 for testing." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.389, + 0.71, + 0.403 + ], + "angle": 0, + "content": "4.2. Evaluation Protocols" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.411, + 0.906, + 0.547 + ], + "angle": 0, + "content": "Following existing DGSS benchmarks [1, 11, 15, 45, 46, 49], we adopt a single-source domain generalization setting, where one dataset is used for training (source domain), and the remaining four datasets are treated as unseen target domains for evaluation. For a fair comparison, all experiments are built on DeepLabV3Plus [7] (with ResNet-50 [21]) and Mask2Former [10] (with Swin-T and Swin-L [40]). We use the mean intersection over the union (mIoU) [17] to measure the performance of segmentation." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.561, + 0.674, + 0.577 + ], + "angle": 0, + "content": "4.3. Training Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.81 + ], + "angle": 0, + "content": "To facilitate a fair evaluation, we exclusively employ Photometric Augmentation as the data augmentation strategy to generate pseudo-target domain samples. The model is trained with Adam Optimizer [31] with an initial learning rate of \\( 1 \\times 10^{-5} \\), momentum parameters \\( (\\beta_{1},\\beta_{2}) = (0.9,0.999) \\), and a batch size of 4 for 100 epochs. The channel of LDP is set to \\( c' = 4 \\). In DPE, we adopt a total diffusion timestep of \\( T = 4 \\), with the noise variance \\( \\beta_{t} \\) linearly increasing from \\( \\beta_{1} = 0.1 \\) to \\( \\beta_{T} = 0.99 \\). The loss coefficients are adjusted to \\( (\\lambda_{\\mathrm{task}},\\lambda_{\\mathrm{sc}},\\lambda_{\\mathrm{prior}}) = (0.5,0.5,1.0) \\), balancing the contributions of different loss terms. All experiments are implemented in PyTorch and conducted on a single NVIDIA RTX 4090 GPU. More details of the implementation and hyperparameters analysis are provided in the Supplementary Material." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.824, + 0.646, + 0.84 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.849, + 0.815, + 0.864 + ], + "angle": 0, + "content": "5.1. Comparison with State-of-the-arts" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.902 + ], + "angle": 0, + "content": "To access the effectiveness of PDAF, we compare our approach against a comprehensive set of state-of-the-art meth" + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21760" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.088, + 0.196, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.157, + 0.195, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.226, + 0.195, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.13, + 0.294, + 0.16, + 0.306 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.089, + 0.3, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.199, + 0.157, + 0.299, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.198, + 0.225, + 0.3, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.196, + 0.293, + 0.302, + 0.306 + ], + "angle": 0, + "content": "DeepLabV3Plus [7]" + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.089, + 0.403, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.157, + 0.402, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.303, + 0.225, + 0.402, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.328, + 0.294, + 0.379, + 0.306 + ], + "angle": 0, + "content": "ISW [11]" + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.089, + 0.503, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.157, + 0.503, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.403, + 0.225, + 0.503, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.418, + 0.294, + 0.489, + 0.306 + ], + "angle": 0, + "content": "SHADE [75]" + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.089, + 0.603, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.157, + 0.603, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.504, + 0.225, + 0.603, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.518, + 0.294, + 0.59, + 0.306 + ], + "angle": 0, + "content": "WildNet [34]" + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.089, + 0.704, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.157, + 0.703, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.604, + 0.225, + 0.704, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.62, + 0.294, + 0.69, + 0.306 + ], + "angle": 0, + "content": "BlindNet [1]" + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.089, + 0.805, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.157, + 0.805, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.705, + 0.225, + 0.805, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.737, + 0.294, + 0.773, + 0.305 + ], + "angle": 0, + "content": "PDAF" + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.089, + 0.905, + 0.156 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.157, + 0.905, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.806, + 0.225, + 0.905, + 0.291 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.818, + 0.294, + 0.892, + 0.305 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.318, + 0.904, + 0.347 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparison between DGSS methods trained on GTAV (G) [52] using DeepLabV3Plus [7] with ResNet-50 [21]. The first, second, and third rows show the predictions for Cityscapes (C) [12], BDD-100K (B) [67], and Mapillary (M) [42], respectively." + }, + { + "type": "table", + "bbox": [ + 0.111, + 0.37, + 0.468, + 0.619 + ], + "angle": 0, + "content": "
MethodBackboneTrained on Cityscapes (C)
BMGSAvg.
DeepLabV3Plus [7]Res5044.9651.6842.5523.2940.62
IBN [45]48.5657.0445.0626.1444.20
IW [46]48.4955.8244.8726.1043.82
Itemorm [25]49.2356.2645.7325.9844.30
DRPC [68]49.8656.3445.6226.5844.60
ISW [11]50.7358.6445.0026.2045.14
GTR [47]50.7557.1645.7926.4745.04
DIRL [65]51.80-46.5226.50-
SHADE [75]50.9560.6748.6127.6246.96
SAW [48]52.9559.8147.2828.3247.09
WildNet [34]50.9458.7947.0127.9546.17
DPCL [66]50.9758.5946.0025.8545.34
BlindNet [1]51.8460.1847.9728.5147.13
PDAF53.5062.9350.5430.6849.41
Mask2Former [10]51.3065.3050.6034.0050.30
HGFformer [15]Swin-T53.4066.9051.3033.6051.30
PDAF56.4066.4055.2034.8053.20
Mask2Former [10]60.1072.2057.8042.4058.13
HGFformer [15]Swin-L61.5072.1059.4041.3058.58
CMFormer [5]62.6073.6060.7043.0059.98
PDAF63.0074.1063.2044.0061.08
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.671 + ], + "angle": 0, + "content": "Table 1. Comparison with existing methods trained on Cityscapes (C) [12]. The best and second-best mIoU scores are blue and magenta." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.7, + 0.483, + 0.776 + ], + "angle": 0, + "content": "ods including both CNN-based techniques [1, 7, 11, 25, 26, 34, 45-48, 65, 66, 68, 75, 76] and Mask2Former-based approaches [5, 10, 15]. We adhere to standard DGSS evaluation protocols and report results directly cited from prior works [1, 5, 11, 46, 48]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.78, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Cityscapes Source Domain. Table 1 shows that PDAF achieves the best overall performance across four unseen datasets, demonstrating strong generalization to diverse domains. PDAF effectively transfers from real-world datasets (C) [12] to target domains with distinct visual distributions, such as game-rendered synthetic scenes (G) [52] and photo-realistic simulations (S) [54]. Notably, PDAF consistently improves performance across CNN-based and" + }, + { + "type": "table", + "bbox": [ + 0.536, + 0.37, + 0.887, + 0.6 + ], + "angle": 0, + "content": "
MethodTrained on GTAV (G)
CBMSAvg.
DeepLabV3Plus [7]28.9525.1428.1826.2327.13
IBN [45]33.8532.3037.7527.9032.95
IW [46]29.9127.4829.7127.6128.68
Iternorm [25]31.8132.7033.8827.0731.37
DRPC [68]37.4232.1434.1228.0632.94
ISW [11]36.5835.2040.3328.3035.10
GTR [47]37.5333.7534.5228.1733.49
DIRL [65]41.0439.1541.60--
SHADE [75]44.6539.2843.34--
SAW [48]39.7537.3441.8630.7937.44
WildNet [34]44.6238.4246.0931.3440.12
AdvStyle [76]39.6235.5437.00--
SPC [26]44.1040.4645.51--
DPCL [66]44.7440.5946.3330.8140.62
BlindNet [1]45.7241.3247.0831.3941.38
PDAF45.4043.5649.4532.9842.85
" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.611, + 0.907, + 0.64 + ], + "angle": 0, + "content": "Table 2. Comparison with existing DGSS methods using a CNN-based network trained on GTAV (G) [52]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.667, + 0.905, + 0.698 + ], + "angle": 0, + "content": "Mask2Former-based methods, highlighting its flexibility in enhancing existing segmentation networks." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.7, + 0.907, + 0.866 + ], + "angle": 0, + "content": "GTAV Source Domain. Table 2 demonstrates the effectiveness of PDAF in bridging the domain gap between synthetic and real-world datasets. Unlike existing DGSS methods that rely primarily on feature normalization, PDAF explicitly models latent domain priors, providing structured guidance for feature alignment and enhancing domain generalization. As shown in Figure 4, qualitative comparisons indicate that PDAF effectively preserves detailed structural information—such as road boundaries and object contours—resulting in more precise and consistent segmentation outcomes." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Clear to Corruptions. We evaluate PDAF on ACDC [55], which introduces four real-world adverse conditions (rain," + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "21761" + } + ], + [ + { + "type": "table", + "bbox": [ + 0.107, + 0.089, + 0.472, + 0.239 + ], + "angle": 0, + "content": "
MethodBackboneTrained on Cityscape (C)
FoggyNightRainSnowAvg.
IBN [45]63.8021.2050.4049.6043.70
Itemorm [25]63.3023.8050.1049.9045.30
IW [46]Res5062.4021.8052.4047.6046.60
ISW [11]64.3024.3056.0049.8048.10
PDAF68.2830.1459.2353.5852.81
Mask2Former [10]56.4039.1058.9058.2053.15
HGFormer [15]Swin-T58.5043.3062.0058.3055.53
PDAF73.7547.7568.1262.2562.97
Mask2Former [10]69.1053.1068.3065.2063.93
HGFormer [15]Swin-L69.9052.7072.0068.6065.80
PDAF80.7255.1273.1371.4370.10
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.249, + 0.483, + 0.276 + ], + "angle": 0, + "content": "Table 3. Comparison with existing methods evaluated on ACDC [55]." + }, + { + "type": "table", + "bbox": [ + 0.141, + 0.291, + 0.435, + 0.364 + ], + "angle": 0, + "content": "
MethodsBMGSAvg.
Baseline44.9651.6842.5523.2940.62
PDAF w/o LPE49.5957.6346.3427.1945.19
PDAF w/o DCM51.5160.7949.4229.5547.82
PDAF w/o DPE51.9860.3349.3828.8647.64
PDAF53.5062.9350.5430.6849.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.122, + 0.375, + 0.452, + 0.389 + ], + "angle": 0, + "content": "Table 4. Effectiveness of proposed modules in PDAF." + }, + { + "type": "table", + "bbox": [ + 0.137, + 0.403, + 0.44, + 0.466 + ], + "angle": 0, + "content": "
\\( \\mathcal{L}_{\\text{task}} \\)\\( \\mathcal{L}_{\\text{sc}} \\)\\( \\mathcal{L}_{\\text{prior}} \\)BMGSAvg.
✓--49.7658.7347.3227.5145.83
✓✓-51.2859.5848.3827.9446.80
✓-✓53.3662.1149.6930.1048.82
✓✓✓53.5062.9350.5430.6849.41
" + }, + { + "type": "table_caption", + "bbox": [ + 0.135, + 0.477, + 0.439, + 0.491 + ], + "angle": 0, + "content": "Table 5. Analysis on the proposed loss functions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.518, + 0.484, + 0.622 + ], + "angle": 0, + "content": "fog, night and snow). We employ DeepLabV3Plus [7] and Mask2Former [10] as the segmentation network and train our method on Cityscapes [12]. As shown in Table 3, PDAF achieves consistent improvement under all adversarial conditions, demonstrating that LDP effectively compensates for degraded features caused by image corruptions and enhances robustness in varied environments." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.633, + 0.255, + 0.648 + ], + "angle": 0, + "content": "5.2. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.656, + 0.483, + 0.761 + ], + "angle": 0, + "content": "To validate the effectiveness of the proposed modules, we perform an ablation study. All experiments are conducted using DeepLabV3Plus [7] with ResNet-50 [21] as the segmentation network. The model is trained on Cityscapes (C) [12] and evaluated on BDD-100K (B) [67], Mapillary (M) [42], GTAV (G) [52], and SYNTHIA (S) [54] to assess generalization across diverse domain shifts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of Proposed Modules. We consider several experimental configurations: (i) Baseline: vanilla DeepLabV3Plus [7] trained without PDAF; (ii) PDAF w/o LPE: where variational posterior is removed, resulting in LDP estimation without cross-domain guidance; (iii) PDAF w/o DCM: where LDP is directly added to the original features without employing DCM; (iv) PDAF w/o DPE: where DPE is replaced by a single-source LPE, omitting the diffusion process; and (v) PDAF: which integrates all proposed" + }, + { + "type": "table", + "bbox": [ + 0.54, + 0.09, + 0.882, + 0.122 + ], + "angle": 0, + "content": "
Number of c'1416642561024
Avg. of mIoU48.7449.4148.6648.3546.1445.69
" + }, + { + "type": "table_caption", + "bbox": [ + 0.522, + 0.132, + 0.898, + 0.146 + ], + "angle": 0, + "content": "Table 6. Ablation study on the number of channels for LDP." + }, + { + "type": "table", + "bbox": [ + 0.532, + 0.161, + 0.888, + 0.193 + ], + "angle": 0, + "content": "
Number of Steps12481632
Avg. of mIoU48.8849.1249.4149.4349.4649.48
" + }, + { + "type": "table_caption", + "bbox": [ + 0.525, + 0.204, + 0.894, + 0.218 + ], + "angle": 0, + "content": "Table 7. Ablation study on the number of timestep in DPE." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.246, + 0.905, + 0.291 + ], + "angle": 0, + "content": "modules. As shown in Table 4, each module enhances segmentation performance, with LPE playing a pivotal role in posterior regularization for structured LDP modeling." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.294, + 0.906, + 0.43 + ], + "angle": 0, + "content": "Analysis of Loss Functions. Table 5 demonstrates that each loss term contributes to improved performance. The prior constraint loss, \\(\\mathcal{L}_{\\mathrm{prior}}\\), enforces consistency between the LDPs estimated by LPE and DPE, enabling DPE to accurately infer LDP solely from target samples. Additionally, the semantic consistency loss, \\(\\mathcal{L}_{\\mathrm{sc}}\\), enhances feature alignment by constraining predictions from both source and pseudo-target samples, further bolstering domain generalization." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.434, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Impact of LDP Channel Dimensions. LDP serve as a variational prior for modeling domain shifts while guiding feature alignment in DGSS. The channel dimension of LDP is crucial for capturing domain-specific variations. As shown in Table 6, a channel dimension of 4 yields the best overall performance. In contrast, higher channel dimensions introduce feature redundancy and increase the risk of overfitting, which limits further performance improvements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.559, + 0.906, + 0.65 + ], + "angle": 0, + "content": "Impact of steps for DPE. DPE iteratively refines LDP through the denoising process, facilitating precise alignment across diverse target domains. As shown in Table 7, performance gradually improves with an increasing number of diffusion steps. To achieve an optimal trade-off between efficiency and accuracy, we set the diffusion timestep to 4." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.664, + 0.634, + 0.68 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.689, + 0.906, + 0.901 + ], + "angle": 0, + "content": "This paper introduces PDAF, a probabilistic diffusion alignment framework designed for DGSS. By explicitly modeling latent domain prior (LDP), PDAF captures domain-specific variations to facilitate robust feature alignment across unseen target domains. To achieve this, PDAF involves a latent prior extractor to infer cross-domain relationships, a domain compensation module to condition segmentation features on LDP and a diffusion prior estimator to estimate LDP through a denoising process, allowing robust generalization to unseen domains. Extensive experiments demonstrate the effectiveness of PDAF, improving the generalization of existing segmentation models across diverse environments, highlighting the robustness of LDP modeling in domain generalized semantic segmentation." + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21762" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.207 + ], + "angle": 0, + "content": "We thank to National Center for High-performance Computing (NCHC) for providing computational and storage resources. This research was supported by Taiwan's National Science and Technology Council under Grant NSTC 111-2221-E-002-136-MY3 and NSTC 114-2221-E-002-067-MY3." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.221, + 0.188, + 0.237 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.246, + 0.484, + 0.302 + ], + "angle": 0, + "content": "[1] Woojin Ahn, Geun Yeong Yang, Hyun Duck Choi, and Myo Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.303, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. In NeurIPS, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.346, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[3] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on robot learning, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.403, + 0.484, + 0.459 + ], + "angle": 0, + "content": "[4] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.46, + 0.484, + 0.502 + ], + "angle": 0, + "content": "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.504, + 0.484, + 0.559 + ], + "angle": 0, + "content": "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.561, + 0.484, + 0.615 + ], + "angle": 0, + "content": "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.617, + 0.484, + 0.685 + ], + "angle": 0, + "content": "[8] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Hua-En Chang, Jian-Jiun Ding, and Sy-Yen Kuo. Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning. In ECCV, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[9] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In AAAI, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.745, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2021. 1, 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.858 + ], + "angle": 0, + "content": "[11] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.86, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.246, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "dataset for semantic urban scene understanding. In CVPR, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[14] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[15] Jian Ding, Nan Xue, Guisong Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[16] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[17] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.344, + 0.905, + 0.372 + ], + "angle": 0, + "content": "[18] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.372, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[19] Muhammad Ghifary, W. Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In ICCV, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[20] Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Mangas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.51 + ], + "angle": 0, + "content": "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2015. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.511, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[22] Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.54, + 0.905, + 0.594 + ], + "angle": 0, + "content": "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.595, + 0.905, + 0.637 + ], + "angle": 0, + "content": "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.637, + 0.905, + 0.678 + ], + "angle": 0, + "content": "[25] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, 2019. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.679, + 0.905, + 0.734 + ], + "angle": 0, + "content": "[26] Wei Huang, Chang Wen Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.735, + 0.905, + 0.775 + ], + "angle": 0, + "content": "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[28] Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. arXiv preprint arXiv:2406.00777, 2024.3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[29] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.479, + 0.945, + 0.521, + 0.957 + ], + "angle": 0, + "content": "21763" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.178, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[32] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2013. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.208, + 0.483, + 0.248 + ], + "angle": 0, + "content": "[33] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.25, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[34] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.483, + 0.333 + ], + "angle": 0, + "content": "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.335, + 0.483, + 0.376 + ], + "angle": 0, + "content": "[36] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex Chichung Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.483, + 0.431 + ], + "angle": 0, + "content": "[37] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.434, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[38] Qingguo Liu, Chenyi Zhuang, Pan Gao, and Jie Qin. Cdformer: when degradation prediction embraces diffusion model for blind image super-resolution. In CVPR, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.491, + 0.483, + 0.544 + ], + "angle": 0, + "content": "[39] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.547, + 0.483, + 0.601 + ], + "angle": 0, + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.604, + 0.483, + 0.644 + ], + "angle": 0, + "content": "[41] Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.647, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[42] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.689, + 0.483, + 0.73 + ], + "angle": 0, + "content": "[43] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In AAAI, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.483, + 0.786 + ], + "angle": 0, + "content": "[44] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In Int. Conf. on Intel. Robots and Systems, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.483, + 0.828 + ], + "angle": 0, + "content": "[45] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 2, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.831, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[46] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 3, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.874, + 0.483, + 0.899 + ], + "angle": 0, + "content": "[47] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liua. Global and local texture randomization for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.486, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.548, + 0.093, + 0.905, + 0.119 + ], + "angle": 0, + "content": "synthetic-to-real semantic segmentation. IEEE TIP, 2021. 2,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.162 + ], + "angle": 0, + "content": "[48] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[49] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.207, + 0.905, + 0.262 + ], + "angle": 0, + "content": "[50] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637v2, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.264, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[51] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.905, + 0.36 + ], + "angle": 0, + "content": "[52] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.363, + 0.905, + 0.403 + ], + "angle": 0, + "content": "[53] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[54] Germán Ros, Laura Sellart, Joanna Materzynska, David Vázquez, and Antonio M. López. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.463, + 0.905, + 0.503 + ], + "angle": 0, + "content": "[55] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.505, + 0.905, + 0.545 + ], + "angle": 0, + "content": "[56] Seonguk Seo, Yumin Suh, Dongwan Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.587 + ], + "angle": 0, + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.591, + 0.905, + 0.631 + ], + "angle": 0, + "content": "[58] Naveen Venkat, Jogendra Nath Kundu, Durgesh Singh, Ambareesh Revanur, et al. Your classifier can secretly suffice multi-source domain adaptation. NIPS, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.634, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.677, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[60] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.719, + 0.905, + 0.772 + ], + "angle": 0, + "content": "[61] Zongbin Wang, Bin Pan, Shiyu Shen, Tianyang Shi, and Zhenwei Shi. Domain generalization guided by large-scale pre-trained priors. arXiv preprint arXiv:2406.05628, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.776, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[62] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[63] Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. NeurIPS, 2023. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "21764" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[64] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[65] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[66] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.263, + 0.482, + 0.319 + ], + "angle": 0, + "content": "[67] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2018. 1, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.32, + 0.482, + 0.387 + ], + "angle": 0, + "content": "[68] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto L. Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 2, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.39, + 0.482, + 0.444 + ], + "angle": 0, + "content": "[69] Haojie Zhang, Yongyi Su, Xun Xu, and Kui Jia. Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation. In CVPR, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[70] Juntao Zhang, Yuehuai Liu, Yu-Wing Tai, and Chi-Keung Tang. C3net: Compound conditioned control for multimodal content generation. In CVPR, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.482, + 0.53 + ], + "angle": 0, + "content": "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.532, + 0.482, + 0.585 + ], + "angle": 0, + "content": "[72] Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.588, + 0.482, + 0.643 + ], + "angle": 0, + "content": "[73] Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher-guided robust tuning of vision foundation models for domain generalized segmentation. In CVPR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.646, + 0.482, + 0.699 + ], + "angle": 0, + "content": "[74] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.702, + 0.482, + 0.743 + ], + "angle": 0, + "content": "[75] Yuyang Zhao, Zhun Zhong, Na Zhao, N. Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 2, 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.482, + 0.786 + ], + "angle": 0, + "content": "[76] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and N. Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 2, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.786 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.48, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "21765" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_origin.pdf b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..c67693453b3b50d32daa37aeda3b399eb2e7c981 --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/293d6594-cb1e-4cf4-8140-8fc705f162e7_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6fb9be89454ce7eb7990cf4db3cc4b831d4f9ead3214229c3e29d17f5b5a145 +size 10351530 diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/full.md b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eebb39ef448a67689120b2148df8f4c68c9c5819 --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/full.md @@ -0,0 +1,465 @@ +# Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation + +I-Hsiang Chen $^{1,2*,\dagger}$ Hua-En Chang $^{1*}$ Wei-Ting Chen $^{3}$ Jenq-Neng Hwang $^{2}$ Sy-Yen Kuo $^{1,4}$ + +$^{1}$ National Taiwan University $^{2}$ University of Washington $^{3}$ Microsoft $^{4}$ Chang Gung University + +![](images/09a34eda5f6c4fe065d992174737b340116711ec76ef6d676ea86718baf40cf9.jpg) +Input + +![](images/9fc9d664f8f71143c10bc4a5a310ce3883de73fb8b322b157ec49ada266d79d7.jpg) + +![](images/fa5ed57ff3e71523a9139e5e6a776588803ef8a89286046b9bfc81e9a3357704.jpg) +Input + +![](images/c552f6f053db62a4452791fa0779bfda6ee78d20ec52c81e9b3da0b00c110554.jpg) +(a) Qualitative comparison between PDAF and its backbones +Figure 1. The proposed PDAF enhances existing segmentation backbones (DeepLabV3Plus [7] and Mask2Former [10]) by leveraging latent domain prior modeling as a compensation mechanism, yielding consistent improvements on unseen target domains such as BDD100K [67]. + +![](images/1ccf65a3e979c1efa3f87f4a80c775ccf6abaffe7cdd4205df920a765d7c43a7.jpg) +DeepLabV3Plus + +![](images/fd2b5b3668205a89ecab4283c519de64ca492e9b4754935639eea3d3fc944012.jpg) + +![](images/d5a937519db47f9b13822931eb3beb981eb3162b56b91af2175d6acd9b0b8818.jpg) +Mask2Former + +![](images/81a198799bc201b612de3fcf01effc14e864f5427cc2cb7c0676be2d9268eff3.jpg) + +![](images/164e3eb26b432fe92f43a126670cb49175598f495392ce375bf3ae46bd4ef25e.jpg) +PDAF + +![](images/5fd74c7ea111b540e9617995901819f4984a0d784310a7f8420e67043579e499.jpg) + +![](images/92841ec651ba45a3669e930584f5f9733e806c1048485f2e3c729cf29945d525.jpg) +PDAF + +![](images/673bee209971741d0e944c3dce10903d6b2f930684fef2d00e51fb6f44e57ef4.jpg) + +![](images/a55e8a41870fb6d49f656d7aa262bfb4693910d392cce831e4342d56a0e207fc.jpg) +Ground Truth + +![](images/bb4bffc791e9ab59c5451e3d92bb594f0cae9926cfb49009948c0e05e2208695.jpg) + +![](images/cf3791240a2ed224c10be1b25c91c245291c9efc75d381ef4cf1e9c9c84f50d4.jpg) +Ground Truth + +![](images/36c8a7eb6ae3a6ddd6562d7ba8e0928b0e47c3b0b3af2225746ea0374390dfb4.jpg) + +![](images/7b917f6661e98c0f6acf641fe67c6d21206000df8e99ff2d6e2adad19cba6c8f.jpg) + +![](images/04b5fe741d99e0b12de6641b8a9984d634e738379b0d0920e9d5d237e432fbb7.jpg) + +![](images/d812e8e0d1942657943ec70329e413a0e8483150ccf91f0f92bf72594643b33b.jpg) +(b) Overview of PDAF framework. + +# Abstract + +Domain Generalized Semantic Segmentation (DGSS) is a critical yet challenging task, as domain shifts in unseen environments can severely compromise model performance. While recent studies enhance feature alignment by projecting features into the source domain, they often neglect intrinsic latent domain priors, leading to suboptimal results. In this paper, we introduce PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. PDAF introduces a Latent Domain Prior + +(LDP) to capture domain shifts and uses this prior as a conditioning factor to align both source and unseen target domains. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired source and pseudotarget images to simulate latent domain shifts, enabling LDP modeling. The framework comprises three modules: the Latent Prior Extractor (LPE) predicts the LDP by supervising domain shifts; the Domain Compensation Module (DCM) adjusts feature representations to mitigate domain shifts; and the Diffusion Prior Estimator (DPE) leverages a diffusion process to estimate the LDP without requiring paired samples. This design enables PDAF to iteratively model domain shifts, progressively refining feature representations to enhance generalization under complex target conditions. Extensive experiments validate the effectiveness of PDAF across diverse and challenging urban scenes. + +# 1. Introduction + +Domain Generalized Semantic Segmentation (DGSS) remains a fundamental challenge in computer vision, primarily due to inevitable domain shifts encountered in unseen environments [51, 69]. Segmentation models trained on a source domain frequently degrade on the unseen target domains [2, 8, 37]. This challenge is particularly critical in real-world applications such as autonomous driving [3, 23] and robotic systems [43, 44], where accurate semantic segmentation is essential for ensuring operational safety. Factors such as lighting variations, weather changes, and other underrepresented conditions in training data exacerbate the generalization problem. Addressing the domain shifts thus remains a key obstacle in developing robust and reliable semantic segmentation models. + +Within the field of DGSS, two primary methodologies have emerged: data augmentation-based approaches and domain-invariant representation learning-based approaches. Data augmentation-based methods enhance the diversity of training data by introducing synthetic variations, thereby exposing models to a broader spectrum of potential domain shifts and improving their generalization ability [24, 75, 76]. Nevertheless, a crucial limitation of data augmentation-based methods is their significant dependence on auxiliary domains or generation models [18, 29, 39]. Conversely, domain-invariant representation learning methods aim to train models that extract consistent features across different domains, effectively mitigating the impact of domain shift [2, 9, 16, 19, 35, 36, 41, 56]. However, the inherent entanglement between style and content poses a significant challenge, often resulting in the loss of critical semantic information during feature decomposition [1, 26, 66]. Recent works achieve feature alignment by projecting features into a constrained feature space [1, 26, 66], but they often overlook the intrinsic properties of latent domain priors, leading to suboptimal performance [58, 61]. + +To overcome this limitation, we propose PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. Specifically, PDAF introduces a Latent Domain Prior (LDP), which models domain-specific variations and guides feature alignment between source and unseen target domains, as illustrated in Figure 1. Specifically, PDAF integrates into a pre-trained segmentation model and leverages paired source and pseudotarget images to simulate domain shifts. This process enables LDP modeling, which serves as a foundation for robust domain alignment. To leverage these priors effectively, PDAF first utilizes the Latent Prior Extractor (LPE) to capture cross-domain relationships between the source and pseudo-target samples, generating an optimal LDP that encodes domain-specific variations. Building upon this, the + +Domain Compensation Module (DCM) conditions feature representations on the extracted LDP, ensuring that domain priors guide the adaptation process to mitigate domain shifts while preserving task-relevant information. However, the direct estimation of LDP from explicit source-target pairs may be infeasible in real-world scenarios; to overcome this, the Diffusion Prior Estimator (DPE) employs probabilistic diffusion modeling to infer LDP without requiring paired samples, enhancing the model's adaptability to unseen domains. Through this progressive refinement, PDAF not only aligns features across domains but also maintains critical variations essential for generalization, leading to improved segmentation performance in complex environments. + +Extensive experiments on benchmark datasets validate the effectiveness of PDAF, demonstrating that probabilistic modeling of LDP successfully enhances the domain generalization of existing segmentation networks. + +Our main contributions are as follows: + +- We introduce PDAF, a novel probabilistic diffusion alignment framework that leverages probabilistic diffusion modeling of latent domain priors (LDP) to enhance the robustness of existing segmentation models against unseen domain shifts. +- We propose three components for PDAF: the latent prior extractor, the domain compensation module, and the diffusion prior estimator. These modules collectively estimate and leverage the LDP for robust feature alignment between source and unseen target domains. +- Extensive experiments across diverse datasets and degradation scenarios demonstrate that our method effectively improves the domain generalization of two widely adopted segmentation models (e.g., DeepLabV3Plus [7] and Mask2Former [10]), achieving state-of-the-art performance. + +# 2. Related Work + +Domain Generalized Semantic Segmentation. Domain Generalized Semantic Segmentation (DGSS) enhances model robustness to unseen domains without requiring target data. Existing approaches fall into data augmentation-based methods and domain-invariant representation learning-based methods. + +Data augmentation-based methods improve generalization by synthesizing or randomizing styles. DRPC [68] enforces pyramid consistency across stylized versions using real images as style references. GTR-LTR [47] replaces source styles with artistic paintings for global and local texture variation, while WildNet [34] enhances feature-level representations with ImageNet [13] as an external data source. Domain-invariant representation learning-based methods remove domain-specific information via normalization and whitening. IBN-Net [45] integrates Instance Normalization (IN) and Batch Normalization (BN) [27], + +while Switchable Whitening (SW) [46] adaptively merges whitening and standardization. ISW [11] and DIRL [65] reduce style-sensitive covariance components, whereas SHADE [75] and SAN-SAW [48] introduce style hallucination and semantic-aware alignment. HGFormer [15] and CMFormer [5] leverage self-attention's visual grouping to handle style variations, demonstrating the strengths of transformer-based models. + +Recent studies reveal that excessive style regularization can weaken content discrimination. SPC [26] projects target styles into a source-style space, while DPCL [66] applies multi-level contrastive learning on projected source features. BlindNet [1] integrates covariance alignment and semantic consistency contrastive learning to reduce style sensitivity in the encoder while enhancing decoder robustness. In addition, some methods explore Vision Foundation Models (VFMs) to improve robustness and transferability in semantic segmentation [4, 62, 72, 73]. Despite progress, DGSS methods focus on feature alignment and augmentation, leaving domain prior modeling underexplored. To bridge this gap, we introduce probabilistic diffusion modeling, enabling structured domain prior estimation and improved generalization to unseen domains. + +Diffusion Models. Diffusion models (DMs) [22] have recently gained significant attention for their outstanding performance in generative modeling. By parameterizing a Markov chain and optimizing the variational lower bound on the likelihood function, DMs effectively capture complex data distributions and produce high-quality samples[22, 30, 33, 59, 70, 71, 74], outperforming conventional frameworks like GANs [14]. To reduce computational overhead, LDM [53] and Wüstchen [50] conduct the diffusion process in latent space while retaining high-quality outputs. In addition to this, several studies have leveraged DMs to encode prior knowledge that facilitate downstream tasks. For instance, DiffIR [64] introduces a diffusion process applied to a compact image prior representation, enhancing both efficiency and stability under challenging conditions, while CDFormer [38] jointly reconstructs content and degraded representations to improve textural fidelity, and UniRestore [6] further bridges perceptual and task-oriented image restoration by leveraging diffusion priors. These advancements highlight the role of diffusion-based methods in capturing complex data distribution, enhancing robustness and adaptability in generative modeling + +Diffusion Models for Domain Generalization. Recent works explore diffusion models for domain generalization in semantic segmentation. DatasetDM [63] synthesizes diverse image-annotation pairs, while DGIn-Style [29] employs semantic guidance to produce consistently paired image-label samples. Gong et al. [20] adopt prompt randomization with pretrained diffusion features, and DIFF [28] integrates sampling and fusion from pre + +trained diffusion models to learn universal feature representations. These methods demonstrate the effectiveness of external diffusion-based models in enhancing domain generalization, but they typically incur substantial computational overhead. By contrast, our approach adopts an efficient probabilistic diffusion modeling framework to explicitly model latent domain priors, capturing diverse domain variations without imposing excessive resource demands. Moreover, this design seamlessly integrates with existing segmentation network, enabling robust out-of-distribution performance with minimal additional overhead. + +# 3. Method + +We first introduce the Latent Domain Prior (LDP) Modeling (Sec. 3.1), where we formulate domain generalized semantic segmentation (DGSS) as a probabilistic learning problem and derive the variational inference framework for LDP estimation. Based on this, we develop the Probabilistic Diffusion Alignment Framework (Sec. 3.2), which leverages LDPs to model domain shifts and guide the existing segmentation model in adapting feature representations. Finally, we describe Model Optimization (Sec. 3.3), detailing the training objectives and loss functions designed to enhance robust domain generalization. + +# 3.1. Latent Domain Prior Modeling + +We propose a probabilistic modeling framework that explicitly integrates LDP to model domain shifts in the DGSS task. By formulating LDP within a variational inference framework, our method enables principled estimation of domain prior that serves as conditioning variables, thereby guiding segmentation models to generalize more effectively on unseen target domains. + +Probabilistic Problem Formulation. To generalize across domain shifts, we cast DGSS as a probabilistic learning problem. We introduce a Latent Domain Prior (LDP) as a latent variable $z$ to capture domain-specific variations that are not directly observable. Given the target domain distribution $p(x_{t},y_{t})$ , where $y_{t}$ denotes the ground truth, we infer $z$ solely from the available target-domain image $x_{t}$ . Consequently, the prediction function of PDAF is formulated as: + +$$ +p _ {\theta , \phi} \left(y _ {t} \mid x _ {t}\right) = \int p _ {\theta} \left(y _ {t} \mid x _ {t}, z\right) p _ {\phi} \left(z \mid x _ {t}\right) d z, \tag {1} +$$ + +where $\theta$ and $\phi$ represent the parameters of the segmentation model and the network designed to estimate $z$ (i.e., LDP), respectively. By incorporating LDP, PDAF regularizes feature adaptation, consequently mitigating domain shifts and enhancing model robustness. + +Variational Inference for LDP Estimation. To facilitate tractable inference of LDP, we introduce a variational posterior $q_{\varphi}(z|x_t,x_s)$ to approximate the true posterior distribution $p(z|x_t,x_s)$ , where $\varphi$ represents the parameters + +![](images/a7408cf0f9b1ee702682731fafed0759b7d3fa7eef7fffd1c251e877422905bc.jpg) +Figure 2. Overall architecture of PDAF. PDAF augments a pre-trained segmentation network by introducing LDP modeling to enhance its domain generalization. The LPE learns the optimal LDP by modeling cross-domain relationships between source and pseudo-target domains, while DPE is employed to estimate LDP using only target inputs. Finally, the DCM enhances segmentation network with LDP guidance, refining feature alignment and improving domain generalization. + +of the inference network and $x_{s}$ indicates the source domain counterpart of target domain image $x_{t}$ . The pair of $x_{s}$ and $x_{t}$ preserves semantic consistency while exhibiting domain-specific variations. By leveraging variational inference, we construct a principled optimization framework that estimates LDP to enhance domain alignment. The Evidence Lower Bound (ELBO) of the predictive function is formulated as: + +$$ +\begin{array}{l} \log p _ {\theta , \phi} \left(y _ {t} \mid x _ {t}\right) \geq \mathbb {E} _ {q _ {\varphi} \left(z \mid x _ {t}, x _ {s}\right)} \left[ \log p _ {\theta} \left(y _ {t} \mid x _ {t}, z\right) \right] \tag {2} \\ - \mathbb {K L} \left[ q _ {\varphi} (z | x _ {t}, x _ {s}) | | p _ {\phi} (z | x _ {t}) \right]. \\ \end{array} +$$ + +The variational posterior $q_{\varphi}(z|x_t,x_s)$ explicitly encodes cross-domain feature correlations from source-target pairs, enabling precise modeling of latent domain variation critical for effective generalization. The KL divergence term encourages the learned posterior to approximate the prior, ensuring structured regularization. A detailed derivation of ELBO is provided in the supplementary material. + +# 3.2. Probabilistic Diffusion Alignment Framework + +Overview of the Framework. We introduce the Probabilistic Diffusion Alignment Framework (PDAF), which instantiates LDP modeling to obtain domain shift, thereby enhancing generalization performance of segmentation models. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired images generated + +through data augmentation, effectively simulating domain shifts. This process preserves semantic consistency while injecting domain-specific variations, thus facilitating the principled ELBO approximation. The optimization function of PDAF is described as follows: + +$$ +\begin{array}{l} \mathcal {L} _ {\mathrm {P D A F}} = - \mathbb {E} _ {q _ {\varphi} \left(z ^ {\prime} \mid x _ {t ^ {\prime}}, x _ {s}\right)} \left[ \log p _ {\theta} \left(y _ {t ^ {\prime}} \mid x _ {t ^ {\prime}}, z ^ {\prime}\right) \right] \\ + \mathbb {K L} \left(q _ {\varphi} \left(z ^ {\prime} | x _ {t ^ {\prime}}, x _ {s}\right) | | p _ {\phi} \left(z ^ {\prime} | x _ {t}\right)\right), \tag {3} \\ \end{array} +$$ + +where $x_{t'}$ represents pseudo-target images generated from source images $x_s$ via data augmentation, while $y_{t'}$ is the corresponding ground truth. $z'$ denotes the LDP, inferred through cross-domain relationships between source and pseudo-target domain samples. The prediction model $p_{\theta}(.)$ , variational posterior $q_{\varphi}(.)$ and prior $p_{\phi}(.)$ jointly optimize segmentation accuracy while modeling domain shifts by maximizing likelihood and minimizing KL divergence, respectively. + +Corresponding to the optimization function in Eq. 3, PDAF introduces three components: (i) a Latent Prior Extractor (LPE), which serves as $q_{\varphi}(.)$ and infers the optimal LDP from source and pseudo-target pairs; (ii) a Domain Compensation Module (DCM), integrated into a pretrained segmentation network to implement the prediction model $p_{\theta}(.)$ , which leverages the extracted LDP to refine feature representations, thereby mitigating domain shifts and improving segmentation accuracy; and (iii) a Diffusion Prior Estimator (DPE), implementing $p_{\phi}(.)$ through probabilistic + +![](images/93dc5a1cfb1bea9e7d80a33f89adc6a69bdf0661c46140c3c34e2c0571bdaa34.jpg) +(a) Details of LPE + +![](images/054ffc7bd43f082a8bf16f571b7a5c7806895e7e37b6ef7a2ef99d30ae01037b.jpg) +(b) Details of DCM +Figure 3. Schematic diagrams of (a) the Latent Prior Extractor, (b) the Domain Compensation Module and (c) the Diffusion Prior Estimator. + +![](images/e8a729b231389f50f222bfb3619eda4a414d5500bfd1c94813761632b2e8eb6d.jpg) +(c) Details of DPE + +diffusion modeling to estimate LDP without paired source and target data. Notably, PDAF is compatible with any existing segmentation model, allowing frozen pre-trained encoders $E_{\vartheta}$ and decoders $D_{\vartheta}$ to discover LDP, and then finetuning the target network $E_{\theta}$ and $D_{\theta}$ for DGSS. The overall architecture of PDAF is illustrated in Figure 2. + +Latent Prior Extractor. The LPE is designed to estimate the optimal LDP by supervising cross-domain feature relationships, providing effective guidance for feature alignment. As illustrated in Figure 3a, the LPE first concatenates source features $h_{\vartheta,s}$ and pseudo-target features $h_{\vartheta,t'}$ , then utilizes residual blocks to model cross-domain feature relationships. The extracted features are then processed through two projection layers to derive the probabilistic modeling parameters, namely the mean $(\mu)$ and variance $(\sigma)$ . To ensure stable and effective training, we constrain the optimal LDP $\tilde{z}$ to follow a standard normal distribution [32], enhancing latent space regularization and enabling more robust prior estimation. The procedure can be formulated as: + +$$ +\begin{array}{l} h _ {\vartheta , s} = E _ {\vartheta} \left(x _ {s}\right), h _ {\vartheta , t ^ {\prime}} = E _ {\vartheta} \left(x _ {t ^ {\prime}}\right), \\ \mu , \sigma = F _ {\varphi} \left(h _ {\vartheta , s}, h _ {\vartheta , t ^ {\prime}}\right), \tag {4} \\ \tilde {z} = \mu + \epsilon \sigma , \epsilon \sim \mathcal {N} (0, I), \\ \end{array} +$$ + +where $F_{\varphi}$ indicates LPE parameterized by $\varphi$ . The dimension of LDP is $\tilde{z} \in \mathbb{R}^{c' \times h \times w}$ , where $c'$ denotes a predefined number of the channel of LDP and $h, w$ correspond to the spatial dimensions of the feature maps. + +Domain Compensation Module. The DCM enhances segmentation models for domain generalization. Drawing inspiration from Spatial Feature Transformation (SFT) [60], DCM leverages LDP as a compensation mechanism, providing domain-aware modulation to improve feature alignment. As illustrated in Figure 3b, a straightforward approach involves projecting LDP onto affine transformation parameters using dedicated projection layers $(F_{\Psi})$ , yielding scale $(\tilde{\gamma})$ and shift $(\tilde{\beta})$ parameters. These parameters are then used to rescale and shift the feature representations, effectively normalizing domain discrepancies across spatial regions. The refined features $\tilde{h}_{\theta ,t^{\prime}}$ are subsequently pro + +cessed by the segmentation head $D_{\theta}$ to generate accurate segmentation maps $\tilde{y}_{t'}$ . The process can be formulated as: + +$$ +\tilde {\gamma}, \tilde {\beta} = F _ {\Psi} (\tilde {z}), +$$ + +$$ +\tilde {h} _ {\theta , t ^ {\prime}} = \tilde {\gamma} \odot h _ {\theta , t ^ {\prime}} \oplus \tilde {\beta}, \tag {5} +$$ + +$$ +\tilde {y} _ {t ^ {\prime}} = D _ {\theta} \left(\tilde {h} _ {\vartheta , t ^ {\prime}}\right), +$$ + +where $\odot$ and $\oplus$ denote element-wise multiplication and element-wise addition, respectively. The dimensions of the affine transformation parameters correspond to the feature map, with $\tilde{\gamma},\tilde{\beta}\in \mathbb{R}^{c\times h\times w}$ . + +Diffusion Prior Estimator. Diffusion Models (DMs) have demonstrated their effectiveness in sampling from complex target domains, significantly improving various computer vision tasks as evidenced in previous studies [38, 50, 53]. In DPE, we leverage probabilistic diffusion modeling to estimate the LDP, enabling arbitrary target domain and accurate prior estimation for domain alignment. As illustrated in Figure 3c, we first apply the forward diffusion process to an optimal LDP $\tilde{z}$ , iteratively adding Gaussian noise to obtain an intermediate latent state $z_{T}$ : + +$$ +q \left(z _ {T} | \tilde {z}\right) = \mathcal {N} \left(z _ {T}; \sqrt {\bar {\alpha} _ {T}} \tilde {z}, (1 - \bar {\alpha} _ {T}) I\right), \tag {6} +$$ + +where $\alpha_{T} = 1 - \beta_{T}$ and $\bar{\alpha}_{T} = \prod_{i=1}^{T} \alpha_{i}$ . For the reverse diffusion process, the DPE initializes from $\hat{z}_{T}$ and performs T-step denoising process to estimate the LDP $\hat{z}_{0}$ , which is formulated as: + +$$ +\hat {z} _ {T - 1} = \frac {1}{\sqrt {\alpha_ {T}}} \left(\hat {z} _ {T} - \epsilon \frac {1 - \alpha_ {T}}{\sqrt {1 - \bar {\alpha} _ {T}}}\right). \tag {7} +$$ + +Specifically, to estimate LDP for target images, we condition the DPE on the extracted target features $h_{\theta ,t'}$ . DPE $(F_{\phi})$ then performs a denoising operation at timestep $T$ , represented as $F_{\phi}(\hat{z}_{T},T,h_{\theta ,t'})$ . To accommodate the reduced channel dimension of LDP, we employ accelerated diffusion-based optimization [57] for efficient and rapid inference. Moreover, this design allows for the joint training of DPE and the segmentation head $D_{\theta}$ , thereby optimizing + +LDP directly for the segmentation task. This process is defined as follows: + +$$ +\hat {z} _ {0} = F _ {\phi} (\hat {z} _ {T}, T, h _ {\theta , t ^ {\prime}}), +$$ + +$$ +\hat {\gamma}, \hat {\beta} = F _ {\Psi} (\hat {z} _ {0}), \tag {8} +$$ + +$$ +\hat {h} _ {\theta , t ^ {\prime}} = \hat {\gamma} \odot h _ {\theta , t ^ {\prime}} \oplus \hat {\beta}, +$$ + +$$ +\hat {y} _ {t ^ {\prime}} = D _ {\theta} (\hat {h} _ {\vartheta , t ^ {\prime}}), +$$ + +where $\hat{y}_{t'}$ represents the predicted segmentation maps conditions on the estimated LDP $\hat{z}_0$ . The dimensions of the affine transformation parameters correspond to the feature map, with $\hat{\gamma}, \hat{\beta} \in \mathbb{R}^{c \times h \times w}$ . + +Inference. During inference, only target domain images $x_{t}$ are available, so the variational posterior is omitted. Instead, we extract target features $h_{\theta ,t}$ using the segmentation backbone $E_{\theta}(\cdot)$ as the conditioning input and sample a Gaussian noise vector $\hat{z}_T$ . The DPE then iteratively refines this noise over $T$ denoising steps to produce the estimated LDP $\hat{z}_0$ . Finally, the DCM leverages the estimated LDP to refine the feature representations, which are used for robust segmentation prediction. + +# 3.3. Model Optimization + +We train our framework by optimizing the likelihood estimation and enforcing posterior regularization to enhance domain generalization, as defined in Eq. (3). To achieve maximum likelihood estimation, we employ task-specific loss functions $\mathcal{L}_{\mathrm{task}}$ tailored to different segmentation models. Specifically, we use weighted cross-entropy loss for DeepLabV3+ [7] and focal loss for Mask2Former [10]. To further enhance feature consistency, we introduce a semantic consistency loss $\mathcal{L}_{\mathrm{sc}}$ that measures the discrepancy between the predictions of the source and pseudo-target images. This loss enforces semantic alignment in feature space and accelerates model convergence: + +$$ +\mathcal {L} _ {\mathrm {s c}} = \left\| f _ {\vartheta} \left(x _ {s}\right) - \tilde {y} _ {t ^ {\prime}} \right\| _ {2} + \left\| f _ {\vartheta} \left(x _ {s}\right) - \hat {y} _ {t ^ {\prime}} \right\| _ {2}, \tag {9} +$$ + +where $f_{\vartheta}$ denotes the frozen segmentation network $E_{\vartheta}(.)$ and $D_{\vartheta}(.)$ , $\tilde{y}_{t'}$ and $\hat{y}_{t'}$ represent the predictions from Eq. (5) and Eq. (8), respectively. To minimize the KL divergence and enforce posterior regularization, we adopt a prior constraint loss $\mathcal{L}_{\mathrm{prior}}$ to align the outputs of LPE and DPE, thereby implicitly reducing the distributional discrepancy between the posterior and the prior: + +$$ +\mathcal {L} _ {\text {p r i o r}} = \| \hat {z} _ {0} - \tilde {z} \| _ {2}. \tag {10} +$$ + +where $\hat{z}_0$ and $\tilde{z}$ indicate the optimal LDP and estimated LDP from Eq. (4) and Eq. (8), respectively. The overall loss is: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \lambda_ {\text {t a s k}} \mathcal {L} _ {\text {t a s k}} + \lambda_ {\mathrm {s c}} \mathcal {L} _ {\mathrm {s c}} + \lambda_ {\text {p r i o r}} \mathcal {L} _ {\text {p r i o r}}, \tag {11} +$$ + +where $\lambda_{\mathrm{task}}$ , $\lambda_{\mathrm{sc}}$ , and $\lambda_{\mathrm{prior}}$ represents the scaling factors. + +# 4. Implementation Details + +# 4.1. Datasets + +We conduct experiments on five urban scene semantic segmentation datasets that share 19 common scene categories. + +Cityscapes (C) [12] consists of high-resolution images captured from 50 distinct cities, where we exclusively use the finely annotated subset containing 2,975 training and 500 validation images. + +BDD-100K (B) [67] provides 7,000 training images and 1,000 validation images collected from various locations conditions. + +Mapillary (M) [42] is a large-scale dataset containing 25,000 high-resolution street-view images sourced globally. GTA5 (G) [52] includes 24,966 images simulated using the Grand Theft Auto V game engine, divided into 12,403 training, 6,382 validation, and 6,181 testing samples. + +SYNTHIA (S) [54] consists of 9,400 photorealistic synthetic images, including 2,830 for testing. + +# 4.2. Evaluation Protocols + +Following existing DGSS benchmarks [1, 11, 15, 45, 46, 49], we adopt a single-source domain generalization setting, where one dataset is used for training (source domain), and the remaining four datasets are treated as unseen target domains for evaluation. For a fair comparison, all experiments are built on DeepLabV3Plus [7] (with ResNet-50 [21]) and Mask2Former [10] (with Swin-T and Swin-L [40]). We use the mean intersection over the union (mIoU) [17] to measure the performance of segmentation. + +# 4.3. Training Details + +To facilitate a fair evaluation, we exclusively employ Photometric Augmentation as the data augmentation strategy to generate pseudo-target domain samples. The model is trained with Adam Optimizer [31] with an initial learning rate of $1 \times 10^{-5}$ , momentum parameters $(\beta_{1},\beta_{2}) = (0.9,0.999)$ , and a batch size of 4 for 100 epochs. The channel of LDP is set to $c' = 4$ . In DPE, we adopt a total diffusion timestep of $T = 4$ , with the noise variance $\beta_{t}$ linearly increasing from $\beta_{1} = 0.1$ to $\beta_{T} = 0.99$ . The loss coefficients are adjusted to $(\lambda_{\mathrm{task}},\lambda_{\mathrm{sc}},\lambda_{\mathrm{prior}}) = (0.5,0.5,1.0)$ , balancing the contributions of different loss terms. All experiments are implemented in PyTorch and conducted on a single NVIDIA RTX 4090 GPU. More details of the implementation and hyperparameters analysis are provided in the Supplementary Material. + +# 5. Experiments + +# 5.1. Comparison with State-of-the-arts + +To access the effectiveness of PDAF, we compare our approach against a comprehensive set of state-of-the-art meth + +![](images/f3fc0319fb6e225f02f4cb33b4c699e2ca28d1e079e01165c94f1395e9b208c3.jpg) + +![](images/7d163fe516cfbe098b12324d14ec783f727dee55adcd7326740f1c62e0d8ae9d.jpg) + +![](images/165c7e4eef114c0eb381cd64c25aa2a144536e4dfd5382ee83e7486be19aff84.jpg) +Input + +![](images/b2660d3a756f451534a98804057300856c424c0c2667a911b881219344688ba3.jpg) + +![](images/a1e5e81f42ac5523e1a2b48932c8ced41f546414eb152a0553d8607dad422c22.jpg) + +![](images/0a3dfe1443b9652ca045eb7d75b02eef45d29391636c82cdffee6291507cf607.jpg) +DeepLabV3Plus [7] + +![](images/6f8a960ec99f3b14a4f911c4e6d25b9d0645bb81677247ca7faeb42e1969e0ba.jpg) + +![](images/19e597b728a49a813bfd76748a2ff58996ff97b047c5593fa85cdf0440c88908.jpg) + +![](images/2f53b884f22fba0c6a87b5c3b9fe8c566f050df8e709256db53c93e49b4a85a7.jpg) +ISW [11] + +![](images/562139c18de1ab8f1bed582fad83f1f88eb0faaff0659106a763b66f242c95d5.jpg) + +![](images/ed2d78f2b6b74e645b4b3a688bd37f10354064915aab39660226ce0405c50b6d.jpg) + +![](images/7106d2d10e74c9140b485a7d656e585a5dea84283c55394f8acc70d648ceaa62.jpg) +SHADE [75] +Figure 4. Qualitative comparison between DGSS methods trained on GTAV (G) [52] using DeepLabV3Plus [7] with ResNet-50 [21]. The first, second, and third rows show the predictions for Cityscapes (C) [12], BDD-100K (B) [67], and Mapillary (M) [42], respectively. + +![](images/9a1088d99bfe9667cec19e2e700240293e0ed35a079ad70362ecf43c14d0534c.jpg) + +![](images/327e9b3ee22ccb2177af0b192b53920e204b035251f40fcd4ec46adfdccbf248.jpg) + +![](images/4d7be68d2f9913c97b84a58107b44f1147e586a00bd76dd439c42318fcd4c0ad.jpg) +WildNet [34] + +![](images/299c53624a2e292d5470ebadf3995512803b23b9e28abe14677d7f1f6a8c9d8a.jpg) + +![](images/f68322b4936bbe68b29d9fc4a6b4e5d6d3d0ee7daa6612046dc2a9d042f61cf7.jpg) + +![](images/e4dd0f0daeb0c5e63f477d5c82b73e1d9c2b5319d122d86d60a4e038888249bf.jpg) +BlindNet [1] + +![](images/03ed4bbfe0227a35fb998b549d91ae0457fba1f65823d7f6bd9efa1360371e32.jpg) + +![](images/9a372b15018a9575633c2f9d738dda67deb7f5bf490d02fe9953887193a22eb5.jpg) + +![](images/01bb4ee1a8fa13ba1cfbda9a377008786cb85755bc24e5949030fec56a42f499.jpg) +PDAF + +![](images/438216f6227b4b0cf09247f11276d9d72b594effa57324bc9c41f387057cd51c.jpg) + +![](images/7bf49b21301531a8bdf53d7671b5637f06f4d2005f9f35a92dba010e1f0e1973.jpg) + +![](images/2810a97e5384679a90d4e0594ffefec88c46f4d64cf16a503cc88a5716f8fba6.jpg) +Ground Truth + +
MethodBackboneTrained on Cityscapes (C)
BMGSAvg.
DeepLabV3Plus [7]Res5044.9651.6842.5523.2940.62
IBN [45]48.5657.0445.0626.1444.20
IW [46]48.4955.8244.8726.1043.82
Itemorm [25]49.2356.2645.7325.9844.30
DRPC [68]49.8656.3445.6226.5844.60
ISW [11]50.7358.6445.0026.2045.14
GTR [47]50.7557.1645.7926.4745.04
DIRL [65]51.80-46.5226.50-
SHADE [75]50.9560.6748.6127.6246.96
SAW [48]52.9559.8147.2828.3247.09
WildNet [34]50.9458.7947.0127.9546.17
DPCL [66]50.9758.5946.0025.8545.34
BlindNet [1]51.8460.1847.9728.5147.13
PDAF53.5062.9350.5430.6849.41
Mask2Former [10]51.3065.3050.6034.0050.30
HGFformer [15]Swin-T53.4066.9051.3033.6051.30
PDAF56.4066.4055.2034.8053.20
Mask2Former [10]60.1072.2057.8042.4058.13
HGFformer [15]Swin-L61.5072.1059.4041.3058.58
CMFormer [5]62.6073.6060.7043.0059.98
PDAF63.0074.1063.2044.0061.08
+ +ods including both CNN-based techniques [1, 7, 11, 25, 26, 34, 45-48, 65, 66, 68, 75, 76] and Mask2Former-based approaches [5, 10, 15]. We adhere to standard DGSS evaluation protocols and report results directly cited from prior works [1, 5, 11, 46, 48]. + +Cityscapes Source Domain. Table 1 shows that PDAF achieves the best overall performance across four unseen datasets, demonstrating strong generalization to diverse domains. PDAF effectively transfers from real-world datasets (C) [12] to target domains with distinct visual distributions, such as game-rendered synthetic scenes (G) [52] and photo-realistic simulations (S) [54]. Notably, PDAF consistently improves performance across CNN-based and + +Table 1. Comparison with existing methods trained on Cityscapes (C) [12]. The best and second-best mIoU scores are blue and magenta. + +
MethodTrained on GTAV (G)
CBMSAvg.
DeepLabV3Plus [7]28.9525.1428.1826.2327.13
IBN [45]33.8532.3037.7527.9032.95
IW [46]29.9127.4829.7127.6128.68
Iternorm [25]31.8132.7033.8827.0731.37
DRPC [68]37.4232.1434.1228.0632.94
ISW [11]36.5835.2040.3328.3035.10
GTR [47]37.5333.7534.5228.1733.49
DIRL [65]41.0439.1541.60--
SHADE [75]44.6539.2843.34--
SAW [48]39.7537.3441.8630.7937.44
WildNet [34]44.6238.4246.0931.3440.12
AdvStyle [76]39.6235.5437.00--
SPC [26]44.1040.4645.51--
DPCL [66]44.7440.5946.3330.8140.62
BlindNet [1]45.7241.3247.0831.3941.38
PDAF45.4043.5649.4532.9842.85
+ +Table 2. Comparison with existing DGSS methods using a CNN-based network trained on GTAV (G) [52]. + +Mask2Former-based methods, highlighting its flexibility in enhancing existing segmentation networks. + +GTAV Source Domain. Table 2 demonstrates the effectiveness of PDAF in bridging the domain gap between synthetic and real-world datasets. Unlike existing DGSS methods that rely primarily on feature normalization, PDAF explicitly models latent domain priors, providing structured guidance for feature alignment and enhancing domain generalization. As shown in Figure 4, qualitative comparisons indicate that PDAF effectively preserves detailed structural information—such as road boundaries and object contours—resulting in more precise and consistent segmentation outcomes. + +Clear to Corruptions. We evaluate PDAF on ACDC [55], which introduces four real-world adverse conditions (rain, + +
MethodBackboneTrained on Cityscape (C)
FoggyNightRainSnowAvg.
IBN [45]63.8021.2050.4049.6043.70
Itemorm [25]63.3023.8050.1049.9045.30
IW [46]Res5062.4021.8052.4047.6046.60
ISW [11]64.3024.3056.0049.8048.10
PDAF68.2830.1459.2353.5852.81
Mask2Former [10]56.4039.1058.9058.2053.15
HGFormer [15]Swin-T58.5043.3062.0058.3055.53
PDAF73.7547.7568.1262.2562.97
Mask2Former [10]69.1053.1068.3065.2063.93
HGFormer [15]Swin-L69.9052.7072.0068.6065.80
PDAF80.7255.1273.1371.4370.10
+ +Table 3. Comparison with existing methods evaluated on ACDC [55]. + +
MethodsBMGSAvg.
Baseline44.9651.6842.5523.2940.62
PDAF w/o LPE49.5957.6346.3427.1945.19
PDAF w/o DCM51.5160.7949.4229.5547.82
PDAF w/o DPE51.9860.3349.3828.8647.64
PDAF53.5062.9350.5430.6849.41
+ +Table 4. Effectiveness of proposed modules in PDAF. + +
\( \mathcal{L}_{\text{task}} \)\( \mathcal{L}_{\text{sc}} \)\( \mathcal{L}_{\text{prior}} \)BMGSAvg.
✓--49.7658.7347.3227.5145.83
✓✓-51.2859.5848.3827.9446.80
✓-✓53.3662.1149.6930.1048.82
✓✓✓53.5062.9350.5430.6849.41
+ +fog, night and snow). We employ DeepLabV3Plus [7] and Mask2Former [10] as the segmentation network and train our method on Cityscapes [12]. As shown in Table 3, PDAF achieves consistent improvement under all adversarial conditions, demonstrating that LDP effectively compensates for degraded features caused by image corruptions and enhances robustness in varied environments. + +# 5.2. Ablation Studies + +To validate the effectiveness of the proposed modules, we perform an ablation study. All experiments are conducted using DeepLabV3Plus [7] with ResNet-50 [21] as the segmentation network. The model is trained on Cityscapes (C) [12] and evaluated on BDD-100K (B) [67], Mapillary (M) [42], GTAV (G) [52], and SYNTHIA (S) [54] to assess generalization across diverse domain shifts. + +Effectiveness of Proposed Modules. We consider several experimental configurations: (i) Baseline: vanilla DeepLabV3Plus [7] trained without PDAF; (ii) PDAF w/o LPE: where variational posterior is removed, resulting in LDP estimation without cross-domain guidance; (iii) PDAF w/o DCM: where LDP is directly added to the original features without employing DCM; (iv) PDAF w/o DPE: where DPE is replaced by a single-source LPE, omitting the diffusion process; and (v) PDAF: which integrates all proposed + +Table 5. Analysis on the proposed loss functions. + +
Number of c'1416642561024
Avg. of mIoU48.7449.4148.6648.3546.1445.69
+ +Table 6. Ablation study on the number of channels for LDP. + +
Number of Steps12481632
Avg. of mIoU48.8849.1249.4149.4349.4649.48
+ +Table 7. Ablation study on the number of timestep in DPE. + +modules. As shown in Table 4, each module enhances segmentation performance, with LPE playing a pivotal role in posterior regularization for structured LDP modeling. + +Analysis of Loss Functions. Table 5 demonstrates that each loss term contributes to improved performance. The prior constraint loss, $\mathcal{L}_{\mathrm{prior}}$ , enforces consistency between the LDPs estimated by LPE and DPE, enabling DPE to accurately infer LDP solely from target samples. Additionally, the semantic consistency loss, $\mathcal{L}_{\mathrm{sc}}$ , enhances feature alignment by constraining predictions from both source and pseudo-target samples, further bolstering domain generalization. + +Impact of LDP Channel Dimensions. LDP serve as a variational prior for modeling domain shifts while guiding feature alignment in DGSS. The channel dimension of LDP is crucial for capturing domain-specific variations. As shown in Table 6, a channel dimension of 4 yields the best overall performance. In contrast, higher channel dimensions introduce feature redundancy and increase the risk of overfitting, which limits further performance improvements. + +Impact of steps for DPE. DPE iteratively refines LDP through the denoising process, facilitating precise alignment across diverse target domains. As shown in Table 7, performance gradually improves with an increasing number of diffusion steps. To achieve an optimal trade-off between efficiency and accuracy, we set the diffusion timestep to 4. + +# 6. Conclusion + +This paper introduces PDAF, a probabilistic diffusion alignment framework designed for DGSS. By explicitly modeling latent domain prior (LDP), PDAF captures domain-specific variations to facilitate robust feature alignment across unseen target domains. To achieve this, PDAF involves a latent prior extractor to infer cross-domain relationships, a domain compensation module to condition segmentation features on LDP and a diffusion prior estimator to estimate LDP through a denoising process, allowing robust generalization to unseen domains. Extensive experiments demonstrate the effectiveness of PDAF, improving the generalization of existing segmentation models across diverse environments, highlighting the robustness of LDP modeling in domain generalized semantic segmentation. + +# Acknowledgements + +We thank to National Center for High-performance Computing (NCHC) for providing computational and storage resources. This research was supported by Taiwan's National Science and Technology Council under Grant NSTC 111-2221-E-002-136-MY3 and NSTC 114-2221-E-002-067-MY3. + +# References + +[1] Woojin Ahn, Geun Yeong Yang, Hyun Duck Choi, and Myo Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3, 6, 7 +[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. In NeurIPS, 2018. 2 +[3] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on robot learning, 2023. 2 +[4] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3 +[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 3, 7 +[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 3 +[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2, 6, 7, 8 +[8] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Hua-En Chang, Jian-Jiun Ding, and Sy-Yen Kuo. Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning. In ECCV, 2022. 2 +[9] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In AAAI, 2022. 2 +[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2021. 1, 2, 6, 7, 8 +[11] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 3, 6, 7, 8 +[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes + +dataset for semantic urban scene understanding. In CVPR, 2016. 6, 7, 8 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2 +[14] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3 +[15] Jian Ding, Nan Xue, Guisong Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 6, 7, 8 +[16] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 2 +[17] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2014. 6 +[18] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2 +[19] Muhammad Ghifary, W. Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In ICCV, 2015. 2 +[20] Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Mangas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3 +[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2015. 6, 7, 8 +[22] Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3 +[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2 +[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2 +[25] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, 2019. 7, 8 +[26] Wei Huang, Chang Wen Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 2, 3, 7 +[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2 +[28] Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. arXiv preprint arXiv:2406.00777, 2024.3 +[29] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 2, 3 + +[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR, 2024. 3 +[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6 +[32] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2013. 5 +[33] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 3 +[34] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 2, 7 +[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2017. 2 +[36] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex Chichung Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 2 +[37] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2 +[38] Qingguo Liu, Chenyi Zhuang, Pan Gao, and Jie Qin. Cdformer: when degradation prediction embraces diffusion model for blind image super-resolution. In CVPR, 2024. 3, 5 +[39] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 2 +[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6 +[41] Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, 2017. 2 +[42] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6, 7, 8 +[43] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In AAAI, 2021. 2 +[44] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In Int. Conf. on Intel. Robots and Systems, 2021. 2 +[45] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 2, 6, 7, 8 +[46] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 3, 6, 7, 8 +[47] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liua. Global and local texture randomization for + +synthetic-to-real semantic segmentation. IEEE TIP, 2021. 2,7 +[48] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 3, 7 +[49] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 6 +[50] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637v2, 2023. 3, 5 +[51] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In CVPR, 2024. 2 +[52] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 6, 7, 8 +[53] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 3, 5 +[54] Germán Ros, Laura Sellart, Joanna Materzynska, David Vázquez, and Antonio M. López. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 6, 7, 8 +[55] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 7, 8 +[56] Seonguk Seo, Yumin Suh, Dongwan Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, 2019. 2 +[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5 +[58] Naveen Venkat, Jogendra Nath Kundu, Durgesh Singh, Ambareesh Revanur, et al. Your classifier can secretly suffice multi-source domain adaptation. NIPS, 2020. 2 +[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3 +[60] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5 +[61] Zongbin Wang, Bin Pan, Shiyu Shen, Tianyang Shi, and Zhenwei Shi. Domain generalization guided by large-scale pre-trained priors. arXiv preprint arXiv:2406.05628, 2024. 2 +[62] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3 +[63] Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. NeurIPS, 2023. 3 + +[64] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 3 +[65] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3, 7 +[66] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 2, 3, 7 +[67] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2018. 1, 6, 7, 8 +[68] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto L. Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 2, 7 +[69] Haojie Zhang, Yongyi Su, Xun Xu, and Kui Jia. Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation. In CVPR, 2024. 2 +[70] Juntao Zhang, Yuehuai Liu, Yu-Wing Tai, and Chi-Keung Tang. C3net: Compound conditioned control for multimodal content generation. In CVPR, 2024. 3 +[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3 +[72] Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In CVPR, 2025. 3 +[73] Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher-guided robust tuning of vision foundation models for domain generalized segmentation. In CVPR, 2025. 3 +[74] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3 +[75] Yuyang Zhao, Zhun Zhong, Na Zhao, N. Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 2, 3, 7 +[76] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and N. Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 2, 7 \ No newline at end of file diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/images.zip b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..dabd1004d63663bbd2fa1609eac9ab478d2e14bb --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d304b0c807aa3b6c330c9fdbe4fc1a10daaec427dfbee199b29c657ffcb25379 +size 653462 diff --git a/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/layout.json b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..3392270c2555a9013ba8009ff5f18162e9696378 --- /dev/null +++ b/2025/Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation/layout.json @@ -0,0 +1,12221 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "spans": [ + { + "bbox": [ + 62, + 102, + 549, + 140 + ], + "type": "text", + "content": "Exploring Probabilistic Modeling Beyond Domain Generalization for Semantic Segmentation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "spans": [ + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "content": "I-Hsiang Chen" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1,2*,\\dagger}" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "content": " Hua-En Chang" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "content": " Wei-Ting Chen" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "content": " Jenq-Neng Hwang" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "text", + "content": " Sy-Yen Kuo" + }, + { + "bbox": [ + 67, + 159, + 542, + 177 + ], + "type": "inline_equation", + "content": "^{1,4}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "spans": [ + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "text", + "content": "National Taiwan University " + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "text", + "content": "University of Washington " + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "text", + "content": "Microsoft " + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 67, + 188, + 541, + 205 + ], + "type": "text", + "content": "Chang Gung University" + } + ] + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 59, + 243, + 138, + 290 + ], + "blocks": [ + { + "bbox": [ + 88, + 234, + 107, + 243 + ], + "lines": [ + { + "bbox": [ + 88, + 234, + 107, + 243 + ], + "spans": [ + { + "bbox": [ + 88, + 234, + 107, + 243 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 243, + 138, + 290 + ], + "lines": [ + { + "bbox": [ + 59, + 243, + 138, + 290 + ], + "spans": [ + { + "bbox": [ + 59, + 243, + 138, + 290 + ], + "type": "image", + "image_path": "09a34eda5f6c4fe065d992174737b340116711ec76ef6d676ea86718baf40cf9.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 59, + 290, + 137, + 338 + ], + "blocks": [ + { + "bbox": [ + 59, + 290, + 137, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 290, + 137, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 290, + 137, + 338 + ], + "type": "image", + "image_path": "9fc9d664f8f71143c10bc4a5a310ce3883de73fb8b322b157ec49ada266d79d7.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 59, + 348, + 137, + 396 + ], + "blocks": [ + { + "bbox": [ + 89, + 339, + 107, + 348 + ], + "lines": [ + { + "bbox": [ + 89, + 339, + 107, + 348 + ], + "spans": [ + { + "bbox": [ + 89, + 339, + 107, + 348 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 348, + 137, + 396 + ], + "lines": [ + { + "bbox": [ + 59, + 348, + 137, + 396 + ], + "spans": [ + { + "bbox": [ + 59, + 348, + 137, + 396 + ], + "type": "image", + "image_path": "fa5ed57ff3e71523a9139e5e6a776588803ef8a89286046b9bfc81e9a3357704.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 59, + 396, + 137, + 444 + ], + "blocks": [ + { + "bbox": [ + 59, + 396, + 137, + 444 + ], + "lines": [ + { + "bbox": [ + 59, + 396, + 137, + 444 + ], + "spans": [ + { + "bbox": [ + 59, + 396, + 137, + 444 + ], + "type": "image", + "image_path": "c552f6f053db62a4452791fa0779bfda6ee78d20ec52c81e9b3da0b00c110554.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 116, + 448, + 313, + 458 + ], + "lines": [ + { + "bbox": [ + 116, + 448, + 313, + 458 + ], + "spans": [ + { + "bbox": [ + 116, + 448, + 313, + 458 + ], + "type": "text", + "content": "(a) Qualitative comparison between PDAF and its backbones" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "lines": [ + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 555, + 491 + ], + "type": "text", + "content": "Figure 1. The proposed PDAF enhances existing segmentation backbones (DeepLabV3Plus [7] and Mask2Former [10]) by leveraging latent domain prior modeling as a compensation mechanism, yielding consistent improvements on unseen target domains such as BDD100K [67]." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 139, + 243, + 218, + 290 + ], + "blocks": [ + { + "bbox": [ + 152, + 234, + 206, + 243 + ], + "lines": [ + { + "bbox": [ + 152, + 234, + 206, + 243 + ], + "spans": [ + { + "bbox": [ + 152, + 234, + 206, + 243 + ], + "type": "text", + "content": "DeepLabV3Plus" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 243, + 218, + 290 + ], + "lines": [ + { + "bbox": [ + 139, + 243, + 218, + 290 + ], + "spans": [ + { + "bbox": [ + 139, + 243, + 218, + 290 + ], + "type": "image", + "image_path": "1ccf65a3e979c1efa3f87f4a80c775ccf6abaffe7cdd4205df920a765d7c43a7.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 139, + 290, + 218, + 338 + ], + "blocks": [ + { + "bbox": [ + 139, + 290, + 218, + 338 + ], + "lines": [ + { + "bbox": [ + 139, + 290, + 218, + 338 + ], + "spans": [ + { + "bbox": [ + 139, + 290, + 218, + 338 + ], + "type": "image", + "image_path": "fd2b5b3668205a89ecab4283c519de64ca492e9b4754935639eea3d3fc944012.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 139, + 348, + 218, + 396 + ], + "blocks": [ + { + "bbox": [ + 156, + 339, + 203, + 348 + ], + "lines": [ + { + "bbox": [ + 156, + 339, + 203, + 348 + ], + "spans": [ + { + "bbox": [ + 156, + 339, + 203, + 348 + ], + "type": "text", + "content": "Mask2Former" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 139, + 348, + 218, + 396 + ], + "lines": [ + { + "bbox": [ + 139, + 348, + 218, + 396 + ], + "spans": [ + { + "bbox": [ + 139, + 348, + 218, + 396 + ], + "type": "image", + "image_path": "d5a937519db47f9b13822931eb3beb981eb3162b56b91af2175d6acd9b0b8818.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 139, + 396, + 218, + 444 + ], + "blocks": [ + { + "bbox": [ + 139, + 396, + 218, + 444 + ], + "lines": [ + { + "bbox": [ + 139, + 396, + 218, + 444 + ], + "spans": [ + { + "bbox": [ + 139, + 396, + 218, + 444 + ], + "type": "image", + "image_path": "81a198799bc201b612de3fcf01effc14e864f5427cc2cb7c0676be2d9268eff3.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 220, + 243, + 299, + 290 + ], + "blocks": [ + { + "bbox": [ + 249, + 234, + 271, + 242 + ], + "lines": [ + { + "bbox": [ + 249, + 234, + 271, + 242 + ], + "spans": [ + { + "bbox": [ + 249, + 234, + 271, + 242 + ], + "type": "text", + "content": "PDAF" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 243, + 299, + 290 + ], + "lines": [ + { + "bbox": [ + 220, + 243, + 299, + 290 + ], + "spans": [ + { + "bbox": [ + 220, + 243, + 299, + 290 + ], + "type": "image", + "image_path": "164e3eb26b432fe92f43a126670cb49175598f495392ce375bf3ae46bd4ef25e.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 220, + 290, + 299, + 338 + ], + "blocks": [ + { + "bbox": [ + 220, + 290, + 299, + 338 + ], + "lines": [ + { + "bbox": [ + 220, + 290, + 299, + 338 + ], + "spans": [ + { + "bbox": [ + 220, + 290, + 299, + 338 + ], + "type": "image", + "image_path": "5fd74c7ea111b540e9617995901819f4984a0d784310a7f8420e67043579e499.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 220, + 348, + 300, + 396 + ], + "blocks": [ + { + "bbox": [ + 249, + 339, + 271, + 348 + ], + "lines": [ + { + "bbox": [ + 249, + 339, + 271, + 348 + ], + "spans": [ + { + "bbox": [ + 249, + 339, + 271, + 348 + ], + "type": "text", + "content": "PDAF" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 220, + 348, + 300, + 396 + ], + "lines": [ + { + "bbox": [ + 220, + 348, + 300, + 396 + ], + "spans": [ + { + "bbox": [ + 220, + 348, + 300, + 396 + ], + "type": "image", + "image_path": "92841ec651ba45a3669e930584f5f9733e806c1048485f2e3c729cf29945d525.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 220, + 396, + 300, + 444 + ], + "blocks": [ + { + "bbox": [ + 220, + 396, + 300, + 444 + ], + "lines": [ + { + "bbox": [ + 220, + 396, + 300, + 444 + ], + "spans": [ + { + "bbox": [ + 220, + 396, + 300, + 444 + ], + "type": "image", + "image_path": "673bee209971741d0e944c3dce10903d6b2f930684fef2d00e51fb6f44e57ef4.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 302, + 242, + 381, + 290 + ], + "blocks": [ + { + "bbox": [ + 318, + 234, + 363, + 242 + ], + "lines": [ + { + "bbox": [ + 318, + 234, + 363, + 242 + ], + "spans": [ + { + "bbox": [ + 318, + 234, + 363, + 242 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 242, + 381, + 290 + ], + "lines": [ + { + "bbox": [ + 302, + 242, + 381, + 290 + ], + "spans": [ + { + "bbox": [ + 302, + 242, + 381, + 290 + ], + "type": "image", + "image_path": "a55e8a41870fb6d49f656d7aa262bfb4693910d392cce831e4342d56a0e207fc.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 302, + 290, + 380, + 338 + ], + "blocks": [ + { + "bbox": [ + 302, + 290, + 380, + 338 + ], + "lines": [ + { + "bbox": [ + 302, + 290, + 380, + 338 + ], + "spans": [ + { + "bbox": [ + 302, + 290, + 380, + 338 + ], + "type": "image", + "image_path": "bb4bffc791e9ab59c5451e3d92bb594f0cae9926cfb49009948c0e05e2208695.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 302, + 348, + 380, + 396 + ], + "blocks": [ + { + "bbox": [ + 318, + 339, + 363, + 348 + ], + "lines": [ + { + "bbox": [ + 318, + 339, + 363, + 348 + ], + "spans": [ + { + "bbox": [ + 318, + 339, + 363, + 348 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 302, + 348, + 380, + 396 + ], + "lines": [ + { + "bbox": [ + 302, + 348, + 380, + 396 + ], + "spans": [ + { + "bbox": [ + 302, + 348, + 380, + 396 + ], + "type": "image", + "image_path": "cf3791240a2ed224c10be1b25c91c245291c9efc75d381ef4cf1e9c9c84f50d4.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 302, + 396, + 380, + 444 + ], + "blocks": [ + { + "bbox": [ + 302, + 396, + 380, + 444 + ], + "lines": [ + { + "bbox": [ + 302, + 396, + 380, + 444 + ], + "spans": [ + { + "bbox": [ + 302, + 396, + 380, + 444 + ], + "type": "image", + "image_path": "36c8a7eb6ae3a6ddd6562d7ba8e0928b0e47c3b0b3af2225746ea0374390dfb4.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 402, + 234, + 510, + 289 + ], + "blocks": [ + { + "bbox": [ + 402, + 234, + 510, + 289 + ], + "lines": [ + { + "bbox": [ + 402, + 234, + 510, + 289 + ], + "spans": [ + { + "bbox": [ + 402, + 234, + 510, + 289 + ], + "type": "image", + "image_path": "7b917f6661e98c0f6acf641fe67c6d21206000df8e99ff2d6e2adad19cba6c8f.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 402, + 291, + 549, + 388 + ], + "blocks": [ + { + "bbox": [ + 402, + 291, + 549, + 388 + ], + "lines": [ + { + "bbox": [ + 402, + 291, + 549, + 388 + ], + "spans": [ + { + "bbox": [ + 402, + 291, + 549, + 388 + ], + "type": "image", + "image_path": "04b5fe741d99e0b12de6641b8a9984d634e738379b0d0920e9d5d237e432fbb7.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 402, + 388, + 509, + 439 + ], + "blocks": [ + { + "bbox": [ + 402, + 388, + 509, + 439 + ], + "lines": [ + { + "bbox": [ + 402, + 388, + 509, + 439 + ], + "spans": [ + { + "bbox": [ + 402, + 388, + 509, + 439 + ], + "type": "image", + "image_path": "d812e8e0d1942657943ec70329e413a0e8483150ccf91f0f92bf72594643b33b.jpg" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 419, + 448, + 533, + 457 + ], + "lines": [ + { + "bbox": [ + 419, + 448, + 533, + 457 + ], + "spans": [ + { + "bbox": [ + 419, + 448, + 533, + 457 + ], + "type": "text", + "content": "(b) Overview of PDAF framework." + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_caption" + } + ], + "index": 32 + }, + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "spans": [ + { + "bbox": [ + 152, + 514, + 200, + 526 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 542, + 297, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 542, + 297, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 542, + 297, + 662 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS) is a critical yet challenging task, as domain shifts in unseen environments can severely compromise model performance. While recent studies enhance feature alignment by projecting features into the source domain, they often neglect intrinsic latent domain priors, leading to suboptimal results. In this paper, we introduce PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. PDAF introduces a Latent Domain Prior" + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 504, + 556, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 556, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 556, + 696 + ], + "type": "text", + "content": "(LDP) to capture domain shifts and uses this prior as a conditioning factor to align both source and unseen target domains. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired source and pseudotarget images to simulate latent domain shifts, enabling LDP modeling. The framework comprises three modules: the Latent Prior Extractor (LPE) predicts the LDP by supervising domain shifts; the Domain Compensation Module (DCM) adjusts feature representations to mitigate domain shifts; and the Diffusion Prior Estimator (DPE) leverages a diffusion process to estimate the LDP without requiring paired samples. This design enables PDAF to iteratively model domain shifts, progressively refining feature representations to enhance generalization under complex target conditions. Extensive experiments validate the effectiveness of PDAF across diverse and challenging urban scenes." + } + ] + } + ], + "index": 37 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 69, + 674, + 247, + 685 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 674, + 247, + 685 + ], + "spans": [ + { + "bbox": [ + 69, + 674, + 247, + 685 + ], + "type": "text", + "content": "Project page: https://pdfa-f.iccv.github.io" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 71, + 685, + 169, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 71, + 685, + 169, + 693 + ], + "spans": [ + { + "bbox": [ + 71, + 685, + 169, + 693 + ], + "type": "text", + "content": "* Indicates equal contribution." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 693, + 295, + 712 + ], + "type": "text", + "content": "† This work was conducted while I-Hsiang Chen was an intern at University of Washington." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21755" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 136, + 83 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 294, + 258 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation (DGSS) remains a fundamental challenge in computer vision, primarily due to inevitable domain shifts encountered in unseen environments [51, 69]. Segmentation models trained on a source domain frequently degrade on the unseen target domains [2, 8, 37]. This challenge is particularly critical in real-world applications such as autonomous driving [3, 23] and robotic systems [43, 44], where accurate semantic segmentation is essential for ensuring operational safety. Factors such as lighting variations, weather changes, and other underrepresented conditions in training data exacerbate the generalization problem. Addressing the domain shifts thus remains a key obstacle in developing robust and reliable semantic segmentation models." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 259, + 296, + 521 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 259, + 296, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 259, + 296, + 521 + ], + "type": "text", + "content": "Within the field of DGSS, two primary methodologies have emerged: data augmentation-based approaches and domain-invariant representation learning-based approaches. Data augmentation-based methods enhance the diversity of training data by introducing synthetic variations, thereby exposing models to a broader spectrum of potential domain shifts and improving their generalization ability [24, 75, 76]. Nevertheless, a crucial limitation of data augmentation-based methods is their significant dependence on auxiliary domains or generation models [18, 29, 39]. Conversely, domain-invariant representation learning methods aim to train models that extract consistent features across different domains, effectively mitigating the impact of domain shift [2, 9, 16, 19, 35, 36, 41, 56]. However, the inherent entanglement between style and content poses a significant challenge, often resulting in the loss of critical semantic information during feature decomposition [1, 26, 66]. Recent works achieve feature alignment by projecting features into a constrained feature space [1, 26, 66], but they often overlook the intrinsic properties of latent domain priors, leading to suboptimal performance [58, 61]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 713 + ], + "type": "text", + "content": "To overcome this limitation, we propose PDAF, a Probabilistic Diffusion Alignment Framework that enhances the generalization of existing segmentation networks through probabilistic diffusion modeling. Specifically, PDAF introduces a Latent Domain Prior (LDP), which models domain-specific variations and guides feature alignment between source and unseen target domains, as illustrated in Figure 1. Specifically, PDAF integrates into a pre-trained segmentation model and leverages paired source and pseudotarget images to simulate domain shifts. This process enables LDP modeling, which serves as a foundation for robust domain alignment. To leverage these priors effectively, PDAF first utilizes the Latent Prior Extractor (LPE) to capture cross-domain relationships between the source and pseudo-target samples, generating an optimal LDP that encodes domain-specific variations. Building upon this, the" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "Domain Compensation Module (DCM) conditions feature representations on the extracted LDP, ensuring that domain priors guide the adaptation process to mitigate domain shifts while preserving task-relevant information. However, the direct estimation of LDP from explicit source-target pairs may be infeasible in real-world scenarios; to overcome this, the Diffusion Prior Estimator (DPE) employs probabilistic diffusion modeling to infer LDP without requiring paired samples, enhancing the model's adaptability to unseen domains. Through this progressive refinement, PDAF not only aligns features across domains but also maintains critical variations essential for generalization, leading to improved segmentation performance in complex environments." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 275 + ], + "type": "text", + "content": "Extensive experiments on benchmark datasets validate the effectiveness of PDAF, demonstrating that probabilistic modeling of LDP successfully enhances the domain generalization of existing segmentation networks." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 326, + 276, + 482, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 276, + 482, + 286 + ], + "spans": [ + { + "bbox": [ + 326, + 276, + 482, + 286 + ], + "type": "text", + "content": "Our main contributions are as follows:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 326, + 288, + 553, + 479 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 326, + 288, + 553, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 288, + 553, + 346 + ], + "spans": [ + { + "bbox": [ + 326, + 288, + 553, + 346 + ], + "type": "text", + "content": "- We introduce PDAF, a novel probabilistic diffusion alignment framework that leverages probabilistic diffusion modeling of latent domain priors (LDP) to enhance the robustness of existing segmentation models against unseen domain shifts." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 326, + 348, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 348, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 326, + 348, + 553, + 407 + ], + "type": "text", + "content": "- We propose three components for PDAF: the latent prior extractor, the domain compensation module, and the diffusion prior estimator. These modules collectively estimate and leverage the LDP for robust feature alignment between source and unseen target domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 326, + 407, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 407, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 326, + 407, + 553, + 479 + ], + "type": "text", + "content": "- Extensive experiments across diverse datasets and degradation scenarios demonstrate that our method effectively improves the domain generalization of two widely adopted segmentation models (e.g., DeepLabV3Plus [7] and Mask2Former [10]), achieving state-of-the-art performance." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 400, + 502 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 510, + 553, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 581 + ], + "type": "text", + "content": "Domain Generalized Semantic Segmentation. Domain Generalized Semantic Segmentation (DGSS) enhances model robustness to unseen domains without requiring target data. Existing approaches fall into data augmentation-based methods and domain-invariant representation learning-based methods." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 555, + 713 + ], + "type": "text", + "content": "Data augmentation-based methods improve generalization by synthesizing or randomizing styles. DRPC [68] enforces pyramid consistency across stylized versions using real images as style references. GTR-LTR [47] replaces source styles with artistic paintings for global and local texture variation, while WildNet [34] enhances feature-level representations with ImageNet [13] as an external data source. Domain-invariant representation learning-based methods remove domain-specific information via normalization and whitening. IBN-Net [45] integrates Instance Normalization (IN) and Batch Normalization (BN) [27]," + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 758 + ], + "type": "text", + "content": "21756" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 167 + ], + "type": "text", + "content": "while Switchable Whitening (SW) [46] adaptively merges whitening and standardization. ISW [11] and DIRL [65] reduce style-sensitive covariance components, whereas SHADE [75] and SAN-SAW [48] introduce style hallucination and semantic-aware alignment. HGFormer [15] and CMFormer [5] leverage self-attention's visual grouping to handle style variations, demonstrating the strengths of transformer-based models." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 171, + 295, + 350 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 171, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 55, + 171, + 295, + 350 + ], + "type": "text", + "content": "Recent studies reveal that excessive style regularization can weaken content discrimination. SPC [26] projects target styles into a source-style space, while DPCL [66] applies multi-level contrastive learning on projected source features. BlindNet [1] integrates covariance alignment and semantic consistency contrastive learning to reduce style sensitivity in the encoder while enhancing decoder robustness. In addition, some methods explore Vision Foundation Models (VFMs) to improve robustness and transferability in semantic segmentation [4, 62, 72, 73]. Despite progress, DGSS methods focus on feature alignment and augmentation, leaving domain prior modeling underexplored. To bridge this gap, we introduce probabilistic diffusion modeling, enabling structured domain prior estimation and improved generalization to unseen domains." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 352, + 295, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 352, + 295, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 352, + 295, + 616 + ], + "type": "text", + "content": "Diffusion Models. Diffusion models (DMs) [22] have recently gained significant attention for their outstanding performance in generative modeling. By parameterizing a Markov chain and optimizing the variational lower bound on the likelihood function, DMs effectively capture complex data distributions and produce high-quality samples[22, 30, 33, 59, 70, 71, 74], outperforming conventional frameworks like GANs [14]. To reduce computational overhead, LDM [53] and Wüstchen [50] conduct the diffusion process in latent space while retaining high-quality outputs. In addition to this, several studies have leveraged DMs to encode prior knowledge that facilitate downstream tasks. For instance, DiffIR [64] introduces a diffusion process applied to a compact image prior representation, enhancing both efficiency and stability under challenging conditions, while CDFormer [38] jointly reconstructs content and degraded representations to improve textural fidelity, and UniRestore [6] further bridges perceptual and task-oriented image restoration by leveraging diffusion priors. These advancements highlight the role of diffusion-based methods in capturing complex data distribution, enhancing robustness and adaptability in generative modeling" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 714 + ], + "type": "text", + "content": "Diffusion Models for Domain Generalization. Recent works explore diffusion models for domain generalization in semantic segmentation. DatasetDM [63] synthesizes diverse image-annotation pairs, while DGIn-Style [29] employs semantic guidance to produce consistently paired image-label samples. Gong et al. [20] adopt prompt randomization with pretrained diffusion features, and DIFF [28] integrates sampling and fusion from pre" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 204 + ], + "type": "text", + "content": "trained diffusion models to learn universal feature representations. These methods demonstrate the effectiveness of external diffusion-based models in enhancing domain generalization, but they typically incur substantial computational overhead. By contrast, our approach adopts an efficient probabilistic diffusion modeling framework to explicitly model latent domain priors, capturing diverse domain variations without imposing excessive resource demands. Moreover, this design seamlessly integrates with existing segmentation network, enabling robust out-of-distribution performance with minimal additional overhead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "spans": [ + { + "bbox": [ + 314, + 214, + 370, + 227 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 234, + 555, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 234, + 555, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 234, + 555, + 365 + ], + "type": "text", + "content": "We first introduce the Latent Domain Prior (LDP) Modeling (Sec. 3.1), where we formulate domain generalized semantic segmentation (DGSS) as a probabilistic learning problem and derive the variational inference framework for LDP estimation. Based on this, we develop the Probabilistic Diffusion Alignment Framework (Sec. 3.2), which leverages LDPs to model domain shifts and guide the existing segmentation model in adapting feature representations. Finally, we describe Model Optimization (Sec. 3.3), detailing the training objectives and loss functions designed to enhance robust domain generalization." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 373, + 484, + 385 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 373, + 484, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 373, + 484, + 385 + ], + "type": "text", + "content": "3.1. Latent Domain Prior Modeling" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "spans": [ + { + "bbox": [ + 313, + 390, + 554, + 474 + ], + "type": "text", + "content": "We propose a probabilistic modeling framework that explicitly integrates LDP to model domain shifts in the DGSS task. By formulating LDP within a variational inference framework, our method enables principled estimation of domain prior that serves as conditioning variables, thereby guiding segmentation models to generalize more effectively on unseen target domains." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": "Probabilistic Problem Formulation. To generalize across domain shifts, we cast DGSS as a probabilistic learning problem. We introduce a Latent Domain Prior (LDP) as a latent variable " + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": " to capture domain-specific variations that are not directly observable. Given the target domain distribution " + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "inline_equation", + "content": "p(x_{t},y_{t})" + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "inline_equation", + "content": "y_{t}" + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": " denotes the ground truth, we infer " + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": " solely from the available target-domain image " + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 313, + 474, + 554, + 570 + ], + "type": "text", + "content": ". Consequently, the prediction function of PDAF is formulated as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 349, + 578, + 555, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 578, + 555, + 602 + ], + "spans": [ + { + "bbox": [ + 349, + 578, + 555, + 602 + ], + "type": "interline_equation", + "content": "p _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) = \\int p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) p _ {\\phi} \\left(z \\mid x _ {t}\\right) d z, \\tag {1}", + "image_path": "3563c6b441727e45225262b28d142dc72999b251ac1529ef740550b604671fb1.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " represent the parameters of the segmentation model and the network designed to estimate " + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 313, + 605, + 554, + 665 + ], + "type": "text", + "content": " (i.e., LDP), respectively. By incorporating LDP, PDAF regularizes feature adaptation, consequently mitigating domain shifts and enhancing model robustness." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": "Variational Inference for LDP Estimation. To facilitate tractable inference of LDP, we introduce a variational posterior " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(z|x_t,x_s)" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " to approximate the true posterior distribution " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "p(z|x_t,x_s)" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 313, + 665, + 554, + 714 + ], + "type": "text", + "content": " represents the parameters" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21757" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 65, + 73, + 537, + 312 + ], + "blocks": [ + { + "bbox": [ + 65, + 73, + 537, + 312 + ], + "lines": [ + { + "bbox": [ + 65, + 73, + 537, + 312 + ], + "spans": [ + { + "bbox": [ + 65, + 73, + 537, + 312 + ], + "type": "image", + "image_path": "a7408cf0f9b1ee702682731fafed0759b7d3fa7eef7fffd1c251e877422905bc.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "lines": [ + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "spans": [ + { + "bbox": [ + 54, + 323, + 555, + 369 + ], + "type": "text", + "content": "Figure 2. Overall architecture of PDAF. PDAF augments a pre-trained segmentation network by introducing LDP modeling to enhance its domain generalization. The LPE learns the optimal LDP by modeling cross-domain relationships between source and pseudo-target domains, while DPE is employed to estimate LDP using only target inputs. Finally, the DCM enhances segmentation network with LDP guidance, refining feature alignment and improving domain generalization." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "spans": [ + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "content": "of the inference network and " + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "content": " indicates the source domain counterpart of target domain image " + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "content": ". The pair of " + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "inline_equation", + "content": "x_{s}" + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 54, + 388, + 296, + 483 + ], + "type": "text", + "content": " preserves semantic consistency while exhibiting domain-specific variations. By leveraging variational inference, we construct a principled optimization framework that estimates LDP to enhance domain alignment. The Evidence Lower Bound (ELBO) of the predictive function is formulated as:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 79, + 491, + 295, + 521 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 79, + 491, + 295, + 521 + ], + "spans": [ + { + "bbox": [ + 79, + 491, + 295, + 521 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\log p _ {\\theta , \\phi} \\left(y _ {t} \\mid x _ {t}\\right) \\geq \\mathbb {E} _ {q _ {\\varphi} \\left(z \\mid x _ {t}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t} \\mid x _ {t}, z\\right) \\right] \\tag {2} \\\\ - \\mathbb {K L} \\left[ q _ {\\varphi} (z | x _ {t}, x _ {s}) | | p _ {\\phi} (z | x _ {t}) \\right]. \\\\ \\end{array}", + "image_path": "107946a4a65ca7d90a28a42c5f85cf45534a5198e03d06f54f54d1af283026a9.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 529, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 529, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 54, + 529, + 296, + 615 + ], + "type": "text", + "content": "The variational posterior " + }, + { + "bbox": [ + 54, + 529, + 296, + 615 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(z|x_t,x_s)" + }, + { + "bbox": [ + 54, + 529, + 296, + 615 + ], + "type": "text", + "content": " explicitly encodes cross-domain feature correlations from source-target pairs, enabling precise modeling of latent domain variation critical for effective generalization. The KL divergence term encourages the learned posterior to approximate the prior, ensuring structured regularization. A detailed derivation of ELBO is provided in the supplementary material." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 621, + 294, + 634 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 621, + 294, + 634 + ], + "spans": [ + { + "bbox": [ + 55, + 621, + 294, + 634 + ], + "type": "text", + "content": "3.2. Probabilistic Diffusion Alignment Framework" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 641, + 296, + 714 + ], + "type": "text", + "content": "Overview of the Framework. We introduce the Probabilistic Diffusion Alignment Framework (PDAF), which instantiates LDP modeling to obtain domain shift, thereby enhancing generalization performance of segmentation models. To achieve this, PDAF integrates into a pre-trained segmentation model and utilizes paired images generated" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "spans": [ + { + "bbox": [ + 313, + 388, + 555, + 448 + ], + "type": "text", + "content": "through data augmentation, effectively simulating domain shifts. This process preserves semantic consistency while injecting domain-specific variations, thus facilitating the principled ELBO approximation. The optimization function of PDAF is described as follows:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 343, + 452, + 553, + 482 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 343, + 452, + 553, + 482 + ], + "spans": [ + { + "bbox": [ + 343, + 452, + 553, + 482 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {\\mathrm {P D A F}} = - \\mathbb {E} _ {q _ {\\varphi} \\left(z ^ {\\prime} \\mid x _ {t ^ {\\prime}}, x _ {s}\\right)} \\left[ \\log p _ {\\theta} \\left(y _ {t ^ {\\prime}} \\mid x _ {t ^ {\\prime}}, z ^ {\\prime}\\right) \\right] \\\\ + \\mathbb {K L} \\left(q _ {\\varphi} \\left(z ^ {\\prime} | x _ {t ^ {\\prime}}, x _ {s}\\right) | | p _ {\\phi} \\left(z ^ {\\prime} | x _ {t}\\right)\\right), \\tag {3} \\\\ \\end{array}", + "image_path": "0529d1f392e2c3dcdd2f0c136da8d0b77b4056a5fdfc856c3294a93e8ab46616.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "x_{t'}" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " represents pseudo-target images generated from source images " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "x_s" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " via data augmentation, while " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "y_{t'}" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " is the corresponding ground truth. " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "z'" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " denotes the LDP, inferred through cross-domain relationships between source and pseudo-target domain samples. The prediction model " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "p_{\\theta}(.)" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": ", variational posterior " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(.)" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " and prior " + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "inline_equation", + "content": "p_{\\phi}(.)" + }, + { + "bbox": [ + 313, + 486, + 555, + 594 + ], + "type": "text", + "content": " jointly optimize segmentation accuracy while modeling domain shifts by maximizing likelihood and minimizing KL divergence, respectively." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": "Corresponding to the optimization function in Eq. 3, PDAF introduces three components: (i) a Latent Prior Extractor (LPE), which serves as " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "q_{\\varphi}(.)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " and infers the optimal LDP from source and pseudo-target pairs; (ii) a Domain Compensation Module (DCM), integrated into a pretrained segmentation network to implement the prediction model " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "p_{\\theta}(.)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": ", which leverages the extracted LDP to refine feature representations, thereby mitigating domain shifts and improving segmentation accuracy; and (iii) a Diffusion Prior Estimator (DPE), implementing " + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "inline_equation", + "content": "p_{\\phi}(.)" + }, + { + "bbox": [ + 313, + 594, + 556, + 715 + ], + "type": "text", + "content": " through probabilistic" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21758" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 77, + 219, + 153 + ], + "blocks": [ + { + "bbox": [ + 60, + 77, + 219, + 153 + ], + "lines": [ + { + "bbox": [ + 60, + 77, + 219, + 153 + ], + "spans": [ + { + "bbox": [ + 60, + 77, + 219, + 153 + ], + "type": "image", + "image_path": "93dc5a1cfb1bea9e7d80a33f89adc6a69bdf0661c46140c3c34e2c0571bdaa34.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 109, + 159, + 170, + 169 + ], + "lines": [ + { + "bbox": [ + 109, + 159, + 170, + 169 + ], + "spans": [ + { + "bbox": [ + 109, + 159, + 170, + 169 + ], + "type": "text", + "content": "(a) Details of LPE" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 228, + 76, + 386, + 153 + ], + "blocks": [ + { + "bbox": [ + 228, + 76, + 386, + 153 + ], + "lines": [ + { + "bbox": [ + 228, + 76, + 386, + 153 + ], + "spans": [ + { + "bbox": [ + 228, + 76, + 386, + 153 + ], + "type": "image", + "image_path": "054ffc7bd43f082a8bf16f571b7a5c7806895e7e37b6ef7a2ef99d30ae01037b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 272, + 159, + 337, + 169 + ], + "lines": [ + { + "bbox": [ + 272, + 159, + 337, + 169 + ], + "spans": [ + { + "bbox": [ + 272, + 159, + 337, + 169 + ], + "type": "text", + "content": "(b) Details of DCM" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "lines": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "spans": [ + { + "bbox": [ + 55, + 178, + 555, + 200 + ], + "type": "text", + "content": "Figure 3. Schematic diagrams of (a) the Latent Prior Extractor, (b) the Domain Compensation Module and (c) the Diffusion Prior Estimator." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 394, + 77, + 548, + 153 + ], + "blocks": [ + { + "bbox": [ + 394, + 77, + 548, + 153 + ], + "lines": [ + { + "bbox": [ + 394, + 77, + 548, + 153 + ], + "spans": [ + { + "bbox": [ + 394, + 77, + 548, + 153 + ], + "type": "image", + "image_path": "e8a729b231389f50f222bfb3619eda4a414d5500bfd1c94813761632b2e8eb6d.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 441, + 159, + 501, + 169 + ], + "lines": [ + { + "bbox": [ + 441, + 159, + 501, + 169 + ], + "spans": [ + { + "bbox": [ + 441, + 159, + 501, + 169 + ], + "type": "text", + "content": "(c) Details of DPE" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "content": "diffusion modeling to estimate LDP without paired source and target data. Notably, PDAF is compatible with any existing segmentation model, allowing frozen pre-trained encoders " + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "inline_equation", + "content": "E_{\\vartheta}" + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "content": " and decoders " + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "inline_equation", + "content": "D_{\\vartheta}" + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "content": " to discover LDP, and then finetuning the target network " + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "inline_equation", + "content": "E_{\\theta}" + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 54, + 222, + 294, + 294 + ], + "type": "text", + "content": " for DGSS. The overall architecture of PDAF is illustrated in Figure 2." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "spans": [ + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": "Latent Prior Extractor. The LPE is designed to estimate the optimal LDP by supervising cross-domain feature relationships, providing effective guidance for feature alignment. As illustrated in Figure 3a, the LPE first concatenates source features " + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "inline_equation", + "content": "h_{\\vartheta,s}" + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": " and pseudo-target features " + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "inline_equation", + "content": "h_{\\vartheta,t'}" + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": ", then utilizes residual blocks to model cross-domain feature relationships. The extracted features are then processed through two projection layers to derive the probabilistic modeling parameters, namely the mean " + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "inline_equation", + "content": "(\\mu)" + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": " and variance " + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "inline_equation", + "content": "(\\sigma)" + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": ". To ensure stable and effective training, we constrain the optimal LDP " + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 55, + 298, + 296, + 453 + ], + "type": "text", + "content": " to follow a standard normal distribution [32], enhancing latent space regularization and enabling more robust prior estimation. The procedure can be formulated as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 106, + 466, + 294, + 508 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 106, + 466, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 106, + 466, + 294, + 508 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} h _ {\\vartheta , s} = E _ {\\vartheta} \\left(x _ {s}\\right), h _ {\\vartheta , t ^ {\\prime}} = E _ {\\vartheta} \\left(x _ {t ^ {\\prime}}\\right), \\\\ \\mu , \\sigma = F _ {\\varphi} \\left(h _ {\\vartheta , s}, h _ {\\vartheta , t ^ {\\prime}}\\right), \\tag {4} \\\\ \\tilde {z} = \\mu + \\epsilon \\sigma , \\epsilon \\sim \\mathcal {N} (0, I), \\\\ \\end{array}", + "image_path": "3ccc80aea282dfb3302302ed81d12684cd23af9df06420c71df9ca5d1b455030.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "inline_equation", + "content": "F_{\\varphi}" + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": " indicates LPE parameterized by " + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\varphi" + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": ". The dimension of LDP is " + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "inline_equation", + "content": "\\tilde{z} \\in \\mathbb{R}^{c' \\times h \\times w}" + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "inline_equation", + "content": "c'" + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": " denotes a predefined number of the channel of LDP and " + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "inline_equation", + "content": "h, w" + }, + { + "bbox": [ + 55, + 517, + 295, + 567 + ], + "type": "text", + "content": " correspond to the spatial dimensions of the feature maps." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": "Domain Compensation Module. The DCM enhances segmentation models for domain generalization. Drawing inspiration from Spatial Feature Transformation (SFT) [60], DCM leverages LDP as a compensation mechanism, providing domain-aware modulation to improve feature alignment. As illustrated in Figure 3b, a straightforward approach involves projecting LDP onto affine transformation parameters using dedicated projection layers " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "(F_{\\Psi})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": ", yielding scale " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\tilde{\\gamma})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " and shift " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\tilde{\\beta})" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " parameters. These parameters are then used to rescale and shift the feature representations, effectively normalizing domain discrepancies across spatial regions. The refined features " + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "inline_equation", + "content": "\\tilde{h}_{\\theta ,t^{\\prime}}" + }, + { + "bbox": [ + 55, + 570, + 296, + 714 + ], + "type": "text", + "content": " are subsequently pro" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "text", + "content": "cessed by the segmentation head " + }, + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "text", + "content": " to generate accurate segmentation maps " + }, + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{t'}" + }, + { + "bbox": [ + 313, + 223, + 554, + 247 + ], + "type": "text", + "content": ". The process can be formulated as:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 389, + 258, + 450, + 270 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 258, + 450, + 270 + ], + "spans": [ + { + "bbox": [ + 389, + 258, + 450, + 270 + ], + "type": "interline_equation", + "content": "\\tilde {\\gamma}, \\tilde {\\beta} = F _ {\\Psi} (\\tilde {z}),", + "image_path": "756f86e395948c64c3bab0569b4f03afcdb948e5b2e14af6b99fbd73ad88032e.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 388, + 274, + 553, + 289 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 274, + 553, + 289 + ], + "spans": [ + { + "bbox": [ + 388, + 274, + 553, + 289 + ], + "type": "interline_equation", + "content": "\\tilde {h} _ {\\theta , t ^ {\\prime}} = \\tilde {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\tilde {\\beta}, \\tag {5}", + "image_path": "04fe96219955a4fdd8b150751a94338be9e9fa6a57cdca1ad58a3bdd5bf264f0.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 395, + 292, + 465, + 305 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 395, + 292, + 465, + 305 + ], + "spans": [ + { + "bbox": [ + 395, + 292, + 465, + 305 + ], + "type": "interline_equation", + "content": "\\tilde {y} _ {t ^ {\\prime}} = D _ {\\theta} \\left(\\tilde {h} _ {\\vartheta , t ^ {\\prime}}\\right),", + "image_path": "69a182029bf02ceba9b04af3f43ef9c99153043b282b002285f0022e9a4dba62.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\oplus" + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "text", + "content": " denote element-wise multiplication and element-wise addition, respectively. The dimensions of the affine transformation parameters correspond to the feature map, with " + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "inline_equation", + "content": "\\tilde{\\gamma},\\tilde{\\beta}\\in \\mathbb{R}^{c\\times h\\times w}" + }, + { + "bbox": [ + 313, + 316, + 554, + 365 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "text", + "content": "Diffusion Prior Estimator. Diffusion Models (DMs) have demonstrated their effectiveness in sampling from complex target domains, significantly improving various computer vision tasks as evidenced in previous studies [38, 50, 53]. In DPE, we leverage probabilistic diffusion modeling to estimate the LDP, enabling arbitrary target domain and accurate prior estimation for domain alignment. As illustrated in Figure 3c, we first apply the forward diffusion process to an optimal LDP " + }, + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "text", + "content": ", iteratively adding Gaussian noise to obtain an intermediate latent state " + }, + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "inline_equation", + "content": "z_{T}" + }, + { + "bbox": [ + 313, + 368, + 555, + 487 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 355, + 498, + 553, + 512 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 355, + 498, + 553, + 512 + ], + "spans": [ + { + "bbox": [ + 355, + 498, + 553, + 512 + ], + "type": "interline_equation", + "content": "q \\left(z _ {T} | \\tilde {z}\\right) = \\mathcal {N} \\left(z _ {T}; \\sqrt {\\bar {\\alpha} _ {T}} \\tilde {z}, (1 - \\bar {\\alpha} _ {T}) I\\right), \\tag {6}", + "image_path": "6446aae5567bc4666a06311f35942d5634c21e6d75ab37fc05d02e57573be786.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\alpha_{T} = 1 - \\beta_{T}" + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_{T} = \\prod_{i=1}^{T} \\alpha_{i}" + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "content": ". For the reverse diffusion process, the DPE initializes from " + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\hat{z}_{T}" + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "content": " and performs T-step denoising process to estimate the LDP " + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "inline_equation", + "content": "\\hat{z}_{0}" + }, + { + "bbox": [ + 313, + 523, + 554, + 571 + ], + "type": "text", + "content": ", which is formulated as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 364, + 580, + 553, + 607 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 364, + 580, + 553, + 607 + ], + "spans": [ + { + "bbox": [ + 364, + 580, + 553, + 607 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {T - 1} = \\frac {1}{\\sqrt {\\alpha_ {T}}} \\left(\\hat {z} _ {T} - \\epsilon \\frac {1 - \\alpha_ {T}}{\\sqrt {1 - \\bar {\\alpha} _ {T}}}\\right). \\tag {7}", + "image_path": "bb3a676c399bd426b8775ebf369b3d169496e7b42f0d5a34dd178c9fa394a506.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": "Specifically, to estimate LDP for target images, we condition the DPE on the extracted target features " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "h_{\\theta ,t'}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ". DPE " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "(F_{\\phi})" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": " then performs a denoising operation at timestep " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ", represented as " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "F_{\\phi}(\\hat{z}_{T},T,h_{\\theta ,t'})" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ". To accommodate the reduced channel dimension of LDP, we employ accelerated diffusion-based optimization [57] for efficient and rapid inference. Moreover, this design allows for the joint training of DPE and the segmentation head " + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "inline_equation", + "content": "D_{\\theta}" + }, + { + "bbox": [ + 313, + 617, + 555, + 714 + ], + "type": "text", + "content": ", thereby optimizing" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21759" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 95 + ], + "type": "text", + "content": "LDP directly for the segmentation task. This process is defined as follows:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 135, + 102, + 226, + 116 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 135, + 102, + 226, + 116 + ], + "spans": [ + { + "bbox": [ + 135, + 102, + 226, + 116 + ], + "type": "interline_equation", + "content": "\\hat {z} _ {0} = F _ {\\phi} (\\hat {z} _ {T}, T, h _ {\\theta , t ^ {\\prime}}),", + "image_path": "e0df5b7f855aec4bbe3a10515b96aacdeff56e1cfdc74fe91f17ee7d9b11b8bf.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 127, + 118, + 295, + 140 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 118, + 295, + 140 + ], + "spans": [ + { + "bbox": [ + 127, + 118, + 295, + 140 + ], + "type": "interline_equation", + "content": "\\hat {\\gamma}, \\hat {\\beta} = F _ {\\Psi} (\\hat {z} _ {0}), \\tag {8}", + "image_path": "50b4f419972123e3877c73c25c8e5a9d35be69ba14992582c2b3c8ac12611164.jpg" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 126, + 135, + 217, + 150 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 126, + 135, + 217, + 150 + ], + "spans": [ + { + "bbox": [ + 126, + 135, + 217, + 150 + ], + "type": "interline_equation", + "content": "\\hat {h} _ {\\theta , t ^ {\\prime}} = \\hat {\\gamma} \\odot h _ {\\theta , t ^ {\\prime}} \\oplus \\hat {\\beta},", + "image_path": "c9b23620223d777a156ac2ba715acb92bcf4848dbdebdb90142c7920881e38b0.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 132, + 153, + 202, + 167 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 153, + 202, + 167 + ], + "spans": [ + { + "bbox": [ + 132, + 153, + 202, + 167 + ], + "type": "interline_equation", + "content": "\\hat {y} _ {t ^ {\\prime}} = D _ {\\theta} (\\hat {h} _ {\\vartheta , t ^ {\\prime}}),", + "image_path": "8f2b43b34c8f15e90823449dfa92ab93b339bbec5df2f5006284c4da179decd3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t'}" + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": " represents the predicted segmentation maps conditions on the estimated LDP " + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": ". The dimensions of the affine transformation parameters correspond to the feature map, with " + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "inline_equation", + "content": "\\hat{\\gamma}, \\hat{\\beta} \\in \\mathbb{R}^{c \\times h \\times w}" + }, + { + "bbox": [ + 55, + 175, + 296, + 223 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "spans": [ + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": "Inference. During inference, only target domain images " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "x_{t}" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": " are available, so the variational posterior is omitted. Instead, we extract target features " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "h_{\\theta ,t}" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": " using the segmentation backbone " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "E_{\\theta}(\\cdot)" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": " as the conditioning input and sample a Gaussian noise vector " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\hat{z}_T" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": ". The DPE then iteratively refines this noise over " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": " denoising steps to produce the estimated LDP " + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 226, + 296, + 334 + ], + "type": "text", + "content": ". Finally, the DCM leverages the estimated LDP to refine the feature representations, which are used for robust segmentation prediction." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 341, + 173, + 354 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 341, + 173, + 354 + ], + "spans": [ + { + "bbox": [ + 55, + 341, + 173, + 354 + ], + "type": "text", + "content": "3.3. Model Optimization" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "text", + "content": "We train our framework by optimizing the likelihood estimation and enforcing posterior regularization to enhance domain generalization, as defined in Eq. (3). To achieve maximum likelihood estimation, we employ task-specific loss functions " + }, + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{task}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "text", + "content": " tailored to different segmentation models. Specifically, we use weighted cross-entropy loss for DeepLabV3+ [7] and focal loss for Mask2Former [10]. To further enhance feature consistency, we introduce a semantic consistency loss " + }, + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sc}}" + }, + { + "bbox": [ + 55, + 358, + 296, + 502 + ], + "type": "text", + "content": " that measures the discrepancy between the predictions of the source and pseudo-target images. This loss enforces semantic alignment in feature space and accelerates model convergence:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 87, + 510, + 295, + 525 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 87, + 510, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 87, + 510, + 295, + 525 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {s c}} = \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\tilde {y} _ {t ^ {\\prime}} \\right\\| _ {2} + \\left\\| f _ {\\vartheta} \\left(x _ {s}\\right) - \\hat {y} _ {t ^ {\\prime}} \\right\\| _ {2}, \\tag {9}", + "image_path": "5d78f4f6a9a95784ab0a65fd5b9aba218bb42bb0c4193d1c72c4618624eefa46.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "f_{\\vartheta}" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": " denotes the frozen segmentation network " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "E_{\\vartheta}(.)" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "D_{\\vartheta}(.)" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\tilde{y}_{t'}" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\hat{y}_{t'}" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": " represent the predictions from Eq. (5) and Eq. (8), respectively. To minimize the KL divergence and enforce posterior regularization, we adopt a prior constraint loss " + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prior}}" + }, + { + "bbox": [ + 55, + 532, + 296, + 616 + ], + "type": "text", + "content": " to align the outputs of LPE and DPE, thereby implicitly reducing the distributional discrepancy between the posterior and the prior:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 134, + 625, + 295, + 639 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 134, + 625, + 295, + 639 + ], + "spans": [ + { + "bbox": [ + 134, + 625, + 295, + 639 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {p r i o r}} = \\| \\hat {z} _ {0} - \\tilde {z} \\| _ {2}. \\tag {10}", + "image_path": "5d4c695c779fc2e72a0d19fe9a78e822dbfc45812daec5cf4d33067740d1bf3f.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "inline_equation", + "content": "\\hat{z}_0" + }, + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "inline_equation", + "content": "\\tilde{z}" + }, + { + "bbox": [ + 55, + 647, + 296, + 671 + ], + "type": "text", + "content": " indicate the optimal LDP and estimated LDP from Eq. (4) and Eq. (8), respectively. The overall loss is:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 93, + 680, + 295, + 693 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 680, + 295, + 693 + ], + "spans": [ + { + "bbox": [ + 93, + 680, + 295, + 693 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\lambda_ {\\text {t a s k}} \\mathcal {L} _ {\\text {t a s k}} + \\lambda_ {\\mathrm {s c}} \\mathcal {L} _ {\\mathrm {s c}} + \\lambda_ {\\text {p r i o r}} \\mathcal {L} _ {\\text {p r i o r}}, \\tag {11}", + "image_path": "eb3c7534f7ef9b53012df107d96e8b4d0d2a49a250c3e5320e52a80f22dd8712.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{task}}" + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{sc}}" + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{prior}}" + }, + { + "bbox": [ + 55, + 701, + 282, + 714 + ], + "type": "text", + "content": " represents the scaling factors." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 71, + 450, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 71, + 450, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 71, + 450, + 84 + ], + "type": "text", + "content": "4. Implementation Details" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "spans": [ + { + "bbox": [ + 314, + 91, + 377, + 102 + ], + "type": "text", + "content": "4.1. Datasets" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "spans": [ + { + "bbox": [ + 313, + 108, + 553, + 132 + ], + "type": "text", + "content": "We conduct experiments on five urban scene semantic segmentation datasets that share 19 common scene categories." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 133, + 553, + 180 + ], + "type": "text", + "content": "Cityscapes (C) [12] consists of high-resolution images captured from 50 distinct cities, where we exclusively use the finely annotated subset containing 2,975 training and 500 validation images." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 553, + 215 + ], + "type": "text", + "content": "BDD-100K (B) [67] provides 7,000 training images and 1,000 validation images collected from various locations conditions." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 216, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 276 + ], + "type": "text", + "content": "Mapillary (M) [42] is a large-scale dataset containing 25,000 high-resolution street-view images sourced globally. GTA5 (G) [52] includes 24,966 images simulated using the Grand Theft Auto V game engine, divided into 12,403 training, 6,382 validation, and 6,181 testing samples." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "spans": [ + { + "bbox": [ + 313, + 276, + 553, + 300 + ], + "type": "text", + "content": "SYNTHIA (S) [54] consists of 9,400 photorealistic synthetic images, including 2,830 for testing." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 314, + 308, + 434, + 319 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 308, + 434, + 319 + ], + "spans": [ + { + "bbox": [ + 314, + 308, + 434, + 319 + ], + "type": "text", + "content": "4.2. Evaluation Protocols" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 325, + 554, + 433 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 325, + 554, + 433 + ], + "spans": [ + { + "bbox": [ + 313, + 325, + 554, + 433 + ], + "type": "text", + "content": "Following existing DGSS benchmarks [1, 11, 15, 45, 46, 49], we adopt a single-source domain generalization setting, where one dataset is used for training (source domain), and the remaining four datasets are treated as unseen target domains for evaluation. For a fair comparison, all experiments are built on DeepLabV3Plus [7] (with ResNet-50 [21]) and Mask2Former [10] (with Swin-T and Swin-L [40]). We use the mean intersection over the union (mIoU) [17] to measure the performance of segmentation." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 314, + 444, + 412, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 444, + 412, + 456 + ], + "spans": [ + { + "bbox": [ + 314, + 444, + 412, + 456 + ], + "type": "text", + "content": "4.3. Training Details" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": "To facilitate a fair evaluation, we exclusively employ Photometric Augmentation as the data augmentation strategy to generate pseudo-target domain samples. The model is trained with Adam Optimizer [31] with an initial learning rate of " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "1 \\times 10^{-5}" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ", momentum parameters " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "(\\beta_{1},\\beta_{2}) = (0.9,0.999)" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ", and a batch size of 4 for 100 epochs. The channel of LDP is set to " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "c' = 4" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ". In DPE, we adopt a total diffusion timestep of " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "T = 4" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ", with the noise variance " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\beta_{t}" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": " linearly increasing from " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.1" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "\\beta_{T} = 0.99" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ". The loss coefficients are adjusted to " + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "inline_equation", + "content": "(\\lambda_{\\mathrm{task}},\\lambda_{\\mathrm{sc}},\\lambda_{\\mathrm{prior}}) = (0.5,0.5,1.0)" + }, + { + "bbox": [ + 313, + 461, + 555, + 641 + ], + "type": "text", + "content": ", balancing the contributions of different loss terms. All experiments are implemented in PyTorch and conducted on a single NVIDIA RTX 4090 GPU. More details of the implementation and hyperparameters analysis are provided in the Supplementary Material." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 314, + 652, + 395, + 665 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 652, + 395, + 665 + ], + "spans": [ + { + "bbox": [ + 314, + 652, + 395, + 665 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 672, + 498, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 672, + 498, + 684 + ], + "spans": [ + { + "bbox": [ + 313, + 672, + 498, + 684 + ], + "type": "text", + "content": "5.1. Comparison with State-of-the-arts" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 714 + ], + "type": "text", + "content": "To access the effectiveness of PDAF, we compare our approach against a comprehensive set of state-of-the-art meth" + } + ] + } + ], + "index": 28 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21760" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 69, + 119, + 123 + ], + "blocks": [ + { + "bbox": [ + 58, + 69, + 119, + 123 + ], + "lines": [ + { + "bbox": [ + 58, + 69, + 119, + 123 + ], + "spans": [ + { + "bbox": [ + 58, + 69, + 119, + 123 + ], + "type": "image", + "image_path": "f3fc0319fb6e225f02f4cb33b4c699e2ca28d1e079e01165c94f1395e9b208c3.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 58, + 124, + 119, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 124, + 119, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 124, + 119, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 124, + 119, + 177 + ], + "type": "image", + "image_path": "7d163fe516cfbe098b12324d14ec783f727dee55adcd7326740f1c62e0d8ae9d.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 58, + 178, + 119, + 230 + ], + "blocks": [ + { + "bbox": [ + 58, + 178, + 119, + 230 + ], + "lines": [ + { + "bbox": [ + 58, + 178, + 119, + 230 + ], + "spans": [ + { + "bbox": [ + 58, + 178, + 119, + 230 + ], + "type": "image", + "image_path": "165c7e4eef114c0eb381cd64c25aa2a144536e4dfd5382ee83e7486be19aff84.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 79, + 232, + 97, + 242 + ], + "lines": [ + { + "bbox": [ + 79, + 232, + 97, + 242 + ], + "spans": [ + { + "bbox": [ + 79, + 232, + 97, + 242 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 121, + 70, + 183, + 123 + ], + "blocks": [ + { + "bbox": [ + 121, + 70, + 183, + 123 + ], + "lines": [ + { + "bbox": [ + 121, + 70, + 183, + 123 + ], + "spans": [ + { + "bbox": [ + 121, + 70, + 183, + 123 + ], + "type": "image", + "image_path": "b2660d3a756f451534a98804057300856c424c0c2667a911b881219344688ba3.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 121, + 124, + 182, + 177 + ], + "blocks": [ + { + "bbox": [ + 121, + 124, + 182, + 177 + ], + "lines": [ + { + "bbox": [ + 121, + 124, + 182, + 177 + ], + "spans": [ + { + "bbox": [ + 121, + 124, + 182, + 177 + ], + "type": "image", + "image_path": "a1e5e81f42ac5523e1a2b48932c8ced41f546414eb152a0553d8607dad422c22.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 121, + 178, + 183, + 230 + ], + "blocks": [ + { + "bbox": [ + 121, + 178, + 183, + 230 + ], + "lines": [ + { + "bbox": [ + 121, + 178, + 183, + 230 + ], + "spans": [ + { + "bbox": [ + 121, + 178, + 183, + 230 + ], + "type": "image", + "image_path": "0a3dfe1443b9652ca045eb7d75b02eef45d29391636c82cdffee6291507cf607.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 119, + 232, + 184, + 242 + ], + "lines": [ + { + "bbox": [ + 119, + 232, + 184, + 242 + ], + "spans": [ + { + "bbox": [ + 119, + 232, + 184, + 242 + ], + "type": "text", + "content": "DeepLabV3Plus [7]" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 185, + 70, + 246, + 123 + ], + "blocks": [ + { + "bbox": [ + 185, + 70, + 246, + 123 + ], + "lines": [ + { + "bbox": [ + 185, + 70, + 246, + 123 + ], + "spans": [ + { + "bbox": [ + 185, + 70, + 246, + 123 + ], + "type": "image", + "image_path": "6f8a960ec99f3b14a4f911c4e6d25b9d0645bb81677247ca7faeb42e1969e0ba.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 185, + 124, + 246, + 177 + ], + "blocks": [ + { + "bbox": [ + 185, + 124, + 246, + 177 + ], + "lines": [ + { + "bbox": [ + 185, + 124, + 246, + 177 + ], + "spans": [ + { + "bbox": [ + 185, + 124, + 246, + 177 + ], + "type": "image", + "image_path": "19e597b728a49a813bfd76748a2ff58996ff97b047c5593fa85cdf0440c88908.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 185, + 178, + 246, + 230 + ], + "blocks": [ + { + "bbox": [ + 185, + 178, + 246, + 230 + ], + "lines": [ + { + "bbox": [ + 185, + 178, + 246, + 230 + ], + "spans": [ + { + "bbox": [ + 185, + 178, + 246, + 230 + ], + "type": "image", + "image_path": "2f53b884f22fba0c6a87b5c3b9fe8c566f050df8e709256db53c93e49b4a85a7.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 200, + 232, + 231, + 242 + ], + "lines": [ + { + "bbox": [ + 200, + 232, + 231, + 242 + ], + "spans": [ + { + "bbox": [ + 200, + 232, + 231, + 242 + ], + "type": "text", + "content": "ISW [11]" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 246, + 70, + 307, + 123 + ], + "blocks": [ + { + "bbox": [ + 246, + 70, + 307, + 123 + ], + "lines": [ + { + "bbox": [ + 246, + 70, + 307, + 123 + ], + "spans": [ + { + "bbox": [ + 246, + 70, + 307, + 123 + ], + "type": "image", + "image_path": "562139c18de1ab8f1bed582fad83f1f88eb0faaff0659106a763b66f242c95d5.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 246, + 124, + 307, + 177 + ], + "blocks": [ + { + "bbox": [ + 246, + 124, + 307, + 177 + ], + "lines": [ + { + "bbox": [ + 246, + 124, + 307, + 177 + ], + "spans": [ + { + "bbox": [ + 246, + 124, + 307, + 177 + ], + "type": "image", + "image_path": "ed2d78f2b6b74e645b4b3a688bd37f10354064915aab39660226ce0405c50b6d.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 246, + 178, + 307, + 230 + ], + "blocks": [ + { + "bbox": [ + 246, + 178, + 307, + 230 + ], + "lines": [ + { + "bbox": [ + 246, + 178, + 307, + 230 + ], + "spans": [ + { + "bbox": [ + 246, + 178, + 307, + 230 + ], + "type": "image", + "image_path": "7106d2d10e74c9140b485a7d656e585a5dea84283c55394f8acc70d648ceaa62.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 255, + 232, + 299, + 242 + ], + "lines": [ + { + "bbox": [ + 255, + 232, + 299, + 242 + ], + "spans": [ + { + "bbox": [ + 255, + 232, + 299, + 242 + ], + "type": "text", + "content": "SHADE [75]" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 251, + 553, + 274 + ], + "lines": [ + { + "bbox": [ + 55, + 251, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 553, + 274 + ], + "type": "text", + "content": "Figure 4. Qualitative comparison between DGSS methods trained on GTAV (G) [52] using DeepLabV3Plus [7] with ResNet-50 [21]. The first, second, and third rows show the predictions for Cityscapes (C) [12], BDD-100K (B) [67], and Mapillary (M) [42], respectively." + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 308, + 70, + 369, + 123 + ], + "blocks": [ + { + "bbox": [ + 308, + 70, + 369, + 123 + ], + "lines": [ + { + "bbox": [ + 308, + 70, + 369, + 123 + ], + "spans": [ + { + "bbox": [ + 308, + 70, + 369, + 123 + ], + "type": "image", + "image_path": "9a1088d99bfe9667cec19e2e700240293e0ed35a079ad70362ecf43c14d0534c.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 308, + 124, + 369, + 177 + ], + "blocks": [ + { + "bbox": [ + 308, + 124, + 369, + 177 + ], + "lines": [ + { + "bbox": [ + 308, + 124, + 369, + 177 + ], + "spans": [ + { + "bbox": [ + 308, + 124, + 369, + 177 + ], + "type": "image", + "image_path": "327e9b3ee22ccb2177af0b192b53920e204b035251f40fcd4ec46adfdccbf248.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 308, + 178, + 369, + 230 + ], + "blocks": [ + { + "bbox": [ + 308, + 178, + 369, + 230 + ], + "lines": [ + { + "bbox": [ + 308, + 178, + 369, + 230 + ], + "spans": [ + { + "bbox": [ + 308, + 178, + 369, + 230 + ], + "type": "image", + "image_path": "4d7be68d2f9913c97b84a58107b44f1147e586a00bd76dd439c42318fcd4c0ad.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 317, + 232, + 361, + 242 + ], + "lines": [ + { + "bbox": [ + 317, + 232, + 361, + 242 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 361, + 242 + ], + "type": "text", + "content": "WildNet [34]" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 369, + 70, + 430, + 123 + ], + "blocks": [ + { + "bbox": [ + 369, + 70, + 430, + 123 + ], + "lines": [ + { + "bbox": [ + 369, + 70, + 430, + 123 + ], + "spans": [ + { + "bbox": [ + 369, + 70, + 430, + 123 + ], + "type": "image", + "image_path": "299c53624a2e292d5470ebadf3995512803b23b9e28abe14677d7f1f6a8c9d8a.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 369, + 124, + 430, + 177 + ], + "blocks": [ + { + "bbox": [ + 369, + 124, + 430, + 177 + ], + "lines": [ + { + "bbox": [ + 369, + 124, + 430, + 177 + ], + "spans": [ + { + "bbox": [ + 369, + 124, + 430, + 177 + ], + "type": "image", + "image_path": "f68322b4936bbe68b29d9fc4a6b4e5d6d3d0ee7daa6612046dc2a9d042f61cf7.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 369, + 178, + 430, + 230 + ], + "blocks": [ + { + "bbox": [ + 369, + 178, + 430, + 230 + ], + "lines": [ + { + "bbox": [ + 369, + 178, + 430, + 230 + ], + "spans": [ + { + "bbox": [ + 369, + 178, + 430, + 230 + ], + "type": "image", + "image_path": "e4dd0f0daeb0c5e63f477d5c82b73e1d9c2b5319d122d86d60a4e038888249bf.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 379, + 232, + 422, + 242 + ], + "lines": [ + { + "bbox": [ + 379, + 232, + 422, + 242 + ], + "spans": [ + { + "bbox": [ + 379, + 232, + 422, + 242 + ], + "type": "text", + "content": "BlindNet [1]" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 431, + 70, + 492, + 123 + ], + "blocks": [ + { + "bbox": [ + 431, + 70, + 492, + 123 + ], + "lines": [ + { + "bbox": [ + 431, + 70, + 492, + 123 + ], + "spans": [ + { + "bbox": [ + 431, + 70, + 492, + 123 + ], + "type": "image", + "image_path": "03ed4bbfe0227a35fb998b549d91ae0457fba1f65823d7f6bd9efa1360371e32.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 431, + 124, + 492, + 177 + ], + "blocks": [ + { + "bbox": [ + 431, + 124, + 492, + 177 + ], + "lines": [ + { + "bbox": [ + 431, + 124, + 492, + 177 + ], + "spans": [ + { + "bbox": [ + 431, + 124, + 492, + 177 + ], + "type": "image", + "image_path": "9a372b15018a9575633c2f9d738dda67deb7f5bf490d02fe9953887193a22eb5.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 431, + 178, + 492, + 230 + ], + "blocks": [ + { + "bbox": [ + 431, + 178, + 492, + 230 + ], + "lines": [ + { + "bbox": [ + 431, + 178, + 492, + 230 + ], + "spans": [ + { + "bbox": [ + 431, + 178, + 492, + 230 + ], + "type": "image", + "image_path": "01bb4ee1a8fa13ba1cfbda9a377008786cb85755bc24e5949030fec56a42f499.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 451, + 232, + 473, + 241 + ], + "lines": [ + { + "bbox": [ + 451, + 232, + 473, + 241 + ], + "spans": [ + { + "bbox": [ + 451, + 232, + 473, + 241 + ], + "type": "text", + "content": "PDAF" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 493, + 70, + 553, + 123 + ], + "blocks": [ + { + "bbox": [ + 493, + 70, + 553, + 123 + ], + "lines": [ + { + "bbox": [ + 493, + 70, + 553, + 123 + ], + "spans": [ + { + "bbox": [ + 493, + 70, + 553, + 123 + ], + "type": "image", + "image_path": "438216f6227b4b0cf09247f11276d9d72b594effa57324bc9c41f387057cd51c.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 493, + 124, + 553, + 177 + ], + "blocks": [ + { + "bbox": [ + 493, + 124, + 553, + 177 + ], + "lines": [ + { + "bbox": [ + 493, + 124, + 553, + 177 + ], + "spans": [ + { + "bbox": [ + 493, + 124, + 553, + 177 + ], + "type": "image", + "image_path": "7bf49b21301531a8bdf53d7671b5637f06f4d2005f9f35a92dba010e1f0e1973.jpg" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_body" + } + ], + "index": 29 + }, + { + "type": "image", + "bbox": [ + 493, + 178, + 553, + 230 + ], + "blocks": [ + { + "bbox": [ + 493, + 178, + 553, + 230 + ], + "lines": [ + { + "bbox": [ + 493, + 178, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 493, + 178, + 553, + 230 + ], + "type": "image", + "image_path": "2810a97e5384679a90d4e0594ffefec88c46f4d64cf16a503cc88a5716f8fba6.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 500, + 232, + 545, + 241 + ], + "lines": [ + { + "bbox": [ + 500, + 232, + 545, + 241 + ], + "spans": [ + { + "bbox": [ + 500, + 232, + 545, + 241 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "type": "table", + "bbox": [ + 67, + 293, + 286, + 490 + ], + "blocks": [ + { + "bbox": [ + 67, + 293, + 286, + 490 + ], + "lines": [ + { + "bbox": [ + 67, + 293, + 286, + 490 + ], + "spans": [ + { + "bbox": [ + 67, + 293, + 286, + 490 + ], + "type": "table", + "html": "
MethodBackboneTrained on Cityscapes (C)
BMGSAvg.
DeepLabV3Plus [7]Res5044.9651.6842.5523.2940.62
IBN [45]48.5657.0445.0626.1444.20
IW [46]48.4955.8244.8726.1043.82
Itemorm [25]49.2356.2645.7325.9844.30
DRPC [68]49.8656.3445.6226.5844.60
ISW [11]50.7358.6445.0026.2045.14
GTR [47]50.7557.1645.7926.4745.04
DIRL [65]51.80-46.5226.50-
SHADE [75]50.9560.6748.6127.6246.96
SAW [48]52.9559.8147.2828.3247.09
WildNet [34]50.9458.7947.0127.9546.17
DPCL [66]50.9758.5946.0025.8545.34
BlindNet [1]51.8460.1847.9728.5147.13
PDAF53.5062.9350.5430.6849.41
Mask2Former [10]51.3065.3050.6034.0050.30
HGFformer [15]Swin-T53.4066.9051.3033.6051.30
PDAF56.4066.4055.2034.8053.20
Mask2Former [10]60.1072.2057.8042.4058.13
HGFformer [15]Swin-L61.5072.1059.4041.3058.58
CMFormer [5]62.6073.6060.7043.0059.98
PDAF63.0074.1063.2044.0061.08
", + "image_path": "56f45b6371d1e7f25f35323918153a50ccb862ad30e3255afac7831b667815f3.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "table_body" + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 554, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 554, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 55, + 554, + 295, + 614 + ], + "type": "text", + "content": "ods including both CNN-based techniques [1, 7, 11, 25, 26, 34, 45-48, 65, 66, 68, 75, 76] and Mask2Former-based approaches [5, 10, 15]. We adhere to standard DGSS evaluation protocols and report results directly cited from prior works [1, 5, 11, 46, 48]." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 617, + 295, + 713 + ], + "type": "text", + "content": "Cityscapes Source Domain. Table 1 shows that PDAF achieves the best overall performance across four unseen datasets, demonstrating strong generalization to diverse domains. PDAF effectively transfers from real-world datasets (C) [12] to target domains with distinct visual distributions, such as game-rendered synthetic scenes (G) [52] and photo-realistic simulations (S) [54]. Notably, PDAF consistently improves performance across CNN-based and" + } + ] + } + ], + "index": 36 + }, + { + "type": "table", + "bbox": [ + 328, + 293, + 542, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 498, + 295, + 531 + ], + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 531 + ], + "type": "text", + "content": "Table 1. Comparison with existing methods trained on Cityscapes (C) [12]. The best and second-best mIoU scores are blue and magenta." + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 328, + 293, + 542, + 475 + ], + "lines": [ + { + "bbox": [ + 328, + 293, + 542, + 475 + ], + "spans": [ + { + "bbox": [ + 328, + 293, + 542, + 475 + ], + "type": "table", + "html": "
MethodTrained on GTAV (G)
CBMSAvg.
DeepLabV3Plus [7]28.9525.1428.1826.2327.13
IBN [45]33.8532.3037.7527.9032.95
IW [46]29.9127.4829.7127.6128.68
Iternorm [25]31.8132.7033.8827.0731.37
DRPC [68]37.4232.1434.1228.0632.94
ISW [11]36.5835.2040.3328.3035.10
GTR [47]37.5333.7534.5228.1733.49
DIRL [65]41.0439.1541.60--
SHADE [75]44.6539.2843.34--
SAW [48]39.7537.3441.8630.7937.44
WildNet [34]44.6238.4246.0931.3440.12
AdvStyle [76]39.6235.5437.00--
SPC [26]44.1040.4645.51--
DPCL [66]44.7440.5946.3330.8140.62
BlindNet [1]45.7241.3247.0831.3941.38
PDAF45.4043.5649.4532.9842.85
", + "image_path": "508dd63bfe76a9f96acf8ba00d2328afef5acb26975ad2e906aaa0557c24e4ba.jpg" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "table_body" + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "lines": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "spans": [ + { + "bbox": [ + 313, + 483, + 555, + 506 + ], + "type": "text", + "content": "Table 2. Comparison with existing DGSS methods using a CNN-based network trained on GTAV (G) [52]." + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "spans": [ + { + "bbox": [ + 313, + 528, + 553, + 552 + ], + "type": "text", + "content": "Mask2Former-based methods, highlighting its flexibility in enhancing existing segmentation networks." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 554, + 555, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 554, + 555, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 554, + 555, + 685 + ], + "type": "text", + "content": "GTAV Source Domain. Table 2 demonstrates the effectiveness of PDAF in bridging the domain gap between synthetic and real-world datasets. Unlike existing DGSS methods that rely primarily on feature normalization, PDAF explicitly models latent domain priors, providing structured guidance for feature alignment and enhancing domain generalization. As shown in Figure 4, qualitative comparisons indicate that PDAF effectively preserves detailed structural information—such as road boundaries and object contours—resulting in more precise and consistent segmentation outcomes." + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Clear to Corruptions. We evaluate PDAF on ACDC [55], which introduces four real-world adverse conditions (rain," + } + ] + } + ], + "index": 41 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21761" + } + ] + } + ], + "index": 42 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 65, + 70, + 288, + 189 + ], + "blocks": [ + { + "bbox": [ + 65, + 70, + 288, + 189 + ], + "lines": [ + { + "bbox": [ + 65, + 70, + 288, + 189 + ], + "spans": [ + { + "bbox": [ + 65, + 70, + 288, + 189 + ], + "type": "table", + "html": "
MethodBackboneTrained on Cityscape (C)
FoggyNightRainSnowAvg.
IBN [45]63.8021.2050.4049.6043.70
Itemorm [25]63.3023.8050.1049.9045.30
IW [46]Res5062.4021.8052.4047.6046.60
ISW [11]64.3024.3056.0049.8048.10
PDAF68.2830.1459.2353.5852.81
Mask2Former [10]56.4039.1058.9058.2053.15
HGFormer [15]Swin-T58.5043.3062.0058.3055.53
PDAF73.7547.7568.1262.2562.97
Mask2Former [10]69.1053.1068.3065.2063.93
HGFormer [15]Swin-L69.9052.7072.0068.6065.80
PDAF80.7255.1273.1371.4370.10
", + "image_path": "4e8b5728f79b9cde1cc64dd29b8e56124570eee358a353a3d89498f8f55e9829.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_body" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 86, + 230, + 266, + 288 + ], + "blocks": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "lines": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "spans": [ + { + "bbox": [ + 55, + 197, + 295, + 218 + ], + "type": "text", + "content": "Table 3. Comparison with existing methods evaluated on ACDC [55]." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 230, + 266, + 288 + ], + "lines": [ + { + "bbox": [ + 86, + 230, + 266, + 288 + ], + "spans": [ + { + "bbox": [ + 86, + 230, + 266, + 288 + ], + "type": "table", + "html": "
MethodsBMGSAvg.
Baseline44.9651.6842.5523.2940.62
PDAF w/o LPE49.5957.6346.3427.1945.19
PDAF w/o DCM51.5160.7949.4229.5547.82
PDAF w/o DPE51.9860.3349.3828.8647.64
PDAF53.5062.9350.5430.6849.41
", + "image_path": "aa80e0164b820852f7369f3711a772f2cf38e3c1267e808954b59b3ff47d006b.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_body" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 83, + 319, + 269, + 369 + ], + "blocks": [ + { + "bbox": [ + 74, + 297, + 276, + 308 + ], + "lines": [ + { + "bbox": [ + 74, + 297, + 276, + 308 + ], + "spans": [ + { + "bbox": [ + 74, + 297, + 276, + 308 + ], + "type": "text", + "content": "Table 4. Effectiveness of proposed modules in PDAF." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 319, + 269, + 369 + ], + "lines": [ + { + "bbox": [ + 83, + 319, + 269, + 369 + ], + "spans": [ + { + "bbox": [ + 83, + 319, + 269, + 369 + ], + "type": "table", + "html": "
\\( \\mathcal{L}_{\\text{task}} \\)\\( \\mathcal{L}_{\\text{sc}} \\)\\( \\mathcal{L}_{\\text{prior}} \\)BMGSAvg.
✓--49.7658.7347.3227.5145.83
✓✓-51.2859.5848.3827.9446.80
✓-✓53.3662.1149.6930.1048.82
✓✓✓53.5062.9350.5430.6849.41
", + "image_path": "1111b4f5692d14c77f47be2cac94e84076e8713e44461386868d09b63b03dbde.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 410, + 296, + 492 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 410, + 296, + 492 + ], + "spans": [ + { + "bbox": [ + 55, + 410, + 296, + 492 + ], + "type": "text", + "content": "fog, night and snow). We employ DeepLabV3Plus [7] and Mask2Former [10] as the segmentation network and train our method on Cityscapes [12]. As shown in Table 3, PDAF achieves consistent improvement under all adversarial conditions, demonstrating that LDP effectively compensates for degraded features caused by image corruptions and enhances robustness in varied environments." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 501, + 156, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 501, + 156, + 513 + ], + "spans": [ + { + "bbox": [ + 55, + 501, + 156, + 513 + ], + "type": "text", + "content": "5.2. Ablation Studies" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 519, + 295, + 602 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 519, + 295, + 602 + ], + "spans": [ + { + "bbox": [ + 55, + 519, + 295, + 602 + ], + "type": "text", + "content": "To validate the effectiveness of the proposed modules, we perform an ablation study. All experiments are conducted using DeepLabV3Plus [7] with ResNet-50 [21] as the segmentation network. The model is trained on Cityscapes (C) [12] and evaluated on BDD-100K (B) [67], Mapillary (M) [42], GTAV (G) [52], and SYNTHIA (S) [54] to assess generalization across diverse domain shifts." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 295, + 714 + ], + "type": "text", + "content": "Effectiveness of Proposed Modules. We consider several experimental configurations: (i) Baseline: vanilla DeepLabV3Plus [7] trained without PDAF; (ii) PDAF w/o LPE: where variational posterior is removed, resulting in LDP estimation without cross-domain guidance; (iii) PDAF w/o DCM: where LDP is directly added to the original features without employing DCM; (iv) PDAF w/o DPE: where DPE is replaced by a single-source LPE, omitting the diffusion process; and (v) PDAF: which integrates all proposed" + } + ] + } + ], + "index": 9 + }, + { + "type": "table", + "bbox": [ + 330, + 71, + 539, + 96 + ], + "blocks": [ + { + "bbox": [ + 82, + 377, + 268, + 388 + ], + "lines": [ + { + "bbox": [ + 82, + 377, + 268, + 388 + ], + "spans": [ + { + "bbox": [ + 82, + 377, + 268, + 388 + ], + "type": "text", + "content": "Table 5. Analysis on the proposed loss functions." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 330, + 71, + 539, + 96 + ], + "lines": [ + { + "bbox": [ + 330, + 71, + 539, + 96 + ], + "spans": [ + { + "bbox": [ + 330, + 71, + 539, + 96 + ], + "type": "table", + "html": "
Number of c'1416642561024
Avg. of mIoU48.7449.4148.6648.3546.1445.69
", + "image_path": "729addfbddbdff75fa2083a9320fad2169336eed4c5336d9e34d12d4092fa4d5.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "type": "table", + "bbox": [ + 325, + 127, + 543, + 152 + ], + "blocks": [ + { + "bbox": [ + 319, + 104, + 549, + 115 + ], + "lines": [ + { + "bbox": [ + 319, + 104, + 549, + 115 + ], + "spans": [ + { + "bbox": [ + 319, + 104, + 549, + 115 + ], + "type": "text", + "content": "Table 6. Ablation study on the number of channels for LDP." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 325, + 127, + 543, + 152 + ], + "lines": [ + { + "bbox": [ + 325, + 127, + 543, + 152 + ], + "spans": [ + { + "bbox": [ + 325, + 127, + 543, + 152 + ], + "type": "table", + "html": "
Number of Steps12481632
Avg. of mIoU48.8849.1249.4149.4349.4649.48
", + "image_path": "61fd6feec8e15941998b6bcc414e90cbf160fa22435a2ef597b9edfd3c6af50e.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "table_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 161, + 547, + 172 + ], + "lines": [ + { + "bbox": [ + 321, + 161, + 547, + 172 + ], + "spans": [ + { + "bbox": [ + 321, + 161, + 547, + 172 + ], + "type": "text", + "content": "Table 7. Ablation study on the number of timestep in DPE." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 313, + 194, + 553, + 230 + ], + "type": "text", + "content": "modules. As shown in Table 4, each module enhances segmentation performance, with LPE playing a pivotal role in posterior regularization for structured LDP modeling." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "spans": [ + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "text", + "content": "Analysis of Loss Functions. Table 5 demonstrates that each loss term contributes to improved performance. The prior constraint loss, " + }, + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{prior}}" + }, + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "text", + "content": ", enforces consistency between the LDPs estimated by LPE and DPE, enabling DPE to accurately infer LDP solely from target samples. Additionally, the semantic consistency loss, " + }, + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{sc}}" + }, + { + "bbox": [ + 313, + 232, + 554, + 340 + ], + "type": "text", + "content": ", enhances feature alignment by constraining predictions from both source and pseudo-target samples, further bolstering domain generalization." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 343, + 554, + 440 + ], + "type": "text", + "content": "Impact of LDP Channel Dimensions. LDP serve as a variational prior for modeling domain shifts while guiding feature alignment in DGSS. The channel dimension of LDP is crucial for capturing domain-specific variations. As shown in Table 6, a channel dimension of 4 yields the best overall performance. In contrast, higher channel dimensions introduce feature redundancy and increase the risk of overfitting, which limits further performance improvements." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 442, + 554, + 514 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 442, + 554, + 514 + ], + "spans": [ + { + "bbox": [ + 313, + 442, + 554, + 514 + ], + "type": "text", + "content": "Impact of steps for DPE. DPE iteratively refines LDP through the denoising process, facilitating precise alignment across diverse target domains. As shown in Table 7, performance gradually improves with an increasing number of diffusion steps. To achieve an optimal trade-off between efficiency and accuracy, we set the diffusion timestep to 4." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "spans": [ + { + "bbox": [ + 313, + 525, + 388, + 538 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 554, + 713 + ], + "type": "text", + "content": "This paper introduces PDAF, a probabilistic diffusion alignment framework designed for DGSS. By explicitly modeling latent domain prior (LDP), PDAF captures domain-specific variations to facilitate robust feature alignment across unseen target domains. To achieve this, PDAF involves a latent prior extractor to infer cross-domain relationships, a domain compensation module to condition segmentation features on LDP and a diffusion prior estimator to estimate LDP through a denoising process, allowing robust generalization to unseen domains. Extensive experiments demonstrate the effectiveness of PDAF, improving the generalization of existing segmentation models across diverse environments, highlighting the robustness of LDP modeling in domain generalized semantic segmentation." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21762" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 163 + ], + "type": "text", + "content": "We thank to National Center for High-performance Computing (NCHC) for providing computational and storage resources. This research was supported by Taiwan's National Science and Technology Council under Grant NSTC 111-2221-E-002-136-MY3 and NSTC 114-2221-E-002-067-MY3." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 115, + 187 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 194, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "spans": [ + { + "bbox": [ + 61, + 194, + 296, + 239 + ], + "type": "text", + "content": "[1] Woojin Ahn, Geun Yeong Yang, Hyun Duck Choi, and Myo Taeg Lim. Style blind domain generalized semantic segmentation via covariance alignment and semantic consistence contrastive learning. In CVPR, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 239, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 239, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 239, + 296, + 273 + ], + "type": "text", + "content": "[2] Yogesh Balaji, Swami Sankaranarayanan, and Rama Chellappa. Metareg: Towards domain generalization using metaregularization. In NeurIPS, 2018. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 319 + ], + "type": "text", + "content": "[3] Florent Bartoccioni, Éloi Zablocki, Andrei Bursuc, Patrick Pérez, Matthieu Cord, and Karteek Alahari. Lara: Latents and rays for multi-camera bird's-eye-view semantic segmentation. In Conference on robot learning, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "spans": [ + { + "bbox": [ + 62, + 319, + 296, + 363 + ], + "type": "text", + "content": "[4] Yasser Benigmim, Subhankar Roy, Slim Essid, Vicky Kalogeiton, and Stephane Lathuilière. Collaborating foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 364, + 296, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 364, + 296, + 397 + ], + "spans": [ + { + "bbox": [ + 62, + 364, + 296, + 397 + ], + "type": "text", + "content": "[5] Qi Bi, Shaodi You, and Theo Gevers. Learning content-enhanced mask transformer for domain generalized urban-scene segmentation. In AAAI, 2024. 3, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "spans": [ + { + "bbox": [ + 62, + 399, + 296, + 442 + ], + "type": "text", + "content": "[6] I Chen, Wei-Ting Chen, Yu-Wei Liu, Yuan-Chun Chiang, Sy-Yen Kuo, Ming-Hsuan Yang, et al. Unirestore: Unified perceptual and task-oriented image restoration model using diffusion prior. In CVPR, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "spans": [ + { + "bbox": [ + 62, + 444, + 296, + 487 + ], + "type": "text", + "content": "[7] Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "spans": [ + { + "bbox": [ + 62, + 488, + 296, + 542 + ], + "type": "text", + "content": "[8] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Hua-En Chang, Jian-Jiun Ding, and Sy-Yen Kuo. Rvsl: Robust vehicle similarity learning in real hazy scenes based on semi-supervised learning. In ECCV, 2022. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 62, + 544, + 296, + 588 + ], + "type": "text", + "content": "[9] Wei-Ting Chen, I-Hsiang Chen, Chih-Yuan Yeh, Hao-Hsiang Yang, Jian-Jiun Ding, and Sy-Yen Kuo. Sjdl-vehicle: Semi-supervised joint defogging learning for foggy vehicle re-identification. In AAAI, 2022. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 590, + 295, + 634 + ], + "type": "text", + "content": "[10] Bowen Cheng, Ishan Misra, Alexander G. Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In CVPR, 2021. 1, 2, 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 679 + ], + "type": "text", + "content": "[11] Sungha Choi, Sanghun Jung, Huiwon Yun, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via instance selective whitening. In CVPR, 2021. 3, 6, 7, 8" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 681, + 295, + 714 + ], + "type": "text", + "content": "[12] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 34, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "dataset for semantic urban scene understanding. In CVPR, 2016. 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 128 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, 2009. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 150 + ], + "type": "text", + "content": "[14] Prafulla Dhariwal and Alex Nichol. Diffusion models beat gans on image synthesis. In NeurIPS, 2021. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 553, + 194 + ], + "type": "text", + "content": "[15] Jian Ding, Nan Xue, Guisong Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In CVPR, 2023. 3, 6, 7, 8" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 227 + ], + "type": "text", + "content": "[16] Qi Dou, Daniel Coelho de Castro, Konstantinos Kamnitsas, and Ben Glocker. Domain generalization via model-agnostic learning of semantic features. In NeurIPS, 2019. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 228, + 553, + 271 + ], + "type": "text", + "content": "[17] Mark Everingham, S. M. Ali Eslami, Luc Van Gool, Christopher K. I. Williams, John M. Winn, and Andrew Zisserman. The pascal visual object classes challenge: A retrospective. *IJCV*, 2014. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 272, + 553, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 272, + 553, + 294 + ], + "spans": [ + { + "bbox": [ + 316, + 272, + 553, + 294 + ], + "type": "text", + "content": "[18] Yaroslav Ganin and Victor Lempitsky. Unsupervised domain adaptation by backpropagation. In ICML, 2015. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 294, + 553, + 327 + ], + "type": "text", + "content": "[19] Muhammad Ghifary, W. Kleijn, Mengjie Zhang, and David Balduzzi. Domain generalization for object recognition with multi-task autoencoders. In ICCV, 2015. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 370 + ], + "type": "text", + "content": "[20] Rui Gong, Martin Danelljan, Han Sun, Julio Delgado Mangas, and Luc Van Gool. Prompting diffusion representations for cross-domain semantic segmentation. arXiv preprint arXiv:2307.02138, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 403 + ], + "type": "text", + "content": "[21] Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, 2015. 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 404, + 553, + 426 + ], + "type": "text", + "content": "[22] Jonathan Ho, Ajay Jain, and P. Abbeel. Denoising diffusion probabilistic models. In NeurIPS, 2020. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "spans": [ + { + "bbox": [ + 316, + 427, + 553, + 470 + ], + "type": "text", + "content": "[23] Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In CVPR, 2023. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "spans": [ + { + "bbox": [ + 316, + 471, + 553, + 504 + ], + "type": "text", + "content": "[24] Jiaxing Huang, Dayan Guan, Aoran Xiao, and Shijian Lu. Fsdr: Frequency space domain randomization for domain generalization. In CVPR, 2021. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "spans": [ + { + "bbox": [ + 316, + 504, + 553, + 536 + ], + "type": "text", + "content": "[25] Lei Huang, Yi Zhou, Fan Zhu, Li Liu, and Ling Shao. Iterative normalization: Beyond standardization towards efficient whitening. In CVPR, 2019. 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "spans": [ + { + "bbox": [ + 316, + 537, + 553, + 581 + ], + "type": "text", + "content": "[26] Wei Huang, Chang Wen Chen, Yong Li, Jiacheng Li, Cheng Li, Fenglong Song, Youliang Yan, and Zhiwei Xiong. Style projected clustering for domain generalized semantic segmentation. In CVPR, 2023. 2, 3, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 582, + 553, + 613 + ], + "type": "text", + "content": "[27] Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. In ICML, 2015. 2" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 657 + ], + "type": "text", + "content": "[28] Yuxiang Ji, Boyong He, Chenyuan Qu, Zhuoyue Tan, Chuan Qin, and Liaoni Wu. Diffusion features to bridge domain gap for semantic segmentation. arXiv preprint arXiv:2406.00777, 2024.3" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 713 + ], + "type": "text", + "content": "[29] Yuru Jia, Lukas Hoyer, Shengyu Huang, Tianfu Wang, Luc Van Gool, Konrad Schindler, and Anton Obukhov. Dgin-style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In ECCV, 2024. 2, 3" + } + ] + } + ], + "index": 33 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21763" + } + ] + } + ], + "index": 35 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 297, + 116 + ], + "type": "text", + "content": "[30] Zeyinzi Jiang, Chaojie Mao, Yulin Pan, Zhen Han, and Jingfeng Zhang. Scedit: Efficient and controllable image diffusion generation via skip connection editing. In CVPR, 2024. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 296, + 138 + ], + "type": "text", + "content": "[31] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2014. 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 162 + ], + "type": "text", + "content": "[32] Diederik P. Kingma and Max Welling. Auto-encoding variational bayes. In ICLR, 2013. 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "spans": [ + { + "bbox": [ + 56, + 164, + 295, + 196 + ], + "type": "text", + "content": "[33] Diederik P. Kingma, Tim Salimans, Ben Poole, and Jonathan Ho. Variational diffusion models. arXiv preprint arXiv:2107.00630, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 295, + 229 + ], + "type": "text", + "content": "[34] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. Wildnet: Learning domain generalized semantic segmentation from the wild. In CVPR, 2022. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 231, + 295, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 295, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 295, + 263 + ], + "type": "text", + "content": "[35] Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M. Hospedales. Learning to generalize: Meta-learning for domain generalization. In AAAI, 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 265, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 265, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 265, + 295, + 297 + ], + "type": "text", + "content": "[36] Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex Chichung Kot. Domain generalization with adversarial feature learning. In CVPR, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 299, + 295, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 295, + 341 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 295, + 341 + ], + "type": "text", + "content": "[37] Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In ECCV, 2018. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 295, + 386 + ], + "type": "text", + "content": "[38] Qingguo Liu, Chenyi Zhuang, Pan Gao, and Jie Qin. Cdformer: when degradation prediction embraces diffusion model for blind image super-resolution. In CVPR, 2024. 3, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 295, + 430 + ], + "type": "text", + "content": "[39] Yajing Liu, Shijun Zhou, Xiyao Liu, Chunhui Hao, Baojie Fan, and Jiandong Tian. Unbiased faster r-cnn for single-source domain generalized object detection. In CVPR, 2024. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 433, + 295, + 475 + ], + "type": "text", + "content": "[40] Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In ICCV, 2021. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 478, + 295, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 478, + 295, + 510 + ], + "spans": [ + { + "bbox": [ + 56, + 478, + 295, + 510 + ], + "type": "text", + "content": "[41] Saeid Motiian, Marco Piccirilli, Donald A. Adjeroh, and Gianfranco Doretto. Unified deep supervised domain adaptation and generalization. In ICCV, 2017. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 512, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 512, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 512, + 295, + 544 + ], + "type": "text", + "content": "[42] Gerhard Neuhold, Tobias Ollmann, Samuel Rota Bulò, and Peter Kontschieder. The mapillary vistas dataset for semantic understanding of street scenes. In ICCV, 2017. 6, 7, 8" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 545, + 295, + 578 + ], + "type": "text", + "content": "[43] David Nilsson, Aleksis Pirinen, Erik Gartner, and Cristian Sminchisescu. Embodied visual active learning for semantic segmentation. In AAAI, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 295, + 622 + ], + "type": "text", + "content": "[44] Yuya Onozuka, Ryosuke Matsumi, and Motoki Shino. Autonomous mobile robot navigation independent of road boundary using driving recommendation map. In Int. Conf. on Intel. Robots and Systems, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 624, + 295, + 655 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 295, + 655 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 295, + 655 + ], + "type": "text", + "content": "[45] Xingang Pan, Ping Luo, Jianping Shi, and Xiaou Tang. Two at once: Enhancing learning and generalization capacities via ibn-net. In ECCV, 2018. 2, 6, 7, 8" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 295, + 689 + ], + "type": "text", + "content": "[46] Xingang Pan, Xiaohang Zhan, Jianping Shi, Xiaou Tang, and Ping Luo. Switchable whitening for deep representation learning. In ICCV, 2019. 3, 6, 7, 8" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 56, + 692, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 692, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 692, + 295, + 712 + ], + "type": "text", + "content": "[47] Duo Peng, Yinjie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liua. Global and local texture randomization for" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 94 + ], + "type": "text", + "content": "synthetic-to-real semantic segmentation. IEEE TIP, 2021. 2,7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 128 + ], + "type": "text", + "content": "[48] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 3, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 130, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 161 + ], + "type": "text", + "content": "[49] Duo Peng, Yinjie Lei, Munawar Hayat, Yulan Guo, and Wen Li. Semantic-aware domain generalized segmentation. In CVPR, 2022. 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 163, + 553, + 207 + ], + "type": "text", + "content": "[50] Pablo Pernias, Dominic Rampas, Mats L. Richter, Christopher J. Pal, and Marc Aubreville. Wuerstchen: An efficient architecture for large-scale text-to-image diffusion models. arXiv preprint arXiv:2306.00637v2, 2023. 3, 5" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 209, + 553, + 251 + ], + "type": "text", + "content": "[51] Sanqing Qu, Tianpei Zou, Lianghua He, Florian Röhrbein, Alois Knoll, Guang Chen, and Changjun Jiang. Lead: Learning decomposition for source-free universal domain adaptation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 553, + 285 + ], + "type": "text", + "content": "[52] Stephan R. Richter, Vibhav Vineet, Stefan Roth, and Vladlen Koltun. Playing for data: Ground truth from computer games. In ECCV, 2016. 6, 7, 8" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 287, + 553, + 319 + ], + "type": "text", + "content": "[53] Robin Rombach, A. Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, 2021. 3, 5" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 364 + ], + "type": "text", + "content": "[54] Germán Ros, Laura Sellart, Joanna Materzynska, David Vázquez, and Antonio M. López. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In CVPR, 2016. 6, 7, 8" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 366, + 553, + 398 + ], + "type": "text", + "content": "[55] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In CVPR, 2021. 7, 8" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 553, + 431 + ], + "type": "text", + "content": "[56] Seonguk Seo, Yumin Suh, Dongwan Kim, Jongwoo Han, and Bohyung Han. Learning to optimize domain specific normalization for domain generalization. In ECCV, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 434, + 553, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 464 + ], + "type": "text", + "content": "[57] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502, 2020. 5" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "spans": [ + { + "bbox": [ + 316, + 468, + 553, + 499 + ], + "type": "text", + "content": "[58] Naveen Venkat, Jogendra Nath Kundu, Durgesh Singh, Ambareesh Revanur, et al. Your classifier can secretly suffice multi-source domain adaptation. NIPS, 2020. 2" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 502, + 553, + 533 + ], + "type": "text", + "content": "[59] Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. Exploiting diffusion prior for real-world image super-resolution. IJCV, 2024. 3" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 316, + 536, + 553, + 567 + ], + "type": "text", + "content": "[60] Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. Recovering realistic texture in image super-resolution by deep spatial feature transform. In CVPR, 2018. 5" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "spans": [ + { + "bbox": [ + 316, + 569, + 553, + 611 + ], + "type": "text", + "content": "[61] Zongbin Wang, Bin Pan, Shiyu Shen, Tianyang Shi, and Zhenwei Shi. Domain generalization guided by large-scale pre-trained priors. arXiv preprint arXiv:2406.05628, 2024. 2" + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 614, + 553, + 667 + ], + "type": "text", + "content": "[62] Zhixiang Wei, Lin Chen, Yi Jin, Xiaoxiao Ma, Tianle Liu, Pengyang Ling, Ben Wang, Huaian Chen, and Jinjin Zheng. Stronger fewer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 553, + 712 + ], + "type": "text", + "content": "[63] Weijia Wu, Yuzhong Zhao, Hao Chen, Yuchao Gu, Rui Zhao, Yefei He, Hong Zhou, Mike Zheng Shou, and Chunhua Shen. Datasetdm: Synthesizing data with perception annotations using diffusion models. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 318, + 757 + ], + "type": "text", + "content": "21764" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 622 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 116 + ], + "type": "text", + "content": "[64] Bin Xia, Yulun Zhang, Shiyin Wang, Yitong Wang, Xinglong Wu, Yapeng Tian, Wenming Yang, and Luc Van Gool. Diffir: Efficient diffusion model for image restoration. In ICCV, 2023. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[65] Qi Xu, Lili Yao, Zhengkai Jiang, Guannan Jiang, Wenqing Chu, Wenhui Han, Wei Zhang, Chengjie Wang, and Ying Tai. Dirl: Domain-invariant representation learning for generalizable semantic segmentation. In AAAI, 2022. 3, 7" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 206 + ], + "type": "text", + "content": "[66] Liwei Yang, Xiang Gu, and Jian Sun. Generalized semantic segmentation by self-supervised source domain projection and multi-level contrastive learning. In AAAI, 2023. 2, 3, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "spans": [ + { + "bbox": [ + 56, + 208, + 294, + 252 + ], + "type": "text", + "content": "[67] Fisher Yu, Haofeng Chen, Xin Wang, Wenqi Xian, Yingying Chen, Fangchen Liu, Vashisht Madhavan, and Trevor Darryll. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In CVPR, 2018. 1, 6, 7, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "spans": [ + { + "bbox": [ + 56, + 253, + 294, + 306 + ], + "type": "text", + "content": "[68] Xiangyu Yue, Yang Zhang, Sicheng Zhao, Alberto L. Sangiovanni-Vincentelli, Kurt Keutzer, and Boqing Gong. Domain randomization and pyramid consistency: Simulation-to-real generalization without accessing target domain data. In ICCV, 2019. 2, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "spans": [ + { + "bbox": [ + 56, + 308, + 294, + 351 + ], + "type": "text", + "content": "[69] Haojie Zhang, Yongyi Su, Xun Xu, and Kui Jia. Improving the generalization of segmentation foundation model under distribution shift via weakly supervised adaptation. In CVPR, 2024. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[70] Juntao Zhang, Yuehuai Liu, Yu-Wing Tai, and Chi-Keung Tang. C3net: Compound conditioned control for multimodal content generation. In CVPR, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 388, + 294, + 419 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 294, + 419 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 294, + 419 + ], + "type": "text", + "content": "[71] Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "spans": [ + { + "bbox": [ + 56, + 421, + 294, + 463 + ], + "type": "text", + "content": "[72] Xin Zhang and Robby T Tan. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In CVPR, 2025. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 465, + 294, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 465, + 294, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 465, + 294, + 509 + ], + "type": "text", + "content": "[73] Dong Zhao, Jinlong Li, Shuang Wang, Mengyao Wu, Qi Zang, Nicu Sebe, and Zhun Zhong. Fishertune: Fisher-guided robust tuning of vision foundation models for domain generalized segmentation. In CVPR, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "spans": [ + { + "bbox": [ + 56, + 511, + 294, + 553 + ], + "type": "text", + "content": "[74] Shihao Zhao, Dongdong Chen, Yen-Chun Chen, Jianmin Bao, Shaozhe Hao, Lu Yuan, and Kwan-Yee K Wong. Uni-controlnet: All-in-one control to text-to-image diffusion models. NeurIPS, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 555, + 294, + 588 + ], + "type": "text", + "content": "[75] Yuyang Zhao, Zhun Zhong, Na Zhao, N. Sebe, and Gim Hee Lee. Style-hallucinated dual consistency learning for domain generalized semantic segmentation. In ECCV, 2022. 2, 3, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 294, + 622 + ], + "type": "text", + "content": "[76] Zhun Zhong, Yuyang Zhao, Gim Hee Lee, and N. Sebe. Adversarial style augmentation for domain generalized urban-scene segmentation. In NeurIPS, 2022. 2, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 293, + 748, + 317, + 757 + ], + "type": "text", + "content": "21765" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_content_list.json b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..6b7ccd090fa91f4603c0a5eccd555b8e1fd5d09b --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_content_list.json @@ -0,0 +1,1286 @@ +[ + { + "type": "text", + "text": "Exploring The Visual Feature Space for Multimodal Neural Decoding", + "text_level": 1, + "bbox": [ + 148, + 130, + 849, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Weihao Xia $\\boxtimes$ Cengiz Oztireli", + "bbox": [ + 367, + 179, + 630, + 200 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "University of Cambridge", + "bbox": [ + 397, + 205, + 594, + 224 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "https://weihaox.github.io/VINDEX", + "bbox": [ + 354, + 229, + 640, + 244 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 263, + 326, + 280 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "The intrication of brain signals drives research that leverages multimodal AI to align brain modalities with visual and textual data for explainable descriptions. However, most existing studies are limited to coarse interpretations, lacking essential details on object descriptions, locations, attributes, and their relationships. This leads to imprecise and ambiguous reconstructions when using such cues for visual decoding. To address this, we analyze different choices of vision feature spaces from pre-trained visual components within Multimodal Large Language Models (MLLMs) and introduce a zero-shot multimodal brain decoding method that interacts with these models to decode across multiple levels of granularities. To assess a model's ability to decode fine details from brain signals, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub). This benchmark includes two key tasks: detailed descriptions and salient question-answering, with metrics highlighting key visual elements like objects, attributes, and relationships. Our approach enhances neural decoding precision and supports more accurate neuro-decoding applications. Code is available at https://github.com/weihaox/VINDEX.", + "bbox": [ + 89, + 296, + 485, + 613 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 636, + 220, + 651 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Human brain signals are inherently complex and not directly interpretable, posing a significant challenge for decoding brain activity. To address this, recent studies have focused on translating brain signals associated visual stimuli into multimodal explanations, such as textual descriptions [13, 42], depth information [54], bounding boxes [53], or direct image reconstruction [35, 41, 42]. Although these efforts have made notable strides, achieving a fine-grained understanding of brain modalities remains an unresolved challenge.", + "bbox": [ + 89, + 659, + 483, + 794 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "One line of research [42, 49, 53] has attempted to translate brain activations from visual stimuli into textual captions. However, existing methods typically generate one-sentence descriptions with limited details, failing to capture the full spectrum of visual perception. The most straightforward approach [43] is to use off-the-shelf MLLMs [25, 29] to generate detailed descriptions from paired images for each", + "bbox": [ + 89, + 795, + 483, + 901 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "brain recording, then train a model to decode these pseudo-annotations from brain signals. However, this practice is computationally inefficient and may introduce bias or errors, making it impractical for generalizable brain decoding. The other key issue is the evaluation of brain-decoded detailed captions, which can be hundreds of words long, rendering conventional metrics unsuitable. Specifically, traditional captioning metrics [4, 36, 51] which mostly rely on $n$ -gram matching between candidate and reference captions, are highly sensitive to stylistic variations, leading to inconsistent evaluations. Model-based metrics either use outdated text encoders [3] or struggle with long text due to token length limitations [15, 38]. While concurrent studies [43] attempt to generate detailed captions for visual decoding, they lack reliable evaluations of caption quality and accuracy.", + "bbox": [ + 511, + 266, + 906, + 492 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In this paper, we address the two critical issues. For the method, we investigate the exploitation of various feature spaces from the visual components within MLLMs to enable fine-grained multimodal explanations from neural activations. We present VINDEX (Visual experts for multimodal Neural DEcoding eXploration), which leverages diverse feature spaces for multimodal neural decoding. We further introduce a denoiser-based alignment strategy that efficiently maps brain tokens to image tokens, allowing for integration into pre-trained MLLMs. This approach facilitates decoding across multiple levels of granularities without the need for additional textual or spatial annotations. For evaluation, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub) to discern a model's ability to decode fine-grained details from brain signals. MG-BrainDub consists of two key tasks: detailed descriptions and salient question-answering, with metrics focusing on key visual elements, including objects, attributes, and relationships, derived from both candidate and ground truth captions. Our contributions are summarized as follows:", + "bbox": [ + 511, + 493, + 908, + 794 + ], + "page_idx": 0 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- Through systematic analysis, we exploit representative feature spaces of visual components within MLLMs for brain signal alignment and multimodal brain interaction.", + "- We introduce a zero-shot multimodal brain decoding method for multi-granular decoding without the need for additional textual or spatial annotations during training.", + "- We construct a benchmark with metrics for fine-grained" + ], + "bbox": [ + 511, + 795, + 903, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "4370", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "brain interpretation, evaluating a model's ability to decode detailed descriptions and perform complex reasoning.", + "bbox": [ + 104, + 90, + 482, + 122 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 137, + 232, + 152 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Brain Perception. Recent advancements in decoding multimodal cues from brain signals, facilitated by MLLMs, have achieved unprecedented performance [35, 41, 49, 54]. These methods typically map brain responses, captured through functional magnetic resonance imaging (fMRI), to more commonly used modalities that can be processed by pre-trained models for subsequent information extraction [20, 40, 56]. For instance, Takagi and Nishimoto [49] connect fMRI data with CLIP text embeddings and the latent space of Stable Diffusion (SD) [40] using ridge regression to reconstruct visual stimuli. Xia et al. [54] extract semantic, depth, and color features from fMRI data to reconstruct images using a depth- and color-conditioned SD model. Several methods [13, 42, 49, 54] aim to create readable multimodal explanations of visual stimuli, such as textual description or depth estimation, but often rely on corresponding external supervision to guide the multimodal training. In contrast, UMBRAE [53] has recently demonstrated the ability to directly decode textual descriptions and bounding boxes from fMRI data using only image-fMRI pairs during training, without requiring additional annotations. This method aligns image features with visual encoders and integrates them into MLLMs. Building on this foundation, our proposed approach aims to further elucidate feature space representations, extracting more granular information while introducing metrics for evaluation.", + "bbox": [ + 89, + 167, + 485, + 559 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal Large Language Models. Expanding Large Language Models (LLMs) to incorporate additional modalities, such as images, has recently attracted significant interest [10, 29, 50, 57]. These models typically comprise three components: a frozen image encoder, a trainable connector, and a frozen or finetuned LLM. The connector bridges the gap from the image feature space to the LLM token space, which can be implemented as a linear layer [7], a multilayer perceptron (MLP) [29], or a lightweight transformer [17]. Recent studies have shown that using pre-trained CLIP [38] as the visual component can lead to compromised performance in visual tasks, particularly in grounding capabilities [7, 50]. To overcome this limitation, researchers have explored ways to more effectively utilize the visual component to enhance perception across diverse scenarios. This includes employing a mixture of vision encoders [19, 44, 50, 59] pretrained on various tasks as expert models and investigating the feature spaces [6, 57] within CLIP. Drawing inspiration from these efforts, we exploit brain alignment with feature spaces in MLLM's visual component for multi-granular representation to facilitate zero-shot multimodal brain decoding.", + "bbox": [ + 89, + 563, + 486, + 882 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Multimodal-Brain Alignment. The prevailing practice", + "bbox": [ + 89, + 885, + 483, + 901 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "for brain alignment is to map neural modalities into a shared multimodal space [13, 35, 41, 53], which can be categorized into generative alignment and discriminative alignment. The first category, exemplified by OneLLM [13], employs generative training to learn multimodal alignment, connecting multimodal inputs, including brain signals, with an LLM. Methods in the second category align brain signals within a pre-trained embedding space, such as CLIP [38], using techniques like linear regression [35, 49], contrastive learning [54], diffusion priors [41], or feature reconstruction [53]. MEVOX [55] ensembles multiple task-specific experts to capture distinct aspects of brain perception. The brain encoder learns to align with omni-contextual explanations. Most of these alignment strategies require per-subject training or subject-specific annotations, which can lead to scalability issues in practice. While UMBRAE [53] performs zero-shot multimodal brain decoding, its regression-based alignment is deterministic. We analyze the limitations of this strategy and propose incorporating a novel denoising optimization objective to enhance the brain-feature alignment.", + "bbox": [ + 511, + 90, + 908, + 393 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 409, + 604, + 422 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Current MLLMs [7, 29] comprise three components: a pretrained vision encoder, a pre-trained LLM, and a connector that bridges vision and language models. Given significant discrepancies in data scale and challenges associated with acquiring large-scale brain datasets, we argue that aligning brain signals with pre-trained visual components, rather than training from scratch [13, 43], is a more effective and efficient approach. This not only alleviates the need to construct extensive text for training MLLMs but also enables zero-shot capability in decoding multimodal cues, allowing for versatile usage across different MLLMs. Thus, we present the rationale behind our method design in three key aspects: (1) the selection of vision feature spaces (Sec. 3.1), (2) the alignment strategy to connect activations with pre-trained MLLMs (Sec. 3.2), and (3) the development of a benchmark necessary for fine-grained brain perception (Sec. 4).", + "bbox": [ + 511, + 435, + 908, + 678 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Exploring Feature Spaces", + "text_level": 1, + "bbox": [ + 511, + 688, + 746, + 704 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "We explore four distinct feature spaces, drawing inspiration from prior MLLM research that leverage different feature representations. By aligning brain signals across these feature spaces, we aim to capture multi-granular representations in MLLMs, facilitating zero-shot multimodal brain decoding. The four feature space choices are illustrated in Fig. 1.", + "bbox": [ + 511, + 710, + 908, + 801 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Single Encoder. This represents using features from a single pre-trained vision encoder (e.g. CLIP [38]) for brain alignment, which is the most commonly used in MMLMs.", + "bbox": [ + 511, + 806, + 906, + 851 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Mixture of Encoders. Encoders trained on diverse tasks allow us to explore the distinct advantages of different vision experts. These include contrastive image-text learning", + "bbox": [ + 511, + 854, + 908, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "4371", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/d7063f93fb0d34da34b86c9124be5332e74ddbfc04b09f97e00259521a8d18c5.jpg", + "image_caption": [ + "(a) Single Encoder" + ], + "image_footnote": [], + "bbox": [ + 143, + 123, + 250, + 169 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4fbd9e0bc2aec6b8c6909c82d5ddf698d59355c3bfb5d90a0a282a959983fb03.jpg", + "image_caption": [ + "(b) Mixture of Encoders", + "Figure 1. Feature Spaces. We aim to capture multi-granular representations for zero-shot multimodal brain decoding by aligning brain signals across different feature spaces in MLLM's visual component. The explored representative feature spaces include: (a) Single Encoder, which utilizes a single encoder to extract selected features; (b) Mixture of Encoders, using hybrid featured from different vision experts specialized in task-specific domains; (c) Aggregated Feature, fusing features from different layers within the same image encoder; and (d) Nested Features, downscaling visual representations in a hierarchical coarse-to-fine structure with different perceptual granularities." + ], + "image_footnote": [], + "bbox": [ + 277, + 95, + 413, + 186 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/91f012f67e2d4caa75382e2cadf6f3fc76e64fc5b64867878874707aa97a0dea.jpg", + "image_caption": [ + "(c) Aggregated Feature" + ], + "image_footnote": [], + "bbox": [ + 437, + 95, + 651, + 185 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/dd5c07bcccffaadb95c3bc2fc875de2399b0624dd7fb86f15bd30a290456899b.jpg", + "image_caption": [ + "(d) Nested Features" + ], + "image_footnote": [], + "bbox": [ + 681, + 121, + 861, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "CLIP [38], self-supervised learning DINO [34], and object-centric models pre-trained for various tasks such as segmentation (SAM [22]), detection (EVA [11]), and recognition (Pix2Struct [24]). These task-specific vision encoders allow MLLMs to achieve optimal performance within respective pre-training domains [19, 45]. Aligning brain signals with features from these models allows for the capture of distinct facets of visual perception. Here, we explore the integrated features from CLIP [38] and DINO [34] encoders for simplicity while demonstrating the core concept of aligning brain signals with a mixture of vision experts. CLIP effectively encodes semantic information, while DINO excels in capturing spatial details such as object location and boundaries.", + "bbox": [ + 88, + 311, + 485, + 507 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Aggregated Feature. Existing MLLMs typically use features from a selected layer in the visual component as the visual representation input for LLMs. Features from different levels—such as shallow, middle, and deep layers—are recognized for capturing different characteristics of the image [57]. Therefore, by encapsulating information from multiple layers, we can enrich the visual input for the LLM. Specifically, we partition features from selected layers into several groups. Features within each group from adjacent layers are integrated along the channel dimension to reduce redundancy and mitigate high dimensionality. The features from each group are then concatenated with the features from the final layer before being fed into the connector.", + "bbox": [ + 89, + 515, + 483, + 712 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Nested Features. MLLMs typically project feature images from the vision encoder into a fixed-length set of visual tokens [29]. Longer token sequences generally yield better results but require more resources. Recent studies have found that the optimal token length is task-specific, laying the foundation for visual token pruning and merging to dynamically adjust the sequence length [5, 21]. Therefore, we downscale selected features using a set of downsampling factors and produce nested features. This approach offers a flexible and controllable method to encode visual content as nested sets of visual tokens with a hierarchical structure, from coarse to fine details—much like a Matryoshka doll [6, 23]. The", + "bbox": [ + 89, + 719, + 485, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "nested features capture structured information across multiple coarse-to-fine levels, enabling adaptable token lengths that align with different perceptual granularities.", + "bbox": [ + 511, + 311, + 906, + 357 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Denoising for Multimodal Brain Alignment", + "text_level": 1, + "bbox": [ + 511, + 366, + 880, + 382 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "This section outlines training a brain encoder using features from Sec. 3.1 as ground truth. Given brain responses $s \\in \\mathbb{R}^{1 \\times L_s}$ (with $L_{s}$ indicating the dimension of input brain data) and associated visual stimuli $v \\in \\mathbb{R}^{W \\times H \\times C}$ (representing image width, height, and channels), we train the brain encoder $\\mathcal{B}$ to minimize the distance between brain features $\\mathbf{b}$ and target image features $\\mathbf{v}$ from the vision encoder $\\mathcal{V}$ , aiming for a close approximation $\\mathcal{B}(s) \\approx \\mathcal{V}(v)$ . Toward the objective, the most straightforward alignment is through element-wise feature reconstruction [53]:", + "bbox": [ + 511, + 388, + 908, + 541 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {R}} = \\mathbb {E} _ {\\mathbf {b} \\sim \\mathbf {B}, \\mathbf {v} \\sim \\mathbf {V}} [ \\| \\mathcal {V} (v) - \\mathcal {B} (s)) \\| _ {2} ^ {2} ], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 583, + 550, + 906, + 570 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where the brain encoder $\\mathcal{B}$ learns the alignment between source brain space $\\mathbf{B}$ and target image space $\\mathbf{V}$ .", + "bbox": [ + 511, + 580, + 905, + 611 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Regression or Denoising. Regression using the Mean Squared Error (MSE, i.e., L2) loss directly learns a deterministic mapping from brain data points to those of images [53]. However, one downside of this deterministic objective is that it can lead to overfitting, as the model may memorize specific instances rather than generalizing well to unseen data—especially considering the data scarcity in brain modalities. This issue becomes particularly problematic when dealing with noisy or incomplete data, as is often the case with brain signals, making the model sensitive to minor variations in the input. Denoising [16, 47, 48], on the other hand, has been shown to be more effective than regression, as introducing noise into training data acts as an implicit form of data augmentation and regularization [26, 52]. The denoising process encourages the model to focus on the underlying data manifold rather than memorizing specific instance values [16, 40, 47, 48]. Furthermore, visual signals suffer from heavy spatial redundancy [14]; recent work has shown that such redundancy also exists in visual features [6].", + "bbox": [ + 511, + 613, + 908, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "4372", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/464f58031a8542d0ff44551a0e2e5f7fddbcefe8c1d459acb61d12dc6baae7c6.jpg", + "image_caption": [ + "Figure 2. Training Procedure with Denoiser." + ], + "image_footnote": [], + "bbox": [ + 135, + 93, + 433, + 303 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Therefore, we introduce a masked denoising optimization objective as shown in Fig. 2. The training procedure of the $\\pi$ -parameterized denoiser $\\mathcal{J}_{\\pi}$ follows a diffusion process:", + "bbox": [ + 89, + 344, + 482, + 390 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {\\varepsilon , t} \\| \\mathcal {J} _ {\\pi} (\\mathbf {v} _ {t}; \\mathbf {b}, t) - \\varepsilon \\| ^ {2}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 397, + 483, + 416 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Here, $\\varepsilon$ represents the ground truth noise sampled from the standard normal distribution $\\mathcal{N}(\\mathbf{0},\\mathbf{I})$ . The noise-corruption is defined as $\\mathbf{v}_t\\coloneqq \\sqrt{\\bar{\\alpha}_t}\\mathbf{v} + \\sqrt{1 - \\bar{\\alpha}_t\\varepsilon}$ , where $\\bar{\\alpha}_t$ is a noise schedule [16, 33] at timestep $t$ , and $\\mathbf{v} = \\mathcal{V}(v)$ maps input images $v$ to its representation. The noise estimator $\\mathcal{J}_{\\pi}(\\mathbf{v}_t;\\mathbf{b},t)$ predicts the noise level in $\\mathbf{v}_t$ at timestep $t$ , conditioned on brain predictions $\\mathbf{b}$ . The positional embedding [26] is added to the masked features to indicate the positions that need to be predicted, with the loss computed only on the unknown ones [14, 26]. The final loss as shown in Fig. 2 is the weighted sum of the regression loss and the denoising loss.", + "bbox": [ + 88, + 424, + 483, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "These features can be extracted from images on the fly during training or pre-calculated in advance to accelerate the process. During inference, the features predicted by the brain encoder, given brain signals as input, are fed into the remaining MLLM components (connector and LLM) for instruction-following tasks [29, 53], such as continuous dialogues, detailed descriptions, and complex reasoning.", + "bbox": [ + 89, + 590, + 483, + 696 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. MG-BrainDUB", + "text_level": 1, + "bbox": [ + 89, + 709, + 245, + 724 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To better evaluate the different choices of feature spaces in decoding fine-grained information, we propose a benchmark for detailed brain understanding. This involves curating high-quality evaluation datasets for two tasks: detailed captioning and salient question answering, each with corresponding metrics. We name this benchmark the Multi-Granularity Brain Descriptive Understanding Benchmark (MG-BrainDUB).", + "bbox": [ + 89, + 734, + 483, + 840 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.1. Descriptive Caption Understanding", + "text_level": 1, + "bbox": [ + 89, + 848, + 398, + 864 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Current brain captioning methods [12, 13, 49, 53] face two primary challenges. First, they rely on COCO [28] annota", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "tions, which provide only short captions with limited classes, lacking the granularity and vocabulary needed for detailed descriptions. Second, existing evaluation metrics are inadequate for assessing fine-grained captions. Traditional rule-based metrics, such as BLEU [36], CIDEr [51], and METEOR [4], compute $n$ -gram matching scores between candidate and reference captions, making them highly sensitive to stylistic variations and leading to inconsistent evaluations. Model-based metrics, such as SPICE [3], suffer from outdated text encoders as backbones, while CLIP-based alternatives [15, 38] enforce a strict 77-token input limit, further restricting their effectiveness in evaluating detailed captions.", + "bbox": [ + 511, + 90, + 906, + 272 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We tackle both challenges through two key strategies. For detailed caption annotation, we construct ground truth descriptions using state-of-the-art MLLMs [29], followed by manual error correction, missing element addition, and hallucination removal through expert human intervention. For evaluation metrics, in addition to leveraging CAPTURE [9], a metric specifically designed for long captions, we introduce a structured evaluation framework that assesses captions based on precision, recall, and F1 scores of three core visual elements: objects, attributes, and relations [9, 32]. To achieve this, we develop an automated pipeline with two key steps: visual element extraction and structured matching.", + "bbox": [ + 511, + 273, + 908, + 455 + ], + "page_idx": 3 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "(a) Element Extraction. Given detailed captions, we first split them into separate sentences using NLTK toolkit [30] and extract key visual elements—objects, attributes, and their relations—using a T5-based factual parser [27, 39].", + "(b) Structured Matching. The matching process aligns extracted objects, attributes, and relations in three steps: exact matching (identifying precise word matches), synonym matching (matching words with similar meanings), and semantic matching (computing cosine similarity of word embeddings for any remaining unmatched elements)." + ], + "bbox": [ + 511, + 455, + 908, + 611 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Finally, we compute the precision, recall, and F1 scores for visual elements based on the matching results. These metrics capture semantic discrepancies: precision penalizes hallucinations and recall captures omissions, reducing confounds from pretrained generative model biases and offering a more accurate measure of genuine brain-based decoding.", + "bbox": [ + 511, + 611, + 906, + 703 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4.2. Question-Answering Annotation", + "text_level": 1, + "bbox": [ + 511, + 712, + 799, + 728 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Previous studies have emphasized the multimodal evaluation for decoded salient objects, considering semantic selectivities in the higher visual cortex of the human brain [8, 18, 37]. Tiny objects in stimuli may not capture the subject's attention or relevant brain activities may not be effectively recorded during experiments [2, 53]. Thus, we further design evaluation focused on salient objects using annotated question-answering pairs. First, we extract salient objects using SAM [22] and input the cropped regions into the MLLM to identify objects, attributes, and relationships within each image using pipelines introduced in Sec. 4.1. We use SAM", + "bbox": [ + 511, + 734, + 908, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "4373", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/00ae8b4bbaf2cf2e4c766c1818c40f5c672dae6971f83d8c5c4ce7618240a9e8.jpg", + "table_caption": [ + "Table 1. Concept Localization. We present brain grounding [53] results derived directly from brain signals, reflecting spatial information encoded in brain activities. Image-based results are included as references, comprising visual grounding from Shikra [7] using associated visual stimuli and grounding performance based on reconstructed images from state-of-the-art visual decoding methods [35, 41, 53, 54]. We highlight the best, second-best, and third-best performance averaged across all four subjects." + ], + "table_footnote": [], + "table_body": "
MethodAll acc@0.5IoUSalient acc@0.5IoUSalient Creatures acc@0.5IoUSalient Objects acc@0.5IoUInconspicuous acc@0.5IoU
VisionShikra [7]51.9647.2262.9256.4466.7159.3458.7953.2738.2935.71
BrainDiffuser [35]17.4919.3427.1827.4638.7134.6314.6219.665.399.20
MindEye [41]15.3418.6523.8326.9629.2931.6417.8821.864.748.28
DREAM [54]16.2118.6526.5127.3534.4333.8517.8820.283.357.78
UMBRAE [53]16.8318.6927.1027.5534.1433.6519.4420.924.007.64
BrainUMBRAE-S [53]14.3118.1322.2825.8226.3229.5817.8821.724.378.53
UMBRAE [53]18.0320.6328.2229.3236.9335.6118.7422.475.329.79
VINDEX-S14.9218.5523.3826.2828.0030.2918.3521.924.378.91
VINDEX18.1020.8928.5029.7837.0036.2019.2522.785.119.80
", + "bbox": [ + 174, + 151, + 823, + 311 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "instead of the original COCO location annotations or Referring Expression Comprehension [7] with COCO classes [28] to extract such information as pseudolabels, since COCO often overlooks salient categories outside its 80 common classes. For instance, an image may feature a large church, but only two small clocks on it are annotated with their spatial locations and linguistic descriptions. For each identified object, we design questions such as \"which description best fits the in the image\" and \"what is the relationship between and to evaluate the model's ability to recognize objects and their interactions within the image. Additionally, we include in-depth reasoning questions [29] that require step-by-step logical processing, based on GPT annotations [1, 43], to assess deeper reasoning capabilities.", + "bbox": [ + 88, + 337, + 485, + 549 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "To mitigate the MLLM's tendency to default to affirmative responses such as \"Yes, I can see it\" rather than genuinely recognizing an object, we follow [32] and introduce an equivalent number of question-answer pairs for objects not present in the image as hallucination components. We then ask the model to select the correct description from multiple choices, where one option is a correct caption, and two are incorrect. The attribute annotations of fine-grained objects serve as the answers, while erroneous linguistic annotations are manually modified to ensure they contradict the actual visual content. We refer to this as salient QA (SQA) for complex reasoning and report the response accuracy.", + "bbox": [ + 89, + 554, + 485, + 736 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 763, + 223, + 781 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In our experiments, we demonstrate multimodal interaction with fMRI using MLLMs to tackle four tasks: concept localization (Sec. 5.2), concise captioning (Sec. 5.3), descriptive captioning (Sec. 5.4), and complex reasoning (Sec. 5.5). While the generated multimodal explanations have been shown to enhance reconstruction performance [43, 49, 53], we prioritize brain perception over reconstruction.", + "bbox": [ + 89, + 794, + 485, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 337, + 732, + 353 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Architecture. For the mixture of vision encoders, we use pre-trained CLIP [38] and DINOv2 [34] as visual encoders. The target image features are extracted from both transformer encoders and then interleaved to spatially mix visual tokens before being subsequently processed by the connector and LLM [50]. The feature dimension $D$ is 4,096 for 7B models and 5,120 for 13B models. For aggregated features, we utilize SigLIP [58] to extract visual features from multiple layers [57]. The connector first encapsulates multi-layer visual features and then maps the integrated visual features to the token space. For nested features, we use CLIP [38] to encode visual cues in a coarse-to-fine manner [6], where an image is represented as patch-based visual tokens. We create $M$ sets of tokens, with the visual tokens at the coarser level derived directly from those at the finer level. Specifically, starting from the initial visual tokens, we sequentially apply $2 \\times 2$ pooling with a stride of 2, resulting in $12 \\times 12$ , $6 \\times 6$ , and $3 \\times 3$ visual tokens. Finally, we apply $3 \\times 3$ pooling to obtain the most condensed single visual token. Details on encoders and MLLMs can be found in the supplementary.", + "bbox": [ + 511, + 364, + 908, + 669 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training Details. Our models are trained on a single A100 GPU. CLIP-224 [38] and DINO-224 [34] feature alignment with a batch size of 128 takes approximately 8 hours, CLIP-336 [38] with a batch size of 64 takes about 23 hours, while SigLIP-384 [58] with a batch size of 4 requires around 40 hours to converge. We use AdamW [31] as the optimizer with $\\beta_{1} = 0.9$ , $\\beta_{2} = 0.95$ , weight decay of 0.01, and apply the one-cycle strategy [46] with an initial learning rate of 3e-4 for the learning rate scheduler. Following visual decoding studies [41, 49, 54], we use the standard train and test splits for four subjects (sub01, sub02, sub05, sub07), each with 24,980 training samples and a shared 982 test samples. For evaluation, we report the average of the three repetitions of the same image in the test set, totaling 982 samples per subject. We implement the denoiser as an MLP with a default", + "bbox": [ + 511, + 672, + 910, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "4374", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/ea7d29e1251f0befef875fe7b29d3e3e000d6875cd943d1a684696aa6d79186e.jpg", + "table_caption": [ + "Table 2. Brain Captioning. We evaluate short captions on BrainHub [53] and detailed captions on MG-BrainDUB. Models with '-S' refers to those with cross-subject support but are trained in a single-subject setting for comparison. MindEye2 results are based on a model adapted to one subject using a pre-trained model that was trained on the remaining seven subjects with the full dataset. The mark† denotes zero-shot methods, indicating that no external captions were used during training. 'Shikra-image' refers to image captioning results from Shikra [7] using corresponding visual stimuli as input, acting as an approximate performance upper bound. Colors denote best, second-best, and third-best performance. Bold and underline represent the best and second best results under identical protocols." + ], + "table_footnote": [], + "table_body": "
MethodBLEU1BLEU2BLEU3BLEU4METEORROUGECIDErSPICECLIP-SRefCLIP-S
Concise Captioning (Sec. 5.3)
Shikra-img [7]82.3869.9058.6349.6635.6065.49161.4327.6280.6085.92
SDRecon [49]36.2117.117.723.4310.0325.1313.835.0261.0766.36
OneLLM [13]47.0426.9715.499.5113.5535.0522.996.2654.8061.28
BrainCap [12]55.9636.2122.7014.5116.6840.6941.309.0664.3169.90
NeuroVLA [43]57.1937.1723.7815.8518.6036.6749.5112.3965.49-
MindEye2 [42]54.8238.6026.4918.1617.5443.7755.7010.9767.5473.73
MEVOX [55]†58.5640.3628.0920.1119.2044.4754.3711.0564.2370.36
UMBRAE-S†57.6338.0225.0016.7618.4142.1551.9311.8366.4472.12
UMBRAE [53]†59.4440.4827.6619.0319.4543.7161.0612.7967.7873.54
VINDEX-S†57.9939.0026.0417.8318.5642.2753.9611.9266.9472.64
VINDEX†60.0040.7227.5718.9119.5143.7860.3212.8468.2673.88
Descriptive Captioning (Sec. 5.4)
NeuroVLA [43]38.9124.0215.2412.4118.4427.8342.5818.4156.16-
VINDEX
w/ SE22.7112.897.214.1113.9525.4314.058.4362.9067.72
w/ ME21.3911.866.313.3911.3117.606.046.4360.8565.96
w/ AF0.090.010.000.000.041.890.030.0153.1759.86
w/ NF120.1011.326.143.4014.6721.151.479.8560.3563.01
w/ NF950.7733.2521.1713.2018.2641.7246.5011.4565.4071.20
w/ NF3638.8723.8714.398.8014.7233.4031.078.9464.0269.80
w/ NF1443.742.221.270.743.4610.456.581.6957.1463.77
", + "bbox": [ + 174, + 176, + 826, + 510 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "depth of 1, a latent dimension (width) of 1024, and a weight multiplier of 1.0 for the denoising loss (indicating the same contribution as the regression loss during training).", + "bbox": [ + 89, + 534, + 482, + 583 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.2. Concept Localization", + "text_level": 1, + "bbox": [ + 89, + 590, + 292, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "This section evaluates the model's ability for concept localization through the brain grounding task [53]. Similar to visual referring expression comprehension, brain grounding uses the prompt \"Find the coordinates of $<\\text{expr}>$ \" and evaluates performance using accuracy and IoU metrics. The accuracy metric, acc@m, measures the percentage of correctly labeled instances with an IoU greater than the threshold $m$ . We assess grounding results derived directly from brain signals, with image-based results provided for reference. These include visual grounding using ground truth stimuli and reconstructed images from state-of-the-art (SOTA) visual decoding methods [35, 41, 54] via Shikra [7]. We provide a comparison with SOTAs and our methods in Tab. 1.", + "bbox": [ + 88, + 612, + 485, + 808 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The results demonstrate that our method VINDEX outperforms all baselines across all metrics in the four salient-based categories in BrainHub [53]. Notably, VINDEX surpasses both direct brain signal-based predictions from UM-BRAE [53] and image-based grounding results, which use reconstructed images from brain signals as input [35, 41, 53,", + "bbox": [ + 89, + 810, + 485, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "54]. This emphasizes that directly extracting location information from the captured brain activities is more accurate than relying on reconstructed images. This observation suggests that reconstruction methods relying on image and text semantic encoding have inherent limitations in decoding spatial information. It also implies that incorporating grounding information in future models could help enhance the accuracy of object location prediction when reconstructing the visual stimuli from neural activations.", + "bbox": [ + 511, + 534, + 908, + 672 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.3. Concise Captioning", + "text_level": 1, + "bbox": [ + 511, + 681, + 702, + 698 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We evaluate short captions on BrainHub [53], demonstrating our method's performance on multimodal language tasks. Tab. 2 presents an evaluation of our brain captioning compared to state-of-the-art baselines, including three subject-specific methods SDRecon [49], BrainCap [12], OneLLM [13], and four cross-subject methods NeuroVLA [43], MindEye2 [42], MEVOX [55], and UMBRAE [53]. Our method, VINDEX, for concise captioning is built upon the same base MLLM Shikra [7] as in UMBRAE [53] for a fair comparison. MindEye2 is first pre-trained on seven subjects and then adapted to the remaining subject using all available data. NeuroVLA [43], UMBRAE [53], and VINDEX are trained on four subjects", + "bbox": [ + 511, + 703, + 910, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "4375", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/97777d63fe5de5fb2fcceec3e4421e2ab66ef750cb44b1f5f125496c0bb9b141.jpg", + "table_caption": [ + "Table 3. Descriptive Captioning and Complex Reasoning. We evaluate detailed descriptions and complex reasoning with salient question-answering on MG-BrainDub under different model settings. 'Space' refers to the feature space elaborated in Sec. 3.1, while 'size' and 'encoder' indicate the size of ground truth images and the encoder used for feature extraction. Colors denote best, second-best, and third-best performance. Models marked with * correspond to those with the same settings as in Tab. 2." + ], + "table_footnote": [], + "table_body": "
SpaceModel SettingCAPTUREObjectAttributeRelationRuntime (s)SQA Acc (%)
SizeEncoderMLLMPRF1PRF1PRF1
SE*224CLIPLLaVA-1.5 7B0.435657.2249.7651.8837.9036.0835.7345.2540.7042.3214.0575.50
SE224CLIPLLaVA-1.5 13B0.473361.8152.9155.6040.6841.2139.3849.2845.0046.528.7077.92
SE224CLIPLLaVA-1.6 7B0.468451.0258.7853.5136.0154.0141.8041.0045.6342.7510.9176.80
SE224CLIPLLaVA-1.6 13B0.464550.4457.8352.8135.8954.0241.7240.4245.4442.3824.1878.69
ME*224CLIP, DINOLLaVA-1.5 13B0.304149.1734.3038.0726.1023.0923.3733.0826.8628.865.0652.66
AF*384SigLIPLLaVA-1.5 7B0.00581.891.071.270.150.110.130.010.000.005.6915.17
NF1*224CLIPLLaVA-1.5 7B0.487759.1957.1957.2442.7442.7641.1847.3246.7046.575.4582.59
NF9*224CLIPLLaVA-1.5 7B0.502162.4359.2659.7441.9743.2040.9550.5649.4949.545.3183.83
NF36*224CLIPLLaVA-1.5 7B0.467262.2354.8956.3740.2438.5637.5847.9044.4145.474.3366.35
NF144*224CLIPLLaVA-1.5 7B0.161452.8422.4728.738.675.556.3212.887.679.221.8331.35
NF1224CLIPLLaVA-1.6 7B0.435949.4751.3749.4335.6549.5940.2336.7538.9637.408.6265.58
NF1336CLIPLLaVA-1.5 7B0.412047.9946.7446.8943.0139.5039.5331.1531.9631.165.0662.27
", + "bbox": [ + 125, + 148, + 870, + 316 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "instead of eight. Caption results from Shikra [7] using associated visual stimuli for captioning act as an approximate upper bound. VINDEX achieves top performance across most metrics for concise brain captioning, even surpassing methods trained with more subject data [42] or external annotations [43]. The other brain-feature alignment methods UMBRAE [53] and MEVOX [55] ranked the second and the third. This highlights the effectiveness of the cross-subject brain feature alignment strategy, showcasing superior zero-shot performance without the need for external training data and generalization across subjects without requiring separate models or subject-specific parameters.", + "bbox": [ + 88, + 342, + 485, + 523 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "NeuroVLA, MindEye2, UMBRAE, and VINDEX generate fluent, complete, and information-rich sentences, highlighting the advantages of using LLMs. Notably, VINDEX and UMBRAE achieve superior performance in a zero-shot setting, without relying on additional annotations, outperforming models trained with real captions from COCO [28] or pseudo captions from MLLMs [29]. This demonstrates that directly aligning brain signals with image features offers a better way than using external visual, spatial, or linguistic annotations. It preserves more accurate semantic and spatial cues decoded from brain signals and avoids factual errors or stylistic inconsistencies from external annotations.", + "bbox": [ + 89, + 523, + 485, + 705 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.4. Descriptive Captioning", + "text_level": 1, + "bbox": [ + 89, + 712, + 305, + 729 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We present a quantitative comparison of descriptive captioning results for NeuroVLA [43] and our method with different feature space choices: single encoder (SE), mixture of encoders (ME), aggregated feature (AF), and nested features $\\mathrm{(NF_n)}$ where the token size $n$ is selected from $\\{1, 9, 36, 144\\}$ after downsampling. The latter Tab. 2 presents evaluations using regular caption metrics from [53]. Considering the limitations of traditional rule-based metrics (sensitive to writing style) and model-based metrics (unable to process long sequences), we evaluate descriptive captioning on MG-BrainDUB (Sec. 4.1). Results in Tab. 3 show that using", + "bbox": [ + 89, + 734, + 485, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "$\\mathrm{NF}_9$ (with 9 visual tokens) as supervision for training brain encoders achieves the best performance on most metrics. Surprisingly, using dense feature aggregation (AF) instead yields the worst results, and models trained with vanilla CLIP features [38] (SE setting at first four rows) remain inferior, even with larger and more sophisticated MLLMs (scaling from v1.5 to v1.6, 7B to 13B).", + "bbox": [ + 511, + 342, + 906, + 449 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training with feature spaces from larger CLIP models (336 vs. 224) does not improve performance but significantly increases training time and computational costs. The brain encoder aligned with CLIP-ViT-L-224 features (corresponding to $16 \\times 16$ tokens) generally transfers well to MLLMs designed with the CLIP-ViT-L-336 model ( $24 \\times 24$ tokens) as the visual encoder, often yielding better results. AF performs the worst but with the largest token size (729 visual tokens for SigLIP-384 [58]). In contrast, aligning brain encoders with CLIP-224 requires only one-fifth of SigLIP-384's training time and one-third of CLIP-336's. Recall that NF downsamples the predicted CLIP features (the same features used in SE/ME) using pooling operations. Notably, 9-token alignment yields the best results, while 1-token alignment ranks second, suggesting that fewer tokens, though less expressive, align better with brain signals, which encode less information than real images. This also explains the inferior results for ME, where mixing brain-CLIP and brain-DINO features amplifies prediction imprecisions, leading to worse performance than using brain-CLIP features alone.", + "bbox": [ + 511, + 455, + 908, + 758 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "The performance drop with increasing feature magnitude suggests that, while richer image features benefit visual perception tasks [6, 50, 57], brain signals capture relatively less comprehensive visual elements. This may due to subjects' selective attention [8, 18, 37] or limitations in capturing brain activities [2, 53]. We expect additional features to enhance the perception of subtler elements, but they mainly introduce unnecessary hallucinations and increase training costs, as brain signals predominantly encode basic visual elements.", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "4376", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Results in Tab. 3 also indicate that traditional caption metrics [4, 36, 51] are not reliable when scaling to longer, more detailed captions. The rule-based metrics show that ME has higher scores than $\\mathrm{NF}_1$ , even though $\\mathrm{NF}_1$ predicts more accurate objects, attributes, and relations. CLIP-based metrics [15] also show deficiencies, as they present high similarities for all generated captions, even for AF, which in most cases lead to nonsensical outputs, such as garbled text and blank spaces. The inference time does not increase as expected with the model complexity and token length; instead, it exhibits irregularities. $\\mathrm{NF}_9^*$ presents the best results across most metrics while maintaining low computation cost in terms of inference time and memory consumption.", + "bbox": [ + 89, + 90, + 480, + 287 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.5. Complex Reasoning", + "text_level": 1, + "bbox": [ + 89, + 296, + 279, + 313 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We evaluate the model's ability to perform complex reasoning on salient objects using Salient Question Answering (SQA) (Sec. 5.5), with accuracy reported in Tab. 3. Unlike previous sections focused on visual content, SQA involves in-depth reasoning that demands step-by-step logical processing to provide the right answer. This experiment highlights the model's ability to decode intricate characteristics of salient elements, meaning the prominent objects most likely to capture the subject's attention during brain activity capture [2], which have been shown to be better preserved in brain data [53]. From the table, $\\mathrm{NF}_9^*$ achieves the best SQA performance across image sizes, visual encoders, and MLLM settings, likely benefiting from a balance between rich visual features and brain data constraints, consistent with detailed description results. In contrast, ME and AF sometimes produce meaningless outputs due to improper feature-brain signal alignment, caused by the added complexity of features during training. They also suffer from hallucinations, generating non-existent objects and misrepresenting attributes and relationships, resulting in accuracy rates at or below random guessing.", + "bbox": [ + 89, + 318, + 482, + 635 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Ablation Studies", + "text_level": 1, + "bbox": [ + 89, + 648, + 253, + 664 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Model-Agnostic Evaluation. VINDEX is a versatile, model-agnostic brain decoding framework that can be easily adapted to various subjects, image sizes, visual encoders, MLLM settings, input prompts, and downstream tasks. Due to the architectural consistency of LLaVA-based variants [6, 7, 29, 50, 57], our trained brain encoders can be directly integrated into different MLLMs without requiring additional modifications. Specifically, for model setups presented in Tab. 3, we only need to train four brain encoders using CLIP-224, CLIP-336, DINO-224, and SigLIP-384. The same brain encoder trained with CLIP-224 can be used across SE, ME, and NF (except for the last one with a 336 image size). This brain-CLIP features are then passed to the remaining MLLMs components, including connectors and LLMs, with task prompts supported by the base MLLM for", + "bbox": [ + 89, + 672, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b4c282c672643ae5b96e98c235425ad72dc89e802443530c26f5df822fb63cc6.jpg", + "table_caption": [ + "Table 4. Ablation Study on Denoiser. CAPTURE [9] is reported across architectural depth $d$ , width $w$ , and weight multiplier $\\beta$ ." + ], + "table_footnote": [], + "table_body": "
SetupS0S1S2S3S4S5S6S7
β-1.01.01.01.00.51.52.0
d-1231333
w-5121,0241,0241,0241,0241,0241024
C0.41790.43080.42570.41830.43560.41930.41970.4096
", + "bbox": [ + 521, + 121, + 898, + 193 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "multimodal interactions, such as interactive dialogue, brain captioning, complex reasoning, and concept localization. The brain encoder aligned with CLIP-224 features adapts well to LLaVA-1.6-7B/13B models, which are designed to use CLIP-336 as the visual encoder, leading to improved performance. This versatility enables our method to efficiently align brain signals with multiple MLLMs at minimal cost.", + "bbox": [ + 511, + 222, + 903, + 328 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Denoising Optimization Objective. We conduct ablation studies on the denoiser in Tab. 4 to evaluate the effectiveness of the denoising objective over vanilla regression across different settings, focusing on architectural depth $d$ , width $w$ , and the weight multiplier $\\beta$ . Here, 'Sn' represents different configurations of specific depth, width, and weight multiplier; S0 refers to the base setup with only the regression loss, without the denoising loss. The results are reported as CAPTURE scores [9] (C) for detailed descriptions across each setup. Total training hours for 300 epochs across all settings is approximately eight hours, with no notable variations. We found that a simple one-depth MLP with 1,024 model channels yields the best results. Results show that training with the denoising loss setup incurs no extra cost but leads to improved performance and more stable training (further analysis in the supplementary material).", + "bbox": [ + 511, + 332, + 906, + 574 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "7. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 590, + 633, + 607 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a zero-shot multimodal brain decoding method that explores representative feature spaces from the visual components within MLLMs. We elucidate different vision feature space choices and introduce a denoiser-based strategy to enhance brain alignment with pre-trained models, improving the accuracy and robustness of brain-encoded feature representations. Our method facilitates efficient and flexible decoding across multiple levels of granularity, eliminating the need for additional textual or spatial annotations during training. Furthermore, we present MG-BrainDUB, a benchmark that provides a more reliable and comprehensive evaluation for multigranular multimodal brain decoding. It comprises two core tasks—detailed description generation and salient question answering—along with a novel metric that prioritizes key visual elements including objects, attributes, and relationships, offering a robust framework for assessing fine-grained brain perception.", + "bbox": [ + 511, + 618, + 908, + 875 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by a UKRI Future Leaders Fellowship [grant number G104084] and an Imminent Research Grant.", + "bbox": [ + 511, + 877, + 903, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "4377", + "bbox": [ + 480, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5", + "[2] Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116-126, 2022. 4, 7, 8", + "[3] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 1, 4", + "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 1, 4, 8", + "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3", + "[6] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. In ICLR, 2025. 2, 3, 5, 7, 8", + "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8", + "[8] Robert Desimone, Thomas D Albright, Charles G Gross, and Charles Bruce. Stimulus-selective properties of inferior temporal neurons in the macaque. Journal of Neuroscience, 4 (8):2051-2062, 1984. 4, 7", + "[9] Hongyuan Dong, Jiawen Li, Bohong Wu, Jiacong Wang, Yuan Zhang, and Haoyuan Guo. Benchmarking and improving detail image caption. arXiv preprint arXiv:2405.19092, 2024. 4, 8", + "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2", + "[11] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 3", + "[12] Matteo Ferrante, Furkan Ozcelik, Tommaso Boccato, Rufin VanRullen, and Nicola Toschi. Brain captioning: Decoding human brain activity into images and text. arXiv preprint arXiv:2305.11560, 2023. 4, 6", + "[13] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In CVPR, 2024. 1, 2, 4, 6", + "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 3, 4", + "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras," + ], + "bbox": [ + 93, + 114, + 483, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 1, 4, 8", + "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3, 4", + "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, pages 4651-4664. PMLR, 2021. 2", + "[18] Nancy Kanwisher, Josh McDermott, and Marvin M Chun. The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of neuroscience, 17 (11):4302-4311, 1997. 4, 7", + "[19] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. In ECCV, pages 113-132. Springer, 2024. 2, 3", + "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, pages 8107-8116, 2020. 2", + "[21] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In WACV, pages 1383-1392, 2024. 3", + "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3, 4", + "[23] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning. In NeurIPS, pages 30233-30249, 2022. 3", + "[24] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understanding. In ICML, 2023. 3", + "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 1", + "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS, 37:56424-56445, 2024. 3, 4", + "[27] Zhuang Li, Yuyang Chai, Terry Yue Zhuo, Lizhen Qu, Gholamreza Haffari, Fei Li, Donghong Ji, and Quan Hung Tran. Factual: A benchmark for faithful and consistent textual scene graph parsing. In ACL, 2023. 4", + "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 4, 5, 7", + "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3, 4, 5, 7, 8" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "4378", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[30] Edward Loper and Steven Bird. Nltk: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, 2004. 4", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5", + "[32] Fan Lu, Wei Wu, Kecheng Zheng, Shuailei Ma, Biao Gong, Jiawei Liu, Wei Zhai, Yang Cao, Yujun Shen, and Zheng-Jun Zha. Benchmarking large vision-language models via directed scene graph for comprehensive image captioning. In CVPR, 2025. 4, 5", + "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 4", + "[34] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2023. 3, 5", + "[35] Furkan Ozcelik and Rufin VanRullen. Brain-Diffuser: Natural scene reconstruction from fMRI signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023. 1, 2, 5, 6", + "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 1, 4, 8", + "[37] Aina Puce, Truett Allison, Maryam Asgari, John C Gore, and Gregory McCarthy. Differential sensitivity of human visual cortex to faces, letterstrings, and textures: a functional magnetic resonance imaging study. Journal of neuroscience, 16(16):5205-5215, 1996. 4, 7", + "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7", + "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 4", + "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3", + "[41] Paul S Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Ethan Cohen, Aidan J Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. In NeurIPS, 2023. 1, 2, 5, 6", + "[42] Paul S Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, and Tanishq Mathew Abraham. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML, 2024. 1, 2, 6, 7", + "[43] Guobin Shen, Dongcheng Zhao, Xiang He, Linghao Feng, Yiting Dong, Jihang Wang, Qian Zhang, and Yi Zeng. Neurovision to language: Enhancing brain recording-based visual" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction and language interaction. In NeurIPS, pages 98083-98110, 2024. 1, 2, 5, 6, 7", + "[44] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In NeurIPS, pages 42048-42070, 2025. 2", + "[45] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, Bryan Catanzaro, Andrew Tao, Jan Kautz, Zhiding Yu, and Guilin Liu. Eagle: Exploring the design space for multimodal llms with mixture of encoders. In ICLR, 2025. 3", + "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5", + "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3", + "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3", + "[49] Yu Takagi and Shinji Nishimoto. Improving visual image reconstruction from human brain activity using latent diffusion models via multiple decoded inputs. arXiv preprint arXiv:2306.11536, 2023. 1, 2, 4, 5, 6", + "[50] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, pages 9568-9578, 2024. 2, 5, 7, 8", + "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 1, 4, 8", + "[52] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Zheng Ge, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In ICLR, 2025. 3", + "[53] Weihao Xia, Raoul de Charette, Cengiz Oztireli, and JingHao Xue. Umbrae: Unified multimodal brain decoding. In ECCV, pages 242-259, 2024. 1, 2, 3, 4, 5, 6, 7, 8", + "[54] Weihao Xia, Raoul de Charette, Cengiz Özireli, and Jing-Hao Xue. Dream: Visual decoding from reversing human visual system. In WACV, pages 8226-8235, 2024. 1, 2, 5, 6", + "[55] Weihao Xia and Cengiz Öztireli. Mevox: Multi-task vision experts for brain captioning. In CVPR Workshop, 2025. 2, 6, 7", + "[56] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 2", + "[57] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms. In NeurIPS, pages 33108-33140, 2025. 2, 3, 5, 7, 8", + "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 7", + "[59] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In NeurIPS, 2024. 2" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "4379", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_model.json b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_model.json new file mode 100644 index 0000000000000000000000000000000000000000..bbdf8f8fdcf8d6f24233906cf7fbcfc6fa2f5163 --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_model.json @@ -0,0 +1,2002 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.149, + 0.131, + 0.851, + 0.154 + ], + "angle": 0, + "content": "Exploring The Visual Feature Space for Multimodal Neural Decoding" + }, + { + "type": "text", + "bbox": [ + 0.368, + 0.18, + 0.631, + 0.201 + ], + "angle": 0, + "content": "Weihao Xia \\(\\boxtimes\\) Cengiz Oztireli" + }, + { + "type": "text", + "bbox": [ + 0.398, + 0.206, + 0.596, + 0.225 + ], + "angle": 0, + "content": "University of Cambridge" + }, + { + "type": "text", + "bbox": [ + 0.356, + 0.231, + 0.642, + 0.245 + ], + "angle": 0, + "content": "https://weihaox.github.io/VINDEX" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.265, + 0.327, + 0.281 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.297, + 0.486, + 0.614 + ], + "angle": 0, + "content": "The intrication of brain signals drives research that leverages multimodal AI to align brain modalities with visual and textual data for explainable descriptions. However, most existing studies are limited to coarse interpretations, lacking essential details on object descriptions, locations, attributes, and their relationships. This leads to imprecise and ambiguous reconstructions when using such cues for visual decoding. To address this, we analyze different choices of vision feature spaces from pre-trained visual components within Multimodal Large Language Models (MLLMs) and introduce a zero-shot multimodal brain decoding method that interacts with these models to decode across multiple levels of granularities. To assess a model's ability to decode fine details from brain signals, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub). This benchmark includes two key tasks: detailed descriptions and salient question-answering, with metrics highlighting key visual elements like objects, attributes, and relationships. Our approach enhances neural decoding precision and supports more accurate neuro-decoding applications. Code is available at https://github.com/weihaox/VINDEX." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.637, + 0.222, + 0.652 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.484, + 0.795 + ], + "angle": 0, + "content": "Human brain signals are inherently complex and not directly interpretable, posing a significant challenge for decoding brain activity. To address this, recent studies have focused on translating brain signals associated visual stimuli into multimodal explanations, such as textual descriptions [13, 42], depth information [54], bounding boxes [53], or direct image reconstruction [35, 41, 42]. Although these efforts have made notable strides, achieving a fine-grained understanding of brain modalities remains an unresolved challenge." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.484, + 0.902 + ], + "angle": 0, + "content": "One line of research [42, 49, 53] has attempted to translate brain activations from visual stimuli into textual captions. However, existing methods typically generate one-sentence descriptions with limited details, failing to capture the full spectrum of visual perception. The most straightforward approach [43] is to use off-the-shelf MLLMs [25, 29] to generate detailed descriptions from paired images for each" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.908, + 0.493 + ], + "angle": 0, + "content": "brain recording, then train a model to decode these pseudo-annotations from brain signals. However, this practice is computationally inefficient and may introduce bias or errors, making it impractical for generalizable brain decoding. The other key issue is the evaluation of brain-decoded detailed captions, which can be hundreds of words long, rendering conventional metrics unsuitable. Specifically, traditional captioning metrics [4, 36, 51] which mostly rely on \\( n \\)-gram matching between candidate and reference captions, are highly sensitive to stylistic variations, leading to inconsistent evaluations. Model-based metrics either use outdated text encoders [3] or struggle with long text due to token length limitations [15, 38]. While concurrent studies [43] attempt to generate detailed captions for visual decoding, they lack reliable evaluations of caption quality and accuracy." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.494, + 0.909, + 0.795 + ], + "angle": 0, + "content": "In this paper, we address the two critical issues. For the method, we investigate the exploitation of various feature spaces from the visual components within MLLMs to enable fine-grained multimodal explanations from neural activations. We present VINDEX (Visual experts for multimodal Neural DEcoding eXploration), which leverages diverse feature spaces for multimodal neural decoding. We further introduce a denoiser-based alignment strategy that efficiently maps brain tokens to image tokens, allowing for integration into pre-trained MLLMs. This approach facilitates decoding across multiple levels of granularities without the need for additional textual or spatial annotations. For evaluation, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub) to discern a model's ability to decode fine-grained details from brain signals. MG-BrainDub consists of two key tasks: detailed descriptions and salient question-answering, with metrics focusing on key visual elements, including objects, attributes, and relationships, derived from both candidate and ground truth captions. Our contributions are summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.796, + 0.905, + 0.841 + ], + "angle": 0, + "content": "- Through systematic analysis, we exploit representative feature spaces of visual components within MLLMs for brain signal alignment and multimodal brain interaction." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.905, + 0.886 + ], + "angle": 0, + "content": "- We introduce a zero-shot multimodal brain decoding method for multi-granular decoding without the need for additional textual or spatial annotations during training." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.887, + 0.905, + 0.902 + ], + "angle": 0, + "content": "- We construct a benchmark with metrics for fine-grained" + }, + { + "type": "list", + "bbox": [ + 0.513, + 0.796, + 0.905, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4370" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.105, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "brain interpretation, evaluating a model's ability to decode detailed descriptions and perform complex reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.138, + 0.233, + 0.154 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.486, + 0.56 + ], + "angle": 0, + "content": "Multimodal Brain Perception. Recent advancements in decoding multimodal cues from brain signals, facilitated by MLLMs, have achieved unprecedented performance [35, 41, 49, 54]. These methods typically map brain responses, captured through functional magnetic resonance imaging (fMRI), to more commonly used modalities that can be processed by pre-trained models for subsequent information extraction [20, 40, 56]. For instance, Takagi and Nishimoto [49] connect fMRI data with CLIP text embeddings and the latent space of Stable Diffusion (SD) [40] using ridge regression to reconstruct visual stimuli. Xia et al. [54] extract semantic, depth, and color features from fMRI data to reconstruct images using a depth- and color-conditioned SD model. Several methods [13, 42, 49, 54] aim to create readable multimodal explanations of visual stimuli, such as textual description or depth estimation, but often rely on corresponding external supervision to guide the multimodal training. In contrast, UMBRAE [53] has recently demonstrated the ability to directly decode textual descriptions and bounding boxes from fMRI data using only image-fMRI pairs during training, without requiring additional annotations. This method aligns image features with visual encoders and integrates them into MLLMs. Building on this foundation, our proposed approach aims to further elucidate feature space representations, extracting more granular information while introducing metrics for evaluation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.564, + 0.487, + 0.883 + ], + "angle": 0, + "content": "Multimodal Large Language Models. Expanding Large Language Models (LLMs) to incorporate additional modalities, such as images, has recently attracted significant interest [10, 29, 50, 57]. These models typically comprise three components: a frozen image encoder, a trainable connector, and a frozen or finetuned LLM. The connector bridges the gap from the image feature space to the LLM token space, which can be implemented as a linear layer [7], a multilayer perceptron (MLP) [29], or a lightweight transformer [17]. Recent studies have shown that using pre-trained CLIP [38] as the visual component can lead to compromised performance in visual tasks, particularly in grounding capabilities [7, 50]. To overcome this limitation, researchers have explored ways to more effectively utilize the visual component to enhance perception across diverse scenarios. This includes employing a mixture of vision encoders [19, 44, 50, 59] pretrained on various tasks as expert models and investigating the feature spaces [6, 57] within CLIP. Drawing inspiration from these efforts, we exploit brain alignment with feature spaces in MLLM's visual component for multi-granular representation to facilitate zero-shot multimodal brain decoding." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.485, + 0.902 + ], + "angle": 0, + "content": "Multimodal-Brain Alignment. The prevailing practice" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.909, + 0.395 + ], + "angle": 0, + "content": "for brain alignment is to map neural modalities into a shared multimodal space [13, 35, 41, 53], which can be categorized into generative alignment and discriminative alignment. The first category, exemplified by OneLLM [13], employs generative training to learn multimodal alignment, connecting multimodal inputs, including brain signals, with an LLM. Methods in the second category align brain signals within a pre-trained embedding space, such as CLIP [38], using techniques like linear regression [35, 49], contrastive learning [54], diffusion priors [41], or feature reconstruction [53]. MEVOX [55] ensembles multiple task-specific experts to capture distinct aspects of brain perception. The brain encoder learns to align with omni-contextual explanations. Most of these alignment strategies require per-subject training or subject-specific annotations, which can lead to scalability issues in practice. While UMBRAE [53] performs zero-shot multimodal brain decoding, its regression-based alignment is deterministic. We analyze the limitations of this strategy and propose incorporating a novel denoising optimization objective to enhance the brain-feature alignment." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.41, + 0.605, + 0.424 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.909, + 0.679 + ], + "angle": 0, + "content": "Current MLLMs [7, 29] comprise three components: a pretrained vision encoder, a pre-trained LLM, and a connector that bridges vision and language models. Given significant discrepancies in data scale and challenges associated with acquiring large-scale brain datasets, we argue that aligning brain signals with pre-trained visual components, rather than training from scratch [13, 43], is a more effective and efficient approach. This not only alleviates the need to construct extensive text for training MLLMs but also enables zero-shot capability in decoding multimodal cues, allowing for versatile usage across different MLLMs. Thus, we present the rationale behind our method design in three key aspects: (1) the selection of vision feature spaces (Sec. 3.1), (2) the alignment strategy to connect activations with pre-trained MLLMs (Sec. 3.2), and (3) the development of a benchmark necessary for fine-grained brain perception (Sec. 4)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.689, + 0.748, + 0.705 + ], + "angle": 0, + "content": "3.1. Exploring Feature Spaces" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.909, + 0.803 + ], + "angle": 0, + "content": "We explore four distinct feature spaces, drawing inspiration from prior MLLM research that leverage different feature representations. By aligning brain signals across these feature spaces, we aim to capture multi-granular representations in MLLMs, facilitating zero-shot multimodal brain decoding. The four feature space choices are illustrated in Fig. 1." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.807, + 0.908, + 0.852 + ], + "angle": 0, + "content": "Single Encoder. This represents using features from a single pre-trained vision encoder (e.g. CLIP [38]) for brain alignment, which is the most commonly used in MMLMs." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Mixture of Encoders. Encoders trained on diverse tasks allow us to explore the distinct advantages of different vision experts. These include contrastive image-text learning" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "4371" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.145, + 0.124, + 0.251, + 0.17 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.139, + 0.191, + 0.248, + 0.204 + ], + "angle": 0, + "content": "(a) Single Encoder" + }, + { + "type": "image", + "bbox": [ + 0.278, + 0.096, + 0.415, + 0.187 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.275, + 0.191, + 0.412, + 0.204 + ], + "angle": 0, + "content": "(b) Mixture of Encoders" + }, + { + "type": "image", + "bbox": [ + 0.439, + 0.096, + 0.653, + 0.186 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.191, + 0.608, + 0.204 + ], + "angle": 0, + "content": "(c) Aggregated Feature" + }, + { + "type": "image", + "bbox": [ + 0.682, + 0.122, + 0.862, + 0.178 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.71, + 0.19, + 0.822, + 0.203 + ], + "angle": 0, + "content": "(d) Nested Features" + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.215, + 0.907, + 0.286 + ], + "angle": 0, + "content": "Figure 1. Feature Spaces. We aim to capture multi-granular representations for zero-shot multimodal brain decoding by aligning brain signals across different feature spaces in MLLM's visual component. The explored representative feature spaces include: (a) Single Encoder, which utilizes a single encoder to extract selected features; (b) Mixture of Encoders, using hybrid featured from different vision experts specialized in task-specific domains; (c) Aggregated Feature, fusing features from different layers within the same image encoder; and (d) Nested Features, downscaling visual representations in a hierarchical coarse-to-fine structure with different perceptual granularities." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.312, + 0.486, + 0.508 + ], + "angle": 0, + "content": "CLIP [38], self-supervised learning DINO [34], and object-centric models pre-trained for various tasks such as segmentation (SAM [22]), detection (EVA [11]), and recognition (Pix2Struct [24]). These task-specific vision encoders allow MLLMs to achieve optimal performance within respective pre-training domains [19, 45]. Aligning brain signals with features from these models allows for the capture of distinct facets of visual perception. Here, we explore the integrated features from CLIP [38] and DINO [34] encoders for simplicity while demonstrating the core concept of aligning brain signals with a mixture of vision experts. CLIP effectively encodes semantic information, while DINO excels in capturing spatial details such as object location and boundaries." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.485, + 0.713 + ], + "angle": 0, + "content": "Aggregated Feature. Existing MLLMs typically use features from a selected layer in the visual component as the visual representation input for LLMs. Features from different levels—such as shallow, middle, and deep layers—are recognized for capturing different characteristics of the image [57]. Therefore, by encapsulating information from multiple layers, we can enrich the visual input for the LLM. Specifically, we partition features from selected layers into several groups. Features within each group from adjacent layers are integrated along the channel dimension to reduce redundancy and mitigate high dimensionality. The features from each group are then concatenated with the features from the final layer before being fed into the connector." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.486, + 0.902 + ], + "angle": 0, + "content": "Nested Features. MLLMs typically project feature images from the vision encoder into a fixed-length set of visual tokens [29]. Longer token sequences generally yield better results but require more resources. Recent studies have found that the optimal token length is task-specific, laying the foundation for visual token pruning and merging to dynamically adjust the sequence length [5, 21]. Therefore, we downscale selected features using a set of downsampling factors and produce nested features. This approach offers a flexible and controllable method to encode visual content as nested sets of visual tokens with a hierarchical structure, from coarse to fine details—much like a Matryoshka doll [6, 23]. The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.358 + ], + "angle": 0, + "content": "nested features capture structured information across multiple coarse-to-fine levels, enabling adaptable token lengths that align with different perceptual granularities." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.367, + 0.882, + 0.383 + ], + "angle": 0, + "content": "3.2. Denoising for Multimodal Brain Alignment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.389, + 0.909, + 0.542 + ], + "angle": 0, + "content": "This section outlines training a brain encoder using features from Sec. 3.1 as ground truth. Given brain responses \\( s \\in \\mathbb{R}^{1 \\times L_s} \\) (with \\( L_{s} \\) indicating the dimension of input brain data) and associated visual stimuli \\( v \\in \\mathbb{R}^{W \\times H \\times C} \\) (representing image width, height, and channels), we train the brain encoder \\( \\mathcal{B} \\) to minimize the distance between brain features \\( \\mathbf{b} \\) and target image features \\( \\mathbf{v} \\) from the vision encoder \\( \\mathcal{V} \\), aiming for a close approximation \\( \\mathcal{B}(s) \\approx \\mathcal{V}(v) \\). Toward the objective, the most straightforward alignment is through element-wise feature reconstruction [53]:" + }, + { + "type": "equation", + "bbox": [ + 0.584, + 0.551, + 0.907, + 0.571 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {R}} = \\mathbb {E} _ {\\mathbf {b} \\sim \\mathbf {B}, \\mathbf {v} \\sim \\mathbf {V}} [ \\| \\mathcal {V} (v) - \\mathcal {B} (s)) \\| _ {2} ^ {2} ], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.581, + 0.906, + 0.612 + ], + "angle": 0, + "content": "where the brain encoder \\(\\mathcal{B}\\) learns the alignment between source brain space \\(\\mathbf{B}\\) and target image space \\(\\mathbf{V}\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.614, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Regression or Denoising. Regression using the Mean Squared Error (MSE, i.e., L2) loss directly learns a deterministic mapping from brain data points to those of images [53]. However, one downside of this deterministic objective is that it can lead to overfitting, as the model may memorize specific instances rather than generalizing well to unseen data—especially considering the data scarcity in brain modalities. This issue becomes particularly problematic when dealing with noisy or incomplete data, as is often the case with brain signals, making the model sensitive to minor variations in the input. Denoising [16, 47, 48], on the other hand, has been shown to be more effective than regression, as introducing noise into training data acts as an implicit form of data augmentation and regularization [26, 52]. The denoising process encourages the model to focus on the underlying data manifold rather than memorizing specific instance values [16, 40, 47, 48]. Furthermore, visual signals suffer from heavy spatial redundancy [14]; recent work has shown that such redundancy also exists in visual features [6]." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4372" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.136, + 0.094, + 0.434, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.148, + 0.306, + 0.427, + 0.321 + ], + "angle": 0, + "content": "Figure 2. Training Procedure with Denoiser." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.345, + 0.483, + 0.391 + ], + "angle": 0, + "content": "Therefore, we introduce a masked denoising optimization objective as shown in Fig. 2. The training procedure of the \\(\\pi\\)-parameterized denoiser \\(\\mathcal{J}_{\\pi}\\) follows a diffusion process:" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.398, + 0.484, + 0.417 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {\\varepsilon , t} \\| \\mathcal {J} _ {\\pi} (\\mathbf {v} _ {t}; \\mathbf {b}, t) - \\varepsilon \\| ^ {2}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.425, + 0.484, + 0.59 + ], + "angle": 0, + "content": "Here, \\(\\varepsilon\\) represents the ground truth noise sampled from the standard normal distribution \\(\\mathcal{N}(\\mathbf{0},\\mathbf{I})\\). The noise-corruption is defined as \\(\\mathbf{v}_t\\coloneqq \\sqrt{\\bar{\\alpha}_t}\\mathbf{v} + \\sqrt{1 - \\bar{\\alpha}_t\\varepsilon}\\), where \\(\\bar{\\alpha}_t\\) is a noise schedule [16, 33] at timestep \\(t\\), and \\(\\mathbf{v} = \\mathcal{V}(v)\\) maps input images \\(v\\) to its representation. The noise estimator \\(\\mathcal{J}_{\\pi}(\\mathbf{v}_t;\\mathbf{b},t)\\) predicts the noise level in \\(\\mathbf{v}_t\\) at timestep \\(t\\), conditioned on brain predictions \\(\\mathbf{b}\\). The positional embedding [26] is added to the masked features to indicate the positions that need to be predicted, with the loss computed only on the unknown ones [14, 26]. The final loss as shown in Fig. 2 is the weighted sum of the regression loss and the denoising loss." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.591, + 0.484, + 0.698 + ], + "angle": 0, + "content": "These features can be extracted from images on the fly during training or pre-calculated in advance to accelerate the process. During inference, the features predicted by the brain encoder, given brain signals as input, are fed into the remaining MLLM components (connector and LLM) for instruction-following tasks [29, 53], such as continuous dialogues, detailed descriptions, and complex reasoning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.71, + 0.246, + 0.725 + ], + "angle": 0, + "content": "4. MG-BrainDUB" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.484, + 0.841 + ], + "angle": 0, + "content": "To better evaluate the different choices of feature spaces in decoding fine-grained information, we propose a benchmark for detailed brain understanding. This involves curating high-quality evaluation datasets for two tasks: detailed captioning and salient question answering, each with corresponding metrics. We name this benchmark the Multi-Granularity Brain Descriptive Understanding Benchmark (MG-BrainDUB)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.4, + 0.866 + ], + "angle": 0, + "content": "4.1. Descriptive Caption Understanding" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Current brain captioning methods [12, 13, 49, 53] face two primary challenges. First, they rely on COCO [28] annota" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.908, + 0.273 + ], + "angle": 0, + "content": "tions, which provide only short captions with limited classes, lacking the granularity and vocabulary needed for detailed descriptions. Second, existing evaluation metrics are inadequate for assessing fine-grained captions. Traditional rule-based metrics, such as BLEU [36], CIDEr [51], and METEOR [4], compute \\( n \\)-gram matching scores between candidate and reference captions, making them highly sensitive to stylistic variations and leading to inconsistent evaluations. Model-based metrics, such as SPICE [3], suffer from outdated text encoders as backbones, while CLIP-based alternatives [15, 38] enforce a strict 77-token input limit, further restricting their effectiveness in evaluating detailed captions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.274, + 0.909, + 0.456 + ], + "angle": 0, + "content": "We tackle both challenges through two key strategies. For detailed caption annotation, we construct ground truth descriptions using state-of-the-art MLLMs [29], followed by manual error correction, missing element addition, and hallucination removal through expert human intervention. For evaluation metrics, in addition to leveraging CAPTURE [9], a metric specifically designed for long captions, we introduce a structured evaluation framework that assesses captions based on precision, recall, and F1 scores of three core visual elements: objects, attributes, and relations [9, 32]. To achieve this, we develop an automated pipeline with two key steps: visual element extraction and structured matching." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.457, + 0.907, + 0.518 + ], + "angle": 0, + "content": "(a) Element Extraction. Given detailed captions, we first split them into separate sentences using NLTK toolkit [30] and extract key visual elements—objects, attributes, and their relations—using a T5-based factual parser [27, 39]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.521, + 0.909, + 0.612 + ], + "angle": 0, + "content": "(b) Structured Matching. The matching process aligns extracted objects, attributes, and relations in three steps: exact matching (identifying precise word matches), synonym matching (matching words with similar meanings), and semantic matching (computing cosine similarity of word embeddings for any remaining unmatched elements)." + }, + { + "type": "list", + "bbox": [ + 0.512, + 0.457, + 0.909, + 0.612 + ], + "angle": 0, + "content": null + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.612, + 0.907, + 0.704 + ], + "angle": 0, + "content": "Finally, we compute the precision, recall, and F1 scores for visual elements based on the matching results. These metrics capture semantic discrepancies: precision penalizes hallucinations and recall captures omissions, reducing confounds from pretrained generative model biases and offering a more accurate measure of genuine brain-based decoding." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.713, + 0.8, + 0.729 + ], + "angle": 0, + "content": "4.2. Question-Answering Annotation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.735, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Previous studies have emphasized the multimodal evaluation for decoded salient objects, considering semantic selectivities in the higher visual cortex of the human brain [8, 18, 37]. Tiny objects in stimuli may not capture the subject's attention or relevant brain activities may not be effectively recorded during experiments [2, 53]. Thus, we further design evaluation focused on salient objects using annotated question-answering pairs. First, we extract salient objects using SAM [22] and input the cropped regions into the MLLM to identify objects, attributes, and relationships within each image using pipelines introduced in Sec. 4.1. We use SAM" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4373" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.148 + ], + "angle": 0, + "content": "Table 1. Concept Localization. We present brain grounding [53] results derived directly from brain signals, reflecting spatial information encoded in brain activities. Image-based results are included as references, comprising visual grounding from Shikra [7] using associated visual stimuli and grounding performance based on reconstructed images from state-of-the-art visual decoding methods [35, 41, 53, 54]. We highlight the best, second-best, and third-best performance averaged across all four subjects." + }, + { + "type": "table", + "bbox": [ + 0.176, + 0.152, + 0.825, + 0.313 + ], + "angle": 0, + "content": "
MethodAll acc@0.5IoUSalient acc@0.5IoUSalient Creatures acc@0.5IoUSalient Objects acc@0.5IoUInconspicuous acc@0.5IoU
VisionShikra [7]51.9647.2262.9256.4466.7159.3458.7953.2738.2935.71
BrainDiffuser [35]17.4919.3427.1827.4638.7134.6314.6219.665.399.20
MindEye [41]15.3418.6523.8326.9629.2931.6417.8821.864.748.28
DREAM [54]16.2118.6526.5127.3534.4333.8517.8820.283.357.78
UMBRAE [53]16.8318.6927.1027.5534.1433.6519.4420.924.007.64
BrainUMBRAE-S [53]14.3118.1322.2825.8226.3229.5817.8821.724.378.53
UMBRAE [53]18.0320.6328.2229.3236.9335.6118.7422.475.329.79
VINDEX-S14.9218.5523.3826.2828.0030.2918.3521.924.378.91
VINDEX18.1020.8928.5029.7837.0036.2019.2522.785.119.80
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.338, + 0.486, + 0.55 + ], + "angle": 0, + "content": "instead of the original COCO location annotations or Referring Expression Comprehension [7] with COCO classes [28] to extract such information as pseudolabels, since COCO often overlooks salient categories outside its 80 common classes. For instance, an image may feature a large church, but only two small clocks on it are annotated with their spatial locations and linguistic descriptions. For each identified object, we design questions such as \"which description best fits the in the image\" and \"what is the relationship between and to evaluate the model's ability to recognize objects and their interactions within the image. Additionally, we include in-depth reasoning questions [29] that require step-by-step logical processing, based on GPT annotations [1, 43], to assess deeper reasoning capabilities." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.486, + 0.737 + ], + "angle": 0, + "content": "To mitigate the MLLM's tendency to default to affirmative responses such as \"Yes, I can see it\" rather than genuinely recognizing an object, we follow [32] and introduce an equivalent number of question-answer pairs for objects not present in the image as hallucination components. We then ask the model to select the correct description from multiple choices, where one option is a correct caption, and two are incorrect. The attribute annotations of fine-grained objects serve as the answers, while erroneous linguistic annotations are manually modified to ensure they contradict the actual visual content. We refer to this as salient QA (SQA) for complex reasoning and report the response accuracy." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.765, + 0.224, + 0.782 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.486, + 0.903 + ], + "angle": 0, + "content": "In our experiments, we demonstrate multimodal interaction with fMRI using MLLMs to tackle four tasks: concept localization (Sec. 5.2), concise captioning (Sec. 5.3), descriptive captioning (Sec. 5.4), and complex reasoning (Sec. 5.5). While the generated multimodal explanations have been shown to enhance reconstruction performance [43, 49, 53], we prioritize brain perception over reconstruction." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.338, + 0.733, + 0.354 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.366, + 0.909, + 0.67 + ], + "angle": 0, + "content": "Architecture. For the mixture of vision encoders, we use pre-trained CLIP [38] and DINOv2 [34] as visual encoders. The target image features are extracted from both transformer encoders and then interleaved to spatially mix visual tokens before being subsequently processed by the connector and LLM [50]. The feature dimension \\(D\\) is 4,096 for 7B models and 5,120 for 13B models. For aggregated features, we utilize SigLIP [58] to extract visual features from multiple layers [57]. The connector first encapsulates multi-layer visual features and then maps the integrated visual features to the token space. For nested features, we use CLIP [38] to encode visual cues in a coarse-to-fine manner [6], where an image is represented as patch-based visual tokens. We create \\(M\\) sets of tokens, with the visual tokens at the coarser level derived directly from those at the finer level. Specifically, starting from the initial visual tokens, we sequentially apply \\(2 \\times 2\\) pooling with a stride of 2, resulting in \\(12 \\times 12\\), \\(6 \\times 6\\), and \\(3 \\times 3\\) visual tokens. Finally, we apply \\(3 \\times 3\\) pooling to obtain the most condensed single visual token. Details on encoders and MLLMs can be found in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.674, + 0.911, + 0.903 + ], + "angle": 0, + "content": "Training Details. Our models are trained on a single A100 GPU. CLIP-224 [38] and DINO-224 [34] feature alignment with a batch size of 128 takes approximately 8 hours, CLIP-336 [38] with a batch size of 64 takes about 23 hours, while SigLIP-384 [58] with a batch size of 4 requires around 40 hours to converge. We use AdamW [31] as the optimizer with \\(\\beta_{1} = 0.9\\), \\(\\beta_{2} = 0.95\\), weight decay of 0.01, and apply the one-cycle strategy [46] with an initial learning rate of 3e-4 for the learning rate scheduler. Following visual decoding studies [41, 49, 54], we use the standard train and test splits for four subjects (sub01, sub02, sub05, sub07), each with 24,980 training samples and a shared 982 test samples. For evaluation, we report the average of the three repetitions of the same image in the test set, totaling 982 samples per subject. We implement the denoiser as an MLP with a default" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4374" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.174 + ], + "angle": 0, + "content": "Table 2. Brain Captioning. We evaluate short captions on BrainHub [53] and detailed captions on MG-BrainDUB. Models with '-S' refers to those with cross-subject support but are trained in a single-subject setting for comparison. MindEye2 results are based on a model adapted to one subject using a pre-trained model that was trained on the remaining seven subjects with the full dataset. The mark† denotes zero-shot methods, indicating that no external captions were used during training. 'Shikra-image' refers to image captioning results from Shikra [7] using corresponding visual stimuli as input, acting as an approximate performance upper bound. Colors denote best, second-best, and third-best performance. Bold and underline represent the best and second best results under identical protocols." + }, + { + "type": "table", + "bbox": [ + 0.175, + 0.178, + 0.827, + 0.511 + ], + "angle": 0, + "content": "
MethodBLEU1BLEU2BLEU3BLEU4METEORROUGECIDErSPICECLIP-SRefCLIP-S
Concise Captioning (Sec. 5.3)
Shikra-img [7]82.3869.9058.6349.6635.6065.49161.4327.6280.6085.92
SDRecon [49]36.2117.117.723.4310.0325.1313.835.0261.0766.36
OneLLM [13]47.0426.9715.499.5113.5535.0522.996.2654.8061.28
BrainCap [12]55.9636.2122.7014.5116.6840.6941.309.0664.3169.90
NeuroVLA [43]57.1937.1723.7815.8518.6036.6749.5112.3965.49-
MindEye2 [42]54.8238.6026.4918.1617.5443.7755.7010.9767.5473.73
MEVOX [55]†58.5640.3628.0920.1119.2044.4754.3711.0564.2370.36
UMBRAE-S†57.6338.0225.0016.7618.4142.1551.9311.8366.4472.12
UMBRAE [53]†59.4440.4827.6619.0319.4543.7161.0612.7967.7873.54
VINDEX-S†57.9939.0026.0417.8318.5642.2753.9611.9266.9472.64
VINDEX†60.0040.7227.5718.9119.5143.7860.3212.8468.2673.88
Descriptive Captioning (Sec. 5.4)
NeuroVLA [43]38.9124.0215.2412.4118.4427.8342.5818.4156.16-
VINDEX
w/ SE22.7112.897.214.1113.9525.4314.058.4362.9067.72
w/ ME21.3911.866.313.3911.3117.606.046.4360.8565.96
w/ AF0.090.010.000.000.041.890.030.0153.1759.86
w/ NF120.1011.326.143.4014.6721.151.479.8560.3563.01
w/ NF950.7733.2521.1713.2018.2641.7246.5011.4565.4071.20
w/ NF3638.8723.8714.398.8014.7233.4031.078.9464.0269.80
w/ NF1443.742.221.270.743.4610.456.581.6957.1463.77
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.584 + ], + "angle": 0, + "content": "depth of 1, a latent dimension (width) of 1024, and a weight multiplier of 1.0 for the denoising loss (indicating the same contribution as the regression loss during training)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.591, + 0.293, + 0.607 + ], + "angle": 0, + "content": "5.2. Concept Localization" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.613, + 0.486, + 0.809 + ], + "angle": 0, + "content": "This section evaluates the model's ability for concept localization through the brain grounding task [53]. Similar to visual referring expression comprehension, brain grounding uses the prompt \"Find the coordinates of \\(<\\text{expr}>\\)\" and evaluates performance using accuracy and IoU metrics. The accuracy metric, acc@m, measures the percentage of correctly labeled instances with an IoU greater than the threshold \\(m\\). We assess grounding results derived directly from brain signals, with image-based results provided for reference. These include visual grounding using ground truth stimuli and reconstructed images from state-of-the-art (SOTA) visual decoding methods [35, 41, 54] via Shikra [7]. We provide a comparison with SOTAs and our methods in Tab. 1." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.486, + 0.902 + ], + "angle": 0, + "content": "The results demonstrate that our method VINDEX outperforms all baselines across all metrics in the four salient-based categories in BrainHub [53]. Notably, VINDEX surpasses both direct brain signal-based predictions from UM-BRAE [53] and image-based grounding results, which use reconstructed images from brain signals as input [35, 41, 53," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.535, + 0.909, + 0.673 + ], + "angle": 0, + "content": "54]. This emphasizes that directly extracting location information from the captured brain activities is more accurate than relying on reconstructed images. This observation suggests that reconstruction methods relying on image and text semantic encoding have inherent limitations in decoding spatial information. It also implies that incorporating grounding information in future models could help enhance the accuracy of object location prediction when reconstructing the visual stimuli from neural activations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.682, + 0.704, + 0.699 + ], + "angle": 0, + "content": "5.3. Concise Captioning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.704, + 0.911, + 0.903 + ], + "angle": 0, + "content": "We evaluate short captions on BrainHub [53], demonstrating our method's performance on multimodal language tasks. Tab. 2 presents an evaluation of our brain captioning compared to state-of-the-art baselines, including three subject-specific methods SDRecon [49], BrainCap [12], OneLLM [13], and four cross-subject methods NeuroVLA [43], MindEye2 [42], MEVOX [55], and UMBRAE [53]. Our method, VINDEX, for concise captioning is built upon the same base MLLM Shikra [7] as in UMBRAE [53] for a fair comparison. MindEye2 is first pre-trained on seven subjects and then adapted to the remaining subject using all available data. NeuroVLA [43], UMBRAE [53], and VINDEX are trained on four subjects" + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4375" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.146 + ], + "angle": 0, + "content": "Table 3. Descriptive Captioning and Complex Reasoning. We evaluate detailed descriptions and complex reasoning with salient question-answering on MG-BrainDub under different model settings. 'Space' refers to the feature space elaborated in Sec. 3.1, while 'size' and 'encoder' indicate the size of ground truth images and the encoder used for feature extraction. Colors denote best, second-best, and third-best performance. Models marked with * correspond to those with the same settings as in Tab. 2." + }, + { + "type": "table", + "bbox": [ + 0.127, + 0.15, + 0.872, + 0.318 + ], + "angle": 0, + "content": "
SpaceModel SettingCAPTUREObjectAttributeRelationRuntime (s)SQA Acc (%)
SizeEncoderMLLMPRF1PRF1PRF1
SE*224CLIPLLaVA-1.5 7B0.435657.2249.7651.8837.9036.0835.7345.2540.7042.3214.0575.50
SE224CLIPLLaVA-1.5 13B0.473361.8152.9155.6040.6841.2139.3849.2845.0046.528.7077.92
SE224CLIPLLaVA-1.6 7B0.468451.0258.7853.5136.0154.0141.8041.0045.6342.7510.9176.80
SE224CLIPLLaVA-1.6 13B0.464550.4457.8352.8135.8954.0241.7240.4245.4442.3824.1878.69
ME*224CLIP, DINOLLaVA-1.5 13B0.304149.1734.3038.0726.1023.0923.3733.0826.8628.865.0652.66
AF*384SigLIPLLaVA-1.5 7B0.00581.891.071.270.150.110.130.010.000.005.6915.17
NF1*224CLIPLLaVA-1.5 7B0.487759.1957.1957.2442.7442.7641.1847.3246.7046.575.4582.59
NF9*224CLIPLLaVA-1.5 7B0.502162.4359.2659.7441.9743.2040.9550.5649.4949.545.3183.83
NF36*224CLIPLLaVA-1.5 7B0.467262.2354.8956.3740.2438.5637.5847.9044.4145.474.3366.35
NF144*224CLIPLLaVA-1.5 7B0.161452.8422.4728.738.675.556.3212.887.679.221.8331.35
NF1224CLIPLLaVA-1.6 7B0.435949.4751.3749.4335.6549.5940.2336.7538.9637.408.6265.58
NF1336CLIPLLaVA-1.5 7B0.412047.9946.7446.8943.0139.5039.5331.1531.9631.165.0662.27
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.343, + 0.486, + 0.525 + ], + "angle": 0, + "content": "instead of eight. Caption results from Shikra [7] using associated visual stimuli for captioning act as an approximate upper bound. VINDEX achieves top performance across most metrics for concise brain captioning, even surpassing methods trained with more subject data [42] or external annotations [43]. The other brain-feature alignment methods UMBRAE [53] and MEVOX [55] ranked the second and the third. This highlights the effectiveness of the cross-subject brain feature alignment strategy, showcasing superior zero-shot performance without the need for external training data and generalization across subjects without requiring separate models or subject-specific parameters." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.525, + 0.486, + 0.707 + ], + "angle": 0, + "content": "NeuroVLA, MindEye2, UMBRAE, and VINDEX generate fluent, complete, and information-rich sentences, highlighting the advantages of using LLMs. Notably, VINDEX and UMBRAE achieve superior performance in a zero-shot setting, without relying on additional annotations, outperforming models trained with real captions from COCO [28] or pseudo captions from MLLMs [29]. This demonstrates that directly aligning brain signals with image features offers a better way than using external visual, spatial, or linguistic annotations. It preserves more accurate semantic and spatial cues decoded from brain signals and avoids factual errors or stylistic inconsistencies from external annotations." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.713, + 0.307, + 0.73 + ], + "angle": 0, + "content": "5.4. Descriptive Captioning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.486, + 0.902 + ], + "angle": 0, + "content": "We present a quantitative comparison of descriptive captioning results for NeuroVLA [43] and our method with different feature space choices: single encoder (SE), mixture of encoders (ME), aggregated feature (AF), and nested features \\(\\mathrm{(NF_n)}\\) where the token size \\(n\\) is selected from \\(\\{1, 9, 36, 144\\}\\) after downsampling. The latter Tab. 2 presents evaluations using regular caption metrics from [53]. Considering the limitations of traditional rule-based metrics (sensitive to writing style) and model-based metrics (unable to process long sequences), we evaluate descriptive captioning on MG-BrainDUB (Sec. 4.1). Results in Tab. 3 show that using" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.343, + 0.907, + 0.45 + ], + "angle": 0, + "content": "\\(\\mathrm{NF}_9\\) (with 9 visual tokens) as supervision for training brain encoders achieves the best performance on most metrics. Surprisingly, using dense feature aggregation (AF) instead yields the worst results, and models trained with vanilla CLIP features [38] (SE setting at first four rows) remain inferior, even with larger and more sophisticated MLLMs (scaling from v1.5 to v1.6, 7B to 13B)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.456, + 0.909, + 0.76 + ], + "angle": 0, + "content": "Training with feature spaces from larger CLIP models (336 vs. 224) does not improve performance but significantly increases training time and computational costs. The brain encoder aligned with CLIP-ViT-L-224 features (corresponding to \\(16 \\times 16\\) tokens) generally transfers well to MLLMs designed with the CLIP-ViT-L-336 model (\\(24 \\times 24\\) tokens) as the visual encoder, often yielding better results. AF performs the worst but with the largest token size (729 visual tokens for SigLIP-384 [58]). In contrast, aligning brain encoders with CLIP-224 requires only one-fifth of SigLIP-384's training time and one-third of CLIP-336's. Recall that NF downsamples the predicted CLIP features (the same features used in SE/ME) using pooling operations. Notably, 9-token alignment yields the best results, while 1-token alignment ranks second, suggesting that fewer tokens, though less expressive, align better with brain signals, which encode less information than real images. This also explains the inferior results for ME, where mixing brain-CLIP and brain-DINO features amplifies prediction imprecisions, leading to worse performance than using brain-CLIP features alone." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "The performance drop with increasing feature magnitude suggests that, while richer image features benefit visual perception tasks [6, 50, 57], brain signals capture relatively less comprehensive visual elements. This may due to subjects' selective attention [8, 18, 37] or limitations in capturing brain activities [2, 53]. We expect additional features to enhance the perception of subtler elements, but they mainly introduce unnecessary hallucinations and increase training costs, as brain signals predominantly encode basic visual elements." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "4376" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.288 + ], + "angle": 0, + "content": "Results in Tab. 3 also indicate that traditional caption metrics [4, 36, 51] are not reliable when scaling to longer, more detailed captions. The rule-based metrics show that ME has higher scores than \\(\\mathrm{NF}_1\\), even though \\(\\mathrm{NF}_1\\) predicts more accurate objects, attributes, and relations. CLIP-based metrics [15] also show deficiencies, as they present high similarities for all generated captions, even for AF, which in most cases lead to nonsensical outputs, such as garbled text and blank spaces. The inference time does not increase as expected with the model complexity and token length; instead, it exhibits irregularities. \\(\\mathrm{NF}_9^*\\) presents the best results across most metrics while maintaining low computation cost in terms of inference time and memory consumption." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.297, + 0.281, + 0.314 + ], + "angle": 0, + "content": "5.5. Complex Reasoning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.319, + 0.483, + 0.636 + ], + "angle": 0, + "content": "We evaluate the model's ability to perform complex reasoning on salient objects using Salient Question Answering (SQA) (Sec. 5.5), with accuracy reported in Tab. 3. Unlike previous sections focused on visual content, SQA involves in-depth reasoning that demands step-by-step logical processing to provide the right answer. This experiment highlights the model's ability to decode intricate characteristics of salient elements, meaning the prominent objects most likely to capture the subject's attention during brain activity capture [2], which have been shown to be better preserved in brain data [53]. From the table, \\(\\mathrm{NF}_9^*\\) achieves the best SQA performance across image sizes, visual encoders, and MLLM settings, likely benefiting from a balance between rich visual features and brain data constraints, consistent with detailed description results. In contrast, ME and AF sometimes produce meaningless outputs due to improper feature-brain signal alignment, caused by the added complexity of features during training. They also suffer from hallucinations, generating non-existent objects and misrepresenting attributes and relationships, resulting in accuracy rates at or below random guessing." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.649, + 0.254, + 0.665 + ], + "angle": 0, + "content": "6. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.674, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Model-Agnostic Evaluation. VINDEX is a versatile, model-agnostic brain decoding framework that can be easily adapted to various subjects, image sizes, visual encoders, MLLM settings, input prompts, and downstream tasks. Due to the architectural consistency of LLaVA-based variants [6, 7, 29, 50, 57], our trained brain encoders can be directly integrated into different MLLMs without requiring additional modifications. Specifically, for model setups presented in Tab. 3, we only need to train four brain encoders using CLIP-224, CLIP-336, DINO-224, and SigLIP-384. The same brain encoder trained with CLIP-224 can be used across SE, ME, and NF (except for the last one with a 336 image size). This brain-CLIP features are then passed to the remaining MLLMs components, including connectors and LLMs, with task prompts supported by the base MLLM for" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.09, + 0.905, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation Study on Denoiser. CAPTURE [9] is reported across architectural depth \\(d\\), width \\(w\\), and weight multiplier \\(\\beta\\)." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.122, + 0.9, + 0.194 + ], + "angle": 0, + "content": "
SetupS0S1S2S3S4S5S6S7
β-1.01.01.01.00.51.52.0
d-1231333
w-5121,0241,0241,0241,0241,0241024
C0.41790.43080.42570.41830.43560.41930.41970.4096
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.223, + 0.905, + 0.329 + ], + "angle": 0, + "content": "multimodal interactions, such as interactive dialogue, brain captioning, complex reasoning, and concept localization. The brain encoder aligned with CLIP-224 features adapts well to LLaVA-1.6-7B/13B models, which are designed to use CLIP-336 as the visual encoder, leading to improved performance. This versatility enables our method to efficiently align brain signals with multiple MLLMs at minimal cost." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.333, + 0.908, + 0.575 + ], + "angle": 0, + "content": "Denoising Optimization Objective. We conduct ablation studies on the denoiser in Tab. 4 to evaluate the effectiveness of the denoising objective over vanilla regression across different settings, focusing on architectural depth \\(d\\), width \\(w\\), and the weight multiplier \\(\\beta\\). Here, 'Sn' represents different configurations of specific depth, width, and weight multiplier; S0 refers to the base setup with only the regression loss, without the denoising loss. The results are reported as CAPTURE scores [9] (C) for detailed descriptions across each setup. Total training hours for 300 epochs across all settings is approximately eight hours, with no notable variations. We found that a simple one-depth MLP with 1,024 model channels yields the best results. Results show that training with the denoising loss setup incurs no extra cost but leads to improved performance and more stable training (further analysis in the supplementary material)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.592, + 0.634, + 0.608 + ], + "angle": 0, + "content": "7. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.619, + 0.909, + 0.875 + ], + "angle": 0, + "content": "In this paper, we propose a zero-shot multimodal brain decoding method that explores representative feature spaces from the visual components within MLLMs. We elucidate different vision feature space choices and introduce a denoiser-based strategy to enhance brain alignment with pre-trained models, improving the accuracy and robustness of brain-encoded feature representations. Our method facilitates efficient and flexible decoding across multiple levels of granularity, eliminating the need for additional textual or spatial annotations during training. Furthermore, we present MG-BrainDUB, a benchmark that provides a more reliable and comprehensive evaluation for multigranular multimodal brain decoding. It comprises two core tasks—detailed description generation and salient question answering—along with a novel metric that prioritizes key visual elements including objects, attributes, and relationships, offering a robust framework for assessing fine-grained brain perception." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.878, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by a UKRI Future Leaders Fellowship [grant number G104084] and an Imminent Research Grant." + }, + { + "type": "page_number", + "bbox": [ + 0.482, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "4377" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.171 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.173, + 0.484, + 0.255 + ], + "angle": 0, + "content": "[2] Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116-126, 2022. 4, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.258, + 0.483, + 0.299 + ], + "angle": 0, + "content": "[3] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 1, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.484, + 0.355 + ], + "angle": 0, + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.358, + 0.484, + 0.398 + ], + "angle": 0, + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.401, + 0.484, + 0.429 + ], + "angle": 0, + "content": "[6] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. In ICLR, 2025. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.431, + 0.484, + 0.485 + ], + "angle": 0, + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.488, + 0.484, + 0.543 + ], + "angle": 0, + "content": "[8] Robert Desimone, Thomas D Albright, Charles G Gross, and Charles Bruce. Stimulus-selective properties of inferior temporal neurons in the macaque. Journal of Neuroscience, 4 (8):2051-2062, 1984. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.545, + 0.484, + 0.599 + ], + "angle": 0, + "content": "[9] Hongyuan Dong, Jiawen Li, Bohong Wu, Jiacong Wang, Yuan Zhang, and Haoyuan Guo. Benchmarking and improving detail image caption. arXiv preprint arXiv:2405.19092, 2024. 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.602, + 0.484, + 0.671 + ], + "angle": 0, + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.673, + 0.484, + 0.728 + ], + "angle": 0, + "content": "[11] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.73, + 0.483, + 0.785 + ], + "angle": 0, + "content": "[12] Matteo Ferrante, Furkan Ozcelik, Tommaso Boccato, Rufin VanRullen, and Nicola Toschi. Brain captioning: Decoding human brain activity into images and text. arXiv preprint arXiv:2305.11560, 2023. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.787, + 0.483, + 0.842 + ], + "angle": 0, + "content": "[13] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In CVPR, 2024. 1, 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.845, + 0.483, + 0.886 + ], + "angle": 0, + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.888, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras," + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.906, + 0.121 + ], + "angle": 0, + "content": "and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.124, + 0.908, + 0.164 + ], + "angle": 0, + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.167, + 0.908, + 0.222 + ], + "angle": 0, + "content": "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, pages 4651-4664. PMLR, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.224, + 0.907, + 0.279 + ], + "angle": 0, + "content": "[18] Nancy Kanwisher, Josh McDermott, and Marvin M Chun. The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of neuroscience, 17 (11):4302-4311, 1997. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.282, + 0.907, + 0.338 + ], + "angle": 0, + "content": "[19] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. In ECCV, pages 113-132. Springer, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.34, + 0.907, + 0.394 + ], + "angle": 0, + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, pages 8107-8116, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.398, + 0.907, + 0.453 + ], + "angle": 0, + "content": "[21] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In WACV, pages 1383-1392, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.907, + 0.512 + ], + "angle": 0, + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.514, + 0.907, + 0.581 + ], + "angle": 0, + "content": "[23] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning. In NeurIPS, pages 30233-30249, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.584, + 0.907, + 0.655 + ], + "angle": 0, + "content": "[24] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understanding. In ICML, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.657, + 0.907, + 0.711 + ], + "angle": 0, + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.714, + 0.907, + 0.756 + ], + "angle": 0, + "content": "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS, 37:56424-56445, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.758, + 0.907, + 0.814 + ], + "angle": 0, + "content": "[27] Zhuang Li, Yuyang Chai, Terry Yue Zhuo, Lizhen Qu, Gholamreza Haffari, Fei Li, Donghong Ji, and Quan Hung Tran. Factual: A benchmark for faithful and consistent textual scene graph parsing. In ACL, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.816, + 0.907, + 0.871 + ], + "angle": 0, + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3, 4, 5, 7, 8" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4378" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[30] Edward Loper and Steven Bird. Nltk: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, 2004. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.484, + 0.163 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.484, + 0.232 + ], + "angle": 0, + "content": "[32] Fan Lu, Wei Wu, Kecheng Zheng, Shuailei Ma, Biao Gong, Jiawei Liu, Wei Zhai, Yang Cao, Yujun Shen, and Zheng-Jun Zha. Benchmarking large vision-language models via directed scene graph for comprehensive image captioning. In CVPR, 2025. 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.484, + 0.273 + ], + "angle": 0, + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.274, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[34] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2023. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.343, + 0.484, + 0.399 + ], + "angle": 0, + "content": "[35] Furkan Ozcelik and Rufin VanRullen. Brain-Diffuser: Natural scene reconstruction from fMRI signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.4, + 0.484, + 0.441 + ], + "angle": 0, + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.441, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[37] Aina Puce, Truett Allison, Maryam Asgari, John C Gore, and Gregory McCarthy. Differential sensitivity of human visual cortex to faces, letterstrings, and textures: a functional magnetic resonance imaging study. Journal of neuroscience, 16(16):5205-5215, 1996. 4, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.58 + ], + "angle": 0, + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.484, + 0.649 + ], + "angle": 0, + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.484, + 0.705 + ], + "angle": 0, + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.484, + 0.775 + ], + "angle": 0, + "content": "[41] Paul S Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Ethan Cohen, Aidan J Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. In NeurIPS, 2023. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.484, + 0.858 + ], + "angle": 0, + "content": "[42] Paul S Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, and Tanishq Mathew Abraham. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML, 2024. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[43] Guobin Shen, Dongcheng Zhao, Xiang He, Linghao Feng, Yiting Dong, Jihang Wang, Qian Zhang, and Yi Zeng. Neurovision to language: Enhancing brain recording-based visual" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "reconstruction and language interaction. In NeurIPS, pages 98083-98110, 2024. 1, 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.12, + 0.907, + 0.176 + ], + "angle": 0, + "content": "[44] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In NeurIPS, pages 42048-42070, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.177, + 0.907, + 0.259 + ], + "angle": 0, + "content": "[45] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, Bryan Catanzaro, Andrew Tao, Jan Kautz, Zhiding Yu, and Guilin Liu. Eagle: Exploring the design space for multimodal llms with mixture of encoders. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.26, + 0.906, + 0.316 + ], + "angle": 0, + "content": "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.316, + 0.906, + 0.343 + ], + "angle": 0, + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.907, + 0.383 + ], + "angle": 0, + "content": "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.385, + 0.907, + 0.441 + ], + "angle": 0, + "content": "[49] Yu Takagi and Shinji Nishimoto. Improving visual image reconstruction from human brain activity using latent diffusion models via multiple decoded inputs. arXiv preprint arXiv:2306.11536, 2023. 1, 2, 4, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.441, + 0.906, + 0.495 + ], + "angle": 0, + "content": "[50] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, pages 9568-9578, 2024. 2, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.907, + 0.538 + ], + "angle": 0, + "content": "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 1, 4, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.539, + 0.907, + 0.58 + ], + "angle": 0, + "content": "[52] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Zheng Ge, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In ICLR, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.58, + 0.907, + 0.622 + ], + "angle": 0, + "content": "[53] Weihao Xia, Raoul de Charette, Cengiz Oztireli, and JingHao Xue. Umbrae: Unified multimodal brain decoding. In ECCV, pages 242-259, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.907, + 0.664 + ], + "angle": 0, + "content": "[54] Weihao Xia, Raoul de Charette, Cengiz Özireli, and Jing-Hao Xue. Dream: Visual decoding from reversing human visual system. In WACV, pages 8226-8235, 2024. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.664, + 0.907, + 0.704 + ], + "angle": 0, + "content": "[55] Weihao Xia and Cengiz Öztireli. Mevox: Multi-task vision experts for brain captioning. In CVPR Workshop, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.706, + 0.907, + 0.748 + ], + "angle": 0, + "content": "[56] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.907, + 0.803 + ], + "angle": 0, + "content": "[57] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms. In NeurIPS, pages 33108-33140, 2025. 2, 3, 5, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.907, + 0.845 + ], + "angle": 0, + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.907, + 0.9 + ], + "angle": 0, + "content": "[59] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In NeurIPS, 2024. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "4379" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_origin.pdf b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..a6e893b7bb703abd5725bbd516102ce51510bfd1 --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/cf7c857e-63f8-4a94-a425-38a73f920d8b_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:8cd715ab8458a8b78857fcc49bf549eec5f993d6005e2d857ed7aa4a7bceb305 +size 515531 diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/full.md b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/full.md new file mode 100644 index 0000000000000000000000000000000000000000..5dadfe2a3d63951f5b56c741bd5775c91643aad1 --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/full.md @@ -0,0 +1,261 @@ +# Exploring The Visual Feature Space for Multimodal Neural Decoding + +Weihao Xia $\boxtimes$ Cengiz Oztireli + +University of Cambridge + +https://weihaox.github.io/VINDEX + +# Abstract + +The intrication of brain signals drives research that leverages multimodal AI to align brain modalities with visual and textual data for explainable descriptions. However, most existing studies are limited to coarse interpretations, lacking essential details on object descriptions, locations, attributes, and their relationships. This leads to imprecise and ambiguous reconstructions when using such cues for visual decoding. To address this, we analyze different choices of vision feature spaces from pre-trained visual components within Multimodal Large Language Models (MLLMs) and introduce a zero-shot multimodal brain decoding method that interacts with these models to decode across multiple levels of granularities. To assess a model's ability to decode fine details from brain signals, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub). This benchmark includes two key tasks: detailed descriptions and salient question-answering, with metrics highlighting key visual elements like objects, attributes, and relationships. Our approach enhances neural decoding precision and supports more accurate neuro-decoding applications. Code is available at https://github.com/weihaox/VINDEX. + +# 1. Introduction + +Human brain signals are inherently complex and not directly interpretable, posing a significant challenge for decoding brain activity. To address this, recent studies have focused on translating brain signals associated visual stimuli into multimodal explanations, such as textual descriptions [13, 42], depth information [54], bounding boxes [53], or direct image reconstruction [35, 41, 42]. Although these efforts have made notable strides, achieving a fine-grained understanding of brain modalities remains an unresolved challenge. + +One line of research [42, 49, 53] has attempted to translate brain activations from visual stimuli into textual captions. However, existing methods typically generate one-sentence descriptions with limited details, failing to capture the full spectrum of visual perception. The most straightforward approach [43] is to use off-the-shelf MLLMs [25, 29] to generate detailed descriptions from paired images for each + +brain recording, then train a model to decode these pseudo-annotations from brain signals. However, this practice is computationally inefficient and may introduce bias or errors, making it impractical for generalizable brain decoding. The other key issue is the evaluation of brain-decoded detailed captions, which can be hundreds of words long, rendering conventional metrics unsuitable. Specifically, traditional captioning metrics [4, 36, 51] which mostly rely on $n$ -gram matching between candidate and reference captions, are highly sensitive to stylistic variations, leading to inconsistent evaluations. Model-based metrics either use outdated text encoders [3] or struggle with long text due to token length limitations [15, 38]. While concurrent studies [43] attempt to generate detailed captions for visual decoding, they lack reliable evaluations of caption quality and accuracy. + +In this paper, we address the two critical issues. For the method, we investigate the exploitation of various feature spaces from the visual components within MLLMs to enable fine-grained multimodal explanations from neural activations. We present VINDEX (Visual experts for multimodal Neural DEcoding eXploration), which leverages diverse feature spaces for multimodal neural decoding. We further introduce a denoiser-based alignment strategy that efficiently maps brain tokens to image tokens, allowing for integration into pre-trained MLLMs. This approach facilitates decoding across multiple levels of granularities without the need for additional textual or spatial annotations. For evaluation, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub) to discern a model's ability to decode fine-grained details from brain signals. MG-BrainDub consists of two key tasks: detailed descriptions and salient question-answering, with metrics focusing on key visual elements, including objects, attributes, and relationships, derived from both candidate and ground truth captions. Our contributions are summarized as follows: + +- Through systematic analysis, we exploit representative feature spaces of visual components within MLLMs for brain signal alignment and multimodal brain interaction. +- We introduce a zero-shot multimodal brain decoding method for multi-granular decoding without the need for additional textual or spatial annotations during training. +- We construct a benchmark with metrics for fine-grained + +brain interpretation, evaluating a model's ability to decode detailed descriptions and perform complex reasoning. + +# 2. Related Work + +Multimodal Brain Perception. Recent advancements in decoding multimodal cues from brain signals, facilitated by MLLMs, have achieved unprecedented performance [35, 41, 49, 54]. These methods typically map brain responses, captured through functional magnetic resonance imaging (fMRI), to more commonly used modalities that can be processed by pre-trained models for subsequent information extraction [20, 40, 56]. For instance, Takagi and Nishimoto [49] connect fMRI data with CLIP text embeddings and the latent space of Stable Diffusion (SD) [40] using ridge regression to reconstruct visual stimuli. Xia et al. [54] extract semantic, depth, and color features from fMRI data to reconstruct images using a depth- and color-conditioned SD model. Several methods [13, 42, 49, 54] aim to create readable multimodal explanations of visual stimuli, such as textual description or depth estimation, but often rely on corresponding external supervision to guide the multimodal training. In contrast, UMBRAE [53] has recently demonstrated the ability to directly decode textual descriptions and bounding boxes from fMRI data using only image-fMRI pairs during training, without requiring additional annotations. This method aligns image features with visual encoders and integrates them into MLLMs. Building on this foundation, our proposed approach aims to further elucidate feature space representations, extracting more granular information while introducing metrics for evaluation. + +Multimodal Large Language Models. Expanding Large Language Models (LLMs) to incorporate additional modalities, such as images, has recently attracted significant interest [10, 29, 50, 57]. These models typically comprise three components: a frozen image encoder, a trainable connector, and a frozen or finetuned LLM. The connector bridges the gap from the image feature space to the LLM token space, which can be implemented as a linear layer [7], a multilayer perceptron (MLP) [29], or a lightweight transformer [17]. Recent studies have shown that using pre-trained CLIP [38] as the visual component can lead to compromised performance in visual tasks, particularly in grounding capabilities [7, 50]. To overcome this limitation, researchers have explored ways to more effectively utilize the visual component to enhance perception across diverse scenarios. This includes employing a mixture of vision encoders [19, 44, 50, 59] pretrained on various tasks as expert models and investigating the feature spaces [6, 57] within CLIP. Drawing inspiration from these efforts, we exploit brain alignment with feature spaces in MLLM's visual component for multi-granular representation to facilitate zero-shot multimodal brain decoding. + +Multimodal-Brain Alignment. The prevailing practice + +for brain alignment is to map neural modalities into a shared multimodal space [13, 35, 41, 53], which can be categorized into generative alignment and discriminative alignment. The first category, exemplified by OneLLM [13], employs generative training to learn multimodal alignment, connecting multimodal inputs, including brain signals, with an LLM. Methods in the second category align brain signals within a pre-trained embedding space, such as CLIP [38], using techniques like linear regression [35, 49], contrastive learning [54], diffusion priors [41], or feature reconstruction [53]. MEVOX [55] ensembles multiple task-specific experts to capture distinct aspects of brain perception. The brain encoder learns to align with omni-contextual explanations. Most of these alignment strategies require per-subject training or subject-specific annotations, which can lead to scalability issues in practice. While UMBRAE [53] performs zero-shot multimodal brain decoding, its regression-based alignment is deterministic. We analyze the limitations of this strategy and propose incorporating a novel denoising optimization objective to enhance the brain-feature alignment. + +# 3. Method + +Current MLLMs [7, 29] comprise three components: a pretrained vision encoder, a pre-trained LLM, and a connector that bridges vision and language models. Given significant discrepancies in data scale and challenges associated with acquiring large-scale brain datasets, we argue that aligning brain signals with pre-trained visual components, rather than training from scratch [13, 43], is a more effective and efficient approach. This not only alleviates the need to construct extensive text for training MLLMs but also enables zero-shot capability in decoding multimodal cues, allowing for versatile usage across different MLLMs. Thus, we present the rationale behind our method design in three key aspects: (1) the selection of vision feature spaces (Sec. 3.1), (2) the alignment strategy to connect activations with pre-trained MLLMs (Sec. 3.2), and (3) the development of a benchmark necessary for fine-grained brain perception (Sec. 4). + +# 3.1. Exploring Feature Spaces + +We explore four distinct feature spaces, drawing inspiration from prior MLLM research that leverage different feature representations. By aligning brain signals across these feature spaces, we aim to capture multi-granular representations in MLLMs, facilitating zero-shot multimodal brain decoding. The four feature space choices are illustrated in Fig. 1. + +Single Encoder. This represents using features from a single pre-trained vision encoder (e.g. CLIP [38]) for brain alignment, which is the most commonly used in MMLMs. + +Mixture of Encoders. Encoders trained on diverse tasks allow us to explore the distinct advantages of different vision experts. These include contrastive image-text learning + +![](images/d7063f93fb0d34da34b86c9124be5332e74ddbfc04b09f97e00259521a8d18c5.jpg) +(a) Single Encoder + +![](images/4fbd9e0bc2aec6b8c6909c82d5ddf698d59355c3bfb5d90a0a282a959983fb03.jpg) +(b) Mixture of Encoders +Figure 1. Feature Spaces. We aim to capture multi-granular representations for zero-shot multimodal brain decoding by aligning brain signals across different feature spaces in MLLM's visual component. The explored representative feature spaces include: (a) Single Encoder, which utilizes a single encoder to extract selected features; (b) Mixture of Encoders, using hybrid featured from different vision experts specialized in task-specific domains; (c) Aggregated Feature, fusing features from different layers within the same image encoder; and (d) Nested Features, downscaling visual representations in a hierarchical coarse-to-fine structure with different perceptual granularities. + +![](images/91f012f67e2d4caa75382e2cadf6f3fc76e64fc5b64867878874707aa97a0dea.jpg) +(c) Aggregated Feature + +![](images/dd5c07bcccffaadb95c3bc2fc875de2399b0624dd7fb86f15bd30a290456899b.jpg) +(d) Nested Features + +CLIP [38], self-supervised learning DINO [34], and object-centric models pre-trained for various tasks such as segmentation (SAM [22]), detection (EVA [11]), and recognition (Pix2Struct [24]). These task-specific vision encoders allow MLLMs to achieve optimal performance within respective pre-training domains [19, 45]. Aligning brain signals with features from these models allows for the capture of distinct facets of visual perception. Here, we explore the integrated features from CLIP [38] and DINO [34] encoders for simplicity while demonstrating the core concept of aligning brain signals with a mixture of vision experts. CLIP effectively encodes semantic information, while DINO excels in capturing spatial details such as object location and boundaries. + +Aggregated Feature. Existing MLLMs typically use features from a selected layer in the visual component as the visual representation input for LLMs. Features from different levels—such as shallow, middle, and deep layers—are recognized for capturing different characteristics of the image [57]. Therefore, by encapsulating information from multiple layers, we can enrich the visual input for the LLM. Specifically, we partition features from selected layers into several groups. Features within each group from adjacent layers are integrated along the channel dimension to reduce redundancy and mitigate high dimensionality. The features from each group are then concatenated with the features from the final layer before being fed into the connector. + +Nested Features. MLLMs typically project feature images from the vision encoder into a fixed-length set of visual tokens [29]. Longer token sequences generally yield better results but require more resources. Recent studies have found that the optimal token length is task-specific, laying the foundation for visual token pruning and merging to dynamically adjust the sequence length [5, 21]. Therefore, we downscale selected features using a set of downsampling factors and produce nested features. This approach offers a flexible and controllable method to encode visual content as nested sets of visual tokens with a hierarchical structure, from coarse to fine details—much like a Matryoshka doll [6, 23]. The + +nested features capture structured information across multiple coarse-to-fine levels, enabling adaptable token lengths that align with different perceptual granularities. + +# 3.2. Denoising for Multimodal Brain Alignment + +This section outlines training a brain encoder using features from Sec. 3.1 as ground truth. Given brain responses $s \in \mathbb{R}^{1 \times L_s}$ (with $L_{s}$ indicating the dimension of input brain data) and associated visual stimuli $v \in \mathbb{R}^{W \times H \times C}$ (representing image width, height, and channels), we train the brain encoder $\mathcal{B}$ to minimize the distance between brain features $\mathbf{b}$ and target image features $\mathbf{v}$ from the vision encoder $\mathcal{V}$ , aiming for a close approximation $\mathcal{B}(s) \approx \mathcal{V}(v)$ . Toward the objective, the most straightforward alignment is through element-wise feature reconstruction [53]: + +$$ +\mathcal {L} _ {\mathrm {R}} = \mathbb {E} _ {\mathbf {b} \sim \mathbf {B}, \mathbf {v} \sim \mathbf {V}} [ \| \mathcal {V} (v) - \mathcal {B} (s)) \| _ {2} ^ {2} ], \tag {1} +$$ + +where the brain encoder $\mathcal{B}$ learns the alignment between source brain space $\mathbf{B}$ and target image space $\mathbf{V}$ . + +Regression or Denoising. Regression using the Mean Squared Error (MSE, i.e., L2) loss directly learns a deterministic mapping from brain data points to those of images [53]. However, one downside of this deterministic objective is that it can lead to overfitting, as the model may memorize specific instances rather than generalizing well to unseen data—especially considering the data scarcity in brain modalities. This issue becomes particularly problematic when dealing with noisy or incomplete data, as is often the case with brain signals, making the model sensitive to minor variations in the input. Denoising [16, 47, 48], on the other hand, has been shown to be more effective than regression, as introducing noise into training data acts as an implicit form of data augmentation and regularization [26, 52]. The denoising process encourages the model to focus on the underlying data manifold rather than memorizing specific instance values [16, 40, 47, 48]. Furthermore, visual signals suffer from heavy spatial redundancy [14]; recent work has shown that such redundancy also exists in visual features [6]. + +![](images/464f58031a8542d0ff44551a0e2e5f7fddbcefe8c1d459acb61d12dc6baae7c6.jpg) +Figure 2. Training Procedure with Denoiser. + +Therefore, we introduce a masked denoising optimization objective as shown in Fig. 2. The training procedure of the $\pi$ -parameterized denoiser $\mathcal{J}_{\pi}$ follows a diffusion process: + +$$ +\mathcal {L} _ {\mathrm {D}} = \mathbb {E} _ {\varepsilon , t} \| \mathcal {J} _ {\pi} (\mathbf {v} _ {t}; \mathbf {b}, t) - \varepsilon \| ^ {2}. \tag {2} +$$ + +Here, $\varepsilon$ represents the ground truth noise sampled from the standard normal distribution $\mathcal{N}(\mathbf{0},\mathbf{I})$ . The noise-corruption is defined as $\mathbf{v}_t\coloneqq \sqrt{\bar{\alpha}_t}\mathbf{v} + \sqrt{1 - \bar{\alpha}_t\varepsilon}$ , where $\bar{\alpha}_t$ is a noise schedule [16, 33] at timestep $t$ , and $\mathbf{v} = \mathcal{V}(v)$ maps input images $v$ to its representation. The noise estimator $\mathcal{J}_{\pi}(\mathbf{v}_t;\mathbf{b},t)$ predicts the noise level in $\mathbf{v}_t$ at timestep $t$ , conditioned on brain predictions $\mathbf{b}$ . The positional embedding [26] is added to the masked features to indicate the positions that need to be predicted, with the loss computed only on the unknown ones [14, 26]. The final loss as shown in Fig. 2 is the weighted sum of the regression loss and the denoising loss. + +These features can be extracted from images on the fly during training or pre-calculated in advance to accelerate the process. During inference, the features predicted by the brain encoder, given brain signals as input, are fed into the remaining MLLM components (connector and LLM) for instruction-following tasks [29, 53], such as continuous dialogues, detailed descriptions, and complex reasoning. + +# 4. MG-BrainDUB + +To better evaluate the different choices of feature spaces in decoding fine-grained information, we propose a benchmark for detailed brain understanding. This involves curating high-quality evaluation datasets for two tasks: detailed captioning and salient question answering, each with corresponding metrics. We name this benchmark the Multi-Granularity Brain Descriptive Understanding Benchmark (MG-BrainDUB). + +# 4.1. Descriptive Caption Understanding + +Current brain captioning methods [12, 13, 49, 53] face two primary challenges. First, they rely on COCO [28] annota + +tions, which provide only short captions with limited classes, lacking the granularity and vocabulary needed for detailed descriptions. Second, existing evaluation metrics are inadequate for assessing fine-grained captions. Traditional rule-based metrics, such as BLEU [36], CIDEr [51], and METEOR [4], compute $n$ -gram matching scores between candidate and reference captions, making them highly sensitive to stylistic variations and leading to inconsistent evaluations. Model-based metrics, such as SPICE [3], suffer from outdated text encoders as backbones, while CLIP-based alternatives [15, 38] enforce a strict 77-token input limit, further restricting their effectiveness in evaluating detailed captions. + +We tackle both challenges through two key strategies. For detailed caption annotation, we construct ground truth descriptions using state-of-the-art MLLMs [29], followed by manual error correction, missing element addition, and hallucination removal through expert human intervention. For evaluation metrics, in addition to leveraging CAPTURE [9], a metric specifically designed for long captions, we introduce a structured evaluation framework that assesses captions based on precision, recall, and F1 scores of three core visual elements: objects, attributes, and relations [9, 32]. To achieve this, we develop an automated pipeline with two key steps: visual element extraction and structured matching. + +(a) Element Extraction. Given detailed captions, we first split them into separate sentences using NLTK toolkit [30] and extract key visual elements—objects, attributes, and their relations—using a T5-based factual parser [27, 39]. +(b) Structured Matching. The matching process aligns extracted objects, attributes, and relations in three steps: exact matching (identifying precise word matches), synonym matching (matching words with similar meanings), and semantic matching (computing cosine similarity of word embeddings for any remaining unmatched elements). + +Finally, we compute the precision, recall, and F1 scores for visual elements based on the matching results. These metrics capture semantic discrepancies: precision penalizes hallucinations and recall captures omissions, reducing confounds from pretrained generative model biases and offering a more accurate measure of genuine brain-based decoding. + +# 4.2. Question-Answering Annotation + +Previous studies have emphasized the multimodal evaluation for decoded salient objects, considering semantic selectivities in the higher visual cortex of the human brain [8, 18, 37]. Tiny objects in stimuli may not capture the subject's attention or relevant brain activities may not be effectively recorded during experiments [2, 53]. Thus, we further design evaluation focused on salient objects using annotated question-answering pairs. First, we extract salient objects using SAM [22] and input the cropped regions into the MLLM to identify objects, attributes, and relationships within each image using pipelines introduced in Sec. 4.1. We use SAM + +Table 1. Concept Localization. We present brain grounding [53] results derived directly from brain signals, reflecting spatial information encoded in brain activities. Image-based results are included as references, comprising visual grounding from Shikra [7] using associated visual stimuli and grounding performance based on reconstructed images from state-of-the-art visual decoding methods [35, 41, 53, 54]. We highlight the best, second-best, and third-best performance averaged across all four subjects. + +
MethodAll acc@0.5IoUSalient acc@0.5IoUSalient Creatures acc@0.5IoUSalient Objects acc@0.5IoUInconspicuous acc@0.5IoU
VisionShikra [7]51.9647.2262.9256.4466.7159.3458.7953.2738.2935.71
BrainDiffuser [35]17.4919.3427.1827.4638.7134.6314.6219.665.399.20
MindEye [41]15.3418.6523.8326.9629.2931.6417.8821.864.748.28
DREAM [54]16.2118.6526.5127.3534.4333.8517.8820.283.357.78
UMBRAE [53]16.8318.6927.1027.5534.1433.6519.4420.924.007.64
BrainUMBRAE-S [53]14.3118.1322.2825.8226.3229.5817.8821.724.378.53
UMBRAE [53]18.0320.6328.2229.3236.9335.6118.7422.475.329.79
VINDEX-S14.9218.5523.3826.2828.0030.2918.3521.924.378.91
VINDEX18.1020.8928.5029.7837.0036.2019.2522.785.119.80
+ +instead of the original COCO location annotations or Referring Expression Comprehension [7] with COCO classes [28] to extract such information as pseudolabels, since COCO often overlooks salient categories outside its 80 common classes. For instance, an image may feature a large church, but only two small clocks on it are annotated with their spatial locations and linguistic descriptions. For each identified object, we design questions such as "which description best fits the in the image" and "what is the relationship between and to evaluate the model's ability to recognize objects and their interactions within the image. Additionally, we include in-depth reasoning questions [29] that require step-by-step logical processing, based on GPT annotations [1, 43], to assess deeper reasoning capabilities. + +To mitigate the MLLM's tendency to default to affirmative responses such as "Yes, I can see it" rather than genuinely recognizing an object, we follow [32] and introduce an equivalent number of question-answer pairs for objects not present in the image as hallucination components. We then ask the model to select the correct description from multiple choices, where one option is a correct caption, and two are incorrect. The attribute annotations of fine-grained objects serve as the answers, while erroneous linguistic annotations are manually modified to ensure they contradict the actual visual content. We refer to this as salient QA (SQA) for complex reasoning and report the response accuracy. + +# 5. Experiments + +In our experiments, we demonstrate multimodal interaction with fMRI using MLLMs to tackle four tasks: concept localization (Sec. 5.2), concise captioning (Sec. 5.3), descriptive captioning (Sec. 5.4), and complex reasoning (Sec. 5.5). While the generated multimodal explanations have been shown to enhance reconstruction performance [43, 49, 53], we prioritize brain perception over reconstruction. + +# 5.1. Implementation Details + +Architecture. For the mixture of vision encoders, we use pre-trained CLIP [38] and DINOv2 [34] as visual encoders. The target image features are extracted from both transformer encoders and then interleaved to spatially mix visual tokens before being subsequently processed by the connector and LLM [50]. The feature dimension $D$ is 4,096 for 7B models and 5,120 for 13B models. For aggregated features, we utilize SigLIP [58] to extract visual features from multiple layers [57]. The connector first encapsulates multi-layer visual features and then maps the integrated visual features to the token space. For nested features, we use CLIP [38] to encode visual cues in a coarse-to-fine manner [6], where an image is represented as patch-based visual tokens. We create $M$ sets of tokens, with the visual tokens at the coarser level derived directly from those at the finer level. Specifically, starting from the initial visual tokens, we sequentially apply $2 \times 2$ pooling with a stride of 2, resulting in $12 \times 12$ , $6 \times 6$ , and $3 \times 3$ visual tokens. Finally, we apply $3 \times 3$ pooling to obtain the most condensed single visual token. Details on encoders and MLLMs can be found in the supplementary. + +Training Details. Our models are trained on a single A100 GPU. CLIP-224 [38] and DINO-224 [34] feature alignment with a batch size of 128 takes approximately 8 hours, CLIP-336 [38] with a batch size of 64 takes about 23 hours, while SigLIP-384 [58] with a batch size of 4 requires around 40 hours to converge. We use AdamW [31] as the optimizer with $\beta_{1} = 0.9$ , $\beta_{2} = 0.95$ , weight decay of 0.01, and apply the one-cycle strategy [46] with an initial learning rate of 3e-4 for the learning rate scheduler. Following visual decoding studies [41, 49, 54], we use the standard train and test splits for four subjects (sub01, sub02, sub05, sub07), each with 24,980 training samples and a shared 982 test samples. For evaluation, we report the average of the three repetitions of the same image in the test set, totaling 982 samples per subject. We implement the denoiser as an MLP with a default + +Table 2. Brain Captioning. We evaluate short captions on BrainHub [53] and detailed captions on MG-BrainDUB. Models with '-S' refers to those with cross-subject support but are trained in a single-subject setting for comparison. MindEye2 results are based on a model adapted to one subject using a pre-trained model that was trained on the remaining seven subjects with the full dataset. The mark† denotes zero-shot methods, indicating that no external captions were used during training. 'Shikra-image' refers to image captioning results from Shikra [7] using corresponding visual stimuli as input, acting as an approximate performance upper bound. Colors denote best, second-best, and third-best performance. Bold and underline represent the best and second best results under identical protocols. + +
MethodBLEU1BLEU2BLEU3BLEU4METEORROUGECIDErSPICECLIP-SRefCLIP-S
Concise Captioning (Sec. 5.3)
Shikra-img [7]82.3869.9058.6349.6635.6065.49161.4327.6280.6085.92
SDRecon [49]36.2117.117.723.4310.0325.1313.835.0261.0766.36
OneLLM [13]47.0426.9715.499.5113.5535.0522.996.2654.8061.28
BrainCap [12]55.9636.2122.7014.5116.6840.6941.309.0664.3169.90
NeuroVLA [43]57.1937.1723.7815.8518.6036.6749.5112.3965.49-
MindEye2 [42]54.8238.6026.4918.1617.5443.7755.7010.9767.5473.73
MEVOX [55]†58.5640.3628.0920.1119.2044.4754.3711.0564.2370.36
UMBRAE-S†57.6338.0225.0016.7618.4142.1551.9311.8366.4472.12
UMBRAE [53]†59.4440.4827.6619.0319.4543.7161.0612.7967.7873.54
VINDEX-S†57.9939.0026.0417.8318.5642.2753.9611.9266.9472.64
VINDEX†60.0040.7227.5718.9119.5143.7860.3212.8468.2673.88
Descriptive Captioning (Sec. 5.4)
NeuroVLA [43]38.9124.0215.2412.4118.4427.8342.5818.4156.16-
VINDEX
w/ SE22.7112.897.214.1113.9525.4314.058.4362.9067.72
w/ ME21.3911.866.313.3911.3117.606.046.4360.8565.96
w/ AF0.090.010.000.000.041.890.030.0153.1759.86
w/ NF120.1011.326.143.4014.6721.151.479.8560.3563.01
w/ NF950.7733.2521.1713.2018.2641.7246.5011.4565.4071.20
w/ NF3638.8723.8714.398.8014.7233.4031.078.9464.0269.80
w/ NF1443.742.221.270.743.4610.456.581.6957.1463.77
+ +depth of 1, a latent dimension (width) of 1024, and a weight multiplier of 1.0 for the denoising loss (indicating the same contribution as the regression loss during training). + +# 5.2. Concept Localization + +This section evaluates the model's ability for concept localization through the brain grounding task [53]. Similar to visual referring expression comprehension, brain grounding uses the prompt "Find the coordinates of $<\text{expr}>$ " and evaluates performance using accuracy and IoU metrics. The accuracy metric, acc@m, measures the percentage of correctly labeled instances with an IoU greater than the threshold $m$ . We assess grounding results derived directly from brain signals, with image-based results provided for reference. These include visual grounding using ground truth stimuli and reconstructed images from state-of-the-art (SOTA) visual decoding methods [35, 41, 54] via Shikra [7]. We provide a comparison with SOTAs and our methods in Tab. 1. + +The results demonstrate that our method VINDEX outperforms all baselines across all metrics in the four salient-based categories in BrainHub [53]. Notably, VINDEX surpasses both direct brain signal-based predictions from UM-BRAE [53] and image-based grounding results, which use reconstructed images from brain signals as input [35, 41, 53, + +54]. This emphasizes that directly extracting location information from the captured brain activities is more accurate than relying on reconstructed images. This observation suggests that reconstruction methods relying on image and text semantic encoding have inherent limitations in decoding spatial information. It also implies that incorporating grounding information in future models could help enhance the accuracy of object location prediction when reconstructing the visual stimuli from neural activations. + +# 5.3. Concise Captioning + +We evaluate short captions on BrainHub [53], demonstrating our method's performance on multimodal language tasks. Tab. 2 presents an evaluation of our brain captioning compared to state-of-the-art baselines, including three subject-specific methods SDRecon [49], BrainCap [12], OneLLM [13], and four cross-subject methods NeuroVLA [43], MindEye2 [42], MEVOX [55], and UMBRAE [53]. Our method, VINDEX, for concise captioning is built upon the same base MLLM Shikra [7] as in UMBRAE [53] for a fair comparison. MindEye2 is first pre-trained on seven subjects and then adapted to the remaining subject using all available data. NeuroVLA [43], UMBRAE [53], and VINDEX are trained on four subjects + +Table 3. Descriptive Captioning and Complex Reasoning. We evaluate detailed descriptions and complex reasoning with salient question-answering on MG-BrainDub under different model settings. 'Space' refers to the feature space elaborated in Sec. 3.1, while 'size' and 'encoder' indicate the size of ground truth images and the encoder used for feature extraction. Colors denote best, second-best, and third-best performance. Models marked with * correspond to those with the same settings as in Tab. 2. + +
SpaceModel SettingCAPTUREObjectAttributeRelationRuntime (s)SQA Acc (%)
SizeEncoderMLLMPRF1PRF1PRF1
SE*224CLIPLLaVA-1.5 7B0.435657.2249.7651.8837.9036.0835.7345.2540.7042.3214.0575.50
SE224CLIPLLaVA-1.5 13B0.473361.8152.9155.6040.6841.2139.3849.2845.0046.528.7077.92
SE224CLIPLLaVA-1.6 7B0.468451.0258.7853.5136.0154.0141.8041.0045.6342.7510.9176.80
SE224CLIPLLaVA-1.6 13B0.464550.4457.8352.8135.8954.0241.7240.4245.4442.3824.1878.69
ME*224CLIP, DINOLLaVA-1.5 13B0.304149.1734.3038.0726.1023.0923.3733.0826.8628.865.0652.66
AF*384SigLIPLLaVA-1.5 7B0.00581.891.071.270.150.110.130.010.000.005.6915.17
NF1*224CLIPLLaVA-1.5 7B0.487759.1957.1957.2442.7442.7641.1847.3246.7046.575.4582.59
NF9*224CLIPLLaVA-1.5 7B0.502162.4359.2659.7441.9743.2040.9550.5649.4949.545.3183.83
NF36*224CLIPLLaVA-1.5 7B0.467262.2354.8956.3740.2438.5637.5847.9044.4145.474.3366.35
NF144*224CLIPLLaVA-1.5 7B0.161452.8422.4728.738.675.556.3212.887.679.221.8331.35
NF1224CLIPLLaVA-1.6 7B0.435949.4751.3749.4335.6549.5940.2336.7538.9637.408.6265.58
NF1336CLIPLLaVA-1.5 7B0.412047.9946.7446.8943.0139.5039.5331.1531.9631.165.0662.27
+ +instead of eight. Caption results from Shikra [7] using associated visual stimuli for captioning act as an approximate upper bound. VINDEX achieves top performance across most metrics for concise brain captioning, even surpassing methods trained with more subject data [42] or external annotations [43]. The other brain-feature alignment methods UMBRAE [53] and MEVOX [55] ranked the second and the third. This highlights the effectiveness of the cross-subject brain feature alignment strategy, showcasing superior zero-shot performance without the need for external training data and generalization across subjects without requiring separate models or subject-specific parameters. + +NeuroVLA, MindEye2, UMBRAE, and VINDEX generate fluent, complete, and information-rich sentences, highlighting the advantages of using LLMs. Notably, VINDEX and UMBRAE achieve superior performance in a zero-shot setting, without relying on additional annotations, outperforming models trained with real captions from COCO [28] or pseudo captions from MLLMs [29]. This demonstrates that directly aligning brain signals with image features offers a better way than using external visual, spatial, or linguistic annotations. It preserves more accurate semantic and spatial cues decoded from brain signals and avoids factual errors or stylistic inconsistencies from external annotations. + +# 5.4. Descriptive Captioning + +We present a quantitative comparison of descriptive captioning results for NeuroVLA [43] and our method with different feature space choices: single encoder (SE), mixture of encoders (ME), aggregated feature (AF), and nested features $\mathrm{(NF_n)}$ where the token size $n$ is selected from $\{1, 9, 36, 144\}$ after downsampling. The latter Tab. 2 presents evaluations using regular caption metrics from [53]. Considering the limitations of traditional rule-based metrics (sensitive to writing style) and model-based metrics (unable to process long sequences), we evaluate descriptive captioning on MG-BrainDUB (Sec. 4.1). Results in Tab. 3 show that using + +$\mathrm{NF}_9$ (with 9 visual tokens) as supervision for training brain encoders achieves the best performance on most metrics. Surprisingly, using dense feature aggregation (AF) instead yields the worst results, and models trained with vanilla CLIP features [38] (SE setting at first four rows) remain inferior, even with larger and more sophisticated MLLMs (scaling from v1.5 to v1.6, 7B to 13B). + +Training with feature spaces from larger CLIP models (336 vs. 224) does not improve performance but significantly increases training time and computational costs. The brain encoder aligned with CLIP-ViT-L-224 features (corresponding to $16 \times 16$ tokens) generally transfers well to MLLMs designed with the CLIP-ViT-L-336 model ( $24 \times 24$ tokens) as the visual encoder, often yielding better results. AF performs the worst but with the largest token size (729 visual tokens for SigLIP-384 [58]). In contrast, aligning brain encoders with CLIP-224 requires only one-fifth of SigLIP-384's training time and one-third of CLIP-336's. Recall that NF downsamples the predicted CLIP features (the same features used in SE/ME) using pooling operations. Notably, 9-token alignment yields the best results, while 1-token alignment ranks second, suggesting that fewer tokens, though less expressive, align better with brain signals, which encode less information than real images. This also explains the inferior results for ME, where mixing brain-CLIP and brain-DINO features amplifies prediction imprecisions, leading to worse performance than using brain-CLIP features alone. + +The performance drop with increasing feature magnitude suggests that, while richer image features benefit visual perception tasks [6, 50, 57], brain signals capture relatively less comprehensive visual elements. This may due to subjects' selective attention [8, 18, 37] or limitations in capturing brain activities [2, 53]. We expect additional features to enhance the perception of subtler elements, but they mainly introduce unnecessary hallucinations and increase training costs, as brain signals predominantly encode basic visual elements. + +Results in Tab. 3 also indicate that traditional caption metrics [4, 36, 51] are not reliable when scaling to longer, more detailed captions. The rule-based metrics show that ME has higher scores than $\mathrm{NF}_1$ , even though $\mathrm{NF}_1$ predicts more accurate objects, attributes, and relations. CLIP-based metrics [15] also show deficiencies, as they present high similarities for all generated captions, even for AF, which in most cases lead to nonsensical outputs, such as garbled text and blank spaces. The inference time does not increase as expected with the model complexity and token length; instead, it exhibits irregularities. $\mathrm{NF}_9^*$ presents the best results across most metrics while maintaining low computation cost in terms of inference time and memory consumption. + +# 5.5. Complex Reasoning + +We evaluate the model's ability to perform complex reasoning on salient objects using Salient Question Answering (SQA) (Sec. 5.5), with accuracy reported in Tab. 3. Unlike previous sections focused on visual content, SQA involves in-depth reasoning that demands step-by-step logical processing to provide the right answer. This experiment highlights the model's ability to decode intricate characteristics of salient elements, meaning the prominent objects most likely to capture the subject's attention during brain activity capture [2], which have been shown to be better preserved in brain data [53]. From the table, $\mathrm{NF}_9^*$ achieves the best SQA performance across image sizes, visual encoders, and MLLM settings, likely benefiting from a balance between rich visual features and brain data constraints, consistent with detailed description results. In contrast, ME and AF sometimes produce meaningless outputs due to improper feature-brain signal alignment, caused by the added complexity of features during training. They also suffer from hallucinations, generating non-existent objects and misrepresenting attributes and relationships, resulting in accuracy rates at or below random guessing. + +# 6. Ablation Studies + +Model-Agnostic Evaluation. VINDEX is a versatile, model-agnostic brain decoding framework that can be easily adapted to various subjects, image sizes, visual encoders, MLLM settings, input prompts, and downstream tasks. Due to the architectural consistency of LLaVA-based variants [6, 7, 29, 50, 57], our trained brain encoders can be directly integrated into different MLLMs without requiring additional modifications. Specifically, for model setups presented in Tab. 3, we only need to train four brain encoders using CLIP-224, CLIP-336, DINO-224, and SigLIP-384. The same brain encoder trained with CLIP-224 can be used across SE, ME, and NF (except for the last one with a 336 image size). This brain-CLIP features are then passed to the remaining MLLMs components, including connectors and LLMs, with task prompts supported by the base MLLM for + +Table 4. Ablation Study on Denoiser. CAPTURE [9] is reported across architectural depth $d$ , width $w$ , and weight multiplier $\beta$ . + +
SetupS0S1S2S3S4S5S6S7
β-1.01.01.01.00.51.52.0
d-1231333
w-5121,0241,0241,0241,0241,0241024
C0.41790.43080.42570.41830.43560.41930.41970.4096
+ +multimodal interactions, such as interactive dialogue, brain captioning, complex reasoning, and concept localization. The brain encoder aligned with CLIP-224 features adapts well to LLaVA-1.6-7B/13B models, which are designed to use CLIP-336 as the visual encoder, leading to improved performance. This versatility enables our method to efficiently align brain signals with multiple MLLMs at minimal cost. + +Denoising Optimization Objective. We conduct ablation studies on the denoiser in Tab. 4 to evaluate the effectiveness of the denoising objective over vanilla regression across different settings, focusing on architectural depth $d$ , width $w$ , and the weight multiplier $\beta$ . Here, 'Sn' represents different configurations of specific depth, width, and weight multiplier; S0 refers to the base setup with only the regression loss, without the denoising loss. The results are reported as CAPTURE scores [9] (C) for detailed descriptions across each setup. Total training hours for 300 epochs across all settings is approximately eight hours, with no notable variations. We found that a simple one-depth MLP with 1,024 model channels yields the best results. Results show that training with the denoising loss setup incurs no extra cost but leads to improved performance and more stable training (further analysis in the supplementary material). + +# 7. Conclusion + +In this paper, we propose a zero-shot multimodal brain decoding method that explores representative feature spaces from the visual components within MLLMs. We elucidate different vision feature space choices and introduce a denoiser-based strategy to enhance brain alignment with pre-trained models, improving the accuracy and robustness of brain-encoded feature representations. Our method facilitates efficient and flexible decoding across multiple levels of granularity, eliminating the need for additional textual or spatial annotations during training. Furthermore, we present MG-BrainDUB, a benchmark that provides a more reliable and comprehensive evaluation for multigranular multimodal brain decoding. It comprises two core tasks—detailed description generation and salient question answering—along with a novel metric that prioritizes key visual elements including objects, attributes, and relationships, offering a robust framework for assessing fine-grained brain perception. + +Acknowledgements. This work was supported by a UKRI Future Leaders Fellowship [grant number G104084] and an Imminent Research Grant. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5 +[2] Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116-126, 2022. 4, 7, 8 +[3] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 1, 4 +[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 1, 4, 8 +[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3 +[6] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. In ICLR, 2025. 2, 3, 5, 7, 8 +[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8 +[8] Robert Desimone, Thomas D Albright, Charles G Gross, and Charles Bruce. Stimulus-selective properties of inferior temporal neurons in the macaque. Journal of Neuroscience, 4 (8):2051-2062, 1984. 4, 7 +[9] Hongyuan Dong, Jiawen Li, Bohong Wu, Jiacong Wang, Yuan Zhang, and Haoyuan Guo. Benchmarking and improving detail image caption. arXiv preprint arXiv:2405.19092, 2024. 4, 8 +[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2 +[11] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 3 +[12] Matteo Ferrante, Furkan Ozcelik, Tommaso Boccato, Rufin VanRullen, and Nicola Toschi. Brain captioning: Decoding human brain activity into images and text. arXiv preprint arXiv:2305.11560, 2023. 4, 6 +[13] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In CVPR, 2024. 1, 2, 4, 6 +[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 3, 4 +[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, + +and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 1, 4, 8 +[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3, 4 +[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, pages 4651-4664. PMLR, 2021. 2 +[18] Nancy Kanwisher, Josh McDermott, and Marvin M Chun. The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of neuroscience, 17 (11):4302-4311, 1997. 4, 7 +[19] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. In ECCV, pages 113-132. Springer, 2024. 2, 3 +[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, pages 8107-8116, 2020. 2 +[21] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In WACV, pages 1383-1392, 2024. 3 +[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3, 4 +[23] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning. In NeurIPS, pages 30233-30249, 2022. 3 +[24] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understanding. In ICML, 2023. 3 +[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 1 +[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS, 37:56424-56445, 2024. 3, 4 +[27] Zhuang Li, Yuyang Chai, Terry Yue Zhuo, Lizhen Qu, Gholamreza Haffari, Fei Li, Donghong Ji, and Quan Hung Tran. Factual: A benchmark for faithful and consistent textual scene graph parsing. In ACL, 2023. 4 +[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 4, 5, 7 +[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3, 4, 5, 7, 8 + +[30] Edward Loper and Steven Bird. Nltk: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, 2004. 4 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5 +[32] Fan Lu, Wei Wu, Kecheng Zheng, Shuailei Ma, Biao Gong, Jiawei Liu, Wei Zhai, Yang Cao, Yujun Shen, and Zheng-Jun Zha. Benchmarking large vision-language models via directed scene graph for comprehensive image captioning. In CVPR, 2025. 4, 5 +[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 4 +[34] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2023. 3, 5 +[35] Furkan Ozcelik and Rufin VanRullen. Brain-Diffuser: Natural scene reconstruction from fMRI signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023. 1, 2, 5, 6 +[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 1, 4, 8 +[37] Aina Puce, Truett Allison, Maryam Asgari, John C Gore, and Gregory McCarthy. Differential sensitivity of human visual cortex to faces, letterstrings, and textures: a functional magnetic resonance imaging study. Journal of neuroscience, 16(16):5205-5215, 1996. 4, 7 +[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7 +[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 4 +[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3 +[41] Paul S Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Ethan Cohen, Aidan J Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. In NeurIPS, 2023. 1, 2, 5, 6 +[42] Paul S Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, and Tanishq Mathew Abraham. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML, 2024. 1, 2, 6, 7 +[43] Guobin Shen, Dongcheng Zhao, Xiang He, Linghao Feng, Yiting Dong, Jihang Wang, Qian Zhang, and Yi Zeng. Neurovision to language: Enhancing brain recording-based visual + +reconstruction and language interaction. In NeurIPS, pages 98083-98110, 2024. 1, 2, 5, 6, 7 +[44] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In NeurIPS, pages 42048-42070, 2025. 2 +[45] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, Bryan Catanzaro, Andrew Tao, Jan Kautz, Zhiding Yu, and Guilin Liu. Eagle: Exploring the design space for multimodal llms with mixture of encoders. In ICLR, 2025. 3 +[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5 +[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3 +[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3 +[49] Yu Takagi and Shinji Nishimoto. Improving visual image reconstruction from human brain activity using latent diffusion models via multiple decoded inputs. arXiv preprint arXiv:2306.11536, 2023. 1, 2, 4, 5, 6 +[50] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, pages 9568-9578, 2024. 2, 5, 7, 8 +[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 1, 4, 8 +[52] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Zheng Ge, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In ICLR, 2025. 3 +[53] Weihao Xia, Raoul de Charette, Cengiz Oztireli, and JingHao Xue. Umbrae: Unified multimodal brain decoding. In ECCV, pages 242-259, 2024. 1, 2, 3, 4, 5, 6, 7, 8 +[54] Weihao Xia, Raoul de Charette, Cengiz Özireli, and Jing-Hao Xue. Dream: Visual decoding from reversing human visual system. In WACV, pages 8226-8235, 2024. 1, 2, 5, 6 +[55] Weihao Xia and Cengiz Öztireli. Mevox: Multi-task vision experts for brain captioning. In CVPR Workshop, 2025. 2, 6, 7 +[56] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 2 +[57] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms. In NeurIPS, pages 33108-33140, 2025. 2, 3, 5, 7, 8 +[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 7 +[59] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In NeurIPS, 2024. 2 \ No newline at end of file diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/images.zip b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a7e282ff072177adf250fa7a3793e6954a484f2e --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:075a3d372273e62f401f8d839fefca46ade2e20adb06bd515ea548e15dc7e113 +size 401275 diff --git a/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/layout.json b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..abb68f68f64de36b11a092c208640e169880a678 --- /dev/null +++ b/2025/Exploring The Visual Feature Space for Multimodal Neural Decoding/layout.json @@ -0,0 +1,7126 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 103, + 520, + 121 + ], + "type": "text", + "content": "Exploring The Visual Feature Space for Multimodal Neural Decoding" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 225, + 142, + 386, + 159 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 225, + 142, + 386, + 159 + ], + "spans": [ + { + "bbox": [ + 225, + 142, + 386, + 159 + ], + "type": "text", + "content": "Weihao Xia " + }, + { + "bbox": [ + 225, + 142, + 386, + 159 + ], + "type": "inline_equation", + "content": "\\boxtimes" + }, + { + "bbox": [ + 225, + 142, + 386, + 159 + ], + "type": "text", + "content": " Cengiz Oztireli" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 243, + 163, + 364, + 178 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 243, + 163, + 364, + 178 + ], + "spans": [ + { + "bbox": [ + 243, + 163, + 364, + 178 + ], + "type": "text", + "content": "University of Cambridge" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 217, + 182, + 392, + 194 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 217, + 182, + 392, + 194 + ], + "spans": [ + { + "bbox": [ + 217, + 182, + 392, + 194 + ], + "type": "text", + "content": "https://weihaox.github.io/VINDEX" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "spans": [ + { + "bbox": [ + 151, + 209, + 200, + 222 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 235, + 297, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 297, + 486 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 297, + 486 + ], + "type": "text", + "content": "The intrication of brain signals drives research that leverages multimodal AI to align brain modalities with visual and textual data for explainable descriptions. However, most existing studies are limited to coarse interpretations, lacking essential details on object descriptions, locations, attributes, and their relationships. This leads to imprecise and ambiguous reconstructions when using such cues for visual decoding. To address this, we analyze different choices of vision feature spaces from pre-trained visual components within Multimodal Large Language Models (MLLMs) and introduce a zero-shot multimodal brain decoding method that interacts with these models to decode across multiple levels of granularities. To assess a model's ability to decode fine details from brain signals, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub). This benchmark includes two key tasks: detailed descriptions and salient question-answering, with metrics highlighting key visual elements like objects, attributes, and relationships. Our approach enhances neural decoding precision and supports more accurate neuro-decoding applications. Code is available at https://github.com/weihaox/VINDEX." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 135, + 516 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 522, + 296, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 296, + 629 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 296, + 629 + ], + "type": "text", + "content": "Human brain signals are inherently complex and not directly interpretable, posing a significant challenge for decoding brain activity. To address this, recent studies have focused on translating brain signals associated visual stimuli into multimodal explanations, such as textual descriptions [13, 42], depth information [54], bounding boxes [53], or direct image reconstruction [35, 41, 42]. Although these efforts have made notable strides, achieving a fine-grained understanding of brain modalities remains an unresolved challenge." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 296, + 714 + ], + "type": "text", + "content": "One line of research [42, 49, 53] has attempted to translate brain activations from visual stimuli into textual captions. However, existing methods typically generate one-sentence descriptions with limited details, failing to capture the full spectrum of visual perception. The most straightforward approach [43] is to use off-the-shelf MLLMs [25, 29] to generate detailed descriptions from paired images for each" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "text", + "content": "brain recording, then train a model to decode these pseudo-annotations from brain signals. However, this practice is computationally inefficient and may introduce bias or errors, making it impractical for generalizable brain decoding. The other key issue is the evaluation of brain-decoded detailed captions, which can be hundreds of words long, rendering conventional metrics unsuitable. Specifically, traditional captioning metrics [4, 36, 51] which mostly rely on " + }, + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 211, + 555, + 390 + ], + "type": "text", + "content": "-gram matching between candidate and reference captions, are highly sensitive to stylistic variations, leading to inconsistent evaluations. Model-based metrics either use outdated text encoders [3] or struggle with long text due to token length limitations [15, 38]. While concurrent studies [43] attempt to generate detailed captions for visual decoding, they lack reliable evaluations of caption quality and accuracy." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 391, + 556, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 391, + 556, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 391, + 556, + 629 + ], + "type": "text", + "content": "In this paper, we address the two critical issues. For the method, we investigate the exploitation of various feature spaces from the visual components within MLLMs to enable fine-grained multimodal explanations from neural activations. We present VINDEX (Visual experts for multimodal Neural DEcoding eXploration), which leverages diverse feature spaces for multimodal neural decoding. We further introduce a denoiser-based alignment strategy that efficiently maps brain tokens to image tokens, allowing for integration into pre-trained MLLMs. This approach facilitates decoding across multiple levels of granularities without the need for additional textual or spatial annotations. For evaluation, we propose the Multi-Granularity Brain Detail Understanding Benchmark (MG-BrainDub) to discern a model's ability to decode fine-grained details from brain signals. MG-BrainDub consists of two key tasks: detailed descriptions and salient question-answering, with metrics focusing on key visual elements, including objects, attributes, and relationships, derived from both candidate and ground truth captions. Our contributions are summarized as follows:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 630, + 553, + 714 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 314, + 630, + 553, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 630, + 553, + 666 + ], + "spans": [ + { + "bbox": [ + 314, + 630, + 553, + 666 + ], + "type": "text", + "content": "- Through systematic analysis, we exploit representative feature spaces of visual components within MLLMs for brain signal alignment and multimodal brain interaction." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 666, + 553, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 553, + 701 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 553, + 701 + ], + "type": "text", + "content": "- We introduce a zero-shot multimodal brain decoding method for multi-granular decoding without the need for additional textual or spatial annotations during training." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 702, + 553, + 714 + ], + "type": "text", + "content": "- We construct a benchmark with metrics for fine-grained" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4370" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 64, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 64, + 72, + 295, + 97 + ], + "type": "text", + "content": "brain interpretation, evaluating a model's ability to decode detailed descriptions and perform complex reasoning." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 109, + 142, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 109, + 142, + 121 + ], + "spans": [ + { + "bbox": [ + 55, + 109, + 142, + 121 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 297, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 297, + 443 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 297, + 443 + ], + "type": "text", + "content": "Multimodal Brain Perception. Recent advancements in decoding multimodal cues from brain signals, facilitated by MLLMs, have achieved unprecedented performance [35, 41, 49, 54]. These methods typically map brain responses, captured through functional magnetic resonance imaging (fMRI), to more commonly used modalities that can be processed by pre-trained models for subsequent information extraction [20, 40, 56]. For instance, Takagi and Nishimoto [49] connect fMRI data with CLIP text embeddings and the latent space of Stable Diffusion (SD) [40] using ridge regression to reconstruct visual stimuli. Xia et al. [54] extract semantic, depth, and color features from fMRI data to reconstruct images using a depth- and color-conditioned SD model. Several methods [13, 42, 49, 54] aim to create readable multimodal explanations of visual stimuli, such as textual description or depth estimation, but often rely on corresponding external supervision to guide the multimodal training. In contrast, UMBRAE [53] has recently demonstrated the ability to directly decode textual descriptions and bounding boxes from fMRI data using only image-fMRI pairs during training, without requiring additional annotations. This method aligns image features with visual encoders and integrates them into MLLMs. Building on this foundation, our proposed approach aims to further elucidate feature space representations, extracting more granular information while introducing metrics for evaluation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 446, + 298, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 446, + 298, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 446, + 298, + 699 + ], + "type": "text", + "content": "Multimodal Large Language Models. Expanding Large Language Models (LLMs) to incorporate additional modalities, such as images, has recently attracted significant interest [10, 29, 50, 57]. These models typically comprise three components: a frozen image encoder, a trainable connector, and a frozen or finetuned LLM. The connector bridges the gap from the image feature space to the LLM token space, which can be implemented as a linear layer [7], a multilayer perceptron (MLP) [29], or a lightweight transformer [17]. Recent studies have shown that using pre-trained CLIP [38] as the visual component can lead to compromised performance in visual tasks, particularly in grounding capabilities [7, 50]. To overcome this limitation, researchers have explored ways to more effectively utilize the visual component to enhance perception across diverse scenarios. This includes employing a mixture of vision encoders [19, 44, 50, 59] pretrained on various tasks as expert models and investigating the feature spaces [6, 57] within CLIP. Drawing inspiration from these efforts, we exploit brain alignment with feature spaces in MLLM's visual component for multi-granular representation to facilitate zero-shot multimodal brain decoding." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 296, + 714 + ], + "type": "text", + "content": "Multimodal-Brain Alignment. The prevailing practice" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 72, + 556, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 556, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 556, + 312 + ], + "type": "text", + "content": "for brain alignment is to map neural modalities into a shared multimodal space [13, 35, 41, 53], which can be categorized into generative alignment and discriminative alignment. The first category, exemplified by OneLLM [13], employs generative training to learn multimodal alignment, connecting multimodal inputs, including brain signals, with an LLM. Methods in the second category align brain signals within a pre-trained embedding space, such as CLIP [38], using techniques like linear regression [35, 49], contrastive learning [54], diffusion priors [41], or feature reconstruction [53]. MEVOX [55] ensembles multiple task-specific experts to capture distinct aspects of brain perception. The brain encoder learns to align with omni-contextual explanations. Most of these alignment strategies require per-subject training or subject-specific annotations, which can lead to scalability issues in practice. While UMBRAE [53] performs zero-shot multimodal brain decoding, its regression-based alignment is deterministic. We analyze the limitations of this strategy and propose incorporating a novel denoising optimization objective to enhance the brain-feature alignment." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 324, + 370, + 335 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 324, + 370, + 335 + ], + "spans": [ + { + "bbox": [ + 314, + 324, + 370, + 335 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 345, + 556, + 537 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 556, + 537 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 556, + 537 + ], + "type": "text", + "content": "Current MLLMs [7, 29] comprise three components: a pretrained vision encoder, a pre-trained LLM, and a connector that bridges vision and language models. Given significant discrepancies in data scale and challenges associated with acquiring large-scale brain datasets, we argue that aligning brain signals with pre-trained visual components, rather than training from scratch [13, 43], is a more effective and efficient approach. This not only alleviates the need to construct extensive text for training MLLMs but also enables zero-shot capability in decoding multimodal cues, allowing for versatile usage across different MLLMs. Thus, we present the rationale behind our method design in three key aspects: (1) the selection of vision feature spaces (Sec. 3.1), (2) the alignment strategy to connect activations with pre-trained MLLMs (Sec. 3.2), and (3) the development of a benchmark necessary for fine-grained brain perception (Sec. 4)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 545, + 457, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 545, + 457, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 545, + 457, + 558 + ], + "type": "text", + "content": "3.1. Exploring Feature Spaces" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 563, + 556, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 556, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 556, + 635 + ], + "type": "text", + "content": "We explore four distinct feature spaces, drawing inspiration from prior MLLM research that leverage different feature representations. By aligning brain signals across these feature spaces, we aim to capture multi-granular representations in MLLMs, facilitating zero-shot multimodal brain decoding. The four feature space choices are illustrated in Fig. 1." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "spans": [ + { + "bbox": [ + 313, + 639, + 555, + 674 + ], + "type": "text", + "content": "Single Encoder. This represents using features from a single pre-trained vision encoder (e.g. CLIP [38]) for brain alignment, which is the most commonly used in MMLMs." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 556, + 714 + ], + "type": "text", + "content": "Mixture of Encoders. Encoders trained on diverse tasks allow us to explore the distinct advantages of different vision experts. These include contrastive image-text learning" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "4371" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 88, + 98, + 153, + 134 + ], + "blocks": [ + { + "bbox": [ + 88, + 98, + 153, + 134 + ], + "lines": [ + { + "bbox": [ + 88, + 98, + 153, + 134 + ], + "spans": [ + { + "bbox": [ + 88, + 98, + 153, + 134 + ], + "type": "image", + "image_path": "d7063f93fb0d34da34b86c9124be5332e74ddbfc04b09f97e00259521a8d18c5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 85, + 151, + 151, + 161 + ], + "lines": [ + { + "bbox": [ + 85, + 151, + 151, + 161 + ], + "spans": [ + { + "bbox": [ + 85, + 151, + 151, + 161 + ], + "type": "text", + "content": "(a) Single Encoder" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 170, + 76, + 253, + 148 + ], + "blocks": [ + { + "bbox": [ + 170, + 76, + 253, + 148 + ], + "lines": [ + { + "bbox": [ + 170, + 76, + 253, + 148 + ], + "spans": [ + { + "bbox": [ + 170, + 76, + 253, + 148 + ], + "type": "image", + "image_path": "4fbd9e0bc2aec6b8c6909c82d5ddf698d59355c3bfb5d90a0a282a959983fb03.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 168, + 151, + 252, + 161 + ], + "lines": [ + { + "bbox": [ + 168, + 151, + 252, + 161 + ], + "spans": [ + { + "bbox": [ + 168, + 151, + 252, + 161 + ], + "type": "text", + "content": "(b) Mixture of Encoders" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 54, + 170, + 555, + 226 + ], + "lines": [ + { + "bbox": [ + 54, + 170, + 555, + 226 + ], + "spans": [ + { + "bbox": [ + 54, + 170, + 555, + 226 + ], + "type": "text", + "content": "Figure 1. Feature Spaces. We aim to capture multi-granular representations for zero-shot multimodal brain decoding by aligning brain signals across different feature spaces in MLLM's visual component. The explored representative feature spaces include: (a) Single Encoder, which utilizes a single encoder to extract selected features; (b) Mixture of Encoders, using hybrid featured from different vision experts specialized in task-specific domains; (c) Aggregated Feature, fusing features from different layers within the same image encoder; and (d) Nested Features, downscaling visual representations in a hierarchical coarse-to-fine structure with different perceptual granularities." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 268, + 76, + 399, + 147 + ], + "blocks": [ + { + "bbox": [ + 268, + 76, + 399, + 147 + ], + "lines": [ + { + "bbox": [ + 268, + 76, + 399, + 147 + ], + "spans": [ + { + "bbox": [ + 268, + 76, + 399, + 147 + ], + "type": "image", + "image_path": "91f012f67e2d4caa75382e2cadf6f3fc76e64fc5b64867878874707aa97a0dea.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 151, + 372, + 161 + ], + "lines": [ + { + "bbox": [ + 291, + 151, + 372, + 161 + ], + "spans": [ + { + "bbox": [ + 291, + 151, + 372, + 161 + ], + "type": "text", + "content": "(c) Aggregated Feature" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 417, + 96, + 527, + 140 + ], + "blocks": [ + { + "bbox": [ + 417, + 96, + 527, + 140 + ], + "lines": [ + { + "bbox": [ + 417, + 96, + 527, + 140 + ], + "spans": [ + { + "bbox": [ + 417, + 96, + 527, + 140 + ], + "type": "image", + "image_path": "dd5c07bcccffaadb95c3bc2fc875de2399b0624dd7fb86f15bd30a290456899b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 434, + 150, + 503, + 160 + ], + "lines": [ + { + "bbox": [ + 434, + 150, + 503, + 160 + ], + "spans": [ + { + "bbox": [ + 434, + 150, + 503, + 160 + ], + "type": "text", + "content": "(d) Nested Features" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 247, + 297, + 402 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 247, + 297, + 402 + ], + "spans": [ + { + "bbox": [ + 54, + 247, + 297, + 402 + ], + "type": "text", + "content": "CLIP [38], self-supervised learning DINO [34], and object-centric models pre-trained for various tasks such as segmentation (SAM [22]), detection (EVA [11]), and recognition (Pix2Struct [24]). These task-specific vision encoders allow MLLMs to achieve optimal performance within respective pre-training domains [19, 45]. Aligning brain signals with features from these models allows for the capture of distinct facets of visual perception. Here, we explore the integrated features from CLIP [38] and DINO [34] encoders for simplicity while demonstrating the core concept of aligning brain signals with a mixture of vision experts. CLIP effectively encodes semantic information, while DINO excels in capturing spatial details such as object location and boundaries." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 564 + ], + "type": "text", + "content": "Aggregated Feature. Existing MLLMs typically use features from a selected layer in the visual component as the visual representation input for LLMs. Features from different levels—such as shallow, middle, and deep layers—are recognized for capturing different characteristics of the image [57]. Therefore, by encapsulating information from multiple layers, we can enrich the visual input for the LLM. Specifically, we partition features from selected layers into several groups. Features within each group from adjacent layers are integrated along the channel dimension to reduce redundancy and mitigate high dimensionality. The features from each group are then concatenated with the features from the final layer before being fed into the connector." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 297, + 714 + ], + "type": "text", + "content": "Nested Features. MLLMs typically project feature images from the vision encoder into a fixed-length set of visual tokens [29]. Longer token sequences generally yield better results but require more resources. Recent studies have found that the optimal token length is task-specific, laying the foundation for visual token pruning and merging to dynamically adjust the sequence length [5, 21]. Therefore, we downscale selected features using a set of downsampling factors and produce nested features. This approach offers a flexible and controllable method to encode visual content as nested sets of visual tokens with a hierarchical structure, from coarse to fine details—much like a Matryoshka doll [6, 23]. The" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 247, + 555, + 283 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 283 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 283 + ], + "type": "text", + "content": "nested features capture structured information across multiple coarse-to-fine levels, enabling adaptable token lengths that align with different perceptual granularities." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 290, + 539, + 303 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 290, + 539, + 303 + ], + "spans": [ + { + "bbox": [ + 313, + 290, + 539, + 303 + ], + "type": "text", + "content": "3.2. Denoising for Multimodal Brain Alignment" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": "This section outlines training a brain encoder using features from Sec. 3.1 as ground truth. Given brain responses " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "s \\in \\mathbb{R}^{1 \\times L_s}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " (with " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "L_{s}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " indicating the dimension of input brain data) and associated visual stimuli " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "v \\in \\mathbb{R}^{W \\times H \\times C}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " (representing image width, height, and channels), we train the brain encoder " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " to minimize the distance between brain features " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " and target image features " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\mathbf{v}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": " from the vision encoder " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{V}" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": ", aiming for a close approximation " + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "inline_equation", + "content": "\\mathcal{B}(s) \\approx \\mathcal{V}(v)" + }, + { + "bbox": [ + 313, + 308, + 556, + 429 + ], + "type": "text", + "content": ". Toward the objective, the most straightforward alignment is through element-wise feature reconstruction [53]:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 357, + 436, + 555, + 452 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 357, + 436, + 555, + 452 + ], + "spans": [ + { + "bbox": [ + 357, + 436, + 555, + 452 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {R}} = \\mathbb {E} _ {\\mathbf {b} \\sim \\mathbf {B}, \\mathbf {v} \\sim \\mathbf {V}} [ \\| \\mathcal {V} (v) - \\mathcal {B} (s)) \\| _ {2} ^ {2} ], \\tag {1}", + "image_path": "223582881b2d6c20cf9ab1ec72d0b062ac8a8b9c3d676275ab05ec4c2330d790.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "text", + "content": "where the brain encoder " + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathcal{B}" + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "text", + "content": " learns the alignment between source brain space " + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{B}" + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "text", + "content": " and target image space " + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "inline_equation", + "content": "\\mathbf{V}" + }, + { + "bbox": [ + 313, + 460, + 554, + 484 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 486, + 556, + 713 + ], + "type": "text", + "content": "Regression or Denoising. Regression using the Mean Squared Error (MSE, i.e., L2) loss directly learns a deterministic mapping from brain data points to those of images [53]. However, one downside of this deterministic objective is that it can lead to overfitting, as the model may memorize specific instances rather than generalizing well to unseen data—especially considering the data scarcity in brain modalities. This issue becomes particularly problematic when dealing with noisy or incomplete data, as is often the case with brain signals, making the model sensitive to minor variations in the input. Denoising [16, 47, 48], on the other hand, has been shown to be more effective than regression, as introducing noise into training data acts as an implicit form of data augmentation and regularization [26, 52]. The denoising process encourages the model to focus on the underlying data manifold rather than memorizing specific instance values [16, 40, 47, 48]. Furthermore, visual signals suffer from heavy spatial redundancy [14]; recent work has shown that such redundancy also exists in visual features [6]." + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4372" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 83, + 74, + 265, + 240 + ], + "blocks": [ + { + "bbox": [ + 83, + 74, + 265, + 240 + ], + "lines": [ + { + "bbox": [ + 83, + 74, + 265, + 240 + ], + "spans": [ + { + "bbox": [ + 83, + 74, + 265, + 240 + ], + "type": "image", + "image_path": "464f58031a8542d0ff44551a0e2e5f7fddbcefe8c1d459acb61d12dc6baae7c6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 90, + 242, + 261, + 254 + ], + "lines": [ + { + "bbox": [ + 90, + 242, + 261, + 254 + ], + "spans": [ + { + "bbox": [ + 90, + 242, + 261, + 254 + ], + "type": "text", + "content": "Figure 2. Training Procedure with Denoiser." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "spans": [ + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "content": "Therefore, we introduce a masked denoising optimization objective as shown in Fig. 2. The training procedure of the " + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "inline_equation", + "content": "\\pi" + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "content": "-parameterized denoiser " + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_{\\pi}" + }, + { + "bbox": [ + 55, + 273, + 295, + 309 + ], + "type": "text", + "content": " follows a diffusion process:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 109, + 315, + 296, + 330 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 315, + 296, + 330 + ], + "spans": [ + { + "bbox": [ + 109, + 315, + 296, + 330 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {D}} = \\mathbb {E} _ {\\varepsilon , t} \\| \\mathcal {J} _ {\\pi} (\\mathbf {v} _ {t}; \\mathbf {b}, t) - \\varepsilon \\| ^ {2}. \\tag {2}", + "image_path": "4b3a9269418ef1301d72682da79dd1745335004ebfae3f3a7a64bdf2a86c214c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "spans": [ + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\varepsilon" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " represents the ground truth noise sampled from the standard normal distribution " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(\\mathbf{0},\\mathbf{I})" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": ". The noise-corruption is defined as " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t\\coloneqq \\sqrt{\\bar{\\alpha}_t}\\mathbf{v} + \\sqrt{1 - \\bar{\\alpha}_t\\varepsilon}" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\bar{\\alpha}_t" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " is a noise schedule [16, 33] at timestep " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{v} = \\mathcal{V}(v)" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " maps input images " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " to its representation. The noise estimator " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathcal{J}_{\\pi}(\\mathbf{v}_t;\\mathbf{b},t)" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " predicts the noise level in " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{v}_t" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": " at timestep " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "t" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": ", conditioned on brain predictions " + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "inline_equation", + "content": "\\mathbf{b}" + }, + { + "bbox": [ + 54, + 336, + 296, + 467 + ], + "type": "text", + "content": ". The positional embedding [26] is added to the masked features to indicate the positions that need to be predicted, with the loss computed only on the unknown ones [14, 26]. The final loss as shown in Fig. 2 is the weighted sum of the regression loss and the denoising loss." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 468, + 296, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 296, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 296, + 552 + ], + "type": "text", + "content": "These features can be extracted from images on the fly during training or pre-calculated in advance to accelerate the process. During inference, the features predicted by the brain encoder, given brain signals as input, are fed into the remaining MLLM components (connector and LLM) for instruction-following tasks [29, 53], such as continuous dialogues, detailed descriptions, and complex reasoning." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 562, + 150, + 574 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 562, + 150, + 574 + ], + "spans": [ + { + "bbox": [ + 55, + 562, + 150, + 574 + ], + "type": "text", + "content": "4. MG-BrainDUB" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 666 + ], + "type": "text", + "content": "To better evaluate the different choices of feature spaces in decoding fine-grained information, we propose a benchmark for detailed brain understanding. This involves curating high-quality evaluation datasets for two tasks: detailed captioning and salient question answering, each with corresponding metrics. We name this benchmark the Multi-Granularity Brain Descriptive Understanding Benchmark (MG-BrainDUB)." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 244, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 244, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 244, + 685 + ], + "type": "text", + "content": "4.1. Descriptive Caption Understanding" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Current brain captioning methods [12, 13, 49, 53] face two primary challenges. First, they rely on COCO [28] annota" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "tions, which provide only short captions with limited classes, lacking the granularity and vocabulary needed for detailed descriptions. Second, existing evaluation metrics are inadequate for assessing fine-grained captions. Traditional rule-based metrics, such as BLEU [36], CIDEr [51], and METEOR [4], compute " + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 72, + 555, + 216 + ], + "type": "text", + "content": "-gram matching scores between candidate and reference captions, making them highly sensitive to stylistic variations and leading to inconsistent evaluations. Model-based metrics, such as SPICE [3], suffer from outdated text encoders as backbones, while CLIP-based alternatives [15, 38] enforce a strict 77-token input limit, further restricting their effectiveness in evaluating detailed captions." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 217, + 556, + 361 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 217, + 556, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 217, + 556, + 361 + ], + "type": "text", + "content": "We tackle both challenges through two key strategies. For detailed caption annotation, we construct ground truth descriptions using state-of-the-art MLLMs [29], followed by manual error correction, missing element addition, and hallucination removal through expert human intervention. For evaluation metrics, in addition to leveraging CAPTURE [9], a metric specifically designed for long captions, we introduce a structured evaluation framework that assesses captions based on precision, recall, and F1 scores of three core visual elements: objects, attributes, and relations [9, 32]. To achieve this, we develop an automated pipeline with two key steps: visual element extraction and structured matching." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 361, + 556, + 484 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 313, + 361, + 555, + 410 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 555, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 555, + 410 + ], + "type": "text", + "content": "(a) Element Extraction. Given detailed captions, we first split them into separate sentences using NLTK toolkit [30] and extract key visual elements—objects, attributes, and their relations—using a T5-based factual parser [27, 39]." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 412, + 556, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 412, + 556, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 412, + 556, + 484 + ], + "type": "text", + "content": "(b) Structured Matching. The matching process aligns extracted objects, attributes, and relations in three steps: exact matching (identifying precise word matches), synonym matching (matching words with similar meanings), and semantic matching (computing cosine similarity of word embeddings for any remaining unmatched elements)." + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 484, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 484, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 484, + 555, + 557 + ], + "type": "text", + "content": "Finally, we compute the precision, recall, and F1 scores for visual elements based on the matching results. These metrics capture semantic discrepancies: precision penalizes hallucinations and recall captures omissions, reducing confounds from pretrained generative model biases and offering a more accurate measure of genuine brain-based decoding." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 564, + 489, + 577 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 564, + 489, + 577 + ], + "spans": [ + { + "bbox": [ + 313, + 564, + 489, + 577 + ], + "type": "text", + "content": "4.2. Question-Answering Annotation" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 556, + 713 + ], + "type": "text", + "content": "Previous studies have emphasized the multimodal evaluation for decoded salient objects, considering semantic selectivities in the higher visual cortex of the human brain [8, 18, 37]. Tiny objects in stimuli may not capture the subject's attention or relevant brain activities may not be effectively recorded during experiments [2, 53]. Thus, we further design evaluation focused on salient objects using annotated question-answering pairs. First, we extract salient objects using SAM [22] and input the cropped regions into the MLLM to identify objects, attributes, and relationships within each image using pipelines introduced in Sec. 4.1. We use SAM" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4373" + } + ] + } + ], + "index": 18 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 120, + 504, + 247 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 117 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 117 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 117 + ], + "type": "text", + "content": "Table 1. Concept Localization. We present brain grounding [53] results derived directly from brain signals, reflecting spatial information encoded in brain activities. Image-based results are included as references, comprising visual grounding from Shikra [7] using associated visual stimuli and grounding performance based on reconstructed images from state-of-the-art visual decoding methods [35, 41, 53, 54]. We highlight the best, second-best, and third-best performance averaged across all four subjects." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 120, + 504, + 247 + ], + "lines": [ + { + "bbox": [ + 107, + 120, + 504, + 247 + ], + "spans": [ + { + "bbox": [ + 107, + 120, + 504, + 247 + ], + "type": "table", + "html": "
MethodAll acc@0.5IoUSalient acc@0.5IoUSalient Creatures acc@0.5IoUSalient Objects acc@0.5IoUInconspicuous acc@0.5IoU
VisionShikra [7]51.9647.2262.9256.4466.7159.3458.7953.2738.2935.71
BrainDiffuser [35]17.4919.3427.1827.4638.7134.6314.6219.665.399.20
MindEye [41]15.3418.6523.8326.9629.2931.6417.8821.864.748.28
DREAM [54]16.2118.6526.5127.3534.4333.8517.8820.283.357.78
UMBRAE [53]16.8318.6927.1027.5534.1433.6519.4420.924.007.64
BrainUMBRAE-S [53]14.3118.1322.2825.8226.3229.5817.8821.724.378.53
UMBRAE [53]18.0320.6328.2229.3236.9335.6118.7422.475.329.79
VINDEX-S14.9218.5523.3826.2828.0030.2918.3521.924.378.91
VINDEX18.1020.8928.5029.7837.0036.2019.2522.785.119.80
", + "image_path": "00ae8b4bbaf2cf2e4c766c1818c40f5c672dae6971f83d8c5c4ce7618240a9e8.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 267, + 297, + 435 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 267, + 297, + 435 + ], + "spans": [ + { + "bbox": [ + 54, + 267, + 297, + 435 + ], + "type": "text", + "content": "instead of the original COCO location annotations or Referring Expression Comprehension [7] with COCO classes [28] to extract such information as pseudolabels, since COCO often overlooks salient categories outside its 80 common classes. For instance, an image may feature a large church, but only two small clocks on it are annotated with their spatial locations and linguistic descriptions. For each identified object, we design questions such as \"which description best fits the in the image\" and \"what is the relationship between and to evaluate the model's ability to recognize objects and their interactions within the image. Additionally, we include in-depth reasoning questions [29] that require step-by-step logical processing, based on GPT annotations [1, 43], to assess deeper reasoning capabilities." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 439, + 297, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 297, + 583 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 297, + 583 + ], + "type": "text", + "content": "To mitigate the MLLM's tendency to default to affirmative responses such as \"Yes, I can see it\" rather than genuinely recognizing an object, we follow [32] and introduce an equivalent number of question-answer pairs for objects not present in the image as hallucination components. We then ask the model to select the correct description from multiple choices, where one option is a correct caption, and two are incorrect. The attribute annotations of fine-grained objects serve as the answers, while erroneous linguistic annotations are manually modified to ensure they contradict the actual visual content. We refer to this as salient QA (SQA) for complex reasoning and report the response accuracy." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 605, + 137, + 619 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 137, + 619 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 137, + 619 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 297, + 715 + ], + "type": "text", + "content": "In our experiments, we demonstrate multimodal interaction with fMRI using MLLMs to tackle four tasks: concept localization (Sec. 5.2), concise captioning (Sec. 5.3), descriptive captioning (Sec. 5.4), and complex reasoning (Sec. 5.5). While the generated multimodal explanations have been shown to enhance reconstruction performance [43, 49, 53], we prioritize brain perception over reconstruction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 267, + 448, + 280 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 267, + 448, + 280 + ], + "spans": [ + { + "bbox": [ + 313, + 267, + 448, + 280 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "spans": [ + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": "Architecture. For the mixture of vision encoders, we use pre-trained CLIP [38] and DINOv2 [34] as visual encoders. The target image features are extracted from both transformer encoders and then interleaved to spatially mix visual tokens before being subsequently processed by the connector and LLM [50]. The feature dimension " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "D" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": " is 4,096 for 7B models and 5,120 for 13B models. For aggregated features, we utilize SigLIP [58] to extract visual features from multiple layers [57]. The connector first encapsulates multi-layer visual features and then maps the integrated visual features to the token space. For nested features, we use CLIP [38] to encode visual cues in a coarse-to-fine manner [6], where an image is represented as patch-based visual tokens. We create " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": " sets of tokens, with the visual tokens at the coarser level derived directly from those at the finer level. Specifically, starting from the initial visual tokens, we sequentially apply " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "2 \\times 2" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": " pooling with a stride of 2, resulting in " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "12 \\times 12" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "6 \\times 6" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": " visual tokens. Finally, we apply " + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "inline_equation", + "content": "3 \\times 3" + }, + { + "bbox": [ + 313, + 289, + 556, + 530 + ], + "type": "text", + "content": " pooling to obtain the most condensed single visual token. Details on encoders and MLLMs can be found in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "content": "Training Details. Our models are trained on a single A100 GPU. CLIP-224 [38] and DINO-224 [34] feature alignment with a batch size of 128 takes approximately 8 hours, CLIP-336 [38] with a batch size of 64 takes about 23 hours, while SigLIP-384 [58] with a batch size of 4 requires around 40 hours to converge. We use AdamW [31] as the optimizer with " + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.95" + }, + { + "bbox": [ + 313, + 533, + 557, + 715 + ], + "type": "text", + "content": ", weight decay of 0.01, and apply the one-cycle strategy [46] with an initial learning rate of 3e-4 for the learning rate scheduler. Following visual decoding studies [41, 49, 54], we use the standard train and test splits for four subjects (sub01, sub02, sub05, sub07), each with 24,980 training samples and a shared 982 test samples. For evaluation, we report the average of the three repetitions of the same image in the test set, totaling 982 samples per subject. We implement the denoiser as an MLP with a default" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4374" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 107, + 140, + 506, + 404 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 137 + ], + "type": "text", + "content": "Table 2. Brain Captioning. We evaluate short captions on BrainHub [53] and detailed captions on MG-BrainDUB. Models with '-S' refers to those with cross-subject support but are trained in a single-subject setting for comparison. MindEye2 results are based on a model adapted to one subject using a pre-trained model that was trained on the remaining seven subjects with the full dataset. The mark† denotes zero-shot methods, indicating that no external captions were used during training. 'Shikra-image' refers to image captioning results from Shikra [7] using corresponding visual stimuli as input, acting as an approximate performance upper bound. Colors denote best, second-best, and third-best performance. Bold and underline represent the best and second best results under identical protocols." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 107, + 140, + 506, + 404 + ], + "lines": [ + { + "bbox": [ + 107, + 140, + 506, + 404 + ], + "spans": [ + { + "bbox": [ + 107, + 140, + 506, + 404 + ], + "type": "table", + "html": "
MethodBLEU1BLEU2BLEU3BLEU4METEORROUGECIDErSPICECLIP-SRefCLIP-S
Concise Captioning (Sec. 5.3)
Shikra-img [7]82.3869.9058.6349.6635.6065.49161.4327.6280.6085.92
SDRecon [49]36.2117.117.723.4310.0325.1313.835.0261.0766.36
OneLLM [13]47.0426.9715.499.5113.5535.0522.996.2654.8061.28
BrainCap [12]55.9636.2122.7014.5116.6840.6941.309.0664.3169.90
NeuroVLA [43]57.1937.1723.7815.8518.6036.6749.5112.3965.49-
MindEye2 [42]54.8238.6026.4918.1617.5443.7755.7010.9767.5473.73
MEVOX [55]†58.5640.3628.0920.1119.2044.4754.3711.0564.2370.36
UMBRAE-S†57.6338.0225.0016.7618.4142.1551.9311.8366.4472.12
UMBRAE [53]†59.4440.4827.6619.0319.4543.7161.0612.7967.7873.54
VINDEX-S†57.9939.0026.0417.8318.5642.2753.9611.9266.9472.64
VINDEX†60.0040.7227.5718.9119.5143.7860.3212.8468.2673.88
Descriptive Captioning (Sec. 5.4)
NeuroVLA [43]38.9124.0215.2412.4118.4427.8342.5818.4156.16-
VINDEX
w/ SE22.7112.897.214.1113.9525.4314.058.4362.9067.72
w/ ME21.3911.866.313.3911.3117.606.046.4360.8565.96
w/ AF0.090.010.000.000.041.890.030.0153.1759.86
w/ NF120.1011.326.143.4014.6721.151.479.8560.3563.01
w/ NF950.7733.2521.1713.2018.2641.7246.5011.4565.4071.20
w/ NF3638.8723.8714.398.8014.7233.4031.078.9464.0269.80
w/ NF1443.742.221.270.743.4610.456.581.6957.1463.77
", + "image_path": "ea7d29e1251f0befef875fe7b29d3e3e000d6875cd943d1a684696aa6d79186e.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 423, + 295, + 462 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 462 + ], + "type": "text", + "content": "depth of 1, a latent dimension (width) of 1024, and a weight multiplier of 1.0 for the denoising loss (indicating the same contribution as the regression loss during training)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 468, + 179, + 480 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 468, + 179, + 480 + ], + "spans": [ + { + "bbox": [ + 55, + 468, + 179, + 480 + ], + "type": "text", + "content": "5.2. Concept Localization" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "spans": [ + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "text", + "content": "This section evaluates the model's ability for concept localization through the brain grounding task [53]. Similar to visual referring expression comprehension, brain grounding uses the prompt \"Find the coordinates of " + }, + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "inline_equation", + "content": "<\\text{expr}>" + }, + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "text", + "content": "\" and evaluates performance using accuracy and IoU metrics. The accuracy metric, acc@m, measures the percentage of correctly labeled instances with an IoU greater than the threshold " + }, + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 54, + 485, + 297, + 640 + ], + "type": "text", + "content": ". We assess grounding results derived directly from brain signals, with image-based results provided for reference. These include visual grounding using ground truth stimuli and reconstructed images from state-of-the-art (SOTA) visual decoding methods [35, 41, 54] via Shikra [7]. We provide a comparison with SOTAs and our methods in Tab. 1." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 297, + 714 + ], + "type": "text", + "content": "The results demonstrate that our method VINDEX outperforms all baselines across all metrics in the four salient-based categories in BrainHub [53]. Notably, VINDEX surpasses both direct brain signal-based predictions from UM-BRAE [53] and image-based grounding results, which use reconstructed images from brain signals as input [35, 41, 53," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 423, + 556, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 423, + 556, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 423, + 556, + 533 + ], + "type": "text", + "content": "54]. This emphasizes that directly extracting location information from the captured brain activities is more accurate than relying on reconstructed images. This observation suggests that reconstruction methods relying on image and text semantic encoding have inherent limitations in decoding spatial information. It also implies that incorporating grounding information in future models could help enhance the accuracy of object location prediction when reconstructing the visual stimuli from neural activations." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 430, + 553 + ], + "type": "text", + "content": "5.3. Concise Captioning" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 557, + 557, + 715 + ], + "type": "text", + "content": "We evaluate short captions on BrainHub [53], demonstrating our method's performance on multimodal language tasks. Tab. 2 presents an evaluation of our brain captioning compared to state-of-the-art baselines, including three subject-specific methods SDRecon [49], BrainCap [12], OneLLM [13], and four cross-subject methods NeuroVLA [43], MindEye2 [42], MEVOX [55], and UMBRAE [53]. Our method, VINDEX, for concise captioning is built upon the same base MLLM Shikra [7] as in UMBRAE [53] for a fair comparison. MindEye2 is first pre-trained on seven subjects and then adapted to the remaining subject using all available data. NeuroVLA [43], UMBRAE [53], and VINDEX are trained on four subjects" + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4375" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 77, + 118, + 533, + 251 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 115 + ], + "type": "text", + "content": "Table 3. Descriptive Captioning and Complex Reasoning. We evaluate detailed descriptions and complex reasoning with salient question-answering on MG-BrainDub under different model settings. 'Space' refers to the feature space elaborated in Sec. 3.1, while 'size' and 'encoder' indicate the size of ground truth images and the encoder used for feature extraction. Colors denote best, second-best, and third-best performance. Models marked with * correspond to those with the same settings as in Tab. 2." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 77, + 118, + 533, + 251 + ], + "lines": [ + { + "bbox": [ + 77, + 118, + 533, + 251 + ], + "spans": [ + { + "bbox": [ + 77, + 118, + 533, + 251 + ], + "type": "table", + "html": "
SpaceModel SettingCAPTUREObjectAttributeRelationRuntime (s)SQA Acc (%)
SizeEncoderMLLMPRF1PRF1PRF1
SE*224CLIPLLaVA-1.5 7B0.435657.2249.7651.8837.9036.0835.7345.2540.7042.3214.0575.50
SE224CLIPLLaVA-1.5 13B0.473361.8152.9155.6040.6841.2139.3849.2845.0046.528.7077.92
SE224CLIPLLaVA-1.6 7B0.468451.0258.7853.5136.0154.0141.8041.0045.6342.7510.9176.80
SE224CLIPLLaVA-1.6 13B0.464550.4457.8352.8135.8954.0241.7240.4245.4442.3824.1878.69
ME*224CLIP, DINOLLaVA-1.5 13B0.304149.1734.3038.0726.1023.0923.3733.0826.8628.865.0652.66
AF*384SigLIPLLaVA-1.5 7B0.00581.891.071.270.150.110.130.010.000.005.6915.17
NF1*224CLIPLLaVA-1.5 7B0.487759.1957.1957.2442.7442.7641.1847.3246.7046.575.4582.59
NF9*224CLIPLLaVA-1.5 7B0.502162.4359.2659.7441.9743.2040.9550.5649.4949.545.3183.83
NF36*224CLIPLLaVA-1.5 7B0.467262.2354.8956.3740.2438.5637.5847.9044.4145.474.3366.35
NF144*224CLIPLLaVA-1.5 7B0.161452.8422.4728.738.675.556.3212.887.679.221.8331.35
NF1224CLIPLLaVA-1.6 7B0.435949.4751.3749.4335.6549.5940.2336.7538.9637.408.6265.58
NF1336CLIPLLaVA-1.5 7B0.412047.9946.7446.8943.0139.5039.5331.1531.9631.165.0662.27
", + "image_path": "97777d63fe5de5fb2fcceec3e4421e2ab66ef750cb44b1f5f125496c0bb9b141.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "spans": [ + { + "bbox": [ + 54, + 271, + 297, + 415 + ], + "type": "text", + "content": "instead of eight. Caption results from Shikra [7] using associated visual stimuli for captioning act as an approximate upper bound. VINDEX achieves top performance across most metrics for concise brain captioning, even surpassing methods trained with more subject data [42] or external annotations [43]. The other brain-feature alignment methods UMBRAE [53] and MEVOX [55] ranked the second and the third. This highlights the effectiveness of the cross-subject brain feature alignment strategy, showcasing superior zero-shot performance without the need for external training data and generalization across subjects without requiring separate models or subject-specific parameters." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 415, + 297, + 559 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 415, + 297, + 559 + ], + "spans": [ + { + "bbox": [ + 55, + 415, + 297, + 559 + ], + "type": "text", + "content": "NeuroVLA, MindEye2, UMBRAE, and VINDEX generate fluent, complete, and information-rich sentences, highlighting the advantages of using LLMs. Notably, VINDEX and UMBRAE achieve superior performance in a zero-shot setting, without relying on additional annotations, outperforming models trained with real captions from COCO [28] or pseudo captions from MLLMs [29]. This demonstrates that directly aligning brain signals with image features offers a better way than using external visual, spatial, or linguistic annotations. It preserves more accurate semantic and spatial cues decoded from brain signals and avoids factual errors or stylistic inconsistencies from external annotations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 564, + 187, + 578 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 564, + 187, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 564, + 187, + 578 + ], + "type": "text", + "content": "5.4. Descriptive Captioning" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": "We present a quantitative comparison of descriptive captioning results for NeuroVLA [43] and our method with different feature space choices: single encoder (SE), mixture of encoders (ME), aggregated feature (AF), and nested features " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\mathrm{(NF_n)}" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " where the token size " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " is selected from " + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "inline_equation", + "content": "\\{1, 9, 36, 144\\}" + }, + { + "bbox": [ + 55, + 582, + 297, + 714 + ], + "type": "text", + "content": " after downsampling. The latter Tab. 2 presents evaluations using regular caption metrics from [53]. Considering the limitations of traditional rule-based metrics (sensitive to writing style) and model-based metrics (unable to process long sequences), we evaluate descriptive captioning on MG-BrainDUB (Sec. 4.1). Results in Tab. 3 show that using" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 271, + 555, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 555, + 356 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 555, + 356 + ], + "type": "inline_equation", + "content": "\\mathrm{NF}_9" + }, + { + "bbox": [ + 313, + 271, + 555, + 356 + ], + "type": "text", + "content": " (with 9 visual tokens) as supervision for training brain encoders achieves the best performance on most metrics. Surprisingly, using dense feature aggregation (AF) instead yields the worst results, and models trained with vanilla CLIP features [38] (SE setting at first four rows) remain inferior, even with larger and more sophisticated MLLMs (scaling from v1.5 to v1.6, 7B to 13B)." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "spans": [ + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "text", + "content": "Training with feature spaces from larger CLIP models (336 vs. 224) does not improve performance but significantly increases training time and computational costs. The brain encoder aligned with CLIP-ViT-L-224 features (corresponding to " + }, + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "inline_equation", + "content": "16 \\times 16" + }, + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "text", + "content": " tokens) generally transfers well to MLLMs designed with the CLIP-ViT-L-336 model (" + }, + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "inline_equation", + "content": "24 \\times 24" + }, + { + "bbox": [ + 313, + 361, + 556, + 601 + ], + "type": "text", + "content": " tokens) as the visual encoder, often yielding better results. AF performs the worst but with the largest token size (729 visual tokens for SigLIP-384 [58]). In contrast, aligning brain encoders with CLIP-224 requires only one-fifth of SigLIP-384's training time and one-third of CLIP-336's. Recall that NF downsamples the predicted CLIP features (the same features used in SE/ME) using pooling operations. Notably, 9-token alignment yields the best results, while 1-token alignment ranks second, suggesting that fewer tokens, though less expressive, align better with brain signals, which encode less information than real images. This also explains the inferior results for ME, where mixing brain-CLIP and brain-DINO features amplifies prediction imprecisions, leading to worse performance than using brain-CLIP features alone." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "The performance drop with increasing feature magnitude suggests that, while richer image features benefit visual perception tasks [6, 50, 57], brain signals capture relatively less comprehensive visual elements. This may due to subjects' selective attention [8, 18, 37] or limitations in capturing brain activities [2, 53]. We expect additional features to enhance the perception of subtler elements, but they mainly introduce unnecessary hallucinations and increase training costs, as brain signals predominantly encode basic visual elements." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "4376" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": "Results in Tab. 3 also indicate that traditional caption metrics [4, 36, 51] are not reliable when scaling to longer, more detailed captions. The rule-based metrics show that ME has higher scores than " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{NF}_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": ", even though " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{NF}_1" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " predicts more accurate objects, attributes, and relations. CLIP-based metrics [15] also show deficiencies, as they present high similarities for all generated captions, even for AF, which in most cases lead to nonsensical outputs, such as garbled text and blank spaces. The inference time does not increase as expected with the model complexity and token length; instead, it exhibits irregularities. " + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "inline_equation", + "content": "\\mathrm{NF}_9^*" + }, + { + "bbox": [ + 55, + 72, + 294, + 228 + ], + "type": "text", + "content": " presents the best results across most metrics while maintaining low computation cost in terms of inference time and memory consumption." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 235, + 171, + 248 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 235, + 171, + 248 + ], + "spans": [ + { + "bbox": [ + 55, + 235, + 171, + 248 + ], + "type": "text", + "content": "5.5. Complex Reasoning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 252, + 295, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 252, + 295, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 252, + 295, + 503 + ], + "type": "text", + "content": "We evaluate the model's ability to perform complex reasoning on salient objects using Salient Question Answering (SQA) (Sec. 5.5), with accuracy reported in Tab. 3. Unlike previous sections focused on visual content, SQA involves in-depth reasoning that demands step-by-step logical processing to provide the right answer. This experiment highlights the model's ability to decode intricate characteristics of salient elements, meaning the prominent objects most likely to capture the subject's attention during brain activity capture [2], which have been shown to be better preserved in brain data [53]. From the table, " + }, + { + "bbox": [ + 55, + 252, + 295, + 503 + ], + "type": "inline_equation", + "content": "\\mathrm{NF}_9^*" + }, + { + "bbox": [ + 55, + 252, + 295, + 503 + ], + "type": "text", + "content": " achieves the best SQA performance across image sizes, visual encoders, and MLLM settings, likely benefiting from a balance between rich visual features and brain data constraints, consistent with detailed description results. In contrast, ME and AF sometimes produce meaningless outputs due to improper feature-brain signal alignment, caused by the added complexity of features during training. They also suffer from hallucinations, generating non-existent objects and misrepresenting attributes and relationships, resulting in accuracy rates at or below random guessing." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 514, + 155, + 526 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 514, + 155, + 526 + ], + "spans": [ + { + "bbox": [ + 55, + 514, + 155, + 526 + ], + "type": "text", + "content": "6. Ablation Studies" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 713 + ], + "type": "text", + "content": "Model-Agnostic Evaluation. VINDEX is a versatile, model-agnostic brain decoding framework that can be easily adapted to various subjects, image sizes, visual encoders, MLLM settings, input prompts, and downstream tasks. Due to the architectural consistency of LLaVA-based variants [6, 7, 29, 50, 57], our trained brain encoders can be directly integrated into different MLLMs without requiring additional modifications. Specifically, for model setups presented in Tab. 3, we only need to train four brain encoders using CLIP-224, CLIP-336, DINO-224, and SigLIP-384. The same brain encoder trained with CLIP-224 can be used across SE, ME, and NF (except for the last one with a 336 image size). This brain-CLIP features are then passed to the remaining MLLMs components, including connectors and LLMs, with task prompts supported by the base MLLM for" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 319, + 96, + 550, + 153 + ], + "blocks": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": "Table 4. Ablation Study on Denoiser. CAPTURE [9] is reported across architectural depth " + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": ", and weight multiplier " + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 71, + 553, + 93 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 96, + 550, + 153 + ], + "lines": [ + { + "bbox": [ + 319, + 96, + 550, + 153 + ], + "spans": [ + { + "bbox": [ + 319, + 96, + 550, + 153 + ], + "type": "table", + "html": "
SetupS0S1S2S3S4S5S6S7
β-1.01.01.01.00.51.52.0
d-1231333
w-5121,0241,0241,0241,0241,0241024
C0.41790.43080.42570.41830.43560.41930.41970.4096
", + "image_path": "b4c282c672643ae5b96e98c235425ad72dc89e802443530c26f5df822fb63cc6.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 176, + 553, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 176, + 553, + 260 + ], + "spans": [ + { + "bbox": [ + 313, + 176, + 553, + 260 + ], + "type": "text", + "content": "multimodal interactions, such as interactive dialogue, brain captioning, complex reasoning, and concept localization. The brain encoder aligned with CLIP-224 features adapts well to LLaVA-1.6-7B/13B models, which are designed to use CLIP-336 as the visual encoder, leading to improved performance. This versatility enables our method to efficiently align brain signals with multiple MLLMs at minimal cost." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "spans": [ + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "text", + "content": "Denoising Optimization Objective. We conduct ablation studies on the denoiser in Tab. 4 to evaluate the effectiveness of the denoising objective over vanilla regression across different settings, focusing on architectural depth " + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "text", + "content": ", width " + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "text", + "content": ", and the weight multiplier " + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 263, + 555, + 455 + ], + "type": "text", + "content": ". Here, 'Sn' represents different configurations of specific depth, width, and weight multiplier; S0 refers to the base setup with only the regression loss, without the denoising loss. The results are reported as CAPTURE scores [9] (C) for detailed descriptions across each setup. Total training hours for 300 epochs across all settings is approximately eight hours, with no notable variations. We found that a simple one-depth MLP with 1,024 model channels yields the best results. Results show that training with the denoising loss setup incurs no extra cost but leads to improved performance and more stable training (further analysis in the supplementary material)." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "spans": [ + { + "bbox": [ + 313, + 468, + 388, + 481 + ], + "type": "text", + "content": "7. Conclusion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 490, + 556, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 490, + 556, + 693 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 556, + 693 + ], + "type": "text", + "content": "In this paper, we propose a zero-shot multimodal brain decoding method that explores representative feature spaces from the visual components within MLLMs. We elucidate different vision feature space choices and introduce a denoiser-based strategy to enhance brain alignment with pre-trained models, improving the accuracy and robustness of brain-encoded feature representations. Our method facilitates efficient and flexible decoding across multiple levels of granularity, eliminating the need for additional textual or spatial annotations during training. Furthermore, we present MG-BrainDUB, a benchmark that provides a more reliable and comprehensive evaluation for multigranular multimodal brain decoding. It comprises two core tasks—detailed description generation and salient question answering—along with a novel metric that prioritizes key visual elements including objects, attributes, and relationships, offering a robust framework for assessing fine-grained brain perception." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 695, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 695, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 695, + 553, + 713 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by a UKRI Future Leaders Fellowship [grant number G104084] and an Imminent Research Grant." + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 315, + 757 + ], + "type": "text", + "content": "4377" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 135 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "spans": [ + { + "bbox": [ + 61, + 137, + 296, + 201 + ], + "type": "text", + "content": "[2] Emily J Allen, Ghislain St-Yves, Yihan Wu, Jesse L Breedlove, Jacob S Prince, Logan T Dowdle, Matthias Nau, Brad Caron, Franco Pestilli, Ian Charest, et al. A massive 7t fmri dataset to bridge cognitive neuroscience and artificial intelligence. Nature neuroscience, 25(1):116-126, 2022. 4, 7, 8" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 204, + 295, + 236 + ], + "type": "text", + "content": "[3] Peter Anderson, Basura Fernando, Mark Johnson, and Stephen Gould. Spice: Semantic propositional image caption evaluation. In ECCV, pages 382-398. Springer, 2016. 1, 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 296, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 296, + 281 + ], + "type": "text", + "content": "[4] Satanjeev Banerjee and Alon Lavie. Meteor: An automatic metric for mt evaluation with improved correlation with human judgments. In ACL Workshop, pages 65-72, 2005. 1, 4, 8" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 283, + 296, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 283, + 296, + 315 + ], + "spans": [ + { + "bbox": [ + 62, + 283, + 296, + 315 + ], + "type": "text", + "content": "[5] Daniel Bolya, Cheng-Yang Fu, Xiaoliang Dai, Peizhao Zhang, Christoph Feichtenhofer, and Judy Hoffman. Token merging: Your vit but faster. In ICLR, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 317, + 296, + 339 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 317, + 296, + 339 + ], + "spans": [ + { + "bbox": [ + 62, + 317, + 296, + 339 + ], + "type": "text", + "content": "[6] Mu Cai, Jianwei Yang, Jianfeng Gao, and Yong Jae Lee. Matryoshka multimodal models. In ICLR, 2025. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "spans": [ + { + "bbox": [ + 62, + 341, + 296, + 384 + ], + "type": "text", + "content": "[7] Keqin Chen, Zhao Zhang, Weili Zeng, Richong Zhang, Feng Zhu, and Rui Zhao. Shikra: Unleashing multimodal llm's referential dialogue magic. arXiv preprint arXiv:2306.15195, 2023. 2, 5, 6, 7, 8" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "spans": [ + { + "bbox": [ + 62, + 386, + 296, + 430 + ], + "type": "text", + "content": "[8] Robert Desimone, Thomas D Albright, Charles G Gross, and Charles Bruce. Stimulus-selective properties of inferior temporal neurons in the macaque. Journal of Neuroscience, 4 (8):2051-2062, 1984. 4, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 431, + 296, + 474 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 431, + 296, + 474 + ], + "spans": [ + { + "bbox": [ + 62, + 431, + 296, + 474 + ], + "type": "text", + "content": "[9] Hongyuan Dong, Jiawen Li, Bohong Wu, Jiacong Wang, Yuan Zhang, and Haoyuan Guo. Benchmarking and improving detail image caption. arXiv preprint arXiv:2405.19092, 2024. 4, 8" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 476, + 296, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 476, + 296, + 531 + ], + "spans": [ + { + "bbox": [ + 57, + 476, + 296, + 531 + ], + "type": "text", + "content": "[10] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2021. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "spans": [ + { + "bbox": [ + 57, + 533, + 296, + 576 + ], + "type": "text", + "content": "[11] Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. EVA: Exploring the limits of masked visual representation learning at scale. In CVPR, 2023. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "spans": [ + { + "bbox": [ + 57, + 578, + 295, + 621 + ], + "type": "text", + "content": "[12] Matteo Ferrante, Furkan Ozcelik, Tommaso Boccato, Rufin VanRullen, and Nicola Toschi. Brain captioning: Decoding human brain activity into images and text. arXiv preprint arXiv:2305.11560, 2023. 4, 6" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 623, + 295, + 666 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 623, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 57, + 623, + 295, + 666 + ], + "type": "text", + "content": "[13] Jiaming Han, Kaixiong Gong, Yiyuan Zhang, Jiaqi Wang, Kaipeng Zhang, Dahua Lin, Yu Qiao, Peng Gao, and Xiangyu Yue. Onellm: One framework to align all modalities with language. In CVPR, 2024. 1, 2, 4, 6" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 669, + 295, + 701 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 669, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 57, + 669, + 295, + 701 + ], + "type": "text", + "content": "[14] Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollar, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000-16009, 2022. 3, 4" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 57, + 703, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 703, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 57, + 703, + 296, + 713 + ], + "type": "text", + "content": "[15] Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras," + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 554, + 95 + ], + "type": "text", + "content": "and Yejin Choi. Clipscore: A reference-free evaluation metric for image captioning. In EMNLP, 2021. 1, 4, 8" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 98, + 555, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 98, + 555, + 129 + ], + "spans": [ + { + "bbox": [ + 317, + 98, + 555, + 129 + ], + "type": "text", + "content": "[16] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840-6851, 2020. 3, 4" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 132, + 555, + 175 + ], + "type": "text", + "content": "[17] Andrew Jaegle, Felix Gimeno, Andy Brock, Oriol Vinyals, Andrew Zisserman, and Joao Carreira. Perceiver: General perception with iterative attention. In ICML, pages 4651-4664. PMLR, 2021. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 177, + 555, + 220 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 177, + 555, + 220 + ], + "spans": [ + { + "bbox": [ + 317, + 177, + 555, + 220 + ], + "type": "text", + "content": "[18] Nancy Kanwisher, Josh McDermott, and Marvin M Chun. The fusiform face area: a module in human extrastriate cortex specialized for face perception. Journal of neuroscience, 17 (11):4302-4311, 1997. 4, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 223, + 555, + 267 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 223, + 555, + 267 + ], + "spans": [ + { + "bbox": [ + 316, + 223, + 555, + 267 + ], + "type": "text", + "content": "[19] Oğuzhan Fatih Kar, Alessio Tonioni, Petra Poklukar, Achin Kulshrestha, Amir Zamir, and Federico Tombari. Brave: Broadening the visual encoding of vision-language models. In ECCV, pages 113-132. Springer, 2024. 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 269, + 555, + 312 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 269, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 317, + 269, + 555, + 312 + ], + "type": "text", + "content": "[20] Tero Karras, Samuli Laine, Miika Aittala, Janne Hellsten, Jaakko Lehtinen, and Timo Aila. Analyzing and improving the image quality of StyleGAN. In CVPR, pages 8107-8116, 2020. 2" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 315, + 555, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 315, + 555, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 315, + 555, + 358 + ], + "type": "text", + "content": "[21] Minchul Kim, Shangqian Gao, Yen-Chang Hsu, Yilin Shen, and Hongxia Jin. Token fusion: Bridging the gap between token pruning and token merging. In WACV, pages 1383-1392, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 360, + 555, + 405 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 555, + 405 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 555, + 405 + ], + "type": "text", + "content": "[22] Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, Piotr Dollar, and Ross Girshick. Segment anything. In ICCV, 2023. 3, 4" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 407, + 555, + 460 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 407, + 555, + 460 + ], + "spans": [ + { + "bbox": [ + 317, + 407, + 555, + 460 + ], + "type": "text", + "content": "[23] Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, et al. Matryoshka representation learning. In NeurIPS, pages 30233-30249, 2022. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 462, + 555, + 518 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 462, + 555, + 518 + ], + "spans": [ + { + "bbox": [ + 317, + 462, + 555, + 518 + ], + "type": "text", + "content": "[24] Kenton Lee, Mandar Joshi, Iulia Raluca Turc, Hexiang Hu, Fangyu Liu, Julian Martin Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, and Kristina Toutanova. Pix2Struct: Screenshot parsing as pretraining for visual language understanding. In ICML, 2023. 3" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 520, + 555, + 563 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 520, + 555, + 563 + ], + "spans": [ + { + "bbox": [ + 317, + 520, + 555, + 563 + ], + "type": "text", + "content": "[25] Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888-12900. PMLR, 2022. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 565, + 555, + 598 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 565, + 555, + 598 + ], + "spans": [ + { + "bbox": [ + 317, + 565, + 555, + 598 + ], + "type": "text", + "content": "[26] Tianhong Li, Yonglong Tian, He Li, Mingyang Deng, and Kaiming He. Autoregressive image generation without vector quantization. NeurIPS, 37:56424-56445, 2024. 3, 4" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 600, + 555, + 644 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 600, + 555, + 644 + ], + "spans": [ + { + "bbox": [ + 317, + 600, + 555, + 644 + ], + "type": "text", + "content": "[27] Zhuang Li, Yuyang Chai, Terry Yue Zhuo, Lizhen Qu, Gholamreza Haffari, Fei Li, Donghong Ji, and Quan Hung Tran. Factual: A benchmark for faithful and consistent textual scene graph parsing. In ACL, 2023. 4" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 317, + 646, + 555, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 646, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 646, + 555, + 689 + ], + "type": "text", + "content": "[28] Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Dólar, and C. Lawrence Zitnick. Microsoft COCO: Common objects in context. In ECCV, pages 740-755, 2014. 4, 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[29] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurIPS, 2023. 1, 2, 3, 4, 5, 7, 8" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "4378" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 14, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[30] Edward Loper and Steven Bird. Nltk: The natural language toolkit. In Proceedings of the ACL Interactive Poster and Demonstration Sessions, 2004. 4" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 296, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 296, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 296, + 129 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In ICLR, 2019. 5" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 296, + 183 + ], + "type": "text", + "content": "[32] Fan Lu, Wei Wu, Kecheng Zheng, Shuailei Ma, Biao Gong, Jiawei Liu, Wei Zhai, Yang Cao, Yujun Shen, and Zheng-Jun Zha. Benchmarking large vision-language models via directed scene graph for comprehensive image captioning. In CVPR, 2025. 4, 5" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 183, + 296, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 296, + 216 + ], + "type": "text", + "content": "[33] Alexander Quinn Nichol and Prafulla Dhariwal. Improved denoising diffusion probabilistic models. In ICML, pages 8162-8171, 2021. 4" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 217, + 296, + 271 + ], + "type": "text", + "content": "[34] Maxime Oquab, Timothee Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. DINOv2: Learning robust visual features without supervision. TMLR, 2023. 3, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "spans": [ + { + "bbox": [ + 56, + 271, + 296, + 316 + ], + "type": "text", + "content": "[35] Furkan Ozcelik and Rufin VanRullen. Brain-Diffuser: Natural scene reconstruction from fMRI signals using generative latent diffusion. Scientific Reports, 13(1):15666, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 296, + 349 + ], + "type": "text", + "content": "[36] Kishore Papineni, Salim Roukos, Todd Ward, and Wei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In ACL, pages 311-318, 2002. 1, 4, 8" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 349, + 296, + 403 + ], + "type": "text", + "content": "[37] Aina Puce, Truett Allison, Maryam Asgari, John C Gore, and Gregory McCarthy. Differential sensitivity of human visual cortex to faces, letterstrings, and textures: a functional magnetic resonance imaging study. Journal of neuroscience, 16(16):5205-5215, 1996. 4, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 459 + ], + "type": "text", + "content": "[38] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763. PMLR, 2021. 1, 2, 3, 4, 5, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 296, + 514 + ], + "type": "text", + "content": "[39] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer. JMLR, 21(140):1-67, 2020. 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 514, + 296, + 558 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 296, + 558 + ], + "type": "text", + "content": "[40] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684-10695, 2022. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 296, + 613 + ], + "type": "text", + "content": "[41] Paul S Scotti, Atmadeep Banerjee, Jimmie Goode, Stepan Shabalin, Alex Nguyen, Ethan Cohen, Aidan J Dempster, Nathalie Verlinde, Elad Yundler, David Weisberg, et al. Reconstructing the mind's eye: fmri-to-image with contrastive learning and diffusion priors. In NeurIPS, 2023. 1, 2, 5, 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 614, + 296, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 296, + 679 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 296, + 679 + ], + "type": "text", + "content": "[42] Paul S Scotti, Mihir Tripathy, Cesar Kadir Torrico Villanueva, Reese Kneeland, Tong Chen, Ashutosh Narang, Charan Santhirasegaran, Jonathan Xu, Thomas Naselaris, Kenneth A Norman, and Tanishq Mathew Abraham. Mindeye2: Shared-subject models enable fmri-to-image with 1 hour of data. In ICML, 2024. 1, 2, 6, 7" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 296, + 713 + ], + "type": "text", + "content": "[43] Guobin Shen, Dongcheng Zhao, Xiang He, Linghao Feng, Yiting Dong, Jihang Wang, Qian Zhang, and Yi Zeng. Neurovision to language: Enhancing brain recording-based visual" + } + ] + } + ], + "index": 13 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "reconstruction and language interaction. In NeurIPS, pages 98083-98110, 2024. 1, 2, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 555, + 139 + ], + "type": "text", + "content": "[44] Leyang Shen, Gongwei Chen, Rui Shao, Weili Guan, and Liqiang Nie. Mone: Mixture of multimodal experts for generalist multimodal large language models. In NeurIPS, pages 42048-42070, 2025. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 555, + 205 + ], + "type": "text", + "content": "[45] Min Shi, Fuxiao Liu, Shihao Wang, Shijia Liao, Subhashree Radhakrishnan, De-An Huang, Hongxu Yin, Karan Sapra, Yaser Yacoob, Humphrey Shi, Bryan Catanzaro, Andrew Tao, Jan Kautz, Zhiding Yu, and Guilin Liu. Eagle: Exploring the design space for multimodal llms with mixture of encoders. In ICLR, 2025. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 205, + 554, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 205, + 554, + 250 + ], + "spans": [ + { + "bbox": [ + 316, + 205, + 554, + 250 + ], + "type": "text", + "content": "[46] Leslie N Smith and Nicholay Topin. Super-convergence: Very fast training of neural networks using large learning rates. In Artificial intelligence and machine learning for multi-domain operations applications, pages 369-386. SPIE, 2019. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 250, + 554, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 250, + 554, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 250, + 554, + 271 + ], + "type": "text", + "content": "[47] Jiaming Song, Chenlin Meng, and Stefano Ermon. Denoising diffusion implicit models. In ICLR, 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 271, + 555, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 555, + 303 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 555, + 303 + ], + "type": "text", + "content": "[48] Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, 2019. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "spans": [ + { + "bbox": [ + 316, + 304, + 555, + 349 + ], + "type": "text", + "content": "[49] Yu Takagi and Shinji Nishimoto. Improving visual image reconstruction from human brain activity using latent diffusion models via multiple decoded inputs. arXiv preprint arXiv:2306.11536, 2023. 1, 2, 4, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 349, + 554, + 392 + ], + "type": "text", + "content": "[50] Shengbang Tong, Zhuang Liu, Yuexiang Zhai, Yi Ma, Yann LeCun, and Saining Xie. Eyes wide shut? exploring the visual shortcomings of multimodal llms. In CVPR, pages 9568-9578, 2024. 2, 5, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 393, + 555, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 555, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 555, + 426 + ], + "type": "text", + "content": "[51] Ramakrishna Vedantam, C Lawrence Zitnick, and Devi Parikh. Cider: Consensus-based image description evaluation. In CVPR, pages 4566-4575, 2015. 1, 4, 8" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "spans": [ + { + "bbox": [ + 316, + 426, + 555, + 459 + ], + "type": "text", + "content": "[52] Haochen Wang, Anlin Zheng, Yucheng Zhao, Tiancai Wang, Zheng Ge, Xiangyu Zhang, and Zhaoxiang Zhang. Reconstructive visual instruction tuning. In ICLR, 2025. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 459, + 555, + 492 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 459, + 555, + 492 + ], + "spans": [ + { + "bbox": [ + 316, + 459, + 555, + 492 + ], + "type": "text", + "content": "[53] Weihao Xia, Raoul de Charette, Cengiz Oztireli, and JingHao Xue. Umbrae: Unified multimodal brain decoding. In ECCV, pages 242-259, 2024. 1, 2, 3, 4, 5, 6, 7, 8" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 492, + 555, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 555, + 525 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 555, + 525 + ], + "type": "text", + "content": "[54] Weihao Xia, Raoul de Charette, Cengiz Özireli, and Jing-Hao Xue. Dream: Visual decoding from reversing human visual system. In WACV, pages 8226-8235, 2024. 1, 2, 5, 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 525, + 555, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 525, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 316, + 525, + 555, + 557 + ], + "type": "text", + "content": "[55] Weihao Xia and Cengiz Öztireli. Mevox: Multi-task vision experts for brain captioning. In CVPR Workshop, 2025. 2, 6, 7" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 555, + 592 + ], + "type": "text", + "content": "[56] Xingqian Xu, Zhangyang Wang, Eric Zhang, Kai Wang, and Humphrey Shi. Versatile diffusion: Text, images and variations all in one diffusion model. In ICCV, 2023. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 555, + 635 + ], + "type": "text", + "content": "[57] Huanjin Yao, Wenhao Wu, Taojiannan Yang, YuXin Song, Mengxi Zhang, Haocheng Feng, Yifan Sun, Zhiheng Li, Wanli Ouyang, and Jingdong Wang. Dense connector for mllms. In NeurIPS, pages 33108-33140, 2025. 2, 3, 5, 7, 8" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 555, + 669 + ], + "type": "text", + "content": "[58] Xiaohua Zhai, Basil Mustafa, Alexander Kolesnikov, and Lucas Beyer. Sigmoid loss for language image pre-training. In ICCV, pages 11975-11986, 2023. 5, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 555, + 712 + ], + "type": "text", + "content": "[59] Zhuofan Zong, Bingqi Ma, Dazhong Shen, Guanglu Song, Hao Shao, Dongzhi Jiang, Hongsheng Li, and Yu Liu. Mova: Adapting mixture of vision experts to multimodal context. In NeurIPS, 2024. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "4379" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_content_list.json b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..97eb7d0a73f6a71be02915c7a6e34efec04ef326 --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_content_list.json @@ -0,0 +1,1521 @@ +[ + { + "type": "text", + "text": "Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement", + "text_level": 1, + "bbox": [ + 200, + 130, + 797, + 175 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Shuo Zhang, Chen Gao, Youfang Lin* \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science& Technology, Beijing Jiaotong University, Beijing, China.", + "bbox": [ + 151, + 203, + 843, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhangshuo, gaochen, yflin}@bjtu.edu.cn", + "bbox": [ + 323, + 258, + 671, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 325 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Light Field (LF) images captured under low illumination conditions typically exhibit low quality. Recent learning-based methods for low-light LF images enhancement are generally tailored to specific illumination inputs, limiting their performance in real-world scenes. Moreover, how to maintain the inherent view-consistency in the enhanced images also remains a difficult problem. In this paper, we propose to explore the view consistency for scene-adaptive low-light LF enhancement. We first analyze the view consistency for LF illumination maps and design a self-supervised view-consistent loss to keep the consistency between the illumination maps of different views in LFs. To enhance the model's perception of illumination, we combine both global and local information to estimate the illumination map, which is easily plugged into other models. Subsequently, we use the illumination maps to light up the low-light LF images and restore the corruption to produce the final enhanced image. Experiments demonstrate that our View-Consistent Network (VCNet) outperforms state-of-the-art methods on real-world low-light LF datasets in both fixed lighting conditions and dynamic lighting conditions. Our proposed illumination adjustment is also demonstrated that can comprehensively improve the performance of existing methods in terms of both image quality and view consistency. Code is available at: https://github.com/shuozh/VCNet", + "bbox": [ + 89, + 340, + 485, + 734 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 762, + 220, + 777 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In contrast to conventional cameras, Light Field (LF) cameras have the capability to simultaneously capture both directional and intensity of all light rays in a single shot. Because of the structural information inherent in LF images, they are well-suited for various applications, including digital refocus [2], depth estimation [8], and 3D reconstruc", + "bbox": [ + 89, + 786, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/91e0ebca1dd1fc459bb03b5f1845d8ce6416d152f9409e24f7445440f00dfe8b.jpg", + "image_caption": [ + "Figure 1. Performance comparison in dynamic illumination tasks. The size of the points represents the run time to process one LF. Our VCNet achieves the best results in terms of detail restoration with higher PSNR and views consistency preservation with higher AP. By adding our proposed illumination adjustment module, labeled as “+E”, the performances of all methods are improved." + ], + "image_footnote": [], + "bbox": [ + 581, + 306, + 838, + 444 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tion [25]. Nevertheless, the low-quality LF images severely compromised the performance of these applications, especially for images captured in low-light environments.", + "bbox": [ + 511, + 560, + 906, + 606 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In recent years, several methods are developed for enhancing low-light LF images. Some methods [15, 28, 37] process stacked views of LF images to implicitly extract disparity information between views to take redundant information for reconstructing LF. Other methods employ advanced spatial-angular separation convolution [35, 37, 38] to extract redundant information between views for recovery. Although these methods are effective in handling LF data, their performance is limited since they do not take the view consistency into consideration, breaking the view consistency of the recovered image. Moreover, these methods learn a low-light image to normal-light image mapping to restore the low-light LF images. For different levels of illumination degradation, they require separate learning processes to train different models for recovery. Since the degradation levels are usually unknown for real-world images, it is difficult for these models to select appropriate parameters for restoration.", + "bbox": [ + 509, + 609, + 908, + 881 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To deal with dynamically degraded inputs, a simple yet", + "bbox": [ + 532, + 885, + 906, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: yflin@bjtu.edu.cn", + "bbox": [ + 109, + 886, + 334, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "7340", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "principled solution relying on the Retinex theory is proposed, which has been widely used for single-image [22, 23, 30]. These Retinex-based methods usually divide the light-up process into two stages: low contrast adjustment and corruptions (e.g., noise, artifact, color distortion, etc.) recovery. By estimating the illumination map of the scene for low contrast adjustment, the model is able to deal with dynamic illumination degradation. For LF images, the most straightforward way is to learn a brute-force mapping function for each view to apply one-to-one illumination adjustment. However, processing each LF view individually tends to recover each view with quite different illumination maps, resulting in varying degrees of color variation across different views. Another possible way is to estimate only one identical illumination adjustment gain for each view [39] or use a learnable parameter to compute the full illumination map [18], which is also improper since there exist disparities between views in one LF. This severely breaks the structural information inherent in LF images and limits the application of the following tasks. Therefore, how to keep the view-consistency in the estimated illumination maps in one LF becomes an important problem for scene-adaptive low-light LF image enhancement.", + "bbox": [ + 89, + 90, + 480, + 436 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this paper, we explore the view-consistency problem in scene-adaptive low-light LF image enhancement. We first analyze the disparity consistency between the illumination adjustment maps of different views. Based on the photometric consistency theory, we propose a self-supervised view-consistent illumination loss to keep the consistency between the estimated illumination maps. These illumination maps are estimated by combining the global brightness of the scene and the local brightness of each view, which are used to light up the low-light LF images. Subsequently, we perform corruption recovery on the light-up LF images to further suppress noise, artifacts, under-/overexposure, and color distortion. Comprehensive experiments are conduct in dynamic and fixed illumination environments. The related results are shown in Fig. 1, in which the proposed View-Consistent Network (VCNet) outperforms others over 7 db. More importantly, we also add our proposed illumination adjustment module to other models, labeled as “+E”. The experiments show that our proposed illumination adjustment module can be embedded into other methods and greatly improve their performance in terms of both image quality and view consistency preservation without increasing a large number of parameters.", + "bbox": [ + 91, + 439, + 482, + 785 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 799, + 232, + 814 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-Light Image Enhancement. The low-light image enhancement methods are mainly divided into two categories. The first category directly learns a brute-force mapping function from the low-light image to its normal-light counterpart. An exemplary illustration is the SID", + "bbox": [ + 89, + 824, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "approach [3], which utilizes an encoder-decoder architecture to effectively restore raw images captured under extremely low-light conditions. DID [19] introduced an innovative residual learning network explicitly crafted for the denoising of low-light images, surpassing the capabilities of SID. Guo et al. [9] designed a Zero-DCE network to employ dynamic range adjustment through pixel-wise and high-order curve estimations. However, these methods lack interpretability and theoretically proven properties. The methods in the second category are designed based on Retinex theory for restoration. Wei et al. [34] and follow-up works [40, 41] combine the Retinex decomposition with a multi-stage training pipeline. In the first training stage, these methods learn a network to decompose the images. Then in the second stage, they employ a different network to denoise the reflectance and adjust the illumination, respectively.", + "bbox": [ + 511, + 90, + 903, + 348 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-Light Video Enhancement. Different from low-light image enhancement, methods that enhance videos recorded under low-light conditions extract valuable complementary details from neighboring frames for better restoration. Wang et al. [26] delved into the field of physical modeling, utilizing an LSTM-based network to improve the quality of low-light videos. Based on Retinex theory, Wang et al. [24] introduced a comprehensive methodology to accomplish simultaneous enhancement of low-light conditions and noise reduction in video sequences. Even though these methods achieved remarkable results in enhancing two-dimensional low-light images and videos, it is still difficult for them to handle the four-dimensional LF dataset with regular sampling in the angular domain.", + "bbox": [ + 511, + 352, + 903, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Low-Light LF Image Enhancement. For LF images, some methods directly perform the restoration of low-light to normal illumination images. Lamba et al. [15] made substantial contributions in this regard by introducing a lowlight LF images dataset and a two-stage L3Fnet architecture. This architecture consists of a global representation block, responsible for encoding the intricate LF geometry, and a view reconstruction block that concentrates on restoring each distinct view. To address the challenge of enhancing dim LF images, they also introduced a three-stage network [14] that effectively utilizes global, local, and view-specific features inherent to low-light LF images. Wang et al. [28] introduced a coarse-to-fine restoration network to progressively restore the low-light LF input. In their network, three types of view stacks are fed into three streams to explore the redundant information of LFs. Recently, Wang et al. [29] proposed a novel feature extraction approach that parallelly processes features across multiple spatial scales. Lyu et al. [39] proposed a multi-stage architecture that mimics the optimization process of solving an inverse imaging problem, in which the illumination maps are estimated and employed in the unfolding process to produce", + "bbox": [ + 511, + 568, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "7341", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ae23a6b98edde3010d7d9a93ffe200085fe5003214ac1676168f34096dabf606.jpg", + "image_caption": [ + "Figure 2. Illustration of our self-supervised view-consistent loss. The reference point $\\mathbf{x}$ has the disparity $d(\\mathbf{x}) = 1$ . Compared with other points (labeled as red and purple), the corresponding point (labeled as orange) has the minimum variance. We propose to directly use the minimum variance as our objective function $\\ell_{vc}$ to minimize." + ], + "image_footnote": [], + "bbox": [ + 133, + 85, + 869, + 176 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the new enhanced result. There are also approaches to divide complex tasks into several intermediate tasks. Inspired by Retinex theory, Zhang et al. [37] proposed a learning-based decomposition-enhancement methodology that systematically dissects the intricate enhancement task into illumination adjustment and reflection denoising. Zhang et al. [39] separate the complex task into several intermediate tasks and use multiple heads to restore low-light images. They also formulate a synthesis pipeline to enrich the available LF data pairs, in which the ground truth illumination maps are generated and used to guide illumination adjustment. Although they consider the dynamically degraded inputs, their model needs the ground truth illumination value to give full supervision in the training process, which is unavailable for real-world images. Moreover, these methods do not consider the consistency constraints of the illumination maps, so that the view consistency of the recovered image is easily broken.", + "bbox": [ + 91, + 243, + 483, + 515 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. View-Consistent Network", + "text_level": 1, + "bbox": [ + 89, + 527, + 328, + 544 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "According to the two-plane model [16], a LF can be formulated as a 4D tensor $\\mathcal{L}(u,v,x,y)\\in \\mathbb{R}^{U\\times V\\times H\\times W}$ , where $\\mathbf{u} = (u,v)$ denote the angular coordinates, and $\\mathbf{x} = (x,y)$ represent the spatial coordinates. Commonly, the LF image is organized as a $U\\times V$ array of views that captured the same scene in one shot. One specific view at angular coordinate $(u,v)$ can be denoted as $\\mathcal{L}_{\\mathbf{u}}\\in \\mathbb{R}^{H\\times W}$ .", + "bbox": [ + 89, + 553, + 483, + 660 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. View Consistency Analyses", + "text_level": 1, + "bbox": [ + 89, + 667, + 334, + 684 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.1. Retinex Theory for Single View", + "text_level": 1, + "bbox": [ + 89, + 689, + 354, + 704 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Drawing upon the Retinex theory, each low-light view in LF can be mathematically expressed as the product of its reflectance and illumination components as:", + "bbox": [ + 89, + 708, + 483, + 753 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathbf {u}} = L _ {\\mathbf {u}} \\odot R _ {\\mathbf {u}}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 758, + 482, + 771 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $L_{\\mathbf{u}}$ , $R_{\\mathbf{u}}$ represents the illumination components, reflectance components with angular coordinates $\\mathbf{u}$ , respectively. $\\odot$ denotes the element-wise multiplication. Similar to [7, 11, 22], we regard $R_{\\mathbf{u}}$ as a well-exposed LF view. To light up $I_{\\mathbf{u}}$ , we element-wisely multiply the two sides of Eq. 1 by an illumination adjustment map $\\overline{L}_{\\mathbf{u}}$ such that $\\overline{L}_{\\mathbf{u}} \\odot L_{\\mathbf{u}} = 1$ as follow:", + "bbox": [ + 89, + 777, + 483, + 882 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathbf {u}} = I _ {\\mathbf {u}} \\odot \\bar {L} _ {\\mathbf {u}}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 238, + 886, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, for LF images, we need to estimate the illumination adjustment map $\\overline{L}_{\\mathbf{u}}$ for all views in one LF.", + "bbox": [ + 511, + 243, + 903, + 273 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1.2. View Consistency in Illumination Maps", + "text_level": 1, + "bbox": [ + 511, + 281, + 834, + 296 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since views in one LF are taken from different angulars of the same scene, in which the depth of the scene is fixed, the views are supposed to satisfy the disparity constraint. Suppose that the parameter of the camera is known, the disparity of two views $I_{\\mathbf{u_j}}$ and $I_{\\mathbf{u_i}}$ is calculated as:", + "bbox": [ + 511, + 299, + 906, + 377 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nD _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {i}}} (\\mathbf {x}) = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * \\frac {B * f}{Z (\\mathbf {x})} = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * d (\\mathbf {x}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 552, + 381, + 906, + 409 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B$ and $d$ represent the baseline length and the disparity between adjacent views, respectively. $f$ represents the focal length of the camera. $Z$ represents the depth of the scene. Therefore, the two captured low-light views satisfy the disparity constraint as:", + "bbox": [ + 511, + 412, + 905, + 488 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nI _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(I _ {\\mathbf {u} _ {\\mathrm {j}}}, D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 635, + 493, + 906, + 508 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $W$ denotes the warp operations according to the disparity $D_{\\mathbf{u_j}\\rightarrow \\mathbf{u_i}}$ . Similarly, the well-exposed LF views also satisfy the disparity constraint as:", + "bbox": [ + 511, + 513, + 905, + 559 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nR _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(R _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 632, + 563, + 905, + 579 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Therefore, according to Eq. 2, the illumination adjustment map should also satisfy the disparity constraint.", + "bbox": [ + 511, + 583, + 905, + 613 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 633, + 617, + 905, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In order to constrain the consistency of the illumination adjustment maps, the direct way is to minimize the distance between the illumination map $\\overline{L}_{\\mathbf{u_i}}$ and the warped illumination map $W(\\overline{L}_{\\mathbf{u_j}};D_{\\mathbf{u_j}\\to \\mathbf{u_i}})$ from different angular $\\mathbf{u_j}$ as:", + "bbox": [ + 511, + 638, + 905, + 700 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\sum_ {\\mathbf {u} _ {\\mathrm {j}}} \\sum_ {\\mathbf {u} _ {\\mathrm {i}}} | | W (\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}) - \\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} | |, \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 709, + 905, + 738 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{u_j},\\mathbf{u_i}\\in \\{(0,0),(0,1),\\dots,(U,V)\\}$ . To simplify the computation, we warp the illumination adjustment maps of different views towards the central view $\\overline{L}_{\\mathbf{u_c}}$ and construct these aligned views to a 4D array $\\mathbf{A}\\in \\mathbb{R}^{U\\times V\\times H\\times W}$ ..", + "bbox": [ + 511, + 739, + 905, + 801 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} \\left(\\mathbf {u} _ {\\mathbf {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathbf {j}}}; D _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {c}}}\\right), \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 619, + 805, + 905, + 821 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\mathbf{u}_{\\mathbf{c}} = \\left(\\frac{U}{2},\\frac{V}{2}\\right)$ . Then, we instead calculate the variance of the warped illumination adjustment maps to assess their disparity consistency as:", + "bbox": [ + 511, + 825, + 903, + 872 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\min \\sum_ {\\mathbf {x}} \\operatorname {V a r} (\\mathbf {A} (:, \\mathbf {x})), \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 640, + 878, + 905, + 904 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "7342", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/4b6f7d130f8f003ae6ffa931942e1e2661c72893d72447eea5d374d25cf36be6.jpg", + "image_caption": [ + "Figure 3. The proposed VCNet network consists of an illumination estimator VCAIA and a corruption restorer. In VCAIA, the illumination map is estimated and multiplied to the original input to light up the illumination. VCRM further recovers the details of the light-up image and outputs the final results." + ], + "image_footnote": [], + "bbox": [ + 96, + 88, + 480, + 210 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $Var(\\mathbf{A}(:,\\mathbf{x}))$ calculates the variance for each pixel along the angular dimension.", + "bbox": [ + 89, + 316, + 482, + 347 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "However, in real scenes, both the ground truth illumination adjustment map $\\overline{L}_{\\mathbf{u}}$ and the corresponding disparity $d$ is unavailable. Moreover, the disparity is also difficult to estimate using low-light images with much noise [12]. Therefore, the disparity constraint of the illumination adjustment maps as in Eq. 9 is difficult to implement.", + "bbox": [ + 89, + 347, + 483, + 439 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Self-Supervised View-Consistent Loss", + "text_level": 1, + "bbox": [ + 89, + 446, + 421, + 464 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we relax the disparity constraint and design a self-supervised view-consistent illumination loss to ensure the view consistency of illumination adjustment maps. Instead of constraining each point according to its true disparity as in Eq. 9, we only constrain each point according to one specific disparity value, no matter whether it is the true disparity or not. Specifically, we first set $N$ discrete candidate disparity values:", + "bbox": [ + 89, + 469, + 483, + 590 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd ^ {n} = d _ {\\min } + n * \\Delta d, n \\leq N \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 601, + 480, + 618 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\Delta d = (d_{max} - d_{min}) / N$ denotes the granularity of $d$ . $d_{min}$ and $d_{max}$ denote the minimum and maximum disparity of the scene. Subsequently, we warp each view to the center view according to all the possible disparity values $d^n$ , and the related 4D array is denoted as $\\mathbf{A}^n \\in \\mathbb{R}^{U \\times V \\times H \\times W}$ :", + "bbox": [ + 89, + 628, + 483, + 704 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {A} ^ {n} \\left(\\mathbf {u} _ {\\mathrm {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {c}}} ^ {n}\\right), \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 187, + 717, + 482, + 736 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $D_{\\mathbf{u}_i - > \\mathbf{u_c}}^n (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^n$", + "bbox": [ + 89, + 738, + 348, + 753 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then calculate the variance of the warped view for each point $\\mathbf{x}$ and for each possible disparity $n$ as $Var(\\mathbf{A}^n (:, \\mathbf{x}))$ , constructing the commonly known cost volume as in the disparity estimation methods [4, 13]. According to the photometric consistency, the variance under the true disparity is supposed to have the minimum value compared with other disparities. Therefore, we first find the specific disparity $d^{*}$ for each point that has the minimum variance:", + "bbox": [ + 89, + 753, + 483, + 875 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd ^ {*} (\\mathbf {x}) = \\arg \\min _ {d ^ {n}} (V a r \\left(\\mathbf {A} ^ {n} (:, \\mathbf {x}))\\right)). \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 176, + 882, + 482, + 904 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/4029e28c722ecfd037184699b092a95911c66b7f899962fb454d4021cc4d4286.jpg", + "image_caption": [ + "Figure 4. The estimated illumination EPI w/ and w/o $loss_{vc}$ . The former keeps the view consistent as the true EPI with normal light." + ], + "image_footnote": [], + "bbox": [ + 573, + 88, + 848, + 183 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We then minimize the variance related to the $d^{*}(\\mathbf{x})$ to make sure that all views are aligned in the disparity $d^{*}(\\mathbf{x})$ :", + "bbox": [ + 511, + 250, + 905, + 279 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\ell_ {v c} = \\sum_ {\\mathbf {x}} \\operatorname {V a r} \\left(\\mathbf {A} ^ {*} (:, \\mathbf {x})\\right), \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 625, + 289, + 905, + 318 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathbf{A}^*$ represents the 4D array with aligned views according $D_{\\mathbf{u_j - > u_c}}^* (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^* (\\mathbf{x})$ as in Eq. 11. An example is shown in Fig. 2, where the minimum variance is set as our objective function. In such a way, each pixel is assigned a disparity value that has the minimum variance to make the illumination adjustment value of each pixel as consistent as possible with other views.", + "bbox": [ + 511, + 328, + 905, + 434 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Adaptive Illumination Adjustment", + "text_level": 1, + "bbox": [ + 511, + 444, + 815, + 460 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Beginning with the input low-light LF image $\\mathcal{L}^{in}$ , we design a simple illumination estimator, View Consistent Adaptive Illumination Adjustment (VCAIA), $\\mathcal{E}$ for LFs to estimate the illumination adjustment map. Specifically, as in Fig. 3, we first estimate the global illumination adjustment matrix of the scene by integrating information from all views. The local brightness refinement is then developed to adjust the matrix according to the details of each view.", + "bbox": [ + 511, + 465, + 905, + 585 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to integrate the abundant information from all views and further reduce the effect of noise, we roughly calculate the average of all views across all color channels in LF image as $I_w \\in \\mathbb{R}^{H \\times W \\times 1}$ . Then following [1], we calculate the global illumination adjustment matrix as:", + "bbox": [ + 511, + 587, + 905, + 662 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{I ^ {l}}} _ {\\text {g l o b a l}} (\\mathbf {x}) = \\frac {\\log \\left(I _ {w} (\\mathbf {x}) / \\mu + 1\\right)}{\\log \\left(\\max \\left(I _ {w} (\\mathbf {x})\\right) / \\mu + 1\\right) \\cdot I _ {w} (\\mathbf {x})}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 542, + 672, + 903, + 704 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Subsequently, we replicate $\\overline{I}^{l}_{global}$ to all views and obtain the global illumination adjustment matrix $\\overline{\\mathcal{L}}^{l}_{global}$ . To further refine illumination estimation for each view, we employ a lightweight learnable network. Specifically, we first stack all views of the input low-light LF image and use the cascaded 3D spatial convolution layers to obtain the local brightness adjustment matrix $\\overline{\\mathcal{L}}^{l}_{local}$ of each view.", + "bbox": [ + 511, + 715, + 905, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "The local and global illumination adjustment matrix is combined to derive the final adjustment matrix and is then multiplied to $\\mathcal{L}^{in}$ to obtain the adjusted high-illumination image: $\\overline{\\mathcal{L}} = (\\overline{\\mathcal{L}^t}_{global} + \\overline{\\mathcal{L}^t}_{local})$ . By minimizing the proposed self-supervised loss $\\ell_{vc}$ on the estimated $\\overline{\\mathcal{L}}$ according", + "bbox": [ + 511, + 825, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "7343", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "table", + "img_path": "images/71514050662cd5af1c0d94bcc604859f5af552ec3700d411a0a7a26e1aa21a6b.jpg", + "table_caption": [ + "Table 1. Quantitative Comparisons of Low Light LF Image Enhancement Methods in L3F-Dynamic Task w/ w/o VCAIA." + ], + "table_footnote": [], + "table_body": "
MethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LFFT-DCU23.170.80030.095819.600.72010.170219.820.67180.2074
DRLF19.100.69470.188817.420.66000.240217.040.63220.2608
DRLF+E22.19 (+3.09)0.7636 (+0.0689)0.1124 (-0.0764)21.38 (+3.96)0.6976 (+0.0376)0.1676 (-0.0726)20.29 (+3.25)0.6326 (+0.0004)0.2265 (-0.0343)
MSP21.370.75600.133018.640.68980.207917.620.65430.2407
MSP+E22.80 (+1.43)0.7936 (+0.0376)0.1261 (-0.0069)21.81 (+3.17)0.7436 (+0.0538)0.1673 (-0.0406)21.10 (+3.48)0.6916 (+0.0373)0.2128 (-0.0279)
VCNet-E22.540.78350.125120.830.71030.201519.920.64500.2666
VCNet24.58 (+2.04)0.8220 (+0.0385)0.0897 (-0.0354)24.06 (+3.23)0.7652 (+0.0549)0.1292 (-0.0723)23.09 (+3.17)0.7032 (-0.0582)0.1981 (-0.0685)
", + "bbox": [ + 96, + 114, + 898, + 229 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "to Eq. 13, we constrain the calculated illumination adjustment maps to satisfy the view consistency even without the true disparities. Since no supervised images are needed, the illumination adjustment can be applied to various low-light LF enhancement networks. We show one estimated example in Fig. 4, where the Epipolar Plane Images (EPIs) are calculated by fixing $v$ in angular domain and $w$ in the spatial domain. Without the proposed $\\ell_{vc}$ , the lines in the EPIs of the estimated illumination maps become ambiguous and are inconsistent with the ground truth EPIs of the normallight LFs. In contrast, when we use $\\ell_{vc}$ to constrain the view consistency, the lines in estimated EPIs of illumination maps retain the direction as the ground truth EPIs.", + "bbox": [ + 89, + 247, + 480, + 444 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Recovery Module", + "text_level": 1, + "bbox": [ + 89, + 463, + 261, + 479 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "After illumination adjustment, the image $\\mathcal{L}^a = \\overline{\\mathcal{L}}\\odot \\mathcal{L}^{in}$ with a standard illumination level is obtained. However, the light-up process also amplifies the noise and artifacts, causing under-/over-exposure and color distortion. Therefore, similar to other retinae-based methods, a recovery module is needed to suppress noise and restore scene details. Some existing networks, such as the LF denoising model DRLF and the LF low-light enhancement model MSP, can be used as the recovery module.", + "bbox": [ + 89, + 488, + 482, + 625 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We also design a simple View Consistent Recovery Module (VCRM) to further import the inherent disparity constraint into the recovery process to enhance the recovery effect. VCRM adopts a multi-scale architecture. In each scale, the view consistent feature aggregation unit is implemented to explore the redundant information of the views to recover details. In each unit, we employ deformable convolution layers on EPIs to further integrate information between different views. Since the slope of lines in EPIs corresponds to disparity, we introduce the disparity constraint to help the model identify correct and consistent complementary information. Specifically, we use uniform offsets in both V-H and U-W spaces, so that the angular information can be efficiently incorporated and maintain epipolar consistency. Finally, we concatenate multi-scale features and use a decoder with a similar structure to the encoder to aggregate them and obtain the refined result $\\mathcal{L}^{out}$ . More detailed design is provided in the supplementary details.", + "bbox": [ + 89, + 628, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 246, + 730, + 263 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As in Fig. 3, the overview architecture of our VCNet primarily employs two VCAIA and VCRM. We train our VCNet by minimizing a hybrid loss: $\\ell_{total} = \\lambda_1\\ell_{vc} + \\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}$ . $\\ell_{rec}$ is the reconstruction loss, which calculates the L1 loss between the ground truth of well-posed image $\\mathcal{L}^{gt}$ and the recovered results $\\mathcal{L}^{out}$ of VCNet. The SSIM loss [33] $\\ell_{ssim}$ measures the structural similarity between $\\mathcal{L}^{gt}$ and $\\mathcal{L}^{out}$ . $\\ell_{vc}$ is the proposed self-supervised illumination loss in Sec.3, which is used to keep the view consistent of the estimated illumination maps $\\overline{\\mathcal{L}}^l$ . The three hyperparameters $\\lambda_1$ , $\\lambda_2$ and $\\lambda_3$ are empirically set to 0.5, 1, 0.5.", + "bbox": [ + 511, + 268, + 903, + 434 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Our model is implemented using the PyTorch framework and trained on a NVIDIA RTX A4000 GPU. The model is optimized using Adam optimization with $\\beta_{1} = 0.9$ and $\\beta_{2} = 0.999$ . The batch size is set to 1 and the learning rate is initially set to 0.0001. Data augmentation techniques such as random cropping, vertical and horizontal flipping are used. During training, the views are randomly cropped in patches of size $128 \\times 128$ . During testing, we use the original resolution LF data without applying data augmentation.", + "bbox": [ + 511, + 435, + 905, + 571 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 582, + 643, + 599 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 511, + 606, + 607, + 619 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We evaluate the proposed method using the publicly accessible Low-Light-Light-Field (L3F) Dataset and L3F-Wild Dataset from [15]. The L3F dataset comprises 27 distinct scenes, each of which was captured under three low-light conditions, labeled L3F-20, 50, 100, and one with normal lighting. The L3F-Wild is captured in almost 0 lux conditions and captured with $\\frac{1}{4} -\\frac{1}{15}$ second exposure time.", + "bbox": [ + 511, + 628, + 903, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "L3F-Fixed Task: Previous LF-based low-light enhancement methods trained their models under fixed illumination conditions. Specifically, following [15], 18 scenes under the same illumination condition are selected for training, while the remaining 9 scenes are reserved for testing. In the fixed illumination conditions, 3 models are specifically trained for L3F-20, 50, 100 in different illumination conditions. We denote it as the L3F-Fixed task.", + "bbox": [ + 511, + 734, + 903, + 853 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "L3F-Dynamic Task: To further assess the model's capability to adapt to various illumination conditions, we establish a dataset with dynamic illumination as L3F-Dynamic.", + "bbox": [ + 511, + 854, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "7344", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e2e38008af38050d261af8f589c35515d89bd0039a5ecd1d99ba4aee0b88eb7.jpg", + "image_caption": [ + "Figure 5. Visual comparison on $L3F$ -Dynamic task, where the central view and EPIs of the recovery LFs are shown. Other algorithms either generate over-exposed and noisy images or introduce black spot corruptions and unnatural artifacts. Our VCNet restores well-exposed structural contents and textures. By adding our VCAIA module, the recovered results of all models are closer to ground truth." + ], + "image_footnote": [], + "bbox": [ + 109, + 88, + 883, + 393 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Specifically, we combine images from the L3F-20, 50, 100 datasets. We use all training sets from the three different illumination conditions (including $18 \\times 3$ scenes) to train a single model. The single trained model is supposed to handle dynamic illumination conditions. We use the same testing dataset as L3F-Fixed to evaluate its performance. Compared to L3F-Fixed task, L3F-Dynamic task is more difficult, which needs to train one single model that is able to handle all different illumination conditions.", + "bbox": [ + 88, + 462, + 482, + 597 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Compare with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 89, + 606, + 434, + 622 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our VCNet and several SOTA methods, including 3 recent published single-frame retinae-based methods (RFormer [22], Pydiff [42] and DarkIR [6]) and 9 LF-based methods (L3F-Net [15], FER-Net [14], DRLF [10], MSP-3 [27], IINet [17], DistgSSR [31], PMSNet [29], LRT [39], LFLL-DCU [18]). The illumination estimator $\\mathcal{E}$ is applied to the LF denoising model DRLF and low-light enhancement model MSP, denoted as DRLF+ $\\mathcal{E}$ , MSP+ $\\mathcal{E}$ . Specifically, we embed the illumination estimator $\\mathcal{E}$ into the beginning of these two models and add the $\\ell_{vc}$ in their training strategy. Our VCNet- $\\mathcal{E}$ is trained by deleting $\\mathcal{E}$ module and only minimizing $\\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}$ . The reference-based metric PSNR, SSIM [32] and LPIPS [36] and nonreference metric NIQE [21] and BRISQUE [20] are chosen as the evaluation metrics.", + "bbox": [ + 91, + 628, + 482, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dynamic Illumination Task. We first assess our model's ability to adapt to dynamic illumination. Specifically, we train each method using the L3F-Dynamic training dataset", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/9caccb28d4d86c68714d74d4651bdeb81a096c22f82b4f93305d351e5f652c1b.jpg", + "table_caption": [ + "Table 2. Quantitative Comparisons with Single-Frame-based Low Light Enhancement Methods on the L3F-dynamic Task." + ], + "table_footnote": [], + "table_body": "
MethodL3F-20L3F-50L3F-100
PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑
RFormer20.36 / 0.77 / 0.342 / 50.9419.30 / 0.70 / 0.493 / 39.1917.13 / 0.62 / 0.615 / 30.79
RFormer*27.29 / 0.87 / 0.163 / 67.2623.81 / 0.77 / 0.294 / 42.3122.51 / 0.70 / 0.374 / 34.84
Pydiff19.74 / 0.71 / 0.406 / 48.7219.13 / 0.62 / 0.535 / 36.3018.47 / 0.49 / 0.757 / 31.14
Pydiff*23.42 / 0.82 / 0.256 / 60.1122.74 / 0.76 / 0.331 / 40.9622.24 / 0.70 / 0.437 / 33.51
DarkIR20.11 / 0.76 / 0.315 / 51.6317.24 / 0.65 / 0.474 / 35.3715.09 / 0.58 / 0.621 / 15.31
VCNet24.58 / 0.82 / 0.089 / 79.0724.06 / 0.77 / 0.129 / 78.4623.09 / 0.70 / 0.198 / 76.80
", + "bbox": [ + 522, + 498, + 893, + 579 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "in Sec. 4.1 and the results are presented in Table 1. By adding the illumination estimator $\\mathcal{E}$ , all three models exhibit around 3dB PSNR improvement by considering the adaptive illumination estimation. Since LFLL-DCU only uses one learnable parameter to estimate the illumination map, their model performs well on the L3F-20 data but poorly on L3F-50 and L3F-100. Our VCNet outperforms the current SOTA methods MSP, DRLF, and LFLL-DCU with more than $2\\sim 6$ dB PSNR improvement. Fig. 5 displays the visual results comparison, where all visual effects are improved by combining the $\\mathcal{E}$ . This clearly demonstrates the effectiveness of our illumination estimator $\\mathcal{E}$ in adaptively adjusting to varying illumination conditions. The results of VCNet stand out by producing more realistic overall color compared with the ground truth. Other methods restore blurred scene details, while our VCNet recovers the clearer and richer detailed texture in the recovered images.", + "bbox": [ + 509, + 597, + 906, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We then make a comparison with single-frame-based methods in Table 2, which claim to handle dynamic lighting conditions. Their models are trained on a large-scale", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "7345", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/3c230fdd3242aa0ff6476bda06fd48908e7c979cccee6a3fbac76470bb7d2f61.jpg", + "image_caption": [ + "Figure 6. Visual comparison on the $L3F$ -Wild dataset, where the central view of LFs is shown. Other methods either collapse by noise, distort color, or under-/over-exposed images. Our VCNet effectively removes the noise and reconstructs well-exposed image details." + ], + "image_footnote": [], + "bbox": [ + 133, + 88, + 866, + 232 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/45ea8be0c053691e908e1184c7cc022222a04589727adb62dc007f39b9b6b3e2.jpg", + "table_caption": [ + "Table 3. Quantitative Comparison on Real-world L3F-Wild Data." + ], + "table_footnote": [], + "table_body": "
MethodsRFormer*Pydiff*DRLFDRLF+EMSPMSP+EVCNet-EVCNet
NIQE ↓15.6717.8219.9718.7818.8218.2317.8714.22
BRISQUE↓31.2036.1232.8030.5430.2029.4230.1125.98
TIMES↓39.84100.1216.2216.263.893.935.996.33
", + "bbox": [ + 93, + 295, + 483, + 354 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "single-image dataset and we carefully fine-tune their methods using L3F-Dynamic training dataset (labelled as *), where each view of LF is considered a sample during the training. Our method outperforms others under most lighting conditions, especially in perceptual quality. Although RFormer* achieves relatively high PSNR and SSIM values in the L3F-20 dataset, it suffers from noticeable blurry artifacts, as shown in Fig. 5.", + "bbox": [ + 89, + 369, + 482, + 489 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "To further demonstrate the generalization ability, we applied the methods trained on the $L3F$ -Dynamic dataset to another real-world low-light LF data $L3F$ -Wild. The results are shown in Fig. 6 and Table 3. Due to the absence of redundant information from different views, the results of single-frame methods are relatively dark, and the restored details suffer from blurring. For LF-based methods, DRLF is observed to cause a noticeable color shift. The results of MSP and LFLL-DCU suffer from residual noise and local over-expose. Combining with the $\\mathcal{E}$ , the recovered results of MSP and DNLF have higher color fidelity, clearer structural details and less noise. In contrast, our results achieve high perceptual quality with clear structures, realistic color and less noises. Our VCAIA module $\\mathcal{E}$ is also proved that can enhance the perceptual quality of other methods.", + "bbox": [ + 89, + 491, + 482, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Fixed Illumination Task. We then evaluate all the methods under the fixed illumination condition by separately training 3 modals using the $L3F$ -Fixed dataset. The results using LFs with angular resolution $5 \\times 5$ and $8 \\times 8$ are illustrated in Table 4, respectively. We retrain the LRT [39] and LFLL-DCU [18] for comparison. When compared with single-frame-based methods, LF-based methods demonstrate higher performance, highlighting the utility of redundant LF information in low-light enhancement. With the increase of the input image views, the recovery effect of the LF models is better because it provides more redundant information to the network. Note that in PMSNet,", + "bbox": [ + 89, + 719, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/8a6a9570eeddb4a08d5b6e823744d97548ea560e7b6e75b1fa664337b85fa1d7.jpg", + "table_caption": [ + "Table 4. Quantitative Comparisons on L3F-fixed Task." + ], + "table_footnote": [], + "table_body": "
Ang ResMethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
8L3F-Net24.780.82623.210.75221.830.683
IINet25.400.86224.140.81122.130.721
DistgSSR25.440.86424.180.81022.960.755
PMSNet27.030.86624.520.81222.740.762
VCNet26.840.87125.410.82323.580.772
5L3F-Net24.980.81423.370.75321.950.669
FER-Net25.310.82323.400.76621.730.688
LRT25.970.81523.990.74222.150.691
LFLL-DCU25.420.84224.030.78722.280.702
DRLF24.780.806422.520.721721.510.6543
DRLF+ε24.990.817923.060.732922.190.6681
MSP25.480.830124.100.775922.750.7133
MSP+ε25.730.843024.680.783123.060.7191
VCNet-ε25.870.837224.350.789722.690.7204
VCNet26.610.847325.130.792123.240.7223
", + "bbox": [ + 555, + 297, + 867, + 500 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in addition to the original $8 \\times 8$ views, views from two more rows and columns are included in the feature extraction. Among LF-based methods, our VCNet significantly outperforms SOTA methods on most of the $L3F$ -fixed task, demonstrating its ability to better handle low-illumination degradation and produce superior restoration results. After adding the illumination estimator $\\mathcal{E}$ , all models also achieve around a 0.2-0.8 dB PSNR improvement.", + "bbox": [ + 511, + 521, + 906, + 642 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "View Consistency Comparison. View consistency holds great significance in LF enhancement as it directly impacts the performance of subsequent LF-based applications like depth estimation. To verify the effectiveness of our VCAIA $\\mathcal{E}$ in maintaining consistency in recovery results, we present the parallax content Precision-Recall (PR) curves used in [5] for evaluation. In one LF, each view is subtracted from the central view, and different threshold values are chosen to calculate binary edge maps. The PR curves show how closely the binary edge maps of the reconstructed LFs are match to the ground truth. We use models trained on dynamic illumination for comparison. In Fig. 7, when all the methods add the illumination adjustment, the PR curve of the recovered image is improved, demonstrating that our $\\mathcal{E}$ can achieve view-consistent illumination adjustment. We further calculate the Average Precision (AP) by measuring the area under the PR curves for quantitative", + "bbox": [ + 511, + 643, + 908, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "7346", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/79a106eca2f9b6bb151d61406fd641448592f227c717d3b1b3ddac630ae229b9.jpg", + "image_caption": [ + "Figure 7. The PR curves map comparison of L3F-20, 50, 100 dataset, in which how well the views consistency preserved is shown. With $\\mathcal{E}$ , the PR curves of all methods are better." + ], + "image_footnote": [], + "bbox": [ + 93, + 87, + 478, + 162 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "comparison. As in Table 2, for the single-frame methods, it is difficult to recover views with consistent illumination, so that the AP is relatively low even with high PSNR and SSIM. By contrast, our model exhibits notably higher AP values, especially under the challenging low-light conditions of the L3F-100 dataset. The comparison with LF-based methods is shown in Fig. 1. Our model consistently achieves the highest AP values, showcasing the superiority of VCNet in terms of low-light enhancement and consistency preservation. The EPIs of the recovered images are also visualized in Fig. 5, in which the orientation of lines corresponds to the disparity information. It is also clear that our model excels in maintaining the view consistency. More results are provided in the supplementary materials.", + "bbox": [ + 88, + 228, + 482, + 439 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Running time. We further show the running time of the models using the size of the scatter in Fig. 1 and Table 3. Our model runs faster (6.3s) than DRLF (16.2s) but is slightly slower than MSP (3.9s). The computation time of $\\mathcal{E}$ is 0.04s on the A5000 for a $5 \\times 5 \\times 642 \\times 432$ input, which is lightweight to add $\\mathcal{E}$ to different models.", + "bbox": [ + 89, + 439, + 483, + 530 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 540, + 243, + 556 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/424d34fe8162df3c97d89c49d477fc6b10a1710b8024c3d2d11a6928af53c14f.jpg", + "table_caption": [ + "Table 5. The Ablation Study on L3F-Dynamic Task." + ], + "table_footnote": [], + "table_body": "
No.LocalGlobal\\( \\ell_{vc} \\)PSNRSSIMLPIPS
(1)×××20.760.700.2036
(2)×√×22.780.750.1469
(3)√√×23.590.750.1420
(4)√√√23.840.760.1395
", + "bbox": [ + 169, + 594, + 405, + 665 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conduct a series of ablation studies under dynamic illumination conditions with $5 \\times 5$ LF images. To ensure fairness, we keep the parameters of the variant models close to VCNet by adjusting the feature channels. It is clear that the performance of the VCNet without $\\mathcal{E}$ (model-1) is the worst. Compared with model-1, model-2 using the global brightness refinement performs better. Meanwhile, with the local brightness adaptive adjustment branch, model-3 increases $0.8\\mathrm{dB}$ in the PSNR values than model-2. We further validate the effectiveness of the loss $\\ell_{vc}$ . Comparing model-4 with model-3, the PSNR increases nearly $0.4\\mathrm{db}$ by adding the $\\ell_{vc}$ to supervise the illumination adjustment matrix $\\overline{\\mathcal{L}^l}$ . This shows that $\\ell_{vc}$ can perform consistency constraints on $\\overline{\\mathcal{L}^l}$ to provide higher-quality light-up images.", + "bbox": [ + 89, + 674, + 482, + 885 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further visualize the adaptive illumination adjust", + "bbox": [ + 109, + 885, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c7db31cb773bc26ebebcca0c4e7cc7a1a92e5c1f644f02c9423305cdc95a46b5.jpg", + "image_caption": [ + "Figure 8. Visualization of the illumination maps in $L3F$ -Dynamic task. For darker scenes, the values in the estimated illumination adjustment map is larger." + ], + "image_footnote": [], + "bbox": [ + 516, + 87, + 898, + 200 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/c1c25df274ae9ab220fd65b2f8f71395366a054a2b5cbe128495740dc1b1ce1c.jpg", + "image_caption": [ + "Figure 9. Visual comparison of the enhanced result (top row) and input central view (second row) of one scene with different low-light levels in the L3F-Wild dataset." + ], + "image_footnote": [], + "bbox": [ + 517, + 250, + 901, + 361 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "ment matrix $\\overline{\\mathcal{L}_l}$ and the related recovered results in different illumination environments in Fig. 8. The model is trained using the $L3F$ -dynamic dataset. For input with an extended exposure time, the generated illumination adjustment matrix tends to be relatively darker. Conversely, as the exposure time diminishes, the input images become darker and the values of the adjustment matrix gradually increase. We also illustrate the performance of our VCNet in processing images with disparate brightness levels from the L3F-Wild dataset in Fig. 9. Our method consistently achieves superior enhancement quality across input scenes with varying levels of brightness. This consistency ensures effective low-light image restoration across varying illumination conditions.", + "bbox": [ + 511, + 421, + 906, + 619 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 632, + 633, + 647 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduced a novel framework for scene-adaptively low-light enhancement in light field(LF) images. First, by analyzing Retinex theory and extending it to the LF, we introduce a self-supervised loss to ensure that the calculated illumination map adheres to the consistency relationship. Then, we design an illumination estimator with global adjustment and local refinement to perceive the illumination of the scene. Finally, we design a corruption restorer and import the consistency constraint into the feature aggregation process for better recovery. Extensive quantitative and qualitative experiments demonstrate that our VC-Net significantly outperforms state-of-the-art methods on both fixed and dynamic illumination enhancement tasks.", + "bbox": [ + 511, + 657, + 906, + 854 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032).", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "7347", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 89, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Hyunchan Ahn, Byungjik Keum, Daehoon Kim, and Hwang Soo Lee. Adaptive local tone mapping based on retina for high dynamic range images. In Proceedings of the International Conference on Consumer Electronics (ICCE), pages 153-156, 2013. 4", + "[2] Tom E. Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):972-986, 2012. 1", + "[3] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 3291-3300, 2018. 2", + "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4", + "[5] Jie Chen, Junhui Hou, and Lap-Pui Chau. Light field denoising via anisotropic parallax analysis in a cnn framework. IEEE Signal Processing Letters, 25(9):1403-1407, 2018. 7", + "[6] Daniel Feijoo, Juan C. Benito, Alvaro Garcia, and Marcos V. Conde. Darkir: Robust low-light image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 10879-10889, 2025. 6", + "[7] Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2782-2790, 2016. 3", + "[8] Chunle Guo, Jing Jin, Junhui Hou, and Jie Chen. Accurate light field depth estimation via an occlusion-aware network. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1", + "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1777-1786, 2020. 2", + "[10] Mantang Guo, Junhui Hou, Jing Jin, Jie Chen, and LapPui Chau. Deep spatial-angular regularization for light field imaging, denoising, and super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6094-6110, 2022. 6", + "[11] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing, 26(2):982-993, 2017. 3", + "[12] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 19-34, 2017. 4", + "[13] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 4", + "[14] Mohit Lamba and Kaushik Mitra. Fast and efficient restoration of extremely dark light fields. In Proceedings of the Win-" + ], + "bbox": [ + 93, + 114, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ter Conference on Applications of Computer Vision (WACV), pages 3152-3161, 2022. 2, 6", + "[15] Mohit Lamba, Kranthi Kumar Rachavarapu, and Kaushik Mitra. Harnessing multi-view perspective of light fields for low-light imaging. IEEE Transactions on Image Processing, 30:1501-1513, 2021. 1, 2, 5, 6", + "[16] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. ACM, 1996. 3", + "[17] Gaosheng Liu, Huanjing Yue, Jiamin Wu, and Jingyu Yang. Intra-inter view interaction network for light field image super-resolution. IEEE Transactions on Multimedia, 25: 256-266, 2023. 6", + "[18] Xianqiang Lyu and Junhui Hou. Enhancing low-light light field images with a deep compensation unfolding network. IEEE Transactions on Image Processing, 33:4131-4144, 2024. 2, 6, 7", + "[19] Paras Maharjan, Li Li, Zhu Li, Ning Xu, Chongyang Ma, and Yue Li. Improving extreme low-light image denoising via residual learning. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 916-921, 2019. 2", + "[20] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6", + "[21] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6", + "[22] Junxiang Ruan, Xiangtao Kong, Wenqi Huang, and Wenming Yang. Retformer: Retinex-based enhancement in transformer for low-light image. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2, 3, 6", + "[23] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 6842-6850, 2019. 2", + "[24] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the International Conference on Computer Vision (ICCV), pages 9680-9689, 2021. 2", + "[25] Tingxian Wang, Alexei A. Efros, and Ravi Ramamoorthi. Depth estimation with occlusion modeling using light-field cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:2170-2181, 2016. 1", + "[26] Wei Wang, Xin Chen, Cheng Yang, Xiang Li, Xuemei Hu, and Tao Yue. Enhancing low light videos by exploring high sensitivity camera noise. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4110-4118, 2019. 2", + "[27] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 6" + ], + "bbox": [ + 517, + 93, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "7348", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[28] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 1, 2", + "[29] Xingzheng Wang, Kaiqiang Chen, Zixuan Wang, and Henhao Huang. Pmsnet: Parallel multi-scale network for accurate low-light light-field image enhancement. IEEE Transactions on Multimedia, 26:2041-2055, 2024. 2, 6", + "[30] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the International Conference on Computer Vision (ICCV), pages 13128-13137, 2023. 2", + "[31] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2023. 6", + "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6", + "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5", + "[34] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2", + "[35] Henry Wing Fung Yeung, Junhui Hou, Xiaoming Chen, Jie Chen, Zhibo Chen, and Yuk Ying Chung. Light field spatial super-resolution using deep efficient spatial-angular separable convolution. IEEE Transactions on Image Processing, 28(5):2319-2330, 2018. 1", + "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2018. 6", + "[37] Shansi Zhang and Edmund Y. Lam. An effective decomposition-enhancement method to restore light field images captured in the dark. Signal Processing, 189:108279, 2021. 1, 3", + "[38] Shuo Zhang, Youfang Lin, and Hao Sheng. Residual networks for light field image super-resolution. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 11046-11055, 2019. 1", + "[39] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Lrt: An efficient low-light restoration transformer for dark light field images. IEEE Transactions on Image Processing, 32:4314-4326, 2023. 2, 3, 6, 7", + "[40] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1632-1640, New York, NY, USA, 2019. ACM. 2" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[41] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision., 129(4):1013-1037, 2021. 2", + "[42] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 6" + ], + "bbox": [ + 514, + 92, + 905, + 175 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "7349", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + } +] \ No newline at end of file diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_model.json b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_model.json new file mode 100644 index 0000000000000000000000000000000000000000..1a0901564f6cae14a60d2a52e672f8ac63026211 --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_model.json @@ -0,0 +1,2024 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.202, + 0.131, + 0.798, + 0.176 + ], + "angle": 0, + "content": "Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement" + }, + { + "type": "text", + "bbox": [ + 0.152, + 0.204, + 0.844, + 0.258 + ], + "angle": 0, + "content": "Shuo Zhang, Chen Gao, Youfang Lin* \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science& Technology, Beijing Jiaotong University, Beijing, China." + }, + { + "type": "text", + "bbox": [ + 0.324, + 0.26, + 0.672, + 0.275 + ], + "angle": 0, + "content": "{zhangshuo, gaochen, yflin}@bjtu.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.327 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.341, + 0.486, + 0.735 + ], + "angle": 0, + "content": "Light Field (LF) images captured under low illumination conditions typically exhibit low quality. Recent learning-based methods for low-light LF images enhancement are generally tailored to specific illumination inputs, limiting their performance in real-world scenes. Moreover, how to maintain the inherent view-consistency in the enhanced images also remains a difficult problem. In this paper, we propose to explore the view consistency for scene-adaptive low-light LF enhancement. We first analyze the view consistency for LF illumination maps and design a self-supervised view-consistent loss to keep the consistency between the illumination maps of different views in LFs. To enhance the model's perception of illumination, we combine both global and local information to estimate the illumination map, which is easily plugged into other models. Subsequently, we use the illumination maps to light up the low-light LF images and restore the corruption to produce the final enhanced image. Experiments demonstrate that our View-Consistent Network (VCNet) outperforms state-of-the-art methods on real-world low-light LF datasets in both fixed lighting conditions and dynamic lighting conditions. Our proposed illumination adjustment is also demonstrated that can comprehensively improve the performance of existing methods in terms of both image quality and view consistency. Code is available at: https://github.com/shuozh/VCNet" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.763, + 0.222, + 0.779 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.787, + 0.485, + 0.879 + ], + "angle": 0, + "content": "In contrast to conventional cameras, Light Field (LF) cameras have the capability to simultaneously capture both directional and intensity of all light rays in a single shot. Because of the structural information inherent in LF images, they are well-suited for various applications, including digital refocus [2], depth estimation [8], and 3D reconstruc" + }, + { + "type": "image", + "bbox": [ + 0.582, + 0.308, + 0.839, + 0.445 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.454, + 0.908, + 0.54 + ], + "angle": 0, + "content": "Figure 1. Performance comparison in dynamic illumination tasks. The size of the points represents the run time to process one LF. Our VCNet achieves the best results in terms of detail restoration with higher PSNR and views consistency preservation with higher AP. By adding our proposed illumination adjustment module, labeled as “+E”, the performances of all methods are improved." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.561, + 0.907, + 0.607 + ], + "angle": 0, + "content": "tion [25]. Nevertheless, the low-quality LF images severely compromised the performance of these applications, especially for images captured in low-light environments." + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.61, + 0.909, + 0.882 + ], + "angle": 0, + "content": "In recent years, several methods are developed for enhancing low-light LF images. Some methods [15, 28, 37] process stacked views of LF images to implicitly extract disparity information between views to take redundant information for reconstructing LF. Other methods employ advanced spatial-angular separation convolution [35, 37, 38] to extract redundant information between views for recovery. Although these methods are effective in handling LF data, their performance is limited since they do not take the view consistency into consideration, breaking the view consistency of the recovered image. Moreover, these methods learn a low-light image to normal-light image mapping to restore the low-light LF images. For different levels of illumination degradation, they require separate learning processes to train different models for recovery. Since the degradation levels are usually unknown for real-world images, it is difficult for these models to select appropriate parameters for restoration." + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.886, + 0.907, + 0.902 + ], + "angle": 0, + "content": "To deal with dynamically degraded inputs, a simple yet" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.887, + 0.336, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: yflin@bjtu.edu.cn" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "7340" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.438 + ], + "angle": 0, + "content": "principled solution relying on the Retinex theory is proposed, which has been widely used for single-image [22, 23, 30]. These Retinex-based methods usually divide the light-up process into two stages: low contrast adjustment and corruptions (e.g., noise, artifact, color distortion, etc.) recovery. By estimating the illumination map of the scene for low contrast adjustment, the model is able to deal with dynamic illumination degradation. For LF images, the most straightforward way is to learn a brute-force mapping function for each view to apply one-to-one illumination adjustment. However, processing each LF view individually tends to recover each view with quite different illumination maps, resulting in varying degrees of color variation across different views. Another possible way is to estimate only one identical illumination adjustment gain for each view [39] or use a learnable parameter to compute the full illumination map [18], which is also improper since there exist disparities between views in one LF. This severely breaks the structural information inherent in LF images and limits the application of the following tasks. Therefore, how to keep the view-consistency in the estimated illumination maps in one LF becomes an important problem for scene-adaptive low-light LF image enhancement." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.44, + 0.483, + 0.786 + ], + "angle": 0, + "content": "In this paper, we explore the view-consistency problem in scene-adaptive low-light LF image enhancement. We first analyze the disparity consistency between the illumination adjustment maps of different views. Based on the photometric consistency theory, we propose a self-supervised view-consistent illumination loss to keep the consistency between the estimated illumination maps. These illumination maps are estimated by combining the global brightness of the scene and the local brightness of each view, which are used to light up the low-light LF images. Subsequently, we perform corruption recovery on the light-up LF images to further suppress noise, artifacts, under-/overexposure, and color distortion. Comprehensive experiments are conduct in dynamic and fixed illumination environments. The related results are shown in Fig. 1, in which the proposed View-Consistent Network (VCNet) outperforms others over 7 db. More importantly, we also add our proposed illumination adjustment module to other models, labeled as “+E”. The experiments show that our proposed illumination adjustment module can be embedded into other methods and greatly improve their performance in terms of both image quality and view consistency preservation without increasing a large number of parameters." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.8, + 0.233, + 0.815 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Low-Light Image Enhancement. The low-light image enhancement methods are mainly divided into two categories. The first category directly learns a brute-force mapping function from the low-light image to its normal-light counterpart. An exemplary illustration is the SID" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.349 + ], + "angle": 0, + "content": "approach [3], which utilizes an encoder-decoder architecture to effectively restore raw images captured under extremely low-light conditions. DID [19] introduced an innovative residual learning network explicitly crafted for the denoising of low-light images, surpassing the capabilities of SID. Guo et al. [9] designed a Zero-DCE network to employ dynamic range adjustment through pixel-wise and high-order curve estimations. However, these methods lack interpretability and theoretically proven properties. The methods in the second category are designed based on Retinex theory for restoration. Wei et al. [34] and follow-up works [40, 41] combine the Retinex decomposition with a multi-stage training pipeline. In the first training stage, these methods learn a network to decompose the images. Then in the second stage, they employ a different network to denoise the reflectance and adjust the illumination, respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.353, + 0.905, + 0.563 + ], + "angle": 0, + "content": "Low-Light Video Enhancement. Different from low-light image enhancement, methods that enhance videos recorded under low-light conditions extract valuable complementary details from neighboring frames for better restoration. Wang et al. [26] delved into the field of physical modeling, utilizing an LSTM-based network to improve the quality of low-light videos. Based on Retinex theory, Wang et al. [24] introduced a comprehensive methodology to accomplish simultaneous enhancement of low-light conditions and noise reduction in video sequences. Even though these methods achieved remarkable results in enhancing two-dimensional low-light images and videos, it is still difficult for them to handle the four-dimensional LF dataset with regular sampling in the angular domain." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.569, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Low-Light LF Image Enhancement. For LF images, some methods directly perform the restoration of low-light to normal illumination images. Lamba et al. [15] made substantial contributions in this regard by introducing a lowlight LF images dataset and a two-stage L3Fnet architecture. This architecture consists of a global representation block, responsible for encoding the intricate LF geometry, and a view reconstruction block that concentrates on restoring each distinct view. To address the challenge of enhancing dim LF images, they also introduced a three-stage network [14] that effectively utilizes global, local, and view-specific features inherent to low-light LF images. Wang et al. [28] introduced a coarse-to-fine restoration network to progressively restore the low-light LF input. In their network, three types of view stacks are fed into three streams to explore the redundant information of LFs. Recently, Wang et al. [29] proposed a novel feature extraction approach that parallelly processes features across multiple spatial scales. Lyu et al. [39] proposed a multi-stage architecture that mimics the optimization process of solving an inverse imaging problem, in which the illumination maps are estimated and employed in the unfolding process to produce" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.515, + 0.957 + ], + "angle": 0, + "content": "7341" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.087, + 0.87, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.186, + 0.908, + 0.229 + ], + "angle": 0, + "content": "Figure 2. Illustration of our self-supervised view-consistent loss. The reference point \\(\\mathbf{x}\\) has the disparity \\(d(\\mathbf{x}) = 1\\). Compared with other points (labeled as red and purple), the corresponding point (labeled as orange) has the minimum variance. We propose to directly use the minimum variance as our objective function \\(\\ell_{vc}\\) to minimize." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.244, + 0.485, + 0.516 + ], + "angle": 0, + "content": "the new enhanced result. There are also approaches to divide complex tasks into several intermediate tasks. Inspired by Retinex theory, Zhang et al. [37] proposed a learning-based decomposition-enhancement methodology that systematically dissects the intricate enhancement task into illumination adjustment and reflection denoising. Zhang et al. [39] separate the complex task into several intermediate tasks and use multiple heads to restore low-light images. They also formulate a synthesis pipeline to enrich the available LF data pairs, in which the ground truth illumination maps are generated and used to guide illumination adjustment. Although they consider the dynamically degraded inputs, their model needs the ground truth illumination value to give full supervision in the training process, which is unavailable for real-world images. Moreover, these methods do not consider the consistency constraints of the illumination maps, so that the view consistency of the recovered image is easily broken." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.528, + 0.33, + 0.545 + ], + "angle": 0, + "content": "3. View-Consistent Network" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.554, + 0.484, + 0.661 + ], + "angle": 0, + "content": "According to the two-plane model [16], a LF can be formulated as a 4D tensor \\(\\mathcal{L}(u,v,x,y)\\in \\mathbb{R}^{U\\times V\\times H\\times W}\\), where \\(\\mathbf{u} = (u,v)\\) denote the angular coordinates, and \\(\\mathbf{x} = (x,y)\\) represent the spatial coordinates. Commonly, the LF image is organized as a \\(U\\times V\\) array of views that captured the same scene in one shot. One specific view at angular coordinate \\((u,v)\\) can be denoted as \\(\\mathcal{L}_{\\mathbf{u}}\\in \\mathbb{R}^{H\\times W}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.669, + 0.336, + 0.685 + ], + "angle": 0, + "content": "3.1. View Consistency Analyses" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.69, + 0.356, + 0.705 + ], + "angle": 0, + "content": "3.1.1. Retinex Theory for Single View" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.709, + 0.484, + 0.755 + ], + "angle": 0, + "content": "Drawing upon the Retinex theory, each low-light view in LF can be mathematically expressed as the product of its reflectance and illumination components as:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.759, + 0.483, + 0.772 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathbf {u}} = L _ {\\mathbf {u}} \\odot R _ {\\mathbf {u}}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.778, + 0.484, + 0.883 + ], + "angle": 0, + "content": "where \\(L_{\\mathbf{u}}\\), \\(R_{\\mathbf{u}}\\) represents the illumination components, reflectance components with angular coordinates \\(\\mathbf{u}\\), respectively. \\(\\odot\\) denotes the element-wise multiplication. Similar to [7, 11, 22], we regard \\(R_{\\mathbf{u}}\\) as a well-exposed LF view. To light up \\(I_{\\mathbf{u}}\\), we element-wisely multiply the two sides of Eq. 1 by an illumination adjustment map \\(\\overline{L}_{\\mathbf{u}}\\) such that \\(\\overline{L}_{\\mathbf{u}} \\odot L_{\\mathbf{u}} = 1\\) as follow:" + }, + { + "type": "equation", + "bbox": [ + 0.24, + 0.887, + 0.483, + 0.902 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathbf {u}} = I _ {\\mathbf {u}} \\odot \\bar {L} _ {\\mathbf {u}}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.244, + 0.905, + 0.275 + ], + "angle": 0, + "content": "Therefore, for LF images, we need to estimate the illumination adjustment map \\(\\overline{L}_{\\mathbf{u}}\\) for all views in one LF." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.282, + 0.835, + 0.297 + ], + "angle": 0, + "content": "3.1.2. View Consistency in Illumination Maps" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.3, + 0.907, + 0.378 + ], + "angle": 0, + "content": "Since views in one LF are taken from different angulars of the same scene, in which the depth of the scene is fixed, the views are supposed to satisfy the disparity constraint. Suppose that the parameter of the camera is known, the disparity of two views \\( I_{\\mathbf{u_j}} \\) and \\( I_{\\mathbf{u_i}} \\) is calculated as:" + }, + { + "type": "equation", + "bbox": [ + 0.553, + 0.382, + 0.907, + 0.41 + ], + "angle": 0, + "content": "\\[\nD _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {i}}} (\\mathbf {x}) = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * \\frac {B * f}{Z (\\mathbf {x})} = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * d (\\mathbf {x}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.414, + 0.906, + 0.489 + ], + "angle": 0, + "content": "where \\( B \\) and \\( d \\) represent the baseline length and the disparity between adjacent views, respectively. \\( f \\) represents the focal length of the camera. \\( Z \\) represents the depth of the scene. Therefore, the two captured low-light views satisfy the disparity constraint as:" + }, + { + "type": "equation", + "bbox": [ + 0.637, + 0.494, + 0.907, + 0.51 + ], + "angle": 0, + "content": "\\[\nI _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(I _ {\\mathbf {u} _ {\\mathrm {j}}}, D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.514, + 0.906, + 0.56 + ], + "angle": 0, + "content": "where \\(W\\) denotes the warp operations according to the disparity \\(D_{\\mathbf{u_j}\\rightarrow \\mathbf{u_i}}\\). Similarly, the well-exposed LF views also satisfy the disparity constraint as:" + }, + { + "type": "equation", + "bbox": [ + 0.633, + 0.564, + 0.906, + 0.58 + ], + "angle": 0, + "content": "\\[\nR _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(R _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.906, + 0.614 + ], + "angle": 0, + "content": "Therefore, according to Eq. 2, the illumination adjustment map should also satisfy the disparity constraint." + }, + { + "type": "equation", + "bbox": [ + 0.634, + 0.618, + 0.906, + 0.634 + ], + "angle": 0, + "content": "\\[\n\\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.639, + 0.906, + 0.702 + ], + "angle": 0, + "content": "In order to constrain the consistency of the illumination adjustment maps, the direct way is to minimize the distance between the illumination map \\(\\overline{L}_{\\mathbf{u_i}}\\) and the warped illumination map \\(W(\\overline{L}_{\\mathbf{u_j}};D_{\\mathbf{u_j}\\to \\mathbf{u_i}})\\) from different angular \\(\\mathbf{u_j}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.71, + 0.906, + 0.739 + ], + "angle": 0, + "content": "\\[\n\\min \\sum_ {\\mathbf {u} _ {\\mathrm {j}}} \\sum_ {\\mathbf {u} _ {\\mathrm {i}}} | | W (\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}) - \\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} | |, \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.741, + 0.906, + 0.802 + ], + "angle": 0, + "content": "where \\(\\mathbf{u_j},\\mathbf{u_i}\\in \\{(0,0),(0,1),\\dots,(U,V)\\}\\). To simplify the computation, we warp the illumination adjustment maps of different views towards the central view \\(\\overline{L}_{\\mathbf{u_c}}\\) and construct these aligned views to a 4D array \\(\\mathbf{A}\\in \\mathbb{R}^{U\\times V\\times H\\times W}\\) .." + }, + { + "type": "equation", + "bbox": [ + 0.62, + 0.806, + 0.906, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} \\left(\\mathbf {u} _ {\\mathbf {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathbf {j}}}; D _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {c}}}\\right), \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.827, + 0.905, + 0.873 + ], + "angle": 0, + "content": "where \\(\\mathbf{u}_{\\mathbf{c}} = \\left(\\frac{U}{2},\\frac{V}{2}\\right)\\). Then, we instead calculate the variance of the warped illumination adjustment maps to assess their disparity consistency as:" + }, + { + "type": "equation", + "bbox": [ + 0.642, + 0.879, + 0.906, + 0.905 + ], + "angle": 0, + "content": "\\[\n\\min \\sum_ {\\mathbf {x}} \\operatorname {V a r} (\\mathbf {A} (:, \\mathbf {x})), \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7342" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.089, + 0.482, + 0.212 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.222, + 0.485, + 0.292 + ], + "angle": 0, + "content": "Figure 3. The proposed VCNet network consists of an illumination estimator VCAIA and a corruption restorer. In VCAIA, the illumination map is estimated and multiplied to the original input to light up the illumination. VCRM further recovers the details of the light-up image and outputs the final results." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.317, + 0.483, + 0.348 + ], + "angle": 0, + "content": "where \\( Var(\\mathbf{A}(:,\\mathbf{x})) \\) calculates the variance for each pixel along the angular dimension." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.348, + 0.484, + 0.44 + ], + "angle": 0, + "content": "However, in real scenes, both the ground truth illumination adjustment map \\(\\overline{L}_{\\mathbf{u}}\\) and the corresponding disparity \\(d\\) is unavailable. Moreover, the disparity is also difficult to estimate using low-light images with much noise [12]. Therefore, the disparity constraint of the illumination adjustment maps as in Eq. 9 is difficult to implement." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.448, + 0.423, + 0.465 + ], + "angle": 0, + "content": "3.2. Self-Supervised View-Consistent Loss" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.484, + 0.591 + ], + "angle": 0, + "content": "In this section, we relax the disparity constraint and design a self-supervised view-consistent illumination loss to ensure the view consistency of illumination adjustment maps. Instead of constraining each point according to its true disparity as in Eq. 9, we only constrain each point according to one specific disparity value, no matter whether it is the true disparity or not. Specifically, we first set \\( N \\) discrete candidate disparity values:" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.602, + 0.482, + 0.619 + ], + "angle": 0, + "content": "\\[\nd ^ {n} = d _ {\\min } + n * \\Delta d, n \\leq N \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.485, + 0.705 + ], + "angle": 0, + "content": "where \\(\\Delta d = (d_{max} - d_{min}) / N\\) denotes the granularity of \\(d\\). \\(d_{min}\\) and \\(d_{max}\\) denote the minimum and maximum disparity of the scene. Subsequently, we warp each view to the center view according to all the possible disparity values \\(d^n\\), and the related 4D array is denoted as \\(\\mathbf{A}^n \\in \\mathbb{R}^{U \\times V \\times H \\times W}\\):" + }, + { + "type": "equation", + "bbox": [ + 0.188, + 0.718, + 0.483, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathbf {A} ^ {n} \\left(\\mathbf {u} _ {\\mathrm {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {c}}} ^ {n}\\right), \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.739, + 0.349, + 0.755 + ], + "angle": 0, + "content": "where \\(D_{\\mathbf{u}_i - > \\mathbf{u_c}}^n (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^n\\)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.755, + 0.484, + 0.875 + ], + "angle": 0, + "content": "We then calculate the variance of the warped view for each point \\(\\mathbf{x}\\) and for each possible disparity \\(n\\) as \\(Var(\\mathbf{A}^n (:, \\mathbf{x}))\\), constructing the commonly known cost volume as in the disparity estimation methods [4, 13]. According to the photometric consistency, the variance under the true disparity is supposed to have the minimum value compared with other disparities. Therefore, we first find the specific disparity \\(d^{*}\\) for each point that has the minimum variance:" + }, + { + "type": "equation", + "bbox": [ + 0.177, + 0.883, + 0.483, + 0.905 + ], + "angle": 0, + "content": "\\[\nd ^ {*} (\\mathbf {x}) = \\arg \\min _ {d ^ {n}} (V a r \\left(\\mathbf {A} ^ {n} (:, \\mathbf {x}))\\right)). \\tag {12}\n\\]" + }, + { + "type": "image", + "bbox": [ + 0.575, + 0.089, + 0.849, + 0.184 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.194, + 0.906, + 0.223 + ], + "angle": 0, + "content": "Figure 4. The estimated illumination EPI w/ and w/o \\( loss_{vc} \\). The former keeps the view consistent as the true EPI with normal light." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.25, + 0.906, + 0.28 + ], + "angle": 0, + "content": "We then minimize the variance related to the \\(d^{*}(\\mathbf{x})\\) to make sure that all views are aligned in the disparity \\(d^{*}(\\mathbf{x})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.627, + 0.29, + 0.906, + 0.319 + ], + "angle": 0, + "content": "\\[\n\\ell_ {v c} = \\sum_ {\\mathbf {x}} \\operatorname {V a r} \\left(\\mathbf {A} ^ {*} (:, \\mathbf {x})\\right), \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.329, + 0.906, + 0.435 + ], + "angle": 0, + "content": "where \\(\\mathbf{A}^*\\) represents the 4D array with aligned views according \\(D_{\\mathbf{u_j - > u_c}}^* (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^* (\\mathbf{x})\\) as in Eq. 11. An example is shown in Fig. 2, where the minimum variance is set as our objective function. In such a way, each pixel is assigned a disparity value that has the minimum variance to make the illumination adjustment value of each pixel as consistent as possible with other views." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.445, + 0.816, + 0.461 + ], + "angle": 0, + "content": "3.3. Adaptive Illumination Adjustment" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.906, + 0.587 + ], + "angle": 0, + "content": "Beginning with the input low-light LF image \\(\\mathcal{L}^{in}\\), we design a simple illumination estimator, View Consistent Adaptive Illumination Adjustment (VCAIA), \\(\\mathcal{E}\\) for LFs to estimate the illumination adjustment map. Specifically, as in Fig. 3, we first estimate the global illumination adjustment matrix of the scene by integrating information from all views. The local brightness refinement is then developed to adjust the matrix according to the details of each view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.588, + 0.906, + 0.664 + ], + "angle": 0, + "content": "In order to integrate the abundant information from all views and further reduce the effect of noise, we roughly calculate the average of all views across all color channels in LF image as \\( I_w \\in \\mathbb{R}^{H \\times W \\times 1} \\). Then following [1], we calculate the global illumination adjustment matrix as:" + }, + { + "type": "equation", + "bbox": [ + 0.544, + 0.673, + 0.905, + 0.705 + ], + "angle": 0, + "content": "\\[\n\\overline {{I ^ {l}}} _ {\\text {g l o b a l}} (\\mathbf {x}) = \\frac {\\log \\left(I _ {w} (\\mathbf {x}) / \\mu + 1\\right)}{\\log \\left(\\max \\left(I _ {w} (\\mathbf {x})\\right) / \\mu + 1\\right) \\cdot I _ {w} (\\mathbf {x})}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.717, + 0.906, + 0.825 + ], + "angle": 0, + "content": "Subsequently, we replicate \\(\\overline{I}^{l}_{global}\\) to all views and obtain the global illumination adjustment matrix \\(\\overline{\\mathcal{L}}^{l}_{global}\\). To further refine illumination estimation for each view, we employ a lightweight learnable network. Specifically, we first stack all views of the input low-light LF image and use the cascaded 3D spatial convolution layers to obtain the local brightness adjustment matrix \\(\\overline{\\mathcal{L}}^{l}_{local}\\) of each view." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.902 + ], + "angle": 0, + "content": "The local and global illumination adjustment matrix is combined to derive the final adjustment matrix and is then multiplied to \\(\\mathcal{L}^{in}\\) to obtain the adjusted high-illumination image: \\(\\overline{\\mathcal{L}} = (\\overline{\\mathcal{L}^t}_{global} + \\overline{\\mathcal{L}^t}_{local})\\). By minimizing the proposed self-supervised loss \\(\\ell_{vc}\\) on the estimated \\(\\overline{\\mathcal{L}}\\) according" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7343" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.144, + 0.089, + 0.852, + 0.104 + ], + "angle": 0, + "content": "Table 1. Quantitative Comparisons of Low Light LF Image Enhancement Methods in L3F-Dynamic Task w/ w/o VCAIA." + }, + { + "type": "table", + "bbox": [ + 0.097, + 0.115, + 0.9, + 0.23 + ], + "angle": 0, + "content": "
MethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LFFT-DCU23.170.80030.095819.600.72010.170219.820.67180.2074
DRLF19.100.69470.188817.420.66000.240217.040.63220.2608
DRLF+E22.19 (+3.09)0.7636 (+0.0689)0.1124 (-0.0764)21.38 (+3.96)0.6976 (+0.0376)0.1676 (-0.0726)20.29 (+3.25)0.6326 (+0.0004)0.2265 (-0.0343)
MSP21.370.75600.133018.640.68980.207917.620.65430.2407
MSP+E22.80 (+1.43)0.7936 (+0.0376)0.1261 (-0.0069)21.81 (+3.17)0.7436 (+0.0538)0.1673 (-0.0406)21.10 (+3.48)0.6916 (+0.0373)0.2128 (-0.0279)
VCNet-E22.540.78350.125120.830.71030.201519.920.64500.2666
VCNet24.58 (+2.04)0.8220 (+0.0385)0.0897 (-0.0354)24.06 (+3.23)0.7652 (+0.0549)0.1292 (-0.0723)23.09 (+3.17)0.7032 (-0.0582)0.1981 (-0.0685)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.248, + 0.482, + 0.445 + ], + "angle": 0, + "content": "to Eq. 13, we constrain the calculated illumination adjustment maps to satisfy the view consistency even without the true disparities. Since no supervised images are needed, the illumination adjustment can be applied to various low-light LF enhancement networks. We show one estimated example in Fig. 4, where the Epipolar Plane Images (EPIs) are calculated by fixing \\( v \\) in angular domain and \\( w \\) in the spatial domain. Without the proposed \\( \\ell_{vc} \\), the lines in the EPIs of the estimated illumination maps become ambiguous and are inconsistent with the ground truth EPIs of the normallight LFs. In contrast, when we use \\( \\ell_{vc} \\) to constrain the view consistency, the lines in estimated EPIs of illumination maps retain the direction as the ground truth EPIs." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.464, + 0.263, + 0.48 + ], + "angle": 0, + "content": "3.4. Recovery Module" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.489, + 0.483, + 0.626 + ], + "angle": 0, + "content": "After illumination adjustment, the image \\(\\mathcal{L}^a = \\overline{\\mathcal{L}}\\odot \\mathcal{L}^{in}\\) with a standard illumination level is obtained. However, the light-up process also amplifies the noise and artifacts, causing under-/over-exposure and color distortion. Therefore, similar to other retinae-based methods, a recovery module is needed to suppress noise and restore scene details. Some existing networks, such as the LF denoising model DRLF and the LF low-light enhancement model MSP, can be used as the recovery module." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.629, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We also design a simple View Consistent Recovery Module (VCRM) to further import the inherent disparity constraint into the recovery process to enhance the recovery effect. VCRM adopts a multi-scale architecture. In each scale, the view consistent feature aggregation unit is implemented to explore the redundant information of the views to recover details. In each unit, we employ deformable convolution layers on EPIs to further integrate information between different views. Since the slope of lines in EPIs corresponds to disparity, we introduce the disparity constraint to help the model identify correct and consistent complementary information. Specifically, we use uniform offsets in both V-H and U-W spaces, so that the angular information can be efficiently incorporated and maintain epipolar consistency. Finally, we concatenate multi-scale features and use a decoder with a similar structure to the encoder to aggregate them and obtain the refined result \\(\\mathcal{L}^{out}\\). More detailed design is provided in the supplementary details." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.247, + 0.731, + 0.264 + ], + "angle": 0, + "content": "3.5. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.269, + 0.905, + 0.435 + ], + "angle": 0, + "content": "As in Fig. 3, the overview architecture of our VCNet primarily employs two VCAIA and VCRM. We train our VCNet by minimizing a hybrid loss: \\(\\ell_{total} = \\lambda_1\\ell_{vc} + \\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}\\). \\(\\ell_{rec}\\) is the reconstruction loss, which calculates the L1 loss between the ground truth of well-posed image \\(\\mathcal{L}^{gt}\\) and the recovered results \\(\\mathcal{L}^{out}\\) of VCNet. The SSIM loss [33] \\(\\ell_{ssim}\\) measures the structural similarity between \\(\\mathcal{L}^{gt}\\) and \\(\\mathcal{L}^{out}\\). \\(\\ell_{vc}\\) is the proposed self-supervised illumination loss in Sec.3, which is used to keep the view consistent of the estimated illumination maps \\(\\overline{\\mathcal{L}}^l\\). The three hyperparameters \\(\\lambda_1\\), \\(\\lambda_2\\) and \\(\\lambda_3\\) are empirically set to 0.5, 1, 0.5." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.436, + 0.906, + 0.572 + ], + "angle": 0, + "content": "Our model is implemented using the PyTorch framework and trained on a NVIDIA RTX A4000 GPU. The model is optimized using Adam optimization with \\(\\beta_{1} = 0.9\\) and \\(\\beta_{2} = 0.999\\). The batch size is set to 1 and the learning rate is initially set to 0.0001. Data augmentation techniques such as random cropping, vertical and horizontal flipping are used. During training, the views are randomly cropped in patches of size \\(128 \\times 128\\). During testing, we use the original resolution LF data without applying data augmentation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.583, + 0.645, + 0.6 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.607, + 0.609, + 0.621 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.629, + 0.905, + 0.736 + ], + "angle": 0, + "content": "We evaluate the proposed method using the publicly accessible Low-Light-Light-Field (L3F) Dataset and L3F-Wild Dataset from [15]. The L3F dataset comprises 27 distinct scenes, each of which was captured under three low-light conditions, labeled L3F-20, 50, 100, and one with normal lighting. The L3F-Wild is captured in almost 0 lux conditions and captured with \\(\\frac{1}{4} -\\frac{1}{15}\\) second exposure time." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.736, + 0.905, + 0.854 + ], + "angle": 0, + "content": "L3F-Fixed Task: Previous LF-based low-light enhancement methods trained their models under fixed illumination conditions. Specifically, following [15], 18 scenes under the same illumination condition are selected for training, while the remaining 9 scenes are reserved for testing. In the fixed illumination conditions, 3 models are specifically trained for L3F-20, 50, 100 in different illumination conditions. We denote it as the L3F-Fixed task." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.905, + 0.901 + ], + "angle": 0, + "content": "L3F-Dynamic Task: To further assess the model's capability to adapt to various illumination conditions, we establish a dataset with dynamic illumination as L3F-Dynamic." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7344" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.111, + 0.089, + 0.885, + 0.395 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.405, + 0.908, + 0.448 + ], + "angle": 0, + "content": "Figure 5. Visual comparison on \\( L3F \\)-Dynamic task, where the central view and EPIs of the recovery LFs are shown. Other algorithms either generate over-exposed and noisy images or introduce black spot corruptions and unnatural artifacts. Our VCNet restores well-exposed structural contents and textures. By adding our VCAIA module, the recovered results of all models are closer to ground truth." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.463, + 0.483, + 0.598 + ], + "angle": 0, + "content": "Specifically, we combine images from the L3F-20, 50, 100 datasets. We use all training sets from the three different illumination conditions (including \\(18 \\times 3\\) scenes) to train a single model. The single trained model is supposed to handle dynamic illumination conditions. We use the same testing dataset as L3F-Fixed to evaluate its performance. Compared to L3F-Fixed task, L3F-Dynamic task is more difficult, which needs to train one single model that is able to handle all different illumination conditions." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.607, + 0.436, + 0.623 + ], + "angle": 0, + "content": "4.2. Compare with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.629, + 0.483, + 0.854 + ], + "angle": 0, + "content": "We compare our VCNet and several SOTA methods, including 3 recent published single-frame retinae-based methods (RFormer [22], Pydiff [42] and DarkIR [6]) and 9 LF-based methods (L3F-Net [15], FER-Net [14], DRLF [10], MSP-3 [27], IINet [17], DistgSSR [31], PMSNet [29], LRT [39], LFLL-DCU [18]). The illumination estimator \\(\\mathcal{E}\\) is applied to the LF denoising model DRLF and low-light enhancement model MSP, denoted as DRLF+ \\(\\mathcal{E}\\), MSP+ \\(\\mathcal{E}\\). Specifically, we embed the illumination estimator \\(\\mathcal{E}\\) into the beginning of these two models and add the \\(\\ell_{vc}\\) in their training strategy. Our VCNet-\\(\\mathcal{E}\\) is trained by deleting \\(\\mathcal{E}\\) module and only minimizing \\(\\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}\\). The reference-based metric PSNR, SSIM [32] and LPIPS [36] and nonreference metric NIQE [21] and BRISQUE [20] are chosen as the evaluation metrics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Dynamic Illumination Task. We first assess our model's ability to adapt to dynamic illumination. Specifically, we train each method using the L3F-Dynamic training dataset" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.461, + 0.906, + 0.489 + ], + "angle": 0, + "content": "Table 2. Quantitative Comparisons with Single-Frame-based Low Light Enhancement Methods on the L3F-dynamic Task." + }, + { + "type": "table", + "bbox": [ + 0.523, + 0.499, + 0.894, + 0.58 + ], + "angle": 0, + "content": "
MethodL3F-20L3F-50L3F-100
PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑
RFormer20.36 / 0.77 / 0.342 / 50.9419.30 / 0.70 / 0.493 / 39.1917.13 / 0.62 / 0.615 / 30.79
RFormer*27.29 / 0.87 / 0.163 / 67.2623.81 / 0.77 / 0.294 / 42.3122.51 / 0.70 / 0.374 / 34.84
Pydiff19.74 / 0.71 / 0.406 / 48.7219.13 / 0.62 / 0.535 / 36.3018.47 / 0.49 / 0.757 / 31.14
Pydiff*23.42 / 0.82 / 0.256 / 60.1122.74 / 0.76 / 0.331 / 40.9622.24 / 0.70 / 0.437 / 33.51
DarkIR20.11 / 0.76 / 0.315 / 51.6317.24 / 0.65 / 0.474 / 35.3715.09 / 0.58 / 0.621 / 15.31
VCNet24.58 / 0.82 / 0.089 / 79.0724.06 / 0.77 / 0.129 / 78.4623.09 / 0.70 / 0.198 / 76.80
" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.598, + 0.907, + 0.854 + ], + "angle": 0, + "content": "in Sec. 4.1 and the results are presented in Table 1. By adding the illumination estimator \\(\\mathcal{E}\\), all three models exhibit around 3dB PSNR improvement by considering the adaptive illumination estimation. Since LFLL-DCU only uses one learnable parameter to estimate the illumination map, their model performs well on the L3F-20 data but poorly on L3F-50 and L3F-100. Our VCNet outperforms the current SOTA methods MSP, DRLF, and LFLL-DCU with more than \\(2\\sim 6\\) dB PSNR improvement. Fig. 5 displays the visual results comparison, where all visual effects are improved by combining the \\(\\mathcal{E}\\). This clearly demonstrates the effectiveness of our illumination estimator \\(\\mathcal{E}\\) in adaptively adjusting to varying illumination conditions. The results of VCNet stand out by producing more realistic overall color compared with the ground truth. Other methods restore blurred scene details, while our VCNet recovers the clearer and richer detailed texture in the recovered images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "We then make a comparison with single-frame-based methods in Table 2, which claim to handle dynamic lighting conditions. Their models are trained on a large-scale" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7345" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.134, + 0.089, + 0.867, + 0.233 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.242, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Figure 6. Visual comparison on the \\(L3F\\)-Wild dataset, where the central view of LFs is shown. Other methods either collapse by noise, distort color, or under-/over-exposed images. Our VCNet effectively removes the noise and reconstructs well-exposed image details." + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.277, + 0.482, + 0.292 + ], + "angle": 0, + "content": "Table 3. Quantitative Comparison on Real-world L3F-Wild Data." + }, + { + "type": "table", + "bbox": [ + 0.094, + 0.296, + 0.485, + 0.356 + ], + "angle": 0, + "content": "
MethodsRFormer*Pydiff*DRLFDRLF+EMSPMSP+EVCNet-EVCNet
NIQE ↓15.6717.8219.9718.7818.8218.2317.8714.22
BRISQUE↓31.2036.1232.8030.5430.2029.4230.1125.98
TIMES↓39.84100.1216.2216.263.893.935.996.33
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.37, + 0.483, + 0.491 + ], + "angle": 0, + "content": "single-image dataset and we carefully fine-tune their methods using L3F-Dynamic training dataset (labelled as *), where each view of LF is considered a sample during the training. Our method outperforms others under most lighting conditions, especially in perceptual quality. Although RFormer* achieves relatively high PSNR and SSIM values in the L3F-20 dataset, it suffers from noticeable blurry artifacts, as shown in Fig. 5." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.483, + 0.719 + ], + "angle": 0, + "content": "To further demonstrate the generalization ability, we applied the methods trained on the \\(L3F\\)-Dynamic dataset to another real-world low-light LF data \\(L3F\\)-Wild. The results are shown in Fig. 6 and Table 3. Due to the absence of redundant information from different views, the results of single-frame methods are relatively dark, and the restored details suffer from blurring. For LF-based methods, DRLF is observed to cause a noticeable color shift. The results of MSP and LFLL-DCU suffer from residual noise and local over-expose. Combining with the \\(\\mathcal{E}\\), the recovered results of MSP and DNLF have higher color fidelity, clearer structural details and less noise. In contrast, our results achieve high perceptual quality with clear structures, realistic color and less noises. Our VCAIA module \\(\\mathcal{E}\\) is also proved that can enhance the perceptual quality of other methods." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.72, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Fixed Illumination Task. We then evaluate all the methods under the fixed illumination condition by separately training 3 modals using the \\(L3F\\)-Fixed dataset. The results using LFs with angular resolution \\(5 \\times 5\\) and \\(8 \\times 8\\) are illustrated in Table 4, respectively. We retrain the LRT [39] and LFLL-DCU [18] for comparison. When compared with single-frame-based methods, LF-based methods demonstrate higher performance, highlighting the utility of redundant LF information in low-light enhancement. With the increase of the input image views, the recovery effect of the LF models is better because it provides more redundant information to the network. Note that in PMSNet," + }, + { + "type": "table_caption", + "bbox": [ + 0.548, + 0.277, + 0.871, + 0.292 + ], + "angle": 0, + "content": "Table 4. Quantitative Comparisons on L3F-fixed Task." + }, + { + "type": "table", + "bbox": [ + 0.557, + 0.298, + 0.869, + 0.501 + ], + "angle": 0, + "content": "
Ang ResMethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
8L3F-Net24.780.82623.210.75221.830.683
IINet25.400.86224.140.81122.130.721
DistgSSR25.440.86424.180.81022.960.755
PMSNet27.030.86624.520.81222.740.762
VCNet26.840.87125.410.82323.580.772
5L3F-Net24.980.81423.370.75321.950.669
FER-Net25.310.82323.400.76621.730.688
LRT25.970.81523.990.74222.150.691
LFLL-DCU25.420.84224.030.78722.280.702
DRLF24.780.806422.520.721721.510.6543
DRLF+ε24.990.817923.060.732922.190.6681
MSP25.480.830124.100.775922.750.7133
MSP+ε25.730.843024.680.783123.060.7191
VCNet-ε25.870.837224.350.789722.690.7204
VCNet26.610.847325.130.792123.240.7223
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.522, + 0.907, + 0.643 + ], + "angle": 0, + "content": "in addition to the original \\(8 \\times 8\\) views, views from two more rows and columns are included in the feature extraction. Among LF-based methods, our VCNet significantly outperforms SOTA methods on most of the \\(L3F\\)-fixed task, demonstrating its ability to better handle low-illumination degradation and produce superior restoration results. After adding the illumination estimator \\(\\mathcal{E}\\), all models also achieve around a 0.2-0.8 dB PSNR improvement." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.909, + 0.903 + ], + "angle": 0, + "content": "View Consistency Comparison. View consistency holds great significance in LF enhancement as it directly impacts the performance of subsequent LF-based applications like depth estimation. To verify the effectiveness of our VCAIA \\(\\mathcal{E}\\) in maintaining consistency in recovery results, we present the parallax content Precision-Recall (PR) curves used in [5] for evaluation. In one LF, each view is subtracted from the central view, and different threshold values are chosen to calculate binary edge maps. The PR curves show how closely the binary edge maps of the reconstructed LFs are match to the ground truth. We use models trained on dynamic illumination for comparison. In Fig. 7, when all the methods add the illumination adjustment, the PR curve of the recovered image is improved, demonstrating that our \\(\\mathcal{E}\\) can achieve view-consistent illumination adjustment. We further calculate the Average Precision (AP) by measuring the area under the PR curves for quantitative" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7346" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.088, + 0.48, + 0.164 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.212 + ], + "angle": 0, + "content": "Figure 7. The PR curves map comparison of L3F-20, 50, 100 dataset, in which how well the views consistency preserved is shown. With \\(\\mathcal{E}\\), the PR curves of all methods are better." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.229, + 0.483, + 0.44 + ], + "angle": 0, + "content": "comparison. As in Table 2, for the single-frame methods, it is difficult to recover views with consistent illumination, so that the AP is relatively low even with high PSNR and SSIM. By contrast, our model exhibits notably higher AP values, especially under the challenging low-light conditions of the L3F-100 dataset. The comparison with LF-based methods is shown in Fig. 1. Our model consistently achieves the highest AP values, showcasing the superiority of VCNet in terms of low-light enhancement and consistency preservation. The EPIs of the recovered images are also visualized in Fig. 5, in which the orientation of lines corresponds to the disparity information. It is also clear that our model excels in maintaining the view consistency. More results are provided in the supplementary materials." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.44, + 0.484, + 0.531 + ], + "angle": 0, + "content": "Running time. We further show the running time of the models using the size of the scatter in Fig. 1 and Table 3. Our model runs faster (6.3s) than DRLF (16.2s) but is slightly slower than MSP (3.9s). The computation time of \\(\\mathcal{E}\\) is 0.04s on the A5000 for a \\(5 \\times 5 \\times 642 \\times 432\\) input, which is lightweight to add \\(\\mathcal{E}\\) to different models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.541, + 0.245, + 0.558 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "table_caption", + "bbox": [ + 0.132, + 0.57, + 0.443, + 0.584 + ], + "angle": 0, + "content": "Table 5. The Ablation Study on L3F-Dynamic Task." + }, + { + "type": "table", + "bbox": [ + 0.171, + 0.595, + 0.406, + 0.666 + ], + "angle": 0, + "content": "
No.LocalGlobal\\( \\ell_{vc} \\)PSNRSSIMLPIPS
(1)×××20.760.700.2036
(2)×√×22.780.750.1469
(3)√√×23.590.750.1420
(4)√√√23.840.760.1395
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.675, + 0.483, + 0.886 + ], + "angle": 0, + "content": "We conduct a series of ablation studies under dynamic illumination conditions with \\(5 \\times 5\\) LF images. To ensure fairness, we keep the parameters of the variant models close to VCNet by adjusting the feature channels. It is clear that the performance of the VCNet without \\(\\mathcal{E}\\) (model-1) is the worst. Compared with model-1, model-2 using the global brightness refinement performs better. Meanwhile, with the local brightness adaptive adjustment branch, model-3 increases \\(0.8\\mathrm{dB}\\) in the PSNR values than model-2. We further validate the effectiveness of the loss \\(\\ell_{vc}\\). Comparing model-4 with model-3, the PSNR increases nearly \\(0.4\\mathrm{db}\\) by adding the \\(\\ell_{vc}\\) to supervise the illumination adjustment matrix \\(\\overline{\\mathcal{L}^l}\\). This shows that \\(\\ell_{vc}\\) can perform consistency constraints on \\(\\overline{\\mathcal{L}^l}\\) to provide higher-quality light-up images." + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "We further visualize the adaptive illumination adjust" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.088, + 0.9, + 0.201 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.208, + 0.907, + 0.248 + ], + "angle": 0, + "content": "Figure 8. Visualization of the illumination maps in \\(L3F\\)-Dynamic task. For darker scenes, the values in the estimated illumination adjustment map is larger." + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.25, + 0.902, + 0.362 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.372, + 0.907, + 0.414 + ], + "angle": 0, + "content": "Figure 9. Visual comparison of the enhanced result (top row) and input central view (second row) of one scene with different low-light levels in the L3F-Wild dataset." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.422, + 0.907, + 0.62 + ], + "angle": 0, + "content": "ment matrix \\(\\overline{\\mathcal{L}_l}\\) and the related recovered results in different illumination environments in Fig. 8. The model is trained using the \\(L3F\\)-dynamic dataset. For input with an extended exposure time, the generated illumination adjustment matrix tends to be relatively darker. Conversely, as the exposure time diminishes, the input images become darker and the values of the adjustment matrix gradually increase. We also illustrate the performance of our VCNet in processing images with disparate brightness levels from the L3F-Wild dataset in Fig. 9. Our method consistently achieves superior enhancement quality across input scenes with varying levels of brightness. This consistency ensures effective low-light image restoration across varying illumination conditions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.633, + 0.634, + 0.648 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.659, + 0.907, + 0.855 + ], + "angle": 0, + "content": "In this paper, we introduced a novel framework for scene-adaptively low-light enhancement in light field(LF) images. First, by analyzing Retinex theory and extending it to the LF, we introduce a self-supervised loss to ensure that the calculated illumination map adheres to the consistency relationship. Then, we design an illumination estimator with global adjustment and local refinement to perceive the illumination of the scene. Finally, we design a corruption restorer and import the consistency constraint into the feature aggregation process for better recovery. Extensive quantitative and qualitative experiments demonstrate that our VC-Net significantly outperforms state-of-the-art methods on both fixed and dynamic illumination enhancement tasks." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032)." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "7347" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.09, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Hyunchan Ahn, Byungjik Keum, Daehoon Kim, and Hwang Soo Lee. Adaptive local tone mapping based on retina for high dynamic range images. In Proceedings of the International Conference on Consumer Electronics (ICCE), pages 153-156, 2013. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.187, + 0.484, + 0.241 + ], + "angle": 0, + "content": "[2] Tom E. Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):972-986, 2012. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.245, + 0.484, + 0.298 + ], + "angle": 0, + "content": "[3] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 3291-3300, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.301, + 0.483, + 0.342 + ], + "angle": 0, + "content": "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.344, + 0.483, + 0.385 + ], + "angle": 0, + "content": "[5] Jie Chen, Junhui Hou, and Lap-Pui Chau. Light field denoising via anisotropic parallax analysis in a cnn framework. IEEE Signal Processing Letters, 25(9):1403-1407, 2018. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.388, + 0.483, + 0.443 + ], + "angle": 0, + "content": "[6] Daniel Feijoo, Juan C. Benito, Alvaro Garcia, and Marcos V. Conde. Darkir: Robust low-light image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 10879-10889, 2025. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.445, + 0.483, + 0.514 + ], + "angle": 0, + "content": "[7] Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2782-2790, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.517, + 0.483, + 0.572 + ], + "angle": 0, + "content": "[8] Chunle Guo, Jing Jin, Junhui Hou, and Jie Chen. Accurate light field depth estimation via an occlusion-aware network. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.574, + 0.483, + 0.643 + ], + "angle": 0, + "content": "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1777-1786, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.645, + 0.483, + 0.713 + ], + "angle": 0, + "content": "[10] Mantang Guo, Junhui Hou, Jing Jin, Jie Chen, and LapPui Chau. Deep spatial-angular regularization for light field imaging, denoising, and super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6094-6110, 2022. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.716, + 0.483, + 0.757 + ], + "angle": 0, + "content": "[11] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing, 26(2):982-993, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.759, + 0.483, + 0.827 + ], + "angle": 0, + "content": "[12] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 19-34, 2017. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.83, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[13] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.483, + 0.902 + ], + "angle": 0, + "content": "[14] Mohit Lamba and Kaushik Mitra. Fast and efficient restoration of extremely dark light fields. In Proceedings of the Win-" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.094, + 0.905, + 0.121 + ], + "angle": 0, + "content": "ter Conference on Applications of Computer Vision (WACV), pages 3152-3161, 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.177 + ], + "angle": 0, + "content": "[15] Mohit Lamba, Kranthi Kumar Rachavarapu, and Kaushik Mitra. Harnessing multi-view perspective of light fields for low-light imaging. IEEE Transactions on Image Processing, 30:1501-1513, 2021. 1, 2, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.179, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[16] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. ACM, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.222, + 0.905, + 0.276 + ], + "angle": 0, + "content": "[17] Gaosheng Liu, Huanjing Yue, Jiamin Wu, and Jingyu Yang. Intra-inter view interaction network for light field image super-resolution. IEEE Transactions on Multimedia, 25: 256-266, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.279, + 0.905, + 0.333 + ], + "angle": 0, + "content": "[18] Xianqiang Lyu and Junhui Hou. Enhancing low-light light field images with a deep compensation unfolding network. IEEE Transactions on Image Processing, 33:4131-4144, 2024. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.336, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[19] Paras Maharjan, Li Li, Zhu Li, Ning Xu, Chongyang Ma, and Yue Li. Improving extreme low-light image denoising via residual learning. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 916-921, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.407, + 0.905, + 0.46 + ], + "angle": 0, + "content": "[20] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.463, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[21] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.507, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[22] Junxiang Ruan, Xiangtao Kong, Wenqi Huang, and Wenming Yang. Retformer: Retinex-based enhancement in transformer for low-light image. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2, 3, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.577, + 0.907, + 0.646 + ], + "angle": 0, + "content": "[23] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 6842-6850, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.648, + 0.905, + 0.716 + ], + "angle": 0, + "content": "[24] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the International Conference on Computer Vision (ICCV), pages 9680-9689, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.905, + 0.773 + ], + "angle": 0, + "content": "[25] Tingxian Wang, Alexei A. Efros, and Ravi Ramamoorthi. Depth estimation with occlusion modeling using light-field cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:2170-2181, 2016. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.775, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[26] Wei Wang, Xin Chen, Cheng Yang, Xiang Li, Xuemei Hu, and Tao Yue. Enhancing low light videos by exploring high sensitivity camera noise. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4110-4118, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.846, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[27] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.518, + 0.094, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7348" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[28] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.15, + 0.482, + 0.205 + ], + "angle": 0, + "content": "[29] Xingzheng Wang, Kaiqiang Chen, Zixuan Wang, and Henhao Huang. Pmsnet: Parallel multi-scale network for accurate low-light light-field image enhancement. IEEE Transactions on Multimedia, 26:2041-2055, 2024. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.207, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[30] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the International Conference on Computer Vision (ICCV), pages 13128-13137, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.359 + ], + "angle": 0, + "content": "[31] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.362, + 0.482, + 0.417 + ], + "angle": 0, + "content": "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.482, + 0.473 + ], + "angle": 0, + "content": "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.476, + 0.482, + 0.517 + ], + "angle": 0, + "content": "[34] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.52, + 0.482, + 0.587 + ], + "angle": 0, + "content": "[35] Henry Wing Fung Yeung, Junhui Hou, Xiaoming Chen, Jie Chen, Zhibo Chen, and Yuk Ying Chung. Light field spatial super-resolution using deep efficient spatial-angular separable convolution. IEEE Transactions on Image Processing, 28(5):2319-2330, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.591, + 0.482, + 0.657 + ], + "angle": 0, + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.715 + ], + "angle": 0, + "content": "[37] Shansi Zhang and Edmund Y. Lam. An effective decomposition-enhancement method to restore light field images captured in the dark. Signal Processing, 189:108279, 2021. 1, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[38] Shuo Zhang, Youfang Lin, and Hao Sheng. Residual networks for light field image super-resolution. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 11046-11055, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[39] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Lrt: An efficient low-light restoration transformer for dark light field images. IEEE Transactions on Image Processing, 32:4314-4326, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[40] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1632-1640, New York, NY, USA, 2019. ACM. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.134 + ], + "angle": 0, + "content": "[41] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision., 129(4):1013-1037, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.906, + 0.176 + ], + "angle": 0, + "content": "[42] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 6" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.093, + 0.906, + 0.176 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "7349" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_origin.pdf b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..e40b0aa260186f4925d70a8702090290f0416892 --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/2b9eae5a-6bd0-47a6-941b-39ead0bda71c_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:cab306171b61db439c6d9b275d5cc4a95c69ecd353084234bd5e5a084f976afc +size 6630627 diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/full.md b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/full.md new file mode 100644 index 0000000000000000000000000000000000000000..3d316ccb3151eaf40fbea1f89ca05824cdfdf436 --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/full.md @@ -0,0 +1,310 @@ +# Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement + +Shuo Zhang, Chen Gao, Youfang Lin* +Beijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science& Technology, Beijing Jiaotong University, Beijing, China. + +{zhangshuo, gaochen, yflin}@bjtu.edu.cn + +# Abstract + +Light Field (LF) images captured under low illumination conditions typically exhibit low quality. Recent learning-based methods for low-light LF images enhancement are generally tailored to specific illumination inputs, limiting their performance in real-world scenes. Moreover, how to maintain the inherent view-consistency in the enhanced images also remains a difficult problem. In this paper, we propose to explore the view consistency for scene-adaptive low-light LF enhancement. We first analyze the view consistency for LF illumination maps and design a self-supervised view-consistent loss to keep the consistency between the illumination maps of different views in LFs. To enhance the model's perception of illumination, we combine both global and local information to estimate the illumination map, which is easily plugged into other models. Subsequently, we use the illumination maps to light up the low-light LF images and restore the corruption to produce the final enhanced image. Experiments demonstrate that our View-Consistent Network (VCNet) outperforms state-of-the-art methods on real-world low-light LF datasets in both fixed lighting conditions and dynamic lighting conditions. Our proposed illumination adjustment is also demonstrated that can comprehensively improve the performance of existing methods in terms of both image quality and view consistency. Code is available at: https://github.com/shuozh/VCNet + +# 1. Introduction + +In contrast to conventional cameras, Light Field (LF) cameras have the capability to simultaneously capture both directional and intensity of all light rays in a single shot. Because of the structural information inherent in LF images, they are well-suited for various applications, including digital refocus [2], depth estimation [8], and 3D reconstruc + +![](images/91e0ebca1dd1fc459bb03b5f1845d8ce6416d152f9409e24f7445440f00dfe8b.jpg) +Figure 1. Performance comparison in dynamic illumination tasks. The size of the points represents the run time to process one LF. Our VCNet achieves the best results in terms of detail restoration with higher PSNR and views consistency preservation with higher AP. By adding our proposed illumination adjustment module, labeled as “+E”, the performances of all methods are improved. + +tion [25]. Nevertheless, the low-quality LF images severely compromised the performance of these applications, especially for images captured in low-light environments. + +In recent years, several methods are developed for enhancing low-light LF images. Some methods [15, 28, 37] process stacked views of LF images to implicitly extract disparity information between views to take redundant information for reconstructing LF. Other methods employ advanced spatial-angular separation convolution [35, 37, 38] to extract redundant information between views for recovery. Although these methods are effective in handling LF data, their performance is limited since they do not take the view consistency into consideration, breaking the view consistency of the recovered image. Moreover, these methods learn a low-light image to normal-light image mapping to restore the low-light LF images. For different levels of illumination degradation, they require separate learning processes to train different models for recovery. Since the degradation levels are usually unknown for real-world images, it is difficult for these models to select appropriate parameters for restoration. + +To deal with dynamically degraded inputs, a simple yet + +principled solution relying on the Retinex theory is proposed, which has been widely used for single-image [22, 23, 30]. These Retinex-based methods usually divide the light-up process into two stages: low contrast adjustment and corruptions (e.g., noise, artifact, color distortion, etc.) recovery. By estimating the illumination map of the scene for low contrast adjustment, the model is able to deal with dynamic illumination degradation. For LF images, the most straightforward way is to learn a brute-force mapping function for each view to apply one-to-one illumination adjustment. However, processing each LF view individually tends to recover each view with quite different illumination maps, resulting in varying degrees of color variation across different views. Another possible way is to estimate only one identical illumination adjustment gain for each view [39] or use a learnable parameter to compute the full illumination map [18], which is also improper since there exist disparities between views in one LF. This severely breaks the structural information inherent in LF images and limits the application of the following tasks. Therefore, how to keep the view-consistency in the estimated illumination maps in one LF becomes an important problem for scene-adaptive low-light LF image enhancement. + +In this paper, we explore the view-consistency problem in scene-adaptive low-light LF image enhancement. We first analyze the disparity consistency between the illumination adjustment maps of different views. Based on the photometric consistency theory, we propose a self-supervised view-consistent illumination loss to keep the consistency between the estimated illumination maps. These illumination maps are estimated by combining the global brightness of the scene and the local brightness of each view, which are used to light up the low-light LF images. Subsequently, we perform corruption recovery on the light-up LF images to further suppress noise, artifacts, under-/overexposure, and color distortion. Comprehensive experiments are conduct in dynamic and fixed illumination environments. The related results are shown in Fig. 1, in which the proposed View-Consistent Network (VCNet) outperforms others over 7 db. More importantly, we also add our proposed illumination adjustment module to other models, labeled as “+E”. The experiments show that our proposed illumination adjustment module can be embedded into other methods and greatly improve their performance in terms of both image quality and view consistency preservation without increasing a large number of parameters. + +# 2. Related Work + +Low-Light Image Enhancement. The low-light image enhancement methods are mainly divided into two categories. The first category directly learns a brute-force mapping function from the low-light image to its normal-light counterpart. An exemplary illustration is the SID + +approach [3], which utilizes an encoder-decoder architecture to effectively restore raw images captured under extremely low-light conditions. DID [19] introduced an innovative residual learning network explicitly crafted for the denoising of low-light images, surpassing the capabilities of SID. Guo et al. [9] designed a Zero-DCE network to employ dynamic range adjustment through pixel-wise and high-order curve estimations. However, these methods lack interpretability and theoretically proven properties. The methods in the second category are designed based on Retinex theory for restoration. Wei et al. [34] and follow-up works [40, 41] combine the Retinex decomposition with a multi-stage training pipeline. In the first training stage, these methods learn a network to decompose the images. Then in the second stage, they employ a different network to denoise the reflectance and adjust the illumination, respectively. + +Low-Light Video Enhancement. Different from low-light image enhancement, methods that enhance videos recorded under low-light conditions extract valuable complementary details from neighboring frames for better restoration. Wang et al. [26] delved into the field of physical modeling, utilizing an LSTM-based network to improve the quality of low-light videos. Based on Retinex theory, Wang et al. [24] introduced a comprehensive methodology to accomplish simultaneous enhancement of low-light conditions and noise reduction in video sequences. Even though these methods achieved remarkable results in enhancing two-dimensional low-light images and videos, it is still difficult for them to handle the four-dimensional LF dataset with regular sampling in the angular domain. + +Low-Light LF Image Enhancement. For LF images, some methods directly perform the restoration of low-light to normal illumination images. Lamba et al. [15] made substantial contributions in this regard by introducing a lowlight LF images dataset and a two-stage L3Fnet architecture. This architecture consists of a global representation block, responsible for encoding the intricate LF geometry, and a view reconstruction block that concentrates on restoring each distinct view. To address the challenge of enhancing dim LF images, they also introduced a three-stage network [14] that effectively utilizes global, local, and view-specific features inherent to low-light LF images. Wang et al. [28] introduced a coarse-to-fine restoration network to progressively restore the low-light LF input. In their network, three types of view stacks are fed into three streams to explore the redundant information of LFs. Recently, Wang et al. [29] proposed a novel feature extraction approach that parallelly processes features across multiple spatial scales. Lyu et al. [39] proposed a multi-stage architecture that mimics the optimization process of solving an inverse imaging problem, in which the illumination maps are estimated and employed in the unfolding process to produce + +![](images/ae23a6b98edde3010d7d9a93ffe200085fe5003214ac1676168f34096dabf606.jpg) +Figure 2. Illustration of our self-supervised view-consistent loss. The reference point $\mathbf{x}$ has the disparity $d(\mathbf{x}) = 1$ . Compared with other points (labeled as red and purple), the corresponding point (labeled as orange) has the minimum variance. We propose to directly use the minimum variance as our objective function $\ell_{vc}$ to minimize. + +the new enhanced result. There are also approaches to divide complex tasks into several intermediate tasks. Inspired by Retinex theory, Zhang et al. [37] proposed a learning-based decomposition-enhancement methodology that systematically dissects the intricate enhancement task into illumination adjustment and reflection denoising. Zhang et al. [39] separate the complex task into several intermediate tasks and use multiple heads to restore low-light images. They also formulate a synthesis pipeline to enrich the available LF data pairs, in which the ground truth illumination maps are generated and used to guide illumination adjustment. Although they consider the dynamically degraded inputs, their model needs the ground truth illumination value to give full supervision in the training process, which is unavailable for real-world images. Moreover, these methods do not consider the consistency constraints of the illumination maps, so that the view consistency of the recovered image is easily broken. + +# 3. View-Consistent Network + +According to the two-plane model [16], a LF can be formulated as a 4D tensor $\mathcal{L}(u,v,x,y)\in \mathbb{R}^{U\times V\times H\times W}$ , where $\mathbf{u} = (u,v)$ denote the angular coordinates, and $\mathbf{x} = (x,y)$ represent the spatial coordinates. Commonly, the LF image is organized as a $U\times V$ array of views that captured the same scene in one shot. One specific view at angular coordinate $(u,v)$ can be denoted as $\mathcal{L}_{\mathbf{u}}\in \mathbb{R}^{H\times W}$ . + +# 3.1. View Consistency Analyses + +# 3.1.1. Retinex Theory for Single View + +Drawing upon the Retinex theory, each low-light view in LF can be mathematically expressed as the product of its reflectance and illumination components as: + +$$ +I _ {\mathbf {u}} = L _ {\mathbf {u}} \odot R _ {\mathbf {u}}, \tag {1} +$$ + +where $L_{\mathbf{u}}$ , $R_{\mathbf{u}}$ represents the illumination components, reflectance components with angular coordinates $\mathbf{u}$ , respectively. $\odot$ denotes the element-wise multiplication. Similar to [7, 11, 22], we regard $R_{\mathbf{u}}$ as a well-exposed LF view. To light up $I_{\mathbf{u}}$ , we element-wisely multiply the two sides of Eq. 1 by an illumination adjustment map $\overline{L}_{\mathbf{u}}$ such that $\overline{L}_{\mathbf{u}} \odot L_{\mathbf{u}} = 1$ as follow: + +$$ +R _ {\mathbf {u}} = I _ {\mathbf {u}} \odot \bar {L} _ {\mathbf {u}}. \tag {2} +$$ + +Therefore, for LF images, we need to estimate the illumination adjustment map $\overline{L}_{\mathbf{u}}$ for all views in one LF. + +# 3.1.2. View Consistency in Illumination Maps + +Since views in one LF are taken from different angulars of the same scene, in which the depth of the scene is fixed, the views are supposed to satisfy the disparity constraint. Suppose that the parameter of the camera is known, the disparity of two views $I_{\mathbf{u_j}}$ and $I_{\mathbf{u_i}}$ is calculated as: + +$$ +D _ {\mathbf {u} _ {\mathbf {j}} \rightarrow \mathbf {u} _ {\mathbf {i}}} (\mathbf {x}) = \left(\mathbf {u} _ {\mathbf {j}} - \mathbf {u} _ {\mathbf {i}}\right) * \frac {B * f}{Z (\mathbf {x})} = \left(\mathbf {u} _ {\mathbf {j}} - \mathbf {u} _ {\mathbf {i}}\right) * d (\mathbf {x}), \tag {3} +$$ + +where $B$ and $d$ represent the baseline length and the disparity between adjacent views, respectively. $f$ represents the focal length of the camera. $Z$ represents the depth of the scene. Therefore, the two captured low-light views satisfy the disparity constraint as: + +$$ +I _ {\mathbf {u} _ {\mathrm {i}}} = W \left(I _ {\mathbf {u} _ {\mathrm {j}}}, D _ {\mathbf {u} _ {\mathrm {j}} \rightarrow \mathbf {u} _ {\mathrm {i}}}\right), \tag {4} +$$ + +where $W$ denotes the warp operations according to the disparity $D_{\mathbf{u_j}\rightarrow \mathbf{u_i}}$ . Similarly, the well-exposed LF views also satisfy the disparity constraint as: + +$$ +R _ {\mathbf {u} _ {\mathrm {i}}} = W \left(R _ {\mathbf {u} _ {\mathrm {j}}}; D _ {\mathbf {u} _ {\mathrm {j}} \rightarrow \mathbf {u} _ {\mathrm {i}}}\right). \tag {5} +$$ + +Therefore, according to Eq. 2, the illumination adjustment map should also satisfy the disparity constraint. + +$$ +\bar {L} _ {\mathbf {u} _ {\mathrm {i}}} = W \left(\bar {L} _ {\mathbf {u} _ {\mathrm {j}}}; D _ {\mathbf {u} _ {\mathrm {j}} \rightarrow \mathbf {u} _ {\mathrm {i}}}\right). \tag {6} +$$ + +In order to constrain the consistency of the illumination adjustment maps, the direct way is to minimize the distance between the illumination map $\overline{L}_{\mathbf{u_i}}$ and the warped illumination map $W(\overline{L}_{\mathbf{u_j}};D_{\mathbf{u_j}\to \mathbf{u_i}})$ from different angular $\mathbf{u_j}$ as: + +$$ +\min \sum_ {\mathbf {u} _ {\mathrm {j}}} \sum_ {\mathbf {u} _ {\mathrm {i}}} | | W (\bar {L} _ {\mathbf {u} _ {\mathrm {j}}}; D _ {\mathbf {u} _ {\mathrm {j}} \rightarrow \mathbf {u} _ {\mathrm {i}}}) - \bar {L} _ {\mathbf {u} _ {\mathrm {i}}} | |, \tag {7} +$$ + +where $\mathbf{u_j},\mathbf{u_i}\in \{(0,0),(0,1),\dots,(U,V)\}$ . To simplify the computation, we warp the illumination adjustment maps of different views towards the central view $\overline{L}_{\mathbf{u_c}}$ and construct these aligned views to a 4D array $\mathbf{A}\in \mathbb{R}^{U\times V\times H\times W}$ .. + +$$ +\mathbf {A} \left(\mathbf {u} _ {\mathbf {j}}, :\right) = W \left(\bar {L} _ {\mathbf {u} _ {\mathbf {j}}}; D _ {\mathbf {u} _ {\mathbf {j}} \rightarrow \mathbf {u} _ {\mathbf {c}}}\right), \tag {8} +$$ + +where $\mathbf{u}_{\mathbf{c}} = \left(\frac{U}{2},\frac{V}{2}\right)$ . Then, we instead calculate the variance of the warped illumination adjustment maps to assess their disparity consistency as: + +$$ +\min \sum_ {\mathbf {x}} \operatorname {V a r} (\mathbf {A} (:, \mathbf {x})), \tag {9} +$$ + +![](images/4b6f7d130f8f003ae6ffa931942e1e2661c72893d72447eea5d374d25cf36be6.jpg) +Figure 3. The proposed VCNet network consists of an illumination estimator VCAIA and a corruption restorer. In VCAIA, the illumination map is estimated and multiplied to the original input to light up the illumination. VCRM further recovers the details of the light-up image and outputs the final results. + +where $Var(\mathbf{A}(:,\mathbf{x}))$ calculates the variance for each pixel along the angular dimension. + +However, in real scenes, both the ground truth illumination adjustment map $\overline{L}_{\mathbf{u}}$ and the corresponding disparity $d$ is unavailable. Moreover, the disparity is also difficult to estimate using low-light images with much noise [12]. Therefore, the disparity constraint of the illumination adjustment maps as in Eq. 9 is difficult to implement. + +# 3.2. Self-Supervised View-Consistent Loss + +In this section, we relax the disparity constraint and design a self-supervised view-consistent illumination loss to ensure the view consistency of illumination adjustment maps. Instead of constraining each point according to its true disparity as in Eq. 9, we only constrain each point according to one specific disparity value, no matter whether it is the true disparity or not. Specifically, we first set $N$ discrete candidate disparity values: + +$$ +d ^ {n} = d _ {\min } + n * \Delta d, n \leq N \tag {10} +$$ + +where $\Delta d = (d_{max} - d_{min}) / N$ denotes the granularity of $d$ . $d_{min}$ and $d_{max}$ denote the minimum and maximum disparity of the scene. Subsequently, we warp each view to the center view according to all the possible disparity values $d^n$ , and the related 4D array is denoted as $\mathbf{A}^n \in \mathbb{R}^{U \times V \times H \times W}$ : + +$$ +\mathbf {A} ^ {n} \left(\mathbf {u} _ {\mathrm {j}}, :\right) = W \left(\bar {L} _ {\mathbf {u} _ {\mathrm {j}}}; D _ {\mathbf {u} _ {\mathrm {j}} \rightarrow \mathbf {u} _ {\mathrm {c}}} ^ {n}\right), \tag {11} +$$ + +where $D_{\mathbf{u}_i - > \mathbf{u_c}}^n (\mathbf{x}) = (\mathbf{u_j} - \mathbf{u_c})*d^n$ + +We then calculate the variance of the warped view for each point $\mathbf{x}$ and for each possible disparity $n$ as $Var(\mathbf{A}^n (:, \mathbf{x}))$ , constructing the commonly known cost volume as in the disparity estimation methods [4, 13]. According to the photometric consistency, the variance under the true disparity is supposed to have the minimum value compared with other disparities. Therefore, we first find the specific disparity $d^{*}$ for each point that has the minimum variance: + +$$ +d ^ {*} (\mathbf {x}) = \arg \min _ {d ^ {n}} (V a r \left(\mathbf {A} ^ {n} (:, \mathbf {x}))\right)). \tag {12} +$$ + +![](images/4029e28c722ecfd037184699b092a95911c66b7f899962fb454d4021cc4d4286.jpg) +Figure 4. The estimated illumination EPI w/ and w/o $loss_{vc}$ . The former keeps the view consistent as the true EPI with normal light. + +We then minimize the variance related to the $d^{*}(\mathbf{x})$ to make sure that all views are aligned in the disparity $d^{*}(\mathbf{x})$ : + +$$ +\ell_ {v c} = \sum_ {\mathbf {x}} \operatorname {V a r} \left(\mathbf {A} ^ {*} (:, \mathbf {x})\right), \tag {13} +$$ + +where $\mathbf{A}^*$ represents the 4D array with aligned views according $D_{\mathbf{u_j - > u_c}}^* (\mathbf{x}) = (\mathbf{u_j} - \mathbf{u_c})*d^* (\mathbf{x})$ as in Eq. 11. An example is shown in Fig. 2, where the minimum variance is set as our objective function. In such a way, each pixel is assigned a disparity value that has the minimum variance to make the illumination adjustment value of each pixel as consistent as possible with other views. + +# 3.3. Adaptive Illumination Adjustment + +Beginning with the input low-light LF image $\mathcal{L}^{in}$ , we design a simple illumination estimator, View Consistent Adaptive Illumination Adjustment (VCAIA), $\mathcal{E}$ for LFs to estimate the illumination adjustment map. Specifically, as in Fig. 3, we first estimate the global illumination adjustment matrix of the scene by integrating information from all views. The local brightness refinement is then developed to adjust the matrix according to the details of each view. + +In order to integrate the abundant information from all views and further reduce the effect of noise, we roughly calculate the average of all views across all color channels in LF image as $I_w \in \mathbb{R}^{H \times W \times 1}$ . Then following [1], we calculate the global illumination adjustment matrix as: + +$$ +\overline {{I ^ {l}}} _ {\text {g l o b a l}} (\mathbf {x}) = \frac {\log \left(I _ {w} (\mathbf {x}) / \mu + 1\right)}{\log \left(\max \left(I _ {w} (\mathbf {x})\right) / \mu + 1\right) \cdot I _ {w} (\mathbf {x})}, \tag {14} +$$ + +Subsequently, we replicate $\overline{I}^{l}_{global}$ to all views and obtain the global illumination adjustment matrix $\overline{\mathcal{L}}^{l}_{global}$ . To further refine illumination estimation for each view, we employ a lightweight learnable network. Specifically, we first stack all views of the input low-light LF image and use the cascaded 3D spatial convolution layers to obtain the local brightness adjustment matrix $\overline{\mathcal{L}}^{l}_{local}$ of each view. + +The local and global illumination adjustment matrix is combined to derive the final adjustment matrix and is then multiplied to $\mathcal{L}^{in}$ to obtain the adjusted high-illumination image: $\overline{\mathcal{L}} = (\overline{\mathcal{L}^t}_{global} + \overline{\mathcal{L}^t}_{local})$ . By minimizing the proposed self-supervised loss $\ell_{vc}$ on the estimated $\overline{\mathcal{L}}$ according + +Table 1. Quantitative Comparisons of Low Light LF Image Enhancement Methods in L3F-Dynamic Task w/ w/o VCAIA. + +
MethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LFFT-DCU23.170.80030.095819.600.72010.170219.820.67180.2074
DRLF19.100.69470.188817.420.66000.240217.040.63220.2608
DRLF+E22.19 (+3.09)0.7636 (+0.0689)0.1124 (-0.0764)21.38 (+3.96)0.6976 (+0.0376)0.1676 (-0.0726)20.29 (+3.25)0.6326 (+0.0004)0.2265 (-0.0343)
MSP21.370.75600.133018.640.68980.207917.620.65430.2407
MSP+E22.80 (+1.43)0.7936 (+0.0376)0.1261 (-0.0069)21.81 (+3.17)0.7436 (+0.0538)0.1673 (-0.0406)21.10 (+3.48)0.6916 (+0.0373)0.2128 (-0.0279)
VCNet-E22.540.78350.125120.830.71030.201519.920.64500.2666
VCNet24.58 (+2.04)0.8220 (+0.0385)0.0897 (-0.0354)24.06 (+3.23)0.7652 (+0.0549)0.1292 (-0.0723)23.09 (+3.17)0.7032 (-0.0582)0.1981 (-0.0685)
+ +to Eq. 13, we constrain the calculated illumination adjustment maps to satisfy the view consistency even without the true disparities. Since no supervised images are needed, the illumination adjustment can be applied to various low-light LF enhancement networks. We show one estimated example in Fig. 4, where the Epipolar Plane Images (EPIs) are calculated by fixing $v$ in angular domain and $w$ in the spatial domain. Without the proposed $\ell_{vc}$ , the lines in the EPIs of the estimated illumination maps become ambiguous and are inconsistent with the ground truth EPIs of the normallight LFs. In contrast, when we use $\ell_{vc}$ to constrain the view consistency, the lines in estimated EPIs of illumination maps retain the direction as the ground truth EPIs. + +# 3.4. Recovery Module + +After illumination adjustment, the image $\mathcal{L}^a = \overline{\mathcal{L}}\odot \mathcal{L}^{in}$ with a standard illumination level is obtained. However, the light-up process also amplifies the noise and artifacts, causing under-/over-exposure and color distortion. Therefore, similar to other retinae-based methods, a recovery module is needed to suppress noise and restore scene details. Some existing networks, such as the LF denoising model DRLF and the LF low-light enhancement model MSP, can be used as the recovery module. + +We also design a simple View Consistent Recovery Module (VCRM) to further import the inherent disparity constraint into the recovery process to enhance the recovery effect. VCRM adopts a multi-scale architecture. In each scale, the view consistent feature aggregation unit is implemented to explore the redundant information of the views to recover details. In each unit, we employ deformable convolution layers on EPIs to further integrate information between different views. Since the slope of lines in EPIs corresponds to disparity, we introduce the disparity constraint to help the model identify correct and consistent complementary information. Specifically, we use uniform offsets in both V-H and U-W spaces, so that the angular information can be efficiently incorporated and maintain epipolar consistency. Finally, we concatenate multi-scale features and use a decoder with a similar structure to the encoder to aggregate them and obtain the refined result $\mathcal{L}^{out}$ . More detailed design is provided in the supplementary details. + +# 3.5. Implementation Details + +As in Fig. 3, the overview architecture of our VCNet primarily employs two VCAIA and VCRM. We train our VCNet by minimizing a hybrid loss: $\ell_{total} = \lambda_1\ell_{vc} + \lambda_2\ell_{rec} + \lambda_3\ell_{ssim}$ . $\ell_{rec}$ is the reconstruction loss, which calculates the L1 loss between the ground truth of well-posed image $\mathcal{L}^{gt}$ and the recovered results $\mathcal{L}^{out}$ of VCNet. The SSIM loss [33] $\ell_{ssim}$ measures the structural similarity between $\mathcal{L}^{gt}$ and $\mathcal{L}^{out}$ . $\ell_{vc}$ is the proposed self-supervised illumination loss in Sec.3, which is used to keep the view consistent of the estimated illumination maps $\overline{\mathcal{L}}^l$ . The three hyperparameters $\lambda_1$ , $\lambda_2$ and $\lambda_3$ are empirically set to 0.5, 1, 0.5. + +Our model is implemented using the PyTorch framework and trained on a NVIDIA RTX A4000 GPU. The model is optimized using Adam optimization with $\beta_{1} = 0.9$ and $\beta_{2} = 0.999$ . The batch size is set to 1 and the learning rate is initially set to 0.0001. Data augmentation techniques such as random cropping, vertical and horizontal flipping are used. During training, the views are randomly cropped in patches of size $128 \times 128$ . During testing, we use the original resolution LF data without applying data augmentation. + +# 4. Experiments + +# 4.1. Dataset + +We evaluate the proposed method using the publicly accessible Low-Light-Light-Field (L3F) Dataset and L3F-Wild Dataset from [15]. The L3F dataset comprises 27 distinct scenes, each of which was captured under three low-light conditions, labeled L3F-20, 50, 100, and one with normal lighting. The L3F-Wild is captured in almost 0 lux conditions and captured with $\frac{1}{4} -\frac{1}{15}$ second exposure time. + +L3F-Fixed Task: Previous LF-based low-light enhancement methods trained their models under fixed illumination conditions. Specifically, following [15], 18 scenes under the same illumination condition are selected for training, while the remaining 9 scenes are reserved for testing. In the fixed illumination conditions, 3 models are specifically trained for L3F-20, 50, 100 in different illumination conditions. We denote it as the L3F-Fixed task. + +L3F-Dynamic Task: To further assess the model's capability to adapt to various illumination conditions, we establish a dataset with dynamic illumination as L3F-Dynamic. + +![](images/6e2e38008af38050d261af8f589c35515d89bd0039a5ecd1d99ba4aee0b88eb7.jpg) +Figure 5. Visual comparison on $L3F$ -Dynamic task, where the central view and EPIs of the recovery LFs are shown. Other algorithms either generate over-exposed and noisy images or introduce black spot corruptions and unnatural artifacts. Our VCNet restores well-exposed structural contents and textures. By adding our VCAIA module, the recovered results of all models are closer to ground truth. + +Specifically, we combine images from the L3F-20, 50, 100 datasets. We use all training sets from the three different illumination conditions (including $18 \times 3$ scenes) to train a single model. The single trained model is supposed to handle dynamic illumination conditions. We use the same testing dataset as L3F-Fixed to evaluate its performance. Compared to L3F-Fixed task, L3F-Dynamic task is more difficult, which needs to train one single model that is able to handle all different illumination conditions. + +# 4.2. Compare with State-of-the-Art Methods + +We compare our VCNet and several SOTA methods, including 3 recent published single-frame retinae-based methods (RFormer [22], Pydiff [42] and DarkIR [6]) and 9 LF-based methods (L3F-Net [15], FER-Net [14], DRLF [10], MSP-3 [27], IINet [17], DistgSSR [31], PMSNet [29], LRT [39], LFLL-DCU [18]). The illumination estimator $\mathcal{E}$ is applied to the LF denoising model DRLF and low-light enhancement model MSP, denoted as DRLF+ $\mathcal{E}$ , MSP+ $\mathcal{E}$ . Specifically, we embed the illumination estimator $\mathcal{E}$ into the beginning of these two models and add the $\ell_{vc}$ in their training strategy. Our VCNet- $\mathcal{E}$ is trained by deleting $\mathcal{E}$ module and only minimizing $\lambda_2\ell_{rec} + \lambda_3\ell_{ssim}$ . The reference-based metric PSNR, SSIM [32] and LPIPS [36] and nonreference metric NIQE [21] and BRISQUE [20] are chosen as the evaluation metrics. + +Dynamic Illumination Task. We first assess our model's ability to adapt to dynamic illumination. Specifically, we train each method using the L3F-Dynamic training dataset + +Table 2. Quantitative Comparisons with Single-Frame-based Low Light Enhancement Methods on the L3F-dynamic Task. + +
MethodL3F-20L3F-50L3F-100
PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑
RFormer20.36 / 0.77 / 0.342 / 50.9419.30 / 0.70 / 0.493 / 39.1917.13 / 0.62 / 0.615 / 30.79
RFormer*27.29 / 0.87 / 0.163 / 67.2623.81 / 0.77 / 0.294 / 42.3122.51 / 0.70 / 0.374 / 34.84
Pydiff19.74 / 0.71 / 0.406 / 48.7219.13 / 0.62 / 0.535 / 36.3018.47 / 0.49 / 0.757 / 31.14
Pydiff*23.42 / 0.82 / 0.256 / 60.1122.74 / 0.76 / 0.331 / 40.9622.24 / 0.70 / 0.437 / 33.51
DarkIR20.11 / 0.76 / 0.315 / 51.6317.24 / 0.65 / 0.474 / 35.3715.09 / 0.58 / 0.621 / 15.31
VCNet24.58 / 0.82 / 0.089 / 79.0724.06 / 0.77 / 0.129 / 78.4623.09 / 0.70 / 0.198 / 76.80
+ +in Sec. 4.1 and the results are presented in Table 1. By adding the illumination estimator $\mathcal{E}$ , all three models exhibit around 3dB PSNR improvement by considering the adaptive illumination estimation. Since LFLL-DCU only uses one learnable parameter to estimate the illumination map, their model performs well on the L3F-20 data but poorly on L3F-50 and L3F-100. Our VCNet outperforms the current SOTA methods MSP, DRLF, and LFLL-DCU with more than $2\sim 6$ dB PSNR improvement. Fig. 5 displays the visual results comparison, where all visual effects are improved by combining the $\mathcal{E}$ . This clearly demonstrates the effectiveness of our illumination estimator $\mathcal{E}$ in adaptively adjusting to varying illumination conditions. The results of VCNet stand out by producing more realistic overall color compared with the ground truth. Other methods restore blurred scene details, while our VCNet recovers the clearer and richer detailed texture in the recovered images. + +We then make a comparison with single-frame-based methods in Table 2, which claim to handle dynamic lighting conditions. Their models are trained on a large-scale + +![](images/3c230fdd3242aa0ff6476bda06fd48908e7c979cccee6a3fbac76470bb7d2f61.jpg) +Figure 6. Visual comparison on the $L3F$ -Wild dataset, where the central view of LFs is shown. Other methods either collapse by noise, distort color, or under-/over-exposed images. Our VCNet effectively removes the noise and reconstructs well-exposed image details. + +Table 3. Quantitative Comparison on Real-world L3F-Wild Data. + +
MethodsRFormer*Pydiff*DRLFDRLF+EMSPMSP+EVCNet-EVCNet
NIQE ↓15.6717.8219.9718.7818.8218.2317.8714.22
BRISQUE↓31.2036.1232.8030.5430.2029.4230.1125.98
TIMES↓39.84100.1216.2216.263.893.935.996.33
+ +single-image dataset and we carefully fine-tune their methods using L3F-Dynamic training dataset (labelled as *), where each view of LF is considered a sample during the training. Our method outperforms others under most lighting conditions, especially in perceptual quality. Although RFormer* achieves relatively high PSNR and SSIM values in the L3F-20 dataset, it suffers from noticeable blurry artifacts, as shown in Fig. 5. + +To further demonstrate the generalization ability, we applied the methods trained on the $L3F$ -Dynamic dataset to another real-world low-light LF data $L3F$ -Wild. The results are shown in Fig. 6 and Table 3. Due to the absence of redundant information from different views, the results of single-frame methods are relatively dark, and the restored details suffer from blurring. For LF-based methods, DRLF is observed to cause a noticeable color shift. The results of MSP and LFLL-DCU suffer from residual noise and local over-expose. Combining with the $\mathcal{E}$ , the recovered results of MSP and DNLF have higher color fidelity, clearer structural details and less noise. In contrast, our results achieve high perceptual quality with clear structures, realistic color and less noises. Our VCAIA module $\mathcal{E}$ is also proved that can enhance the perceptual quality of other methods. + +Fixed Illumination Task. We then evaluate all the methods under the fixed illumination condition by separately training 3 modals using the $L3F$ -Fixed dataset. The results using LFs with angular resolution $5 \times 5$ and $8 \times 8$ are illustrated in Table 4, respectively. We retrain the LRT [39] and LFLL-DCU [18] for comparison. When compared with single-frame-based methods, LF-based methods demonstrate higher performance, highlighting the utility of redundant LF information in low-light enhancement. With the increase of the input image views, the recovery effect of the LF models is better because it provides more redundant information to the network. Note that in PMSNet, + +Table 4. Quantitative Comparisons on L3F-fixed Task. + +
Ang ResMethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
8L3F-Net24.780.82623.210.75221.830.683
IINet25.400.86224.140.81122.130.721
DistgSSR25.440.86424.180.81022.960.755
PMSNet27.030.86624.520.81222.740.762
VCNet26.840.87125.410.82323.580.772
5L3F-Net24.980.81423.370.75321.950.669
FER-Net25.310.82323.400.76621.730.688
LRT25.970.81523.990.74222.150.691
LFLL-DCU25.420.84224.030.78722.280.702
DRLF24.780.806422.520.721721.510.6543
DRLF+ε24.990.817923.060.732922.190.6681
MSP25.480.830124.100.775922.750.7133
MSP+ε25.730.843024.680.783123.060.7191
VCNet-ε25.870.837224.350.789722.690.7204
VCNet26.610.847325.130.792123.240.7223
+ +in addition to the original $8 \times 8$ views, views from two more rows and columns are included in the feature extraction. Among LF-based methods, our VCNet significantly outperforms SOTA methods on most of the $L3F$ -fixed task, demonstrating its ability to better handle low-illumination degradation and produce superior restoration results. After adding the illumination estimator $\mathcal{E}$ , all models also achieve around a 0.2-0.8 dB PSNR improvement. + +View Consistency Comparison. View consistency holds great significance in LF enhancement as it directly impacts the performance of subsequent LF-based applications like depth estimation. To verify the effectiveness of our VCAIA $\mathcal{E}$ in maintaining consistency in recovery results, we present the parallax content Precision-Recall (PR) curves used in [5] for evaluation. In one LF, each view is subtracted from the central view, and different threshold values are chosen to calculate binary edge maps. The PR curves show how closely the binary edge maps of the reconstructed LFs are match to the ground truth. We use models trained on dynamic illumination for comparison. In Fig. 7, when all the methods add the illumination adjustment, the PR curve of the recovered image is improved, demonstrating that our $\mathcal{E}$ can achieve view-consistent illumination adjustment. We further calculate the Average Precision (AP) by measuring the area under the PR curves for quantitative + +![](images/79a106eca2f9b6bb151d61406fd641448592f227c717d3b1b3ddac630ae229b9.jpg) +Figure 7. The PR curves map comparison of L3F-20, 50, 100 dataset, in which how well the views consistency preserved is shown. With $\mathcal{E}$ , the PR curves of all methods are better. + +comparison. As in Table 2, for the single-frame methods, it is difficult to recover views with consistent illumination, so that the AP is relatively low even with high PSNR and SSIM. By contrast, our model exhibits notably higher AP values, especially under the challenging low-light conditions of the L3F-100 dataset. The comparison with LF-based methods is shown in Fig. 1. Our model consistently achieves the highest AP values, showcasing the superiority of VCNet in terms of low-light enhancement and consistency preservation. The EPIs of the recovered images are also visualized in Fig. 5, in which the orientation of lines corresponds to the disparity information. It is also clear that our model excels in maintaining the view consistency. More results are provided in the supplementary materials. + +Running time. We further show the running time of the models using the size of the scatter in Fig. 1 and Table 3. Our model runs faster (6.3s) than DRLF (16.2s) but is slightly slower than MSP (3.9s). The computation time of $\mathcal{E}$ is 0.04s on the A5000 for a $5 \times 5 \times 642 \times 432$ input, which is lightweight to add $\mathcal{E}$ to different models. + +# 4.3. Ablation Study + +Table 5. The Ablation Study on L3F-Dynamic Task. + +
No.LocalGlobal\( \ell_{vc} \)PSNRSSIMLPIPS
(1)×××20.760.700.2036
(2)×√×22.780.750.1469
(3)√√×23.590.750.1420
(4)√√√23.840.760.1395
+ +We conduct a series of ablation studies under dynamic illumination conditions with $5 \times 5$ LF images. To ensure fairness, we keep the parameters of the variant models close to VCNet by adjusting the feature channels. It is clear that the performance of the VCNet without $\mathcal{E}$ (model-1) is the worst. Compared with model-1, model-2 using the global brightness refinement performs better. Meanwhile, with the local brightness adaptive adjustment branch, model-3 increases $0.8\mathrm{dB}$ in the PSNR values than model-2. We further validate the effectiveness of the loss $\ell_{vc}$ . Comparing model-4 with model-3, the PSNR increases nearly $0.4\mathrm{db}$ by adding the $\ell_{vc}$ to supervise the illumination adjustment matrix $\overline{\mathcal{L}^l}$ . This shows that $\ell_{vc}$ can perform consistency constraints on $\overline{\mathcal{L}^l}$ to provide higher-quality light-up images. + +We further visualize the adaptive illumination adjust + +![](images/c7db31cb773bc26ebebcca0c4e7cc7a1a92e5c1f644f02c9423305cdc95a46b5.jpg) +Figure 8. Visualization of the illumination maps in $L3F$ -Dynamic task. For darker scenes, the values in the estimated illumination adjustment map is larger. + +![](images/c1c25df274ae9ab220fd65b2f8f71395366a054a2b5cbe128495740dc1b1ce1c.jpg) +Figure 9. Visual comparison of the enhanced result (top row) and input central view (second row) of one scene with different low-light levels in the L3F-Wild dataset. + +ment matrix $\overline{\mathcal{L}_l}$ and the related recovered results in different illumination environments in Fig. 8. The model is trained using the $L3F$ -dynamic dataset. For input with an extended exposure time, the generated illumination adjustment matrix tends to be relatively darker. Conversely, as the exposure time diminishes, the input images become darker and the values of the adjustment matrix gradually increase. We also illustrate the performance of our VCNet in processing images with disparate brightness levels from the L3F-Wild dataset in Fig. 9. Our method consistently achieves superior enhancement quality across input scenes with varying levels of brightness. This consistency ensures effective low-light image restoration across varying illumination conditions. + +# 5. Conclusion + +In this paper, we introduced a novel framework for scene-adaptively low-light enhancement in light field(LF) images. First, by analyzing Retinex theory and extending it to the LF, we introduce a self-supervised loss to ensure that the calculated illumination map adheres to the consistency relationship. Then, we design an illumination estimator with global adjustment and local refinement to perceive the illumination of the scene. Finally, we design a corruption restorer and import the consistency constraint into the feature aggregation process for better recovery. Extensive quantitative and qualitative experiments demonstrate that our VC-Net significantly outperforms state-of-the-art methods on both fixed and dynamic illumination enhancement tasks. + +Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032). + +# References + +[1] Hyunchan Ahn, Byungjik Keum, Daehoon Kim, and Hwang Soo Lee. Adaptive local tone mapping based on retina for high dynamic range images. In Proceedings of the International Conference on Consumer Electronics (ICCE), pages 153-156, 2013. 4 +[2] Tom E. Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):972-986, 2012. 1 +[3] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 3291-3300, 2018. 2 +[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4 +[5] Jie Chen, Junhui Hou, and Lap-Pui Chau. Light field denoising via anisotropic parallax analysis in a cnn framework. IEEE Signal Processing Letters, 25(9):1403-1407, 2018. 7 +[6] Daniel Feijoo, Juan C. Benito, Alvaro Garcia, and Marcos V. Conde. Darkir: Robust low-light image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 10879-10889, 2025. 6 +[7] Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2782-2790, 2016. 3 +[8] Chunle Guo, Jing Jin, Junhui Hou, and Jie Chen. Accurate light field depth estimation via an occlusion-aware network. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1 +[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1777-1786, 2020. 2 +[10] Mantang Guo, Junhui Hou, Jing Jin, Jie Chen, and LapPui Chau. Deep spatial-angular regularization for light field imaging, denoising, and super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6094-6110, 2022. 6 +[11] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing, 26(2):982-993, 2017. 3 +[12] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 19-34, 2017. 4 +[13] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 4 +[14] Mohit Lamba and Kaushik Mitra. Fast and efficient restoration of extremely dark light fields. In Proceedings of the Win- + +ter Conference on Applications of Computer Vision (WACV), pages 3152-3161, 2022. 2, 6 +[15] Mohit Lamba, Kranthi Kumar Rachavarapu, and Kaushik Mitra. Harnessing multi-view perspective of light fields for low-light imaging. IEEE Transactions on Image Processing, 30:1501-1513, 2021. 1, 2, 5, 6 +[16] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. ACM, 1996. 3 +[17] Gaosheng Liu, Huanjing Yue, Jiamin Wu, and Jingyu Yang. Intra-inter view interaction network for light field image super-resolution. IEEE Transactions on Multimedia, 25: 256-266, 2023. 6 +[18] Xianqiang Lyu and Junhui Hou. Enhancing low-light light field images with a deep compensation unfolding network. IEEE Transactions on Image Processing, 33:4131-4144, 2024. 2, 6, 7 +[19] Paras Maharjan, Li Li, Zhu Li, Ning Xu, Chongyang Ma, and Yue Li. Improving extreme low-light image denoising via residual learning. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 916-921, 2019. 2 +[20] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6 +[21] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a "completely blind" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6 +[22] Junxiang Ruan, Xiangtao Kong, Wenqi Huang, and Wenming Yang. Retformer: Retinex-based enhancement in transformer for low-light image. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2, 3, 6 +[23] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 6842-6850, 2019. 2 +[24] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the International Conference on Computer Vision (ICCV), pages 9680-9689, 2021. 2 +[25] Tingxian Wang, Alexei A. Efros, and Ravi Ramamoorthi. Depth estimation with occlusion modeling using light-field cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:2170-2181, 2016. 1 +[26] Wei Wang, Xin Chen, Cheng Yang, Xiang Li, Xuemei Hu, and Tao Yue. Enhancing low light videos by exploring high sensitivity camera noise. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4110-4118, 2019. 2 +[27] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 6 + +[28] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 1, 2 +[29] Xingzheng Wang, Kaiqiang Chen, Zixuan Wang, and Henhao Huang. Pmsnet: Parallel multi-scale network for accurate low-light light-field image enhancement. IEEE Transactions on Multimedia, 26:2041-2055, 2024. 2, 6 +[30] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the International Conference on Computer Vision (ICCV), pages 13128-13137, 2023. 2 +[31] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2023. 6 +[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6 +[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5 +[34] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2 +[35] Henry Wing Fung Yeung, Junhui Hou, Xiaoming Chen, Jie Chen, Zhibo Chen, and Yuk Ying Chung. Light field spatial super-resolution using deep efficient spatial-angular separable convolution. IEEE Transactions on Image Processing, 28(5):2319-2330, 2018. 1 +[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2018. 6 +[37] Shansi Zhang and Edmund Y. Lam. An effective decomposition-enhancement method to restore light field images captured in the dark. Signal Processing, 189:108279, 2021. 1, 3 +[38] Shuo Zhang, Youfang Lin, and Hao Sheng. Residual networks for light field image super-resolution. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 11046-11055, 2019. 1 +[39] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Lrt: An efficient low-light restoration transformer for dark light field images. IEEE Transactions on Image Processing, 32:4314-4326, 2023. 2, 3, 6, 7 +[40] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1632-1640, New York, NY, USA, 2019. ACM. 2 + +[41] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision., 129(4):1013-1037, 2021. 2 +[42] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 6 \ No newline at end of file diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/images.zip b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..4d5357da764b55641e95caefe9f6eec57570b23f --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:721ed3dd59e3029c7473a660f3cc8a971ccf5c451eb11cc09baba4fb922655b0 +size 733895 diff --git a/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/layout.json b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..9646ad78a28156fe066751b527f1d57edd2c6d41 --- /dev/null +++ b/2025/Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement/layout.json @@ -0,0 +1,8783 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 123, + 103, + 488, + 139 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 103, + 488, + 139 + ], + "spans": [ + { + "bbox": [ + 123, + 103, + 488, + 139 + ], + "type": "text", + "content": "Exploring View Consistency for Scene-Adaptive Low-Light Light Field Image Enhancement" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 93, + 161, + 516, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 161, + 516, + 204 + ], + "spans": [ + { + "bbox": [ + 93, + 161, + 516, + 204 + ], + "type": "text", + "content": "Shuo Zhang, Chen Gao, Youfang Lin* \nBeijing Key Laboratory of Traffic Data Mining and Embodied Intelligence, School of Computer Science& Technology, Beijing Jiaotong University, Beijing, China." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 198, + 205, + 411, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 198, + 205, + 411, + 217 + ], + "spans": [ + { + "bbox": [ + 198, + 205, + 411, + 217 + ], + "type": "text", + "content": "{zhangshuo, gaochen, yflin}@bjtu.edu.cn" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 258 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 270, + 297, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 270, + 297, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 270, + 297, + 582 + ], + "type": "text", + "content": "Light Field (LF) images captured under low illumination conditions typically exhibit low quality. Recent learning-based methods for low-light LF images enhancement are generally tailored to specific illumination inputs, limiting their performance in real-world scenes. Moreover, how to maintain the inherent view-consistency in the enhanced images also remains a difficult problem. In this paper, we propose to explore the view consistency for scene-adaptive low-light LF enhancement. We first analyze the view consistency for LF illumination maps and design a self-supervised view-consistent loss to keep the consistency between the illumination maps of different views in LFs. To enhance the model's perception of illumination, we combine both global and local information to estimate the illumination map, which is easily plugged into other models. Subsequently, we use the illumination maps to light up the low-light LF images and restore the corruption to produce the final enhanced image. Experiments demonstrate that our View-Consistent Network (VCNet) outperforms state-of-the-art methods on real-world low-light LF datasets in both fixed lighting conditions and dynamic lighting conditions. Our proposed illumination adjustment is also demonstrated that can comprehensively improve the performance of existing methods in terms of both image quality and view consistency. Code is available at: https://github.com/shuozh/VCNet" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "spans": [ + { + "bbox": [ + 55, + 604, + 135, + 616 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 623, + 296, + 696 + ], + "type": "text", + "content": "In contrast to conventional cameras, Light Field (LF) cameras have the capability to simultaneously capture both directional and intensity of all light rays in a single shot. Because of the structural information inherent in LF images, they are well-suited for various applications, including digital refocus [2], depth estimation [8], and 3D reconstruc" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 356, + 243, + 513, + 352 + ], + "blocks": [ + { + "bbox": [ + 356, + 243, + 513, + 352 + ], + "lines": [ + { + "bbox": [ + 356, + 243, + 513, + 352 + ], + "spans": [ + { + "bbox": [ + 356, + 243, + 513, + 352 + ], + "type": "image", + "image_path": "91e0ebca1dd1fc459bb03b5f1845d8ce6416d152f9409e24f7445440f00dfe8b.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 359, + 555, + 427 + ], + "lines": [ + { + "bbox": [ + 313, + 359, + 555, + 427 + ], + "spans": [ + { + "bbox": [ + 313, + 359, + 555, + 427 + ], + "type": "text", + "content": "Figure 1. Performance comparison in dynamic illumination tasks. The size of the points represents the run time to process one LF. Our VCNet achieves the best results in terms of detail restoration with higher PSNR and views consistency preservation with higher AP. By adding our proposed illumination adjustment module, labeled as “+E”, the performances of all methods are improved." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 444, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 555, + 480 + ], + "type": "text", + "content": "tion [25]. Nevertheless, the low-quality LF images severely compromised the performance of these applications, especially for images captured in low-light environments." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 312, + 483, + 556, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 483, + 556, + 698 + ], + "spans": [ + { + "bbox": [ + 312, + 483, + 556, + 698 + ], + "type": "text", + "content": "In recent years, several methods are developed for enhancing low-light LF images. Some methods [15, 28, 37] process stacked views of LF images to implicitly extract disparity information between views to take redundant information for reconstructing LF. Other methods employ advanced spatial-angular separation convolution [35, 37, 38] to extract redundant information between views for recovery. Although these methods are effective in handling LF data, their performance is limited since they do not take the view consistency into consideration, breaking the view consistency of the recovered image. Moreover, these methods learn a low-light image to normal-light image mapping to restore the low-light LF images. For different levels of illumination degradation, they require separate learning processes to train different models for recovery. Since the degradation levels are usually unknown for real-world images, it is difficult for these models to select appropriate parameters for restoration." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 326, + 701, + 555, + 714 + ], + "type": "text", + "content": "To deal with dynamically degraded inputs, a simple yet" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 702, + 205, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 702, + 205, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 702, + 205, + 713 + ], + "type": "text", + "content": "*Corresponding author: yflin@bjtu.edu.cn" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "7340" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 346 + ], + "type": "text", + "content": "principled solution relying on the Retinex theory is proposed, which has been widely used for single-image [22, 23, 30]. These Retinex-based methods usually divide the light-up process into two stages: low contrast adjustment and corruptions (e.g., noise, artifact, color distortion, etc.) recovery. By estimating the illumination map of the scene for low contrast adjustment, the model is able to deal with dynamic illumination degradation. For LF images, the most straightforward way is to learn a brute-force mapping function for each view to apply one-to-one illumination adjustment. However, processing each LF view individually tends to recover each view with quite different illumination maps, resulting in varying degrees of color variation across different views. Another possible way is to estimate only one identical illumination adjustment gain for each view [39] or use a learnable parameter to compute the full illumination map [18], which is also improper since there exist disparities between views in one LF. This severely breaks the structural information inherent in LF images and limits the application of the following tasks. Therefore, how to keep the view-consistency in the estimated illumination maps in one LF becomes an important problem for scene-adaptive low-light LF image enhancement." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 348, + 295, + 622 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 295, + 622 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 295, + 622 + ], + "type": "text", + "content": "In this paper, we explore the view-consistency problem in scene-adaptive low-light LF image enhancement. We first analyze the disparity consistency between the illumination adjustment maps of different views. Based on the photometric consistency theory, we propose a self-supervised view-consistent illumination loss to keep the consistency between the estimated illumination maps. These illumination maps are estimated by combining the global brightness of the scene and the local brightness of each view, which are used to light up the low-light LF images. Subsequently, we perform corruption recovery on the light-up LF images to further suppress noise, artifacts, under-/overexposure, and color distortion. Comprehensive experiments are conduct in dynamic and fixed illumination environments. The related results are shown in Fig. 1, in which the proposed View-Consistent Network (VCNet) outperforms others over 7 db. More importantly, we also add our proposed illumination adjustment module to other models, labeled as “+E”. The experiments show that our proposed illumination adjustment module can be embedded into other methods and greatly improve their performance in terms of both image quality and view consistency preservation without increasing a large number of parameters." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "spans": [ + { + "bbox": [ + 55, + 633, + 142, + 645 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 294, + 713 + ], + "type": "text", + "content": "Low-Light Image Enhancement. The low-light image enhancement methods are mainly divided into two categories. The first category directly learns a brute-force mapping function from the low-light image to its normal-light counterpart. An exemplary illustration is the SID" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 276 + ], + "type": "text", + "content": "approach [3], which utilizes an encoder-decoder architecture to effectively restore raw images captured under extremely low-light conditions. DID [19] introduced an innovative residual learning network explicitly crafted for the denoising of low-light images, surpassing the capabilities of SID. Guo et al. [9] designed a Zero-DCE network to employ dynamic range adjustment through pixel-wise and high-order curve estimations. However, these methods lack interpretability and theoretically proven properties. The methods in the second category are designed based on Retinex theory for restoration. Wei et al. [34] and follow-up works [40, 41] combine the Retinex decomposition with a multi-stage training pipeline. In the first training stage, these methods learn a network to decompose the images. Then in the second stage, they employ a different network to denoise the reflectance and adjust the illumination, respectively." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 279, + 553, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 279, + 553, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 279, + 553, + 445 + ], + "type": "text", + "content": "Low-Light Video Enhancement. Different from low-light image enhancement, methods that enhance videos recorded under low-light conditions extract valuable complementary details from neighboring frames for better restoration. Wang et al. [26] delved into the field of physical modeling, utilizing an LSTM-based network to improve the quality of low-light videos. Based on Retinex theory, Wang et al. [24] introduced a comprehensive methodology to accomplish simultaneous enhancement of low-light conditions and noise reduction in video sequences. Even though these methods achieved remarkable results in enhancing two-dimensional low-light images and videos, it is still difficult for them to handle the four-dimensional LF dataset with regular sampling in the angular domain." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 553, + 713 + ], + "type": "text", + "content": "Low-Light LF Image Enhancement. For LF images, some methods directly perform the restoration of low-light to normal illumination images. Lamba et al. [15] made substantial contributions in this regard by introducing a lowlight LF images dataset and a two-stage L3Fnet architecture. This architecture consists of a global representation block, responsible for encoding the intricate LF geometry, and a view reconstruction block that concentrates on restoring each distinct view. To address the challenge of enhancing dim LF images, they also introduced a three-stage network [14] that effectively utilizes global, local, and view-specific features inherent to low-light LF images. Wang et al. [28] introduced a coarse-to-fine restoration network to progressively restore the low-light LF input. In their network, three types of view stacks are fed into three streams to explore the redundant information of LFs. Recently, Wang et al. [29] proposed a novel feature extraction approach that parallelly processes features across multiple spatial scales. Lyu et al. [39] proposed a multi-stage architecture that mimics the optimization process of solving an inverse imaging problem, in which the illumination maps are estimated and employed in the unfolding process to produce" + } + ] + } + ], + "index": 6 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7341" + } + ] + } + ], + "index": 7 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 68, + 532, + 140 + ], + "blocks": [ + { + "bbox": [ + 82, + 68, + 532, + 140 + ], + "lines": [ + { + "bbox": [ + 82, + 68, + 532, + 140 + ], + "spans": [ + { + "bbox": [ + 82, + 68, + 532, + 140 + ], + "type": "image", + "image_path": "ae23a6b98edde3010d7d9a93ffe200085fe5003214ac1676168f34096dabf606.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "lines": [ + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "spans": [ + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "text", + "content": "Figure 2. Illustration of our self-supervised view-consistent loss. The reference point " + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "text", + "content": " has the disparity " + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "inline_equation", + "content": "d(\\mathbf{x}) = 1" + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "text", + "content": ". Compared with other points (labeled as red and purple), the corresponding point (labeled as orange) has the minimum variance. We propose to directly use the minimum variance as our objective function " + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 54, + 147, + 555, + 181 + ], + "type": "text", + "content": " to minimize." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 193, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 193, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 56, + 193, + 296, + 408 + ], + "type": "text", + "content": "the new enhanced result. There are also approaches to divide complex tasks into several intermediate tasks. Inspired by Retinex theory, Zhang et al. [37] proposed a learning-based decomposition-enhancement methodology that systematically dissects the intricate enhancement task into illumination adjustment and reflection denoising. Zhang et al. [39] separate the complex task into several intermediate tasks and use multiple heads to restore low-light images. They also formulate a synthesis pipeline to enrich the available LF data pairs, in which the ground truth illumination maps are generated and used to guide illumination adjustment. Although they consider the dynamically degraded inputs, their model needs the ground truth illumination value to give full supervision in the training process, which is unavailable for real-world images. Moreover, these methods do not consider the consistency constraints of the illumination maps, so that the view consistency of the recovered image is easily broken." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 418, + 201, + 431 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 418, + 201, + 431 + ], + "spans": [ + { + "bbox": [ + 55, + 418, + 201, + 431 + ], + "type": "text", + "content": "3. View-Consistent Network" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": "According to the two-plane model [16], a LF can be formulated as a 4D tensor " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{L}(u,v,x,y)\\in \\mathbb{R}^{U\\times V\\times H\\times W}" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{u} = (u,v)" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": " denote the angular coordinates, and " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathbf{x} = (x,y)" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": " represent the spatial coordinates. Commonly, the LF image is organized as a " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "U\\times V" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": " array of views that captured the same scene in one shot. One specific view at angular coordinate " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "(u,v)" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": " can be denoted as " + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathbf{u}}\\in \\mathbb{R}^{H\\times W}" + }, + { + "bbox": [ + 55, + 438, + 296, + 523 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 529, + 205, + 542 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 529, + 205, + 542 + ], + "spans": [ + { + "bbox": [ + 55, + 529, + 205, + 542 + ], + "type": "text", + "content": "3.1. View Consistency Analyses" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 546, + 217, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 217, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 217, + 558 + ], + "type": "text", + "content": "3.1.1. Retinex Theory for Single View" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 561, + 296, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 561, + 296, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 561, + 296, + 597 + ], + "type": "text", + "content": "Drawing upon the Retinex theory, each low-light view in LF can be mathematically expressed as the product of its reflectance and illumination components as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 146, + 601, + 295, + 611 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 601, + 295, + 611 + ], + "spans": [ + { + "bbox": [ + 146, + 601, + 295, + 611 + ], + "type": "interline_equation", + "content": "I _ {\\mathbf {u}} = L _ {\\mathbf {u}} \\odot R _ {\\mathbf {u}}, \\tag {1}", + "image_path": "7e2e2b5589613cc59115859eeeba618e4fba3cc00162485991dd0c28fd018222.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "spans": [ + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "L_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": " represents the illumination components, reflectance components with angular coordinates " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\mathbf{u}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": ", respectively. " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\odot" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": " denotes the element-wise multiplication. Similar to [7, 11, 22], we regard " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "R_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": " as a well-exposed LF view. To light up " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "I_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": ", we element-wisely multiply the two sides of Eq. 1 by an illumination adjustment map " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": " such that " + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u}} \\odot L_{\\mathbf{u}} = 1" + }, + { + "bbox": [ + 55, + 616, + 296, + 699 + ], + "type": "text", + "content": " as follow:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 146, + 702, + 295, + 714 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 702, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 146, + 702, + 295, + 714 + ], + "type": "interline_equation", + "content": "R _ {\\mathbf {u}} = I _ {\\mathbf {u}} \\odot \\bar {L} _ {\\mathbf {u}}. \\tag {2}", + "image_path": "c615c38d8b1686f49d1f3bab211bea7996a2765ab74a65811838a1a82ac0280b.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 193, + 553, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 193, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 313, + 193, + 553, + 217 + ], + "type": "text", + "content": "Therefore, for LF images, we need to estimate the illumination adjustment map " + }, + { + "bbox": [ + 313, + 193, + 553, + 217 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u}}" + }, + { + "bbox": [ + 313, + 193, + 553, + 217 + ], + "type": "text", + "content": " for all views in one LF." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 223, + 511, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 223, + 511, + 235 + ], + "spans": [ + { + "bbox": [ + 313, + 223, + 511, + 235 + ], + "type": "text", + "content": "3.1.2. View Consistency in Illumination Maps" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "spans": [ + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "text", + "content": "Since views in one LF are taken from different angulars of the same scene, in which the depth of the scene is fixed, the views are supposed to satisfy the disparity constraint. Suppose that the parameter of the camera is known, the disparity of two views " + }, + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "inline_equation", + "content": "I_{\\mathbf{u_j}}" + }, + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "inline_equation", + "content": "I_{\\mathbf{u_i}}" + }, + { + "bbox": [ + 313, + 237, + 555, + 299 + ], + "type": "text", + "content": " is calculated as:" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 338, + 302, + 555, + 324 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 338, + 302, + 555, + 324 + ], + "spans": [ + { + "bbox": [ + 338, + 302, + 555, + 324 + ], + "type": "interline_equation", + "content": "D _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {i}}} (\\mathbf {x}) = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * \\frac {B * f}{Z (\\mathbf {x})} = \\left(\\mathbf {u} _ {\\mathbf {j}} - \\mathbf {u} _ {\\mathbf {i}}\\right) * d (\\mathbf {x}), \\tag {3}", + "image_path": "3137760db2c3bd1b4f260a339d7e6c033e6f3eb010204fc6dd781483b5948c46.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "inline_equation", + "content": "B" + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "content": " represent the baseline length and the disparity between adjacent views, respectively. " + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "inline_equation", + "content": "f" + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "content": " represents the focal length of the camera. " + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "inline_equation", + "content": "Z" + }, + { + "bbox": [ + 313, + 327, + 554, + 387 + ], + "type": "text", + "content": " represents the depth of the scene. Therefore, the two captured low-light views satisfy the disparity constraint as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 389, + 391, + 555, + 403 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 389, + 391, + 555, + 403 + ], + "spans": [ + { + "bbox": [ + 389, + 391, + 555, + 403 + ], + "type": "interline_equation", + "content": "I _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(I _ {\\mathbf {u} _ {\\mathrm {j}}}, D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right), \\tag {4}", + "image_path": "669167ecd198968183158cf525de2c92c07955ed5de165e30d8b5e298660327d.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "spans": [ + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "inline_equation", + "content": "W" + }, + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "text", + "content": " denotes the warp operations according to the disparity " + }, + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{u_j}\\rightarrow \\mathbf{u_i}}" + }, + { + "bbox": [ + 313, + 407, + 554, + 443 + ], + "type": "text", + "content": ". Similarly, the well-exposed LF views also satisfy the disparity constraint as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 387, + 446, + 554, + 459 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 387, + 446, + 554, + 459 + ], + "spans": [ + { + "bbox": [ + 387, + 446, + 554, + 459 + ], + "type": "interline_equation", + "content": "R _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(R _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {5}", + "image_path": "0c6d191110d2e69ed557b254eda07626966d62577a981d8d03fa10a1852a544f.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 554, + 486 + ], + "type": "text", + "content": "Therefore, according to Eq. 2, the illumination adjustment map should also satisfy the disparity constraint." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 388, + 489, + 554, + 502 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 388, + 489, + 554, + 502 + ], + "spans": [ + { + "bbox": [ + 388, + 489, + 554, + 502 + ], + "type": "interline_equation", + "content": "\\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}\\right). \\tag {6}", + "image_path": "3827a7309e88c7a2feacb7dc93ab95a97c89cd461b4d1de9590fb7424f4db922.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "content": "In order to constrain the consistency of the illumination adjustment maps, the direct way is to minimize the distance between the illumination map " + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u_i}}" + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "content": " and the warped illumination map " + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "inline_equation", + "content": "W(\\overline{L}_{\\mathbf{u_j}};D_{\\mathbf{u_j}\\to \\mathbf{u_i}})" + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "content": " from different angular " + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{u_j}" + }, + { + "bbox": [ + 313, + 506, + 554, + 555 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 361, + 562, + 554, + 585 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 562, + 554, + 585 + ], + "spans": [ + { + "bbox": [ + 361, + 562, + 554, + 585 + ], + "type": "interline_equation", + "content": "\\min \\sum_ {\\mathbf {u} _ {\\mathrm {j}}} \\sum_ {\\mathbf {u} _ {\\mathrm {i}}} | | W (\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {i}}}) - \\bar {L} _ {\\mathbf {u} _ {\\mathrm {i}}} | |, \\tag {7}", + "image_path": "0cf8ce9ea9fa41ecae3e22e3731ddc1c8cf97c50b801a760f50b5dfa7bc3b963.jpg" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{u_j},\\mathbf{u_i}\\in \\{(0,0),(0,1),\\dots,(U,V)\\}" + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "text", + "content": ". To simplify the computation, we warp the illumination adjustment maps of different views towards the central view " + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u_c}}" + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "text", + "content": " and construct these aligned views to a 4D array " + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "inline_equation", + "content": "\\mathbf{A}\\in \\mathbb{R}^{U\\times V\\times H\\times W}" + }, + { + "bbox": [ + 313, + 586, + 554, + 635 + ], + "type": "text", + "content": " .." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 379, + 638, + 554, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 379, + 638, + 554, + 651 + ], + "spans": [ + { + "bbox": [ + 379, + 638, + 554, + 651 + ], + "type": "interline_equation", + "content": "\\mathbf {A} \\left(\\mathbf {u} _ {\\mathbf {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathbf {j}}}; D _ {\\mathbf {u} _ {\\mathbf {j}} \\rightarrow \\mathbf {u} _ {\\mathbf {c}}}\\right), \\tag {8}", + "image_path": "21d972b84ffef73f422a4670305d7eecd7660e9b53d4dcbbd0338b92b2ab4bb6.jpg" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "inline_equation", + "content": "\\mathbf{u}_{\\mathbf{c}} = \\left(\\frac{U}{2},\\frac{V}{2}\\right)" + }, + { + "bbox": [ + 313, + 654, + 553, + 691 + ], + "type": "text", + "content": ". Then, we instead calculate the variance of the warped illumination adjustment maps to assess their disparity consistency as:" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 392, + 696, + 554, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 392, + 696, + 554, + 716 + ], + "spans": [ + { + "bbox": [ + 392, + 696, + 554, + 716 + ], + "type": "interline_equation", + "content": "\\min \\sum_ {\\mathbf {x}} \\operatorname {V a r} (\\mathbf {A} (:, \\mathbf {x})), \\tag {9}", + "image_path": "b3c7c3a797c2814ae228262fe82a12c1118ad2107846216b8260a90632184a95.jpg" + } + ] + } + ], + "index": 26 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7342" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 70, + 294, + 167 + ], + "blocks": [ + { + "bbox": [ + 59, + 70, + 294, + 167 + ], + "lines": [ + { + "bbox": [ + 59, + 70, + 294, + 167 + ], + "spans": [ + { + "bbox": [ + 59, + 70, + 294, + 167 + ], + "type": "image", + "image_path": "4b6f7d130f8f003ae6ffa931942e1e2661c72893d72447eea5d374d25cf36be6.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 175, + 296, + 231 + ], + "lines": [ + { + "bbox": [ + 55, + 175, + 296, + 231 + ], + "spans": [ + { + "bbox": [ + 55, + 175, + 296, + 231 + ], + "type": "text", + "content": "Figure 3. The proposed VCNet network consists of an illumination estimator VCAIA and a corruption restorer. In VCAIA, the illumination map is estimated and multiplied to the original input to light up the illumination. VCRM further recovers the details of the light-up image and outputs the final results." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 251, + 295, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 251, + 295, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 251, + 295, + 275 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 251, + 295, + 275 + ], + "type": "inline_equation", + "content": "Var(\\mathbf{A}(:,\\mathbf{x}))" + }, + { + "bbox": [ + 55, + 251, + 295, + 275 + ], + "type": "text", + "content": " calculates the variance for each pixel along the angular dimension." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "content": "However, in real scenes, both the ground truth illumination adjustment map " + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "inline_equation", + "content": "\\overline{L}_{\\mathbf{u}}" + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "content": " and the corresponding disparity " + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 275, + 296, + 348 + ], + "type": "text", + "content": " is unavailable. Moreover, the disparity is also difficult to estimate using low-light images with much noise [12]. Therefore, the disparity constraint of the illumination adjustment maps as in Eq. 9 is difficult to implement." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 354, + 258, + 368 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 354, + 258, + 368 + ], + "spans": [ + { + "bbox": [ + 55, + 354, + 258, + 368 + ], + "type": "text", + "content": "3.2. Self-Supervised View-Consistent Loss" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 296, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 296, + 468 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 296, + 468 + ], + "type": "text", + "content": "In this section, we relax the disparity constraint and design a self-supervised view-consistent illumination loss to ensure the view consistency of illumination adjustment maps. Instead of constraining each point according to its true disparity as in Eq. 9, we only constrain each point according to one specific disparity value, no matter whether it is the true disparity or not. Specifically, we first set " + }, + { + "bbox": [ + 55, + 372, + 296, + 468 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 372, + 296, + 468 + ], + "type": "text", + "content": " discrete candidate disparity values:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 115, + 476, + 294, + 490 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 476, + 294, + 490 + ], + "spans": [ + { + "bbox": [ + 115, + 476, + 294, + 490 + ], + "type": "interline_equation", + "content": "d ^ {n} = d _ {\\min } + n * \\Delta d, n \\leq N \\tag {10}", + "image_path": "174d6c1fdf4ec31af51cbdab21a730c4f8134251dbc88f98401ffd61d3ce617d.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\Delta d = (d_{max} - d_{min}) / N" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": " denotes the granularity of " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "d_{min}" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "d_{max}" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": " denote the minimum and maximum disparity of the scene. Subsequently, we warp each view to the center view according to all the possible disparity values " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "d^n" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": ", and the related 4D array is denoted as " + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^n \\in \\mathbb{R}^{U \\times V \\times H \\times W}" + }, + { + "bbox": [ + 55, + 498, + 296, + 558 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 115, + 568, + 295, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 115, + 568, + 295, + 583 + ], + "spans": [ + { + "bbox": [ + 115, + 568, + 295, + 583 + ], + "type": "interline_equation", + "content": "\\mathbf {A} ^ {n} \\left(\\mathbf {u} _ {\\mathrm {j}}, :\\right) = W \\left(\\bar {L} _ {\\mathbf {u} _ {\\mathrm {j}}}; D _ {\\mathbf {u} _ {\\mathrm {j}} \\rightarrow \\mathbf {u} _ {\\mathrm {c}}} ^ {n}\\right), \\tag {11}", + "image_path": "f7e789d63b9de8317ee938bea69b035b99dc27176828debd61ec55804f1aae37.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 585, + 213, + 597 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 213, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 213, + 597 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 585, + 213, + 597 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{u}_i - > \\mathbf{u_c}}^n (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^n" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": "We then calculate the variance of the warped view for each point " + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "inline_equation", + "content": "\\mathbf{x}" + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": " and for each possible disparity " + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "inline_equation", + "content": "Var(\\mathbf{A}^n (:, \\mathbf{x}))" + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": ", constructing the commonly known cost volume as in the disparity estimation methods [4, 13]. According to the photometric consistency, the variance under the true disparity is supposed to have the minimum value compared with other disparities. Therefore, we first find the specific disparity " + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "inline_equation", + "content": "d^{*}" + }, + { + "bbox": [ + 55, + 597, + 296, + 693 + ], + "type": "text", + "content": " for each point that has the minimum variance:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 108, + 699, + 295, + 716 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 108, + 699, + 295, + 716 + ], + "spans": [ + { + "bbox": [ + 108, + 699, + 295, + 716 + ], + "type": "interline_equation", + "content": "d ^ {*} (\\mathbf {x}) = \\arg \\min _ {d ^ {n}} (V a r \\left(\\mathbf {A} ^ {n} (:, \\mathbf {x}))\\right)). \\tag {12}", + "image_path": "dffd31cbec663885720a571f94a4f0639b3bdad3dd52f5df7c218860c81dd6d8.jpg" + } + ] + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 351, + 70, + 519, + 145 + ], + "blocks": [ + { + "bbox": [ + 351, + 70, + 519, + 145 + ], + "lines": [ + { + "bbox": [ + 351, + 70, + 519, + 145 + ], + "spans": [ + { + "bbox": [ + 351, + 70, + 519, + 145 + ], + "type": "image", + "image_path": "4029e28c722ecfd037184699b092a95911c66b7f899962fb454d4021cc4d4286.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 153, + 554, + 176 + ], + "lines": [ + { + "bbox": [ + 313, + 153, + 554, + 176 + ], + "spans": [ + { + "bbox": [ + 313, + 153, + 554, + 176 + ], + "type": "text", + "content": "Figure 4. The estimated illumination EPI w/ and w/o " + }, + { + "bbox": [ + 313, + 153, + 554, + 176 + ], + "type": "inline_equation", + "content": "loss_{vc}" + }, + { + "bbox": [ + 313, + 153, + 554, + 176 + ], + "type": "text", + "content": ". The former keeps the view consistent as the true EPI with normal light." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "spans": [ + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "text", + "content": "We then minimize the variance related to the " + }, + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "inline_equation", + "content": "d^{*}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "text", + "content": " to make sure that all views are aligned in the disparity " + }, + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "inline_equation", + "content": "d^{*}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 198, + 554, + 221 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 383, + 229, + 554, + 252 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 383, + 229, + 554, + 252 + ], + "spans": [ + { + "bbox": [ + 383, + 229, + 554, + 252 + ], + "type": "interline_equation", + "content": "\\ell_ {v c} = \\sum_ {\\mathbf {x}} \\operatorname {V a r} \\left(\\mathbf {A} ^ {*} (:, \\mathbf {x})\\right), \\tag {13}", + "image_path": "2556814df7a5fb87cfc8efa0e9b66dec524b9e0c624f23ffd12b09a9e993c455.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "inline_equation", + "content": "\\mathbf{A}^*" + }, + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "text", + "content": " represents the 4D array with aligned views according " + }, + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "inline_equation", + "content": "D_{\\mathbf{u_j - > u_c}}^* (\\mathbf{x}) = (\\mathbf{u_j} - \\mathbf{u_c})*d^* (\\mathbf{x})" + }, + { + "bbox": [ + 313, + 260, + 554, + 344 + ], + "type": "text", + "content": " as in Eq. 11. An example is shown in Fig. 2, where the minimum variance is set as our objective function. In such a way, each pixel is assigned a disparity value that has the minimum variance to make the illumination adjustment value of each pixel as consistent as possible with other views." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 352, + 499, + 365 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 352, + 499, + 365 + ], + "spans": [ + { + "bbox": [ + 313, + 352, + 499, + 365 + ], + "type": "text", + "content": "3.3. Adaptive Illumination Adjustment" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "text", + "content": "Beginning with the input low-light LF image " + }, + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{in}" + }, + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "text", + "content": ", we design a simple illumination estimator, View Consistent Adaptive Illumination Adjustment (VCAIA), " + }, + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 369, + 554, + 464 + ], + "type": "text", + "content": " for LFs to estimate the illumination adjustment map. Specifically, as in Fig. 3, we first estimate the global illumination adjustment matrix of the scene by integrating information from all views. The local brightness refinement is then developed to adjust the matrix according to the details of each view." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "text", + "content": "In order to integrate the abundant information from all views and further reduce the effect of noise, we roughly calculate the average of all views across all color channels in LF image as " + }, + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "inline_equation", + "content": "I_w \\in \\mathbb{R}^{H \\times W \\times 1}" + }, + { + "bbox": [ + 313, + 465, + 554, + 525 + ], + "type": "text", + "content": ". Then following [1], we calculate the global illumination adjustment matrix as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 332, + 533, + 553, + 558 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 332, + 533, + 553, + 558 + ], + "spans": [ + { + "bbox": [ + 332, + 533, + 553, + 558 + ], + "type": "interline_equation", + "content": "\\overline {{I ^ {l}}} _ {\\text {g l o b a l}} (\\mathbf {x}) = \\frac {\\log \\left(I _ {w} (\\mathbf {x}) / \\mu + 1\\right)}{\\log \\left(\\max \\left(I _ {w} (\\mathbf {x})\\right) / \\mu + 1\\right) \\cdot I _ {w} (\\mathbf {x})}, \\tag {14}", + "image_path": "efda896f8934bb46fb372e63442357e293eecd9b49317fa796b22a68315d9b16.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "text", + "content": "Subsequently, we replicate " + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\overline{I}^{l}_{global}" + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "text", + "content": " to all views and obtain the global illumination adjustment matrix " + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}}^{l}_{global}" + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "text", + "content": ". To further refine illumination estimation for each view, we employ a lightweight learnable network. Specifically, we first stack all views of the input low-light LF image and use the cascaded 3D spatial convolution layers to obtain the local brightness adjustment matrix " + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}}^{l}_{local}" + }, + { + "bbox": [ + 313, + 567, + 554, + 653 + ], + "type": "text", + "content": " of each view." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": "The local and global illumination adjustment matrix is combined to derive the final adjustment matrix and is then multiplied to " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{in}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " to obtain the adjusted high-illumination image: " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}} = (\\overline{\\mathcal{L}^t}_{global} + \\overline{\\mathcal{L}^t}_{local})" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": ". By minimizing the proposed self-supervised loss " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " on the estimated " + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}}" + }, + { + "bbox": [ + 313, + 654, + 554, + 714 + ], + "type": "text", + "content": " according" + } + ] + } + ], + "index": 22 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7343" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 59, + 91, + 550, + 182 + ], + "blocks": [ + { + "bbox": [ + 88, + 70, + 521, + 82 + ], + "lines": [ + { + "bbox": [ + 88, + 70, + 521, + 82 + ], + "spans": [ + { + "bbox": [ + 88, + 70, + 521, + 82 + ], + "type": "text", + "content": "Table 1. Quantitative Comparisons of Low Light LF Image Enhancement Methods in L3F-Dynamic Task w/ w/o VCAIA." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 91, + 550, + 182 + ], + "lines": [ + { + "bbox": [ + 59, + 91, + 550, + 182 + ], + "spans": [ + { + "bbox": [ + 59, + 91, + 550, + 182 + ], + "type": "table", + "html": "
MethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓PSNR↑SSIM↑LPIPS↓
LFFT-DCU23.170.80030.095819.600.72010.170219.820.67180.2074
DRLF19.100.69470.188817.420.66000.240217.040.63220.2608
DRLF+E22.19 (+3.09)0.7636 (+0.0689)0.1124 (-0.0764)21.38 (+3.96)0.6976 (+0.0376)0.1676 (-0.0726)20.29 (+3.25)0.6326 (+0.0004)0.2265 (-0.0343)
MSP21.370.75600.133018.640.68980.207917.620.65430.2407
MSP+E22.80 (+1.43)0.7936 (+0.0376)0.1261 (-0.0069)21.81 (+3.17)0.7436 (+0.0538)0.1673 (-0.0406)21.10 (+3.48)0.6916 (+0.0373)0.2128 (-0.0279)
VCNet-E22.540.78350.125120.830.71030.201519.920.64500.2666
VCNet24.58 (+2.04)0.8220 (+0.0385)0.0897 (-0.0354)24.06 (+3.23)0.7652 (+0.0549)0.1292 (-0.0723)23.09 (+3.17)0.7032 (-0.0582)0.1981 (-0.0685)
", + "image_path": "71514050662cd5af1c0d94bcc604859f5af552ec3700d411a0a7a26e1aa21a6b.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "content": "to Eq. 13, we constrain the calculated illumination adjustment maps to satisfy the view consistency even without the true disparities. Since no supervised images are needed, the illumination adjustment can be applied to various low-light LF enhancement networks. We show one estimated example in Fig. 4, where the Epipolar Plane Images (EPIs) are calculated by fixing " + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "inline_equation", + "content": "v" + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "content": " in angular domain and " + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "inline_equation", + "content": "w" + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "content": " in the spatial domain. Without the proposed " + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "content": ", the lines in the EPIs of the estimated illumination maps become ambiguous and are inconsistent with the ground truth EPIs of the normallight LFs. In contrast, when we use " + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 55, + 196, + 294, + 352 + ], + "type": "text", + "content": " to constrain the view consistency, the lines in estimated EPIs of illumination maps retain the direction as the ground truth EPIs." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 367, + 160, + 380 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 367, + 160, + 380 + ], + "spans": [ + { + "bbox": [ + 55, + 367, + 160, + 380 + ], + "type": "text", + "content": "3.4. Recovery Module" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 387, + 295, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 387, + 295, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 387, + 295, + 495 + ], + "type": "text", + "content": "After illumination adjustment, the image " + }, + { + "bbox": [ + 55, + 387, + 295, + 495 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^a = \\overline{\\mathcal{L}}\\odot \\mathcal{L}^{in}" + }, + { + "bbox": [ + 55, + 387, + 295, + 495 + ], + "type": "text", + "content": " with a standard illumination level is obtained. However, the light-up process also amplifies the noise and artifacts, causing under-/over-exposure and color distortion. Therefore, similar to other retinae-based methods, a recovery module is needed to suppress noise and restore scene details. Some existing networks, such as the LF denoising model DRLF and the LF low-light enhancement model MSP, can be used as the recovery module." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": "We also design a simple View Consistent Recovery Module (VCRM) to further import the inherent disparity constraint into the recovery process to enhance the recovery effect. VCRM adopts a multi-scale architecture. In each scale, the view consistent feature aggregation unit is implemented to explore the redundant information of the views to recover details. In each unit, we employ deformable convolution layers on EPIs to further integrate information between different views. Since the slope of lines in EPIs corresponds to disparity, we introduce the disparity constraint to help the model identify correct and consistent complementary information. Specifically, we use uniform offsets in both V-H and U-W spaces, so that the angular information can be efficiently incorporated and maintain epipolar consistency. Finally, we concatenate multi-scale features and use a decoder with a similar structure to the encoder to aggregate them and obtain the refined result " + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 55, + 498, + 295, + 714 + ], + "type": "text", + "content": ". More detailed design is provided in the supplementary details." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 195, + 447, + 209 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 195, + 447, + 209 + ], + "spans": [ + { + "bbox": [ + 313, + 195, + 447, + 209 + ], + "type": "text", + "content": "3.5. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "spans": [ + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": "As in Fig. 3, the overview architecture of our VCNet primarily employs two VCAIA and VCRM. We train our VCNet by minimizing a hybrid loss: " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\ell_{total} = \\lambda_1\\ell_{vc} + \\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\ell_{rec}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " is the reconstruction loss, which calculates the L1 loss between the ground truth of well-posed image " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{gt}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " and the recovered results " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " of VCNet. The SSIM loss [33] " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\ell_{ssim}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " measures the structural similarity between " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{gt}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^{out}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": ". " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " is the proposed self-supervised illumination loss in Sec.3, which is used to keep the view consistent of the estimated illumination maps " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}}^l" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": ". The three hyperparameters " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_1" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_2" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "inline_equation", + "content": "\\lambda_3" + }, + { + "bbox": [ + 313, + 213, + 553, + 344 + ], + "type": "text", + "content": " are empirically set to 0.5, 1, 0.5." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "spans": [ + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "text", + "content": "Our model is implemented using the PyTorch framework and trained on a NVIDIA RTX A4000 GPU. The model is optimized using Adam optimization with " + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\beta_{1} = 0.9" + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "inline_equation", + "content": "\\beta_{2} = 0.999" + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "text", + "content": ". The batch size is set to 1 and the learning rate is initially set to 0.0001. Data augmentation techniques such as random cropping, vertical and horizontal flipping are used. During training, the views are randomly cropped in patches of size " + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "inline_equation", + "content": "128 \\times 128" + }, + { + "bbox": [ + 313, + 345, + 554, + 453 + ], + "type": "text", + "content": ". During testing, we use the original resolution LF data without applying data augmentation." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 394, + 475 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 480, + 372, + 491 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 480, + 372, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 480, + 372, + 491 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "spans": [ + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "text", + "content": "We evaluate the proposed method using the publicly accessible Low-Light-Light-Field (L3F) Dataset and L3F-Wild Dataset from [15]. The L3F dataset comprises 27 distinct scenes, each of which was captured under three low-light conditions, labeled L3F-20, 50, 100, and one with normal lighting. The L3F-Wild is captured in almost 0 lux conditions and captured with " + }, + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "inline_equation", + "content": "\\frac{1}{4} -\\frac{1}{15}" + }, + { + "bbox": [ + 313, + 498, + 553, + 582 + ], + "type": "text", + "content": " second exposure time." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 582, + 553, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 582, + 553, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 582, + 553, + 676 + ], + "type": "text", + "content": "L3F-Fixed Task: Previous LF-based low-light enhancement methods trained their models under fixed illumination conditions. Specifically, following [15], 18 scenes under the same illumination condition are selected for training, while the remaining 9 scenes are reserved for testing. In the fixed illumination conditions, 3 models are specifically trained for L3F-20, 50, 100 in different illumination conditions. We denote it as the L3F-Fixed task." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 553, + 713 + ], + "type": "text", + "content": "L3F-Dynamic Task: To further assess the model's capability to adapt to various illumination conditions, we establish a dataset with dynamic illumination as L3F-Dynamic." + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7344" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 67, + 70, + 541, + 312 + ], + "blocks": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "lines": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "spans": [ + { + "bbox": [ + 67, + 70, + 541, + 312 + ], + "type": "image", + "image_path": "6e2e38008af38050d261af8f589c35515d89bd0039a5ecd1d99ba4aee0b88eb7.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "lines": [ + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "spans": [ + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "text", + "content": "Figure 5. Visual comparison on " + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 54, + 320, + 555, + 354 + ], + "type": "text", + "content": "-Dynamic task, where the central view and EPIs of the recovery LFs are shown. Other algorithms either generate over-exposed and noisy images or introduce black spot corruptions and unnatural artifacts. Our VCNet restores well-exposed structural contents and textures. By adding our VCAIA module, the recovered results of all models are closer to ground truth." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 366, + 295, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 366, + 295, + 473 + ], + "spans": [ + { + "bbox": [ + 54, + 366, + 295, + 473 + ], + "type": "text", + "content": "Specifically, we combine images from the L3F-20, 50, 100 datasets. We use all training sets from the three different illumination conditions (including " + }, + { + "bbox": [ + 54, + 366, + 295, + 473 + ], + "type": "inline_equation", + "content": "18 \\times 3" + }, + { + "bbox": [ + 54, + 366, + 295, + 473 + ], + "type": "text", + "content": " scenes) to train a single model. The single trained model is supposed to handle dynamic illumination conditions. We use the same testing dataset as L3F-Fixed to evaluate its performance. Compared to L3F-Fixed task, L3F-Dynamic task is more difficult, which needs to train one single model that is able to handle all different illumination conditions." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 480, + 266, + 493 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 480, + 266, + 493 + ], + "spans": [ + { + "bbox": [ + 55, + 480, + 266, + 493 + ], + "type": "text", + "content": "4.2. Compare with State-of-the-Art Methods" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "spans": [ + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": "We compare our VCNet and several SOTA methods, including 3 recent published single-frame retinae-based methods (RFormer [22], Pydiff [42] and DarkIR [6]) and 9 LF-based methods (L3F-Net [15], FER-Net [14], DRLF [10], MSP-3 [27], IINet [17], DistgSSR [31], PMSNet [29], LRT [39], LFLL-DCU [18]). The illumination estimator " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": " is applied to the LF denoising model DRLF and low-light enhancement model MSP, denoted as DRLF+ " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": ", MSP+ " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": ". Specifically, we embed the illumination estimator " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": " into the beginning of these two models and add the " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": " in their training strategy. Our VCNet-" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": " is trained by deleting " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": " module and only minimizing " + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "inline_equation", + "content": "\\lambda_2\\ell_{rec} + \\lambda_3\\ell_{ssim}" + }, + { + "bbox": [ + 56, + 498, + 295, + 676 + ], + "type": "text", + "content": ". The reference-based metric PSNR, SSIM [32] and LPIPS [36] and nonreference metric NIQE [21] and BRISQUE [20] are chosen as the evaluation metrics." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "Dynamic Illumination Task. We first assess our model's ability to adapt to dynamic illumination. Specifically, we train each method using the L3F-Dynamic training dataset" + } + ] + } + ], + "index": 5 + }, + { + "type": "table", + "bbox": [ + 320, + 395, + 547, + 459 + ], + "blocks": [ + { + "bbox": [ + 313, + 365, + 554, + 387 + ], + "lines": [ + { + "bbox": [ + 313, + 365, + 554, + 387 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 554, + 387 + ], + "type": "text", + "content": "Table 2. Quantitative Comparisons with Single-Frame-based Low Light Enhancement Methods on the L3F-dynamic Task." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 320, + 395, + 547, + 459 + ], + "lines": [ + { + "bbox": [ + 320, + 395, + 547, + 459 + ], + "spans": [ + { + "bbox": [ + 320, + 395, + 547, + 459 + ], + "type": "table", + "html": "
MethodL3F-20L3F-50L3F-100
PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑PSNR↑/SSIM↑/LPIPS↓/AP↑
RFormer20.36 / 0.77 / 0.342 / 50.9419.30 / 0.70 / 0.493 / 39.1917.13 / 0.62 / 0.615 / 30.79
RFormer*27.29 / 0.87 / 0.163 / 67.2623.81 / 0.77 / 0.294 / 42.3122.51 / 0.70 / 0.374 / 34.84
Pydiff19.74 / 0.71 / 0.406 / 48.7219.13 / 0.62 / 0.535 / 36.3018.47 / 0.49 / 0.757 / 31.14
Pydiff*23.42 / 0.82 / 0.256 / 60.1122.74 / 0.76 / 0.331 / 40.9622.24 / 0.70 / 0.437 / 33.51
DarkIR20.11 / 0.76 / 0.315 / 51.6317.24 / 0.65 / 0.474 / 35.3715.09 / 0.58 / 0.621 / 15.31
VCNet24.58 / 0.82 / 0.089 / 79.0724.06 / 0.77 / 0.129 / 78.4623.09 / 0.70 / 0.198 / 76.80
", + "image_path": "9caccb28d4d86c68714d74d4651bdeb81a096c22f82b4f93305d351e5f652c1b.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_body" + } + ], + "index": 7 + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "spans": [ + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "content": "in Sec. 4.1 and the results are presented in Table 1. By adding the illumination estimator " + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "content": ", all three models exhibit around 3dB PSNR improvement by considering the adaptive illumination estimation. Since LFLL-DCU only uses one learnable parameter to estimate the illumination map, their model performs well on the L3F-20 data but poorly on L3F-50 and L3F-100. Our VCNet outperforms the current SOTA methods MSP, DRLF, and LFLL-DCU with more than " + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "inline_equation", + "content": "2\\sim 6" + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "content": " dB PSNR improvement. Fig. 5 displays the visual results comparison, where all visual effects are improved by combining the " + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "content": ". This clearly demonstrates the effectiveness of our illumination estimator " + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 312, + 473, + 555, + 676 + ], + "type": "text", + "content": " in adaptively adjusting to varying illumination conditions. The results of VCNet stand out by producing more realistic overall color compared with the ground truth. Other methods restore blurred scene details, while our VCNet recovers the clearer and richer detailed texture in the recovered images." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "We then make a comparison with single-frame-based methods in Table 2, which claim to handle dynamic lighting conditions. Their models are trained on a large-scale" + } + ] + } + ], + "index": 9 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7345" + } + ] + } + ], + "index": 10 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 82, + 70, + 530, + 184 + ], + "blocks": [ + { + "bbox": [ + 82, + 70, + 530, + 184 + ], + "lines": [ + { + "bbox": [ + 82, + 70, + 530, + 184 + ], + "spans": [ + { + "bbox": [ + 82, + 70, + 530, + 184 + ], + "type": "image", + "image_path": "3c230fdd3242aa0ff6476bda06fd48908e7c979cccee6a3fbac76470bb7d2f61.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "text", + "content": "Figure 6. Visual comparison on the " + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 55, + 191, + 555, + 214 + ], + "type": "text", + "content": "-Wild dataset, where the central view of LFs is shown. Other methods either collapse by noise, distort color, or under-/over-exposed images. Our VCNet effectively removes the noise and reconstructs well-exposed image details." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "table", + "bbox": [ + 57, + 234, + 296, + 281 + ], + "blocks": [ + { + "bbox": [ + 56, + 219, + 294, + 231 + ], + "lines": [ + { + "bbox": [ + 56, + 219, + 294, + 231 + ], + "spans": [ + { + "bbox": [ + 56, + 219, + 294, + 231 + ], + "type": "text", + "content": "Table 3. Quantitative Comparison on Real-world L3F-Wild Data." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 57, + 234, + 296, + 281 + ], + "lines": [ + { + "bbox": [ + 57, + 234, + 296, + 281 + ], + "spans": [ + { + "bbox": [ + 57, + 234, + 296, + 281 + ], + "type": "table", + "html": "
MethodsRFormer*Pydiff*DRLFDRLF+EMSPMSP+EVCNet-EVCNet
NIQE ↓15.6717.8219.9718.7818.8218.2317.8714.22
BRISQUE↓31.2036.1232.8030.5430.2029.4230.1125.98
TIMES↓39.84100.1216.2216.263.893.935.996.33
", + "image_path": "45ea8be0c053691e908e1184c7cc022222a04589727adb62dc007f39b9b6b3e2.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 293, + 295, + 388 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 293, + 295, + 388 + ], + "spans": [ + { + "bbox": [ + 55, + 293, + 295, + 388 + ], + "type": "text", + "content": "single-image dataset and we carefully fine-tune their methods using L3F-Dynamic training dataset (labelled as *), where each view of LF is considered a sample during the training. Our method outperforms others under most lighting conditions, especially in perceptual quality. Although RFormer* achieves relatively high PSNR and SSIM values in the L3F-20 dataset, it suffers from noticeable blurry artifacts, as shown in Fig. 5." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "content": "To further demonstrate the generalization ability, we applied the methods trained on the " + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "content": "-Dynamic dataset to another real-world low-light LF data " + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "content": "-Wild. The results are shown in Fig. 6 and Table 3. Due to the absence of redundant information from different views, the results of single-frame methods are relatively dark, and the restored details suffer from blurring. For LF-based methods, DRLF is observed to cause a noticeable color shift. The results of MSP and LFLL-DCU suffer from residual noise and local over-expose. Combining with the " + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "content": ", the recovered results of MSP and DNLF have higher color fidelity, clearer structural details and less noise. In contrast, our results achieve high perceptual quality with clear structures, realistic color and less noises. Our VCAIA module " + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 389, + 295, + 569 + ], + "type": "text", + "content": " is also proved that can enhance the perceptual quality of other methods." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "Fixed Illumination Task. We then evaluate all the methods under the fixed illumination condition by separately training 3 modals using the " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": "-Fixed dataset. The results using LFs with angular resolution " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 55, + 570, + 295, + 713 + ], + "type": "text", + "content": " are illustrated in Table 4, respectively. We retrain the LRT [39] and LFLL-DCU [18] for comparison. When compared with single-frame-based methods, LF-based methods demonstrate higher performance, highlighting the utility of redundant LF information in low-light enhancement. With the increase of the input image views, the recovery effect of the LF models is better because it provides more redundant information to the network. Note that in PMSNet," + } + ] + } + ], + "index": 6 + }, + { + "type": "table", + "bbox": [ + 340, + 236, + 531, + 396 + ], + "blocks": [ + { + "bbox": [ + 335, + 219, + 533, + 231 + ], + "lines": [ + { + "bbox": [ + 335, + 219, + 533, + 231 + ], + "spans": [ + { + "bbox": [ + 335, + 219, + 533, + 231 + ], + "type": "text", + "content": "Table 4. Quantitative Comparisons on L3F-fixed Task." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 340, + 236, + 531, + 396 + ], + "lines": [ + { + "bbox": [ + 340, + 236, + 531, + 396 + ], + "spans": [ + { + "bbox": [ + 340, + 236, + 531, + 396 + ], + "type": "table", + "html": "
Ang ResMethodL3F-20L3F-50L3F-100
PSNR↑SSIM↑PSNR↑SSIM↑PSNR↑SSIM↑
8L3F-Net24.780.82623.210.75221.830.683
IINet25.400.86224.140.81122.130.721
DistgSSR25.440.86424.180.81022.960.755
PMSNet27.030.86624.520.81222.740.762
VCNet26.840.87125.410.82323.580.772
5L3F-Net24.980.81423.370.75321.950.669
FER-Net25.310.82323.400.76621.730.688
LRT25.970.81523.990.74222.150.691
LFLL-DCU25.420.84224.030.78722.280.702
DRLF24.780.806422.520.721721.510.6543
DRLF+ε24.990.817923.060.732922.190.6681
MSP25.480.830124.100.775922.750.7133
MSP+ε25.730.843024.680.783123.060.7191
VCNet-ε25.870.837224.350.789722.690.7204
VCNet26.610.847325.130.792123.240.7223
", + "image_path": "8a6a9570eeddb4a08d5b6e823744d97548ea560e7b6e75b1fa664337b85fa1d7.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "text", + "content": "in addition to the original " + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "inline_equation", + "content": "8 \\times 8" + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "text", + "content": " views, views from two more rows and columns are included in the feature extraction. Among LF-based methods, our VCNet significantly outperforms SOTA methods on most of the " + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "text", + "content": "-fixed task, demonstrating its ability to better handle low-illumination degradation and produce superior restoration results. After adding the illumination estimator " + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 413, + 555, + 509 + ], + "type": "text", + "content": ", all models also achieve around a 0.2-0.8 dB PSNR improvement." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": "View Consistency Comparison. View consistency holds great significance in LF enhancement as it directly impacts the performance of subsequent LF-based applications like depth estimation. To verify the effectiveness of our VCAIA " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " in maintaining consistency in recovery results, we present the parallax content Precision-Recall (PR) curves used in [5] for evaluation. In one LF, each view is subtracted from the central view, and different threshold values are chosen to calculate binary edge maps. The PR curves show how closely the binary edge maps of the reconstructed LFs are match to the ground truth. We use models trained on dynamic illumination for comparison. In Fig. 7, when all the methods add the illumination adjustment, the PR curve of the recovered image is improved, demonstrating that our " + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 313, + 510, + 556, + 715 + ], + "type": "text", + "content": " can achieve view-consistent illumination adjustment. We further calculate the Average Precision (AP) by measuring the area under the PR curves for quantitative" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7346" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 69, + 293, + 129 + ], + "blocks": [ + { + "bbox": [ + 57, + 69, + 293, + 129 + ], + "lines": [ + { + "bbox": [ + 57, + 69, + 293, + 129 + ], + "spans": [ + { + "bbox": [ + 57, + 69, + 293, + 129 + ], + "type": "image", + "image_path": "79a106eca2f9b6bb151d61406fd641448592f227c717d3b1b3ddac630ae229b9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "type": "text", + "content": "Figure 7. The PR curves map comparison of L3F-20, 50, 100 dataset, in which how well the views consistency preserved is shown. With " + }, + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 134, + 295, + 167 + ], + "type": "text", + "content": ", the PR curves of all methods are better." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 181, + 295, + 348 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 181, + 295, + 348 + ], + "spans": [ + { + "bbox": [ + 54, + 181, + 295, + 348 + ], + "type": "text", + "content": "comparison. As in Table 2, for the single-frame methods, it is difficult to recover views with consistent illumination, so that the AP is relatively low even with high PSNR and SSIM. By contrast, our model exhibits notably higher AP values, especially under the challenging low-light conditions of the L3F-100 dataset. The comparison with LF-based methods is shown in Fig. 1. Our model consistently achieves the highest AP values, showcasing the superiority of VCNet in terms of low-light enhancement and consistency preservation. The EPIs of the recovered images are also visualized in Fig. 5, in which the orientation of lines corresponds to the disparity information. It is also clear that our model excels in maintaining the view consistency. More results are provided in the supplementary materials." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "content": "Running time. We further show the running time of the models using the size of the scatter in Fig. 1 and Table 3. Our model runs faster (6.3s) than DRLF (16.2s) but is slightly slower than MSP (3.9s). The computation time of " + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "content": " is 0.04s on the A5000 for a " + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "inline_equation", + "content": "5 \\times 5 \\times 642 \\times 432" + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "content": " input, which is lightweight to add " + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 348, + 296, + 420 + ], + "type": "text", + "content": " to different models." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 428, + 149, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 149, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 149, + 441 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 104, + 471, + 248, + 527 + ], + "blocks": [ + { + "bbox": [ + 80, + 451, + 271, + 462 + ], + "lines": [ + { + "bbox": [ + 80, + 451, + 271, + 462 + ], + "spans": [ + { + "bbox": [ + 80, + 451, + 271, + 462 + ], + "type": "text", + "content": "Table 5. The Ablation Study on L3F-Dynamic Task." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 104, + 471, + 248, + 527 + ], + "lines": [ + { + "bbox": [ + 104, + 471, + 248, + 527 + ], + "spans": [ + { + "bbox": [ + 104, + 471, + 248, + 527 + ], + "type": "table", + "html": "
No.LocalGlobal\\( \\ell_{vc} \\)PSNRSSIMLPIPS
(1)×××20.760.700.2036
(2)×√×22.780.750.1469
(3)√√×23.590.750.1420
(4)√√√23.840.760.1395
", + "image_path": "424d34fe8162df3c97d89c49d477fc6b10a1710b8024c3d2d11a6928af53c14f.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "spans": [ + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": "We conduct a series of ablation studies under dynamic illumination conditions with " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "5 \\times 5" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " LF images. To ensure fairness, we keep the parameters of the variant models close to VCNet by adjusting the feature channels. It is clear that the performance of the VCNet without " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\mathcal{E}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " (model-1) is the worst. Compared with model-1, model-2 using the global brightness refinement performs better. Meanwhile, with the local brightness adaptive adjustment branch, model-3 increases " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "0.8\\mathrm{dB}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " in the PSNR values than model-2. We further validate the effectiveness of the loss " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": ". Comparing model-4 with model-3, the PSNR increases nearly " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "0.4\\mathrm{db}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " by adding the " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " to supervise the illumination adjustment matrix " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}^l}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": ". This shows that " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\ell_{vc}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " can perform consistency constraints on " + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}^l}" + }, + { + "bbox": [ + 55, + 534, + 295, + 701 + ], + "type": "text", + "content": " to provide higher-quality light-up images." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 67, + 701, + 295, + 714 + ], + "type": "text", + "content": "We further visualize the adaptive illumination adjust" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 69, + 550, + 159 + ], + "blocks": [ + { + "bbox": [ + 316, + 69, + 550, + 159 + ], + "lines": [ + { + "bbox": [ + 316, + 69, + 550, + 159 + ], + "spans": [ + { + "bbox": [ + 316, + 69, + 550, + 159 + ], + "type": "image", + "image_path": "c7db31cb773bc26ebebcca0c4e7cc7a1a92e5c1f644f02c9423305cdc95a46b5.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 164, + 555, + 196 + ], + "lines": [ + { + "bbox": [ + 313, + 164, + 555, + 196 + ], + "spans": [ + { + "bbox": [ + 313, + 164, + 555, + 196 + ], + "type": "text", + "content": "Figure 8. Visualization of the illumination maps in " + }, + { + "bbox": [ + 313, + 164, + 555, + 196 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 313, + 164, + 555, + 196 + ], + "type": "text", + "content": "-Dynamic task. For darker scenes, the values in the estimated illumination adjustment map is larger." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 317, + 198, + 552, + 286 + ], + "blocks": [ + { + "bbox": [ + 317, + 198, + 552, + 286 + ], + "lines": [ + { + "bbox": [ + 317, + 198, + 552, + 286 + ], + "spans": [ + { + "bbox": [ + 317, + 198, + 552, + 286 + ], + "type": "image", + "image_path": "c1c25df274ae9ab220fd65b2f8f71395366a054a2b5cbe128495740dc1b1ce1c.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 294, + 555, + 327 + ], + "lines": [ + { + "bbox": [ + 313, + 294, + 555, + 327 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 555, + 327 + ], + "type": "text", + "content": "Figure 9. Visual comparison of the enhanced result (top row) and input central view (second row) of one scene with different low-light levels in the L3F-Wild dataset." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "spans": [ + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "text", + "content": "ment matrix " + }, + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "inline_equation", + "content": "\\overline{\\mathcal{L}_l}" + }, + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "text", + "content": " and the related recovered results in different illumination environments in Fig. 8. The model is trained using the " + }, + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "inline_equation", + "content": "L3F" + }, + { + "bbox": [ + 313, + 334, + 555, + 491 + ], + "type": "text", + "content": "-dynamic dataset. For input with an extended exposure time, the generated illumination adjustment matrix tends to be relatively darker. Conversely, as the exposure time diminishes, the input images become darker and the values of the adjustment matrix gradually increase. We also illustrate the performance of our VCNet in processing images with disparate brightness levels from the L3F-Wild dataset in Fig. 9. Our method consistently achieves superior enhancement quality across input scenes with varying levels of brightness. This consistency ensures effective low-light image restoration across varying illumination conditions." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "spans": [ + { + "bbox": [ + 314, + 501, + 388, + 513 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 521, + 555, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 521, + 555, + 677 + ], + "spans": [ + { + "bbox": [ + 313, + 521, + 555, + 677 + ], + "type": "text", + "content": "In this paper, we introduced a novel framework for scene-adaptively low-light enhancement in light field(LF) images. First, by analyzing Retinex theory and extending it to the LF, we introduce a self-supervised loss to ensure that the calculated illumination map adheres to the consistency relationship. Then, we design an illumination estimator with global adjustment and local refinement to perceive the illumination of the scene. Finally, we design a corruption restorer and import the consistency constraint into the feature aggregation process for better recovery. Extensive quantitative and qualitative experiments demonstrate that our VC-Net significantly outperforms state-of-the-art methods on both fixed and dynamic illumination enhancement tasks." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "Acknowledgement. This work was supported by the National Natural Science Foundation of China (No.62372032)." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7347" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Hyunchan Ahn, Byungjik Keum, Daehoon Kim, and Hwang Soo Lee. Adaptive local tone mapping based on retina for high dynamic range images. In Proceedings of the International Conference on Consumer Electronics (ICCE), pages 153-156, 2013. 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "spans": [ + { + "bbox": [ + 61, + 148, + 296, + 190 + ], + "type": "text", + "content": "[2] Tom E. Bishop and Paolo Favaro. The light field camera: Extended depth of field, aliasing, and superresolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 34(5):972-986, 2012. 1" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "spans": [ + { + "bbox": [ + 62, + 194, + 296, + 236 + ], + "type": "text", + "content": "[3] Chen Chen, Qifeng Chen, Jia Xu, and Vladlen Koltun. Learning to see in the dark. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 3291-3300, 2018. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "spans": [ + { + "bbox": [ + 62, + 238, + 295, + 270 + ], + "type": "text", + "content": "[4] Haoyu Chen, Jinjin Gu, and Zhi Zhang. Attention in attention network for image super-resolution. arXiv preprint arXiv:2104.09497, 2021. 4" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "spans": [ + { + "bbox": [ + 62, + 272, + 295, + 304 + ], + "type": "text", + "content": "[5] Jie Chen, Junhui Hou, and Lap-Pui Chau. Light field denoising via anisotropic parallax analysis in a cnn framework. IEEE Signal Processing Letters, 25(9):1403-1407, 2018. 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "spans": [ + { + "bbox": [ + 62, + 307, + 295, + 350 + ], + "type": "text", + "content": "[6] Daniel Feijoo, Juan C. Benito, Alvaro Garcia, and Marcos V. Conde. Darkir: Robust low-light image restoration. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 10879-10889, 2025. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "spans": [ + { + "bbox": [ + 62, + 352, + 295, + 407 + ], + "type": "text", + "content": "[7] Xueyang Fu, Delu Zeng, Yue Huang, Xiao-Ping Zhang, and Xinghao Ding. A weighted variational model for simultaneous reflectance and illumination estimation. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 2782-2790, 2016. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "spans": [ + { + "bbox": [ + 62, + 409, + 295, + 453 + ], + "type": "text", + "content": "[8] Chunle Guo, Jing Jin, Junhui Hou, and Jie Chen. Accurate light field depth estimation via an occlusion-aware network. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 1-6, 2020. 1" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 454, + 295, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 454, + 295, + 509 + ], + "spans": [ + { + "bbox": [ + 62, + 454, + 295, + 509 + ], + "type": "text", + "content": "[9] Chunle Guo, Chongyi Li, Jichang Guo, Chen Change Loy, Junhui Hou, Sam Kwong, and Runmin Cong. Zero-reference deep curve estimation for low-light image enhancement. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), pages 1777-1786, 2020. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 510, + 295, + 564 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 510, + 295, + 564 + ], + "spans": [ + { + "bbox": [ + 57, + 510, + 295, + 564 + ], + "type": "text", + "content": "[10] Mantang Guo, Junhui Hou, Jing Jin, Jie Chen, and LapPui Chau. Deep spatial-angular regularization for light field imaging, denoising, and super-resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(10): 6094-6110, 2022. 6" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "spans": [ + { + "bbox": [ + 57, + 567, + 295, + 599 + ], + "type": "text", + "content": "[11] Xiaojie Guo, Yu Li, and Haibin Ling. Lime: Low-light image enhancement via illumination map estimation. IEEE Transactions on Image Processing, 26(2):982-993, 2017. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "spans": [ + { + "bbox": [ + 57, + 601, + 295, + 654 + ], + "type": "text", + "content": "[12] Katrin Honauer, Ole Johannsen, Daniel Kondermann, and Bastian Goldluecke. A dataset and evaluation methodology for depth estimation on 4d light fields. In Proceedings of the Asian Conference on Computer Vision (ACCV), pages 19-34, 2017. 4" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 57, + 657, + 295, + 689 + ], + "type": "text", + "content": "[13] Jing Jin and Junhui Hou. Occlusion-aware unsupervised learning of depth from 4-d light fields. IEEE Transactions on Image Processing, 31:2216-2228, 2022. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 295, + 714 + ], + "type": "text", + "content": "[14] Mohit Lamba and Kaushik Mitra. Fast and efficient restoration of extremely dark light fields. In Proceedings of the Win-" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 317, + 74, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 30, + "blocks": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 74, + 553, + 95 + ], + "type": "text", + "content": "ter Conference on Applications of Computer Vision (WACV), pages 3152-3161, 2022. 2, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 140 + ], + "type": "text", + "content": "[15] Mohit Lamba, Kranthi Kumar Rachavarapu, and Kaushik Mitra. Harnessing multi-view perspective of light fields for low-light imaging. IEEE Transactions on Image Processing, 30:1501-1513, 2021. 1, 2, 5, 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 317, + 141, + 553, + 174 + ], + "type": "text", + "content": "[16] Marc Levoy and Pat Hanrahan. Light field rendering. In Proceedings of the 23rd annual conference on Computer graphics and interactive techniques. ACM, 1996. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "spans": [ + { + "bbox": [ + 317, + 175, + 553, + 218 + ], + "type": "text", + "content": "[17] Gaosheng Liu, Huanjing Yue, Jiamin Wu, and Jingyu Yang. Intra-inter view interaction network for light field image super-resolution. IEEE Transactions on Multimedia, 25: 256-266, 2023. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 220, + 553, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 220, + 553, + 263 + ], + "spans": [ + { + "bbox": [ + 317, + 220, + 553, + 263 + ], + "type": "text", + "content": "[18] Xianqiang Lyu and Junhui Hou. Enhancing low-light light field images with a deep compensation unfolding network. IEEE Transactions on Image Processing, 33:4131-4144, 2024. 2, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 317, + 266, + 553, + 319 + ], + "type": "text", + "content": "[19] Paras Maharjan, Li Li, Zhu Li, Ning Xu, Chongyang Ma, and Yue Li. Improving extreme low-light image denoising via residual learning. In Proceedings of the International Conference on Multimedia and Expo (ICME), pages 916-921, 2019. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "spans": [ + { + "bbox": [ + 317, + 322, + 553, + 364 + ], + "type": "text", + "content": "[20] Anish Mittal, Anush Krishna Moorthy, and Alan Conrad Bovik. No-reference image quality assessment in the spatial domain. IEEE Transactions on Image Processing, 21 (12):4695-4708, 2012. 6" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 366, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 366, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 317, + 366, + 553, + 399 + ], + "type": "text", + "content": "[21] Anish Mittal, Rajiv Soundararajan, and Alan C. Bovik. Making a \"completely blind\" image quality analyzer. IEEE Signal Processing Letters, 20(3):209-212, 2013. 6" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 401, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 401, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 317, + 401, + 553, + 455 + ], + "type": "text", + "content": "[22] Junxiang Ruan, Xiangtao Kong, Wenqi Huang, and Wenming Yang. Retformer: Retinex-based enhancement in transformer for low-light image. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5, 2023. 2, 3, 6" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 456, + 555, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 456, + 555, + 511 + ], + "spans": [ + { + "bbox": [ + 317, + 456, + 555, + 511 + ], + "type": "text", + "content": "[23] Ruixing Wang, Qing Zhang, Chi-Wing Fu, Xiaoyong Shen, Wei-Shi Zheng, and Jiaya Jia. Underexposed photo enhancement using deep illumination estimation. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 6842-6850, 2019. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 513, + 553, + 567 + ], + "type": "text", + "content": "[24] Ruixing Wang, Xiaogang Xu, Chi-Wing Fu, Jiangbo Lu, Bei Yu, and Jiaya Jia. Seeing dynamic scene in the dark: A high-quality video dataset with mechatronic alignment. In Proceedings of the International Conference on Computer Vision (ICCV), pages 9680-9689, 2021. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 553, + 612 + ], + "type": "text", + "content": "[25] Tingxian Wang, Alexei A. Efros, and Ravi Ramamoorthi. Depth estimation with occlusion modeling using light-field cameras. IEEE Transactions on Pattern Analysis and Machine Intelligence, 38:2170-2181, 2016. 1" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 613, + 553, + 667 + ], + "type": "text", + "content": "[26] Wei Wang, Xin Chen, Cheng Yang, Xiang Li, Xuemei Hu, and Tao Yue. Enhancing low light videos by exploring high sensitivity camera noise. In Proceedings of the International Conference on Computer Vision (ICCV), pages 4110-4118, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 317, + 670, + 553, + 713 + ], + "type": "text", + "content": "[27] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 6" + } + ] + } + ], + "index": 29 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7348" + } + ] + } + ], + "index": 31 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[28] Xianglang Wang, Youfang Lin, and Shuo Zhang. Multi-stream progressive restoration for low-light light field enhancement and denoising. IEEE Transactions on Computational Imaging, 9:70-82, 2023. 1, 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 162 + ], + "type": "text", + "content": "[29] Xingzheng Wang, Kaiqiang Chen, Zixuan Wang, and Henhao Huang. Pmsnet: Parallel multi-scale network for accurate low-light light-field image enhancement. IEEE Transactions on Multimedia, 26:2041-2055, 2024. 2, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 294, + 228 + ], + "type": "text", + "content": "[30] Yinglong Wang, Zhen Liu, Jianzhuang Liu, Songcen Xu, and Shuaicheng Liu. Low-light image enhancement with illumination-aware gamma correction and complete image modelling network. In Proceedings of the International Conference on Computer Vision (ICCV), pages 13128-13137, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 284 + ], + "type": "text", + "content": "[31] Yingqian Wang, Longguang Wang, Gaochang Wu, Jungang Yang, Wei An, Jingyi Yu, and Yulan Guo. Disentangling light fields for super-resolution and disparity estimation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(1):425-443, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "spans": [ + { + "bbox": [ + 56, + 286, + 294, + 330 + ], + "type": "text", + "content": "[32] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 294, + 374 + ], + "type": "text", + "content": "[33] Zhou Wang, A.C. Bovik, H.R. Sheikh, and E.P. Simoncelli. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600-612, 2004. 5" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "spans": [ + { + "bbox": [ + 56, + 376, + 294, + 409 + ], + "type": "text", + "content": "[34] Chen Wei, Wenjing Wang, Wenhan Yang, and Jiaying Liu. Deep retina decomposition for low-light enhancement. arXiv preprint arXiv:1808.04560, 2018. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 411, + 294, + 464 + ], + "type": "text", + "content": "[35] Henry Wing Fung Yeung, Junhui Hou, Xiaoming Chen, Jie Chen, Zhibo Chen, and Yuk Ying Chung. Light field spatial super-resolution using deep efficient spatial-angular separable convolution. IEEE Transactions on Image Processing, 28(5):2319-2330, 2018. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 468, + 294, + 520 + ], + "type": "text", + "content": "[36] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the Computer Vision and Pattern Recognition Conference (CVPR), 2018. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 566 + ], + "type": "text", + "content": "[37] Shansi Zhang and Edmund Y. Lam. An effective decomposition-enhancement method to restore light field images captured in the dark. Signal Processing, 189:108279, 2021. 1, 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[38] Shuo Zhang, Youfang Lin, and Hao Sheng. Residual networks for light field image super-resolution. In Proceedings of the Conference on Computer Vision and Pattern Recognition (CVPR), pages 11046-11055, 2019. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 656 + ], + "type": "text", + "content": "[39] Shansi Zhang, Nan Meng, and Edmund Y. Lam. Lrt: An efficient low-light restoration transformer for dark light field images. IEEE Transactions on Image Processing, 32:4314-4326, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[40] Yonghua Zhang, Jiawan Zhang, and Xiaojie Guo. Kindling the darkness: A practical low-light image enhancer. In Proceedings of the 27th ACM International Conference on Multimedia, pages 1632-1640, New York, NY, USA, 2019. ACM. 2" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 73, + 554, + 139 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 73, + 554, + 106 + ], + "type": "text", + "content": "[41] Yonghua Zhang, Xiaojie Guo, Jiayi Ma, Wei Liu, and Jiawan Zhang. Beyond brightening low-light images. International Journal of Computer Vision., 129(4):1013-1037, 2021. 2" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 554, + 139 + ], + "type": "text", + "content": "[42] Dewei Zhou, Zongxin Yang, and Yi Yang. Pyramid diffusion models for low-light image enhancement. arXiv preprint arXiv:2305.10028, 2023. 6" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "7349" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_content_list.json b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..e5aba2c5dc1fae6501e0837197e5b9beac960194 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_content_list.json @@ -0,0 +1,1600 @@ +[ + { + "type": "text", + "text": "Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions", + "text_level": 1, + "bbox": [ + 202, + 130, + 795, + 176 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yuwen Pan $^{1}$ Rui Sun $^{1}$ Wangkai Li $^{1}$ Tianzhu Zhang $^{1,2*}$", + "bbox": [ + 256, + 202, + 720, + 220 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 305, + 220, + 691, + 238 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory", + "bbox": [ + 138, + 239, + 854, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{panyw, issunrui, lwklwk}@mail.ustc.edu.cn, tzzhang@ustc.edu.cn", + "bbox": [ + 217, + 258, + 774, + 273 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 308, + 326, + 324 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation under adverse conditions is critical for reliable visual perception in challenging weather environments. These extreme scenarios introduce distortions, such as low contrast and reduced visibility, making traditional segmentation models struggle. The scarcity of labeled data in such conditions makes it difficult to train models directly for these environments. Unsupervised domain adaptation (UDA) has been proposed as a solution to transfer knowledge from labeled source domains (normal weather) to unlabeled target domains (adverse weather). However, existing methods face significant challenges, particularly due to weather unawareness and feature heterogeneity. Many models fail to account for the unique characteristics of different weather conditions, and the significant feature discrepancies between normal and adverse weather images hinder effective adaptation. In this paper, we propose a novel weather-aware aggregation and adaptation network that leverages characteristic knowledge to achieve weather homogenization and enhance scene perception. Specifically, we introduce amplitude prompt aggregation to capture essential characteristics from the Fourier frequency domain that are indicative of different weather conditions. Additionally, we employ weather heterogeneity adaptation to mitigate the inter-domain heterogeneity, thereby achieving feature homogenization across diverse environments. Extensive experimental results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions.", + "bbox": [ + 88, + 340, + 485, + 779 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 806, + 220, + 821 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Semantic segmentation involves assigning a class label to each pixel in an image [30, 35, 36, 38, 47], which plays a crucial role in various real-world applications such as au", + "bbox": [ + 89, + 832, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/4014c906c279a86dcddccedca7e5c2ec5874bdba89befc81d7f9d53ca8590a3a.jpg", + "image_caption": [ + "(a) Traditional Cross-weather Paradigm" + ], + "image_footnote": [], + "bbox": [ + 545, + 306, + 877, + 434 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/eb4ce93c6987c17e3eba03460d11ee818b4e92033b34bd47c7584f7a3f8bc3f1.jpg", + "image_caption": [ + "(b) Weather-aware Aggregation and Adaptation (ours)", + "Figure 1. Comparison of the traditional cross-weather paradigm and our weather-aware aggregation and adaptation. To mitigate the existing issues (unawareness of weather characteristics and severe cross-weather heterogeneity), we aim to aggregate weather-aware attribute knowledge and homogenize domain discrepancies." + ], + "image_footnote": [], + "bbox": [ + 545, + 446, + 875, + 594 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "tonomous driving [34, 37, 59] and surveillance [13, 32, 60]. However, the performance of segmentation models often degrades significantly when faced with adverse weather conditions [22, 45], including night, fog, rain, and snow. The deterioration in visibility and scene appearance poses a substantial challenge for traditional segmentation algorithms, which are typically trained on datasets collected under normal weather conditions. Therefore, exploring an effective way for semantic parsing under extreme weather conditions is highly anticipated.", + "bbox": [ + 511, + 717, + 906, + 868 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Extensive research efforts have been devoted to semantic segmentation in adverse weather conditions [7, 26, 41-", + "bbox": [ + 511, + 869, + 908, + 901 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author.", + "bbox": [ + 109, + 887, + 235, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "13952", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "43, 51]. Due to the scarcity of labeled data in such adverse conditions, existing methods strive to employ an unsupervised domain adaptation (UDA) paradigm utilizing both labeled normal weather data and images without annotations under extreme scenarios to address the challenge [16, 44, 50, 53, 54]. Kerim et al. [20] propose a dual-branch network based on DeepLabV3+ [1], enhancing training knowledge for extreme weather conditions by incorporating a synthetic multi-scene dataset. However, discerning similar-looking categories in the target domain with limited real annotations remains a challenge for domain adaptation methods. To address this, MIC [18] introduces the masked image consistency, leveraging spatial context relations in the target domain as additional cues to enhance the robustness of UDA. The emergence of these methods has benefited the understanding of scenes in extreme weather conditions.", + "bbox": [ + 89, + 90, + 480, + 347 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Despite considerable progress, the task of distinguishing and analyzing scenes under different extreme weather conditions remains challenging. Whether through the assistance of synthetic data or exploration of spatial context clues, existing methods struggle with accurately identifying weather-aware semantics in multiple extreme scenarios. We attribute the predicament of this cross-weather paradigm to two inherent issues (Figure 1): (1) Unawareness of Weather Characteristics. The style characteristics under adverse weathers, such as illumination conditions and contrast, are primary culprits responsible for the inter-weather gap among diverse scenes. Although recognizing this crucial aspect, existing works [10, 58] generally employ low-level enhancements or implicit learning for inter-domain style transfer, failing to fully exploit weather attribute representation. Therefore, designing a strategy to aggregate weather-aware scene knowledge across diverse conditions is worthy of exploration. (2) Severe Cross-Weather Heterogeneity. Early works [16, 18] tend to scene images from different weather conditions uniformly, ignoring the substantial differences across different weather types. The significant heterogeneity across weather conditions presents another challenge, stemming from the diverse and complex nature of weather phenomena that introduce distinct visual distortions and scene alterations. Even though some methods [2, 11] attempt to address this challenge, the significant differences in characteristics across different weather conditions still pose considerable difficulties. These variations lead to a significant disparity between features learned from the normal weather and those required for effective interpretation in adverse scenes. Addressing how to leverage characteristic knowledge to achieve weather homogenization to facilitate scene perception in adverse scenes is an essential question.", + "bbox": [ + 89, + 351, + 486, + 866 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Motivated by the above discussions, we propose a Weather-aware Aggregation and Adaptation NETwork", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "(WA $^2$ Net) to aggregate weather-aware attribute knowledge and homogenize domain discrepancies, including amplitude prompt aggregation and weather heterogeneity adaptation. In the amplitude prompt aggregation, to acquire weather-aware characteristic knowledge, we utilize the amplitude operation of Fourier transform to capture essential characteristics indicative of different weather conditions. These amplitude features serve as discriminative cues for weather-related variations in the scene. To efficiently aggregate the crucial characteristic knowledge from the frequency domain into weather-adaptive prompts, we prepend a set of learnable prompts in the prompt bank to interact with frequency features by cross-attention mechanism to obtain weather-aware prompts. These prompts encapsulate the intrinsic characteristics of various weather conditions, providing essential guidance for the adaptation process. In the weather heterogeneity adaptation, to realize inter-weather homogenization for effective semantic parsing, the resultant weather-aware prompts are leveraged to generate adapted characteristics after modulation. Then, we propose a novel homogenization attention mechanism to acquire homogenized features across diverse weather conditions. This mechanism identifies the discrepancy between features of the original image and adapted characteristics, decoupling weather characteristics from valuable target information and thereby filtering out attribute interference. In this way, our WA $^2$ Net achieves effective weather-aware knowledge aggregation and robust adaptation, ensuring improved performance for semantic segmentation in adverse weather conditions.", + "bbox": [ + 511, + 90, + 906, + 542 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To sum up, our contributions can be summarized as follows:", + "bbox": [ + 511, + 545, + 906, + 573 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We propose a novel weather-aware aggregation and adaptation network for semantic segmentation under adverse conditions. Our paper exploits a practical way to acquire weather-adaptive knowledge and enhance cross-weather perception via an aggregation and adaptation manner.", + "- We design the amplitude prompt aggregation to acquire weather-aware prompts from weather characteristics through amplitude extraction, and the weather heterogeneity adaptation to inherit weather-aware knowledge and mitigate the inter-weather heterogeneity.", + "- Extensive experiment results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions." + ], + "bbox": [ + 513, + 577, + 903, + 786 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 513, + 805, + 653, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. Semantic Segmentation under Adverse Conditions", + "text_level": 1, + "bbox": [ + 511, + 832, + 903, + 862 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "The field of semantic segmentation [3, 4, 23, 27, 28, 39, 48] has made significant advancements with the emergence", + "bbox": [ + 511, + 869, + 906, + 901 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "13953", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "of convolutional neural networks (CNNs) [21] and transformers [52]. However, traditional semantic segmentation models experience a notable performance drop under adverse conditions such as rain, fog and night-time environments [45, 46, 61]. Therefore, semantic segmentation under adverse conditions aims to develop robust models capable of accurately delineating objects and regions in challenging environments. Upon the scarcity of labeled data in extreme weathers, the unsupervised domain adaptation paradigm is utilized to enhance the generalization of models across different weather conditions with the aid of normal weather data [24, 54]. Kerim et al. [20] propose a dual-branch network and expand the training dataset for extreme weather conditions by integrating a synthetic multiscene dataset. Nevertheless, domain adaptation methods encounter challenges due to the difficulty in distinguishing similar-looking categories in the target domain with limited real annotations. To tackle this issue, MIC [18] introduces Masked Image Consistency, utilizing spatial context relations in the target domain as additional cues to enhance vision perception. Additionally, to mitigate knowledge forgetting in new scenes, Yang et al. [57] introduce adaptive knowledge acquisition and replay of weather-condition combinations, aiming to alleviate model-forgetting issues in diverse scenes. Differently, we strive to tackle the challenge of weather-unaware perception, leveraging weather-aware prompts in an aggregation and adaptation manner.", + "bbox": [ + 89, + 90, + 483, + 500 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "2.2. Prompt Learning", + "text_level": 1, + "bbox": [ + 89, + 510, + 264, + 526 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Prompting [29] incorporates task-specific instructions into input text to guide model behavior, enhancing performance without retraining, particularly in few-shot learning [9]. Recent work extends prompting to vision tasks, where learnable prompts enable adaptation without tuning all parameters [25] and improve segmentation generalization [14]. For semantic segmentation under adverse weather, weather-adaptive prompts modulate visual embeddings to mitigate distortions like lighting and visibility changes, enhancing robustness without extensive fine-tuning. Integrating weather-specific knowledge into prompts improves generalization across diverse conditions, making models more effective in real-world applications.", + "bbox": [ + 89, + 532, + 483, + 731 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 89, + 746, + 181, + 762 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminaries", + "text_level": 1, + "bbox": [ + 89, + 771, + 232, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Typically, in the field of unsupervised domain adaptation [16, 18], the source domain data can be formulated as $D^{s} = \\{X^{s}, Y^{s}\\}$ , where $X^{s}$ and $Y^{s}$ represent the corresponding image and label space, respectively. Each source domain image $x_{k}^{s} \\in X^{s}$ is associated with a ground truth segmentation mask $y_{k}^{s} \\in Y^{s}$ . Similarly, let $X^{t}$ be the target domain image space, and $Y^{t}$ be the label space. However,", + "bbox": [ + 89, + 794, + 483, + 902 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "unlike the source domain, the target domain has no access to annotated segmentation masks. The supervised loss $\\mathcal{L}^s$ can be formulated with a neural network $f_{\\mu}$ as:", + "bbox": [ + 511, + 90, + 905, + 137 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {s} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {s}\\right), y _ {k} ^ {s}\\right), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 598, + 150, + 906, + 191 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{s}$ denotes the image number of the source domain, $\\mathcal{L}_{ce}$ denotes the pixel-wise cross-entropy loss as:", + "bbox": [ + 511, + 203, + 905, + 234 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c e} (\\hat {y}, y) = - \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {C} y _ {h w c} \\log \\hat {y} _ {h w c}. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 566, + 244, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "The model trained on the source domain $D^{s}$ normally exhibits performance degradation when tested on the target domain $D^{t}$ . Therefore, the self-training paradigm generates pseudo-labels for the target domain images $X^{t}$ with a teacher network $g_{\\phi}$ instead of the student network $f_{\\mu}$ . Specifically, the model predicts segmentation masks for unlabeled target domain images, producing pseudo-labels $\\hat{y}_k^t$ for each image $x_k^t$ in the target domain. The unsupervised loss $\\mathcal{L}^t$ for the target domain is formulated as:", + "bbox": [ + 511, + 300, + 906, + 436 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} ^ {t} = \\frac {1}{N _ {t}} \\sum_ {k = 1} ^ {N _ {t}} \\lambda^ {t} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {t}\\right), \\hat {y} _ {k} ^ {t}\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 449, + 905, + 491 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $N_{t}$ denotes the image number of the target domain, $\\lambda^t$ is a weight coefficient for the unsupervised loss. The student network updates the parameters of the teacher network through backpropagation, utilizing the teacher network's output as targets to minimize prediction errors, while also employing Exponential Moving Average (EMA) update to stabilize the training process and improve convergence.", + "bbox": [ + 511, + 502, + 905, + 609 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Overview", + "text_level": 1, + "bbox": [ + 511, + 619, + 624, + 633 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "As shown in Figure 2, given an input image $x^{s}$ from the source domain and $x^{t}$ from the target domain under adverse conditions, we denote the image under different weather condition $\\theta$ as $x^{\\theta}$ for clarity, where $\\theta \\in (0, \\dots, \\Theta)$ , $\\Theta$ denotes the total number of adverse weather conditions ( $x^{0}$ refers to the source domain image under the normal weather condition). The Fourier amplitude features $f_{a}^{\\theta}$ are decoupled from the frequency domain. Then, a set of prompts, i.e., prompt bank $\\mathbf{P}$ , is learned during the prompt aggregation process. The acquired weather-aware prompts $\\hat{\\mathbf{P}}$ further interact with amplitude features $f_{a}^{\\theta}$ to evolve into adapted characteristic $\\hat{f}_{a}^{\\theta}$ . Then, to achieve homogenization for features from diverse weather conditions, we further employ homogenization attention to mitigate the influence of weather heterogeneity. Finally, the modulated features $f_{h}^{\\theta}$ are fed into the teacher-student network for comprehensive training.", + "bbox": [ + 511, + 642, + 906, + 901 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "13954", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/61fe4643d51294a0265b4d8679594ece20fc2870f68416906e2d33b97ff393a2.jpg", + "image_caption": [ + "Figure 2. Illustration of our proposed $\\mathrm{WA}^2\\mathrm{Net}$ . Our network contains the amplitude prompt aggregation to acquire weather-aware prompts (Section 3.3) and weather heterogeneity adaptation to realize feature homogenization (Section 3.4)." + ], + "image_footnote": [], + "bbox": [ + 94, + 88, + 908, + 258 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Amplitude Prompt Aggregation", + "text_level": 1, + "bbox": [ + 89, + 321, + 372, + 338 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In order to disentangle the latent weather characteristics and target information fused in the RGB images, we first model the representation of weather characteristics by exploring the Fourier frequency domain. Then, we collect the weather characteristic properties in Fourier amplitude features and acquire weather-aware prompts during the prompt aggregation process.", + "bbox": [ + 89, + 343, + 483, + 449 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Amplitude Characteristic Extraction. In the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information [33, 58], which reflects the energy distribution of the frequency components of the image and can be understood as a representation of the \"characteristics\" or \"style\" of the image. Therefore, we can utilize the amplitude of Fourier spectrum to extract weather characteristic knowledge in diverse scenes. Specifically, in the amplitude extractor, given an image $x^{\\theta} \\in \\mathbb{R}^{H \\times W \\times 3}$ , we apply a two-dimensional Fourier transform $\\mathcal{F}(x)$ as:", + "bbox": [ + 89, + 450, + 483, + 616 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x _ {h, w} ^ {\\theta} e ^ {- J 2 \\pi \\left(\\frac {w h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 140, + 633, + 482, + 676 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $J$ refers to the imaginary unit. Then we can acquire the corresponding amplitude $\\mathcal{A}$ and phase $\\Phi$ as:", + "bbox": [ + 89, + 681, + 482, + 710 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {A} \\left(x ^ {\\theta}\\right) _ {u, v} = \\left| \\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} \\right|, \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 200, + 715, + 480, + 736 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\Phi \\left(x ^ {\\theta}\\right) _ {u, v} = \\arg \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v}\\right) = \\arctan \\left[ \\frac {\\operatorname {I m} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)}{\\operatorname {R e} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)} \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 739, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\arg$ represents the argument (or phase angle) of the complex number, Im and Re represent the imaginary and real part of $\\mathcal{F}(x)$ , respectively. In this way, to finally generate our amplitude characteristic map, we fix the phase to an average constant $c^p$ and apply inverse Fourier transform to acquire the amplitude characteristic map as", + "bbox": [ + 89, + 787, + 482, + 878 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\overline {{\\mathcal {A}}} (x ^ {\\theta}) = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x ^ {\\theta}) _ {u, v} e ^ {- J c ^ {p}} ], \\tag {7}\n$$\n", + "text_format": "latex", + "bbox": [ + 178, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathcal{F}^{-1}$ denotes the inverse Fourier transformation. Then, we can obtain the amplitude characteristics $\\mathbf{F}_a = \\{f_a^\\theta\\}_{\\theta=0}^\\Theta$ through a shared-weight lightweight encoder (e.g., ResNet-18 [15]), preparing for the following aggregation of weather-aware prompts.", + "bbox": [ + 511, + 321, + 906, + 398 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Weather-aware Prompt Aggregation. To effectively aggregate characteristic information across different weather conditions, we aim to develop a medium capable of absorbing cross-weather knowledge. Inspired by [19], we introduce a set of prompts, i.e., prompt bank $\\mathbf{P} = \\{\\pmb{p}_n\\}_{n=1}^N$ , where $\\pmb{p}_n \\in \\mathbb{R}^{1 \\times L}$ and $N$ denotes the number of prototypes. These learnable prompts will be leveraged to extract characteristic knowledge by interaction with amplitude features $f_a^\\theta$ , which highly represents the style condition of the current scene. In specific, for the acquired amplitude characteristic $f_a^\\theta$ , the queries arise from the prompts $\\mathbf{P}$ , and keys and values arise from the amplitude feature $\\mathbf{F}_a = \\{f_a^0, \\dots, f_a^\\theta, \\dots, f_a^\\Theta\\}$ . Formally,", + "bbox": [ + 511, + 398, + 906, + 597 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {Q} _ {n} = \\boldsymbol {p} _ {n} \\mathbf {W} ^ {Q}, \\mathbf {K} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {K}, \\mathbf {V} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {V}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 607, + 906, + 627 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $n\\in [1,\\dots ,N]$ , and $\\mathbf{W}^Q\\in \\mathbb{R}^{C\\times C_k}$ $\\mathbf{W}^K\\in$ $\\mathbb{R}^{C\\times C_k}$ $\\mathbf{W}^V\\in \\mathbb{R}^{C\\times C_v}$ are linear projections. The attention weights are calculated based on the dot-product similarity between each query and key:", + "bbox": [ + 511, + 637, + 905, + 700 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ns _ {n, \\theta} = \\frac {\\exp (\\beta_ {n , \\theta})}{\\sum_ {\\theta = 0} ^ {\\Theta} \\exp (\\beta_ {n , \\theta})}, \\beta_ {n, \\theta} = \\frac {\\mathbf {Q} _ {n} \\mathbf {K} ^ {\\theta} {} ^ {\\top}}{\\sqrt {C _ {k}}}, \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 562, + 712, + 905, + 753 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {\\boldsymbol {p}} _ {n} = \\operatorname {F F N} (\\operatorname {A t t} \\left(\\mathbf {Q} _ {n}, \\mathbf {K}, \\mathbf {V}\\right)) = \\operatorname {F F N} \\left(\\sum_ {\\theta = 0} ^ {\\Theta} s _ {n, \\theta} \\mathbf {V} ^ {\\theta}\\right), \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 768, + 903, + 821 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\mathrm{FFN}()$ denotes the feed forward network. In this way, weather-aware prompts $\\hat{\\mathbf{P}} = \\{\\hat{p}_n\\}_{n=1}^N$ can distill rich characteristic knowledge from the amplitude features $\\mathbf{F}_a$ .", + "bbox": [ + 511, + 824, + 903, + 869 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Besides, in order to enable prompts to carry diverse and comprehensive knowledge of various weather conditions,", + "bbox": [ + 511, + 869, + 905, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "13955", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "we apply the diversity loss on the prompt bank $\\mathbf{P}$ . Formally,", + "bbox": [ + 89, + 90, + 482, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {d i v} = \\frac {1}{N (N - 1)} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1, i \\neq j} ^ {N} \\left(\\frac {\\langle \\boldsymbol {p} _ {i} , \\boldsymbol {p} _ {j} \\rangle}{\\| \\boldsymbol {p} _ {i} \\| _ {2} \\| \\boldsymbol {p} _ {j} \\| _ {2}}\\right). \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 107, + 114, + 482, + 159 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Applying diversity loss to different prompts offers the advantage of encouraging each prompt $\\pmb{p}_n$ to focus on distinct attributes of weather conditions, thereby preventing redundancy in aggregation learning. Diversity loss enhances weather-specific prompts by encouraging each to capture unique traits. Integrating the prompt bank and diversity loss within domain adaptation allows the model to adapt to varied weathers. Splitting this into two tasks is limited by the lack of direct supervision on weather-adaptive prompts; only foreground masks from the source domain are available. Diversity loss thus broadens the prompt bank's coverage of style features for better domain adaptation.", + "bbox": [ + 89, + 167, + 483, + 349 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.4. Weather Heterogeneity Adaptation", + "text_level": 1, + "bbox": [ + 89, + 357, + 397, + 373 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In adverse weather conditions, weather-induced distortions such as glare, fog, and raindrops introduce complexities that can degrade the performance of vision perception. To address the weather heterogeneity challenges posed by varying weather conditions, we introduce the weather heterogeneity adaptation, which utilizes acquired weather-adaptive prompts $\\hat{\\mathbf{P}}$ as references to modulate the weather characteristic in the current scene and homogenizes weather features via homogenization attention.", + "bbox": [ + 89, + 378, + 483, + 515 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Prompt-guided Modulation. In fact, the introduction of the prompt bank can be deemed as a guidepost in artificially solving incidents, i.e., guiding information, which absorbs weather-aware knowledge across various scenes and can transfer reliable weather-aware properties into modulated characteristic features. For the acquired weather amplitude feature $f_{a}^{\\theta}$ , the modulated characteristic can be formulated as:", + "bbox": [ + 89, + 515, + 483, + 635 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\hat {f} _ {a} ^ {\\theta} = \\operatorname {S o f t m a x} \\left(\\frac {f _ {a} ^ {\\theta} \\hat {\\mathbf {P}} ^ {\\top}}{\\sqrt {C}}\\right) \\hat {\\mathbf {P}}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 197, + 642, + 482, + 678 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In this way, the modulated characteristic $\\hat{f}_a^\\theta$ inherits rich cross-weather knowledge from the weather-aware prompts $\\hat{\\mathbf{P}}$ , thereby evolving into an adapted representation poised for subsequent homogenization.", + "bbox": [ + 89, + 688, + 482, + 748 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Homogenization Attention Mechanism. With the acquired representation of modulated characteristic $\\hat{f}_a^\\theta$ , we first extract the vision features of image $x^{\\theta}$ from the RGB domain through feature encoding as $f_v^\\theta$ . To homogenize vision features from diverse weather conditions, we aim to find an appropriate way to mitigate the influence of $\\hat{f}_a^\\theta$ on $f_v^\\theta$ . Inspired by the success of Transformer [52] architecture in discovering local regions, we further explore the potential of attention mechanism in characteristic homogenization for different weather conditions.", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Given an image feature $f_{v}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times c}$ extracted from the vision feature extractor (e.g., MiT-B5 [56]), we use the extracted $\\hat{f}_{a}^{\\theta}$ to remove the weather characteristics through a novel homogenization attention mechanism. First, we employ two convolution layers to map $f_{v}^{\\theta}$ and $\\hat{f}_{a}^{\\theta}$ to the same dimension $C$ , then we can obtain $\\overline{\\mathbf{f}}_{\\mathbf{v}}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C}$ and $\\overline{f}_{a}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C}$ , respectively. The disparity between image features and weather characteristics signifies the saliency details of the target objects, thereby aiding in accentuating the relevant information about the targets. Rather than vanilla transformer attention mechanism [52], we custom design the homogenization attention to reduce heterogeneity in a finer manner, prompting the vision system to highlight salient regions across diverse scenes. Specifically, we acquire the homogenized map $\\mathcal{H} \\in \\mathbb{R}^{h \\times w}$ from pixel-level features and weather characteristic as:", + "bbox": [ + 511, + 90, + 906, + 339 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {H} _ {i, j} ^ {\\theta} = \\sum_ {c = 1} ^ {C} \\left(\\overline {{\\mathbf {f}}} _ {\\mathbf {v} _ {i, j, c}} ^ {\\theta} - \\overline {{f}} _ {a i, j, c} ^ {\\theta}\\right) ^ {2}, \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 343, + 906, + 386 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $i, j,$ and $c$ are the index of height, width, and channel, respectively. Finally, we can get the homogenized pixel features $\\mathbf{F}^a$ by weighting the homogenized map $\\mathcal{H}$ to original image features in a pixel-wise manner as:", + "bbox": [ + 511, + 388, + 905, + 450 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {f} _ {h} ^ {\\theta} = \\mathbf {f} ^ {\\theta} \\circ \\mathcal {H} ^ {\\theta}, \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 658, + 455, + 906, + 474 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\circ$ denotes the element-wise product. Besides, we use a self-attention layer for further aggregating target information across different pixels. The obtained visual features are fed into the conventional teacher-student network, wherein the images from the source domain are input to the student network for supervised constraints, while the images from the target domain are separately input to both the student and teacher networks to generate predictions and pseudolabels for unsupervised constraints.", + "bbox": [ + 511, + 479, + 905, + 617 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.5. Training Supervision", + "text_level": 1, + "bbox": [ + 513, + 625, + 712, + 640 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "As illustrated in Section 3.1, the acquired adapted vision features from the source and target are input to the teacher-student network to acquire the supervised loss $\\mathcal{L}^s$ and unsupervised loss $\\mathcal{L}^t$ . Together with the diversity loss $\\mathcal{L}_{div}$ in the prompt bank, we can finally obtain the segmentation map as follows:", + "bbox": [ + 511, + 645, + 905, + 736 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} ^ {s} + \\mathcal {L} ^ {t} + \\lambda_ {\\text {d i v}} \\mathcal {L} _ {\\text {d i v}}, \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 606, + 742, + 903, + 760 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\lambda_{div}$ is the trade-off weight.", + "bbox": [ + 511, + 767, + 743, + 782 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiment", + "text_level": 1, + "bbox": [ + 511, + 792, + 638, + 810 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Experiment Setup", + "text_level": 1, + "bbox": [ + 511, + 818, + 689, + 834 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. In our experiments, following the conventional practice [16, 18, 20], we utilize 4 datasets including CityScapes [6], ACDC [45], NightCity [49] and DarkZurich [44]. For the source domain, we use the training", + "bbox": [ + 511, + 839, + 905, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "13956", + "bbox": [ + 480, + 944, + 519, + 957 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/2a294b57cab2367a98b901c947534bfa845d7c16205ecfc1f99bdf793cabfb5c.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons of semantic segmentation performance under adverse conditions on CityScapes [6]→ACDC [45], where the best performance is marked as **bold** and the second best is **underlined**. † denotes the performance of our reproduction." + ], + "table_footnote": [], + "table_body": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]72.914.340.516.621.29.317.421.263.823.818.332.619.569.536.234.546.226.936.132.7
\\(MGCDA^†\\) [44]73.428.769.919.326.336.853.053.375.432.084.651.026.177.643.245.953.932.741.548.7
\\(DANNet^†\\) [54]84.354.277.638.030.018.941.635.271.339.486.648.729.276.241.643.058.632.643.950.0
DAFormer [16]58.451.384.042.735.150.730.057.074.852.851.358.332.682.758.354.982.444.150.755.4
\\(SSAC^†\\) [20]76.559.882.752.634.551.258.161.263.958.167.261.540.081.859.261.783.447.555.659.6
HRDA [17]88.357.988.155.236.756.362.965.374.257.785.968.845.788.576.482.487.752.760.468.0
MIC [18]90.867.189.254.540.557.262.068.476.361.887.071.349.489.775.786.889.156.963.070.4
HALO [8]94.279.888.260.251.164.178.265.687.955.795.566.320.788.982.289.387.950.459.071.9
CoDA [12]93.172.790.757.347.456.869.970.087.359.895.471.447.690.377.183.889.154.764.172.6
\\(WA^2Net (ours)\\)94.577.191.661.947.657.568.570.479.264.095.672.850.290.575.588.790.957.562.673.8
", + "bbox": [ + 99, + 128, + 903, + 309 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/b34a12fef0e95e5b836cf27a0420c84c8978007fd2186173e0b2d384a3b6e91e.jpg", + "table_caption": [ + "Table 2. Quantitative comparisons of performance on CityScapes [6] $\\rightarrow$ NightCity [49]. Considering the NightCity [49] dataset is newly introduced, we reproduce the results of state-of-the-art methods for better comparison." + ], + "table_footnote": [], + "table_body": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
DANNet [54]57.536.452.342.539.533.227.846.148.918.467.049.528.357.244.050.443.229.137.444.0
DAFormer [16]64.241.059.647.043.235.731.952.050.320.373.151.829.762.049.858.945.331.239.846.8
SSAC [20]84.546.775.553.149.638.331.048.856.820.281.552.630.275.853.561.146.830.237.548.4
HRDA [17]83.550.578.651.948.340.535.457.858.219.981.856.329.780.565.880.750.431.040.852.2
MIC [18]84.951.678.453.848.641.838.860.557.122.083.460.231.981.464.870.852.237.843.853.8
HALO [8]85.752.177.952.947.239.337.563.255.621.582.161.430.780.563.972.353.439.144.754.2
CoDA [12]86.150.979.151.749.340.136.965.458.323.282.762.833.180.262.574.654.938.345.654.9
WA2Net (ours)87.454.279.354.349.940.642.267.859.422.883.968.234.383.565.176.156.846.448.256.5
", + "bbox": [ + 99, + 363, + 901, + 518 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "split of CityScapes; for the target domain, we use either the test of ACDC/DarkZurich or the validation set of NightCity. (1) CityScapes' training set contains 2,975 images of 19 classes with resolution of $2048 \\times 1024$ . (2) ACDC contains weather conditions including fog, nighttime, rain, and snow. The test set contains 2,000 images of $1920 \\times 1080$ resolution. (3) NightCity includes 4,297 real night-time images, divided into 2,998 train images and 1,299 val images. (4) DarkZurich contains 2,416 training and 151 test images for nighttime.", + "bbox": [ + 88, + 544, + 480, + 696 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Implementation Details. We adopt Pytorch [40] and Detectron2 [55] to implement the proposed method. 4 NVIDIA GeForce RTX 3090 GPUs are used for training. The extractor of amplitude characteristic is ResNet-18 [15]. During the training stage, our model is trained with a batch size of 4, using the Adam optimizer [31] with an initial learning rate of 0.0001. The input images are rescaled and randomly cropped to $512 \\times 512$ following the same data augmentation in DAFormer [16], and the EMA coefficient for updating the teacher net is set to be 0.999. We set the prompt number as $N = 16$ , and the coefficient of the diversity loss as $\\lambda_{div} = 0.2$ .", + "bbox": [ + 88, + 699, + 482, + 881 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. For a fair comparison, we adopt the", + "bbox": [ + 89, + 885, + 482, + 901 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "metric of mean intersection over union (mIOU). It measures the similarity between the prediction and the ground truth by computing the ratio of the intersection area to the union area of the two masks for each object class.", + "bbox": [ + 511, + 544, + 906, + 606 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with State-of-the-Art Methods", + "text_level": 1, + "bbox": [ + 511, + 616, + 880, + 633 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our method demonstrates superior performance for semantic segmentation under adverse conditions, outperforming state-of-the-art methods in Table 1, 2& 3. As can be observed, our method achieves significant improvements in challenging classes such as road, traffic light and person. Besides, as shown in Figure 3, our method shows promising segmentation performance in diverse weather scenes. In specific, our method performs great in most scenarios, while MIC [18] may generate confusing activation due to the inability to explicitly perceive weather conditions, our $\\mathrm{WA}^2\\mathrm{Net}$ can adaptively focus on accurate foreground targets owing to weather-aware aggregation and adaptation.", + "bbox": [ + 511, + 638, + 906, + 821 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 511, + 832, + 666, + 848 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Effectiveness of the Amplitude Prompt Aggregation. The amplitude prompt aggregation consists of two processes: amplitude characteristic extraction and weather-", + "bbox": [ + 511, + 854, + 905, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "13957", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/5312aaebc363028fd095fb4665655d438aa82084b47f00db5936d93539adf0d1.jpg", + "table_caption": [ + "Table 3. Quantitative comparisons of performance on CityScapes [6] $\\rightarrow$ DarkZurich [44], where the best performance is marked as bold and the second is underlined. $^\\dagger$ denotes the performance of our reproduction of the corresponding model." + ], + "table_footnote": [], + "table_body": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]85.837.955.527.714.523.114.021.132.18.72.039.916.664.013.80.058.828.520.729.7
\\( MGCDA^† \\)[44]80.349.366.27.811.041.438.939.064.118.055.852.153.574.766.00.037.529.122.742.5
\\( DANNet^† \\)[54]90.054.074.841.021.125.026.830.272.026.284.047.033.968.219.00.366.438.323.644.3
DAFormer [16]93.565.573.339.419.253.344.144.059.534.566.653.452.782.152.79.589.350.538.553.8
HRDA [17]90.456.372.039.519.557.852.743.159.329.170.560.058.684.075.511.290.551.640.955.9
MIC [18]94.875.084.055.128.462.035.552.659.246.870.065.261.782.164.218.591.352.644.060.2
\\( HALO^† \\)[8]94.574.583.955.528.661.236.052.059.045.570.864.561.281.564.018.290.852.344.260.0
\\( CoDA^† \\)[12]95.374.285.157.329.161.539.853.458.745.372.666.162.283.065.719.390.553.245.361.4
\\( WA^2Net (ours) \\)95.777.986.759.631.263.451.955.760.947.276.367.963.184.569.818.191.954.846.162.8
", + "bbox": [ + 98, + 128, + 903, + 296 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a139027fda751d9b3b1062f89b17f3ac752b11cf3361acd094660e047642e9ee.jpg", + "image_caption": [ + "Figure 3. Qualitative results of our $\\mathrm{WA}^2\\mathrm{Net}$ and MIC [18] under adverse conditions on ACDC [45], NightCity [49] and DarkZurich [44]." + ], + "image_footnote": [], + "bbox": [ + 91, + 310, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a22e947c7ff632d946fd8d851888fb2061f7a0650b6a780aa9e80ebb07849b40.jpg", + "table_caption": [ + "Table 4. Ablation on main components of our network." + ], + "table_footnote": [], + "table_body": "
Amplitude Prompt AggregationWeather Heterogeneity AdaptationACDCNightCity
Charac. ExtractionPrompt Agg.Prompt Modu.Homo. Attention
XXXX68.652.4
✓XX✓70.753.7
✓X✓✓71.254.9
✓✓✓✓73.856.5
", + "bbox": [ + 96, + 635, + 480, + 718 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "aware prompt aggregation. To evaluate the effectiveness of the amplitude prompt aggregation, we conduct ablation studies as shown in Table 4. As shown in the $3^{rd}$ row in Table 4, not using prompt aggregation alone means that the prompt bank does not undergo knowledge learning of amplitude characteristics but instead self-initialized and only updated during network training. Without aggregating amplitude characteristic information, prompts cannot perceive weather-aware knowledge, failing to effectively modulate features in the prompt-guided modulation with a significant", + "bbox": [ + 89, + 750, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/d3bb0d37b9897b9c6539fd0e44b8aabd104556c36ec121ef3e6c2c821ac193e2.jpg", + "image_caption": [ + "Figure 4. Qualitative results on the daytime $\\rightarrow$ nighttime setting." + ], + "image_footnote": [], + "bbox": [ + 516, + 609, + 903, + 806 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "performance decline.", + "bbox": [ + 513, + 867, + 656, + 878 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Effectiveness of the Weather Heterogeneity Adaptation.", + "bbox": [ + 513, + 885, + 903, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "13958", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b3b6e63ea06dcad60b26e5e96898d8178c07ae106e9c56622edd1175756f10e0.jpg", + "image_caption": [ + "Figure 5. Ablation experiments on super-parameters: $\\lambda_{div}$ and $N$ ." + ], + "image_footnote": [], + "bbox": [ + 93, + 88, + 279, + 184 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c794cbebdacf63ba504da70ac30668cd0fc793e71fc500ec69ebd066441fbad.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 88, + 480, + 184 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/b331d193ae5e6e82bc7897e1812e75b860d75fd3e0625ba37ca1501a848fd00c.jpg", + "table_caption": [ + "Table 5. Performance w/ and w/o the diversity loss ${\\mathcal{L}}_{div}$ ." + ], + "table_footnote": [], + "table_body": "
Diversity LossACDC [45]NightCity [49]
w/o L_divw/ L_div
✓72.456.0
✓73.856.5
", + "bbox": [ + 135, + 250, + 437, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "The weather heterogeneity adaptation includes the prompt-guided modulation and homogenization attention. We assess the impact of the weather heterogeneity adaptation by disabling this process and observing the changes in segmentation accuracy. Simultaneously discarding both prompt aggregation and modulation, as shown in the $2^{nd}$ row of Table 4 (i.e., abandoning the generation and use of prompts), means that the extracted amplitude features are directly fed into the homogenization attention mechanism. This approach results in $f_{a}^{\\theta}$ being unable to perceive multi-weather scene characteristics, adversely affecting the model's generalization in extreme scenarios.", + "bbox": [ + 88, + 340, + 482, + 521 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of the Homogenization Attention. We also analyze the effectiveness of our designed homogenization attention with comparisons against other attention mechanisms as shown in Table 6. The homogenization attention mechanism is designed to address the significant variability in weather conditions by dynamically aligning features from different weather domains. Unlike crossattention, which focuses on explicit interactions between tokens across feature spaces, and masked attention, which selectively attends to spatial regions, our homogenization attention incorporates a domain adaptation strategy that directly mitigates the feature discrepancies caused by adverse weather conditions. As can be observed, the heterogeneity adaptation mechanism proved to be essential to dynamically adapt to varying weather patterns, thereby improving overall performance and reliability. Besides, Figure 6 displays activation maps of different classes with and without weather-aware aggregation and adaptation, illustrating the impact of our design on class-specific segmentation quality.", + "bbox": [ + 91, + 523, + 482, + 810 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Analysis of Hyperparameters. We conducted a comprehensive analysis of hyperparameters to understand their influence on the model's performance. The impact of the choice of diversity loss $\\lambda_{div}$ and the prompt number $N$ is shown in Figure 5. Too few prompts hinder the ability to capture rich scene information, while too many may capture", + "bbox": [ + 89, + 810, + 482, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/bd7dd7d5c66d460fe34e30a6074bb875c83c1e53e6ee5feb2ed7aca241cfaa98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 514, + 89, + 903, + 145 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5d1f1ad9ec2f931d833e7e0b94d7373a07315dd06483ebb5df0164856ec05ec8.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 146, + 903, + 199 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/4d149bd444e680be6c14d67f2e65d1b72417dc843867d4116f4c65ddd036e7f2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 516, + 200, + 903, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8fe8b14f43974e59a52c5762f3295bd2d55378d75c7d85cc650e75583bc552aa.jpg", + "image_caption": [ + "Figure 6. Comparison of activation maps for different classes with (w/) and without(w/o) weather-aware aggregation and adaptation." + ], + "image_footnote": [], + "bbox": [ + 516, + 256, + 903, + 310 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/dd323b684dacd469efcd3daa9ecc21de2f12f6e5c780fe4e7341edeab4efb3e0.jpg", + "table_caption": [ + "Table 6. Comparison of different attention mechanisms with our homogenization attention." + ], + "table_footnote": [], + "table_body": "
Attention MechanismACDC [45]NightCity [49]
cross-attention [52]70.954.1
masked attention [5]71.855.0
Homogenization Attention (ours)73.856.5
", + "bbox": [ + 521, + 406, + 901, + 470 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "irrelevant noise information. It is worth noting that the improvement brought by the diversity loss in Table 5 is more significant when ACDC [45] is the target domain. This is due to ACDC containing more weather diversities (night, rain, snow and fog) compared to the night-only NightCity [49], prompts on ACDC can aggregate more diverse knowledge with $\\mathcal{L}_{div}$ . Therefore, the diversity in prompt combinations contributes to diverse scene analysis under multiple weather conditions.", + "bbox": [ + 511, + 498, + 906, + 633 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 650, + 633, + 666 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we propose a weather-aware aggregation and adaptation network (WA $^2$ Net) for semantic segmentation under adverse conditions. We design the amplitude prompt aggregation to acquire weather-aware characteristic knowledge and weather heterogeneity adaptation to realize interweather homogenization for effective parsing. Extensive experiments demonstrate our effectiveness.", + "bbox": [ + 511, + 676, + 906, + 782 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 511, + 799, + 671, + 815 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "This work was partially supported by the National Key Research and Development Program of China (No.2024YFB3909902), National Nature Science Foundation of China (62121002), Youth Innovation Promotion Association of CAS.", + "bbox": [ + 511, + 825, + 905, + 898 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "13959", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 90, + 187, + 104 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2", + "[2] Yang Chen, Liang Shan, Yi Qu, Weilong Zhang, and Lu Chang. Amsc: Adaptive masking and structure-constraint learning for domain adaptive semantic segmentation under adverse conditions. IEEE Signal Processing Letters, 2023. 2", + "[3] Yujia Chen, Wangkai Li, Zhaoyang Li, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Sam-glomeruli: Enhanced segment anything model for precise glomeruli segmentation. In International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis, pages 182-191. Springer, 2024. 2", + "[4] Yujia Chen, Rui Sun, Wangkai Li, Huayu Mai, Naisong Luo, Yuwen Pan, and Tianzhu Zhang. Alleviate and mining: Rethinking unsupervised domain adaptation for mitochondria segmentation from pseudo-label perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2339-2347, 2025. 2", + "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 8", + "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 7", + "[7] Ozgur Erkent and Christian Laugier. Semantic segmentation with unsupervised domain adaptation under varying weather conditions for autonomous vehicles. IEEE Robotics and Automation Letters, 5(2):3580-3587, 2020. 1", + "[8] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. In International Conference on Machine Learning, pages 13864-13884. PMLR, 2024. 6, 7", + "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3", + "[10] Rui Gong, Qin Wang, Martin Danelljan, Dengxin Dai, and Luc Van Gool. Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7225-7235, 2023. 2", + "[11] Ziyang Gong, Fuhao Li, Yupeng Deng, Wenjun Shen, Xi-anzheng Ma, Zhenming Ji, and Nan Xia. Train one, generalize to all: Generalizable semantic segmentation from single-scene to all adverse scenes. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2275-2284, 2023. 2" + ], + "bbox": [ + 93, + 114, + 483, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] Ziyang Gong, Fuhao Li, Yupeng Deng, Deblina Bhattacharjee, Xianzheng Ma, Xiangwei Zhu, and Zhenming Ji. Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning. In European Conference on Computer Vision, pages 130-148. Springer, 2024. 6, 7", + "[13] Monica Gruosso, Nicola Capecce, and Ugo Erra. Human segmentation in surveillance video with deep learning. *Multimedia Tools and Applications*, 80(1):1175-1199, 2021. 1", + "[14] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. Ppt: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, 2022. 3", + "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6", + "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2, 3, 5, 6, 7", + "[17] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision, pages 372-391. Springer, 2022. 6, 7", + "[18] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721-11732, 2023. 2, 3, 5, 6, 7", + "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4", + "[20] Abdulrahman Kerim, Felipe Chamone, Washington Ramos, Leandro Soriano Marcolino, Erickson R Nascimento, and Richard Jiang. Semantic segmentation under adverse conditions: a weather and nighttime-aware synthetic data-based approach. arXiv preprint arXiv:2210.05626, 2022. 2, 3, 5, 6", + "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3", + "[22] Sohyun Lee, Namyup Kim, Sungyeon Kim, and Suha Kwak. Frest: Feature restoration for semantic segmentation under multiple adverse conditions. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 1", + "[23] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2", + "[24] Wangkai Li, Rui Sun, Bohao Liao, Zhaoyang Li, and Tianzhu Zhang. Balanced learning for domain adaptive semantic segmentation. In *Forty-second International Conference on Machine Learning*. 3" + ], + "bbox": [ + 93, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "13960", + "bbox": [ + 480, + 945, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3", + "[26] Zhi Li, Xing Wu, Jianjia Wang, and Yike Guo. Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing, 127: 104554, 2022. 1", + "[27] Zhaoyang Li, Yuan Wang, Wangkai Li, Rui Sun, and Tianzhu Zhang. Localization and expansion: A decoupled framework for point cloud few-shot semantic segmentation. In European Conference on Computer Vision, pages 18-34. Springer, 2024. 2", + "[28] Zhaoyang Li, Yuan Wang, Wangkai Li, Tianzhu Zhang, and Xiang Liu. Dual-agent optimization framework for cross-domain few-shot segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9849-9859, 2025. 2", + "[29] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 3", + "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1", + "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6", + "[32] Ning Lv, Zenghui Zhang, Cong Li, Jiaxuan Deng, Tao Su, Chen Chen, and Yang Zhou. A hybrid-attention semantic segmentation network for remote sensing interpretation in land-use surveillance. International Journal of Machine Learning and Cybernetics, 14(2):395–406, 2023. 1", + "[33] Claudia Menzel, Gregor U Hayn-Leichsenring, Oliver Langner, Holger Wiese, and Christoph Redies. Fourier power spectrum characteristics of face photographs: attractiveness perception depends on low-level image properties. PLoS One, 10(4):e0122801, 2015. 4", + "[34] Khan Muhammad, Tanveer Hussain, Hayat Ullah, Javier Del Ser, Mahdi Rezaei, Neeraj Kumar, Mohammad Hijji, Paolo Bellavista, and Victor Hugo C de Albuquerque. Vision-based semantic segmentation in scene understanding for autonomous driving: Recent achievements, challenges, and outlooks. IEEE Transactions on Intelligent Transportation Systems, 23(12):22694-22715, 2022. 1", + "[35] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1520-1528, 2015. 1", + "[36] Yuwen Pan, Naisong Luo, Rui Sun, Meng Meng, Tianzhu Zhang, Zhiwei Xiong, and Yongdong Zhang. Adaptive template transformer for mitochondria segmentation in electron microscopy images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21474-21484, 2023. 1", + "[37] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase" + ], + "bbox": [ + 91, + 92, + 483, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "enhancement for night-time semantic segmentation. In European Conference on Computer Vision, pages 408-424. Springer, 2024. 1", + "[38] Yuwen Pan, Rui Sun, Yuan Wang, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Purify then guide: A bidirectional bridge network for open-vocabulary semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1", + "[39] Yuwen Pan, Rui Sun, Yuan Wang, Tianzhu Zhang, and Yongdong Zhang. Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2031-2040, 2024. 2", + "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6", + "[41] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In 2019 22th International Conference on Information Fusion (FUSION), pages 1-8. IEEE, 2019. 1", + "[42] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of fast video-sequence segmentation. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020.", + "[43] Nikhil Reddy, Abhinav Singhal, Abhishek Kumar, Mahsa Baktashmotlagh, and Chetan Arora. Master of all: Simultaneous generalization of urban-scene segmentation to all adverse weather conditions. In European Conference on Computer Vision, pages 51-69. Springer, 2022. 2", + "[44] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3139-3153, 2020. 2, 5, 6, 7", + "[45] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1, 3, 5, 6, 7, 8", + "[46] Rui Sun, Naisong Luo, Yuan Wang, Yuwen Pan, Huayu Mai, Zhe Zhang, and Tianzhu Zhang. 1st place solution for youtuves challenge 2022: Video object segmentation. In Proceedings of the Conference on Computer Vision and Pattern Recognition Workshops, 2022. 3", + "[47] Rui Sun, Naisong Luo, Yuwen Pan, Huayu Mai, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Appearance prompt vision transformer for connectome reconstruction. In IJCAI, pages 1423-1431, 2023. 1", + "[48] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: exploring the better weighting function for semi-supervised semantic segmentation. Advances in neural information processing systems, 36:61792-61805, 2023. 2", + "[49] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson WH Lau. Night-time scene parsing with a large" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "13961", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "real dataset. IEEE Transactions on Image Processing, 30: 9085-9098, 2021. 5, 6, 7, 8", + "[50] Yingjie Tian and Siyu Zhu. Partial domain adaptation on semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3798-3809, 2021. 2", + "[51] Abhinav Valada, Johan Vertens, Ankit Dhall, and Wolfram Burgard. Adapnet: Adaptive semantic segmentation in adverse environmental conditions. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4644-4651. IEEE, 2017. 2", + "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 8", + "[53] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 6, 7", + "[54] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2, 3, 6, 7", + "[55] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6", + "[56] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5", + "[57] Xin Yang, Wending Yan, Yuan Yuan, Michael Bi Mi, and Robby T Tan. Semantic segmentation in multiple adverse weather conditions with domain knowledge retention. arXiv preprint arXiv:2401.07459, 2024. 3", + "[58] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 2, 4", + "[59] Chunhui Zhang, Guanjie Huang, Li Liu, Shan Huang, Yinan Yang, Xiang Wan, Shiming Ge, and Dacheng Tao. Webuav-3m: A benchmark for unveiling the power of million-scale deep uav tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9186-9205, 2022. 1", + "[60] Chunhui Zhang, Li Liu, Guanjie Huang, Hao Wen, Xi Zhou, and Yanfeng Wang. Webuot-1m: Advancing deep underwater object tracking with a million-scale benchmark. Advances in Neural Information Processing Systems, 37:50152-50167, 2024. 1", + "[61] Chunhui Zhang, Li Liu, Hao Wen, Xi Zhou, and Yanfeng Wang. Mambatrack: Exploiting dual-enhancement for night uav tracking. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3" + ], + "bbox": [ + 91, + 90, + 482, + 869 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "13962", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_model.json b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_model.json new file mode 100644 index 0000000000000000000000000000000000000000..9c4807d899fcb8935f01ad8bddd04f0b8d15bcd5 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_model.json @@ -0,0 +1,2323 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.204, + 0.131, + 0.797, + 0.177 + ], + "angle": 0, + "content": "Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions" + }, + { + "type": "text", + "bbox": [ + 0.258, + 0.203, + 0.722, + 0.221 + ], + "angle": 0, + "content": "Yuwen Pan\\(^{1}\\) Rui Sun\\(^{1}\\) Wangkai Li\\(^{1}\\) Tianzhu Zhang\\(^{1,2*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.307, + 0.222, + 0.692, + 0.239 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.14, + 0.24, + 0.856, + 0.258 + ], + "angle": 0, + "content": "\\(^{2}\\)National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.26, + 0.775, + 0.275 + ], + "angle": 0, + "content": "{panyw, issunrui, lwklwk}@mail.ustc.edu.cn, tzzhang@ustc.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.309, + 0.327, + 0.325 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.342, + 0.486, + 0.78 + ], + "angle": 0, + "content": "Semantic segmentation under adverse conditions is critical for reliable visual perception in challenging weather environments. These extreme scenarios introduce distortions, such as low contrast and reduced visibility, making traditional segmentation models struggle. The scarcity of labeled data in such conditions makes it difficult to train models directly for these environments. Unsupervised domain adaptation (UDA) has been proposed as a solution to transfer knowledge from labeled source domains (normal weather) to unlabeled target domains (adverse weather). However, existing methods face significant challenges, particularly due to weather unawareness and feature heterogeneity. Many models fail to account for the unique characteristics of different weather conditions, and the significant feature discrepancies between normal and adverse weather images hinder effective adaptation. In this paper, we propose a novel weather-aware aggregation and adaptation network that leverages characteristic knowledge to achieve weather homogenization and enhance scene perception. Specifically, we introduce amplitude prompt aggregation to capture essential characteristics from the Fourier frequency domain that are indicative of different weather conditions. Additionally, we employ weather heterogeneity adaptation to mitigate the inter-domain heterogeneity, thereby achieving feature homogenization across diverse environments. Extensive experimental results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.808, + 0.222, + 0.823 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.833, + 0.485, + 0.879 + ], + "angle": 0, + "content": "Semantic segmentation involves assigning a class label to each pixel in an image [30, 35, 36, 38, 47], which plays a crucial role in various real-world applications such as au" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.308, + 0.878, + 0.435 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.612, + 0.435, + 0.81, + 0.446 + ], + "angle": 0, + "content": "(a) Traditional Cross-weather Paradigm" + }, + { + "type": "image", + "bbox": [ + 0.546, + 0.447, + 0.877, + 0.595 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.576, + 0.596, + 0.847, + 0.608 + ], + "angle": 0, + "content": "(b) Weather-aware Aggregation and Adaptation (ours)" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.619, + 0.908, + 0.689 + ], + "angle": 0, + "content": "Figure 1. Comparison of the traditional cross-weather paradigm and our weather-aware aggregation and adaptation. To mitigate the existing issues (unawareness of weather characteristics and severe cross-weather heterogeneity), we aim to aggregate weather-aware attribute knowledge and homogenize domain discrepancies." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.718, + 0.907, + 0.869 + ], + "angle": 0, + "content": "tonomous driving [34, 37, 59] and surveillance [13, 32, 60]. However, the performance of segmentation models often degrades significantly when faced with adverse weather conditions [22, 45], including night, fog, rain, and snow. The deterioration in visibility and scene appearance poses a substantial challenge for traditional segmentation algorithms, which are typically trained on datasets collected under normal weather conditions. Therefore, exploring an effective way for semantic parsing under extreme weather conditions is highly anticipated." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Extensive research efforts have been devoted to semantic segmentation in adverse weather conditions [7, 26, 41-" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.236, + 0.9 + ], + "angle": 0, + "content": "*Corresponding author." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.482, + 0.348 + ], + "angle": 0, + "content": "43, 51]. Due to the scarcity of labeled data in such adverse conditions, existing methods strive to employ an unsupervised domain adaptation (UDA) paradigm utilizing both labeled normal weather data and images without annotations under extreme scenarios to address the challenge [16, 44, 50, 53, 54]. Kerim et al. [20] propose a dual-branch network based on DeepLabV3+ [1], enhancing training knowledge for extreme weather conditions by incorporating a synthetic multi-scene dataset. However, discerning similar-looking categories in the target domain with limited real annotations remains a challenge for domain adaptation methods. To address this, MIC [18] introduces the masked image consistency, leveraging spatial context relations in the target domain as additional cues to enhance the robustness of UDA. The emergence of these methods has benefited the understanding of scenes in extreme weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.352, + 0.487, + 0.867 + ], + "angle": 0, + "content": "Despite considerable progress, the task of distinguishing and analyzing scenes under different extreme weather conditions remains challenging. Whether through the assistance of synthetic data or exploration of spatial context clues, existing methods struggle with accurately identifying weather-aware semantics in multiple extreme scenarios. We attribute the predicament of this cross-weather paradigm to two inherent issues (Figure 1): (1) Unawareness of Weather Characteristics. The style characteristics under adverse weathers, such as illumination conditions and contrast, are primary culprits responsible for the inter-weather gap among diverse scenes. Although recognizing this crucial aspect, existing works [10, 58] generally employ low-level enhancements or implicit learning for inter-domain style transfer, failing to fully exploit weather attribute representation. Therefore, designing a strategy to aggregate weather-aware scene knowledge across diverse conditions is worthy of exploration. (2) Severe Cross-Weather Heterogeneity. Early works [16, 18] tend to scene images from different weather conditions uniformly, ignoring the substantial differences across different weather types. The significant heterogeneity across weather conditions presents another challenge, stemming from the diverse and complex nature of weather phenomena that introduce distinct visual distortions and scene alterations. Even though some methods [2, 11] attempt to address this challenge, the significant differences in characteristics across different weather conditions still pose considerable difficulties. These variations lead to a significant disparity between features learned from the normal weather and those required for effective interpretation in adverse scenes. Addressing how to leverage characteristic knowledge to achieve weather homogenization to facilitate scene perception in adverse scenes is an essential question." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.871, + 0.485, + 0.903 + ], + "angle": 0, + "content": "Motivated by the above discussions, we propose a Weather-aware Aggregation and Adaptation NETwork" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.908, + 0.543 + ], + "angle": 0, + "content": "(WA\\(^2\\)Net) to aggregate weather-aware attribute knowledge and homogenize domain discrepancies, including amplitude prompt aggregation and weather heterogeneity adaptation. In the amplitude prompt aggregation, to acquire weather-aware characteristic knowledge, we utilize the amplitude operation of Fourier transform to capture essential characteristics indicative of different weather conditions. These amplitude features serve as discriminative cues for weather-related variations in the scene. To efficiently aggregate the crucial characteristic knowledge from the frequency domain into weather-adaptive prompts, we prepend a set of learnable prompts in the prompt bank to interact with frequency features by cross-attention mechanism to obtain weather-aware prompts. These prompts encapsulate the intrinsic characteristics of various weather conditions, providing essential guidance for the adaptation process. In the weather heterogeneity adaptation, to realize inter-weather homogenization for effective semantic parsing, the resultant weather-aware prompts are leveraged to generate adapted characteristics after modulation. Then, we propose a novel homogenization attention mechanism to acquire homogenized features across diverse weather conditions. This mechanism identifies the discrepancy between features of the original image and adapted characteristics, decoupling weather characteristics from valuable target information and thereby filtering out attribute interference. In this way, our WA\\(^2\\)Net achieves effective weather-aware knowledge aggregation and robust adaptation, ensuring improved performance for semantic segmentation in adverse weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.546, + 0.907, + 0.574 + ], + "angle": 0, + "content": "To sum up, our contributions can be summarized as follows:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.578, + 0.905, + 0.653 + ], + "angle": 0, + "content": "- We propose a novel weather-aware aggregation and adaptation network for semantic segmentation under adverse conditions. Our paper exploits a practical way to acquire weather-adaptive knowledge and enhance cross-weather perception via an aggregation and adaptation manner." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.654, + 0.905, + 0.728 + ], + "angle": 0, + "content": "- We design the amplitude prompt aggregation to acquire weather-aware prompts from weather characteristics through amplitude extraction, and the weather heterogeneity adaptation to inherit weather-aware knowledge and mitigate the inter-weather heterogeneity." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.729, + 0.905, + 0.787 + ], + "angle": 0, + "content": "- Extensive experiment results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.578, + 0.905, + 0.787 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.806, + 0.655, + 0.822 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.905, + 0.863 + ], + "angle": 0, + "content": "2.1. Semantic Segmentation under Adverse Conditions" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.902 + ], + "angle": 0, + "content": "The field of semantic segmentation [3, 4, 23, 27, 28, 39, 48] has made significant advancements with the emergence" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13953" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.501 + ], + "angle": 0, + "content": "of convolutional neural networks (CNNs) [21] and transformers [52]. However, traditional semantic segmentation models experience a notable performance drop under adverse conditions such as rain, fog and night-time environments [45, 46, 61]. Therefore, semantic segmentation under adverse conditions aims to develop robust models capable of accurately delineating objects and regions in challenging environments. Upon the scarcity of labeled data in extreme weathers, the unsupervised domain adaptation paradigm is utilized to enhance the generalization of models across different weather conditions with the aid of normal weather data [24, 54]. Kerim et al. [20] propose a dual-branch network and expand the training dataset for extreme weather conditions by integrating a synthetic multiscene dataset. Nevertheless, domain adaptation methods encounter challenges due to the difficulty in distinguishing similar-looking categories in the target domain with limited real annotations. To tackle this issue, MIC [18] introduces Masked Image Consistency, utilizing spatial context relations in the target domain as additional cues to enhance vision perception. Additionally, to mitigate knowledge forgetting in new scenes, Yang et al. [57] introduce adaptive knowledge acquisition and replay of weather-condition combinations, aiming to alleviate model-forgetting issues in diverse scenes. Differently, we strive to tackle the challenge of weather-unaware perception, leveraging weather-aware prompts in an aggregation and adaptation manner." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.511, + 0.265, + 0.527 + ], + "angle": 0, + "content": "2.2. Prompt Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.534, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Prompting [29] incorporates task-specific instructions into input text to guide model behavior, enhancing performance without retraining, particularly in few-shot learning [9]. Recent work extends prompting to vision tasks, where learnable prompts enable adaptation without tuning all parameters [25] and improve segmentation generalization [14]. For semantic segmentation under adverse weather, weather-adaptive prompts modulate visual embeddings to mitigate distortions like lighting and visibility changes, enhancing robustness without extensive fine-tuning. Integrating weather-specific knowledge into prompts improves generalization across diverse conditions, making models more effective in real-world applications." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.747, + 0.183, + 0.763 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.772, + 0.233, + 0.787 + ], + "angle": 0, + "content": "3.1. Preliminaries" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Typically, in the field of unsupervised domain adaptation [16, 18], the source domain data can be formulated as \\( D^{s} = \\{X^{s}, Y^{s}\\} \\), where \\( X^{s} \\) and \\( Y^{s} \\) represent the corresponding image and label space, respectively. Each source domain image \\( x_{k}^{s} \\in X^{s} \\) is associated with a ground truth segmentation mask \\( y_{k}^{s} \\in Y^{s} \\). Similarly, let \\( X^{t} \\) be the target domain image space, and \\( Y^{t} \\) be the label space. However," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.138 + ], + "angle": 0, + "content": "unlike the source domain, the target domain has no access to annotated segmentation masks. The supervised loss \\(\\mathcal{L}^s\\) can be formulated with a neural network \\(f_{\\mu}\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.599, + 0.151, + 0.907, + 0.193 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {s} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {s}\\right), y _ {k} ^ {s}\\right), \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.204, + 0.906, + 0.235 + ], + "angle": 0, + "content": "where \\(N_{s}\\) denotes the image number of the source domain, \\(\\mathcal{L}_{ce}\\) denotes the pixel-wise cross-entropy loss as:" + }, + { + "type": "equation", + "bbox": [ + 0.568, + 0.246, + 0.907, + 0.289 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c e} (\\hat {y}, y) = - \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {C} y _ {h w c} \\log \\hat {y} _ {h w c}. \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.301, + 0.907, + 0.438 + ], + "angle": 0, + "content": "The model trained on the source domain \\(D^{s}\\) normally exhibits performance degradation when tested on the target domain \\(D^{t}\\). Therefore, the self-training paradigm generates pseudo-labels for the target domain images \\(X^{t}\\) with a teacher network \\(g_{\\phi}\\) instead of the student network \\(f_{\\mu}\\). Specifically, the model predicts segmentation masks for unlabeled target domain images, producing pseudo-labels \\(\\hat{y}_k^t\\) for each image \\(x_k^t\\) in the target domain. The unsupervised loss \\(\\mathcal{L}^t\\) for the target domain is formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.591, + 0.45, + 0.906, + 0.492 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} ^ {t} = \\frac {1}{N _ {t}} \\sum_ {k = 1} ^ {N _ {t}} \\lambda^ {t} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {t}\\right), \\hat {y} _ {k} ^ {t}\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.503, + 0.906, + 0.61 + ], + "angle": 0, + "content": "where \\(N_{t}\\) denotes the image number of the target domain, \\(\\lambda^t\\) is a weight coefficient for the unsupervised loss. The student network updates the parameters of the teacher network through backpropagation, utilizing the teacher network's output as targets to minimize prediction errors, while also employing Exponential Moving Average (EMA) update to stabilize the training process and improve convergence." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.62, + 0.625, + 0.635 + ], + "angle": 0, + "content": "3.2. Overview" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.902 + ], + "angle": 0, + "content": "As shown in Figure 2, given an input image \\( x^{s} \\) from the source domain and \\( x^{t} \\) from the target domain under adverse conditions, we denote the image under different weather condition \\( \\theta \\) as \\( x^{\\theta} \\) for clarity, where \\( \\theta \\in (0, \\dots, \\Theta) \\), \\( \\Theta \\) denotes the total number of adverse weather conditions (\\( x^{0} \\) refers to the source domain image under the normal weather condition). The Fourier amplitude features \\( f_{a}^{\\theta} \\) are decoupled from the frequency domain. Then, a set of prompts, i.e., prompt bank \\( \\mathbf{P} \\), is learned during the prompt aggregation process. The acquired weather-aware prompts \\( \\hat{\\mathbf{P}} \\) further interact with amplitude features \\( f_{a}^{\\theta} \\) to evolve into adapted characteristic \\( \\hat{f}_{a}^{\\theta} \\). Then, to achieve homogenization for features from diverse weather conditions, we further employ homogenization attention to mitigate the influence of weather heterogeneity. Finally, the modulated features \\( f_{h}^{\\theta} \\) are fed into the teacher-student network for comprehensive training." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.095, + 0.089, + 0.91, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.268, + 0.908, + 0.297 + ], + "angle": 0, + "content": "Figure 2. Illustration of our proposed \\(\\mathrm{WA}^2\\mathrm{Net}\\). Our network contains the amplitude prompt aggregation to acquire weather-aware prompts (Section 3.3) and weather heterogeneity adaptation to realize feature homogenization (Section 3.4)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.323, + 0.373, + 0.339 + ], + "angle": 0, + "content": "3.3. Amplitude Prompt Aggregation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.344, + 0.484, + 0.45 + ], + "angle": 0, + "content": "In order to disentangle the latent weather characteristics and target information fused in the RGB images, we first model the representation of weather characteristics by exploring the Fourier frequency domain. Then, we collect the weather characteristic properties in Fourier amplitude features and acquire weather-aware prompts during the prompt aggregation process." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.451, + 0.484, + 0.617 + ], + "angle": 0, + "content": "Amplitude Characteristic Extraction. In the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information [33, 58], which reflects the energy distribution of the frequency components of the image and can be understood as a representation of the \"characteristics\" or \"style\" of the image. Therefore, we can utilize the amplitude of Fourier spectrum to extract weather characteristic knowledge in diverse scenes. Specifically, in the amplitude extractor, given an image \\( x^{\\theta} \\in \\mathbb{R}^{H \\times W \\times 3} \\), we apply a two-dimensional Fourier transform \\( \\mathcal{F}(x) \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.142, + 0.634, + 0.483, + 0.677 + ], + "angle": 0, + "content": "\\[\n\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x _ {h, w} ^ {\\theta} e ^ {- J 2 \\pi \\left(\\frac {w h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.682, + 0.483, + 0.711 + ], + "angle": 0, + "content": "where \\(J\\) refers to the imaginary unit. Then we can acquire the corresponding amplitude \\(\\mathcal{A}\\) and phase \\(\\Phi\\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.202, + 0.717, + 0.482, + 0.737 + ], + "angle": 0, + "content": "\\[\n\\mathcal {A} \\left(x ^ {\\theta}\\right) _ {u, v} = \\left| \\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} \\right|, \\tag {5}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.741, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\Phi \\left(x ^ {\\theta}\\right) _ {u, v} = \\arg \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v}\\right) = \\arctan \\left[ \\frac {\\operatorname {I m} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)}{\\operatorname {R e} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)} \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.789, + 0.483, + 0.879 + ], + "angle": 0, + "content": "where \\( \\arg \\) represents the argument (or phase angle) of the complex number, Im and Re represent the imaginary and real part of \\( \\mathcal{F}(x) \\), respectively. In this way, to finally generate our amplitude characteristic map, we fix the phase to an average constant \\( c^p \\) and apply inverse Fourier transform to acquire the amplitude characteristic map as" + }, + { + "type": "equation", + "bbox": [ + 0.179, + 0.884, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\n\\overline {{\\mathcal {A}}} (x ^ {\\theta}) = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x ^ {\\theta}) _ {u, v} e ^ {- J c ^ {p}} ], \\tag {7}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.323, + 0.907, + 0.399 + ], + "angle": 0, + "content": "where \\(\\mathcal{F}^{-1}\\) denotes the inverse Fourier transformation. Then, we can obtain the amplitude characteristics \\(\\mathbf{F}_a = \\{f_a^\\theta\\}_{\\theta=0}^\\Theta\\) through a shared-weight lightweight encoder (e.g., ResNet-18 [15]), preparing for the following aggregation of weather-aware prompts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.907, + 0.598 + ], + "angle": 0, + "content": "Weather-aware Prompt Aggregation. To effectively aggregate characteristic information across different weather conditions, we aim to develop a medium capable of absorbing cross-weather knowledge. Inspired by [19], we introduce a set of prompts, i.e., prompt bank \\(\\mathbf{P} = \\{\\pmb{p}_n\\}_{n=1}^N\\), where \\(\\pmb{p}_n \\in \\mathbb{R}^{1 \\times L}\\) and \\(N\\) denotes the number of prototypes. These learnable prompts will be leveraged to extract characteristic knowledge by interaction with amplitude features \\(f_a^\\theta\\), which highly represents the style condition of the current scene. In specific, for the acquired amplitude characteristic \\(f_a^\\theta\\), the queries arise from the prompts \\(\\mathbf{P}\\), and keys and values arise from the amplitude feature \\(\\mathbf{F}_a = \\{f_a^0, \\dots, f_a^\\theta, \\dots, f_a^\\Theta\\}\\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.608, + 0.907, + 0.628 + ], + "angle": 0, + "content": "\\[\n\\mathbf {Q} _ {n} = \\boldsymbol {p} _ {n} \\mathbf {W} ^ {Q}, \\mathbf {K} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {K}, \\mathbf {V} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {V}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.638, + 0.906, + 0.702 + ], + "angle": 0, + "content": "where \\(n\\in [1,\\dots ,N]\\) , and \\(\\mathbf{W}^Q\\in \\mathbb{R}^{C\\times C_k}\\) \\(\\mathbf{W}^K\\in\\) \\(\\mathbb{R}^{C\\times C_k}\\) \\(\\mathbf{W}^V\\in \\mathbb{R}^{C\\times C_v}\\) are linear projections. The attention weights are calculated based on the dot-product similarity between each query and key:" + }, + { + "type": "equation", + "bbox": [ + 0.563, + 0.713, + 0.906, + 0.755 + ], + "angle": 0, + "content": "\\[\ns _ {n, \\theta} = \\frac {\\exp (\\beta_ {n , \\theta})}{\\sum_ {\\theta = 0} ^ {\\Theta} \\exp (\\beta_ {n , \\theta})}, \\beta_ {n, \\theta} = \\frac {\\mathbf {Q} _ {n} \\mathbf {K} ^ {\\theta} {} ^ {\\top}}{\\sqrt {C _ {k}}}, \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.526, + 0.77, + 0.905, + 0.823 + ], + "angle": 0, + "content": "\\[\n\\hat {\\boldsymbol {p}} _ {n} = \\operatorname {F F N} (\\operatorname {A t t} \\left(\\mathbf {Q} _ {n}, \\mathbf {K}, \\mathbf {V}\\right)) = \\operatorname {F F N} \\left(\\sum_ {\\theta = 0} ^ {\\Theta} s _ {n, \\theta} \\mathbf {V} ^ {\\theta}\\right), \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.825, + 0.905, + 0.87 + ], + "angle": 0, + "content": "where \\(\\mathrm{FFN}()\\) denotes the feed forward network. In this way, weather-aware prompts \\(\\hat{\\mathbf{P}} = \\{\\hat{p}_n\\}_{n=1}^N\\) can distill rich characteristic knowledge from the amplitude features \\(\\mathbf{F}_a\\)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Besides, in order to enable prompts to carry diverse and comprehensive knowledge of various weather conditions," + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13955" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.108 + ], + "angle": 0, + "content": "we apply the diversity loss on the prompt bank \\(\\mathbf{P}\\). Formally," + }, + { + "type": "equation", + "bbox": [ + 0.108, + 0.116, + 0.483, + 0.16 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {d i v} = \\frac {1}{N (N - 1)} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1, i \\neq j} ^ {N} \\left(\\frac {\\langle \\boldsymbol {p} _ {i} , \\boldsymbol {p} _ {j} \\rangle}{\\| \\boldsymbol {p} _ {i} \\| _ {2} \\| \\boldsymbol {p} _ {j} \\| _ {2}}\\right). \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.168, + 0.484, + 0.351 + ], + "angle": 0, + "content": "Applying diversity loss to different prompts offers the advantage of encouraging each prompt \\( \\pmb{p}_n \\) to focus on distinct attributes of weather conditions, thereby preventing redundancy in aggregation learning. Diversity loss enhances weather-specific prompts by encouraging each to capture unique traits. Integrating the prompt bank and diversity loss within domain adaptation allows the model to adapt to varied weathers. Splitting this into two tasks is limited by the lack of direct supervision on weather-adaptive prompts; only foreground masks from the source domain are available. Diversity loss thus broadens the prompt bank's coverage of style features for better domain adaptation." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.358, + 0.398, + 0.374 + ], + "angle": 0, + "content": "3.4. Weather Heterogeneity Adaptation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.38, + 0.484, + 0.516 + ], + "angle": 0, + "content": "In adverse weather conditions, weather-induced distortions such as glare, fog, and raindrops introduce complexities that can degrade the performance of vision perception. To address the weather heterogeneity challenges posed by varying weather conditions, we introduce the weather heterogeneity adaptation, which utilizes acquired weather-adaptive prompts \\(\\hat{\\mathbf{P}}\\) as references to modulate the weather characteristic in the current scene and homogenizes weather features via homogenization attention." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.516, + 0.484, + 0.636 + ], + "angle": 0, + "content": "Prompt-guided Modulation. In fact, the introduction of the prompt bank can be deemed as a guidepost in artificially solving incidents, i.e., guiding information, which absorbs weather-aware knowledge across various scenes and can transfer reliable weather-aware properties into modulated characteristic features. For the acquired weather amplitude feature \\( f_{a}^{\\theta} \\), the modulated characteristic can be formulated as:" + }, + { + "type": "equation", + "bbox": [ + 0.198, + 0.643, + 0.483, + 0.679 + ], + "angle": 0, + "content": "\\[\n\\hat {f} _ {a} ^ {\\theta} = \\operatorname {S o f t m a x} \\left(\\frac {f _ {a} ^ {\\theta} \\hat {\\mathbf {P}} ^ {\\top}}{\\sqrt {C}}\\right) \\hat {\\mathbf {P}}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.483, + 0.749 + ], + "angle": 0, + "content": "In this way, the modulated characteristic \\(\\hat{f}_a^\\theta\\) inherits rich cross-weather knowledge from the weather-aware prompts \\(\\hat{\\mathbf{P}}\\), thereby evolving into an adapted representation poised for subsequent homogenization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Homogenization Attention Mechanism. With the acquired representation of modulated characteristic \\(\\hat{f}_a^\\theta\\), we first extract the vision features of image \\(x^{\\theta}\\) from the RGB domain through feature encoding as \\(f_v^\\theta\\). To homogenize vision features from diverse weather conditions, we aim to find an appropriate way to mitigate the influence of \\(\\hat{f}_a^\\theta\\) on \\(f_v^\\theta\\). Inspired by the success of Transformer [52] architecture in discovering local regions, we further explore the potential of attention mechanism in characteristic homogenization for different weather conditions." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.091, + 0.907, + 0.34 + ], + "angle": 0, + "content": "Given an image feature \\( f_{v}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times c} \\) extracted from the vision feature extractor (e.g., MiT-B5 [56]), we use the extracted \\( \\hat{f}_{a}^{\\theta} \\) to remove the weather characteristics through a novel homogenization attention mechanism. First, we employ two convolution layers to map \\( f_{v}^{\\theta} \\) and \\( \\hat{f}_{a}^{\\theta} \\) to the same dimension \\( C \\), then we can obtain \\( \\overline{\\mathbf{f}}_{\\mathbf{v}}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C} \\) and \\( \\overline{f}_{a}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C} \\), respectively. The disparity between image features and weather characteristics signifies the saliency details of the target objects, thereby aiding in accentuating the relevant information about the targets. Rather than vanilla transformer attention mechanism [52], we custom design the homogenization attention to reduce heterogeneity in a finer manner, prompting the vision system to highlight salient regions across diverse scenes. Specifically, we acquire the homogenized map \\( \\mathcal{H} \\in \\mathbb{R}^{h \\times w} \\) from pixel-level features and weather characteristic as:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.344, + 0.907, + 0.387 + ], + "angle": 0, + "content": "\\[\n\\mathcal {H} _ {i, j} ^ {\\theta} = \\sum_ {c = 1} ^ {C} \\left(\\overline {{\\mathbf {f}}} _ {\\mathbf {v} _ {i, j, c}} ^ {\\theta} - \\overline {{f}} _ {a i, j, c} ^ {\\theta}\\right) ^ {2}, \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.39, + 0.906, + 0.452 + ], + "angle": 0, + "content": "where \\(i, j,\\) and \\(c\\) are the index of height, width, and channel, respectively. Finally, we can get the homogenized pixel features \\(\\mathbf{F}^a\\) by weighting the homogenized map \\(\\mathcal{H}\\) to original image features in a pixel-wise manner as:" + }, + { + "type": "equation", + "bbox": [ + 0.66, + 0.456, + 0.907, + 0.475 + ], + "angle": 0, + "content": "\\[\n\\mathbf {f} _ {h} ^ {\\theta} = \\mathbf {f} ^ {\\theta} \\circ \\mathcal {H} ^ {\\theta}, \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.481, + 0.906, + 0.618 + ], + "angle": 0, + "content": "where \\( \\circ \\) denotes the element-wise product. Besides, we use a self-attention layer for further aggregating target information across different pixels. The obtained visual features are fed into the conventional teacher-student network, wherein the images from the source domain are input to the student network for supervised constraints, while the images from the target domain are separately input to both the student and teacher networks to generate predictions and pseudolabels for unsupervised constraints." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.625, + 0.714, + 0.641 + ], + "angle": 0, + "content": "3.5. Training Supervision" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.646, + 0.906, + 0.737 + ], + "angle": 0, + "content": "As illustrated in Section 3.1, the acquired adapted vision features from the source and target are input to the teacher-student network to acquire the supervised loss \\(\\mathcal{L}^s\\) and unsupervised loss \\(\\mathcal{L}^t\\). Together with the diversity loss \\(\\mathcal{L}_{div}\\) in the prompt bank, we can finally obtain the segmentation map as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.607, + 0.743, + 0.905, + 0.761 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} ^ {s} + \\mathcal {L} ^ {t} + \\lambda_ {\\text {d i v}} \\mathcal {L} _ {\\text {d i v}}, \\tag {15}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.768, + 0.744, + 0.783 + ], + "angle": 0, + "content": "where \\(\\lambda_{div}\\) is the trade-off weight." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.794, + 0.64, + 0.811 + ], + "angle": 0, + "content": "4. Experiment" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.819, + 0.691, + 0.835 + ], + "angle": 0, + "content": "4.1. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Datasets. In our experiments, following the conventional practice [16, 18, 20], we utilize 4 datasets including CityScapes [6], ACDC [45], NightCity [49] and DarkZurich [44]. For the source domain, we use the training" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.958 + ], + "angle": 0, + "content": "13956" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.119 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons of semantic segmentation performance under adverse conditions on CityScapes [6]→ACDC [45], where the best performance is marked as **bold** and the second best is **underlined**. † denotes the performance of our reproduction." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.13, + 0.905, + 0.31 + ], + "angle": 0, + "content": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]72.914.340.516.621.29.317.421.263.823.818.332.619.569.536.234.546.226.936.132.7
\\(MGCDA^†\\) [44]73.428.769.919.326.336.853.053.375.432.084.651.026.177.643.245.953.932.741.548.7
\\(DANNet^†\\) [54]84.354.277.638.030.018.941.635.271.339.486.648.729.276.241.643.058.632.643.950.0
DAFormer [16]58.451.384.042.735.150.730.057.074.852.851.358.332.682.758.354.982.444.150.755.4
\\(SSAC^†\\) [20]76.559.882.752.634.551.258.161.263.958.167.261.540.081.859.261.783.447.555.659.6
HRDA [17]88.357.988.155.236.756.362.965.374.257.785.968.845.788.576.482.487.752.760.468.0
MIC [18]90.867.189.254.540.557.262.068.476.361.887.071.349.489.775.786.889.156.963.070.4
HALO [8]94.279.888.260.251.164.178.265.687.955.795.566.320.788.982.289.387.950.459.071.9
CoDA [12]93.172.790.757.347.456.869.970.087.359.895.471.447.690.377.183.889.154.764.172.6
\\(WA^2Net (ours)\\)94.577.191.661.947.657.568.570.479.264.095.672.850.290.575.588.790.957.562.673.8
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.324, + 0.907, + 0.355 + ], + "angle": 0, + "content": "Table 2. Quantitative comparisons of performance on CityScapes [6] \\(\\rightarrow\\) NightCity [49]. Considering the NightCity [49] dataset is newly introduced, we reproduce the results of state-of-the-art methods for better comparison." + }, + { + "type": "table", + "bbox": [ + 0.1, + 0.364, + 0.903, + 0.519 + ], + "angle": 0, + "content": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
DANNet [54]57.536.452.342.539.533.227.846.148.918.467.049.528.357.244.050.443.229.137.444.0
DAFormer [16]64.241.059.647.043.235.731.952.050.320.373.151.829.762.049.858.945.331.239.846.8
SSAC [20]84.546.775.553.149.638.331.048.856.820.281.552.630.275.853.561.146.830.237.548.4
HRDA [17]83.550.578.651.948.340.535.457.858.219.981.856.329.780.565.880.750.431.040.852.2
MIC [18]84.951.678.453.848.641.838.860.557.122.083.460.231.981.464.870.852.237.843.853.8
HALO [8]85.752.177.952.947.239.337.563.255.621.582.161.430.780.563.972.353.439.144.754.2
CoDA [12]86.150.979.151.749.340.136.965.458.323.282.762.833.180.262.574.654.938.345.654.9
WA2Net (ours)87.454.279.354.349.940.642.267.859.422.883.968.234.383.565.176.156.846.448.256.5
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.545, + 0.482, + 0.697 + ], + "angle": 0, + "content": "split of CityScapes; for the target domain, we use either the test of ACDC/DarkZurich or the validation set of NightCity. (1) CityScapes' training set contains 2,975 images of 19 classes with resolution of \\(2048 \\times 1024\\). (2) ACDC contains weather conditions including fog, nighttime, rain, and snow. The test set contains 2,000 images of \\(1920 \\times 1080\\) resolution. (3) NightCity includes 4,297 real night-time images, divided into 2,998 train images and 1,299 val images. (4) DarkZurich contains 2,416 training and 151 test images for nighttime." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.7, + 0.483, + 0.882 + ], + "angle": 0, + "content": "Implementation Details. We adopt Pytorch [40] and Detectron2 [55] to implement the proposed method. 4 NVIDIA GeForce RTX 3090 GPUs are used for training. The extractor of amplitude characteristic is ResNet-18 [15]. During the training stage, our model is trained with a batch size of 4, using the Adam optimizer [31] with an initial learning rate of 0.0001. The input images are rescaled and randomly cropped to \\(512 \\times 512\\) following the same data augmentation in DAFormer [16], and the EMA coefficient for updating the teacher net is set to be 0.999. We set the prompt number as \\(N = 16\\), and the coefficient of the diversity loss as \\(\\lambda_{div} = 0.2\\)." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.886, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Evaluation Metric. For a fair comparison, we adopt the" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.545, + 0.907, + 0.607 + ], + "angle": 0, + "content": "metric of mean intersection over union (mIOU). It measures the similarity between the prediction and the ground truth by computing the ratio of the intersection area to the union area of the two masks for each object class." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.617, + 0.882, + 0.634 + ], + "angle": 0, + "content": "4.2. Comparison with State-of-the-Art Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.64, + 0.907, + 0.822 + ], + "angle": 0, + "content": "Our method demonstrates superior performance for semantic segmentation under adverse conditions, outperforming state-of-the-art methods in Table 1, 2& 3. As can be observed, our method achieves significant improvements in challenging classes such as road, traffic light and person. Besides, as shown in Figure 3, our method shows promising segmentation performance in diverse weather scenes. In specific, our method performs great in most scenarios, while MIC [18] may generate confusing activation due to the inability to explicitly perceive weather conditions, our \\(\\mathrm{WA}^2\\mathrm{Net}\\) can adaptively focus on accurate foreground targets owing to weather-aware aggregation and adaptation." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.833, + 0.667, + 0.849 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.903 + ], + "angle": 0, + "content": "Effectiveness of the Amplitude Prompt Aggregation. The amplitude prompt aggregation consists of two processes: amplitude characteristic extraction and weather-" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13957" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.09, + 0.908, + 0.119 + ], + "angle": 0, + "content": "Table 3. Quantitative comparisons of performance on CityScapes [6] \\(\\rightarrow\\) DarkZurich [44], where the best performance is marked as bold and the second is underlined. \\(^\\dagger\\) denotes the performance of our reproduction of the corresponding model." + }, + { + "type": "table", + "bbox": [ + 0.099, + 0.13, + 0.905, + 0.297 + ], + "angle": 0, + "content": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]85.837.955.527.714.523.114.021.132.18.72.039.916.664.013.80.058.828.520.729.7
\\( MGCDA^† \\)[44]80.349.366.27.811.041.438.939.064.118.055.852.153.574.766.00.037.529.122.742.5
\\( DANNet^† \\)[54]90.054.074.841.021.125.026.830.272.026.284.047.033.968.219.00.366.438.323.644.3
DAFormer [16]93.565.573.339.419.253.344.144.059.534.566.653.452.782.152.79.589.350.538.553.8
HRDA [17]90.456.372.039.519.557.852.743.159.329.170.560.058.684.075.511.290.551.640.955.9
MIC [18]94.875.084.055.128.462.035.552.659.246.870.065.261.782.164.218.591.352.644.060.2
\\( HALO^† \\)[8]94.574.583.955.528.661.236.052.059.045.570.864.561.281.564.018.290.852.344.260.0
\\( CoDA^† \\)[12]95.374.285.157.329.161.539.853.458.745.372.666.162.283.065.719.390.553.245.361.4
\\( WA^2Net (ours) \\)95.777.986.759.631.263.451.955.760.947.276.367.963.184.569.818.191.954.846.162.8
" + }, + { + "type": "image", + "bbox": [ + 0.093, + 0.311, + 0.907, + 0.562 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.092, + 0.571, + 0.904, + 0.586 + ], + "angle": 0, + "content": "Figure 3. Qualitative results of our \\(\\mathrm{WA}^2\\mathrm{Net}\\) and MIC [18] under adverse conditions on ACDC [45], NightCity [49] and DarkZurich [44]." + }, + { + "type": "table_caption", + "bbox": [ + 0.124, + 0.61, + 0.451, + 0.624 + ], + "angle": 0, + "content": "Table 4. Ablation on main components of our network." + }, + { + "type": "table", + "bbox": [ + 0.098, + 0.636, + 0.481, + 0.719 + ], + "angle": 0, + "content": "
Amplitude Prompt AggregationWeather Heterogeneity AdaptationACDCNightCity
Charac. ExtractionPrompt Agg.Prompt Modu.Homo. Attention
XXXX68.652.4
✓XX✓70.753.7
✓X✓✓71.254.9
✓✓✓✓73.856.5
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.751, + 0.484, + 0.901 + ], + "angle": 0, + "content": "aware prompt aggregation. To evaluate the effectiveness of the amplitude prompt aggregation, we conduct ablation studies as shown in Table 4. As shown in the \\(3^{rd}\\) row in Table 4, not using prompt aggregation alone means that the prompt bank does not undergo knowledge learning of amplitude characteristics but instead self-initialized and only updated during network training. Without aggregating amplitude characteristic information, prompts cannot perceive weather-aware knowledge, failing to effectively modulate features in the prompt-guided modulation with a significant" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.61, + 0.905, + 0.808 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.516, + 0.818, + 0.902, + 0.833 + ], + "angle": 0, + "content": "Figure 4. Qualitative results on the daytime \\(\\rightarrow\\) nighttime setting." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.868, + 0.657, + 0.88 + ], + "angle": 0, + "content": "performance decline." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.886, + 0.905, + 0.902 + ], + "angle": 0, + "content": "Effectiveness of the Weather Heterogeneity Adaptation." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13958" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.094, + 0.089, + 0.28, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.089, + 0.482, + 0.185 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.196, + 0.483, + 0.211 + ], + "angle": 0, + "content": "Figure 5. Ablation experiments on super-parameters: \\(\\lambda_{div}\\) and \\(N\\)." + }, + { + "type": "table_caption", + "bbox": [ + 0.118, + 0.226, + 0.456, + 0.24 + ], + "angle": 0, + "content": "Table 5. Performance w/ and w/o the diversity loss \\( {\\mathcal{L}}_{div} \\) ." + }, + { + "type": "table", + "bbox": [ + 0.136, + 0.251, + 0.439, + 0.315 + ], + "angle": 0, + "content": "
Diversity LossACDC [45]NightCity [49]
w/o L_divw/ L_div
✓72.456.0
✓73.856.5
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.341, + 0.483, + 0.522 + ], + "angle": 0, + "content": "The weather heterogeneity adaptation includes the prompt-guided modulation and homogenization attention. We assess the impact of the weather heterogeneity adaptation by disabling this process and observing the changes in segmentation accuracy. Simultaneously discarding both prompt aggregation and modulation, as shown in the \\(2^{nd}\\) row of Table 4 (i.e., abandoning the generation and use of prompts), means that the extracted amplitude features are directly fed into the homogenization attention mechanism. This approach results in \\(f_{a}^{\\theta}\\) being unable to perceive multi-weather scene characteristics, adversely affecting the model's generalization in extreme scenarios." + }, + { + "type": "text", + "bbox": [ + 0.092, + 0.524, + 0.483, + 0.811 + ], + "angle": 0, + "content": "Analysis of the Homogenization Attention. We also analyze the effectiveness of our designed homogenization attention with comparisons against other attention mechanisms as shown in Table 6. The homogenization attention mechanism is designed to address the significant variability in weather conditions by dynamically aligning features from different weather domains. Unlike crossattention, which focuses on explicit interactions between tokens across feature spaces, and masked attention, which selectively attends to spatial regions, our homogenization attention incorporates a domain adaptation strategy that directly mitigates the feature discrepancies caused by adverse weather conditions. As can be observed, the heterogeneity adaptation mechanism proved to be essential to dynamically adapt to varying weather patterns, thereby improving overall performance and reliability. Besides, Figure 6 displays activation maps of different classes with and without weather-aware aggregation and adaptation, illustrating the impact of our design on class-specific segmentation quality." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Analysis of Hyperparameters. We conducted a comprehensive analysis of hyperparameters to understand their influence on the model's performance. The impact of the choice of diversity loss \\(\\lambda_{div}\\) and the prompt number \\(N\\) is shown in Figure 5. Too few prompts hinder the ability to capture rich scene information, while too many may capture" + }, + { + "type": "image", + "bbox": [ + 0.515, + 0.09, + 0.905, + 0.146 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.147, + 0.905, + 0.2 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.201, + 0.905, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.257, + 0.905, + 0.311 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.323, + 0.907, + 0.352 + ], + "angle": 0, + "content": "Figure 6. Comparison of activation maps for different classes with (w/) and without(w/o) weather-aware aggregation and adaptation." + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.368, + 0.907, + 0.395 + ], + "angle": 0, + "content": "Table 6. Comparison of different attention mechanisms with our homogenization attention." + }, + { + "type": "table", + "bbox": [ + 0.522, + 0.407, + 0.903, + 0.471 + ], + "angle": 0, + "content": "
Attention MechanismACDC [45]NightCity [49]
cross-attention [52]70.954.1
masked attention [5]71.855.0
Homogenization Attention (ours)73.856.5
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.499, + 0.907, + 0.635 + ], + "angle": 0, + "content": "irrelevant noise information. It is worth noting that the improvement brought by the diversity loss in Table 5 is more significant when ACDC [45] is the target domain. This is due to ACDC containing more weather diversities (night, rain, snow and fog) compared to the night-only NightCity [49], prompts on ACDC can aggregate more diverse knowledge with \\(\\mathcal{L}_{div}\\). Therefore, the diversity in prompt combinations contributes to diverse scene analysis under multiple weather conditions." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.651, + 0.634, + 0.667 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.678, + 0.907, + 0.783 + ], + "angle": 0, + "content": "In this paper, we propose a weather-aware aggregation and adaptation network (WA\\(^2\\)Net) for semantic segmentation under adverse conditions. We design the amplitude prompt aggregation to acquire weather-aware characteristic knowledge and weather heterogeneity adaptation to realize interweather homogenization for effective parsing. Extensive experiments demonstrate our effectiveness." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.8, + 0.673, + 0.816 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.906, + 0.9 + ], + "angle": 0, + "content": "This work was partially supported by the National Key Research and Development Program of China (No.2024YFB3909902), National Nature Science Foundation of China (62121002), Youth Innovation Promotion Association of CAS." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13959" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.093, + 0.091, + 0.188, + 0.106 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.115, + 0.484, + 0.184 + ], + "angle": 0, + "content": "[1] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.186, + 0.484, + 0.242 + ], + "angle": 0, + "content": "[2] Yang Chen, Liang Shan, Yi Qu, Weilong Zhang, and Lu Chang. Amsc: Adaptive masking and structure-constraint learning for domain adaptive semantic segmentation under adverse conditions. IEEE Signal Processing Letters, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.243, + 0.483, + 0.325 + ], + "angle": 0, + "content": "[3] Yujia Chen, Wangkai Li, Zhaoyang Li, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Sam-glomeruli: Enhanced segment anything model for precise glomeruli segmentation. In International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis, pages 182-191. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.327, + 0.483, + 0.408 + ], + "angle": 0, + "content": "[4] Yujia Chen, Rui Sun, Wangkai Li, Huayu Mai, Naisong Luo, Yuwen Pan, and Tianzhu Zhang. Alleviate and mining: Rethinking unsupervised domain adaptation for mitochondria segmentation from pseudo-label perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2339-2347, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.41, + 0.483, + 0.48 + ], + "angle": 0, + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.481, + 0.483, + 0.564 + ], + "angle": 0, + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.565, + 0.483, + 0.62 + ], + "angle": 0, + "content": "[7] Ozgur Erkent and Christian Laugier. Semantic segmentation with unsupervised domain adaptation under varying weather conditions for autonomous vehicles. IEEE Robotics and Automation Letters, 5(2):3580-3587, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.621, + 0.483, + 0.69 + ], + "angle": 0, + "content": "[8] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. In International Conference on Machine Learning, pages 13864-13884. PMLR, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.691, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.735, + 0.483, + 0.816 + ], + "angle": 0, + "content": "[10] Rui Gong, Qin Wang, Martin Danelljan, Dengxin Dai, and Luc Van Gool. Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7225-7235, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.818, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Ziyang Gong, Fuhao Li, Yupeng Deng, Wenjun Shen, Xi-anzheng Ma, Zhenming Ji, and Nan Xia. Train one, generalize to all: Generalizable semantic segmentation from single-scene to all adverse scenes. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2275-2284, 2023. 2" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.115, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.906, + 0.163 + ], + "angle": 0, + "content": "[12] Ziyang Gong, Fuhao Li, Yupeng Deng, Deblina Bhattacharjee, Xianzheng Ma, Xiangwei Zhu, and Zhenming Ji. Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning. In European Conference on Computer Vision, pages 130-148. Springer, 2024. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[13] Monica Gruosso, Nicola Capecce, and Ugo Erra. Human segmentation in surveillance video with deep learning. *Multimedia Tools and Applications*, 80(1):1175-1199, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.209, + 0.905, + 0.277 + ], + "angle": 0, + "content": "[14] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. Ppt: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.281, + 0.905, + 0.337 + ], + "angle": 0, + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.339, + 0.907, + 0.409 + ], + "angle": 0, + "content": "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.411, + 0.905, + 0.467 + ], + "angle": 0, + "content": "[17] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision, pages 372-391. Springer, 2022. 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.469, + 0.905, + 0.539 + ], + "angle": 0, + "content": "[18] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721-11732, 2023. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.541, + 0.905, + 0.597 + ], + "angle": 0, + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.599, + 0.905, + 0.669 + ], + "angle": 0, + "content": "[20] Abdulrahman Kerim, Felipe Chamone, Washington Ramos, Leandro Soriano Marcolino, Erickson R Nascimento, and Richard Jiang. Semantic segmentation under adverse conditions: a weather and nighttime-aware synthetic data-based approach. arXiv preprint arXiv:2210.05626, 2022. 2, 3, 5, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.671, + 0.905, + 0.725 + ], + "angle": 0, + "content": "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.728, + 0.905, + 0.785 + ], + "angle": 0, + "content": "[22] Sohyun Lee, Namyup Kim, Sungyeon Kim, and Suha Kwak. Frest: Feature restoration for semantic segmentation under multiple adverse conditions. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[23] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Wangkai Li, Rui Sun, Bohao Liao, Zhaoyang Li, and Tianzhu Zhang. Balanced learning for domain adaptive semantic segmentation. In *Forty-second International Conference on Machine Learning*. 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.52, + 0.957 + ], + "angle": 0, + "content": "13960" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.134 + ], + "angle": 0, + "content": "[25] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.484, + 0.19 + ], + "angle": 0, + "content": "[26] Zhi Li, Xing Wu, Jianjia Wang, and Yike Guo. Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing, 127: 104554, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.193, + 0.484, + 0.262 + ], + "angle": 0, + "content": "[27] Zhaoyang Li, Yuan Wang, Wangkai Li, Rui Sun, and Tianzhu Zhang. Localization and expansion: A decoupled framework for point cloud few-shot semantic segmentation. In European Conference on Computer Vision, pages 18-34. Springer, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.264, + 0.484, + 0.332 + ], + "angle": 0, + "content": "[28] Zhaoyang Li, Yuan Wang, Wangkai Li, Tianzhu Zhang, and Xiang Liu. Dual-agent optimization framework for cross-domain few-shot segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9849-9859, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.403 + ], + "angle": 0, + "content": "[29] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.405, + 0.484, + 0.461 + ], + "angle": 0, + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.462, + 0.484, + 0.49 + ], + "angle": 0, + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.492, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[32] Ning Lv, Zenghui Zhang, Cong Li, Jiaxuan Deng, Tao Su, Chen Chen, and Yang Zhou. A hybrid-attention semantic segmentation network for remote sensing interpretation in land-use surveillance. International Journal of Machine Learning and Cybernetics, 14(2):395–406, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.563, + 0.484, + 0.631 + ], + "angle": 0, + "content": "[33] Claudia Menzel, Gregor U Hayn-Leichsenring, Oliver Langner, Holger Wiese, and Christoph Redies. Fourier power spectrum characteristics of face photographs: attractiveness perception depends on low-level image properties. PLoS One, 10(4):e0122801, 2015. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.484, + 0.73 + ], + "angle": 0, + "content": "[34] Khan Muhammad, Tanveer Hussain, Hayat Ullah, Javier Del Ser, Mahdi Rezaei, Neeraj Kumar, Mohammad Hijji, Paolo Bellavista, and Victor Hugo C de Albuquerque. Vision-based semantic segmentation in scene understanding for autonomous driving: Recent achievements, challenges, and outlooks. IEEE Transactions on Intelligent Transportation Systems, 23(12):22694-22715, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.732, + 0.484, + 0.788 + ], + "angle": 0, + "content": "[35] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1520-1528, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.789, + 0.484, + 0.87 + ], + "angle": 0, + "content": "[36] Yuwen Pan, Naisong Luo, Rui Sun, Meng Meng, Tianzhu Zhang, Zhiwei Xiong, and Yongdong Zhang. Adaptive template transformer for mitochondria segmentation in electron microscopy images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21474-21484, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[37] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.134 + ], + "angle": 0, + "content": "enhancement for night-time semantic segmentation. In European Conference on Computer Vision, pages 408-424. Springer, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.204 + ], + "angle": 0, + "content": "[38] Yuwen Pan, Rui Sun, Yuan Wang, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Purify then guide: A bidirectional bridge network for open-vocabulary semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.205, + 0.905, + 0.273 + ], + "angle": 0, + "content": "[39] Yuwen Pan, Rui Sun, Yuan Wang, Tianzhu Zhang, and Yongdong Zhang. Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2031-2040, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.274, + 0.905, + 0.355 + ], + "angle": 0, + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.357, + 0.905, + 0.413 + ], + "angle": 0, + "content": "[41] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In 2019 22th International Conference on Information Fusion (FUSION), pages 1-8. IEEE, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.414, + 0.905, + 0.482 + ], + "angle": 0, + "content": "[42] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of fast video-sequence segmentation. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.484, + 0.905, + 0.552 + ], + "angle": 0, + "content": "[43] Nikhil Reddy, Abhinav Singhal, Abhishek Kumar, Mahsa Baktashmotlagh, and Chetan Arora. Master of all: Simultaneous generalization of urban-scene segmentation to all adverse weather conditions. In European Conference on Computer Vision, pages 51-69. Springer, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[44] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3139-3153, 2020. 2, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.622, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[45] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1, 3, 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[46] Rui Sun, Naisong Luo, Yuan Wang, Yuwen Pan, Huayu Mai, Zhe Zhang, and Tianzhu Zhang. 1st place solution for youtuves challenge 2022: Video object segmentation. In Proceedings of the Conference on Computer Vision and Pattern Recognition Workshops, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.762, + 0.905, + 0.817 + ], + "angle": 0, + "content": "[47] Rui Sun, Naisong Luo, Yuwen Pan, Huayu Mai, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Appearance prompt vision transformer for connectome reconstruction. In IJCAI, pages 1423-1431, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.818, + 0.905, + 0.872 + ], + "angle": 0, + "content": "[48] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: exploring the better weighting function for semi-supervised semantic segmentation. Advances in neural information processing systems, 36:61792-61805, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson WH Lau. Night-time scene parsing with a large" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.518, + 0.957 + ], + "angle": 0, + "content": "13961" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.482, + 0.12 + ], + "angle": 0, + "content": "real dataset. IEEE Transactions on Image Processing, 30: 9085-9098, 2021. 5, 6, 7, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.483, + 0.163 + ], + "angle": 0, + "content": "[50] Yingjie Tian and Siyu Zhu. Partial domain adaptation on semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3798-3809, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.234 + ], + "angle": 0, + "content": "[51] Abhinav Valada, Johan Vertens, Ankit Dhall, and Wolfram Burgard. Adapnet: Adaptive semantic segmentation in adverse environmental conditions. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4644-4651. IEEE, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.482, + 0.373 + ], + "angle": 0, + "content": "[53] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.375, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[54] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.446, + 0.482, + 0.488 + ], + "angle": 0, + "content": "[55] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.489, + 0.482, + 0.544 + ], + "angle": 0, + "content": "[56] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.546, + 0.482, + 0.601 + ], + "angle": 0, + "content": "[57] Xin Yang, Wending Yan, Yuan Yuan, Michael Bi Mi, and Robby T Tan. Semantic segmentation in multiple adverse weather conditions with domain knowledge retention. arXiv preprint arXiv:2401.07459, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.602, + 0.483, + 0.657 + ], + "angle": 0, + "content": "[58] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 2, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.659, + 0.482, + 0.728 + ], + "angle": 0, + "content": "[59] Chunhui Zhang, Guanjie Huang, Li Liu, Shan Huang, Yinan Yang, Xiang Wan, Shiming Ge, and Dacheng Tao. Webuav-3m: A benchmark for unveiling the power of million-scale deep uav tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9186-9205, 2022. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.729, + 0.482, + 0.797 + ], + "angle": 0, + "content": "[60] Chunhui Zhang, Li Liu, Guanjie Huang, Hao Wen, Xi Zhou, and Yanfeng Wang. Webuot-1m: Advancing deep underwater object tracking with a million-scale benchmark. Advances in Neural Information Processing Systems, 37:50152-50167, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.8, + 0.482, + 0.87 + ], + "angle": 0, + "content": "[61] Chunhui Zhang, Li Liu, Hao Wen, Xi Zhou, and Yanfeng Wang. Mambatrack: Exploiting dual-enhancement for night uav tracking. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.483, + 0.87 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "13962" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_origin.pdf b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..28ba43ba36462e0a7797e0eef76f89575d782561 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/bcf1a36a-351d-4939-805f-0e0dec332561_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:31184b89bac63fc516bbe5793f76779b2f9bb60153fad3d42e472ba84ba4511e +size 5827875 diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/full.md b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ded0f1569f53d26495692276ecf5df7f6a72b417 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/full.md @@ -0,0 +1,339 @@ +# Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions + +Yuwen Pan $^{1}$ Rui Sun $^{1}$ Wangkai Li $^{1}$ Tianzhu Zhang $^{1,2*}$ + +1University of Science and Technology of China + +$^{2}$ National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory + +{panyw, issunrui, lwklwk}@mail.ustc.edu.cn, tzzhang@ustc.edu.cn + +# Abstract + +Semantic segmentation under adverse conditions is critical for reliable visual perception in challenging weather environments. These extreme scenarios introduce distortions, such as low contrast and reduced visibility, making traditional segmentation models struggle. The scarcity of labeled data in such conditions makes it difficult to train models directly for these environments. Unsupervised domain adaptation (UDA) has been proposed as a solution to transfer knowledge from labeled source domains (normal weather) to unlabeled target domains (adverse weather). However, existing methods face significant challenges, particularly due to weather unawareness and feature heterogeneity. Many models fail to account for the unique characteristics of different weather conditions, and the significant feature discrepancies between normal and adverse weather images hinder effective adaptation. In this paper, we propose a novel weather-aware aggregation and adaptation network that leverages characteristic knowledge to achieve weather homogenization and enhance scene perception. Specifically, we introduce amplitude prompt aggregation to capture essential characteristics from the Fourier frequency domain that are indicative of different weather conditions. Additionally, we employ weather heterogeneity adaptation to mitigate the inter-domain heterogeneity, thereby achieving feature homogenization across diverse environments. Extensive experimental results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions. + +# 1. Introduction + +Semantic segmentation involves assigning a class label to each pixel in an image [30, 35, 36, 38, 47], which plays a crucial role in various real-world applications such as au + +![](images/4014c906c279a86dcddccedca7e5c2ec5874bdba89befc81d7f9d53ca8590a3a.jpg) +(a) Traditional Cross-weather Paradigm + +![](images/eb4ce93c6987c17e3eba03460d11ee818b4e92033b34bd47c7584f7a3f8bc3f1.jpg) +(b) Weather-aware Aggregation and Adaptation (ours) +Figure 1. Comparison of the traditional cross-weather paradigm and our weather-aware aggregation and adaptation. To mitigate the existing issues (unawareness of weather characteristics and severe cross-weather heterogeneity), we aim to aggregate weather-aware attribute knowledge and homogenize domain discrepancies. + +tonomous driving [34, 37, 59] and surveillance [13, 32, 60]. However, the performance of segmentation models often degrades significantly when faced with adverse weather conditions [22, 45], including night, fog, rain, and snow. The deterioration in visibility and scene appearance poses a substantial challenge for traditional segmentation algorithms, which are typically trained on datasets collected under normal weather conditions. Therefore, exploring an effective way for semantic parsing under extreme weather conditions is highly anticipated. + +Extensive research efforts have been devoted to semantic segmentation in adverse weather conditions [7, 26, 41- + +43, 51]. Due to the scarcity of labeled data in such adverse conditions, existing methods strive to employ an unsupervised domain adaptation (UDA) paradigm utilizing both labeled normal weather data and images without annotations under extreme scenarios to address the challenge [16, 44, 50, 53, 54]. Kerim et al. [20] propose a dual-branch network based on DeepLabV3+ [1], enhancing training knowledge for extreme weather conditions by incorporating a synthetic multi-scene dataset. However, discerning similar-looking categories in the target domain with limited real annotations remains a challenge for domain adaptation methods. To address this, MIC [18] introduces the masked image consistency, leveraging spatial context relations in the target domain as additional cues to enhance the robustness of UDA. The emergence of these methods has benefited the understanding of scenes in extreme weather conditions. + +Despite considerable progress, the task of distinguishing and analyzing scenes under different extreme weather conditions remains challenging. Whether through the assistance of synthetic data or exploration of spatial context clues, existing methods struggle with accurately identifying weather-aware semantics in multiple extreme scenarios. We attribute the predicament of this cross-weather paradigm to two inherent issues (Figure 1): (1) Unawareness of Weather Characteristics. The style characteristics under adverse weathers, such as illumination conditions and contrast, are primary culprits responsible for the inter-weather gap among diverse scenes. Although recognizing this crucial aspect, existing works [10, 58] generally employ low-level enhancements or implicit learning for inter-domain style transfer, failing to fully exploit weather attribute representation. Therefore, designing a strategy to aggregate weather-aware scene knowledge across diverse conditions is worthy of exploration. (2) Severe Cross-Weather Heterogeneity. Early works [16, 18] tend to scene images from different weather conditions uniformly, ignoring the substantial differences across different weather types. The significant heterogeneity across weather conditions presents another challenge, stemming from the diverse and complex nature of weather phenomena that introduce distinct visual distortions and scene alterations. Even though some methods [2, 11] attempt to address this challenge, the significant differences in characteristics across different weather conditions still pose considerable difficulties. These variations lead to a significant disparity between features learned from the normal weather and those required for effective interpretation in adverse scenes. Addressing how to leverage characteristic knowledge to achieve weather homogenization to facilitate scene perception in adverse scenes is an essential question. + +Motivated by the above discussions, we propose a Weather-aware Aggregation and Adaptation NETwork + +(WA $^2$ Net) to aggregate weather-aware attribute knowledge and homogenize domain discrepancies, including amplitude prompt aggregation and weather heterogeneity adaptation. In the amplitude prompt aggregation, to acquire weather-aware characteristic knowledge, we utilize the amplitude operation of Fourier transform to capture essential characteristics indicative of different weather conditions. These amplitude features serve as discriminative cues for weather-related variations in the scene. To efficiently aggregate the crucial characteristic knowledge from the frequency domain into weather-adaptive prompts, we prepend a set of learnable prompts in the prompt bank to interact with frequency features by cross-attention mechanism to obtain weather-aware prompts. These prompts encapsulate the intrinsic characteristics of various weather conditions, providing essential guidance for the adaptation process. In the weather heterogeneity adaptation, to realize inter-weather homogenization for effective semantic parsing, the resultant weather-aware prompts are leveraged to generate adapted characteristics after modulation. Then, we propose a novel homogenization attention mechanism to acquire homogenized features across diverse weather conditions. This mechanism identifies the discrepancy between features of the original image and adapted characteristics, decoupling weather characteristics from valuable target information and thereby filtering out attribute interference. In this way, our WA $^2$ Net achieves effective weather-aware knowledge aggregation and robust adaptation, ensuring improved performance for semantic segmentation in adverse weather conditions. + +To sum up, our contributions can be summarized as follows: + +- We propose a novel weather-aware aggregation and adaptation network for semantic segmentation under adverse conditions. Our paper exploits a practical way to acquire weather-adaptive knowledge and enhance cross-weather perception via an aggregation and adaptation manner. +- We design the amplitude prompt aggregation to acquire weather-aware prompts from weather characteristics through amplitude extraction, and the weather heterogeneity adaptation to inherit weather-aware knowledge and mitigate the inter-weather heterogeneity. +- Extensive experiment results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions. + +# 2. Related Work + +# 2.1. Semantic Segmentation under Adverse Conditions + +The field of semantic segmentation [3, 4, 23, 27, 28, 39, 48] has made significant advancements with the emergence + +of convolutional neural networks (CNNs) [21] and transformers [52]. However, traditional semantic segmentation models experience a notable performance drop under adverse conditions such as rain, fog and night-time environments [45, 46, 61]. Therefore, semantic segmentation under adverse conditions aims to develop robust models capable of accurately delineating objects and regions in challenging environments. Upon the scarcity of labeled data in extreme weathers, the unsupervised domain adaptation paradigm is utilized to enhance the generalization of models across different weather conditions with the aid of normal weather data [24, 54]. Kerim et al. [20] propose a dual-branch network and expand the training dataset for extreme weather conditions by integrating a synthetic multiscene dataset. Nevertheless, domain adaptation methods encounter challenges due to the difficulty in distinguishing similar-looking categories in the target domain with limited real annotations. To tackle this issue, MIC [18] introduces Masked Image Consistency, utilizing spatial context relations in the target domain as additional cues to enhance vision perception. Additionally, to mitigate knowledge forgetting in new scenes, Yang et al. [57] introduce adaptive knowledge acquisition and replay of weather-condition combinations, aiming to alleviate model-forgetting issues in diverse scenes. Differently, we strive to tackle the challenge of weather-unaware perception, leveraging weather-aware prompts in an aggregation and adaptation manner. + +# 2.2. Prompt Learning + +Prompting [29] incorporates task-specific instructions into input text to guide model behavior, enhancing performance without retraining, particularly in few-shot learning [9]. Recent work extends prompting to vision tasks, where learnable prompts enable adaptation without tuning all parameters [25] and improve segmentation generalization [14]. For semantic segmentation under adverse weather, weather-adaptive prompts modulate visual embeddings to mitigate distortions like lighting and visibility changes, enhancing robustness without extensive fine-tuning. Integrating weather-specific knowledge into prompts improves generalization across diverse conditions, making models more effective in real-world applications. + +# 3. Method + +# 3.1. Preliminaries + +Typically, in the field of unsupervised domain adaptation [16, 18], the source domain data can be formulated as $D^{s} = \{X^{s}, Y^{s}\}$ , where $X^{s}$ and $Y^{s}$ represent the corresponding image and label space, respectively. Each source domain image $x_{k}^{s} \in X^{s}$ is associated with a ground truth segmentation mask $y_{k}^{s} \in Y^{s}$ . Similarly, let $X^{t}$ be the target domain image space, and $Y^{t}$ be the label space. However, + +unlike the source domain, the target domain has no access to annotated segmentation masks. The supervised loss $\mathcal{L}^s$ can be formulated with a neural network $f_{\mu}$ as: + +$$ +\mathcal {L} ^ {s} = \frac {1}{N _ {s}} \sum_ {k = 1} ^ {N _ {s}} \mathcal {L} _ {c e} \left(f _ {\mu} \left(x _ {k} ^ {s}\right), y _ {k} ^ {s}\right), \tag {1} +$$ + +where $N_{s}$ denotes the image number of the source domain, $\mathcal{L}_{ce}$ denotes the pixel-wise cross-entropy loss as: + +$$ +\mathcal {L} _ {c e} (\hat {y}, y) = - \sum_ {h = 1} ^ {H} \sum_ {w = 1} ^ {W} \sum_ {c = 1} ^ {C} y _ {h w c} \log \hat {y} _ {h w c}. \tag {2} +$$ + +The model trained on the source domain $D^{s}$ normally exhibits performance degradation when tested on the target domain $D^{t}$ . Therefore, the self-training paradigm generates pseudo-labels for the target domain images $X^{t}$ with a teacher network $g_{\phi}$ instead of the student network $f_{\mu}$ . Specifically, the model predicts segmentation masks for unlabeled target domain images, producing pseudo-labels $\hat{y}_k^t$ for each image $x_k^t$ in the target domain. The unsupervised loss $\mathcal{L}^t$ for the target domain is formulated as: + +$$ +\mathcal {L} ^ {t} = \frac {1}{N _ {t}} \sum_ {k = 1} ^ {N _ {t}} \lambda^ {t} \mathcal {L} _ {c e} \left(f _ {\mu} \left(x _ {k} ^ {t}\right), \hat {y} _ {k} ^ {t}\right), \tag {3} +$$ + +where $N_{t}$ denotes the image number of the target domain, $\lambda^t$ is a weight coefficient for the unsupervised loss. The student network updates the parameters of the teacher network through backpropagation, utilizing the teacher network's output as targets to minimize prediction errors, while also employing Exponential Moving Average (EMA) update to stabilize the training process and improve convergence. + +# 3.2. Overview + +As shown in Figure 2, given an input image $x^{s}$ from the source domain and $x^{t}$ from the target domain under adverse conditions, we denote the image under different weather condition $\theta$ as $x^{\theta}$ for clarity, where $\theta \in (0, \dots, \Theta)$ , $\Theta$ denotes the total number of adverse weather conditions ( $x^{0}$ refers to the source domain image under the normal weather condition). The Fourier amplitude features $f_{a}^{\theta}$ are decoupled from the frequency domain. Then, a set of prompts, i.e., prompt bank $\mathbf{P}$ , is learned during the prompt aggregation process. The acquired weather-aware prompts $\hat{\mathbf{P}}$ further interact with amplitude features $f_{a}^{\theta}$ to evolve into adapted characteristic $\hat{f}_{a}^{\theta}$ . Then, to achieve homogenization for features from diverse weather conditions, we further employ homogenization attention to mitigate the influence of weather heterogeneity. Finally, the modulated features $f_{h}^{\theta}$ are fed into the teacher-student network for comprehensive training. + +![](images/61fe4643d51294a0265b4d8679594ece20fc2870f68416906e2d33b97ff393a2.jpg) +Figure 2. Illustration of our proposed $\mathrm{WA}^2\mathrm{Net}$ . Our network contains the amplitude prompt aggregation to acquire weather-aware prompts (Section 3.3) and weather heterogeneity adaptation to realize feature homogenization (Section 3.4). + +# 3.3. Amplitude Prompt Aggregation + +In order to disentangle the latent weather characteristics and target information fused in the RGB images, we first model the representation of weather characteristics by exploring the Fourier frequency domain. Then, we collect the weather characteristic properties in Fourier amplitude features and acquire weather-aware prompts during the prompt aggregation process. + +Amplitude Characteristic Extraction. In the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information [33, 58], which reflects the energy distribution of the frequency components of the image and can be understood as a representation of the "characteristics" or "style" of the image. Therefore, we can utilize the amplitude of Fourier spectrum to extract weather characteristic knowledge in diverse scenes. Specifically, in the amplitude extractor, given an image $x^{\theta} \in \mathbb{R}^{H \times W \times 3}$ , we apply a two-dimensional Fourier transform $\mathcal{F}(x)$ as: + +$$ +\mathcal {F} \left(x ^ {\theta}\right) _ {u, v} = \sum_ {h = 0} ^ {H - 1} \sum_ {w = 0} ^ {W - 1} x _ {h, w} ^ {\theta} e ^ {- J 2 \pi \left(\frac {w h}{H} + \frac {v w}{W}\right)}, \tag {4} +$$ + +where $J$ refers to the imaginary unit. Then we can acquire the corresponding amplitude $\mathcal{A}$ and phase $\Phi$ as: + +$$ +\mathcal {A} \left(x ^ {\theta}\right) _ {u, v} = \left| \mathcal {F} \left(x ^ {\theta}\right) _ {u, v} \right|, \tag {5} +$$ + +$$ +\Phi \left(x ^ {\theta}\right) _ {u, v} = \arg \left(\mathcal {F} \left(x ^ {\theta}\right) _ {u, v}\right) = \arctan \left[ \frac {\operatorname {I m} \left(\mathcal {F} \left(x ^ {\theta}\right) _ {u , v}\right)}{\operatorname {R e} \left(\mathcal {F} \left(x ^ {\theta}\right) _ {u , v}\right)} \right], \tag {6} +$$ + +where $\arg$ represents the argument (or phase angle) of the complex number, Im and Re represent the imaginary and real part of $\mathcal{F}(x)$ , respectively. In this way, to finally generate our amplitude characteristic map, we fix the phase to an average constant $c^p$ and apply inverse Fourier transform to acquire the amplitude characteristic map as + +$$ +\overline {{\mathcal {A}}} (x ^ {\theta}) = \mathcal {F} ^ {- 1} [ \mathcal {A} (x ^ {\theta}) _ {u, v} e ^ {- J c ^ {p}} ], \tag {7} +$$ + +where $\mathcal{F}^{-1}$ denotes the inverse Fourier transformation. Then, we can obtain the amplitude characteristics $\mathbf{F}_a = \{f_a^\theta\}_{\theta=0}^\Theta$ through a shared-weight lightweight encoder (e.g., ResNet-18 [15]), preparing for the following aggregation of weather-aware prompts. + +Weather-aware Prompt Aggregation. To effectively aggregate characteristic information across different weather conditions, we aim to develop a medium capable of absorbing cross-weather knowledge. Inspired by [19], we introduce a set of prompts, i.e., prompt bank $\mathbf{P} = \{\pmb{p}_n\}_{n=1}^N$ , where $\pmb{p}_n \in \mathbb{R}^{1 \times L}$ and $N$ denotes the number of prototypes. These learnable prompts will be leveraged to extract characteristic knowledge by interaction with amplitude features $f_a^\theta$ , which highly represents the style condition of the current scene. In specific, for the acquired amplitude characteristic $f_a^\theta$ , the queries arise from the prompts $\mathbf{P}$ , and keys and values arise from the amplitude feature $\mathbf{F}_a = \{f_a^0, \dots, f_a^\theta, \dots, f_a^\Theta\}$ . Formally, + +$$ +\mathbf {Q} _ {n} = \boldsymbol {p} _ {n} \mathbf {W} ^ {Q}, \mathbf {K} ^ {\theta} = f _ {a} ^ {\theta} \mathbf {W} ^ {K}, \mathbf {V} ^ {\theta} = f _ {a} ^ {\theta} \mathbf {W} ^ {V}, \tag {8} +$$ + +where $n\in [1,\dots ,N]$ , and $\mathbf{W}^Q\in \mathbb{R}^{C\times C_k}$ $\mathbf{W}^K\in$ $\mathbb{R}^{C\times C_k}$ $\mathbf{W}^V\in \mathbb{R}^{C\times C_v}$ are linear projections. The attention weights are calculated based on the dot-product similarity between each query and key: + +$$ +s _ {n, \theta} = \frac {\exp (\beta_ {n , \theta})}{\sum_ {\theta = 0} ^ {\Theta} \exp (\beta_ {n , \theta})}, \beta_ {n, \theta} = \frac {\mathbf {Q} _ {n} \mathbf {K} ^ {\theta} {} ^ {\top}}{\sqrt {C _ {k}}}, \tag {9} +$$ + +$$ +\hat {\boldsymbol {p}} _ {n} = \operatorname {F F N} (\operatorname {A t t} \left(\mathbf {Q} _ {n}, \mathbf {K}, \mathbf {V}\right)) = \operatorname {F F N} \left(\sum_ {\theta = 0} ^ {\Theta} s _ {n, \theta} \mathbf {V} ^ {\theta}\right), \tag {10} +$$ + +where $\mathrm{FFN}()$ denotes the feed forward network. In this way, weather-aware prompts $\hat{\mathbf{P}} = \{\hat{p}_n\}_{n=1}^N$ can distill rich characteristic knowledge from the amplitude features $\mathbf{F}_a$ . + +Besides, in order to enable prompts to carry diverse and comprehensive knowledge of various weather conditions, + +we apply the diversity loss on the prompt bank $\mathbf{P}$ . Formally, + +$$ +\mathcal {L} _ {d i v} = \frac {1}{N (N - 1)} \sum_ {i = 1} ^ {N} \sum_ {j = 1, i \neq j} ^ {N} \left(\frac {\langle \boldsymbol {p} _ {i} , \boldsymbol {p} _ {j} \rangle}{\| \boldsymbol {p} _ {i} \| _ {2} \| \boldsymbol {p} _ {j} \| _ {2}}\right). \tag {11} +$$ + +Applying diversity loss to different prompts offers the advantage of encouraging each prompt $\pmb{p}_n$ to focus on distinct attributes of weather conditions, thereby preventing redundancy in aggregation learning. Diversity loss enhances weather-specific prompts by encouraging each to capture unique traits. Integrating the prompt bank and diversity loss within domain adaptation allows the model to adapt to varied weathers. Splitting this into two tasks is limited by the lack of direct supervision on weather-adaptive prompts; only foreground masks from the source domain are available. Diversity loss thus broadens the prompt bank's coverage of style features for better domain adaptation. + +# 3.4. Weather Heterogeneity Adaptation + +In adverse weather conditions, weather-induced distortions such as glare, fog, and raindrops introduce complexities that can degrade the performance of vision perception. To address the weather heterogeneity challenges posed by varying weather conditions, we introduce the weather heterogeneity adaptation, which utilizes acquired weather-adaptive prompts $\hat{\mathbf{P}}$ as references to modulate the weather characteristic in the current scene and homogenizes weather features via homogenization attention. + +Prompt-guided Modulation. In fact, the introduction of the prompt bank can be deemed as a guidepost in artificially solving incidents, i.e., guiding information, which absorbs weather-aware knowledge across various scenes and can transfer reliable weather-aware properties into modulated characteristic features. For the acquired weather amplitude feature $f_{a}^{\theta}$ , the modulated characteristic can be formulated as: + +$$ +\hat {f} _ {a} ^ {\theta} = \operatorname {S o f t m a x} \left(\frac {f _ {a} ^ {\theta} \hat {\mathbf {P}} ^ {\top}}{\sqrt {C}}\right) \hat {\mathbf {P}}. \tag {12} +$$ + +In this way, the modulated characteristic $\hat{f}_a^\theta$ inherits rich cross-weather knowledge from the weather-aware prompts $\hat{\mathbf{P}}$ , thereby evolving into an adapted representation poised for subsequent homogenization. + +Homogenization Attention Mechanism. With the acquired representation of modulated characteristic $\hat{f}_a^\theta$ , we first extract the vision features of image $x^{\theta}$ from the RGB domain through feature encoding as $f_v^\theta$ . To homogenize vision features from diverse weather conditions, we aim to find an appropriate way to mitigate the influence of $\hat{f}_a^\theta$ on $f_v^\theta$ . Inspired by the success of Transformer [52] architecture in discovering local regions, we further explore the potential of attention mechanism in characteristic homogenization for different weather conditions. + +Given an image feature $f_{v}^{\theta} \in \mathbb{R}^{h \times w \times c}$ extracted from the vision feature extractor (e.g., MiT-B5 [56]), we use the extracted $\hat{f}_{a}^{\theta}$ to remove the weather characteristics through a novel homogenization attention mechanism. First, we employ two convolution layers to map $f_{v}^{\theta}$ and $\hat{f}_{a}^{\theta}$ to the same dimension $C$ , then we can obtain $\overline{\mathbf{f}}_{\mathbf{v}}^{\theta} \in \mathbb{R}^{h \times w \times C}$ and $\overline{f}_{a}^{\theta} \in \mathbb{R}^{h \times w \times C}$ , respectively. The disparity between image features and weather characteristics signifies the saliency details of the target objects, thereby aiding in accentuating the relevant information about the targets. Rather than vanilla transformer attention mechanism [52], we custom design the homogenization attention to reduce heterogeneity in a finer manner, prompting the vision system to highlight salient regions across diverse scenes. Specifically, we acquire the homogenized map $\mathcal{H} \in \mathbb{R}^{h \times w}$ from pixel-level features and weather characteristic as: + +$$ +\mathcal {H} _ {i, j} ^ {\theta} = \sum_ {c = 1} ^ {C} \left(\overline {{\mathbf {f}}} _ {\mathbf {v} _ {i, j, c}} ^ {\theta} - \overline {{f}} _ {a i, j, c} ^ {\theta}\right) ^ {2}, \tag {13} +$$ + +where $i, j,$ and $c$ are the index of height, width, and channel, respectively. Finally, we can get the homogenized pixel features $\mathbf{F}^a$ by weighting the homogenized map $\mathcal{H}$ to original image features in a pixel-wise manner as: + +$$ +\mathbf {f} _ {h} ^ {\theta} = \mathbf {f} ^ {\theta} \circ \mathcal {H} ^ {\theta}, \tag {14} +$$ + +where $\circ$ denotes the element-wise product. Besides, we use a self-attention layer for further aggregating target information across different pixels. The obtained visual features are fed into the conventional teacher-student network, wherein the images from the source domain are input to the student network for supervised constraints, while the images from the target domain are separately input to both the student and teacher networks to generate predictions and pseudolabels for unsupervised constraints. + +# 3.5. Training Supervision + +As illustrated in Section 3.1, the acquired adapted vision features from the source and target are input to the teacher-student network to acquire the supervised loss $\mathcal{L}^s$ and unsupervised loss $\mathcal{L}^t$ . Together with the diversity loss $\mathcal{L}_{div}$ in the prompt bank, we can finally obtain the segmentation map as follows: + +$$ +\mathcal {L} _ {\text {t o t a l}} = \mathcal {L} ^ {s} + \mathcal {L} ^ {t} + \lambda_ {\text {d i v}} \mathcal {L} _ {\text {d i v}}, \tag {15} +$$ + +where $\lambda_{div}$ is the trade-off weight. + +# 4. Experiment + +# 4.1. Experiment Setup + +Datasets. In our experiments, following the conventional practice [16, 18, 20], we utilize 4 datasets including CityScapes [6], ACDC [45], NightCity [49] and DarkZurich [44]. For the source domain, we use the training + +Table 1. Quantitative comparisons of semantic segmentation performance under adverse conditions on CityScapes [6]→ACDC [45], where the best performance is marked as **bold** and the second best is **underlined**. † denotes the performance of our reproduction. + +
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]72.914.340.516.621.29.317.421.263.823.818.332.619.569.536.234.546.226.936.132.7
\(MGCDA^†\) [44]73.428.769.919.326.336.853.053.375.432.084.651.026.177.643.245.953.932.741.548.7
\(DANNet^†\) [54]84.354.277.638.030.018.941.635.271.339.486.648.729.276.241.643.058.632.643.950.0
DAFormer [16]58.451.384.042.735.150.730.057.074.852.851.358.332.682.758.354.982.444.150.755.4
\(SSAC^†\) [20]76.559.882.752.634.551.258.161.263.958.167.261.540.081.859.261.783.447.555.659.6
HRDA [17]88.357.988.155.236.756.362.965.374.257.785.968.845.788.576.482.487.752.760.468.0
MIC [18]90.867.189.254.540.557.262.068.476.361.887.071.349.489.775.786.889.156.963.070.4
HALO [8]94.279.888.260.251.164.178.265.687.955.795.566.320.788.982.289.387.950.459.071.9
CoDA [12]93.172.790.757.347.456.869.970.087.359.895.471.447.690.377.183.889.154.764.172.6
\(WA^2Net (ours)\)94.577.191.661.947.657.568.570.479.264.095.672.850.290.575.588.790.957.562.673.8
+ +Table 2. Quantitative comparisons of performance on CityScapes [6] $\rightarrow$ NightCity [49]. Considering the NightCity [49] dataset is newly introduced, we reproduce the results of state-of-the-art methods for better comparison. + +
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
DANNet [54]57.536.452.342.539.533.227.846.148.918.467.049.528.357.244.050.443.229.137.444.0
DAFormer [16]64.241.059.647.043.235.731.952.050.320.373.151.829.762.049.858.945.331.239.846.8
SSAC [20]84.546.775.553.149.638.331.048.856.820.281.552.630.275.853.561.146.830.237.548.4
HRDA [17]83.550.578.651.948.340.535.457.858.219.981.856.329.780.565.880.750.431.040.852.2
MIC [18]84.951.678.453.848.641.838.860.557.122.083.460.231.981.464.870.852.237.843.853.8
HALO [8]85.752.177.952.947.239.337.563.255.621.582.161.430.780.563.972.353.439.144.754.2
CoDA [12]86.150.979.151.749.340.136.965.458.323.282.762.833.180.262.574.654.938.345.654.9
WA2Net (ours)87.454.279.354.349.940.642.267.859.422.883.968.234.383.565.176.156.846.448.256.5
+ +split of CityScapes; for the target domain, we use either the test of ACDC/DarkZurich or the validation set of NightCity. (1) CityScapes' training set contains 2,975 images of 19 classes with resolution of $2048 \times 1024$ . (2) ACDC contains weather conditions including fog, nighttime, rain, and snow. The test set contains 2,000 images of $1920 \times 1080$ resolution. (3) NightCity includes 4,297 real night-time images, divided into 2,998 train images and 1,299 val images. (4) DarkZurich contains 2,416 training and 151 test images for nighttime. + +Implementation Details. We adopt Pytorch [40] and Detectron2 [55] to implement the proposed method. 4 NVIDIA GeForce RTX 3090 GPUs are used for training. The extractor of amplitude characteristic is ResNet-18 [15]. During the training stage, our model is trained with a batch size of 4, using the Adam optimizer [31] with an initial learning rate of 0.0001. The input images are rescaled and randomly cropped to $512 \times 512$ following the same data augmentation in DAFormer [16], and the EMA coefficient for updating the teacher net is set to be 0.999. We set the prompt number as $N = 16$ , and the coefficient of the diversity loss as $\lambda_{div} = 0.2$ . + +Evaluation Metric. For a fair comparison, we adopt the + +metric of mean intersection over union (mIOU). It measures the similarity between the prediction and the ground truth by computing the ratio of the intersection area to the union area of the two masks for each object class. + +# 4.2. Comparison with State-of-the-Art Methods + +Our method demonstrates superior performance for semantic segmentation under adverse conditions, outperforming state-of-the-art methods in Table 1, 2& 3. As can be observed, our method achieves significant improvements in challenging classes such as road, traffic light and person. Besides, as shown in Figure 3, our method shows promising segmentation performance in diverse weather scenes. In specific, our method performs great in most scenarios, while MIC [18] may generate confusing activation due to the inability to explicitly perceive weather conditions, our $\mathrm{WA}^2\mathrm{Net}$ can adaptively focus on accurate foreground targets owing to weather-aware aggregation and adaptation. + +# 4.3. Ablation Study + +Effectiveness of the Amplitude Prompt Aggregation. The amplitude prompt aggregation consists of two processes: amplitude characteristic extraction and weather- + +Table 3. Quantitative comparisons of performance on CityScapes [6] $\rightarrow$ DarkZurich [44], where the best performance is marked as bold and the second is underlined. $^\dagger$ denotes the performance of our reproduction of the corresponding model. + +
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]85.837.955.527.714.523.114.021.132.18.72.039.916.664.013.80.058.828.520.729.7
\( MGCDA^† \)[44]80.349.366.27.811.041.438.939.064.118.055.852.153.574.766.00.037.529.122.742.5
\( DANNet^† \)[54]90.054.074.841.021.125.026.830.272.026.284.047.033.968.219.00.366.438.323.644.3
DAFormer [16]93.565.573.339.419.253.344.144.059.534.566.653.452.782.152.79.589.350.538.553.8
HRDA [17]90.456.372.039.519.557.852.743.159.329.170.560.058.684.075.511.290.551.640.955.9
MIC [18]94.875.084.055.128.462.035.552.659.246.870.065.261.782.164.218.591.352.644.060.2
\( HALO^† \)[8]94.574.583.955.528.661.236.052.059.045.570.864.561.281.564.018.290.852.344.260.0
\( CoDA^† \)[12]95.374.285.157.329.161.539.853.458.745.372.666.162.283.065.719.390.553.245.361.4
\( WA^2Net (ours) \)95.777.986.759.631.263.451.955.760.947.276.367.963.184.569.818.191.954.846.162.8
+ +![](images/a139027fda751d9b3b1062f89b17f3ac752b11cf3361acd094660e047642e9ee.jpg) +Figure 3. Qualitative results of our $\mathrm{WA}^2\mathrm{Net}$ and MIC [18] under adverse conditions on ACDC [45], NightCity [49] and DarkZurich [44]. + +Table 4. Ablation on main components of our network. + +
Amplitude Prompt AggregationWeather Heterogeneity AdaptationACDCNightCity
Charac. ExtractionPrompt Agg.Prompt Modu.Homo. Attention
XXXX68.652.4
✓XX✓70.753.7
✓X✓✓71.254.9
✓✓✓✓73.856.5
+ +aware prompt aggregation. To evaluate the effectiveness of the amplitude prompt aggregation, we conduct ablation studies as shown in Table 4. As shown in the $3^{rd}$ row in Table 4, not using prompt aggregation alone means that the prompt bank does not undergo knowledge learning of amplitude characteristics but instead self-initialized and only updated during network training. Without aggregating amplitude characteristic information, prompts cannot perceive weather-aware knowledge, failing to effectively modulate features in the prompt-guided modulation with a significant + +![](images/d3bb0d37b9897b9c6539fd0e44b8aabd104556c36ec121ef3e6c2c821ac193e2.jpg) +Figure 4. Qualitative results on the daytime $\rightarrow$ nighttime setting. + +performance decline. + +Effectiveness of the Weather Heterogeneity Adaptation. + +![](images/b3b6e63ea06dcad60b26e5e96898d8178c07ae106e9c56622edd1175756f10e0.jpg) +Figure 5. Ablation experiments on super-parameters: $\lambda_{div}$ and $N$ . + +![](images/3c794cbebdacf63ba504da70ac30668cd0fc793e71fc500ec69ebd066441fbad.jpg) + +Table 5. Performance w/ and w/o the diversity loss ${\mathcal{L}}_{div}$ . + +
Diversity LossACDC [45]NightCity [49]
w/o L_divw/ L_div
✓72.456.0
✓73.856.5
+ +The weather heterogeneity adaptation includes the prompt-guided modulation and homogenization attention. We assess the impact of the weather heterogeneity adaptation by disabling this process and observing the changes in segmentation accuracy. Simultaneously discarding both prompt aggregation and modulation, as shown in the $2^{nd}$ row of Table 4 (i.e., abandoning the generation and use of prompts), means that the extracted amplitude features are directly fed into the homogenization attention mechanism. This approach results in $f_{a}^{\theta}$ being unable to perceive multi-weather scene characteristics, adversely affecting the model's generalization in extreme scenarios. + +Analysis of the Homogenization Attention. We also analyze the effectiveness of our designed homogenization attention with comparisons against other attention mechanisms as shown in Table 6. The homogenization attention mechanism is designed to address the significant variability in weather conditions by dynamically aligning features from different weather domains. Unlike crossattention, which focuses on explicit interactions between tokens across feature spaces, and masked attention, which selectively attends to spatial regions, our homogenization attention incorporates a domain adaptation strategy that directly mitigates the feature discrepancies caused by adverse weather conditions. As can be observed, the heterogeneity adaptation mechanism proved to be essential to dynamically adapt to varying weather patterns, thereby improving overall performance and reliability. Besides, Figure 6 displays activation maps of different classes with and without weather-aware aggregation and adaptation, illustrating the impact of our design on class-specific segmentation quality. + +Analysis of Hyperparameters. We conducted a comprehensive analysis of hyperparameters to understand their influence on the model's performance. The impact of the choice of diversity loss $\lambda_{div}$ and the prompt number $N$ is shown in Figure 5. Too few prompts hinder the ability to capture rich scene information, while too many may capture + +![](images/bd7dd7d5c66d460fe34e30a6074bb875c83c1e53e6ee5feb2ed7aca241cfaa98.jpg) + +![](images/5d1f1ad9ec2f931d833e7e0b94d7373a07315dd06483ebb5df0164856ec05ec8.jpg) + +![](images/4d149bd444e680be6c14d67f2e65d1b72417dc843867d4116f4c65ddd036e7f2.jpg) + +![](images/8fe8b14f43974e59a52c5762f3295bd2d55378d75c7d85cc650e75583bc552aa.jpg) +Figure 6. Comparison of activation maps for different classes with (w/) and without(w/o) weather-aware aggregation and adaptation. + +Table 6. Comparison of different attention mechanisms with our homogenization attention. + +
Attention MechanismACDC [45]NightCity [49]
cross-attention [52]70.954.1
masked attention [5]71.855.0
Homogenization Attention (ours)73.856.5
+ +irrelevant noise information. It is worth noting that the improvement brought by the diversity loss in Table 5 is more significant when ACDC [45] is the target domain. This is due to ACDC containing more weather diversities (night, rain, snow and fog) compared to the night-only NightCity [49], prompts on ACDC can aggregate more diverse knowledge with $\mathcal{L}_{div}$ . Therefore, the diversity in prompt combinations contributes to diverse scene analysis under multiple weather conditions. + +# 5. Conclusion + +In this paper, we propose a weather-aware aggregation and adaptation network (WA $^2$ Net) for semantic segmentation under adverse conditions. We design the amplitude prompt aggregation to acquire weather-aware characteristic knowledge and weather heterogeneity adaptation to realize interweather homogenization for effective parsing. Extensive experiments demonstrate our effectiveness. + +# Acknowledgments + +This work was partially supported by the National Key Research and Development Program of China (No.2024YFB3909902), National Nature Science Foundation of China (62121002), Youth Innovation Promotion Association of CAS. + +# References + +[1] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2 +[2] Yang Chen, Liang Shan, Yi Qu, Weilong Zhang, and Lu Chang. Amsc: Adaptive masking and structure-constraint learning for domain adaptive semantic segmentation under adverse conditions. IEEE Signal Processing Letters, 2023. 2 +[3] Yujia Chen, Wangkai Li, Zhaoyang Li, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Sam-glomeruli: Enhanced segment anything model for precise glomeruli segmentation. In International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis, pages 182-191. Springer, 2024. 2 +[4] Yujia Chen, Rui Sun, Wangkai Li, Huayu Mai, Naisong Luo, Yuwen Pan, and Tianzhu Zhang. Alleviate and mining: Rethinking unsupervised domain adaptation for mitochondria segmentation from pseudo-label perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2339-2347, 2025. 2 +[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 8 +[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 7 +[7] Ozgur Erkent and Christian Laugier. Semantic segmentation with unsupervised domain adaptation under varying weather conditions for autonomous vehicles. IEEE Robotics and Automation Letters, 5(2):3580-3587, 2020. 1 +[8] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. In International Conference on Machine Learning, pages 13864-13884. PMLR, 2024. 6, 7 +[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3 +[10] Rui Gong, Qin Wang, Martin Danelljan, Dengxin Dai, and Luc Van Gool. Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7225-7235, 2023. 2 +[11] Ziyang Gong, Fuhao Li, Yupeng Deng, Wenjun Shen, Xi-anzheng Ma, Zhenming Ji, and Nan Xia. Train one, generalize to all: Generalizable semantic segmentation from single-scene to all adverse scenes. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2275-2284, 2023. 2 + +[12] Ziyang Gong, Fuhao Li, Yupeng Deng, Deblina Bhattacharjee, Xianzheng Ma, Xiangwei Zhu, and Zhenming Ji. Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning. In European Conference on Computer Vision, pages 130-148. Springer, 2024. 6, 7 +[13] Monica Gruosso, Nicola Capecce, and Ugo Erra. Human segmentation in surveillance video with deep learning. *Multimedia Tools and Applications*, 80(1):1175-1199, 2021. 1 +[14] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. Ppt: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, 2022. 3 +[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6 +[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2, 3, 5, 6, 7 +[17] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision, pages 372-391. Springer, 2022. 6, 7 +[18] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721-11732, 2023. 2, 3, 5, 6, 7 +[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4 +[20] Abdulrahman Kerim, Felipe Chamone, Washington Ramos, Leandro Soriano Marcolino, Erickson R Nascimento, and Richard Jiang. Semantic segmentation under adverse conditions: a weather and nighttime-aware synthetic data-based approach. arXiv preprint arXiv:2210.05626, 2022. 2, 3, 5, 6 +[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3 +[22] Sohyun Lee, Namyup Kim, Sungyeon Kim, and Suha Kwak. Frest: Feature restoration for semantic segmentation under multiple adverse conditions. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 1 +[23] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2 +[24] Wangkai Li, Rui Sun, Bohao Liao, Zhaoyang Li, and Tianzhu Zhang. Balanced learning for domain adaptive semantic segmentation. In *Forty-second International Conference on Machine Learning*. 3 + +[25] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3 +[26] Zhi Li, Xing Wu, Jianjia Wang, and Yike Guo. Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing, 127: 104554, 2022. 1 +[27] Zhaoyang Li, Yuan Wang, Wangkai Li, Rui Sun, and Tianzhu Zhang. Localization and expansion: A decoupled framework for point cloud few-shot semantic segmentation. In European Conference on Computer Vision, pages 18-34. Springer, 2024. 2 +[28] Zhaoyang Li, Yuan Wang, Wangkai Li, Tianzhu Zhang, and Xiang Liu. Dual-agent optimization framework for cross-domain few-shot segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9849-9859, 2025. 2 +[29] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 3 +[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1 +[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6 +[32] Ning Lv, Zenghui Zhang, Cong Li, Jiaxuan Deng, Tao Su, Chen Chen, and Yang Zhou. A hybrid-attention semantic segmentation network for remote sensing interpretation in land-use surveillance. International Journal of Machine Learning and Cybernetics, 14(2):395–406, 2023. 1 +[33] Claudia Menzel, Gregor U Hayn-Leichsenring, Oliver Langner, Holger Wiese, and Christoph Redies. Fourier power spectrum characteristics of face photographs: attractiveness perception depends on low-level image properties. PLoS One, 10(4):e0122801, 2015. 4 +[34] Khan Muhammad, Tanveer Hussain, Hayat Ullah, Javier Del Ser, Mahdi Rezaei, Neeraj Kumar, Mohammad Hijji, Paolo Bellavista, and Victor Hugo C de Albuquerque. Vision-based semantic segmentation in scene understanding for autonomous driving: Recent achievements, challenges, and outlooks. IEEE Transactions on Intelligent Transportation Systems, 23(12):22694-22715, 2022. 1 +[35] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1520-1528, 2015. 1 +[36] Yuwen Pan, Naisong Luo, Rui Sun, Meng Meng, Tianzhu Zhang, Zhiwei Xiong, and Yongdong Zhang. Adaptive template transformer for mitochondria segmentation in electron microscopy images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21474-21484, 2023. 1 +[37] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase + +enhancement for night-time semantic segmentation. In European Conference on Computer Vision, pages 408-424. Springer, 2024. 1 +[38] Yuwen Pan, Rui Sun, Yuan Wang, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Purify then guide: A bidirectional bridge network for open-vocabulary semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1 +[39] Yuwen Pan, Rui Sun, Yuan Wang, Tianzhu Zhang, and Yongdong Zhang. Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2031-2040, 2024. 2 +[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6 +[41] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In 2019 22th International Conference on Information Fusion (FUSION), pages 1-8. IEEE, 2019. 1 +[42] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of fast video-sequence segmentation. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020. +[43] Nikhil Reddy, Abhinav Singhal, Abhishek Kumar, Mahsa Baktashmotlagh, and Chetan Arora. Master of all: Simultaneous generalization of urban-scene segmentation to all adverse weather conditions. In European Conference on Computer Vision, pages 51-69. Springer, 2022. 2 +[44] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3139-3153, 2020. 2, 5, 6, 7 +[45] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1, 3, 5, 6, 7, 8 +[46] Rui Sun, Naisong Luo, Yuan Wang, Yuwen Pan, Huayu Mai, Zhe Zhang, and Tianzhu Zhang. 1st place solution for youtuves challenge 2022: Video object segmentation. In Proceedings of the Conference on Computer Vision and Pattern Recognition Workshops, 2022. 3 +[47] Rui Sun, Naisong Luo, Yuwen Pan, Huayu Mai, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Appearance prompt vision transformer for connectome reconstruction. In IJCAI, pages 1423-1431, 2023. 1 +[48] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: exploring the better weighting function for semi-supervised semantic segmentation. Advances in neural information processing systems, 36:61792-61805, 2023. 2 +[49] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson WH Lau. Night-time scene parsing with a large + +real dataset. IEEE Transactions on Image Processing, 30: 9085-9098, 2021. 5, 6, 7, 8 +[50] Yingjie Tian and Siyu Zhu. Partial domain adaptation on semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3798-3809, 2021. 2 +[51] Abhinav Valada, Johan Vertens, Ankit Dhall, and Wolfram Burgard. Adapnet: Adaptive semantic segmentation in adverse environmental conditions. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4644-4651. IEEE, 2017. 2 +[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 8 +[53] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 6, 7 +[54] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2, 3, 6, 7 +[55] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6 +[56] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5 +[57] Xin Yang, Wending Yan, Yuan Yuan, Michael Bi Mi, and Robby T Tan. Semantic segmentation in multiple adverse weather conditions with domain knowledge retention. arXiv preprint arXiv:2401.07459, 2024. 3 +[58] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 2, 4 +[59] Chunhui Zhang, Guanjie Huang, Li Liu, Shan Huang, Yinan Yang, Xiang Wan, Shiming Ge, and Dacheng Tao. Webuav-3m: A benchmark for unveiling the power of million-scale deep uav tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9186-9205, 2022. 1 +[60] Chunhui Zhang, Li Liu, Guanjie Huang, Hao Wen, Xi Zhou, and Yanfeng Wang. Webuot-1m: Advancing deep underwater object tracking with a million-scale benchmark. Advances in Neural Information Processing Systems, 37:50152-50167, 2024. 1 +[61] Chunhui Zhang, Li Liu, Hao Wen, Xi Zhou, and Yanfeng Wang. Mambatrack: Exploiting dual-enhancement for night uav tracking. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3 \ No newline at end of file diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/images.zip b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..a948310793b843bf1a2f86a9a526546efcdeeec4 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:838378676791754c3075d9ced7f9445135cb2a58325d12b98c92d17867da833b +size 846573 diff --git a/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/layout.json b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..48f8e06d9d51ed4fe2757e5d691c02ce331edb33 --- /dev/null +++ b/2025/Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions/layout.json @@ -0,0 +1,9405 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 124, + 103, + 487, + 140 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 124, + 103, + 487, + 140 + ], + "spans": [ + { + "bbox": [ + 124, + 103, + 487, + 140 + ], + "type": "text", + "content": "Exploring Weather-aware Aggregation and Adaptation for Semantic Segmentation under Adverse Conditions" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "spans": [ + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "text", + "content": "Yuwen Pan" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "text", + "content": " Rui Sun" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "text", + "content": " Wangkai Li" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "text", + "content": " Tianzhu Zhang" + }, + { + "bbox": [ + 157, + 160, + 441, + 175 + ], + "type": "inline_equation", + "content": "^{1,2*}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 187, + 175, + 423, + 189 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 187, + 175, + 423, + 189 + ], + "spans": [ + { + "bbox": [ + 187, + 175, + 423, + 189 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 85, + 190, + 523, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 85, + 190, + 523, + 204 + ], + "spans": [ + { + "bbox": [ + 85, + 190, + 523, + 204 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 85, + 190, + 523, + 204 + ], + "type": "text", + "content": "National Key Laboratory of Deep Space Exploration, Deep Space Exploration Laboratory" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "spans": [ + { + "bbox": [ + 133, + 205, + 474, + 217 + ], + "type": "text", + "content": "{panyw, issunrui, lwklwk}@mail.ustc.edu.cn, tzzhang@ustc.edu.cn" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "spans": [ + { + "bbox": [ + 151, + 244, + 200, + 257 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 270, + 297, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 297, + 617 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 297, + 617 + ], + "type": "text", + "content": "Semantic segmentation under adverse conditions is critical for reliable visual perception in challenging weather environments. These extreme scenarios introduce distortions, such as low contrast and reduced visibility, making traditional segmentation models struggle. The scarcity of labeled data in such conditions makes it difficult to train models directly for these environments. Unsupervised domain adaptation (UDA) has been proposed as a solution to transfer knowledge from labeled source domains (normal weather) to unlabeled target domains (adverse weather). However, existing methods face significant challenges, particularly due to weather unawareness and feature heterogeneity. Many models fail to account for the unique characteristics of different weather conditions, and the significant feature discrepancies between normal and adverse weather images hinder effective adaptation. In this paper, we propose a novel weather-aware aggregation and adaptation network that leverages characteristic knowledge to achieve weather homogenization and enhance scene perception. Specifically, we introduce amplitude prompt aggregation to capture essential characteristics from the Fourier frequency domain that are indicative of different weather conditions. Additionally, we employ weather heterogeneity adaptation to mitigate the inter-domain heterogeneity, thereby achieving feature homogenization across diverse environments. Extensive experimental results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 639, + 135, + 651 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 296, + 696 + ], + "type": "text", + "content": "Semantic segmentation involves assigning a class label to each pixel in an image [30, 35, 36, 38, 47], which plays a crucial role in various real-world applications such as au" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 334, + 243, + 537, + 344 + ], + "blocks": [ + { + "bbox": [ + 334, + 243, + 537, + 344 + ], + "lines": [ + { + "bbox": [ + 334, + 243, + 537, + 344 + ], + "spans": [ + { + "bbox": [ + 334, + 243, + 537, + 344 + ], + "type": "image", + "image_path": "4014c906c279a86dcddccedca7e5c2ec5874bdba89befc81d7f9d53ca8590a3a.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 374, + 344, + 495, + 353 + ], + "lines": [ + { + "bbox": [ + 374, + 344, + 495, + 353 + ], + "spans": [ + { + "bbox": [ + 374, + 344, + 495, + 353 + ], + "type": "text", + "content": "(a) Traditional Cross-weather Paradigm" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 334, + 354, + 536, + 471 + ], + "blocks": [ + { + "bbox": [ + 334, + 354, + 536, + 471 + ], + "lines": [ + { + "bbox": [ + 334, + 354, + 536, + 471 + ], + "spans": [ + { + "bbox": [ + 334, + 354, + 536, + 471 + ], + "type": "image", + "image_path": "eb4ce93c6987c17e3eba03460d11ee818b4e92033b34bd47c7584f7a3f8bc3f1.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 352, + 472, + 518, + 481 + ], + "lines": [ + { + "bbox": [ + 352, + 472, + 518, + 481 + ], + "spans": [ + { + "bbox": [ + 352, + 472, + 518, + 481 + ], + "type": "text", + "content": "(b) Weather-aware Aggregation and Adaptation (ours)" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 490, + 555, + 545 + ], + "lines": [ + { + "bbox": [ + 313, + 490, + 555, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 490, + 555, + 545 + ], + "type": "text", + "content": "Figure 1. Comparison of the traditional cross-weather paradigm and our weather-aware aggregation and adaptation. To mitigate the existing issues (unawareness of weather characteristics and severe cross-weather heterogeneity), we aim to aggregate weather-aware attribute knowledge and homogenize domain discrepancies." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "spans": [ + { + "bbox": [ + 313, + 568, + 555, + 688 + ], + "type": "text", + "content": "tonomous driving [34, 37, 59] and surveillance [13, 32, 60]. However, the performance of segmentation models often degrades significantly when faced with adverse weather conditions [22, 45], including night, fog, rain, and snow. The deterioration in visibility and scene appearance poses a substantial challenge for traditional segmentation algorithms, which are typically trained on datasets collected under normal weather conditions. Therefore, exploring an effective way for semantic parsing under extreme weather conditions is highly anticipated." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 556, + 714 + ], + "type": "text", + "content": "Extensive research efforts have been devoted to semantic segmentation in adverse weather conditions [7, 26, 41-" + } + ] + } + ], + "index": 17 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 144, + 712 + ], + "type": "text", + "content": "*Corresponding author." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13952" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 275 + ], + "type": "text", + "content": "43, 51]. Due to the scarcity of labeled data in such adverse conditions, existing methods strive to employ an unsupervised domain adaptation (UDA) paradigm utilizing both labeled normal weather data and images without annotations under extreme scenarios to address the challenge [16, 44, 50, 53, 54]. Kerim et al. [20] propose a dual-branch network based on DeepLabV3+ [1], enhancing training knowledge for extreme weather conditions by incorporating a synthetic multi-scene dataset. However, discerning similar-looking categories in the target domain with limited real annotations remains a challenge for domain adaptation methods. To address this, MIC [18] introduces the masked image consistency, leveraging spatial context relations in the target domain as additional cues to enhance the robustness of UDA. The emergence of these methods has benefited the understanding of scenes in extreme weather conditions." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 278, + 298, + 686 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 278, + 298, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 278, + 298, + 686 + ], + "type": "text", + "content": "Despite considerable progress, the task of distinguishing and analyzing scenes under different extreme weather conditions remains challenging. Whether through the assistance of synthetic data or exploration of spatial context clues, existing methods struggle with accurately identifying weather-aware semantics in multiple extreme scenarios. We attribute the predicament of this cross-weather paradigm to two inherent issues (Figure 1): (1) Unawareness of Weather Characteristics. The style characteristics under adverse weathers, such as illumination conditions and contrast, are primary culprits responsible for the inter-weather gap among diverse scenes. Although recognizing this crucial aspect, existing works [10, 58] generally employ low-level enhancements or implicit learning for inter-domain style transfer, failing to fully exploit weather attribute representation. Therefore, designing a strategy to aggregate weather-aware scene knowledge across diverse conditions is worthy of exploration. (2) Severe Cross-Weather Heterogeneity. Early works [16, 18] tend to scene images from different weather conditions uniformly, ignoring the substantial differences across different weather types. The significant heterogeneity across weather conditions presents another challenge, stemming from the diverse and complex nature of weather phenomena that introduce distinct visual distortions and scene alterations. Even though some methods [2, 11] attempt to address this challenge, the significant differences in characteristics across different weather conditions still pose considerable difficulties. These variations lead to a significant disparity between features learned from the normal weather and those required for effective interpretation in adverse scenes. Addressing how to leverage characteristic knowledge to achieve weather homogenization to facilitate scene perception in adverse scenes is an essential question." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Motivated by the above discussions, we propose a Weather-aware Aggregation and Adaptation NETwork" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "text", + "content": "(WA" + }, + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "text", + "content": "Net) to aggregate weather-aware attribute knowledge and homogenize domain discrepancies, including amplitude prompt aggregation and weather heterogeneity adaptation. In the amplitude prompt aggregation, to acquire weather-aware characteristic knowledge, we utilize the amplitude operation of Fourier transform to capture essential characteristics indicative of different weather conditions. These amplitude features serve as discriminative cues for weather-related variations in the scene. To efficiently aggregate the crucial characteristic knowledge from the frequency domain into weather-adaptive prompts, we prepend a set of learnable prompts in the prompt bank to interact with frequency features by cross-attention mechanism to obtain weather-aware prompts. These prompts encapsulate the intrinsic characteristics of various weather conditions, providing essential guidance for the adaptation process. In the weather heterogeneity adaptation, to realize inter-weather homogenization for effective semantic parsing, the resultant weather-aware prompts are leveraged to generate adapted characteristics after modulation. Then, we propose a novel homogenization attention mechanism to acquire homogenized features across diverse weather conditions. This mechanism identifies the discrepancy between features of the original image and adapted characteristics, decoupling weather characteristics from valuable target information and thereby filtering out attribute interference. In this way, our WA" + }, + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 72, + 555, + 430 + ], + "type": "text", + "content": "Net achieves effective weather-aware knowledge aggregation and robust adaptation, ensuring improved performance for semantic segmentation in adverse weather conditions." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 432, + 555, + 454 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 432, + 555, + 454 + ], + "spans": [ + { + "bbox": [ + 313, + 432, + 555, + 454 + ], + "type": "text", + "content": "To sum up, our contributions can be summarized as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 457, + 553, + 623 + ], + "type": "list", + "angle": 0, + "index": 8, + "blocks": [ + { + "bbox": [ + 314, + 457, + 553, + 517 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 457, + 553, + 517 + ], + "spans": [ + { + "bbox": [ + 314, + 457, + 553, + 517 + ], + "type": "text", + "content": "- We propose a novel weather-aware aggregation and adaptation network for semantic segmentation under adverse conditions. Our paper exploits a practical way to acquire weather-adaptive knowledge and enhance cross-weather perception via an aggregation and adaptation manner." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "spans": [ + { + "bbox": [ + 314, + 517, + 553, + 576 + ], + "type": "text", + "content": "- We design the amplitude prompt aggregation to acquire weather-aware prompts from weather characteristics through amplitude extraction, and the weather heterogeneity adaptation to inherit weather-aware knowledge and mitigate the inter-weather heterogeneity." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 577, + 553, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 577, + 553, + 623 + ], + "spans": [ + { + "bbox": [ + 314, + 577, + 553, + 623 + ], + "type": "text", + "content": "- Extensive experiment results on multiple challenging benchmarks demonstrate that our method achieves consistent improvements for semantic segmentation under adverse conditions." + } + ] + } + ], + "index": 7 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "spans": [ + { + "bbox": [ + 314, + 638, + 400, + 651 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 553, + 683 + ], + "type": "text", + "content": "2.1. Semantic Segmentation under Adverse Conditions" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 714 + ], + "type": "text", + "content": "The field of semantic segmentation [3, 4, 23, 27, 28, 39, 48] has made significant advancements with the emergence" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13953" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 396 + ], + "type": "text", + "content": "of convolutional neural networks (CNNs) [21] and transformers [52]. However, traditional semantic segmentation models experience a notable performance drop under adverse conditions such as rain, fog and night-time environments [45, 46, 61]. Therefore, semantic segmentation under adverse conditions aims to develop robust models capable of accurately delineating objects and regions in challenging environments. Upon the scarcity of labeled data in extreme weathers, the unsupervised domain adaptation paradigm is utilized to enhance the generalization of models across different weather conditions with the aid of normal weather data [24, 54]. Kerim et al. [20] propose a dual-branch network and expand the training dataset for extreme weather conditions by integrating a synthetic multiscene dataset. Nevertheless, domain adaptation methods encounter challenges due to the difficulty in distinguishing similar-looking categories in the target domain with limited real annotations. To tackle this issue, MIC [18] introduces Masked Image Consistency, utilizing spatial context relations in the target domain as additional cues to enhance vision perception. Additionally, to mitigate knowledge forgetting in new scenes, Yang et al. [57] introduce adaptive knowledge acquisition and replay of weather-condition combinations, aiming to alleviate model-forgetting issues in diverse scenes. Differently, we strive to tackle the challenge of weather-unaware perception, leveraging weather-aware prompts in an aggregation and adaptation manner." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 404, + 162, + 417 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 404, + 162, + 417 + ], + "spans": [ + { + "bbox": [ + 55, + 404, + 162, + 417 + ], + "type": "text", + "content": "2.2. Prompt Learning" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 422, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 296, + 579 + ], + "type": "text", + "content": "Prompting [29] incorporates task-specific instructions into input text to guide model behavior, enhancing performance without retraining, particularly in few-shot learning [9]. Recent work extends prompting to vision tasks, where learnable prompts enable adaptation without tuning all parameters [25] and improve segmentation generalization [14]. For semantic segmentation under adverse weather, weather-adaptive prompts modulate visual embeddings to mitigate distortions like lighting and visibility changes, enhancing robustness without extensive fine-tuning. Integrating weather-specific knowledge into prompts improves generalization across diverse conditions, making models more effective in real-world applications." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 591, + 111, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 591, + 111, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 591, + 111, + 604 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 611, + 142, + 623 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 142, + 623 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 142, + 623 + ], + "type": "text", + "content": "3.1. Preliminaries" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": "Typically, in the field of unsupervised domain adaptation [16, 18], the source domain data can be formulated as " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "D^{s} = \\{X^{s}, Y^{s}\\}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "X^{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "Y^{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " represent the corresponding image and label space, respectively. Each source domain image " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "x_{k}^{s} \\in X^{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " is associated with a ground truth segmentation mask " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "y_{k}^{s} \\in Y^{s}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": ". Similarly, let " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "X^{t}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " be the target domain image space, and " + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "inline_equation", + "content": "Y^{t}" + }, + { + "bbox": [ + 55, + 629, + 296, + 715 + ], + "type": "text", + "content": " be the label space. However," + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": "unlike the source domain, the target domain has no access to annotated segmentation masks. The supervised loss " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^s" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " can be formulated with a neural network " + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "inline_equation", + "content": "f_{\\mu}" + }, + { + "bbox": [ + 313, + 72, + 554, + 109 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 366, + 119, + 555, + 152 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 366, + 119, + 555, + 152 + ], + "spans": [ + { + "bbox": [ + 366, + 119, + 555, + 152 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {s} = \\frac {1}{N _ {s}} \\sum_ {k = 1} ^ {N _ {s}} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {s}\\right), y _ {k} ^ {s}\\right), \\tag {1}", + "image_path": "2ceca08ccc8a085394d96cd93451f3d8c2c5bcc4ed97f3237540748168665350.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "spans": [ + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "inline_equation", + "content": "N_{s}" + }, + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "text", + "content": " denotes the image number of the source domain, " + }, + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{ce}" + }, + { + "bbox": [ + 313, + 161, + 554, + 186 + ], + "type": "text", + "content": " denotes the pixel-wise cross-entropy loss as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 347, + 194, + 555, + 228 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 347, + 194, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 347, + 194, + 555, + 228 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c e} (\\hat {y}, y) = - \\sum_ {h = 1} ^ {H} \\sum_ {w = 1} ^ {W} \\sum_ {c = 1} ^ {C} y _ {h w c} \\log \\hat {y} _ {h w c}. \\tag {2}", + "image_path": "a056dd20a1b18bac7ed5f3822cfeaadae2efbaf213c41da279184cb4825bffd4.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": "The model trained on the source domain " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "D^{s}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " normally exhibits performance degradation when tested on the target domain " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "D^{t}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": ". Therefore, the self-training paradigm generates pseudo-labels for the target domain images " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "X^{t}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " with a teacher network " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "g_{\\phi}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " instead of the student network " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "f_{\\mu}" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": ". Specifically, the model predicts segmentation masks for unlabeled target domain images, producing pseudo-labels " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\hat{y}_k^t" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " for each image " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "x_k^t" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " in the target domain. The unsupervised loss " + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^t" + }, + { + "bbox": [ + 313, + 238, + 555, + 346 + ], + "type": "text", + "content": " for the target domain is formulated as:" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 361, + 356, + 554, + 389 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 356, + 554, + 389 + ], + "spans": [ + { + "bbox": [ + 361, + 356, + 554, + 389 + ], + "type": "interline_equation", + "content": "\\mathcal {L} ^ {t} = \\frac {1}{N _ {t}} \\sum_ {k = 1} ^ {N _ {t}} \\lambda^ {t} \\mathcal {L} _ {c e} \\left(f _ {\\mu} \\left(x _ {k} ^ {t}\\right), \\hat {y} _ {k} ^ {t}\\right), \\tag {3}", + "image_path": "8cc6d283e61f9560d09a96c52bbeb555be2794d691f1b834f9cef74354c9bfe2.jpg" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "spans": [ + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "inline_equation", + "content": "N_{t}" + }, + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "text", + "content": " denotes the image number of the target domain, " + }, + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "inline_equation", + "content": "\\lambda^t" + }, + { + "bbox": [ + 313, + 398, + 554, + 483 + ], + "type": "text", + "content": " is a weight coefficient for the unsupervised loss. The student network updates the parameters of the teacher network through backpropagation, utilizing the teacher network's output as targets to minimize prediction errors, while also employing Exponential Moving Average (EMA) update to stabilize the training process and improve convergence." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 491, + 382, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 382, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 382, + 502 + ], + "type": "text", + "content": "3.2. Overview" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": "As shown in Figure 2, given an input image " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "x^{s}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " from the source domain and " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "x^{t}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " from the target domain under adverse conditions, we denote the image under different weather condition " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " as " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "x^{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " for clarity, where " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\theta \\in (0, \\dots, \\Theta)" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\Theta" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " denotes the total number of adverse weather conditions (" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "x^{0}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " refers to the source domain image under the normal weather condition). The Fourier amplitude features " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{a}^{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " are decoupled from the frequency domain. Then, a set of prompts, i.e., prompt bank " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": ", is learned during the prompt aggregation process. The acquired weather-aware prompts " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " further interact with amplitude features " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{a}^{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " to evolve into adapted characteristic " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "\\hat{f}_{a}^{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": ". Then, to achieve homogenization for features from diverse weather conditions, we further employ homogenization attention to mitigate the influence of weather heterogeneity. Finally, the modulated features " + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "inline_equation", + "content": "f_{h}^{\\theta}" + }, + { + "bbox": [ + 313, + 509, + 555, + 714 + ], + "type": "text", + "content": " are fed into the teacher-student network for comprehensive training." + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13954" + } + ] + } + ], + "index": 15 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 70, + 556, + 205 + ], + "blocks": [ + { + "bbox": [ + 58, + 70, + 556, + 205 + ], + "lines": [ + { + "bbox": [ + 58, + 70, + 556, + 205 + ], + "spans": [ + { + "bbox": [ + 58, + 70, + 556, + 205 + ], + "type": "image", + "image_path": "61fe4643d51294a0265b4d8679594ece20fc2870f68416906e2d33b97ff393a2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 212, + 555, + 235 + ], + "lines": [ + { + "bbox": [ + 55, + 212, + 555, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 212, + 555, + 235 + ], + "type": "text", + "content": "Figure 2. Illustration of our proposed " + }, + { + "bbox": [ + 55, + 212, + 555, + 235 + ], + "type": "inline_equation", + "content": "\\mathrm{WA}^2\\mathrm{Net}" + }, + { + "bbox": [ + 55, + 212, + 555, + 235 + ], + "type": "text", + "content": ". Our network contains the amplitude prompt aggregation to acquire weather-aware prompts (Section 3.3) and weather heterogeneity adaptation to realize feature homogenization (Section 3.4)." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 255, + 228, + 268 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 228, + 268 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 228, + 268 + ], + "type": "text", + "content": "3.3. Amplitude Prompt Aggregation" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 272, + 296, + 356 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 272, + 296, + 356 + ], + "spans": [ + { + "bbox": [ + 55, + 272, + 296, + 356 + ], + "type": "text", + "content": "In order to disentangle the latent weather characteristics and target information fused in the RGB images, we first model the representation of weather characteristics by exploring the Fourier frequency domain. Then, we collect the weather characteristic properties in Fourier amplitude features and acquire weather-aware prompts during the prompt aggregation process." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "spans": [ + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "text", + "content": "Amplitude Characteristic Extraction. In the frequency domain, it is known that the amplitude component of Fourier spectrum preserves low-level statistics information [33, 58], which reflects the energy distribution of the frequency components of the image and can be understood as a representation of the \"characteristics\" or \"style\" of the image. Therefore, we can utilize the amplitude of Fourier spectrum to extract weather characteristic knowledge in diverse scenes. Specifically, in the amplitude extractor, given an image " + }, + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "inline_equation", + "content": "x^{\\theta} \\in \\mathbb{R}^{H \\times W \\times 3}" + }, + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "text", + "content": ", we apply a two-dimensional Fourier transform " + }, + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 55, + 357, + 296, + 488 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 86, + 502, + 295, + 536 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 86, + 502, + 295, + 536 + ], + "spans": [ + { + "bbox": [ + 86, + 502, + 295, + 536 + ], + "type": "interline_equation", + "content": "\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} = \\sum_ {h = 0} ^ {H - 1} \\sum_ {w = 0} ^ {W - 1} x _ {h, w} ^ {\\theta} e ^ {- J 2 \\pi \\left(\\frac {w h}{H} + \\frac {v w}{W}\\right)}, \\tag {4}", + "image_path": "e916bcb4370fc9ef121c664a7e12d6a03dc69a69d4aacbeab59809821f4a0d6b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "spans": [ + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "text", + "content": " refers to the imaginary unit. Then we can acquire the corresponding amplitude " + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\mathcal{A}" + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "text", + "content": " and phase " + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 540, + 295, + 563 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 123, + 567, + 294, + 583 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 123, + 567, + 294, + 583 + ], + "spans": [ + { + "bbox": [ + 123, + 567, + 294, + 583 + ], + "type": "interline_equation", + "content": "\\mathcal {A} \\left(x ^ {\\theta}\\right) _ {u, v} = \\left| \\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v} \\right|, \\tag {5}", + "image_path": "54453cc7d610db622e80a4468f0f0192ace7e620eff1bc4da32c0c476a6d4a24.jpg" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 60, + 586, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 586, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 60, + 586, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\Phi \\left(x ^ {\\theta}\\right) _ {u, v} = \\arg \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u, v}\\right) = \\arctan \\left[ \\frac {\\operatorname {I m} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)}{\\operatorname {R e} \\left(\\mathcal {F} \\left(x ^ {\\theta}\\right) _ {u , v}\\right)} \\right], \\tag {6}", + "image_path": "979250d8f8a542e700eb6975413e33a58b9293ee74503c081ec94e3add96b3fe.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\arg" + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "content": " represents the argument (or phase angle) of the complex number, Im and Re represent the imaginary and real part of " + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)" + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "content": ", respectively. In this way, to finally generate our amplitude characteristic map, we fix the phase to an average constant " + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "inline_equation", + "content": "c^p" + }, + { + "bbox": [ + 55, + 624, + 295, + 696 + ], + "type": "text", + "content": " and apply inverse Fourier transform to acquire the amplitude characteristic map as" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 109, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "\\overline {{\\mathcal {A}}} (x ^ {\\theta}) = \\mathcal {F} ^ {- 1} [ \\mathcal {A} (x ^ {\\theta}) _ {u, v} e ^ {- J c ^ {p}} ], \\tag {7}", + "image_path": "865decc0090a161275b58b9f747347a48ce791739595938582d05f3b0252e1a2.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathcal{F}^{-1}" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": " denotes the inverse Fourier transformation. Then, we can obtain the amplitude characteristics " + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a = \\{f_a^\\theta\\}_{\\theta=0}^\\Theta" + }, + { + "bbox": [ + 313, + 255, + 555, + 316 + ], + "type": "text", + "content": " through a shared-weight lightweight encoder (e.g., ResNet-18 [15]), preparing for the following aggregation of weather-aware prompts." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": "Weather-aware Prompt Aggregation. To effectively aggregate characteristic information across different weather conditions, we aim to develop a medium capable of absorbing cross-weather knowledge. Inspired by [19], we introduce a set of prompts, i.e., prompt bank " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{\\pmb{p}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n \\in \\mathbb{R}^{1 \\times L}" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": " denotes the number of prototypes. These learnable prompts will be leveraged to extract characteristic knowledge by interaction with amplitude features " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "f_a^\\theta" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": ", which highly represents the style condition of the current scene. In specific, for the acquired amplitude characteristic " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "f_a^\\theta" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": ", the queries arise from the prompts " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": ", and keys and values arise from the amplitude feature " + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a = \\{f_a^0, \\dots, f_a^\\theta, \\dots, f_a^\\Theta\\}" + }, + { + "bbox": [ + 313, + 316, + 555, + 473 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 341, + 481, + 555, + 497 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 481, + 555, + 497 + ], + "spans": [ + { + "bbox": [ + 341, + 481, + 555, + 497 + ], + "type": "interline_equation", + "content": "\\mathbf {Q} _ {n} = \\boldsymbol {p} _ {n} \\mathbf {W} ^ {Q}, \\mathbf {K} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {K}, \\mathbf {V} ^ {\\theta} = f _ {a} ^ {\\theta} \\mathbf {W} ^ {V}, \\tag {8}", + "image_path": "b7276c4df70ec4650bd3ae7ea76ceb2d0fa203701322c657d482fab71c2878cb.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "spans": [ + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "inline_equation", + "content": "n\\in [1,\\dots ,N]" + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "text", + "content": " , and " + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^Q\\in \\mathbb{R}^{C\\times C_k}" + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^K\\in" + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbb{R}^{C\\times C_k}" + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "inline_equation", + "content": "\\mathbf{W}^V\\in \\mathbb{R}^{C\\times C_v}" + }, + { + "bbox": [ + 313, + 505, + 554, + 555 + ], + "type": "text", + "content": " are linear projections. The attention weights are calculated based on the dot-product similarity between each query and key:" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 344, + 564, + 554, + 597 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 344, + 564, + 554, + 597 + ], + "spans": [ + { + "bbox": [ + 344, + 564, + 554, + 597 + ], + "type": "interline_equation", + "content": "s _ {n, \\theta} = \\frac {\\exp (\\beta_ {n , \\theta})}{\\sum_ {\\theta = 0} ^ {\\Theta} \\exp (\\beta_ {n , \\theta})}, \\beta_ {n, \\theta} = \\frac {\\mathbf {Q} _ {n} \\mathbf {K} ^ {\\theta} {} ^ {\\top}}{\\sqrt {C _ {k}}}, \\tag {9}", + "image_path": "2a0e4a97c28357ddcba64537014ae27e2b78c8bac4110448f648e2a3e868291c.jpg" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 321, + 609, + 553, + 651 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 609, + 553, + 651 + ], + "spans": [ + { + "bbox": [ + 321, + 609, + 553, + 651 + ], + "type": "interline_equation", + "content": "\\hat {\\boldsymbol {p}} _ {n} = \\operatorname {F F N} (\\operatorname {A t t} \\left(\\mathbf {Q} _ {n}, \\mathbf {K}, \\mathbf {V}\\right)) = \\operatorname {F F N} \\left(\\sum_ {\\theta = 0} ^ {\\Theta} s _ {n, \\theta} \\mathbf {V} ^ {\\theta}\\right), \\tag {10}", + "image_path": "91d707c20b295e254c98ede004dd83068fc42ca40eb8ed7e4c69720d405c3351.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathrm{FFN}()" + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "text", + "content": " denotes the feed forward network. In this way, weather-aware prompts " + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}} = \\{\\hat{p}_n\\}_{n=1}^N" + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "text", + "content": " can distill rich characteristic knowledge from the amplitude features " + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "inline_equation", + "content": "\\mathbf{F}_a" + }, + { + "bbox": [ + 313, + 653, + 553, + 689 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 554, + 713 + ], + "type": "text", + "content": "Besides, in order to enable prompts to carry diverse and comprehensive knowledge of various weather conditions," + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13955" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": "we apply the diversity loss on the prompt bank " + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 72, + 295, + 85 + ], + "type": "text", + "content": ". Formally," + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 66, + 91, + 295, + 126 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 91, + 295, + 126 + ], + "spans": [ + { + "bbox": [ + 66, + 91, + 295, + 126 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {d i v} = \\frac {1}{N (N - 1)} \\sum_ {i = 1} ^ {N} \\sum_ {j = 1, i \\neq j} ^ {N} \\left(\\frac {\\langle \\boldsymbol {p} _ {i} , \\boldsymbol {p} _ {j} \\rangle}{\\| \\boldsymbol {p} _ {i} \\| _ {2} \\| \\boldsymbol {p} _ {j} \\| _ {2}}\\right). \\tag {11}", + "image_path": "cd0e7b213277ffceb49952ffd6d54ed347e444f32d13152054884009f3d2499c.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 133, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 133, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 133, + 296, + 277 + ], + "type": "text", + "content": "Applying diversity loss to different prompts offers the advantage of encouraging each prompt " + }, + { + "bbox": [ + 55, + 133, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\pmb{p}_n" + }, + { + "bbox": [ + 55, + 133, + 296, + 277 + ], + "type": "text", + "content": " to focus on distinct attributes of weather conditions, thereby preventing redundancy in aggregation learning. Diversity loss enhances weather-specific prompts by encouraging each to capture unique traits. Integrating the prompt bank and diversity loss within domain adaptation allows the model to adapt to varied weathers. Splitting this into two tasks is limited by the lack of direct supervision on weather-adaptive prompts; only foreground masks from the source domain are available. Diversity loss thus broadens the prompt bank's coverage of style features for better domain adaptation." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 283, + 243, + 296 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 283, + 243, + 296 + ], + "spans": [ + { + "bbox": [ + 55, + 283, + 243, + 296 + ], + "type": "text", + "content": "3.4. Weather Heterogeneity Adaptation" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "spans": [ + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "text", + "content": "In adverse weather conditions, weather-induced distortions such as glare, fog, and raindrops introduce complexities that can degrade the performance of vision perception. To address the weather heterogeneity challenges posed by varying weather conditions, we introduce the weather heterogeneity adaptation, which utilizes acquired weather-adaptive prompts " + }, + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 300, + 296, + 408 + ], + "type": "text", + "content": " as references to modulate the weather characteristic in the current scene and homogenizes weather features via homogenization attention." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 408, + 296, + 503 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 408, + 296, + 503 + ], + "spans": [ + { + "bbox": [ + 55, + 408, + 296, + 503 + ], + "type": "text", + "content": "Prompt-guided Modulation. In fact, the introduction of the prompt bank can be deemed as a guidepost in artificially solving incidents, i.e., guiding information, which absorbs weather-aware knowledge across various scenes and can transfer reliable weather-aware properties into modulated characteristic features. For the acquired weather amplitude feature " + }, + { + "bbox": [ + 55, + 408, + 296, + 503 + ], + "type": "inline_equation", + "content": "f_{a}^{\\theta}" + }, + { + "bbox": [ + 55, + 408, + 296, + 503 + ], + "type": "text", + "content": ", the modulated characteristic can be formulated as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 121, + 509, + 295, + 537 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 121, + 509, + 295, + 537 + ], + "spans": [ + { + "bbox": [ + 121, + 509, + 295, + 537 + ], + "type": "interline_equation", + "content": "\\hat {f} _ {a} ^ {\\theta} = \\operatorname {S o f t m a x} \\left(\\frac {f _ {a} ^ {\\theta} \\hat {\\mathbf {P}} ^ {\\top}}{\\sqrt {C}}\\right) \\hat {\\mathbf {P}}. \\tag {12}", + "image_path": "814dfbf3402f9d9b58dff6194de41a68f905e212ea9d87870f034fe7dc615fe0.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "content": "In this way, the modulated characteristic " + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\hat{f}_a^\\theta" + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "content": " inherits rich cross-weather knowledge from the weather-aware prompts " + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "inline_equation", + "content": "\\hat{\\mathbf{P}}" + }, + { + "bbox": [ + 55, + 545, + 295, + 593 + ], + "type": "text", + "content": ", thereby evolving into an adapted representation poised for subsequent homogenization." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "Homogenization Attention Mechanism. With the acquired representation of modulated characteristic " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}_a^\\theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ", we first extract the vision features of image " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "x^{\\theta}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " from the RGB domain through feature encoding as " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_v^\\theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". To homogenize vision features from diverse weather conditions, we aim to find an appropriate way to mitigate the influence of " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\hat{f}_a^\\theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " on " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "f_v^\\theta" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": ". Inspired by the success of Transformer [52] architecture in discovering local regions, we further explore the potential of attention mechanism in characteristic homogenization for different weather conditions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": "Given an image feature " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "f_{v}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times c}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " extracted from the vision feature extractor (e.g., MiT-B5 [56]), we use the extracted " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\hat{f}_{a}^{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " to remove the weather characteristics through a novel homogenization attention mechanism. First, we employ two convolution layers to map " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "f_{v}^{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\hat{f}_{a}^{\\theta}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " to the same dimension " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "C" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": ", then we can obtain " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\overline{\\mathbf{f}}_{\\mathbf{v}}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\overline{f}_{a}^{\\theta} \\in \\mathbb{R}^{h \\times w \\times C}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": ", respectively. The disparity between image features and weather characteristics signifies the saliency details of the target objects, thereby aiding in accentuating the relevant information about the targets. Rather than vanilla transformer attention mechanism [52], we custom design the homogenization attention to reduce heterogeneity in a finer manner, prompting the vision system to highlight salient regions across diverse scenes. Specifically, we acquire the homogenized map " + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "inline_equation", + "content": "\\mathcal{H} \\in \\mathbb{R}^{h \\times w}" + }, + { + "bbox": [ + 313, + 72, + 555, + 269 + ], + "type": "text", + "content": " from pixel-level features and weather characteristic as:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 370, + 272, + 555, + 306 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 272, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 370, + 272, + 555, + 306 + ], + "type": "interline_equation", + "content": "\\mathcal {H} _ {i, j} ^ {\\theta} = \\sum_ {c = 1} ^ {C} \\left(\\overline {{\\mathbf {f}}} _ {\\mathbf {v} _ {i, j, c}} ^ {\\theta} - \\overline {{f}} _ {a i, j, c} ^ {\\theta}\\right) ^ {2}, \\tag {13}", + "image_path": "7c16937224c8a249e0d48120a6fdd891243b65fab8c9ecdabd55dc5f52575409.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "inline_equation", + "content": "i, j," + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "inline_equation", + "content": "c" + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "content": " are the index of height, width, and channel, respectively. Finally, we can get the homogenized pixel features " + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\mathbf{F}^a" + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "content": " by weighting the homogenized map " + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 313, + 308, + 554, + 357 + ], + "type": "text", + "content": " to original image features in a pixel-wise manner as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 403, + 361, + 555, + 376 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 403, + 361, + 555, + 376 + ], + "spans": [ + { + "bbox": [ + 403, + 361, + 555, + 376 + ], + "type": "interline_equation", + "content": "\\mathbf {f} _ {h} ^ {\\theta} = \\mathbf {f} ^ {\\theta} \\circ \\mathcal {H} ^ {\\theta}, \\tag {14}", + "image_path": "3fcdb15700226767c90923b055524566714391b0757ad97f94d1d787a2ecce47.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 380, + 554, + 489 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 380, + 554, + 489 + ], + "spans": [ + { + "bbox": [ + 313, + 380, + 554, + 489 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 380, + 554, + 489 + ], + "type": "inline_equation", + "content": "\\circ" + }, + { + "bbox": [ + 313, + 380, + 554, + 489 + ], + "type": "text", + "content": " denotes the element-wise product. Besides, we use a self-attention layer for further aggregating target information across different pixels. The obtained visual features are fed into the conventional teacher-student network, wherein the images from the source domain are input to the student network for supervised constraints, while the images from the target domain are separately input to both the student and teacher networks to generate predictions and pseudolabels for unsupervised constraints." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 314, + 495, + 436, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 495, + 436, + 507 + ], + "spans": [ + { + "bbox": [ + 314, + 495, + 436, + 507 + ], + "type": "text", + "content": "3.5. Training Supervision" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "text", + "content": "As illustrated in Section 3.1, the acquired adapted vision features from the source and target are input to the teacher-student network to acquire the supervised loss " + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^s" + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "text", + "content": " and unsupervised loss " + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}^t" + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "text", + "content": ". Together with the diversity loss " + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div}" + }, + { + "bbox": [ + 313, + 511, + 554, + 583 + ], + "type": "text", + "content": " in the prompt bank, we can finally obtain the segmentation map as follows:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 371, + 588, + 553, + 602 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 371, + 588, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 371, + 588, + 553, + 602 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {t o t a l}} = \\mathcal {L} ^ {s} + \\mathcal {L} ^ {t} + \\lambda_ {\\text {d i v}} \\mathcal {L} _ {\\text {d i v}}, \\tag {15}", + "image_path": "aa96b2dbdd65699e8640fdcd720a8777426190bdc64b1248ccdc55ad31ec135f.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "inline_equation", + "content": "\\lambda_{div}" + }, + { + "bbox": [ + 313, + 608, + 455, + 620 + ], + "type": "text", + "content": " is the trade-off weight." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 628, + 391, + 642 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 628, + 391, + 642 + ], + "spans": [ + { + "bbox": [ + 313, + 628, + 391, + 642 + ], + "type": "text", + "content": "4. Experiment" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 648, + 422, + 661 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 648, + 422, + 661 + ], + "spans": [ + { + "bbox": [ + 313, + 648, + 422, + 661 + ], + "type": "text", + "content": "4.1. Experiment Setup" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 554, + 715 + ], + "type": "text", + "content": "Datasets. In our experiments, following the conventional practice [16, 18, 20], we utilize 4 datasets including CityScapes [6], ACDC [45], NightCity [49] and DarkZurich [44]. For the source domain, we use the training" + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 758 + ], + "type": "text", + "content": "13956" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 61, + 102, + 553, + 245 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 94 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons of semantic segmentation performance under adverse conditions on CityScapes [6]→ACDC [45], where the best performance is marked as **bold** and the second best is **underlined**. † denotes the performance of our reproduction." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 102, + 553, + 245 + ], + "lines": [ + { + "bbox": [ + 61, + 102, + 553, + 245 + ], + "spans": [ + { + "bbox": [ + 61, + 102, + 553, + 245 + ], + "type": "table", + "html": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]72.914.340.516.621.29.317.421.263.823.818.332.619.569.536.234.546.226.936.132.7
\\(MGCDA^†\\) [44]73.428.769.919.326.336.853.053.375.432.084.651.026.177.643.245.953.932.741.548.7
\\(DANNet^†\\) [54]84.354.277.638.030.018.941.635.271.339.486.648.729.276.241.643.058.632.643.950.0
DAFormer [16]58.451.384.042.735.150.730.057.074.852.851.358.332.682.758.354.982.444.150.755.4
\\(SSAC^†\\) [20]76.559.882.752.634.551.258.161.263.958.167.261.540.081.859.261.783.447.555.659.6
HRDA [17]88.357.988.155.236.756.362.965.374.257.785.968.845.788.576.482.487.752.760.468.0
MIC [18]90.867.189.254.540.557.262.068.476.361.887.071.349.489.775.786.889.156.963.070.4
HALO [8]94.279.888.260.251.164.178.265.687.955.795.566.320.788.982.289.387.950.459.071.9
CoDA [12]93.172.790.757.347.456.869.970.087.359.895.471.447.690.377.183.889.154.764.172.6
\\(WA^2Net (ours)\\)94.577.191.661.947.657.568.570.479.264.095.672.850.290.575.588.790.957.562.673.8
", + "image_path": "2a294b57cab2367a98b901c947534bfa845d7c16205ecfc1f99bdf793cabfb5c.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 61, + 288, + 552, + 411 + ], + "blocks": [ + { + "bbox": [ + 55, + 256, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 55, + 256, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 55, + 256, + 555, + 281 + ], + "type": "text", + "content": "Table 2. Quantitative comparisons of performance on CityScapes [6] " + }, + { + "bbox": [ + 55, + 256, + 555, + 281 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 256, + 555, + 281 + ], + "type": "text", + "content": " NightCity [49]. Considering the NightCity [49] dataset is newly introduced, we reproduce the results of state-of-the-art methods for better comparison." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 61, + 288, + 552, + 411 + ], + "lines": [ + { + "bbox": [ + 61, + 288, + 552, + 411 + ], + "spans": [ + { + "bbox": [ + 61, + 288, + 552, + 411 + ], + "type": "table", + "html": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
DANNet [54]57.536.452.342.539.533.227.846.148.918.467.049.528.357.244.050.443.229.137.444.0
DAFormer [16]64.241.059.647.043.235.731.952.050.320.373.151.829.762.049.858.945.331.239.846.8
SSAC [20]84.546.775.553.149.638.331.048.856.820.281.552.630.275.853.561.146.830.237.548.4
HRDA [17]83.550.578.651.948.340.535.457.858.219.981.856.329.780.565.880.750.431.040.852.2
MIC [18]84.951.678.453.848.641.838.860.557.122.083.460.231.981.464.870.852.237.843.853.8
HALO [8]85.752.177.952.947.239.337.563.255.621.582.161.430.780.563.972.353.439.144.754.2
CoDA [12]86.150.979.151.749.340.136.965.458.323.282.762.833.180.262.574.654.938.345.654.9
WA2Net (ours)87.454.279.354.349.940.642.267.859.422.883.968.234.383.565.176.156.846.448.256.5
", + "image_path": "b34a12fef0e95e5b836cf27a0420c84c8978007fd2186173e0b2d384a3b6e91e.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "spans": [ + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "text", + "content": "split of CityScapes; for the target domain, we use either the test of ACDC/DarkZurich or the validation set of NightCity. (1) CityScapes' training set contains 2,975 images of 19 classes with resolution of " + }, + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "inline_equation", + "content": "2048 \\times 1024" + }, + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "text", + "content": ". (2) ACDC contains weather conditions including fog, nighttime, rain, and snow. The test set contains 2,000 images of " + }, + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "inline_equation", + "content": "1920 \\times 1080" + }, + { + "bbox": [ + 54, + 431, + 294, + 552 + ], + "type": "text", + "content": " resolution. (3) NightCity includes 4,297 real night-time images, divided into 2,998 train images and 1,299 val images. (4) DarkZurich contains 2,416 training and 151 test images for nighttime." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "spans": [ + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "text", + "content": "Implementation Details. We adopt Pytorch [40] and Detectron2 [55] to implement the proposed method. 4 NVIDIA GeForce RTX 3090 GPUs are used for training. The extractor of amplitude characteristic is ResNet-18 [15]. During the training stage, our model is trained with a batch size of 4, using the Adam optimizer [31] with an initial learning rate of 0.0001. The input images are rescaled and randomly cropped to " + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "inline_equation", + "content": "512 \\times 512" + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "text", + "content": " following the same data augmentation in DAFormer [16], and the EMA coefficient for updating the teacher net is set to be 0.999. We set the prompt number as " + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "inline_equation", + "content": "N = 16" + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "text", + "content": ", and the coefficient of the diversity loss as " + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "inline_equation", + "content": "\\lambda_{div} = 0.2" + }, + { + "bbox": [ + 54, + 554, + 295, + 698 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 701, + 295, + 714 + ], + "type": "text", + "content": "Evaluation Metric. For a fair comparison, we adopt the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 431, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 431, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 431, + 555, + 480 + ], + "type": "text", + "content": "metric of mean intersection over union (mIOU). It measures the similarity between the prediction and the ground truth by computing the ratio of the intersection area to the union area of the two masks for each object class." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 488, + 539, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 488, + 539, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 488, + 539, + 502 + ], + "type": "text", + "content": "4.2. Comparison with State-of-the-Art Methods" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 506, + 555, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 506, + 555, + 651 + ], + "spans": [ + { + "bbox": [ + 313, + 506, + 555, + 651 + ], + "type": "text", + "content": "Our method demonstrates superior performance for semantic segmentation under adverse conditions, outperforming state-of-the-art methods in Table 1, 2& 3. As can be observed, our method achieves significant improvements in challenging classes such as road, traffic light and person. Besides, as shown in Figure 3, our method shows promising segmentation performance in diverse weather scenes. In specific, our method performs great in most scenarios, while MIC [18] may generate confusing activation due to the inability to explicitly perceive weather conditions, our " + }, + { + "bbox": [ + 313, + 506, + 555, + 651 + ], + "type": "inline_equation", + "content": "\\mathrm{WA}^2\\mathrm{Net}" + }, + { + "bbox": [ + 313, + 506, + 555, + 651 + ], + "type": "text", + "content": " can adaptively focus on accurate foreground targets owing to weather-aware aggregation and adaptation." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 659, + 408, + 672 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 659, + 408, + 672 + ], + "spans": [ + { + "bbox": [ + 313, + 659, + 408, + 672 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 715 + ], + "type": "text", + "content": "Effectiveness of the Amplitude Prompt Aggregation. The amplitude prompt aggregation consists of two processes: amplitude characteristic extraction and weather-" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "13957" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 60, + 102, + 553, + 235 + ], + "blocks": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "lines": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "spans": [ + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": "Table 3. Quantitative comparisons of performance on CityScapes [6] " + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": " DarkZurich [44], where the best performance is marked as bold and the second is underlined. " + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "inline_equation", + "content": "^\\dagger" + }, + { + "bbox": [ + 55, + 71, + 555, + 94 + ], + "type": "text", + "content": " denotes the performance of our reproduction of the corresponding model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 60, + 102, + 553, + 235 + ], + "lines": [ + { + "bbox": [ + 60, + 102, + 553, + 235 + ], + "spans": [ + { + "bbox": [ + 60, + 102, + 553, + 235 + ], + "type": "table", + "html": "
MethodRoadS.walkBuild.WallFencePoleTr.LightSignVeget.TerrainSkyPersonRiderCarTruckBusTrainM.bikeBikeOverall
ADVENT [53]85.837.955.527.714.523.114.021.132.18.72.039.916.664.013.80.058.828.520.729.7
\\( MGCDA^† \\)[44]80.349.366.27.811.041.438.939.064.118.055.852.153.574.766.00.037.529.122.742.5
\\( DANNet^† \\)[54]90.054.074.841.021.125.026.830.272.026.284.047.033.968.219.00.366.438.323.644.3
DAFormer [16]93.565.573.339.419.253.344.144.059.534.566.653.452.782.152.79.589.350.538.553.8
HRDA [17]90.456.372.039.519.557.852.743.159.329.170.560.058.684.075.511.290.551.640.955.9
MIC [18]94.875.084.055.128.462.035.552.659.246.870.065.261.782.164.218.591.352.644.060.2
\\( HALO^† \\)[8]94.574.583.955.528.661.236.052.059.045.570.864.561.281.564.018.290.852.344.260.0
\\( CoDA^† \\)[12]95.374.285.157.329.161.539.853.458.745.372.666.162.283.065.719.390.553.245.361.4
\\( WA^2Net (ours) \\)95.777.986.759.631.263.451.955.760.947.276.367.963.184.569.818.191.954.846.162.8
", + "image_path": "5312aaebc363028fd095fb4665655d438aa82084b47f00db5936d93539adf0d1.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 56, + 246, + 555, + 445 + ], + "blocks": [ + { + "bbox": [ + 56, + 246, + 555, + 445 + ], + "lines": [ + { + "bbox": [ + 56, + 246, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 56, + 246, + 555, + 445 + ], + "type": "image", + "image_path": "a139027fda751d9b3b1062f89b17f3ac752b11cf3361acd094660e047642e9ee.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 56, + 452, + 553, + 464 + ], + "lines": [ + { + "bbox": [ + 56, + 452, + 553, + 464 + ], + "spans": [ + { + "bbox": [ + 56, + 452, + 553, + 464 + ], + "type": "text", + "content": "Figure 3. Qualitative results of our " + }, + { + "bbox": [ + 56, + 452, + 553, + 464 + ], + "type": "inline_equation", + "content": "\\mathrm{WA}^2\\mathrm{Net}" + }, + { + "bbox": [ + 56, + 452, + 553, + 464 + ], + "type": "text", + "content": " and MIC [18] under adverse conditions on ACDC [45], NightCity [49] and DarkZurich [44]." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "table", + "bbox": [ + 59, + 503, + 294, + 569 + ], + "blocks": [ + { + "bbox": [ + 75, + 483, + 276, + 494 + ], + "lines": [ + { + "bbox": [ + 75, + 483, + 276, + 494 + ], + "spans": [ + { + "bbox": [ + 75, + 483, + 276, + 494 + ], + "type": "text", + "content": "Table 4. Ablation on main components of our network." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 59, + 503, + 294, + 569 + ], + "lines": [ + { + "bbox": [ + 59, + 503, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 59, + 503, + 294, + 569 + ], + "type": "table", + "html": "
Amplitude Prompt AggregationWeather Heterogeneity AdaptationACDCNightCity
Charac. ExtractionPrompt Agg.Prompt Modu.Homo. Attention
XXXX68.652.4
✓XX✓70.753.7
✓X✓✓71.254.9
✓✓✓✓73.856.5
", + "image_path": "a22e947c7ff632d946fd8d851888fb2061f7a0650b6a780aa9e80ebb07849b40.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_body" + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": "aware prompt aggregation. To evaluate the effectiveness of the amplitude prompt aggregation, we conduct ablation studies as shown in Table 4. As shown in the " + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "inline_equation", + "content": "3^{rd}" + }, + { + "bbox": [ + 55, + 594, + 296, + 713 + ], + "type": "text", + "content": " row in Table 4, not using prompt aggregation alone means that the prompt bank does not undergo knowledge learning of amplitude characteristics but instead self-initialized and only updated during network training. Without aggregating amplitude characteristic information, prompts cannot perceive weather-aware knowledge, failing to effectively modulate features in the prompt-guided modulation with a significant" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 316, + 483, + 553, + 639 + ], + "blocks": [ + { + "bbox": [ + 316, + 483, + 553, + 639 + ], + "lines": [ + { + "bbox": [ + 316, + 483, + 553, + 639 + ], + "spans": [ + { + "bbox": [ + 316, + 483, + 553, + 639 + ], + "type": "image", + "image_path": "d3bb0d37b9897b9c6539fd0e44b8aabd104556c36ec121ef3e6c2c821ac193e2.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 315, + 647, + 552, + 659 + ], + "lines": [ + { + "bbox": [ + 315, + 647, + 552, + 659 + ], + "spans": [ + { + "bbox": [ + 315, + 647, + 552, + 659 + ], + "type": "text", + "content": "Figure 4. Qualitative results on the daytime " + }, + { + "bbox": [ + 315, + 647, + 552, + 659 + ], + "type": "inline_equation", + "content": "\\rightarrow" + }, + { + "bbox": [ + 315, + 647, + 552, + 659 + ], + "type": "text", + "content": " nighttime setting." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 687, + 402, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 687, + 402, + 696 + ], + "spans": [ + { + "bbox": [ + 314, + 687, + 402, + 696 + ], + "type": "text", + "content": "performance decline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 314, + 701, + 553, + 714 + ], + "type": "text", + "content": "Effectiveness of the Weather Heterogeneity Adaptation." + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13958" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 57, + 70, + 171, + 146 + ], + "blocks": [ + { + "bbox": [ + 57, + 70, + 171, + 146 + ], + "lines": [ + { + "bbox": [ + 57, + 70, + 171, + 146 + ], + "spans": [ + { + "bbox": [ + 57, + 70, + 171, + 146 + ], + "type": "image", + "image_path": "b3b6e63ea06dcad60b26e5e96898d8178c07ae106e9c56622edd1175756f10e0.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "lines": [ + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "spans": [ + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "type": "text", + "content": "Figure 5. Ablation experiments on super-parameters: " + }, + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "type": "inline_equation", + "content": "\\lambda_{div}" + }, + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 155, + 295, + 167 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 173, + 70, + 294, + 146 + ], + "blocks": [ + { + "bbox": [ + 173, + 70, + 294, + 146 + ], + "lines": [ + { + "bbox": [ + 173, + 70, + 294, + 146 + ], + "spans": [ + { + "bbox": [ + 173, + 70, + 294, + 146 + ], + "type": "image", + "image_path": "3c794cbebdacf63ba504da70ac30668cd0fc793e71fc500ec69ebd066441fbad.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 83, + 198, + 268, + 249 + ], + "blocks": [ + { + "bbox": [ + 72, + 178, + 279, + 190 + ], + "lines": [ + { + "bbox": [ + 72, + 178, + 279, + 190 + ], + "spans": [ + { + "bbox": [ + 72, + 178, + 279, + 190 + ], + "type": "text", + "content": "Table 5. Performance w/ and w/o the diversity loss " + }, + { + "bbox": [ + 72, + 178, + 279, + 190 + ], + "type": "inline_equation", + "content": "{\\mathcal{L}}_{div}" + }, + { + "bbox": [ + 72, + 178, + 279, + 190 + ], + "type": "text", + "content": " ." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 83, + 198, + 268, + 249 + ], + "lines": [ + { + "bbox": [ + 83, + 198, + 268, + 249 + ], + "spans": [ + { + "bbox": [ + 83, + 198, + 268, + 249 + ], + "type": "table", + "html": "
Diversity LossACDC [45]NightCity [49]
w/o L_divw/ L_div
✓72.456.0
✓73.856.5
", + "image_path": "b331d193ae5e6e82bc7897e1812e75b860d75fd3e0625ba37ca1501a848fd00c.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_body" + } + ], + "index": 4 + }, + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "spans": [ + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "text", + "content": "The weather heterogeneity adaptation includes the prompt-guided modulation and homogenization attention. We assess the impact of the weather heterogeneity adaptation by disabling this process and observing the changes in segmentation accuracy. Simultaneously discarding both prompt aggregation and modulation, as shown in the " + }, + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "inline_equation", + "content": "2^{nd}" + }, + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "text", + "content": " row of Table 4 (i.e., abandoning the generation and use of prompts), means that the extracted amplitude features are directly fed into the homogenization attention mechanism. This approach results in " + }, + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "inline_equation", + "content": "f_{a}^{\\theta}" + }, + { + "bbox": [ + 54, + 270, + 295, + 413 + ], + "type": "text", + "content": " being unable to perceive multi-weather scene characteristics, adversely affecting the model's generalization in extreme scenarios." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "spans": [ + { + "bbox": [ + 56, + 415, + 295, + 642 + ], + "type": "text", + "content": "Analysis of the Homogenization Attention. We also analyze the effectiveness of our designed homogenization attention with comparisons against other attention mechanisms as shown in Table 6. The homogenization attention mechanism is designed to address the significant variability in weather conditions by dynamically aligning features from different weather domains. Unlike crossattention, which focuses on explicit interactions between tokens across feature spaces, and masked attention, which selectively attends to spatial regions, our homogenization attention incorporates a domain adaptation strategy that directly mitigates the feature discrepancies caused by adverse weather conditions. As can be observed, the heterogeneity adaptation mechanism proved to be essential to dynamically adapt to varying weather patterns, thereby improving overall performance and reliability. Besides, Figure 6 displays activation maps of different classes with and without weather-aware aggregation and adaptation, illustrating the impact of our design on class-specific segmentation quality." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": "Analysis of Hyperparameters. We conducted a comprehensive analysis of hyperparameters to understand their influence on the model's performance. The impact of the choice of diversity loss " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "\\lambda_{div}" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " and the prompt number " + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "inline_equation", + "content": "N" + }, + { + "bbox": [ + 55, + 642, + 295, + 714 + ], + "type": "text", + "content": " is shown in Figure 5. Too few prompts hinder the ability to capture rich scene information, while too many may capture" + } + ] + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 315, + 71, + 553, + 115 + ], + "blocks": [ + { + "bbox": [ + 315, + 71, + 553, + 115 + ], + "lines": [ + { + "bbox": [ + 315, + 71, + 553, + 115 + ], + "spans": [ + { + "bbox": [ + 315, + 71, + 553, + 115 + ], + "type": "image", + "image_path": "bd7dd7d5c66d460fe34e30a6074bb875c83c1e53e6ee5feb2ed7aca241cfaa98.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 116, + 553, + 158 + ], + "blocks": [ + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "lines": [ + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "spans": [ + { + "bbox": [ + 316, + 116, + 553, + 158 + ], + "type": "image", + "image_path": "5d1f1ad9ec2f931d833e7e0b94d7373a07315dd06483ebb5df0164856ec05ec8.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 316, + 159, + 553, + 202 + ], + "blocks": [ + { + "bbox": [ + 316, + 159, + 553, + 202 + ], + "lines": [ + { + "bbox": [ + 316, + 159, + 553, + 202 + ], + "spans": [ + { + "bbox": [ + 316, + 159, + 553, + 202 + ], + "type": "image", + "image_path": "4d149bd444e680be6c14d67f2e65d1b72417dc843867d4116f4c65ddd036e7f2.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 316, + 203, + 553, + 246 + ], + "blocks": [ + { + "bbox": [ + 316, + 203, + 553, + 246 + ], + "lines": [ + { + "bbox": [ + 316, + 203, + 553, + 246 + ], + "spans": [ + { + "bbox": [ + 316, + 203, + 553, + 246 + ], + "type": "image", + "image_path": "8fe8b14f43974e59a52c5762f3295bd2d55378d75c7d85cc650e75583bc552aa.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 255, + 555, + 278 + ], + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 278 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 278 + ], + "type": "text", + "content": "Figure 6. Comparison of activation maps for different classes with (w/) and without(w/o) weather-aware aggregation and adaptation." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "table", + "bbox": [ + 319, + 322, + 552, + 373 + ], + "blocks": [ + { + "bbox": [ + 313, + 291, + 555, + 312 + ], + "lines": [ + { + "bbox": [ + 313, + 291, + 555, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 291, + 555, + 312 + ], + "type": "text", + "content": "Table 6. Comparison of different attention mechanisms with our homogenization attention." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 319, + 322, + 552, + 373 + ], + "lines": [ + { + "bbox": [ + 319, + 322, + 552, + 373 + ], + "spans": [ + { + "bbox": [ + 319, + 322, + 552, + 373 + ], + "type": "table", + "html": "
Attention MechanismACDC [45]NightCity [49]
cross-attention [52]70.954.1
masked attention [5]71.855.0
Homogenization Attention (ours)73.856.5
", + "image_path": "dd323b684dacd469efcd3daa9ecc21de2f12f6e5c780fe4e7341edeab4efb3e0.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 395, + 555, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 395, + 555, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 395, + 555, + 502 + ], + "type": "text", + "content": "irrelevant noise information. It is worth noting that the improvement brought by the diversity loss in Table 5 is more significant when ACDC [45] is the target domain. This is due to ACDC containing more weather diversities (night, rain, snow and fog) compared to the night-only NightCity [49], prompts on ACDC can aggregate more diverse knowledge with " + }, + { + "bbox": [ + 313, + 395, + 555, + 502 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{div}" + }, + { + "bbox": [ + 313, + 395, + 555, + 502 + ], + "type": "text", + "content": ". Therefore, the diversity in prompt combinations contributes to diverse scene analysis under multiple weather conditions." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 515, + 388, + 528 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 515, + 388, + 528 + ], + "spans": [ + { + "bbox": [ + 313, + 515, + 388, + 528 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 536, + 555, + 620 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 536, + 555, + 620 + ], + "spans": [ + { + "bbox": [ + 313, + 536, + 555, + 620 + ], + "type": "text", + "content": "In this paper, we propose a weather-aware aggregation and adaptation network (WA" + }, + { + "bbox": [ + 313, + 536, + 555, + 620 + ], + "type": "inline_equation", + "content": "^2" + }, + { + "bbox": [ + 313, + 536, + 555, + 620 + ], + "type": "text", + "content": "Net) for semantic segmentation under adverse conditions. We design the amplitude prompt aggregation to acquire weather-aware characteristic knowledge and weather heterogeneity adaptation to realize interweather homogenization for effective parsing. Extensive experiments demonstrate our effectiveness." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 633, + 411, + 646 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 633, + 411, + 646 + ], + "spans": [ + { + "bbox": [ + 313, + 633, + 411, + 646 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 554, + 712 + ], + "type": "text", + "content": "This work was partially supported by the National Key Research and Development Program of China (No.2024YFB3909902), National Nature Science Foundation of China (62121002), Youth Innovation Promotion Association of CAS." + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "13959" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 115, + 83 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "spans": [ + { + "bbox": [ + 61, + 91, + 296, + 145 + ], + "type": "text", + "content": "[1] Liang-Chieh Chen, Yukun Zhu, George Papandreou, Florian Schroff, and Hartwig Adam. Encoder-decoder with atrous separable convolution for semantic image segmentation. In Proceedings of the European conference on computer vision (ECCV), pages 801-818, 2018. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 61, + 147, + 296, + 191 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 147, + 296, + 191 + ], + "spans": [ + { + "bbox": [ + 61, + 147, + 296, + 191 + ], + "type": "text", + "content": "[2] Yang Chen, Liang Shan, Yi Qu, Weilong Zhang, and Lu Chang. Amsc: Adaptive masking and structure-constraint learning for domain adaptive semantic segmentation under adverse conditions. IEEE Signal Processing Letters, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "spans": [ + { + "bbox": [ + 62, + 192, + 295, + 257 + ], + "type": "text", + "content": "[3] Yujia Chen, Wangkai Li, Zhaoyang Li, Rui Sun, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Sam-glomeruli: Enhanced segment anything model for precise glomeruli segmentation. In International Workshop on Medical Optical Imaging and Virtual Microscopy Image Analysis, pages 182-191. Springer, 2024. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "spans": [ + { + "bbox": [ + 62, + 258, + 295, + 323 + ], + "type": "text", + "content": "[4] Yujia Chen, Rui Sun, Wangkai Li, Huayu Mai, Naisong Luo, Yuwen Pan, and Tianzhu Zhang. Alleviate and mining: Rethinking unsupervised domain adaptation for mitochondria segmentation from pseudo-label perspective. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2339-2347, 2025. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 324, + 295, + 380 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 324, + 295, + 380 + ], + "spans": [ + { + "bbox": [ + 62, + 324, + 295, + 380 + ], + "type": "text", + "content": "[5] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexander Kirillov, and Rohit Girdhar. Masked-attention mask transformer for universal image segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1290–1299, 2022. 8" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "spans": [ + { + "bbox": [ + 62, + 380, + 295, + 446 + ], + "type": "text", + "content": "[6] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3213-3223, 2016. 5, 6, 7" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "spans": [ + { + "bbox": [ + 62, + 447, + 295, + 491 + ], + "type": "text", + "content": "[7] Ozgur Erkent and Christian Laugier. Semantic segmentation with unsupervised domain adaptation under varying weather conditions for autonomous vehicles. IEEE Robotics and Automation Letters, 5(2):3580-3587, 2020. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "spans": [ + { + "bbox": [ + 62, + 491, + 295, + 546 + ], + "type": "text", + "content": "[8] Luca Franco, Paolo Mandica, Konstantinos Kallidromitis, Devin Guillory, Yu-Teng Li, Trevor Darrell, and Fabio Galasso. Hyperbolic active learning for semantic segmentation under domain shift. In International Conference on Machine Learning, pages 13864-13884. PMLR, 2024. 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 547, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 547, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 62, + 547, + 295, + 581 + ], + "type": "text", + "content": "[9] Tianyu Gao, Adam Fisch, and Danqi Chen. Making pretrained language models better few-shot learners. arXiv preprint arXiv:2012.15723, 2020. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 57, + 582, + 295, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 582, + 295, + 646 + ], + "spans": [ + { + "bbox": [ + 57, + 582, + 295, + 646 + ], + "type": "text", + "content": "[10] Rui Gong, Qin Wang, Martin Danelljan, Dengxin Dai, and Luc Van Gool. Continuous pseudo-label rectified domain adaptive semantic segmentation with implicit neural representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7225-7235, 2023. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 647, + 295, + 712 + ], + "type": "text", + "content": "[11] Ziyang Gong, Fuhao Li, Yupeng Deng, Wenjun Shen, Xi-anzheng Ma, Zhenming Ji, and Nan Xia. Train one, generalize to all: Generalizable semantic segmentation from single-scene to all adverse scenes. In Proceedings of the 31st ACM International Conference on Multimedia, pages 2275-2284, 2023. 2" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 57, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 316, + 73, + 554, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 554, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 554, + 129 + ], + "type": "text", + "content": "[12] Ziyang Gong, Fuhao Li, Yupeng Deng, Deblina Bhattacharjee, Xianzheng Ma, Xiangwei Zhu, and Zhenming Ji. Coda: Instructive chain-of-domain adaptation with severity-aware visual prompt tuning. In European Conference on Computer Vision, pages 130-148. Springer, 2024. 6, 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 163 + ], + "type": "text", + "content": "[13] Monica Gruosso, Nicola Capecce, and Ugo Erra. Human segmentation in surveillance video with deep learning. *Multimedia Tools and Applications*, 80(1):1175-1199, 2021. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 165, + 553, + 219 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 165, + 553, + 219 + ], + "spans": [ + { + "bbox": [ + 317, + 165, + 553, + 219 + ], + "type": "text", + "content": "[14] Yuxian Gu, Xu Han, Zhiyuan Liu, and Minlie Huang. Ppt: Pre-trained prompt tuning for few-shot learning. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 8410-8423, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "spans": [ + { + "bbox": [ + 316, + 222, + 553, + 266 + ], + "type": "text", + "content": "[15] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 770-778, 2016. 4, 6" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 268, + 555, + 323 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 268, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 317, + 268, + 555, + 323 + ], + "type": "text", + "content": "[16] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Daformer: Improving network architectures and training strategies for domain-adaptive semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9924-9935, 2022. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 325, + 553, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 325, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 316, + 325, + 553, + 369 + ], + "type": "text", + "content": "[17] Lukas Hoyer, Dengxin Dai, and Luc Van Gool. Hrda: Context-aware high-resolution domain-adaptive semantic segmentation. In European conference on computer vision, pages 372-391. Springer, 2022. 6, 7" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "spans": [ + { + "bbox": [ + 316, + 371, + 553, + 426 + ], + "type": "text", + "content": "[18] Lukas Hoyer, Dengxin Dai, Haoran Wang, and Luc Van Gool. Mic: Masked image consistency for context-enhanced domain adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11721-11732, 2023. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "spans": [ + { + "bbox": [ + 317, + 428, + 553, + 472 + ], + "type": "text", + "content": "[19] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In European Conference on Computer Vision, pages 709-727. Springer, 2022. 4" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 474, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 474, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 317, + 474, + 553, + 529 + ], + "type": "text", + "content": "[20] Abdulrahman Kerim, Felipe Chamone, Washington Ramos, Leandro Soriano Marcolino, Erickson R Nascimento, and Richard Jiang. Semantic segmentation under adverse conditions: a weather and nighttime-aware synthetic data-based approach. arXiv preprint arXiv:2210.05626, 2022. 2, 3, 5, 6" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 531, + 553, + 574 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 531, + 553, + 574 + ], + "spans": [ + { + "bbox": [ + 317, + 531, + 553, + 574 + ], + "type": "text", + "content": "[21] Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278-2324, 1998. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 576, + 553, + 621 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 576, + 553, + 621 + ], + "spans": [ + { + "bbox": [ + 317, + 576, + 553, + 621 + ], + "type": "text", + "content": "[22] Sohyun Lee, Namyup Kim, Sungyeon Kim, and Suha Kwak. Frest: Feature restoration for semantic segmentation under multiple adverse conditions. In European Conference on Computer Vision, pages 1-18. Springer, 2024. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 667 + ], + "type": "text", + "content": "[23] Liulei Li, Tianfei Zhou, Wenguan Wang, Jianwu Li, and Yi Yang. Deep hierarchical semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1246-1257, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 713 + ], + "type": "text", + "content": "[24] Wangkai Li, Rui Sun, Bohao Liao, Zhaoyang Li, and Tianzhu Zhang. Balanced learning for domain adaptive semantic segmentation. In *Forty-second International Conference on Machine Learning*. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 318, + 757 + ], + "type": "text", + "content": "13960" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 296, + 106 + ], + "type": "text", + "content": "[25] Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation. arXiv preprint arXiv:2101.00190, 2021. 3" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 296, + 150 + ], + "type": "text", + "content": "[26] Zhi Li, Xing Wu, Jianjia Wang, and Yike Guo. Weather-degraded image semantic segmentation with multi-task knowledge distillation. Image and Vision Computing, 127: 104554, 2022. 1" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "spans": [ + { + "bbox": [ + 56, + 152, + 296, + 207 + ], + "type": "text", + "content": "[27] Zhaoyang Li, Yuan Wang, Wangkai Li, Rui Sun, and Tianzhu Zhang. Localization and expansion: A decoupled framework for point cloud few-shot semantic segmentation. In European Conference on Computer Vision, pages 18-34. Springer, 2024. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "spans": [ + { + "bbox": [ + 56, + 209, + 296, + 262 + ], + "type": "text", + "content": "[28] Zhaoyang Li, Yuan Wang, Wangkai Li, Tianzhu Zhang, and Xiang Liu. Dual-agent optimization framework for cross-domain few-shot segmentation. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 9849-9859, 2025. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 319 + ], + "type": "text", + "content": "[29] Pengfei Liu, Weizhe Yuan, Jinlan Fu, Zhengbao Jiang, Hiroaki Hayashi, and Graham Neubig. Pre-train, prompt, and predict: A systematic survey of prompting methods in natural language processing. ACM Computing Surveys, 55(9): 1-35, 2023. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 320, + 296, + 365 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 320, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 56, + 320, + 296, + 365 + ], + "type": "text", + "content": "[30] Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully convolutional networks for semantic segmentation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 3431-3440, 2015. 1" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 56, + 365, + 296, + 388 + ], + "type": "text", + "content": "[31] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. arXiv preprint arXiv:1711.05101, 2017. 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 56, + 389, + 296, + 443 + ], + "type": "text", + "content": "[32] Ning Lv, Zenghui Zhang, Cong Li, Jiaxuan Deng, Tao Su, Chen Chen, and Yang Zhou. A hybrid-attention semantic segmentation network for remote sensing interpretation in land-use surveillance. International Journal of Machine Learning and Cybernetics, 14(2):395–406, 2023. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 296, + 499 + ], + "type": "text", + "content": "[33] Claudia Menzel, Gregor U Hayn-Leichsenring, Oliver Langner, Holger Wiese, and Christoph Redies. Fourier power spectrum characteristics of face photographs: attractiveness perception depends on low-level image properties. PLoS One, 10(4):e0122801, 2015. 4" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 501, + 296, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 296, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 296, + 578 + ], + "type": "text", + "content": "[34] Khan Muhammad, Tanveer Hussain, Hayat Ullah, Javier Del Ser, Mahdi Rezaei, Neeraj Kumar, Mohammad Hijji, Paolo Bellavista, and Victor Hugo C de Albuquerque. Vision-based semantic segmentation in scene understanding for autonomous driving: Recent achievements, challenges, and outlooks. IEEE Transactions on Intelligent Transportation Systems, 23(12):22694-22715, 2022. 1" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 579, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 579, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 579, + 296, + 624 + ], + "type": "text", + "content": "[35] Hyeonwoo Noh, Seunghoon Hong, and Bohyung Han. Learning deconvolution network for semantic segmentation. In Proceedings of the IEEE international conference on computer vision, pages 1520-1528, 2015. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 624, + 296, + 689 + ], + "type": "text", + "content": "[36] Yuwen Pan, Naisong Luo, Rui Sun, Meng Meng, Tianzhu Zhang, Zhiwei Xiong, and Yongdong Zhang. Adaptive template transformer for mitochondria segmentation in electron microscopy images. In Proceedings of the IEEE/CVF international conference on computer vision, pages 21474-21484, 2023. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[37] Yuwen Pan, Rui Sun, Naisong Luo, Tianzhu Zhang, and Yongdong Zhang. Exploring reliable matching with phase" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 106 + ], + "type": "text", + "content": "enhancement for night-time semantic segmentation. In European Conference on Computer Vision, pages 408-424. Springer, 2024. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 161 + ], + "type": "text", + "content": "[38] Yuwen Pan, Rui Sun, Yuan Wang, Wenfei Yang, Tianzhu Zhang, and Yongdong Zhang. Purify then guide: A bidirectional bridge network for open-vocabulary semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 2024. 1" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "spans": [ + { + "bbox": [ + 316, + 162, + 553, + 216 + ], + "type": "text", + "content": "[39] Yuwen Pan, Rui Sun, Yuan Wang, Tianzhu Zhang, and Yongdong Zhang. Rethinking the implicit optimization paradigm with dual alignments for referring remote sensing image segmentation. In Proceedings of the 32nd ACM International Conference on Multimedia, pages 2031-2040, 2024. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "spans": [ + { + "bbox": [ + 316, + 217, + 553, + 281 + ], + "type": "text", + "content": "[40] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. Advances in neural information processing systems, 32, 2019. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 282, + 553, + 327 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 282, + 553, + 327 + ], + "spans": [ + { + "bbox": [ + 316, + 282, + 553, + 327 + ], + "type": "text", + "content": "[41] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of sensor data fusion. In 2019 22th International Conference on Information Fusion (FUSION), pages 1-8. IEEE, 2019. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "spans": [ + { + "bbox": [ + 316, + 327, + 553, + 381 + ], + "type": "text", + "content": "[42] Andreas Pfeuffer and Klaus Dietmayer. Robust semantic segmentation in adverse weather conditions by means of fast video-sequence segmentation. In 2020 IEEE 23rd International Conference on Intelligent Transportation Systems (ITSC), pages 1-6. IEEE, 2020." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 316, + 383, + 553, + 437 + ], + "type": "text", + "content": "[43] Nikhil Reddy, Abhinav Singhal, Abhishek Kumar, Mahsa Baktashmotlagh, and Chetan Arora. Master of all: Simultaneous generalization of urban-scene segmentation to all adverse weather conditions. In European Conference on Computer Vision, pages 51-69. Springer, 2022. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[44] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Map-guided curriculum domain adaptation and uncertainty-aware evaluation for semantic nighttime image segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(6):3139-3153, 2020. 2, 5, 6, 7" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 492, + 553, + 547 + ], + "type": "text", + "content": "[45] Christos Sakaridis, Dengxin Dai, and Luc Van Gool. Acdc: The adverse conditions dataset with correspondences for semantic driving scene understanding. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 10765-10775, 2021. 1, 3, 5, 6, 7, 8" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 602 + ], + "type": "text", + "content": "[46] Rui Sun, Naisong Luo, Yuan Wang, Yuwen Pan, Huayu Mai, Zhe Zhang, and Tianzhu Zhang. 1st place solution for youtuves challenge 2022: Video object segmentation. In Proceedings of the Conference on Computer Vision and Pattern Recognition Workshops, 2022. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "spans": [ + { + "bbox": [ + 316, + 603, + 553, + 647 + ], + "type": "text", + "content": "[47] Rui Sun, Naisong Luo, Yuwen Pan, Huayu Mai, Tianzhu Zhang, Zhiwei Xiong, and Feng Wu. Appearance prompt vision transformer for connectome reconstruction. In IJCAI, pages 1423-1431, 2023. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 647, + 553, + 690 + ], + "type": "text", + "content": "[48] Rui Sun, Huayu Mai, Tianzhu Zhang, and Feng Wu. Daw: exploring the better weighting function for semi-supervised semantic segmentation. Advances in neural information processing systems, 36:61792-61805, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 553, + 713 + ], + "type": "text", + "content": "[49] Xin Tan, Ke Xu, Ying Cao, Yiheng Zhang, Lizhuang Ma, and Rynson WH Lau. Night-time scene parsing with a large" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13961" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 295, + 689 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 294, + 95 + ], + "type": "text", + "content": "real dataset. IEEE Transactions on Image Processing, 30: 9085-9098, 2021. 5, 6, 7, 8" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 295, + 129 + ], + "type": "text", + "content": "[50] Yingjie Tian and Siyu Zhu. Partial domain adaptation on semantic segmentation. IEEE Transactions on Circuits and Systems for Video Technology, 32(6):3798-3809, 2021. 2" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 294, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 185 + ], + "type": "text", + "content": "[51] Abhinav Valada, Johan Vertens, Ankit Dhall, and Wolfram Burgard. Adapnet: Adaptive semantic segmentation in adverse environmental conditions. In 2017 IEEE International Conference on Robotics and Automation (ICRA), pages 4644-4651. IEEE, 2017. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 295, + 229 + ], + "type": "text", + "content": "[52] Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Lukasz Kaiser, and Illia Polosukhin. Attention is all you need. Advances in neural information processing systems, 30, 2017. 3, 5, 8" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 294, + 295 + ], + "type": "text", + "content": "[53] Tuan-Hung Vu, Himalaya Jain, Maxime Bucher, Matthieu Cord, and Patrick Pérez. Advent: Adversarial entropy minimization for domain adaptation in semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2517-2526, 2019. 2, 6, 7" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 56, + 297, + 294, + 352 + ], + "type": "text", + "content": "[54] Xinyi Wu, Zhenyao Wu, Hao Guo, Lili Ju, and Song Wang. Darnet: A one-stage domain adaptation network for unsupervised nighttime semantic segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15769-15778, 2021. 2, 3, 6, 7" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 353, + 294, + 386 + ], + "type": "text", + "content": "[55] Yuxin Wu, Alexander Kirillov, Francisco Massa, Wan-Yen Lo, and Ross Girshick. Detector2. https://github.com/facebookresearch/detectron2, 2019.6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "spans": [ + { + "bbox": [ + 56, + 387, + 294, + 430 + ], + "type": "text", + "content": "[56] E. Xie, W. Wang, Z. Yu, A. Anandkumar, J. M. Alvarez, and P. Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. In Neural Information Processing Systems (NeurIPS), 2021. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "spans": [ + { + "bbox": [ + 56, + 432, + 294, + 475 + ], + "type": "text", + "content": "[57] Xin Yang, Wending Yan, Yuan Yuan, Michael Bi Mi, and Robby T Tan. Semantic segmentation in multiple adverse weather conditions with domain knowledge retention. arXiv preprint arXiv:2401.07459, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "spans": [ + { + "bbox": [ + 56, + 476, + 295, + 520 + ], + "type": "text", + "content": "[58] Yanchao Yang and Stefano Soatto. Fda: Fourier domain adaptation for semantic segmentation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4085-4095, 2020. 2, 4" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "spans": [ + { + "bbox": [ + 56, + 521, + 294, + 576 + ], + "type": "text", + "content": "[59] Chunhui Zhang, Guanjie Huang, Li Liu, Shan Huang, Yinan Yang, Xiang Wan, Shiming Ge, and Dacheng Tao. Webuav-3m: A benchmark for unveiling the power of million-scale deep uav tracking. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(7):9186-9205, 2022. 1" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "spans": [ + { + "bbox": [ + 56, + 577, + 294, + 631 + ], + "type": "text", + "content": "[60] Chunhui Zhang, Li Liu, Guanjie Huang, Hao Wen, Xi Zhou, and Yanfeng Wang. Webuot-1m: Advancing deep underwater object tracking with a million-scale benchmark. Advances in Neural Information Processing Systems, 37:50152-50167, 2024. 1" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 633, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 633, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 633, + 294, + 689 + ], + "type": "text", + "content": "[61] Chunhui Zhang, Li Liu, Hao Wen, Xi Zhou, and Yanfeng Wang. Mambatrack: Exploiting dual-enhancement for night uav tracking. In ICASSP 2025-2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), pages 1-5. IEEE, 2025. 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "13962" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_content_list.json b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..09eea9dd3e8f7d89de0f41f8e3a9f47bcd499831 --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_content_list.json @@ -0,0 +1,1817 @@ +[ + { + "type": "text", + "text": "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics", + "text_level": 1, + "bbox": [ + 112, + 128, + 885, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Taowen Wang $^{1*}$ , Cheng Han $^{2*}$ , James Liang $^{3*}$ , \nWenhao Yang $^{4}$ , Dongfang Liu $^{1\\dagger}$ , Luna Xinyu Zhang $^{1\\dagger}$ , Qifan Wang $^{5}$ , \nJiebo Luo $^{6}$ , Ruixiang Tang $^{7\\dagger}$", + "bbox": [ + 217, + 202, + 779, + 258 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{1}$ Rochester Institute of Technology, $^{2}$ University of Missouri - Kansas City, $^{3}$ U.S. Naval Research Laboratory, $^{4}$ Lamar University, $^{5}$ Meta AI, $^{6}$ University of Rochester, $^{7}$ Rutgers University", + "bbox": [ + 207, + 273, + 790, + 328 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 361, + 326, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a $100\\%$ reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments1.", + "bbox": [ + 89, + 393, + 485, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 91, + 773, + 220, + 789 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "\"First directive: A robot cannot harm a human or, through inaction, allow a human to be harmed.\"", + "bbox": [ + 104, + 794, + 478, + 823 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "—Finch [75]", + "bbox": [ + 385, + 825, + 478, + 840 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a27d1366dc9413109af42625ca8f73a705e6f2d57e087ea6e0769e728e603a08.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 362, + 897, + 531 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/a7f1a9cd58f23d1dfa2cd5f851f69b1c8dae7758bd7a5bcc21a17523d63813b5.jpg", + "image_caption": [ + "Figure 1. Adversarial Vulnerabilities induced by malicious manipulation. (A). Illustration of adversarial threats in robotic task execution. (B). Example of semantic-rich adversarial patches generated by proposed methods. (C). Comparison of failure rates across different attack schemes (UADA, UPA, and TMA)." + ], + "image_footnote": [], + "bbox": [ + 521, + 534, + 704, + 638 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/08f45c5ade4c1077e04a483cdb86841d235414718c7f84a3b13bde22084bdd57.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 709, + 532, + 897, + 638 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "In the movie Finch, set in a post-apocalyptic future, an intelligent robot navigates complex themes, underscoring the importance of interactive safety with its human master. Once speculative, this portrayal is now increasingly plausible with the rise of Large Vision Language Models (LVLMs) [70, 79], capable of seamlessly interpreting both visual and linguistic contexts. A notable realization of this potential can be seen in Vision-Language-Action (VLA) models [16, 34, 73, 74], which integrate LVLMs into robotic systems to enable end-to-end learning that encompasses both high-level trajectory planning and low-level robot action control. While VLA models demonstrate a promising step toward generalist", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution.", + "bbox": [ + 107, + 849, + 220, + 862 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "† Corresponding author.", + "bbox": [ + 109, + 862, + 235, + 875 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "$\\ddagger$ Work done during Luna Xinyu Zhang's internship at RIT.", + "bbox": [ + 109, + 875, + 418, + 887 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "1 Project homepage.", + "bbox": [ + 109, + 887, + 215, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "6948", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "robotic intelligence, they also introduce potential vulnerabilities that remain largely underexplored. This work, in light of this view, pioneers the understanding of the pressing need to investigate the new attack surfaces associated with VLA-based robotic systems.", + "bbox": [ + 89, + 90, + 480, + 167 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To gain deeper insights into these vulnerabilities, we analyze VLA-based models and general robotic operations, highlighting two key characteristics crucial for designing effective adversarial attacks. First, creating attack objectives for robotic systems requires consideration of the physical dynamics and constraints intrinsic to robotic movement. Conventional adversarial attacks, in this case, often fail to produce significant deviations in intended actions because they disregard these constraints. Second, VLA models generate control signals that function as a time series of $K$ -class predictions (see §3.1), making it essential to design attacks that exploit the temporal dependencies within these sequences to cause substantial disruptions in robotic behavior. Achieving attacks that are effective in both digital simulations and real-world environments remains critical yet challenging.", + "bbox": [ + 89, + 169, + 482, + 396 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, our work intensifies the adversarial threats posed to VLA-based systems by both developing specialized attack objectives and designing effective attack methods. Specifically, we formulate an Action Discrepancy Objective aimed at maximizing the action discrepancy within VLA-based robotic systems. This strategy ensures that, at each decision point, the robot's behavior can diverge from the optimal trajectory. Additionally, we introduce Geometry-Aware Objective that considers the robot's movement in three-dimensional space, characterized by three degrees of freedom. By optimizing the cosine similarity between adversarial and ground-truth directions, we induce deviations in the robot's movement direction from its intended path, increasing the likelihood of task failure. To achieve these attack objectives, we develop straightforward yet effective Patch-Based Attacks targeting VLA-based robotic systems. This approach enables adversarial attacks in both digital and physical settings, revealing substantial vulnerabilities within the VLA-based system.", + "bbox": [ + 89, + 398, + 482, + 686 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Altogether, these innovations yield several significant contributions: ① This work presents a pioneering and comprehensive analysis of vulnerabilities within VLA-based robotic systems, a new paradigm for training generalized robotic policies using generative foundation models. We reveal critical threats to adversarial attacks, emphasizing the urgent need to enhance robustness before real-world deployment; ② To the best of our knowledge, we are the first work to define attack objectives specific to the powerful VLA models and to employ a straightforward adversarial patch against it (see §3). This offers valuable insights for the research community to explore systemic failures in similar concurrent generative foundation models; ③ We rigorously evaluate our approach in both simulated and real-world environments across four", + "bbox": [ + 89, + 689, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "distinct robotic tasks, observing significant raises in task failure rates of $100\\%$ and $43\\%$ , respectively. This highlights the effectiveness of our attack strategies (see $\\S 4.3$ ).", + "bbox": [ + 511, + 90, + 903, + 137 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related work", + "text_level": 1, + "bbox": [ + 513, + 143, + 651, + 157 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.1. AI-driven Generalist Robot", + "text_level": 1, + "bbox": [ + 513, + 167, + 761, + 183 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Developing generalist robots [6, 14, 27, 37, 60] requires models to not only handle varied interactions but also maintain robustness in unstructured environments. Early intelligent robots [51, 52, 54, 55, 83] generally relied on rule-based approaches, effective only in controlled settings and needing extensive reprogramming for new tasks. The advent of deep learning [19, 31, 40, 56] and reinforcement learning [1, 63, 72] has shifted this paradigm to adapt through the data-driven pipeline, increasing robot versatility and reducing manual reconfiguration.", + "bbox": [ + 511, + 189, + 906, + 339 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [7, 53, 57] have further advanced the development and potential of Large Vision Language Models (LVLMs) as key enablers for generalist robots, demonstrating promising generalization across a variety of scenes [15, 25, 32, 41, 64, 84]. One of the notable examples is OpenVLA [34], which integrates dual visual encoders with a pre-trained large language model to enable observation-to-action capabilities through instruction-following [10, 33, 45, 46, 70, 82]. By aligning visual and textual semantics, OpenVLA facilitates complex scenario understanding and end-to-end action generation, showing impressive generalization when executing previously unseen task instructions.", + "bbox": [ + 511, + 340, + 908, + 536 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2.2. Adversarial Attacks in Robotic", + "text_level": 1, + "bbox": [ + 513, + 546, + 785, + 561 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Adversarial attacks serve as critical tools for assessing model vulnerabilities, particularly in robotics, where models operate in dynamic, real-world settings. Traditional gradient-based, pixel-level attacks [17, 21, 43, 49, 61, 71] exploit model gradients to calculate malicious perturbations, achieving high attack success rates in digital environments. For physical-world attacks, patch-based methods [3, 8, 59, 76, 77] offer a practical alternative for real-world applications by introducing physically realizable perturbations that retain efficacy under varied conditions, such as angle or lighting changes, making them highly applicable for robotic systems. Therefore, considering VLA-based models enable robot execution in real-world scenarios, this work employs a patch-based attack strategy.", + "bbox": [ + 511, + 568, + 908, + 779 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "VLA models [13, 26, 64] couple visual and linguistic modalities for perception-action alignment. The continuous, high-dimensional nature of visual inputs [49, 58, 62] makes them the most likely target for adversarial perturbations, as attackers can subtly alter visual data in ways that are difficult to detect [2, 9, 65]. Consequently, this work applies attacks on the visual modality to achieve robust manipulations across diverse environments. To the best of our knowledge, this", + "bbox": [ + 511, + 780, + 908, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "6949", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/c1f482ac55f606457a94bb97d63f25131deeaa895b2426f984ac68d6b5d941b2.jpg", + "image_caption": [ + "Figure 2. Overall Adversarial Framework. The robot captures an input image, processes it through a vision-language model to generate tokens representing actions, and then uses an action de-tokenizer for discrete bin prediction. The model is optimized with adversarial objectives focusing on various discrepancies and geometries (i.e., UADA, UPA, TMA). Forward propagation is shown in black, and backpropagation is highlighted in pink. These objectives aim to maximize errors and minimize task performance, with visual emphasis on 3D-space manipulation and a focus on generating adversarial perturbation $\\delta$ during task execution, such as picking up a can." + ], + "image_footnote": [], + "bbox": [ + 112, + 92, + 883, + 311 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "study is one of the earliest attempts to formally define adversarial objectives for VLA models, targeting gaps in adversarial robustness through the model's geometric constraints (e.g., spatial dependencies in visual input) and architectural nuances (e.g., cross-modal information integration).", + "bbox": [ + 89, + 390, + 483, + 465 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Methodology", + "text_level": 1, + "bbox": [ + 89, + 476, + 225, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this section, we first review the core algorithmic principles of VLA in §3.1, which serves as the foundation for the adversarial scheme designs. We then present two types of untargeted adversarial attacks, including Untargeted Action Discrepancy Attack (UADA) in §3.2 and Untargeted Position-aware Attack (UPA) in §3.3. In §3.4, we further include Targeted Manipulation Attack (TMA), designed to direct the robot toward specific erroneous execution. Finally, we introduce the Normalized Action Discrepancy (NAD) metric in §3.5. The overall framework is shown in Fig. 2.", + "bbox": [ + 89, + 502, + 483, + 654 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Preliminary", + "text_level": 1, + "bbox": [ + 89, + 666, + 220, + 681 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Vision-Language-Action models [34, 84] (see Fig. 2) are built on large language models integrated with visual encoders, enabling robots to interpret human instructions and process visual input from a camera to perform context-aware actions. VLA models [34, 42, 84] generally abstract continuous action predictions into a classification problem by discretizing robot actions. Specifically, they first discretize the continuous probabilities into class labels $y = \\arg \\max_{x} \\mathcal{F}(x)$ where $\\mathcal{F}(\\cdot)$ is the VLA model. An action de-tokenizer then generates actions $A = DT(y)$ . By categorizing action values into discrete class labels, the model converts continuous probability outputs into discrete signals, this simplification facilitates quicker convergence and faster training times, and is commonly used for VLA-based models [34, 42, 84].", + "bbox": [ + 89, + 688, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this work, we focus on a 7 degree-of-freedoms (DoFs) robotic arm [23]. At each step, an action consists of 7 DoFs with specific physical significance in three-dimensional Cartesian space, represented by:", + "bbox": [ + 511, + 390, + 906, + 450 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nA = \\left[ \\Delta P _ {x}, \\Delta P _ {y}, \\Delta P _ {z}, \\Delta R _ {x}, \\Delta R _ {y}, \\Delta R _ {z}, \\text {g r i p p e r} \\right], \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 550, + 460, + 903, + 476 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\Delta P_{x,y,z}$ and $\\Delta R_{x,y,z}$ denote relative positional and rotational changes along the x-, y-, and z-axis, discretized into 256 uniform bins across each DoFs' action bounds $[y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i]$ [34]. The gripper state is binary, indicating its open or closed state. This control design presents a unique challenge for adversarial attacks, as finely divided bins result in minimal action discrepancies between neighboring bins (e.g., $\\pm 0.007/\\mathrm{bin}$ ). This means that even if an attack shifts the classification to an adjacent bin, the resulting action discrepancy has minimal impact on real-world performance.", + "bbox": [ + 511, + 486, + 906, + 637 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Untargeted Action Discrepancy Attack", + "text_level": 1, + "bbox": [ + 511, + 647, + 846, + 662 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To exacerbate action discrepancies, we introduce the Untargeted Action Discrepancy Attack (UADA), which aims to maximize deviations in robot actions. This attack is based on the observation that larger robot actions usually correlate with intense physical movements, which, in turn, may amplify the potential for real-world hazards [28-30]. For UADA, the attack target is one or a combination of 7 DoFs, defined as $y_{gt}^{I} = \\{y_{gt}^{i}|i\\in [1,\\dots ,7]\\}$ . To define UADA's objective, we first identify the most distant action $y_{adv}^{i}$ , which maximizes the discrepancy from the $i$ -th DoF ground truth action $y^{i}$ . This action, for each DoF, corresponds to its own action bound $[y_{\\min}^{i},y_{\\max}^{i}]$ as:", + "bbox": [ + 511, + 669, + 908, + 853 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\ny _ {a d v} ^ {i} = \\left\\{ \\begin{array}{l l} y _ {\\max } ^ {i}, & \\text {i f} | y _ {\\max } ^ {i} - y _ {g t} ^ {i} | \\geq | y _ {\\min } ^ {i} - y _ {g t} ^ {i} | \\\\ y _ {\\min } ^ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 532, + 863, + 906, + 905 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "6950", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Instead of directly using $y_{adv}^{i}$ as the misclassification target, we introduce a soft attack objective to capture the discrepancy between actions, ensuring smooth gradient optimization and stable attack performance. Specifically, due to the physical action magnitude information contains in bin labels, we reweight the output probability $\\mathcal{F}(x)_{bins}^{i} \\in \\mathcal{R}^{1 \\times bins}$ using normalized bin labels $y_{bins}^{i} = [\\frac{1}{bins}, \\dots, 1]$ , where $\\lfloor y_{bins}^{i} \\rfloor$ and $\\lceil y_{bins}^{i} \\rceil$ correspond to the normalized values of $y_{\\min}^{i}$ and $y_{\\max}^{i}$ , respectively. The reweighting process is defined as:", + "bbox": [ + 89, + 90, + 483, + 227 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\ny _ {s o f t} ^ {i} = \\sum_ {b i n s = 1} ^ {n} \\mathcal {F} (x + \\delta) _ {b i n s} ^ {i} \\otimes y _ {b i n s} ^ {i}, \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 158, + 239, + 482, + 280 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\otimes$ denotes Hadamard Product, $\\delta$ is the adversarial patch and $y_{\\text{soft}}^i$ represents the $i$ -th DoF's soft action. Finally, the objective of UADA is to minimize the discrepancy between the soft and the most distant actions, which is:", + "bbox": [ + 89, + 289, + 483, + 349 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U A D A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\sum_ {i} ^ {I} \\left(y _ {s o f t} ^ {i} - y _ {a d v} ^ {i}\\right) ^ {2}, \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 153, + 361, + 482, + 402 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "UADA defines the objective function $\\mathcal{L}_{\\mathrm{UADA}}$ tailored for attacking the robot's action space, allowing adversarial patches to create significant, lasting disruptions in task performance.", + "bbox": [ + 89, + 412, + 483, + 459 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.3. Untargeted Position-aware Attack", + "text_level": 1, + "bbox": [ + 89, + 468, + 388, + 484 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We explore untargeted adversarial attacks in parallel by considering the positional dynamics within VLA models. In typical task execution, precise and directed movements of the end-effector towards designated goals are essential, with actions mapped in three-dimensional space. The positional DoFs, $A^p = [\\Delta P_x, \\Delta P_y, \\Delta P_z]$ encapsulates the directional movements within 3D for effective robotic control.", + "bbox": [ + 89, + 491, + 483, + 595 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Recognizing the importance of $A^p = DT(y^p)$ in controlling the end-effector's path, we introduce a position-aware attack to disrupt the intended movement trajectory. This type of adversarial vulnerability remains largely unexplored but has significant implications for task failure: the attack objective can steer the end-effector away from its intended path by introducing directional perturbations, amplifying errors and causing substantial trajectory distortions. Formally, we define the Untargeted Position-aware Attack (UPA) objective as:", + "bbox": [ + 89, + 597, + 483, + 733 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U P A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ \\alpha \\frac {y _ {a d v} ^ {p} \\cdot y ^ {p}}{\\| y _ {a d v} ^ {p} \\| \\| y ^ {p} \\|} + \\beta \\frac {1}{\\| y _ {a d v} ^ {p} - y ^ {p} \\| _ {2}} \\right], \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 109, + 753, + 482, + 786 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\alpha$ and $\\beta$ are the hyperparameters that balance between the directionality and magnitude of the perturbations. By integrating geometric awareness into our attack, this approach generates perturbations that induce cumulative deviations in the robot's trajectory, even with minimal adjustments. $L_{2}$ -normalization $\\| \\cdot \\| _2$ further intensifies these deviations, rendering the attack highly effective and disruptive.", + "bbox": [ + 89, + 794, + 483, + 900 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.4. Targeted Manipulation Attack", + "text_level": 1, + "bbox": [ + 513, + 90, + 784, + 106 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In addition to the aforementioned untargeted attacks, we also explore the adversarial vulnerabilities of VLA models with a targeted strategy. This attack aims to directly mislead the models into predicting specific trajectories, resulting in precise alterations. We design this attack since the targeted alterations can drive malicious behaviors or induce task failure. The objective of the targeted manipulation attack is:", + "bbox": [ + 511, + 112, + 906, + 218 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {T M A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {T} ^ {I}\\right) \\right], \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 581, + 223, + 906, + 242 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $y_{T}^{I} = \\{y_{T}^{i} = t|i\\in [1,\\dots ,7],t\\in [y_{\\min}^{i},y_{\\max}^{i}] \\}$ represents the adversarial target(s); and $CE(\\cdot ,\\cdot)$ denotes the Cross-Entropy loss, which measures the error between the predicted state under adversarial perturbation and the desired target state. By steering the robot toward an adversarial target across successive time steps, our approach manipulates the trajectory and undermines task performance, ultimately leading to a substantial disruption of intended operations. Furthermore, manipulating any individual DoF or a combination thereof can also severely compromise task success, given that the errors in one dimension tend to propagate and magnify over time during execution (see Fig. 3).", + "bbox": [ + 511, + 247, + 908, + 431 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.5. Metric: Normalized Action Discrepancy", + "text_level": 1, + "bbox": [ + 511, + 438, + 857, + 455 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To assess the potential physical impact of adversarial patches, we introduce Normalized Action Discrepancy (NAD) as a fine-grained metric for evaluating discrepancies at the action level. NAD serves as a complementary measurement alongside the coarse-grained Failure Rate (FR) at the task level, providing a detailed analysis of deviations during action execution. To define the NAD, we first determine the maximum action discrepancy $d_{\\mathrm{max}}^i$ , which represents the largest deviation that can occur within the allowed range for $i$ -th DoF. This is computed by measuring the L1 distance between the ground truth $y_{gt}^i$ and its action bound $[y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i]$ as:", + "bbox": [ + 511, + 460, + 908, + 628 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nd _ {\\mathrm {m a x}} ^ {i} (y) = \\max \\left[ D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m i n}} ^ {i} | \\big), D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m a x}} ^ {i} | \\big) \\right]. (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 526, + 636, + 906, + 655 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "We next calculate the applied action discrepancy $d_{\\mathrm{applied}}^i (x,y)$ $= |DT(\\mathcal{F}(x)^i) - DT(y_{gt}^i)|$ to measure the deviation between the model's output and ground truth. We define the NAD as the drifting ratio between $d_{\\mathrm{applied}}^i$ and $d_{\\mathrm{max}}^i$ as:", + "bbox": [ + 511, + 660, + 906, + 726 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathrm {N A D} = \\frac {1}{\\mathrm {I}} \\sum_ {\\mathrm {i}} ^ {\\mathrm {I}} \\frac {\\mathrm {d} _ {\\text {a p p l i e d}} ^ {\\mathrm {i}}}{\\mathrm {d} _ {\\max } ^ {\\mathrm {i}}}, \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 630, + 734, + 906, + 773 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $I \\in [1, \\dots, 7]$ represents a single or a combination of DoF(s). In this way, NAD provides normalized measurement of action deviations, allowing for a consistent evaluation of adversarial impact across varying DoF(s).", + "bbox": [ + 511, + 784, + 908, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 511, + 851, + 643, + 867 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In this section, we first detail the implementation of our adversarial framework and baseline methods in §4.1. Next, we", + "bbox": [ + 511, + 869, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "6951", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/053d65c210e902c1eb3f997270755a01597bfa93840980fb3420e9367d440e72.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 161, + 113, + 589, + 243 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/56b3f6c182dadf720ec5078e7ad1dc0489d0d62a8e2794270f803036b9f8ec51.jpg", + "image_caption": [ + "TASK (UADA) : Pick up the black bowl between the plate and the ramekin and place it on the plate" + ], + "image_footnote": [], + "bbox": [ + 601, + 112, + 810, + 244 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/e568da881f28597e9b6b015d091a52624fbabdb43f3356c6d1cf874c6c3ffe94.jpg", + "image_caption": [ + "TASK (UPA) : Pick up the black bowl on the ramekin and place it on the plate" + ], + "image_footnote": [], + "bbox": [ + 158, + 265, + 589, + 397 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/17fa09dc76817dc102176fa4d847ebf1f5c9b294e44ed926ac405b8cb258b33b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 604, + 266, + 810, + 397 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/68bbb29a18557a64d22e434219a261b2da93f2cdbfab928084518b2d8d9efed1.jpg", + "image_caption": [ + "TASK (TMA) : Pick up the alphabet soup and place it in the bask" + ], + "image_footnote": [], + "bbox": [ + 161, + 419, + 589, + 551 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/aeae4fff3aed917ce687bb61bf292a422d815551dcb583cffeda6f27f537b83e.jpg", + "image_caption": [ + "Figure 3. Qualitative Results of adversarial vulnerabilities over OpenVLA-7B [34] and OpenVLA-7B-LIBERO [34] with objectives of UADA, UPA, and TMA, respectively. We visualize the overall 3D trajectories and 2D trajectories of benign $\\bullet$ and adversarial $\\bullet$ scenarios at each time step to compare the impact of the generated adversarial patch in affecting them. The untargeted trajectory $\\bullet$ is visualized in UADA task. All trajectories start with $\\triangle$ , and we plot the success end point, marked as $\\star$ ." + ], + "image_footnote": [], + "bbox": [ + 607, + 420, + 812, + 554 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "outline the experimental setup in $\\S 4.2$ and present quantitative and qualitative results in both simulation and physicalworld scenarios (§4.3). We then conduct diagnostic experiments (§4.4) to analyze the impact of key components. Additionally, we assess the robustness of our method against various defense mechanisms in $\\S 4.5$ and conclude with a comprehensive systemic discussion in $\\S 4.6$ .", + "bbox": [ + 89, + 623, + 485, + 729 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 738, + 307, + 753 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "The implementation of our adversarial patch attack pipeline is detailed in Algorithm 1. We incorporate two key modifications to enhance the training stability of the generated patch. Inner-loops. Following previous work [49], we add $k$ inner-loops optimize steps during each iteration in line 4, aiming to reduce data variance and ensure more consistent updates. Geometric Transformations. To improve the robustness of our attack under real-world scenarios, we employ random geometric transformations $\\mathcal{T}[(\\cdot), (sh_x, sh_y, \\theta)]$ in line 6 with", + "bbox": [ + 89, + 763, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Algorithm 1: Adversarial Patch Attack.", + "text_level": 1, + "bbox": [ + 524, + 626, + 787, + 640 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/76ee417d291e6b51af26087be62d338fdb778dbe3d37a36813b0ad31817ecf74.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
1:Input: X: dataset; δ: patch; L0: attack objective; \nF: VLA model; T(·): transformations; φ,ψ: \ntransformation parameters; T: attack steps; k: \ninner-loop steps.
2:Initialize: δ ~ U[0,1), L0 ∈ {LUADA, LUPA, LTMA}
3:for i = 1,2,···, T do
4:for i = 1,2,···, k do
5:shx, shy ~ U(-φ,φ), θ ~ U(-ψ,ψ)
6:ypred ← F[T((x + δ), (shx, shy, θ))]
7:δ ← ∂L0/∂δ{▷ Update}
8:end for
9:end for
", + "bbox": [ + 542, + 643, + 903, + 825 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "affine and rotation transformations [3] as:", + "bbox": [ + 513, + 854, + 790, + 869 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {T} \\left[ (\\cdot), \\left(s h _ {x}, s h _ {y}, \\theta\\right) \\right] = \\left[ \\begin{array}{c c} 1 & s h _ {y} \\\\ s h _ {x} & 1 \\end{array} \\right] \\cdot \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right], \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 522, + 873, + 903, + 905 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "6952", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $sh_x, sh_y \\sim \\mathcal{U}(-\\phi, \\phi)$ are shear factors, and $\\theta \\sim \\mathcal{U}(-\\psi, \\psi)$ denotes the rotation angle, both sampled from uniform distributions during inner loop optimization step $k$ . Untargeted Manipulation Attack (UMA) Baseline. As a pioneering work, we could not find a directly comparable baseline for our attacks. Therefore, we adapt prior work in adversarial learning as one of our baseline methods [66]. This attack is designed to mislead classifiers; we integrate it into our framework, leading to the following objective:", + "bbox": [ + 89, + 90, + 483, + 229 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {U M A}} = - \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {g t} ^ {I}\\right) \\right]. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 151, + 237, + 482, + 257 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "By minimizing the negative CE loss, this attack pushes the predicted action away from the ground truth action.", + "bbox": [ + 89, + 266, + 482, + 296 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Random Noise Baseline. We generate random noise patches as an additional baseline, representing an unstructured intervention without any learned adversarial intent. Specifically, the noise is sampled from a Gaussian distribution $\\mathcal{N}(0,1)$ and evaluated with the same setting as proposed attacks.", + "bbox": [ + 89, + 297, + 483, + 387 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Experiment Setup", + "text_level": 1, + "bbox": [ + 89, + 396, + 267, + 412 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Datasets. We conduct attacks on BridgeData V2 [69] and LIBERO [44] with corresponding VLAs. BridgeData V2 [69] is a real-world dataset comprising 24 diverse environments and 13 distinct skills, such as grasping, placing, and object rearrangement, with a total of 60,096 trajectories. LIBERO [44] is a simulation dataset designed to evaluate models across four distinct task types (Spatial, Object, Goal, Long). Notably, LIBERO-Long combines diverse objects, layouts, and long-horizon tasks, making them challenging for complex, multi-step planning, thus we choose LIBERO-Long to conduct UADA, UPA and TMA.", + "bbox": [ + 89, + 416, + 483, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Victim VLAs. In our study, we select publicly available and current most influential VLAs as victim models for evaluation. Specifically, we focus on four variants of the OpenVLA model, each trained independently on different task suites within the LIBERO dataset [44] (i.e., Spatial, Object, Goal, Long). To evaluate the effectiveness of our methods, we craft adversarial patches using three distinct generating setups: Simulation Setting involves a model trained in a simulated environment using the LIBERO Long suite with the openvla-7B-libero-long variant [34]. Physical Setting uses a model trained on physical world data from the BridgeData v2 [69] with the openvla-7B model [34]. This approach allows us to assess our methods on both simulated and real-world data. Subsequently, we evaluate the performance of generated adversarial patches on victim models (i.e., OpenVLA LIBERO variants) trained on different tasks suites to rigorously prove the robustness and effectiveness of our method. These models are trained with distinct data sources and task objectives.", + "bbox": [ + 89, + 583, + 483, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Robot Setups (Real-world Setting). For the real-world tasks, we adopt a robotic system consisting of a Universal Robots UR10e equipped with a Robotiq Hand-E Gripper to", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "provide a 7DoF motion. The sensing system includes one RGB webcam in the fixed position with a shoulder view.", + "bbox": [ + 511, + 90, + 903, + 119 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Details. To assess the effectiveness and robustness of our methods, we conduct experiments on the LIBERO dataset [44]. Each suite consists of 10 tasks, with each task executed for 50 trials, resulting in a total of 500 rollouts, following Kim et al. [34]. For reproducibility, we carefully select patch paste locations for each suite, ensuring they do not obscure objects or robots in the test environment.", + "bbox": [ + 511, + 121, + 906, + 226 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Evaluation Metric. Regarding the task execution evaluation, we take the maximum steps of each task suite in the LIBERO training dataset as the timeout failure condition to reduce computational overhead. Furthermore, building on the concept of Success Rate (SR) introduced in LIBERO [44], we adopt Failure Rate (FR), defined as $1 - \\mathrm{SR}$ , as the primary evaluation metric. To further quantify action discrepancy, we employ NAD (see Eq. 8) for untargeted attacks on corresponding DoF(s), as it considers relative distances to quantify deviation from the optimal action. A higher NAD value indicates a greater discrepancy between the ground truth and the predicted action. Additionally, we report the standard deviation of FR across tasks within each task suite.", + "bbox": [ + 511, + 227, + 908, + 436 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.3. Main Result", + "text_level": 1, + "bbox": [ + 513, + 443, + 643, + 455 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Quantitative Results. For UADA and UPA, our methods effectively amplify action discrepancies, leading to a notable transfer attack ability in increasing failure rates (see Tab. 1a). Specifically, while attacking $\\mathrm{DoF}_1$ and $\\mathrm{DoF}_{1\\sim 3}$ in the Simulation setup, UADA and UPA achieve NAD of $21.0\\%$ and $14.5\\%$ , significantly outperforming UMA scenarios with increments of $6.9\\%$ and $3.1\\%$ , respectively. Both UADA and UPA effectively disrupt robot execution, yielding maximum average failure rates of $100\\%$ and $89.7\\%$ , respectively. Regarding the Physical setup (see Tab. 1b), our attack methods demonstrate strong transferability, as adversarial patches generated in the physical setting significantly impact the performance of LIBERO-X test. Specifically, both UADA and UPA generate malicious actions with large action discrepancies, which are $32.6\\%$ ( $\\mathrm{DoF}_1$ ) and $26.9\\%$ ( $\\mathrm{DoF}_{1\\sim 3}$ ) with an increase of $5.9\\%$ and $3.1\\%$ compared to $26.7\\%$ ( $\\mathrm{DoF}_1$ ) and $23.8\\%$ ( $\\mathrm{DoF}_{1\\sim 3}$ ) in UMA scenarios, respectively. Compare the two settings (Tab. 1a v.s. Tab. 1b), we observe a general variance in NAD (e.g. $32.6\\%$ v.s. $21.0\\%$ for UADA). This discrepancy can be attributed to the fundamental differences between the simulation and real-world datasets used for conducting attacks (Bridge V2 [69] v.s. LIBERO [44]). The increased variability in real-world data, including environmental complexity, object diversity, and task difficulty, allows the robot more opportunities to generate larger action discrepancies within the validation dataset.", + "bbox": [ + 511, + 463, + 908, + 853 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "For TMA task, we evaluate the effectiveness of our method by manipulating all DoF(s) to 0 (i.e., $y_{T}^{i} = 0$ in Eq. 6). Our method demonstrates significant effectiveness", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "6953", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/8e3470a77711c197f73486545df09d246217489a8292513a8ffd640ad8d75122.jpg", + "table_caption": [ + "Table 1. Untargeted Results. We report FR and NAD in LIBERO simulation. * denotes the in-domain victim model and dataset aligned with the patch generation model and dataset, △ denotes a transfer attack evaluation involving a distinct victim model and dataset (see §4.2). The FR (↑) is highlighted in best and second best for each task suite." + ], + "table_footnote": [ + "(a) Simulation Setup. Adversarial patch generated on OpenVLA-LIBERO-10 victim model with LIBERO-Long dataset and evaluated on LIBERO-X." + ], + "table_body": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long*Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF114.1100±0.0%99.0±3.0%100±0.0%100±0.0%99.8%
DoF1~311.4100±0.0%98.8±2.6%99.0±2.4%100±0.0%99.5%
UADADoF121.0100±0.0%99.2±2.4%100±0.0%100±0.0%99.8%
DoF1~317.9100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~314.596.2±11.4%77.8±12.5%88.0±10.4%96.8±3.0%89.7%
", + "bbox": [ + 91, + 133, + 493, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2ccd3461e905ec0aa04f3a3175f0e13ec21363d6a253c4f469ec22f9bc917f8a.jpg", + "table_caption": [], + "table_footnote": [ + "(b) Physical Setup. Adversarial patch generated on OpenVLA-7b victim model with BV2 dataset and evaluated on LIBERO-X." + ], + "table_body": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long△Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF126.789.6±15.3%60.8±16.8%64.8±26.8%80.0±20.8%73.8%
DoF1~323.896.6±6.2%56.4±20.4%80.0±20.5%82.0±17.9%78.8%
UADADoF132.699.2±1.8%98.8±2.4%92.6±15.2%96.6±4.8%96.8%
DoF1~327.7100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~326.995.6±9.5%57.0±23.6%57.2±24.4%72.8±19.7%70.7%
", + "bbox": [ + 500, + 133, + 901, + 239 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/a15e18d92c0fcb6be49c990cb7fed52fec6e594e9d86a4ee69ec12aba4592436.jpg", + "table_caption": [ + "Table 2. Targeted Manipulation Attack Results. Failure Rate (FR, ↑) and its standard deviation across tasks within LIBERO [44] suite are reported. The performance of task suits and DoFs is averaged separately, with the best and second-best values separately." + ], + "table_footnote": [], + "table_body": "
Targeted DoF(s)MetricSimulation (Attacked on OpenVLA-LIBERO-10)Physical (Attacked on OpenVLA-7b)
Spatial△Object△Goal△Long*AvgSpatial△Object△Goal△Long△Avg
BenignFR15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
DoF1FR100±0.0%97.8±3.0%100±0.0%100±0.0%99.5%90.6±19.9%36.6±18.1%61.0±21.7%86.4±12.7%68.7%
DoF2100±0.0%99.2±1.3%100±0.0%100±0.0%99.8%99.6±0.8%69.2±11.1%76.2±26.1%91.8±9.9%84.2%
DoF3100±0.0%97.2±6.0%100±0.0%100±0.0%99.3%72.0±14.8%30.2±17.4%39.2±20.9%62.4±17.5%51.0%
DoF494.0±11.0%37.4±23.2%51.2±26.5%87.6±14.9%67.6%93.6±9.7%31.8±20.3%45.0±18.3%63.8±17.0%58.6%
DoF5100±0.0%92.6±8.4%100±0.0%100±0.0%98.2%97.8±4.9%36.0±22.5%62.0±24.8%79.4±21.3%68.8%
DoF6100±0.0%80.0±9.5%92.6±13.4%100±0.0%93.2%79.0±14.2%22.2±14.4%41.0±21.5%62.6±15.7%51.2%
DoF799.2±1.3%86.6±8.3%67.2±22.5%96.4±6.4%87.4%91.8±15.3%98.8±1.3%70.6±23.4%93.2±10.2%88.6%
DoF1~3100±0.0%100±0.0%100±0.0%100±0.0%100%96.4±8.0%83.6±16.8%74.4±26.3%91.8±11.9%86.6%
Task Avg99.2%86.4%88.9%98.0%-90.1%51.1%58.7%78.9%-
", + "bbox": [ + 91, + 304, + 903, + 465 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "in manipulating robotic trajectory and increasing FR (see Tab. 2). Specifically, when applied to the different generation settings, our approach yields notable increments across various tasks, including a max average failure rate $100\\%$ v.s. $23.5\\%$ of benign performance in Simulation, and $88.6\\%$ in Physical, respectively. Moreover, during attack $\\mathrm{DoF}_4$ , we observe low FR and high task deviation. This failure can be attributed to the fact that $\\mathrm{DoF}_4$ controls the orientation along the x-axis, which can be redundant DoF in tasks. As a result, attacks targeting redundant DoF(s) [38] are less likely to disrupt execution effectively. This observation underscores the importance of task-specific considerations when designing adversarial attacks on robotic systems.", + "bbox": [ + 88, + 472, + 485, + 667 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Qualitative Results. We qualitatively analyze robot movement trajectories under the three proposed attacks in Fig. 3. For UADA, we compare the trajectory deviations in the same trail with patches generated from UADA and UMA attacks. As seen, the UMA induces small deviations in the trajectory. UADA, on the other hand, produces significantly larger trajectory deviations (also supported by NAD metric in Tab. 1b), which is attributed to UADA's capability of incorporating action discrepancies. In our observation, UADA significantly amplifies the effect on overall task execution, thereby increasing the potential for robotic hazards. For UPA, we observe chaotic and irregular behaviors, including instances where the end-effector moves out of the camera's field of view. We attribute this phenomenon to the efficacy of the adversarial patch in perturbing the model's spatial perception,", + "bbox": [ + 88, + 674, + 486, + 901 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "inducing a consistent deviation from the intended direction of movement, ultimately resulting in failure in a long run. For TMA, we observe a notable reduction in the range of motion along the x-axis corresponding to the targeted attack axis. This suggests that our proposed targeted attack can effectively constrained robot movements.", + "bbox": [ + 511, + 472, + 906, + 561 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "In summary, the qualitative analysis shows that both Untarget attack, UADA, UPA, and TMA can effectively disrupt the robot actions generated from VLA models. These findings underscore a pressing security concern during the deployment of generalist robots, especially when considering application scenes that require reliable operations [11, 68].", + "bbox": [ + 511, + 563, + 908, + 654 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0161ac097af27979f5c686f293cee7b1fa3c957783c496e98982d3444ce4ada4.jpg", + "image_caption": [ + "Figure 4. Qualitative Results of the physical world. The first/second row show benign and adversarial cases respectively." + ], + "image_footnote": [], + "bbox": [ + 537, + 657, + 879, + 779 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Real-world Performance. In addition to the digital simulation results, we also conduct a comprehensive evaluation of the performance of our generated adversarial patches in real-world scenarios. The evaluation encompassed 100 trials across three distinct tasks: object grasping, placement, and manipulation. As shown in Fig. 4, the adversarial patch", + "bbox": [ + 511, + 810, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "6954", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/3fde9a9844a89860e0754a29d91bdf04c8255d59b85cad5673fd228251be5c83.jpg", + "image_caption": [ + "(a) Inner-loop" + ], + "image_footnote": [], + "bbox": [ + 114, + 89, + 202, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/809061dfd5fe3958fdfd83eb8711dd71b12036eafb386de14c8b3a2b22dae332.jpg", + "image_caption": [ + "(b) Patch Size" + ], + "image_footnote": [], + "bbox": [ + 202, + 89, + 351, + 190 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aa8c52338edb1b5e3959b050bc5389712291477bbcde871f240f2d15a3a4a905.jpg", + "image_caption": [ + "(c) JPEG" + ], + "image_footnote": [], + "bbox": [ + 352, + 89, + 488, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/50e2ebe37466de40a562310a4612b267b5cf1fe5328d4545b824eccc09b63551.jpg", + "image_caption": [ + "(d) Noise" + ], + "image_footnote": [], + "bbox": [ + 488, + 89, + 616, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/aec35d748df36db7070db39a02bd5670bfdacc54ab9f44352ab116361f175606.jpg", + "image_caption": [ + "(e) Median Blur" + ], + "image_footnote": [], + "bbox": [ + 616, + 89, + 743, + 191 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/eb35dcf69370a82268f417f57d53cb4d9376908e211032ace0d6e0c65cfb30b6.jpg", + "image_caption": [ + "(f) Bit-Depth" + ], + "image_footnote": [], + "bbox": [ + 745, + 89, + 875, + 191 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "generated with UADA demonstrated the ability to manipulate the robot effectively, achieving an attack success rate exceeding $43\\%$ . Although this success rate is lower than the corresponding digital-world performance (i.e., $100\\%$ ), it highlights the effectiveness of our patches in physical-world applications as well without the need for further adaptations. Crucially, we observed that the induced erratic movements (similar to simulation scenario) of the robot during successful attacks pose significant risks to human safety and surrounding environments. This finding emphasizes the severe threat posed by VLA models in real-world applications, consistent with similar observations in digital settings.", + "bbox": [ + 88, + 252, + 485, + 434 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.4. Diagnostic Experiment", + "text_level": 1, + "bbox": [ + 89, + 439, + 303, + 455 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Inner-loops. We discuss the impact of inner-loop steps to model performance in Fig. 5(a). The results show that NAD first improves when inner-loop steps continue to increase. This is due to the reduced data variance and gradually stabilized gradients. However, further increasing the steps results in a slower performance increase, with a noticeably longer training schedule. We thus choose inner-loop steps=50 as it balances between performance and scale.", + "bbox": [ + 89, + 458, + 483, + 578 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Impact of Patch Size. We study TMA's performance with different patch sizes and report the average FR across four LIBERO [44] tasks in Fig. 5(b). The patch sizes examined are $[3\\%, 5\\%, 7\\%, 10\\%]$ of input image. Our findings indicate an inversely proportional relationship between the $L_{1}$ distance of predicted actions and the target action as the patch size increased. This increase in patch size also correlated with a marked rise in FR. The observed trend suggests larger patches give adversaries more optimization space to influence the model, aligning with prior work [12, 81].", + "bbox": [ + 89, + 579, + 483, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.5. Robustness Evaluation", + "text_level": 1, + "bbox": [ + 89, + 734, + 302, + 750 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We further examine whether concurrent defense strategies [18, 78, 80] can resist our adversarial examples within VLA. Specifically, we applied four prior defense techniques and reported their FR for both benign and adversarial samples in Fig. 5(c-f), respectively. The results show that our adversarial attack bypasses most defense strategies.", + "bbox": [ + 89, + 755, + 483, + 845 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.6. Systemic Discussion", + "text_level": 1, + "bbox": [ + 89, + 849, + 279, + 866 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Patch Pattern Analysis. In Fig. 6, we present the adversarial patches generated for a range of targets. A particularly", + "bbox": [ + 89, + 869, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6647818771e16f42e30d226e9343a943174dc32d516535cce66e32304da2161e.jpg", + "image_caption": [ + "Figure 5. Impact of Inner-loop, Patch Size and Defense Discussion. The figure shows how varying Inner-loop affects NAD in UADA, and patch sizes affect $L1$ distance and the failure rates in TMA, both targeting at DoF $_1$ . (a) Impact of Inner-loop, (b) Impact of Patch Size and (c-f) the effect of four different defenses on failure rates." + ], + "image_footnote": [], + "bbox": [ + 521, + 253, + 576, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/b0278a6162a00b426c5c683f1356a894d55279d8a5fa69166b1d9654d68b6fb3.jpg", + "image_caption": [ + "(a) BridgeData V2 [69]" + ], + "image_footnote": [], + "bbox": [ + 584, + 253, + 640, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/270053252a7867145c80093655706037c8dda7c601624b1ae2fb649b325f4f39.jpg", + "image_caption": [ + "Figure 6. Patch Visualization. Semantic-rich patches." + ], + "image_footnote": [], + "bbox": [ + 648, + 253, + 704, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ee3a2ee1d1681df590d280ecb23433001c527085eda6cfa90dfa87f3741023b3.jpg", + "image_caption": [ + "(b) LIBERO [44]" + ], + "image_footnote": [], + "bbox": [ + 710, + 253, + 767, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/3c4da73883a24268bd55d0e3459b8a31f977fd15624a301f0ce345ffc9426d13.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 774, + 253, + 831, + 297 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6c71575b7ab846bd94392342ea4be2ad4acfe3eb7f9991a7caa58911db48c16b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 838, + 253, + 893, + 297 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "noteworthy observation is that some of these patches bear a striking resemblance to the structural joints of a robotic arm. Given the strong similarity between these patches and robotic arms in appearance, we hypothesize that these VLA-based models, in their current training paradigms, are often limited to a narrow range of robotic systems operating within restricted camera views. This constrained perspective camera pose may inadvertently induce a learned representation bias, where adversarial perturbations align with prominent structural features of the training robot's appearance in order to deceive current victim models. Given that these patches are model-specific and reflect the spatial and visual patterns of the training environment, we have a reasonable concern that they may compromise VLA-based models' generalizability and robustness in physical-world deployments.", + "bbox": [ + 511, + 338, + 906, + 564 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Possible Defense Strategies. Our work highlights the pressing need for a paradigm shift in the VLA-based models' training strategies in order to successfully defend our attacks. Specifically, future approaches might incorporate more training scenarios that involve complex multi-robot interactions to mitigate the current patch bias. Another possible solution is to leverage the logical relationships within the robot arm's physical structure to estimate the actual position of each joint. By predicting joint positions deviated from physical constraints, we can eliminate unreasonable adversarial samples, even if they are similar to actual joint(s).", + "bbox": [ + 511, + 565, + 908, + 731 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 741, + 633, + 756 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "While VLA models have gained significant popularity due to their substantial robot capabilities, this study pioneers specific robotic attack objectives and demonstrates that these models are in fact vulnerable to various types of attacks. Experimental results demonstrate that our attacks expose significant vulnerabilities to VLA models, raising concerns regarding impetuous real-world deployments. We believe our framework provides pioneering and foundational contributions to the reliability of AI-enhanced robotic systems.", + "bbox": [ + 511, + 763, + 906, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "6955", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements", + "text_level": 1, + "bbox": [ + 91, + 90, + 258, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This research was supported by the National Science Foundation under Grant No. 2450068 and No. 2242243, as well as start-up resources provided by Rutgers University. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of U.S. Naval Research Laboratory (NRL) or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein.", + "bbox": [ + 89, + 114, + 485, + 267 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 281, + 187, + 296 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Mokhaled NA Al-Hamadani, Mohammed A Fadhel, Laith Alzubaidi, and Balazs Harangi. Reinforcement learning algorithms and applications in healthcare and robotics: A comprehensive and systematic review. Sensors, 24(8):2461, 2024.", + "[2] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018.", + "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In ICML, 2018.", + "[4] Lefteris Benos, Avital Bechar, and Dionysis Bochtis. Safety and ergonomics in human-robot interactive agricultural operations. Biosystems Engineering, 200:55-72, 2020.", + "[5] Nicole Berx, Wilm Decré, Ido Morag, Peter Chemweno, and Liliane Pintelon. Identification and classification of risk factors for human-robot collaboration from a system-wide perspective. Computers & Industrial Engineering, 163:107827, 2022.", + "[6] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446), 2019.", + "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In Robotics: Science and Systems XIX, 2023.", + "[8] Tom B. Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gil mer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2018.", + "[9] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 3-14, 2017." + ], + "bbox": [ + 99, + 306, + 485, + 900 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023.", + "[11] Shen Cheng and BS Dhillon. Reliability and availability analysis of a robot-safety system. Journal of Quality in Maintenance Engineering, 17:203-232, 2011.", + "[12] Zhiyuan Cheng, Cheng Han, James Liang, Qifan Wang, Xi-angyu Zhang, and Dongfang Liu. Self-supervised adversarial training of monocular depth estimation against physical-world attacks. arXiv preprint arXiv:2406.05857, 2024.", + "[13] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility via: Multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024.", + "[14] Jinda Cui and Jeff Trinkle. Toward next-generation learned robot manipulation. Sci. Robotics, 6(54):9461, 2021.", + "[15] Yiming Cui, Cheng Han, and Dongfang Liu. Collaborative multi-task learning for multi-object tracking and segmentation. Journal on Autonomous Transportation Systems, 1:1-23, 2024.", + "[16] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. *Quar-vla: Vision-language-action model for quadruped robots*. In *ECCV*, 2025.", + "[17] Andrew Du, Bo Chen, Tat-Jun Chin, Yee Wei Law, Michele Sasdelli, Ramesh Rajasegaran, and Dillon Campbell. Physical adversarial attacks on an aerial imagery object detector. In WACV, 2022.", + "[18] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M. Roy. A study of the effect of JPG compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016.", + "[19] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In CoRL, 2017.", + "[20] Saskia Golz, Christian Osendorfer, and Sami Haddadin. Using tactile sensation for learning contact knowledge: Discriminate collision from physical interaction. In ICRA, 2015.", + "[21] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015.", + "[22] Jérémie Guiochet, Mathilde Machin, and Hélène Waeselynck. Safety-critical advanced robots: A survey. Robotics Auton. Syst., 94:43-52, 2017.", + "[23] Sami Haddadin, Sven Parusel, Lars Johannsmeier, Saskia Golz, Simon Gabl, Florian Walch, Mohamadreza Sabaghian, Christoph Jähne, Lukas Hausperger, and Simon Haddadin. The franka emika robot: A reference platform for robotics research and education. IEEE Robotics & Automation Magazine, 29:46-64, 2022.", + "[24] Roni-Jussi Halme, Minna Lanz, Joni Kämäräinen, Roel Pieters, Jyrki Latokartano, and Antti Hietanen. Review of vision-based safety systems for human-robot collaboration. Procedia Cirp, 72:111-116, 2018." + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "6956", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[25] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023.", + "[26] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. In CoRL, 2023.", + "[27] Félix Ingrand and Malik Ghallab. Deliberation for autonomous robots: A survey. Artif. Intell., 247:10-44, 2017.", + "[28] ISO 10218-1/2:2011. Robots and Robotic Devices Safety Requirements for Industrial Robots Part 1: Robots/Part 2: Robot Systems and Integration. Standard, International Organization for Standardization, 2011.", + "[29] ISO13482. Robots and robotic devices — safety requirements for personal care robots. International Organization for Standardization, 2014.", + "[30] ISO/TS 15066:2016. Robots and Robotic Devices Collaborative Robots. Standard, International Organization for Standardization, 2016.", + "[31] Stephen James, Michael Bloesch, and Andrew J. Davison. Task-embedded control networks for few-shot imitation learning. In CoRL, 2018.", + "[32] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: general robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022.", + "[33] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024.", + "[34] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024.", + "[35] Hema Swetha Koppula and Ashutosh Saxena. Learning spatio-temporal structure from RGB-D videos for human activity detection and anticipation. In ICML, 2013.", + "[36] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE Trans. Pattern Anal. Mach. Intell., 38(1): 14-29, 2016.", + "[37] Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learning for manipulation: Challenges, representations, and algorithms. J. Mach. Learn. Res., 22:30:1-30:82, 2021.", + "[38] Andrew J. Kurdila and Pinhas Ben-Tzvi. Dynamics and control of robotic systems. John Wiley & Sons, 2019.", + "[39] Przemyslaw A. Lasota, Terrence Fong, and Julie A. Shah. A survey of methods for safe human-robot interaction. Found. Trends Robotics, 5(4):261–349, 2017.", + "[40] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for" + ], + "bbox": [ + 91, + 92, + 485, + 900 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "robotic grasping with deep learning and large-scale data collection. Int. J. Robotics Res., 37(4-5):421-436, 2018.", + "[41] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023.", + "[42] Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, and Michael S. Ryoo. Llara: Supercharging robot learning data for vision-language policy. arXiv preprint arXiv:2406.20095, 2024.", + "[43] Aishan Liu, Xianglong Liu, Jiaxin Fan, Yuqing Ma, Anlan Zhang, Huiyuan Xie, and Dacheng Tao. Perceptual-sensitive GAN for generating adversarial patches. In AAAI, 2019.", + "[44] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2023.", + "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024.", + "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurlPS, 2024.", + "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019.", + "[48] Alessandro De Luca, Alin Albu-Schäffer, Sami Haddadin, and Gerd Hirzinger. Collision detection and safe reaction with the DLR-III lightweight manipulator arm. In IROS, 2006.", + "[49] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018.", + "[50] Jim Mainprice and Dmitry Berenson. Human-robot collaborative manipulation planning using early prediction of human motion. In IROS, 2013.", + "[51] Matthew T. Mason. Mechanics of Robotic Manipulation. MIT Press, 2001.", + "[52] Matthew T. Mason. Toward robotic manipulation. Annu. Rev. Control. Robotics Auton. Syst., 1:1-28, 2018.", + "[53] Oier Mees, Lukás Hermann, and Wolfram Burgard. What matters in language conditioned robotic imitation learning over unstructured data. IEEE Robotics Autom. Lett., 7(4): 11205-11212, 2022.", + "[54] Vivek G Moudgal, Kevin M Passino, and Stephen Yurkovich. Rule-based control for a flexible-link robot. IEEE Trans. on Control Sys. Tech., 2(4):392-405, 1994.", + "[55] Richard M. Murray, Zexiang Li, and S. Shankar Sastry. A mathematical introduction to robotics manipulation. CRC Press, 1994.", + "[56] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016.", + "[57] Vitchyr Pong, Murtaza Dalal, Steven Lin, Ashvin Nair, Shikhar Bahl, and Sergey Levine. Skew-fit: State-covering self-supervised reinforcement learning. In ICML, 2020.", + "[58] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "6957", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "examples jailbreak aligned large language models. In AAAI, 2024.", + "[59] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In ACM SIGSAC, 2016.", + "[60] Michael H. Smith and L. Stephen Coles. Design of a low cost, general purpose robot. In IJCAI, 1973.", + "[61] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Trans. Evol. Comput., 23(5):828-841, 2019.", + "[62] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014.", + "[63] Chen Tang, Ben Abbatematteo, Jiaheng Hu, Rohan Chandra, Roberto Martin-Martin, and Peter Stone. Deep reinforcement learning for robotics: A survey of real-world successes. arXiv preprint arXiv:2408.03539, 2024.", + "[64] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024.", + "[65] Florian Tramér. Detecting adversarial examples is (nearly) as hard as classifying them. In ICML, 2022.", + "[66] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020.", + "[67] Valeria Villani, Fabio Pini, Francesco Leali, and Cristian Secchi. Survey on human-robot collaboration in industrial settings: Safety, intuitive interfaces and applications. Mechatronics, 55:248-266, 2018.", + "[68] Christian Vogel, Christoph Walter, and Norbert Elkmann. A projection-based sensor system for safe physical human-robot collaboration. In IROS, 2013.", + "[69] Homer Rich Walke, Kevin Black, Tony Z. Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata V2: A dataset for robot learning at scale. In CoRL, 2023.", + "[70] Taowen Wang, Yiyang Liu, James Chenhao Liang, Yiming Cui, Yuning Mao, Shaoliang Nie, Jiahao Liu, Fuli Feng, Zenglin Xu, Cheng Han, et al. $\\mathbf{M}^2$ pt: Multimodal prompt tuning for zero-shot instruction learning. In EMNLP, 2024.", + "[71] Zhibo Wang, Siyan Zheng, Mengkai Song, Qian Wang, Alireza Rahimpour, and Hairong Qi. advpattern: Physical-world attacks on deep person re-identification via adversarially transformable patterns. In ICCV, 2019.", + "[72] Zixiang Wang, Hao Yan, Zhuoyue Wang, Zhengjia Xu, Zhizhong Wu, and Yining Wang. Research on autonomous robots navigation based on reinforcement learning. In RAIIC, 2024.", + "[73] Zhijie Wang, Zhehua Zhou, Jiayang Song, Yuheng Huang, Zhan Shu, and Lei Ma. Towards testing and evaluating vision-language-action models for robotic manipulation: An empirical study. arXiv preprint arXiv:2409.12894, 2024." + ], + "bbox": [ + 91, + 90, + 483, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[74] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024.", + "[75] Wikipedia contributors. Finch (film) — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Finch_(film) & oldid=1249481595, 2024. [Online; accessed 5-November-2024].", + "[76] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024.", + "[77] Kaidi Xu, Gaoyuan Zhang, Sijia Liu, Quanfu Fan, Mengshu Sun, Hongge Chen, Pin-Yu Chen, Yanzhi Wang, and Xue Lin. Adversarial t-shirt! evading person detectors in a physical world. In ECCV, 2020.", + "[78] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018.", + "[79] Zhiyang Xu, Chao Feng, Rulin Shao, Trevor Ashby, Ying Shen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. Visionplan: Scaling human-labeled tasks in visual instruction tuning. arXiv preprint arXiv:2402.11690, 2024.", + "[80] Yuchen Zhang and Percy Liang. Defending against whitebox adversarial attacks via randomized discretization. In AISTATS, 2019.", + "[81] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quanxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, 2022.", + "[82] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023.", + "[83] Yuanyang Zhu, Zhi Wang, Chunlin Chen, and Daoyi Dong. Rule-based reinforcement learning for efficient robot navigation with space reduction. IEEE/ASME Trans. on Mechatronics, 27:846-857, 2021.", + "[84] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023." + ], + "bbox": [ + 516, + 90, + 908, + 838 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "6958", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + } +] \ No newline at end of file diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_model.json b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_model.json new file mode 100644 index 0000000000000000000000000000000000000000..cfaf493c39262321c8c353177edfd536cc1d44da --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_model.json @@ -0,0 +1,2796 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.113, + 0.13, + 0.887, + 0.175 + ], + "angle": 0, + "content": "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.203, + 0.78, + 0.259 + ], + "angle": 0, + "content": "Taowen Wang\\(^{1*}\\), Cheng Han\\(^{2*}\\), James Liang\\(^{3*}\\), \nWenhao Yang\\(^{4}\\), Dongfang Liu\\(^{1\\dagger}\\), Luna Xinyu Zhang\\(^{1\\dagger}\\), Qifan Wang\\(^{5}\\), \nJiebo Luo\\(^{6}\\), Ruixiang Tang\\(^{7\\dagger}\\)" + }, + { + "type": "text", + "bbox": [ + 0.208, + 0.274, + 0.792, + 0.329 + ], + "angle": 0, + "content": "\\(^{1}\\)Rochester Institute of Technology, \\(^{2}\\)University of Missouri - Kansas City, \\(^{3}\\)U.S. Naval Research Laboratory, \\(^{4}\\)Lamar University, \\(^{5}\\)Meta AI, \\(^{6}\\)University of Rochester, \\(^{7}\\)Rutgers University" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.362, + 0.327, + 0.379 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.394, + 0.486, + 0.759 + ], + "angle": 0, + "content": "Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a \\(100\\%\\) reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments1." + }, + { + "type": "title", + "bbox": [ + 0.092, + 0.774, + 0.222, + 0.79 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.106, + 0.795, + 0.48, + 0.824 + ], + "angle": 0, + "content": "\"First directive: A robot cannot harm a human or, through inaction, allow a human to be harmed.\"" + }, + { + "type": "text", + "bbox": [ + 0.386, + 0.826, + 0.48, + 0.841 + ], + "angle": 0, + "content": "—Finch [75]" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.363, + 0.898, + 0.532 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.535, + 0.705, + 0.639 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.534, + 0.898, + 0.64 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.644, + 0.908, + 0.714 + ], + "angle": 0, + "content": "Figure 1. Adversarial Vulnerabilities induced by malicious manipulation. (A). Illustration of adversarial threats in robotic task execution. (B). Example of semantic-rich adversarial patches generated by proposed methods. (C). Comparison of failure rates across different attack schemes (UADA, UPA, and TMA)." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "In the movie Finch, set in a post-apocalyptic future, an intelligent robot navigates complex themes, underscoring the importance of interactive safety with its human master. Once speculative, this portrayal is now increasingly plausible with the rise of Large Vision Language Models (LVLMs) [70, 79], capable of seamlessly interpreting both visual and linguistic contexts. A notable realization of this potential can be seen in Vision-Language-Action (VLA) models [16, 34, 73, 74], which integrate LVLMs into robotic systems to enable end-to-end learning that encompasses both high-level trajectory planning and low-level robot action control. While VLA models demonstrate a promising step toward generalist" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.851, + 0.221, + 0.863 + ], + "angle": 0, + "content": "*Equal contribution." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.863, + 0.236, + 0.876 + ], + "angle": 0, + "content": "† Corresponding author." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.876, + 0.419, + 0.888 + ], + "angle": 0, + "content": "\\(\\ddagger\\)Work done during Luna Xinyu Zhang's internship at RIT." + }, + { + "type": "page_footnote", + "bbox": [ + 0.111, + 0.888, + 0.216, + 0.901 + ], + "angle": 0, + "content": "1 Project homepage." + }, + { + "type": "list", + "bbox": [ + 0.109, + 0.851, + 0.419, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6948" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.168 + ], + "angle": 0, + "content": "robotic intelligence, they also introduce potential vulnerabilities that remain largely underexplored. This work, in light of this view, pioneers the understanding of the pressing need to investigate the new attack surfaces associated with VLA-based robotic systems." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.17, + 0.483, + 0.397 + ], + "angle": 0, + "content": "To gain deeper insights into these vulnerabilities, we analyze VLA-based models and general robotic operations, highlighting two key characteristics crucial for designing effective adversarial attacks. First, creating attack objectives for robotic systems requires consideration of the physical dynamics and constraints intrinsic to robotic movement. Conventional adversarial attacks, in this case, often fail to produce significant deviations in intended actions because they disregard these constraints. Second, VLA models generate control signals that function as a time series of \\( K \\)-class predictions (see §3.1), making it essential to design attacks that exploit the temporal dependencies within these sequences to cause substantial disruptions in robotic behavior. Achieving attacks that are effective in both digital simulations and real-world environments remains critical yet challenging." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.4, + 0.483, + 0.687 + ], + "angle": 0, + "content": "To address these challenges, our work intensifies the adversarial threats posed to VLA-based systems by both developing specialized attack objectives and designing effective attack methods. Specifically, we formulate an Action Discrepancy Objective aimed at maximizing the action discrepancy within VLA-based robotic systems. This strategy ensures that, at each decision point, the robot's behavior can diverge from the optimal trajectory. Additionally, we introduce Geometry-Aware Objective that considers the robot's movement in three-dimensional space, characterized by three degrees of freedom. By optimizing the cosine similarity between adversarial and ground-truth directions, we induce deviations in the robot's movement direction from its intended path, increasing the likelihood of task failure. To achieve these attack objectives, we develop straightforward yet effective Patch-Based Attacks targeting VLA-based robotic systems. This approach enables adversarial attacks in both digital and physical settings, revealing substantial vulnerabilities within the VLA-based system." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.69, + 0.483, + 0.901 + ], + "angle": 0, + "content": "Altogether, these innovations yield several significant contributions: ① This work presents a pioneering and comprehensive analysis of vulnerabilities within VLA-based robotic systems, a new paradigm for training generalized robotic policies using generative foundation models. We reveal critical threats to adversarial attacks, emphasizing the urgent need to enhance robustness before real-world deployment; ② To the best of our knowledge, we are the first work to define attack objectives specific to the powerful VLA models and to employ a straightforward adversarial patch against it (see §3). This offers valuable insights for the research community to explore systemic failures in similar concurrent generative foundation models; ③ We rigorously evaluate our approach in both simulated and real-world environments across four" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.138 + ], + "angle": 0, + "content": "distinct robotic tasks, observing significant raises in task failure rates of \\(100\\%\\) and \\(43\\%\\), respectively. This highlights the effectiveness of our attack strategies (see \\(\\S 4.3\\))." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.144, + 0.652, + 0.159 + ], + "angle": 0, + "content": "2. Related work" + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.169, + 0.763, + 0.184 + ], + "angle": 0, + "content": "2.1. AI-driven Generalist Robot" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.19, + 0.908, + 0.34 + ], + "angle": 0, + "content": "Developing generalist robots [6, 14, 27, 37, 60] requires models to not only handle varied interactions but also maintain robustness in unstructured environments. Early intelligent robots [51, 52, 54, 55, 83] generally relied on rule-based approaches, effective only in controlled settings and needing extensive reprogramming for new tasks. The advent of deep learning [19, 31, 40, 56] and reinforcement learning [1, 63, 72] has shifted this paradigm to adapt through the data-driven pipeline, increasing robot versatility and reducing manual reconfiguration." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.342, + 0.909, + 0.537 + ], + "angle": 0, + "content": "Recent studies [7, 53, 57] have further advanced the development and potential of Large Vision Language Models (LVLMs) as key enablers for generalist robots, demonstrating promising generalization across a variety of scenes [15, 25, 32, 41, 64, 84]. One of the notable examples is OpenVLA [34], which integrates dual visual encoders with a pre-trained large language model to enable observation-to-action capabilities through instruction-following [10, 33, 45, 46, 70, 82]. By aligning visual and textual semantics, OpenVLA facilitates complex scenario understanding and end-to-end action generation, showing impressive generalization when executing previously unseen task instructions." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.547, + 0.787, + 0.562 + ], + "angle": 0, + "content": "2.2. Adversarial Attacks in Robotic" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.569, + 0.909, + 0.78 + ], + "angle": 0, + "content": "Adversarial attacks serve as critical tools for assessing model vulnerabilities, particularly in robotics, where models operate in dynamic, real-world settings. Traditional gradient-based, pixel-level attacks [17, 21, 43, 49, 61, 71] exploit model gradients to calculate malicious perturbations, achieving high attack success rates in digital environments. For physical-world attacks, patch-based methods [3, 8, 59, 76, 77] offer a practical alternative for real-world applications by introducing physically realizable perturbations that retain efficacy under varied conditions, such as angle or lighting changes, making them highly applicable for robotic systems. Therefore, considering VLA-based models enable robot execution in real-world scenarios, this work employs a patch-based attack strategy." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.781, + 0.909, + 0.901 + ], + "angle": 0, + "content": "VLA models [13, 26, 64] couple visual and linguistic modalities for perception-action alignment. The continuous, high-dimensional nature of visual inputs [49, 58, 62] makes them the most likely target for adversarial perturbations, as attackers can subtly alter visual data in ways that are difficult to detect [2, 9, 65]. Consequently, this work applies attacks on the visual modality to achieve robust manipulations across diverse environments. To the best of our knowledge, this" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6949" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.113, + 0.093, + 0.885, + 0.312 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.315, + 0.907, + 0.385 + ], + "angle": 0, + "content": "Figure 2. Overall Adversarial Framework. The robot captures an input image, processes it through a vision-language model to generate tokens representing actions, and then uses an action de-tokenizer for discrete bin prediction. The model is optimized with adversarial objectives focusing on various discrepancies and geometries (i.e., UADA, UPA, TMA). Forward propagation is shown in black, and backpropagation is highlighted in pink. These objectives aim to maximize errors and minimize task performance, with visual emphasis on 3D-space manipulation and a focus on generating adversarial perturbation \\(\\delta\\) during task execution, such as picking up a can." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.391, + 0.484, + 0.466 + ], + "angle": 0, + "content": "study is one of the earliest attempts to formally define adversarial objectives for VLA models, targeting gaps in adversarial robustness through the model's geometric constraints (e.g., spatial dependencies in visual input) and architectural nuances (e.g., cross-modal information integration)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.477, + 0.226, + 0.493 + ], + "angle": 0, + "content": "3. Methodology" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.503, + 0.484, + 0.655 + ], + "angle": 0, + "content": "In this section, we first review the core algorithmic principles of VLA in §3.1, which serves as the foundation for the adversarial scheme designs. We then present two types of untargeted adversarial attacks, including Untargeted Action Discrepancy Attack (UADA) in §3.2 and Untargeted Position-aware Attack (UPA) in §3.3. In §3.4, we further include Targeted Manipulation Attack (TMA), designed to direct the robot toward specific erroneous execution. Finally, we introduce the Normalized Action Discrepancy (NAD) metric in §3.5. The overall framework is shown in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.667, + 0.221, + 0.683 + ], + "angle": 0, + "content": "3.1. Preliminary" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.689, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Vision-Language-Action models [34, 84] (see Fig. 2) are built on large language models integrated with visual encoders, enabling robots to interpret human instructions and process visual input from a camera to perform context-aware actions. VLA models [34, 42, 84] generally abstract continuous action predictions into a classification problem by discretizing robot actions. Specifically, they first discretize the continuous probabilities into class labels \\( y = \\arg \\max_{x} \\mathcal{F}(x) \\) where \\( \\mathcal{F}(\\cdot) \\) is the VLA model. An action de-tokenizer then generates actions \\( A = DT(y) \\). By categorizing action values into discrete class labels, the model converts continuous probability outputs into discrete signals, this simplification facilitates quicker convergence and faster training times, and is commonly used for VLA-based models [34, 42, 84]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.391, + 0.907, + 0.452 + ], + "angle": 0, + "content": "In this work, we focus on a 7 degree-of-freedoms (DoFs) robotic arm [23]. At each step, an action consists of 7 DoFs with specific physical significance in three-dimensional Cartesian space, represented by:" + }, + { + "type": "equation", + "bbox": [ + 0.552, + 0.462, + 0.905, + 0.477 + ], + "angle": 0, + "content": "\\[\nA = \\left[ \\Delta P _ {x}, \\Delta P _ {y}, \\Delta P _ {z}, \\Delta R _ {x}, \\Delta R _ {y}, \\Delta R _ {z}, \\text {g r i p p e r} \\right], \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.487, + 0.907, + 0.638 + ], + "angle": 0, + "content": "where \\(\\Delta P_{x,y,z}\\) and \\(\\Delta R_{x,y,z}\\) denote relative positional and rotational changes along the x-, y-, and z-axis, discretized into 256 uniform bins across each DoFs' action bounds \\([y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i]\\) [34]. The gripper state is binary, indicating its open or closed state. This control design presents a unique challenge for adversarial attacks, as finely divided bins result in minimal action discrepancies between neighboring bins (e.g., \\(\\pm 0.007/\\mathrm{bin}\\)). This means that even if an attack shifts the classification to an adjacent bin, the resulting action discrepancy has minimal impact on real-world performance." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.648, + 0.848, + 0.664 + ], + "angle": 0, + "content": "3.2. Untargeted Action Discrepancy Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.67, + 0.909, + 0.854 + ], + "angle": 0, + "content": "To exacerbate action discrepancies, we introduce the Untargeted Action Discrepancy Attack (UADA), which aims to maximize deviations in robot actions. This attack is based on the observation that larger robot actions usually correlate with intense physical movements, which, in turn, may amplify the potential for real-world hazards [28-30]. For UADA, the attack target is one or a combination of 7 DoFs, defined as \\( y_{gt}^{I} = \\{y_{gt}^{i}|i\\in [1,\\dots ,7]\\} \\). To define UADA's objective, we first identify the most distant action \\( y_{adv}^{i} \\), which maximizes the discrepancy from the \\( i \\)-th DoF ground truth action \\( y^{i} \\). This action, for each DoF, corresponds to its own action bound \\( [y_{\\min}^{i},y_{\\max}^{i}] \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.534, + 0.864, + 0.907, + 0.906 + ], + "angle": 0, + "content": "\\[\ny _ {a d v} ^ {i} = \\left\\{ \\begin{array}{l l} y _ {\\max } ^ {i}, & \\text {i f} | y _ {\\max } ^ {i} - y _ {g t} ^ {i} | \\geq | y _ {\\min } ^ {i} - y _ {g t} ^ {i} | \\\\ y _ {\\min } ^ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6950" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.091, + 0.485, + 0.228 + ], + "angle": 0, + "content": "Instead of directly using \\( y_{adv}^{i} \\) as the misclassification target, we introduce a soft attack objective to capture the discrepancy between actions, ensuring smooth gradient optimization and stable attack performance. Specifically, due to the physical action magnitude information contains in bin labels, we reweight the output probability \\( \\mathcal{F}(x)_{bins}^{i} \\in \\mathcal{R}^{1 \\times bins} \\) using normalized bin labels \\( y_{bins}^{i} = [\\frac{1}{bins}, \\dots, 1] \\), where \\( \\lfloor y_{bins}^{i} \\rfloor \\) and \\( \\lceil y_{bins}^{i} \\rceil \\) correspond to the normalized values of \\( y_{\\min}^{i} \\) and \\( y_{\\max}^{i} \\), respectively. The reweighting process is defined as:" + }, + { + "type": "equation", + "bbox": [ + 0.16, + 0.24, + 0.483, + 0.281 + ], + "angle": 0, + "content": "\\[\ny _ {s o f t} ^ {i} = \\sum_ {b i n s = 1} ^ {n} \\mathcal {F} (x + \\delta) _ {b i n s} ^ {i} \\otimes y _ {b i n s} ^ {i}, \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.29, + 0.484, + 0.351 + ], + "angle": 0, + "content": "where \\(\\otimes\\) denotes Hadamard Product, \\(\\delta\\) is the adversarial patch and \\(y_{\\text{soft}}^i\\) represents the \\(i\\)-th DoF's soft action. Finally, the objective of UADA is to minimize the discrepancy between the soft and the most distant actions, which is:" + }, + { + "type": "equation", + "bbox": [ + 0.154, + 0.362, + 0.483, + 0.403 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U A D A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\sum_ {i} ^ {I} \\left(y _ {s o f t} ^ {i} - y _ {a d v} ^ {i}\\right) ^ {2}, \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.485, + 0.46 + ], + "angle": 0, + "content": "UADA defines the objective function \\(\\mathcal{L}_{\\mathrm{UADA}}\\) tailored for attacking the robot's action space, allowing adversarial patches to create significant, lasting disruptions in task performance." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.469, + 0.389, + 0.485 + ], + "angle": 0, + "content": "3.3. Untargeted Position-aware Attack" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.492, + 0.484, + 0.596 + ], + "angle": 0, + "content": "We explore untargeted adversarial attacks in parallel by considering the positional dynamics within VLA models. In typical task execution, precise and directed movements of the end-effector towards designated goals are essential, with actions mapped in three-dimensional space. The positional DoFs, \\( A^p = [\\Delta P_x, \\Delta P_y, \\Delta P_z] \\) encapsulates the directional movements within 3D for effective robotic control." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.598, + 0.484, + 0.734 + ], + "angle": 0, + "content": "Recognizing the importance of \\( A^p = DT(y^p) \\) in controlling the end-effector's path, we introduce a position-aware attack to disrupt the intended movement trajectory. This type of adversarial vulnerability remains largely unexplored but has significant implications for task failure: the attack objective can steer the end-effector away from its intended path by introducing directional perturbations, amplifying errors and causing substantial trajectory distortions. Formally, we define the Untargeted Position-aware Attack (UPA) objective as:" + }, + { + "type": "equation", + "bbox": [ + 0.11, + 0.755, + 0.483, + 0.787 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U P A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ \\alpha \\frac {y _ {a d v} ^ {p} \\cdot y ^ {p}}{\\| y _ {a d v} ^ {p} \\| \\| y ^ {p} \\|} + \\beta \\frac {1}{\\| y _ {a d v} ^ {p} - y ^ {p} \\| _ {2}} \\right], \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.795, + 0.484, + 0.901 + ], + "angle": 0, + "content": "where \\(\\alpha\\) and \\(\\beta\\) are the hyperparameters that balance between the directionality and magnitude of the perturbations. By integrating geometric awareness into our attack, this approach generates perturbations that induce cumulative deviations in the robot's trajectory, even with minimal adjustments. \\(L_{2}\\)-normalization \\(\\| \\cdot \\| _2\\) further intensifies these deviations, rendering the attack highly effective and disruptive." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.091, + 0.785, + 0.107 + ], + "angle": 0, + "content": "3.4. Targeted Manipulation Attack" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.113, + 0.907, + 0.219 + ], + "angle": 0, + "content": "In addition to the aforementioned untargeted attacks, we also explore the adversarial vulnerabilities of VLA models with a targeted strategy. This attack aims to directly mislead the models into predicting specific trajectories, resulting in precise alterations. We design this attack since the targeted alterations can drive malicious behaviors or induce task failure. The objective of the targeted manipulation attack is:" + }, + { + "type": "equation", + "bbox": [ + 0.582, + 0.224, + 0.907, + 0.243 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {T M A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {T} ^ {I}\\right) \\right], \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.248, + 0.909, + 0.432 + ], + "angle": 0, + "content": "where \\( y_{T}^{I} = \\{y_{T}^{i} = t|i\\in [1,\\dots ,7],t\\in [y_{\\min}^{i},y_{\\max}^{i}] \\} \\) represents the adversarial target(s); and \\( CE(\\cdot ,\\cdot) \\) denotes the Cross-Entropy loss, which measures the error between the predicted state under adversarial perturbation and the desired target state. By steering the robot toward an adversarial target across successive time steps, our approach manipulates the trajectory and undermines task performance, ultimately leading to a substantial disruption of intended operations. Furthermore, manipulating any individual DoF or a combination thereof can also severely compromise task success, given that the errors in one dimension tend to propagate and magnify over time during execution (see Fig. 3)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.439, + 0.859, + 0.456 + ], + "angle": 0, + "content": "3.5. Metric: Normalized Action Discrepancy" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.461, + 0.909, + 0.63 + ], + "angle": 0, + "content": "To assess the potential physical impact of adversarial patches, we introduce Normalized Action Discrepancy (NAD) as a fine-grained metric for evaluating discrepancies at the action level. NAD serves as a complementary measurement alongside the coarse-grained Failure Rate (FR) at the task level, providing a detailed analysis of deviations during action execution. To define the NAD, we first determine the maximum action discrepancy \\( d_{\\mathrm{max}}^i \\), which represents the largest deviation that can occur within the allowed range for \\( i \\)-th DoF. This is computed by measuring the L1 distance between the ground truth \\( y_{gt}^i \\) and its action bound \\( [y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i] \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.527, + 0.637, + 0.907, + 0.656 + ], + "angle": 0, + "content": "\\[\nd _ {\\mathrm {m a x}} ^ {i} (y) = \\max \\left[ D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m i n}} ^ {i} | \\big), D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m a x}} ^ {i} | \\big) \\right]. (7)\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.661, + 0.908, + 0.727 + ], + "angle": 0, + "content": "We next calculate the applied action discrepancy \\( d_{\\mathrm{applied}}^i (x,y) \\) \\( = |DT(\\mathcal{F}(x)^i) - DT(y_{gt}^i)| \\) to measure the deviation between the model's output and ground truth. We define the NAD as the drifting ratio between \\( d_{\\mathrm{applied}}^i \\) and \\( d_{\\mathrm{max}}^i \\) as:" + }, + { + "type": "equation", + "bbox": [ + 0.631, + 0.735, + 0.907, + 0.775 + ], + "angle": 0, + "content": "\\[\n\\mathrm {N A D} = \\frac {1}{\\mathrm {I}} \\sum_ {\\mathrm {i}} ^ {\\mathrm {I}} \\frac {\\mathrm {d} _ {\\text {a p p l i e d}} ^ {\\mathrm {i}}}{\\mathrm {d} _ {\\max } ^ {\\mathrm {i}}}, \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.785, + 0.909, + 0.847 + ], + "angle": 0, + "content": "where \\( I \\in [1, \\dots, 7] \\) represents a single or a combination of DoF(s). In this way, NAD provides normalized measurement of action deviations, allowing for a consistent evaluation of adversarial impact across varying DoF(s)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.852, + 0.645, + 0.868 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.871, + 0.908, + 0.901 + ], + "angle": 0, + "content": "In this section, we first detail the implementation of our adversarial framework and baseline methods in §4.1. Next, we" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "6951" + } + ], + [ + { + "type": "image_caption", + "bbox": [ + 0.168, + 0.094, + 0.826, + 0.109 + ], + "angle": 0, + "content": "TASK (UADA) : Pick up the black bowl between the plate and the ramekin and place it on the plate" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.114, + 0.591, + 0.244 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.602, + 0.113, + 0.812, + 0.245 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.243, + 0.248, + 0.755, + 0.264 + ], + "angle": 0, + "content": "TASK (UPA) : Pick up the black bowl on the ramekin and place it on the plate" + }, + { + "type": "image", + "bbox": [ + 0.16, + 0.266, + 0.591, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.606, + 0.267, + 0.811, + 0.398 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.285, + 0.403, + 0.713, + 0.418 + ], + "angle": 0, + "content": "TASK (TMA) : Pick up the alphabet soup and place it in the bask" + }, + { + "type": "image", + "bbox": [ + 0.162, + 0.42, + 0.591, + 0.552 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.609, + 0.421, + 0.813, + 0.555 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.56, + 0.908, + 0.616 + ], + "angle": 0, + "content": "Figure 3. Qualitative Results of adversarial vulnerabilities over OpenVLA-7B [34] and OpenVLA-7B-LIBERO [34] with objectives of UADA, UPA, and TMA, respectively. We visualize the overall 3D trajectories and 2D trajectories of benign \\(\\bullet\\) and adversarial \\(\\bullet\\) scenarios at each time step to compare the impact of the generated adversarial patch in affecting them. The untargeted trajectory \\(\\bullet\\) is visualized in UADA task. All trajectories start with \\(\\triangle\\), and we plot the success end point, marked as \\(\\star\\)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.624, + 0.486, + 0.73 + ], + "angle": 0, + "content": "outline the experimental setup in \\(\\S 4.2\\) and present quantitative and qualitative results in both simulation and physicalworld scenarios (§4.3). We then conduct diagnostic experiments (§4.4) to analyze the impact of key components. Additionally, we assess the robustness of our method against various defense mechanisms in \\(\\S 4.5\\) and conclude with a comprehensive systemic discussion in \\(\\S 4.6\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.739, + 0.308, + 0.755 + ], + "angle": 0, + "content": "4.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.485, + 0.903 + ], + "angle": 0, + "content": "The implementation of our adversarial patch attack pipeline is detailed in Algorithm 1. We incorporate two key modifications to enhance the training stability of the generated patch. Inner-loops. Following previous work [49], we add \\( k \\) inner-loops optimize steps during each iteration in line 4, aiming to reduce data variance and ensure more consistent updates. Geometric Transformations. To improve the robustness of our attack under real-world scenarios, we employ random geometric transformations \\( \\mathcal{T}[(\\cdot), (sh_x, sh_y, \\theta)] \\) in line 6 with" + }, + { + "type": "title", + "bbox": [ + 0.525, + 0.627, + 0.788, + 0.641 + ], + "angle": 0, + "content": "Algorithm 1: Adversarial Patch Attack." + }, + { + "type": "table", + "bbox": [ + 0.544, + 0.645, + 0.905, + 0.826 + ], + "angle": 0, + "content": "
1:Input: X: dataset; δ: patch; L0: attack objective; \nF: VLA model; T(·): transformations; φ,ψ: \ntransformation parameters; T: attack steps; k: \ninner-loop steps.
2:Initialize: δ ~ U[0,1), L0 ∈ {LUADA, LUPA, LTMA}
3:for i = 1,2,···, T do
4:for i = 1,2,···, k do
5:shx, shy ~ U(-φ,φ), θ ~ U(-ψ,ψ)
6:ypred ← F[T((x + δ), (shx, shy, θ))]
7:δ ← ∂L0/∂δ{▷ Update}
8:end for
9:end for
" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.856, + 0.791, + 0.87 + ], + "angle": 0, + "content": "affine and rotation transformations [3] as:" + }, + { + "type": "equation", + "bbox": [ + 0.524, + 0.874, + 0.905, + 0.906 + ], + "angle": 0, + "content": "\\[\n\\mathcal {T} \\left[ (\\cdot), \\left(s h _ {x}, s h _ {y}, \\theta\\right) \\right] = \\left[ \\begin{array}{c c} 1 & s h _ {y} \\\\ s h _ {x} & 1 \\end{array} \\right] \\cdot \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right], \\tag {9}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6952" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.23 + ], + "angle": 0, + "content": "where \\(sh_x, sh_y \\sim \\mathcal{U}(-\\phi, \\phi)\\) are shear factors, and \\(\\theta \\sim \\mathcal{U}(-\\psi, \\psi)\\) denotes the rotation angle, both sampled from uniform distributions during inner loop optimization step \\(k\\). Untargeted Manipulation Attack (UMA) Baseline. As a pioneering work, we could not find a directly comparable baseline for our attacks. Therefore, we adapt prior work in adversarial learning as one of our baseline methods [66]. This attack is designed to mislead classifiers; we integrate it into our framework, leading to the following objective:" + }, + { + "type": "equation", + "bbox": [ + 0.152, + 0.238, + 0.483, + 0.258 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {U M A}} = - \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {g t} ^ {I}\\right) \\right]. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.267, + 0.483, + 0.297 + ], + "angle": 0, + "content": "By minimizing the negative CE loss, this attack pushes the predicted action away from the ground truth action." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.298, + 0.484, + 0.388 + ], + "angle": 0, + "content": "Random Noise Baseline. We generate random noise patches as an additional baseline, representing an unstructured intervention without any learned adversarial intent. Specifically, the noise is sampled from a Gaussian distribution \\(\\mathcal{N}(0,1)\\) and evaluated with the same setting as proposed attacks." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.397, + 0.268, + 0.413 + ], + "angle": 0, + "content": "4.2. Experiment Setup" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.417, + 0.484, + 0.581 + ], + "angle": 0, + "content": "Datasets. We conduct attacks on BridgeData V2 [69] and LIBERO [44] with corresponding VLAs. BridgeData V2 [69] is a real-world dataset comprising 24 diverse environments and 13 distinct skills, such as grasping, placing, and object rearrangement, with a total of 60,096 trajectories. LIBERO [44] is a simulation dataset designed to evaluate models across four distinct task types (Spatial, Object, Goal, Long). Notably, LIBERO-Long combines diverse objects, layouts, and long-horizon tasks, making them challenging for complex, multi-step planning, thus we choose LIBERO-Long to conduct UADA, UPA and TMA." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.584, + 0.484, + 0.854 + ], + "angle": 0, + "content": "Victim VLAs. In our study, we select publicly available and current most influential VLAs as victim models for evaluation. Specifically, we focus on four variants of the OpenVLA model, each trained independently on different task suites within the LIBERO dataset [44] (i.e., Spatial, Object, Goal, Long). To evaluate the effectiveness of our methods, we craft adversarial patches using three distinct generating setups: Simulation Setting involves a model trained in a simulated environment using the LIBERO Long suite with the openvla-7B-libero-long variant [34]. Physical Setting uses a model trained on physical world data from the BridgeData v2 [69] with the openvla-7B model [34]. This approach allows us to assess our methods on both simulated and real-world data. Subsequently, we evaluate the performance of generated adversarial patches on victim models (i.e., OpenVLA LIBERO variants) trained on different tasks suites to rigorously prove the robustness and effectiveness of our method. These models are trained with distinct data sources and task objectives." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Robot Setups (Real-world Setting). For the real-world tasks, we adopt a robotic system consisting of a Universal Robots UR10e equipped with a Robotiq Hand-E Gripper to" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.121 + ], + "angle": 0, + "content": "provide a 7DoF motion. The sensing system includes one RGB webcam in the fixed position with a shoulder view." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.122, + 0.907, + 0.227 + ], + "angle": 0, + "content": "Evaluation Details. To assess the effectiveness and robustness of our methods, we conduct experiments on the LIBERO dataset [44]. Each suite consists of 10 tasks, with each task executed for 50 trials, resulting in a total of 500 rollouts, following Kim et al. [34]. For reproducibility, we carefully select patch paste locations for each suite, ensuring they do not obscure objects or robots in the test environment." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.228, + 0.909, + 0.438 + ], + "angle": 0, + "content": "Evaluation Metric. Regarding the task execution evaluation, we take the maximum steps of each task suite in the LIBERO training dataset as the timeout failure condition to reduce computational overhead. Furthermore, building on the concept of Success Rate (SR) introduced in LIBERO [44], we adopt Failure Rate (FR), defined as \\(1 - \\mathrm{SR}\\), as the primary evaluation metric. To further quantify action discrepancy, we employ NAD (see Eq. 8) for untargeted attacks on corresponding DoF(s), as it considers relative distances to quantify deviation from the optimal action. A higher NAD value indicates a greater discrepancy between the ground truth and the predicted action. Additionally, we report the standard deviation of FR across tasks within each task suite." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.444, + 0.645, + 0.457 + ], + "angle": 0, + "content": "4.3. Main Result" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.464, + 0.909, + 0.854 + ], + "angle": 0, + "content": "Quantitative Results. For UADA and UPA, our methods effectively amplify action discrepancies, leading to a notable transfer attack ability in increasing failure rates (see Tab. 1a). Specifically, while attacking \\(\\mathrm{DoF}_1\\) and \\(\\mathrm{DoF}_{1\\sim 3}\\) in the Simulation setup, UADA and UPA achieve NAD of \\(21.0\\%\\) and \\(14.5\\%\\), significantly outperforming UMA scenarios with increments of \\(6.9\\%\\) and \\(3.1\\%\\), respectively. Both UADA and UPA effectively disrupt robot execution, yielding maximum average failure rates of \\(100\\%\\) and \\(89.7\\%\\), respectively. Regarding the Physical setup (see Tab. 1b), our attack methods demonstrate strong transferability, as adversarial patches generated in the physical setting significantly impact the performance of LIBERO-X test. Specifically, both UADA and UPA generate malicious actions with large action discrepancies, which are \\(32.6\\%\\) (\\(\\mathrm{DoF}_1\\)) and \\(26.9\\%\\) (\\(\\mathrm{DoF}_{1\\sim 3}\\)) with an increase of \\(5.9\\%\\) and \\(3.1\\%\\) compared to \\(26.7\\%\\) (\\(\\mathrm{DoF}_1\\)) and \\(23.8\\%\\) (\\(\\mathrm{DoF}_{1\\sim 3}\\)) in UMA scenarios, respectively. Compare the two settings (Tab. 1a v.s. Tab. 1b), we observe a general variance in NAD (e.g. \\(32.6\\%\\) v.s. \\(21.0\\%\\) for UADA). This discrepancy can be attributed to the fundamental differences between the simulation and real-world datasets used for conducting attacks (Bridge V2 [69] v.s. LIBERO [44]). The increased variability in real-world data, including environmental complexity, object diversity, and task difficulty, allows the robot more opportunities to generate larger action discrepancies within the validation dataset." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.908, + 0.901 + ], + "angle": 0, + "content": "For TMA task, we evaluate the effectiveness of our method by manipulating all DoF(s) to 0 (i.e., \\( y_{T}^{i} = 0 \\) in Eq. 6). Our method demonstrates significant effectiveness" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6953" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.908, + 0.131 + ], + "angle": 0, + "content": "Table 1. Untargeted Results. We report FR and NAD in LIBERO simulation. * denotes the in-domain victim model and dataset aligned with the patch generation model and dataset, △ denotes a transfer attack evaluation involving a distinct victim model and dataset (see §4.2). The FR (↑) is highlighted in best and second best for each task suite." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.134, + 0.495, + 0.24 + ], + "angle": 0, + "content": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long*Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF114.1100±0.0%99.0±3.0%100±0.0%100±0.0%99.8%
DoF1~311.4100±0.0%98.8±2.6%99.0±2.4%100±0.0%99.5%
UADADoF121.0100±0.0%99.2±2.4%100±0.0%100±0.0%99.8%
DoF1~317.9100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~314.596.2±11.4%77.8±12.5%88.0±10.4%96.8±3.0%89.7%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.087, + 0.242, + 0.493, + 0.267 + ], + "angle": 0, + "content": "(a) Simulation Setup. Adversarial patch generated on OpenVLA-LIBERO-10 victim model with LIBERO-Long dataset and evaluated on LIBERO-X." + }, + { + "type": "table", + "bbox": [ + 0.5, + 0.134, + 0.903, + 0.24 + ], + "angle": 0, + "content": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long△Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF126.789.6±15.3%60.8±16.8%64.8±26.8%80.0±20.8%73.8%
DoF1~323.896.6±6.2%56.4±20.4%80.0±20.5%82.0±17.9%78.8%
UADADoF132.699.2±1.8%98.8±2.4%92.6±15.2%96.6±4.8%96.8%
DoF1~327.7100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~326.995.6±9.5%57.0±23.6%57.2±24.4%72.8±19.7%70.7%
" + }, + { + "type": "table_footnote", + "bbox": [ + 0.495, + 0.242, + 0.9, + 0.267 + ], + "angle": 0, + "content": "(b) Physical Setup. Adversarial patch generated on OpenVLA-7b victim model with BV2 dataset and evaluated on LIBERO-X." + }, + { + "type": "table_caption", + "bbox": [ + 0.089, + 0.273, + 0.908, + 0.303 + ], + "angle": 0, + "content": "Table 2. Targeted Manipulation Attack Results. Failure Rate (FR, ↑) and its standard deviation across tasks within LIBERO [44] suite are reported. The performance of task suits and DoFs is averaged separately, with the best and second-best values separately." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.305, + 0.905, + 0.466 + ], + "angle": 0, + "content": "
Targeted DoF(s)MetricSimulation (Attacked on OpenVLA-LIBERO-10)Physical (Attacked on OpenVLA-7b)
Spatial△Object△Goal△Long*AvgSpatial△Object△Goal△Long△Avg
BenignFR15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
DoF1FR100±0.0%97.8±3.0%100±0.0%100±0.0%99.5%90.6±19.9%36.6±18.1%61.0±21.7%86.4±12.7%68.7%
DoF2100±0.0%99.2±1.3%100±0.0%100±0.0%99.8%99.6±0.8%69.2±11.1%76.2±26.1%91.8±9.9%84.2%
DoF3100±0.0%97.2±6.0%100±0.0%100±0.0%99.3%72.0±14.8%30.2±17.4%39.2±20.9%62.4±17.5%51.0%
DoF494.0±11.0%37.4±23.2%51.2±26.5%87.6±14.9%67.6%93.6±9.7%31.8±20.3%45.0±18.3%63.8±17.0%58.6%
DoF5100±0.0%92.6±8.4%100±0.0%100±0.0%98.2%97.8±4.9%36.0±22.5%62.0±24.8%79.4±21.3%68.8%
DoF6100±0.0%80.0±9.5%92.6±13.4%100±0.0%93.2%79.0±14.2%22.2±14.4%41.0±21.5%62.6±15.7%51.2%
DoF799.2±1.3%86.6±8.3%67.2±22.5%96.4±6.4%87.4%91.8±15.3%98.8±1.3%70.6±23.4%93.2±10.2%88.6%
DoF1~3100±0.0%100±0.0%100±0.0%100±0.0%100%96.4±8.0%83.6±16.8%74.4±26.3%91.8±11.9%86.6%
Task Avg99.2%86.4%88.9%98.0%-90.1%51.1%58.7%78.9%-
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.473, + 0.486, + 0.669 + ], + "angle": 0, + "content": "in manipulating robotic trajectory and increasing FR (see Tab. 2). Specifically, when applied to the different generation settings, our approach yields notable increments across various tasks, including a max average failure rate \\(100\\%\\) v.s. \\(23.5\\%\\) of benign performance in Simulation, and \\(88.6\\%\\) in Physical, respectively. Moreover, during attack \\(\\mathrm{DoF}_4\\), we observe low FR and high task deviation. This failure can be attributed to the fact that \\(\\mathrm{DoF}_4\\) controls the orientation along the x-axis, which can be redundant DoF in tasks. As a result, attacks targeting redundant DoF(s) [38] are less likely to disrupt execution effectively. This observation underscores the importance of task-specific considerations when designing adversarial attacks on robotic systems." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.675, + 0.487, + 0.902 + ], + "angle": 0, + "content": "Qualitative Results. We qualitatively analyze robot movement trajectories under the three proposed attacks in Fig. 3. For UADA, we compare the trajectory deviations in the same trail with patches generated from UADA and UMA attacks. As seen, the UMA induces small deviations in the trajectory. UADA, on the other hand, produces significantly larger trajectory deviations (also supported by NAD metric in Tab. 1b), which is attributed to UADA's capability of incorporating action discrepancies. In our observation, UADA significantly amplifies the effect on overall task execution, thereby increasing the potential for robotic hazards. For UPA, we observe chaotic and irregular behaviors, including instances where the end-effector moves out of the camera's field of view. We attribute this phenomenon to the efficacy of the adversarial patch in perturbing the model's spatial perception," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.473, + 0.907, + 0.563 + ], + "angle": 0, + "content": "inducing a consistent deviation from the intended direction of movement, ultimately resulting in failure in a long run. For TMA, we observe a notable reduction in the range of motion along the x-axis corresponding to the targeted attack axis. This suggests that our proposed targeted attack can effectively constrained robot movements." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.564, + 0.909, + 0.655 + ], + "angle": 0, + "content": "In summary, the qualitative analysis shows that both Untarget attack, UADA, UPA, and TMA can effectively disrupt the robot actions generated from VLA models. These findings underscore a pressing security concern during the deployment of generalist robots, especially when considering application scenes that require reliable operations [11, 68]." + }, + { + "type": "image", + "bbox": [ + 0.539, + 0.659, + 0.88, + 0.78 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.782, + 0.908, + 0.81 + ], + "angle": 0, + "content": "Figure 4. Qualitative Results of the physical world. The first/second row show benign and adversarial cases respectively." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Real-world Performance. In addition to the digital simulation results, we also conduct a comprehensive evaluation of the performance of our generated adversarial patches in real-world scenarios. The evaluation encompassed 100 trials across three distinct tasks: object grasping, placement, and manipulation. As shown in Fig. 4, the adversarial patch" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6954" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.115, + 0.09, + 0.203, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.132, + 0.195, + 0.203, + 0.207 + ], + "angle": 0, + "content": "(a) Inner-loop" + }, + { + "type": "image", + "bbox": [ + 0.203, + 0.09, + 0.352, + 0.191 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.242, + 0.195, + 0.313, + 0.206 + ], + "angle": 0, + "content": "(b) Patch Size" + }, + { + "type": "image", + "bbox": [ + 0.353, + 0.09, + 0.49, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.4, + 0.195, + 0.446, + 0.206 + ], + "angle": 0, + "content": "(c) JPEG" + }, + { + "type": "image", + "bbox": [ + 0.49, + 0.09, + 0.617, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.534, + 0.195, + 0.581, + 0.206 + ], + "angle": 0, + "content": "(d) Noise" + }, + { + "type": "image", + "bbox": [ + 0.617, + 0.09, + 0.745, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.646, + 0.195, + 0.726, + 0.206 + ], + "angle": 0, + "content": "(e) Median Blur" + }, + { + "type": "image", + "bbox": [ + 0.746, + 0.09, + 0.877, + 0.192 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.782, + 0.195, + 0.849, + 0.206 + ], + "angle": 0, + "content": "(f) Bit-Depth" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.208, + 0.907, + 0.249 + ], + "angle": 0, + "content": "Figure 5. Impact of Inner-loop, Patch Size and Defense Discussion. The figure shows how varying Inner-loop affects NAD in UADA, and patch sizes affect \\( L1 \\) distance and the failure rates in TMA, both targeting at DoF\\(_1\\). (a) Impact of Inner-loop, (b) Impact of Patch Size and (c-f) the effect of four different defenses on failure rates." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.253, + 0.486, + 0.435 + ], + "angle": 0, + "content": "generated with UADA demonstrated the ability to manipulate the robot effectively, achieving an attack success rate exceeding \\(43\\%\\). Although this success rate is lower than the corresponding digital-world performance (i.e., \\(100\\%\\)), it highlights the effectiveness of our patches in physical-world applications as well without the need for further adaptations. Crucially, we observed that the induced erratic movements (similar to simulation scenario) of the robot during successful attacks pose significant risks to human safety and surrounding environments. This finding emphasizes the severe threat posed by VLA models in real-world applications, consistent with similar observations in digital settings." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.44, + 0.304, + 0.456 + ], + "angle": 0, + "content": "4.4. Diagnostic Experiment" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.579 + ], + "angle": 0, + "content": "Impact of Inner-loops. We discuss the impact of inner-loop steps to model performance in Fig. 5(a). The results show that NAD first improves when inner-loop steps continue to increase. This is due to the reduced data variance and gradually stabilized gradients. However, further increasing the steps results in a slower performance increase, with a noticeably longer training schedule. We thus choose inner-loop steps=50 as it balances between performance and scale." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.58, + 0.484, + 0.732 + ], + "angle": 0, + "content": "Impact of Patch Size. We study TMA's performance with different patch sizes and report the average FR across four LIBERO [44] tasks in Fig. 5(b). The patch sizes examined are \\([3\\%, 5\\%, 7\\%, 10\\%]\\) of input image. Our findings indicate an inversely proportional relationship between the \\(L_{1}\\) distance of predicted actions and the target action as the patch size increased. This increase in patch size also correlated with a marked rise in FR. The observed trend suggests larger patches give adversaries more optimization space to influence the model, aligning with prior work [12, 81]." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.736, + 0.303, + 0.751 + ], + "angle": 0, + "content": "4.5. Robustness Evaluation" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.756, + 0.485, + 0.847 + ], + "angle": 0, + "content": "We further examine whether concurrent defense strategies [18, 78, 80] can resist our adversarial examples within VLA. Specifically, we applied four prior defense techniques and reported their FR for both benign and adversarial samples in Fig. 5(c-f), respectively. The results show that our adversarial attack bypasses most defense strategies." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.851, + 0.281, + 0.867 + ], + "angle": 0, + "content": "4.6. Systemic Discussion" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Patch Pattern Analysis. In Fig. 6, we present the adversarial patches generated for a range of targets. A particularly" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.254, + 0.578, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.585, + 0.255, + 0.641, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.649, + 0.255, + 0.705, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.712, + 0.255, + 0.768, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.775, + 0.255, + 0.832, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.839, + 0.255, + 0.895, + 0.298 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.551, + 0.303, + 0.676, + 0.316 + ], + "angle": 0, + "content": "(a) BridgeData V2 [69]" + }, + { + "type": "image_caption", + "bbox": [ + 0.757, + 0.303, + 0.849, + 0.316 + ], + "angle": 0, + "content": "(b) LIBERO [44]" + }, + { + "type": "image_caption", + "bbox": [ + 0.55, + 0.318, + 0.868, + 0.332 + ], + "angle": 0, + "content": "Figure 6. Patch Visualization. Semantic-rich patches." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.339, + 0.907, + 0.565 + ], + "angle": 0, + "content": "noteworthy observation is that some of these patches bear a striking resemblance to the structural joints of a robotic arm. Given the strong similarity between these patches and robotic arms in appearance, we hypothesize that these VLA-based models, in their current training paradigms, are often limited to a narrow range of robotic systems operating within restricted camera views. This constrained perspective camera pose may inadvertently induce a learned representation bias, where adversarial perturbations align with prominent structural features of the training robot's appearance in order to deceive current victim models. Given that these patches are model-specific and reflect the spatial and visual patterns of the training environment, we have a reasonable concern that they may compromise VLA-based models' generalizability and robustness in physical-world deployments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.566, + 0.909, + 0.732 + ], + "angle": 0, + "content": "Possible Defense Strategies. Our work highlights the pressing need for a paradigm shift in the VLA-based models' training strategies in order to successfully defend our attacks. Specifically, future approaches might incorporate more training scenarios that involve complex multi-robot interactions to mitigate the current patch bias. Another possible solution is to leverage the logical relationships within the robot arm's physical structure to estimate the actual position of each joint. By predicting joint positions deviated from physical constraints, we can eliminate unreasonable adversarial samples, even if they are similar to actual joint(s)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.742, + 0.634, + 0.757 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.907, + 0.901 + ], + "angle": 0, + "content": "While VLA models have gained significant popularity due to their substantial robot capabilities, this study pioneers specific robotic attack objectives and demonstrates that these models are in fact vulnerable to various types of attacks. Experimental results demonstrate that our attacks expose significant vulnerabilities to VLA models, raising concerns regarding impetuous real-world deployments. We believe our framework provides pioneering and foundational contributions to the reliability of AI-enhanced robotic systems." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "6955" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.259, + 0.108 + ], + "angle": 0, + "content": "Acknowledgements" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.116, + 0.486, + 0.268 + ], + "angle": 0, + "content": "This research was supported by the National Science Foundation under Grant No. 2450068 and No. 2242243, as well as start-up resources provided by Rutgers University. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of U.S. Naval Research Laboratory (NRL) or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.282, + 0.188, + 0.297 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.307, + 0.484, + 0.362 + ], + "angle": 0, + "content": "[1] Mokhaled NA Al-Hamadani, Mohammed A Fadhel, Laith Alzubaidi, and Balazs Harangi. Reinforcement learning algorithms and applications in healthcare and robotics: A comprehensive and systematic review. Sensors, 24(8):2461, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.364, + 0.484, + 0.405 + ], + "angle": 0, + "content": "[2] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.407, + 0.484, + 0.448 + ], + "angle": 0, + "content": "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In ICML, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.451, + 0.484, + 0.491 + ], + "angle": 0, + "content": "[4] Lefteris Benos, Avital Bechar, and Dionysis Bochtis. Safety and ergonomics in human-robot interactive agricultural operations. Biosystems Engineering, 200:55-72, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.493, + 0.484, + 0.56 + ], + "angle": 0, + "content": "[5] Nicole Berx, Wilm Decré, Ido Morag, Peter Chemweno, and Liliane Pintelon. Identification and classification of risk factors for human-robot collaboration from a system-wide perspective. Computers & Industrial Engineering, 163:107827, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.564, + 0.482, + 0.592 + ], + "angle": 0, + "content": "[6] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446), 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.594, + 0.486, + 0.8 + ], + "angle": 0, + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In Robotics: Science and Systems XIX, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.803, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[8] Tom B. Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gil mer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.846, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[9] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 3-14, 2017." + }, + { + "type": "list", + "bbox": [ + 0.101, + 0.307, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.161 + ], + "angle": 0, + "content": "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.907, + 0.205 + ], + "angle": 0, + "content": "[11] Shen Cheng and BS Dhillon. Reliability and availability analysis of a robot-safety system. Journal of Quality in Maintenance Engineering, 17:203-232, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.207, + 0.907, + 0.261 + ], + "angle": 0, + "content": "[12] Zhiyuan Cheng, Cheng Han, James Liang, Qifan Wang, Xi-angyu Zhang, and Dongfang Liu. Self-supervised adversarial training of monocular depth estimation against physical-world attacks. arXiv preprint arXiv:2406.05857, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.263, + 0.906, + 0.344 + ], + "angle": 0, + "content": "[13] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility via: Multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.347, + 0.906, + 0.375 + ], + "angle": 0, + "content": "[14] Jinda Cui and Jeff Trinkle. Toward next-generation learned robot manipulation. Sci. Robotics, 6(54):9461, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.377, + 0.907, + 0.431 + ], + "angle": 0, + "content": "[15] Yiming Cui, Cheng Han, and Dongfang Liu. Collaborative multi-task learning for multi-object tracking and segmentation. Journal on Autonomous Transportation Systems, 1:1-23, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.434, + 0.907, + 0.488 + ], + "angle": 0, + "content": "[16] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. *Quar-vla: Vision-language-action model for quadruped robots*. In *ECCV*, 2025." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.49, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[17] Andrew Du, Bo Chen, Tat-Jun Chin, Yee Wei Law, Michele Sasdelli, Ramesh Rajasegaran, and Dillon Campbell. Physical adversarial attacks on an aerial imagery object detector. In WACV, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.547, + 0.906, + 0.588 + ], + "angle": 0, + "content": "[18] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M. Roy. A study of the effect of JPG compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.59, + 0.907, + 0.631 + ], + "angle": 0, + "content": "[19] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In CoRL, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.633, + 0.907, + 0.674 + ], + "angle": 0, + "content": "[20] Saskia Golz, Christian Osendorfer, and Sami Haddadin. Using tactile sensation for learning contact knowledge: Discriminate collision from physical interaction. In ICRA, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.676, + 0.907, + 0.716 + ], + "angle": 0, + "content": "[21] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.719, + 0.907, + 0.759 + ], + "angle": 0, + "content": "[22] Jérémie Guiochet, Mathilde Machin, and Hélène Waeselynck. Safety-critical advanced robots: A survey. Robotics Auton. Syst., 94:43-52, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.762, + 0.907, + 0.843 + ], + "angle": 0, + "content": "[23] Sami Haddadin, Sven Parusel, Lars Johannsmeier, Saskia Golz, Simon Gabl, Florian Walch, Mohamadreza Sabaghian, Christoph Jähne, Lukas Hausperger, and Simon Haddadin. The franka emika robot: A reference platform for robotics research and education. IEEE Robotics & Automation Magazine, 29:46-64, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.846, + 0.907, + 0.901 + ], + "angle": 0, + "content": "[24] Roni-Jussi Halme, Minna Lanz, Joni Kämäräinen, Roel Pieters, Jyrki Latokartano, and Antti Hietanen. Review of vision-based safety systems for human-robot collaboration. Procedia Cirp, 72:111-116, 2018." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6956" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.147 + ], + "angle": 0, + "content": "[25] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.483, + 0.204 + ], + "angle": 0, + "content": "[26] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. In CoRL, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.206, + 0.484, + 0.234 + ], + "angle": 0, + "content": "[27] Félix Ingrand and Malik Ghallab. Deliberation for autonomous robots: A survey. Artif. Intell., 247:10-44, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.235, + 0.484, + 0.289 + ], + "angle": 0, + "content": "[28] ISO 10218-1/2:2011. Robots and Robotic Devices Safety Requirements for Industrial Robots Part 1: Robots/Part 2: Robot Systems and Integration. Standard, International Organization for Standardization, 2011." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.291, + 0.484, + 0.333 + ], + "angle": 0, + "content": "[29] ISO13482. Robots and robotic devices — safety requirements for personal care robots. International Organization for Standardization, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.334, + 0.484, + 0.375 + ], + "angle": 0, + "content": "[30] ISO/TS 15066:2016. Robots and Robotic Devices Collaborative Robots. Standard, International Organization for Standardization, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.377, + 0.484, + 0.418 + ], + "angle": 0, + "content": "[31] Stephen James, Michael Bloesch, and Andrew J. Davison. Task-embedded control networks for few-shot imitation learning. In CoRL, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.42, + 0.484, + 0.488 + ], + "angle": 0, + "content": "[32] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: general robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.49, + 0.484, + 0.546 + ], + "angle": 0, + "content": "[33] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.548, + 0.484, + 0.643 + ], + "angle": 0, + "content": "[34] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.645, + 0.484, + 0.687 + ], + "angle": 0, + "content": "[35] Hema Swetha Koppula and Ashutosh Saxena. Learning spatio-temporal structure from RGB-D videos for human activity detection and anticipation. In ICML, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.688, + 0.484, + 0.743 + ], + "angle": 0, + "content": "[36] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE Trans. Pattern Anal. Mach. Intell., 38(1): 14-29, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.745, + 0.484, + 0.799 + ], + "angle": 0, + "content": "[37] Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learning for manipulation: Challenges, representations, and algorithms. J. Mach. Learn. Res., 22:30:1-30:82, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.802, + 0.484, + 0.829 + ], + "angle": 0, + "content": "[38] Andrew J. Kurdila and Pinhas Ben-Tzvi. Dynamics and control of robotic systems. John Wiley & Sons, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.83, + 0.484, + 0.871 + ], + "angle": 0, + "content": "[39] Przemyslaw A. Lasota, Terrence Fong, and Julie A. Shah. A survey of methods for safe human-robot interaction. Found. Trends Robotics, 5(4):261–349, 2017." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[40] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.486, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.908, + 0.12 + ], + "angle": 0, + "content": "robotic grasping with deep learning and large-scale data collection. Int. J. Robotics Res., 37(4-5):421-436, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.908, + 0.189 + ], + "angle": 0, + "content": "[41] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.191, + 0.908, + 0.261 + ], + "angle": 0, + "content": "[42] Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, and Michael S. Ryoo. Llara: Supercharging robot learning data for vision-language policy. arXiv preprint arXiv:2406.20095, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.262, + 0.906, + 0.304 + ], + "angle": 0, + "content": "[43] Aishan Liu, Xianglong Liu, Jiaxin Fan, Yuqing Ma, Anlan Zhang, Huiyuan Xie, and Dacheng Tao. Perceptual-sensitive GAN for generating adversarial patches. In AAAI, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.305, + 0.908, + 0.347 + ], + "angle": 0, + "content": "[44] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.347, + 0.908, + 0.387 + ], + "angle": 0, + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.389, + 0.908, + 0.418 + ], + "angle": 0, + "content": "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurlPS, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.419, + 0.906, + 0.446 + ], + "angle": 0, + "content": "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.448, + 0.908, + 0.502 + ], + "angle": 0, + "content": "[48] Alessandro De Luca, Alin Albu-Schäffer, Sami Haddadin, and Gerd Hirzinger. Collision detection and safe reaction with the DLR-III lightweight manipulator arm. In IROS, 2006." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.906, + 0.544 + ], + "angle": 0, + "content": "[49] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.908, + 0.587 + ], + "angle": 0, + "content": "[50] Jim Mainprice and Dmitry Berenson. Human-robot collaborative manipulation planning using early prediction of human motion. In IROS, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.589, + 0.906, + 0.616 + ], + "angle": 0, + "content": "[51] Matthew T. Mason. Mechanics of Robotic Manipulation. MIT Press, 2001." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.617, + 0.908, + 0.645 + ], + "angle": 0, + "content": "[52] Matthew T. Mason. Toward robotic manipulation. Annu. Rev. Control. Robotics Auton. Syst., 1:1-28, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.647, + 0.908, + 0.701 + ], + "angle": 0, + "content": "[53] Oier Mees, Lukás Hermann, and Wolfram Burgard. What matters in language conditioned robotic imitation learning over unstructured data. IEEE Robotics Autom. Lett., 7(4): 11205-11212, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.702, + 0.908, + 0.744 + ], + "angle": 0, + "content": "[54] Vivek G Moudgal, Kevin M Passino, and Stephen Yurkovich. Rule-based control for a flexible-link robot. IEEE Trans. on Control Sys. Tech., 2(4):392-405, 1994." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.745, + 0.906, + 0.786 + ], + "angle": 0, + "content": "[55] Richard M. Murray, Zexiang Li, and S. Shankar Sastry. A mathematical introduction to robotics manipulation. CRC Press, 1994." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.788, + 0.908, + 0.829 + ], + "angle": 0, + "content": "[56] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.83, + 0.908, + 0.872 + ], + "angle": 0, + "content": "[57] Vitchyr Pong, Murtaza Dalal, Steven Lin, Ashvin Nair, Shikhar Bahl, and Sergey Levine. Skew-fit: State-covering self-supervised reinforcement learning. In ICML, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.873, + 0.908, + 0.901 + ], + "angle": 0, + "content": "[58] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6957" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.092, + 0.484, + 0.119 + ], + "angle": 0, + "content": "examples jailbreak aligned large language models. In AAAI, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.121, + 0.484, + 0.175 + ], + "angle": 0, + "content": "[59] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In ACM SIGSAC, 2016." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.484, + 0.204 + ], + "angle": 0, + "content": "[60] Michael H. Smith and L. Stephen Coles. Design of a low cost, general purpose robot. In IJCAI, 1973." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.484, + 0.247 + ], + "angle": 0, + "content": "[61] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Trans. Evol. Comput., 23(5):828-841, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.248, + 0.484, + 0.288 + ], + "angle": 0, + "content": "[62] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.289, + 0.484, + 0.343 + ], + "angle": 0, + "content": "[63] Chen Tang, Ben Abbatematteo, Jiaheng Hu, Rohan Chandra, Roberto Martin-Martin, and Peter Stone. Deep reinforcement learning for robotics: A survey of real-world successes. arXiv preprint arXiv:2408.03539, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.344, + 0.484, + 0.44 + ], + "angle": 0, + "content": "[64] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.484, + 0.469 + ], + "angle": 0, + "content": "[65] Florian Tramér. Detecting adversarial examples is (nearly) as hard as classifying them. In ICML, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.47, + 0.484, + 0.51 + ], + "angle": 0, + "content": "[66] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.484, + 0.566 + ], + "angle": 0, + "content": "[67] Valeria Villani, Fabio Pini, Francesco Leali, and Cristian Secchi. Survey on human-robot collaboration in industrial settings: Safety, intuitive interfaces and applications. Mechatronics, 55:248-266, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.567, + 0.484, + 0.608 + ], + "angle": 0, + "content": "[68] Christian Vogel, Christoph Walter, and Norbert Elkmann. A projection-based sensor system for safe physical human-robot collaboration. In IROS, 2013." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.609, + 0.484, + 0.678 + ], + "angle": 0, + "content": "[69] Homer Rich Walke, Kevin Black, Tony Z. Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata V2: A dataset for robot learning at scale. In CoRL, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.679, + 0.484, + 0.734 + ], + "angle": 0, + "content": "[70] Taowen Wang, Yiyang Liu, James Chenhao Liang, Yiming Cui, Yuning Mao, Shaoliang Nie, Jiahao Liu, Fuli Feng, Zenglin Xu, Cheng Han, et al. \\(\\mathbf{M}^2\\)pt: Multimodal prompt tuning for zero-shot instruction learning. In EMNLP, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.484, + 0.789 + ], + "angle": 0, + "content": "[71] Zhibo Wang, Siyan Zheng, Mengkai Song, Qian Wang, Alireza Rahimpour, and Hairong Qi. advpattern: Physical-world attacks on deep person re-identification via adversarially transformable patterns. In ICCV, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.484, + 0.843 + ], + "angle": 0, + "content": "[72] Zixiang Wang, Hao Yan, Zhuoyue Wang, Zhengjia Xu, Zhizhong Wu, and Yining Wang. Research on autonomous robots navigation based on reinforcement learning. In RAIIC, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.484, + 0.901 + ], + "angle": 0, + "content": "[73] Zhijie Wang, Zhehua Zhou, Jiayang Song, Yuheng Huang, Zhan Shu, and Lei Ma. Towards testing and evaluating vision-language-action models for robotic manipulation: An empirical study. arXiv preprint arXiv:2409.12894, 2024." + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.484, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.908, + 0.162 + ], + "angle": 0, + "content": "[74] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.163, + 0.909, + 0.219 + ], + "angle": 0, + "content": "[75] Wikipedia contributors. Finch (film) — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Finch_(film) & oldid=1249481595, 2024. [Online; accessed 5-November-2024]." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.908, + 0.262 + ], + "angle": 0, + "content": "[76] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.263, + 0.908, + 0.317 + ], + "angle": 0, + "content": "[77] Kaidi Xu, Gaoyuan Zhang, Sijia Liu, Quanfu Fan, Mengshu Sun, Hongge Chen, Pin-Yu Chen, Yanzhi Wang, and Xue Lin. Adversarial t-shirt! evading person detectors in a physical world. In ECCV, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.319, + 0.908, + 0.36 + ], + "angle": 0, + "content": "[78] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.362, + 0.908, + 0.417 + ], + "angle": 0, + "content": "[79] Zhiyang Xu, Chao Feng, Rulin Shao, Trevor Ashby, Ying Shen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. Visionplan: Scaling human-labeled tasks in visual instruction tuning. arXiv preprint arXiv:2402.11690, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.418, + 0.908, + 0.459 + ], + "angle": 0, + "content": "[80] Yuchen Zhang and Percy Liang. Defending against whitebox adversarial attacks via randomized discretization. In AISTATS, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.461, + 0.908, + 0.503 + ], + "angle": 0, + "content": "[81] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quanxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, 2022." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.504, + 0.908, + 0.544 + ], + "angle": 0, + "content": "[82] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.546, + 0.908, + 0.602 + ], + "angle": 0, + "content": "[83] Yuanyang Zhu, Zhi Wang, Chunlin Chen, and Daoyi Dong. Rule-based reinforcement learning for efficient robot navigation with space reduction. IEEE/ASME Trans. on Mechatronics, 27:846-857, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.603, + 0.909, + 0.839 + ], + "angle": 0, + "content": "[84] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023." + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.909, + 0.839 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "6958" + } + ] +] \ No newline at end of file diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_origin.pdf b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..2538f44ef74c0cb5ba992fd60287ee7593d89719 --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/654da609-1a74-4dd3-8326-891848884c54_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d6eaf70aaea9258b0a6defc971804c3716ba06bbaa154b0084d6e22def2d5bd1 +size 6528740 diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/full.md b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/full.md new file mode 100644 index 0000000000000000000000000000000000000000..739289403ae30fa6093fa846b2a87e8b2114d4ae --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/full.md @@ -0,0 +1,383 @@ +# Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics + +Taowen Wang $^{1*}$ , Cheng Han $^{2*}$ , James Liang $^{3*}$ , +Wenhao Yang $^{4}$ , Dongfang Liu $^{1\dagger}$ , Luna Xinyu Zhang $^{1\dagger}$ , Qifan Wang $^{5}$ , +Jiebo Luo $^{6}$ , Ruixiang Tang $^{7\dagger}$ + +$^{1}$ Rochester Institute of Technology, $^{2}$ University of Missouri - Kansas City, $^{3}$ U.S. Naval Research Laboratory, $^{4}$ Lamar University, $^{5}$ Meta AI, $^{6}$ University of Rochester, $^{7}$ Rutgers University + +# Abstract + +Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a $100\%$ reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments1. + +# 1. Introduction + +"First directive: A robot cannot harm a human or, through inaction, allow a human to be harmed." + +—Finch [75] + +![](images/a27d1366dc9413109af42625ca8f73a705e6f2d57e087ea6e0769e728e603a08.jpg) + +![](images/a7f1a9cd58f23d1dfa2cd5f851f69b1c8dae7758bd7a5bcc21a17523d63813b5.jpg) +Figure 1. Adversarial Vulnerabilities induced by malicious manipulation. (A). Illustration of adversarial threats in robotic task execution. (B). Example of semantic-rich adversarial patches generated by proposed methods. (C). Comparison of failure rates across different attack schemes (UADA, UPA, and TMA). + +![](images/08f45c5ade4c1077e04a483cdb86841d235414718c7f84a3b13bde22084bdd57.jpg) + +In the movie Finch, set in a post-apocalyptic future, an intelligent robot navigates complex themes, underscoring the importance of interactive safety with its human master. Once speculative, this portrayal is now increasingly plausible with the rise of Large Vision Language Models (LVLMs) [70, 79], capable of seamlessly interpreting both visual and linguistic contexts. A notable realization of this potential can be seen in Vision-Language-Action (VLA) models [16, 34, 73, 74], which integrate LVLMs into robotic systems to enable end-to-end learning that encompasses both high-level trajectory planning and low-level robot action control. While VLA models demonstrate a promising step toward generalist + +robotic intelligence, they also introduce potential vulnerabilities that remain largely underexplored. This work, in light of this view, pioneers the understanding of the pressing need to investigate the new attack surfaces associated with VLA-based robotic systems. + +To gain deeper insights into these vulnerabilities, we analyze VLA-based models and general robotic operations, highlighting two key characteristics crucial for designing effective adversarial attacks. First, creating attack objectives for robotic systems requires consideration of the physical dynamics and constraints intrinsic to robotic movement. Conventional adversarial attacks, in this case, often fail to produce significant deviations in intended actions because they disregard these constraints. Second, VLA models generate control signals that function as a time series of $K$ -class predictions (see §3.1), making it essential to design attacks that exploit the temporal dependencies within these sequences to cause substantial disruptions in robotic behavior. Achieving attacks that are effective in both digital simulations and real-world environments remains critical yet challenging. + +To address these challenges, our work intensifies the adversarial threats posed to VLA-based systems by both developing specialized attack objectives and designing effective attack methods. Specifically, we formulate an Action Discrepancy Objective aimed at maximizing the action discrepancy within VLA-based robotic systems. This strategy ensures that, at each decision point, the robot's behavior can diverge from the optimal trajectory. Additionally, we introduce Geometry-Aware Objective that considers the robot's movement in three-dimensional space, characterized by three degrees of freedom. By optimizing the cosine similarity between adversarial and ground-truth directions, we induce deviations in the robot's movement direction from its intended path, increasing the likelihood of task failure. To achieve these attack objectives, we develop straightforward yet effective Patch-Based Attacks targeting VLA-based robotic systems. This approach enables adversarial attacks in both digital and physical settings, revealing substantial vulnerabilities within the VLA-based system. + +Altogether, these innovations yield several significant contributions: ① This work presents a pioneering and comprehensive analysis of vulnerabilities within VLA-based robotic systems, a new paradigm for training generalized robotic policies using generative foundation models. We reveal critical threats to adversarial attacks, emphasizing the urgent need to enhance robustness before real-world deployment; ② To the best of our knowledge, we are the first work to define attack objectives specific to the powerful VLA models and to employ a straightforward adversarial patch against it (see §3). This offers valuable insights for the research community to explore systemic failures in similar concurrent generative foundation models; ③ We rigorously evaluate our approach in both simulated and real-world environments across four + +distinct robotic tasks, observing significant raises in task failure rates of $100\%$ and $43\%$ , respectively. This highlights the effectiveness of our attack strategies (see $\S 4.3$ ). + +# 2. Related work + +# 2.1. AI-driven Generalist Robot + +Developing generalist robots [6, 14, 27, 37, 60] requires models to not only handle varied interactions but also maintain robustness in unstructured environments. Early intelligent robots [51, 52, 54, 55, 83] generally relied on rule-based approaches, effective only in controlled settings and needing extensive reprogramming for new tasks. The advent of deep learning [19, 31, 40, 56] and reinforcement learning [1, 63, 72] has shifted this paradigm to adapt through the data-driven pipeline, increasing robot versatility and reducing manual reconfiguration. + +Recent studies [7, 53, 57] have further advanced the development and potential of Large Vision Language Models (LVLMs) as key enablers for generalist robots, demonstrating promising generalization across a variety of scenes [15, 25, 32, 41, 64, 84]. One of the notable examples is OpenVLA [34], which integrates dual visual encoders with a pre-trained large language model to enable observation-to-action capabilities through instruction-following [10, 33, 45, 46, 70, 82]. By aligning visual and textual semantics, OpenVLA facilitates complex scenario understanding and end-to-end action generation, showing impressive generalization when executing previously unseen task instructions. + +# 2.2. Adversarial Attacks in Robotic + +Adversarial attacks serve as critical tools for assessing model vulnerabilities, particularly in robotics, where models operate in dynamic, real-world settings. Traditional gradient-based, pixel-level attacks [17, 21, 43, 49, 61, 71] exploit model gradients to calculate malicious perturbations, achieving high attack success rates in digital environments. For physical-world attacks, patch-based methods [3, 8, 59, 76, 77] offer a practical alternative for real-world applications by introducing physically realizable perturbations that retain efficacy under varied conditions, such as angle or lighting changes, making them highly applicable for robotic systems. Therefore, considering VLA-based models enable robot execution in real-world scenarios, this work employs a patch-based attack strategy. + +VLA models [13, 26, 64] couple visual and linguistic modalities for perception-action alignment. The continuous, high-dimensional nature of visual inputs [49, 58, 62] makes them the most likely target for adversarial perturbations, as attackers can subtly alter visual data in ways that are difficult to detect [2, 9, 65]. Consequently, this work applies attacks on the visual modality to achieve robust manipulations across diverse environments. To the best of our knowledge, this + +![](images/c1f482ac55f606457a94bb97d63f25131deeaa895b2426f984ac68d6b5d941b2.jpg) +Figure 2. Overall Adversarial Framework. The robot captures an input image, processes it through a vision-language model to generate tokens representing actions, and then uses an action de-tokenizer for discrete bin prediction. The model is optimized with adversarial objectives focusing on various discrepancies and geometries (i.e., UADA, UPA, TMA). Forward propagation is shown in black, and backpropagation is highlighted in pink. These objectives aim to maximize errors and minimize task performance, with visual emphasis on 3D-space manipulation and a focus on generating adversarial perturbation $\delta$ during task execution, such as picking up a can. + +study is one of the earliest attempts to formally define adversarial objectives for VLA models, targeting gaps in adversarial robustness through the model's geometric constraints (e.g., spatial dependencies in visual input) and architectural nuances (e.g., cross-modal information integration). + +# 3. Methodology + +In this section, we first review the core algorithmic principles of VLA in §3.1, which serves as the foundation for the adversarial scheme designs. We then present two types of untargeted adversarial attacks, including Untargeted Action Discrepancy Attack (UADA) in §3.2 and Untargeted Position-aware Attack (UPA) in §3.3. In §3.4, we further include Targeted Manipulation Attack (TMA), designed to direct the robot toward specific erroneous execution. Finally, we introduce the Normalized Action Discrepancy (NAD) metric in §3.5. The overall framework is shown in Fig. 2. + +# 3.1. Preliminary + +Vision-Language-Action models [34, 84] (see Fig. 2) are built on large language models integrated with visual encoders, enabling robots to interpret human instructions and process visual input from a camera to perform context-aware actions. VLA models [34, 42, 84] generally abstract continuous action predictions into a classification problem by discretizing robot actions. Specifically, they first discretize the continuous probabilities into class labels $y = \arg \max_{x} \mathcal{F}(x)$ where $\mathcal{F}(\cdot)$ is the VLA model. An action de-tokenizer then generates actions $A = DT(y)$ . By categorizing action values into discrete class labels, the model converts continuous probability outputs into discrete signals, this simplification facilitates quicker convergence and faster training times, and is commonly used for VLA-based models [34, 42, 84]. + +In this work, we focus on a 7 degree-of-freedoms (DoFs) robotic arm [23]. At each step, an action consists of 7 DoFs with specific physical significance in three-dimensional Cartesian space, represented by: + +$$ +A = \left[ \Delta P _ {x}, \Delta P _ {y}, \Delta P _ {z}, \Delta R _ {x}, \Delta R _ {y}, \Delta R _ {z}, \text {g r i p p e r} \right], \tag {1} +$$ + +where $\Delta P_{x,y,z}$ and $\Delta R_{x,y,z}$ denote relative positional and rotational changes along the x-, y-, and z-axis, discretized into 256 uniform bins across each DoFs' action bounds $[y_{\mathrm{min}}^i, y_{\mathrm{max}}^i]$ [34]. The gripper state is binary, indicating its open or closed state. This control design presents a unique challenge for adversarial attacks, as finely divided bins result in minimal action discrepancies between neighboring bins (e.g., $\pm 0.007/\mathrm{bin}$ ). This means that even if an attack shifts the classification to an adjacent bin, the resulting action discrepancy has minimal impact on real-world performance. + +# 3.2. Untargeted Action Discrepancy Attack + +To exacerbate action discrepancies, we introduce the Untargeted Action Discrepancy Attack (UADA), which aims to maximize deviations in robot actions. This attack is based on the observation that larger robot actions usually correlate with intense physical movements, which, in turn, may amplify the potential for real-world hazards [28-30]. For UADA, the attack target is one or a combination of 7 DoFs, defined as $y_{gt}^{I} = \{y_{gt}^{i}|i\in [1,\dots ,7]\}$ . To define UADA's objective, we first identify the most distant action $y_{adv}^{i}$ , which maximizes the discrepancy from the $i$ -th DoF ground truth action $y^{i}$ . This action, for each DoF, corresponds to its own action bound $[y_{\min}^{i},y_{\max}^{i}]$ as: + +$$ +y _ {a d v} ^ {i} = \left\{ \begin{array}{l l} y _ {\max } ^ {i}, & \text {i f} | y _ {\max } ^ {i} - y _ {g t} ^ {i} | \geq | y _ {\min } ^ {i} - y _ {g t} ^ {i} | \\ y _ {\min } ^ {i}, & \text {o t h e r w i s e} \end{array} . \right. \tag {2} +$$ + +Instead of directly using $y_{adv}^{i}$ as the misclassification target, we introduce a soft attack objective to capture the discrepancy between actions, ensuring smooth gradient optimization and stable attack performance. Specifically, due to the physical action magnitude information contains in bin labels, we reweight the output probability $\mathcal{F}(x)_{bins}^{i} \in \mathcal{R}^{1 \times bins}$ using normalized bin labels $y_{bins}^{i} = [\frac{1}{bins}, \dots, 1]$ , where $\lfloor y_{bins}^{i} \rfloor$ and $\lceil y_{bins}^{i} \rceil$ correspond to the normalized values of $y_{\min}^{i}$ and $y_{\max}^{i}$ , respectively. The reweighting process is defined as: + +$$ +y _ {s o f t} ^ {i} = \sum_ {b i n s = 1} ^ {n} \mathcal {F} (x + \delta) _ {b i n s} ^ {i} \otimes y _ {b i n s} ^ {i}, \tag {3} +$$ + +where $\otimes$ denotes Hadamard Product, $\delta$ is the adversarial patch and $y_{\text{soft}}^i$ represents the $i$ -th DoF's soft action. Finally, the objective of UADA is to minimize the discrepancy between the soft and the most distant actions, which is: + +$$ +\mathcal {L} _ {\mathrm {U A D A}} = \mathbb {E} _ {(x, y) \sim \mathcal {X}} \sum_ {i} ^ {I} \left(y _ {s o f t} ^ {i} - y _ {a d v} ^ {i}\right) ^ {2}, \tag {4} +$$ + +UADA defines the objective function $\mathcal{L}_{\mathrm{UADA}}$ tailored for attacking the robot's action space, allowing adversarial patches to create significant, lasting disruptions in task performance. + +# 3.3. Untargeted Position-aware Attack + +We explore untargeted adversarial attacks in parallel by considering the positional dynamics within VLA models. In typical task execution, precise and directed movements of the end-effector towards designated goals are essential, with actions mapped in three-dimensional space. The positional DoFs, $A^p = [\Delta P_x, \Delta P_y, \Delta P_z]$ encapsulates the directional movements within 3D for effective robotic control. + +Recognizing the importance of $A^p = DT(y^p)$ in controlling the end-effector's path, we introduce a position-aware attack to disrupt the intended movement trajectory. This type of adversarial vulnerability remains largely unexplored but has significant implications for task failure: the attack objective can steer the end-effector away from its intended path by introducing directional perturbations, amplifying errors and causing substantial trajectory distortions. Formally, we define the Untargeted Position-aware Attack (UPA) objective as: + +$$ +\mathcal {L} _ {\mathrm {U P A}} = \mathbb {E} _ {(x, y) \sim \mathcal {X}} \left[ \alpha \frac {y _ {a d v} ^ {p} \cdot y ^ {p}}{\| y _ {a d v} ^ {p} \| \| y ^ {p} \|} + \beta \frac {1}{\| y _ {a d v} ^ {p} - y ^ {p} \| _ {2}} \right], \tag {5} +$$ + +where $\alpha$ and $\beta$ are the hyperparameters that balance between the directionality and magnitude of the perturbations. By integrating geometric awareness into our attack, this approach generates perturbations that induce cumulative deviations in the robot's trajectory, even with minimal adjustments. $L_{2}$ -normalization $\| \cdot \| _2$ further intensifies these deviations, rendering the attack highly effective and disruptive. + +# 3.4. Targeted Manipulation Attack + +In addition to the aforementioned untargeted attacks, we also explore the adversarial vulnerabilities of VLA models with a targeted strategy. This attack aims to directly mislead the models into predicting specific trajectories, resulting in precise alterations. We design this attack since the targeted alterations can drive malicious behaviors or induce task failure. The objective of the targeted manipulation attack is: + +$$ +\mathcal {L} _ {\mathrm {T M A}} = \mathbb {E} _ {(x, y) \sim \mathcal {X}} \left[ C E \left(\mathcal {F} (x + \delta) ^ {I}, y _ {T} ^ {I}\right) \right], \tag {6} +$$ + +where $y_{T}^{I} = \{y_{T}^{i} = t|i\in [1,\dots ,7],t\in [y_{\min}^{i},y_{\max}^{i}] \}$ represents the adversarial target(s); and $CE(\cdot ,\cdot)$ denotes the Cross-Entropy loss, which measures the error between the predicted state under adversarial perturbation and the desired target state. By steering the robot toward an adversarial target across successive time steps, our approach manipulates the trajectory and undermines task performance, ultimately leading to a substantial disruption of intended operations. Furthermore, manipulating any individual DoF or a combination thereof can also severely compromise task success, given that the errors in one dimension tend to propagate and magnify over time during execution (see Fig. 3). + +# 3.5. Metric: Normalized Action Discrepancy + +To assess the potential physical impact of adversarial patches, we introduce Normalized Action Discrepancy (NAD) as a fine-grained metric for evaluating discrepancies at the action level. NAD serves as a complementary measurement alongside the coarse-grained Failure Rate (FR) at the task level, providing a detailed analysis of deviations during action execution. To define the NAD, we first determine the maximum action discrepancy $d_{\mathrm{max}}^i$ , which represents the largest deviation that can occur within the allowed range for $i$ -th DoF. This is computed by measuring the L1 distance between the ground truth $y_{gt}^i$ and its action bound $[y_{\mathrm{min}}^i, y_{\mathrm{max}}^i]$ as: + +$$ +d _ {\mathrm {m a x}} ^ {i} (y) = \max \left[ D T \big (| y _ {g t} ^ {i} - y _ {\mathrm {m i n}} ^ {i} | \big), D T \big (| y _ {g t} ^ {i} - y _ {\mathrm {m a x}} ^ {i} | \big) \right]. (7) +$$ + +We next calculate the applied action discrepancy $d_{\mathrm{applied}}^i (x,y)$ $= |DT(\mathcal{F}(x)^i) - DT(y_{gt}^i)|$ to measure the deviation between the model's output and ground truth. We define the NAD as the drifting ratio between $d_{\mathrm{applied}}^i$ and $d_{\mathrm{max}}^i$ as: + +$$ +\mathrm {N A D} = \frac {1}{\mathrm {I}} \sum_ {\mathrm {i}} ^ {\mathrm {I}} \frac {\mathrm {d} _ {\text {a p p l i e d}} ^ {\mathrm {i}}}{\mathrm {d} _ {\max } ^ {\mathrm {i}}}, \tag {8} +$$ + +where $I \in [1, \dots, 7]$ represents a single or a combination of DoF(s). In this way, NAD provides normalized measurement of action deviations, allowing for a consistent evaluation of adversarial impact across varying DoF(s). + +# 4. Experiments + +In this section, we first detail the implementation of our adversarial framework and baseline methods in §4.1. Next, we + +![](images/053d65c210e902c1eb3f997270755a01597bfa93840980fb3420e9367d440e72.jpg) + +![](images/56b3f6c182dadf720ec5078e7ad1dc0489d0d62a8e2794270f803036b9f8ec51.jpg) +TASK (UADA) : Pick up the black bowl between the plate and the ramekin and place it on the plate + +![](images/e568da881f28597e9b6b015d091a52624fbabdb43f3356c6d1cf874c6c3ffe94.jpg) +TASK (UPA) : Pick up the black bowl on the ramekin and place it on the plate + +![](images/17fa09dc76817dc102176fa4d847ebf1f5c9b294e44ed926ac405b8cb258b33b.jpg) + +![](images/68bbb29a18557a64d22e434219a261b2da93f2cdbfab928084518b2d8d9efed1.jpg) +TASK (TMA) : Pick up the alphabet soup and place it in the bask + +![](images/aeae4fff3aed917ce687bb61bf292a422d815551dcb583cffeda6f27f537b83e.jpg) +Figure 3. Qualitative Results of adversarial vulnerabilities over OpenVLA-7B [34] and OpenVLA-7B-LIBERO [34] with objectives of UADA, UPA, and TMA, respectively. We visualize the overall 3D trajectories and 2D trajectories of benign $\bullet$ and adversarial $\bullet$ scenarios at each time step to compare the impact of the generated adversarial patch in affecting them. The untargeted trajectory $\bullet$ is visualized in UADA task. All trajectories start with $\triangle$ , and we plot the success end point, marked as $\star$ . + +outline the experimental setup in $\S 4.2$ and present quantitative and qualitative results in both simulation and physicalworld scenarios (§4.3). We then conduct diagnostic experiments (§4.4) to analyze the impact of key components. Additionally, we assess the robustness of our method against various defense mechanisms in $\S 4.5$ and conclude with a comprehensive systemic discussion in $\S 4.6$ . + +# 4.1. Implementation Details + +The implementation of our adversarial patch attack pipeline is detailed in Algorithm 1. We incorporate two key modifications to enhance the training stability of the generated patch. Inner-loops. Following previous work [49], we add $k$ inner-loops optimize steps during each iteration in line 4, aiming to reduce data variance and ensure more consistent updates. Geometric Transformations. To improve the robustness of our attack under real-world scenarios, we employ random geometric transformations $\mathcal{T}[(\cdot), (sh_x, sh_y, \theta)]$ in line 6 with + +# Algorithm 1: Adversarial Patch Attack. + +
1:Input: X: dataset; δ: patch; L0: attack objective; +F: VLA model; T(·): transformations; φ,ψ: +transformation parameters; T: attack steps; k: +inner-loop steps.
2:Initialize: δ ~ U[0,1), L0 ∈ {LUADA, LUPA, LTMA}
3:for i = 1,2,···, T do
4:for i = 1,2,···, k do
5:shx, shy ~ U(-φ,φ), θ ~ U(-ψ,ψ)
6:ypred ← F[T((x + δ), (shx, shy, θ))]
7:δ ← ∂L0/∂δ{▷ Update}
8:end for
9:end for
+ +affine and rotation transformations [3] as: + +$$ +\mathcal {T} \left[ (\cdot), \left(s h _ {x}, s h _ {y}, \theta\right) \right] = \left[ \begin{array}{c c} 1 & s h _ {y} \\ s h _ {x} & 1 \end{array} \right] \cdot \left[ \begin{array}{c c} \cos \theta & - \sin \theta \\ \sin \theta & \cos \theta \end{array} \right], \tag {9} +$$ + +where $sh_x, sh_y \sim \mathcal{U}(-\phi, \phi)$ are shear factors, and $\theta \sim \mathcal{U}(-\psi, \psi)$ denotes the rotation angle, both sampled from uniform distributions during inner loop optimization step $k$ . Untargeted Manipulation Attack (UMA) Baseline. As a pioneering work, we could not find a directly comparable baseline for our attacks. Therefore, we adapt prior work in adversarial learning as one of our baseline methods [66]. This attack is designed to mislead classifiers; we integrate it into our framework, leading to the following objective: + +$$ +\mathcal {L} _ {\mathrm {U M A}} = - \mathbb {E} _ {(x, y) \sim \mathcal {X}} \left[ C E \left(\mathcal {F} (x + \delta) ^ {I}, y _ {g t} ^ {I}\right) \right]. \tag {10} +$$ + +By minimizing the negative CE loss, this attack pushes the predicted action away from the ground truth action. + +Random Noise Baseline. We generate random noise patches as an additional baseline, representing an unstructured intervention without any learned adversarial intent. Specifically, the noise is sampled from a Gaussian distribution $\mathcal{N}(0,1)$ and evaluated with the same setting as proposed attacks. + +# 4.2. Experiment Setup + +Datasets. We conduct attacks on BridgeData V2 [69] and LIBERO [44] with corresponding VLAs. BridgeData V2 [69] is a real-world dataset comprising 24 diverse environments and 13 distinct skills, such as grasping, placing, and object rearrangement, with a total of 60,096 trajectories. LIBERO [44] is a simulation dataset designed to evaluate models across four distinct task types (Spatial, Object, Goal, Long). Notably, LIBERO-Long combines diverse objects, layouts, and long-horizon tasks, making them challenging for complex, multi-step planning, thus we choose LIBERO-Long to conduct UADA, UPA and TMA. + +Victim VLAs. In our study, we select publicly available and current most influential VLAs as victim models for evaluation. Specifically, we focus on four variants of the OpenVLA model, each trained independently on different task suites within the LIBERO dataset [44] (i.e., Spatial, Object, Goal, Long). To evaluate the effectiveness of our methods, we craft adversarial patches using three distinct generating setups: Simulation Setting involves a model trained in a simulated environment using the LIBERO Long suite with the openvla-7B-libero-long variant [34]. Physical Setting uses a model trained on physical world data from the BridgeData v2 [69] with the openvla-7B model [34]. This approach allows us to assess our methods on both simulated and real-world data. Subsequently, we evaluate the performance of generated adversarial patches on victim models (i.e., OpenVLA LIBERO variants) trained on different tasks suites to rigorously prove the robustness and effectiveness of our method. These models are trained with distinct data sources and task objectives. + +Robot Setups (Real-world Setting). For the real-world tasks, we adopt a robotic system consisting of a Universal Robots UR10e equipped with a Robotiq Hand-E Gripper to + +provide a 7DoF motion. The sensing system includes one RGB webcam in the fixed position with a shoulder view. + +Evaluation Details. To assess the effectiveness and robustness of our methods, we conduct experiments on the LIBERO dataset [44]. Each suite consists of 10 tasks, with each task executed for 50 trials, resulting in a total of 500 rollouts, following Kim et al. [34]. For reproducibility, we carefully select patch paste locations for each suite, ensuring they do not obscure objects or robots in the test environment. + +Evaluation Metric. Regarding the task execution evaluation, we take the maximum steps of each task suite in the LIBERO training dataset as the timeout failure condition to reduce computational overhead. Furthermore, building on the concept of Success Rate (SR) introduced in LIBERO [44], we adopt Failure Rate (FR), defined as $1 - \mathrm{SR}$ , as the primary evaluation metric. To further quantify action discrepancy, we employ NAD (see Eq. 8) for untargeted attacks on corresponding DoF(s), as it considers relative distances to quantify deviation from the optimal action. A higher NAD value indicates a greater discrepancy between the ground truth and the predicted action. Additionally, we report the standard deviation of FR across tasks within each task suite. + +# 4.3. Main Result + +Quantitative Results. For UADA and UPA, our methods effectively amplify action discrepancies, leading to a notable transfer attack ability in increasing failure rates (see Tab. 1a). Specifically, while attacking $\mathrm{DoF}_1$ and $\mathrm{DoF}_{1\sim 3}$ in the Simulation setup, UADA and UPA achieve NAD of $21.0\%$ and $14.5\%$ , significantly outperforming UMA scenarios with increments of $6.9\%$ and $3.1\%$ , respectively. Both UADA and UPA effectively disrupt robot execution, yielding maximum average failure rates of $100\%$ and $89.7\%$ , respectively. Regarding the Physical setup (see Tab. 1b), our attack methods demonstrate strong transferability, as adversarial patches generated in the physical setting significantly impact the performance of LIBERO-X test. Specifically, both UADA and UPA generate malicious actions with large action discrepancies, which are $32.6\%$ ( $\mathrm{DoF}_1$ ) and $26.9\%$ ( $\mathrm{DoF}_{1\sim 3}$ ) with an increase of $5.9\%$ and $3.1\%$ compared to $26.7\%$ ( $\mathrm{DoF}_1$ ) and $23.8\%$ ( $\mathrm{DoF}_{1\sim 3}$ ) in UMA scenarios, respectively. Compare the two settings (Tab. 1a v.s. Tab. 1b), we observe a general variance in NAD (e.g. $32.6\%$ v.s. $21.0\%$ for UADA). This discrepancy can be attributed to the fundamental differences between the simulation and real-world datasets used for conducting attacks (Bridge V2 [69] v.s. LIBERO [44]). The increased variability in real-world data, including environmental complexity, object diversity, and task difficulty, allows the robot more opportunities to generate larger action discrepancies within the validation dataset. + +For TMA task, we evaluate the effectiveness of our method by manipulating all DoF(s) to 0 (i.e., $y_{T}^{i} = 0$ in Eq. 6). Our method demonstrates significant effectiveness + +Table 1. Untargeted Results. We report FR and NAD in LIBERO simulation. * denotes the in-domain victim model and dataset aligned with the patch generation model and dataset, △ denotes a transfer attack evaluation involving a distinct victim model and dataset (see §4.2). The FR (↑) is highlighted in best and second best for each task suite. + +
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long*Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF114.1100±0.0%99.0±3.0%100±0.0%100±0.0%99.8%
DoF1~311.4100±0.0%98.8±2.6%99.0±2.4%100±0.0%99.5%
UADADoF121.0100±0.0%99.2±2.4%100±0.0%100±0.0%99.8%
DoF1~317.9100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~314.596.2±11.4%77.8±12.5%88.0±10.4%96.8±3.0%89.7%
+ +(a) Simulation Setup. Adversarial patch generated on OpenVLA-LIBERO-10 victim model with LIBERO-Long dataset and evaluated on LIBERO-X. + +
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long△Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF126.789.6±15.3%60.8±16.8%64.8±26.8%80.0±20.8%73.8%
DoF1~323.896.6±6.2%56.4±20.4%80.0±20.5%82.0±17.9%78.8%
UADADoF132.699.2±1.8%98.8±2.4%92.6±15.2%96.6±4.8%96.8%
DoF1~327.7100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~326.995.6±9.5%57.0±23.6%57.2±24.4%72.8±19.7%70.7%
+ +(b) Physical Setup. Adversarial patch generated on OpenVLA-7b victim model with BV2 dataset and evaluated on LIBERO-X. + +Table 2. Targeted Manipulation Attack Results. Failure Rate (FR, ↑) and its standard deviation across tasks within LIBERO [44] suite are reported. The performance of task suits and DoFs is averaged separately, with the best and second-best values separately. + +
Targeted DoF(s)MetricSimulation (Attacked on OpenVLA-LIBERO-10)Physical (Attacked on OpenVLA-7b)
Spatial△Object△Goal△Long*AvgSpatial△Object△Goal△Long△Avg
BenignFR15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
DoF1FR100±0.0%97.8±3.0%100±0.0%100±0.0%99.5%90.6±19.9%36.6±18.1%61.0±21.7%86.4±12.7%68.7%
DoF2100±0.0%99.2±1.3%100±0.0%100±0.0%99.8%99.6±0.8%69.2±11.1%76.2±26.1%91.8±9.9%84.2%
DoF3100±0.0%97.2±6.0%100±0.0%100±0.0%99.3%72.0±14.8%30.2±17.4%39.2±20.9%62.4±17.5%51.0%
DoF494.0±11.0%37.4±23.2%51.2±26.5%87.6±14.9%67.6%93.6±9.7%31.8±20.3%45.0±18.3%63.8±17.0%58.6%
DoF5100±0.0%92.6±8.4%100±0.0%100±0.0%98.2%97.8±4.9%36.0±22.5%62.0±24.8%79.4±21.3%68.8%
DoF6100±0.0%80.0±9.5%92.6±13.4%100±0.0%93.2%79.0±14.2%22.2±14.4%41.0±21.5%62.6±15.7%51.2%
DoF799.2±1.3%86.6±8.3%67.2±22.5%96.4±6.4%87.4%91.8±15.3%98.8±1.3%70.6±23.4%93.2±10.2%88.6%
DoF1~3100±0.0%100±0.0%100±0.0%100±0.0%100%96.4±8.0%83.6±16.8%74.4±26.3%91.8±11.9%86.6%
Task Avg99.2%86.4%88.9%98.0%-90.1%51.1%58.7%78.9%-
+ +in manipulating robotic trajectory and increasing FR (see Tab. 2). Specifically, when applied to the different generation settings, our approach yields notable increments across various tasks, including a max average failure rate $100\%$ v.s. $23.5\%$ of benign performance in Simulation, and $88.6\%$ in Physical, respectively. Moreover, during attack $\mathrm{DoF}_4$ , we observe low FR and high task deviation. This failure can be attributed to the fact that $\mathrm{DoF}_4$ controls the orientation along the x-axis, which can be redundant DoF in tasks. As a result, attacks targeting redundant DoF(s) [38] are less likely to disrupt execution effectively. This observation underscores the importance of task-specific considerations when designing adversarial attacks on robotic systems. + +Qualitative Results. We qualitatively analyze robot movement trajectories under the three proposed attacks in Fig. 3. For UADA, we compare the trajectory deviations in the same trail with patches generated from UADA and UMA attacks. As seen, the UMA induces small deviations in the trajectory. UADA, on the other hand, produces significantly larger trajectory deviations (also supported by NAD metric in Tab. 1b), which is attributed to UADA's capability of incorporating action discrepancies. In our observation, UADA significantly amplifies the effect on overall task execution, thereby increasing the potential for robotic hazards. For UPA, we observe chaotic and irregular behaviors, including instances where the end-effector moves out of the camera's field of view. We attribute this phenomenon to the efficacy of the adversarial patch in perturbing the model's spatial perception, + +inducing a consistent deviation from the intended direction of movement, ultimately resulting in failure in a long run. For TMA, we observe a notable reduction in the range of motion along the x-axis corresponding to the targeted attack axis. This suggests that our proposed targeted attack can effectively constrained robot movements. + +In summary, the qualitative analysis shows that both Untarget attack, UADA, UPA, and TMA can effectively disrupt the robot actions generated from VLA models. These findings underscore a pressing security concern during the deployment of generalist robots, especially when considering application scenes that require reliable operations [11, 68]. + +![](images/0161ac097af27979f5c686f293cee7b1fa3c957783c496e98982d3444ce4ada4.jpg) +Figure 4. Qualitative Results of the physical world. The first/second row show benign and adversarial cases respectively. + +Real-world Performance. In addition to the digital simulation results, we also conduct a comprehensive evaluation of the performance of our generated adversarial patches in real-world scenarios. The evaluation encompassed 100 trials across three distinct tasks: object grasping, placement, and manipulation. As shown in Fig. 4, the adversarial patch + +![](images/3fde9a9844a89860e0754a29d91bdf04c8255d59b85cad5673fd228251be5c83.jpg) +(a) Inner-loop + +![](images/809061dfd5fe3958fdfd83eb8711dd71b12036eafb386de14c8b3a2b22dae332.jpg) +(b) Patch Size + +![](images/aa8c52338edb1b5e3959b050bc5389712291477bbcde871f240f2d15a3a4a905.jpg) +(c) JPEG + +![](images/50e2ebe37466de40a562310a4612b267b5cf1fe5328d4545b824eccc09b63551.jpg) +(d) Noise + +![](images/aec35d748df36db7070db39a02bd5670bfdacc54ab9f44352ab116361f175606.jpg) +(e) Median Blur + +![](images/eb35dcf69370a82268f417f57d53cb4d9376908e211032ace0d6e0c65cfb30b6.jpg) +(f) Bit-Depth + +generated with UADA demonstrated the ability to manipulate the robot effectively, achieving an attack success rate exceeding $43\%$ . Although this success rate is lower than the corresponding digital-world performance (i.e., $100\%$ ), it highlights the effectiveness of our patches in physical-world applications as well without the need for further adaptations. Crucially, we observed that the induced erratic movements (similar to simulation scenario) of the robot during successful attacks pose significant risks to human safety and surrounding environments. This finding emphasizes the severe threat posed by VLA models in real-world applications, consistent with similar observations in digital settings. + +# 4.4. Diagnostic Experiment + +Impact of Inner-loops. We discuss the impact of inner-loop steps to model performance in Fig. 5(a). The results show that NAD first improves when inner-loop steps continue to increase. This is due to the reduced data variance and gradually stabilized gradients. However, further increasing the steps results in a slower performance increase, with a noticeably longer training schedule. We thus choose inner-loop steps=50 as it balances between performance and scale. + +Impact of Patch Size. We study TMA's performance with different patch sizes and report the average FR across four LIBERO [44] tasks in Fig. 5(b). The patch sizes examined are $[3\%, 5\%, 7\%, 10\%]$ of input image. Our findings indicate an inversely proportional relationship between the $L_{1}$ distance of predicted actions and the target action as the patch size increased. This increase in patch size also correlated with a marked rise in FR. The observed trend suggests larger patches give adversaries more optimization space to influence the model, aligning with prior work [12, 81]. + +# 4.5. Robustness Evaluation + +We further examine whether concurrent defense strategies [18, 78, 80] can resist our adversarial examples within VLA. Specifically, we applied four prior defense techniques and reported their FR for both benign and adversarial samples in Fig. 5(c-f), respectively. The results show that our adversarial attack bypasses most defense strategies. + +# 4.6. Systemic Discussion + +Patch Pattern Analysis. In Fig. 6, we present the adversarial patches generated for a range of targets. A particularly + +![](images/6647818771e16f42e30d226e9343a943174dc32d516535cce66e32304da2161e.jpg) +Figure 5. Impact of Inner-loop, Patch Size and Defense Discussion. The figure shows how varying Inner-loop affects NAD in UADA, and patch sizes affect $L1$ distance and the failure rates in TMA, both targeting at DoF $_1$ . (a) Impact of Inner-loop, (b) Impact of Patch Size and (c-f) the effect of four different defenses on failure rates. + +![](images/b0278a6162a00b426c5c683f1356a894d55279d8a5fa69166b1d9654d68b6fb3.jpg) +(a) BridgeData V2 [69] + +![](images/270053252a7867145c80093655706037c8dda7c601624b1ae2fb649b325f4f39.jpg) +Figure 6. Patch Visualization. Semantic-rich patches. + +![](images/ee3a2ee1d1681df590d280ecb23433001c527085eda6cfa90dfa87f3741023b3.jpg) +(b) LIBERO [44] + +![](images/3c4da73883a24268bd55d0e3459b8a31f977fd15624a301f0ce345ffc9426d13.jpg) + +![](images/6c71575b7ab846bd94392342ea4be2ad4acfe3eb7f9991a7caa58911db48c16b.jpg) + +noteworthy observation is that some of these patches bear a striking resemblance to the structural joints of a robotic arm. Given the strong similarity between these patches and robotic arms in appearance, we hypothesize that these VLA-based models, in their current training paradigms, are often limited to a narrow range of robotic systems operating within restricted camera views. This constrained perspective camera pose may inadvertently induce a learned representation bias, where adversarial perturbations align with prominent structural features of the training robot's appearance in order to deceive current victim models. Given that these patches are model-specific and reflect the spatial and visual patterns of the training environment, we have a reasonable concern that they may compromise VLA-based models' generalizability and robustness in physical-world deployments. + +Possible Defense Strategies. Our work highlights the pressing need for a paradigm shift in the VLA-based models' training strategies in order to successfully defend our attacks. Specifically, future approaches might incorporate more training scenarios that involve complex multi-robot interactions to mitigate the current patch bias. Another possible solution is to leverage the logical relationships within the robot arm's physical structure to estimate the actual position of each joint. By predicting joint positions deviated from physical constraints, we can eliminate unreasonable adversarial samples, even if they are similar to actual joint(s). + +# 5. Conclusion + +While VLA models have gained significant popularity due to their substantial robot capabilities, this study pioneers specific robotic attack objectives and demonstrates that these models are in fact vulnerable to various types of attacks. Experimental results demonstrate that our attacks expose significant vulnerabilities to VLA models, raising concerns regarding impetuous real-world deployments. We believe our framework provides pioneering and foundational contributions to the reliability of AI-enhanced robotic systems. + +# Acknowledgements + +This research was supported by the National Science Foundation under Grant No. 2450068 and No. 2242243, as well as start-up resources provided by Rutgers University. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of U.S. Naval Research Laboratory (NRL) or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein. + +# References + +[1] Mokhaled NA Al-Hamadani, Mohammed A Fadhel, Laith Alzubaidi, and Balazs Harangi. Reinforcement learning algorithms and applications in healthcare and robotics: A comprehensive and systematic review. Sensors, 24(8):2461, 2024. +[2] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018. +[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In ICML, 2018. +[4] Lefteris Benos, Avital Bechar, and Dionysis Bochtis. Safety and ergonomics in human-robot interactive agricultural operations. Biosystems Engineering, 200:55-72, 2020. +[5] Nicole Berx, Wilm Decré, Ido Morag, Peter Chemweno, and Liliane Pintelon. Identification and classification of risk factors for human-robot collaboration from a system-wide perspective. Computers & Industrial Engineering, 163:107827, 2022. +[6] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446), 2019. +[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In Robotics: Science and Systems XIX, 2023. +[8] Tom B. Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gil mer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2018. +[9] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 3-14, 2017. + +[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023. +[11] Shen Cheng and BS Dhillon. Reliability and availability analysis of a robot-safety system. Journal of Quality in Maintenance Engineering, 17:203-232, 2011. +[12] Zhiyuan Cheng, Cheng Han, James Liang, Qifan Wang, Xi-angyu Zhang, and Dongfang Liu. Self-supervised adversarial training of monocular depth estimation against physical-world attacks. arXiv preprint arXiv:2406.05857, 2024. +[13] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility via: Multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024. +[14] Jinda Cui and Jeff Trinkle. Toward next-generation learned robot manipulation. Sci. Robotics, 6(54):9461, 2021. +[15] Yiming Cui, Cheng Han, and Dongfang Liu. Collaborative multi-task learning for multi-object tracking and segmentation. Journal on Autonomous Transportation Systems, 1:1-23, 2024. +[16] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. *Quar-vla: Vision-language-action model for quadruped robots*. In *ECCV*, 2025. +[17] Andrew Du, Bo Chen, Tat-Jun Chin, Yee Wei Law, Michele Sasdelli, Ramesh Rajasegaran, and Dillon Campbell. Physical adversarial attacks on an aerial imagery object detector. In WACV, 2022. +[18] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M. Roy. A study of the effect of JPG compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016. +[19] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In CoRL, 2017. +[20] Saskia Golz, Christian Osendorfer, and Sami Haddadin. Using tactile sensation for learning contact knowledge: Discriminate collision from physical interaction. In ICRA, 2015. +[21] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015. +[22] Jérémie Guiochet, Mathilde Machin, and Hélène Waeselynck. Safety-critical advanced robots: A survey. Robotics Auton. Syst., 94:43-52, 2017. +[23] Sami Haddadin, Sven Parusel, Lars Johannsmeier, Saskia Golz, Simon Gabl, Florian Walch, Mohamadreza Sabaghian, Christoph Jähne, Lukas Hausperger, and Simon Haddadin. The franka emika robot: A reference platform for robotics research and education. IEEE Robotics & Automation Magazine, 29:46-64, 2022. +[24] Roni-Jussi Halme, Minna Lanz, Joni Kämäräinen, Roel Pieters, Jyrki Latokartano, and Antti Hietanen. Review of vision-based safety systems for human-robot collaboration. Procedia Cirp, 72:111-116, 2018. + +[25] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023. +[26] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. In CoRL, 2023. +[27] Félix Ingrand and Malik Ghallab. Deliberation for autonomous robots: A survey. Artif. Intell., 247:10-44, 2017. +[28] ISO 10218-1/2:2011. Robots and Robotic Devices Safety Requirements for Industrial Robots Part 1: Robots/Part 2: Robot Systems and Integration. Standard, International Organization for Standardization, 2011. +[29] ISO13482. Robots and robotic devices — safety requirements for personal care robots. International Organization for Standardization, 2014. +[30] ISO/TS 15066:2016. Robots and Robotic Devices Collaborative Robots. Standard, International Organization for Standardization, 2016. +[31] Stephen James, Michael Bloesch, and Andrew J. Davison. Task-embedded control networks for few-shot imitation learning. In CoRL, 2018. +[32] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: general robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022. +[33] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024. +[34] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024. +[35] Hema Swetha Koppula and Ashutosh Saxena. Learning spatio-temporal structure from RGB-D videos for human activity detection and anticipation. In ICML, 2013. +[36] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE Trans. Pattern Anal. Mach. Intell., 38(1): 14-29, 2016. +[37] Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learning for manipulation: Challenges, representations, and algorithms. J. Mach. Learn. Res., 22:30:1-30:82, 2021. +[38] Andrew J. Kurdila and Pinhas Ben-Tzvi. Dynamics and control of robotic systems. John Wiley & Sons, 2019. +[39] Przemyslaw A. Lasota, Terrence Fong, and Julie A. Shah. A survey of methods for safe human-robot interaction. Found. Trends Robotics, 5(4):261–349, 2017. +[40] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for + +robotic grasping with deep learning and large-scale data collection. Int. J. Robotics Res., 37(4-5):421-436, 2018. +[41] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023. +[42] Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, and Michael S. Ryoo. Llara: Supercharging robot learning data for vision-language policy. arXiv preprint arXiv:2406.20095, 2024. +[43] Aishan Liu, Xianglong Liu, Jiaxin Fan, Yuqing Ma, Anlan Zhang, Huiyuan Xie, and Dacheng Tao. Perceptual-sensitive GAN for generating adversarial patches. In AAAI, 2019. +[44] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2023. +[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024. +[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurlPS, 2024. +[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019. +[48] Alessandro De Luca, Alin Albu-Schäffer, Sami Haddadin, and Gerd Hirzinger. Collision detection and safe reaction with the DLR-III lightweight manipulator arm. In IROS, 2006. +[49] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018. +[50] Jim Mainprice and Dmitry Berenson. Human-robot collaborative manipulation planning using early prediction of human motion. In IROS, 2013. +[51] Matthew T. Mason. Mechanics of Robotic Manipulation. MIT Press, 2001. +[52] Matthew T. Mason. Toward robotic manipulation. Annu. Rev. Control. Robotics Auton. Syst., 1:1-28, 2018. +[53] Oier Mees, Lukás Hermann, and Wolfram Burgard. What matters in language conditioned robotic imitation learning over unstructured data. IEEE Robotics Autom. Lett., 7(4): 11205-11212, 2022. +[54] Vivek G Moudgal, Kevin M Passino, and Stephen Yurkovich. Rule-based control for a flexible-link robot. IEEE Trans. on Control Sys. Tech., 2(4):392-405, 1994. +[55] Richard M. Murray, Zexiang Li, and S. Shankar Sastry. A mathematical introduction to robotics manipulation. CRC Press, 1994. +[56] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016. +[57] Vitchyr Pong, Murtaza Dalal, Steven Lin, Ashvin Nair, Shikhar Bahl, and Sergey Levine. Skew-fit: State-covering self-supervised reinforcement learning. In ICML, 2020. +[58] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial + +examples jailbreak aligned large language models. In AAAI, 2024. +[59] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In ACM SIGSAC, 2016. +[60] Michael H. Smith and L. Stephen Coles. Design of a low cost, general purpose robot. In IJCAI, 1973. +[61] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Trans. Evol. Comput., 23(5):828-841, 2019. +[62] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014. +[63] Chen Tang, Ben Abbatematteo, Jiaheng Hu, Rohan Chandra, Roberto Martin-Martin, and Peter Stone. Deep reinforcement learning for robotics: A survey of real-world successes. arXiv preprint arXiv:2408.03539, 2024. +[64] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024. +[65] Florian Tramér. Detecting adversarial examples is (nearly) as hard as classifying them. In ICML, 2022. +[66] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020. +[67] Valeria Villani, Fabio Pini, Francesco Leali, and Cristian Secchi. Survey on human-robot collaboration in industrial settings: Safety, intuitive interfaces and applications. Mechatronics, 55:248-266, 2018. +[68] Christian Vogel, Christoph Walter, and Norbert Elkmann. A projection-based sensor system for safe physical human-robot collaboration. In IROS, 2013. +[69] Homer Rich Walke, Kevin Black, Tony Z. Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata V2: A dataset for robot learning at scale. In CoRL, 2023. +[70] Taowen Wang, Yiyang Liu, James Chenhao Liang, Yiming Cui, Yuning Mao, Shaoliang Nie, Jiahao Liu, Fuli Feng, Zenglin Xu, Cheng Han, et al. $\mathbf{M}^2$ pt: Multimodal prompt tuning for zero-shot instruction learning. In EMNLP, 2024. +[71] Zhibo Wang, Siyan Zheng, Mengkai Song, Qian Wang, Alireza Rahimpour, and Hairong Qi. advpattern: Physical-world attacks on deep person re-identification via adversarially transformable patterns. In ICCV, 2019. +[72] Zixiang Wang, Hao Yan, Zhuoyue Wang, Zhengjia Xu, Zhizhong Wu, and Yining Wang. Research on autonomous robots navigation based on reinforcement learning. In RAIIC, 2024. +[73] Zhijie Wang, Zhehua Zhou, Jiayang Song, Yuheng Huang, Zhan Shu, and Lei Ma. Towards testing and evaluating vision-language-action models for robotic manipulation: An empirical study. arXiv preprint arXiv:2409.12894, 2024. + +[74] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024. +[75] Wikipedia contributors. Finch (film) — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Finch_(film) & oldid=1249481595, 2024. [Online; accessed 5-November-2024]. +[76] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024. +[77] Kaidi Xu, Gaoyuan Zhang, Sijia Liu, Quanfu Fan, Mengshu Sun, Hongge Chen, Pin-Yu Chen, Yanzhi Wang, and Xue Lin. Adversarial t-shirt! evading person detectors in a physical world. In ECCV, 2020. +[78] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018. +[79] Zhiyang Xu, Chao Feng, Rulin Shao, Trevor Ashby, Ying Shen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. Visionplan: Scaling human-labeled tasks in visual instruction tuning. arXiv preprint arXiv:2402.11690, 2024. +[80] Yuchen Zhang and Percy Liang. Defending against whitebox adversarial attacks via randomized discretization. In AISTATS, 2019. +[81] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quanxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, 2022. +[82] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023. +[83] Yuanyang Zhu, Zhi Wang, Chunlin Chen, and Daoyi Dong. Rule-based reinforcement learning for efficient robot navigation with space reduction. IEEE/ASME Trans. on Mechatronics, 27:846-857, 2021. +[84] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023. \ No newline at end of file diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/images.zip b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..7b705c37fc9be17bc31a473ffd13346f950e2bc5 --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ccfc007273fcd8175512cb510af46833a81a6aab0be0dc07f0d9d5b7f8a092e4 +size 742940 diff --git a/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/layout.json b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..1e3049fa40888fc6dee99821318bc6db0b603c07 --- /dev/null +++ b/2025/Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics/layout.json @@ -0,0 +1,10834 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 69, + 102, + 542, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 69, + 102, + 542, + 138 + ], + "spans": [ + { + "bbox": [ + 69, + 102, + 542, + 138 + ], + "type": "text", + "content": "Exploring the Adversarial Vulnerabilities of Vision-Language-Action Models in Robotics" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "spans": [ + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": "Taowen Wang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", Cheng Han" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", James Liang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{3*}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", \nWenhao Yang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", Dongfang Liu" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", Luna Xinyu Zhang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{1\\dagger}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", Qifan Wang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", \nJiebo Luo" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "text", + "content": ", Ruixiang Tang" + }, + { + "bbox": [ + 133, + 160, + 477, + 205 + ], + "type": "inline_equation", + "content": "^{7\\dagger}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "spans": [ + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "Rochester Institute of Technology, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "University of Missouri - Kansas City, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{3}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "U.S. Naval Research Laboratory, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{4}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "Lamar University, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{5}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "Meta AI, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{6}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "University of Rochester, " + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "inline_equation", + "content": "^{7}" + }, + { + "bbox": [ + 127, + 217, + 484, + 260 + ], + "type": "text", + "content": "Rutgers University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "spans": [ + { + "bbox": [ + 151, + 286, + 200, + 300 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "content": "Recently in robotics, Vision-Language-Action (VLA) models have emerged as a transformative approach, enabling robots to execute complex tasks by integrating visual and linguistic inputs within an end-to-end learning framework. Despite their significant capabilities, VLA models introduce new attack surfaces. This paper systematically evaluates their robustness. Recognizing the unique demands of robotic execution, our attack objectives target the inherent spatial and functional characteristics of robotic systems. In particular, we introduce two untargeted attack objectives that leverage spatial foundations to destabilize robotic actions, and a targeted attack objective that manipulates the robotic trajectory. Additionally, we design an adversarial patch generation approach that places a small, colorful patch within the camera's view, effectively executing the attack in both digital and physical environments. Our evaluation reveals a marked degradation in task success rates, with up to a " + }, + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 55, + 312, + 297, + 601 + ], + "type": "text", + "content": " reduction across a suite of simulated robotic tasks, highlighting critical security gaps in current VLA architectures. By unveiling these vulnerabilities and proposing actionable evaluation metrics, we advance both the understanding and enhancement of safety for VLA-based robotic systems, underscoring the necessity for continuously developing robust defense strategies prior to physical-world deployments1." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 135, + 625 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 64, + 629, + 293, + 652 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 64, + 629, + 293, + 652 + ], + "spans": [ + { + "bbox": [ + 64, + 629, + 293, + 652 + ], + "type": "text", + "content": "\"First directive: A robot cannot harm a human or, through inaction, allow a human to be harmed.\"" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 236, + 654, + 293, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 236, + 654, + 293, + 666 + ], + "spans": [ + { + "bbox": [ + 236, + 654, + 293, + 666 + ], + "type": "text", + "content": "—Finch [75]" + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 318, + 287, + 549, + 421 + ], + "blocks": [ + { + "bbox": [ + 318, + 287, + 549, + 421 + ], + "lines": [ + { + "bbox": [ + 318, + 287, + 549, + 421 + ], + "spans": [ + { + "bbox": [ + 318, + 287, + 549, + 421 + ], + "type": "image", + "image_path": "a27d1366dc9413109af42625ca8f73a705e6f2d57e087ea6e0769e728e603a08.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 319, + 423, + 431, + 506 + ], + "blocks": [ + { + "bbox": [ + 319, + 423, + 431, + 506 + ], + "lines": [ + { + "bbox": [ + 319, + 423, + 431, + 506 + ], + "spans": [ + { + "bbox": [ + 319, + 423, + 431, + 506 + ], + "type": "image", + "image_path": "a7f1a9cd58f23d1dfa2cd5f851f69b1c8dae7758bd7a5bcc21a17523d63813b5.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 565 + ], + "type": "text", + "content": "Figure 1. Adversarial Vulnerabilities induced by malicious manipulation. (A). Illustration of adversarial threats in robotic task execution. (B). Example of semantic-rich adversarial patches generated by proposed methods. (C). Comparison of failure rates across different attack schemes (UADA, UPA, and TMA)." + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 434, + 422, + 549, + 506 + ], + "blocks": [ + { + "bbox": [ + 434, + 422, + 549, + 506 + ], + "lines": [ + { + "bbox": [ + 434, + 422, + 549, + 506 + ], + "spans": [ + { + "bbox": [ + 434, + 422, + 549, + 506 + ], + "type": "image", + "image_path": "08f45c5ade4c1077e04a483cdb86841d235414718c7f84a3b13bde22084bdd57.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "In the movie Finch, set in a post-apocalyptic future, an intelligent robot navigates complex themes, underscoring the importance of interactive safety with its human master. Once speculative, this portrayal is now increasingly plausible with the rise of Large Vision Language Models (LVLMs) [70, 79], capable of seamlessly interpreting both visual and linguistic contexts. A notable realization of this potential can be seen in Vision-Language-Action (VLA) models [16, 34, 73, 74], which integrate LVLMs into robotic systems to enable end-to-end learning that encompasses both high-level trajectory planning and low-level robot action control. While VLA models demonstrate a promising step toward generalist" + } + ] + } + ], + "index": 14 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 673, + 135, + 683 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 673, + 135, + 683 + ], + "spans": [ + { + "bbox": [ + 66, + 673, + 135, + 683 + ], + "type": "text", + "content": "*Equal contribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 67, + 683, + 144, + 693 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 683, + 144, + 693 + ], + "spans": [ + { + "bbox": [ + 67, + 683, + 144, + 693 + ], + "type": "text", + "content": "† Corresponding author." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 67, + 693, + 256, + 703 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 693, + 256, + 703 + ], + "spans": [ + { + "bbox": [ + 67, + 693, + 256, + 703 + ], + "type": "inline_equation", + "content": "\\ddagger" + }, + { + "bbox": [ + 67, + 693, + 256, + 703 + ], + "type": "text", + "content": "Work done during Luna Xinyu Zhang's internship at RIT." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 67, + 703, + 132, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 132, + 713 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 132, + 713 + ], + "type": "text", + "content": "1 Project homepage." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6948" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 133 + ], + "type": "text", + "content": "robotic intelligence, they also introduce potential vulnerabilities that remain largely underexplored. This work, in light of this view, pioneers the understanding of the pressing need to investigate the new attack surfaces associated with VLA-based robotic systems." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "spans": [ + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": "To gain deeper insights into these vulnerabilities, we analyze VLA-based models and general robotic operations, highlighting two key characteristics crucial for designing effective adversarial attacks. First, creating attack objectives for robotic systems requires consideration of the physical dynamics and constraints intrinsic to robotic movement. Conventional adversarial attacks, in this case, often fail to produce significant deviations in intended actions because they disregard these constraints. Second, VLA models generate control signals that function as a time series of " + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "inline_equation", + "content": "K" + }, + { + "bbox": [ + 55, + 134, + 295, + 314 + ], + "type": "text", + "content": "-class predictions (see §3.1), making it essential to design attacks that exploit the temporal dependencies within these sequences to cause substantial disruptions in robotic behavior. Achieving attacks that are effective in both digital simulations and real-world environments remains critical yet challenging." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 295, + 544 + ], + "type": "text", + "content": "To address these challenges, our work intensifies the adversarial threats posed to VLA-based systems by both developing specialized attack objectives and designing effective attack methods. Specifically, we formulate an Action Discrepancy Objective aimed at maximizing the action discrepancy within VLA-based robotic systems. This strategy ensures that, at each decision point, the robot's behavior can diverge from the optimal trajectory. Additionally, we introduce Geometry-Aware Objective that considers the robot's movement in three-dimensional space, characterized by three degrees of freedom. By optimizing the cosine similarity between adversarial and ground-truth directions, we induce deviations in the robot's movement direction from its intended path, increasing the likelihood of task failure. To achieve these attack objectives, we develop straightforward yet effective Patch-Based Attacks targeting VLA-based robotic systems. This approach enables adversarial attacks in both digital and physical settings, revealing substantial vulnerabilities within the VLA-based system." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 546, + 295, + 713 + ], + "type": "text", + "content": "Altogether, these innovations yield several significant contributions: ① This work presents a pioneering and comprehensive analysis of vulnerabilities within VLA-based robotic systems, a new paradigm for training generalized robotic policies using generative foundation models. We reveal critical threats to adversarial attacks, emphasizing the urgent need to enhance robustness before real-world deployment; ② To the best of our knowledge, we are the first work to define attack objectives specific to the powerful VLA models and to employ a straightforward adversarial patch against it (see §3). This offers valuable insights for the research community to explore systemic failures in similar concurrent generative foundation models; ③ We rigorously evaluate our approach in both simulated and real-world environments across four" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": "distinct robotic tasks, observing significant raises in task failure rates of " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": ", respectively. This highlights the effectiveness of our attack strategies (see " + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "inline_equation", + "content": "\\S 4.3" + }, + { + "bbox": [ + 313, + 72, + 553, + 109 + ], + "type": "text", + "content": ")." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 314, + 114, + 399, + 125 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 114, + 399, + 125 + ], + "spans": [ + { + "bbox": [ + 314, + 114, + 399, + 125 + ], + "type": "text", + "content": "2. Related work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 133, + 466, + 145 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 133, + 466, + 145 + ], + "spans": [ + { + "bbox": [ + 314, + 133, + 466, + 145 + ], + "type": "text", + "content": "2.1. AI-driven Generalist Robot" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 150, + 555, + 269 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 150, + 555, + 269 + ], + "spans": [ + { + "bbox": [ + 313, + 150, + 555, + 269 + ], + "type": "text", + "content": "Developing generalist robots [6, 14, 27, 37, 60] requires models to not only handle varied interactions but also maintain robustness in unstructured environments. Early intelligent robots [51, 52, 54, 55, 83] generally relied on rule-based approaches, effective only in controlled settings and needing extensive reprogramming for new tasks. The advent of deep learning [19, 31, 40, 56] and reinforcement learning [1, 63, 72] has shifted this paradigm to adapt through the data-driven pipeline, increasing robot versatility and reducing manual reconfiguration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 270, + 556, + 425 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 556, + 425 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 556, + 425 + ], + "type": "text", + "content": "Recent studies [7, 53, 57] have further advanced the development and potential of Large Vision Language Models (LVLMs) as key enablers for generalist robots, demonstrating promising generalization across a variety of scenes [15, 25, 32, 41, 64, 84]. One of the notable examples is OpenVLA [34], which integrates dual visual encoders with a pre-trained large language model to enable observation-to-action capabilities through instruction-following [10, 33, 45, 46, 70, 82]. By aligning visual and textual semantics, OpenVLA facilitates complex scenario understanding and end-to-end action generation, showing impressive generalization when executing previously unseen task instructions." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 433, + 481, + 445 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 433, + 481, + 445 + ], + "spans": [ + { + "bbox": [ + 314, + 433, + 481, + 445 + ], + "type": "text", + "content": "2.2. Adversarial Attacks in Robotic" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "spans": [ + { + "bbox": [ + 313, + 450, + 556, + 617 + ], + "type": "text", + "content": "Adversarial attacks serve as critical tools for assessing model vulnerabilities, particularly in robotics, where models operate in dynamic, real-world settings. Traditional gradient-based, pixel-level attacks [17, 21, 43, 49, 61, 71] exploit model gradients to calculate malicious perturbations, achieving high attack success rates in digital environments. For physical-world attacks, patch-based methods [3, 8, 59, 76, 77] offer a practical alternative for real-world applications by introducing physically realizable perturbations that retain efficacy under varied conditions, such as angle or lighting changes, making them highly applicable for robotic systems. Therefore, considering VLA-based models enable robot execution in real-world scenarios, this work employs a patch-based attack strategy." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 618, + 556, + 713 + ], + "type": "text", + "content": "VLA models [13, 26, 64] couple visual and linguistic modalities for perception-action alignment. The continuous, high-dimensional nature of visual inputs [49, 58, 62] makes them the most likely target for adversarial perturbations, as attackers can subtly alter visual data in ways that are difficult to detect [2, 9, 65]. Consequently, this work applies attacks on the visual modality to achieve robust manipulations across diverse environments. To the best of our knowledge, this" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6949" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 73, + 541, + 247 + ], + "blocks": [ + { + "bbox": [ + 69, + 73, + 541, + 247 + ], + "lines": [ + { + "bbox": [ + 69, + 73, + 541, + 247 + ], + "spans": [ + { + "bbox": [ + 69, + 73, + 541, + 247 + ], + "type": "image", + "image_path": "c1f482ac55f606457a94bb97d63f25131deeaa895b2426f984ac68d6b5d941b2.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 249, + 555, + 304 + ], + "lines": [ + { + "bbox": [ + 55, + 249, + 555, + 304 + ], + "spans": [ + { + "bbox": [ + 55, + 249, + 555, + 304 + ], + "type": "text", + "content": "Figure 2. Overall Adversarial Framework. The robot captures an input image, processes it through a vision-language model to generate tokens representing actions, and then uses an action de-tokenizer for discrete bin prediction. The model is optimized with adversarial objectives focusing on various discrepancies and geometries (i.e., UADA, UPA, TMA). Forward propagation is shown in black, and backpropagation is highlighted in pink. These objectives aim to maximize errors and minimize task performance, with visual emphasis on 3D-space manipulation and a focus on generating adversarial perturbation " + }, + { + "bbox": [ + 55, + 249, + 555, + 304 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 249, + 555, + 304 + ], + "type": "text", + "content": " during task execution, such as picking up a can." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "spans": [ + { + "bbox": [ + 55, + 309, + 296, + 369 + ], + "type": "text", + "content": "study is one of the earliest attempts to formally define adversarial objectives for VLA models, targeting gaps in adversarial robustness through the model's geometric constraints (e.g., spatial dependencies in visual input) and architectural nuances (e.g., cross-modal information integration)." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 377, + 138, + 390 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 377, + 138, + 390 + ], + "spans": [ + { + "bbox": [ + 55, + 377, + 138, + 390 + ], + "type": "text", + "content": "3. Methodology" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "spans": [ + { + "bbox": [ + 55, + 398, + 296, + 518 + ], + "type": "text", + "content": "In this section, we first review the core algorithmic principles of VLA in §3.1, which serves as the foundation for the adversarial scheme designs. We then present two types of untargeted adversarial attacks, including Untargeted Action Discrepancy Attack (UADA) in §3.2 and Untargeted Position-aware Attack (UPA) in §3.3. In §3.4, we further include Targeted Manipulation Attack (TMA), designed to direct the robot toward specific erroneous execution. Finally, we introduce the Normalized Action Discrepancy (NAD) metric in §3.5. The overall framework is shown in Fig. 2." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "spans": [ + { + "bbox": [ + 55, + 528, + 135, + 540 + ], + "type": "text", + "content": "3.1. Preliminary" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": "Vision-Language-Action models [34, 84] (see Fig. 2) are built on large language models integrated with visual encoders, enabling robots to interpret human instructions and process visual input from a camera to perform context-aware actions. VLA models [34, 42, 84] generally abstract continuous action predictions into a classification problem by discretizing robot actions. Specifically, they first discretize the continuous probabilities into class labels " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "y = \\arg \\max_{x} \\mathcal{F}(x)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " where " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(\\cdot)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": " is the VLA model. An action de-tokenizer then generates actions " + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "inline_equation", + "content": "A = DT(y)" + }, + { + "bbox": [ + 55, + 545, + 296, + 713 + ], + "type": "text", + "content": ". By categorizing action values into discrete class labels, the model converts continuous probability outputs into discrete signals, this simplification facilitates quicker convergence and faster training times, and is commonly used for VLA-based models [34, 42, 84]." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "spans": [ + { + "bbox": [ + 313, + 309, + 555, + 357 + ], + "type": "text", + "content": "In this work, we focus on a 7 degree-of-freedoms (DoFs) robotic arm [23]. At each step, an action consists of 7 DoFs with specific physical significance in three-dimensional Cartesian space, represented by:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 337, + 365, + 553, + 377 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 337, + 365, + 553, + 377 + ], + "spans": [ + { + "bbox": [ + 337, + 365, + 553, + 377 + ], + "type": "interline_equation", + "content": "A = \\left[ \\Delta P _ {x}, \\Delta P _ {y}, \\Delta P _ {z}, \\Delta R _ {x}, \\Delta R _ {y}, \\Delta R _ {z}, \\text {g r i p p e r} \\right], \\tag {1}", + "image_path": "4a6ddfe80a8e3de328a4d22f8e88957d853c043c172037026a36555bf2604e22.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "spans": [ + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\Delta P_{x,y,z}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\Delta R_{x,y,z}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " denote relative positional and rotational changes along the x-, y-, and z-axis, discretized into 256 uniform bins across each DoFs' action bounds " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "[y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i]" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": " [34]. The gripper state is binary, indicating its open or closed state. This control design presents a unique challenge for adversarial attacks, as finely divided bins result in minimal action discrepancies between neighboring bins (e.g., " + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "inline_equation", + "content": "\\pm 0.007/\\mathrm{bin}" + }, + { + "bbox": [ + 313, + 385, + 555, + 505 + ], + "type": "text", + "content": "). This means that even if an attack shifts the classification to an adjacent bin, the resulting action discrepancy has minimal impact on real-world performance." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 513, + 518, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 513, + 518, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 513, + 518, + 525 + ], + "type": "text", + "content": "3.2. Untargeted Action Discrepancy Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": "To exacerbate action discrepancies, we introduce the Untargeted Action Discrepancy Attack (UADA), which aims to maximize deviations in robot actions. This attack is based on the observation that larger robot actions usually correlate with intense physical movements, which, in turn, may amplify the potential for real-world hazards [28-30]. For UADA, the attack target is one or a combination of 7 DoFs, defined as " + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "inline_equation", + "content": "y_{gt}^{I} = \\{y_{gt}^{i}|i\\in [1,\\dots ,7]\\}" + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": ". To define UADA's objective, we first identify the most distant action " + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "inline_equation", + "content": "y_{adv}^{i}" + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": ", which maximizes the discrepancy from the " + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": "-th DoF ground truth action " + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "inline_equation", + "content": "y^{i}" + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": ". This action, for each DoF, corresponds to its own action bound " + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "inline_equation", + "content": "[y_{\\min}^{i},y_{\\max}^{i}]" + }, + { + "bbox": [ + 313, + 530, + 556, + 676 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 326, + 684, + 555, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 684, + 555, + 717 + ], + "spans": [ + { + "bbox": [ + 326, + 684, + 555, + 717 + ], + "type": "interline_equation", + "content": "y _ {a d v} ^ {i} = \\left\\{ \\begin{array}{l l} y _ {\\max } ^ {i}, & \\text {i f} | y _ {\\max } ^ {i} - y _ {g t} ^ {i} | \\geq | y _ {\\min } ^ {i} - y _ {g t} ^ {i} | \\\\ y _ {\\min } ^ {i}, & \\text {o t h e r w i s e} \\end{array} . \\right. \\tag {2}", + "image_path": "a4555a27fc84afcefc844b2c1352cdf3f5ad9c7a64fc4337f2715a9532886388.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6950" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": "Instead of directly using " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "y_{adv}^{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " as the misclassification target, we introduce a soft attack objective to capture the discrepancy between actions, ensuring smooth gradient optimization and stable attack performance. Specifically, due to the physical action magnitude information contains in bin labels, we reweight the output probability " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\mathcal{F}(x)_{bins}^{i} \\in \\mathcal{R}^{1 \\times bins}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " using normalized bin labels " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "y_{bins}^{i} = [\\frac{1}{bins}, \\dots, 1]" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\lfloor y_{bins}^{i} \\rfloor" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "\\lceil y_{bins}^{i} \\rceil" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " correspond to the normalized values of " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "y_{\\min}^{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "inline_equation", + "content": "y_{\\max}^{i}" + }, + { + "bbox": [ + 55, + 72, + 296, + 180 + ], + "type": "text", + "content": ", respectively. The reweighting process is defined as:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 97, + 190, + 295, + 222 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 97, + 190, + 295, + 222 + ], + "spans": [ + { + "bbox": [ + 97, + 190, + 295, + 222 + ], + "type": "interline_equation", + "content": "y _ {s o f t} ^ {i} = \\sum_ {b i n s = 1} ^ {n} \\mathcal {F} (x + \\delta) _ {b i n s} ^ {i} \\otimes y _ {b i n s} ^ {i}, \\tag {3}", + "image_path": "76887832c82837ddb418f44c9e7ee93bb286448044a16ef1eedc5f1762fc664b.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "spans": [ + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\otimes" + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "content": " denotes Hadamard Product, " + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "inline_equation", + "content": "\\delta" + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "content": " is the adversarial patch and " + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "inline_equation", + "content": "y_{\\text{soft}}^i" + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "content": " represents the " + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 55, + 229, + 296, + 277 + ], + "type": "text", + "content": "-th DoF's soft action. Finally, the objective of UADA is to minimize the discrepancy between the soft and the most distant actions, which is:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 94, + 286, + 295, + 319 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 286, + 295, + 319 + ], + "spans": [ + { + "bbox": [ + 94, + 286, + 295, + 319 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U A D A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\sum_ {i} ^ {I} \\left(y _ {s o f t} ^ {i} - y _ {a d v} ^ {i}\\right) ^ {2}, \\tag {4}", + "image_path": "8a834a84e42e59be0f65b8759382fa7930293ea7e1f3e1527682de90fa09977c.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 327, + 296, + 364 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 364 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 364 + ], + "type": "text", + "content": "UADA defines the objective function " + }, + { + "bbox": [ + 55, + 327, + 296, + 364 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{UADA}}" + }, + { + "bbox": [ + 55, + 327, + 296, + 364 + ], + "type": "text", + "content": " tailored for attacking the robot's action space, allowing adversarial patches to create significant, lasting disruptions in task performance." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 371, + 238, + 384 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 371, + 238, + 384 + ], + "spans": [ + { + "bbox": [ + 55, + 371, + 238, + 384 + ], + "type": "text", + "content": "3.3. Untargeted Position-aware Attack" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 389, + 296, + 472 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 389, + 296, + 472 + ], + "spans": [ + { + "bbox": [ + 55, + 389, + 296, + 472 + ], + "type": "text", + "content": "We explore untargeted adversarial attacks in parallel by considering the positional dynamics within VLA models. In typical task execution, precise and directed movements of the end-effector towards designated goals are essential, with actions mapped in three-dimensional space. The positional DoFs, " + }, + { + "bbox": [ + 55, + 389, + 296, + 472 + ], + "type": "inline_equation", + "content": "A^p = [\\Delta P_x, \\Delta P_y, \\Delta P_z]" + }, + { + "bbox": [ + 55, + 389, + 296, + 472 + ], + "type": "text", + "content": " encapsulates the directional movements within 3D for effective robotic control." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "text", + "content": "Recognizing the importance of " + }, + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "inline_equation", + "content": "A^p = DT(y^p)" + }, + { + "bbox": [ + 55, + 473, + 296, + 581 + ], + "type": "text", + "content": " in controlling the end-effector's path, we introduce a position-aware attack to disrupt the intended movement trajectory. This type of adversarial vulnerability remains largely unexplored but has significant implications for task failure: the attack objective can steer the end-effector away from its intended path by introducing directional perturbations, amplifying errors and causing substantial trajectory distortions. Formally, we define the Untargeted Position-aware Attack (UPA) objective as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 67, + 597, + 295, + 623 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 597, + 295, + 623 + ], + "spans": [ + { + "bbox": [ + 67, + 597, + 295, + 623 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U P A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ \\alpha \\frac {y _ {a d v} ^ {p} \\cdot y ^ {p}}{\\| y _ {a d v} ^ {p} \\| \\| y ^ {p} \\|} + \\beta \\frac {1}{\\| y _ {a d v} ^ {p} - y ^ {p} \\| _ {2}} \\right], \\tag {5}", + "image_path": "d90cbb7b0a1d4bc118f83c95fc47a06692fa293abe025c21ad433c2bf12eb224.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " are the hyperparameters that balance between the directionality and magnitude of the perturbations. By integrating geometric awareness into our attack, this approach generates perturbations that induce cumulative deviations in the robot's trajectory, even with minimal adjustments. " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "L_{2}" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": "-normalization " + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "inline_equation", + "content": "\\| \\cdot \\| _2" + }, + { + "bbox": [ + 55, + 629, + 296, + 713 + ], + "type": "text", + "content": " further intensifies these deviations, rendering the attack highly effective and disruptive." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 314, + 72, + 480, + 84 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 72, + 480, + 84 + ], + "spans": [ + { + "bbox": [ + 314, + 72, + 480, + 84 + ], + "type": "text", + "content": "3.4. Targeted Manipulation Attack" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "spans": [ + { + "bbox": [ + 313, + 89, + 555, + 173 + ], + "type": "text", + "content": "In addition to the aforementioned untargeted attacks, we also explore the adversarial vulnerabilities of VLA models with a targeted strategy. This attack aims to directly mislead the models into predicting specific trajectories, resulting in precise alterations. We design this attack since the targeted alterations can drive malicious behaviors or induce task failure. The objective of the targeted manipulation attack is:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 356, + 177, + 555, + 192 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 356, + 177, + 555, + 192 + ], + "spans": [ + { + "bbox": [ + 356, + 177, + 555, + 192 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {T M A}} = \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {T} ^ {I}\\right) \\right], \\tag {6}", + "image_path": "08d6dfb52a18c070b9322f79d8f14785cda36da8f62a7d03904a20521d75487e.jpg" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "spans": [ + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "inline_equation", + "content": "y_{T}^{I} = \\{y_{T}^{i} = t|i\\in [1,\\dots ,7],t\\in [y_{\\min}^{i},y_{\\max}^{i}] \\}" + }, + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "text", + "content": " represents the adversarial target(s); and " + }, + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "inline_equation", + "content": "CE(\\cdot ,\\cdot)" + }, + { + "bbox": [ + 313, + 196, + 556, + 342 + ], + "type": "text", + "content": " denotes the Cross-Entropy loss, which measures the error between the predicted state under adversarial perturbation and the desired target state. By steering the robot toward an adversarial target across successive time steps, our approach manipulates the trajectory and undermines task performance, ultimately leading to a substantial disruption of intended operations. Furthermore, manipulating any individual DoF or a combination thereof can also severely compromise task success, given that the errors in one dimension tend to propagate and magnify over time during execution (see Fig. 3)." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 347, + 525, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 347, + 525, + 361 + ], + "spans": [ + { + "bbox": [ + 313, + 347, + 525, + 361 + ], + "type": "text", + "content": "3.5. Metric: Normalized Action Discrepancy" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "spans": [ + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "content": "To assess the potential physical impact of adversarial patches, we introduce Normalized Action Discrepancy (NAD) as a fine-grained metric for evaluating discrepancies at the action level. NAD serves as a complementary measurement alongside the coarse-grained Failure Rate (FR) at the task level, providing a detailed analysis of deviations during action execution. To define the NAD, we first determine the maximum action discrepancy " + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{max}}^i" + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "content": ", which represents the largest deviation that can occur within the allowed range for " + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "content": "-th DoF. This is computed by measuring the L1 distance between the ground truth " + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "inline_equation", + "content": "y_{gt}^i" + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "content": " and its action bound " + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "inline_equation", + "content": "[y_{\\mathrm{min}}^i, y_{\\mathrm{max}}^i]" + }, + { + "bbox": [ + 313, + 365, + 556, + 498 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 322, + 504, + 555, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 322, + 504, + 555, + 519 + ], + "spans": [ + { + "bbox": [ + 322, + 504, + 555, + 519 + ], + "type": "interline_equation", + "content": "d _ {\\mathrm {m a x}} ^ {i} (y) = \\max \\left[ D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m i n}} ^ {i} | \\big), D T \\big (| y _ {g t} ^ {i} - y _ {\\mathrm {m a x}} ^ {i} | \\big) \\right]. (7)", + "image_path": "8938b0c0b1eba641d2afe0998d1bd1c57100f1accafc9610794412c6214620d4.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "spans": [ + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "text", + "content": "We next calculate the applied action discrepancy " + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{applied}}^i (x,y)" + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "inline_equation", + "content": "= |DT(\\mathcal{F}(x)^i) - DT(y_{gt}^i)|" + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "text", + "content": " to measure the deviation between the model's output and ground truth. We define the NAD as the drifting ratio between " + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{applied}}^i" + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "inline_equation", + "content": "d_{\\mathrm{max}}^i" + }, + { + "bbox": [ + 313, + 523, + 555, + 575 + ], + "type": "text", + "content": " as:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 386, + 582, + 555, + 613 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 386, + 582, + 555, + 613 + ], + "spans": [ + { + "bbox": [ + 386, + 582, + 555, + 613 + ], + "type": "interline_equation", + "content": "\\mathrm {N A D} = \\frac {1}{\\mathrm {I}} \\sum_ {\\mathrm {i}} ^ {\\mathrm {I}} \\frac {\\mathrm {d} _ {\\text {a p p l i e d}} ^ {\\mathrm {i}}}{\\mathrm {d} _ {\\max } ^ {\\mathrm {i}}}, \\tag {8}", + "image_path": "21aa2035d49d292f6a2a47307f01d52f8c3a4e6914c8a387198c2324f5e9cc28.jpg" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 621, + 556, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 621, + 556, + 670 + ], + "spans": [ + { + "bbox": [ + 313, + 621, + 556, + 670 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 621, + 556, + 670 + ], + "type": "inline_equation", + "content": "I \\in [1, \\dots, 7]" + }, + { + "bbox": [ + 313, + 621, + 556, + 670 + ], + "type": "text", + "content": " represents a single or a combination of DoF(s). In this way, NAD provides normalized measurement of action deviations, allowing for a consistent evaluation of adversarial impact across varying DoF(s)." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 674, + 394, + 687 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 674, + 394, + 687 + ], + "spans": [ + { + "bbox": [ + 313, + 674, + 394, + 687 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 713 + ], + "type": "text", + "content": "In this section, we first detail the implementation of our adversarial framework and baseline methods in §4.1. Next, we" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "6951" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 99, + 90, + 361, + 193 + ], + "blocks": [ + { + "bbox": [ + 99, + 90, + 361, + 193 + ], + "lines": [ + { + "bbox": [ + 99, + 90, + 361, + 193 + ], + "spans": [ + { + "bbox": [ + 99, + 90, + 361, + 193 + ], + "type": "image", + "image_path": "053d65c210e902c1eb3f997270755a01597bfa93840980fb3420e9367d440e72.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 368, + 89, + 496, + 194 + ], + "blocks": [ + { + "bbox": [ + 102, + 74, + 505, + 86 + ], + "lines": [ + { + "bbox": [ + 102, + 74, + 505, + 86 + ], + "spans": [ + { + "bbox": [ + 102, + 74, + 505, + 86 + ], + "type": "text", + "content": "TASK (UADA) : Pick up the black bowl between the plate and the ramekin and place it on the plate" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 368, + 89, + 496, + 194 + ], + "lines": [ + { + "bbox": [ + 368, + 89, + 496, + 194 + ], + "spans": [ + { + "bbox": [ + 368, + 89, + 496, + 194 + ], + "type": "image", + "image_path": "56b3f6c182dadf720ec5078e7ad1dc0489d0d62a8e2794270f803036b9f8ec51.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 97, + 210, + 361, + 315 + ], + "blocks": [ + { + "bbox": [ + 148, + 196, + 462, + 209 + ], + "lines": [ + { + "bbox": [ + 148, + 196, + 462, + 209 + ], + "spans": [ + { + "bbox": [ + 148, + 196, + 462, + 209 + ], + "type": "text", + "content": "TASK (UPA) : Pick up the black bowl on the ramekin and place it on the plate" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 97, + 210, + 361, + 315 + ], + "lines": [ + { + "bbox": [ + 97, + 210, + 361, + 315 + ], + "spans": [ + { + "bbox": [ + 97, + 210, + 361, + 315 + ], + "type": "image", + "image_path": "e568da881f28597e9b6b015d091a52624fbabdb43f3356c6d1cf874c6c3ffe94.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 370, + 211, + 496, + 315 + ], + "blocks": [ + { + "bbox": [ + 370, + 211, + 496, + 315 + ], + "lines": [ + { + "bbox": [ + 370, + 211, + 496, + 315 + ], + "spans": [ + { + "bbox": [ + 370, + 211, + 496, + 315 + ], + "type": "image", + "image_path": "17fa09dc76817dc102176fa4d847ebf1f5c9b294e44ed926ac405b8cb258b33b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 99, + 332, + 361, + 437 + ], + "blocks": [ + { + "bbox": [ + 174, + 319, + 436, + 331 + ], + "lines": [ + { + "bbox": [ + 174, + 319, + 436, + 331 + ], + "spans": [ + { + "bbox": [ + 174, + 319, + 436, + 331 + ], + "type": "text", + "content": "TASK (TMA) : Pick up the alphabet soup and place it in the bask" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 99, + 332, + 361, + 437 + ], + "lines": [ + { + "bbox": [ + 99, + 332, + 361, + 437 + ], + "spans": [ + { + "bbox": [ + 99, + 332, + 361, + 437 + ], + "type": "image", + "image_path": "68bbb29a18557a64d22e434219a261b2da93f2cdbfab928084518b2d8d9efed1.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 372, + 333, + 497, + 439 + ], + "blocks": [ + { + "bbox": [ + 372, + 333, + 497, + 439 + ], + "lines": [ + { + "bbox": [ + 372, + 333, + 497, + 439 + ], + "spans": [ + { + "bbox": [ + 372, + 333, + 497, + 439 + ], + "type": "image", + "image_path": "aeae4fff3aed917ce687bb61bf292a422d815551dcb583cffeda6f27f537b83e.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "lines": [ + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "spans": [ + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": "Figure 3. Qualitative Results of adversarial vulnerabilities over OpenVLA-7B [34] and OpenVLA-7B-LIBERO [34] with objectives of UADA, UPA, and TMA, respectively. We visualize the overall 3D trajectories and 2D trajectories of benign " + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": " and adversarial " + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": " scenarios at each time step to compare the impact of the generated adversarial patch in affecting them. The untargeted trajectory " + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\bullet" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": " is visualized in UADA task. All trajectories start with " + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\triangle" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": ", and we plot the success end point, marked as " + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "inline_equation", + "content": "\\star" + }, + { + "bbox": [ + 55, + 443, + 555, + 487 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "spans": [ + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "text", + "content": "outline the experimental setup in " + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "inline_equation", + "content": "\\S 4.2" + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "text", + "content": " and present quantitative and qualitative results in both simulation and physicalworld scenarios (§4.3). We then conduct diagnostic experiments (§4.4) to analyze the impact of key components. Additionally, we assess the robustness of our method against various defense mechanisms in " + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "inline_equation", + "content": "\\S 4.5" + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "text", + "content": " and conclude with a comprehensive systemic discussion in " + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "inline_equation", + "content": "\\S 4.6" + }, + { + "bbox": [ + 55, + 494, + 297, + 578 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 585, + 188, + 597 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 585, + 188, + 597 + ], + "spans": [ + { + "bbox": [ + 55, + 585, + 188, + 597 + ], + "type": "text", + "content": "4.1. Implementation Details" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": "The implementation of our adversarial patch attack pipeline is detailed in Algorithm 1. We incorporate two key modifications to enhance the training stability of the generated patch. Inner-loops. Following previous work [49], we add " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " inner-loops optimize steps during each iteration in line 4, aiming to reduce data variance and ensure more consistent updates. Geometric Transformations. To improve the robustness of our attack under real-world scenarios, we employ random geometric transformations " + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathcal{T}[(\\cdot), (sh_x, sh_y, \\theta)]" + }, + { + "bbox": [ + 55, + 605, + 296, + 715 + ], + "type": "text", + "content": " in line 6 with" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 321, + 496, + 482, + 507 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 496, + 482, + 507 + ], + "spans": [ + { + "bbox": [ + 321, + 496, + 482, + 507 + ], + "type": "text", + "content": "Algorithm 1: Adversarial Patch Attack." + } + ] + } + ], + "index": 13 + }, + { + "type": "table", + "bbox": [ + 332, + 510, + 553, + 654 + ], + "blocks": [ + { + "bbox": [ + 332, + 510, + 553, + 654 + ], + "lines": [ + { + "bbox": [ + 332, + 510, + 553, + 654 + ], + "spans": [ + { + "bbox": [ + 332, + 510, + 553, + 654 + ], + "type": "table", + "html": "
1:Input: X: dataset; δ: patch; L0: attack objective; \nF: VLA model; T(·): transformations; φ,ψ: \ntransformation parameters; T: attack steps; k: \ninner-loop steps.
2:Initialize: δ ~ U[0,1), L0 ∈ {LUADA, LUPA, LTMA}
3:for i = 1,2,···, T do
4:for i = 1,2,···, k do
5:shx, shy ~ U(-φ,φ), θ ~ U(-ψ,ψ)
6:ypred ← F[T((x + δ), (shx, shy, θ))]
7:δ ← ∂L0/∂δ{▷ Update}
8:end for
9:end for
", + "image_path": "76ee417d291e6b51af26087be62d338fdb778dbe3d37a36813b0ad31817ecf74.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "table_body" + } + ], + "index": 14 + }, + { + "bbox": [ + 314, + 677, + 484, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 677, + 484, + 689 + ], + "spans": [ + { + "bbox": [ + 314, + 677, + 484, + 689 + ], + "type": "text", + "content": "affine and rotation transformations [3] as:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 320, + 692, + 553, + 717 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 692, + 553, + 717 + ], + "spans": [ + { + "bbox": [ + 320, + 692, + 553, + 717 + ], + "type": "interline_equation", + "content": "\\mathcal {T} \\left[ (\\cdot), \\left(s h _ {x}, s h _ {y}, \\theta\\right) \\right] = \\left[ \\begin{array}{c c} 1 & s h _ {y} \\\\ s h _ {x} & 1 \\end{array} \\right] \\cdot \\left[ \\begin{array}{c c} \\cos \\theta & - \\sin \\theta \\\\ \\sin \\theta & \\cos \\theta \\end{array} \\right], \\tag {9}", + "image_path": "c49f312dd65ba2ce4fa50a7d997301e1e05796bfff6cc9f7d99ba6e96c509957.jpg" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6952" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "sh_x, sh_y \\sim \\mathcal{U}(-\\phi, \\phi)" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " are shear factors, and " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "\\theta \\sim \\mathcal{U}(-\\psi, \\psi)" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": " denotes the rotation angle, both sampled from uniform distributions during inner loop optimization step " + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 72, + 296, + 182 + ], + "type": "text", + "content": ". Untargeted Manipulation Attack (UMA) Baseline. As a pioneering work, we could not find a directly comparable baseline for our attacks. Therefore, we adapt prior work in adversarial learning as one of our baseline methods [66]. This attack is designed to mislead classifiers; we integrate it into our framework, leading to the following objective:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 93, + 188, + 295, + 204 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 93, + 188, + 295, + 204 + ], + "spans": [ + { + "bbox": [ + 93, + 188, + 295, + 204 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {U M A}} = - \\mathbb {E} _ {(x, y) \\sim \\mathcal {X}} \\left[ C E \\left(\\mathcal {F} (x + \\delta) ^ {I}, y _ {g t} ^ {I}\\right) \\right]. \\tag {10}", + "image_path": "9b7baafea76146b18461b6a58ddcf75194bdecb6c59df4d83daedd08ded19981.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "spans": [ + { + "bbox": [ + 55, + 211, + 295, + 235 + ], + "type": "text", + "content": "By minimizing the negative CE loss, this attack pushes the predicted action away from the ground truth action." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "text", + "content": "Random Noise Baseline. We generate random noise patches as an additional baseline, representing an unstructured intervention without any learned adversarial intent. Specifically, the noise is sampled from a Gaussian distribution " + }, + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "inline_equation", + "content": "\\mathcal{N}(0,1)" + }, + { + "bbox": [ + 55, + 236, + 296, + 307 + ], + "type": "text", + "content": " and evaluated with the same setting as proposed attacks." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 314, + 164, + 327 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 314, + 164, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 314, + 164, + 327 + ], + "type": "text", + "content": "4.2. Experiment Setup" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 330, + 296, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 330, + 296, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 330, + 296, + 460 + ], + "type": "text", + "content": "Datasets. We conduct attacks on BridgeData V2 [69] and LIBERO [44] with corresponding VLAs. BridgeData V2 [69] is a real-world dataset comprising 24 diverse environments and 13 distinct skills, such as grasping, placing, and object rearrangement, with a total of 60,096 trajectories. LIBERO [44] is a simulation dataset designed to evaluate models across four distinct task types (Spatial, Object, Goal, Long). Notably, LIBERO-Long combines diverse objects, layouts, and long-horizon tasks, making them challenging for complex, multi-step planning, thus we choose LIBERO-Long to conduct UADA, UPA and TMA." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 462, + 296, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 462, + 296, + 676 + ], + "spans": [ + { + "bbox": [ + 55, + 462, + 296, + 676 + ], + "type": "text", + "content": "Victim VLAs. In our study, we select publicly available and current most influential VLAs as victim models for evaluation. Specifically, we focus on four variants of the OpenVLA model, each trained independently on different task suites within the LIBERO dataset [44] (i.e., Spatial, Object, Goal, Long). To evaluate the effectiveness of our methods, we craft adversarial patches using three distinct generating setups: Simulation Setting involves a model trained in a simulated environment using the LIBERO Long suite with the openvla-7B-libero-long variant [34]. Physical Setting uses a model trained on physical world data from the BridgeData v2 [69] with the openvla-7B model [34]. This approach allows us to assess our methods on both simulated and real-world data. Subsequently, we evaluate the performance of generated adversarial patches on victim models (i.e., OpenVLA LIBERO variants) trained on different tasks suites to rigorously prove the robustness and effectiveness of our method. These models are trained with distinct data sources and task objectives." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Robot Setups (Real-world Setting). For the real-world tasks, we adopt a robotic system consisting of a Universal Robots UR10e equipped with a Robotiq Hand-E Gripper to" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 95 + ], + "type": "text", + "content": "provide a 7DoF motion. The sensing system includes one RGB webcam in the fixed position with a shoulder view." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 96, + 555, + 179 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 96, + 555, + 179 + ], + "spans": [ + { + "bbox": [ + 313, + 96, + 555, + 179 + ], + "type": "text", + "content": "Evaluation Details. To assess the effectiveness and robustness of our methods, we conduct experiments on the LIBERO dataset [44]. Each suite consists of 10 tasks, with each task executed for 50 trials, resulting in a total of 500 rollouts, following Kim et al. [34]. For reproducibility, we carefully select patch paste locations for each suite, ensuring they do not obscure objects or robots in the test environment." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 180, + 556, + 346 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 180, + 556, + 346 + ], + "spans": [ + { + "bbox": [ + 313, + 180, + 556, + 346 + ], + "type": "text", + "content": "Evaluation Metric. Regarding the task execution evaluation, we take the maximum steps of each task suite in the LIBERO training dataset as the timeout failure condition to reduce computational overhead. Furthermore, building on the concept of Success Rate (SR) introduced in LIBERO [44], we adopt Failure Rate (FR), defined as " + }, + { + "bbox": [ + 313, + 180, + 556, + 346 + ], + "type": "inline_equation", + "content": "1 - \\mathrm{SR}" + }, + { + "bbox": [ + 313, + 180, + 556, + 346 + ], + "type": "text", + "content": ", as the primary evaluation metric. To further quantify action discrepancy, we employ NAD (see Eq. 8) for untargeted attacks on corresponding DoF(s), as it considers relative distances to quantify deviation from the optimal action. A higher NAD value indicates a greater discrepancy between the ground truth and the predicted action. Additionally, we report the standard deviation of FR across tasks within each task suite." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 314, + 351, + 394, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 351, + 394, + 361 + ], + "spans": [ + { + "bbox": [ + 314, + 351, + 394, + 361 + ], + "type": "text", + "content": "4.3. Main Result" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "spans": [ + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": "Quantitative Results. For UADA and UPA, our methods effectively amplify action discrepancies, leading to a notable transfer attack ability in increasing failure rates (see Tab. 1a). Specifically, while attacking " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_1" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_{1\\sim 3}" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " in the Simulation setup, UADA and UPA achieve NAD of " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "21.0\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "14.5\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ", significantly outperforming UMA scenarios with increments of " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "6.9\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ", respectively. Both UADA and UPA effectively disrupt robot execution, yielding maximum average failure rates of " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "89.7\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ", respectively. Regarding the Physical setup (see Tab. 1b), our attack methods demonstrate strong transferability, as adversarial patches generated in the physical setting significantly impact the performance of LIBERO-X test. Specifically, both UADA and UPA generate malicious actions with large action discrepancies, which are " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_1" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "26.9\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_{1\\sim 3}" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ") with an increase of " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "5.9\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "3.1\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "26.7\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_1" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ") and " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "23.8\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " (" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_{1\\sim 3}" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": ") in UMA scenarios, respectively. Compare the two settings (Tab. 1a v.s. Tab. 1b), we observe a general variance in NAD (e.g. " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "32.6\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " v.s. " + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "inline_equation", + "content": "21.0\\%" + }, + { + "bbox": [ + 313, + 367, + 556, + 676 + ], + "type": "text", + "content": " for UADA). This discrepancy can be attributed to the fundamental differences between the simulation and real-world datasets used for conducting attacks (Bridge V2 [69] v.s. LIBERO [44]). The increased variability in real-world data, including environmental complexity, object diversity, and task difficulty, allows the robot more opportunities to generate larger action discrepancies within the validation dataset." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "For TMA task, we evaluate the effectiveness of our method by manipulating all DoF(s) to 0 (i.e., " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "y_{T}^{i} = 0" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " in Eq. 6). Our method demonstrates significant effectiveness" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6953" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 106, + 302, + 190 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 103 + ], + "type": "text", + "content": "Table 1. Untargeted Results. We report FR and NAD in LIBERO simulation. * denotes the in-domain victim model and dataset aligned with the patch generation model and dataset, △ denotes a transfer attack evaluation involving a distinct victim model and dataset (see §4.2). The FR (↑) is highlighted in best and second best for each task suite." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 106, + 302, + 190 + ], + "lines": [ + { + "bbox": [ + 56, + 106, + 302, + 190 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 302, + 190 + ], + "type": "table", + "html": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long*Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF114.1100±0.0%99.0±3.0%100±0.0%100±0.0%99.8%
DoF1~311.4100±0.0%98.8±2.6%99.0±2.4%100±0.0%99.5%
UADADoF121.0100±0.0%99.2±2.4%100±0.0%100±0.0%99.8%
DoF1~317.9100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~314.596.2±11.4%77.8±12.5%88.0±10.4%96.8±3.0%89.7%
", + "image_path": "8e3470a77711c197f73486545df09d246217489a8292513a8ffd640ad8d75122.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 53, + 191, + 301, + 211 + ], + "lines": [ + { + "bbox": [ + 53, + 191, + 301, + 211 + ], + "spans": [ + { + "bbox": [ + 53, + 191, + 301, + 211 + ], + "type": "text", + "content": "(a) Simulation Setup. Adversarial patch generated on OpenVLA-LIBERO-10 victim model with LIBERO-Long dataset and evaluated on LIBERO-X." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 306, + 106, + 552, + 190 + ], + "blocks": [ + { + "bbox": [ + 306, + 106, + 552, + 190 + ], + "lines": [ + { + "bbox": [ + 306, + 106, + 552, + 190 + ], + "spans": [ + { + "bbox": [ + 306, + 106, + 552, + 190 + ], + "type": "table", + "html": "
ObjectiveAction(s)NAD(%)Victim Model
Spatial△Object△Goal△Long△Avg
Benign--15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
UMADoF126.789.6±15.3%60.8±16.8%64.8±26.8%80.0±20.8%73.8%
DoF1~323.896.6±6.2%56.4±20.4%80.0±20.5%82.0±17.9%78.8%
UADADoF132.699.2±1.8%98.8±2.4%92.6±15.2%96.6±4.8%96.8%
DoF1~327.7100±0.0%100±0.0%100±0.0%100±0.0%100%
UPADoF1~326.995.6±9.5%57.0±23.6%57.2±24.4%72.8±19.7%70.7%
", + "image_path": "2ccd3461e905ec0aa04f3a3175f0e13ec21363d6a253c4f469ec22f9bc917f8a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + }, + { + "bbox": [ + 302, + 191, + 550, + 211 + ], + "lines": [ + { + "bbox": [ + 302, + 191, + 550, + 211 + ], + "spans": [ + { + "bbox": [ + 302, + 191, + 550, + 211 + ], + "type": "text", + "content": "(b) Physical Setup. Adversarial patch generated on OpenVLA-7b victim model with BV2 dataset and evaluated on LIBERO-X." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "table_footnote" + } + ], + "index": 3 + }, + { + "type": "table", + "bbox": [ + 56, + 241, + 553, + 369 + ], + "blocks": [ + { + "bbox": [ + 54, + 216, + 555, + 239 + ], + "lines": [ + { + "bbox": [ + 54, + 216, + 555, + 239 + ], + "spans": [ + { + "bbox": [ + 54, + 216, + 555, + 239 + ], + "type": "text", + "content": "Table 2. Targeted Manipulation Attack Results. Failure Rate (FR, ↑) and its standard deviation across tasks within LIBERO [44] suite are reported. The performance of task suits and DoFs is averaged separately, with the best and second-best values separately." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 241, + 553, + 369 + ], + "lines": [ + { + "bbox": [ + 56, + 241, + 553, + 369 + ], + "spans": [ + { + "bbox": [ + 56, + 241, + 553, + 369 + ], + "type": "table", + "html": "
Targeted DoF(s)MetricSimulation (Attacked on OpenVLA-LIBERO-10)Physical (Attacked on OpenVLA-7b)
Spatial△Object△Goal△Long*AvgSpatial△Object△Goal△Long△Avg
BenignFR15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%15.3±10.2%11.6±10.0%20.8±12.0%46.3±18.6%23.5%
Random Noise28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%28.8±24.2%14.8±7.9%21.0±15.5%48.4±14.8%28.3%
DoF1FR100±0.0%97.8±3.0%100±0.0%100±0.0%99.5%90.6±19.9%36.6±18.1%61.0±21.7%86.4±12.7%68.7%
DoF2100±0.0%99.2±1.3%100±0.0%100±0.0%99.8%99.6±0.8%69.2±11.1%76.2±26.1%91.8±9.9%84.2%
DoF3100±0.0%97.2±6.0%100±0.0%100±0.0%99.3%72.0±14.8%30.2±17.4%39.2±20.9%62.4±17.5%51.0%
DoF494.0±11.0%37.4±23.2%51.2±26.5%87.6±14.9%67.6%93.6±9.7%31.8±20.3%45.0±18.3%63.8±17.0%58.6%
DoF5100±0.0%92.6±8.4%100±0.0%100±0.0%98.2%97.8±4.9%36.0±22.5%62.0±24.8%79.4±21.3%68.8%
DoF6100±0.0%80.0±9.5%92.6±13.4%100±0.0%93.2%79.0±14.2%22.2±14.4%41.0±21.5%62.6±15.7%51.2%
DoF799.2±1.3%86.6±8.3%67.2±22.5%96.4±6.4%87.4%91.8±15.3%98.8±1.3%70.6±23.4%93.2±10.2%88.6%
DoF1~3100±0.0%100±0.0%100±0.0%100±0.0%100%96.4±8.0%83.6±16.8%74.4±26.3%91.8±11.9%86.6%
Task Avg99.2%86.4%88.9%98.0%-90.1%51.1%58.7%78.9%-
", + "image_path": "a15e18d92c0fcb6be49c990cb7fed52fec6e594e9d86a4ee69ec12aba4592436.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "table_body" + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "spans": [ + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": "in manipulating robotic trajectory and increasing FR (see Tab. 2). Specifically, when applied to the different generation settings, our approach yields notable increments across various tasks, including a max average failure rate " + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": " v.s. " + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "inline_equation", + "content": "23.5\\%" + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": " of benign performance in Simulation, and " + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "inline_equation", + "content": "88.6\\%" + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": " in Physical, respectively. Moreover, during attack " + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_4" + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": ", we observe low FR and high task deviation. This failure can be attributed to the fact that " + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "inline_equation", + "content": "\\mathrm{DoF}_4" + }, + { + "bbox": [ + 54, + 374, + 297, + 529 + ], + "type": "text", + "content": " controls the orientation along the x-axis, which can be redundant DoF in tasks. As a result, attacks targeting redundant DoF(s) [38] are less likely to disrupt execution effectively. This observation underscores the importance of task-specific considerations when designing adversarial attacks on robotic systems." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 54, + 534, + 298, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 534, + 298, + 714 + ], + "spans": [ + { + "bbox": [ + 54, + 534, + 298, + 714 + ], + "type": "text", + "content": "Qualitative Results. We qualitatively analyze robot movement trajectories under the three proposed attacks in Fig. 3. For UADA, we compare the trajectory deviations in the same trail with patches generated from UADA and UMA attacks. As seen, the UMA induces small deviations in the trajectory. UADA, on the other hand, produces significantly larger trajectory deviations (also supported by NAD metric in Tab. 1b), which is attributed to UADA's capability of incorporating action discrepancies. In our observation, UADA significantly amplifies the effect on overall task execution, thereby increasing the potential for robotic hazards. For UPA, we observe chaotic and irregular behaviors, including instances where the end-effector moves out of the camera's field of view. We attribute this phenomenon to the efficacy of the adversarial patch in perturbing the model's spatial perception," + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 374, + 555, + 445 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 374, + 555, + 445 + ], + "spans": [ + { + "bbox": [ + 313, + 374, + 555, + 445 + ], + "type": "text", + "content": "inducing a consistent deviation from the intended direction of movement, ultimately resulting in failure in a long run. For TMA, we observe a notable reduction in the range of motion along the x-axis corresponding to the targeted attack axis. This suggests that our proposed targeted attack can effectively constrained robot movements." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 446, + 556, + 518 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 446, + 556, + 518 + ], + "spans": [ + { + "bbox": [ + 313, + 446, + 556, + 518 + ], + "type": "text", + "content": "In summary, the qualitative analysis shows that both Untarget attack, UADA, UPA, and TMA can effectively disrupt the robot actions generated from VLA models. These findings underscore a pressing security concern during the deployment of generalist robots, especially when considering application scenes that require reliable operations [11, 68]." + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 329, + 521, + 538, + 617 + ], + "blocks": [ + { + "bbox": [ + 329, + 521, + 538, + 617 + ], + "lines": [ + { + "bbox": [ + 329, + 521, + 538, + 617 + ], + "spans": [ + { + "bbox": [ + 329, + 521, + 538, + 617 + ], + "type": "image", + "image_path": "0161ac097af27979f5c686f293cee7b1fa3c957783c496e98982d3444ce4ada4.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 619, + 555, + 641 + ], + "lines": [ + { + "bbox": [ + 313, + 619, + 555, + 641 + ], + "spans": [ + { + "bbox": [ + 313, + 619, + 555, + 641 + ], + "type": "text", + "content": "Figure 4. Qualitative Results of the physical world. The first/second row show benign and adversarial cases respectively." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 556, + 714 + ], + "type": "text", + "content": "Real-world Performance. In addition to the digital simulation results, we also conduct a comprehensive evaluation of the performance of our generated adversarial patches in real-world scenarios. The evaluation encompassed 100 trials across three distinct tasks: object grasping, placement, and manipulation. As shown in Fig. 4, the adversarial patch" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6954" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 70, + 71, + 124, + 151 + ], + "blocks": [ + { + "bbox": [ + 70, + 71, + 124, + 151 + ], + "lines": [ + { + "bbox": [ + 70, + 71, + 124, + 151 + ], + "spans": [ + { + "bbox": [ + 70, + 71, + 124, + 151 + ], + "type": "image", + "image_path": "3fde9a9844a89860e0754a29d91bdf04c8255d59b85cad5673fd228251be5c83.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 80, + 154, + 124, + 163 + ], + "lines": [ + { + "bbox": [ + 80, + 154, + 124, + 163 + ], + "spans": [ + { + "bbox": [ + 80, + 154, + 124, + 163 + ], + "type": "text", + "content": "(a) Inner-loop" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 124, + 71, + 215, + 151 + ], + "blocks": [ + { + "bbox": [ + 124, + 71, + 215, + 151 + ], + "lines": [ + { + "bbox": [ + 124, + 71, + 215, + 151 + ], + "spans": [ + { + "bbox": [ + 124, + 71, + 215, + 151 + ], + "type": "image", + "image_path": "809061dfd5fe3958fdfd83eb8711dd71b12036eafb386de14c8b3a2b22dae332.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 148, + 154, + 191, + 163 + ], + "lines": [ + { + "bbox": [ + 148, + 154, + 191, + 163 + ], + "spans": [ + { + "bbox": [ + 148, + 154, + 191, + 163 + ], + "type": "text", + "content": "(b) Patch Size" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 216, + 71, + 299, + 152 + ], + "blocks": [ + { + "bbox": [ + 216, + 71, + 299, + 152 + ], + "lines": [ + { + "bbox": [ + 216, + 71, + 299, + 152 + ], + "spans": [ + { + "bbox": [ + 216, + 71, + 299, + 152 + ], + "type": "image", + "image_path": "aa8c52338edb1b5e3959b050bc5389712291477bbcde871f240f2d15a3a4a905.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 244, + 154, + 272, + 163 + ], + "lines": [ + { + "bbox": [ + 244, + 154, + 272, + 163 + ], + "spans": [ + { + "bbox": [ + 244, + 154, + 272, + 163 + ], + "type": "text", + "content": "(c) JPEG" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 299, + 71, + 377, + 152 + ], + "blocks": [ + { + "bbox": [ + 299, + 71, + 377, + 152 + ], + "lines": [ + { + "bbox": [ + 299, + 71, + 377, + 152 + ], + "spans": [ + { + "bbox": [ + 299, + 71, + 377, + 152 + ], + "type": "image", + "image_path": "50e2ebe37466de40a562310a4612b267b5cf1fe5328d4545b824eccc09b63551.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 326, + 154, + 355, + 163 + ], + "lines": [ + { + "bbox": [ + 326, + 154, + 355, + 163 + ], + "spans": [ + { + "bbox": [ + 326, + 154, + 355, + 163 + ], + "type": "text", + "content": "(d) Noise" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 377, + 71, + 455, + 152 + ], + "blocks": [ + { + "bbox": [ + 377, + 71, + 455, + 152 + ], + "lines": [ + { + "bbox": [ + 377, + 71, + 455, + 152 + ], + "spans": [ + { + "bbox": [ + 377, + 71, + 455, + 152 + ], + "type": "image", + "image_path": "aec35d748df36db7070db39a02bd5670bfdacc54ab9f44352ab116361f175606.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 395, + 154, + 444, + 163 + ], + "lines": [ + { + "bbox": [ + 395, + 154, + 444, + 163 + ], + "spans": [ + { + "bbox": [ + 395, + 154, + 444, + 163 + ], + "type": "text", + "content": "(e) Median Blur" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 456, + 71, + 536, + 152 + ], + "blocks": [ + { + "bbox": [ + 456, + 71, + 536, + 152 + ], + "lines": [ + { + "bbox": [ + 456, + 71, + 536, + 152 + ], + "spans": [ + { + "bbox": [ + 456, + 71, + 536, + 152 + ], + "type": "image", + "image_path": "eb35dcf69370a82268f417f57d53cb4d9376908e211032ace0d6e0c65cfb30b6.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 478, + 154, + 519, + 163 + ], + "lines": [ + { + "bbox": [ + 478, + 154, + 519, + 163 + ], + "spans": [ + { + "bbox": [ + 478, + 154, + 519, + 163 + ], + "type": "text", + "content": "(f) Bit-Depth" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "text", + "content": "generated with UADA demonstrated the ability to manipulate the robot effectively, achieving an attack success rate exceeding " + }, + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "inline_equation", + "content": "43\\%" + }, + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "text", + "content": ". Although this success rate is lower than the corresponding digital-world performance (i.e., " + }, + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "inline_equation", + "content": "100\\%" + }, + { + "bbox": [ + 54, + 200, + 297, + 344 + ], + "type": "text", + "content": "), it highlights the effectiveness of our patches in physical-world applications as well without the need for further adaptations. Crucially, we observed that the induced erratic movements (similar to simulation scenario) of the robot during successful attacks pose significant risks to human safety and surrounding environments. This finding emphasizes the severe threat posed by VLA models in real-world applications, consistent with similar observations in digital settings." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 348, + 186, + 361 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 348, + 186, + 361 + ], + "spans": [ + { + "bbox": [ + 55, + 348, + 186, + 361 + ], + "type": "text", + "content": "4.4. Diagnostic Experiment" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 55, + 363, + 296, + 458 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 458 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 458 + ], + "type": "text", + "content": "Impact of Inner-loops. We discuss the impact of inner-loop steps to model performance in Fig. 5(a). The results show that NAD first improves when inner-loop steps continue to increase. This is due to the reduced data variance and gradually stabilized gradients. However, further increasing the steps results in a slower performance increase, with a noticeably longer training schedule. We thus choose inner-loop steps=50 as it balances between performance and scale." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "spans": [ + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "text", + "content": "Impact of Patch Size. We study TMA's performance with different patch sizes and report the average FR across four LIBERO [44] tasks in Fig. 5(b). The patch sizes examined are " + }, + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "inline_equation", + "content": "[3\\%, 5\\%, 7\\%, 10\\%]" + }, + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "text", + "content": " of input image. Our findings indicate an inversely proportional relationship between the " + }, + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "inline_equation", + "content": "L_{1}" + }, + { + "bbox": [ + 55, + 459, + 296, + 579 + ], + "type": "text", + "content": " distance of predicted actions and the target action as the patch size increased. This increase in patch size also correlated with a marked rise in FR. The observed trend suggests larger patches give adversaries more optimization space to influence the model, aligning with prior work [12, 81]." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 55, + 582, + 185, + 594 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 185, + 594 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 185, + 594 + ], + "type": "text", + "content": "4.5. Robustness Evaluation" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "spans": [ + { + "bbox": [ + 55, + 598, + 296, + 670 + ], + "type": "text", + "content": "We further examine whether concurrent defense strategies [18, 78, 80] can resist our adversarial examples within VLA. Specifically, we applied four prior defense techniques and reported their FR for both benign and adversarial samples in Fig. 5(c-f), respectively. The results show that our adversarial attack bypasses most defense strategies." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 55, + 673, + 171, + 686 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 673, + 171, + 686 + ], + "spans": [ + { + "bbox": [ + 55, + 673, + 171, + 686 + ], + "type": "text", + "content": "4.6. Systemic Discussion" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 714 + ], + "type": "text", + "content": "Patch Pattern Analysis. In Fig. 6, we present the adversarial patches generated for a range of targets. A particularly" + } + ] + } + ], + "index": 20 + }, + { + "type": "image", + "bbox": [ + 319, + 201, + 353, + 236 + ], + "blocks": [ + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "lines": [ + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "spans": [ + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "type": "text", + "content": "Figure 5. Impact of Inner-loop, Patch Size and Defense Discussion. The figure shows how varying Inner-loop affects NAD in UADA, and patch sizes affect " + }, + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "type": "inline_equation", + "content": "L1" + }, + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "type": "text", + "content": " distance and the failure rates in TMA, both targeting at DoF" + }, + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "type": "inline_equation", + "content": "_1" + }, + { + "bbox": [ + 55, + 164, + 555, + 197 + ], + "type": "text", + "content": ". (a) Impact of Inner-loop, (b) Impact of Patch Size and (c-f) the effect of four different defenses on failure rates." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 319, + 201, + 353, + 236 + ], + "lines": [ + { + "bbox": [ + 319, + 201, + 353, + 236 + ], + "spans": [ + { + "bbox": [ + 319, + 201, + 353, + 236 + ], + "type": "image", + "image_path": "6647818771e16f42e30d226e9343a943174dc32d516535cce66e32304da2161e.jpg" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_body" + } + ], + "index": 21 + }, + { + "type": "image", + "bbox": [ + 358, + 201, + 392, + 236 + ], + "blocks": [ + { + "bbox": [ + 358, + 201, + 392, + 236 + ], + "lines": [ + { + "bbox": [ + 358, + 201, + 392, + 236 + ], + "spans": [ + { + "bbox": [ + 358, + 201, + 392, + 236 + ], + "type": "image", + "image_path": "b0278a6162a00b426c5c683f1356a894d55279d8a5fa69166b1d9654d68b6fb3.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 337, + 239, + 413, + 250 + ], + "lines": [ + { + "bbox": [ + 337, + 239, + 413, + 250 + ], + "spans": [ + { + "bbox": [ + 337, + 239, + 413, + 250 + ], + "type": "text", + "content": "(a) BridgeData V2 [69]" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 397, + 201, + 431, + 236 + ], + "blocks": [ + { + "bbox": [ + 397, + 201, + 431, + 236 + ], + "lines": [ + { + "bbox": [ + 397, + 201, + 431, + 236 + ], + "spans": [ + { + "bbox": [ + 397, + 201, + 431, + 236 + ], + "type": "image", + "image_path": "270053252a7867145c80093655706037c8dda7c601624b1ae2fb649b325f4f39.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 336, + 251, + 531, + 262 + ], + "lines": [ + { + "bbox": [ + 336, + 251, + 531, + 262 + ], + "spans": [ + { + "bbox": [ + 336, + 251, + 531, + 262 + ], + "type": "text", + "content": "Figure 6. Patch Visualization. Semantic-rich patches." + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 23 + }, + { + "type": "image", + "bbox": [ + 435, + 201, + 470, + 236 + ], + "blocks": [ + { + "bbox": [ + 435, + 201, + 470, + 236 + ], + "lines": [ + { + "bbox": [ + 435, + 201, + 470, + 236 + ], + "spans": [ + { + "bbox": [ + 435, + 201, + 470, + 236 + ], + "type": "image", + "image_path": "ee3a2ee1d1681df590d280ecb23433001c527085eda6cfa90dfa87f3741023b3.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 463, + 239, + 519, + 250 + ], + "lines": [ + { + "bbox": [ + 463, + 239, + 519, + 250 + ], + "spans": [ + { + "bbox": [ + 463, + 239, + 519, + 250 + ], + "type": "text", + "content": "(b) LIBERO [44]" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 474, + 201, + 509, + 236 + ], + "blocks": [ + { + "bbox": [ + 474, + 201, + 509, + 236 + ], + "lines": [ + { + "bbox": [ + 474, + 201, + 509, + 236 + ], + "spans": [ + { + "bbox": [ + 474, + 201, + 509, + 236 + ], + "type": "image", + "image_path": "3c4da73883a24268bd55d0e3459b8a31f977fd15624a301f0ce345ffc9426d13.jpg" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_body" + } + ], + "index": 25 + }, + { + "type": "image", + "bbox": [ + 513, + 201, + 547, + 236 + ], + "blocks": [ + { + "bbox": [ + 513, + 201, + 547, + 236 + ], + "lines": [ + { + "bbox": [ + 513, + 201, + 547, + 236 + ], + "spans": [ + { + "bbox": [ + 513, + 201, + 547, + 236 + ], + "type": "image", + "image_path": "6c71575b7ab846bd94392342ea4be2ad4acfe3eb7f9991a7caa58911db48c16b.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + } + ], + "index": 26 + }, + { + "bbox": [ + 313, + 268, + 555, + 447 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 268, + 555, + 447 + ], + "spans": [ + { + "bbox": [ + 313, + 268, + 555, + 447 + ], + "type": "text", + "content": "noteworthy observation is that some of these patches bear a striking resemblance to the structural joints of a robotic arm. Given the strong similarity between these patches and robotic arms in appearance, we hypothesize that these VLA-based models, in their current training paradigms, are often limited to a narrow range of robotic systems operating within restricted camera views. This constrained perspective camera pose may inadvertently induce a learned representation bias, where adversarial perturbations align with prominent structural features of the training robot's appearance in order to deceive current victim models. Given that these patches are model-specific and reflect the spatial and visual patterns of the training environment, we have a reasonable concern that they may compromise VLA-based models' generalizability and robustness in physical-world deployments." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 313, + 448, + 556, + 579 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 448, + 556, + 579 + ], + "spans": [ + { + "bbox": [ + 313, + 448, + 556, + 579 + ], + "type": "text", + "content": "Possible Defense Strategies. Our work highlights the pressing need for a paradigm shift in the VLA-based models' training strategies in order to successfully defend our attacks. Specifically, future approaches might incorporate more training scenarios that involve complex multi-robot interactions to mitigate the current patch bias. Another possible solution is to leverage the logical relationships within the robot arm's physical structure to estimate the actual position of each joint. By predicting joint positions deviated from physical constraints, we can eliminate unreasonable adversarial samples, even if they are similar to actual joint(s)." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 313, + 587, + 388, + 599 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 587, + 388, + 599 + ], + "spans": [ + { + "bbox": [ + 313, + 587, + 388, + 599 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 555, + 713 + ], + "type": "text", + "content": "While VLA models have gained significant popularity due to their substantial robot capabilities, this study pioneers specific robotic attack objectives and demonstrates that these models are in fact vulnerable to various types of attacks. Experimental results demonstrate that our attacks expose significant vulnerabilities to VLA models, raising concerns regarding impetuous real-world deployments. We believe our framework provides pioneering and foundational contributions to the reliability of AI-enhanced robotic systems." + } + ] + } + ], + "index": 33 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6955" + } + ] + } + ], + "index": 34 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 158, + 85 + ], + "type": "text", + "content": "Acknowledgements" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 297, + 212 + ], + "type": "text", + "content": "This research was supported by the National Science Foundation under Grant No. 2450068 and No. 2242243, as well as start-up resources provided by Rutgers University. The views and conclusions contained herein are those of the authors and should not be interpreted as necessarily representing the official policies or endorsements, either expressed or implied, of U.S. Naval Research Laboratory (NRL) or the U.S. Government. The U.S. Government is authorized to reproduce and distribute reprints for Government purposes notwithstanding any copyright notation herein." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 223, + 115, + 235 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 223, + 115, + 235 + ], + "spans": [ + { + "bbox": [ + 56, + 223, + 115, + 235 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 243, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 243, + 296, + 286 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 243, + 296, + 286 + ], + "spans": [ + { + "bbox": [ + 61, + 243, + 296, + 286 + ], + "type": "text", + "content": "[1] Mokhaled NA Al-Hamadani, Mohammed A Fadhel, Laith Alzubaidi, and Balazs Harangi. Reinforcement learning algorithms and applications in healthcare and robotics: A comprehensive and systematic review. Sensors, 24(8):2461, 2024." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "spans": [ + { + "bbox": [ + 62, + 288, + 296, + 320 + ], + "type": "text", + "content": "[2] Anish Athalye, Nicholas Carlini, and David A. Wagner. Obfuscated gradients give a false sense of security: Circumventing defenses to adversarial examples. In ICML, 2018." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 322, + 296, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 322, + 296, + 354 + ], + "spans": [ + { + "bbox": [ + 62, + 322, + 296, + 354 + ], + "type": "text", + "content": "[3] Anish Athalye, Logan Engstrom, Andrew Ilyas, and Kevin Kwok. Synthesizing robust adversarial examples. In ICML, 2018." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 357, + 296, + 388 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 357, + 296, + 388 + ], + "spans": [ + { + "bbox": [ + 62, + 357, + 296, + 388 + ], + "type": "text", + "content": "[4] Lefteris Benos, Avital Bechar, and Dionysis Bochtis. Safety and ergonomics in human-robot interactive agricultural operations. Biosystems Engineering, 200:55-72, 2020." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "spans": [ + { + "bbox": [ + 62, + 390, + 296, + 443 + ], + "type": "text", + "content": "[5] Nicole Berx, Wilm Decré, Ido Morag, Peter Chemweno, and Liliane Pintelon. Identification and classification of risk factors for human-robot collaboration from a system-wide perspective. Computers & Industrial Engineering, 163:107827, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 446, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 446, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 446, + 294, + 468 + ], + "type": "text", + "content": "[6] Aude Billard and Danica Kragic. Trends and challenges in robot manipulation. Science, 364(6446), 2019." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 470, + 297, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 470, + 297, + 633 + ], + "spans": [ + { + "bbox": [ + 62, + 470, + 297, + 633 + ], + "type": "text", + "content": "[7] Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Joseph Dabis, Chelsea Finn, Keerthana Gopalakrishnan, Karol Hausman, Alexander Herzog, Jasmine Hsu, Julian Ibarz, Brian Ichter, Alex Irpan, Tomas Jackson, Sally Jesmonth, Nikhil J. Joshi, Ryan Julian, Dmitry Kalashnikov, Yuheng Kuang, Isabel Leal, Kuang-Huei Lee, Sergey Levine, Yao Lu, Utsav Malla, Deeksha Manjunath, Igor Mordatch, Ofir Nachum, Carolina Parada, Jodilyn Peralta, Emily Perez, Karl Pertsch, Jornell Quiambao, Kanishka Rao, Michael S. Ryoo, Grecia Salazar, Pannag R. Sanketi, Kevin Sayed, Jaspiar Singh, Sumedh Sontakke, Austin Stone, Clayton Tan, Huong T. Tran, Vincent Vanhoucke, Steve Vega, Quan Vuong, Fei Xia, Ted Xiao, Peng Xu, Sichun Xu, Tianhe Yu, and Brianna Zitkovich. RT-1: robotics transformer for real-world control at scale. In Robotics: Science and Systems XIX, 2023." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 635, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 635, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 62, + 635, + 296, + 667 + ], + "type": "text", + "content": "[8] Tom B. Brown, Dandelion Mane, Aurko Roy, Martin Abadi, and Justin Gil mer. Adversarial patch. arXiv preprint arXiv:1712.09665, 2018." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 62, + 670, + 296, + 713 + ], + "type": "text", + "content": "[9] Nicholas Carlini and David A. Wagner. Adversarial examples are not easily detected: Bypassing ten detection methods. In Proceedings of the 10th ACM Workshop on Artificial Intelligence and Security, pages 3-14, 2017." + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 555, + 127 + ], + "type": "text", + "content": "[10] Xi Chen, Xiao Wang, Lucas Beyer, Alexander Kolesnikov, Jialin Wu, Paul Voigtlaender, Basil Mustafa, Sebastian Goodman, Ibrahim Alabdulmohsin, Piotr Padlewski, et al. Pali-3 vision language models: Smaller, faster, stronger. arXiv preprint arXiv:2310.09199, 2023." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 555, + 162 + ], + "type": "text", + "content": "[11] Shen Cheng and BS Dhillon. Reliability and availability analysis of a robot-safety system. Journal of Quality in Maintenance Engineering, 17:203-232, 2011." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 317, + 163, + 555, + 206 + ], + "type": "text", + "content": "[12] Zhiyuan Cheng, Cheng Han, James Liang, Qifan Wang, Xi-angyu Zhang, and Dongfang Liu. Self-supervised adversarial training of monocular depth estimation against physical-world attacks. arXiv preprint arXiv:2406.05857, 2024." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 208, + 554, + 272 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 208, + 554, + 272 + ], + "spans": [ + { + "bbox": [ + 317, + 208, + 554, + 272 + ], + "type": "text", + "content": "[13] Hao-Tien Lewis Chiang, Zhuo Xu, Zipeng Fu, Mithun George Jacob, Tingnan Zhang, Tsang-Wei Edward Lee, Wenhao Yu, Connor Schenck, David Rendleman, Dhruv Shah, et al. Mobility via: Multimodal instruction navigation with long-context vlms and topological graphs. arXiv preprint arXiv:2407.07775, 2024." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 274, + 554, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 274, + 554, + 297 + ], + "spans": [ + { + "bbox": [ + 317, + 274, + 554, + 297 + ], + "type": "text", + "content": "[14] Jinda Cui and Jeff Trinkle. Toward next-generation learned robot manipulation. Sci. Robotics, 6(54):9461, 2021." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 298, + 555, + 341 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 298, + 555, + 341 + ], + "spans": [ + { + "bbox": [ + 317, + 298, + 555, + 341 + ], + "type": "text", + "content": "[15] Yiming Cui, Cheng Han, and Dongfang Liu. Collaborative multi-task learning for multi-object tracking and segmentation. Journal on Autonomous Transportation Systems, 1:1-23, 2024." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 343, + 555, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 343, + 555, + 386 + ], + "spans": [ + { + "bbox": [ + 317, + 343, + 555, + 386 + ], + "type": "text", + "content": "[16] Pengxiang Ding, Han Zhao, Wenjie Zhang, Wenxuan Song, Min Zhang, Siteng Huang, Ningxi Yang, and Donglin Wang. *Quar-vla: Vision-language-action model for quadruped robots*. In *ECCV*, 2025." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 388, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 388, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 317, + 388, + 554, + 430 + ], + "type": "text", + "content": "[17] Andrew Du, Bo Chen, Tat-Jun Chin, Yee Wei Law, Michele Sasdelli, Ramesh Rajasegaran, and Dillon Campbell. Physical adversarial attacks on an aerial imagery object detector. In WACV, 2022." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 433, + 554, + 465 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 433, + 554, + 465 + ], + "spans": [ + { + "bbox": [ + 316, + 433, + 554, + 465 + ], + "type": "text", + "content": "[18] Gintare Karolina Dziugaite, Zoubin Ghahramani, and Daniel M. Roy. A study of the effect of JPG compression on adversarial images. arXiv preprint arXiv:1608.00853, 2016." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 317, + 467, + 555, + 499 + ], + "type": "text", + "content": "[19] Chelsea Finn, Tianhe Yu, Tianhao Zhang, Pieter Abbeel, and Sergey Levine. One-shot visual imitation learning via meta-learning. In CoRL, 2017." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 501, + 555, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 501, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 317, + 501, + 555, + 533 + ], + "type": "text", + "content": "[20] Saskia Golz, Christian Osendorfer, and Sami Haddadin. Using tactile sensation for learning contact knowledge: Discriminate collision from physical interaction. In ICRA, 2015." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 535, + 555, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 535, + 555, + 567 + ], + "spans": [ + { + "bbox": [ + 317, + 535, + 555, + 567 + ], + "type": "text", + "content": "[21] Ian J. Goodfellow, Jonathon Shlens, and Christian Szegedy. Explaining and harnessing adversarial examples. In ICLR, 2015." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 569, + 555, + 601 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 569, + 555, + 601 + ], + "spans": [ + { + "bbox": [ + 317, + 569, + 555, + 601 + ], + "type": "text", + "content": "[22] Jérémie Guiochet, Mathilde Machin, and Hélène Waeselynck. Safety-critical advanced robots: A survey. Robotics Auton. Syst., 94:43-52, 2017." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 603, + 555, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 603, + 555, + 667 + ], + "spans": [ + { + "bbox": [ + 317, + 603, + 555, + 667 + ], + "type": "text", + "content": "[23] Sami Haddadin, Sven Parusel, Lars Johannsmeier, Saskia Golz, Simon Gabl, Florian Walch, Mohamadreza Sabaghian, Christoph Jähne, Lukas Hausperger, and Simon Haddadin. The franka emika robot: A reference platform for robotics research and education. IEEE Robotics & Automation Magazine, 29:46-64, 2022." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 670, + 555, + 713 + ], + "type": "text", + "content": "[24] Roni-Jussi Halme, Minna Lanz, Joni Kämäräinen, Roel Pieters, Jyrki Latokartano, and Antti Hietanen. Review of vision-based safety systems for human-robot collaboration. Procedia Cirp, 72:111-116, 2018." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "6956" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 297, + 116 + ], + "type": "text", + "content": "[25] Jiangyong Huang, Silong Yong, Xiaojian Ma, Xiongkun Linghu, Puhao Li, Yan Wang, Qing Li, Song-Chun Zhu, Baoxiong Jia, and Siyuan Huang. An embodied generalist agent in 3d world. arXiv preprint arXiv:2311.12871, 2023." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 295, + 161 + ], + "type": "text", + "content": "[26] Wenlong Huang, Chen Wang, Ruohan Zhang, Yunzhu Li, Jiajun Wu, and Li Fei-Fei. Voxposer: Composable 3d value maps for robotic manipulation with language models. In CoRL, 2023." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 163, + 296, + 185 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 163, + 296, + 185 + ], + "spans": [ + { + "bbox": [ + 56, + 163, + 296, + 185 + ], + "type": "text", + "content": "[27] Félix Ingrand and Malik Ghallab. Deliberation for autonomous robots: A survey. Artif. Intell., 247:10-44, 2017." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 186, + 296, + 228 + ], + "type": "text", + "content": "[28] ISO 10218-1/2:2011. Robots and Robotic Devices Safety Requirements for Industrial Robots Part 1: Robots/Part 2: Robot Systems and Integration. Standard, International Organization for Standardization, 2011." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "spans": [ + { + "bbox": [ + 56, + 230, + 296, + 263 + ], + "type": "text", + "content": "[29] ISO13482. Robots and robotic devices — safety requirements for personal care robots. International Organization for Standardization, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 264, + 296, + 297 + ], + "type": "text", + "content": "[30] ISO/TS 15066:2016. Robots and Robotic Devices Collaborative Robots. Standard, International Organization for Standardization, 2016." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "spans": [ + { + "bbox": [ + 56, + 298, + 296, + 331 + ], + "type": "text", + "content": "[31] Stephen James, Michael Bloesch, and Andrew J. Davison. Task-embedded control networks for few-shot imitation learning. In CoRL, 2018." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "spans": [ + { + "bbox": [ + 56, + 332, + 296, + 386 + ], + "type": "text", + "content": "[32] Yunfan Jiang, Agrim Gupta, Zichen Zhang, Guanzhi Wang, Yongqiang Dou, Yanjun Chen, Li Fei-Fei, Anima Anandkumar, Yuke Zhu, and Linxi Fan. VIMA: general robot manipulation with multimodal prompts. arXiv preprint arXiv:2210.03094, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "spans": [ + { + "bbox": [ + 56, + 388, + 296, + 432 + ], + "type": "text", + "content": "[33] Siddharth Karamcheti, Suraj Nair, Ashwin Balakrishna, Percy Liang, Thomas Kollar, and Dorsa Sadigh. Prismatic vlms: Investigating the design space of visually-conditioned language models. arXiv preprint arXiv:2402.07865, 2024." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 434, + 296, + 509 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 434, + 296, + 509 + ], + "spans": [ + { + "bbox": [ + 56, + 434, + 296, + 509 + ], + "type": "text", + "content": "[34] Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, Quan Vuong, Thomas Kollar, Benjamin Burchfiel, Russ Tedrake, Dorsa Sadigh, Sergey Levine, Percy Liang, and Chelsea Finn. Openvla: An open-source vision-language-action model. arXiv preprint arXiv:2406.09246, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 510, + 296, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 510, + 296, + 544 + ], + "spans": [ + { + "bbox": [ + 56, + 510, + 296, + 544 + ], + "type": "text", + "content": "[35] Hema Swetha Koppula and Ashutosh Saxena. Learning spatio-temporal structure from RGB-D videos for human activity detection and anticipation. In ICML, 2013." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "spans": [ + { + "bbox": [ + 56, + 544, + 296, + 588 + ], + "type": "text", + "content": "[36] Hema Swetha Koppula and Ashutosh Saxena. Anticipating human activities using object affordances for reactive robotic response. IEEE Trans. Pattern Anal. Mach. Intell., 38(1): 14-29, 2016." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 590, + 296, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 590, + 296, + 632 + ], + "spans": [ + { + "bbox": [ + 56, + 590, + 296, + 632 + ], + "type": "text", + "content": "[37] Oliver Kroemer, Scott Niekum, and George Konidaris. A review of robot learning for manipulation: Challenges, representations, and algorithms. J. Mach. Learn. Res., 22:30:1-30:82, 2021." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 635, + 296, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 635, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 635, + 296, + 656 + ], + "type": "text", + "content": "[38] Andrew J. Kurdila and Pinhas Ben-Tzvi. Dynamics and control of robotic systems. John Wiley & Sons, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 657, + 296, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 657, + 296, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 657, + 296, + 689 + ], + "type": "text", + "content": "[39] Przemyslaw A. Lasota, Terrence Fong, and Julie A. Shah. A survey of methods for safe human-robot interaction. Found. Trends Robotics, 5(4):261–349, 2017." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 296, + 713 + ], + "type": "text", + "content": "[40] Sergey Levine, Peter Pastor, Alex Krizhevsky, Julian Ibarz, and Deirdre Quillen. Learning hand-eye coordination for" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 36, + "blocks": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 555, + 95 + ], + "type": "text", + "content": "robotic grasping with deep learning and large-scale data collection. Int. J. Robotics Res., 37(4-5):421-436, 2018." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 555, + 149 + ], + "type": "text", + "content": "[41] Xinghang Li, Minghuan Liu, Hanbo Zhang, Cunjun Yu, Jie Xu, Hongtao Wu, Chilam Cheang, Ya Jing, Weinan Zhang, Huaping Liu, et al. Vision-language foundation models as effective robot imitators. arXiv preprint arXiv:2311.01378, 2023." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 151, + 555, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 151, + 555, + 206 + ], + "spans": [ + { + "bbox": [ + 316, + 151, + 555, + 206 + ], + "type": "text", + "content": "[42] Xiang Li, Cristina Mata, Jongwoo Park, Kumara Kahatapitiya, Yoo Sung Jang, Jinghuan Shang, Kanchana Ranasinghe, Ryan Burgert, Mu Cai, Yong Jae Lee, and Michael S. Ryoo. Llara: Supercharging robot learning data for vision-language policy. arXiv preprint arXiv:2406.20095, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 207, + 554, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 207, + 554, + 240 + ], + "spans": [ + { + "bbox": [ + 316, + 207, + 554, + 240 + ], + "type": "text", + "content": "[43] Aishan Liu, Xianglong Liu, Jiaxin Fan, Yuqing Ma, Anlan Zhang, Huiyuan Xie, and Dacheng Tao. Perceptual-sensitive GAN for generating adversarial patches. In AAAI, 2019." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 241, + 555, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 241, + 555, + 274 + ], + "spans": [ + { + "bbox": [ + 316, + 241, + 555, + 274 + ], + "type": "text", + "content": "[44] Bo Liu, Yifeng Zhu, Chongkai Gao, Yihao Feng, Qiang Liu, Yuke Zhu, and Peter Stone. LIBERO: benchmarking knowledge transfer for lifelong robot learning. In NeurIPS, 2023." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 274, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 274, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 274, + 555, + 306 + ], + "type": "text", + "content": "[45] Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In CVPR, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 308, + 555, + 331 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 555, + 331 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 555, + 331 + ], + "type": "text", + "content": "[46] Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In NeurlPS, 2024." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 331, + 554, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 554, + 353 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 554, + 353 + ], + "type": "text", + "content": "[47] Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In *ICLR*, 2019." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 354, + 555, + 397 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 354, + 555, + 397 + ], + "spans": [ + { + "bbox": [ + 316, + 354, + 555, + 397 + ], + "type": "text", + "content": "[48] Alessandro De Luca, Alin Albu-Schäffer, Sami Haddadin, and Gerd Hirzinger. Collision detection and safe reaction with the DLR-III lightweight manipulator arm. In IROS, 2006." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 399, + 554, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 554, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 554, + 430 + ], + "type": "text", + "content": "[49] Aleksander Madry, Aleksandar Makelov, Ludwig Schmidt, Dimitris Tsipras, and Adrian Vladu. Towards deep learning models resistant to adversarial attacks. In ICLR, 2018." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 464 + ], + "type": "text", + "content": "[50] Jim Mainprice and Dmitry Berenson. Human-robot collaborative manipulation planning using early prediction of human motion. In IROS, 2013." + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 466, + 554, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 466, + 554, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 466, + 554, + 487 + ], + "type": "text", + "content": "[51] Matthew T. Mason. Mechanics of Robotic Manipulation. MIT Press, 2001." + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 488, + 555, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 488, + 555, + 510 + ], + "spans": [ + { + "bbox": [ + 316, + 488, + 555, + 510 + ], + "type": "text", + "content": "[52] Matthew T. Mason. Toward robotic manipulation. Annu. Rev. Control. Robotics Auton. Syst., 1:1-28, 2018." + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "spans": [ + { + "bbox": [ + 316, + 512, + 555, + 555 + ], + "type": "text", + "content": "[53] Oier Mees, Lukás Hermann, and Wolfram Burgard. What matters in language conditioned robotic imitation learning over unstructured data. IEEE Robotics Autom. Lett., 7(4): 11205-11212, 2022." + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 555, + 555, + 589 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 555, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 316, + 555, + 555, + 589 + ], + "type": "text", + "content": "[54] Vivek G Moudgal, Kevin M Passino, and Stephen Yurkovich. Rule-based control for a flexible-link robot. IEEE Trans. on Control Sys. Tech., 2(4):392-405, 1994." + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 316, + 590, + 554, + 622 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 590, + 554, + 622 + ], + "spans": [ + { + "bbox": [ + 316, + 590, + 554, + 622 + ], + "type": "text", + "content": "[55] Richard M. Murray, Zexiang Li, and S. Shankar Sastry. A mathematical introduction to robotics manipulation. CRC Press, 1994." + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 555, + 656 + ], + "type": "text", + "content": "[56] Lerrel Pinto and Abhinav Gupta. Supersizing self-supervision: Learning to grasp from 50k tries and 700 robot hours. In ICRA, 2016." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "spans": [ + { + "bbox": [ + 316, + 657, + 555, + 690 + ], + "type": "text", + "content": "[57] Vitchyr Pong, Murtaza Dalal, Steven Lin, Ashvin Nair, Shikhar Bahl, and Sergey Levine. Skew-fit: State-covering self-supervised reinforcement learning. In ICML, 2020." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 691, + 555, + 713 + ], + "type": "text", + "content": "[58] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mittal. Visual adversarial" + } + ] + } + ], + "index": 35 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6957" + } + ] + } + ], + "index": 37 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 296, + 713 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "spans": [ + { + "bbox": [ + 76, + 72, + 296, + 94 + ], + "type": "text", + "content": "examples jailbreak aligned large language models. In AAAI, 2024." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "spans": [ + { + "bbox": [ + 56, + 95, + 296, + 138 + ], + "type": "text", + "content": "[59] Mahmood Sharif, Sruti Bhagavatula, Lujo Bauer, and Michael K. Reiter. Accessorize to a crime: Real and stealthy attacks on state-of-the-art face recognition. In ACM SIGSAC, 2016." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 140, + 296, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 296, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 296, + 161 + ], + "type": "text", + "content": "[60] Michael H. Smith and L. Stephen Coles. Design of a low cost, general purpose robot. In IJCAI, 1973." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 296, + 195 + ], + "type": "text", + "content": "[61] Jiawei Su, Danilo Vasconcellos Vargas, and Kouichi Sakurai. One pixel attack for fooling deep neural networks. IEEE Trans. Evol. Comput., 23(5):828-841, 2019." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 196, + 296, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 196, + 296, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 196, + 296, + 228 + ], + "type": "text", + "content": "[62] Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian J. Goodfellow, and Rob Fergus. Intriguing properties of neural networks. In ICLR, 2014." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 296, + 271 + ], + "type": "text", + "content": "[63] Chen Tang, Ben Abbatematteo, Jiaheng Hu, Rohan Chandra, Roberto Martin-Martin, and Peter Stone. Deep reinforcement learning for robotics: A survey of real-world successes. arXiv preprint arXiv:2408.03539, 2024." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 272, + 296, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 272, + 296, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 272, + 296, + 348 + ], + "type": "text", + "content": "[64] Octo Model Team, Dibya Ghosh, Homer Walke, Karl Pertsch, Kevin Black, Oier Mees, Sudeep Dasari, Joey Hejna, Tobias Kreiman, Charles Xu, Jianlan Luo, You Liang Tan, Lawrence Yunliang Chen, Pannag Sanketi, Quan Vuong, Ted Xiao, Dorsa Sadigh, Chelsea Finn, and Sergey Levine. Octo: An open-source generalist robot policy. arXiv preprint arXiv:2405.12213, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 350, + 296, + 371 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 296, + 371 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 296, + 371 + ], + "type": "text", + "content": "[65] Florian Tramér. Detecting adversarial examples is (nearly) as hard as classifying them. In ICML, 2022." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 372, + 296, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 372, + 296, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 372, + 296, + 403 + ], + "type": "text", + "content": "[66] Florian Tramér, Nicholas Carlini, Wieland Brendel, and Aleksander Madry. On adaptive attacks to adversarial example defenses. In NeurIPS, 2020." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 296, + 448 + ], + "type": "text", + "content": "[67] Valeria Villani, Fabio Pini, Francesco Leali, and Cristian Secchi. Survey on human-robot collaboration in industrial settings: Safety, intuitive interfaces and applications. Mechatronics, 55:248-266, 2018." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 449, + 296, + 481 + ], + "type": "text", + "content": "[68] Christian Vogel, Christoph Walter, and Norbert Elkmann. A projection-based sensor system for safe physical human-robot collaboration. In IROS, 2013." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "spans": [ + { + "bbox": [ + 56, + 482, + 296, + 536 + ], + "type": "text", + "content": "[69] Homer Rich Walke, Kevin Black, Tony Z. Zhao, Quan Vuong, Chongyi Zheng, Philippe Hansen-Estruch, Andre Wang He, Vivek Myers, Moo Jin Kim, Max Du, Abraham Lee, Kuan Fang, Chelsea Finn, and Sergey Levine. Bridgedata V2: A dataset for robot learning at scale. In CoRL, 2023." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "[70] Taowen Wang, Yiyang Liu, James Chenhao Liang, Yiming Cui, Yuning Mao, Shaoliang Nie, Jiahao Liu, Fuli Feng, Zenglin Xu, Cheng Han, et al. " + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "inline_equation", + "content": "\\mathbf{M}^2" + }, + { + "bbox": [ + 56, + 537, + 296, + 581 + ], + "type": "text", + "content": "pt: Multimodal prompt tuning for zero-shot instruction learning. In EMNLP, 2024." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 296, + 624 + ], + "type": "text", + "content": "[71] Zhibo Wang, Siyan Zheng, Mengkai Song, Qian Wang, Alireza Rahimpour, and Hairong Qi. advpattern: Physical-world attacks on deep person re-identification via adversarially transformable patterns. In ICCV, 2019." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 296, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 296, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 296, + 667 + ], + "type": "text", + "content": "[72] Zixiang Wang, Hao Yan, Zhuoyue Wang, Zhengjia Xu, Zhizhong Wu, and Yining Wang. Research on autonomous robots navigation based on reinforcement learning. In RAIIC, 2024." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 296, + 713 + ], + "type": "text", + "content": "[73] Zhijie Wang, Zhehua Zhou, Jiayang Song, Yuheng Huang, Zhan Shu, and Lei Ma. Towards testing and evaluating vision-language-action models for robotic manipulation: An empirical study. arXiv preprint arXiv:2409.12894, 2024." + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 556, + 664 + ], + "type": "list", + "angle": 0, + "index": 28, + "blocks": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 555, + 128 + ], + "type": "text", + "content": "[74] Junjie Wen, Yichen Zhu, Jinming Li, Minjie Zhu, Kun Wu, Zhiyuan Xu, Ran Cheng, Chaomin Shen, Yaxin Peng, Feifei Feng, et al. Tinyvla: Towards fast, data-efficient vision-language-action models for robotic manipulation. arXiv preprint arXiv:2409.12514, 2024." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 129, + 556, + 173 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 556, + 173 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 556, + 173 + ], + "type": "text", + "content": "[75] Wikipedia contributors. Finch (film) — Wikipedia, the free encyclopedia. https://en.wikipedia.org/w/index.php?title=Finch_(film) & oldid=1249481595, 2024. [Online; accessed 5-November-2024]." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 174, + 555, + 207 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 555, + 207 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 555, + 207 + ], + "type": "text", + "content": "[76] Chen Henry Wu, Jing Yu Koh, Ruslan Salakhutdinov, Daniel Fried, and Aditi Raghunathan. Adversarial attacks on multimodal agents. arXiv preprint arXiv:2406.12814, 2024." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 208, + 555, + 251 + ], + "type": "text", + "content": "[77] Kaidi Xu, Gaoyuan Zhang, Sijia Liu, Quanfu Fan, Mengshu Sun, Hongge Chen, Pin-Yu Chen, Yanzhi Wang, and Xue Lin. Adversarial t-shirt! evading person detectors in a physical world. In ECCV, 2020." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "spans": [ + { + "bbox": [ + 316, + 252, + 555, + 285 + ], + "type": "text", + "content": "[78] Weilin Xu, David Evans, and Yanjun Qi. Feature squeezing: Detecting adversarial examples in deep neural networks. In NDSS, 2018." + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "spans": [ + { + "bbox": [ + 316, + 286, + 555, + 330 + ], + "type": "text", + "content": "[79] Zhiyang Xu, Chao Feng, Rulin Shao, Trevor Ashby, Ying Shen, Di Jin, Yu Cheng, Qifan Wang, and Lifu Huang. Visionplan: Scaling human-labeled tasks in visual instruction tuning. arXiv preprint arXiv:2402.11690, 2024." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 331, + 555, + 363 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 331, + 555, + 363 + ], + "spans": [ + { + "bbox": [ + 316, + 331, + 555, + 363 + ], + "type": "text", + "content": "[80] Yuchen Zhang and Percy Liang. Defending against whitebox adversarial attacks via randomized discretization. In AISTATS, 2019." + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 365, + 555, + 398 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 365, + 555, + 398 + ], + "spans": [ + { + "bbox": [ + 316, + 365, + 555, + 398 + ], + "type": "text", + "content": "[81] Yaoyuan Zhang, Yu-an Tan, Tian Chen, Xinrui Liu, Quanxin Zhang, and Yuanzhang Li. Enhancing the transferability of adversarial examples with random patch. In IJCAI, 2022." + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 316, + 399, + 555, + 430 + ], + "type": "text", + "content": "[82] Bo Zhao, Boya Wu, Muyang He, and Tiejun Huang. Svit: Scaling up visual instruction tuning. arXiv preprint arXiv:2307.04087, 2023." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "spans": [ + { + "bbox": [ + 316, + 432, + 555, + 476 + ], + "type": "text", + "content": "[83] Yuanyang Zhu, Zhi Wang, Chunlin Chen, and Daoyi Dong. Rule-based reinforcement learning for efficient robot navigation with space reduction. IEEE/ASME Trans. on Mechatronics, 27:846-857, 2021." + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 477, + 556, + 664 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 477, + 556, + 664 + ], + "spans": [ + { + "bbox": [ + 316, + 477, + 556, + 664 + ], + "type": "text", + "content": "[84] Brianna Zitkovich, Tianhe Yu, Sichun Xu, Peng Xu, Ted Xiao, Fei Xia, Jialin Wu, Paul Wohlhart, Stefan Welker, Ayzaan Wahid, Quan Vuong, Vincent Vanhoucke, Huong T. Tran, Radu Soricut, Anikait Singh, Jaspiar Singh, Pierre Sermanet, Pannag R. Sanketi, Grecia Salazar, Michael S. Ryoo, Krista Reymann, Kanishka Rao, Karl Pertsch, Igor Mordatch, Henryk Michalewski, Yao Lu, Sergey Levine, Lisa Lee, Tsang-Wei Edward Lee, Isabel Leal, Yuheng Kuang, Dmitry Kalashnikov, Ryan Julian, Nikhil J. Joshi, Alex Irpan, Brian Ichter, Jasmine Hsu, Alexander Herzog, Karol Hausman, Keerthana Gopalakrishnan, Chuyuan Fu, Pete Florence, Chelsea Finn, Kumar Avinava Dubey, Danny Driess, Tianli Ding, Krzysztof Marcin Choromanski, Xi Chen, Yevgen Chebotar, Justice Carbajal, Noah Brown, Anthony Brohan, Montserrat Gonzalez Arenas, and Kehang Han. RT-2: vision-language-action models transfer web knowledge to robotic control. In CoRL, 2023." + } + ] + } + ], + "index": 27 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "6958" + } + ] + } + ], + "index": 29 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_content_list.json b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..16149bd53c7948dc8193d0c65b37db4f815967ff --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_content_list.json @@ -0,0 +1,2362 @@ +[ + { + "type": "text", + "text": "Expressive Talking Human from Single-Image with Imperfect Priors", + "text_level": 1, + "bbox": [ + 153, + 130, + 844, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Jun Xiang $^{1,2}$", + "bbox": [ + 217, + 179, + 323, + 199 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yudong Guo $^{1*}$", + "bbox": [ + 362, + 180, + 480, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Yancheng Yuan2", + "bbox": [ + 351, + 199, + 483, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1University of Science and Technology of China", + "bbox": [ + 215, + 217, + 504, + 233 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Leipeng Hu", + "bbox": [ + 519, + 180, + 624, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Boyang Guo", + "bbox": [ + 665, + 180, + 774, + 198 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Juyong Zhang", + "bbox": [ + 522, + 199, + 643, + 217 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ The Hong Kong Polytechnic University", + "bbox": [ + 532, + 218, + 777, + 233 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/c16c23a4569ecbb3eeaed8bf07a5e9408122317482dbf5661749dc23a15510a8.jpg", + "image_caption": [ + "One-Shot Photos", + "Figure 1. Given a one-shot image (e.g., your favorite photo) as input, our method reconstructs a fully expressive whole-body talking avatar that captures personalized details and supports realistic animation, including vivid body gestures and natural expression changes." + ], + "image_footnote": [], + "bbox": [ + 94, + 273, + 338, + 484 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/64c18b8bcd57950a232f1518eae3b0bb4624793e605782833db13b035c89dac8.jpg", + "image_caption": [ + "Animatable Expressive Talking Avatars" + ], + "image_footnote": [], + "bbox": [ + 344, + 271, + 906, + 484 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 248, + 559, + 326, + 574 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Building realistic and animatable avatars still requires minutes of multi-view or monocular self-rotating videos, and most methods lack precise control over gestures and expressions. To push this boundary, we address the challenge of constructing a whole-body talking avatar from a single image. We propose a novel pipeline that tackles two critical issues: 1) complex dynamic modeling and 2) generalization to novel gestures and expressions. To achieve seamless generalization, we leverage recent pose-guided image-to-video diffusion models to generate imperfect video frames as pseudo-labels. To overcome the dynamic modeling challenge posed by inconsistent and noisy pseudo-frames, we introduce a tightly coupled 3DGS-mesh hybrid avatar representation and apply several key regularizations to mitigate inconsistencies caused by imperfect labels. Extensive experiments on diverse subjects demonstrate that our method enables the creation of a photorealistic, precisely animatable, and expressive whole-body", + "bbox": [ + 88, + 593, + 485, + 867 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "talking avatar from just a single image. Project page: https://ustc3dv.github.io/OneShotOneTalk/.", + "bbox": [ + 511, + 561, + 906, + 590 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 513, + 618, + 645, + 633 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Realistic rendering and precise control of gestures and expressions for whole-body talking avatars hold significant potential for AR/VR applications, such as telepresence and immersive remote conferencing. While extensively studied in the fields of 3D Vision and Graphics, most production-level body avatars still require a light stage [11, 65] to capture hours or minutes of multi-view frames, making them unsuitable for typical consumer usage. To address this, we focus on the challenging task of constructing an animatable whole-body avatar that supports both photo-realistic rendering and precise 3D control of gestures and expressions, from only a single image. This one-shot pipeline primarily faces the following challenges.", + "bbox": [ + 509, + 643, + 906, + 839 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Complex dynamic modeling. Humans exhibit complex gestures and facial movements when communicating with each other. To model the motion space of the entire body, SMPL-X [46] integrates several previous mod", + "bbox": [ + 511, + 839, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Corresponding author: Yudong Guo.", + "bbox": [ + 91, + 886, + 294, + 900 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "10398", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "els [37, 40, 54] to capture body, hand, and facial movements. With the help of these parametric geometric models, recent personalized human avatars learn to model dynamic geometry and appearance by integrating neural textures [13, 38], hybrid representation [5, 17, 27], neural radiance fields [29, 47, 49, 64, 79], and 3D Gaussians [20, 22, 33, 35, 51] for dynamic photo-realistic rendering. To capture the full appearance of a whole-body avatar, these methods require complete observations, typically relying on dense input data for supervision, such as multi-view videos or self-rotating monocular videos. Moreover, these approaches primarily focus on body motion and novel view synthesis, falling short in accurately capturing and animating diverse facial expressions and hand movements, which limits their practical applicability. Due to the reliance on dense inputs, realistic and expressive whole-body modeling from a single image remains an unsolved challenge in this field.", + "bbox": [ + 89, + 90, + 480, + 361 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Generalization to novel gestures and expressions. Another key challenge lies in the limited ability to generalize to a wide variety of gestures and facial expressions, especially those that deviate significantly from the poses seen during training. This limitation arises from the data-driven nature of current methods, where the training data typically encompasses only a finite set of gestures and expressions. Consequently, when animating avatars, regions or motions that were underrepresented in the training data—such as dynamic clothing, inner mouth movements, or intricate hand gestures—are often poorly synthesized or omitted entirely. These issues are further compounded when working with a single image input, as the limited visual cues often fail to capture the full range of motion and expression. Recently, several approaches [19, 59, 62, 63, 71, 77, 83] have leveraged image and video diffusion models [3, 53] to achieve image-to-video generation guided by whole-body landmarks [73]. Although various model architectures and training techniques have been explored to improve generation quality, temporal consistency remains a challenge, and these methods have several key limitations. For example, body landmarks fail to disentangle identity and pose, leading to poor cross-identity animation results, with facial and body distortions and identity mismatches. Moreover, the sparse 2D landmarks are insufficient for precisely controlling gestures and facial movements, as they do not accurately capture the underlying dynamic geometry.", + "bbox": [ + 91, + 368, + 482, + 775 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In this work, we propose a novel pipeline to address both challenges in the one-shot setting. To generalize to diverse gestures and facial movements, we preprocess the large-scale TED Gesture Dataset [74] to build a comprehensive whole-body motion space for people talking. We then use these motion sequences as guidance to drive the input single image with a pre-trained whole-body video diffusion model [77] and a 3D face animation model [7]. This", + "bbox": [ + 89, + 780, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enables us to generate various video sequences of the target person performing different gestures and expressions. However, directly using these pseudo-frames to train a body avatar leads to unsatisfactory results. First, as noted above, current diffusion-based body animation methods struggle with temporal consistency, identity preservation, and motion alignment. Directly employing them as supervision results in significant blurring, distortion, and identity degradation. Furthermore, as we adapt two individual approaches for generating pseudo-frames for body gestures and facial expressions, respectively. Due to inconsistencies in camera spaces and rendering procedures, merging them into a unified avatar representation poses additional challenges.", + "bbox": [ + 511, + 90, + 903, + 287 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To tackle the challenge of complex dynamic modeling, we utilize both the single input image and imperfect pseudoframes to train a hybrid mesh-3DGS avatar representation, constrained by several carefully designed regularizations. The single input image provides an accurate, though incomplete, appearance for the body avatar, while the pseudo videos offer imperfect but more complete visual cues. For the pseudo video labels, instead of using per-pixel losses, we employ a perceptual-based loss term [76], which helps achieve reasonable appearance modeling while alleviating misalignment in the pseudo-frames.", + "bbox": [ + 511, + 287, + 903, + 452 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To further alleviate the inconsistencies caused by the pseudo-frames, we adopt a tightly coupled mesh-3DGS hybrid avatar representation. By introducing Laplacian smoothing and normal-consistency regularization on the deformed body mesh, we ensure that the structure of the 3D Gaussians used for rendering is well-constrained. Finally, we supervise the avatar representation using both gesture and head pseudo-frames, enabling the creation of a photorealistic, precisely animatable, and expressive whole-body avatar.", + "bbox": [ + 511, + 454, + 903, + 602 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In summary, our contributions include the following aspects:", + "bbox": [ + 511, + 604, + 903, + 633 + ], + "page_idx": 1 + }, + { + "type": "list", + "sub_type": "text", + "list_items": [ + "- We introduce a novel pipeline that overcomes the key challenges of building a whole-body expressive talking avatar from a single image.", + "- Our diffusion guidance strategy effectively extracts valuable knowledge from imperfect diffusion outputs and combines it with the limited information from the input image, enabling complete modeling of the talking avatar.", + "- Our carefully designed 3DGS-mesh coupled avatar representation, along with essential regularization techniques, facilitates accurate modeling of diverse subjects and stabilizes the optimization process." + ], + "bbox": [ + 513, + 635, + 903, + 800 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 511, + 814, + 653, + 830 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Human Gaussian Splatting. 3D Gaussian Splatting [31] is the state-of-the-art method for scene reconstruction and novel view synthesis (NVS), offering superior rendering speed and visual quality. This has significantly influ", + "bbox": [ + 511, + 839, + 903, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "10399", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "enced human avatar studies. Methods using multi-view videos [30, 43, 45, 51, 85] demonstrate excellent performance, with unique designs such as ASH [45], which achieves efficient Gaussian learning via mesh UV parameterization. For monocular video input, most human Gaussian splatting methods [18, 20, 22, 35, 42, 57] link Gaussian fields to parametric mesh models, often using additional regularization terms. ExAvatar [42] applies connectivity-based regularizers to short, casually captured videos. For sparse-view images, GPS-Gaussian [81] achieves real-time human NVs by encoding human priors into the network, while HumanSplat [44] generates high-quality static reconstructions from one-shot inputs. Concurrent works [52, 61] introduce multi-view diffusion models but are limited to the head or body areas. IDOL [84] leverages a large-scale dataset and shows great performance in single-view human reconstruction. However, it still struggles with handling facial dynamic. By contrast, our method is capable of recovering a realistic, animatable whole-body talking avatar from just a single image.", + "bbox": [ + 89, + 90, + 480, + 393 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Avatar Reconstruction from Few-Shot Images. Some works [8, 28, 67] apply 3D GAN inversion for one-shot human reconstruction, but they struggle with preserving personal details and generalization. PIFu [55] and subsequent works [6, 16, 56, 68, 69, 82] introduce pixel-aligned features and neural fields for image-based human reconstruction. An alternative approach leverages diffusion priors to fill in missing details, such as training human-centered diffusion models [2, 15, 44], using novel-view diffusion results for additional supervision [36, 39, 78], and employing Score Distillation Sampling (SDS) [50] to generate 3D avatars from 2D priors [24, 70, 72, 75]. However, these methods focus on static scenes and overlook dynamic human motion, limiting their ability to capture human dynamic. The SMPL-based [40] representation enables ELICIT [23] and SHERF [21] to model human dynamic, but overlooks hand and facial motions, restricting expressive animation capabilities. In contrast, our approach leverages priors from a pose-guided human video diffusion model, capturing both human appearance and dynamic, and enabling expressive full-body animations, particularly in the hands and face.", + "bbox": [ + 91, + 398, + 482, + 729 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Pose-Guided Human Video Diffusion. Pose-guided human video diffusion models [9, 19, 59, 63, 71, 77, 83] directly generate animated videos from a reference image and pose sequence, bypassing traditional 3D reconstruction and rendering processes. The success of these models depends on the quality of training data, model design, and pose guidance. Some approaches [9, 19, 71, 83] incorporate temporal layers to ensure smooth transitions, inspired byAnimateDiff [12], while others [48, 59, 77] use video diffusion models [3, 4] for dynamic sequences. Pose guidance is typically provided by OpenPose [19, 63, 77], DensePose [71],", + "bbox": [ + 89, + 734, + 482, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "depth maps [9], or SMPL [83]. StableAnimator [62] proposes HJB-based face optimization that further enhances face quality, and Make-Your-Anchors [25] personalizes outputs by fine-tuning models on identity-specific images. We adopt MimicMotion [77] for its superior performance in handling hand regions. Despite their strengths, these 2D models still face challenges, such as image distortion, identity changes, and pose misalignment, due to the lack of 3D understanding. To address these, we leverage optimized avatar representations with carefully designed constraints, improving consistency and naturalness in the generated animations.", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 511, + 286, + 604, + 301 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Given a single image of the target person, we aim to reconstruct a 3D talking avatar that fully inherits the identity and enables natural animation. To address the challenge of complex dynamic modeling from imperfect pseudo-frames, we adopt a tightly coupled 3DGS-mesh hybrid avatar representation (Sec. 3.1). To generalize well to diverse gestures and facial movements, we generate imperfect sequences of the target person driven by various motion sequences (Sec. 3.2). Finally, we introduce the carefully designed constraints and loss terms to train the representation from noisy videos effectively (Sec. 3.3). The entire pipeline is illustrated in Fig. 2.", + "bbox": [ + 511, + 311, + 906, + 492 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.1. Coupled 3DGS-Mesh Avatar", + "text_level": 1, + "bbox": [ + 511, + 501, + 769, + 517 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Whole-body parametric mesh models facilitate human animation and provide good initialization, while 3DGS offers enhanced expressiveness and realistic rendering. To address the challenges of the one-shot task, we design a novel coupled 3DGS-mesh representation, which effectively integrates the geometric priors and surface regularization of the mesh without diminishing the expressive capability of the Gaussian field.", + "bbox": [ + 511, + 523, + 903, + 642 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "We couple the 3DGS field with the typical SMPL-X model, which is formulated as follows:", + "bbox": [ + 511, + 643, + 903, + 672 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nM (\\beta , \\theta , \\phi) = W (T (\\beta , \\theta , \\phi), J (\\beta), \\theta , \\mathcal {W}), \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 685, + 903, + 702 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\nT (\\beta , \\theta , \\phi) = \\bar {T} + B _ {S} (\\beta) + B _ {E} (\\phi) + B _ {P} (\\theta), \\tag {2}\n$$\n", + "text_format": "latex", + "bbox": [ + 557, + 726, + 903, + 743 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $B_{S}(\\cdot), B_{E}(\\cdot)$ and $B_{P}(\\cdot)$ denote shape, expression, and pose blend functions respectively, with $\\beta, \\phi$ and $\\theta$ representing the corresponding parameters. $\\bar{T}$ is the template mesh, and $W(\\cdot)$ is the standard LBS function that rotates the vertices in $T$ around the estimated joints $J$ , smoothed by the blend weights $\\mathcal{W}$ . Since the shape code $\\beta$ is fixed once registered, we will omit it in the later statements and denote $T = T(\\beta, \\theta, \\phi)$ and $J = J(\\beta)$ for simplicity. Inspired by [1, 45, 66], we initialize the 3D Gaussians on the canonical mesh surface using UV parameterization. For a", + "bbox": [ + 511, + 750, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "10400", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/136c8b08ef624783e8463e48c14d08b62a627a88e3190fa1058e033d6fd3d924.jpg", + "image_caption": [ + "Figure 2. Overview. Our method constructs an expressive whole-body talking avatar from a single image. We begin by generating pseudo body and head frames using pre-trained generative models, driven by a collected video dataset with diverse poses. Per-pixel supervision on the input image, perceptual supervision on imperfect pseudo-frames, and mesh-related constraints are then applied to guide the 3DGS-mesh coupled avatar representation, ensuring realistic and expressive avatar reconstruction and animation." + ], + "image_footnote": [], + "bbox": [ + 96, + 92, + 898, + 364 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3D Gaussian located on a triangle $k = \\{\\mathbf{V}_1, \\mathbf{V}_2, \\mathbf{V}_3\\}$ of mesh $T$ with barycentric coordinates $(u, v)$ , the surface position is given by:", + "bbox": [ + 89, + 458, + 483, + 505 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\operatorname {B a r y} (T) = \\mathcal {V} (k, u, v) = u \\mathbf {V} _ {\\mathbf {1}} + v \\mathbf {V} _ {\\mathbf {2}} + (1 - u - v) \\mathbf {V} _ {\\mathbf {3}}. \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 99, + 513, + 482, + 531 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Coupled 3DGS-Mesh Deformation. To model clothes, haircuts, and other complex regions that the SMPL-X mesh fails to handle, the deformation of 3D Gaussians is crucial. While previous works can achieve correct deformation with sufficient supervision, we need to impose additional constraints on the deformation field to prevent overfitting to the single input image and imperfect pseudo-frames, while also facilitating the integration of other modules. Based on these considerations, we design two deformation fields in the canonical space and enforce their proximity: one represents the conventional Gaussian deformation, and the other represents the critical mesh deformation. This approach allows us to indirectly influence the Gaussian deformation by applying soft constraints on the deformed mesh. Moreover, since there is no strict binding between the deformed mesh surface and the Gaussians, complex regions are still effectively handled. The two deformation fields, together with the full-body animation and Gaussian Splitting process, are formulated as follows:", + "bbox": [ + 89, + 541, + 483, + 828 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG (\\theta , \\phi , \\mathbf {P}) = S P (W (B a r y (T) + d X, J, \\theta , B a r y (\\mathcal {W})), \\mathbf {P}). \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 89, + 839, + 483, + 871 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nM (\\theta , \\phi) = W (T + d T, J, \\theta , \\mathcal {W}), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 174, + 883, + 482, + 902 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $G(\\cdot)$ represents the final rendered image, $SP(\\cdot)$ and $\\mathbf{P}$ denote the splatting rendering process and the properties of 3D Gaussians, and $dX$ and $dT$ represent the optimized deformations of the Gaussians and mesh vertices, respectively. We apply key regularizers on $dT$ and propagate soft constraints to $dX$ by encouraging $dX$ to align with $Bary(dT)$ .", + "bbox": [ + 511, + 458, + 906, + 566 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Pseudo-Frames Generation", + "text_level": 1, + "bbox": [ + 511, + 575, + 759, + 589 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the one-shot image setting, many regions are unseen or occluded. To construct a complete avatar and ensure generalization to novel gestures and expressions, we turn our attention to recent advances in human motion diffusion models [19, 77] and head animation techniques [7, 10].", + "bbox": [ + 511, + 598, + 906, + 672 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "For leveraging diffusion-based generative models, the SDS loss [50] is widely used in text/image-to-3D works [24, 60, 72, 75]. Although diffusion models may introduce moderate deviations, the SDS technique remains effective in these studies. However, for tasks such as realistic dynamic avatar animation from a one-shot image, pose alignment is critical [18, 42], and the misalignment introduced by 2D diffusion models cannot be overlooked. Therefore, directly applying SDS in our setting is suboptimal. Instead, we focus on extracting accurate information from pseudo-frames synthesized by these generative models.", + "bbox": [ + 511, + 674, + 908, + 839 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To generate whole-body pseudo-frames, we collect a set of SMPL-X pose sequences $\\{(\\theta_{N_i},\\phi_{N_i})_{i = 1}^F\\}_{j = 1}^K$ from the TED Gesture Dataset [74] as input to the generative models. Since no unified generative model exists for both body and", + "bbox": [ + 511, + 840, + 908, + 902 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "10401", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/d060dc60fdd93b5cfc853f8f1c7920ab535fa33c8a270767db6dad919d8e1634.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 94, + 282, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fa99557ef1b3e52b7ecce646f40d38e31ed3bbae2a4c532fd8b26a74f394f931.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 112, + 233, + 169, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/dc5746401070b29c558c127a26c625f1906bad85c1790a4ed23eb00c1571576f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 192, + 227, + 279, + 358 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4a1241aa7d6933c48913f53f7cb4c94724184651c5bfad9471798eb075aabe0f.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 282, + 94, + 346, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/8c6d967c4b15c3c87b2c7bd4dec887c95f4fe9ae1d8e7d672a3c063f7ff2fab4.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 284, + 148, + 346, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/4eaab04a0012cae545c06daefa546ce43b242f18efe28f6376ce16d097c46287.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 367, + 95, + 415, + 222 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/f41e91d1723bab134428e064568faa9309a028d0082080b22aee1f20ee478750.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 349, + 228, + 431, + 358 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/77fa1ef8cfc5c11e6e54eb6fc8bc32501396ac5eab30731991ef1c3d09185311.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 437, + 95, + 593, + 223 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/6e2f399b3ae495ed6e4747d3cd740ab0bab5bb3442ecfc9200cf3c4e74f81d09.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 439, + 227, + 501, + 276 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ce4303bb171cd4202d2c921a368427dcd6d997659aef6e71f430e22680c811fd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 501, + 227, + 589, + 358 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/5134cb783d35dac4a79c5a83f194f78f74fcc0764c90609a7391fee3d9484042.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 95, + 658, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/37e59c5f3e0eb9c9a0f70969bcceb18348ded7f1fad84121cb415c055fdeaec2.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 148, + 658, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/3a383c78e917258b4da566b6ce601465c703c962949a9091542b5c64a7fa5af0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 95, + 732, + 222 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a636c897480b0534ce5c8b847029b1fb418e05e8d86ed0f6e3612f3d3ec48ee0.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 596, + 227, + 743, + 358 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c4921b9a9dfd82c7dfe992d4b3ff6f4b05a326e356c2dd5e889a108b8b60e326.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 95, + 815, + 143 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/82e338918f226678a8764f87d87b32e758ec31a7b3bd019625da8cfa4a12697e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 150, + 815, + 196 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b11488b81a7d183fc2d9156071a7680ebc12986c20a63bd0dc957fb56e8adb98.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 833, + 95, + 883, + 220 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/c8b9ccd8fd22c8e0d514f531e08c7c03fc013d004967758bfe96f6ea801b24a5.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 754, + 227, + 815, + 275 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/d97f9030bf1e648ce910525a839387844c6dbd2fb54d26d7b0d96d97ad2264ec.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 816, + 228, + 898, + 356 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/fbe806b5bc89ac407db8b293fe92098464e8941abf07cdfcbc1add316ec4eccb.jpg", + "image_caption": [ + "Input Image", + "Driving Image", + "Figure 3. Qualitative comparisons with representative methods [23, 42, 62] in the cross-identity motion reenactment task. Our method achieves accurate and realistic animation with almost all fine details preserved and identity unchanged." + ], + "image_footnote": [], + "bbox": [ + 99, + 368, + 281, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/a53440c716403b9ee9a262321371d9c588bc5e1d2ccb6262e79ad2a6b5cccab0.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 297, + 368, + 426, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/bcaa1bcd7e6b534197cbfe1f07683ad7a36b4ae8cf54cfec5812a81cf3fc1a63.jpg", + "image_caption": [ + "StableAnimator" + ], + "image_footnote": [], + "bbox": [ + 450, + 368, + 578, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/b197a1cab344f42dcfaf217c3357ad5bc9c21306589dc0dacf50a081ad4f203a.jpg", + "image_caption": [ + "ExAvatar-1shot" + ], + "image_footnote": [], + "bbox": [ + 611, + 368, + 741, + 465 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/ee77569f867464bf3251e8c18cf2febd89b461511a6f71b88f540cee2914c579.jpg", + "image_caption": [ + "ELICIT" + ], + "image_footnote": [], + "bbox": [ + 769, + 368, + 897, + 465 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "head simultaneously, we employ two separate approaches to generate pseudo-frames for body gestures and facial expressions, respectively. Given a source image $I_{S}$ registered with the SMPL-X parameters $(\\beta_{S},\\theta_{S},\\phi_{S})$ , we adopt MimicMotion [77] with the collected pose sequences to generate various pseudo body frames:", + "bbox": [ + 89, + 540, + 482, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nI _ {N} = \\operatorname {M o t i o n} \\left(I _ {S}, D \\left(\\beta_ {S}, \\theta_ {N}, \\phi_ {N}\\right)\\right), \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 163, + 637, + 482, + 655 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $D(\\cdot)$ denotes the mapping from SMPL-X parameters to DWPose [73]. We randomly set the root body pose in $\\theta_N$ for each segment to increase viewpoint generalization, with pitch $\\in (-30^{\\circ}, 30^{\\circ})$ and yaw $\\in (-10^{\\circ}, 10^{\\circ})$ . After that, we re-track $I_N$ to obtain more accurate pose parameters $(\\hat{\\theta}_N, \\hat{\\phi}_N)$ , while keeping the shape code $\\beta_S$ fixed. The re-tracking process is crucial, as it overcomes the misalignment introduced by 2D diffusion models and helps achieve precise and realistic results. For the head region, we adopt Portrait4D-v2 [7] to generate various pseudo frames of the target person performing diverse expressions, also driven by the videos from the TED Gesture Dataset.", + "bbox": [ + 89, + 659, + 482, + 840 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "3.3. Objective Functions", + "text_level": 1, + "bbox": [ + 89, + 848, + 282, + 864 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Assisted by our novel hybrid 3DGS-mesh avatar representation, we introduce several carefully designed regularization", + "bbox": [ + 89, + 869, + 482, + 901 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "terms to stabilize the avatar reconstruction process and effectively extract the correct information from both the one-shot input and the imperfect pseudo-frames.", + "bbox": [ + 511, + 540, + 906, + 585 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Mesh-related Constraints. We use the following loss functions to apply soft constraints on the Gaussian field based on the mesh:", + "bbox": [ + 511, + 585, + 906, + 630 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} \\mathcal {L} _ {S C} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {M G C}} \\mathcal {L} _ {\\mathrm {M G C}} + \\lambda_ {\\mathrm {l a p}} \\mathcal {L} _ {\\mathrm {l a p}}. \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 607, + 638, + 808, + 672 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "(7)", + "bbox": [ + 883, + 648, + 903, + 662 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Here, $\\mathcal{L}_{\\mathrm{normal}}$ is the normal consistency loss applied to the deformed mesh surface, ensuring surface normal consistency post-deformation. $\\mathcal{L}_M$ is the mask loss, which measures the discrepancy between the ground truth mask and the mask of the deformed mesh rendered via [34]. These two losses work together to regulate the behavior of $dT$ , while influencing Gaussian deformation $dX$ through the mesh-Gaussian consistency loss:", + "bbox": [ + 511, + 681, + 906, + 803 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\mathrm {M G C}} = \\| d X - B a r y (d T) \\| _ {1}. \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 604, + 813, + 903, + 830 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Additionally, we compute the Laplacian smoothing loss $\\mathcal{L}_{\\mathrm{lap}}$ for $dX$ along with the scaling and RGBs of the 3D Gaussians, based on their initial connectivity state on the canonical mesh surface.", + "bbox": [ + 511, + 840, + 906, + 898 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "10402", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Perceptual Supervision of Pseudo-Frames. The synthesized pseudo-frames exhibit noticeable artifacts, such as image distortion and lack of 3D consistency, especially in the results generated with MimicMotion. These artifacts cannot be fully corrected through pose alignment alone. As a result, pixel-aligned losses like L1 or MSE often lead to issues like texture flickering, blurring, and identity changes. However, the high-level human structure is consistently well-preserved in the source image, which aids human recognition and is empirically referred to as perceptual similarity. Previous works [41, 76] have used deep neural networks to capture these perceptual structure features. We adopt the LPIPS perceptual loss [76], using VGG [58] as the backbone. By learning deep perceptual human features from dynamic poses and preserving detailed information from the source image, our method is able to conduct realistic and complete talking avatar animations. Thus, for the pseudoframes, we primarily use the following perceptual loss:", + "bbox": [ + 89, + 90, + 483, + 363 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {\\text {d i f f}} = \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {N}, G \\left(\\hat {\\theta} _ {N}, \\hat {\\phi} _ {N}, \\mathbf {P}\\right)\\right). \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 371, + 482, + 388 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Per-pixel Supervision of the Input Image. For the source image $I_S$ with pose $(\\theta_S, \\phi_S)$ , we use the regular L1 loss, SSIM loss, and mask loss between $I_S$ and the splatted image $G(\\theta_S, \\phi_S, \\mathbf{P})$ :", + "bbox": [ + 89, + 398, + 483, + 460 + ], + "page_idx": 5 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {S} = \\lambda_ {\\mathrm {R G B}} \\mathcal {L} _ {\\mathrm {R G B}} + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} + \\lambda_ {\\mathrm {a l p h a}} \\mathcal {L} _ {\\mathrm {a l p h a}}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 119, + 468, + 482, + 484 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "The final loss function is the sum of all the aforementioned losses.", + "bbox": [ + 89, + 493, + 483, + 523 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "3.4. Implementation Details", + "text_level": 1, + "bbox": [ + 89, + 532, + 308, + 547 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We set the UV map resolution to 512 and use approximately 150,000 Gaussians in total. We train our models using the Adam optimizer [32] with $\\beta = (0.9, 0.999)$ . The learning rates for the Gaussian parameters are the same as those in the official implementation, while the learning rates for both the Gaussian and mesh deformation fields are set to $\\eta = 1e^{-4}$ . For the loss weights, we set $\\lambda_{\\mathrm{normal}}$ , $\\lambda_{\\mathrm{M}}$ , $\\lambda_{\\mathrm{MGC}}$ , $\\lambda_{\\mathrm{lap}}$ , $\\lambda_{\\mathrm{LPIPS}}$ , $\\lambda_{\\mathrm{RGB}}$ , $\\lambda_{\\mathrm{SSIM}}$ , $\\lambda_{\\mathrm{alpha}}$ to $1e^{-2}$ , $1e^{-1}$ , $1e^{1}$ , $1e^{2}$ , $2e^{-1}$ , $8e^{-1}$ , $1e^{-1}$ , and $4e^{-1}$ , respectively. We start adding the perceptual loss at the 2000th step, while the other losses are used throughout the entire training process with a maximum number of training steps of 20000. We adopt joint training with separate supervision for body and head pseudo-frames. Training takes 3 hours on single 4090 due to critical but inefficient re-tracking. But avatar reconstruction and rendering are highly efficient with an inference speed of 50FPS+ at 1K resolution.", + "bbox": [ + 89, + 554, + 483, + 811 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 823, + 223, + 840 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.1. Dataset", + "text_level": 1, + "bbox": [ + 89, + 848, + 186, + 863 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We use the poses and expressions processed from 100 videos of the TED Gesture Dataset [74] as the motion", + "bbox": [ + 89, + 869, + 483, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sequences during training. We separately sampled about 1,800 frames with the most distinct variations in poses and expressions. For evaluation, the one-shot input and driving poses are primarily sourced from the ActorsHQ [26], NeuMan [29] and the Casual Conversations Dataset [14]. All videos and images are cropped and resized to a fixed aspect ratio of 9:16. For foreground segmentation, we use BiRefNet [80], and for pose tracking, we employ the custom fitting procedure from ExAvatar [42].", + "bbox": [ + 511, + 90, + 905, + 227 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "4.2. Comparison with Representative Methods", + "text_level": 1, + "bbox": [ + 511, + 239, + 872, + 256 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "We compare our method with several representative works, including: (1) ExAvatar [42], a recent SOTA work that models human avatars with a mesh-based 3D Gaussian field. (2) ELICIT [23], a one-shot NeRF-based animatable avatar. (3) StableAnimator [62], a SOTA general 2D pose-guided human video diffusion model. (4) Make-Your-Anchors [25], a 2D diffusion-based method that is fine-tuned on a one-minute video clip of the individual to enhance identity information.", + "bbox": [ + 511, + 262, + 905, + 397 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Note that for ExAvatar, since it takes short videos as input, we compare with two versions of it: ExAvatar-40shot, which uses 40-shot images, and ExAvatar-1shot, which uses one-shot images as input. For Make-Your-Ananchor, as we find it does not perform well on one-shot input, we only compare it with available video input by fine-tuning it on a short video clip.", + "bbox": [ + 511, + 398, + 905, + 503 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Qualitative Comparisons. Fig. 3 and Fig. 4 present a qualitative comparison between our method and the other representative approaches. For Fig. 3, we use the pose of a different identity as the driving signal. For Fig. 4, we compare the performance of these methods on the self-driven pose reenactment task, using subjects that have corresponding video data.", + "bbox": [ + 511, + 506, + 905, + 611 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "As observed, ExAvatar-1shot tends to overfit the input image and fails to recover accurate textures in regions occluded by hands. Even with 40-shot input, ExAvatar still struggles with incomplete knowledge and fails to handle novel gestures effectively. StableAnimator generates relatively reasonable results but is constrained by its training distribution and struggles with identity consistency across frames, often leading to appearance mismatches and identity changes. ELICIT, which uses a NeRF-based representation, ensures 3D consistency but relies on SMPL for geometry, which neglects the hand region. This coarse semantic proxy fails to support complex hand reconstruction or animation. Make-Your-AnAnchor, although pre-trained on multiple identities to learn human motion priors, requires a long video with sufficient movement and appearance data to adapt to new identities. It struggles with short fine-tuning videos and fails to recover fine details, especially for gestures outside its fine-tuning distribution.", + "bbox": [ + 511, + 612, + 905, + 883 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In contrast, our method generates expressive talking", + "bbox": [ + 531, + 885, + 905, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "10403", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b88b938fcf40d67c0b10fdeae381ff0e72dbf0628c6bb89c5b442f4b425570d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 98, + 89, + 251, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/461498c5fb7d9605afc9d9c8ca35e9599e577bf2b3d15ea52272df3c758b13d9.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 254, + 93, + 349, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f21d406557de3ecffb7058f27f10929fa928621f5494ca1310d8a66cf21b5d99.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 354, + 92, + 462, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/28a8cf84ba456eb9ea039e3d45c9367be52a28b7c430b7b21ec0f4355264cb27.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 467, + 93, + 565, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/99e84388e162bfebe18a5f3d40ac5f8d784ba4c19d7f2840e6961196ed978559.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 571, + 93, + 669, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/603e0207c2597979f3e0b266b2a3d0835e8a6649914f651ae6c822328f06fdaa.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 678, + 93, + 776, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/f2b48159c359213127748e10c4429d218473e4b0f66919f5755d40609f04d78b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 781, + 92, + 898, + 218 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8d1925219951917a259deea15803b4159199f78824f3b9c167b154303fe1ca06.jpg", + "image_caption": [ + "Input Image", + "Ground Truth" + ], + "image_footnote": [], + "bbox": [ + 98, + 228, + 251, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b80ac9d6ceecda1eb17aa851b2a3dbc00a354cf57015e734d516d82de5e2dd14.jpg", + "image_caption": [ + "Ours" + ], + "image_footnote": [], + "bbox": [ + 254, + 227, + 352, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/1097f9c3c141bcda207f1e39e9040ba895e4cfa234f218fc346f4f7e5038130f.jpg", + "image_caption": [ + "StableAnimator" + ], + "image_footnote": [], + "bbox": [ + 354, + 227, + 464, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a920911ef28c9476a6151cbaa5b50d642c681765b59fc218a48a08294cc93791.jpg", + "image_caption": [ + "ExAvatar-40shot" + ], + "image_footnote": [], + "bbox": [ + 467, + 226, + 566, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25adc51e724a59fe944e88ac02cb50868217549157043c4d43e3e3c1a000be86.jpg", + "image_caption": [ + "ExAvatar-1shot" + ], + "image_footnote": [], + "bbox": [ + 570, + 227, + 671, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/0eb6d0a8e929f7e2c12a04ef6aa26fedc961025d4895aca9b6610f10e9db7284.jpg", + "image_caption": [ + "ELICIT" + ], + "image_footnote": [], + "bbox": [ + 678, + 227, + 779, + 357 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/b50d0db4bb81c52a06542b2c337530d79305a08481a0b3aff61bfd1a4021628d.jpg", + "image_caption": [ + "Make-Your-Anchors" + ], + "image_footnote": [], + "bbox": [ + 782, + 227, + 898, + 357 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/af943ac667e3478e64e3d11ac72c69662b8a0a13f62b04d0be13ebf7469e7200.jpg", + "table_caption": [], + "table_footnote": [], + "table_body": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
ELICIT5.65522.5590.9210.0660
ExAvatar3.93124.2150.9240.0409
StableAnimator1.88327.3500.9350.0327
Ours1.21629.3110.9430.0299
", + "bbox": [ + 94, + 435, + 478, + 515 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 1. Quantitative comparisons with representative methods [23, 42, 62] on self-driven data. Our method outperforms others in pixel-wise error metrics, realism evaluation metrics and perceptual quality metrics. (ExAvatar here denotes ExAvatar-40shot.)", + "bbox": [ + 89, + 526, + 482, + 583 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "avatars from a single input image. Our results preserve fine human details and achieve natural animation with excellent rendering quality. We provide additional examples of cross-identity pose reenactment in Fig. 5. Using the same driving pose, identities with completely different attributes can be driven in the same way, thanks to the SMPL-X model and the 3DGS-mesh coupled representation.", + "bbox": [ + 89, + 595, + 483, + 700 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Quantitative Comparison. Tab. 1 presents a quantitative comparison between our model and other methods. Although the one-shot image animation task typically lacks a strict test set for quantitative comparison, we perform the evaluation on the self-driven task and use four common metrics for comparison: Mean Squared Error, PSNR, SSIM, and LPIPS. Our method outperforms all others across these metrics, demonstrating superior realism and 3D consistency in the results. However, we believe that these metrics do not fully capture the quality or capabilities of the methods, especially for the one-shot image animation task. We encourage readers to refer to the video results of our method for a more comprehensive and objective evaluation.", + "bbox": [ + 89, + 704, + 483, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c28295904109f88173963caecf54b2d211b0e4b1f0ecaa2b55b2912765d94f1e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 519, + 436, + 609, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/25699cc9b99af0fbe2b2e70a26eb4a3e9654ffe701167d3ff592ec2f5d7ad951.jpg", + "image_caption": [ + "Driving Pose" + ], + "image_footnote": [], + "bbox": [ + 519, + 517, + 609, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8564196c94e1af9126a725ed2d3f0f212c97e3230d44b508a219b94a44b98a43.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 436, + 684, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/6f275feae8d6b4fadb57e84f8bcd5e4048ba6e5deaf015a08c9642828bacc781.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 614, + 517, + 684, + 595 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/dab3b0faa9d80bd568e02481b13fc246038cf21c99950a2eae3501056e60a0db.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 686, + 436, + 736, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/10fbd9de7b0cd184fc6baedfa8a046b30f1d26c55fb37ea4c3f895859c673c1b.jpg", + "image_caption": [ + "Inputs" + ], + "image_footnote": [], + "bbox": [ + 686, + 517, + 733, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/59412e687f8130757fd32dea8b2e7fd53b3b61ca0168eec2b808b4d5fb31e9cd.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 772, + 436, + 820, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/8cf9c90109f0a6f56607c552f22ed7dd3e80d0055fdad7bce1fedbfcd4003387.jpg", + "image_caption": [ + "Result1" + ], + "image_footnote": [], + "bbox": [ + 761, + 521, + 823, + 597 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/16bf92d9cffe3145086c9ad41e4595d922e8ad0cd2724b58aa7d042bfafb358e.jpg", + "image_caption": [ + "Figure 4. Qualitative comparisons with representative methods [23, 25, 42, 62] in the self-driven motion reenactment task. Our method well models facial and hand regions, which matches the input image most in global identity preservation and local details modeling, even compared with some methods trained on captured videos." + ], + "image_footnote": [], + "bbox": [ + 844, + 436, + 893, + 515 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/87b15cc5dabf966138227fadf700527c0fc0db1b5ddc633b97ac0625b19a23e5.jpg", + "image_caption": [ + "Result2" + ], + "image_footnote": [], + "bbox": [ + 841, + 521, + 897, + 597 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/2c9acb02520e8db975ed1f0783911add374fc595005baab19e39fcdddb03d957.jpg", + "table_caption": [ + "Figure 5. More examples of cross-identity pose reenactment. Different subjects can be accurately animated with the same poses." + ], + "table_footnote": [], + "table_body": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
w/o-mesh-SC1.5327.990.93800.0299
L1&SSIM-PFG1.4228.400.94160.0347
w/o-retrack1.5227.990.93820.0296
w/o-pseudo-face1.4628.240.93910.0293
Full-model1.2229.310.94310.0299
", + "bbox": [ + 516, + 657, + 901, + 752 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Table 2. Quantitative results of ablation studies.", + "bbox": [ + 566, + 762, + 849, + 775 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.3. Ablation Studies", + "text_level": 1, + "bbox": [ + 511, + 787, + 678, + 801 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Perceptual-based Pseudo-Frames Guidance. Pseudoframes guidance is the core concept behind this work, enabling the challenging task of generating animatable and expressive talking avatars from a single input image. As shown in Fig. 6, without diffusion guidance, we are unable to inpaint unseen regions effectively, and the result suffers", + "bbox": [ + 511, + 809, + 906, + 900 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "10404", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/71b77f527cbfc7ac8fc58e98448da6cc85c9c4f745f84965659f4f8b222045ae.jpg", + "image_caption": [ + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 98, + 92, + 207, + 258 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ada814de8703553ec9d213fc8807f4f1224032926ff8c2ac37f52c62d3da9de2.jpg", + "image_caption": [ + "Driving Pose" + ], + "image_footnote": [], + "bbox": [ + 207, + 93, + 310, + 255 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/feb678e3053189aac911b764ddd593fad38389147ed7a940c4107e59ded03cf2.jpg", + "image_caption": [ + "Full Model" + ], + "image_footnote": [], + "bbox": [ + 323, + 90, + 473, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/639f29e821f3d0267212d23c053479a9625d2d0092f227147beb2c6ea32d806e.jpg", + "image_caption": [ + "L1&SSIM PFG" + ], + "image_footnote": [], + "bbox": [ + 486, + 90, + 637, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/58e5f24b793c74a898161b94e4999da9f35e1ab808a0e14fbc74a123fc436825.jpg", + "image_caption": [ + "w/o PFG" + ], + "image_footnote": [], + "bbox": [ + 650, + 95, + 763, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/a48351aba68138112958e91e6a898b1dfd2d9daed9f34873145c2daa9c5b977b.jpg", + "image_caption": [ + "w/o Re-Tracking" + ], + "image_footnote": [], + "bbox": [ + 794, + 95, + 895, + 257 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/71f3ade00f3ed16e031235d599840d32f89cbc1e1be13cff8650b6d3fe4c92e8.jpg", + "image_caption": [ + "Figure 6. Visualization of ablation studies on perceptual-based pseudo-frames guidance (PFG) and re-tracking step.", + "Input Image" + ], + "image_footnote": [], + "bbox": [ + 96, + 309, + 169, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/370405f1fa9eb19c7ff0d136f340c106528b32491f89e6404a2b5b0a6bd0239e.jpg", + "image_caption": [ + "Full Model" + ], + "image_footnote": [], + "bbox": [ + 178, + 309, + 274, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/fcb35dfad717117a501e77ca41be6e1cbd88267cb4a72de4ecf7ac578f0a9ece.jpg", + "image_caption": [ + "w/o Pseudo Face" + ], + "image_footnote": [], + "bbox": [ + 284, + 309, + 375, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/8b49a100e6a31db0220f63bfa0b56060a244a48fa5ed83dbb7863177fb592e97.jpg", + "image_caption": [ + "w/o mesh SC" + ], + "image_footnote": [], + "bbox": [ + 380, + 309, + 477, + 426 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ac02986e65c94cd3a15d2445c3662e3c6cd1fbafeff3d224c5984264937fe32d.jpg", + "image_caption": [ + "Figure 7. Visualization of ablation studies on pseudo face supervision and mesh-related soft constraints.", + "Input Image", + "Figure 8. Novel view synthesis. Our method demonstrates robust performance within a view range of -30 to 30 degrees. However, its performance degrades for larger viewing angles" + ], + "image_footnote": [], + "bbox": [ + 94, + 482, + 176, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/5654ae2390b62f8a82cd304d4677cd7dca27b03ad241d1b732b0e8dbc8d1bbf7.jpg", + "image_caption": [ + "$20^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 181, + 484, + 254, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/350ef94e58b722de0ca470dc007f096871fbc3bd569f76d5cdc2cd77d3c6ecbe.jpg", + "image_caption": [ + "$30^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 266, + 484, + 331, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/52c748fc6f6b9d9492f6039731a0b85fbf154817d21bcacd2b8b672517468b0a.jpg", + "image_caption": [ + "$45^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 348, + 484, + 408, + 590 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/0cf30b73d88edcde663adc98bd9328734d4fd761dd28854b6a5400009861cad7.jpg", + "image_caption": [ + "$60^{\\circ}$" + ], + "image_footnote": [], + "bbox": [ + 428, + 484, + 480, + 590 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "from severe overfitting to the input image. Additionally, due to the misalignment and inconsistency of the pseudo-output generated by motion diffusion models, using pixel-based losses like L1 for diffusion guidance results in overly smooth outputs and struggles to capture fine details, especially in facial and hand regions. In contrast, the perceptual diffusion guidance we employ not only preserves the full personalized attributes of the input image but also inpaints unseen regions with more natural and consistent textures.", + "bbox": [ + 89, + 672, + 482, + 808 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Re-Tracking Step. The re-tracking procedure applied to the pseudo-output of motion diffusion models helps mitigate the misalignment introduced by 2D diffusion models, preventing texture errors and detail loss. As shown in Fig. 6, the re-tracking step successfully recovers more accurate texture structures.", + "bbox": [ + 89, + 810, + 482, + 898 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Mesh-related Constraints. Our designed hybrid 3DGS-mesh avatar representation and the corresponding regularizations significantly enhance the expressiveness and integrity of the final results, as demonstrated in Fig. 7. When soft constraints for mesh deformation are omitted, geometric artifacts appear in specific regions and the Gaussian field fails to capture fine details.", + "bbox": [ + 511, + 311, + 903, + 416 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Pseudo Face Supervision. While perceptual-based guidance and hybrid 3DGS-mesh avatar representation enable pseudo human supervision to recover coarse head textures, incorporating targeted information from pseudo face frames further enhances facial quality, as presented in Fig. 7.", + "bbox": [ + 511, + 417, + 903, + 494 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Besides the significant visual results shown in Figs. 6 and 7, quantitative evaluation is also conduct, as seen in Tab. 2.", + "bbox": [ + 511, + 494, + 903, + 525 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Conclusion", + "text_level": 1, + "bbox": [ + 511, + 541, + 633, + 556 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "In this paper, we introduce a novel pipeline for creating expressive talking avatars from a single image. We propose a coupled 3DGS-Mesh avatar representation, incorporating several key constraints and a carefully designed hybrid learning framework that combines information from both the input image and pseudo-frames. Experimental results demonstrate that our method outperforms existing techniques, with our one-shot avatar even surpassing state-of-the-art methods that require video input. Considering its simplicity in construction and ability to generate vivid animations, our method shows significant potential for practical applications of talking avatars across various fields.", + "bbox": [ + 511, + 566, + 906, + 748 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. Severe mismatches between the input image and the registered parametric human mesh, especially in regions like fingers, can cause optimization issues and result in incorrect textures. Additionally, rendering large views or extending to full $360^{\\circ}$ human reconstruction remains difficult due to current limitations in human motion diffusion models and the lack of data for large, novel viewpoints, as seen in Fig. 8. Future work will explore integrating semantic information from large language models and static priors from 3D reconstruction to address these limitations.", + "bbox": [ + 511, + 750, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "10405", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62441224, No. 62272433, No. 62402468), The Hong Kong Polytechnic University (Project No. P0051019), the Research Center for Intelligent Operations Research of the PolyU and the advanced computing resources provided by the Supercomputing Center of the USTC.", + "bbox": [ + 89, + 90, + 480, + 189 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 215, + 187, + 231 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gordon Wetzstein. Gaussian shell maps for efficient 3d human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9441-9451, 2024. 3", + "[2] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3", + "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3", + "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3", + "[5] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 2", + "[6] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3", + "[7] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. In Proceedings of the European conference on computer vision (ECCV), 2024. 2, 4, 5", + "[8] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14916-14927, 2023. 3", + "[9] Mengyang Feng, Jinlin Liu, Kai Yu, Yuan Yao, Zheng Hui, Xiefan Guo, Xianhui Lin, Haolan Xue, Chen Shi, Xiaowen Li, et al. Dreamoving: A human video generation framework based on diffusion models. arXiv e-prints, pages arXiv-2312, 2023. 3", + "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + ], + "bbox": [ + 93, + 241, + 482, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4", + "[11] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 1", + "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning, 2023. 3", + "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 2", + "[14] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in ai: the casual conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021. 6", + "[15] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 3", + "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. Stereopifu: Depth aware clothed human digitization via stereo vision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3", + "[17] Chen Honghu, Peng Bo, Tao Yunfan, and Zhang Juyong. D $^3$ human: Dynamic disentangled digital human from monocular video. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2", + "[18] Hezhen Hu, Zhiwen Fan, Tianhao Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. Expressive gaussian human avatars from monocular rgb video. In NeurIPS, 2024. 3, 4", + "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2, 3, 4", + "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3", + "[21] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9352-9364, 2023. 3", + "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2, 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "10406", + "bbox": [ + 480, + 944, + 519, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[23] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3, 5, 6, 7", + "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531–1542. IEEE, 2024. 3, 4", + "[25] Ziyao Huang, Fan Tang, Yong Zhang, Xiaodong Cun, Juan Cao, Jintao Li, and Tong-Yee Lee. Make-your-anchor: A diffusion-based 2d avatar generation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2024. 3, 6, 7", + "[26] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6", + "[27] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2", + "[28] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3", + "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 6", + "[30] Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19734-19745, 2024. 3", + "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2", + "[32] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6", + "[33] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: Human gaussian splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2", + "[34] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5", + "[35] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 3" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[36] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 3", + "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (ToG), 36 (6):194-1, 2017. 2", + "[38] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. 2", + "[39] Zhibin Liu, Haoye Dong, Aviral Chharia, and Hefeng Wu. Human-vdm: Learning single-image 3d human gaussian splatting from video diffusion models. arXiv preprint arXiv:2409.02851, 2024. 3", + "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (ToG), 34(6), 2015. 2, 3", + "[41] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 6", + "[42] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In ECCV, 2024. 3, 4, 5, 6, 7", + "[43] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3", + "[44] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splattering with structure priors. arXiv preprint arXiv:2406.12459, 2024. 3", + "[45] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1165-1175, 2024. 3", + "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1", + "[47] Bo Peng, Jun Hu, Jingtao Zhou, Xuan Gao, and Juyong Zhang. Intrinsicngp: Intrinsic coordinate based hash encoding for human nerf. IEEE Transactions on Visualization and Computer Graphics, 2023. 2", + "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 3" + ], + "bbox": [ + 516, + 90, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "10407", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2", + "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 4", + "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-athlete: Animatable avatars via deformable 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3", + "[52] Lingteng Qiu, Shenhao Zhu, Qi Zuo, Xiaodong Gu, Yuan Dong, Junfei Zhang, Chao Xu, Zhe Li, Weihao Yuan, Liefeng Bo, Guanying Chen, and Zilong Dong. Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction, 2024. 3", + "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2", + "[54] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6), 2017. 2", + "[55] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3", + "[56] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3", + "[57] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3", + "[58] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6", + "[59] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. ICLR 2025, 2025. 2, 3", + "[60] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.4", + "[61] Jiapeng Tang, Davide Davoli, Tobias Kirschstein, Liam Schoneveld, and Matthias Niessner. Gaf: Gaussian avatar" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "reconstruction from monocular videos via multi-view diffusion, 2024. 3", + "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024. 2, 3, 5, 6, 7", + "[63] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2, 3", + "[64] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2", + "[65] Donglai Xiang, Fabian Prada, Zhe Cao, Kaiwen Guo, Chenglei Wu, Jessica Hodgins, and Timur Bagautdinov. Drivable avatar clothing: Faithful full-body telepresence with dynamic clothing driven by sparse rgb-d input. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1", + "[66] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1802-1812, 2024. 3", + "[67] Zhangyang Xiong, Di Kang, Derong Jin, Weikai Chen, Linchao Bao, Shuguang Cui, and Xiaoguang Han. Get3dhuman: Lifting stylegan-human into a 3d generative model using pixel-aligned reconstruction priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3", + "[68] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3", + "[69] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 512-523, 2023. 3", + "[70] Yuliang Xiu, Yufei Ye, Zhen Liu, Dimitrios Tzionas, and Michael J Black. Puzzle avatar: Assembling 3d avatars from personal albums. ACM Transactions on Graphics (TOG), 2024. 3", + "[71] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2, 3", + "[72] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3, 4" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "10408", + "bbox": [ + 480, + 944, + 517, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 2, 5", + "[74] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 4, 6", + "[75] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844-1854, 2024. 3, 4", + "[76] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 6", + "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 3, 4, 5", + "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9947, 2024. 3", + "[79] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2", + "[80] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6", + "[81] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3", + "[82] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 2021. 3", + "[83] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2, 3", + "[84] Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. idol: Instant photorealistic 3d human creation from a single image. arXiv preprint arXiv:2412.14963, 2024. 3" + ], + "bbox": [ + 91, + 92, + 482, + 898 + ], + "page_idx": 11 + }, + { + "type": "ref_text", + "text": "[85] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3", + "bbox": [ + 516, + 92, + 903, + 146 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "10409", + "bbox": [ + 480, + 945, + 517, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_model.json b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_model.json new file mode 100644 index 0000000000000000000000000000000000000000..be315d948f4781b1f9ac08073cc4fa872632974a --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_model.json @@ -0,0 +1,3491 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.154, + 0.131, + 0.846, + 0.154 + ], + "angle": 0, + "content": "Expressive Talking Human from Single-Image with Imperfect Priors" + }, + { + "type": "text", + "bbox": [ + 0.218, + 0.18, + 0.324, + 0.2 + ], + "angle": 0, + "content": "Jun Xiang\\(^{1,2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.364, + 0.181, + 0.482, + 0.199 + ], + "angle": 0, + "content": "Yudong Guo\\(^{1*}\\)" + }, + { + "type": "text", + "bbox": [ + 0.352, + 0.2, + 0.484, + 0.217 + ], + "angle": 0, + "content": "Yancheng Yuan2" + }, + { + "type": "text", + "bbox": [ + 0.217, + 0.218, + 0.506, + 0.234 + ], + "angle": 0, + "content": "1University of Science and Technology of China" + }, + { + "type": "text", + "bbox": [ + 0.521, + 0.181, + 0.625, + 0.199 + ], + "angle": 0, + "content": "Leipeng Hu" + }, + { + "type": "text", + "bbox": [ + 0.666, + 0.181, + 0.775, + 0.199 + ], + "angle": 0, + "content": "Boyang Guo" + }, + { + "type": "text", + "bbox": [ + 0.524, + 0.2, + 0.645, + 0.218 + ], + "angle": 0, + "content": "Juyong Zhang" + }, + { + "type": "text", + "bbox": [ + 0.534, + 0.219, + 0.778, + 0.234 + ], + "angle": 0, + "content": "\\(^{2}\\)The Hong Kong Polytechnic University" + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.275, + 0.339, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.171, + 0.492, + 0.264, + 0.503 + ], + "angle": 0, + "content": "One-Shot Photos" + }, + { + "type": "image", + "bbox": [ + 0.346, + 0.272, + 0.907, + 0.485 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.496, + 0.491, + 0.708, + 0.505 + ], + "angle": 0, + "content": "Animatable Expressive Talking Avatars" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.513, + 0.908, + 0.543 + ], + "angle": 0, + "content": "Figure 1. Given a one-shot image (e.g., your favorite photo) as input, our method reconstructs a fully expressive whole-body talking avatar that captures personalized details and supports realistic animation, including vivid body gestures and natural expression changes." + }, + { + "type": "title", + "bbox": [ + 0.249, + 0.56, + 0.327, + 0.575 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.594, + 0.486, + 0.868 + ], + "angle": 0, + "content": "Building realistic and animatable avatars still requires minutes of multi-view or monocular self-rotating videos, and most methods lack precise control over gestures and expressions. To push this boundary, we address the challenge of constructing a whole-body talking avatar from a single image. We propose a novel pipeline that tackles two critical issues: 1) complex dynamic modeling and 2) generalization to novel gestures and expressions. To achieve seamless generalization, we leverage recent pose-guided image-to-video diffusion models to generate imperfect video frames as pseudo-labels. To overcome the dynamic modeling challenge posed by inconsistent and noisy pseudo-frames, we introduce a tightly coupled 3DGS-mesh hybrid avatar representation and apply several key regularizations to mitigate inconsistencies caused by imperfect labels. Extensive experiments on diverse subjects demonstrate that our method enables the creation of a photorealistic, precisely animatable, and expressive whole-body" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.562, + 0.907, + 0.592 + ], + "angle": 0, + "content": "talking avatar from just a single image. Project page: https://ustc3dv.github.io/OneShotOneTalk/." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.619, + 0.646, + 0.635 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.511, + 0.644, + 0.907, + 0.84 + ], + "angle": 0, + "content": "Realistic rendering and precise control of gestures and expressions for whole-body talking avatars hold significant potential for AR/VR applications, such as telepresence and immersive remote conferencing. While extensively studied in the fields of 3D Vision and Graphics, most production-level body avatars still require a light stage [11, 65] to capture hours or minutes of multi-view frames, making them unsuitable for typical consumer usage. To address this, we focus on the challenging task of constructing an animatable whole-body avatar that supports both photo-realistic rendering and precise 3D control of gestures and expressions, from only a single image. This one-shot pipeline primarily faces the following challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Complex dynamic modeling. Humans exhibit complex gestures and facial movements when communicating with each other. To model the motion space of the entire body, SMPL-X [46] integrates several previous mod" + }, + { + "type": "page_footnote", + "bbox": [ + 0.092, + 0.887, + 0.295, + 0.901 + ], + "angle": 0, + "content": "*Corresponding author: Yudong Guo." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10398" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.362 + ], + "angle": 0, + "content": "els [37, 40, 54] to capture body, hand, and facial movements. With the help of these parametric geometric models, recent personalized human avatars learn to model dynamic geometry and appearance by integrating neural textures [13, 38], hybrid representation [5, 17, 27], neural radiance fields [29, 47, 49, 64, 79], and 3D Gaussians [20, 22, 33, 35, 51] for dynamic photo-realistic rendering. To capture the full appearance of a whole-body avatar, these methods require complete observations, typically relying on dense input data for supervision, such as multi-view videos or self-rotating monocular videos. Moreover, these approaches primarily focus on body motion and novel view synthesis, falling short in accurately capturing and animating diverse facial expressions and hand movements, which limits their practical applicability. Due to the reliance on dense inputs, realistic and expressive whole-body modeling from a single image remains an unsolved challenge in this field." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.369, + 0.483, + 0.776 + ], + "angle": 0, + "content": "Generalization to novel gestures and expressions. Another key challenge lies in the limited ability to generalize to a wide variety of gestures and facial expressions, especially those that deviate significantly from the poses seen during training. This limitation arises from the data-driven nature of current methods, where the training data typically encompasses only a finite set of gestures and expressions. Consequently, when animating avatars, regions or motions that were underrepresented in the training data—such as dynamic clothing, inner mouth movements, or intricate hand gestures—are often poorly synthesized or omitted entirely. These issues are further compounded when working with a single image input, as the limited visual cues often fail to capture the full range of motion and expression. Recently, several approaches [19, 59, 62, 63, 71, 77, 83] have leveraged image and video diffusion models [3, 53] to achieve image-to-video generation guided by whole-body landmarks [73]. Although various model architectures and training techniques have been explored to improve generation quality, temporal consistency remains a challenge, and these methods have several key limitations. For example, body landmarks fail to disentangle identity and pose, leading to poor cross-identity animation results, with facial and body distortions and identity mismatches. Moreover, the sparse 2D landmarks are insufficient for precisely controlling gestures and facial movements, as they do not accurately capture the underlying dynamic geometry." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.482, + 0.901 + ], + "angle": 0, + "content": "In this work, we propose a novel pipeline to address both challenges in the one-shot setting. To generalize to diverse gestures and facial movements, we preprocess the large-scale TED Gesture Dataset [74] to build a comprehensive whole-body motion space for people talking. We then use these motion sequences as guidance to drive the input single image with a pre-trained whole-body video diffusion model [77] and a 3D face animation model [7]. This" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.288 + ], + "angle": 0, + "content": "enables us to generate various video sequences of the target person performing different gestures and expressions. However, directly using these pseudo-frames to train a body avatar leads to unsatisfactory results. First, as noted above, current diffusion-based body animation methods struggle with temporal consistency, identity preservation, and motion alignment. Directly employing them as supervision results in significant blurring, distortion, and identity degradation. Furthermore, as we adapt two individual approaches for generating pseudo-frames for body gestures and facial expressions, respectively. Due to inconsistencies in camera spaces and rendering procedures, merging them into a unified avatar representation poses additional challenges." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.289, + 0.905, + 0.453 + ], + "angle": 0, + "content": "To tackle the challenge of complex dynamic modeling, we utilize both the single input image and imperfect pseudoframes to train a hybrid mesh-3DGS avatar representation, constrained by several carefully designed regularizations. The single input image provides an accurate, though incomplete, appearance for the body avatar, while the pseudo videos offer imperfect but more complete visual cues. For the pseudo video labels, instead of using per-pixel losses, we employ a perceptual-based loss term [76], which helps achieve reasonable appearance modeling while alleviating misalignment in the pseudo-frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.455, + 0.905, + 0.603 + ], + "angle": 0, + "content": "To further alleviate the inconsistencies caused by the pseudo-frames, we adopt a tightly coupled mesh-3DGS hybrid avatar representation. By introducing Laplacian smoothing and normal-consistency regularization on the deformed body mesh, we ensure that the structure of the 3D Gaussians used for rendering is well-constrained. Finally, we supervise the avatar representation using both gesture and head pseudo-frames, enabling the creation of a photorealistic, precisely animatable, and expressive whole-body avatar." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.605, + 0.904, + 0.635 + ], + "angle": 0, + "content": "In summary, our contributions include the following aspects:" + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.636, + 0.904, + 0.68 + ], + "angle": 0, + "content": "- We introduce a novel pipeline that overcomes the key challenges of building a whole-body expressive talking avatar from a single image." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.681, + 0.905, + 0.741 + ], + "angle": 0, + "content": "- Our diffusion guidance strategy effectively extracts valuable knowledge from imperfect diffusion outputs and combines it with the limited information from the input image, enabling complete modeling of the talking avatar." + }, + { + "type": "text", + "bbox": [ + 0.514, + 0.741, + 0.905, + 0.801 + ], + "angle": 0, + "content": "- Our carefully designed 3DGS-mesh coupled avatar representation, along with essential regularization techniques, facilitates accurate modeling of diverse subjects and stabilizes the optimization process." + }, + { + "type": "list", + "bbox": [ + 0.514, + 0.636, + 0.905, + 0.801 + ], + "angle": 0, + "content": null + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.815, + 0.655, + 0.831 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.84, + 0.905, + 0.901 + ], + "angle": 0, + "content": "Human Gaussian Splatting. 3D Gaussian Splatting [31] is the state-of-the-art method for scene reconstruction and novel view synthesis (NVS), offering superior rendering speed and visual quality. This has significantly influ" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.956 + ], + "angle": 0, + "content": "10399" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.394 + ], + "angle": 0, + "content": "enced human avatar studies. Methods using multi-view videos [30, 43, 45, 51, 85] demonstrate excellent performance, with unique designs such as ASH [45], which achieves efficient Gaussian learning via mesh UV parameterization. For monocular video input, most human Gaussian splatting methods [18, 20, 22, 35, 42, 57] link Gaussian fields to parametric mesh models, often using additional regularization terms. ExAvatar [42] applies connectivity-based regularizers to short, casually captured videos. For sparse-view images, GPS-Gaussian [81] achieves real-time human NVs by encoding human priors into the network, while HumanSplat [44] generates high-quality static reconstructions from one-shot inputs. Concurrent works [52, 61] introduce multi-view diffusion models but are limited to the head or body areas. IDOL [84] leverages a large-scale dataset and shows great performance in single-view human reconstruction. However, it still struggles with handling facial dynamic. By contrast, our method is capable of recovering a realistic, animatable whole-body talking avatar from just a single image." + }, + { + "type": "text", + "bbox": [ + 0.093, + 0.399, + 0.483, + 0.73 + ], + "angle": 0, + "content": "Avatar Reconstruction from Few-Shot Images. Some works [8, 28, 67] apply 3D GAN inversion for one-shot human reconstruction, but they struggle with preserving personal details and generalization. PIFu [55] and subsequent works [6, 16, 56, 68, 69, 82] introduce pixel-aligned features and neural fields for image-based human reconstruction. An alternative approach leverages diffusion priors to fill in missing details, such as training human-centered diffusion models [2, 15, 44], using novel-view diffusion results for additional supervision [36, 39, 78], and employing Score Distillation Sampling (SDS) [50] to generate 3D avatars from 2D priors [24, 70, 72, 75]. However, these methods focus on static scenes and overlook dynamic human motion, limiting their ability to capture human dynamic. The SMPL-based [40] representation enables ELICIT [23] and SHERF [21] to model human dynamic, but overlooks hand and facial motions, restricting expressive animation capabilities. In contrast, our approach leverages priors from a pose-guided human video diffusion model, capturing both human appearance and dynamic, and enabling expressive full-body animations, particularly in the hands and face." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.735, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Pose-Guided Human Video Diffusion. Pose-guided human video diffusion models [9, 19, 59, 63, 71, 77, 83] directly generate animated videos from a reference image and pose sequence, bypassing traditional 3D reconstruction and rendering processes. The success of these models depends on the quality of training data, model design, and pose guidance. Some approaches [9, 19, 71, 83] incorporate temporal layers to ensure smooth transitions, inspired byAnimateDiff [12], while others [48, 59, 77] use video diffusion models [3, 4] for dynamic sequences. Pose guidance is typically provided by OpenPose [19, 63, 77], DensePose [71]," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "depth maps [9], or SMPL [83]. StableAnimator [62] proposes HJB-based face optimization that further enhances face quality, and Make-Your-Anchors [25] personalizes outputs by fine-tuning models on identity-specific images. We adopt MimicMotion [77] for its superior performance in handling hand regions. Despite their strengths, these 2D models still face challenges, such as image distortion, identity changes, and pose misalignment, due to the lack of 3D understanding. To address these, we leverage optimized avatar representations with carefully designed constraints, improving consistency and naturalness in the generated animations." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.287, + 0.605, + 0.302 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.907, + 0.493 + ], + "angle": 0, + "content": "Given a single image of the target person, we aim to reconstruct a 3D talking avatar that fully inherits the identity and enables natural animation. To address the challenge of complex dynamic modeling from imperfect pseudo-frames, we adopt a tightly coupled 3DGS-mesh hybrid avatar representation (Sec. 3.1). To generalize well to diverse gestures and facial movements, we generate imperfect sequences of the target person driven by various motion sequences (Sec. 3.2). Finally, we introduce the carefully designed constraints and loss terms to train the representation from noisy videos effectively (Sec. 3.3). The entire pipeline is illustrated in Fig. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.502, + 0.771, + 0.518 + ], + "angle": 0, + "content": "3.1. Coupled 3DGS-Mesh Avatar" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.524, + 0.905, + 0.643 + ], + "angle": 0, + "content": "Whole-body parametric mesh models facilitate human animation and provide good initialization, while 3DGS offers enhanced expressiveness and realistic rendering. To address the challenges of the one-shot task, we design a novel coupled 3DGS-mesh representation, which effectively integrates the geometric priors and surface regularization of the mesh without diminishing the expressive capability of the Gaussian field." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.645, + 0.905, + 0.674 + ], + "angle": 0, + "content": "We couple the 3DGS field with the typical SMPL-X model, which is formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.686, + 0.905, + 0.703 + ], + "angle": 0, + "content": "\\[\nM (\\beta , \\theta , \\phi) = W (T (\\beta , \\theta , \\phi), J (\\beta), \\theta , \\mathcal {W}), \\tag {1}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.558, + 0.727, + 0.905, + 0.744 + ], + "angle": 0, + "content": "\\[\nT (\\beta , \\theta , \\phi) = \\bar {T} + B _ {S} (\\beta) + B _ {E} (\\phi) + B _ {P} (\\theta), \\tag {2}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\( B_{S}(\\cdot), B_{E}(\\cdot) \\) and \\( B_{P}(\\cdot) \\) denote shape, expression, and pose blend functions respectively, with \\( \\beta, \\phi \\) and \\( \\theta \\) representing the corresponding parameters. \\( \\bar{T} \\) is the template mesh, and \\( W(\\cdot) \\) is the standard LBS function that rotates the vertices in \\( T \\) around the estimated joints \\( J \\), smoothed by the blend weights \\( \\mathcal{W} \\). Since the shape code \\( \\beta \\) is fixed once registered, we will omit it in the later statements and denote \\( T = T(\\beta, \\theta, \\phi) \\) and \\( J = J(\\beta) \\) for simplicity. Inspired by [1, 45, 66], we initialize the 3D Gaussians on the canonical mesh surface using UV parameterization. For a" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10400" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.098, + 0.093, + 0.9, + 0.365 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.376, + 0.908, + 0.435 + ], + "angle": 0, + "content": "Figure 2. Overview. Our method constructs an expressive whole-body talking avatar from a single image. We begin by generating pseudo body and head frames using pre-trained generative models, driven by a collected video dataset with diverse poses. Per-pixel supervision on the input image, perceptual supervision on imperfect pseudo-frames, and mesh-related constraints are then applied to guide the 3DGS-mesh coupled avatar representation, ensuring realistic and expressive avatar reconstruction and animation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.459, + 0.484, + 0.506 + ], + "angle": 0, + "content": "3D Gaussian located on a triangle \\( k = \\{\\mathbf{V}_1, \\mathbf{V}_2, \\mathbf{V}_3\\} \\) of mesh \\( T \\) with barycentric coordinates \\( (u, v) \\), the surface position is given by:" + }, + { + "type": "equation", + "bbox": [ + 0.1, + 0.515, + 0.483, + 0.532 + ], + "angle": 0, + "content": "\\[\n\\operatorname {B a r y} (T) = \\mathcal {V} (k, u, v) = u \\mathbf {V} _ {\\mathbf {1}} + v \\mathbf {V} _ {\\mathbf {2}} + (1 - u - v) \\mathbf {V} _ {\\mathbf {3}}. \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.542, + 0.484, + 0.829 + ], + "angle": 0, + "content": "Coupled 3DGS-Mesh Deformation. To model clothes, haircuts, and other complex regions that the SMPL-X mesh fails to handle, the deformation of 3D Gaussians is crucial. While previous works can achieve correct deformation with sufficient supervision, we need to impose additional constraints on the deformation field to prevent overfitting to the single input image and imperfect pseudo-frames, while also facilitating the integration of other modules. Based on these considerations, we design two deformation fields in the canonical space and enforce their proximity: one represents the conventional Gaussian deformation, and the other represents the critical mesh deformation. This approach allows us to indirectly influence the Gaussian deformation by applying soft constraints on the deformed mesh. Moreover, since there is no strict binding between the deformed mesh surface and the Gaussians, complex regions are still effectively handled. The two deformation fields, together with the full-body animation and Gaussian Splitting process, are formulated as follows:" + }, + { + "type": "equation", + "bbox": [ + 0.09, + 0.84, + 0.484, + 0.872 + ], + "angle": 0, + "content": "\\[\nG (\\theta , \\phi , \\mathbf {P}) = S P (W (B a r y (T) + d X, J, \\theta , B a r y (\\mathcal {W})), \\mathbf {P}). \\tag {4}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.175, + 0.885, + 0.483, + 0.903 + ], + "angle": 0, + "content": "\\[\nM (\\theta , \\phi) = W (T + d T, J, \\theta , \\mathcal {W}), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.459, + 0.907, + 0.568 + ], + "angle": 0, + "content": "where \\( G(\\cdot) \\) represents the final rendered image, \\( SP(\\cdot) \\) and \\( \\mathbf{P} \\) denote the splatting rendering process and the properties of 3D Gaussians, and \\( dX \\) and \\( dT \\) represent the optimized deformations of the Gaussians and mesh vertices, respectively. We apply key regularizers on \\( dT \\) and propagate soft constraints to \\( dX \\) by encouraging \\( dX \\) to align with \\( Bary(dT) \\)." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.576, + 0.761, + 0.59 + ], + "angle": 0, + "content": "3.2. Pseudo-Frames Generation" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.599, + 0.907, + 0.674 + ], + "angle": 0, + "content": "In the one-shot image setting, many regions are unseen or occluded. To construct a complete avatar and ensure generalization to novel gestures and expressions, we turn our attention to recent advances in human motion diffusion models [19, 77] and head animation techniques [7, 10]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.675, + 0.909, + 0.84 + ], + "angle": 0, + "content": "For leveraging diffusion-based generative models, the SDS loss [50] is widely used in text/image-to-3D works [24, 60, 72, 75]. Although diffusion models may introduce moderate deviations, the SDS technique remains effective in these studies. However, for tasks such as realistic dynamic avatar animation from a one-shot image, pose alignment is critical [18, 42], and the misalignment introduced by 2D diffusion models cannot be overlooked. Therefore, directly applying SDS in our setting is suboptimal. Instead, we focus on extracting accurate information from pseudo-frames synthesized by these generative models." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.841, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To generate whole-body pseudo-frames, we collect a set of SMPL-X pose sequences \\(\\{(\\theta_{N_i},\\phi_{N_i})_{i = 1}^F\\}_{j = 1}^K\\) from the TED Gesture Dataset [74] as input to the generative models. Since no unified generative model exists for both body and" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.518, + 0.957 + ], + "angle": 0, + "content": "10401" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.095, + 0.283, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.114, + 0.234, + 0.171, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.194, + 0.228, + 0.281, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.284, + 0.095, + 0.348, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.15, + 0.347, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.368, + 0.096, + 0.416, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.35, + 0.229, + 0.433, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.438, + 0.096, + 0.594, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.441, + 0.228, + 0.503, + 0.277 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.503, + 0.228, + 0.591, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.096, + 0.659, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.15, + 0.659, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.096, + 0.733, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.598, + 0.228, + 0.745, + 0.359 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.096, + 0.816, + 0.144 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.756, + 0.151, + 0.816, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.834, + 0.096, + 0.885, + 0.221 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.755, + 0.228, + 0.816, + 0.276 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.818, + 0.229, + 0.899, + 0.357 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.1, + 0.369, + 0.282, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.109, + 0.475, + 0.181, + 0.488 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.195, + 0.475, + 0.28, + 0.488 + ], + "angle": 0, + "content": "Driving Image" + }, + { + "type": "image", + "bbox": [ + 0.299, + 0.369, + 0.427, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.346, + 0.476, + 0.376, + 0.487 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.452, + 0.369, + 0.58, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.462, + 0.476, + 0.555, + 0.487 + ], + "angle": 0, + "content": "StableAnimator" + }, + { + "type": "image", + "bbox": [ + 0.612, + 0.369, + 0.742, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.631, + 0.476, + 0.721, + 0.487 + ], + "angle": 0, + "content": "ExAvatar-1shot" + }, + { + "type": "image", + "bbox": [ + 0.771, + 0.369, + 0.898, + 0.466 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.812, + 0.475, + 0.859, + 0.486 + ], + "angle": 0, + "content": "ELICIT" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.504, + 0.907, + 0.533 + ], + "angle": 0, + "content": "Figure 3. Qualitative comparisons with representative methods [23, 42, 62] in the cross-identity motion reenactment task. Our method achieves accurate and realistic animation with almost all fine details preserved and identity unchanged." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.541, + 0.483, + 0.631 + ], + "angle": 0, + "content": "head simultaneously, we employ two separate approaches to generate pseudo-frames for body gestures and facial expressions, respectively. Given a source image \\( I_{S} \\) registered with the SMPL-X parameters \\( (\\beta_{S},\\theta_{S},\\phi_{S}) \\), we adopt MimicMotion [77] with the collected pose sequences to generate various pseudo body frames:" + }, + { + "type": "equation", + "bbox": [ + 0.165, + 0.638, + 0.483, + 0.656 + ], + "angle": 0, + "content": "\\[\nI _ {N} = \\operatorname {M o t i o n} \\left(I _ {S}, D \\left(\\beta_ {S}, \\theta_ {N}, \\phi_ {N}\\right)\\right), \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.66, + 0.483, + 0.841 + ], + "angle": 0, + "content": "where \\(D(\\cdot)\\) denotes the mapping from SMPL-X parameters to DWPose [73]. We randomly set the root body pose in \\(\\theta_N\\) for each segment to increase viewpoint generalization, with pitch \\(\\in (-30^{\\circ}, 30^{\\circ})\\) and yaw \\(\\in (-10^{\\circ}, 10^{\\circ})\\). After that, we re-track \\(I_N\\) to obtain more accurate pose parameters \\((\\hat{\\theta}_N, \\hat{\\phi}_N)\\), while keeping the shape code \\(\\beta_S\\) fixed. The re-tracking process is crucial, as it overcomes the misalignment introduced by 2D diffusion models and helps achieve precise and realistic results. For the head region, we adopt Portrait4D-v2 [7] to generate various pseudo frames of the target person performing diverse expressions, also driven by the videos from the TED Gesture Dataset." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.283, + 0.866 + ], + "angle": 0, + "content": "3.3. Objective Functions" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.902 + ], + "angle": 0, + "content": "Assisted by our novel hybrid 3DGS-mesh avatar representation, we introduce several carefully designed regularization" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.541, + 0.907, + 0.586 + ], + "angle": 0, + "content": "terms to stabilize the avatar reconstruction process and effectively extract the correct information from both the one-shot input and the imperfect pseudo-frames." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.586, + 0.907, + 0.631 + ], + "angle": 0, + "content": "Mesh-related Constraints. We use the following loss functions to apply soft constraints on the Gaussian field based on the mesh:" + }, + { + "type": "equation", + "bbox": [ + 0.608, + 0.64, + 0.81, + 0.674 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} \\mathcal {L} _ {S C} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {M G C}} \\mathcal {L} _ {\\mathrm {M G C}} + \\lambda_ {\\mathrm {l a p}} \\mathcal {L} _ {\\mathrm {l a p}}. \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.884, + 0.65, + 0.905, + 0.663 + ], + "angle": 0, + "content": "(7)" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.683, + 0.907, + 0.804 + ], + "angle": 0, + "content": "Here, \\(\\mathcal{L}_{\\mathrm{normal}}\\) is the normal consistency loss applied to the deformed mesh surface, ensuring surface normal consistency post-deformation. \\(\\mathcal{L}_M\\) is the mask loss, which measures the discrepancy between the ground truth mask and the mask of the deformed mesh rendered via [34]. These two losses work together to regulate the behavior of \\(dT\\), while influencing Gaussian deformation \\(dX\\) through the mesh-Gaussian consistency loss:" + }, + { + "type": "equation", + "bbox": [ + 0.606, + 0.814, + 0.905, + 0.831 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\mathrm {M G C}} = \\| d X - B a r y (d T) \\| _ {1}. \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.841, + 0.907, + 0.9 + ], + "angle": 0, + "content": "Additionally, we compute the Laplacian smoothing loss \\(\\mathcal{L}_{\\mathrm{lap}}\\) for \\(dX\\) along with the scaling and RGBs of the 3D Gaussians, based on their initial connectivity state on the canonical mesh surface." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10402" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.364 + ], + "angle": 0, + "content": "Perceptual Supervision of Pseudo-Frames. The synthesized pseudo-frames exhibit noticeable artifacts, such as image distortion and lack of 3D consistency, especially in the results generated with MimicMotion. These artifacts cannot be fully corrected through pose alignment alone. As a result, pixel-aligned losses like L1 or MSE often lead to issues like texture flickering, blurring, and identity changes. However, the high-level human structure is consistently well-preserved in the source image, which aids human recognition and is empirically referred to as perceptual similarity. Previous works [41, 76] have used deep neural networks to capture these perceptual structure features. We adopt the LPIPS perceptual loss [76], using VGG [58] as the backbone. By learning deep perceptual human features from dynamic poses and preserving detailed information from the source image, our method is able to conduct realistic and complete talking avatar animations. Thus, for the pseudoframes, we primarily use the following perceptual loss:" + }, + { + "type": "equation", + "bbox": [ + 0.149, + 0.372, + 0.483, + 0.39 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {\\text {d i f f}} = \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {N}, G \\left(\\hat {\\theta} _ {N}, \\hat {\\phi} _ {N}, \\mathbf {P}\\right)\\right). \\tag {9}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.399, + 0.484, + 0.461 + ], + "angle": 0, + "content": "Per-pixel Supervision of the Input Image. For the source image \\(I_S\\) with pose \\((\\theta_S, \\phi_S)\\), we use the regular L1 loss, SSIM loss, and mask loss between \\(I_S\\) and the splatted image \\(G(\\theta_S, \\phi_S, \\mathbf{P})\\):" + }, + { + "type": "equation", + "bbox": [ + 0.12, + 0.469, + 0.483, + 0.486 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {S} = \\lambda_ {\\mathrm {R G B}} \\mathcal {L} _ {\\mathrm {R G B}} + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} + \\lambda_ {\\mathrm {a l p h a}} \\mathcal {L} _ {\\mathrm {a l p h a}}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.494, + 0.484, + 0.524 + ], + "angle": 0, + "content": "The final loss function is the sum of all the aforementioned losses." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.533, + 0.31, + 0.549 + ], + "angle": 0, + "content": "3.4. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.555, + 0.484, + 0.812 + ], + "angle": 0, + "content": "We set the UV map resolution to 512 and use approximately 150,000 Gaussians in total. We train our models using the Adam optimizer [32] with \\(\\beta = (0.9, 0.999)\\). The learning rates for the Gaussian parameters are the same as those in the official implementation, while the learning rates for both the Gaussian and mesh deformation fields are set to \\(\\eta = 1e^{-4}\\). For the loss weights, we set \\(\\lambda_{\\mathrm{normal}}\\), \\(\\lambda_{\\mathrm{M}}\\), \\(\\lambda_{\\mathrm{MGC}}\\), \\(\\lambda_{\\mathrm{lap}}\\), \\(\\lambda_{\\mathrm{LPIPS}}\\), \\(\\lambda_{\\mathrm{RGB}}\\), \\(\\lambda_{\\mathrm{SSIM}}\\), \\(\\lambda_{\\mathrm{alpha}}\\) to \\(1e^{-2}\\), \\(1e^{-1}\\), \\(1e^{1}\\), \\(1e^{2}\\), \\(2e^{-1}\\), \\(8e^{-1}\\), \\(1e^{-1}\\), and \\(4e^{-1}\\), respectively. We start adding the perceptual loss at the 2000th step, while the other losses are used throughout the entire training process with a maximum number of training steps of 20000. We adopt joint training with separate supervision for body and head pseudo-frames. Training takes 3 hours on single 4090 due to critical but inefficient re-tracking. But avatar reconstruction and rendering are highly efficient with an inference speed of 50FPS+ at 1K resolution." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.824, + 0.224, + 0.841 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.187, + 0.864 + ], + "angle": 0, + "content": "4.1. Dataset" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.901 + ], + "angle": 0, + "content": "We use the poses and expressions processed from 100 videos of the TED Gesture Dataset [74] as the motion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.906, + 0.228 + ], + "angle": 0, + "content": "sequences during training. We separately sampled about 1,800 frames with the most distinct variations in poses and expressions. For evaluation, the one-shot input and driving poses are primarily sourced from the ActorsHQ [26], NeuMan [29] and the Casual Conversations Dataset [14]. All videos and images are cropped and resized to a fixed aspect ratio of 9:16. For foreground segmentation, we use BiRefNet [80], and for pose tracking, we employ the custom fitting procedure from ExAvatar [42]." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.24, + 0.874, + 0.257 + ], + "angle": 0, + "content": "4.2. Comparison with Representative Methods" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.263, + 0.906, + 0.398 + ], + "angle": 0, + "content": "We compare our method with several representative works, including: (1) ExAvatar [42], a recent SOTA work that models human avatars with a mesh-based 3D Gaussian field. (2) ELICIT [23], a one-shot NeRF-based animatable avatar. (3) StableAnimator [62], a SOTA general 2D pose-guided human video diffusion model. (4) Make-Your-Anchors [25], a 2D diffusion-based method that is fine-tuned on a one-minute video clip of the individual to enhance identity information." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.4, + 0.906, + 0.505 + ], + "angle": 0, + "content": "Note that for ExAvatar, since it takes short videos as input, we compare with two versions of it: ExAvatar-40shot, which uses 40-shot images, and ExAvatar-1shot, which uses one-shot images as input. For Make-Your-Ananchor, as we find it does not perform well on one-shot input, we only compare it with available video input by fine-tuning it on a short video clip." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.507, + 0.906, + 0.612 + ], + "angle": 0, + "content": "Qualitative Comparisons. Fig. 3 and Fig. 4 present a qualitative comparison between our method and the other representative approaches. For Fig. 3, we use the pose of a different identity as the driving signal. For Fig. 4, we compare the performance of these methods on the self-driven pose reenactment task, using subjects that have corresponding video data." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.613, + 0.906, + 0.884 + ], + "angle": 0, + "content": "As observed, ExAvatar-1shot tends to overfit the input image and fails to recover accurate textures in regions occluded by hands. Even with 40-shot input, ExAvatar still struggles with incomplete knowledge and fails to handle novel gestures effectively. StableAnimator generates relatively reasonable results but is constrained by its training distribution and struggles with identity consistency across frames, often leading to appearance mismatches and identity changes. ELICIT, which uses a NeRF-based representation, ensures 3D consistency but relies on SMPL for geometry, which neglects the hand region. This coarse semantic proxy fails to support complex hand reconstruction or animation. Make-Your-AnAnchor, although pre-trained on multiple identities to learn human motion priors, requires a long video with sufficient movement and appearance data to adapt to new identities. It struggles with short fine-tuning videos and fails to recover fine details, especially for gestures outside its fine-tuning distribution." + }, + { + "type": "text", + "bbox": [ + 0.532, + 0.886, + 0.906, + 0.902 + ], + "angle": 0, + "content": "In contrast, our method generates expressive talking" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10403" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.09, + 0.252, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.094, + 0.35, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.093, + 0.463, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.094, + 0.566, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.572, + 0.094, + 0.671, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.094, + 0.777, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.782, + 0.093, + 0.899, + 0.219 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.099, + 0.229, + 0.253, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.107, + 0.363, + 0.165, + 0.373 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image_caption", + "bbox": [ + 0.179, + 0.363, + 0.243, + 0.373 + ], + "angle": 0, + "content": "Ground Truth" + }, + { + "type": "image", + "bbox": [ + 0.255, + 0.228, + 0.353, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.294, + 0.363, + 0.319, + 0.372 + ], + "angle": 0, + "content": "Ours" + }, + { + "type": "image", + "bbox": [ + 0.355, + 0.228, + 0.465, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.375, + 0.363, + 0.448, + 0.372 + ], + "angle": 0, + "content": "StableAnimator" + }, + { + "type": "image", + "bbox": [ + 0.468, + 0.227, + 0.567, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.477, + 0.363, + 0.556, + 0.372 + ], + "angle": 0, + "content": "ExAvatar-40shot" + }, + { + "type": "image", + "bbox": [ + 0.571, + 0.228, + 0.673, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.588, + 0.363, + 0.66, + 0.372 + ], + "angle": 0, + "content": "ExAvatar-1shot" + }, + { + "type": "image", + "bbox": [ + 0.679, + 0.228, + 0.78, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.709, + 0.363, + 0.746, + 0.372 + ], + "angle": 0, + "content": "ELICIT" + }, + { + "type": "image", + "bbox": [ + 0.783, + 0.228, + 0.899, + 0.358 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.794, + 0.363, + 0.886, + 0.372 + ], + "angle": 0, + "content": "Make-Your-Anchors" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.388, + 0.908, + 0.43 + ], + "angle": 0, + "content": "Figure 4. Qualitative comparisons with representative methods [23, 25, 42, 62] in the self-driven motion reenactment task. Our method well models facial and hand regions, which matches the input image most in global identity preservation and local details modeling, even compared with some methods trained on captured videos." + }, + { + "type": "table", + "bbox": [ + 0.095, + 0.436, + 0.48, + 0.516 + ], + "angle": 0, + "content": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
ELICIT5.65522.5590.9210.0660
ExAvatar3.93124.2150.9240.0409
StableAnimator1.88327.3500.9350.0327
Ours1.21629.3110.9430.0299
" + }, + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.527, + 0.483, + 0.584 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons with representative methods [23, 42, 62] on self-driven data. Our method outperforms others in pixel-wise error metrics, realism evaluation metrics and perceptual quality metrics. (ExAvatar here denotes ExAvatar-40shot.)" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.596, + 0.484, + 0.702 + ], + "angle": 0, + "content": "avatars from a single input image. Our results preserve fine human details and achieve natural animation with excellent rendering quality. We provide additional examples of cross-identity pose reenactment in Fig. 5. Using the same driving pose, identities with completely different attributes can be driven in the same way, thanks to the SMPL-X model and the 3DGS-mesh coupled representation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.705, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Quantitative Comparison. Tab. 1 presents a quantitative comparison between our model and other methods. Although the one-shot image animation task typically lacks a strict test set for quantitative comparison, we perform the evaluation on the self-driven task and use four common metrics for comparison: Mean Squared Error, PSNR, SSIM, and LPIPS. Our method outperforms all others across these metrics, demonstrating superior realism and 3D consistency in the results. However, we believe that these metrics do not fully capture the quality or capabilities of the methods, especially for the one-shot image animation task. We encourage readers to refer to the video results of our method for a more comprehensive and objective evaluation." + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.438, + 0.611, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.518, + 0.61, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.536, + 0.599, + 0.596, + 0.609 + ], + "angle": 0, + "content": "Driving Pose" + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.438, + 0.685, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.616, + 0.518, + 0.685, + 0.597 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.664, + 0.599, + 0.693, + 0.609 + ], + "angle": 0, + "content": "Inputs" + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.438, + 0.738, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.687, + 0.518, + 0.735, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.773, + 0.438, + 0.821, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.763, + 0.522, + 0.825, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.773, + 0.599, + 0.807, + 0.608 + ], + "angle": 0, + "content": "Result1" + }, + { + "type": "image", + "bbox": [ + 0.845, + 0.438, + 0.894, + 0.516 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.843, + 0.522, + 0.898, + 0.598 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.855, + 0.599, + 0.89, + 0.608 + ], + "angle": 0, + "content": "Result2" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.622, + 0.905, + 0.652 + ], + "angle": 0, + "content": "Figure 5. More examples of cross-identity pose reenactment. Different subjects can be accurately animated with the same poses." + }, + { + "type": "table", + "bbox": [ + 0.517, + 0.659, + 0.903, + 0.753 + ], + "angle": 0, + "content": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
w/o-mesh-SC1.5327.990.93800.0299
L1&SSIM-PFG1.4228.400.94160.0347
w/o-retrack1.5227.990.93820.0296
w/o-pseudo-face1.4628.240.93910.0293
Full-model1.2229.310.94310.0299
" + }, + { + "type": "table_caption", + "bbox": [ + 0.567, + 0.763, + 0.851, + 0.776 + ], + "angle": 0, + "content": "Table 2. Quantitative results of ablation studies." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.788, + 0.679, + 0.802 + ], + "angle": 0, + "content": "4.3. Ablation Studies" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.907, + 0.901 + ], + "angle": 0, + "content": "Perceptual-based Pseudo-Frames Guidance. Pseudoframes guidance is the core concept behind this work, enabling the challenging task of generating animatable and expressive talking avatars from a single input image. As shown in Fig. 6, without diffusion guidance, we are unable to inpaint unseen regions effectively, and the result suffers" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10404" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.099, + 0.093, + 0.208, + 0.259 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.119, + 0.261, + 0.187, + 0.273 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.209, + 0.094, + 0.312, + 0.256 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.235, + 0.261, + 0.307, + 0.273 + ], + "angle": 0, + "content": "Driving Pose" + }, + { + "type": "image", + "bbox": [ + 0.325, + 0.092, + 0.474, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.373, + 0.261, + 0.434, + 0.272 + ], + "angle": 0, + "content": "Full Model" + }, + { + "type": "image", + "bbox": [ + 0.488, + 0.092, + 0.638, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.52, + 0.261, + 0.607, + 0.272 + ], + "angle": 0, + "content": "L1&SSIM PFG" + }, + { + "type": "image", + "bbox": [ + 0.651, + 0.097, + 0.764, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.688, + 0.261, + 0.738, + 0.271 + ], + "angle": 0, + "content": "w/o PFG" + }, + { + "type": "image", + "bbox": [ + 0.795, + 0.097, + 0.897, + 0.258 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.8, + 0.261, + 0.894, + 0.273 + ], + "angle": 0, + "content": "w/o Re-Tracking" + }, + { + "type": "image_caption", + "bbox": [ + 0.156, + 0.289, + 0.839, + 0.304 + ], + "angle": 0, + "content": "Figure 6. Visualization of ablation studies on perceptual-based pseudo-frames guidance (PFG) and re-tracking step." + }, + { + "type": "image", + "bbox": [ + 0.097, + 0.31, + 0.171, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.108, + 0.431, + 0.16, + 0.441 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.179, + 0.31, + 0.276, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.202, + 0.431, + 0.248, + 0.44 + ], + "angle": 0, + "content": "Full Model" + }, + { + "type": "image", + "bbox": [ + 0.285, + 0.31, + 0.376, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.291, + 0.431, + 0.361, + 0.44 + ], + "angle": 0, + "content": "w/o Pseudo Face" + }, + { + "type": "image", + "bbox": [ + 0.382, + 0.31, + 0.478, + 0.428 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.398, + 0.431, + 0.454, + 0.44 + ], + "angle": 0, + "content": "w/o mesh SC" + }, + { + "type": "image_caption", + "bbox": [ + 0.091, + 0.453, + 0.483, + 0.479 + ], + "angle": 0, + "content": "Figure 7. Visualization of ablation studies on pseudo face supervision and mesh-related soft constraints." + }, + { + "type": "image", + "bbox": [ + 0.096, + 0.483, + 0.177, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.106, + 0.595, + 0.167, + 0.606 + ], + "angle": 0, + "content": "Input Image" + }, + { + "type": "image", + "bbox": [ + 0.182, + 0.485, + 0.255, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.204, + 0.594, + 0.223, + 0.604 + ], + "angle": 0, + "content": "\\(20^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.267, + 0.485, + 0.332, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.284, + 0.594, + 0.303, + 0.604 + ], + "angle": 0, + "content": "\\(30^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.349, + 0.485, + 0.409, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.359, + 0.594, + 0.377, + 0.604 + ], + "angle": 0, + "content": "\\(45^{\\circ}\\)" + }, + { + "type": "image", + "bbox": [ + 0.429, + 0.485, + 0.482, + 0.592 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.439, + 0.594, + 0.456, + 0.604 + ], + "angle": 0, + "content": "\\(60^{\\circ}\\)" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.619, + 0.483, + 0.661 + ], + "angle": 0, + "content": "Figure 8. Novel view synthesis. Our method demonstrates robust performance within a view range of -30 to 30 degrees. However, its performance degrades for larger viewing angles" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.673, + 0.483, + 0.809 + ], + "angle": 0, + "content": "from severe overfitting to the input image. Additionally, due to the misalignment and inconsistency of the pseudo-output generated by motion diffusion models, using pixel-based losses like L1 for diffusion guidance results in overly smooth outputs and struggles to capture fine details, especially in facial and hand regions. In contrast, the perceptual diffusion guidance we employ not only preserves the full personalized attributes of the input image but also inpaints unseen regions with more natural and consistent textures." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.9 + ], + "angle": 0, + "content": "Re-Tracking Step. The re-tracking procedure applied to the pseudo-output of motion diffusion models helps mitigate the misalignment introduced by 2D diffusion models, preventing texture errors and detail loss. As shown in Fig. 6, the re-tracking step successfully recovers more accurate texture structures." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.312, + 0.905, + 0.417 + ], + "angle": 0, + "content": "Mesh-related Constraints. Our designed hybrid 3DGS-mesh avatar representation and the corresponding regularizations significantly enhance the expressiveness and integrity of the final results, as demonstrated in Fig. 7. When soft constraints for mesh deformation are omitted, geometric artifacts appear in specific regions and the Gaussian field fails to capture fine details." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.419, + 0.905, + 0.495 + ], + "angle": 0, + "content": "Pseudo Face Supervision. While perceptual-based guidance and hybrid 3DGS-mesh avatar representation enable pseudo human supervision to recover coarse head textures, incorporating targeted information from pseudo face frames further enhances facial quality, as presented in Fig. 7." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.495, + 0.905, + 0.526 + ], + "angle": 0, + "content": "Besides the significant visual results shown in Figs. 6 and 7, quantitative evaluation is also conduct, as seen in Tab. 2." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.542, + 0.634, + 0.558 + ], + "angle": 0, + "content": "5. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.568, + 0.907, + 0.749 + ], + "angle": 0, + "content": "In this paper, we introduce a novel pipeline for creating expressive talking avatars from a single image. We propose a coupled 3DGS-Mesh avatar representation, incorporating several key constraints and a carefully designed hybrid learning framework that combines information from both the input image and pseudo-frames. Experimental results demonstrate that our method outperforms existing techniques, with our one-shot avatar even surpassing state-of-the-art methods that require video input. Considering its simplicity in construction and ability to generate vivid animations, our method shows significant potential for practical applications of talking avatars across various fields." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations. Severe mismatches between the input image and the registered parametric human mesh, especially in regions like fingers, can cause optimization issues and result in incorrect textures. Additionally, rendering large views or extending to full \\(360^{\\circ}\\) human reconstruction remains difficult due to current limitations in human motion diffusion models and the lack of data for large, novel viewpoints, as seen in Fig. 8. Future work will explore integrating semantic information from large language models and static priors from 3D reconstruction to address these limitations." + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10405" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.19 + ], + "angle": 0, + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62441224, No. 62272433, No. 62402468), The Hong Kong Polytechnic University (Project No. P0051019), the Research Center for Intelligent Operations Research of the PolyU and the advanced computing resources provided by the Supercomputing Center of the USTC." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.216, + 0.188, + 0.232 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.242, + 0.482, + 0.324 + ], + "angle": 0, + "content": "[1] Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gordon Wetzstein. Gaussian shell maps for efficient 3d human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9441-9451, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.326, + 0.482, + 0.393 + ], + "angle": 0, + "content": "[2] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.396, + 0.482, + 0.466 + ], + "angle": 0, + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.467, + 0.482, + 0.535 + ], + "angle": 0, + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.537, + 0.482, + 0.591 + ], + "angle": 0, + "content": "[5] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.593, + 0.482, + 0.674 + ], + "angle": 0, + "content": "[6] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.677, + 0.482, + 0.732 + ], + "angle": 0, + "content": "[7] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. In Proceedings of the European conference on computer vision (ECCV), 2024. 2, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.734, + 0.483, + 0.802 + ], + "angle": 0, + "content": "[8] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14916-14927, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.804, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[9] Mengyang Feng, Jinlin Liu, Kai Yu, Yuan Yao, Zheng Hui, Xiefan Guo, Xianhui Lin, Haolan Xue, Chen Shi, Xiaowen Li, et al. Dreamoving: A human video generation framework based on diffusion models. arXiv e-prints, pages arXiv-2312, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.873, + 0.482, + 0.902 + ], + "angle": 0, + "content": "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.242, + 0.483, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.203 + ], + "angle": 0, + "content": "[11] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.205, + 0.905, + 0.26 + ], + "angle": 0, + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.261, + 0.905, + 0.315 + ], + "angle": 0, + "content": "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.317, + 0.905, + 0.385 + ], + "angle": 0, + "content": "[14] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in ai: the casual conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.386, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[15] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.455, + 0.905, + 0.511 + ], + "angle": 0, + "content": "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. Stereopifu: Depth aware clothed human digitization via stereo vision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.512, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[17] Chen Honghu, Peng Bo, Tao Yunfan, and Zhang Juyong. D\\(^3\\) human: Dynamic disentangled digital human from monocular video. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.568, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[18] Hezhen Hu, Zhiwen Fan, Tianhao Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. Expressive gaussian human avatars from monocular rgb video. In NeurIPS, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.623, + 0.907, + 0.678 + ], + "angle": 0, + "content": "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2, 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.679, + 0.905, + 0.761 + ], + "angle": 0, + "content": "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.763, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[21] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9352-9364, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.832, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.52, + 0.957 + ], + "angle": 0, + "content": "10406" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[23] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.232 + ], + "angle": 0, + "content": "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531–1542. IEEE, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.232, + 0.482, + 0.302 + ], + "angle": 0, + "content": "[25] Ziyao Huang, Fan Tang, Yong Zhang, Xiaodong Cun, Juan Cao, Jintao Li, and Tong-Yee Lee. Make-your-anchor: A diffusion-based 2d avatar generation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2024. 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.302, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[26] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.371, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[27] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.426, + 0.482, + 0.496 + ], + "angle": 0, + "content": "[28] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.497, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[30] Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19734-19745, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.636, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.691, + 0.482, + 0.72 + ], + "angle": 0, + "content": "[32] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.72, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[33] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: Human gaussian splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.831 + ], + "angle": 0, + "content": "[34] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[35] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.161 + ], + "angle": 0, + "content": "[36] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.162, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (ToG), 36 (6):194-1, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.218, + 0.905, + 0.285 + ], + "angle": 0, + "content": "[38] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.287, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[39] Zhibin Liu, Haoye Dong, Aviral Chharia, and Hefeng Wu. Human-vdm: Learning single-image 3d human gaussian splatting from video diffusion models. arXiv preprint arXiv:2409.02851, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.343, + 0.905, + 0.398 + ], + "angle": 0, + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (ToG), 34(6), 2015. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.399, + 0.905, + 0.455 + ], + "angle": 0, + "content": "[41] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.455, + 0.905, + 0.496 + ], + "angle": 0, + "content": "[42] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In ECCV, 2024. 3, 4, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.497, + 0.905, + 0.567 + ], + "angle": 0, + "content": "[43] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.567, + 0.905, + 0.636 + ], + "angle": 0, + "content": "[44] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splattering with structure priors. arXiv preprint arXiv:2406.12459, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.636, + 0.905, + 0.707 + ], + "angle": 0, + "content": "[45] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1165-1175, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.707, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.789, + 0.905, + 0.845 + ], + "angle": 0, + "content": "[47] Bo Peng, Jun Hu, Jingtao Zhou, Xuan Gao, and Juyong Zhang. Intrinsicngp: Intrinsic coordinate based hash encoding for human nerf. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.092, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10407" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.176 + ], + "angle": 0, + "content": "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.177, + 0.482, + 0.218 + ], + "angle": 0, + "content": "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.219, + 0.482, + 0.287 + ], + "angle": 0, + "content": "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-athlete: Animatable avatars via deformable 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.288, + 0.482, + 0.357 + ], + "angle": 0, + "content": "[52] Lingteng Qiu, Shenhao Zhu, Qi Zuo, Xiaodong Gu, Yuan Dong, Junfei Zhang, Chao Xu, Zhe Li, Weihao Yuan, Liefeng Bo, Guanying Chen, and Zilong Dong. Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.358, + 0.482, + 0.426 + ], + "angle": 0, + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.427, + 0.482, + 0.482 + ], + "angle": 0, + "content": "[54] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6), 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.484, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[55] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.621 + ], + "angle": 0, + "content": "[56] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.622, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[57] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.706, + 0.482, + 0.746 + ], + "angle": 0, + "content": "[58] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.748, + 0.482, + 0.816 + ], + "angle": 0, + "content": "[59] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. ICLR 2025, 2025. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.818, + 0.482, + 0.871 + ], + "angle": 0, + "content": "[60] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[61] Jiapeng Tang, Davide Davoli, Tobias Kirschstein, Liam Schoneveld, and Matthias Niessner. Gaf: Gaussian avatar" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "reconstruction from monocular videos via multi-view diffusion, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.121, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.178, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[63] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.247, + 0.905, + 0.328 + ], + "angle": 0, + "content": "[64] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.33, + 0.905, + 0.399 + ], + "angle": 0, + "content": "[65] Donglai Xiang, Fabian Prada, Zhe Cao, Kaiwen Guo, Chenglei Wu, Jessica Hodgins, and Timur Bagautdinov. Drivable avatar clothing: Faithful full-body telepresence with dynamic clothing driven by sparse rgb-d input. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.401, + 0.905, + 0.468 + ], + "angle": 0, + "content": "[66] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1802-1812, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.47, + 0.905, + 0.55 + ], + "angle": 0, + "content": "[67] Zhangyang Xiong, Di Kang, Derong Jin, Weikai Chen, Linchao Bao, Shuguang Cui, and Xiaoguang Han. Get3dhuman: Lifting stylegan-human into a 3d generative model using pixel-aligned reconstruction priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.553, + 0.905, + 0.621 + ], + "angle": 0, + "content": "[68] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.623, + 0.905, + 0.691 + ], + "angle": 0, + "content": "[69] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 512-523, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.692, + 0.905, + 0.747 + ], + "angle": 0, + "content": "[70] Yuliang Xiu, Yufei Ye, Zhen Liu, Dimitrios Tzionas, and Michael J Black. Puzzle avatar: Assembling 3d avatars from personal albums. ACM Transactions on Graphics (TOG), 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.748, + 0.905, + 0.831 + ], + "angle": 0, + "content": "[71] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[72] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3, 4" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.945, + 0.519, + 0.956 + ], + "angle": 0, + "content": "10408" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.147 + ], + "angle": 0, + "content": "[73] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 2, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.149, + 0.482, + 0.231 + ], + "angle": 0, + "content": "[74] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 4, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.302 + ], + "angle": 0, + "content": "[75] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844-1854, 2024. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.303, + 0.482, + 0.37 + ], + "angle": 0, + "content": "[76] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.44 + ], + "angle": 0, + "content": "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 3, 4, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.442, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9947, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.512, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[79] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.635 + ], + "angle": 0, + "content": "[80] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.719 + ], + "angle": 0, + "content": "[81] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.721, + 0.482, + 0.775 + ], + "angle": 0, + "content": "[82] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.777, + 0.482, + 0.844 + ], + "angle": 0, + "content": "[83] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.846, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[84] Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. idol: Instant photorealistic 3d human creation from a single image. arXiv preprint arXiv:2412.14963, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "[85] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + }, + { + "type": "page_number", + "bbox": [ + 0.481, + 0.946, + 0.519, + 0.957 + ], + "angle": 0, + "content": "10409" + } + ] +] \ No newline at end of file diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_origin.pdf b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..7239ae577735e75998d287e5c768458bb705d73d --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/d3058941-5a3c-47b1-a217-ebdb5b94b775_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2abe3b60e5a84d6bb536651913e838f0c7f31f3cb0e1405a1d24ab8d0ef6c552 +size 8485330 diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/full.md b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/full.md new file mode 100644 index 0000000000000000000000000000000000000000..eb47b5008c825a418ecb3f4a3c7231e40eae3ad0 --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/full.md @@ -0,0 +1,480 @@ +# Expressive Talking Human from Single-Image with Imperfect Priors + +Jun Xiang $^{1,2}$ + +Yudong Guo $^{1*}$ + +Yancheng Yuan2 + +1University of Science and Technology of China + +Leipeng Hu + +Boyang Guo + +Juyong Zhang + +$^{2}$ The Hong Kong Polytechnic University + +![](images/c16c23a4569ecbb3eeaed8bf07a5e9408122317482dbf5661749dc23a15510a8.jpg) +One-Shot Photos +Figure 1. Given a one-shot image (e.g., your favorite photo) as input, our method reconstructs a fully expressive whole-body talking avatar that captures personalized details and supports realistic animation, including vivid body gestures and natural expression changes. + +![](images/64c18b8bcd57950a232f1518eae3b0bb4624793e605782833db13b035c89dac8.jpg) +Animatable Expressive Talking Avatars + +# Abstract + +Building realistic and animatable avatars still requires minutes of multi-view or monocular self-rotating videos, and most methods lack precise control over gestures and expressions. To push this boundary, we address the challenge of constructing a whole-body talking avatar from a single image. We propose a novel pipeline that tackles two critical issues: 1) complex dynamic modeling and 2) generalization to novel gestures and expressions. To achieve seamless generalization, we leverage recent pose-guided image-to-video diffusion models to generate imperfect video frames as pseudo-labels. To overcome the dynamic modeling challenge posed by inconsistent and noisy pseudo-frames, we introduce a tightly coupled 3DGS-mesh hybrid avatar representation and apply several key regularizations to mitigate inconsistencies caused by imperfect labels. Extensive experiments on diverse subjects demonstrate that our method enables the creation of a photorealistic, precisely animatable, and expressive whole-body + +talking avatar from just a single image. Project page: https://ustc3dv.github.io/OneShotOneTalk/. + +# 1. Introduction + +Realistic rendering and precise control of gestures and expressions for whole-body talking avatars hold significant potential for AR/VR applications, such as telepresence and immersive remote conferencing. While extensively studied in the fields of 3D Vision and Graphics, most production-level body avatars still require a light stage [11, 65] to capture hours or minutes of multi-view frames, making them unsuitable for typical consumer usage. To address this, we focus on the challenging task of constructing an animatable whole-body avatar that supports both photo-realistic rendering and precise 3D control of gestures and expressions, from only a single image. This one-shot pipeline primarily faces the following challenges. + +Complex dynamic modeling. Humans exhibit complex gestures and facial movements when communicating with each other. To model the motion space of the entire body, SMPL-X [46] integrates several previous mod + +els [37, 40, 54] to capture body, hand, and facial movements. With the help of these parametric geometric models, recent personalized human avatars learn to model dynamic geometry and appearance by integrating neural textures [13, 38], hybrid representation [5, 17, 27], neural radiance fields [29, 47, 49, 64, 79], and 3D Gaussians [20, 22, 33, 35, 51] for dynamic photo-realistic rendering. To capture the full appearance of a whole-body avatar, these methods require complete observations, typically relying on dense input data for supervision, such as multi-view videos or self-rotating monocular videos. Moreover, these approaches primarily focus on body motion and novel view synthesis, falling short in accurately capturing and animating diverse facial expressions and hand movements, which limits their practical applicability. Due to the reliance on dense inputs, realistic and expressive whole-body modeling from a single image remains an unsolved challenge in this field. + +Generalization to novel gestures and expressions. Another key challenge lies in the limited ability to generalize to a wide variety of gestures and facial expressions, especially those that deviate significantly from the poses seen during training. This limitation arises from the data-driven nature of current methods, where the training data typically encompasses only a finite set of gestures and expressions. Consequently, when animating avatars, regions or motions that were underrepresented in the training data—such as dynamic clothing, inner mouth movements, or intricate hand gestures—are often poorly synthesized or omitted entirely. These issues are further compounded when working with a single image input, as the limited visual cues often fail to capture the full range of motion and expression. Recently, several approaches [19, 59, 62, 63, 71, 77, 83] have leveraged image and video diffusion models [3, 53] to achieve image-to-video generation guided by whole-body landmarks [73]. Although various model architectures and training techniques have been explored to improve generation quality, temporal consistency remains a challenge, and these methods have several key limitations. For example, body landmarks fail to disentangle identity and pose, leading to poor cross-identity animation results, with facial and body distortions and identity mismatches. Moreover, the sparse 2D landmarks are insufficient for precisely controlling gestures and facial movements, as they do not accurately capture the underlying dynamic geometry. + +In this work, we propose a novel pipeline to address both challenges in the one-shot setting. To generalize to diverse gestures and facial movements, we preprocess the large-scale TED Gesture Dataset [74] to build a comprehensive whole-body motion space for people talking. We then use these motion sequences as guidance to drive the input single image with a pre-trained whole-body video diffusion model [77] and a 3D face animation model [7]. This + +enables us to generate various video sequences of the target person performing different gestures and expressions. However, directly using these pseudo-frames to train a body avatar leads to unsatisfactory results. First, as noted above, current diffusion-based body animation methods struggle with temporal consistency, identity preservation, and motion alignment. Directly employing them as supervision results in significant blurring, distortion, and identity degradation. Furthermore, as we adapt two individual approaches for generating pseudo-frames for body gestures and facial expressions, respectively. Due to inconsistencies in camera spaces and rendering procedures, merging them into a unified avatar representation poses additional challenges. + +To tackle the challenge of complex dynamic modeling, we utilize both the single input image and imperfect pseudoframes to train a hybrid mesh-3DGS avatar representation, constrained by several carefully designed regularizations. The single input image provides an accurate, though incomplete, appearance for the body avatar, while the pseudo videos offer imperfect but more complete visual cues. For the pseudo video labels, instead of using per-pixel losses, we employ a perceptual-based loss term [76], which helps achieve reasonable appearance modeling while alleviating misalignment in the pseudo-frames. + +To further alleviate the inconsistencies caused by the pseudo-frames, we adopt a tightly coupled mesh-3DGS hybrid avatar representation. By introducing Laplacian smoothing and normal-consistency regularization on the deformed body mesh, we ensure that the structure of the 3D Gaussians used for rendering is well-constrained. Finally, we supervise the avatar representation using both gesture and head pseudo-frames, enabling the creation of a photorealistic, precisely animatable, and expressive whole-body avatar. + +In summary, our contributions include the following aspects: + +- We introduce a novel pipeline that overcomes the key challenges of building a whole-body expressive talking avatar from a single image. +- Our diffusion guidance strategy effectively extracts valuable knowledge from imperfect diffusion outputs and combines it with the limited information from the input image, enabling complete modeling of the talking avatar. +- Our carefully designed 3DGS-mesh coupled avatar representation, along with essential regularization techniques, facilitates accurate modeling of diverse subjects and stabilizes the optimization process. + +# 2. Related Work + +Human Gaussian Splatting. 3D Gaussian Splatting [31] is the state-of-the-art method for scene reconstruction and novel view synthesis (NVS), offering superior rendering speed and visual quality. This has significantly influ + +enced human avatar studies. Methods using multi-view videos [30, 43, 45, 51, 85] demonstrate excellent performance, with unique designs such as ASH [45], which achieves efficient Gaussian learning via mesh UV parameterization. For monocular video input, most human Gaussian splatting methods [18, 20, 22, 35, 42, 57] link Gaussian fields to parametric mesh models, often using additional regularization terms. ExAvatar [42] applies connectivity-based regularizers to short, casually captured videos. For sparse-view images, GPS-Gaussian [81] achieves real-time human NVs by encoding human priors into the network, while HumanSplat [44] generates high-quality static reconstructions from one-shot inputs. Concurrent works [52, 61] introduce multi-view diffusion models but are limited to the head or body areas. IDOL [84] leverages a large-scale dataset and shows great performance in single-view human reconstruction. However, it still struggles with handling facial dynamic. By contrast, our method is capable of recovering a realistic, animatable whole-body talking avatar from just a single image. + +Avatar Reconstruction from Few-Shot Images. Some works [8, 28, 67] apply 3D GAN inversion for one-shot human reconstruction, but they struggle with preserving personal details and generalization. PIFu [55] and subsequent works [6, 16, 56, 68, 69, 82] introduce pixel-aligned features and neural fields for image-based human reconstruction. An alternative approach leverages diffusion priors to fill in missing details, such as training human-centered diffusion models [2, 15, 44], using novel-view diffusion results for additional supervision [36, 39, 78], and employing Score Distillation Sampling (SDS) [50] to generate 3D avatars from 2D priors [24, 70, 72, 75]. However, these methods focus on static scenes and overlook dynamic human motion, limiting their ability to capture human dynamic. The SMPL-based [40] representation enables ELICIT [23] and SHERF [21] to model human dynamic, but overlooks hand and facial motions, restricting expressive animation capabilities. In contrast, our approach leverages priors from a pose-guided human video diffusion model, capturing both human appearance and dynamic, and enabling expressive full-body animations, particularly in the hands and face. + +Pose-Guided Human Video Diffusion. Pose-guided human video diffusion models [9, 19, 59, 63, 71, 77, 83] directly generate animated videos from a reference image and pose sequence, bypassing traditional 3D reconstruction and rendering processes. The success of these models depends on the quality of training data, model design, and pose guidance. Some approaches [9, 19, 71, 83] incorporate temporal layers to ensure smooth transitions, inspired byAnimateDiff [12], while others [48, 59, 77] use video diffusion models [3, 4] for dynamic sequences. Pose guidance is typically provided by OpenPose [19, 63, 77], DensePose [71], + +depth maps [9], or SMPL [83]. StableAnimator [62] proposes HJB-based face optimization that further enhances face quality, and Make-Your-Anchors [25] personalizes outputs by fine-tuning models on identity-specific images. We adopt MimicMotion [77] for its superior performance in handling hand regions. Despite their strengths, these 2D models still face challenges, such as image distortion, identity changes, and pose misalignment, due to the lack of 3D understanding. To address these, we leverage optimized avatar representations with carefully designed constraints, improving consistency and naturalness in the generated animations. + +# 3. Method + +Given a single image of the target person, we aim to reconstruct a 3D talking avatar that fully inherits the identity and enables natural animation. To address the challenge of complex dynamic modeling from imperfect pseudo-frames, we adopt a tightly coupled 3DGS-mesh hybrid avatar representation (Sec. 3.1). To generalize well to diverse gestures and facial movements, we generate imperfect sequences of the target person driven by various motion sequences (Sec. 3.2). Finally, we introduce the carefully designed constraints and loss terms to train the representation from noisy videos effectively (Sec. 3.3). The entire pipeline is illustrated in Fig. 2. + +# 3.1. Coupled 3DGS-Mesh Avatar + +Whole-body parametric mesh models facilitate human animation and provide good initialization, while 3DGS offers enhanced expressiveness and realistic rendering. To address the challenges of the one-shot task, we design a novel coupled 3DGS-mesh representation, which effectively integrates the geometric priors and surface regularization of the mesh without diminishing the expressive capability of the Gaussian field. + +We couple the 3DGS field with the typical SMPL-X model, which is formulated as follows: + +$$ +M (\beta , \theta , \phi) = W (T (\beta , \theta , \phi), J (\beta), \theta , \mathcal {W}), \tag {1} +$$ + +$$ +T (\beta , \theta , \phi) = \bar {T} + B _ {S} (\beta) + B _ {E} (\phi) + B _ {P} (\theta), \tag {2} +$$ + +where $B_{S}(\cdot), B_{E}(\cdot)$ and $B_{P}(\cdot)$ denote shape, expression, and pose blend functions respectively, with $\beta, \phi$ and $\theta$ representing the corresponding parameters. $\bar{T}$ is the template mesh, and $W(\cdot)$ is the standard LBS function that rotates the vertices in $T$ around the estimated joints $J$ , smoothed by the blend weights $\mathcal{W}$ . Since the shape code $\beta$ is fixed once registered, we will omit it in the later statements and denote $T = T(\beta, \theta, \phi)$ and $J = J(\beta)$ for simplicity. Inspired by [1, 45, 66], we initialize the 3D Gaussians on the canonical mesh surface using UV parameterization. For a + +![](images/136c8b08ef624783e8463e48c14d08b62a627a88e3190fa1058e033d6fd3d924.jpg) +Figure 2. Overview. Our method constructs an expressive whole-body talking avatar from a single image. We begin by generating pseudo body and head frames using pre-trained generative models, driven by a collected video dataset with diverse poses. Per-pixel supervision on the input image, perceptual supervision on imperfect pseudo-frames, and mesh-related constraints are then applied to guide the 3DGS-mesh coupled avatar representation, ensuring realistic and expressive avatar reconstruction and animation. + +3D Gaussian located on a triangle $k = \{\mathbf{V}_1, \mathbf{V}_2, \mathbf{V}_3\}$ of mesh $T$ with barycentric coordinates $(u, v)$ , the surface position is given by: + +$$ +\operatorname {B a r y} (T) = \mathcal {V} (k, u, v) = u \mathbf {V} _ {\mathbf {1}} + v \mathbf {V} _ {\mathbf {2}} + (1 - u - v) \mathbf {V} _ {\mathbf {3}}. \tag {3} +$$ + +Coupled 3DGS-Mesh Deformation. To model clothes, haircuts, and other complex regions that the SMPL-X mesh fails to handle, the deformation of 3D Gaussians is crucial. While previous works can achieve correct deformation with sufficient supervision, we need to impose additional constraints on the deformation field to prevent overfitting to the single input image and imperfect pseudo-frames, while also facilitating the integration of other modules. Based on these considerations, we design two deformation fields in the canonical space and enforce their proximity: one represents the conventional Gaussian deformation, and the other represents the critical mesh deformation. This approach allows us to indirectly influence the Gaussian deformation by applying soft constraints on the deformed mesh. Moreover, since there is no strict binding between the deformed mesh surface and the Gaussians, complex regions are still effectively handled. The two deformation fields, together with the full-body animation and Gaussian Splitting process, are formulated as follows: + +$$ +G (\theta , \phi , \mathbf {P}) = S P (W (B a r y (T) + d X, J, \theta , B a r y (\mathcal {W})), \mathbf {P}). \tag {4} +$$ + +$$ +M (\theta , \phi) = W (T + d T, J, \theta , \mathcal {W}), \tag {5} +$$ + +where $G(\cdot)$ represents the final rendered image, $SP(\cdot)$ and $\mathbf{P}$ denote the splatting rendering process and the properties of 3D Gaussians, and $dX$ and $dT$ represent the optimized deformations of the Gaussians and mesh vertices, respectively. We apply key regularizers on $dT$ and propagate soft constraints to $dX$ by encouraging $dX$ to align with $Bary(dT)$ . + +# 3.2. Pseudo-Frames Generation + +In the one-shot image setting, many regions are unseen or occluded. To construct a complete avatar and ensure generalization to novel gestures and expressions, we turn our attention to recent advances in human motion diffusion models [19, 77] and head animation techniques [7, 10]. + +For leveraging diffusion-based generative models, the SDS loss [50] is widely used in text/image-to-3D works [24, 60, 72, 75]. Although diffusion models may introduce moderate deviations, the SDS technique remains effective in these studies. However, for tasks such as realistic dynamic avatar animation from a one-shot image, pose alignment is critical [18, 42], and the misalignment introduced by 2D diffusion models cannot be overlooked. Therefore, directly applying SDS in our setting is suboptimal. Instead, we focus on extracting accurate information from pseudo-frames synthesized by these generative models. + +To generate whole-body pseudo-frames, we collect a set of SMPL-X pose sequences $\{(\theta_{N_i},\phi_{N_i})_{i = 1}^F\}_{j = 1}^K$ from the TED Gesture Dataset [74] as input to the generative models. Since no unified generative model exists for both body and + +![](images/d060dc60fdd93b5cfc853f8f1c7920ab535fa33c8a270767db6dad919d8e1634.jpg) + +![](images/fa99557ef1b3e52b7ecce646f40d38e31ed3bbae2a4c532fd8b26a74f394f931.jpg) + +![](images/dc5746401070b29c558c127a26c625f1906bad85c1790a4ed23eb00c1571576f.jpg) + +![](images/4a1241aa7d6933c48913f53f7cb4c94724184651c5bfad9471798eb075aabe0f.jpg) + +![](images/8c6d967c4b15c3c87b2c7bd4dec887c95f4fe9ae1d8e7d672a3c063f7ff2fab4.jpg) + +![](images/4eaab04a0012cae545c06daefa546ce43b242f18efe28f6376ce16d097c46287.jpg) + +![](images/f41e91d1723bab134428e064568faa9309a028d0082080b22aee1f20ee478750.jpg) + +![](images/77fa1ef8cfc5c11e6e54eb6fc8bc32501396ac5eab30731991ef1c3d09185311.jpg) + +![](images/6e2f399b3ae495ed6e4747d3cd740ab0bab5bb3442ecfc9200cf3c4e74f81d09.jpg) + +![](images/ce4303bb171cd4202d2c921a368427dcd6d997659aef6e71f430e22680c811fd.jpg) + +![](images/5134cb783d35dac4a79c5a83f194f78f74fcc0764c90609a7391fee3d9484042.jpg) + +![](images/37e59c5f3e0eb9c9a0f70969bcceb18348ded7f1fad84121cb415c055fdeaec2.jpg) + +![](images/3a383c78e917258b4da566b6ce601465c703c962949a9091542b5c64a7fa5af0.jpg) + +![](images/a636c897480b0534ce5c8b847029b1fb418e05e8d86ed0f6e3612f3d3ec48ee0.jpg) + +![](images/c4921b9a9dfd82c7dfe992d4b3ff6f4b05a326e356c2dd5e889a108b8b60e326.jpg) + +![](images/82e338918f226678a8764f87d87b32e758ec31a7b3bd019625da8cfa4a12697e.jpg) + +![](images/b11488b81a7d183fc2d9156071a7680ebc12986c20a63bd0dc957fb56e8adb98.jpg) + +![](images/c8b9ccd8fd22c8e0d514f531e08c7c03fc013d004967758bfe96f6ea801b24a5.jpg) + +![](images/d97f9030bf1e648ce910525a839387844c6dbd2fb54d26d7b0d96d97ad2264ec.jpg) + +![](images/fbe806b5bc89ac407db8b293fe92098464e8941abf07cdfcbc1add316ec4eccb.jpg) +Input Image +Driving Image +Figure 3. Qualitative comparisons with representative methods [23, 42, 62] in the cross-identity motion reenactment task. Our method achieves accurate and realistic animation with almost all fine details preserved and identity unchanged. + +![](images/a53440c716403b9ee9a262321371d9c588bc5e1d2ccb6262e79ad2a6b5cccab0.jpg) +Ours + +![](images/bcaa1bcd7e6b534197cbfe1f07683ad7a36b4ae8cf54cfec5812a81cf3fc1a63.jpg) +StableAnimator + +![](images/b197a1cab344f42dcfaf217c3357ad5bc9c21306589dc0dacf50a081ad4f203a.jpg) +ExAvatar-1shot + +![](images/ee77569f867464bf3251e8c18cf2febd89b461511a6f71b88f540cee2914c579.jpg) +ELICIT + +head simultaneously, we employ two separate approaches to generate pseudo-frames for body gestures and facial expressions, respectively. Given a source image $I_{S}$ registered with the SMPL-X parameters $(\beta_{S},\theta_{S},\phi_{S})$ , we adopt MimicMotion [77] with the collected pose sequences to generate various pseudo body frames: + +$$ +I _ {N} = \operatorname {M o t i o n} \left(I _ {S}, D \left(\beta_ {S}, \theta_ {N}, \phi_ {N}\right)\right), \tag {6} +$$ + +where $D(\cdot)$ denotes the mapping from SMPL-X parameters to DWPose [73]. We randomly set the root body pose in $\theta_N$ for each segment to increase viewpoint generalization, with pitch $\in (-30^{\circ}, 30^{\circ})$ and yaw $\in (-10^{\circ}, 10^{\circ})$ . After that, we re-track $I_N$ to obtain more accurate pose parameters $(\hat{\theta}_N, \hat{\phi}_N)$ , while keeping the shape code $\beta_S$ fixed. The re-tracking process is crucial, as it overcomes the misalignment introduced by 2D diffusion models and helps achieve precise and realistic results. For the head region, we adopt Portrait4D-v2 [7] to generate various pseudo frames of the target person performing diverse expressions, also driven by the videos from the TED Gesture Dataset. + +# 3.3. Objective Functions + +Assisted by our novel hybrid 3DGS-mesh avatar representation, we introduce several carefully designed regularization + +terms to stabilize the avatar reconstruction process and effectively extract the correct information from both the one-shot input and the imperfect pseudo-frames. + +Mesh-related Constraints. We use the following loss functions to apply soft constraints on the Gaussian field based on the mesh: + +$$ +\begin{array}{l} \mathcal {L} _ {S C} = \lambda_ {\text {n o r m a l}} \mathcal {L} _ {\text {n o r m a l}} + \lambda_ {\mathrm {M}} \mathcal {L} _ {\mathrm {M}} \\ + \lambda_ {\mathrm {M G C}} \mathcal {L} _ {\mathrm {M G C}} + \lambda_ {\mathrm {l a p}} \mathcal {L} _ {\mathrm {l a p}}. \\ \end{array} +$$ + +(7) + +Here, $\mathcal{L}_{\mathrm{normal}}$ is the normal consistency loss applied to the deformed mesh surface, ensuring surface normal consistency post-deformation. $\mathcal{L}_M$ is the mask loss, which measures the discrepancy between the ground truth mask and the mask of the deformed mesh rendered via [34]. These two losses work together to regulate the behavior of $dT$ , while influencing Gaussian deformation $dX$ through the mesh-Gaussian consistency loss: + +$$ +\mathcal {L} _ {\mathrm {M G C}} = \| d X - B a r y (d T) \| _ {1}. \tag {8} +$$ + +Additionally, we compute the Laplacian smoothing loss $\mathcal{L}_{\mathrm{lap}}$ for $dX$ along with the scaling and RGBs of the 3D Gaussians, based on their initial connectivity state on the canonical mesh surface. + +Perceptual Supervision of Pseudo-Frames. The synthesized pseudo-frames exhibit noticeable artifacts, such as image distortion and lack of 3D consistency, especially in the results generated with MimicMotion. These artifacts cannot be fully corrected through pose alignment alone. As a result, pixel-aligned losses like L1 or MSE often lead to issues like texture flickering, blurring, and identity changes. However, the high-level human structure is consistently well-preserved in the source image, which aids human recognition and is empirically referred to as perceptual similarity. Previous works [41, 76] have used deep neural networks to capture these perceptual structure features. We adopt the LPIPS perceptual loss [76], using VGG [58] as the backbone. By learning deep perceptual human features from dynamic poses and preserving detailed information from the source image, our method is able to conduct realistic and complete talking avatar animations. Thus, for the pseudoframes, we primarily use the following perceptual loss: + +$$ +\mathcal {L} _ {\text {d i f f}} = \lambda_ {\mathrm {L P I P S}} \mathcal {L} _ {\mathrm {L P I P S}} \left(I _ {N}, G \left(\hat {\theta} _ {N}, \hat {\phi} _ {N}, \mathbf {P}\right)\right). \tag {9} +$$ + +Per-pixel Supervision of the Input Image. For the source image $I_S$ with pose $(\theta_S, \phi_S)$ , we use the regular L1 loss, SSIM loss, and mask loss between $I_S$ and the splatted image $G(\theta_S, \phi_S, \mathbf{P})$ : + +$$ +\mathcal {L} _ {S} = \lambda_ {\mathrm {R G B}} \mathcal {L} _ {\mathrm {R G B}} + \lambda_ {\mathrm {S S I M}} \mathcal {L} _ {\mathrm {S S I M}} + \lambda_ {\mathrm {a l p h a}} \mathcal {L} _ {\mathrm {a l p h a}}. \tag {10} +$$ + +The final loss function is the sum of all the aforementioned losses. + +# 3.4. Implementation Details + +We set the UV map resolution to 512 and use approximately 150,000 Gaussians in total. We train our models using the Adam optimizer [32] with $\beta = (0.9, 0.999)$ . The learning rates for the Gaussian parameters are the same as those in the official implementation, while the learning rates for both the Gaussian and mesh deformation fields are set to $\eta = 1e^{-4}$ . For the loss weights, we set $\lambda_{\mathrm{normal}}$ , $\lambda_{\mathrm{M}}$ , $\lambda_{\mathrm{MGC}}$ , $\lambda_{\mathrm{lap}}$ , $\lambda_{\mathrm{LPIPS}}$ , $\lambda_{\mathrm{RGB}}$ , $\lambda_{\mathrm{SSIM}}$ , $\lambda_{\mathrm{alpha}}$ to $1e^{-2}$ , $1e^{-1}$ , $1e^{1}$ , $1e^{2}$ , $2e^{-1}$ , $8e^{-1}$ , $1e^{-1}$ , and $4e^{-1}$ , respectively. We start adding the perceptual loss at the 2000th step, while the other losses are used throughout the entire training process with a maximum number of training steps of 20000. We adopt joint training with separate supervision for body and head pseudo-frames. Training takes 3 hours on single 4090 due to critical but inefficient re-tracking. But avatar reconstruction and rendering are highly efficient with an inference speed of 50FPS+ at 1K resolution. + +# 4. Experiments + +# 4.1. Dataset + +We use the poses and expressions processed from 100 videos of the TED Gesture Dataset [74] as the motion + +sequences during training. We separately sampled about 1,800 frames with the most distinct variations in poses and expressions. For evaluation, the one-shot input and driving poses are primarily sourced from the ActorsHQ [26], NeuMan [29] and the Casual Conversations Dataset [14]. All videos and images are cropped and resized to a fixed aspect ratio of 9:16. For foreground segmentation, we use BiRefNet [80], and for pose tracking, we employ the custom fitting procedure from ExAvatar [42]. + +# 4.2. Comparison with Representative Methods + +We compare our method with several representative works, including: (1) ExAvatar [42], a recent SOTA work that models human avatars with a mesh-based 3D Gaussian field. (2) ELICIT [23], a one-shot NeRF-based animatable avatar. (3) StableAnimator [62], a SOTA general 2D pose-guided human video diffusion model. (4) Make-Your-Anchors [25], a 2D diffusion-based method that is fine-tuned on a one-minute video clip of the individual to enhance identity information. + +Note that for ExAvatar, since it takes short videos as input, we compare with two versions of it: ExAvatar-40shot, which uses 40-shot images, and ExAvatar-1shot, which uses one-shot images as input. For Make-Your-Ananchor, as we find it does not perform well on one-shot input, we only compare it with available video input by fine-tuning it on a short video clip. + +Qualitative Comparisons. Fig. 3 and Fig. 4 present a qualitative comparison between our method and the other representative approaches. For Fig. 3, we use the pose of a different identity as the driving signal. For Fig. 4, we compare the performance of these methods on the self-driven pose reenactment task, using subjects that have corresponding video data. + +As observed, ExAvatar-1shot tends to overfit the input image and fails to recover accurate textures in regions occluded by hands. Even with 40-shot input, ExAvatar still struggles with incomplete knowledge and fails to handle novel gestures effectively. StableAnimator generates relatively reasonable results but is constrained by its training distribution and struggles with identity consistency across frames, often leading to appearance mismatches and identity changes. ELICIT, which uses a NeRF-based representation, ensures 3D consistency but relies on SMPL for geometry, which neglects the hand region. This coarse semantic proxy fails to support complex hand reconstruction or animation. Make-Your-AnAnchor, although pre-trained on multiple identities to learn human motion priors, requires a long video with sufficient movement and appearance data to adapt to new identities. It struggles with short fine-tuning videos and fails to recover fine details, especially for gestures outside its fine-tuning distribution. + +In contrast, our method generates expressive talking + +![](images/b88b938fcf40d67c0b10fdeae381ff0e72dbf0628c6bb89c5b442f4b425570d9.jpg) + +![](images/461498c5fb7d9605afc9d9c8ca35e9599e577bf2b3d15ea52272df3c758b13d9.jpg) + +![](images/f21d406557de3ecffb7058f27f10929fa928621f5494ca1310d8a66cf21b5d99.jpg) + +![](images/28a8cf84ba456eb9ea039e3d45c9367be52a28b7c430b7b21ec0f4355264cb27.jpg) + +![](images/99e84388e162bfebe18a5f3d40ac5f8d784ba4c19d7f2840e6961196ed978559.jpg) + +![](images/603e0207c2597979f3e0b266b2a3d0835e8a6649914f651ae6c822328f06fdaa.jpg) + +![](images/f2b48159c359213127748e10c4429d218473e4b0f66919f5755d40609f04d78b.jpg) + +![](images/8d1925219951917a259deea15803b4159199f78824f3b9c167b154303fe1ca06.jpg) +Input Image +Ground Truth + +![](images/b80ac9d6ceecda1eb17aa851b2a3dbc00a354cf57015e734d516d82de5e2dd14.jpg) +Ours + +![](images/1097f9c3c141bcda207f1e39e9040ba895e4cfa234f218fc346f4f7e5038130f.jpg) +StableAnimator + +![](images/a920911ef28c9476a6151cbaa5b50d642c681765b59fc218a48a08294cc93791.jpg) +ExAvatar-40shot + +![](images/25adc51e724a59fe944e88ac02cb50868217549157043c4d43e3e3c1a000be86.jpg) +ExAvatar-1shot + +![](images/0eb6d0a8e929f7e2c12a04ef6aa26fedc961025d4895aca9b6610f10e9db7284.jpg) +ELICIT + +![](images/b50d0db4bb81c52a06542b2c337530d79305a08481a0b3aff61bfd1a4021628d.jpg) +Make-Your-Anchors + +
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
ELICIT5.65522.5590.9210.0660
ExAvatar3.93124.2150.9240.0409
StableAnimator1.88327.3500.9350.0327
Ours1.21629.3110.9430.0299
+ +Table 1. Quantitative comparisons with representative methods [23, 42, 62] on self-driven data. Our method outperforms others in pixel-wise error metrics, realism evaluation metrics and perceptual quality metrics. (ExAvatar here denotes ExAvatar-40shot.) + +avatars from a single input image. Our results preserve fine human details and achieve natural animation with excellent rendering quality. We provide additional examples of cross-identity pose reenactment in Fig. 5. Using the same driving pose, identities with completely different attributes can be driven in the same way, thanks to the SMPL-X model and the 3DGS-mesh coupled representation. + +Quantitative Comparison. Tab. 1 presents a quantitative comparison between our model and other methods. Although the one-shot image animation task typically lacks a strict test set for quantitative comparison, we perform the evaluation on the self-driven task and use four common metrics for comparison: Mean Squared Error, PSNR, SSIM, and LPIPS. Our method outperforms all others across these metrics, demonstrating superior realism and 3D consistency in the results. However, we believe that these metrics do not fully capture the quality or capabilities of the methods, especially for the one-shot image animation task. We encourage readers to refer to the video results of our method for a more comprehensive and objective evaluation. + +![](images/c28295904109f88173963caecf54b2d211b0e4b1f0ecaa2b55b2912765d94f1e.jpg) + +![](images/25699cc9b99af0fbe2b2e70a26eb4a3e9654ffe701167d3ff592ec2f5d7ad951.jpg) +Driving Pose + +![](images/8564196c94e1af9126a725ed2d3f0f212c97e3230d44b508a219b94a44b98a43.jpg) + +![](images/6f275feae8d6b4fadb57e84f8bcd5e4048ba6e5deaf015a08c9642828bacc781.jpg) + +![](images/dab3b0faa9d80bd568e02481b13fc246038cf21c99950a2eae3501056e60a0db.jpg) + +![](images/10fbd9de7b0cd184fc6baedfa8a046b30f1d26c55fb37ea4c3f895859c673c1b.jpg) +Inputs + +![](images/59412e687f8130757fd32dea8b2e7fd53b3b61ca0168eec2b808b4d5fb31e9cd.jpg) + +![](images/8cf9c90109f0a6f56607c552f22ed7dd3e80d0055fdad7bce1fedbfcd4003387.jpg) +Result1 + +![](images/16bf92d9cffe3145086c9ad41e4595d922e8ad0cd2724b58aa7d042bfafb358e.jpg) +Figure 4. Qualitative comparisons with representative methods [23, 25, 42, 62] in the self-driven motion reenactment task. Our method well models facial and hand regions, which matches the input image most in global identity preservation and local details modeling, even compared with some methods trained on captured videos. + +![](images/87b15cc5dabf966138227fadf700527c0fc0db1b5ddc633b97ac0625b19a23e5.jpg) +Result2 + +Figure 5. More examples of cross-identity pose reenactment. Different subjects can be accurately animated with the same poses. + +
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
w/o-mesh-SC1.5327.990.93800.0299
L1&SSIM-PFG1.4228.400.94160.0347
w/o-retrack1.5227.990.93820.0296
w/o-pseudo-face1.4628.240.93910.0293
Full-model1.2229.310.94310.0299
+ +Table 2. Quantitative results of ablation studies. + +# 4.3. Ablation Studies + +Perceptual-based Pseudo-Frames Guidance. Pseudoframes guidance is the core concept behind this work, enabling the challenging task of generating animatable and expressive talking avatars from a single input image. As shown in Fig. 6, without diffusion guidance, we are unable to inpaint unseen regions effectively, and the result suffers + +![](images/71b77f527cbfc7ac8fc58e98448da6cc85c9c4f745f84965659f4f8b222045ae.jpg) +Input Image + +![](images/ada814de8703553ec9d213fc8807f4f1224032926ff8c2ac37f52c62d3da9de2.jpg) +Driving Pose + +![](images/feb678e3053189aac911b764ddd593fad38389147ed7a940c4107e59ded03cf2.jpg) +Full Model + +![](images/639f29e821f3d0267212d23c053479a9625d2d0092f227147beb2c6ea32d806e.jpg) +L1&SSIM PFG + +![](images/58e5f24b793c74a898161b94e4999da9f35e1ab808a0e14fbc74a123fc436825.jpg) +w/o PFG + +![](images/a48351aba68138112958e91e6a898b1dfd2d9daed9f34873145c2daa9c5b977b.jpg) +w/o Re-Tracking + +![](images/71f3ade00f3ed16e031235d599840d32f89cbc1e1be13cff8650b6d3fe4c92e8.jpg) +Figure 6. Visualization of ablation studies on perceptual-based pseudo-frames guidance (PFG) and re-tracking step. +Input Image + +![](images/370405f1fa9eb19c7ff0d136f340c106528b32491f89e6404a2b5b0a6bd0239e.jpg) +Full Model + +![](images/fcb35dfad717117a501e77ca41be6e1cbd88267cb4a72de4ecf7ac578f0a9ece.jpg) +w/o Pseudo Face + +![](images/8b49a100e6a31db0220f63bfa0b56060a244a48fa5ed83dbb7863177fb592e97.jpg) +w/o mesh SC + +![](images/ac02986e65c94cd3a15d2445c3662e3c6cd1fbafeff3d224c5984264937fe32d.jpg) +Figure 7. Visualization of ablation studies on pseudo face supervision and mesh-related soft constraints. +Input Image +Figure 8. Novel view synthesis. Our method demonstrates robust performance within a view range of -30 to 30 degrees. However, its performance degrades for larger viewing angles + +![](images/5654ae2390b62f8a82cd304d4677cd7dca27b03ad241d1b732b0e8dbc8d1bbf7.jpg) +$20^{\circ}$ + +![](images/350ef94e58b722de0ca470dc007f096871fbc3bd569f76d5cdc2cd77d3c6ecbe.jpg) +$30^{\circ}$ + +![](images/52c748fc6f6b9d9492f6039731a0b85fbf154817d21bcacd2b8b672517468b0a.jpg) +$45^{\circ}$ + +![](images/0cf30b73d88edcde663adc98bd9328734d4fd761dd28854b6a5400009861cad7.jpg) +$60^{\circ}$ + +from severe overfitting to the input image. Additionally, due to the misalignment and inconsistency of the pseudo-output generated by motion diffusion models, using pixel-based losses like L1 for diffusion guidance results in overly smooth outputs and struggles to capture fine details, especially in facial and hand regions. In contrast, the perceptual diffusion guidance we employ not only preserves the full personalized attributes of the input image but also inpaints unseen regions with more natural and consistent textures. + +Re-Tracking Step. The re-tracking procedure applied to the pseudo-output of motion diffusion models helps mitigate the misalignment introduced by 2D diffusion models, preventing texture errors and detail loss. As shown in Fig. 6, the re-tracking step successfully recovers more accurate texture structures. + +Mesh-related Constraints. Our designed hybrid 3DGS-mesh avatar representation and the corresponding regularizations significantly enhance the expressiveness and integrity of the final results, as demonstrated in Fig. 7. When soft constraints for mesh deformation are omitted, geometric artifacts appear in specific regions and the Gaussian field fails to capture fine details. + +Pseudo Face Supervision. While perceptual-based guidance and hybrid 3DGS-mesh avatar representation enable pseudo human supervision to recover coarse head textures, incorporating targeted information from pseudo face frames further enhances facial quality, as presented in Fig. 7. + +Besides the significant visual results shown in Figs. 6 and 7, quantitative evaluation is also conduct, as seen in Tab. 2. + +# 5. Conclusion + +In this paper, we introduce a novel pipeline for creating expressive talking avatars from a single image. We propose a coupled 3DGS-Mesh avatar representation, incorporating several key constraints and a carefully designed hybrid learning framework that combines information from both the input image and pseudo-frames. Experimental results demonstrate that our method outperforms existing techniques, with our one-shot avatar even surpassing state-of-the-art methods that require video input. Considering its simplicity in construction and ability to generate vivid animations, our method shows significant potential for practical applications of talking avatars across various fields. + +Limitations. Severe mismatches between the input image and the registered parametric human mesh, especially in regions like fingers, can cause optimization issues and result in incorrect textures. Additionally, rendering large views or extending to full $360^{\circ}$ human reconstruction remains difficult due to current limitations in human motion diffusion models and the lack of data for large, novel viewpoints, as seen in Fig. 8. Future work will explore integrating semantic information from large language models and static priors from 3D reconstruction to address these limitations. + +Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62441224, No. 62272433, No. 62402468), The Hong Kong Polytechnic University (Project No. P0051019), the Research Center for Intelligent Operations Research of the PolyU and the advanced computing resources provided by the Supercomputing Center of the USTC. + +# References + +[1] Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gordon Wetzstein. Gaussian shell maps for efficient 3d human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9441-9451, 2024. 3 +[2] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3 +[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3 +[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3 +[5] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 2 +[6] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3 +[7] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. In Proceedings of the European conference on computer vision (ECCV), 2024. 2, 4, 5 +[8] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14916-14927, 2023. 3 +[9] Mengyang Feng, Jinlin Liu, Kai Yu, Yuan Yao, Zheng Hui, Xiefan Guo, Xianhui Lin, Haolan Xue, Chen Shi, Xiaowen Li, et al. Dreamoving: A human video generation framework based on diffusion models. arXiv e-prints, pages arXiv-2312, 2023. 3 +[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor + +trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4 +[11] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 1 +[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning, 2023. 3 +[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 2 +[14] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in ai: the casual conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021. 6 +[15] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 3 +[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. Stereopifu: Depth aware clothed human digitization via stereo vision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3 +[17] Chen Honghu, Peng Bo, Tao Yunfan, and Zhang Juyong. D $^3$ human: Dynamic disentangled digital human from monocular video. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2 +[18] Hezhen Hu, Zhiwen Fan, Tianhao Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. Expressive gaussian human avatars from monocular rgb video. In NeurIPS, 2024. 3, 4 +[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2, 3, 4 +[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3 +[21] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9352-9364, 2023. 3 +[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2, 3 + +[23] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3, 5, 6, 7 +[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531–1542. IEEE, 2024. 3, 4 +[25] Ziyao Huang, Fan Tang, Yong Zhang, Xiaodong Cun, Juan Cao, Jintao Li, and Tong-Yee Lee. Make-your-anchor: A diffusion-based 2d avatar generation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2024. 3, 6, 7 +[26] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6 +[27] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2 +[28] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3 +[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 6 +[30] Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19734-19745, 2024. 3 +[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2 +[32] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6 +[33] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: Human gaussian splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2 +[34] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5 +[35] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 3 + +[36] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 3 +[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (ToG), 36 (6):194-1, 2017. 2 +[38] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. 2 +[39] Zhibin Liu, Haoye Dong, Aviral Chharia, and Hefeng Wu. Human-vdm: Learning single-image 3d human gaussian splatting from video diffusion models. arXiv preprint arXiv:2409.02851, 2024. 3 +[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (ToG), 34(6), 2015. 2, 3 +[41] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 6 +[42] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In ECCV, 2024. 3, 4, 5, 6, 7 +[43] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3 +[44] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splattering with structure priors. arXiv preprint arXiv:2406.12459, 2024. 3 +[45] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1165-1175, 2024. 3 +[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1 +[47] Bo Peng, Jun Hu, Jingtao Zhou, Xuan Gao, and Juyong Zhang. Intrinsicngp: Intrinsic coordinate based hash encoding for human nerf. IEEE Transactions on Visualization and Computer Graphics, 2023. 2 +[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 3 + +[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2 +[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 4 +[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-athlete: Animatable avatars via deformable 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3 +[52] Lingteng Qiu, Shenhao Zhu, Qi Zuo, Xiaodong Gu, Yuan Dong, Junfei Zhang, Chao Xu, Zhe Li, Weihao Yuan, Liefeng Bo, Guanying Chen, and Zilong Dong. Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction, 2024. 3 +[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2 +[54] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6), 2017. 2 +[55] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3 +[56] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3 +[57] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3 +[58] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6 +[59] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. ICLR 2025, 2025. 2, 3 +[60] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.4 +[61] Jiapeng Tang, Davide Davoli, Tobias Kirschstein, Liam Schoneveld, and Matthias Niessner. Gaf: Gaussian avatar + +reconstruction from monocular videos via multi-view diffusion, 2024. 3 +[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024. 2, 3, 5, 6, 7 +[63] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2, 3 +[64] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2 +[65] Donglai Xiang, Fabian Prada, Zhe Cao, Kaiwen Guo, Chenglei Wu, Jessica Hodgins, and Timur Bagautdinov. Drivable avatar clothing: Faithful full-body telepresence with dynamic clothing driven by sparse rgb-d input. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1 +[66] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1802-1812, 2024. 3 +[67] Zhangyang Xiong, Di Kang, Derong Jin, Weikai Chen, Linchao Bao, Shuguang Cui, and Xiaoguang Han. Get3dhuman: Lifting stylegan-human into a 3d generative model using pixel-aligned reconstruction priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3 +[68] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3 +[69] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 512-523, 2023. 3 +[70] Yuliang Xiu, Yufei Ye, Zhen Liu, Dimitrios Tzionas, and Michael J Black. Puzzle avatar: Assembling 3d avatars from personal albums. ACM Transactions on Graphics (TOG), 2024. 3 +[71] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2, 3 +[72] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3, 4 + +[73] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 2, 5 +[74] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 4, 6 +[75] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844-1854, 2024. 3, 4 +[76] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 6 +[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 3, 4, 5 +[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9947, 2024. 3 +[79] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2 +[80] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6 +[81] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3 +[82] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 2021. 3 +[83] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2, 3 +[84] Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. idol: Instant photorealistic 3d human creation from a single image. arXiv preprint arXiv:2412.14963, 2024. 3 + +[85] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3 \ No newline at end of file diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/images.zip b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..ebf16eef88ec1c79325983599023dd7103dd8be0 --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2cbd5e0421f4bf013f206a63131650c4d8078b072b075f868f9905bf309cf001 +size 676478 diff --git a/2025/Expressive Talking Human from Single-Image with Imperfect Priors/layout.json b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..5b1909fba02d8dc405185d4434ef7336bf1f6737 --- /dev/null +++ b/2025/Expressive Talking Human from Single-Image with Imperfect Priors/layout.json @@ -0,0 +1,12519 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 94, + 103, + 517, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 94, + 103, + 517, + 121 + ], + "spans": [ + { + "bbox": [ + 94, + 103, + 517, + 121 + ], + "type": "text", + "content": "Expressive Talking Human from Single-Image with Imperfect Priors" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 133, + 142, + 198, + 158 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 133, + 142, + 198, + 158 + ], + "spans": [ + { + "bbox": [ + 133, + 142, + 198, + 158 + ], + "type": "text", + "content": "Jun Xiang" + }, + { + "bbox": [ + 133, + 142, + 198, + 158 + ], + "type": "inline_equation", + "content": "^{1,2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 222, + 143, + 294, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 222, + 143, + 294, + 157 + ], + "spans": [ + { + "bbox": [ + 222, + 143, + 294, + 157 + ], + "type": "text", + "content": "Yudong Guo" + }, + { + "bbox": [ + 222, + 143, + 294, + 157 + ], + "type": "inline_equation", + "content": "^{1*}" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 215, + 158, + 296, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 215, + 158, + 296, + 171 + ], + "spans": [ + { + "bbox": [ + 215, + 158, + 296, + 171 + ], + "type": "text", + "content": "Yancheng Yuan2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 132, + 172, + 309, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 132, + 172, + 309, + 185 + ], + "spans": [ + { + "bbox": [ + 132, + 172, + 309, + 185 + ], + "type": "text", + "content": "1University of Science and Technology of China" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 318, + 143, + 382, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 318, + 143, + 382, + 157 + ], + "spans": [ + { + "bbox": [ + 318, + 143, + 382, + 157 + ], + "type": "text", + "content": "Leipeng Hu" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 407, + 143, + 474, + 157 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 407, + 143, + 474, + 157 + ], + "spans": [ + { + "bbox": [ + 407, + 143, + 474, + 157 + ], + "type": "text", + "content": "Boyang Guo" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 320, + 158, + 394, + 172 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 320, + 158, + 394, + 172 + ], + "spans": [ + { + "bbox": [ + 320, + 158, + 394, + 172 + ], + "type": "text", + "content": "Juyong Zhang" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 326, + 173, + 476, + 185 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 326, + 173, + 476, + 185 + ], + "spans": [ + { + "bbox": [ + 326, + 173, + 476, + 185 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 326, + 173, + 476, + 185 + ], + "type": "text", + "content": "The Hong Kong Polytechnic University" + } + ] + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 58, + 217, + 207, + 384 + ], + "blocks": [ + { + "bbox": [ + 58, + 217, + 207, + 384 + ], + "lines": [ + { + "bbox": [ + 58, + 217, + 207, + 384 + ], + "spans": [ + { + "bbox": [ + 58, + 217, + 207, + 384 + ], + "type": "image", + "image_path": "c16c23a4569ecbb3eeaed8bf07a5e9408122317482dbf5661749dc23a15510a8.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 104, + 389, + 161, + 398 + ], + "lines": [ + { + "bbox": [ + 104, + 389, + 161, + 398 + ], + "spans": [ + { + "bbox": [ + 104, + 389, + 161, + 398 + ], + "type": "text", + "content": "One-Shot Photos" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "lines": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "spans": [ + { + "bbox": [ + 55, + 406, + 555, + 430 + ], + "type": "text", + "content": "Figure 1. Given a one-shot image (e.g., your favorite photo) as input, our method reconstructs a fully expressive whole-body talking avatar that captures personalized details and supports realistic animation, including vivid body gestures and natural expression changes." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 211, + 215, + 555, + 384 + ], + "blocks": [ + { + "bbox": [ + 211, + 215, + 555, + 384 + ], + "lines": [ + { + "bbox": [ + 211, + 215, + 555, + 384 + ], + "spans": [ + { + "bbox": [ + 211, + 215, + 555, + 384 + ], + "type": "image", + "image_path": "64c18b8bcd57950a232f1518eae3b0bb4624793e605782833db13b035c89dac8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 303, + 388, + 433, + 399 + ], + "lines": [ + { + "bbox": [ + 303, + 388, + 433, + 399 + ], + "spans": [ + { + "bbox": [ + 303, + 388, + 433, + 399 + ], + "type": "text", + "content": "Animatable Expressive Talking Avatars" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 152, + 443, + 200, + 455 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 152, + 443, + 200, + 455 + ], + "spans": [ + { + "bbox": [ + 152, + 443, + 200, + 455 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 54, + 470, + 297, + 687 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 470, + 297, + 687 + ], + "spans": [ + { + "bbox": [ + 54, + 470, + 297, + 687 + ], + "type": "text", + "content": "Building realistic and animatable avatars still requires minutes of multi-view or monocular self-rotating videos, and most methods lack precise control over gestures and expressions. To push this boundary, we address the challenge of constructing a whole-body talking avatar from a single image. We propose a novel pipeline that tackles two critical issues: 1) complex dynamic modeling and 2) generalization to novel gestures and expressions. To achieve seamless generalization, we leverage recent pose-guided image-to-video diffusion models to generate imperfect video frames as pseudo-labels. To overcome the dynamic modeling challenge posed by inconsistent and noisy pseudo-frames, we introduce a tightly coupled 3DGS-mesh hybrid avatar representation and apply several key regularizations to mitigate inconsistencies caused by imperfect labels. Extensive experiments on diverse subjects demonstrate that our method enables the creation of a photorealistic, precisely animatable, and expressive whole-body" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "spans": [ + { + "bbox": [ + 313, + 445, + 555, + 468 + ], + "type": "text", + "content": "talking avatar from just a single image. Project page: https://ustc3dv.github.io/OneShotOneTalk/." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 314, + 490, + 395, + 502 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 490, + 395, + 502 + ], + "spans": [ + { + "bbox": [ + 314, + 490, + 395, + 502 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "spans": [ + { + "bbox": [ + 312, + 510, + 555, + 665 + ], + "type": "text", + "content": "Realistic rendering and precise control of gestures and expressions for whole-body talking avatars hold significant potential for AR/VR applications, such as telepresence and immersive remote conferencing. While extensively studied in the fields of 3D Vision and Graphics, most production-level body avatars still require a light stage [11, 65] to capture hours or minutes of multi-view frames, making them unsuitable for typical consumer usage. To address this, we focus on the challenging task of constructing an animatable whole-body avatar that supports both photo-realistic rendering and precise 3D control of gestures and expressions, from only a single image. This one-shot pipeline primarily faces the following challenges." + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 556, + 715 + ], + "type": "text", + "content": "Complex dynamic modeling. Humans exhibit complex gestures and facial movements when communicating with each other. To model the motion space of the entire body, SMPL-X [46] integrates several previous mod" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 702, + 180, + 713 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 702, + 180, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 702, + 180, + 713 + ], + "type": "text", + "content": "*Corresponding author: Yudong Guo." + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10398" + } + ] + } + ], + "index": 23 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 286 + ], + "type": "text", + "content": "els [37, 40, 54] to capture body, hand, and facial movements. With the help of these parametric geometric models, recent personalized human avatars learn to model dynamic geometry and appearance by integrating neural textures [13, 38], hybrid representation [5, 17, 27], neural radiance fields [29, 47, 49, 64, 79], and 3D Gaussians [20, 22, 33, 35, 51] for dynamic photo-realistic rendering. To capture the full appearance of a whole-body avatar, these methods require complete observations, typically relying on dense input data for supervision, such as multi-view videos or self-rotating monocular videos. Moreover, these approaches primarily focus on body motion and novel view synthesis, falling short in accurately capturing and animating diverse facial expressions and hand movements, which limits their practical applicability. Due to the reliance on dense inputs, realistic and expressive whole-body modeling from a single image remains an unsolved challenge in this field." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 292, + 295, + 614 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 292, + 295, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 292, + 295, + 614 + ], + "type": "text", + "content": "Generalization to novel gestures and expressions. Another key challenge lies in the limited ability to generalize to a wide variety of gestures and facial expressions, especially those that deviate significantly from the poses seen during training. This limitation arises from the data-driven nature of current methods, where the training data typically encompasses only a finite set of gestures and expressions. Consequently, when animating avatars, regions or motions that were underrepresented in the training data—such as dynamic clothing, inner mouth movements, or intricate hand gestures—are often poorly synthesized or omitted entirely. These issues are further compounded when working with a single image input, as the limited visual cues often fail to capture the full range of motion and expression. Recently, several approaches [19, 59, 62, 63, 71, 77, 83] have leveraged image and video diffusion models [3, 53] to achieve image-to-video generation guided by whole-body landmarks [73]. Although various model architectures and training techniques have been explored to improve generation quality, temporal consistency remains a challenge, and these methods have several key limitations. For example, body landmarks fail to disentangle identity and pose, leading to poor cross-identity animation results, with facial and body distortions and identity mismatches. Moreover, the sparse 2D landmarks are insufficient for precisely controlling gestures and facial movements, as they do not accurately capture the underlying dynamic geometry." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 294, + 713 + ], + "type": "text", + "content": "In this work, we propose a novel pipeline to address both challenges in the one-shot setting. To generalize to diverse gestures and facial movements, we preprocess the large-scale TED Gesture Dataset [74] to build a comprehensive whole-body motion space for people talking. We then use these motion sequences as guidance to drive the input single image with a pre-trained whole-body video diffusion model [77] and a 3D face animation model [7]. This" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 228 + ], + "type": "text", + "content": "enables us to generate various video sequences of the target person performing different gestures and expressions. However, directly using these pseudo-frames to train a body avatar leads to unsatisfactory results. First, as noted above, current diffusion-based body animation methods struggle with temporal consistency, identity preservation, and motion alignment. Directly employing them as supervision results in significant blurring, distortion, and identity degradation. Furthermore, as we adapt two individual approaches for generating pseudo-frames for body gestures and facial expressions, respectively. Due to inconsistencies in camera spaces and rendering procedures, merging them into a unified avatar representation poses additional challenges." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 228, + 553, + 358 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 358 + ], + "type": "text", + "content": "To tackle the challenge of complex dynamic modeling, we utilize both the single input image and imperfect pseudoframes to train a hybrid mesh-3DGS avatar representation, constrained by several carefully designed regularizations. The single input image provides an accurate, though incomplete, appearance for the body avatar, while the pseudo videos offer imperfect but more complete visual cues. For the pseudo video labels, instead of using per-pixel losses, we employ a perceptual-based loss term [76], which helps achieve reasonable appearance modeling while alleviating misalignment in the pseudo-frames." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 360, + 553, + 477 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 360, + 553, + 477 + ], + "spans": [ + { + "bbox": [ + 313, + 360, + 553, + 477 + ], + "type": "text", + "content": "To further alleviate the inconsistencies caused by the pseudo-frames, we adopt a tightly coupled mesh-3DGS hybrid avatar representation. By introducing Laplacian smoothing and normal-consistency regularization on the deformed body mesh, we ensure that the structure of the 3D Gaussians used for rendering is well-constrained. Finally, we supervise the avatar representation using both gesture and head pseudo-frames, enabling the creation of a photorealistic, precisely animatable, and expressive whole-body avatar." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "spans": [ + { + "bbox": [ + 313, + 479, + 553, + 502 + ], + "type": "text", + "content": "In summary, our contributions include the following aspects:" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 314, + 503, + 553, + 634 + ], + "type": "list", + "angle": 0, + "index": 10, + "blocks": [ + { + "bbox": [ + 314, + 503, + 553, + 538 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 503, + 553, + 538 + ], + "spans": [ + { + "bbox": [ + 314, + 503, + 553, + 538 + ], + "type": "text", + "content": "- We introduce a novel pipeline that overcomes the key challenges of building a whole-body expressive talking avatar from a single image." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 314, + 539, + 553, + 586 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 539, + 553, + 586 + ], + "spans": [ + { + "bbox": [ + 314, + 539, + 553, + 586 + ], + "type": "text", + "content": "- Our diffusion guidance strategy effectively extracts valuable knowledge from imperfect diffusion outputs and combines it with the limited information from the input image, enabling complete modeling of the talking avatar." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 314, + 586, + 553, + 634 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 586, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 314, + 586, + 553, + 634 + ], + "type": "text", + "content": "- Our carefully designed 3DGS-mesh coupled avatar representation, along with essential regularization techniques, facilitates accurate modeling of diverse subjects and stabilizes the optimization process." + } + ] + } + ], + "index": 9 + } + ], + "sub_type": "text" + }, + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "spans": [ + { + "bbox": [ + 313, + 645, + 400, + 658 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 665, + 553, + 713 + ], + "type": "text", + "content": "Human Gaussian Splatting. 3D Gaussian Splatting [31] is the state-of-the-art method for scene reconstruction and novel view synthesis (NVS), offering superior rendering speed and visual quality. This has significantly influ" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10399" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 312 + ], + "type": "text", + "content": "enced human avatar studies. Methods using multi-view videos [30, 43, 45, 51, 85] demonstrate excellent performance, with unique designs such as ASH [45], which achieves efficient Gaussian learning via mesh UV parameterization. For monocular video input, most human Gaussian splatting methods [18, 20, 22, 35, 42, 57] link Gaussian fields to parametric mesh models, often using additional regularization terms. ExAvatar [42] applies connectivity-based regularizers to short, casually captured videos. For sparse-view images, GPS-Gaussian [81] achieves real-time human NVs by encoding human priors into the network, while HumanSplat [44] generates high-quality static reconstructions from one-shot inputs. Concurrent works [52, 61] introduce multi-view diffusion models but are limited to the head or body areas. IDOL [84] leverages a large-scale dataset and shows great performance in single-view human reconstruction. However, it still struggles with handling facial dynamic. By contrast, our method is capable of recovering a realistic, animatable whole-body talking avatar from just a single image." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 316, + 295, + 578 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 316, + 295, + 578 + ], + "spans": [ + { + "bbox": [ + 56, + 316, + 295, + 578 + ], + "type": "text", + "content": "Avatar Reconstruction from Few-Shot Images. Some works [8, 28, 67] apply 3D GAN inversion for one-shot human reconstruction, but they struggle with preserving personal details and generalization. PIFu [55] and subsequent works [6, 16, 56, 68, 69, 82] introduce pixel-aligned features and neural fields for image-based human reconstruction. An alternative approach leverages diffusion priors to fill in missing details, such as training human-centered diffusion models [2, 15, 44], using novel-view diffusion results for additional supervision [36, 39, 78], and employing Score Distillation Sampling (SDS) [50] to generate 3D avatars from 2D priors [24, 70, 72, 75]. However, these methods focus on static scenes and overlook dynamic human motion, limiting their ability to capture human dynamic. The SMPL-based [40] representation enables ELICIT [23] and SHERF [21] to model human dynamic, but overlooks hand and facial motions, restricting expressive animation capabilities. In contrast, our approach leverages priors from a pose-guided human video diffusion model, capturing both human appearance and dynamic, and enabling expressive full-body animations, particularly in the hands and face." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 295, + 714 + ], + "type": "text", + "content": "Pose-Guided Human Video Diffusion. Pose-guided human video diffusion models [9, 19, 59, 63, 71, 77, 83] directly generate animated videos from a reference image and pose sequence, bypassing traditional 3D reconstruction and rendering processes. The success of these models depends on the quality of training data, model design, and pose guidance. Some approaches [9, 19, 71, 83] incorporate temporal layers to ensure smooth transitions, inspired byAnimateDiff [12], while others [48, 59, 77] use video diffusion models [3, 4] for dynamic sequences. Pose guidance is typically provided by OpenPose [19, 63, 77], DensePose [71]," + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "depth maps [9], or SMPL [83]. StableAnimator [62] proposes HJB-based face optimization that further enhances face quality, and Make-Your-Anchors [25] personalizes outputs by fine-tuning models on identity-specific images. We adopt MimicMotion [77] for its superior performance in handling hand regions. Despite their strengths, these 2D models still face challenges, such as image distortion, identity changes, and pose misalignment, due to the lack of 3D understanding. To address these, we leverage optimized avatar representations with carefully designed constraints, improving consistency and naturalness in the generated animations." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 227, + 370, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 227, + 370, + 239 + ], + "spans": [ + { + "bbox": [ + 313, + 227, + 370, + 239 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 247, + 555, + 390 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 390 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 390 + ], + "type": "text", + "content": "Given a single image of the target person, we aim to reconstruct a 3D talking avatar that fully inherits the identity and enables natural animation. To address the challenge of complex dynamic modeling from imperfect pseudo-frames, we adopt a tightly coupled 3DGS-mesh hybrid avatar representation (Sec. 3.1). To generalize well to diverse gestures and facial movements, we generate imperfect sequences of the target person driven by various motion sequences (Sec. 3.2). Finally, we introduce the carefully designed constraints and loss terms to train the representation from noisy videos effectively (Sec. 3.3). The entire pipeline is illustrated in Fig. 2." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 397, + 471, + 410 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 471, + 410 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 471, + 410 + ], + "type": "text", + "content": "3.1. Coupled 3DGS-Mesh Avatar" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 415, + 553, + 509 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 415, + 553, + 509 + ], + "spans": [ + { + "bbox": [ + 313, + 415, + 553, + 509 + ], + "type": "text", + "content": "Whole-body parametric mesh models facilitate human animation and provide good initialization, while 3DGS offers enhanced expressiveness and realistic rendering. To address the challenges of the one-shot task, we design a novel coupled 3DGS-mesh representation, which effectively integrates the geometric priors and surface regularization of the mesh without diminishing the expressive capability of the Gaussian field." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 553, + 533 + ], + "type": "text", + "content": "We couple the 3DGS field with the typical SMPL-X model, which is formulated as follows:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 348, + 543, + 553, + 556 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 543, + 553, + 556 + ], + "spans": [ + { + "bbox": [ + 348, + 543, + 553, + 556 + ], + "type": "interline_equation", + "content": "M (\\beta , \\theta , \\phi) = W (T (\\beta , \\theta , \\phi), J (\\beta), \\theta , \\mathcal {W}), \\tag {1}", + "image_path": "2dc4585a07dc83b1c126f9010367421a690cc888d0f3611d85db4f9ca8fd7ccd.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 341, + 575, + 553, + 589 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 341, + 575, + 553, + 589 + ], + "spans": [ + { + "bbox": [ + 341, + 575, + 553, + 589 + ], + "type": "interline_equation", + "content": "T (\\beta , \\theta , \\phi) = \\bar {T} + B _ {S} (\\beta) + B _ {E} (\\phi) + B _ {P} (\\theta), \\tag {2}", + "image_path": "77c14a4c1890bcc57a36039273751a52be644f7f81f57f40dc005a61e4ef1e18.jpg" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "B_{S}(\\cdot), B_{E}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "B_{P}(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " denote shape, expression, and pose blend functions respectively, with " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta, \\phi" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " representing the corresponding parameters. " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\bar{T}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " is the template mesh, and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "W(\\cdot)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " is the standard LBS function that rotates the vertices in " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " around the estimated joints " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "J" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ", smoothed by the blend weights " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathcal{W}" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": ". Since the shape code " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\beta" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " is fixed once registered, we will omit it in the later statements and denote " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "T = T(\\beta, \\theta, \\phi)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "inline_equation", + "content": "J = J(\\beta)" + }, + { + "bbox": [ + 313, + 594, + 555, + 713 + ], + "type": "text", + "content": " for simplicity. Inspired by [1, 45, 66], we initialize the 3D Gaussians on the canonical mesh surface using UV parameterization. For a" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10400" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 73, + 550, + 289 + ], + "blocks": [ + { + "bbox": [ + 59, + 73, + 550, + 289 + ], + "lines": [ + { + "bbox": [ + 59, + 73, + 550, + 289 + ], + "spans": [ + { + "bbox": [ + 59, + 73, + 550, + 289 + ], + "type": "image", + "image_path": "136c8b08ef624783e8463e48c14d08b62a627a88e3190fa1058e033d6fd3d924.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "lines": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "spans": [ + { + "bbox": [ + 54, + 297, + 555, + 344 + ], + "type": "text", + "content": "Figure 2. Overview. Our method constructs an expressive whole-body talking avatar from a single image. We begin by generating pseudo body and head frames using pre-trained generative models, driven by a collected video dataset with diverse poses. Per-pixel supervision on the input image, perceptual supervision on imperfect pseudo-frames, and mesh-related constraints are then applied to guide the 3DGS-mesh coupled avatar representation, ensuring realistic and expressive avatar reconstruction and animation." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "spans": [ + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "text", + "content": "3D Gaussian located on a triangle " + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "inline_equation", + "content": "k = \\{\\mathbf{V}_1, \\mathbf{V}_2, \\mathbf{V}_3\\}" + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "text", + "content": " of mesh " + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "text", + "content": " with barycentric coordinates " + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "inline_equation", + "content": "(u, v)" + }, + { + "bbox": [ + 55, + 363, + 296, + 400 + ], + "type": "text", + "content": ", the surface position is given by:" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 407, + 295, + 421 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 407, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 61, + 407, + 295, + 421 + ], + "type": "interline_equation", + "content": "\\operatorname {B a r y} (T) = \\mathcal {V} (k, u, v) = u \\mathbf {V} _ {\\mathbf {1}} + v \\mathbf {V} _ {\\mathbf {2}} + (1 - u - v) \\mathbf {V} _ {\\mathbf {3}}. \\tag {3}", + "image_path": "de56aa592f8aa08c1407d5589e4242e3b9d7775e15254370dd4ac3fcd0f20c0b.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 429, + 296, + 656 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 429, + 296, + 656 + ], + "spans": [ + { + "bbox": [ + 55, + 429, + 296, + 656 + ], + "type": "text", + "content": "Coupled 3DGS-Mesh Deformation. To model clothes, haircuts, and other complex regions that the SMPL-X mesh fails to handle, the deformation of 3D Gaussians is crucial. While previous works can achieve correct deformation with sufficient supervision, we need to impose additional constraints on the deformation field to prevent overfitting to the single input image and imperfect pseudo-frames, while also facilitating the integration of other modules. Based on these considerations, we design two deformation fields in the canonical space and enforce their proximity: one represents the conventional Gaussian deformation, and the other represents the critical mesh deformation. This approach allows us to indirectly influence the Gaussian deformation by applying soft constraints on the deformed mesh. Moreover, since there is no strict binding between the deformed mesh surface and the Gaussians, complex regions are still effectively handled. The two deformation fields, together with the full-body animation and Gaussian Splitting process, are formulated as follows:" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 665, + 296, + 690 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 665, + 296, + 690 + ], + "spans": [ + { + "bbox": [ + 55, + 665, + 296, + 690 + ], + "type": "interline_equation", + "content": "G (\\theta , \\phi , \\mathbf {P}) = S P (W (B a r y (T) + d X, J, \\theta , B a r y (\\mathcal {W})), \\mathbf {P}). \\tag {4}", + "image_path": "d1281621758c73e87de8783ba03aa9eb5e5528a336195a6d4945e90b0a21251b.jpg" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 107, + 700, + 295, + 715 + ], + "type": "interline_equation", + "content": "M (\\theta , \\phi) = W (T + d T, J, \\theta , \\mathcal {W}), \\tag {5}", + "image_path": "4433e95575273ab325ac1eaa109b2f62faceaae9bcba0408f61dd3bf90c0cec6.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "spans": [ + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "G(\\cdot)" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " represents the final rendered image, " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "SP(\\cdot)" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " denote the splatting rendering process and the properties of 3D Gaussians, and " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "dX" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "dT" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " represent the optimized deformations of the Gaussians and mesh vertices, respectively. We apply key regularizers on " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "dT" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " and propagate soft constraints to " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "dX" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " by encouraging " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "dX" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": " to align with " + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "inline_equation", + "content": "Bary(dT)" + }, + { + "bbox": [ + 313, + 363, + 555, + 449 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 456, + 465, + 467 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 456, + 465, + 467 + ], + "spans": [ + { + "bbox": [ + 313, + 456, + 465, + 467 + ], + "type": "text", + "content": "3.2. Pseudo-Frames Generation" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 474, + 555, + 533 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 474, + 555, + 533 + ], + "spans": [ + { + "bbox": [ + 313, + 474, + 555, + 533 + ], + "type": "text", + "content": "In the one-shot image setting, many regions are unseen or occluded. To construct a complete avatar and ensure generalization to novel gestures and expressions, we turn our attention to recent advances in human motion diffusion models [19, 77] and head animation techniques [7, 10]." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "spans": [ + { + "bbox": [ + 313, + 534, + 556, + 665 + ], + "type": "text", + "content": "For leveraging diffusion-based generative models, the SDS loss [50] is widely used in text/image-to-3D works [24, 60, 72, 75]. Although diffusion models may introduce moderate deviations, the SDS technique remains effective in these studies. However, for tasks such as realistic dynamic avatar animation from a one-shot image, pose alignment is critical [18, 42], and the misalignment introduced by 2D diffusion models cannot be overlooked. Therefore, directly applying SDS in our setting is suboptimal. Instead, we focus on extracting accurate information from pseudo-frames synthesized by these generative models." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": "To generate whole-body pseudo-frames, we collect a set of SMPL-X pose sequences " + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "inline_equation", + "content": "\\{(\\theta_{N_i},\\phi_{N_i})_{i = 1}^F\\}_{j = 1}^K" + }, + { + "bbox": [ + 313, + 666, + 556, + 715 + ], + "type": "text", + "content": " from the TED Gesture Dataset [74] as input to the generative models. Since no unified generative model exists for both body and" + } + ] + } + ], + "index": 11 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10401" + } + ] + } + ], + "index": 12 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 75, + 173, + 177 + ], + "blocks": [ + { + "bbox": [ + 60, + 75, + 173, + 177 + ], + "lines": [ + { + "bbox": [ + 60, + 75, + 173, + 177 + ], + "spans": [ + { + "bbox": [ + 60, + 75, + 173, + 177 + ], + "type": "image", + "image_path": "d060dc60fdd93b5cfc853f8f1c7920ab535fa33c8a270767db6dad919d8e1634.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 69, + 185, + 104, + 282 + ], + "blocks": [ + { + "bbox": [ + 69, + 185, + 104, + 282 + ], + "lines": [ + { + "bbox": [ + 69, + 185, + 104, + 282 + ], + "spans": [ + { + "bbox": [ + 69, + 185, + 104, + 282 + ], + "type": "image", + "image_path": "fa99557ef1b3e52b7ecce646f40d38e31ed3bbae2a4c532fd8b26a74f394f931.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 118, + 180, + 171, + 284 + ], + "blocks": [ + { + "bbox": [ + 118, + 180, + 171, + 284 + ], + "lines": [ + { + "bbox": [ + 118, + 180, + 171, + 284 + ], + "spans": [ + { + "bbox": [ + 118, + 180, + 171, + 284 + ], + "type": "image", + "image_path": "dc5746401070b29c558c127a26c625f1906bad85c1790a4ed23eb00c1571576f.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 173, + 75, + 212, + 114 + ], + "blocks": [ + { + "bbox": [ + 173, + 75, + 212, + 114 + ], + "lines": [ + { + "bbox": [ + 173, + 75, + 212, + 114 + ], + "spans": [ + { + "bbox": [ + 173, + 75, + 212, + 114 + ], + "type": "image", + "image_path": "4a1241aa7d6933c48913f53f7cb4c94724184651c5bfad9471798eb075aabe0f.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 174, + 118, + 212, + 156 + ], + "blocks": [ + { + "bbox": [ + 174, + 118, + 212, + 156 + ], + "lines": [ + { + "bbox": [ + 174, + 118, + 212, + 156 + ], + "spans": [ + { + "bbox": [ + 174, + 118, + 212, + 156 + ], + "type": "image", + "image_path": "8c6d967c4b15c3c87b2c7bd4dec887c95f4fe9ae1d8e7d672a3c063f7ff2fab4.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 225, + 76, + 254, + 176 + ], + "blocks": [ + { + "bbox": [ + 225, + 76, + 254, + 176 + ], + "lines": [ + { + "bbox": [ + 225, + 76, + 254, + 176 + ], + "spans": [ + { + "bbox": [ + 225, + 76, + 254, + 176 + ], + "type": "image", + "image_path": "4eaab04a0012cae545c06daefa546ce43b242f18efe28f6376ce16d097c46287.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 214, + 181, + 264, + 284 + ], + "blocks": [ + { + "bbox": [ + 214, + 181, + 264, + 284 + ], + "lines": [ + { + "bbox": [ + 214, + 181, + 264, + 284 + ], + "spans": [ + { + "bbox": [ + 214, + 181, + 264, + 284 + ], + "type": "image", + "image_path": "f41e91d1723bab134428e064568faa9309a028d0082080b22aee1f20ee478750.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 268, + 76, + 363, + 177 + ], + "blocks": [ + { + "bbox": [ + 268, + 76, + 363, + 177 + ], + "lines": [ + { + "bbox": [ + 268, + 76, + 363, + 177 + ], + "spans": [ + { + "bbox": [ + 268, + 76, + 363, + 177 + ], + "type": "image", + "image_path": "77fa1ef8cfc5c11e6e54eb6fc8bc32501396ac5eab30731991ef1c3d09185311.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 269, + 180, + 307, + 219 + ], + "blocks": [ + { + "bbox": [ + 269, + 180, + 307, + 219 + ], + "lines": [ + { + "bbox": [ + 269, + 180, + 307, + 219 + ], + "spans": [ + { + "bbox": [ + 269, + 180, + 307, + 219 + ], + "type": "image", + "image_path": "6e2f399b3ae495ed6e4747d3cd740ab0bab5bb3442ecfc9200cf3c4e74f81d09.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 307, + 180, + 361, + 284 + ], + "blocks": [ + { + "bbox": [ + 307, + 180, + 361, + 284 + ], + "lines": [ + { + "bbox": [ + 307, + 180, + 361, + 284 + ], + "spans": [ + { + "bbox": [ + 307, + 180, + 361, + 284 + ], + "type": "image", + "image_path": "ce4303bb171cd4202d2c921a368427dcd6d997659aef6e71f430e22680c811fd.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 365, + 76, + 403, + 114 + ], + "blocks": [ + { + "bbox": [ + 365, + 76, + 403, + 114 + ], + "lines": [ + { + "bbox": [ + 365, + 76, + 403, + 114 + ], + "spans": [ + { + "bbox": [ + 365, + 76, + 403, + 114 + ], + "type": "image", + "image_path": "5134cb783d35dac4a79c5a83f194f78f74fcc0764c90609a7391fee3d9484042.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 365, + 118, + 403, + 156 + ], + "blocks": [ + { + "bbox": [ + 365, + 118, + 403, + 156 + ], + "lines": [ + { + "bbox": [ + 365, + 118, + 403, + 156 + ], + "spans": [ + { + "bbox": [ + 365, + 118, + 403, + 156 + ], + "type": "image", + "image_path": "37e59c5f3e0eb9c9a0f70969bcceb18348ded7f1fad84121cb415c055fdeaec2.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 415, + 76, + 448, + 176 + ], + "blocks": [ + { + "bbox": [ + 415, + 76, + 448, + 176 + ], + "lines": [ + { + "bbox": [ + 415, + 76, + 448, + 176 + ], + "spans": [ + { + "bbox": [ + 415, + 76, + 448, + 176 + ], + "type": "image", + "image_path": "3a383c78e917258b4da566b6ce601465c703c962949a9091542b5c64a7fa5af0.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 365, + 180, + 455, + 284 + ], + "blocks": [ + { + "bbox": [ + 365, + 180, + 455, + 284 + ], + "lines": [ + { + "bbox": [ + 365, + 180, + 455, + 284 + ], + "spans": [ + { + "bbox": [ + 365, + 180, + 455, + 284 + ], + "type": "image", + "image_path": "a636c897480b0534ce5c8b847029b1fb418e05e8d86ed0f6e3612f3d3ec48ee0.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 462, + 76, + 499, + 114 + ], + "blocks": [ + { + "bbox": [ + 462, + 76, + 499, + 114 + ], + "lines": [ + { + "bbox": [ + 462, + 76, + 499, + 114 + ], + "spans": [ + { + "bbox": [ + 462, + 76, + 499, + 114 + ], + "type": "image", + "image_path": "c4921b9a9dfd82c7dfe992d4b3ff6f4b05a326e356c2dd5e889a108b8b60e326.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 462, + 119, + 499, + 156 + ], + "blocks": [ + { + "bbox": [ + 462, + 119, + 499, + 156 + ], + "lines": [ + { + "bbox": [ + 462, + 119, + 499, + 156 + ], + "spans": [ + { + "bbox": [ + 462, + 119, + 499, + 156 + ], + "type": "image", + "image_path": "82e338918f226678a8764f87d87b32e758ec31a7b3bd019625da8cfa4a12697e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 510, + 76, + 541, + 175 + ], + "blocks": [ + { + "bbox": [ + 510, + 76, + 541, + 175 + ], + "lines": [ + { + "bbox": [ + 510, + 76, + 541, + 175 + ], + "spans": [ + { + "bbox": [ + 510, + 76, + 541, + 175 + ], + "type": "image", + "image_path": "b11488b81a7d183fc2d9156071a7680ebc12986c20a63bd0dc957fb56e8adb98.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 462, + 180, + 499, + 218 + ], + "blocks": [ + { + "bbox": [ + 462, + 180, + 499, + 218 + ], + "lines": [ + { + "bbox": [ + 462, + 180, + 499, + 218 + ], + "spans": [ + { + "bbox": [ + 462, + 180, + 499, + 218 + ], + "type": "image", + "image_path": "c8b9ccd8fd22c8e0d514f531e08c7c03fc013d004967758bfe96f6ea801b24a5.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 500, + 181, + 550, + 282 + ], + "blocks": [ + { + "bbox": [ + 500, + 181, + 550, + 282 + ], + "lines": [ + { + "bbox": [ + 500, + 181, + 550, + 282 + ], + "spans": [ + { + "bbox": [ + 500, + 181, + 550, + 282 + ], + "type": "image", + "image_path": "d97f9030bf1e648ce910525a839387844c6dbd2fb54d26d7b0d96d97ad2264ec.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 61, + 292, + 172, + 369 + ], + "blocks": [ + { + "bbox": [ + 61, + 292, + 172, + 369 + ], + "lines": [ + { + "bbox": [ + 61, + 292, + 172, + 369 + ], + "spans": [ + { + "bbox": [ + 61, + 292, + 172, + 369 + ], + "type": "image", + "image_path": "fbe806b5bc89ac407db8b293fe92098464e8941abf07cdfcbc1add316ec4eccb.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 376, + 110, + 386 + ], + "lines": [ + { + "bbox": [ + 66, + 376, + 110, + 386 + ], + "spans": [ + { + "bbox": [ + 66, + 376, + 110, + 386 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 119, + 376, + 171, + 386 + ], + "lines": [ + { + "bbox": [ + 119, + 376, + 171, + 386 + ], + "spans": [ + { + "bbox": [ + 119, + 376, + 171, + 386 + ], + "type": "text", + "content": "Driving Image" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "lines": [ + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 555, + 422 + ], + "type": "text", + "content": "Figure 3. Qualitative comparisons with representative methods [23, 42, 62] in the cross-identity motion reenactment task. Our method achieves accurate and realistic animation with almost all fine details preserved and identity unchanged." + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 182, + 292, + 261, + 369 + ], + "blocks": [ + { + "bbox": [ + 182, + 292, + 261, + 369 + ], + "lines": [ + { + "bbox": [ + 182, + 292, + 261, + 369 + ], + "spans": [ + { + "bbox": [ + 182, + 292, + 261, + 369 + ], + "type": "image", + "image_path": "a53440c716403b9ee9a262321371d9c588bc5e1d2ccb6262e79ad2a6b5cccab0.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 211, + 376, + 230, + 385 + ], + "lines": [ + { + "bbox": [ + 211, + 376, + 230, + 385 + ], + "spans": [ + { + "bbox": [ + 211, + 376, + 230, + 385 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 276, + 292, + 354, + 369 + ], + "blocks": [ + { + "bbox": [ + 276, + 292, + 354, + 369 + ], + "lines": [ + { + "bbox": [ + 276, + 292, + 354, + 369 + ], + "spans": [ + { + "bbox": [ + 276, + 292, + 354, + 369 + ], + "type": "image", + "image_path": "bcaa1bcd7e6b534197cbfe1f07683ad7a36b4ae8cf54cfec5812a81cf3fc1a63.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 282, + 376, + 339, + 385 + ], + "lines": [ + { + "bbox": [ + 282, + 376, + 339, + 385 + ], + "spans": [ + { + "bbox": [ + 282, + 376, + 339, + 385 + ], + "type": "text", + "content": "StableAnimator" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 374, + 292, + 454, + 369 + ], + "blocks": [ + { + "bbox": [ + 374, + 292, + 454, + 369 + ], + "lines": [ + { + "bbox": [ + 374, + 292, + 454, + 369 + ], + "spans": [ + { + "bbox": [ + 374, + 292, + 454, + 369 + ], + "type": "image", + "image_path": "b197a1cab344f42dcfaf217c3357ad5bc9c21306589dc0dacf50a081ad4f203a.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 386, + 376, + 441, + 385 + ], + "lines": [ + { + "bbox": [ + 386, + 376, + 441, + 385 + ], + "spans": [ + { + "bbox": [ + 386, + 376, + 441, + 385 + ], + "type": "text", + "content": "ExAvatar-1shot" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 471, + 292, + 549, + 369 + ], + "blocks": [ + { + "bbox": [ + 471, + 292, + 549, + 369 + ], + "lines": [ + { + "bbox": [ + 471, + 292, + 549, + 369 + ], + "spans": [ + { + "bbox": [ + 471, + 292, + 549, + 369 + ], + "type": "image", + "image_path": "ee77569f867464bf3251e8c18cf2febd89b461511a6f71b88f540cee2914c579.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 496, + 376, + 525, + 384 + ], + "lines": [ + { + "bbox": [ + 496, + 376, + 525, + 384 + ], + "spans": [ + { + "bbox": [ + 496, + 376, + 525, + 384 + ], + "type": "text", + "content": "ELICIT" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "spans": [ + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "text", + "content": "head simultaneously, we employ two separate approaches to generate pseudo-frames for body gestures and facial expressions, respectively. Given a source image " + }, + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "inline_equation", + "content": "I_{S}" + }, + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "text", + "content": " registered with the SMPL-X parameters " + }, + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "inline_equation", + "content": "(\\beta_{S},\\theta_{S},\\phi_{S})" + }, + { + "bbox": [ + 55, + 428, + 295, + 499 + ], + "type": "text", + "content": ", we adopt MimicMotion [77] with the collected pose sequences to generate various pseudo body frames:" + } + ] + } + ], + "index": 31 + }, + { + "bbox": [ + 100, + 505, + 295, + 519 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 100, + 505, + 295, + 519 + ], + "spans": [ + { + "bbox": [ + 100, + 505, + 295, + 519 + ], + "type": "interline_equation", + "content": "I _ {N} = \\operatorname {M o t i o n} \\left(I _ {S}, D \\left(\\beta_ {S}, \\theta_ {N}, \\phi_ {N}\\right)\\right), \\tag {6}", + "image_path": "4454ed35f33e3e95fec715b1b03d7fa52a2c7d5777f6de8f8aec2877494bd347.jpg" + } + ] + } + ], + "index": 32 + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "D(\\cdot)" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": " denotes the mapping from SMPL-X parameters to DWPose [73]. We randomly set the root body pose in " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\theta_N" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": " for each segment to increase viewpoint generalization, with pitch " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\in (-30^{\\circ}, 30^{\\circ})" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": " and yaw " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\in (-10^{\\circ}, 10^{\\circ})" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": ". After that, we re-track " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "I_N" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": " to obtain more accurate pose parameters " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "(\\hat{\\theta}_N, \\hat{\\phi}_N)" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": ", while keeping the shape code " + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "inline_equation", + "content": "\\beta_S" + }, + { + "bbox": [ + 55, + 522, + 295, + 666 + ], + "type": "text", + "content": " fixed. The re-tracking process is crucial, as it overcomes the misalignment introduced by 2D diffusion models and helps achieve precise and realistic results. For the head region, we adopt Portrait4D-v2 [7] to generate various pseudo frames of the target person performing diverse expressions, also driven by the videos from the TED Gesture Dataset." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 173, + 685 + ], + "type": "text", + "content": "3.3. Objective Functions" + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 714 + ], + "type": "text", + "content": "Assisted by our novel hybrid 3DGS-mesh avatar representation, we introduce several carefully designed regularization" + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 428, + 555, + 464 + ], + "type": "text", + "content": "terms to stabilize the avatar reconstruction process and effectively extract the correct information from both the one-shot input and the imperfect pseudo-frames." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 464, + 555, + 499 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 464, + 555, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 464, + 555, + 499 + ], + "type": "text", + "content": "Mesh-related Constraints. We use the following loss functions to apply soft constraints on the Gaussian field based on the mesh:" + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 372, + 506, + 495, + 533 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 372, + 506, + 495, + 533 + ], + "spans": [ + { + "bbox": [ + 372, + 506, + 495, + 533 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} \\mathcal {L} _ {S C} = \\lambda_ {\\text {n o r m a l}} \\mathcal {L} _ {\\text {n o r m a l}} + \\lambda_ {\\mathrm {M}} \\mathcal {L} _ {\\mathrm {M}} \\\\ + \\lambda_ {\\mathrm {M G C}} \\mathcal {L} _ {\\mathrm {M G C}} + \\lambda_ {\\mathrm {l a p}} \\mathcal {L} _ {\\mathrm {l a p}}. \\\\ \\end{array}", + "image_path": "f5ebe42d41ecb5c245fad4de69b77c244d0c94237ce4cbb7390cb3f66225527d.jpg" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 541, + 514, + 553, + 525 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 541, + 514, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 541, + 514, + 553, + 525 + ], + "type": "text", + "content": "(7)" + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "spans": [ + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "content": "Here, " + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{normal}}" + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "content": " is the normal consistency loss applied to the deformed mesh surface, ensuring surface normal consistency post-deformation. " + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_M" + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "content": " is the mask loss, which measures the discrepancy between the ground truth mask and the mask of the deformed mesh rendered via [34]. These two losses work together to regulate the behavior of " + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "inline_equation", + "content": "dT" + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "content": ", while influencing Gaussian deformation " + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "inline_equation", + "content": "dX" + }, + { + "bbox": [ + 313, + 540, + 555, + 636 + ], + "type": "text", + "content": " through the mesh-Gaussian consistency loss:" + } + ] + } + ], + "index": 40 + }, + { + "bbox": [ + 370, + 644, + 553, + 658 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 370, + 644, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 370, + 644, + 553, + 658 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\mathrm {M G C}} = \\| d X - B a r y (d T) \\| _ {1}. \\tag {8}", + "image_path": "4ce2b0fee2b2c510fb4e90db276f6612243e6405dfb993eb8b411b8d601695d9.jpg" + } + ] + } + ], + "index": 41 + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "spans": [ + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": "Additionally, we compute the Laplacian smoothing loss " + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_{\\mathrm{lap}}" + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "inline_equation", + "content": "dX" + }, + { + "bbox": [ + 313, + 666, + 555, + 712 + ], + "type": "text", + "content": " along with the scaling and RGBs of the 3D Gaussians, based on their initial connectivity state on the canonical mesh surface." + } + ] + } + ], + "index": 42 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10402" + } + ] + } + ], + "index": 43 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 288 + ], + "type": "text", + "content": "Perceptual Supervision of Pseudo-Frames. The synthesized pseudo-frames exhibit noticeable artifacts, such as image distortion and lack of 3D consistency, especially in the results generated with MimicMotion. These artifacts cannot be fully corrected through pose alignment alone. As a result, pixel-aligned losses like L1 or MSE often lead to issues like texture flickering, blurring, and identity changes. However, the high-level human structure is consistently well-preserved in the source image, which aids human recognition and is empirically referred to as perceptual similarity. Previous works [41, 76] have used deep neural networks to capture these perceptual structure features. We adopt the LPIPS perceptual loss [76], using VGG [58] as the backbone. By learning deep perceptual human features from dynamic poses and preserving detailed information from the source image, our method is able to conduct realistic and complete talking avatar animations. Thus, for the pseudoframes, we primarily use the following perceptual loss:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 294, + 295, + 308 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 294, + 295, + 308 + ], + "spans": [ + { + "bbox": [ + 91, + 294, + 295, + 308 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {\\text {d i f f}} = \\lambda_ {\\mathrm {L P I P S}} \\mathcal {L} _ {\\mathrm {L P I P S}} \\left(I _ {N}, G \\left(\\hat {\\theta} _ {N}, \\hat {\\phi} _ {N}, \\mathbf {P}\\right)\\right). \\tag {9}", + "image_path": "f04ae415b62196fa66a32e8cc115140d5858999061b0a3fd1d485c54c7ec653e.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "spans": [ + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": "Per-pixel Supervision of the Input Image. For the source image " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "I_S" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": " with pose " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "(\\theta_S, \\phi_S)" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": ", we use the regular L1 loss, SSIM loss, and mask loss between " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "I_S" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": " and the splatted image " + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "inline_equation", + "content": "G(\\theta_S, \\phi_S, \\mathbf{P})" + }, + { + "bbox": [ + 55, + 316, + 296, + 365 + ], + "type": "text", + "content": ":" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 73, + 371, + 295, + 384 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 371, + 295, + 384 + ], + "spans": [ + { + "bbox": [ + 73, + 371, + 295, + 384 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {S} = \\lambda_ {\\mathrm {R G B}} \\mathcal {L} _ {\\mathrm {R G B}} + \\lambda_ {\\mathrm {S S I M}} \\mathcal {L} _ {\\mathrm {S S I M}} + \\lambda_ {\\mathrm {a l p h a}} \\mathcal {L} _ {\\mathrm {a l p h a}}. \\tag {10}", + "image_path": "0457698990f60f0226a62070c1352ab8f71feb575795e001ce7d169c25d26600.jpg" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 391, + 296, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 391, + 296, + 415 + ], + "spans": [ + { + "bbox": [ + 55, + 391, + 296, + 415 + ], + "type": "text", + "content": "The final loss function is the sum of all the aforementioned losses." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 422, + 189, + 434 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 422, + 189, + 434 + ], + "spans": [ + { + "bbox": [ + 55, + 422, + 189, + 434 + ], + "type": "text", + "content": "3.4. Implementation Details" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "spans": [ + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": "We set the UV map resolution to 512 and use approximately 150,000 Gaussians in total. We train our models using the Adam optimizer [32] with " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\beta = (0.9, 0.999)" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ". The learning rates for the Gaussian parameters are the same as those in the official implementation, while the learning rates for both the Gaussian and mesh deformation fields are set to " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\eta = 1e^{-4}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ". For the loss weights, we set " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{normal}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{M}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{MGC}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{lap}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{LPIPS}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{RGB}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{SSIM}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "\\lambda_{\\mathrm{alpha}}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": " to " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "1e^{-2}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "1e^{-1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "1e^{1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "1e^{2}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "2e^{-1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "8e^{-1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "1e^{-1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "inline_equation", + "content": "4e^{-1}" + }, + { + "bbox": [ + 55, + 439, + 296, + 643 + ], + "type": "text", + "content": ", respectively. We start adding the perceptual loss at the 2000th step, while the other losses are used throughout the entire training process with a maximum number of training steps of 20000. We adopt joint training with separate supervision for body and head pseudo-frames. Training takes 3 hours on single 4090 due to critical but inefficient re-tracking. But avatar reconstruction and rendering are highly efficient with an inference speed of 50FPS+ at 1K resolution." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "spans": [ + { + "bbox": [ + 55, + 652, + 137, + 666 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 672, + 114, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 114, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 114, + 684 + ], + "type": "text", + "content": "4.1. Dataset" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 713 + ], + "type": "text", + "content": "We use the poses and expressions processed from 100 videos of the TED Gesture Dataset [74] as the motion" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 554, + 180 + ], + "type": "text", + "content": "sequences during training. We separately sampled about 1,800 frames with the most distinct variations in poses and expressions. For evaluation, the one-shot input and driving poses are primarily sourced from the ActorsHQ [26], NeuMan [29] and the Casual Conversations Dataset [14]. All videos and images are cropped and resized to a fixed aspect ratio of 9:16. For foreground segmentation, we use BiRefNet [80], and for pose tracking, we employ the custom fitting procedure from ExAvatar [42]." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 190, + 534, + 203 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 190, + 534, + 203 + ], + "spans": [ + { + "bbox": [ + 313, + 190, + 534, + 203 + ], + "type": "text", + "content": "4.2. Comparison with Representative Methods" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 208, + 554, + 315 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 208, + 554, + 315 + ], + "spans": [ + { + "bbox": [ + 313, + 208, + 554, + 315 + ], + "type": "text", + "content": "We compare our method with several representative works, including: (1) ExAvatar [42], a recent SOTA work that models human avatars with a mesh-based 3D Gaussian field. (2) ELICIT [23], a one-shot NeRF-based animatable avatar. (3) StableAnimator [62], a SOTA general 2D pose-guided human video diffusion model. (4) Make-Your-Anchors [25], a 2D diffusion-based method that is fine-tuned on a one-minute video clip of the individual to enhance identity information." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 316, + 554, + 399 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 316, + 554, + 399 + ], + "spans": [ + { + "bbox": [ + 313, + 316, + 554, + 399 + ], + "type": "text", + "content": "Note that for ExAvatar, since it takes short videos as input, we compare with two versions of it: ExAvatar-40shot, which uses 40-shot images, and ExAvatar-1shot, which uses one-shot images as input. For Make-Your-Ananchor, as we find it does not perform well on one-shot input, we only compare it with available video input by fine-tuning it on a short video clip." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 401, + 554, + 484 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 401, + 554, + 484 + ], + "spans": [ + { + "bbox": [ + 313, + 401, + 554, + 484 + ], + "type": "text", + "content": "Qualitative Comparisons. Fig. 3 and Fig. 4 present a qualitative comparison between our method and the other representative approaches. For Fig. 3, we use the pose of a different identity as the driving signal. For Fig. 4, we compare the performance of these methods on the self-driven pose reenactment task, using subjects that have corresponding video data." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 485, + 554, + 700 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 485, + 554, + 700 + ], + "spans": [ + { + "bbox": [ + 313, + 485, + 554, + 700 + ], + "type": "text", + "content": "As observed, ExAvatar-1shot tends to overfit the input image and fails to recover accurate textures in regions occluded by hands. Even with 40-shot input, ExAvatar still struggles with incomplete knowledge and fails to handle novel gestures effectively. StableAnimator generates relatively reasonable results but is constrained by its training distribution and struggles with identity consistency across frames, often leading to appearance mismatches and identity changes. ELICIT, which uses a NeRF-based representation, ensures 3D consistency but relies on SMPL for geometry, which neglects the hand region. This coarse semantic proxy fails to support complex hand reconstruction or animation. Make-Your-AnAnchor, although pre-trained on multiple identities to learn human motion priors, requires a long video with sufficient movement and appearance data to adapt to new identities. It struggles with short fine-tuning videos and fails to recover fine details, especially for gestures outside its fine-tuning distribution." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 325, + 701, + 554, + 714 + ], + "type": "text", + "content": "In contrast, our method generates expressive talking" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10403" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 71, + 154, + 173 + ], + "blocks": [ + { + "bbox": [ + 60, + 71, + 154, + 173 + ], + "lines": [ + { + "bbox": [ + 60, + 71, + 154, + 173 + ], + "spans": [ + { + "bbox": [ + 60, + 71, + 154, + 173 + ], + "type": "image", + "image_path": "b88b938fcf40d67c0b10fdeae381ff0e72dbf0628c6bb89c5b442f4b425570d9.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 156, + 74, + 214, + 173 + ], + "blocks": [ + { + "bbox": [ + 156, + 74, + 214, + 173 + ], + "lines": [ + { + "bbox": [ + 156, + 74, + 214, + 173 + ], + "spans": [ + { + "bbox": [ + 156, + 74, + 214, + 173 + ], + "type": "image", + "image_path": "461498c5fb7d9605afc9d9c8ca35e9599e577bf2b3d15ea52272df3c758b13d9.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 217, + 73, + 283, + 173 + ], + "blocks": [ + { + "bbox": [ + 217, + 73, + 283, + 173 + ], + "lines": [ + { + "bbox": [ + 217, + 73, + 283, + 173 + ], + "spans": [ + { + "bbox": [ + 217, + 73, + 283, + 173 + ], + "type": "image", + "image_path": "f21d406557de3ecffb7058f27f10929fa928621f5494ca1310d8a66cf21b5d99.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 286, + 74, + 346, + 173 + ], + "blocks": [ + { + "bbox": [ + 286, + 74, + 346, + 173 + ], + "lines": [ + { + "bbox": [ + 286, + 74, + 346, + 173 + ], + "spans": [ + { + "bbox": [ + 286, + 74, + 346, + 173 + ], + "type": "image", + "image_path": "28a8cf84ba456eb9ea039e3d45c9367be52a28b7c430b7b21ec0f4355264cb27.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 350, + 74, + 410, + 173 + ], + "blocks": [ + { + "bbox": [ + 350, + 74, + 410, + 173 + ], + "lines": [ + { + "bbox": [ + 350, + 74, + 410, + 173 + ], + "spans": [ + { + "bbox": [ + 350, + 74, + 410, + 173 + ], + "type": "image", + "image_path": "99e84388e162bfebe18a5f3d40ac5f8d784ba4c19d7f2840e6961196ed978559.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 415, + 74, + 475, + 173 + ], + "blocks": [ + { + "bbox": [ + 415, + 74, + 475, + 173 + ], + "lines": [ + { + "bbox": [ + 415, + 74, + 475, + 173 + ], + "spans": [ + { + "bbox": [ + 415, + 74, + 475, + 173 + ], + "type": "image", + "image_path": "603e0207c2597979f3e0b266b2a3d0835e8a6649914f651ae6c822328f06fdaa.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 478, + 73, + 550, + 173 + ], + "blocks": [ + { + "bbox": [ + 478, + 73, + 550, + 173 + ], + "lines": [ + { + "bbox": [ + 478, + 73, + 550, + 173 + ], + "spans": [ + { + "bbox": [ + 478, + 73, + 550, + 173 + ], + "type": "image", + "image_path": "f2b48159c359213127748e10c4429d218473e4b0f66919f5755d40609f04d78b.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 60, + 181, + 154, + 283 + ], + "blocks": [ + { + "bbox": [ + 60, + 181, + 154, + 283 + ], + "lines": [ + { + "bbox": [ + 60, + 181, + 154, + 283 + ], + "spans": [ + { + "bbox": [ + 60, + 181, + 154, + 283 + ], + "type": "image", + "image_path": "8d1925219951917a259deea15803b4159199f78824f3b9c167b154303fe1ca06.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 65, + 287, + 100, + 295 + ], + "lines": [ + { + "bbox": [ + 65, + 287, + 100, + 295 + ], + "spans": [ + { + "bbox": [ + 65, + 287, + 100, + 295 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 109, + 287, + 148, + 295 + ], + "lines": [ + { + "bbox": [ + 109, + 287, + 148, + 295 + ], + "spans": [ + { + "bbox": [ + 109, + 287, + 148, + 295 + ], + "type": "text", + "content": "Ground Truth" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 156, + 180, + 216, + 283 + ], + "blocks": [ + { + "bbox": [ + 156, + 180, + 216, + 283 + ], + "lines": [ + { + "bbox": [ + 156, + 180, + 216, + 283 + ], + "spans": [ + { + "bbox": [ + 156, + 180, + 216, + 283 + ], + "type": "image", + "image_path": "b80ac9d6ceecda1eb17aa851b2a3dbc00a354cf57015e734d516d82de5e2dd14.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 179, + 287, + 195, + 294 + ], + "lines": [ + { + "bbox": [ + 179, + 287, + 195, + 294 + ], + "spans": [ + { + "bbox": [ + 179, + 287, + 195, + 294 + ], + "type": "text", + "content": "Ours" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 217, + 180, + 284, + 283 + ], + "blocks": [ + { + "bbox": [ + 217, + 180, + 284, + 283 + ], + "lines": [ + { + "bbox": [ + 217, + 180, + 284, + 283 + ], + "spans": [ + { + "bbox": [ + 217, + 180, + 284, + 283 + ], + "type": "image", + "image_path": "1097f9c3c141bcda207f1e39e9040ba895e4cfa234f218fc346f4f7e5038130f.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 229, + 287, + 274, + 294 + ], + "lines": [ + { + "bbox": [ + 229, + 287, + 274, + 294 + ], + "spans": [ + { + "bbox": [ + 229, + 287, + 274, + 294 + ], + "type": "text", + "content": "StableAnimator" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_caption" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 286, + 179, + 347, + 283 + ], + "blocks": [ + { + "bbox": [ + 286, + 179, + 347, + 283 + ], + "lines": [ + { + "bbox": [ + 286, + 179, + 347, + 283 + ], + "spans": [ + { + "bbox": [ + 286, + 179, + 347, + 283 + ], + "type": "image", + "image_path": "a920911ef28c9476a6151cbaa5b50d642c681765b59fc218a48a08294cc93791.jpg" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 291, + 287, + 340, + 294 + ], + "lines": [ + { + "bbox": [ + 291, + 287, + 340, + 294 + ], + "spans": [ + { + "bbox": [ + 291, + 287, + 340, + 294 + ], + "type": "text", + "content": "ExAvatar-40shot" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 14 + }, + { + "type": "image", + "bbox": [ + 349, + 180, + 411, + 283 + ], + "blocks": [ + { + "bbox": [ + 349, + 180, + 411, + 283 + ], + "lines": [ + { + "bbox": [ + 349, + 180, + 411, + 283 + ], + "spans": [ + { + "bbox": [ + 349, + 180, + 411, + 283 + ], + "type": "image", + "image_path": "25adc51e724a59fe944e88ac02cb50868217549157043c4d43e3e3c1a000be86.jpg" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 359, + 287, + 403, + 294 + ], + "lines": [ + { + "bbox": [ + 359, + 287, + 403, + 294 + ], + "spans": [ + { + "bbox": [ + 359, + 287, + 403, + 294 + ], + "type": "text", + "content": "ExAvatar-1shot" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 16 + }, + { + "type": "image", + "bbox": [ + 415, + 180, + 477, + 283 + ], + "blocks": [ + { + "bbox": [ + 415, + 180, + 477, + 283 + ], + "lines": [ + { + "bbox": [ + 415, + 180, + 477, + 283 + ], + "spans": [ + { + "bbox": [ + 415, + 180, + 477, + 283 + ], + "type": "image", + "image_path": "0eb6d0a8e929f7e2c12a04ef6aa26fedc961025d4895aca9b6610f10e9db7284.jpg" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 433, + 287, + 456, + 294 + ], + "lines": [ + { + "bbox": [ + 433, + 287, + 456, + 294 + ], + "spans": [ + { + "bbox": [ + 433, + 287, + 456, + 294 + ], + "type": "text", + "content": "ELICIT" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_caption" + } + ], + "index": 18 + }, + { + "type": "image", + "bbox": [ + 479, + 180, + 550, + 283 + ], + "blocks": [ + { + "bbox": [ + 479, + 180, + 550, + 283 + ], + "lines": [ + { + "bbox": [ + 479, + 180, + 550, + 283 + ], + "spans": [ + { + "bbox": [ + 479, + 180, + 550, + 283 + ], + "type": "image", + "image_path": "b50d0db4bb81c52a06542b2c337530d79305a08481a0b3aff61bfd1a4021628d.jpg" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 485, + 287, + 542, + 294 + ], + "lines": [ + { + "bbox": [ + 485, + 287, + 542, + 294 + ], + "spans": [ + { + "bbox": [ + 485, + 287, + 542, + 294 + ], + "type": "text", + "content": "Make-Your-Anchors" + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + } + ], + "index": 20 + }, + { + "type": "table", + "bbox": [ + 58, + 345, + 293, + 408 + ], + "blocks": [ + { + "bbox": [ + 58, + 345, + 293, + 408 + ], + "lines": [ + { + "bbox": [ + 58, + 345, + 293, + 408 + ], + "spans": [ + { + "bbox": [ + 58, + 345, + 293, + 408 + ], + "type": "table", + "html": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
ELICIT5.65522.5590.9210.0660
ExAvatar3.93124.2150.9240.0409
StableAnimator1.88327.3500.9350.0327
Ours1.21629.3110.9430.0299
", + "image_path": "af943ac667e3478e64e3d11ac72c69662b8a0a13f62b04d0be13ebf7469e7200.jpg" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "table_body" + } + ], + "index": 23 + }, + { + "bbox": [ + 55, + 417, + 295, + 462 + ], + "lines": [ + { + "bbox": [ + 55, + 417, + 295, + 462 + ], + "spans": [ + { + "bbox": [ + 55, + 417, + 295, + 462 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons with representative methods [23, 42, 62] on self-driven data. Our method outperforms others in pixel-wise error metrics, realism evaluation metrics and perceptual quality metrics. (ExAvatar here denotes ExAvatar-40shot.)" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 55, + 472, + 296, + 555 + ], + "type": "text", + "content": "avatars from a single input image. Our results preserve fine human details and achieve natural animation with excellent rendering quality. We provide additional examples of cross-identity pose reenactment in Fig. 5. Using the same driving pose, identities with completely different attributes can be driven in the same way, thanks to the SMPL-X model and the 3DGS-mesh coupled representation." + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 558, + 296, + 713 + ], + "type": "text", + "content": "Quantitative Comparison. Tab. 1 presents a quantitative comparison between our model and other methods. Although the one-shot image animation task typically lacks a strict test set for quantitative comparison, we perform the evaluation on the self-driven task and use four common metrics for comparison: Mean Squared Error, PSNR, SSIM, and LPIPS. Our method outperforms all others across these metrics, demonstrating superior realism and 3D consistency in the results. However, we believe that these metrics do not fully capture the quality or capabilities of the methods, especially for the one-shot image animation task. We encourage readers to refer to the video results of our method for a more comprehensive and objective evaluation." + } + ] + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 318, + 346, + 373, + 408 + ], + "blocks": [ + { + "bbox": [ + 318, + 346, + 373, + 408 + ], + "lines": [ + { + "bbox": [ + 318, + 346, + 373, + 408 + ], + "spans": [ + { + "bbox": [ + 318, + 346, + 373, + 408 + ], + "type": "image", + "image_path": "c28295904109f88173963caecf54b2d211b0e4b1f0ecaa2b55b2912765d94f1e.jpg" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_body" + } + ], + "index": 27 + }, + { + "type": "image", + "bbox": [ + 318, + 410, + 373, + 473 + ], + "blocks": [ + { + "bbox": [ + 318, + 410, + 373, + 473 + ], + "lines": [ + { + "bbox": [ + 318, + 410, + 373, + 473 + ], + "spans": [ + { + "bbox": [ + 318, + 410, + 373, + 473 + ], + "type": "image", + "image_path": "25699cc9b99af0fbe2b2e70a26eb4a3e9654ffe701167d3ff592ec2f5d7ad951.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 328, + 474, + 364, + 482 + ], + "lines": [ + { + "bbox": [ + 328, + 474, + 364, + 482 + ], + "spans": [ + { + "bbox": [ + 328, + 474, + 364, + 482 + ], + "type": "text", + "content": "Driving Pose" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 376, + 346, + 419, + 408 + ], + "blocks": [ + { + "bbox": [ + 376, + 346, + 419, + 408 + ], + "lines": [ + { + "bbox": [ + 376, + 346, + 419, + 408 + ], + "spans": [ + { + "bbox": [ + 376, + 346, + 419, + 408 + ], + "type": "image", + "image_path": "8564196c94e1af9126a725ed2d3f0f212c97e3230d44b508a219b94a44b98a43.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + } + ], + "index": 30 + }, + { + "type": "image", + "bbox": [ + 376, + 410, + 419, + 472 + ], + "blocks": [ + { + "bbox": [ + 376, + 410, + 419, + 472 + ], + "lines": [ + { + "bbox": [ + 376, + 410, + 419, + 472 + ], + "spans": [ + { + "bbox": [ + 376, + 410, + 419, + 472 + ], + "type": "image", + "image_path": "6f275feae8d6b4fadb57e84f8bcd5e4048ba6e5deaf015a08c9642828bacc781.jpg" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_body" + } + ], + "index": 31 + }, + { + "type": "image", + "bbox": [ + 420, + 346, + 451, + 408 + ], + "blocks": [ + { + "bbox": [ + 420, + 346, + 451, + 408 + ], + "lines": [ + { + "bbox": [ + 420, + 346, + 451, + 408 + ], + "spans": [ + { + "bbox": [ + 420, + 346, + 451, + 408 + ], + "type": "image", + "image_path": "dab3b0faa9d80bd568e02481b13fc246038cf21c99950a2eae3501056e60a0db.jpg" + } + ] + } + ], + "index": 33, + "angle": 0, + "type": "image_body" + } + ], + "index": 33 + }, + { + "type": "image", + "bbox": [ + 420, + 410, + 449, + 473 + ], + "blocks": [ + { + "bbox": [ + 406, + 474, + 424, + 482 + ], + "lines": [ + { + "bbox": [ + 406, + 474, + 424, + 482 + ], + "spans": [ + { + "bbox": [ + 406, + 474, + 424, + 482 + ], + "type": "text", + "content": "Inputs" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 420, + 410, + 449, + 473 + ], + "lines": [ + { + "bbox": [ + 420, + 410, + 449, + 473 + ], + "spans": [ + { + "bbox": [ + 420, + 410, + 449, + 473 + ], + "type": "image", + "image_path": "10fbd9de7b0cd184fc6baedfa8a046b30f1d26c55fb37ea4c3f895859c673c1b.jpg" + } + ] + } + ], + "index": 34, + "angle": 0, + "type": "image_body" + } + ], + "index": 34 + }, + { + "type": "image", + "bbox": [ + 473, + 346, + 502, + 408 + ], + "blocks": [ + { + "bbox": [ + 473, + 346, + 502, + 408 + ], + "lines": [ + { + "bbox": [ + 473, + 346, + 502, + 408 + ], + "spans": [ + { + "bbox": [ + 473, + 346, + 502, + 408 + ], + "type": "image", + "image_path": "59412e687f8130757fd32dea8b2e7fd53b3b61ca0168eec2b808b4d5fb31e9cd.jpg" + } + ] + } + ], + "index": 35, + "angle": 0, + "type": "image_body" + } + ], + "index": 35 + }, + { + "type": "image", + "bbox": [ + 466, + 413, + 504, + 473 + ], + "blocks": [ + { + "bbox": [ + 466, + 413, + 504, + 473 + ], + "lines": [ + { + "bbox": [ + 466, + 413, + 504, + 473 + ], + "spans": [ + { + "bbox": [ + 466, + 413, + 504, + 473 + ], + "type": "image", + "image_path": "8cf9c90109f0a6f56607c552f22ed7dd3e80d0055fdad7bce1fedbfcd4003387.jpg" + } + ] + } + ], + "index": 36, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 473, + 474, + 493, + 481 + ], + "lines": [ + { + "bbox": [ + 473, + 474, + 493, + 481 + ], + "spans": [ + { + "bbox": [ + 473, + 474, + 493, + 481 + ], + "type": "text", + "content": "Result1" + } + ] + } + ], + "index": 37, + "angle": 0, + "type": "image_caption" + } + ], + "index": 36 + }, + { + "type": "image", + "bbox": [ + 517, + 346, + 547, + 408 + ], + "blocks": [ + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "lines": [ + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "spans": [ + { + "bbox": [ + 55, + 307, + 555, + 340 + ], + "type": "text", + "content": "Figure 4. Qualitative comparisons with representative methods [23, 25, 42, 62] in the self-driven motion reenactment task. Our method well models facial and hand regions, which matches the input image most in global identity preservation and local details modeling, even compared with some methods trained on captured videos." + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 517, + 346, + 547, + 408 + ], + "lines": [ + { + "bbox": [ + 517, + 346, + 547, + 408 + ], + "spans": [ + { + "bbox": [ + 517, + 346, + 547, + 408 + ], + "type": "image", + "image_path": "16bf92d9cffe3145086c9ad41e4595d922e8ad0cd2724b58aa7d042bfafb358e.jpg" + } + ] + } + ], + "index": 38, + "angle": 0, + "type": "image_body" + } + ], + "index": 38 + }, + { + "type": "image", + "bbox": [ + 515, + 413, + 549, + 473 + ], + "blocks": [ + { + "bbox": [ + 515, + 413, + 549, + 473 + ], + "lines": [ + { + "bbox": [ + 515, + 413, + 549, + 473 + ], + "spans": [ + { + "bbox": [ + 515, + 413, + 549, + 473 + ], + "type": "image", + "image_path": "87b15cc5dabf966138227fadf700527c0fc0db1b5ddc633b97ac0625b19a23e5.jpg" + } + ] + } + ], + "index": 39, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 523, + 474, + 544, + 481 + ], + "lines": [ + { + "bbox": [ + 523, + 474, + 544, + 481 + ], + "spans": [ + { + "bbox": [ + 523, + 474, + 544, + 481 + ], + "type": "text", + "content": "Result2" + } + ] + } + ], + "index": 40, + "angle": 0, + "type": "image_caption" + } + ], + "index": 39 + }, + { + "type": "table", + "bbox": [ + 316, + 521, + 552, + 596 + ], + "blocks": [ + { + "bbox": [ + 313, + 492, + 553, + 516 + ], + "lines": [ + { + "bbox": [ + 313, + 492, + 553, + 516 + ], + "spans": [ + { + "bbox": [ + 313, + 492, + 553, + 516 + ], + "type": "text", + "content": "Figure 5. More examples of cross-identity pose reenactment. Different subjects can be accurately animated with the same poses." + } + ] + } + ], + "index": 41, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 316, + 521, + 552, + 596 + ], + "lines": [ + { + "bbox": [ + 316, + 521, + 552, + 596 + ], + "spans": [ + { + "bbox": [ + 316, + 521, + 552, + 596 + ], + "type": "table", + "html": "
MetricsMSE(×10-3)↓PSNR↑SSIM↑LPIPS↓
w/o-mesh-SC1.5327.990.93800.0299
L1&SSIM-PFG1.4228.400.94160.0347
w/o-retrack1.5227.990.93820.0296
w/o-pseudo-face1.4628.240.93910.0293
Full-model1.2229.310.94310.0299
", + "image_path": "2c9acb02520e8db975ed1f0783911add374fc595005baab19e39fcdddb03d957.jpg" + } + ] + } + ], + "index": 42, + "angle": 0, + "type": "table_body" + } + ], + "index": 42 + }, + { + "bbox": [ + 347, + 604, + 520, + 614 + ], + "lines": [ + { + "bbox": [ + 347, + 604, + 520, + 614 + ], + "spans": [ + { + "bbox": [ + 347, + 604, + 520, + 614 + ], + "type": "text", + "content": "Table 2. Quantitative results of ablation studies." + } + ] + } + ], + "index": 43, + "angle": 0, + "type": "text" + }, + { + "bbox": [ + 313, + 624, + 415, + 635 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 624, + 415, + 635 + ], + "spans": [ + { + "bbox": [ + 313, + 624, + 415, + 635 + ], + "type": "text", + "content": "4.3. Ablation Studies" + } + ] + } + ], + "index": 44 + }, + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 555, + 713 + ], + "type": "text", + "content": "Perceptual-based Pseudo-Frames Guidance. Pseudoframes guidance is the core concept behind this work, enabling the challenging task of generating animatable and expressive talking avatars from a single input image. As shown in Fig. 6, without diffusion guidance, we are unable to inpaint unseen regions effectively, and the result suffers" + } + ] + } + ], + "index": 45 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10404" + } + ] + } + ], + "index": 46 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 60, + 73, + 127, + 205 + ], + "blocks": [ + { + "bbox": [ + 60, + 73, + 127, + 205 + ], + "lines": [ + { + "bbox": [ + 60, + 73, + 127, + 205 + ], + "spans": [ + { + "bbox": [ + 60, + 73, + 127, + 205 + ], + "type": "image", + "image_path": "71b77f527cbfc7ac8fc58e98448da6cc85c9c4f745f84965659f4f8b222045ae.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 72, + 206, + 114, + 216 + ], + "lines": [ + { + "bbox": [ + 72, + 206, + 114, + 216 + ], + "spans": [ + { + "bbox": [ + 72, + 206, + 114, + 216 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 127, + 74, + 190, + 202 + ], + "blocks": [ + { + "bbox": [ + 127, + 74, + 190, + 202 + ], + "lines": [ + { + "bbox": [ + 127, + 74, + 190, + 202 + ], + "spans": [ + { + "bbox": [ + 127, + 74, + 190, + 202 + ], + "type": "image", + "image_path": "ada814de8703553ec9d213fc8807f4f1224032926ff8c2ac37f52c62d3da9de2.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 143, + 206, + 187, + 216 + ], + "lines": [ + { + "bbox": [ + 143, + 206, + 187, + 216 + ], + "spans": [ + { + "bbox": [ + 143, + 206, + 187, + 216 + ], + "type": "text", + "content": "Driving Pose" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 198, + 72, + 290, + 204 + ], + "blocks": [ + { + "bbox": [ + 198, + 72, + 290, + 204 + ], + "lines": [ + { + "bbox": [ + 198, + 72, + 290, + 204 + ], + "spans": [ + { + "bbox": [ + 198, + 72, + 290, + 204 + ], + "type": "image", + "image_path": "feb678e3053189aac911b764ddd593fad38389147ed7a940c4107e59ded03cf2.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 228, + 206, + 265, + 215 + ], + "lines": [ + { + "bbox": [ + 228, + 206, + 265, + 215 + ], + "spans": [ + { + "bbox": [ + 228, + 206, + 265, + 215 + ], + "type": "text", + "content": "Full Model" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 298, + 72, + 390, + 204 + ], + "blocks": [ + { + "bbox": [ + 298, + 72, + 390, + 204 + ], + "lines": [ + { + "bbox": [ + 298, + 72, + 390, + 204 + ], + "spans": [ + { + "bbox": [ + 298, + 72, + 390, + 204 + ], + "type": "image", + "image_path": "639f29e821f3d0267212d23c053479a9625d2d0092f227147beb2c6ea32d806e.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 318, + 206, + 371, + 215 + ], + "lines": [ + { + "bbox": [ + 318, + 206, + 371, + 215 + ], + "spans": [ + { + "bbox": [ + 318, + 206, + 371, + 215 + ], + "type": "text", + "content": "L1&SSIM PFG" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_caption" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 398, + 76, + 467, + 204 + ], + "blocks": [ + { + "bbox": [ + 398, + 76, + 467, + 204 + ], + "lines": [ + { + "bbox": [ + 398, + 76, + 467, + 204 + ], + "spans": [ + { + "bbox": [ + 398, + 76, + 467, + 204 + ], + "type": "image", + "image_path": "58e5f24b793c74a898161b94e4999da9f35e1ab808a0e14fbc74a123fc436825.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 206, + 451, + 214 + ], + "lines": [ + { + "bbox": [ + 421, + 206, + 451, + 214 + ], + "spans": [ + { + "bbox": [ + 421, + 206, + 451, + 214 + ], + "type": "text", + "content": "w/o PFG" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_caption" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 486, + 76, + 548, + 204 + ], + "blocks": [ + { + "bbox": [ + 486, + 76, + 548, + 204 + ], + "lines": [ + { + "bbox": [ + 486, + 76, + 548, + 204 + ], + "spans": [ + { + "bbox": [ + 486, + 76, + 548, + 204 + ], + "type": "image", + "image_path": "a48351aba68138112958e91e6a898b1dfd2d9daed9f34873145c2daa9c5b977b.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 489, + 206, + 547, + 216 + ], + "lines": [ + { + "bbox": [ + 489, + 206, + 547, + 216 + ], + "spans": [ + { + "bbox": [ + 489, + 206, + 547, + 216 + ], + "type": "text", + "content": "w/o Re-Tracking" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 59, + 245, + 104, + 338 + ], + "blocks": [ + { + "bbox": [ + 95, + 228, + 513, + 240 + ], + "lines": [ + { + "bbox": [ + 95, + 228, + 513, + 240 + ], + "spans": [ + { + "bbox": [ + 95, + 228, + 513, + 240 + ], + "type": "text", + "content": "Figure 6. Visualization of ablation studies on perceptual-based pseudo-frames guidance (PFG) and re-tracking step." + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 59, + 245, + 104, + 338 + ], + "lines": [ + { + "bbox": [ + 59, + 245, + 104, + 338 + ], + "spans": [ + { + "bbox": [ + 59, + 245, + 104, + 338 + ], + "type": "image", + "image_path": "71f3ade00f3ed16e031235d599840d32f89cbc1e1be13cff8650b6d3fe4c92e8.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 66, + 341, + 97, + 349 + ], + "lines": [ + { + "bbox": [ + 66, + 341, + 97, + 349 + ], + "spans": [ + { + "bbox": [ + 66, + 341, + 97, + 349 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 109, + 245, + 168, + 338 + ], + "blocks": [ + { + "bbox": [ + 109, + 245, + 168, + 338 + ], + "lines": [ + { + "bbox": [ + 109, + 245, + 168, + 338 + ], + "spans": [ + { + "bbox": [ + 109, + 245, + 168, + 338 + ], + "type": "image", + "image_path": "370405f1fa9eb19c7ff0d136f340c106528b32491f89e6404a2b5b0a6bd0239e.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 123, + 341, + 151, + 348 + ], + "lines": [ + { + "bbox": [ + 123, + 341, + 151, + 348 + ], + "spans": [ + { + "bbox": [ + 123, + 341, + 151, + 348 + ], + "type": "text", + "content": "Full Model" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "type": "image", + "bbox": [ + 174, + 245, + 230, + 338 + ], + "blocks": [ + { + "bbox": [ + 174, + 245, + 230, + 338 + ], + "lines": [ + { + "bbox": [ + 174, + 245, + 230, + 338 + ], + "spans": [ + { + "bbox": [ + 174, + 245, + 230, + 338 + ], + "type": "image", + "image_path": "fcb35dfad717117a501e77ca41be6e1cbd88267cb4a72de4ecf7ac578f0a9ece.jpg" + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 178, + 341, + 220, + 348 + ], + "lines": [ + { + "bbox": [ + 178, + 341, + 220, + 348 + ], + "spans": [ + { + "bbox": [ + 178, + 341, + 220, + 348 + ], + "type": "text", + "content": "w/o Pseudo Face" + } + ] + } + ], + "index": 18, + "angle": 0, + "type": "image_caption" + } + ], + "index": 17 + }, + { + "type": "image", + "bbox": [ + 233, + 245, + 292, + 338 + ], + "blocks": [ + { + "bbox": [ + 233, + 245, + 292, + 338 + ], + "lines": [ + { + "bbox": [ + 233, + 245, + 292, + 338 + ], + "spans": [ + { + "bbox": [ + 233, + 245, + 292, + 338 + ], + "type": "image", + "image_path": "8b49a100e6a31db0220f63bfa0b56060a244a48fa5ed83dbb7863177fb592e97.jpg" + } + ] + } + ], + "index": 19, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 243, + 341, + 277, + 348 + ], + "lines": [ + { + "bbox": [ + 243, + 341, + 277, + 348 + ], + "spans": [ + { + "bbox": [ + 243, + 341, + 277, + 348 + ], + "type": "text", + "content": "w/o mesh SC" + } + ] + } + ], + "index": 20, + "angle": 0, + "type": "image_caption" + } + ], + "index": 19 + }, + { + "type": "image", + "bbox": [ + 58, + 382, + 108, + 468 + ], + "blocks": [ + { + "bbox": [ + 55, + 358, + 295, + 379 + ], + "lines": [ + { + "bbox": [ + 55, + 358, + 295, + 379 + ], + "spans": [ + { + "bbox": [ + 55, + 358, + 295, + 379 + ], + "type": "text", + "content": "Figure 7. Visualization of ablation studies on pseudo face supervision and mesh-related soft constraints." + } + ] + } + ], + "index": 21, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 58, + 382, + 108, + 468 + ], + "lines": [ + { + "bbox": [ + 58, + 382, + 108, + 468 + ], + "spans": [ + { + "bbox": [ + 58, + 382, + 108, + 468 + ], + "type": "image", + "image_path": "ac02986e65c94cd3a15d2445c3662e3c6cd1fbafeff3d224c5984264937fe32d.jpg" + } + ] + } + ], + "index": 22, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 64, + 471, + 102, + 479 + ], + "lines": [ + { + "bbox": [ + 64, + 471, + 102, + 479 + ], + "spans": [ + { + "bbox": [ + 64, + 471, + 102, + 479 + ], + "type": "text", + "content": "Input Image" + } + ] + } + ], + "index": 23, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 490, + 295, + 523 + ], + "lines": [ + { + "bbox": [ + 55, + 490, + 295, + 523 + ], + "spans": [ + { + "bbox": [ + 55, + 490, + 295, + 523 + ], + "type": "text", + "content": "Figure 8. Novel view synthesis. Our method demonstrates robust performance within a view range of -30 to 30 degrees. However, its performance degrades for larger viewing angles" + } + ] + } + ], + "index": 32, + "angle": 0, + "type": "image_caption" + } + ], + "index": 22 + }, + { + "type": "image", + "bbox": [ + 111, + 384, + 156, + 468 + ], + "blocks": [ + { + "bbox": [ + 111, + 384, + 156, + 468 + ], + "lines": [ + { + "bbox": [ + 111, + 384, + 156, + 468 + ], + "spans": [ + { + "bbox": [ + 111, + 384, + 156, + 468 + ], + "type": "image", + "image_path": "5654ae2390b62f8a82cd304d4677cd7dca27b03ad241d1b732b0e8dbc8d1bbf7.jpg" + } + ] + } + ], + "index": 24, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 124, + 470, + 136, + 478 + ], + "lines": [ + { + "bbox": [ + 124, + 470, + 136, + 478 + ], + "spans": [ + { + "bbox": [ + 124, + 470, + 136, + 478 + ], + "type": "inline_equation", + "content": "20^{\\circ}" + } + ] + } + ], + "index": 25, + "angle": 0, + "type": "image_caption" + } + ], + "index": 24 + }, + { + "type": "image", + "bbox": [ + 163, + 384, + 203, + 468 + ], + "blocks": [ + { + "bbox": [ + 163, + 384, + 203, + 468 + ], + "lines": [ + { + "bbox": [ + 163, + 384, + 203, + 468 + ], + "spans": [ + { + "bbox": [ + 163, + 384, + 203, + 468 + ], + "type": "image", + "image_path": "350ef94e58b722de0ca470dc007f096871fbc3bd569f76d5cdc2cd77d3c6ecbe.jpg" + } + ] + } + ], + "index": 26, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 173, + 470, + 185, + 478 + ], + "lines": [ + { + "bbox": [ + 173, + 470, + 185, + 478 + ], + "spans": [ + { + "bbox": [ + 173, + 470, + 185, + 478 + ], + "type": "inline_equation", + "content": "30^{\\circ}" + } + ] + } + ], + "index": 27, + "angle": 0, + "type": "image_caption" + } + ], + "index": 26 + }, + { + "type": "image", + "bbox": [ + 213, + 384, + 250, + 468 + ], + "blocks": [ + { + "bbox": [ + 213, + 384, + 250, + 468 + ], + "lines": [ + { + "bbox": [ + 213, + 384, + 250, + 468 + ], + "spans": [ + { + "bbox": [ + 213, + 384, + 250, + 468 + ], + "type": "image", + "image_path": "52c748fc6f6b9d9492f6039731a0b85fbf154817d21bcacd2b8b672517468b0a.jpg" + } + ] + } + ], + "index": 28, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 219, + 470, + 230, + 478 + ], + "lines": [ + { + "bbox": [ + 219, + 470, + 230, + 478 + ], + "spans": [ + { + "bbox": [ + 219, + 470, + 230, + 478 + ], + "type": "inline_equation", + "content": "45^{\\circ}" + } + ] + } + ], + "index": 29, + "angle": 0, + "type": "image_caption" + } + ], + "index": 28 + }, + { + "type": "image", + "bbox": [ + 262, + 384, + 294, + 468 + ], + "blocks": [ + { + "bbox": [ + 262, + 384, + 294, + 468 + ], + "lines": [ + { + "bbox": [ + 262, + 384, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 262, + 384, + 294, + 468 + ], + "type": "image", + "image_path": "0cf30b73d88edcde663adc98bd9328734d4fd761dd28854b6a5400009861cad7.jpg" + } + ] + } + ], + "index": 30, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 268, + 470, + 279, + 478 + ], + "lines": [ + { + "bbox": [ + 268, + 470, + 279, + 478 + ], + "spans": [ + { + "bbox": [ + 268, + 470, + 279, + 478 + ], + "type": "inline_equation", + "content": "60^{\\circ}" + } + ] + } + ], + "index": 31, + "angle": 0, + "type": "image_caption" + } + ], + "index": 30 + }, + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "spans": [ + { + "bbox": [ + 55, + 533, + 295, + 640 + ], + "type": "text", + "content": "from severe overfitting to the input image. Additionally, due to the misalignment and inconsistency of the pseudo-output generated by motion diffusion models, using pixel-based losses like L1 for diffusion guidance results in overly smooth outputs and struggles to capture fine details, especially in facial and hand regions. In contrast, the perceptual diffusion guidance we employ not only preserves the full personalized attributes of the input image but also inpaints unseen regions with more natural and consistent textures." + } + ] + } + ], + "index": 33 + }, + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 712 + ], + "type": "text", + "content": "Re-Tracking Step. The re-tracking procedure applied to the pseudo-output of motion diffusion models helps mitigate the misalignment introduced by 2D diffusion models, preventing texture errors and detail loss. As shown in Fig. 6, the re-tracking step successfully recovers more accurate texture structures." + } + ] + } + ], + "index": 34 + }, + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 553, + 330 + ], + "type": "text", + "content": "Mesh-related Constraints. Our designed hybrid 3DGS-mesh avatar representation and the corresponding regularizations significantly enhance the expressiveness and integrity of the final results, as demonstrated in Fig. 7. When soft constraints for mesh deformation are omitted, geometric artifacts appear in specific regions and the Gaussian field fails to capture fine details." + } + ] + } + ], + "index": 35 + }, + { + "bbox": [ + 313, + 331, + 553, + 392 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 331, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 313, + 331, + 553, + 392 + ], + "type": "text", + "content": "Pseudo Face Supervision. While perceptual-based guidance and hybrid 3DGS-mesh avatar representation enable pseudo human supervision to recover coarse head textures, incorporating targeted information from pseudo face frames further enhances facial quality, as presented in Fig. 7." + } + ] + } + ], + "index": 36 + }, + { + "bbox": [ + 313, + 392, + 553, + 416 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 392, + 553, + 416 + ], + "spans": [ + { + "bbox": [ + 313, + 392, + 553, + 416 + ], + "type": "text", + "content": "Besides the significant visual results shown in Figs. 6 and 7, quantitative evaluation is also conduct, as seen in Tab. 2." + } + ] + } + ], + "index": 37 + }, + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "spans": [ + { + "bbox": [ + 313, + 429, + 388, + 441 + ], + "type": "text", + "content": "5. Conclusion" + } + ] + } + ], + "index": 38 + }, + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 449, + 555, + 593 + ], + "type": "text", + "content": "In this paper, we introduce a novel pipeline for creating expressive talking avatars from a single image. We propose a coupled 3DGS-Mesh avatar representation, incorporating several key constraints and a carefully designed hybrid learning framework that combines information from both the input image and pseudo-frames. Experimental results demonstrate that our method outperforms existing techniques, with our one-shot avatar even surpassing state-of-the-art methods that require video input. Considering its simplicity in construction and ability to generate vivid animations, our method shows significant potential for practical applications of talking avatars across various fields." + } + ] + } + ], + "index": 39 + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": "Limitations. Severe mismatches between the input image and the registered parametric human mesh, especially in regions like fingers, can cause optimization issues and result in incorrect textures. Additionally, rendering large views or extending to full " + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "inline_equation", + "content": "360^{\\circ}" + }, + { + "bbox": [ + 313, + 594, + 556, + 713 + ], + "type": "text", + "content": " human reconstruction remains difficult due to current limitations in human motion diffusion models and the lack of data for large, novel viewpoints, as seen in Fig. 8. Future work will explore integrating semantic information from large language models and static priors from 3D reconstruction to address these limitations." + } + ] + } + ], + "index": 40 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10405" + } + ] + } + ], + "index": 41 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 150 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 150 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 150 + ], + "type": "text", + "content": "Acknowledgements. This work was supported by the National Natural Science Foundation of China (No. 62441224, No. 62272433, No. 62402468), The Hong Kong Polytechnic University (Project No. P0051019), the Research Center for Intelligent Operations Research of the PolyU and the advanced computing resources provided by the Supercomputing Center of the USTC." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 171, + 115, + 183 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 191, + 295, + 714 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 61, + 191, + 294, + 256 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 294, + 256 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 294, + 256 + ], + "type": "text", + "content": "[1] Rameen Abdal, Wang Yifan, Zifan Shi, Yinghao Xu, Ryan Po, Zhengfei Kuang, Qifeng Chen, Dit-Yan Yeung, and Gordon Wetzstein. Gaussian shell maps for efficient 3d human generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9441-9451, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 258, + 294, + 311 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 258, + 294, + 311 + ], + "spans": [ + { + "bbox": [ + 61, + 258, + 294, + 311 + ], + "type": "text", + "content": "[2] Badour AlBahar, Shunsuke Saito, Hung-Yu Tseng, Changil Kim, Johannes Kopf, and Jia-Bin Huang. Single-image 3d human digitization with shape-guided diffusion. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 313, + 294, + 369 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 313, + 294, + 369 + ], + "spans": [ + { + "bbox": [ + 62, + 313, + 294, + 369 + ], + "type": "text", + "content": "[3] Andreas Blattmann, Tim Dockhorn, Sumith Kulal, Daniel Mendelevitch, Maciej Kilian, Dominik Lorenz, Yam Levi, Zion English, Vikram Voleti, Adam Letts, et al. Stable video diffusion: Scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127, 2023. 2, 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "spans": [ + { + "bbox": [ + 62, + 369, + 294, + 423 + ], + "type": "text", + "content": "[4] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "spans": [ + { + "bbox": [ + 62, + 425, + 294, + 468 + ], + "type": "text", + "content": "[5] Xu Chen, Yufeng Zheng, Michael J Black, Otmar Hilliges, and Andreas Geiger. Snarf: Differentiable forward skinning for animating non-rigid neural implicit shapes. In International Conference on Computer Vision (ICCV), 2021. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 469, + 294, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 469, + 294, + 533 + ], + "spans": [ + { + "bbox": [ + 62, + 469, + 294, + 533 + ], + "type": "text", + "content": "[6] Enric Corona, Mihai Zanfir, Thiemo Alldieck, Eduard Gabriel Bazavan, Andrei Zanfir, and Cristian Sminchisescu. Structured 3d features for reconstructing controllable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 16954-16964, 2023. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 536, + 294, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 536, + 294, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 536, + 294, + 579 + ], + "type": "text", + "content": "[7] Yu Deng, Duomin Wang, and Baoyuan Wang. Portrait4d-v2: Pseudo multi-view data creates better 4d head synthesizer. In Proceedings of the European conference on computer vision (ECCV), 2024. 2, 4, 5" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 62, + 581, + 295, + 635 + ], + "type": "text", + "content": "[8] Zijian Dong, Xu Chen, Jinlong Yang, Michael J Black, Otmar Hilliges, and Andreas Geiger. Ag3d: Learning to generate 3d avatars from 2d image collections. In Proceedings of the IEEE/CVF international conference on computer vision, pages 14916-14927, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 62, + 636, + 294, + 689 + ], + "type": "text", + "content": "[9] Mengyang Feng, Jinlin Liu, Kai Yu, Yuan Yao, Zheng Hui, Xiefan Guo, Xianhui Lin, Haolan Xue, Chen Shi, Xiaowen Li, et al. Dreamoving: A human video generation framework based on diffusion models. arXiv e-prints, pages arXiv-2312, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 691, + 294, + 714 + ], + "type": "text", + "content": "[10] Jianzhu Guo, Dingyun Zhang, Xiaogiang Liu, Zhizhou Zhong, Yuan Zhang, Pengfei Wan, and Di Zhang. Livepor" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "trait: Efficient portrait animation with stitching and retargeting control. arXiv preprint arXiv:2407.03168, 2024. 4" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 160 + ], + "type": "text", + "content": "[11] Kaiwen Guo, Peter Lincoln, Philip Davidson, Jay Busch, Xueming Yu, Matt Whalen, Geoff Harvey, Sergio Orts-Escolano, Rohit Pandey, Jason Dourgarian, et al. The relightables: Volumetric performance capture of humans with realistic relighting. ACM Transactions on Graphics (ToG), 38(6):1–19, 2019. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 317, + 162, + 553, + 205 + ], + "type": "text", + "content": "[12] Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Imagine your personalized text-to-image diffusion models without specific tuning, 2023. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "spans": [ + { + "bbox": [ + 317, + 206, + 553, + 249 + ], + "type": "text", + "content": "[13] Marc Habermann, Lingjie Liu, Weipeng Xu, Michael Zollhoefer, Gerard Pons-Moll, and Christian Theobalt. Real-time deep dynamic characters. ACM Transactions on Graphics (ToG), 40(4):1-16, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "spans": [ + { + "bbox": [ + 317, + 251, + 553, + 304 + ], + "type": "text", + "content": "[14] Caner Hazirbas, Joanna Bitton, Brian Dolhansky, Jacqueline Pan, Albert Gordo, and Cristian Canton Ferrer. Towards measuring fairness in ai: the casual conversations dataset. IEEE Transactions on Biometrics, Behavior, and Identity Science, 4(3):324-332, 2021. 6" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 317, + 305, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 305, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 317, + 305, + 553, + 358 + ], + "type": "text", + "content": "[15] I Ho, Jie Song, Otmar Hilliges, et al. Sith: Single-view textured human reconstruction with image-conditioned diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 538-549, 2024. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 317, + 360, + 553, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 360, + 553, + 404 + ], + "spans": [ + { + "bbox": [ + 317, + 360, + 553, + 404 + ], + "type": "text", + "content": "[16] Yang Hong, Juyong Zhang, Boyi Jiang, Yudong Guo, Ligang Liu, and Hujun Bao. Stereopifu: Depth aware clothed human digitization via stereo vision. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "text", + "content": "[17] Chen Honghu, Peng Bo, Tao Yunfan, and Zhang Juyong. D" + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "inline_equation", + "content": "^3" + }, + { + "bbox": [ + 317, + 405, + 553, + 449 + ], + "type": "text", + "content": " human: Dynamic disentangled digital human from monocular video. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 317, + 449, + 553, + 491 + ], + "type": "text", + "content": "[18] Hezhen Hu, Zhiwen Fan, Tianhao Wu, Yihan Xi, Seoyoung Lee, Georgios Pavlakos, and Zhangyang Wang. Expressive gaussian human avatars from monocular rgb video. In NeurIPS, 2024. 3, 4" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "spans": [ + { + "bbox": [ + 317, + 493, + 555, + 536 + ], + "type": "text", + "content": "[19] Li Hu. Animate anyone: Consistent and controllable image-to-video synthesis for character animation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 8153-8163, 2024. 2, 3, 4" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "spans": [ + { + "bbox": [ + 317, + 537, + 553, + 602 + ], + "type": "text", + "content": "[20] Liangxiao Hu, Hongwen Zhang, Yuxiang Zhang, Boyao Zhou, Boning Liu, Shengping Zhang, and Liqiang Nie. Gaussianavatar: Towards realistic human avatar modeling from a single video via animatable 3d gaussians. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2, 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 604, + 553, + 657 + ], + "type": "text", + "content": "[21] Shoukang Hu, Fangzhou Hong, Liang Pan, Haiyi Mei, Lei Yang, and Ziwei Liu. Sherf: Generalizable human nerf from a single image. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 9352-9364, 2023. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 317, + 658, + 553, + 712 + ], + "type": "text", + "content": "[22] Shoukang Hu, Tao Hu, and Ziwei Liu. Gauhuman: Articulated gaussian splatting from monocular human videos. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20418-20431, 2024. 2, 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 318, + 757 + ], + "type": "text", + "content": "10406" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 127 + ], + "type": "text", + "content": "[23] Yangyi Huang, Hongwei Yi, Weiyang Liu, Haofan Wang, Boxi Wu, Wenxiao Wang, Binbin Lin, Debing Zhang, and Deng Cai. One-shot implicit animatable avatars with model-based priors. In IEEE Conference on Computer Vision (ICCV), 2023. 3, 5, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 183 + ], + "type": "text", + "content": "[24] Yangyi Huang, Hongwei Yi, Yuliang Xiu, Tingting Liao, Ji-axiang Tang, Deng Cai, and Justus Thies. Tech: Text-guided reconstruction of lifelike clothed humans. In 2024 International Conference on 3D Vision (3DV), pages 1531–1542. IEEE, 2024. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 183, + 294, + 239 + ], + "type": "text", + "content": "[25] Ziyao Huang, Fan Tang, Yong Zhang, Xiaodong Cun, Juan Cao, Jintao Li, and Tong-Yee Lee. Make-your-anchor: A diffusion-based 2d avatar generation framework. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6997-7006, 2024. 3, 6, 7" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[26] Mustafa Isik, Martin Rünz, Markos Georgopoulos, Taras Khakhulin, Jonathan Starck, Lourdes Agapito, and Matthias Nießner. Humanrf: High-fidelity neural radiance fields for humans in motion. ACM Transactions on Graphics (TOG), 42(4):1-12, 2023. 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 293, + 294, + 337 + ], + "type": "text", + "content": "[27] Boyi Jiang, Yang Hong, Hujun Bao, and Juyong Zhang. Self-con: Self reconstruction your digital avatar from monocular video. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 337, + 294, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 337, + 294, + 392 + ], + "spans": [ + { + "bbox": [ + 56, + 337, + 294, + 392 + ], + "type": "text", + "content": "[28] Suyi Jiang, Haoran Jiang, Ziyu Wang, Haimin Luo, Wenzheng Chen, and Lan Xu. Humangen: Generating human radiance fields with explicit priors. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12543-12554, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 393, + 294, + 437 + ], + "type": "text", + "content": "[29] Wei Jiang, Kwang Moo Yi, Golnoosh Samei, Oncel Tuzel, and Anurag Ranjan. Neuman: Neural human radiance field from a single video. In Proceedings of the European conference on computer vision (ECCV), 2022. 2, 6" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 502 + ], + "type": "text", + "content": "[30] Yuheng Jiang, Zhehao Shen, Penghao Wang, Zhuo Su, Yu Hong, Yingliang Zhang, Jingyi Yu, and Lan Xu. Hifi4g: High-fidelity human performance rendering via compact gaussian splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19734-19745, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 503, + 294, + 546 + ], + "type": "text", + "content": "[31] Bernhard Kerbl, Georgios Kopanas, Thomas Leimkuhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139-1, 2023. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 547, + 294, + 570 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 547, + 294, + 570 + ], + "spans": [ + { + "bbox": [ + 56, + 547, + 294, + 570 + ], + "type": "text", + "content": "[32] Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 570, + 294, + 613 + ], + "type": "text", + "content": "[33] Muhammed Kocabas, Jen-Hao Rick Chang, James Gabriel, Oncel Tuzel, and Anurag Ranjan. HUGS: Human gaussian splatting. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 658 + ], + "type": "text", + "content": "[34] Samuli Laine, Janne Hellsten, Tero Karras, Yeongho Seol, Jaakko Lehtinen, and Timo Aila. Modular primitives for high-performance differentiable rendering. ACM Transactions on Graphics, 39(6), 2020. 5" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[35] Jiahui Lei, Yufu Wang, Georgios Pavlakos, Lingjie Liu, and Kostas Daniilidis. Gart: Gaussian articulated template models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19876-19887, 2024. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 72, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "spans": [ + { + "bbox": [ + 316, + 72, + 553, + 127 + ], + "type": "text", + "content": "[36] Peng Li, Wangguandong Zheng, Yuan Liu, Tao Yu, Yangguang Li, Xingqun Qi, Mengfei Li, Xiaowei Chi, Siyu Xia, Wei Xue, et al. Pshuman: Photorealistic single-view human reconstruction using cross-scale diffusion. arXiv preprint arXiv:2409.10141, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 128, + 553, + 171 + ], + "type": "text", + "content": "[37] Tianye Li, Timo Bolkart, Michael J Black, Hao Li, and Javier Romero. Learning a model of facial shape and expression from 4d scans. ACM Transactions on Graphics (ToG), 36 (6):194-1, 2017. 2" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 172, + 553, + 225 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 172, + 553, + 225 + ], + "spans": [ + { + "bbox": [ + 316, + 172, + 553, + 225 + ], + "type": "text", + "content": "[38] Lingjie Liu, Marc Habermann, Viktor Rudnev, Kripasindhu Sarkar, Jiatao Gu, and Christian Theobalt. Neural actor: Neural free-view synthesis of human actors with pose control. ACM Transactions on Graphics (TOG), 40(6):1-16, 2021. 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 227, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 227, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 316, + 227, + 553, + 271 + ], + "type": "text", + "content": "[39] Zhibin Liu, Haoye Dong, Aviral Chharia, and Hefeng Wu. Human-vdm: Learning single-image 3d human gaussian splatting from video diffusion models. arXiv preprint arXiv:2409.02851, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "spans": [ + { + "bbox": [ + 316, + 271, + 553, + 315 + ], + "type": "text", + "content": "[40] Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. Smpl: a skinned multiperson linear model. ACM Transactions on Graphics (ToG), 34(6), 2015. 2, 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "spans": [ + { + "bbox": [ + 316, + 316, + 553, + 360 + ], + "type": "text", + "content": "[41] Roey Mechrez, Itamar Talmi, and Lihi Zelnik-Manor. The contextual loss for image transformation with non-aligned data. In Proceedings of the European conference on computer vision (ECCV), pages 768-783, 2018. 6" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 316, + 360, + 553, + 392 + ], + "type": "text", + "content": "[42] Gyeongsik Moon, Takaaki Shiratori, and Shunsuke Saito. Expressive whole-body 3D gaussian avatar. In ECCV, 2024. 3, 4, 5, 6, 7" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "spans": [ + { + "bbox": [ + 316, + 393, + 553, + 449 + ], + "type": "text", + "content": "[43] Arthur Moreau, Jifei Song, Helisa Dhamo, Richard Shaw, Yiren Zhou, and Eduardo Pérez-Pellitero. Human gaussian splatting: Real-time rendering of animatable avatars. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 788-798, 2024. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "spans": [ + { + "bbox": [ + 316, + 449, + 553, + 503 + ], + "type": "text", + "content": "[44] Panwang Pan, Zhuo Su, Chenguo Lin, Zhen Fan, Yongjie Zhang, Zeming Li, Tingting Shen, Yadong Mu, and Yebin Liu. Humansplat: Generalizable single-image human gaussian splattering with structure priors. arXiv preprint arXiv:2406.12459, 2024. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "spans": [ + { + "bbox": [ + 316, + 503, + 553, + 559 + ], + "type": "text", + "content": "[45] Haokai Pang, Heming Zhu, Adam Kortylewski, Christian Theobalt, and Marc Habermann. Ash: Animatable gaussian splats for efficient and photoreal human rendering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 1165-1175, 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 316, + 559, + 553, + 624 + ], + "type": "text", + "content": "[46] Georgios Pavlakos, Vasileios Choutas, Nima Ghorbani, Timo Bolkart, Ahmed A. A. Osman, Dimitrios Tzionas, and Michael J. Black. Expressive body capture: 3D hands, face, and body from a single image. In Proceedings IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), pages 10975-10985, 2019. 1" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "spans": [ + { + "bbox": [ + 316, + 624, + 553, + 669 + ], + "type": "text", + "content": "[47] Bo Peng, Jun Hu, Jingtao Zhou, Xuan Gao, and Juyong Zhang. Intrinsicngp: Intrinsic coordinate based hash encoding for human nerf. IEEE Transactions on Visualization and Computer Graphics, 2023. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[48] Bohao Peng, Jian Wang, Yuechen Zhang, Wenbo Li, MingChang Yang, and Jiaya Jia. Controlnext: Powerful and efficient control for image and video generation. arXiv preprint arXiv:2408.06070, 2024. 3" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10407" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 139 + ], + "type": "text", + "content": "[49] Sida Peng, Yuanqing Zhang, Yinghao Xu, Qianqian Wang, Qing Shuai, Hujun Bao, and Xiaowei Zhou. Neural body: Implicit neural representations with structured latent codes for novel view synthesis of dynamic humans. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9054-9063, 2021. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "spans": [ + { + "bbox": [ + 56, + 140, + 294, + 172 + ], + "type": "text", + "content": "[50] Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Mildenhall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "spans": [ + { + "bbox": [ + 56, + 173, + 294, + 227 + ], + "type": "text", + "content": "[51] Zhiyin Qian, Shaofei Wang, Marko Mihajlovic, Andreas Geiger, and Siyu Tang. 3dgs-athlete: Animatable avatars via deformable 3d gaussian splattering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024. 2, 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "spans": [ + { + "bbox": [ + 56, + 228, + 294, + 282 + ], + "type": "text", + "content": "[52] Lingteng Qiu, Shenhao Zhu, Qi Zuo, Xiaodong Gu, Yuan Dong, Junfei Zhang, Chao Xu, Zhe Li, Weihao Yuan, Liefeng Bo, Guanying Chen, and Zilong Dong. Anigs: Animatable gaussian avatar from a single image with inconsistent gaussian reconstruction, 2024. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "spans": [ + { + "bbox": [ + 56, + 283, + 294, + 337 + ], + "type": "text", + "content": "[53] Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684-10695, 2022. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "spans": [ + { + "bbox": [ + 56, + 338, + 294, + 381 + ], + "type": "text", + "content": "[54] Javier Romero, Dimitrios Tzionas, and Michael J. Black. Embodied hands: modeling and capturing hands and bodies together. ACM Transactions on Graphics (ToG), 36(6), 2017. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 383, + 294, + 437 + ], + "type": "text", + "content": "[55] Shunsuke Saito, Zeng Huang, Ryota Natsume, Shigeo Morishima, Angjoo Kanazawa, and Hao Li. Pifu: Pixel-aligned implicit function for high-resolution clothed human digitization. In Proceedings of the IEEE/CVF international conference on computer vision, pages 2304-2314, 2019. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 491 + ], + "type": "text", + "content": "[56] Shunsuke Saito, Tomas Simon, Jason Saragih, and Hanbyul Joo. Pifuhd: Multi-level pixel-aligned implicit function for high-resolution 3d human digitization. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 84-93, 2020. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 492, + 294, + 557 + ], + "type": "text", + "content": "[57] Zhijing Shao, Zhaolong Wang, Zhuang Li, Duotun Wang, Xiangru Lin, Yu Zhang, Mingming Fan, and Zeyu Wang. SplattingAvatar: Realistic Real-Time Human Avatars with Mesh-Embedded Gaussian Splatting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "spans": [ + { + "bbox": [ + 56, + 559, + 294, + 590 + ], + "type": "text", + "content": "[58] Karen Simonyan. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "spans": [ + { + "bbox": [ + 56, + 592, + 294, + 646 + ], + "type": "text", + "content": "[59] Shuai Tan, Biao Gong, Xiang Wang, Shiwei Zhang, Dandan Zheng, Ruobing Zheng, Kecheng Zheng, Jingdong Chen, and Ming Yang. Animate-x: Universal character image animation with enhanced motion representation. ICLR 2025, 2025. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 647, + 294, + 689 + ], + "type": "text", + "content": "[60] Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for efficient 3d content creation. arXiv preprint arXiv:2309.16653, 2023.4" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 294, + 712 + ], + "type": "text", + "content": "[61] Jiapeng Tang, Davide Davoli, Tobias Kirschstein, Liam Schoneveld, and Matthias Niessner. Gaf: Gaussian avatar" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 95 + ], + "type": "text", + "content": "reconstruction from monocular videos via multi-view diffusion, 2024. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 316, + 95, + 553, + 139 + ], + "type": "text", + "content": "[62] Shuyuan Tu, Zhen Xing, Xintong Han, Zhi-Qi Cheng, Qi Dai, Chong Luo, and Zuxuan Wu. Stableanimator: High-quality identity-preserving human image animation. arXiv preprint arXiv:2411.17697, 2024. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 140, + 553, + 194 + ], + "type": "text", + "content": "[63] Xiang Wang, Shiwei Zhang, Changxin Gao, Jiayu Wang, Xiaojiang Zhou, Yingya Zhang, Luxin Yan, and Nong Sang. Unianimate: Taming unified video diffusion models for consistent human image animation. arXiv preprint arXiv:2406.01188, 2024. 2, 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "spans": [ + { + "bbox": [ + 316, + 195, + 553, + 259 + ], + "type": "text", + "content": "[64] Chung-Yi Weng, Brian Curless, Pratul P Srinivasan, Jonathan T Barron, and Ira Kemelmacher-Shlizerman. Humanerf: Free-viewpoint rendering of moving people from monocular video. In Proceedings of the IEEE/CVF conference on computer vision and pattern Recognition, pages 16210-16220, 2022. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "spans": [ + { + "bbox": [ + 316, + 261, + 553, + 316 + ], + "type": "text", + "content": "[65] Donglai Xiang, Fabian Prada, Zhe Cao, Kaiwen Guo, Chenglei Wu, Jessica Hodgins, and Timur Bagautdinov. Drivable avatar clothing: Faithful full-body telepresence with dynamic clothing driven by sparse rgb-d input. In SIGGRAPH Asia 2023 Conference Papers, pages 1-11, 2023. 1" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 317, + 553, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 317, + 553, + 370 + ], + "spans": [ + { + "bbox": [ + 316, + 317, + 553, + 370 + ], + "type": "text", + "content": "[66] Jun Xiang, Xuan Gao, Yudong Guo, and Juyong Zhang. Flashavatar: High-fidelity head avatar with efficient gaussian embedding. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1802-1812, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "spans": [ + { + "bbox": [ + 316, + 372, + 553, + 435 + ], + "type": "text", + "content": "[67] Zhangyang Xiong, Di Kang, Derong Jin, Weikai Chen, Linchao Bao, Shuguang Cui, and Xiaoguang Han. Get3dhuman: Lifting stylegan-human into a 3d generative model using pixel-aligned reconstruction priors. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 316, + 437, + 553, + 491 + ], + "type": "text", + "content": "[68] Yuliang Xiu, Jinlong Yang, Dimitrios Tzionas, and Michael J Black. Icon: Implicit clothed humans obtained from normals. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 13286-13296. IEEE, 2022. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "spans": [ + { + "bbox": [ + 316, + 493, + 553, + 547 + ], + "type": "text", + "content": "[69] Yuliang Xiu, Jinlong Yang, Xu Cao, Dimitrios Tzionas, and Michael J Black. Econ: Explicit clothed humans optimized via normal integration. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 512-523, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "spans": [ + { + "bbox": [ + 316, + 548, + 553, + 591 + ], + "type": "text", + "content": "[70] Yuliang Xiu, Yufei Ye, Zhen Liu, Dimitrios Tzionas, and Michael J Black. Puzzle avatar: Assembling 3d avatars from personal albums. ACM Transactions on Graphics (TOG), 2024. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "spans": [ + { + "bbox": [ + 316, + 592, + 553, + 658 + ], + "type": "text", + "content": "[71] Zhongcong Xu, Jianfeng Zhang, Jun Hao Liew, Hanshu Yan, Jia-Wei Liu, Chenxu Zhang, Jiashi Feng, and Mike Zheng Shou. Magicanimate: Temporally consistent human image animation using diffusion model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1481–1490, 2024. 2, 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 712 + ], + "type": "text", + "content": "[72] Xihe Yang, Xingyu Chen, Daiheng Gao, Shaohui Wang, Xiaoguang Han, and Baoyuan Wang. Have-fun: Human avatar reconstruction from few-shot unconstrained images. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 742-752, 2024. 3, 4" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 748, + 317, + 757 + ], + "type": "text", + "content": "10408" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 712 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 116 + ], + "type": "text", + "content": "[73] Zhendong Yang, Ailing Zeng, Chun Yuan, and Yu Li. Effective whole-body pose estimation with two-stages distillation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 4210-4220, 2023. 2, 5" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "spans": [ + { + "bbox": [ + 56, + 118, + 294, + 182 + ], + "type": "text", + "content": "[74] Youngwoo Yoon, Woo-Ri Ko, Minsu Jang, Jaeyeon Lee, Jae-hong Kim, and Geehyuk Lee. Robots learn social skills: End-to-end learning of co-speech gesture generation for humanoid robots. In 2019 International Conference on Robotics and Automation (ICRA), pages 4303-4309. IEEE, 2019. 2, 4, 6" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 239 + ], + "type": "text", + "content": "[75] Jingbo Zhang, Xiaoyu Li, Qi Zhang, Yanpei Cao, Ying Shan, and Jing Liao. Humanref: Single image to 3d human generation via reference-guided diffusion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1844-1854, 2024. 3, 4" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 239, + 294, + 293 + ], + "type": "text", + "content": "[76] Richard Zhang, Phillip Isola, Alexei A Efros, Eli Shechtman, and Oliver Wang. The unreasonable effectiveness of deep features as a perceptual metric. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 586-595, 2018. 2, 6" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 348 + ], + "type": "text", + "content": "[77] Yuang Zhang, Jiaxi Gu, Li-Wen Wang, Han Wang, Junqi Cheng, Yuefeng Zhu, and Fangyuan Zou. Mimicmotion: High-quality human motion video generation with confidence-aware pose guidance. arXiv preprint arXiv:2406.19680, 2024. 2, 3, 4, 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 350, + 294, + 403 + ], + "type": "text", + "content": "[78] Zechuan Zhang, Zongxin Yang, and Yi Yang. Sifu: Side-view conditioned implicit function for real-world usable clothed human reconstruction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 9936-9947, 2024. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 405, + 294, + 459 + ], + "type": "text", + "content": "[79] Fuqiang Zhao, Wei Yang, Jiakai Zhang, Pei Lin, Yingliang Zhang, Jingyi Yu, and Lan Xu. Humanerf: Efficiently generated human radiance field from sparse inputs. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7743-7753, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 502 + ], + "type": "text", + "content": "[80] Peng Zheng, Dehong Gao, Deng-Ping Fan, Li Liu, Jorma Laaksonen, Wanli Ouyang, and Nicu Sebe. Bilateral reference for high-resolution dichotomous image segmentation. CAAI Artificial Intelligence Research, 3:9150038, 2024. 6" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 569 + ], + "type": "text", + "content": "[81] Shunyuan Zheng, Boyao Zhou, Ruizhi Shao, Boning Liu, Shengping Zhang, Liqiang Nie, and Yebin Liu. Gps-gaussian: Generalizable pixel-wise 3d gaussian splatting for real-time human novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 19680-19690, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "spans": [ + { + "bbox": [ + 56, + 571, + 294, + 613 + ], + "type": "text", + "content": "[82] Zerong Zheng, Tao Yu, Yebin Liu, and Qionghai Dai. Pamir: Parametric model-conditioned implicit representation for image-based human reconstruction. IEEE transactions on pattern analysis and machine intelligence, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 615, + 294, + 668 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 615, + 294, + 668 + ], + "spans": [ + { + "bbox": [ + 56, + 615, + 294, + 668 + ], + "type": "text", + "content": "[83] Shenhao Zhu, Junming Leo Chen, Zuozhuo Dai, Qingkun Su, Yinghui Xu, Xun Cao, Yao Yao, Hao Zhu, and Siyu Zhu. Champ: Controllable and consistent human image animation with 3d parametric guidance. arXiv preprint arXiv:2403.14781, 2024. 2, 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 670, + 294, + 712 + ], + "type": "text", + "content": "[84] Yiyu Zhuang, Jiaxi Lv, Hao Wen, Qing Shuai, Ailing Zeng, Hao Zhu, Shifeng Chen, Yujiu Yang, Xun Cao, and Wei Liu. idol: Instant photorealistic 3d human creation from a single image. arXiv preprint arXiv:2412.14963, 2024. 3" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 116 + ], + "type": "text", + "content": "[85] Wojciech Zielonka, Timur Bagautdinov, Shunsuke Saito, Michael Zollhöfer, Justus Thies, and Javier Romero. Drivable 3d gaussian avatars. arXiv preprint arXiv:2311.08581, 2023. 3" + } + ] + } + ], + "index": 13 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "spans": [ + { + "bbox": [ + 294, + 749, + 317, + 757 + ], + "type": "text", + "content": "10409" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_content_list.json b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..c8d1dbcf72f52c4e3886d8ff9f9a650d05f6a1e9 --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_content_list.json @@ -0,0 +1,1547 @@ +[ + { + "type": "text", + "text": "Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens", + "text_level": 1, + "bbox": [ + 125, + 128, + 872, + 174 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Suchisrit Gangopadhyay $^{1*}$ , Jung-Hee Kim $^{2*}$ , Xien Chen $^{1*}$ , Patrick Rim $^{1}$ , Hyoungseob Park $^{1}$ , Alex Wong $^{1}$ , Yale University $^{1}$ , Michigan State University $^{2}$", + "bbox": [ + 102, + 202, + 929, + 239 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{rit.gangopadhyay,xien.chen,patrick.rim,hyoungseob.park,alex.wong}@yale.edu, kimjun84@msu.edu", + "bbox": [ + 99, + 241, + 919, + 257 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 291, + 326, + 306 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: github.com/JungHeeKim29/calibration-token.", + "bbox": [ + 88, + 323, + 485, + 717 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 744, + 220, + 762 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Three-dimensional (3D) reconstruction is a fundamental component in many spatial applications, including autonomous vehicles, extended reality (XR), robotic manipulation. Each of these applications has unique demands for the field of view (FOV), often wider than the standard (perspective) camera. To meet this need, these applications tend to be deployed on systems equipped with fish", + "bbox": [ + 89, + 771, + 483, + 878 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/2282c912a8566b27ce86728225f5bf2e37e6035ecabe4966b19728f050f9f6f7.jpg", + "image_caption": [ + "Input", + "Figure 1. Foundational monocular depth estimators fail on fisheye images. Despite being trained on large-scale datasets, foundational monocular depth estimators (FMDEs) models produce erroneous outputs. The inaccurate, blurry estimates are caused by a covariate shift that stem from fisheye distortion." + ], + "image_footnote": [], + "bbox": [ + 516, + 290, + 643, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/804b813bf9fbcce9d675e2d349bdb96da517c4c862c4295f3cc474b8e2c3ef58.jpg", + "image_caption": [ + "MiDaS" + ], + "image_footnote": [], + "bbox": [ + 645, + 290, + 772, + 378 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/53b519b73404b54cfb261275430f67dcb413ef6b839dc536cb7098b940114650.jpg", + "image_caption": [ + "DepthAnything" + ], + "image_footnote": [], + "bbox": [ + 774, + 290, + 905, + 378 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "eye or other wide-angle cameras, which allows for wider coverage of the 3D environment. However, images captured by these cameras also come with substantial distortion, which arise from differences in projective geometry, where straight lines within the 3D environment or the 3D scene are preserved in perspective images but may appear curved in fisheye images.", + "bbox": [ + 511, + 477, + 906, + 582 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Foundational monocular depth estimators (FMDEs) [44, 48, 73] are trained on orders of tens of millions of images, enabling them to generalize across a wide range 3D scenes. However, their training data is comprised of internet images, which are predominantly captured using perspective cameras. Hence, despite being trained on large-scale datasets, FMDEs produce erroneous estimates when transferred to fisheye images (see Fig. 1). These errors stem from a covariate shift, which can be characterized by changes in camera calibration (intrinsic, distortion) parameters – leading to differences in object appearance and their perceived depth or distance from the camera.", + "bbox": [ + 511, + 583, + 908, + 763 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "To address fisheye distortion, one solution is to recalibrate and undistort images or perform a map projection to some canonical reference frame. In principle, if one has the correct calibration, it is possible to re-project a fisheye image into a perspective-like view (or vice versa). In practice, however, there are several problems: (1) The calibration process itself can be error-prone and sensitive to physical perturbations in the camera system. Minor bumps, focus changes, or lens replacements can degrade or invalid", + "bbox": [ + 511, + 763, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "*Equal contribution", + "bbox": [ + 109, + 887, + 217, + 898 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "5198", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "idate previously computed intrinsic parameters. (2) Even when re-projection is performed accurately, the transformation introduces latency and spatial artifacts (e.g., stretching, cropping, aliasing, loss). When used as a preprocessing step for existing pretrained depth estimators, these artifacts still present a covariate shift and can degrade performance.", + "bbox": [ + 89, + 90, + 480, + 181 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Another solution is to train a separate monocular depth estimator specifically for fisheye images. However, publicly available image datasets for fisheye cameras are orders of tens to hundreds of times smaller than those for perspective cameras. Hence, it is difficult to assemble sufficient data to reach the large-scale training requirement of an FMDE. Nonetheless, one can adapt or finetune existing FMDEs for fisheye imagery. While this can improve performance on fisheye images, it introduces the risk of parameter drift, where the resulting FMDEs may lose their generalizability across 3D scenes. Moreover, the resulting finetuned model becomes specialized to fisheye cameras, limiting its applicability to other camera types, which adds operational overhead in applications involving mixed camera systems, e.g., autonomous vehicles or robotics.", + "bbox": [ + 89, + 185, + 482, + 411 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose a novel approach termed Calibration Tokens. Our key insight is that existing FMDEs are already capable of estimating depth for perspective images, and that errors on fisheye images are caused by a covariate shift due to differences in camera calibration and distortion. Hence, rather than retraining or finetuning the entire model, we aim to \"recalibrate\" the fisheye latent embeddings such that they become more conducive to an FMDE originally trained on perspective images. Leveraging the fact that many FMDEs [44, 48, 49, 73] follow a Transformer-based architecture [11], we will exploit the (self- and cross-) attention mechanism to modulate the latent (token) embeddings by inserting Calibration Tokens as part of the input. Therefore, the existing FMDE will remain effectively unchanged, while Calibration Tokens serve to adapt their internal representations to mitigate the covariate shift by aligning the latent embeddings of fisheye images to the distribution of latent embeddings of perspective images. This design also allows us to preserve the original image content without performing any spatial re-projection, ensuring the process is lossless in terms of the raw pixels. Our hypothesis is that by adding a small set of trainable tokens to encode the fisheye camera calibration information and utilizing them to recalibrate the latent embeddings, we will be able to reuse existing FMDEs trained on perspective images and adapt them to fisheye images without sacrificing their generalizability across diverse 3D scenes.", + "bbox": [ + 89, + 414, + 482, + 821 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To train these Calibration Tokens, we propose a self-supervised objective that leverages inverse warping in the input and output spaces. FMDEs can infer high-fidelity depth maps for perspective images, so we use the perspective image depth estimates as our training target. We then", + "bbox": [ + 89, + 824, + 482, + 900 + ], + "page_idx": 1 + }, + { + "type": "image", + "img_path": "images/ad025756a0561445ab4028e80b266e274c2565dbefc8b3f28f87d3d6944e7ba9.jpg", + "image_caption": [ + "Figure 2. Inference on different cameras. Calibration Tokens enable foundational monocular depth estimators to adapt to fisheye images while maintaining performance on perspective images." + ], + "image_footnote": [], + "bbox": [ + 519, + 89, + 906, + 303 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "induce artificial distortion on perspective images to create pairs of perspective and synthetic fisheye images with diverse fisheye distortions. However, rather than doing the same in the output space, we undistort the fisheye depth maps to the original perspective reference frame to compute a self-supervised loss between the undistorted fisheye and perspective depth maps. By minimizing this self-supervised loss, the Calibration Tokens learn to align fisheye image embeddings to those of perspective images in the latent space, without any labels. Additionally, computing the loss in the original perspective frame allows our method to preserve the supervision signal instead of introducing artifacts.", + "bbox": [ + 511, + 371, + 903, + 551 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our approach allows us to bypass the need to compile large-scale fisheye datasets by exploiting the abundance of perspective image datasets. As our method operates in the reference frame of the input, we also avoid transformation artifacts at inference time, whether in the input or output space. Furthermore, our method preserves compatibility with perspective images: One simply needs to append or remove Calibration Tokens for FMDEs to be applied to fisheye or perspective images. We demonstrate our method on indoors and outdoors across several recent FMDEs and consistently improve over baselines.", + "bbox": [ + 511, + 551, + 905, + 717 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Our contributions: (1) We propose a novel approach to extend foundational monocular depth estimators (FMDEs) trained on perspective images to fisheye images. (2) We introduce Calibration Tokens that modulate the latent embeddings of fisheye images towards the distribution of perspective image embeddings. (3) We introduce a self-supervised training objective that recalibrates input perspective image to fisheye images, but \"undos\" the transformation in the output to enable loss computation on high-fidelity (perspective) depth maps inferred by FMDEs. (4) Our approach only requires training one set of tokens to achieve state-of-the-art performance for both indoors and outdoors.", + "bbox": [ + 511, + 719, + 905, + 900 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "5199", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Works", + "text_level": 1, + "bbox": [ + 91, + 89, + 240, + 104 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Monocular Depth Estimation can be trained in a supervised or unsupervised manner. Supervised methods [13, 14, 29-32, 37, 71, 77] minimize the difference between depth estimates and ground-truth depth maps. [18] re-formulated the problem as ordinal regression while other methods proposed architectures innovations. [2] partitions depth ranges into adaptive bins. [5] incorporates an attention-based up-sample block. [33] employs hierarchical aggregation and heterogeneous interaction modules. [80] uses neural window fully-connected CRFs to compute energy. [59] synthesizes perspectively accurate images to enrich training data. Additional inputs e.g., language [81-83], lidar [6, 15, 51, 74], radar [50, 54], are used to enable metric-scale depth estimates. Unsupervised methods [7, 8, 23, 38, 42, 47, 57, 62-67, 84] minimizes photometric reconstruction error. [19] frames depth estimation as a novel view synthesis problem. [20] introduces a left-right consistency loss. [88] uses a pose network to enable unsupervised training on video sequences. [21] introduced auto-masking and min-reprojection loss. Additional loss terms based on visual odometry [16, 60], iterative closest point [40], surface normals [75], trinocular assumption [46], and semantic segmentation [22, 28] were also introduced. [39] redesigned the skip connection and decoders to extract high-resolution features, [86] combined global and local representations and [85] introduced a lightweight architecture with dilated convolution and attention. AugUndo [68] leveraged invertibility of transformation groups for data augmentation.", + "bbox": [ + 93, + 114, + 480, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Foundational Monocular Depth Estimators are trained with supervised or semi-supervised learning on large-scale datasets. MiDaS [49] is the first to demonstrate generalizable monocular depth estimation by compiling datasets for large-scale training. DPT [48] extended the approach and introduced transformers for dense predictions. Marigold [26] repurposes diffusion models for monocular depth estimation. DepthAnything [73] proposes a pseudo-labeling method to curate a large-scale dataset. Additionally, UniDepth [44] employs a camera self-prompting module and a pseudo-spherical output space, enabling metric-scale depth prediction across diverse 3D scenes without relying on external camera parameters. DepthPro [3] proposes a multi-scale vision transformer for metric-scale depth estimation. As all of these FMDEs are trained on perspective images, they fail to generalize to fisheye cameras.", + "bbox": [ + 93, + 553, + 480, + 792 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Fisheye Images. Images taken by a fisheye camera are distorted and unsuitable for use in a perspective image encoder. Existing distortion correction algorithms [12, 25] rely on different camera projection models [27, 41, 55] to undistort images into a perspective view. However, these methods depend on camera calibration parameters, which can introduce artifacts due to calibration inaccuracies. Recent approaches", + "bbox": [ + 89, + 795, + 480, + 901 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "[24, 36] demonstrate training a separate model to perform depth estimation with different camera types. They utilize an equirectangular projection to project points from different reference frames to a canonical equirectangular frame, but this can incur transformation artifacts and distortions. Additionally, deep-learning-based methods [17, 34] that aim to rectify distortion have been introduced. However, these methods require a large number of parameters with limited accuracy and field of view. Consequently, many recent works targeted for fisheye images involve training an entire network [1, 79, 87] exclusively on fisheye images. Our method extends foundational monocular depth estimators to fisheye images instead.", + "bbox": [ + 511, + 90, + 903, + 286 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Token-Based Methods. Recent transformer-based architectures represent input images [11] (or other modalities [69, 70]) as sequences of tokens. In many cases, an additional token (e.g. [CLS] token in BERT [10] or the distillation token in DeiT [58]) is employed to aggregate information across all tokens. Such tokens can be adapted to various purposes, acting as a compact representation that \"binds\" or fuses information, e.g., [72] uses tokens to learn synthetic and real tactile response maps. Inspired by these advances, we introduce minimal trainable tokens appended to the fisheye embedding, enabling the model to \"bind\" or reconcile image distortions within a frozen backbone. Our approach is lightweight and requires no major architectural modifications, but extends foundational monocular depth estimators trained on perspective images to fisheye images.", + "bbox": [ + 511, + 287, + 903, + 513 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3. Method", + "text_level": 1, + "bbox": [ + 513, + 526, + 604, + 542 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Let $I:\\Omega \\mapsto \\mathbb{R}^3$ denote an RGB image obtained from a calibrated camera and $\\Omega \\subset \\mathbb{R}^2$ the image space. Monocular depth estimation aims to learn a parameterized function $h_{\\omega ,\\psi}:\\mathbb{R}^3\\to \\mathbb{R}_+$ that maps an image to a depth map $d:\\Omega \\mapsto \\mathbb{R}_{+}$ . We assume access to a foundational monocular depth estimator (FMDE) pretrained on some large-scale dataset of perspective images. We will pair each image with (pseudo)ground truth $\\tilde{d} = h_{\\omega ,\\psi}(I)$ inferred by the FMDE to obtain a training dataset $\\mathcal{D} = \\{(I^{(n)},\\tilde{d}^{(n)})\\}_{n = 1}^{N}$ .", + "bbox": [ + 511, + 551, + 903, + 688 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To extend FMDEs, trained on perspective images, to fisheye images, we introduce Calibration Tokens as an adaptation mechanism. Due to the prevalence of Transformer architectures in many FMDEs, we train a set of lightweight tokens to model the change in calibration between a perspective camera and different fisheye cameras. The goal of our Calibration Token is to recalibrate or translate latent embeddings of fisheye images back to those of perspective images. Our method takes advantage of the attention mechanism inherent in FMDEs and enable Calibration Tokens to modulate the latent embeddings, thus facilitating latent alignment. The outcome is an FMDE that is capable of inferring depth for fisheye images with Calibration Tokens, and perspective images without.", + "bbox": [ + 511, + 689, + 903, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "5200", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/5197d726d466f8831043af4a4243e6c876dc1adc5d75e3d73f923b858c7340f4.jpg", + "image_caption": [ + "Figure 3. Overview of our method. We introduce a set of trainable Calibration Tokens, which is appended to the input sequence of the fisheye image tokens. The Calibration Tokens are trained to adapt the model to produce accurate depth maps for images with various fisheye distortions. A unique fisheye calibration token is appended to the input of each new layer of the encoder." + ], + "image_footnote": [], + "bbox": [ + 94, + 93, + 903, + 340 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.1. Extending FMDEs with Calibration Tokens", + "text_level": 1, + "bbox": [ + 89, + 409, + 460, + 425 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Specifically, let $\\phi \\in \\mathbb{R}^{M\\times F}$ represent a set of Calibration Tokens, where $M$ denotes the number of tokens and $F$ their number of dimensions. For a given layer within the encoder $f_{\\omega}$ of an FMDE, we will concatenate Calibration Tokens to the input sequence of patches or (embeddings) of the vision transformer: $f_{\\omega}([I;\\phi]) = [z^{(L)};f_{\\omega}(\\phi)]$ , where $z$ denotes the latent embeddings recalibrated by $\\phi$ , $L$ the last layer, and $[\\cdot ;]$ the concatenation operation.", + "bbox": [ + 89, + 434, + 483, + 556 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "As each layer denotes a separate latent space, we extend our approach to a multi-layer modulation scheme. Let $\\Phi \\in \\mathbb{R}^{L\\times M\\times F}$ be the set of Calibration Tokens for each layer $l\\in \\{1,\\dots,L\\}$ . A unique set of Calibration Tokens $\\phi^{(l)}\\in \\mathbb{R}^{M\\times F}$ is appended at each encoder layer: $f_{\\omega}^{(l)}[(z^{(l - 1)};\\phi^{(l)})] = [z^{(l)};f_{\\omega}^{(l)}(\\phi^{(l)})]$ for a layer $l$ . Each set of Calibration Tokens will modulate the patch embeddings for a specific layer through the attention mechanism; hence, following the convention in existing works [4, 9], we discard Calibration Tokens from the encoder output. A key insight is that the FMDE is already able to estimate high-fidelity depth maps for perspective images. We posit that the covariate shift exists in the encodings of fisheye image. Hence, once the latent embeddings of fisheye images have been recalibrated to those of perspective images, the decoder will be able to estimate depth to similar fidelity as perspective images. Therefore, we do not utilize Calibration Tokens to modulate the decoder layers. The final estimate is obtained by $\\hat{d} = g_{\\psi}(z^{(L)})$ , where $g_{\\psi}$ denotes the decoder.", + "bbox": [ + 89, + 561, + 483, + 851 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Since our method does not apply spatial transformations during inference, it remains entirely lossless for input images. Additionally, it offers several efficiencies: (1) The", + "bbox": [ + 89, + 854, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "only trainable parameters in our method are the light-weight Calibration Tokens, which consist of significantly fewer parameters than vision transformer models. Our method introduces minimal computational overhead and results in lower time and space complexity than training or finetuning a full model. (2) Our approach is backward-compatible with perspective images. By omitting our Calibration Tokens, an FMDE maintains its original depth estimation performance on perspective images. (3) At inference, camera intrinsics are not required, as the training process allows generalization across various fisheye camera intrinsics. As a result, our method eliminates the need for the arduous and error-prone calibration process after training.", + "bbox": [ + 511, + 409, + 906, + 606 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "3.2. Learning Calibration Tokens", + "text_level": 1, + "bbox": [ + 511, + 614, + 772, + 630 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "To train Calibration Tokens, we will leverage the abundance of publicly available perspective image datasets. During our training, we synthesize fisheye images from perspective images by recalibrating them using artificial fisheye intrinsic and distortion parameters. This will produce pairs of perspective and fisheye images from which we can leverage self-supervision, and allows us to use a much larger training dataset than exclusively training with real fisheye images. We follow previous approaches [17, 78] to obtain synthetic fisheye images from the calibrated perspective images. Our synthesized fisheye images follow the distortion model introduced by Kannala & Brandt [25]:", + "bbox": [ + 511, + 636, + 906, + 818 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nr (\\theta) = k _ {1} \\theta + k _ {2} \\theta^ {3} + k _ {3} \\theta^ {5} + k _ {4} \\theta^ {7}, \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 589, + 827, + 903, + 845 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where $\\theta$ denotes the angle between the ray and the optical axis, and $\\{k_i\\}_{i = 1}^4$ are distortion coefficients that can represent most of the real world fisheye distortion models. The", + "bbox": [ + 511, + 854, + 906, + 900 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "5201", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "change in coordinate between $(x,y)$ in the perspective image and $(x^{\\prime},y^{\\prime})$ in the fisheye image can be formulated as", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} x ^ {\\prime} = r (\\theta) \\cos (\\varphi), y ^ {\\prime} = r (\\theta) \\sin (\\varphi), \\tag {2} \\\\ \\varphi = \\arctan ((y - c _ {x}) / (x - c _ {y})), \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 128, + 482, + 164 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $(c_x, c_y)$ is the principal point in the perspective image. We define the transformation from the perspective to the fisheye reference frame as $T$ and its inverse transformation as $T^{-1}$ . Our training dataset is composed of perspective images and synthesized distorted image pairs with corresponding forward and inverse transformations.", + "bbox": [ + 89, + 171, + 482, + 261 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Loss Function. Inspired by AugUndo [68] and their use of invertible transformations to preserve the supervision signal by undoing data augmentation, we propose to synthesize fisheye images from the abundance of perspective images as inputs, but undistort the output to facilitate loss computation. By remapping depth estimates of synthetic fisheye images to the perspective frame, we enable the use of high-fidelity estimates inferred by FMDEs on perspective images as supervision. Hence, we can optimize Calibration Tokens with the following self-supervised loss:", + "bbox": [ + 89, + 262, + 483, + 412 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\arg \\min _ {\\Psi} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {x \\in \\Omega} \\log \\left(\\left| \\tilde {d} ^ {(n)} (x) - T ^ {- 1} \\circ \\hat {d} ^ {(n)} (x) \\right| + 1\\right), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 98, + 419, + 482, + 460 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where $\\tilde{d} = h_{\\omega, \\psi}(I)$ and $\\hat{d} = h_{\\omega, \\psi}(T \\circ I; \\Phi)$ denotes the predicted depth map from the given perspective image and synthesized fisheye image, respectively. $\\Phi$ denotes the proposed trainable Calibration Token appended to the patch embeddings. Calibration Tokens are trained to minimize the difference between the perspective output and the fisheye output re-projected into the perspective reference frame. Eq. (3) follows the log of absolute differences ( $\\log L1$ ) proposed in [43], which enhances training stability and empirically outperforms L1 loss, especially in border regions where discrepancies between perspective and fisheye images are most significant (see Sec. 4.3 for details).", + "bbox": [ + 89, + 469, + 482, + 651 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "It is important to note that attempting to instead transform the perspective depth maps outputted by FMDEs to the fisheye reference frame for the loss computation introduces information loss in the training objective. This will lead to re-projection artifacts in the supervision and result in learning inaccuracies during training. In Section A of the Supp. Mat., we further demonstrate comparison results between training in fisheye image space and perspective image space. Our training scheme is self-supervised and requires only calibrated perspective images, which can be easily obtained, making our approach both scalable and practical.", + "bbox": [ + 89, + 652, + 482, + 819 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4. Experiments", + "text_level": 1, + "bbox": [ + 89, + 830, + 222, + 847 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Datasets. Training our Calibration Tokens requires only calibrated perspective images, enabling us to leverage significantly more data compared to training solely on fisheye", + "bbox": [ + 89, + 854, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "images. Moreover, since our loss is computed based on comparisons with perspective image outputs, ground truth is not required for our training pipeline.", + "bbox": [ + 511, + 90, + 903, + 136 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Training datasets: NYUv2 [53] has a variety of perspective indoor scenes; VOID [65] contains indoor perspective office, classroom and stairwell scenes; IRS [61] contains rendered perspective scenes of home, restaurant, and store settings; Hypersim [52] comprises photorealistic, synthetic images of indoor residential and commercial environments in a perspective reference frame. Waymo [56] dataset consists of a diverse set of urban driving scenes.", + "bbox": [ + 511, + 137, + 903, + 258 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Test datasets: ScanNet++ [76] offers 3D reconstructions of diverse indoor scenes, captured using laser scanning and DSLR imaging with a fisheye lens, which allows us to evaluate with real fisheye images and ground truth depth maps. KITTI-360 [35] includes suburban driving scenes captured with a multi-sensor setup, including fisheye cameras. Notably, it features a different field of view compared to ScanNet++ [76], allowing us to assess the generalization capability of the Calibration Tokens.", + "bbox": [ + 511, + 258, + 903, + 393 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Models. Calibration Tokens do not require specific settings and can be integrated into any model utilizing a vision transformer. We evaluate the effectiveness of our approach using by extending MiDaS [49], DepthAnything [73], and UniDepth [45] to fisheye images. Note that we used 8 tokens per layer for each of the model experiments.", + "bbox": [ + 511, + 396, + 903, + 486 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Evaluation Metrics. We evaluate depth prediction accuracy using standard metrics from monocular depth estimation of root mean squared error (RMSE) and $\\delta_{1}$ . Details on these metrics can be found in the Supp. Mat.", + "bbox": [ + 511, + 487, + 903, + 547 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Implementation Details. We trained our Calibration Tokens based on 3 different FMDEs (MiDas[49], DepthAnything[49], UniDepth[45]). We utilized the pretrained ViT-L backbone for MiDas [49] and DepthAnything [49], and the ViT-S backbone for UniDepth [45]. We trained our model on 4 NVIDIA 3090 GPUs for 40k iterations with a batch size of 16. For input, we used images in the resolution of $518 \\times 518$ . For testing, we used $462 \\times 616$ resolution on the ScanNet++ dataset [76], and $700 \\times 700$ on the KITTI-360 dataset [35] to preserve its aspect ratio. We also synthesize random fisheye distortions in the training images. Our Calibration Tokens are trained with a joint dataset consisting of indoor and outdoor datasets totaling up to only 200K samples, and as shown in Tab. 1, obtain comparable results on both domains with fewer samples than existing methods [24] that are trained specifically for each.", + "bbox": [ + 511, + 547, + 905, + 790 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.1. Main Result", + "text_level": 1, + "bbox": [ + 511, + 801, + 643, + 816 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We conduct experiments to analyze the impact of Calibration Tokens on model performance. As a baseline, we compare our model to DepthAnyCamera [24], the state-of-the-art monocular depth estimation (MDE) method for fisheye images. Here, we evaluate the DepthAnyCamera model us", + "bbox": [ + 511, + 825, + 903, + 900 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "5202", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "image", + "img_path": "images/2a1a8d350a4e1bbf69bf451f2a6d930a3517314311135404ace9feff9315652c.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 129, + 88, + 316, + 196 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/75e471bbad96869ca92b9930f57ca65cc4de1012673c81c1c817f009822aef74.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 318, + 90, + 496, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/ae2a4402312e4dc25abe2b4a7b6556c65361b0085f311ffa75c1896822ac0122.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 90, + 679, + 195 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/422ef3057227da8502a4891b9f65d9c235f97ba7228b6b93be73e3059ecf4a33.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 142, + 862, + 251 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/dfeae9c5e7edf804c7761b8b69d90cefa9e5711439075cdbf0cf3b5924d0d70e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 196, + 316, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/b0b0b6e11d73c86503357c3d2d3b3a6db8080d345c5b5c32a961f28e6483615b.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 196, + 496, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/2b3f7a48ee2048cc6c906db186e9a7c5e67434fb7063e090aa3ecacedc4999fb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 196, + 679, + 301 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/c34f49219ae3e9362b1626a5b4587a3c28cc171924b7b793e9bad63ad95ee183.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 679, + 369, + 861, + 510 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/fe7c3c5300dd2e5ceb79b29dc4c4f0cc82053d288d3baa899c79d4791bea9dfb.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 133, + 301, + 316, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/68e2eac299c0ef1e34659f2a6045307b0a6df232fdda33a9ab17a6819e1bba54.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 301, + 496, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/8adea137841a19e9c19410f14b3210afd3a95816f103e14160eb7c30a084059e.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 301, + 679, + 441 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/e570e0b6610a08e75690aedfe06d81a3d1f4e08a1ce7bc60eb7308d9a3f87eac.jpg", + "image_caption": [ + "Figure 4. Comparison on ScanNet++(Indoor) [76] and KITTI-360 [35] dataset. Qualitative comparison results on ScanNet++ and KITTI-360 datasets. Here, +C. T. indicates prediction results by appending Calibration Tokens to patch embeddings of the model located above. Calibration Tokens enable models to adapt to different fisheye cameras, especially in regions with large distortions." + ], + "image_footnote": [], + "bbox": [ + 133, + 441, + 316, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/72ec0f8a27bc4a256edb7f916d4ee22ecc24cea759d66ebd9b2723c385fde619.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 316, + 441, + 496, + 580 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/aa0e006ec4d39685618949322c0684b01e48554f5e09c5a52f680af7f3bb4160.jpg", + "image_caption": [], + "image_footnote": [], + "bbox": [ + 500, + 441, + 679, + 580 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "ing a ResNet101 backbone, and compare both its pretrained model and a model trained on our dataset for any fairness concerns. We also compare with FoVA-Depth [36], which is equirectangular projection based, like DepthAnyCamera. Indoor Evaluation. Among the pretrained foundational monocular depth estimators, UniDepth achieves the best performance with our Calibration Tokens on the ScanNet++ indoor dataset as shown in Tab. 1. Notably, our Calibration Tokens enable MiDAS to improve $12\\%$ and DepthAnything to achieve a $17\\%$ improvement in the RMSE metric compared to the model without Calibration Tokens. Similarly, UniDepth benefits from a $13\\%$ improvement in the RMSE metric. Furthermore, compared to the comparison baselines, pretrained DepthAnyCamera and FoVA-Depth, UniDepth with Calibration Tokens surpasses their performance by $11\\%$ and $14\\%$ in the RMSE metric, respectively. DepthAnyCamera and FoVA-Depth utilize camera intrin", + "bbox": [ + 89, + 643, + 482, + 900 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "sics for input images at test time, requiring image transformations back and forth from the equirectangular reference frame, which makes them more error-prone than our direct learning-based approach.", + "bbox": [ + 511, + 643, + 906, + 703 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Outdoor Evaluation. We evaluate FMDEs with our proposed Calibration Tokens against state-of-the-art methods in outdoor environments. The results show that Calibration Tokens consistently improve accuracy across different FMDEs in outdoor scenarios. Specifically, MiDaS and DepthAnything achieve improvement in the RMSE metric. UniDepth also improves $2\\%$ in the RMSE metric, outperforming the comparison baselines. Given that the KITTI-360 dataset contains highly distorted images with a field of view exceeding 180 degrees, our Calibration Tokens demonstrate robustness across various distortion models.", + "bbox": [ + 511, + 704, + 906, + 869 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Our Calibration Tokens are able to outperform DepthAnyCamera and FoVA-Depth without separate", + "bbox": [ + 511, + 869, + 906, + 902 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "5203", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/046b4600cd7c9180711a78149ef27a112c7ddbcd33159b7bb8229bde2caaa163.jpg", + "table_caption": [ + "Table 1. Quantitative comparisons on indoors (ScanNet++) and outdoors (KITTI-360) benchmarks. We evaluated zero-shot monocular depth estimation by incorporating trained Calibration Tokens into recent foundational monocular depth estimators models. Note: Our method uses the same training set for both the indoor and outdoor settings; whereas existing methods train separate models for each setting." + ], + "table_footnote": [], + "table_body": "
TestsetExperimentModelTrain DatasetRMSE ↓δ1 ↑
ScanNet++ [76]BaselineMiDaS [49]Mix 1.4M0.5060.563
+ Calibration TokensMiDaS [49]+ Mix 200K0.4460.569
BaselineDepthAnything [73]Mix 63.5M0.7310.463
+ Calibration TokensDepthAnything [73]+ Mix 200K0.6070.506
BaselineUniDepth [45]Mix 16M0.2790.720
+ Calibration TokensUniDepth [45]+ Mix 200K0.2440.766
ComparisonsDepthAnyCamera [24]Indoor 670K0.2750.761
DepthAnyCamera [24]Mix 200K0.7610.255
FoVA-Depth [36]Indoor 190K0.2850.548
KITTI-360 [35]BaselineMiDaS [49]Mix 1.4M3.3120.586
+ Calibration TokensMiDaS[49]+ Mix 200K2.3480.658
BaselineDepthAnything [73]Mix 63.5M2.2140.839
+ Calibration TokensDepthAnything [73]+ Mix 200K2.0430.810
BaselineUniDepth [45]Mix 16M2.0850.663
+ Calibration TokensUniDepth [45]+ Mix 200K2.0400.664
ComparisonsDepthAnyCamera [24]Outdoor 130K2.0670.852
DepthAnyCamera [24]Mix 200K5.6750.348
FoVA-Depth [36]Outdoor 80K3.0960.632
", + "bbox": [ + 140, + 135, + 854, + 458 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/c058ec082d15bf9c4f1865100451108c72e6e7452e3e17a9960933886b4dac5d.jpg", + "image_caption": [ + "Figure 5. t-SNE plot of fisheye and perspective embeddings. Fisheye embeddings become closer to those of perspective images after being modulated by Calibration Tokens." + ], + "image_footnote": [], + "bbox": [ + 132, + 470, + 437, + 655 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "indoor and outdoor training sets, suggesting the generalization potential of our method to wide ranges of fisheye distortions. Also, the KITTI-360 ground truth points are significantly sparser and more concentrated in ground regions as compared to ScanNet++, which may explain the discrepancy in evaluation metrics. Nonetheless, our method performs comparably without needing to train specialized sets of Calibration Tokens for different fisheye models.", + "bbox": [ + 89, + 714, + 483, + 835 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "4.2. Analysis", + "text_level": 1, + "bbox": [ + 89, + 847, + 192, + 863 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Feature Modulation. To visualize how our Calibration Tokens affect fisheye embeddings, Fig. 5 shows a two-", + "bbox": [ + 89, + 869, + 483, + 902 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/7843bff95324b3bd2dbd6be87ee4039d24e6309aeb6c579829603f949bc0236c.jpg", + "image_caption": [ + "Figure 6. Comparison of attention maps for single and multi-layer tokens. We visualize the attention map of Calibration Tokens to the encoded patch embeddings. Calibration Tokens attend highly to distorted border regions: \"Single Token\" (top) has little effect in most layers due to lower attention as only a set of tokens are append to the input. The proposed multi-layer or \"Layer-wise Tokens\" scheme (bottom) attends to embeddings across all layers." + ], + "image_footnote": [], + "bbox": [ + 521, + 470, + 906, + 612 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "dimensional tSNE reduction to both fisheye and perspective image embeddings from the same set of images. After adding Calibration Tokens to the fisheye embeddings, they are modulated towards the perspective image distribution.", + "bbox": [ + 511, + 733, + 905, + 794 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Layer-wise Tokens. As observed in Fig. 6, when we append only a single set of tokens (\"Single Token\") at the initial transformer block of the pre-trained model, the Calibration Tokens exhibit limited attention to the patch embeddings across most layers. As a result, the patch embeddings of most layers are unchanged. However, when we attach unique tokens at every layer (\"Layer-wise Tokens\"), we see", + "bbox": [ + 511, + 795, + 906, + 902 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "5204", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/3c76a5002ab4c804f854decc70731b3c19f73c7ffb503c9989858d8838835d33.jpg", + "table_caption": [ + "Table 2. Comparison results with finetuning. We conducted experiments comparing finetuning (F.T.) with the use of Calibration Tokens (C.T.) added to the baseline model." + ], + "table_footnote": [], + "table_body": "
DatasetsModelsExp.RMSEδ1
ScanNet++MiDASF.T.2.1780.129
C.T.0.4460.569
DepthAnythingF.T.1.4590.462
C.T.0.6070.506
UniDepthF.T.0.4320.574
C.T.0.2440.766
KITTI-360MiDASF.T.9.2890.098
C.T.2.3480.658
DepthAnythingF.T.4.3620.636
C.T.2.0430.810
UniDepthF.T.3.2170.403
C.T.2.0400.664
", + "bbox": [ + 91, + 135, + 480, + 378 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/77d820d70e09da2f9b8d2ce8923aadcef2fc2105e8a0a4c8e698b49ead21b3e4.jpg", + "table_caption": [ + "Table 3. Analysis on computational cost. We analyze the computational overhead introduced by Calibration Tokens. Values in parentheses indicate the relative increase as a percentage." + ], + "table_footnote": [], + "table_body": "
ModelsModel memoryTokens memoryInference time
MiDAS1.7G0.8M(0.05%)0.6ms(0.8%)
DepthAnything1.7G0.8M(0.05%)0.8ms(0.8%)
UniDepth0.7G0.2M(0.02%)0.4ms(0.7%)
", + "bbox": [ + 91, + 433, + 478, + 518 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "higher attention at more layers. Thus, we opt to use the \"Layer-wise\" approach to better modulate the fisheye patch embeddings toward the distribution of perspective images.", + "bbox": [ + 89, + 534, + 482, + 580 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Comparison with Finetuning. To further analyze the robustness of the Calibration Tokens, we conducted experiments comparing our method with a finetuning approach. We trained the model with a fixed learning rate of $10^{-6}$ on our synthetic fisheye dataset for the same number of iterations. As shown in Tab. 2, the finetuning approach leads to a significant performance drop, highlighting the importance of using Calibration Tokens, which preserve the original model's training on perspective images.", + "bbox": [ + 89, + 580, + 482, + 715 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Computational Cost. Tab. 3 shows the impact of Calibration Tokens on computational costs across different FMDEs. Incorporating Calibration Tokens results in only a $0.05\\%$ and $0.02\\%$ increase in memory usage, less than 1 MB and a $0.8\\%$ and $0.7\\%$ increase in inference time, with an added latency of less than 1 ms. This analysis highlights the efficiency of our proposed Calibration Tokens.", + "bbox": [ + 89, + 715, + 482, + 821 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "4.3. Ablation Study", + "text_level": 1, + "bbox": [ + 89, + 832, + 243, + 849 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We conducted an ablation study on different Calibration Token configurations to validate our contributions. Note that \"Single Token\" refers to a single set of Calibration Tokens", + "bbox": [ + 89, + 854, + 482, + 900 + ], + "page_idx": 7 + }, + { + "type": "table", + "img_path": "images/91b1ac53317f49b2bafc07bd3cc4a8672cbfe8b09ac304a53ce518facadd6628.jpg", + "table_caption": [ + "Table 4. Ablation study. We ablate the training objective and modulate scheme for our proposed Calibration Tokens." + ], + "table_footnote": [], + "table_body": "
DatasetMethodRMSEδ1
ScanNet++Single token0.2600.741
+ LogL1 Loss0.2540.752
+ Layer-wise Tokens0.2440.766
KITTI-360Single token2.0850.656
+ LogL1 Loss2.0650.665
+ Layer-wise Tokens2.0400.664
", + "bbox": [ + 519, + 121, + 895, + 250 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "appended in the first layer of the vision transformer without removal, with L1 loss applied. In this configuration, $\\phi \\in \\mathbb{R}^{M\\times F}$ as opposed to $\\Phi \\in \\mathbb{R}^{L\\times M\\times F}$ in the layer-wise setting. The ablation study on the ScanNet++ and KITTI-360 datasets is performed using the UniDepth model.", + "bbox": [ + 511, + 266, + 903, + 340 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "LogL1 Loss. We observed stable improvements with LogL1 loss compared to baseline L1 loss. As shown in Tab. 4, the LogL1 loss improves both metrics across indoor and outdoor datasets. Qualitative comparisons between L1 and LogL1 objectives are shown in the Supp. Mat.", + "bbox": [ + 511, + 342, + 903, + 417 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Layer-wise Tokens. Tab. 4 demonstrates the advantages of using layer-wise tokens over a single set of Calibration Tokens in the first layer. Even when the same number of tokens is fed to the visual transformer layers, we observed a significant improvement in the contribution of layer-wise tokens. This supports our hypothesis about how the influence of Calibration Tokens diminishes through a forward pass as observed in our experiments by appending a single set of tokens at the first layer. Fig. 6 visualizes attention.", + "bbox": [ + 511, + 419, + 905, + 555 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5. Discussion", + "text_level": 1, + "bbox": [ + 511, + 571, + 627, + 585 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Calibration Tokens enable FMEs to adapt to images captured by fisheye cameras. Empirically, our method improves on monocular depth estimation on fisheye cameras. While our method trains only one set of tokens for both indoor and outdoor settings, our promising results motivate this as a general approach to adapting vision foundational models. Furthermore, a convenience afforded by our method is in the reuse and backward-compatibility of FMDEs with perspective images. This reduces the operational overhead of multi-camera systems by enabling a single FMDE to handle multiple camera inputs - adding cameras become as easy as appending tokens.", + "bbox": [ + 511, + 597, + 906, + 777 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations. While we offer a light-weight method of extending FMDEs to fisheye images, its success inherently depends on the quality and representational power of the underlying FMDEs. If the pretrained model struggles with certain 3D scenes or lighting conditions for perspective images, then these issues carry over. Nonetheless, as novel FMDEs emerge, our framework can be readily transferred to new models using transformer-based architectures.", + "bbox": [ + 511, + 779, + 908, + 900 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "5205", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments This work is supported by NSF 2112562 Athena AI Institute and the Rosenfeld Science Scholars Program.", + "bbox": [ + 89, + 90, + 485, + 137 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 151, + 187, + 167 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Bruno Arsenali, Prashanth Viswanath, and Jelena Novosel. Rotinvmtl: Rotation invariant multinet on fisheye images for autonomous driving applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3", + "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3", + "[3] Aleksei Bochkovskii, AmaAgl Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In International Conference on Learning Representations, 2025. 3", + "[4] Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer. arXiv preprint arXiv:2006.11527, 2020. 4", + "[5] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Transformer-based monocular depth estimation with attention supervision. In 32nd British Machine Vision Conference (BMVC 2021), 2021. 3", + "[6] Xien Chen, Suchisrit Gangopadhyay, Michael Chu, Patrick Rim, Hyoungseob Park, and Alex Wong. Uncle: Unsupervised continual learning of depth completion. arXiv preprint arXiv:2410.18074, 2024. 3", + "[7] Hyesong Choi, Hunsang Lee, Sunkyung Kim, Sunok Kim, Seungryong Kim, Kwanghoon Sohn, and Dongbo Min. Adaptive confidence thresholding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12808-12818, 2021. 3", + "[8] Younjoon Chung, Hyoungseob Park, Patrick Rim, Xiaoran Zhang, Jihe He, Ziyao Zeng, Safa Cicek, Byung-Woo Hong, James S. Duncan, and Alex Wong. Eta: Energy-based test-time adaptation for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3", + "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4", + "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 3", + "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + ], + "bbox": [ + 93, + 176, + 485, + 898 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[12] C Brown Duane. Close-range camera calibration. Photogramm. Eng, 37(8):855-866, 1971. 3", + "[13] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 3", + "[14] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3", + "[15] Vadim Ezhov, Hyoungseob Park, Zhaoyang Zhang, Rishi Upadhyay, Howard Zhang, Chethan Chinder Chandrappa, Achuta Kadambi, Yunhao Ba, Julie Dorsey, and Alex Wong. All-day depth completion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 3", + "[16] Xiaohan Fei, Alex Wong, and Stefano Soatto. Geosupervised visual depth prediction. IEEE Robotics and Automation Letters, 4(2):1661-1668, 2019. 3", + "[17] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 3, 4", + "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3", + "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 3", + "[20] Clément Godard, Oisin Mac Aodha, and Gabriel J Brossow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3", + "[21] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3", + "[22] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 3", + "[23] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 3", + "[24] Yuliang Guo, Sparsh Garg, S Mahdi H Miangoleh, Xinyu Huang, and Liu Ren. Depth any camera: Zero-shot met" + ], + "bbox": [ + 516, + 92, + 903, + 900 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "5206", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "ric depth estimation from any camera. arXiv preprint arXiv:2501.02464, 2025.3,5,7", + "[25] Juho Kannala and Sami Brandt. A generic camera calibration method for fish-eye lenses. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., pages 10-13. IEEE, 2004. 3, 4", + "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9492-9502, 2024. 3", + "[27] R Kingslake. A history of the photographic lens. University of Rochester NY, 145, 1989. 3", + "[28] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 61-71, 2021. 3", + "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 3", + "[30] Dong Lao, Yangchao Wu, Tian Yu Liu, Alex Wong, and Stefano Soatto. Sub-token vit embedding via stochastic resonance transformers. In International Conference on Machine Learning. PMLR, 2024.", + "[31] Dong Lao, Fengyu Yang, Daniel Wang, Hyoungseob Park, Samuel Lu, Alex Wong, and Stefano Soatto. On the viability of monocular depth pre-training for semantic segmentation. In European Conference on Computer Vision. Springer, 2024.", + "[32] Bo Li, Chunhua Shen, Yuchao Dai, Anton Van Den Hengel, and Mingyi He. Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1119-1127, 2015. 3", + "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 3", + "[34] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology, 30(3):725-733, 2019. 3", + "[35] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 5, 6, 7", + "[36] Daniel Lichy, Hang Su, Abhishek Badki, Jan Kautz, and Orazio Gallo. Fova-depth: Field-of-view agnostic depth estimation for cross-dataset generalization. In 2024 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2024. 3, 6, 7" + ], + "bbox": [ + 91, + 92, + 483, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[37] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 3", + "[38] Tian Yu Liu, Parth Agrawal, Allison Chen, Byung-Woo Hong, and Alex Wong. Monitored distillation for positive congruent depth completion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 35-53. Springer, 2022. 3", + "[39] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2301, 2021. 3", + "[40] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5667-5675, 2018. 3", + "[41] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 3", + "[42] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20519-20529, 2024. 3", + "[43] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 5", + "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 3", + "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 5, 7", + "[46] Matteo Poggi, Fabio Tosi, and Stefano Mattoccia. Learning monocular depth estimation with unsupervised trinocular assumptions. In 2018 International conference on 3d vision (3DV), pages 324-333. IEEE, 2018. 3", + "[47] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3227-3237, 2020. 3", + "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3", + "[49] Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross" + ], + "bbox": [ + 516, + 92, + 906, + 900 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "5207", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Dataset Transfer. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(03):1623-1637, 2022. 2, 3, 5, 7", + "[50] Patrick Rim, Hyoungseob Park, Vadim Ezhov, Jeffrey Moon, and Alex Wong. Radar-guided polynomial fitting for metric depth estimation. arXiv preprint arXiv:2503.17182, 2025. 3", + "[51] Patrick Rim, Hyoungseob Park, Ziyao Zeng, Younjoon Chung, and Alex Wong. Protodepth: Unsupervised continual depth completion with prototypes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6304-6316, 2025. 3", + "[52] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5", + "[53] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, 2012. 5", + "[54] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 3", + "[55] Daniel E Stevenson and Margaret M Fleck. Nonparametric correction of distortion. In Proceedings Third IEEE Workshop on Applications of Computer Vision. WACV'96, pages 214-219. IEEE, 1996. 3", + "[56] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 5", + "[57] Fabio Tosi, Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning monocular depth estimation infusing traditional stereo knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9799-9809, 2019. 3", + "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3", + "[59] Rishi Upadhyay, Howard Zhang, Yunhao Ba, Ethan Yang, Blake Gella, Sicheng Jiang, Alex Wong, and Achuta Kadambi. Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 3", + "[60] Chaoyang Wang, José Miguel Buenaposada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 3", + "[61] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic" + ], + "bbox": [ + 91, + 92, + 482, + 900 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5", + "[62] Jamie Watson, Michael Firman, Gabriel J Brostow, and Daniyar Turmukhambetov. Self-supervised monocular depth hints. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2162-2171, 2019. 3", + "[63] Alex Wong and Stefano Soatto. Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5644-5653, 2019.", + "[64] Alex Wong, Safa Cicek, and Stefano Soatto. Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems, 33:8486-8497, 2020.", + "[65] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 5", + "[66] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 6(2):1495-1502, 2021.", + "[67] Alex Wong, Xiaohan Fei, Byung-Woo Hong, and Stefano Soatto. An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters, 6 (2):3120-3127, 2021. 3", + "[68] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision, pages 274-293. Springer, 2024. 3, 5", + "[69] Chao Xia, Chenfeng Xu, Patrick Rim, Mingyu Ding, Nanning Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Quadric representations for lidar odometry, mapping and localization. IEEE Robotics and Automation Letters, 8 (8):5023-5030, 2023. 3", + "[70] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3", + "[71] Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5354-5362, 2017. 3", + "[72] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26340-26353, 2024. 3" + ], + "bbox": [ + 516, + 92, + 906, + 898 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "5208", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 7", + "[74] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3353-3362, 2019. 3", + "[75] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3", + "[76] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 5, 6, 7", + "[77] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5684-5693, 2019. 3", + "[78] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. Fisheyerecnet: A multi-context collaborative deep network for fisheye image rectification. In Proceedings of the European conference on computer vision (ECCV), pages 469-484, 2018. 4", + "[79] Senthil Yogamani, David Unger, Venkatraman Narayanan, and Varun Ravi Kumar. Fisheyebevseg: Surround view fisheye cameras based bird's-eye view segmentation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1331-1334, 2024. 3", + "[80] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3916-3925, 2022. 3", + "[81] Ziyao Zeng, Jingcheng Ni, Daniel Wang, Patrick Rim, Younjoon Chung, Fengyu Yang, Byung-Woo Hong, and Alex Wong. Priordiffusion: Leverage language prior in diffusion models for monocular depth estimation. arXiv preprint arXiv:2411.16750, 2024. 3", + "[82] Ziyao Zeng, Daniel Wang, Fengyu Yang, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Wordepth: Variational language prior for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9708-9719, 2024.", + "[83] Ziyao Zeng, Yangchao Wu, Hyoungseob Park, Daniel Wang, Fengyu Yang, Stefano Soatto, Dong Lao, Byung-Woo Hong, and Alex Wong. Rsa: Resolving scale ambiguities in monocular depth estimators through language descriptions. Advances in neural information processing systems, 37, 2024. 3", + "[84] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with" + ], + "bbox": [ + 91, + 92, + 480, + 900 + ], + "page_idx": 11 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3", + "[85] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 3", + "[86] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 International Conference on 3D Vision (3DV), pages 668-678. IEEE, 2022. 3", + "[87] Guoyang Zhao, Yuxuan Liu, Weiqing Qi, Fulong Ma, Ming Liu, and Jun Ma. Fisheyedgepth: A real scale self-supervised depth estimation model for fisheye camera. arXiv preprint arXiv:2409.15054, 2024. 3", + "[88] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + ], + "bbox": [ + 516, + 92, + 903, + 415 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "5209", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_model.json b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_model.json new file mode 100644 index 0000000000000000000000000000000000000000..980a6150fc4013c4949745d727edd97e9a60606f --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_model.json @@ -0,0 +1,2523 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.043 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.126, + 0.13, + 0.873, + 0.175 + ], + "angle": 0, + "content": "Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens" + }, + { + "type": "text", + "bbox": [ + 0.104, + 0.203, + 0.93, + 0.24 + ], + "angle": 0, + "content": "Suchisrit Gangopadhyay\\(^{1*}\\), Jung-Hee Kim\\(^{2*}\\), Xien Chen\\(^{1*}\\), Patrick Rim\\(^{1}\\), Hyoungseob Park\\(^{1}\\), Alex Wong\\(^{1}\\), Yale University\\(^{1}\\), Michigan State University\\(^{2}\\)" + }, + { + "type": "text", + "bbox": [ + 0.101, + 0.242, + 0.921, + 0.258 + ], + "angle": 0, + "content": "{rit.gangopadhyay,xien.chen,patrick.rim,hyoungseob.park,alex.wong}@yale.edu, kimjun84@msu.edu" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.292, + 0.327, + 0.308 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.324, + 0.486, + 0.718 + ], + "angle": 0, + "content": "We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: github.com/JungHeeKim29/calibration-token." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.746, + 0.222, + 0.763 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.772, + 0.484, + 0.879 + ], + "angle": 0, + "content": "Three-dimensional (3D) reconstruction is a fundamental component in many spatial applications, including autonomous vehicles, extended reality (XR), robotic manipulation. Each of these applications has unique demands for the field of view (FOV), often wider than the standard (perspective) camera. To meet this need, these applications tend to be deployed on systems equipped with fish" + }, + { + "type": "image", + "bbox": [ + 0.517, + 0.291, + 0.644, + 0.38 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.379, + 0.596, + 0.391 + ], + "angle": 0, + "content": "Input" + }, + { + "type": "image", + "bbox": [ + 0.646, + 0.291, + 0.774, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.689, + 0.379, + 0.732, + 0.39 + ], + "angle": 0, + "content": "MiDaS" + }, + { + "type": "image", + "bbox": [ + 0.776, + 0.291, + 0.906, + 0.379 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.795, + 0.379, + 0.887, + 0.391 + ], + "angle": 0, + "content": "DepthAnything" + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.395, + 0.907, + 0.465 + ], + "angle": 0, + "content": "Figure 1. Foundational monocular depth estimators fail on fisheye images. Despite being trained on large-scale datasets, foundational monocular depth estimators (FMDEs) models produce erroneous outputs. The inaccurate, blurry estimates are caused by a covariate shift that stem from fisheye distortion." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.478, + 0.907, + 0.583 + ], + "angle": 0, + "content": "eye or other wide-angle cameras, which allows for wider coverage of the 3D environment. However, images captured by these cameras also come with substantial distortion, which arise from differences in projective geometry, where straight lines within the 3D environment or the 3D scene are preserved in perspective images but may appear curved in fisheye images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.584, + 0.909, + 0.765 + ], + "angle": 0, + "content": "Foundational monocular depth estimators (FMDEs) [44, 48, 73] are trained on orders of tens of millions of images, enabling them to generalize across a wide range 3D scenes. However, their training data is comprised of internet images, which are predominantly captured using perspective cameras. Hence, despite being trained on large-scale datasets, FMDEs produce erroneous estimates when transferred to fisheye images (see Fig. 1). These errors stem from a covariate shift, which can be characterized by changes in camera calibration (intrinsic, distortion) parameters – leading to differences in object appearance and their perceived depth or distance from the camera." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.765, + 0.909, + 0.903 + ], + "angle": 0, + "content": "To address fisheye distortion, one solution is to recalibrate and undistort images or perform a map projection to some canonical reference frame. In principle, if one has the correct calibration, it is possible to re-project a fisheye image into a perspective-like view (or vice versa). In practice, however, there are several problems: (1) The calibration process itself can be error-prone and sensitive to physical perturbations in the camera system. Minor bumps, focus changes, or lens replacements can degrade or invalid" + }, + { + "type": "page_footnote", + "bbox": [ + 0.11, + 0.888, + 0.218, + 0.9 + ], + "angle": 0, + "content": "*Equal contribution" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5198" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.482, + 0.183 + ], + "angle": 0, + "content": "idate previously computed intrinsic parameters. (2) Even when re-projection is performed accurately, the transformation introduces latency and spatial artifacts (e.g., stretching, cropping, aliasing, loss). When used as a preprocessing step for existing pretrained depth estimators, these artifacts still present a covariate shift and can degrade performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.186, + 0.483, + 0.412 + ], + "angle": 0, + "content": "Another solution is to train a separate monocular depth estimator specifically for fisheye images. However, publicly available image datasets for fisheye cameras are orders of tens to hundreds of times smaller than those for perspective cameras. Hence, it is difficult to assemble sufficient data to reach the large-scale training requirement of an FMDE. Nonetheless, one can adapt or finetune existing FMDEs for fisheye imagery. While this can improve performance on fisheye images, it introduces the risk of parameter drift, where the resulting FMDEs may lose their generalizability across 3D scenes. Moreover, the resulting finetuned model becomes specialized to fisheye cameras, limiting its applicability to other camera types, which adds operational overhead in applications involving mixed camera systems, e.g., autonomous vehicles or robotics." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.415, + 0.483, + 0.822 + ], + "angle": 0, + "content": "To address these challenges, we propose a novel approach termed Calibration Tokens. Our key insight is that existing FMDEs are already capable of estimating depth for perspective images, and that errors on fisheye images are caused by a covariate shift due to differences in camera calibration and distortion. Hence, rather than retraining or finetuning the entire model, we aim to \"recalibrate\" the fisheye latent embeddings such that they become more conducive to an FMDE originally trained on perspective images. Leveraging the fact that many FMDEs [44, 48, 49, 73] follow a Transformer-based architecture [11], we will exploit the (self- and cross-) attention mechanism to modulate the latent (token) embeddings by inserting Calibration Tokens as part of the input. Therefore, the existing FMDE will remain effectively unchanged, while Calibration Tokens serve to adapt their internal representations to mitigate the covariate shift by aligning the latent embeddings of fisheye images to the distribution of latent embeddings of perspective images. This design also allows us to preserve the original image content without performing any spatial re-projection, ensuring the process is lossless in terms of the raw pixels. Our hypothesis is that by adding a small set of trainable tokens to encode the fisheye camera calibration information and utilizing them to recalibrate the latent embeddings, we will be able to reuse existing FMDEs trained on perspective images and adapt them to fisheye images without sacrificing their generalizability across diverse 3D scenes." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.825, + 0.483, + 0.901 + ], + "angle": 0, + "content": "To train these Calibration Tokens, we propose a self-supervised objective that leverages inverse warping in the input and output spaces. FMDEs can infer high-fidelity depth maps for perspective images, so we use the perspective image depth estimates as our training target. We then" + }, + { + "type": "image", + "bbox": [ + 0.52, + 0.09, + 0.907, + 0.304 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.313, + 0.907, + 0.356 + ], + "angle": 0, + "content": "Figure 2. Inference on different cameras. Calibration Tokens enable foundational monocular depth estimators to adapt to fisheye images while maintaining performance on perspective images." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.372, + 0.905, + 0.552 + ], + "angle": 0, + "content": "induce artificial distortion on perspective images to create pairs of perspective and synthetic fisheye images with diverse fisheye distortions. However, rather than doing the same in the output space, we undistort the fisheye depth maps to the original perspective reference frame to compute a self-supervised loss between the undistorted fisheye and perspective depth maps. By minimizing this self-supervised loss, the Calibration Tokens learn to align fisheye image embeddings to those of perspective images in the latent space, without any labels. Additionally, computing the loss in the original perspective frame allows our method to preserve the supervision signal instead of introducing artifacts." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.553, + 0.906, + 0.718 + ], + "angle": 0, + "content": "Our approach allows us to bypass the need to compile large-scale fisheye datasets by exploiting the abundance of perspective image datasets. As our method operates in the reference frame of the input, we also avoid transformation artifacts at inference time, whether in the input or output space. Furthermore, our method preserves compatibility with perspective images: One simply needs to append or remove Calibration Tokens for FMDEs to be applied to fisheye or perspective images. We demonstrate our method on indoors and outdoors across several recent FMDEs and consistently improve over baselines." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.906, + 0.901 + ], + "angle": 0, + "content": "Our contributions: (1) We propose a novel approach to extend foundational monocular depth estimators (FMDEs) trained on perspective images to fisheye images. (2) We introduce Calibration Tokens that modulate the latent embeddings of fisheye images towards the distribution of perspective image embeddings. (3) We introduce a self-supervised training objective that recalibrates input perspective image to fisheye images, but \"undos\" the transformation in the output to enable loss computation on high-fidelity (perspective) depth maps inferred by FMDEs. (4) Our approach only requires training one set of tokens to achieve state-of-the-art performance for both indoors and outdoors." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5199" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.09, + 0.241, + 0.106 + ], + "angle": 0, + "content": "2. Related Works" + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.116, + 0.482, + 0.552 + ], + "angle": 0, + "content": "Monocular Depth Estimation can be trained in a supervised or unsupervised manner. Supervised methods [13, 14, 29-32, 37, 71, 77] minimize the difference between depth estimates and ground-truth depth maps. [18] re-formulated the problem as ordinal regression while other methods proposed architectures innovations. [2] partitions depth ranges into adaptive bins. [5] incorporates an attention-based up-sample block. [33] employs hierarchical aggregation and heterogeneous interaction modules. [80] uses neural window fully-connected CRFs to compute energy. [59] synthesizes perspectively accurate images to enrich training data. Additional inputs e.g., language [81-83], lidar [6, 15, 51, 74], radar [50, 54], are used to enable metric-scale depth estimates. Unsupervised methods [7, 8, 23, 38, 42, 47, 57, 62-67, 84] minimizes photometric reconstruction error. [19] frames depth estimation as a novel view synthesis problem. [20] introduces a left-right consistency loss. [88] uses a pose network to enable unsupervised training on video sequences. [21] introduced auto-masking and min-reprojection loss. Additional loss terms based on visual odometry [16, 60], iterative closest point [40], surface normals [75], trinocular assumption [46], and semantic segmentation [22, 28] were also introduced. [39] redesigned the skip connection and decoders to extract high-resolution features, [86] combined global and local representations and [85] introduced a lightweight architecture with dilated convolution and attention. AugUndo [68] leveraged invertibility of transformation groups for data augmentation." + }, + { + "type": "text", + "bbox": [ + 0.094, + 0.554, + 0.482, + 0.794 + ], + "angle": 0, + "content": "Foundational Monocular Depth Estimators are trained with supervised or semi-supervised learning on large-scale datasets. MiDaS [49] is the first to demonstrate generalizable monocular depth estimation by compiling datasets for large-scale training. DPT [48] extended the approach and introduced transformers for dense predictions. Marigold [26] repurposes diffusion models for monocular depth estimation. DepthAnything [73] proposes a pseudo-labeling method to curate a large-scale dataset. Additionally, UniDepth [44] employs a camera self-prompting module and a pseudo-spherical output space, enabling metric-scale depth prediction across diverse 3D scenes without relying on external camera parameters. DepthPro [3] proposes a multi-scale vision transformer for metric-scale depth estimation. As all of these FMDEs are trained on perspective images, they fail to generalize to fisheye cameras." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.796, + 0.482, + 0.902 + ], + "angle": 0, + "content": "Fisheye Images. Images taken by a fisheye camera are distorted and unsuitable for use in a perspective image encoder. Existing distortion correction algorithms [12, 25] rely on different camera projection models [27, 41, 55] to undistort images into a perspective view. However, these methods depend on camera calibration parameters, which can introduce artifacts due to calibration inaccuracies. Recent approaches" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.905, + 0.287 + ], + "angle": 0, + "content": "[24, 36] demonstrate training a separate model to perform depth estimation with different camera types. They utilize an equirectangular projection to project points from different reference frames to a canonical equirectangular frame, but this can incur transformation artifacts and distortions. Additionally, deep-learning-based methods [17, 34] that aim to rectify distortion have been introduced. However, these methods require a large number of parameters with limited accuracy and field of view. Consequently, many recent works targeted for fisheye images involve training an entire network [1, 79, 87] exclusively on fisheye images. Our method extends foundational monocular depth estimators to fisheye images instead." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.288, + 0.905, + 0.515 + ], + "angle": 0, + "content": "Token-Based Methods. Recent transformer-based architectures represent input images [11] (or other modalities [69, 70]) as sequences of tokens. In many cases, an additional token (e.g. [CLS] token in BERT [10] or the distillation token in DeiT [58]) is employed to aggregate information across all tokens. Such tokens can be adapted to various purposes, acting as a compact representation that \"binds\" or fuses information, e.g., [72] uses tokens to learn synthetic and real tactile response maps. Inspired by these advances, we introduce minimal trainable tokens appended to the fisheye embedding, enabling the model to \"bind\" or reconcile image distortions within a frozen backbone. Our approach is lightweight and requires no major architectural modifications, but extends foundational monocular depth estimators trained on perspective images to fisheye images." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.527, + 0.605, + 0.543 + ], + "angle": 0, + "content": "3. Method" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.552, + 0.905, + 0.689 + ], + "angle": 0, + "content": "Let \\(I:\\Omega \\mapsto \\mathbb{R}^3\\) denote an RGB image obtained from a calibrated camera and \\(\\Omega \\subset \\mathbb{R}^2\\) the image space. Monocular depth estimation aims to learn a parameterized function \\(h_{\\omega ,\\psi}:\\mathbb{R}^3\\to \\mathbb{R}_+\\) that maps an image to a depth map \\(d:\\Omega \\mapsto \\mathbb{R}_{+}\\). We assume access to a foundational monocular depth estimator (FMDE) pretrained on some large-scale dataset of perspective images. We will pair each image with (pseudo)ground truth \\(\\tilde{d} = h_{\\omega ,\\psi}(I)\\) inferred by the FMDE to obtain a training dataset \\(\\mathcal{D} = \\{(I^{(n)},\\tilde{d}^{(n)})\\}_{n = 1}^{N}\\)." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.69, + 0.905, + 0.901 + ], + "angle": 0, + "content": "To extend FMDEs, trained on perspective images, to fisheye images, we introduce Calibration Tokens as an adaptation mechanism. Due to the prevalence of Transformer architectures in many FMDEs, we train a set of lightweight tokens to model the change in calibration between a perspective camera and different fisheye cameras. The goal of our Calibration Token is to recalibrate or translate latent embeddings of fisheye images back to those of perspective images. Our method takes advantage of the attention mechanism inherent in FMDEs and enable Calibration Tokens to modulate the latent embeddings, thus facilitating latent alignment. The outcome is an FMDE that is capable of inferring depth for fisheye images with Calibration Tokens, and perspective images without." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5200" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.094, + 0.904, + 0.341 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.089, + 0.352, + 0.908, + 0.396 + ], + "angle": 0, + "content": "Figure 3. Overview of our method. We introduce a set of trainable Calibration Tokens, which is appended to the input sequence of the fisheye image tokens. The Calibration Tokens are trained to adapt the model to produce accurate depth maps for images with various fisheye distortions. A unique fisheye calibration token is appended to the input of each new layer of the encoder." + }, + { + "type": "title", + "bbox": [ + 0.09, + 0.41, + 0.461, + 0.426 + ], + "angle": 0, + "content": "3.1. Extending FMDEs with Calibration Tokens" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.435, + 0.484, + 0.558 + ], + "angle": 0, + "content": "Specifically, let \\(\\phi \\in \\mathbb{R}^{M\\times F}\\) represent a set of Calibration Tokens, where \\(M\\) denotes the number of tokens and \\(F\\) their number of dimensions. For a given layer within the encoder \\(f_{\\omega}\\) of an FMDE, we will concatenate Calibration Tokens to the input sequence of patches or (embeddings) of the vision transformer: \\(f_{\\omega}([I;\\phi]) = [z^{(L)};f_{\\omega}(\\phi)]\\), where \\(z\\) denotes the latent embeddings recalibrated by \\(\\phi\\), \\(L\\) the last layer, and \\([\\cdot ;]\\) the concatenation operation." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.563, + 0.484, + 0.852 + ], + "angle": 0, + "content": "As each layer denotes a separate latent space, we extend our approach to a multi-layer modulation scheme. Let \\(\\Phi \\in \\mathbb{R}^{L\\times M\\times F}\\) be the set of Calibration Tokens for each layer \\(l\\in \\{1,\\dots,L\\}\\). A unique set of Calibration Tokens \\(\\phi^{(l)}\\in \\mathbb{R}^{M\\times F}\\) is appended at each encoder layer: \\(f_{\\omega}^{(l)}[(z^{(l - 1)};\\phi^{(l)})] = [z^{(l)};f_{\\omega}^{(l)}(\\phi^{(l)})]\\) for a layer \\(l\\). Each set of Calibration Tokens will modulate the patch embeddings for a specific layer through the attention mechanism; hence, following the convention in existing works [4, 9], we discard Calibration Tokens from the encoder output. A key insight is that the FMDE is already able to estimate high-fidelity depth maps for perspective images. We posit that the covariate shift exists in the encodings of fisheye image. Hence, once the latent embeddings of fisheye images have been recalibrated to those of perspective images, the decoder will be able to estimate depth to similar fidelity as perspective images. Therefore, we do not utilize Calibration Tokens to modulate the decoder layers. The final estimate is obtained by \\(\\hat{d} = g_{\\psi}(z^{(L)})\\), where \\(g_{\\psi}\\) denotes the decoder." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Since our method does not apply spatial transformations during inference, it remains entirely lossless for input images. Additionally, it offers several efficiencies: (1) The" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.41, + 0.907, + 0.607 + ], + "angle": 0, + "content": "only trainable parameters in our method are the light-weight Calibration Tokens, which consist of significantly fewer parameters than vision transformer models. Our method introduces minimal computational overhead and results in lower time and space complexity than training or finetuning a full model. (2) Our approach is backward-compatible with perspective images. By omitting our Calibration Tokens, an FMDE maintains its original depth estimation performance on perspective images. (3) At inference, camera intrinsics are not required, as the training process allows generalization across various fisheye camera intrinsics. As a result, our method eliminates the need for the arduous and error-prone calibration process after training." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.615, + 0.774, + 0.631 + ], + "angle": 0, + "content": "3.2. Learning Calibration Tokens" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.637, + 0.907, + 0.819 + ], + "angle": 0, + "content": "To train Calibration Tokens, we will leverage the abundance of publicly available perspective image datasets. During our training, we synthesize fisheye images from perspective images by recalibrating them using artificial fisheye intrinsic and distortion parameters. This will produce pairs of perspective and fisheye images from which we can leverage self-supervision, and allows us to use a much larger training dataset than exclusively training with real fisheye images. We follow previous approaches [17, 78] to obtain synthetic fisheye images from the calibrated perspective images. Our synthesized fisheye images follow the distortion model introduced by Kannala & Brandt [25]:" + }, + { + "type": "equation", + "bbox": [ + 0.59, + 0.828, + 0.905, + 0.846 + ], + "angle": 0, + "content": "\\[\nr (\\theta) = k _ {1} \\theta + k _ {2} \\theta^ {3} + k _ {3} \\theta^ {5} + k _ {4} \\theta^ {7}, \\tag {1}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.856, + 0.907, + 0.901 + ], + "angle": 0, + "content": "where \\(\\theta\\) denotes the angle between the ray and the optical axis, and \\(\\{k_i\\}_{i = 1}^4\\) are distortion coefficients that can represent most of the real world fisheye distortion models. The" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "5201" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "change in coordinate between \\((x,y)\\) in the perspective image and \\((x^{\\prime},y^{\\prime})\\) in the fisheye image can be formulated as" + }, + { + "type": "equation", + "bbox": [ + 0.163, + 0.129, + 0.483, + 0.165 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} x ^ {\\prime} = r (\\theta) \\cos (\\varphi), y ^ {\\prime} = r (\\theta) \\sin (\\varphi), \\tag {2} \\\\ \\varphi = \\arctan ((y - c _ {x}) / (x - c _ {y})), \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.172, + 0.483, + 0.262 + ], + "angle": 0, + "content": "where \\((c_x, c_y)\\) is the principal point in the perspective image. We define the transformation from the perspective to the fisheye reference frame as \\(T\\) and its inverse transformation as \\(T^{-1}\\). Our training dataset is composed of perspective images and synthesized distorted image pairs with corresponding forward and inverse transformations." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.263, + 0.484, + 0.414 + ], + "angle": 0, + "content": "Loss Function. Inspired by AugUndo [68] and their use of invertible transformations to preserve the supervision signal by undoing data augmentation, we propose to synthesize fisheye images from the abundance of perspective images as inputs, but undistort the output to facilitate loss computation. By remapping depth estimates of synthetic fisheye images to the perspective frame, we enable the use of high-fidelity estimates inferred by FMDEs on perspective images as supervision. Hence, we can optimize Calibration Tokens with the following self-supervised loss:" + }, + { + "type": "equation", + "bbox": [ + 0.099, + 0.42, + 0.483, + 0.462 + ], + "angle": 0, + "content": "\\[\n\\arg \\min _ {\\Psi} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {x \\in \\Omega} \\log \\left(\\left| \\tilde {d} ^ {(n)} (x) - T ^ {- 1} \\circ \\hat {d} ^ {(n)} (x) \\right| + 1\\right), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.47, + 0.483, + 0.652 + ], + "angle": 0, + "content": "where \\(\\tilde{d} = h_{\\omega, \\psi}(I)\\) and \\(\\hat{d} = h_{\\omega, \\psi}(T \\circ I; \\Phi)\\) denotes the predicted depth map from the given perspective image and synthesized fisheye image, respectively. \\(\\Phi\\) denotes the proposed trainable Calibration Token appended to the patch embeddings. Calibration Tokens are trained to minimize the difference between the perspective output and the fisheye output re-projected into the perspective reference frame. Eq. (3) follows the log of absolute differences (\\(\\log L1\\)) proposed in [43], which enhances training stability and empirically outperforms L1 loss, especially in border regions where discrepancies between perspective and fisheye images are most significant (see Sec. 4.3 for details)." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.653, + 0.483, + 0.82 + ], + "angle": 0, + "content": "It is important to note that attempting to instead transform the perspective depth maps outputted by FMDEs to the fisheye reference frame for the loss computation introduces information loss in the training objective. This will lead to re-projection artifacts in the supervision and result in learning inaccuracies during training. In Section A of the Supp. Mat., we further demonstrate comparison results between training in fisheye image space and perspective image space. Our training scheme is self-supervised and requires only calibrated perspective images, which can be easily obtained, making our approach both scalable and practical." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.831, + 0.223, + 0.848 + ], + "angle": 0, + "content": "4. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Datasets. Training our Calibration Tokens requires only calibrated perspective images, enabling us to leverage significantly more data compared to training solely on fisheye" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.137 + ], + "angle": 0, + "content": "images. Moreover, since our loss is computed based on comparisons with perspective image outputs, ground truth is not required for our training pipeline." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.138, + 0.905, + 0.259 + ], + "angle": 0, + "content": "Training datasets: NYUv2 [53] has a variety of perspective indoor scenes; VOID [65] contains indoor perspective office, classroom and stairwell scenes; IRS [61] contains rendered perspective scenes of home, restaurant, and store settings; Hypersim [52] comprises photorealistic, synthetic images of indoor residential and commercial environments in a perspective reference frame. Waymo [56] dataset consists of a diverse set of urban driving scenes." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.26, + 0.905, + 0.395 + ], + "angle": 0, + "content": "Test datasets: ScanNet++ [76] offers 3D reconstructions of diverse indoor scenes, captured using laser scanning and DSLR imaging with a fisheye lens, which allows us to evaluate with real fisheye images and ground truth depth maps. KITTI-360 [35] includes suburban driving scenes captured with a multi-sensor setup, including fisheye cameras. Notably, it features a different field of view compared to ScanNet++ [76], allowing us to assess the generalization capability of the Calibration Tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.397, + 0.905, + 0.487 + ], + "angle": 0, + "content": "Models. Calibration Tokens do not require specific settings and can be integrated into any model utilizing a vision transformer. We evaluate the effectiveness of our approach using by extending MiDaS [49], DepthAnything [73], and UniDepth [45] to fisheye images. Note that we used 8 tokens per layer for each of the model experiments." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.488, + 0.905, + 0.548 + ], + "angle": 0, + "content": "Evaluation Metrics. We evaluate depth prediction accuracy using standard metrics from monocular depth estimation of root mean squared error (RMSE) and \\(\\delta_{1}\\). Details on these metrics can be found in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.549, + 0.906, + 0.791 + ], + "angle": 0, + "content": "Implementation Details. We trained our Calibration Tokens based on 3 different FMDEs (MiDas[49], DepthAnything[49], UniDepth[45]). We utilized the pretrained ViT-L backbone for MiDas [49] and DepthAnything [49], and the ViT-S backbone for UniDepth [45]. We trained our model on 4 NVIDIA 3090 GPUs for 40k iterations with a batch size of 16. For input, we used images in the resolution of \\(518 \\times 518\\). For testing, we used \\(462 \\times 616\\) resolution on the ScanNet++ dataset [76], and \\(700 \\times 700\\) on the KITTI-360 dataset [35] to preserve its aspect ratio. We also synthesize random fisheye distortions in the training images. Our Calibration Tokens are trained with a joint dataset consisting of indoor and outdoor datasets totaling up to only 200K samples, and as shown in Tab. 1, obtain comparable results on both domains with fewer samples than existing methods [24] that are trained specifically for each." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.803, + 0.645, + 0.817 + ], + "angle": 0, + "content": "4.1. Main Result" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.905, + 0.901 + ], + "angle": 0, + "content": "We conduct experiments to analyze the impact of Calibration Tokens on model performance. As a baseline, we compare our model to DepthAnyCamera [24], the state-of-the-art monocular depth estimation (MDE) method for fisheye images. Here, we evaluate the DepthAnyCamera model us" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5202" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.13, + 0.089, + 0.317, + 0.197 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.319, + 0.091, + 0.497, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.091, + 0.68, + 0.196 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.143, + 0.864, + 0.252 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.197, + 0.318, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.197, + 0.497, + 0.303 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.197, + 0.68, + 0.302 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.68, + 0.37, + 0.862, + 0.511 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.303, + 0.317, + 0.443 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.303, + 0.497, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.303, + 0.68, + 0.442 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.134, + 0.443, + 0.317, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.318, + 0.443, + 0.497, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image", + "bbox": [ + 0.5, + 0.443, + 0.68, + 0.582 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.586, + 0.908, + 0.629 + ], + "angle": 0, + "content": "Figure 4. Comparison on ScanNet++(Indoor) [76] and KITTI-360 [35] dataset. Qualitative comparison results on ScanNet++ and KITTI-360 datasets. Here, +C. T. indicates prediction results by appending Calibration Tokens to patch embeddings of the model located above. Calibration Tokens enable models to adapt to different fisheye cameras, especially in regions with large distortions." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.644, + 0.483, + 0.901 + ], + "angle": 0, + "content": "ing a ResNet101 backbone, and compare both its pretrained model and a model trained on our dataset for any fairness concerns. We also compare with FoVA-Depth [36], which is equirectangular projection based, like DepthAnyCamera. Indoor Evaluation. Among the pretrained foundational monocular depth estimators, UniDepth achieves the best performance with our Calibration Tokens on the ScanNet++ indoor dataset as shown in Tab. 1. Notably, our Calibration Tokens enable MiDAS to improve \\(12\\%\\) and DepthAnything to achieve a \\(17\\%\\) improvement in the RMSE metric compared to the model without Calibration Tokens. Similarly, UniDepth benefits from a \\(13\\%\\) improvement in the RMSE metric. Furthermore, compared to the comparison baselines, pretrained DepthAnyCamera and FoVA-Depth, UniDepth with Calibration Tokens surpasses their performance by \\(11\\%\\) and \\(14\\%\\) in the RMSE metric, respectively. DepthAnyCamera and FoVA-Depth utilize camera intrin" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.704 + ], + "angle": 0, + "content": "sics for input images at test time, requiring image transformations back and forth from the equirectangular reference frame, which makes them more error-prone than our direct learning-based approach." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.705, + 0.907, + 0.87 + ], + "angle": 0, + "content": "Outdoor Evaluation. We evaluate FMDEs with our proposed Calibration Tokens against state-of-the-art methods in outdoor environments. The results show that Calibration Tokens consistently improve accuracy across different FMDEs in outdoor scenarios. Specifically, MiDaS and DepthAnything achieve improvement in the RMSE metric. UniDepth also improves \\(2\\%\\) in the RMSE metric, outperforming the comparison baselines. Given that the KITTI-360 dataset contains highly distorted images with a field of view exceeding 180 degrees, our Calibration Tokens demonstrate robustness across various distortion models." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.871, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Our Calibration Tokens are able to outperform DepthAnyCamera and FoVA-Depth without separate" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5203" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.089, + 0.907, + 0.133 + ], + "angle": 0, + "content": "Table 1. Quantitative comparisons on indoors (ScanNet++) and outdoors (KITTI-360) benchmarks. We evaluated zero-shot monocular depth estimation by incorporating trained Calibration Tokens into recent foundational monocular depth estimators models. Note: Our method uses the same training set for both the indoor and outdoor settings; whereas existing methods train separate models for each setting." + }, + { + "type": "table", + "bbox": [ + 0.142, + 0.136, + 0.856, + 0.459 + ], + "angle": 0, + "content": "
TestsetExperimentModelTrain DatasetRMSE ↓δ1 ↑
ScanNet++ [76]BaselineMiDaS [49]Mix 1.4M0.5060.563
+ Calibration TokensMiDaS [49]+ Mix 200K0.4460.569
BaselineDepthAnything [73]Mix 63.5M0.7310.463
+ Calibration TokensDepthAnything [73]+ Mix 200K0.6070.506
BaselineUniDepth [45]Mix 16M0.2790.720
+ Calibration TokensUniDepth [45]+ Mix 200K0.2440.766
ComparisonsDepthAnyCamera [24]Indoor 670K0.2750.761
DepthAnyCamera [24]Mix 200K0.7610.255
FoVA-Depth [36]Indoor 190K0.2850.548
KITTI-360 [35]BaselineMiDaS [49]Mix 1.4M3.3120.586
+ Calibration TokensMiDaS[49]+ Mix 200K2.3480.658
BaselineDepthAnything [73]Mix 63.5M2.2140.839
+ Calibration TokensDepthAnything [73]+ Mix 200K2.0430.810
BaselineUniDepth [45]Mix 16M2.0850.663
+ Calibration TokensUniDepth [45]+ Mix 200K2.0400.664
ComparisonsDepthAnyCamera [24]Outdoor 130K2.0670.852
DepthAnyCamera [24]Mix 200K5.6750.348
FoVA-Depth [36]Outdoor 80K3.0960.632
" + }, + { + "type": "image", + "bbox": [ + 0.133, + 0.472, + 0.439, + 0.656 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.657, + 0.483, + 0.698 + ], + "angle": 0, + "content": "Figure 5. t-SNE plot of fisheye and perspective embeddings. Fisheye embeddings become closer to those of perspective images after being modulated by Calibration Tokens." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.715, + 0.484, + 0.837 + ], + "angle": 0, + "content": "indoor and outdoor training sets, suggesting the generalization potential of our method to wide ranges of fisheye distortions. Also, the KITTI-360 ground truth points are significantly sparser and more concentrated in ground regions as compared to ScanNet++, which may explain the discrepancy in evaluation metrics. Nonetheless, our method performs comparably without needing to train specialized sets of Calibration Tokens for different fisheye models." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.848, + 0.194, + 0.864 + ], + "angle": 0, + "content": "4.2. Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.484, + 0.903 + ], + "angle": 0, + "content": "Feature Modulation. To visualize how our Calibration Tokens affect fisheye embeddings, Fig. 5 shows a two-" + }, + { + "type": "image", + "bbox": [ + 0.522, + 0.472, + 0.907, + 0.613 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.616, + 0.907, + 0.715 + ], + "angle": 0, + "content": "Figure 6. Comparison of attention maps for single and multi-layer tokens. We visualize the attention map of Calibration Tokens to the encoded patch embeddings. Calibration Tokens attend highly to distorted border regions: \"Single Token\" (top) has little effect in most layers due to lower attention as only a set of tokens are append to the input. The proposed multi-layer or \"Layer-wise Tokens\" scheme (bottom) attends to embeddings across all layers." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.734, + 0.906, + 0.795 + ], + "angle": 0, + "content": "dimensional tSNE reduction to both fisheye and perspective image embeddings from the same set of images. After adding Calibration Tokens to the fisheye embeddings, they are modulated towards the perspective image distribution." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.796, + 0.907, + 0.903 + ], + "angle": 0, + "content": "Layer-wise Tokens. As observed in Fig. 6, when we append only a single set of tokens (\"Single Token\") at the initial transformer block of the pre-trained model, the Calibration Tokens exhibit limited attention to the patch embeddings across most layers. As a result, the patch embeddings of most layers are unchanged. However, when we attach unique tokens at every layer (\"Layer-wise Tokens\"), we see" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5204" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.089, + 0.483, + 0.131 + ], + "angle": 0, + "content": "Table 2. Comparison results with finetuning. We conducted experiments comparing finetuning (F.T.) with the use of Calibration Tokens (C.T.) added to the baseline model." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.379 + ], + "angle": 0, + "content": "
DatasetsModelsExp.RMSEδ1
ScanNet++MiDASF.T.2.1780.129
C.T.0.4460.569
DepthAnythingF.T.1.4590.462
C.T.0.6070.506
UniDepthF.T.0.4320.574
C.T.0.2440.766
KITTI-360MiDASF.T.9.2890.098
C.T.2.3480.658
DepthAnythingF.T.4.3620.636
C.T.2.0430.810
UniDepthF.T.3.2170.403
C.T.2.0400.664
" + }, + { + "type": "table_caption", + "bbox": [ + 0.092, + 0.387, + 0.483, + 0.43 + ], + "angle": 0, + "content": "Table 3. Analysis on computational cost. We analyze the computational overhead introduced by Calibration Tokens. Values in parentheses indicate the relative increase as a percentage." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.434, + 0.48, + 0.52 + ], + "angle": 0, + "content": "
ModelsModel memoryTokens memoryInference time
MiDAS1.7G0.8M(0.05%)0.6ms(0.8%)
DepthAnything1.7G0.8M(0.05%)0.8ms(0.8%)
UniDepth0.7G0.2M(0.02%)0.4ms(0.7%)
" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.535, + 0.483, + 0.581 + ], + "angle": 0, + "content": "higher attention at more layers. Thus, we opt to use the \"Layer-wise\" approach to better modulate the fisheye patch embeddings toward the distribution of perspective images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.581, + 0.483, + 0.717 + ], + "angle": 0, + "content": "Comparison with Finetuning. To further analyze the robustness of the Calibration Tokens, we conducted experiments comparing our method with a finetuning approach. We trained the model with a fixed learning rate of \\(10^{-6}\\) on our synthetic fisheye dataset for the same number of iterations. As shown in Tab. 2, the finetuning approach leads to a significant performance drop, highlighting the importance of using Calibration Tokens, which preserve the original model's training on perspective images." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.717, + 0.483, + 0.822 + ], + "angle": 0, + "content": "Computational Cost. Tab. 3 shows the impact of Calibration Tokens on computational costs across different FMDEs. Incorporating Calibration Tokens results in only a \\(0.05\\%\\) and \\(0.02\\%\\) increase in memory usage, less than 1 MB and a \\(0.8\\%\\) and \\(0.7\\%\\) increase in inference time, with an added latency of less than 1 ms. This analysis highlights the efficiency of our proposed Calibration Tokens." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.833, + 0.245, + 0.85 + ], + "angle": 0, + "content": "4.3. Ablation Study" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.855, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We conducted an ablation study on different Calibration Token configurations to validate our contributions. Note that \"Single Token\" refers to a single set of Calibration Tokens" + }, + { + "type": "table_caption", + "bbox": [ + 0.513, + 0.089, + 0.907, + 0.118 + ], + "angle": 0, + "content": "Table 4. Ablation study. We ablate the training objective and modulate scheme for our proposed Calibration Tokens." + }, + { + "type": "table", + "bbox": [ + 0.521, + 0.122, + 0.897, + 0.251 + ], + "angle": 0, + "content": "
DatasetMethodRMSEδ1
ScanNet++Single token0.2600.741
+ LogL1 Loss0.2540.752
+ Layer-wise Tokens0.2440.766
KITTI-360Single token2.0850.656
+ LogL1 Loss2.0650.665
+ Layer-wise Tokens2.0400.664
" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.267, + 0.905, + 0.342 + ], + "angle": 0, + "content": "appended in the first layer of the vision transformer without removal, with L1 loss applied. In this configuration, \\(\\phi \\in \\mathbb{R}^{M\\times F}\\) as opposed to \\(\\Phi \\in \\mathbb{R}^{L\\times M\\times F}\\) in the layer-wise setting. The ablation study on the ScanNet++ and KITTI-360 datasets is performed using the UniDepth model." + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.343, + 0.905, + 0.419 + ], + "angle": 0, + "content": "LogL1 Loss. We observed stable improvements with LogL1 loss compared to baseline L1 loss. As shown in Tab. 4, the LogL1 loss improves both metrics across indoor and outdoor datasets. Qualitative comparisons between L1 and LogL1 objectives are shown in the Supp. Mat." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.42, + 0.906, + 0.556 + ], + "angle": 0, + "content": "Layer-wise Tokens. Tab. 4 demonstrates the advantages of using layer-wise tokens over a single set of Calibration Tokens in the first layer. Even when the same number of tokens is fed to the visual transformer layers, we observed a significant improvement in the contribution of layer-wise tokens. This supports our hypothesis about how the influence of Calibration Tokens diminishes through a forward pass as observed in our experiments by appending a single set of tokens at the first layer. Fig. 6 visualizes attention." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.572, + 0.628, + 0.587 + ], + "angle": 0, + "content": "5. Discussion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.598, + 0.907, + 0.779 + ], + "angle": 0, + "content": "Calibration Tokens enable FMEs to adapt to images captured by fisheye cameras. Empirically, our method improves on monocular depth estimation on fisheye cameras. While our method trains only one set of tokens for both indoor and outdoor settings, our promising results motivate this as a general approach to adapting vision foundational models. Furthermore, a convenience afforded by our method is in the reuse and backward-compatibility of FMDEs with perspective images. This reduces the operational overhead of multi-camera systems by enabling a single FMDE to handle multiple camera inputs - adding cameras become as easy as appending tokens." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.78, + 0.909, + 0.901 + ], + "angle": 0, + "content": "Limitations. While we offer a light-weight method of extending FMDEs to fisheye images, its success inherently depends on the quality and representational power of the underlying FMDEs. If the pretrained model struggles with certain 3D scenes or lighting conditions for perspective images, then these issues carry over. Nonetheless, as novel FMDEs emerge, our framework can be readily transferred to new models using transformer-based architectures." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5205" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.486, + 0.138 + ], + "angle": 0, + "content": "Acknowledgments This work is supported by NSF 2112562 Athena AI Institute and the Rosenfeld Science Scholars Program." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.152, + 0.188, + 0.168 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.177, + 0.486, + 0.247 + ], + "angle": 0, + "content": "[1] Bruno Arsenali, Prashanth Viswanath, and Jelena Novosel. Rotinvmtl: Rotation invariant multinet on fisheye images for autonomous driving applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.249, + 0.484, + 0.304 + ], + "angle": 0, + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.307, + 0.483, + 0.375 + ], + "angle": 0, + "content": "[3] Aleksei Bochkovskii, AmaAgl Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In International Conference on Learning Representations, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.377, + 0.483, + 0.417 + ], + "angle": 0, + "content": "[4] Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer. arXiv preprint arXiv:2006.11527, 2020. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.42, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[5] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Transformer-based monocular depth estimation with attention supervision. In 32nd British Machine Vision Conference (BMVC 2021), 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.477, + 0.483, + 0.532 + ], + "angle": 0, + "content": "[6] Xien Chen, Suchisrit Gangopadhyay, Michael Chu, Patrick Rim, Hyoungseob Park, and Alex Wong. Uncle: Unsupervised continual learning of depth completion. arXiv preprint arXiv:2410.18074, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.534, + 0.483, + 0.604 + ], + "angle": 0, + "content": "[7] Hyesong Choi, Hunsang Lee, Sunkyung Kim, Sunok Kim, Seungryong Kim, Kwanghoon Sohn, and Dongbo Min. Adaptive confidence thresholding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12808-12818, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.606, + 0.483, + 0.687 + ], + "angle": 0, + "content": "[8] Younjoon Chung, Hyoungseob Park, Patrick Rim, Xiaoran Zhang, Jihe He, Ziyao Zeng, Safa Cicek, Byung-Woo Hong, James S. Duncan, and Alex Wong. Eta: Energy-based test-time adaptation for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.69, + 0.483, + 0.732 + ], + "angle": 0, + "content": "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.734, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.804, + 0.483, + 0.9 + ], + "angle": 0, + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.177, + 0.486, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "[12] C Brown Duane. Close-range camera calibration. Photogramm. Eng, 37(8):855-866, 1971. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.123, + 0.905, + 0.19 + ], + "angle": 0, + "content": "[13] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.193, + 0.905, + 0.247 + ], + "angle": 0, + "content": "[14] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.25, + 0.905, + 0.331 + ], + "angle": 0, + "content": "[15] Vadim Ezhov, Hyoungseob Park, Zhaoyang Zhang, Rishi Upadhyay, Howard Zhang, Chethan Chinder Chandrappa, Achuta Kadambi, Yunhao Ba, Julie Dorsey, and Alex Wong. All-day depth completion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.904, + 0.375 + ], + "angle": 0, + "content": "[16] Xiaohan Fei, Alex Wong, and Stefano Soatto. Geosupervised visual depth prediction. IEEE Robotics and Automation Letters, 4(2):1661-1668, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.378, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[17] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.449, + 0.905, + 0.517 + ], + "angle": 0, + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.52, + 0.905, + 0.602 + ], + "angle": 0, + "content": "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.604, + 0.905, + 0.671 + ], + "angle": 0, + "content": "[20] Clément Godard, Oisin Mac Aodha, and Gabriel J Brossow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.675, + 0.905, + 0.742 + ], + "angle": 0, + "content": "[21] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.746, + 0.905, + 0.801 + ], + "angle": 0, + "content": "[22] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.803, + 0.905, + 0.871 + ], + "angle": 0, + "content": "[23] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.874, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[24] Yuliang Guo, Sparsh Garg, S Mahdi H Miangoleh, Xinyu Huang, and Liu Ren. Depth any camera: Zero-shot met" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "5206" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.125, + 0.093, + 0.484, + 0.12 + ], + "angle": 0, + "content": "ric depth estimation from any camera. arXiv preprint arXiv:2501.02464, 2025.3,5,7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.122, + 0.484, + 0.178 + ], + "angle": 0, + "content": "[25] Juho Kannala and Sami Brandt. A generic camera calibration method for fish-eye lenses. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., pages 10-13. IEEE, 2004. 3, 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.179, + 0.482, + 0.261 + ], + "angle": 0, + "content": "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9492-9502, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.263, + 0.483, + 0.29 + ], + "angle": 0, + "content": "[27] R Kingslake. A history of the photographic lens. University of Rochester NY, 145, 1989. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.293, + 0.482, + 0.376 + ], + "angle": 0, + "content": "[28] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 61-71, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.378, + 0.482, + 0.446 + ], + "angle": 0, + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.449, + 0.482, + 0.504 + ], + "angle": 0, + "content": "[30] Dong Lao, Yangchao Wu, Tian Yu Liu, Alex Wong, and Stefano Soatto. Sub-token vit embedding via stochastic resonance transformers. In International Conference on Machine Learning. PMLR, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.505, + 0.482, + 0.573 + ], + "angle": 0, + "content": "[31] Dong Lao, Fengyu Yang, Daniel Wang, Hyoungseob Park, Samuel Lu, Alex Wong, and Stefano Soatto. On the viability of monocular depth pre-training for semantic segmentation. In European Conference on Computer Vision. Springer, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.576, + 0.482, + 0.658 + ], + "angle": 0, + "content": "[32] Bo Li, Chunhua Shen, Yuchao Dai, Anton Van Den Hengel, and Mingyi He. Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1119-1127, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.661, + 0.482, + 0.716 + ], + "angle": 0, + "content": "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.718, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[34] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology, 30(3):725-733, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.829 + ], + "angle": 0, + "content": "[35] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.832, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[36] Daniel Lichy, Hang Su, Abhishek Badki, Jan Kautz, and Orazio Gallo. Fova-depth: Field-of-view agnostic depth estimation for cross-dataset generalization. In 2024 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2024. 3, 6, 7" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.484, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.16 + ], + "angle": 0, + "content": "[37] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.164, + 0.905, + 0.245 + ], + "angle": 0, + "content": "[38] Tian Yu Liu, Parth Agrawal, Allison Chen, Byung-Woo Hong, and Alex Wong. Monitored distillation for positive congruent depth completion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 35-53. Springer, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.249, + 0.905, + 0.317 + ], + "angle": 0, + "content": "[39] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2301, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.32, + 0.907, + 0.387 + ], + "angle": 0, + "content": "[40] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5667-5675, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.39, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[41] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.42, + 0.907, + 0.474 + ], + "angle": 0, + "content": "[42] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20519-20529, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.476, + 0.907, + 0.545 + ], + "angle": 0, + "content": "[43] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.548, + 0.905, + 0.616 + ], + "angle": 0, + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.618, + 0.905, + 0.673 + ], + "angle": 0, + "content": "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.676, + 0.905, + 0.73 + ], + "angle": 0, + "content": "[46] Matteo Poggi, Fabio Tosi, and Stefano Mattoccia. Learning monocular depth estimation with unsupervised trinocular assumptions. In 2018 International conference on 3d vision (3DV), pages 324-333. IEEE, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.732, + 0.905, + 0.8 + ], + "angle": 0, + "content": "[47] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3227-3237, 2020. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.802, + 0.905, + 0.857 + ], + "angle": 0, + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.86, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[49] Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "5207" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.126, + 0.093, + 0.482, + 0.12 + ], + "angle": 0, + "content": "Dataset Transfer. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(03):1623-1637, 2022. 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.122, + 0.483, + 0.162 + ], + "angle": 0, + "content": "[50] Patrick Rim, Hyoungseob Park, Vadim Ezhov, Jeffrey Moon, and Alex Wong. Radar-guided polynomial fitting for metric depth estimation. arXiv preprint arXiv:2503.17182, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.163, + 0.483, + 0.232 + ], + "angle": 0, + "content": "[51] Patrick Rim, Hyoungseob Park, Ziyao Zeng, Younjoon Chung, and Alex Wong. Protodepth: Unsupervised continual depth completion with prototypes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6304-6316, 2025. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.233, + 0.483, + 0.315 + ], + "angle": 0, + "content": "[52] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.483, + 0.37 + ], + "angle": 0, + "content": "[53] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, 2012. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.483, + 0.455 + ], + "angle": 0, + "content": "[54] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.456, + 0.483, + 0.51 + ], + "angle": 0, + "content": "[55] Daniel E Stevenson and Margaret M Fleck. Nonparametric correction of distortion. In Proceedings Third IEEE Workshop on Applications of Computer Vision. WACV'96, pages 214-219. IEEE, 1996. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.483, + 0.594 + ], + "angle": 0, + "content": "[56] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.595, + 0.483, + 0.663 + ], + "angle": 0, + "content": "[57] Fabio Tosi, Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning monocular depth estimation infusing traditional stereo knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9799-9809, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.665, + 0.483, + 0.734 + ], + "angle": 0, + "content": "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.483, + 0.803 + ], + "angle": 0, + "content": "[59] Rishi Upadhyay, Howard Zhang, Yunhao Ba, Ethan Yang, Blake Gella, Sicheng Jiang, Alex Wong, and Achuta Kadambi. Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.804, + 0.483, + 0.871 + ], + "angle": 0, + "content": "[60] Chaoyang Wang, José Miguel Buenaposada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.873, + 0.483, + 0.901 + ], + "angle": 0, + "content": "[61] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.483, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.147 + ], + "angle": 0, + "content": "indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.15, + 0.905, + 0.207 + ], + "angle": 0, + "content": "[62] Jamie Watson, Michael Firman, Gabriel J Brostow, and Daniyar Turmukhambetov. Self-supervised monocular depth hints. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2162-2171, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.208, + 0.905, + 0.275 + ], + "angle": 0, + "content": "[63] Alex Wong and Stefano Soatto. Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5644-5653, 2019." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.278, + 0.905, + 0.332 + ], + "angle": 0, + "content": "[64] Alex Wong, Safa Cicek, and Stefano Soatto. Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems, 33:8486-8497, 2020." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.334, + 0.905, + 0.39 + ], + "angle": 0, + "content": "[65] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.392, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[66] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 6(2):1495-1502, 2021." + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.905, + 0.504 + ], + "angle": 0, + "content": "[67] Alex Wong, Xiaohan Fei, Byung-Woo Hong, and Stefano Soatto. An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters, 6 (2):3120-3127, 2021. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.507, + 0.905, + 0.575 + ], + "angle": 0, + "content": "[68] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision, pages 274-293. Springer, 2024. 3, 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.578, + 0.905, + 0.646 + ], + "angle": 0, + "content": "[69] Chao Xia, Chenfeng Xu, Patrick Rim, Mingyu Ding, Nanning Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Quadric representations for lidar odometry, mapping and localization. IEEE Robotics and Automation Letters, 8 (8):5023-5030, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.649, + 0.905, + 0.732 + ], + "angle": 0, + "content": "[70] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.734, + 0.907, + 0.802 + ], + "angle": 0, + "content": "[71] Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5354-5362, 2017. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.804, + 0.905, + 0.899 + ], + "angle": 0, + "content": "[72] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26340-26353, 2024. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.907, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5208" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.164, + 0.482, + 0.22 + ], + "angle": 0, + "content": "[74] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3353-3362, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.221, + 0.482, + 0.288 + ], + "angle": 0, + "content": "[75] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.292, + 0.482, + 0.348 + ], + "angle": 0, + "content": "[76] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.349, + 0.482, + 0.404 + ], + "angle": 0, + "content": "[77] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5684-5693, 2019. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.406, + 0.482, + 0.475 + ], + "angle": 0, + "content": "[78] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. Fisheyerecnet: A multi-context collaborative deep network for fisheye image rectification. In Proceedings of the European conference on computer vision (ECCV), pages 469-484, 2018. 4" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.478, + 0.482, + 0.559 + ], + "angle": 0, + "content": "[79] Senthil Yogamani, David Unger, Venkatraman Narayanan, and Varun Ravi Kumar. Fisheyebevseg: Surround view fisheye cameras based bird's-eye view segmentation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1331-1334, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.562, + 0.482, + 0.63 + ], + "angle": 0, + "content": "[80] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3916-3925, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.633, + 0.482, + 0.701 + ], + "angle": 0, + "content": "[81] Ziyao Zeng, Jingcheng Ni, Daniel Wang, Patrick Rim, Younjoon Chung, Fengyu Yang, Byung-Woo Hong, and Alex Wong. Priordiffusion: Leverage language prior in diffusion models for monocular depth estimation. arXiv preprint arXiv:2411.16750, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.704, + 0.482, + 0.773 + ], + "angle": 0, + "content": "[82] Ziyao Zeng, Daniel Wang, Fengyu Yang, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Wordepth: Variational language prior for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9708-9719, 2024." + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.775, + 0.482, + 0.856 + ], + "angle": 0, + "content": "[83] Ziyao Zeng, Yangchao Wu, Hyoungseob Park, Daniel Wang, Fengyu Yang, Stefano Soatto, Dong Lao, Byung-Woo Hong, and Alex Wong. Rsa: Resolving scale ambiguities in monocular depth estimators through language descriptions. Advances in neural information processing systems, 37, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.859, + 0.482, + 0.901 + ], + "angle": 0, + "content": "[84] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.545, + 0.093, + 0.905, + 0.133 + ], + "angle": 0, + "content": "deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.136, + 0.905, + 0.217 + ], + "angle": 0, + "content": "[85] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.22, + 0.905, + 0.302 + ], + "angle": 0, + "content": "[86] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 International Conference on 3D Vision (3DV), pages 668-678. IEEE, 2022. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.304, + 0.905, + 0.358 + ], + "angle": 0, + "content": "[87] Guoyang Zhao, Yuxuan Liu, Weiqing Qi, Fulong Ma, Ming Liu, and Jun Ma. Fisheyedgepth: A real scale self-supervised depth estimation model for fisheye camera. arXiv preprint arXiv:2409.15054, 2024. 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.361, + 0.905, + 0.416 + ], + "angle": 0, + "content": "[88] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.416 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "5209" + } + ] +] \ No newline at end of file diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_origin.pdf b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..16bbead0963db48ccf986e1f72a2b6923156751b --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/ff8cd9c3-7a4b-4914-bf82-9a51ee69f9ce_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96903f58b5603d3d47e50aa482a99a3fb8a1e7810f5b288d0182762c4264d108 +size 2587905 diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/full.md b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/full.md new file mode 100644 index 0000000000000000000000000000000000000000..8276df8a74bc7147bc467bc86b0b187d653932f3 --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/full.md @@ -0,0 +1,320 @@ +# Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens + +Suchisrit Gangopadhyay $^{1*}$ , Jung-Hee Kim $^{2*}$ , Xien Chen $^{1*}$ , Patrick Rim $^{1}$ , Hyoungseob Park $^{1}$ , Alex Wong $^{1}$ , Yale University $^{1}$ , Michigan State University $^{2}$ + +{rit.gangopadhyay,xien.chen,patrick.rim,hyoungseob.park,alex.wong}@yale.edu, kimjun84@msu.edu + +# Abstract + +We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: github.com/JungHeeKim29/calibration-token. + +# 1. Introduction + +Three-dimensional (3D) reconstruction is a fundamental component in many spatial applications, including autonomous vehicles, extended reality (XR), robotic manipulation. Each of these applications has unique demands for the field of view (FOV), often wider than the standard (perspective) camera. To meet this need, these applications tend to be deployed on systems equipped with fish + +![](images/2282c912a8566b27ce86728225f5bf2e37e6035ecabe4966b19728f050f9f6f7.jpg) +Input +Figure 1. Foundational monocular depth estimators fail on fisheye images. Despite being trained on large-scale datasets, foundational monocular depth estimators (FMDEs) models produce erroneous outputs. The inaccurate, blurry estimates are caused by a covariate shift that stem from fisheye distortion. + +![](images/804b813bf9fbcce9d675e2d349bdb96da517c4c862c4295f3cc474b8e2c3ef58.jpg) +MiDaS + +![](images/53b519b73404b54cfb261275430f67dcb413ef6b839dc536cb7098b940114650.jpg) +DepthAnything + +eye or other wide-angle cameras, which allows for wider coverage of the 3D environment. However, images captured by these cameras also come with substantial distortion, which arise from differences in projective geometry, where straight lines within the 3D environment or the 3D scene are preserved in perspective images but may appear curved in fisheye images. + +Foundational monocular depth estimators (FMDEs) [44, 48, 73] are trained on orders of tens of millions of images, enabling them to generalize across a wide range 3D scenes. However, their training data is comprised of internet images, which are predominantly captured using perspective cameras. Hence, despite being trained on large-scale datasets, FMDEs produce erroneous estimates when transferred to fisheye images (see Fig. 1). These errors stem from a covariate shift, which can be characterized by changes in camera calibration (intrinsic, distortion) parameters – leading to differences in object appearance and their perceived depth or distance from the camera. + +To address fisheye distortion, one solution is to recalibrate and undistort images or perform a map projection to some canonical reference frame. In principle, if one has the correct calibration, it is possible to re-project a fisheye image into a perspective-like view (or vice versa). In practice, however, there are several problems: (1) The calibration process itself can be error-prone and sensitive to physical perturbations in the camera system. Minor bumps, focus changes, or lens replacements can degrade or invalid + +idate previously computed intrinsic parameters. (2) Even when re-projection is performed accurately, the transformation introduces latency and spatial artifacts (e.g., stretching, cropping, aliasing, loss). When used as a preprocessing step for existing pretrained depth estimators, these artifacts still present a covariate shift and can degrade performance. + +Another solution is to train a separate monocular depth estimator specifically for fisheye images. However, publicly available image datasets for fisheye cameras are orders of tens to hundreds of times smaller than those for perspective cameras. Hence, it is difficult to assemble sufficient data to reach the large-scale training requirement of an FMDE. Nonetheless, one can adapt or finetune existing FMDEs for fisheye imagery. While this can improve performance on fisheye images, it introduces the risk of parameter drift, where the resulting FMDEs may lose their generalizability across 3D scenes. Moreover, the resulting finetuned model becomes specialized to fisheye cameras, limiting its applicability to other camera types, which adds operational overhead in applications involving mixed camera systems, e.g., autonomous vehicles or robotics. + +To address these challenges, we propose a novel approach termed Calibration Tokens. Our key insight is that existing FMDEs are already capable of estimating depth for perspective images, and that errors on fisheye images are caused by a covariate shift due to differences in camera calibration and distortion. Hence, rather than retraining or finetuning the entire model, we aim to "recalibrate" the fisheye latent embeddings such that they become more conducive to an FMDE originally trained on perspective images. Leveraging the fact that many FMDEs [44, 48, 49, 73] follow a Transformer-based architecture [11], we will exploit the (self- and cross-) attention mechanism to modulate the latent (token) embeddings by inserting Calibration Tokens as part of the input. Therefore, the existing FMDE will remain effectively unchanged, while Calibration Tokens serve to adapt their internal representations to mitigate the covariate shift by aligning the latent embeddings of fisheye images to the distribution of latent embeddings of perspective images. This design also allows us to preserve the original image content without performing any spatial re-projection, ensuring the process is lossless in terms of the raw pixels. Our hypothesis is that by adding a small set of trainable tokens to encode the fisheye camera calibration information and utilizing them to recalibrate the latent embeddings, we will be able to reuse existing FMDEs trained on perspective images and adapt them to fisheye images without sacrificing their generalizability across diverse 3D scenes. + +To train these Calibration Tokens, we propose a self-supervised objective that leverages inverse warping in the input and output spaces. FMDEs can infer high-fidelity depth maps for perspective images, so we use the perspective image depth estimates as our training target. We then + +![](images/ad025756a0561445ab4028e80b266e274c2565dbefc8b3f28f87d3d6944e7ba9.jpg) +Figure 2. Inference on different cameras. Calibration Tokens enable foundational monocular depth estimators to adapt to fisheye images while maintaining performance on perspective images. + +induce artificial distortion on perspective images to create pairs of perspective and synthetic fisheye images with diverse fisheye distortions. However, rather than doing the same in the output space, we undistort the fisheye depth maps to the original perspective reference frame to compute a self-supervised loss between the undistorted fisheye and perspective depth maps. By minimizing this self-supervised loss, the Calibration Tokens learn to align fisheye image embeddings to those of perspective images in the latent space, without any labels. Additionally, computing the loss in the original perspective frame allows our method to preserve the supervision signal instead of introducing artifacts. + +Our approach allows us to bypass the need to compile large-scale fisheye datasets by exploiting the abundance of perspective image datasets. As our method operates in the reference frame of the input, we also avoid transformation artifacts at inference time, whether in the input or output space. Furthermore, our method preserves compatibility with perspective images: One simply needs to append or remove Calibration Tokens for FMDEs to be applied to fisheye or perspective images. We demonstrate our method on indoors and outdoors across several recent FMDEs and consistently improve over baselines. + +Our contributions: (1) We propose a novel approach to extend foundational monocular depth estimators (FMDEs) trained on perspective images to fisheye images. (2) We introduce Calibration Tokens that modulate the latent embeddings of fisheye images towards the distribution of perspective image embeddings. (3) We introduce a self-supervised training objective that recalibrates input perspective image to fisheye images, but "undos" the transformation in the output to enable loss computation on high-fidelity (perspective) depth maps inferred by FMDEs. (4) Our approach only requires training one set of tokens to achieve state-of-the-art performance for both indoors and outdoors. + +# 2. Related Works + +Monocular Depth Estimation can be trained in a supervised or unsupervised manner. Supervised methods [13, 14, 29-32, 37, 71, 77] minimize the difference between depth estimates and ground-truth depth maps. [18] re-formulated the problem as ordinal regression while other methods proposed architectures innovations. [2] partitions depth ranges into adaptive bins. [5] incorporates an attention-based up-sample block. [33] employs hierarchical aggregation and heterogeneous interaction modules. [80] uses neural window fully-connected CRFs to compute energy. [59] synthesizes perspectively accurate images to enrich training data. Additional inputs e.g., language [81-83], lidar [6, 15, 51, 74], radar [50, 54], are used to enable metric-scale depth estimates. Unsupervised methods [7, 8, 23, 38, 42, 47, 57, 62-67, 84] minimizes photometric reconstruction error. [19] frames depth estimation as a novel view synthesis problem. [20] introduces a left-right consistency loss. [88] uses a pose network to enable unsupervised training on video sequences. [21] introduced auto-masking and min-reprojection loss. Additional loss terms based on visual odometry [16, 60], iterative closest point [40], surface normals [75], trinocular assumption [46], and semantic segmentation [22, 28] were also introduced. [39] redesigned the skip connection and decoders to extract high-resolution features, [86] combined global and local representations and [85] introduced a lightweight architecture with dilated convolution and attention. AugUndo [68] leveraged invertibility of transformation groups for data augmentation. + +Foundational Monocular Depth Estimators are trained with supervised or semi-supervised learning on large-scale datasets. MiDaS [49] is the first to demonstrate generalizable monocular depth estimation by compiling datasets for large-scale training. DPT [48] extended the approach and introduced transformers for dense predictions. Marigold [26] repurposes diffusion models for monocular depth estimation. DepthAnything [73] proposes a pseudo-labeling method to curate a large-scale dataset. Additionally, UniDepth [44] employs a camera self-prompting module and a pseudo-spherical output space, enabling metric-scale depth prediction across diverse 3D scenes without relying on external camera parameters. DepthPro [3] proposes a multi-scale vision transformer for metric-scale depth estimation. As all of these FMDEs are trained on perspective images, they fail to generalize to fisheye cameras. + +Fisheye Images. Images taken by a fisheye camera are distorted and unsuitable for use in a perspective image encoder. Existing distortion correction algorithms [12, 25] rely on different camera projection models [27, 41, 55] to undistort images into a perspective view. However, these methods depend on camera calibration parameters, which can introduce artifacts due to calibration inaccuracies. Recent approaches + +[24, 36] demonstrate training a separate model to perform depth estimation with different camera types. They utilize an equirectangular projection to project points from different reference frames to a canonical equirectangular frame, but this can incur transformation artifacts and distortions. Additionally, deep-learning-based methods [17, 34] that aim to rectify distortion have been introduced. However, these methods require a large number of parameters with limited accuracy and field of view. Consequently, many recent works targeted for fisheye images involve training an entire network [1, 79, 87] exclusively on fisheye images. Our method extends foundational monocular depth estimators to fisheye images instead. + +Token-Based Methods. Recent transformer-based architectures represent input images [11] (or other modalities [69, 70]) as sequences of tokens. In many cases, an additional token (e.g. [CLS] token in BERT [10] or the distillation token in DeiT [58]) is employed to aggregate information across all tokens. Such tokens can be adapted to various purposes, acting as a compact representation that "binds" or fuses information, e.g., [72] uses tokens to learn synthetic and real tactile response maps. Inspired by these advances, we introduce minimal trainable tokens appended to the fisheye embedding, enabling the model to "bind" or reconcile image distortions within a frozen backbone. Our approach is lightweight and requires no major architectural modifications, but extends foundational monocular depth estimators trained on perspective images to fisheye images. + +# 3. Method + +Let $I:\Omega \mapsto \mathbb{R}^3$ denote an RGB image obtained from a calibrated camera and $\Omega \subset \mathbb{R}^2$ the image space. Monocular depth estimation aims to learn a parameterized function $h_{\omega ,\psi}:\mathbb{R}^3\to \mathbb{R}_+$ that maps an image to a depth map $d:\Omega \mapsto \mathbb{R}_{+}$ . We assume access to a foundational monocular depth estimator (FMDE) pretrained on some large-scale dataset of perspective images. We will pair each image with (pseudo)ground truth $\tilde{d} = h_{\omega ,\psi}(I)$ inferred by the FMDE to obtain a training dataset $\mathcal{D} = \{(I^{(n)},\tilde{d}^{(n)})\}_{n = 1}^{N}$ . + +To extend FMDEs, trained on perspective images, to fisheye images, we introduce Calibration Tokens as an adaptation mechanism. Due to the prevalence of Transformer architectures in many FMDEs, we train a set of lightweight tokens to model the change in calibration between a perspective camera and different fisheye cameras. The goal of our Calibration Token is to recalibrate or translate latent embeddings of fisheye images back to those of perspective images. Our method takes advantage of the attention mechanism inherent in FMDEs and enable Calibration Tokens to modulate the latent embeddings, thus facilitating latent alignment. The outcome is an FMDE that is capable of inferring depth for fisheye images with Calibration Tokens, and perspective images without. + +![](images/5197d726d466f8831043af4a4243e6c876dc1adc5d75e3d73f923b858c7340f4.jpg) +Figure 3. Overview of our method. We introduce a set of trainable Calibration Tokens, which is appended to the input sequence of the fisheye image tokens. The Calibration Tokens are trained to adapt the model to produce accurate depth maps for images with various fisheye distortions. A unique fisheye calibration token is appended to the input of each new layer of the encoder. + +# 3.1. Extending FMDEs with Calibration Tokens + +Specifically, let $\phi \in \mathbb{R}^{M\times F}$ represent a set of Calibration Tokens, where $M$ denotes the number of tokens and $F$ their number of dimensions. For a given layer within the encoder $f_{\omega}$ of an FMDE, we will concatenate Calibration Tokens to the input sequence of patches or (embeddings) of the vision transformer: $f_{\omega}([I;\phi]) = [z^{(L)};f_{\omega}(\phi)]$ , where $z$ denotes the latent embeddings recalibrated by $\phi$ , $L$ the last layer, and $[\cdot ;]$ the concatenation operation. + +As each layer denotes a separate latent space, we extend our approach to a multi-layer modulation scheme. Let $\Phi \in \mathbb{R}^{L\times M\times F}$ be the set of Calibration Tokens for each layer $l\in \{1,\dots,L\}$ . A unique set of Calibration Tokens $\phi^{(l)}\in \mathbb{R}^{M\times F}$ is appended at each encoder layer: $f_{\omega}^{(l)}[(z^{(l - 1)};\phi^{(l)})] = [z^{(l)};f_{\omega}^{(l)}(\phi^{(l)})]$ for a layer $l$ . Each set of Calibration Tokens will modulate the patch embeddings for a specific layer through the attention mechanism; hence, following the convention in existing works [4, 9], we discard Calibration Tokens from the encoder output. A key insight is that the FMDE is already able to estimate high-fidelity depth maps for perspective images. We posit that the covariate shift exists in the encodings of fisheye image. Hence, once the latent embeddings of fisheye images have been recalibrated to those of perspective images, the decoder will be able to estimate depth to similar fidelity as perspective images. Therefore, we do not utilize Calibration Tokens to modulate the decoder layers. The final estimate is obtained by $\hat{d} = g_{\psi}(z^{(L)})$ , where $g_{\psi}$ denotes the decoder. + +Since our method does not apply spatial transformations during inference, it remains entirely lossless for input images. Additionally, it offers several efficiencies: (1) The + +only trainable parameters in our method are the light-weight Calibration Tokens, which consist of significantly fewer parameters than vision transformer models. Our method introduces minimal computational overhead and results in lower time and space complexity than training or finetuning a full model. (2) Our approach is backward-compatible with perspective images. By omitting our Calibration Tokens, an FMDE maintains its original depth estimation performance on perspective images. (3) At inference, camera intrinsics are not required, as the training process allows generalization across various fisheye camera intrinsics. As a result, our method eliminates the need for the arduous and error-prone calibration process after training. + +# 3.2. Learning Calibration Tokens + +To train Calibration Tokens, we will leverage the abundance of publicly available perspective image datasets. During our training, we synthesize fisheye images from perspective images by recalibrating them using artificial fisheye intrinsic and distortion parameters. This will produce pairs of perspective and fisheye images from which we can leverage self-supervision, and allows us to use a much larger training dataset than exclusively training with real fisheye images. We follow previous approaches [17, 78] to obtain synthetic fisheye images from the calibrated perspective images. Our synthesized fisheye images follow the distortion model introduced by Kannala & Brandt [25]: + +$$ +r (\theta) = k _ {1} \theta + k _ {2} \theta^ {3} + k _ {3} \theta^ {5} + k _ {4} \theta^ {7}, \tag {1} +$$ + +where $\theta$ denotes the angle between the ray and the optical axis, and $\{k_i\}_{i = 1}^4$ are distortion coefficients that can represent most of the real world fisheye distortion models. The + +change in coordinate between $(x,y)$ in the perspective image and $(x^{\prime},y^{\prime})$ in the fisheye image can be formulated as + +$$ +\begin{array}{l} x ^ {\prime} = r (\theta) \cos (\varphi), y ^ {\prime} = r (\theta) \sin (\varphi), \tag {2} \\ \varphi = \arctan ((y - c _ {x}) / (x - c _ {y})), \\ \end{array} +$$ + +where $(c_x, c_y)$ is the principal point in the perspective image. We define the transformation from the perspective to the fisheye reference frame as $T$ and its inverse transformation as $T^{-1}$ . Our training dataset is composed of perspective images and synthesized distorted image pairs with corresponding forward and inverse transformations. + +Loss Function. Inspired by AugUndo [68] and their use of invertible transformations to preserve the supervision signal by undoing data augmentation, we propose to synthesize fisheye images from the abundance of perspective images as inputs, but undistort the output to facilitate loss computation. By remapping depth estimates of synthetic fisheye images to the perspective frame, we enable the use of high-fidelity estimates inferred by FMDEs on perspective images as supervision. Hence, we can optimize Calibration Tokens with the following self-supervised loss: + +$$ +\arg \min _ {\Psi} \frac {1}{N} \sum_ {n = 1} ^ {N} \sum_ {x \in \Omega} \log \left(\left| \tilde {d} ^ {(n)} (x) - T ^ {- 1} \circ \hat {d} ^ {(n)} (x) \right| + 1\right), \tag {3} +$$ + +where $\tilde{d} = h_{\omega, \psi}(I)$ and $\hat{d} = h_{\omega, \psi}(T \circ I; \Phi)$ denotes the predicted depth map from the given perspective image and synthesized fisheye image, respectively. $\Phi$ denotes the proposed trainable Calibration Token appended to the patch embeddings. Calibration Tokens are trained to minimize the difference between the perspective output and the fisheye output re-projected into the perspective reference frame. Eq. (3) follows the log of absolute differences ( $\log L1$ ) proposed in [43], which enhances training stability and empirically outperforms L1 loss, especially in border regions where discrepancies between perspective and fisheye images are most significant (see Sec. 4.3 for details). + +It is important to note that attempting to instead transform the perspective depth maps outputted by FMDEs to the fisheye reference frame for the loss computation introduces information loss in the training objective. This will lead to re-projection artifacts in the supervision and result in learning inaccuracies during training. In Section A of the Supp. Mat., we further demonstrate comparison results between training in fisheye image space and perspective image space. Our training scheme is self-supervised and requires only calibrated perspective images, which can be easily obtained, making our approach both scalable and practical. + +# 4. Experiments + +Datasets. Training our Calibration Tokens requires only calibrated perspective images, enabling us to leverage significantly more data compared to training solely on fisheye + +images. Moreover, since our loss is computed based on comparisons with perspective image outputs, ground truth is not required for our training pipeline. + +Training datasets: NYUv2 [53] has a variety of perspective indoor scenes; VOID [65] contains indoor perspective office, classroom and stairwell scenes; IRS [61] contains rendered perspective scenes of home, restaurant, and store settings; Hypersim [52] comprises photorealistic, synthetic images of indoor residential and commercial environments in a perspective reference frame. Waymo [56] dataset consists of a diverse set of urban driving scenes. + +Test datasets: ScanNet++ [76] offers 3D reconstructions of diverse indoor scenes, captured using laser scanning and DSLR imaging with a fisheye lens, which allows us to evaluate with real fisheye images and ground truth depth maps. KITTI-360 [35] includes suburban driving scenes captured with a multi-sensor setup, including fisheye cameras. Notably, it features a different field of view compared to ScanNet++ [76], allowing us to assess the generalization capability of the Calibration Tokens. + +Models. Calibration Tokens do not require specific settings and can be integrated into any model utilizing a vision transformer. We evaluate the effectiveness of our approach using by extending MiDaS [49], DepthAnything [73], and UniDepth [45] to fisheye images. Note that we used 8 tokens per layer for each of the model experiments. + +Evaluation Metrics. We evaluate depth prediction accuracy using standard metrics from monocular depth estimation of root mean squared error (RMSE) and $\delta_{1}$ . Details on these metrics can be found in the Supp. Mat. + +Implementation Details. We trained our Calibration Tokens based on 3 different FMDEs (MiDas[49], DepthAnything[49], UniDepth[45]). We utilized the pretrained ViT-L backbone for MiDas [49] and DepthAnything [49], and the ViT-S backbone for UniDepth [45]. We trained our model on 4 NVIDIA 3090 GPUs for 40k iterations with a batch size of 16. For input, we used images in the resolution of $518 \times 518$ . For testing, we used $462 \times 616$ resolution on the ScanNet++ dataset [76], and $700 \times 700$ on the KITTI-360 dataset [35] to preserve its aspect ratio. We also synthesize random fisheye distortions in the training images. Our Calibration Tokens are trained with a joint dataset consisting of indoor and outdoor datasets totaling up to only 200K samples, and as shown in Tab. 1, obtain comparable results on both domains with fewer samples than existing methods [24] that are trained specifically for each. + +# 4.1. Main Result + +We conduct experiments to analyze the impact of Calibration Tokens on model performance. As a baseline, we compare our model to DepthAnyCamera [24], the state-of-the-art monocular depth estimation (MDE) method for fisheye images. Here, we evaluate the DepthAnyCamera model us + +![](images/2a1a8d350a4e1bbf69bf451f2a6d930a3517314311135404ace9feff9315652c.jpg) + +![](images/75e471bbad96869ca92b9930f57ca65cc4de1012673c81c1c817f009822aef74.jpg) + +![](images/ae2a4402312e4dc25abe2b4a7b6556c65361b0085f311ffa75c1896822ac0122.jpg) + +![](images/422ef3057227da8502a4891b9f65d9c235f97ba7228b6b93be73e3059ecf4a33.jpg) + +![](images/dfeae9c5e7edf804c7761b8b69d90cefa9e5711439075cdbf0cf3b5924d0d70e.jpg) + +![](images/b0b0b6e11d73c86503357c3d2d3b3a6db8080d345c5b5c32a961f28e6483615b.jpg) + +![](images/2b3f7a48ee2048cc6c906db186e9a7c5e67434fb7063e090aa3ecacedc4999fb.jpg) + +![](images/c34f49219ae3e9362b1626a5b4587a3c28cc171924b7b793e9bad63ad95ee183.jpg) + +![](images/fe7c3c5300dd2e5ceb79b29dc4c4f0cc82053d288d3baa899c79d4791bea9dfb.jpg) + +![](images/68e2eac299c0ef1e34659f2a6045307b0a6df232fdda33a9ab17a6819e1bba54.jpg) + +![](images/8adea137841a19e9c19410f14b3210afd3a95816f103e14160eb7c30a084059e.jpg) + +![](images/e570e0b6610a08e75690aedfe06d81a3d1f4e08a1ce7bc60eb7308d9a3f87eac.jpg) +Figure 4. Comparison on ScanNet++(Indoor) [76] and KITTI-360 [35] dataset. Qualitative comparison results on ScanNet++ and KITTI-360 datasets. Here, +C. T. indicates prediction results by appending Calibration Tokens to patch embeddings of the model located above. Calibration Tokens enable models to adapt to different fisheye cameras, especially in regions with large distortions. + +![](images/72ec0f8a27bc4a256edb7f916d4ee22ecc24cea759d66ebd9b2723c385fde619.jpg) + +![](images/aa0e006ec4d39685618949322c0684b01e48554f5e09c5a52f680af7f3bb4160.jpg) + +ing a ResNet101 backbone, and compare both its pretrained model and a model trained on our dataset for any fairness concerns. We also compare with FoVA-Depth [36], which is equirectangular projection based, like DepthAnyCamera. Indoor Evaluation. Among the pretrained foundational monocular depth estimators, UniDepth achieves the best performance with our Calibration Tokens on the ScanNet++ indoor dataset as shown in Tab. 1. Notably, our Calibration Tokens enable MiDAS to improve $12\%$ and DepthAnything to achieve a $17\%$ improvement in the RMSE metric compared to the model without Calibration Tokens. Similarly, UniDepth benefits from a $13\%$ improvement in the RMSE metric. Furthermore, compared to the comparison baselines, pretrained DepthAnyCamera and FoVA-Depth, UniDepth with Calibration Tokens surpasses their performance by $11\%$ and $14\%$ in the RMSE metric, respectively. DepthAnyCamera and FoVA-Depth utilize camera intrin + +sics for input images at test time, requiring image transformations back and forth from the equirectangular reference frame, which makes them more error-prone than our direct learning-based approach. + +Outdoor Evaluation. We evaluate FMDEs with our proposed Calibration Tokens against state-of-the-art methods in outdoor environments. The results show that Calibration Tokens consistently improve accuracy across different FMDEs in outdoor scenarios. Specifically, MiDaS and DepthAnything achieve improvement in the RMSE metric. UniDepth also improves $2\%$ in the RMSE metric, outperforming the comparison baselines. Given that the KITTI-360 dataset contains highly distorted images with a field of view exceeding 180 degrees, our Calibration Tokens demonstrate robustness across various distortion models. + +Our Calibration Tokens are able to outperform DepthAnyCamera and FoVA-Depth without separate + +Table 1. Quantitative comparisons on indoors (ScanNet++) and outdoors (KITTI-360) benchmarks. We evaluated zero-shot monocular depth estimation by incorporating trained Calibration Tokens into recent foundational monocular depth estimators models. Note: Our method uses the same training set for both the indoor and outdoor settings; whereas existing methods train separate models for each setting. + +
TestsetExperimentModelTrain DatasetRMSE ↓δ1 ↑
ScanNet++ [76]BaselineMiDaS [49]Mix 1.4M0.5060.563
+ Calibration TokensMiDaS [49]+ Mix 200K0.4460.569
BaselineDepthAnything [73]Mix 63.5M0.7310.463
+ Calibration TokensDepthAnything [73]+ Mix 200K0.6070.506
BaselineUniDepth [45]Mix 16M0.2790.720
+ Calibration TokensUniDepth [45]+ Mix 200K0.2440.766
ComparisonsDepthAnyCamera [24]Indoor 670K0.2750.761
DepthAnyCamera [24]Mix 200K0.7610.255
FoVA-Depth [36]Indoor 190K0.2850.548
KITTI-360 [35]BaselineMiDaS [49]Mix 1.4M3.3120.586
+ Calibration TokensMiDaS[49]+ Mix 200K2.3480.658
BaselineDepthAnything [73]Mix 63.5M2.2140.839
+ Calibration TokensDepthAnything [73]+ Mix 200K2.0430.810
BaselineUniDepth [45]Mix 16M2.0850.663
+ Calibration TokensUniDepth [45]+ Mix 200K2.0400.664
ComparisonsDepthAnyCamera [24]Outdoor 130K2.0670.852
DepthAnyCamera [24]Mix 200K5.6750.348
FoVA-Depth [36]Outdoor 80K3.0960.632
+ +![](images/c058ec082d15bf9c4f1865100451108c72e6e7452e3e17a9960933886b4dac5d.jpg) +Figure 5. t-SNE plot of fisheye and perspective embeddings. Fisheye embeddings become closer to those of perspective images after being modulated by Calibration Tokens. + +indoor and outdoor training sets, suggesting the generalization potential of our method to wide ranges of fisheye distortions. Also, the KITTI-360 ground truth points are significantly sparser and more concentrated in ground regions as compared to ScanNet++, which may explain the discrepancy in evaluation metrics. Nonetheless, our method performs comparably without needing to train specialized sets of Calibration Tokens for different fisheye models. + +# 4.2. Analysis + +Feature Modulation. To visualize how our Calibration Tokens affect fisheye embeddings, Fig. 5 shows a two- + +![](images/7843bff95324b3bd2dbd6be87ee4039d24e6309aeb6c579829603f949bc0236c.jpg) +Figure 6. Comparison of attention maps for single and multi-layer tokens. We visualize the attention map of Calibration Tokens to the encoded patch embeddings. Calibration Tokens attend highly to distorted border regions: "Single Token" (top) has little effect in most layers due to lower attention as only a set of tokens are append to the input. The proposed multi-layer or "Layer-wise Tokens" scheme (bottom) attends to embeddings across all layers. + +dimensional tSNE reduction to both fisheye and perspective image embeddings from the same set of images. After adding Calibration Tokens to the fisheye embeddings, they are modulated towards the perspective image distribution. + +Layer-wise Tokens. As observed in Fig. 6, when we append only a single set of tokens ("Single Token") at the initial transformer block of the pre-trained model, the Calibration Tokens exhibit limited attention to the patch embeddings across most layers. As a result, the patch embeddings of most layers are unchanged. However, when we attach unique tokens at every layer ("Layer-wise Tokens"), we see + +Table 2. Comparison results with finetuning. We conducted experiments comparing finetuning (F.T.) with the use of Calibration Tokens (C.T.) added to the baseline model. + +
DatasetsModelsExp.RMSEδ1
ScanNet++MiDASF.T.2.1780.129
C.T.0.4460.569
DepthAnythingF.T.1.4590.462
C.T.0.6070.506
UniDepthF.T.0.4320.574
C.T.0.2440.766
KITTI-360MiDASF.T.9.2890.098
C.T.2.3480.658
DepthAnythingF.T.4.3620.636
C.T.2.0430.810
UniDepthF.T.3.2170.403
C.T.2.0400.664
+ +Table 3. Analysis on computational cost. We analyze the computational overhead introduced by Calibration Tokens. Values in parentheses indicate the relative increase as a percentage. + +
ModelsModel memoryTokens memoryInference time
MiDAS1.7G0.8M(0.05%)0.6ms(0.8%)
DepthAnything1.7G0.8M(0.05%)0.8ms(0.8%)
UniDepth0.7G0.2M(0.02%)0.4ms(0.7%)
+ +higher attention at more layers. Thus, we opt to use the "Layer-wise" approach to better modulate the fisheye patch embeddings toward the distribution of perspective images. + +Comparison with Finetuning. To further analyze the robustness of the Calibration Tokens, we conducted experiments comparing our method with a finetuning approach. We trained the model with a fixed learning rate of $10^{-6}$ on our synthetic fisheye dataset for the same number of iterations. As shown in Tab. 2, the finetuning approach leads to a significant performance drop, highlighting the importance of using Calibration Tokens, which preserve the original model's training on perspective images. + +Computational Cost. Tab. 3 shows the impact of Calibration Tokens on computational costs across different FMDEs. Incorporating Calibration Tokens results in only a $0.05\%$ and $0.02\%$ increase in memory usage, less than 1 MB and a $0.8\%$ and $0.7\%$ increase in inference time, with an added latency of less than 1 ms. This analysis highlights the efficiency of our proposed Calibration Tokens. + +# 4.3. Ablation Study + +We conducted an ablation study on different Calibration Token configurations to validate our contributions. Note that "Single Token" refers to a single set of Calibration Tokens + +Table 4. Ablation study. We ablate the training objective and modulate scheme for our proposed Calibration Tokens. + +
DatasetMethodRMSEδ1
ScanNet++Single token0.2600.741
+ LogL1 Loss0.2540.752
+ Layer-wise Tokens0.2440.766
KITTI-360Single token2.0850.656
+ LogL1 Loss2.0650.665
+ Layer-wise Tokens2.0400.664
+ +appended in the first layer of the vision transformer without removal, with L1 loss applied. In this configuration, $\phi \in \mathbb{R}^{M\times F}$ as opposed to $\Phi \in \mathbb{R}^{L\times M\times F}$ in the layer-wise setting. The ablation study on the ScanNet++ and KITTI-360 datasets is performed using the UniDepth model. + +LogL1 Loss. We observed stable improvements with LogL1 loss compared to baseline L1 loss. As shown in Tab. 4, the LogL1 loss improves both metrics across indoor and outdoor datasets. Qualitative comparisons between L1 and LogL1 objectives are shown in the Supp. Mat. + +Layer-wise Tokens. Tab. 4 demonstrates the advantages of using layer-wise tokens over a single set of Calibration Tokens in the first layer. Even when the same number of tokens is fed to the visual transformer layers, we observed a significant improvement in the contribution of layer-wise tokens. This supports our hypothesis about how the influence of Calibration Tokens diminishes through a forward pass as observed in our experiments by appending a single set of tokens at the first layer. Fig. 6 visualizes attention. + +# 5. Discussion + +Calibration Tokens enable FMEs to adapt to images captured by fisheye cameras. Empirically, our method improves on monocular depth estimation on fisheye cameras. While our method trains only one set of tokens for both indoor and outdoor settings, our promising results motivate this as a general approach to adapting vision foundational models. Furthermore, a convenience afforded by our method is in the reuse and backward-compatibility of FMDEs with perspective images. This reduces the operational overhead of multi-camera systems by enabling a single FMDE to handle multiple camera inputs - adding cameras become as easy as appending tokens. + +Limitations. While we offer a light-weight method of extending FMDEs to fisheye images, its success inherently depends on the quality and representational power of the underlying FMDEs. If the pretrained model struggles with certain 3D scenes or lighting conditions for perspective images, then these issues carry over. Nonetheless, as novel FMDEs emerge, our framework can be readily transferred to new models using transformer-based architectures. + +Acknowledgments This work is supported by NSF 2112562 Athena AI Institute and the Rosenfeld Science Scholars Program. + +# References + +[1] Bruno Arsenali, Prashanth Viswanath, and Jelena Novosel. Rotinvmtl: Rotation invariant multinet on fisheye images for autonomous driving applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3 +[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3 +[3] Aleksei Bochkovskii, AmaAgl Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In International Conference on Learning Representations, 2025. 3 +[4] Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer. arXiv preprint arXiv:2006.11527, 2020. 4 +[5] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Transformer-based monocular depth estimation with attention supervision. In 32nd British Machine Vision Conference (BMVC 2021), 2021. 3 +[6] Xien Chen, Suchisrit Gangopadhyay, Michael Chu, Patrick Rim, Hyoungseob Park, and Alex Wong. Uncle: Unsupervised continual learning of depth completion. arXiv preprint arXiv:2410.18074, 2024. 3 +[7] Hyesong Choi, Hunsang Lee, Sunkyung Kim, Sunok Kim, Seungryong Kim, Kwanghoon Sohn, and Dongbo Min. Adaptive confidence thresholding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12808-12818, 2021. 3 +[8] Younjoon Chung, Hyoungseob Park, Patrick Rim, Xiaoran Zhang, Jihe He, Ziyao Zeng, Safa Cicek, Byung-Woo Hong, James S. Duncan, and Alex Wong. Eta: Energy-based test-time adaptation for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3 +[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4 +[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 3 +[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3 + +[12] C Brown Duane. Close-range camera calibration. Photogramm. Eng, 37(8):855-866, 1971. 3 +[13] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 3 +[14] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3 +[15] Vadim Ezhov, Hyoungseob Park, Zhaoyang Zhang, Rishi Upadhyay, Howard Zhang, Chethan Chinder Chandrappa, Achuta Kadambi, Yunhao Ba, Julie Dorsey, and Alex Wong. All-day depth completion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 3 +[16] Xiaohan Fei, Alex Wong, and Stefano Soatto. Geosupervised visual depth prediction. IEEE Robotics and Automation Letters, 4(2):1661-1668, 2019. 3 +[17] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 3, 4 +[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3 +[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 3 +[20] Clément Godard, Oisin Mac Aodha, and Gabriel J Brossow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3 +[21] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3 +[22] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 3 +[23] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 3 +[24] Yuliang Guo, Sparsh Garg, S Mahdi H Miangoleh, Xinyu Huang, and Liu Ren. Depth any camera: Zero-shot met + +ric depth estimation from any camera. arXiv preprint arXiv:2501.02464, 2025.3,5,7 +[25] Juho Kannala and Sami Brandt. A generic camera calibration method for fish-eye lenses. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., pages 10-13. IEEE, 2004. 3, 4 +[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9492-9502, 2024. 3 +[27] R Kingslake. A history of the photographic lens. University of Rochester NY, 145, 1989. 3 +[28] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 61-71, 2021. 3 +[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 3 +[30] Dong Lao, Yangchao Wu, Tian Yu Liu, Alex Wong, and Stefano Soatto. Sub-token vit embedding via stochastic resonance transformers. In International Conference on Machine Learning. PMLR, 2024. +[31] Dong Lao, Fengyu Yang, Daniel Wang, Hyoungseob Park, Samuel Lu, Alex Wong, and Stefano Soatto. On the viability of monocular depth pre-training for semantic segmentation. In European Conference on Computer Vision. Springer, 2024. +[32] Bo Li, Chunhua Shen, Yuchao Dai, Anton Van Den Hengel, and Mingyi He. Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1119-1127, 2015. 3 +[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 3 +[34] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology, 30(3):725-733, 2019. 3 +[35] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 5, 6, 7 +[36] Daniel Lichy, Hang Su, Abhishek Badki, Jan Kautz, and Orazio Gallo. Fova-depth: Field-of-view agnostic depth estimation for cross-dataset generalization. In 2024 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2024. 3, 6, 7 + +[37] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 3 +[38] Tian Yu Liu, Parth Agrawal, Allison Chen, Byung-Woo Hong, and Alex Wong. Monitored distillation for positive congruent depth completion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 35-53. Springer, 2022. 3 +[39] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2301, 2021. 3 +[40] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5667-5675, 2018. 3 +[41] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 3 +[42] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20519-20529, 2024. 3 +[43] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 5 +[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 3 +[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 5, 7 +[46] Matteo Poggi, Fabio Tosi, and Stefano Mattoccia. Learning monocular depth estimation with unsupervised trinocular assumptions. In 2018 International conference on 3d vision (3DV), pages 324-333. IEEE, 2018. 3 +[47] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3227-3237, 2020. 3 +[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3 +[49] Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross + +Dataset Transfer. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(03):1623-1637, 2022. 2, 3, 5, 7 +[50] Patrick Rim, Hyoungseob Park, Vadim Ezhov, Jeffrey Moon, and Alex Wong. Radar-guided polynomial fitting for metric depth estimation. arXiv preprint arXiv:2503.17182, 2025. 3 +[51] Patrick Rim, Hyoungseob Park, Ziyao Zeng, Younjoon Chung, and Alex Wong. Protodepth: Unsupervised continual depth completion with prototypes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6304-6316, 2025. 3 +[52] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5 +[53] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, 2012. 5 +[54] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 3 +[55] Daniel E Stevenson and Margaret M Fleck. Nonparametric correction of distortion. In Proceedings Third IEEE Workshop on Applications of Computer Vision. WACV'96, pages 214-219. IEEE, 1996. 3 +[56] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 5 +[57] Fabio Tosi, Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning monocular depth estimation infusing traditional stereo knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9799-9809, 2019. 3 +[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3 +[59] Rishi Upadhyay, Howard Zhang, Yunhao Ba, Ethan Yang, Blake Gella, Sicheng Jiang, Alex Wong, and Achuta Kadambi. Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 3 +[60] Chaoyang Wang, José Miguel Buenaposada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 3 +[61] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic + +indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5 +[62] Jamie Watson, Michael Firman, Gabriel J Brostow, and Daniyar Turmukhambetov. Self-supervised monocular depth hints. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2162-2171, 2019. 3 +[63] Alex Wong and Stefano Soatto. Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5644-5653, 2019. +[64] Alex Wong, Safa Cicek, and Stefano Soatto. Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems, 33:8486-8497, 2020. +[65] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 5 +[66] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 6(2):1495-1502, 2021. +[67] Alex Wong, Xiaohan Fei, Byung-Woo Hong, and Stefano Soatto. An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters, 6 (2):3120-3127, 2021. 3 +[68] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision, pages 274-293. Springer, 2024. 3, 5 +[69] Chao Xia, Chenfeng Xu, Patrick Rim, Mingyu Ding, Nanning Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Quadric representations for lidar odometry, mapping and localization. IEEE Robotics and Automation Letters, 8 (8):5023-5030, 2023. 3 +[70] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3 +[71] Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5354-5362, 2017. 3 +[72] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26340-26353, 2024. 3 + +[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 7 +[74] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3353-3362, 2019. 3 +[75] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3 +[76] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 5, 6, 7 +[77] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5684-5693, 2019. 3 +[78] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. Fisheyerecnet: A multi-context collaborative deep network for fisheye image rectification. In Proceedings of the European conference on computer vision (ECCV), pages 469-484, 2018. 4 +[79] Senthil Yogamani, David Unger, Venkatraman Narayanan, and Varun Ravi Kumar. Fisheyebevseg: Surround view fisheye cameras based bird's-eye view segmentation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1331-1334, 2024. 3 +[80] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3916-3925, 2022. 3 +[81] Ziyao Zeng, Jingcheng Ni, Daniel Wang, Patrick Rim, Younjoon Chung, Fengyu Yang, Byung-Woo Hong, and Alex Wong. Priordiffusion: Leverage language prior in diffusion models for monocular depth estimation. arXiv preprint arXiv:2411.16750, 2024. 3 +[82] Ziyao Zeng, Daniel Wang, Fengyu Yang, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Wordepth: Variational language prior for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9708-9719, 2024. +[83] Ziyao Zeng, Yangchao Wu, Hyoungseob Park, Daniel Wang, Fengyu Yang, Stefano Soatto, Dong Lao, Byung-Woo Hong, and Alex Wong. Rsa: Resolving scale ambiguities in monocular depth estimators through language descriptions. Advances in neural information processing systems, 37, 2024. 3 +[84] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with + +deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3 +[85] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 3 +[86] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 International Conference on 3D Vision (3DV), pages 668-678. IEEE, 2022. 3 +[87] Guoyang Zhao, Yuxuan Liu, Weiqing Qi, Fulong Ma, Ming Liu, and Jun Ma. Fisheyedgepth: A real scale self-supervised depth estimation model for fisheye camera. arXiv preprint arXiv:2409.15054, 2024. 3 +[88] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3 \ No newline at end of file diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/images.zip b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..478a0b886839e6acf299348179c094a7c1c969f1 --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:fed7b8b6bfa4bd1e4b4e4a812ea380046d969aa43d0035e64cb6cc9cd0c5c7e2 +size 537801 diff --git a/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/layout.json b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..6ad2a3ef7dc98ca250f4eb5e4ecb0e23fdca8c79 --- /dev/null +++ b/2025/Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens/layout.json @@ -0,0 +1,8934 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "spans": [ + { + "bbox": [ + 77, + 102, + 534, + 138 + ], + "type": "text", + "content": "Extending Foundational Monocular Depth Estimators to Fisheye Cameras with Calibration Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "spans": [ + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": "Suchisrit Gangopadhyay" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Jung-Hee Kim" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{2*}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Xien Chen" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1*}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Patrick Rim" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Hyoungseob Park" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Alex Wong" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Yale University" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "text", + "content": ", Michigan State University" + }, + { + "bbox": [ + 63, + 160, + 569, + 190 + ], + "type": "inline_equation", + "content": "^{2}" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 191, + 563, + 204 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 191, + 563, + 204 + ], + "spans": [ + { + "bbox": [ + 61, + 191, + 563, + 204 + ], + "type": "text", + "content": "{rit.gangopadhyay,xien.chen,patrick.rim,hyoungseob.park,alex.wong}@yale.edu, kimjun84@msu.edu" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "spans": [ + { + "bbox": [ + 151, + 231, + 200, + 243 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "spans": [ + { + "bbox": [ + 54, + 256, + 297, + 568 + ], + "type": "text", + "content": "We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: github.com/JungHeeKim29/calibration-token." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 590, + 135, + 604 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 590, + 135, + 604 + ], + "spans": [ + { + "bbox": [ + 55, + 590, + 135, + 604 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "spans": [ + { + "bbox": [ + 55, + 611, + 296, + 696 + ], + "type": "text", + "content": "Three-dimensional (3D) reconstruction is a fundamental component in many spatial applications, including autonomous vehicles, extended reality (XR), robotic manipulation. Each of these applications has unique demands for the field of view (FOV), often wider than the standard (perspective) camera. To meet this need, these applications tend to be deployed on systems equipped with fish" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 316, + 230, + 394, + 300 + ], + "blocks": [ + { + "bbox": [ + 316, + 230, + 394, + 300 + ], + "lines": [ + { + "bbox": [ + 316, + 230, + 394, + 300 + ], + "spans": [ + { + "bbox": [ + 316, + 230, + 394, + 300 + ], + "type": "image", + "image_path": "2282c912a8566b27ce86728225f5bf2e37e6035ecabe4966b19728f050f9f6f7.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 300, + 364, + 309 + ], + "lines": [ + { + "bbox": [ + 344, + 300, + 364, + 309 + ], + "spans": [ + { + "bbox": [ + 344, + 300, + 364, + 309 + ], + "type": "text", + "content": "Input" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 313, + 312, + 555, + 368 + ], + "lines": [ + { + "bbox": [ + 313, + 312, + 555, + 368 + ], + "spans": [ + { + "bbox": [ + 313, + 312, + 555, + 368 + ], + "type": "text", + "content": "Figure 1. Foundational monocular depth estimators fail on fisheye images. Despite being trained on large-scale datasets, foundational monocular depth estimators (FMDEs) models produce erroneous outputs. The inaccurate, blurry estimates are caused by a covariate shift that stem from fisheye distortion." + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 395, + 230, + 473, + 300 + ], + "blocks": [ + { + "bbox": [ + 395, + 230, + 473, + 300 + ], + "lines": [ + { + "bbox": [ + 395, + 230, + 473, + 300 + ], + "spans": [ + { + "bbox": [ + 395, + 230, + 473, + 300 + ], + "type": "image", + "image_path": "804b813bf9fbcce9d675e2d349bdb96da517c4c862c4295f3cc474b8e2c3ef58.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 421, + 300, + 447, + 308 + ], + "lines": [ + { + "bbox": [ + 421, + 300, + 447, + 308 + ], + "spans": [ + { + "bbox": [ + 421, + 300, + 447, + 308 + ], + "type": "text", + "content": "MiDaS" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 474, + 230, + 554, + 300 + ], + "blocks": [ + { + "bbox": [ + 474, + 230, + 554, + 300 + ], + "lines": [ + { + "bbox": [ + 474, + 230, + 554, + 300 + ], + "spans": [ + { + "bbox": [ + 474, + 230, + 554, + 300 + ], + "type": "image", + "image_path": "53b519b73404b54cfb261275430f67dcb413ef6b839dc536cb7098b940114650.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 486, + 300, + 542, + 309 + ], + "lines": [ + { + "bbox": [ + 486, + 300, + 542, + 309 + ], + "spans": [ + { + "bbox": [ + 486, + 300, + 542, + 309 + ], + "type": "text", + "content": "DepthAnything" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 378, + 555, + 461 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 378, + 555, + 461 + ], + "spans": [ + { + "bbox": [ + 313, + 378, + 555, + 461 + ], + "type": "text", + "content": "eye or other wide-angle cameras, which allows for wider coverage of the 3D environment. However, images captured by these cameras also come with substantial distortion, which arise from differences in projective geometry, where straight lines within the 3D environment or the 3D scene are preserved in perspective images but may appear curved in fisheye images." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 462, + 556, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 462, + 556, + 605 + ], + "spans": [ + { + "bbox": [ + 313, + 462, + 556, + 605 + ], + "type": "text", + "content": "Foundational monocular depth estimators (FMDEs) [44, 48, 73] are trained on orders of tens of millions of images, enabling them to generalize across a wide range 3D scenes. However, their training data is comprised of internet images, which are predominantly captured using perspective cameras. Hence, despite being trained on large-scale datasets, FMDEs produce erroneous estimates when transferred to fisheye images (see Fig. 1). These errors stem from a covariate shift, which can be characterized by changes in camera calibration (intrinsic, distortion) parameters – leading to differences in object appearance and their perceived depth or distance from the camera." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 605, + 556, + 715 + ], + "type": "text", + "content": "To address fisheye distortion, one solution is to recalibrate and undistort images or perform a map projection to some canonical reference frame. In principle, if one has the correct calibration, it is possible to re-project a fisheye image into a perspective-like view (or vice versa). In practice, however, there are several problems: (1) The calibration process itself can be error-prone and sensitive to physical perturbations in the camera system. Minor bumps, focus changes, or lens replacements can degrade or invalid" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "spans": [ + { + "bbox": [ + 67, + 703, + 133, + 712 + ], + "type": "text", + "content": "*Equal contribution" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5198" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 294, + 144 + ], + "type": "text", + "content": "idate previously computed intrinsic parameters. (2) Even when re-projection is performed accurately, the transformation introduces latency and spatial artifacts (e.g., stretching, cropping, aliasing, loss). When used as a preprocessing step for existing pretrained depth estimators, these artifacts still present a covariate shift and can degrade performance." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 147, + 295, + 326 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 147, + 295, + 326 + ], + "spans": [ + { + "bbox": [ + 55, + 147, + 295, + 326 + ], + "type": "text", + "content": "Another solution is to train a separate monocular depth estimator specifically for fisheye images. However, publicly available image datasets for fisheye cameras are orders of tens to hundreds of times smaller than those for perspective cameras. Hence, it is difficult to assemble sufficient data to reach the large-scale training requirement of an FMDE. Nonetheless, one can adapt or finetune existing FMDEs for fisheye imagery. While this can improve performance on fisheye images, it introduces the risk of parameter drift, where the resulting FMDEs may lose their generalizability across 3D scenes. Moreover, the resulting finetuned model becomes specialized to fisheye cameras, limiting its applicability to other camera types, which adds operational overhead in applications involving mixed camera systems, e.g., autonomous vehicles or robotics." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 328, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 328, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 328, + 295, + 651 + ], + "type": "text", + "content": "To address these challenges, we propose a novel approach termed Calibration Tokens. Our key insight is that existing FMDEs are already capable of estimating depth for perspective images, and that errors on fisheye images are caused by a covariate shift due to differences in camera calibration and distortion. Hence, rather than retraining or finetuning the entire model, we aim to \"recalibrate\" the fisheye latent embeddings such that they become more conducive to an FMDE originally trained on perspective images. Leveraging the fact that many FMDEs [44, 48, 49, 73] follow a Transformer-based architecture [11], we will exploit the (self- and cross-) attention mechanism to modulate the latent (token) embeddings by inserting Calibration Tokens as part of the input. Therefore, the existing FMDE will remain effectively unchanged, while Calibration Tokens serve to adapt their internal representations to mitigate the covariate shift by aligning the latent embeddings of fisheye images to the distribution of latent embeddings of perspective images. This design also allows us to preserve the original image content without performing any spatial re-projection, ensuring the process is lossless in terms of the raw pixels. Our hypothesis is that by adding a small set of trainable tokens to encode the fisheye camera calibration information and utilizing them to recalibrate the latent embeddings, we will be able to reuse existing FMDEs trained on perspective images and adapt them to fisheye images without sacrificing their generalizability across diverse 3D scenes." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 653, + 295, + 713 + ], + "type": "text", + "content": "To train these Calibration Tokens, we propose a self-supervised objective that leverages inverse warping in the input and output spaces. FMDEs can infer high-fidelity depth maps for perspective images, so we use the perspective image depth estimates as our training target. We then" + } + ] + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 318, + 71, + 555, + 240 + ], + "blocks": [ + { + "bbox": [ + 318, + 71, + 555, + 240 + ], + "lines": [ + { + "bbox": [ + 318, + 71, + 555, + 240 + ], + "spans": [ + { + "bbox": [ + 318, + 71, + 555, + 240 + ], + "type": "image", + "image_path": "ad025756a0561445ab4028e80b266e274c2565dbefc8b3f28f87d3d6944e7ba9.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "lines": [ + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "spans": [ + { + "bbox": [ + 313, + 247, + 555, + 281 + ], + "type": "text", + "content": "Figure 2. Inference on different cameras. Calibration Tokens enable foundational monocular depth estimators to adapt to fisheye images while maintaining performance on perspective images." + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 294, + 553, + 437 + ], + "type": "text", + "content": "induce artificial distortion on perspective images to create pairs of perspective and synthetic fisheye images with diverse fisheye distortions. However, rather than doing the same in the output space, we undistort the fisheye depth maps to the original perspective reference frame to compute a self-supervised loss between the undistorted fisheye and perspective depth maps. By minimizing this self-supervised loss, the Calibration Tokens learn to align fisheye image embeddings to those of perspective images in the latent space, without any labels. Additionally, computing the loss in the original perspective frame allows our method to preserve the supervision signal instead of introducing artifacts." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 554, + 568 + ], + "type": "text", + "content": "Our approach allows us to bypass the need to compile large-scale fisheye datasets by exploiting the abundance of perspective image datasets. As our method operates in the reference frame of the input, we also avoid transformation artifacts at inference time, whether in the input or output space. Furthermore, our method preserves compatibility with perspective images: One simply needs to append or remove Calibration Tokens for FMDEs to be applied to fisheye or perspective images. We demonstrate our method on indoors and outdoors across several recent FMDEs and consistently improve over baselines." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 554, + 713 + ], + "type": "text", + "content": "Our contributions: (1) We propose a novel approach to extend foundational monocular depth estimators (FMDEs) trained on perspective images to fisheye images. (2) We introduce Calibration Tokens that modulate the latent embeddings of fisheye images towards the distribution of perspective image embeddings. (3) We introduce a self-supervised training objective that recalibrates input perspective image to fisheye images, but \"undos\" the transformation in the output to enable loss computation on high-fidelity (perspective) depth maps inferred by FMDEs. (4) Our approach only requires training one set of tokens to achieve state-of-the-art performance for both indoors and outdoors." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5199" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "spans": [ + { + "bbox": [ + 56, + 71, + 147, + 83 + ], + "type": "text", + "content": "2. Related Works" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 91, + 294, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 91, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 57, + 91, + 294, + 437 + ], + "type": "text", + "content": "Monocular Depth Estimation can be trained in a supervised or unsupervised manner. Supervised methods [13, 14, 29-32, 37, 71, 77] minimize the difference between depth estimates and ground-truth depth maps. [18] re-formulated the problem as ordinal regression while other methods proposed architectures innovations. [2] partitions depth ranges into adaptive bins. [5] incorporates an attention-based up-sample block. [33] employs hierarchical aggregation and heterogeneous interaction modules. [80] uses neural window fully-connected CRFs to compute energy. [59] synthesizes perspectively accurate images to enrich training data. Additional inputs e.g., language [81-83], lidar [6, 15, 51, 74], radar [50, 54], are used to enable metric-scale depth estimates. Unsupervised methods [7, 8, 23, 38, 42, 47, 57, 62-67, 84] minimizes photometric reconstruction error. [19] frames depth estimation as a novel view synthesis problem. [20] introduces a left-right consistency loss. [88] uses a pose network to enable unsupervised training on video sequences. [21] introduced auto-masking and min-reprojection loss. Additional loss terms based on visual odometry [16, 60], iterative closest point [40], surface normals [75], trinocular assumption [46], and semantic segmentation [22, 28] were also introduced. [39] redesigned the skip connection and decoders to extract high-resolution features, [86] combined global and local representations and [85] introduced a lightweight architecture with dilated convolution and attention. AugUndo [68] leveraged invertibility of transformation groups for data augmentation." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 438, + 294, + 628 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 438, + 294, + 628 + ], + "spans": [ + { + "bbox": [ + 57, + 438, + 294, + 628 + ], + "type": "text", + "content": "Foundational Monocular Depth Estimators are trained with supervised or semi-supervised learning on large-scale datasets. MiDaS [49] is the first to demonstrate generalizable monocular depth estimation by compiling datasets for large-scale training. DPT [48] extended the approach and introduced transformers for dense predictions. Marigold [26] repurposes diffusion models for monocular depth estimation. DepthAnything [73] proposes a pseudo-labeling method to curate a large-scale dataset. Additionally, UniDepth [44] employs a camera self-prompting module and a pseudo-spherical output space, enabling metric-scale depth prediction across diverse 3D scenes without relying on external camera parameters. DepthPro [3] proposes a multi-scale vision transformer for metric-scale depth estimation. As all of these FMDEs are trained on perspective images, they fail to generalize to fisheye cameras." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 630, + 294, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 630, + 294, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 630, + 294, + 714 + ], + "type": "text", + "content": "Fisheye Images. Images taken by a fisheye camera are distorted and unsuitable for use in a perspective image encoder. Existing distortion correction algorithms [12, 25] rely on different camera projection models [27, 41, 55] to undistort images into a perspective view. However, these methods depend on camera calibration parameters, which can introduce artifacts due to calibration inaccuracies. Recent approaches" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 227 + ], + "type": "text", + "content": "[24, 36] demonstrate training a separate model to perform depth estimation with different camera types. They utilize an equirectangular projection to project points from different reference frames to a canonical equirectangular frame, but this can incur transformation artifacts and distortions. Additionally, deep-learning-based methods [17, 34] that aim to rectify distortion have been introduced. However, these methods require a large number of parameters with limited accuracy and field of view. Consequently, many recent works targeted for fisheye images involve training an entire network [1, 79, 87] exclusively on fisheye images. Our method extends foundational monocular depth estimators to fisheye images instead." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "spans": [ + { + "bbox": [ + 313, + 228, + 553, + 407 + ], + "type": "text", + "content": "Token-Based Methods. Recent transformer-based architectures represent input images [11] (or other modalities [69, 70]) as sequences of tokens. In many cases, an additional token (e.g. [CLS] token in BERT [10] or the distillation token in DeiT [58]) is employed to aggregate information across all tokens. Such tokens can be adapted to various purposes, acting as a compact representation that \"binds\" or fuses information, e.g., [72] uses tokens to learn synthetic and real tactile response maps. Inspired by these advances, we introduce minimal trainable tokens appended to the fisheye embedding, enabling the model to \"bind\" or reconcile image distortions within a frozen backbone. Our approach is lightweight and requires no major architectural modifications, but extends foundational monocular depth estimators trained on perspective images to fisheye images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 314, + 417, + 370, + 430 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 417, + 370, + 430 + ], + "spans": [ + { + "bbox": [ + 314, + 417, + 370, + 430 + ], + "type": "text", + "content": "3. Method" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "spans": [ + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": "Let " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "I:\\Omega \\mapsto \\mathbb{R}^3" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": " denote an RGB image obtained from a calibrated camera and " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\Omega \\subset \\mathbb{R}^2" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": " the image space. Monocular depth estimation aims to learn a parameterized function " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "h_{\\omega ,\\psi}:\\mathbb{R}^3\\to \\mathbb{R}_+" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": " that maps an image to a depth map " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "d:\\Omega \\mapsto \\mathbb{R}_{+}" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": ". We assume access to a foundational monocular depth estimator (FMDE) pretrained on some large-scale dataset of perspective images. We will pair each image with (pseudo)ground truth " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\tilde{d} = h_{\\omega ,\\psi}(I)" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": " inferred by the FMDE to obtain a training dataset " + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "inline_equation", + "content": "\\mathcal{D} = \\{(I^{(n)},\\tilde{d}^{(n)})\\}_{n = 1}^{N}" + }, + { + "bbox": [ + 313, + 437, + 553, + 545 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 553, + 713 + ], + "type": "text", + "content": "To extend FMDEs, trained on perspective images, to fisheye images, we introduce Calibration Tokens as an adaptation mechanism. Due to the prevalence of Transformer architectures in many FMDEs, we train a set of lightweight tokens to model the change in calibration between a perspective camera and different fisheye cameras. The goal of our Calibration Token is to recalibrate or translate latent embeddings of fisheye images back to those of perspective images. Our method takes advantage of the attention mechanism inherent in FMDEs and enable Calibration Tokens to modulate the latent embeddings, thus facilitating latent alignment. The outcome is an FMDE that is capable of inferring depth for fisheye images with Calibration Tokens, and perspective images without." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5200" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 74, + 553, + 270 + ], + "blocks": [ + { + "bbox": [ + 58, + 74, + 553, + 270 + ], + "lines": [ + { + "bbox": [ + 58, + 74, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 58, + 74, + 553, + 270 + ], + "type": "image", + "image_path": "5197d726d466f8831043af4a4243e6c876dc1adc5d75e3d73f923b858c7340f4.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 54, + 278, + 555, + 313 + ], + "lines": [ + { + "bbox": [ + 54, + 278, + 555, + 313 + ], + "spans": [ + { + "bbox": [ + 54, + 278, + 555, + 313 + ], + "type": "text", + "content": "Figure 3. Overview of our method. We introduce a set of trainable Calibration Tokens, which is appended to the input sequence of the fisheye image tokens. The Calibration Tokens are trained to adapt the model to produce accurate depth maps for images with various fisheye distortions. A unique fisheye calibration token is appended to the input of each new layer of the encoder." + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 324, + 282, + 337 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 324, + 282, + 337 + ], + "spans": [ + { + "bbox": [ + 55, + 324, + 282, + 337 + ], + "type": "text", + "content": "3.1. Extending FMDEs with Calibration Tokens" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "spans": [ + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": "Specifically, let " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " represent a set of Calibration Tokens, where " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "M" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " denotes the number of tokens and " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "F" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " their number of dimensions. For a given layer within the encoder " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "f_{\\omega}" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " of an FMDE, we will concatenate Calibration Tokens to the input sequence of patches or (embeddings) of the vision transformer: " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "f_{\\omega}([I;\\phi]) = [z^{(L)};f_{\\omega}(\\phi)]" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "z" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " denotes the latent embeddings recalibrated by " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "\\phi" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "L" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " the last layer, and " + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "inline_equation", + "content": "[\\cdot ;]" + }, + { + "bbox": [ + 55, + 344, + 296, + 441 + ], + "type": "text", + "content": " the concatenation operation." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "spans": [ + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": "As each layer denotes a separate latent space, we extend our approach to a multi-layer modulation scheme. Let " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{L\\times M\\times F}" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": " be the set of Calibration Tokens for each layer " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "l\\in \\{1,\\dots,L\\}" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": ". A unique set of Calibration Tokens " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\phi^{(l)}\\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": " is appended at each encoder layer: " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "f_{\\omega}^{(l)}[(z^{(l - 1)};\\phi^{(l)})] = [z^{(l)};f_{\\omega}^{(l)}(\\phi^{(l)})]" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": " for a layer " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "l" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": ". Each set of Calibration Tokens will modulate the patch embeddings for a specific layer through the attention mechanism; hence, following the convention in existing works [4, 9], we discard Calibration Tokens from the encoder output. A key insight is that the FMDE is already able to estimate high-fidelity depth maps for perspective images. We posit that the covariate shift exists in the encodings of fisheye image. Hence, once the latent embeddings of fisheye images have been recalibrated to those of perspective images, the decoder will be able to estimate depth to similar fidelity as perspective images. Therefore, we do not utilize Calibration Tokens to modulate the decoder layers. The final estimate is obtained by " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "\\hat{d} = g_{\\psi}(z^{(L)})" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "inline_equation", + "content": "g_{\\psi}" + }, + { + "bbox": [ + 55, + 445, + 296, + 674 + ], + "type": "text", + "content": " denotes the decoder." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 714 + ], + "type": "text", + "content": "Since our method does not apply spatial transformations during inference, it remains entirely lossless for input images. Additionally, it offers several efficiencies: (1) The" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 324, + 555, + 480 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 324, + 555, + 480 + ], + "spans": [ + { + "bbox": [ + 313, + 324, + 555, + 480 + ], + "type": "text", + "content": "only trainable parameters in our method are the light-weight Calibration Tokens, which consist of significantly fewer parameters than vision transformer models. Our method introduces minimal computational overhead and results in lower time and space complexity than training or finetuning a full model. (2) Our approach is backward-compatible with perspective images. By omitting our Calibration Tokens, an FMDE maintains its original depth estimation performance on perspective images. (3) At inference, camera intrinsics are not required, as the training process allows generalization across various fisheye camera intrinsics. As a result, our method eliminates the need for the arduous and error-prone calibration process after training." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 487, + 473, + 499 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 487, + 473, + 499 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 473, + 499 + ], + "type": "text", + "content": "3.2. Learning Calibration Tokens" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 504, + 555, + 648 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 504, + 555, + 648 + ], + "spans": [ + { + "bbox": [ + 313, + 504, + 555, + 648 + ], + "type": "text", + "content": "To train Calibration Tokens, we will leverage the abundance of publicly available perspective image datasets. During our training, we synthesize fisheye images from perspective images by recalibrating them using artificial fisheye intrinsic and distortion parameters. This will produce pairs of perspective and fisheye images from which we can leverage self-supervision, and allows us to use a much larger training dataset than exclusively training with real fisheye images. We follow previous approaches [17, 78] to obtain synthetic fisheye images from the calibrated perspective images. Our synthesized fisheye images follow the distortion model introduced by Kannala & Brandt [25]:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 361, + 655, + 553, + 670 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 361, + 655, + 553, + 670 + ], + "spans": [ + { + "bbox": [ + 361, + 655, + 553, + 670 + ], + "type": "interline_equation", + "content": "r (\\theta) = k _ {1} \\theta + k _ {2} \\theta^ {3} + k _ {3} \\theta^ {5} + k _ {4} \\theta^ {7}, \\tag {1}", + "image_path": "3fa077e06146c9cfb915b2e4a497e0303b1de4e918c6fa39d9bfac9b1afc39a8.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\theta" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " denotes the angle between the ray and the optical axis, and " + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\{k_i\\}_{i = 1}^4" + }, + { + "bbox": [ + 313, + 677, + 555, + 713 + ], + "type": "text", + "content": " are distortion coefficients that can represent most of the real world fisheye distortion models. The" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "5201" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "change in coordinate between " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "(x,y)" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in the perspective image and " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "(x^{\\prime},y^{\\prime})" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": " in the fisheye image can be formulated as" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 99, + 102, + 295, + 130 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 102, + 295, + 130 + ], + "spans": [ + { + "bbox": [ + 99, + 102, + 295, + 130 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} x ^ {\\prime} = r (\\theta) \\cos (\\varphi), y ^ {\\prime} = r (\\theta) \\sin (\\varphi), \\tag {2} \\\\ \\varphi = \\arctan ((y - c _ {x}) / (x - c _ {y})), \\\\ \\end{array}", + "image_path": "72a588d9949e0cb84d6474f5301a1464849459f4f3dc839c6affaef9ee77b419.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "spans": [ + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "inline_equation", + "content": "(c_x, c_y)" + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "text", + "content": " is the principal point in the perspective image. We define the transformation from the perspective to the fisheye reference frame as " + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "inline_equation", + "content": "T" + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "text", + "content": " and its inverse transformation as " + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "inline_equation", + "content": "T^{-1}" + }, + { + "bbox": [ + 55, + 136, + 295, + 207 + ], + "type": "text", + "content": ". Our training dataset is composed of perspective images and synthesized distorted image pairs with corresponding forward and inverse transformations." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 208, + 296, + 327 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 208, + 296, + 327 + ], + "spans": [ + { + "bbox": [ + 55, + 208, + 296, + 327 + ], + "type": "text", + "content": "Loss Function. Inspired by AugUndo [68] and their use of invertible transformations to preserve the supervision signal by undoing data augmentation, we propose to synthesize fisheye images from the abundance of perspective images as inputs, but undistort the output to facilitate loss computation. By remapping depth estimates of synthetic fisheye images to the perspective frame, we enable the use of high-fidelity estimates inferred by FMDEs on perspective images as supervision. Hence, we can optimize Calibration Tokens with the following self-supervised loss:" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 60, + 332, + 295, + 365 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 60, + 332, + 295, + 365 + ], + "spans": [ + { + "bbox": [ + 60, + 332, + 295, + 365 + ], + "type": "interline_equation", + "content": "\\arg \\min _ {\\Psi} \\frac {1}{N} \\sum_ {n = 1} ^ {N} \\sum_ {x \\in \\Omega} \\log \\left(\\left| \\tilde {d} ^ {(n)} (x) - T ^ {- 1} \\circ \\hat {d} ^ {(n)} (x) \\right| + 1\\right), \\tag {3}", + "image_path": "643c6e89a795a44c068e5d74d837061fa34dbe8072f3649d39adc0ed82812cb3.jpg" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "spans": [ + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\tilde{d} = h_{\\omega, \\psi}(I)" + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\hat{d} = h_{\\omega, \\psi}(T \\circ I; \\Phi)" + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "content": " denotes the predicted depth map from the given perspective image and synthesized fisheye image, respectively. " + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\Phi" + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "content": " denotes the proposed trainable Calibration Token appended to the patch embeddings. Calibration Tokens are trained to minimize the difference between the perspective output and the fisheye output re-projected into the perspective reference frame. Eq. (3) follows the log of absolute differences (" + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "inline_equation", + "content": "\\log L1" + }, + { + "bbox": [ + 55, + 372, + 295, + 516 + ], + "type": "text", + "content": ") proposed in [43], which enhances training stability and empirically outperforms L1 loss, especially in border regions where discrepancies between perspective and fisheye images are most significant (see Sec. 4.3 for details)." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "spans": [ + { + "bbox": [ + 55, + 517, + 295, + 649 + ], + "type": "text", + "content": "It is important to note that attempting to instead transform the perspective depth maps outputted by FMDEs to the fisheye reference frame for the loss computation introduces information loss in the training objective. This will lead to re-projection artifacts in the supervision and result in learning inaccuracies during training. In Section A of the Supp. Mat., we further demonstrate comparison results between training in fisheye image space and perspective image space. Our training scheme is self-supervised and requires only calibrated perspective images, which can be easily obtained, making our approach both scalable and practical." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 658, + 136, + 671 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 658, + 136, + 671 + ], + "spans": [ + { + "bbox": [ + 55, + 658, + 136, + 671 + ], + "type": "text", + "content": "4. Experiments" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 715 + ], + "type": "text", + "content": "Datasets. Training our Calibration Tokens requires only calibrated perspective images, enabling us to leverage significantly more data compared to training solely on fisheye" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 108 + ], + "type": "text", + "content": "images. Moreover, since our loss is computed based on comparisons with perspective image outputs, ground truth is not required for our training pipeline." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "spans": [ + { + "bbox": [ + 313, + 109, + 553, + 205 + ], + "type": "text", + "content": "Training datasets: NYUv2 [53] has a variety of perspective indoor scenes; VOID [65] contains indoor perspective office, classroom and stairwell scenes; IRS [61] contains rendered perspective scenes of home, restaurant, and store settings; Hypersim [52] comprises photorealistic, synthetic images of indoor residential and commercial environments in a perspective reference frame. Waymo [56] dataset consists of a diverse set of urban driving scenes." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "spans": [ + { + "bbox": [ + 313, + 205, + 553, + 312 + ], + "type": "text", + "content": "Test datasets: ScanNet++ [76] offers 3D reconstructions of diverse indoor scenes, captured using laser scanning and DSLR imaging with a fisheye lens, which allows us to evaluate with real fisheye images and ground truth depth maps. KITTI-360 [35] includes suburban driving scenes captured with a multi-sensor setup, including fisheye cameras. Notably, it features a different field of view compared to ScanNet++ [76], allowing us to assess the generalization capability of the Calibration Tokens." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 314, + 553, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 314, + 553, + 385 + ], + "spans": [ + { + "bbox": [ + 313, + 314, + 553, + 385 + ], + "type": "text", + "content": "Models. Calibration Tokens do not require specific settings and can be integrated into any model utilizing a vision transformer. We evaluate the effectiveness of our approach using by extending MiDaS [49], DepthAnything [73], and UniDepth [45] to fisheye images. Note that we used 8 tokens per layer for each of the model experiments." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 386, + 553, + 434 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 386, + 553, + 434 + ], + "spans": [ + { + "bbox": [ + 313, + 386, + 553, + 434 + ], + "type": "text", + "content": "Evaluation Metrics. We evaluate depth prediction accuracy using standard metrics from monocular depth estimation of root mean squared error (RMSE) and " + }, + { + "bbox": [ + 313, + 386, + 553, + 434 + ], + "type": "inline_equation", + "content": "\\delta_{1}" + }, + { + "bbox": [ + 313, + 386, + 553, + 434 + ], + "type": "text", + "content": ". Details on these metrics can be found in the Supp. Mat." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "spans": [ + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "text", + "content": "Implementation Details. We trained our Calibration Tokens based on 3 different FMDEs (MiDas[49], DepthAnything[49], UniDepth[45]). We utilized the pretrained ViT-L backbone for MiDas [49] and DepthAnything [49], and the ViT-S backbone for UniDepth [45]. We trained our model on 4 NVIDIA 3090 GPUs for 40k iterations with a batch size of 16. For input, we used images in the resolution of " + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "inline_equation", + "content": "518 \\times 518" + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "text", + "content": ". For testing, we used " + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "inline_equation", + "content": "462 \\times 616" + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "text", + "content": " resolution on the ScanNet++ dataset [76], and " + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "inline_equation", + "content": "700 \\times 700" + }, + { + "bbox": [ + 313, + 434, + 554, + 626 + ], + "type": "text", + "content": " on the KITTI-360 dataset [35] to preserve its aspect ratio. We also synthesize random fisheye distortions in the training images. Our Calibration Tokens are trained with a joint dataset consisting of indoor and outdoor datasets totaling up to only 200K samples, and as shown in Tab. 1, obtain comparable results on both domains with fewer samples than existing methods [24] that are trained specifically for each." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 635, + 394, + 647 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 635, + 394, + 647 + ], + "spans": [ + { + "bbox": [ + 313, + 635, + 394, + 647 + ], + "type": "text", + "content": "4.1. Main Result" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 553, + 713 + ], + "type": "text", + "content": "We conduct experiments to analyze the impact of Calibration Tokens on model performance. As a baseline, we compare our model to DepthAnyCamera [24], the state-of-the-art monocular depth estimation (MDE) method for fisheye images. Here, we evaluate the DepthAnyCamera model us" + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5202" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 79, + 70, + 194, + 156 + ], + "blocks": [ + { + "bbox": [ + 79, + 70, + 194, + 156 + ], + "lines": [ + { + "bbox": [ + 79, + 70, + 194, + 156 + ], + "spans": [ + { + "bbox": [ + 79, + 70, + 194, + 156 + ], + "type": "image", + "image_path": "2a1a8d350a4e1bbf69bf451f2a6d930a3517314311135404ace9feff9315652c.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 195, + 72, + 304, + 155 + ], + "blocks": [ + { + "bbox": [ + 195, + 72, + 304, + 155 + ], + "lines": [ + { + "bbox": [ + 195, + 72, + 304, + 155 + ], + "spans": [ + { + "bbox": [ + 195, + 72, + 304, + 155 + ], + "type": "image", + "image_path": "75e471bbad96869ca92b9930f57ca65cc4de1012673c81c1c817f009822aef74.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 306, + 72, + 416, + 155 + ], + "blocks": [ + { + "bbox": [ + 306, + 72, + 416, + 155 + ], + "lines": [ + { + "bbox": [ + 306, + 72, + 416, + 155 + ], + "spans": [ + { + "bbox": [ + 306, + 72, + 416, + 155 + ], + "type": "image", + "image_path": "ae2a4402312e4dc25abe2b4a7b6556c65361b0085f311ffa75c1896822ac0122.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 416, + 113, + 528, + 199 + ], + "blocks": [ + { + "bbox": [ + 416, + 113, + 528, + 199 + ], + "lines": [ + { + "bbox": [ + 416, + 113, + 528, + 199 + ], + "spans": [ + { + "bbox": [ + 416, + 113, + 528, + 199 + ], + "type": "image", + "image_path": "422ef3057227da8502a4891b9f65d9c235f97ba7228b6b93be73e3059ecf4a33.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_body" + } + ], + "index": 3 + }, + { + "type": "image", + "bbox": [ + 82, + 156, + 194, + 239 + ], + "blocks": [ + { + "bbox": [ + 82, + 156, + 194, + 239 + ], + "lines": [ + { + "bbox": [ + 82, + 156, + 194, + 239 + ], + "spans": [ + { + "bbox": [ + 82, + 156, + 194, + 239 + ], + "type": "image", + "image_path": "dfeae9c5e7edf804c7761b8b69d90cefa9e5711439075cdbf0cf3b5924d0d70e.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + } + ], + "index": 4 + }, + { + "type": "image", + "bbox": [ + 194, + 156, + 304, + 239 + ], + "blocks": [ + { + "bbox": [ + 194, + 156, + 304, + 239 + ], + "lines": [ + { + "bbox": [ + 194, + 156, + 304, + 239 + ], + "spans": [ + { + "bbox": [ + 194, + 156, + 304, + 239 + ], + "type": "image", + "image_path": "b0b0b6e11d73c86503357c3d2d3b3a6db8080d345c5b5c32a961f28e6483615b.jpg" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_body" + } + ], + "index": 5 + }, + { + "type": "image", + "bbox": [ + 306, + 156, + 416, + 239 + ], + "blocks": [ + { + "bbox": [ + 306, + 156, + 416, + 239 + ], + "lines": [ + { + "bbox": [ + 306, + 156, + 416, + 239 + ], + "spans": [ + { + "bbox": [ + 306, + 156, + 416, + 239 + ], + "type": "image", + "image_path": "2b3f7a48ee2048cc6c906db186e9a7c5e67434fb7063e090aa3ecacedc4999fb.jpg" + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_body" + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 416, + 293, + 527, + 404 + ], + "blocks": [ + { + "bbox": [ + 416, + 293, + 527, + 404 + ], + "lines": [ + { + "bbox": [ + 416, + 293, + 527, + 404 + ], + "spans": [ + { + "bbox": [ + 416, + 293, + 527, + 404 + ], + "type": "image", + "image_path": "c34f49219ae3e9362b1626a5b4587a3c28cc171924b7b793e9bad63ad95ee183.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + } + ], + "index": 7 + }, + { + "type": "image", + "bbox": [ + 82, + 239, + 194, + 350 + ], + "blocks": [ + { + "bbox": [ + 82, + 239, + 194, + 350 + ], + "lines": [ + { + "bbox": [ + 82, + 239, + 194, + 350 + ], + "spans": [ + { + "bbox": [ + 82, + 239, + 194, + 350 + ], + "type": "image", + "image_path": "fe7c3c5300dd2e5ceb79b29dc4c4f0cc82053d288d3baa899c79d4791bea9dfb.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_body" + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 194, + 239, + 304, + 350 + ], + "blocks": [ + { + "bbox": [ + 194, + 239, + 304, + 350 + ], + "lines": [ + { + "bbox": [ + 194, + 239, + 304, + 350 + ], + "spans": [ + { + "bbox": [ + 194, + 239, + 304, + 350 + ], + "type": "image", + "image_path": "68e2eac299c0ef1e34659f2a6045307b0a6df232fdda33a9ab17a6819e1bba54.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 306, + 239, + 416, + 350 + ], + "blocks": [ + { + "bbox": [ + 306, + 239, + 416, + 350 + ], + "lines": [ + { + "bbox": [ + 306, + 239, + 416, + 350 + ], + "spans": [ + { + "bbox": [ + 306, + 239, + 416, + 350 + ], + "type": "image", + "image_path": "8adea137841a19e9c19410f14b3210afd3a95816f103e14160eb7c30a084059e.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + } + ], + "index": 10 + }, + { + "type": "image", + "bbox": [ + 82, + 350, + 194, + 460 + ], + "blocks": [ + { + "bbox": [ + 82, + 350, + 194, + 460 + ], + "lines": [ + { + "bbox": [ + 82, + 350, + 194, + 460 + ], + "spans": [ + { + "bbox": [ + 82, + 350, + 194, + 460 + ], + "type": "image", + "image_path": "e570e0b6610a08e75690aedfe06d81a3d1f4e08a1ce7bc60eb7308d9a3f87eac.jpg" + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "lines": [ + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "spans": [ + { + "bbox": [ + 55, + 464, + 555, + 498 + ], + "type": "text", + "content": "Figure 4. Comparison on ScanNet++(Indoor) [76] and KITTI-360 [35] dataset. Qualitative comparison results on ScanNet++ and KITTI-360 datasets. Here, +C. T. indicates prediction results by appending Calibration Tokens to patch embeddings of the model located above. Calibration Tokens enable models to adapt to different fisheye cameras, especially in regions with large distortions." + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 11 + }, + { + "type": "image", + "bbox": [ + 194, + 350, + 304, + 460 + ], + "blocks": [ + { + "bbox": [ + 194, + 350, + 304, + 460 + ], + "lines": [ + { + "bbox": [ + 194, + 350, + 304, + 460 + ], + "spans": [ + { + "bbox": [ + 194, + 350, + 304, + 460 + ], + "type": "image", + "image_path": "72ec0f8a27bc4a256edb7f916d4ee22ecc24cea759d66ebd9b2723c385fde619.jpg" + } + ] + } + ], + "index": 12, + "angle": 0, + "type": "image_body" + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 306, + 350, + 416, + 460 + ], + "blocks": [ + { + "bbox": [ + 306, + 350, + 416, + 460 + ], + "lines": [ + { + "bbox": [ + 306, + 350, + 416, + 460 + ], + "spans": [ + { + "bbox": [ + 306, + 350, + 416, + 460 + ], + "type": "image", + "image_path": "aa0e006ec4d39685618949322c0684b01e48554f5e09c5a52f680af7f3bb4160.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + } + ], + "index": 13 + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": "ing a ResNet101 backbone, and compare both its pretrained model and a model trained on our dataset for any fairness concerns. We also compare with FoVA-Depth [36], which is equirectangular projection based, like DepthAnyCamera. Indoor Evaluation. Among the pretrained foundational monocular depth estimators, UniDepth achieves the best performance with our Calibration Tokens on the ScanNet++ indoor dataset as shown in Tab. 1. Notably, our Calibration Tokens enable MiDAS to improve " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "12\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and DepthAnything to achieve a " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "17\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " improvement in the RMSE metric compared to the model without Calibration Tokens. Similarly, UniDepth benefits from a " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "13\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " improvement in the RMSE metric. Furthermore, compared to the comparison baselines, pretrained DepthAnyCamera and FoVA-Depth, UniDepth with Calibration Tokens surpasses their performance by " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "11\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "inline_equation", + "content": "14\\%" + }, + { + "bbox": [ + 55, + 510, + 295, + 713 + ], + "type": "text", + "content": " in the RMSE metric, respectively. DepthAnyCamera and FoVA-Depth utilize camera intrin" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 510, + 555, + 557 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 557 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 557 + ], + "type": "text", + "content": "sics for input images at test time, requiring image transformations back and forth from the equirectangular reference frame, which makes them more error-prone than our direct learning-based approach." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 558, + 555, + 689 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 558, + 555, + 689 + ], + "spans": [ + { + "bbox": [ + 313, + 558, + 555, + 689 + ], + "type": "text", + "content": "Outdoor Evaluation. We evaluate FMDEs with our proposed Calibration Tokens against state-of-the-art methods in outdoor environments. The results show that Calibration Tokens consistently improve accuracy across different FMDEs in outdoor scenarios. Specifically, MiDaS and DepthAnything achieve improvement in the RMSE metric. UniDepth also improves " + }, + { + "bbox": [ + 313, + 558, + 555, + 689 + ], + "type": "inline_equation", + "content": "2\\%" + }, + { + "bbox": [ + 313, + 558, + 555, + 689 + ], + "type": "text", + "content": " in the RMSE metric, outperforming the comparison baselines. Given that the KITTI-360 dataset contains highly distorted images with a field of view exceeding 180 degrees, our Calibration Tokens demonstrate robustness across various distortion models." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 689, + 555, + 715 + ], + "type": "text", + "content": "Our Calibration Tokens are able to outperform DepthAnyCamera and FoVA-Depth without separate" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5203" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 86, + 107, + 523, + 363 + ], + "blocks": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "lines": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "spans": [ + { + "bbox": [ + 55, + 70, + 555, + 105 + ], + "type": "text", + "content": "Table 1. Quantitative comparisons on indoors (ScanNet++) and outdoors (KITTI-360) benchmarks. We evaluated zero-shot monocular depth estimation by incorporating trained Calibration Tokens into recent foundational monocular depth estimators models. Note: Our method uses the same training set for both the indoor and outdoor settings; whereas existing methods train separate models for each setting." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 86, + 107, + 523, + 363 + ], + "lines": [ + { + "bbox": [ + 86, + 107, + 523, + 363 + ], + "spans": [ + { + "bbox": [ + 86, + 107, + 523, + 363 + ], + "type": "table", + "html": "
TestsetExperimentModelTrain DatasetRMSE ↓δ1 ↑
ScanNet++ [76]BaselineMiDaS [49]Mix 1.4M0.5060.563
+ Calibration TokensMiDaS [49]+ Mix 200K0.4460.569
BaselineDepthAnything [73]Mix 63.5M0.7310.463
+ Calibration TokensDepthAnything [73]+ Mix 200K0.6070.506
BaselineUniDepth [45]Mix 16M0.2790.720
+ Calibration TokensUniDepth [45]+ Mix 200K0.2440.766
ComparisonsDepthAnyCamera [24]Indoor 670K0.2750.761
DepthAnyCamera [24]Mix 200K0.7610.255
FoVA-Depth [36]Indoor 190K0.2850.548
KITTI-360 [35]BaselineMiDaS [49]Mix 1.4M3.3120.586
+ Calibration TokensMiDaS[49]+ Mix 200K2.3480.658
BaselineDepthAnything [73]Mix 63.5M2.2140.839
+ Calibration TokensDepthAnything [73]+ Mix 200K2.0430.810
BaselineUniDepth [45]Mix 16M2.0850.663
+ Calibration TokensUniDepth [45]+ Mix 200K2.0400.664
ComparisonsDepthAnyCamera [24]Outdoor 130K2.0670.852
DepthAnyCamera [24]Mix 200K5.6750.348
FoVA-Depth [36]Outdoor 80K3.0960.632
", + "image_path": "046b4600cd7c9180711a78149ef27a112c7ddbcd33159b7bb8229bde2caaa163.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "image", + "bbox": [ + 81, + 373, + 268, + 519 + ], + "blocks": [ + { + "bbox": [ + 81, + 373, + 268, + 519 + ], + "lines": [ + { + "bbox": [ + 81, + 373, + 268, + 519 + ], + "spans": [ + { + "bbox": [ + 81, + 373, + 268, + 519 + ], + "type": "image", + "image_path": "c058ec082d15bf9c4f1865100451108c72e6e7452e3e17a9960933886b4dac5d.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 55, + 520, + 295, + 552 + ], + "lines": [ + { + "bbox": [ + 55, + 520, + 295, + 552 + ], + "spans": [ + { + "bbox": [ + 55, + 520, + 295, + 552 + ], + "type": "text", + "content": "Figure 5. t-SNE plot of fisheye and perspective embeddings. Fisheye embeddings become closer to those of perspective images after being modulated by Calibration Tokens." + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "spans": [ + { + "bbox": [ + 55, + 566, + 296, + 662 + ], + "type": "text", + "content": "indoor and outdoor training sets, suggesting the generalization potential of our method to wide ranges of fisheye distortions. Also, the KITTI-360 ground truth points are significantly sparser and more concentrated in ground regions as compared to ScanNet++, which may explain the discrepancy in evaluation metrics. Nonetheless, our method performs comparably without needing to train specialized sets of Calibration Tokens for different fisheye models." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 671, + 118, + 684 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 671, + 118, + 684 + ], + "spans": [ + { + "bbox": [ + 55, + 671, + 118, + 684 + ], + "type": "text", + "content": "4.2. Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 296, + 715 + ], + "type": "text", + "content": "Feature Modulation. To visualize how our Calibration Tokens affect fisheye embeddings, Fig. 5 shows a two-" + } + ] + } + ], + "index": 6 + }, + { + "type": "image", + "bbox": [ + 319, + 373, + 555, + 485 + ], + "blocks": [ + { + "bbox": [ + 319, + 373, + 555, + 485 + ], + "lines": [ + { + "bbox": [ + 319, + 373, + 555, + 485 + ], + "spans": [ + { + "bbox": [ + 319, + 373, + 555, + 485 + ], + "type": "image", + "image_path": "7843bff95324b3bd2dbd6be87ee4039d24e6309aeb6c579829603f949bc0236c.jpg" + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 487, + 555, + 566 + ], + "lines": [ + { + "bbox": [ + 313, + 487, + 555, + 566 + ], + "spans": [ + { + "bbox": [ + 313, + 487, + 555, + 566 + ], + "type": "text", + "content": "Figure 6. Comparison of attention maps for single and multi-layer tokens. We visualize the attention map of Calibration Tokens to the encoded patch embeddings. Calibration Tokens attend highly to distorted border regions: \"Single Token\" (top) has little effect in most layers due to lower attention as only a set of tokens are append to the input. The proposed multi-layer or \"Layer-wise Tokens\" scheme (bottom) attends to embeddings across all layers." + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "image_caption" + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "spans": [ + { + "bbox": [ + 313, + 581, + 554, + 629 + ], + "type": "text", + "content": "dimensional tSNE reduction to both fisheye and perspective image embeddings from the same set of images. After adding Calibration Tokens to the fisheye embeddings, they are modulated towards the perspective image distribution." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 630, + 555, + 715 + ], + "type": "text", + "content": "Layer-wise Tokens. As observed in Fig. 6, when we append only a single set of tokens (\"Single Token\") at the initial transformer block of the pre-trained model, the Calibration Tokens exhibit limited attention to the patch embeddings across most layers. As a result, the patch embeddings of most layers are unchanged. However, when we attach unique tokens at every layer (\"Layer-wise Tokens\"), we see" + } + ] + } + ], + "index": 10 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5204" + } + ] + } + ], + "index": 11 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 56, + 107, + 294, + 300 + ], + "blocks": [ + { + "bbox": [ + 56, + 70, + 295, + 103 + ], + "lines": [ + { + "bbox": [ + 56, + 70, + 295, + 103 + ], + "spans": [ + { + "bbox": [ + 56, + 70, + 295, + 103 + ], + "type": "text", + "content": "Table 2. Comparison results with finetuning. We conducted experiments comparing finetuning (F.T.) with the use of Calibration Tokens (C.T.) added to the baseline model." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 107, + 294, + 300 + ], + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 300 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 300 + ], + "type": "table", + "html": "
DatasetsModelsExp.RMSEδ1
ScanNet++MiDASF.T.2.1780.129
C.T.0.4460.569
DepthAnythingF.T.1.4590.462
C.T.0.6070.506
UniDepthF.T.0.4320.574
C.T.0.2440.766
KITTI-360MiDASF.T.9.2890.098
C.T.2.3480.658
DepthAnythingF.T.4.3620.636
C.T.2.0430.810
UniDepthF.T.3.2170.403
C.T.2.0400.664
", + "image_path": "3c76a5002ab4c804f854decc70731b3c19f73c7ffb503c9989858d8838835d33.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 343, + 293, + 411 + ], + "blocks": [ + { + "bbox": [ + 56, + 306, + 295, + 340 + ], + "lines": [ + { + "bbox": [ + 56, + 306, + 295, + 340 + ], + "spans": [ + { + "bbox": [ + 56, + 306, + 295, + 340 + ], + "type": "text", + "content": "Table 3. Analysis on computational cost. We analyze the computational overhead introduced by Calibration Tokens. Values in parentheses indicate the relative increase as a percentage." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 343, + 293, + 411 + ], + "lines": [ + { + "bbox": [ + 56, + 343, + 293, + 411 + ], + "spans": [ + { + "bbox": [ + 56, + 343, + 293, + 411 + ], + "type": "table", + "html": "
ModelsModel memoryTokens memoryInference time
MiDAS1.7G0.8M(0.05%)0.6ms(0.8%)
DepthAnything1.7G0.8M(0.05%)0.8ms(0.8%)
UniDepth0.7G0.2M(0.02%)0.4ms(0.7%)
", + "image_path": "77d820d70e09da2f9b8d2ce8923aadcef2fc2105e8a0a4c8e698b49ead21b3e4.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 423, + 295, + 460 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 423, + 295, + 460 + ], + "spans": [ + { + "bbox": [ + 55, + 423, + 295, + 460 + ], + "type": "text", + "content": "higher attention at more layers. Thus, we opt to use the \"Layer-wise\" approach to better modulate the fisheye patch embeddings toward the distribution of perspective images." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": "Comparison with Finetuning. To further analyze the robustness of the Calibration Tokens, we conducted experiments comparing our method with a finetuning approach. We trained the model with a fixed learning rate of " + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "inline_equation", + "content": "10^{-6}" + }, + { + "bbox": [ + 55, + 460, + 295, + 567 + ], + "type": "text", + "content": " on our synthetic fisheye dataset for the same number of iterations. As shown in Tab. 2, the finetuning approach leads to a significant performance drop, highlighting the importance of using Calibration Tokens, which preserve the original model's training on perspective images." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "spans": [ + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": "Computational Cost. Tab. 3 shows the impact of Calibration Tokens on computational costs across different FMDEs. Incorporating Calibration Tokens results in only a " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "0.05\\%" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "0.02\\%" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " increase in memory usage, less than 1 MB and a " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "0.8\\%" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "inline_equation", + "content": "0.7\\%" + }, + { + "bbox": [ + 55, + 567, + 295, + 651 + ], + "type": "text", + "content": " increase in inference time, with an added latency of less than 1 ms. This analysis highlights the efficiency of our proposed Calibration Tokens." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 659, + 149, + 673 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 659, + 149, + 673 + ], + "spans": [ + { + "bbox": [ + 55, + 659, + 149, + 673 + ], + "type": "text", + "content": "4.3. Ablation Study" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 295, + 713 + ], + "type": "text", + "content": "We conducted an ablation study on different Calibration Token configurations to validate our contributions. Note that \"Single Token\" refers to a single set of Calibration Tokens" + } + ] + } + ], + "index": 8 + }, + { + "type": "table", + "bbox": [ + 318, + 96, + 548, + 198 + ], + "blocks": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "lines": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "spans": [ + { + "bbox": [ + 313, + 70, + 555, + 93 + ], + "type": "text", + "content": "Table 4. Ablation study. We ablate the training objective and modulate scheme for our proposed Calibration Tokens." + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 318, + 96, + 548, + 198 + ], + "lines": [ + { + "bbox": [ + 318, + 96, + 548, + 198 + ], + "spans": [ + { + "bbox": [ + 318, + 96, + 548, + 198 + ], + "type": "table", + "html": "
DatasetMethodRMSEδ1
ScanNet++Single token0.2600.741
+ LogL1 Loss0.2540.752
+ Layer-wise Tokens0.2440.766
KITTI-360Single token2.0850.656
+ LogL1 Loss2.0650.665
+ Layer-wise Tokens2.0400.664
", + "image_path": "91b1ac53317f49b2bafc07bd3cc4a8672cbfe8b09ac304a53ce518facadd6628.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "table_body" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "spans": [ + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "text", + "content": "appended in the first layer of the vision transformer without removal, with L1 loss applied. In this configuration, " + }, + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\phi \\in \\mathbb{R}^{M\\times F}" + }, + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "text", + "content": " as opposed to " + }, + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "inline_equation", + "content": "\\Phi \\in \\mathbb{R}^{L\\times M\\times F}" + }, + { + "bbox": [ + 313, + 211, + 553, + 270 + ], + "type": "text", + "content": " in the layer-wise setting. The ablation study on the ScanNet++ and KITTI-360 datasets is performed using the UniDepth model." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "spans": [ + { + "bbox": [ + 313, + 271, + 553, + 331 + ], + "type": "text", + "content": "LogL1 Loss. We observed stable improvements with LogL1 loss compared to baseline L1 loss. As shown in Tab. 4, the LogL1 loss improves both metrics across indoor and outdoor datasets. Qualitative comparisons between L1 and LogL1 objectives are shown in the Supp. Mat." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 332, + 554, + 440 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 332, + 554, + 440 + ], + "spans": [ + { + "bbox": [ + 313, + 332, + 554, + 440 + ], + "type": "text", + "content": "Layer-wise Tokens. Tab. 4 demonstrates the advantages of using layer-wise tokens over a single set of Calibration Tokens in the first layer. Even when the same number of tokens is fed to the visual transformer layers, we observed a significant improvement in the contribution of layer-wise tokens. This supports our hypothesis about how the influence of Calibration Tokens diminishes through a forward pass as observed in our experiments by appending a single set of tokens at the first layer. Fig. 6 visualizes attention." + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 453, + 384, + 464 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 453, + 384, + 464 + ], + "spans": [ + { + "bbox": [ + 313, + 453, + 384, + 464 + ], + "type": "text", + "content": "5. Discussion" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "spans": [ + { + "bbox": [ + 313, + 473, + 555, + 616 + ], + "type": "text", + "content": "Calibration Tokens enable FMEs to adapt to images captured by fisheye cameras. Empirically, our method improves on monocular depth estimation on fisheye cameras. While our method trains only one set of tokens for both indoor and outdoor settings, our promising results motivate this as a general approach to adapting vision foundational models. Furthermore, a convenience afforded by our method is in the reuse and backward-compatibility of FMDEs with perspective images. This reduces the operational overhead of multi-camera systems by enabling a single FMDE to handle multiple camera inputs - adding cameras become as easy as appending tokens." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 617, + 556, + 713 + ], + "type": "text", + "content": "Limitations. While we offer a light-weight method of extending FMDEs to fisheye images, its success inherently depends on the quality and representational power of the underlying FMDEs. If the pretrained model struggles with certain 3D scenes or lighting conditions for perspective images, then these issues carry over. Nonetheless, as novel FMDEs emerge, our framework can be readily transferred to new models using transformer-based architectures." + } + ] + } + ], + "index": 16 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5205" + } + ] + } + ], + "index": 17 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 297, + 109 + ], + "type": "text", + "content": "Acknowledgments This work is supported by NSF 2112562 Athena AI Institute and the Rosenfeld Science Scholars Program." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "spans": [ + { + "bbox": [ + 56, + 120, + 115, + 133 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 57, + 140, + 297, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 61, + 140, + 297, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 140, + 297, + 195 + ], + "spans": [ + { + "bbox": [ + 61, + 140, + 297, + 195 + ], + "type": "text", + "content": "[1] Bruno Arsenali, Prashanth Viswanath, and Jelena Novosel. Rotinvmtl: Rotation invariant multinet on fisheye images for autonomous driving applications. In Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops, pages 0–0, 2019. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 61, + 197, + 296, + 240 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 197, + 296, + 240 + ], + "spans": [ + { + "bbox": [ + 61, + 197, + 296, + 240 + ], + "type": "text", + "content": "[2] Shariq Farooq Bhat, Ibrahim Alhashim, and Peter Wonka. Adabins: Depth estimation using adaptive bins. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4009-4018, 2021. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "spans": [ + { + "bbox": [ + 62, + 243, + 295, + 297 + ], + "type": "text", + "content": "[3] Aleksei Bochkovskii, AmaAgl Delaunoy, Hugo Germain, Marcel Santos, Yichao Zhou, Stephan R Richter, and Vladlen Koltun. Depth pro: Sharp monocular metric depth in less than a second. In International Conference on Learning Representations, 2025. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "spans": [ + { + "bbox": [ + 62, + 298, + 295, + 330 + ], + "type": "text", + "content": "[4] Mikhail S Burtsev, Yuri Kuratov, Anton Peganov, and Grigory V Sapunov. Memory transformer. arXiv preprint arXiv:2006.11527, 2020. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 332, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 332, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 62, + 332, + 294, + 376 + ], + "type": "text", + "content": "[5] Wenjie Chang, Yueyi Zhang, and Zhiwei Xiong. Transformer-based monocular depth estimation with attention supervision. In 32nd British Machine Vision Conference (BMVC 2021), 2021. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 377, + 295, + 421 + ], + "type": "text", + "content": "[6] Xien Chen, Suchisrit Gangopadhyay, Michael Chu, Patrick Rim, Hyoungseob Park, and Alex Wong. Uncle: Unsupervised continual learning of depth completion. arXiv preprint arXiv:2410.18074, 2024. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 295, + 478 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 295, + 478 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 295, + 478 + ], + "type": "text", + "content": "[7] Hyesong Choi, Hunsang Lee, Sunkyung Kim, Sunok Kim, Seungryong Kim, Kwanghoon Sohn, and Dongbo Min. Adaptive confidence thresholding for monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12808-12818, 2021. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "spans": [ + { + "bbox": [ + 62, + 479, + 295, + 544 + ], + "type": "text", + "content": "[8] Younjoon Chung, Hyoungseob Park, Patrick Rim, Xiaoran Zhang, Jihe He, Ziyao Zeng, Safa Cicek, Byung-Woo Hong, James S. Duncan, and Alex Wong. Eta: Energy-based test-time adaptation for depth completion. In Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 546, + 295, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 546, + 295, + 579 + ], + "spans": [ + { + "bbox": [ + 62, + 546, + 295, + 579 + ], + "type": "text", + "content": "[9] Timothee Darcet, Maxime Oquab, Julien Mairal, and Piotr Bojanowski. Vision transformers need registers. arXiv preprint arXiv:2309.16588, 2023. 4" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 581, + 295, + 634 + ], + "type": "text", + "content": "[10] Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. In North American Chapter of the Association for Computational Linguistics, 2019. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "spans": [ + { + "bbox": [ + 57, + 636, + 295, + 712 + ], + "type": "text", + "content": "[11] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, 2021. 2, 3" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 95 + ], + "type": "text", + "content": "[12] C Brown Duane. Close-range camera calibration. Photogramm. Eng, 37(8):855-866, 1971. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "spans": [ + { + "bbox": [ + 317, + 97, + 553, + 150 + ], + "type": "text", + "content": "[13] David Eigen and Rob Fergus. Predicting depth, surface normals and semantic labels with a common multi-scale convolutional architecture. In Proceedings of the IEEE international conference on computer vision, pages 2650-2658, 2015. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "spans": [ + { + "bbox": [ + 316, + 152, + 553, + 195 + ], + "type": "text", + "content": "[14] David Eigen, Christian Puhrsch, and Rob Fergus. Depth map prediction from a single image using a multi-scale deep network. Advances in neural information processing systems, 27, 2014. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 198, + 553, + 262 + ], + "type": "text", + "content": "[15] Vadim Ezhov, Hyoungseob Park, Zhaoyang Zhang, Rishi Upadhyay, Howard Zhang, Chethan Chinder Chandrappa, Achuta Kadambi, Yunhao Ba, Julie Dorsey, and Alex Wong. All-day depth completion. In 2024 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2024. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 297 + ], + "type": "text", + "content": "[16] Xiaohan Fei, Alex Wong, and Stefano Soatto. Geosupervised visual depth prediction. IEEE Robotics and Automation Letters, 4(2):1661-1668, 2019. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 299, + 553, + 354 + ], + "type": "text", + "content": "[17] Hao Feng, Wendi Wang, Jiajun Deng, Wengang Zhou, Li Li, and Houqiang Li. Simfir: A simple framework for fisheye image rectification with self-supervised representation learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12418-12427, 2023. 3, 4" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "spans": [ + { + "bbox": [ + 317, + 355, + 553, + 409 + ], + "type": "text", + "content": "[18] Huan Fu, Mingming Gong, Chaohui Wang, Kayhan Batmanghelich, and Dacheng Tao. Deep ordinal regression network for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2002-2011, 2018. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 317, + 411, + 553, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 411, + 553, + 476 + ], + "spans": [ + { + "bbox": [ + 317, + 411, + 553, + 476 + ], + "type": "text", + "content": "[19] Ravi Garg, Vijay Kumar Bg, Gustavo Carneiro, and Ian Reid. Unsupervised cnn for single view depth estimation: Geometry to the rescue. In Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11-14, 2016, Proceedings, Part VIII 14, pages 740-756. Springer, 2016. 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "spans": [ + { + "bbox": [ + 317, + 478, + 553, + 531 + ], + "type": "text", + "content": "[20] Clément Godard, Oisin Mac Aodha, and Gabriel J Brossow. Unsupervised monocular depth estimation with left-right consistency. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 270-279, 2017. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "spans": [ + { + "bbox": [ + 317, + 534, + 553, + 587 + ], + "type": "text", + "content": "[21] Clément Godard, Oisin Mac Aodha, Michael Firman, and Gabriel J Brostow. Digging into self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF international conference on computer vision, pages 3828-3838, 2019. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "spans": [ + { + "bbox": [ + 317, + 590, + 553, + 634 + ], + "type": "text", + "content": "[22] Vitor Guizilini, Rui Hou, Jie Li, Rares Ambrus, and Adrien Gaidon. Semantically-guided representation learning for self-supervised monocular depth. In International Conference on Learning Representations, 2019. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "spans": [ + { + "bbox": [ + 317, + 635, + 553, + 689 + ], + "type": "text", + "content": "[23] Vitor Guizilini, Rares Ambrus, Sudeep Pillai, Allan Raventos, and Adrien Gaidon. 3d packing for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2485-2494, 2020. 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 692, + 553, + 713 + ], + "type": "text", + "content": "[24] Yuliang Guo, Sparsh Garg, S Mahdi H Miangoleh, Xinyu Huang, and Liu Ren. Depth any camera: Zero-shot met" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "5206" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 296, + 712 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "spans": [ + { + "bbox": [ + 76, + 73, + 296, + 95 + ], + "type": "text", + "content": "ric depth estimation from any camera. arXiv preprint arXiv:2501.02464, 2025.3,5,7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 96, + 296, + 140 + ], + "type": "text", + "content": "[25] Juho Kannala and Sami Brandt. A generic camera calibration method for fish-eye lenses. In Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004., pages 10-13. IEEE, 2004. 3, 4" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "spans": [ + { + "bbox": [ + 56, + 141, + 294, + 206 + ], + "type": "text", + "content": "[26] Bingxin Ke, Anton Obukhov, Shengyu Huang, Nando Metzger, Rodrigo Caye Daudt, and Konrad Schindler. Repurposing diffusion-based image generators for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 9492-9502, 2024. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 208, + 295, + 229 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 208, + 295, + 229 + ], + "spans": [ + { + "bbox": [ + 57, + 208, + 295, + 229 + ], + "type": "text", + "content": "[27] R Kingslake. A history of the photographic lens. University of Rochester NY, 145, 1989. 3" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 232, + 294, + 297 + ], + "type": "text", + "content": "[28] Varun Ravi Kumar, Marvin Klingner, Senthil Yogamani, Stefan Milz, Tim Fingscheidt, and Patrick Mader. Syndistnet: Self-supervised monocular fisheye camera distance estimation synergized with semantic segmentation for autonomous driving. In Proceedings of the IEEE/CVF winter conference on applications of computer vision, pages 61-71, 2021. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "spans": [ + { + "bbox": [ + 56, + 299, + 294, + 353 + ], + "type": "text", + "content": "[29] Iro Laina, Christian Rupprecht, Vasileios Belagiannis, Federico Tombari, and Nassir Navab. Deeper depth prediction with fully convolutional residual networks. In 2016 Fourth international conference on 3D vision (3DV), pages 239-248. IEEE, 2016. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "spans": [ + { + "bbox": [ + 56, + 355, + 294, + 399 + ], + "type": "text", + "content": "[30] Dong Lao, Yangchao Wu, Tian Yu Liu, Alex Wong, and Stefano Soatto. Sub-token vit embedding via stochastic resonance transformers. In International Conference on Machine Learning. PMLR, 2024." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "spans": [ + { + "bbox": [ + 56, + 399, + 294, + 453 + ], + "type": "text", + "content": "[31] Dong Lao, Fengyu Yang, Daniel Wang, Hyoungseob Park, Samuel Lu, Alex Wong, and Stefano Soatto. On the viability of monocular depth pre-training for semantic segmentation. In European Conference on Computer Vision. Springer, 2024." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "spans": [ + { + "bbox": [ + 56, + 456, + 294, + 521 + ], + "type": "text", + "content": "[32] Bo Li, Chunhua Shen, Yuchao Dai, Anton Van Den Hengel, and Mingyi He. Depth and surface normal estimation from monocular images using regression on deep features and hierarchical crfs. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1119-1127, 2015. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "spans": [ + { + "bbox": [ + 56, + 523, + 294, + 567 + ], + "type": "text", + "content": "[33] Zhenyu Li, Zehui Chen, Xianming Liu, and Junjun Jiang. Depthformer: Exploiting long-range correlation and local information for accurate monocular depth estimation. Machine Intelligence Research, 20(6):837-854, 2023. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 568, + 294, + 612 + ], + "type": "text", + "content": "[34] Kang Liao, Chunyu Lin, Yao Zhao, and Moncef Gabbouj. Dr-gan: Automatic radial distortion rectification using conditional gan in real-time. IEEE Transactions on Circuits and Systems for Video Technology, 30(3):725-733, 2019. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 656 + ], + "type": "text", + "content": "[35] Yiyi Liao, Jun Xie, and Andreas Geiger. Kitti-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(3):3292-3310, 2022. 5, 6, 7" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 658, + 294, + 712 + ], + "type": "text", + "content": "[36] Daniel Lichy, Hang Su, Abhishek Badki, Jan Kautz, and Orazio Gallo. Fova-depth: Field-of-view agnostic depth estimation for cross-dataset generalization. In 2024 International Conference on 3D Vision (3DV), pages 1-10. IEEE, 2024. 3, 6, 7" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 713 + ], + "type": "list", + "angle": 0, + "index": 27, + "blocks": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "spans": [ + { + "bbox": [ + 316, + 73, + 553, + 126 + ], + "type": "text", + "content": "[37] Fayao Liu, Chunhua Shen, Guosheng Lin, and Ian Reid. Learning depth from single monocular images using deep convolutional neural fields. IEEE transactions on pattern analysis and machine intelligence, 38(10):2024-2039, 2015. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "spans": [ + { + "bbox": [ + 316, + 129, + 553, + 194 + ], + "type": "text", + "content": "[38] Tian Yu Liu, Parth Agrawal, Allison Chen, Byung-Woo Hong, and Alex Wong. Monitored distillation for positive congruent depth completion. In Computer Vision-ECCV 2022: 17th European Conference, Tel Aviv, Israel, October 23-27, 2022, Proceedings, Part II, pages 35-53. Springer, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 316, + 197, + 553, + 251 + ], + "type": "text", + "content": "[39] Xiaoyang Lyu, Liang Liu, Mengmeng Wang, Xin Kong, Lina Liu, Yong Liu, Xinxin Chen, and Yi Yuan. Hr-depth: High resolution self-supervised monocular depth estimation. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 2294-2301, 2021. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "spans": [ + { + "bbox": [ + 316, + 253, + 555, + 306 + ], + "type": "text", + "content": "[40] Reza Mahjourian, Martin Wicke, and Anelia Angelova. Unsupervised learning of depth and ego-motion from monocular video using 3d geometric constraints. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5667-5675, 2018. 3" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 308, + 553, + 329 + ], + "type": "text", + "content": "[41] Kenro Miyamoto. Fish eye lens. JOSA, 54(8):1060-1061, 1964. 3" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "spans": [ + { + "bbox": [ + 316, + 332, + 555, + 375 + ], + "type": "text", + "content": "[42] Hyoungseob Park, Anjali Gupta, and Alex Wong. Test-time adaptation for depth completion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 20519-20529, 2024. 3" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 376, + 555, + 431 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 376, + 555, + 431 + ], + "spans": [ + { + "bbox": [ + 316, + 376, + 555, + 431 + ], + "type": "text", + "content": "[43] Rui Peng, Ronggang Wang, Yawen Lai, Luyang Tang, and Yangang Cai. Excavating the potential capacity of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 15560-15569, 2021. 5" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "spans": [ + { + "bbox": [ + 316, + 434, + 553, + 487 + ], + "type": "text", + "content": "[44] Luigi Piccinelli, Yung-Hsu Yang, Christos Sakaridis, Mattia Segu, Siyuan Li, Luc Van Gool, and Fisher Yu. Undepth: Universal monocular metric depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10106-10116, 2024. 1, 2, 3" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "spans": [ + { + "bbox": [ + 316, + 489, + 553, + 533 + ], + "type": "text", + "content": "[45] Luigi Piccinelli, Christos Sakaridis, Yung-Hsu Yang, Mattia Segu, Siyuan Li, Wim Abbeloos, and Luc Van Gool. Undepthv2: Universal monocular metric depth estimation made simpler. arXiv preprint arXiv:2502.20110, 2025. 5, 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 535, + 553, + 578 + ], + "type": "text", + "content": "[46] Matteo Poggi, Fabio Tosi, and Stefano Mattoccia. Learning monocular depth estimation with unsupervised trinocular assumptions. In 2018 International conference on 3d vision (3DV), pages 324-333. IEEE, 2018. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "spans": [ + { + "bbox": [ + 316, + 579, + 553, + 633 + ], + "type": "text", + "content": "[47] Matteo Poggi, Filippo Aleotti, Fabio Tosi, and Stefano Mattoccia. On the uncertainty of self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3227-3237, 2020. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "spans": [ + { + "bbox": [ + 316, + 635, + 553, + 678 + ], + "type": "text", + "content": "[48] René Ranftl, Alexey Bochkovskiy, and Vladlen Koltun. Vision transformers for dense prediction. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 2, 3" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 316, + 681, + 553, + 713 + ], + "type": "text", + "content": "[49] Rene Ranftl, Katrin Lasinger, David Hafner, Konrad Schindler, and Vladlen Koltun. Towards Robust Monocular Depth Estimation: Mixing Datasets for Zero-Shot Cross" + } + ] + } + ], + "index": 26 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5207" + } + ] + } + ], + "index": 28 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 295, + 713 + ], + "type": "list", + "angle": 0, + "index": 13, + "blocks": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "spans": [ + { + "bbox": [ + 77, + 73, + 294, + 95 + ], + "type": "text", + "content": "Dataset Transfer. IEEE Transactions on Pattern Analysis & Machine Intelligence, 44(03):1623-1637, 2022. 2, 3, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "spans": [ + { + "bbox": [ + 57, + 96, + 295, + 128 + ], + "type": "text", + "content": "[50] Patrick Rim, Hyoungseob Park, Vadim Ezhov, Jeffrey Moon, and Alex Wong. Radar-guided polynomial fitting for metric depth estimation. arXiv preprint arXiv:2503.17182, 2025. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 295, + 183 + ], + "type": "text", + "content": "[51] Patrick Rim, Hyoungseob Park, Ziyao Zeng, Younjoon Chung, and Alex Wong. Protodepth: Unsupervised continual depth completion with prototypes. In Proceedings of the Computer Vision and Pattern Recognition Conference, pages 6304-6316, 2025. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "spans": [ + { + "bbox": [ + 56, + 184, + 295, + 249 + ], + "type": "text", + "content": "[52] Mike Roberts, Jason Ramapuram, Anurag Ranjan, Atulit Kumar, Miguel Angel Bautista, Nathan Paczan, Russ Webb, and Joshua M Susskind. Hypersim: A photorealistic synthetic dataset for holistic indoor scene understanding. In Proceedings of the IEEE/CVF international conference on computer vision, pages 10912-10922, 2021. 5" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 250, + 295, + 293 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 295, + 293 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 295, + 293 + ], + "type": "text", + "content": "[53] Nathan Silberman, Derek Hoiem, Pushmeet Kohli, and Rob Fergus. Indoor segmentation and support inference from rgbd images. In European Conference on Computer Vision, 2012. 5" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 295, + 360 + ], + "type": "text", + "content": "[54] Akash Deep Singh, Yunhao Ba, Ankur Sarker, Howard Zhang, Achuta Kadambi, Stefano Soatto, Mani Srivastava, and Alex Wong. Depth estimation from camera image and mmwave radar point cloud. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9275-9285, 2023. 3" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 361, + 295, + 403 + ], + "type": "text", + "content": "[55] Daniel E Stevenson and Margaret M Fleck. Nonparametric correction of distortion. In Proceedings Third IEEE Workshop on Applications of Computer Vision. WACV'96, pages 214-219. IEEE, 1996. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 295, + 470 + ], + "type": "text", + "content": "[56] Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, et al. Scalability in perception for autonomous driving: Waymo open dataset. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2446-2454, 2020. 5" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "spans": [ + { + "bbox": [ + 56, + 471, + 295, + 525 + ], + "type": "text", + "content": "[57] Fabio Tosi, Filippo Aleotti, Matteo Poggi, and Stefano Mattoccia. Learning monocular depth estimation infusing traditional stereo knowledge. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9799-9809, 2019. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 526, + 295, + 581 + ], + "type": "text", + "content": "[58] Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herve Jegou. Training data-efficient image transformers and distillation through attention. In Proceedings of the 38th International Conference on Machine Learning, pages 10347-10357. PMLR, 2021. 3" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 295, + 635 + ], + "type": "text", + "content": "[59] Rishi Upadhyay, Howard Zhang, Yunhao Ba, Ethan Yang, Blake Gella, Sicheng Jiang, Alex Wong, and Achuta Kadambi. Enhancing diffusion models with 3d perspective geometry constraints. ACM Transactions on Graphics (TOG), 42(6):1-15, 2023. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "spans": [ + { + "bbox": [ + 56, + 636, + 295, + 689 + ], + "type": "text", + "content": "[60] Chaoyang Wang, José Miguel Buenaposada, Rui Zhu, and Simon Lucey. Learning depth from monocular videos using direct methods. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 2022-2030, 2018. 3" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 691, + 295, + 713 + ], + "type": "text", + "content": "[61] Qiang Wang, Shizhen Zheng, Qingsong Yan, Fei Deng, Kaiyong Zhao, and Xiaowen Chu. IRS: A large naturalistic" + } + ] + } + ], + "index": 12 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 712 + ], + "type": "list", + "angle": 0, + "index": 26, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 116 + ], + "type": "text", + "content": "indoor robotics stereo dataset to train deep models for disparity and surface normal estimation. In 2021 IEEE International Conference on Multimedia and Expo (ICME), pages 1-6. IEEE, 2021. 5" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "spans": [ + { + "bbox": [ + 316, + 118, + 553, + 163 + ], + "type": "text", + "content": "[62] Jamie Watson, Michael Firman, Gabriel J Brostow, and Daniyar Turmukhambetov. Self-supervised monocular depth hints. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2162-2171, 2019. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 316, + 164, + 553, + 217 + ], + "type": "text", + "content": "[63] Alex Wong and Stefano Soatto. Bilateral cyclic constraint and adaptive regularization for unsupervised monocular depth prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5644-5653, 2019." + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "spans": [ + { + "bbox": [ + 316, + 220, + 553, + 262 + ], + "type": "text", + "content": "[64] Alex Wong, Safa Cicek, and Stefano Soatto. Targeted adversarial perturbations for monocular depth prediction. Advances in neural information processing systems, 33:8486-8497, 2020." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "spans": [ + { + "bbox": [ + 316, + 264, + 553, + 308 + ], + "type": "text", + "content": "[65] Alex Wong, Xiaohan Fei, Stephanie Tsuei, and Stefano Soatto. Unsupervised depth completion from visual inertial odometry. IEEE Robotics and Automation Letters, 5(2): 1899-1906, 2020. 5" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 310, + 553, + 354 + ], + "type": "text", + "content": "[66] Alex Wong, Safa Cicek, and Stefano Soatto. Learning topology from synthetic data for unsupervised depth completion. IEEE Robotics and Automation Letters, 6(2):1495-1502, 2021." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 553, + 399 + ], + "type": "text", + "content": "[67] Alex Wong, Xiaohan Fei, Byung-Woo Hong, and Stefano Soatto. An adaptive framework for learning unsupervised depth completion. IEEE Robotics and Automation Letters, 6 (2):3120-3127, 2021. 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 401, + 553, + 455 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 401, + 553, + 455 + ], + "spans": [ + { + "bbox": [ + 316, + 401, + 553, + 455 + ], + "type": "text", + "content": "[68] Yangchao Wu, Tian Yu Liu, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Augundo: Scaling up augmentations for monocular depth completion and estimation. In European Conference on Computer Vision, pages 274-293. Springer, 2024. 3, 5" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 457, + 553, + 511 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 457, + 553, + 511 + ], + "spans": [ + { + "bbox": [ + 316, + 457, + 553, + 511 + ], + "type": "text", + "content": "[69] Chao Xia, Chenfeng Xu, Patrick Rim, Mingyu Ding, Nanning Zheng, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Quadric representations for lidar odometry, mapping and localization. IEEE Robotics and Automation Letters, 8 (8):5023-5030, 2023. 3" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 514, + 553, + 579 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 514, + 553, + 579 + ], + "spans": [ + { + "bbox": [ + 316, + 514, + 553, + 579 + ], + "type": "text", + "content": "[70] Yichen Xie, Chenfeng Xu, Marie-Julie Rakotosaona, Patrick Rim, Federico Tombari, Kurt Keutzer, Masayoshi Tomizuka, and Wei Zhan. Sparsefusion: Fusing multi-modal sparse representations for multi-sensor 3d object detection. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 17591-17602, 2023. 3" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "spans": [ + { + "bbox": [ + 316, + 581, + 555, + 635 + ], + "type": "text", + "content": "[71] Dan Xu, Elisa Ricci, Wanli Ouyang, Xiaogang Wang, and Nicu Sebe. Multi-scale continuous crfs as sequential deep networks for monocular depth estimation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 5354-5362, 2017. 3" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 636, + 553, + 712 + ], + "type": "text", + "content": "[72] Fengyu Yang, Chao Feng, Ziyang Chen, Hyoungseob Park, Daniel Wang, Yiming Dou, Ziyao Zeng, Xien Chen, Rit Gangopadhyay, Andrew Owens, and Alex Wong. Binding touch to everything: Learning unified multimodal tactile representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26340-26353, 2024. 3" + } + ] + } + ], + "index": 25 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "5208" + } + ] + } + ], + "index": 27 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 713 + ], + "type": "list", + "angle": 0, + "index": 12, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[73] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2024. 1, 2, 3, 5, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "spans": [ + { + "bbox": [ + 56, + 129, + 294, + 174 + ], + "type": "text", + "content": "[74] Yanchao Yang, Alex Wong, and Stefano Soatto. Dense depth posterior (ddp) from single image and sparse range. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3353-3362, 2019. 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "spans": [ + { + "bbox": [ + 56, + 175, + 294, + 228 + ], + "type": "text", + "content": "[75] Zhenheng Yang, Peng Wang, Yang Wang, Wei Xu, and Ram Nevatia. Lego: Learning edge with geometry all at once by watching videos. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 225-234, 2018. 3" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "spans": [ + { + "bbox": [ + 56, + 231, + 294, + 275 + ], + "type": "text", + "content": "[76] Chandan Yeshwanth, Yueh-Cheng Liu, Matthias Nießner, and Angela Dai. Scannet++: A high-fidelity dataset of 3d indoor scenes. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 12-22, 2023. 5, 6, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "spans": [ + { + "bbox": [ + 56, + 276, + 294, + 319 + ], + "type": "text", + "content": "[77] Wei Yin, Yifan Liu, Chunhua Shen, and Youliang Yan. Enforcing geometric constraints of virtual normal for depth prediction. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 5684-5693, 2019. 3" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "spans": [ + { + "bbox": [ + 56, + 321, + 294, + 376 + ], + "type": "text", + "content": "[78] Xiaoqing Yin, Xinchao Wang, Jun Yu, Maojun Zhang, Pascal Fua, and Dacheng Tao. Fisheyerecnet: A multi-context collaborative deep network for fisheye image rectification. In Proceedings of the European conference on computer vision (ECCV), pages 469-484, 2018. 4" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 378, + 294, + 442 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 378, + 294, + 442 + ], + "spans": [ + { + "bbox": [ + 56, + 378, + 294, + 442 + ], + "type": "text", + "content": "[79] Senthil Yogamani, David Unger, Venkatraman Narayanan, and Varun Ravi Kumar. Fisheyebevseg: Surround view fisheye cameras based bird's-eye view segmentation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 1331-1334, 2024. 3" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "spans": [ + { + "bbox": [ + 56, + 445, + 294, + 498 + ], + "type": "text", + "content": "[80] Weihao Yuan, Xiaodong Gu, Zuozhuo Dai, Siyu Zhu, and Ping Tan. Neural window fully-connected crfs for monocular depth estimation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3916-3925, 2022. 3" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "spans": [ + { + "bbox": [ + 56, + 501, + 294, + 555 + ], + "type": "text", + "content": "[81] Ziyao Zeng, Jingcheng Ni, Daniel Wang, Patrick Rim, Younjoon Chung, Fengyu Yang, Byung-Woo Hong, and Alex Wong. Priordiffusion: Leverage language prior in diffusion models for monocular depth estimation. arXiv preprint arXiv:2411.16750, 2024. 3" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "spans": [ + { + "bbox": [ + 56, + 557, + 294, + 612 + ], + "type": "text", + "content": "[82] Ziyao Zeng, Daniel Wang, Fengyu Yang, Hyoungseob Park, Stefano Soatto, Dong Lao, and Alex Wong. Wordepth: Variational language prior for monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 9708-9719, 2024." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 613, + 294, + 677 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 613, + 294, + 677 + ], + "spans": [ + { + "bbox": [ + 56, + 613, + 294, + 677 + ], + "type": "text", + "content": "[83] Ziyao Zeng, Yangchao Wu, Hyoungseob Park, Daniel Wang, Fengyu Yang, Stefano Soatto, Dong Lao, Byung-Woo Hong, and Alex Wong. Rsa: Resolving scale ambiguities in monocular depth estimators through language descriptions. Advances in neural information processing systems, 37, 2024. 3" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 56, + 680, + 294, + 713 + ], + "type": "text", + "content": "[84] Huangying Zhan, Ravi Garg, Chamara Saroj Weerasekera, Kejie Li, Harsh Agarwal, and Ian Reid. Unsupervised learning of monocular depth estimation and visual odometry with" + } + ] + } + ], + "index": 11 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 329 + ], + "type": "list", + "angle": 0, + "index": 18, + "blocks": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "spans": [ + { + "bbox": [ + 333, + 73, + 553, + 105 + ], + "type": "text", + "content": "deep feature reconstruction. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 340-349, 2018. 3" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "spans": [ + { + "bbox": [ + 316, + 107, + 553, + 171 + ], + "type": "text", + "content": "[85] Ning Zhang, Francesco Nex, George Vosselman, and Norman Kerle. Lite-mono: A lightweight cnn and transformer architecture for self-supervised monocular depth estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 18537-18546, 2023. 3" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "spans": [ + { + "bbox": [ + 316, + 174, + 553, + 239 + ], + "type": "text", + "content": "[86] Chaoqiang Zhao, Youmin Zhang, Matteo Poggi, Fabio Tosi, Xianda Guo, Zheng Zhu, Guan Huang, Yang Tang, and Stefano Mattoccia. Monovit: Self-supervised monocular depth estimation with a vision transformer. In 2022 International Conference on 3D Vision (3DV), pages 668-678. IEEE, 2022. 3" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "spans": [ + { + "bbox": [ + 316, + 240, + 553, + 283 + ], + "type": "text", + "content": "[87] Guoyang Zhao, Yuxuan Liu, Weiqing Qi, Fulong Ma, Ming Liu, and Jun Ma. Fisheyedgepth: A real scale self-supervised depth estimation model for fisheye camera. arXiv preprint arXiv:2409.15054, 2024. 3" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "spans": [ + { + "bbox": [ + 316, + 285, + 553, + 329 + ], + "type": "text", + "content": "[88] Tinghui Zhou, Matthew Brown, Noah Snavely, and David G Lowe. Unsupervised learning of depth and ego-motion from video. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1851-1858, 2017. 3" + } + ] + } + ], + "index": 17 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "5209" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_content_list.json b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_content_list.json new file mode 100644 index 0000000000000000000000000000000000000000..fc6e518cfd15d20459102e6928be91d48f21c9b2 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_content_list.json @@ -0,0 +1,1721 @@ +[ + { + "type": "text", + "text": "External Knowledge Injection for CLIP-Based Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 119, + 130, + 879, + 152 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Da-Wei Zhou $^{1,2}$ , Kai-Wen Li $^{1,2}$ , Jingyi Ning $^{2}$ , Han-Jia Ye $^{1,2(\\boxtimes)}$ , Lijun Zhang $^{1,2}$ , De-Chuan Zhan $^{1,2}$ $^{1}$ School of Artificial Intelligence, Nanjing University", + "bbox": [ + 109, + 179, + 885, + 215 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "$^{2}$ National Key Laboratory for Novel Software Technology, Nanjing University", + "bbox": [ + 184, + 215, + 810, + 234 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "{zhoudw,likw,yehj,zhanglj,zhandc}@lamda.nju.edu.cn，ningjy@nju.edu.cn", + "bbox": [ + 189, + 236, + 800, + 252 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Abstract", + "text_level": 1, + "bbox": [ + 246, + 286, + 326, + 301 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of \"cat\" can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledgeGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features as data evolves. Extensive experiments demonstrate its state-of-the-art performance. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE.", + "bbox": [ + 88, + 319, + 485, + 712 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "1. Introduction", + "text_level": 1, + "bbox": [ + 89, + 742, + 223, + 758 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "Recent advancements in deep learning have significantly impacted various aspects of life [13, 21, 72]. However, real-world data often presents substantial challenges to these models, particularly streaming data that requires continual learning [2, 34]. Class-Incremental Learning (CIL) [51] has been proposed to address this challenge by enabling models to absorb new knowledge incrementally.", + "bbox": [ + 89, + 768, + 483, + 875 + ], + "page_idx": 0 + }, + { + "type": "image", + "img_path": "images/5b4db64d4b85d37e186f514ef85db0823d9b7c8d92f9561e3fec5787c5d3fe69.jpg", + "image_caption": [ + "Figure 1. Last accuracy comparison on nine benchmark datasets. All methods are deployed with the same CLIP for fair comparison. ENGINE outperforms other competitors with a substantial margin. A major issue in CIL is catastrophic forgetting [18, 19], where learning new classes leads to the loss of previously acquired knowledge. As a result, much research has focused on mitigating forgetting in deep learning [12, 42]. In contrast to traditional approaches that train models from scratch, recent progress in pre-training has shifted attention towards using pre-trained models (PTMs) [65], such as Vision Transformers[15] and CLIP [49]. With extensive training data serving as pre-learned knowledge, PTMs offer a more generalizable starting point for CIL, showing promise for real-world applications." + ], + "image_footnote": [], + "bbox": [ + 532, + 266, + 888, + 503 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "CLIP [49], as a pioneering pre-trained vision-language model [27, 33, 71, 75, 78], exhibits strong zero-shot performance by matching the embeddings of query images to class names. However, since visual features often contain fine-grained information [45], these detailed descriptors are neglected when using class names as matching targets. For example, the text concept of \"a photo of a cat\" can be further decomposed into features like \"whiskers,\" \"tail,\" and \"fur\" in order to match the visual embeddings, while these informative features are not well utilized. Furthermore, as the model is updated in CIL, the sub-features can be overwritten during the learning process, resulting in", + "bbox": [ + 511, + 719, + 908, + 902 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "CVF", + "bbox": [ + 106, + 2, + 181, + 42 + ], + "page_idx": 0 + }, + { + "type": "header", + "text": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore.", + "bbox": [ + 238, + 0, + 807, + 46 + ], + "page_idx": 0 + }, + { + "type": "page_footnote", + "text": "†Correspondence to: Han-Jia Ye (yehj@lamda.nju.edu.cn)", + "bbox": [ + 107, + 886, + 419, + 901 + ], + "page_idx": 0 + }, + { + "type": "page_number", + "text": "3314", + "bbox": [ + 482, + 944, + 516, + 955 + ], + "page_idx": 0 + }, + { + "type": "text", + "text": "forgetting and a mismatch between cross-modal features.", + "bbox": [ + 89, + 90, + 467, + 104 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Recent studies [62, 65, 73, 86] have highlighted the limitations of using fixed template text in vision-language models, which often leads to a loss of rich contextual information. To address this, researchers have proposed learnable text inputs, such as task-specific prompts [88], which allow the model to encode additional, task-relevant information. However, these learnable inputs can become biased towards the visual features in the training set, ultimately compromising the model's generalizability [87].", + "bbox": [ + 88, + 106, + 480, + 241 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In CLIP, a more detailed description in the textual space can provide a more informative matching target, enhancing prediction accuracy. For example, instead of using the generic template \"a photo of a cat,\" directly describing its features, such as \"whiskers\" and \"tail,\" can significantly improve recognition. In this way, these detailed features serve as anchors in the embedding space, and aligning to them helps reduce the risk of catastrophic forgetting.", + "bbox": [ + 88, + 242, + 480, + 362 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Learning with informative descriptors offers a promising approach to enhancing CLIP's continual learning capabilities. However, two major challenges remain: 1) acquiring informative features. Given the diversity of classes in downstream tasks, it is impractical for human experts to annotate all relevant features for recognition. 2) mitigating forgetting when utilizing external knowledge. As CIL involves continuous data streams, the model risks forgetting previously acquired knowledge after each update. Therefore, it is crucial to efficiently extract and retain external knowledge while preventing forgetting during sequential updates.", + "bbox": [ + 88, + 363, + 480, + 527 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "To address these challenges, we propose ExterNal knowledgeGe InjEction (ENGINE) for CLIP-based CIL. To acquire informative features, we enhance knowledge transfer by incorporating external knowledge from both visual and textual modalities through a dual-branch tuning framework. The visual branch is enhanced via data augmentation to enrich visual features, while the textual branch uses GPT-4 to rewrite discriminative prompts. In this way, we can encode diverse and informative class descriptions in the continual updating process. In addition to on-the-fly knowledge injection, we also apply post-tuning by re-ranking predictions during inference. Post-tuning is done by considering the local pair-wise features, which can further calibrate the wrong predictions to boost incremental learning.", + "bbox": [ + 88, + 529, + 480, + 741 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "2. Related Work", + "text_level": 1, + "bbox": [ + 89, + 753, + 232, + 768 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Class-Incremental Learning (CIL): is a long-standing problem, aiming to absorb new knowledge without forgetting [12, 42]. Typical CIL algorithms train a model from scratch, which can be divided into several groups. Knowledge distillation-based methods aim to build the mapping target [23] between old and new models to resist forgetting [16, 36, 51], which involve logit-wise alignment [36, 51], feature-wise alignment [24, 39, 46], and", + "bbox": [ + 88, + 779, + 480, + 900 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "group-wise alignment [14, 20, 57]. Replay-based methods aim to recover previous knowledge by saving and replaying the subset of seen classes [5, 8, 9, 38, 40, 81]. Besides, parameter regularization-based methods estimate the importance of parameters and restrict important ones not to change [3, 4, 30, 79]. Model rectification-based methods observe and rectify the inductive bias in the model, e.g., the biased logits [52, 77, 80] and classifier weights [67, 80]. Model expansion-based methods [60, 70, 83] adjust the network structure to fit the characteristics of the evolving data, e.g., expanding neurons [69, 74], the whole backbone [60, 61, 70, 82, 83], and lightweight modules [17].", + "bbox": [ + 511, + 90, + 903, + 271 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "Pre-Trained Model-Based CIL: improves the capability of CIL by starting with generalizable pre-trained models [35, 48, 56, 85]. To maintain the pre-trained knowledge, most works seek to freeze the backbone and append lightweight modules, e.g., prompts [53, 63-65, 88] and adapters [10, 50, 76]. For example, L2P [65] and DualPrompt [64] design the visual prompt [28] pool to select the instance-specific prompts when learning with the pre-trained Vision Transformer. Several works build complex prompt combination/generation targets with attention mechanism [53] or generative network [29]. Other works directly utilize the pre-trained features to construct strong classifiers by matching class prototypes to the embeddings [44, 54, 84]. When the pre-trained CLIP is available as initialization, existing works seek to enhance cross-modal matching information via learning multi-modal prompts [62, 63]. MOE-Adapter [76] further extends the lightweight module selection process via mixture-of-experts [43], while PROOF [86] extends the representation ability of CLIP by appending new projection layers for new tasks. RAPF [25] learns the adapter modules for new tasks by decomposed parameter fusion to resist forgetting in the updating process.", + "bbox": [ + 511, + 272, + 903, + 604 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3. Preliminaries", + "text_level": 1, + "bbox": [ + 511, + 619, + 651, + 636 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "3.1. Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 511, + 645, + 766, + 662 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "CIL is designed to continually build a unified classifier for all seen classes of the data stream [51]. We denote training sets as $\\{\\mathcal{D}^1,\\mathcal{D}^2,\\dots ,\\mathcal{D}^B\\}$ , where each task $\\mathcal{D}^b = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{n_b}$ contains $n_b$ instances. Each training instance $\\mathbf{x}_i\\in \\mathbb{R}^D$ belongs to class $y_{i}\\in Y_{b}$ , i.e., the label space of task $b$ is $Y_{b}$ . We have $Y_{b}\\cap Y_{b^{\\prime}} = \\emptyset$ for $b\\neq b^{\\prime}$ . This paper follows the exemplar-free CIL setting [64, 65, 89], where the model cannot hold any historical instances in the memory. In other words, we can only access data from $\\mathcal{D}^b$ for model training when learning the $b$ -th incremental task. In CIL, the target is to build a unified classifier for all seen classes $\\mathcal{V}_b = Y_1\\cup \\dots Y_b$ , i.e., find a model $f(\\mathbf{x}):X\\to \\mathcal{V}_b$ that minimizes the expected risk:", + "bbox": [ + 511, + 667, + 903, + 864 + ], + "page_idx": 1 + }, + { + "type": "equation", + "text": "\n$$\nf ^ {*} = \\underset {f \\in \\mathcal {H}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {t} ^ {1} \\cup \\dots \\mathcal {D} _ {t} ^ {b}} \\mathbb {I} (y \\neq f (\\mathbf {x})). \\tag {1}\n$$\n", + "text_format": "latex", + "bbox": [ + 558, + 877, + 903, + 902 + ], + "page_idx": 1 + }, + { + "type": "page_number", + "text": "3315", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 1 + }, + { + "type": "text", + "text": "In Eq. 1, $\\mathcal{H}$ represents the hypothesis space and $\\mathbb{I}(\\cdot)$ is the indicator function. $\\mathcal{D}_t^b$ denotes the $b$ -th task's data distribution. In this paper, we follow [25, 76, 86] and assume that a pre-trained CLIP model (contrastive language-image pre-training) [49] is available as the initialization for $f(\\mathbf{x})$ . Specifically, CLIP contains the visual and textual encoder, i.e., $g_i(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^d$ , $g_t(\\cdot): \\mathbb{R}^{Dt} \\to \\mathbb{R}^d$ , where the $D / Dt$ -dimensional images and texts are projected to the same $d$ -dimensional embedding space. During inference, it constructs the template text $\\mathbf{t}_i$ using the class name, i.e., \"a photo of a $[\\mathrm{CLASS}]_i$ \", and matches the query image to text embeddings of all classes:", + "bbox": [ + 89, + 90, + 483, + 272 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\begin{array}{l} f _ {y _ {i}} (\\mathbf {x}, \\mathbf {t} _ {i}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {j}\\right) / \\tau\\right)} \\tag {2} \\\\ = \\frac {\\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}, \\\\ \\end{array}\n$$\n", + "text_format": "latex", + "bbox": [ + 122, + 279, + 482, + 358 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\cos (\\cdot ,\\cdot)$ is cosine similarity and $\\tau$ is temperature. $\\mathbf{z} = g_{i}(\\mathbf{x})$ and $\\mathbf{w}_i = g_t(\\mathbf{t}_i)$ are visual and textual features, respectively. In Eq. 2, the logit is assigned by the relative similarity of the visual embedding to textual embedding.", + "bbox": [ + 89, + 364, + 483, + 426 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "3.2. Baselines in Class-Incremental Learning", + "text_level": 1, + "bbox": [ + 89, + 433, + 437, + 450 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "To overcome forgetting, there are two typical solutions for CIL with pre-trained models by learning prompts [28, 88].", + "bbox": [ + 89, + 455, + 482, + 486 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Visual Prompts: As fully-finetuning will harm the generalizability of the PTM, several works [28, 53, 64, 65] propose to freeze the pre-trained visual encoder and append visual prompts for subsequent tasks. They design a prompt pool to select instance-specific prompts, with which the visual embedding is represented as:", + "bbox": [ + 89, + 486, + 483, + 575 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {z} = \\bar {g} _ {i} (\\mathbf {x}, \\mathcal {P}), \\tag {3}\n$$\n", + "text_format": "latex", + "bbox": [ + 233, + 585, + 482, + 601 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{g}_i$ is the frozen image encoder, $\\mathcal{P}$ is the prompt pool for prompt selection. By replacing $g_i(\\mathbf{x})$ in Eq. 2 with Eq. 3, the model efficiently encodes downstream information into the visual prompts and tackles forgetting.", + "bbox": [ + 89, + 609, + 482, + 670 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Learning Textual Prompts: Since CLIP contains two branches for the visual and textual encoder, there are also works learning textual prompts instead of using the template text. Specifically, the textual input is formulated into a set of learnable prompts in CoOp [88], i.e., $\\tilde{\\mathbf{t}}_i = [\\mathrm{V}]_1[\\mathrm{V}]_2\\cdots[\\mathrm{V}]_M[\\mathrm{CLASS}]_i$ . In this way, the textual embedding is denoted as:", + "bbox": [ + 89, + 671, + 483, + 776 + ], + "page_idx": 2 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {w} _ {i} = \\bar {g} _ {t} (\\tilde {\\mathbf {t}} _ {i}), \\tag {4}\n$$\n", + "text_format": "latex", + "bbox": [ + 236, + 784, + 482, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "where $\\bar{g}_t$ is the frozen text encoder. Similarly, we can replace the textual embedding in Eq. 2 via Eq. 4 to encode task information into these learnable prompts.", + "bbox": [ + 89, + 810, + 482, + 854 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Discussions: Eq. 3 and Eq. 4 focus on different aspects to adjust the pre-trained CLIP into downstream tasks. However, when adapting the model with only visual prompts,", + "bbox": [ + 89, + 854, + 483, + 900 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "the problem occurs since the model has to decompose the class name into fine-grained concepts to match the visual features. The sequential updating process will cause the degradation of such decomposition ability, resulting in the mismatch between visual and textual features. By contrast, if we utilize textual prompts, the model is only focused on the specific training instance and lacks holistic information about the specific class. When the testing data contains the image of the same class with different distributions, the model shall be confused due to the prompt-level overfitting. Hence, the ideal text input should be general and holistic to match the visual features and convey all informative features that the target class possesses.", + "bbox": [ + 511, + 90, + 906, + 287 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4. ENGINE: External Knowledge Injection", + "text_level": 1, + "bbox": [ + 511, + 300, + 872, + 316 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Noting that prompt-based methods are highly limited by the training instances, we need to extract comprehensive information for the model to adapt to downstream tasks while preventing forgetting. To encode holistic information into the model, we seek help from large language models like GPT-4 [1] to provide the general descriptions instead of using the template features. Apart from using detailed descriptions in the training process, we also make full use of them during inference by seeking pair-wise discriminative features. On the other hand, as external knowledge is continually injected into the model, we also design the injection unit expansion and prototype replay strategy. In the following sections, we first introduce how to inject external knowledge during model training and then introduce how to utilize it during inference.", + "bbox": [ + 511, + 325, + 906, + 551 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "4.1. On-the-fly Knowledge Injection", + "text_level": 1, + "bbox": [ + 511, + 560, + 794, + 575 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Since the target is to encode general information into the CLIP during incremental learning, we first design the model training pipeline that can enhance recognition. For example, a \"cat\" contains visual features like \"soft, short fur\", \"long, thin tail\", and \"round face with large eyes\". It would be better to match the cat image to these visual features than \"a photo of a cat\", since those features contain more discriminative and fine-grained information. Correspondingly, to enhance CLIP's knowledge of general class description, we seek help from GPT-4 [1] to provide discriminative visual features:", + "bbox": [ + 511, + 582, + 906, + 748 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "Q: What are unique visual features of $[\\mathrm{CLASS}]_i$ in a photo? Focus on the key visual features.", + "bbox": [ + 544, + 756, + 870, + 786 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "A: 1. Long, thin tail that aids in balance. 2.", + "bbox": [ + 544, + 786, + 849, + 801 + ], + "page_idx": 2 + }, + { + "type": "text", + "text": "In this way, we can get a set of visual descriptions for each class, denoted as $\\mathbf{d}_i$ . A naive solution is to replace the template text $\\mathbf{t}_i$ into $\\mathbf{d}_i$ during training, while we propose a more intuitive way to inject the knowledge into the model. Specifically, we design a knowledge injection unit $u_t$ to dynamically encode these textual descriptors. We implement", + "bbox": [ + 511, + 810, + 906, + 900 + ], + "page_idx": 2 + }, + { + "type": "page_number", + "text": "3316", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 2 + }, + { + "type": "image", + "img_path": "images/6908e6de466c635a8ac1e49078c1e9325cb9036237e3cecd611ac12ae3ebb1a5.jpg", + "image_caption": [ + "(a) On-the-fly Knowledge Injection" + ], + "image_footnote": [], + "bbox": [ + 112, + 68, + 612, + 252 + ], + "page_idx": 3 + }, + { + "type": "image", + "img_path": "images/a28fd7ad1910454bf814df56f2412adf9cbad31e548508ae144600f62b9d0d14.jpg", + "image_caption": [ + "(b) Post-tuning Knowledge Injection", + "Figure 2. Illustration of ENGINE. Left: On-the-fly knowledge injection. We utilize GPT-4 to provide visual features of each class, and append injection units to encode external knowledge into the model. We also adopt random augmentation to enhance image diversity using the same network structure. To overcome forgetting, we append the prototypes of previous classes in each learning stage. Right: Post-tuning knowledge injection. We extract top- $k$ predictions and generate pair-wise discriminative features locally. The input image is forwarded with every textual feature to further refine the prediction." + ], + "image_footnote": [], + "bbox": [ + 622, + 73, + 885, + 253 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "$u_{t}$ with a linear layer, i.e., $u_{t}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}$ , and append it after the textual encoder. We have the similarity loss:", + "bbox": [ + 89, + 345, + 483, + 377 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {t} = - \\operatorname {S i m} \\left(u _ {t} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right), \\bar {g} _ {t} \\left(\\mathbf {d} _ {i}\\right)\\right), \\tag {5}\n$$\n", + "text_format": "latex", + "bbox": [ + 168, + 386, + 482, + 405 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "where we use cosine similarity to calculate $\\mathrm{Sim}(\\cdot, \\cdot)$ . In Eq. 5, we maximize the similarity between $u_{t}\\left(\\bar{g}_{t}\\left(\\mathbf{t}_{i}\\right)\\right)$ and $\\bar{g}_{t}\\left(\\mathbf{d}_{i}\\right)$ , i.e., making the adapted textual feature similar to external descriptions. Since the textual encoder $\\bar{g}_{t}$ is frozen, we can inject this external information into the injection unit $u_{t}$ . In this way, the extracted textual features will contain detailed descriptions even using the template text, e.g., highlighting the features of \"long thin tail\" and \"round face\" when the input is \"a photo of a cat\". In the implementation, since GPT-4 will output a set of descriptions, we randomly choose one of the descriptions $\\mathbf{d}_{i}$ to calculate Eq. 5 in each iteration.", + "bbox": [ + 89, + 412, + 483, + 593 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning external visual knowledge: Similar to the textual branch, we can also inject visual information into the model during training. Given that we cannot fetch external data in the learning process, we utilize random data augmentation to adjust the input images, and denote the augmented input as $\\mathcal{A}(\\mathbf{x})$ . Correspondingly, we can learn the image knowledge injection unit $u_{i}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}$ , and maximize the similarity between augmented and vanilla input:", + "bbox": [ + 89, + 594, + 483, + 715 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {i} = - \\operatorname {S i m} \\left(u _ {i} (\\bar {g} _ {i} (\\mathbf {x})), \\bar {g} _ {i} (\\mathcal {A} (\\mathbf {x}))\\right). \\tag {6}\n$$\n", + "text_format": "latex", + "bbox": [ + 161, + 724, + 482, + 741 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Similar to Eq. 5, Eq. 6 freezes the visual encoder $\\bar{g}_i$ , and only optimizes the image injection unit $u_i$ . In the matching process, the model maximizes the similarity of the adapted visual feature to the augmented view, which helps the model extract diverse visual information.", + "bbox": [ + 89, + 750, + 482, + 824 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Learning task-specific injection unit: In CIL, new tasks will emerge as data evolves, and sequentially tuning a single injection unit will result in forgetting previous knowledge. Consequently, we conduct the injection process for each incremental task to avoid forgetting. Specifically, we", + "bbox": [ + 89, + 825, + 483, + 901 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "initialize a new injection unit $u_{t}^{b}, u_{i}^{b}$ when learning the $b$ -th task, and freeze all previous injection units when learning the current task. With the set of visual and textual units, we utilize the summation of them for the final representation:", + "bbox": [ + 511, + 345, + 906, + 407 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {i} (\\mathbf {x}) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {i} ^ {p} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) + u _ {i} ^ {b} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) \\qquad (7)\n$$\n", + "text_format": "latex", + "bbox": [ + 571, + 412, + 906, + 436 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nG _ {t} \\left(\\mathbf {t} _ {i}\\right) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {t} ^ {p} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) + u _ {t} ^ {b} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) \\tag {8}\n$$\n", + "text_format": "latex", + "bbox": [ + 568, + 441, + 905, + 462 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In the original CLIP, we have the cross-modal contrastive loss to match the image to the corresponding text description. Correspondingly, we utilize the injected visual and textual features to calculate the contrastive loss:", + "bbox": [ + 511, + 464, + 905, + 523 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathcal {L} _ {c} = \\ell \\left(f _ {\\text {i n j}} (\\mathbf {x}), y\\right) \\quad \\text {w h e r e} \\tag {9}\n$$\n", + "text_format": "latex", + "bbox": [ + 570, + 532, + 905, + 550 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\text {i n j}, y _ {i}} (\\mathbf {x}) = \\frac {\\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}) , G _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {V} _ {b} |} \\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}), G _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}. \\tag {10}\n$$\n", + "text_format": "latex", + "bbox": [ + 529, + 551, + 903, + 590 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "In Eq. 9, we utilize the injected visual and textual features to calculate the contrastive loss, aiming to increase the similarity of pair-wise injected features.", + "bbox": [ + 511, + 595, + 905, + 641 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "Preventing forgetting of injection unit In the updating process, we combine Eq. 5, Eq. 6, and Eq. 9 to update the model. Since we cannot hold historical instances for replay, the model may still forget previous concepts due to the sequential updating. To this end, utilizing the substitution of previous concepts can create a calibration across all seen classes. Observing that the visual backbone $g_{i}$ is not updating throughout the learning process, we generate a subset of previous classes using visual prototypes. Specifically, during the learning process, we can calculate the average visual embedding of each class as:", + "bbox": [ + 511, + 642, + 906, + 808 + ], + "page_idx": 3 + }, + { + "type": "equation", + "text": "\n$$\n\\mathbf {p} _ {k} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right) g _ {i} \\left(\\mathbf {x} _ {j}\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right)}, \\tag {11}\n$$\n", + "text_format": "latex", + "bbox": [ + 620, + 814, + 903, + 849 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "which stands for the most representative features of the corresponding class. We can treat them as the visual embedding of previous classes when learning new", + "bbox": [ + 511, + 854, + 905, + 901 + ], + "page_idx": 3 + }, + { + "type": "page_number", + "text": "3317", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 3 + }, + { + "type": "text", + "text": "ones, i.e., we construct the auxiliary training set $\\mathbf{P} = \\{(\\mathbf{p}_1, 1), (\\mathbf{p}_2, 2), \\dots, (\\mathbf{p}_{|\\mathcal{Y}_b|}, |\\mathcal{Y}_b|)\\}$ , and optimize:", + "bbox": [ + 89, + 90, + 482, + 122 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\min _ {\\left\\{u _ {t} ^ {b}, u _ {i} ^ {b} \\right\\}} \\sum_ {\\left(\\mathbf {x}, y\\right) \\in \\mathcal {D} ^ {b} \\cup \\mathbf {P}} \\mathcal {L} _ {t} + \\mathcal {L} _ {i} + \\mathcal {L} _ {c}. \\tag {12}\n$$\n", + "text_format": "latex", + "bbox": [ + 148, + 132, + 482, + 152 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Specifically, the prototype set $\\mathbf{P}$ is only adopted in calculating $\\mathcal{L}_c$ . To enhance diversity across training stages, we also add Gaussian noise on the prototype embeddings, i.e., $\\mathbf{p}' = \\mathbf{p} + \\epsilon$ , $\\epsilon \\in \\mathcal{N}(0,\\alpha^2\\mathbf{I})$ , and utilize $\\mathbf{p}'$ in Eq. 12. In this way, we can revisit previous knowledge to prevent forgetting when learning new injection units. Additionally, we also consider matching the visual features to these prototypes to further assist recognition [66, 86].", + "bbox": [ + 89, + 159, + 483, + 280 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Summary of on-the-fly injection: We visualize the training stage knowledge injection in Figure 2 (left). ENGINE learns two sets of knowledge injection units, aiming to encode the extra knowledge into the unit module. We utilize GPT-4 to extract informative class descriptors, and utilize similarity loss to encode them into the injection units. Similarly, data augmentation is utilized to provide visual information. Hence, the model can extract more informative features as it sees more tasks. To prevent forgetting, we also design the learning objective considering class prototypes. Since injection units are implemented with a single layer, the appended parameter size $(d\\times d)$ per unit is negligible for CLIP. Besides, since injection units are linear layers, they can be further re-parameterized by aggregating weights into a single one, i.e., $\\sum_{p}u_{i}^{p}$ and $\\sum_{p}u_{t}^{p}$ .", + "bbox": [ + 89, + 281, + 483, + 511 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.2. Post-tuning Knowledge Injection", + "text_level": 1, + "bbox": [ + 89, + 516, + 377, + 532 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "We have introduced the way to inject external knowledge during training in Section 4.1. Apart from the model training stage, we can also utilize external knowledge to calibrate the prediction results during inference, and we call this step \"post-tuning knowledge injection\". For example, after the model makes inference by matching visual and textual features, we can treat the current prediction results as a preliminary result, which needs further rectification. For the top- $k$ prediction results in this preliminary result, we treat them as competitors to the ground truth that could result in confusion. Hence, we need to construct a further prediction considering the image and all local competitors, which can focus more on the top- $k$ classes to refine the prediction.", + "bbox": [ + 89, + 537, + 482, + 734 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Refining the results: After learning the injection units, we can get the prediction results using the injected features for more comprehensive features. However, the model may still make wrong predictions, e.g., classifying a \"cat\" into a similar class \"lion\" since they have similar visual features. To refine the prediction results, we design an extra consolidation process using external knowledge. Denoting the prediction results of injected feature matching in Eq. 10 as $f_{\\mathrm{inj}}(\\mathbf{x})$ , we first extract the top- $k$ predictions as the competitor candidates, i.e., $\\{f_{\\mathrm{inj},i_1}(\\mathbf{x}), f_{\\mathrm{inj},i_2}(\\mathbf{x}), \\dots, f_{\\mathrm{inj},i_k}(\\mathbf{x})\\} = \\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x}))$ , where $\\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x}))$ indicates selecting the", + "bbox": [ + 89, + 734, + 483, + 902 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "top- $k$ largest value in $f_{\\mathrm{inj}}(\\mathbf{x})$ . Their labels are denoted as:", + "bbox": [ + 511, + 90, + 892, + 107 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\n\\left\\{y _ {i _ {1}}, y _ {i _ {2}}, \\dots , y _ {i _ {k}} \\right\\} = \\left\\{y _ {i} \\mid f _ {\\text {i n j}, i} (\\mathbf {x}) \\in \\operatorname {T o p} _ {k} \\left(f _ {\\text {i n j}} (\\mathbf {x})\\right) \\right\\}. \\tag {13}\n$$\n", + "text_format": "latex", + "bbox": [ + 524, + 112, + 903, + 143 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In Eq. 13, the label set $\\{y_{i_1}, y_{i_2}, \\ldots, y_{i_k}\\}$ corresponds to the possible confusing labels for the input image. To further refine the predictions, we aim for pair-wise disambiguation among this label set. We again utilize GPT-4 to generate pair-wise discriminative features:", + "bbox": [ + 511, + 143, + 903, + 219 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Q: What are unique visual features of $[\\mathrm{CLASS}]_i$ compared to $[\\mathrm{CLASS}]_j$ in a photo? Focus on their key visual differences.", + "bbox": [ + 544, + 224, + 872, + 270 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "A: [CLASS]i: 1. Long, thin tail that aids in balance... [CLASS]j: 1. Stockier, heavier tail with a tuft of hair at the end ...", + "bbox": [ + 544, + 271, + 872, + 315 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "With this process, GPT-4 provides the most discriminative features that can differentiate two classes, e.g., the differences in \"tail\" to differentiate \"cat\" from \"lion\" and vice versa. Given the top- $k$ classes, we can generate their pairwise descriptors as $\\mathbf{D} = [\\mathbf{d}_{ij}]_{i,j=1}^{k}$ , where $\\mathbf{d}_{ij}$ denotes the descriptions to differ class $i$ from class $j$ . We feed these fine-grained descriptions as the textual input, and match the query embedding to them:", + "bbox": [ + 511, + 321, + 906, + 445 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf _ {\\mathrm {p t}, y _ {i}} (\\mathbf {x}) = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {j \\neq i} f _ {y _ {i}} \\left(\\mathbf {x}, \\mathbf {d} _ {i j}\\right), \\tag {14}\n$$\n", + "text_format": "latex", + "bbox": [ + 591, + 450, + 903, + 494 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "where we utilize zero-shot CLIP in Eq. 2 to match the query image to these pair-wise discriminative features, and average the similarities as the post-tuning logits.", + "bbox": [ + 511, + 501, + 903, + 545 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "Effect of post-tuning knowledge injection: We visualize the post-tuning knowledge injection in Figure 2 (right). To discriminate similar concepts and refine the wrong predictions, we first extract the top- $k$ predictions of the model, and utilize GPT-4 to generate pair-wise discriminative features locally. Eq. 14 provides the fine-grained local prediction among these top- $k$ classes, which can further help refine the results as data evolves. We adopt the zero-shot CLIP for Eq. 14, which does not require incremental updates.", + "bbox": [ + 511, + 546, + 905, + 681 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "4.3. Summary of ENGINE", + "text_level": 1, + "bbox": [ + 511, + 689, + 705, + 704 + ], + "page_idx": 4 + }, + { + "type": "text", + "text": "In ENGINE, we utilize external knowledge to help continual learning in two aspects, i.e., on-the-fly injection and post-tuning injection. With the help of GPT-4, we can extract more informative descriptors suitable for the incremental learning classes, and inject the external knowledge into the knowledge units. Besides, we also extract local descriptors to discriminate pair-wise classes in the post-tuning stage to help refine the predictions. During training, we update the model via Eq. 12 to inject knowledge into the knowledge units. During inference, we aggregate the following outputs as the prediction:", + "bbox": [ + 511, + 710, + 906, + 878 + ], + "page_idx": 4 + }, + { + "type": "equation", + "text": "\n$$\nf (\\mathbf {x}) = f _ {\\text {i n j}} (\\mathbf {x}) + f _ {\\mathrm {p t}} (\\mathbf {x}). \\tag {15}\n$$\n", + "text_format": "latex", + "bbox": [ + 624, + 885, + 903, + 902 + ], + "page_idx": 4 + }, + { + "type": "page_number", + "text": "3318", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 4 + }, + { + "type": "table", + "img_path": "images/57a5e018adaf9e4df21f499ae70dd72b6b99039adf90e25ed74554d6a9f8aaef.jpg", + "table_caption": [ + "Table 1. Average and last performance comparison of different methods. The best performance is shown in bold. All methods are initialized with the same pre-trained CLIP without exemplars for a fair comparison." + ], + "table_footnote": [], + "table_body": "
MethodAircraftCIFAR100Cars
B0 Inc10B50 Inc10B0 Inc10B50 Inc10B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)
Finetune3.160.961.721.057.844.445.302.463.141.101.541.13
CoOp [88]14.547.1413.057.7747.0024.2441.2324.1236.4621.6537.4020.87
SimpleCIL [84]59.2448.0953.0548.0984.1576.6380.2076.6392.0486.8588.9686.85
ZS-CLIP [49]26.6617.2221.7017.2281.8171.3876.4971.3882.6076.3778.3276.37
L2P [65]47.1928.2944.0732.1382.7473.0381.1473.6176.6361.8276.3765.64
DualPrompt [64]44.3025.8346.0733.5781.6372.4480.1272.5776.2662.9476.8867.55
CODA-Prompt [53]45.9827.6945.1432.2882.4373.4378.6971.5880.2166.4775.0664.19
RAPF [25]50.3823.6140.4725.4486.1478.0482.1777.9382.8962.8575.8763.19
ENGINE69.6958.6964.3859.0286.9279.2283.1579.4794.1490.0891.6190.03
ImageNet-RCUBUCF
MethodB0 Inc20B100 Inc20B0 Inc20B100 Inc20B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(A_B\\)
Finetune1.370.431.010.882.060.640.560.474.511.591.210.80
CoOp [88]60.7337.5254.2039.7727.618.5724.0310.1447.8533.4642.0224.74
SimpleCIL [84]81.0674.4876.8474.4883.8177.5279.7577.5290.4485.6888.1285.68
ZS-CLIP [49]83.3777.1779.5777.1774.3863.0667.9663.0675.5067.6471.4467.64
L2P [65]75.9766.5272.8266.7770.8757.9375.6466.1286.3476.4383.9576.62
DualPrompt [64]76.2166.6573.2267.5869.8957.4674.4064.8485.2175.8284.3176.35
CODA-Prompt [53]77.6968.9573.7168.0573.1262.9873.9562.2187.7680.1483.0475.03
RAPF [25]81.2670.4876.1070.2379.0962.7772.8262.9392.2880.3390.3181.55
ENGINE86.2280.3783.6380.9886.6580.2082.5979.3094.3590.0392.5189.58
SUNFoodObjectNet
MethodB0 Inc30B150 Inc30B0 Inc10B50 Inc10B0 Inc20B100 Inc20
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{\\mathcal{A}}\\)\\(\\mathcal{A}_B\\)
Finetune4.511.590.780.723.491.712.141.521.340.470.690.54
CoOp [88]45.9323.1139.3324.8936.0114.1833.1318.6721.246.2916.216.82
SimpleCIL [84]82.1375.5878.6275.5887.8981.6584.7381.6552.0640.1345.1140.13
ZS-CLIP [49]79.4272.1174.9572.1187.8681.9284.7581.9238.4326.4331.1226.43
L2P [65]82.8274.5479.5773.1085.6677.3380.4273.1351.4039.3948.9142.83
DualPrompt [64]82.4674.4079.3773.0284.9277.2980.0072.7552.6240.7249.0842.92
CODA-Prompt [53]83.3475.7180.3874.1786.1878.7880.9874.1346.4934.1340.5734.13
RAPF [25]82.1372.4778.0473.1088.5781.1585.5381.1748.6727.4339.2828.73
ENGINE85.0478.5481.5778.4589.8183.8986.8983.9459.1145.1951.3244.99
", + "bbox": [ + 127, + 93, + 861, + 599 + ], + "page_idx": 5 + }, + { + "type": "table", + "img_path": "images/bb7df7ed0b60ebf1ca741f73e2c1f217a5d9ea06c838ff37f070c1c2a7dfb23a.jpg", + "table_caption": [ + "Table 2. Comparison to traditional exemplar-based CIL methods. ENGINE does not use any exemplars." + ], + "table_footnote": [], + "table_body": "
MethodExemplarsAircraft B0 Inc10SUN B0 Inc30
AABAAB
iCaRL [51]20 / class53.6043.9878.5667.30
MEMO [83]20 / class42.2425.4181.4873.45
PROOF [86]20 / class61.0053.5983.5777.28
ENGINE069.6958.6985.0478.54
", + "bbox": [ + 91, + 642, + 473, + 734 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5. Experiments", + "text_level": 1, + "bbox": [ + 89, + 752, + 223, + 768 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "In this section, we conduct extensive experiments on nine benchmark datasets and compare ENGINE to state-of-the-art methods. We visualize the incremental learning curve and provide ablation studies on prompt engineering and parameter sensitivity to test the model's robustness. Besides, we also visualize the learned embeddings and logit changes after injecting external knowledge to show its effectiveness. We provide more results in the supplementary.", + "bbox": [ + 89, + 780, + 482, + 902 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "5.1. Implementation Details", + "text_level": 1, + "bbox": [ + 511, + 625, + 730, + 641 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset: We follow [65, 86, 88] to evaluate the performance on nine benchmark datasets that have domain gap to CLIP's pre-training dataset, i.e., CIFAR100 [32], CUB200 [59], ObjectNet [6], ImageNet-R [22], FGVCAircraft [41], StanfordCars [31], Food101 [7], SUN397 [68] and UCF101 [55]. Following [86], we utilize the sampled 100 classes from CIFAR100, Aircraft, Cars, Food, UCF, 200 classes from CUB200, ObjectNet, ImageNet-R, and 300 classes from SUN to ease the data split. More details are reported in the supplementary.", + "bbox": [ + 511, + 652, + 906, + 805 + ], + "page_idx": 5 + }, + { + "type": "text", + "text": "Dataset split: Following [51, 65], we use 'B- $m$ Inc- $n$ ' to split the classes in CIL. $m$ indicates the number of classes in the first stage, and $n$ represents that of every following stage. We follow [51] to randomly shuffle the class order with random seed 1993 for all compared methods, and keep this same for every method.", + "bbox": [ + 511, + 809, + 908, + 901 + ], + "page_idx": 5 + }, + { + "type": "page_number", + "text": "3319", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 5 + }, + { + "type": "image", + "img_path": "images/224d58b8d4e13ed74089cdd54c4c30a3bb1888d6cb98c92389b491af76175e92.jpg", + "image_caption": [ + "(a) Aircraft Base0 Inc10" + ], + "image_footnote": [], + "bbox": [ + 94, + 65, + 359, + 223 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/facb6e73795a87063d68e79b97490b31620b7ff86f510bfc888d025613de2a1a.jpg", + "image_caption": [ + "(b) Cars Base0 Inc10" + ], + "image_footnote": [], + "bbox": [ + 366, + 65, + 633, + 222 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/42cd25a8b36a24d9279ae8e65edd0c7c3670e31cd0f8f2a681017f3cd0d1e4ce.jpg", + "image_caption": [ + "(c) UCF Base0 Inc10" + ], + "image_footnote": [], + "bbox": [ + 638, + 66, + 901, + 222 + ], + "page_idx": 6 + }, + { + "type": "table", + "img_path": "images/1de9d8b665a18aa85c4883a0af9618edf24763af17eedfb38e5abafe85e6036a.jpg", + "table_caption": [ + "Table 3. Results when all methods use the same textual description generated by GPT-4 and data augmentation." + ], + "table_footnote": [], + "table_body": "
MethodTemplateImageNet-R AB0 Inc20 A_BCIFAR AB0 Inc10 A_B
ZS-CLIPGPT Generated83.6877.5982.2671.70
RAPFGPT Generated81.8971.0486.4978.52
ENGINEGPT Generated86.2280.3786.9279.22
", + "bbox": [ + 91, + 303, + 475, + 375 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Comparison methods: We first compare to SOTA pretrained model-based CIL algorithms, e.g., L2P [65], DualPrompt [64], CODA-Prompt [53], SimpleCIL [84]. Besides, we also compare to SOTA CLIP-based CIL algorithms, e.g., CoOp [88], PROOF [86], RAPF [25]. The baseline that finetunes CLIP for incremental tasks is denoted as Finetune. All methods are deployed with the same CLIP as initialization.", + "bbox": [ + 88, + 383, + 482, + 502 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Training details: The experiments are deployed with NVIDIA 4090 using PyTorch [47]. We follow [25, 86] to consider CLIP with ViT-B/16 for all compared methods for fair comparison. For vision-based methods that cannot utilize the textual prompt (e.g., L2P, DualPrompt, CODA-Prompt), we utilize CLIP's visual branch as their initialization. We report the results using LAION-400M pre-trained CLIP [26] in the main paper, and report results using OpenAI [49] in the supplementary. In ENGINE, we use SGD optimizer with a batch size of 64 to optimize the model for 10 epochs. The learning rate decays from 0.05 with cosine annealing. We set the prototype noise ratio $\\alpha$ to 0.25 and $k$ in post-tuning to 5. We use OpenAI GPT-4o mini [1] to generate the prompts for textual knowledge and AutoAugment [11] to augment the visual features. We explore other LLMs to extract external knowledge in the supplementary. The source code will be publicly available upon acceptance. Evaluation metric: Following [51, 86], we use $\\mathcal{A}_b$ to represent the model's accuracy after the $b$ -th stage. Specifically, we adopt $\\mathcal{A}_B$ (the performance after the last stage) and $\\bar{\\mathcal{A}} = \\frac{1}{B}\\sum_{b=1}^{B}\\mathcal{A}_b$ (average performance along incremental stages) as measurements.", + "bbox": [ + 89, + 503, + 483, + 838 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "5.2. Benchmark Comparison", + "text_level": 1, + "bbox": [ + 89, + 848, + 316, + 864 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "We first compare ENGINE to other state-of-the-art methods on the benchmark datasets and report the results in Table 1", + "bbox": [ + 89, + 869, + 482, + 900 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/ec1ee823ba1959f74917c9e66162db42f244bde62fd25d9b84e8cc63a2e008f6.jpg", + "image_caption": [ + "Figure 3. Incremental performance of different methods. We report the performance gap after the last incremental stage of ENGINE and the runner-up method at the end of the line. All methods utilize the same CLIP pre-trained weight. More figures are shown in supplementary.", + "(a) Ablation study" + ], + "image_footnote": [], + "bbox": [ + 517, + 275, + 705, + 390 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/a3393fed7b92a9574c5b020a2468f0c4ee24cf6e8d599c845be95894a3165b7f.jpg", + "image_caption": [ + "(b) Parameter sensitivity", + "Figure 4. Ablation study and parameter sensitivity." + ], + "image_footnote": [], + "bbox": [ + 709, + 275, + 898, + 391 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "and Figure 3. Specifically, we find ENGINE outperforms current SOTA by $3\\% \\sim 10\\%$ on these benchmark datasets. We find the baseline method Finetune has the worst performance, indicating that the decomposition ability to tackle the class template into fine-grained features is totally forgotten. For visual prompt-based methods like L2P, DualPrompt, and CODA-Prompt, the performance is limited due to the inability to utilize textual information. For the textual prompt tuning method CoOp, we find that it achieves poor performance due to the forgetting of textual prompts. For other CLIP-based methods like RAPF, our method still shows substantial improvement, verifying its strong continual learning ability without forgetting.", + "bbox": [ + 511, + 448, + 906, + 647 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Apart from these non-exemplar-based methods, we also consider typical exemplar-based CIL methods for comparison. We report the comparison results in Table 2, where all methods are initialized with the same CLIP. As we can infer from the table, ENGINE still outperforms them even without using exemplars.", + "bbox": [ + 511, + 652, + 906, + 743 + ], + "page_idx": 6 + }, + { + "type": "text", + "text": "Finally, what if all CLIP-based methods utilize the same external knowledge? Correspondingly, we also consider using the same generated prompt and data augmentation for ZS-CLIP and RAPF, and report the results in Table 3. In this table, we replace the template text of all methods using the same text in Section 4.1 and adopt the same data augmentation for image input. We find ENGINE still outperforms them even using the same prompt, indicating that ENGINE can better utilize the external information for incremental learning.", + "bbox": [ + 511, + 750, + 908, + 901 + ], + "page_idx": 6 + }, + { + "type": "page_number", + "text": "3320", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 6 + }, + { + "type": "image", + "img_path": "images/de85726447bf967dad8ed2129b75a3a536f4fd270e8d47c2bc670b116618223b.jpg", + "image_caption": [ + "(a) ZS-CLIP, single stage" + ], + "image_footnote": [], + "bbox": [ + 96, + 71, + 282, + 181 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/ab615e14df2c4a9549f04dd7caf47bb2ce428e22913d931db6fec55654ca7a54.jpg", + "image_caption": [ + "(b) ENGINE, two stages", + "Figure 5. t-SNE [58] visualizations of zero-shot CLIP and ENGINE. ENGINE aligns the cross-modal features without forgetting. We represent the visual features of the first task with dots and classes of the second task with squares. The textual features are represented by triangles, and we utilize the shadow area to denote the decision boundary." + ], + "image_footnote": [], + "bbox": [ + 287, + 71, + 475, + 181 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "5.3. Further Analysis", + "text_level": 1, + "bbox": [ + 89, + 297, + 256, + 314 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Ablation study: We first conduct ablations to investigate the effectiveness of each component in ENGINE. We report the incremental performance on CIFAR100 B0 Inc10 in Figure 4a. We can infer that 'ZS-CLIP' has the worst performance since the downstream distribution is different from the pre-training stage, and we view it as the baseline. Correspondingly, we first equip the model with visual knowledge injection (Eq. 6), and denote the model as 'w/ Visual Injection'. It shows that the performance drastically improves as the model has been adapted to extract more informative visual features. Similarly, we then append the textual information using Eq. 5, denoted as 'w/ Dual Injection.' As shown in the figure, using external knowledge for both modalities can further improve the performance than visual knowledge only. Finally, we equip the model with post-tuning knowledge injection, and utilize Eq. 15 for inference (i.e., 'w/ Post-tuning Injection'). This corresponds to the full version of ENGINE, and we find it achieves the best performance. Ablations verify that every component in ENGINE boosts the CIL performance.", + "bbox": [ + 89, + 321, + 483, + 625 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Parameter robustness: There are two hyperparameters in ENGINE, i.e., the prototype noise scale $\\alpha$ in Section 4.1 and post-tuning candidate $k$ in Eq. 14. We conduct experiments on CUB B0 Inc20 to investigate the robustness by changing these parameters. Specifically, we choose $\\alpha$ among $\\{0.15, 0.2, 0.25, 0.3, 0.35\\}$ , and $k$ among $\\{3, 4, 5, 6, 7\\}$ . We report the last performance in Figure 4b. As shown in the figure, the performance is robust with the change of these parameters, and we suggest $\\alpha = 0.25$ , $k = 5$ as default.", + "bbox": [ + 89, + 627, + 483, + 763 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Visualizations: In this section, we utilize t-SNE [58] to visualize the cross-modal features learned by ENGINE on CIFAR100 B0 Inc5. We represent the visual features of the first task with dots $(\\bigcirc)$ and classes of the second task with squares $(\\square)$ . The textual features are represented by triangles $(\\triangle)$ , and we utilize the shadow area to denote the decision boundary divided by the textual features. Specifically, we first visualize the instances of the first stage in zero-shot CLIP's embedding space in Figure 5a. The results indicate", + "bbox": [ + 89, + 763, + 483, + 901 + ], + "page_idx": 7 + }, + { + "type": "image", + "img_path": "images/6f24666e94e7d1c0d2e221513cfe0f00eee579da00e50e37e3cfe638fc1ccc28.jpg", + "image_caption": [ + "Figure 6. Left: Input. Middle: Top-5 predictions before post-tuning knowledge injection. Right: Top-5 predictions after post-tuning. More cases are shown in the supplementary." + ], + "image_footnote": [], + "bbox": [ + 517, + 71, + 903, + 273 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "that pre-trained CLIP has difficulty in aligning the visual and textual features since they are situated in two distinct areas in the embedding space [37]. While for Figure 5b, we visualize the embedding of ENGINE across two incremental stages. As we can infer from the figure, ENGINE has two advantages: 1) The visual and textual features are aligned in the same cluster, and 2) The model has not forgotten previous classes when learning new ones.", + "bbox": [ + 511, + 340, + 906, + 462 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "We also visualize the model output before and after post-tuning knowledge injection on ImageNet-R in Figure 6. We utilize bars with red edges to denote the ground-truth class. As we can infer from the figure, post-tuning knowledge injection can help the model refine the predictions by generating local comparisons. For the case on top of Figure 6, the model can focus on the differences between \"acorn\" and \"birdhouse\" to make the final prediction.", + "bbox": [ + 511, + 465, + 908, + 588 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "6. Conclusion", + "text_level": 1, + "bbox": [ + 513, + 614, + 633, + 628 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Class-incremental learning is essential to real-world applications. This paper aims to enable CLIP with CIL ability by injecting external knowledge. Specifically, we consider on-the-fly knowledge injection by learning dual-branch injection units. We augment images for the visual branch and generate detailed feature descriptions with GPT-4 for the textual branch. By maximizing the similarity between external information and the injection module, we encode informative features into the pre-trained CLIP. Besides, we also consider post-tuning knowledge injection with local class information to calibrate the predictions. Extensive experiments verify ENGINE's effectiveness.", + "bbox": [ + 511, + 643, + 906, + 824 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Limitations and future works: This paper relies on GPT-4 to provide expert knowledge for specific tasks, which may be impossible for the specific cases where GPT-4 cannot generalize. Future works include designing other formats of external knowledge, e.g., knowledge graphs.", + "bbox": [ + 511, + 825, + 908, + 902 + ], + "page_idx": 7 + }, + { + "type": "page_number", + "text": "3321", + "bbox": [ + 482, + 944, + 513, + 955 + ], + "page_idx": 7 + }, + { + "type": "text", + "text": "Acknowledgments", + "text_level": 1, + "bbox": [ + 91, + 90, + 250, + 107 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "This work is partially supported by NSFC (U23A20382, 62476123, 62376118, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund (RAGR20240101), Postdoctoral Fellowship Program of CPSF (GZB20250389), China Postdoctoral Science Foundation (2025M771535), and Collaborative Innovation Center of Novel Software Technology and Industrialization.", + "bbox": [ + 89, + 114, + 483, + 236 + ], + "page_idx": 8 + }, + { + "type": "text", + "text": "References", + "text_level": 1, + "bbox": [ + 91, + 250, + 187, + 266 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7", + "[2] Charu C Aggarwal. A survey of stream clustering algorithms. In Data Clustering, pages 231-258. Chapman and Hall/CRC, 2018. 1", + "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2", + "[4] Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars. Task-free continual learning. In CVPR, pages 11254-11263, 2019. 2", + "[5] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In NeurIPS, pages 11816-11825, 2019. 2", + "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, pages 9448-9458, 2019. 6", + "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461, 2014. 6", + "[8] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 2", + "[9] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2018. 2", + "[10] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, pages 16664-16678, 2022. 2", + "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In CVPR, pages 113-123, 2019. 7", + "[12] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern" + ], + "bbox": [ + 93, + 275, + 483, + 901 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "analysis and machine intelligence, 44(7):3366-3385, 2021. 1, 2", + "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1", + "[14] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class-incremental learning via relation knowledge distillation. In AAAI, pages 1255-1263, 2021. 2", + "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1", + "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020. 2", + "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285–9295, 2022. 2", + "[18] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 1", + "[19] Robert M French and André Ferrara. Modeling time perception in rats: Evidence for catastrophic interference in animal learning. In Proceedings of the 21st Annual Conference of the Cognitive Science Conference, pages 173-178. CiteSeer, 1999. 1", + "[20] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-DFCIL: relation-guided representation learning for data-free class incremental learning. In ECCV, pages 423–439, 2022. 2", + "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2015. 1", + "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6", + "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2", + "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2", + "[25] Linlan Huang, Xusheng Cao, Haori Lu, and Xialei Liu. Class-incremental learning with clip: Adaptive representation adjustment and parameter fusion. In ECCV, pages 214-231, 2024. 2, 3, 6, 7", + "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + ], + "bbox": [ + 516, + 92, + 906, + 901 + ], + "page_idx": 8 + }, + { + "type": "page_number", + "text": "3322", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 8 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1", + "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727, 2022. 2, 3", + "[29] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In ICCV, pages 11847-11857, 2023. 2", + "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2", + "[31] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, pages 554–561, 2013. 6", + "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6", + "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1", + "[34] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 1", + "[35] Lan Li, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Addressing imbalanced domain-incremental learning through dual-balance collaborative experts. In ICML, 2025. 2", + "[36] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, pages 614-629, 2016. 2", + "[37] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, pages 17612-17625, 2022. 8", + "[38] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, pages 12245-12254, 2020. 2", + "[39] Yichen Lu, Mei Wang, and Weihong Deng. Augmented geometric distillation for data-free incremental person reid. In CVPR, pages 7329-7338, 2022. 2", + "[40] Zilin Luo, Yaoyao Liu, Bernt Schiele, and Qianru Sun. Class-incremental exemplar compression for class-incremental learning. In CVPR, pages 11371-11380, 2023. 2", + "[41] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6", + "[42] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern" + ], + "bbox": [ + 91, + 92, + 480, + 898 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1, 2", + "[43] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2", + "[44] Mark D McDonnell, Dong Gong, Amin Parveneh, Ehsan Abbasnejad, and Anton van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. In NeurIPS, 2023. 2", + "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1", + "[46] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, pages 13698-13707, 2021. 2", + "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 7", + "[48] Zhi-Hong Qi, Da-Wei Zhou, Yiran Yao, Han-Jia Ye, and De-Chuan Zhan. Adaptive adapter routing for long-tailed class incremental learning. Machine Learning, 114(3):1-20, 2025. 2", + "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 3, 6, 7", + "[50] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In NIPS, pages 506-516, 2017. 2", + "[51] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2, 6, 7", + "[52] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, pages 16722-16731, 2022. 2", + "[53] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, pages 11909-11919, 2023. 2, 3, 6, 7", + "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NIPS, pages 4080-4090, 2017. 2", + "[55] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6", + "[56] Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, DeChuan Zhan, and Han-Jia Ye. Mos: Model surgery for pretrained model-based class-incremental learning. In AAAI, pages 20699-20707, 2025. 2" + ], + "bbox": [ + 516, + 92, + 903, + 898 + ], + "page_idx": 9 + }, + { + "type": "page_number", + "text": "3323", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 9 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[57] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, pages 254-270, 2020. 2", + "[58] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8", + "[59] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6", + "[60] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022. 2", + "[61] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In ICLR, 2023. 2", + "[62] Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lu, and Baochang Zhang. Attriclip: A non-incremental learner for incremental knowledge learning. In CVPR, pages 3654-3663, 2023. 2", + "[63] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. NeurIPS, pages 5682-5695, 2022. 2", + "[64] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, pages 631-648, 2022. 2, 3, 6, 7", + "[65] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. 1, 2, 3, 6, 7", + "[66] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free clip-based adaptation. In ICLR, 2024. 5", + "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019. 2", + "[68] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6", + "[69] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, pages 899–908, 2018. 2", + "[70] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021. 2", + "[71] Xu Yang, Zhangzikang Li, Haiyang Xu, Hanwang Zhang, Qinghao Ye, Chenliang Li, Ming Yan, Yu Zhang, Fei Huang, and Songfang Huang. Learning trajectory-word alignments for video-language tasks. In ICCV, pages 2504-2514, 2023. 1", + "[72] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. In NeurIPS, pages 40924-40943, 2023. 1" + ], + "bbox": [ + 91, + 90, + 480, + 898 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[73] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved clip classification. In CVPR, pages 27402-27411, 2024. 2", + "[74] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018. 2", + "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2024. 1", + "[76] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2, 3", + "[77] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, pages 6982-6991, 2020. 2", + "[78] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1", + "[79] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017. 2", + "[80] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13208-13217, 2020. 2", + "[81] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In ICML, pages 61649–61663, 2024. 2", + "[82] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Task-agnostic guided feature expansion for class incremental learning. In CVPR, pages 10099-10109, 2025. 2", + "[83] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In ICLR, 2023. 2, 6", + "[84] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision, 133:1012-1032, 2025. 2, 6, 7", + "[85] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, Lijun Zhang, and De-Chuan Zhan. Dual consolidation for pre-trained model-based domain-incremental learning. In CVPR, pages 20547–20557, 2025. 2", + "[86] Da-Wei Zhou, Yuanhan Zhang, Yan Wang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(6): 4489-4504, 2025. 2, 3, 5, 6, 7", + "[87] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2" + ], + "bbox": [ + 514, + 90, + 903, + 900 + ], + "page_idx": 10 + }, + { + "type": "page_number", + "text": "3324", + "bbox": [ + 482, + 944, + 514, + 955 + ], + "page_idx": 10 + }, + { + "type": "list", + "sub_type": "ref_text", + "list_items": [ + "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 6, 7", + "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, pages 5871-5880, 2021. 2" + ], + "bbox": [ + 91, + 90, + 480, + 176 + ], + "page_idx": 11 + }, + { + "type": "page_number", + "text": "3325", + "bbox": [ + 482, + 945, + 514, + 955 + ], + "page_idx": 11 + } +] \ No newline at end of file diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_model.json b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_model.json new file mode 100644 index 0000000000000000000000000000000000000000..106551f2cb7ec19b8699149e12e8f3c756e17b06 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_model.json @@ -0,0 +1,2743 @@ +[ + [ + { + "type": "header", + "bbox": [ + 0.107, + 0.003, + 0.182, + 0.044 + ], + "angle": 0, + "content": "CVF" + }, + { + "type": "header", + "bbox": [ + 0.239, + 0.001, + 0.808, + 0.047 + ], + "angle": 0, + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + }, + { + "type": "title", + "bbox": [ + 0.12, + 0.131, + 0.88, + 0.154 + ], + "angle": 0, + "content": "External Knowledge Injection for CLIP-Based Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.11, + 0.18, + 0.886, + 0.217 + ], + "angle": 0, + "content": "Da-Wei Zhou\\(^{1,2}\\), Kai-Wen Li\\(^{1,2}\\), Jingyi Ning\\(^{2}\\), Han-Jia Ye\\(^{1,2(\\boxtimes)}\\), Lijun Zhang\\(^{1,2}\\), De-Chuan Zhan\\(^{1,2}\\) \n\\(^{1}\\) School of Artificial Intelligence, Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.186, + 0.217, + 0.812, + 0.235 + ], + "angle": 0, + "content": "\\(^{2}\\) National Key Laboratory for Novel Software Technology, Nanjing University" + }, + { + "type": "text", + "bbox": [ + 0.191, + 0.237, + 0.802, + 0.253 + ], + "angle": 0, + "content": "{zhoudw,likw,yehj,zhanglj,zhandc}@lamda.nju.edu.cn，ningjy@nju.edu.cn" + }, + { + "type": "title", + "bbox": [ + 0.248, + 0.287, + 0.327, + 0.303 + ], + "angle": 0, + "content": "Abstract" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.32, + 0.486, + 0.713 + ], + "angle": 0, + "content": "Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of \"cat\" can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledgeGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features as data evolves. Extensive experiments demonstrate its state-of-the-art performance. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.743, + 0.224, + 0.759 + ], + "angle": 0, + "content": "1. Introduction" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.77, + 0.484, + 0.876 + ], + "angle": 0, + "content": "Recent advancements in deep learning have significantly impacted various aspects of life [13, 21, 72]. However, real-world data often presents substantial challenges to these models, particularly streaming data that requires continual learning [2, 34]. Class-Incremental Learning (CIL) [51] has been proposed to address this challenge by enabling models to absorb new knowledge incrementally." + }, + { + "type": "image", + "bbox": [ + 0.534, + 0.267, + 0.889, + 0.504 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.512, + 0.51, + 0.907, + 0.72 + ], + "angle": 0, + "content": "Figure 1. Last accuracy comparison on nine benchmark datasets. All methods are deployed with the same CLIP for fair comparison. ENGINE outperforms other competitors with a substantial margin. A major issue in CIL is catastrophic forgetting [18, 19], where learning new classes leads to the loss of previously acquired knowledge. As a result, much research has focused on mitigating forgetting in deep learning [12, 42]. In contrast to traditional approaches that train models from scratch, recent progress in pre-training has shifted attention towards using pre-trained models (PTMs) [65], such as Vision Transformers[15] and CLIP [49]. With extensive training data serving as pre-learned knowledge, PTMs offer a more generalizable starting point for CIL, showing promise for real-world applications." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.72, + 0.909, + 0.903 + ], + "angle": 0, + "content": "CLIP [49], as a pioneering pre-trained vision-language model [27, 33, 71, 75, 78], exhibits strong zero-shot performance by matching the embeddings of query images to class names. However, since visual features often contain fine-grained information [45], these detailed descriptors are neglected when using class names as matching targets. For example, the text concept of \"a photo of a cat\" can be further decomposed into features like \"whiskers,\" \"tail,\" and \"fur\" in order to match the visual embeddings, while these informative features are not well utilized. Furthermore, as the model is updated in CIL, the sub-features can be overwritten during the learning process, resulting in" + }, + { + "type": "page_footnote", + "bbox": [ + 0.109, + 0.887, + 0.42, + 0.902 + ], + "angle": 0, + "content": "†Correspondence to: Han-Jia Ye (yehj@lamda.nju.edu.cn)" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.517, + 0.957 + ], + "angle": 0, + "content": "3314" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.468, + 0.106 + ], + "angle": 0, + "content": "forgetting and a mismatch between cross-modal features." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.107, + 0.482, + 0.242 + ], + "angle": 0, + "content": "Recent studies [62, 65, 73, 86] have highlighted the limitations of using fixed template text in vision-language models, which often leads to a loss of rich contextual information. To address this, researchers have proposed learnable text inputs, such as task-specific prompts [88], which allow the model to encode additional, task-relevant information. However, these learnable inputs can become biased towards the visual features in the training set, ultimately compromising the model's generalizability [87]." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.243, + 0.482, + 0.363 + ], + "angle": 0, + "content": "In CLIP, a more detailed description in the textual space can provide a more informative matching target, enhancing prediction accuracy. For example, instead of using the generic template \"a photo of a cat,\" directly describing its features, such as \"whiskers\" and \"tail,\" can significantly improve recognition. In this way, these detailed features serve as anchors in the embedding space, and aligning to them helps reduce the risk of catastrophic forgetting." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.364, + 0.482, + 0.529 + ], + "angle": 0, + "content": "Learning with informative descriptors offers a promising approach to enhancing CLIP's continual learning capabilities. However, two major challenges remain: 1) acquiring informative features. Given the diversity of classes in downstream tasks, it is impractical for human experts to annotate all relevant features for recognition. 2) mitigating forgetting when utilizing external knowledge. As CIL involves continuous data streams, the model risks forgetting previously acquired knowledge after each update. Therefore, it is crucial to efficiently extract and retain external knowledge while preventing forgetting during sequential updates." + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.53, + 0.482, + 0.742 + ], + "angle": 0, + "content": "To address these challenges, we propose ExterNal knowledgeGe InjEction (ENGINE) for CLIP-based CIL. To acquire informative features, we enhance knowledge transfer by incorporating external knowledge from both visual and textual modalities through a dual-branch tuning framework. The visual branch is enhanced via data augmentation to enrich visual features, while the textual branch uses GPT-4 to rewrite discriminative prompts. In this way, we can encode diverse and informative class descriptions in the continual updating process. In addition to on-the-fly knowledge injection, we also apply post-tuning by re-ranking predictions during inference. Post-tuning is done by considering the local pair-wise features, which can further calibrate the wrong predictions to boost incremental learning." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.755, + 0.233, + 0.77 + ], + "angle": 0, + "content": "2. Related Work" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.78, + 0.482, + 0.901 + ], + "angle": 0, + "content": "Class-Incremental Learning (CIL): is a long-standing problem, aiming to absorb new knowledge without forgetting [12, 42]. Typical CIL algorithms train a model from scratch, which can be divided into several groups. Knowledge distillation-based methods aim to build the mapping target [23] between old and new models to resist forgetting [16, 36, 51], which involve logit-wise alignment [36, 51], feature-wise alignment [24, 39, 46], and" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.905, + 0.272 + ], + "angle": 0, + "content": "group-wise alignment [14, 20, 57]. Replay-based methods aim to recover previous knowledge by saving and replaying the subset of seen classes [5, 8, 9, 38, 40, 81]. Besides, parameter regularization-based methods estimate the importance of parameters and restrict important ones not to change [3, 4, 30, 79]. Model rectification-based methods observe and rectify the inductive bias in the model, e.g., the biased logits [52, 77, 80] and classifier weights [67, 80]. Model expansion-based methods [60, 70, 83] adjust the network structure to fit the characteristics of the evolving data, e.g., expanding neurons [69, 74], the whole backbone [60, 61, 70, 82, 83], and lightweight modules [17]." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.273, + 0.905, + 0.606 + ], + "angle": 0, + "content": "Pre-Trained Model-Based CIL: improves the capability of CIL by starting with generalizable pre-trained models [35, 48, 56, 85]. To maintain the pre-trained knowledge, most works seek to freeze the backbone and append lightweight modules, e.g., prompts [53, 63-65, 88] and adapters [10, 50, 76]. For example, L2P [65] and DualPrompt [64] design the visual prompt [28] pool to select the instance-specific prompts when learning with the pre-trained Vision Transformer. Several works build complex prompt combination/generation targets with attention mechanism [53] or generative network [29]. Other works directly utilize the pre-trained features to construct strong classifiers by matching class prototypes to the embeddings [44, 54, 84]. When the pre-trained CLIP is available as initialization, existing works seek to enhance cross-modal matching information via learning multi-modal prompts [62, 63]. MOE-Adapter [76] further extends the lightweight module selection process via mixture-of-experts [43], while PROOF [86] extends the representation ability of CLIP by appending new projection layers for new tasks. RAPF [25] learns the adapter modules for new tasks by decomposed parameter fusion to resist forgetting in the updating process." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.621, + 0.652, + 0.637 + ], + "angle": 0, + "content": "3. Preliminaries" + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.646, + 0.767, + 0.663 + ], + "angle": 0, + "content": "3.1. Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.668, + 0.905, + 0.865 + ], + "angle": 0, + "content": "CIL is designed to continually build a unified classifier for all seen classes of the data stream [51]. We denote training sets as \\(\\{\\mathcal{D}^1,\\mathcal{D}^2,\\dots ,\\mathcal{D}^B\\}\\), where each task \\(\\mathcal{D}^b = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{n_b}\\) contains \\(n_b\\) instances. Each training instance \\(\\mathbf{x}_i\\in \\mathbb{R}^D\\) belongs to class \\(y_{i}\\in Y_{b}\\), i.e., the label space of task \\(b\\) is \\(Y_{b}\\). We have \\(Y_{b}\\cap Y_{b^{\\prime}} = \\emptyset\\) for \\(b\\neq b^{\\prime}\\). This paper follows the exemplar-free CIL setting [64, 65, 89], where the model cannot hold any historical instances in the memory. In other words, we can only access data from \\(\\mathcal{D}^b\\) for model training when learning the \\(b\\)-th incremental task. In CIL, the target is to build a unified classifier for all seen classes \\(\\mathcal{V}_b = Y_1\\cup \\dots Y_b\\), i.e., find a model \\(f(\\mathbf{x}):X\\to \\mathcal{V}_b\\) that minimizes the expected risk:" + }, + { + "type": "equation", + "bbox": [ + 0.56, + 0.878, + 0.905, + 0.904 + ], + "angle": 0, + "content": "\\[\nf ^ {*} = \\underset {f \\in \\mathcal {H}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {t} ^ {1} \\cup \\dots \\mathcal {D} _ {t} ^ {b}} \\mathbb {I} (y \\neq f (\\mathbf {x})). \\tag {1}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3315" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.09, + 0.092, + 0.485, + 0.273 + ], + "angle": 0, + "content": "In Eq. 1, \\(\\mathcal{H}\\) represents the hypothesis space and \\(\\mathbb{I}(\\cdot)\\) is the indicator function. \\(\\mathcal{D}_t^b\\) denotes the \\(b\\)-th task's data distribution. In this paper, we follow [25, 76, 86] and assume that a pre-trained CLIP model (contrastive language-image pre-training) [49] is available as the initialization for \\(f(\\mathbf{x})\\). Specifically, CLIP contains the visual and textual encoder, i.e., \\(g_i(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^d\\), \\(g_t(\\cdot): \\mathbb{R}^{Dt} \\to \\mathbb{R}^d\\), where the \\(D / Dt\\)-dimensional images and texts are projected to the same \\(d\\)-dimensional embedding space. During inference, it constructs the template text \\(\\mathbf{t}_i\\) using the class name, i.e., \"a photo of a \\([\\mathrm{CLASS}]_i\\)\", and matches the query image to text embeddings of all classes:" + }, + { + "type": "equation", + "bbox": [ + 0.124, + 0.28, + 0.483, + 0.359 + ], + "angle": 0, + "content": "\\[\n\\begin{array}{l} f _ {y _ {i}} (\\mathbf {x}, \\mathbf {t} _ {i}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {j}\\right) / \\tau\\right)} \\tag {2} \\\\ = \\frac {\\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}, \\\\ \\end{array}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.366, + 0.484, + 0.427 + ], + "angle": 0, + "content": "where \\(\\cos (\\cdot ,\\cdot)\\) is cosine similarity and \\(\\tau\\) is temperature. \\(\\mathbf{z} = g_{i}(\\mathbf{x})\\) and \\(\\mathbf{w}_i = g_t(\\mathbf{t}_i)\\) are visual and textual features, respectively. In Eq. 2, the logit is assigned by the relative similarity of the visual embedding to textual embedding." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.434, + 0.439, + 0.451 + ], + "angle": 0, + "content": "3.2. Baselines in Class-Incremental Learning" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.457, + 0.483, + 0.487 + ], + "angle": 0, + "content": "To overcome forgetting, there are two typical solutions for CIL with pre-trained models by learning prompts [28, 88]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.487, + 0.484, + 0.577 + ], + "angle": 0, + "content": "Learning Visual Prompts: As fully-finetuning will harm the generalizability of the PTM, several works [28, 53, 64, 65] propose to freeze the pre-trained visual encoder and append visual prompts for subsequent tasks. They design a prompt pool to select instance-specific prompts, with which the visual embedding is represented as:" + }, + { + "type": "equation", + "bbox": [ + 0.235, + 0.587, + 0.483, + 0.602 + ], + "angle": 0, + "content": "\\[\n\\mathbf {z} = \\bar {g} _ {i} (\\mathbf {x}, \\mathcal {P}), \\tag {3}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.611, + 0.483, + 0.671 + ], + "angle": 0, + "content": "where \\(\\bar{g}_i\\) is the frozen image encoder, \\(\\mathcal{P}\\) is the prompt pool for prompt selection. By replacing \\(g_i(\\mathbf{x})\\) in Eq. 2 with Eq. 3, the model efficiently encodes downstream information into the visual prompts and tackles forgetting." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.672, + 0.484, + 0.777 + ], + "angle": 0, + "content": "Learning Textual Prompts: Since CLIP contains two branches for the visual and textual encoder, there are also works learning textual prompts instead of using the template text. Specifically, the textual input is formulated into a set of learnable prompts in CoOp [88], i.e., \\(\\tilde{\\mathbf{t}}_i = [\\mathrm{V}]_1[\\mathrm{V}]_2\\cdots[\\mathrm{V}]_M[\\mathrm{CLASS}]_i\\). In this way, the textual embedding is denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.237, + 0.785, + 0.483, + 0.803 + ], + "angle": 0, + "content": "\\[\n\\mathbf {w} _ {i} = \\bar {g} _ {t} (\\tilde {\\mathbf {t}} _ {i}), \\tag {4}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.811, + 0.483, + 0.856 + ], + "angle": 0, + "content": "where \\(\\bar{g}_t\\) is the frozen text encoder. Similarly, we can replace the textual embedding in Eq. 2 via Eq. 4 to encode task information into these learnable prompts." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.856, + 0.484, + 0.901 + ], + "angle": 0, + "content": "Discussions: Eq. 3 and Eq. 4 focus on different aspects to adjust the pre-trained CLIP into downstream tasks. However, when adapting the model with only visual prompts," + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.092, + 0.907, + 0.289 + ], + "angle": 0, + "content": "the problem occurs since the model has to decompose the class name into fine-grained concepts to match the visual features. The sequential updating process will cause the degradation of such decomposition ability, resulting in the mismatch between visual and textual features. By contrast, if we utilize textual prompts, the model is only focused on the specific training instance and lacks holistic information about the specific class. When the testing data contains the image of the same class with different distributions, the model shall be confused due to the prompt-level overfitting. Hence, the ideal text input should be general and holistic to match the visual features and convey all informative features that the target class possesses." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.301, + 0.874, + 0.318 + ], + "angle": 0, + "content": "4. ENGINE: External Knowledge Injection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.326, + 0.907, + 0.553 + ], + "angle": 0, + "content": "Noting that prompt-based methods are highly limited by the training instances, we need to extract comprehensive information for the model to adapt to downstream tasks while preventing forgetting. To encode holistic information into the model, we seek help from large language models like GPT-4 [1] to provide the general descriptions instead of using the template features. Apart from using detailed descriptions in the training process, we also make full use of them during inference by seeking pair-wise discriminative features. On the other hand, as external knowledge is continually injected into the model, we also design the injection unit expansion and prototype replay strategy. In the following sections, we first introduce how to inject external knowledge during model training and then introduce how to utilize it during inference." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.561, + 0.795, + 0.577 + ], + "angle": 0, + "content": "4.1. On-the-fly Knowledge Injection" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.583, + 0.907, + 0.749 + ], + "angle": 0, + "content": "Since the target is to encode general information into the CLIP during incremental learning, we first design the model training pipeline that can enhance recognition. For example, a \"cat\" contains visual features like \"soft, short fur\", \"long, thin tail\", and \"round face with large eyes\". It would be better to match the cat image to these visual features than \"a photo of a cat\", since those features contain more discriminative and fine-grained information. Correspondingly, to enhance CLIP's knowledge of general class description, we seek help from GPT-4 [1] to provide discriminative visual features:" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.757, + 0.871, + 0.787 + ], + "angle": 0, + "content": "Q: What are unique visual features of \\([\\mathrm{CLASS}]_i\\) in a photo? Focus on the key visual features." + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.787, + 0.851, + 0.802 + ], + "angle": 0, + "content": "A: 1. Long, thin tail that aids in balance. 2." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.811, + 0.907, + 0.901 + ], + "angle": 0, + "content": "In this way, we can get a set of visual descriptions for each class, denoted as \\(\\mathbf{d}_i\\). A naive solution is to replace the template text \\(\\mathbf{t}_i\\) into \\(\\mathbf{d}_i\\) during training, while we propose a more intuitive way to inject the knowledge into the model. Specifically, we design a knowledge injection unit \\(u_t\\) to dynamically encode these textual descriptors. We implement" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3316" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.114, + 0.069, + 0.614, + 0.253 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.254, + 0.255, + 0.468, + 0.269 + ], + "angle": 0, + "content": "(a) On-the-fly Knowledge Injection" + }, + { + "type": "image", + "bbox": [ + 0.624, + 0.074, + 0.887, + 0.254 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.648, + 0.255, + 0.869, + 0.269 + ], + "angle": 0, + "content": "(b) Post-tuning Knowledge Injection" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.276, + 0.907, + 0.345 + ], + "angle": 0, + "content": "Figure 2. Illustration of ENGINE. Left: On-the-fly knowledge injection. We utilize GPT-4 to provide visual features of each class, and append injection units to encode external knowledge into the model. We also adopt random augmentation to enhance image diversity using the same network structure. To overcome forgetting, we append the prototypes of previous classes in each learning stage. Right: Post-tuning knowledge injection. We extract top-\\(k\\) predictions and generate pair-wise discriminative features locally. The input image is forwarded with every textual feature to further refine the prediction." + }, + { + "type": "text", + "bbox": [ + 0.091, + 0.347, + 0.484, + 0.378 + ], + "angle": 0, + "content": "\\(u_{t}\\) with a linear layer, i.e., \\(u_{t}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}\\), and append it after the textual encoder. We have the similarity loss:" + }, + { + "type": "equation", + "bbox": [ + 0.169, + 0.387, + 0.483, + 0.406 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {t} = - \\operatorname {S i m} \\left(u _ {t} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right), \\bar {g} _ {t} \\left(\\mathbf {d} _ {i}\\right)\\right), \\tag {5}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.413, + 0.484, + 0.594 + ], + "angle": 0, + "content": "where we use cosine similarity to calculate \\(\\mathrm{Sim}(\\cdot, \\cdot)\\). In Eq. 5, we maximize the similarity between \\(u_{t}\\left(\\bar{g}_{t}\\left(\\mathbf{t}_{i}\\right)\\right)\\) and \\(\\bar{g}_{t}\\left(\\mathbf{d}_{i}\\right)\\), i.e., making the adapted textual feature similar to external descriptions. Since the textual encoder \\(\\bar{g}_{t}\\) is frozen, we can inject this external information into the injection unit \\(u_{t}\\). In this way, the extracted textual features will contain detailed descriptions even using the template text, e.g., highlighting the features of \"long thin tail\" and \"round face\" when the input is \"a photo of a cat\". In the implementation, since GPT-4 will output a set of descriptions, we randomly choose one of the descriptions \\(\\mathbf{d}_{i}\\) to calculate Eq. 5 in each iteration." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.595, + 0.484, + 0.716 + ], + "angle": 0, + "content": "Learning external visual knowledge: Similar to the textual branch, we can also inject visual information into the model during training. Given that we cannot fetch external data in the learning process, we utilize random data augmentation to adjust the input images, and denote the augmented input as \\(\\mathcal{A}(\\mathbf{x})\\). Correspondingly, we can learn the image knowledge injection unit \\(u_{i}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}\\), and maximize the similarity between augmented and vanilla input:" + }, + { + "type": "equation", + "bbox": [ + 0.162, + 0.725, + 0.483, + 0.742 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {i} = - \\operatorname {S i m} \\left(u _ {i} (\\bar {g} _ {i} (\\mathbf {x})), \\bar {g} _ {i} (\\mathcal {A} (\\mathbf {x}))\\right). \\tag {6}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.75, + 0.483, + 0.825 + ], + "angle": 0, + "content": "Similar to Eq. 5, Eq. 6 freezes the visual encoder \\(\\bar{g}_i\\), and only optimizes the image injection unit \\(u_i\\). In the matching process, the model maximizes the similarity of the adapted visual feature to the augmented view, which helps the model extract diverse visual information." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.826, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Learning task-specific injection unit: In CIL, new tasks will emerge as data evolves, and sequentially tuning a single injection unit will result in forgetting previous knowledge. Consequently, we conduct the injection process for each incremental task to avoid forgetting. Specifically, we" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.347, + 0.907, + 0.409 + ], + "angle": 0, + "content": "initialize a new injection unit \\( u_{t}^{b}, u_{i}^{b} \\) when learning the \\( b \\)-th task, and freeze all previous injection units when learning the current task. With the set of visual and textual units, we utilize the summation of them for the final representation:" + }, + { + "type": "equation", + "bbox": [ + 0.573, + 0.414, + 0.907, + 0.437 + ], + "angle": 0, + "content": "\\[\nG _ {i} (\\mathbf {x}) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {i} ^ {p} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) + u _ {i} ^ {b} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) \\qquad (7)\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.569, + 0.442, + 0.906, + 0.463 + ], + "angle": 0, + "content": "\\[\nG _ {t} \\left(\\mathbf {t} _ {i}\\right) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {t} ^ {p} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) + u _ {t} ^ {b} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) \\tag {8}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.465, + 0.906, + 0.525 + ], + "angle": 0, + "content": "In the original CLIP, we have the cross-modal contrastive loss to match the image to the corresponding text description. Correspondingly, we utilize the injected visual and textual features to calculate the contrastive loss:" + }, + { + "type": "equation", + "bbox": [ + 0.571, + 0.533, + 0.906, + 0.551 + ], + "angle": 0, + "content": "\\[\n\\mathcal {L} _ {c} = \\ell \\left(f _ {\\text {i n j}} (\\mathbf {x}), y\\right) \\quad \\text {w h e r e} \\tag {9}\n\\]" + }, + { + "type": "equation", + "bbox": [ + 0.53, + 0.553, + 0.905, + 0.591 + ], + "angle": 0, + "content": "\\[\nf _ {\\text {i n j}, y _ {i}} (\\mathbf {x}) = \\frac {\\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}) , G _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {V} _ {b} |} \\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}), G _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}. \\tag {10}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.597, + 0.906, + 0.642 + ], + "angle": 0, + "content": "In Eq. 9, we utilize the injected visual and textual features to calculate the contrastive loss, aiming to increase the similarity of pair-wise injected features." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.643, + 0.907, + 0.809 + ], + "angle": 0, + "content": "Preventing forgetting of injection unit In the updating process, we combine Eq. 5, Eq. 6, and Eq. 9 to update the model. Since we cannot hold historical instances for replay, the model may still forget previous concepts due to the sequential updating. To this end, utilizing the substitution of previous concepts can create a calibration across all seen classes. Observing that the visual backbone \\( g_{i} \\) is not updating throughout the learning process, we generate a subset of previous classes using visual prototypes. Specifically, during the learning process, we can calculate the average visual embedding of each class as:" + }, + { + "type": "equation", + "bbox": [ + 0.621, + 0.815, + 0.905, + 0.851 + ], + "angle": 0, + "content": "\\[\n\\mathbf {p} _ {k} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right) g _ {i} \\left(\\mathbf {x} _ {j}\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right)}, \\tag {11}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.856, + 0.906, + 0.902 + ], + "angle": 0, + "content": "which stands for the most representative features of the corresponding class. We can treat them as the visual embedding of previous classes when learning new" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3317" + } + ], + [ + { + "type": "text", + "bbox": [ + 0.091, + 0.092, + 0.483, + 0.123 + ], + "angle": 0, + "content": "ones, i.e., we construct the auxiliary training set \\(\\mathbf{P} = \\{(\\mathbf{p}_1, 1), (\\mathbf{p}_2, 2), \\dots, (\\mathbf{p}_{|\\mathcal{Y}_b|}, |\\mathcal{Y}_b|)\\}\\), and optimize:" + }, + { + "type": "equation", + "bbox": [ + 0.15, + 0.133, + 0.483, + 0.153 + ], + "angle": 0, + "content": "\\[\n\\min _ {\\left\\{u _ {t} ^ {b}, u _ {i} ^ {b} \\right\\}} \\sum_ {\\left(\\mathbf {x}, y\\right) \\in \\mathcal {D} ^ {b} \\cup \\mathbf {P}} \\mathcal {L} _ {t} + \\mathcal {L} _ {i} + \\mathcal {L} _ {c}. \\tag {12}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.16, + 0.484, + 0.281 + ], + "angle": 0, + "content": "Specifically, the prototype set \\(\\mathbf{P}\\) is only adopted in calculating \\(\\mathcal{L}_c\\). To enhance diversity across training stages, we also add Gaussian noise on the prototype embeddings, i.e., \\(\\mathbf{p}' = \\mathbf{p} + \\epsilon\\), \\(\\epsilon \\in \\mathcal{N}(0,\\alpha^2\\mathbf{I})\\), and utilize \\(\\mathbf{p}'\\) in Eq. 12. In this way, we can revisit previous knowledge to prevent forgetting when learning new injection units. Additionally, we also consider matching the visual features to these prototypes to further assist recognition [66, 86]." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.282, + 0.484, + 0.512 + ], + "angle": 0, + "content": "Summary of on-the-fly injection: We visualize the training stage knowledge injection in Figure 2 (left). ENGINE learns two sets of knowledge injection units, aiming to encode the extra knowledge into the unit module. We utilize GPT-4 to extract informative class descriptors, and utilize similarity loss to encode them into the injection units. Similarly, data augmentation is utilized to provide visual information. Hence, the model can extract more informative features as it sees more tasks. To prevent forgetting, we also design the learning objective considering class prototypes. Since injection units are implemented with a single layer, the appended parameter size \\((d\\times d)\\) per unit is negligible for CLIP. Besides, since injection units are linear layers, they can be further re-parameterized by aggregating weights into a single one, i.e., \\(\\sum_{p}u_{i}^{p}\\) and \\(\\sum_{p}u_{t}^{p}\\)." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.517, + 0.379, + 0.533 + ], + "angle": 0, + "content": "4.2. Post-tuning Knowledge Injection" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.539, + 0.483, + 0.735 + ], + "angle": 0, + "content": "We have introduced the way to inject external knowledge during training in Section 4.1. Apart from the model training stage, we can also utilize external knowledge to calibrate the prediction results during inference, and we call this step \"post-tuning knowledge injection\". For example, after the model makes inference by matching visual and textual features, we can treat the current prediction results as a preliminary result, which needs further rectification. For the top-\\(k\\) prediction results in this preliminary result, we treat them as competitors to the ground truth that could result in confusion. Hence, we need to construct a further prediction considering the image and all local competitors, which can focus more on the top-\\(k\\) classes to refine the prediction." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.736, + 0.484, + 0.904 + ], + "angle": 0, + "content": "Refining the results: After learning the injection units, we can get the prediction results using the injected features for more comprehensive features. However, the model may still make wrong predictions, e.g., classifying a \"cat\" into a similar class \"lion\" since they have similar visual features. To refine the prediction results, we design an extra consolidation process using external knowledge. Denoting the prediction results of injected feature matching in Eq. 10 as \\( f_{\\mathrm{inj}}(\\mathbf{x}) \\), we first extract the top-\\( k \\) predictions as the competitor candidates, i.e., \\( \\{f_{\\mathrm{inj},i_1}(\\mathbf{x}), f_{\\mathrm{inj},i_2}(\\mathbf{x}), \\dots, f_{\\mathrm{inj},i_k}(\\mathbf{x})\\} = \\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x})) \\), where \\( \\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x})) \\) indicates selecting the" + }, + { + "type": "text", + "bbox": [ + 0.513, + 0.092, + 0.893, + 0.108 + ], + "angle": 0, + "content": "top- \\(k\\) largest value in \\(f_{\\mathrm{inj}}(\\mathbf{x})\\). Their labels are denoted as:" + }, + { + "type": "equation", + "bbox": [ + 0.525, + 0.113, + 0.905, + 0.144 + ], + "angle": 0, + "content": "\\[\n\\left\\{y _ {i _ {1}}, y _ {i _ {2}}, \\dots , y _ {i _ {k}} \\right\\} = \\left\\{y _ {i} \\mid f _ {\\text {i n j}, i} (\\mathbf {x}) \\in \\operatorname {T o p} _ {k} \\left(f _ {\\text {i n j}} (\\mathbf {x})\\right) \\right\\}. \\tag {13}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.145, + 0.905, + 0.22 + ], + "angle": 0, + "content": "In Eq. 13, the label set \\(\\{y_{i_1}, y_{i_2}, \\ldots, y_{i_k}\\}\\) corresponds to the possible confusing labels for the input image. To further refine the predictions, we aim for pair-wise disambiguation among this label set. We again utilize GPT-4 to generate pair-wise discriminative features:" + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.226, + 0.874, + 0.271 + ], + "angle": 0, + "content": "Q: What are unique visual features of \\([\\mathrm{CLASS}]_i\\) compared to \\([\\mathrm{CLASS}]_j\\) in a photo? Focus on their key visual differences." + }, + { + "type": "text", + "bbox": [ + 0.545, + 0.272, + 0.873, + 0.316 + ], + "angle": 0, + "content": "A: [CLASS]i: 1. Long, thin tail that aids in balance... [CLASS]j: 1. Stockier, heavier tail with a tuft of hair at the end ..." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.323, + 0.907, + 0.446 + ], + "angle": 0, + "content": "With this process, GPT-4 provides the most discriminative features that can differentiate two classes, e.g., the differences in \"tail\" to differentiate \"cat\" from \"lion\" and vice versa. Given the top-\\(k\\) classes, we can generate their pairwise descriptors as \\(\\mathbf{D} = [\\mathbf{d}_{ij}]_{i,j=1}^{k}\\), where \\(\\mathbf{d}_{ij}\\) denotes the descriptions to differ class \\(i\\) from class \\(j\\). We feed these fine-grained descriptions as the textual input, and match the query embedding to them:" + }, + { + "type": "equation", + "bbox": [ + 0.593, + 0.452, + 0.905, + 0.495 + ], + "angle": 0, + "content": "\\[\nf _ {\\mathrm {p t}, y _ {i}} (\\mathbf {x}) = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {j \\neq i} f _ {y _ {i}} \\left(\\mathbf {x}, \\mathbf {d} _ {i j}\\right), \\tag {14}\n\\]" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.502, + 0.905, + 0.546 + ], + "angle": 0, + "content": "where we utilize zero-shot CLIP in Eq. 2 to match the query image to these pair-wise discriminative features, and average the similarities as the post-tuning logits." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.547, + 0.906, + 0.683 + ], + "angle": 0, + "content": "Effect of post-tuning knowledge injection: We visualize the post-tuning knowledge injection in Figure 2 (right). To discriminate similar concepts and refine the wrong predictions, we first extract the top-\\(k\\) predictions of the model, and utilize GPT-4 to generate pair-wise discriminative features locally. Eq. 14 provides the fine-grained local prediction among these top-\\(k\\) classes, which can further help refine the results as data evolves. We adopt the zero-shot CLIP for Eq. 14, which does not require incremental updates." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.69, + 0.706, + 0.705 + ], + "angle": 0, + "content": "4.3. Summary of ENGINE" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.712, + 0.907, + 0.879 + ], + "angle": 0, + "content": "In ENGINE, we utilize external knowledge to help continual learning in two aspects, i.e., on-the-fly injection and post-tuning injection. With the help of GPT-4, we can extract more informative descriptors suitable for the incremental learning classes, and inject the external knowledge into the knowledge units. Besides, we also extract local descriptors to discriminate pair-wise classes in the post-tuning stage to help refine the predictions. During training, we update the model via Eq. 12 to inject knowledge into the knowledge units. During inference, we aggregate the following outputs as the prediction:" + }, + { + "type": "equation", + "bbox": [ + 0.625, + 0.886, + 0.905, + 0.903 + ], + "angle": 0, + "content": "\\[\nf (\\mathbf {x}) = f _ {\\text {i n j}} (\\mathbf {x}) + f _ {\\mathrm {p t}} (\\mathbf {x}). \\tag {15}\n\\]" + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3318" + } + ], + [ + { + "type": "table_caption", + "bbox": [ + 0.09, + 0.064, + 0.907, + 0.093 + ], + "angle": 0, + "content": "Table 1. Average and last performance comparison of different methods. The best performance is shown in bold. All methods are initialized with the same pre-trained CLIP without exemplars for a fair comparison." + }, + { + "type": "table", + "bbox": [ + 0.129, + 0.094, + 0.862, + 0.601 + ], + "angle": 0, + "content": "
MethodAircraftCIFAR100Cars
B0 Inc10B50 Inc10B0 Inc10B50 Inc10B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)
Finetune3.160.961.721.057.844.445.302.463.141.101.541.13
CoOp [88]14.547.1413.057.7747.0024.2441.2324.1236.4621.6537.4020.87
SimpleCIL [84]59.2448.0953.0548.0984.1576.6380.2076.6392.0486.8588.9686.85
ZS-CLIP [49]26.6617.2221.7017.2281.8171.3876.4971.3882.6076.3778.3276.37
L2P [65]47.1928.2944.0732.1382.7473.0381.1473.6176.6361.8276.3765.64
DualPrompt [64]44.3025.8346.0733.5781.6372.4480.1272.5776.2662.9476.8867.55
CODA-Prompt [53]45.9827.6945.1432.2882.4373.4378.6971.5880.2166.4775.0664.19
RAPF [25]50.3823.6140.4725.4486.1478.0482.1777.9382.8962.8575.8763.19
ENGINE69.6958.6964.3859.0286.9279.2283.1579.4794.1490.0891.6190.03
ImageNet-RCUBUCF
MethodB0 Inc20B100 Inc20B0 Inc20B100 Inc20B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(A_B\\)
Finetune1.370.431.010.882.060.640.560.474.511.591.210.80
CoOp [88]60.7337.5254.2039.7727.618.5724.0310.1447.8533.4642.0224.74
SimpleCIL [84]81.0674.4876.8474.4883.8177.5279.7577.5290.4485.6888.1285.68
ZS-CLIP [49]83.3777.1779.5777.1774.3863.0667.9663.0675.5067.6471.4467.64
L2P [65]75.9766.5272.8266.7770.8757.9375.6466.1286.3476.4383.9576.62
DualPrompt [64]76.2166.6573.2267.5869.8957.4674.4064.8485.2175.8284.3176.35
CODA-Prompt [53]77.6968.9573.7168.0573.1262.9873.9562.2187.7680.1483.0475.03
RAPF [25]81.2670.4876.1070.2379.0962.7772.8262.9392.2880.3390.3181.55
ENGINE86.2280.3783.6380.9886.6580.2082.5979.3094.3590.0392.5189.58
SUNFoodObjectNet
MethodB0 Inc30B150 Inc30B0 Inc10B50 Inc10B0 Inc20B100 Inc20
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{\\mathcal{A}}\\)\\(\\mathcal{A}_B\\)
Finetune4.511.590.780.723.491.712.141.521.340.470.690.54
CoOp [88]45.9323.1139.3324.8936.0114.1833.1318.6721.246.2916.216.82
SimpleCIL [84]82.1375.5878.6275.5887.8981.6584.7381.6552.0640.1345.1140.13
ZS-CLIP [49]79.4272.1174.9572.1187.8681.9284.7581.9238.4326.4331.1226.43
L2P [65]82.8274.5479.5773.1085.6677.3380.4273.1351.4039.3948.9142.83
DualPrompt [64]82.4674.4079.3773.0284.9277.2980.0072.7552.6240.7249.0842.92
CODA-Prompt [53]83.3475.7180.3874.1786.1878.7880.9874.1346.4934.1340.5734.13
RAPF [25]82.1372.4778.0473.1088.5781.1585.5381.1748.6727.4339.2828.73
ENGINE85.0478.5481.5778.4589.8183.8986.8983.9459.1145.1951.3244.99
" + }, + { + "type": "table_caption", + "bbox": [ + 0.093, + 0.614, + 0.482, + 0.642 + ], + "angle": 0, + "content": "Table 2. Comparison to traditional exemplar-based CIL methods. ENGINE does not use any exemplars." + }, + { + "type": "table", + "bbox": [ + 0.093, + 0.643, + 0.474, + 0.736 + ], + "angle": 0, + "content": "
MethodExemplarsAircraft B0 Inc10SUN B0 Inc30
AABAAB
iCaRL [51]20 / class53.6043.9878.5667.30
MEMO [83]20 / class42.2425.4181.4873.45
PROOF [86]20 / class61.0053.5983.5777.28
ENGINE069.6958.6985.0478.54
" + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.753, + 0.224, + 0.77 + ], + "angle": 0, + "content": "5. Experiments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.781, + 0.483, + 0.903 + ], + "angle": 0, + "content": "In this section, we conduct extensive experiments on nine benchmark datasets and compare ENGINE to state-of-the-art methods. We visualize the incremental learning curve and provide ablation studies on prompt engineering and parameter sensitivity to test the model's robustness. Besides, we also visualize the learned embeddings and logit changes after injecting external knowledge to show its effectiveness. We provide more results in the supplementary." + }, + { + "type": "title", + "bbox": [ + 0.513, + 0.626, + 0.731, + 0.642 + ], + "angle": 0, + "content": "5.1. Implementation Details" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.806 + ], + "angle": 0, + "content": "Dataset: We follow [65, 86, 88] to evaluate the performance on nine benchmark datasets that have domain gap to CLIP's pre-training dataset, i.e., CIFAR100 [32], CUB200 [59], ObjectNet [6], ImageNet-R [22], FGVCAircraft [41], StanfordCars [31], Food101 [7], SUN397 [68] and UCF101 [55]. Following [86], we utilize the sampled 100 classes from CIFAR100, Aircraft, Cars, Food, UCF, 200 classes from CUB200, ObjectNet, ImageNet-R, and 300 classes from SUN to ease the data split. More details are reported in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.81, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Dataset split: Following [51, 65], we use 'B-\\(m\\) Inc-\\(n\\)' to split the classes in CIL. \\(m\\) indicates the number of classes in the first stage, and \\(n\\) represents that of every following stage. We follow [51] to randomly shuffle the class order with random seed 1993 for all compared methods, and keep this same for every method." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3319" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.096, + 0.066, + 0.36, + 0.224 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.163, + 0.229, + 0.294, + 0.241 + ], + "angle": 0, + "content": "(a) Aircraft Base0 Inc10" + }, + { + "type": "image", + "bbox": [ + 0.367, + 0.066, + 0.634, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.442, + 0.229, + 0.556, + 0.24 + ], + "angle": 0, + "content": "(b) Cars Base0 Inc10" + }, + { + "type": "image", + "bbox": [ + 0.639, + 0.067, + 0.903, + 0.223 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.713, + 0.229, + 0.827, + 0.24 + ], + "angle": 0, + "content": "(c) UCF Base0 Inc10" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.244, + 0.907, + 0.271 + ], + "angle": 0, + "content": "Figure 3. Incremental performance of different methods. We report the performance gap after the last incremental stage of ENGINE and the runner-up method at the end of the line. All methods utilize the same CLIP pre-trained weight. More figures are shown in supplementary." + }, + { + "type": "table_caption", + "bbox": [ + 0.091, + 0.275, + 0.483, + 0.303 + ], + "angle": 0, + "content": "Table 3. Results when all methods use the same textual description generated by GPT-4 and data augmentation." + }, + { + "type": "table", + "bbox": [ + 0.092, + 0.304, + 0.476, + 0.376 + ], + "angle": 0, + "content": "
MethodTemplateImageNet-R AB0 Inc20 A_BCIFAR AB0 Inc10 A_B
ZS-CLIPGPT Generated83.6877.5982.2671.70
RAPFGPT Generated81.8971.0486.4978.52
ENGINEGPT Generated86.2280.3786.9279.22
" + }, + { + "type": "text", + "bbox": [ + 0.089, + 0.384, + 0.483, + 0.503 + ], + "angle": 0, + "content": "Comparison methods: We first compare to SOTA pretrained model-based CIL algorithms, e.g., L2P [65], DualPrompt [64], CODA-Prompt [53], SimpleCIL [84]. Besides, we also compare to SOTA CLIP-based CIL algorithms, e.g., CoOp [88], PROOF [86], RAPF [25]. The baseline that finetunes CLIP for incremental tasks is denoted as Finetune. All methods are deployed with the same CLIP as initialization." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.505, + 0.484, + 0.839 + ], + "angle": 0, + "content": "Training details: The experiments are deployed with NVIDIA 4090 using PyTorch [47]. We follow [25, 86] to consider CLIP with ViT-B/16 for all compared methods for fair comparison. For vision-based methods that cannot utilize the textual prompt (e.g., L2P, DualPrompt, CODA-Prompt), we utilize CLIP's visual branch as their initialization. We report the results using LAION-400M pre-trained CLIP [26] in the main paper, and report results using OpenAI [49] in the supplementary. In ENGINE, we use SGD optimizer with a batch size of 64 to optimize the model for 10 epochs. The learning rate decays from 0.05 with cosine annealing. We set the prototype noise ratio \\(\\alpha\\) to 0.25 and \\(k\\) in post-tuning to 5. We use OpenAI GPT-4o mini [1] to generate the prompts for textual knowledge and AutoAugment [11] to augment the visual features. We explore other LLMs to extract external knowledge in the supplementary. The source code will be publicly available upon acceptance. Evaluation metric: Following [51, 86], we use \\(\\mathcal{A}_b\\) to represent the model's accuracy after the \\(b\\)-th stage. Specifically, we adopt \\(\\mathcal{A}_B\\) (the performance after the last stage) and \\(\\bar{\\mathcal{A}} = \\frac{1}{B}\\sum_{b=1}^{B}\\mathcal{A}_b\\) (average performance along incremental stages) as measurements." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.849, + 0.317, + 0.865 + ], + "angle": 0, + "content": "5.2. Benchmark Comparison" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.871, + 0.483, + 0.901 + ], + "angle": 0, + "content": "We first compare ENGINE to other state-of-the-art methods on the benchmark datasets and report the results in Table 1" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.276, + 0.707, + 0.391 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.563, + 0.395, + 0.661, + 0.408 + ], + "angle": 0, + "content": "(a) Ablation study" + }, + { + "type": "image", + "bbox": [ + 0.71, + 0.276, + 0.9, + 0.392 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.741, + 0.395, + 0.871, + 0.408 + ], + "angle": 0, + "content": "(b) Parameter sensitivity" + }, + { + "type": "image_caption", + "bbox": [ + 0.558, + 0.41, + 0.862, + 0.424 + ], + "angle": 0, + "content": "Figure 4. Ablation study and parameter sensitivity." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.449, + 0.907, + 0.648 + ], + "angle": 0, + "content": "and Figure 3. Specifically, we find ENGINE outperforms current SOTA by \\(3\\% \\sim 10\\%\\) on these benchmark datasets. We find the baseline method Finetune has the worst performance, indicating that the decomposition ability to tackle the class template into fine-grained features is totally forgotten. For visual prompt-based methods like L2P, DualPrompt, and CODA-Prompt, the performance is limited due to the inability to utilize textual information. For the textual prompt tuning method CoOp, we find that it achieves poor performance due to the forgetting of textual prompts. For other CLIP-based methods like RAPF, our method still shows substantial improvement, verifying its strong continual learning ability without forgetting." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.653, + 0.907, + 0.744 + ], + "angle": 0, + "content": "Apart from these non-exemplar-based methods, we also consider typical exemplar-based CIL methods for comparison. We report the comparison results in Table 2, where all methods are initialized with the same CLIP. As we can infer from the table, ENGINE still outperforms them even without using exemplars." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.75, + 0.909, + 0.902 + ], + "angle": 0, + "content": "Finally, what if all CLIP-based methods utilize the same external knowledge? Correspondingly, we also consider using the same generated prompt and data augmentation for ZS-CLIP and RAPF, and report the results in Table 3. In this table, we replace the template text of all methods using the same text in Section 4.1 and adopt the same data augmentation for image input. We find ENGINE still outperforms them even using the same prompt, indicating that ENGINE can better utilize the external information for incremental learning." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3320" + } + ], + [ + { + "type": "image", + "bbox": [ + 0.097, + 0.073, + 0.283, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.122, + 0.186, + 0.258, + 0.198 + ], + "angle": 0, + "content": "(a) ZS-CLIP, single stage" + }, + { + "type": "image", + "bbox": [ + 0.289, + 0.073, + 0.477, + 0.182 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.319, + 0.186, + 0.446, + 0.198 + ], + "angle": 0, + "content": "(b) ENGINE, two stages" + }, + { + "type": "image_caption", + "bbox": [ + 0.09, + 0.2, + 0.483, + 0.284 + ], + "angle": 0, + "content": "Figure 5. t-SNE [58] visualizations of zero-shot CLIP and ENGINE. ENGINE aligns the cross-modal features without forgetting. We represent the visual features of the first task with dots and classes of the second task with squares. The textual features are represented by triangles, and we utilize the shadow area to denote the decision boundary." + }, + { + "type": "title", + "bbox": [ + 0.091, + 0.299, + 0.258, + 0.315 + ], + "angle": 0, + "content": "5.3. Further Analysis" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.323, + 0.485, + 0.625 + ], + "angle": 0, + "content": "Ablation study: We first conduct ablations to investigate the effectiveness of each component in ENGINE. We report the incremental performance on CIFAR100 B0 Inc10 in Figure 4a. We can infer that 'ZS-CLIP' has the worst performance since the downstream distribution is different from the pre-training stage, and we view it as the baseline. Correspondingly, we first equip the model with visual knowledge injection (Eq. 6), and denote the model as 'w/ Visual Injection'. It shows that the performance drastically improves as the model has been adapted to extract more informative visual features. Similarly, we then append the textual information using Eq. 5, denoted as 'w/ Dual Injection.' As shown in the figure, using external knowledge for both modalities can further improve the performance than visual knowledge only. Finally, we equip the model with post-tuning knowledge injection, and utilize Eq. 15 for inference (i.e., 'w/ Post-tuning Injection'). This corresponds to the full version of ENGINE, and we find it achieves the best performance. Ablations verify that every component in ENGINE boosts the CIL performance." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.628, + 0.484, + 0.764 + ], + "angle": 0, + "content": "Parameter robustness: There are two hyperparameters in ENGINE, i.e., the prototype noise scale \\(\\alpha\\) in Section 4.1 and post-tuning candidate \\(k\\) in Eq. 14. We conduct experiments on CUB B0 Inc20 to investigate the robustness by changing these parameters. Specifically, we choose \\(\\alpha\\) among \\(\\{0.15, 0.2, 0.25, 0.3, 0.35\\}\\), and \\(k\\) among \\(\\{3, 4, 5, 6, 7\\}\\). We report the last performance in Figure 4b. As shown in the figure, the performance is robust with the change of these parameters, and we suggest \\(\\alpha = 0.25\\), \\(k = 5\\) as default." + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.765, + 0.484, + 0.902 + ], + "angle": 0, + "content": "Visualizations: In this section, we utilize t-SNE [58] to visualize the cross-modal features learned by ENGINE on CIFAR100 B0 Inc5. We represent the visual features of the first task with dots \\((\\bigcirc)\\) and classes of the second task with squares \\((\\square)\\). The textual features are represented by triangles \\((\\triangle)\\), and we utilize the shadow area to denote the decision boundary divided by the textual features. Specifically, we first visualize the instances of the first stage in zero-shot CLIP's embedding space in Figure 5a. The results indicate" + }, + { + "type": "image", + "bbox": [ + 0.518, + 0.072, + 0.905, + 0.275 + ], + "angle": 0, + "content": null + }, + { + "type": "image_caption", + "bbox": [ + 0.513, + 0.276, + 0.905, + 0.318 + ], + "angle": 0, + "content": "Figure 6. Left: Input. Middle: Top-5 predictions before post-tuning knowledge injection. Right: Top-5 predictions after post-tuning. More cases are shown in the supplementary." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.342, + 0.907, + 0.463 + ], + "angle": 0, + "content": "that pre-trained CLIP has difficulty in aligning the visual and textual features since they are situated in two distinct areas in the embedding space [37]. While for Figure 5b, we visualize the embedding of ENGINE across two incremental stages. As we can infer from the figure, ENGINE has two advantages: 1) The visual and textual features are aligned in the same cluster, and 2) The model has not forgotten previous classes when learning new ones." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.467, + 0.909, + 0.589 + ], + "angle": 0, + "content": "We also visualize the model output before and after post-tuning knowledge injection on ImageNet-R in Figure 6. We utilize bars with red edges to denote the ground-truth class. As we can infer from the figure, post-tuning knowledge injection can help the model refine the predictions by generating local comparisons. For the case on top of Figure 6, the model can focus on the differences between \"acorn\" and \"birdhouse\" to make the final prediction." + }, + { + "type": "title", + "bbox": [ + 0.514, + 0.615, + 0.634, + 0.63 + ], + "angle": 0, + "content": "6. Conclusion" + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.644, + 0.907, + 0.825 + ], + "angle": 0, + "content": "Class-incremental learning is essential to real-world applications. This paper aims to enable CLIP with CIL ability by injecting external knowledge. Specifically, we consider on-the-fly knowledge injection by learning dual-branch injection units. We augment images for the visual branch and generate detailed feature descriptions with GPT-4 for the textual branch. By maximizing the similarity between external information and the injection module, we encode informative features into the pre-trained CLIP. Besides, we also consider post-tuning knowledge injection with local class information to calibrate the predictions. Extensive experiments verify ENGINE's effectiveness." + }, + { + "type": "text", + "bbox": [ + 0.512, + 0.826, + 0.909, + 0.903 + ], + "angle": 0, + "content": "Limitations and future works: This paper relies on GPT-4 to provide expert knowledge for specific tasks, which may be impossible for the specific cases where GPT-4 cannot generalize. Future works include designing other formats of external knowledge, e.g., knowledge graphs." + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.514, + 0.957 + ], + "angle": 0, + "content": "3321" + } + ], + [ + { + "type": "title", + "bbox": [ + 0.092, + 0.091, + 0.251, + 0.108 + ], + "angle": 0, + "content": "Acknowledgments" + }, + { + "type": "text", + "bbox": [ + 0.09, + 0.115, + 0.484, + 0.237 + ], + "angle": 0, + "content": "This work is partially supported by NSFC (U23A20382, 62476123, 62376118, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund (RAGR20240101), Postdoctoral Fellowship Program of CPSF (GZB20250389), China Postdoctoral Science Foundation (2025M771535), and Collaborative Innovation Center of Novel Software Technology and Industrialization." + }, + { + "type": "title", + "bbox": [ + 0.093, + 0.251, + 0.188, + 0.267 + ], + "angle": 0, + "content": "References" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.276, + 0.484, + 0.345 + ], + "angle": 0, + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.1, + 0.347, + 0.484, + 0.387 + ], + "angle": 0, + "content": "[2] Charu C Aggarwal. A survey of stream clustering algorithms. In Data Clustering, pages 231-258. Chapman and Hall/CRC, 2018. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.101, + 0.39, + 0.482, + 0.445 + ], + "angle": 0, + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.447, + 0.483, + 0.488 + ], + "angle": 0, + "content": "[4] Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars. Task-free continual learning. In CVPR, pages 11254-11263, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.49, + 0.484, + 0.532 + ], + "angle": 0, + "content": "[5] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In NeurIPS, pages 11816-11825, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.533, + 0.484, + 0.602 + ], + "angle": 0, + "content": "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, pages 9448-9458, 2019. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.604, + 0.484, + 0.645 + ], + "angle": 0, + "content": "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461, 2014. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.647, + 0.484, + 0.702 + ], + "angle": 0, + "content": "[8] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.102, + 0.704, + 0.484, + 0.745 + ], + "angle": 0, + "content": "[9] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.747, + 0.483, + 0.801 + ], + "angle": 0, + "content": "[10] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, pages 16664-16678, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.803, + 0.483, + 0.845 + ], + "angle": 0, + "content": "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In CVPR, pages 113-123, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.094, + 0.847, + 0.484, + 0.902 + ], + "angle": 0, + "content": "[12] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern" + }, + { + "type": "list", + "bbox": [ + 0.094, + 0.276, + 0.484, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "analysis and machine intelligence, 44(7):3366-3385, 2021. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.123, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.165, + 0.905, + 0.22 + ], + "angle": 0, + "content": "[14] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class-incremental learning via relation knowledge distillation. In AAAI, pages 1255-1263, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.222, + 0.905, + 0.291 + ], + "angle": 0, + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.293, + 0.905, + 0.347 + ], + "angle": 0, + "content": "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.349, + 0.905, + 0.404 + ], + "angle": 0, + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285–9295, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.406, + 0.905, + 0.447 + ], + "angle": 0, + "content": "[18] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.449, + 0.908, + 0.516 + ], + "angle": 0, + "content": "[19] Robert M French and André Ferrara. Modeling time perception in rats: Evidence for catastrophic interference in animal learning. In Proceedings of the 21st Annual Conference of the Cognitive Science Conference, pages 173-178. CiteSeer, 1999. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.519, + 0.905, + 0.573 + ], + "angle": 0, + "content": "[20] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-DFCIL: relation-guided representation learning for data-free class incremental learning. In ECCV, pages 423–439, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.576, + 0.905, + 0.617 + ], + "angle": 0, + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2015. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.619, + 0.905, + 0.688 + ], + "angle": 0, + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.69, + 0.905, + 0.731 + ], + "angle": 0, + "content": "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.733, + 0.905, + 0.774 + ], + "angle": 0, + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.518, + 0.776, + 0.905, + 0.83 + ], + "angle": 0, + "content": "[25] Linlan Huang, Xusheng Cao, Haori Lu, and Xialei Liu. Class-incremental learning with clip: Adaptive representation adjustment and parameter fusion. In ECCV, pages 214-231, 2024. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.832, + 0.905, + 0.902 + ], + "angle": 0, + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.908, + 0.902 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3322" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.161 + ], + "angle": 0, + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.204 + ], + "angle": 0, + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727, 2022. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.205, + 0.482, + 0.247 + ], + "angle": 0, + "content": "[29] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In ICCV, pages 11847-11857, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.247, + 0.482, + 0.328 + ], + "angle": 0, + "content": "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.33, + 0.482, + 0.372 + ], + "angle": 0, + "content": "[31] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, pages 554–561, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.372, + 0.482, + 0.412 + ], + "angle": 0, + "content": "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.413, + 0.482, + 0.468 + ], + "angle": 0, + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.469, + 0.482, + 0.511 + ], + "angle": 0, + "content": "[34] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.552 + ], + "angle": 0, + "content": "[35] Lan Li, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Addressing imbalanced domain-incremental learning through dual-balance collaborative experts. In ICML, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.553, + 0.482, + 0.58 + ], + "angle": 0, + "content": "[36] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, pages 614-629, 2016. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.581, + 0.482, + 0.636 + ], + "angle": 0, + "content": "[37] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, pages 17612-17625, 2022. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.637, + 0.482, + 0.69 + ], + "angle": 0, + "content": "[38] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, pages 12245-12254, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.692, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[39] Yichen Lu, Mei Wang, and Weihong Deng. Augmented geometric distillation for data-free incremental person reid. In CVPR, pages 7329-7338, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.735, + 0.482, + 0.788 + ], + "angle": 0, + "content": "[40] Zilin Luo, Yaoyao Liu, Bernt Schiele, and Qianru Sun. Class-incremental exemplar compression for class-incremental learning. In CVPR, pages 11371-11380, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.79, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[41] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.9 + ], + "angle": 0, + "content": "[42] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.093, + 0.482, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.549, + 0.093, + 0.905, + 0.12 + ], + "angle": 0, + "content": "Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1, 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.122, + 0.905, + 0.164 + ], + "angle": 0, + "content": "[43] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.166, + 0.905, + 0.222 + ], + "angle": 0, + "content": "[44] Mark D McDonnell, Dong Gong, Amin Parveneh, Ehsan Abbasnejad, and Anton van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. In NeurIPS, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.224, + 0.905, + 0.252 + ], + "angle": 0, + "content": "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.254, + 0.905, + 0.296 + ], + "angle": 0, + "content": "[46] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, pages 13698-13707, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.298, + 0.905, + 0.368 + ], + "angle": 0, + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.37, + 0.905, + 0.424 + ], + "angle": 0, + "content": "[48] Zhi-Hong Qi, Da-Wei Zhou, Yiran Yao, Han-Jia Ye, and De-Chuan Zhan. Adaptive adapter routing for long-tailed class incremental learning. Machine Learning, 114(3):1-20, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.427, + 0.905, + 0.497 + ], + "angle": 0, + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.499, + 0.905, + 0.541 + ], + "angle": 0, + "content": "[50] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In NIPS, pages 506-516, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.543, + 0.905, + 0.598 + ], + "angle": 0, + "content": "[51] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.601, + 0.905, + 0.668 + ], + "angle": 0, + "content": "[52] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, pages 16722-16731, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.672, + 0.905, + 0.755 + ], + "angle": 0, + "content": "[53] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, pages 11909-11919, 2023. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.758, + 0.905, + 0.798 + ], + "angle": 0, + "content": "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NIPS, pages 4080-4090, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.801, + 0.905, + 0.843 + ], + "angle": 0, + "content": "[55] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.517, + 0.845, + 0.905, + 0.9 + ], + "angle": 0, + "content": "[56] Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, DeChuan Zhan, and Han-Jia Ye. Mos: Model surgery for pretrained model-based class-incremental learning. In AAAI, pages 20699-20707, 2025. 2" + }, + { + "type": "list", + "bbox": [ + 0.517, + 0.093, + 0.905, + 0.9 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.957 + ], + "angle": 0, + "content": "3323" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.134 + ], + "angle": 0, + "content": "[57] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, pages 254-270, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.134, + 0.482, + 0.162 + ], + "angle": 0, + "content": "[58] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.162, + 0.482, + 0.217 + ], + "angle": 0, + "content": "[59] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.218, + 0.482, + 0.26 + ], + "angle": 0, + "content": "[60] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.26, + 0.482, + 0.316 + ], + "angle": 0, + "content": "[61] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In ICLR, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.316, + 0.482, + 0.383 + ], + "angle": 0, + "content": "[62] Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lu, and Baochang Zhang. Attriclip: A non-incremental learner for incremental knowledge learning. In CVPR, pages 3654-3663, 2023. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.385, + 0.482, + 0.439 + ], + "angle": 0, + "content": "[63] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. NeurIPS, pages 5682-5695, 2022. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.44, + 0.482, + 0.51 + ], + "angle": 0, + "content": "[64] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, pages 631-648, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.511, + 0.482, + 0.566 + ], + "angle": 0, + "content": "[65] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. 1, 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.566, + 0.482, + 0.608 + ], + "angle": 0, + "content": "[66] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free clip-based adaptation. In ICLR, 2024. 5" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.608, + 0.482, + 0.65 + ], + "angle": 0, + "content": "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.65, + 0.482, + 0.704 + ], + "angle": 0, + "content": "[68] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.705, + 0.482, + 0.734 + ], + "angle": 0, + "content": "[69] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, pages 899–908, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.734, + 0.482, + 0.776 + ], + "angle": 0, + "content": "[70] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.776, + 0.482, + 0.843 + ], + "angle": 0, + "content": "[71] Xu Yang, Zhangzikang Li, Haiyang Xu, Hanwang Zhang, Qinghao Ye, Chenliang Li, Ming Yan, Yu Zhang, Fei Huang, and Songfang Huang. Learning trajectory-word alignments for video-language tasks. In ICCV, pages 2504-2514, 2023. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.845, + 0.482, + 0.899 + ], + "angle": 0, + "content": "[72] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. In NeurIPS, pages 40924-40943, 2023. 1" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.899 + ], + "angle": 0, + "content": null + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.134 + ], + "angle": 0, + "content": "[73] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved clip classification. In CVPR, pages 27402-27411, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.134, + 0.905, + 0.176 + ], + "angle": 0, + "content": "[74] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.176, + 0.905, + 0.232 + ], + "angle": 0, + "content": "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2024. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.232, + 0.905, + 0.288 + ], + "angle": 0, + "content": "[76] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2, 3" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.288, + 0.905, + 0.343 + ], + "angle": 0, + "content": "[77] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, pages 6982-6991, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.343, + 0.905, + 0.412 + ], + "angle": 0, + "content": "[78] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.412, + 0.905, + 0.453 + ], + "angle": 0, + "content": "[79] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.454, + 0.905, + 0.509 + ], + "angle": 0, + "content": "[80] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13208-13217, 2020. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.51, + 0.905, + 0.565 + ], + "angle": 0, + "content": "[81] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In ICML, pages 61649–61663, 2024. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.566, + 0.905, + 0.62 + ], + "angle": 0, + "content": "[82] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Task-agnostic guided feature expansion for class incremental learning. In CVPR, pages 10099-10109, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.621, + 0.905, + 0.664 + ], + "angle": 0, + "content": "[83] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In ICLR, 2023. 2, 6" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.664, + 0.905, + 0.733 + ], + "angle": 0, + "content": "[84] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision, 133:1012-1032, 2025. 2, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.733, + 0.905, + 0.788 + ], + "angle": 0, + "content": "[85] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, Lijun Zhang, and De-Chuan Zhan. Dual consolidation for pre-trained model-based domain-incremental learning. In CVPR, pages 20547–20557, 2025. 2" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.788, + 0.905, + 0.858 + ], + "angle": 0, + "content": "[86] Da-Wei Zhou, Yuanhan Zhang, Yan Wang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(6): 4489-4504, 2025. 2, 3, 5, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.516, + 0.859, + 0.905, + 0.901 + ], + "angle": 0, + "content": "[87] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2" + }, + { + "type": "list", + "bbox": [ + 0.516, + 0.092, + 0.905, + 0.901 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.945, + 0.516, + 0.956 + ], + "angle": 0, + "content": "3324" + } + ], + [ + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.133 + ], + "angle": 0, + "content": "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 6, 7" + }, + { + "type": "ref_text", + "bbox": [ + 0.093, + 0.136, + 0.482, + 0.177 + ], + "angle": 0, + "content": "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, pages 5871-5880, 2021. 2" + }, + { + "type": "list", + "bbox": [ + 0.093, + 0.092, + 0.482, + 0.177 + ], + "angle": 0, + "content": null + }, + { + "type": "page_number", + "bbox": [ + 0.483, + 0.946, + 0.515, + 0.956 + ], + "angle": 0, + "content": "3325" + } + ] +] \ No newline at end of file diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_origin.pdf b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_origin.pdf new file mode 100644 index 0000000000000000000000000000000000000000..856b3ef0a6ac947ab77c6022b63ac4b4c3a61124 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/59db92d7-ff41-4079-a08f-c630a2445462_origin.pdf @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:b3b8efe01b5335cba383e636a06c376c66a95d43f75c4f79ad1f497277dda164 +size 1306898 diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/full.md b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/full.md new file mode 100644 index 0000000000000000000000000000000000000000..ad930bbfcd086a5db2c6de0c6d24b5e8ee2439b7 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/full.md @@ -0,0 +1,380 @@ +# External Knowledge Injection for CLIP-Based Class-Incremental Learning + +Da-Wei Zhou $^{1,2}$ , Kai-Wen Li $^{1,2}$ , Jingyi Ning $^{2}$ , Han-Jia Ye $^{1,2(\boxtimes)}$ , Lijun Zhang $^{1,2}$ , De-Chuan Zhan $^{1,2}$ $^{1}$ School of Artificial Intelligence, Nanjing University + +$^{2}$ National Key Laboratory for Novel Software Technology, Nanjing University + +{zhoudw,likw,yehj,zhanglj,zhandc}@lamda.nju.edu.cn，ningjy@nju.edu.cn + +# Abstract + +Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of "cat" can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledgeGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features as data evolves. Extensive experiments demonstrate its state-of-the-art performance. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE. + +# 1. Introduction + +Recent advancements in deep learning have significantly impacted various aspects of life [13, 21, 72]. However, real-world data often presents substantial challenges to these models, particularly streaming data that requires continual learning [2, 34]. Class-Incremental Learning (CIL) [51] has been proposed to address this challenge by enabling models to absorb new knowledge incrementally. + +![](images/5b4db64d4b85d37e186f514ef85db0823d9b7c8d92f9561e3fec5787c5d3fe69.jpg) +Figure 1. Last accuracy comparison on nine benchmark datasets. All methods are deployed with the same CLIP for fair comparison. ENGINE outperforms other competitors with a substantial margin. A major issue in CIL is catastrophic forgetting [18, 19], where learning new classes leads to the loss of previously acquired knowledge. As a result, much research has focused on mitigating forgetting in deep learning [12, 42]. In contrast to traditional approaches that train models from scratch, recent progress in pre-training has shifted attention towards using pre-trained models (PTMs) [65], such as Vision Transformers[15] and CLIP [49]. With extensive training data serving as pre-learned knowledge, PTMs offer a more generalizable starting point for CIL, showing promise for real-world applications. + +CLIP [49], as a pioneering pre-trained vision-language model [27, 33, 71, 75, 78], exhibits strong zero-shot performance by matching the embeddings of query images to class names. However, since visual features often contain fine-grained information [45], these detailed descriptors are neglected when using class names as matching targets. For example, the text concept of "a photo of a cat" can be further decomposed into features like "whiskers," "tail," and "fur" in order to match the visual embeddings, while these informative features are not well utilized. Furthermore, as the model is updated in CIL, the sub-features can be overwritten during the learning process, resulting in + +forgetting and a mismatch between cross-modal features. + +Recent studies [62, 65, 73, 86] have highlighted the limitations of using fixed template text in vision-language models, which often leads to a loss of rich contextual information. To address this, researchers have proposed learnable text inputs, such as task-specific prompts [88], which allow the model to encode additional, task-relevant information. However, these learnable inputs can become biased towards the visual features in the training set, ultimately compromising the model's generalizability [87]. + +In CLIP, a more detailed description in the textual space can provide a more informative matching target, enhancing prediction accuracy. For example, instead of using the generic template "a photo of a cat," directly describing its features, such as "whiskers" and "tail," can significantly improve recognition. In this way, these detailed features serve as anchors in the embedding space, and aligning to them helps reduce the risk of catastrophic forgetting. + +Learning with informative descriptors offers a promising approach to enhancing CLIP's continual learning capabilities. However, two major challenges remain: 1) acquiring informative features. Given the diversity of classes in downstream tasks, it is impractical for human experts to annotate all relevant features for recognition. 2) mitigating forgetting when utilizing external knowledge. As CIL involves continuous data streams, the model risks forgetting previously acquired knowledge after each update. Therefore, it is crucial to efficiently extract and retain external knowledge while preventing forgetting during sequential updates. + +To address these challenges, we propose ExterNal knowledgeGe InjEction (ENGINE) for CLIP-based CIL. To acquire informative features, we enhance knowledge transfer by incorporating external knowledge from both visual and textual modalities through a dual-branch tuning framework. The visual branch is enhanced via data augmentation to enrich visual features, while the textual branch uses GPT-4 to rewrite discriminative prompts. In this way, we can encode diverse and informative class descriptions in the continual updating process. In addition to on-the-fly knowledge injection, we also apply post-tuning by re-ranking predictions during inference. Post-tuning is done by considering the local pair-wise features, which can further calibrate the wrong predictions to boost incremental learning. + +# 2. Related Work + +Class-Incremental Learning (CIL): is a long-standing problem, aiming to absorb new knowledge without forgetting [12, 42]. Typical CIL algorithms train a model from scratch, which can be divided into several groups. Knowledge distillation-based methods aim to build the mapping target [23] between old and new models to resist forgetting [16, 36, 51], which involve logit-wise alignment [36, 51], feature-wise alignment [24, 39, 46], and + +group-wise alignment [14, 20, 57]. Replay-based methods aim to recover previous knowledge by saving and replaying the subset of seen classes [5, 8, 9, 38, 40, 81]. Besides, parameter regularization-based methods estimate the importance of parameters and restrict important ones not to change [3, 4, 30, 79]. Model rectification-based methods observe and rectify the inductive bias in the model, e.g., the biased logits [52, 77, 80] and classifier weights [67, 80]. Model expansion-based methods [60, 70, 83] adjust the network structure to fit the characteristics of the evolving data, e.g., expanding neurons [69, 74], the whole backbone [60, 61, 70, 82, 83], and lightweight modules [17]. + +Pre-Trained Model-Based CIL: improves the capability of CIL by starting with generalizable pre-trained models [35, 48, 56, 85]. To maintain the pre-trained knowledge, most works seek to freeze the backbone and append lightweight modules, e.g., prompts [53, 63-65, 88] and adapters [10, 50, 76]. For example, L2P [65] and DualPrompt [64] design the visual prompt [28] pool to select the instance-specific prompts when learning with the pre-trained Vision Transformer. Several works build complex prompt combination/generation targets with attention mechanism [53] or generative network [29]. Other works directly utilize the pre-trained features to construct strong classifiers by matching class prototypes to the embeddings [44, 54, 84]. When the pre-trained CLIP is available as initialization, existing works seek to enhance cross-modal matching information via learning multi-modal prompts [62, 63]. MOE-Adapter [76] further extends the lightweight module selection process via mixture-of-experts [43], while PROOF [86] extends the representation ability of CLIP by appending new projection layers for new tasks. RAPF [25] learns the adapter modules for new tasks by decomposed parameter fusion to resist forgetting in the updating process. + +# 3. Preliminaries + +# 3.1. Class-Incremental Learning + +CIL is designed to continually build a unified classifier for all seen classes of the data stream [51]. We denote training sets as $\{\mathcal{D}^1,\mathcal{D}^2,\dots ,\mathcal{D}^B\}$ , where each task $\mathcal{D}^b = \{(\mathbf{x}_i,y_i)\}_{i = 1}^{n_b}$ contains $n_b$ instances. Each training instance $\mathbf{x}_i\in \mathbb{R}^D$ belongs to class $y_{i}\in Y_{b}$ , i.e., the label space of task $b$ is $Y_{b}$ . We have $Y_{b}\cap Y_{b^{\prime}} = \emptyset$ for $b\neq b^{\prime}$ . This paper follows the exemplar-free CIL setting [64, 65, 89], where the model cannot hold any historical instances in the memory. In other words, we can only access data from $\mathcal{D}^b$ for model training when learning the $b$ -th incremental task. In CIL, the target is to build a unified classifier for all seen classes $\mathcal{V}_b = Y_1\cup \dots Y_b$ , i.e., find a model $f(\mathbf{x}):X\to \mathcal{V}_b$ that minimizes the expected risk: + +$$ +f ^ {*} = \underset {f \in \mathcal {H}} {\operatorname {a r g m i n}} \mathbb {E} _ {(\mathbf {x}, y) \sim \mathcal {D} _ {t} ^ {1} \cup \dots \mathcal {D} _ {t} ^ {b}} \mathbb {I} (y \neq f (\mathbf {x})). \tag {1} +$$ + +In Eq. 1, $\mathcal{H}$ represents the hypothesis space and $\mathbb{I}(\cdot)$ is the indicator function. $\mathcal{D}_t^b$ denotes the $b$ -th task's data distribution. In this paper, we follow [25, 76, 86] and assume that a pre-trained CLIP model (contrastive language-image pre-training) [49] is available as the initialization for $f(\mathbf{x})$ . Specifically, CLIP contains the visual and textual encoder, i.e., $g_i(\cdot): \mathbb{R}^D \to \mathbb{R}^d$ , $g_t(\cdot): \mathbb{R}^{Dt} \to \mathbb{R}^d$ , where the $D / Dt$ -dimensional images and texts are projected to the same $d$ -dimensional embedding space. During inference, it constructs the template text $\mathbf{t}_i$ using the class name, i.e., "a photo of a $[\mathrm{CLASS}]_i$ ", and matches the query image to text embeddings of all classes: + +$$ +\begin{array}{l} f _ {y _ {i}} (\mathbf {x}, \mathbf {t} _ {i}) = \frac {\exp \left(\cos \left(\mathbf {z} , \mathbf {w} _ {i}\right) / \tau\right)}{\sum_ {j = 1} ^ {| \mathcal {Y} _ {b} |} \exp \left(\cos \left(\mathbf {z} , \mathbf {w} _ {j}\right) / \tau\right)} \tag {2} \\ = \frac {\exp \left(\cos \left(g _ {i} (\mathbf {x}) , g _ {t} (\mathbf {t} _ {i})\right) / \tau\right)}{\sum_ {j = 1} ^ {| \mathcal {Y} _ {b} |} \exp \left(\cos \left(g _ {i} (\mathbf {x}) , g _ {t} (\mathbf {t} _ {j})\right) / \tau\right)}, \\ \end{array} +$$ + +where $\cos (\cdot ,\cdot)$ is cosine similarity and $\tau$ is temperature. $\mathbf{z} = g_{i}(\mathbf{x})$ and $\mathbf{w}_i = g_t(\mathbf{t}_i)$ are visual and textual features, respectively. In Eq. 2, the logit is assigned by the relative similarity of the visual embedding to textual embedding. + +# 3.2. Baselines in Class-Incremental Learning + +To overcome forgetting, there are two typical solutions for CIL with pre-trained models by learning prompts [28, 88]. + +Learning Visual Prompts: As fully-finetuning will harm the generalizability of the PTM, several works [28, 53, 64, 65] propose to freeze the pre-trained visual encoder and append visual prompts for subsequent tasks. They design a prompt pool to select instance-specific prompts, with which the visual embedding is represented as: + +$$ +\mathbf {z} = \bar {g} _ {i} (\mathbf {x}, \mathcal {P}), \tag {3} +$$ + +where $\bar{g}_i$ is the frozen image encoder, $\mathcal{P}$ is the prompt pool for prompt selection. By replacing $g_i(\mathbf{x})$ in Eq. 2 with Eq. 3, the model efficiently encodes downstream information into the visual prompts and tackles forgetting. + +Learning Textual Prompts: Since CLIP contains two branches for the visual and textual encoder, there are also works learning textual prompts instead of using the template text. Specifically, the textual input is formulated into a set of learnable prompts in CoOp [88], i.e., $\tilde{\mathbf{t}}_i = [\mathrm{V}]_1[\mathrm{V}]_2\cdots[\mathrm{V}]_M[\mathrm{CLASS}]_i$ . In this way, the textual embedding is denoted as: + +$$ +\mathbf {w} _ {i} = \bar {g} _ {t} (\tilde {\mathbf {t}} _ {i}), \tag {4} +$$ + +where $\bar{g}_t$ is the frozen text encoder. Similarly, we can replace the textual embedding in Eq. 2 via Eq. 4 to encode task information into these learnable prompts. + +Discussions: Eq. 3 and Eq. 4 focus on different aspects to adjust the pre-trained CLIP into downstream tasks. However, when adapting the model with only visual prompts, + +the problem occurs since the model has to decompose the class name into fine-grained concepts to match the visual features. The sequential updating process will cause the degradation of such decomposition ability, resulting in the mismatch between visual and textual features. By contrast, if we utilize textual prompts, the model is only focused on the specific training instance and lacks holistic information about the specific class. When the testing data contains the image of the same class with different distributions, the model shall be confused due to the prompt-level overfitting. Hence, the ideal text input should be general and holistic to match the visual features and convey all informative features that the target class possesses. + +# 4. ENGINE: External Knowledge Injection + +Noting that prompt-based methods are highly limited by the training instances, we need to extract comprehensive information for the model to adapt to downstream tasks while preventing forgetting. To encode holistic information into the model, we seek help from large language models like GPT-4 [1] to provide the general descriptions instead of using the template features. Apart from using detailed descriptions in the training process, we also make full use of them during inference by seeking pair-wise discriminative features. On the other hand, as external knowledge is continually injected into the model, we also design the injection unit expansion and prototype replay strategy. In the following sections, we first introduce how to inject external knowledge during model training and then introduce how to utilize it during inference. + +# 4.1. On-the-fly Knowledge Injection + +Since the target is to encode general information into the CLIP during incremental learning, we first design the model training pipeline that can enhance recognition. For example, a "cat" contains visual features like "soft, short fur", "long, thin tail", and "round face with large eyes". It would be better to match the cat image to these visual features than "a photo of a cat", since those features contain more discriminative and fine-grained information. Correspondingly, to enhance CLIP's knowledge of general class description, we seek help from GPT-4 [1] to provide discriminative visual features: + +Q: What are unique visual features of $[\mathrm{CLASS}]_i$ in a photo? Focus on the key visual features. + +A: 1. Long, thin tail that aids in balance. 2. + +In this way, we can get a set of visual descriptions for each class, denoted as $\mathbf{d}_i$ . A naive solution is to replace the template text $\mathbf{t}_i$ into $\mathbf{d}_i$ during training, while we propose a more intuitive way to inject the knowledge into the model. Specifically, we design a knowledge injection unit $u_t$ to dynamically encode these textual descriptors. We implement + +![](images/6908e6de466c635a8ac1e49078c1e9325cb9036237e3cecd611ac12ae3ebb1a5.jpg) +(a) On-the-fly Knowledge Injection + +![](images/a28fd7ad1910454bf814df56f2412adf9cbad31e548508ae144600f62b9d0d14.jpg) +(b) Post-tuning Knowledge Injection +Figure 2. Illustration of ENGINE. Left: On-the-fly knowledge injection. We utilize GPT-4 to provide visual features of each class, and append injection units to encode external knowledge into the model. We also adopt random augmentation to enhance image diversity using the same network structure. To overcome forgetting, we append the prototypes of previous classes in each learning stage. Right: Post-tuning knowledge injection. We extract top- $k$ predictions and generate pair-wise discriminative features locally. The input image is forwarded with every textual feature to further refine the prediction. + +$u_{t}$ with a linear layer, i.e., $u_{t}(\cdot):\mathbb{R}^{d}\to \mathbb{R}^{d}$ , and append it after the textual encoder. We have the similarity loss: + +$$ +\mathcal {L} _ {t} = - \operatorname {S i m} \left(u _ {t} \left(\bar {g} _ {t} \left(\mathbf {t} _ {i}\right)\right), \bar {g} _ {t} \left(\mathbf {d} _ {i}\right)\right), \tag {5} +$$ + +where we use cosine similarity to calculate $\mathrm{Sim}(\cdot, \cdot)$ . In Eq. 5, we maximize the similarity between $u_{t}\left(\bar{g}_{t}\left(\mathbf{t}_{i}\right)\right)$ and $\bar{g}_{t}\left(\mathbf{d}_{i}\right)$ , i.e., making the adapted textual feature similar to external descriptions. Since the textual encoder $\bar{g}_{t}$ is frozen, we can inject this external information into the injection unit $u_{t}$ . In this way, the extracted textual features will contain detailed descriptions even using the template text, e.g., highlighting the features of "long thin tail" and "round face" when the input is "a photo of a cat". In the implementation, since GPT-4 will output a set of descriptions, we randomly choose one of the descriptions $\mathbf{d}_{i}$ to calculate Eq. 5 in each iteration. + +Learning external visual knowledge: Similar to the textual branch, we can also inject visual information into the model during training. Given that we cannot fetch external data in the learning process, we utilize random data augmentation to adjust the input images, and denote the augmented input as $\mathcal{A}(\mathbf{x})$ . Correspondingly, we can learn the image knowledge injection unit $u_{i}(\cdot):\mathbb{R}^{d}\to \mathbb{R}^{d}$ , and maximize the similarity between augmented and vanilla input: + +$$ +\mathcal {L} _ {i} = - \operatorname {S i m} \left(u _ {i} (\bar {g} _ {i} (\mathbf {x})), \bar {g} _ {i} (\mathcal {A} (\mathbf {x}))\right). \tag {6} +$$ + +Similar to Eq. 5, Eq. 6 freezes the visual encoder $\bar{g}_i$ , and only optimizes the image injection unit $u_i$ . In the matching process, the model maximizes the similarity of the adapted visual feature to the augmented view, which helps the model extract diverse visual information. + +Learning task-specific injection unit: In CIL, new tasks will emerge as data evolves, and sequentially tuning a single injection unit will result in forgetting previous knowledge. Consequently, we conduct the injection process for each incremental task to avoid forgetting. Specifically, we + +initialize a new injection unit $u_{t}^{b}, u_{i}^{b}$ when learning the $b$ -th task, and freeze all previous injection units when learning the current task. With the set of visual and textual units, we utilize the summation of them for the final representation: + +$$ +G _ {i} (\mathbf {x}) = \sum_ {p = 1} ^ {b - 1} \bar {u} _ {i} ^ {p} \left(\bar {g} _ {i} (\mathbf {x})\right) + u _ {i} ^ {b} \left(\bar {g} _ {i} (\mathbf {x})\right) \qquad (7) +$$ + +$$ +G _ {t} \left(\mathbf {t} _ {i}\right) = \sum_ {p = 1} ^ {b - 1} \bar {u} _ {t} ^ {p} \left(\bar {g} _ {t} \left(\mathbf {t} _ {i}\right)\right) + u _ {t} ^ {b} \left(\bar {g} _ {t} \left(\mathbf {t} _ {i}\right)\right) \tag {8} +$$ + +In the original CLIP, we have the cross-modal contrastive loss to match the image to the corresponding text description. Correspondingly, we utilize the injected visual and textual features to calculate the contrastive loss: + +$$ +\mathcal {L} _ {c} = \ell \left(f _ {\text {i n j}} (\mathbf {x}), y\right) \quad \text {w h e r e} \tag {9} +$$ + +$$ +f _ {\text {i n j}, y _ {i}} (\mathbf {x}) = \frac {\exp \left(\cos \left(G _ {i} (\mathbf {x}) , G _ {t} (\mathbf {t} _ {i})\right) / \tau\right)}{\sum_ {j = 1} ^ {| \mathcal {V} _ {b} |} \exp \left(\cos \left(G _ {i} (\mathbf {x}), G _ {t} (\mathbf {t} _ {j})\right) / \tau\right)}. \tag {10} +$$ + +In Eq. 9, we utilize the injected visual and textual features to calculate the contrastive loss, aiming to increase the similarity of pair-wise injected features. + +Preventing forgetting of injection unit In the updating process, we combine Eq. 5, Eq. 6, and Eq. 9 to update the model. Since we cannot hold historical instances for replay, the model may still forget previous concepts due to the sequential updating. To this end, utilizing the substitution of previous concepts can create a calibration across all seen classes. Observing that the visual backbone $g_{i}$ is not updating throughout the learning process, we generate a subset of previous classes using visual prototypes. Specifically, during the learning process, we can calculate the average visual embedding of each class as: + +$$ +\mathbf {p} _ {k} = \frac {\sum_ {j = 1} ^ {| \mathcal {D} ^ {b} |} \mathbb {I} \left(y _ {j} = k\right) g _ {i} \left(\mathbf {x} _ {j}\right)}{\sum_ {j = 1} ^ {| \mathcal {D} ^ {b} |} \mathbb {I} \left(y _ {j} = k\right)}, \tag {11} +$$ + +which stands for the most representative features of the corresponding class. We can treat them as the visual embedding of previous classes when learning new + +ones, i.e., we construct the auxiliary training set $\mathbf{P} = \{(\mathbf{p}_1, 1), (\mathbf{p}_2, 2), \dots, (\mathbf{p}_{|\mathcal{Y}_b|}, |\mathcal{Y}_b|)\}$ , and optimize: + +$$ +\min _ {\left\{u _ {t} ^ {b}, u _ {i} ^ {b} \right\}} \sum_ {\left(\mathbf {x}, y\right) \in \mathcal {D} ^ {b} \cup \mathbf {P}} \mathcal {L} _ {t} + \mathcal {L} _ {i} + \mathcal {L} _ {c}. \tag {12} +$$ + +Specifically, the prototype set $\mathbf{P}$ is only adopted in calculating $\mathcal{L}_c$ . To enhance diversity across training stages, we also add Gaussian noise on the prototype embeddings, i.e., $\mathbf{p}' = \mathbf{p} + \epsilon$ , $\epsilon \in \mathcal{N}(0,\alpha^2\mathbf{I})$ , and utilize $\mathbf{p}'$ in Eq. 12. In this way, we can revisit previous knowledge to prevent forgetting when learning new injection units. Additionally, we also consider matching the visual features to these prototypes to further assist recognition [66, 86]. + +Summary of on-the-fly injection: We visualize the training stage knowledge injection in Figure 2 (left). ENGINE learns two sets of knowledge injection units, aiming to encode the extra knowledge into the unit module. We utilize GPT-4 to extract informative class descriptors, and utilize similarity loss to encode them into the injection units. Similarly, data augmentation is utilized to provide visual information. Hence, the model can extract more informative features as it sees more tasks. To prevent forgetting, we also design the learning objective considering class prototypes. Since injection units are implemented with a single layer, the appended parameter size $(d\times d)$ per unit is negligible for CLIP. Besides, since injection units are linear layers, they can be further re-parameterized by aggregating weights into a single one, i.e., $\sum_{p}u_{i}^{p}$ and $\sum_{p}u_{t}^{p}$ . + +# 4.2. Post-tuning Knowledge Injection + +We have introduced the way to inject external knowledge during training in Section 4.1. Apart from the model training stage, we can also utilize external knowledge to calibrate the prediction results during inference, and we call this step "post-tuning knowledge injection". For example, after the model makes inference by matching visual and textual features, we can treat the current prediction results as a preliminary result, which needs further rectification. For the top- $k$ prediction results in this preliminary result, we treat them as competitors to the ground truth that could result in confusion. Hence, we need to construct a further prediction considering the image and all local competitors, which can focus more on the top- $k$ classes to refine the prediction. + +Refining the results: After learning the injection units, we can get the prediction results using the injected features for more comprehensive features. However, the model may still make wrong predictions, e.g., classifying a "cat" into a similar class "lion" since they have similar visual features. To refine the prediction results, we design an extra consolidation process using external knowledge. Denoting the prediction results of injected feature matching in Eq. 10 as $f_{\mathrm{inj}}(\mathbf{x})$ , we first extract the top- $k$ predictions as the competitor candidates, i.e., $\{f_{\mathrm{inj},i_1}(\mathbf{x}), f_{\mathrm{inj},i_2}(\mathbf{x}), \dots, f_{\mathrm{inj},i_k}(\mathbf{x})\} = \mathrm{Top}_k(f_{\mathrm{inj}}(\mathbf{x}))$ , where $\mathrm{Top}_k(f_{\mathrm{inj}}(\mathbf{x}))$ indicates selecting the + +top- $k$ largest value in $f_{\mathrm{inj}}(\mathbf{x})$ . Their labels are denoted as: + +$$ +\left\{y _ {i _ {1}}, y _ {i _ {2}}, \dots , y _ {i _ {k}} \right\} = \left\{y _ {i} \mid f _ {\text {i n j}, i} (\mathbf {x}) \in \operatorname {T o p} _ {k} \left(f _ {\text {i n j}} (\mathbf {x})\right) \right\}. \tag {13} +$$ + +In Eq. 13, the label set $\{y_{i_1}, y_{i_2}, \ldots, y_{i_k}\}$ corresponds to the possible confusing labels for the input image. To further refine the predictions, we aim for pair-wise disambiguation among this label set. We again utilize GPT-4 to generate pair-wise discriminative features: + +Q: What are unique visual features of $[\mathrm{CLASS}]_i$ compared to $[\mathrm{CLASS}]_j$ in a photo? Focus on their key visual differences. + +A: [CLASS]i: 1. Long, thin tail that aids in balance... [CLASS]j: 1. Stockier, heavier tail with a tuft of hair at the end ... + +With this process, GPT-4 provides the most discriminative features that can differentiate two classes, e.g., the differences in "tail" to differentiate "cat" from "lion" and vice versa. Given the top- $k$ classes, we can generate their pairwise descriptors as $\mathbf{D} = [\mathbf{d}_{ij}]_{i,j=1}^{k}$ , where $\mathbf{d}_{ij}$ denotes the descriptions to differ class $i$ from class $j$ . We feed these fine-grained descriptions as the textual input, and match the query embedding to them: + +$$ +f _ {\mathrm {p t}, y _ {i}} (\mathbf {x}) = \frac {1}{k - 1} \sum_ {j = 1} ^ {j \neq i} f _ {y _ {i}} \left(\mathbf {x}, \mathbf {d} _ {i j}\right), \tag {14} +$$ + +where we utilize zero-shot CLIP in Eq. 2 to match the query image to these pair-wise discriminative features, and average the similarities as the post-tuning logits. + +Effect of post-tuning knowledge injection: We visualize the post-tuning knowledge injection in Figure 2 (right). To discriminate similar concepts and refine the wrong predictions, we first extract the top- $k$ predictions of the model, and utilize GPT-4 to generate pair-wise discriminative features locally. Eq. 14 provides the fine-grained local prediction among these top- $k$ classes, which can further help refine the results as data evolves. We adopt the zero-shot CLIP for Eq. 14, which does not require incremental updates. + +# 4.3. Summary of ENGINE + +In ENGINE, we utilize external knowledge to help continual learning in two aspects, i.e., on-the-fly injection and post-tuning injection. With the help of GPT-4, we can extract more informative descriptors suitable for the incremental learning classes, and inject the external knowledge into the knowledge units. Besides, we also extract local descriptors to discriminate pair-wise classes in the post-tuning stage to help refine the predictions. During training, we update the model via Eq. 12 to inject knowledge into the knowledge units. During inference, we aggregate the following outputs as the prediction: + +$$ +f (\mathbf {x}) = f _ {\text {i n j}} (\mathbf {x}) + f _ {\mathrm {p t}} (\mathbf {x}). \tag {15} +$$ + +Table 1. Average and last performance comparison of different methods. The best performance is shown in bold. All methods are initialized with the same pre-trained CLIP without exemplars for a fair comparison. + +
MethodAircraftCIFAR100Cars
B0 Inc10B50 Inc10B0 Inc10B50 Inc10B0 Inc10B50 Inc10
\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)
Finetune3.160.961.721.057.844.445.302.463.141.101.541.13
CoOp [88]14.547.1413.057.7747.0024.2441.2324.1236.4621.6537.4020.87
SimpleCIL [84]59.2448.0953.0548.0984.1576.6380.2076.6392.0486.8588.9686.85
ZS-CLIP [49]26.6617.2221.7017.2281.8171.3876.4971.3882.6076.3778.3276.37
L2P [65]47.1928.2944.0732.1382.7473.0381.1473.6176.6361.8276.3765.64
DualPrompt [64]44.3025.8346.0733.5781.6372.4480.1272.5776.2662.9476.8867.55
CODA-Prompt [53]45.9827.6945.1432.2882.4373.4378.6971.5880.2166.4775.0664.19
RAPF [25]50.3823.6140.4725.4486.1478.0482.1777.9382.8962.8575.8763.19
ENGINE69.6958.6964.3859.0286.9279.2283.1579.4794.1490.0891.6190.03
ImageNet-RCUBUCF
MethodB0 Inc20B100 Inc20B0 Inc20B100 Inc20B0 Inc10B50 Inc10
\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(A_B\)
Finetune1.370.431.010.882.060.640.560.474.511.591.210.80
CoOp [88]60.7337.5254.2039.7727.618.5724.0310.1447.8533.4642.0224.74
SimpleCIL [84]81.0674.4876.8474.4883.8177.5279.7577.5290.4485.6888.1285.68
ZS-CLIP [49]83.3777.1779.5777.1774.3863.0667.9663.0675.5067.6471.4467.64
L2P [65]75.9766.5272.8266.7770.8757.9375.6466.1286.3476.4383.9576.62
DualPrompt [64]76.2166.6573.2267.5869.8957.4674.4064.8485.2175.8284.3176.35
CODA-Prompt [53]77.6968.9573.7168.0573.1262.9873.9562.2187.7680.1483.0475.03
RAPF [25]81.2670.4876.1070.2379.0962.7772.8262.9392.2880.3390.3181.55
ENGINE86.2280.3783.6380.9886.6580.2082.5979.3094.3590.0392.5189.58
SUNFoodObjectNet
MethodB0 Inc30B150 Inc30B0 Inc10B50 Inc10B0 Inc20B100 Inc20
\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{A}\)\(\mathcal{A}_B\)\(\bar{\mathcal{A}}\)\(\mathcal{A}_B\)
Finetune4.511.590.780.723.491.712.141.521.340.470.690.54
CoOp [88]45.9323.1139.3324.8936.0114.1833.1318.6721.246.2916.216.82
SimpleCIL [84]82.1375.5878.6275.5887.8981.6584.7381.6552.0640.1345.1140.13
ZS-CLIP [49]79.4272.1174.9572.1187.8681.9284.7581.9238.4326.4331.1226.43
L2P [65]82.8274.5479.5773.1085.6677.3380.4273.1351.4039.3948.9142.83
DualPrompt [64]82.4674.4079.3773.0284.9277.2980.0072.7552.6240.7249.0842.92
CODA-Prompt [53]83.3475.7180.3874.1786.1878.7880.9874.1346.4934.1340.5734.13
RAPF [25]82.1372.4778.0473.1088.5781.1585.5381.1748.6727.4339.2828.73
ENGINE85.0478.5481.5778.4589.8183.8986.8983.9459.1145.1951.3244.99
+ +Table 2. Comparison to traditional exemplar-based CIL methods. ENGINE does not use any exemplars. + +
MethodExemplarsAircraft B0 Inc10SUN B0 Inc30
AABAAB
iCaRL [51]20 / class53.6043.9878.5667.30
MEMO [83]20 / class42.2425.4181.4873.45
PROOF [86]20 / class61.0053.5983.5777.28
ENGINE069.6958.6985.0478.54
+ +# 5. Experiments + +In this section, we conduct extensive experiments on nine benchmark datasets and compare ENGINE to state-of-the-art methods. We visualize the incremental learning curve and provide ablation studies on prompt engineering and parameter sensitivity to test the model's robustness. Besides, we also visualize the learned embeddings and logit changes after injecting external knowledge to show its effectiveness. We provide more results in the supplementary. + +# 5.1. Implementation Details + +Dataset: We follow [65, 86, 88] to evaluate the performance on nine benchmark datasets that have domain gap to CLIP's pre-training dataset, i.e., CIFAR100 [32], CUB200 [59], ObjectNet [6], ImageNet-R [22], FGVCAircraft [41], StanfordCars [31], Food101 [7], SUN397 [68] and UCF101 [55]. Following [86], we utilize the sampled 100 classes from CIFAR100, Aircraft, Cars, Food, UCF, 200 classes from CUB200, ObjectNet, ImageNet-R, and 300 classes from SUN to ease the data split. More details are reported in the supplementary. + +Dataset split: Following [51, 65], we use 'B- $m$ Inc- $n$ ' to split the classes in CIL. $m$ indicates the number of classes in the first stage, and $n$ represents that of every following stage. We follow [51] to randomly shuffle the class order with random seed 1993 for all compared methods, and keep this same for every method. + +![](images/224d58b8d4e13ed74089cdd54c4c30a3bb1888d6cb98c92389b491af76175e92.jpg) +(a) Aircraft Base0 Inc10 + +![](images/facb6e73795a87063d68e79b97490b31620b7ff86f510bfc888d025613de2a1a.jpg) +(b) Cars Base0 Inc10 + +![](images/42cd25a8b36a24d9279ae8e65edd0c7c3670e31cd0f8f2a681017f3cd0d1e4ce.jpg) +(c) UCF Base0 Inc10 + +Table 3. Results when all methods use the same textual description generated by GPT-4 and data augmentation. + +
MethodTemplateImageNet-R AB0 Inc20 A_BCIFAR AB0 Inc10 A_B
ZS-CLIPGPT Generated83.6877.5982.2671.70
RAPFGPT Generated81.8971.0486.4978.52
ENGINEGPT Generated86.2280.3786.9279.22
+ +Comparison methods: We first compare to SOTA pretrained model-based CIL algorithms, e.g., L2P [65], DualPrompt [64], CODA-Prompt [53], SimpleCIL [84]. Besides, we also compare to SOTA CLIP-based CIL algorithms, e.g., CoOp [88], PROOF [86], RAPF [25]. The baseline that finetunes CLIP for incremental tasks is denoted as Finetune. All methods are deployed with the same CLIP as initialization. + +Training details: The experiments are deployed with NVIDIA 4090 using PyTorch [47]. We follow [25, 86] to consider CLIP with ViT-B/16 for all compared methods for fair comparison. For vision-based methods that cannot utilize the textual prompt (e.g., L2P, DualPrompt, CODA-Prompt), we utilize CLIP's visual branch as their initialization. We report the results using LAION-400M pre-trained CLIP [26] in the main paper, and report results using OpenAI [49] in the supplementary. In ENGINE, we use SGD optimizer with a batch size of 64 to optimize the model for 10 epochs. The learning rate decays from 0.05 with cosine annealing. We set the prototype noise ratio $\alpha$ to 0.25 and $k$ in post-tuning to 5. We use OpenAI GPT-4o mini [1] to generate the prompts for textual knowledge and AutoAugment [11] to augment the visual features. We explore other LLMs to extract external knowledge in the supplementary. The source code will be publicly available upon acceptance. Evaluation metric: Following [51, 86], we use $\mathcal{A}_b$ to represent the model's accuracy after the $b$ -th stage. Specifically, we adopt $\mathcal{A}_B$ (the performance after the last stage) and $\bar{\mathcal{A}} = \frac{1}{B}\sum_{b=1}^{B}\mathcal{A}_b$ (average performance along incremental stages) as measurements. + +# 5.2. Benchmark Comparison + +We first compare ENGINE to other state-of-the-art methods on the benchmark datasets and report the results in Table 1 + +![](images/ec1ee823ba1959f74917c9e66162db42f244bde62fd25d9b84e8cc63a2e008f6.jpg) +Figure 3. Incremental performance of different methods. We report the performance gap after the last incremental stage of ENGINE and the runner-up method at the end of the line. All methods utilize the same CLIP pre-trained weight. More figures are shown in supplementary. +(a) Ablation study + +![](images/a3393fed7b92a9574c5b020a2468f0c4ee24cf6e8d599c845be95894a3165b7f.jpg) +(b) Parameter sensitivity +Figure 4. Ablation study and parameter sensitivity. + +and Figure 3. Specifically, we find ENGINE outperforms current SOTA by $3\% \sim 10\%$ on these benchmark datasets. We find the baseline method Finetune has the worst performance, indicating that the decomposition ability to tackle the class template into fine-grained features is totally forgotten. For visual prompt-based methods like L2P, DualPrompt, and CODA-Prompt, the performance is limited due to the inability to utilize textual information. For the textual prompt tuning method CoOp, we find that it achieves poor performance due to the forgetting of textual prompts. For other CLIP-based methods like RAPF, our method still shows substantial improvement, verifying its strong continual learning ability without forgetting. + +Apart from these non-exemplar-based methods, we also consider typical exemplar-based CIL methods for comparison. We report the comparison results in Table 2, where all methods are initialized with the same CLIP. As we can infer from the table, ENGINE still outperforms them even without using exemplars. + +Finally, what if all CLIP-based methods utilize the same external knowledge? Correspondingly, we also consider using the same generated prompt and data augmentation for ZS-CLIP and RAPF, and report the results in Table 3. In this table, we replace the template text of all methods using the same text in Section 4.1 and adopt the same data augmentation for image input. We find ENGINE still outperforms them even using the same prompt, indicating that ENGINE can better utilize the external information for incremental learning. + +![](images/de85726447bf967dad8ed2129b75a3a536f4fd270e8d47c2bc670b116618223b.jpg) +(a) ZS-CLIP, single stage + +![](images/ab615e14df2c4a9549f04dd7caf47bb2ce428e22913d931db6fec55654ca7a54.jpg) +(b) ENGINE, two stages +Figure 5. t-SNE [58] visualizations of zero-shot CLIP and ENGINE. ENGINE aligns the cross-modal features without forgetting. We represent the visual features of the first task with dots and classes of the second task with squares. The textual features are represented by triangles, and we utilize the shadow area to denote the decision boundary. + +# 5.3. Further Analysis + +Ablation study: We first conduct ablations to investigate the effectiveness of each component in ENGINE. We report the incremental performance on CIFAR100 B0 Inc10 in Figure 4a. We can infer that 'ZS-CLIP' has the worst performance since the downstream distribution is different from the pre-training stage, and we view it as the baseline. Correspondingly, we first equip the model with visual knowledge injection (Eq. 6), and denote the model as 'w/ Visual Injection'. It shows that the performance drastically improves as the model has been adapted to extract more informative visual features. Similarly, we then append the textual information using Eq. 5, denoted as 'w/ Dual Injection.' As shown in the figure, using external knowledge for both modalities can further improve the performance than visual knowledge only. Finally, we equip the model with post-tuning knowledge injection, and utilize Eq. 15 for inference (i.e., 'w/ Post-tuning Injection'). This corresponds to the full version of ENGINE, and we find it achieves the best performance. Ablations verify that every component in ENGINE boosts the CIL performance. + +Parameter robustness: There are two hyperparameters in ENGINE, i.e., the prototype noise scale $\alpha$ in Section 4.1 and post-tuning candidate $k$ in Eq. 14. We conduct experiments on CUB B0 Inc20 to investigate the robustness by changing these parameters. Specifically, we choose $\alpha$ among $\{0.15, 0.2, 0.25, 0.3, 0.35\}$ , and $k$ among $\{3, 4, 5, 6, 7\}$ . We report the last performance in Figure 4b. As shown in the figure, the performance is robust with the change of these parameters, and we suggest $\alpha = 0.25$ , $k = 5$ as default. + +Visualizations: In this section, we utilize t-SNE [58] to visualize the cross-modal features learned by ENGINE on CIFAR100 B0 Inc5. We represent the visual features of the first task with dots $(\bigcirc)$ and classes of the second task with squares $(\square)$ . The textual features are represented by triangles $(\triangle)$ , and we utilize the shadow area to denote the decision boundary divided by the textual features. Specifically, we first visualize the instances of the first stage in zero-shot CLIP's embedding space in Figure 5a. The results indicate + +![](images/6f24666e94e7d1c0d2e221513cfe0f00eee579da00e50e37e3cfe638fc1ccc28.jpg) +Figure 6. Left: Input. Middle: Top-5 predictions before post-tuning knowledge injection. Right: Top-5 predictions after post-tuning. More cases are shown in the supplementary. + +that pre-trained CLIP has difficulty in aligning the visual and textual features since they are situated in two distinct areas in the embedding space [37]. While for Figure 5b, we visualize the embedding of ENGINE across two incremental stages. As we can infer from the figure, ENGINE has two advantages: 1) The visual and textual features are aligned in the same cluster, and 2) The model has not forgotten previous classes when learning new ones. + +We also visualize the model output before and after post-tuning knowledge injection on ImageNet-R in Figure 6. We utilize bars with red edges to denote the ground-truth class. As we can infer from the figure, post-tuning knowledge injection can help the model refine the predictions by generating local comparisons. For the case on top of Figure 6, the model can focus on the differences between "acorn" and "birdhouse" to make the final prediction. + +# 6. Conclusion + +Class-incremental learning is essential to real-world applications. This paper aims to enable CLIP with CIL ability by injecting external knowledge. Specifically, we consider on-the-fly knowledge injection by learning dual-branch injection units. We augment images for the visual branch and generate detailed feature descriptions with GPT-4 for the textual branch. By maximizing the similarity between external information and the injection module, we encode informative features into the pre-trained CLIP. Besides, we also consider post-tuning knowledge injection with local class information to calibrate the predictions. Extensive experiments verify ENGINE's effectiveness. + +Limitations and future works: This paper relies on GPT-4 to provide expert knowledge for specific tasks, which may be impossible for the specific cases where GPT-4 cannot generalize. Future works include designing other formats of external knowledge, e.g., knowledge graphs. + +# Acknowledgments + +This work is partially supported by NSFC (U23A20382, 62476123, 62376118, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund (RAGR20240101), Postdoctoral Fellowship Program of CPSF (GZB20250389), China Postdoctoral Science Foundation (2025M771535), and Collaborative Innovation Center of Novel Software Technology and Industrialization. + +# References + +[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7 +[2] Charu C Aggarwal. A survey of stream clustering algorithms. In Data Clustering, pages 231-258. Chapman and Hall/CRC, 2018. 1 +[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2 +[4] Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars. Task-free continual learning. In CVPR, pages 11254-11263, 2019. 2 +[5] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In NeurIPS, pages 11816-11825, 2019. 2 +[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, pages 9448-9458, 2019. 6 +[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461, 2014. 6 +[8] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 2 +[9] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2018. 2 +[10] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, pages 16664-16678, 2022. 2 +[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In CVPR, pages 113-123, 2019. 7 +[12] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern + +analysis and machine intelligence, 44(7):3366-3385, 2021. 1, 2 +[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1 +[14] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class-incremental learning via relation knowledge distillation. In AAAI, pages 1255-1263, 2021. 2 +[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1 +[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020. 2 +[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285–9295, 2022. 2 +[18] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 1 +[19] Robert M French and André Ferrara. Modeling time perception in rats: Evidence for catastrophic interference in animal learning. In Proceedings of the 21st Annual Conference of the Cognitive Science Conference, pages 173-178. CiteSeer, 1999. 1 +[20] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-DFCIL: relation-guided representation learning for data-free class incremental learning. In ECCV, pages 423–439, 2022. 2 +[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2015. 1 +[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6 +[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2 +[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2 +[25] Linlan Huang, Xusheng Cao, Haori Lu, and Xialei Liu. Class-incremental learning with clip: Adaptive representation adjustment and parameter fusion. In ECCV, pages 214-231, 2024. 2, 3, 6, 7 +[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7 + +[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1 +[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727, 2022. 2, 3 +[29] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In ICCV, pages 11847-11857, 2023. 2 +[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2 +[31] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, pages 554–561, 2013. 6 +[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6 +[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1 +[34] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 1 +[35] Lan Li, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Addressing imbalanced domain-incremental learning through dual-balance collaborative experts. In ICML, 2025. 2 +[36] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, pages 614-629, 2016. 2 +[37] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, pages 17612-17625, 2022. 8 +[38] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, pages 12245-12254, 2020. 2 +[39] Yichen Lu, Mei Wang, and Weihong Deng. Augmented geometric distillation for data-free incremental person reid. In CVPR, pages 7329-7338, 2022. 2 +[40] Zilin Luo, Yaoyao Liu, Bernt Schiele, and Qianru Sun. Class-incremental exemplar compression for class-incremental learning. In CVPR, pages 11371-11380, 2023. 2 +[41] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6 +[42] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern + +Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1, 2 +[43] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2 +[44] Mark D McDonnell, Dong Gong, Amin Parveneh, Ehsan Abbasnejad, and Anton van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. In NeurIPS, 2023. 2 +[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1 +[46] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, pages 13698-13707, 2021. 2 +[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 7 +[48] Zhi-Hong Qi, Da-Wei Zhou, Yiran Yao, Han-Jia Ye, and De-Chuan Zhan. Adaptive adapter routing for long-tailed class incremental learning. Machine Learning, 114(3):1-20, 2025. 2 +[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 3, 6, 7 +[50] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In NIPS, pages 506-516, 2017. 2 +[51] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2, 6, 7 +[52] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, pages 16722-16731, 2022. 2 +[53] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, pages 11909-11919, 2023. 2, 3, 6, 7 +[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NIPS, pages 4080-4090, 2017. 2 +[55] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6 +[56] Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, DeChuan Zhan, and Han-Jia Ye. Mos: Model surgery for pretrained model-based class-incremental learning. In AAAI, pages 20699-20707, 2025. 2 + +[57] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, pages 254-270, 2020. 2 +[58] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8 +[59] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6 +[60] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022. 2 +[61] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In ICLR, 2023. 2 +[62] Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lu, and Baochang Zhang. Attriclip: A non-incremental learner for incremental knowledge learning. In CVPR, pages 3654-3663, 2023. 2 +[63] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. NeurIPS, pages 5682-5695, 2022. 2 +[64] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, pages 631-648, 2022. 2, 3, 6, 7 +[65] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. 1, 2, 3, 6, 7 +[66] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free clip-based adaptation. In ICLR, 2024. 5 +[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019. 2 +[68] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6 +[69] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, pages 899–908, 2018. 2 +[70] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021. 2 +[71] Xu Yang, Zhangzikang Li, Haiyang Xu, Hanwang Zhang, Qinghao Ye, Chenliang Li, Ming Yan, Yu Zhang, Fei Huang, and Songfang Huang. Learning trajectory-word alignments for video-language tasks. In ICCV, pages 2504-2514, 2023. 1 +[72] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. In NeurIPS, pages 40924-40943, 2023. 1 + +[73] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved clip classification. In CVPR, pages 27402-27411, 2024. 2 +[74] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018. 2 +[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2024. 1 +[76] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2, 3 +[77] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, pages 6982-6991, 2020. 2 +[78] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1 +[79] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017. 2 +[80] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13208-13217, 2020. 2 +[81] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In ICML, pages 61649–61663, 2024. 2 +[82] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Task-agnostic guided feature expansion for class incremental learning. In CVPR, pages 10099-10109, 2025. 2 +[83] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In ICLR, 2023. 2, 6 +[84] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision, 133:1012-1032, 2025. 2, 6, 7 +[85] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, Lijun Zhang, and De-Chuan Zhan. Dual consolidation for pre-trained model-based domain-incremental learning. In CVPR, pages 20547–20557, 2025. 2 +[86] Da-Wei Zhou, Yuanhan Zhang, Yan Wang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(6): 4489-4504, 2025. 2, 3, 5, 6, 7 +[87] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2 + +[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 6, 7 +[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, pages 5871-5880, 2021. 2 \ No newline at end of file diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/images.zip b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/images.zip new file mode 100644 index 0000000000000000000000000000000000000000..506d49022550455ae043b4c0f6401bfd8cdba110 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/images.zip @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:96fbc0562b920d841382c2b6769915a5ba60849957135b149ee8103ea06c0078 +size 677851 diff --git a/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/layout.json b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/layout.json new file mode 100644 index 0000000000000000000000000000000000000000..01f182f9d991313607e00b406e5c5d49313025c2 --- /dev/null +++ b/2025/External Knowledge Injection for CLIP-Based Class-Incremental Learning/layout.json @@ -0,0 +1,10532 @@ +{ + "pdf_info": [ + { + "para_blocks": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "spans": [ + { + "bbox": [ + 73, + 103, + 538, + 121 + ], + "type": "text", + "content": "External Knowledge Injection for CLIP-Based Class-Incremental Learning" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "spans": [ + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": "Da-Wei Zhou" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": ", Kai-Wen Li" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": ", Jingyi Ning" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": ", Han-Jia Ye" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1,2(\\boxtimes)}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": ", Lijun Zhang" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": ", De-Chuan Zhan" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1,2}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "inline_equation", + "content": "^{1}" + }, + { + "bbox": [ + 67, + 142, + 542, + 171 + ], + "type": "text", + "content": " School of Artificial Intelligence, Nanjing University" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 113, + 171, + 496, + 186 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 113, + 171, + 496, + 186 + ], + "spans": [ + { + "bbox": [ + 113, + 171, + 496, + 186 + ], + "type": "inline_equation", + "content": "^{2}" + }, + { + "bbox": [ + 113, + 171, + 496, + 186 + ], + "type": "text", + "content": " National Key Laboratory for Novel Software Technology, Nanjing University" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 116, + 187, + 490, + 200 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 116, + 187, + 490, + 200 + ], + "spans": [ + { + "bbox": [ + 116, + 187, + 490, + 200 + ], + "type": "text", + "content": "{zhoudw,likw,yehj,zhanglj,zhandc}@lamda.nju.edu.cn，ningjy@nju.edu.cn" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "spans": [ + { + "bbox": [ + 151, + 227, + 200, + 239 + ], + "type": "text", + "content": "Abstract" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 54, + 253, + 297, + 564 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 253, + 297, + 564 + ], + "spans": [ + { + "bbox": [ + 54, + 253, + 297, + 564 + ], + "type": "text", + "content": "Class-Incremental Learning (CIL) enables learning systems to continuously adapt to evolving data streams. With the advancement of pre-training, leveraging pre-trained vision-language models (e.g., CLIP) offers a promising starting point for CIL. However, CLIP makes decisions by matching visual embeddings to class names, overlooking the rich contextual information conveyed through language. For instance, the concept of \"cat\" can be decomposed into features like tail, fur, and face for recognition. Besides, since the model is continually updated, these detailed features are overwritten in CIL, requiring external knowledge for compensation. In this paper, we introduce ExterNal knowledgeGe INjEction (ENGINE) for CLIP-based CIL. To enhance knowledge transfer from outside the dataset, we propose a dual-branch injection tuning framework that encodes informative knowledge from both visual and textual modalities. The visual branch is enhanced with data augmentation to enrich the visual features, while the textual branch leverages GPT-4 to rewrite discriminative descriptors. In addition to this on-the-fly knowledge injection, we also implement post-tuning knowledge by re-ranking the prediction results during inference. With the injected knowledge, the model can better capture informative features as data evolves. Extensive experiments demonstrate its state-of-the-art performance. Code is available at: https://github.com/LAMDA-CL/ICCV25-ENGINE." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "spans": [ + { + "bbox": [ + 55, + 588, + 137, + 601 + ], + "type": "text", + "content": "1. Introduction" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "spans": [ + { + "bbox": [ + 55, + 609, + 296, + 693 + ], + "type": "text", + "content": "Recent advancements in deep learning have significantly impacted various aspects of life [13, 21, 72]. However, real-world data often presents substantial challenges to these models, particularly streaming data that requires continual learning [2, 34]. Class-Incremental Learning (CIL) [51] has been proposed to address this challenge by enabling models to absorb new knowledge incrementally." + } + ] + } + ], + "index": 9 + }, + { + "type": "image", + "bbox": [ + 326, + 211, + 544, + 399 + ], + "blocks": [ + { + "bbox": [ + 326, + 211, + 544, + 399 + ], + "lines": [ + { + "bbox": [ + 326, + 211, + 544, + 399 + ], + "spans": [ + { + "bbox": [ + 326, + 211, + 544, + 399 + ], + "type": "image", + "image_path": "5b4db64d4b85d37e186f514ef85db0823d9b7c8d92f9561e3fec5787c5d3fe69.jpg" + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 403, + 555, + 570 + ], + "lines": [ + { + "bbox": [ + 313, + 403, + 555, + 570 + ], + "spans": [ + { + "bbox": [ + 313, + 403, + 555, + 570 + ], + "type": "text", + "content": "Figure 1. Last accuracy comparison on nine benchmark datasets. All methods are deployed with the same CLIP for fair comparison. ENGINE outperforms other competitors with a substantial margin. A major issue in CIL is catastrophic forgetting [18, 19], where learning new classes leads to the loss of previously acquired knowledge. As a result, much research has focused on mitigating forgetting in deep learning [12, 42]. In contrast to traditional approaches that train models from scratch, recent progress in pre-training has shifted attention towards using pre-trained models (PTMs) [65], such as Vision Transformers[15] and CLIP [49]. With extensive training data serving as pre-learned knowledge, PTMs offer a more generalizable starting point for CIL, showing promise for real-world applications." + } + ] + } + ], + "index": 11, + "angle": 0, + "type": "image_caption" + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 570, + 556, + 715 + ], + "type": "text", + "content": "CLIP [49], as a pioneering pre-trained vision-language model [27, 33, 71, 75, 78], exhibits strong zero-shot performance by matching the embeddings of query images to class names. However, since visual features often contain fine-grained information [45], these detailed descriptors are neglected when using class names as matching targets. For example, the text concept of \"a photo of a cat\" can be further decomposed into features like \"whiskers,\" \"tail,\" and \"fur\" in order to match the visual embeddings, while these informative features are not well utilized. Furthermore, as the model is updated in CIL, the sub-features can be overwritten during the learning process, resulting in" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "spans": [ + { + "bbox": [ + 65, + 2, + 111, + 34 + ], + "type": "text", + "content": "CVF" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "header", + "angle": 0, + "lines": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "spans": [ + { + "bbox": [ + 146, + 0, + 494, + 37 + ], + "type": "text", + "content": "This ICCV paper is the Open Access version, provided by the Computer Vision Foundation. Except for this watermark, it is identical to the accepted version; the final published version of the proceedings is available on IEEE Xplore." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 66, + 702, + 257, + 714 + ], + "type": "page_footnote", + "angle": 0, + "lines": [ + { + "bbox": [ + 66, + 702, + 257, + 714 + ], + "spans": [ + { + "bbox": [ + 66, + 702, + 257, + 714 + ], + "type": "text", + "content": "†Correspondence to: Han-Jia Ye (yehj@lamda.nju.edu.cn)" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 316, + 757 + ], + "type": "text", + "content": "3314" + } + ] + } + ], + "index": 14 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 0 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 286, + 83 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 286, + 83 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 286, + 83 + ], + "type": "text", + "content": "forgetting and a mismatch between cross-modal features." + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "spans": [ + { + "bbox": [ + 54, + 84, + 294, + 191 + ], + "type": "text", + "content": "Recent studies [62, 65, 73, 86] have highlighted the limitations of using fixed template text in vision-language models, which often leads to a loss of rich contextual information. To address this, researchers have proposed learnable text inputs, such as task-specific prompts [88], which allow the model to encode additional, task-relevant information. However, these learnable inputs can become biased towards the visual features in the training set, ultimately compromising the model's generalizability [87]." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 54, + 192, + 294, + 287 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 192, + 294, + 287 + ], + "spans": [ + { + "bbox": [ + 54, + 192, + 294, + 287 + ], + "type": "text", + "content": "In CLIP, a more detailed description in the textual space can provide a more informative matching target, enhancing prediction accuracy. For example, instead of using the generic template \"a photo of a cat,\" directly describing its features, such as \"whiskers\" and \"tail,\" can significantly improve recognition. In this way, these detailed features serve as anchors in the embedding space, and aligning to them helps reduce the risk of catastrophic forgetting." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 54, + 288, + 294, + 418 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 288, + 294, + 418 + ], + "spans": [ + { + "bbox": [ + 54, + 288, + 294, + 418 + ], + "type": "text", + "content": "Learning with informative descriptors offers a promising approach to enhancing CLIP's continual learning capabilities. However, two major challenges remain: 1) acquiring informative features. Given the diversity of classes in downstream tasks, it is impractical for human experts to annotate all relevant features for recognition. 2) mitigating forgetting when utilizing external knowledge. As CIL involves continuous data streams, the model risks forgetting previously acquired knowledge after each update. Therefore, it is crucial to efficiently extract and retain external knowledge while preventing forgetting during sequential updates." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 54, + 419, + 294, + 587 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 419, + 294, + 587 + ], + "spans": [ + { + "bbox": [ + 54, + 419, + 294, + 587 + ], + "type": "text", + "content": "To address these challenges, we propose ExterNal knowledgeGe InjEction (ENGINE) for CLIP-based CIL. To acquire informative features, we enhance knowledge transfer by incorporating external knowledge from both visual and textual modalities through a dual-branch tuning framework. The visual branch is enhanced via data augmentation to enrich visual features, while the textual branch uses GPT-4 to rewrite discriminative prompts. In this way, we can encode diverse and informative class descriptions in the continual updating process. In addition to on-the-fly knowledge injection, we also apply post-tuning by re-ranking predictions during inference. Post-tuning is done by considering the local pair-wise features, which can further calibrate the wrong predictions to boost incremental learning." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 597, + 142, + 609 + ], + "type": "text", + "content": "2. Related Work" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 54, + 617, + 294, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 617, + 294, + 713 + ], + "spans": [ + { + "bbox": [ + 54, + 617, + 294, + 713 + ], + "type": "text", + "content": "Class-Incremental Learning (CIL): is a long-standing problem, aiming to absorb new knowledge without forgetting [12, 42]. Typical CIL algorithms train a model from scratch, which can be divided into several groups. Knowledge distillation-based methods aim to build the mapping target [23] between old and new models to resist forgetting [16, 36, 51], which involve logit-wise alignment [36, 51], feature-wise alignment [24, 39, 46], and" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 553, + 215 + ], + "type": "text", + "content": "group-wise alignment [14, 20, 57]. Replay-based methods aim to recover previous knowledge by saving and replaying the subset of seen classes [5, 8, 9, 38, 40, 81]. Besides, parameter regularization-based methods estimate the importance of parameters and restrict important ones not to change [3, 4, 30, 79]. Model rectification-based methods observe and rectify the inductive bias in the model, e.g., the biased logits [52, 77, 80] and classifier weights [67, 80]. Model expansion-based methods [60, 70, 83] adjust the network structure to fit the characteristics of the evolving data, e.g., expanding neurons [69, 74], the whole backbone [60, 61, 70, 82, 83], and lightweight modules [17]." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 216, + 553, + 479 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 216, + 553, + 479 + ], + "spans": [ + { + "bbox": [ + 313, + 216, + 553, + 479 + ], + "type": "text", + "content": "Pre-Trained Model-Based CIL: improves the capability of CIL by starting with generalizable pre-trained models [35, 48, 56, 85]. To maintain the pre-trained knowledge, most works seek to freeze the backbone and append lightweight modules, e.g., prompts [53, 63-65, 88] and adapters [10, 50, 76]. For example, L2P [65] and DualPrompt [64] design the visual prompt [28] pool to select the instance-specific prompts when learning with the pre-trained Vision Transformer. Several works build complex prompt combination/generation targets with attention mechanism [53] or generative network [29]. Other works directly utilize the pre-trained features to construct strong classifiers by matching class prototypes to the embeddings [44, 54, 84]. When the pre-trained CLIP is available as initialization, existing works seek to enhance cross-modal matching information via learning multi-modal prompts [62, 63]. MOE-Adapter [76] further extends the lightweight module selection process via mixture-of-experts [43], while PROOF [86] extends the representation ability of CLIP by appending new projection layers for new tasks. RAPF [25] learns the adapter modules for new tasks by decomposed parameter fusion to resist forgetting in the updating process." + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 491, + 399, + 504 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 491, + 399, + 504 + ], + "spans": [ + { + "bbox": [ + 313, + 491, + 399, + 504 + ], + "type": "text", + "content": "3. Preliminaries" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 511, + 469, + 525 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 511, + 469, + 525 + ], + "spans": [ + { + "bbox": [ + 313, + 511, + 469, + 525 + ], + "type": "text", + "content": "3.1. Class-Incremental Learning" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "spans": [ + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": "CIL is designed to continually build a unified classifier for all seen classes of the data stream [51]. We denote training sets as " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "\\{\\mathcal{D}^1,\\mathcal{D}^2,\\dots ,\\mathcal{D}^B\\}" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": ", where each task " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^b = \\{(\\mathbf{x}_i,y_i)\\}_{i = 1}^{n_b}" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " contains " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "n_b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " instances. Each training instance " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "\\mathbf{x}_i\\in \\mathbb{R}^D" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " belongs to class " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "y_{i}\\in Y_{b}" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": ", i.e., the label space of task " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " is " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "Y_{b}" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": ". We have " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "Y_{b}\\cap Y_{b^{\\prime}} = \\emptyset" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " for " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "b\\neq b^{\\prime}" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": ". This paper follows the exemplar-free CIL setting [64, 65, 89], where the model cannot hold any historical instances in the memory. In other words, we can only access data from " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{D}^b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " for model training when learning the " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": "-th incremental task. In CIL, the target is to build a unified classifier for all seen classes " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "\\mathcal{V}_b = Y_1\\cup \\dots Y_b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": ", i.e., find a model " + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x}):X\\to \\mathcal{V}_b" + }, + { + "bbox": [ + 313, + 529, + 553, + 685 + ], + "type": "text", + "content": " that minimizes the expected risk:" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 342, + 695, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 342, + 695, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 342, + 695, + 553, + 715 + ], + "type": "interline_equation", + "content": "f ^ {*} = \\underset {f \\in \\mathcal {H}} {\\operatorname {a r g m i n}} \\mathbb {E} _ {(\\mathbf {x}, y) \\sim \\mathcal {D} _ {t} ^ {1} \\cup \\dots \\mathcal {D} _ {t} ^ {b}} \\mathbb {I} (y \\neq f (\\mathbf {x})). \\tag {1}", + "image_path": "dcbe34a18e19d57babbbc7715a748f1aef688056c93a286d6906bd2100a8b26d.jpg" + } + ] + } + ], + "index": 12 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3315" + } + ] + } + ], + "index": 13 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 1 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "In Eq. 1, " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{H}" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " represents the hypothesis space and " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbb{I}(\\cdot)" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " is the indicator function. " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathcal{D}_t^b" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " denotes the " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-th task's data distribution. In this paper, we follow [25, 76, 86] and assume that a pre-trained CLIP model (contrastive language-image pre-training) [49] is available as the initialization for " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "f(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ". Specifically, CLIP contains the visual and textual encoder, i.e., " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "g_i(\\cdot): \\mathbb{R}^D \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "g_t(\\cdot): \\mathbb{R}^{Dt} \\to \\mathbb{R}^d" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": ", where the " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "D / Dt" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-dimensional images and texts are projected to the same " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "d" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "-dimensional embedding space. During inference, it constructs the template text " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": " using the class name, i.e., \"a photo of a " + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLASS}]_i" + }, + { + "bbox": [ + 55, + 72, + 296, + 216 + ], + "type": "text", + "content": "\", and matches the query image to text embeddings of all classes:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 75, + 221, + 295, + 284 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 75, + 221, + 295, + 284 + ], + "spans": [ + { + "bbox": [ + 75, + 221, + 295, + 284 + ], + "type": "interline_equation", + "content": "\\begin{array}{l} f _ {y _ {i}} (\\mathbf {x}, \\mathbf {t} _ {i}) = \\frac {\\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {i}\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(\\mathbf {z} , \\mathbf {w} _ {j}\\right) / \\tau\\right)} \\tag {2} \\\\ = \\frac {\\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {Y} _ {b} |} \\exp \\left(\\cos \\left(g _ {i} (\\mathbf {x}) , g _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}, \\\\ \\end{array}", + "image_path": "97c5d30284df57faeea6aff8f523c2a390ab1a03cb62f7fe49cbca50e50fd085.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "spans": [ + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\cos (\\cdot ,\\cdot)" + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": " is cosine similarity and " + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\tau" + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": " is temperature. " + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{z} = g_{i}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "inline_equation", + "content": "\\mathbf{w}_i = g_t(\\mathbf{t}_i)" + }, + { + "bbox": [ + 55, + 289, + 296, + 338 + ], + "type": "text", + "content": " are visual and textual features, respectively. In Eq. 2, the logit is assigned by the relative similarity of the visual embedding to textual embedding." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 343, + 268, + 357 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 343, + 268, + 357 + ], + "spans": [ + { + "bbox": [ + 55, + 343, + 268, + 357 + ], + "type": "text", + "content": "3.2. Baselines in Class-Incremental Learning" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "spans": [ + { + "bbox": [ + 55, + 361, + 295, + 385 + ], + "type": "text", + "content": "To overcome forgetting, there are two typical solutions for CIL with pre-trained models by learning prompts [28, 88]." + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "spans": [ + { + "bbox": [ + 55, + 385, + 296, + 456 + ], + "type": "text", + "content": "Learning Visual Prompts: As fully-finetuning will harm the generalizability of the PTM, several works [28, 53, 64, 65] propose to freeze the pre-trained visual encoder and append visual prompts for subsequent tasks. They design a prompt pool to select instance-specific prompts, with which the visual embedding is represented as:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 143, + 464, + 295, + 476 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 143, + 464, + 295, + 476 + ], + "spans": [ + { + "bbox": [ + 143, + 464, + 295, + 476 + ], + "type": "interline_equation", + "content": "\\mathbf {z} = \\bar {g} _ {i} (\\mathbf {x}, \\mathcal {P}), \\tag {3}", + "image_path": "728c4cda7ce94c23c316d6bc28e43981c6fcc6f26233f5f7b3baa62704ab201a.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "spans": [ + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\bar{g}_i" + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "text", + "content": " is the frozen image encoder, " + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "inline_equation", + "content": "\\mathcal{P}" + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "text", + "content": " is the prompt pool for prompt selection. By replacing " + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "inline_equation", + "content": "g_i(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 483, + 295, + 531 + ], + "type": "text", + "content": " in Eq. 2 with Eq. 3, the model efficiently encodes downstream information into the visual prompts and tackles forgetting." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 532, + 296, + 615 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 532, + 296, + 615 + ], + "spans": [ + { + "bbox": [ + 55, + 532, + 296, + 615 + ], + "type": "text", + "content": "Learning Textual Prompts: Since CLIP contains two branches for the visual and textual encoder, there are also works learning textual prompts instead of using the template text. Specifically, the textual input is formulated into a set of learnable prompts in CoOp [88], i.e., " + }, + { + "bbox": [ + 55, + 532, + 296, + 615 + ], + "type": "inline_equation", + "content": "\\tilde{\\mathbf{t}}_i = [\\mathrm{V}]_1[\\mathrm{V}]_2\\cdots[\\mathrm{V}]_M[\\mathrm{CLASS}]_i" + }, + { + "bbox": [ + 55, + 532, + 296, + 615 + ], + "type": "text", + "content": ". In this way, the textual embedding is denoted as:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 145, + 621, + 295, + 635 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 145, + 621, + 295, + 635 + ], + "spans": [ + { + "bbox": [ + 145, + 621, + 295, + 635 + ], + "type": "interline_equation", + "content": "\\mathbf {w} _ {i} = \\bar {g} _ {t} (\\tilde {\\mathbf {t}} _ {i}), \\tag {4}", + "image_path": "6d9e5711b52151dbd8187a4b6140b98562d0120df415c9d1cc7066d6462c9d14.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "spans": [ + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "content": "where " + }, + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "inline_equation", + "content": "\\bar{g}_t" + }, + { + "bbox": [ + 55, + 642, + 295, + 677 + ], + "type": "text", + "content": " is the frozen text encoder. Similarly, we can replace the textual embedding in Eq. 2 via Eq. 4 to encode task information into these learnable prompts." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 677, + 296, + 713 + ], + "type": "text", + "content": "Discussions: Eq. 3 and Eq. 4 focus on different aspects to adjust the pre-trained CLIP into downstream tasks. However, when adapting the model with only visual prompts," + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 555, + 228 + ], + "type": "text", + "content": "the problem occurs since the model has to decompose the class name into fine-grained concepts to match the visual features. The sequential updating process will cause the degradation of such decomposition ability, resulting in the mismatch between visual and textual features. By contrast, if we utilize textual prompts, the model is only focused on the specific training instance and lacks holistic information about the specific class. When the testing data contains the image of the same class with different distributions, the model shall be confused due to the prompt-level overfitting. Hence, the ideal text input should be general and holistic to match the visual features and convey all informative features that the target class possesses." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 313, + 238, + 534, + 251 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 238, + 534, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 238, + 534, + 251 + ], + "type": "text", + "content": "4. ENGINE: External Knowledge Injection" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 258, + 555, + 437 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 258, + 555, + 437 + ], + "spans": [ + { + "bbox": [ + 313, + 258, + 555, + 437 + ], + "type": "text", + "content": "Noting that prompt-based methods are highly limited by the training instances, we need to extract comprehensive information for the model to adapt to downstream tasks while preventing forgetting. To encode holistic information into the model, we seek help from large language models like GPT-4 [1] to provide the general descriptions instead of using the template features. Apart from using detailed descriptions in the training process, we also make full use of them during inference by seeking pair-wise discriminative features. On the other hand, as external knowledge is continually injected into the model, we also design the injection unit expansion and prototype replay strategy. In the following sections, we first introduce how to inject external knowledge during model training and then introduce how to utilize it during inference." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 444, + 486, + 456 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 444, + 486, + 456 + ], + "spans": [ + { + "bbox": [ + 313, + 444, + 486, + 456 + ], + "type": "text", + "content": "4.1. On-the-fly Knowledge Injection" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 461, + 555, + 593 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 461, + 555, + 593 + ], + "spans": [ + { + "bbox": [ + 313, + 461, + 555, + 593 + ], + "type": "text", + "content": "Since the target is to encode general information into the CLIP during incremental learning, we first design the model training pipeline that can enhance recognition. For example, a \"cat\" contains visual features like \"soft, short fur\", \"long, thin tail\", and \"round face with large eyes\". It would be better to match the cat image to these visual features than \"a photo of a cat\", since those features contain more discriminative and fine-grained information. Correspondingly, to enhance CLIP's knowledge of general class description, we seek help from GPT-4 [1] to provide discriminative visual features:" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 333, + 599, + 533, + 623 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 599, + 533, + 623 + ], + "spans": [ + { + "bbox": [ + 333, + 599, + 533, + 623 + ], + "type": "text", + "content": "Q: What are unique visual features of " + }, + { + "bbox": [ + 333, + 599, + 533, + 623 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLASS}]_i" + }, + { + "bbox": [ + 333, + 599, + 533, + 623 + ], + "type": "text", + "content": " in a photo? Focus on the key visual features." + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 333, + 623, + 520, + 635 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 623, + 520, + 635 + ], + "spans": [ + { + "bbox": [ + 333, + 623, + 520, + 635 + ], + "type": "text", + "content": "A: 1. Long, thin tail that aids in balance. 2." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "spans": [ + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": "In this way, we can get a set of visual descriptions for each class, denoted as " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": ". A naive solution is to replace the template text " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{t}_i" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " into " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_i" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " during training, while we propose a more intuitive way to inject the knowledge into the model. Specifically, we design a knowledge injection unit " + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "inline_equation", + "content": "u_t" + }, + { + "bbox": [ + 313, + 642, + 555, + 713 + ], + "type": "text", + "content": " to dynamically encode these textual descriptors. We implement" + } + ] + } + ], + "index": 19 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3316" + } + ] + } + ], + "index": 20 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 2 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 69, + 54, + 375, + 200 + ], + "blocks": [ + { + "bbox": [ + 69, + 54, + 375, + 200 + ], + "lines": [ + { + "bbox": [ + 69, + 54, + 375, + 200 + ], + "spans": [ + { + "bbox": [ + 69, + 54, + 375, + 200 + ], + "type": "image", + "image_path": "6908e6de466c635a8ac1e49078c1e9325cb9036237e3cecd611ac12ae3ebb1a5.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 155, + 201, + 286, + 213 + ], + "lines": [ + { + "bbox": [ + 155, + 201, + 286, + 213 + ], + "spans": [ + { + "bbox": [ + 155, + 201, + 286, + 213 + ], + "type": "text", + "content": "(a) On-the-fly Knowledge Injection" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 381, + 58, + 542, + 201 + ], + "blocks": [ + { + "bbox": [ + 381, + 58, + 542, + 201 + ], + "lines": [ + { + "bbox": [ + 381, + 58, + 542, + 201 + ], + "spans": [ + { + "bbox": [ + 381, + 58, + 542, + 201 + ], + "type": "image", + "image_path": "a28fd7ad1910454bf814df56f2412adf9cbad31e548508ae144600f62b9d0d14.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 396, + 201, + 531, + 213 + ], + "lines": [ + { + "bbox": [ + 396, + 201, + 531, + 213 + ], + "spans": [ + { + "bbox": [ + 396, + 201, + 531, + 213 + ], + "type": "text", + "content": "(b) Post-tuning Knowledge Injection" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 218, + 555, + 273 + ], + "lines": [ + { + "bbox": [ + 55, + 218, + 555, + 273 + ], + "spans": [ + { + "bbox": [ + 55, + 218, + 555, + 273 + ], + "type": "text", + "content": "Figure 2. Illustration of ENGINE. Left: On-the-fly knowledge injection. We utilize GPT-4 to provide visual features of each class, and append injection units to encode external knowledge into the model. We also adopt random augmentation to enhance image diversity using the same network structure. To overcome forgetting, we append the prototypes of previous classes in each learning stage. Right: Post-tuning knowledge injection. We extract top-" + }, + { + "bbox": [ + 55, + 218, + 555, + 273 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 218, + 555, + 273 + ], + "type": "text", + "content": " predictions and generate pair-wise discriminative features locally. The input image is forwarded with every textual feature to further refine the prediction." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "spans": [ + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "type": "inline_equation", + "content": "u_{t}" + }, + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "type": "text", + "content": " with a linear layer, i.e., " + }, + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "type": "inline_equation", + "content": "u_{t}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 274, + 296, + 299 + ], + "type": "text", + "content": ", and append it after the textual encoder. We have the similarity loss:" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 103, + 306, + 295, + 321 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 103, + 306, + 295, + 321 + ], + "spans": [ + { + "bbox": [ + 103, + 306, + 295, + 321 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {t} = - \\operatorname {S i m} \\left(u _ {t} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right), \\bar {g} _ {t} \\left(\\mathbf {d} _ {i}\\right)\\right), \\tag {5}", + "image_path": "fbd49a1fd6988f01704272c7fc1a856ae56961861fe9b08cdb19ffb454d07dec.jpg" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "spans": [ + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": "where we use cosine similarity to calculate " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathrm{Sim}(\\cdot, \\cdot)" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": ". In Eq. 5, we maximize the similarity between " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "u_{t}\\left(\\bar{g}_{t}\\left(\\mathbf{t}_{i}\\right)\\right)" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\bar{g}_{t}\\left(\\mathbf{d}_{i}\\right)" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": ", i.e., making the adapted textual feature similar to external descriptions. Since the textual encoder " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\bar{g}_{t}" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": " is frozen, we can inject this external information into the injection unit " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "u_{t}" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": ". In this way, the extracted textual features will contain detailed descriptions even using the template text, e.g., highlighting the features of \"long thin tail\" and \"round face\" when the input is \"a photo of a cat\". In the implementation, since GPT-4 will output a set of descriptions, we randomly choose one of the descriptions " + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{i}" + }, + { + "bbox": [ + 55, + 327, + 296, + 470 + ], + "type": "text", + "content": " to calculate Eq. 5 in each iteration." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "spans": [ + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": "Learning external visual knowledge: Similar to the textual branch, we can also inject visual information into the model during training. Given that we cannot fetch external data in the learning process, we utilize random data augmentation to adjust the input images, and denote the augmented input as " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "\\mathcal{A}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": ". Correspondingly, we can learn the image knowledge injection unit " + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "inline_equation", + "content": "u_{i}(\\cdot):\\mathbb{R}^{d}\\to \\mathbb{R}^{d}" + }, + { + "bbox": [ + 55, + 471, + 296, + 567 + ], + "type": "text", + "content": ", and maximize the similarity between augmented and vanilla input:" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 99, + 574, + 295, + 587 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 99, + 574, + 295, + 587 + ], + "spans": [ + { + "bbox": [ + 99, + 574, + 295, + 587 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {i} = - \\operatorname {S i m} \\left(u _ {i} (\\bar {g} _ {i} (\\mathbf {x})), \\bar {g} _ {i} (\\mathcal {A} (\\mathbf {x}))\\right). \\tag {6}", + "image_path": "00bc0a799d8dcd9494d6347d70407fb54bdede494a2cc0cb801bb8d095a45789.jpg" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "spans": [ + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": "Similar to Eq. 5, Eq. 6 freezes the visual encoder " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "\\bar{g}_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": ", and only optimizes the image injection unit " + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "inline_equation", + "content": "u_i" + }, + { + "bbox": [ + 55, + 594, + 295, + 653 + ], + "type": "text", + "content": ". In the matching process, the model maximizes the similarity of the adapted visual feature to the augmented view, which helps the model extract diverse visual information." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 654, + 296, + 714 + ], + "type": "text", + "content": "Learning task-specific injection unit: In CIL, new tasks will emerge as data evolves, and sequentially tuning a single injection unit will result in forgetting previous knowledge. Consequently, we conduct the injection process for each incremental task to avoid forgetting. Specifically, we" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "spans": [ + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "text", + "content": "initialize a new injection unit " + }, + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "inline_equation", + "content": "u_{t}^{b}, u_{i}^{b}" + }, + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "text", + "content": " when learning the " + }, + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 313, + 274, + 555, + 323 + ], + "type": "text", + "content": "-th task, and freeze all previous injection units when learning the current task. With the set of visual and textual units, we utilize the summation of them for the final representation:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 350, + 327, + 555, + 346 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 350, + 327, + 555, + 346 + ], + "spans": [ + { + "bbox": [ + 350, + 327, + 555, + 346 + ], + "type": "interline_equation", + "content": "G _ {i} (\\mathbf {x}) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {i} ^ {p} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) + u _ {i} ^ {b} \\left(\\bar {g} _ {i} (\\mathbf {x})\\right) \\qquad (7)", + "image_path": "16d2c3a11a08b34340de184af2ab038db2d07783e7314744c81a0d3f2ba114c3.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 348, + 350, + 554, + 366 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 348, + 350, + 554, + 366 + ], + "spans": [ + { + "bbox": [ + 348, + 350, + 554, + 366 + ], + "type": "interline_equation", + "content": "G _ {t} \\left(\\mathbf {t} _ {i}\\right) = \\sum_ {p = 1} ^ {b - 1} \\bar {u} _ {t} ^ {p} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) + u _ {t} ^ {b} \\left(\\bar {g} _ {t} \\left(\\mathbf {t} _ {i}\\right)\\right) \\tag {8}", + "image_path": "e0cd628f6205544fe10c16290098d597fd0df578d0571f278281c8752eef6858.jpg" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 368, + 554, + 415 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 368, + 554, + 415 + ], + "spans": [ + { + "bbox": [ + 313, + 368, + 554, + 415 + ], + "type": "text", + "content": "In the original CLIP, we have the cross-modal contrastive loss to match the image to the corresponding text description. Correspondingly, we utilize the injected visual and textual features to calculate the contrastive loss:" + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 349, + 422, + 554, + 436 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 349, + 422, + 554, + 436 + ], + "spans": [ + { + "bbox": [ + 349, + 422, + 554, + 436 + ], + "type": "interline_equation", + "content": "\\mathcal {L} _ {c} = \\ell \\left(f _ {\\text {i n j}} (\\mathbf {x}), y\\right) \\quad \\text {w h e r e} \\tag {9}", + "image_path": "dad9cc44af4ed1806b8df97ff79d99588f652c7aaa4fda9d8f45205b2756a3a6.jpg" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 324, + 437, + 553, + 468 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 324, + 437, + 553, + 468 + ], + "spans": [ + { + "bbox": [ + 324, + 437, + 553, + 468 + ], + "type": "interline_equation", + "content": "f _ {\\text {i n j}, y _ {i}} (\\mathbf {x}) = \\frac {\\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}) , G _ {t} (\\mathbf {t} _ {i})\\right) / \\tau\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {V} _ {b} |} \\exp \\left(\\cos \\left(G _ {i} (\\mathbf {x}), G _ {t} (\\mathbf {t} _ {j})\\right) / \\tau\\right)}. \\tag {10}", + "image_path": "5f04d38c8abc65295d77599133c87cc83e4292c233a9e8cf03a12f1971a74142.jpg" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 472, + 554, + 508 + ], + "type": "text", + "content": "In Eq. 9, we utilize the injected visual and textual features to calculate the contrastive loss, aiming to increase the similarity of pair-wise injected features." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "spans": [ + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": "Preventing forgetting of injection unit In the updating process, we combine Eq. 5, Eq. 6, and Eq. 9 to update the model. Since we cannot hold historical instances for replay, the model may still forget previous concepts due to the sequential updating. To this end, utilizing the substitution of previous concepts can create a calibration across all seen classes. Observing that the visual backbone " + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "inline_equation", + "content": "g_{i}" + }, + { + "bbox": [ + 313, + 509, + 555, + 640 + ], + "type": "text", + "content": " is not updating throughout the learning process, we generate a subset of previous classes using visual prototypes. Specifically, during the learning process, we can calculate the average visual embedding of each class as:" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 380, + 645, + 553, + 673 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 380, + 645, + 553, + 673 + ], + "spans": [ + { + "bbox": [ + 380, + 645, + 553, + 673 + ], + "type": "interline_equation", + "content": "\\mathbf {p} _ {k} = \\frac {\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right) g _ {i} \\left(\\mathbf {x} _ {j}\\right)}{\\sum_ {j = 1} ^ {| \\mathcal {D} ^ {b} |} \\mathbb {I} \\left(y _ {j} = k\\right)}, \\tag {11}", + "image_path": "4561618e1e8c09ce7115794d2147fd8a92fc4c96d1f4d1376eb14ad1fb1c1257.jpg" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 677, + 554, + 714 + ], + "type": "text", + "content": "which stands for the most representative features of the corresponding class. We can treat them as the visual embedding of previous classes when learning new" + } + ] + } + ], + "index": 21 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3317" + } + ] + } + ], + "index": 22 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 3 + }, + { + "para_blocks": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "spans": [ + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": "ones, i.e., we construct the auxiliary training set " + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "inline_equation", + "content": "\\mathbf{P} = \\{(\\mathbf{p}_1, 1), (\\mathbf{p}_2, 2), \\dots, (\\mathbf{p}_{|\\mathcal{Y}_b|}, |\\mathcal{Y}_b|)\\}" + }, + { + "bbox": [ + 55, + 72, + 295, + 97 + ], + "type": "text", + "content": ", and optimize:" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 91, + 105, + 295, + 121 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 91, + 105, + 295, + 121 + ], + "spans": [ + { + "bbox": [ + 91, + 105, + 295, + 121 + ], + "type": "interline_equation", + "content": "\\min _ {\\left\\{u _ {t} ^ {b}, u _ {i} ^ {b} \\right\\}} \\sum_ {\\left(\\mathbf {x}, y\\right) \\in \\mathcal {D} ^ {b} \\cup \\mathbf {P}} \\mathcal {L} _ {t} + \\mathcal {L} _ {i} + \\mathcal {L} _ {c}. \\tag {12}", + "image_path": "c64f64c5275cfd79ed5e50362b53e67d82c16d441981a8b0fa1d22607896f388.jpg" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "spans": [ + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": "Specifically, the prototype set " + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{P}" + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": " is only adopted in calculating " + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathcal{L}_c" + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": ". To enhance diversity across training stages, we also add Gaussian noise on the prototype embeddings, i.e., " + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{p}' = \\mathbf{p} + \\epsilon" + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\epsilon \\in \\mathcal{N}(0,\\alpha^2\\mathbf{I})" + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": ", and utilize " + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "inline_equation", + "content": "\\mathbf{p}'" + }, + { + "bbox": [ + 55, + 126, + 296, + 222 + ], + "type": "text", + "content": " in Eq. 12. In this way, we can revisit previous knowledge to prevent forgetting when learning new injection units. Additionally, we also consider matching the visual features to these prototypes to further assist recognition [66, 86]." + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "spans": [ + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "text", + "content": "Summary of on-the-fly injection: We visualize the training stage knowledge injection in Figure 2 (left). ENGINE learns two sets of knowledge injection units, aiming to encode the extra knowledge into the unit module. We utilize GPT-4 to extract informative class descriptors, and utilize similarity loss to encode them into the injection units. Similarly, data augmentation is utilized to provide visual information. Hence, the model can extract more informative features as it sees more tasks. To prevent forgetting, we also design the learning objective considering class prototypes. Since injection units are implemented with a single layer, the appended parameter size " + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "inline_equation", + "content": "(d\\times d)" + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "text", + "content": " per unit is negligible for CLIP. Besides, since injection units are linear layers, they can be further re-parameterized by aggregating weights into a single one, i.e., " + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\sum_{p}u_{i}^{p}" + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "text", + "content": " and " + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "inline_equation", + "content": "\\sum_{p}u_{t}^{p}" + }, + { + "bbox": [ + 55, + 223, + 296, + 405 + ], + "type": "text", + "content": "." + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 409, + 231, + 422 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 409, + 231, + 422 + ], + "spans": [ + { + "bbox": [ + 55, + 409, + 231, + 422 + ], + "type": "text", + "content": "4.2. Post-tuning Knowledge Injection" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "spans": [ + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": "We have introduced the way to inject external knowledge during training in Section 4.1. Apart from the model training stage, we can also utilize external knowledge to calibrate the prediction results during inference, and we call this step \"post-tuning knowledge injection\". For example, after the model makes inference by matching visual and textual features, we can treat the current prediction results as a preliminary result, which needs further rectification. For the top-" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " prediction results in this preliminary result, we treat them as competitors to the ground truth that could result in confusion. Hence, we need to construct a further prediction considering the image and all local competitors, which can focus more on the top-" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 426, + 295, + 582 + ], + "type": "text", + "content": " classes to refine the prediction." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": "Refining the results: After learning the injection units, we can get the prediction results using the injected features for more comprehensive features. However, the model may still make wrong predictions, e.g., classifying a \"cat\" into a similar class \"lion\" since they have similar visual features. To refine the prediction results, we design an extra consolidation process using external knowledge. Denoting the prediction results of injected feature matching in Eq. 10 as " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{inj}}(\\mathbf{x})" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", we first extract the top-" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " predictions as the competitor candidates, i.e., " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\{f_{\\mathrm{inj},i_1}(\\mathbf{x}), f_{\\mathrm{inj},i_2}(\\mathbf{x}), \\dots, f_{\\mathrm{inj},i_k}(\\mathbf{x})\\} = \\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x}))" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "inline_equation", + "content": "\\mathrm{Top}_k(f_{\\mathrm{inj}}(\\mathbf{x}))" + }, + { + "bbox": [ + 55, + 582, + 296, + 715 + ], + "type": "text", + "content": " indicates selecting the" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "spans": [ + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "text", + "content": "top- " + }, + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "text", + "content": " largest value in " + }, + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "inline_equation", + "content": "f_{\\mathrm{inj}}(\\mathbf{x})" + }, + { + "bbox": [ + 313, + 72, + 546, + 85 + ], + "type": "text", + "content": ". Their labels are denoted as:" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 321, + 89, + 553, + 114 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 321, + 89, + 553, + 114 + ], + "spans": [ + { + "bbox": [ + 321, + 89, + 553, + 114 + ], + "type": "interline_equation", + "content": "\\left\\{y _ {i _ {1}}, y _ {i _ {2}}, \\dots , y _ {i _ {k}} \\right\\} = \\left\\{y _ {i} \\mid f _ {\\text {i n j}, i} (\\mathbf {x}) \\in \\operatorname {T o p} _ {k} \\left(f _ {\\text {i n j}} (\\mathbf {x})\\right) \\right\\}. \\tag {13}", + "image_path": "84061915e95195572d35848fad9b2d9c4111085c734afcdf1b0a4c93f04bac63.jpg" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "text", + "content": "In Eq. 13, the label set " + }, + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "inline_equation", + "content": "\\{y_{i_1}, y_{i_2}, \\ldots, y_{i_k}\\}" + }, + { + "bbox": [ + 313, + 114, + 553, + 174 + ], + "type": "text", + "content": " corresponds to the possible confusing labels for the input image. To further refine the predictions, we aim for pair-wise disambiguation among this label set. We again utilize GPT-4 to generate pair-wise discriminative features:" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "spans": [ + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "text", + "content": "Q: What are unique visual features of " + }, + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLASS}]_i" + }, + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "text", + "content": " compared to " + }, + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "inline_equation", + "content": "[\\mathrm{CLASS}]_j" + }, + { + "bbox": [ + 333, + 178, + 534, + 214 + ], + "type": "text", + "content": " in a photo? Focus on their key visual differences." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 333, + 215, + 534, + 250 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 333, + 215, + 534, + 250 + ], + "spans": [ + { + "bbox": [ + 333, + 215, + 534, + 250 + ], + "type": "text", + "content": "A: [CLASS]i: 1. Long, thin tail that aids in balance... [CLASS]j: 1. Stockier, heavier tail with a tuft of hair at the end ..." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "spans": [ + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": "With this process, GPT-4 provides the most discriminative features that can differentiate two classes, e.g., the differences in \"tail\" to differentiate \"cat\" from \"lion\" and vice versa. Given the top-" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": " classes, we can generate their pairwise descriptors as " + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{D} = [\\mathbf{d}_{ij}]_{i,j=1}^{k}" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": ", where " + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "inline_equation", + "content": "\\mathbf{d}_{ij}" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": " denotes the descriptions to differ class " + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "inline_equation", + "content": "i" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": " from class " + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "inline_equation", + "content": "j" + }, + { + "bbox": [ + 313, + 255, + 555, + 353 + ], + "type": "text", + "content": ". We feed these fine-grained descriptions as the textual input, and match the query embedding to them:" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 362, + 357, + 553, + 392 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 362, + 357, + 553, + 392 + ], + "spans": [ + { + "bbox": [ + 362, + 357, + 553, + 392 + ], + "type": "interline_equation", + "content": "f _ {\\mathrm {p t}, y _ {i}} (\\mathbf {x}) = \\frac {1}{k - 1} \\sum_ {j = 1} ^ {j \\neq i} f _ {y _ {i}} \\left(\\mathbf {x}, \\mathbf {d} _ {i j}\\right), \\tag {14}", + "image_path": "f993970c66a8aba747e054873aa2c933e0691f84fb312cd8a79bb86d54bab656.jpg" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 397, + 553, + 432 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 397, + 553, + 432 + ], + "spans": [ + { + "bbox": [ + 313, + 397, + 553, + 432 + ], + "type": "text", + "content": "where we utilize zero-shot CLIP in Eq. 2 to match the query image to these pair-wise discriminative features, and average the similarities as the post-tuning logits." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "spans": [ + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "content": "Effect of post-tuning knowledge injection: We visualize the post-tuning knowledge injection in Figure 2 (right). To discriminate similar concepts and refine the wrong predictions, we first extract the top-" + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "content": " predictions of the model, and utilize GPT-4 to generate pair-wise discriminative features locally. Eq. 14 provides the fine-grained local prediction among these top-" + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 313, + 433, + 554, + 540 + ], + "type": "text", + "content": " classes, which can further help refine the results as data evolves. We adopt the zero-shot CLIP for Eq. 14, which does not require incremental updates." + } + ] + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 546, + 432, + 558 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 546, + 432, + 558 + ], + "spans": [ + { + "bbox": [ + 313, + 546, + 432, + 558 + ], + "type": "text", + "content": "4.3. Summary of ENGINE" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "spans": [ + { + "bbox": [ + 313, + 563, + 555, + 696 + ], + "type": "text", + "content": "In ENGINE, we utilize external knowledge to help continual learning in two aspects, i.e., on-the-fly injection and post-tuning injection. With the help of GPT-4, we can extract more informative descriptors suitable for the incremental learning classes, and inject the external knowledge into the knowledge units. Besides, we also extract local descriptors to discriminate pair-wise classes in the post-tuning stage to help refine the predictions. During training, we update the model via Eq. 12 to inject knowledge into the knowledge units. During inference, we aggregate the following outputs as the prediction:" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 382, + 701, + 553, + 715 + ], + "type": "interline_equation", + "angle": 0, + "lines": [ + { + "bbox": [ + 382, + 701, + 553, + 715 + ], + "spans": [ + { + "bbox": [ + 382, + 701, + 553, + 715 + ], + "type": "interline_equation", + "content": "f (\\mathbf {x}) = f _ {\\text {i n j}} (\\mathbf {x}) + f _ {\\mathrm {p t}} (\\mathbf {x}). \\tag {15}", + "image_path": "5fc88fd76722d2a3c633d4a64be0a39cbdc6f7da9b17082cdd9b802217107659.jpg" + } + ] + } + ], + "index": 18 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3318" + } + ] + } + ], + "index": 19 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 4 + }, + { + "para_blocks": [ + { + "type": "table", + "bbox": [ + 78, + 74, + 527, + 475 + ], + "blocks": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "lines": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "spans": [ + { + "bbox": [ + 55, + 50, + 555, + 73 + ], + "type": "text", + "content": "Table 1. Average and last performance comparison of different methods. The best performance is shown in bold. All methods are initialized with the same pre-trained CLIP without exemplars for a fair comparison." + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 78, + 74, + 527, + 475 + ], + "lines": [ + { + "bbox": [ + 78, + 74, + 527, + 475 + ], + "spans": [ + { + "bbox": [ + 78, + 74, + 527, + 475 + ], + "type": "table", + "html": "
MethodAircraftCIFAR100Cars
B0 Inc10B50 Inc10B0 Inc10B50 Inc10B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)
Finetune3.160.961.721.057.844.445.302.463.141.101.541.13
CoOp [88]14.547.1413.057.7747.0024.2441.2324.1236.4621.6537.4020.87
SimpleCIL [84]59.2448.0953.0548.0984.1576.6380.2076.6392.0486.8588.9686.85
ZS-CLIP [49]26.6617.2221.7017.2281.8171.3876.4971.3882.6076.3778.3276.37
L2P [65]47.1928.2944.0732.1382.7473.0381.1473.6176.6361.8276.3765.64
DualPrompt [64]44.3025.8346.0733.5781.6372.4480.1272.5776.2662.9476.8867.55
CODA-Prompt [53]45.9827.6945.1432.2882.4373.4378.6971.5880.2166.4775.0664.19
RAPF [25]50.3823.6140.4725.4486.1478.0482.1777.9382.8962.8575.8763.19
ENGINE69.6958.6964.3859.0286.9279.2283.1579.4794.1490.0891.6190.03
ImageNet-RCUBUCF
MethodB0 Inc20B100 Inc20B0 Inc20B100 Inc20B0 Inc10B50 Inc10
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(A_B\\)
Finetune1.370.431.010.882.060.640.560.474.511.591.210.80
CoOp [88]60.7337.5254.2039.7727.618.5724.0310.1447.8533.4642.0224.74
SimpleCIL [84]81.0674.4876.8474.4883.8177.5279.7577.5290.4485.6888.1285.68
ZS-CLIP [49]83.3777.1779.5777.1774.3863.0667.9663.0675.5067.6471.4467.64
L2P [65]75.9766.5272.8266.7770.8757.9375.6466.1286.3476.4383.9576.62
DualPrompt [64]76.2166.6573.2267.5869.8957.4674.4064.8485.2175.8284.3176.35
CODA-Prompt [53]77.6968.9573.7168.0573.1262.9873.9562.2187.7680.1483.0475.03
RAPF [25]81.2670.4876.1070.2379.0962.7772.8262.9392.2880.3390.3181.55
ENGINE86.2280.3783.6380.9886.6580.2082.5979.3094.3590.0392.5189.58
SUNFoodObjectNet
MethodB0 Inc30B150 Inc30B0 Inc10B50 Inc10B0 Inc20B100 Inc20
\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{A}\\)\\(\\mathcal{A}_B\\)\\(\\bar{\\mathcal{A}}\\)\\(\\mathcal{A}_B\\)
Finetune4.511.590.780.723.491.712.141.521.340.470.690.54
CoOp [88]45.9323.1139.3324.8936.0114.1833.1318.6721.246.2916.216.82
SimpleCIL [84]82.1375.5878.6275.5887.8981.6584.7381.6552.0640.1345.1140.13
ZS-CLIP [49]79.4272.1174.9572.1187.8681.9284.7581.9238.4326.4331.1226.43
L2P [65]82.8274.5479.5773.1085.6677.3380.4273.1351.4039.3948.9142.83
DualPrompt [64]82.4674.4079.3773.0284.9277.2980.0072.7552.6240.7249.0842.92
CODA-Prompt [53]83.3475.7180.3874.1786.1878.7880.9874.1346.4934.1340.5734.13
RAPF [25]82.1372.4778.0473.1088.5781.1585.5381.1748.6727.4339.2828.73
ENGINE85.0478.5481.5778.4589.8183.8986.8983.9459.1145.1951.3244.99
", + "image_path": "57a5e018adaf9e4df21f499ae70dd72b6b99039adf90e25ed74554d6a9f8aaef.jpg" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "table_body" + } + ], + "index": 1 + }, + { + "type": "table", + "bbox": [ + 56, + 509, + 290, + 582 + ], + "blocks": [ + { + "bbox": [ + 56, + 486, + 294, + 508 + ], + "lines": [ + { + "bbox": [ + 56, + 486, + 294, + 508 + ], + "spans": [ + { + "bbox": [ + 56, + 486, + 294, + 508 + ], + "type": "text", + "content": "Table 2. Comparison to traditional exemplar-based CIL methods. ENGINE does not use any exemplars." + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 509, + 290, + 582 + ], + "lines": [ + { + "bbox": [ + 56, + 509, + 290, + 582 + ], + "spans": [ + { + "bbox": [ + 56, + 509, + 290, + 582 + ], + "type": "table", + "html": "
MethodExemplarsAircraft B0 Inc10SUN B0 Inc30
AABAAB
iCaRL [51]20 / class53.6043.9878.5667.30
MEMO [83]20 / class42.2425.4181.4873.45
PROOF [86]20 / class61.0053.5983.5777.28
ENGINE069.6958.6985.0478.54
", + "image_path": "bb7df7ed0b60ebf1ca741f73e2c1f217a5d9ea06c838ff37f070c1c2a7dfb23a.jpg" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "table_body" + } + ], + "index": 3 + }, + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "spans": [ + { + "bbox": [ + 55, + 596, + 137, + 609 + ], + "type": "text", + "content": "5. Experiments" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "spans": [ + { + "bbox": [ + 55, + 618, + 295, + 715 + ], + "type": "text", + "content": "In this section, we conduct extensive experiments on nine benchmark datasets and compare ENGINE to state-of-the-art methods. We visualize the incremental learning curve and provide ablation studies on prompt engineering and parameter sensitivity to test the model's robustness. Besides, we also visualize the learned embeddings and logit changes after injecting external knowledge to show its effectiveness. We provide more results in the supplementary." + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 313, + 495, + 447, + 508 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 495, + 447, + 508 + ], + "spans": [ + { + "bbox": [ + 313, + 495, + 447, + 508 + ], + "type": "text", + "content": "5.1. Implementation Details" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 313, + 517, + 555, + 638 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 638 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 638 + ], + "type": "text", + "content": "Dataset: We follow [65, 86, 88] to evaluate the performance on nine benchmark datasets that have domain gap to CLIP's pre-training dataset, i.e., CIFAR100 [32], CUB200 [59], ObjectNet [6], ImageNet-R [22], FGVCAircraft [41], StanfordCars [31], Food101 [7], SUN397 [68] and UCF101 [55]. Following [86], we utilize the sampled 100 classes from CIFAR100, Aircraft, Cars, Food, UCF, 200 classes from CUB200, ObjectNet, ImageNet-R, and 300 classes from SUN to ease the data split. More details are reported in the supplementary." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "Dataset split: Following [51, 65], we use 'B-" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " Inc-" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": "' to split the classes in CIL. " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "m" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " indicates the number of classes in the first stage, and " + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "inline_equation", + "content": "n" + }, + { + "bbox": [ + 313, + 641, + 556, + 714 + ], + "type": "text", + "content": " represents that of every following stage. We follow [51] to randomly shuffle the class order with random seed 1993 for all compared methods, and keep this same for every method." + } + ] + } + ], + "index": 8 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3319" + } + ] + } + ], + "index": 9 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 5 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 58, + 52, + 220, + 177 + ], + "blocks": [ + { + "bbox": [ + 58, + 52, + 220, + 177 + ], + "lines": [ + { + "bbox": [ + 58, + 52, + 220, + 177 + ], + "spans": [ + { + "bbox": [ + 58, + 52, + 220, + 177 + ], + "type": "image", + "image_path": "224d58b8d4e13ed74089cdd54c4c30a3bb1888d6cb98c92389b491af76175e92.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 99, + 181, + 179, + 190 + ], + "lines": [ + { + "bbox": [ + 99, + 181, + 179, + 190 + ], + "spans": [ + { + "bbox": [ + 99, + 181, + 179, + 190 + ], + "type": "text", + "content": "(a) Aircraft Base0 Inc10" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 224, + 52, + 388, + 176 + ], + "blocks": [ + { + "bbox": [ + 224, + 52, + 388, + 176 + ], + "lines": [ + { + "bbox": [ + 224, + 52, + 388, + 176 + ], + "spans": [ + { + "bbox": [ + 224, + 52, + 388, + 176 + ], + "type": "image", + "image_path": "facb6e73795a87063d68e79b97490b31620b7ff86f510bfc888d025613de2a1a.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 270, + 181, + 340, + 190 + ], + "lines": [ + { + "bbox": [ + 270, + 181, + 340, + 190 + ], + "spans": [ + { + "bbox": [ + 270, + 181, + 340, + 190 + ], + "type": "text", + "content": "(b) Cars Base0 Inc10" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "type": "image", + "bbox": [ + 391, + 53, + 552, + 176 + ], + "blocks": [ + { + "bbox": [ + 391, + 53, + 552, + 176 + ], + "lines": [ + { + "bbox": [ + 391, + 53, + 552, + 176 + ], + "spans": [ + { + "bbox": [ + 391, + 53, + 552, + 176 + ], + "type": "image", + "image_path": "42cd25a8b36a24d9279ae8e65edd0c7c3670e31cd0f8f2a681017f3cd0d1e4ce.jpg" + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 436, + 181, + 506, + 190 + ], + "lines": [ + { + "bbox": [ + 436, + 181, + 506, + 190 + ], + "spans": [ + { + "bbox": [ + 436, + 181, + 506, + 190 + ], + "type": "text", + "content": "(c) UCF Base0 Inc10" + } + ] + } + ], + "index": 5, + "angle": 0, + "type": "image_caption" + } + ], + "index": 4 + }, + { + "type": "table", + "bbox": [ + 56, + 240, + 291, + 297 + ], + "blocks": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "lines": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "spans": [ + { + "bbox": [ + 55, + 217, + 295, + 239 + ], + "type": "text", + "content": "Table 3. Results when all methods use the same textual description generated by GPT-4 and data augmentation." + } + ] + } + ], + "index": 7, + "angle": 0, + "type": "table_caption" + }, + { + "bbox": [ + 56, + 240, + 291, + 297 + ], + "lines": [ + { + "bbox": [ + 56, + 240, + 291, + 297 + ], + "spans": [ + { + "bbox": [ + 56, + 240, + 291, + 297 + ], + "type": "table", + "html": "
MethodTemplateImageNet-R AB0 Inc20 A_BCIFAR AB0 Inc10 A_B
ZS-CLIPGPT Generated83.6877.5982.2671.70
RAPFGPT Generated81.8971.0486.4978.52
ENGINEGPT Generated86.2280.3786.9279.22
", + "image_path": "1de9d8b665a18aa85c4883a0af9618edf24763af17eedfb38e5abafe85e6036a.jpg" + } + ] + } + ], + "index": 8, + "angle": 0, + "type": "table_body" + } + ], + "index": 8 + }, + { + "bbox": [ + 54, + 304, + 295, + 398 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 54, + 304, + 295, + 398 + ], + "spans": [ + { + "bbox": [ + 54, + 304, + 295, + 398 + ], + "type": "text", + "content": "Comparison methods: We first compare to SOTA pretrained model-based CIL algorithms, e.g., L2P [65], DualPrompt [64], CODA-Prompt [53], SimpleCIL [84]. Besides, we also compare to SOTA CLIP-based CIL algorithms, e.g., CoOp [88], PROOF [86], RAPF [25]. The baseline that finetunes CLIP for incremental tasks is denoted as Finetune. All methods are deployed with the same CLIP as initialization." + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "spans": [ + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": "Training details: The experiments are deployed with NVIDIA 4090 using PyTorch [47]. We follow [25, 86] to consider CLIP with ViT-B/16 for all compared methods for fair comparison. For vision-based methods that cannot utilize the textual prompt (e.g., L2P, DualPrompt, CODA-Prompt), we utilize CLIP's visual branch as their initialization. We report the results using LAION-400M pre-trained CLIP [26] in the main paper, and report results using OpenAI [49] in the supplementary. In ENGINE, we use SGD optimizer with a batch size of 64 to optimize the model for 10 epochs. The learning rate decays from 0.05 with cosine annealing. We set the prototype noise ratio " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": " to 0.25 and " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": " in post-tuning to 5. We use OpenAI GPT-4o mini [1] to generate the prompts for textual knowledge and AutoAugment [11] to augment the visual features. We explore other LLMs to extract external knowledge in the supplementary. The source code will be publicly available upon acceptance. Evaluation metric: Following [51, 86], we use " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_b" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": " to represent the model's accuracy after the " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "b" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": "-th stage. Specifically, we adopt " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\mathcal{A}_B" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": " (the performance after the last stage) and " + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "inline_equation", + "content": "\\bar{\\mathcal{A}} = \\frac{1}{B}\\sum_{b=1}^{B}\\mathcal{A}_b" + }, + { + "bbox": [ + 55, + 399, + 296, + 664 + ], + "type": "text", + "content": " (average performance along incremental stages) as measurements." + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 55, + 672, + 194, + 685 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 672, + 194, + 685 + ], + "spans": [ + { + "bbox": [ + 55, + 672, + 194, + 685 + ], + "type": "text", + "content": "5.2. Benchmark Comparison" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "spans": [ + { + "bbox": [ + 55, + 689, + 295, + 713 + ], + "type": "text", + "content": "We first compare ENGINE to other state-of-the-art methods on the benchmark datasets and report the results in Table 1" + } + ] + } + ], + "index": 12 + }, + { + "type": "image", + "bbox": [ + 317, + 218, + 432, + 309 + ], + "blocks": [ + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "lines": [ + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "spans": [ + { + "bbox": [ + 55, + 193, + 555, + 214 + ], + "type": "text", + "content": "Figure 3. Incremental performance of different methods. We report the performance gap after the last incremental stage of ENGINE and the runner-up method at the end of the line. All methods utilize the same CLIP pre-trained weight. More figures are shown in supplementary." + } + ] + } + ], + "index": 6, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 317, + 218, + 432, + 309 + ], + "lines": [ + { + "bbox": [ + 317, + 218, + 432, + 309 + ], + "spans": [ + { + "bbox": [ + 317, + 218, + 432, + 309 + ], + "type": "image", + "image_path": "ec1ee823ba1959f74917c9e66162db42f244bde62fd25d9b84e8cc63a2e008f6.jpg" + } + ] + } + ], + "index": 13, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 344, + 312, + 404, + 323 + ], + "lines": [ + { + "bbox": [ + 344, + 312, + 404, + 323 + ], + "spans": [ + { + "bbox": [ + 344, + 312, + 404, + 323 + ], + "type": "text", + "content": "(a) Ablation study" + } + ] + } + ], + "index": 14, + "angle": 0, + "type": "image_caption" + } + ], + "index": 13 + }, + { + "type": "image", + "bbox": [ + 434, + 218, + 550, + 310 + ], + "blocks": [ + { + "bbox": [ + 434, + 218, + 550, + 310 + ], + "lines": [ + { + "bbox": [ + 434, + 218, + 550, + 310 + ], + "spans": [ + { + "bbox": [ + 434, + 218, + 550, + 310 + ], + "type": "image", + "image_path": "a3393fed7b92a9574c5b020a2468f0c4ee24cf6e8d599c845be95894a3165b7f.jpg" + } + ] + } + ], + "index": 15, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 453, + 312, + 533, + 323 + ], + "lines": [ + { + "bbox": [ + 453, + 312, + 533, + 323 + ], + "spans": [ + { + "bbox": [ + 453, + 312, + 533, + 323 + ], + "type": "text", + "content": "(b) Parameter sensitivity" + } + ] + } + ], + "index": 16, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 341, + 324, + 527, + 335 + ], + "lines": [ + { + "bbox": [ + 341, + 324, + 527, + 335 + ], + "spans": [ + { + "bbox": [ + 341, + 324, + 527, + 335 + ], + "type": "text", + "content": "Figure 4. Ablation study and parameter sensitivity." + } + ] + } + ], + "index": 17, + "angle": 0, + "type": "image_caption" + } + ], + "index": 15 + }, + { + "bbox": [ + 313, + 355, + 555, + 513 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 355, + 555, + 513 + ], + "spans": [ + { + "bbox": [ + 313, + 355, + 555, + 513 + ], + "type": "text", + "content": "and Figure 3. Specifically, we find ENGINE outperforms current SOTA by " + }, + { + "bbox": [ + 313, + 355, + 555, + 513 + ], + "type": "inline_equation", + "content": "3\\% \\sim 10\\%" + }, + { + "bbox": [ + 313, + 355, + 555, + 513 + ], + "type": "text", + "content": " on these benchmark datasets. We find the baseline method Finetune has the worst performance, indicating that the decomposition ability to tackle the class template into fine-grained features is totally forgotten. For visual prompt-based methods like L2P, DualPrompt, and CODA-Prompt, the performance is limited due to the inability to utilize textual information. For the textual prompt tuning method CoOp, we find that it achieves poor performance due to the forgetting of textual prompts. For other CLIP-based methods like RAPF, our method still shows substantial improvement, verifying its strong continual learning ability without forgetting." + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "spans": [ + { + "bbox": [ + 313, + 517, + 555, + 589 + ], + "type": "text", + "content": "Apart from these non-exemplar-based methods, we also consider typical exemplar-based CIL methods for comparison. We report the comparison results in Table 2, where all methods are initialized with the same CLIP. As we can infer from the table, ENGINE still outperforms them even without using exemplars." + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "spans": [ + { + "bbox": [ + 313, + 594, + 556, + 714 + ], + "type": "text", + "content": "Finally, what if all CLIP-based methods utilize the same external knowledge? Correspondingly, we also consider using the same generated prompt and data augmentation for ZS-CLIP and RAPF, and report the results in Table 3. In this table, we replace the template text of all methods using the same text in Section 4.1 and adopt the same data augmentation for image input. We find ENGINE still outperforms them even using the same prompt, indicating that ENGINE can better utilize the external information for incremental learning." + } + ] + } + ], + "index": 20 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3320" + } + ] + } + ], + "index": 21 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 6 + }, + { + "para_blocks": [ + { + "type": "image", + "bbox": [ + 59, + 57, + 173, + 144 + ], + "blocks": [ + { + "bbox": [ + 59, + 57, + 173, + 144 + ], + "lines": [ + { + "bbox": [ + 59, + 57, + 173, + 144 + ], + "spans": [ + { + "bbox": [ + 59, + 57, + 173, + 144 + ], + "type": "image", + "image_path": "de85726447bf967dad8ed2129b75a3a536f4fd270e8d47c2bc670b116618223b.jpg" + } + ] + } + ], + "index": 0, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 74, + 147, + 157, + 156 + ], + "lines": [ + { + "bbox": [ + 74, + 147, + 157, + 156 + ], + "spans": [ + { + "bbox": [ + 74, + 147, + 157, + 156 + ], + "type": "text", + "content": "(a) ZS-CLIP, single stage" + } + ] + } + ], + "index": 1, + "angle": 0, + "type": "image_caption" + } + ], + "index": 0 + }, + { + "type": "image", + "bbox": [ + 176, + 57, + 291, + 144 + ], + "blocks": [ + { + "bbox": [ + 176, + 57, + 291, + 144 + ], + "lines": [ + { + "bbox": [ + 176, + 57, + 291, + 144 + ], + "spans": [ + { + "bbox": [ + 176, + 57, + 291, + 144 + ], + "type": "image", + "image_path": "ab615e14df2c4a9549f04dd7caf47bb2ce428e22913d931db6fec55654ca7a54.jpg" + } + ] + } + ], + "index": 2, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 195, + 147, + 272, + 156 + ], + "lines": [ + { + "bbox": [ + 195, + 147, + 272, + 156 + ], + "spans": [ + { + "bbox": [ + 195, + 147, + 272, + 156 + ], + "type": "text", + "content": "(b) ENGINE, two stages" + } + ] + } + ], + "index": 3, + "angle": 0, + "type": "image_caption" + }, + { + "bbox": [ + 55, + 158, + 295, + 224 + ], + "lines": [ + { + "bbox": [ + 55, + 158, + 295, + 224 + ], + "spans": [ + { + "bbox": [ + 55, + 158, + 295, + 224 + ], + "type": "text", + "content": "Figure 5. t-SNE [58] visualizations of zero-shot CLIP and ENGINE. ENGINE aligns the cross-modal features without forgetting. We represent the visual features of the first task with dots and classes of the second task with squares. The textual features are represented by triangles, and we utilize the shadow area to denote the decision boundary." + } + ] + } + ], + "index": 4, + "angle": 0, + "type": "image_caption" + } + ], + "index": 2 + }, + { + "bbox": [ + 55, + 236, + 157, + 249 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 236, + 157, + 249 + ], + "spans": [ + { + "bbox": [ + 55, + 236, + 157, + 249 + ], + "type": "text", + "content": "5.3. Further Analysis" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "spans": [ + { + "bbox": [ + 55, + 255, + 296, + 495 + ], + "type": "text", + "content": "Ablation study: We first conduct ablations to investigate the effectiveness of each component in ENGINE. We report the incremental performance on CIFAR100 B0 Inc10 in Figure 4a. We can infer that 'ZS-CLIP' has the worst performance since the downstream distribution is different from the pre-training stage, and we view it as the baseline. Correspondingly, we first equip the model with visual knowledge injection (Eq. 6), and denote the model as 'w/ Visual Injection'. It shows that the performance drastically improves as the model has been adapted to extract more informative visual features. Similarly, we then append the textual information using Eq. 5, denoted as 'w/ Dual Injection.' As shown in the figure, using external knowledge for both modalities can further improve the performance than visual knowledge only. Finally, we equip the model with post-tuning knowledge injection, and utilize Eq. 15 for inference (i.e., 'w/ Post-tuning Injection'). This corresponds to the full version of ENGINE, and we find it achieves the best performance. Ablations verify that every component in ENGINE boosts the CIL performance." + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "spans": [ + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": "Parameter robustness: There are two hyperparameters in ENGINE, i.e., the prototype noise scale " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": " in Section 4.1 and post-tuning candidate " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": " in Eq. 14. We conduct experiments on CUB B0 Inc20 to investigate the robustness by changing these parameters. Specifically, we choose " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\alpha" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": " among " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{0.15, 0.2, 0.25, 0.3, 0.35\\}" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": ", and " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "k" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": " among " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\{3, 4, 5, 6, 7\\}" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": ". We report the last performance in Figure 4b. As shown in the figure, the performance is robust with the change of these parameters, and we suggest " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "\\alpha = 0.25" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": ", " + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "inline_equation", + "content": "k = 5" + }, + { + "bbox": [ + 55, + 497, + 296, + 605 + ], + "type": "text", + "content": " as default." + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": "Visualizations: In this section, we utilize t-SNE [58] to visualize the cross-modal features learned by ENGINE on CIFAR100 B0 Inc5. We represent the visual features of the first task with dots " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\bigcirc)" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": " and classes of the second task with squares " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\square)" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ". The textual features are represented by triangles " + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "inline_equation", + "content": "(\\triangle)" + }, + { + "bbox": [ + 55, + 605, + 296, + 714 + ], + "type": "text", + "content": ", and we utilize the shadow area to denote the decision boundary divided by the textual features. Specifically, we first visualize the instances of the first stage in zero-shot CLIP's embedding space in Figure 5a. The results indicate" + } + ] + } + ], + "index": 8 + }, + { + "type": "image", + "bbox": [ + 317, + 57, + 553, + 217 + ], + "blocks": [ + { + "bbox": [ + 317, + 57, + 553, + 217 + ], + "lines": [ + { + "bbox": [ + 317, + 57, + 553, + 217 + ], + "spans": [ + { + "bbox": [ + 317, + 57, + 553, + 217 + ], + "type": "image", + "image_path": "6f24666e94e7d1c0d2e221513cfe0f00eee579da00e50e37e3cfe638fc1ccc28.jpg" + } + ] + } + ], + "index": 9, + "angle": 0, + "type": "image_body" + }, + { + "bbox": [ + 313, + 218, + 553, + 251 + ], + "lines": [ + { + "bbox": [ + 313, + 218, + 553, + 251 + ], + "spans": [ + { + "bbox": [ + 313, + 218, + 553, + 251 + ], + "type": "text", + "content": "Figure 6. Left: Input. Middle: Top-5 predictions before post-tuning knowledge injection. Right: Top-5 predictions after post-tuning. More cases are shown in the supplementary." + } + ] + } + ], + "index": 10, + "angle": 0, + "type": "image_caption" + } + ], + "index": 9 + }, + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "spans": [ + { + "bbox": [ + 313, + 270, + 555, + 366 + ], + "type": "text", + "content": "that pre-trained CLIP has difficulty in aligning the visual and textual features since they are situated in two distinct areas in the embedding space [37]. While for Figure 5b, we visualize the embedding of ENGINE across two incremental stages. As we can infer from the figure, ENGINE has two advantages: 1) The visual and textual features are aligned in the same cluster, and 2) The model has not forgotten previous classes when learning new ones." + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 313, + 369, + 556, + 466 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 369, + 556, + 466 + ], + "spans": [ + { + "bbox": [ + 313, + 369, + 556, + 466 + ], + "type": "text", + "content": "We also visualize the model output before and after post-tuning knowledge injection on ImageNet-R in Figure 6. We utilize bars with red edges to denote the ground-truth class. As we can infer from the figure, post-tuning knowledge injection can help the model refine the predictions by generating local comparisons. For the case on top of Figure 6, the model can focus on the differences between \"acorn\" and \"birdhouse\" to make the final prediction." + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 314, + 487, + 388, + 498 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 314, + 487, + 388, + 498 + ], + "spans": [ + { + "bbox": [ + 314, + 487, + 388, + 498 + ], + "type": "text", + "content": "6. Conclusion" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "spans": [ + { + "bbox": [ + 313, + 510, + 555, + 653 + ], + "type": "text", + "content": "Class-incremental learning is essential to real-world applications. This paper aims to enable CLIP with CIL ability by injecting external knowledge. Specifically, we consider on-the-fly knowledge injection by learning dual-branch injection units. We augment images for the visual branch and generate detailed feature descriptions with GPT-4 for the textual branch. By maximizing the similarity between external information and the injection module, we encode informative features into the pre-trained CLIP. Besides, we also consider post-tuning knowledge injection with local class information to calibrate the predictions. Extensive experiments verify ENGINE's effectiveness." + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "spans": [ + { + "bbox": [ + 313, + 654, + 556, + 715 + ], + "type": "text", + "content": "Limitations and future works: This paper relies on GPT-4 to provide expert knowledge for specific tasks, which may be impossible for the specific cases where GPT-4 cannot generalize. Future works include designing other formats of external knowledge, e.g., knowledge graphs." + } + ] + } + ], + "index": 15 + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 314, + 757 + ], + "type": "text", + "content": "3321" + } + ] + } + ], + "index": 16 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 7 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 153, + 85 + ], + "type": "text", + "content": "Acknowledgments" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "angle": 0, + "lines": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "spans": [ + { + "bbox": [ + 55, + 91, + 296, + 187 + ], + "type": "text", + "content": "This work is partially supported by NSFC (U23A20382, 62476123, 62376118, 62250069), Fundamental Research Funds for the Central Universities (2024300373,14380021), CCF-Tencent Rhino-Bird Open Research Fund (RAGR20240101), Postdoctoral Fellowship Program of CPSF (GZB20250389), China Postdoctoral Science Foundation (2025M771535), and Collaborative Innovation Center of Novel Software Technology and Industrialization." + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "title", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "spans": [ + { + "bbox": [ + 56, + 198, + 115, + 211 + ], + "type": "text", + "content": "References" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 57, + 218, + 296, + 714 + ], + "type": "list", + "angle": 0, + "index": 15, + "blocks": [ + { + "bbox": [ + 61, + 218, + 296, + 273 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 218, + 296, + 273 + ], + "spans": [ + { + "bbox": [ + 61, + 218, + 296, + 273 + ], + "type": "text", + "content": "[1] Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023. 3, 7" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "spans": [ + { + "bbox": [ + 61, + 274, + 296, + 306 + ], + "type": "text", + "content": "[2] Charu C Aggarwal. A survey of stream clustering algorithms. In Data Clustering, pages 231-258. Chapman and Hall/CRC, 2018. 1" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 61, + 308, + 294, + 352 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 61, + 308, + 294, + 352 + ], + "spans": [ + { + "bbox": [ + 61, + 308, + 294, + 352 + ], + "type": "text", + "content": "[3] Rahaf Aljundi, Francesca Babiloni, Mohamed Elhoseiny, Marcus Rohrbach, and Tinne Tuytelaars. Memory aware synapses: Learning what (not) to forget. In ECCV, pages 139-154, 2018. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "spans": [ + { + "bbox": [ + 62, + 354, + 295, + 386 + ], + "type": "text", + "content": "[4] Rahaf Aljundi, Klaas Kelchtermans, and Tinne Tuytelaars. Task-free continual learning. In CVPR, pages 11254-11263, 2019. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "spans": [ + { + "bbox": [ + 62, + 388, + 296, + 421 + ], + "type": "text", + "content": "[5] Rahaf Aljundi, Min Lin, Baptiste Goujaud, and Yoshua Bengio. Gradient based sample selection for online continual learning. In NeurIPS, pages 11816-11825, 2019. 2" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "spans": [ + { + "bbox": [ + 62, + 422, + 296, + 476 + ], + "type": "text", + "content": "[6] Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models. In NeurIPS, pages 9448-9458, 2019. 6" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "spans": [ + { + "bbox": [ + 62, + 478, + 296, + 510 + ], + "type": "text", + "content": "[7] Lukas Bossard, Matthieu Guillaumin, and Luc Van Gool. Food-101-mining discriminative components with random forests. In ECCV, pages 446–461, 2014. 6" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "spans": [ + { + "bbox": [ + 62, + 512, + 296, + 555 + ], + "type": "text", + "content": "[8] Arslan Chaudhry, Puneet K Dokania, Thalaiyasingam Ajthan, and Philip HS Torr. Riemannian walk for incremental learning: Understanding forgetting and intransigence. In ECCV, pages 532-547, 2018. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "spans": [ + { + "bbox": [ + 62, + 557, + 296, + 590 + ], + "type": "text", + "content": "[9] Arslan Chaudhry, Marc'Aurelio Ranzato, Marcus Rohrbach, and Mohamed Elhoseiny. Efficient lifelong learning with a gem. In ICLR, 2018. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "spans": [ + { + "bbox": [ + 57, + 591, + 295, + 634 + ], + "type": "text", + "content": "[10] Shoufa Chen, Chongjian Ge, Zhan Tong, Jiangliu Wang, Yibing Song, Jue Wang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recognition. NeurIPS, pages 16664-16678, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "spans": [ + { + "bbox": [ + 57, + 635, + 295, + 669 + ], + "type": "text", + "content": "[11] Ekin D Cubuk, Barret Zoph, Dandelion Mane, Vijay Vasudevan, and Quoc V Le. Autoaugment: Learning augmentation strategies from data. In CVPR, pages 113-123, 2019. 7" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "spans": [ + { + "bbox": [ + 57, + 670, + 296, + 714 + ], + "type": "text", + "content": "[12] Matthias De Lange, Rahaf Aljundi, Marc Masana, Sarah Parisot, Xu Jia, Aleš Leonardis, Gregory Slabaugh, and Tinne Tuytelaars. A continual learning survey: Defying forgetting in classification tasks. IEEE transactions on pattern" + } + ] + } + ], + "index": 14 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 555, + 714 + ], + "type": "list", + "angle": 0, + "index": 31, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "analysis and machine intelligence, 44(7):3366-3385, 2021. 1, 2" + } + ] + } + ], + "index": 16 + }, + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 97, + 553, + 129 + ], + "type": "text", + "content": "[13] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009. 1" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "spans": [ + { + "bbox": [ + 316, + 130, + 553, + 174 + ], + "type": "text", + "content": "[14] Songlin Dong, Xiaopeng Hong, Xiaoyu Tao, Xinyuan Chang, Xing Wei, and Yihong Gong. Few-shot class-incremental learning via relation knowledge distillation. In AAAI, pages 1255-1263, 2021. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "spans": [ + { + "bbox": [ + 316, + 175, + 553, + 230 + ], + "type": "text", + "content": "[15] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In ICLR, 2020. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "spans": [ + { + "bbox": [ + 317, + 232, + 553, + 274 + ], + "type": "text", + "content": "[16] Arthur Douillard, Matthieu Cord, Charles Ollion, Thomas Robert, and Eduardo Valle. Podnet: Pooled outputs distillation for small-tasks incremental learning. In ECCV, pages 86-102, 2020. 2" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "spans": [ + { + "bbox": [ + 316, + 276, + 553, + 319 + ], + "type": "text", + "content": "[17] Arthur Douillard, Alexandre Ramé, Guillaume Couairon, and Matthieu Cord. Dytox: Transformers for continual learning with dynamic token expansion. In CVPR, pages 9285–9295, 2022. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "spans": [ + { + "bbox": [ + 316, + 321, + 553, + 354 + ], + "type": "text", + "content": "[18] Robert M French. Catastrophic forgetting in connectionist networks. Trends in cognitive sciences, 3(4):128-135, 1999. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 355, + 555, + 408 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 355, + 555, + 408 + ], + "spans": [ + { + "bbox": [ + 316, + 355, + 555, + 408 + ], + "type": "text", + "content": "[19] Robert M French and André Ferrara. Modeling time perception in rats: Evidence for catastrophic interference in animal learning. In Proceedings of the 21st Annual Conference of the Cognitive Science Conference, pages 173-178. CiteSeer, 1999. 1" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "spans": [ + { + "bbox": [ + 316, + 411, + 553, + 453 + ], + "type": "text", + "content": "[20] Qiankun Gao, Chen Zhao, Bernard Ghanem, and Jian Zhang. R-DFCIL: relation-guided representation learning for data-free class incremental learning. In ECCV, pages 423–439, 2022. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "spans": [ + { + "bbox": [ + 316, + 456, + 553, + 488 + ], + "type": "text", + "content": "[21] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770-778, 2015. 1" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "spans": [ + { + "bbox": [ + 317, + 490, + 553, + 544 + ], + "type": "text", + "content": "[22] Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kadavath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340-8349, 2021. 6" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 546, + 553, + 578 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 546, + 553, + 578 + ], + "spans": [ + { + "bbox": [ + 316, + 546, + 553, + 578 + ], + "type": "text", + "content": "[23] Geoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 580, + 553, + 613 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 580, + 553, + 613 + ], + "spans": [ + { + "bbox": [ + 316, + 580, + 553, + 613 + ], + "type": "text", + "content": "[24] Saihui Hou, Xinyu Pan, Chen Change Loy, Zilei Wang, and Dahua Lin. Learning a unified classifier incrementally via rebalancing. In CVPR, pages 831-839, 2019. 2" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "spans": [ + { + "bbox": [ + 317, + 614, + 553, + 657 + ], + "type": "text", + "content": "[25] Linlan Huang, Xusheng Cao, Haori Lu, and Xialei Liu. Class-incremental learning with clip: Adaptive representation adjustment and parameter fusion. In ECCV, pages 214-231, 2024. 2, 3, 6, 7" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "spans": [ + { + "bbox": [ + 316, + 658, + 553, + 714 + ], + "type": "text", + "content": "[26] Gabriel Ilharco, Mitchell Wortsman, Ross Wightman, Cade Gordon, Nicholas Carlini, Rohan Taori, Achal Dave, Vaishaal Shankar, Hongseok Namkoong, John Miller, Hannaneh Hajishirzi, Ali Farhadi, and Ludwig Schmidt. Openclip, 2021. 7" + } + ] + } + ], + "index": 30 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3322" + } + ] + } + ], + "index": 32 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 8 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "spans": [ + { + "bbox": [ + 56, + 73, + 294, + 127 + ], + "type": "text", + "content": "[27] Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904-4916, 2021. 1" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 161 + ], + "type": "text", + "content": "[28] Menglin Jia, Luming Tang, Bor-Chun Chen, Claire Cardie, Serge J. Belongie, Bharath Hariharan, and Ser-Nam Lim. Visual prompt tuning. In ECCV, pages 709-727, 2022. 2, 3" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "spans": [ + { + "bbox": [ + 56, + 162, + 294, + 195 + ], + "type": "text", + "content": "[29] Dahuin Jung, Dongyoon Han, Jihwan Bang, and Hwanjun Song. Generating instance-level prompts for rehearsal-free continual learning. In ICCV, pages 11847-11857, 2023. 2" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 195, + 294, + 259 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 195, + 294, + 259 + ], + "spans": [ + { + "bbox": [ + 56, + 195, + 294, + 259 + ], + "type": "text", + "content": "[30] James Kirkpatrick, Razvan Pascanu, Neil Rabinowitz, Joel Veness, Guillaume Desjardins, Andrei A Rusu, Kieran Milan, John Quan, Tiago Ramalho, Agnieszka Grabska-Barwinska, et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences, 114(13):3521-3526, 2017. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "spans": [ + { + "bbox": [ + 56, + 261, + 294, + 294 + ], + "type": "text", + "content": "[31] Jonathan Krause, Michael Stark, Jia Deng, and Li Fei-Fei. 3d object representations for fine-grained categorization. In ICCV Workshop, pages 554–561, 2013. 6" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "spans": [ + { + "bbox": [ + 56, + 294, + 294, + 326 + ], + "type": "text", + "content": "[32] Alex Krizhevsky, Geoffrey Hinton, et al. Learning multiple layers of features from tiny images. Technical report, 2009. 6" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "spans": [ + { + "bbox": [ + 56, + 327, + 294, + 370 + ], + "type": "text", + "content": "[33] Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, pages 19730–19742, 2023. 1" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "spans": [ + { + "bbox": [ + 56, + 371, + 294, + 404 + ], + "type": "text", + "content": "[34] Li Li, Jiawei Peng, Huiyi Chen, Chongyang Gao, and Xu Yang. How to configure good in-context sequence for visual question answering. In CVPR, pages 26710-26720, 2024. 1" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 437 + ], + "type": "text", + "content": "[35] Lan Li, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Addressing imbalanced domain-incremental learning through dual-balance collaborative experts. In ICML, 2025. 2" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "spans": [ + { + "bbox": [ + 56, + 437, + 294, + 459 + ], + "type": "text", + "content": "[36] Zhizhong Li and Derek Hoiem. Learning without forgetting. In ECCV, pages 614-629, 2016. 2" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "spans": [ + { + "bbox": [ + 56, + 460, + 294, + 503 + ], + "type": "text", + "content": "[37] Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. NeurIPS, pages 17612-17625, 2022. 8" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "spans": [ + { + "bbox": [ + 56, + 504, + 294, + 546 + ], + "type": "text", + "content": "[38] Yaoyao Liu, Yuting Su, An-An Liu, Bernt Schiele, and Qianru Sun. Mnemonics training: Multi-class incremental learning without forgetting. In CVPR, pages 12245-12254, 2020. 2" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 548, + 294, + 581 + ], + "type": "text", + "content": "[39] Yichen Lu, Mei Wang, and Weihong Deng. Augmented geometric distillation for data-free incremental person reid. In CVPR, pages 7329-7338, 2022. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "spans": [ + { + "bbox": [ + 56, + 582, + 294, + 624 + ], + "type": "text", + "content": "[40] Zilin Luo, Yaoyao Liu, Bernt Schiele, and Qianru Sun. Class-incremental exemplar compression for class-incremental learning. In CVPR, pages 11371-11380, 2023. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 625, + 294, + 667 + ], + "type": "text", + "content": "[41] Subhransu Maji, Esa Rahtu, Juho Kannala, Matthew Blaschko, and Andrea Vedaldi. Fine-grained visual classification of aircraft. arXiv preprint arXiv:1306.5151, 2013. 6" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[42] Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D Bagdanov, and Joost Van De Weijer. Class-incremental learning: survey and performance evaluation on image classification. IEEE Transactions on Pattern" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 316, + 73, + 553, + 712 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "spans": [ + { + "bbox": [ + 335, + 73, + 553, + 95 + ], + "type": "text", + "content": "Analysis and Machine Intelligence, 45(5):5513-5533, 2022. 1, 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "spans": [ + { + "bbox": [ + 316, + 96, + 553, + 129 + ], + "type": "text", + "content": "[43] Saeed Masoudnia and Reza Ebrahimpour. Mixture of experts: a literature survey. Artificial Intelligence Review, 42: 275-293, 2014. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "spans": [ + { + "bbox": [ + 316, + 131, + 553, + 175 + ], + "type": "text", + "content": "[44] Mark D McDonnell, Dong Gong, Amin Parveneh, Ehsan Abbasnejad, and Anton van den Hengel. Ranpac: Random projections and pre-trained models for continual learning. In NeurIPS, 2023. 2" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 316, + 177, + 553, + 199 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 177, + 553, + 199 + ], + "spans": [ + { + "bbox": [ + 316, + 177, + 553, + 199 + ], + "type": "text", + "content": "[45] Sachit Menon and Carl Vondrick. Visual classification via description from large language models. In ICLR, 2023. 1" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 316, + 201, + 553, + 234 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 201, + 553, + 234 + ], + "spans": [ + { + "bbox": [ + 316, + 201, + 553, + 234 + ], + "type": "text", + "content": "[46] Jaeyoo Park, Minsoo Kang, and Bohyung Han. Class incremental learning for action recognition in videos. In ICCV, pages 13698-13707, 2021. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 316, + 236, + 553, + 291 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 236, + 553, + 291 + ], + "spans": [ + { + "bbox": [ + 316, + 236, + 553, + 291 + ], + "type": "text", + "content": "[47] Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, et al. Pytorch: An imperative style, high-performance deep learning library. In NeurIPS, pages 8026-8037, 2019. 7" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 316, + 293, + 553, + 335 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 293, + 553, + 335 + ], + "spans": [ + { + "bbox": [ + 316, + 293, + 553, + 335 + ], + "type": "text", + "content": "[48] Zhi-Hong Qi, Da-Wei Zhou, Yiran Yao, Han-Jia Ye, and De-Chuan Zhan. Adaptive adapter routing for long-tailed class incremental learning. Machine Learning, 114(3):1-20, 2025. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "spans": [ + { + "bbox": [ + 316, + 338, + 553, + 393 + ], + "type": "text", + "content": "[49] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervision. In ICML, pages 8748-8763, 2021. 1, 3, 6, 7" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "spans": [ + { + "bbox": [ + 316, + 395, + 553, + 428 + ], + "type": "text", + "content": "[50] Sylvestre-Alvise Rebuffi, Hakan Bilen, and Andrea Vedaldi. Learning multiple visual domains with residual adapters. In NIPS, pages 506-516, 2017. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "spans": [ + { + "bbox": [ + 316, + 430, + 553, + 473 + ], + "type": "text", + "content": "[51] Sylvestre-Alvise Rebuffi, Alexander Kolesnikov, Georg Sperl, and Christoph H Lampert. icarl: Incremental classifier and representation learning. In CVPR, pages 2001–2010, 2017. 1, 2, 6, 7" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "spans": [ + { + "bbox": [ + 316, + 475, + 553, + 529 + ], + "type": "text", + "content": "[52] Yujun Shi, Kuangqi Zhou, Jian Liang, Zihang Jiang, Jiashi Feng, Philip HS Torr, Song Bai, and Vincent YF Tan. Mimicking the oracle: An initial phase decorrelation approach for class incremental learning. In CVPR, pages 16722-16731, 2022. 2" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 316, + 532, + 553, + 597 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 532, + 553, + 597 + ], + "spans": [ + { + "bbox": [ + 316, + 532, + 553, + 597 + ], + "type": "text", + "content": "[53] James Seale Smith, Leonid Karlinsky, Vyshnavi Gutta, Paola Cascante-Bonilla, Donghyun Kim, Assaf Arbelle, Rameswar Panda, Rogerio Feris, and Zsolt Kira. Coda-prompt: Continual decomposed attention-based prompting for rehearsal-free continual learning. In CVPR, pages 11909-11919, 2023. 2, 3, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "spans": [ + { + "bbox": [ + 316, + 600, + 553, + 632 + ], + "type": "text", + "content": "[54] Jake Snell, Kevin Swersky, and Richard Zemel. Prototypical networks for few-shot learning. In NIPS, pages 4080-4090, 2017. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "spans": [ + { + "bbox": [ + 316, + 634, + 553, + 667 + ], + "type": "text", + "content": "[55] Khurram Soomro, Amir Roshan Zamir, and Mubarak Shah. Ucf101: A dataset of 101 human actions classes from videos in the wild. arXiv preprint arXiv:1212.0402, 2012. 6" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "spans": [ + { + "bbox": [ + 316, + 669, + 553, + 712 + ], + "type": "text", + "content": "[56] Hai-Long Sun, Da-Wei Zhou, Hanbin Zhao, Le Gan, DeChuan Zhan, and Han-Jia Ye. Mos: Model surgery for pretrained model-based class-incremental learning. In AAAI, pages 20699-20707, 2025. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3323" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 9 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 712 + ], + "type": "list", + "angle": 0, + "index": 16, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 106 + ], + "type": "text", + "content": "[57] Xiaoyu Tao, Xinyuan Chang, Xiaopeng Hong, Xing Wei, and Yihong Gong. Topology-preserving class-incremental learning. In ECCV, pages 254-270, 2020. 2" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "spans": [ + { + "bbox": [ + 56, + 106, + 294, + 128 + ], + "type": "text", + "content": "[58] Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. JMLR, 9(11), 2008. 8" + } + ] + } + ], + "index": 1 + }, + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "spans": [ + { + "bbox": [ + 56, + 128, + 294, + 171 + ], + "type": "text", + "content": "[59] C. Wah, S. Branson, P. Welinder, P. Perona, and S. Belongie. The Caltech-UCSD Birds-200-2011 Dataset. Technical Report CNS-TR-2011-001, California Institute of Technology, 2011. 6" + } + ] + } + ], + "index": 2 + }, + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "spans": [ + { + "bbox": [ + 56, + 172, + 294, + 205 + ], + "type": "text", + "content": "[60] Fu-Yun Wang, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Foster: Feature boosting and compression for class incremental learning. In ECCV, pages 398–414, 2022. 2" + } + ] + } + ], + "index": 3 + }, + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "spans": [ + { + "bbox": [ + 56, + 205, + 294, + 250 + ], + "type": "text", + "content": "[61] Fu-Yun Wang, Da-Wei Zhou, Liu Liu, Han-Jia Ye, Yatao Bian, De-Chuan Zhan, and Peilin Zhao. BEEF: Bi-compatible class-incremental learning via energy-based expansion and fusion. In ICLR, 2023. 2" + } + ] + } + ], + "index": 4 + }, + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "spans": [ + { + "bbox": [ + 56, + 250, + 294, + 303 + ], + "type": "text", + "content": "[62] Runqi Wang, Xiaoyue Duan, Guoliang Kang, Jianzhuang Liu, Shaohui Lin, Songcen Xu, Jinhu Lu, and Baochang Zhang. Attriclip: A non-incremental learner for incremental knowledge learning. In CVPR, pages 3654-3663, 2023. 2" + } + ] + } + ], + "index": 5 + }, + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "spans": [ + { + "bbox": [ + 56, + 304, + 294, + 347 + ], + "type": "text", + "content": "[63] Yabin Wang, Zhiwu Huang, and Xiaopeng Hong. S-prompts learning with pre-trained transformers: An occam's razor for domain incremental learning. NeurIPS, pages 5682-5695, 2022. 2" + } + ] + } + ], + "index": 6 + }, + { + "bbox": [ + 56, + 348, + 294, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 348, + 294, + 403 + ], + "spans": [ + { + "bbox": [ + 56, + 348, + 294, + 403 + ], + "type": "text", + "content": "[64] Zifeng Wang, Zizhao Zhang, Sayna Ebrahimi, Ruoxi Sun, Han Zhang, Chen-Yu Lee, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, et al. Dualprompt: Complementary prompting for rehearsal-free continual learning. In ECCV, pages 631-648, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 7 + }, + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "spans": [ + { + "bbox": [ + 56, + 404, + 294, + 448 + ], + "type": "text", + "content": "[65] Zifeng Wang, Zizhao Zhang, Chen-Yu Lee, Han Zhang, Ruoxi Sun, Xiaqi Ren, Guolong Su, Vincent Perot, Jennifer Dy, and Tomas Pfister. Learning to prompt for continual learning. In CVPR, pages 139-149, 2022. 1, 2, 3, 6, 7" + } + ] + } + ], + "index": 8 + }, + { + "bbox": [ + 56, + 448, + 294, + 481 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 448, + 294, + 481 + ], + "spans": [ + { + "bbox": [ + 56, + 448, + 294, + 481 + ], + "type": "text", + "content": "[66] Zhengbo Wang, Jian Liang, Lijun Sheng, Ran He, Zilei Wang, and Tieniu Tan. A hard-to-beat baseline for training-free clip-based adaptation. In ICLR, 2024. 5" + } + ] + } + ], + "index": 9 + }, + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "spans": [ + { + "bbox": [ + 56, + 481, + 294, + 514 + ], + "type": "text", + "content": "[67] Yue Wu, Yinpeng Chen, Lijuan Wang, Yuancheng Ye, Zicheng Liu, Yandong Guo, and Yun Fu. Large scale incremental learning. In CVPR, pages 374-382, 2019. 2" + } + ] + } + ], + "index": 10 + }, + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "spans": [ + { + "bbox": [ + 56, + 514, + 294, + 557 + ], + "type": "text", + "content": "[68] Jianxiong Xiao, James Hays, Krista A Ehinger, Aude Oliva, and Antonio Torralba. Sun database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485-3492, 2010. 6" + } + ] + } + ], + "index": 11 + }, + { + "bbox": [ + 56, + 558, + 294, + 581 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 558, + 294, + 581 + ], + "spans": [ + { + "bbox": [ + 56, + 558, + 294, + 581 + ], + "type": "text", + "content": "[69] Ju Xu and Zhanxing Zhu. Reinforced continual learning. In NeurIPS, pages 899–908, 2018. 2" + } + ] + } + ], + "index": 12 + }, + { + "bbox": [ + 56, + 581, + 294, + 614 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 581, + 294, + 614 + ], + "spans": [ + { + "bbox": [ + 56, + 581, + 294, + 614 + ], + "type": "text", + "content": "[70] Shipeng Yan, Jiangwei Xie, and Xuming He. Der: Dynamically expandable representation for class incremental learning. In CVPR, pages 3014-3023, 2021. 2" + } + ] + } + ], + "index": 13 + }, + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "spans": [ + { + "bbox": [ + 56, + 614, + 294, + 667 + ], + "type": "text", + "content": "[71] Xu Yang, Zhangzikang Li, Haiyang Xu, Hanwang Zhang, Qinghao Ye, Chenliang Li, Ming Yan, Yu Zhang, Fei Huang, and Songfang Huang. Learning trajectory-word alignments for video-language tasks. In ICCV, pages 2504-2514, 2023. 1" + } + ] + } + ], + "index": 14 + }, + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "spans": [ + { + "bbox": [ + 56, + 669, + 294, + 712 + ], + "type": "text", + "content": "[72] Xu Yang, Yongliang Wu, Mingzhuo Yang, Haokun Chen, and Xin Geng. Exploring diverse in-context configurations for image captioning. In NeurIPS, pages 40924-40943, 2023. 1" + } + ] + } + ], + "index": 15 + } + ], + "sub_type": "ref_text" + }, + { + "bbox": [ + 315, + 72, + 553, + 713 + ], + "type": "list", + "angle": 0, + "index": 32, + "blocks": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "spans": [ + { + "bbox": [ + 315, + 72, + 553, + 106 + ], + "type": "text", + "content": "[73] Chao Yi, Lu Ren, De-Chuan Zhan, and Han-Jia Ye. Leveraging cross-modal neighbor representation for improved clip classification. In CVPR, pages 27402-27411, 2024. 2" + } + ] + } + ], + "index": 17 + }, + { + "bbox": [ + 315, + 106, + 553, + 139 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 106, + 553, + 139 + ], + "spans": [ + { + "bbox": [ + 315, + 106, + 553, + 139 + ], + "type": "text", + "content": "[74] Jaehong Yoon, Eunho Yang, Jeongtae Lee, and Sung Ju Hwang. Lifelong learning with dynamically expandable networks. In ICLR, 2018. 2" + } + ] + } + ], + "index": 18 + }, + { + "bbox": [ + 315, + 139, + 553, + 183 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 139, + 553, + 183 + ], + "spans": [ + { + "bbox": [ + 315, + 139, + 553, + 183 + ], + "type": "text", + "content": "[75] Jiahui Yu, Zirui Wang, Vijay Vasudevan, Legg Yeung, Mojtaba Seyedhosseini, and Yonghui Wu. Coca: Contrastive captioners are image-text foundation models. Transactions on Machine Learning Research, 2024. 1" + } + ] + } + ], + "index": 19 + }, + { + "bbox": [ + 315, + 183, + 553, + 228 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 183, + 553, + 228 + ], + "spans": [ + { + "bbox": [ + 315, + 183, + 553, + 228 + ], + "type": "text", + "content": "[76] Jiazuo Yu, Yunzhi Zhuge, Lu Zhang, Ping Hu, Dong Wang, Huchuan Lu, and You He. Boosting continual learning of vision-language models via mixture-of-experts adapters. In CVPR, pages 23219-23230, 2024. 2, 3" + } + ] + } + ], + "index": 20 + }, + { + "bbox": [ + 315, + 228, + 553, + 271 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 228, + 553, + 271 + ], + "spans": [ + { + "bbox": [ + 315, + 228, + 553, + 271 + ], + "type": "text", + "content": "[77] Lu Yu, Bartlomiej Twardowski, Xialei Liu, Luis Herranz, Kai Wang, Yongmei Cheng, Shangling Jui, and Joost van de Weijer. Semantic drift compensation for class-incremental learning. In CVPR, pages 6982-6991, 2020. 2" + } + ] + } + ], + "index": 21 + }, + { + "bbox": [ + 315, + 271, + 553, + 326 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 271, + 553, + 326 + ], + "spans": [ + { + "bbox": [ + 315, + 271, + 553, + 326 + ], + "type": "text", + "content": "[78] Lu Yuan, Dongdong Chen, Yi-Ling Chen, Noel Codella, Xiyang Dai, Jianfeng Gao, Houdong Hu, Xuedong Huang, Boxin Li, Chunyuan Li, et al. Florence: A new foundation model for computer vision. arXiv preprint arXiv:2111.11432, 2021. 1" + } + ] + } + ], + "index": 22 + }, + { + "bbox": [ + 315, + 326, + 553, + 358 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 326, + 553, + 358 + ], + "spans": [ + { + "bbox": [ + 315, + 326, + 553, + 358 + ], + "type": "text", + "content": "[79] Friedemann Zenke, Ben Poole, and Surya Ganguli. Continual learning through synaptic intelligence. In ICML, pages 3987-3995, 2017. 2" + } + ] + } + ], + "index": 23 + }, + { + "bbox": [ + 315, + 359, + 553, + 403 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 359, + 553, + 403 + ], + "spans": [ + { + "bbox": [ + 315, + 359, + 553, + 403 + ], + "type": "text", + "content": "[80] Bowen Zhao, Xi Xiao, Guojun Gan, Bin Zhang, and Shu-Tao Xia. Maintaining discrimination and fairness in class incremental learning. In CVPR, pages 13208-13217, 2020. 2" + } + ] + } + ], + "index": 24 + }, + { + "bbox": [ + 315, + 403, + 553, + 447 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 403, + 553, + 447 + ], + "spans": [ + { + "bbox": [ + 315, + 403, + 553, + 447 + ], + "type": "text", + "content": "[81] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Multi-layer rehearsal feature augmentation for class incremental learning. In ICML, pages 61649–61663, 2024. 2" + } + ] + } + ], + "index": 25 + }, + { + "bbox": [ + 315, + 448, + 553, + 491 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 448, + 553, + 491 + ], + "spans": [ + { + "bbox": [ + 315, + 448, + 553, + 491 + ], + "type": "text", + "content": "[82] Bowen Zheng, Da-Wei Zhou, Han-Jia Ye, and De-Chuan Zhan. Task-agnostic guided feature expansion for class incremental learning. In CVPR, pages 10099-10109, 2025. 2" + } + ] + } + ], + "index": 26 + }, + { + "bbox": [ + 315, + 491, + 553, + 525 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 491, + 553, + 525 + ], + "spans": [ + { + "bbox": [ + 315, + 491, + 553, + 525 + ], + "type": "text", + "content": "[83] Da-Wei Zhou, Qi-Wei Wang, Han-Jia Ye, and De-Chuan Zhan. A model or 603 exemplars: Towards memory-efficient class-incremental learning. In ICLR, 2023. 2, 6" + } + ] + } + ], + "index": 27 + }, + { + "bbox": [ + 315, + 525, + 553, + 580 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 525, + 553, + 580 + ], + "spans": [ + { + "bbox": [ + 315, + 525, + 553, + 580 + ], + "type": "text", + "content": "[84] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Revisiting class-incremental learning with pretrained models: Generalizability and adaptivity are all you need. International Journal of Computer Vision, 133:1012-1032, 2025. 2, 6, 7" + } + ] + } + ], + "index": 28 + }, + { + "bbox": [ + 315, + 580, + 553, + 624 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 580, + 553, + 624 + ], + "spans": [ + { + "bbox": [ + 315, + 580, + 553, + 624 + ], + "type": "text", + "content": "[85] Da-Wei Zhou, Zi-Wen Cai, Han-Jia Ye, Lijun Zhang, and De-Chuan Zhan. Dual consolidation for pre-trained model-based domain-incremental learning. In CVPR, pages 20547–20557, 2025. 2" + } + ] + } + ], + "index": 29 + }, + { + "bbox": [ + 315, + 624, + 553, + 679 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 624, + 553, + 679 + ], + "spans": [ + { + "bbox": [ + 315, + 624, + 553, + 679 + ], + "type": "text", + "content": "[86] Da-Wei Zhou, Yuanhan Zhang, Yan Wang, Jingyi Ning, Han-Jia Ye, De-Chuan Zhan, and Ziwei Liu. Learning without forgetting for vision-language models. IEEE Transactions on Pattern Analysis and Machine Intelligence, 47(6): 4489-4504, 2025. 2, 3, 5, 6, 7" + } + ] + } + ], + "index": 30 + }, + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "spans": [ + { + "bbox": [ + 315, + 680, + 553, + 713 + ], + "type": "text", + "content": "[87] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Zwei Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816-16825, 2022. 2" + } + ] + } + ], + "index": 31 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 748, + 315, + 757 + ], + "type": "text", + "content": "3324" + } + ] + } + ], + "index": 33 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 10 + }, + { + "para_blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 140 + ], + "type": "list", + "angle": 0, + "index": 2, + "blocks": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "spans": [ + { + "bbox": [ + 56, + 72, + 294, + 105 + ], + "type": "text", + "content": "[88] Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models. IJCV, 130(9):2337-2348, 2022. 2, 3, 6, 7" + } + ] + } + ], + "index": 0 + }, + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "ref_text", + "angle": 0, + "lines": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "spans": [ + { + "bbox": [ + 56, + 107, + 294, + 140 + ], + "type": "text", + "content": "[89] Fei Zhu, Xu-Yao Zhang, Chuang Wang, Fei Yin, and ChengLin Liu. Prototype augmentation and self-supervision for incremental learning. In CVPR, pages 5871-5880, 2021. 2" + } + ] + } + ], + "index": 1 + } + ], + "sub_type": "ref_text" + } + ], + "discarded_blocks": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "page_number", + "angle": 0, + "lines": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "spans": [ + { + "bbox": [ + 295, + 749, + 315, + 757 + ], + "type": "text", + "content": "3325" + } + ] + } + ], + "index": 3 + } + ], + "page_size": [ + 612, + 792 + ], + "page_idx": 11 + } + ], + "_backend": "vlm", + "_version_name": "2.6.4" +} \ No newline at end of file